1 /* Subroutines used for code generation on IBM RS/6000.
2 Copyright (C) 1991, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
3 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007
4 Free Software Foundation, Inc.
5 Contributed by Richard Kenner ([email protected])
6
7 This file is part of GCC.
8
9 GCC is free software; you can redistribute it and/or modify it
10 under the terms of the GNU General Public License as published
11 by the Free Software Foundation; either version 2, or (at your
12 option) any later version.
13
14 GCC is distributed in the hope that it will be useful, but WITHOUT
15 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
16 or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public
17 License for more details.
18
19 You should have received a copy of the GNU General Public License
20 along with GCC; see the file COPYING. If not, write to the
21 Free Software Foundation, 51 Franklin Street, Fifth Floor, Boston,
22 MA 02110-1301, USA. */
23
24 #include "config.h"
25 #include "system.h"
26 #include "coretypes.h"
27 #include "tm.h"
28 #include "rtl.h"
29 #include "regs.h"
30 #include "hard-reg-set.h"
31 #include "real.h"
32 #include "insn-config.h"
33 #include "conditions.h"
34 #include "insn-attr.h"
35 #include "flags.h"
36 #include "recog.h"
37 #include "obstack.h"
38 #include "tree.h"
39 #include "expr.h"
40 #include "optabs.h"
41 #include "except.h"
42 #include "function.h"
43 #include "output.h"
44 #include "basic-block.h"
45 #include "integrate.h"
46 #include "toplev.h"
47 #include "ggc.h"
48 #include "hashtab.h"
49 #include "tm_p.h"
50 #include "target.h"
51 #include "target-def.h"
52 #include "langhooks.h"
53 #include "reload.h"
54 #include "cfglayout.h"
55 #include "sched-int.h"
56 #include "tree-gimple.h"
57 #include "intl.h"
58 #include "params.h"
59 #include "tm-constrs.h"
60 #if TARGET_XCOFF
61 #include "xcoffout.h" /* get declarations of xcoff_*_section_name */
62 #endif
63 #if TARGET_MACHO
64 #include "gstab.h" /* for N_SLINE */
65 #endif
66
67 #ifndef TARGET_NO_PROTOTYPE
68 #define TARGET_NO_PROTOTYPE 0
69 #endif
70
71 #define min(A,B) ((A) < (B) ? (A) : (B))
72 #define max(A,B) ((A) > (B) ? (A) : (B))
73
74 /* Structure used to define the rs6000 stack */
75 typedef struct rs6000_stack {
76 int first_gp_reg_save; /* first callee saved GP register used */
77 int first_fp_reg_save; /* first callee saved FP register used */
78 int first_altivec_reg_save; /* first callee saved AltiVec register used */
79 int lr_save_p; /* true if the link reg needs to be saved */
80 int cr_save_p; /* true if the CR reg needs to be saved */
81 unsigned int vrsave_mask; /* mask of vec registers to save */
82 int push_p; /* true if we need to allocate stack space */
83 int calls_p; /* true if the function makes any calls */
84 int world_save_p; /* true if we're saving *everything*:
85 r13-r31, cr, f14-f31, vrsave, v20-v31 */
86 enum rs6000_abi abi; /* which ABI to use */
87 int gp_save_offset; /* offset to save GP regs from initial SP */
88 int fp_save_offset; /* offset to save FP regs from initial SP */
89 int altivec_save_offset; /* offset to save AltiVec regs from initial SP */
90 int lr_save_offset; /* offset to save LR from initial SP */
91 int cr_save_offset; /* offset to save CR from initial SP */
92 int vrsave_save_offset; /* offset to save VRSAVE from initial SP */
93 int spe_gp_save_offset; /* offset to save spe 64-bit gprs */
94 int varargs_save_offset; /* offset to save the varargs registers */
95 int ehrd_offset; /* offset to EH return data */
96 int reg_size; /* register size (4 or 8) */
97 HOST_WIDE_INT vars_size; /* variable save area size */
98 int parm_size; /* outgoing parameter size */
99 int save_size; /* save area size */
100 int fixed_size; /* fixed size of stack frame */
101 int gp_size; /* size of saved GP registers */
102 int fp_size; /* size of saved FP registers */
103 int altivec_size; /* size of saved AltiVec registers */
104 int cr_size; /* size to hold CR if not in save_size */
105 int vrsave_size; /* size to hold VRSAVE if not in save_size */
106 int altivec_padding_size; /* size of altivec alignment padding if
107 not in save_size */
108 int spe_gp_size; /* size of 64-bit GPR save size for SPE */
109 int spe_padding_size;
110 HOST_WIDE_INT total_size; /* total bytes allocated for stack */
111 int spe_64bit_regs_used;
112 } rs6000_stack_t;
113
114 /* A C structure for machine-specific, per-function data.
115 This is added to the cfun structure. */
116 typedef struct machine_function GTY(())
117 {
118 /* Flags if __builtin_return_address (n) with n >= 1 was used. */
119 int ra_needs_full_frame;
120 /* Some local-dynamic symbol. */
121 const char *some_ld_name;
122 /* Whether the instruction chain has been scanned already. */
123 int insn_chain_scanned_p;
124 /* Flags if __builtin_return_address (0) was used. */
125 int ra_need_lr;
126 /* Offset from virtual_stack_vars_rtx to the start of the ABI_V4
127 varargs save area. */
128 HOST_WIDE_INT varargs_save_offset;
129 } machine_function;
130
131 /* Target cpu type */
132
133 enum processor_type rs6000_cpu;
134 struct rs6000_cpu_select rs6000_select[3] =
135 {
136 /* switch name, tune arch */
137 { (const char *)0, "--with-cpu=", 1, 1 },
138 { (const char *)0, "-mcpu=", 1, 1 },
139 { (const char *)0, "-mtune=", 1, 0 },
140 };
141
142 /* Always emit branch hint bits. */
143 static GTY(()) bool rs6000_always_hint;
144
145 /* Schedule instructions for group formation. */
146 static GTY(()) bool rs6000_sched_groups;
147
148 /* Support for -msched-costly-dep option. */
149 const char *rs6000_sched_costly_dep_str;
150 enum rs6000_dependence_cost rs6000_sched_costly_dep;
151
152 /* Support for -minsert-sched-nops option. */
153 const char *rs6000_sched_insert_nops_str;
154 enum rs6000_nop_insertion rs6000_sched_insert_nops;
155
156 /* Support targetm.vectorize.builtin_mask_for_load. */
157 static GTY(()) tree altivec_builtin_mask_for_load;
158
159 /* Size of long double. */
160 int rs6000_long_double_type_size;
161
162 /* IEEE quad extended precision long double. */
163 int rs6000_ieeequad;
164
165 /* Whether -mabi=altivec has appeared. */
166 int rs6000_altivec_abi;
167
168 /* Nonzero if we want SPE ABI extensions. */
169 int rs6000_spe_abi;
170
171 /* Nonzero if floating point operations are done in the GPRs. */
172 int rs6000_float_gprs = 0;
173
174 /* Nonzero if we want Darwin's struct-by-value-in-regs ABI. */
175 int rs6000_darwin64_abi;
176
177 /* Set to nonzero once AIX common-mode calls have been defined. */
178 static GTY(()) int common_mode_defined;
179
180 /* Save information from a "cmpxx" operation until the branch or scc is
181 emitted. */
182 rtx rs6000_compare_op0, rs6000_compare_op1;
183 int rs6000_compare_fp_p;
184
185 /* Label number of label created for -mrelocatable, to call to so we can
186 get the address of the GOT section */
187 int rs6000_pic_labelno;
188
189 #ifdef USING_ELFOS_H
190 /* Which abi to adhere to */
191 const char *rs6000_abi_name;
192
193 /* Semantics of the small data area */
194 enum rs6000_sdata_type rs6000_sdata = SDATA_DATA;
195
196 /* Which small data model to use */
197 const char *rs6000_sdata_name = (char *)0;
198
199 /* Counter for labels which are to be placed in .fixup. */
200 int fixuplabelno = 0;
201 #endif
202
203 /* Bit size of immediate TLS offsets and string from which it is decoded. */
204 int rs6000_tls_size = 32;
205 const char *rs6000_tls_size_string;
206
207 /* ABI enumeration available for subtarget to use. */
208 enum rs6000_abi rs6000_current_abi;
209
210 /* Whether to use variant of AIX ABI for PowerPC64 Linux. */
211 int dot_symbols;
212
213 /* Debug flags */
214 const char *rs6000_debug_name;
215 int rs6000_debug_stack; /* debug stack applications */
216 int rs6000_debug_arg; /* debug argument handling */
217
218 /* Value is TRUE if register/mode pair is acceptable. */
219 bool rs6000_hard_regno_mode_ok_p[NUM_MACHINE_MODES][FIRST_PSEUDO_REGISTER];
220
221 /* Built in types. */
222
223 tree rs6000_builtin_types[RS6000_BTI_MAX];
224 tree rs6000_builtin_decls[RS6000_BUILTIN_COUNT];
225
226 const char *rs6000_traceback_name;
227 static enum {
228 traceback_default = 0,
229 traceback_none,
230 traceback_part,
231 traceback_full
232 } rs6000_traceback;
233
234 /* Flag to say the TOC is initialized */
235 int toc_initialized;
236 char toc_label_name[10];
237
238 static GTY(()) section *read_only_data_section;
239 static GTY(()) section *private_data_section;
240 static GTY(()) section *read_only_private_data_section;
241 static GTY(()) section *sdata2_section;
242 static GTY(()) section *toc_section;
243
244 /* Control alignment for fields within structures. */
245 /* String from -malign-XXXXX. */
246 int rs6000_alignment_flags;
247
248 /* True for any options that were explicitly set. */
249 struct {
250 bool aix_struct_ret; /* True if -maix-struct-ret was used. */
251 bool alignment; /* True if -malign- was used. */
252 bool abi; /* True if -mabi=spe/nospe was used. */
253 bool spe; /* True if -mspe= was used. */
254 bool float_gprs; /* True if -mfloat-gprs= was used. */
255 bool isel; /* True if -misel was used. */
256 bool long_double; /* True if -mlong-double- was used. */
257 bool ieee; /* True if -mabi=ieee/ibmlongdouble used. */
258 } rs6000_explicit_options;
259
260 struct builtin_description
261 {
262 /* mask is not const because we're going to alter it below. This
263 nonsense will go away when we rewrite the -march infrastructure
264 to give us more target flag bits. */
265 unsigned int mask;
266 const enum insn_code icode;
267 const char *const name;
268 const enum rs6000_builtins code;
269 };
270
271 /* Target cpu costs. */
272
273 struct processor_costs {
274 const int mulsi; /* cost of SImode multiplication. */
275 const int mulsi_const; /* cost of SImode multiplication by constant. */
276 const int mulsi_const9; /* cost of SImode mult by short constant. */
277 const int muldi; /* cost of DImode multiplication. */
278 const int divsi; /* cost of SImode division. */
279 const int divdi; /* cost of DImode division. */
280 const int fp; /* cost of simple SFmode and DFmode insns. */
281 const int dmul; /* cost of DFmode multiplication (and fmadd). */
282 const int sdiv; /* cost of SFmode division (fdivs). */
283 const int ddiv; /* cost of DFmode division (fdiv). */
284 };
285
286 const struct processor_costs *rs6000_cost;
287
288 /* Processor costs (relative to an add) */
289
290 /* Instruction size costs on 32bit processors. */
291 static const
292 struct processor_costs size32_cost = {
293 COSTS_N_INSNS (1), /* mulsi */
294 COSTS_N_INSNS (1), /* mulsi_const */
295 COSTS_N_INSNS (1), /* mulsi_const9 */
296 COSTS_N_INSNS (1), /* muldi */
297 COSTS_N_INSNS (1), /* divsi */
298 COSTS_N_INSNS (1), /* divdi */
299 COSTS_N_INSNS (1), /* fp */
300 COSTS_N_INSNS (1), /* dmul */
301 COSTS_N_INSNS (1), /* sdiv */
302 COSTS_N_INSNS (1), /* ddiv */
303 };
304
305 /* Instruction size costs on 64bit processors. */
306 static const
307 struct processor_costs size64_cost = {
308 COSTS_N_INSNS (1), /* mulsi */
309 COSTS_N_INSNS (1), /* mulsi_const */
310 COSTS_N_INSNS (1), /* mulsi_const9 */
311 COSTS_N_INSNS (1), /* muldi */
312 COSTS_N_INSNS (1), /* divsi */
313 COSTS_N_INSNS (1), /* divdi */
314 COSTS_N_INSNS (1), /* fp */
315 COSTS_N_INSNS (1), /* dmul */
316 COSTS_N_INSNS (1), /* sdiv */
317 COSTS_N_INSNS (1), /* ddiv */
318 };
319
320 /* Instruction costs on RIOS1 processors. */
321 static const
322 struct processor_costs rios1_cost = {
323 COSTS_N_INSNS (5), /* mulsi */
324 COSTS_N_INSNS (4), /* mulsi_const */
325 COSTS_N_INSNS (3), /* mulsi_const9 */
326 COSTS_N_INSNS (5), /* muldi */
327 COSTS_N_INSNS (19), /* divsi */
328 COSTS_N_INSNS (19), /* divdi */
329 COSTS_N_INSNS (2), /* fp */
330 COSTS_N_INSNS (2), /* dmul */
331 COSTS_N_INSNS (19), /* sdiv */
332 COSTS_N_INSNS (19), /* ddiv */
333 };
334
335 /* Instruction costs on RIOS2 processors. */
336 static const
337 struct processor_costs rios2_cost = {
338 COSTS_N_INSNS (2), /* mulsi */
339 COSTS_N_INSNS (2), /* mulsi_const */
340 COSTS_N_INSNS (2), /* mulsi_const9 */
341 COSTS_N_INSNS (2), /* muldi */
342 COSTS_N_INSNS (13), /* divsi */
343 COSTS_N_INSNS (13), /* divdi */
344 COSTS_N_INSNS (2), /* fp */
345 COSTS_N_INSNS (2), /* dmul */
346 COSTS_N_INSNS (17), /* sdiv */
347 COSTS_N_INSNS (17), /* ddiv */
348 };
349
350 /* Instruction costs on RS64A processors. */
351 static const
352 struct processor_costs rs64a_cost = {
353 COSTS_N_INSNS (20), /* mulsi */
354 COSTS_N_INSNS (12), /* mulsi_const */
355 COSTS_N_INSNS (8), /* mulsi_const9 */
356 COSTS_N_INSNS (34), /* muldi */
357 COSTS_N_INSNS (65), /* divsi */
358 COSTS_N_INSNS (67), /* divdi */
359 COSTS_N_INSNS (4), /* fp */
360 COSTS_N_INSNS (4), /* dmul */
361 COSTS_N_INSNS (31), /* sdiv */
362 COSTS_N_INSNS (31), /* ddiv */
363 };
364
365 /* Instruction costs on MPCCORE processors. */
366 static const
367 struct processor_costs mpccore_cost = {
368 COSTS_N_INSNS (2), /* mulsi */
369 COSTS_N_INSNS (2), /* mulsi_const */
370 COSTS_N_INSNS (2), /* mulsi_const9 */
371 COSTS_N_INSNS (2), /* muldi */
372 COSTS_N_INSNS (6), /* divsi */
373 COSTS_N_INSNS (6), /* divdi */
374 COSTS_N_INSNS (4), /* fp */
375 COSTS_N_INSNS (5), /* dmul */
376 COSTS_N_INSNS (10), /* sdiv */
377 COSTS_N_INSNS (17), /* ddiv */
378 };
379
380 /* Instruction costs on PPC403 processors. */
381 static const
382 struct processor_costs ppc403_cost = {
383 COSTS_N_INSNS (4), /* mulsi */
384 COSTS_N_INSNS (4), /* mulsi_const */
385 COSTS_N_INSNS (4), /* mulsi_const9 */
386 COSTS_N_INSNS (4), /* muldi */
387 COSTS_N_INSNS (33), /* divsi */
388 COSTS_N_INSNS (33), /* divdi */
389 COSTS_N_INSNS (11), /* fp */
390 COSTS_N_INSNS (11), /* dmul */
391 COSTS_N_INSNS (11), /* sdiv */
392 COSTS_N_INSNS (11), /* ddiv */
393 };
394
395 /* Instruction costs on PPC405 processors. */
396 static const
397 struct processor_costs ppc405_cost = {
398 COSTS_N_INSNS (5), /* mulsi */
399 COSTS_N_INSNS (4), /* mulsi_const */
400 COSTS_N_INSNS (3), /* mulsi_const9 */
401 COSTS_N_INSNS (5), /* muldi */
402 COSTS_N_INSNS (35), /* divsi */
403 COSTS_N_INSNS (35), /* divdi */
404 COSTS_N_INSNS (11), /* fp */
405 COSTS_N_INSNS (11), /* dmul */
406 COSTS_N_INSNS (11), /* sdiv */
407 COSTS_N_INSNS (11), /* ddiv */
408 };
409
410 /* Instruction costs on PPC440 processors. */
411 static const
412 struct processor_costs ppc440_cost = {
413 COSTS_N_INSNS (3), /* mulsi */
414 COSTS_N_INSNS (2), /* mulsi_const */
415 COSTS_N_INSNS (2), /* mulsi_const9 */
416 COSTS_N_INSNS (3), /* muldi */
417 COSTS_N_INSNS (34), /* divsi */
418 COSTS_N_INSNS (34), /* divdi */
419 COSTS_N_INSNS (5), /* fp */
420 COSTS_N_INSNS (5), /* dmul */
421 COSTS_N_INSNS (19), /* sdiv */
422 COSTS_N_INSNS (33), /* ddiv */
423 };
424
425 /* Instruction costs on PPC601 processors. */
426 static const
427 struct processor_costs ppc601_cost = {
428 COSTS_N_INSNS (5), /* mulsi */
429 COSTS_N_INSNS (5), /* mulsi_const */
430 COSTS_N_INSNS (5), /* mulsi_const9 */
431 COSTS_N_INSNS (5), /* muldi */
432 COSTS_N_INSNS (36), /* divsi */
433 COSTS_N_INSNS (36), /* divdi */
434 COSTS_N_INSNS (4), /* fp */
435 COSTS_N_INSNS (5), /* dmul */
436 COSTS_N_INSNS (17), /* sdiv */
437 COSTS_N_INSNS (31), /* ddiv */
438 };
439
440 /* Instruction costs on PPC603 processors. */
441 static const
442 struct processor_costs ppc603_cost = {
443 COSTS_N_INSNS (5), /* mulsi */
444 COSTS_N_INSNS (3), /* mulsi_const */
445 COSTS_N_INSNS (2), /* mulsi_const9 */
446 COSTS_N_INSNS (5), /* muldi */
447 COSTS_N_INSNS (37), /* divsi */
448 COSTS_N_INSNS (37), /* divdi */
449 COSTS_N_INSNS (3), /* fp */
450 COSTS_N_INSNS (4), /* dmul */
451 COSTS_N_INSNS (18), /* sdiv */
452 COSTS_N_INSNS (33), /* ddiv */
453 };
454
455 /* Instruction costs on PPC604 processors. */
456 static const
457 struct processor_costs ppc604_cost = {
458 COSTS_N_INSNS (4), /* mulsi */
459 COSTS_N_INSNS (4), /* mulsi_const */
460 COSTS_N_INSNS (4), /* mulsi_const9 */
461 COSTS_N_INSNS (4), /* muldi */
462 COSTS_N_INSNS (20), /* divsi */
463 COSTS_N_INSNS (20), /* divdi */
464 COSTS_N_INSNS (3), /* fp */
465 COSTS_N_INSNS (3), /* dmul */
466 COSTS_N_INSNS (18), /* sdiv */
467 COSTS_N_INSNS (32), /* ddiv */
468 };
469
470 /* Instruction costs on PPC604e processors. */
471 static const
472 struct processor_costs ppc604e_cost = {
473 COSTS_N_INSNS (2), /* mulsi */
474 COSTS_N_INSNS (2), /* mulsi_const */
475 COSTS_N_INSNS (2), /* mulsi_const9 */
476 COSTS_N_INSNS (2), /* muldi */
477 COSTS_N_INSNS (20), /* divsi */
478 COSTS_N_INSNS (20), /* divdi */
479 COSTS_N_INSNS (3), /* fp */
480 COSTS_N_INSNS (3), /* dmul */
481 COSTS_N_INSNS (18), /* sdiv */
482 COSTS_N_INSNS (32), /* ddiv */
483 };
484
485 /* Instruction costs on PPC620 processors. */
486 static const
487 struct processor_costs ppc620_cost = {
488 COSTS_N_INSNS (5), /* mulsi */
489 COSTS_N_INSNS (4), /* mulsi_const */
490 COSTS_N_INSNS (3), /* mulsi_const9 */
491 COSTS_N_INSNS (7), /* muldi */
492 COSTS_N_INSNS (21), /* divsi */
493 COSTS_N_INSNS (37), /* divdi */
494 COSTS_N_INSNS (3), /* fp */
495 COSTS_N_INSNS (3), /* dmul */
496 COSTS_N_INSNS (18), /* sdiv */
497 COSTS_N_INSNS (32), /* ddiv */
498 };
499
500 /* Instruction costs on PPC630 processors. */
501 static const
502 struct processor_costs ppc630_cost = {
503 COSTS_N_INSNS (5), /* mulsi */
504 COSTS_N_INSNS (4), /* mulsi_const */
505 COSTS_N_INSNS (3), /* mulsi_const9 */
506 COSTS_N_INSNS (7), /* muldi */
507 COSTS_N_INSNS (21), /* divsi */
508 COSTS_N_INSNS (37), /* divdi */
509 COSTS_N_INSNS (3), /* fp */
510 COSTS_N_INSNS (3), /* dmul */
511 COSTS_N_INSNS (17), /* sdiv */
512 COSTS_N_INSNS (21), /* ddiv */
513 };
514
515 /* Instruction costs on PPC750 and PPC7400 processors. */
516 static const
517 struct processor_costs ppc750_cost = {
518 COSTS_N_INSNS (5), /* mulsi */
519 COSTS_N_INSNS (3), /* mulsi_const */
520 COSTS_N_INSNS (2), /* mulsi_const9 */
521 COSTS_N_INSNS (5), /* muldi */
522 COSTS_N_INSNS (17), /* divsi */
523 COSTS_N_INSNS (17), /* divdi */
524 COSTS_N_INSNS (3), /* fp */
525 COSTS_N_INSNS (3), /* dmul */
526 COSTS_N_INSNS (17), /* sdiv */
527 COSTS_N_INSNS (31), /* ddiv */
528 };
529
530 /* Instruction costs on PPC7450 processors. */
531 static const
532 struct processor_costs ppc7450_cost = {
533 COSTS_N_INSNS (4), /* mulsi */
534 COSTS_N_INSNS (3), /* mulsi_const */
535 COSTS_N_INSNS (3), /* mulsi_const9 */
536 COSTS_N_INSNS (4), /* muldi */
537 COSTS_N_INSNS (23), /* divsi */
538 COSTS_N_INSNS (23), /* divdi */
539 COSTS_N_INSNS (5), /* fp */
540 COSTS_N_INSNS (5), /* dmul */
541 COSTS_N_INSNS (21), /* sdiv */
542 COSTS_N_INSNS (35), /* ddiv */
543 };
544
545 /* Instruction costs on PPC8540 processors. */
546 static const
547 struct processor_costs ppc8540_cost = {
548 COSTS_N_INSNS (4), /* mulsi */
549 COSTS_N_INSNS (4), /* mulsi_const */
550 COSTS_N_INSNS (4), /* mulsi_const9 */
551 COSTS_N_INSNS (4), /* muldi */
552 COSTS_N_INSNS (19), /* divsi */
553 COSTS_N_INSNS (19), /* divdi */
554 COSTS_N_INSNS (4), /* fp */
555 COSTS_N_INSNS (4), /* dmul */
556 COSTS_N_INSNS (29), /* sdiv */
557 COSTS_N_INSNS (29), /* ddiv */
558 };
559
560 /* Instruction costs on POWER4 and POWER5 processors. */
561 static const
562 struct processor_costs power4_cost = {
563 COSTS_N_INSNS (3), /* mulsi */
564 COSTS_N_INSNS (2), /* mulsi_const */
565 COSTS_N_INSNS (2), /* mulsi_const9 */
566 COSTS_N_INSNS (4), /* muldi */
567 COSTS_N_INSNS (18), /* divsi */
568 COSTS_N_INSNS (34), /* divdi */
569 COSTS_N_INSNS (3), /* fp */
570 COSTS_N_INSNS (3), /* dmul */
571 COSTS_N_INSNS (17), /* sdiv */
572 COSTS_N_INSNS (17), /* ddiv */
573 };
574
575
576 static bool rs6000_function_ok_for_sibcall (tree, tree);
577 static const char *rs6000_invalid_within_doloop (rtx);
578 static rtx rs6000_generate_compare (enum rtx_code);
579 static void rs6000_maybe_dead (rtx);
580 static void rs6000_emit_stack_tie (void);
581 static void rs6000_frame_related (rtx, rtx, HOST_WIDE_INT, rtx, rtx);
582 static rtx spe_synthesize_frame_save (rtx);
583 static bool spe_func_has_64bit_regs_p (void);
584 static void emit_frame_save (rtx, rtx, enum machine_mode, unsigned int,
585 int, HOST_WIDE_INT);
586 static rtx gen_frame_mem_offset (enum machine_mode, rtx, int);
587 static void rs6000_emit_allocate_stack (HOST_WIDE_INT, int);
588 static unsigned rs6000_hash_constant (rtx);
589 static unsigned toc_hash_function (const void *);
590 static int toc_hash_eq (const void *, const void *);
591 static int constant_pool_expr_1 (rtx, int *, int *);
592 static bool constant_pool_expr_p (rtx);
593 static bool legitimate_small_data_p (enum machine_mode, rtx);
594 static bool legitimate_indexed_address_p (rtx, int);
595 static bool legitimate_lo_sum_address_p (enum machine_mode, rtx, int);
596 static struct machine_function * rs6000_init_machine_status (void);
597 static bool rs6000_assemble_integer (rtx, unsigned int, int);
598 static bool no_global_regs_above (int);
599 #ifdef HAVE_GAS_HIDDEN
600 static void rs6000_assemble_visibility (tree, int);
601 #endif
602 static int rs6000_ra_ever_killed (void);
603 static tree rs6000_handle_longcall_attribute (tree *, tree, tree, int, bool *);
604 static tree rs6000_handle_altivec_attribute (tree *, tree, tree, int, bool *);
605 static bool rs6000_ms_bitfield_layout_p (tree);
606 static tree rs6000_handle_struct_attribute (tree *, tree, tree, int, bool *);
607 static void rs6000_eliminate_indexed_memrefs (rtx operands[2]);
608 static const char *rs6000_mangle_fundamental_type (tree);
609 extern const struct attribute_spec rs6000_attribute_table[];
610 static void rs6000_set_default_type_attributes (tree);
611 static void rs6000_output_function_prologue (FILE *, HOST_WIDE_INT);
612 static void rs6000_output_function_epilogue (FILE *, HOST_WIDE_INT);
613 static void rs6000_output_mi_thunk (FILE *, tree, HOST_WIDE_INT, HOST_WIDE_INT,
614 tree);
615 static rtx rs6000_emit_set_long_const (rtx, HOST_WIDE_INT, HOST_WIDE_INT);
616 static bool rs6000_return_in_memory (tree, tree);
617 static void rs6000_file_start (void);
618 #if TARGET_ELF
619 static int rs6000_elf_reloc_rw_mask (void);
620 static void rs6000_elf_asm_out_constructor (rtx, int);
621 static void rs6000_elf_asm_out_destructor (rtx, int);
622 static void rs6000_elf_end_indicate_exec_stack (void) ATTRIBUTE_UNUSED;
623 static void rs6000_elf_asm_init_sections (void);
624 static section *rs6000_elf_select_rtx_section (enum machine_mode, rtx,
625 unsigned HOST_WIDE_INT);
626 static void rs6000_elf_encode_section_info (tree, rtx, int)
627 ATTRIBUTE_UNUSED;
628 #endif
629 static bool rs6000_use_blocks_for_constant_p (enum machine_mode, rtx);
630 #if TARGET_XCOFF
631 static void rs6000_xcoff_asm_output_anchor (rtx);
632 static void rs6000_xcoff_asm_globalize_label (FILE *, const char *);
633 static void rs6000_xcoff_asm_init_sections (void);
634 static int rs6000_xcoff_reloc_rw_mask (void);
635 static void rs6000_xcoff_asm_named_section (const char *, unsigned int, tree);
636 static section *rs6000_xcoff_select_section (tree, int,
637 unsigned HOST_WIDE_INT);
638 static void rs6000_xcoff_unique_section (tree, int);
639 static section *rs6000_xcoff_select_rtx_section
640 (enum machine_mode, rtx, unsigned HOST_WIDE_INT);
641 static const char * rs6000_xcoff_strip_name_encoding (const char *);
642 static unsigned int rs6000_xcoff_section_type_flags (tree, const char *, int);
643 static void rs6000_xcoff_file_start (void);
644 static void rs6000_xcoff_file_end (void);
645 #endif
646 static int rs6000_variable_issue (FILE *, int, rtx, int);
647 static bool rs6000_rtx_costs (rtx, int, int, int *);
648 static int rs6000_adjust_cost (rtx, rtx, rtx, int);
649 static bool is_microcoded_insn (rtx);
650 static int is_dispatch_slot_restricted (rtx);
651 static bool is_cracked_insn (rtx);
652 static bool is_branch_slot_insn (rtx);
653 static int rs6000_adjust_priority (rtx, int);
654 static int rs6000_issue_rate (void);
655 static bool rs6000_is_costly_dependence (rtx, rtx, rtx, int, int);
656 static rtx get_next_active_insn (rtx, rtx);
657 static bool insn_terminates_group_p (rtx , enum group_termination);
658 static bool is_costly_group (rtx *, rtx);
659 static int force_new_group (int, FILE *, rtx *, rtx, bool *, int, int *);
660 static int redefine_groups (FILE *, int, rtx, rtx);
661 static int pad_groups (FILE *, int, rtx, rtx);
662 static void rs6000_sched_finish (FILE *, int);
663 static int rs6000_use_sched_lookahead (void);
664 static tree rs6000_builtin_mask_for_load (void);
665
666 static void def_builtin (int, const char *, tree, int);
667 static bool rs6000_vector_alignment_reachable (tree, bool);
668 static void rs6000_init_builtins (void);
669 static rtx rs6000_expand_unop_builtin (enum insn_code, tree, rtx);
670 static rtx rs6000_expand_binop_builtin (enum insn_code, tree, rtx);
671 static rtx rs6000_expand_ternop_builtin (enum insn_code, tree, rtx);
672 static rtx rs6000_expand_builtin (tree, rtx, rtx, enum machine_mode, int);
673 static void altivec_init_builtins (void);
674 static void rs6000_common_init_builtins (void);
675 static void rs6000_init_libfuncs (void);
676
677 static void enable_mask_for_builtins (struct builtin_description *, int,
678 enum rs6000_builtins,
679 enum rs6000_builtins);
680 static tree build_opaque_vector_type (tree, int);
681 static void spe_init_builtins (void);
682 static rtx spe_expand_builtin (tree, rtx, bool *);
683 static rtx spe_expand_stv_builtin (enum insn_code, tree);
684 static rtx spe_expand_predicate_builtin (enum insn_code, tree, rtx);
685 static rtx spe_expand_evsel_builtin (enum insn_code, tree, rtx);
686 static int rs6000_emit_int_cmove (rtx, rtx, rtx, rtx);
687 static rs6000_stack_t *rs6000_stack_info (void);
688 static void debug_stack_info (rs6000_stack_t *);
689
690 static rtx altivec_expand_builtin (tree, rtx, bool *);
691 static rtx altivec_expand_ld_builtin (tree, rtx, bool *);
692 static rtx altivec_expand_st_builtin (tree, rtx, bool *);
693 static rtx altivec_expand_dst_builtin (tree, rtx, bool *);
694 static rtx altivec_expand_abs_builtin (enum insn_code, tree, rtx);
695 static rtx altivec_expand_predicate_builtin (enum insn_code,
696 const char *, tree, rtx);
697 static rtx altivec_expand_lv_builtin (enum insn_code, tree, rtx);
698 static rtx altivec_expand_stv_builtin (enum insn_code, tree);
699 static rtx altivec_expand_vec_init_builtin (tree, tree, rtx);
700 static rtx altivec_expand_vec_set_builtin (tree);
701 static rtx altivec_expand_vec_ext_builtin (tree, rtx);
702 static int get_element_number (tree, tree);
703 static bool rs6000_handle_option (size_t, const char *, int);
704 static void rs6000_parse_tls_size_option (void);
705 static void rs6000_parse_yes_no_option (const char *, const char *, int *);
706 static int first_altivec_reg_to_save (void);
707 static unsigned int compute_vrsave_mask (void);
708 static void compute_save_world_info (rs6000_stack_t *info_ptr);
709 static void is_altivec_return_reg (rtx, void *);
710 static rtx generate_set_vrsave (rtx, rs6000_stack_t *, int);
711 int easy_vector_constant (rtx, enum machine_mode);
712 static bool rs6000_is_opaque_type (tree);
713 static rtx rs6000_dwarf_register_span (rtx);
714 static rtx rs6000_legitimize_tls_address (rtx, enum tls_model);
715 static void rs6000_output_dwarf_dtprel (FILE *, int, rtx) ATTRIBUTE_UNUSED;
716 static rtx rs6000_tls_get_addr (void);
717 static rtx rs6000_got_sym (void);
718 static int rs6000_tls_symbol_ref_1 (rtx *, void *);
719 static const char *rs6000_get_some_local_dynamic_name (void);
720 static int rs6000_get_some_local_dynamic_name_1 (rtx *, void *);
721 static rtx rs6000_complex_function_value (enum machine_mode);
722 static rtx rs6000_spe_function_arg (CUMULATIVE_ARGS *,
723 enum machine_mode, tree);
724 static void rs6000_darwin64_record_arg_advance_flush (CUMULATIVE_ARGS *,
725 HOST_WIDE_INT);
726 static void rs6000_darwin64_record_arg_advance_recurse (CUMULATIVE_ARGS *,
727 tree, HOST_WIDE_INT);
728 static void rs6000_darwin64_record_arg_flush (CUMULATIVE_ARGS *,
729 HOST_WIDE_INT,
730 rtx[], int *);
731 static void rs6000_darwin64_record_arg_recurse (CUMULATIVE_ARGS *,
732 tree, HOST_WIDE_INT,
733 rtx[], int *);
734 static rtx rs6000_darwin64_record_arg (CUMULATIVE_ARGS *, tree, int, bool);
735 static rtx rs6000_mixed_function_arg (enum machine_mode, tree, int);
736 static void rs6000_move_block_from_reg (int regno, rtx x, int nregs);
737 static void setup_incoming_varargs (CUMULATIVE_ARGS *,
738 enum machine_mode, tree,
739 int *, int);
740 static bool rs6000_pass_by_reference (CUMULATIVE_ARGS *, enum machine_mode,
741 tree, bool);
742 static int rs6000_arg_partial_bytes (CUMULATIVE_ARGS *, enum machine_mode,
743 tree, bool);
744 static const char *invalid_arg_for_unprototyped_fn (tree, tree, tree);
745 #if TARGET_MACHO
746 static void macho_branch_islands (void);
747 static int no_previous_def (tree function_name);
748 static tree get_prev_label (tree function_name);
749 static void rs6000_darwin_file_start (void);
750 #endif
751
752 static tree rs6000_build_builtin_va_list (void);
753 static tree rs6000_gimplify_va_arg (tree, tree, tree *, tree *);
754 static bool rs6000_must_pass_in_stack (enum machine_mode, tree);
755 static bool rs6000_scalar_mode_supported_p (enum machine_mode);
756 static bool rs6000_vector_mode_supported_p (enum machine_mode);
757 static int get_vec_cmp_insn (enum rtx_code, enum machine_mode,
758 enum machine_mode);
759 static rtx rs6000_emit_vector_compare (enum rtx_code, rtx, rtx,
760 enum machine_mode);
761 static int get_vsel_insn (enum machine_mode);
762 static void rs6000_emit_vector_select (rtx, rtx, rtx, rtx);
763 static tree rs6000_stack_protect_fail (void);
764
765 const int INSN_NOT_AVAILABLE = -1;
766 static enum machine_mode rs6000_eh_return_filter_mode (void);
767
768 /* Hash table stuff for keeping track of TOC entries. */
769
770 struct toc_hash_struct GTY(())
771 {
772 /* `key' will satisfy CONSTANT_P; in fact, it will satisfy
773 ASM_OUTPUT_SPECIAL_POOL_ENTRY_P. */
774 rtx key;
775 enum machine_mode key_mode;
776 int labelno;
777 };
778
779 static GTY ((param_is (struct toc_hash_struct))) htab_t toc_hash_table;
780
781 /* Default register names. */
782 char rs6000_reg_names[][8] =
783 {
784 "0", "1", "2", "3", "4", "5", "6", "7",
785 "8", "9", "10", "11", "12", "13", "14", "15",
786 "16", "17", "18", "19", "20", "21", "22", "23",
787 "24", "25", "26", "27", "28", "29", "30", "31",
788 "0", "1", "2", "3", "4", "5", "6", "7",
789 "8", "9", "10", "11", "12", "13", "14", "15",
790 "16", "17", "18", "19", "20", "21", "22", "23",
791 "24", "25", "26", "27", "28", "29", "30", "31",
792 "mq", "lr", "ctr","ap",
793 "0", "1", "2", "3", "4", "5", "6", "7",
794 "xer",
795 /* AltiVec registers. */
796 "0", "1", "2", "3", "4", "5", "6", "7",
797 "8", "9", "10", "11", "12", "13", "14", "15",
798 "16", "17", "18", "19", "20", "21", "22", "23",
799 "24", "25", "26", "27", "28", "29", "30", "31",
800 "vrsave", "vscr",
801 /* SPE registers. */
802 "spe_acc", "spefscr",
803 /* Soft frame pointer. */
804 "sfp"
805 };
806
807 #ifdef TARGET_REGNAMES
808 static const char alt_reg_names[][8] =
809 {
810 "%r0", "%r1", "%r2", "%r3", "%r4", "%r5", "%r6", "%r7",
811 "%r8", "%r9", "%r10", "%r11", "%r12", "%r13", "%r14", "%r15",
812 "%r16", "%r17", "%r18", "%r19", "%r20", "%r21", "%r22", "%r23",
813 "%r24", "%r25", "%r26", "%r27", "%r28", "%r29", "%r30", "%r31",
814 "%f0", "%f1", "%f2", "%f3", "%f4", "%f5", "%f6", "%f7",
815 "%f8", "%f9", "%f10", "%f11", "%f12", "%f13", "%f14", "%f15",
816 "%f16", "%f17", "%f18", "%f19", "%f20", "%f21", "%f22", "%f23",
817 "%f24", "%f25", "%f26", "%f27", "%f28", "%f29", "%f30", "%f31",
818 "mq", "lr", "ctr", "ap",
819 "%cr0", "%cr1", "%cr2", "%cr3", "%cr4", "%cr5", "%cr6", "%cr7",
820 "xer",
821 /* AltiVec registers. */
822 "%v0", "%v1", "%v2", "%v3", "%v4", "%v5", "%v6", "%v7",
823 "%v8", "%v9", "%v10", "%v11", "%v12", "%v13", "%v14", "%v15",
824 "%v16", "%v17", "%v18", "%v19", "%v20", "%v21", "%v22", "%v23",
825 "%v24", "%v25", "%v26", "%v27", "%v28", "%v29", "%v30", "%v31",
826 "vrsave", "vscr",
827 /* SPE registers. */
828 "spe_acc", "spefscr",
829 /* Soft frame pointer. */
830 "sfp"
831 };
832 #endif
833
834 #ifndef MASK_STRICT_ALIGN
835 #define MASK_STRICT_ALIGN 0
836 #endif
837 #ifndef TARGET_PROFILE_KERNEL
838 #define TARGET_PROFILE_KERNEL 0
839 #endif
840
841 /* The VRSAVE bitmask puts bit %v0 as the most significant bit. */
842 #define ALTIVEC_REG_BIT(REGNO) (0x80000000 >> ((REGNO) - FIRST_ALTIVEC_REGNO))
843
844 /* Initialize the GCC target structure. */
845 #undef TARGET_ATTRIBUTE_TABLE
846 #define TARGET_ATTRIBUTE_TABLE rs6000_attribute_table
847 #undef TARGET_SET_DEFAULT_TYPE_ATTRIBUTES
848 #define TARGET_SET_DEFAULT_TYPE_ATTRIBUTES rs6000_set_default_type_attributes
849
850 #undef TARGET_ASM_ALIGNED_DI_OP
851 #define TARGET_ASM_ALIGNED_DI_OP DOUBLE_INT_ASM_OP
852
853 /* Default unaligned ops are only provided for ELF. Find the ops needed
854 for non-ELF systems. */
855 #ifndef OBJECT_FORMAT_ELF
856 #if TARGET_XCOFF
857 /* For XCOFF. rs6000_assemble_integer will handle unaligned DIs on
858 64-bit targets. */
859 #undef TARGET_ASM_UNALIGNED_HI_OP
860 #define TARGET_ASM_UNALIGNED_HI_OP "\t.vbyte\t2,"
861 #undef TARGET_ASM_UNALIGNED_SI_OP
862 #define TARGET_ASM_UNALIGNED_SI_OP "\t.vbyte\t4,"
863 #undef TARGET_ASM_UNALIGNED_DI_OP
864 #define TARGET_ASM_UNALIGNED_DI_OP "\t.vbyte\t8,"
865 #else
866 /* For Darwin. */
867 #undef TARGET_ASM_UNALIGNED_HI_OP
868 #define TARGET_ASM_UNALIGNED_HI_OP "\t.short\t"
869 #undef TARGET_ASM_UNALIGNED_SI_OP
870 #define TARGET_ASM_UNALIGNED_SI_OP "\t.long\t"
871 #undef TARGET_ASM_UNALIGNED_DI_OP
872 #define TARGET_ASM_UNALIGNED_DI_OP "\t.quad\t"
873 #undef TARGET_ASM_ALIGNED_DI_OP
874 #define TARGET_ASM_ALIGNED_DI_OP "\t.quad\t"
875 #endif
876 #endif
877
878 /* This hook deals with fixups for relocatable code and DI-mode objects
879 in 64-bit code. */
880 #undef TARGET_ASM_INTEGER
881 #define TARGET_ASM_INTEGER rs6000_assemble_integer
882
883 #ifdef HAVE_GAS_HIDDEN
884 #undef TARGET_ASM_ASSEMBLE_VISIBILITY
885 #define TARGET_ASM_ASSEMBLE_VISIBILITY rs6000_assemble_visibility
886 #endif
887
888 #undef TARGET_HAVE_TLS
889 #define TARGET_HAVE_TLS HAVE_AS_TLS
890
891 #undef TARGET_CANNOT_FORCE_CONST_MEM
892 #define TARGET_CANNOT_FORCE_CONST_MEM rs6000_tls_referenced_p
893
894 #undef TARGET_ASM_FUNCTION_PROLOGUE
895 #define TARGET_ASM_FUNCTION_PROLOGUE rs6000_output_function_prologue
896 #undef TARGET_ASM_FUNCTION_EPILOGUE
897 #define TARGET_ASM_FUNCTION_EPILOGUE rs6000_output_function_epilogue
898
899 #undef TARGET_SCHED_VARIABLE_ISSUE
900 #define TARGET_SCHED_VARIABLE_ISSUE rs6000_variable_issue
901
902 #undef TARGET_SCHED_ISSUE_RATE
903 #define TARGET_SCHED_ISSUE_RATE rs6000_issue_rate
904 #undef TARGET_SCHED_ADJUST_COST
905 #define TARGET_SCHED_ADJUST_COST rs6000_adjust_cost
906 #undef TARGET_SCHED_ADJUST_PRIORITY
907 #define TARGET_SCHED_ADJUST_PRIORITY rs6000_adjust_priority
908 #undef TARGET_SCHED_IS_COSTLY_DEPENDENCE
909 #define TARGET_SCHED_IS_COSTLY_DEPENDENCE rs6000_is_costly_dependence
910 #undef TARGET_SCHED_FINISH
911 #define TARGET_SCHED_FINISH rs6000_sched_finish
912
913 #undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD
914 #define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD rs6000_use_sched_lookahead
915
916 #undef TARGET_VECTORIZE_BUILTIN_MASK_FOR_LOAD
917 #define TARGET_VECTORIZE_BUILTIN_MASK_FOR_LOAD rs6000_builtin_mask_for_load
918
919 #undef TARGET_VECTOR_ALIGNMENT_REACHABLE
920 #define TARGET_VECTOR_ALIGNMENT_REACHABLE rs6000_vector_alignment_reachable
921
922 #undef TARGET_INIT_BUILTINS
923 #define TARGET_INIT_BUILTINS rs6000_init_builtins
924
925 #undef TARGET_EXPAND_BUILTIN
926 #define TARGET_EXPAND_BUILTIN rs6000_expand_builtin
927
928 #undef TARGET_MANGLE_FUNDAMENTAL_TYPE
929 #define TARGET_MANGLE_FUNDAMENTAL_TYPE rs6000_mangle_fundamental_type
930
931 #undef TARGET_INIT_LIBFUNCS
932 #define TARGET_INIT_LIBFUNCS rs6000_init_libfuncs
933
934 #if TARGET_MACHO
935 #undef TARGET_BINDS_LOCAL_P
936 #define TARGET_BINDS_LOCAL_P darwin_binds_local_p
937 #endif
938
939 #undef TARGET_MS_BITFIELD_LAYOUT_P
940 #define TARGET_MS_BITFIELD_LAYOUT_P rs6000_ms_bitfield_layout_p
941
942 #undef TARGET_ASM_OUTPUT_MI_THUNK
943 #define TARGET_ASM_OUTPUT_MI_THUNK rs6000_output_mi_thunk
944
945 #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK
946 #define TARGET_ASM_CAN_OUTPUT_MI_THUNK hook_bool_tree_hwi_hwi_tree_true
947
948 #undef TARGET_FUNCTION_OK_FOR_SIBCALL
949 #define TARGET_FUNCTION_OK_FOR_SIBCALL rs6000_function_ok_for_sibcall
950
951 #undef TARGET_INVALID_WITHIN_DOLOOP
952 #define TARGET_INVALID_WITHIN_DOLOOP rs6000_invalid_within_doloop
953
954 #undef TARGET_RTX_COSTS
955 #define TARGET_RTX_COSTS rs6000_rtx_costs
956 #undef TARGET_ADDRESS_COST
957 #define TARGET_ADDRESS_COST hook_int_rtx_0
958
959 #undef TARGET_VECTOR_OPAQUE_P
960 #define TARGET_VECTOR_OPAQUE_P rs6000_is_opaque_type
961
962 #undef TARGET_DWARF_REGISTER_SPAN
963 #define TARGET_DWARF_REGISTER_SPAN rs6000_dwarf_register_span
964
965 /* On rs6000, function arguments are promoted, as are function return
966 values. */
967 #undef TARGET_PROMOTE_FUNCTION_ARGS
968 #define TARGET_PROMOTE_FUNCTION_ARGS hook_bool_tree_true
969 #undef TARGET_PROMOTE_FUNCTION_RETURN
970 #define TARGET_PROMOTE_FUNCTION_RETURN hook_bool_tree_true
971
972 #undef TARGET_RETURN_IN_MEMORY
973 #define TARGET_RETURN_IN_MEMORY rs6000_return_in_memory
974
975 #undef TARGET_SETUP_INCOMING_VARARGS
976 #define TARGET_SETUP_INCOMING_VARARGS setup_incoming_varargs
977
978 /* Always strict argument naming on rs6000. */
979 #undef TARGET_STRICT_ARGUMENT_NAMING
980 #define TARGET_STRICT_ARGUMENT_NAMING hook_bool_CUMULATIVE_ARGS_true
981 #undef TARGET_PRETEND_OUTGOING_VARARGS_NAMED
982 #define TARGET_PRETEND_OUTGOING_VARARGS_NAMED hook_bool_CUMULATIVE_ARGS_true
983 #undef TARGET_SPLIT_COMPLEX_ARG
984 #define TARGET_SPLIT_COMPLEX_ARG hook_bool_tree_true
985 #undef TARGET_MUST_PASS_IN_STACK
986 #define TARGET_MUST_PASS_IN_STACK rs6000_must_pass_in_stack
987 #undef TARGET_PASS_BY_REFERENCE
988 #define TARGET_PASS_BY_REFERENCE rs6000_pass_by_reference
989 #undef TARGET_ARG_PARTIAL_BYTES
990 #define TARGET_ARG_PARTIAL_BYTES rs6000_arg_partial_bytes
991
992 #undef TARGET_BUILD_BUILTIN_VA_LIST
993 #define TARGET_BUILD_BUILTIN_VA_LIST rs6000_build_builtin_va_list
994
995 #undef TARGET_GIMPLIFY_VA_ARG_EXPR
996 #define TARGET_GIMPLIFY_VA_ARG_EXPR rs6000_gimplify_va_arg
997
998 #undef TARGET_EH_RETURN_FILTER_MODE
999 #define TARGET_EH_RETURN_FILTER_MODE rs6000_eh_return_filter_mode
1000
1001 #undef TARGET_SCALAR_MODE_SUPPORTED_P
1002 #define TARGET_SCALAR_MODE_SUPPORTED_P rs6000_scalar_mode_supported_p
1003
1004 #undef TARGET_VECTOR_MODE_SUPPORTED_P
1005 #define TARGET_VECTOR_MODE_SUPPORTED_P rs6000_vector_mode_supported_p
1006
1007 #undef TARGET_INVALID_ARG_FOR_UNPROTOTYPED_FN
1008 #define TARGET_INVALID_ARG_FOR_UNPROTOTYPED_FN invalid_arg_for_unprototyped_fn
1009
1010 #undef TARGET_HANDLE_OPTION
1011 #define TARGET_HANDLE_OPTION rs6000_handle_option
1012
1013 #undef TARGET_DEFAULT_TARGET_FLAGS
1014 #define TARGET_DEFAULT_TARGET_FLAGS \
1015 (TARGET_DEFAULT)
1016
1017 #undef TARGET_STACK_PROTECT_FAIL
1018 #define TARGET_STACK_PROTECT_FAIL rs6000_stack_protect_fail
1019
1020 /* MPC604EUM 3.5.2 Weak Consistency between Multiple Processors
1021 The PowerPC architecture requires only weak consistency among
1022 processors--that is, memory accesses between processors need not be
1023 sequentially consistent and memory accesses among processors can occur
1024 in any order. The ability to order memory accesses weakly provides
1025 opportunities for more efficient use of the system bus. Unless a
1026 dependency exists, the 604e allows read operations to precede store
1027 operations. */
1028 #undef TARGET_RELAXED_ORDERING
1029 #define TARGET_RELAXED_ORDERING true
1030
1031 #ifdef HAVE_AS_TLS
1032 #undef TARGET_ASM_OUTPUT_DWARF_DTPREL
1033 #define TARGET_ASM_OUTPUT_DWARF_DTPREL rs6000_output_dwarf_dtprel
1034 #endif
1035
1036 /* Use a 32-bit anchor range. This leads to sequences like:
1037
1038 addis tmp,anchor,high
1039 add dest,tmp,low
1040
1041 where tmp itself acts as an anchor, and can be shared between
1042 accesses to the same 64k page. */
1043 #undef TARGET_MIN_ANCHOR_OFFSET
1044 #define TARGET_MIN_ANCHOR_OFFSET -0x7fffffff - 1
1045 #undef TARGET_MAX_ANCHOR_OFFSET
1046 #define TARGET_MAX_ANCHOR_OFFSET 0x7fffffff
1047 #undef TARGET_USE_BLOCKS_FOR_CONSTANT_P
1048 #define TARGET_USE_BLOCKS_FOR_CONSTANT_P rs6000_use_blocks_for_constant_p
1049
1050 struct gcc_target targetm = TARGET_INITIALIZER;
1051
1052
1053 /* Value is 1 if hard register REGNO can hold a value of machine-mode
1054 MODE. */
1055 static int
rs6000_hard_regno_mode_ok(int regno,enum machine_mode mode)1056 rs6000_hard_regno_mode_ok (int regno, enum machine_mode mode)
1057 {
1058 /* The GPRs can hold any mode, but values bigger than one register
1059 cannot go past R31. */
1060 if (INT_REGNO_P (regno))
1061 return INT_REGNO_P (regno + HARD_REGNO_NREGS (regno, mode) - 1);
1062
1063 /* The float registers can only hold floating modes and DImode.
1064 This also excludes decimal float modes. */
1065 if (FP_REGNO_P (regno))
1066 return
1067 (SCALAR_FLOAT_MODE_P (mode)
1068 && !DECIMAL_FLOAT_MODE_P (mode)
1069 && FP_REGNO_P (regno + HARD_REGNO_NREGS (regno, mode) - 1))
1070 || (GET_MODE_CLASS (mode) == MODE_INT
1071 && GET_MODE_SIZE (mode) == UNITS_PER_FP_WORD);
1072
1073 /* The CR register can only hold CC modes. */
1074 if (CR_REGNO_P (regno))
1075 return GET_MODE_CLASS (mode) == MODE_CC;
1076
1077 if (XER_REGNO_P (regno))
1078 return mode == PSImode;
1079
1080 /* AltiVec only in AldyVec registers. */
1081 if (ALTIVEC_REGNO_P (regno))
1082 return ALTIVEC_VECTOR_MODE (mode);
1083
1084 /* ...but GPRs can hold SIMD data on the SPE in one register. */
1085 if (SPE_SIMD_REGNO_P (regno) && TARGET_SPE && SPE_VECTOR_MODE (mode))
1086 return 1;
1087
1088 /* We cannot put TImode anywhere except general register and it must be
1089 able to fit within the register set. */
1090
1091 return GET_MODE_SIZE (mode) <= UNITS_PER_WORD;
1092 }
1093
1094 /* Initialize rs6000_hard_regno_mode_ok_p table. */
1095 static void
rs6000_init_hard_regno_mode_ok(void)1096 rs6000_init_hard_regno_mode_ok (void)
1097 {
1098 int r, m;
1099
1100 for (r = 0; r < FIRST_PSEUDO_REGISTER; ++r)
1101 for (m = 0; m < NUM_MACHINE_MODES; ++m)
1102 if (rs6000_hard_regno_mode_ok (r, m))
1103 rs6000_hard_regno_mode_ok_p[m][r] = true;
1104 }
1105
1106 /* If not otherwise specified by a target, make 'long double' equivalent to
1107 'double'. */
1108
1109 #ifndef RS6000_DEFAULT_LONG_DOUBLE_SIZE
1110 #define RS6000_DEFAULT_LONG_DOUBLE_SIZE 64
1111 #endif
1112
1113 /* Override command line options. Mostly we process the processor
1114 type and sometimes adjust other TARGET_ options. */
1115
1116 void
rs6000_override_options(const char * default_cpu)1117 rs6000_override_options (const char *default_cpu)
1118 {
1119 size_t i, j;
1120 struct rs6000_cpu_select *ptr;
1121 int set_masks;
1122
1123 /* Simplifications for entries below. */
1124
1125 enum {
1126 POWERPC_BASE_MASK = MASK_POWERPC | MASK_NEW_MNEMONICS,
1127 POWERPC_7400_MASK = POWERPC_BASE_MASK | MASK_PPC_GFXOPT | MASK_ALTIVEC
1128 };
1129
1130 /* This table occasionally claims that a processor does not support
1131 a particular feature even though it does, but the feature is slower
1132 than the alternative. Thus, it shouldn't be relied on as a
1133 complete description of the processor's support.
1134
1135 Please keep this list in order, and don't forget to update the
1136 documentation in invoke.texi when adding a new processor or
1137 flag. */
1138 static struct ptt
1139 {
1140 const char *const name; /* Canonical processor name. */
1141 const enum processor_type processor; /* Processor type enum value. */
1142 const int target_enable; /* Target flags to enable. */
1143 } const processor_target_table[]
1144 = {{"401", PROCESSOR_PPC403, POWERPC_BASE_MASK | MASK_SOFT_FLOAT},
1145 {"403", PROCESSOR_PPC403,
1146 POWERPC_BASE_MASK | MASK_SOFT_FLOAT | MASK_STRICT_ALIGN},
1147 {"405", PROCESSOR_PPC405,
1148 POWERPC_BASE_MASK | MASK_SOFT_FLOAT | MASK_MULHW | MASK_DLMZB},
1149 {"405fp", PROCESSOR_PPC405,
1150 POWERPC_BASE_MASK | MASK_MULHW | MASK_DLMZB},
1151 {"440", PROCESSOR_PPC440,
1152 POWERPC_BASE_MASK | MASK_SOFT_FLOAT | MASK_MULHW | MASK_DLMZB},
1153 {"440fp", PROCESSOR_PPC440,
1154 POWERPC_BASE_MASK | MASK_MULHW | MASK_DLMZB},
1155 {"505", PROCESSOR_MPCCORE, POWERPC_BASE_MASK},
1156 {"601", PROCESSOR_PPC601,
1157 MASK_POWER | POWERPC_BASE_MASK | MASK_MULTIPLE | MASK_STRING},
1158 {"602", PROCESSOR_PPC603, POWERPC_BASE_MASK | MASK_PPC_GFXOPT},
1159 {"603", PROCESSOR_PPC603, POWERPC_BASE_MASK | MASK_PPC_GFXOPT},
1160 {"603e", PROCESSOR_PPC603, POWERPC_BASE_MASK | MASK_PPC_GFXOPT},
1161 {"604", PROCESSOR_PPC604, POWERPC_BASE_MASK | MASK_PPC_GFXOPT},
1162 {"604e", PROCESSOR_PPC604e, POWERPC_BASE_MASK | MASK_PPC_GFXOPT},
1163 {"620", PROCESSOR_PPC620,
1164 POWERPC_BASE_MASK | MASK_PPC_GFXOPT | MASK_POWERPC64},
1165 {"630", PROCESSOR_PPC630,
1166 POWERPC_BASE_MASK | MASK_PPC_GFXOPT | MASK_POWERPC64},
1167 {"740", PROCESSOR_PPC750, POWERPC_BASE_MASK | MASK_PPC_GFXOPT},
1168 {"7400", PROCESSOR_PPC7400, POWERPC_7400_MASK},
1169 {"7450", PROCESSOR_PPC7450, POWERPC_7400_MASK},
1170 {"750", PROCESSOR_PPC750, POWERPC_BASE_MASK | MASK_PPC_GFXOPT},
1171 {"801", PROCESSOR_MPCCORE, POWERPC_BASE_MASK | MASK_SOFT_FLOAT},
1172 {"821", PROCESSOR_MPCCORE, POWERPC_BASE_MASK | MASK_SOFT_FLOAT},
1173 {"823", PROCESSOR_MPCCORE, POWERPC_BASE_MASK | MASK_SOFT_FLOAT},
1174 {"8540", PROCESSOR_PPC8540, POWERPC_BASE_MASK | MASK_STRICT_ALIGN},
1175 /* 8548 has a dummy entry for now. */
1176 {"8548", PROCESSOR_PPC8540, POWERPC_BASE_MASK | MASK_STRICT_ALIGN},
1177 {"860", PROCESSOR_MPCCORE, POWERPC_BASE_MASK | MASK_SOFT_FLOAT},
1178 {"970", PROCESSOR_POWER4,
1179 POWERPC_7400_MASK | MASK_PPC_GPOPT | MASK_MFCRF | MASK_POWERPC64},
1180 {"common", PROCESSOR_COMMON, MASK_NEW_MNEMONICS},
1181 {"ec603e", PROCESSOR_PPC603, POWERPC_BASE_MASK | MASK_SOFT_FLOAT},
1182 {"G3", PROCESSOR_PPC750, POWERPC_BASE_MASK | MASK_PPC_GFXOPT},
1183 {"G4", PROCESSOR_PPC7450, POWERPC_7400_MASK},
1184 {"G5", PROCESSOR_POWER4,
1185 POWERPC_7400_MASK | MASK_PPC_GPOPT | MASK_MFCRF | MASK_POWERPC64},
1186 {"power", PROCESSOR_POWER, MASK_POWER | MASK_MULTIPLE | MASK_STRING},
1187 {"power2", PROCESSOR_POWER,
1188 MASK_POWER | MASK_POWER2 | MASK_MULTIPLE | MASK_STRING},
1189 {"power3", PROCESSOR_PPC630,
1190 POWERPC_BASE_MASK | MASK_PPC_GFXOPT | MASK_POWERPC64},
1191 {"power4", PROCESSOR_POWER4,
1192 POWERPC_BASE_MASK | MASK_PPC_GFXOPT | MASK_MFCRF | MASK_POWERPC64},
1193 {"power5", PROCESSOR_POWER5,
1194 POWERPC_BASE_MASK | MASK_POWERPC64 | MASK_PPC_GFXOPT
1195 | MASK_MFCRF | MASK_POPCNTB},
1196 {"power5+", PROCESSOR_POWER5,
1197 POWERPC_BASE_MASK | MASK_POWERPC64 | MASK_PPC_GFXOPT
1198 | MASK_MFCRF | MASK_POPCNTB | MASK_FPRND},
1199 {"power6", PROCESSOR_POWER5,
1200 POWERPC_7400_MASK | MASK_POWERPC64 | MASK_MFCRF | MASK_POPCNTB
1201 | MASK_FPRND},
1202 {"powerpc", PROCESSOR_POWERPC, POWERPC_BASE_MASK},
1203 {"powerpc64", PROCESSOR_POWERPC64,
1204 POWERPC_BASE_MASK | MASK_PPC_GFXOPT | MASK_POWERPC64},
1205 {"rios", PROCESSOR_RIOS1, MASK_POWER | MASK_MULTIPLE | MASK_STRING},
1206 {"rios1", PROCESSOR_RIOS1, MASK_POWER | MASK_MULTIPLE | MASK_STRING},
1207 {"rios2", PROCESSOR_RIOS2,
1208 MASK_POWER | MASK_POWER2 | MASK_MULTIPLE | MASK_STRING},
1209 {"rsc", PROCESSOR_PPC601, MASK_POWER | MASK_MULTIPLE | MASK_STRING},
1210 {"rsc1", PROCESSOR_PPC601, MASK_POWER | MASK_MULTIPLE | MASK_STRING},
1211 {"rs64", PROCESSOR_RS64A,
1212 POWERPC_BASE_MASK | MASK_PPC_GFXOPT | MASK_POWERPC64}
1213 };
1214
1215 const size_t ptt_size = ARRAY_SIZE (processor_target_table);
1216
1217 /* Some OSs don't support saving the high part of 64-bit registers on
1218 context switch. Other OSs don't support saving Altivec registers.
1219 On those OSs, we don't touch the MASK_POWERPC64 or MASK_ALTIVEC
1220 settings; if the user wants either, the user must explicitly specify
1221 them and we won't interfere with the user's specification. */
1222
1223 enum {
1224 POWER_MASKS = MASK_POWER | MASK_POWER2 | MASK_MULTIPLE | MASK_STRING,
1225 POWERPC_MASKS = (POWERPC_BASE_MASK | MASK_PPC_GPOPT | MASK_STRICT_ALIGN
1226 | MASK_PPC_GFXOPT | MASK_POWERPC64 | MASK_ALTIVEC
1227 | MASK_MFCRF | MASK_POPCNTB | MASK_FPRND | MASK_MULHW
1228 | MASK_DLMZB)
1229 };
1230
1231 rs6000_init_hard_regno_mode_ok ();
1232
1233 set_masks = POWER_MASKS | POWERPC_MASKS | MASK_SOFT_FLOAT;
1234 #ifdef OS_MISSING_POWERPC64
1235 if (OS_MISSING_POWERPC64)
1236 set_masks &= ~MASK_POWERPC64;
1237 #endif
1238 #ifdef OS_MISSING_ALTIVEC
1239 if (OS_MISSING_ALTIVEC)
1240 set_masks &= ~MASK_ALTIVEC;
1241 #endif
1242
1243 /* Don't override by the processor default if given explicitly. */
1244 set_masks &= ~target_flags_explicit;
1245
1246 /* Identify the processor type. */
1247 rs6000_select[0].string = default_cpu;
1248 rs6000_cpu = TARGET_POWERPC64 ? PROCESSOR_DEFAULT64 : PROCESSOR_DEFAULT;
1249
1250 for (i = 0; i < ARRAY_SIZE (rs6000_select); i++)
1251 {
1252 ptr = &rs6000_select[i];
1253 if (ptr->string != (char *)0 && ptr->string[0] != '\0')
1254 {
1255 for (j = 0; j < ptt_size; j++)
1256 if (! strcmp (ptr->string, processor_target_table[j].name))
1257 {
1258 if (ptr->set_tune_p)
1259 rs6000_cpu = processor_target_table[j].processor;
1260
1261 if (ptr->set_arch_p)
1262 {
1263 target_flags &= ~set_masks;
1264 target_flags |= (processor_target_table[j].target_enable
1265 & set_masks);
1266 }
1267 break;
1268 }
1269
1270 if (j == ptt_size)
1271 error ("bad value (%s) for %s switch", ptr->string, ptr->name);
1272 }
1273 }
1274
1275 if (TARGET_E500)
1276 rs6000_isel = 1;
1277
1278 /* If we are optimizing big endian systems for space, use the load/store
1279 multiple and string instructions. */
1280 if (BYTES_BIG_ENDIAN && optimize_size)
1281 target_flags |= ~target_flags_explicit & (MASK_MULTIPLE | MASK_STRING);
1282
1283 /* Don't allow -mmultiple or -mstring on little endian systems
1284 unless the cpu is a 750, because the hardware doesn't support the
1285 instructions used in little endian mode, and causes an alignment
1286 trap. The 750 does not cause an alignment trap (except when the
1287 target is unaligned). */
1288
1289 if (!BYTES_BIG_ENDIAN && rs6000_cpu != PROCESSOR_PPC750)
1290 {
1291 if (TARGET_MULTIPLE)
1292 {
1293 target_flags &= ~MASK_MULTIPLE;
1294 if ((target_flags_explicit & MASK_MULTIPLE) != 0)
1295 warning (0, "-mmultiple is not supported on little endian systems");
1296 }
1297
1298 if (TARGET_STRING)
1299 {
1300 target_flags &= ~MASK_STRING;
1301 if ((target_flags_explicit & MASK_STRING) != 0)
1302 warning (0, "-mstring is not supported on little endian systems");
1303 }
1304 }
1305
1306 /* Set debug flags */
1307 if (rs6000_debug_name)
1308 {
1309 if (! strcmp (rs6000_debug_name, "all"))
1310 rs6000_debug_stack = rs6000_debug_arg = 1;
1311 else if (! strcmp (rs6000_debug_name, "stack"))
1312 rs6000_debug_stack = 1;
1313 else if (! strcmp (rs6000_debug_name, "arg"))
1314 rs6000_debug_arg = 1;
1315 else
1316 error ("unknown -mdebug-%s switch", rs6000_debug_name);
1317 }
1318
1319 if (rs6000_traceback_name)
1320 {
1321 if (! strncmp (rs6000_traceback_name, "full", 4))
1322 rs6000_traceback = traceback_full;
1323 else if (! strncmp (rs6000_traceback_name, "part", 4))
1324 rs6000_traceback = traceback_part;
1325 else if (! strncmp (rs6000_traceback_name, "no", 2))
1326 rs6000_traceback = traceback_none;
1327 else
1328 error ("unknown -mtraceback arg %qs; expecting %<full%>, %<partial%> or %<none%>",
1329 rs6000_traceback_name);
1330 }
1331
1332 if (!rs6000_explicit_options.long_double)
1333 rs6000_long_double_type_size = RS6000_DEFAULT_LONG_DOUBLE_SIZE;
1334
1335 #ifndef POWERPC_LINUX
1336 if (!rs6000_explicit_options.ieee)
1337 rs6000_ieeequad = 1;
1338 #endif
1339
1340 /* Set Altivec ABI as default for powerpc64 linux. */
1341 if (TARGET_ELF && TARGET_64BIT)
1342 {
1343 rs6000_altivec_abi = 1;
1344 TARGET_ALTIVEC_VRSAVE = 1;
1345 }
1346
1347 /* Set the Darwin64 ABI as default for 64-bit Darwin. */
1348 if (DEFAULT_ABI == ABI_DARWIN && TARGET_64BIT)
1349 {
1350 rs6000_darwin64_abi = 1;
1351 #if TARGET_MACHO
1352 darwin_one_byte_bool = 1;
1353 #endif
1354 /* Default to natural alignment, for better performance. */
1355 rs6000_alignment_flags = MASK_ALIGN_NATURAL;
1356 }
1357
1358 /* Place FP constants in the constant pool instead of TOC
1359 if section anchors enabled. */
1360 if (flag_section_anchors)
1361 TARGET_NO_FP_IN_TOC = 1;
1362
1363 /* Handle -mtls-size option. */
1364 rs6000_parse_tls_size_option ();
1365
1366 #ifdef SUBTARGET_OVERRIDE_OPTIONS
1367 SUBTARGET_OVERRIDE_OPTIONS;
1368 #endif
1369 #ifdef SUBSUBTARGET_OVERRIDE_OPTIONS
1370 SUBSUBTARGET_OVERRIDE_OPTIONS;
1371 #endif
1372 #ifdef SUB3TARGET_OVERRIDE_OPTIONS
1373 SUB3TARGET_OVERRIDE_OPTIONS;
1374 #endif
1375
1376 if (TARGET_E500)
1377 {
1378 if (TARGET_ALTIVEC)
1379 error ("AltiVec and E500 instructions cannot coexist");
1380
1381 /* The e500 does not have string instructions, and we set
1382 MASK_STRING above when optimizing for size. */
1383 if ((target_flags & MASK_STRING) != 0)
1384 target_flags = target_flags & ~MASK_STRING;
1385 }
1386 else if (rs6000_select[1].string != NULL)
1387 {
1388 /* For the powerpc-eabispe configuration, we set all these by
1389 default, so let's unset them if we manually set another
1390 CPU that is not the E500. */
1391 if (!rs6000_explicit_options.abi)
1392 rs6000_spe_abi = 0;
1393 if (!rs6000_explicit_options.spe)
1394 rs6000_spe = 0;
1395 if (!rs6000_explicit_options.float_gprs)
1396 rs6000_float_gprs = 0;
1397 if (!rs6000_explicit_options.isel)
1398 rs6000_isel = 0;
1399 if (!rs6000_explicit_options.long_double)
1400 rs6000_long_double_type_size = RS6000_DEFAULT_LONG_DOUBLE_SIZE;
1401 }
1402
1403 rs6000_always_hint = (rs6000_cpu != PROCESSOR_POWER4
1404 && rs6000_cpu != PROCESSOR_POWER5);
1405 rs6000_sched_groups = (rs6000_cpu == PROCESSOR_POWER4
1406 || rs6000_cpu == PROCESSOR_POWER5);
1407
1408 rs6000_sched_restricted_insns_priority
1409 = (rs6000_sched_groups ? 1 : 0);
1410
1411 /* Handle -msched-costly-dep option. */
1412 rs6000_sched_costly_dep
1413 = (rs6000_sched_groups ? store_to_load_dep_costly : no_dep_costly);
1414
1415 if (rs6000_sched_costly_dep_str)
1416 {
1417 if (! strcmp (rs6000_sched_costly_dep_str, "no"))
1418 rs6000_sched_costly_dep = no_dep_costly;
1419 else if (! strcmp (rs6000_sched_costly_dep_str, "all"))
1420 rs6000_sched_costly_dep = all_deps_costly;
1421 else if (! strcmp (rs6000_sched_costly_dep_str, "true_store_to_load"))
1422 rs6000_sched_costly_dep = true_store_to_load_dep_costly;
1423 else if (! strcmp (rs6000_sched_costly_dep_str, "store_to_load"))
1424 rs6000_sched_costly_dep = store_to_load_dep_costly;
1425 else
1426 rs6000_sched_costly_dep = atoi (rs6000_sched_costly_dep_str);
1427 }
1428
1429 /* Handle -minsert-sched-nops option. */
1430 rs6000_sched_insert_nops
1431 = (rs6000_sched_groups ? sched_finish_regroup_exact : sched_finish_none);
1432
1433 if (rs6000_sched_insert_nops_str)
1434 {
1435 if (! strcmp (rs6000_sched_insert_nops_str, "no"))
1436 rs6000_sched_insert_nops = sched_finish_none;
1437 else if (! strcmp (rs6000_sched_insert_nops_str, "pad"))
1438 rs6000_sched_insert_nops = sched_finish_pad_groups;
1439 else if (! strcmp (rs6000_sched_insert_nops_str, "regroup_exact"))
1440 rs6000_sched_insert_nops = sched_finish_regroup_exact;
1441 else
1442 rs6000_sched_insert_nops = atoi (rs6000_sched_insert_nops_str);
1443 }
1444
1445 #ifdef TARGET_REGNAMES
1446 /* If the user desires alternate register names, copy in the
1447 alternate names now. */
1448 if (TARGET_REGNAMES)
1449 memcpy (rs6000_reg_names, alt_reg_names, sizeof (rs6000_reg_names));
1450 #endif
1451
1452 /* Set aix_struct_return last, after the ABI is determined.
1453 If -maix-struct-return or -msvr4-struct-return was explicitly
1454 used, don't override with the ABI default. */
1455 if (!rs6000_explicit_options.aix_struct_ret)
1456 aix_struct_return = (DEFAULT_ABI != ABI_V4 || DRAFT_V4_STRUCT_RET);
1457
1458 if (TARGET_LONG_DOUBLE_128 && !TARGET_IEEEQUAD)
1459 REAL_MODE_FORMAT (TFmode) = &ibm_extended_format;
1460
1461 if (TARGET_TOC)
1462 ASM_GENERATE_INTERNAL_LABEL (toc_label_name, "LCTOC", 1);
1463
1464 /* We can only guarantee the availability of DI pseudo-ops when
1465 assembling for 64-bit targets. */
1466 if (!TARGET_64BIT)
1467 {
1468 targetm.asm_out.aligned_op.di = NULL;
1469 targetm.asm_out.unaligned_op.di = NULL;
1470 }
1471
1472 /* Set branch target alignment, if not optimizing for size. */
1473 if (!optimize_size)
1474 {
1475 if (rs6000_sched_groups)
1476 {
1477 if (align_functions <= 0)
1478 align_functions = 16;
1479 if (align_jumps <= 0)
1480 align_jumps = 16;
1481 if (align_loops <= 0)
1482 align_loops = 16;
1483 }
1484 if (align_jumps_max_skip <= 0)
1485 align_jumps_max_skip = 15;
1486 if (align_loops_max_skip <= 0)
1487 align_loops_max_skip = 15;
1488 }
1489
1490 /* Arrange to save and restore machine status around nested functions. */
1491 init_machine_status = rs6000_init_machine_status;
1492
1493 /* We should always be splitting complex arguments, but we can't break
1494 Linux and Darwin ABIs at the moment. For now, only AIX is fixed. */
1495 if (DEFAULT_ABI != ABI_AIX)
1496 targetm.calls.split_complex_arg = NULL;
1497
1498 /* Initialize rs6000_cost with the appropriate target costs. */
1499 if (optimize_size)
1500 rs6000_cost = TARGET_POWERPC64 ? &size64_cost : &size32_cost;
1501 else
1502 switch (rs6000_cpu)
1503 {
1504 case PROCESSOR_RIOS1:
1505 rs6000_cost = &rios1_cost;
1506 break;
1507
1508 case PROCESSOR_RIOS2:
1509 rs6000_cost = &rios2_cost;
1510 break;
1511
1512 case PROCESSOR_RS64A:
1513 rs6000_cost = &rs64a_cost;
1514 break;
1515
1516 case PROCESSOR_MPCCORE:
1517 rs6000_cost = &mpccore_cost;
1518 break;
1519
1520 case PROCESSOR_PPC403:
1521 rs6000_cost = &ppc403_cost;
1522 break;
1523
1524 case PROCESSOR_PPC405:
1525 rs6000_cost = &ppc405_cost;
1526 break;
1527
1528 case PROCESSOR_PPC440:
1529 rs6000_cost = &ppc440_cost;
1530 break;
1531
1532 case PROCESSOR_PPC601:
1533 rs6000_cost = &ppc601_cost;
1534 break;
1535
1536 case PROCESSOR_PPC603:
1537 rs6000_cost = &ppc603_cost;
1538 break;
1539
1540 case PROCESSOR_PPC604:
1541 rs6000_cost = &ppc604_cost;
1542 break;
1543
1544 case PROCESSOR_PPC604e:
1545 rs6000_cost = &ppc604e_cost;
1546 break;
1547
1548 case PROCESSOR_PPC620:
1549 rs6000_cost = &ppc620_cost;
1550 break;
1551
1552 case PROCESSOR_PPC630:
1553 rs6000_cost = &ppc630_cost;
1554 break;
1555
1556 case PROCESSOR_PPC750:
1557 case PROCESSOR_PPC7400:
1558 rs6000_cost = &ppc750_cost;
1559 break;
1560
1561 case PROCESSOR_PPC7450:
1562 rs6000_cost = &ppc7450_cost;
1563 break;
1564
1565 case PROCESSOR_PPC8540:
1566 rs6000_cost = &ppc8540_cost;
1567 break;
1568
1569 case PROCESSOR_POWER4:
1570 case PROCESSOR_POWER5:
1571 rs6000_cost = &power4_cost;
1572 break;
1573
1574 default:
1575 gcc_unreachable ();
1576 }
1577 }
1578
1579 /* Implement targetm.vectorize.builtin_mask_for_load. */
1580 static tree
rs6000_builtin_mask_for_load(void)1581 rs6000_builtin_mask_for_load (void)
1582 {
1583 if (TARGET_ALTIVEC)
1584 return altivec_builtin_mask_for_load;
1585 else
1586 return 0;
1587 }
1588
1589
1590 /* Return true iff, data reference of TYPE can reach vector alignment (16)
1591 after applying N number of iterations. This routine does not determine
1592 how may iterations are required to reach desired alignment. */
1593
1594 static bool
rs6000_vector_alignment_reachable(tree type ATTRIBUTE_UNUSED,bool is_packed)1595 rs6000_vector_alignment_reachable (tree type ATTRIBUTE_UNUSED, bool is_packed)
1596 {
1597 if (is_packed)
1598 return false;
1599
1600 if (TARGET_32BIT)
1601 {
1602 if (rs6000_alignment_flags == MASK_ALIGN_NATURAL)
1603 return true;
1604
1605 if (rs6000_alignment_flags == MASK_ALIGN_POWER)
1606 return true;
1607
1608 return false;
1609 }
1610 else
1611 {
1612 if (TARGET_MACHO)
1613 return false;
1614
1615 /* Assuming that all other types are naturally aligned. CHECKME! */
1616 return true;
1617 }
1618 }
1619
1620 /* Handle generic options of the form -mfoo=yes/no.
1621 NAME is the option name.
1622 VALUE is the option value.
1623 FLAG is the pointer to the flag where to store a 1 or 0, depending on
1624 whether the option value is 'yes' or 'no' respectively. */
1625 static void
rs6000_parse_yes_no_option(const char * name,const char * value,int * flag)1626 rs6000_parse_yes_no_option (const char *name, const char *value, int *flag)
1627 {
1628 if (value == 0)
1629 return;
1630 else if (!strcmp (value, "yes"))
1631 *flag = 1;
1632 else if (!strcmp (value, "no"))
1633 *flag = 0;
1634 else
1635 error ("unknown -m%s= option specified: '%s'", name, value);
1636 }
1637
1638 /* Validate and record the size specified with the -mtls-size option. */
1639
1640 static void
rs6000_parse_tls_size_option(void)1641 rs6000_parse_tls_size_option (void)
1642 {
1643 if (rs6000_tls_size_string == 0)
1644 return;
1645 else if (strcmp (rs6000_tls_size_string, "16") == 0)
1646 rs6000_tls_size = 16;
1647 else if (strcmp (rs6000_tls_size_string, "32") == 0)
1648 rs6000_tls_size = 32;
1649 else if (strcmp (rs6000_tls_size_string, "64") == 0)
1650 rs6000_tls_size = 64;
1651 else
1652 error ("bad value %qs for -mtls-size switch", rs6000_tls_size_string);
1653 }
1654
1655 void
optimization_options(int level ATTRIBUTE_UNUSED,int size ATTRIBUTE_UNUSED)1656 optimization_options (int level ATTRIBUTE_UNUSED, int size ATTRIBUTE_UNUSED)
1657 {
1658 if (DEFAULT_ABI == ABI_DARWIN)
1659 /* The Darwin libraries never set errno, so we might as well
1660 avoid calling them when that's the only reason we would. */
1661 flag_errno_math = 0;
1662
1663 /* Double growth factor to counter reduced min jump length. */
1664 set_param_value ("max-grow-copy-bb-insns", 16);
1665
1666 /* Enable section anchors by default.
1667 Skip section anchors for Objective C and Objective C++
1668 until front-ends fixed. */
1669 if (!TARGET_MACHO && lang_hooks.name[4] != 'O')
1670 flag_section_anchors = 1;
1671 }
1672
1673 /* Implement TARGET_HANDLE_OPTION. */
1674
1675 static bool
rs6000_handle_option(size_t code,const char * arg,int value)1676 rs6000_handle_option (size_t code, const char *arg, int value)
1677 {
1678 switch (code)
1679 {
1680 case OPT_mno_power:
1681 target_flags &= ~(MASK_POWER | MASK_POWER2
1682 | MASK_MULTIPLE | MASK_STRING);
1683 target_flags_explicit |= (MASK_POWER | MASK_POWER2
1684 | MASK_MULTIPLE | MASK_STRING);
1685 break;
1686 case OPT_mno_powerpc:
1687 target_flags &= ~(MASK_POWERPC | MASK_PPC_GPOPT
1688 | MASK_PPC_GFXOPT | MASK_POWERPC64);
1689 target_flags_explicit |= (MASK_POWERPC | MASK_PPC_GPOPT
1690 | MASK_PPC_GFXOPT | MASK_POWERPC64);
1691 break;
1692 case OPT_mfull_toc:
1693 target_flags &= ~MASK_MINIMAL_TOC;
1694 TARGET_NO_FP_IN_TOC = 0;
1695 TARGET_NO_SUM_IN_TOC = 0;
1696 target_flags_explicit |= MASK_MINIMAL_TOC;
1697 #ifdef TARGET_USES_SYSV4_OPT
1698 /* Note, V.4 no longer uses a normal TOC, so make -mfull-toc, be
1699 just the same as -mminimal-toc. */
1700 target_flags |= MASK_MINIMAL_TOC;
1701 target_flags_explicit |= MASK_MINIMAL_TOC;
1702 #endif
1703 break;
1704
1705 #ifdef TARGET_USES_SYSV4_OPT
1706 case OPT_mtoc:
1707 /* Make -mtoc behave like -mminimal-toc. */
1708 target_flags |= MASK_MINIMAL_TOC;
1709 target_flags_explicit |= MASK_MINIMAL_TOC;
1710 break;
1711 #endif
1712
1713 #ifdef TARGET_USES_AIX64_OPT
1714 case OPT_maix64:
1715 #else
1716 case OPT_m64:
1717 #endif
1718 target_flags |= MASK_POWERPC64 | MASK_POWERPC;
1719 target_flags |= ~target_flags_explicit & MASK_PPC_GFXOPT;
1720 target_flags_explicit |= MASK_POWERPC64 | MASK_POWERPC;
1721 break;
1722
1723 #ifdef TARGET_USES_AIX64_OPT
1724 case OPT_maix32:
1725 #else
1726 case OPT_m32:
1727 #endif
1728 target_flags &= ~MASK_POWERPC64;
1729 target_flags_explicit |= MASK_POWERPC64;
1730 break;
1731
1732 case OPT_minsert_sched_nops_:
1733 rs6000_sched_insert_nops_str = arg;
1734 break;
1735
1736 case OPT_mminimal_toc:
1737 if (value == 1)
1738 {
1739 TARGET_NO_FP_IN_TOC = 0;
1740 TARGET_NO_SUM_IN_TOC = 0;
1741 }
1742 break;
1743
1744 case OPT_mpower:
1745 if (value == 1)
1746 {
1747 target_flags |= (MASK_MULTIPLE | MASK_STRING);
1748 target_flags_explicit |= (MASK_MULTIPLE | MASK_STRING);
1749 }
1750 break;
1751
1752 case OPT_mpower2:
1753 if (value == 1)
1754 {
1755 target_flags |= (MASK_POWER | MASK_MULTIPLE | MASK_STRING);
1756 target_flags_explicit |= (MASK_POWER | MASK_MULTIPLE | MASK_STRING);
1757 }
1758 break;
1759
1760 case OPT_mpowerpc_gpopt:
1761 case OPT_mpowerpc_gfxopt:
1762 if (value == 1)
1763 {
1764 target_flags |= MASK_POWERPC;
1765 target_flags_explicit |= MASK_POWERPC;
1766 }
1767 break;
1768
1769 case OPT_maix_struct_return:
1770 case OPT_msvr4_struct_return:
1771 rs6000_explicit_options.aix_struct_ret = true;
1772 break;
1773
1774 case OPT_mvrsave_:
1775 rs6000_parse_yes_no_option ("vrsave", arg, &(TARGET_ALTIVEC_VRSAVE));
1776 break;
1777
1778 case OPT_misel_:
1779 rs6000_explicit_options.isel = true;
1780 rs6000_parse_yes_no_option ("isel", arg, &(rs6000_isel));
1781 break;
1782
1783 case OPT_mspe_:
1784 rs6000_explicit_options.spe = true;
1785 rs6000_parse_yes_no_option ("spe", arg, &(rs6000_spe));
1786 /* No SPE means 64-bit long doubles, even if an E500. */
1787 if (!rs6000_spe)
1788 rs6000_long_double_type_size = 64;
1789 break;
1790
1791 case OPT_mdebug_:
1792 rs6000_debug_name = arg;
1793 break;
1794
1795 #ifdef TARGET_USES_SYSV4_OPT
1796 case OPT_mcall_:
1797 rs6000_abi_name = arg;
1798 break;
1799
1800 case OPT_msdata_:
1801 rs6000_sdata_name = arg;
1802 break;
1803
1804 case OPT_mtls_size_:
1805 rs6000_tls_size_string = arg;
1806 break;
1807
1808 case OPT_mrelocatable:
1809 if (value == 1)
1810 {
1811 target_flags |= MASK_MINIMAL_TOC;
1812 target_flags_explicit |= MASK_MINIMAL_TOC;
1813 TARGET_NO_FP_IN_TOC = 1;
1814 }
1815 break;
1816
1817 case OPT_mrelocatable_lib:
1818 if (value == 1)
1819 {
1820 target_flags |= MASK_RELOCATABLE | MASK_MINIMAL_TOC;
1821 target_flags_explicit |= MASK_RELOCATABLE | MASK_MINIMAL_TOC;
1822 TARGET_NO_FP_IN_TOC = 1;
1823 }
1824 else
1825 {
1826 target_flags &= ~MASK_RELOCATABLE;
1827 target_flags_explicit |= MASK_RELOCATABLE;
1828 }
1829 break;
1830 #endif
1831
1832 case OPT_mabi_:
1833 if (!strcmp (arg, "altivec"))
1834 {
1835 rs6000_explicit_options.abi = true;
1836 rs6000_altivec_abi = 1;
1837 rs6000_spe_abi = 0;
1838 }
1839 else if (! strcmp (arg, "no-altivec"))
1840 {
1841 /* ??? Don't set rs6000_explicit_options.abi here, to allow
1842 the default for rs6000_spe_abi to be chosen later. */
1843 rs6000_altivec_abi = 0;
1844 }
1845 else if (! strcmp (arg, "spe"))
1846 {
1847 rs6000_explicit_options.abi = true;
1848 rs6000_spe_abi = 1;
1849 rs6000_altivec_abi = 0;
1850 if (!TARGET_SPE_ABI)
1851 error ("not configured for ABI: '%s'", arg);
1852 }
1853 else if (! strcmp (arg, "no-spe"))
1854 {
1855 rs6000_explicit_options.abi = true;
1856 rs6000_spe_abi = 0;
1857 }
1858
1859 /* These are here for testing during development only, do not
1860 document in the manual please. */
1861 else if (! strcmp (arg, "d64"))
1862 {
1863 rs6000_darwin64_abi = 1;
1864 warning (0, "Using darwin64 ABI");
1865 }
1866 else if (! strcmp (arg, "d32"))
1867 {
1868 rs6000_darwin64_abi = 0;
1869 warning (0, "Using old darwin ABI");
1870 }
1871
1872 else if (! strcmp (arg, "ibmlongdouble"))
1873 {
1874 rs6000_explicit_options.ieee = true;
1875 rs6000_ieeequad = 0;
1876 warning (0, "Using IBM extended precision long double");
1877 }
1878 else if (! strcmp (arg, "ieeelongdouble"))
1879 {
1880 rs6000_explicit_options.ieee = true;
1881 rs6000_ieeequad = 1;
1882 warning (0, "Using IEEE extended precision long double");
1883 }
1884
1885 else
1886 {
1887 error ("unknown ABI specified: '%s'", arg);
1888 return false;
1889 }
1890 break;
1891
1892 case OPT_mcpu_:
1893 rs6000_select[1].string = arg;
1894 break;
1895
1896 case OPT_mtune_:
1897 rs6000_select[2].string = arg;
1898 break;
1899
1900 case OPT_mtraceback_:
1901 rs6000_traceback_name = arg;
1902 break;
1903
1904 case OPT_mfloat_gprs_:
1905 rs6000_explicit_options.float_gprs = true;
1906 if (! strcmp (arg, "yes") || ! strcmp (arg, "single"))
1907 rs6000_float_gprs = 1;
1908 else if (! strcmp (arg, "double"))
1909 rs6000_float_gprs = 2;
1910 else if (! strcmp (arg, "no"))
1911 rs6000_float_gprs = 0;
1912 else
1913 {
1914 error ("invalid option for -mfloat-gprs: '%s'", arg);
1915 return false;
1916 }
1917 break;
1918
1919 case OPT_mlong_double_:
1920 rs6000_explicit_options.long_double = true;
1921 rs6000_long_double_type_size = RS6000_DEFAULT_LONG_DOUBLE_SIZE;
1922 if (value != 64 && value != 128)
1923 {
1924 error ("Unknown switch -mlong-double-%s", arg);
1925 rs6000_long_double_type_size = RS6000_DEFAULT_LONG_DOUBLE_SIZE;
1926 return false;
1927 }
1928 else
1929 rs6000_long_double_type_size = value;
1930 break;
1931
1932 case OPT_msched_costly_dep_:
1933 rs6000_sched_costly_dep_str = arg;
1934 break;
1935
1936 case OPT_malign_:
1937 rs6000_explicit_options.alignment = true;
1938 if (! strcmp (arg, "power"))
1939 {
1940 /* On 64-bit Darwin, power alignment is ABI-incompatible with
1941 some C library functions, so warn about it. The flag may be
1942 useful for performance studies from time to time though, so
1943 don't disable it entirely. */
1944 if (DEFAULT_ABI == ABI_DARWIN && TARGET_64BIT)
1945 warning (0, "-malign-power is not supported for 64-bit Darwin;"
1946 " it is incompatible with the installed C and C++ libraries");
1947 rs6000_alignment_flags = MASK_ALIGN_POWER;
1948 }
1949 else if (! strcmp (arg, "natural"))
1950 rs6000_alignment_flags = MASK_ALIGN_NATURAL;
1951 else
1952 {
1953 error ("unknown -malign-XXXXX option specified: '%s'", arg);
1954 return false;
1955 }
1956 break;
1957 }
1958 return true;
1959 }
1960
1961 /* Do anything needed at the start of the asm file. */
1962
1963 static void
rs6000_file_start(void)1964 rs6000_file_start (void)
1965 {
1966 size_t i;
1967 char buffer[80];
1968 const char *start = buffer;
1969 struct rs6000_cpu_select *ptr;
1970 const char *default_cpu = TARGET_CPU_DEFAULT;
1971 FILE *file = asm_out_file;
1972
1973 default_file_start ();
1974
1975 #ifdef TARGET_BI_ARCH
1976 if ((TARGET_DEFAULT ^ target_flags) & MASK_64BIT)
1977 default_cpu = 0;
1978 #endif
1979
1980 if (flag_verbose_asm)
1981 {
1982 sprintf (buffer, "\n%s rs6000/powerpc options:", ASM_COMMENT_START);
1983 rs6000_select[0].string = default_cpu;
1984
1985 for (i = 0; i < ARRAY_SIZE (rs6000_select); i++)
1986 {
1987 ptr = &rs6000_select[i];
1988 if (ptr->string != (char *)0 && ptr->string[0] != '\0')
1989 {
1990 fprintf (file, "%s %s%s", start, ptr->name, ptr->string);
1991 start = "";
1992 }
1993 }
1994
1995 if (PPC405_ERRATUM77)
1996 {
1997 fprintf (file, "%s PPC405CR_ERRATUM77", start);
1998 start = "";
1999 }
2000
2001 #ifdef USING_ELFOS_H
2002 switch (rs6000_sdata)
2003 {
2004 case SDATA_NONE: fprintf (file, "%s -msdata=none", start); start = ""; break;
2005 case SDATA_DATA: fprintf (file, "%s -msdata=data", start); start = ""; break;
2006 case SDATA_SYSV: fprintf (file, "%s -msdata=sysv", start); start = ""; break;
2007 case SDATA_EABI: fprintf (file, "%s -msdata=eabi", start); start = ""; break;
2008 }
2009
2010 if (rs6000_sdata && g_switch_value)
2011 {
2012 fprintf (file, "%s -G " HOST_WIDE_INT_PRINT_UNSIGNED, start,
2013 g_switch_value);
2014 start = "";
2015 }
2016 #endif
2017
2018 if (*start == '\0')
2019 putc ('\n', file);
2020 }
2021
2022 if (DEFAULT_ABI == ABI_AIX || (TARGET_ELF && flag_pic == 2))
2023 {
2024 switch_to_section (toc_section);
2025 switch_to_section (text_section);
2026 }
2027 }
2028
2029
2030 /* Return nonzero if this function is known to have a null epilogue. */
2031
2032 int
direct_return(void)2033 direct_return (void)
2034 {
2035 if (reload_completed)
2036 {
2037 rs6000_stack_t *info = rs6000_stack_info ();
2038
2039 if (info->first_gp_reg_save == 32
2040 && info->first_fp_reg_save == 64
2041 && info->first_altivec_reg_save == LAST_ALTIVEC_REGNO + 1
2042 && ! info->lr_save_p
2043 && ! info->cr_save_p
2044 && info->vrsave_mask == 0
2045 && ! info->push_p)
2046 return 1;
2047 }
2048
2049 return 0;
2050 }
2051
2052 /* Return the number of instructions it takes to form a constant in an
2053 integer register. */
2054
2055 int
num_insns_constant_wide(HOST_WIDE_INT value)2056 num_insns_constant_wide (HOST_WIDE_INT value)
2057 {
2058 /* signed constant loadable with {cal|addi} */
2059 if ((unsigned HOST_WIDE_INT) (value + 0x8000) < 0x10000)
2060 return 1;
2061
2062 /* constant loadable with {cau|addis} */
2063 else if ((value & 0xffff) == 0
2064 && (value >> 31 == -1 || value >> 31 == 0))
2065 return 1;
2066
2067 #if HOST_BITS_PER_WIDE_INT == 64
2068 else if (TARGET_POWERPC64)
2069 {
2070 HOST_WIDE_INT low = ((value & 0xffffffff) ^ 0x80000000) - 0x80000000;
2071 HOST_WIDE_INT high = value >> 31;
2072
2073 if (high == 0 || high == -1)
2074 return 2;
2075
2076 high >>= 1;
2077
2078 if (low == 0)
2079 return num_insns_constant_wide (high) + 1;
2080 else
2081 return (num_insns_constant_wide (high)
2082 + num_insns_constant_wide (low) + 1);
2083 }
2084 #endif
2085
2086 else
2087 return 2;
2088 }
2089
2090 int
num_insns_constant(rtx op,enum machine_mode mode)2091 num_insns_constant (rtx op, enum machine_mode mode)
2092 {
2093 HOST_WIDE_INT low, high;
2094
2095 switch (GET_CODE (op))
2096 {
2097 case CONST_INT:
2098 #if HOST_BITS_PER_WIDE_INT == 64
2099 if ((INTVAL (op) >> 31) != 0 && (INTVAL (op) >> 31) != -1
2100 && mask64_operand (op, mode))
2101 return 2;
2102 else
2103 #endif
2104 return num_insns_constant_wide (INTVAL (op));
2105
2106 case CONST_DOUBLE:
2107 if (mode == SFmode)
2108 {
2109 long l;
2110 REAL_VALUE_TYPE rv;
2111
2112 REAL_VALUE_FROM_CONST_DOUBLE (rv, op);
2113 REAL_VALUE_TO_TARGET_SINGLE (rv, l);
2114 return num_insns_constant_wide ((HOST_WIDE_INT) l);
2115 }
2116
2117 if (mode == VOIDmode || mode == DImode)
2118 {
2119 high = CONST_DOUBLE_HIGH (op);
2120 low = CONST_DOUBLE_LOW (op);
2121 }
2122 else
2123 {
2124 long l[2];
2125 REAL_VALUE_TYPE rv;
2126
2127 REAL_VALUE_FROM_CONST_DOUBLE (rv, op);
2128 REAL_VALUE_TO_TARGET_DOUBLE (rv, l);
2129 high = l[WORDS_BIG_ENDIAN == 0];
2130 low = l[WORDS_BIG_ENDIAN != 0];
2131 }
2132
2133 if (TARGET_32BIT)
2134 return (num_insns_constant_wide (low)
2135 + num_insns_constant_wide (high));
2136 else
2137 {
2138 if ((high == 0 && low >= 0)
2139 || (high == -1 && low < 0))
2140 return num_insns_constant_wide (low);
2141
2142 else if (mask64_operand (op, mode))
2143 return 2;
2144
2145 else if (low == 0)
2146 return num_insns_constant_wide (high) + 1;
2147
2148 else
2149 return (num_insns_constant_wide (high)
2150 + num_insns_constant_wide (low) + 1);
2151 }
2152
2153 default:
2154 gcc_unreachable ();
2155 }
2156 }
2157
2158 /* Interpret element ELT of the CONST_VECTOR OP as an integer value.
2159 If the mode of OP is MODE_VECTOR_INT, this simply returns the
2160 corresponding element of the vector, but for V4SFmode and V2SFmode,
2161 the corresponding "float" is interpreted as an SImode integer. */
2162
2163 static HOST_WIDE_INT
const_vector_elt_as_int(rtx op,unsigned int elt)2164 const_vector_elt_as_int (rtx op, unsigned int elt)
2165 {
2166 rtx tmp = CONST_VECTOR_ELT (op, elt);
2167 if (GET_MODE (op) == V4SFmode
2168 || GET_MODE (op) == V2SFmode)
2169 tmp = gen_lowpart (SImode, tmp);
2170 return INTVAL (tmp);
2171 }
2172
2173 /* Return true if OP can be synthesized with a particular vspltisb, vspltish
2174 or vspltisw instruction. OP is a CONST_VECTOR. Which instruction is used
2175 depends on STEP and COPIES, one of which will be 1. If COPIES > 1,
2176 all items are set to the same value and contain COPIES replicas of the
2177 vsplt's operand; if STEP > 1, one in STEP elements is set to the vsplt's
2178 operand and the others are set to the value of the operand's msb. */
2179
2180 static bool
vspltis_constant(rtx op,unsigned step,unsigned copies)2181 vspltis_constant (rtx op, unsigned step, unsigned copies)
2182 {
2183 enum machine_mode mode = GET_MODE (op);
2184 enum machine_mode inner = GET_MODE_INNER (mode);
2185
2186 unsigned i;
2187 unsigned nunits = GET_MODE_NUNITS (mode);
2188 unsigned bitsize = GET_MODE_BITSIZE (inner);
2189 unsigned mask = GET_MODE_MASK (inner);
2190
2191 HOST_WIDE_INT val = const_vector_elt_as_int (op, nunits - 1);
2192 HOST_WIDE_INT splat_val = val;
2193 HOST_WIDE_INT msb_val = val > 0 ? 0 : -1;
2194
2195 /* Construct the value to be splatted, if possible. If not, return 0. */
2196 for (i = 2; i <= copies; i *= 2)
2197 {
2198 HOST_WIDE_INT small_val;
2199 bitsize /= 2;
2200 small_val = splat_val >> bitsize;
2201 mask >>= bitsize;
2202 if (splat_val != ((small_val << bitsize) | (small_val & mask)))
2203 return false;
2204 splat_val = small_val;
2205 }
2206
2207 /* Check if SPLAT_VAL can really be the operand of a vspltis[bhw]. */
2208 if (EASY_VECTOR_15 (splat_val))
2209 ;
2210
2211 /* Also check if we can splat, and then add the result to itself. Do so if
2212 the value is positive, of if the splat instruction is using OP's mode;
2213 for splat_val < 0, the splat and the add should use the same mode. */
2214 else if (EASY_VECTOR_15_ADD_SELF (splat_val)
2215 && (splat_val >= 0 || (step == 1 && copies == 1)))
2216 ;
2217
2218 else
2219 return false;
2220
2221 /* Check if VAL is present in every STEP-th element, and the
2222 other elements are filled with its most significant bit. */
2223 for (i = 0; i < nunits - 1; ++i)
2224 {
2225 HOST_WIDE_INT desired_val;
2226 if (((i + 1) & (step - 1)) == 0)
2227 desired_val = val;
2228 else
2229 desired_val = msb_val;
2230
2231 if (desired_val != const_vector_elt_as_int (op, i))
2232 return false;
2233 }
2234
2235 return true;
2236 }
2237
2238
2239 /* Return true if OP is of the given MODE and can be synthesized
2240 with a vspltisb, vspltish or vspltisw. */
2241
2242 bool
easy_altivec_constant(rtx op,enum machine_mode mode)2243 easy_altivec_constant (rtx op, enum machine_mode mode)
2244 {
2245 unsigned step, copies;
2246
2247 if (mode == VOIDmode)
2248 mode = GET_MODE (op);
2249 else if (mode != GET_MODE (op))
2250 return false;
2251
2252 /* Start with a vspltisw. */
2253 step = GET_MODE_NUNITS (mode) / 4;
2254 copies = 1;
2255
2256 if (vspltis_constant (op, step, copies))
2257 return true;
2258
2259 /* Then try with a vspltish. */
2260 if (step == 1)
2261 copies <<= 1;
2262 else
2263 step >>= 1;
2264
2265 if (vspltis_constant (op, step, copies))
2266 return true;
2267
2268 /* And finally a vspltisb. */
2269 if (step == 1)
2270 copies <<= 1;
2271 else
2272 step >>= 1;
2273
2274 if (vspltis_constant (op, step, copies))
2275 return true;
2276
2277 return false;
2278 }
2279
2280 /* Generate a VEC_DUPLICATE representing a vspltis[bhw] instruction whose
2281 result is OP. Abort if it is not possible. */
2282
2283 rtx
gen_easy_altivec_constant(rtx op)2284 gen_easy_altivec_constant (rtx op)
2285 {
2286 enum machine_mode mode = GET_MODE (op);
2287 int nunits = GET_MODE_NUNITS (mode);
2288 rtx last = CONST_VECTOR_ELT (op, nunits - 1);
2289 unsigned step = nunits / 4;
2290 unsigned copies = 1;
2291
2292 /* Start with a vspltisw. */
2293 if (vspltis_constant (op, step, copies))
2294 return gen_rtx_VEC_DUPLICATE (V4SImode, gen_lowpart (SImode, last));
2295
2296 /* Then try with a vspltish. */
2297 if (step == 1)
2298 copies <<= 1;
2299 else
2300 step >>= 1;
2301
2302 if (vspltis_constant (op, step, copies))
2303 return gen_rtx_VEC_DUPLICATE (V8HImode, gen_lowpart (HImode, last));
2304
2305 /* And finally a vspltisb. */
2306 if (step == 1)
2307 copies <<= 1;
2308 else
2309 step >>= 1;
2310
2311 if (vspltis_constant (op, step, copies))
2312 return gen_rtx_VEC_DUPLICATE (V16QImode, gen_lowpart (QImode, last));
2313
2314 gcc_unreachable ();
2315 }
2316
2317 const char *
output_vec_const_move(rtx * operands)2318 output_vec_const_move (rtx *operands)
2319 {
2320 int cst, cst2;
2321 enum machine_mode mode;
2322 rtx dest, vec;
2323
2324 dest = operands[0];
2325 vec = operands[1];
2326 mode = GET_MODE (dest);
2327
2328 if (TARGET_ALTIVEC)
2329 {
2330 rtx splat_vec;
2331 if (zero_constant (vec, mode))
2332 return "vxor %0,%0,%0";
2333
2334 splat_vec = gen_easy_altivec_constant (vec);
2335 gcc_assert (GET_CODE (splat_vec) == VEC_DUPLICATE);
2336 operands[1] = XEXP (splat_vec, 0);
2337 if (!EASY_VECTOR_15 (INTVAL (operands[1])))
2338 return "#";
2339
2340 switch (GET_MODE (splat_vec))
2341 {
2342 case V4SImode:
2343 return "vspltisw %0,%1";
2344
2345 case V8HImode:
2346 return "vspltish %0,%1";
2347
2348 case V16QImode:
2349 return "vspltisb %0,%1";
2350
2351 default:
2352 gcc_unreachable ();
2353 }
2354 }
2355
2356 gcc_assert (TARGET_SPE);
2357
2358 /* Vector constant 0 is handled as a splitter of V2SI, and in the
2359 pattern of V1DI, V4HI, and V2SF.
