1//===-- FLATInstructions.td - FLAT Instruction Definitions ----------------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8
9def FLATOffset : ComplexPattern<i64, 2, "SelectFlatOffset<false>", [], [SDNPWantRoot], -10>;
10def FLATOffsetSigned : ComplexPattern<i64, 2, "SelectFlatOffset<true>", [], [SDNPWantRoot], -10>;
11
12def GlobalSAddr : ComplexPattern<i64, 3, "SelectGlobalSAddr", [], [SDNPWantRoot], -10>;
13
14//===----------------------------------------------------------------------===//
15// FLAT classes
16//===----------------------------------------------------------------------===//
17
18class FLAT_Pseudo<string opName, dag outs, dag ins,
19                  string asmOps, list<dag> pattern=[]> :
20  InstSI<outs, ins, "", pattern>,
21  SIMCInstr<opName, SIEncodingFamily.NONE> {
22
23  let isPseudo = 1;
24  let isCodeGenOnly = 1;
25
26  let FLAT = 1;
27
28  let UseNamedOperandTable = 1;
29  let hasSideEffects = 0;
30  let SchedRW = [WriteVMEM];
31
32  string Mnemonic = opName;
33  string AsmOperands = asmOps;
34
35  bits<1> is_flat_global = 0;
36  bits<1> is_flat_scratch = 0;
37
38  bits<1> has_vdst = 1;
39
40  // We need to distinguish having saddr and enabling saddr because
41  // saddr is only valid for scratch and global instructions. Pre-gfx9
42  // these bits were reserved, so we also don't necessarily want to
43  // set these bits to the disabled value for the original flat
44  // segment instructions.
45  bits<1> has_saddr = 0;
46  bits<1> enabled_saddr = 0;
47  bits<7> saddr_value = 0;
48  bits<1> has_vaddr = 1;
49
50  bits<1> has_data = 1;
51  bits<1> has_glc  = 1;
52  bits<1> glcValue = 0;
53  bits<1> has_dlc  = 1;
54  bits<1> dlcValue = 0;
55
56  let SubtargetPredicate = !if(is_flat_global, HasFlatGlobalInsts,
57    !if(is_flat_scratch, HasFlatScratchInsts, HasFlatAddressSpace));
58
59  // TODO: M0 if it could possibly access LDS (before gfx9? only)?
60  let Uses = !if(is_flat_global, [EXEC], [EXEC, FLAT_SCR]);
61
62  // Internally, FLAT instruction are executed as both an LDS and a
63  // Buffer instruction; so, they increment both VM_CNT and LGKM_CNT
64  // and are not considered done until both have been decremented.
65  let VM_CNT = 1;
66  let LGKM_CNT = !if(!or(is_flat_global, is_flat_scratch), 0, 1);
67
68  let IsNonFlatSeg = !if(!or(is_flat_global, is_flat_scratch), 1, 0);
69}
70
71class FLAT_Real <bits<7> op, FLAT_Pseudo ps> :
72  InstSI <ps.OutOperandList, ps.InOperandList, ps.Mnemonic # ps.AsmOperands, []>,
73  Enc64 {
74
75  let isPseudo = 0;
76  let isCodeGenOnly = 0;
77
78  // copy relevant pseudo op flags
79  let SubtargetPredicate = ps.SubtargetPredicate;
80  let AsmMatchConverter  = ps.AsmMatchConverter;
81  let TSFlags = ps.TSFlags;
82  let UseNamedOperandTable = ps.UseNamedOperandTable;
83
84  // encoding fields
85  bits<8> vaddr;
86  bits<8> vdata;
87  bits<7> saddr;
88  bits<8> vdst;
89
90  bits<1> slc;
91  bits<1> glc;
92  bits<1> dlc;
93
94  // Only valid on gfx9
95  bits<1> lds = 0; // XXX - What does this actually do?
96
97  // Segment, 00=flat, 01=scratch, 10=global, 11=reserved
98  bits<2> seg = !if(ps.is_flat_global, 0b10,
99                  !if(ps.is_flat_scratch, 0b01, 0));
100
101  // Signed offset. Highest bit ignored for flat and treated as 12-bit
102  // unsigned for flat accesses.
103  bits<13> offset;
104  bits<1> nv = 0; // XXX - What does this actually do?
105
106  // We don't use tfe right now, and it was removed in gfx9.
107  bits<1> tfe = 0;
108
109  // Only valid on GFX9+
110  let Inst{12-0} = offset;
111  let Inst{13} = lds;
112  let Inst{15-14} = seg;
113
114  let Inst{16}    = !if(ps.has_glc, glc, ps.glcValue);
115  let Inst{17}    = slc;
116  let Inst{24-18} = op;
117  let Inst{31-26} = 0x37; // Encoding.
118  let Inst{39-32} = !if(ps.has_vaddr, vaddr, ?);
119  let Inst{47-40} = !if(ps.has_data, vdata, ?);
120  let Inst{54-48} = !if(ps.has_saddr, !if(ps.enabled_saddr, saddr, 0x7f), 0);
121
122  // 54-48 is reserved.
123  let Inst{55}    = nv; // nv on GFX9+, TFE before.
124  let Inst{63-56} = !if(ps.has_vdst, vdst, ?);
125}
126
127class GlobalSaddrTable <bit is_saddr, string Name = ""> {
128  bit IsSaddr = is_saddr;
129  string SaddrOp = Name;
130}
131
132// TODO: Is exec allowed for saddr? The disabled value 0x7f is the
133// same encoding value as exec_hi, so it isn't possible to use that if
134// saddr is 32-bit (which isn't handled here yet).
135class FLAT_Load_Pseudo <string opName, RegisterClass regClass,
136  bit HasTiedOutput = 0,
137  bit HasSaddr = 0, bit EnableSaddr = 0> : FLAT_Pseudo<
138  opName,
139  (outs regClass:$vdst),
140  !con(
141    !con(
142      !if(EnableSaddr,
143        (ins SReg_64:$saddr, VGPR_32:$vaddr),
144        (ins VReg_64:$vaddr)),
145        (ins flat_offset:$offset)),
146        // FIXME: Operands with default values do not work with following non-optional operands.
147        !if(HasTiedOutput, (ins GLC:$glc, SLC:$slc, DLC:$dlc, regClass:$vdst_in),
148                           (ins GLC_0:$glc, SLC_0:$slc, DLC_0:$dlc))),
149  " $vdst, $vaddr"#!if(HasSaddr, !if(EnableSaddr, ", $saddr", ", off"), "")#"$offset$glc$slc$dlc"> {
150  let has_data = 0;
151  let mayLoad = 1;
152  let has_saddr = HasSaddr;
153  let enabled_saddr = EnableSaddr;
154  let PseudoInstr = opName#!if(!and(HasSaddr, EnableSaddr), "_SADDR", "");
155  let maybeAtomic = 1;
156
157  let Constraints = !if(HasTiedOutput, "$vdst = $vdst_in", "");
158  let DisableEncoding = !if(HasTiedOutput, "$vdst_in", "");
159}
160
161class FLAT_Store_Pseudo <string opName, RegisterClass vdataClass,
162  bit HasSaddr = 0, bit EnableSaddr = 0> : FLAT_Pseudo<
163  opName,
164  (outs),
165  !con(
166    !if(EnableSaddr,
167      (ins VGPR_32:$vaddr, vdataClass:$vdata, SReg_64:$saddr),
168      (ins VReg_64:$vaddr, vdataClass:$vdata)),
169      (ins flat_offset:$offset, GLC_0:$glc, SLC_0:$slc, DLC_0:$dlc)),
170  " $vaddr, $vdata"#!if(HasSaddr, !if(EnableSaddr, ", $saddr", ", off"), "")#"$offset$glc$slc$dlc"> {
171  let mayLoad  = 0;
172  let mayStore = 1;
173  let has_vdst = 0;
174  let has_saddr = HasSaddr;
175  let enabled_saddr = EnableSaddr;
176  let PseudoInstr = opName#!if(!and(HasSaddr, EnableSaddr), "_SADDR", "");
177  let maybeAtomic = 1;
178}
179
180multiclass FLAT_Global_Load_Pseudo<string opName, RegisterClass regClass, bit HasTiedInput = 0> {
181  let is_flat_global = 1, SubtargetPredicate = HasFlatGlobalInsts in {
182    def "" : FLAT_Load_Pseudo<opName, regClass, HasTiedInput, 1>,
183      GlobalSaddrTable<0, opName>;
184    def _SADDR : FLAT_Load_Pseudo<opName, regClass, HasTiedInput, 1, 1>,
185      GlobalSaddrTable<1, opName>;
186  }
187}
188
189class FLAT_Global_Load_AddTid_Pseudo <string opName, RegisterClass regClass,
190  bit HasTiedOutput = 0, bit HasSignedOffset = 0> : FLAT_Pseudo<
191  opName,
192  (outs regClass:$vdst),
193  !con((ins SReg_64:$saddr, flat_offset:$offset, GLC_0:$glc, SLC_0:$slc, DLC_0:$dlc),
194    !if(HasTiedOutput, (ins regClass:$vdst_in), (ins))),
195  " $vdst, $saddr$offset$glc$slc$dlc"> {
196  let is_flat_global = 1;
197  let has_data = 0;
198  let mayLoad = 1;
199  let has_vaddr = 0;
200  let has_saddr = 1;
201  let enabled_saddr = 1;
202  let maybeAtomic = 1;
203
204  let Constraints = !if(HasTiedOutput, "$vdst = $vdst_in", "");
205  let DisableEncoding = !if(HasTiedOutput, "$vdst_in", "");
206}
207
208multiclass FLAT_Global_Store_Pseudo<string opName, RegisterClass regClass> {
209  let is_flat_global = 1, SubtargetPredicate = HasFlatGlobalInsts in {
210    def "" : FLAT_Store_Pseudo<opName, regClass, 1>,
211      GlobalSaddrTable<0, opName>;
212    def _SADDR : FLAT_Store_Pseudo<opName, regClass, 1, 1>,
213      GlobalSaddrTable<1, opName>;
214  }
215}
216
217class FLAT_Global_Store_AddTid_Pseudo <string opName, RegisterClass vdataClass,
218  bit HasSignedOffset = 0> : FLAT_Pseudo<
219  opName,
220  (outs),
221  !con(
222    (ins vdataClass:$vdata, SReg_64:$saddr),
223      (ins flat_offset:$offset, GLC:$glc, SLC:$slc, DLC:$dlc)),
224  " $vdata, $saddr$offset$glc$slc$dlc"> {
225  let is_flat_global = 1;
226  let mayLoad  = 0;
227  let mayStore = 1;
228  let has_vdst = 0;
229  let has_vaddr = 0;
230  let has_saddr = 1;
231  let enabled_saddr = 1;
232  let maybeAtomic = 1;
233}
234
235class FLAT_Scratch_Load_Pseudo <string opName, RegisterClass regClass,
236  bit EnableSaddr = 0>: FLAT_Pseudo<
237  opName,
238  (outs regClass:$vdst),
239  !if(EnableSaddr,
240      (ins SReg_32_XEXEC_HI:$saddr, flat_offset:$offset, GLC:$glc, SLC:$slc, DLC:$dlc),
241      (ins VGPR_32:$vaddr, flat_offset:$offset, GLC:$glc, SLC:$slc, DLC:$dlc)),
242  " $vdst, "#!if(EnableSaddr, "off", "$vaddr")#!if(EnableSaddr, ", $saddr", ", off")#"$offset$glc$slc$dlc"> {
243  let has_data = 0;
244  let mayLoad = 1;
245  let has_saddr = 1;
246  let enabled_saddr = EnableSaddr;
247  let has_vaddr = !if(EnableSaddr, 0, 1);
248  let PseudoInstr = opName#!if(EnableSaddr, "_SADDR", "");
249  let maybeAtomic = 1;
250}
251
252class FLAT_Scratch_Store_Pseudo <string opName, RegisterClass vdataClass, bit EnableSaddr = 0> : FLAT_Pseudo<
253  opName,
254  (outs),
255  !if(EnableSaddr,
256    (ins vdataClass:$vdata, SReg_32_XEXEC_HI:$saddr, flat_offset:$offset, GLC:$glc, SLC:$slc, DLC:$dlc),
257    (ins vdataClass:$vdata, VGPR_32:$vaddr, flat_offset:$offset, GLC:$glc, SLC:$slc, DLC:$dlc)),
258  " "#!if(EnableSaddr, "off", "$vaddr")#", $vdata, "#!if(EnableSaddr, "$saddr", "off")#"$offset$glc$slc$dlc"> {
259  let mayLoad  = 0;
260  let mayStore = 1;
261  let has_vdst = 0;
262  let has_saddr = 1;
263  let enabled_saddr = EnableSaddr;
264  let has_vaddr = !if(EnableSaddr, 0, 1);
265  let PseudoInstr = opName#!if(EnableSaddr, "_SADDR", "");
266  let maybeAtomic = 1;
267}
268
269multiclass FLAT_Scratch_Load_Pseudo<string opName, RegisterClass regClass> {
270  let is_flat_scratch = 1 in {
271    def "" : FLAT_Scratch_Load_Pseudo<opName, regClass>;
272    def _SADDR : FLAT_Scratch_Load_Pseudo<opName, regClass, 1>;
273  }
274}
275
276multiclass FLAT_Scratch_Store_Pseudo<string opName, RegisterClass regClass> {
277  let is_flat_scratch = 1 in {
278    def "" : FLAT_Scratch_Store_Pseudo<opName, regClass>;
279    def _SADDR : FLAT_Scratch_Store_Pseudo<opName, regClass, 1>;
280  }
281}
282
283class FLAT_AtomicNoRet_Pseudo<string opName, dag outs, dag ins,
284                               string asm, list<dag> pattern = []> :
285  FLAT_Pseudo<opName, outs, ins, asm, pattern> {
286    let mayLoad = 1;
287    let mayStore = 1;
288    let has_glc  = 0;
289    let glcValue = 0;
290    let has_dlc  = 0;