2360
2361 FIXME: We should probably return # and add post reload
2362 splitters for these, but this way is so easy ;-). */
2363 cst = INTVAL (CONST_VECTOR_ELT (vec, 0));
2364 cst2 = INTVAL (CONST_VECTOR_ELT (vec, 1));
2365 operands[1] = CONST_VECTOR_ELT (vec, 0);
2366 operands[2] = CONST_VECTOR_ELT (vec, 1);
2367 if (cst == cst2)
2368 return "li %0,%1\n\tevmergelo %0,%0,%0";
2369 else
2370 return "li %0,%1\n\tevmergelo %0,%0,%0\n\tli %0,%2";
2371 }
2372
2373 /* Initialize vector TARGET to VALS. */
2374
2375 void
rs6000_expand_vector_init(rtx target,rtx vals)2376 rs6000_expand_vector_init (rtx target, rtx vals)
2377 {
2378 enum machine_mode mode = GET_MODE (target);
2379 enum machine_mode inner_mode = GET_MODE_INNER (mode);
2380 int n_elts = GET_MODE_NUNITS (mode);
2381 int n_var = 0, one_var = -1;
2382 bool all_same = true, all_const_zero = true;
2383 rtx x, mem;
2384 int i;
2385
2386 for (i = 0; i < n_elts; ++i)
2387 {
2388 x = XVECEXP (vals, 0, i);
2389 if (!CONSTANT_P (x))
2390 ++n_var, one_var = i;
2391 else if (x != CONST0_RTX (inner_mode))
2392 all_const_zero = false;
2393
2394 if (i > 0 && !rtx_equal_p (x, XVECEXP (vals, 0, 0)))
2395 all_same = false;
2396 }
2397
2398 if (n_var == 0)
2399 {
2400 if (mode != V4SFmode && all_const_zero)
2401 {
2402 /* Zero register. */
2403 emit_insn (gen_rtx_SET (VOIDmode, target,
2404 gen_rtx_XOR (mode, target, target)));
2405 return;
2406 }
2407 else if (mode != V4SFmode && easy_vector_constant (vals, mode))
2408 {
2409 /* Splat immediate. */
2410 emit_insn (gen_rtx_SET (VOIDmode, target, vals));
2411 return;
2412 }
2413 else if (all_same)
2414 ; /* Splat vector element. */
2415 else
2416 {
2417 /* Load from constant pool. */
2418 emit_move_insn (target, gen_rtx_CONST_VECTOR (mode, XVEC (vals, 0)));
2419 return;
2420 }
2421 }
2422
2423 /* Store value to stack temp. Load vector element. Splat. */
2424 if (all_same)
2425 {
2426 mem = assign_stack_temp (mode, GET_MODE_SIZE (inner_mode), 0);
2427 emit_move_insn (adjust_address_nv (mem, inner_mode, 0),
2428 XVECEXP (vals, 0, 0));
2429 x = gen_rtx_UNSPEC (VOIDmode,
2430 gen_rtvec (1, const0_rtx), UNSPEC_LVE);
2431 emit_insn (gen_rtx_PARALLEL (VOIDmode,
2432 gen_rtvec (2,
2433 gen_rtx_SET (VOIDmode,
2434 target, mem),
2435 x)));
2436 x = gen_rtx_VEC_SELECT (inner_mode, target,
2437 gen_rtx_PARALLEL (VOIDmode,
2438 gen_rtvec (1, const0_rtx)));
2439 emit_insn (gen_rtx_SET (VOIDmode, target,
2440 gen_rtx_VEC_DUPLICATE (mode, x)));
2441 return;
2442 }
2443
2444 /* One field is non-constant. Load constant then overwrite
2445 varying field. */
2446 if (n_var == 1)
2447 {
2448 rtx copy = copy_rtx (vals);
2449
2450 /* Load constant part of vector, substitute neighboring value for
2451 varying element. */
2452 XVECEXP (copy, 0, one_var) = XVECEXP (vals, 0, (one_var + 1) % n_elts);
2453 rs6000_expand_vector_init (target, copy);
2454
2455 /* Insert variable. */
2456 rs6000_expand_vector_set (target, XVECEXP (vals, 0, one_var), one_var);
2457 return;
2458 }
2459
2460 /* Construct the vector in memory one field at a time
2461 and load the whole vector. */
2462 mem = assign_stack_temp (mode, GET_MODE_SIZE (mode), 0);
2463 for (i = 0; i < n_elts; i++)
2464 emit_move_insn (adjust_address_nv (mem, inner_mode,
2465 i * GET_MODE_SIZE (inner_mode)),
2466 XVECEXP (vals, 0, i));
2467 emit_move_insn (target, mem);
2468 }
2469
2470 /* Set field ELT of TARGET to VAL. */
2471
2472 void
rs6000_expand_vector_set(rtx target,rtx val,int elt)2473 rs6000_expand_vector_set (rtx target, rtx val, int elt)
2474 {
2475 enum machine_mode mode = GET_MODE (target);
2476 enum machine_mode inner_mode = GET_MODE_INNER (mode);
2477 rtx reg = gen_reg_rtx (mode);
2478 rtx mask, mem, x;
2479 int width = GET_MODE_SIZE (inner_mode);
2480 int i;
2481
2482 /* Load single variable value. */
2483 mem = assign_stack_temp (mode, GET_MODE_SIZE (inner_mode), 0);
2484 emit_move_insn (adjust_address_nv (mem, inner_mode, 0), val);
2485 x = gen_rtx_UNSPEC (VOIDmode,
2486 gen_rtvec (1, const0_rtx), UNSPEC_LVE);
2487 emit_insn (gen_rtx_PARALLEL (VOIDmode,
2488 gen_rtvec (2,
2489 gen_rtx_SET (VOIDmode,
2490 reg, mem),
2491 x)));
2492
2493 /* Linear sequence. */
2494 mask = gen_rtx_PARALLEL (V16QImode, rtvec_alloc (16));
2495 for (i = 0; i < 16; ++i)
2496 XVECEXP (mask, 0, i) = GEN_INT (i);
2497
2498 /* Set permute mask to insert element into target. */
2499 for (i = 0; i < width; ++i)
2500 XVECEXP (mask, 0, elt*width + i)
2501 = GEN_INT (i + 0x10);
2502 x = gen_rtx_CONST_VECTOR (V16QImode, XVEC (mask, 0));
2503 x = gen_rtx_UNSPEC (mode,
2504 gen_rtvec (3, target, reg,
2505 force_reg (V16QImode, x)),
2506 UNSPEC_VPERM);
2507 emit_insn (gen_rtx_SET (VOIDmode, target, x));
2508 }
2509
2510 /* Extract field ELT from VEC into TARGET. */
2511
2512 void
rs6000_expand_vector_extract(rtx target,rtx vec,int elt)2513 rs6000_expand_vector_extract (rtx target, rtx vec, int elt)
2514 {
2515 enum machine_mode mode = GET_MODE (vec);
2516 enum machine_mode inner_mode = GET_MODE_INNER (mode);
2517 rtx mem, x;
2518
2519 /* Allocate mode-sized buffer. */
2520 mem = assign_stack_temp (mode, GET_MODE_SIZE (mode), 0);
2521
2522 /* Add offset to field within buffer matching vector element. */
2523 mem = adjust_address_nv (mem, mode, elt * GET_MODE_SIZE (inner_mode));
2524
2525 /* Store single field into mode-sized buffer. */
2526 x = gen_rtx_UNSPEC (VOIDmode,
2527 gen_rtvec (1, const0_rtx), UNSPEC_STVE);
2528 emit_insn (gen_rtx_PARALLEL (VOIDmode,
2529 gen_rtvec (2,
2530 gen_rtx_SET (VOIDmode,
2531 mem, vec),
2532 x)));
2533 emit_move_insn (target, adjust_address_nv (mem, inner_mode, 0));
2534 }
2535
2536 /* Generates shifts and masks for a pair of rldicl or rldicr insns to
2537 implement ANDing by the mask IN. */
2538 void
build_mask64_2_operands(rtx in,rtx * out)2539 build_mask64_2_operands (rtx in, rtx *out)
2540 {
2541 #if HOST_BITS_PER_WIDE_INT >= 64
2542 unsigned HOST_WIDE_INT c, lsb, m1, m2;
2543 int shift;
2544
2545 gcc_assert (GET_CODE (in) == CONST_INT);
2546
2547 c = INTVAL (in);
2548 if (c & 1)
2549 {
2550 /* Assume c initially something like 0x00fff000000fffff. The idea
2551 is to rotate the word so that the middle ^^^^^^ group of zeros
2552 is at the MS end and can be cleared with an rldicl mask. We then
2553 rotate back and clear off the MS ^^ group of zeros with a
2554 second rldicl. */
2555 c = ~c; /* c == 0xff000ffffff00000 */
2556 lsb = c & -c; /* lsb == 0x0000000000100000 */
2557 m1 = -lsb; /* m1 == 0xfffffffffff00000 */
2558 c = ~c; /* c == 0x00fff000000fffff */
2559 c &= -lsb; /* c == 0x00fff00000000000 */
2560 lsb = c & -c; /* lsb == 0x0000100000000000 */
2561 c = ~c; /* c == 0xff000fffffffffff */
2562 c &= -lsb; /* c == 0xff00000000000000 */
2563 shift = 0;
2564 while ((lsb >>= 1) != 0)
2565 shift++; /* shift == 44 on exit from loop */
2566 m1 <<= 64 - shift; /* m1 == 0xffffff0000000000 */
2567 m1 = ~m1; /* m1 == 0x000000ffffffffff */
2568 m2 = ~c; /* m2 == 0x00ffffffffffffff */
2569 }
2570 else
2571 {
2572 /* Assume c initially something like 0xff000f0000000000. The idea
2573 is to rotate the word so that the ^^^ middle group of zeros
2574 is at the LS end and can be cleared with an rldicr mask. We then
2575 rotate back and clear off the LS group of ^^^^^^^^^^ zeros with
2576 a second rldicr. */
2577 lsb = c & -c; /* lsb == 0x0000010000000000 */
2578 m2 = -lsb; /* m2 == 0xffffff0000000000 */
2579 c = ~c; /* c == 0x00fff0ffffffffff */
2580 c &= -lsb; /* c == 0x00fff00000000000 */
2581 lsb = c & -c; /* lsb == 0x0000100000000000 */
2582 c = ~c; /* c == 0xff000fffffffffff */
2583 c &= -lsb; /* c == 0xff00000000000000 */
2584 shift = 0;
2585 while ((lsb >>= 1) != 0)
2586 shift++; /* shift == 44 on exit from loop */
2587 m1 = ~c; /* m1 == 0x00ffffffffffffff */
2588 m1 >>= shift; /* m1 == 0x0000000000000fff */
2589 m1 = ~m1; /* m1 == 0xfffffffffffff000 */
2590 }
2591
2592 /* Note that when we only have two 0->1 and 1->0 transitions, one of the
2593 masks will be all 1's. We are guaranteed more than one transition. */
2594 out[0] = GEN_INT (64 - shift);
2595 out[1] = GEN_INT (m1);
2596 out[2] = GEN_INT (shift);
2597 out[3] = GEN_INT (m2);
2598 #else
2599 (void)in;
2600 (void)out;
2601 gcc_unreachable ();
2602 #endif
2603 }
2604
2605 /* Return TRUE if OP is an invalid SUBREG operation on the e500. */
2606
2607 bool
invalid_e500_subreg(rtx op,enum machine_mode mode)2608 invalid_e500_subreg (rtx op, enum machine_mode mode)
2609 {
2610 if (TARGET_E500_DOUBLE)
2611 {
2612 /* Reject (subreg:SI (reg:DF)). */
2613 if (GET_CODE (op) == SUBREG
2614 && mode == SImode
2615 && REG_P (SUBREG_REG (op))
2616 && GET_MODE (SUBREG_REG (op)) == DFmode)
2617 return true;
2618
2619 /* Reject (subreg:DF (reg:DI)). */
2620 if (GET_CODE (op) == SUBREG
2621 && mode == DFmode
2622 && REG_P (SUBREG_REG (op))
2623 && GET_MODE (SUBREG_REG (op)) == DImode)
2624 return true;
2625 }
2626
2627 if (TARGET_SPE
2628 && GET_CODE (op) == SUBREG
2629 && mode == SImode
2630 && REG_P (SUBREG_REG (op))
2631 && SPE_VECTOR_MODE (GET_MODE (SUBREG_REG (op))))
2632 return true;
2633
2634 return false;
2635 }
2636
2637 /* Darwin, AIX increases natural record alignment to doubleword if the first
2638 field is an FP double while the FP fields remain word aligned. */
2639
2640 unsigned int
rs6000_special_round_type_align(tree type,unsigned int computed,unsigned int specified)2641 rs6000_special_round_type_align (tree type, unsigned int computed,
2642 unsigned int specified)
2643 {
2644 unsigned int align = MAX (computed, specified);
2645 tree field = TYPE_FIELDS (type);
2646
2647 /* Skip all non field decls */
2648 while (field != NULL && TREE_CODE (field) != FIELD_DECL)
2649 field = TREE_CHAIN (field);
2650
2651 if (field != NULL && field != type)
2652 {
2653 type = TREE_TYPE (field);
2654 while (TREE_CODE (type) == ARRAY_TYPE)
2655 type = TREE_TYPE (type);
2656
2657 if (type != error_mark_node && TYPE_MODE (type) == DFmode)
2658 align = MAX (align, 64);
2659 }
2660
2661 return align;
2662 }
2663
2664 /* Return 1 for an operand in small memory on V.4/eabi. */
2665
2666 int
small_data_operand(rtx op ATTRIBUTE_UNUSED,enum machine_mode mode ATTRIBUTE_UNUSED)2667 small_data_operand (rtx op ATTRIBUTE_UNUSED,
2668 enum machine_mode mode ATTRIBUTE_UNUSED)
2669 {
2670 #if TARGET_ELF
2671 rtx sym_ref;
2672
2673 if (rs6000_sdata == SDATA_NONE || rs6000_sdata == SDATA_DATA)
2674 return 0;
2675
2676 if (DEFAULT_ABI != ABI_V4)
2677 return 0;
2678
2679 if (GET_CODE (op) == SYMBOL_REF)
2680 sym_ref = op;
2681
2682 else if (GET_CODE (op) != CONST
2683 || GET_CODE (XEXP (op, 0)) != PLUS
2684 || GET_CODE (XEXP (XEXP (op, 0), 0)) != SYMBOL_REF
2685 || GET_CODE (XEXP (XEXP (op, 0), 1)) != CONST_INT)
2686 return 0;
2687
2688 else
2689 {
2690 rtx sum = XEXP (op, 0);
2691 HOST_WIDE_INT summand;
2692
2693 /* We have to be careful here, because it is the referenced address
2694 that must be 32k from _SDA_BASE_, not just the symbol. */
2695 summand = INTVAL (XEXP (sum, 1));
2696 if (summand < 0 || (unsigned HOST_WIDE_INT) summand > g_switch_value)
2697 return 0;
2698
2699 sym_ref = XEXP (sum, 0);
2700 }
2701
2702 return SYMBOL_REF_SMALL_P (sym_ref);
2703 #else
2704 return 0;
2705 #endif
2706 }
2707
2708 /* Return true if either operand is a general purpose register. */
2709
2710 bool
gpr_or_gpr_p(rtx op0,rtx op1)2711 gpr_or_gpr_p (rtx op0, rtx op1)
2712 {
2713 return ((REG_P (op0) && INT_REGNO_P (REGNO (op0)))
2714 || (REG_P (op1) && INT_REGNO_P (REGNO (op1))));
2715 }
2716
2717
2718 /* Subroutines of rs6000_legitimize_address and rs6000_legitimate_address. */
2719
2720 static int
constant_pool_expr_1(rtx op,int * have_sym,int * have_toc)2721 constant_pool_expr_1 (rtx op, int *have_sym, int *have_toc)
2722 {
2723 switch (GET_CODE (op))
2724 {
2725 case SYMBOL_REF:
2726 if (RS6000_SYMBOL_REF_TLS_P (op))
2727 return 0;
2728 else if (CONSTANT_POOL_ADDRESS_P (op))
2729 {
2730 if (ASM_OUTPUT_SPECIAL_POOL_ENTRY_P (get_pool_constant (op), Pmode))
2731 {
2732 *have_sym = 1;
2733 return 1;
2734 }
2735 else
2736 return 0;
2737 }
2738 else if (! strcmp (XSTR (op, 0), toc_label_name))
2739 {
2740 *have_toc = 1;
2741 return 1;
2742 }
2743 else
2744 return 0;
2745 case PLUS:
2746 case MINUS:
2747 return (constant_pool_expr_1 (XEXP (op, 0), have_sym, have_toc)
2748 && constant_pool_expr_1 (XEXP (op, 1), have_sym, have_toc));
2749 case CONST:
2750 return constant_pool_expr_1 (XEXP (op, 0), have_sym, have_toc);
2751 case CONST_INT:
2752 return 1;
2753 default:
2754 return 0;
2755 }
2756 }
2757
2758 static bool
constant_pool_expr_p(rtx op)2759 constant_pool_expr_p (rtx op)
2760 {
2761 int have_sym = 0;
2762 int have_toc = 0;
2763 return constant_pool_expr_1 (op, &have_sym, &have_toc) && have_sym;
2764 }
2765
2766 bool
toc_relative_expr_p(rtx op)2767 toc_relative_expr_p (rtx op)
2768 {
2769 int have_sym = 0;
2770 int have_toc = 0;
2771 return constant_pool_expr_1 (op, &have_sym, &have_toc) && have_toc;
2772 }
2773
2774 bool
legitimate_constant_pool_address_p(rtx x)2775 legitimate_constant_pool_address_p (rtx x)
2776 {
2777 return (TARGET_TOC
2778 && GET_CODE (x) == PLUS
2779 && GET_CODE (XEXP (x, 0)) == REG
2780 && (TARGET_MINIMAL_TOC || REGNO (XEXP (x, 0)) == TOC_REGISTER)
2781 && constant_pool_expr_p (XEXP (x, 1)));
2782 }
2783
2784 static bool
legitimate_small_data_p(enum machine_mode mode,rtx x)2785 legitimate_small_data_p (enum machine_mode mode, rtx x)
2786 {
2787 return (DEFAULT_ABI == ABI_V4
2788 && !flag_pic && !TARGET_TOC
2789 && (GET_CODE (x) == SYMBOL_REF || GET_CODE (x) == CONST)
2790 && small_data_operand (x, mode));
2791 }
2792
2793 /* SPE offset addressing is limited to 5-bits worth of double words. */
2794 #define SPE_CONST_OFFSET_OK(x) (((x) & ~0xf8) == 0)
2795
2796 bool
rs6000_legitimate_offset_address_p(enum machine_mode mode,rtx x,int strict)2797 rs6000_legitimate_offset_address_p (enum machine_mode mode, rtx x, int strict)
2798 {
2799 unsigned HOST_WIDE_INT offset, extra;
2800
2801 if (GET_CODE (x) != PLUS)
2802 return false;
2803 if (GET_CODE (XEXP (x, 0)) != REG)
2804 return false;
2805 if (!INT_REG_OK_FOR_BASE_P (XEXP (x, 0), strict))
2806 return false;
2807 if (legitimate_constant_pool_address_p (x))
2808 return true;
2809 if (GET_CODE (XEXP (x, 1)) != CONST_INT)
2810 return false;
2811
2812 offset = INTVAL (XEXP (x, 1));
2813 extra = 0;
2814 switch (mode)
2815 {
2816 case V16QImode:
2817 case V8HImode:
2818 case V4SFmode:
2819 case V4SImode:
2820 /* AltiVec vector modes. Only reg+reg addressing is valid and
2821 constant offset zero should not occur due to canonicalization.
2822 Allow any offset when not strict before reload. */
2823 return !strict;
2824
2825 case V4HImode:
2826 case V2SImode:
2827 case V1DImode:
2828 case V2SFmode:
2829 /* SPE vector modes. */
2830 return SPE_CONST_OFFSET_OK (offset);
2831
2832 case DFmode:
2833 if (TARGET_E500_DOUBLE)
2834 return SPE_CONST_OFFSET_OK (offset);
2835
2836 case DImode:
2837 /* On e500v2, we may have:
2838
2839 (subreg:DF (mem:DI (plus (reg) (const_int))) 0).
2840
2841 Which gets addressed with evldd instructions. */
2842 if (TARGET_E500_DOUBLE)
2843 return SPE_CONST_OFFSET_OK (offset);
2844
2845 if (mode == DFmode || !TARGET_POWERPC64)
2846 extra = 4;
2847 else if (offset & 3)
2848 return false;
2849 break;
2850
2851 case TFmode:
2852 case TImode:
2853 if (mode == TFmode || !TARGET_POWERPC64)
2854 extra = 12;
2855 else if (offset & 3)
2856 return false;
2857 else
2858 extra = 8;
2859 break;
2860
2861 default:
2862 break;
2863 }
2864
2865 offset += 0x8000;
2866 return (offset < 0x10000) && (offset + extra < 0x10000);
2867 }
2868
2869 static bool
legitimate_indexed_address_p(rtx x,int strict)2870 legitimate_indexed_address_p (rtx x, int strict)
2871 {
2872 rtx op0, op1;
2873
2874 if (GET_CODE (x) != PLUS)
2875 return false;
2876
2877 op0 = XEXP (x, 0);
2878 op1 = XEXP (x, 1);
2879
2880 /* Recognize the rtl generated by reload which we know will later be
2881 replaced with proper base and index regs. */
2882 if (!strict
2883 && reload_in_progress
2884 && (REG_P (op0) || GET_CODE (op0) == PLUS)
2885 && REG_P (op1))
2886 return true;
2887
2888 return (REG_P (op0) && REG_P (op1)
2889 && ((INT_REG_OK_FOR_BASE_P (op0, strict)
2890 && INT_REG_OK_FOR_INDEX_P (op1, strict))
2891 || (INT_REG_OK_FOR_BASE_P (op1, strict)
2892 && INT_REG_OK_FOR_INDEX_P (op0, strict))));
2893 }
2894
2895 inline bool
legitimate_indirect_address_p(rtx x,int strict)2896 legitimate_indirect_address_p (rtx x, int strict)
2897 {
2898 return GET_CODE (x) == REG && INT_REG_OK_FOR_BASE_P (x, strict);
2899 }
2900
2901 bool
macho_lo_sum_memory_operand(rtx x,enum machine_mode mode)2902 macho_lo_sum_memory_operand (rtx x, enum machine_mode mode)
2903 {
2904 if (!TARGET_MACHO || !flag_pic
2905 || mode != SImode || GET_CODE (x) != MEM)
2906 return false;
2907 x = XEXP (x, 0);
2908
2909 if (GET_CODE (x) != LO_SUM)
2910 return false;
2911 if (GET_CODE (XEXP (x, 0)) != REG)
2912 return false;
2913 if (!INT_REG_OK_FOR_BASE_P (XEXP (x, 0), 0))
2914 return false;
2915 x = XEXP (x, 1);
2916
2917 return CONSTANT_P (x);
2918 }
2919
2920 static bool
legitimate_lo_sum_address_p(enum machine_mode mode,rtx x,int strict)2921 legitimate_lo_sum_address_p (enum machine_mode mode, rtx x, int strict)
2922 {
2923 if (GET_CODE (x) != LO_SUM)
2924 return false;
2925 if (GET_CODE (XEXP (x, 0)) != REG)
2926 return false;
2927 if (!INT_REG_OK_FOR_BASE_P (XEXP (x, 0), strict))
2928 return false;
2929 /* Restrict addressing for DI because of our SUBREG hackery. */
2930 if (TARGET_E500_DOUBLE && (mode == DFmode || mode == DImode))
2931 return false;
2932 x = XEXP (x, 1);
2933
2934 if (TARGET_ELF || TARGET_MACHO)
2935 {
2936 if (DEFAULT_ABI != ABI_AIX && DEFAULT_ABI != ABI_DARWIN && flag_pic)
2937 return false;
2938 if (TARGET_TOC)
2939 return false;
2940 if (GET_MODE_NUNITS (mode) != 1)
2941 return false;
2942 if (GET_MODE_BITSIZE (mode) > 64
2943 || (GET_MODE_BITSIZE (mode) > 32 && !TARGET_POWERPC64
2944 && !(TARGET_HARD_FLOAT && TARGET_FPRS && mode == DFmode)))
2945 return false;
2946
2947 return CONSTANT_P (x);
2948 }
2949
2950 return false;
2951 }
2952
2953
2954 /* Try machine-dependent ways of modifying an illegitimate address
2955 to be legitimate. If we find one, return the new, valid address.
2956 This is used from only one place: `memory_address' in explow.c.
2957
2958 OLDX is the address as it was before break_out_memory_refs was
2959 called. In some cases it is useful to look at this to decide what
2960 needs to be done.
2961
2962 MODE is passed so that this function can use GO_IF_LEGITIMATE_ADDRESS.
2963
2964 It is always safe for this function to do nothing. It exists to
2965 recognize opportunities to optimize the output.
2966
2967 On RS/6000, first check for the sum of a register with a constant
2968 integer that is out of range. If so, generate code to add the
2969 constant with the low-order 16 bits masked to the register and force
2970 this result into another register (this can be done with `cau').
2971 Then generate an address of REG+(CONST&0xffff), allowing for the
2972 possibility of bit 16 being a one.
2973
2974 Then check for the sum of a register and something not constant, try to
2975 load the other things into a register and return the sum. */
2976
2977 rtx
rs6000_legitimize_address(rtx x,rtx oldx ATTRIBUTE_UNUSED,enum machine_mode mode)2978 rs6000_legitimize_address (rtx x, rtx oldx ATTRIBUTE_UNUSED,
2979 enum machine_mode mode)
2980 {
2981 if (GET_CODE (x) == SYMBOL_REF)
2982 {
2983 enum tls_model model = SYMBOL_REF_TLS_MODEL (x);
2984 if (model != 0)
2985 return rs6000_legitimize_tls_address (x, model);
2986 }
2987
2988 if (GET_CODE (x) == PLUS
2989 && GET_CODE (XEXP (x, 0)) == REG
2990 && GET_CODE (XEXP (x, 1)) == CONST_INT
2991 && (unsigned HOST_WIDE_INT) (INTVAL (XEXP (x, 1)) + 0x8000) >= 0x10000)
2992 {
2993 HOST_WIDE_INT high_int, low_int;
2994 rtx sum;
2995 low_int = ((INTVAL (XEXP (x, 1)) & 0xffff) ^ 0x8000) - 0x8000;
2996 high_int = INTVAL (XEXP (x, 1)) - low_int;
2997 sum = force_operand (gen_rtx_PLUS (Pmode, XEXP (x, 0),
2998 GEN_INT (high_int)), 0);
2999 return gen_rtx_PLUS (Pmode, sum, GEN_INT (low_int));
3000 }
3001 else if (GET_CODE (x) == PLUS
3002 && GET_CODE (XEXP (x, 0)) == REG
3003 && GET_CODE (XEXP (x, 1)) != CONST_INT
3004 && GET_MODE_NUNITS (mode) == 1
3005 && ((TARGET_HARD_FLOAT && TARGET_FPRS)
3006 || TARGET_POWERPC64
3007 || (((mode != DImode && mode != DFmode) || TARGET_E500_DOUBLE)
3008 && mode != TFmode))
3009 && (TARGET_POWERPC64 || mode != DImode)
3010 && mode != TImode)
3011 {
3012 return gen_rtx_PLUS (Pmode, XEXP (x, 0),
3013 force_reg (Pmode, force_operand (XEXP (x, 1), 0)));
3014 }
3015 else if (ALTIVEC_VECTOR_MODE (mode))
3016 {
3017 rtx reg;
3018
3019 /* Make sure both operands are registers. */
3020 if (GET_CODE (x) == PLUS)
3021 return gen_rtx_PLUS (Pmode, force_reg (Pmode, XEXP (x, 0)),
3022 force_reg (Pmode, XEXP (x, 1)));
3023
3024 reg = force_reg (Pmode, x);
3025 return reg;
3026 }
3027 else if (SPE_VECTOR_MODE (mode)
3028 || (TARGET_E500_DOUBLE && (mode == DFmode
3029 || mode == DImode)))
3030 {
3031 if (mode == DImode)
3032 return NULL_RTX;
3033 /* We accept [reg + reg] and [reg + OFFSET]. */
3034
3035 if (GET_CODE (x) == PLUS)
3036 {
3037 rtx op1 = XEXP (x, 0);
3038 rtx op2 = XEXP (x, 1);
3039
3040 op1 = force_reg (Pmode, op1);
3041
3042 if (GET_CODE (op2) != REG
3043 && (GET_CODE (op2) != CONST_INT
3044 || !SPE_CONST_OFFSET_OK (INTVAL (op2))))
3045 op2 = force_reg (Pmode, op2);
3046
3047 return gen_rtx_PLUS (Pmode, op1, op2);
3048 }
3049
3050 return force_reg (Pmode, x);
3051 }
3052 else if (TARGET_ELF
3053 && TARGET_32BIT
3054 && TARGET_NO_TOC
3055 && ! flag_pic
3056 && GET_CODE (x) != CONST_INT
3057 && GET_CODE (x) != CONST_DOUBLE
3058 && CONSTANT_P (x)
3059 && GET_MODE_NUNITS (mode) == 1
3060 && (GET_MODE_BITSIZE (mode) <= 32
3061 || ((TARGET_HARD_FLOAT && TARGET_FPRS) && mode == DFmode)))
3062 {
3063 rtx reg = gen_reg_rtx (Pmode);
3064 emit_insn (gen_elf_high (reg, x));
3065 return gen_rtx_LO_SUM (Pmode, reg, x);
3066 }
3067 else if (TARGET_MACHO && TARGET_32BIT && TARGET_NO_TOC
3068 && ! flag_pic
3069 #if TARGET_MACHO
3070 && ! MACHO_DYNAMIC_NO_PIC_P
3071 #endif
3072 && GET_CODE (x) != CONST_INT
3073 && GET_CODE (x) != CONST_DOUBLE
3074 && CONSTANT_P (x)
3075 && ((TARGET_HARD_FLOAT && TARGET_FPRS) || mode != DFmode)
3076 && mode != DImode
3077 && mode != TImode)
3078 {
3079 rtx reg = gen_reg_rtx (Pmode);
3080 emit_insn (gen_macho_high (reg, x));
3081 return gen_rtx_LO_SUM (Pmode, reg, x);
3082 }
3083 else if (TARGET_TOC
3084 && constant_pool_expr_p (x)
3085 && ASM_OUTPUT_SPECIAL_POOL_ENTRY_P (get_pool_constant (x), Pmode))
3086 {
3087 return create_TOC_reference (x);
3088 }
3089 else
3090 return NULL_RTX;
3091 }
3092
3093 /* This is called from dwarf2out.c via TARGET_ASM_OUTPUT_DWARF_DTPREL.
3094 We need to emit DTP-relative relocations. */
3095
3096 static void
rs6000_output_dwarf_dtprel(FILE * file,int size,rtx x)3097 rs6000_output_dwarf_dtprel (FILE *file, int size, rtx x)
3098 {
3099 switch (size)
3100 {
3101 case 4:
3102 fputs ("\t.long\t", file);
3103 break;
3104 case 8:
3105 fputs (DOUBLE_INT_ASM_OP, file);
3106 break;
3107 default:
3108 gcc_unreachable ();
3109 }
3110 output_addr_const (file, x);
3111 fputs ("@dtprel+0x8000", file);
3112 }
3113
3114 /* Construct the SYMBOL_REF for the tls_get_addr function. */
3115
3116 static GTY(()) rtx rs6000_tls_symbol;
3117 static rtx
rs6000_tls_get_addr(void)3118 rs6000_tls_get_addr (void)
3119 {
3120 if (!rs6000_tls_symbol)
3121 rs6000_tls_symbol = init_one_libfunc ("__tls_get_addr");
3122
3123 return rs6000_tls_symbol;
3124 }
3125
3126 /* Construct the SYMBOL_REF for TLS GOT references. */
3127
3128 static GTY(()) rtx rs6000_got_symbol;
3129 static rtx
rs6000_got_sym(void)3130 rs6000_got_sym (void)
3131 {
3132 if (!rs6000_got_symbol)
3133 {
3134 rs6000_got_symbol = gen_rtx_SYMBOL_REF (Pmode, "_GLOBAL_OFFSET_TABLE_");
3135 SYMBOL_REF_FLAGS (rs6000_got_symbol) |= SYMBOL_FLAG_LOCAL;
3136 SYMBOL_REF_FLAGS (rs6000_got_symbol) |= SYMBOL_FLAG_EXTERNAL;
3137 }
3138
3139 return rs6000_got_symbol;
3140 }
3141
3142 /* ADDR contains a thread-local SYMBOL_REF. Generate code to compute
3143 this (thread-local) address. */
3144
3145 static rtx
rs6000_legitimize_tls_address(rtx addr,enum tls_model model)3146 rs6000_legitimize_tls_address (rtx addr, enum tls_model model)
3147 {
3148 rtx dest, insn;
3149
3150 dest = gen_reg_rtx (Pmode);
3151 if (model == TLS_MODEL_LOCAL_EXEC && rs6000_tls_size == 16)
3152 {
3153 rtx tlsreg;
3154
3155 if (TARGET_64BIT)
3156 {
3157 tlsreg = gen_rtx_REG (Pmode, 13);
3158 insn = gen_tls_tprel_64 (dest, tlsreg, addr);
3159 }
3160 else
3161 {
3162 tlsreg = gen_rtx_REG (Pmode, 2);
3163 insn = gen_tls_tprel_32 (dest, tlsreg, addr);
3164 }
3165 emit_insn (insn);
3166 }
3167 else if (model == TLS_MODEL_LOCAL_EXEC && rs6000_tls_size == 32)
3168 {
3169 rtx tlsreg, tmp;
3170
3171 tmp = gen_reg_rtx (Pmode);
3172 if (TARGET_64BIT)
3173 {
3174 tlsreg = gen_rtx_REG (Pmode, 13);
3175 insn = gen_tls_tprel_ha_64 (tmp, tlsreg, addr);
3176 }
3177 else
3178 {
3179 tlsreg = gen_rtx_REG (Pmode, 2);
3180 insn = gen_tls_tprel_ha_32 (tmp, tlsreg, addr);
3181 }
3182 emit_insn (insn);
3183 if (TARGET_64BIT)
3184 insn = gen_tls_tprel_lo_64 (dest, tmp, addr);
3185 else
3186 insn = gen_tls_tprel_lo_32 (dest, tmp, addr);
3187 emit_insn (insn);
3188 }
3189 else
3190 {
3191 rtx r3, got, tga, tmp1, tmp2, eqv;
3192
3193 /* We currently use relocations like @got@tlsgd for tls, which
3194 means the linker will handle allocation of tls entries, placing
3195 them in the .got section. So use a pointer to the .got section,
3196 not one to secondary TOC sections used by 64-bit -mminimal-toc,
3197 or to secondary GOT sections used by 32-bit -fPIC. */
3198 if (TARGET_64BIT)
3199 got = gen_rtx_REG (Pmode, 2);
3200 else
3201 {
3202 if (flag_pic == 1)
3203 got = gen_rtx_REG (Pmode, RS6000_PIC_OFFSET_TABLE_REGNUM);
3204 else
3205 {
3206 rtx gsym = rs6000_got_sym ();
3207 got = gen_reg_rtx (Pmode);
3208 if (flag_pic == 0)
3209 rs6000_emit_move (got, gsym, Pmode);
3210 else
3211 {
3212 rtx tempLR, tmp3, mem;
3213 rtx first, last;
3214
3215 tempLR = gen_reg_rtx (Pmode);
3216 tmp1 = gen_reg_rtx (Pmode);
3217 tmp2 = gen_reg_rtx (Pmode);
3218 tmp3 = gen_reg_rtx (Pmode);
3219 mem = gen_const_mem (Pmode, tmp1);
3220
3221 first = emit_insn (gen_load_toc_v4_PIC_1b (tempLR, gsym));
3222 emit_move_insn (tmp1, tempLR);
3223 emit_move_insn (tmp2, mem);
3224 emit_insn (gen_addsi3 (tmp3, tmp1, tmp2));
3225 last = emit_move_insn (got, tmp3);
3226 REG_NOTES (last) = gen_rtx_EXPR_LIST (REG_EQUAL, gsym,
3227 REG_NOTES (last));
3228 REG_NOTES (first) = gen_rtx_INSN_LIST (REG_LIBCALL, last,
3229 REG_NOTES (first));
3230 REG_NOTES (last) = gen_rtx_INSN_LIST (REG_RETVAL, first,
3231 REG_NOTES (last));
3232 }
3233 }
3234 }
3235
3236 if (model == TLS_MODEL_GLOBAL_DYNAMIC)
3237 {
3238 r3 = gen_rtx_REG (Pmode, 3);
3239 if (TARGET_64BIT)
3240 insn = gen_tls_gd_64 (r3, got, addr);
3241 else
3242 insn = gen_tls_gd_32 (r3, got, addr);
3243 start_sequence ();
3244 emit_insn (insn);
3245 tga = gen_rtx_MEM (Pmode, rs6000_tls_get_addr ());
3246 insn = gen_call_value (r3, tga, const0_rtx, const0_rtx);
3247 insn = emit_call_insn (insn);
3248 CONST_OR_PURE_CALL_P (insn) = 1;
3249 use_reg (&CALL_INSN_FUNCTION_USAGE (insn), r3);
3250 insn = get_insns ();
3251 end_sequence ();
3252 emit_libcall_block (insn, dest, r3, addr);
3253 }
3254 else if (model == TLS_MODEL_LOCAL_DYNAMIC)
3255 {
3256 r3 = gen_rtx_REG (Pmode, 3);
3257 if (TARGET_64BIT)
3258 insn = gen_tls_ld_64 (r3, got);
3259 else
3260 insn = gen_tls_ld_32 (r3, got);
3261 start_sequence ();
3262 emit_insn (insn);
3263 tga = gen_rtx_MEM (Pmode, rs6000_tls_get_addr ());
3264 insn = gen_call_value (r3, tga, const0_rtx, const0_rtx);
3265 insn = emit_call_insn (insn);
3266 CONST_OR_PURE_CALL_P (insn) = 1;
3267 use_reg (&CALL_INSN_FUNCTION_USAGE (insn), r3);
3268 insn = get_insns ();
3269 end_sequence ();
3270 tmp1 = gen_reg_rtx (Pmode);
3271 eqv = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, const0_rtx),
3272 UNSPEC_TLSLD);
3273 emit_libcall_block (insn, tmp1, r3, eqv);
3274 if (rs6000_tls_size == 16)
3275 {
3276 if (TARGET_64BIT)
3277 insn = gen_tls_dtprel_64 (dest, tmp1, addr);
3278 else
3279 insn = gen_tls_dtprel_32 (dest, tmp1, addr);
3280 }
3281 else if (rs6000_tls_size == 32)
3282 {
3283 tmp2 = gen_reg_rtx (Pmode);
3284 if (TARGET_64BIT)
3285 insn = gen_tls_dtprel_ha_64 (tmp2, tmp1, addr);
3286 else
3287 insn = gen_tls_dtprel_ha_32 (tmp2, tmp1, addr);
3288 emit_insn (insn);
3289 if (TARGET_64BIT)
3290 insn = gen_tls_dtprel_lo_64 (dest, tmp2, addr);
3291 else
3292 insn = gen_tls_dtprel_lo_32 (dest, tmp2, addr);
3293 }
3294 else
3295 {
3296 tmp2 = gen_reg_rtx (Pmode);
3297 if (TARGET_64BIT)
3298 insn = gen_tls_got_dtprel_64 (tmp2, got, addr);
3299 else
3300 insn = gen_tls_got_dtprel_32 (tmp2, got, addr);
3301 emit_insn (insn);
3302 insn = gen_rtx_SET (Pmode, dest,
3303 gen_rtx_PLUS (Pmode, tmp2, tmp1));
3304 }
3305 emit_insn (insn);
3306 }
3307 else
3308 {
3309 /* IE, or 64 bit offset LE. */
3310 tmp2 = gen_reg_rtx (Pmode);
3311 if (TARGET_64BIT)
3312 insn = gen_tls_got_tprel_64 (tmp2, got, addr);
3313 else
3314 insn = gen_tls_got_tprel_32 (tmp2, got, addr);
3315 emit_insn (insn);
3316 if (TARGET_64BIT)
3317 insn = gen_tls_tls_64 (dest, tmp2, addr);
3318 else
3319 insn = gen_tls_tls_32 (dest, tmp2, addr);
3320 emit_insn (insn);
3321 }
3322 }
3323
3324 return dest;
3325 }
3326
3327 /* Return 1 if X contains a thread-local symbol. */
3328
3329 bool
rs6000_tls_referenced_p(rtx x)3330 rs6000_tls_referenced_p (rtx x)
3331 {
3332 if (! TARGET_HAVE_TLS)
3333 return false;
3334
3335 return for_each_rtx (&x, &rs6000_tls_symbol_ref_1, 0);
3336 }
3337
3338 /* Return 1 if *X is a thread-local symbol. This is the same as
3339 rs6000_tls_symbol_ref except for the type of the unused argument. */
3340
3341 static int
rs6000_tls_symbol_ref_1(rtx * x,void * data ATTRIBUTE_UNUSED)3342 rs6000_tls_symbol_ref_1 (rtx *x, void *data ATTRIBUTE_UNUSED)
3343 {
3344 return RS6000_SYMBOL_REF_TLS_P (*x);
3345 }
3346
3347 /* The convention appears to be to define this wherever it is used.
3348 With legitimize_reload_address now defined here, REG_MODE_OK_FOR_BASE_P
3349 is now used here. */
3350 #ifndef REG_MODE_OK_FOR_BASE_P
3351 #define REG_MODE_OK_FOR_BASE_P(REGNO, MODE) REG_OK_FOR_BASE_P (REGNO)
3352 #endif
3353
3354 /* Our implementation of LEGITIMIZE_RELOAD_ADDRESS. Returns a value to
3355 replace the input X, or the original X if no replacement is called for.
3356 The output parameter *WIN is 1 if the calling macro should goto WIN,
3357 0 if it should not.
3358
3359 For RS/6000, we wish to handle large displacements off a base
3360 register by splitting the addend across an addiu/addis and the mem insn.
3361 This cuts number of extra insns needed from 3 to 1.
3362
3363 On Darwin, we use this to generate code for floating point constants.
3364 A movsf_low is generated so we wind up with 2 instructions rather than 3.
3365 The Darwin code is inside #if TARGET_MACHO because only then is
3366 machopic_function_base_name() defined. */
3367 rtx
rs6000_legitimize_reload_address(rtx x,enum machine_mode mode,int opnum,int type,int ind_levels ATTRIBUTE_UNUSED,int * win)3368 rs6000_legitimize_reload_address (rtx x, enum machine_mode mode,
3369 int opnum, int type,
3370 int ind_levels ATTRIBUTE_UNUSED, int *win)
3371 {
3372 /* We must recognize output that we have already generated ourselves. */
3373 if (GET_CODE (x) == PLUS
3374 && GET_CODE (XEXP (x, 0)) == PLUS
3375 && GET_CODE (XEXP (XEXP (x, 0), 0)) == REG
3376 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
3377 && GET_CODE (XEXP (x, 1)) == CONST_INT)
3378 {
3379 push_reload (XEXP (x, 0), NULL_RTX, &XEXP (x, 0), NULL,
3380 BASE_REG_CLASS, GET_MODE (x), VOIDmode, 0, 0,
3381 opnum, (enum reload_type)type);
3382 *win = 1;
3383 return x;
3384 }
3385
3386 #if TARGET_MACHO
3387 if (DEFAULT_ABI == ABI_DARWIN && flag_pic
3388 && GET_CODE (x) == LO_SUM
3389 && GET_CODE (XEXP (x, 0)) == PLUS
3390 && XEXP (XEXP (x, 0), 0) == pic_offset_table_rtx
3391 && GET_CODE (XEXP (XEXP (x, 0), 1)) == HIGH
3392 && GET_CODE (XEXP (XEXP (XEXP (x, 0), 1), 0)) == CONST
3393 && XEXP (XEXP (XEXP (x, 0), 1), 0) == XEXP (x, 1)
3394 && GET_CODE (XEXP (XEXP (x, 1), 0)) == MINUS
3395 && GET_CODE (XEXP (XEXP (XEXP (x, 1), 0), 0)) == SYMBOL_REF
3396 && GET_CODE (XEXP (XEXP (XEXP (x, 1), 0), 1)) == SYMBOL_REF)
3397 {
3398 /* Result of previous invocation of this function on Darwin
3399 floating point constant. */
3400 push_reload (XEXP (x, 0), NULL_RTX, &XEXP (x, 0), NULL,
3401 BASE_REG_CLASS, Pmode, VOIDmode, 0, 0,
3402 opnum, (enum reload_type)type);
3403 *win = 1;
3404 return x;
3405 }
3406 #endif
3407
3408 /* Force ld/std non-word aligned offset into base register by wrapping
3409 in offset 0. */
3410 if (GET_CODE (x) == PLUS
3411 && GET_CODE (XEXP (x, 0)) == REG
3412 && REGNO (XEXP (x, 0)) < 32
3413 && REG_MODE_OK_FOR_BASE_P (XEXP (x, 0), mode)
3414 && GET_CODE (XEXP (x, 1)) == CONST_INT
3415 && (INTVAL (XEXP (x, 1)) & 3) != 0
3416 && !ALTIVEC_VECTOR_MODE (mode)
3417 && GET_MODE_SIZE (mode) >= UNITS_PER_WORD
3418 && TARGET_POWERPC64)
3419 {
3420 x = gen_rtx_PLUS (GET_MODE (x), x, GEN_INT (0));
3421 push_reload (XEXP (x, 0), NULL_RTX, &XEXP (x, 0), NULL,
3422 BASE_REG_CLASS, GET_MODE (x), VOIDmode, 0, 0,
3423 opnum, (enum reload_type) type);
3424 *win = 1;
3425 return x;
3426 }
3427
3428 if (GET_CODE (x) == PLUS
3429 && GET_CODE (XEXP (x, 0)) == REG
3430 && REGNO (XEXP (x, 0)) < FIRST_PSEUDO_REGISTER
3431 && REG_MODE_OK_FOR_BASE_P (XEXP (x, 0), mode)
3432 && GET_CODE (XEXP (x, 1)) == CONST_INT
3433 && !SPE_VECTOR_MODE (mode)
3434 && !(TARGET_E500_DOUBLE && (mode == DFmode
3435 || mode == DImode))
3436 && !ALTIVEC_VECTOR_MODE (mode))
3437 {
3438 HOST_WIDE_INT val = INTVAL (XEXP (x, 1));
3439 HOST_WIDE_INT low = ((val & 0xffff) ^ 0x8000) - 0x8000;
3440 HOST_WIDE_INT high
3441 = (((val - low) & 0xffffffff) ^ 0x80000000) - 0x80000000;
3442
3443 /* Check for 32-bit overflow. */
3444 if (high + low != val)
3445 {
3446 *win = 0;
3447 return x;
3448 }
3449
3450 /* Reload the high part into a base reg; leave the low part
3451 in the mem directly. */
3452
3453 x = gen_rtx_PLUS (GET_MODE (x),
3454 gen_rtx_PLUS (GET_MODE (x), XEXP (x, 0),
3455 GEN_INT (high)),
3456 GEN_INT (low));
3457
3458 push_reload (XEXP (x, 0), NULL_RTX, &XEXP (x, 0), NULL,
3459 BASE_REG_CLASS, GET_MODE (x), VOIDmode, 0, 0,
3460 opnum, (enum reload_type)type);
3461 *win = 1;
3462 return x;
3463 }
3464
3465 if (GET_CODE (x) == SYMBOL_REF
3466 && !ALTIVEC_VECTOR_MODE (mode)
3467 && !SPE_VECTOR_MODE (mode)
3468 #if TARGET_MACHO
3469 && DEFAULT_ABI == ABI_DARWIN
3470 && (flag_pic || MACHO_DYNAMIC_NO_PIC_P)
3471 #else
3472 && DEFAULT_ABI == ABI_V4
3473 && !flag_pic
3474 #endif
3475 /* Don't do this for TFmode, since the result isn't offsettable.
3476 The same goes for DImode without 64-bit gprs and DFmode
3477 without fprs. */
3478 && mode != TFmode
3479 && (mode != DImode || TARGET_POWERPC64)
3480 && (mode != DFmode || TARGET_POWERPC64
3481 || (TARGET_FPRS && TARGET_HARD_FLOAT)))
3482 {
3483 #if TARGET_MACHO
3484 if (flag_pic)
3485 {
3486 rtx offset = gen_rtx_CONST (Pmode,
3487 gen_rtx_MINUS (Pmode, x,
3488 machopic_function_base_sym ()));
3489 x = gen_rtx_LO_SUM (GET_MODE (x),
3490 gen_rtx_PLUS (Pmode, pic_offset_table_rtx,
3491 gen_rtx_HIGH (Pmode, offset)), offset);
3492 }
3493 else
3494 #endif
3495 x = gen_rtx_LO_SUM (GET_MODE (x),
3496 gen_rtx_HIGH (Pmode, x), x);
3497
3498 push_reload (XEXP (x, 0), NULL_RTX, &XEXP (x, 0), NULL,
3499 BASE_REG_CLASS, Pmode, VOIDmode, 0, 0,
3500 opnum, (enum reload_type)type);
3501 *win = 1;
3502 return x;
3503 }
3504
3505 /* Reload an offset address wrapped by an AND that represents the
3506 masking of the lower bits. Strip the outer AND and let reload
3507 convert the offset address into an indirect address. */
3508 if (TARGET_ALTIVEC
3509 && ALTIVEC_VECTOR_MODE (mode)
3510 && GET_CODE (x) == AND
3511 && GET_CODE (XEXP (x, 0)) == PLUS
3512 && GET_CODE (XEXP (XEXP (x, 0), 0)) == REG
3513 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
3514 && GET_CODE (XEXP (x, 1)) == CONST_INT
3515 && INTVAL (XEXP (x, 1)) == -16)
3516 {
3517 x = XEXP (x, 0);
3518 *win = 1;
3519 return x;
3520 }
3521
3522 if (TARGET_TOC
3523 && constant_pool_expr_p (x)
3524 && ASM_OUTPUT_SPECIAL_POOL_ENTRY_P (get_pool_constant (x), mode))
3525 {
3526 x = create_TOC_reference (x);
3527 *win = 1;
3528 return x;
3529 }
3530 *win = 0;
3531 return x;
3532 }
3533
3534 /* GO_IF_LEGITIMATE_ADDRESS recognizes an RTL expression
3535 that is a valid memory address for an instruction.
3536 The MODE argument is the machine mode for the MEM expression
3537 that wants to use this address.
3538
3539 On the RS/6000, there are four valid address: a SYMBOL_REF that
3540 refers to a constant pool entry of an address (or the sum of it
3541 plus a constant), a short (16-bit signed) constant plus a register,
3542 the sum of two registers, or a register indirect, possibly with an
3543 auto-increment. For DFmode and DImode with a constant plus register,
3544 we must ensure that both words are addressable or PowerPC64 with offset
3545 word aligned.
3546
3547 For modes spanning multiple registers (DFmode in 32-bit GPRs,
3548 32-bit DImode, TImode, TFmode), indexed addressing cannot be used because
3549 adjacent memory cells are accessed by adding word-sized offsets
3550 during assembly output. */
3551 int
rs6000_legitimate_address(enum machine_mode mode,rtx x,int reg_ok_strict)3552 rs6000_legitimate_address (enum machine_mode mode, rtx x, int reg_ok_strict)
3553 {
3554 /* If this is an unaligned stvx/ldvx type address, discard the outer AND. */
3555 if (TARGET_ALTIVEC
3556 && ALTIVEC_VECTOR_MODE (mode)
3557 && GET_CODE (x) == AND
3558 && GET_CODE (XEXP (x, 1)) == CONST_INT
3559 && INTVAL (XEXP (x, 1)) == -16)
3560 x = XEXP (x, 0);
3561
3562 if (RS6000_SYMBOL_REF_TLS_P (x))
3563 return 0;
3564 if (legitimate_indirect_address_p (x, reg_ok_strict))
3565 return 1;
3566 if ((GET_CODE (x) == PRE_INC || GET_CODE (x) == PRE_DEC)
3567 && !ALTIVEC_VECTOR_MODE (mode)
3568 && !SPE_VECTOR_MODE (mode)
3569 && mode != TFmode
3570 /* Restrict addressing for DI because of our SUBREG hackery. */
3571 && !(TARGET_E500_DOUBLE && (mode == DFmode || mode == DImode))
3572 && TARGET_UPDATE
3573 && legitimate_indirect_address_p (XEXP (x, 0), reg_ok_strict))
3574 return 1;
3575 if (legitimate_small_data_p (mode, x))
3576 return 1;
3577 if (legitimate_constant_pool_address_p (x))
3578 return 1;
3579 /* If not REG_OK_STRICT (before reload) let pass any stack offset. */
3580 if (! reg_ok_strict
3581 && GET_CODE (x) == PLUS
3582 && GET_CODE (XEXP (x, 0)) == REG
3583 && (XEXP (x, 0) == virtual_stack_vars_rtx
3584 || XEXP (x, 0) == arg_pointer_rtx)
3585 && GET_CODE (XEXP (x, 1)) == CONST_INT)
3586 return 1;
3587 if (rs6000_legitimate_offset_address_p (mode, x, reg_ok_strict))
3588 return 1;
3589 if (mode != TImode
3590 && mode != TFmode
3591 && ((TARGET_HARD_FLOAT && TARGET_FPRS)
3592 || TARGET_POWERPC64
3593 || ((mode != DFmode || TARGET_E500_DOUBLE) && mode != TFmode))
3594 && (TARGET_POWERPC64 || mode != DImode)
3595 && legitimate_indexed_address_p (x, reg_ok_strict))
3596 return 1;
3597 if (legitimate_lo_sum_address_p (mode, x, reg_ok_strict))
3598 return 1;
3599 return 0;
3600 }
3601
3602 /* Go to LABEL if ADDR (a legitimate address expression)
3603 has an effect that depends on the machine mode it is used for.
3604
3605 On the RS/6000 this is true of all integral offsets (since AltiVec
3606 modes don't allow them) or is a pre-increment or decrement.
3607
3608 ??? Except that due to conceptual problems in offsettable_address_p
3609 we can't really report the problems of integral offsets. So leave
3610 this assuming that the adjustable offset must be valid for the
3611 sub-words of a TFmode operand, which is what we had before. */
3612
3613 bool
rs6000_mode_dependent_address(rtx addr)3614 rs6000_mode_dependent_address (rtx addr)
3615 {
3616 switch (GET_CODE (addr))
3617 {
3618 case PLUS:
3619 if (GET_CODE (XEXP (addr, 1)) == CONST_INT)
3620 {
3621 unsigned HOST_WIDE_INT val = INTVAL (XEXP (addr, 1));
3622 return val + 12 + 0x8000 >= 0x10000;
3623 }
3624 break;
3625
3626 case LO_SUM:
3627 return true;
3628
3629 case PRE_INC:
3630 case PRE_DEC:
3631 return TARGET_UPDATE;
3632
3633 default:
3634 break;
3635 }
3636
3637 return false;
3638 }
3639
3640 /* More elaborate version of recog's offsettable_memref_p predicate
3641 that works around the ??? note of rs6000_mode_dependent_address.
3642 In particular it accepts
3643
3644 (mem:DI (plus:SI (reg/f:SI 31 31) (const_int 32760 [0x7ff8])))
3645
3646 in 32-bit mode, that the recog predicate rejects. */
3647
3648 bool
rs6000_offsettable_memref_p(rtx op)3649 rs6000_offsettable_memref_p (rtx op)
3650 {
3651 if (!MEM_P (op))
3652 return false;
3653
3654 /* First mimic offsettable_memref_p. */
3655 if (offsettable_address_p (1, GET_MODE (op), XEXP (op, 0)))
3656 return true;
3657
3658 /* offsettable_address_p invokes rs6000_mode_dependent_address, but
3659 the latter predicate knows nothing about the mode of the memory
3660 reference and, therefore, assumes that it is the largest supported
3661 mode (TFmode). As a consequence, legitimate offsettable memory
3662 references are rejected. rs6000_legitimate_offset_address_p contains
3663 the correct logic for the PLUS case of rs6000_mode_dependent_address. */
3664 return rs6000_legitimate_offset_address_p (GET_MODE (op), XEXP (op, 0), 1);
3665 }
3666
3667 /* Return number of consecutive hard regs needed starting at reg REGNO
3668 to hold something of mode MODE.
3669 This is ordinarily the length in words of a value of mode MODE
3670 but can be less for certain modes in special long registers.
3671
3672 For the SPE, GPRs are 64 bits but only 32 bits are visible in
3673 scalar instructions. The upper 32 bits are only available to the
3674 SIMD instructions.
3675
3676 POWER and PowerPC GPRs hold 32 bits worth;
3677 PowerPC64 GPRs and FPRs point register holds 64 bits worth. */
3678
3679 int
rs6000_hard_regno_nregs(int regno,enum machine_mode mode)3680 rs6000_hard_regno_nregs (int regno, enum machine_mode mode)
3681 {
3682 if (FP_REGNO_P (regno))
3683 return (GET_MODE_SIZE (mode) + UNITS_PER_FP_WORD - 1) / UNITS_PER_FP_WORD;
3684
3685 if (SPE_SIMD_REGNO_P (regno) && TARGET_SPE && SPE_VECTOR_MODE (mode))
3686 return (GET_MODE_SIZE (mode) + UNITS_PER_SPE_WORD - 1) / UNITS_PER_SPE_WORD;
3687
3688 if (ALTIVEC_REGNO_P (regno))
3689 return
3690 (GET_MODE_SIZE (mode) + UNITS_PER_ALTIVEC_WORD - 1) / UNITS_PER_ALTIVEC_WORD;
3691
3692 /* The value returned for SCmode in the E500 double case is 2 for
3693 ABI compatibility; storing an SCmode value in a single register
3694 would require function_arg and rs6000_spe_function_arg to handle
3695 SCmode so as to pass the value correctly in a pair of
3696 registers. */
3697 if (TARGET_E500_DOUBLE && FLOAT_MODE_P (mode) && mode != SCmode)
3698 return (GET_MODE_SIZE (mode) + UNITS_PER_FP_WORD - 1) / UNITS_PER_FP_WORD;
3699
3700 return (GET_MODE_SIZE (mode) + UNITS_PER_WORD - 1) / UNITS_PER_WORD;
3701 }
3702
3703 /* Change register usage conditional on target flags. */
3704 void
rs6000_conditional_register_usage(void)3705 rs6000_conditional_register_usage (void)
3706 {
3707 int i;
3708
3709 /* Set MQ register fixed (already call_used) if not POWER
3710 architecture (RIOS1, RIOS2, RSC, and PPC601) so that it will not
3711 be allocated. */
3712 if (! TARGET_POWER)
3713 fixed_regs[64] = 1;
3714
3715 /* 64-bit AIX and Linux reserve GPR13 for thread-private data. */
3716 if (TARGET_64BIT)
3717 fixed_regs[13] = call_used_regs[13]
3718 = call_really_used_regs[13] = 1;
3719
3720 /* Conditionally disable FPRs. */
3721 if (TARGET_SOFT_FLOAT || !TARGET_FPRS)
3722 for (i = 32; i < 64; i++)
3723 fixed_regs[i] = call_used_regs[i]
3724 = call_really_used_regs[i] = 1;
3725
3726 /* The TOC register is not killed across calls in a way that is
3727 visible to the compiler. */
3728 if (DEFAULT_ABI == ABI_AIX)
3729 call_really_used_regs[2] = 0;
3730
3731 if (DEFAULT_ABI == ABI_V4
3732 && PIC_OFFSET_TABLE_REGNUM != INVALID_REGNUM
3733 && flag_pic == 2)
3734 fixed_regs[RS6000_PIC_OFFSET_TABLE_REGNUM] = 1;
3735
3736 if (DEFAULT_ABI == ABI_V4
3737 && PIC_OFFSET_TABLE_REGNUM != INVALID_REGNUM
3738 && flag_pic == 1)
3739 fixed_regs[RS6000_PIC_OFFSET_TABLE_REGNUM]
3740 = call_used_regs[RS6000_PIC_OFFSET_TABLE_REGNUM]
3741 = call_really_used_regs[RS6000_PIC_OFFSET_TABLE_REGNUM] = 1;
3742
3743 if (DEFAULT_ABI == ABI_DARWIN
3744 && PIC_OFFSET_TABLE_REGNUM != INVALID_REGNUM)
3745 fixed_regs[RS6000_PIC_OFFSET_TABLE_REGNUM]
3746 = call_used_regs[RS6000_PIC_OFFSET_TABLE_REGNUM]
3747 = call_really_used_regs[RS6000_PIC_OFFSET_TABLE_REGNUM] = 1;
3748
3749 if (TARGET_TOC && TARGET_MINIMAL_TOC)
3750 fixed_regs[RS6000_PIC_OFFSET_TABLE_REGNUM]
3751 = call_used_regs[RS6000_PIC_OFFSET_TABLE_REGNUM] = 1;
3752
3753 if (TARGET_ALTIVEC)
3754 global_regs[VSCR_REGNO] = 1;
3755
3756 if (TARGET_SPE)
3757 {
3758 global_regs[SPEFSCR_REGNO] = 1;
3759 fixed_regs[FIXED_SCRATCH]
3760 = call_used_regs[FIXED_SCRATCH]
3761 = call_really_used_regs[FIXED_SCRATCH] = 1;
3762 }
3763
3764 if (! TARGET_ALTIVEC)
3765 {
3766 for (i = FIRST_ALTIVEC_REGNO; i <= LAST_ALTIVEC_REGNO; ++i)
3767 fixed_regs[i] = call_used_regs[i] = call_really_used_regs[i] = 1;
3768 call_really_used_regs[VRSAVE_REGNO] = 1;
3769 }
3770
3771 if (TARGET_ALTIVEC_ABI)
3772 for (i = FIRST_ALTIVEC_REGNO; i < FIRST_ALTIVEC_REGNO + 20; ++i)
3773 call_used_regs[i] = call_really_used_regs[i] = 1;
3774 }
3775
3776 /* Try to output insns to set TARGET equal to the constant C if it can
3777 be done in less than N insns. Do all computations in MODE.
3778 Returns the place where the output has been placed if it can be
3779 done and the insns have been emitted. If it would take more than N
3780 insns, zero is returned and no insns and emitted. */
3781
3782 rtx
rs6000_emit_set_const(rtx dest,enum machine_mode mode,rtx source,int n ATTRIBUTE_UNUSED)3783 rs6000_emit_set_const (rtx dest, enum machine_mode mode,
3784 rtx source, int n ATTRIBUTE_UNUSED)
3785 {
3786 rtx result, insn, set;
3787 HOST_WIDE_INT c0, c1;
3788
3789 switch (mode)
3790 {
3791 case QImode:
3792 case HImode:
3793 if (dest == NULL)
3794 dest = gen_reg_rtx (mode);
3795 emit_insn (gen_rtx_SET (VOIDmode, dest, source));
3796 return dest;
3797
3798 case SImode:
3799 result = no_new_pseudos ? dest : gen_reg_rtx (SImode);
3800
3801 emit_insn (gen_rtx_SET (VOIDmode, result,
3802 GEN_INT (INTVAL (source)
3803 & (~ (HOST_WIDE_INT) 0xffff))));
3804 emit_insn (gen_rtx_SET (VOIDmode, dest,
3805 gen_rtx_IOR (SImode, result,
3806 GEN_INT (INTVAL (source) & 0xffff))));
3807 result = dest;
3808 break;
3809
3810 case DImode:
3811 switch (GET_CODE (source))
3812 {
3813 case CONST_INT:
3814 c0 = INTVAL (source);
3815 c1 = -(c0 < 0);
3816 break;
3817
3818 case CONST_DOUBLE:
3819 #if HOST_BITS_PER_WIDE_INT >= 64
3820 c0 = CONST_DOUBLE_LOW (source);
3821 c1 = -(c0 < 0);
3822 #else
3823 c0 = CONST_DOUBLE_LOW (source);
3824 c1 = CONST_DOUBLE_HIGH (source);
3825 #endif
3826 break;
3827
3828 default:
3829 gcc_unreachable ();
3830 }
3831
3832 result = rs6000_emit_set_long_const (dest, c0, c1);
3833 break;
3834
3835 default:
3836 gcc_unreachable ();
3837 }
3838
3839 insn = get_last_insn ();
3840 set = single_set (insn);
3841 if (! CONSTANT_P (SET_SRC (set)))
3842 set_unique_reg_note (insn, REG_EQUAL, source);
3843
3844 return result;
3845 }
3846
3847 /* Having failed to find a 3 insn sequence in rs6000_emit_set_const,
3848 fall back to a straight forward decomposition. We do this to avoid
3849 exponential run times encountered when looking for longer sequences
3850 with rs6000_emit_set_const. */
3851 static rtx
rs6000_emit_set_long_const(rtx dest,HOST_WIDE_INT c1,HOST_WIDE_INT c2)3852 rs6000_emit_set_long_const (rtx dest, HOST_WIDE_INT c1, HOST_WIDE_INT c2)
3853 {
3854 if (!TARGET_POWERPC64)
3855 {
3856 rtx operand1, operand2;
3857
3858 operand1 = operand_subword_force (dest, WORDS_BIG_ENDIAN == 0,
3859 DImode);
3860 operand2 = operand_subword_force (dest, WORDS_BIG_ENDIAN != 0,
3861 DImode);
3862 emit_move_insn (operand1, GEN_INT (c1));
3863 emit_move_insn (operand2, GEN_INT (c2));
3864 }
3865 else
3866 {
3867 HOST_WIDE_INT ud1, ud2, ud3, ud4;
3868
3869 ud1 = c1 & 0xffff;
3870 ud2 = (c1 & 0xffff0000) >> 16;
3871 #if HOST_BITS_PER_WIDE_INT >= 64
3872 c2 = c1 >> 32;
3873 #endif
3874 ud3 = c2 & 0xffff;
3875 ud4 = (c2 & 0xffff0000) >> 16;
3876
3877 if ((ud4 == 0xffff && ud3 == 0xffff && ud2 == 0xffff && (ud1 & 0x8000))
3878 || (ud4 == 0 && ud3 == 0 && ud2 == 0 && ! (ud1 & 0x8000)))
3879 {
3880 if (ud1 & 0x8000)
3881 emit_move_insn (dest, GEN_INT (((ud1 ^ 0x8000) - 0x8000)));
3882 else
3883 emit_move_insn (dest, GEN_INT (ud1));
3884 }
3885
3886 else if ((ud4 == 0xffff && ud3 == 0xffff && (ud2 & 0x8000))
3887 || (ud4 == 0 && ud3 == 0 && ! (ud2 & 0x8000)))
3888 {
3889 if (ud2 & 0x8000)
3890 emit_move_insn (dest, GEN_INT (((ud2 << 16) ^ 0x80000000)
3891 - 0x80000000));
3892 else
3893 emit_move_insn (dest, GEN_INT (ud2 << 16));
3894 if (ud1 != 0)
3895 emit_move_insn (dest, gen_rtx_IOR (DImode, dest, GEN_INT (ud1)));
3896 }
3897 else if ((ud4 == 0xffff && (ud3 & 0x8000))
3898 || (ud4 == 0 && ! (ud3 & 0x8000)))
3899 {
3900 if (ud3 & 0x8000)
3901 emit_move_insn (dest, GEN_INT (((ud3 << 16) ^ 0x80000000)
3902 - 0x80000000));
3903 else
3904 emit_move_insn (dest, GEN_INT (ud3 << 16));
3905
3906 if (ud2 != 0)
3907 emit_move_insn (dest, gen_rtx_IOR (DImode, dest, GEN_INT (ud2)));
3908 emit_move_insn (dest, gen_rtx_ASHIFT (DImode, dest, GEN_INT (16)));
3909 if (ud1 != 0)
3910 emit_move_insn (dest, gen_rtx_IOR (DImode, dest, GEN_INT (ud1)));
3911 }
3912 else
3913 {
3914 if (ud4 & 0x8000)
3915 emit_move_insn (dest, GEN_INT (((ud4 << 16) ^ 0x80000000)
3916 - 0x80000000));
3917 else
3918 emit_move_insn (dest, GEN_INT (ud4 << 16));
3919
3920 if (ud3 != 0)
3921 emit_move_insn (dest, gen_rtx_IOR (DImode, dest, GEN_INT (ud3)));
3922
3923 emit_move_insn (dest, gen_rtx_ASHIFT (DImode, dest, GEN_INT (32)));
3924 if (ud2 != 0)
3925 emit_move_insn (dest, gen_rtx_IOR (DImode, dest,
3926 GEN_INT (ud2 << 16)));
3927 if (ud1 != 0)
3928 emit_move_insn (dest, gen_rtx_IOR (DImode, dest, GEN_INT (ud1)));
3929 }
3930 }
3931 return dest;
3932 }
3933
3934 /* Helper for the following. Get rid of [r+r] memory refs
3935 in cases where it won't work (TImode, TFmode). */
3936
3937 static void
rs6000_eliminate_indexed_memrefs(rtx operands[2])3938 rs6000_eliminate_indexed_memrefs (rtx operands[2])
3939 {
3940 if (GET_CODE (operands[0]) == MEM
3941 && GET_CODE (XEXP (operands[0], 0)) != REG
3942 && ! legitimate_constant_pool_address_p (XEXP (operands[0], 0))
3943 && ! reload_in_progress)
3944 operands[0]
3945 = replace_equiv_address (operands[0],
3946 copy_addr_to_reg (XEXP (operands[0], 0)));
3947
3948 if (GET_CODE (operands[1]) == MEM
3949 && GET_CODE (XEXP (operands[1], 0)) != REG
3950 && ! legitimate_constant_pool_address_p (XEXP (operands[1], 0))
3951 && ! reload_in_progress)
3952 operands[1]
3953 = replace_equiv_address (operands[1],
3954 copy_addr_to_reg (XEXP (operands[1], 0)));
3955 }
3956
3957 /* Emit a move from SOURCE to DEST in mode MODE. */
3958 void
rs6000_emit_move(rtx dest,rtx source,enum machine_mode mode)3959 rs6000_emit_move (rtx dest, rtx source, enum machine_mode mode)
3960 {
3961 rtx operands[2];
3962 operands[0] = dest;
3963 operands[1] = source;
3964
3965 /* Sanity checks. Check that we get CONST_DOUBLE only when we should. */
3966 if (GET_CODE (operands[1]) == CONST_DOUBLE
3967 && ! FLOAT_MODE_P (mode)
3968 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT)
3969 {
3970 /* FIXME. This should never happen. */
3971 /* Since it seems that it does, do the safe thing and convert
3972 to a CONST_INT. */
3973 operands[1] = gen_int_mode (CONST_DOUBLE_LOW (operands[1]), mode);
3974 }
3975 gcc_assert (GET_CODE (operands[1]) != CONST_DOUBLE
3976 || FLOAT_MODE_P (mode)
3977 || ((CONST_DOUBLE_HIGH (operands[1]) != 0
3978 || CONST_DOUBLE_LOW (operands[1]) < 0)
3979 && (CONST_DOUBLE_HIGH (operands[1]) != -1
3980 || CONST_DOUBLE_LOW (operands[1]) >= 0)));
3981
3982 /* Check if GCC is setting up a block move that will end up using FP
3983 registers as temporaries. We must make sure this is acceptable. */
3984 if (GET_CODE (operands[0]) == MEM
3985 && GET_CODE (operands[1]) == MEM
3986 && mode == DImode
3987 && (SLOW_UNALIGNED_ACCESS (DImode, MEM_ALIGN (operands[0]))
3988 || SLOW_UNALIGNED_ACCESS (DImode, MEM_ALIGN (operands[1])))
3989 && ! (SLOW_UNALIGNED_ACCESS (SImode, (MEM_ALIGN (operands[0]) > 32
3990 ? 32 : MEM_ALIGN (operands[0])))
3991 || SLOW_UNALIGNED_ACCESS (SImode, (MEM_ALIGN (operands[1]) > 32
3992 ? 32
3993 : MEM_ALIGN (operands[1]))))
3994 && ! MEM_VOLATILE_P (operands [0])
3995 && ! MEM_VOLATILE_P (operands [1]))
3996 {
3997 emit_move_insn (adjust_address (operands[0], SImode, 0),
3998 adjust_address (operands[1], SImode, 0));
3999 emit_move_insn (adjust_address (operands[0], SImode, 4),
4000 adjust_address (operands[1], SImode, 4));
4001 return;
4002 }
4003
4004 if (!no_new_pseudos && GET_CODE (operands[0]) == MEM
4005 && !gpc_reg_operand (operands[1], mode))
4006 operands[1] = force_reg (mode, operands[1]);
4007
4008 if (mode == SFmode && ! TARGET_POWERPC
4009 && TARGET_HARD_FLOAT && TARGET_FPRS
4010 && GET_CODE (operands[0]) == MEM)
4011 {
4012 int regnum;
4013
4014 if (reload_in_progress || reload_completed)
4015 regnum = true_regnum (operands[1]);
4016 else if (GET_CODE (operands[1]) == REG)
4017 regnum = REGNO (operands[1]);
4018 else
4019 regnum = -1;
4020
4021 /* If operands[1] is a register, on POWER it may have
4022 double-precision data in it, so truncate it to single
4023 precision. */
4024 if (FP_REGNO_P (regnum) || regnum >= FIRST_PSEUDO_REGISTER)
4025 {
4026 rtx newreg;
4027 newreg = (no_new_pseudos ? operands[1] : gen_reg_rtx (mode));
4028 emit_insn (gen_aux_truncdfsf2 (newreg, operands[1]));
4029 operands[1] = newreg;
4030 }
4031 }
4032
4033 /* Recognize the case where operand[1] is a reference to thread-local
4034 data and load its address to a register. */
4035 if (rs6000_tls_referenced_p (operands[1]))
4036 {
4037 enum tls_model model;
4038 rtx tmp = operands[1];
4039 rtx addend = NULL;
4040
4041 if (GET_CODE (tmp) == CONST && GET_CODE (XEXP (tmp, 0)) == PLUS)
4042 {
4043 addend = XEXP (XEXP (tmp, 0), 1);
4044 tmp = XEXP (XEXP (tmp, 0), 0);
4045 }
4046
4047 gcc_assert (GET_CODE (tmp) == SYMBOL_REF);
4048 model = SYMBOL_REF_TLS_MODEL (tmp);
4049 gcc_assert (model != 0);
4050
4051 tmp = rs6000_legitimize_tls_address (tmp, model);
4052 if (addend)
4053 {
4054 tmp = gen_rtx_PLUS (mode, tmp, addend);
4055 tmp = force_operand (tmp, operands[0]);
4056 }
4057 operands[1] = tmp;
4058 }
4059
4060 /* Handle the case where reload calls us with an invalid address. */
4061 if (reload_in_progress && mode == Pmode
4062 && (! general_operand (operands[1], mode)
4063 || ! nonimmediate_operand (operands[0], mode)))
4064 goto emit_set;
4065
4066 /* 128-bit constant floating-point values on Darwin should really be
4067 loaded as two parts. */
4068 if (!TARGET_IEEEQUAD && TARGET_LONG_DOUBLE_128
4069 && mode == TFmode && GET_CODE (operands[1]) == CONST_DOUBLE)
4070 {
4071 /* DImode is used, not DFmode, because simplify_gen_subreg doesn't
4072 know how to get a DFmode SUBREG of a TFmode. */
4073 rs6000_emit_move (simplify_gen_subreg (DImode, operands[0], mode, 0),
4074 simplify_gen_subreg (DImode, operands[1], mode, 0),
4075 DImode);
4076 rs6000_emit_move (simplify_gen_subreg (DImode, operands[0], mode,
4077 GET_MODE_SIZE (DImode)),
4078 simplify_gen_subreg (DImode, operands[1], mode,
4079 GET_MODE_SIZE (DImode)),
4080 DImode);
4081 return;
4082 }
4083
4084 /* FIXME: In the long term, this switch statement should go away
4085 and be replaced by a sequence of tests based on things like
4086 mode == Pmode. */
4087 switch (mode)
4088 {
4089 case HImode:
4090 case QImode:
4091 if (CONSTANT_P (operands[1])
4092 && GET_CODE (operands[1]) != CONST_INT)
4093 operands[1] = force_const_mem (mode, operands[1]);
4094 break;
4095
4096 case TFmode:
4097 rs6000_eliminate_indexed_memrefs (operands);
4098 /* fall through */
4099
4100 case DFmode:
4101 case SFmode:
4102 if (CONSTANT_P (operands[1])
4103 && ! easy_fp_constant (operands[1], mode))
4104 operands[1] = force_const_mem (mode, operands[1]);
4105 break;
4106
4107 case V16QImode:
4108 case V8HImode:
4109 case V4SFmode:
4110 case V4SImode:
4111 case V4HImode:
4112 case V2SFmode:
4113 case V2SImode:
4114 case V1DImode:
4115 if (CONSTANT_P (operands[1])
4116 && !easy_vector_constant (operands[1], mode))
4117 operands[1] = force_const_mem (mode, operands[1]);
4118 break;
4119
4120 case SImode:
4121 case DImode:
4122 /* Use default pattern for address of ELF small data */
4123 if (TARGET_ELF
4124 && mode == Pmode
4125 && DEFAULT_ABI == ABI_V4
4126 && (GET_CODE (operands[1]) == SYMBOL_REF
4127 || GET_CODE (operands[1]) == CONST)
4128 && small_data_operand (operands[1], mode))
4129 {
4130 emit_insn (gen_rtx_SET (VOIDmode, operands[0], operands[1]));
4131 return;
4132 }
4133
4134 if (DEFAULT_ABI == ABI_V4
4135 && mode == Pmode && mode == SImode
4136 && flag_pic == 1 && got_operand (operands[1], mode))
4137 {
4138 emit_insn (gen_movsi_got (operands[0], operands[1]));
4139 return;
4140 }
4141
4142 if ((TARGET_ELF || DEFAULT_ABI == ABI_DARWIN)
4143 && TARGET_NO_TOC
4144 && ! flag_pic
4145 && mode == Pmode
4146 && CONSTANT_P (operands[1])
4147 && GET_CODE (operands[1]) != HIGH
4148 && GET_CODE (operands[1]) != CONST_INT)
4149 {
4150 rtx target = (no_new_pseudos ? operands[0] : gen_reg_rtx (mode));
4151
4152 /* If this is a function address on -mcall-aixdesc,
4153 convert it to the address of the descriptor. */
4154 if (DEFAULT_ABI == ABI_AIX
4155 && GET_CODE (operands[1]) == SYMBOL_REF
4156 && XSTR (operands[1], 0)[0] == '.')
4157 {
4158 const char *name = XSTR (operands[1], 0);
4159 rtx new_ref;
4160 while (*name == '.')
4161 name++;
4162 new_ref = gen_rtx_SYMBOL_REF (Pmode, name);
4163 CONSTANT_POOL_ADDRESS_P (new_ref)
4164 = CONSTANT_POOL_ADDRESS_P (operands[1]);
4165 SYMBOL_REF_FLAGS (new_ref) = SYMBOL_REF_FLAGS (operands[1]);
4166 SYMBOL_REF_USED (new_ref) = SYMBOL_REF_USED (operands[1]);
4167 SYMBOL_REF_DATA (new_ref) = SYMBOL_REF_DATA (operands[1]);
4168 operands[1] = new_ref;
4169 }
4170
4171 if (DEFAULT_ABI == ABI_DARWIN)
4172 {
4173 #if TARGET_MACHO
4174 if (MACHO_DYNAMIC_NO_PIC_P)
4175 {
4176 /* Take care of any required data indirection. */
4177 operands[1] = rs6000_machopic_legitimize_pic_address (
4178 operands[1], mode, operands[0]);
4179 if (operands[0] != operands[1])
4180 emit_insn (gen_rtx_SET (VOIDmode,
4181 operands[0], operands[1]));
4182 return;
4183 }
4184 #endif
4185 emit_insn (gen_macho_high (target, operands[1]));
4186 emit_insn (gen_macho_low (operands[0], target, operands[1]));
4187 return;
4188 }
4189
4190 emit_insn (gen_elf_high (target, operands[1]));
4191 emit_insn (gen_elf_low (operands[0], target, operands[1]));
4192 return;
4193 }
4194
4195 /* If this is a SYMBOL_REF that refers to a constant pool entry,
4196 and we have put it in the TOC, we just need to make a TOC-relative
4197 reference to it. */
4198 if (TARGET_TOC
4199 && GET_CODE (operands[1]) == SYMBOL_REF
4200 && constant_pool_expr_p (operands[1])
4201 && ASM_OUTPUT_SPECIAL_POOL_ENTRY_P (get_pool_constant (operands[1]),
4202 get_pool_mode (operands[1])))
4203 {
4204 operands[1] = create_TOC_reference (operands[1]);
4205 }
4206 else if (mode == Pmode
4207 && CONSTANT_P (operands[1])
4208 && ((GET_CODE (operands[1]) != CONST_INT
4209 && ! easy_fp_constant (operands[1], mode))
4210 || (GET_CODE (operands[1]) == CONST_INT
4211 && num_insns_constant (operands[1], mode) > 2)
4212 || (GET_CODE (operands[0]) == REG
4213 && FP_REGNO_P (REGNO (operands[0]))))
4214 && GET_CODE (operands[1]) != HIGH
4215 && ! legitimate_constant_pool_address_p (operands[1])
4216 && ! toc_relative_expr_p (operands[1]))
4217 {
4218 /* Emit a USE operation so that the constant isn't deleted if
4219 expensive optimizations are turned on because nobody
4220 references it. This should only be done for operands that
4221 contain SYMBOL_REFs with CONSTANT_POOL_ADDRESS_P set.
4222 This should not be done for operands that contain LABEL_REFs.
4223 For now, we just handle the obvious case. */
4224 if (GET_CODE (operands[1]) != LABEL_REF)
4225 emit_insn (gen_rtx_USE (VOIDmode, operands[1]));
4226
4227 #if TARGET_MACHO
4228 /* Darwin uses a special PIC legitimizer. */
4229 if (DEFAULT_ABI == ABI_DARWIN && MACHOPIC_INDIRECT)
4230 {
4231 operands[1] =
4232 rs6000_machopic_legitimize_pic_address (operands[1], mode,
4233 operands[0]);
4234 if (operands[0] != operands[1])
4235 emit_insn (gen_rtx_SET (VOIDmode, operands[0], operands[1]));
4236 return;
4237 }
4238 #endif
4239
4240 /* If we are to limit the number of things we put in the TOC and
4241 this is a symbol plus a constant we can add in one insn,
4242 just put the symbol in the TOC and add the constant. Don't do
4243 this if reload is in progress. */
4244 if (GET_CODE (operands[1]) == CONST
4245 && TARGET_NO_SUM_IN_TOC && ! reload_in_progress
4246 && GET_CODE (XEXP (operands[1], 0)) == PLUS
4247 && add_operand (XEXP (XEXP (operands[1], 0), 1), mode)
4248 && (GET_CODE (XEXP (XEXP (operands[1], 0), 0)) == LABEL_REF
4249 || GET_CODE (XEXP (XEXP (operands[1], 0), 0)) == SYMBOL_REF)
4250 && ! side_effects_p (operands[0]))
4251 {
4252 rtx sym =
4253 force_const_mem (mode, XEXP (XEXP (operands[1], 0), 0));
4254 rtx other = XEXP (XEXP (operands[1], 0), 1);
4255
4256 sym = force_reg (mode, sym);
4257 if (mode == SImode)
4258 emit_insn (gen_addsi3 (operands[0], sym, other));
4259 else
4260 emit_insn (gen_adddi3 (operands[0], sym, other));
4261 return;
4262 }
4263
4264 operands[1] = force_const_mem (mode, operands[1]);
4265
4266 if (TARGET_TOC
4267 && constant_pool_expr_p (XEXP (operands[1], 0))
4268 && ASM_OUTPUT_SPECIAL_POOL_ENTRY_P (
4269 get_pool_constant (XEXP (operands[1], 0)),
4270 get_pool_mode (XEXP (operands[1], 0))))
4271 {
4272 operands[1]
4273 = gen_const_mem (mode,
4274 create_TOC_reference (XEXP (operands[1], 0)));
4275 set_mem_alias_set (operands[1], get_TOC_alias_set ());
4276 }
4277 }
4278 break;
4279
4280 case TImode:
4281 rs6000_eliminate_indexed_memrefs (operands);
4282
4283 if (TARGET_POWER)
4284 {
4285 emit_insn (gen_rtx_PARALLEL (VOIDmode,
4286 gen_rtvec (2,
4287 gen_rtx_SET (VOIDmode,
4288 operands[0], operands[1]),
4289 gen_rtx_CLOBBER (VOIDmode,
4290 gen_rtx_SCRATCH (SImode)))));
4291 return;
4292 }
4293 break;
4294
4295 default:
4296 gcc_unreachable ();
4297 }
4298
4299 /* Above, we may have called force_const_mem which may have returned
4300 an invalid address. If we can, fix this up; otherwise, reload will
4301 have to deal with it. */
4302 if (GET_CODE (operands[1]) == MEM && ! reload_in_progress)
4303 operands[1] = validize_mem (operands[1]);
4304
4305 emit_set:
4306 emit_insn (gen_rtx_SET (VOIDmode, operands[0], operands[1]));
4307 }
4308
4309 /* Nonzero if we can use a floating-point register to pass this arg. */
4310 #define USE_FP_FOR_ARG_P(CUM,MODE,TYPE) \
4311 (SCALAR_FLOAT_MODE_P (MODE) \
4312 && !DECIMAL_FLOAT_MODE_P (MODE) \
4313 && (CUM)->fregno <= FP_ARG_MAX_REG \
4314 && TARGET_HARD_FLOAT && TARGET_FPRS)
4315
4316 /* Nonzero if we can use an AltiVec register to pass this arg. */
4317 #define USE_ALTIVEC_FOR_ARG_P(CUM,MODE,TYPE,NAMED) \
4318 (ALTIVEC_VECTOR_MODE (MODE) \
4319 && (CUM)->vregno <= ALTIVEC_ARG_MAX_REG \
4320 && TARGET_ALTIVEC_ABI \
4321 && (NAMED))
4322
4323 /* Return a nonzero value to say to return the function value in
4324 memory, just as large structures are always returned. TYPE will be
4325 the data type of the value, and FNTYPE will be the type of the
4326 function doing the returning, or @code{NULL} for libcalls.
4327
4328 The AIX ABI for the RS/6000 specifies that all structures are
4329 returned in memory. The Darwin ABI does the same. The SVR4 ABI
4330 specifies that structures <= 8 bytes are returned in r3/r4, but a
4331 draft put them in memory, and GCC used to implement the draft
4332 instead of the final standard. Therefore, aix_struct_return
4333 controls this instead of DEFAULT_ABI; V.4 targets needing backward
4334 compatibility can change DRAFT_V4_STRUCT_RET to override the
4335 default, and -m switches get the final word. See
4336 rs6000_override_options for more details.
4337
4338 The PPC32 SVR4 ABI uses IEEE double extended for long double, if 128-bit
4339 long double support is enabled. These values are returned in memory.
4340
4341 int_size_in_bytes returns -1 for variable size objects, which go in
4342 memory always. The cast to unsigned makes -1 > 8. */
4343
4344 static bool
rs6000_return_in_memory(tree type,tree fntype ATTRIBUTE_UNUSED)4345 rs6000_return_in_memory (tree type, tree fntype ATTRIBUTE_UNUSED)
4346 {
4347 /* In the darwin64 abi, try to use registers for larger structs
4348 if possible. */
4349 if (rs6000_darwin64_abi
4350 && TREE_CODE (type) == RECORD_TYPE
4351 && int_size_in_bytes (type) > 0)
4352 {
4353 CUMULATIVE_ARGS valcum;
4354 rtx valret;
4355
4356 valcum.words = 0;
4357 valcum.fregno = FP_ARG_MIN_REG;
4358 valcum.vregno = ALTIVEC_ARG_MIN_REG;
4359 /* Do a trial code generation as if this were going to be passed
4360 as an argument; if any part goes in memory, we return NULL. */
4361 valret = rs6000_darwin64_record_arg (&valcum, type, 1, true);
4362 if (valret)
4363 return false;
4364 /* Otherwise fall through to more conventional ABI rules. */
4365 }
4366
4367 if (AGGREGATE_TYPE_P (type)
4368 && (aix_struct_return
4369 || (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 8))
4370 return true;
4371
4372 /* Allow -maltivec -mabi=no-altivec without warning. Altivec vector
4373 modes only exist for GCC vector types if -maltivec. */
4374 if (TARGET_32BIT && !TARGET_ALTIVEC_ABI
4375 && ALTIVEC_VECTOR_MODE (TYPE_MODE (type)))
4376 return false;
4377
4378 /* Return synthetic vectors in memory. */
4379 if (TREE_CODE (type) == VECTOR_TYPE
4380 && int_size_in_bytes (type) > (TARGET_ALTIVEC_ABI ? 16 : 8))
4381 {
4382 static bool warned_for_return_big_vectors = false;
4383 if (!warned_for_return_big_vectors)
4384 {
4385 warning (0, "GCC vector returned by reference: "
4386 "non-standard ABI extension with no compatibility guarantee");
4387 warned_for_return_big_vectors = true;
4388 }
4389 return true;
4390 }
4391
4392 if (DEFAULT_ABI == ABI_V4 && TARGET_IEEEQUAD && TYPE_MODE (type) == TFmode)
4393 return true;
4394
4395 return false;
4396 }
4397
4398 /* Initialize a variable CUM of type CUMULATIVE_ARGS
4399 for a call to a function whose data type is FNTYPE.
4400 For a library call, FNTYPE is 0.
4401
4402 For incoming args we set the number of arguments in the prototype large
4403 so we never return a PARALLEL. */
4404
4405 void
init_cumulative_args(CUMULATIVE_ARGS * cum,tree fntype,rtx libname ATTRIBUTE_UNUSED,int incoming,int libcall,int n_named_args)4406 init_cumulative_args (CUMULATIVE_ARGS *cum, tree fntype,
4407 rtx libname ATTRIBUTE_UNUSED, int incoming,
4408 int libcall, int n_named_args)
4409 {
4410 static CUMULATIVE_ARGS zero_cumulative;
4411
4412 *cum = zero_cumulative;
4413 cum->words = 0;
4414 cum->fregno = FP_ARG_MIN_REG;
4415 cum->vregno = ALTIVEC_ARG_MIN_REG;
4416 cum->prototype = (fntype && TYPE_ARG_TYPES (fntype));
4417 cum->call_cookie = ((DEFAULT_ABI == ABI_V4 && libcall)
4418 ? CALL_LIBCALL : CALL_NORMAL);
4419 cum->sysv_gregno = GP_ARG_MIN_REG;
4420 cum->stdarg = fntype
4421 && (TYPE_ARG_TYPES (fntype) != 0
4422 && (TREE_VALUE (tree_last (TYPE_ARG_TYPES (fntype)))
4423 != void_type_node));
4424
4425 cum->nargs_prototype = 0;
4426 if (incoming || cum->prototype)
4427 cum->nargs_prototype = n_named_args;
4428
4429 /* Check for a longcall attribute. */
4430 if ((!fntype && rs6000_default_long_calls)
4431 || (fntype
4432 && lookup_attribute ("longcall", TYPE_ATTRIBUTES (fntype))
4433 && !lookup_attribute ("shortcall", TYPE_ATTRIBUTES (fntype))))
4434 cum->call_cookie |= CALL_LONG;
4435
4436 if (TARGET_DEBUG_ARG)
4437 {
4438 fprintf (stderr, "\ninit_cumulative_args:");
4439 if (fntype)
4440 {
4441 tree ret_type = TREE_TYPE (fntype);
4442 fprintf (stderr, " ret code = %s,",
4443 tree_code_name[ (int)TREE_CODE (ret_type) ]);
4444 }
4445
4446 if (cum->call_cookie & CALL_LONG)
4447 fprintf (stderr, " longcall,");
4448
4449 fprintf (stderr, " proto = %d, nargs = %d\n",
4450 cum->prototype, cum->nargs_prototype);
4451 }
4452
4453 if (fntype
4454 && !TARGET_ALTIVEC
4455 && TARGET_ALTIVEC_ABI
4456 && ALTIVEC_VECTOR_MODE (TYPE_MODE (TREE_TYPE (fntype))))
4457 {
4458 error ("cannot return value in vector register because"
4459 " altivec instructions are disabled, use -maltivec"
4460 " to enable them");
4461 }
4462 }
4463
4464 /* Return true if TYPE must be passed on the stack and not in registers. */
4465
4466 static bool
rs6000_must_pass_in_stack(enum machine_mode mode,tree type)4467 rs6000_must_pass_in_stack (enum machine_mode mode, tree type)
4468 {
4469 if (DEFAULT_ABI == ABI_AIX || TARGET_64BIT)
4470 return must_pass_in_stack_var_size (mode, type);
4471 else
4472 return must_pass_in_stack_var_size_or_pad (mode, type);
4473 }
4474
4475 /* If defined, a C expression which determines whether, and in which
4476 direction, to pad out an argument with extra space. The value
4477 should be of type `enum direction': either `upward' to pad above
4478 the argument, `downward' to pad below, or `none' to inhibit
4479 padding.
4480
4481 For the AIX ABI structs are always stored left shifted in their
4482 argument slot. */
4483
4484 enum direction
function_arg_padding(enum machine_mode mode,tree type)4485 function_arg_padding (enum machine_mode mode, tree type)
4486 {
4487 #ifndef AGGREGATE_PADDING_FIXED
4488 #define AGGREGATE_PADDING_FIXED 0
4489 #endif
4490 #ifndef AGGREGATES_PAD_UPWARD_ALWAYS
4491 #define AGGREGATES_PAD_UPWARD_ALWAYS 0
4492 #endif
4493
4494 if (!AGGREGATE_PADDING_FIXED)
4495 {
4496 /* GCC used to pass structures of the same size as integer types as
4497 if they were in fact integers, ignoring FUNCTION_ARG_PADDING.
4498 i.e. Structures of size 1 or 2 (or 4 when TARGET_64BIT) were
4499 passed padded downward, except that -mstrict-align further
4500 muddied the water in that multi-component structures of 2 and 4
4501 bytes in size were passed padded upward.
4502
4503 The following arranges for best compatibility with previous
4504 versions of gcc, but removes the -mstrict-align dependency. */
4505 if (BYTES_BIG_ENDIAN)
4506 {
4507 HOST_WIDE_INT size = 0;
4508
4509 if (mode == BLKmode)
4510 {
4511 if (type && TREE_CODE (TYPE_SIZE (type)) == INTEGER_CST)
4512 size = int_size_in_bytes (type);
4513 }
4514 else
4515 size = GET_MODE_SIZE (mode);
4516
4517 if (size == 1 || size == 2 || size == 4)
4518 return downward;
4519 }
4520 return upward;
4521 }
4522
4523 if (AGGREGATES_PAD_UPWARD_ALWAYS)
4524 {
4525 if (type != 0 && AGGREGATE_TYPE_P (type))
4526 return upward;
4527 }
4528
4529 /* Fall back to the default. */
4530 return DEFAULT_FUNCTION_ARG_PADDING (mode, type);
4531 }
4532
4533 /* If defined, a C expression that gives the alignment boundary, in bits,
4534 of an argument with the specified mode and type. If it is not defined,
4535 PARM_BOUNDARY is used for all arguments.
4536
4537 V.4 wants long longs and doubles to be double word aligned. Just
4538 testing the mode size is a boneheaded way to do this as it means
4539 that other types such as complex int are also double word aligned.
4540 However, we're stuck with this because changing the ABI might break
4541 existing library interfaces.
4542
4543 Doubleword align SPE vectors.
4544 Quadword align Altivec vectors.
4545 Quadword align large synthetic vector types. */
4546
4547 int
function_arg_boundary(enum machine_mode mode,tree type)4548 function_arg_boundary (enum machine_mode mode, tree type)
4549 {
4550 if (DEFAULT_ABI == ABI_V4
4551 && (GET_MODE_SIZE (mode) == 8
4552 || (TARGET_HARD_FLOAT
4553 && TARGET_FPRS
4554 && mode == TFmode)))
4555 return 64;
4556 else if (SPE_VECTOR_MODE (mode)
4557 || (type && TREE_CODE (type) == VECTOR_TYPE
4558 && int_size_in_bytes (type) >= 8
4559 && int_size_in_bytes (type) < 16))
4560 return 64;
4561 else if (ALTIVEC_VECTOR_MODE (mode)
4562 || (type && TREE_CODE (type) == VECTOR_TYPE
4563 && int_size_in_bytes (type) >= 16))
4564 return 128;
4565 else if (rs6000_darwin64_abi && mode == BLKmode
4566 && type && TYPE_ALIGN (type) > 64)
4567 return 128;
4568 else
4569 return PARM_BOUNDARY;
4570 }
4571
4572 /* For a function parm of MODE and TYPE, return the starting word in
4573 the parameter area. NWORDS of the parameter area are already used. */
4574
4575 static unsigned int
rs6000_parm_start(enum machine_mode mode,tree type,unsigned int nwords)4576 rs6000_parm_start (enum machine_mode mode, tree type, unsigned int nwords)
4577 {
4578 unsigned int align;
4579 unsigned int parm_offset;
4580
4581 align = function_arg_boundary (mode, type) / PARM_BOUNDARY - 1;
4582 parm_offset = DEFAULT_ABI == ABI_V4 ? 2 : 6;
4583 return nwords + (-(parm_offset + nwords) & align);
4584 }
4585
4586 /* Compute the size (in words) of a function argument. */
4587
4588 static unsigned long
rs6000_arg_size(enum machine_mode mode,tree type)4589 rs6000_arg_size (enum machine_mode mode, tree type)
4590 {
4591 unsigned long size;
4592
4593 if (mode != BLKmode)
4594 size = GET_MODE_SIZE (mode);
4595 else
4596 size = int_size_in_bytes (type);
4597
4598 if (TARGET_32BIT)
4599 return (size + 3) >> 2;
4600 else
4601 return (size + 7) >> 3;
4602 }
4603
4604 /* Use this to flush pending int fields. */
4605
4606 static void
rs6000_darwin64_record_arg_advance_flush(CUMULATIVE_ARGS * cum,HOST_WIDE_INT bitpos)4607 rs6000_darwin64_record_arg_advance_flush (CUMULATIVE_ARGS *cum,
4608 HOST_WIDE_INT bitpos)
4609 {
4610 unsigned int startbit, endbit;
4611 int intregs, intoffset;
4612 enum machine_mode mode;
4613
4614 if (cum->intoffset == -1)
4615 return;
4616
4617 intoffset = cum->intoffset;
4618 cum->intoffset = -1;
4619
4620 if (intoffset % BITS_PER_WORD != 0)
4621 {
4622 mode = mode_for_size (BITS_PER_WORD - intoffset % BITS_PER_WORD,
4623 MODE_INT, 0);
4624 if (mode == BLKmode)
4625 {
4626 /* We couldn't find an appropriate mode, which happens,
4627 e.g., in packed structs when there are 3 bytes to load.
4628 Back intoffset back to the beginning of the word in this
4629 case. */
4630 intoffset = intoffset & -BITS_PER_WORD;
4631 }
4632 }
4633
4634 startbit = intoffset & -BITS_PER_WORD;
4635 endbit = (bitpos + BITS_PER_WORD - 1) & -BITS_PER_WORD;
4636 intregs = (endbit - startbit) / BITS_PER_WORD;
4637 cum->words += intregs;
4638 }
4639
4640 /* The darwin64 ABI calls for us to recurse down through structs,
4641 looking for elements passed in registers. Unfortunately, we have
4642 to track int register count here also because of misalignments
4643 in powerpc alignment mode. */
4644
4645 static void
rs6000_darwin64_record_arg_advance_recurse(CUMULATIVE_ARGS * cum,tree type,HOST_WIDE_INT startbitpos)4646 rs6000_darwin64_record_arg_advance_recurse (CUMULATIVE_ARGS *cum,
4647 tree type,
4648 HOST_WIDE_INT startbitpos)
4649 {
4650 tree f;
4651
4652 for (f = TYPE_FIELDS (type); f ; f = TREE_CHAIN (f))
4653 if (TREE_CODE (f) == FIELD_DECL)
4654 {
4655 HOST_WIDE_INT bitpos = startbitpos;
4656 tree ftype = TREE_TYPE (f);
4657 enum machine_mode mode;
4658 if (ftype == error_mark_node)
4659 continue;
4660 mode = TYPE_MODE (ftype);
4661
4662 if (DECL_SIZE (f) != 0
4663 && host_integerp (bit_position (f), 1))
4664 bitpos += int_bit_position (f);
4665
4666 /* ??? FIXME: else assume zero offset. */
4667
4668 if (TREE_CODE (ftype) == RECORD_TYPE)
4669 rs6000_darwin64_record_arg_advance_recurse (cum, ftype, bitpos);
4670 else if (USE_FP_FOR_ARG_P (cum, mode, ftype))
4671 {
4672 rs6000_darwin64_record_arg_advance_flush (cum, bitpos);
4673 cum->fregno += (GET_MODE_SIZE (mode) + 7) >> 3;
4674 cum->words += (GET_MODE_SIZE (mode) + 7) >> 3;
4675 }
4676 else if (USE_ALTIVEC_FOR_ARG_P (cum, mode, type, 1))
4677 {
4678 rs6000_darwin64_record_arg_advance_flush (cum, bitpos);
4679 cum->vregno++;
4680 cum->words += 2;
4681 }
4682 else if (cum->intoffset == -1)
4683 cum->intoffset = bitpos;
4684 }
4685 }
4686
4687 /* Update the data in CUM to advance over an argument
4688 of mode MODE and data type TYPE.
4689 (TYPE is null for libcalls where that information may not be available.)
4690
4691 Note that for args passed by reference, function_arg will be called
4692 with MODE and TYPE set to that of the pointer to the arg, not the arg
4693 itself. */
4694
4695 void
function_arg_advance(CUMULATIVE_ARGS * cum,enum machine_mode mode,tree type,int named,int depth)4696 function_arg_advance (CUMULATIVE_ARGS *cum, enum machine_mode mode,
4697 tree type, int named, int depth)
4698 {
4699 int size;
4700
4701 /* Only tick off an argument if we're not recursing. */
4702 if (depth == 0)
4703 cum->nargs_prototype--;
4704
4705 if (TARGET_ALTIVEC_ABI
4706 && (ALTIVEC_VECTOR_MODE (mode)
4707 || (type && TREE_CODE (type) == VECTOR_TYPE
4708 && int_size_in_bytes (type) == 16)))
4709 {
4710 bool stack = false;
4711
4712 if (USE_ALTIVEC_FOR_ARG_P (cum, mode, type, named))
4713 {
4714 cum->vregno++;
4715 if (!TARGET_ALTIVEC)
4716 error ("cannot pass argument in vector register because"
4717 " altivec instructions are disabled, use -maltivec"
4718 " to enable them");
4719
4720 /* PowerPC64 Linux and AIX allocate GPRs for a vector argument
4721 even if it is going to be passed in a vector register.
4722 Darwin does the same for variable-argument functions. */
4723 if ((DEFAULT_ABI == ABI_AIX && TARGET_64BIT)
4724 || (cum->stdarg && DEFAULT_ABI != ABI_V4))
4725 stack = true;
4726 }
4727 else
4728 stack = true;
4729
4730 if (stack)
4731 {
4732 int align;
4733
4734 /* Vector parameters must be 16-byte aligned. This places
4735 them at 2 mod 4 in terms of words in 32-bit mode, since
4736 the parameter save area starts at offset 24 from the
4737 stack. In 64-bit mode, they just have to start on an
4738 even word, since the parameter save area is 16-byte
4739 aligned. Space for GPRs is reserved even if the argument
4740 will be passed in memory. */
4741 if (TARGET_32BIT)
4742 align = (2 - cum->words) & 3;
4743 else
4744 align = cum->words & 1;
4745 cum->words += align + rs6000_arg_size (mode, type);
4746
4747 if (TARGET_DEBUG_ARG)
4748 {
4749 fprintf (stderr, "function_adv: words = %2d, align=%d, ",
4750 cum->words, align);
4751 fprintf (stderr, "nargs = %4d, proto = %d, mode = %4s\n",
4752 cum->nargs_prototype, cum->prototype,
4753 GET_MODE_NAME (mode));
4754 }
4755 }
4756 }
4757 else if (TARGET_SPE_ABI && TARGET_SPE && SPE_VECTOR_MODE (mode)
4758 && !cum->stdarg
4759 && cum->sysv_gregno <= GP_ARG_MAX_REG)
4760 cum->sysv_gregno++;
4761
4762 else if (rs6000_darwin64_abi
4763 && mode == BLKmode
4764 && TREE_CODE (type) == RECORD_TYPE
4765 && (size = int_size_in_bytes (type)) > 0)
4766 {
4767 /* Variable sized types have size == -1 and are
4768 treated as if consisting entirely of ints.
4769 Pad to 16 byte boundary if needed. */
4770 if (TYPE_ALIGN (type) >= 2 * BITS_PER_WORD
4771 && (cum->words % 2) != 0)
4772 cum->words++;
4773 /* For varargs, we can just go up by the size of the struct. */
4774 if (!named)
4775 cum->words += (size + 7) / 8;
4776 else
4777 {
4778 /* It is tempting to say int register count just goes up by
4779 sizeof(type)/8, but this is wrong in a case such as
4780 { int; double; int; } [powerpc alignment]. We have to
4781 grovel through the fields for these too. */
4782 cum->intoffset = 0;
4783 rs6000_darwin64_record_arg_advance_recurse (cum, type, 0);
4784 rs6000_darwin64_record_arg_advance_flush (cum,
4785 size * BITS_PER_UNIT);
4786 }
4787 }
4788 else if (DEFAULT_ABI == ABI_V4)
4789 {
4790 if (TARGET_HARD_FLOAT && TARGET_FPRS
4791 && (mode == SFmode || mode == DFmode
4792 || (mode == TFmode && !TARGET_IEEEQUAD)))
4793 {
4794 if (cum->fregno + (mode == TFmode ? 1 : 0) <= FP_ARG_V4_MAX_REG)
4795 cum->fregno += (GET_MODE_SIZE (mode) + 7) >> 3;
4796 else
4797 {
4798 cum->fregno = FP_ARG_V4_MAX_REG + 1;
4799 if (mode == DFmode || mode == TFmode)
4800 cum->words += cum->words & 1;
4801 cum->words += rs6000_arg_size (mode, type);
4802 }
4803 }
4804 else
4805 {
4806 int n_words = rs6000_arg_size (mode, type);
4807 int gregno = cum->sysv_gregno;
4808
4809 /* Long long and SPE vectors are put in (r3,r4), (r5,r6),
4810 (r7,r8) or (r9,r10). As does any other 2 word item such
4811 as complex int due to a historical mistake. */
4812 if (n_words == 2)
4813 gregno += (1 - gregno) & 1;
4814
4815 /* Multi-reg args are not split between registers and stack. */
4816 if (gregno + n_words - 1 > GP_ARG_MAX_REG)
4817 {
4818 /* Long long and SPE vectors are aligned on the stack.
4819 So are other 2 word items such as complex int due to
4820 a historical mistake. */
4821 if (n_words == 2)
4822 cum->words += cum->words & 1;
4823 cum->words += n_words;
4824 }
4825
4826 /* Note: continuing to accumulate gregno past when we've started
4827 spilling to the stack indicates the fact that we've started
4828 spilling to the stack to expand_builtin_saveregs. */
4829 cum->sysv_gregno = gregno + n_words;
4830 }
4831
4832 if (TARGET_DEBUG_ARG)
4833 {
4834 fprintf (stderr, "function_adv: words = %2d, fregno = %2d, ",
4835 cum->words, cum->fregno);
4836 fprintf (stderr, "gregno = %2d, nargs = %4d, proto = %d, ",
4837 cum->sysv_gregno, cum->nargs_prototype, cum->prototype);
4838 fprintf (stderr, "mode = %4s, named = %d\n",
4839 GET_MODE_NAME (mode), named);
4840 }
4841 }
4842 else
4843 {
4844 int n_words = rs6000_arg_size (mode, type);
4845 int start_words = cum->words;
4846 int align_words = rs6000_parm_start (mode, type, start_words);
4847
4848 cum->words = align_words + n_words;
4849
4850 if (SCALAR_FLOAT_MODE_P (mode)
4851 && !DECIMAL_FLOAT_MODE_P (mode)
4852 && TARGET_HARD_FLOAT && TARGET_FPRS)
4853 cum->fregno += (GET_MODE_SIZE (mode) + 7) >> 3;
4854
4855 if (TARGET_DEBUG_ARG)
4856 {
4857 fprintf (stderr, "function_adv: words = %2d, fregno = %2d, ",
4858 cum->words, cum->fregno);
4859 fprintf (stderr, "nargs = %4d, proto = %d, mode = %4s, ",
4860 cum->nargs_prototype, cum->prototype, GET_MODE_NAME (mode));
4861 fprintf (stderr, "named = %d, align = %d, depth = %d\n",
4862 named, align_words - start_words, depth);
4863 }
4864 }
4865 }
4866
4867 static rtx
spe_build_register_parallel(enum machine_mode mode,int gregno)4868 spe_build_register_parallel (enum machine_mode mode, int gregno)
4869 {
4870 rtx r1, r3;
4871
4872 switch (mode)
4873 {
4874 case DFmode:
4875 r1 = gen_rtx_REG (DImode, gregno);
4876 r1 = gen_rtx_EXPR_LIST (VOIDmode, r1, const0_rtx);
4877 return gen_rtx_PARALLEL (mode, gen_rtvec (1, r1));
4878
4879 case DCmode:
4880 r1 = gen_rtx_REG (DImode, gregno);
4881 r1 = gen_rtx_EXPR_LIST (VOIDmode, r1, const0_rtx);
4882 r3 = gen_rtx_REG (DImode, gregno + 2);
4883 r3 = gen_rtx_EXPR_LIST (VOIDmode, r3, GEN_INT (8));
4884 return gen_rtx_PARALLEL (mode, gen_rtvec (2, r1, r3));
4885
4886 default:
4887 gcc_unreachable ();
4888 }
4889 }
4890
4891 /* Determine where to put a SIMD argument on the SPE. */
4892 static rtx
rs6000_spe_function_arg(CUMULATIVE_ARGS * cum,enum machine_mode mode,tree type)4893 rs6000_spe_function_arg (CUMULATIVE_ARGS *cum, enum machine_mode mode,
4894 tree type)
4895 {
4896 int gregno = cum->sysv_gregno;
4897
4898 /* On E500 v2, double arithmetic is done on the full 64-bit GPR, but
4899 are passed and returned in a pair of GPRs for ABI compatibility. */
4900 if (TARGET_E500_DOUBLE && (mode == DFmode || mode == DCmode))
4901 {
4902 int n_words = rs6000_arg_size (mode, type);
4903
4904 /* Doubles go in an odd/even register pair (r5/r6, etc). */
4905 if (mode == DFmode)
4906 gregno += (1 - gregno) & 1;
4907
4908 /* Multi-reg args are not split between registers and stack. */
4909 if (gregno + n_words - 1 > GP_ARG_MAX_REG)
4910 return NULL_RTX;
4911
4912 return spe_build_register_parallel (mode, gregno);
4913 }
4914 if (cum->stdarg)
4915 {
4916 int n_words = rs6000_arg_size (mode, type);
4917
4918 /* SPE vectors are put in odd registers. */
4919 if (n_words == 2 && (gregno & 1) == 0)
4920 gregno += 1;
4921
4922 if (gregno + n_words - 1 <= GP_ARG_MAX_REG)
4923 {
4924 rtx r1, r2;
4925 enum machine_mode m = SImode;
4926
4927 r1 = gen_rtx_REG (m, gregno);
4928 r1 = gen_rtx_EXPR_LIST (m, r1, const0_rtx);
4929 r2 = gen_rtx_REG (m, gregno + 1);
4930 r2 = gen_rtx_EXPR_LIST (m, r2, GEN_INT (4));
4931 return gen_rtx_PARALLEL (mode, gen_rtvec (2, r1, r2));
4932 }
4933 else
4934 return NULL_RTX;
4935 }
4936 else
4937 {
4938 if (gregno <= GP_ARG_MAX_REG)
4939 return gen_rtx_REG (mode, gregno);
4940 else
4941 return NULL_RTX;
4942 }
4943 }
4944
4945 /* A subroutine of rs6000_darwin64_record_arg. Assign the bits of the
4946 structure between cum->intoffset and bitpos to integer registers. */
4947
4948 static void
rs6000_darwin64_record_arg_flush(CUMULATIVE_ARGS * cum,HOST_WIDE_INT bitpos,rtx rvec[],int * k)4949 rs6000_darwin64_record_arg_flush (CUMULATIVE_ARGS *cum,
4950 HOST_WIDE_INT bitpos, rtx rvec[], int *k)
4951 {
4952 enum machine_mode mode;
4953 unsigned int regno;
4954 unsigned int startbit, endbit;
4955 int this_regno, intregs, intoffset;
4956 rtx reg;
4957
4958 if (cum->intoffset == -1)
4959 return;
4960
4961 intoffset = cum->intoffset;
4962 cum->intoffset = -1;
4963
4964 /* If this is the trailing part of a word, try to only load that
4965 much into the register. Otherwise load the whole register. Note
4966 that in the latter case we may pick up unwanted bits. It's not a
4967 problem at the moment but may wish to revisit. */
4968
4969 if (intoffset % BITS_PER_WORD != 0)
4970 {
4971 mode = mode_for_size (BITS_PER_WORD - intoffset % BITS_PER_WORD,
4972 MODE_INT, 0);
4973 if (mode == BLKmode)
4974 {
4975 /* We couldn't find an appropriate mode, which happens,
4976 e.g., in packed structs when there are 3 bytes to load.
4977 Back intoffset back to the beginning of the word in this
4978 case. */
4979 intoffset = intoffset & -BITS_PER_WORD;
4980 mode = word_mode;
4981 }
4982 }
4983 else
4984 mode = word_mode;
4985
4986 startbit = intoffset & -BITS_PER_WORD;
4987 endbit = (bitpos + BITS_PER_WORD - 1) & -BITS_PER_WORD;
4988 intregs = (endbit - startbit) / BITS_PER_WORD;
4989 this_regno = cum->words + intoffset / BITS_PER_WORD;
4990
4991 if (intregs > 0 && intregs > GP_ARG_NUM_REG - this_regno)
4992 cum->use_stack = 1;
4993
4994 intregs = MIN (intregs, GP_ARG_NUM_REG - this_regno);
4995 if (intregs <= 0)
4996 return;
4997
4998 intoffset /= BITS_PER_UNIT;
4999 do
5000 {
5001 regno = GP_ARG_MIN_REG + this_regno;
5002 reg = gen_rtx_REG (mode, regno);
5003 rvec[(*k)++] =
5004 gen_rtx_EXPR_LIST (VOIDmode, reg, GEN_INT (intoffset));
5005
5006 this_regno += 1;
5007 intoffset = (intoffset | (UNITS_PER_WORD-1)) + 1;
5008 mode = word_mode;
5009 intregs -= 1;
5010 }
5011 while (intregs > 0);
5012 }
5013
5014 /* Recursive workhorse for the following. */
5015
5016 static void
rs6000_darwin64_record_arg_recurse(CUMULATIVE_ARGS * cum,tree type,HOST_WIDE_INT startbitpos,rtx rvec[],int * k)5017 rs6000_darwin64_record_arg_recurse (CUMULATIVE_ARGS *cum, tree type,
5018 HOST_WIDE_INT startbitpos, rtx rvec[],
5019 int *k)
5020 {
5021 tree f;
5022
5023 for (f = TYPE_FIELDS (type); f ; f = TREE_CHAIN (f))
5024 if (TREE_CODE (f) == FIELD_DECL)
5025 {
5026 HOST_WIDE_INT bitpos = startbitpos;
5027 tree ftype = TREE_TYPE (f);
5028 enum machine_mode mode;
5029 if (ftype == error_mark_node)
5030 continue;
5031 mode = TYPE_MODE (ftype);
5032
5033 if (DECL_SIZE (f) != 0
5034 && host_integerp (bit_position (f), 1))
5035 bitpos += int_bit_position (f);
5036
5037 /* ??? FIXME: else assume zero offset. */
5038
5039 if (TREE_CODE (ftype) == RECORD_TYPE)
5040 rs6000_darwin64_record_arg_recurse (cum, ftype, bitpos, rvec, k);
5041 else if (cum->named && USE_FP_FOR_ARG_P (cum, mode, ftype))
5042 {
5043 #if 0
5044 switch (mode)
5045 {
5046 case SCmode: mode = SFmode; break;
5047 case DCmode: mode = DFmode; break;
5048 case TCmode: mode = TFmode; break;
5049 default: break;
5050 }
5051 #endif
5052 rs6000_darwin64_record_arg_flush (cum, bitpos, rvec, k);
5053 rvec[(*k)++]
5054 = gen_rtx_EXPR_LIST (VOIDmode,
5055 gen_rtx_REG (mode, cum->fregno++),
5056 GEN_INT (bitpos / BITS_PER_UNIT));
5057 if (mode == TFmode)
5058 cum->fregno++;
5059 }
5060 else if (cum->named && USE_ALTIVEC_FOR_ARG_P (cum, mode, ftype, 1))
5061 {
5062 rs6000_darwin64_record_arg_flush (cum, bitpos, rvec, k);
5063 rvec[(*k)++]
5064 = gen_rtx_EXPR_LIST (VOIDmode,
5065 gen_rtx_REG (mode, cum->vregno++),
5066 GEN_INT (bitpos / BITS_PER_UNIT));
5067 }
5068 else if (cum->intoffset == -1)
5069 cum->intoffset = bitpos;
5070 }
5071 }
5072
5073 /* For the darwin64 ABI, we want to construct a PARALLEL consisting of
5074 the register(s) to be used for each field and subfield of a struct
5075 being passed by value, along with the offset of where the
5076 register's value may be found in the block. FP fields go in FP
5077 register, vector fields go in vector registers, and everything
5078 else goes in int registers, packed as in memory.