291    let dlcValue = 0;
292    let has_vdst = 0;
293    let maybeAtomic = 1;
294}
295
296class FLAT_AtomicRet_Pseudo<string opName, dag outs, dag ins,
297                            string asm, list<dag> pattern = []>
298  : FLAT_AtomicNoRet_Pseudo<opName, outs, ins, asm, pattern> {
299  let hasPostISelHook = 1;
300  let has_vdst = 1;
301  let glcValue = 1;
302  let dlcValue = 0;
303  let PseudoInstr = NAME # "_RTN";
304}
305
306multiclass FLAT_Atomic_Pseudo<
307  string opName,
308  RegisterClass vdst_rc,
309  ValueType vt,
310  SDPatternOperator atomic = null_frag,
311  ValueType data_vt = vt,
312  RegisterClass data_rc = vdst_rc,
313  bit isFP = isFloatType<data_vt>.ret> {
314  def "" : FLAT_AtomicNoRet_Pseudo <opName,
315    (outs),
316    (ins VReg_64:$vaddr, data_rc:$vdata, flat_offset:$offset, SLC_0:$slc),
317    " $vaddr, $vdata$offset$slc">,
318    GlobalSaddrTable<0, opName>,
319    AtomicNoRet <opName, 0> {
320    let PseudoInstr = NAME;
321    let FPAtomic = isFP;
322    let AddedComplexity = -1; // Prefer global atomics if available
323  }
324
325  def _RTN : FLAT_AtomicRet_Pseudo <opName,
326    (outs vdst_rc:$vdst),
327    (ins VReg_64:$vaddr, data_rc:$vdata, flat_offset:$offset, SLC_0:$slc),
328    " $vdst, $vaddr, $vdata$offset glc$slc",
329    [(set vt:$vdst,
330      (atomic (FLATOffset i64:$vaddr, i16:$offset), data_vt:$vdata))]>,
331       GlobalSaddrTable<0, opName#"_rtn">,
332       AtomicNoRet <opName, 1>{
333    let FPAtomic = isFP;
334    let AddedComplexity = -1; // Prefer global atomics if available
335  }
336}
337
338multiclass FLAT_Global_Atomic_Pseudo_NO_RTN<
339  string opName,
340  RegisterClass vdst_rc,
341  ValueType vt,
342  SDPatternOperator atomic = null_frag,
343  ValueType data_vt = vt,
344  RegisterClass data_rc = vdst_rc,
345  bit isFP = isFloatType<data_vt>.ret> {
346
347  def "" : FLAT_AtomicNoRet_Pseudo <opName,
348    (outs),
349    (ins VReg_64:$vaddr, data_rc:$vdata, flat_offset:$offset, SLC_0:$slc),
350    " $vaddr, $vdata, off$offset$slc">,
351    GlobalSaddrTable<0, opName>,
352    AtomicNoRet <opName, 0> {
353    let has_saddr = 1;
354    let PseudoInstr = NAME;
355    let FPAtomic = isFP;
356  }
357
358  def _SADDR : FLAT_AtomicNoRet_Pseudo <opName,
359    (outs),
360    (ins VGPR_32:$vaddr, data_rc:$vdata, SReg_64:$saddr, flat_offset:$offset, SLC_0:$slc),
361    " $vaddr, $vdata, $saddr$offset$slc">,
362    GlobalSaddrTable<1, opName>,
363    AtomicNoRet <opName#"_saddr", 0> {
364    let has_saddr = 1;
365    let enabled_saddr = 1;
366    let PseudoInstr = NAME#"_SADDR";
367    let FPAtomic = isFP;
368  }
369}
370
371multiclass FLAT_Global_Atomic_Pseudo_RTN<
372  string opName,
373  RegisterClass vdst_rc,
374  ValueType vt,
375  SDPatternOperator atomic = null_frag,
376  ValueType data_vt = vt,
377  RegisterClass data_rc = vdst_rc,
378  bit isFP = isFloatType<data_vt>.ret> {
379
380  def _RTN : FLAT_AtomicRet_Pseudo <opName,
381    (outs vdst_rc:$vdst),
382      (ins VReg_64:$vaddr, data_rc:$vdata, flat_offset:$offset, SLC_0:$slc),
383    " $vdst, $vaddr, $vdata, off$offset glc$slc",
384    [(set vt:$vdst,
385      (atomic (FLATOffsetSigned i64:$vaddr, i16:$offset), data_vt:$vdata))]>,
386      GlobalSaddrTable<0, opName#"_rtn">,
387      AtomicNoRet <opName, 1> {
388    let has_saddr = 1;
389    let FPAtomic = isFP;
390  }
391
392  def _SADDR_RTN : FLAT_AtomicRet_Pseudo <opName,
393    (outs vdst_rc:$vdst),
394      (ins VGPR_32:$vaddr, data_rc:$vdata, SReg_64:$saddr, flat_offset:$offset, SLC_0:$slc),
395    " $vdst, $vaddr, $vdata, $saddr$offset glc$slc">,
396    GlobalSaddrTable<1, opName#"_rtn">,
397    AtomicNoRet <opName#"_saddr", 1> {
398     let has_saddr = 1;
399     let enabled_saddr = 1;
400     let PseudoInstr = NAME#"_SADDR_RTN";
401     let FPAtomic = isFP;
402  }
403}
404
405multiclass FLAT_Global_Atomic_Pseudo<
406  string opName,
407  RegisterClass vdst_rc,
408  ValueType vt,
409  SDPatternOperator atomic_rtn = null_frag,
410  SDPatternOperator atomic_no_rtn = null_frag,
411  ValueType data_vt = vt,
412  RegisterClass data_rc = vdst_rc> {
413  let is_flat_global = 1, SubtargetPredicate = HasFlatGlobalInsts in {
414    defm "" : FLAT_Global_Atomic_Pseudo_NO_RTN<opName, vdst_rc, vt, atomic_no_rtn, data_vt, data_rc>;
415    defm "" : FLAT_Global_Atomic_Pseudo_RTN<opName, vdst_rc, vt, atomic_rtn, data_vt, data_rc>;
416  }
417}
418
419//===----------------------------------------------------------------------===//
420// Flat Instructions
421//===----------------------------------------------------------------------===//
422
423def FLAT_LOAD_UBYTE    : FLAT_Load_Pseudo <"flat_load_ubyte", VGPR_32>;
424def FLAT_LOAD_SBYTE    : FLAT_Load_Pseudo <"flat_load_sbyte", VGPR_32>;
425def FLAT_LOAD_USHORT   : FLAT_Load_Pseudo <"flat_load_ushort", VGPR_32>;
426def FLAT_LOAD_SSHORT   : FLAT_Load_Pseudo <"flat_load_sshort", VGPR_32>;
427def FLAT_LOAD_DWORD    : FLAT_Load_Pseudo <"flat_load_dword", VGPR_32>;
428def FLAT_LOAD_DWORDX2  : FLAT_Load_Pseudo <"flat_load_dwordx2", VReg_64>;
429def FLAT_LOAD_DWORDX4  : FLAT_Load_Pseudo <"flat_load_dwordx4", VReg_128>;
430def FLAT_LOAD_DWORDX3  : FLAT_Load_Pseudo <"flat_load_dwordx3", VReg_96>;
431
432def FLAT_STORE_BYTE    : FLAT_Store_Pseudo <"flat_store_byte", VGPR_32>;
433def FLAT_STORE_SHORT   : FLAT_Store_Pseudo <"flat_store_short", VGPR_32>;
434def FLAT_STORE_DWORD   : FLAT_Store_Pseudo <"flat_store_dword", VGPR_32>;
435def FLAT_STORE_DWORDX2 : FLAT_Store_Pseudo <"flat_store_dwordx2", VReg_64>;
436def FLAT_STORE_DWORDX4 : FLAT_Store_Pseudo <"flat_store_dwordx4", VReg_128>;
437def FLAT_STORE_DWORDX3 : FLAT_Store_Pseudo <"flat_store_dwordx3", VReg_96>;
438
439let SubtargetPredicate = HasD16LoadStore in {
440def FLAT_LOAD_UBYTE_D16     : FLAT_Load_Pseudo <"flat_load_ubyte_d16", VGPR_32, 1>;
441def FLAT_LOAD_UBYTE_D16_HI  : FLAT_Load_Pseudo <"flat_load_ubyte_d16_hi", VGPR_32, 1>;
442def FLAT_LOAD_SBYTE_D16     : FLAT_Load_Pseudo <"flat_load_sbyte_d16", VGPR_32, 1>;
443def FLAT_LOAD_SBYTE_D16_HI  : FLAT_Load_Pseudo <"flat_load_sbyte_d16_hi", VGPR_32, 1>;
444def FLAT_LOAD_SHORT_D16     : FLAT_Load_Pseudo <"flat_load_short_d16", VGPR_32, 1>;
445def FLAT_LOAD_SHORT_D16_HI  : FLAT_Load_Pseudo <"flat_load_short_d16_hi", VGPR_32, 1>;
446
447def FLAT_STORE_BYTE_D16_HI  : FLAT_Store_Pseudo <"flat_store_byte_d16_hi", VGPR_32>;
448def FLAT_STORE_SHORT_D16_HI : FLAT_Store_Pseudo <"flat_store_short_d16_hi", VGPR_32>;
449}
450
451defm FLAT_ATOMIC_CMPSWAP    : FLAT_Atomic_Pseudo <"flat_atomic_cmpswap",
452                                VGPR_32, i32, AMDGPUatomic_cmp_swap_flat_32,
453                                v2i32, VReg_64>;
454
455defm FLAT_ATOMIC_CMPSWAP_X2 : FLAT_Atomic_Pseudo <"flat_atomic_cmpswap_x2",
456                                VReg_64, i64, AMDGPUatomic_cmp_swap_flat_64,
457                                v2i64, VReg_128>;
458
459defm FLAT_ATOMIC_SWAP       : FLAT_Atomic_Pseudo <"flat_atomic_swap",
460                                VGPR_32, i32, atomic_swap_flat_32>;
461
462defm FLAT_ATOMIC_SWAP_X2    : FLAT_Atomic_Pseudo <"flat_atomic_swap_x2",
463                                VReg_64, i64, atomic_swap_flat_64>;
464
465defm FLAT_ATOMIC_ADD        : FLAT_Atomic_Pseudo <"flat_atomic_add",
466                                VGPR_32, i32, atomic_load_add_flat_32>;
467
468defm FLAT_ATOMIC_SUB        : FLAT_Atomic_Pseudo <"flat_atomic_sub",
469                                VGPR_32, i32, atomic_load_sub_flat_32>;
470
471defm FLAT_ATOMIC_SMIN       : FLAT_Atomic_Pseudo <"flat_atomic_smin",
472                                VGPR_32, i32, atomic_load_min_flat_32>;
473
474defm FLAT_ATOMIC_UMIN       : FLAT_Atomic_Pseudo <"flat_atomic_umin",
475                                VGPR_32, i32, atomic_load_umin_flat_32>;
476
477defm FLAT_ATOMIC_SMAX       : FLAT_Atomic_Pseudo <"flat_atomic_smax",
478                                VGPR_32, i32, atomic_load_max_flat_32>;
479
480defm FLAT_ATOMIC_UMAX       : FLAT_Atomic_Pseudo <"flat_atomic_umax",
481                                VGPR_32, i32, atomic_load_umax_flat_32>;
482
483defm FLAT_ATOMIC_AND        : FLAT_Atomic_Pseudo <"flat_atomic_and",
484                                VGPR_32, i32, atomic_load_and_flat_32>;
485
486defm FLAT_ATOMIC_OR         : FLAT_Atomic_Pseudo <"flat_atomic_or",
487                                VGPR_32, i32, atomic_load_or_flat_32>;
488
489defm FLAT_ATOMIC_XOR        : FLAT_Atomic_Pseudo <"flat_atomic_xor",
490                                VGPR_32, i32, atomic_load_xor_flat_32>;
491
492defm FLAT_ATOMIC_INC        : FLAT_Atomic_Pseudo <"flat_atomic_inc",
493                                VGPR_32, i32, atomic_inc_flat_32>;
494
495defm FLAT_ATOMIC_DEC        : FLAT_Atomic_Pseudo <"flat_atomic_dec",
496                                VGPR_32, i32, atomic_dec_flat_32>;
497
498defm FLAT_ATOMIC_ADD_X2     : FLAT_Atomic_Pseudo <"flat_atomic_add_x2",
499                                VReg_64, i64, atomic_load_add_flat_64>;
500
501defm FLAT_ATOMIC_SUB_X2     : FLAT_Atomic_Pseudo <"flat_atomic_sub_x2",
502                                VReg_64, i64, atomic_load_sub_flat_64>;
503
504defm FLAT_ATOMIC_SMIN_X2    : FLAT_Atomic_Pseudo <"flat_atomic_smin_x2",
505                                VReg_64, i64, atomic_load_min_flat_64>;
506
507defm FLAT_ATOMIC_UMIN_X2    : FLAT_Atomic_Pseudo <"flat_atomic_umin_x2",
508                                VReg_64, i64, atomic_load_umin_flat_64>;
509
510defm FLAT_ATOMIC_SMAX_X2    : FLAT_Atomic_Pseudo <"flat_atomic_smax_x2",
511                                VReg_64, i64, atomic_load_max_flat_64>;
512
513defm FLAT_ATOMIC_UMAX_X2    : FLAT_Atomic_Pseudo <"flat_atomic_umax_x2",
514                                VReg_64, i64, atomic_load_umax_flat_64>;
515
516defm FLAT_ATOMIC_AND_X2     : FLAT_Atomic_Pseudo <"flat_atomic_and_x2",
517                                VReg_64, i64, atomic_load_and_flat_64>;
518
519defm FLAT_ATOMIC_OR_X2      : FLAT_Atomic_Pseudo <"flat_atomic_or_x2",
520                                VReg_64, i64, atomic_load_or_flat_64>;
521
522defm FLAT_ATOMIC_XOR_X2     : FLAT_Atomic_Pseudo <"flat_atomic_xor_x2",
523                                VReg_64, i64, atomic_load_xor_flat_64>;
524
525defm FLAT_ATOMIC_INC_X2     : FLAT_Atomic_Pseudo <"flat_atomic_inc_x2",
526                                VReg_64, i64, atomic_inc_flat_64>;
527
528defm FLAT_ATOMIC_DEC_X2     : FLAT_Atomic_Pseudo <"flat_atomic_dec_x2",
529                                VReg_64, i64, atomic_dec_flat_64>;
530
531// GFX7-, GFX10-only flat instructions.
532let SubtargetPredicate = isGFX7GFX10 in {
533
534defm FLAT_ATOMIC_FCMPSWAP    : FLAT_Atomic_Pseudo <"flat_atomic_fcmpswap",
535                                VGPR_32, f32, null_frag, v2f32, VReg_64>;
536
537defm FLAT_ATOMIC_FCMPSWAP_X2 : FLAT_Atomic_Pseudo <"flat_atomic_fcmpswap_x2",
538                                VReg_64, f64, null_frag, v2f64, VReg_128>;
539
540defm FLAT_ATOMIC_FMIN        : FLAT_Atomic_Pseudo <"flat_atomic_fmin",
541                                VGPR_32, f32>;
542
543defm FLAT_ATOMIC_FMAX        : FLAT_Atomic_Pseudo <"flat_atomic_fmax",
544                                VGPR_32, f32>;
545
546defm FLAT_ATOMIC_FMIN_X2     : FLAT_Atomic_Pseudo <"flat_atomic_fmin_x2",
547                                VReg_64, f64>;
548
549defm FLAT_ATOMIC_FMAX_X2     : FLAT_Atomic_Pseudo <"flat_atomic_fmax_x2",
550                                VReg_64, f64>;
551
552} // End SubtargetPredicate = isGFX7GFX10
553
554defm GLOBAL_LOAD_UBYTE    : FLAT_Global_Load_Pseudo <"global_load_ubyte", VGPR_32>;
555defm GLOBAL_LOAD_SBYTE    : FLAT_Global_Load_Pseudo <"global_load_sbyte", VGPR_32>;
556defm GLOBAL_LOAD_USHORT   : FLAT_Global_Load_Pseudo <"global_load_ushort", VGPR_32>;
557defm GLOBAL_LOAD_SSHORT   : FLAT_Global_Load_Pseudo <"global_load_sshort", VGPR_32>;
558defm GLOBAL_LOAD_DWORD    : FLAT_Global_Load_Pseudo <"global_load_dword", VGPR_32>;
559defm GLOBAL_LOAD_DWORDX2  : FLAT_Global_Load_Pseudo <"global_load_dwordx2", VReg_64>;
560defm GLOBAL_LOAD_DWORDX3  : FLAT_Global_Load_Pseudo <"global_load_dwordx3", VReg_96>;
561defm GLOBAL_LOAD_DWORDX4  : FLAT_Global_Load_Pseudo <"global_load_dwordx4", VReg_128>;
562
563defm GLOBAL_LOAD_UBYTE_D16    : FLAT_Global_Load_Pseudo <"global_load_ubyte_d16", VGPR_32, 1>;
564defm GLOBAL_LOAD_UBYTE_D16_HI : FLAT_Global_Load_Pseudo <"global_load_ubyte_d16_hi", VGPR_32, 1>;
565defm GLOBAL_LOAD_SBYTE_D16    : FLAT_Global_Load_Pseudo <"global_load_sbyte_d16", VGPR_32, 1>;
566defm GLOBAL_LOAD_SBYTE_D16_HI : FLAT_Global_Load_Pseudo <"global_load_sbyte_d16_hi", VGPR_32, 1>;
567defm GLOBAL_LOAD_SHORT_D16    : FLAT_Global_Load_Pseudo <"global_load_short_d16", VGPR_32, 1>;
568defm GLOBAL_LOAD_SHORT_D16_HI : FLAT_Global_Load_Pseudo <"global_load_short_d16_hi", VGPR_32, 1>;
569let OtherPredicates = [HasGFX10_BEncoding] in
570def  GLOBAL_LOAD_DWORD_ADDTID : FLAT_Global_Load_AddTid_Pseudo <"global_load_dword_addtid", VGPR_32>;
571
572defm GLOBAL_STORE_BYTE    : FLAT_Global_Store_Pseudo <"global_store_byte", VGPR_32>;
573defm GLOBAL_STORE_SHORT   : FLAT_Global_Store_Pseudo <"global_store_short", VGPR_32>;
574defm GLOBAL_STORE_DWORD   : FLAT_Global_Store_Pseudo <"global_store_dword", VGPR_32>;
575defm GLOBAL_STORE_DWORDX2 : FLAT_Global_Store_Pseudo <"global_store_dwordx2", VReg_64>;
576defm GLOBAL_STORE_DWORDX3 : FLAT_Global_Store_Pseudo <"global_store_dwordx3", VReg_96>;
577defm GLOBAL_STORE_DWORDX4 : FLAT_Global_Store_Pseudo <"global_store_dwordx4", VReg_128>;
578let OtherPredicates = [HasGFX10_BEncoding] in
579def  GLOBAL_STORE_DWORD_ADDTID : FLAT_Global_Store_AddTid_Pseudo <"global_store_dword_addtid", VGPR_32>;
580
581defm GLOBAL_STORE_BYTE_D16_HI  : FLAT_Global_Store_Pseudo <"global_store_byte_d16_hi", VGPR_32>;
582defm GLOBAL_STORE_SHORT_D16_HI : FLAT_Global_Store_Pseudo <"global_store_short_d16_hi", VGPR_32>;
583
584let is_flat_global = 1 in {
585defm GLOBAL_ATOMIC_CMPSWAP : FLAT_Global_Atomic_Pseudo <"global_atomic_cmpswap",
586                               VGPR_32, i32, AMDGPUatomic_cmp_swap_global_32, null_frag,
587                               v2i32, VReg_64>;
588
589defm GLOBAL_ATOMIC_CMPSWAP_X2 : FLAT_Global_Atomic_Pseudo <"global_atomic_cmpswap_x2",
590                                  VReg_64, i64, AMDGPUatomic_cmp_swap_global_64,
591                                  null_frag,
592                                  v2i64, VReg_128>;
593
594defm GLOBAL_ATOMIC_SWAP : FLAT_Global_Atomic_Pseudo <"global_atomic_swap",
595                             VGPR_32, i32, atomic_swap_global_32>;
596
597defm GLOBAL_ATOMIC_SWAP_X2 : FLAT_Global_Atomic_Pseudo <"global_atomic_swap_x2",
598                                VReg_64, i64, atomic_swap_global_64>;
599
600defm GLOBAL_ATOMIC_ADD : FLAT_Global_Atomic_Pseudo <"global_atomic_add",
601                           VGPR_32, i32, atomic_load_add_global_32>;
602
603defm GLOBAL_ATOMIC_SUB : FLAT_Global_Atomic_Pseudo <"global_atomic_sub",
604                           VGPR_32, i32, atomic_load_sub_global_32>;
605
606defm GLOBAL_ATOMIC_SMIN : FLAT_Global_Atomic_Pseudo <"global_atomic_smin",
607                            VGPR_32, i32, atomic_load_min_global_32>;
608
609defm GLOBAL_ATOMIC_UMIN : FLAT_Global_Atomic_Pseudo <"global_atomic_umin",
610                            VGPR_32, i32, atomic_load_umin_global_32>;
611
612defm GLOBAL_ATOMIC_SMAX : FLAT_Global_Atomic_Pseudo <"global_atomic_smax",
613                            VGPR_32, i32, atomic_load_max_global_32>;
614
615defm GLOBAL_ATOMIC_UMAX : FLAT_Global_Atomic_Pseudo <"global_atomic_umax",
616                            VGPR_32, i32, atomic_load_umax_global_32>;
617
618defm GLOBAL_ATOMIC_AND : FLAT_Global_Atomic_Pseudo <"global_atomic_and",
619                           VGPR_32, i32, atomic_load_and_global_32>;
620
621defm GLOBAL_ATOMIC_OR : FLAT_Global_Atomic_Pseudo <"global_atomic_or",
622                          VGPR_32, i32, atomic_load_or_global_32>;
623
624defm GLOBAL_ATOMIC_XOR : FLAT_Global_Atomic_Pseudo <"global_atomic_xor",
625                           VGPR_32, i32, atomic_load_xor_global_32>;
626
627defm GLOBAL_ATOMIC_INC : FLAT_Global_Atomic_Pseudo <"global_atomic_inc",
628                           VGPR_32, i32, atomic_inc_global_32>;
629
630defm GLOBAL_ATOMIC_DEC : FLAT_Global_Atomic_Pseudo <"global_atomic_dec",
631                           VGPR_32, i32, atomic_dec_global_32>;
632
633defm GLOBAL_ATOMIC_ADD_X2 : FLAT_Global_Atomic_Pseudo <"global_atomic_add_x2",
634                              VReg_64, i64, atomic_load_add_global_64>;
635
636defm GLOBAL_ATOMIC_SUB_X2 : FLAT_Global_Atomic_Pseudo <"global_atomic_sub_x2",
637                              VReg_64, i64, atomic_load_sub_global_64>;
638
639defm GLOBAL_ATOMIC_SMIN_X2 : FLAT_Global_Atomic_Pseudo <"global_atomic_smin_x2",
640                               VReg_64, i64, atomic_load_min_global_64>;
641
642defm GLOBAL_ATOMIC_UMIN_X2 : FLAT_Global_Atomic_Pseudo <"global_atomic_umin_x2",
643                               VReg_64, i64, atomic_load_umin_global_64>;
644
645defm GLOBAL_ATOMIC_SMAX_X2 : FLAT_Global_Atomic_Pseudo <"global_atomic_smax_x2",
646                               VReg_64, i64, atomic_load_max_global_64>;
647
648defm GLOBAL_ATOMIC_UMAX_X2 : FLAT_Global_Atomic_Pseudo <"global_atomic_umax_x2",
649                               VReg_64, i64, atomic_load_umax_global_64>;
650
651defm GLOBAL_ATOMIC_AND_X2 : FLAT_Global_Atomic_Pseudo <"global_atomic_and_x2",
652                              VReg_64, i64, atomic_load_and_global_64>;
653
654defm GLOBAL_ATOMIC_OR_X2 : FLAT_Global_Atomic_Pseudo <"global_atomic_or_x2",
655                             VReg_64, i64, atomic_load_or_global_64>;
656
657defm GLOBAL_ATOMIC_XOR_X2 : FLAT_Global_Atomic_Pseudo <"global_atomic_xor_x2",
658                              VReg_64, i64, atomic_load_xor_global_64>;
659
660defm GLOBAL_ATOMIC_INC_X2 : FLAT_Global_Atomic_Pseudo <"global_atomic_inc_x2",
661                              VReg_64, i64, atomic_inc_global_64>;
662
663defm GLOBAL_ATOMIC_DEC_X2 : FLAT_Global_Atomic_Pseudo <"global_atomic_dec_x2",
664                              VReg_64, i64, atomic_dec_global_64>;
665
666let SubtargetPredicate = HasGFX10_BEncoding in
667defm GLOBAL_ATOMIC_CSUB : FLAT_Global_Atomic_Pseudo_RTN <"global_atomic_csub",
668                              VGPR_32, i32, int_amdgcn_global_atomic_csub>;
669} // End is_flat_global = 1
670
671
672
673let SubtargetPredicate = HasFlatScratchInsts in {
674defm SCRATCH_LOAD_UBYTE    : FLAT_Scratch_Load_Pseudo <"scratch_load_ubyte", VGPR_32>;
675defm SCRATCH_LOAD_SBYTE    : FLAT_Scratch_Load_Pseudo <"scratch_load_sbyte", VGPR_32>;
676defm SCRATCH_LOAD_USHORT   : FLAT_Scratch_Load_Pseudo <"scratch_load_ushort", VGPR_32>;
677defm SCRATCH_LOAD_SSHORT   : FLAT_Scratch_Load_Pseudo <"scratch_load_sshort", VGPR_32>;
678defm SCRATCH_LOAD_DWORD    : FLAT_Scratch_Load_Pseudo <"scratch_load_dword", VGPR_32>;
679defm SCRATCH_LOAD_DWORDX2  : FLAT_Scratch_Load_Pseudo <"scratch_load_dwordx2", VReg_64>;
680defm SCRATCH_LOAD_DWORDX3  : FLAT_Scratch_Load_Pseudo <"scratch_load_dwordx3", VReg_96>;
681defm SCRATCH_LOAD_DWORDX4  : FLAT_Scratch_Load_Pseudo <"scratch_load_dwordx4", VReg_128>;
682
683defm SCRATCH_LOAD_UBYTE_D16    : FLAT_Scratch_Load_Pseudo <"scratch_load_ubyte_d16", VGPR_32>;
684defm SCRATCH_LOAD_UBYTE_D16_HI : FLAT_Scratch_Load_Pseudo <"scratch_load_ubyte_d16_hi", VGPR_32>;
685defm SCRATCH_LOAD_SBYTE_D16    : FLAT_Scratch_Load_Pseudo <"scratch_load_sbyte_d16", VGPR_32>;
686defm SCRATCH_LOAD_SBYTE_D16_HI : FLAT_Scratch_Load_Pseudo <"scratch_load_sbyte_d16_hi", VGPR_32>;
687defm SCRATCH_LOAD_SHORT_D16    : FLAT_Scratch_Load_Pseudo <"scratch_load_short_d16", VGPR_32>;
688defm SCRATCH_LOAD_SHORT_D16_HI : FLAT_Scratch_Load_Pseudo <"scratch_load_short_d16_hi", VGPR_32>;
689
690defm SCRATCH_STORE_BYTE    : FLAT_Scratch_Store_Pseudo <"scratch_store_byte", VGPR_32>;
691defm SCRATCH_STORE_SHORT   : FLAT_Scratch_Store_Pseudo <"scratch_store_short", VGPR_32>;
692defm SCRATCH_STORE_DWORD   : FLAT_Scratch_Store_Pseudo <"scratch_store_dword", VGPR_32>;
693defm SCRATCH_STORE_DWORDX2 : FLAT_Scratch_Store_Pseudo <"scratch_store_dwordx2", VReg_64>;
694defm SCRATCH_STORE_DWORDX3 : FLAT_Scratch_Store_Pseudo <"scratch_store_dwordx3", VReg_96>;
695defm SCRATCH_STORE_DWORDX4 : FLAT_Scratch_Store_Pseudo <"scratch_store_dwordx4", VReg_128>;
696
697defm SCRATCH_STORE_BYTE_D16_HI : FLAT_Scratch_Store_Pseudo <"scratch_store_byte_d16_hi", VGPR_32>;
698defm SCRATCH_STORE_SHORT_D16_HI : FLAT_Scratch_Store_Pseudo <"scratch_store_short_d16_hi", VGPR_32>;
699
700} // End SubtargetPredicate = HasFlatScratchInsts
701
702let SubtargetPredicate = isGFX10Plus, is_flat_global = 1 in {
703  defm GLOBAL_ATOMIC_FCMPSWAP :
704    FLAT_Global_Atomic_Pseudo<"global_atomic_fcmpswap", VGPR_32, f32>;
705  defm GLOBAL_ATOMIC_FMIN :
706    FLAT_Global_Atomic_Pseudo<"global_atomic_fmin", VGPR_32, f32>;
707  defm GLOBAL_ATOMIC_FMAX :
708    FLAT_Global_Atomic_Pseudo<"global_atomic_fmax", VGPR_32, f32>;
709  defm GLOBAL_ATOMIC_FCMPSWAP_X2 :
710    FLAT_Global_Atomic_Pseudo<"global_atomic_fcmpswap_x2", VReg_64, f64>;
711  defm GLOBAL_ATOMIC_FMIN_X2 :
712    FLAT_Global_Atomic_Pseudo<"global_atomic_fmin_x2", VReg_64, f64>;
713  defm GLOBAL_ATOMIC_FMAX_X2 :
714    FLAT_Global_Atomic_Pseudo<"global_atomic_fmax_x2", VReg_64, f64>;
715} // End SubtargetPredicate = isGFX10Plus, is_flat_global = 1
716
717let SubtargetPredicate = HasAtomicFaddInsts, is_flat_global = 1 in {
718
719defm GLOBAL_ATOMIC_ADD_F32 : FLAT_Global_Atomic_Pseudo_NO_RTN <
720  "global_atomic_add_f32", VGPR_32, f32, atomic_fadd_global_noret
721>;
722defm GLOBAL_ATOMIC_PK_ADD_F16 : FLAT_Global_Atomic_Pseudo_NO_RTN <
723  "global_atomic_pk_add_f16", VGPR_32, v2f16, atomic_fadd_global_noret
724>;
725
726} // End SubtargetPredicate = HasAtomicFaddInsts
727
728//===----------------------------------------------------------------------===//
729// Flat Patterns
730//===----------------------------------------------------------------------===//
731
732// Patterns for global loads with no offset.