5079
5080 This code is also used for function return values. RETVAL indicates
5081 whether this is the case.
5082
5083 Much of this is taken from the SPARC V9 port, which has a similar
5084 calling convention. */
5085
5086 static rtx
rs6000_darwin64_record_arg(CUMULATIVE_ARGS * orig_cum,tree type,int named,bool retval)5087 rs6000_darwin64_record_arg (CUMULATIVE_ARGS *orig_cum, tree type,
5088 int named, bool retval)
5089 {
5090 rtx rvec[FIRST_PSEUDO_REGISTER];
5091 int k = 1, kbase = 1;
5092 HOST_WIDE_INT typesize = int_size_in_bytes (type);
5093 /* This is a copy; modifications are not visible to our caller. */
5094 CUMULATIVE_ARGS copy_cum = *orig_cum;
5095 CUMULATIVE_ARGS *cum = ©_cum;
5096
5097 /* Pad to 16 byte boundary if needed. */
5098 if (!retval && TYPE_ALIGN (type) >= 2 * BITS_PER_WORD
5099 && (cum->words % 2) != 0)
5100 cum->words++;
5101
5102 cum->intoffset = 0;
5103 cum->use_stack = 0;
5104 cum->named = named;
5105
5106 /* Put entries into rvec[] for individual FP and vector fields, and
5107 for the chunks of memory that go in int regs. Note we start at
5108 element 1; 0 is reserved for an indication of using memory, and
5109 may or may not be filled in below. */
5110 rs6000_darwin64_record_arg_recurse (cum, type, 0, rvec, &k);
5111 rs6000_darwin64_record_arg_flush (cum, typesize * BITS_PER_UNIT, rvec, &k);
5112
5113 /* If any part of the struct went on the stack put all of it there.
5114 This hack is because the generic code for
5115 FUNCTION_ARG_PARTIAL_NREGS cannot handle cases where the register
5116 parts of the struct are not at the beginning. */
5117 if (cum->use_stack)
5118 {
5119 if (retval)
5120 return NULL_RTX; /* doesn't go in registers at all */
5121 kbase = 0;
5122 rvec[0] = gen_rtx_EXPR_LIST (VOIDmode, NULL_RTX, const0_rtx);
5123 }
5124 if (k > 1 || cum->use_stack)
5125 return gen_rtx_PARALLEL (BLKmode, gen_rtvec_v (k - kbase, &rvec[kbase]));
5126 else
5127 return NULL_RTX;
5128 }
5129
5130 /* Determine where to place an argument in 64-bit mode with 32-bit ABI. */
5131
5132 static rtx
rs6000_mixed_function_arg(enum machine_mode mode,tree type,int align_words)5133 rs6000_mixed_function_arg (enum machine_mode mode, tree type, int align_words)
5134 {
5135 int n_units;
5136 int i, k;
5137 rtx rvec[GP_ARG_NUM_REG + 1];
5138
5139 if (align_words >= GP_ARG_NUM_REG)
5140 return NULL_RTX;
5141
5142 n_units = rs6000_arg_size (mode, type);
5143
5144 /* Optimize the simple case where the arg fits in one gpr, except in
5145 the case of BLKmode due to assign_parms assuming that registers are
5146 BITS_PER_WORD wide. */
5147 if (n_units == 0
5148 || (n_units == 1 && mode != BLKmode))
5149 return gen_rtx_REG (mode, GP_ARG_MIN_REG + align_words);
5150
5151 k = 0;
5152 if (align_words + n_units > GP_ARG_NUM_REG)
5153 /* Not all of the arg fits in gprs. Say that it goes in memory too,
5154 using a magic NULL_RTX component.
5155 This is not strictly correct. Only some of the arg belongs in
5156 memory, not all of it. However, the normal scheme using
5157 function_arg_partial_nregs can result in unusual subregs, eg.
5158 (subreg:SI (reg:DF) 4), which are not handled well. The code to
5159 store the whole arg to memory is often more efficient than code
5160 to store pieces, and we know that space is available in the right
5161 place for the whole arg. */
5162 rvec[k++] = gen_rtx_EXPR_LIST (VOIDmode, NULL_RTX, const0_rtx);
5163
5164 i = 0;
5165 do
5166 {
5167 rtx r = gen_rtx_REG (SImode, GP_ARG_MIN_REG + align_words);
5168 rtx off = GEN_INT (i++ * 4);
5169 rvec[k++] = gen_rtx_EXPR_LIST (VOIDmode, r, off);
5170 }
5171 while (++align_words < GP_ARG_NUM_REG && --n_units != 0);
5172
5173 return gen_rtx_PARALLEL (mode, gen_rtvec_v (k, rvec));
5174 }
5175
5176 /* Determine where to put an argument to a function.
5177 Value is zero to push the argument on the stack,
5178 or a hard register in which to store the argument.
5179
5180 MODE is the argument's machine mode.
5181 TYPE is the data type of the argument (as a tree).
5182 This is null for libcalls where that information may
5183 not be available.
5184 CUM is a variable of type CUMULATIVE_ARGS which gives info about
5185 the preceding args and about the function being called. It is
5186 not modified in this routine.
5187 NAMED is nonzero if this argument is a named parameter
5188 (otherwise it is an extra parameter matching an ellipsis).
5189
5190 On RS/6000 the first eight words of non-FP are normally in registers
5191 and the rest are pushed. Under AIX, the first 13 FP args are in registers.
5192 Under V.4, the first 8 FP args are in registers.
5193
5194 If this is floating-point and no prototype is specified, we use
5195 both an FP and integer register (or possibly FP reg and stack). Library
5196 functions (when CALL_LIBCALL is set) always have the proper types for args,
5197 so we can pass the FP value just in one register. emit_library_function
5198 doesn't support PARALLEL anyway.
5199
5200 Note that for args passed by reference, function_arg will be called
5201 with MODE and TYPE set to that of the pointer to the arg, not the arg
5202 itself. */
5203
5204 rtx
function_arg(CUMULATIVE_ARGS * cum,enum machine_mode mode,tree type,int named)5205 function_arg (CUMULATIVE_ARGS *cum, enum machine_mode mode,
5206 tree type, int named)
5207 {
5208 enum rs6000_abi abi = DEFAULT_ABI;
5209
5210 /* Return a marker to indicate whether CR1 needs to set or clear the
5211 bit that V.4 uses to say fp args were passed in registers.
5212 Assume that we don't need the marker for software floating point,
5213 or compiler generated library calls. */
5214 if (mode == VOIDmode)
5215 {
5216 if (abi == ABI_V4
5217 && (cum->call_cookie & CALL_LIBCALL) == 0
5218 && (cum->stdarg
5219 || (cum->nargs_prototype < 0
5220 && (cum->prototype || TARGET_NO_PROTOTYPE))))
5221 {
5222 /* For the SPE, we need to crxor CR6 always. */
5223 if (TARGET_SPE_ABI)
5224 return GEN_INT (cum->call_cookie | CALL_V4_SET_FP_ARGS);
5225 else if (TARGET_HARD_FLOAT && TARGET_FPRS)
5226 return GEN_INT (cum->call_cookie
5227 | ((cum->fregno == FP_ARG_MIN_REG)
5228 ? CALL_V4_SET_FP_ARGS
5229 : CALL_V4_CLEAR_FP_ARGS));
5230 }
5231
5232 return GEN_INT (cum->call_cookie);
5233 }
5234
5235 if (rs6000_darwin64_abi && mode == BLKmode
5236 && TREE_CODE (type) == RECORD_TYPE)
5237 {
5238 rtx rslt = rs6000_darwin64_record_arg (cum, type, named, false);
5239 if (rslt != NULL_RTX)
5240 return rslt;
5241 /* Else fall through to usual handling. */
5242 }
5243
5244 if (USE_ALTIVEC_FOR_ARG_P (cum, mode, type, named))
5245 if (TARGET_64BIT && ! cum->prototype)
5246 {
5247 /* Vector parameters get passed in vector register
5248 and also in GPRs or memory, in absence of prototype. */
5249 int align_words;
5250 rtx slot;
5251 align_words = (cum->words + 1) & ~1;
5252
5253 if (align_words >= GP_ARG_NUM_REG)
5254 {
5255 slot = NULL_RTX;
5256 }
5257 else
5258 {
5259 slot = gen_rtx_REG (mode, GP_ARG_MIN_REG + align_words);
5260 }
5261 return gen_rtx_PARALLEL (mode,
5262 gen_rtvec (2,
5263 gen_rtx_EXPR_LIST (VOIDmode,
5264 slot, const0_rtx),
5265 gen_rtx_EXPR_LIST (VOIDmode,
5266 gen_rtx_REG (mode, cum->vregno),
5267 const0_rtx)));
5268 }
5269 else
5270 return gen_rtx_REG (mode, cum->vregno);
5271 else if (TARGET_ALTIVEC_ABI
5272 && (ALTIVEC_VECTOR_MODE (mode)
5273 || (type && TREE_CODE (type) == VECTOR_TYPE
5274 && int_size_in_bytes (type) == 16)))
5275 {
5276 if (named || abi == ABI_V4)
5277 return NULL_RTX;
5278 else
5279 {
5280 /* Vector parameters to varargs functions under AIX or Darwin
5281 get passed in memory and possibly also in GPRs. */
5282 int align, align_words, n_words;
5283 enum machine_mode part_mode;
5284
5285 /* Vector parameters must be 16-byte aligned. This places them at
5286 2 mod 4 in terms of words in 32-bit mode, since the parameter
5287 save area starts at offset 24 from the stack. In 64-bit mode,
5288 they just have to start on an even word, since the parameter
5289 save area is 16-byte aligned. */
5290 if (TARGET_32BIT)
5291 align = (2 - cum->words) & 3;
5292 else
5293 align = cum->words & 1;
5294 align_words = cum->words + align;
5295
5296 /* Out of registers? Memory, then. */
5297 if (align_words >= GP_ARG_NUM_REG)
5298 return NULL_RTX;
5299
5300 if (TARGET_32BIT && TARGET_POWERPC64)
5301 return rs6000_mixed_function_arg (mode, type, align_words);
5302
5303 /* The vector value goes in GPRs. Only the part of the
5304 value in GPRs is reported here. */
5305 part_mode = mode;
5306 n_words = rs6000_arg_size (mode, type);
5307 if (align_words + n_words > GP_ARG_NUM_REG)
5308 /* Fortunately, there are only two possibilities, the value
5309 is either wholly in GPRs or half in GPRs and half not. */
5310 part_mode = DImode;
5311
5312 return gen_rtx_REG (part_mode, GP_ARG_MIN_REG + align_words);
5313 }
5314 }
5315 else if (TARGET_SPE_ABI && TARGET_SPE
5316 && (SPE_VECTOR_MODE (mode)
5317 || (TARGET_E500_DOUBLE && (mode == DFmode
5318 || mode == DCmode))))
5319 return rs6000_spe_function_arg (cum, mode, type);
5320
5321 else if (abi == ABI_V4)
5322 {
5323 if (TARGET_HARD_FLOAT && TARGET_FPRS
5324 && (mode == SFmode || mode == DFmode
5325 || (mode == TFmode && !TARGET_IEEEQUAD)))
5326 {
5327 if (cum->fregno + (mode == TFmode ? 1 : 0) <= FP_ARG_V4_MAX_REG)
5328 return gen_rtx_REG (mode, cum->fregno);
5329 else
5330 return NULL_RTX;
5331 }
5332 else
5333 {
5334 int n_words = rs6000_arg_size (mode, type);
5335 int gregno = cum->sysv_gregno;
5336
5337 /* Long long and SPE vectors are put in (r3,r4), (r5,r6),
5338 (r7,r8) or (r9,r10). As does any other 2 word item such
5339 as complex int due to a historical mistake. */
5340 if (n_words == 2)
5341 gregno += (1 - gregno) & 1;
5342
5343 /* Multi-reg args are not split between registers and stack. */
5344 if (gregno + n_words - 1 > GP_ARG_MAX_REG)
5345 return NULL_RTX;
5346
5347 if (TARGET_32BIT && TARGET_POWERPC64)
5348 return rs6000_mixed_function_arg (mode, type,
5349 gregno - GP_ARG_MIN_REG);
5350 return gen_rtx_REG (mode, gregno);
5351 }
5352 }
5353 else
5354 {
5355 int align_words = rs6000_parm_start (mode, type, cum->words);
5356
5357 if (USE_FP_FOR_ARG_P (cum, mode, type))
5358 {
5359 rtx rvec[GP_ARG_NUM_REG + 1];
5360 rtx r;
5361 int k;
5362 bool needs_psave;
5363 enum machine_mode fmode = mode;
5364 unsigned long n_fpreg = (GET_MODE_SIZE (mode) + 7) >> 3;
5365
5366 if (cum->fregno + n_fpreg > FP_ARG_MAX_REG + 1)
5367 {
5368 /* Currently, we only ever need one reg here because complex
5369 doubles are split. */
5370 gcc_assert (cum->fregno == FP_ARG_MAX_REG && fmode == TFmode);
5371
5372 /* Long double split over regs and memory. */
5373 fmode = DFmode;
5374 }
5375
5376 /* Do we also need to pass this arg in the parameter save
5377 area? */
5378 needs_psave = (type
5379 && (cum->nargs_prototype <= 0
5380 || (DEFAULT_ABI == ABI_AIX
5381 && TARGET_XL_COMPAT
5382 && align_words >= GP_ARG_NUM_REG)));
5383
5384 if (!needs_psave && mode == fmode)
5385 return gen_rtx_REG (fmode, cum->fregno);
5386
5387 k = 0;
5388 if (needs_psave)
5389 {
5390 /* Describe the part that goes in gprs or the stack.
5391 This piece must come first, before the fprs. */
5392 if (align_words < GP_ARG_NUM_REG)
5393 {
5394 unsigned long n_words = rs6000_arg_size (mode, type);
5395
5396 if (align_words + n_words > GP_ARG_NUM_REG
5397 || (TARGET_32BIT && TARGET_POWERPC64))
5398 {
5399 /* If this is partially on the stack, then we only
5400 include the portion actually in registers here. */
5401 enum machine_mode rmode = TARGET_32BIT ? SImode : DImode;
5402 rtx off;
5403 int i = 0;
5404 if (align_words + n_words > GP_ARG_NUM_REG)
5405 /* Not all of the arg fits in gprs. Say that it
5406 goes in memory too, using a magic NULL_RTX
5407 component. Also see comment in
5408 rs6000_mixed_function_arg for why the normal
5409 function_arg_partial_nregs scheme doesn't work
5410 in this case. */
5411 rvec[k++] = gen_rtx_EXPR_LIST (VOIDmode, NULL_RTX,
5412 const0_rtx);
5413 do
5414 {
5415 r = gen_rtx_REG (rmode,
5416 GP_ARG_MIN_REG + align_words);
5417 off = GEN_INT (i++ * GET_MODE_SIZE (rmode));
5418 rvec[k++] = gen_rtx_EXPR_LIST (VOIDmode, r, off);
5419 }
5420 while (++align_words < GP_ARG_NUM_REG && --n_words != 0);
5421 }
5422 else
5423 {
5424 /* The whole arg fits in gprs. */
5425 r = gen_rtx_REG (mode, GP_ARG_MIN_REG + align_words);
5426 rvec[k++] = gen_rtx_EXPR_LIST (VOIDmode, r, const0_rtx);
5427 }
5428 }
5429 else
5430 /* It's entirely in memory. */
5431 rvec[k++] = gen_rtx_EXPR_LIST (VOIDmode, NULL_RTX, const0_rtx);
5432 }
5433
5434 /* Describe where this piece goes in the fprs. */
5435 r = gen_rtx_REG (fmode, cum->fregno);
5436 rvec[k++] = gen_rtx_EXPR_LIST (VOIDmode, r, const0_rtx);
5437
5438 return gen_rtx_PARALLEL (mode, gen_rtvec_v (k, rvec));
5439 }
5440 else if (align_words < GP_ARG_NUM_REG)
5441 {
5442 if (TARGET_32BIT && TARGET_POWERPC64)
5443 return rs6000_mixed_function_arg (mode, type, align_words);
5444
5445 if (mode == BLKmode)
5446 mode = Pmode;
5447
5448 return gen_rtx_REG (mode, GP_ARG_MIN_REG + align_words);
5449 }
5450 else
5451 return NULL_RTX;
5452 }
5453 }
5454
5455 /* For an arg passed partly in registers and partly in memory, this is
5456 the number of bytes passed in registers. For args passed entirely in
5457 registers or entirely in memory, zero. When an arg is described by a
5458 PARALLEL, perhaps using more than one register type, this function
5459 returns the number of bytes used by the first element of the PARALLEL. */
5460
5461 static int
rs6000_arg_partial_bytes(CUMULATIVE_ARGS * cum,enum machine_mode mode,tree type,bool named)5462 rs6000_arg_partial_bytes (CUMULATIVE_ARGS *cum, enum machine_mode mode,
5463 tree type, bool named)
5464 {
5465 int ret = 0;
5466 int align_words;
5467
5468 if (DEFAULT_ABI == ABI_V4)
5469 return 0;
5470
5471 if (USE_ALTIVEC_FOR_ARG_P (cum, mode, type, named)
5472 && cum->nargs_prototype >= 0)
5473 return 0;
5474
5475 /* In this complicated case we just disable the partial_nregs code. */
5476 if (rs6000_darwin64_abi && mode == BLKmode
5477 && TREE_CODE (type) == RECORD_TYPE
5478 && int_size_in_bytes (type) > 0)
5479 return 0;
5480
5481 align_words = rs6000_parm_start (mode, type, cum->words);
5482
5483 if (USE_FP_FOR_ARG_P (cum, mode, type))
5484 {
5485 /* If we are passing this arg in the fixed parameter save area
5486 (gprs or memory) as well as fprs, then this function should
5487 return the number of partial bytes passed in the parameter
5488 save area rather than partial bytes passed in fprs. */
5489 if (type
5490 && (cum->nargs_prototype <= 0
5491 || (DEFAULT_ABI == ABI_AIX
5492 && TARGET_XL_COMPAT
5493 && align_words >= GP_ARG_NUM_REG)))
5494 return 0;
5495 else if (cum->fregno + ((GET_MODE_SIZE (mode) + 7) >> 3)
5496 > FP_ARG_MAX_REG + 1)
5497 ret = (FP_ARG_MAX_REG + 1 - cum->fregno) * 8;
5498 else if (cum->nargs_prototype >= 0)
5499 return 0;
5500 }
5501
5502 if (align_words < GP_ARG_NUM_REG
5503 && GP_ARG_NUM_REG < align_words + rs6000_arg_size (mode, type))
5504 ret = (GP_ARG_NUM_REG - align_words) * (TARGET_32BIT ? 4 : 8);
5505
5506 if (ret != 0 && TARGET_DEBUG_ARG)
5507 fprintf (stderr, "rs6000_arg_partial_bytes: %d\n", ret);
5508
5509 return ret;
5510 }
5511
5512 /* A C expression that indicates when an argument must be passed by
5513 reference. If nonzero for an argument, a copy of that argument is
5514 made in memory and a pointer to the argument is passed instead of
5515 the argument itself. The pointer is passed in whatever way is
5516 appropriate for passing a pointer to that type.
5517
5518 Under V.4, aggregates and long double are passed by reference.
5519
5520 As an extension to all 32-bit ABIs, AltiVec vectors are passed by
5521 reference unless the AltiVec vector extension ABI is in force.
5522
5523 As an extension to all ABIs, variable sized types are passed by
5524 reference. */
5525
5526 static bool
rs6000_pass_by_reference(CUMULATIVE_ARGS * cum ATTRIBUTE_UNUSED,enum machine_mode mode,tree type,bool named ATTRIBUTE_UNUSED)5527 rs6000_pass_by_reference (CUMULATIVE_ARGS *cum ATTRIBUTE_UNUSED,
5528 enum machine_mode mode, tree type,
5529 bool named ATTRIBUTE_UNUSED)
5530 {
5531 if (DEFAULT_ABI == ABI_V4 && TARGET_IEEEQUAD && mode == TFmode)
5532 {
5533 if (TARGET_DEBUG_ARG)
5534 fprintf (stderr, "function_arg_pass_by_reference: V4 long double\n");
5535 return 1;
5536 }
5537
5538 if (!type)
5539 return 0;
5540
5541 if (DEFAULT_ABI == ABI_V4 && AGGREGATE_TYPE_P (type))
5542 {
5543 if (TARGET_DEBUG_ARG)
5544 fprintf (stderr, "function_arg_pass_by_reference: V4 aggregate\n");
5545 return 1;
5546 }
5547
5548 if (int_size_in_bytes (type) < 0)
5549 {
5550 if (TARGET_DEBUG_ARG)
5551 fprintf (stderr, "function_arg_pass_by_reference: variable size\n");
5552 return 1;
5553 }
5554
5555 /* Allow -maltivec -mabi=no-altivec without warning. Altivec vector
5556 modes only exist for GCC vector types if -maltivec. */
5557 if (TARGET_32BIT && !TARGET_ALTIVEC_ABI && ALTIVEC_VECTOR_MODE (mode))
5558 {
5559 if (TARGET_DEBUG_ARG)
5560 fprintf (stderr, "function_arg_pass_by_reference: AltiVec\n");
5561 return 1;
5562 }
5563
5564 /* Pass synthetic vectors in memory. */
5565 if (TREE_CODE (type) == VECTOR_TYPE
5566 && int_size_in_bytes (type) > (TARGET_ALTIVEC_ABI ? 16 : 8))
5567 {
5568 static bool warned_for_pass_big_vectors = false;
5569 if (TARGET_DEBUG_ARG)
5570 fprintf (stderr, "function_arg_pass_by_reference: synthetic vector\n");
5571 if (!warned_for_pass_big_vectors)
5572 {
5573 warning (0, "GCC vector passed by reference: "
5574 "non-standard ABI extension with no compatibility guarantee");
5575 warned_for_pass_big_vectors = true;
5576 }
5577 return 1;
5578 }
5579
5580 return 0;
5581 }
5582
5583 static void
rs6000_move_block_from_reg(int regno,rtx x,int nregs)5584 rs6000_move_block_from_reg (int regno, rtx x, int nregs)
5585 {
5586 int i;
5587 enum machine_mode reg_mode = TARGET_32BIT ? SImode : DImode;
5588
5589 if (nregs == 0)
5590 return;
5591
5592 for (i = 0; i < nregs; i++)
5593 {
5594 rtx tem = adjust_address_nv (x, reg_mode, i * GET_MODE_SIZE (reg_mode));
5595 if (reload_completed)
5596 {
5597 if (! strict_memory_address_p (reg_mode, XEXP (tem, 0)))
5598 tem = NULL_RTX;
5599 else
5600 tem = simplify_gen_subreg (reg_mode, x, BLKmode,
5601 i * GET_MODE_SIZE (reg_mode));
5602 }
5603 else
5604 tem = replace_equiv_address (tem, XEXP (tem, 0));
5605
5606 gcc_assert (tem);
5607
5608 emit_move_insn (tem, gen_rtx_REG (reg_mode, regno + i));
5609 }
5610 }
5611
5612 /* Perform any needed actions needed for a function that is receiving a
5613 variable number of arguments.
5614
5615 CUM is as above.
5616
5617 MODE and TYPE are the mode and type of the current parameter.
5618
5619 PRETEND_SIZE is a variable that should be set to the amount of stack
5620 that must be pushed by the prolog to pretend that our caller pushed
5621 it.
5622
5623 Normally, this macro will push all remaining incoming registers on the
5624 stack and set PRETEND_SIZE to the length of the registers pushed. */
5625
5626 static void
setup_incoming_varargs(CUMULATIVE_ARGS * cum,enum machine_mode mode,tree type,int * pretend_size ATTRIBUTE_UNUSED,int no_rtl)5627 setup_incoming_varargs (CUMULATIVE_ARGS *cum, enum machine_mode mode,
5628 tree type, int *pretend_size ATTRIBUTE_UNUSED,
5629 int no_rtl)
5630 {
5631 CUMULATIVE_ARGS next_cum;
5632 int reg_size = TARGET_32BIT ? 4 : 8;
5633 rtx save_area = NULL_RTX, mem;
5634 int first_reg_offset, set;
5635
5636 /* Skip the last named argument. */
5637 next_cum = *cum;
5638 function_arg_advance (&next_cum, mode, type, 1, 0);
5639
5640 if (DEFAULT_ABI == ABI_V4)
5641 {
5642 first_reg_offset = next_cum.sysv_gregno - GP_ARG_MIN_REG;
5643
5644 if (! no_rtl)
5645 {
5646 int gpr_reg_num = 0, gpr_size = 0, fpr_size = 0;
5647 HOST_WIDE_INT offset = 0;
5648
5649 /* Try to optimize the size of the varargs save area.
5650 The ABI requires that ap.reg_save_area is doubleword
5651 aligned, but we don't need to allocate space for all
5652 the bytes, only those to which we actually will save
5653 anything. */
5654 if (cfun->va_list_gpr_size && first_reg_offset < GP_ARG_NUM_REG)
5655 gpr_reg_num = GP_ARG_NUM_REG - first_reg_offset;
5656 if (TARGET_HARD_FLOAT && TARGET_FPRS
5657 && next_cum.fregno <= FP_ARG_V4_MAX_REG
5658 && cfun->va_list_fpr_size)
5659 {
5660 if (gpr_reg_num)
5661 fpr_size = (next_cum.fregno - FP_ARG_MIN_REG)
5662 * UNITS_PER_FP_WORD;
5663 if (cfun->va_list_fpr_size
5664 < FP_ARG_V4_MAX_REG + 1 - next_cum.fregno)
5665 fpr_size += cfun->va_list_fpr_size * UNITS_PER_FP_WORD;
5666 else
5667 fpr_size += (FP_ARG_V4_MAX_REG + 1 - next_cum.fregno)
5668 * UNITS_PER_FP_WORD;
5669 }
5670 if (gpr_reg_num)
5671 {
5672 offset = -((first_reg_offset * reg_size) & ~7);
5673 if (!fpr_size && gpr_reg_num > cfun->va_list_gpr_size)
5674 {
5675 gpr_reg_num = cfun->va_list_gpr_size;
5676 if (reg_size == 4 && (first_reg_offset & 1))
5677 gpr_reg_num++;
5678 }
5679 gpr_size = (gpr_reg_num * reg_size + 7) & ~7;
5680 }
5681 else if (fpr_size)
5682 offset = - (int) (next_cum.fregno - FP_ARG_MIN_REG)
5683 * UNITS_PER_FP_WORD
5684 - (int) (GP_ARG_NUM_REG * reg_size);
5685
5686 if (gpr_size + fpr_size)
5687 {
5688 rtx reg_save_area
5689 = assign_stack_local (BLKmode, gpr_size + fpr_size, 64);
5690 gcc_assert (GET_CODE (reg_save_area) == MEM);
5691 reg_save_area = XEXP (reg_save_area, 0);
5692 if (GET_CODE (reg_save_area) == PLUS)
5693 {
5694 gcc_assert (XEXP (reg_save_area, 0)
5695 == virtual_stack_vars_rtx);
5696 gcc_assert (GET_CODE (XEXP (reg_save_area, 1)) == CONST_INT);
5697 offset += INTVAL (XEXP (reg_save_area, 1));
5698 }
5699 else
5700 gcc_assert (reg_save_area == virtual_stack_vars_rtx);
5701 }
5702
5703 cfun->machine->varargs_save_offset = offset;
5704 save_area = plus_constant (virtual_stack_vars_rtx, offset);
5705 }
5706 }
5707 else
5708 {
5709 first_reg_offset = next_cum.words;
5710 save_area = virtual_incoming_args_rtx;
5711
5712 if (targetm.calls.must_pass_in_stack (mode, type))
5713 first_reg_offset += rs6000_arg_size (TYPE_MODE (type), type);
5714 }
5715
5716 set = get_varargs_alias_set ();
5717 if (! no_rtl && first_reg_offset < GP_ARG_NUM_REG
5718 && cfun->va_list_gpr_size)
5719 {
5720 int nregs = GP_ARG_NUM_REG - first_reg_offset;
5721
5722 if (va_list_gpr_counter_field)
5723 {
5724 /* V4 va_list_gpr_size counts number of registers needed. */
5725 if (nregs > cfun->va_list_gpr_size)
5726 nregs = cfun->va_list_gpr_size;
5727 }
5728 else
5729 {
5730 /* char * va_list instead counts number of bytes needed. */
5731 if (nregs > cfun->va_list_gpr_size / reg_size)
5732 nregs = cfun->va_list_gpr_size / reg_size;
5733 }
5734
5735 mem = gen_rtx_MEM (BLKmode,
5736 plus_constant (save_area,
5737 first_reg_offset * reg_size));
5738 MEM_NOTRAP_P (mem) = 1;
5739 set_mem_alias_set (mem, set);
5740 set_mem_align (mem, BITS_PER_WORD);
5741
5742 rs6000_move_block_from_reg (GP_ARG_MIN_REG + first_reg_offset, mem,
5743 nregs);
5744 }
5745
5746 /* Save FP registers if needed. */
5747 if (DEFAULT_ABI == ABI_V4
5748 && TARGET_HARD_FLOAT && TARGET_FPRS
5749 && ! no_rtl
5750 && next_cum.fregno <= FP_ARG_V4_MAX_REG
5751 && cfun->va_list_fpr_size)
5752 {
5753 int fregno = next_cum.fregno, nregs;
5754 rtx cr1 = gen_rtx_REG (CCmode, CR1_REGNO);
5755 rtx lab = gen_label_rtx ();
5756 int off = (GP_ARG_NUM_REG * reg_size) + ((fregno - FP_ARG_MIN_REG)
5757 * UNITS_PER_FP_WORD);
5758
5759 emit_jump_insn
5760 (gen_rtx_SET (VOIDmode,
5761 pc_rtx,
5762 gen_rtx_IF_THEN_ELSE (VOIDmode,
5763 gen_rtx_NE (VOIDmode, cr1,
5764 const0_rtx),
5765 gen_rtx_LABEL_REF (VOIDmode, lab),
5766 pc_rtx)));
5767
5768 for (nregs = 0;
5769 fregno <= FP_ARG_V4_MAX_REG && nregs < cfun->va_list_fpr_size;
5770 fregno++, off += UNITS_PER_FP_WORD, nregs++)
5771 {
5772 mem = gen_rtx_MEM (DFmode, plus_constant (save_area, off));
5773 MEM_NOTRAP_P (mem) = 1;
5774 set_mem_alias_set (mem, set);
5775 set_mem_align (mem, GET_MODE_ALIGNMENT (DFmode));
5776 emit_move_insn (mem, gen_rtx_REG (DFmode, fregno));
5777 }
5778
5779 emit_label (lab);
5780 }
5781 }
5782
5783 /* Create the va_list data type. */
5784
5785 static tree
rs6000_build_builtin_va_list(void)5786 rs6000_build_builtin_va_list (void)
5787 {
5788 tree f_gpr, f_fpr, f_res, f_ovf, f_sav, record, type_decl;
5789
5790 /* For AIX, prefer 'char *' because that's what the system
5791 header files like. */
5792 if (DEFAULT_ABI != ABI_V4)
5793 return build_pointer_type (char_type_node);
5794
5795 record = (*lang_hooks.types.make_type) (RECORD_TYPE);
5796 type_decl = build_decl (TYPE_DECL, get_identifier ("__va_list_tag"), record);
5797
5798 f_gpr = build_decl (FIELD_DECL, get_identifier ("gpr"),
5799 unsigned_char_type_node);
5800 f_fpr = build_decl (FIELD_DECL, get_identifier ("fpr"),
5801 unsigned_char_type_node);
5802 /* Give the two bytes of padding a name, so that -Wpadded won't warn on
5803 every user file. */
5804 f_res = build_decl (FIELD_DECL, get_identifier ("reserved"),
5805 short_unsigned_type_node);
5806 f_ovf = build_decl (FIELD_DECL, get_identifier ("overflow_arg_area"),
5807 ptr_type_node);
5808 f_sav = build_decl (FIELD_DECL, get_identifier ("reg_save_area"),
5809 ptr_type_node);
5810
5811 va_list_gpr_counter_field = f_gpr;
5812 va_list_fpr_counter_field = f_fpr;
5813
5814 DECL_FIELD_CONTEXT (f_gpr) = record;
5815 DECL_FIELD_CONTEXT (f_fpr) = record;
5816 DECL_FIELD_CONTEXT (f_res) = record;
5817 DECL_FIELD_CONTEXT (f_ovf) = record;
5818 DECL_FIELD_CONTEXT (f_sav) = record;
5819
5820 TREE_CHAIN (record) = type_decl;
5821 TYPE_NAME (record) = type_decl;
5822 TYPE_FIELDS (record) = f_gpr;
5823 TREE_CHAIN (f_gpr) = f_fpr;
5824 TREE_CHAIN (f_fpr) = f_res;
5825 TREE_CHAIN (f_res) = f_ovf;
5826 TREE_CHAIN (f_ovf) = f_sav;
5827
5828 layout_type (record);
5829
5830 /* The correct type is an array type of one element. */
5831 return build_array_type (record, build_index_type (size_zero_node));
5832 }
5833
5834 /* Implement va_start. */
5835
5836 void
rs6000_va_start(tree valist,rtx nextarg)5837 rs6000_va_start (tree valist, rtx nextarg)
5838 {
5839 HOST_WIDE_INT words, n_gpr, n_fpr;
5840 tree f_gpr, f_fpr, f_res, f_ovf, f_sav;
5841 tree gpr, fpr, ovf, sav, t;
5842
5843 /* Only SVR4 needs something special. */
5844 if (DEFAULT_ABI != ABI_V4)
5845 {
5846 std_expand_builtin_va_start (valist, nextarg);
5847 return;
5848 }
5849
5850 f_gpr = TYPE_FIELDS (TREE_TYPE (va_list_type_node));
5851 f_fpr = TREE_CHAIN (f_gpr);
5852 f_res = TREE_CHAIN (f_fpr);
5853 f_ovf = TREE_CHAIN (f_res);
5854 f_sav = TREE_CHAIN (f_ovf);
5855
5856 valist = build_va_arg_indirect_ref (valist);
5857 gpr = build3 (COMPONENT_REF, TREE_TYPE (f_gpr), valist, f_gpr, NULL_TREE);
5858 fpr = build3 (COMPONENT_REF, TREE_TYPE (f_fpr), valist, f_fpr, NULL_TREE);
5859 ovf = build3 (COMPONENT_REF, TREE_TYPE (f_ovf), valist, f_ovf, NULL_TREE);
5860 sav = build3 (COMPONENT_REF, TREE_TYPE (f_sav), valist, f_sav, NULL_TREE);
5861
5862 /* Count number of gp and fp argument registers used. */
5863 words = current_function_args_info.words;
5864 n_gpr = MIN (current_function_args_info.sysv_gregno - GP_ARG_MIN_REG,
5865 GP_ARG_NUM_REG);
5866 n_fpr = MIN (current_function_args_info.fregno - FP_ARG_MIN_REG,
5867 FP_ARG_NUM_REG);
5868
5869 if (TARGET_DEBUG_ARG)
5870 fprintf (stderr, "va_start: words = "HOST_WIDE_INT_PRINT_DEC", n_gpr = "
5871 HOST_WIDE_INT_PRINT_DEC", n_fpr = "HOST_WIDE_INT_PRINT_DEC"\n",
5872 words, n_gpr, n_fpr);
5873
5874 if (cfun->va_list_gpr_size)
5875 {
5876 t = build2 (MODIFY_EXPR, TREE_TYPE (gpr), gpr,
5877 build_int_cst (NULL_TREE, n_gpr));
5878 TREE_SIDE_EFFECTS (t) = 1;
5879 expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL);
5880 }
5881
5882 if (cfun->va_list_fpr_size)
5883 {
5884 t = build2 (MODIFY_EXPR, TREE_TYPE (fpr), fpr,
5885 build_int_cst (NULL_TREE, n_fpr));
5886 TREE_SIDE_EFFECTS (t) = 1;
5887 expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL);
5888 }
5889
5890 /* Find the overflow area. */
5891 t = make_tree (TREE_TYPE (ovf), virtual_incoming_args_rtx);
5892 if (words != 0)
5893 t = build2 (PLUS_EXPR, TREE_TYPE (ovf), t,
5894 build_int_cst (NULL_TREE, words * UNITS_PER_WORD));
5895 t = build2 (MODIFY_EXPR, TREE_TYPE (ovf), ovf, t);
5896 TREE_SIDE_EFFECTS (t) = 1;
5897 expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL);
5898
5899 /* If there were no va_arg invocations, don't set up the register
5900 save area. */
5901 if (!cfun->va_list_gpr_size
5902 && !cfun->va_list_fpr_size
5903 && n_gpr < GP_ARG_NUM_REG
5904 && n_fpr < FP_ARG_V4_MAX_REG)
5905 return;
5906
5907 /* Find the register save area. */
5908 t = make_tree (TREE_TYPE (sav), virtual_stack_vars_rtx);
5909 if (cfun->machine->varargs_save_offset)
5910 t = build2 (PLUS_EXPR, TREE_TYPE (sav), t,
5911 build_int_cst (NULL_TREE, cfun->machine->varargs_save_offset));
5912 t = build2 (MODIFY_EXPR, TREE_TYPE (sav), sav, t);
5913 TREE_SIDE_EFFECTS (t) = 1;
5914 expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL);
5915 }
5916
5917 /* Implement va_arg. */
5918
5919 tree
rs6000_gimplify_va_arg(tree valist,tree type,tree * pre_p,tree * post_p)5920 rs6000_gimplify_va_arg (tree valist, tree type, tree *pre_p, tree *post_p)
5921 {
5922 tree f_gpr, f_fpr, f_res, f_ovf, f_sav;
5923 tree gpr, fpr, ovf, sav, reg, t, u;
5924 int size, rsize, n_reg, sav_ofs, sav_scale;
5925 tree lab_false, lab_over, addr;
5926 int align;
5927 tree ptrtype = build_pointer_type (type);
5928
5929 if (pass_by_reference (NULL, TYPE_MODE (type), type, false))
5930 {
5931 t = rs6000_gimplify_va_arg (valist, ptrtype, pre_p, post_p);
5932 return build_va_arg_indirect_ref (t);
5933 }
5934
5935 if (DEFAULT_ABI != ABI_V4)
5936 {
5937 if (targetm.calls.split_complex_arg && TREE_CODE (type) == COMPLEX_TYPE)
5938 {
5939 tree elem_type = TREE_TYPE (type);
5940 enum machine_mode elem_mode = TYPE_MODE (elem_type);
5941 int elem_size = GET_MODE_SIZE (elem_mode);
5942
5943 if (elem_size < UNITS_PER_WORD)
5944 {
5945 tree real_part, imag_part;
5946 tree post = NULL_TREE;
5947
5948 real_part = rs6000_gimplify_va_arg (valist, elem_type, pre_p,
5949 &post);
5950 /* Copy the value into a temporary, lest the formal temporary
5951 be reused out from under us. */
5952 real_part = get_initialized_tmp_var (real_part, pre_p, &post);
5953 append_to_statement_list (post, pre_p);
5954
5955 imag_part = rs6000_gimplify_va_arg (valist, elem_type, pre_p,
5956 post_p);
5957
5958 return build2 (COMPLEX_EXPR, type, real_part, imag_part);
5959 }
5960 }
5961
5962 return std_gimplify_va_arg_expr (valist, type, pre_p, post_p);
5963 }
5964
5965 f_gpr = TYPE_FIELDS (TREE_TYPE (va_list_type_node));
5966 f_fpr = TREE_CHAIN (f_gpr);
5967 f_res = TREE_CHAIN (f_fpr);
5968 f_ovf = TREE_CHAIN (f_res);
5969 f_sav = TREE_CHAIN (f_ovf);
5970
5971 valist = build_va_arg_indirect_ref (valist);
5972 gpr = build3 (COMPONENT_REF, TREE_TYPE (f_gpr), valist, f_gpr, NULL_TREE);
5973 fpr = build3 (COMPONENT_REF, TREE_TYPE (f_fpr), valist, f_fpr, NULL_TREE);
5974 ovf = build3 (COMPONENT_REF, TREE_TYPE (f_ovf), valist, f_ovf, NULL_TREE);
5975 sav = build3 (COMPONENT_REF, TREE_TYPE (f_sav), valist, f_sav, NULL_TREE);
5976
5977 size = int_size_in_bytes (type);
5978 rsize = (size + 3) / 4;
5979 align = 1;
5980
5981 if (TARGET_HARD_FLOAT && TARGET_FPRS
5982 && (TYPE_MODE (type) == SFmode
5983 || TYPE_MODE (type) == DFmode
5984 || TYPE_MODE (type) == TFmode))
5985 {
5986 /* FP args go in FP registers, if present. */
5987 reg = fpr;
5988 n_reg = (size + 7) / 8;
5989 sav_ofs = 8*4;
5990 sav_scale = 8;
5991 if (TYPE_MODE (type) != SFmode)
5992 align = 8;
5993 }
5994 else
5995 {
5996 /* Otherwise into GP registers. */
5997 reg = gpr;
5998 n_reg = rsize;
5999 sav_ofs = 0;
6000 sav_scale = 4;
6001 if (n_reg == 2)
6002 align = 8;
6003 }
6004
6005 /* Pull the value out of the saved registers.... */
6006
6007 lab_over = NULL;
6008 addr = create_tmp_var (ptr_type_node, "addr");
6009 DECL_POINTER_ALIAS_SET (addr) = get_varargs_alias_set ();
6010
6011 /* AltiVec vectors never go in registers when -mabi=altivec. */
6012 if (TARGET_ALTIVEC_ABI && ALTIVEC_VECTOR_MODE (TYPE_MODE (type)))
6013 align = 16;
6014 else
6015 {
6016 lab_false = create_artificial_label ();
6017 lab_over = create_artificial_label ();
6018
6019 /* Long long and SPE vectors are aligned in the registers.
6020 As are any other 2 gpr item such as complex int due to a
6021 historical mistake. */
6022 u = reg;
6023 if (n_reg == 2 && reg == gpr)
6024 {
6025 u = build2 (BIT_AND_EXPR, TREE_TYPE (reg), reg,
6026 size_int (n_reg - 1));
6027 u = build2 (POSTINCREMENT_EXPR, TREE_TYPE (reg), reg, u);
6028 }
6029
6030 t = fold_convert (TREE_TYPE (reg), size_int (8 - n_reg + 1));
6031 t = build2 (GE_EXPR, boolean_type_node, u, t);
6032 u = build1 (GOTO_EXPR, void_type_node, lab_false);
6033 t = build3 (COND_EXPR, void_type_node, t, u, NULL_TREE);
6034 gimplify_and_add (t, pre_p);
6035
6036 t = sav;
6037 if (sav_ofs)
6038 t = build2 (PLUS_EXPR, ptr_type_node, sav, size_int (sav_ofs));
6039
6040 u = build2 (POSTINCREMENT_EXPR, TREE_TYPE (reg), reg, size_int (n_reg));
6041 u = build1 (CONVERT_EXPR, integer_type_node, u);
6042 u = build2 (MULT_EXPR, integer_type_node, u, size_int (sav_scale));
6043 t = build2 (PLUS_EXPR, ptr_type_node, t, u);
6044
6045 t = build2 (MODIFY_EXPR, void_type_node, addr, t);
6046 gimplify_and_add (t, pre_p);
6047
6048 t = build1 (GOTO_EXPR, void_type_node, lab_over);
6049 gimplify_and_add (t, pre_p);
6050
6051 t = build1 (LABEL_EXPR, void_type_node, lab_false);
6052 append_to_statement_list (t, pre_p);
6053
6054 if ((n_reg == 2 && reg != gpr) || n_reg > 2)
6055 {
6056 /* Ensure that we don't find any more args in regs.
6057 Alignment has taken care of the n_reg == 2 gpr case. */
6058 t = build2 (MODIFY_EXPR, TREE_TYPE (reg), reg, size_int (8));
6059 gimplify_and_add (t, pre_p);
6060 }
6061 }
6062
6063 /* ... otherwise out of the overflow area. */
6064
6065 /* Care for on-stack alignment if needed. */
6066 t = ovf;
6067 if (align != 1)
6068 {
6069 t = build2 (PLUS_EXPR, TREE_TYPE (t), t, size_int (align - 1));
6070 t = build2 (BIT_AND_EXPR, TREE_TYPE (t), t,
6071 build_int_cst (NULL_TREE, -align));
6072 }
6073 gimplify_expr (&t, pre_p, NULL, is_gimple_val, fb_rvalue);
6074
6075 u = build2 (MODIFY_EXPR, void_type_node, addr, t);
6076 gimplify_and_add (u, pre_p);
6077
6078 t = build2 (PLUS_EXPR, TREE_TYPE (t), t, size_int (size));
6079 t = build2 (MODIFY_EXPR, TREE_TYPE (ovf), ovf, t);
6080 gimplify_and_add (t, pre_p);
6081
6082 if (lab_over)
6083 {
6084 t = build1 (LABEL_EXPR, void_type_node, lab_over);
6085 append_to_statement_list (t, pre_p);
6086 }
6087
6088 if (STRICT_ALIGNMENT
6089 && (TYPE_ALIGN (type)
6090 > (unsigned) BITS_PER_UNIT * (align < 4 ? 4 : align)))
6091 {
6092 /* The value (of type complex double, for example) may not be
6093 aligned in memory in the saved registers, so copy via a
6094 temporary. (This is the same code as used for SPARC.) */
6095 tree tmp = create_tmp_var (type, "va_arg_tmp");
6096 tree dest_addr = build_fold_addr_expr (tmp);
6097
6098 tree copy = build_function_call_expr
6099 (implicit_built_in_decls[BUILT_IN_MEMCPY],
6100 tree_cons (NULL_TREE, dest_addr,
6101 tree_cons (NULL_TREE, addr,
6102 tree_cons (NULL_TREE, size_int (rsize * 4),
6103 NULL_TREE))));
6104
6105 gimplify_and_add (copy, pre_p);
6106 addr = dest_addr;
6107 }
6108
6109 addr = fold_convert (ptrtype, addr);
6110 return build_va_arg_indirect_ref (addr);
6111 }
6112
6113 /* Builtins. */
6114
6115 static void
def_builtin(int mask,const char * name,tree type,int code)6116 def_builtin (int mask, const char *name, tree type, int code)
6117 {
6118 if (mask & target_flags)
6119 {
6120 if (rs6000_builtin_decls[code])
6121 abort ();
6122
6123 rs6000_builtin_decls[code] =
6124 lang_hooks.builtin_function (name, type, code, BUILT_IN_MD,
6125 NULL, NULL_TREE);
6126 }
6127 }
6128
6129 /* Simple ternary operations: VECd = foo (VECa, VECb, VECc). */
6130
6131 static const struct builtin_description bdesc_3arg[] =
6132 {
6133 { MASK_ALTIVEC, CODE_FOR_altivec_vmaddfp, "__builtin_altivec_vmaddfp", ALTIVEC_BUILTIN_VMADDFP },
6134 { MASK_ALTIVEC, CODE_FOR_altivec_vmhaddshs, "__builtin_altivec_vmhaddshs", ALTIVEC_BUILTIN_VMHADDSHS },
6135 { MASK_ALTIVEC, CODE_FOR_altivec_vmhraddshs, "__builtin_altivec_vmhraddshs", ALTIVEC_BUILTIN_VMHRADDSHS },
6136 { MASK_ALTIVEC, CODE_FOR_altivec_vmladduhm, "__builtin_altivec_vmladduhm", ALTIVEC_BUILTIN_VMLADDUHM},
6137 { MASK_ALTIVEC, CODE_FOR_altivec_vmsumubm, "__builtin_altivec_vmsumubm", ALTIVEC_BUILTIN_VMSUMUBM },
6138 { MASK_ALTIVEC, CODE_FOR_altivec_vmsummbm, "__builtin_altivec_vmsummbm", ALTIVEC_BUILTIN_VMSUMMBM },
6139 { MASK_ALTIVEC, CODE_FOR_altivec_vmsumuhm, "__builtin_altivec_vmsumuhm", ALTIVEC_BUILTIN_VMSUMUHM },
6140 { MASK_ALTIVEC, CODE_FOR_altivec_vmsumshm, "__builtin_altivec_vmsumshm", ALTIVEC_BUILTIN_VMSUMSHM },
6141 { MASK_ALTIVEC, CODE_FOR_altivec_vmsumuhs, "__builtin_altivec_vmsumuhs", ALTIVEC_BUILTIN_VMSUMUHS },
6142 { MASK_ALTIVEC, CODE_FOR_altivec_vmsumshs, "__builtin_altivec_vmsumshs", ALTIVEC_BUILTIN_VMSUMSHS },
6143 { MASK_ALTIVEC, CODE_FOR_altivec_vnmsubfp, "__builtin_altivec_vnmsubfp", ALTIVEC_BUILTIN_VNMSUBFP },
6144 { MASK_ALTIVEC, CODE_FOR_altivec_vperm_v4sf, "__builtin_altivec_vperm_4sf", ALTIVEC_BUILTIN_VPERM_4SF },
6145 { MASK_ALTIVEC, CODE_FOR_altivec_vperm_v4si, "__builtin_altivec_vperm_4si", ALTIVEC_BUILTIN_VPERM_4SI },
6146 { MASK_ALTIVEC, CODE_FOR_altivec_vperm_v8hi, "__builtin_altivec_vperm_8hi", ALTIVEC_BUILTIN_VPERM_8HI },
6147 { MASK_ALTIVEC, CODE_FOR_altivec_vperm_v16qi, "__builtin_altivec_vperm_16qi", ALTIVEC_BUILTIN_VPERM_16QI },
6148 { MASK_ALTIVEC, CODE_FOR_altivec_vsel_v4sf, "__builtin_altivec_vsel_4sf", ALTIVEC_BUILTIN_VSEL_4SF },
6149 { MASK_ALTIVEC, CODE_FOR_altivec_vsel_v4si, "__builtin_altivec_vsel_4si", ALTIVEC_BUILTIN_VSEL_4SI },
6150 { MASK_ALTIVEC, CODE_FOR_altivec_vsel_v8hi, "__builtin_altivec_vsel_8hi", ALTIVEC_BUILTIN_VSEL_8HI },
6151 { MASK_ALTIVEC, CODE_FOR_altivec_vsel_v16qi, "__builtin_altivec_vsel_16qi", ALTIVEC_BUILTIN_VSEL_16QI },
6152 { MASK_ALTIVEC, CODE_FOR_altivec_vsldoi_v16qi, "__builtin_altivec_vsldoi_16qi", ALTIVEC_BUILTIN_VSLDOI_16QI },
6153 { MASK_ALTIVEC, CODE_FOR_altivec_vsldoi_v8hi, "__builtin_altivec_vsldoi_8hi", ALTIVEC_BUILTIN_VSLDOI_8HI },
6154 { MASK_ALTIVEC, CODE_FOR_altivec_vsldoi_v4si, "__builtin_altivec_vsldoi_4si", ALTIVEC_BUILTIN_VSLDOI_4SI },
6155 { MASK_ALTIVEC, CODE_FOR_altivec_vsldoi_v4sf, "__builtin_altivec_vsldoi_4sf", ALTIVEC_BUILTIN_VSLDOI_4SF },
6156
6157 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_madd", ALTIVEC_BUILTIN_VEC_MADD },
6158 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_madds", ALTIVEC_BUILTIN_VEC_MADDS },
6159 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_mladd", ALTIVEC_BUILTIN_VEC_MLADD },
6160 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_mradds", ALTIVEC_BUILTIN_VEC_MRADDS },
6161 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_msum", ALTIVEC_BUILTIN_VEC_MSUM },
6162 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vmsumshm", ALTIVEC_BUILTIN_VEC_VMSUMSHM },
6163 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vmsumuhm", ALTIVEC_BUILTIN_VEC_VMSUMUHM },
6164 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vmsummbm", ALTIVEC_BUILTIN_VEC_VMSUMMBM },
6165 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vmsumubm", ALTIVEC_BUILTIN_VEC_VMSUMUBM },
6166 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_msums", ALTIVEC_BUILTIN_VEC_MSUMS },
6167 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vmsumshs", ALTIVEC_BUILTIN_VEC_VMSUMSHS },
6168 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vmsumuhs", ALTIVEC_BUILTIN_VEC_VMSUMUHS },
6169 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_nmsub", ALTIVEC_BUILTIN_VEC_NMSUB },
6170 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_perm", ALTIVEC_BUILTIN_VEC_PERM },
6171 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_sel", ALTIVEC_BUILTIN_VEC_SEL },
6172 };
6173
6174 /* DST operations: void foo (void *, const int, const char). */
6175
6176 static const struct builtin_description bdesc_dst[] =
6177 {
6178 { MASK_ALTIVEC, CODE_FOR_altivec_dst, "__builtin_altivec_dst", ALTIVEC_BUILTIN_DST },
6179 { MASK_ALTIVEC, CODE_FOR_altivec_dstt, "__builtin_altivec_dstt", ALTIVEC_BUILTIN_DSTT },
6180 { MASK_ALTIVEC, CODE_FOR_altivec_dstst, "__builtin_altivec_dstst", ALTIVEC_BUILTIN_DSTST },
6181 { MASK_ALTIVEC, CODE_FOR_altivec_dststt, "__builtin_altivec_dststt", ALTIVEC_BUILTIN_DSTSTT },
6182
6183 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_dst", ALTIVEC_BUILTIN_VEC_DST },
6184 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_dstt", ALTIVEC_BUILTIN_VEC_DSTT },
6185 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_dstst", ALTIVEC_BUILTIN_VEC_DSTST },
6186 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_dststt", ALTIVEC_BUILTIN_VEC_DSTSTT }
6187 };
6188
6189 /* Simple binary operations: VECc = foo (VECa, VECb). */
6190
6191 static struct builtin_description bdesc_2arg[] =
6192 {
6193 { MASK_ALTIVEC, CODE_FOR_addv16qi3, "__builtin_altivec_vaddubm", ALTIVEC_BUILTIN_VADDUBM },
6194 { MASK_ALTIVEC, CODE_FOR_addv8hi3, "__builtin_altivec_vadduhm", ALTIVEC_BUILTIN_VADDUHM },
6195 { MASK_ALTIVEC, CODE_FOR_addv4si3, "__builtin_altivec_vadduwm", ALTIVEC_BUILTIN_VADDUWM },
6196 { MASK_ALTIVEC, CODE_FOR_addv4sf3, "__builtin_altivec_vaddfp", ALTIVEC_BUILTIN_VADDFP },
6197 { MASK_ALTIVEC, CODE_FOR_altivec_vaddcuw, "__builtin_altivec_vaddcuw", ALTIVEC_BUILTIN_VADDCUW },
6198 { MASK_ALTIVEC, CODE_FOR_altivec_vaddubs, "__builtin_altivec_vaddubs", ALTIVEC_BUILTIN_VADDUBS },
6199 { MASK_ALTIVEC, CODE_FOR_altivec_vaddsbs, "__builtin_altivec_vaddsbs", ALTIVEC_BUILTIN_VADDSBS },
6200 { MASK_ALTIVEC, CODE_FOR_altivec_vadduhs, "__builtin_altivec_vadduhs", ALTIVEC_BUILTIN_VADDUHS },
6201 { MASK_ALTIVEC, CODE_FOR_altivec_vaddshs, "__builtin_altivec_vaddshs", ALTIVEC_BUILTIN_VADDSHS },
6202 { MASK_ALTIVEC, CODE_FOR_altivec_vadduws, "__builtin_altivec_vadduws", ALTIVEC_BUILTIN_VADDUWS },
6203 { MASK_ALTIVEC, CODE_FOR_altivec_vaddsws, "__builtin_altivec_vaddsws", ALTIVEC_BUILTIN_VADDSWS },
6204 { MASK_ALTIVEC, CODE_FOR_andv4si3, "__builtin_altivec_vand", ALTIVEC_BUILTIN_VAND },
6205 { MASK_ALTIVEC, CODE_FOR_andcv4si3, "__builtin_altivec_vandc", ALTIVEC_BUILTIN_VANDC },
6206 { MASK_ALTIVEC, CODE_FOR_altivec_vavgub, "__builtin_altivec_vavgub", ALTIVEC_BUILTIN_VAVGUB },
6207 { MASK_ALTIVEC, CODE_FOR_altivec_vavgsb, "__builtin_altivec_vavgsb", ALTIVEC_BUILTIN_VAVGSB },
6208 { MASK_ALTIVEC, CODE_FOR_altivec_vavguh, "__builtin_altivec_vavguh", ALTIVEC_BUILTIN_VAVGUH },
6209 { MASK_ALTIVEC, CODE_FOR_altivec_vavgsh, "__builtin_altivec_vavgsh", ALTIVEC_BUILTIN_VAVGSH },
6210 { MASK_ALTIVEC, CODE_FOR_altivec_vavguw, "__builtin_altivec_vavguw", ALTIVEC_BUILTIN_VAVGUW },
6211 { MASK_ALTIVEC, CODE_FOR_altivec_vavgsw, "__builtin_altivec_vavgsw", ALTIVEC_BUILTIN_VAVGSW },
6212 { MASK_ALTIVEC, CODE_FOR_altivec_vcfux, "__builtin_altivec_vcfux", ALTIVEC_BUILTIN_VCFUX },
6213 { MASK_ALTIVEC, CODE_FOR_altivec_vcfsx, "__builtin_altivec_vcfsx", ALTIVEC_BUILTIN_VCFSX },
6214 { MASK_ALTIVEC, CODE_FOR_altivec_vcmpbfp, "__builtin_altivec_vcmpbfp", ALTIVEC_BUILTIN_VCMPBFP },
6215 { MASK_ALTIVEC, CODE_FOR_altivec_vcmpequb, "__builtin_altivec_vcmpequb", ALTIVEC_BUILTIN_VCMPEQUB },
6216 { MASK_ALTIVEC, CODE_FOR_altivec_vcmpequh, "__builtin_altivec_vcmpequh", ALTIVEC_BUILTIN_VCMPEQUH },
6217 { MASK_ALTIVEC, CODE_FOR_altivec_vcmpequw, "__builtin_altivec_vcmpequw", ALTIVEC_BUILTIN_VCMPEQUW },
6218 { MASK_ALTIVEC, CODE_FOR_altivec_vcmpeqfp, "__builtin_altivec_vcmpeqfp", ALTIVEC_BUILTIN_VCMPEQFP },
6219 { MASK_ALTIVEC, CODE_FOR_altivec_vcmpgefp, "__builtin_altivec_vcmpgefp", ALTIVEC_BUILTIN_VCMPGEFP },
6220 { MASK_ALTIVEC, CODE_FOR_altivec_vcmpgtub, "__builtin_altivec_vcmpgtub", ALTIVEC_BUILTIN_VCMPGTUB },
6221 { MASK_ALTIVEC, CODE_FOR_altivec_vcmpgtsb, "__builtin_altivec_vcmpgtsb", ALTIVEC_BUILTIN_VCMPGTSB },
6222 { MASK_ALTIVEC, CODE_FOR_altivec_vcmpgtuh, "__builtin_altivec_vcmpgtuh", ALTIVEC_BUILTIN_VCMPGTUH },
6223 { MASK_ALTIVEC, CODE_FOR_altivec_vcmpgtsh, "__builtin_altivec_vcmpgtsh", ALTIVEC_BUILTIN_VCMPGTSH },
6224 { MASK_ALTIVEC, CODE_FOR_altivec_vcmpgtuw, "__builtin_altivec_vcmpgtuw", ALTIVEC_BUILTIN_VCMPGTUW },
6225 { MASK_ALTIVEC, CODE_FOR_altivec_vcmpgtsw, "__builtin_altivec_vcmpgtsw", ALTIVEC_BUILTIN_VCMPGTSW },
6226 { MASK_ALTIVEC, CODE_FOR_altivec_vcmpgtfp, "__builtin_altivec_vcmpgtfp", ALTIVEC_BUILTIN_VCMPGTFP },
6227 { MASK_ALTIVEC, CODE_FOR_altivec_vctsxs, "__builtin_altivec_vctsxs", ALTIVEC_BUILTIN_VCTSXS },
6228 { MASK_ALTIVEC, CODE_FOR_altivec_vctuxs, "__builtin_altivec_vctuxs", ALTIVEC_BUILTIN_VCTUXS },
6229 { MASK_ALTIVEC, CODE_FOR_umaxv16qi3, "__builtin_altivec_vmaxub", ALTIVEC_BUILTIN_VMAXUB },
6230 { MASK_ALTIVEC, CODE_FOR_smaxv16qi3, "__builtin_altivec_vmaxsb", ALTIVEC_BUILTIN_VMAXSB },
6231 { MASK_ALTIVEC, CODE_FOR_umaxv8hi3, "__builtin_altivec_vmaxuh", ALTIVEC_BUILTIN_VMAXUH },
6232 { MASK_ALTIVEC, CODE_FOR_smaxv8hi3, "__builtin_altivec_vmaxsh", ALTIVEC_BUILTIN_VMAXSH },
6233 { MASK_ALTIVEC, CODE_FOR_umaxv4si3, "__builtin_altivec_vmaxuw", ALTIVEC_BUILTIN_VMAXUW },
6234 { MASK_ALTIVEC, CODE_FOR_smaxv4si3, "__builtin_altivec_vmaxsw", ALTIVEC_BUILTIN_VMAXSW },
6235 { MASK_ALTIVEC, CODE_FOR_smaxv4sf3, "__builtin_altivec_vmaxfp", ALTIVEC_BUILTIN_VMAXFP },
6236 { MASK_ALTIVEC, CODE_FOR_altivec_vmrghb, "__builtin_altivec_vmrghb", ALTIVEC_BUILTIN_VMRGHB },
6237 { MASK_ALTIVEC, CODE_FOR_altivec_vmrghh, "__builtin_altivec_vmrghh", ALTIVEC_BUILTIN_VMRGHH },
6238 { MASK_ALTIVEC, CODE_FOR_altivec_vmrghw, "__builtin_altivec_vmrghw", ALTIVEC_BUILTIN_VMRGHW },
6239 { MASK_ALTIVEC, CODE_FOR_altivec_vmrglb, "__builtin_altivec_vmrglb", ALTIVEC_BUILTIN_VMRGLB },
6240 { MASK_ALTIVEC, CODE_FOR_altivec_vmrglh, "__builtin_altivec_vmrglh", ALTIVEC_BUILTIN_VMRGLH },
6241 { MASK_ALTIVEC, CODE_FOR_altivec_vmrglw, "__builtin_altivec_vmrglw", ALTIVEC_BUILTIN_VMRGLW },
6242 { MASK_ALTIVEC, CODE_FOR_uminv16qi3, "__builtin_altivec_vminub", ALTIVEC_BUILTIN_VMINUB },
6243 { MASK_ALTIVEC, CODE_FOR_sminv16qi3, "__builtin_altivec_vminsb", ALTIVEC_BUILTIN_VMINSB },
6244 { MASK_ALTIVEC, CODE_FOR_uminv8hi3, "__builtin_altivec_vminuh", ALTIVEC_BUILTIN_VMINUH },
6245 { MASK_ALTIVEC, CODE_FOR_sminv8hi3, "__builtin_altivec_vminsh", ALTIVEC_BUILTIN_VMINSH },
6246 { MASK_ALTIVEC, CODE_FOR_uminv4si3, "__builtin_altivec_vminuw", ALTIVEC_BUILTIN_VMINUW },
6247 { MASK_ALTIVEC, CODE_FOR_sminv4si3, "__builtin_altivec_vminsw", ALTIVEC_BUILTIN_VMINSW },
6248 { MASK_ALTIVEC, CODE_FOR_sminv4sf3, "__builtin_altivec_vminfp", ALTIVEC_BUILTIN_VMINFP },
6249 { MASK_ALTIVEC, CODE_FOR_altivec_vmuleub, "__builtin_altivec_vmuleub", ALTIVEC_BUILTIN_VMULEUB },
6250 { MASK_ALTIVEC, CODE_FOR_altivec_vmulesb, "__builtin_altivec_vmulesb", ALTIVEC_BUILTIN_VMULESB },
6251 { MASK_ALTIVEC, CODE_FOR_altivec_vmuleuh, "__builtin_altivec_vmuleuh", ALTIVEC_BUILTIN_VMULEUH },
6252 { MASK_ALTIVEC, CODE_FOR_altivec_vmulesh, "__builtin_altivec_vmulesh", ALTIVEC_BUILTIN_VMULESH },
6253 { MASK_ALTIVEC, CODE_FOR_altivec_vmuloub, "__builtin_altivec_vmuloub", ALTIVEC_BUILTIN_VMULOUB },
6254 { MASK_ALTIVEC, CODE_FOR_altivec_vmulosb, "__builtin_altivec_vmulosb", ALTIVEC_BUILTIN_VMULOSB },
6255 { MASK_ALTIVEC, CODE_FOR_altivec_vmulouh, "__builtin_altivec_vmulouh", ALTIVEC_BUILTIN_VMULOUH },
6256 { MASK_ALTIVEC, CODE_FOR_altivec_vmulosh, "__builtin_altivec_vmulosh", ALTIVEC_BUILTIN_VMULOSH },
6257 { MASK_ALTIVEC, CODE_FOR_altivec_norv4si3, "__builtin_altivec_vnor", ALTIVEC_BUILTIN_VNOR },
6258 { MASK_ALTIVEC, CODE_FOR_iorv4si3, "__builtin_altivec_vor", ALTIVEC_BUILTIN_VOR },
6259 { MASK_ALTIVEC, CODE_FOR_altivec_vpkuhum, "__builtin_altivec_vpkuhum", ALTIVEC_BUILTIN_VPKUHUM },
6260 { MASK_ALTIVEC, CODE_FOR_altivec_vpkuwum, "__builtin_altivec_vpkuwum", ALTIVEC_BUILTIN_VPKUWUM },
6261 { MASK_ALTIVEC, CODE_FOR_altivec_vpkpx, "__builtin_altivec_vpkpx", ALTIVEC_BUILTIN_VPKPX },
6262 { MASK_ALTIVEC, CODE_FOR_altivec_vpkshss, "__builtin_altivec_vpkshss", ALTIVEC_BUILTIN_VPKSHSS },
6263 { MASK_ALTIVEC, CODE_FOR_altivec_vpkswss, "__builtin_altivec_vpkswss", ALTIVEC_BUILTIN_VPKSWSS },
6264 { MASK_ALTIVEC, CODE_FOR_altivec_vpkuhus, "__builtin_altivec_vpkuhus", ALTIVEC_BUILTIN_VPKUHUS },
6265 { MASK_ALTIVEC, CODE_FOR_altivec_vpkshus, "__builtin_altivec_vpkshus", ALTIVEC_BUILTIN_VPKSHUS },
6266 { MASK_ALTIVEC, CODE_FOR_altivec_vpkuwus, "__builtin_altivec_vpkuwus", ALTIVEC_BUILTIN_VPKUWUS },
6267 { MASK_ALTIVEC, CODE_FOR_altivec_vpkswus, "__builtin_altivec_vpkswus", ALTIVEC_BUILTIN_VPKSWUS },
6268 { MASK_ALTIVEC, CODE_FOR_altivec_vrlb, "__builtin_altivec_vrlb", ALTIVEC_BUILTIN_VRLB },
6269 { MASK_ALTIVEC, CODE_FOR_altivec_vrlh, "__builtin_altivec_vrlh", ALTIVEC_BUILTIN_VRLH },
6270 { MASK_ALTIVEC, CODE_FOR_altivec_vrlw, "__builtin_altivec_vrlw", ALTIVEC_BUILTIN_VRLW },
6271 { MASK_ALTIVEC, CODE_FOR_altivec_vslb, "__builtin_altivec_vslb", ALTIVEC_BUILTIN_VSLB },
6272 { MASK_ALTIVEC, CODE_FOR_altivec_vslh, "__builtin_altivec_vslh", ALTIVEC_BUILTIN_VSLH },
6273 { MASK_ALTIVEC, CODE_FOR_altivec_vslw, "__builtin_altivec_vslw", ALTIVEC_BUILTIN_VSLW },
6274 { MASK_ALTIVEC, CODE_FOR_altivec_vsl, "__builtin_altivec_vsl", ALTIVEC_BUILTIN_VSL },
6275 { MASK_ALTIVEC, CODE_FOR_altivec_vslo, "__builtin_altivec_vslo", ALTIVEC_BUILTIN_VSLO },
6276 { MASK_ALTIVEC, CODE_FOR_altivec_vspltb, "__builtin_altivec_vspltb", ALTIVEC_BUILTIN_VSPLTB },
6277 { MASK_ALTIVEC, CODE_FOR_altivec_vsplth, "__builtin_altivec_vsplth", ALTIVEC_BUILTIN_VSPLTH },
6278 { MASK_ALTIVEC, CODE_FOR_altivec_vspltw, "__builtin_altivec_vspltw", ALTIVEC_BUILTIN_VSPLTW },
6279 { MASK_ALTIVEC, CODE_FOR_lshrv16qi3, "__builtin_altivec_vsrb", ALTIVEC_BUILTIN_VSRB },
6280 { MASK_ALTIVEC, CODE_FOR_lshrv8hi3, "__builtin_altivec_vsrh", ALTIVEC_BUILTIN_VSRH },
6281 { MASK_ALTIVEC, CODE_FOR_lshrv4si3, "__builtin_altivec_vsrw", ALTIVEC_BUILTIN_VSRW },
6282 { MASK_ALTIVEC, CODE_FOR_ashrv16qi3, "__builtin_altivec_vsrab", ALTIVEC_BUILTIN_VSRAB },
6283 { MASK_ALTIVEC, CODE_FOR_ashrv8hi3, "__builtin_altivec_vsrah", ALTIVEC_BUILTIN_VSRAH },
6284 { MASK_ALTIVEC, CODE_FOR_ashrv4si3, "__builtin_altivec_vsraw", ALTIVEC_BUILTIN_VSRAW },
6285 { MASK_ALTIVEC, CODE_FOR_altivec_vsr, "__builtin_altivec_vsr", ALTIVEC_BUILTIN_VSR },
6286 { MASK_ALTIVEC, CODE_FOR_altivec_vsro, "__builtin_altivec_vsro", ALTIVEC_BUILTIN_VSRO },
6287 { MASK_ALTIVEC, CODE_FOR_subv16qi3, "__builtin_altivec_vsububm", ALTIVEC_BUILTIN_VSUBUBM },
6288 { MASK_ALTIVEC, CODE_FOR_subv8hi3, "__builtin_altivec_vsubuhm", ALTIVEC_BUILTIN_VSUBUHM },
6289 { MASK_ALTIVEC, CODE_FOR_subv4si3, "__builtin_altivec_vsubuwm", ALTIVEC_BUILTIN_VSUBUWM },
6290 { MASK_ALTIVEC, CODE_FOR_subv4sf3, "__builtin_altivec_vsubfp", ALTIVEC_BUILTIN_VSUBFP },
6291 { MASK_ALTIVEC, CODE_FOR_altivec_vsubcuw, "__builtin_altivec_vsubcuw", ALTIVEC_BUILTIN_VSUBCUW },
6292 { MASK_ALTIVEC, CODE_FOR_altivec_vsububs, "__builtin_altivec_vsububs", ALTIVEC_BUILTIN_VSUBUBS },
6293 { MASK_ALTIVEC, CODE_FOR_altivec_vsubsbs, "__builtin_altivec_vsubsbs", ALTIVEC_BUILTIN_VSUBSBS },
6294 { MASK_ALTIVEC, CODE_FOR_altivec_vsubuhs, "__builtin_altivec_vsubuhs", ALTIVEC_BUILTIN_VSUBUHS },
6295 { MASK_ALTIVEC, CODE_FOR_altivec_vsubshs, "__builtin_altivec_vsubshs", ALTIVEC_BUILTIN_VSUBSHS },
6296 { MASK_ALTIVEC, CODE_FOR_altivec_vsubuws, "__builtin_altivec_vsubuws", ALTIVEC_BUILTIN_VSUBUWS },
6297 { MASK_ALTIVEC, CODE_FOR_altivec_vsubsws, "__builtin_altivec_vsubsws", ALTIVEC_BUILTIN_VSUBSWS },
6298 { MASK_ALTIVEC, CODE_FOR_altivec_vsum4ubs, "__builtin_altivec_vsum4ubs", ALTIVEC_BUILTIN_VSUM4UBS },
6299 { MASK_ALTIVEC, CODE_FOR_altivec_vsum4sbs, "__builtin_altivec_vsum4sbs", ALTIVEC_BUILTIN_VSUM4SBS },
6300 { MASK_ALTIVEC, CODE_FOR_altivec_vsum4shs, "__builtin_altivec_vsum4shs", ALTIVEC_BUILTIN_VSUM4SHS },
6301 { MASK_ALTIVEC, CODE_FOR_altivec_vsum2sws, "__builtin_altivec_vsum2sws", ALTIVEC_BUILTIN_VSUM2SWS },
6302 { MASK_ALTIVEC, CODE_FOR_altivec_vsumsws, "__builtin_altivec_vsumsws", ALTIVEC_BUILTIN_VSUMSWS },
6303 { MASK_ALTIVEC, CODE_FOR_xorv4si3, "__builtin_altivec_vxor", ALTIVEC_BUILTIN_VXOR },
6304
6305 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_add", ALTIVEC_BUILTIN_VEC_ADD },
6306 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vaddfp", ALTIVEC_BUILTIN_VEC_VADDFP },
6307 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vadduwm", ALTIVEC_BUILTIN_VEC_VADDUWM },
6308 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vadduhm", ALTIVEC_BUILTIN_VEC_VADDUHM },
6309 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vaddubm", ALTIVEC_BUILTIN_VEC_VADDUBM },
6310 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_addc", ALTIVEC_BUILTIN_VEC_ADDC },
6311 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_adds", ALTIVEC_BUILTIN_VEC_ADDS },
6312 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vaddsws", ALTIVEC_BUILTIN_VEC_VADDSWS },
6313 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vadduws", ALTIVEC_BUILTIN_VEC_VADDUWS },
6314 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vaddshs", ALTIVEC_BUILTIN_VEC_VADDSHS },
6315 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vadduhs", ALTIVEC_BUILTIN_VEC_VADDUHS },
6316 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vaddsbs", ALTIVEC_BUILTIN_VEC_VADDSBS },
6317 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vaddubs", ALTIVEC_BUILTIN_VEC_VADDUBS },
6318 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_and", ALTIVEC_BUILTIN_VEC_AND },
6319 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_andc", ALTIVEC_BUILTIN_VEC_ANDC },
6320 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_avg", ALTIVEC_BUILTIN_VEC_AVG },
6321 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vavgsw", ALTIVEC_BUILTIN_VEC_VAVGSW },
6322 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vavguw", ALTIVEC_BUILTIN_VEC_VAVGUW },
6323 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vavgsh", ALTIVEC_BUILTIN_VEC_VAVGSH },
6324 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vavguh", ALTIVEC_BUILTIN_VEC_VAVGUH },
6325 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vavgsb", ALTIVEC_BUILTIN_VEC_VAVGSB },
6326 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vavgub", ALTIVEC_BUILTIN_VEC_VAVGUB },
6327 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_cmpb", ALTIVEC_BUILTIN_VEC_CMPB },
6328 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_cmpeq", ALTIVEC_BUILTIN_VEC_CMPEQ },
6329 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vcmpeqfp", ALTIVEC_BUILTIN_VEC_VCMPEQFP },
6330 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vcmpequw", ALTIVEC_BUILTIN_VEC_VCMPEQUW },
6331 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vcmpequh", ALTIVEC_BUILTIN_VEC_VCMPEQUH },
6332 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vcmpequb", ALTIVEC_BUILTIN_VEC_VCMPEQUB },
6333 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_cmpge", ALTIVEC_BUILTIN_VEC_CMPGE },
6334 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_cmpgt", ALTIVEC_BUILTIN_VEC_CMPGT },
6335 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vcmpgtfp", ALTIVEC_BUILTIN_VEC_VCMPGTFP },
6336 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vcmpgtsw", ALTIVEC_BUILTIN_VEC_VCMPGTSW },
6337 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vcmpgtuw", ALTIVEC_BUILTIN_VEC_VCMPGTUW },
6338 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vcmpgtsh", ALTIVEC_BUILTIN_VEC_VCMPGTSH },
6339 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vcmpgtuh", ALTIVEC_BUILTIN_VEC_VCMPGTUH },
6340 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vcmpgtsb", ALTIVEC_BUILTIN_VEC_VCMPGTSB },
6341 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vcmpgtub", ALTIVEC_BUILTIN_VEC_VCMPGTUB },
6342 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_cmple", ALTIVEC_BUILTIN_VEC_CMPLE },
6343 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_cmplt", ALTIVEC_BUILTIN_VEC_CMPLT },
6344 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_max", ALTIVEC_BUILTIN_VEC_MAX },
6345 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vmaxfp", ALTIVEC_BUILTIN_VEC_VMAXFP },
6346 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vmaxsw", ALTIVEC_BUILTIN_VEC_VMAXSW },
6347 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vmaxuw", ALTIVEC_BUILTIN_VEC_VMAXUW },
6348 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vmaxsh", ALTIVEC_BUILTIN_VEC_VMAXSH },
6349 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vmaxuh", ALTIVEC_BUILTIN_VEC_VMAXUH },
6350 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vmaxsb", ALTIVEC_BUILTIN_VEC_VMAXSB },
6351 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vmaxub", ALTIVEC_BUILTIN_VEC_VMAXUB },
6352 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_mergeh", ALTIVEC_BUILTIN_VEC_MERGEH },
6353 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vmrghw", ALTIVEC_BUILTIN_VEC_VMRGHW },
6354 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vmrghh", ALTIVEC_BUILTIN_VEC_VMRGHH },
6355 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vmrghb", ALTIVEC_BUILTIN_VEC_VMRGHB },
6356 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_mergel", ALTIVEC_BUILTIN_VEC_MERGEL },
6357 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vmrglw", ALTIVEC_BUILTIN_VEC_VMRGLW },
6358 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vmrglh", ALTIVEC_BUILTIN_VEC_VMRGLH },
6359 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vmrglb", ALTIVEC_BUILTIN_VEC_VMRGLB },
6360 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_min", ALTIVEC_BUILTIN_VEC_MIN },
6361 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vminfp", ALTIVEC_BUILTIN_VEC_VMINFP },
6362 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vminsw", ALTIVEC_BUILTIN_VEC_VMINSW },
6363 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vminuw", ALTIVEC_BUILTIN_VEC_VMINUW },
6364 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vminsh", ALTIVEC_BUILTIN_VEC_VMINSH },
6365 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vminuh", ALTIVEC_BUILTIN_VEC_VMINUH },
6366 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vminsb", ALTIVEC_BUILTIN_VEC_VMINSB },
6367 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vminub", ALTIVEC_BUILTIN_VEC_VMINUB },
6368 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_mule", ALTIVEC_BUILTIN_VEC_MULE },
6369 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vmuleub", ALTIVEC_BUILTIN_VEC_VMULEUB },
6370 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vmulesb", ALTIVEC_BUILTIN_VEC_VMULESB },
6371 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vmuleuh", ALTIVEC_BUILTIN_VEC_VMULEUH },
6372 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vmulesh", ALTIVEC_BUILTIN_VEC_VMULESH },
6373 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_mulo", ALTIVEC_BUILTIN_VEC_MULO },
6374 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vmulosh", ALTIVEC_BUILTIN_VEC_VMULOSH },
6375 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vmulouh", ALTIVEC_BUILTIN_VEC_VMULOUH },
6376 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vmulosb", ALTIVEC_BUILTIN_VEC_VMULOSB },
6377 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vmuloub", ALTIVEC_BUILTIN_VEC_VMULOUB },
6378 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_nor", ALTIVEC_BUILTIN_VEC_NOR },
6379 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_or", ALTIVEC_BUILTIN_VEC_OR },
6380 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_pack", ALTIVEC_BUILTIN_VEC_PACK },
6381 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vpkuwum", ALTIVEC_BUILTIN_VEC_VPKUWUM },
6382 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vpkuhum", ALTIVEC_BUILTIN_VEC_VPKUHUM },
6383 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_packpx", ALTIVEC_BUILTIN_VEC_PACKPX },
6384 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_packs", ALTIVEC_BUILTIN_VEC_PACKS },
6385 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vpkswss", ALTIVEC_BUILTIN_VEC_VPKSWSS },
6386 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vpkuwus", ALTIVEC_BUILTIN_VEC_VPKUWUS },
6387 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vpkshss", ALTIVEC_BUILTIN_VEC_VPKSHSS },
6388 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vpkuhus", ALTIVEC_BUILTIN_VEC_VPKUHUS },
6389 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_packsu", ALTIVEC_BUILTIN_VEC_PACKSU },
6390 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vpkswus", ALTIVEC_BUILTIN_VEC_VPKSWUS },
6391 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vpkshus", ALTIVEC_BUILTIN_VEC_VPKSHUS },
6392 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_rl", ALTIVEC_BUILTIN_VEC_RL },
6393 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vrlw", ALTIVEC_BUILTIN_VEC_VRLW },
6394 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vrlh", ALTIVEC_BUILTIN_VEC_VRLH },
6395 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vrlb", ALTIVEC_BUILTIN_VEC_VRLB },
6396 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_sl", ALTIVEC_BUILTIN_VEC_SL },
6397 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vslw", ALTIVEC_BUILTIN_VEC_VSLW },
6398 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vslh", ALTIVEC_BUILTIN_VEC_VSLH },
6399 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vslb", ALTIVEC_BUILTIN_VEC_VSLB },
6400 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_sll", ALTIVEC_BUILTIN_VEC_SLL },
6401 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_slo", ALTIVEC_BUILTIN_VEC_SLO },
6402 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_sr", ALTIVEC_BUILTIN_VEC_SR },
6403 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vsrw", ALTIVEC_BUILTIN_VEC_VSRW },
6404 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vsrh", ALTIVEC_BUILTIN_VEC_VSRH },
6405 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vsrb", ALTIVEC_BUILTIN_VEC_VSRB },
6406 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_sra", ALTIVEC_BUILTIN_VEC_SRA },
6407 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vsraw", ALTIVEC_BUILTIN_VEC_VSRAW },
6408 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vsrah", ALTIVEC_BUILTIN_VEC_VSRAH },
6409 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vsrab", ALTIVEC_BUILTIN_VEC_VSRAB },
6410 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_srl", ALTIVEC_BUILTIN_VEC_SRL },
6411 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_sro", ALTIVEC_BUILTIN_VEC_SRO },
6412 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_sub", ALTIVEC_BUILTIN_VEC_SUB },
6413 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vsubfp", ALTIVEC_BUILTIN_VEC_VSUBFP },
6414 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vsubuwm", ALTIVEC_BUILTIN_VEC_VSUBUWM },
6415 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vsubuhm", ALTIVEC_BUILTIN_VEC_VSUBUHM },
6416 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vsububm", ALTIVEC_BUILTIN_VEC_VSUBUBM },
6417 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_subc", ALTIVEC_BUILTIN_VEC_SUBC },
6418 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_subs", ALTIVEC_BUILTIN_VEC_SUBS },
6419 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vsubsws", ALTIVEC_BUILTIN_VEC_VSUBSWS },
6420 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vsubuws", ALTIVEC_BUILTIN_VEC_VSUBUWS },
6421 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vsubshs", ALTIVEC_BUILTIN_VEC_VSUBSHS },
6422 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vsubuhs", ALTIVEC_BUILTIN_VEC_VSUBUHS },
6423 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vsubsbs", ALTIVEC_BUILTIN_VEC_VSUBSBS },
6424 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vsububs", ALTIVEC_BUILTIN_VEC_VSUBUBS },
6425 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_sum4s", ALTIVEC_BUILTIN_VEC_SUM4S },
6426 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vsum4shs", ALTIVEC_BUILTIN_VEC_VSUM4SHS },
6427 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vsum4sbs", ALTIVEC_BUILTIN_VEC_VSUM4SBS },
6428 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vsum4ubs", ALTIVEC_BUILTIN_VEC_VSUM4UBS },
6429 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_sum2s", ALTIVEC_BUILTIN_VEC_SUM2S },
6430 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_sums", ALTIVEC_BUILTIN_VEC_SUMS },
6431 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_xor", ALTIVEC_BUILTIN_VEC_XOR },
6432
6433 /* Place holder, leave as first spe builtin. */
6434 { 0, CODE_FOR_spe_evaddw, "__builtin_spe_evaddw", SPE_BUILTIN_EVADDW },
6435 { 0, CODE_FOR_spe_evand, "__builtin_spe_evand", SPE_BUILTIN_EVAND },
6436 { 0, CODE_FOR_spe_evandc, "__builtin_spe_evandc", SPE_BUILTIN_EVANDC },
6437 { 0, CODE_FOR_spe_evdivws, "__builtin_spe_evdivws", SPE_BUILTIN_EVDIVWS },
6438 { 0, CODE_FOR_spe_evdivwu, "__builtin_spe_evdivwu", SPE_BUILTIN_EVDIVWU },
6439 { 0, CODE_FOR_spe_eveqv, "__builtin_spe_eveqv", SPE_BUILTIN_EVEQV },
6440 { 0, CODE_FOR_spe_evfsadd, "__builtin_spe_evfsadd", SPE_BUILTIN_EVFSADD },
6441 { 0, CODE_FOR_spe_evfsdiv, "__builtin_spe_evfsdiv", SPE_BUILTIN_EVFSDIV },
6442 { 0, CODE_FOR_spe_evfsmul, "__builtin_spe_evfsmul", SPE_BUILTIN_EVFSMUL },
6443 { 0, CODE_FOR_spe_evfssub, "__builtin_spe_evfssub", SPE_BUILTIN_EVFSSUB },
6444 { 0, CODE_FOR_spe_evmergehi, "__builtin_spe_evmergehi", SPE_BUILTIN_EVMERGEHI },
6445 { 0, CODE_FOR_spe_evmergehilo, "__builtin_spe_evmergehilo", SPE_BUILTIN_EVMERGEHILO },
6446 { 0, CODE_FOR_spe_evmergelo, "__builtin_spe_evmergelo", SPE_BUILTIN_EVMERGELO },
6447 { 0, CODE_FOR_spe_evmergelohi, "__builtin_spe_evmergelohi", SPE_BUILTIN_EVMERGELOHI },
6448 { 0, CODE_FOR_spe_evmhegsmfaa, "__builtin_spe_evmhegsmfaa", SPE_BUILTIN_EVMHEGSMFAA },
6449 { 0, CODE_FOR_spe_evmhegsmfan, "__builtin_spe_evmhegsmfan", SPE_BUILTIN_EVMHEGSMFAN },
6450 { 0, CODE_FOR_spe_evmhegsmiaa, "__builtin_spe_evmhegsmiaa", SPE_BUILTIN_EVMHEGSMIAA },
6451 { 0, CODE_FOR_spe_evmhegsmian, "__builtin_spe_evmhegsmian", SPE_BUILTIN_EVMHEGSMIAN },
6452 { 0, CODE_FOR_spe_evmhegumiaa, "__builtin_spe_evmhegumiaa", SPE_BUILTIN_EVMHEGUMIAA },
6453 { 0, CODE_FOR_spe_evmhegumian, "__builtin_spe_evmhegumian", SPE_BUILTIN_EVMHEGUMIAN },
6454 { 0, CODE_FOR_spe_evmhesmf, "__builtin_spe_evmhesmf", SPE_BUILTIN_EVMHESMF },
6455 { 0, CODE_FOR_spe_evmhesmfa, "__builtin_spe_evmhesmfa", SPE_BUILTIN_EVMHESMFA },
6456 { 0, CODE_FOR_spe_evmhesmfaaw, "__builtin_spe_evmhesmfaaw", SPE_BUILTIN_EVMHESMFAAW },
6457 { 0, CODE_FOR_spe_evmhesmfanw, "__builtin_spe_evmhesmfanw", SPE_BUILTIN_EVMHESMFANW },
6458 { 0, CODE_FOR_spe_evmhesmi, "__builtin_spe_evmhesmi", SPE_BUILTIN_EVMHESMI },
6459 { 0, CODE_FOR_spe_evmhesmia, "__builtin_spe_evmhesmia", SPE_BUILTIN_EVMHESMIA },
6460 { 0, CODE_FOR_spe_evmhesmiaaw, "__builtin_spe_evmhesmiaaw", SPE_BUILTIN_EVMHESMIAAW },
6461 { 0, CODE_FOR_spe_evmhesmianw, "__builtin_spe_evmhesmianw", SPE_BUILTIN_EVMHESMIANW },
6462 { 0, CODE_FOR_spe_evmhessf, "__builtin_spe_evmhessf", SPE_BUILTIN_EVMHESSF },
6463 { 0, CODE_FOR_spe_evmhessfa, "__builtin_spe_evmhessfa", SPE_BUILTIN_EVMHESSFA },
6464 { 0, CODE_FOR_spe_evmhessfaaw, "__builtin_spe_evmhessfaaw", SPE_BUILTIN_EVMHESSFAAW },
6465 { 0, CODE_FOR_spe_evmhessfanw, "__builtin_spe_evmhessfanw", SPE_BUILTIN_EVMHESSFANW },
6466 { 0, CODE_FOR_spe_evmhessiaaw, "__builtin_spe_evmhessiaaw", SPE_BUILTIN_EVMHESSIAAW },
6467 { 0, CODE_FOR_spe_evmhessianw, "__builtin_spe_evmhessianw", SPE_BUILTIN_EVMHESSIANW },
6468 { 0, CODE_FOR_spe_evmheumi, "__builtin_spe_evmheumi", SPE_BUILTIN_EVMHEUMI },
6469 { 0, CODE_FOR_spe_evmheumia, "__builtin_spe_evmheumia", SPE_BUILTIN_EVMHEUMIA },
6470 { 0, CODE_FOR_spe_evmheumiaaw, "__builtin_spe_evmheumiaaw", SPE_BUILTIN_EVMHEUMIAAW },
6471 { 0, CODE_FOR_spe_evmheumianw, "__builtin_spe_evmheumianw", SPE_BUILTIN_EVMHEUMIANW },
6472 { 0, CODE_FOR_spe_evmheusiaaw, "__builtin_spe_evmheusiaaw", SPE_BUILTIN_EVMHEUSIAAW },
6473 { 0, CODE_FOR_spe_evmheusianw, "__builtin_spe_evmheusianw", SPE_BUILTIN_EVMHEUSIANW },
6474 { 0, CODE_FOR_spe_evmhogsmfaa, "__builtin_spe_evmhogsmfaa", SPE_BUILTIN_EVMHOGSMFAA },
6475 { 0, CODE_FOR_spe_evmhogsmfan, "__builtin_spe_evmhogsmfan", SPE_BUILTIN_EVMHOGSMFAN },
6476 { 0, CODE_FOR_spe_evmhogsmiaa, "__builtin_spe_evmhogsmiaa", SPE_BUILTIN_EVMHOGSMIAA },
6477 { 0, CODE_FOR_spe_evmhogsmian, "__builtin_spe_evmhogsmian", SPE_BUILTIN_EVMHOGSMIAN },
6478 { 0, CODE_FOR_spe_evmhogumiaa, "__builtin_spe_evmhogumiaa", SPE_BUILTIN_EVMHOGUMIAA },
6479 { 0, CODE_FOR_spe_evmhogumian, "__builtin_spe_evmhogumian", SPE_BUILTIN_EVMHOGUMIAN },
6480 { 0, CODE_FOR_spe_evmhosmf, "__builtin_spe_evmhosmf", SPE_BUILTIN_EVMHOSMF },
6481 { 0, CODE_FOR_spe_evmhosmfa, "__builtin_spe_evmhosmfa", SPE_BUILTIN_EVMHOSMFA },
6482 { 0, CODE_FOR_spe_evmhosmfaaw, "__builtin_spe_evmhosmfaaw", SPE_BUILTIN_EVMHOSMFAAW },
6483 { 0, CODE_FOR_spe_evmhosmfanw, "__builtin_spe_evmhosmfanw", SPE_BUILTIN_EVMHOSMFANW },
6484 { 0, CODE_FOR_spe_evmhosmi, "__builtin_spe_evmhosmi", SPE_BUILTIN_EVMHOSMI },
6485 { 0, CODE_FOR_spe_evmhosmia, "__builtin_spe_evmhosmia", SPE_BUILTIN_EVMHOSMIA },
6486 { 0, CODE_FOR_spe_evmhosmiaaw, "__builtin_spe_evmhosmiaaw", SPE_BUILTIN_EVMHOSMIAAW },
6487 { 0, CODE_FOR_spe_evmhosmianw, "__builtin_spe_evmhosmianw", SPE_BUILTIN_EVMHOSMIANW },
6488 { 0, CODE_FOR_spe_evmhossf, "__builtin_spe_evmhossf", SPE_BUILTIN_EVMHOSSF },
6489 { 0, CODE_FOR_spe_evmhossfa, "__builtin_spe_evmhossfa", SPE_BUILTIN_EVMHOSSFA },
6490 { 0, CODE_FOR_spe_evmhossfaaw, "__builtin_spe_evmhossfaaw", SPE_BUILTIN_EVMHOSSFAAW },
6491 { 0, CODE_FOR_spe_evmhossfanw, "__builtin_spe_evmhossfanw", SPE_BUILTIN_EVMHOSSFANW },
6492 { 0, CODE_FOR_spe_evmhossiaaw, "__builtin_spe_evmhossiaaw", SPE_BUILTIN_EVMHOSSIAAW },
6493 { 0, CODE_FOR_spe_evmhossianw, "__builtin_spe_evmhossianw", SPE_BUILTIN_EVMHOSSIANW },
6494 { 0, CODE_FOR_spe_evmhoumi, "__builtin_spe_evmhoumi", SPE_BUILTIN_EVMHOUMI },
6495 { 0, CODE_FOR_spe_evmhoumia, "__builtin_spe_evmhoumia", SPE_BUILTIN_EVMHOUMIA },
6496 { 0, CODE_FOR_spe_evmhoumiaaw, "__builtin_spe_evmhoumiaaw", SPE_BUILTIN_EVMHOUMIAAW },
6497 { 0, CODE_FOR_spe_evmhoumianw, "__builtin_spe_evmhoumianw", SPE_BUILTIN_EVMHOUMIANW },
6498 { 0, CODE_FOR_spe_evmhousiaaw, "__builtin_spe_evmhousiaaw", SPE_BUILTIN_EVMHOUSIAAW },
6499 { 0, CODE_FOR_spe_evmhousianw, "__builtin_spe_evmhousianw", SPE_BUILTIN_EVMHOUSIANW },
6500 { 0, CODE_FOR_spe_evmwhsmf, "__builtin_spe_evmwhsmf", SPE_BUILTIN_EVMWHSMF },
6501 { 0, CODE_FOR_spe_evmwhsmfa, "__builtin_spe_evmwhsmfa", SPE_BUILTIN_EVMWHSMFA },
6502 { 0, CODE_FOR_spe_evmwhsmi, "__builtin_spe_evmwhsmi", SPE_BUILTIN_EVMWHSMI },
6503 { 0, CODE_FOR_spe_evmwhsmia, "__builtin_spe_evmwhsmia", SPE_BUILTIN_EVMWHSMIA },
6504 { 0, CODE_FOR_spe_evmwhssf, "__builtin_spe_evmwhssf", SPE_BUILTIN_EVMWHSSF },
6505 { 0, CODE_FOR_spe_evmwhssfa, "__builtin_spe_evmwhssfa", SPE_BUILTIN_EVMWHSSFA },
6506 { 0, CODE_FOR_spe_evmwhumi, "__builtin_spe_evmwhumi", SPE_BUILTIN_EVMWHUMI },
6507 { 0, CODE_FOR_spe_evmwhumia, "__builtin_spe_evmwhumia", SPE_BUILTIN_EVMWHUMIA },
6508 { 0, CODE_FOR_spe_evmwlsmiaaw, "__builtin_spe_evmwlsmiaaw", SPE_BUILTIN_EVMWLSMIAAW },
6509 { 0, CODE_FOR_spe_evmwlsmianw, "__builtin_spe_evmwlsmianw", SPE_BUILTIN_EVMWLSMIANW },
6510 { 0, CODE_FOR_spe_evmwlssiaaw, "__builtin_spe_evmwlssiaaw", SPE_BUILTIN_EVMWLSSIAAW },
6511 { 0, CODE_FOR_spe_evmwlssianw, "__builtin_spe_evmwlssianw", SPE_BUILTIN_EVMWLSSIANW },
6512 { 0, CODE_FOR_spe_evmwlumi, "__builtin_spe_evmwlumi", SPE_BUILTIN_EVMWLUMI },
6513 { 0, CODE_FOR_spe_evmwlumia, "__builtin_spe_evmwlumia", SPE_BUILTIN_EVMWLUMIA },
6514 { 0, CODE_FOR_spe_evmwlumiaaw, "__builtin_spe_evmwlumiaaw", SPE_BUILTIN_EVMWLUMIAAW },
6515 { 0, CODE_FOR_spe_evmwlumianw, "__builtin_spe_evmwlumianw", SPE_BUILTIN_EVMWLUMIANW },
6516 { 0, CODE_FOR_spe_evmwlusiaaw, "__builtin_spe_evmwlusiaaw", SPE_BUILTIN_EVMWLUSIAAW },
6517 { 0, CODE_FOR_spe_evmwlusianw, "__builtin_spe_evmwlusianw", SPE_BUILTIN_EVMWLUSIANW },
6518 { 0, CODE_FOR_spe_evmwsmf, "__builtin_spe_evmwsmf", SPE_BUILTIN_EVMWSMF },
6519 { 0, CODE_FOR_spe_evmwsmfa, "__builtin_spe_evmwsmfa", SPE_BUILTIN_EVMWSMFA },
6520 { 0, CODE_FOR_spe_evmwsmfaa, "__builtin_spe_evmwsmfaa", SPE_BUILTIN_EVMWSMFAA },
6521 { 0, CODE_FOR_spe_evmwsmfan, "__builtin_spe_evmwsmfan", SPE_BUILTIN_EVMWSMFAN },
6522 { 0, CODE_FOR_spe_evmwsmi, "__builtin_spe_evmwsmi", SPE_BUILTIN_EVMWSMI },
6523 { 0, CODE_FOR_spe_evmwsmia, "__builtin_spe_evmwsmia", SPE_BUILTIN_EVMWSMIA },
6524 { 0, CODE_FOR_spe_evmwsmiaa, "__builtin_spe_evmwsmiaa", SPE_BUILTIN_EVMWSMIAA },
6525 { 0, CODE_FOR_spe_evmwsmian, "__builtin_spe_evmwsmian", SPE_BUILTIN_EVMWSMIAN },
6526 { 0, CODE_FOR_spe_evmwssf, "__builtin_spe_evmwssf", SPE_BUILTIN_EVMWSSF },
6527 { 0, CODE_FOR_spe_evmwssfa, "__builtin_spe_evmwssfa", SPE_BUILTIN_EVMWSSFA },
6528 { 0, CODE_FOR_spe_evmwssfaa, "__builtin_spe_evmwssfaa", SPE_BUILTIN_EVMWSSFAA },
6529 { 0, CODE_FOR_spe_evmwssfan, "__builtin_spe_evmwssfan", SPE_BUILTIN_EVMWSSFAN },
6530 { 0, CODE_FOR_spe_evmwumi, "__builtin_spe_evmwumi", SPE_BUILTIN_EVMWUMI },
6531 { 0, CODE_FOR_spe_evmwumia, "__builtin_spe_evmwumia", SPE_BUILTIN_EVMWUMIA },
6532 { 0, CODE_FOR_spe_evmwumiaa, "__builtin_spe_evmwumiaa", SPE_BUILTIN_EVMWUMIAA },
6533 { 0, CODE_FOR_spe_evmwumian, "__builtin_spe_evmwumian", SPE_BUILTIN_EVMWUMIAN },
6534 { 0, CODE_FOR_spe_evnand, "__builtin_spe_evnand", SPE_BUILTIN_EVNAND },
6535 { 0, CODE_FOR_spe_evnor, "__builtin_spe_evnor", SPE_BUILTIN_EVNOR },
6536 { 0, CODE_FOR_spe_evor, "__builtin_spe_evor", SPE_BUILTIN_EVOR },
6537 { 0, CODE_FOR_spe_evorc, "__builtin_spe_evorc", SPE_BUILTIN_EVORC },
6538 { 0, CODE_FOR_spe_evrlw, "__builtin_spe_evrlw", SPE_BUILTIN_EVRLW },
6539 { 0, CODE_FOR_spe_evslw, "__builtin_spe_evslw", SPE_BUILTIN_EVSLW },
6540 { 0, CODE_FOR_spe_evsrws, "__builtin_spe_evsrws", SPE_BUILTIN_EVSRWS },
6541 { 0, CODE_FOR_spe_evsrwu, "__builtin_spe_evsrwu", SPE_BUILTIN_EVSRWU },
6542 { 0, CODE_FOR_spe_evsubfw, "__builtin_spe_evsubfw", SPE_BUILTIN_EVSUBFW },
6543
6544 /* SPE binary operations expecting a 5-bit unsigned literal. */
6545 { 0, CODE_FOR_spe_evaddiw, "__builtin_spe_evaddiw", SPE_BUILTIN_EVADDIW },
6546
6547 { 0, CODE_FOR_spe_evrlwi, "__builtin_spe_evrlwi", SPE_BUILTIN_EVRLWI },
6548 { 0, CODE_FOR_spe_evslwi, "__builtin_spe_evslwi", SPE_BUILTIN_EVSLWI },
6549 { 0, CODE_FOR_spe_evsrwis, "__builtin_spe_evsrwis", SPE_BUILTIN_EVSRWIS },
6550 { 0, CODE_FOR_spe_evsrwiu, "__builtin_spe_evsrwiu", SPE_BUILTIN_EVSRWIU },
6551 { 0, CODE_FOR_spe_evsubifw, "__builtin_spe_evsubifw", SPE_BUILTIN_EVSUBIFW },
6552 { 0, CODE_FOR_spe_evmwhssfaa, "__builtin_spe_evmwhssfaa", SPE_BUILTIN_EVMWHSSFAA },
6553 { 0, CODE_FOR_spe_evmwhssmaa, "__builtin_spe_evmwhssmaa", SPE_BUILTIN_EVMWHSSMAA },
6554 { 0, CODE_FOR_spe_evmwhsmfaa, "__builtin_spe_evmwhsmfaa", SPE_BUILTIN_EVMWHSMFAA },
6555 { 0, CODE_FOR_spe_evmwhsmiaa, "__builtin_spe_evmwhsmiaa", SPE_BUILTIN_EVMWHSMIAA },
6556 { 0, CODE_FOR_spe_evmwhusiaa, "__builtin_spe_evmwhusiaa", SPE_BUILTIN_EVMWHUSIAA },
6557 { 0, CODE_FOR_spe_evmwhumiaa, "__builtin_spe_evmwhumiaa", SPE_BUILTIN_EVMWHUMIAA },
6558 { 0, CODE_FOR_spe_evmwhssfan, "__builtin_spe_evmwhssfan", SPE_BUILTIN_EVMWHSSFAN },
6559 { 0, CODE_FOR_spe_evmwhssian, "__builtin_spe_evmwhssian", SPE_BUILTIN_EVMWHSSIAN },
6560 { 0, CODE_FOR_spe_evmwhsmfan, "__builtin_spe_evmwhsmfan", SPE_BUILTIN_EVMWHSMFAN },
6561 { 0, CODE_FOR_spe_evmwhsmian, "__builtin_spe_evmwhsmian", SPE_BUILTIN_EVMWHSMIAN },
6562 { 0, CODE_FOR_spe_evmwhusian, "__builtin_spe_evmwhusian", SPE_BUILTIN_EVMWHUSIAN },
6563 { 0, CODE_FOR_spe_evmwhumian, "__builtin_spe_evmwhumian", SPE_BUILTIN_EVMWHUMIAN },
6564 { 0, CODE_FOR_spe_evmwhgssfaa, "__builtin_spe_evmwhgssfaa", SPE_BUILTIN_EVMWHGSSFAA },
6565 { 0, CODE_FOR_spe_evmwhgsmfaa, "__builtin_spe_evmwhgsmfaa", SPE_BUILTIN_EVMWHGSMFAA },
6566 { 0, CODE_FOR_spe_evmwhgsmiaa, "__builtin_spe_evmwhgsmiaa", SPE_BUILTIN_EVMWHGSMIAA },
6567 { 0, CODE_FOR_spe_evmwhgumiaa, "__builtin_spe_evmwhgumiaa", SPE_BUILTIN_EVMWHGUMIAA },
6568 { 0, CODE_FOR_spe_evmwhgssfan, "__builtin_spe_evmwhgssfan", SPE_BUILTIN_EVMWHGSSFAN },
6569 { 0, CODE_FOR_spe_evmwhgsmfan, "__builtin_spe_evmwhgsmfan", SPE_BUILTIN_EVMWHGSMFAN },
6570 { 0, CODE_FOR_spe_evmwhgsmian, "__builtin_spe_evmwhgsmian", SPE_BUILTIN_EVMWHGSMIAN },
6571 { 0, CODE_FOR_spe_evmwhgumian, "__builtin_spe_evmwhgumian", SPE_BUILTIN_EVMWHGUMIAN },
6572 { 0, CODE_FOR_spe_brinc, "__builtin_spe_brinc", SPE_BUILTIN_BRINC },
6573
6574 /* Place-holder. Leave as last binary SPE builtin. */
6575 { 0, CODE_FOR_xorv2si3, "__builtin_spe_evxor", SPE_BUILTIN_EVXOR }
6576 };
6577
6578 /* AltiVec predicates. */
6579
6580 struct builtin_description_predicates
6581 {
6582 const unsigned int mask;
6583 const enum insn_code icode;
6584 const char *opcode;
6585 const char *const name;
6586 const enum rs6000_builtins code;
6587 };
6588
6589 static const struct builtin_description_predicates bdesc_altivec_preds[] =
6590 {
6591 { MASK_ALTIVEC, CODE_FOR_altivec_predicate_v4sf, "*vcmpbfp.", "__builtin_altivec_vcmpbfp_p", ALTIVEC_BUILTIN_VCMPBFP_P },
6592 { MASK_ALTIVEC, CODE_FOR_altivec_predicate_v4sf, "*vcmpeqfp.", "__builtin_altivec_vcmpeqfp_p", ALTIVEC_BUILTIN_VCMPEQFP_P },
6593 { MASK_ALTIVEC, CODE_FOR_altivec_predicate_v4sf, "*vcmpgefp.", "__builtin_altivec_vcmpgefp_p", ALTIVEC_BUILTIN_VCMPGEFP_P },
6594 { MASK_ALTIVEC, CODE_FOR_altivec_predicate_v4sf, "*vcmpgtfp.", "__builtin_altivec_vcmpgtfp_p", ALTIVEC_BUILTIN_VCMPGTFP_P },
6595 { MASK_ALTIVEC, CODE_FOR_altivec_predicate_v4si, "*vcmpequw.", "__builtin_altivec_vcmpequw_p", ALTIVEC_BUILTIN_VCMPEQUW_P },
6596 { MASK_ALTIVEC, CODE_FOR_altivec_predicate_v4si, "*vcmpgtsw.", "__builtin_altivec_vcmpgtsw_p", ALTIVEC_BUILTIN_VCMPGTSW_P },
6597 { MASK_ALTIVEC, CODE_FOR_altivec_predicate_v4si, "*vcmpgtuw.", "__builtin_altivec_vcmpgtuw_p", ALTIVEC_BUILTIN_VCMPGTUW_P },
6598 { MASK_ALTIVEC, CODE_FOR_altivec_predicate_v8hi, "*vcmpgtuh.", "__builtin_altivec_vcmpgtuh_p", ALTIVEC_BUILTIN_VCMPGTUH_P },
6599 { MASK_ALTIVEC, CODE_FOR_altivec_predicate_v8hi, "*vcmpgtsh.", "__builtin_altivec_vcmpgtsh_p", ALTIVEC_BUILTIN_VCMPGTSH_P },
6600 { MASK_ALTIVEC, CODE_FOR_altivec_predicate_v8hi, "*vcmpequh.", "__builtin_altivec_vcmpequh_p", ALTIVEC_BUILTIN_VCMPEQUH_P },
6601 { MASK_ALTIVEC, CODE_FOR_altivec_predicate_v16qi, "*vcmpequb.", "__builtin_altivec_vcmpequb_p", ALTIVEC_BUILTIN_VCMPEQUB_P },
6602 { MASK_ALTIVEC, CODE_FOR_altivec_predicate_v16qi, "*vcmpgtsb.", "__builtin_altivec_vcmpgtsb_p", ALTIVEC_BUILTIN_VCMPGTSB_P },
6603 { MASK_ALTIVEC, CODE_FOR_altivec_predicate_v16qi, "*vcmpgtub.", "__builtin_altivec_vcmpgtub_p", ALTIVEC_BUILTIN_VCMPGTUB_P },
6604
6605 { MASK_ALTIVEC, 0, NULL, "__builtin_vec_vcmpeq_p", ALTIVEC_BUILTIN_VCMPEQ_P },
6606 { MASK_ALTIVEC, 0, NULL, "__builtin_vec_vcmpgt_p", ALTIVEC_BUILTIN_VCMPGT_P },
6607 { MASK_ALTIVEC, 0, NULL, "__builtin_vec_vcmpge_p", ALTIVEC_BUILTIN_VCMPGE_P }
6608 };
6609
6610 /* SPE predicates. */
6611 static struct builtin_description bdesc_spe_predicates[] =
6612 {
6613 /* Place-holder. Leave as first. */
6614 { 0, CODE_FOR_spe_evcmpeq, "__builtin_spe_evcmpeq", SPE_BUILTIN_EVCMPEQ },
6615 { 0, CODE_FOR_spe_evcmpgts, "__builtin_spe_evcmpgts", SPE_BUILTIN_EVCMPGTS },
6616 { 0, CODE_FOR_spe_evcmpgtu, "__builtin_spe_evcmpgtu", SPE_BUILTIN_EVCMPGTU },
6617 { 0, CODE_FOR_spe_evcmplts, "__builtin_spe_evcmplts", SPE_BUILTIN_EVCMPLTS },
6618 { 0, CODE_FOR_spe_evcmpltu, "__builtin_spe_evcmpltu", SPE_BUILTIN_EVCMPLTU },
6619 { 0, CODE_FOR_spe_evfscmpeq, "__builtin_spe_evfscmpeq", SPE_BUILTIN_EVFSCMPEQ },
6620 { 0, CODE_FOR_spe_evfscmpgt, "__builtin_spe_evfscmpgt", SPE_BUILTIN_EVFSCMPGT },
6621 { 0, CODE_FOR_spe_evfscmplt, "__builtin_spe_evfscmplt", SPE_BUILTIN_EVFSCMPLT },
6622 { 0, CODE_FOR_spe_evfststeq, "__builtin_spe_evfststeq", SPE_BUILTIN_EVFSTSTEQ },
6623 { 0, CODE_FOR_spe_evfststgt, "__builtin_spe_evfststgt", SPE_BUILTIN_EVFSTSTGT },
6624 /* Place-holder. Leave as last. */
6625 { 0, CODE_FOR_spe_evfststlt, "__builtin_spe_evfststlt", SPE_BUILTIN_EVFSTSTLT },
6626 };
6627
6628 /* SPE evsel predicates. */
6629 static struct builtin_description bdesc_spe_evsel[] =
6630 {
6631 /* Place-holder. Leave as first. */
6632 { 0, CODE_FOR_spe_evcmpgts, "__builtin_spe_evsel_gts", SPE_BUILTIN_EVSEL_CMPGTS },
6633 { 0, CODE_FOR_spe_evcmpgtu, "__builtin_spe_evsel_gtu", SPE_BUILTIN_EVSEL_CMPGTU },
6634 { 0, CODE_FOR_spe_evcmplts, "__builtin_spe_evsel_lts", SPE_BUILTIN_EVSEL_CMPLTS },
6635 { 0, CODE_FOR_spe_evcmpltu, "__builtin_spe_evsel_ltu", SPE_BUILTIN_EVSEL_CMPLTU },
6636 { 0, CODE_FOR_spe_evcmpeq, "__builtin_spe_evsel_eq", SPE_BUILTIN_EVSEL_CMPEQ },
6637 { 0, CODE_FOR_spe_evfscmpgt, "__builtin_spe_evsel_fsgt", SPE_BUILTIN_EVSEL_FSCMPGT },
6638 { 0, CODE_FOR_spe_evfscmplt, "__builtin_spe_evsel_fslt", SPE_BUILTIN_EVSEL_FSCMPLT },
6639 { 0, CODE_FOR_spe_evfscmpeq, "__builtin_spe_evsel_fseq", SPE_BUILTIN_EVSEL_FSCMPEQ },
6640 { 0, CODE_FOR_spe_evfststgt, "__builtin_spe_evsel_fststgt", SPE_BUILTIN_EVSEL_FSTSTGT },
6641 { 0, CODE_FOR_spe_evfststlt, "__builtin_spe_evsel_fststlt", SPE_BUILTIN_EVSEL_FSTSTLT },
6642 /* Place-holder. Leave as last. */
6643 { 0, CODE_FOR_spe_evfststeq, "__builtin_spe_evsel_fststeq", SPE_BUILTIN_EVSEL_FSTSTEQ },
6644 };
6645
6646 /* ABS* operations. */
6647
6648 static const struct builtin_description bdesc_abs[] =
6649 {
6650 { MASK_ALTIVEC, CODE_FOR_absv4si2, "__builtin_altivec_abs_v4si", ALTIVEC_BUILTIN_ABS_V4SI },
6651 { MASK_ALTIVEC, CODE_FOR_absv8hi2, "__builtin_altivec_abs_v8hi", ALTIVEC_BUILTIN_ABS_V8HI },
6652 { MASK_ALTIVEC, CODE_FOR_absv4sf2, "__builtin_altivec_abs_v4sf", ALTIVEC_BUILTIN_ABS_V4SF },
6653 { MASK_ALTIVEC, CODE_FOR_absv16qi2, "__builtin_altivec_abs_v16qi", ALTIVEC_BUILTIN_ABS_V16QI },
6654 { MASK_ALTIVEC, CODE_FOR_altivec_abss_v4si, "__builtin_altivec_abss_v4si", ALTIVEC_BUILTIN_ABSS_V4SI },
6655 { MASK_ALTIVEC, CODE_FOR_altivec_abss_v8hi, "__builtin_altivec_abss_v8hi", ALTIVEC_BUILTIN_ABSS_V8HI },
6656 { MASK_ALTIVEC, CODE_FOR_altivec_abss_v16qi, "__builtin_altivec_abss_v16qi", ALTIVEC_BUILTIN_ABSS_V16QI }
6657 };
6658
6659 /* Simple unary operations: VECb = foo (unsigned literal) or VECb =
6660 foo (VECa). */
6661
6662 static struct builtin_description bdesc_1arg[] =
6663 {
6664 { MASK_ALTIVEC, CODE_FOR_altivec_vexptefp, "__builtin_altivec_vexptefp", ALTIVEC_BUILTIN_VEXPTEFP },
6665 { MASK_ALTIVEC, CODE_FOR_altivec_vlogefp, "__builtin_altivec_vlogefp", ALTIVEC_BUILTIN_VLOGEFP },
6666 { MASK_ALTIVEC, CODE_FOR_altivec_vrefp, "__builtin_altivec_vrefp", ALTIVEC_BUILTIN_VREFP },
6667 { MASK_ALTIVEC, CODE_FOR_altivec_vrfim, "__builtin_altivec_vrfim", ALTIVEC_BUILTIN_VRFIM },
6668 { MASK_ALTIVEC, CODE_FOR_altivec_vrfin, "__builtin_altivec_vrfin", ALTIVEC_BUILTIN_VRFIN },
6669 { MASK_ALTIVEC, CODE_FOR_altivec_vrfip, "__builtin_altivec_vrfip", ALTIVEC_BUILTIN_VRFIP },
6670 { MASK_ALTIVEC, CODE_FOR_ftruncv4sf2, "__builtin_altivec_vrfiz", ALTIVEC_BUILTIN_VRFIZ },
6671 { MASK_ALTIVEC, CODE_FOR_altivec_vrsqrtefp, "__builtin_altivec_vrsqrtefp", ALTIVEC_BUILTIN_VRSQRTEFP },
6672 { MASK_ALTIVEC, CODE_FOR_altivec_vspltisb, "__builtin_altivec_vspltisb", ALTIVEC_BUILTIN_VSPLTISB },
6673 { MASK_ALTIVEC, CODE_FOR_altivec_vspltish, "__builtin_altivec_vspltish", ALTIVEC_BUILTIN_VSPLTISH },
6674 { MASK_ALTIVEC, CODE_FOR_altivec_vspltisw, "__builtin_altivec_vspltisw", ALTIVEC_BUILTIN_VSPLTISW },
6675 { MASK_ALTIVEC, CODE_FOR_altivec_vupkhsb, "__builtin_altivec_vupkhsb", ALTIVEC_BUILTIN_VUPKHSB },
6676 { MASK_ALTIVEC, CODE_FOR_altivec_vupkhpx, "__builtin_altivec_vupkhpx", ALTIVEC_BUILTIN_VUPKHPX },
6677 { MASK_ALTIVEC, CODE_FOR_altivec_vupkhsh, "__builtin_altivec_vupkhsh", ALTIVEC_BUILTIN_VUPKHSH },
6678 { MASK_ALTIVEC, CODE_FOR_altivec_vupklsb, "__builtin_altivec_vupklsb", ALTIVEC_BUILTIN_VUPKLSB },
6679 { MASK_ALTIVEC, CODE_FOR_altivec_vupklpx, "__builtin_altivec_vupklpx", ALTIVEC_BUILTIN_VUPKLPX },
6680 { MASK_ALTIVEC, CODE_FOR_altivec_vupklsh, "__builtin_altivec_vupklsh", ALTIVEC_BUILTIN_VUPKLSH },
6681
6682 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_abs", ALTIVEC_BUILTIN_VEC_ABS },
6683 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_abss", ALTIVEC_BUILTIN_VEC_ABSS },
6684 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_ceil", ALTIVEC_BUILTIN_VEC_CEIL },
6685 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_expte", ALTIVEC_BUILTIN_VEC_EXPTE },
6686 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_floor", ALTIVEC_BUILTIN_VEC_FLOOR },
6687 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_loge", ALTIVEC_BUILTIN_VEC_LOGE },
6688 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_mtvscr", ALTIVEC_BUILTIN_VEC_MTVSCR },
6689 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_re", ALTIVEC_BUILTIN_VEC_RE },
6690 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_round", ALTIVEC_BUILTIN_VEC_ROUND },
6691 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_rsqrte", ALTIVEC_BUILTIN_VEC_RSQRTE },
6692 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_trunc", ALTIVEC_BUILTIN_VEC_TRUNC },
6693 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_unpackh", ALTIVEC_BUILTIN_VEC_UNPACKH },
6694 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vupkhsh", ALTIVEC_BUILTIN_VEC_VUPKHSH },
6695 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vupkhpx", ALTIVEC_BUILTIN_VEC_VUPKHPX },
6696 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vupkhsb", ALTIVEC_BUILTIN_VEC_VUPKHSB },
6697 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_unpackl", ALTIVEC_BUILTIN_VEC_UNPACKL },
6698 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vupklpx", ALTIVEC_BUILTIN_VEC_VUPKLPX },
6699 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vupklsh", ALTIVEC_BUILTIN_VEC_VUPKLSH },
6700 { MASK_ALTIVEC, CODE_FOR_nothing, "__builtin_vec_vupklsb", ALTIVEC_BUILTIN_VEC_VUPKLSB },
6701
6702 /* The SPE unary builtins must start with SPE_BUILTIN_EVABS and
6703 end with SPE_BUILTIN_EVSUBFUSIAAW. */
6704 { 0, CODE_FOR_spe_evabs, "__builtin_spe_evabs", SPE_BUILTIN_EVABS },
6705 { 0, CODE_FOR_spe_evaddsmiaaw, "__builtin_spe_evaddsmiaaw", SPE_BUILTIN_EVADDSMIAAW },
6706 { 0, CODE_FOR_spe_evaddssiaaw, "__builtin_spe_evaddssiaaw", SPE_BUILTIN_EVADDSSIAAW },
6707 { 0, CODE_FOR_spe_evaddumiaaw, "__builtin_spe_evaddumiaaw", SPE_BUILTIN_EVADDUMIAAW },
6708 { 0, CODE_FOR_spe_evaddusiaaw, "__builtin_spe_evaddusiaaw", SPE_BUILTIN_EVADDUSIAAW },
6709 { 0, CODE_FOR_spe_evcntlsw, "__builtin_spe_evcntlsw", SPE_BUILTIN_EVCNTLSW },
6710 { 0, CODE_FOR_spe_evcntlzw, "__builtin_spe_evcntlzw", SPE_BUILTIN_EVCNTLZW },
6711 { 0, CODE_FOR_spe_evextsb, "__builtin_spe_evextsb", SPE_BUILTIN_EVEXTSB },
6712 { 0, CODE_FOR_spe_evextsh, "__builtin_spe_evextsh", SPE_BUILTIN_EVEXTSH },
6713 { 0, CODE_FOR_spe_evfsabs, "__builtin_spe_evfsabs", SPE_BUILTIN_EVFSABS },
6714 { 0, CODE_FOR_spe_evfscfsf, "__builtin_spe_evfscfsf", SPE_BUILTIN_EVFSCFSF },
6715 { 0, CODE_FOR_spe_evfscfsi, "__builtin_spe_evfscfsi", SPE_BUILTIN_EVFSCFSI },
6716 { 0, CODE_FOR_spe_evfscfuf, "__builtin_spe_evfscfuf", SPE_BUILTIN_EVFSCFUF },
6717 { 0, CODE_FOR_spe_evfscfui, "__builtin_spe_evfscfui", SPE_BUILTIN_EVFSCFUI },
6718 { 0, CODE_FOR_spe_evfsctsf, "__builtin_spe_evfsctsf", SPE_BUILTIN_EVFSCTSF },
6719 { 0, CODE_FOR_spe_evfsctsi, "__builtin_spe_evfsctsi", SPE_BUILTIN_EVFSCTSI },
6720 { 0, CODE_FOR_spe_evfsctsiz, "__builtin_spe_evfsctsiz", SPE_BUILTIN_EVFSCTSIZ },
6721 { 0, CODE_FOR_spe_evfsctuf, "__builtin_spe_evfsctuf", SPE_BUILTIN_EVFSCTUF },
6722 { 0, CODE_FOR_spe_evfsctui, "__builtin_spe_evfsctui", SPE_BUILTIN_EVFSCTUI },
6723 { 0, CODE_FOR_spe_evfsctuiz, "__builtin_spe_evfsctuiz", SPE_BUILTIN_EVFSCTUIZ },
6724 { 0, CODE_FOR_spe_evfsnabs, "__builtin_spe_evfsnabs", SPE_BUILTIN_EVFSNABS },
6725 { 0, CODE_FOR_spe_evfsneg, "__builtin_spe_evfsneg", SPE_BUILTIN_EVFSNEG },
6726 { 0, CODE_FOR_spe_evmra, "__builtin_spe_evmra", SPE_BUILTIN_EVMRA },
6727 { 0, CODE_FOR_negv2si2, "__builtin_spe_evneg", SPE_BUILTIN_EVNEG },
6728 { 0, CODE_FOR_spe_evrndw, "__builtin_spe_evrndw", SPE_BUILTIN_EVRNDW },
6729 { 0, CODE_FOR_spe_evsubfsmiaaw, "__builtin_spe_evsubfsmiaaw", SPE_BUILTIN_EVSUBFSMIAAW },
6730 { 0, CODE_FOR_spe_evsubfssiaaw, "__builtin_spe_evsubfssiaaw", SPE_BUILTIN_EVSUBFSSIAAW },
6731 { 0, CODE_FOR_spe_evsubfumiaaw, "__builtin_spe_evsubfumiaaw", SPE_BUILTIN_EVSUBFUMIAAW },
6732
6733 /* Place-holder. Leave as last unary SPE builtin. */
6734 { 0, CODE_FOR_spe_evsubfusiaaw, "__builtin_spe_evsubfusiaaw", SPE_BUILTIN_EVSUBFUSIAAW }
6735 };
6736
6737 static rtx
rs6000_expand_unop_builtin(enum insn_code icode,tree arglist,rtx target)6738 rs6000_expand_unop_builtin (enum insn_code icode, tree arglist, rtx target)
6739 {
6740 rtx pat;
6741 tree arg0 = TREE_VALUE (arglist);
6742 rtx op0 = expand_normal (arg0);
6743 enum machine_mode tmode = insn_data[icode].operand[0].mode;
6744 enum machine_mode mode0 = insn_data[icode].operand[1].mode;
6745
6746 if (icode == CODE_FOR_nothing)
6747 /* Builtin not supported on this processor. */
6748 return 0;
6749
6750 /* If we got invalid arguments bail out before generating bad rtl. */
6751 if (arg0 == error_mark_node)
6752 return const0_rtx;
6753
6754 if (icode == CODE_FOR_altivec_vspltisb
6755 || icode == CODE_FOR_altivec_vspltish
6756 || icode == CODE_FOR_altivec_vspltisw
6757 || icode == CODE_FOR_spe_evsplatfi
6758 || icode == CODE_FOR_spe_evsplati)
6759 {
6760 /* Only allow 5-bit *signed* literals. */
6761 if (GET_CODE (op0) != CONST_INT
6762 || INTVAL (op0) > 15
6763 || INTVAL (op0) < -16)
6764 {
6765 error ("argument 1 must be a 5-bit signed literal");
6766 return const0_rtx;
6767 }
6768 }
6769
6770 if (target == 0
6771 || GET_MODE (target) != tmode
6772 || ! (*insn_data[icode].operand[0].predicate) (target, tmode))
6773 target = gen_reg_rtx (tmode);
6774
6775 if (! (*insn_data[icode].operand[1].predicate) (op0, mode0))
6776 op0 = copy_to_mode_reg (mode0, op0);
6777
6778 pat = GEN_FCN (icode) (target, op0);
6779 if (! pat)
6780 return 0;
6781 emit_insn (pat);
6782
6783 return target;
6784 }
6785
6786 static rtx
altivec_expand_abs_builtin(enum insn_code icode,tree arglist,rtx target)6787 altivec_expand_abs_builtin (enum insn_code icode, tree arglist, rtx target)
6788 {
6789 rtx pat, scratch1, scratch2;
6790 tree arg0 = TREE_VALUE (arglist);
6791 rtx op0 = expand_normal (arg0);
6792 enum machine_mode tmode = insn_data[icode].operand[0].mode;
6793 enum machine_mode mode0 = insn_data[icode].operand[1].mode;
6794
6795 /* If we have invalid arguments, bail out before generating bad rtl. */
6796 if (arg0 == error_mark_node)
6797 return const0_rtx;
6798
6799 if (target == 0
6800 || GET_MODE (target) != tmode
6801 || ! (*insn_data[icode].operand[0].predicate) (target, tmode))
6802 target = gen_reg_rtx (tmode);
6803
6804 if (! (*insn_data[icode].operand[1].predicate) (op0, mode0))
6805 op0 = copy_to_mode_reg (mode0, op0);
6806
6807 scratch1 = gen_reg_rtx (mode0);
6808 scratch2 = gen_reg_rtx (mode0);
6809
6810 pat = GEN_FCN (icode) (target, op0, scratch1, scratch2);
6811 if (! pat)
6812 return 0;
6813 emit_insn (pat);
6814
6815 return target;
6816 }
6817
6818 static rtx
rs6000_expand_binop_builtin(enum insn_code icode,tree arglist,rtx target)6819 rs6000_expand_binop_builtin (enum insn_code icode, tree arglist, rtx target)
6820 {
6821 rtx pat;
6822 tree arg0 = TREE_VALUE (arglist);
6823 tree arg1 = TREE_VALUE (TREE_CHAIN (arglist));
6824 rtx op0 = expand_normal (arg0);
6825 rtx op1 = expand_normal (arg1);
6826 enum machine_mode tmode = insn_data[icode].operand[0].mode;
6827 enum machine_mode mode0 = insn_data[icode].operand[1].mode;
6828 enum machine_mode mode1 = insn_data[icode].operand[2].mode;
6829
6830 if (icode == CODE_FOR_nothing)
6831 /* Builtin not supported on this processor. */
6832 return 0;
6833
6834 /* If we got invalid arguments bail out before generating bad rtl. */
6835 if (arg0 == error_mark_node || arg1 == error_mark_node)
6836 return const0_rtx;
6837
6838 if (icode == CODE_FOR_altivec_vcfux
6839 || icode == CODE_FOR_altivec_vcfsx
6840 || icode == CODE_FOR_altivec_vctsxs
6841 || icode == CODE_FOR_altivec_vctuxs
6842 || icode == CODE_FOR_altivec_vspltb
6843 || icode == CODE_FOR_altivec_vsplth
6844 || icode == CODE_FOR_altivec_vspltw
6845 || icode == CODE_FOR_spe_evaddiw
6846 || icode == CODE_FOR_spe_evldd
6847 || icode == CODE_FOR_spe_evldh
6848 || icode == CODE_FOR_spe_evldw
6849 || icode == CODE_FOR_spe_evlhhesplat
6850 || icode == CODE_FOR_spe_evlhhossplat
6851 || icode == CODE_FOR_spe_evlhhousplat
6852 || icode == CODE_FOR_spe_evlwhe
6853 || icode == CODE_FOR_spe_evlwhos
6854 || icode == CODE_FOR_spe_evlwhou
6855 || icode == CODE_FOR_spe_evlwhsplat
6856 || icode == CODE_FOR_spe_evlwwsplat
6857 || icode == CODE_FOR_spe_evrlwi
6858 || icode == CODE_FOR_spe_evslwi
6859 || icode == CODE_FOR_spe_evsrwis
6860 || icode == CODE_FOR_spe_evsubifw
6861 || icode == CODE_FOR_spe_evsrwiu)
6862 {
6863 /* Only allow 5-bit unsigned literals. */
6864 STRIP_NOPS (arg1);
6865 if (TREE_CODE (arg1) != INTEGER_CST
6866 || TREE_INT_CST_LOW (arg1) & ~0x1f)
6867 {
6868 error ("argument 2 must be a 5-bit unsigned literal");
6869 return const0_rtx;
6870 }
6871 }
6872
6873 if (target == 0
6874 || GET_MODE (target) != tmode
6875 || ! (*insn_data[icode].operand[0].predicate) (target, tmode))
6876 target = gen_reg_rtx (tmode);
6877
6878 if (! (*insn_data[icode].operand[1].predicate) (op0, mode0))
6879 op0 = copy_to_mode_reg (mode0, op0);
6880 if (! (*insn_data[icode].operand[2].predicate) (op1, mode1))
6881 op1 = copy_to_mode_reg (mode1, op1);
6882
6883 pat = GEN_FCN (icode) (target, op0, op1);
6884 if (! pat)
6885 return 0;
6886 emit_insn (pat);
6887
6888 return target;
6889 }
6890
6891 static rtx
altivec_expand_predicate_builtin(enum insn_code icode,const char * opcode,tree arglist,rtx target)6892 altivec_expand_predicate_builtin (enum insn_code icode, const char *opcode,
6893 tree arglist, rtx target)
6894 {
6895 rtx pat, scratch;
6896 tree cr6_form = TREE_VALUE (arglist);
6897 tree arg0 = TREE_VALUE (TREE_CHAIN (arglist));
6898 tree arg1 = TREE_VALUE (TREE_CHAIN (TREE_CHAIN (arglist)));
6899 rtx op0 = expand_normal (arg0);
6900 rtx op1 = expand_normal (arg1);
6901 enum machine_mode tmode = SImode;
6902 enum machine_mode mode0 = insn_data[icode].operand[1].mode;
6903 enum machine_mode mode1 = insn_data[icode].operand[2].mode;
6904 int cr6_form_int;
6905
6906 if (TREE_CODE (cr6_form) != INTEGER_CST)
6907 {
6908 error ("argument 1 of __builtin_altivec_predicate must be a constant");
6909 return const0_rtx;
6910 }
6911 else
6912 cr6_form_int = TREE_INT_CST_LOW (cr6_form);
6913
6914 gcc_assert (mode0 == mode1);
6915
6916 /* If we have invalid arguments, bail out before generating bad rtl. */
6917 if (arg0 == error_mark_node || arg1 == error_mark_node)
6918 return const0_rtx;
6919
6920 if (target == 0
6921 || GET_MODE (target) != tmode
6922 || ! (*insn_data[icode].operand[0].predicate) (target, tmode))
6923 target = gen_reg_rtx (tmode);
6924
6925 if (! (*insn_data[icode].operand[1].predicate) (op0, mode0))
6926 op0 = copy_to_mode_reg (mode0, op0);
6927 if (! (*insn_data[icode].operand[2].predicate) (op1, mode1))
6928 op1 = copy_to_mode_reg (mode1, op1);
6929
6930 scratch = gen_reg_rtx (mode0);
6931
6932 pat = GEN_FCN (icode) (scratch, op0, op1,
6933 gen_rtx_SYMBOL_REF (Pmode, opcode));
6934 if (! pat)
6935 return 0;
6936 emit_insn (pat);
6937
6938 /* The vec_any* and vec_all* predicates use the same opcodes for two
6939 different operations, but the bits in CR6 will be different
6940 depending on what information we want. So we have to play tricks
6941 with CR6 to get the right bits out.
6942
6943 If you think this is disgusting, look at the specs for the
6944 AltiVec predicates. */
6945
6946 switch (cr6_form_int)
6947 {
6948 case 0:
6949 emit_insn (gen_cr6_test_for_zero (target));
6950 break;
6951 case 1:
6952 emit_insn (gen_cr6_test_for_zero_reverse (target));
6953 break;
6954 case 2:
6955 emit_insn (gen_cr6_test_for_lt (target));
6956 break;
6957 case 3:
6958 emit_insn (gen_cr6_test_for_lt_reverse (target));
6959 break;
6960 default:
6961 error ("argument 1 of __builtin_altivec_predicate is out of range");
6962 break;
6963 }
6964
6965 return target;
6966 }
6967
6968 static rtx
altivec_expand_lv_builtin(enum insn_code icode,tree arglist,rtx target)6969 altivec_expand_lv_builtin (enum insn_code icode, tree arglist, rtx target)
6970 {
6971 rtx pat, addr;
6972 tree arg0 = TREE_VALUE (arglist);
6973 tree arg1 = TREE_VALUE (TREE_CHAIN (arglist));
6974 enum machine_mode tmode = insn_data[icode].operand[0].mode;
6975 enum machine_mode mode0 = Pmode;
6976 enum machine_mode mode1 = Pmode;
6977 rtx op0 = expand_normal (arg0);
6978 rtx op1 = expand_normal (arg1);
6979
6980 if (icode == CODE_FOR_nothing)
6981 /* Builtin not supported on this processor. */
6982 return 0;
6983
6984 /* If we got invalid arguments bail out before generating bad rtl. */
6985 if (arg0 == error_mark_node || arg1 == error_mark_node)
6986 return const0_rtx;
6987
6988 if (target == 0
6989 || GET_MODE (target) != tmode
6990 || ! (*insn_data[icode].operand[0].predicate) (target, tmode))
6991 target = gen_reg_rtx (tmode);
6992
6993 op1 = copy_to_mode_reg (mode1, op1);
6994
6995 if (op0 == const0_rtx)
6996 {
6997 addr = gen_rtx_MEM (tmode, op1);
6998 }
6999 else
7000 {
7001 op0 = copy_to_mode_reg (mode0, op0);
7002 addr = gen_rtx_MEM (tmode, gen_rtx_PLUS (Pmode, op0, op1));
7003 }
7004
7005 pat = GEN_FCN (icode) (target, addr);
7006
7007 if (! pat)
7008 return 0;
7009 emit_insn (pat);
7010
7011 return target;
7012 }
7013
7014 static rtx
spe_expand_stv_builtin(enum insn_code icode,tree arglist)7015 spe_expand_stv_builtin (enum insn_code icode, tree arglist)
7016 {
7017 tree arg0 = TREE_VALUE (arglist);
7018 tree arg1 = TREE_VALUE (TREE_CHAIN (arglist));
7019 tree arg2 = TREE_VALUE (TREE_CHAIN (TREE_CHAIN (arglist)));
7020 rtx op0 = expand_normal (arg0);
7021 rtx op1 = expand_normal (arg1);
7022 rtx op2 = expand_normal (arg2);
7023 rtx pat;
7024 enum machine_mode mode0 = insn_data[icode].operand[0].mode;
7025 enum machine_mode mode1 = insn_data[icode].operand[1].mode;
7026 enum machine_mode mode2 = insn_data[icode].operand[2].mode;
7027
7028 /* Invalid arguments. Bail before doing anything stoopid! */
7029 if (arg0 == error_mark_node
7030 || arg1 == error_mark_node
7031 || arg2 == error_mark_node)
7032 return const0_rtx;
7033
7034 if (! (*insn_data[icode].operand[2].predicate) (op0, mode2))
7035 op0 = copy_to_mode_reg (mode2, op0);
7036 if (! (*insn_data[icode].operand[0].predicate) (op1, mode0))
7037 op1 = copy_to_mode_reg (mode0, op1);
7038 if (! (*insn_data[icode].operand[1].predicate) (op2, mode1))
7039 op2 = copy_to_mode_reg (mode1, op2);
7040
7041 pat = GEN_FCN (icode) (op1, op2, op0);
7042 if (pat)
7043 emit_insn (pat);
7044 return NULL_RTX;
7045 }
7046
7047 static rtx
altivec_expand_stv_builtin(enum insn_code icode,tree arglist)7048 altivec_expand_stv_builtin (enum insn_code icode, tree arglist)
7049 {
7050 tree arg0 = TREE_VALUE (arglist);
7051 tree arg1 = TREE_VALUE (TREE_CHAIN (arglist));
7052 tree arg2 = TREE_VALUE (TREE_CHAIN (TREE_CHAIN (arglist)));
7053 rtx op0 = expand_normal (arg0);
7054 rtx op1 = expand_normal (arg1);
7055 rtx op2 = expand_normal (arg2);
7056 rtx pat, addr;
7057 enum machine_mode tmode = insn_data[icode].operand[0].mode;
7058 enum machine_mode mode1 = Pmode;
7059 enum machine_mode mode2 = Pmode;
7060
7061 /* Invalid arguments. Bail before doing anything stoopid! */
7062 if (arg0 == error_mark_node
7063 || arg1 == error_mark_node
7064 || arg2 == error_mark_node)
7065 return const0_rtx;
7066
7067 if (! (*insn_data[icode].operand[1].predicate) (op0, tmode))
7068 op0 = copy_to_mode_reg (tmode, op0);
7069
7070 op2 = copy_to_mode_reg (mode2, op2);
7071
7072 if (op1 == const0_rtx)
7073 {
7074 addr = gen_rtx_MEM (tmode, op2);
7075 }
7076 else
7077 {
7078 op1 = copy_to_mode_reg (mode1, op1);
7079 addr = gen_rtx_MEM (tmode, gen_rtx_PLUS (Pmode, op1, op2));
7080 }
7081
7082 pat = GEN_FCN (icode) (addr, op0);
7083 if (pat)
7084 emit_insn (pat);
7085 return NULL_RTX;
7086 }
7087
7088 static rtx
rs6000_expand_ternop_builtin(enum insn_code icode,tree arglist,rtx target)7089 rs6000_expand_ternop_builtin (enum insn_code icode, tree arglist, rtx target)
7090 {
7091 rtx pat;
7092 tree arg0 = TREE_VALUE (arglist);
7093 tree arg1 = TREE_VALUE (TREE_CHAIN (arglist));
7094 tree arg2 = TREE_VALUE (TREE_CHAIN (TREE_CHAIN (arglist)));
7095 rtx op0 = expand_normal (arg0);
7096 rtx op1 = expand_normal (arg1);
7097 rtx op2 = expand_normal (arg2);
7098 enum machine_mode tmode = insn_data[icode].operand[0].mode;
7099 enum machine_mode mode0 = insn_data[icode].operand[1].mode;
7100 enum machine_mode mode1 = insn_data[icode].operand[2].mode;
7101 enum machine_mode mode2 = insn_data[icode].operand[3].mode;
7102
7103 if (icode == CODE_FOR_nothing)
7104 /* Builtin not supported on this processor. */
7105 return 0;
7106
7107 /* If we got invalid arguments bail out before generating bad rtl. */
7108 if (arg0 == error_mark_node
7109 || arg1 == error_mark_node
7110 || arg2 == error_mark_node)
7111 return const0_rtx;
7112
7113 if (icode == CODE_FOR_altivec_vsldoi_v4sf
7114 || icode == CODE_FOR_altivec_vsldoi_v4si
7115 || icode == CODE_FOR_altivec_vsldoi_v8hi
7116 || icode == CODE_FOR_altivec_vsldoi_v16qi)
7117 {
7118 /* Only allow 4-bit unsigned literals. */
7119 STRIP_NOPS (arg2);
7120 if (TREE_CODE (arg2) != INTEGER_CST
7121 || TREE_INT_CST_LOW (arg2) & ~0xf)
7122 {
7123 error ("argument 3 must be a 4-bit unsigned literal");
7124 return const0_rtx;
7125 }
7126 }
7127
7128 if (target == 0
7129 || GET_MODE (target) != tmode
7130 || ! (*insn_data[icode].operand[0].predicate) (target, tmode))
7131 target = gen_reg_rtx (tmode);
7132
7133 if (! (*insn_data[icode].operand[1].predicate) (op0, mode0))
7134 op0 = copy_to_mode_reg (mode0, op0);
7135 if (! (*insn_data[icode].operand[2].predicate) (op1, mode1))
7136 op1 = copy_to_mode_reg (mode1, op1);
7137 if (! (*insn_data[icode].operand[3].predicate) (op2, mode2))
7138 op2 = copy_to_mode_reg (mode2, op2);
7139
7140 pat = GEN_FCN (icode) (target, op0, op1, op2);
7141 if (! pat)
7142 return 0;
7143 emit_insn (pat);
7144
7145 return target;
7146 }
7147
7148 /* Expand the lvx builtins. */
7149 static rtx
altivec_expand_ld_builtin(tree exp,rtx target,bool * expandedp)7150 altivec_expand_ld_builtin (tree exp, rtx target, bool *expandedp)
7151 {
7152 tree fndecl = TREE_OPERAND (TREE_OPERAND (exp, 0), 0);
7153 tree arglist = TREE_OPERAND (exp, 1);
7154 unsigned int fcode = DECL_FUNCTION_CODE (fndecl);
7155 tree arg0;
7156 enum machine_mode tmode, mode0;
7157 rtx pat, op0;
7158 enum insn_code icode;
7159
7160 switch (fcode)
7161 {
7162 case ALTIVEC_BUILTIN_LD_INTERNAL_16qi:
7163 icode = CODE_FOR_altivec_lvx_v16qi;
7164 break;
7165 case ALTIVEC_BUILTIN_LD_INTERNAL_8hi:
7166 icode = CODE_FOR_altivec_lvx_v8hi;
7167 break;
7168 case ALTIVEC_BUILTIN_LD_INTERNAL_4si:
7169 icode = CODE_FOR_altivec_lvx_v4si;
7170 break;
7171 case ALTIVEC_BUILTIN_LD_INTERNAL_4sf:
7172 icode = CODE_FOR_altivec_lvx_v4sf;
7173 break;
7174 default:
7175 *expandedp = false;
7176 return NULL_RTX;
7177 }
7178
7179 *expandedp = true;
7180
7181 arg0 = TREE_VALUE (arglist);
7182 op0 = expand_normal (arg0);
7183 tmode = insn_data[icode].operand[0].mode;
7184 mode0 = insn_data[icode].operand[1].mode;
7185
7186 if (target == 0
7187 || GET_MODE (target) != tmode
7188 || ! (*insn_data[icode].operand[0].predicate) (target, tmode))
7189 target = gen_reg_rtx (tmode);
7190
7191 if (! (*insn_data[icode].operand[1].predicate) (op0, mode0))
7192 op0 = gen_rtx_MEM (mode0, copy_to_mode_reg (Pmode, op0));
7193
7194 pat = GEN_FCN (icode) (target, op0);
7195 if (! pat)
7196 return 0;
7197 emit_insn (pat);
7198 return target;
7199 }
7200
7201 /* Expand the stvx builtins. */
7202 static rtx
altivec_expand_st_builtin(tree exp,rtx target ATTRIBUTE_UNUSED,bool * expandedp)7203 altivec_expand_st_builtin (tree exp, rtx target ATTRIBUTE_UNUSED,
7204 bool *expandedp)
7205 {
7206 tree fndecl = TREE_OPERAND (TREE_OPERAND (exp, 0), 0);
7207 tree arglist = TREE_OPERAND (exp, 1);
7208 unsigned int fcode = DECL_FUNCTION_CODE (fndecl);
7209 tree arg0, arg1;
7210 enum machine_mode mode0, mode1;
7211 rtx pat, op0, op1;
7212 enum insn_code icode;
7213
7214 switch (fcode)
7215 {
7216 case ALTIVEC_BUILTIN_ST_INTERNAL_16qi:
7217 icode = CODE_FOR_altivec_stvx_v16qi;
7218 break;
7219 case ALTIVEC_BUILTIN_ST_INTERNAL_8hi:
7220 icode = CODE_FOR_altivec_stvx_v8hi;
7221 break;
7222 case ALTIVEC_BUILTIN_ST_INTERNAL_4si:
7223 icode = CODE_FOR_altivec_stvx_v4si;
7224 break;
7225 case ALTIVEC_BUILTIN_ST_INTERNAL_4sf:
7226 icode = CODE_FOR_altivec_stvx_v4sf;
7227 break;
7228 default:
7229 *expandedp = false;
7230 return NULL_RTX;
7231 }
7232
7233 arg0 = TREE_VALUE (arglist);
7234 arg1 = TREE_VALUE (TREE_CHAIN (arglist));
7235 op0 = expand_normal (arg0);
7236 op1 = expand_normal (arg1);
7237 mode0 = insn_data[icode].operand[0].mode;
7238 mode1 = insn_data[icode].operand[1].mode;
7239
7240 if (! (*insn_data[icode].operand[0].predicate) (op0, mode0))
7241 op0 = gen_rtx_MEM (mode0, copy_to_mode_reg (Pmode, op0));
7242 if (! (*insn_data[icode].operand[1].predicate) (op1, mode1))
7243 op1 = copy_to_mode_reg (mode1, op1);
7244
7245 pat = GEN_FCN (icode) (op0, op1);
7246 if (pat)
7247 emit_insn (pat);
7248
7249 *expandedp = true;
7250 return NULL_RTX;
7251 }
7252
7253 /* Expand the dst builtins. */
7254 static rtx
altivec_expand_dst_builtin(tree exp,rtx target ATTRIBUTE_UNUSED,bool * expandedp)7255 altivec_expand_dst_builtin (tree exp, rtx target ATTRIBUTE_UNUSED,
7256 bool *expandedp)
7257 {
7258 tree fndecl = TREE_OPERAND (TREE_OPERAND (exp, 0), 0);
7259 tree arglist = TREE_OPERAND (exp, 1);
7260 unsigned int fcode = DECL_FUNCTION_CODE (fndecl);
7261 tree arg0, arg1, arg2;
7262 enum machine_mode mode0, mode1, mode2;
7263 rtx pat, op0, op1, op2;
7264 struct builtin_description *d;
7265 size_t i;
7266
7267 *expandedp = false;
7268
7269 /* Handle DST variants. */
7270 d = (struct builtin_description *) bdesc_dst;
7271 for (i = 0; i < ARRAY_SIZE (bdesc_dst); i++, d++)
7272 if (d->code == fcode)
7273 {
7274 arg0 = TREE_VALUE (arglist);
7275 arg1 = TREE_VALUE (TREE_CHAIN (arglist));
7276 arg2 = TREE_VALUE (TREE_CHAIN (TREE_CHAIN (arglist)));
7277 op0 = expand_normal (arg0);
7278 op1 = expand_normal (arg1);
7279 op2 = expand_normal (arg2);
7280 mode0 = insn_data[d->icode].operand[0].mode;
7281 mode1 = insn_data[d->icode].operand[1].mode;
7282 mode2 = insn_data[d->icode].operand[2].mode;
7283
7284 /* Invalid arguments, bail out before generating bad rtl. */
7285 if (arg0 == error_mark_node
7286 || arg1 == error_mark_node
7287 || arg2 == error_mark_node)
7288 return const0_rtx;
7289
7290 *expandedp = true;
7291 STRIP_NOPS (arg2);
7292 if (TREE_CODE (arg2) != INTEGER_CST
7293 || TREE_INT_CST_LOW (arg2) & ~0x3)
7294 {
7295 error ("argument to %qs must be a 2-bit unsigned literal", d->name);
7296 return const0_rtx;
7297 }
7298
7299 if (! (*insn_data[d->icode].operand[0].predicate) (op0, mode0))
7300 op0 = copy_to_mode_reg (Pmode, op0);
7301 if (! (*insn_data[d->icode].operand[1].predicate) (op1, mode1))
7302 op1 = copy_to_mode_reg (mode1, op1);
7303
7304 pat = GEN_FCN (d->icode) (op0, op1, op2);
7305 if (pat != 0)
7306 emit_insn (pat);
7307
7308 return NULL_RTX;
7309 }
7310
7311 return NULL_RTX;
7312 }
7313
7314 /* Expand vec_init builtin. */
7315 static rtx
altivec_expand_vec_init_builtin(tree type,tree arglist,rtx target)7316 altivec_expand_vec_init_builtin (tree type, tree arglist, rtx target)
7317 {
7318 enum machine_mode tmode = TYPE_MODE (type);
7319 enum machine_mode inner_mode = GET_MODE_INNER (tmode);
7320 int i, n_elt = GET_MODE_NUNITS (tmode);
7321 rtvec v = rtvec_alloc (n_elt);
7322
7323 gcc_assert (VECTOR_MODE_P (tmode));
7324
7325 for (i = 0; i < n_elt; ++i, arglist = TREE_CHAIN (arglist))
7326 {
7327 rtx x = expand_normal (TREE_VALUE (arglist));
7328 RTVEC_ELT (v, i) = gen_lowpart (inner_mode, x);
7329 }
7330
7331 gcc_assert (arglist == NULL);
7332
7333 if (!target || !register_operand (target, tmode))
7334 target = gen_reg_rtx (tmode);
7335
7336 rs6000_expand_vector_init (target, gen_rtx_PARALLEL (tmode, v));
7337 return target;
7338 }
7339
7340 /* Return the integer constant in ARG. Constrain it to be in the range
7341 of the subparts of VEC_TYPE; issue an error if not. */
7342
7343 static int
get_element_number(tree vec_type,tree arg)7344 get_element_number (tree vec_type, tree arg)
7345 {
7346 unsigned HOST_WIDE_INT elt, max = TYPE_VECTOR_SUBPARTS (vec_type) - 1;
7347
7348 if (!host_integerp (arg, 1)
7349 || (elt = tree_low_cst (arg, 1), elt > max))
7350 {
7351 error ("selector must be an integer constant in the range 0..%wi", max);
7352 return 0;
7353 }
7354
7355 return elt;
7356 }
7357
7358 /* Expand vec_set builtin. */
7359 static rtx
altivec_expand_vec_set_builtin(tree arglist)7360 altivec_expand_vec_set_builtin (tree arglist)
7361 {
7362 enum machine_mode tmode, mode1;
7363 tree arg0, arg1, arg2;
7364 int elt;
7365 rtx op0, op1;
7366
7367 arg0 = TREE_VALUE (arglist);
7368 arg1 = TREE_VALUE (TREE_CHAIN (arglist));
7369 arg2 = TREE_VALUE (TREE_CHAIN (TREE_CHAIN (arglist)));
7370
7371 tmode = TYPE_MODE (TREE_TYPE (arg0));
7372 mode1 = TYPE_MODE (TREE_TYPE (TREE_TYPE (arg0)));
7373 gcc_assert (VECTOR_MODE_P (tmode));
7374
7375 op0 = expand_expr (arg0, NULL_RTX, tmode, 0);
7376 op1 = expand_expr (arg1, NULL_RTX, mode1, 0);
7377 elt = get_element_number (TREE_TYPE (arg0), arg2);
7378
7379 if (GET_MODE (op1) != mode1 && GET_MODE (op1) != VOIDmode)
7380 op1 = convert_modes (mode1, GET_MODE (op1), op1, true);
7381
7382 op0 = force_reg (tmode, op0);
7383 op1 = force_reg (mode1, op1);
7384
7385 rs6000_expand_vector_set (op0, op1, elt);
7386
7387 return op0;
7388 }
7389
7390 /* Expand vec_ext builtin. */
7391 static rtx
altivec_expand_vec_ext_builtin(tree arglist,rtx target)7392 altivec_expand_vec_ext_builtin (tree arglist, rtx target)
7393 {
7394 enum machine_mode tmode, mode0;
7395 tree arg0, arg1;
7396 int elt;
7397 rtx op0;
7398
7399 arg0 = TREE_VALUE (arglist);
7400 arg1 = TREE_VALUE (TREE_CHAIN (arglist));
7401
7402 op0 = expand_normal (arg0);
7403 elt = get_element_number (TREE_TYPE (arg0), arg1);
7404
7405 tmode = TYPE_MODE (TREE_TYPE (TREE_TYPE (arg0)));
7406 mode0 = TYPE_MODE (TREE_TYPE (arg0));
7407 gcc_assert (VECTOR_MODE_P (mode0));
7408
7409 op0 = force_reg (mode0, op0);
7410
7411 if (optimize || !target || !register_operand (target, tmode))
7412 target = gen_reg_rtx (tmode);
7413
7414 rs6000_expand_vector_extract (target, op0, elt);
7415
7416 return target;
7417 }
7418
7419 /* Expand the builtin in EXP and store the result in TARGET. Store
7420 true in *EXPANDEDP if we found a builtin to expand. */
7421 static rtx
altivec_expand_builtin(tree exp,rtx target,bool * expandedp)7422 altivec_expand_builtin (tree exp, rtx target, bool *expandedp)
7423 {
7424 struct builtin_description *d;
7425 struct builtin_description_predicates *dp;
7426 size_t i;
7427 enum insn_code icode;
7428 tree fndecl = TREE_OPERAND (TREE_OPERAND (exp, 0), 0);
7429 tree arglist = TREE_OPERAND (exp, 1);
7430 tree arg0;
7431 rtx op0, pat;
7432 enum machine_mode tmode, mode0;
7433 unsigned int fcode = DECL_FUNCTION_CODE (fndecl);
7434
7435 if (fcode >= ALTIVEC_BUILTIN_OVERLOADED_FIRST
7436 && fcode <= ALTIVEC_BUILTIN_OVERLOADED_LAST)
7437 {
7438 *expandedp = true;
7439 error ("unresolved overload for Altivec builtin %qF", fndecl);
7440 return const0_rtx;
7441 }
7442
7443 target = altivec_expand_ld_builtin (exp, target, expandedp);
7444 if (*expandedp)
7445 return target;
7446
7447 target = altivec_expand_st_builtin (exp, target, expandedp);
7448 if (*expandedp)
7449 return target;
7450
7451 target = altivec_expand_dst_builtin (exp, target, expandedp);
7452 if (*expandedp)
7453 return target;
7454
7455 *expandedp = true;
7456
7457 switch (fcode)
7458 {
7459 case ALTIVEC_BUILTIN_STVX:
7460 return altivec_expand_stv_builtin (CODE_FOR_altivec_stvx, arglist);
7461 case ALTIVEC_BUILTIN_STVEBX:
7462 return altivec_expand_stv_builtin (CODE_FOR_altivec_stvebx, arglist);
7463 case ALTIVEC_BUILTIN_STVEHX:
7464 return altivec_expand_stv_builtin (CODE_FOR_altivec_stvehx, arglist);
7465 case ALTIVEC_BUILTIN_STVEWX:
7466 return altivec_expand_stv_builtin (CODE_FOR_altivec_stvewx, arglist);
7467 case ALTIVEC_BUILTIN_STVXL:
7468 return altivec_expand_stv_builtin (CODE_FOR_altivec_stvxl, arglist);
7469
7470 case ALTIVEC_BUILTIN_MFVSCR:
7471 icode = CODE_FOR_altivec_mfvscr;
7472 tmode = insn_data[icode].operand[0].mode;
7473
7474 if (target == 0
7475 || GET_MODE (target) != tmode
7476 || ! (*insn_data[icode].operand[0].predicate) (target, tmode))
7477 target = gen_reg_rtx (tmode);
7478
7479 pat = GEN_FCN (icode) (target);
7480 if (! pat)
7481 return 0;
7482 emit_insn (pat);
7483 return target;
7484
7485 case ALTIVEC_BUILTIN_MTVSCR:
7486 icode = CODE_FOR_altivec_mtvscr;
7487 arg0 = TREE_VALUE (arglist);
7488 op0 = expand_normal (arg0);
7489 mode0 = insn_data[icode].operand[0].mode;
7490
7491 /* If we got invalid arguments bail out before generating bad rtl. */
7492 if (arg0 == error_mark_node)
7493 return const0_rtx;
7494
7495 if (! (*insn_data[icode].operand[0].predicate) (op0, mode0))
7496 op0 = copy_to_mode_reg (mode0, op0);
7497
7498 pat = GEN_FCN (icode) (op0);
7499 if (pat)
7500 emit_insn (pat);
7501 return NULL_RTX;
7502
7503 case ALTIVEC_BUILTIN_DSSALL:
7504 emit_insn (gen_altivec_dssall ());
7505 return NULL_RTX;
7506
7507 case ALTIVEC_BUILTIN_DSS:
7508 icode = CODE_FOR_altivec_dss;
7509 arg0 = TREE_VALUE (arglist);
7510 STRIP_NOPS (arg0);
7511 op0 = expand_normal (arg0);
7512 mode0 = insn_data[icode].operand[0].mode;
7513
7514 /* If we got invalid arguments bail out before generating bad rtl. */
7515 if (arg0 == error_mark_node)
7516 return const0_rtx;
7517
7518 if (TREE_CODE (arg0) != INTEGER_CST
7519 || TREE_INT_CST_LOW (arg0) & ~0x3)
7520 {
7521 error ("argument to dss must be a 2-bit unsigned literal");
7522 return const0_rtx;
7523 }
7524
7525 if (! (*insn_data[icode].operand[0].predicate) (op0, mode0))
7526 op0 = copy_to_mode_reg (mode0, op0);
7527
7528 emit_insn (gen_altivec_dss (op0));
7529 return NULL_RTX;
7530
7531 case ALTIVEC_BUILTIN_VEC_INIT_V4SI:
7532 case ALTIVEC_BUILTIN_VEC_INIT_V8HI:
7533 case ALTIVEC_BUILTIN_VEC_INIT_V16QI:
7534 case ALTIVEC_BUILTIN_VEC_INIT_V4SF:
7535 return altivec_expand_vec_init_builtin (TREE_TYPE (exp), arglist, target);
7536
7537 case ALTIVEC_BUILTIN_VEC_SET_V4SI:
7538 case ALTIVEC_BUILTIN_VEC_SET_V8HI:
7539 case ALTIVEC_BUILTIN_VEC_SET_V16QI:
7540 case ALTIVEC_BUILTIN_VEC_SET_V4SF:
7541 return altivec_expand_vec_set_builtin (arglist);
7542
7543 case ALTIVEC_BUILTIN_VEC_EXT_V4SI:
7544 case ALTIVEC_BUILTIN_VEC_EXT_V8HI:
7545 case ALTIVEC_BUILTIN_VEC_EXT_V16QI:
7546 case ALTIVEC_BUILTIN_VEC_EXT_V4SF:
7547 return altivec_expand_vec_ext_builtin (arglist, target);
7548
7549 default:
7550 break;
7551 /* Fall through. */
7552 }
7553
7554 /* Expand abs* operations. */
7555 d = (struct builtin_description *) bdesc_abs;
7556 for (i = 0; i < ARRAY_SIZE (bdesc_abs); i++, d++)
7557 if (d->code == fcode)
7558 return altivec_expand_abs_builtin (d->icode, arglist, target);
7559
7560 /* Expand the AltiVec predicates. */
7561 dp = (struct builtin_description_predicates *) bdesc_altivec_preds;
7562 for (i = 0; i < ARRAY_SIZE (bdesc_altivec_preds); i++, dp++)
7563 if (dp->code == fcode)
7564 return altivec_expand_predicate_builtin (dp->icode, dp->opcode,
7565 arglist, target);
7566
7567 /* LV* are funky. We initialized them differently. */
7568 switch (fcode)
7569 {
7570 case ALTIVEC_BUILTIN_LVSL:
7571 return altivec_expand_lv_builtin (CODE_FOR_altivec_lvsl,
7572 arglist, target);
7573 case ALTIVEC_BUILTIN_LVSR:
7574 return altivec_expand_lv_builtin (CODE_FOR_altivec_lvsr,
7575 arglist, target);
7576 case ALTIVEC_BUILTIN_LVEBX:
7577 return altivec_expand_lv_builtin (CODE_FOR_altivec_lvebx,
7578 arglist, target);
7579 case ALTIVEC_BUILTIN_LVEHX:
7580 return altivec_expand_lv_builtin (CODE_FOR_altivec_lvehx,
7581 arglist, target);
7582 case ALTIVEC_BUILTIN_LVEWX:
7583 return altivec_expand_lv_builtin (CODE_FOR_altivec_lvewx,
7584 arglist, target);
7585 case ALTIVEC_BUILTIN_LVXL:
7586 return altivec_expand_lv_builtin (CODE_FOR_altivec_lvxl,
7587 arglist, target);
7588 case ALTIVEC_BUILTIN_LVX:
7589 return altivec_expand_lv_builtin (CODE_FOR_altivec_lvx,
7590 arglist, target);
7591 default:
7592 break;
7593 /* Fall through. */
7594 }
7595
7596 *expandedp = false;
7597 return NULL_RTX;
7598 }
7599
7600 /* Binops that need to be initialized manually, but can be expanded
7601 automagically by rs6000_expand_binop_builtin. */
7602 static struct builtin_description bdesc_2arg_spe[] =
7603 {
7604 { 0, CODE_FOR_spe_evlddx, "__builtin_spe_evlddx", SPE_BUILTIN_EVLDDX },
7605 { 0, CODE_FOR_spe_evldwx, "__builtin_spe_evldwx", SPE_BUILTIN_EVLDWX },
7606 { 0, CODE_FOR_spe_evldhx, "__builtin_spe_evldhx", SPE_BUILTIN_EVLDHX },
7607 { 0, CODE_FOR_spe_evlwhex, "__builtin_spe_evlwhex", SPE_BUILTIN_EVLWHEX },
7608 { 0, CODE_FOR_spe_evlwhoux, "__builtin_spe_evlwhoux", SPE_BUILTIN_EVLWHOUX },
7609 { 0, CODE_FOR_spe_evlwhosx, "__builtin_spe_evlwhosx", SPE_BUILTIN_EVLWHOSX },
7610 { 0, CODE_FOR_spe_evlwwsplatx, "__builtin_spe_evlwwsplatx", SPE_BUILTIN_EVLWWSPLATX },
7611 { 0, CODE_FOR_spe_evlwhsplatx, "__builtin_spe_evlwhsplatx", SPE_BUILTIN_EVLWHSPLATX },
7612 { 0, CODE_FOR_spe_evlhhesplatx, "__builtin_spe_evlhhesplatx", SPE_BUILTIN_EVLHHESPLATX },
7613 { 0, CODE_FOR_spe_evlhhousplatx, "__builtin_spe_evlhhousplatx", SPE_BUILTIN_EVLHHOUSPLATX },
7614 { 0, CODE_FOR_spe_evlhhossplatx, "__builtin_spe_evlhhossplatx", SPE_BUILTIN_EVLHHOSSPLATX },
7615 { 0, CODE_FOR_spe_evldd, "__builtin_spe_evldd", SPE_BUILTIN_EVLDD },
7616 { 0, CODE_FOR_spe_evldw, "__builtin_spe_evldw", SPE_BUILTIN_EVLDW },
7617 { 0, CODE_FOR_spe_evldh, "__builtin_spe_evldh", SPE_BUILTIN_EVLDH },
7618 { 0, CODE_FOR_spe_evlwhe, "__builtin_spe_evlwhe", SPE_BUILTIN_EVLWHE },
7619 { 0, CODE_FOR_spe_evlwhou, "__builtin_spe_evlwhou", SPE_BUILTIN_EVLWHOU },
7620 { 0, CODE_FOR_spe_evlwhos, "__builtin_spe_evlwhos", SPE_BUILTIN_EVLWHOS },
7621 { 0, CODE_FOR_spe_evlwwsplat, "__builtin_spe_evlwwsplat", SPE_BUILTIN_EVLWWSPLAT },
7622 { 0, CODE_FOR_spe_evlwhsplat, "__builtin_spe_evlwhsplat", SPE_BUILTIN_EVLWHSPLAT },
7623 { 0, CODE_FOR_spe_evlhhesplat, "__builtin_spe_evlhhesplat", SPE_BUILTIN_EVLHHESPLAT },
7624 { 0, CODE_FOR_spe_evlhhousplat, "__builtin_spe_evlhhousplat", SPE_BUILTIN_EVLHHOUSPLAT },
7625 { 0, CODE_FOR_spe_evlhhossplat, "__builtin_spe_evlhhossplat", SPE_BUILTIN_EVLHHOSSPLAT }
7626 };
7627
7628 /* Expand the builtin in EXP and store the result in TARGET. Store
7629 true in *EXPANDEDP if we found a builtin to expand.
7630
7631 This expands the SPE builtins that are not simple unary and binary
7632 operations. */
7633 static rtx
spe_expand_builtin(tree exp,rtx target,bool * expandedp)7634 spe_expand_builtin (tree exp, rtx target, bool *expandedp)
7635 {
7636 tree fndecl = TREE_OPERAND (TREE_OPERAND (exp, 0), 0);
7637 tree arglist = TREE_OPERAND (exp, 1);
7638 tree arg1, arg0;
7639 unsigned int fcode = DECL_FUNCTION_CODE (fndecl);
7640 enum insn_code icode;
7641 enum machine_mode tmode, mode0;
7642 rtx pat, op0;
7643 struct builtin_description *d;
7644 size_t i;
7645
7646 *expandedp = true;
7647
7648 /* Syntax check for a 5-bit unsigned immediate. */
7649 switch (fcode)
7650 {
7651 case SPE_BUILTIN_EVSTDD:
7652 case SPE_BUILTIN_EVSTDH:
7653 case SPE_BUILTIN_EVSTDW:
7654 case SPE_BUILTIN_EVSTWHE:
7655 case SPE_BUILTIN_EVSTWHO:
7656 case SPE_BUILTIN_EVSTWWE:
7657 case SPE_BUILTIN_EVSTWWO:
7658 arg1 = TREE_VALUE (TREE_CHAIN (TREE_CHAIN (arglist)));
7659 if (TREE_CODE (arg1) != INTEGER_CST
7660 || TREE_INT_CST_LOW (arg1) & ~0x1f)
7661 {
7662 error ("argument 2 must be a 5-bit unsigned literal");
7663 return const0_rtx;
7664 }
7665 break;
7666 default:
7667 break;
7668 }
7669
7670 /* The evsplat*i instructions are not quite generic. */
7671 switch (fcode)
7672 {
7673 case SPE_BUILTIN_EVSPLATFI:
7674 return rs6000_expand_unop_builtin (CODE_FOR_spe_evsplatfi,
7675 arglist, target);
7676 case SPE_BUILTIN_EVSPLATI:
7677 return rs6000_expand_unop_builtin (CODE_FOR_spe_evsplati,
7678 arglist, target);
7679 default:
7680 break;
7681 }
7682
7683 d = (struct builtin_description *) bdesc_2arg_spe;
7684 for (i = 0; i < ARRAY_SIZE (bdesc_2arg_spe); ++i, ++d)
7685 if (d->code == fcode)
7686 return rs6000_expand_binop_builtin (d->icode, arglist, target);
7687
7688 d = (struct builtin_description *) bdesc_spe_predicates;
7689 for (i = 0; i < ARRAY_SIZE (bdesc_spe_predicates); ++i, ++d)
7690 if (d->code == fcode)
7691 return spe_expand_predicate_builtin (d->icode, arglist, target);
7692
7693 d = (struct builtin_description *) bdesc_spe_evsel;
7694 for (i = 0; i < ARRAY_SIZE (bdesc_spe_evsel); ++i, ++d)
7695 if (d->code == fcode)
7696 return spe_expand_evsel_builtin (d->icode, arglist, target);
7697
7698 switch (fcode)
7699 {
7700 case SPE_BUILTIN_EVSTDDX:
7701 return spe_expand_stv_builtin (CODE_FOR_spe_evstddx, arglist);
7702 case SPE_BUILTIN_EVSTDHX:
7703 return spe_expand_stv_builtin (CODE_FOR_spe_evstdhx, arglist);
7704 case SPE_BUILTIN_EVSTDWX:
7705 return spe_expand_stv_builtin (CODE_FOR_spe_evstdwx, arglist);
7706 case SPE_BUILTIN_EVSTWHEX:
7707 return spe_expand_stv_builtin (CODE_FOR_spe_evstwhex, arglist);
7708 case SPE_BUILTIN_EVSTWHOX:
7709 return spe_expand_stv_builtin (CODE_FOR_spe_evstwhox, arglist);
7710 case SPE_BUILTIN_EVSTWWEX:
7711 return spe_expand_stv_builtin (CODE_FOR_spe_evstwwex, arglist);
7712 case SPE_BUILTIN_EVSTWWOX:
7713 return spe_expand_stv_builtin (CODE_FOR_spe_evstwwox, arglist);
7714 case SPE_BUILTIN_EVSTDD:
7715 return spe_expand_stv_builtin (CODE_FOR_spe_evstdd, arglist);
7716 case SPE_BUILTIN_EVSTDH:
7717 return spe_expand_stv_builtin (CODE_FOR_spe_evstdh, arglist);
7718 case SPE_BUILTIN_EVSTDW:
7719 return spe_expand_stv_builtin (CODE_FOR_spe_evstdw, arglist);
7720 case SPE_BUILTIN_EVSTWHE:
7721 return spe_expand_stv_builtin (CODE_FOR_spe_evstwhe, arglist);
7722 case SPE_BUILTIN_EVSTWHO:
7723 return spe_expand_stv_builtin (CODE_FOR_spe_evstwho, arglist);
7724 case SPE_BUILTIN_EVSTWWE:
7725 return spe_expand_stv_builtin (CODE_FOR_spe_evstwwe, arglist);
7726 case SPE_BUILTIN_EVSTWWO:
7727 return spe_expand_stv_builtin (CODE_FOR_spe_evstwwo, arglist);
7728 case SPE_BUILTIN_MFSPEFSCR:
7729 icode = CODE_FOR_spe_mfspefscr;
7730 tmode = insn_data[icode].operand[0].mode;
7731
7732 if (target == 0
7733 || GET_MODE (target) != tmode
7734 || ! (*insn_data[icode].operand[0].predicate) (target, tmode))
7735 target = gen_reg_rtx (tmode);
7736
7737 pat = GEN_FCN (icode) (target);
7738 if (! pat)
7739 return 0;
7740 emit_insn (pat);
7741 return target;
7742 case SPE_BUILTIN_MTSPEFSCR:
7743 icode = CODE_FOR_spe_mtspefscr;
7744 arg0 = TREE_VALUE (arglist);
7745 op0 = expand_normal (arg0);
7746 mode0 = insn_data[icode].operand[0].mode;
7747
7748 if (arg0 == error_mark_node)
7749 return const0_rtx;
7750
7751 if (! (*insn_data[icode].operand[0].predicate) (op0, mode0))
7752 op0 = copy_to_mode_reg (mode0, op0);
7753
7754 pat = GEN_FCN (icode) (op0);
7755 if (pat)
7756 emit_insn (pat);
7757 return NULL_RTX;
7758 default:
7759 break;
7760 }
7761
7762 *expandedp = false;
7763 return NULL_RTX;
7764 }
7765
7766 static rtx
spe_expand_predicate_builtin(enum insn_code icode,tree arglist,rtx target)7767 spe_expand_predicate_builtin (enum insn_code icode, tree arglist, rtx target)
7768 {
7769 rtx pat, scratch, tmp;
7770 tree form = TREE_VALUE (arglist);
7771 tree arg0 = TREE_VALUE (TREE_CHAIN (arglist));
7772 tree arg1 = TREE_VALUE (TREE_CHAIN (TREE_CHAIN (arglist)));
7773 rtx op0 = expand_normal (arg0);
7774 rtx op1 = expand_normal (arg1);
7775 enum machine_mode mode0 = insn_data[icode].operand[1].mode;
7776 enum machine_mode mode1 = insn_data[icode].operand[2].mode;
7777 int form_int;
7778 enum rtx_code code;
7779
7780 if (TREE_CODE (form) != INTEGER_CST)
7781 {
7782 error ("argument 1 of __builtin_spe_predicate must be a constant");
7783 return const0_rtx;
7784 }
7785 else
7786 form_int = TREE_INT_CST_LOW (form);
7787
7788 gcc_assert (mode0 == mode1);
7789
7790 if (arg0 == error_mark_node || arg1 == error_mark_node)
7791 return const0_rtx;
7792
7793 if (target == 0
7794 || GET_MODE (target) != SImode
7795 || ! (*insn_data[icode].operand[0].predicate) (target, SImode))
7796 target = gen_reg_rtx (SImode);
7797
7798 if (! (*insn_data[icode].operand[1].predicate) (op0, mode0))
7799 op0 = copy_to_mode_reg (mode0, op0);
7800 if (! (*insn_data[icode].operand[2].predicate) (op1, mode1))
7801 op1 = copy_to_mode_reg (mode1, op1);
7802
7803 scratch = gen_reg_rtx (CCmode);
7804
7805 pat = GEN_FCN (icode) (scratch, op0, op1);
7806 if (! pat)
7807 return const0_rtx;
7808 emit_insn (pat);
7809
7810 /* There are 4 variants for each predicate: _any_, _all_, _upper_,
7811 _lower_. We use one compare, but look in different bits of the
7812 CR for each variant.
7813
7814 There are 2 elements in each SPE simd type (upper/lower). The CR
7815 bits are set as follows:
7816
7817 BIT0 | BIT 1 | BIT 2 | BIT 3
7818 U | L | (U | L) | (U & L)
7819
7820 So, for an "all" relationship, BIT 3 would be set.
7821 For an "any" relationship, BIT 2 would be set. Etc.
7822
7823 Following traditional nomenclature, these bits map to:
7824
7825 BIT0 | BIT 1 | BIT 2 | BIT 3
7826 LT | GT | EQ | OV
7827
7828 Later, we will generate rtl to look in the LT/EQ/EQ/OV bits.
7829 */
7830
7831 switch (form_int)
7832 {
7833 /* All variant. OV bit. */
7834 case 0:
7835 /* We need to get to the OV bit, which is the ORDERED bit. We
7836 could generate (ordered:SI (reg:CC xx) (const_int 0)), but
7837 that's ugly and will make validate_condition_mode die.
7838 So let's just use another pattern. */
7839 emit_insn (gen_move_from_CR_ov_bit (target, scratch));
7840 return target;
7841 /* Any variant. EQ bit. */
7842 case 1:
7843 code = EQ;
7844 break;
7845 /* Upper variant. LT bit. */
7846 case 2:
7847 code = LT;
7848 break;
7849 /* Lower variant. GT bit. */
7850 case 3:
7851 code = GT;
7852 break;
7853 default:
7854 error ("argument 1 of __builtin_spe_predicate is out of range");
7855 return const0_rtx;
7856 }
7857
7858 tmp = gen_rtx_fmt_ee (code, SImode, scratch, const0_rtx);
7859 emit_move_insn (target, tmp);
7860
7861 return target;
7862 }
7863
7864 /* The evsel builtins look like this:
7865
7866 e = __builtin_spe_evsel_OP (a, b, c, d);
7867
7868 and work like this:
7869
7870 e[upper] = a[upper] *OP* b[upper] ? c[upper] : d[upper];
7871 e[lower] = a[lower] *OP* b[lower] ? c[lower] : d[lower];
7872 */
7873
7874 static rtx
spe_expand_evsel_builtin(enum insn_code icode,tree arglist,rtx target)7875 spe_expand_evsel_builtin (enum insn_code icode, tree arglist, rtx target)
7876 {
7877 rtx pat, scratch;
7878 tree arg0 = TREE_VALUE (arglist);
7879 tree arg1 = TREE_VALUE (TREE_CHAIN (arglist));
7880 tree arg2 = TREE_VALUE (TREE_CHAIN (TREE_CHAIN (arglist)));
7881 tree arg3 = TREE_VALUE (TREE_CHAIN (TREE_CHAIN (TREE_CHAIN (arglist))));
7882 rtx op0 = expand_normal (arg0);
7883 rtx op1 = expand_normal (arg1);
7884 rtx op2 = expand_normal (arg2);
7885 rtx op3 = expand_normal (arg3);
7886 enum machine_mode mode0 = insn_data[icode].operand[1].mode;
7887 enum machine_mode mode1 = insn_data[icode].operand[2].mode;
7888
7889 gcc_assert (mode0 == mode1);
7890
7891 if (arg0 == error_mark_node || arg1 == error_mark_node
7892 || arg2 == error_mark_node || arg3 == error_mark_node)
7893 return const0_rtx;
7894
7895 if (target == 0
7896 || GET_MODE (target) != mode0
7897 || ! (*insn_data[icode].operand[0].predicate) (target, mode0))
7898 target = gen_reg_rtx (mode0);
7899
7900 if (! (*insn_data[icode].operand[1].predicate) (op0, mode0))
7901 op0 = copy_to_mode_reg (mode0, op0);
7902 if (! (*insn_data[icode].operand[1].predicate) (op1, mode1))
7903 op1 = copy_to_mode_reg (mode0, op1);
7904 if (! (*insn_data[icode].operand[1].predicate) (op2, mode1))
7905 op2 = copy_to_mode_reg (mode0, op2);
7906 if (! (*insn_data[icode].operand[1].predicate) (op3, mode1))
7907 op3 = copy_to_mode_reg (mode0, op3);
7908
7909 /* Generate the compare. */
7910 scratch = gen_reg_rtx (CCmode);
7911 pat = GEN_FCN (icode) (scratch, op0, op1);
7912 if (! pat)
7913 return const0_rtx;
7914 emit_insn (pat);
7915
7916 if (mode0 == V2SImode)
7917 emit_insn (gen_spe_evsel (target, op2, op3, scratch));
7918 else
7919 emit_insn (gen_spe_evsel_fs (target, op2, op3, scratch));
7920
7921 return target;
7922 }
7923
7924 /* Expand an expression EXP that calls a built-in function,
7925 with result going to TARGET if that's convenient
7926 (and in mode MODE if that's convenient).
7927 SUBTARGET may be used as the target for computing one of EXP's operands.
7928 IGNORE is nonzero if the value is to be ignored. */
7929
7930 static rtx
rs6000_expand_builtin(tree exp,rtx target,rtx subtarget ATTRIBUTE_UNUSED,enum machine_mode mode ATTRIBUTE_UNUSED,int ignore ATTRIBUTE_UNUSED)7931 rs6000_expand_builtin (tree exp, rtx target, rtx subtarget ATTRIBUTE_UNUSED,
7932 enum machine_mode mode ATTRIBUTE_UNUSED,
7933 int ignore ATTRIBUTE_UNUSED)
7934 {
7935 tree fndecl = TREE_OPERAND (TREE_OPERAND (exp, 0), 0);
7936 tree arglist = TREE_OPERAND (exp, 1);
7937 unsigned int fcode = DECL_FUNCTION_CODE (fndecl);
7938 struct builtin_description *d;
7939 size_t i;
7940 rtx ret;
7941 bool success;
7942
7943 if (fcode == ALTIVEC_BUILTIN_MASK_FOR_LOAD
7944 || fcode == ALTIVEC_BUILTIN_MASK_FOR_STORE)
7945 {
7946 int icode = (int) CODE_FOR_altivec_lvsr;
7947 enum machine_mode tmode = insn_data[icode].operand[0].mode;
7948 enum machine_mode mode = insn_data[icode].operand[1].mode;
7949 tree arg;
7950 rtx op, addr, pat;
7951
7952 gcc_assert (TARGET_ALTIVEC);
7953
7954 arg = TREE_VALUE (arglist);
7955 gcc_assert (TREE_CODE (TREE_TYPE (arg)) == POINTER_TYPE);
7956 op = expand_expr (arg, NULL_RTX, Pmode, EXPAND_NORMAL);
7957 addr = memory_address (mode, op);
7958 if (fcode == ALTIVEC_BUILTIN_MASK_FOR_STORE)
7959 op = addr;
7960 else
7961 {
7962 /* For the load case need to negate the address. */
7963 op = gen_reg_rtx (GET_MODE (addr));
7964 emit_insn (gen_rtx_SET (VOIDmode, op,
7965 gen_rtx_NEG (GET_MODE (addr), addr)));
7966 }
7967 op = gen_rtx_MEM (mode, op);
7968
7969 if (target == 0
7970 || GET_MODE (target) != tmode
7971 || ! (*insn_data[icode].operand[0].predicate) (target, tmode))
7972 target = gen_reg_rtx (tmode);
7973
7974 /*pat = gen_altivec_lvsr (target, op);*/
7975 pat = GEN_FCN (icode) (target, op);
7976 if (!pat)
7977 return 0;
7978 emit_insn (pat);
7979
7980 return target;
7981 }
7982
7983 if (TARGET_ALTIVEC)
7984 {
7985 ret = altivec_expand_builtin (exp, target, &success);
7986
7987 if (success)
7988 return ret;
7989 }
7990 if (TARGET_SPE)
7991 {
7992 ret = spe_expand_builtin (exp, target, &success);
7993
7994 if (success)
7995 return ret;
7996 }
7997
7998 gcc_assert (TARGET_ALTIVEC || TARGET_SPE);
7999
8000 /* Handle simple unary operations. */
8001 d = (struct builtin_description *) bdesc_1arg;
8002 for (i = 0; i < ARRAY_SIZE (bdesc_1arg); i++, d++)
8003 if (d->code == fcode)
8004 return rs6000_expand_unop_builtin (d->icode, arglist, target);
8005
8006 /* Handle simple binary operations. */
8007 d = (struct builtin_description *) bdesc_2arg;
8008 for (i = 0; i < ARRAY_SIZE (bdesc_2arg); i++, d++)
8009 if (d->code == fcode)
8010 return rs6000_expand_binop_builtin (d->icode, arglist, target);
8011
8012 /* Handle simple ternary operations. */
8013 d = (struct builtin_description *) bdesc_3arg;
8014 for (i = 0; i < ARRAY_SIZE (bdesc_3arg); i++, d++)
8015 if (d->code == fcode)
8016 return rs6000_expand_ternop_builtin (d->icode, arglist, target);
8017
8018 gcc_unreachable ();
8019 }
8020
8021 static tree
build_opaque_vector_type(tree node,int nunits)8022 build_opaque_vector_type (tree node, int nunits)
8023 {
8024 node = copy_node (node);
8025 TYPE_MAIN_VARIANT (node) = node;
8026 return build_vector_type (node, nunits);
8027 }
8028
8029 static void
rs6000_init_builtins(void)8030 rs6000_init_builtins (void)
8031 {
8032 V2SI_type_node = build_vector_type (intSI_type_node, 2);
8033 V2SF_type_node = build_vector_type (float_type_node, 2);
8034 V4HI_type_node = build_vector_type (intHI_type_node, 4);
8035 V4SI_type_node = build_vector_type (intSI_type_node, 4);
8036 V4SF_type_node = build_vector_type (float_type_node, 4);
8037 V8HI_type_node = build_vector_type (intHI_type_node, 8);
8038 V16QI_type_node = build_vector_type (intQI_type_node, 16);
8039
8040 unsigned_V16QI_type_node = build_vector_type (unsigned_intQI_type_node, 16);
8041 unsigned_V8HI_type_node = build_vector_type (unsigned_intHI_type_node, 8);
8042 unsigned_V4SI_type_node = build_vector_type (unsigned_intSI_type_node, 4);
8043
8044 opaque_V2SF_type_node = build_opaque_vector_type (float_type_node, 2);
8045 opaque_V2SI_type_node = build_opaque_vector_type (intSI_type_node, 2);
8046 opaque_p_V2SI_type_node = build_pointer_type (opaque_V2SI_type_node);
8047 opaque_V4SI_type_node = copy_node (V4SI_type_node);
8048
8049 /* The 'vector bool ...' types must be kept distinct from 'vector unsigned ...'
8050 types, especially in C++ land. Similarly, 'vector pixel' is distinct from
8051 'vector unsigned short'. */
8052
8053 bool_char_type_node = build_distinct_type_copy (unsigned_intQI_type_node);
8054 bool_short_type_node = build_distinct_type_copy (unsigned_intHI_type_node);
8055 bool_int_type_node = build_distinct_type_copy (unsigned_intSI_type_node);
8056 pixel_type_node = build_distinct_type_copy (unsigned_intHI_type_node);
8057
8058 long_integer_type_internal_node = long_integer_type_node;
8059 long_unsigned_type_internal_node = long_unsigned_type_node;
8060 intQI_type_internal_node = intQI_type_node;
8061 uintQI_type_internal_node = unsigned_intQI_type_node;
8062 intHI_type_internal_node = intHI_type_node;
8063 uintHI_type_internal_node = unsigned_intHI_type_node;
8064 intSI_type_internal_node = intSI_type_node;
8065 uintSI_type_internal_node = unsigned_intSI_type_node;
8066 float_type_internal_node = float_type_node;
8067 void_type_internal_node = void_type_node;
8068
8069 (*lang_hooks.decls.pushdecl) (build_decl (TYPE_DECL,
8070 get_identifier ("__bool char"),
8071 bool_char_type_node));
8072 (*lang_hooks.decls.pushdecl) (build_decl (TYPE_DECL,
8073 get_identifier ("__bool short"),
8074 bool_short_type_node));
8075 (*lang_hooks.decls.pushdecl) (build_decl (TYPE_DECL,
8076 get_identifier ("__bool int"),
8077 bool_int_type_node));
8078 (*lang_hooks.decls.pushdecl) (build_decl (TYPE_DECL,
8079 get_identifier ("__pixel"),
8080 pixel_type_node));
8081
8082 bool_V16QI_type_node = build_vector_type (bool_char_type_node, 16);
8083 bool_V8HI_type_node = build_vector_type (bool_short_type_node, 8);
8084 bool_V4SI_type_node = build_vector_type (bool_int_type_node, 4);
8085 pixel_V8HI_type_node = build_vector_type (pixel_type_node, 8);
8086
8087 (*lang_hooks.decls.pushdecl) (build_decl (TYPE_DECL,
8088 get_identifier ("__vector unsigned char"),
8089 unsigned_V16QI_type_node));
8090 (*lang_hooks.decls.pushdecl) (build_decl (TYPE_DECL,
8091 get_identifier ("__vector signed char"),
8092 V16QI_type_node));
8093 (*lang_hooks.decls.pushdecl) (build_decl (TYPE_DECL,
8094 get_identifier ("__vector __bool char"),
8095 bool_V16QI_type_node));
8096
8097 (*lang_hooks.decls.pushdecl) (build_decl (TYPE_DECL,
8098 get_identifier ("__vector unsigned short"),
8099 unsigned_V8HI_type_node));
8100 (*lang_hooks.decls.pushdecl) (build_decl (TYPE_DECL,
8101 get_identifier ("__vector signed short"),
8102 V8HI_type_node));
8103 (*lang_hooks.decls.pushdecl) (build_decl (TYPE_DECL,
8104 get_identifier ("__vector __bool short"),
8105 bool_V8HI_type_node));
8106
8107 (*lang_hooks.decls.pushdecl) (build_decl (TYPE_DECL,
8108 get_identifier ("__vector unsigned int"),
8109 unsigned_V4SI_type_node));
8110 (*lang_hooks.decls.pushdecl) (build_decl (TYPE_DECL,
8111 get_identifier ("__vector signed int"),
8112 V4SI_type_node));
8113 (*lang_hooks.decls.pushdecl) (build_decl (TYPE_DECL,
8114 get_identifier ("__vector __bool int"),
8115 bool_V4SI_type_node));
8116
8117 (*lang_hooks.decls.pushdecl) (build_decl (TYPE_DECL,
8118 get_identifier ("__vector float"),
8119 V4SF_type_node));
8120 (*lang_hooks.decls.pushdecl) (build_decl (TYPE_DECL,
8121 get_identifier ("__vector __pixel"),
8122 pixel_V8HI_type_node));
8123
8124 if (TARGET_SPE)
8125 spe_init_builtins ();
8126 if (TARGET_ALTIVEC)
8127 altivec_init_builtins ();
8128 if (TARGET_ALTIVEC || TARGET_SPE)
8129 rs6000_common_init_builtins ();
8130
8131 #if TARGET_XCOFF
8132 /* AIX libm provides clog as __clog. */
8133 if (built_in_decls [BUILT_IN_CLOG])
8134 set_user_assembler_name (built_in_decls [BUILT_IN_CLOG], "__clog");
8135 #endif
8136 }
8137
8138 /* Search through a set of builtins and enable the mask bits.
8139 DESC is an array of builtins.
8140 SIZE is the total number of builtins.
8141 START is the builtin enum at which to start.
8142 END is the builtin enum at which to end. */
8143 static void
enable_mask_for_builtins(struct builtin_description * desc,int size,enum rs6000_builtins start,enum rs6000_builtins end)8144 enable_mask_for_builtins (struct builtin_description *desc, int size,
8145 enum rs6000_builtins start,
8146 enum rs6000_builtins end)
8147 {
8148 int i;
8149
8150 for (i = 0; i < size; ++i)
8151 if (desc[i].code == start)
8152 break;
8153
8154 if (i == size)
8155 return;
8156
8157 for (; i < size; ++i)
8158 {
8159 /* Flip all the bits on. */
8160 desc[i].mask = target_flags;
8161 if (desc[i].code == end)
8162 break;
8163 }
8164 }
8165
8166 static void
spe_init_builtins(void)8167 spe_init_builtins (void)
8168 {
8169 tree endlink = void_list_node;
8170 tree puint_type_node = build_pointer_type (unsigned_type_node);
8171 tree pushort_type_node = build_pointer_type (short_unsigned_type_node);
8172 struct builtin_description *d;
8173 size_t i;
8174
8175 tree v2si_ftype_4_v2si
8176 = build_function_type
8177 (opaque_V2SI_type_node,
8178 tree_cons (NULL_TREE, opaque_V2SI_type_node,
8179 tree_cons (NULL_TREE, opaque_V2SI_type_node,
8180 tree_cons (NULL_TREE, opaque_V2SI_type_node,
8181 tree_cons (NULL_TREE, opaque_V2SI_type_node,
8182 endlink)))));
8183
8184 tree v2sf_ftype_4_v2sf
8185 = build_function_type
8186 (opaque_V2SF_type_node,
8187 tree_cons (NULL_TREE, opaque_V2SF_type_node,
8188 tree_cons (NULL_TREE, opaque_V2SF_type_node,
8189 tree_cons (NULL_TREE, opaque_V2SF_type_node,
8190 tree_cons (NULL_TREE, opaque_V2SF_type_node,
8191 endlink)))));
8192
8193 tree int_ftype_int_v2si_v2si
8194 = build_function_type
8195 (integer_type_node,
8196 tree_cons (NULL_TREE, integer_type_node,
8197 tree_cons (NULL_TREE, opaque_V2SI_type_node,
8198 tree_cons (NULL_TREE, opaque_V2SI_type_node,
8199 endlink))));
8200
8201 tree int_ftype_int_v2sf_v2sf
8202 = build_function_type
8203 (integer_type_node,
8204 tree_cons (NULL_TREE, integer_type_node,
8205 tree_cons (NULL_TREE, opaque_V2SF_type_node,
8206 tree_cons (NULL_TREE, opaque_V2SF_type_node,
8207 endlink))));
8208
8209 tree void_ftype_v2si_puint_int
8210 = build_function_type (void_type_node,
8211 tree_cons (NULL_TREE, opaque_V2SI_type_node,
8212 tree_cons (NULL_TREE, puint_type_node,
8213 tree_cons (NULL_TREE,
8214 integer_type_node,
8215 endlink))));
8216
8217 tree void_ftype_v2si_puint_char
8218 = build_function_type (void_type_node,
8219 tree_cons (NULL_TREE, opaque_V2SI_type_node,
8220 tree_cons (NULL_TREE, puint_type_node,
8221 tree_cons (NULL_TREE,
8222 char_type_node,
8223 endlink))));
8224
8225 tree void_ftype_v2si_pv2si_int
8226 = build_function_type (void_type_node,
8227 tree_cons (NULL_TREE, opaque_V2SI_type_node,
8228 tree_cons (NULL_TREE, opaque_p_V2SI_type_node,
8229 tree_cons (NULL_TREE,
8230 integer_type_node,
8231 endlink))));
8232
8233 tree void_ftype_v2si_pv2si_char
8234 = build_function_type (void_type_node,
8235 tree_cons (NULL_TREE, opaque_V2SI_type_node,
8236 tree_cons (NULL_TREE, opaque_p_V2SI_type_node,
8237 tree_cons (NULL_TREE,
8238 char_type_node,
8239 endlink))));
8240
8241 tree void_ftype_int
8242 = build_function_type (void_type_node,
8243 tree_cons (NULL_TREE, integer_type_node, endlink));
8244
8245 tree int_ftype_void
8246 = build_function_type (integer_type_node, endlink);
8247
8248 tree v2si_ftype_pv2si_int
8249 = build_function_type (opaque_V2SI_type_node,
8250 tree_cons (NULL_TREE, opaque_p_V2SI_type_node,
8251 tree_cons (NULL_TREE, integer_type_node,
8252 endlink)));
8253
8254 tree v2si_ftype_puint_int
8255 = build_function_type (opaque_V2SI_type_node,
8256 tree_cons (NULL_TREE, puint_type_node,
8257 tree_cons (NULL_TREE, integer_type_node,
8258 endlink)));
8259
8260 tree v2si_ftype_pushort_int
8261 = build_function_type (opaque_V2SI_type_node,
8262 tree_cons (NULL_TREE, pushort_type_node,
8263 tree_cons (NULL_TREE, integer_type_node,
8264 endlink)));
8265
8266 tree v2si_ftype_signed_char
8267 = build_function_type (opaque_V2SI_type_node,
8268 tree_cons (NULL_TREE, signed_char_type_node,
8269 endlink));
8270
8271 /* The initialization of the simple binary and unary builtins is
8272 done in rs6000_common_init_builtins, but we have to enable the
8273 mask bits here manually because we have run out of `target_flags'
8274 bits. We really need to redesign this mask business. */
8275
8276 enable_mask_for_builtins ((struct builtin_description *) bdesc_2arg,
8277 ARRAY_SIZE (bdesc_2arg),
8278 SPE_BUILTIN_EVADDW,
8279 SPE_BUILTIN_EVXOR);
8280 enable_mask_for_builtins ((struct builtin_description *) bdesc_1arg,
8281 ARRAY_SIZE (bdesc_1arg),
8282 SPE_BUILTIN_EVABS,
8283 SPE_BUILTIN_EVSUBFUSIAAW);
8284 enable_mask_for_builtins ((struct builtin_description *) bdesc_spe_predicates,
8285 ARRAY_SIZE (bdesc_spe_predicates),
8286 SPE_BUILTIN_EVCMPEQ,
8287 SPE_BUILTIN_EVFSTSTLT);
8288 enable_mask_for_builtins ((struct builtin_description *) bdesc_spe_evsel,
8289 ARRAY_SIZE (bdesc_spe_evsel),
8290 SPE_BUILTIN_EVSEL_CMPGTS,
8291 SPE_BUILTIN_EVSEL_FSTSTEQ);
8292
8293 (*lang_hooks.decls.pushdecl)
8294 (build_decl (TYPE_DECL, get_identifier ("__ev64_opaque__"),
8295 opaque_V2SI_type_node));
8296
8297 /* Initialize irregular SPE builtins. */
8298
8299 def_builtin (target_flags, "__builtin_spe_mtspefscr", void_ftype_int, SPE_BUILTIN_MTSPEFSCR);
8300 def_builtin (target_flags, "__builtin_spe_mfspefscr", int_ftype_void, SPE_BUILTIN_MFSPEFSCR);
8301 def_builtin (target_flags, "__builtin_spe_evstddx", void_ftype_v2si_pv2si_int, SPE_BUILTIN_EVSTDDX);
8302 def_builtin (target_flags, "__builtin_spe_evstdhx", void_ftype_v2si_pv2si_int, SPE_BUILTIN_EVSTDHX);
8303 def_builtin (target_flags, "__builtin_spe_evstdwx", void_ftype_v2si_pv2si_int, SPE_BUILTIN_EVSTDWX);
8304 def_builtin (target_flags, "__builtin_spe_evstwhex", void_ftype_v2si_puint_int, SPE_BUILTIN_EVSTWHEX);
8305 def_builtin (target_flags, "__builtin_spe_evstwhox", void_ftype_v2si_puint_int, SPE_BUILTIN_EVSTWHOX);
8306 def_builtin (target_flags, "__builtin_spe_evstwwex", void_ftype_v2si_puint_int, SPE_BUILTIN_EVSTWWEX);
8307 def_builtin (target_flags, "__builtin_spe_evstwwox", void_ftype_v2si_puint_int, SPE_BUILTIN_EVSTWWOX);
8308 def_builtin (target_flags, "__builtin_spe_evstdd", void_ftype_v2si_pv2si_char, SPE_BUILTIN_EVSTDD);
8309 def_builtin (target_flags, "__builtin_spe_evstdh", void_ftype_v2si_pv2si_char, SPE_BUILTIN_EVSTDH);
8310 def_builtin (target_flags, "__builtin_spe_evstdw", void_ftype_v2si_pv2si_char, SPE_BUILTIN_EVSTDW);
8311 def_builtin (target_flags, "__builtin_spe_evstwhe", void_ftype_v2si_puint_char, SPE_BUILTIN_EVSTWHE);
8312 def_builtin (target_flags, "__builtin_spe_evstwho", void_ftype_v2si_puint_char, SPE_BUILTIN_EVSTWHO);
8313 def_builtin (target_flags, "__builtin_spe_evstwwe", void_ftype_v2si_puint_char, SPE_BUILTIN_EVSTWWE);
8314 def_builtin (target_flags, "__builtin_spe_evstwwo", void_ftype_v2si_puint_char, SPE_BUILTIN_EVSTWWO);
8315 def_builtin (target_flags, "__builtin_spe_evsplatfi", v2si_ftype_signed_char, SPE_BUILTIN_EVSPLATFI);
8316 def_builtin (target_flags, "__builtin_spe_evsplati", v2si_ftype_signed_char, SPE_BUILTIN_EVSPLATI);
8317
8318 /* Loads. */
8319 def_builtin (target_flags, "__builtin_spe_evlddx", v2si_ftype_pv2si_int, SPE_BUILTIN_EVLDDX);
8320 def_builtin (target_flags, "__builtin_spe_evldwx", v2si_ftype_pv2si_int, SPE_BUILTIN_EVLDWX);
8321 def_builtin (target_flags, "__builtin_spe_evldhx", v2si_ftype_pv2si_int, SPE_BUILTIN_EVLDHX);
8322 def_builtin (target_flags, "__builtin_spe_evlwhex", v2si_ftype_puint_int, SPE_BUILTIN_EVLWHEX);
8323 def_builtin (target_flags, "__builtin_spe_evlwhoux", v2si_ftype_puint_int, SPE_BUILTIN_EVLWHOUX);
8324 def_builtin (target_flags, "__builtin_spe_evlwhosx", v2si_ftype_puint_int, SPE_BUILTIN_EVLWHOSX);
8325 def_builtin (target_flags, "__builtin_spe_evlwwsplatx", v2si_ftype_puint_int, SPE_BUILTIN_EVLWWSPLATX);
8326 def_builtin (target_flags, "__builtin_spe_evlwhsplatx", v2si_ftype_puint_int, SPE_BUILTIN_EVLWHSPLATX);
8327 def_builtin (target_flags, "__builtin_spe_evlhhesplatx", v2si_ftype_pushort_int, SPE_BUILTIN_EVLHHESPLATX);
8328 def_builtin (target_flags, "__builtin_spe_evlhhousplatx", v2si_ftype_pushort_int, SPE_BUILTIN_EVLHHOUSPLATX);
8329 def_builtin (target_flags, "__builtin_spe_evlhhossplatx", v2si_ftype_pushort_int, SPE_BUILTIN_EVLHHOSSPLATX);
8330 def_builtin (target_flags, "__builtin_spe_evldd", v2si_ftype_pv2si_int, SPE_BUILTIN_EVLDD);
8331 def_builtin (target_flags, "__builtin_spe_evldw", v2si_ftype_pv2si_int, SPE_BUILTIN_EVLDW);
8332 def_builtin (target_flags, "__builtin_spe_evldh", v2si_ftype_pv2si_int, SPE_BUILTIN_EVLDH);
8333 def_builtin (target_flags, "__builtin_spe_evlhhesplat", v2si_ftype_pushort_int, SPE_BUILTIN_EVLHHESPLAT);
8334 def_builtin (target_flags, "__builtin_spe_evlhhossplat", v2si_ftype_pushort_int, SPE_BUILTIN_EVLHHOSSPLAT);
8335 def_builtin (target_flags, "__builtin_spe_evlhhousplat", v2si_ftype_pushort_int, SPE_BUILTIN_EVLHHOUSPLAT);
8336 def_builtin (target_flags, "__builtin_spe_evlwhe", v2si_ftype_puint_int, SPE_BUILTIN_EVLWHE);
8337 def_builtin (target_flags, "__builtin_spe_evlwhos", v2si_ftype_puint_int, SPE_BUILTIN_EVLWHOS);
8338 def_builtin (target_flags, "__builtin_spe_evlwhou", v2si_ftype_puint_int, SPE_BUILTIN_EVLWHOU);
8339 def_builtin (target_flags, "__builtin_spe_evlwhsplat", v2si_ftype_puint_int, SPE_BUILTIN_EVLWHSPLAT);
8340 def_builtin (target_flags, "__builtin_spe_evlwwsplat", v2si_ftype_puint_int, SPE_BUILTIN_EVLWWSPLAT);
8341
8342 /* Predicates. */
8343 d = (struct builtin_description *) bdesc_spe_predicates;
8344 for (i = 0; i < ARRAY_SIZE (bdesc_spe_predicates); ++i, d++)
8345 {
8346 tree type;
8347
8348 switch (insn_data[d->icode].operand[1].mode)
8349 {
8350 case V2SImode:
8351 type = int_ftype_int_v2si_v2si;
8352 break;
8353 case V2SFmode:
8354 type = int_ftype_int_v2sf_v2sf;
8355 break;
8356 default:
8357 gcc_unreachable ();
8358 }
8359
8360 def_builtin (d->mask, d->name, type, d->code);
8361 }
8362
8363 /* Evsel predicates. */
8364 d = (struct builtin_description *) bdesc_spe_evsel;
8365 for (i = 0; i < ARRAY_SIZE (bdesc_spe_evsel); ++i, d++)
8366 {
8367 tree type;
8368
8369 switch (insn_data[d->icode].operand[1].mode)
8370 {
8371 case V2SImode:
8372 type = v2si_ftype_4_v2si;
8373 break;
8374 case V2SFmode:
8375 type = v2sf_ftype_4_v2sf;
8376 break;
8377 default:
8378 gcc_unreachable ();
8379 }
8380
8381 def_builtin (d->mask, d->name, type, d->code);
8382 }
8383 }
8384
8385 static void
altivec_init_builtins(void)8386 altivec_init_builtins (void)
8387 {
8388 struct builtin_description *d;
8389 struct builtin_description_predicates *dp;
8390 size_t i;
8391 tree ftype;
8392
8393 tree pfloat_type_node = build_pointer_type (float_type_node);
8394 tree pint_type_node = build_pointer_type (integer_type_node);
8395 tree pshort_type_node = build_pointer_type (short_integer_type_node);
8396 tree pchar_type_node = build_pointer_type (char_type_node);
8397
8398 tree pvoid_type_node = build_pointer_type (void_type_node);
8399
8400 tree pcfloat_type_node = build_pointer_type (build_qualified_type (float_type_node, TYPE_QUAL_CONST));
8401 tree pcint_type_node = build_pointer_type (build_qualified_type (integer_type_node, TYPE_QUAL_CONST));
8402 tree pcshort_type_node = build_pointer_type (build_qualified_type (short_integer_type_node, TYPE_QUAL_CONST));
8403 tree pcchar_type_node = build_pointer_type (build_qualified_type (char_type_node, TYPE_QUAL_CONST));
8404
8405 tree pcvoid_type_node = build_pointer_type (build_qualified_type (void_type_node, TYPE_QUAL_CONST));
8406
8407 tree int_ftype_opaque
8408 = build_function_type_list (integer_type_node,
8409 opaque_V4SI_type_node, NULL_TREE);
8410
8411 tree opaque_ftype_opaque_int
8412 = build_function_type_list (opaque_V4SI_type_node,
8413 opaque_V4SI_type_node, integer_type_node, NULL_TREE);
8414 tree opaque_ftype_opaque_opaque_int
8415 = build_function_type_list (opaque_V4SI_type_node,
8416 opaque_V4SI_type_node, opaque_V4SI_type_node,
8417 integer_type_node, NULL_TREE);
8418 tree int_ftype_int_opaque_opaque
8419 = build_function_type_list (integer_type_node,
8420 integer_type_node, opaque_V4SI_type_node,
8421 opaque_V4SI_type_node, NULL_TREE);
8422 tree int_ftype_int_v4si_v4si
8423 = build_function_type_list (integer_type_node,
8424 integer_type_node, V4SI_type_node,
8425 V4SI_type_node, NULL_TREE);
8426 tree v4sf_ftype_pcfloat
8427 = build_function_type_list (V4SF_type_node, pcfloat_type_node, NULL_TREE);
8428 tree void_ftype_pfloat_v4sf
8429 = build_function_type_list (void_type_node,
8430 pfloat_type_node, V4SF_type_node, NULL_TREE);
8431 tree v4si_ftype_pcint
8432 = build_function_type_list (V4SI_type_node, pcint_type_node, NULL_TREE);
8433 tree void_ftype_pint_v4si
8434 = build_function_type_list (void_type_node,
8435 pint_type_node, V4SI_type_node, NULL_TREE);
8436 tree v8hi_ftype_pcshort
8437 = build_function_type_list (V8HI_type_node, pcshort_type_node, NULL_TREE);
8438 tree void_ftype_pshort_v8hi
8439 = build_function_type_list (void_type_node,
8440 pshort_type_node, V8HI_type_node, NULL_TREE);
8441 tree v16qi_ftype_pcchar
8442 = build_function_type_list (V16QI_type_node, pcchar_type_node, NULL_TREE);
8443 tree void_ftype_pchar_v16qi
8444 = build_function_type_list (void_type_node,
8445 pchar_type_node, V16QI_type_node, NULL_TREE);
8446 tree void_ftype_v4si
8447 = build_function_type_list (void_type_node, V4SI_type_node, NULL_TREE);
8448 tree v8hi_ftype_void
8449 = build_function_type (V8HI_type_node, void_list_node);
8450 tree void_ftype_void
8451 = build_function_type (void_type_node, void_list_node);
8452 tree void_ftype_int
8453 = build_function_type_list (void_type_node, integer_type_node, NULL_TREE);
8454
8455 tree opaque_ftype_long_pcvoid
8456 = build_function_type_list (opaque_V4SI_type_node,
8457 long_integer_type_node, pcvoid_type_node, NULL_TREE);
8458 tree v16qi_ftype_long_pcvoid
8459 = build_function_type_list (V16QI_type_node,
8460 long_integer_type_node, pcvoid_type_node, NULL_TREE);
8461 tree v8hi_ftype_long_pcvoid
8462 = build_function_type_list (V8HI_type_node,
8463 long_integer_type_node, pcvoid_type_node, NULL_TREE);
8464 tree v4si_ftype_long_pcvoid
8465 = build_function_type_list (V4SI_type_node,
8466 long_integer_type_node, pcvoid_type_node, NULL_TREE);
8467
8468 tree void_ftype_opaque_long_pvoid
8469 = build_function_type_list (void_type_node,
8470 opaque_V4SI_type_node, long_integer_type_node,
8471 pvoid_type_node, NULL_TREE);
8472 tree void_ftype_v4si_long_pvoid
8473 = build_function_type_list (void_type_node,
8474 V4SI_type_node, long_integer_type_node,
8475 pvoid_type_node, NULL_TREE);
8476 tree void_ftype_v16qi_long_pvoid
8477 = build_function_type_list (void_type_node,
8478 V16QI_type_node, long_integer_type_node,
8479 pvoid_type_node, NULL_TREE);
8480 tree void_ftype_v8hi_long_pvoid
8481 = build_function_type_list (void_type_node,
8482 V8HI_type_node, long_integer_type_node,
8483 pvoid_type_node, NULL_TREE);
8484 tree int_ftype_int_v8hi_v8hi
8485 = build_function_type_list (integer_type_node,
8486 integer_type_node, V8HI_type_node,
8487 V8HI_type_node, NULL_TREE);
8488 tree int_ftype_int_v16qi_v16qi
8489 = build_function_type_list (integer_type_node,
8490 integer_type_node, V16QI_type_node,
8491 V16QI_type_node, NULL_TREE);
8492 tree int_ftype_int_v4sf_v4sf
8493 = build_function_type_list (integer_type_node,
8494 integer_type_node, V4SF_type_node,
8495 V4SF_type_node, NULL_TREE);
8496 tree v4si_ftype_v4si
8497 = build_function_type_list (V4SI_type_node, V4SI_type_node, NULL_TREE);
8498 tree v8hi_ftype_v8hi
8499 = build_function_type_list (V8HI_type_node, V8HI_type_node, NULL_TREE);
8500 tree v16qi_ftype_v16qi
8501 = build_function_type_list (V16QI_type_node, V16QI_type_node, NULL_TREE);
8502 tree v4sf_ftype_v4sf
8503 = build_function_type_list (V4SF_type_node, V4SF_type_node, NULL_TREE);
8504 tree void_ftype_pcvoid_int_int
8505 = build_function_type_list (void_type_node,
8506 pcvoid_type_node, integer_type_node,
8507 integer_type_node, NULL_TREE);
8508
8509 def_builtin (MASK_ALTIVEC, "__builtin_altivec_ld_internal_4sf", v4sf_ftype_pcfloat,
8510 ALTIVEC_BUILTIN_LD_INTERNAL_4sf);
8511 def_builtin (MASK_ALTIVEC, "__builtin_altivec_st_internal_4sf", void_ftype_pfloat_v4sf,
8512 ALTIVEC_BUILTIN_ST_INTERNAL_4sf);
8513 def_builtin (MASK_ALTIVEC, "__builtin_altivec_ld_internal_4si", v4si_ftype_pcint,
8514 ALTIVEC_BUILTIN_LD_INTERNAL_4si);
8515 def_builtin (MASK_ALTIVEC, "__builtin_altivec_st_internal_4si", void_ftype_pint_v4si,
8516 ALTIVEC_BUILTIN_ST_INTERNAL_4si);
8517 def_builtin (MASK_ALTIVEC, "__builtin_altivec_ld_internal_8hi", v8hi_ftype_pcshort,
8518 ALTIVEC_BUILTIN_LD_INTERNAL_8hi);
8519 def_builtin (MASK_ALTIVEC, "__builtin_altivec_st_internal_8hi", void_ftype_pshort_v8hi,
8520 ALTIVEC_BUILTIN_ST_INTERNAL_8hi);
8521 def_builtin (MASK_ALTIVEC, "__builtin_altivec_ld_internal_16qi", v16qi_ftype_pcchar,
8522 ALTIVEC_BUILTIN_LD_INTERNAL_16qi);
8523 def_builtin (MASK_ALTIVEC, "__builtin_altivec_st_internal_16qi", void_ftype_pchar_v16qi,
8524 ALTIVEC_BUILTIN_ST_INTERNAL_16qi);
8525 def_builtin (MASK_ALTIVEC, "__builtin_altivec_mtvscr", void_ftype_v4si, ALTIVEC_BUILTIN_MTVSCR);
8526 def_builtin (MASK_ALTIVEC, "__builtin_altivec_mfvscr", v8hi_ftype_void, ALTIVEC_BUILTIN_MFVSCR);
8527 def_builtin (MASK_ALTIVEC, "__builtin_altivec_dssall", void_ftype_void, ALTIVEC_BUILTIN_DSSALL);
8528 def_builtin (MASK_ALTIVEC, "__builtin_altivec_dss", void_ftype_int, ALTIVEC_BUILTIN_DSS);
8529 def_builtin (MASK_ALTIVEC, "__builtin_altivec_lvsl", v16qi_ftype_long_pcvoid, ALTIVEC_BUILTIN_LVSL);
8530 def_builtin (MASK_ALTIVEC, "__builtin_altivec_lvsr", v16qi_ftype_long_pcvoid, ALTIVEC_BUILTIN_LVSR);
8531 def_builtin (MASK_ALTIVEC, "__builtin_altivec_lvebx", v16qi_ftype_long_pcvoid, ALTIVEC_BUILTIN_LVEBX);
8532 def_builtin (MASK_ALTIVEC, "__builtin_altivec_lvehx", v8hi_ftype_long_pcvoid, ALTIVEC_BUILTIN_LVEHX);
8533 def_builtin (MASK_ALTIVEC, "__builtin_altivec_lvewx", v4si_ftype_long_pcvoid, ALTIVEC_BUILTIN_LVEWX);
8534 def_builtin (MASK_ALTIVEC, "__builtin_altivec_lvxl", v4si_ftype_long_pcvoid, ALTIVEC_BUILTIN_LVXL);
8535 def_builtin (MASK_ALTIVEC, "__builtin_altivec_lvx", v4si_ftype_long_pcvoid, ALTIVEC_BUILTIN_LVX);
8536 def_builtin (MASK_ALTIVEC, "__builtin_altivec_stvx", void_ftype_v4si_long_pvoid, ALTIVEC_BUILTIN_STVX);
8537 def_builtin (MASK_ALTIVEC, "__builtin_altivec_stvewx", void_ftype_v4si_long_pvoid, ALTIVEC_BUILTIN_STVEWX);
8538 def_builtin (MASK_ALTIVEC, "__builtin_altivec_stvxl", void_ftype_v4si_long_pvoid, ALTIVEC_BUILTIN_STVXL);
8539 def_builtin (MASK_ALTIVEC, "__builtin_altivec_stvebx", void_ftype_v16qi_long_pvoid, ALTIVEC_BUILTIN_STVEBX);
8540 def_builtin (MASK_ALTIVEC, "__builtin_altivec_stvehx", void_ftype_v8hi_long_pvoid, ALTIVEC_BUILTIN_STVEHX);
8541 def_builtin (MASK_ALTIVEC, "__builtin_vec_ld", opaque_ftype_long_pcvoid, ALTIVEC_BUILTIN_VEC_LD);
8542 def_builtin (MASK_ALTIVEC, "__builtin_vec_lde", opaque_ftype_long_pcvoid, ALTIVEC_BUILTIN_VEC_LDE);
8543 def_builtin (MASK_ALTIVEC, "__builtin_vec_ldl", opaque_ftype_long_pcvoid, ALTIVEC_BUILTIN_VEC_LDL);
8544 def_builtin (MASK_ALTIVEC, "__builtin_vec_lvsl", v16qi_ftype_long_pcvoid, ALTIVEC_BUILTIN_VEC_LVSL);
8545 def_builtin (MASK_ALTIVEC, "__builtin_vec_lvsr", v16qi_ftype_long_pcvoid, ALTIVEC_BUILTIN_VEC_LVSR);
8546 def_builtin (MASK_ALTIVEC, "__builtin_vec_lvebx", v16qi_ftype_long_pcvoid, ALTIVEC_BUILTIN_VEC_LVEBX);
8547 def_builtin (MASK_ALTIVEC, "__builtin_vec_lvehx", v8hi_ftype_long_pcvoid, ALTIVEC_BUILTIN_VEC_LVEHX);
8548 def_builtin (MASK_ALTIVEC, "__builtin_vec_lvewx", v4si_ftype_long_pcvoid, ALTIVEC_BUILTIN_VEC_LVEWX);
8549 def_builtin (MASK_ALTIVEC, "__builtin_vec_st", void_ftype_opaque_long_pvoid, ALTIVEC_BUILTIN_VEC_ST);
8550 def_builtin (MASK_ALTIVEC, "__builtin_vec_ste", void_ftype_opaque_long_pvoid, ALTIVEC_BUILTIN_VEC_STE);
8551 def_builtin (MASK_ALTIVEC, "__builtin_vec_stl", void_ftype_opaque_long_pvoid, ALTIVEC_BUILTIN_VEC_STL);
8552 def_builtin (MASK_ALTIVEC, "__builtin_vec_stvewx", void_ftype_opaque_long_pvoid, ALTIVEC_BUILTIN_VEC_STVEWX);
8553 def_builtin (MASK_ALTIVEC, "__builtin_vec_stvebx", void_ftype_opaque_long_pvoid, ALTIVEC_BUILTIN_VEC_STVEBX);
8554 def_builtin (MASK_ALTIVEC, "__builtin_vec_stvehx", void_ftype_opaque_long_pvoid, ALTIVEC_BUILTIN_VEC_STVEHX);
8555
8556 def_builtin (MASK_ALTIVEC, "__builtin_vec_step", int_ftype_opaque, ALTIVEC_BUILTIN_VEC_STEP);
8557
8558 def_builtin (MASK_ALTIVEC, "__builtin_vec_sld", opaque_ftype_opaque_opaque_int, ALTIVEC_BUILTIN_VEC_SLD);
8559 def_builtin (MASK_ALTIVEC, "__builtin_vec_splat", opaque_ftype_opaque_int, ALTIVEC_BUILTIN_VEC_SPLAT);
8560 def_builtin (MASK_ALTIVEC, "__builtin_vec_vspltw", opaque_ftype_opaque_int, ALTIVEC_BUILTIN_VEC_VSPLTW);
8561 def_builtin (MASK_ALTIVEC, "__builtin_vec_vsplth", opaque_ftype_opaque_int, ALTIVEC_BUILTIN_VEC_VSPLTH);
8562 def_builtin (MASK_ALTIVEC, "__builtin_vec_vspltb", opaque_ftype_opaque_int, ALTIVEC_BUILTIN_VEC_VSPLTB);
8563 def_builtin (MASK_ALTIVEC, "__builtin_vec_ctf", opaque_ftype_opaque_int, ALTIVEC_BUILTIN_VEC_CTF);
8564 def_builtin (MASK_ALTIVEC, "__builtin_vec_vcfsx", opaque_ftype_opaque_int, ALTIVEC_BUILTIN_VEC_VCFSX);
8565 def_builtin (MASK_ALTIVEC, "__builtin_vec_vcfux", opaque_ftype_opaque_int, ALTIVEC_BUILTIN_VEC_VCFUX);
8566 def_builtin (MASK_ALTIVEC, "__builtin_vec_cts", opaque_ftype_opaque_int, ALTIVEC_BUILTIN_VEC_CTS);
8567 def_builtin (MASK_ALTIVEC, "__builtin_vec_ctu", opaque_ftype_opaque_int, ALTIVEC_BUILTIN_VEC_CTU);
8568
8569 /* Add the DST variants. */
8570 d = (struct builtin_description *) bdesc_dst;
8571 for (i = 0; i < ARRAY_SIZE (bdesc_dst); i++, d++)
8572 def_builtin (d->mask, d->name, void_ftype_pcvoid_int_int, d->code);
8573
8574 /* Initialize the predicates. */
8575 dp = (struct builtin_description_predicates *) bdesc_altivec_preds;
8576 for (i = 0; i < ARRAY_SIZE (bdesc_altivec_preds); i++, dp++)
8577 {
8578 enum machine_mode mode1;
8579 tree type;
8580 bool is_overloaded = dp->code >= ALTIVEC_BUILTIN_OVERLOADED_FIRST
8581 && dp->code <= ALTIVEC_BUILTIN_OVERLOADED_LAST;
8582
8583 if (is_overloaded)
8584 mode1 = VOIDmode;
8585 else
8586 mode1 = insn_data[dp->icode].operand[1].mode;
8587
8588 switch (mode1)
8589 {
8590 case VOIDmode:
8591 type = int_ftype_int_opaque_opaque;
8592 break;
8593 case V4SImode:
8594 type = int_ftype_int_v4si_v4si;
8595 break;
8596 case V8HImode:
8597 type = int_ftype_int_v8hi_v8hi;
8598 break;
8599 case V16QImode:
8600 type = int_ftype_int_v16qi_v16qi;
8601 break;
8602 case V4SFmode:
8603 type = int_ftype_int_v4sf_v4sf;
8604 break;
8605 default:
8606 gcc_unreachable ();
8607 }
8608
8609 def_builtin (dp->mask, dp->name, type, dp->code);
8610 }
8611
8612 /* Initialize the abs* operators. */
8613 d = (struct builtin_description *) bdesc_abs;
8614 for (i = 0; i < ARRAY_SIZE (bdesc_abs); i++, d++)
8615 {
8616 enum machine_mode mode0;
8617 tree type;
8618
8619 mode0 = insn_data[d->icode].operand[0].mode;
8620
8621 switch (mode0)
8622 {
8623 case V4SImode:
8624 type = v4si_ftype_v4si;
8625 break;
8626 case V8HImode:
8627 type = v8hi_ftype_v8hi;
8628 break;
8629 case V16QImode:
8630 type = v16qi_ftype_v16qi;
8631 break;
8632 case V4SFmode:
8633 type = v4sf_ftype_v4sf;
8634 break;
8635 default:
8636 gcc_unreachable ();
8637 }
8638
8639 def_builtin (d->mask, d->name, type, d->code);
8640 }
8641
8642 if (TARGET_ALTIVEC)
8643 {
8644 tree decl;
8645
8646 /* Initialize target builtin that implements
8647 targetm.vectorize.builtin_mask_for_load. */
8648
8649 decl = lang_hooks.builtin_function ("__builtin_altivec_mask_for_load",
8650 v16qi_ftype_long_pcvoid,
8651 ALTIVEC_BUILTIN_MASK_FOR_LOAD,
8652 BUILT_IN_MD, NULL,
8653 tree_cons (get_identifier ("const"),
8654 NULL_TREE, NULL_TREE));
8655 /* Record the decl. Will be used by rs6000_builtin_mask_for_load. */
8656 altivec_builtin_mask_for_load = decl;
8657 }
8658
8659 /* Access to the vec_init patterns. */
8660 ftype = build_function_type_list (V4SI_type_node, integer_type_node,
8661 integer_type_node, integer_type_node,
8662 integer_type_node, NULL_TREE);
8663 def_builtin (MASK_ALTIVEC, "__builtin_vec_init_v4si", ftype,
8664 ALTIVEC_BUILTIN_VEC_INIT_V4SI);
8665
8666 ftype = build_function_type_list (V8HI_type_node, short_integer_type_node,
8667 short_integer_type_node,
8668 short_integer_type_node,
8669 short_integer_type_node,
8670 short_integer_type_node,
8671 short_integer_type_node,
8672 short_integer_type_node,
8673 short_integer_type_node, NULL_TREE);
8674 def_builtin (MASK_ALTIVEC, "__builtin_vec_init_v8hi", ftype,
8675 ALTIVEC_BUILTIN_VEC_INIT_V8HI);
8676
8677 ftype = build_function_type_list (V16QI_type_node, char_type_node,
8678 char_type_node, char_type_node,
8679 char_type_node, char_type_node,
8680 char_type_node, char_type_node,
8681 char_type_node, char_type_node,
8682 char_type_node, char_type_node,
8683 char_type_node, char_type_node,
8684 char_type_node, char_type_node,
8685 char_type_node, NULL_TREE);
8686 def_builtin (MASK_ALTIVEC, "__builtin_vec_init_v16qi", ftype,
8687 ALTIVEC_BUILTIN_VEC_INIT_V16QI);
8688
8689 ftype = build_function_type_list (V4SF_type_node, float_type_node,
8690 float_type_node, float_type_node,
8691 float_type_node, NULL_TREE);
8692 def_builtin (MASK_ALTIVEC, "__builtin_vec_init_v4sf", ftype,
8693 ALTIVEC_BUILTIN_VEC_INIT_V4SF);
8694
8695 /* Access to the vec_set patterns. */
8696 ftype = build_function_type_list (V4SI_type_node, V4SI_type_node,
8697 intSI_type_node,
8698 integer_type_node, NULL_TREE);
8699 def_builtin (MASK_ALTIVEC, "__builtin_vec_set_v4si", ftype,
8700 ALTIVEC_BUILTIN_VEC_SET_V4SI);
8701
8702 ftype = build_function_type_list (V8HI_type_node, V8HI_type_node,
8703 intHI_type_node,
8704 integer_type_node, NULL_TREE);
8705 def_builtin (MASK_ALTIVEC, "__builtin_vec_set_v8hi", ftype,
8706 ALTIVEC_BUILTIN_VEC_SET_V8HI);
8707
8708 ftype = build_function_type_list (V8HI_type_node, V16QI_type_node,
8709 intQI_type_node,
8710 integer_type_node, NULL_TREE);
8711 def_builtin (MASK_ALTIVEC, "__builtin_vec_set_v16qi", ftype,
8712 ALTIVEC_BUILTIN_VEC_SET_V16QI);
8713
8714 ftype = build_function_type_list (V4SF_type_node, V4SF_type_node,
8715 float_type_node,
8716 integer_type_node, NULL_TREE);
8717 def_builtin (MASK_ALTIVEC, "__builtin_vec_set_v4sf", ftype,
8718 ALTIVEC_BUILTIN_VEC_SET_V4SF);
8719
8720 /* Access to the vec_extract patterns. */
8721 ftype = build_function_type_list (intSI_type_node, V4SI_type_node,
8722 integer_type_node, NULL_TREE);
8723 def_builtin (MASK_ALTIVEC, "__builtin_vec_ext_v4si", ftype,
8724 ALTIVEC_BUILTIN_VEC_EXT_V4SI);
8725
8726 ftype = build_function_type_list (intHI_type_node, V8HI_type_node,
8727 integer_type_node, NULL_TREE);
8728 def_builtin (MASK_ALTIVEC, "__builtin_vec_ext_v8hi", ftype,
8729 ALTIVEC_BUILTIN_VEC_EXT_V8HI);
8730
8731 ftype = build_function_type_list (intQI_type_node, V16QI_type_node,
8732 integer_type_node, NULL_TREE);
8733 def_builtin (MASK_ALTIVEC, "__builtin_vec_ext_v16qi", ftype,
8734 ALTIVEC_BUILTIN_VEC_EXT_V16QI);
8735
8736 ftype = build_function_type_list (float_type_node, V4SF_type_node,
8737 integer_type_node, NULL_TREE);
8738 def_builtin (MASK_ALTIVEC, "__builtin_vec_ext_v4sf", ftype,
8739 ALTIVEC_BUILTIN_VEC_EXT_V4SF);
8740 }
8741
8742 static void
rs6000_common_init_builtins(void)8743 rs6000_common_init_builtins (void)
8744 {
8745 struct builtin_description *d;
8746 size_t i;
8747
8748 tree v4sf_ftype_v4sf_v4sf_v16qi
8749 = build_function_type_list (V4SF_type_node,
8750 V4SF_type_node, V4SF_type_node,
8751 V16QI_type_node, NULL_TREE);
8752 tree v4si_ftype_v4si_v4si_v16qi
8753 = build_function_type_list (V4SI_type_node,
8754 V4SI_type_node, V4SI_type_node,
8755 V16QI_type_node, NULL_TREE);
8756 tree v8hi_ftype_v8hi_v8hi_v16qi
8757 = build_function_type_list (V8HI_type_node,
8758 V8HI_type_node, V8HI_type_node,
8759 V16QI_type_node, NULL_TREE);
8760 tree v16qi_ftype_v16qi_v16qi_v16qi
8761 = build_function_type_list (V16QI_type_node,
8762 V16QI_type_node, V16QI_type_node,
8763 V16QI_type_node, NULL_TREE);
8764 tree v4si_ftype_int
8765 = build_function_type_list (V4SI_type_node, integer_type_node, NULL_TREE);
8766 tree v8hi_ftype_int
8767 = build_function_type_list (V8HI_type_node, integer_type_node, NULL_TREE);
8768 tree v16qi_ftype_int
8769 = build_function_type_list (V16QI_type_node, integer_type_node, NULL_TREE);
8770 tree v8hi_ftype_v16qi
8771 = build_function_type_list (V8HI_type_node, V16QI_type_node, NULL_TREE);
8772 tree v4sf_ftype_v4sf
8773 = build_function_type_list (V4SF_type_node, V4SF_type_node, NULL_TREE);
8774
8775 tree v2si_ftype_v2si_v2si
8776 = build_function_type_list (opaque_V2SI_type_node,
8777 opaque_V2SI_type_node,
8778 opaque_V2SI_type_node, NULL_TREE);
8779
8780 tree v2sf_ftype_v2sf_v2sf
8781 = build_function_type_list (opaque_V2SF_type_node,
8782 opaque_V2SF_type_node,
8783 opaque_V2SF_type_node, NULL_TREE);
8784
8785 tree v2si_ftype_int_int
8786 = build_function_type_list (opaque_V2SI_type_node,
8787 integer_type_node, integer_type_node,
8788 NULL_TREE);
8789
8790 tree opaque_ftype_opaque
8791 = build_function_type_list (opaque_V4SI_type_node,
8792 opaque_V4SI_type_node, NULL_TREE);
8793
8794 tree v2si_ftype_v2si
8795 = build_function_type_list (opaque_V2SI_type_node,
8796 opaque_V2SI_type_node, NULL_TREE);
8797
8798 tree v2sf_ftype_v2sf
8799 = build_function_type_list (opaque_V2SF_type_node,
8800 opaque_V2SF_type_node, NULL_TREE);
8801
8802 tree v2sf_ftype_v2si
8803 = build_function_type_list (opaque_V2SF_type_node,
8804 opaque_V2SI_type_node, NULL_TREE);
8805
8806 tree v2si_ftype_v2sf
8807 = build_function_type_list (opaque_V2SI_type_node,
8808 opaque_V2SF_type_node, NULL_TREE);
8809
8810 tree v2si_ftype_v2si_char
8811 = build_function_type_list (opaque_V2SI_type_node,
8812 opaque_V2SI_type_node,
8813 char_type_node, NULL_TREE);
8814
8815 tree v2si_ftype_int_char
8816 = build_function_type_list (opaque_V2SI_type_node,
8817 integer_type_node, char_type_node, NULL_TREE);
8818
8819 tree v2si_ftype_char
8820 = build_function_type_list (opaque_V2SI_type_node,
8821 char_type_node, NULL_TREE);
8822
8823 tree int_ftype_int_int
8824 = build_function_type_list (integer_type_node,
8825 integer_type_node, integer_type_node,
8826 NULL_TREE);
8827
8828 tree opaque_ftype_opaque_opaque
8829 = build_function_type_list (opaque_V4SI_type_node,
8830 opaque_V4SI_type_node, opaque_V4SI_type_node, NULL_TREE);
8831 tree v4si_ftype_v4si_v4si
8832 = build_function_type_list (V4SI_type_node,
8833 V4SI_type_node, V4SI_type_node, NULL_TREE);
8834 tree v4sf_ftype_v4si_int
8835 = build_function_type_list (V4SF_type_node,
8836 V4SI_type_node, integer_type_node, NULL_TREE);
8837 tree v4si_ftype_v4sf_int
8838 = build_function_type_list (V4SI_type_node,
8839 V4SF_type_node, integer_type_node, NULL_TREE);
8840 tree v4si_ftype_v4si_int
8841 = build_function_type_list (V4SI_type_node,
8842 V4SI_type_node, integer_type_node, NULL_TREE);
8843 tree v8hi_ftype_v8hi_int
8844 = build_function_type_list (V8HI_type_node,
8845 V8HI_type_node, integer_type_node, NULL_TREE);
8846 tree v16qi_ftype_v16qi_int
8847 = build_function_type_list (V16QI_type_node,
8848 V16QI_type_node, integer_type_node, NULL_TREE);
8849 tree v16qi_ftype_v16qi_v16qi_int
8850 = build_function_type_list (V16QI_type_node,
8851 V16QI_type_node, V16QI_type_node,
8852 integer_type_node, NULL_TREE);
8853 tree v8hi_ftype_v8hi_v8hi_int
8854 = build_function_type_list (V8HI_type_node,
8855 V8HI_type_node, V8HI_type_node,
8856 integer_type_node, NULL_TREE);
8857 tree v4si_ftype_v4si_v4si_int
8858 = build_function_type_list (V4SI_type_node,
8859 V4SI_type_node, V4SI_type_node,
8860 integer_type_node, NULL_TREE);
8861 tree v4sf_ftype_v4sf_v4sf_int
8862 = build_function_type_list (V4SF_type_node,
8863 V4SF_type_node, V4SF_type_node,
8864 integer_type_node, NULL_TREE);
8865 tree v4sf_ftype_v4sf_v4sf
8866 = build_function_type_list (V4SF_type_node,
8867 V4SF_type_node, V4SF_type_node, NULL_TREE);
8868 tree opaque_ftype_opaque_opaque_opaque
8869 = build_function_type_list (opaque_V4SI_type_node,
8870 opaque_V4SI_type_node, opaque_V4SI_type_node,
8871 opaque_V4SI_type_node, NULL_TREE);
8872 tree v4sf_ftype_v4sf_v4sf_v4si
8873 = build_function_type_list (V4SF_type_node,
8874 V4SF_type_node, V4SF_type_node,
8875 V4SI_type_node, NULL_TREE);
8876 tree v4sf_ftype_v4sf_v4sf_v4sf
8877 = build_function_type_list (V4SF_type_node,
8878 V4SF_type_node, V4SF_type_node,
8879 V4SF_type_node, NULL_TREE);
8880 tree v4si_ftype_v4si_v4si_v4si
8881 = build_function_type_list (V4SI_type_node,
8882 V4SI_type_node, V4SI_type_node,
8883 V4SI_type_node, NULL_TREE);
8884 tree v8hi_ftype_v8hi_v8hi
8885 = build_function_type_list (V8HI_type_node,
8886 V8HI_type_node, V8HI_type_node, NULL_TREE);
8887 tree v8hi_ftype_v8hi_v8hi_v8hi
8888 = build_function_type_list (V8HI_type_node,
8889 V8HI_type_node, V8HI_type_node,
8890 V8HI_type_node, NULL_TREE);
8891 tree v4si_ftype_v8hi_v8hi_v4si
8892 = build_function_type_list (V4SI_type_node,
8893 V8HI_type_node, V8HI_type_node,
8894 V4SI_type_node, NULL_TREE);
8895 tree v4si_ftype_v16qi_v16qi_v4si
8896 = build_function_type_list (V4SI_type_node,
8897 V16QI_type_node, V16QI_type_node,
8898 V4SI_type_node, NULL_TREE);
8899 tree v16qi_ftype_v16qi_v16qi
8900 = build_function_type_list (V16QI_type_node,
8901 V16QI_type_node, V16QI_type_node, NULL_TREE);
8902 tree v4si_ftype_v4sf_v4sf
8903 = build_function_type_list (V4SI_type_node,
8904 V4SF_type_node, V4SF_type_node, NULL_TREE);
8905 tree v8hi_ftype_v16qi_v16qi
8906 = build_function_type_list (V8HI_type_node,
8907 V16QI_type_node, V16QI_type_node, NULL_TREE);
8908 tree v4si_ftype_v8hi_v8hi
8909 = build_function_type_list (V4SI_type_node,
8910 V8HI_type_node, V8HI_type_node, NULL_TREE);
8911 tree v8hi_ftype_v4si_v4si
8912 = build_function_type_list (V8HI_type_node,
8913 V4SI_type_node, V4SI_type_node, NULL_TREE);
8914 tree v16qi_ftype_v8hi_v8hi
8915 = build_function_type_list (V16QI_type_node,
8916 V8HI_type_node, V8HI_type_node, NULL_TREE);
8917 tree v4si_ftype_v16qi_v4si
8918 = build_function_type_list (V4SI_type_node,
8919 V16QI_type_node, V4SI_type_node, NULL_TREE);
8920 tree v4si_ftype_v16qi_v16qi
8921 = build_function_type_list (V4SI_type_node,
8922 V16QI_type_node, V16QI_type_node, NULL_TREE);
8923 tree v4si_ftype_v8hi_v4si
8924 = build_function_type_list (V4SI_type_node,
8925 V8HI_type_node, V4SI_type_node, NULL_TREE);
8926 tree v4si_ftype_v8hi
8927 = build_function_type_list (V4SI_type_node, V8HI_type_node, NULL_TREE);
8928 tree int_ftype_v4si_v4si
8929 = build_function_type_list (integer_type_node,
8930 V4SI_type_node, V4SI_type_node, NULL_TREE);
8931 tree int_ftype_v4sf_v4sf
8932 = build_function_type_list (integer_type_node,
8933 V4SF_type_node, V4SF_type_node, NULL_TREE);
8934 tree int_ftype_v16qi_v16qi
8935 = build_function_type_list (integer_type_node,
8936 V16QI_type_node, V16QI_type_node, NULL_TREE);
8937 tree int_ftype_v8hi_v8hi
8938 = build_function_type_list (integer_type_node,
8939 V8HI_type_node, V8HI_type_node, NULL_TREE);
8940
8941 /* Add the simple ternary operators. */
8942 d = (struct builtin_description *) bdesc_3arg;
8943 for (i = 0; i < ARRAY_SIZE (bdesc_3arg); i++, d++)
8944 {
8945 enum machine_mode mode0, mode1, mode2, mode3;
8946 tree type;
8947 bool is_overloaded = d->code >= ALTIVEC_BUILTIN_OVERLOADED_FIRST
8948 && d->code <= ALTIVEC_BUILTIN_OVERLOADED_LAST;
8949
8950 if (is_overloaded)
8951 {
8952 mode0 = VOIDmode;
8953 mode1 = VOIDmode;
8954 mode2 = VOIDmode;
8955 mode3 = VOIDmode;
8956 }
8957 else
8958 {
8959 if (d->name == 0 || d->icode == CODE_FOR_nothing)
8960 continue;
8961
8962 mode0 = insn_data[d->icode].operand[0].mode;
8963 mode1 = insn_data[d->icode].operand[1].mode;
8964 mode2 = insn_data[d->icode].operand[2].mode;
8965 mode3 = insn_data[d->icode].operand[3].mode;
8966 }
8967
8968 /* When all four are of the same mode. */
8969 if (mode0 == mode1 && mode1 == mode2 && mode2 == mode3)
8970 {
8971 switch (mode0)
8972 {
8973 case VOIDmode:
8974 type = opaque_ftype_opaque_opaque_opaque;
8975 break;
8976 case V4SImode:
8977 type = v4si_ftype_v4si_v4si_v4si;
8978 break;
8979 case V4SFmode:
8980 type = v4sf_ftype_v4sf_v4sf_v4sf;
8981 break;
8982 case V8HImode:
8983 type = v8hi_ftype_v8hi_v8hi_v8hi;
8984 break;
8985 case V16QImode:
8986 type = v16qi_ftype_v16qi_v16qi_v16qi;
8987 break;
8988 default:
8989 gcc_unreachable ();
8990 }
8991 }
8992 else if (mode0 == mode1 && mode1 == mode2 && mode3 == V16QImode)
8993 {
8994 switch (mode0)
8995 {
8996 case V4SImode:
8997 type = v4si_ftype_v4si_v4si_v16qi;
8998 break;
8999 case V4SFmode:
9000 type = v4sf_ftype_v4sf_v4sf_v16qi;
9001 break;
9002 case V8HImode:
9003 type = v8hi_ftype_v8hi_v8hi_v16qi;
9004 break;
9005 case V16QImode:
9006 type = v16qi_ftype_v16qi_v16qi_v16qi;
9007 break;
9008 default:
9009 gcc_unreachable ();
9010 }
9011 }
9012 else if (mode0 == V4SImode && mode1 == V16QImode && mode2 == V16QImode
9013 && mode3 == V4SImode)
9014 type = v4si_ftype_v16qi_v16qi_v4si;
9015 else if (mode0 == V4SImode && mode1 == V8HImode && mode2 == V8HImode
9016 && mode3 == V4SImode)
9017 type = v4si_ftype_v8hi_v8hi_v4si;
9018 else if (mode0 == V4SFmode && mode1 == V4SFmode && mode2 == V4SFmode
9019 && mode3 == V4SImode)
9020 type = v4sf_ftype_v4sf_v4sf_v4si;
9021
9022 /* vchar, vchar, vchar, 4 bit literal. */
9023 else if (mode0 == V16QImode && mode1 == mode0 && mode2 == mode0
9024 && mode3 == QImode)
9025 type = v16qi_ftype_v16qi_v16qi_int;
9026
9027 /* vshort, vshort, vshort, 4 bit literal. */
9028 else if (mode0 == V8HImode && mode1 == mode0 && mode2 == mode0
9029 && mode3 == QImode)
9030 type = v8hi_ftype_v8hi_v8hi_int;
9031
9032 /* vint, vint, vint, 4 bit literal. */
9033 else if (mode0 == V4SImode && mode1 == mode0 && mode2 == mode0
9034 && mode3 == QImode)
9035 type = v4si_ftype_v4si_v4si_int;
9036
9037 /* vfloat, vfloat, vfloat, 4 bit literal. */
9038 else if (mode0 == V4SFmode && mode1 == mode0 && mode2 == mode0
9039 && mode3 == QImode)
9040 type = v4sf_ftype_v4sf_v4sf_int;
9041
9042 else
9043 gcc_unreachable ();
9044
9045 def_builtin (d->mask, d->name, type, d->code);
9046 }
9047
9048 /* Add the simple binary operators. */
9049 d = (struct builtin_description *) bdesc_2arg;
9050 for (i = 0; i < ARRAY_SIZE (bdesc_2arg); i++, d++)
9051 {
9052 enum machine_mode mode0, mode1, mode2;
9053 tree type;
9054 bool is_overloaded = d->code >= ALTIVEC_BUILTIN_OVERLOADED_FIRST
9055 && d->code <= ALTIVEC_BUILTIN_OVERLOADED_LAST;
9056
9057 if (is_overloaded)
9058 {
9059 mode0 = VOIDmode;
9060 mode1 = VOIDmode;
9061 mode2 = VOIDmode;
9062 }
9063 else
9064 {
9065 if (d->name == 0 || d->icode == CODE_FOR_nothing)
9066 continue;
9067
9068 mode0 = insn_data[d->icode].operand[0].mode;
9069 mode1 = insn_data[d->icode].operand[1].mode;
9070 mode2 = insn_data[d->icode].operand[2].mode;
9071 }
9072
9073 /* When all three operands are of the same mode. */
9074 if (mode0 == mode1 && mode1 == mode2)
9075 {
9076 switch (mode0)
9077 {
9078 case VOIDmode:
9079 type = opaque_ftype_opaque_opaque;
9080 break;
9081 case V4SFmode:
9082 type = v4sf_ftype_v4sf_v4sf;
9083 break;
9084 case V4SImode:
9085 type = v4si_ftype_v4si_v4si;
9086 break;
9087 case V16QImode:
9088 type = v16qi_ftype_v16qi_v16qi;
9089 break;
9090 case V8HImode:
9091 type = v8hi_ftype_v8hi_v8hi;
9092 break;
9093 case V2SImode:
9094 type = v2si_ftype_v2si_v2si;
9095 break;
9096 case V2SFmode:
9097 type = v2sf_ftype_v2sf_v2sf;
9098 break;
9099 case SImode:
9100 type = int_ftype_int_int;
9101 break;
9102 default:
9103 gcc_unreachable ();
9104 }
9105 }
9106
9107 /* A few other combos we really don't want to do manually. */
9108
9109 /* vint, vfloat, vfloat. */
9110 else if (mode0 == V4SImode && mode1 == V4SFmode && mode2 == V4SFmode)
9111 type = v4si_ftype_v4sf_v4sf;
9112
9113 /* vshort, vchar, vchar. */
9114 else if (mode0 == V8HImode && mode1 == V16QImode && mode2 == V16QImode)
9115 type = v8hi_ftype_v16qi_v16qi;
9116
9117 /* vint, vshort, vshort. */
9118 else if (mode0 == V4SImode && mode1 == V8HImode && mode2 == V8HImode)
9119 type = v4si_ftype_v8hi_v8hi;
9120
9121 /* vshort, vint, vint. */
9122 else if (mode0 == V8HImode && mode1 == V4SImode && mode2 == V4SImode)
9123 type = v8hi_ftype_v4si_v4si;
9124
9125 /* vchar, vshort, vshort. */
9126 else if (mode0 == V16QImode && mode1 == V8HImode && mode2 == V8HImode)
9127 type = v16qi_ftype_v8hi_v8hi;
9128
9129 /* vint, vchar, vint. */
9130 else if (mode0 == V4SImode && mode1 == V16QImode && mode2 == V4SImode)
9131 type = v4si_ftype_v16qi_v4si;
9132
9133 /* vint, vchar, vchar. */
9134 else if (mode0 == V4SImode && mode1 == V16QImode && mode2 == V16QImode)
9135 type = v4si_ftype_v16qi_v16qi;
9136
9137 /* vint, vshort, vint. */
9138 else if (mode0 == V4SImode && mode1 == V8HImode && mode2 == V4SImode)
9139 type = v4si_ftype_v8hi_v4si;
9140
9141 /* vint, vint, 5 bit literal. */
9142 else if (mode0 == V4SImode && mode1 == V4SImode && mode2 == QImode)
9143 type = v4si_ftype_v4si_int;
9144
9145 /* vshort, vshort, 5 bit literal. */
9146 else if (mode0 == V8HImode && mode1 == V8HImode && mode2 == QImode)
9147 type = v8hi_ftype_v8hi_int;
9148
9149 /* vchar, vchar, 5 bit literal. */
9150 else if (mode0 == V16QImode && mode1 == V16QImode && mode2 == QImode)
9151 type = v16qi_ftype_v16qi_int;
9152
9153 /* vfloat, vint, 5 bit literal. */
9154 else if (mode0 == V4SFmode && mode1 == V4SImode && mode2 == QImode)
9155 type = v4sf_ftype_v4si_int;
9156
9157 /* vint, vfloat, 5 bit literal. */
9158 else if (mode0 == V4SImode && mode1 == V4SFmode && mode2 == QImode)
9159 type = v4si_ftype_v4sf_int;
9160
9161 else if (mode0 == V2SImode && mode1 == SImode && mode2 == SImode)
9162 type = v2si_ftype_int_int;
9163
9164 else if (mode0 == V2SImode && mode1 == V2SImode && mode2 == QImode)
9165 type = v2si_ftype_v2si_char;
9166
9167 else if (mode0 == V2SImode && mode1 == SImode && mode2 == QImode)
9168 type = v2si_ftype_int_char;
9169
9170 else
9171 {
9172 /* int, x, x. */
9173 gcc_assert (mode0 == SImode);
9174 switch (mode1)
9175 {
9176 case V4SImode:
9177 type = int_ftype_v4si_v4si;
9178 break;
9179 case V4SFmode:
9180 type = int_ftype_v4sf_v4sf;
9181 break;
9182 case V16QImode:
9183 type = int_ftype_v16qi_v16qi;
9184 break;
9185 case V8HImode:
9186 type = int_ftype_v8hi_v8hi;
9187 break;
9188 default:
9189 gcc_unreachable ();
9190 }
9191 }
9192
9193 def_builtin (d->mask, d->name, type, d->code);
9194 }
9195
9196 /* Add the simple unary operators. */
9197 d = (struct builtin_description *) bdesc_1arg;
9198 for (i = 0; i < ARRAY_SIZE (bdesc_1arg); i++, d++)
9199 {
9200 enum machine_mode mode0, mode1;
9201 tree type;
9202 bool is_overloaded = d->code >= ALTIVEC_BUILTIN_OVERLOADED_FIRST
9203 && d->code <= ALTIVEC_BUILTIN_OVERLOADED_LAST;
9204
9205 if (is_overloaded)
9206 {
9207 mode0 = VOIDmode;
9208 mode1 = VOIDmode;
9209 }
9210 else
9211 {
9212 if (d->name == 0 || d->icode == CODE_FOR_nothing)
9213 continue;
9214
9215 mode0 = insn_data[d->icode].operand[0].mode;
9216 mode1 = insn_data[d->icode].operand[1].mode;
9217 }
9218
9219 if (mode0 == V4SImode && mode1 == QImode)
9220 type = v4si_ftype_int;
9221 else if (mode0 == V8HImode && mode1 == QImode)
9222 type = v8hi_ftype_int;
9223 else if (mode0 == V16QImode && mode1 == QImode)
9224 type = v16qi_ftype_int;
9225 else if (mode0 == VOIDmode && mode1 == VOIDmode)
9226 type = opaque_ftype_opaque;
9227 else if (mode0 == V4SFmode && mode1 == V4SFmode)
9228 type = v4sf_ftype_v4sf;
9229 else if (mode0 == V8HImode && mode1 == V16QImode)
9230 type = v8hi_ftype_v16qi;
9231 else if (mode0 == V4SImode && mode1 == V8HImode)
9232 type = v4si_ftype_v8hi;
9233 else if (mode0 == V2SImode && mode1 == V2SImode)
9234 type = v2si_ftype_v2si;
9235 else if (mode0 == V2SFmode && mode1 == V2SFmode)
9236 type = v2sf_ftype_v2sf;
9237 else if (mode0 == V2SFmode && mode1 == V2SImode)
9238 type = v2sf_ftype_v2si;
9239 else if (mode0 == V2SImode && mode1 == V2SFmode)
9240 type = v2si_ftype_v2sf;
9241 else if (mode0 == V2SImode && mode1 == QImode)
9242 type = v2si_ftype_char;
9243 else
9244 gcc_unreachable ();
9245
9246 def_builtin (d->mask, d->name, type, d->code);
9247 }
9248 }
9249
9250 static void
rs6000_init_libfuncs(void)9251 rs6000_init_libfuncs (void)
9252 {
9253 if (DEFAULT_ABI != ABI_V4 && TARGET_XCOFF
9254 && !TARGET_POWER2 && !TARGET_POWERPC)
9255 {
9256 /* AIX library routines for float->int conversion. */
9257 set_conv_libfunc (sfix_optab, SImode, DFmode, "__itrunc");
9258 set_conv_libfunc (ufix_optab, SImode, DFmode, "__uitrunc");
9259 set_conv_libfunc (sfix_optab, SImode, TFmode, "_qitrunc");
9260 set_conv_libfunc (ufix_optab, SImode, TFmode, "_quitrunc");
9261 }
9262
9263 if (!TARGET_IEEEQUAD)
9264 /* AIX/Darwin/64-bit Linux quad floating point routines. */
9265 if (!TARGET_XL_COMPAT)
9266 {
9267 set_optab_libfunc (add_optab, TFmode, "__gcc_qadd");
9268 set_optab_libfunc (sub_optab, TFmode, "__gcc_qsub");
9269 set_optab_libfunc (smul_optab, TFmode, "__gcc_qmul");
9270 set_optab_libfunc (sdiv_optab, TFmode, "__gcc_qdiv");
9271
9272 if (TARGET_SOFT_FLOAT)
9273 {
9274 set_optab_libfunc (neg_optab, TFmode, "__gcc_qneg");
9275 set_optab_libfunc (eq_optab, TFmode, "__gcc_qeq");
9276 set_optab_libfunc (ne_optab, TFmode, "__gcc_qne");
9277 set_optab_libfunc (gt_optab, TFmode, "__gcc_qgt");
9278 set_optab_libfunc (ge_optab, TFmode, "__gcc_qge");
9279 set_optab_libfunc (lt_optab, TFmode, "__gcc_qlt");
9280 set_optab_libfunc (le_optab, TFmode, "__gcc_qle");
9281 set_optab_libfunc (unord_optab, TFmode, "__gcc_qunord");
9282
9283 set_conv_libfunc (sext_optab, TFmode, SFmode, "__gcc_stoq");
9284 set_conv_libfunc (sext_optab, TFmode, DFmode, "__gcc_dtoq");
9285 set_conv_libfunc (trunc_optab, SFmode, TFmode, "__gcc_qtos");
9286 set_conv_libfunc (trunc_optab, DFmode, TFmode, "__gcc_qtod");
9287 set_conv_libfunc (sfix_optab, SImode, TFmode, "__gcc_qtoi");
9288 set_conv_libfunc (ufix_optab, SImode, TFmode, "__gcc_qtou");
9289 set_conv_libfunc (sfloat_optab, TFmode, SImode, "__gcc_itoq");
9290 set_conv_libfunc (ufloat_optab, TFmode, SImode, "__gcc_utoq");
9291 }
9292 }
9293 else
9294 {
9295 set_optab_libfunc (add_optab, TFmode, "_xlqadd");
9296 set_optab_libfunc (sub_optab, TFmode, "_xlqsub");
9297 set_optab_libfunc (smul_optab, TFmode, "_xlqmul");
9298 set_optab_libfunc (sdiv_optab, TFmode, "_xlqdiv");
9299 }
9300 else
9301 {
9302 /* 32-bit SVR4 quad floating point routines. */
9303
9304 set_optab_libfunc (add_optab, TFmode, "_q_add");
9305 set_optab_libfunc (sub_optab, TFmode, "_q_sub");
9306 set_optab_libfunc (neg_optab, TFmode, "_q_neg");
9307 set_optab_libfunc (smul_optab, TFmode, "_q_mul");
9308 set_optab_libfunc (sdiv_optab, TFmode, "_q_div");
9309 if (TARGET_PPC_GPOPT || TARGET_POWER2)
9310 set_optab_libfunc (sqrt_optab, TFmode, "_q_sqrt");
9311
9312 set_optab_libfunc (eq_optab, TFmode, "_q_feq");
9313 set_optab_libfunc (ne_optab, TFmode, "_q_fne");
9314 set_optab_libfunc (gt_optab, TFmode, "_q_fgt");
9315 set_optab_libfunc (ge_optab, TFmode, "_q_fge");
9316 set_optab_libfunc (lt_optab, TFmode, "_q_flt");
9317 set_optab_libfunc (le_optab, TFmode, "_q_fle");
9318
9319 set_conv_libfunc (sext_optab, TFmode, SFmode, "_q_stoq");
9320 set_conv_libfunc (sext_optab, TFmode, DFmode, "_q_dtoq");
9321 set_conv_libfunc (trunc_optab, SFmode, TFmode, "_q_qtos");
9322 set_conv_libfunc (trunc_optab, DFmode, TFmode, "_q_qtod");
9323 set_conv_libfunc (sfix_optab, SImode, TFmode, "_q_qtoi");
9324 set_conv_libfunc (ufix_optab, SImode, TFmode, "_q_qtou");
9325 set_conv_libfunc (sfloat_optab, TFmode, SImode, "_q_itoq");
9326 set_conv_libfunc (ufloat_optab, TFmode, SImode, "_q_utoq");
9327 }
9328 }
9329
9330
9331 /* Expand a block clear operation, and return 1 if successful. Return 0
9332 if we should let the compiler generate normal code.