733class FlatLoadPat <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCNPat <
734  (vt (node (FLATOffset i64:$vaddr, i16:$offset))),
735  (inst $vaddr, $offset)
736>;
737
738class FlatLoadPat_D16 <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCNPat <
739  (node (FLATOffset (i64 VReg_64:$vaddr), i16:$offset), vt:$in),
740  (inst $vaddr, $offset, 0, 0, 0, $in)
741>;
742
743class FlatSignedLoadPat_D16 <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCNPat <
744  (node (FLATOffsetSigned (i64 VReg_64:$vaddr), i16:$offset), vt:$in),
745  (inst $vaddr, $offset, 0, 0, 0, $in)
746>;
747
748class GlobalLoadSaddrPat_D16 <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCNPat <
749  (vt (node (GlobalSAddr (i64 SReg_64:$saddr), (i32 VGPR_32:$voffset), i16:$offset), vt:$in)),
750  (inst $saddr, $voffset, $offset, 0, 0, 0, $in)
751>;
752
753class FlatLoadSignedPat <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCNPat <
754  (vt (node (FLATOffsetSigned (i64 VReg_64:$vaddr), i16:$offset))),
755  (inst $vaddr, $offset)
756>;
757
758class GlobalLoadSaddrPat <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCNPat <
759  (vt (node (GlobalSAddr (i64 SReg_64:$saddr), (i32 VGPR_32:$voffset), i16:$offset))),
760  (inst $saddr, $voffset, $offset, 0, 0, 0)
761>;
762
763class GlobalStoreSaddrPat <FLAT_Pseudo inst, SDPatternOperator node,
764                           ValueType vt> : GCNPat <
765  (node vt:$data, (GlobalSAddr (i64 SReg_64:$saddr), (i32 VGPR_32:$voffset), i16:$offset)),
766  (inst $voffset, getVregSrcForVT<vt>.ret:$data, $saddr, $offset)
767>;
768
769class GlobalAtomicStoreSaddrPat <FLAT_Pseudo inst, SDPatternOperator node,
770                                 ValueType vt> : GCNPat <
771  (node (GlobalSAddr (i64 SReg_64:$saddr), (i32 VGPR_32:$voffset), i16:$offset), vt:$data),
772  (inst $voffset, getVregSrcForVT<vt>.ret:$data, $saddr, $offset)
773>;
774
775class GlobalAtomicSaddrPat <FLAT_Pseudo inst, SDPatternOperator node,
776                            ValueType vt, ValueType data_vt = vt> : GCNPat <
777  (vt (node (GlobalSAddr (i64 SReg_64:$saddr), (i32 VGPR_32:$voffset), i16:$offset), data_vt:$data)),
778  (inst $voffset, getVregSrcForVT<data_vt>.ret:$data, $saddr, $offset)
779>;
780
781class GlobalAtomicNoRtnSaddrPat <FLAT_Pseudo inst, SDPatternOperator node,
782                                 ValueType vt> : GCNPat <
783  (node (GlobalSAddr (i64 SReg_64:$saddr), (i32 VGPR_32:$voffset), i16:$offset), vt:$data),
784  (inst $voffset, getVregSrcForVT<vt>.ret:$data, $saddr, $offset)
785>;
786
787class FlatStorePat <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCNPat <
788  (node vt:$data, (FLATOffset i64:$vaddr, i16:$offset)),
789  (inst $vaddr, getVregSrcForVT<vt>.ret:$data, $offset)
790>;
791
792class FlatStoreSignedPat <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCNPat <
793  (node vt:$data, (FLATOffsetSigned i64:$vaddr, i16:$offset)),
794  (inst $vaddr, getVregSrcForVT<vt>.ret:$data, $offset)
795>;
796
797class FlatStoreAtomicPat <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCNPat <
798  // atomic store follows atomic binop convention so the address comes
799  // first.
800  (node (FLATOffset i64:$vaddr, i16:$offset), vt:$data),
801  (inst $vaddr, getVregSrcForVT<vt>.ret:$data, $offset)
802>;
803
804class FlatStoreSignedAtomicPat <FLAT_Pseudo inst, SDPatternOperator node,
805                                ValueType vt, ValueType data_vt = vt> : GCNPat <
806  // atomic store follows atomic binop convention so the address comes
807  // first.
808  (node (FLATOffset i64:$vaddr, i16:$offset), data_vt:$data),
809  (inst $vaddr, getVregSrcForVT<data_vt>.ret:$data, $offset)
810>;
811
812class FlatAtomicPat <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt,
813                     ValueType data_vt = vt> : GCNPat <
814  (vt (node (FLATOffset i64:$vaddr, i16:$offset), data_vt:$data)),
815  (inst $vaddr, $data, $offset)
816>;
817
818class FlatAtomicPatNoRtn <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCNPat <
819  (node (FLATOffset i64:$vaddr, i16:$offset), vt:$data),
820  (inst VReg_64:$vaddr, getVregSrcForVT<vt>.ret:$data, $offset)
821>;
822
823class FlatSignedAtomicPatNoRtn <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCNPat <
824  (node (FLATOffsetSigned i64:$vaddr, i16:$offset), vt:$data),
825  (inst VReg_64:$vaddr, getVregSrcForVT<vt>.ret:$data, $offset)
826>;
827
828class FlatSignedAtomicPat <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt,
829                           ValueType data_vt = vt> : GCNPat <
830  (vt (node (FLATOffsetSigned i64:$vaddr, i16:$offset), data_vt:$data)),
831  (inst $vaddr, $data, $offset)
832>;
833
834let OtherPredicates = [HasFlatAddressSpace] in {
835
836def : FlatLoadPat <FLAT_LOAD_UBYTE, extloadi8_flat, i32>;
837def : FlatLoadPat <FLAT_LOAD_UBYTE, zextloadi8_flat, i32>;
838def : FlatLoadPat <FLAT_LOAD_SBYTE, sextloadi8_flat, i32>;
839def : FlatLoadPat <FLAT_LOAD_UBYTE, extloadi8_flat, i16>;
840def : FlatLoadPat <FLAT_LOAD_UBYTE, zextloadi8_flat, i16>;
841def : FlatLoadPat <FLAT_LOAD_SBYTE, sextloadi8_flat, i16>;
842def : FlatLoadPat <FLAT_LOAD_USHORT, extloadi16_flat, i32>;
843def : FlatLoadPat <FLAT_LOAD_USHORT, zextloadi16_flat, i32>;
844def : FlatLoadPat <FLAT_LOAD_USHORT, load_flat, i16>;
845def : FlatLoadPat <FLAT_LOAD_SSHORT, sextloadi16_flat, i32>;
846def : FlatLoadPat <FLAT_LOAD_DWORDX3, load_flat, v3i32>;
847
848def : FlatLoadPat <FLAT_LOAD_DWORD, atomic_load_32_flat, i32>;
849def : FlatLoadPat <FLAT_LOAD_DWORDX2, atomic_load_64_flat, i64>;
850
851def : FlatStorePat <FLAT_STORE_BYTE, truncstorei8_flat, i32>;
852def : FlatStorePat <FLAT_STORE_SHORT, truncstorei16_flat, i32>;
853
854foreach vt = Reg32Types.types in {
855def : FlatLoadPat <FLAT_LOAD_DWORD, load_flat, vt>;
856def : FlatStorePat <FLAT_STORE_DWORD, store_flat, vt>;
857}
858
859foreach vt = VReg_64.RegTypes in {
860def : FlatStorePat <FLAT_STORE_DWORDX2, store_flat, vt>;
861def : FlatLoadPat <FLAT_LOAD_DWORDX2, load_flat, vt>;
862}
863
864def : FlatStorePat <FLAT_STORE_DWORDX3, store_flat, v3i32>;
865
866foreach vt = VReg_128.RegTypes in {
867def : FlatLoadPat <FLAT_LOAD_DWORDX4, load_flat, vt>;
868def : FlatStorePat <FLAT_STORE_DWORDX4, store_flat, vt>;
869}
870
871def : FlatStoreAtomicPat <FLAT_STORE_DWORD, atomic_store_flat_32, i32>;
872def : FlatStoreAtomicPat <FLAT_STORE_DWORDX2, atomic_store_flat_64, i64>;
873
874def : FlatAtomicPat <FLAT_ATOMIC_ADD_RTN, atomic_load_add_global_32, i32>;
875def : FlatAtomicPat <FLAT_ATOMIC_SUB_RTN, atomic_load_sub_global_32, i32>;
876def : FlatAtomicPat <FLAT_ATOMIC_INC_RTN, atomic_inc_global_32, i32>;
877def : FlatAtomicPat <FLAT_ATOMIC_DEC_RTN, atomic_dec_global_32, i32>;
878def : FlatAtomicPat <FLAT_ATOMIC_AND_RTN, atomic_load_and_global_32, i32>;
879def : FlatAtomicPat <FLAT_ATOMIC_SMAX_RTN, atomic_load_max_global_32, i32>;
880def : FlatAtomicPat <FLAT_ATOMIC_UMAX_RTN, atomic_load_umax_global_32, i32>;
881def : FlatAtomicPat <FLAT_ATOMIC_SMIN_RTN, atomic_load_min_global_32, i32>;
882def : FlatAtomicPat <FLAT_ATOMIC_UMIN_RTN, atomic_load_umin_global_32, i32>;
883def : FlatAtomicPat <FLAT_ATOMIC_OR_RTN, atomic_load_or_global_32, i32>;
884def : FlatAtomicPat <FLAT_ATOMIC_SWAP_RTN, atomic_swap_global_32, i32>;
885def : FlatAtomicPat <FLAT_ATOMIC_CMPSWAP_RTN, AMDGPUatomic_cmp_swap_global_32, i32, v2i32>;
886def : FlatAtomicPat <FLAT_ATOMIC_XOR_RTN, atomic_load_xor_global_32, i32>;
887
888def : FlatAtomicPat <FLAT_ATOMIC_ADD_X2_RTN, atomic_load_add_global_64, i64>;
889def : FlatAtomicPat <FLAT_ATOMIC_SUB_X2_RTN, atomic_load_sub_global_64, i64>;
890def : FlatAtomicPat <FLAT_ATOMIC_INC_X2_RTN, atomic_inc_global_64, i64>;
891def : FlatAtomicPat <FLAT_ATOMIC_DEC_X2_RTN, atomic_dec_global_64, i64>;
892def : FlatAtomicPat <FLAT_ATOMIC_AND_X2_RTN, atomic_load_and_global_64, i64>;
893def : FlatAtomicPat <FLAT_ATOMIC_SMAX_X2_RTN, atomic_load_max_global_64, i64>;
894def : FlatAtomicPat <FLAT_ATOMIC_UMAX_X2_RTN, atomic_load_umax_global_64, i64>;
895def : FlatAtomicPat <FLAT_ATOMIC_SMIN_X2_RTN, atomic_load_min_global_64, i64>;
896def : FlatAtomicPat <FLAT_ATOMIC_UMIN_X2_RTN, atomic_load_umin_global_64, i64>;
897def : FlatAtomicPat <FLAT_ATOMIC_OR_X2_RTN, atomic_load_or_global_64, i64>;
898def : FlatAtomicPat <FLAT_ATOMIC_SWAP_X2_RTN, atomic_swap_global_64, i64>;
899def : FlatAtomicPat <FLAT_ATOMIC_CMPSWAP_X2_RTN, AMDGPUatomic_cmp_swap_global_64, i64, v2i64>;
900def : FlatAtomicPat <FLAT_ATOMIC_XOR_X2_RTN, atomic_load_xor_global_64, i64>;
901
902def : FlatStorePat <FLAT_STORE_BYTE, truncstorei8_flat, i16>;
903def : FlatStorePat <FLAT_STORE_SHORT, store_flat, i16>;
904
905let OtherPredicates = [D16PreservesUnusedBits] in {
906def : FlatStorePat <FLAT_STORE_SHORT_D16_HI, truncstorei16_hi16_flat, i32>;