9333
9334 operands[0] is the destination
9335 operands[1] is the length
9336 operands[3] is the alignment */
9337
9338 int
expand_block_clear(rtx operands[])9339 expand_block_clear (rtx operands[])
9340 {
9341 rtx orig_dest = operands[0];
9342 rtx bytes_rtx = operands[1];
9343 rtx align_rtx = operands[3];
9344 bool constp = (GET_CODE (bytes_rtx) == CONST_INT);
9345 HOST_WIDE_INT align;
9346 HOST_WIDE_INT bytes;
9347 int offset;
9348 int clear_bytes;
9349 int clear_step;
9350
9351 /* If this is not a fixed size move, just call memcpy */
9352 if (! constp)
9353 return 0;
9354
9355 /* This must be a fixed size alignment */
9356 gcc_assert (GET_CODE (align_rtx) == CONST_INT);
9357 align = INTVAL (align_rtx) * BITS_PER_UNIT;
9358
9359 /* Anything to clear? */
9360 bytes = INTVAL (bytes_rtx);
9361 if (bytes <= 0)
9362 return 1;
9363
9364 /* Use the builtin memset after a point, to avoid huge code bloat.
9365 When optimize_size, avoid any significant code bloat; calling
9366 memset is about 4 instructions, so allow for one instruction to
9367 load zero and three to do clearing. */
9368 if (TARGET_ALTIVEC && align >= 128)
9369 clear_step = 16;
9370 else if (TARGET_POWERPC64 && align >= 32)
9371 clear_step = 8;
9372 else
9373 clear_step = 4;
9374
9375 if (optimize_size && bytes > 3 * clear_step)
9376 return 0;
9377 if (! optimize_size && bytes > 8 * clear_step)
9378 return 0;
9379
9380 for (offset = 0; bytes > 0; offset += clear_bytes, bytes -= clear_bytes)
9381 {
9382 enum machine_mode mode = BLKmode;
9383 rtx dest;
9384
9385 if (bytes >= 16 && TARGET_ALTIVEC && align >= 128)
9386 {
9387 clear_bytes = 16;
9388 mode = V4SImode;
9389 }
9390 else if (bytes >= 8 && TARGET_POWERPC64
9391 /* 64-bit loads and stores require word-aligned
9392 displacements. */
9393 && (align >= 64 || (!STRICT_ALIGNMENT && align >= 32)))
9394 {
9395 clear_bytes = 8;
9396 mode = DImode;
9397 }
9398 else if (bytes >= 4 && (align >= 32 || !STRICT_ALIGNMENT))
9399 { /* move 4 bytes */
9400 clear_bytes = 4;
9401 mode = SImode;
9402 }
9403 else if (bytes >= 2 && (align >= 16 || !STRICT_ALIGNMENT))
9404 { /* move 2 bytes */
9405 clear_bytes = 2;
9406 mode = HImode;
9407 }
9408 else /* move 1 byte at a time */
9409 {
9410 clear_bytes = 1;
9411 mode = QImode;
9412 }
9413
9414 dest = adjust_address (orig_dest, mode, offset);
9415
9416 emit_move_insn (dest, CONST0_RTX (mode));
9417 }
9418
9419 return 1;
9420 }
9421
9422
9423 /* Expand a block move operation, and return 1 if successful. Return 0
9424 if we should let the compiler generate normal code.
9425
9426 operands[0] is the destination
9427 operands[1] is the source
9428 operands[2] is the length
9429 operands[3] is the alignment */
9430
9431 #define MAX_MOVE_REG 4
9432
9433 int
expand_block_move(rtx operands[])9434 expand_block_move (rtx operands[])
9435 {
9436 rtx orig_dest = operands[0];
9437 rtx orig_src = operands[1];
9438 rtx bytes_rtx = operands[2];
9439 rtx align_rtx = operands[3];
9440 int constp = (GET_CODE (bytes_rtx) == CONST_INT);
9441 int align;
9442 int bytes;
9443 int offset;
9444 int move_bytes;
9445 rtx stores[MAX_MOVE_REG];
9446 int num_reg = 0;
9447
9448 /* If this is not a fixed size move, just call memcpy */
9449 if (! constp)
9450 return 0;
9451
9452 /* This must be a fixed size alignment */
9453 gcc_assert (GET_CODE (align_rtx) == CONST_INT);
9454 align = INTVAL (align_rtx) * BITS_PER_UNIT;
9455
9456 /* Anything to move? */
9457 bytes = INTVAL (bytes_rtx);
9458 if (bytes <= 0)
9459 return 1;
9460
9461 /* store_one_arg depends on expand_block_move to handle at least the size of
9462 reg_parm_stack_space. */
9463 if (bytes > (TARGET_POWERPC64 ? 64 : 32))
9464 return 0;
9465
9466 for (offset = 0; bytes > 0; offset += move_bytes, bytes -= move_bytes)
9467 {
9468 union {
9469 rtx (*movmemsi) (rtx, rtx, rtx, rtx);
9470 rtx (*mov) (rtx, rtx);
9471 } gen_func;
9472 enum machine_mode mode = BLKmode;
9473 rtx src, dest;
9474
9475 /* Altivec first, since it will be faster than a string move
9476 when it applies, and usually not significantly larger. */
9477 if (TARGET_ALTIVEC && bytes >= 16 && align >= 128)
9478 {
9479 move_bytes = 16;
9480 mode = V4SImode;
9481 gen_func.mov = gen_movv4si;
9482 }
9483 else if (TARGET_STRING
9484 && bytes > 24 /* move up to 32 bytes at a time */
9485 && ! fixed_regs[5]
9486 && ! fixed_regs[6]
9487 && ! fixed_regs[7]
9488 && ! fixed_regs[8]
9489 && ! fixed_regs[9]
9490 && ! fixed_regs[10]
9491 && ! fixed_regs[11]
9492 && ! fixed_regs[12])
9493 {
9494 move_bytes = (bytes > 32) ? 32 : bytes;
9495 gen_func.movmemsi = gen_movmemsi_8reg;
9496 }
9497 else if (TARGET_STRING
9498 && bytes > 16 /* move up to 24 bytes at a time */
9499 && ! fixed_regs[5]
9500 && ! fixed_regs[6]
9501 && ! fixed_regs[7]
9502 && ! fixed_regs[8]
9503 && ! fixed_regs[9]
9504 && ! fixed_regs[10])
9505 {
9506 move_bytes = (bytes > 24) ? 24 : bytes;
9507 gen_func.movmemsi = gen_movmemsi_6reg;
9508 }
9509 else if (TARGET_STRING
9510 && bytes > 8 /* move up to 16 bytes at a time */
9511 && ! fixed_regs[5]
9512 && ! fixed_regs[6]
9513 && ! fixed_regs[7]
9514 && ! fixed_regs[8])
9515 {
9516 move_bytes = (bytes > 16) ? 16 : bytes;
9517 gen_func.movmemsi = gen_movmemsi_4reg;
9518 }
9519 else if (bytes >= 8 && TARGET_POWERPC64
9520 /* 64-bit loads and stores require word-aligned
9521 displacements. */
9522 && (align >= 64 || (!STRICT_ALIGNMENT && align >= 32)))
9523 {
9524 move_bytes = 8;
9525 mode = DImode;
9526 gen_func.mov = gen_movdi;
9527 }
9528 else if (TARGET_STRING && bytes > 4 && !TARGET_POWERPC64)
9529 { /* move up to 8 bytes at a time */
9530 move_bytes = (bytes > 8) ? 8 : bytes;
9531 gen_func.movmemsi = gen_movmemsi_2reg;
9532 }
9533 else if (bytes >= 4 && (align >= 32 || !STRICT_ALIGNMENT))
9534 { /* move 4 bytes */
9535 move_bytes = 4;
9536 mode = SImode;
9537 gen_func.mov = gen_movsi;
9538 }
9539 else if (bytes >= 2 && (align >= 16 || !STRICT_ALIGNMENT))
9540 { /* move 2 bytes */
9541 move_bytes = 2;
9542 mode = HImode;
9543 gen_func.mov = gen_movhi;
9544 }
9545 else if (TARGET_STRING && bytes > 1)
9546 { /* move up to 4 bytes at a time */
9547 move_bytes = (bytes > 4) ? 4 : bytes;
9548 gen_func.movmemsi = gen_movmemsi_1reg;
9549 }
9550 else /* move 1 byte at a time */
9551 {
9552 move_bytes = 1;
9553 mode = QImode;
9554 gen_func.mov = gen_movqi;
9555 }
9556
9557 src = adjust_address (orig_src, mode, offset);
9558 dest = adjust_address (orig_dest, mode, offset);
9559
9560 if (mode != BLKmode)
9561 {
9562 rtx tmp_reg = gen_reg_rtx (mode);
9563
9564 emit_insn ((*gen_func.mov) (tmp_reg, src));
9565 stores[num_reg++] = (*gen_func.mov) (dest, tmp_reg);
9566 }
9567
9568 if (mode == BLKmode || num_reg >= MAX_MOVE_REG || bytes == move_bytes)
9569 {
9570 int i;
9571 for (i = 0; i < num_reg; i++)
9572 emit_insn (stores[i]);
9573 num_reg = 0;
9574 }
9575
9576 if (mode == BLKmode)
9577 {
9578 /* Move the address into scratch registers. The movmemsi
9579 patterns require zero offset. */
9580 if (!REG_P (XEXP (src, 0)))
9581 {
9582 rtx src_reg = copy_addr_to_reg (XEXP (src, 0));
9583 src = replace_equiv_address (src, src_reg);
9584 }
9585 set_mem_size (src, GEN_INT (move_bytes));
9586
9587 if (!REG_P (XEXP (dest, 0)))
9588 {
9589 rtx dest_reg = copy_addr_to_reg (XEXP (dest, 0));
9590 dest = replace_equiv_address (dest, dest_reg);
9591 }
9592 set_mem_size (dest, GEN_INT (move_bytes));
9593
9594 emit_insn ((*gen_func.movmemsi) (dest, src,
9595 GEN_INT (move_bytes & 31),
9596 align_rtx));
9597 }
9598 }
9599
9600 return 1;
9601 }
9602
9603
9604 /* Return a string to perform a load_multiple operation.
9605 operands[0] is the vector.
9606 operands[1] is the source address.
9607 operands[2] is the first destination register. */
9608
9609 const char *
rs6000_output_load_multiple(rtx operands[3])9610 rs6000_output_load_multiple (rtx operands[3])
9611 {
9612 /* We have to handle the case where the pseudo used to contain the address
9613 is assigned to one of the output registers. */
9614 int i, j;
9615 int words = XVECLEN (operands[0], 0);
9616 rtx xop[10];
9617
9618 if (XVECLEN (operands[0], 0) == 1)
9619 return "{l|lwz} %2,0(%1)";
9620
9621 for (i = 0; i < words; i++)
9622 if (refers_to_regno_p (REGNO (operands[2]) + i,
9623 REGNO (operands[2]) + i + 1, operands[1], 0))
9624 {
9625 if (i == words-1)
9626 {
9627 xop[0] = GEN_INT (4 * (words-1));
9628 xop[1] = operands[1];
9629 xop[2] = operands[2];
9630 output_asm_insn ("{lsi|lswi} %2,%1,%0\n\t{l|lwz} %1,%0(%1)", xop);
9631 return "";
9632 }
9633 else if (i == 0)
9634 {
9635 xop[0] = GEN_INT (4 * (words-1));
9636 xop[1] = operands[1];
9637 xop[2] = gen_rtx_REG (SImode, REGNO (operands[2]) + 1);
9638 output_asm_insn ("{cal %1,4(%1)|addi %1,%1,4}\n\t{lsi|lswi} %2,%1,%0\n\t{l|lwz} %1,-4(%1)", xop);
9639 return "";
9640 }
9641 else
9642 {
9643 for (j = 0; j < words; j++)
9644 if (j != i)
9645 {
9646 xop[0] = GEN_INT (j * 4);
9647 xop[1] = operands[1];
9648 xop[2] = gen_rtx_REG (SImode, REGNO (operands[2]) + j);
9649 output_asm_insn ("{l|lwz} %2,%0(%1)", xop);
9650 }
9651 xop[0] = GEN_INT (i * 4);
9652 xop[1] = operands[1];
9653 output_asm_insn ("{l|lwz} %1,%0(%1)", xop);
9654 return "";
9655 }
9656 }
9657
9658 return "{lsi|lswi} %2,%1,%N0";
9659 }
9660
9661
9662 /* A validation routine: say whether CODE, a condition code, and MODE
9663 match. The other alternatives either don't make sense or should
9664 never be generated. */
9665
9666 void
validate_condition_mode(enum rtx_code code,enum machine_mode mode)9667 validate_condition_mode (enum rtx_code code, enum machine_mode mode)
9668 {
9669 gcc_assert ((GET_RTX_CLASS (code) == RTX_COMPARE
9670 || GET_RTX_CLASS (code) == RTX_COMM_COMPARE)
9671 && GET_MODE_CLASS (mode) == MODE_CC);
9672
9673 /* These don't make sense. */
9674 gcc_assert ((code != GT && code != LT && code != GE && code != LE)
9675 || mode != CCUNSmode);
9676
9677 gcc_assert ((code != GTU && code != LTU && code != GEU && code != LEU)
9678 || mode == CCUNSmode);
9679
9680 gcc_assert (mode == CCFPmode
9681 || (code != ORDERED && code != UNORDERED
9682 && code != UNEQ && code != LTGT
9683 && code != UNGT && code != UNLT
9684 && code != UNGE && code != UNLE));
9685
9686 /* These should never be generated except for
9687 flag_finite_math_only. */
9688 gcc_assert (mode != CCFPmode
9689 || flag_finite_math_only
9690 || (code != LE && code != GE
9691 && code != UNEQ && code != LTGT
9692 && code != UNGT && code != UNLT));
9693
9694 /* These are invalid; the information is not there. */
9695 gcc_assert (mode != CCEQmode || code == EQ || code == NE);
9696 }
9697
9698
9699 /* Return 1 if ANDOP is a mask that has no bits on that are not in the
9700 mask required to convert the result of a rotate insn into a shift
9701 left insn of SHIFTOP bits. Both are known to be SImode CONST_INT. */
9702
9703 int
includes_lshift_p(rtx shiftop,rtx andop)9704 includes_lshift_p (rtx shiftop, rtx andop)
9705 {
9706 unsigned HOST_WIDE_INT shift_mask = ~(unsigned HOST_WIDE_INT) 0;
9707
9708 shift_mask <<= INTVAL (shiftop);
9709
9710 return (INTVAL (andop) & 0xffffffff & ~shift_mask) == 0;
9711 }
9712
9713 /* Similar, but for right shift. */
9714
9715 int
includes_rshift_p(rtx shiftop,rtx andop)9716 includes_rshift_p (rtx shiftop, rtx andop)
9717 {
9718 unsigned HOST_WIDE_INT shift_mask = ~(unsigned HOST_WIDE_INT) 0;
9719
9720 shift_mask >>= INTVAL (shiftop);
9721
9722 return (INTVAL (andop) & 0xffffffff & ~shift_mask) == 0;
9723 }
9724
9725 /* Return 1 if ANDOP is a mask suitable for use with an rldic insn
9726 to perform a left shift. It must have exactly SHIFTOP least
9727 significant 0's, then one or more 1's, then zero or more 0's. */
9728
9729 int
includes_rldic_lshift_p(rtx shiftop,rtx andop)9730 includes_rldic_lshift_p (rtx shiftop, rtx andop)
9731 {
9732 if (GET_CODE (andop) == CONST_INT)
9733 {
9734 HOST_WIDE_INT c, lsb, shift_mask;
9735
9736 c = INTVAL (andop);
9737 if (c == 0 || c == ~0)
9738 return 0;
9739
9740 shift_mask = ~0;
9741 shift_mask <<= INTVAL (shiftop);
9742
9743 /* Find the least significant one bit. */
9744 lsb = c & -c;
9745
9746 /* It must coincide with the LSB of the shift mask. */
9747 if (-lsb != shift_mask)
9748 return 0;
9749
9750 /* Invert to look for the next transition (if any). */
9751 c = ~c;
9752
9753 /* Remove the low group of ones (originally low group of zeros). */
9754 c &= -lsb;
9755
9756 /* Again find the lsb, and check we have all 1's above. */
9757 lsb = c & -c;
9758 return c == -lsb;
9759 }
9760 else if (GET_CODE (andop) == CONST_DOUBLE
9761 && (GET_MODE (andop) == VOIDmode || GET_MODE (andop) == DImode))
9762 {
9763 HOST_WIDE_INT low, high, lsb;
9764 HOST_WIDE_INT shift_mask_low, shift_mask_high;
9765
9766 low = CONST_DOUBLE_LOW (andop);
9767 if (HOST_BITS_PER_WIDE_INT < 64)
9768 high = CONST_DOUBLE_HIGH (andop);
9769
9770 if ((low == 0 && (HOST_BITS_PER_WIDE_INT >= 64 || high == 0))
9771 || (low == ~0 && (HOST_BITS_PER_WIDE_INT >= 64 || high == ~0)))
9772 return 0;
9773
9774 if (HOST_BITS_PER_WIDE_INT < 64 && low == 0)
9775 {
9776 shift_mask_high = ~0;
9777 if (INTVAL (shiftop) > 32)
9778 shift_mask_high <<= INTVAL (shiftop) - 32;
9779
9780 lsb = high & -high;
9781
9782 if (-lsb != shift_mask_high || INTVAL (shiftop) < 32)
9783 return 0;
9784
9785 high = ~high;
9786 high &= -lsb;
9787
9788 lsb = high & -high;
9789 return high == -lsb;
9790 }
9791
9792 shift_mask_low = ~0;
9793 shift_mask_low <<= INTVAL (shiftop);
9794
9795 lsb = low & -low;
9796
9797 if (-lsb != shift_mask_low)
9798 return 0;
9799
9800 if (HOST_BITS_PER_WIDE_INT < 64)
9801 high = ~high;
9802 low = ~low;
9803 low &= -lsb;
9804
9805 if (HOST_BITS_PER_WIDE_INT < 64 && low == 0)
9806 {
9807 lsb = high & -high;
9808 return high == -lsb;
9809 }
9810
9811 lsb = low & -low;
9812 return low == -lsb && (HOST_BITS_PER_WIDE_INT >= 64 || high == ~0);
9813 }
9814 else
9815 return 0;
9816 }
9817
9818 /* Return 1 if ANDOP is a mask suitable for use with an rldicr insn
9819 to perform a left shift. It must have SHIFTOP or more least
9820 significant 0's, with the remainder of the word 1's. */
9821
9822 int
includes_rldicr_lshift_p(rtx shiftop,rtx andop)9823 includes_rldicr_lshift_p (rtx shiftop, rtx andop)
9824 {
9825 if (GET_CODE (andop) == CONST_INT)
9826 {
9827 HOST_WIDE_INT c, lsb, shift_mask;
9828
9829 shift_mask = ~0;
9830 shift_mask <<= INTVAL (shiftop);
9831 c = INTVAL (andop);
9832
9833 /* Find the least significant one bit. */
9834 lsb = c & -c;
9835
9836 /* It must be covered by the shift mask.
9837 This test also rejects c == 0. */
9838 if ((lsb & shift_mask) == 0)
9839 return 0;
9840
9841 /* Check we have all 1's above the transition, and reject all 1's. */
9842 return c == -lsb && lsb != 1;
9843 }
9844 else if (GET_CODE (andop) == CONST_DOUBLE
9845 && (GET_MODE (andop) == VOIDmode || GET_MODE (andop) == DImode))
9846 {
9847 HOST_WIDE_INT low, lsb, shift_mask_low;
9848
9849 low = CONST_DOUBLE_LOW (andop);
9850
9851 if (HOST_BITS_PER_WIDE_INT < 64)
9852 {
9853 HOST_WIDE_INT high, shift_mask_high;
9854
9855 high = CONST_DOUBLE_HIGH (andop);
9856
9857 if (low == 0)
9858 {
9859 shift_mask_high = ~0;
9860 if (INTVAL (shiftop) > 32)
9861 shift_mask_high <<= INTVAL (shiftop) - 32;
9862
9863 lsb = high & -high;
9864
9865 if ((lsb & shift_mask_high) == 0)
9866 return 0;
9867
9868 return high == -lsb;
9869 }
9870 if (high != ~0)
9871 return 0;
9872 }
9873
9874 shift_mask_low = ~0;
9875 shift_mask_low <<= INTVAL (shiftop);
9876
9877 lsb = low & -low;
9878
9879 if ((lsb & shift_mask_low) == 0)
9880 return 0;
9881
9882 return low == -lsb && lsb != 1;
9883 }
9884 else
9885 return 0;
9886 }
9887
9888 /* Return 1 if operands will generate a valid arguments to rlwimi
9889 instruction for insert with right shift in 64-bit mode. The mask may
9890 not start on the first bit or stop on the last bit because wrap-around
9891 effects of instruction do not correspond to semantics of RTL insn. */
9892
9893 int
insvdi_rshift_rlwimi_p(rtx sizeop,rtx startop,rtx shiftop)9894 insvdi_rshift_rlwimi_p (rtx sizeop, rtx startop, rtx shiftop)
9895 {
9896 if (INTVAL (startop) > 32
9897 && INTVAL (startop) < 64
9898 && INTVAL (sizeop) > 1
9899 && INTVAL (sizeop) + INTVAL (startop) < 64
9900 && INTVAL (shiftop) > 0
9901 && INTVAL (sizeop) + INTVAL (shiftop) < 32
9902 && (64 - (INTVAL (shiftop) & 63)) >= INTVAL (sizeop))
9903 return 1;
9904
9905 return 0;
9906 }
9907
9908 /* Return 1 if REGNO (reg1) == REGNO (reg2) - 1 making them candidates
9909 for lfq and stfq insns iff the registers are hard registers. */
9910
9911 int
registers_ok_for_quad_peep(rtx reg1,rtx reg2)9912 registers_ok_for_quad_peep (rtx reg1, rtx reg2)
9913 {
9914 /* We might have been passed a SUBREG. */
9915 if (GET_CODE (reg1) != REG || GET_CODE (reg2) != REG)
9916 return 0;
9917
9918 /* We might have been passed non floating point registers. */
9919 if (!FP_REGNO_P (REGNO (reg1))
9920 || !FP_REGNO_P (REGNO (reg2)))
9921 return 0;
9922
9923 return (REGNO (reg1) == REGNO (reg2) - 1);
9924 }
9925
9926 /* Return 1 if addr1 and addr2 are suitable for lfq or stfq insn.
9927 addr1 and addr2 must be in consecutive memory locations
9928 (addr2 == addr1 + 8). */
9929
9930 int
mems_ok_for_quad_peep(rtx mem1,rtx mem2)9931 mems_ok_for_quad_peep (rtx mem1, rtx mem2)
9932 {
9933 rtx addr1, addr2;
9934 unsigned int reg1, reg2;
9935 int offset1, offset2;
9936
9937 /* The mems cannot be volatile. */
9938 if (MEM_VOLATILE_P (mem1) || MEM_VOLATILE_P (mem2))
9939 return 0;
9940
9941 addr1 = XEXP (mem1, 0);
9942 addr2 = XEXP (mem2, 0);
9943
9944 /* Extract an offset (if used) from the first addr. */
9945 if (GET_CODE (addr1) == PLUS)
9946 {
9947 /* If not a REG, return zero. */
9948 if (GET_CODE (XEXP (addr1, 0)) != REG)
9949 return 0;
9950 else
9951 {
9952 reg1 = REGNO (XEXP (addr1, 0));
9953 /* The offset must be constant! */
9954 if (GET_CODE (XEXP (addr1, 1)) != CONST_INT)
9955 return 0;
9956 offset1 = INTVAL (XEXP (addr1, 1));
9957 }
9958 }
9959 else if (GET_CODE (addr1) != REG)
9960 return 0;
9961 else
9962 {
9963 reg1 = REGNO (addr1);
9964 /* This was a simple (mem (reg)) expression. Offset is 0. */
9965 offset1 = 0;
9966 }
9967
9968 /* And now for the second addr. */
9969 if (GET_CODE (addr2) == PLUS)
9970 {
9971 /* If not a REG, return zero. */
9972 if (GET_CODE (XEXP (addr2, 0)) != REG)
9973 return 0;
9974 else
9975 {
9976 reg2 = REGNO (XEXP (addr2, 0));
9977 /* The offset must be constant. */
9978 if (GET_CODE (XEXP (addr2, 1)) != CONST_INT)
9979 return 0;
9980 offset2 = INTVAL (XEXP (addr2, 1));
9981 }
9982 }
9983 else if (GET_CODE (addr2) != REG)
9984 return 0;
9985 else
9986 {
9987 reg2 = REGNO (addr2);
9988 /* This was a simple (mem (reg)) expression. Offset is 0. */
9989 offset2 = 0;
9990 }
9991
9992 /* Both of these must have the same base register. */
9993 if (reg1 != reg2)
9994 return 0;
9995
9996 /* The offset for the second addr must be 8 more than the first addr. */
9997 if (offset2 != offset1 + 8)
9998 return 0;
9999
10000 /* All the tests passed. addr1 and addr2 are valid for lfq or stfq
10001 instructions. */
10002 return 1;
10003 }
10004
10005 /* Return the register class of a scratch register needed to copy IN into
10006 or out of a register in CLASS in MODE. If it can be done directly,
10007 NO_REGS is returned. */
10008
10009 enum reg_class
rs6000_secondary_reload_class(enum reg_class class,enum machine_mode mode ATTRIBUTE_UNUSED,rtx in)10010 rs6000_secondary_reload_class (enum reg_class class,
10011 enum machine_mode mode ATTRIBUTE_UNUSED,
10012 rtx in)
10013 {
10014 int regno;
10015
10016 if (TARGET_ELF || (DEFAULT_ABI == ABI_DARWIN
10017 #if TARGET_MACHO
10018 && MACHOPIC_INDIRECT
10019 #endif
10020 ))
10021 {
10022 /* We cannot copy a symbolic operand directly into anything
10023 other than BASE_REGS for TARGET_ELF. So indicate that a
10024 register from BASE_REGS is needed as an intermediate
10025 register.
10026
10027 On Darwin, pic addresses require a load from memory, which
10028 needs a base register. */
10029 if (class != BASE_REGS
10030 && (GET_CODE (in) == SYMBOL_REF
10031 || GET_CODE (in) == HIGH
10032 || GET_CODE (in) == LABEL_REF
10033 || GET_CODE (in) == CONST))
10034 return BASE_REGS;
10035 }
10036
10037 if (GET_CODE (in) == REG)
10038 {
10039 regno = REGNO (in);
10040 if (regno >= FIRST_PSEUDO_REGISTER)
10041 {
10042 regno = true_regnum (in);
10043 if (regno >= FIRST_PSEUDO_REGISTER)
10044 regno = -1;
10045 }
10046 }
10047 else if (GET_CODE (in) == SUBREG)
10048 {
10049 regno = true_regnum (in);
10050 if (regno >= FIRST_PSEUDO_REGISTER)
10051 regno = -1;
10052 }
10053 else
10054 regno = -1;
10055
10056 /* We can place anything into GENERAL_REGS and can put GENERAL_REGS
10057 into anything. */
10058 if (class == GENERAL_REGS || class == BASE_REGS
10059 || (regno >= 0 && INT_REGNO_P (regno)))
10060 return NO_REGS;
10061
10062 /* Constants, memory, and FP registers can go into FP registers. */
10063 if ((regno == -1 || FP_REGNO_P (regno))
10064 && (class == FLOAT_REGS || class == NON_SPECIAL_REGS))
10065 return NO_REGS;
10066
10067 /* Memory, and AltiVec registers can go into AltiVec registers. */
10068 if ((regno == -1 || ALTIVEC_REGNO_P (regno))
10069 && class == ALTIVEC_REGS)
10070 return NO_REGS;
10071
10072 /* We can copy among the CR registers. */
10073 if ((class == CR_REGS || class == CR0_REGS)
10074 && regno >= 0 && CR_REGNO_P (regno))
10075 return NO_REGS;
10076
10077 /* Otherwise, we need GENERAL_REGS. */
10078 return GENERAL_REGS;
10079 }
10080
10081 /* Given a comparison operation, return the bit number in CCR to test. We
10082 know this is a valid comparison.
10083
10084 SCC_P is 1 if this is for an scc. That means that %D will have been
10085 used instead of %C, so the bits will be in different places.
10086
10087 Return -1 if OP isn't a valid comparison for some reason. */
10088
10089 int
ccr_bit(rtx op,int scc_p)10090 ccr_bit (rtx op, int scc_p)
10091 {
10092 enum rtx_code code = GET_CODE (op);
10093 enum machine_mode cc_mode;
10094 int cc_regnum;
10095 int base_bit;
10096 rtx reg;
10097
10098 if (!COMPARISON_P (op))
10099 return -1;
10100
10101 reg = XEXP (op, 0);
10102
10103 gcc_assert (GET_CODE (reg) == REG && CR_REGNO_P (REGNO (reg)));
10104
10105 cc_mode = GET_MODE (reg);
10106 cc_regnum = REGNO (reg);
10107 base_bit = 4 * (cc_regnum - CR0_REGNO);
10108
10109 validate_condition_mode (code, cc_mode);
10110
10111 /* When generating a sCOND operation, only positive conditions are
10112 allowed. */
10113 gcc_assert (!scc_p
10114 || code == EQ || code == GT || code == LT || code == UNORDERED
10115 || code == GTU || code == LTU);
10116
10117 switch (code)
10118 {
10119 case NE:
10120 return scc_p ? base_bit + 3 : base_bit + 2;
10121 case EQ:
10122 return base_bit + 2;
10123 case GT: case GTU: case UNLE:
10124 return base_bit + 1;
10125 case LT: case LTU: case UNGE:
10126 return base_bit;
10127 case ORDERED: case UNORDERED:
10128 return base_bit + 3;
10129
10130 case GE: case GEU:
10131 /* If scc, we will have done a cror to put the bit in the
10132 unordered position. So test that bit. For integer, this is ! LT
10133 unless this is an scc insn. */
10134 return scc_p ? base_bit + 3 : base_bit;
10135
10136 case LE: case LEU:
10137 return scc_p ? base_bit + 3 : base_bit + 1;
10138
10139 default:
10140 gcc_unreachable ();
10141 }
10142 }
10143
10144 /* Return the GOT register. */
10145
10146 rtx
rs6000_got_register(rtx value ATTRIBUTE_UNUSED)10147 rs6000_got_register (rtx value ATTRIBUTE_UNUSED)
10148 {
10149 /* The second flow pass currently (June 1999) can't update
10150 regs_ever_live without disturbing other parts of the compiler, so
10151 update it here to make the prolog/epilogue code happy. */
10152 if (no_new_pseudos && ! regs_ever_live[RS6000_PIC_OFFSET_TABLE_REGNUM])
10153 regs_ever_live[RS6000_PIC_OFFSET_TABLE_REGNUM] = 1;
10154
10155 current_function_uses_pic_offset_table = 1;
10156
10157 return pic_offset_table_rtx;
10158 }
10159
10160 /* Function to init struct machine_function.
10161 This will be called, via a pointer variable,
10162 from push_function_context. */
10163
10164 static struct machine_function *
rs6000_init_machine_status(void)10165 rs6000_init_machine_status (void)
10166 {
10167 return ggc_alloc_cleared (sizeof (machine_function));
10168 }
10169
10170 /* These macros test for integers and extract the low-order bits. */
10171 #define INT_P(X) \
10172 ((GET_CODE (X) == CONST_INT || GET_CODE (X) == CONST_DOUBLE) \
10173 && GET_MODE (X) == VOIDmode)
10174
10175 #define INT_LOWPART(X) \
10176 (GET_CODE (X) == CONST_INT ? INTVAL (X) : CONST_DOUBLE_LOW (X))
10177
10178 int
extract_MB(rtx op)10179 extract_MB (rtx op)
10180 {
10181 int i;
10182 unsigned long val = INT_LOWPART (op);
10183
10184 /* If the high bit is zero, the value is the first 1 bit we find
10185 from the left. */
10186 if ((val & 0x80000000) == 0)
10187 {
10188 gcc_assert (val & 0xffffffff);
10189
10190 i = 1;
10191 while (((val <<= 1) & 0x80000000) == 0)
10192 ++i;
10193 return i;
10194 }
10195
10196 /* If the high bit is set and the low bit is not, or the mask is all
10197 1's, the value is zero. */
10198 if ((val & 1) == 0 || (val & 0xffffffff) == 0xffffffff)
10199 return 0;
10200
10201 /* Otherwise we have a wrap-around mask. Look for the first 0 bit
10202 from the right. */
10203 i = 31;
10204 while (((val >>= 1) & 1) != 0)
10205 --i;
10206
10207 return i;
10208 }
10209
10210 int
extract_ME(rtx op)10211 extract_ME (rtx op)
10212 {
10213 int i;
10214 unsigned long val = INT_LOWPART (op);
10215
10216 /* If the low bit is zero, the value is the first 1 bit we find from
10217 the right. */
10218 if ((val & 1) == 0)
10219 {
10220 gcc_assert (val & 0xffffffff);
10221
10222 i = 30;
10223 while (((val >>= 1) & 1) == 0)
10224 --i;
10225
10226 return i;
10227 }
10228
10229 /* If the low bit is set and the high bit is not, or the mask is all
10230 1's, the value is 31. */
10231 if ((val & 0x80000000) == 0 || (val & 0xffffffff) == 0xffffffff)
10232 return 31;
10233
10234 /* Otherwise we have a wrap-around mask. Look for the first 0 bit
10235 from the left. */
10236 i = 0;
10237 while (((val <<= 1) & 0x80000000) != 0)
10238 ++i;
10239
10240 return i;
10241 }
10242
10243 /* Locate some local-dynamic symbol still in use by this function
10244 so that we can print its name in some tls_ld pattern. */
10245
10246 static const char *
rs6000_get_some_local_dynamic_name(void)10247 rs6000_get_some_local_dynamic_name (void)
10248 {
10249 rtx insn;
10250
10251 if (cfun->machine->some_ld_name)
10252 return cfun->machine->some_ld_name;
10253
10254 for (insn = get_insns (); insn ; insn = NEXT_INSN (insn))
10255 if (INSN_P (insn)
10256 && for_each_rtx (&PATTERN (insn),
10257 rs6000_get_some_local_dynamic_name_1, 0))
10258 return cfun->machine->some_ld_name;
10259
10260 gcc_unreachable ();
10261 }
10262
10263 /* Helper function for rs6000_get_some_local_dynamic_name. */
10264
10265 static int
rs6000_get_some_local_dynamic_name_1(rtx * px,void * data ATTRIBUTE_UNUSED)10266 rs6000_get_some_local_dynamic_name_1 (rtx *px, void *data ATTRIBUTE_UNUSED)
10267 {
10268 rtx x = *px;
10269
10270 if (GET_CODE (x) == SYMBOL_REF)
10271 {
10272 const char *str = XSTR (x, 0);
10273 if (SYMBOL_REF_TLS_MODEL (x) == TLS_MODEL_LOCAL_DYNAMIC)
10274 {
10275 cfun->machine->some_ld_name = str;
10276 return 1;
10277 }
10278 }
10279
10280 return 0;
10281 }
10282
10283 /* Write out a function code label. */
10284
10285 void
rs6000_output_function_entry(FILE * file,const char * fname)10286 rs6000_output_function_entry (FILE *file, const char *fname)
10287 {
10288 if (fname[0] != '.')
10289 {
10290 switch (DEFAULT_ABI)
10291 {
10292 default:
10293 gcc_unreachable ();
10294
10295 case ABI_AIX:
10296 if (DOT_SYMBOLS)
10297 putc ('.', file);
10298 else
10299 ASM_OUTPUT_INTERNAL_LABEL_PREFIX (file, "L.");
10300 break;
10301
10302 case ABI_V4:
10303 case ABI_DARWIN:
10304 break;
10305 }
10306 }
10307 if (TARGET_AIX)
10308 RS6000_OUTPUT_BASENAME (file, fname);
10309 else
10310 assemble_name (file, fname);
10311 }
10312
10313 /* Print an operand. Recognize special options, documented below. */
10314
10315 #if TARGET_ELF
10316 #define SMALL_DATA_RELOC ((rs6000_sdata == SDATA_EABI) ? "sda21" : "sdarel")
10317 #define SMALL_DATA_REG ((rs6000_sdata == SDATA_EABI) ? 0 : 13)
10318 #else
10319 #define SMALL_DATA_RELOC "sda21"
10320 #define SMALL_DATA_REG 0
10321 #endif
10322
10323 void
print_operand(FILE * file,rtx x,int code)10324 print_operand (FILE *file, rtx x, int code)
10325 {
10326 int i;
10327 HOST_WIDE_INT val;
10328 unsigned HOST_WIDE_INT uval;
10329
10330 switch (code)
10331 {
10332 case '.':
10333 /* Write out an instruction after the call which may be replaced
10334 with glue code by the loader. This depends on the AIX version. */
10335 asm_fprintf (file, RS6000_CALL_GLUE);
10336 return;
10337
10338 /* %a is output_address. */
10339
10340 case 'A':
10341 /* If X is a constant integer whose low-order 5 bits are zero,
10342 write 'l'. Otherwise, write 'r'. This is a kludge to fix a bug
10343 in the AIX assembler where "sri" with a zero shift count
10344 writes a trash instruction. */
10345 if (GET_CODE (x) == CONST_INT && (INTVAL (x) & 31) == 0)
10346 putc ('l', file);
10347 else
10348 putc ('r', file);
10349 return;
10350
10351 case 'b':
10352 /* If constant, low-order 16 bits of constant, unsigned.
10353 Otherwise, write normally. */
10354 if (INT_P (x))
10355 fprintf (file, HOST_WIDE_INT_PRINT_DEC, INT_LOWPART (x) & 0xffff);
10356 else
10357 print_operand (file, x, 0);
10358 return;
10359
10360 case 'B':
10361 /* If the low-order bit is zero, write 'r'; otherwise, write 'l'
10362 for 64-bit mask direction. */
10363 putc (((INT_LOWPART (x) & 1) == 0 ? 'r' : 'l'), file);
10364 return;
10365
10366 /* %c is output_addr_const if a CONSTANT_ADDRESS_P, otherwise
10367 output_operand. */
10368
10369 case 'c':
10370 /* X is a CR register. Print the number of the GT bit of the CR. */
10371 if (GET_CODE (x) != REG || ! CR_REGNO_P (REGNO (x)))
10372 output_operand_lossage ("invalid %%E value");
10373 else
10374 fprintf (file, "%d", 4 * (REGNO (x) - CR0_REGNO) + 1);
10375 return;
10376
10377 case 'D':
10378 /* Like 'J' but get to the GT bit only. */
10379 gcc_assert (GET_CODE (x) == REG);
10380
10381 /* Bit 1 is GT bit. */
10382 i = 4 * (REGNO (x) - CR0_REGNO) + 1;
10383
10384 /* Add one for shift count in rlinm for scc. */
10385 fprintf (file, "%d", i + 1);
10386 return;
10387
10388 case 'E':
10389 /* X is a CR register. Print the number of the EQ bit of the CR */
10390 if (GET_CODE (x) != REG || ! CR_REGNO_P (REGNO (x)))
10391 output_operand_lossage ("invalid %%E value");
10392 else
10393 fprintf (file, "%d", 4 * (REGNO (x) - CR0_REGNO) + 2);
10394 return;
10395
10396 case 'f':
10397 /* X is a CR register. Print the shift count needed to move it
10398 to the high-order four bits. */
10399 if (GET_CODE (x) != REG || ! CR_REGNO_P (REGNO (x)))
10400 output_operand_lossage ("invalid %%f value");
10401 else
10402 fprintf (file, "%d", 4 * (REGNO (x) - CR0_REGNO));
10403 return;
10404
10405 case 'F':
10406 /* Similar, but print the count for the rotate in the opposite
10407 direction. */
10408 if (GET_CODE (x) != REG || ! CR_REGNO_P (REGNO (x)))
10409 output_operand_lossage ("invalid %%F value");
10410 else
10411 fprintf (file, "%d", 32 - 4 * (REGNO (x) - CR0_REGNO));
10412 return;
10413
10414 case 'G':
10415 /* X is a constant integer. If it is negative, print "m",
10416 otherwise print "z". This is to make an aze or ame insn. */
10417 if (GET_CODE (x) != CONST_INT)
10418 output_operand_lossage ("invalid %%G value");
10419 else if (INTVAL (x) >= 0)
10420 putc ('z', file);
10421 else
10422 putc ('m', file);
10423 return;
10424
10425 case 'h':
10426 /* If constant, output low-order five bits. Otherwise, write
10427 normally. */
10428 if (INT_P (x))
10429 fprintf (file, HOST_WIDE_INT_PRINT_DEC, INT_LOWPART (x) & 31);
10430 else
10431 print_operand (file, x, 0);
10432 return;
10433
10434 case 'H':
10435 /* If constant, output low-order six bits. Otherwise, write
10436 normally. */
10437 if (INT_P (x))
10438 fprintf (file, HOST_WIDE_INT_PRINT_DEC, INT_LOWPART (x) & 63);
10439 else
10440 print_operand (file, x, 0);
10441 return;
10442
10443 case 'I':
10444 /* Print `i' if this is a constant, else nothing. */
10445 if (INT_P (x))
10446 putc ('i', file);
10447 return;
10448
10449 case 'j':
10450 /* Write the bit number in CCR for jump. */
10451 i = ccr_bit (x, 0);
10452 if (i == -1)
10453 output_operand_lossage ("invalid %%j code");
10454 else
10455 fprintf (file, "%d", i);
10456 return;
10457
10458 case 'J':
10459 /* Similar, but add one for shift count in rlinm for scc and pass
10460 scc flag to `ccr_bit'. */
10461 i = ccr_bit (x, 1);
10462 if (i == -1)
10463 output_operand_lossage ("invalid %%J code");
10464 else
10465 /* If we want bit 31, write a shift count of zero, not 32. */
10466 fprintf (file, "%d", i == 31 ? 0 : i + 1);
10467 return;
10468
10469 case 'k':
10470 /* X must be a constant. Write the 1's complement of the
10471 constant. */
10472 if (! INT_P (x))
10473 output_operand_lossage ("invalid %%k value");
10474 else
10475 fprintf (file, HOST_WIDE_INT_PRINT_DEC, ~ INT_LOWPART (x));
10476 return;
10477
10478 case 'K':
10479 /* X must be a symbolic constant on ELF. Write an
10480 expression suitable for an 'addi' that adds in the low 16
10481 bits of the MEM. */
10482 if (GET_CODE (x) != CONST)
10483 {
10484 print_operand_address (file, x);
10485 fputs ("@l", file);
10486 }
10487 else
10488 {
10489 if (GET_CODE (XEXP (x, 0)) != PLUS
10490 || (GET_CODE (XEXP (XEXP (x, 0), 0)) != SYMBOL_REF
10491 && GET_CODE (XEXP (XEXP (x, 0), 0)) != LABEL_REF)
10492 || GET_CODE (XEXP (XEXP (x, 0), 1)) != CONST_INT)
10493 output_operand_lossage ("invalid %%K value");
10494 print_operand_address (file, XEXP (XEXP (x, 0), 0));
10495 fputs ("@l", file);
10496 /* For GNU as, there must be a non-alphanumeric character
10497 between 'l' and the number. The '-' is added by
10498 print_operand() already. */
10499 if (INTVAL (XEXP (XEXP (x, 0), 1)) >= 0)
10500 fputs ("+", file);
10501 print_operand (file, XEXP (XEXP (x, 0), 1), 0);
10502 }
10503 return;
10504
10505 /* %l is output_asm_label. */
10506
10507 case 'L':
10508 /* Write second word of DImode or DFmode reference. Works on register
10509 or non-indexed memory only. */
10510 if (GET_CODE (x) == REG)
10511 fputs (reg_names[REGNO (x) + 1], file);
10512 else if (GET_CODE (x) == MEM)
10513 {
10514 /* Handle possible auto-increment. Since it is pre-increment and
10515 we have already done it, we can just use an offset of word. */
10516 if (GET_CODE (XEXP (x, 0)) == PRE_INC
10517 || GET_CODE (XEXP (x, 0)) == PRE_DEC)
10518 output_address (plus_constant (XEXP (XEXP (x, 0), 0),
10519 UNITS_PER_WORD));
10520 else
10521 output_address (XEXP (adjust_address_nv (x, SImode,
10522 UNITS_PER_WORD),
10523 0));
10524
10525 if (small_data_operand (x, GET_MODE (x)))
10526 fprintf (file, "@%s(%s)", SMALL_DATA_RELOC,
10527 reg_names[SMALL_DATA_REG]);
10528 }
10529 return;
10530
10531 case 'm':
10532 /* MB value for a mask operand. */
10533 if (! mask_operand (x, SImode))
10534 output_operand_lossage ("invalid %%m value");
10535
10536 fprintf (file, "%d", extract_MB (x));
10537 return;
10538
10539 case 'M':
10540 /* ME value for a mask operand. */
10541 if (! mask_operand (x, SImode))
10542 output_operand_lossage ("invalid %%M value");
10543
10544 fprintf (file, "%d", extract_ME (x));
10545 return;
10546
10547 /* %n outputs the negative of its operand. */
10548
10549 case 'N':
10550 /* Write the number of elements in the vector times 4. */
10551 if (GET_CODE (x) != PARALLEL)
10552 output_operand_lossage ("invalid %%N value");
10553 else
10554 fprintf (file, "%d", XVECLEN (x, 0) * 4);
10555 return;
10556
10557 case 'O':
10558 /* Similar, but subtract 1 first. */
10559 if (GET_CODE (x) != PARALLEL)
10560 output_operand_lossage ("invalid %%O value");
10561 else
10562 fprintf (file, "%d", (XVECLEN (x, 0) - 1) * 4);
10563 return;
10564
10565 case 'p':
10566 /* X is a CONST_INT that is a power of two. Output the logarithm. */
10567 if (! INT_P (x)
10568 || INT_LOWPART (x) < 0
10569 || (i = exact_log2 (INT_LOWPART (x))) < 0)
10570 output_operand_lossage ("invalid %%p value");
10571 else
10572 fprintf (file, "%d", i);
10573 return;
10574
10575 case 'P':
10576 /* The operand must be an indirect memory reference. The result
10577 is the register name. */
10578 if (GET_CODE (x) != MEM || GET_CODE (XEXP (x, 0)) != REG
10579 || REGNO (XEXP (x, 0)) >= 32)
10580 output_operand_lossage ("invalid %%P value");
10581 else
10582 fputs (reg_names[REGNO (XEXP (x, 0))], file);
10583 return;
10584
10585 case 'q':
10586 /* This outputs the logical code corresponding to a boolean
10587 expression. The expression may have one or both operands
10588 negated (if one, only the first one). For condition register
10589 logical operations, it will also treat the negated
10590 CR codes as NOTs, but not handle NOTs of them. */
10591 {
10592 const char *const *t = 0;
10593 const char *s;
10594 enum rtx_code code = GET_CODE (x);
10595 static const char * const tbl[3][3] = {
10596 { "and", "andc", "nor" },
10597 { "or", "orc", "nand" },
10598 { "xor", "eqv", "xor" } };
10599
10600 if (code == AND)
10601 t = tbl[0];
10602 else if (code == IOR)
10603 t = tbl[1];
10604 else if (code == XOR)
10605 t = tbl[2];
10606 else
10607 output_operand_lossage ("invalid %%q value");
10608
10609 if (GET_CODE (XEXP (x, 0)) != NOT)
10610 s = t[0];
10611 else
10612 {
10613 if (GET_CODE (XEXP (x, 1)) == NOT)
10614 s = t[2];
10615 else
10616 s = t[1];
10617 }
10618
10619 fputs (s, file);
10620 }
10621 return;
10622
10623 case 'Q':
10624 if (TARGET_MFCRF)
10625 fputc (',', file);
10626 /* FALLTHRU */
10627 else
10628 return;
10629
10630 case 'R':
10631 /* X is a CR register. Print the mask for `mtcrf'. */
10632 if (GET_CODE (x) != REG || ! CR_REGNO_P (REGNO (x)))
10633 output_operand_lossage ("invalid %%R value");
10634 else
10635 fprintf (file, "%d", 128 >> (REGNO (x) - CR0_REGNO));
10636 return;
10637
10638 case 's':
10639 /* Low 5 bits of 32 - value */
10640 if (! INT_P (x))
10641 output_operand_lossage ("invalid %%s value");
10642 else
10643 fprintf (file, HOST_WIDE_INT_PRINT_DEC, (32 - INT_LOWPART (x)) & 31);
10644 return;
10645
10646 case 'S':
10647 /* PowerPC64 mask position. All 0's is excluded.
10648 CONST_INT 32-bit mask is considered sign-extended so any
10649 transition must occur within the CONST_INT, not on the boundary. */
10650 if (! mask64_operand (x, DImode))
10651 output_operand_lossage ("invalid %%S value");
10652
10653 uval = INT_LOWPART (x);
10654
10655 if (uval & 1) /* Clear Left */
10656 {
10657 #if HOST_BITS_PER_WIDE_INT > 64
10658 uval &= ((unsigned HOST_WIDE_INT) 1 << 64) - 1;
10659 #endif
10660 i = 64;
10661 }
10662 else /* Clear Right */
10663 {
10664 uval = ~uval;
10665 #if HOST_BITS_PER_WIDE_INT > 64
10666 uval &= ((unsigned HOST_WIDE_INT) 1 << 64) - 1;
10667 #endif
10668 i = 63;
10669 }
10670 while (uval != 0)
10671 --i, uval >>= 1;
10672 gcc_assert (i >= 0);
10673 fprintf (file, "%d", i);
10674 return;
10675
10676 case 't':
10677 /* Like 'J' but get to the OVERFLOW/UNORDERED bit. */
10678 gcc_assert (GET_CODE (x) == REG && GET_MODE (x) == CCmode);
10679
10680 /* Bit 3 is OV bit. */
10681 i = 4 * (REGNO (x) - CR0_REGNO) + 3;
10682
10683 /* If we want bit 31, write a shift count of zero, not 32. */
10684 fprintf (file, "%d", i == 31 ? 0 : i + 1);
10685 return;
10686
10687 case 'T':
10688 /* Print the symbolic name of a branch target register. */
10689 if (GET_CODE (x) != REG || (REGNO (x) != LINK_REGISTER_REGNUM
10690 && REGNO (x) != COUNT_REGISTER_REGNUM))
10691 output_operand_lossage ("invalid %%T value");
10692 else if (REGNO (x) == LINK_REGISTER_REGNUM)
10693 fputs (TARGET_NEW_MNEMONICS ? "lr" : "r", file);
10694 else
10695 fputs ("ctr", file);
10696 return;
10697
10698 case 'u':
10699 /* High-order 16 bits of constant for use in unsigned operand. */
10700 if (! INT_P (x))
10701 output_operand_lossage ("invalid %%u value");
10702 else
10703 fprintf (file, HOST_WIDE_INT_PRINT_HEX,
10704 (INT_LOWPART (x) >> 16) & 0xffff);
10705 return;
10706
10707 case 'v':
10708 /* High-order 16 bits of constant for use in signed operand. */
10709 if (! INT_P (x))
10710 output_operand_lossage ("invalid %%v value");
10711 else
10712 fprintf (file, HOST_WIDE_INT_PRINT_HEX,
10713 (INT_LOWPART (x) >> 16) & 0xffff);
10714 return;
10715
10716 case 'U':
10717 /* Print `u' if this has an auto-increment or auto-decrement. */
10718 if (GET_CODE (x) == MEM
10719 && (GET_CODE (XEXP (x, 0)) == PRE_INC
10720 || GET_CODE (XEXP (x, 0)) == PRE_DEC))
10721 putc ('u', file);
10722 return;
10723
10724 case 'V':
10725 /* Print the trap code for this operand. */
10726 switch (GET_CODE (x))
10727 {
10728 case EQ:
10729 fputs ("eq", file); /* 4 */
10730 break;
10731 case NE:
10732 fputs ("ne", file); /* 24 */
10733 break;
10734 case LT:
10735 fputs ("lt", file); /* 16 */
10736 break;
10737 case LE:
10738 fputs ("le", file); /* 20 */
10739 break;
10740 case GT:
10741 fputs ("gt", file); /* 8 */
10742 break;
10743 case GE:
10744 fputs ("ge", file); /* 12 */
10745 break;
10746 case LTU:
10747 fputs ("llt", file); /* 2 */
10748 break;
10749 case LEU:
10750 fputs ("lle", file); /* 6 */
10751 break;
10752 case GTU:
10753 fputs ("lgt", file); /* 1 */
10754 break;
10755 case GEU:
10756 fputs ("lge", file); /* 5 */
10757 break;
10758 default:
10759 gcc_unreachable ();
10760 }
10761 break;
10762
10763 case 'w':
10764 /* If constant, low-order 16 bits of constant, signed. Otherwise, write
10765 normally. */
10766 if (INT_P (x))
10767 fprintf (file, HOST_WIDE_INT_PRINT_DEC,
10768 ((INT_LOWPART (x) & 0xffff) ^ 0x8000) - 0x8000);
10769 else
10770 print_operand (file, x, 0);
10771 return;
10772
10773 case 'W':
10774 /* MB value for a PowerPC64 rldic operand. */
10775 val = (GET_CODE (x) == CONST_INT
10776 ? INTVAL (x) : CONST_DOUBLE_HIGH (x));
10777
10778 if (val < 0)
10779 i = -1;
10780 else
10781 for (i = 0; i < HOST_BITS_PER_WIDE_INT; i++)
10782 if ((val <<= 1) < 0)
10783 break;
10784
10785 #if HOST_BITS_PER_WIDE_INT == 32
10786 if (GET_CODE (x) == CONST_INT && i >= 0)
10787 i += 32; /* zero-extend high-part was all 0's */
10788 else if (GET_CODE (x) == CONST_DOUBLE && i == 32)
10789 {
10790 val = CONST_DOUBLE_LOW (x);
10791
10792 gcc_assert (val);
10793 if (val < 0)
10794 --i;
10795 else
10796 for ( ; i < 64; i++)
10797 if ((val <<= 1) < 0)
10798 break;
10799 }
10800 #endif
10801
10802 fprintf (file, "%d", i + 1);
10803 return;
10804
10805 case 'X':
10806 if (GET_CODE (x) == MEM
10807 && legitimate_indexed_address_p (XEXP (x, 0), 0))
10808 putc ('x', file);
10809 return;
10810
10811 case 'Y':
10812 /* Like 'L', for third word of TImode */
10813 if (GET_CODE (x) == REG)
10814 fputs (reg_names[REGNO (x) + 2], file);
10815 else if (GET_CODE (x) == MEM)
10816 {
10817 if (GET_CODE (XEXP (x, 0)) == PRE_INC
10818 || GET_CODE (XEXP (x, 0)) == PRE_DEC)
10819 output_address (plus_constant (XEXP (XEXP (x, 0), 0), 8));
10820 else
10821 output_address (XEXP (adjust_address_nv (x, SImode, 8), 0));
10822 if (small_data_operand (x, GET_MODE (x)))
10823 fprintf (file, "@%s(%s)", SMALL_DATA_RELOC,
10824 reg_names[SMALL_DATA_REG]);
10825 }
10826 return;
10827
10828 case 'z':
10829 /* X is a SYMBOL_REF. Write out the name preceded by a
10830 period and without any trailing data in brackets. Used for function
10831 names. If we are configured for System V (or the embedded ABI) on
10832 the PowerPC, do not emit the period, since those systems do not use
10833 TOCs and the like. */
10834 gcc_assert (GET_CODE (x) == SYMBOL_REF);
10835
10836 /* Mark the decl as referenced so that cgraph will output the
10837 function. */
10838 if (SYMBOL_REF_DECL (x))
10839 mark_decl_referenced (SYMBOL_REF_DECL (x));
10840
10841 /* For macho, check to see if we need a stub. */
10842 if (TARGET_MACHO)
10843 {
10844 const char *name = XSTR (x, 0);
10845 #if TARGET_MACHO
10846 if (MACHOPIC_INDIRECT
10847 && machopic_classify_symbol (x) == MACHOPIC_UNDEFINED_FUNCTION)
10848 name = machopic_indirection_name (x, /*stub_p=*/true);
10849 #endif
10850 assemble_name (file, name);
10851 }
10852 else if (!DOT_SYMBOLS)
10853 assemble_name (file, XSTR (x, 0));
10854 else
10855 rs6000_output_function_entry (file, XSTR (x, 0));
10856 return;
10857
10858 case 'Z':
10859 /* Like 'L', for last word of TImode. */
10860 if (GET_CODE (x) == REG)
10861 fputs (reg_names[REGNO (x) + 3], file);
10862 else if (GET_CODE (x) == MEM)
10863 {
10864 if (GET_CODE (XEXP (x, 0)) == PRE_INC
10865 || GET_CODE (XEXP (x, 0)) == PRE_DEC)
10866 output_address (plus_constant (XEXP (XEXP (x, 0), 0), 12));
10867 else
10868 output_address (XEXP (adjust_address_nv (x, SImode, 12), 0));
10869 if (small_data_operand (x, GET_MODE (x)))
10870 fprintf (file, "@%s(%s)", SMALL_DATA_RELOC,
10871 reg_names[SMALL_DATA_REG]);
10872 }
10873 return;
10874
10875 /* Print AltiVec or SPE memory operand. */
10876 case 'y':
10877 {
10878 rtx tmp;
10879
10880 gcc_assert (GET_CODE (x) == MEM);
10881
10882 tmp = XEXP (x, 0);
10883
10884 /* Ugly hack because %y is overloaded. */
10885 if (TARGET_E500 && GET_MODE_SIZE (GET_MODE (x)) == 8)
10886 {
10887 /* Handle [reg]. */
10888 if (GET_CODE (tmp) == REG)
10889 {
10890 fprintf (file, "0(%s)", reg_names[REGNO (tmp)]);
10891 break;
10892 }
10893 /* Handle [reg+UIMM]. */
10894 else if (GET_CODE (tmp) == PLUS &&
10895 GET_CODE (XEXP (tmp, 1)) == CONST_INT)
10896 {
10897 int x;
10898
10899 gcc_assert (GET_CODE (XEXP (tmp, 0)) == REG);
10900
10901 x = INTVAL (XEXP (tmp, 1));
10902 fprintf (file, "%d(%s)", x, reg_names[REGNO (XEXP (tmp, 0))]);
10903 break;
10904 }
10905
10906 /* Fall through. Must be [reg+reg]. */
10907 }
10908 if (TARGET_ALTIVEC
10909 && GET_CODE (tmp) == AND
10910 && GET_CODE (XEXP (tmp, 1)) == CONST_INT
10911 && INTVAL (XEXP (tmp, 1)) == -16)
10912 tmp = XEXP (tmp, 0);
10913 if (GET_CODE (tmp) == REG)
10914 fprintf (file, "0,%s", reg_names[REGNO (tmp)]);
10915 else
10916 {
10917 gcc_assert (GET_CODE (tmp) == PLUS
10918 && REG_P (XEXP (tmp, 0))
10919 && REG_P (XEXP (tmp, 1)));
10920
10921 if (REGNO (XEXP (tmp, 0)) == 0)
10922 fprintf (file, "%s,%s", reg_names[ REGNO (XEXP (tmp, 1)) ],
10923 reg_names[ REGNO (XEXP (tmp, 0)) ]);
10924 else
10925 fprintf (file, "%s,%s", reg_names[ REGNO (XEXP (tmp, 0)) ],
10926 reg_names[ REGNO (XEXP (tmp, 1)) ]);
10927 }
10928 break;
10929 }
10930
10931 case 0:
10932 if (GET_CODE (x) == REG)
10933 fprintf (file, "%s", reg_names[REGNO (x)]);
10934 else if (GET_CODE (x) == MEM)
10935 {
10936 /* We need to handle PRE_INC and PRE_DEC here, since we need to
10937 know the width from the mode. */
10938 if (GET_CODE (XEXP (x, 0)) == PRE_INC)
10939 fprintf (file, "%d(%s)", GET_MODE_SIZE (GET_MODE (x)),
10940 reg_names[REGNO (XEXP (XEXP (x, 0), 0))]);
10941 else if (GET_CODE (XEXP (x, 0)) == PRE_DEC)
10942 fprintf (file, "%d(%s)", - GET_MODE_SIZE (GET_MODE (x)),
10943 reg_names[REGNO (XEXP (XEXP (x, 0), 0))]);
10944 else
10945 output_address (XEXP (x, 0));
10946 }
10947 else
10948 output_addr_const (file, x);
10949 return;
10950
10951 case '&':
10952 assemble_name (file, rs6000_get_some_local_dynamic_name ());
10953 return;
10954
10955 default:
10956 output_operand_lossage ("invalid %%xn code");
10957 }
10958 }
10959
10960 /* Print the address of an operand. */
10961
10962 void
print_operand_address(FILE * file,rtx x)10963 print_operand_address (FILE *file, rtx x)
10964 {
10965 if (GET_CODE (x) == REG)
10966 fprintf (file, "0(%s)", reg_names[ REGNO (x) ]);
10967 else if (GET_CODE (x) == SYMBOL_REF || GET_CODE (x) == CONST
10968 || GET_CODE (x) == LABEL_REF)
10969 {
10970 output_addr_const (file, x);
10971 if (small_data_operand (x, GET_MODE (x)))
10972 fprintf (file, "@%s(%s)", SMALL_DATA_RELOC,
10973 reg_names[SMALL_DATA_REG]);
10974 else
10975 gcc_assert (!TARGET_TOC);
10976 }
10977 else if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 1)) == REG)
10978 {
10979 gcc_assert (REG_P (XEXP (x, 0)));
10980 if (REGNO (XEXP (x, 0)) == 0)
10981 fprintf (file, "%s,%s", reg_names[ REGNO (XEXP (x, 1)) ],
10982 reg_names[ REGNO (XEXP (x, 0)) ]);
10983 else
10984 fprintf (file, "%s,%s", reg_names[ REGNO (XEXP (x, 0)) ],
10985 reg_names[ REGNO (XEXP (x, 1)) ]);
10986 }
10987 else if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 1)) == CONST_INT)
10988 fprintf (file, HOST_WIDE_INT_PRINT_DEC "(%s)",
10989 INTVAL (XEXP (x, 1)), reg_names[ REGNO (XEXP (x, 0)) ]);
10990 #if TARGET_ELF
10991 else if (GET_CODE (x) == LO_SUM && GET_CODE (XEXP (x, 0)) == REG
10992 && CONSTANT_P (XEXP (x, 1)))
10993 {
10994 output_addr_const (file, XEXP (x, 1));
10995 fprintf (file, "@l(%s)", reg_names[ REGNO (XEXP (x, 0)) ]);
10996 }
10997 #endif
10998 #if TARGET_MACHO
10999 else if (GET_CODE (x) == LO_SUM && GET_CODE (XEXP (x, 0)) == REG
11000 && CONSTANT_P (XEXP (x, 1)))
11001 {
11002 fprintf (file, "lo16(");
11003 output_addr_const (file, XEXP (x, 1));
11004 fprintf (file, ")(%s)", reg_names[ REGNO (XEXP (x, 0)) ]);
11005 }
11006 #endif
11007 else if (legitimate_constant_pool_address_p (x))
11008 {
11009 if (TARGET_AIX && (!TARGET_ELF || !TARGET_MINIMAL_TOC))
11010 {
11011 rtx contains_minus = XEXP (x, 1);
11012 rtx minus, symref;
11013 const char *name;
11014
11015 /* Find the (minus (sym) (toc)) buried in X, and temporarily
11016 turn it into (sym) for output_addr_const. */
11017 while (GET_CODE (XEXP (contains_minus, 0)) != MINUS)
11018 contains_minus = XEXP (contains_minus, 0);
11019
11020 minus = XEXP (contains_minus, 0);
11021 symref = XEXP (minus, 0);
11022 XEXP (contains_minus, 0) = symref;
11023 if (TARGET_ELF)
11024 {
11025 char *newname;
11026
11027 name = XSTR (symref, 0);
11028 newname = alloca (strlen (name) + sizeof ("@toc"));
11029 strcpy (newname, name);
11030 strcat (newname, "@toc");
11031 XSTR (symref, 0) = newname;
11032 }
11033 output_addr_const (file, XEXP (x, 1));
11034 if (TARGET_ELF)
11035 XSTR (symref, 0) = name;
11036 XEXP (contains_minus, 0) = minus;
11037 }
11038 else
11039 output_addr_const (file, XEXP (x, 1));
11040
11041 fprintf (file, "(%s)", reg_names[REGNO (XEXP (x, 0))]);
11042 }
11043 else
11044 gcc_unreachable ();
11045 }
11046
11047 /* Target hook for assembling integer objects. The PowerPC version has
11048 to handle fixup entries for relocatable code if RELOCATABLE_NEEDS_FIXUP
11049 is defined. It also needs to handle DI-mode objects on 64-bit
11050 targets. */
11051
11052 static bool
rs6000_assemble_integer(rtx x,unsigned int size,int aligned_p)11053 rs6000_assemble_integer (rtx x, unsigned int size, int aligned_p)
11054 {
11055 #ifdef RELOCATABLE_NEEDS_FIXUP
11056 /* Special handling for SI values. */
11057 if (RELOCATABLE_NEEDS_FIXUP && size == 4 && aligned_p)
11058 {
11059 static int recurse = 0;
11060
11061 /* For -mrelocatable, we mark all addresses that need to be fixed up
11062 in the .fixup section. */
11063 if (TARGET_RELOCATABLE
11064 && in_section != toc_section
11065 && in_section != text_section
11066 && !unlikely_text_section_p (in_section)
11067 && !recurse
11068 && GET_CODE (x) != CONST_INT
11069 && GET_CODE (x) != CONST_DOUBLE
11070 && CONSTANT_P (x))
11071 {
11072 char buf[256];
11073
11074 recurse = 1;
11075 ASM_GENERATE_INTERNAL_LABEL (buf, "LCP", fixuplabelno);
11076 fixuplabelno++;
11077 ASM_OUTPUT_LABEL (asm_out_file, buf);
11078 fprintf (asm_out_file, "\t.long\t(");
11079 output_addr_const (asm_out_file, x);
11080 fprintf (asm_out_file, ")@fixup\n");
11081 fprintf (asm_out_file, "\t.section\t\".fixup\",\"aw\"\n");
11082 ASM_OUTPUT_ALIGN (asm_out_file, 2);
11083 fprintf (asm_out_file, "\t.long\t");
11084 assemble_name (asm_out_file, buf);
11085 fprintf (asm_out_file, "\n\t.previous\n");
11086 recurse = 0;
11087 return true;
11088 }
11089 /* Remove initial .'s to turn a -mcall-aixdesc function
11090 address into the address of the descriptor, not the function
11091 itself. */
11092 else if (GET_CODE (x) == SYMBOL_REF
11093 && XSTR (x, 0)[0] == '.'
11094 && DEFAULT_ABI == ABI_AIX)
11095 {
11096 const char *name = XSTR (x, 0);
11097 while (*name == '.')
11098 name++;
11099
11100 fprintf (asm_out_file, "\t.long\t%s\n", name);
11101 return true;
11102 }
11103 }
11104 #endif /* RELOCATABLE_NEEDS_FIXUP */
11105 return default_assemble_integer (x, size, aligned_p);
11106 }
11107
11108 #ifdef HAVE_GAS_HIDDEN
11109 /* Emit an assembler directive to set symbol visibility for DECL to
11110 VISIBILITY_TYPE. */
11111
11112 static void
rs6000_assemble_visibility(tree decl,int vis)11113 rs6000_assemble_visibility (tree decl, int vis)
11114 {
11115 /* Functions need to have their entry point symbol visibility set as
11116 well as their descriptor symbol visibility. */
11117 if (DEFAULT_ABI == ABI_AIX
11118 && DOT_SYMBOLS
11119 && TREE_CODE (decl) == FUNCTION_DECL)
11120 {
11121 static const char * const visibility_types[] = {
11122 NULL, "internal", "hidden", "protected"
11123 };
11124
11125 const char *name, *type;
11126
11127 name = ((* targetm.strip_name_encoding)
11128 (IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (decl))));
11129 type = visibility_types[vis];
11130
11131 fprintf (asm_out_file, "\t.%s\t%s\n", type, name);
11132 fprintf (asm_out_file, "\t.%s\t.%s\n", type, name);
11133 }
11134 else
11135 default_assemble_visibility (decl, vis);
11136 }
11137 #endif
11138
11139 enum rtx_code
rs6000_reverse_condition(enum machine_mode mode,enum rtx_code code)11140 rs6000_reverse_condition (enum machine_mode mode, enum rtx_code code)
11141 {
11142 /* Reversal of FP compares takes care -- an ordered compare
11143 becomes an unordered compare and vice versa. */
11144 if (mode == CCFPmode
11145 && (!flag_finite_math_only
11146 || code == UNLT || code == UNLE || code == UNGT || code == UNGE
11147 || code == UNEQ || code == LTGT))
11148 return reverse_condition_maybe_unordered (code);
11149 else
11150 return reverse_condition (code);
11151 }
11152
11153 /* Generate a compare for CODE. Return a brand-new rtx that
11154 represents the result of the compare. */
11155
11156 static rtx
rs6000_generate_compare(enum rtx_code code)11157 rs6000_generate_compare (enum rtx_code code)
11158 {
11159 enum machine_mode comp_mode;
11160 rtx compare_result;
11161
11162 if (rs6000_compare_fp_p)
11163 comp_mode = CCFPmode;
11164 else if (code == GTU || code == LTU
11165 || code == GEU || code == LEU)
11166 comp_mode = CCUNSmode;
11167 else if ((code == EQ || code == NE)
11168 && GET_CODE (rs6000_compare_op0) == SUBREG
11169 && GET_CODE (rs6000_compare_op1) == SUBREG
11170 && SUBREG_PROMOTED_UNSIGNED_P (rs6000_compare_op0)
11171 && SUBREG_PROMOTED_UNSIGNED_P (rs6000_compare_op1))
11172 /* These are unsigned values, perhaps there will be a later
11173 ordering compare that can be shared with this one.
11174 Unfortunately we cannot detect the signedness of the operands
11175 for non-subregs. */
11176 comp_mode = CCUNSmode;
11177 else
11178 comp_mode = CCmode;
11179
11180 /* First, the compare. */
11181 compare_result = gen_reg_rtx (comp_mode);
11182
11183 /* E500 FP compare instructions on the GPRs. Yuck! */
11184 if ((TARGET_E500 && !TARGET_FPRS && TARGET_HARD_FLOAT)
11185 && rs6000_compare_fp_p)
11186 {
11187 rtx cmp, or_result, compare_result2;
11188 enum machine_mode op_mode = GET_MODE (rs6000_compare_op0);
11189
11190 if (op_mode == VOIDmode)
11191 op_mode = GET_MODE (rs6000_compare_op1);
11192
11193 /* The E500 FP compare instructions toggle the GT bit (CR bit 1) only.
11194 This explains the following mess. */
11195
11196 switch (code)
11197 {
11198 case EQ: case UNEQ: case NE: case LTGT:
11199 switch (op_mode)
11200 {
11201 case SFmode:
11202 cmp = flag_unsafe_math_optimizations
11203 ? gen_tstsfeq_gpr (compare_result, rs6000_compare_op0,
11204 rs6000_compare_op1)
11205 : gen_cmpsfeq_gpr (compare_result, rs6000_compare_op0,
11206 rs6000_compare_op1);
11207 break;
11208
11209 case DFmode:
11210 cmp = flag_unsafe_math_optimizations
11211 ? gen_tstdfeq_gpr (compare_result, rs6000_compare_op0,
11212 rs6000_compare_op1)
11213 : gen_cmpdfeq_gpr (compare_result, rs6000_compare_op0,
11214 rs6000_compare_op1);
11215 break;
11216
11217 default:
11218 gcc_unreachable ();
11219 }
11220 break;
11221
11222 case GT: case GTU: case UNGT: case UNGE: case GE: case GEU:
11223 switch (op_mode)
11224 {
11225 case SFmode:
11226 cmp = flag_unsafe_math_optimizations
11227 ? gen_tstsfgt_gpr (compare_result, rs6000_compare_op0,
11228 rs6000_compare_op1)
11229 : gen_cmpsfgt_gpr (compare_result, rs6000_compare_op0,
11230 rs6000_compare_op1);
11231 break;
11232
11233 case DFmode:
11234 cmp = flag_unsafe_math_optimizations
11235 ? gen_tstdfgt_gpr (compare_result, rs6000_compare_op0,
11236 rs6000_compare_op1)
11237 : gen_cmpdfgt_gpr (compare_result, rs6000_compare_op0,
11238 rs6000_compare_op1);
11239 break;
11240
11241 default:
11242 gcc_unreachable ();
11243 }
11244 break;
11245
11246 case LT: case LTU: case UNLT: case UNLE: case LE: case LEU:
11247 switch (op_mode)
11248 {
11249 case SFmode:
11250 cmp = flag_unsafe_math_optimizations
11251 ? gen_tstsflt_gpr (compare_result, rs6000_compare_op0,
11252 rs6000_compare_op1)
11253 : gen_cmpsflt_gpr (compare_result, rs6000_compare_op0,
11254 rs6000_compare_op1);
11255 break;
11256
11257 case DFmode:
11258 cmp = flag_unsafe_math_optimizations
11259 ? gen_tstdflt_gpr (compare_result, rs6000_compare_op0,
11260 rs6000_compare_op1)
11261 : gen_cmpdflt_gpr (compare_result, rs6000_compare_op0,
11262 rs6000_compare_op1);
11263 break;
11264
11265 default:
11266 gcc_unreachable ();
11267 }
11268 break;
11269 default:
11270 gcc_unreachable ();
11271 }
11272
11273 /* Synthesize LE and GE from LT/GT || EQ. */
11274 if (code == LE || code == GE || code == LEU || code == GEU)
11275 {
11276 emit_insn (cmp);
11277
11278 switch (code)
11279 {
11280 case LE: code = LT; break;
11281 case GE: code = GT; break;
11282 case LEU: code = LT; break;
11283 case GEU: code = GT; break;
11284 default: gcc_unreachable ();
11285 }
11286
11287 compare_result2 = gen_reg_rtx (CCFPmode);
11288
11289 /* Do the EQ. */
11290 switch (op_mode)
11291 {
11292 case SFmode:
11293 cmp = flag_unsafe_math_optimizations
11294 ? gen_tstsfeq_gpr (compare_result2, rs6000_compare_op0,
11295 rs6000_compare_op1)
11296 : gen_cmpsfeq_gpr (compare_result2, rs6000_compare_op0,
11297 rs6000_compare_op1);
11298 break;
11299
11300 case DFmode:
11301 cmp = flag_unsafe_math_optimizations
11302 ? gen_tstdfeq_gpr (compare_result2, rs6000_compare_op0,
11303 rs6000_compare_op1)
11304 : gen_cmpdfeq_gpr (compare_result2, rs6000_compare_op0,
11305 rs6000_compare_op1);
11306 break;
11307
11308 default:
11309 gcc_unreachable ();
11310 }
11311 emit_insn (cmp);
11312
11313 /* OR them together. */
11314 or_result = gen_reg_rtx (CCFPmode);
11315 cmp = gen_e500_cr_ior_compare (or_result, compare_result,
11316 compare_result2);
11317 compare_result = or_result;
11318 code = EQ;
11319 }
11320 else
11321 {
11322 if (code == NE || code == LTGT)
11323 code = NE;
11324 else
11325 code = EQ;
11326 }
11327
11328 emit_insn (cmp);
11329 }
11330 else
11331 {
11332 /* Generate XLC-compatible TFmode compare as PARALLEL with extra
11333 CLOBBERs to match cmptf_internal2 pattern. */
11334 if (comp_mode == CCFPmode && TARGET_XL_COMPAT
11335 && GET_MODE (rs6000_compare_op0) == TFmode
11336 && !TARGET_IEEEQUAD
11337 && TARGET_HARD_FLOAT && TARGET_FPRS && TARGET_LONG_DOUBLE_128)
11338 emit_insn (gen_rtx_PARALLEL (VOIDmode,
11339 gen_rtvec (9,
11340 gen_rtx_SET (VOIDmode,
11341 compare_result,
11342 gen_rtx_COMPARE (comp_mode,
11343 rs6000_compare_op0,
11344 rs6000_compare_op1)),
11345 gen_rtx_CLOBBER (VOIDmode, gen_rtx_SCRATCH (DFmode)),
11346 gen_rtx_CLOBBER (VOIDmode, gen_rtx_SCRATCH (DFmode)),
11347 gen_rtx_CLOBBER (VOIDmode, gen_rtx_SCRATCH (DFmode)),
11348 gen_rtx_CLOBBER (VOIDmode, gen_rtx_SCRATCH (DFmode)),
11349 gen_rtx_CLOBBER (VOIDmode, gen_rtx_SCRATCH (DFmode)),
11350 gen_rtx_CLOBBER (VOIDmode, gen_rtx_SCRATCH (DFmode)),
11351 gen_rtx_CLOBBER (VOIDmode, gen_rtx_SCRATCH (DFmode)),
11352 gen_rtx_CLOBBER (VOIDmode, gen_rtx_SCRATCH (DFmode)))));
11353 else if (GET_CODE (rs6000_compare_op1) == UNSPEC
11354 && XINT (rs6000_compare_op1, 1) == UNSPEC_SP_TEST)
11355 {
11356 rtx op1 = XVECEXP (rs6000_compare_op1, 0, 0);
11357 comp_mode = CCEQmode;
11358 compare_result = gen_reg_rtx (CCEQmode);
11359 if (TARGET_64BIT)
11360 emit_insn (gen_stack_protect_testdi (compare_result,
11361 rs6000_compare_op0, op1));
11362 else
11363 emit_insn (gen_stack_protect_testsi (compare_result,
11364 rs6000_compare_op0, op1));
11365 }
11366 else
11367 emit_insn (gen_rtx_SET (VOIDmode, compare_result,
11368 gen_rtx_COMPARE (comp_mode,
11369 rs6000_compare_op0,
11370 rs6000_compare_op1)));
11371 }
11372
11373 /* Some kinds of FP comparisons need an OR operation;
11374 under flag_finite_math_only we don't bother. */
11375 if (rs6000_compare_fp_p
11376 && !flag_finite_math_only
11377 && !(TARGET_HARD_FLOAT && TARGET_E500 && !TARGET_FPRS)
11378 && (code == LE || code == GE
11379 || code == UNEQ || code == LTGT
11380 || code == UNGT || code == UNLT))
11381 {
11382 enum rtx_code or1, or2;
11383 rtx or1_rtx, or2_rtx, compare2_rtx;
11384 rtx or_result = gen_reg_rtx (CCEQmode);
11385
11386 switch (code)
11387 {
11388 case LE: or1 = LT; or2 = EQ; break;
11389 case GE: or1 = GT; or2 = EQ; break;
11390 case UNEQ: or1 = UNORDERED; or2 = EQ; break;
11391 case LTGT: or1 = LT; or2 = GT; break;
11392 case UNGT: or1 = UNORDERED; or2 = GT; break;
11393 case UNLT: or1 = UNORDERED; or2 = LT; break;
11394 default: gcc_unreachable ();
11395 }
11396 validate_condition_mode (or1, comp_mode);
11397 validate_condition_mode (or2, comp_mode);
11398 or1_rtx = gen_rtx_fmt_ee (or1, SImode, compare_result, const0_rtx);
11399 or2_rtx = gen_rtx_fmt_ee (or2, SImode, compare_result, const0_rtx);
11400 compare2_rtx = gen_rtx_COMPARE (CCEQmode,
11401 gen_rtx_IOR (SImode, or1_rtx, or2_rtx),
11402 const_true_rtx);
11403 emit_insn (gen_rtx_SET (VOIDmode, or_result, compare2_rtx));
11404
11405 compare_result = or_result;
11406 code = EQ;
11407 }
11408
11409 validate_condition_mode (code, GET_MODE (compare_result));
11410
11411 return gen_rtx_fmt_ee (code, VOIDmode, compare_result, const0_rtx);
11412 }
11413
11414
11415 /* Emit the RTL for an sCOND pattern. */
11416
11417 void
rs6000_emit_sCOND(enum rtx_code code,rtx result)11418 rs6000_emit_sCOND (enum rtx_code code, rtx result)
11419 {
11420 rtx condition_rtx;
11421 enum machine_mode op_mode;
11422 enum rtx_code cond_code;
11423
11424 condition_rtx = rs6000_generate_compare (code);
11425 cond_code = GET_CODE (condition_rtx);
11426
11427 if (TARGET_E500 && rs6000_compare_fp_p
11428 && !TARGET_FPRS && TARGET_HARD_FLOAT)
11429 {
11430 rtx t;
11431
11432 PUT_MODE (condition_rtx, SImode);
11433 t = XEXP (condition_rtx, 0);
11434
11435 gcc_assert (cond_code == NE || cond_code == EQ);
11436
11437 if (cond_code == NE)
11438 emit_insn (gen_e500_flip_gt_bit (t, t));
11439
11440 emit_insn (gen_move_from_CR_gt_bit (result, t));
11441 return;
11442 }
11443
11444 if (cond_code == NE
11445 || cond_code == GE || cond_code == LE
11446 || cond_code == GEU || cond_code == LEU
11447 || cond_code == ORDERED || cond_code == UNGE || cond_code == UNLE)
11448 {
11449 rtx not_result = gen_reg_rtx (CCEQmode);
11450 rtx not_op, rev_cond_rtx;
11451 enum machine_mode cc_mode;
11452
11453 cc_mode = GET_MODE (XEXP (condition_rtx, 0));
11454
11455 rev_cond_rtx = gen_rtx_fmt_ee (rs6000_reverse_condition (cc_mode, cond_code),
11456 SImode, XEXP (condition_rtx, 0), const0_rtx);
11457 not_op = gen_rtx_COMPARE (CCEQmode, rev_cond_rtx, const0_rtx);
11458 emit_insn (gen_rtx_SET (VOIDmode, not_result, not_op));
11459 condition_rtx = gen_rtx_EQ (VOIDmode, not_result, const0_rtx);
11460 }
11461
11462 op_mode = GET_MODE (rs6000_compare_op0);
11463 if (op_mode == VOIDmode)
11464 op_mode = GET_MODE (rs6000_compare_op1);
11465
11466 if (TARGET_POWERPC64 && (op_mode == DImode || rs6000_compare_fp_p))
11467 {
11468 PUT_MODE (condition_rtx, DImode);
11469 convert_move (result, condition_rtx, 0);
11470 }
11471 else
11472 {
11473 PUT_MODE (condition_rtx, SImode);
11474 emit_insn (gen_rtx_SET (VOIDmode, result, condition_rtx));
11475 }
11476 }
11477
11478 /* Emit a branch of kind CODE to location LOC. */
11479
11480 void
rs6000_emit_cbranch(enum rtx_code code,rtx loc)11481 rs6000_emit_cbranch (enum rtx_code code, rtx loc)
11482 {
11483 rtx condition_rtx, loc_ref;
11484
11485 condition_rtx = rs6000_generate_compare (code);
11486 loc_ref = gen_rtx_LABEL_REF (VOIDmode, loc);
11487 emit_jump_insn (gen_rtx_SET (VOIDmode, pc_rtx,
11488 gen_rtx_IF_THEN_ELSE (VOIDmode, condition_rtx,
11489 loc_ref, pc_rtx)));
11490 }
11491
11492 /* Return the string to output a conditional branch to LABEL, which is
11493 the operand number of the label, or -1 if the branch is really a
11494 conditional return.
11495
11496 OP is the conditional expression. XEXP (OP, 0) is assumed to be a
11497 condition code register and its mode specifies what kind of
11498 comparison we made.
11499
11500 REVERSED is nonzero if we should reverse the sense of the comparison.
11501
11502 INSN is the insn. */
11503
11504 char *
output_cbranch(rtx op,const char * label,int reversed,rtx insn)11505 output_cbranch (rtx op, const char *label, int reversed, rtx insn)
11506 {
11507 static char string[64];
11508 enum rtx_code code = GET_CODE (op);
11509 rtx cc_reg = XEXP (op, 0);
11510 enum machine_mode mode = GET_MODE (cc_reg);
11511 int cc_regno = REGNO (cc_reg) - CR0_REGNO;
11512 int need_longbranch = label != NULL && get_attr_length (insn) == 8;
11513 int really_reversed = reversed ^ need_longbranch;
11514 char *s = string;
11515 const char *ccode;
11516 const char *pred;
11517 rtx note;
11518
11519 validate_condition_mode (code, mode);
11520
11521 /* Work out which way this really branches. We could use
11522 reverse_condition_maybe_unordered here always but this
11523 makes the resulting assembler clearer. */
11524 if (really_reversed)
11525 {
11526 /* Reversal of FP compares takes care -- an ordered compare
11527 becomes an unordered compare and vice versa. */
11528 if (mode == CCFPmode)
11529 code = reverse_condition_maybe_unordered (code);
11530 else
11531 code = reverse_condition (code);
11532 }
11533
11534 if ((TARGET_E500 && !TARGET_FPRS && TARGET_HARD_FLOAT) && mode == CCFPmode)
11535 {
11536 /* The efscmp/tst* instructions twiddle bit 2, which maps nicely
11537 to the GT bit. */
11538 switch (code)
11539 {
11540 case EQ:
11541 /* Opposite of GT. */
11542 code = GT;
11543 break;
11544
11545 case NE:
11546 code = UNLE;
11547 break;
11548
11549 default:
11550 gcc_unreachable ();
11551 }
11552 }
11553
11554 switch (code)
11555 {
11556 /* Not all of these are actually distinct opcodes, but
11557 we distinguish them for clarity of the resulting assembler. */
11558 case NE: case LTGT:
11559 ccode = "ne"; break;
11560 case EQ: case UNEQ:
11561 ccode = "eq"; break;
11562 case GE: case GEU:
11563 ccode = "ge"; break;
11564 case GT: case GTU: case UNGT:
11565 ccode = "gt"; break;
11566 case LE: case LEU:
11567 ccode = "le"; break;
11568 case LT: case LTU: case UNLT:
11569 ccode = "lt"; break;
11570 case UNORDERED: ccode = "un"; break;
11571 case ORDERED: ccode = "nu"; break;
11572 case UNGE: ccode = "nl"; break;
11573 case UNLE: ccode = "ng"; break;
11574 default:
11575 gcc_unreachable ();
11576 }
11577
11578 /* Maybe we have a guess as to how likely the branch is.
11579 The old mnemonics don't have a way to specify this information. */
11580 pred = "";
11581 note = find_reg_note (insn, REG_BR_PROB, NULL_RTX);
11582 if (note != NULL_RTX)
11583 {
11584 /* PROB is the difference from 50%. */
11585 int prob = INTVAL (XEXP (note, 0)) - REG_BR_PROB_BASE / 2;
11586
11587 /* Only hint for highly probable/improbable branches on newer
11588 cpus as static prediction overrides processor dynamic
11589 prediction. For older cpus we may as well always hint, but
11590 assume not taken for branches that are very close to 50% as a
11591 mispredicted taken branch is more expensive than a
11592 mispredicted not-taken branch. */
11593 if (rs6000_always_hint
11594 || (abs (prob) > REG_BR_PROB_BASE / 100 * 48
11595 && br_prob_note_reliable_p (note)))
11596 {
11597 if (abs (prob) > REG_BR_PROB_BASE / 20
11598 && ((prob > 0) ^ need_longbranch))
11599 pred = "+";
11600 else
11601 pred = "-";
11602 }
11603 }
11604
11605 if (label == NULL)
11606 s += sprintf (s, "{b%sr|b%slr%s} ", ccode, ccode, pred);
11607 else
11608 s += sprintf (s, "{b%s|b%s%s} ", ccode, ccode, pred);
11609
11610 /* We need to escape any '%' characters in the reg_names string.
11611 Assume they'd only be the first character.... */
11612 if (reg_names[cc_regno + CR0_REGNO][0] == '%')
11613 *s++ = '%';
11614 s += sprintf (s, "%s", reg_names[cc_regno + CR0_REGNO]);
11615
11616 if (label != NULL)
11617 {
11618 /* If the branch distance was too far, we may have to use an
11619 unconditional branch to go the distance. */
11620 if (need_longbranch)
11621 s += sprintf (s, ",$+8\n\tb %s", label);
11622 else
11623 s += sprintf (s, ",%s", label);
11624 }
11625
11626 return string;
11627 }
11628
11629 /* Return the string to flip the GT bit on a CR. */
11630 char *
output_e500_flip_gt_bit(rtx dst,rtx src)11631 output_e500_flip_gt_bit (rtx dst, rtx src)
11632 {
11633 static char string[64];
11634 int a, b;
11635
11636 gcc_assert (GET_CODE (dst) == REG && CR_REGNO_P (REGNO (dst))
11637 && GET_CODE (src) == REG && CR_REGNO_P (REGNO (src)));
11638
11639 /* GT bit. */
11640 a = 4 * (REGNO (dst) - CR0_REGNO) + 1;
11641 b = 4 * (REGNO (src) - CR0_REGNO) + 1;
11642
11643 sprintf (string, "crnot %d,%d", a, b);
11644 return string;
11645 }
11646
11647 /* Return insn index for the vector compare instruction for given CODE,
11648 and DEST_MODE, OP_MODE. Return INSN_NOT_AVAILABLE if valid insn is
11649 not available. */
11650
11651 static int
get_vec_cmp_insn(enum rtx_code code,enum machine_mode dest_mode,enum machine_mode op_mode)11652 get_vec_cmp_insn (enum rtx_code code,
11653 enum machine_mode dest_mode,
11654 enum machine_mode op_mode)
11655 {
11656 if (!TARGET_ALTIVEC)
11657 return INSN_NOT_AVAILABLE;
11658
11659 switch (code)
11660 {
11661 case EQ:
11662 if (dest_mode == V16QImode && op_mode == V16QImode)
11663 return UNSPEC_VCMPEQUB;
11664 if (dest_mode == V8HImode && op_mode == V8HImode)
11665 return UNSPEC_VCMPEQUH;
11666 if (dest_mode == V4SImode && op_mode == V4SImode)
11667 return UNSPEC_VCMPEQUW;
11668 if (dest_mode == V4SImode && op_mode == V4SFmode)
11669 return UNSPEC_VCMPEQFP;
11670 break;
11671 case GE:
11672 if (dest_mode == V4SImode && op_mode == V4SFmode)
11673 return UNSPEC_VCMPGEFP;
11674 case GT:
11675 if (dest_mode == V16QImode && op_mode == V16QImode)
11676 return UNSPEC_VCMPGTSB;
11677 if (dest_mode == V8HImode && op_mode == V8HImode)
11678 return UNSPEC_VCMPGTSH;
11679 if (dest_mode == V4SImode && op_mode == V4SImode)
11680 return UNSPEC_VCMPGTSW;
11681 if (dest_mode == V4SImode && op_mode == V4SFmode)
11682 return UNSPEC_VCMPGTFP;
11683 break;
11684 case GTU:
11685 if (dest_mode == V16QImode && op_mode == V16QImode)
11686 return UNSPEC_VCMPGTUB;
11687 if (dest_mode == V8HImode && op_mode == V8HImode)
11688 return UNSPEC_VCMPGTUH;
11689 if (dest_mode == V4SImode && op_mode == V4SImode)
11690 return UNSPEC_VCMPGTUW;
11691 break;
11692 default:
11693 break;
11694 }
11695 return INSN_NOT_AVAILABLE;
11696 }
11697
11698 /* Emit vector compare for operands OP0 and OP1 using code RCODE.
11699 DMODE is expected destination mode. This is a recursive function. */
11700
11701 static rtx
rs6000_emit_vector_compare(enum rtx_code rcode,rtx op0,rtx op1,enum machine_mode dmode)11702 rs6000_emit_vector_compare (enum rtx_code rcode,
11703 rtx op0, rtx op1,
11704 enum machine_mode dmode)
11705 {
11706 int vec_cmp_insn;
11707 rtx mask;
11708 enum machine_mode dest_mode;
11709 enum machine_mode op_mode = GET_MODE (op1);
11710
11711 gcc_assert (TARGET_ALTIVEC);
11712 gcc_assert (GET_MODE (op0) == GET_MODE (op1));
11713
11714 /* Floating point vector compare instructions uses destination V4SImode.
11715 Move destination to appropriate mode later. */
11716 if (dmode == V4SFmode)
11717 dest_mode = V4SImode;
11718 else
11719 dest_mode = dmode;
11720
11721 mask = gen_reg_rtx (dest_mode);
11722 vec_cmp_insn = get_vec_cmp_insn (rcode, dest_mode, op_mode);
11723
11724 if (vec_cmp_insn == INSN_NOT_AVAILABLE)
11725 {
11726 bool swap_operands = false;
11727 bool try_again = false;
11728 switch (rcode)
11729 {
11730 case LT:
11731 rcode = GT;
11732 swap_operands = true;
11733 try_again = true;
11734 break;
11735 case LTU:
11736 rcode = GTU;
11737 swap_operands = true;
11738 try_again = true;
11739 break;
11740 case NE:
11741 case UNLE:
11742 case UNLT:
11743 case UNGE:
11744 case UNGT:
11745 /* Invert condition and try again.
11746 e.g., A != B becomes ~(A==B). */
11747 {
11748 enum rtx_code rev_code;
11749 enum insn_code nor_code;
11750 rtx eq_rtx;
11751
11752 rev_code = reverse_condition_maybe_unordered (rcode);
11753 eq_rtx = rs6000_emit_vector_compare (rev_code, op0, op1,
11754 dest_mode);
11755
11756 nor_code = one_cmpl_optab->handlers[(int)dest_mode].insn_code;
11757 gcc_assert (nor_code != CODE_FOR_nothing);
11758 emit_insn (GEN_FCN (nor_code) (mask, eq_rtx));
11759
11760 if (dmode != dest_mode)
11761 {
11762 rtx temp = gen_reg_rtx (dest_mode);
11763 convert_move (temp, mask, 0);
11764 return temp;
11765 }
11766 return mask;
11767 }
11768 break;
11769 case GE:
11770 case GEU:
11771 case LE:
11772 case LEU:
11773 /* Try GT/GTU/LT/LTU OR EQ */
11774 {
11775 rtx c_rtx, eq_rtx;
11776 enum insn_code ior_code;
11777 enum rtx_code new_code;
11778
11779 switch (rcode)
11780 {
11781 case GE:
11782 new_code = GT;
11783 break;
11784
11785 case GEU:
11786 new_code = GTU;
11787 break;
11788
11789 case LE:
11790 new_code = LT;
11791 break;
11792
11793 case LEU:
11794 new_code = LTU;
11795 break;
11796
11797 default:
11798 gcc_unreachable ();
11799 }
11800
11801 c_rtx = rs6000_emit_vector_compare (new_code,
11802 op0, op1, dest_mode);
11803 eq_rtx = rs6000_emit_vector_compare (EQ, op0, op1,
11804 dest_mode);
11805
11806 ior_code = ior_optab->handlers[(int)dest_mode].insn_code;
11807 gcc_assert (ior_code != CODE_FOR_nothing);
11808 emit_insn (GEN_FCN (ior_code) (mask, c_rtx, eq_rtx));
11809 if (dmode != dest_mode)
11810 {
11811 rtx temp = gen_reg_rtx (dest_mode);
11812 convert_move (temp, mask, 0);
11813 return temp;
11814 }
11815 return mask;
11816 }
11817 break;
11818 default:
11819 gcc_unreachable ();
11820 }
11821
11822 if (try_again)
11823 {
11824 vec_cmp_insn = get_vec_cmp_insn (rcode, dest_mode, op_mode);
11825 /* You only get two chances. */
11826 gcc_assert (vec_cmp_insn != INSN_NOT_AVAILABLE);
11827 }
11828
11829 if (swap_operands)
11830 {
11831 rtx tmp;
11832 tmp = op0;
11833 op0 = op1;
11834 op1 = tmp;
11835 }
11836 }
11837
11838 emit_insn (gen_rtx_SET (VOIDmode, mask,
11839 gen_rtx_UNSPEC (dest_mode,
11840 gen_rtvec (2, op0, op1),
11841 vec_cmp_insn)));
11842 if (dmode != dest_mode)
11843 {
11844 rtx temp = gen_reg_rtx (dest_mode);
11845 convert_move (temp, mask, 0);
11846 return temp;
11847 }
11848 return mask;
11849 }
11850
11851 /* Return vector select instruction for MODE. Return INSN_NOT_AVAILABLE, if
11852 valid insn doesn exist for given mode. */
11853
11854 static int
get_vsel_insn(enum machine_mode mode)11855 get_vsel_insn (enum machine_mode mode)
11856 {
11857 switch (mode)
11858 {
11859 case V4SImode:
11860 return UNSPEC_VSEL4SI;
11861 break;
11862 case V4SFmode:
11863 return UNSPEC_VSEL4SF;
11864 break;
11865 case V8HImode:
11866 return UNSPEC_VSEL8HI;
11867 break;
11868 case V16QImode:
11869 return UNSPEC_VSEL16QI;
11870 break;
11871 default:
11872 return INSN_NOT_AVAILABLE;
11873 break;
11874 }
11875 return INSN_NOT_AVAILABLE;
11876 }
11877
11878 /* Emit vector select insn where DEST is destination using
11879 operands OP1, OP2 and MASK. */
11880
11881 static void
rs6000_emit_vector_select(rtx dest,rtx op1,rtx op2,rtx mask)11882 rs6000_emit_vector_select (rtx dest, rtx op1, rtx op2, rtx mask)
11883 {
11884 rtx t, temp;
11885 enum machine_mode dest_mode = GET_MODE (dest);
11886 int vsel_insn_index = get_vsel_insn (GET_MODE (dest));
11887
11888 temp = gen_reg_rtx (dest_mode);
11889
11890 /* For each vector element, select op1 when mask is 1 otherwise
11891 select op2. */
11892 t = gen_rtx_SET (VOIDmode, temp,
11893 gen_rtx_UNSPEC (dest_mode,
11894 gen_rtvec (3, op2, op1, mask),
11895 vsel_insn_index));
11896 emit_insn (t);
11897 emit_move_insn (dest, temp);
11898 return;
11899 }
11900
11901 /* Emit vector conditional expression.
11902 DEST is destination. OP1 and OP2 are two VEC_COND_EXPR operands.
11903 CC_OP0 and CC_OP1 are the two operands for the relation operation COND. */
11904
11905 int
rs6000_emit_vector_cond_expr(rtx dest,rtx op1,rtx op2,rtx cond,rtx cc_op0,rtx cc_op1)11906 rs6000_emit_vector_cond_expr (rtx dest, rtx op1, rtx op2,
11907 rtx cond, rtx cc_op0, rtx cc_op1)
11908 {
11909 enum machine_mode dest_mode = GET_MODE (dest);
11910 enum rtx_code rcode = GET_CODE (cond);
11911 rtx mask;
11912
11913 if (!TARGET_ALTIVEC)
11914 return 0;
11915
11916 /* Get the vector mask for the given relational operations. */
11917 mask = rs6000_emit_vector_compare (rcode, cc_op0, cc_op1, dest_mode);
11918
11919 rs6000_emit_vector_select (dest, op1, op2, mask);
11920
11921 return 1;
11922 }
11923
11924 /* Emit a conditional move: move TRUE_COND to DEST if OP of the
11925 operands of the last comparison is nonzero/true, FALSE_COND if it
11926 is zero/false. Return 0 if the hardware has no such operation. */
11927
11928 int
rs6000_emit_cmove(rtx dest,rtx op,rtx true_cond,rtx false_cond)11929 rs6000_emit_cmove (rtx dest, rtx op, rtx true_cond, rtx false_cond)
11930 {
11931 enum rtx_code code = GET_CODE (op);
11932 rtx op0 = rs6000_compare_op0;
11933 rtx op1 = rs6000_compare_op1;
11934 REAL_VALUE_TYPE c1;
11935 enum machine_mode compare_mode = GET_MODE (op0);
11936 enum machine_mode result_mode = GET_MODE (dest);
11937 rtx temp;
11938 bool is_against_zero;
11939
11940 /* These modes should always match. */
11941 if (GET_MODE (op1) != compare_mode
11942 /* In the isel case however, we can use a compare immediate, so
11943 op1 may be a small constant. */
11944 && (!TARGET_ISEL || !short_cint_operand (op1, VOIDmode)))
11945 return 0;
11946 if (GET_MODE (true_cond) != result_mode)
11947 return 0;
11948 if (GET_MODE (false_cond) != result_mode)
11949 return 0;
11950
11951 /* First, work out if the hardware can do this at all, or
11952 if it's too slow.... */
11953 if (! rs6000_compare_fp_p)
11954 {
11955 if (TARGET_ISEL)
11956 return rs6000_emit_int_cmove (dest, op, true_cond, false_cond);
11957 return 0;
11958 }
11959 else if (TARGET_E500 && TARGET_HARD_FLOAT && !TARGET_FPRS
11960 && SCALAR_FLOAT_MODE_P (compare_mode))
11961 return 0;
11962
11963 is_against_zero = op1 == CONST0_RTX (compare_mode);
11964
11965 /* A floating-point subtract might overflow, underflow, or produce
11966 an inexact result, thus changing the floating-point flags, so it
11967 can't be generated if we care about that. It's safe if one side
11968 of the construct is zero, since then no subtract will be
11969 generated. */
11970 if (SCALAR_FLOAT_MODE_P (compare_mode)
11971 && flag_trapping_math && ! is_against_zero)
11972 return 0;
11973
11974 /* Eliminate half of the comparisons by switching operands, this
11975 makes the remaining code simpler. */
11976 if (code == UNLT || code == UNGT || code == UNORDERED || code == NE
11977 || code == LTGT || code == LT || code == UNLE)
11978 {
11979 code = reverse_condition_maybe_unordered (code);
11980 temp = true_cond;
11981 true_cond = false_cond;
11982 false_cond = temp;
11983 }
11984
11985 /* UNEQ and LTGT take four instructions for a comparison with zero,
11986 it'll probably be faster to use a branch here too. */
11987 if (code == UNEQ && HONOR_NANS (compare_mode))
11988 return 0;
11989
11990 if (GET_CODE (op1) == CONST_DOUBLE)
11991 REAL_VALUE_FROM_CONST_DOUBLE (c1, op1);
11992
11993 /* We're going to try to implement comparisons by performing
11994 a subtract, then comparing against zero. Unfortunately,
11995 Inf - Inf is NaN which is not zero, and so if we don't
11996 know that the operand is finite and the comparison
11997 would treat EQ different to UNORDERED, we can't do it. */
11998 if (HONOR_INFINITIES (compare_mode)
11999 && code != GT && code != UNGE
12000 && (GET_CODE (op1) != CONST_DOUBLE || real_isinf (&c1))
12001 /* Constructs of the form (a OP b ? a : b) are safe. */
12002 && ((! rtx_equal_p (op0, false_cond) && ! rtx_equal_p (op1, false_cond))
12003 || (! rtx_equal_p (op0, true_cond)
12004 && ! rtx_equal_p (op1, true_cond))))
12005 return 0;
12006
12007 /* At this point we know we can use fsel. */
12008
12009 /* Reduce the comparison to a comparison against zero. */
12010 if (! is_against_zero)
12011 {
12012 temp = gen_reg_rtx (compare_mode);
12013 emit_insn (gen_rtx_SET (VOIDmode, temp,
12014 gen_rtx_MINUS (compare_mode, op0, op1)));
12015 op0 = temp;
12016 op1 = CONST0_RTX (compare_mode);
12017 }
12018
12019 /* If we don't care about NaNs we can reduce some of the comparisons
12020 down to faster ones. */
12021 if (! HONOR_NANS (compare_mode))
12022 switch (code)
12023 {
12024 case GT:
12025 code = LE;
12026 temp = true_cond;
12027 true_cond = false_cond;
12028 false_cond = temp;
12029 break;
12030 case UNGE:
12031 code = GE;
12032 break;
12033 case UNEQ:
12034 code = EQ;
12035 break;
12036 default:
12037 break;
12038 }
12039
12040 /* Now, reduce everything down to a GE. */
12041 switch (code)
12042 {
12043 case GE:
12044 break;
12045
12046 case LE:
12047 temp = gen_reg_rtx (compare_mode);
12048 emit_insn (gen_rtx_SET (VOIDmode, temp, gen_rtx_NEG (compare_mode, op0)));
12049 op0 = temp;
12050 break;
12051
12052 case ORDERED:
12053 temp = gen_reg_rtx (compare_mode);
12054 emit_insn (gen_rtx_SET (VOIDmode, temp, gen_rtx_ABS (compare_mode, op0)));
12055 op0 = temp;
12056 break;
12057
12058 case EQ:
12059 temp = gen_reg_rtx (compare_mode);
12060 emit_insn (gen_rtx_SET (VOIDmode, temp,
12061 gen_rtx_NEG (compare_mode,
12062 gen_rtx_ABS (compare_mode, op0))));
12063 op0 = temp;
12064 break;
12065
12066 case UNGE:
12067 /* a UNGE 0 <-> (a GE 0 || -a UNLT 0) */
12068 temp = gen_reg_rtx (result_mode);
12069 emit_insn (gen_rtx_SET (VOIDmode, temp,
12070 gen_rtx_IF_THEN_ELSE (result_mode,
12071 gen_rtx_GE (VOIDmode,
12072 op0, op1),
12073 true_cond, false_cond)));
12074 false_cond = true_cond;
12075 true_cond = temp;
12076
12077 temp = gen_reg_rtx (compare_mode);
12078 emit_insn (gen_rtx_SET (VOIDmode, temp, gen_rtx_NEG (compare_mode, op0)));
12079 op0 = temp;
12080 break;
12081
12082 case GT:
12083 /* a GT 0 <-> (a GE 0 && -a UNLT 0) */
12084 temp = gen_reg_rtx (result_mode);
12085 emit_insn (gen_rtx_SET (VOIDmode, temp,
12086 gen_rtx_IF_THEN_ELSE (result_mode,
12087 gen_rtx_GE (VOIDmode,
12088 op0, op1),
12089 true_cond, false_cond)));
12090 true_cond = false_cond;
12091 false_cond = temp;
12092
12093 temp = gen_reg_rtx (compare_mode);
12094 emit_insn (gen_rtx_SET (VOIDmode, temp, gen_rtx_NEG (compare_mode, op0)));
12095 op0 = temp;
12096 break;
12097
12098 default:
12099 gcc_unreachable ();
12100 }
12101
12102 emit_insn (gen_rtx_SET (VOIDmode, dest,
12103 gen_rtx_IF_THEN_ELSE (result_mode,
12104 gen_rtx_GE (VOIDmode,
12105 op0, op1),
12106 true_cond, false_cond)));
12107 return 1;
12108 }
12109
12110 /* Same as above, but for ints (isel). */
12111
12112 static int
rs6000_emit_int_cmove(rtx dest,rtx op,rtx true_cond,rtx false_cond)12113 rs6000_emit_int_cmove (rtx dest, rtx op, rtx true_cond, rtx false_cond)
12114 {
12115 rtx condition_rtx, cr;
12116
12117 /* All isel implementations thus far are 32-bits. */
12118 if (GET_MODE (rs6000_compare_op0) != SImode)
12119 return 0;
12120
12121 /* We still have to do the compare, because isel doesn't do a
12122 compare, it just looks at the CRx bits set by a previous compare
12123 instruction. */
12124 condition_rtx = rs6000_generate_compare (GET_CODE (op));
12125 cr = XEXP (condition_rtx, 0);
12126
12127 if (GET_MODE (cr) == CCmode)
12128 emit_insn (gen_isel_signed (dest, condition_rtx,
12129 true_cond, false_cond, cr));
12130 else
12131 emit_insn (gen_isel_unsigned (dest, condition_rtx,
12132 true_cond, false_cond, cr));
12133
12134 return 1;
12135 }
12136
12137 const char *
output_isel(rtx * operands)12138 output_isel (rtx *operands)
12139 {
12140 enum rtx_code code;
12141
12142 code = GET_CODE (operands[1]);
12143 if (code == GE || code == GEU || code == LE || code == LEU || code == NE)
12144 {
12145 PUT_CODE (operands[1], reverse_condition (code));
12146 return "isel %0,%3,%2,%j1";
12147 }
12148 else
12149 return "isel %0,%2,%3,%j1";
12150 }
12151
12152 void
rs6000_emit_minmax(rtx dest,enum rtx_code code,rtx op0,rtx op1)12153 rs6000_emit_minmax (rtx dest, enum rtx_code code, rtx op0, rtx op1)
12154 {
12155 enum machine_mode mode = GET_MODE (op0);
12156 enum rtx_code c;
12157 rtx target;
12158
12159 if (code == SMAX || code == SMIN)
12160 c = GE;
12161 else
12162 c = GEU;
12163
12164 if (code == SMAX || code == UMAX)
12165 target = emit_conditional_move (dest, c, op0, op1, mode,
12166 op0, op1, mode, 0);
12167 else
12168 target = emit_conditional_move (dest, c, op0, op1, mode,
12169 op1, op0, mode, 0);
12170 gcc_assert (target);
12171 if (target != dest)
12172 emit_move_insn (dest, target);
12173 }
12174
12175 /* Emit instructions to perform a load-reserved/store-conditional operation.
12176 The operation performed is an atomic
12177 (set M (CODE:MODE M OP))
12178 If not NULL, BEFORE is atomically set to M before the operation, and
12179 AFTER is set to M after the operation (that is, (CODE:MODE M OP)).
12180 If SYNC_P then a memory barrier is emitted before the operation.
12181 Either OP or M may be wrapped in a NOT operation. */
12182
12183 void
rs6000_emit_sync(enum rtx_code code,enum machine_mode mode,rtx m,rtx op,rtx before_param,rtx after_param,bool sync_p)12184 rs6000_emit_sync (enum rtx_code code, enum machine_mode mode,
12185 rtx m, rtx op, rtx before_param, rtx after_param,
12186 bool sync_p)
12187 {
12188 enum machine_mode used_mode;
12189 rtx the_op, set_before, set_after, set_atomic, cc_scratch, before, after;
12190 rtx used_m;
12191 rtvec vec;
12192 HOST_WIDE_INT imask = GET_MODE_MASK (mode);
12193 rtx shift = NULL_RTX;
12194
12195 if (sync_p)
12196 emit_insn (gen_memory_barrier ());
12197
12198 if (GET_CODE (m) == NOT)
12199 used_m = XEXP (m, 0);
12200 else
12201 used_m = m;
12202
12203 /* If this is smaller than SImode, we'll have to use SImode with
12204 adjustments. */
12205 if (mode == QImode || mode == HImode)
12206 {
12207 rtx newop, oldop;
12208
12209 if (MEM_ALIGN (used_m) >= 32)
12210 {
12211 int ishift = 0;
12212 if (BYTES_BIG_ENDIAN)
12213 ishift = GET_MODE_BITSIZE (SImode) - GET_MODE_BITSIZE (mode);
12214
12215 shift = GEN_INT (ishift);
12216 }
12217 else
12218 {
12219 rtx addrSI, aligned_addr;
12220 int shift_mask = mode == QImode ? 0x18 : 0x10;
12221
12222 addrSI = force_reg (SImode, gen_lowpart_common (SImode,
12223 XEXP (used_m, 0)));
12224 shift = gen_reg_rtx (SImode);
12225
12226 emit_insn (gen_rlwinm (shift, addrSI, GEN_INT (3),
12227 GEN_INT (shift_mask)));
12228 emit_insn (gen_xorsi3 (shift, shift, GEN_INT (shift_mask)));
12229
12230 aligned_addr = expand_binop (Pmode, and_optab,
12231 XEXP (used_m, 0),
12232 GEN_INT (-4), NULL_RTX,
12233 1, OPTAB_LIB_WIDEN);
12234 used_m = change_address (used_m, SImode, aligned_addr);
12235 set_mem_align (used_m, 32);
12236 /* It's safe to keep the old alias set of USED_M, because
12237 the operation is atomic and only affects the original
12238 USED_M. */
12239 if (GET_CODE (m) == NOT)
12240 m = gen_rtx_NOT (SImode, used_m);
12241 else
12242 m = used_m;
12243 }
12244
12245 if (GET_CODE (op) == NOT)
12246 {
12247 oldop = lowpart_subreg (SImode, XEXP (op, 0), mode);
12248 oldop = gen_rtx_NOT (SImode, oldop);
12249 }
12250 else
12251 oldop = lowpart_subreg (SImode, op, mode);
12252
12253 switch (code)
12254 {
12255 case IOR:
12256 case XOR:
12257 newop = expand_binop (SImode, and_optab,
12258 oldop, GEN_INT (imask), NULL_RTX,
12259 1, OPTAB_LIB_WIDEN);
12260 emit_insn (gen_ashlsi3 (newop, newop, shift));
12261 break;
12262
12263 case AND:
12264 newop = expand_binop (SImode, ior_optab,
12265 oldop, GEN_INT (~imask), NULL_RTX,
12266 1, OPTAB_LIB_WIDEN);
12267 emit_insn (gen_rotlsi3 (newop, newop, shift));
12268 break;
12269
12270 case PLUS:
12271 case MINUS:
12272 {
12273 rtx mask;
12274
12275 newop = expand_binop (SImode, and_optab,
12276 oldop, GEN_INT (imask), NULL_RTX,
12277 1, OPTAB_LIB_WIDEN);
12278 emit_insn (gen_ashlsi3 (newop, newop, shift));
12279
12280 mask = gen_reg_rtx (SImode);
12281 emit_move_insn (mask, GEN_INT (imask));
12282 emit_insn (gen_ashlsi3 (mask, mask, shift));
12283
12284 if (code == PLUS)
12285 newop = gen_rtx_PLUS (SImode, m, newop);
12286 else
12287 newop = gen_rtx_MINUS (SImode, m, newop);
12288 newop = gen_rtx_AND (SImode, newop, mask);
12289 newop = gen_rtx_IOR (SImode, newop,
12290 gen_rtx_AND (SImode,
12291 gen_rtx_NOT (SImode, mask),
12292 m));
12293 break;
12294 }
12295
12296 default:
12297 gcc_unreachable ();
12298 }
12299
12300 if (GET_CODE (m) == NOT)
12301 {
12302 rtx mask, xorm;
12303
12304 mask = gen_reg_rtx (SImode);
12305 emit_move_insn (mask, GEN_INT (imask));
12306 emit_insn (gen_ashlsi3 (mask, mask, shift));
12307
12308 xorm = gen_rtx_XOR (SImode, used_m, mask);
12309 /* Depending on the value of 'op', the XOR or the operation might
12310 be able to be simplified away. */
12311 newop = simplify_gen_binary (code, SImode, xorm, newop);
12312 }
12313 op = newop;
12314 used_mode = SImode;
12315 before = gen_reg_rtx (used_mode);
12316 after = gen_reg_rtx (used_mode);
12317 }
12318 else
12319 {
12320 used_mode = mode;
12321 before = before_param;
12322 after = after_param;
12323
12324 if (before == NULL_RTX)
12325 before = gen_reg_rtx (used_mode);
12326 if (after == NULL_RTX)
12327 after = gen_reg_rtx (used_mode);
12328 }
12329
12330 if ((code == PLUS || code == MINUS || GET_CODE (m) == NOT)
12331 && used_mode != mode)
12332 the_op = op; /* Computed above. */
12333 else if (GET_CODE (op) == NOT && GET_CODE (m) != NOT)
12334 the_op = gen_rtx_fmt_ee (code, used_mode, op, m);
12335 else
12336 the_op = gen_rtx_fmt_ee (code, used_mode, m, op);
12337
12338 set_after = gen_rtx_SET (VOIDmode, after, the_op);
12339 set_before = gen_rtx_SET (VOIDmode, before, used_m);
12340 set_atomic = gen_rtx_SET (VOIDmode, used_m,
12341 gen_rtx_UNSPEC (used_mode,
12342 gen_rtvec (1, the_op),
12343 UNSPEC_SYNC_OP));
12344 cc_scratch = gen_rtx_CLOBBER (VOIDmode, gen_rtx_SCRATCH (CCmode));
12345
12346 if ((code == PLUS || code == MINUS) && used_mode != mode)
12347 vec = gen_rtvec (5, set_after, set_before, set_atomic, cc_scratch,
12348 gen_rtx_CLOBBER (VOIDmode, gen_rtx_SCRATCH (SImode)));
12349 else
12350 vec = gen_rtvec (4, set_after, set_before, set_atomic, cc_scratch);
12351 emit_insn (gen_rtx_PARALLEL (VOIDmode, vec));
12352
12353 /* Shift and mask the return values properly. */
12354 if (used_mode != mode && before_param)
12355 {
12356 emit_insn (gen_lshrsi3 (before, before, shift));
12357 convert_move (before_param, before, 1);
12358 }
12359
12360 if (used_mode != mode && after_param)
12361 {
12362 emit_insn (gen_lshrsi3 (after, after, shift));
12363 convert_move (after_param, after, 1);
12364 }
12365
12366 /* The previous sequence will end with a branch that's dependent on
12367 the conditional store, so placing an isync will ensure that no
12368 other instructions (especially, no load or store instructions)
12369 can start before the atomic operation completes. */
12370 if (sync_p)
12371 emit_insn (gen_isync ());
12372 }
12373
12374 /* A subroutine of the atomic operation splitters. Jump to LABEL if
12375 COND is true. Mark the jump as unlikely to be taken. */
12376
12377 static void
emit_unlikely_jump(rtx cond,rtx label)12378 emit_unlikely_jump (rtx cond, rtx label)
12379 {
12380 rtx very_unlikely = GEN_INT (REG_BR_PROB_BASE / 100 - 1);
12381 rtx x;
12382
12383 x = gen_rtx_IF_THEN_ELSE (VOIDmode, cond, label, pc_rtx);
12384 x = emit_jump_insn (gen_rtx_SET (VOIDmode, pc_rtx, x));
12385 REG_NOTES (x) = gen_rtx_EXPR_LIST (REG_BR_PROB, very_unlikely, NULL_RTX);
12386 }
12387
12388 /* A subroutine of the atomic operation splitters. Emit a load-locked
12389 instruction in MODE. */
12390
12391 static void
emit_load_locked(enum machine_mode mode,rtx reg,rtx mem)12392 emit_load_locked (enum machine_mode mode, rtx reg, rtx mem)
12393 {
12394 rtx (*fn) (rtx, rtx) = NULL;
12395 if (mode == SImode)
12396 fn = gen_load_locked_si;
12397 else if (mode == DImode)
12398 fn = gen_load_locked_di;
12399 emit_insn (fn (reg, mem));
12400 }
12401
12402 /* A subroutine of the atomic operation splitters. Emit a store-conditional
12403 instruction in MODE. */
12404
12405 static void
emit_store_conditional(enum machine_mode mode,rtx res,rtx mem,rtx val)12406 emit_store_conditional (enum machine_mode mode, rtx res, rtx mem, rtx val)
12407 {
12408 rtx (*fn) (rtx, rtx, rtx) = NULL;
12409 if (mode == SImode)
12410 fn = gen_store_conditional_si;
12411 else if (mode == DImode)
12412 fn = gen_store_conditional_di;
12413
12414 /* Emit sync before stwcx. to address PPC405 Erratum. */
12415 if (PPC405_ERRATUM77)
12416 emit_insn (gen_memory_barrier ());
12417
12418 emit_insn (fn (res, mem, val));
12419 }
12420
12421 /* Expand an an atomic fetch-and-operate pattern. CODE is the binary operation
12422 to perform. MEM is the memory on which to operate. VAL is the second
12423 operand of the binary operator. BEFORE and AFTER are optional locations to
12424 return the value of MEM either before of after the operation. SCRATCH is
12425 a scratch register. */
12426
12427 void
rs6000_split_atomic_op(enum rtx_code code,rtx mem,rtx val,rtx before,rtx after,rtx scratch)12428 rs6000_split_atomic_op (enum rtx_code code, rtx mem, rtx val,
12429 rtx before, rtx after, rtx scratch)
12430 {
12431 enum machine_mode mode = GET_MODE (mem);
12432 rtx label, x, cond = gen_rtx_REG (CCmode, CR0_REGNO);
12433
12434 emit_insn (gen_memory_barrier ());
12435
12436 label = gen_label_rtx ();
12437 emit_label (label);
12438 label = gen_rtx_LABEL_REF (VOIDmode, label);
12439
12440 if (before == NULL_RTX)
12441 before = scratch;
12442 emit_load_locked (mode, before, mem);
12443
12444 if (code == NOT)
12445 x = gen_rtx_AND (mode, gen_rtx_NOT (mode, before), val);
12446 else if (code == AND)
12447 x = gen_rtx_UNSPEC (mode, gen_rtvec (2, before, val), UNSPEC_AND);
12448 else
12449 x = gen_rtx_fmt_ee (code, mode, before, val);
12450
12451 if (after != NULL_RTX)
12452 emit_insn (gen_rtx_SET (VOIDmode, after, copy_rtx (x)));
12453 emit_insn (gen_rtx_SET (VOIDmode, scratch, x));
12454
12455 emit_store_conditional (mode, cond, mem, scratch);
12456
12457 x = gen_rtx_NE (VOIDmode, cond, const0_rtx);
12458 emit_unlikely_jump (x, label);
12459
12460 emit_insn (gen_isync ());
12461 }
12462
12463 /* Expand an atomic compare and swap operation. MEM is the memory on which
12464 to operate. OLDVAL is the old value to be compared. NEWVAL is the new
12465 value to be stored. SCRATCH is a scratch GPR. */
12466
12467 void
rs6000_split_compare_and_swap(rtx retval,rtx mem,rtx oldval,rtx newval,rtx scratch)12468 rs6000_split_compare_and_swap (rtx retval, rtx mem, rtx oldval, rtx newval,
12469 rtx scratch)
12470 {
12471 enum machine_mode mode = GET_MODE (mem);
12472 rtx label1, label2, x, cond = gen_rtx_REG (CCmode, CR0_REGNO);
12473
12474 emit_insn (gen_memory_barrier ());
12475
12476 label1 = gen_rtx_LABEL_REF (VOIDmode, gen_label_rtx ());
12477 label2 = gen_rtx_LABEL_REF (VOIDmode, gen_label_rtx ());
12478 emit_label (XEXP (label1, 0));
12479
12480 emit_load_locked (mode, retval, mem);
12481
12482 x = gen_rtx_COMPARE (CCmode, retval, oldval);
12483 emit_insn (gen_rtx_SET (VOIDmode, cond, x));
12484
12485 x = gen_rtx_NE (VOIDmode, cond, const0_rtx);
12486 emit_unlikely_jump (x, label2);
12487
12488 emit_move_insn (scratch, newval);
12489 emit_store_conditional (mode, cond, mem, scratch);
12490
12491 x = gen_rtx_NE (VOIDmode, cond, const0_rtx);
12492 emit_unlikely_jump (x, label1);
12493
12494 emit_insn (gen_isync ());
12495 emit_label (XEXP (label2, 0));
12496 }
12497
12498 /* Expand an atomic test and set operation. MEM is the memory on which
12499 to operate. VAL is the value set. SCRATCH is a scratch GPR. */
12500
12501 void
rs6000_split_lock_test_and_set(rtx retval,rtx mem,rtx val,rtx scratch)12502 rs6000_split_lock_test_and_set (rtx retval, rtx mem, rtx val, rtx scratch)
12503 {
12504 enum machine_mode mode = GET_MODE (mem);
12505 rtx label, x, cond = gen_rtx_REG (CCmode, CR0_REGNO);
12506
12507 emit_insn (gen_memory_barrier ());
12508
12509 label = gen_rtx_LABEL_REF (VOIDmode, gen_label_rtx ());
12510 emit_label (XEXP (label, 0));
12511
12512 emit_load_locked (mode, retval, mem);
12513 emit_move_insn (scratch, val);
12514 emit_store_conditional (mode, cond, mem, scratch);
12515
12516 x = gen_rtx_NE (VOIDmode, cond, const0_rtx);
12517 emit_unlikely_jump (x, label);
12518
12519 emit_insn (gen_isync ());
12520 }
12521
12522 void
rs6000_expand_compare_and_swapqhi(rtx dst,rtx mem,rtx oldval,rtx newval)12523 rs6000_expand_compare_and_swapqhi (rtx dst, rtx mem, rtx oldval, rtx newval)
12524 {
12525 enum machine_mode mode = GET_MODE (mem);
12526 rtx addrSI, align, wdst, shift, mask;
12527 HOST_WIDE_INT shift_mask = mode == QImode ? 0x18 : 0x10;
12528 HOST_WIDE_INT imask = GET_MODE_MASK (mode);
12529
12530 /* Shift amount for subword relative to aligned word. */
12531 addrSI = force_reg (SImode, gen_lowpart_common (SImode, XEXP (mem, 0)));
12532 shift = gen_reg_rtx (SImode);
12533 emit_insn (gen_rlwinm (shift, addrSI, GEN_INT (3),
12534 GEN_INT (shift_mask)));
12535 emit_insn (gen_xorsi3 (shift, shift, GEN_INT (shift_mask)));
12536
12537 /* Shift and mask old value into position within word. */
12538 oldval = convert_modes (SImode, mode, oldval, 1);
12539 oldval = expand_binop (SImode, and_optab,
12540 oldval, GEN_INT (imask), NULL_RTX,
12541 1, OPTAB_LIB_WIDEN);
12542 emit_insn (gen_ashlsi3 (oldval, oldval, shift));
12543
12544 /* Shift and mask new value into position within word. */
12545 newval = convert_modes (SImode, mode, newval, 1);
12546 newval = expand_binop (SImode, and_optab,
12547 newval, GEN_INT (imask), NULL_RTX,
12548 1, OPTAB_LIB_WIDEN);
12549 emit_insn (gen_ashlsi3 (newval, newval, shift));
12550
12551 /* Mask for insertion. */
12552 mask = gen_reg_rtx (SImode);
12553 emit_move_insn (mask, GEN_INT (imask));
12554 emit_insn (gen_ashlsi3 (mask, mask, shift));
12555
12556 /* Address of aligned word containing subword. */
12557 align = expand_binop (Pmode, and_optab, XEXP (mem, 0), GEN_INT (-4),
12558 NULL_RTX, 1, OPTAB_LIB_WIDEN);
12559 mem = change_address (mem, SImode, align);
12560 set_mem_align (mem, 32);
12561 MEM_VOLATILE_P (mem) = 1;
12562
12563 wdst = gen_reg_rtx (SImode);
12564 emit_insn (gen_sync_compare_and_swapqhi_internal (wdst, mask,
12565 oldval, newval, mem));
12566
12567 emit_move_insn (dst, gen_lowpart (mode, wdst));
12568 }
12569
12570 void
rs6000_split_compare_and_swapqhi(rtx dest,rtx mask,rtx oldval,rtx newval,rtx mem,rtx scratch)12571 rs6000_split_compare_and_swapqhi (rtx dest, rtx mask,
12572 rtx oldval, rtx newval, rtx mem,
12573 rtx scratch)
12574 {
12575 rtx label1, label2, x, cond = gen_rtx_REG (CCmode, CR0_REGNO);
12576
12577 emit_insn (gen_memory_barrier ());
12578 label1 = gen_rtx_LABEL_REF (VOIDmode, gen_label_rtx ());
12579 label2 = gen_rtx_LABEL_REF (VOIDmode, gen_label_rtx ());
12580 emit_label (XEXP (label1, 0));
12581
12582 emit_load_locked (SImode, scratch, mem);
12583
12584 /* Mask subword within loaded value for comparison with oldval.
12585 Use UNSPEC_AND to avoid clobber.*/
12586 emit_insn (gen_rtx_SET (SImode, dest,
12587 gen_rtx_UNSPEC (SImode,
12588 gen_rtvec (2, scratch, mask),
12589 UNSPEC_AND)));
12590
12591 x = gen_rtx_COMPARE (CCmode, dest, oldval);
12592 emit_insn (gen_rtx_SET (VOIDmode, cond, x));
12593
12594 x = gen_rtx_NE (VOIDmode, cond, const0_rtx);
12595 emit_unlikely_jump (x, label2);
12596
12597 /* Clear subword within loaded value for insertion of new value. */
12598 emit_insn (gen_rtx_SET (SImode, scratch,
12599 gen_rtx_AND (SImode,
12600 gen_rtx_NOT (SImode, mask), scratch)));
12601 emit_insn (gen_iorsi3 (scratch, scratch, newval));
12602 emit_store_conditional (SImode, cond, mem, scratch);
12603
12604 x = gen_rtx_NE (VOIDmode, cond, const0_rtx);
12605 emit_unlikely_jump (x, label1);
12606
12607 emit_insn (gen_isync ());
12608 emit_label (XEXP (label2, 0));
12609 }
12610
12611
12612 /* Emit instructions to move SRC to DST. Called by splitters for
12613 multi-register moves. It will emit at most one instruction for
12614 each register that is accessed; that is, it won't emit li/lis pairs
12615 (or equivalent for 64-bit code). One of SRC or DST must be a hard
12616 register. */
12617
12618 void
rs6000_split_multireg_move(rtx dst,rtx src)12619 rs6000_split_multireg_move (rtx dst, rtx src)
12620 {
12621 /* The register number of the first register being moved. */
12622 int reg;
12623 /* The mode that is to be moved. */
12624 enum machine_mode mode;
12625 /* The mode that the move is being done in, and its size. */
12626 enum machine_mode reg_mode;
12627 int reg_mode_size;
12628 /* The number of registers that will be moved. */
12629 int nregs;
12630
12631 reg = REG_P (dst) ? REGNO (dst) : REGNO (src);
12632 mode = GET_MODE (dst);
12633 nregs = hard_regno_nregs[reg][mode];
12634 if (FP_REGNO_P (reg))
12635 reg_mode = DFmode;
12636 else if (ALTIVEC_REGNO_P (reg))
12637 reg_mode = V16QImode;
12638 else if (TARGET_E500_DOUBLE && mode == TFmode)
12639 reg_mode = DFmode;
12640 else
12641 reg_mode = word_mode;
12642 reg_mode_size = GET_MODE_SIZE (reg_mode);
12643
12644 gcc_assert (reg_mode_size * nregs == GET_MODE_SIZE (mode));
12645
12646 if (REG_P (src) && REG_P (dst) && (REGNO (src) < REGNO (dst)))
12647 {
12648 /* Move register range backwards, if we might have destructive
12649 overlap. */
12650 int i;
12651 for (i = nregs - 1; i >= 0; i--)
12652 emit_insn (gen_rtx_SET (VOIDmode,
12653 simplify_gen_subreg (reg_mode, dst, mode,
12654 i * reg_mode_size),
12655 simplify_gen_subreg (reg_mode, src, mode,
12656 i * reg_mode_size)));
12657 }
12658 else
12659 {
12660 int i;
12661 int j = -1;
12662 bool used_update = false;
12663
12664 if (MEM_P (src) && INT_REGNO_P (reg))
12665 {
12666 rtx breg;
12667
12668 if (GET_CODE (XEXP (src, 0)) == PRE_INC
12669 || GET_CODE (XEXP (src, 0)) == PRE_DEC)
12670 {
12671 rtx delta_rtx;
12672 breg = XEXP (XEXP (src, 0), 0);
12673 delta_rtx = (GET_CODE (XEXP (src, 0)) == PRE_INC
12674 ? GEN_INT (GET_MODE_SIZE (GET_MODE (src)))
12675 : GEN_INT (-GET_MODE_SIZE (GET_MODE (src))));
12676 emit_insn (TARGET_32BIT
12677 ? gen_addsi3 (breg, breg, delta_rtx)
12678 : gen_adddi3 (breg, breg, delta_rtx));
12679 src = replace_equiv_address (src, breg);
12680 }
12681 else if (! rs6000_offsettable_memref_p (src))
12682 {
12683 rtx basereg;
12684 basereg = gen_rtx_REG (Pmode, reg);
12685 emit_insn (gen_rtx_SET (VOIDmode, basereg, XEXP (src, 0)));
12686 src = replace_equiv_address (src, basereg);
12687 }
12688
12689 breg = XEXP (src, 0);
12690 if (GET_CODE (breg) == PLUS || GET_CODE (breg) == LO_SUM)
12691 breg = XEXP (breg, 0);
12692
12693 /* If the base register we are using to address memory is
12694 also a destination reg, then change that register last. */
12695 if (REG_P (breg)
12696 && REGNO (breg) >= REGNO (dst)
12697 && REGNO (breg) < REGNO (dst) + nregs)
12698 j = REGNO (breg) - REGNO (dst);
12699 }
12700
12701 if (GET_CODE (dst) == MEM && INT_REGNO_P (reg))
12702 {
12703 rtx breg;
12704
12705 if (GET_CODE (XEXP (dst, 0)) == PRE_INC
12706 || GET_CODE (XEXP (dst, 0)) == PRE_DEC)
12707 {
12708 rtx delta_rtx;
12709 breg = XEXP (XEXP (dst, 0), 0);
12710 delta_rtx = (GET_CODE (XEXP (dst, 0)) == PRE_INC
12711 ? GEN_INT (GET_MODE_SIZE (GET_MODE (dst)))
12712 : GEN_INT (-GET_MODE_SIZE (GET_MODE (dst))));
12713
12714 /* We have to update the breg before doing the store.
12715 Use store with update, if available. */
12716
12717 if (TARGET_UPDATE)
12718 {
12719 rtx nsrc = simplify_gen_subreg (reg_mode, src, mode, 0);
12720 emit_insn (TARGET_32BIT
12721 ? (TARGET_POWERPC64
12722 ? gen_movdi_si_update (breg, breg, delta_rtx, nsrc)
12723 : gen_movsi_update (breg, breg, delta_rtx, nsrc))
12724 : gen_movdi_di_update (breg, breg, delta_rtx, nsrc));
12725 used_update = true;
12726 }
12727 else
12728 emit_insn (TARGET_32BIT
12729 ? gen_addsi3 (breg, breg, delta_rtx)
12730 : gen_adddi3 (breg, breg, delta_rtx));
12731 dst = replace_equiv_address (dst, breg);
12732 }
12733 else
12734 gcc_assert (rs6000_offsettable_memref_p (dst));
12735 }
12736
12737 for (i = 0; i < nregs; i++)
12738 {
12739 /* Calculate index to next subword. */
12740 ++j;
12741 if (j == nregs)
12742 j = 0;
12743
12744 /* If compiler already emitted move of first word by
12745 store with update, no need to do anything. */
12746 if (j == 0 && used_update)
12747 continue;
12748
12749 emit_insn (gen_rtx_SET (VOIDmode,
12750 simplify_gen_subreg (reg_mode, dst, mode,
12751 j * reg_mode_size),
12752 simplify_gen_subreg (reg_mode, src, mode,
12753 j * reg_mode_size)));
12754 }
12755 }
12756 }
12757
12758
12759 /* This page contains routines that are used to determine what the
12760 function prologue and epilogue code will do and write them out. */
12761
12762 /* Return the first fixed-point register that is required to be
12763 saved. 32 if none. */
12764
12765 int
first_reg_to_save(void)12766 first_reg_to_save (void)
12767 {
12768 int first_reg;
12769
12770 /* Find lowest numbered live register. */
12771 for (first_reg = 13; first_reg <= 31; first_reg++)
12772 if (regs_ever_live[first_reg]
12773 && (! call_used_regs[first_reg]
12774 || (first_reg == RS6000_PIC_OFFSET_TABLE_REGNUM
12775 && ((DEFAULT_ABI == ABI_V4 && flag_pic != 0)
12776 || (DEFAULT_ABI == ABI_DARWIN && flag_pic)
12777 || (TARGET_TOC && TARGET_MINIMAL_TOC)))))
12778 break;
12779
12780 #if TARGET_MACHO
12781 if (flag_pic
12782 && current_function_uses_pic_offset_table
12783 && first_reg > RS6000_PIC_OFFSET_TABLE_REGNUM)
12784 return RS6000_PIC_OFFSET_TABLE_REGNUM;
12785 #endif
12786
12787 return first_reg;
12788 }
12789
12790 /* Similar, for FP regs. */
12791
12792 int
first_fp_reg_to_save(void)12793 first_fp_reg_to_save (void)
12794 {
12795 int first_reg;
12796
12797 /* Find lowest numbered live register. */
12798 for (first_reg = 14 + 32; first_reg <= 63; first_reg++)
12799 if (regs_ever_live[first_reg])
12800 break;
12801
12802 return first_reg;
12803 }
12804
12805 /* Similar, for AltiVec regs. */
12806
12807 static int
first_altivec_reg_to_save(void)12808 first_altivec_reg_to_save (void)
12809 {
12810 int i;
12811
12812 /* Stack frame remains as is unless we are in AltiVec ABI. */
12813 if (! TARGET_ALTIVEC_ABI)
12814 return LAST_ALTIVEC_REGNO + 1;
12815
12816 /* On Darwin, the unwind routines are compiled without
12817 TARGET_ALTIVEC, and use save_world to save/restore the
12818 altivec registers when necessary. */
12819 if (DEFAULT_ABI == ABI_DARWIN && current_function_calls_eh_return
12820 && ! TARGET_ALTIVEC)
12821 return FIRST_ALTIVEC_REGNO + 20;
12822
12823 /* Find lowest numbered live register. */
12824 for (i = FIRST_ALTIVEC_REGNO + 20; i <= LAST_ALTIVEC_REGNO; ++i)
12825 if (regs_ever_live[i])
12826 break;
12827
12828 return i;
12829 }
12830
12831 /* Return a 32-bit mask of the AltiVec registers we need to set in
12832 VRSAVE. Bit n of the return value is 1 if Vn is live. The MSB in
12833 the 32-bit word is 0. */
12834
12835 static unsigned int
compute_vrsave_mask(void)12836 compute_vrsave_mask (void)
12837 {
12838 unsigned int i, mask = 0;
12839
12840 /* On Darwin, the unwind routines are compiled without
12841 TARGET_ALTIVEC, and use save_world to save/restore the
12842 call-saved altivec registers when necessary. */
12843 if (DEFAULT_ABI == ABI_DARWIN && current_function_calls_eh_return
12844 && ! TARGET_ALTIVEC)
12845 mask |= 0xFFF;
12846
12847 /* First, find out if we use _any_ altivec registers. */
12848 for (i = FIRST_ALTIVEC_REGNO; i <= LAST_ALTIVEC_REGNO; ++i)
12849 if (regs_ever_live[i])
12850 mask |= ALTIVEC_REG_BIT (i);
12851
12852 if (mask == 0)
12853 return mask;
12854
12855 /* Next, remove the argument registers from the set. These must
12856 be in the VRSAVE mask set by the caller, so we don't need to add
12857 them in again. More importantly, the mask we compute here is
12858 used to generate CLOBBERs in the set_vrsave insn, and we do not
12859 wish the argument registers to die. */
12860 for (i = cfun->args_info.vregno - 1; i >= ALTIVEC_ARG_MIN_REG; --i)
12861 mask &= ~ALTIVEC_REG_BIT (i);
12862
12863 /* Similarly, remove the return value from the set. */
12864 {
12865 bool yes = false;
12866 diddle_return_value (is_altivec_return_reg, &yes);
12867 if (yes)
12868 mask &= ~ALTIVEC_REG_BIT (ALTIVEC_ARG_RETURN);
12869 }
12870
12871 return mask;
12872 }
12873
12874 /* For a very restricted set of circumstances, we can cut down the
12875 size of prologues/epilogues by calling our own save/restore-the-world
12876 routines. */
12877
12878 static void
compute_save_world_info(rs6000_stack_t * info_ptr)12879 compute_save_world_info (rs6000_stack_t *info_ptr)
12880 {
12881 info_ptr->world_save_p = 1;
12882 info_ptr->world_save_p
12883 = (WORLD_SAVE_P (info_ptr)
12884 && DEFAULT_ABI == ABI_DARWIN
12885 && ! (current_function_calls_setjmp && flag_exceptions)
12886 && info_ptr->first_fp_reg_save == FIRST_SAVED_FP_REGNO
12887 && info_ptr->first_gp_reg_save == FIRST_SAVED_GP_REGNO
12888 && info_ptr->first_altivec_reg_save == FIRST_SAVED_ALTIVEC_REGNO
12889 && info_ptr->cr_save_p);
12890
12891 /* This will not work in conjunction with sibcalls. Make sure there
12892 are none. (This check is expensive, but seldom executed.) */
12893 if (WORLD_SAVE_P (info_ptr))
12894 {
12895 rtx insn;
12896 for ( insn = get_last_insn_anywhere (); insn; insn = PREV_INSN (insn))
12897 if ( GET_CODE (insn) == CALL_INSN
12898 && SIBLING_CALL_P (insn))
12899 {
12900 info_ptr->world_save_p = 0;
12901 break;
12902 }
12903 }
12904
12905 if (WORLD_SAVE_P (info_ptr))
12906 {
12907 /* Even if we're not touching VRsave, make sure there's room on the
12908 stack for it, if it looks like we're calling SAVE_WORLD, which
12909 will attempt to save it. */
12910 info_ptr->vrsave_size = 4;
12911
12912 /* "Save" the VRsave register too if we're saving the world. */
12913 if (info_ptr->vrsave_mask == 0)
12914 info_ptr->vrsave_mask = compute_vrsave_mask ();
12915
12916 /* Because the Darwin register save/restore routines only handle
12917 F14 .. F31 and V20 .. V31 as per the ABI, perform a consistency
12918 check. */
12919 gcc_assert (info_ptr->first_fp_reg_save >= FIRST_SAVED_FP_REGNO
12920 && (info_ptr->first_altivec_reg_save
12921 >= FIRST_SAVED_ALTIVEC_REGNO));
12922 }
12923 return;
12924 }
12925
12926
12927 static void
is_altivec_return_reg(rtx reg,void * xyes)12928 is_altivec_return_reg (rtx reg, void *xyes)
12929 {
12930 bool *yes = (bool *) xyes;
12931 if (REGNO (reg) == ALTIVEC_ARG_RETURN)
12932 *yes = true;
12933 }
12934
12935
12936 /* Calculate the stack information for the current function. This is
12937 complicated by having two separate calling sequences, the AIX calling
12938 sequence and the V.4 calling sequence.
12939
12940 AIX (and Darwin/Mac OS X) stack frames look like:
12941 32-bit 64-bit
12942 SP----> +---------------------------------------+
12943 | back chain to caller | 0 0
12944 +---------------------------------------+
12945 | saved CR | 4 8 (8-11)
12946 +---------------------------------------+
12947 | saved LR | 8 16
12948 +---------------------------------------+
12949 | reserved for compilers | 12 24
12950 +---------------------------------------+
12951 | reserved for binders | 16 32
12952 +---------------------------------------+
12953 | saved TOC pointer | 20 40
12954 +---------------------------------------+
12955 | Parameter save area (P) | 24 48
12956 +---------------------------------------+
12957 | Alloca space (A) | 24+P etc.
12958 +---------------------------------------+
12959 | Local variable space (L) | 24+P+A
12960 +---------------------------------------+
12961 | Float/int conversion temporary (X) | 24+P+A+L
12962 +---------------------------------------+
12963 | Save area for AltiVec registers (W) | 24+P+A+L+X
12964 +---------------------------------------+
12965 | AltiVec alignment padding (Y) | 24+P+A+L+X+W
12966 +---------------------------------------+
12967 | Save area for VRSAVE register (Z) | 24+P+A+L+X+W+Y
12968 +---------------------------------------+
12969 | Save area for GP registers (G) | 24+P+A+X+L+X+W+Y+Z
12970 +---------------------------------------+
12971 | Save area for FP registers (F) | 24+P+A+X+L+X+W+Y+Z+G
12972 +---------------------------------------+
12973 old SP->| back chain to caller's caller |
12974 +---------------------------------------+
12975
12976 The required alignment for AIX configurations is two words (i.e., 8
12977 or 16 bytes).
12978
12979
12980 V.4 stack frames look like:
12981
12982 SP----> +---------------------------------------+
12983 | back chain to caller | 0
12984 +---------------------------------------+
12985 | caller's saved LR | 4
12986 +---------------------------------------+
12987 | Parameter save area (P) | 8
12988 +---------------------------------------+
12989 | Alloca space (A) | 8+P
12990 +---------------------------------------+
12991 | Varargs save area (V) | 8+P+A
12992 +---------------------------------------+
12993 | Local variable space (L) | 8+P+A+V
12994 +---------------------------------------+
12995 | Float/int conversion temporary (X) | 8+P+A+V+L
12996 +---------------------------------------+
12997 | Save area for AltiVec registers (W) | 8+P+A+V+L+X
12998 +---------------------------------------+
12999 | AltiVec alignment padding (Y) | 8+P+A+V+L+X+W
13000 +---------------------------------------+
13001 | Save area for VRSAVE register (Z) | 8+P+A+V+L+X+W+Y
13002 +---------------------------------------+
13003 | SPE: area for 64-bit GP registers |
13004 +---------------------------------------+
13005 | SPE alignment padding |
13006 +---------------------------------------+
13007 | saved CR (C) | 8+P+A+V+L+X+W+Y+Z
13008 +---------------------------------------+
13009 | Save area for GP registers (G) | 8+P+A+V+L+X+W+Y+Z+C
13010 +---------------------------------------+
13011 | Save area for FP registers (F) | 8+P+A+V+L+X+W+Y+Z+C+G
13012 +---------------------------------------+
13013 old SP->| back chain to caller's caller |
13014 +---------------------------------------+
13015
13016 The required alignment for V.4 is 16 bytes, or 8 bytes if -meabi is
13017 given. (But note below and in sysv4.h that we require only 8 and
13018 may round up the size of our stack frame anyways. The historical
13019 reason is early versions of powerpc-linux which didn't properly
13020 align the stack at program startup. A happy side-effect is that
13021 -mno-eabi libraries can be used with -meabi programs.)
13022
13023 The EABI configuration defaults to the V.4 layout. However,
13024 the stack alignment requirements may differ. If -mno-eabi is not
13025 given, the required stack alignment is 8 bytes; if -mno-eabi is
13026 given, the required alignment is 16 bytes. (But see V.4 comment
13027 above.) */
13028
13029 #ifndef ABI_STACK_BOUNDARY
13030 #define ABI_STACK_BOUNDARY STACK_BOUNDARY
13031 #endif
13032
13033 static rs6000_stack_t *
rs6000_stack_info(void)13034 rs6000_stack_info (void)
13035 {
13036 static rs6000_stack_t info;
13037 rs6000_stack_t *info_ptr = &info;
13038 int reg_size = TARGET_32BIT ? 4 : 8;
13039 int ehrd_size;
13040 int save_align;
13041 HOST_WIDE_INT non_fixed_size;
13042
13043 memset (&info, 0, sizeof (info));
13044
13045 if (TARGET_SPE)
13046 {
13047 /* Cache value so we don't rescan instruction chain over and over. */
13048 if (cfun->machine->insn_chain_scanned_p == 0)
13049 cfun->machine->insn_chain_scanned_p
13050 = spe_func_has_64bit_regs_p () + 1;
13051 info_ptr->spe_64bit_regs_used = cfun->machine->insn_chain_scanned_p - 1;
13052 }
13053
13054 /* Select which calling sequence. */
13055 info_ptr->abi = DEFAULT_ABI;
13056
13057 /* Calculate which registers need to be saved & save area size. */
13058 info_ptr->first_gp_reg_save = first_reg_to_save ();
13059 /* Assume that we will have to save RS6000_PIC_OFFSET_TABLE_REGNUM,
13060 even if it currently looks like we won't. */
13061 if (((TARGET_TOC && TARGET_MINIMAL_TOC)
13062 || (flag_pic == 1 && DEFAULT_ABI == ABI_V4)
13063 || (flag_pic && DEFAULT_ABI == ABI_DARWIN))
13064 && info_ptr->first_gp_reg_save > RS6000_PIC_OFFSET_TABLE_REGNUM)
13065 info_ptr->gp_size = reg_size * (32 - RS6000_PIC_OFFSET_TABLE_REGNUM);
13066 else
13067 info_ptr->gp_size = reg_size * (32 - info_ptr->first_gp_reg_save);
13068
13069 /* For the SPE, we have an additional upper 32-bits on each GPR.
13070 Ideally we should save the entire 64-bits only when the upper
13071 half is used in SIMD instructions. Since we only record
13072 registers live (not the size they are used in), this proves
13073 difficult because we'd have to traverse the instruction chain at
13074 the right time, taking reload into account. This is a real pain,
13075 so we opt to save the GPRs in 64-bits always if but one register
13076 gets used in 64-bits. Otherwise, all the registers in the frame
13077 get saved in 32-bits.
13078
13079 So... since when we save all GPRs (except the SP) in 64-bits, the
13080 traditional GP save area will be empty. */
13081 if (TARGET_SPE_ABI && info_ptr->spe_64bit_regs_used != 0)
13082 info_ptr->gp_size = 0;
13083
13084 info_ptr->first_fp_reg_save = first_fp_reg_to_save ();
13085 info_ptr->fp_size = 8 * (64 - info_ptr->first_fp_reg_save);
13086
13087 info_ptr->first_altivec_reg_save = first_altivec_reg_to_save ();
13088 info_ptr->altivec_size = 16 * (LAST_ALTIVEC_REGNO + 1
13089 - info_ptr->first_altivec_reg_save);
13090
13091 /* Does this function call anything? */
13092 info_ptr->calls_p = (! current_function_is_leaf
13093 || cfun->machine->ra_needs_full_frame);
13094
13095 /* Determine if we need to save the link register. */
13096 if ((DEFAULT_ABI == ABI_AIX
13097 && current_function_profile
13098 && !TARGET_PROFILE_KERNEL)
13099 #ifdef TARGET_RELOCATABLE
13100 || (TARGET_RELOCATABLE && (get_pool_size () != 0))
13101 #endif
13102 || (info_ptr->first_fp_reg_save != 64
13103 && !FP_SAVE_INLINE (info_ptr->first_fp_reg_save))
13104 || info_ptr->first_altivec_reg_save <= LAST_ALTIVEC_REGNO
13105 || (DEFAULT_ABI == ABI_V4 && current_function_calls_alloca)
13106 || info_ptr->calls_p
13107 || rs6000_ra_ever_killed ())
13108 {
13109 info_ptr->lr_save_p = 1;
13110 regs_ever_live[LINK_REGISTER_REGNUM] = 1;
13111 }
13112
13113 /* Determine if we need to save the condition code registers. */
13114 if (regs_ever_live[CR2_REGNO]
13115 || regs_ever_live[CR3_REGNO]
13116 || regs_ever_live[CR4_REGNO])
13117 {
13118 info_ptr->cr_save_p = 1;
13119 if (DEFAULT_ABI == ABI_V4)
13120 info_ptr->cr_size = reg_size;
13121 }
13122
13123 /* If the current function calls __builtin_eh_return, then we need
13124 to allocate stack space for registers that will hold data for
13125 the exception handler. */
13126 if (current_function_calls_eh_return)
13127 {
13128 unsigned int i;
13129 for (i = 0; EH_RETURN_DATA_REGNO (i) != INVALID_REGNUM; ++i)
13130 continue;
13131
13132 /* SPE saves EH registers in 64-bits. */
13133 ehrd_size = i * (TARGET_SPE_ABI
13134 && info_ptr->spe_64bit_regs_used != 0
13135 ? UNITS_PER_SPE_WORD : UNITS_PER_WORD);
13136 }
13137 else
13138 ehrd_size = 0;
13139
13140 /* Determine various sizes. */
13141 info_ptr->reg_size = reg_size;
13142 info_ptr->fixed_size = RS6000_SAVE_AREA;
13143 info_ptr->vars_size = RS6000_ALIGN (get_frame_size (), 8);
13144 info_ptr->parm_size = RS6000_ALIGN (current_function_outgoing_args_size,
13145 TARGET_ALTIVEC ? 16 : 8);
13146 if (FRAME_GROWS_DOWNWARD)
13147 info_ptr->vars_size
13148 += RS6000_ALIGN (info_ptr->fixed_size + info_ptr->vars_size
13149 + info_ptr->parm_size,
13150 ABI_STACK_BOUNDARY / BITS_PER_UNIT)
13151 - (info_ptr->fixed_size + info_ptr->vars_size
13152 + info_ptr->parm_size);
13153
13154 if (TARGET_SPE_ABI && info_ptr->spe_64bit_regs_used != 0)
13155 info_ptr->spe_gp_size = 8 * (32 - info_ptr->first_gp_reg_save);
13156 else
13157 info_ptr->spe_gp_size = 0;
13158
13159 if (TARGET_ALTIVEC_ABI)
13160 info_ptr->vrsave_mask = compute_vrsave_mask ();
13161 else
13162 info_ptr->vrsave_mask = 0;
13163
13164 if (TARGET_ALTIVEC_VRSAVE && info_ptr->vrsave_mask)
13165 info_ptr->vrsave_size = 4;
13166 else
13167 info_ptr->vrsave_size = 0;
13168
13169 compute_save_world_info (info_ptr);
13170
13171 /* Calculate the offsets. */
13172 switch (DEFAULT_ABI)
13173 {
13174 case ABI_NONE:
13175 default:
13176 gcc_unreachable ();
13177
13178 case ABI_AIX:
13179 case ABI_DARWIN:
13180 info_ptr->fp_save_offset = - info_ptr->fp_size;
13181 info_ptr->gp_save_offset = info_ptr->fp_save_offset - info_ptr->gp_size;
13182
13183 if (TARGET_ALTIVEC_ABI)
13184 {
13185 info_ptr->vrsave_save_offset
13186 = info_ptr->gp_save_offset - info_ptr->vrsave_size;
13187
13188 /* Align stack so vector save area is on a quadword boundary.
13189 The padding goes above the vectors. */
13190 if (info_ptr->altivec_size != 0)
13191 info_ptr->altivec_padding_size
13192 = info_ptr->vrsave_save_offset & 0xF;
13193 else
13194 info_ptr->altivec_padding_size = 0;
13195
13196 info_ptr->altivec_save_offset
13197 = info_ptr->vrsave_save_offset
13198 - info_ptr->altivec_padding_size
13199 - info_ptr->altivec_size;
13200 gcc_assert (info_ptr->altivec_size == 0
13201 || info_ptr->altivec_save_offset % 16 == 0);
13202
13203 /* Adjust for AltiVec case. */
13204 info_ptr->ehrd_offset = info_ptr->altivec_save_offset - ehrd_size;
13205 }
13206 else
13207 info_ptr->ehrd_offset = info_ptr->gp_save_offset - ehrd_size;
13208 info_ptr->cr_save_offset = reg_size; /* first word when 64-bit. */
13209 info_ptr->lr_save_offset = 2*reg_size;
13210 break;
13211
13212 case ABI_V4:
13213 info_ptr->fp_save_offset = - info_ptr->fp_size;
13214 info_ptr->gp_save_offset = info_ptr->fp_save_offset - info_ptr->gp_size;
13215 info_ptr->cr_save_offset = info_ptr->gp_save_offset - info_ptr->cr_size;
13216
13217 if (TARGET_SPE_ABI && info_ptr->spe_64bit_regs_used != 0)
13218 {
13219 /* Align stack so SPE GPR save area is aligned on a
13220 double-word boundary. */
13221 if (info_ptr->spe_gp_size != 0)
13222 info_ptr->spe_padding_size
13223 = 8 - (-info_ptr->cr_save_offset % 8);
13224 else
13225 info_ptr->spe_padding_size = 0;
13226
13227 info_ptr->spe_gp_save_offset
13228 = info_ptr->cr_save_offset
13229 - info_ptr->spe_padding_size
13230 - info_ptr->spe_gp_size;
13231
13232 /* Adjust for SPE case. */
13233 info_ptr->ehrd_offset = info_ptr->spe_gp_save_offset;
13234 }
13235 else if (TARGET_ALTIVEC_ABI)
13236 {
13237 info_ptr->vrsave_save_offset
13238 = info_ptr->cr_save_offset - info_ptr->vrsave_size;
13239
13240 /* Align stack so vector save area is on a quadword boundary. */
13241 if (info_ptr->altivec_size != 0)
13242 info_ptr->altivec_padding_size
13243 = 16 - (-info_ptr->vrsave_save_offset % 16);
13244 else
13245 info_ptr->altivec_padding_size = 0;
13246
13247 info_ptr->altivec_save_offset
13248 = info_ptr->vrsave_save_offset
13249 - info_ptr->altivec_padding_size
13250 - info_ptr->altivec_size;
13251
13252 /* Adjust for AltiVec case. */
13253 info_ptr->ehrd_offset = info_ptr->altivec_save_offset;
13254 }
13255 else
13256 info_ptr->ehrd_offset = info_ptr->cr_save_offset;
13257 info_ptr->ehrd_offset -= ehrd_size;
13258 info_ptr->lr_save_offset = reg_size;
13259 break;
13260 }
13261
13262 save_align = (TARGET_ALTIVEC_ABI || DEFAULT_ABI == ABI_DARWIN) ? 16 : 8;
13263 info_ptr->save_size = RS6000_ALIGN (info_ptr->fp_size
13264 + info_ptr->gp_size
13265 + info_ptr->altivec_size
13266 + info_ptr->altivec_padding_size
13267 + info_ptr->spe_gp_size
13268 + info_ptr->spe_padding_size
13269 + ehrd_size
13270 + info_ptr->cr_size
13271 + info_ptr->vrsave_size,
13272 save_align);
13273
13274 non_fixed_size = (info_ptr->vars_size
13275 + info_ptr->parm_size
13276 + info_ptr->save_size);
13277
13278 info_ptr->total_size = RS6000_ALIGN (non_fixed_size + info_ptr->fixed_size,
13279 ABI_STACK_BOUNDARY / BITS_PER_UNIT);
13280
13281 /* Determine if we need to allocate any stack frame:
13282
13283 For AIX we need to push the stack if a frame pointer is needed
13284 (because the stack might be dynamically adjusted), if we are
13285 debugging, if we make calls, or if the sum of fp_save, gp_save,
13286 and local variables are more than the space needed to save all
13287 non-volatile registers: 32-bit: 18*8 + 19*4 = 220 or 64-bit: 18*8
13288 + 18*8 = 288 (GPR13 reserved).
13289
13290 For V.4 we don't have the stack cushion that AIX uses, but assume
13291 that the debugger can handle stackless frames. */
13292
13293 if (info_ptr->calls_p)
13294 info_ptr->push_p = 1;
13295
13296 else if (DEFAULT_ABI == ABI_V4)
13297 info_ptr->push_p = non_fixed_size != 0;
13298
13299 else if (frame_pointer_needed)
13300 info_ptr->push_p = 1;
13301
13302 else if (TARGET_XCOFF && write_symbols != NO_DEBUG)
13303 info_ptr->push_p = 1;
13304
13305 else
13306 info_ptr->push_p = non_fixed_size > (TARGET_32BIT ? 220 : 288);
13307
13308 /* Zero offsets if we're not saving those registers. */
13309 if (info_ptr->fp_size == 0)
13310 info_ptr->fp_save_offset = 0;
13311
13312 if (info_ptr->gp_size == 0)
13313 info_ptr->gp_save_offset = 0;
13314
13315 if (! TARGET_ALTIVEC_ABI || info_ptr->altivec_size == 0)
13316 info_ptr->altivec_save_offset = 0;
13317
13318 if (! TARGET_ALTIVEC_ABI || info_ptr->vrsave_mask == 0)
13319 info_ptr->vrsave_save_offset = 0;
13320
13321 if (! TARGET_SPE_ABI
13322 || info_ptr->spe_64bit_regs_used == 0
13323 || info_ptr->spe_gp_size == 0)
13324 info_ptr->spe_gp_save_offset = 0;
13325
13326 if (! info_ptr->lr_save_p)
13327 info_ptr->lr_save_offset = 0;
13328
13329 if (! info_ptr->cr_save_p)
13330 info_ptr->cr_save_offset = 0;
13331
13332 return info_ptr;
13333 }
13334
13335 /* Return true if the current function uses any GPRs in 64-bit SIMD
13336 mode. */
13337
13338 static bool
spe_func_has_64bit_regs_p(void)13339 spe_func_has_64bit_regs_p (void)
13340 {
13341 rtx insns, insn;
13342
13343 /* Functions that save and restore all the call-saved registers will
13344 need to save/restore the registers in 64-bits. */
13345 if (current_function_calls_eh_return
13346 || current_function_calls_setjmp
13347 || current_function_has_nonlocal_goto)
13348 return true;
13349
13350 insns = get_insns ();
13351
13352 for (insn = NEXT_INSN (insns); insn != NULL_RTX; insn = NEXT_INSN (insn))
13353 {
13354 if (INSN_P (insn))
13355 {
13356 rtx i;
13357
13358 /* FIXME: This should be implemented with attributes...
13359
13360 (set_attr "spe64" "true")....then,
13361 if (get_spe64(insn)) return true;
13362
13363 It's the only reliable way to do the stuff below. */
13364
13365 i = PATTERN (insn);
13366 if (GET_CODE (i) == SET)
13367 {
13368 enum machine_mode mode = GET_MODE (SET_SRC (i));
13369
13370 if (SPE_VECTOR_MODE (mode))
13371 return true;
13372 if (TARGET_E500_DOUBLE && mode == DFmode)
13373 return true;
13374 }
13375 }
13376 }
13377
13378 return false;
13379 }
13380
13381 static void
debug_stack_info(rs6000_stack_t * info)13382 debug_stack_info (rs6000_stack_t *info)
13383 {
13384 const char *abi_string;
13385
13386 if (! info)
13387 info = rs6000_stack_info ();
13388
13389 fprintf (stderr, "\nStack information for function %s:\n",
13390 ((current_function_decl && DECL_NAME (current_function_decl))
13391 ? IDENTIFIER_POINTER (DECL_NAME (current_function_decl))
13392 : "<unknown>"));
13393
13394 switch (info->abi)
13395 {
13396 default: abi_string = "Unknown"; break;
13397 case ABI_NONE: abi_string = "NONE"; break;
13398 case ABI_AIX: abi_string = "AIX"; break;
13399 case ABI_DARWIN: abi_string = "Darwin"; break;
13400 case ABI_V4: abi_string = "V.4"; break;
13401 }
13402
13403 fprintf (stderr, "\tABI = %5s\n", abi_string);
13404
13405 if (TARGET_ALTIVEC_ABI)
13406 fprintf (stderr, "\tALTIVEC ABI extensions enabled.\n");
13407
13408 if (TARGET_SPE_ABI)
13409 fprintf (stderr, "\tSPE ABI extensions enabled.\n");
13410
13411 if (info->first_gp_reg_save != 32)
13412 fprintf (stderr, "\tfirst_gp_reg_save = %5d\n", info->first_gp_reg_save);
13413
13414 if (info->first_fp_reg_save != 64)
13415 fprintf (stderr, "\tfirst_fp_reg_save = %5d\n", info->first_fp_reg_save);
13416
13417 if (info->first_altivec_reg_save <= LAST_ALTIVEC_REGNO)
13418 fprintf (stderr, "\tfirst_altivec_reg_save = %5d\n",
13419 info->first_altivec_reg_save);
13420
13421 if (info->lr_save_p)
13422 fprintf (stderr, "\tlr_save_p = %5d\n", info->lr_save_p);
13423
13424 if (info->cr_save_p)
13425 fprintf (stderr, "\tcr_save_p = %5d\n", info->cr_save_p);
13426
13427 if (info->vrsave_mask)
13428 fprintf (stderr, "\tvrsave_mask = 0x%x\n", info->vrsave_mask);
13429
13430 if (info->push_p)
13431 fprintf (stderr, "\tpush_p = %5d\n", info->push_p);
13432
13433 if (info->calls_p)
13434 fprintf (stderr, "\tcalls_p = %5d\n", info->calls_p);
13435
13436 if (info->gp_save_offset)
13437 fprintf (stderr, "\tgp_save_offset = %5d\n", info->gp_save_offset);
13438
13439 if (info->fp_save_offset)
13440 fprintf (stderr, "\tfp_save_offset = %5d\n", info->fp_save_offset);
13441
13442 if (info->altivec_save_offset)
13443 fprintf (stderr, "\taltivec_save_offset = %5d\n",
13444 info->altivec_save_offset);
13445
13446 if (info->spe_gp_save_offset)
13447 fprintf (stderr, "\tspe_gp_save_offset = %5d\n",
13448 info->spe_gp_save_offset);
13449
13450 if (info->vrsave_save_offset)
13451 fprintf (stderr, "\tvrsave_save_offset = %5d\n",
13452 info->vrsave_save_offset);
13453
13454 if (info->lr_save_offset)
13455 fprintf (stderr, "\tlr_save_offset = %5d\n", info->lr_save_offset);
13456
13457 if (info->cr_save_offset)
13458 fprintf (stderr, "\tcr_save_offset = %5d\n", info->cr_save_offset);
13459
13460 if (info->varargs_save_offset)
13461 fprintf (stderr, "\tvarargs_save_offset = %5d\n", info->varargs_save_offset);
13462
13463 if (info->total_size)
13464 fprintf (stderr, "\ttotal_size = "HOST_WIDE_INT_PRINT_DEC"\n",
13465 info->total_size);
13466
13467 if (info->vars_size)
13468 fprintf (stderr, "\tvars_size = "HOST_WIDE_INT_PRINT_DEC"\n",
13469 info->vars_size);
13470
13471 if (info->parm_size)
13472 fprintf (stderr, "\tparm_size = %5d\n", info->parm_size);
13473
13474 if (info->fixed_size)
13475 fprintf (stderr, "\tfixed_size = %5d\n", info->fixed_size);
13476
13477 if (info->gp_size)
13478 fprintf (stderr, "\tgp_size = %5d\n", info->gp_size);
13479
13480 if (info->spe_gp_size)
13481 fprintf (stderr, "\tspe_gp_size = %5d\n", info->spe_gp_size);
13482
13483 if (info->fp_size)
13484 fprintf (stderr, "\tfp_size = %5d\n", info->fp_size);
13485
13486 if (info->altivec_size)
13487 fprintf (stderr, "\taltivec_size = %5d\n", info->altivec_size);
13488
13489 if (info->vrsave_size)
13490 fprintf (stderr, "\tvrsave_size = %5d\n", info->vrsave_size);
13491
13492 if (info->altivec_padding_size)
13493 fprintf (stderr, "\taltivec_padding_size= %5d\n",
13494 info->altivec_padding_size);
13495
13496 if (info->spe_padding_size)
13497 fprintf (stderr, "\tspe_padding_size = %5d\n",
13498 info->spe_padding_size);
13499
13500 if (info->cr_size)
13501 fprintf (stderr, "\tcr_size = %5d\n", info->cr_size);
13502
13503 if (info->save_size)
13504 fprintf (stderr, "\tsave_size = %5d\n", info->save_size);
13505
13506 if (info->reg_size != 4)
13507 fprintf (stderr, "\treg_size = %5d\n", info->reg_size);
13508
13509 fprintf (stderr, "\n");
13510 }
13511
13512 rtx
rs6000_return_addr(int count,rtx frame)13513 rs6000_return_addr (int count, rtx frame)
13514 {
13515 /* Currently we don't optimize very well between prolog and body
13516 code and for PIC code the code can be actually quite bad, so
13517 don't try to be too clever here. */
13518 if (count != 0 || (DEFAULT_ABI != ABI_AIX && flag_pic))
13519 {
13520 cfun->machine->ra_needs_full_frame = 1;
13521
13522 return
13523 gen_rtx_MEM
13524 (Pmode,
13525 memory_address
13526 (Pmode,
13527 plus_constant (copy_to_reg
13528 (gen_rtx_MEM (Pmode,
13529 memory_address (Pmode, frame))),
13530 RETURN_ADDRESS_OFFSET)));
13531 }
13532
13533 cfun->machine->ra_need_lr = 1;
13534 return get_hard_reg_initial_val (Pmode, LINK_REGISTER_REGNUM);
13535 }
13536
13537 /* Say whether a function is a candidate for sibcall handling or not.
13538 We do not allow indirect calls to be optimized into sibling calls.
13539 Also, we can't do it if there are any vector parameters; there's
13540 nowhere to put the VRsave code so it works; note that functions with
13541 vector parameters are required to have a prototype, so the argument
13542 type info must be available here. (The tail recursion case can work
13543 with vector parameters, but there's no way to distinguish here.) */
13544 static bool
rs6000_function_ok_for_sibcall(tree decl,tree exp ATTRIBUTE_UNUSED)13545 rs6000_function_ok_for_sibcall (tree decl, tree exp ATTRIBUTE_UNUSED)
13546 {
13547 tree type;
13548 if (decl)
13549 {
13550 if (TARGET_ALTIVEC_VRSAVE)
13551 {
13552 for (type = TYPE_ARG_TYPES (TREE_TYPE (decl));
13553 type; type = TREE_CHAIN (type))
13554 {
13555 if (TREE_CODE (TREE_VALUE (type)) == VECTOR_TYPE)
13556 return false;
13557 }
13558 }
13559 if (DEFAULT_ABI == ABI_DARWIN
13560 || ((*targetm.binds_local_p) (decl)
13561 && (DEFAULT_ABI != ABI_AIX || !DECL_EXTERNAL (decl))))
13562 {
13563 tree attr_list = TYPE_ATTRIBUTES (TREE_TYPE (decl));
13564
13565 if (!lookup_attribute ("longcall", attr_list)
13566 || lookup_attribute ("shortcall", attr_list))
13567 return true;
13568 }
13569 }
13570 return false;
13571 }
13572
13573 /* NULL if INSN insn is valid within a low-overhead loop.
13574 Otherwise return why doloop cannot be applied.
13575 PowerPC uses the COUNT register for branch on table instructions. */
13576
13577 static const char *
rs6000_invalid_within_doloop(rtx insn)13578 rs6000_invalid_within_doloop (rtx insn)
13579 {
13580 if (CALL_P (insn))
13581 return "Function call in the loop.";
13582
13583 if (JUMP_P (insn)
13584 && (GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC
13585 || GET_CODE (PATTERN (insn)) == ADDR_VEC))
13586 return "Computed branch in the loop.";
13587
13588 return NULL;
13589 }
13590
13591 static int
rs6000_ra_ever_killed(void)13592 rs6000_ra_ever_killed (void)
13593 {
13594 rtx top;
13595 rtx reg;
13596 rtx insn;
13597
13598 if (current_function_is_thunk)
13599 return 0;
13600
13601 /* regs_ever_live has LR marked as used if any sibcalls are present,
13602 but this should not force saving and restoring in the
13603 pro/epilogue. Likewise, reg_set_between_p thinks a sibcall
13604 clobbers LR, so that is inappropriate. */
13605
13606 /* Also, the prologue can generate a store into LR that
13607 doesn't really count, like this:
13608
13609 move LR->R0
13610 bcl to set PIC register
13611 move LR->R31
13612 move R0->LR
13613
13614 When we're called from the epilogue, we need to avoid counting
13615 this as a store. */
13616
13617 push_topmost_sequence ();
13618 top = get_insns ();
13619 pop_topmost_sequence ();
13620 reg = gen_rtx_REG (Pmode, LINK_REGISTER_REGNUM);
13621
13622 for (insn = NEXT_INSN (top); insn != NULL_RTX; insn = NEXT_INSN (insn))
13623 {
13624 if (INSN_P (insn))
13625 {
13626 if (CALL_P (insn))
13627 {
13628 if (!SIBLING_CALL_P (insn))
13629 return 1;
13630 }
13631 else if (find_regno_note (insn, REG_INC, LINK_REGISTER_REGNUM))
13632 return 1;
13633 else if (set_of (reg, insn) != NULL_RTX
13634 && !prologue_epilogue_contains (insn))
13635 return 1;
13636 }
13637 }
13638 return 0;
13639 }
13640
13641 /* Add a REG_MAYBE_DEAD note to the insn. */
13642 static void
rs6000_maybe_dead(rtx insn)13643 rs6000_maybe_dead (rtx insn)
13644 {
13645 REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_MAYBE_DEAD,
13646 const0_rtx,
13647 REG_NOTES (insn));
13648 }
13649
13650 /* Emit instructions needed to load the TOC register.
13651 This is only needed when TARGET_TOC, TARGET_MINIMAL_TOC, and there is
13652 a constant pool; or for SVR4 -fpic. */
13653
13654 void
rs6000_emit_load_toc_table(int fromprolog)13655 rs6000_emit_load_toc_table (int fromprolog)
13656 {
13657 rtx dest, insn;
13658 dest = gen_rtx_REG (Pmode, RS6000_PIC_OFFSET_TABLE_REGNUM);
13659
13660 if (TARGET_ELF && TARGET_SECURE_PLT && DEFAULT_ABI != ABI_AIX && flag_pic)
13661 {
13662 char buf[30];
13663 rtx lab, tmp1, tmp2, got, tempLR;
13664
13665 ASM_GENERATE_INTERNAL_LABEL (buf, "LCF", rs6000_pic_labelno);
13666 lab = gen_rtx_SYMBOL_REF (Pmode, ggc_strdup (buf));
13667 if (flag_pic == 2)
13668 got = gen_rtx_SYMBOL_REF (Pmode, toc_label_name);
13669 else
13670 got = rs6000_got_sym ();
13671 tmp1 = tmp2 = dest;
13672 if (!fromprolog)
13673 {
13674 tmp1 = gen_reg_rtx (Pmode);
13675 tmp2 = gen_reg_rtx (Pmode);
13676 }
13677 tempLR = (fromprolog
13678 ? gen_rtx_REG (Pmode, LINK_REGISTER_REGNUM)
13679 : gen_reg_rtx (Pmode));
13680 insn = emit_insn (gen_load_toc_v4_PIC_1 (tempLR, lab));
13681 if (fromprolog)
13682 rs6000_maybe_dead (insn);
13683 insn = emit_move_insn (tmp1, tempLR);
13684 if (fromprolog)
13685 rs6000_maybe_dead (insn);
13686 insn = emit_insn (gen_load_toc_v4_PIC_3b (tmp2, tmp1, got, lab));
13687 if (fromprolog)
13688 rs6000_maybe_dead (insn);
13689 insn = emit_insn (gen_load_toc_v4_PIC_3c (dest, tmp2, got, lab));
13690 if (fromprolog)
13691 rs6000_maybe_dead (insn);
13692 }
13693 else if (TARGET_ELF && DEFAULT_ABI == ABI_V4 && flag_pic == 1)
13694 {
13695 rtx tempLR = (fromprolog
13696 ? gen_rtx_REG (Pmode, LINK_REGISTER_REGNUM)
13697 : gen_reg_rtx (Pmode));
13698
13699 insn = emit_insn (gen_load_toc_v4_pic_si (tempLR));
13700 if (fromprolog)
13701 rs6000_maybe_dead (insn);
13702 insn = emit_move_insn (dest, tempLR);
13703 if (fromprolog)
13704 rs6000_maybe_dead (insn);
13705 }
13706 else if (TARGET_ELF && DEFAULT_ABI != ABI_AIX && flag_pic == 2)
13707 {
13708 char buf[30];
13709 rtx tempLR = (fromprolog
13710 ? gen_rtx_REG (Pmode, LINK_REGISTER_REGNUM)
13711 : gen_reg_rtx (Pmode));
13712 rtx temp0 = (fromprolog
13713 ? gen_rtx_REG (Pmode, 0)
13714 : gen_reg_rtx (Pmode));
13715
13716 if (fromprolog)
13717 {
13718 rtx symF, symL;
13719
13720 ASM_GENERATE_INTERNAL_LABEL (buf, "LCF", rs6000_pic_labelno);
13721 symF = gen_rtx_SYMBOL_REF (Pmode, ggc_strdup (buf));
13722
13723 ASM_GENERATE_INTERNAL_LABEL (buf, "LCL", rs6000_pic_labelno);
13724 symL = gen_rtx_SYMBOL_REF (Pmode, ggc_strdup (buf));
13725
13726 rs6000_maybe_dead (emit_insn (gen_load_toc_v4_PIC_1 (tempLR,
13727 symF)));
13728 rs6000_maybe_dead (emit_move_insn (dest, tempLR));
13729 rs6000_maybe_dead (emit_insn (gen_load_toc_v4_PIC_2 (temp0, dest,
13730 symL,
13731 symF)));
13732 }
13733 else
13734 {
13735 rtx tocsym;
13736
13737 tocsym = gen_rtx_SYMBOL_REF (Pmode, toc_label_name);
13738 emit_insn (gen_load_toc_v4_PIC_1b (tempLR, tocsym));
13739 emit_move_insn (dest, tempLR);
13740 emit_move_insn (temp0, gen_rtx_MEM (Pmode, dest));
13741 }
13742 insn = emit_insn (gen_addsi3 (dest, temp0, dest));
13743 if (fromprolog)
13744 rs6000_maybe_dead (insn);
13745 }
13746 else if (TARGET_ELF && !TARGET_AIX && flag_pic == 0 && TARGET_MINIMAL_TOC)
13747 {
13748 /* This is for AIX code running in non-PIC ELF32. */
13749 char buf[30];
13750 rtx realsym;
13751 ASM_GENERATE_INTERNAL_LABEL (buf, "LCTOC", 1);
13752 realsym = gen_rtx_SYMBOL_REF (Pmode, ggc_strdup (buf));
13753
13754 insn = emit_insn (gen_elf_high (dest, realsym));
13755 if (fromprolog)
13756 rs6000_maybe_dead (insn);
13757 insn = emit_insn (gen_elf_low (dest, dest, realsym));
13758 if (fromprolog)
13759 rs6000_maybe_dead (insn);
13760 }
13761 else
13762 {
13763 gcc_assert (DEFAULT_ABI == ABI_AIX);
13764
13765 if (TARGET_32BIT)
13766 insn = emit_insn (gen_load_toc_aix_si (dest));
13767 else
13768 insn = emit_insn (gen_load_toc_aix_di (dest));
13769 if (fromprolog)
13770 rs6000_maybe_dead (insn);
13771 }
13772 }
13773
13774 /* Emit instructions to restore the link register after determining where
13775 its value has been stored. */
13776
13777 void
rs6000_emit_eh_reg_restore(rtx source,rtx scratch)13778 rs6000_emit_eh_reg_restore (rtx source, rtx scratch)
13779 {
13780 rs6000_stack_t *info = rs6000_stack_info ();
13781 rtx operands[2];
13782
13783 operands[0] = source;
13784 operands[1] = scratch;
13785
13786 if (info->lr_save_p)
13787 {
13788 rtx frame_rtx = stack_pointer_rtx;
13789 HOST_WIDE_INT sp_offset = 0;
13790 rtx tmp;
13791
13792 if (frame_pointer_needed
13793 || current_function_calls_alloca
13794 || info->total_size > 32767)
13795 {
13796 tmp = gen_frame_mem (Pmode, frame_rtx);
13797 emit_move_insn (operands[1], tmp);
13798 frame_rtx = operands[1];
13799 }
13800 else if (info->push_p)
13801 sp_offset = info->total_size;
13802
13803 tmp = plus_constant (frame_rtx, info->lr_save_offset + sp_offset);
13804 tmp = gen_frame_mem (Pmode, tmp);
13805 emit_move_insn (tmp, operands[0]);
13806 }
13807 else
13808 emit_move_insn (gen_rtx_REG (Pmode, LINK_REGISTER_REGNUM), operands[0]);
13809 }
13810
13811 static GTY(()) int set = -1;
13812
13813 int
get_TOC_alias_set(void)13814 get_TOC_alias_set (void)
13815 {
13816 if (set == -1)
13817 set = new_alias_set ();
13818 return set;
13819 }
13820
13821 /* This returns nonzero if the current function uses the TOC. This is
13822 determined by the presence of (use (unspec ... UNSPEC_TOC)), which
13823 is generated by the ABI_V4 load_toc_* patterns. */
13824 #if TARGET_ELF
13825 static int
uses_TOC(void)13826 uses_TOC (void)
13827 {
13828 rtx insn;
13829
13830 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
13831 if (INSN_P (insn))
13832 {
13833 rtx pat = PATTERN (insn);
13834 int i;
13835
13836 if (GET_CODE (pat) == PARALLEL)
13837 for (i = 0; i < XVECLEN (pat, 0); i++)
13838 {
13839 rtx sub = XVECEXP (pat, 0, i);
13840 if (GET_CODE (sub) == USE)
13841 {
13842 sub = XEXP (sub, 0);
13843 if (GET_CODE (sub) == UNSPEC
13844 && XINT (sub, 1) == UNSPEC_TOC)
13845 return 1;
13846 }
13847 }
13848 }
13849 return 0;
13850 }
13851 #endif
13852
13853 rtx
create_TOC_reference(rtx symbol)13854 create_TOC_reference (rtx symbol)
13855 {
13856 if (no_new_pseudos)
13857 regs_ever_live[TOC_REGISTER] = 1;
13858 return gen_rtx_PLUS (Pmode,
13859 gen_rtx_REG (Pmode, TOC_REGISTER),
13860 gen_rtx_CONST (Pmode,
13861 gen_rtx_MINUS (Pmode, symbol,
13862 gen_rtx_SYMBOL_REF (Pmode, toc_label_name))));
13863 }
13864
13865 /* If _Unwind_* has been called from within the same module,
13866 toc register is not guaranteed to be saved to 40(1) on function
13867 entry. Save it there in that case. */
13868
13869 void
rs6000_aix_emit_builtin_unwind_init(void)13870 rs6000_aix_emit_builtin_unwind_init (void)
13871 {
13872 rtx mem;
13873 rtx stack_top = gen_reg_rtx (Pmode);
13874 rtx opcode_addr = gen_reg_rtx (Pmode);
13875 rtx opcode = gen_reg_rtx (SImode);
13876 rtx tocompare = gen_reg_rtx (SImode);
13877 rtx no_toc_save_needed = gen_label_rtx ();
13878
13879 mem = gen_frame_mem (Pmode, hard_frame_pointer_rtx);
13880 emit_move_insn (stack_top, mem);
13881
13882 mem = gen_frame_mem (Pmode,
13883 gen_rtx_PLUS (Pmode, stack_top,
13884 GEN_INT (2 * GET_MODE_SIZE (Pmode))));
13885 emit_move_insn (opcode_addr, mem);
13886 emit_move_insn (opcode, gen_rtx_MEM (SImode, opcode_addr));
13887 emit_move_insn (tocompare, gen_int_mode (TARGET_32BIT ? 0x80410014
13888 : 0xE8410028, SImode));
13889
13890 do_compare_rtx_and_jump (opcode, tocompare, EQ, 1,
13891 SImode, NULL_RTX, NULL_RTX,
13892 no_toc_save_needed);
13893
13894 mem = gen_frame_mem (Pmode,
13895 gen_rtx_PLUS (Pmode, stack_top,
13896 GEN_INT (5 * GET_MODE_SIZE (Pmode))));
13897 emit_move_insn (mem, gen_rtx_REG (Pmode, 2));
13898 emit_label (no_toc_save_needed);
13899 }
13900
13901 /* This ties together stack memory (MEM with an alias set of frame_alias_set)
13902 and the change to the stack pointer. */
13903
13904 static void
rs6000_emit_stack_tie(void)13905 rs6000_emit_stack_tie (void)
13906 {
13907 rtx mem = gen_frame_mem (BLKmode,
13908 gen_rtx_REG (Pmode, STACK_POINTER_REGNUM));
13909
13910 emit_insn (gen_stack_tie (mem));
13911 }
13912
13913 /* Emit the correct code for allocating stack space, as insns.
13914 If COPY_R12, make sure a copy of the old frame is left in r12.
13915 The generated code may use hard register 0 as a temporary. */
13916
13917 static void
rs6000_emit_allocate_stack(HOST_WIDE_INT size,int copy_r12)13918 rs6000_emit_allocate_stack (HOST_WIDE_INT size, int copy_r12)
13919 {
13920 rtx insn;
13921 rtx stack_reg = gen_rtx_REG (Pmode, STACK_POINTER_REGNUM);
13922 rtx tmp_reg = gen_rtx_REG (Pmode, 0);
13923 rtx todec = gen_int_mode (-size, Pmode);
13924
13925 if (INTVAL (todec) != -size)
13926 {
13927 warning (0, "stack frame too large");
13928 emit_insn (gen_trap ());
13929 return;
13930 }
13931
13932 if (current_function_limit_stack)
13933 {
13934 if (REG_P (stack_limit_rtx)
13935 && REGNO (stack_limit_rtx) > 1
13936 && REGNO (stack_limit_rtx) <= 31)
13937 {
13938 emit_insn (TARGET_32BIT
13939 ? gen_addsi3 (tmp_reg,
13940 stack_limit_rtx,
13941 GEN_INT (size))
13942 : gen_adddi3 (tmp_reg,
13943 stack_limit_rtx,
13944 GEN_INT (size)));
13945
13946 emit_insn (gen_cond_trap (LTU, stack_reg, tmp_reg,
13947 const0_rtx));
13948 }
13949 else if (GET_CODE (stack_limit_rtx) == SYMBOL_REF
13950 && TARGET_32BIT
13951 && DEFAULT_ABI == ABI_V4)
13952 {
13953 rtx toload = gen_rtx_CONST (VOIDmode,
13954 gen_rtx_PLUS (Pmode,
13955 stack_limit_rtx,
13956 GEN_INT (size)));
13957
13958 emit_insn (gen_elf_high (tmp_reg, toload));
13959 emit_insn (gen_elf_low (tmp_reg, tmp_reg, toload));
13960 emit_insn (gen_cond_trap (LTU, stack_reg, tmp_reg,
13961 const0_rtx));
13962 }
13963 else
13964 warning (0, "stack limit expression is not supported");
13965 }
13966
13967 if (copy_r12 || ! TARGET_UPDATE)
13968 emit_move_insn (gen_rtx_REG (Pmode, 12), stack_reg);
13969
13970 if (TARGET_UPDATE)
13971 {
13972 if (size > 32767)
13973 {
13974 /* Need a note here so that try_split doesn't get confused. */
13975 if (get_last_insn () == NULL_RTX)
13976 emit_note (NOTE_INSN_DELETED);
13977 insn = emit_move_insn (tmp_reg, todec);
13978 try_split (PATTERN (insn), insn, 0);
13979 todec = tmp_reg;
13980 }
13981
13982 insn = emit_insn (TARGET_32BIT
13983 ? gen_movsi_update (stack_reg, stack_reg,
13984 todec, stack_reg)
13985 : gen_movdi_di_update (stack_reg, stack_reg,
13986 todec, stack_reg));
13987 }
13988 else
13989 {
13990 insn = emit_insn (TARGET_32BIT
13991 ? gen_addsi3 (stack_reg, stack_reg, todec)
13992 : gen_adddi3 (stack_reg, stack_reg, todec));
13993 emit_move_insn (gen_rtx_MEM (Pmode, stack_reg),
13994 gen_rtx_REG (Pmode, 12));
13995 }
13996
13997 RTX_FRAME_RELATED_P (insn) = 1;
13998 REG_NOTES (insn) =
13999 gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR,
14000 gen_rtx_SET (VOIDmode, stack_reg,
14001 gen_rtx_PLUS (Pmode, stack_reg,
14002 GEN_INT (-size))),
14003 REG_NOTES (insn));
14004 }
14005
14006 /* Add to 'insn' a note which is PATTERN (INSN) but with REG replaced
14007 with (plus:P (reg 1) VAL), and with REG2 replaced with RREG if REG2
14008 is not NULL. It would be nice if dwarf2out_frame_debug_expr could
14009 deduce these equivalences by itself so it wasn't necessary to hold
14010 its hand so much. */
14011
14012 static void
rs6000_frame_related(rtx insn,rtx reg,HOST_WIDE_INT val,rtx reg2,rtx rreg)14013 rs6000_frame_related (rtx insn, rtx reg, HOST_WIDE_INT val,
14014 rtx reg2, rtx rreg)
14015 {
14016 rtx real, temp;
14017
14018 /* copy_rtx will not make unique copies of registers, so we need to
14019 ensure we don't have unwanted sharing here. */
14020 if (reg == reg2)
14021 reg = gen_raw_REG (GET_MODE (reg), REGNO (reg));
14022
14023 if (reg == rreg)
14024 reg = gen_raw_REG (GET_MODE (reg), REGNO (reg));
14025
14026 real = copy_rtx (PATTERN (insn));
14027
14028 if (reg2 != NULL_RTX)
14029 real = replace_rtx (real, reg2, rreg);
14030
14031 real = replace_rtx (real, reg,
14032 gen_rtx_PLUS (Pmode, gen_rtx_REG (Pmode,
14033 STACK_POINTER_REGNUM),
14034 GEN_INT (val)));
14035
14036 /* We expect that 'real' is either a SET or a PARALLEL containing
14037 SETs (and possibly other stuff). In a PARALLEL, all the SETs
14038 are important so they all have to be marked RTX_FRAME_RELATED_P. */
14039
14040 if (GET_CODE (real) == SET)
14041 {
14042 rtx set = real;
14043
14044 temp = simplify_rtx (SET_SRC (set));
14045 if (temp)
14046 SET_SRC (set) = temp;
14047 temp = simplify_rtx (SET_DEST (set));
14048 if (temp)
14049 SET_DEST (set) = temp;
14050 if (GET_CODE (SET_DEST (set)) == MEM)
14051 {
14052 temp = simplify_rtx (XEXP (SET_DEST (set), 0));
14053 if (temp)
14054 XEXP (SET_DEST (set), 0) = temp;
14055 }
14056 }
14057 else
14058 {
14059 int i;
14060
14061 gcc_assert (GET_CODE (real) == PARALLEL);
14062 for (i = 0; i < XVECLEN (real, 0); i++)
14063 if (GET_CODE (XVECEXP (real, 0, i)) == SET)
14064 {
14065 rtx set = XVECEXP (real, 0, i);
14066
14067 temp = simplify_rtx (SET_SRC (set));
14068 if (temp)
14069 SET_SRC (set) = temp;
14070 temp = simplify_rtx (SET_DEST (set));
14071 if (temp)
14072 SET_DEST (set) = temp;
14073 if (GET_CODE (SET_DEST (set)) == MEM)
14074 {
14075 temp = simplify_rtx (XEXP (SET_DEST (set), 0));
14076 if (temp)
14077 XEXP (SET_DEST (set), 0) = temp;
14078 }
14079 RTX_FRAME_RELATED_P (set) = 1;
14080 }
14081 }
14082
14083 if (TARGET_SPE)
14084 real = spe_synthesize_frame_save (real);
14085
14086 RTX_FRAME_RELATED_P (insn) = 1;
14087 REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR,
14088 real,
14089 REG_NOTES (insn));
14090 }
14091
14092 /* Given an SPE frame note, return a PARALLEL of SETs with the
14093 original note, plus a synthetic register save. */
14094
14095 static rtx
spe_synthesize_frame_save(rtx real)14096 spe_synthesize_frame_save (rtx real)
14097 {
14098 rtx synth, offset, reg, real2;
14099
14100 if (GET_CODE (real) != SET
14101 || GET_MODE (SET_SRC (real)) != V2SImode)
14102 return real;
14103
14104 /* For the SPE, registers saved in 64-bits, get a PARALLEL for their
14105 frame related note. The parallel contains a set of the register
14106 being saved, and another set to a synthetic register (n+1200).
14107 This is so we can differentiate between 64-bit and 32-bit saves.
14108 Words cannot describe this nastiness. */
14109
14110 gcc_assert (GET_CODE (SET_DEST (real)) == MEM
14111 && GET_CODE (XEXP (SET_DEST (real), 0)) == PLUS
14112 && GET_CODE (SET_SRC (real)) == REG);
14113
14114 /* Transform:
14115 (set (mem (plus (reg x) (const y)))
14116 (reg z))
14117 into:
14118 (set (mem (plus (reg x) (const y+4)))
14119 (reg z+1200))
14120 */
14121
14122 real2 = copy_rtx (real);
14123 PUT_MODE (SET_DEST (real2), SImode);
14124 reg = SET_SRC (real2);
14125 real2 = replace_rtx (real2, reg, gen_rtx_REG (SImode, REGNO (reg)));
14126 synth = copy_rtx (real2);
14127
14128 if (BYTES_BIG_ENDIAN)
14129 {
14130 offset = XEXP (XEXP (SET_DEST (real2), 0), 1);
14131 real2 = replace_rtx (real2, offset, GEN_INT (INTVAL (offset) + 4));
14132 }
14133
14134 reg = SET_SRC (synth);
14135
14136 synth = replace_rtx (synth, reg,
14137 gen_rtx_REG (SImode, REGNO (reg) + 1200));
14138
14139 offset = XEXP (XEXP (SET_DEST (synth), 0), 1);
14140 synth = replace_rtx (synth, offset,
14141 GEN_INT (INTVAL (offset)
14142 + (BYTES_BIG_ENDIAN ? 0 : 4)));
14143
14144 RTX_FRAME_RELATED_P (synth) = 1;
14145 RTX_FRAME_RELATED_P (real2) = 1;
14146 if (BYTES_BIG_ENDIAN)
14147 real = gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, synth, real2));
14148 else
14149 real = gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, real2, synth));
14150
14151 return real;
14152 }
14153
14154 /* Returns an insn that has a vrsave set operation with the
14155 appropriate CLOBBERs. */
14156
14157 static rtx
generate_set_vrsave(rtx reg,rs6000_stack_t * info,int epiloguep)14158 generate_set_vrsave (rtx reg, rs6000_stack_t *info, int epiloguep)
14159 {
14160 int nclobs, i;
14161 rtx insn, clobs[TOTAL_ALTIVEC_REGS + 1];
14162 rtx vrsave = gen_rtx_REG (SImode, VRSAVE_REGNO);
14163
14164 clobs[0]
14165 = gen_rtx_SET (VOIDmode,
14166 vrsave,
14167 gen_rtx_UNSPEC_VOLATILE (SImode,
14168 gen_rtvec (2, reg, vrsave),
14169 UNSPECV_SET_VRSAVE));
14170
14171 nclobs = 1;
14172
14173 /* We need to clobber the registers in the mask so the scheduler
14174 does not move sets to VRSAVE before sets of AltiVec registers.