907def : FlatStorePat <FLAT_STORE_BYTE_D16_HI, truncstorei8_hi16_flat, i32>;
908
909def : FlatLoadPat_D16 <FLAT_LOAD_UBYTE_D16_HI, az_extloadi8_d16_hi_flat, v2i16>;
910def : FlatLoadPat_D16 <FLAT_LOAD_UBYTE_D16_HI, az_extloadi8_d16_hi_flat, v2f16>;
911def : FlatLoadPat_D16 <FLAT_LOAD_SBYTE_D16_HI, sextloadi8_d16_hi_flat, v2i16>;
912def : FlatLoadPat_D16 <FLAT_LOAD_SBYTE_D16_HI, sextloadi8_d16_hi_flat, v2f16>;
913def : FlatLoadPat_D16 <FLAT_LOAD_SHORT_D16_HI, load_d16_hi_flat, v2i16>;
914def : FlatLoadPat_D16 <FLAT_LOAD_SHORT_D16_HI, load_d16_hi_flat, v2f16>;
915
916def : FlatLoadPat_D16 <FLAT_LOAD_UBYTE_D16, az_extloadi8_d16_lo_flat, v2i16>;
917def : FlatLoadPat_D16 <FLAT_LOAD_UBYTE_D16, az_extloadi8_d16_lo_flat, v2f16>;
918def : FlatLoadPat_D16 <FLAT_LOAD_SBYTE_D16, sextloadi8_d16_lo_flat, v2i16>;
919def : FlatLoadPat_D16 <FLAT_LOAD_SBYTE_D16, sextloadi8_d16_lo_flat, v2f16>;
920def : FlatLoadPat_D16 <FLAT_LOAD_SHORT_D16, load_d16_lo_flat, v2i16>;
921def : FlatLoadPat_D16 <FLAT_LOAD_SHORT_D16, load_d16_lo_flat, v2f16>;
922}
923
924} // End OtherPredicates = [HasFlatAddressSpace]
925
926
927multiclass GlobalFLATLoadPats<FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> {
928  def : FlatLoadSignedPat <inst, node, vt> {
929    let AddedComplexity = 10;
930  }
931
932  def : GlobalLoadSaddrPat<!cast<FLAT_Pseudo>(!cast<string>(inst)#"_SADDR"), node, vt> {
933    let AddedComplexity = 11;
934  }
935}
936
937multiclass GlobalFLATLoadPats_D16<FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> {
938  def : FlatSignedLoadPat_D16 <inst, node, vt> {
939    let AddedComplexity = 10;
940  }
941
942  def : GlobalLoadSaddrPat_D16<!cast<FLAT_Pseudo>(!cast<string>(inst)#"_SADDR"), node, vt> {
943    let AddedComplexity = 11;
944  }
945}
946
947multiclass GlobalFLATStorePats<FLAT_Pseudo inst, SDPatternOperator node,
948                               ValueType vt> {
949  def : FlatStoreSignedPat <inst, node, vt> {
950    let AddedComplexity = 10;
951  }
952
953  def : GlobalStoreSaddrPat<!cast<FLAT_Pseudo>(!cast<string>(inst)#"_SADDR"), node, vt> {
954    let AddedComplexity = 11;
955  }
956}
957
958// Deal with swapped operands for atomic_store vs. regular store
959multiclass GlobalFLATAtomicStorePats<FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> {
960  def : FlatStoreSignedAtomicPat <inst, node, vt> {
961    let AddedComplexity = 10;
962  }
963
964  def : GlobalAtomicStoreSaddrPat<!cast<FLAT_Pseudo>(!cast<string>(inst)#"_SADDR"), node, vt> {
965    let AddedComplexity = 11;
966  }
967}
968
969multiclass GlobalFLATAtomicPats<string nortn_inst_name, SDPatternOperator node,
970                               ValueType vt, ValueType data_vt = vt> {
971  def : FlatSignedAtomicPat <!cast<FLAT_Pseudo>(nortn_inst_name#"_RTN"), node, vt, data_vt> {
972    let AddedComplexity = 10;
973  }
974
975  def : GlobalAtomicSaddrPat<!cast<FLAT_Pseudo>(nortn_inst_name#"_SADDR_RTN"), node, vt, data_vt> {
976    let AddedComplexity = 11;
977  }
978}
979
980multiclass GlobalFLATNoRtnAtomicPats<FLAT_Pseudo inst, SDPatternOperator node,
981                                     ValueType vt> {
982  def : FlatSignedAtomicPatNoRtn <inst, node, vt> {
983    let AddedComplexity = 10;
984  }
985
986  def : GlobalAtomicNoRtnSaddrPat<!cast<FLAT_Pseudo>(!cast<string>(inst)#"_SADDR"), node, vt> {
987    let AddedComplexity = 11;
988  }
989}
990
991let OtherPredicates = [HasFlatGlobalInsts] in {
992
993defm : GlobalFLATLoadPats <GLOBAL_LOAD_UBYTE, extloadi8_global, i32>;
994defm : GlobalFLATLoadPats <GLOBAL_LOAD_UBYTE, zextloadi8_global, i32>;
995defm : GlobalFLATLoadPats <GLOBAL_LOAD_SBYTE, sextloadi8_global, i32>;
996defm : GlobalFLATLoadPats <GLOBAL_LOAD_UBYTE, extloadi8_global, i16>;
997defm : GlobalFLATLoadPats <GLOBAL_LOAD_UBYTE, zextloadi8_global, i16>;
998defm : GlobalFLATLoadPats <GLOBAL_LOAD_SBYTE, sextloadi8_global, i16>;
999defm : GlobalFLATLoadPats <GLOBAL_LOAD_USHORT, extloadi16_global, i32>;
1000defm : GlobalFLATLoadPats <GLOBAL_LOAD_USHORT, zextloadi16_global, i32>;
1001defm : GlobalFLATLoadPats <GLOBAL_LOAD_SSHORT, sextloadi16_global, i32>;
1002defm : GlobalFLATLoadPats <GLOBAL_LOAD_USHORT, load_global, i16>;
1003
1004foreach vt = Reg32Types.types in {
1005defm : GlobalFLATLoadPats <GLOBAL_LOAD_DWORD, load_global, vt>;
1006defm : GlobalFLATStorePats <GLOBAL_STORE_DWORD, store_global, vt>;
1007}
1008
1009foreach vt = VReg_64.RegTypes in {
1010defm : GlobalFLATLoadPats <GLOBAL_LOAD_DWORDX2, load_global, vt>;
1011defm : GlobalFLATStorePats <GLOBAL_STORE_DWORDX2, store_global, vt>;
1012}
1013
1014defm : GlobalFLATLoadPats <GLOBAL_LOAD_DWORDX3, load_global, v3i32>;
1015
1016foreach vt = VReg_128.RegTypes in {
1017defm : GlobalFLATLoadPats <GLOBAL_LOAD_DWORDX4, load_global, vt>;
1018defm : GlobalFLATStorePats <GLOBAL_STORE_DWORDX4, store_global, vt>;
1019}
1020
1021// There is no distinction for atomic load lowering during selection;
1022// the memory legalizer will set the cache bits and insert the
1023// appropriate waits.
1024defm : GlobalFLATLoadPats <GLOBAL_LOAD_DWORD, atomic_load_32_global, i32>;
1025defm : GlobalFLATLoadPats <GLOBAL_LOAD_DWORDX2, atomic_load_64_global, i64>;
1026
1027defm : GlobalFLATStorePats <GLOBAL_STORE_BYTE, truncstorei8_global, i32>;
1028defm : GlobalFLATStorePats <GLOBAL_STORE_BYTE, truncstorei8_global, i16>;
1029defm : GlobalFLATStorePats <GLOBAL_STORE_SHORT, truncstorei16_global, i32>;
1030defm : GlobalFLATStorePats <GLOBAL_STORE_SHORT, store_global, i16>;
1031defm : GlobalFLATStorePats <GLOBAL_STORE_DWORDX3, store_global, v3i32>;
1032
1033let OtherPredicates = [D16PreservesUnusedBits] in {
1034defm : GlobalFLATStorePats <GLOBAL_STORE_SHORT_D16_HI, truncstorei16_hi16_global, i32>;
1035defm : GlobalFLATStorePats <GLOBAL_STORE_BYTE_D16_HI, truncstorei8_hi16_global, i32>;
1036
1037defm : GlobalFLATLoadPats_D16 <GLOBAL_LOAD_UBYTE_D16_HI, az_extloadi8_d16_hi_global, v2i16>;
1038defm : GlobalFLATLoadPats_D16 <GLOBAL_LOAD_UBYTE_D16_HI, az_extloadi8_d16_hi_global, v2f16>;
1039defm : GlobalFLATLoadPats_D16 <GLOBAL_LOAD_SBYTE_D16_HI, sextloadi8_d16_hi_global, v2i16>;
1040defm : GlobalFLATLoadPats_D16 <GLOBAL_LOAD_SBYTE_D16_HI, sextloadi8_d16_hi_global, v2f16>;
1041defm : GlobalFLATLoadPats_D16 <GLOBAL_LOAD_SHORT_D16_HI, load_d16_hi_global, v2i16>;
1042defm : GlobalFLATLoadPats_D16 <GLOBAL_LOAD_SHORT_D16_HI, load_d16_hi_global, v2f16>;
1043
1044defm : GlobalFLATLoadPats_D16 <GLOBAL_LOAD_UBYTE_D16, az_extloadi8_d16_lo_global, v2i16>;
1045defm : GlobalFLATLoadPats_D16 <GLOBAL_LOAD_UBYTE_D16, az_extloadi8_d16_lo_global, v2f16>;
1046defm : GlobalFLATLoadPats_D16 <GLOBAL_LOAD_SBYTE_D16, sextloadi8_d16_lo_global, v2i16>;
1047defm : GlobalFLATLoadPats_D16 <GLOBAL_LOAD_SBYTE_D16, sextloadi8_d16_lo_global, v2f16>;
1048defm : GlobalFLATLoadPats_D16 <GLOBAL_LOAD_SHORT_D16, load_d16_lo_global, v2i16>;
1049defm : GlobalFLATLoadPats_D16 <GLOBAL_LOAD_SHORT_D16, load_d16_lo_global, v2f16>;
1050}
1051
1052defm : GlobalFLATAtomicStorePats <GLOBAL_STORE_DWORD, atomic_store_global_32, i32>;
1053defm : GlobalFLATAtomicStorePats <GLOBAL_STORE_DWORDX2, atomic_store_global_64, i64>;
1054
1055defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_ADD", atomic_load_add_global_32, i32>;
1056defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_SUB", atomic_load_sub_global_32, i32>;
1057defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_INC", atomic_inc_global_32, i32>;
1058defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_DEC", atomic_dec_global_32, i32>;
1059defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_AND", atomic_load_and_global_32, i32>;
1060defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_SMAX", atomic_load_max_global_32, i32>;
1061defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_UMAX", atomic_load_umax_global_32, i32>;
1062defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_SMIN", atomic_load_min_global_32, i32>;
1063defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_UMIN", atomic_load_umin_global_32, i32>;
1064defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_OR", atomic_load_or_global_32, i32>;
1065defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_SWAP", atomic_swap_global_32, i32>;
1066defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_CMPSWAP", AMDGPUatomic_cmp_swap_global_32, i32, v2i32>;
1067defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_XOR", atomic_load_xor_global_32, i32>;
1068defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_CSUB", int_amdgcn_global_atomic_csub, i32>;
1069
1070defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_ADD_X2", atomic_load_add_global_64, i64>;
1071defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_SUB_X2", atomic_load_sub_global_64, i64>;
1072defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_INC_X2", atomic_inc_global_64, i64>;
1073defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_DEC_X2", atomic_dec_global_64, i64>;
1074defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_AND_X2", atomic_load_and_global_64, i64>;
1075defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_SMAX_X2", atomic_load_max_global_64, i64>;
1076defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_UMAX_X2", atomic_load_umax_global_64, i64>;
1077defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_SMIN_X2", atomic_load_min_global_64, i64>;
1078defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_UMIN_X2", atomic_load_umin_global_64, i64>;
1079defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_OR_X2", atomic_load_or_global_64, i64>;
1080defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_SWAP_X2", atomic_swap_global_64, i64>;
1081defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_CMPSWAP_X2", AMDGPUatomic_cmp_swap_global_64, i64, v2i64>;
1082defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_XOR_X2", atomic_load_xor_global_64, i64>;
1083
1084defm : GlobalFLATNoRtnAtomicPats <GLOBAL_ATOMIC_ADD_F32,    atomic_fadd_global_noret, f32>;
1085defm : GlobalFLATNoRtnAtomicPats <GLOBAL_ATOMIC_PK_ADD_F16, atomic_fadd_global_noret, v2f16>;
1086
1087} // End OtherPredicates = [HasFlatGlobalInsts], AddedComplexity = 10
1088
1089
1090//===----------------------------------------------------------------------===//
1091// Target
1092//===----------------------------------------------------------------------===//
1093
1094//===----------------------------------------------------------------------===//
1095// CI
1096//===----------------------------------------------------------------------===//
1097
1098class FLAT_Real_ci <bits<7> op, FLAT_Pseudo ps> :
1099  FLAT_Real <op, ps>,
1100  SIMCInstr <ps.PseudoInstr, SIEncodingFamily.SI> {
1101  let AssemblerPredicate = isGFX7Only;
1102  let DecoderNamespace="GFX7";
1103}
1104
1105def FLAT_LOAD_UBYTE_ci         : FLAT_Real_ci <0x8,  FLAT_LOAD_UBYTE>;
1106def FLAT_LOAD_SBYTE_ci         : FLAT_Real_ci <0x9,  FLAT_LOAD_SBYTE>;
1107def FLAT_LOAD_USHORT_ci        : FLAT_Real_ci <0xa,  FLAT_LOAD_USHORT>;
1108def FLAT_LOAD_SSHORT_ci        : FLAT_Real_ci <0xb,  FLAT_LOAD_SSHORT>;
1109def FLAT_LOAD_DWORD_ci         : FLAT_Real_ci <0xc,  FLAT_LOAD_DWORD>;
1110def FLAT_LOAD_DWORDX2_ci       : FLAT_Real_ci <0xd,  FLAT_LOAD_DWORDX2>;
1111def FLAT_LOAD_DWORDX4_ci       : FLAT_Real_ci <0xe,  FLAT_LOAD_DWORDX4>;
1112def FLAT_LOAD_DWORDX3_ci       : FLAT_Real_ci <0xf,  FLAT_LOAD_DWORDX3>;
1113
1114def FLAT_STORE_BYTE_ci         : FLAT_Real_ci <0x18, FLAT_STORE_BYTE>;
1115def FLAT_STORE_SHORT_ci        : FLAT_Real_ci <0x1a, FLAT_STORE_SHORT>;
1116def FLAT_STORE_DWORD_ci        : FLAT_Real_ci <0x1c, FLAT_STORE_DWORD>;
1117def FLAT_STORE_DWORDX2_ci      : FLAT_Real_ci <0x1d, FLAT_STORE_DWORDX2>;
1118def FLAT_STORE_DWORDX4_ci      : FLAT_Real_ci <0x1e, FLAT_STORE_DWORDX4>;
1119def FLAT_STORE_DWORDX3_ci      : FLAT_Real_ci <0x1f, FLAT_STORE_DWORDX3>;
1120
1121multiclass FLAT_Real_Atomics_ci <bits<7> op, FLAT_Pseudo ps> {
1122  def _ci     : FLAT_Real_ci<op, !cast<FLAT_Pseudo>(ps.PseudoInstr)>;
1123  def _RTN_ci : FLAT_Real_ci<op, !cast<FLAT_Pseudo>(ps.PseudoInstr # "_RTN")>;
1124}
1125
1126defm FLAT_ATOMIC_SWAP          : FLAT_Real_Atomics_ci <0x30, FLAT_ATOMIC_SWAP>;
1127defm FLAT_ATOMIC_CMPSWAP       : FLAT_Real_Atomics_ci <0x31, FLAT_ATOMIC_CMPSWAP>;
1128defm FLAT_ATOMIC_ADD           : FLAT_Real_Atomics_ci <0x32, FLAT_ATOMIC_ADD>;
1129defm FLAT_ATOMIC_SUB           : FLAT_Real_Atomics_ci <0x33, FLAT_ATOMIC_SUB>;
1130defm FLAT_ATOMIC_SMIN          : FLAT_Real_Atomics_ci <0x35, FLAT_ATOMIC_SMIN>;
1131defm FLAT_ATOMIC_UMIN          : FLAT_Real_Atomics_ci <0x36, FLAT_ATOMIC_UMIN>;
1132defm FLAT_ATOMIC_SMAX          : FLAT_Real_Atomics_ci <0x37, FLAT_ATOMIC_SMAX>;
1133defm FLAT_ATOMIC_UMAX          : FLAT_Real_Atomics_ci <0x38, FLAT_ATOMIC_UMAX>;
1134defm FLAT_ATOMIC_AND           : FLAT_Real_Atomics_ci <0x39, FLAT_ATOMIC_AND>;
1135defm FLAT_ATOMIC_OR            : FLAT_Real_Atomics_ci <0x3a, FLAT_ATOMIC_OR>;
1136defm FLAT_ATOMIC_XOR           : FLAT_Real_Atomics_ci <0x3b, FLAT_ATOMIC_XOR>;
1137defm FLAT_ATOMIC_INC           : FLAT_Real_Atomics_ci <0x3c, FLAT_ATOMIC_INC>;
1138defm FLAT_ATOMIC_DEC           : FLAT_Real_Atomics_ci <0x3d, FLAT_ATOMIC_DEC>;
1139defm FLAT_ATOMIC_SWAP_X2       : FLAT_Real_Atomics_ci <0x50, FLAT_ATOMIC_SWAP_X2>;
1140defm FLAT_ATOMIC_CMPSWAP_X2    : FLAT_Real_Atomics_ci <0x51, FLAT_ATOMIC_CMPSWAP_X2>;
1141defm FLAT_ATOMIC_ADD_X2        : FLAT_Real_Atomics_ci <0x52, FLAT_ATOMIC_ADD_X2>;
1142defm FLAT_ATOMIC_SUB_X2        : FLAT_Real_Atomics_ci <0x53, FLAT_ATOMIC_SUB_X2>;
1143defm FLAT_ATOMIC_SMIN_X2       : FLAT_Real_Atomics_ci <0x55, FLAT_ATOMIC_SMIN_X2>;
1144defm FLAT_ATOMIC_UMIN_X2       : FLAT_Real_Atomics_ci <0x56, FLAT_ATOMIC_UMIN_X2>;
1145defm FLAT_ATOMIC_SMAX_X2       : FLAT_Real_Atomics_ci <0x57, FLAT_ATOMIC_SMAX_X2>;
1146defm FLAT_ATOMIC_UMAX_X2       : FLAT_Real_Atomics_ci <0x58, FLAT_ATOMIC_UMAX_X2>;
1147defm FLAT_ATOMIC_AND_X2        : FLAT_Real_Atomics_ci <0x59, FLAT_ATOMIC_AND_X2>;
1148defm FLAT_ATOMIC_OR_X2         : FLAT_Real_Atomics_ci <0x5a, FLAT_ATOMIC_OR_X2>;
1149defm FLAT_ATOMIC_XOR_X2        : FLAT_Real_Atomics_ci <0x5b, FLAT_ATOMIC_XOR_X2>;
1150defm FLAT_ATOMIC_INC_X2        : FLAT_Real_Atomics_ci <0x5c, FLAT_ATOMIC_INC_X2>;
1151defm FLAT_ATOMIC_DEC_X2        : FLAT_Real_Atomics_ci <0x5d, FLAT_ATOMIC_DEC_X2>;
1152
1153// CI Only flat instructions
1154defm FLAT_ATOMIC_FCMPSWAP      : FLAT_Real_Atomics_ci <0x3e, FLAT_ATOMIC_FCMPSWAP>;
1155defm FLAT_ATOMIC_FMIN          : FLAT_Real_Atomics_ci <0x3f, FLAT_ATOMIC_FMIN>;
1156defm FLAT_ATOMIC_FMAX          : FLAT_Real_Atomics_ci <0x40, FLAT_ATOMIC_FMAX>;
1157defm FLAT_ATOMIC_FCMPSWAP_X2   : FLAT_Real_Atomics_ci <0x5e, FLAT_ATOMIC_FCMPSWAP_X2>;
1158defm FLAT_ATOMIC_FMIN_X2       : FLAT_Real_Atomics_ci <0x5f, FLAT_ATOMIC_FMIN_X2>;
1159defm FLAT_ATOMIC_FMAX_X2       : FLAT_Real_Atomics_ci <0x60, FLAT_ATOMIC_FMAX_X2>;
1160
1161
1162//===----------------------------------------------------------------------===//
1163// VI
1164//===----------------------------------------------------------------------===//
1165
1166class FLAT_Real_vi <bits<7> op, FLAT_Pseudo ps> :
1167  FLAT_Real <op, ps>,
1168  SIMCInstr <ps.PseudoInstr, SIEncodingFamily.VI> {
1169  let AssemblerPredicate = isGFX8GFX9;
1170  let DecoderNamespace = "GFX8";
1171}
1172
1173multiclass FLAT_Real_AllAddr_vi<bits<7> op> {
1174  def _vi : FLAT_Real_vi<op, !cast<FLAT_Pseudo>(NAME)>;
1175  def _SADDR_vi : FLAT_Real_vi<op, !cast<FLAT_Pseudo>(NAME#"_SADDR")>;
1176}
1177
1178def FLAT_LOAD_UBYTE_vi         : FLAT_Real_vi <0x10, FLAT_LOAD_UBYTE>;
1179def FLAT_LOAD_SBYTE_vi         : FLAT_Real_vi <0x11, FLAT_LOAD_SBYTE>;
1180def FLAT_LOAD_USHORT_vi        : FLAT_Real_vi <0x12, FLAT_LOAD_USHORT>;
1181def FLAT_LOAD_SSHORT_vi        : FLAT_Real_vi <0x13, FLAT_LOAD_SSHORT>;
1182def FLAT_LOAD_DWORD_vi         : FLAT_Real_vi <0x14, FLAT_LOAD_DWORD>;
1183def FLAT_LOAD_DWORDX2_vi       : FLAT_Real_vi <0x15, FLAT_LOAD_DWORDX2>;
1184def FLAT_LOAD_DWORDX4_vi       : FLAT_Real_vi <0x17, FLAT_LOAD_DWORDX4>;
1185def FLAT_LOAD_DWORDX3_vi       : FLAT_Real_vi <0x16, FLAT_LOAD_DWORDX3>;
1186
1187def FLAT_STORE_BYTE_vi         : FLAT_Real_vi <0x18, FLAT_STORE_BYTE>;
1188def FLAT_STORE_BYTE_D16_HI_vi  : FLAT_Real_vi <0x19, FLAT_STORE_BYTE_D16_HI>;
1189def FLAT_STORE_SHORT_vi        : FLAT_Real_vi <0x1a, FLAT_STORE_SHORT>;
1190def FLAT_STORE_SHORT_D16_HI_vi : FLAT_Real_vi <0x1b, FLAT_STORE_SHORT_D16_HI>;
1191def FLAT_STORE_DWORD_vi        : FLAT_Real_vi <0x1c, FLAT_STORE_DWORD>;
1192def FLAT_STORE_DWORDX2_vi      : FLAT_Real_vi <0x1d, FLAT_STORE_DWORDX2>;
1193def FLAT_STORE_DWORDX4_vi      : FLAT_Real_vi <0x1f, FLAT_STORE_DWORDX4>;
1194def FLAT_STORE_DWORDX3_vi      : FLAT_Real_vi <0x1e, FLAT_STORE_DWORDX3>;
1195
1196def FLAT_LOAD_UBYTE_D16_vi    : FLAT_Real_vi <0x20, FLAT_LOAD_UBYTE_D16>;
1197def FLAT_LOAD_UBYTE_D16_HI_vi : FLAT_Real_vi <0x21, FLAT_LOAD_UBYTE_D16_HI>;
1198def FLAT_LOAD_SBYTE_D16_vi    : FLAT_Real_vi <0x22, FLAT_LOAD_SBYTE_D16>;
1199def FLAT_LOAD_SBYTE_D16_HI_vi : FLAT_Real_vi <0x23, FLAT_LOAD_SBYTE_D16_HI>;
1200def FLAT_LOAD_SHORT_D16_vi    : FLAT_Real_vi <0x24, FLAT_LOAD_SHORT_D16>;
1201def FLAT_LOAD_SHORT_D16_HI_vi : FLAT_Real_vi <0x25, FLAT_LOAD_SHORT_D16_HI>;
1202
1203multiclass FLAT_Real_Atomics_vi <bits<7> op, FLAT_Pseudo ps> {
1204  def _vi     : FLAT_Real_vi<op, !cast<FLAT_Pseudo>(ps.PseudoInstr)>;
1205  def _RTN_vi : FLAT_Real_vi<op, !cast<FLAT_Pseudo>(ps.PseudoInstr # "_RTN")>;
1206}
1207
1208multiclass FLAT_Global_Real_Atomics_vi<bits<7> op> :
1209  FLAT_Real_AllAddr_vi<op> {
1210  def _RTN_vi  : FLAT_Real_vi <op, !cast<FLAT_Pseudo>(NAME#"_RTN")>;