14175
14176 However, if the function receives nonlocal gotos, reload will set
14177 all call saved registers live. We will end up with:
14178
14179 (set (reg 999) (mem))
14180 (parallel [ (set (reg vrsave) (unspec blah))
14181 (clobber (reg 999))])
14182
14183 The clobber will cause the store into reg 999 to be dead, and
14184 flow will attempt to delete an epilogue insn. In this case, we
14185 need an unspec use/set of the register. */
14186
14187 for (i = FIRST_ALTIVEC_REGNO; i <= LAST_ALTIVEC_REGNO; ++i)
14188 if (info->vrsave_mask & ALTIVEC_REG_BIT (i))
14189 {
14190 if (!epiloguep || call_used_regs [i])
14191 clobs[nclobs++] = gen_rtx_CLOBBER (VOIDmode,
14192 gen_rtx_REG (V4SImode, i));
14193 else
14194 {
14195 rtx reg = gen_rtx_REG (V4SImode, i);
14196
14197 clobs[nclobs++]
14198 = gen_rtx_SET (VOIDmode,
14199 reg,
14200 gen_rtx_UNSPEC (V4SImode,
14201 gen_rtvec (1, reg), 27));
14202 }
14203 }
14204
14205 insn = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (nclobs));
14206
14207 for (i = 0; i < nclobs; ++i)
14208 XVECEXP (insn, 0, i) = clobs[i];
14209
14210 return insn;
14211 }
14212
14213 /* Save a register into the frame, and emit RTX_FRAME_RELATED_P notes.
14214 Save REGNO into [FRAME_REG + OFFSET] in mode MODE. */
14215
14216 static void
emit_frame_save(rtx frame_reg,rtx frame_ptr,enum machine_mode mode,unsigned int regno,int offset,HOST_WIDE_INT total_size)14217 emit_frame_save (rtx frame_reg, rtx frame_ptr, enum machine_mode mode,
14218 unsigned int regno, int offset, HOST_WIDE_INT total_size)
14219 {
14220 rtx reg, offset_rtx, insn, mem, addr, int_rtx;
14221 rtx replacea, replaceb;
14222
14223 int_rtx = GEN_INT (offset);
14224
14225 /* Some cases that need register indexed addressing. */
14226 if ((TARGET_ALTIVEC_ABI && ALTIVEC_VECTOR_MODE (mode))
14227 || (TARGET_E500_DOUBLE && mode == DFmode)
14228 || (TARGET_SPE_ABI
14229 && SPE_VECTOR_MODE (mode)
14230 && !SPE_CONST_OFFSET_OK (offset)))
14231 {
14232 /* Whomever calls us must make sure r11 is available in the
14233 flow path of instructions in the prologue. */
14234 offset_rtx = gen_rtx_REG (Pmode, 11);
14235 emit_move_insn (offset_rtx, int_rtx);
14236
14237 replacea = offset_rtx;
14238 replaceb = int_rtx;
14239 }
14240 else
14241 {
14242 offset_rtx = int_rtx;
14243 replacea = NULL_RTX;
14244 replaceb = NULL_RTX;
14245 }
14246
14247 reg = gen_rtx_REG (mode, regno);
14248 addr = gen_rtx_PLUS (Pmode, frame_reg, offset_rtx);
14249 mem = gen_frame_mem (mode, addr);
14250
14251 insn = emit_move_insn (mem, reg);
14252
14253 rs6000_frame_related (insn, frame_ptr, total_size, replacea, replaceb);
14254 }
14255
14256 /* Emit an offset memory reference suitable for a frame store, while
14257 converting to a valid addressing mode. */
14258
14259 static rtx
gen_frame_mem_offset(enum machine_mode mode,rtx reg,int offset)14260 gen_frame_mem_offset (enum machine_mode mode, rtx reg, int offset)
14261 {
14262 rtx int_rtx, offset_rtx;
14263
14264 int_rtx = GEN_INT (offset);
14265
14266 if ((TARGET_SPE_ABI && SPE_VECTOR_MODE (mode))
14267 || (TARGET_E500_DOUBLE && mode == DFmode))
14268 {
14269 offset_rtx = gen_rtx_REG (Pmode, FIXED_SCRATCH);
14270 emit_move_insn (offset_rtx, int_rtx);
14271 }
14272 else
14273 offset_rtx = int_rtx;
14274
14275 return gen_frame_mem (mode, gen_rtx_PLUS (Pmode, reg, offset_rtx));
14276 }
14277
14278 /* Look for user-defined global regs. We should not save and restore these,
14279 and cannot use stmw/lmw if there are any in its range. */
14280
14281 static bool
no_global_regs_above(int first_greg)14282 no_global_regs_above (int first_greg)
14283 {
14284 int i;
14285 for (i = 0; i < 32 - first_greg; i++)
14286 if (global_regs[first_greg + i])
14287 return false;
14288 return true;
14289 }
14290
14291 #ifndef TARGET_FIX_AND_CONTINUE
14292 #define TARGET_FIX_AND_CONTINUE 0
14293 #endif
14294
14295 /* Emit function prologue as insns. */
14296
14297 void
rs6000_emit_prologue(void)14298 rs6000_emit_prologue (void)
14299 {
14300 rs6000_stack_t *info = rs6000_stack_info ();
14301 enum machine_mode reg_mode = Pmode;
14302 int reg_size = TARGET_32BIT ? 4 : 8;
14303 rtx sp_reg_rtx = gen_rtx_REG (Pmode, STACK_POINTER_REGNUM);
14304 rtx frame_ptr_rtx = gen_rtx_REG (Pmode, 12);
14305 rtx frame_reg_rtx = sp_reg_rtx;
14306 rtx cr_save_rtx = NULL_RTX;
14307 rtx insn;
14308 int saving_FPRs_inline;
14309 int using_store_multiple;
14310 HOST_WIDE_INT sp_offset = 0;
14311
14312 if (TARGET_FIX_AND_CONTINUE)
14313 {
14314 /* gdb on darwin arranges to forward a function from the old
14315 address by modifying the first 5 instructions of the function
14316 to branch to the overriding function. This is necessary to
14317 permit function pointers that point to the old function to
14318 actually forward to the new function. */
14319 emit_insn (gen_nop ());
14320 emit_insn (gen_nop ());
14321 emit_insn (gen_nop ());
14322 emit_insn (gen_nop ());
14323 emit_insn (gen_nop ());
14324 }
14325
14326 if (TARGET_SPE_ABI && info->spe_64bit_regs_used != 0)
14327 {
14328 reg_mode = V2SImode;
14329 reg_size = 8;
14330 }
14331
14332 using_store_multiple = (TARGET_MULTIPLE && ! TARGET_POWERPC64
14333 && (!TARGET_SPE_ABI
14334 || info->spe_64bit_regs_used == 0)
14335 && info->first_gp_reg_save < 31
14336 && no_global_regs_above (info->first_gp_reg_save));
14337 saving_FPRs_inline = (info->first_fp_reg_save == 64
14338 || FP_SAVE_INLINE (info->first_fp_reg_save)
14339 || current_function_calls_eh_return
14340 || cfun->machine->ra_need_lr);
14341
14342 /* For V.4, update stack before we do any saving and set back pointer. */
14343 if (! WORLD_SAVE_P (info)
14344 && info->push_p
14345 && (DEFAULT_ABI == ABI_V4
14346 || current_function_calls_eh_return))
14347 {
14348 if (info->total_size < 32767)
14349 sp_offset = info->total_size;
14350 else
14351 frame_reg_rtx = frame_ptr_rtx;
14352 rs6000_emit_allocate_stack (info->total_size,
14353 (frame_reg_rtx != sp_reg_rtx
14354 && (info->cr_save_p
14355 || info->lr_save_p
14356 || info->first_fp_reg_save < 64
14357 || info->first_gp_reg_save < 32
14358 )));
14359 if (frame_reg_rtx != sp_reg_rtx)
14360 rs6000_emit_stack_tie ();
14361 }
14362
14363 /* Handle world saves specially here. */
14364 if (WORLD_SAVE_P (info))
14365 {
14366 int i, j, sz;
14367 rtx treg;
14368 rtvec p;
14369 rtx reg0;
14370
14371 /* save_world expects lr in r0. */
14372 reg0 = gen_rtx_REG (Pmode, 0);
14373 if (info->lr_save_p)
14374 {
14375 insn = emit_move_insn (reg0,
14376 gen_rtx_REG (Pmode, LINK_REGISTER_REGNUM));
14377 RTX_FRAME_RELATED_P (insn) = 1;
14378 }
14379
14380 /* The SAVE_WORLD and RESTORE_WORLD routines make a number of
14381 assumptions about the offsets of various bits of the stack
14382 frame. */
14383 gcc_assert (info->gp_save_offset == -220
14384 && info->fp_save_offset == -144
14385 && info->lr_save_offset == 8
14386 && info->cr_save_offset == 4
14387 && info->push_p
14388 && info->lr_save_p
14389 && (!current_function_calls_eh_return
14390 || info->ehrd_offset == -432)
14391 && info->vrsave_save_offset == -224
14392 && info->altivec_save_offset == -416);
14393
14394 treg = gen_rtx_REG (SImode, 11);
14395 emit_move_insn (treg, GEN_INT (-info->total_size));
14396
14397 /* SAVE_WORLD takes the caller's LR in R0 and the frame size
14398 in R11. It also clobbers R12, so beware! */
14399
14400 /* Preserve CR2 for save_world prologues */
14401 sz = 5;
14402 sz += 32 - info->first_gp_reg_save;
14403 sz += 64 - info->first_fp_reg_save;
14404 sz += LAST_ALTIVEC_REGNO - info->first_altivec_reg_save + 1;
14405 p = rtvec_alloc (sz);
14406 j = 0;
14407 RTVEC_ELT (p, j++) = gen_rtx_CLOBBER (VOIDmode,
14408 gen_rtx_REG (Pmode,
14409 LINK_REGISTER_REGNUM));
14410 RTVEC_ELT (p, j++) = gen_rtx_USE (VOIDmode,
14411 gen_rtx_SYMBOL_REF (Pmode,
14412 "*save_world"));
14413 /* We do floats first so that the instruction pattern matches
14414 properly. */
14415 for (i = 0; i < 64 - info->first_fp_reg_save; i++)
14416 {
14417 rtx reg = gen_rtx_REG (DFmode, info->first_fp_reg_save + i);
14418 rtx addr = gen_rtx_PLUS (Pmode, frame_reg_rtx,
14419 GEN_INT (info->fp_save_offset
14420 + sp_offset + 8 * i));
14421 rtx mem = gen_frame_mem (DFmode, addr);
14422
14423 RTVEC_ELT (p, j++) = gen_rtx_SET (VOIDmode, mem, reg);
14424 }
14425 for (i = 0; info->first_altivec_reg_save + i <= LAST_ALTIVEC_REGNO; i++)
14426 {
14427 rtx reg = gen_rtx_REG (V4SImode, info->first_altivec_reg_save + i);
14428 rtx addr = gen_rtx_PLUS (Pmode, frame_reg_rtx,
14429 GEN_INT (info->altivec_save_offset
14430 + sp_offset + 16 * i));
14431 rtx mem = gen_frame_mem (V4SImode, addr);
14432
14433 RTVEC_ELT (p, j++) = gen_rtx_SET (VOIDmode, mem, reg);
14434 }
14435 for (i = 0; i < 32 - info->first_gp_reg_save; i++)
14436 {
14437 rtx reg = gen_rtx_REG (reg_mode, info->first_gp_reg_save + i);
14438 rtx addr = gen_rtx_PLUS (Pmode, frame_reg_rtx,
14439 GEN_INT (info->gp_save_offset
14440 + sp_offset + reg_size * i));
14441 rtx mem = gen_frame_mem (reg_mode, addr);
14442
14443 RTVEC_ELT (p, j++) = gen_rtx_SET (VOIDmode, mem, reg);
14444 }
14445
14446 {
14447 /* CR register traditionally saved as CR2. */
14448 rtx reg = gen_rtx_REG (reg_mode, CR2_REGNO);
14449 rtx addr = gen_rtx_PLUS (Pmode, frame_reg_rtx,
14450 GEN_INT (info->cr_save_offset
14451 + sp_offset));
14452 rtx mem = gen_frame_mem (reg_mode, addr);
14453
14454 RTVEC_ELT (p, j++) = gen_rtx_SET (VOIDmode, mem, reg);
14455 }
14456 /* Explain about use of R0. */
14457 if (info->lr_save_p)
14458 {
14459 rtx addr = gen_rtx_PLUS (Pmode, frame_reg_rtx,
14460 GEN_INT (info->lr_save_offset
14461 + sp_offset));
14462 rtx mem = gen_frame_mem (reg_mode, addr);
14463
14464 RTVEC_ELT (p, j++) = gen_rtx_SET (VOIDmode, mem, reg0);
14465 }
14466 /* Explain what happens to the stack pointer. */
14467 {
14468 rtx newval = gen_rtx_PLUS (Pmode, sp_reg_rtx, treg);
14469 RTVEC_ELT (p, j++) = gen_rtx_SET (VOIDmode, sp_reg_rtx, newval);
14470 }
14471
14472 insn = emit_insn (gen_rtx_PARALLEL (VOIDmode, p));
14473 rs6000_frame_related (insn, frame_ptr_rtx, info->total_size,
14474 treg, GEN_INT (-info->total_size));
14475 sp_offset = info->total_size;
14476 }
14477
14478 /* If we use the link register, get it into r0. */
14479 if (!WORLD_SAVE_P (info) && info->lr_save_p)
14480 {
14481 insn = emit_move_insn (gen_rtx_REG (Pmode, 0),
14482 gen_rtx_REG (Pmode, LINK_REGISTER_REGNUM));
14483 RTX_FRAME_RELATED_P (insn) = 1;
14484 }
14485
14486 /* If we need to save CR, put it into r12. */
14487 if (!WORLD_SAVE_P (info) && info->cr_save_p && frame_reg_rtx != frame_ptr_rtx)
14488 {
14489 rtx set;
14490
14491 cr_save_rtx = gen_rtx_REG (SImode, 12);
14492 insn = emit_insn (gen_movesi_from_cr (cr_save_rtx));
14493 RTX_FRAME_RELATED_P (insn) = 1;
14494 /* Now, there's no way that dwarf2out_frame_debug_expr is going
14495 to understand '(unspec:SI [(reg:CC 68) ...] UNSPEC_MOVESI_FROM_CR)'.
14496 But that's OK. All we have to do is specify that _one_ condition
14497 code register is saved in this stack slot. The thrower's epilogue
14498 will then restore all the call-saved registers.
14499 We use CR2_REGNO (70) to be compatible with gcc-2.95 on Linux. */
14500 set = gen_rtx_SET (VOIDmode, cr_save_rtx,
14501 gen_rtx_REG (SImode, CR2_REGNO));
14502 REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR,
14503 set,
14504 REG_NOTES (insn));
14505 }
14506
14507 /* Do any required saving of fpr's. If only one or two to save, do
14508 it ourselves. Otherwise, call function. */
14509 if (!WORLD_SAVE_P (info) && saving_FPRs_inline)
14510 {
14511 int i;
14512 for (i = 0; i < 64 - info->first_fp_reg_save; i++)
14513 if ((regs_ever_live[info->first_fp_reg_save+i]
14514 && ! call_used_regs[info->first_fp_reg_save+i]))
14515 emit_frame_save (frame_reg_rtx, frame_ptr_rtx, DFmode,
14516 info->first_fp_reg_save + i,
14517 info->fp_save_offset + sp_offset + 8 * i,
14518 info->total_size);
14519 }
14520 else if (!WORLD_SAVE_P (info) && info->first_fp_reg_save != 64)
14521 {
14522 int i;
14523 char rname[30];
14524 const char *alloc_rname;
14525 rtvec p;
14526 p = rtvec_alloc (2 + 64 - info->first_fp_reg_save);
14527
14528 RTVEC_ELT (p, 0) = gen_rtx_CLOBBER (VOIDmode,
14529 gen_rtx_REG (Pmode,
14530 LINK_REGISTER_REGNUM));
14531 sprintf (rname, "%s%d%s", SAVE_FP_PREFIX,
14532 info->first_fp_reg_save - 32, SAVE_FP_SUFFIX);
14533 alloc_rname = ggc_strdup (rname);
14534 RTVEC_ELT (p, 1) = gen_rtx_USE (VOIDmode,
14535 gen_rtx_SYMBOL_REF (Pmode,
14536 alloc_rname));
14537 for (i = 0; i < 64 - info->first_fp_reg_save; i++)
14538 {
14539 rtx addr, reg, mem;
14540 reg = gen_rtx_REG (DFmode, info->first_fp_reg_save + i);
14541 addr = gen_rtx_PLUS (Pmode, frame_reg_rtx,
14542 GEN_INT (info->fp_save_offset
14543 + sp_offset + 8*i));
14544 mem = gen_frame_mem (DFmode, addr);
14545
14546 RTVEC_ELT (p, i + 2) = gen_rtx_SET (VOIDmode, mem, reg);
14547 }
14548 insn = emit_insn (gen_rtx_PARALLEL (VOIDmode, p));
14549 rs6000_frame_related (insn, frame_ptr_rtx, info->total_size,
14550 NULL_RTX, NULL_RTX);
14551 }
14552
14553 /* Save GPRs. This is done as a PARALLEL if we are using
14554 the store-multiple instructions. */
14555 if (!WORLD_SAVE_P (info) && using_store_multiple)
14556 {
14557 rtvec p;
14558 int i;
14559 p = rtvec_alloc (32 - info->first_gp_reg_save);
14560 for (i = 0; i < 32 - info->first_gp_reg_save; i++)
14561 {
14562 rtx addr, reg, mem;
14563 reg = gen_rtx_REG (reg_mode, info->first_gp_reg_save + i);
14564 addr = gen_rtx_PLUS (Pmode, frame_reg_rtx,
14565 GEN_INT (info->gp_save_offset
14566 + sp_offset
14567 + reg_size * i));
14568 mem = gen_frame_mem (reg_mode, addr);
14569
14570 RTVEC_ELT (p, i) = gen_rtx_SET (VOIDmode, mem, reg);
14571 }
14572 insn = emit_insn (gen_rtx_PARALLEL (VOIDmode, p));
14573 rs6000_frame_related (insn, frame_ptr_rtx, info->total_size,
14574 NULL_RTX, NULL_RTX);
14575 }
14576 else if (!WORLD_SAVE_P (info))
14577 {
14578 int i;
14579 for (i = 0; i < 32 - info->first_gp_reg_save; i++)
14580 if ((regs_ever_live[info->first_gp_reg_save + i]
14581 && (!call_used_regs[info->first_gp_reg_save + i]
14582 || (i + info->first_gp_reg_save
14583 == RS6000_PIC_OFFSET_TABLE_REGNUM
14584 && TARGET_TOC && TARGET_MINIMAL_TOC)))
14585 || (i + info->first_gp_reg_save == RS6000_PIC_OFFSET_TABLE_REGNUM
14586 && ((DEFAULT_ABI == ABI_V4 && flag_pic != 0)
14587 || (DEFAULT_ABI == ABI_DARWIN && flag_pic))))
14588 {
14589 rtx addr, reg, mem;
14590 reg = gen_rtx_REG (reg_mode, info->first_gp_reg_save + i);
14591
14592 if (TARGET_SPE_ABI && info->spe_64bit_regs_used != 0)
14593 {
14594 int offset = info->spe_gp_save_offset + sp_offset + 8 * i;
14595 rtx b;
14596
14597 if (!SPE_CONST_OFFSET_OK (offset))
14598 {
14599 b = gen_rtx_REG (Pmode, FIXED_SCRATCH);
14600 emit_move_insn (b, GEN_INT (offset));
14601 }
14602 else
14603 b = GEN_INT (offset);
14604
14605 addr = gen_rtx_PLUS (Pmode, frame_reg_rtx, b);
14606 mem = gen_frame_mem (V2SImode, addr);
14607 insn = emit_move_insn (mem, reg);
14608
14609 if (GET_CODE (b) == CONST_INT)
14610 rs6000_frame_related (insn, frame_ptr_rtx, info->total_size,
14611 NULL_RTX, NULL_RTX);
14612 else
14613 rs6000_frame_related (insn, frame_ptr_rtx, info->total_size,
14614 b, GEN_INT (offset));
14615 }
14616 else
14617 {
14618 addr = gen_rtx_PLUS (Pmode, frame_reg_rtx,
14619 GEN_INT (info->gp_save_offset
14620 + sp_offset
14621 + reg_size * i));
14622 mem = gen_frame_mem (reg_mode, addr);
14623
14624 insn = emit_move_insn (mem, reg);
14625 rs6000_frame_related (insn, frame_ptr_rtx, info->total_size,
14626 NULL_RTX, NULL_RTX);
14627 }
14628 }
14629 }
14630
14631 /* ??? There's no need to emit actual instructions here, but it's the
14632 easiest way to get the frame unwind information emitted. */
14633 if (current_function_calls_eh_return)
14634 {
14635 unsigned int i, regno;
14636
14637 /* In AIX ABI we need to pretend we save r2 here. */
14638 if (TARGET_AIX)
14639 {
14640 rtx addr, reg, mem;
14641
14642 reg = gen_rtx_REG (reg_mode, 2);
14643 addr = gen_rtx_PLUS (Pmode, frame_reg_rtx,
14644 GEN_INT (sp_offset + 5 * reg_size));
14645 mem = gen_frame_mem (reg_mode, addr);
14646
14647 insn = emit_move_insn (mem, reg);
14648 rs6000_frame_related (insn, frame_ptr_rtx, info->total_size,
14649 NULL_RTX, NULL_RTX);
14650 PATTERN (insn) = gen_blockage ();
14651 }
14652
14653 for (i = 0; ; ++i)
14654 {
14655 regno = EH_RETURN_DATA_REGNO (i);
14656 if (regno == INVALID_REGNUM)
14657 break;
14658
14659 emit_frame_save (frame_reg_rtx, frame_ptr_rtx, reg_mode, regno,
14660 info->ehrd_offset + sp_offset
14661 + reg_size * (int) i,
14662 info->total_size);
14663 }
14664 }
14665
14666 /* Save lr if we used it. */
14667 if (!WORLD_SAVE_P (info) && info->lr_save_p)
14668 {
14669 rtx addr = gen_rtx_PLUS (Pmode, frame_reg_rtx,
14670 GEN_INT (info->lr_save_offset + sp_offset));
14671 rtx reg = gen_rtx_REG (Pmode, 0);
14672 rtx mem = gen_rtx_MEM (Pmode, addr);
14673 /* This should not be of frame_alias_set, because of
14674 __builtin_return_address. */
14675
14676 insn = emit_move_insn (mem, reg);
14677 rs6000_frame_related (insn, frame_ptr_rtx, info->total_size,
14678 NULL_RTX, NULL_RTX);
14679 }
14680
14681 /* Save CR if we use any that must be preserved. */
14682 if (!WORLD_SAVE_P (info) && info->cr_save_p)
14683 {
14684 rtx addr = gen_rtx_PLUS (Pmode, frame_reg_rtx,
14685 GEN_INT (info->cr_save_offset + sp_offset));
14686 rtx mem = gen_frame_mem (SImode, addr);
14687 /* See the large comment above about why CR2_REGNO is used. */
14688 rtx magic_eh_cr_reg = gen_rtx_REG (SImode, CR2_REGNO);
14689
14690 /* If r12 was used to hold the original sp, copy cr into r0 now
14691 that it's free. */
14692 if (REGNO (frame_reg_rtx) == 12)
14693 {
14694 rtx set;
14695
14696 cr_save_rtx = gen_rtx_REG (SImode, 0);
14697 insn = emit_insn (gen_movesi_from_cr (cr_save_rtx));
14698 RTX_FRAME_RELATED_P (insn) = 1;
14699 set = gen_rtx_SET (VOIDmode, cr_save_rtx, magic_eh_cr_reg);
14700 REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR,
14701 set,
14702 REG_NOTES (insn));
14703
14704 }
14705 insn = emit_move_insn (mem, cr_save_rtx);
14706
14707 rs6000_frame_related (insn, frame_ptr_rtx, info->total_size,
14708 NULL_RTX, NULL_RTX);
14709 }
14710
14711 /* Update stack and set back pointer unless this is V.4,
14712 for which it was done previously. */
14713 if (!WORLD_SAVE_P (info) && info->push_p
14714 && !(DEFAULT_ABI == ABI_V4 || current_function_calls_eh_return))
14715 {
14716 if (info->total_size < 32767)
14717 sp_offset = info->total_size;
14718 else
14719 frame_reg_rtx = frame_ptr_rtx;
14720 rs6000_emit_allocate_stack (info->total_size,
14721 (frame_reg_rtx != sp_reg_rtx
14722 && ((info->altivec_size != 0)
14723 || (info->vrsave_mask != 0)
14724 )));
14725 if (frame_reg_rtx != sp_reg_rtx)
14726 rs6000_emit_stack_tie ();
14727 }
14728
14729 /* Set frame pointer, if needed. */
14730 if (frame_pointer_needed)
14731 {
14732 insn = emit_move_insn (gen_rtx_REG (Pmode, HARD_FRAME_POINTER_REGNUM),
14733 sp_reg_rtx);
14734 RTX_FRAME_RELATED_P (insn) = 1;
14735 }
14736
14737 /* Save AltiVec registers if needed. Save here because the red zone does
14738 not include AltiVec registers. */
14739 if (!WORLD_SAVE_P (info) && TARGET_ALTIVEC_ABI && info->altivec_size != 0)
14740 {
14741 int i;
14742
14743 /* There should be a non inline version of this, for when we
14744 are saving lots of vector registers. */
14745 for (i = info->first_altivec_reg_save; i <= LAST_ALTIVEC_REGNO; ++i)
14746 if (info->vrsave_mask & ALTIVEC_REG_BIT (i))
14747 {
14748 rtx areg, savereg, mem;
14749 int offset;
14750
14751 offset = info->altivec_save_offset + sp_offset
14752 + 16 * (i - info->first_altivec_reg_save);
14753
14754 savereg = gen_rtx_REG (V4SImode, i);
14755
14756 areg = gen_rtx_REG (Pmode, 0);
14757 emit_move_insn (areg, GEN_INT (offset));
14758
14759 /* AltiVec addressing mode is [reg+reg]. */
14760 mem = gen_frame_mem (V4SImode,
14761 gen_rtx_PLUS (Pmode, frame_reg_rtx, areg));
14762
14763 insn = emit_move_insn (mem, savereg);
14764
14765 rs6000_frame_related (insn, frame_ptr_rtx, info->total_size,
14766 areg, GEN_INT (offset));
14767 }
14768 }
14769
14770 /* VRSAVE is a bit vector representing which AltiVec registers
14771 are used. The OS uses this to determine which vector
14772 registers to save on a context switch. We need to save
14773 VRSAVE on the stack frame, add whatever AltiVec registers we
14774 used in this function, and do the corresponding magic in the
14775 epilogue. */
14776
14777 if (TARGET_ALTIVEC && TARGET_ALTIVEC_VRSAVE
14778 && info->vrsave_mask != 0)
14779 {
14780 rtx reg, mem, vrsave;
14781 int offset;
14782
14783 /* Get VRSAVE onto a GPR. Note that ABI_V4 might be using r12
14784 as frame_reg_rtx and r11 as the static chain pointer for
14785 nested functions. */
14786 reg = gen_rtx_REG (SImode, 0);
14787 vrsave = gen_rtx_REG (SImode, VRSAVE_REGNO);
14788 if (TARGET_MACHO)
14789 emit_insn (gen_get_vrsave_internal (reg));
14790 else
14791 emit_insn (gen_rtx_SET (VOIDmode, reg, vrsave));
14792
14793 if (!WORLD_SAVE_P (info))
14794 {
14795 /* Save VRSAVE. */
14796 offset = info->vrsave_save_offset + sp_offset;
14797 mem = gen_frame_mem (SImode,
14798 gen_rtx_PLUS (Pmode, frame_reg_rtx,
14799 GEN_INT (offset)));
14800 insn = emit_move_insn (mem, reg);
14801 }
14802
14803 /* Include the registers in the mask. */
14804 emit_insn (gen_iorsi3 (reg, reg, GEN_INT ((int) info->vrsave_mask)));
14805
14806 insn = emit_insn (generate_set_vrsave (reg, info, 0));
14807 }
14808
14809 /* If we are using RS6000_PIC_OFFSET_TABLE_REGNUM, we need to set it up. */
14810 if ((TARGET_TOC && TARGET_MINIMAL_TOC && get_pool_size () != 0)
14811 || (DEFAULT_ABI == ABI_V4
14812 && (flag_pic == 1 || (flag_pic && TARGET_SECURE_PLT))
14813 && regs_ever_live[RS6000_PIC_OFFSET_TABLE_REGNUM]))
14814 {
14815 /* If emit_load_toc_table will use the link register, we need to save
14816 it. We use R12 for this purpose because emit_load_toc_table
14817 can use register 0. This allows us to use a plain 'blr' to return
14818 from the procedure more often. */
14819 int save_LR_around_toc_setup = (TARGET_ELF
14820 && DEFAULT_ABI != ABI_AIX
14821 && flag_pic
14822 && ! info->lr_save_p
14823 && EDGE_COUNT (EXIT_BLOCK_PTR->preds) > 0);
14824 if (save_LR_around_toc_setup)
14825 {
14826 rtx lr = gen_rtx_REG (Pmode, LINK_REGISTER_REGNUM);
14827
14828 insn = emit_move_insn (frame_ptr_rtx, lr);
14829 rs6000_maybe_dead (insn);
14830 RTX_FRAME_RELATED_P (insn) = 1;
14831
14832 rs6000_emit_load_toc_table (TRUE);
14833
14834 insn = emit_move_insn (lr, frame_ptr_rtx);
14835 rs6000_maybe_dead (insn);
14836 RTX_FRAME_RELATED_P (insn) = 1;
14837 }
14838 else
14839 rs6000_emit_load_toc_table (TRUE);
14840 }
14841
14842 #if TARGET_MACHO
14843 if (DEFAULT_ABI == ABI_DARWIN
14844 && flag_pic && current_function_uses_pic_offset_table)
14845 {
14846 rtx lr = gen_rtx_REG (Pmode, LINK_REGISTER_REGNUM);
14847 rtx src = machopic_function_base_sym ();
14848
14849 /* Save and restore LR locally around this call (in R0). */
14850 if (!info->lr_save_p)
14851 rs6000_maybe_dead (emit_move_insn (gen_rtx_REG (Pmode, 0), lr));
14852
14853 rs6000_maybe_dead (emit_insn (gen_load_macho_picbase (lr, src)));
14854
14855 insn = emit_move_insn (gen_rtx_REG (Pmode,
14856 RS6000_PIC_OFFSET_TABLE_REGNUM),
14857 lr);
14858 rs6000_maybe_dead (insn);
14859
14860 if (!info->lr_save_p)
14861 rs6000_maybe_dead (emit_move_insn (lr, gen_rtx_REG (Pmode, 0)));
14862 }
14863 #endif
14864 }
14865
14866 /* Write function prologue. */
14867
14868 static void
rs6000_output_function_prologue(FILE * file,HOST_WIDE_INT size ATTRIBUTE_UNUSED)14869 rs6000_output_function_prologue (FILE *file,
14870 HOST_WIDE_INT size ATTRIBUTE_UNUSED)
14871 {
14872 rs6000_stack_t *info = rs6000_stack_info ();
14873
14874 if (TARGET_DEBUG_STACK)
14875 debug_stack_info (info);
14876
14877 /* Write .extern for any function we will call to save and restore
14878 fp values. */
14879 if (info->first_fp_reg_save < 64
14880 && !FP_SAVE_INLINE (info->first_fp_reg_save))
14881 fprintf (file, "\t.extern %s%d%s\n\t.extern %s%d%s\n",
14882 SAVE_FP_PREFIX, info->first_fp_reg_save - 32, SAVE_FP_SUFFIX,
14883 RESTORE_FP_PREFIX, info->first_fp_reg_save - 32,
14884 RESTORE_FP_SUFFIX);
14885
14886 /* Write .extern for AIX common mode routines, if needed. */
14887 if (! TARGET_POWER && ! TARGET_POWERPC && ! common_mode_defined)
14888 {
14889 fputs ("\t.extern __mulh\n", file);
14890 fputs ("\t.extern __mull\n", file);
14891 fputs ("\t.extern __divss\n", file);
14892 fputs ("\t.extern __divus\n", file);
14893 fputs ("\t.extern __quoss\n", file);
14894 fputs ("\t.extern __quous\n", file);
14895 common_mode_defined = 1;
14896 }
14897
14898 if (! HAVE_prologue)
14899 {
14900 start_sequence ();
14901
14902 /* A NOTE_INSN_DELETED is supposed to be at the start and end of
14903 the "toplevel" insn chain. */
14904 emit_note (NOTE_INSN_DELETED);
14905 rs6000_emit_prologue ();
14906 emit_note (NOTE_INSN_DELETED);
14907
14908 /* Expand INSN_ADDRESSES so final() doesn't crash. */
14909 {
14910 rtx insn;
14911 unsigned addr = 0;
14912 for (insn = get_insns (); insn != 0; insn = NEXT_INSN (insn))
14913 {
14914 INSN_ADDRESSES_NEW (insn, addr);
14915 addr += 4;
14916 }
14917 }
14918
14919 if (TARGET_DEBUG_STACK)
14920 debug_rtx_list (get_insns (), 100);
14921 final (get_insns (), file, FALSE);
14922 end_sequence ();
14923 }
14924
14925 rs6000_pic_labelno++;
14926 }
14927
14928 /* Emit function epilogue as insns.
14929
14930 At present, dwarf2out_frame_debug_expr doesn't understand
14931 register restores, so we don't bother setting RTX_FRAME_RELATED_P
14932 anywhere in the epilogue. Most of the insns below would in any case
14933 need special notes to explain where r11 is in relation to the stack. */
14934
14935 void
rs6000_emit_epilogue(int sibcall)14936 rs6000_emit_epilogue (int sibcall)
14937 {
14938 rs6000_stack_t *info;
14939 int restoring_FPRs_inline;
14940 int using_load_multiple;
14941 int using_mfcr_multiple;
14942 int use_backchain_to_restore_sp;
14943 int sp_offset = 0;
14944 rtx sp_reg_rtx = gen_rtx_REG (Pmode, 1);
14945 rtx frame_reg_rtx = sp_reg_rtx;
14946 enum machine_mode reg_mode = Pmode;
14947 int reg_size = TARGET_32BIT ? 4 : 8;
14948 int i;
14949
14950 info = rs6000_stack_info ();
14951
14952 if (TARGET_SPE_ABI && info->spe_64bit_regs_used != 0)
14953 {
14954 reg_mode = V2SImode;
14955 reg_size = 8;
14956 }
14957
14958 using_load_multiple = (TARGET_MULTIPLE && ! TARGET_POWERPC64
14959 && (!TARGET_SPE_ABI
14960 || info->spe_64bit_regs_used == 0)
14961 && info->first_gp_reg_save < 31
14962 && no_global_regs_above (info->first_gp_reg_save));
14963 restoring_FPRs_inline = (sibcall
14964 || current_function_calls_eh_return
14965 || info->first_fp_reg_save == 64
14966 || FP_SAVE_INLINE (info->first_fp_reg_save));
14967 use_backchain_to_restore_sp = (frame_pointer_needed
14968 || current_function_calls_alloca
14969 || info->total_size > 32767);
14970 using_mfcr_multiple = (rs6000_cpu == PROCESSOR_PPC601
14971 || rs6000_cpu == PROCESSOR_PPC603
14972 || rs6000_cpu == PROCESSOR_PPC750
14973 || optimize_size);
14974
14975 if (WORLD_SAVE_P (info))
14976 {
14977 int i, j;
14978 char rname[30];
14979 const char *alloc_rname;
14980 rtvec p;
14981
14982 /* eh_rest_world_r10 will return to the location saved in the LR
14983 stack slot (which is not likely to be our caller.)
14984 Input: R10 -- stack adjustment. Clobbers R0, R11, R12, R7, R8.
14985 rest_world is similar, except any R10 parameter is ignored.
14986 The exception-handling stuff that was here in 2.95 is no
14987 longer necessary. */
14988
14989 p = rtvec_alloc (9
14990 + 1
14991 + 32 - info->first_gp_reg_save
14992 + LAST_ALTIVEC_REGNO + 1 - info->first_altivec_reg_save
14993 + 63 + 1 - info->first_fp_reg_save);
14994
14995 strcpy (rname, ((current_function_calls_eh_return) ?
14996 "*eh_rest_world_r10" : "*rest_world"));
14997 alloc_rname = ggc_strdup (rname);
14998
14999 j = 0;
15000 RTVEC_ELT (p, j++) = gen_rtx_RETURN (VOIDmode);
15001 RTVEC_ELT (p, j++) = gen_rtx_USE (VOIDmode,
15002 gen_rtx_REG (Pmode,
15003 LINK_REGISTER_REGNUM));
15004 RTVEC_ELT (p, j++)
15005 = gen_rtx_USE (VOIDmode, gen_rtx_SYMBOL_REF (Pmode, alloc_rname));
15006 /* The instruction pattern requires a clobber here;
15007 it is shared with the restVEC helper. */
15008 RTVEC_ELT (p, j++)
15009 = gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (Pmode, 11));
15010
15011 {
15012 /* CR register traditionally saved as CR2. */
15013 rtx reg = gen_rtx_REG (reg_mode, CR2_REGNO);
15014 rtx addr = gen_rtx_PLUS (Pmode, frame_reg_rtx,
15015 GEN_INT (info->cr_save_offset));
15016 rtx mem = gen_frame_mem (reg_mode, addr);
15017
15018 RTVEC_ELT (p, j++) = gen_rtx_SET (VOIDmode, reg, mem);
15019 }
15020
15021 for (i = 0; i < 32 - info->first_gp_reg_save; i++)
15022 {
15023 rtx reg = gen_rtx_REG (reg_mode, info->first_gp_reg_save + i);
15024 rtx addr = gen_rtx_PLUS (Pmode, frame_reg_rtx,
15025 GEN_INT (info->gp_save_offset
15026 + reg_size * i));
15027 rtx mem = gen_frame_mem (reg_mode, addr);
15028
15029 RTVEC_ELT (p, j++) = gen_rtx_SET (VOIDmode, reg, mem);
15030 }
15031 for (i = 0; info->first_altivec_reg_save + i <= LAST_ALTIVEC_REGNO; i++)
15032 {
15033 rtx reg = gen_rtx_REG (V4SImode, info->first_altivec_reg_save + i);
15034 rtx addr = gen_rtx_PLUS (Pmode, frame_reg_rtx,
15035 GEN_INT (info->altivec_save_offset
15036 + 16 * i));
15037 rtx mem = gen_frame_mem (V4SImode, addr);
15038
15039 RTVEC_ELT (p, j++) = gen_rtx_SET (VOIDmode, reg, mem);
15040 }
15041 for (i = 0; info->first_fp_reg_save + i <= 63; i++)
15042 {
15043 rtx reg = gen_rtx_REG (DFmode, info->first_fp_reg_save + i);
15044 rtx addr = gen_rtx_PLUS (Pmode, frame_reg_rtx,
15045 GEN_INT (info->fp_save_offset
15046 + 8 * i));
15047 rtx mem = gen_frame_mem (DFmode, addr);
15048
15049 RTVEC_ELT (p, j++) = gen_rtx_SET (VOIDmode, reg, mem);
15050 }
15051 RTVEC_ELT (p, j++)
15052 = gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (Pmode, 0));
15053 RTVEC_ELT (p, j++)
15054 = gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (SImode, 12));
15055 RTVEC_ELT (p, j++)
15056 = gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (SImode, 7));
15057 RTVEC_ELT (p, j++)
15058 = gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (SImode, 8));
15059 RTVEC_ELT (p, j++)
15060 = gen_rtx_USE (VOIDmode, gen_rtx_REG (SImode, 10));
15061 emit_jump_insn (gen_rtx_PARALLEL (VOIDmode, p));
15062
15063 return;
15064 }
15065
15066 /* Set sp_offset based on the stack push from the prologue. */
15067 if (info->total_size < 32767)
15068 sp_offset = info->total_size;
15069
15070 /* Restore AltiVec registers if needed. */
15071 if (TARGET_ALTIVEC_ABI && info->altivec_size != 0)
15072 {
15073 int i;
15074
15075 for (i = info->first_altivec_reg_save; i <= LAST_ALTIVEC_REGNO; ++i)
15076 if (info->vrsave_mask & ALTIVEC_REG_BIT (i))
15077 {
15078 rtx addr, areg, mem;
15079
15080 areg = gen_rtx_REG (Pmode, 0);
15081 emit_move_insn
15082 (areg, GEN_INT (info->altivec_save_offset
15083 + sp_offset
15084 + 16 * (i - info->first_altivec_reg_save)));
15085
15086 /* AltiVec addressing mode is [reg+reg]. */
15087 addr = gen_rtx_PLUS (Pmode, frame_reg_rtx, areg);
15088 mem = gen_frame_mem (V4SImode, addr);
15089
15090 emit_move_insn (gen_rtx_REG (V4SImode, i), mem);
15091 }
15092 }
15093
15094 /* Restore VRSAVE if needed. */
15095 if (TARGET_ALTIVEC && TARGET_ALTIVEC_VRSAVE
15096 && info->vrsave_mask != 0)
15097 {
15098 rtx addr, mem, reg;
15099
15100 addr = gen_rtx_PLUS (Pmode, frame_reg_rtx,
15101 GEN_INT (info->vrsave_save_offset + sp_offset));
15102 mem = gen_frame_mem (SImode, addr);
15103 reg = gen_rtx_REG (SImode, 12);
15104 emit_move_insn (reg, mem);
15105
15106 emit_insn (generate_set_vrsave (reg, info, 1));
15107 }
15108
15109 sp_offset = 0;
15110
15111 /* If we have a frame pointer, a call to alloca, or a large stack
15112 frame, restore the old stack pointer using the backchain. Otherwise,
15113 we know what size to update it with. */
15114 if (use_backchain_to_restore_sp)
15115 {
15116 /* Under V.4, don't reset the stack pointer until after we're done
15117 loading the saved registers. */
15118 if (DEFAULT_ABI == ABI_V4)
15119 frame_reg_rtx = gen_rtx_REG (Pmode, 11);
15120
15121 emit_move_insn (frame_reg_rtx,
15122 gen_rtx_MEM (Pmode, sp_reg_rtx));
15123 }
15124 else if (info->push_p)
15125 {
15126 if (DEFAULT_ABI == ABI_V4
15127 || current_function_calls_eh_return)
15128 sp_offset = info->total_size;
15129 else
15130 {
15131 emit_insn (TARGET_32BIT
15132 ? gen_addsi3 (sp_reg_rtx, sp_reg_rtx,
15133 GEN_INT (info->total_size))
15134 : gen_adddi3 (sp_reg_rtx, sp_reg_rtx,
15135 GEN_INT (info->total_size)));
15136 }
15137 }
15138
15139 /* Get the old lr if we saved it. */
15140 if (info->lr_save_p)
15141 {
15142 rtx mem = gen_frame_mem_offset (Pmode, frame_reg_rtx,
15143 info->lr_save_offset + sp_offset);
15144
15145 emit_move_insn (gen_rtx_REG (Pmode, 0), mem);
15146 }
15147
15148 /* Get the old cr if we saved it. */
15149 if (info->cr_save_p)
15150 {
15151 rtx addr = gen_rtx_PLUS (Pmode, frame_reg_rtx,
15152 GEN_INT (info->cr_save_offset + sp_offset));
15153 rtx mem = gen_frame_mem (SImode, addr);
15154
15155 emit_move_insn (gen_rtx_REG (SImode, 12), mem);
15156 }
15157
15158 /* Set LR here to try to overlap restores below. */
15159 if (info->lr_save_p)
15160 emit_move_insn (gen_rtx_REG (Pmode, LINK_REGISTER_REGNUM),
15161 gen_rtx_REG (Pmode, 0));
15162
15163 /* Load exception handler data registers, if needed. */
15164 if (current_function_calls_eh_return)
15165 {
15166 unsigned int i, regno;
15167
15168 if (TARGET_AIX)
15169 {
15170 rtx addr = gen_rtx_PLUS (Pmode, frame_reg_rtx,
15171 GEN_INT (sp_offset + 5 * reg_size));
15172 rtx mem = gen_frame_mem (reg_mode, addr);
15173
15174 emit_move_insn (gen_rtx_REG (reg_mode, 2), mem);
15175 }
15176
15177 for (i = 0; ; ++i)
15178 {
15179 rtx mem;
15180
15181 regno = EH_RETURN_DATA_REGNO (i);
15182 if (regno == INVALID_REGNUM)
15183 break;
15184
15185 mem = gen_frame_mem_offset (reg_mode, frame_reg_rtx,
15186 info->ehrd_offset + sp_offset
15187 + reg_size * (int) i);
15188
15189 emit_move_insn (gen_rtx_REG (reg_mode, regno), mem);
15190 }
15191 }
15192
15193 /* Restore GPRs. This is done as a PARALLEL if we are using
15194 the load-multiple instructions. */
15195 if (using_load_multiple)
15196 {
15197 rtvec p;
15198 p = rtvec_alloc (32 - info->first_gp_reg_save);
15199 for (i = 0; i < 32 - info->first_gp_reg_save; i++)
15200 {
15201 rtx addr = gen_rtx_PLUS (Pmode, frame_reg_rtx,
15202 GEN_INT (info->gp_save_offset
15203 + sp_offset
15204 + reg_size * i));
15205 rtx mem = gen_frame_mem (reg_mode, addr);
15206
15207 RTVEC_ELT (p, i) =
15208 gen_rtx_SET (VOIDmode,
15209 gen_rtx_REG (reg_mode, info->first_gp_reg_save + i),
15210 mem);
15211 }
15212 emit_insn (gen_rtx_PARALLEL (VOIDmode, p));
15213 }
15214 else
15215 for (i = 0; i < 32 - info->first_gp_reg_save; i++)
15216 if ((regs_ever_live[info->first_gp_reg_save + i]
15217 && (!call_used_regs[info->first_gp_reg_save + i]
15218 || (i + info->first_gp_reg_save == RS6000_PIC_OFFSET_TABLE_REGNUM
15219 && TARGET_TOC && TARGET_MINIMAL_TOC)))
15220 || (i + info->first_gp_reg_save == RS6000_PIC_OFFSET_TABLE_REGNUM
15221 && ((DEFAULT_ABI == ABI_V4 && flag_pic != 0)
15222 || (DEFAULT_ABI == ABI_DARWIN && flag_pic))))
15223 {
15224 rtx addr = gen_rtx_PLUS (Pmode, frame_reg_rtx,
15225 GEN_INT (info->gp_save_offset
15226 + sp_offset
15227 + reg_size * i));
15228 rtx mem = gen_frame_mem (reg_mode, addr);
15229
15230 /* Restore 64-bit quantities for SPE. */
15231 if (TARGET_SPE_ABI && info->spe_64bit_regs_used != 0)
15232 {
15233 int offset = info->spe_gp_save_offset + sp_offset + 8 * i;
15234 rtx b;
15235
15236 if (!SPE_CONST_OFFSET_OK (offset))
15237 {
15238 b = gen_rtx_REG (Pmode, FIXED_SCRATCH);
15239 emit_move_insn (b, GEN_INT (offset));
15240 }
15241 else
15242 b = GEN_INT (offset);
15243
15244 addr = gen_rtx_PLUS (Pmode, frame_reg_rtx, b);
15245 mem = gen_frame_mem (V2SImode, addr);
15246 }
15247
15248 emit_move_insn (gen_rtx_REG (reg_mode,
15249 info->first_gp_reg_save + i), mem);
15250 }
15251
15252 /* Restore fpr's if we need to do it without calling a function. */
15253 if (restoring_FPRs_inline)
15254 for (i = 0; i < 64 - info->first_fp_reg_save; i++)
15255 if ((regs_ever_live[info->first_fp_reg_save+i]
15256 && ! call_used_regs[info->first_fp_reg_save+i]))
15257 {
15258 rtx addr, mem;
15259 addr = gen_rtx_PLUS (Pmode, frame_reg_rtx,
15260 GEN_INT (info->fp_save_offset
15261 + sp_offset
15262 + 8 * i));
15263 mem = gen_frame_mem (DFmode, addr);
15264
15265 emit_move_insn (gen_rtx_REG (DFmode,
15266 info->first_fp_reg_save + i),
15267 mem);
15268 }
15269
15270 /* If we saved cr, restore it here. Just those that were used. */
15271 if (info->cr_save_p)
15272 {
15273 rtx r12_rtx = gen_rtx_REG (SImode, 12);
15274 int count = 0;
15275
15276 if (using_mfcr_multiple)
15277 {
15278 for (i = 0; i < 8; i++)
15279 if (regs_ever_live[CR0_REGNO+i] && ! call_used_regs[CR0_REGNO+i])
15280 count++;
15281 gcc_assert (count);
15282 }
15283
15284 if (using_mfcr_multiple && count > 1)
15285 {
15286 rtvec p;
15287 int ndx;
15288
15289 p = rtvec_alloc (count);
15290
15291 ndx = 0;
15292 for (i = 0; i < 8; i++)
15293 if (regs_ever_live[CR0_REGNO+i] && ! call_used_regs[CR0_REGNO+i])
15294 {
15295 rtvec r = rtvec_alloc (2);
15296 RTVEC_ELT (r, 0) = r12_rtx;
15297 RTVEC_ELT (r, 1) = GEN_INT (1 << (7-i));
15298 RTVEC_ELT (p, ndx) =
15299 gen_rtx_SET (VOIDmode, gen_rtx_REG (CCmode, CR0_REGNO+i),
15300 gen_rtx_UNSPEC (CCmode, r, UNSPEC_MOVESI_TO_CR));
15301 ndx++;
15302 }
15303 emit_insn (gen_rtx_PARALLEL (VOIDmode, p));
15304 gcc_assert (ndx == count);
15305 }
15306 else
15307 for (i = 0; i < 8; i++)
15308 if (regs_ever_live[CR0_REGNO+i] && ! call_used_regs[CR0_REGNO+i])
15309 {
15310 emit_insn (gen_movsi_to_cr_one (gen_rtx_REG (CCmode,
15311 CR0_REGNO+i),
15312 r12_rtx));
15313 }
15314 }
15315
15316 /* If this is V.4, unwind the stack pointer after all of the loads
15317 have been done. */
15318 if (frame_reg_rtx != sp_reg_rtx)
15319 {
15320 /* This blockage is needed so that sched doesn't decide to move
15321 the sp change before the register restores. */
15322 rs6000_emit_stack_tie ();
15323 emit_move_insn (sp_reg_rtx, frame_reg_rtx);
15324 }
15325 else if (sp_offset != 0)
15326 emit_insn (TARGET_32BIT
15327 ? gen_addsi3 (sp_reg_rtx, sp_reg_rtx,
15328 GEN_INT (sp_offset))
15329 : gen_adddi3 (sp_reg_rtx, sp_reg_rtx,
15330 GEN_INT (sp_offset)));
15331
15332 if (current_function_calls_eh_return)
15333 {
15334 rtx sa = EH_RETURN_STACKADJ_RTX;
15335 emit_insn (TARGET_32BIT
15336 ? gen_addsi3 (sp_reg_rtx, sp_reg_rtx, sa)
15337 : gen_adddi3 (sp_reg_rtx, sp_reg_rtx, sa));
15338 }
15339
15340 if (!sibcall)
15341 {
15342 rtvec p;
15343 if (! restoring_FPRs_inline)
15344 p = rtvec_alloc (3 + 64 - info->first_fp_reg_save);
15345 else
15346 p = rtvec_alloc (2);
15347
15348 RTVEC_ELT (p, 0) = gen_rtx_RETURN (VOIDmode);
15349 RTVEC_ELT (p, 1) = gen_rtx_USE (VOIDmode,
15350 gen_rtx_REG (Pmode,
15351 LINK_REGISTER_REGNUM));
15352
15353 /* If we have to restore more than two FP registers, branch to the
15354 restore function. It will return to our caller. */
15355 if (! restoring_FPRs_inline)
15356 {
15357 int i;
15358 char rname[30];
15359 const char *alloc_rname;
15360
15361 sprintf (rname, "%s%d%s", RESTORE_FP_PREFIX,
15362 info->first_fp_reg_save - 32, RESTORE_FP_SUFFIX);
15363 alloc_rname = ggc_strdup (rname);
15364 RTVEC_ELT (p, 2) = gen_rtx_USE (VOIDmode,
15365 gen_rtx_SYMBOL_REF (Pmode,
15366 alloc_rname));
15367
15368 for (i = 0; i < 64 - info->first_fp_reg_save; i++)
15369 {
15370 rtx addr, mem;
15371 addr = gen_rtx_PLUS (Pmode, sp_reg_rtx,
15372 GEN_INT (info->fp_save_offset + 8*i));
15373 mem = gen_frame_mem (DFmode, addr);
15374
15375 RTVEC_ELT (p, i+3) =
15376 gen_rtx_SET (VOIDmode,
15377 gen_rtx_REG (DFmode, info->first_fp_reg_save + i),
15378 mem);
15379 }
15380 }
15381
15382 emit_jump_insn (gen_rtx_PARALLEL (VOIDmode, p));
15383 }
15384 }
15385
15386 /* Write function epilogue. */
15387
15388 static void
rs6000_output_function_epilogue(FILE * file,HOST_WIDE_INT size ATTRIBUTE_UNUSED)15389 rs6000_output_function_epilogue (FILE *file,
15390 HOST_WIDE_INT size ATTRIBUTE_UNUSED)
15391 {
15392 if (! HAVE_epilogue)
15393 {
15394 rtx insn = get_last_insn ();
15395 /* If the last insn was a BARRIER, we don't have to write anything except
15396 the trace table. */
15397 if (GET_CODE (insn) == NOTE)
15398 insn = prev_nonnote_insn (insn);
15399 if (insn == 0 || GET_CODE (insn) != BARRIER)
15400 {
15401 /* This is slightly ugly, but at least we don't have two
15402 copies of the epilogue-emitting code. */
15403 start_sequence ();
15404
15405 /* A NOTE_INSN_DELETED is supposed to be at the start
15406 and end of the "toplevel" insn chain. */
15407 emit_note (NOTE_INSN_DELETED);
15408 rs6000_emit_epilogue (FALSE);
15409 emit_note (NOTE_INSN_DELETED);
15410
15411 /* Expand INSN_ADDRESSES so final() doesn't crash. */
15412 {
15413 rtx insn;
15414 unsigned addr = 0;
15415 for (insn = get_insns (); insn != 0; insn = NEXT_INSN (insn))
15416 {
15417 INSN_ADDRESSES_NEW (insn, addr);
15418 addr += 4;
15419 }
15420 }
15421
15422 if (TARGET_DEBUG_STACK)
15423 debug_rtx_list (get_insns (), 100);
15424 final (get_insns (), file, FALSE);
15425 end_sequence ();
15426 }
15427 }
15428
15429 #if TARGET_MACHO
15430 macho_branch_islands ();
15431 /* Mach-O doesn't support labels at the end of objects, so if
15432 it looks like we might want one, insert a NOP. */
15433 {
15434 rtx insn = get_last_insn ();
15435 while (insn
15436 && NOTE_P (insn)
15437 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_DELETED_LABEL)
15438 insn = PREV_INSN (insn);
15439 if (insn
15440 && (LABEL_P (insn)
15441 || (NOTE_P (insn)
15442 && NOTE_LINE_NUMBER (insn) == NOTE_INSN_DELETED_LABEL)))
15443 fputs ("\tnop\n", file);
15444 }
15445 #endif
15446
15447 /* Output a traceback table here. See /usr/include/sys/debug.h for info
15448 on its format.
15449
15450 We don't output a traceback table if -finhibit-size-directive was
15451 used. The documentation for -finhibit-size-directive reads
15452 ``don't output a @code{.size} assembler directive, or anything
15453 else that would cause trouble if the function is split in the
15454 middle, and the two halves are placed at locations far apart in
15455 memory.'' The traceback table has this property, since it
15456 includes the offset from the start of the function to the
15457 traceback table itself.
15458
15459 System V.4 Powerpc's (and the embedded ABI derived from it) use a
15460 different traceback table. */
15461 if (DEFAULT_ABI == ABI_AIX && ! flag_inhibit_size_directive
15462 && rs6000_traceback != traceback_none && !current_function_is_thunk)
15463 {
15464 const char *fname = NULL;
15465 const char *language_string = lang_hooks.name;
15466 int fixed_parms = 0, float_parms = 0, parm_info = 0;
15467 int i;
15468 int optional_tbtab;
15469 rs6000_stack_t *info = rs6000_stack_info ();
15470
15471 if (rs6000_traceback == traceback_full)
15472 optional_tbtab = 1;
15473 else if (rs6000_traceback == traceback_part)
15474 optional_tbtab = 0;
15475 else
15476 optional_tbtab = !optimize_size && !TARGET_ELF;
15477
15478 if (optional_tbtab)
15479 {
15480 fname = XSTR (XEXP (DECL_RTL (current_function_decl), 0), 0);
15481 while (*fname == '.') /* V.4 encodes . in the name */
15482 fname++;
15483
15484 /* Need label immediately before tbtab, so we can compute
15485 its offset from the function start. */
15486 ASM_OUTPUT_INTERNAL_LABEL_PREFIX (file, "LT");
15487 ASM_OUTPUT_LABEL (file, fname);
15488 }
15489
15490 /* The .tbtab pseudo-op can only be used for the first eight
15491 expressions, since it can't handle the possibly variable
15492 length fields that follow. However, if you omit the optional
15493 fields, the assembler outputs zeros for all optional fields
15494 anyways, giving each variable length field is minimum length
15495 (as defined in sys/debug.h). Thus we can not use the .tbtab
15496 pseudo-op at all. */
15497
15498 /* An all-zero word flags the start of the tbtab, for debuggers
15499 that have to find it by searching forward from the entry
15500 point or from the current pc. */
15501 fputs ("\t.long 0\n", file);
15502
15503 /* Tbtab format type. Use format type 0. */
15504 fputs ("\t.byte 0,", file);
15505
15506 /* Language type. Unfortunately, there does not seem to be any
15507 official way to discover the language being compiled, so we
15508 use language_string.
15509 C is 0. Fortran is 1. Pascal is 2. Ada is 3. C++ is 9.
15510 Java is 13. Objective-C is 14. Objective-C++ isn't assigned
15511 a number, so for now use 9. */
15512 if (! strcmp (language_string, "GNU C"))
15513 i = 0;
15514 else if (! strcmp (language_string, "GNU F77")
15515 || ! strcmp (language_string, "GNU F95"))
15516 i = 1;
15517 else if (! strcmp (language_string, "GNU Pascal"))
15518 i = 2;
15519 else if (! strcmp (language_string, "GNU Ada"))
15520 i = 3;
15521 else if (! strcmp (language_string, "GNU C++")
15522 || ! strcmp (language_string, "GNU Objective-C++"))
15523 i = 9;
15524 else if (! strcmp (language_string, "GNU Java"))
15525 i = 13;
15526 else if (! strcmp (language_string, "GNU Objective-C"))
15527 i = 14;
15528 else
15529 gcc_unreachable ();
15530 fprintf (file, "%d,", i);
15531
15532 /* 8 single bit fields: global linkage (not set for C extern linkage,
15533 apparently a PL/I convention?), out-of-line epilogue/prologue, offset
15534 from start of procedure stored in tbtab, internal function, function
15535 has controlled storage, function has no toc, function uses fp,
15536 function logs/aborts fp operations. */
15537 /* Assume that fp operations are used if any fp reg must be saved. */
15538 fprintf (file, "%d,",
15539 (optional_tbtab << 5) | ((info->first_fp_reg_save != 64) << 1));
15540
15541 /* 6 bitfields: function is interrupt handler, name present in
15542 proc table, function calls alloca, on condition directives
15543 (controls stack walks, 3 bits), saves condition reg, saves
15544 link reg. */
15545 /* The `function calls alloca' bit seems to be set whenever reg 31 is
15546 set up as a frame pointer, even when there is no alloca call. */
15547 fprintf (file, "%d,",
15548 ((optional_tbtab << 6)
15549 | ((optional_tbtab & frame_pointer_needed) << 5)
15550 | (info->cr_save_p << 1)
15551 | (info->lr_save_p)));
15552
15553 /* 3 bitfields: saves backchain, fixup code, number of fpr saved
15554 (6 bits). */
15555 fprintf (file, "%d,",
15556 (info->push_p << 7) | (64 - info->first_fp_reg_save));
15557
15558 /* 2 bitfields: spare bits (2 bits), number of gpr saved (6 bits). */
15559 fprintf (file, "%d,", (32 - first_reg_to_save ()));
15560
15561 if (optional_tbtab)
15562 {
15563 /* Compute the parameter info from the function decl argument
15564 list. */
15565 tree decl;
15566 int next_parm_info_bit = 31;
15567
15568 for (decl = DECL_ARGUMENTS (current_function_decl);
15569 decl; decl = TREE_CHAIN (decl))
15570 {
15571 rtx parameter = DECL_INCOMING_RTL (decl);
15572 enum machine_mode mode = GET_MODE (parameter);
15573
15574 if (GET_CODE (parameter) == REG)
15575 {
15576 if (SCALAR_FLOAT_MODE_P (mode))
15577 {
15578 int bits;
15579
15580 float_parms++;
15581
15582 switch (mode)
15583 {
15584 case SFmode:
15585 bits = 0x2;
15586 break;
15587
15588 case DFmode:
15589 case TFmode:
15590 bits = 0x3;
15591 break;
15592
15593 default:
15594 gcc_unreachable ();
15595 }
15596
15597 /* If only one bit will fit, don't or in this entry. */
15598 if (next_parm_info_bit > 0)
15599 parm_info |= (bits << (next_parm_info_bit - 1));
15600 next_parm_info_bit -= 2;
15601 }
15602 else
15603 {
15604 fixed_parms += ((GET_MODE_SIZE (mode)
15605 + (UNITS_PER_WORD - 1))
15606 / UNITS_PER_WORD);
15607 next_parm_info_bit -= 1;
15608 }
15609 }
15610 }
15611 }
15612
15613 /* Number of fixed point parameters. */
15614 /* This is actually the number of words of fixed point parameters; thus
15615 an 8 byte struct counts as 2; and thus the maximum value is 8. */
15616 fprintf (file, "%d,", fixed_parms);
15617
15618 /* 2 bitfields: number of floating point parameters (7 bits), parameters
15619 all on stack. */
15620 /* This is actually the number of fp registers that hold parameters;
15621 and thus the maximum value is 13. */
15622 /* Set parameters on stack bit if parameters are not in their original
15623 registers, regardless of whether they are on the stack? Xlc
15624 seems to set the bit when not optimizing. */
15625 fprintf (file, "%d\n", ((float_parms << 1) | (! optimize)));
15626
15627 if (! optional_tbtab)
15628 return;
15629
15630 /* Optional fields follow. Some are variable length. */
15631
15632 /* Parameter types, left adjusted bit fields: 0 fixed, 10 single float,
15633 11 double float. */
15634 /* There is an entry for each parameter in a register, in the order that
15635 they occur in the parameter list. Any intervening arguments on the
15636 stack are ignored. If the list overflows a long (max possible length
15637 34 bits) then completely leave off all elements that don't fit. */
15638 /* Only emit this long if there was at least one parameter. */
15639 if (fixed_parms || float_parms)
15640 fprintf (file, "\t.long %d\n", parm_info);
15641
15642 /* Offset from start of code to tb table. */
15643 fputs ("\t.long ", file);
15644 ASM_OUTPUT_INTERNAL_LABEL_PREFIX (file, "LT");
15645 if (TARGET_AIX)
15646 RS6000_OUTPUT_BASENAME (file, fname);
15647 else
15648 assemble_name (file, fname);
15649 putc ('-', file);
15650 rs6000_output_function_entry (file, fname);
15651 putc ('\n', file);
15652
15653 /* Interrupt handler mask. */
15654 /* Omit this long, since we never set the interrupt handler bit
15655 above. */
15656
15657 /* Number of CTL (controlled storage) anchors. */
15658 /* Omit this long, since the has_ctl bit is never set above. */
15659
15660 /* Displacement into stack of each CTL anchor. */
15661 /* Omit this list of longs, because there are no CTL anchors. */
15662
15663 /* Length of function name. */
15664 if (*fname == '*')
15665 ++fname;
15666 fprintf (file, "\t.short %d\n", (int) strlen (fname));
15667
15668 /* Function name. */
15669 assemble_string (fname, strlen (fname));
15670
15671 /* Register for alloca automatic storage; this is always reg 31.
15672 Only emit this if the alloca bit was set above. */
15673 if (frame_pointer_needed)
15674 fputs ("\t.byte 31\n", file);
15675
15676 fputs ("\t.align 2\n", file);
15677 }
15678 }
15679
15680 /* A C compound statement that outputs the assembler code for a thunk
15681 function, used to implement C++ virtual function calls with
15682 multiple inheritance. The thunk acts as a wrapper around a virtual
15683 function, adjusting the implicit object parameter before handing
15684 control off to the real function.
15685
15686 First, emit code to add the integer DELTA to the location that
15687 contains the incoming first argument. Assume that this argument
15688 contains a pointer, and is the one used to pass the `this' pointer
15689 in C++. This is the incoming argument *before* the function
15690 prologue, e.g. `%o0' on a sparc. The addition must preserve the
15691 values of all other incoming arguments.
15692
15693 After the addition, emit code to jump to FUNCTION, which is a
15694 `FUNCTION_DECL'. This is a direct pure jump, not a call, and does
15695 not touch the return address. Hence returning from FUNCTION will
15696 return to whoever called the current `thunk'.
15697
15698 The effect must be as if FUNCTION had been called directly with the
15699 adjusted first argument. This macro is responsible for emitting
15700 all of the code for a thunk function; output_function_prologue()
15701 and output_function_epilogue() are not invoked.
15702
15703 The THUNK_FNDECL is redundant. (DELTA and FUNCTION have already
15704 been extracted from it.) It might possibly be useful on some
15705 targets, but probably not.
15706
15707 If you do not define this macro, the target-independent code in the
15708 C++ frontend will generate a less efficient heavyweight thunk that
15709 calls FUNCTION instead of jumping to it. The generic approach does
15710 not support varargs. */
15711
15712 static void
rs6000_output_mi_thunk(FILE * file,tree thunk_fndecl ATTRIBUTE_UNUSED,HOST_WIDE_INT delta,HOST_WIDE_INT vcall_offset,tree function)15713 rs6000_output_mi_thunk (FILE *file, tree thunk_fndecl ATTRIBUTE_UNUSED,
15714 HOST_WIDE_INT delta, HOST_WIDE_INT vcall_offset,
15715 tree function)
15716 {
15717 rtx this, insn, funexp;
15718
15719 reload_completed = 1;
15720 epilogue_completed = 1;
15721 no_new_pseudos = 1;
15722 reset_block_changes ();
15723
15724 /* Mark the end of the (empty) prologue. */
15725 emit_note (NOTE_INSN_PROLOGUE_END);
15726
15727 /* Find the "this" pointer. If the function returns a structure,
15728 the structure return pointer is in r3. */
15729 if (aggregate_value_p (TREE_TYPE (TREE_TYPE (function)), function))
15730 this = gen_rtx_REG (Pmode, 4);
15731 else
15732 this = gen_rtx_REG (Pmode, 3);
15733
15734 /* Apply the constant offset, if required. */
15735 if (delta)
15736 {
15737 rtx delta_rtx = GEN_INT (delta);
15738 emit_insn (TARGET_32BIT
15739 ? gen_addsi3 (this, this, delta_rtx)
15740 : gen_adddi3 (this, this, delta_rtx));
15741 }
15742
15743 /* Apply the offset from the vtable, if required. */
15744 if (vcall_offset)
15745 {
15746 rtx vcall_offset_rtx = GEN_INT (vcall_offset);
15747 rtx tmp = gen_rtx_REG (Pmode, 12);
15748
15749 emit_move_insn (tmp, gen_rtx_MEM (Pmode, this));
15750 if (((unsigned HOST_WIDE_INT) vcall_offset) + 0x8000 >= 0x10000)
15751 {
15752 emit_insn (TARGET_32BIT
15753 ? gen_addsi3 (tmp, tmp, vcall_offset_rtx)
15754 : gen_adddi3 (tmp, tmp, vcall_offset_rtx));
15755 emit_move_insn (tmp, gen_rtx_MEM (Pmode, tmp));
15756 }
15757 else
15758 {
15759 rtx loc = gen_rtx_PLUS (Pmode, tmp, vcall_offset_rtx);
15760
15761 emit_move_insn (tmp, gen_rtx_MEM (Pmode, loc));
15762 }
15763 emit_insn (TARGET_32BIT
15764 ? gen_addsi3 (this, this, tmp)
15765 : gen_adddi3 (this, this, tmp));
15766 }
15767
15768 /* Generate a tail call to the target function. */
15769 if (!TREE_USED (function))
15770 {
15771 assemble_external (function);
15772 TREE_USED (function) = 1;
15773 }
15774 funexp = XEXP (DECL_RTL (function), 0);
15775 funexp = gen_rtx_MEM (FUNCTION_MODE, funexp);
15776
15777 #if TARGET_MACHO
15778 if (MACHOPIC_INDIRECT)
15779 funexp = machopic_indirect_call_target (funexp);
15780 #endif
15781
15782 /* gen_sibcall expects reload to convert scratch pseudo to LR so we must
15783 generate sibcall RTL explicitly. */
15784 insn = emit_call_insn (
15785 gen_rtx_PARALLEL (VOIDmode,
15786 gen_rtvec (4,
15787 gen_rtx_CALL (VOIDmode,
15788 funexp, const0_rtx),
15789 gen_rtx_USE (VOIDmode, const0_rtx),
15790 gen_rtx_USE (VOIDmode,
15791 gen_rtx_REG (SImode,
15792 LINK_REGISTER_REGNUM)),
15793 gen_rtx_RETURN (VOIDmode))));
15794 SIBLING_CALL_P (insn) = 1;
15795 emit_barrier ();
15796
15797 /* Run just enough of rest_of_compilation to get the insns emitted.
15798 There's not really enough bulk here to make other passes such as
15799 instruction scheduling worth while. Note that use_thunk calls
15800 assemble_start_function and assemble_end_function. */
15801 insn = get_insns ();
15802 insn_locators_initialize ();
15803 shorten_branches (insn);
15804 final_start_function (insn, file, 1);
15805 final (insn, file, 1);
15806 final_end_function ();
15807
15808 reload_completed = 0;
15809 epilogue_completed = 0;
15810 no_new_pseudos = 0;
15811 }
15812
15813 /* A quick summary of the various types of 'constant-pool tables'
15814 under PowerPC:
15815
15816 Target Flags Name One table per
15817 AIX (none) AIX TOC object file
15818 AIX -mfull-toc AIX TOC object file
15819 AIX -mminimal-toc AIX minimal TOC translation unit
15820 SVR4/EABI (none) SVR4 SDATA object file
15821 SVR4/EABI -fpic SVR4 pic object file
15822 SVR4/EABI -fPIC SVR4 PIC translation unit
15823 SVR4/EABI -mrelocatable EABI TOC function
15824 SVR4/EABI -maix AIX TOC object file
15825 SVR4/EABI -maix -mminimal-toc
15826 AIX minimal TOC translation unit
15827
15828 Name Reg. Set by entries contains:
15829 made by addrs? fp? sum?
15830
15831 AIX TOC 2 crt0 as Y option option
15832 AIX minimal TOC 30 prolog gcc Y Y option
15833 SVR4 SDATA 13 crt0 gcc N Y N
15834 SVR4 pic 30 prolog ld Y not yet N
15835 SVR4 PIC 30 prolog gcc Y option option
15836 EABI TOC 30 prolog gcc Y option option
15837
15838 */
15839
15840 /* Hash functions for the hash table. */
15841
15842 static unsigned
rs6000_hash_constant(rtx k)15843 rs6000_hash_constant (rtx k)
15844 {
15845 enum rtx_code code = GET_CODE (k);
15846 enum machine_mode mode = GET_MODE (k);
15847 unsigned result = (code << 3) ^ mode;
15848 const char *format;
15849 int flen, fidx;
15850
15851 format = GET_RTX_FORMAT (code);
15852 flen = strlen (format);
15853 fidx = 0;
15854
15855 switch (code)
15856 {
15857 case LABEL_REF:
15858 return result * 1231 + (unsigned) INSN_UID (XEXP (k, 0));
15859
15860 case CONST_DOUBLE:
15861 if (mode != VOIDmode)
15862 return real_hash (CONST_DOUBLE_REAL_VALUE (k)) * result;
15863 flen = 2;
15864 break;
15865
15866 case CODE_LABEL:
15867 fidx = 3;
15868 break;
15869
15870 default:
15871 break;
15872 }
15873
15874 for (; fidx < flen; fidx++)
15875 switch (format[fidx])
15876 {
15877 case 's':
15878 {
15879 unsigned i, len;
15880 const char *str = XSTR (k, fidx);
15881 len = strlen (str);
15882 result = result * 613 + len;
15883 for (i = 0; i < len; i++)
15884 result = result * 613 + (unsigned) str[i];
15885 break;
15886 }
15887 case 'u':
15888 case 'e':
15889 result = result * 1231 + rs6000_hash_constant (XEXP (k, fidx));
15890 break;
15891 case 'i':
15892 case 'n':
15893 result = result * 613 + (unsigned) XINT (k, fidx);
15894 break;
15895 case 'w':
15896 if (sizeof (unsigned) >= sizeof (HOST_WIDE_INT))
15897 result = result * 613 + (unsigned) XWINT (k, fidx);
15898 else
15899 {
15900 size_t i;
15901 for (i = 0; i < sizeof (HOST_WIDE_INT) / sizeof (unsigned); i++)
15902 result = result * 613 + (unsigned) (XWINT (k, fidx)
15903 >> CHAR_BIT * i);
15904 }
15905 break;
15906 case '0':
15907 break;
15908 default:
15909 gcc_unreachable ();
15910 }
15911
15912 return result;
15913 }
15914
15915 static unsigned
toc_hash_function(const void * hash_entry)15916 toc_hash_function (const void *hash_entry)
15917 {
15918 const struct toc_hash_struct *thc =
15919 (const struct toc_hash_struct *) hash_entry;
15920 return rs6000_hash_constant (thc->key) ^ thc->key_mode;
15921 }
15922
15923 /* Compare H1 and H2 for equivalence. */
15924
15925 static int
toc_hash_eq(const void * h1,const void * h2)15926 toc_hash_eq (const void *h1, const void *h2)
15927 {
15928 rtx r1 = ((const struct toc_hash_struct *) h1)->key;
15929 rtx r2 = ((const struct toc_hash_struct *) h2)->key;
15930
15931 if (((const struct toc_hash_struct *) h1)->key_mode
15932 != ((const struct toc_hash_struct *) h2)->key_mode)
15933 return 0;
15934
15935 return rtx_equal_p (r1, r2);
15936 }
15937
15938 /* These are the names given by the C++ front-end to vtables, and
15939 vtable-like objects. Ideally, this logic should not be here;
15940 instead, there should be some programmatic way of inquiring as
15941 to whether or not an object is a vtable. */
15942
15943 #define VTABLE_NAME_P(NAME) \
15944 (strncmp ("_vt.", name, strlen ("_vt.")) == 0 \
15945 || strncmp ("_ZTV", name, strlen ("_ZTV")) == 0 \
15946 || strncmp ("_ZTT", name, strlen ("_ZTT")) == 0 \
15947 || strncmp ("_ZTI", name, strlen ("_ZTI")) == 0 \
15948 || strncmp ("_ZTC", name, strlen ("_ZTC")) == 0)
15949
15950 void
rs6000_output_symbol_ref(FILE * file,rtx x)15951 rs6000_output_symbol_ref (FILE *file, rtx x)
15952 {
15953 /* Currently C++ toc references to vtables can be emitted before it
15954 is decided whether the vtable is public or private. If this is
15955 the case, then the linker will eventually complain that there is
15956 a reference to an unknown section. Thus, for vtables only,
15957 we emit the TOC reference to reference the symbol and not the
15958 section. */
15959 const char *name = XSTR (x, 0);
15960
15961 if (VTABLE_NAME_P (name))
15962 {
15963 RS6000_OUTPUT_BASENAME (file, name);
15964 }
15965 else
15966 assemble_name (file, name);
15967 }
15968
15969 /* Output a TOC entry. We derive the entry name from what is being
15970 written. */
15971
15972 void
output_toc(FILE * file,rtx x,int labelno,enum machine_mode mode)15973 output_toc (FILE *file, rtx x, int labelno, enum machine_mode mode)
15974 {
15975 char buf[256];
15976 const char *name = buf;
15977 const char *real_name;
15978 rtx base = x;
15979 HOST_WIDE_INT offset = 0;
15980
15981 gcc_assert (!TARGET_NO_TOC);
15982
15983 /* When the linker won't eliminate them, don't output duplicate
15984 TOC entries (this happens on AIX if there is any kind of TOC,
15985 and on SVR4 under -fPIC or -mrelocatable). Don't do this for
15986 CODE_LABELs. */
15987 if (TARGET_TOC && GET_CODE (x) != LABEL_REF)
15988 {
15989 struct toc_hash_struct *h;
15990 void * * found;
15991
15992 /* Create toc_hash_table. This can't be done at OVERRIDE_OPTIONS
15993 time because GGC is not initialized at that point. */
15994 if (toc_hash_table == NULL)
15995 toc_hash_table = htab_create_ggc (1021, toc_hash_function,
15996 toc_hash_eq, NULL);
15997
15998 h = ggc_alloc (sizeof (*h));
15999 h->key = x;
16000 h->key_mode = mode;
16001 h->labelno = labelno;
16002
16003 found = htab_find_slot (toc_hash_table, h, 1);
16004 if (*found == NULL)
16005 *found = h;
16006 else /* This is indeed a duplicate.
16007 Set this label equal to that label. */
16008 {
16009 fputs ("\t.set ", file);
16010 ASM_OUTPUT_INTERNAL_LABEL_PREFIX (file, "LC");
16011 fprintf (file, "%d,", labelno);
16012 ASM_OUTPUT_INTERNAL_LABEL_PREFIX (file, "LC");
16013 fprintf (file, "%d\n", ((*(const struct toc_hash_struct **)
16014 found)->labelno));
16015 return;
16016 }
16017 }
16018
16019 /* If we're going to put a double constant in the TOC, make sure it's
16020 aligned properly when strict alignment is on. */
16021 if (GET_CODE (x) == CONST_DOUBLE
16022 && STRICT_ALIGNMENT
16023 && GET_MODE_BITSIZE (mode) >= 64
16024 && ! (TARGET_NO_FP_IN_TOC && ! TARGET_MINIMAL_TOC)) {
16025 ASM_OUTPUT_ALIGN (file, 3);
16026 }
16027
16028 (*targetm.asm_out.internal_label) (file, "LC", labelno);
16029
16030 /* Handle FP constants specially. Note that if we have a minimal
16031 TOC, things we put here aren't actually in the TOC, so we can allow
16032 FP constants. */
16033 if (GET_CODE (x) == CONST_DOUBLE &&
16034 (GET_MODE (x) == TFmode || GET_MODE (x) == TDmode))
16035 {
16036 REAL_VALUE_TYPE rv;
16037 long k[4];
16038
16039 REAL_VALUE_FROM_CONST_DOUBLE (rv, x);
16040 if (DECIMAL_FLOAT_MODE_P (GET_MODE (x)))
16041 REAL_VALUE_TO_TARGET_DECIMAL128 (rv, k);
16042 else
16043 REAL_VALUE_TO_TARGET_LONG_DOUBLE (rv, k);
16044
16045 if (TARGET_64BIT)
16046 {
16047 if (TARGET_MINIMAL_TOC)
16048 fputs (DOUBLE_INT_ASM_OP, file);
16049 else
16050 fprintf (file, "\t.tc FT_%lx_%lx_%lx_%lx[TC],",
16051 k[0] & 0xffffffff, k[1] & 0xffffffff,
16052 k[2] & 0xffffffff, k[3] & 0xffffffff);
16053 fprintf (file, "0x%lx%08lx,0x%lx%08lx\n",
16054 k[0] & 0xffffffff, k[1] & 0xffffffff,
16055 k[2] & 0xffffffff, k[3] & 0xffffffff);
16056 return;
16057 }
16058 else
16059 {
16060 if (TARGET_MINIMAL_TOC)
16061 fputs ("\t.long ", file);
16062 else
16063 fprintf (file, "\t.tc FT_%lx_%lx_%lx_%lx[TC],",
16064 k[0] & 0xffffffff, k[1] & 0xffffffff,
16065 k[2] & 0xffffffff, k[3] & 0xffffffff);
16066 fprintf (file, "0x%lx,0x%lx,0x%lx,0x%lx\n",
16067 k[0] & 0xffffffff, k[1] & 0xffffffff,
16068 k[2] & 0xffffffff, k[3] & 0xffffffff);
16069 return;
16070 }
16071 }
16072 else if (GET_CODE (x) == CONST_DOUBLE &&
16073 (GET_MODE (x) == DFmode || GET_MODE (x) == DDmode))
16074 {
16075 REAL_VALUE_TYPE rv;
16076 long k[2];
16077
16078 REAL_VALUE_FROM_CONST_DOUBLE (rv, x);
16079
16080 if (DECIMAL_FLOAT_MODE_P (GET_MODE (x)))
16081 REAL_VALUE_TO_TARGET_DECIMAL64 (rv, k);
16082 else
16083 REAL_VALUE_TO_TARGET_DOUBLE (rv, k);
16084
16085 if (TARGET_64BIT)
16086 {
16087 if (TARGET_MINIMAL_TOC)
16088 fputs (DOUBLE_INT_ASM_OP, file);
16089 else
16090 fprintf (file, "\t.tc FD_%lx_%lx[TC],",
16091 k[0] & 0xffffffff, k[1] & 0xffffffff);
16092 fprintf (file, "0x%lx%08lx\n",
16093 k[0] & 0xffffffff, k[1] & 0xffffffff);
16094 return;
16095 }
16096 else
16097 {
16098 if (TARGET_MINIMAL_TOC)
16099 fputs ("\t.long ", file);
16100 else
16101 fprintf (file, "\t.tc FD_%lx_%lx[TC],",
16102 k[0] & 0xffffffff, k[1] & 0xffffffff);
16103 fprintf (file, "0x%lx,0x%lx\n",
16104 k[0] & 0xffffffff, k[1] & 0xffffffff);
16105 return;
16106 }
16107 }
16108 else if (GET_CODE (x) == CONST_DOUBLE &&
16109 (GET_MODE (x) == SFmode || GET_MODE (x) == SDmode))
16110 {
16111 REAL_VALUE_TYPE rv;
16112 long l;
16113
16114 REAL_VALUE_FROM_CONST_DOUBLE (rv, x);
16115 if (DECIMAL_FLOAT_MODE_P (GET_MODE (x)))
16116 REAL_VALUE_TO_TARGET_DECIMAL32 (rv, l);
16117 else
16118 REAL_VALUE_TO_TARGET_SINGLE (rv, l);
16119
16120 if (TARGET_64BIT)
16121 {
16122 if (TARGET_MINIMAL_TOC)
16123 fputs (DOUBLE_INT_ASM_OP, file);
16124 else
16125 fprintf (file, "\t.tc FS_%lx[TC],", l & 0xffffffff);
16126 fprintf (file, "0x%lx00000000\n", l & 0xffffffff);
16127 return;
16128 }
16129 else
16130 {
16131 if (TARGET_MINIMAL_TOC)
16132 fputs ("\t.long ", file);
16133 else
16134 fprintf (file, "\t.tc FS_%lx[TC],", l & 0xffffffff);
16135 fprintf (file, "0x%lx\n", l & 0xffffffff);
16136 return;
16137 }
16138 }
16139 else if (GET_MODE (x) == VOIDmode
16140 && (GET_CODE (x) == CONST_INT || GET_CODE (x) == CONST_DOUBLE))
16141 {
16142 unsigned HOST_WIDE_INT low;
16143 HOST_WIDE_INT high;
16144
16145 if (GET_CODE (x) == CONST_DOUBLE)
16146 {
16147 low = CONST_DOUBLE_LOW (x);
16148 high = CONST_DOUBLE_HIGH (x);
16149 }
16150 else
16151 #if HOST_BITS_PER_WIDE_INT == 32
16152 {
16153 low = INTVAL (x);
16154 high = (low & 0x80000000) ? ~0 : 0;
16155 }
16156 #else
16157 {
16158 low = INTVAL (x) & 0xffffffff;
16159 high = (HOST_WIDE_INT) INTVAL (x) >> 32;
16160 }
16161 #endif
16162
16163 /* TOC entries are always Pmode-sized, but since this
16164 is a bigendian machine then if we're putting smaller
16165 integer constants in the TOC we have to pad them.
16166 (This is still a win over putting the constants in
16167 a separate constant pool, because then we'd have
16168 to have both a TOC entry _and_ the actual constant.)
16169
16170 For a 32-bit target, CONST_INT values are loaded and shifted
16171 entirely within `low' and can be stored in one TOC entry. */
16172
16173 /* It would be easy to make this work, but it doesn't now. */
16174 gcc_assert (!TARGET_64BIT || POINTER_SIZE >= GET_MODE_BITSIZE (mode));
16175
16176 if (POINTER_SIZE > GET_MODE_BITSIZE (mode))
16177 {
16178 #if HOST_BITS_PER_WIDE_INT == 32
16179 lshift_double (low, high, POINTER_SIZE - GET_MODE_BITSIZE (mode),
16180 POINTER_SIZE, &low, &high, 0);
16181 #else
16182 low |= high << 32;
16183 low <<= POINTER_SIZE - GET_MODE_BITSIZE (mode);
16184 high = (HOST_WIDE_INT) low >> 32;
16185 low &= 0xffffffff;
16186 #endif
16187 }
16188
16189 if (TARGET_64BIT)
16190 {
16191 if (TARGET_MINIMAL_TOC)
16192 fputs (DOUBLE_INT_ASM_OP, file);
16193 else
16194 fprintf (file, "\t.tc ID_%lx_%lx[TC],",
16195 (long) high & 0xffffffff, (long) low & 0xffffffff);
16196 fprintf (file, "0x%lx%08lx\n",
16197 (long) high & 0xffffffff, (long) low & 0xffffffff);
16198 return;
16199 }
16200 else
16201 {
16202 if (POINTER_SIZE < GET_MODE_BITSIZE (mode))
16203 {
16204 if (TARGET_MINIMAL_TOC)
16205 fputs ("\t.long ", file);
16206 else
16207 fprintf (file, "\t.tc ID_%lx_%lx[TC],",
16208 (long) high & 0xffffffff, (long) low & 0xffffffff);
16209 fprintf (file, "0x%lx,0x%lx\n",
16210 (long) high & 0xffffffff, (long) low & 0xffffffff);
16211 }
16212 else
16213 {
16214 if (TARGET_MINIMAL_TOC)
16215 fputs ("\t.long ", file);
16216 else
16217 fprintf (file, "\t.tc IS_%lx[TC],", (long) low & 0xffffffff);
16218 fprintf (file, "0x%lx\n", (long) low & 0xffffffff);
16219 }
16220 return;
16221 }
16222 }
16223
16224 if (GET_CODE (x) == CONST)
16225 {
16226 gcc_assert (GET_CODE (XEXP (x, 0)) == PLUS);
16227
16228 base = XEXP (XEXP (x, 0), 0);
16229 offset = INTVAL (XEXP (XEXP (x, 0), 1));
16230 }
16231
16232 switch (GET_CODE (base))
16233 {
16234 case SYMBOL_REF:
16235 name = XSTR (base, 0);
16236 break;
16237
16238 case LABEL_REF:
16239 ASM_GENERATE_INTERNAL_LABEL (buf, "L",
16240 CODE_LABEL_NUMBER (XEXP (base, 0)));
16241 break;
16242
16243 case CODE_LABEL:
16244 ASM_GENERATE_INTERNAL_LABEL (buf, "L", CODE_LABEL_NUMBER (base));
16245 break;
16246
16247 default:
16248 gcc_unreachable ();
16249 }
16250
16251 real_name = (*targetm.strip_name_encoding) (name);
16252 if (TARGET_MINIMAL_TOC)
16253 fputs (TARGET_32BIT ? "\t.long " : DOUBLE_INT_ASM_OP, file);
16254 else
16255 {
16256 fprintf (file, "\t.tc %s", real_name);
16257
16258 if (offset < 0)
16259 fprintf (file, ".N" HOST_WIDE_INT_PRINT_UNSIGNED, - offset);
16260 else if (offset)
16261 fprintf (file, ".P" HOST_WIDE_INT_PRINT_UNSIGNED, offset);
16262
16263 fputs ("[TC],", file);
16264 }
16265
16266 /* Currently C++ toc references to vtables can be emitted before it
16267 is decided whether the vtable is public or private. If this is
16268 the case, then the linker will eventually complain that there is
16269 a TOC reference to an unknown section. Thus, for vtables only,
16270 we emit the TOC reference to reference the symbol and not the
16271 section. */
16272 if (VTABLE_NAME_P (name))
16273 {
16274 RS6000_OUTPUT_BASENAME (file, name);
16275 if (offset < 0)
16276 fprintf (file, HOST_WIDE_INT_PRINT_DEC, offset);
16277 else if (offset > 0)
16278 fprintf (file, "+" HOST_WIDE_INT_PRINT_DEC, offset);
16279 }
16280 else
16281 output_addr_const (file, x);
16282 putc ('\n', file);
16283 }
16284
16285 /* Output an assembler pseudo-op to write an ASCII string of N characters
16286 starting at P to FILE.
16287
16288 On the RS/6000, we have to do this using the .byte operation and
16289 write out special characters outside the quoted string.
16290 Also, the assembler is broken; very long strings are truncated,
16291 so we must artificially break them up early. */
16292
16293 void
output_ascii(FILE * file,const char * p,int n)16294 output_ascii (FILE *file, const char *p, int n)
16295 {
16296 char c;
16297 int i, count_string;
16298 const char *for_string = "\t.byte \"";
16299 const char *for_decimal = "\t.byte ";
16300 const char *to_close = NULL;
16301
16302 count_string = 0;
16303 for (i = 0; i < n; i++)
16304 {
16305 c = *p++;
16306 if (c >= ' ' && c < 0177)
16307 {
16308 if (for_string)
16309 fputs (for_string, file);
16310 putc (c, file);
16311
16312 /* Write two quotes to get one. */
16313 if (c == '"')
16314 {
16315 putc (c, file);
16316 ++count_string;
16317 }
16318
16319 for_string = NULL;
16320 for_decimal = "\"\n\t.byte ";
16321 to_close = "\"\n";
16322 ++count_string;
16323
16324 if (count_string >= 512)
16325 {
16326 fputs (to_close, file);
16327
16328 for_string = "\t.byte \"";
16329 for_decimal = "\t.byte ";
16330 to_close = NULL;
16331 count_string = 0;
16332 }
16333 }
16334 else
16335 {
16336 if (for_decimal)
16337 fputs (for_decimal, file);
16338 fprintf (file, "%d", c);
16339
16340 for_string = "\n\t.byte \"";
16341 for_decimal = ", ";
16342 to_close = "\n";
16343 count_string = 0;
16344 }
16345 }
16346
16347 /* Now close the string if we have written one. Then end the line. */
16348 if (to_close)
16349 fputs (to_close, file);
16350 }
16351
16352 /* Generate a unique section name for FILENAME for a section type
16353 represented by SECTION_DESC. Output goes into BUF.
16354
16355 SECTION_DESC can be any string, as long as it is different for each
16356 possible section type.
16357
16358 We name the section in the same manner as xlc. The name begins with an
16359 underscore followed by the filename (after stripping any leading directory
16360 names) with the last period replaced by the string SECTION_DESC. If
16361 FILENAME does not contain a period, SECTION_DESC is appended to the end of
16362 the name. */
16363
16364 void
rs6000_gen_section_name(char ** buf,const char * filename,const char * section_desc)16365 rs6000_gen_section_name (char **buf, const char *filename,
16366 const char *section_desc)
16367 {
16368 const char *q, *after_last_slash, *last_period = 0;
16369 char *p;
16370 int len;
16371
16372 after_last_slash = filename;
16373 for (q = filename; *q; q++)
16374 {
16375 if (*q == '/')
16376 after_last_slash = q + 1;
16377 else if (*q == '.')
16378 last_period = q;
16379 }
16380
16381 len = strlen (after_last_slash) + strlen (section_desc) + 2;
16382 *buf = (char *) xmalloc (len);
16383
16384 p = *buf;
16385 *p++ = '_';
16386
16387 for (q = after_last_slash; *q; q++)
16388 {
16389 if (q == last_period)
16390 {
16391 strcpy (p, section_desc);
16392 p += strlen (section_desc);
16393 break;
16394 }
16395
16396 else if (ISALNUM (*q))
16397 *p++ = *q;
16398 }
16399
16400 if (last_period == 0)
16401 strcpy (p, section_desc);
16402 else
16403 *p = '\0';
16404 }
16405
16406 /* Emit profile function. */
16407
16408 void
output_profile_hook(int labelno ATTRIBUTE_UNUSED)16409 output_profile_hook (int labelno ATTRIBUTE_UNUSED)
16410 {
16411 /* Non-standard profiling for kernels, which just saves LR then calls
16412 _mcount without worrying about arg saves. The idea is to change
16413 the function prologue as little as possible as it isn't easy to
16414 account for arg save/restore code added just for _mcount. */
16415 if (TARGET_PROFILE_KERNEL)
16416 return;
16417
16418 if (DEFAULT_ABI == ABI_AIX)
16419 {
16420 #ifndef NO_PROFILE_COUNTERS
16421 # define NO_PROFILE_COUNTERS 0
16422 #endif
16423 if (NO_PROFILE_COUNTERS)
16424 emit_library_call (init_one_libfunc (RS6000_MCOUNT), 0, VOIDmode, 0);
16425 else
16426 {
16427 char buf[30];
16428 const char *label_name;
16429 rtx fun;
16430
16431 ASM_GENERATE_INTERNAL_LABEL (buf, "LP", labelno);
16432 label_name = (*targetm.strip_name_encoding) (ggc_strdup (buf));
16433 fun = gen_rtx_SYMBOL_REF (Pmode, label_name);
16434
16435 emit_library_call (init_one_libfunc (RS6000_MCOUNT), 0, VOIDmode, 1,
16436 fun, Pmode);
16437 }
16438 }
16439 else if (DEFAULT_ABI == ABI_DARWIN)
16440 {
16441 const char *mcount_name = RS6000_MCOUNT;
16442 int caller_addr_regno = LINK_REGISTER_REGNUM;
16443
16444 /* Be conservative and always set this, at least for now. */
16445 current_function_uses_pic_offset_table = 1;
16446
16447 #if TARGET_MACHO
16448 /* For PIC code, set up a stub and collect the caller's address
16449 from r0, which is where the prologue puts it. */
16450 if (MACHOPIC_INDIRECT
16451 && current_function_uses_pic_offset_table)
16452 caller_addr_regno = 0;
16453 #endif
16454 emit_library_call (gen_rtx_SYMBOL_REF (Pmode, mcount_name),
16455 0, VOIDmode, 1,
16456 gen_rtx_REG (Pmode, caller_addr_regno), Pmode);
16457 }
16458 }
16459
16460 /* Write function profiler code. */
16461
16462 void
output_function_profiler(FILE * file,int labelno)16463 output_function_profiler (FILE *file, int labelno)
16464 {
16465 char buf[100];
16466
16467 switch (DEFAULT_ABI)
16468 {
16469 default:
16470 gcc_unreachable ();
16471
16472 case ABI_V4:
16473 if (!TARGET_32BIT)
16474 {
16475 warning (0, "no profiling of 64-bit code for this ABI");
16476 return;
16477 }
16478 ASM_GENERATE_INTERNAL_LABEL (buf, "LP", labelno);
16479 fprintf (file, "\tmflr %s\n", reg_names[0]);
16480 if (NO_PROFILE_COUNTERS)
16481 {
16482 asm_fprintf (file, "\t{st|stw} %s,4(%s)\n",
16483 reg_names[0], reg_names[1]);
16484 }
16485 else if (TARGET_SECURE_PLT && flag_pic)
16486 {
16487 asm_fprintf (file, "\tbcl 20,31,1f\n1:\n\t{st|stw} %s,4(%s)\n",
16488 reg_names[0], reg_names[1]);
16489 asm_fprintf (file, "\tmflr %s\n", reg_names[12]);
16490 asm_fprintf (file, "\t{cau|addis} %s,%s,",
16491 reg_names[12], reg_names[12]);
16492 assemble_name (file, buf);
16493 asm_fprintf (file, "-1b@ha\n\t{cal|la} %s,", reg_names[0]);
16494 assemble_name (file, buf);
16495 asm_fprintf (file, "-1b@l(%s)\n", reg_names[12]);
16496 }
16497 else if (flag_pic == 1)
16498 {
16499 fputs ("\tbl _GLOBAL_OFFSET_TABLE_@local-4\n", file);
16500 asm_fprintf (file, "\t{st|stw} %s,4(%s)\n",
16501 reg_names[0], reg_names[1]);
16502 asm_fprintf (file, "\tmflr %s\n", reg_names[12]);
16503 asm_fprintf (file, "\t{l|lwz} %s,", reg_names[0]);
16504 assemble_name (file, buf);
16505 asm_fprintf (file, "@got(%s)\n", reg_names[12]);
16506 }
16507 else if (flag_pic > 1)
16508 {
16509 asm_fprintf (file, "\t{st|stw} %s,4(%s)\n",
16510 reg_names[0], reg_names[1]);
16511 /* Now, we need to get the address of the label. */
16512 fputs ("\tbcl 20,31,1f\n\t.long ", file);
16513 assemble_name (file, buf);
16514 fputs ("-.\n1:", file);
16515 asm_fprintf (file, "\tmflr %s\n", reg_names[11]);
16516 asm_fprintf (file, "\t{l|lwz} %s,0(%s)\n",
16517 reg_names[0], reg_names[11]);
16518 asm_fprintf (file, "\t{cax|add} %s,%s,%s\n",
16519 reg_names[0], reg_names[0], reg_names[11]);
16520 }
16521 else
16522 {
16523 asm_fprintf (file, "\t{liu|lis} %s,", reg_names[12]);
16524 assemble_name (file, buf);
16525 fputs ("@ha\n", file);
16526 asm_fprintf (file, "\t{st|stw} %s,4(%s)\n",
16527 reg_names[0], reg_names[1]);
16528 asm_fprintf (file, "\t{cal|la} %s,", reg_names[0]);
16529 assemble_name (file, buf);
16530 asm_fprintf (file, "@l(%s)\n", reg_names[12]);
16531 }
16532
16533 /* ABI_V4 saves the static chain reg with ASM_OUTPUT_REG_PUSH. */
16534 fprintf (file, "\tbl %s%s\n",
16535 RS6000_MCOUNT, flag_pic ? "@plt" : "");
16536 break;
16537
16538 case ABI_AIX:
16539 case ABI_DARWIN:
16540 if (!TARGET_PROFILE_KERNEL)
16541 {
16542 /* Don't do anything, done in output_profile_hook (). */
16543 }
16544 else
16545 {
16546 gcc_assert (!TARGET_32BIT);
16547
16548 asm_fprintf (file, "\tmflr %s\n", reg_names[0]);
16549 asm_fprintf (file, "\tstd %s,16(%s)\n", reg_names[0], reg_names[1]);
16550
16551 if (cfun->static_chain_decl != NULL)
16552 {
16553 asm_fprintf (file, "\tstd %s,24(%s)\n",
16554 reg_names[STATIC_CHAIN_REGNUM], reg_names[1]);
16555 fprintf (file, "\tbl %s\n", RS6000_MCOUNT);
16556 asm_fprintf (file, "\tld %s,24(%s)\n",
16557 reg_names[STATIC_CHAIN_REGNUM], reg_names[1]);
16558 }
16559 else
16560 fprintf (file, "\tbl %s\n", RS6000_MCOUNT);
16561 }
16562 break;
16563 }
16564 }
16565
16566
16567 /* Power4 load update and store update instructions are cracked into a
16568 load or store and an integer insn which are executed in the same cycle.
16569 Branches have their own dispatch slot which does not count against the
16570 GCC issue rate, but it changes the program flow so there are no other
16571 instructions to issue in this cycle. */
16572
16573 static int
rs6000_variable_issue(FILE * stream ATTRIBUTE_UNUSED,int verbose ATTRIBUTE_UNUSED,rtx insn,int more)16574 rs6000_variable_issue (FILE *stream ATTRIBUTE_UNUSED,
16575 int verbose ATTRIBUTE_UNUSED,
16576 rtx insn, int more)
16577 {
16578 if (GET_CODE (PATTERN (insn)) == USE
16579 || GET_CODE (PATTERN (insn)) == CLOBBER)
16580 return more;
16581
16582 if (rs6000_sched_groups)
16583 {
16584 if (is_microcoded_insn (insn))
16585 return 0;
16586 else if (is_cracked_insn (insn))
16587 return more > 2 ? more - 2 : 0;
16588 }
16589
16590 return more - 1;
16591 }
16592
16593 /* Adjust the cost of a scheduling dependency. Return the new cost of
16594 a dependency LINK or INSN on DEP_INSN. COST is the current cost. */
16595
16596 static int
rs6000_adjust_cost(rtx insn,rtx link,rtx dep_insn,int cost)16597 rs6000_adjust_cost (rtx insn, rtx link, rtx dep_insn, int cost)
16598 {
16599 if (! recog_memoized (insn))
16600 return 0;
16601
16602 if (REG_NOTE_KIND (link) != 0)
16603 return 0;
16604
16605 if (REG_NOTE_KIND (link) == 0)
16606 {
16607 /* Data dependency; DEP_INSN writes a register that INSN reads
16608 some cycles later. */
16609
16610 /* Separate a load from a narrower, dependent store. */
16611 if (rs6000_sched_groups
16612 && GET_CODE (PATTERN (insn)) == SET
16613 && GET_CODE (PATTERN (dep_insn)) == SET
16614 && GET_CODE (XEXP (PATTERN (insn), 1)) == MEM
16615 && GET_CODE (XEXP (PATTERN (dep_insn), 0)) == MEM
16616 && (GET_MODE_SIZE (GET_MODE (XEXP (PATTERN (insn), 1)))
16617 > GET_MODE_SIZE (GET_MODE (XEXP (PATTERN (dep_insn), 0)))))
16618 return cost + 14;
16619
16620 switch (get_attr_type (insn))
16621 {
16622 case TYPE_JMPREG:
16623 /* Tell the first scheduling pass about the latency between
16624 a mtctr and bctr (and mtlr and br/blr). The first
16625 scheduling pass will not know about this latency since
16626 the mtctr instruction, which has the latency associated
16627 to it, will be generated by reload. */
16628 return TARGET_POWER ? 5 : 4;
16629 case TYPE_BRANCH:
16630 /* Leave some extra cycles between a compare and its
16631 dependent branch, to inhibit expensive mispredicts. */
16632 if ((rs6000_cpu_attr == CPU_PPC603
16633 || rs6000_cpu_attr == CPU_PPC604
16634 || rs6000_cpu_attr == CPU_PPC604E
16635 || rs6000_cpu_attr == CPU_PPC620
16636 || rs6000_cpu_attr == CPU_PPC630
16637 || rs6000_cpu_attr == CPU_PPC750
16638 || rs6000_cpu_attr == CPU_PPC7400
16639 || rs6000_cpu_attr == CPU_PPC7450
16640 || rs6000_cpu_attr == CPU_POWER4
16641 || rs6000_cpu_attr == CPU_POWER5)
16642 && recog_memoized (dep_insn)
16643 && (INSN_CODE (dep_insn) >= 0)
16644 && (get_attr_type (dep_insn) == TYPE_CMP
16645 || get_attr_type (dep_insn) == TYPE_COMPARE
16646 || get_attr_type (dep_insn) == TYPE_DELAYED_COMPARE
16647 || get_attr_type (dep_insn) == TYPE_IMUL_COMPARE
16648 || get_attr_type (dep_insn) == TYPE_LMUL_COMPARE
16649 || get_attr_type (dep_insn) == TYPE_FPCOMPARE
16650 || get_attr_type (dep_insn) == TYPE_CR_LOGICAL
16651 || get_attr_type (dep_insn) == TYPE_DELAYED_CR))
16652 return cost + 2;
16653 default:
16654 break;
16655 }
16656 /* Fall out to return default cost. */
16657 }
16658
16659 return cost;
16660 }
16661
16662 /* The function returns a true if INSN is microcoded.
16663 Return false otherwise. */
16664
16665 static bool
is_microcoded_insn(rtx insn)16666 is_microcoded_insn (rtx insn)
16667 {
16668 if (!insn || !INSN_P (insn)
16669 || GET_CODE (PATTERN (insn)) == USE
16670 || GET_CODE (PATTERN (insn)) == CLOBBER)
16671 return false;
16672
16673 if (rs6000_sched_groups)
16674 {
16675 enum attr_type type = get_attr_type (insn);
16676 if (type == TYPE_LOAD_EXT_U
16677 || type == TYPE_LOAD_EXT_UX
16678 || type == TYPE_LOAD_UX
16679 || type == TYPE_STORE_UX
16680 || type == TYPE_MFCR)
16681 return true;
16682 }
16683
16684 return false;
16685 }
16686
16687 /* The function returns a nonzero value if INSN can be scheduled only
16688 as the first insn in a dispatch group ("dispatch-slot restricted").
16689 In this case, the returned value indicates how many dispatch slots
16690 the insn occupies (at the beginning of the group).