1211  def _SADDR_RTN_vi : FLAT_Real_vi <op, !cast<FLAT_Pseudo>(NAME#"_SADDR_RTN")>;
1212}
1213
1214
1215defm FLAT_ATOMIC_SWAP       : FLAT_Real_Atomics_vi <0x40, FLAT_ATOMIC_SWAP>;
1216defm FLAT_ATOMIC_CMPSWAP    : FLAT_Real_Atomics_vi <0x41, FLAT_ATOMIC_CMPSWAP>;
1217defm FLAT_ATOMIC_ADD        : FLAT_Real_Atomics_vi <0x42, FLAT_ATOMIC_ADD>;
1218defm FLAT_ATOMIC_SUB        : FLAT_Real_Atomics_vi <0x43, FLAT_ATOMIC_SUB>;
1219defm FLAT_ATOMIC_SMIN       : FLAT_Real_Atomics_vi <0x44, FLAT_ATOMIC_SMIN>;
1220defm FLAT_ATOMIC_UMIN       : FLAT_Real_Atomics_vi <0x45, FLAT_ATOMIC_UMIN>;
1221defm FLAT_ATOMIC_SMAX       : FLAT_Real_Atomics_vi <0x46, FLAT_ATOMIC_SMAX>;
1222defm FLAT_ATOMIC_UMAX       : FLAT_Real_Atomics_vi <0x47, FLAT_ATOMIC_UMAX>;
1223defm FLAT_ATOMIC_AND        : FLAT_Real_Atomics_vi <0x48, FLAT_ATOMIC_AND>;
1224defm FLAT_ATOMIC_OR         : FLAT_Real_Atomics_vi <0x49, FLAT_ATOMIC_OR>;
1225defm FLAT_ATOMIC_XOR        : FLAT_Real_Atomics_vi <0x4a, FLAT_ATOMIC_XOR>;
1226defm FLAT_ATOMIC_INC        : FLAT_Real_Atomics_vi <0x4b, FLAT_ATOMIC_INC>;
1227defm FLAT_ATOMIC_DEC        : FLAT_Real_Atomics_vi <0x4c, FLAT_ATOMIC_DEC>;
1228defm FLAT_ATOMIC_SWAP_X2    : FLAT_Real_Atomics_vi <0x60, FLAT_ATOMIC_SWAP_X2>;
1229defm FLAT_ATOMIC_CMPSWAP_X2 : FLAT_Real_Atomics_vi <0x61, FLAT_ATOMIC_CMPSWAP_X2>;
1230defm FLAT_ATOMIC_ADD_X2     : FLAT_Real_Atomics_vi <0x62, FLAT_ATOMIC_ADD_X2>;
1231defm FLAT_ATOMIC_SUB_X2     : FLAT_Real_Atomics_vi <0x63, FLAT_ATOMIC_SUB_X2>;
1232defm FLAT_ATOMIC_SMIN_X2    : FLAT_Real_Atomics_vi <0x64, FLAT_ATOMIC_SMIN_X2>;
1233defm FLAT_ATOMIC_UMIN_X2    : FLAT_Real_Atomics_vi <0x65, FLAT_ATOMIC_UMIN_X2>;
1234defm FLAT_ATOMIC_SMAX_X2    : FLAT_Real_Atomics_vi <0x66, FLAT_ATOMIC_SMAX_X2>;
1235defm FLAT_ATOMIC_UMAX_X2    : FLAT_Real_Atomics_vi <0x67, FLAT_ATOMIC_UMAX_X2>;
1236defm FLAT_ATOMIC_AND_X2     : FLAT_Real_Atomics_vi <0x68, FLAT_ATOMIC_AND_X2>;
1237defm FLAT_ATOMIC_OR_X2      : FLAT_Real_Atomics_vi <0x69, FLAT_ATOMIC_OR_X2>;
1238defm FLAT_ATOMIC_XOR_X2     : FLAT_Real_Atomics_vi <0x6a, FLAT_ATOMIC_XOR_X2>;
1239defm FLAT_ATOMIC_INC_X2     : FLAT_Real_Atomics_vi <0x6b, FLAT_ATOMIC_INC_X2>;
1240defm FLAT_ATOMIC_DEC_X2     : FLAT_Real_Atomics_vi <0x6c, FLAT_ATOMIC_DEC_X2>;
1241
1242defm GLOBAL_LOAD_UBYTE : FLAT_Real_AllAddr_vi <0x10>;
1243defm GLOBAL_LOAD_SBYTE : FLAT_Real_AllAddr_vi <0x11>;
1244defm GLOBAL_LOAD_USHORT : FLAT_Real_AllAddr_vi <0x12>;
1245defm GLOBAL_LOAD_SSHORT : FLAT_Real_AllAddr_vi <0x13>;
1246defm GLOBAL_LOAD_DWORD : FLAT_Real_AllAddr_vi <0x14>;
1247defm GLOBAL_LOAD_DWORDX2 : FLAT_Real_AllAddr_vi <0x15>;
1248defm GLOBAL_LOAD_DWORDX3 : FLAT_Real_AllAddr_vi <0x16>;
1249defm GLOBAL_LOAD_DWORDX4 : FLAT_Real_AllAddr_vi <0x17>;
1250
1251defm GLOBAL_LOAD_UBYTE_D16    : FLAT_Real_AllAddr_vi <0x20>;
1252defm GLOBAL_LOAD_UBYTE_D16_HI : FLAT_Real_AllAddr_vi <0x21>;
1253defm GLOBAL_LOAD_SBYTE_D16    : FLAT_Real_AllAddr_vi <0x22>;
1254defm GLOBAL_LOAD_SBYTE_D16_HI : FLAT_Real_AllAddr_vi <0x23>;
1255defm GLOBAL_LOAD_SHORT_D16    : FLAT_Real_AllAddr_vi <0x24>;
1256defm GLOBAL_LOAD_SHORT_D16_HI : FLAT_Real_AllAddr_vi <0x25>;
1257
1258defm GLOBAL_STORE_BYTE : FLAT_Real_AllAddr_vi <0x18>;
1259defm GLOBAL_STORE_BYTE_D16_HI : FLAT_Real_AllAddr_vi <0x19>;
1260defm GLOBAL_STORE_SHORT : FLAT_Real_AllAddr_vi <0x1a>;
1261defm GLOBAL_STORE_SHORT_D16_HI : FLAT_Real_AllAddr_vi <0x1b>;
1262defm GLOBAL_STORE_DWORD : FLAT_Real_AllAddr_vi <0x1c>;
1263defm GLOBAL_STORE_DWORDX2 : FLAT_Real_AllAddr_vi <0x1d>;
1264defm GLOBAL_STORE_DWORDX3 : FLAT_Real_AllAddr_vi <0x1e>;
1265defm GLOBAL_STORE_DWORDX4 : FLAT_Real_AllAddr_vi <0x1f>;
1266
1267
1268defm GLOBAL_ATOMIC_SWAP       : FLAT_Global_Real_Atomics_vi <0x40>;
1269defm GLOBAL_ATOMIC_CMPSWAP    : FLAT_Global_Real_Atomics_vi <0x41>;
1270defm GLOBAL_ATOMIC_ADD        : FLAT_Global_Real_Atomics_vi <0x42>;
1271defm GLOBAL_ATOMIC_SUB        : FLAT_Global_Real_Atomics_vi <0x43>;
1272defm GLOBAL_ATOMIC_SMIN       : FLAT_Global_Real_Atomics_vi <0x44>;
1273defm GLOBAL_ATOMIC_UMIN       : FLAT_Global_Real_Atomics_vi <0x45>;
1274defm GLOBAL_ATOMIC_SMAX       : FLAT_Global_Real_Atomics_vi <0x46>;
1275defm GLOBAL_ATOMIC_UMAX       : FLAT_Global_Real_Atomics_vi <0x47>;
1276defm GLOBAL_ATOMIC_AND        : FLAT_Global_Real_Atomics_vi <0x48>;
1277defm GLOBAL_ATOMIC_OR         : FLAT_Global_Real_Atomics_vi <0x49>;
1278defm GLOBAL_ATOMIC_XOR        : FLAT_Global_Real_Atomics_vi <0x4a>;
1279defm GLOBAL_ATOMIC_INC        : FLAT_Global_Real_Atomics_vi <0x4b>;
1280defm GLOBAL_ATOMIC_DEC        : FLAT_Global_Real_Atomics_vi <0x4c>;
1281defm GLOBAL_ATOMIC_SWAP_X2    : FLAT_Global_Real_Atomics_vi <0x60>;
1282defm GLOBAL_ATOMIC_CMPSWAP_X2 : FLAT_Global_Real_Atomics_vi <0x61>;
1283defm GLOBAL_ATOMIC_ADD_X2     : FLAT_Global_Real_Atomics_vi <0x62>;
1284defm GLOBAL_ATOMIC_SUB_X2     : FLAT_Global_Real_Atomics_vi <0x63>;
1285defm GLOBAL_ATOMIC_SMIN_X2    : FLAT_Global_Real_Atomics_vi <0x64>;
1286defm GLOBAL_ATOMIC_UMIN_X2    : FLAT_Global_Real_Atomics_vi <0x65>;
1287defm GLOBAL_ATOMIC_SMAX_X2    : FLAT_Global_Real_Atomics_vi <0x66>;
1288defm GLOBAL_ATOMIC_UMAX_X2    : FLAT_Global_Real_Atomics_vi <0x67>;
1289defm GLOBAL_ATOMIC_AND_X2     : FLAT_Global_Real_Atomics_vi <0x68>;
1290defm GLOBAL_ATOMIC_OR_X2      : FLAT_Global_Real_Atomics_vi <0x69>;
1291defm GLOBAL_ATOMIC_XOR_X2     : FLAT_Global_Real_Atomics_vi <0x6a>;
1292defm GLOBAL_ATOMIC_INC_X2     : FLAT_Global_Real_Atomics_vi <0x6b>;
1293defm GLOBAL_ATOMIC_DEC_X2     : FLAT_Global_Real_Atomics_vi <0x6c>;
1294
1295defm SCRATCH_LOAD_UBYTE         : FLAT_Real_AllAddr_vi <0x10>;
1296defm SCRATCH_LOAD_SBYTE         : FLAT_Real_AllAddr_vi <0x11>;
1297defm SCRATCH_LOAD_USHORT        : FLAT_Real_AllAddr_vi <0x12>;
1298defm SCRATCH_LOAD_SSHORT        : FLAT_Real_AllAddr_vi <0x13>;
1299defm SCRATCH_LOAD_DWORD         : FLAT_Real_AllAddr_vi <0x14>;
1300defm SCRATCH_LOAD_DWORDX2       : FLAT_Real_AllAddr_vi <0x15>;
1301defm SCRATCH_LOAD_DWORDX3       : FLAT_Real_AllAddr_vi <0x16>;
1302defm SCRATCH_LOAD_DWORDX4       : FLAT_Real_AllAddr_vi <0x17>;
1303defm SCRATCH_STORE_BYTE         : FLAT_Real_AllAddr_vi <0x18>;
1304defm SCRATCH_STORE_BYTE_D16_HI  : FLAT_Real_AllAddr_vi <0x19>;
1305defm SCRATCH_LOAD_UBYTE_D16     : FLAT_Real_AllAddr_vi <0x20>;
1306defm SCRATCH_LOAD_UBYTE_D16_HI  : FLAT_Real_AllAddr_vi <0x21>;
1307defm SCRATCH_LOAD_SBYTE_D16     : FLAT_Real_AllAddr_vi <0x22>;
1308defm SCRATCH_LOAD_SBYTE_D16_HI  : FLAT_Real_AllAddr_vi <0x23>;
1309defm SCRATCH_LOAD_SHORT_D16     : FLAT_Real_AllAddr_vi <0x24>;
1310defm SCRATCH_LOAD_SHORT_D16_HI  : FLAT_Real_AllAddr_vi <0x25>;
1311defm SCRATCH_STORE_SHORT        : FLAT_Real_AllAddr_vi <0x1a>;
1312defm SCRATCH_STORE_SHORT_D16_HI : FLAT_Real_AllAddr_vi <0x1b>;
1313defm SCRATCH_STORE_DWORD        : FLAT_Real_AllAddr_vi <0x1c>;
1314defm SCRATCH_STORE_DWORDX2      : FLAT_Real_AllAddr_vi <0x1d>;
1315defm SCRATCH_STORE_DWORDX3      : FLAT_Real_AllAddr_vi <0x1e>;
1316defm SCRATCH_STORE_DWORDX4      : FLAT_Real_AllAddr_vi <0x1f>;
1317
1318
1319//===----------------------------------------------------------------------===//
1320// GFX10.
1321//===----------------------------------------------------------------------===//
1322
1323class FLAT_Real_gfx10<bits<7> op, FLAT_Pseudo ps> :
1324    FLAT_Real<op, ps>, SIMCInstr<ps.PseudoInstr, SIEncodingFamily.GFX10> {
1325  let AssemblerPredicate = isGFX10Plus;
1326  let DecoderNamespace = "GFX10";
1327
1328  let Inst{11-0}  = offset{11-0};
1329  let Inst{12}    = !if(ps.has_dlc, dlc, ps.dlcValue);
1330  let Inst{54-48} = !if(ps.has_saddr, !if(ps.enabled_saddr, saddr, 0x7d), 0x7d);
1331  let Inst{55}    = 0;
1332}
1333
1334
1335multiclass FLAT_Real_Base_gfx10<bits<7> op> {
1336  def _gfx10 :
1337    FLAT_Real_gfx10<op, !cast<FLAT_Pseudo>(NAME)>;
1338}
1339
1340multiclass FLAT_Real_RTN_gfx10<bits<7> op> {
1341  def _RTN_gfx10 :
1342    FLAT_Real_gfx10<op, !cast<FLAT_Pseudo>(NAME#"_RTN")>;
1343}
1344
1345multiclass FLAT_Real_SADDR_gfx10<bits<7> op> {
1346  def _SADDR_gfx10 :
1347    FLAT_Real_gfx10<op, !cast<FLAT_Pseudo>(NAME#"_SADDR")>;
1348}
1349
1350multiclass FLAT_Real_SADDR_RTN_gfx10<bits<7> op> {
1351  def _SADDR_RTN_gfx10 :
1352    FLAT_Real_gfx10<op, !cast<FLAT_Pseudo>(NAME#"_SADDR_RTN")>;
1353}
1354
1355
1356multiclass FLAT_Real_AllAddr_gfx10<bits<7> op> :
1357  FLAT_Real_Base_gfx10<op>,
1358  FLAT_Real_SADDR_gfx10<op>;
1359
1360multiclass FLAT_Real_Atomics_gfx10<bits<7> op> :
1361  FLAT_Real_Base_gfx10<op>,
1362  FLAT_Real_RTN_gfx10<op>;
1363
1364multiclass FLAT_Real_GlblAtomics_gfx10<bits<7> op> :
1365  FLAT_Real_AllAddr_gfx10<op>,
1366  FLAT_Real_RTN_gfx10<op>,
1367  FLAT_Real_SADDR_RTN_gfx10<op>;
1368
1369multiclass FLAT_Real_GlblAtomics_RTN_gfx10<bits<7> op> :
1370  FLAT_Real_RTN_gfx10<op>,
1371  FLAT_Real_SADDR_RTN_gfx10<op>;
1372
1373// ENC_FLAT.