16691 Return 0 otherwise. */
16692
16693 static int
is_dispatch_slot_restricted(rtx insn)16694 is_dispatch_slot_restricted (rtx insn)
16695 {
16696 enum attr_type type;
16697
16698 if (!rs6000_sched_groups)
16699 return 0;
16700
16701 if (!insn
16702 || insn == NULL_RTX
16703 || GET_CODE (insn) == NOTE
16704 || GET_CODE (PATTERN (insn)) == USE
16705 || GET_CODE (PATTERN (insn)) == CLOBBER)
16706 return 0;
16707
16708 type = get_attr_type (insn);
16709
16710 switch (type)
16711 {
16712 case TYPE_MFCR:
16713 case TYPE_MFCRF:
16714 case TYPE_MTCR:
16715 case TYPE_DELAYED_CR:
16716 case TYPE_CR_LOGICAL:
16717 case TYPE_MTJMPR:
16718 case TYPE_MFJMPR:
16719 return 1;
16720 case TYPE_IDIV:
16721 case TYPE_LDIV:
16722 return 2;
16723 case TYPE_LOAD_L:
16724 case TYPE_STORE_C:
16725 case TYPE_ISYNC:
16726 case TYPE_SYNC:
16727 return 4;
16728 default:
16729 if (rs6000_cpu == PROCESSOR_POWER5
16730 && is_cracked_insn (insn))
16731 return 2;
16732 return 0;
16733 }
16734 }
16735
16736 /* The function returns true if INSN is cracked into 2 instructions
16737 by the processor (and therefore occupies 2 issue slots). */
16738
16739 static bool
is_cracked_insn(rtx insn)16740 is_cracked_insn (rtx insn)
16741 {
16742 if (!insn || !INSN_P (insn)
16743 || GET_CODE (PATTERN (insn)) == USE
16744 || GET_CODE (PATTERN (insn)) == CLOBBER)
16745 return false;
16746
16747 if (rs6000_sched_groups)
16748 {
16749 enum attr_type type = get_attr_type (insn);
16750 if (type == TYPE_LOAD_U || type == TYPE_STORE_U
16751 || type == TYPE_FPLOAD_U || type == TYPE_FPSTORE_U
16752 || type == TYPE_FPLOAD_UX || type == TYPE_FPSTORE_UX
16753 || type == TYPE_LOAD_EXT || type == TYPE_DELAYED_CR
16754 || type == TYPE_COMPARE || type == TYPE_DELAYED_COMPARE
16755 || type == TYPE_IMUL_COMPARE || type == TYPE_LMUL_COMPARE
16756 || type == TYPE_IDIV || type == TYPE_LDIV
16757 || type == TYPE_INSERT_WORD)
16758 return true;
16759 }
16760
16761 return false;
16762 }
16763
16764 /* The function returns true if INSN can be issued only from
16765 the branch slot. */
16766
16767 static bool
is_branch_slot_insn(rtx insn)16768 is_branch_slot_insn (rtx insn)
16769 {
16770 if (!insn || !INSN_P (insn)
16771 || GET_CODE (PATTERN (insn)) == USE
16772 || GET_CODE (PATTERN (insn)) == CLOBBER)
16773 return false;
16774
16775 if (rs6000_sched_groups)
16776 {
16777 enum attr_type type = get_attr_type (insn);
16778 if (type == TYPE_BRANCH || type == TYPE_JMPREG)
16779 return true;
16780 return false;
16781 }
16782
16783 return false;
16784 }
16785
16786 /* A C statement (sans semicolon) to update the integer scheduling
16787 priority INSN_PRIORITY (INSN). Increase the priority to execute the
16788 INSN earlier, reduce the priority to execute INSN later. Do not
16789 define this macro if you do not need to adjust the scheduling
16790 priorities of insns. */
16791
16792 static int
rs6000_adjust_priority(rtx insn ATTRIBUTE_UNUSED,int priority)16793 rs6000_adjust_priority (rtx insn ATTRIBUTE_UNUSED, int priority)
16794 {
16795 /* On machines (like the 750) which have asymmetric integer units,
16796 where one integer unit can do multiply and divides and the other
16797 can't, reduce the priority of multiply/divide so it is scheduled
16798 before other integer operations. */
16799
16800 #if 0
16801 if (! INSN_P (insn))
16802 return priority;
16803
16804 if (GET_CODE (PATTERN (insn)) == USE)
16805 return priority;
16806
16807 switch (rs6000_cpu_attr) {
16808 case CPU_PPC750:
16809 switch (get_attr_type (insn))
16810 {
16811 default:
16812 break;
16813
16814 case TYPE_IMUL:
16815 case TYPE_IDIV:
16816 fprintf (stderr, "priority was %#x (%d) before adjustment\n",
16817 priority, priority);
16818 if (priority >= 0 && priority < 0x01000000)
16819 priority >>= 3;
16820 break;
16821 }
16822 }
16823 #endif
16824
16825 if (is_dispatch_slot_restricted (insn)
16826 && reload_completed
16827 && current_sched_info->sched_max_insns_priority
16828 && rs6000_sched_restricted_insns_priority)
16829 {
16830
16831 /* Prioritize insns that can be dispatched only in the first
16832 dispatch slot. */
16833 if (rs6000_sched_restricted_insns_priority == 1)
16834 /* Attach highest priority to insn. This means that in
16835 haifa-sched.c:ready_sort(), dispatch-slot restriction considerations
16836 precede 'priority' (critical path) considerations. */
16837 return current_sched_info->sched_max_insns_priority;
16838 else if (rs6000_sched_restricted_insns_priority == 2)
16839 /* Increase priority of insn by a minimal amount. This means that in
16840 haifa-sched.c:ready_sort(), only 'priority' (critical path)
16841 considerations precede dispatch-slot restriction considerations. */
16842 return (priority + 1);
16843 }
16844
16845 return priority;
16846 }
16847
16848 /* Return how many instructions the machine can issue per cycle. */
16849
16850 static int
rs6000_issue_rate(void)16851 rs6000_issue_rate (void)
16852 {
16853 /* Use issue rate of 1 for first scheduling pass to decrease degradation. */
16854 if (!reload_completed)
16855 return 1;
16856
16857 switch (rs6000_cpu_attr) {
16858 case CPU_RIOS1: /* ? */
16859 case CPU_RS64A:
16860 case CPU_PPC601: /* ? */
16861 case CPU_PPC7450:
16862 return 3;
16863 case CPU_PPC440:
16864 case CPU_PPC603:
16865 case CPU_PPC750:
16866 case CPU_PPC7400:
16867 case CPU_PPC8540:
16868 return 2;
16869 case CPU_RIOS2:
16870 case CPU_PPC604:
16871 case CPU_PPC604E:
16872 case CPU_PPC620:
16873 case CPU_PPC630:
16874 return 4;
16875 case CPU_POWER4:
16876 case CPU_POWER5:
16877 return 5;
16878 default:
16879 return 1;
16880 }
16881 }
16882
16883 /* Return how many instructions to look ahead for better insn
16884 scheduling. */
16885
16886 static int
rs6000_use_sched_lookahead(void)16887 rs6000_use_sched_lookahead (void)
16888 {
16889 if (rs6000_cpu_attr == CPU_PPC8540)
16890 return 4;
16891 return 0;
16892 }
16893
16894 /* Determine is PAT refers to memory. */
16895
16896 static bool
is_mem_ref(rtx pat)16897 is_mem_ref (rtx pat)
16898 {
16899 const char * fmt;
16900 int i, j;
16901 bool ret = false;
16902
16903 if (GET_CODE (pat) == MEM)
16904 return true;
16905
16906 /* Recursively process the pattern. */
16907 fmt = GET_RTX_FORMAT (GET_CODE (pat));
16908
16909 for (i = GET_RTX_LENGTH (GET_CODE (pat)) - 1; i >= 0 && !ret; i--)
16910 {
16911 if (fmt[i] == 'e')
16912 ret |= is_mem_ref (XEXP (pat, i));
16913 else if (fmt[i] == 'E')
16914 for (j = XVECLEN (pat, i) - 1; j >= 0; j--)
16915 ret |= is_mem_ref (XVECEXP (pat, i, j));
16916 }
16917
16918 return ret;
16919 }
16920
16921 /* Determine if PAT is a PATTERN of a load insn. */
16922
16923 static bool
is_load_insn1(rtx pat)16924 is_load_insn1 (rtx pat)
16925 {
16926 if (!pat || pat == NULL_RTX)
16927 return false;
16928
16929 if (GET_CODE (pat) == SET)
16930 return is_mem_ref (SET_SRC (pat));
16931
16932 if (GET_CODE (pat) == PARALLEL)
16933 {
16934 int i;
16935
16936 for (i = 0; i < XVECLEN (pat, 0); i++)
16937 if (is_load_insn1 (XVECEXP (pat, 0, i)))
16938 return true;
16939 }
16940
16941 return false;
16942 }
16943
16944 /* Determine if INSN loads from memory. */
16945
16946 static bool
is_load_insn(rtx insn)16947 is_load_insn (rtx insn)
16948 {
16949 if (!insn || !INSN_P (insn))
16950 return false;
16951
16952 if (GET_CODE (insn) == CALL_INSN)
16953 return false;
16954
16955 return is_load_insn1 (PATTERN (insn));
16956 }
16957
16958 /* Determine if PAT is a PATTERN of a store insn. */
16959
16960 static bool
is_store_insn1(rtx pat)16961 is_store_insn1 (rtx pat)
16962 {
16963 if (!pat || pat == NULL_RTX)
16964 return false;
16965
16966 if (GET_CODE (pat) == SET)
16967 return is_mem_ref (SET_DEST (pat));
16968
16969 if (GET_CODE (pat) == PARALLEL)
16970 {
16971 int i;
16972
16973 for (i = 0; i < XVECLEN (pat, 0); i++)
16974 if (is_store_insn1 (XVECEXP (pat, 0, i)))
16975 return true;
16976 }
16977
16978 return false;
16979 }
16980
16981 /* Determine if INSN stores to memory. */
16982
16983 static bool
is_store_insn(rtx insn)16984 is_store_insn (rtx insn)
16985 {
16986 if (!insn || !INSN_P (insn))
16987 return false;
16988
16989 return is_store_insn1 (PATTERN (insn));
16990 }
16991
16992 /* Returns whether the dependence between INSN and NEXT is considered
16993 costly by the given target. */
16994
16995 static bool
rs6000_is_costly_dependence(rtx insn,rtx next,rtx link,int cost,int distance)16996 rs6000_is_costly_dependence (rtx insn, rtx next, rtx link, int cost,
16997 int distance)
16998 {
16999 /* If the flag is not enabled - no dependence is considered costly;
17000 allow all dependent insns in the same group.
17001 This is the most aggressive option. */
17002 if (rs6000_sched_costly_dep == no_dep_costly)
17003 return false;
17004
17005 /* If the flag is set to 1 - a dependence is always considered costly;
17006 do not allow dependent instructions in the same group.
17007 This is the most conservative option. */
17008 if (rs6000_sched_costly_dep == all_deps_costly)
17009 return true;
17010
17011 if (rs6000_sched_costly_dep == store_to_load_dep_costly
17012 && is_load_insn (next)
17013 && is_store_insn (insn))
17014 /* Prevent load after store in the same group. */
17015 return true;
17016
17017 if (rs6000_sched_costly_dep == true_store_to_load_dep_costly
17018 && is_load_insn (next)
17019 && is_store_insn (insn)
17020 && (!link || (int) REG_NOTE_KIND (link) == 0))
17021 /* Prevent load after store in the same group if it is a true
17022 dependence. */
17023 return true;
17024
17025 /* The flag is set to X; dependences with latency >= X are considered costly,
17026 and will not be scheduled in the same group. */
17027 if (rs6000_sched_costly_dep <= max_dep_latency
17028 && ((cost - distance) >= (int)rs6000_sched_costly_dep))
17029 return true;
17030
17031 return false;
17032 }
17033
17034 /* Return the next insn after INSN that is found before TAIL is reached,
17035 skipping any "non-active" insns - insns that will not actually occupy
17036 an issue slot. Return NULL_RTX if such an insn is not found. */
17037
17038 static rtx
get_next_active_insn(rtx insn,rtx tail)17039 get_next_active_insn (rtx insn, rtx tail)
17040 {
17041 if (insn == NULL_RTX || insn == tail)
17042 return NULL_RTX;
17043
17044 while (1)
17045 {
17046 insn = NEXT_INSN (insn);
17047 if (insn == NULL_RTX || insn == tail)
17048 return NULL_RTX;
17049
17050 if (CALL_P (insn)
17051 || JUMP_P (insn)
17052 || (NONJUMP_INSN_P (insn)
17053 && GET_CODE (PATTERN (insn)) != USE
17054 && GET_CODE (PATTERN (insn)) != CLOBBER
17055 && INSN_CODE (insn) != CODE_FOR_stack_tie))
17056 break;
17057 }
17058 return insn;
17059 }
17060
17061 /* Return whether the presence of INSN causes a dispatch group termination
17062 of group WHICH_GROUP.
17063
17064 If WHICH_GROUP == current_group, this function will return true if INSN
17065 causes the termination of the current group (i.e, the dispatch group to
17066 which INSN belongs). This means that INSN will be the last insn in the
17067 group it belongs to.
17068
17069 If WHICH_GROUP == previous_group, this function will return true if INSN
17070 causes the termination of the previous group (i.e, the dispatch group that
17071 precedes the group to which INSN belongs). This means that INSN will be
17072 the first insn in the group it belongs to). */
17073
17074 static bool
insn_terminates_group_p(rtx insn,enum group_termination which_group)17075 insn_terminates_group_p (rtx insn, enum group_termination which_group)
17076 {
17077 enum attr_type type;
17078
17079 if (! insn)
17080 return false;
17081
17082 type = get_attr_type (insn);
17083
17084 if (is_microcoded_insn (insn))
17085 return true;
17086
17087 if (which_group == current_group)
17088 {
17089 if (is_branch_slot_insn (insn))
17090 return true;
17091 return false;
17092 }
17093 else if (which_group == previous_group)
17094 {
17095 if (is_dispatch_slot_restricted (insn))
17096 return true;
17097 return false;
17098 }
17099
17100 return false;
17101 }
17102
17103 /* Return true if it is recommended to keep NEXT_INSN "far" (in a separate
17104 dispatch group) from the insns in GROUP_INSNS. Return false otherwise. */
17105
17106 static bool
is_costly_group(rtx * group_insns,rtx next_insn)17107 is_costly_group (rtx *group_insns, rtx next_insn)
17108 {
17109 int i;
17110 rtx link;
17111 int cost;
17112 int issue_rate = rs6000_issue_rate ();
17113
17114 for (i = 0; i < issue_rate; i++)
17115 {
17116 rtx insn = group_insns[i];
17117 if (!insn)
17118 continue;
17119 for (link = INSN_DEPEND (insn); link != 0; link = XEXP (link, 1))
17120 {
17121 rtx next = XEXP (link, 0);
17122 if (next == next_insn)
17123 {
17124 cost = insn_cost (insn, link, next_insn);
17125 if (rs6000_is_costly_dependence (insn, next_insn, link, cost, 0))
17126 return true;
17127 }
17128 }
17129 }
17130
17131 return false;
17132 }
17133
17134 /* Utility of the function redefine_groups.
17135 Check if it is too costly to schedule NEXT_INSN together with GROUP_INSNS
17136 in the same dispatch group. If so, insert nops before NEXT_INSN, in order
17137 to keep it "far" (in a separate group) from GROUP_INSNS, following
17138 one of the following schemes, depending on the value of the flag
17139 -minsert_sched_nops = X:
17140 (1) X == sched_finish_regroup_exact: insert exactly as many nops as needed
17141 in order to force NEXT_INSN into a separate group.
17142 (2) X < sched_finish_regroup_exact: insert exactly X nops.
17143 GROUP_END, CAN_ISSUE_MORE and GROUP_COUNT record the state after nop
17144 insertion (has a group just ended, how many vacant issue slots remain in the
17145 last group, and how many dispatch groups were encountered so far). */
17146
17147 static int
force_new_group(int sched_verbose,FILE * dump,rtx * group_insns,rtx next_insn,bool * group_end,int can_issue_more,int * group_count)17148 force_new_group (int sched_verbose, FILE *dump, rtx *group_insns,
17149 rtx next_insn, bool *group_end, int can_issue_more,
17150 int *group_count)
17151 {
17152 rtx nop;
17153 bool force;
17154 int issue_rate = rs6000_issue_rate ();
17155 bool end = *group_end;
17156 int i;
17157
17158 if (next_insn == NULL_RTX)
17159 return can_issue_more;
17160
17161 if (rs6000_sched_insert_nops > sched_finish_regroup_exact)
17162 return can_issue_more;
17163
17164 force = is_costly_group (group_insns, next_insn);
17165 if (!force)
17166 return can_issue_more;
17167
17168 if (sched_verbose > 6)
17169 fprintf (dump,"force: group count = %d, can_issue_more = %d\n",
17170 *group_count ,can_issue_more);
17171
17172 if (rs6000_sched_insert_nops == sched_finish_regroup_exact)
17173 {
17174 if (*group_end)
17175 can_issue_more = 0;
17176
17177 /* Since only a branch can be issued in the last issue_slot, it is
17178 sufficient to insert 'can_issue_more - 1' nops if next_insn is not
17179 a branch. If next_insn is a branch, we insert 'can_issue_more' nops;
17180 in this case the last nop will start a new group and the branch
17181 will be forced to the new group. */
17182 if (can_issue_more && !is_branch_slot_insn (next_insn))
17183 can_issue_more--;
17184
17185 while (can_issue_more > 0)
17186 {
17187 nop = gen_nop ();
17188 emit_insn_before (nop, next_insn);
17189 can_issue_more--;
17190 }
17191
17192 *group_end = true;
17193 return 0;
17194 }
17195
17196 if (rs6000_sched_insert_nops < sched_finish_regroup_exact)
17197 {
17198 int n_nops = rs6000_sched_insert_nops;
17199
17200 /* Nops can't be issued from the branch slot, so the effective
17201 issue_rate for nops is 'issue_rate - 1'. */
17202 if (can_issue_more == 0)
17203 can_issue_more = issue_rate;
17204 can_issue_more--;
17205 if (can_issue_more == 0)
17206 {
17207 can_issue_more = issue_rate - 1;
17208 (*group_count)++;
17209 end = true;
17210 for (i = 0; i < issue_rate; i++)
17211 {
17212 group_insns[i] = 0;
17213 }
17214 }
17215
17216 while (n_nops > 0)
17217 {
17218 nop = gen_nop ();
17219 emit_insn_before (nop, next_insn);
17220 if (can_issue_more == issue_rate - 1) /* new group begins */
17221 end = false;
17222 can_issue_more--;
17223 if (can_issue_more == 0)
17224 {
17225 can_issue_more = issue_rate - 1;
17226 (*group_count)++;
17227 end = true;
17228 for (i = 0; i < issue_rate; i++)
17229 {
17230 group_insns[i] = 0;
17231 }
17232 }
17233 n_nops--;
17234 }
17235
17236 /* Scale back relative to 'issue_rate' (instead of 'issue_rate - 1'). */
17237 can_issue_more++;
17238
17239 /* Is next_insn going to start a new group? */
17240 *group_end
17241 = (end
17242 || (can_issue_more == 1 && !is_branch_slot_insn (next_insn))
17243 || (can_issue_more <= 2 && is_cracked_insn (next_insn))
17244 || (can_issue_more < issue_rate &&
17245 insn_terminates_group_p (next_insn, previous_group)));
17246 if (*group_end && end)
17247 (*group_count)--;
17248
17249 if (sched_verbose > 6)
17250 fprintf (dump, "done force: group count = %d, can_issue_more = %d\n",
17251 *group_count, can_issue_more);
17252 return can_issue_more;
17253 }
17254
17255 return can_issue_more;
17256 }
17257
17258 /* This function tries to synch the dispatch groups that the compiler "sees"
17259 with the dispatch groups that the processor dispatcher is expected to
17260 form in practice. It tries to achieve this synchronization by forcing the
17261 estimated processor grouping on the compiler (as opposed to the function
17262 'pad_goups' which tries to force the scheduler's grouping on the processor).
17263
17264 The function scans the insn sequence between PREV_HEAD_INSN and TAIL and
17265 examines the (estimated) dispatch groups that will be formed by the processor
17266 dispatcher. It marks these group boundaries to reflect the estimated
17267 processor grouping, overriding the grouping that the scheduler had marked.
17268 Depending on the value of the flag '-minsert-sched-nops' this function can
17269 force certain insns into separate groups or force a certain distance between
17270 them by inserting nops, for example, if there exists a "costly dependence"
17271 between the insns.
17272
17273 The function estimates the group boundaries that the processor will form as
17274 follows: It keeps track of how many vacant issue slots are available after
17275 each insn. A subsequent insn will start a new group if one of the following
17276 4 cases applies:
17277 - no more vacant issue slots remain in the current dispatch group.
17278 - only the last issue slot, which is the branch slot, is vacant, but the next
17279 insn is not a branch.
17280 - only the last 2 or less issue slots, including the branch slot, are vacant,
17281 which means that a cracked insn (which occupies two issue slots) can't be
17282 issued in this group.
17283 - less than 'issue_rate' slots are vacant, and the next insn always needs to
17284 start a new group. */
17285
17286 static int
redefine_groups(FILE * dump,int sched_verbose,rtx prev_head_insn,rtx tail)17287 redefine_groups (FILE *dump, int sched_verbose, rtx prev_head_insn, rtx tail)
17288 {
17289 rtx insn, next_insn;
17290 int issue_rate;
17291 int can_issue_more;
17292 int slot, i;
17293 bool group_end;
17294 int group_count = 0;
17295 rtx *group_insns;
17296
17297 /* Initialize. */
17298 issue_rate = rs6000_issue_rate ();
17299 group_insns = alloca (issue_rate * sizeof (rtx));
17300 for (i = 0; i < issue_rate; i++)
17301 {
17302 group_insns[i] = 0;
17303 }
17304 can_issue_more = issue_rate;
17305 slot = 0;
17306 insn = get_next_active_insn (prev_head_insn, tail);
17307 group_end = false;
17308
17309 while (insn != NULL_RTX)
17310 {
17311 slot = (issue_rate - can_issue_more);
17312 group_insns[slot] = insn;
17313 can_issue_more =
17314 rs6000_variable_issue (dump, sched_verbose, insn, can_issue_more);
17315 if (insn_terminates_group_p (insn, current_group))
17316 can_issue_more = 0;
17317
17318 next_insn = get_next_active_insn (insn, tail);
17319 if (next_insn == NULL_RTX)
17320 return group_count + 1;
17321
17322 /* Is next_insn going to start a new group? */
17323 group_end
17324 = (can_issue_more == 0
17325 || (can_issue_more == 1 && !is_branch_slot_insn (next_insn))
17326 || (can_issue_more <= 2 && is_cracked_insn (next_insn))
17327 || (can_issue_more < issue_rate &&
17328 insn_terminates_group_p (next_insn, previous_group)));
17329
17330 can_issue_more = force_new_group (sched_verbose, dump, group_insns,
17331 next_insn, &group_end, can_issue_more,
17332 &group_count);
17333
17334 if (group_end)
17335 {
17336 group_count++;
17337 can_issue_more = 0;
17338 for (i = 0; i < issue_rate; i++)
17339 {
17340 group_insns[i] = 0;
17341 }
17342 }
17343
17344 if (GET_MODE (next_insn) == TImode && can_issue_more)
17345 PUT_MODE (next_insn, VOIDmode);
17346 else if (!can_issue_more && GET_MODE (next_insn) != TImode)
17347 PUT_MODE (next_insn, TImode);
17348
17349 insn = next_insn;
17350 if (can_issue_more == 0)
17351 can_issue_more = issue_rate;
17352 } /* while */
17353
17354 return group_count;
17355 }
17356
17357 /* Scan the insn sequence between PREV_HEAD_INSN and TAIL and examine the
17358 dispatch group boundaries that the scheduler had marked. Pad with nops
17359 any dispatch groups which have vacant issue slots, in order to force the
17360 scheduler's grouping on the processor dispatcher. The function
17361 returns the number of dispatch groups found. */
17362
17363 static int
pad_groups(FILE * dump,int sched_verbose,rtx prev_head_insn,rtx tail)17364 pad_groups (FILE *dump, int sched_verbose, rtx prev_head_insn, rtx tail)
17365 {
17366 rtx insn, next_insn;
17367 rtx nop;
17368 int issue_rate;
17369 int can_issue_more;
17370 int group_end;
17371 int group_count = 0;
17372
17373 /* Initialize issue_rate. */
17374 issue_rate = rs6000_issue_rate ();
17375 can_issue_more = issue_rate;
17376
17377 insn = get_next_active_insn (prev_head_insn, tail);
17378 next_insn = get_next_active_insn (insn, tail);
17379
17380 while (insn != NULL_RTX)
17381 {
17382 can_issue_more =
17383 rs6000_variable_issue (dump, sched_verbose, insn, can_issue_more);
17384
17385 group_end = (next_insn == NULL_RTX || GET_MODE (next_insn) == TImode);
17386
17387 if (next_insn == NULL_RTX)
17388 break;
17389
17390 if (group_end)
17391 {
17392 /* If the scheduler had marked group termination at this location
17393 (between insn and next_indn), and neither insn nor next_insn will
17394 force group termination, pad the group with nops to force group
17395 termination. */
17396 if (can_issue_more
17397 && (rs6000_sched_insert_nops == sched_finish_pad_groups)
17398 && !insn_terminates_group_p (insn, current_group)
17399 && !insn_terminates_group_p (next_insn, previous_group))
17400 {
17401 if (!is_branch_slot_insn (next_insn))
17402 can_issue_more--;
17403
17404 while (can_issue_more)
17405 {
17406 nop = gen_nop ();
17407 emit_insn_before (nop, next_insn);
17408 can_issue_more--;
17409 }
17410 }
17411
17412 can_issue_more = issue_rate;
17413 group_count++;
17414 }
17415
17416 insn = next_insn;
17417 next_insn = get_next_active_insn (insn, tail);
17418 }
17419
17420 return group_count;
17421 }
17422
17423 /* The following function is called at the end of scheduling BB.
17424 After reload, it inserts nops at insn group bundling. */
17425
17426 static void
rs6000_sched_finish(FILE * dump,int sched_verbose)17427 rs6000_sched_finish (FILE *dump, int sched_verbose)
17428 {
17429 int n_groups;
17430
17431 if (sched_verbose)
17432 fprintf (dump, "=== Finishing schedule.\n");
17433
17434 if (reload_completed && rs6000_sched_groups)
17435 {
17436 if (rs6000_sched_insert_nops == sched_finish_none)
17437 return;
17438
17439 if (rs6000_sched_insert_nops == sched_finish_pad_groups)
17440 n_groups = pad_groups (dump, sched_verbose,
17441 current_sched_info->prev_head,
17442 current_sched_info->next_tail);
17443 else
17444 n_groups = redefine_groups (dump, sched_verbose,
17445 current_sched_info->prev_head,
17446 current_sched_info->next_tail);
17447
17448 if (sched_verbose >= 6)
17449 {
17450 fprintf (dump, "ngroups = %d\n", n_groups);
17451 print_rtl (dump, current_sched_info->prev_head);
17452 fprintf (dump, "Done finish_sched\n");
17453 }
17454 }
17455 }
17456
17457 /* Length in units of the trampoline for entering a nested function. */
17458
17459 int
rs6000_trampoline_size(void)17460 rs6000_trampoline_size (void)
17461 {
17462 int ret = 0;
17463
17464 switch (DEFAULT_ABI)
17465 {
17466 default:
17467 gcc_unreachable ();
17468
17469 case ABI_AIX:
17470 ret = (TARGET_32BIT) ? 12 : 24;
17471 break;
17472
17473 case ABI_DARWIN:
17474 case ABI_V4:
17475 ret = (TARGET_32BIT) ? 40 : 48;
17476 break;
17477 }
17478
17479 return ret;
17480 }
17481
17482 /* Emit RTL insns to initialize the variable parts of a trampoline.
17483 FNADDR is an RTX for the address of the function's pure code.
17484 CXT is an RTX for the static chain value for the function. */
17485
17486 void
rs6000_initialize_trampoline(rtx addr,rtx fnaddr,rtx cxt)17487 rs6000_initialize_trampoline (rtx addr, rtx fnaddr, rtx cxt)
17488 {
17489 int regsize = (TARGET_32BIT) ? 4 : 8;
17490 rtx ctx_reg = force_reg (Pmode, cxt);
17491
17492 switch (DEFAULT_ABI)
17493 {
17494 default:
17495 gcc_unreachable ();
17496
17497 /* Macros to shorten the code expansions below. */
17498 #define MEM_DEREF(addr) gen_rtx_MEM (Pmode, memory_address (Pmode, addr))
17499 #define MEM_PLUS(addr,offset) \
17500 gen_rtx_MEM (Pmode, memory_address (Pmode, plus_constant (addr, offset)))
17501
17502 /* Under AIX, just build the 3 word function descriptor */
17503 case ABI_AIX:
17504 {
17505 rtx fn_reg = gen_reg_rtx (Pmode);
17506 rtx toc_reg = gen_reg_rtx (Pmode);
17507 emit_move_insn (fn_reg, MEM_DEREF (fnaddr));
17508 emit_move_insn (toc_reg, MEM_PLUS (fnaddr, regsize));
17509 emit_move_insn (MEM_DEREF (addr), fn_reg);
17510 emit_move_insn (MEM_PLUS (addr, regsize), toc_reg);
17511 emit_move_insn (MEM_PLUS (addr, 2*regsize), ctx_reg);
17512 }
17513 break;
17514
17515 /* Under V.4/eabi/darwin, __trampoline_setup does the real work. */
17516 case ABI_DARWIN:
17517 case ABI_V4:
17518 emit_library_call (gen_rtx_SYMBOL_REF (Pmode, "__trampoline_setup"),
17519 FALSE, VOIDmode, 4,
17520 addr, Pmode,
17521 GEN_INT (rs6000_trampoline_size ()), SImode,
17522 fnaddr, Pmode,
17523 ctx_reg, Pmode);
17524 break;
17525 }
17526
17527 return;
17528 }
17529
17530
17531 /* Table of valid machine attributes. */
17532
17533 const struct attribute_spec rs6000_attribute_table[] =
17534 {
17535 /* { name, min_len, max_len, decl_req, type_req, fn_type_req, handler } */
17536 { "altivec", 1, 1, false, true, false, rs6000_handle_altivec_attribute },
17537 { "longcall", 0, 0, false, true, true, rs6000_handle_longcall_attribute },
17538 { "shortcall", 0, 0, false, true, true, rs6000_handle_longcall_attribute },
17539 { "ms_struct", 0, 0, false, false, false, rs6000_handle_struct_attribute },
17540 { "gcc_struct", 0, 0, false, false, false, rs6000_handle_struct_attribute },
17541 #ifdef SUBTARGET_ATTRIBUTE_TABLE
17542 SUBTARGET_ATTRIBUTE_TABLE,
17543 #endif
17544 { NULL, 0, 0, false, false, false, NULL }
17545 };
17546
17547 /* Handle the "altivec" attribute. The attribute may have
17548 arguments as follows:
17549
17550 __attribute__((altivec(vector__)))
17551 __attribute__((altivec(pixel__))) (always followed by 'unsigned short')
17552 __attribute__((altivec(bool__))) (always followed by 'unsigned')
17553
17554 and may appear more than once (e.g., 'vector bool char') in a
17555 given declaration. */
17556
17557 static tree
rs6000_handle_altivec_attribute(tree * node,tree name ATTRIBUTE_UNUSED,tree args,int flags ATTRIBUTE_UNUSED,bool * no_add_attrs)17558 rs6000_handle_altivec_attribute (tree *node,
17559 tree name ATTRIBUTE_UNUSED,
17560 tree args,
17561 int flags ATTRIBUTE_UNUSED,
17562 bool *no_add_attrs)
17563 {
17564 tree type = *node, result = NULL_TREE;
17565 enum machine_mode mode;
17566 int unsigned_p;
17567 char altivec_type
17568 = ((args && TREE_CODE (args) == TREE_LIST && TREE_VALUE (args)
17569 && TREE_CODE (TREE_VALUE (args)) == IDENTIFIER_NODE)
17570 ? *IDENTIFIER_POINTER (TREE_VALUE (args))
17571 : '?');
17572
17573 while (POINTER_TYPE_P (type)
17574 || TREE_CODE (type) == FUNCTION_TYPE
17575 || TREE_CODE (type) == METHOD_TYPE
17576 || TREE_CODE (type) == ARRAY_TYPE)
17577 type = TREE_TYPE (type);
17578
17579 mode = TYPE_MODE (type);
17580
17581 /* Check for invalid AltiVec type qualifiers. */
17582 if (type == long_unsigned_type_node || type == long_integer_type_node)
17583 {
17584 if (TARGET_64BIT)
17585 error ("use of %<long%> in AltiVec types is invalid for 64-bit code");
17586 else if (rs6000_warn_altivec_long)
17587 warning (0, "use of %<long%> in AltiVec types is deprecated; use %<int%>");
17588 }
17589 else if (type == long_long_unsigned_type_node
17590 || type == long_long_integer_type_node)
17591 error ("use of %<long long%> in AltiVec types is invalid");
17592 else if (type == double_type_node)
17593 error ("use of %<double%> in AltiVec types is invalid");
17594 else if (type == long_double_type_node)
17595 error ("use of %<long double%> in AltiVec types is invalid");
17596 else if (type == boolean_type_node)
17597 error ("use of boolean types in AltiVec types is invalid");
17598 else if (TREE_CODE (type) == COMPLEX_TYPE)
17599 error ("use of %<complex%> in AltiVec types is invalid");
17600 else if (DECIMAL_FLOAT_MODE_P (mode))
17601 error ("use of decimal floating point types in AltiVec types is invalid");
17602
17603 switch (altivec_type)
17604 {
17605 case 'v':
17606 unsigned_p = TYPE_UNSIGNED (type);
17607 switch (mode)
17608 {
17609 case SImode:
17610 result = (unsigned_p ? unsigned_V4SI_type_node : V4SI_type_node);
17611 break;
17612 case HImode:
17613 result = (unsigned_p ? unsigned_V8HI_type_node : V8HI_type_node);
17614 break;
17615 case QImode:
17616 result = (unsigned_p ? unsigned_V16QI_type_node : V16QI_type_node);
17617 break;
17618 case SFmode: result = V4SF_type_node; break;
17619 /* If the user says 'vector int bool', we may be handed the 'bool'
17620 attribute _before_ the 'vector' attribute, and so select the
17621 proper type in the 'b' case below. */
17622 case V4SImode: case V8HImode: case V16QImode: case V4SFmode:
17623 result = type;
17624 default: break;
17625 }
17626 break;
17627 case 'b':
17628 switch (mode)
17629 {
17630 case SImode: case V4SImode: result = bool_V4SI_type_node; break;
17631 case HImode: case V8HImode: result = bool_V8HI_type_node; break;
17632 case QImode: case V16QImode: result = bool_V16QI_type_node;
17633 default: break;
17634 }
17635 break;
17636 case 'p':
17637 switch (mode)
17638 {
17639 case V8HImode: result = pixel_V8HI_type_node;
17640 default: break;
17641 }
17642 default: break;
17643 }
17644
17645 if (result && result != type && TYPE_READONLY (type))
17646 result = build_qualified_type (result, TYPE_QUAL_CONST);
17647
17648 *no_add_attrs = true; /* No need to hang on to the attribute. */
17649
17650 if (result)
17651 *node = reconstruct_complex_type (*node, result);
17652
17653 return NULL_TREE;
17654 }
17655
17656 /* AltiVec defines four built-in scalar types that serve as vector
17657 elements; we must teach the compiler how to mangle them. */
17658
17659 static const char *
rs6000_mangle_fundamental_type(tree type)17660 rs6000_mangle_fundamental_type (tree type)
17661 {
17662 if (type == bool_char_type_node) return "U6__boolc";
17663 if (type == bool_short_type_node) return "U6__bools";
17664 if (type == pixel_type_node) return "u7__pixel";
17665 if (type == bool_int_type_node) return "U6__booli";
17666
17667 /* Mangle IBM extended float long double as `g' (__float128) on
17668 powerpc*-linux where long-double-64 previously was the default. */
17669 if (TYPE_MAIN_VARIANT (type) == long_double_type_node
17670 && TARGET_ELF
17671 && TARGET_LONG_DOUBLE_128
17672 && !TARGET_IEEEQUAD)
17673 return "g";
17674
17675 /* For all other types, use normal C++ mangling. */
17676 return NULL;
17677 }
17678
17679 /* Handle a "longcall" or "shortcall" attribute; arguments as in
17680 struct attribute_spec.handler. */
17681
17682 static tree
rs6000_handle_longcall_attribute(tree * node,tree name,tree args ATTRIBUTE_UNUSED,int flags ATTRIBUTE_UNUSED,bool * no_add_attrs)17683 rs6000_handle_longcall_attribute (tree *node, tree name,
17684 tree args ATTRIBUTE_UNUSED,
17685 int flags ATTRIBUTE_UNUSED,
17686 bool *no_add_attrs)
17687 {
17688 if (TREE_CODE (*node) != FUNCTION_TYPE
17689 && TREE_CODE (*node) != FIELD_DECL
17690 && TREE_CODE (*node) != TYPE_DECL)
17691 {
17692 warning (OPT_Wattributes, "%qs attribute only applies to functions",
17693 IDENTIFIER_POINTER (name));
17694 *no_add_attrs = true;
17695 }
17696
17697 return NULL_TREE;
17698 }
17699
17700 /* Set longcall attributes on all functions declared when
17701 rs6000_default_long_calls is true. */
17702 static void
rs6000_set_default_type_attributes(tree type)17703 rs6000_set_default_type_attributes (tree type)
17704 {
17705 if (rs6000_default_long_calls
17706 && (TREE_CODE (type) == FUNCTION_TYPE
17707 || TREE_CODE (type) == METHOD_TYPE))
17708 TYPE_ATTRIBUTES (type) = tree_cons (get_identifier ("longcall"),
17709 NULL_TREE,
17710 TYPE_ATTRIBUTES (type));
17711
17712 #if TARGET_MACHO
17713 darwin_set_default_type_attributes (type);
17714 #endif
17715 }
17716
17717 /* Return a reference suitable for calling a function with the
17718 longcall attribute. */
17719
17720 rtx
rs6000_longcall_ref(rtx call_ref)17721 rs6000_longcall_ref (rtx call_ref)
17722 {
17723 const char *call_name;
17724 tree node;
17725
17726 if (GET_CODE (call_ref) != SYMBOL_REF)
17727 return call_ref;
17728
17729 /* System V adds '.' to the internal name, so skip them. */
17730 call_name = XSTR (call_ref, 0);
17731 if (*call_name == '.')
17732 {
17733 while (*call_name == '.')
17734 call_name++;
17735
17736 node = get_identifier (call_name);
17737 call_ref = gen_rtx_SYMBOL_REF (VOIDmode, IDENTIFIER_POINTER (node));
17738 }
17739
17740 return force_reg (Pmode, call_ref);
17741 }
17742
17743 #ifndef TARGET_USE_MS_BITFIELD_LAYOUT
17744 #define TARGET_USE_MS_BITFIELD_LAYOUT 0
17745 #endif
17746
17747 /* Handle a "ms_struct" or "gcc_struct" attribute; arguments as in
17748 struct attribute_spec.handler. */
17749 static tree
rs6000_handle_struct_attribute(tree * node,tree name,tree args ATTRIBUTE_UNUSED,int flags ATTRIBUTE_UNUSED,bool * no_add_attrs)17750 rs6000_handle_struct_attribute (tree *node, tree name,
17751 tree args ATTRIBUTE_UNUSED,
17752 int flags ATTRIBUTE_UNUSED, bool *no_add_attrs)
17753 {
17754 tree *type = NULL;
17755 if (DECL_P (*node))
17756 {
17757 if (TREE_CODE (*node) == TYPE_DECL)
17758 type = &TREE_TYPE (*node);
17759 }
17760 else
17761 type = node;
17762
17763 if (!(type && (TREE_CODE (*type) == RECORD_TYPE
17764 || TREE_CODE (*type) == UNION_TYPE)))
17765 {
17766 warning (OPT_Wattributes, "%qs attribute ignored", IDENTIFIER_POINTER (name));
17767 *no_add_attrs = true;
17768 }
17769
17770 else if ((is_attribute_p ("ms_struct", name)
17771 && lookup_attribute ("gcc_struct", TYPE_ATTRIBUTES (*type)))
17772 || ((is_attribute_p ("gcc_struct", name)
17773 && lookup_attribute ("ms_struct", TYPE_ATTRIBUTES (*type)))))
17774 {
17775 warning (OPT_Wattributes, "%qs incompatible attribute ignored",
17776 IDENTIFIER_POINTER (name));
17777 *no_add_attrs = true;
17778 }
17779
17780 return NULL_TREE;
17781 }
17782
17783 static bool
rs6000_ms_bitfield_layout_p(tree record_type)17784 rs6000_ms_bitfield_layout_p (tree record_type)
17785 {
17786 return (TARGET_USE_MS_BITFIELD_LAYOUT &&
17787 !lookup_attribute ("gcc_struct", TYPE_ATTRIBUTES (record_type)))
17788 || lookup_attribute ("ms_struct", TYPE_ATTRIBUTES (record_type));
17789 }
17790
17791 #ifdef USING_ELFOS_H
17792
17793 /* A get_unnamed_section callback, used for switching to toc_section. */
17794
17795 static void
rs6000_elf_output_toc_section_asm_op(const void * data ATTRIBUTE_UNUSED)17796 rs6000_elf_output_toc_section_asm_op (const void *data ATTRIBUTE_UNUSED)
17797 {
17798 if (DEFAULT_ABI == ABI_AIX
17799 && TARGET_MINIMAL_TOC
17800 && !TARGET_RELOCATABLE)
17801 {
17802 if (!toc_initialized)
17803 {
17804 toc_initialized = 1;
17805 fprintf (asm_out_file, "%s\n", TOC_SECTION_ASM_OP);
17806 (*targetm.asm_out.internal_label) (asm_out_file, "LCTOC", 0);
17807 fprintf (asm_out_file, "\t.tc ");
17808 ASM_OUTPUT_INTERNAL_LABEL_PREFIX (asm_out_file, "LCTOC1[TC],");
17809 ASM_OUTPUT_INTERNAL_LABEL_PREFIX (asm_out_file, "LCTOC1");
17810 fprintf (asm_out_file, "\n");
17811
17812 fprintf (asm_out_file, "%s\n", MINIMAL_TOC_SECTION_ASM_OP);
17813 ASM_OUTPUT_INTERNAL_LABEL_PREFIX (asm_out_file, "LCTOC1");
17814 fprintf (asm_out_file, " = .+32768\n");
17815 }
17816 else
17817 fprintf (asm_out_file, "%s\n", MINIMAL_TOC_SECTION_ASM_OP);
17818 }
17819 else if (DEFAULT_ABI == ABI_AIX && !TARGET_RELOCATABLE)
17820 fprintf (asm_out_file, "%s\n", TOC_SECTION_ASM_OP);
17821 else
17822 {
17823 fprintf (asm_out_file, "%s\n", MINIMAL_TOC_SECTION_ASM_OP);
17824 if (!toc_initialized)
17825 {
17826 ASM_OUTPUT_INTERNAL_LABEL_PREFIX (asm_out_file, "LCTOC1");
17827 fprintf (asm_out_file, " = .+32768\n");
17828 toc_initialized = 1;
17829 }
17830 }
17831 }
17832
17833 /* Implement TARGET_ASM_INIT_SECTIONS. */
17834
17835 static void
rs6000_elf_asm_init_sections(void)17836 rs6000_elf_asm_init_sections (void)
17837 {
17838 toc_section
17839 = get_unnamed_section (0, rs6000_elf_output_toc_section_asm_op, NULL);
17840
17841 sdata2_section
17842 = get_unnamed_section (SECTION_WRITE, output_section_asm_op,
17843 SDATA2_SECTION_ASM_OP);
17844 }
17845
17846 /* Implement TARGET_SELECT_RTX_SECTION. */
17847
17848 static section *
rs6000_elf_select_rtx_section(enum machine_mode mode,rtx x,unsigned HOST_WIDE_INT align)17849 rs6000_elf_select_rtx_section (enum machine_mode mode, rtx x,
17850 unsigned HOST_WIDE_INT align)
17851 {
17852 if (ASM_OUTPUT_SPECIAL_POOL_ENTRY_P (x, mode))
17853 return toc_section;
17854 else
17855 return default_elf_select_rtx_section (mode, x, align);
17856 }
17857
17858 /* For a SYMBOL_REF, set generic flags and then perform some
17859 target-specific processing.
17860
17861 When the AIX ABI is requested on a non-AIX system, replace the
17862 function name with the real name (with a leading .) rather than the
17863 function descriptor name. This saves a lot of overriding code to
17864 read the prefixes. */
17865
17866 static void
rs6000_elf_encode_section_info(tree decl,rtx rtl,int first)17867 rs6000_elf_encode_section_info (tree decl, rtx rtl, int first)
17868 {
17869 default_encode_section_info (decl, rtl, first);
17870
17871 if (first
17872 && TREE_CODE (decl) == FUNCTION_DECL
17873 && !TARGET_AIX
17874 && DEFAULT_ABI == ABI_AIX)
17875 {
17876 rtx sym_ref = XEXP (rtl, 0);
17877 size_t len = strlen (XSTR (sym_ref, 0));
17878 char *str = alloca (len + 2);
17879 str[0] = '.';
17880 memcpy (str + 1, XSTR (sym_ref, 0), len + 1);
17881 XSTR (sym_ref, 0) = ggc_alloc_string (str, len + 1);
17882 }
17883 }
17884
17885 bool
rs6000_elf_in_small_data_p(tree decl)17886 rs6000_elf_in_small_data_p (tree decl)
17887 {
17888 if (rs6000_sdata == SDATA_NONE)
17889 return false;
17890
17891 /* We want to merge strings, so we never consider them small data. */
17892 if (TREE_CODE (decl) == STRING_CST)
17893 return false;
17894
17895 /* Functions are never in the small data area. */
17896 if (TREE_CODE (decl) == FUNCTION_DECL)
17897 return false;
17898
17899 if (TREE_CODE (decl) == VAR_DECL && DECL_SECTION_NAME (decl))
17900 {
17901 const char *section = TREE_STRING_POINTER (DECL_SECTION_NAME (decl));
17902 if (strcmp (section, ".sdata") == 0
17903 || strcmp (section, ".sdata2") == 0
17904 || strcmp (section, ".sbss") == 0
17905 || strcmp (section, ".sbss2") == 0
17906 || strcmp (section, ".PPC.EMB.sdata0") == 0
17907 || strcmp (section, ".PPC.EMB.sbss0") == 0)
17908 return true;
17909 }
17910 else
17911 {
17912 HOST_WIDE_INT size = int_size_in_bytes (TREE_TYPE (decl));
17913
17914 if (size > 0
17915 && (unsigned HOST_WIDE_INT) size <= g_switch_value
17916 /* If it's not public, and we're not going to reference it there,
17917 there's no need to put it in the small data section. */
17918 && (rs6000_sdata != SDATA_DATA || TREE_PUBLIC (decl)))
17919 return true;
17920 }
17921
17922 return false;
17923 }
17924
17925 #endif /* USING_ELFOS_H */
17926
17927 /* Implement TARGET_USE_BLOCKS_FOR_CONSTANT_P. */
17928
17929 static bool
rs6000_use_blocks_for_constant_p(enum machine_mode mode,rtx x)17930 rs6000_use_blocks_for_constant_p (enum machine_mode mode, rtx x)
17931 {
17932 return !ASM_OUTPUT_SPECIAL_POOL_ENTRY_P (x, mode);
17933 }
17934
17935 /* Return a REG that occurs in ADDR with coefficient 1.
17936 ADDR can be effectively incremented by incrementing REG.
17937
17938 r0 is special and we must not select it as an address
17939 register by this routine since our caller will try to
17940 increment the returned register via an "la" instruction. */
17941
17942 rtx
find_addr_reg(rtx addr)17943 find_addr_reg (rtx addr)
17944 {
17945 while (GET_CODE (addr) == PLUS)
17946 {
17947 if (GET_CODE (XEXP (addr, 0)) == REG
17948 && REGNO (XEXP (addr, 0)) != 0)
17949 addr = XEXP (addr, 0);
17950 else if (GET_CODE (XEXP (addr, 1)) == REG
17951 && REGNO (XEXP (addr, 1)) != 0)
17952 addr = XEXP (addr, 1);
17953 else if (CONSTANT_P (XEXP (addr, 0)))
17954 addr = XEXP (addr, 1);
17955 else if (CONSTANT_P (XEXP (addr, 1)))
17956 addr = XEXP (addr, 0);
17957 else
17958 gcc_unreachable ();
17959 }
17960 gcc_assert (GET_CODE (addr) == REG && REGNO (addr) != 0);
17961 return addr;
17962 }
17963
17964 void
rs6000_fatal_bad_address(rtx op)17965 rs6000_fatal_bad_address (rtx op)
17966 {
17967 fatal_insn ("bad address", op);
17968 }
17969
17970 #if TARGET_MACHO
17971
17972 static tree branch_island_list = 0;
17973
17974 /* Remember to generate a branch island for far calls to the given
17975 function. */
17976
17977 static void
add_compiler_branch_island(tree label_name,tree function_name,int line_number)17978 add_compiler_branch_island (tree label_name, tree function_name,
17979 int line_number)
17980 {
17981 tree branch_island = build_tree_list (function_name, label_name);
17982 TREE_TYPE (branch_island) = build_int_cst (NULL_TREE, line_number);
17983 TREE_CHAIN (branch_island) = branch_island_list;
17984 branch_island_list = branch_island;
17985 }
17986
17987 #define BRANCH_ISLAND_LABEL_NAME(BRANCH_ISLAND) TREE_VALUE (BRANCH_ISLAND)
17988 #define BRANCH_ISLAND_FUNCTION_NAME(BRANCH_ISLAND) TREE_PURPOSE (BRANCH_ISLAND)
17989 #define BRANCH_ISLAND_LINE_NUMBER(BRANCH_ISLAND) \
17990 TREE_INT_CST_LOW (TREE_TYPE (BRANCH_ISLAND))
17991
17992 /* Generate far-jump branch islands for everything on the
17993 branch_island_list. Invoked immediately after the last instruction
17994 of the epilogue has been emitted; the branch-islands must be
17995 appended to, and contiguous with, the function body. Mach-O stubs
17996 are generated in machopic_output_stub(). */
17997
17998 static void
macho_branch_islands(void)17999 macho_branch_islands (void)
18000 {
18001 char tmp_buf[512];
18002 tree branch_island;
18003
18004 for (branch_island = branch_island_list;
18005 branch_island;
18006 branch_island = TREE_CHAIN (branch_island))
18007 {
18008 const char *label =
18009 IDENTIFIER_POINTER (BRANCH_ISLAND_LABEL_NAME (branch_island));
18010 const char *name =
18011 IDENTIFIER_POINTER (BRANCH_ISLAND_FUNCTION_NAME (branch_island));
18012 char name_buf[512];
18013 /* Cheap copy of the details from the Darwin ASM_OUTPUT_LABELREF(). */
18014 if (name[0] == '*' || name[0] == '&')
18015 strcpy (name_buf, name+1);
18016 else
18017 {
18018 name_buf[0] = '_';
18019 strcpy (name_buf+1, name);
18020 }
18021 strcpy (tmp_buf, "\n");
18022 strcat (tmp_buf, label);
18023 #if defined (DBX_DEBUGGING_INFO) || defined (XCOFF_DEBUGGING_INFO)
18024 if (write_symbols == DBX_DEBUG || write_symbols == XCOFF_DEBUG)
18025 dbxout_stabd (N_SLINE, BRANCH_ISLAND_LINE_NUMBER (branch_island));
18026 #endif /* DBX_DEBUGGING_INFO || XCOFF_DEBUGGING_INFO */
18027 if (flag_pic)
18028 {
18029 strcat (tmp_buf, ":\n\tmflr r0\n\tbcl 20,31,");
18030 strcat (tmp_buf, label);
18031 strcat (tmp_buf, "_pic\n");
18032 strcat (tmp_buf, label);
18033 strcat (tmp_buf, "_pic:\n\tmflr r11\n");
18034
18035 strcat (tmp_buf, "\taddis r11,r11,ha16(");
18036 strcat (tmp_buf, name_buf);
18037 strcat (tmp_buf, " - ");
18038 strcat (tmp_buf, label);
18039 strcat (tmp_buf, "_pic)\n");
18040
18041 strcat (tmp_buf, "\tmtlr r0\n");
18042
18043 strcat (tmp_buf, "\taddi r12,r11,lo16(");
18044 strcat (tmp_buf, name_buf);
18045 strcat (tmp_buf, " - ");
18046 strcat (tmp_buf, label);
18047 strcat (tmp_buf, "_pic)\n");
18048
18049 strcat (tmp_buf, "\tmtctr r12\n\tbctr\n");
18050 }
18051 else
18052 {
18053 strcat (tmp_buf, ":\nlis r12,hi16(");
18054 strcat (tmp_buf, name_buf);
18055 strcat (tmp_buf, ")\n\tori r12,r12,lo16(");
18056 strcat (tmp_buf, name_buf);
18057 strcat (tmp_buf, ")\n\tmtctr r12\n\tbctr");
18058 }
18059 output_asm_insn (tmp_buf, 0);
18060 #if defined (DBX_DEBUGGING_INFO) || defined (XCOFF_DEBUGGING_INFO)
18061 if (write_symbols == DBX_DEBUG || write_symbols == XCOFF_DEBUG)
18062 dbxout_stabd (N_SLINE, BRANCH_ISLAND_LINE_NUMBER (branch_island));
18063 #endif /* DBX_DEBUGGING_INFO || XCOFF_DEBUGGING_INFO */
18064 }
18065
18066 branch_island_list = 0;
18067 }
18068
18069 /* NO_PREVIOUS_DEF checks in the link list whether the function name is
18070 already there or not. */
18071
18072 static int
no_previous_def(tree function_name)18073 no_previous_def (tree function_name)
18074 {
18075 tree branch_island;
18076 for (branch_island = branch_island_list;
18077 branch_island;
18078 branch_island = TREE_CHAIN (branch_island))
18079 if (function_name == BRANCH_ISLAND_FUNCTION_NAME (branch_island))
18080 return 0;
18081 return 1;
18082 }
18083
18084 /* GET_PREV_LABEL gets the label name from the previous definition of
18085 the function. */
18086
18087 static tree
get_prev_label(tree function_name)18088 get_prev_label (tree function_name)
18089 {
18090 tree branch_island;
18091 for (branch_island = branch_island_list;
18092 branch_island;
18093 branch_island = TREE_CHAIN (branch_island))
18094 if (function_name == BRANCH_ISLAND_FUNCTION_NAME (branch_island))
18095 return BRANCH_ISLAND_LABEL_NAME (branch_island);
18096 return 0;
18097 }
18098
18099 #ifndef DARWIN_LINKER_GENERATES_ISLANDS
18100 #define DARWIN_LINKER_GENERATES_ISLANDS 0
18101 #endif
18102
18103 /* KEXTs still need branch islands. */
18104 #define DARWIN_GENERATE_ISLANDS (!DARWIN_LINKER_GENERATES_ISLANDS \
18105 || flag_mkernel || flag_apple_kext)
18106
18107 /* INSN is either a function call or a millicode call. It may have an
18108 unconditional jump in its delay slot.
18109
18110 CALL_DEST is the routine we are calling. */
18111
18112 char *
output_call(rtx insn,rtx * operands,int dest_operand_number,int cookie_operand_number)18113 output_call (rtx insn, rtx *operands, int dest_operand_number,
18114 int cookie_operand_number)
18115 {
18116 static char buf[256];
18117 if (DARWIN_GENERATE_ISLANDS
18118 && GET_CODE (operands[dest_operand_number]) == SYMBOL_REF
18119 && (INTVAL (operands[cookie_operand_number]) & CALL_LONG))
18120 {
18121 tree labelname;
18122 tree funname = get_identifier (XSTR (operands[dest_operand_number], 0));
18123
18124 if (no_previous_def (funname))
18125 {
18126 int line_number = 0;
18127 rtx label_rtx = gen_label_rtx ();
18128 char *label_buf, temp_buf[256];
18129 ASM_GENERATE_INTERNAL_LABEL (temp_buf, "L",
18130 CODE_LABEL_NUMBER (label_rtx));
18131 label_buf = temp_buf[0] == '*' ? temp_buf + 1 : temp_buf;
18132 labelname = get_identifier (label_buf);
18133 for (; insn && GET_CODE (insn) != NOTE; insn = PREV_INSN (insn));
18134 if (insn)
18135 line_number = NOTE_LINE_NUMBER (insn);
18136 add_compiler_branch_island (labelname, funname, line_number);
18137 }
18138 else
18139 labelname = get_prev_label (funname);
18140
18141 /* "jbsr foo, L42" is Mach-O for "Link as 'bl foo' if a 'bl'
18142 instruction will reach 'foo', otherwise link as 'bl L42'".
18143 "L42" should be a 'branch island', that will do a far jump to
18144 'foo'. Branch islands are generated in
18145 macho_branch_islands(). */
18146 sprintf (buf, "jbsr %%z%d,%.246s",
18147 dest_operand_number, IDENTIFIER_POINTER (labelname));
18148 }
18149 else
18150 sprintf (buf, "bl %%z%d", dest_operand_number);
18151 return buf;
18152 }
18153
18154 /* Generate PIC and indirect symbol stubs. */
18155
18156 void
machopic_output_stub(FILE * file,const char * symb,const char * stub)18157 machopic_output_stub (FILE *file, const char *symb, const char *stub)
18158 {
18159 unsigned int length;
18160 char *symbol_name, *lazy_ptr_name;
18161 char *local_label_0;
18162 static int label = 0;
18163
18164 /* Lose our funky encoding stuff so it doesn't contaminate the stub. */
18165 symb = (*targetm.strip_name_encoding) (symb);
18166
18167
18168 length = strlen (symb);
18169 symbol_name = alloca (length + 32);
18170 GEN_SYMBOL_NAME_FOR_SYMBOL (symbol_name, symb, length);
18171
18172 lazy_ptr_name = alloca (length + 32);
18173 GEN_LAZY_PTR_NAME_FOR_SYMBOL (lazy_ptr_name, symb, length);
18174
18175 if (flag_pic == 2)
18176 switch_to_section (darwin_sections[machopic_picsymbol_stub1_section]);
18177 else
18178 switch_to_section (darwin_sections[machopic_symbol_stub1_section]);
18179
18180 if (flag_pic == 2)
18181 {
18182 fprintf (file, "\t.align 5\n");
18183
18184 fprintf (file, "%s:\n", stub);
18185 fprintf (file, "\t.indirect_symbol %s\n", symbol_name);
18186
18187 label++;
18188 local_label_0 = alloca (sizeof ("\"L00000000000$spb\""));
18189 sprintf (local_label_0, "\"L%011d$spb\"", label);
18190
18191 fprintf (file, "\tmflr r0\n");
18192 fprintf (file, "\tbcl 20,31,%s\n", local_label_0);
18193 fprintf (file, "%s:\n\tmflr r11\n", local_label_0);
18194 fprintf (file, "\taddis r11,r11,ha16(%s-%s)\n",
18195 lazy_ptr_name, local_label_0);
18196 fprintf (file, "\tmtlr r0\n");
18197 fprintf (file, "\t%s r12,lo16(%s-%s)(r11)\n",
18198 (TARGET_64BIT ? "ldu" : "lwzu"),
18199 lazy_ptr_name, local_label_0);
18200 fprintf (file, "\tmtctr r12\n");
18201 fprintf (file, "\tbctr\n");
18202 }
18203 else
18204 {
18205 fprintf (file, "\t.align 4\n");
18206
18207 fprintf (file, "%s:\n", stub);
18208 fprintf (file, "\t.indirect_symbol %s\n", symbol_name);
18209
18210 fprintf (file, "\tlis r11,ha16(%s)\n", lazy_ptr_name);
18211 fprintf (file, "\t%s r12,lo16(%s)(r11)\n",
18212 (TARGET_64BIT ? "ldu" : "lwzu"),
18213 lazy_ptr_name);
18214 fprintf (file, "\tmtctr r12\n");
18215 fprintf (file, "\tbctr\n");
18216 }
18217
18218 switch_to_section (darwin_sections[machopic_lazy_symbol_ptr_section]);
18219 fprintf (file, "%s:\n", lazy_ptr_name);
18220 fprintf (file, "\t.indirect_symbol %s\n", symbol_name);
18221 fprintf (file, "%sdyld_stub_binding_helper\n",
18222 (TARGET_64BIT ? DOUBLE_INT_ASM_OP : "\t.long\t"));
18223 }
18224
18225 /* Legitimize PIC addresses. If the address is already
18226 position-independent, we return ORIG. Newly generated
18227 position-independent addresses go into a reg. This is REG if non
18228 zero, otherwise we allocate register(s) as necessary. */
18229
18230 #define SMALL_INT(X) ((UINTVAL (X) + 0x8000) < 0x10000)
18231
18232 rtx
rs6000_machopic_legitimize_pic_address(rtx orig,enum machine_mode mode,rtx reg)18233 rs6000_machopic_legitimize_pic_address (rtx orig, enum machine_mode mode,
18234 rtx reg)
18235 {
18236 rtx base, offset;
18237
18238 if (reg == NULL && ! reload_in_progress && ! reload_completed)
18239 reg = gen_reg_rtx (Pmode);
18240
18241 if (GET_CODE (orig) == CONST)
18242 {
18243 rtx reg_temp;
18244
18245 if (GET_CODE (XEXP (orig, 0)) == PLUS
18246 && XEXP (XEXP (orig, 0), 0) == pic_offset_table_rtx)
18247 return orig;
18248
18249 gcc_assert (GET_CODE (XEXP (orig, 0)) == PLUS);
18250
18251 /* Use a different reg for the intermediate value, as
18252 it will be marked UNCHANGING. */
18253 reg_temp = no_new_pseudos ? reg : gen_reg_rtx (Pmode);
18254 base = rs6000_machopic_legitimize_pic_address (XEXP (XEXP (orig, 0), 0),
18255 Pmode, reg_temp);
18256 offset =
18257 rs6000_machopic_legitimize_pic_address (XEXP (XEXP (orig, 0), 1),
18258 Pmode, reg);
18259
18260 if (GET_CODE (offset) == CONST_INT)
18261 {
18262 if (SMALL_INT (offset))
18263 return plus_constant (base, INTVAL (offset));
18264 else if (! reload_in_progress && ! reload_completed)
18265 offset = force_reg (Pmode, offset);
18266 else
18267 {
18268 rtx mem = force_const_mem (Pmode, orig);
18269 return machopic_legitimize_pic_address (mem, Pmode, reg);
18270 }
18271 }
18272 return gen_rtx_PLUS (Pmode, base, offset);
18273 }
18274
18275 /* Fall back on generic machopic code. */
18276 return machopic_legitimize_pic_address (orig, mode, reg);
18277 }
18278
18279 /* Output a .machine directive for the Darwin assembler, and call
18280 the generic start_file routine. */
18281
18282 static void
rs6000_darwin_file_start(void)18283 rs6000_darwin_file_start (void)
18284 {
18285 static const struct
18286 {
18287 const char *arg;
18288 const char *name;
18289 int if_set;
18290 } mapping[] = {
18291 { "ppc64", "ppc64", MASK_64BIT },
18292 { "970", "ppc970", MASK_PPC_GPOPT | MASK_MFCRF | MASK_POWERPC64 },
18293 { "power4", "ppc970", 0 },
18294 { "G5", "ppc970", 0 },
18295 { "7450", "ppc7450", 0 },
18296 { "7400", "ppc7400", MASK_ALTIVEC },
18297 { "G4", "ppc7400", 0 },
18298 { "750", "ppc750", 0 },
18299 { "740", "ppc750", 0 },
18300 { "G3", "ppc750", 0 },
18301 { "604e", "ppc604e", 0 },
18302 { "604", "ppc604", 0 },
18303 { "603e", "ppc603", 0 },
18304 { "603", "ppc603", 0 },
18305 { "601", "ppc601", 0 },
18306 { NULL, "ppc", 0 } };
18307 const char *cpu_id = "";
18308 size_t i;
18309
18310 rs6000_file_start ();
18311 darwin_file_start ();
18312
18313 /* Determine the argument to -mcpu=. Default to G3 if not specified. */
18314 for (i = 0; i < ARRAY_SIZE (rs6000_select); i++)
18315 if (rs6000_select[i].set_arch_p && rs6000_select[i].string
18316 && rs6000_select[i].string[0] != '\0')
18317 cpu_id = rs6000_select[i].string;
18318
18319 /* Look through the mapping array. Pick the first name that either
18320 matches the argument, has a bit set in IF_SET that is also set
18321 in the target flags, or has a NULL name. */
18322
18323 i = 0;
18324 while (mapping[i].arg != NULL
18325 && strcmp (mapping[i].arg, cpu_id) != 0
18326 && (mapping[i].if_set & target_flags) == 0)
18327 i++;
18328
18329 fprintf (asm_out_file, "\t.machine %s\n", mapping[i].name);
18330 }
18331
18332 #endif /* TARGET_MACHO */
18333
18334 #if TARGET_ELF
18335 static int
rs6000_elf_reloc_rw_mask(void)18336 rs6000_elf_reloc_rw_mask (void)
18337 {
18338 if (flag_pic)
18339 return 3;
18340 else if (DEFAULT_ABI == ABI_AIX)
18341 return 2;
18342 else
18343 return 0;
18344 }
18345
18346 /* Record an element in the table of global constructors. SYMBOL is
18347 a SYMBOL_REF of the function to be called; PRIORITY is a number
18348 between 0 and MAX_INIT_PRIORITY.
18349
18350 This differs from default_named_section_asm_out_constructor in
18351 that we have special handling for -mrelocatable. */
18352
18353 static void
rs6000_elf_asm_out_constructor(rtx symbol,int priority)18354 rs6000_elf_asm_out_constructor (rtx symbol, int priority)
18355 {
18356 const char *section = ".ctors";
18357 char buf[16];
18358
18359 if (priority != DEFAULT_INIT_PRIORITY)
18360 {
18361 sprintf (buf, ".ctors.%.5u",
18362 /* Invert the numbering so the linker puts us in the proper
18363 order; constructors are run from right to left, and the
18364 linker sorts in increasing order. */
18365 MAX_INIT_PRIORITY - priority);
18366 section = buf;
18367 }
18368
18369 switch_to_section (get_section (section, SECTION_WRITE, NULL));
18370 assemble_align (POINTER_SIZE);
18371
18372 if (TARGET_RELOCATABLE)
18373 {
18374 fputs ("\t.long (", asm_out_file);
18375 output_addr_const (asm_out_file, symbol);
18376 fputs (")@fixup\n", asm_out_file);
18377 }
18378 else
18379 assemble_integer (symbol, POINTER_SIZE / BITS_PER_UNIT, POINTER_SIZE, 1);
18380 }
18381
18382 static void
rs6000_elf_asm_out_destructor(rtx symbol,int priority)18383 rs6000_elf_asm_out_destructor (rtx symbol, int priority)
18384 {
18385 const char *section = ".dtors";
18386 char buf[16];
18387
18388 if (priority != DEFAULT_INIT_PRIORITY)
18389 {
18390 sprintf (buf, ".dtors.%.5u",
18391 /* Invert the numbering so the linker puts us in the proper
18392 order; constructors are run from right to left, and the
18393 linker sorts in increasing order. */
18394 MAX_INIT_PRIORITY - priority);
18395 section = buf;
18396 }
18397
18398 switch_to_section (get_section (section, SECTION_WRITE, NULL));
18399 assemble_align (POINTER_SIZE);
18400
18401 if (TARGET_RELOCATABLE)
18402 {
18403 fputs ("\t.long (", asm_out_file);
18404 output_addr_const (asm_out_file, symbol);
18405 fputs (")@fixup\n", asm_out_file);
18406 }
18407 else
18408 assemble_integer (symbol, POINTER_SIZE / BITS_PER_UNIT, POINTER_SIZE, 1);
18409 }
18410
18411 void
rs6000_elf_declare_function_name(FILE * file,const char * name,tree decl)18412 rs6000_elf_declare_function_name (FILE *file, const char *name, tree decl)
18413 {
18414 if (TARGET_64BIT)
18415 {
18416 fputs ("\t.section\t\".opd\",\"aw\"\n\t.align 3\n", file);
18417 ASM_OUTPUT_LABEL (file, name);
18418 fputs (DOUBLE_INT_ASM_OP, file);
18419 rs6000_output_function_entry (file, name);
18420 fputs (",.TOC.@tocbase,0\n\t.previous\n", file);
18421 if (DOT_SYMBOLS)
18422 {
18423 fputs ("\t.size\t", file);
18424 assemble_name (file, name);
18425 fputs (",24\n\t.type\t.", file);
18426 assemble_name (file, name);
18427 fputs (",@function\n", file);
18428 if (TREE_PUBLIC (decl) && ! DECL_WEAK (decl))
18429 {
18430 fputs ("\t.globl\t.", file);
18431 assemble_name (file, name);
18432 putc ('\n', file);
18433 }
18434 }
18435 else
18436 ASM_OUTPUT_TYPE_DIRECTIVE (file, name, "function");
18437 ASM_DECLARE_RESULT (file, DECL_RESULT (decl));
18438 rs6000_output_function_entry (file, name);
18439 fputs (":\n", file);
18440 return;
18441 }
18442
18443 if (TARGET_RELOCATABLE
18444 && !TARGET_SECURE_PLT
18445 && (get_pool_size () != 0 || current_function_profile)
18446 && uses_TOC ())
18447 {
18448 char buf[256];
18449
18450 (*targetm.asm_out.internal_label) (file, "LCL", rs6000_pic_labelno);
18451
18452 ASM_GENERATE_INTERNAL_LABEL (buf, "LCTOC", 1);
18453 fprintf (file, "\t.long ");
18454 assemble_name (file, buf);
18455 putc ('-', file);
18456 ASM_GENERATE_INTERNAL_LABEL (buf, "LCF", rs6000_pic_labelno);
18457 assemble_name (file, buf);
18458 putc ('\n', file);
18459 }
18460
18461 ASM_OUTPUT_TYPE_DIRECTIVE (file, name, "function");
18462 ASM_DECLARE_RESULT (file, DECL_RESULT (decl));
18463
18464 if (DEFAULT_ABI == ABI_AIX)
18465 {
18466 const char *desc_name, *orig_name;
18467
18468 orig_name = (*targetm.strip_name_encoding) (name);
18469 desc_name = orig_name;
18470 while (*desc_name == '.')
18471 desc_name++;
18472
18473 if (TREE_PUBLIC (decl))
18474 fprintf (file, "\t.globl %s\n", desc_name);
18475
18476 fprintf (file, "%s\n", MINIMAL_TOC_SECTION_ASM_OP);
18477 fprintf (file, "%s:\n", desc_name);
18478 fprintf (file, "\t.long %s\n", orig_name);
18479 fputs ("\t.long _GLOBAL_OFFSET_TABLE_\n", file);
18480 if (DEFAULT_ABI == ABI_AIX)
18481 fputs ("\t.long 0\n", file);
18482 fprintf (file, "\t.previous\n");
18483 }
18484 ASM_OUTPUT_LABEL (file, name);
18485 }
18486
18487 static void
rs6000_elf_end_indicate_exec_stack(void)18488 rs6000_elf_end_indicate_exec_stack (void)
18489 {
18490 if (NEED_INDICATE_EXEC_STACK)
18491 file_end_indicate_exec_stack ();
18492 }
18493 #endif
18494
18495 #if TARGET_XCOFF
18496 static void
rs6000_xcoff_asm_output_anchor(rtx symbol)18497 rs6000_xcoff_asm_output_anchor (rtx symbol)
18498 {
18499 char buffer[100];
18500
18501 sprintf (buffer, "$ + " HOST_WIDE_INT_PRINT_DEC,
18502 SYMBOL_REF_BLOCK_OFFSET (symbol));
18503 ASM_OUTPUT_DEF (asm_out_file, XSTR (symbol, 0), buffer);
18504 }
18505
18506 static void
rs6000_xcoff_asm_globalize_label(FILE * stream,const char * name)18507 rs6000_xcoff_asm_globalize_label (FILE *stream, const char *name)
18508 {
18509 fputs (GLOBAL_ASM_OP, stream);
18510 RS6000_OUTPUT_BASENAME (stream, name);
18511 putc ('\n', stream);
18512 }
18513
18514 /* A get_unnamed_decl callback, used for read-only sections. PTR
18515 points to the section string variable. */
18516
18517 static void
rs6000_xcoff_output_readonly_section_asm_op(const void * directive)18518 rs6000_xcoff_output_readonly_section_asm_op (const void *directive)
18519 {
18520 fprintf (asm_out_file, "\t.csect %s[RO],3\n",
18521 *(const char *const *) directive);
18522 }
18523
18524 /* Likewise for read-write sections. */
18525
18526 static void
rs6000_xcoff_output_readwrite_section_asm_op(const void * directive)18527 rs6000_xcoff_output_readwrite_section_asm_op (const void *directive)
18528 {
18529 fprintf (asm_out_file, "\t.csect %s[RW],3\n",
18530 *(const char *const *) directive);
18531 }
18532
18533 /* A get_unnamed_section callback, used for switching to toc_section. */
18534
18535 static void
rs6000_xcoff_output_toc_section_asm_op(const void * data ATTRIBUTE_UNUSED)18536 rs6000_xcoff_output_toc_section_asm_op (const void *data ATTRIBUTE_UNUSED)
18537 {
18538 if (TARGET_MINIMAL_TOC)
18539 {
18540 /* toc_section is always selected at least once from
18541 rs6000_xcoff_file_start, so this is guaranteed to
18542 always be defined once and only once in each file. */
18543 if (!toc_initialized)
18544 {
18545 fputs ("\t.toc\nLCTOC..1:\n", asm_out_file);
18546 fputs ("\t.tc toc_table[TC],toc_table[RW]\n", asm_out_file);
18547 toc_initialized = 1;
18548 }
18549 fprintf (asm_out_file, "\t.csect toc_table[RW]%s\n",
18550 (TARGET_32BIT ? "" : ",3"));
18551 }
18552 else
18553 fputs ("\t.toc\n", asm_out_file);
18554 }
18555
18556 /* Implement TARGET_ASM_INIT_SECTIONS. */
18557
18558 static void
rs6000_xcoff_asm_init_sections(void)18559 rs6000_xcoff_asm_init_sections (void)
18560 {
18561 read_only_data_section
18562 = get_unnamed_section (0, rs6000_xcoff_output_readonly_section_asm_op,
18563 &xcoff_read_only_section_name);
18564
18565 private_data_section
18566 = get_unnamed_section (SECTION_WRITE,
18567 rs6000_xcoff_output_readwrite_section_asm_op,
18568 &xcoff_private_data_section_name);
18569
18570 read_only_private_data_section
18571 = get_unnamed_section (0, rs6000_xcoff_output_readonly_section_asm_op,
18572 &xcoff_private_data_section_name);
18573
18574 toc_section
18575 = get_unnamed_section (0, rs6000_xcoff_output_toc_section_asm_op, NULL);
18576
18577 readonly_data_section = read_only_data_section;
18578 exception_section = data_section;
18579 }
18580
18581 static int
rs6000_xcoff_reloc_rw_mask(void)18582 rs6000_xcoff_reloc_rw_mask (void)
18583 {
18584 return 3;
18585 }
18586
18587 static void
rs6000_xcoff_asm_named_section(const char * name,unsigned int flags,tree decl ATTRIBUTE_UNUSED)18588 rs6000_xcoff_asm_named_section (const char *name, unsigned int flags,
18589 tree decl ATTRIBUTE_UNUSED)
18590 {
18591 int smclass;
18592 static const char * const suffix[3] = { "PR", "RO", "RW" };
18593
18594 if (flags & SECTION_CODE)
18595 smclass = 0;
18596 else if (flags & SECTION_WRITE)
18597 smclass = 2;
18598 else
18599 smclass = 1;
18600
18601 fprintf (asm_out_file, "\t.csect %s%s[%s],%u\n",
18602 (flags & SECTION_CODE) ? "." : "",
18603 name, suffix[smclass], flags & SECTION_ENTSIZE);
18604 }
18605
18606 static section *
rs6000_xcoff_select_section(tree decl,int reloc,unsigned HOST_WIDE_INT align ATTRIBUTE_UNUSED)18607 rs6000_xcoff_select_section (tree decl, int reloc,
18608 unsigned HOST_WIDE_INT align ATTRIBUTE_UNUSED)
18609 {
18610 if (decl_readonly_section (decl, reloc))
18611 {
18612 if (TREE_PUBLIC (decl))
18613 return read_only_data_section;
18614 else
18615 return read_only_private_data_section;
18616 }
18617 else
18618 {
18619 if (TREE_PUBLIC (decl))
18620 return data_section;
18621 else
18622 return private_data_section;
18623 }
18624 }
18625
18626 static void
rs6000_xcoff_unique_section(tree decl,int reloc ATTRIBUTE_UNUSED)18627 rs6000_xcoff_unique_section (tree decl, int reloc ATTRIBUTE_UNUSED)
18628 {
18629 const char *name;
18630
18631 /* Use select_section for private and uninitialized data. */
18632 if (!TREE_PUBLIC (decl)
18633 || DECL_COMMON (decl)
18634 || DECL_INITIAL (decl) == NULL_TREE
18635 || DECL_INITIAL (decl) == error_mark_node
18636 || (flag_zero_initialized_in_bss
18637 && initializer_zerop (DECL_INITIAL (decl))))
18638 return;
18639
18640 name = IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (decl));
18641 name = (*targetm.strip_name_encoding) (name);
18642 DECL_SECTION_NAME (decl) = build_string (strlen (name), name);
18643 }
18644
18645 /* Select section for constant in constant pool.
18646
18647 On RS/6000, all constants are in the private read-only data area.
18648 However, if this is being placed in the TOC it must be output as a
18649 toc entry. */
18650
18651 static section *
rs6000_xcoff_select_rtx_section(enum machine_mode mode,rtx x,unsigned HOST_WIDE_INT align ATTRIBUTE_UNUSED)18652 rs6000_xcoff_select_rtx_section (enum machine_mode mode, rtx x,
18653 unsigned HOST_WIDE_INT align ATTRIBUTE_UNUSED)
18654 {
18655 if (ASM_OUTPUT_SPECIAL_POOL_ENTRY_P (x, mode))
18656 return toc_section;
18657 else
18658 return read_only_private_data_section;
18659 }
18660
18661 /* Remove any trailing [DS] or the like from the symbol name. */
18662
18663 static const char *
rs6000_xcoff_strip_name_encoding(const char * name)18664 rs6000_xcoff_strip_name_encoding (const char *name)
18665 {
18666 size_t len;
18667 if (*name == '*')
18668 name++;
18669 len = strlen (name);
18670 if (name[len - 1] == ']')
18671 return ggc_alloc_string (name, len - 4);
18672 else
18673 return name;
18674 }
18675
18676 /* Section attributes. AIX is always PIC. */
18677
18678 static unsigned int
rs6000_xcoff_section_type_flags(tree decl,const char * name,int reloc)18679 rs6000_xcoff_section_type_flags (tree decl, const char *name, int reloc)
18680 {
18681 unsigned int align;
18682 unsigned int flags = default_section_type_flags (decl, name, reloc);
18683
18684 /* Align to at least UNIT size. */
18685 if (flags & SECTION_CODE)
18686 align = MIN_UNITS_PER_WORD;
18687 else
18688 /* Increase alignment of large objects if not already stricter. */
18689 align = MAX ((DECL_ALIGN (decl) / BITS_PER_UNIT),
18690 int_size_in_bytes (TREE_TYPE (decl)) > MIN_UNITS_PER_WORD
18691 ? UNITS_PER_FP_WORD : MIN_UNITS_PER_WORD);
18692
18693 return flags | (exact_log2 (align) & SECTION_ENTSIZE);
18694 }
18695
18696 /* Output at beginning of assembler file.
18697
18698 Initialize the section names for the RS/6000 at this point.
18699
18700 Specify filename, including full path, to assembler.
18701
18702 We want to go into the TOC section so at least one .toc will be emitted.
18703 Also, in order to output proper .bs/.es pairs, we need at least one static
18704 [RW] section emitted.
18705
18706 Finally, declare mcount when profiling to make the assembler happy. */
18707
18708 static void
rs6000_xcoff_file_start(void)18709 rs6000_xcoff_file_start (void)
18710 {
18711 rs6000_gen_section_name (&xcoff_bss_section_name,
18712 main_input_filename, ".bss_");
18713 rs6000_gen_section_name (&xcoff_private_data_section_name,
18714 main_input_filename, ".rw_");
18715 rs6000_gen_section_name (&xcoff_read_only_section_name,
18716 main_input_filename, ".ro_");
18717
18718 fputs ("\t.file\t", asm_out_file);
18719 output_quoted_string (asm_out_file, main_input_filename);
18720 fputc ('\n', asm_out_file);
18721 if (write_symbols != NO_DEBUG)
18722 switch_to_section (private_data_section);
18723 switch_to_section (text_section);
18724 if (profile_flag)
18725 fprintf (asm_out_file, "\t.extern %s\n", RS6000_MCOUNT);
18726 rs6000_file_start ();
18727 }
18728
18729 /* Output at end of assembler file.
18730 On the RS/6000, referencing data should automatically pull in text. */
18731
18732 static void
rs6000_xcoff_file_end(void)18733 rs6000_xcoff_file_end (void)
18734 {
18735 switch_to_section (text_section);
18736 fputs ("_section_.text:\n", asm_out_file);
18737 switch_to_section (data_section);
18738 fputs (TARGET_32BIT
18739 ? "\t.long _section_.text\n" : "\t.llong _section_.text\n",
18740 asm_out_file);
18741 }
18742 #endif /* TARGET_XCOFF */
18743
18744 /* Compute a (partial) cost for rtx X. Return true if the complete
18745 cost has been computed, and false if subexpressions should be
18746 scanned. In either case, *TOTAL contains the cost result. */
18747
18748 static bool
rs6000_rtx_costs(rtx x,int code,int outer_code,int * total)18749 rs6000_rtx_costs (rtx x, int code, int outer_code, int *total)
18750 {
18751 enum machine_mode mode = GET_MODE (x);
18752
18753 switch (code)
18754 {
18755 /* On the RS/6000, if it is valid in the insn, it is free. */
18756 case CONST_INT:
18757 if (((outer_code == SET
18758 || outer_code == PLUS
18759 || outer_code == MINUS)
18760 && (satisfies_constraint_I (x)
18761 || satisfies_constraint_L (x)))
18762 || (outer_code == AND
18763 && (satisfies_constraint_K (x)
18764 || (mode == SImode
18765 ? satisfies_constraint_L (x)
18766 : satisfies_constraint_J (x))
18767 || mask_operand (x, mode)
18768 || (mode == DImode
18769 && mask64_operand (x, DImode))))
18770 || ((outer_code == IOR || outer_code == XOR)
18771 && (satisfies_constraint_K (x)
18772 || (mode == SImode
18773 ? satisfies_constraint_L (x)
18774 : satisfies_constraint_J (x))))
18775 || outer_code == ASHIFT
18776 || outer_code == ASHIFTRT
18777 || outer_code == LSHIFTRT
18778 || outer_code == ROTATE
18779 || outer_code == ROTATERT
18780 || outer_code == ZERO_EXTRACT
18781 || (outer_code == MULT
18782 && satisfies_constraint_I (x))
18783 || ((outer_code == DIV || outer_code == UDIV
18784 || outer_code == MOD || outer_code == UMOD)
18785 && exact_log2 (INTVAL (x)) >= 0)
18786 || (outer_code == COMPARE
18787 && (satisfies_constraint_I (x)
18788 || satisfies_constraint_K (x)))
18789 || (outer_code == EQ
18790 && (satisfies_constraint_I (x)
18791 || satisfies_constraint_K (x)
18792 || (mode == SImode
18793 ? satisfies_constraint_L (x)
18794 : satisfies_constraint_J (x))))
18795 || (outer_code == GTU
18796 && satisfies_constraint_I (x))
18797 || (outer_code == LTU
18798 && satisfies_constraint_P (x)))
18799 {
18800 *total = 0;
18801 return true;
18802 }
18803 else if ((outer_code == PLUS
18804 && reg_or_add_cint_operand (x, VOIDmode))
18805 || (outer_code == MINUS
18806 && reg_or_sub_cint_operand (x, VOIDmode))
18807 || ((outer_code == SET
18808 || outer_code == IOR
18809 || outer_code == XOR)
18810 && (INTVAL (x)
18811 & ~ (unsigned HOST_WIDE_INT) 0xffffffff) == 0))
18812 {
18813 *total = COSTS_N_INSNS (1);
18814 return true;
18815 }
18816 /* FALLTHRU */
18817
18818 case CONST_DOUBLE:
18819 if (mode == DImode && code == CONST_DOUBLE)
18820 {
18821 if ((outer_code == IOR || outer_code == XOR)
18822 && CONST_DOUBLE_HIGH (x) == 0
18823 && (CONST_DOUBLE_LOW (x)
18824 & ~ (unsigned HOST_WIDE_INT) 0xffff) == 0)
18825 {
18826 *total = 0;
18827 return true;
18828 }
18829 else if ((outer_code == AND && and64_2_operand (x, DImode))
18830 || ((outer_code == SET
18831 || outer_code == IOR
18832 || outer_code == XOR)
18833 && CONST_DOUBLE_HIGH (x) == 0))
18834 {
18835 *total = COSTS_N_INSNS (1);
18836 return true;
18837 }
18838 }
18839 /* FALLTHRU */
18840
18841 case CONST:
18842 case HIGH:
18843 case SYMBOL_REF:
18844 case MEM:
18845 /* When optimizing for size, MEM should be slightly more expensive
18846 than generating address, e.g., (plus (reg) (const)).
18847 L1 cache latency is about two instructions. */
18848 *total = optimize_size ? COSTS_N_INSNS (1) + 1 : COSTS_N_INSNS (2);
18849 return true;
18850
18851 case LABEL_REF:
18852 *total = 0;
18853 return true;
18854
18855 case PLUS:
18856 if (mode == DFmode)
18857 {
18858 if (GET_CODE (XEXP (x, 0)) == MULT)
18859 {
18860 /* FNMA accounted in outer NEG. */
18861 if (outer_code == NEG)
18862 *total = rs6000_cost->dmul - rs6000_cost->fp;
18863 else
18864 *total = rs6000_cost->dmul;
18865 }
18866 else
18867 *total = rs6000_cost->fp;
18868 }
18869 else if (mode == SFmode)
18870 {
18871 /* FNMA accounted in outer NEG. */
18872 if (outer_code == NEG && GET_CODE (XEXP (x, 0)) == MULT)
18873 *total = 0;
18874 else
18875 *total = rs6000_cost->fp;
18876 }
18877 else
18878 *total = COSTS_N_INSNS (1);
18879 return false;
18880
18881 case MINUS:
18882 if (mode == DFmode)
18883 {
18884 if (GET_CODE (XEXP (x, 0)) == MULT)
18885 {
18886 /* FNMA accounted in outer NEG. */
18887 if (outer_code == NEG)
18888 *total = 0;
18889 else
18890 *total = rs6000_cost->dmul;
18891 }
18892 else
18893 *total = rs6000_cost->fp;
18894 }
18895 else if (mode == SFmode)
18896 {
18897 /* FNMA accounted in outer NEG. */
18898 if (outer_code == NEG && GET_CODE (XEXP (x, 0)) == MULT)
18899 *total = 0;
18900 else
18901 *total = rs6000_cost->fp;
18902 }
18903 else
18904 *total = COSTS_N_INSNS (1);
18905 return false;
18906
18907 case MULT:
18908 if (GET_CODE (XEXP (x, 1)) == CONST_INT
18909 && satisfies_constraint_I (XEXP (x, 1)))
18910 {
18911 if (INTVAL (XEXP (x, 1)) >= -256
18912 && INTVAL (XEXP (x, 1)) <= 255)
18913 *total = rs6000_cost->mulsi_const9;
18914 else
18915 *total = rs6000_cost->mulsi_const;
18916 }
18917 /* FMA accounted in outer PLUS/MINUS. */
18918 else if ((mode == DFmode || mode == SFmode)
18919 && (outer_code == PLUS || outer_code == MINUS))
18920 *total = 0;
18921 else if (mode == DFmode)
18922 *total = rs6000_cost->dmul;
18923 else if (mode == SFmode)
18924 *total = rs6000_cost->fp;
18925 else if (mode == DImode)
18926 *total = rs6000_cost->muldi;
18927 else
18928 *total = rs6000_cost->mulsi;
18929 return false;
18930
18931 case DIV:
18932 case MOD:
18933 if (FLOAT_MODE_P (mode))
18934 {
18935 *total = mode == DFmode ? rs6000_cost->ddiv
18936 : rs6000_cost->sdiv;
18937 return false;
18938 }
18939 /* FALLTHRU */
18940
18941 case UDIV:
18942 case UMOD:
18943 if (GET_CODE (XEXP (x, 1)) == CONST_INT
18944 && exact_log2 (INTVAL (XEXP (x, 1))) >= 0)
18945 {
18946 if (code == DIV || code == MOD)
18947 /* Shift, addze */
18948 *total = COSTS_N_INSNS (2);
18949 else
18950 /* Shift */
18951 *total = COSTS_N_INSNS (1);
18952 }
18953 else
18954 {
18955 if (GET_MODE (XEXP (x, 1)) == DImode)
18956 *total = rs6000_cost->divdi;
18957 else
18958 *total = rs6000_cost->divsi;
18959 }
18960 /* Add in shift and subtract for MOD. */
18961 if (code == MOD || code == UMOD)
18962 *total += COSTS_N_INSNS (2);
18963 return false;
18964
18965 case FFS:
18966 *total = COSTS_N_INSNS (4);
18967 return false;
18968
18969 case NOT:
18970 if (outer_code == AND || outer_code == IOR || outer_code == XOR)
18971 {
18972 *total = 0;
18973 return false;
18974 }
18975 /* FALLTHRU */
18976
18977 case AND:
18978 case IOR:
18979 case XOR:
18980 case ZERO_EXTRACT:
18981 *total = COSTS_N_INSNS (1);
18982 return false;
18983
18984 case ASHIFT:
18985 case ASHIFTRT:
18986 case LSHIFTRT:
18987 case ROTATE:
18988 case ROTATERT:
18989 /* Handle mul_highpart. */
18990 if (outer_code == TRUNCATE
18991 && GET_CODE (XEXP (x, 0)) == MULT)
18992 {
18993 if (mode == DImode)
18994 *total = rs6000_cost->muldi;
18995 else
18996 *total = rs6000_cost->mulsi;
18997 return true;
18998 }
18999 else if (outer_code == AND)
19000 *total = 0;
19001 else
19002 *total = COSTS_N_INSNS (1);
19003 return false;
19004
19005 case SIGN_EXTEND:
19006 case ZERO_EXTEND:
19007 if (GET_CODE (XEXP (x, 0)) == MEM)
19008 *total = 0;
19009 else
19010 *total = COSTS_N_INSNS (1);
19011 return false;
19012
19013 case COMPARE:
19014 case NEG:
19015 case ABS:
19016 if (!FLOAT_MODE_P (mode))
19017 {
19018 *total = COSTS_N_INSNS (1);
19019 return false;
19020 }
19021 /* FALLTHRU */
19022
19023 case FLOAT:
19024 case UNSIGNED_FLOAT:
19025 case FIX:
19026 case UNSIGNED_FIX:
19027 case FLOAT_TRUNCATE:
19028 *total = rs6000_cost->fp;
19029 return false;
19030
19031 case FLOAT_EXTEND:
19032 if (mode == DFmode)
19033 *total = 0;
19034 else
19035 *total = rs6000_cost->fp;
19036 return false;
19037
19038 case UNSPEC:
19039 switch (XINT (x, 1))
19040 {
19041 case UNSPEC_FRSP:
19042 *total = rs6000_cost->fp;
19043 return true;
19044
19045 default:
19046 break;
19047 }
19048 break;
19049
19050 case CALL:
19051 case IF_THEN_ELSE:
19052 if (optimize_size)
19053 {
19054 *total = COSTS_N_INSNS (1);
19055 return true;
19056 }
19057 else if (FLOAT_MODE_P (mode)
19058 && TARGET_PPC_GFXOPT && TARGET_HARD_FLOAT && TARGET_FPRS)
19059 {
19060 *total = rs6000_cost->fp;
19061 return false;
19062 }
19063 break;
19064
19065 case EQ:
19066 case GTU:
19067 case LTU:
19068 /* Carry bit requires mode == Pmode.
19069 NEG or PLUS already counted so only add one. */
19070 if (mode == Pmode
19071 && (outer_code == NEG || outer_code == PLUS))
19072 {
19073 *total = COSTS_N_INSNS (1);
19074 return true;
19075 }
19076 if (outer_code == SET)
19077 {
19078 if (XEXP (x, 1) == const0_rtx)
19079 {
19080 *total = COSTS_N_INSNS (2);
19081 return true;
19082 }
19083 else if (mode == Pmode)
19084 {
19085 *total = COSTS_N_INSNS (3);
19086 return false;
19087 }
19088 }
19089 /* FALLTHRU */
19090
19091 case GT:
19092 case LT:
19093 case UNORDERED:
19094 if (outer_code == SET && (XEXP (x, 1) == const0_rtx))
19095 {
19096 *total = COSTS_N_INSNS (2);
19097 return true;
19098 }
19099 /* CC COMPARE. */
19100 if (outer_code == COMPARE)
19101 {
19102 *total = 0;
19103 return true;
19104 }
19105 break;
19106
19107 default:
19108 break;
19109 }
19110
19111 return false;
19112 }
19113
19114 /* A C expression returning the cost of moving data from a register of class
19115 CLASS1 to one of CLASS2. */
19116
19117 int
rs6000_register_move_cost(enum machine_mode mode,enum reg_class from,enum reg_class to)19118 rs6000_register_move_cost (enum machine_mode mode,
19119 enum reg_class from, enum reg_class to)
19120 {
19121 /* Moves from/to GENERAL_REGS. */
19122 if (reg_classes_intersect_p (to, GENERAL_REGS)
19123 || reg_classes_intersect_p (from, GENERAL_REGS))
19124 {
19125 if (! reg_classes_intersect_p (to, GENERAL_REGS))
19126 from = to;
19127
19128 if (from == FLOAT_REGS || from == ALTIVEC_REGS)
19129 return (rs6000_memory_move_cost (mode, from, 0)
19130 + rs6000_memory_move_cost (mode, GENERAL_REGS, 0));
19131
19132 /* It's more expensive to move CR_REGS than CR0_REGS because of the
19133 shift. */
19134 else if (from == CR_REGS)
19135 return 4;
19136
19137 else
19138 /* A move will cost one instruction per GPR moved. */
19139 return 2 * hard_regno_nregs[0][mode];
19140 }
19141
19142 /* Moving between two similar registers is just one instruction. */
19143 else if (reg_classes_intersect_p (to, from))
19144 return mode == TFmode ? 4 : 2;
19145
19146 /* Everything else has to go through GENERAL_REGS. */
19147 else
19148 return (rs6000_register_move_cost (mode, GENERAL_REGS, to)
19149 + rs6000_register_move_cost (mode, from, GENERAL_REGS));
19150 }
19151
19152 /* A C expressions returning the cost of moving data of MODE from a register to
19153 or from memory. */
19154
19155 int
rs6000_memory_move_cost(enum machine_mode mode,enum reg_class class,int in ATTRIBUTE_UNUSED)19156 rs6000_memory_move_cost (enum machine_mode mode, enum reg_class class,
19157 int in ATTRIBUTE_UNUSED)
19158 {
19159 if (reg_classes_intersect_p (class, GENERAL_REGS))
19160 return 4 * hard_regno_nregs[0][mode];
19161 else if (reg_classes_intersect_p (class, FLOAT_REGS))
19162 return 4 * hard_regno_nregs[32][mode];
19163 else if (reg_classes_intersect_p (class, ALTIVEC_REGS))
19164 return 4 * hard_regno_nregs[FIRST_ALTIVEC_REGNO][mode];
19165 else
19166 return 4 + rs6000_register_move_cost (mode, class, GENERAL_REGS);
19167 }
19168
19169 /* Newton-Raphson approximation of single-precision floating point divide n/d.
19170 Assumes no trapping math and finite arguments. */
19171
19172 void
rs6000_emit_swdivsf(rtx res,rtx n,rtx d)19173 rs6000_emit_swdivsf (rtx res, rtx n, rtx d)
19174 {
19175 rtx x0, e0, e1, y1, u0, v0, one;
19176
19177 x0 = gen_reg_rtx (SFmode);
19178 e0 = gen_reg_rtx (SFmode);
19179 e1 = gen_reg_rtx (SFmode);
19180 y1 = gen_reg_rtx (SFmode);
19181 u0 = gen_reg_rtx (SFmode);
19182 v0 = gen_reg_rtx (SFmode);
19183 one = force_reg (SFmode, CONST_DOUBLE_FROM_REAL_VALUE (dconst1, SFmode));
19184
19185 /* x0 = 1./d estimate */
19186 emit_insn (gen_rtx_SET (VOIDmode, x0,
19187 gen_rtx_UNSPEC (SFmode, gen_rtvec (1, d),
19188 UNSPEC_FRES)));
19189 /* e0 = 1. - d * x0 */
19190 emit_insn (gen_rtx_SET (VOIDmode, e0,
19191 gen_rtx_MINUS (SFmode, one,
19192 gen_rtx_MULT (SFmode, d, x0))));
19193 /* e1 = e0 + e0 * e0 */
19194 emit_insn (gen_rtx_SET (VOIDmode, e1,
19195 gen_rtx_PLUS (SFmode,
19196 gen_rtx_MULT (SFmode, e0, e0), e0)));
19197 /* y1 = x0 + e1 * x0 */
19198 emit_insn (gen_rtx_SET (VOIDmode, y1,
19199 gen_rtx_PLUS (SFmode,
19200 gen_rtx_MULT (SFmode, e1, x0), x0)));
19201 /* u0 = n * y1 */
19202 emit_insn (gen_rtx_SET (VOIDmode, u0,
19203 gen_rtx_MULT (SFmode, n, y1)));
19204 /* v0 = n - d * u0 */
19205 emit_insn (gen_rtx_SET (VOIDmode, v0,
19206 gen_rtx_MINUS (SFmode, n,
19207 gen_rtx_MULT (SFmode, d, u0))));
19208 /* res = u0 + v0 * y1 */
19209 emit_insn (gen_rtx_SET (VOIDmode, res,
19210 gen_rtx_PLUS (SFmode,
19211 gen_rtx_MULT (SFmode, v0, y1), u0)));
19212 }
19213
19214 /* Newton-Raphson approximation of double-precision floating point divide n/d.
19215 Assumes no trapping math and finite arguments. */
19216
19217 void
rs6000_emit_swdivdf(rtx res,rtx n,rtx d)19218 rs6000_emit_swdivdf (rtx res, rtx n, rtx d)
19219 {
19220 rtx x0, e0, e1, e2, y1, y2, y3, u0, v0, one;
19221
19222 x0 = gen_reg_rtx (DFmode);
19223 e0 = gen_reg_rtx (DFmode);
19224 e1 = gen_reg_rtx (DFmode);
19225 e2 = gen_reg_rtx (DFmode);
19226 y1 = gen_reg_rtx (DFmode);
19227 y2 = gen_reg_rtx (DFmode);
19228 y3 = gen_reg_rtx (DFmode);
19229 u0 = gen_reg_rtx (DFmode);
19230 v0 = gen_reg_rtx (DFmode);
19231 one = force_reg (DFmode, CONST_DOUBLE_FROM_REAL_VALUE (dconst1, DFmode));
19232
19233 /* x0 = 1./d estimate */
19234 emit_insn (gen_rtx_SET (VOIDmode, x0,
19235 gen_rtx_UNSPEC (DFmode, gen_rtvec (1, d),
19236 UNSPEC_FRES)));
19237 /* e0 = 1. - d * x0 */
19238 emit_insn (gen_rtx_SET (VOIDmode, e0,
19239 gen_rtx_MINUS (DFmode, one,
19240 gen_rtx_MULT (SFmode, d, x0))));
19241 /* y1 = x0 + e0 * x0 */
19242 emit_insn (gen_rtx_SET (VOIDmode, y1,
19243 gen_rtx_PLUS (DFmode,
19244 gen_rtx_MULT (DFmode, e0, x0), x0)));
19245 /* e1 = e0 * e0 */
19246 emit_insn (gen_rtx_SET (VOIDmode, e1,
19247 gen_rtx_MULT (DFmode, e0, e0)));
19248 /* y2 = y1 + e1 * y1 */
19249 emit_insn (gen_rtx_SET (VOIDmode, y2,
19250 gen_rtx_PLUS (DFmode,
19251 gen_rtx_MULT (DFmode, e1, y1), y1)));
19252 /* e2 = e1 * e1 */
19253 emit_insn (gen_rtx_SET (VOIDmode, e2,
19254 gen_rtx_MULT (DFmode, e1, e1)));
19255 /* y3 = y2 + e2 * y2 */
19256 emit_insn (gen_rtx_SET (VOIDmode, y3,
19257 gen_rtx_PLUS (DFmode,
19258 gen_rtx_MULT (DFmode, e2, y2), y2)));
19259 /* u0 = n * y3 */
19260 emit_insn (gen_rtx_SET (VOIDmode, u0,
19261 gen_rtx_MULT (DFmode, n, y3)));
19262 /* v0 = n - d * u0 */
19263 emit_insn (gen_rtx_SET (VOIDmode, v0,
19264 gen_rtx_MINUS (DFmode, n,
19265 gen_rtx_MULT (DFmode, d, u0))));
19266 /* res = u0 + v0 * y3 */
19267 emit_insn (gen_rtx_SET (VOIDmode, res,
19268 gen_rtx_PLUS (DFmode,
19269 gen_rtx_MULT (DFmode, v0, y3), u0)));
19270 }
19271
19272 /* Return an RTX representing where to find the function value of a
19273 function returning MODE. */
19274 static rtx
rs6000_complex_function_value(enum machine_mode mode)19275 rs6000_complex_function_value (enum machine_mode mode)
19276 {
19277 unsigned int regno;
19278 rtx r1, r2;
19279 enum machine_mode inner = GET_MODE_INNER (mode);
19280 unsigned int inner_bytes = GET_MODE_SIZE (inner);
19281
19282 if (FLOAT_MODE_P (mode) && TARGET_HARD_FLOAT && TARGET_FPRS)
19283 regno = FP_ARG_RETURN;
19284 else
19285 {
19286 regno = GP_ARG_RETURN;
19287
19288 /* 32-bit is OK since it'll go in r3/r4. */
19289 if (TARGET_32BIT && inner_bytes >= 4)
19290 return gen_rtx_REG (mode, regno);
19291 }
19292
19293 if (inner_bytes >= 8)
19294 return gen_rtx_REG (mode, regno);
19295
19296 r1 = gen_rtx_EXPR_LIST (inner, gen_rtx_REG (inner, regno),
19297 const0_rtx);
19298 r2 = gen_rtx_EXPR_LIST (inner, gen_rtx_REG (inner, regno + 1),
19299 GEN_INT (inner_bytes));
19300 return gen_rtx_PARALLEL (mode, gen_rtvec (2, r1, r2));
19301 }
19302
19303 /* Define how to find the value returned by a function.
19304 VALTYPE is the data type of the value (as a tree).
19305 If the precise function being called is known, FUNC is its FUNCTION_DECL;
19306 otherwise, FUNC is 0.
19307
19308 On the SPE, both FPs and vectors are returned in r3.
19309
19310 On RS/6000 an integer value is in r3 and a floating-point value is in
19311 fp1, unless -msoft-float. */
19312
19313 rtx
rs6000_function_value(tree valtype,tree func ATTRIBUTE_UNUSED)19314 rs6000_function_value (tree valtype, tree func ATTRIBUTE_UNUSED)
19315 {
19316 enum machine_mode mode;
19317 unsigned int regno;
19318
19319 /* Special handling for structs in darwin64. */
19320 if (rs6000_darwin64_abi
19321 && TYPE_MODE (valtype) == BLKmode
19322 && TREE_CODE (valtype) == RECORD_TYPE
19323 && int_size_in_bytes (valtype) > 0)
19324 {
19325 CUMULATIVE_ARGS valcum;
19326 rtx valret;
19327
19328 valcum.words = 0;
19329 valcum.fregno = FP_ARG_MIN_REG;
19330 valcum.vregno = ALTIVEC_ARG_MIN_REG;
19331 /* Do a trial code generation as if this were going to be passed as
19332 an argument; if any part goes in memory, we return NULL. */
19333 valret = rs6000_darwin64_record_arg (&valcum, valtype, 1, true);
19334 if (valret)
19335 return valret;
19336 /* Otherwise fall through to standard ABI rules. */
19337 }
19338
19339 if (TARGET_32BIT && TARGET_POWERPC64 && TYPE_MODE (valtype) == DImode)
19340 {
19341 /* Long long return value need be split in -mpowerpc64, 32bit ABI. */
19342 return gen_rtx_PARALLEL (DImode,
19343 gen_rtvec (2,
19344 gen_rtx_EXPR_LIST (VOIDmode,
19345 gen_rtx_REG (SImode, GP_ARG_RETURN),
19346 const0_rtx),
19347 gen_rtx_EXPR_LIST (VOIDmode,
19348 gen_rtx_REG (SImode,
19349 GP_ARG_RETURN + 1),
19350 GEN_INT (4))));
19351 }
19352 if (TARGET_32BIT && TARGET_POWERPC64 && TYPE_MODE (valtype) == DCmode)
19353 {
19354 return gen_rtx_PARALLEL (DCmode,
19355 gen_rtvec (4,
19356 gen_rtx_EXPR_LIST (VOIDmode,
19357 gen_rtx_REG (SImode, GP_ARG_RETURN),
19358 const0_rtx),
19359 gen_rtx_EXPR_LIST (VOIDmode,
19360 gen_rtx_REG (SImode,
19361 GP_ARG_RETURN + 1),
19362 GEN_INT (4)),
19363 gen_rtx_EXPR_LIST (VOIDmode,
19364 gen_rtx_REG (SImode,
19365 GP_ARG_RETURN + 2),
19366 GEN_INT (8)),
19367 gen_rtx_EXPR_LIST (VOIDmode,
19368 gen_rtx_REG (SImode,
19369 GP_ARG_RETURN + 3),
19370 GEN_INT (12))));
19371 }
19372
19373 mode = TYPE_MODE (valtype);
19374 if ((INTEGRAL_TYPE_P (valtype) && GET_MODE_BITSIZE (mode) < BITS_PER_WORD)
19375 || POINTER_TYPE_P (valtype))
19376 mode = TARGET_32BIT ? SImode : DImode;
19377
19378 if (DECIMAL_FLOAT_MODE_P (mode))
19379 regno = GP_ARG_RETURN;
19380 else if (SCALAR_FLOAT_TYPE_P (valtype) && TARGET_HARD_FLOAT && TARGET_FPRS)
19381 regno = FP_ARG_RETURN;
19382 else if (TREE_CODE (valtype) == COMPLEX_TYPE
19383 && targetm.calls.split_complex_arg)
19384 return rs6000_complex_function_value (mode);
19385 else if (TREE_CODE (valtype) == VECTOR_TYPE
19386 && TARGET_ALTIVEC && TARGET_ALTIVEC_ABI
19387 && ALTIVEC_VECTOR_MODE (mode))
19388 regno = ALTIVEC_ARG_RETURN;
19389 else if (TARGET_E500_DOUBLE && TARGET_HARD_FLOAT
19390 && (mode == DFmode || mode == DCmode))
19391 return spe_build_register_parallel (mode, GP_ARG_RETURN);
19392 else
19393 regno = GP_ARG_RETURN;
19394
19395 return gen_rtx_REG (mode, regno);
19396 }
19397
19398 /* Define how to find the value returned by a library function
19399 assuming the value has mode MODE. */
19400 rtx
rs6000_libcall_value(enum machine_mode mode)19401 rs6000_libcall_value (enum machine_mode mode)
19402 {
19403 unsigned int regno;
19404
19405 if (TARGET_32BIT && TARGET_POWERPC64 && mode == DImode)
19406 {
19407 /* Long long return value need be split in -mpowerpc64, 32bit ABI. */
19408 return gen_rtx_PARALLEL (DImode,
19409 gen_rtvec (2,
19410 gen_rtx_EXPR_LIST (VOIDmode,
19411 gen_rtx_REG (SImode, GP_ARG_RETURN),
19412 const0_rtx),
19413 gen_rtx_EXPR_LIST (VOIDmode,
19414 gen_rtx_REG (SImode,
19415 GP_ARG_RETURN + 1),
19416 GEN_INT (4))));
19417 }
19418
19419 if (DECIMAL_FLOAT_MODE_P (mode))
19420 regno = GP_ARG_RETURN;
19421 else if (SCALAR_FLOAT_MODE_P (mode)
19422 && TARGET_HARD_FLOAT && TARGET_FPRS)
19423 regno = FP_ARG_RETURN;
19424 else if (ALTIVEC_VECTOR_MODE (mode)
19425 && TARGET_ALTIVEC && TARGET_ALTIVEC_ABI)
19426 regno = ALTIVEC_ARG_RETURN;
19427 else if (COMPLEX_MODE_P (mode) && targetm.calls.split_complex_arg)
19428 return rs6000_complex_function_value (mode);
19429 else if (TARGET_E500_DOUBLE && TARGET_HARD_FLOAT
19430 && (mode == DFmode || mode == DCmode))
19431 return spe_build_register_parallel (mode, GP_ARG_RETURN);
19432 else
19433 regno = GP_ARG_RETURN;
19434
19435 return gen_rtx_REG (mode, regno);
19436 }
19437
19438 /* Define the offset between two registers, FROM to be eliminated and its
19439 replacement TO, at the start of a routine. */
19440 HOST_WIDE_INT
rs6000_initial_elimination_offset(int from,int to)19441 rs6000_initial_elimination_offset (int from, int to)
19442 {
19443 rs6000_stack_t *info = rs6000_stack_info ();
19444 HOST_WIDE_INT offset;
19445
19446 if (from == HARD_FRAME_POINTER_REGNUM && to == STACK_POINTER_REGNUM)
19447 offset = info->push_p ? 0 : -info->total_size;
19448 else if (from == FRAME_POINTER_REGNUM && to == STACK_POINTER_REGNUM)
19449 {
19450 offset = info->push_p ? 0 : -info->total_size;
19451 if (FRAME_GROWS_DOWNWARD)
19452 offset += info->fixed_size + info->vars_size + info->parm_size;
19453 }
19454 else if (from == FRAME_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM)
19455 offset = FRAME_GROWS_DOWNWARD
19456 ? info->fixed_size + info->vars_size + info->parm_size
19457 : 0;
19458 else if (from == ARG_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM)
19459 offset = info->total_size;
19460 else if (from == ARG_POINTER_REGNUM && to == STACK_POINTER_REGNUM)
19461 offset = info->push_p ? info->total_size : 0;
19462 else if (from == RS6000_PIC_OFFSET_TABLE_REGNUM)
19463 offset = 0;
19464 else
19465 gcc_unreachable ();
19466
19467 return offset;
19468 }
19469
19470 /* Return true if TYPE is a SPE or AltiVec opaque type. */
19471
19472 static bool
rs6000_is_opaque_type(tree type)19473 rs6000_is_opaque_type (tree type)
19474 {
19475 return (type == opaque_V2SI_type_node
19476 || type == opaque_V2SF_type_node
19477 || type == opaque_p_V2SI_type_node
19478 || type == opaque_V4SI_type_node);
19479 }
19480
19481 static rtx
rs6000_dwarf_register_span(rtx reg)19482 rs6000_dwarf_register_span (rtx reg)
19483 {
19484 unsigned regno;
19485
19486 if (TARGET_SPE
19487 && (SPE_VECTOR_MODE (GET_MODE (reg))
19488 || (TARGET_E500_DOUBLE && GET_MODE (reg) == DFmode)))
19489 ;
19490 else
19491 return NULL_RTX;
19492
19493 regno = REGNO (reg);
19494
19495 /* The duality of the SPE register size wreaks all kinds of havoc.
19496 This is a way of distinguishing r0 in 32-bits from r0 in
19497 64-bits. */
19498 return
19499 gen_rtx_PARALLEL (VOIDmode,
19500 BYTES_BIG_ENDIAN
19501 ? gen_rtvec (2,
19502 gen_rtx_REG (SImode, regno + 1200),
19503 gen_rtx_REG (SImode, regno))
19504 : gen_rtvec (2,
19505 gen_rtx_REG (SImode, regno),
19506 gen_rtx_REG (SImode, regno + 1200)));
19507 }
19508
19509 /* Map internal gcc register numbers to DWARF2 register numbers. */
19510
19511 unsigned int
rs6000_dbx_register_number(unsigned int regno)19512 rs6000_dbx_register_number (unsigned int regno)
19513 {
19514 if (regno <= 63 || write_symbols != DWARF2_DEBUG)
19515 return regno;
19516 if (regno == MQ_REGNO)
19517 return 100;
19518 if (regno == LINK_REGISTER_REGNUM)
19519 return 108;
19520 if (regno == COUNT_REGISTER_REGNUM)
19521 return 109;
19522 if (CR_REGNO_P (regno))
19523 return regno - CR0_REGNO + 86;
19524 if (regno == XER_REGNO)
19525 return 101;
19526 if (ALTIVEC_REGNO_P (regno))
19527 return regno - FIRST_ALTIVEC_REGNO + 1124;
19528 if (regno == VRSAVE_REGNO)
19529 return 356;
19530 if (regno == VSCR_REGNO)
19531 return 67;
19532 if (regno == SPE_ACC_REGNO)
19533 return 99;
19534 if (regno == SPEFSCR_REGNO)
19535 return 612;
19536 /* SPE high reg number. We get these values of regno from
19537 rs6000_dwarf_register_span. */
19538 gcc_assert (regno >= 1200 && regno < 1232);
19539 return regno;
19540 }
19541
19542 /* target hook eh_return_filter_mode */
19543 static enum machine_mode
rs6000_eh_return_filter_mode(void)19544 rs6000_eh_return_filter_mode (void)
19545 {
19546 return TARGET_32BIT ? SImode : word_mode;
19547 }
19548
19549 /* Target hook for scalar_mode_supported_p. */
19550 static bool
rs6000_scalar_mode_supported_p(enum machine_mode mode)19551 rs6000_scalar_mode_supported_p (enum machine_mode mode)
19552 {
19553 if (DECIMAL_FLOAT_MODE_P (mode))
19554 return true;
19555 else
19556 return default_scalar_mode_supported_p (mode);
19557 }
19558
19559 /* Target hook for vector_mode_supported_p. */
19560 static bool
rs6000_vector_mode_supported_p(enum machine_mode mode)19561 rs6000_vector_mode_supported_p (enum machine_mode mode)
19562 {
19563
19564 if (TARGET_SPE && SPE_VECTOR_MODE (mode))
19565 return true;
19566
19567 else if (TARGET_ALTIVEC && ALTIVEC_VECTOR_MODE (mode))
19568 return true;
19569
19570 else
19571 return false;
19572 }
19573
19574 /* Target hook for invalid_arg_for_unprototyped_fn. */
19575 static const char *
invalid_arg_for_unprototyped_fn(tree typelist,tree funcdecl,tree val)19576 invalid_arg_for_unprototyped_fn (tree typelist, tree funcdecl, tree val)
19577 {
19578 return (!rs6000_darwin64_abi
19579 && typelist == 0
19580 && TREE_CODE (TREE_TYPE (val)) == VECTOR_TYPE
19581 && (funcdecl == NULL_TREE
19582 || (TREE_CODE (funcdecl) == FUNCTION_DECL
19583 && DECL_BUILT_IN_CLASS (funcdecl) != BUILT_IN_MD)))
19584 ? N_("AltiVec argument passed to unprototyped function")
19585 : NULL;
19586 }
19587
19588 /* For TARGET_SECURE_PLT 32-bit PIC code we can save PIC register
19589 setup by using __stack_chk_fail_local hidden function instead of
19590 calling __stack_chk_fail directly. Otherwise it is better to call
19591 __stack_chk_fail directly. */
19592
19593 static tree
rs6000_stack_protect_fail(void)19594 rs6000_stack_protect_fail (void)
19595 {
19596 return (DEFAULT_ABI == ABI_V4 && TARGET_SECURE_PLT && flag_pic)
19597 ? default_hidden_stack_protect_fail ()
19598 : default_external_stack_protect_fail ();
19599 }
19600
19601 #include "gt-rs6000.h"
19602