1374defm FLAT_LOAD_UBYTE            : FLAT_Real_Base_gfx10<0x008>;
1375defm FLAT_LOAD_SBYTE            : FLAT_Real_Base_gfx10<0x009>;
1376defm FLAT_LOAD_USHORT           : FLAT_Real_Base_gfx10<0x00a>;
1377defm FLAT_LOAD_SSHORT           : FLAT_Real_Base_gfx10<0x00b>;
1378defm FLAT_LOAD_DWORD            : FLAT_Real_Base_gfx10<0x00c>;
1379defm FLAT_LOAD_DWORDX2          : FLAT_Real_Base_gfx10<0x00d>;
1380defm FLAT_LOAD_DWORDX4          : FLAT_Real_Base_gfx10<0x00e>;
1381defm FLAT_LOAD_DWORDX3          : FLAT_Real_Base_gfx10<0x00f>;
1382defm FLAT_STORE_BYTE            : FLAT_Real_Base_gfx10<0x018>;
1383defm FLAT_STORE_BYTE_D16_HI     : FLAT_Real_Base_gfx10<0x019>;
1384defm FLAT_STORE_SHORT           : FLAT_Real_Base_gfx10<0x01a>;
1385defm FLAT_STORE_SHORT_D16_HI    : FLAT_Real_Base_gfx10<0x01b>;
1386defm FLAT_STORE_DWORD           : FLAT_Real_Base_gfx10<0x01c>;
1387defm FLAT_STORE_DWORDX2         : FLAT_Real_Base_gfx10<0x01d>;
1388defm FLAT_STORE_DWORDX4         : FLAT_Real_Base_gfx10<0x01e>;
1389defm FLAT_STORE_DWORDX3         : FLAT_Real_Base_gfx10<0x01f>;
1390defm FLAT_LOAD_UBYTE_D16        : FLAT_Real_Base_gfx10<0x020>;
1391defm FLAT_LOAD_UBYTE_D16_HI     : FLAT_Real_Base_gfx10<0x021>;
1392defm FLAT_LOAD_SBYTE_D16        : FLAT_Real_Base_gfx10<0x022>;
1393defm FLAT_LOAD_SBYTE_D16_HI     : FLAT_Real_Base_gfx10<0x023>;
1394defm FLAT_LOAD_SHORT_D16        : FLAT_Real_Base_gfx10<0x024>;
1395defm FLAT_LOAD_SHORT_D16_HI     : FLAT_Real_Base_gfx10<0x025>;
1396defm FLAT_ATOMIC_SWAP           : FLAT_Real_Atomics_gfx10<0x030>;
1397defm FLAT_ATOMIC_CMPSWAP        : FLAT_Real_Atomics_gfx10<0x031>;
1398defm FLAT_ATOMIC_ADD            : FLAT_Real_Atomics_gfx10<0x032>;
1399defm FLAT_ATOMIC_SUB            : FLAT_Real_Atomics_gfx10<0x033>;
1400defm FLAT_ATOMIC_SMIN           : FLAT_Real_Atomics_gfx10<0x035>;
1401defm FLAT_ATOMIC_UMIN           : FLAT_Real_Atomics_gfx10<0x036>;
1402defm FLAT_ATOMIC_SMAX           : FLAT_Real_Atomics_gfx10<0x037>;
1403defm FLAT_ATOMIC_UMAX           : FLAT_Real_Atomics_gfx10<0x038>;
1404defm FLAT_ATOMIC_AND            : FLAT_Real_Atomics_gfx10<0x039>;
1405defm FLAT_ATOMIC_OR             : FLAT_Real_Atomics_gfx10<0x03a>;
1406defm FLAT_ATOMIC_XOR            : FLAT_Real_Atomics_gfx10<0x03b>;
1407defm FLAT_ATOMIC_INC            : FLAT_Real_Atomics_gfx10<0x03c>;
1408defm FLAT_ATOMIC_DEC            : FLAT_Real_Atomics_gfx10<0x03d>;
1409defm FLAT_ATOMIC_FCMPSWAP       : FLAT_Real_Atomics_gfx10<0x03e>;
1410defm FLAT_ATOMIC_FMIN           : FLAT_Real_Atomics_gfx10<0x03f>;
1411defm FLAT_ATOMIC_FMAX           : FLAT_Real_Atomics_gfx10<0x040>;
1412defm FLAT_ATOMIC_SWAP_X2        : FLAT_Real_Atomics_gfx10<0x050>;
1413defm FLAT_ATOMIC_CMPSWAP_X2     : FLAT_Real_Atomics_gfx10<0x051>;
1414defm FLAT_ATOMIC_ADD_X2         : FLAT_Real_Atomics_gfx10<0x052>;
1415defm FLAT_ATOMIC_SUB_X2         : FLAT_Real_Atomics_gfx10<0x053>;
1416defm FLAT_ATOMIC_SMIN_X2        : FLAT_Real_Atomics_gfx10<0x055>;
1417defm FLAT_ATOMIC_UMIN_X2        : FLAT_Real_Atomics_gfx10<0x056>;
1418defm FLAT_ATOMIC_SMAX_X2        : FLAT_Real_Atomics_gfx10<0x057>;
1419defm FLAT_ATOMIC_UMAX_X2        : FLAT_Real_Atomics_gfx10<0x058>;
1420defm FLAT_ATOMIC_AND_X2         : FLAT_Real_Atomics_gfx10<0x059>;
1421defm FLAT_ATOMIC_OR_X2          : FLAT_Real_Atomics_gfx10<0x05a>;
1422defm FLAT_ATOMIC_XOR_X2         : FLAT_Real_Atomics_gfx10<0x05b>;
1423defm FLAT_ATOMIC_INC_X2         : FLAT_Real_Atomics_gfx10<0x05c>;
1424defm FLAT_ATOMIC_DEC_X2         : FLAT_Real_Atomics_gfx10<0x05d>;
1425defm FLAT_ATOMIC_FCMPSWAP_X2    : FLAT_Real_Atomics_gfx10<0x05e>;
1426defm FLAT_ATOMIC_FMIN_X2        : FLAT_Real_Atomics_gfx10<0x05f>;
1427defm FLAT_ATOMIC_FMAX_X2        : FLAT_Real_Atomics_gfx10<0x060>;
1428
1429
1430// ENC_FLAT_GLBL.
1431defm GLOBAL_LOAD_UBYTE          : FLAT_Real_AllAddr_gfx10<0x008>;
1432defm GLOBAL_LOAD_SBYTE          : FLAT_Real_AllAddr_gfx10<0x009>;
1433defm GLOBAL_LOAD_USHORT         : FLAT_Real_AllAddr_gfx10<0x00a>;
1434defm GLOBAL_LOAD_SSHORT         : FLAT_Real_AllAddr_gfx10<0x00b>;
1435defm GLOBAL_LOAD_DWORD          : FLAT_Real_AllAddr_gfx10<0x00c>;
1436defm GLOBAL_LOAD_DWORDX2        : FLAT_Real_AllAddr_gfx10<0x00d>;
1437defm GLOBAL_LOAD_DWORDX4        : FLAT_Real_AllAddr_gfx10<0x00e>;
1438defm GLOBAL_LOAD_DWORDX3        : FLAT_Real_AllAddr_gfx10<0x00f>;
1439defm GLOBAL_STORE_BYTE          : FLAT_Real_AllAddr_gfx10<0x018>;
1440defm GLOBAL_STORE_BYTE_D16_HI   : FLAT_Real_AllAddr_gfx10<0x019>;
1441defm GLOBAL_STORE_SHORT         : FLAT_Real_AllAddr_gfx10<0x01a>;
1442defm GLOBAL_STORE_SHORT_D16_HI  : FLAT_Real_AllAddr_gfx10<0x01b>;
1443defm GLOBAL_STORE_DWORD         : FLAT_Real_AllAddr_gfx10<0x01c>;
1444defm GLOBAL_STORE_DWORDX2       : FLAT_Real_AllAddr_gfx10<0x01d>;
1445defm GLOBAL_STORE_DWORDX4       : FLAT_Real_AllAddr_gfx10<0x01e>;
1446defm GLOBAL_STORE_DWORDX3       : FLAT_Real_AllAddr_gfx10<0x01f>;
1447defm GLOBAL_LOAD_UBYTE_D16      : FLAT_Real_AllAddr_gfx10<0x020>;
1448defm GLOBAL_LOAD_UBYTE_D16_HI   : FLAT_Real_AllAddr_gfx10<0x021>;
1449defm GLOBAL_LOAD_SBYTE_D16      : FLAT_Real_AllAddr_gfx10<0x022>;
1450defm GLOBAL_LOAD_SBYTE_D16_HI   : FLAT_Real_AllAddr_gfx10<0x023>;
1451defm GLOBAL_LOAD_SHORT_D16      : FLAT_Real_AllAddr_gfx10<0x024>;
1452defm GLOBAL_LOAD_SHORT_D16_HI   : FLAT_Real_AllAddr_gfx10<0x025>;
1453defm GLOBAL_ATOMIC_SWAP         : FLAT_Real_GlblAtomics_gfx10<0x030>;
1454defm GLOBAL_ATOMIC_CMPSWAP      : FLAT_Real_GlblAtomics_gfx10<0x031>;
1455defm GLOBAL_ATOMIC_ADD          : FLAT_Real_GlblAtomics_gfx10<0x032>;
1456defm GLOBAL_ATOMIC_SUB          : FLAT_Real_GlblAtomics_gfx10<0x033>;
1457defm GLOBAL_ATOMIC_CSUB         : FLAT_Real_GlblAtomics_RTN_gfx10<0x034>;
1458defm GLOBAL_ATOMIC_SMIN         : FLAT_Real_GlblAtomics_gfx10<0x035>;
1459defm GLOBAL_ATOMIC_UMIN         : FLAT_Real_GlblAtomics_gfx10<0x036>;
1460defm GLOBAL_ATOMIC_SMAX         : FLAT_Real_GlblAtomics_gfx10<0x037>;
1461defm GLOBAL_ATOMIC_UMAX         : FLAT_Real_GlblAtomics_gfx10<0x038>;
1462defm GLOBAL_ATOMIC_AND          : FLAT_Real_GlblAtomics_gfx10<0x039>;
1463defm GLOBAL_ATOMIC_OR           : FLAT_Real_GlblAtomics_gfx10<0x03a>;
1464defm GLOBAL_ATOMIC_XOR          : FLAT_Real_GlblAtomics_gfx10<0x03b>;
1465defm GLOBAL_ATOMIC_INC          : FLAT_Real_GlblAtomics_gfx10<0x03c>;
1466defm GLOBAL_ATOMIC_DEC          : FLAT_Real_GlblAtomics_gfx10<0x03d>;
1467defm GLOBAL_ATOMIC_FCMPSWAP     : FLAT_Real_GlblAtomics_gfx10<0x03e>;
1468defm GLOBAL_ATOMIC_FMIN         : FLAT_Real_GlblAtomics_gfx10<0x03f>;
1469defm GLOBAL_ATOMIC_FMAX         : FLAT_Real_GlblAtomics_gfx10<0x040>;
1470defm GLOBAL_ATOMIC_SWAP_X2      : FLAT_Real_GlblAtomics_gfx10<0x050>;
1471defm GLOBAL_ATOMIC_CMPSWAP_X2   : FLAT_Real_GlblAtomics_gfx10<0x051>;
1472defm GLOBAL_ATOMIC_ADD_X2       : FLAT_Real_GlblAtomics_gfx10<0x052>;
1473defm GLOBAL_ATOMIC_SUB_X2       : FLAT_Real_GlblAtomics_gfx10<0x053>;
1474defm GLOBAL_ATOMIC_SMIN_X2      : FLAT_Real_GlblAtomics_gfx10<0x055>;
1475defm GLOBAL_ATOMIC_UMIN_X2      : FLAT_Real_GlblAtomics_gfx10<0x056>;
1476defm GLOBAL_ATOMIC_SMAX_X2      : FLAT_Real_GlblAtomics_gfx10<0x057>;
1477defm GLOBAL_ATOMIC_UMAX_X2      : FLAT_Real_GlblAtomics_gfx10<0x058>;
1478defm GLOBAL_ATOMIC_AND_X2       : FLAT_Real_GlblAtomics_gfx10<0x059>;
1479defm GLOBAL_ATOMIC_OR_X2        : FLAT_Real_GlblAtomics_gfx10<0x05a>;
1480defm GLOBAL_ATOMIC_XOR_X2       : FLAT_Real_GlblAtomics_gfx10<0x05b>;
1481defm GLOBAL_ATOMIC_INC_X2       : FLAT_Real_GlblAtomics_gfx10<0x05c>;
1482defm GLOBAL_ATOMIC_DEC_X2       : FLAT_Real_GlblAtomics_gfx10<0x05d>;
1483defm GLOBAL_ATOMIC_FCMPSWAP_X2  : FLAT_Real_GlblAtomics_gfx10<0x05e>;
1484defm GLOBAL_ATOMIC_FMIN_X2      : FLAT_Real_GlblAtomics_gfx10<0x05f>;
1485defm GLOBAL_ATOMIC_FMAX_X2      : FLAT_Real_GlblAtomics_gfx10<0x060>;
1486defm GLOBAL_LOAD_DWORD_ADDTID   : FLAT_Real_Base_gfx10<0x016>;
1487defm GLOBAL_STORE_DWORD_ADDTID  : FLAT_Real_Base_gfx10<0x017>;
1488
1489// ENC_FLAT_SCRATCH.
1490defm SCRATCH_LOAD_UBYTE         : FLAT_Real_AllAddr_gfx10<0x008>;
1491defm SCRATCH_LOAD_SBYTE         : FLAT_Real_AllAddr_gfx10<0x009>;
1492defm SCRATCH_LOAD_USHORT        : FLAT_Real_AllAddr_gfx10<0x00a>;
1493defm SCRATCH_LOAD_SSHORT        : FLAT_Real_AllAddr_gfx10<0x00b>;
1494defm SCRATCH_LOAD_DWORD         : FLAT_Real_AllAddr_gfx10<0x00c>;
1495defm SCRATCH_LOAD_DWORDX2       : FLAT_Real_AllAddr_gfx10<0x00d>;
1496defm SCRATCH_LOAD_DWORDX4       : FLAT_Real_AllAddr_gfx10<0x00e>;
1497defm SCRATCH_LOAD_DWORDX3       : FLAT_Real_AllAddr_gfx10<0x00f>;
1498defm SCRATCH_STORE_BYTE         : FLAT_Real_AllAddr_gfx10<0x018>;
1499defm SCRATCH_STORE_BYTE_D16_HI  : FLAT_Real_AllAddr_gfx10<0x019>;
1500defm SCRATCH_STORE_SHORT        : FLAT_Real_AllAddr_gfx10<0x01a>;
1501defm SCRATCH_STORE_SHORT_D16_HI : FLAT_Real_AllAddr_gfx10<0x01b>;
1502defm SCRATCH_STORE_DWORD        : FLAT_Real_AllAddr_gfx10<0x01c>;
1503defm SCRATCH_STORE_DWORDX2      : FLAT_Real_AllAddr_gfx10<0x01d>;
1504defm SCRATCH_STORE_DWORDX4      : FLAT_Real_AllAddr_gfx10<0x01e>;
1505defm SCRATCH_STORE_DWORDX3      : FLAT_Real_AllAddr_gfx10<0x01f>;
1506defm SCRATCH_LOAD_UBYTE_D16     : FLAT_Real_AllAddr_gfx10<0x020>;
1507defm SCRATCH_LOAD_UBYTE_D16_HI  : FLAT_Real_AllAddr_gfx10<0x021>;
1508defm SCRATCH_LOAD_SBYTE_D16     : FLAT_Real_AllAddr_gfx10<0x022>;
1509defm SCRATCH_LOAD_SBYTE_D16_HI  : FLAT_Real_AllAddr_gfx10<0x023>;
1510defm SCRATCH_LOAD_SHORT_D16     : FLAT_Real_AllAddr_gfx10<0x024>;
1511defm SCRATCH_LOAD_SHORT_D16_HI  : FLAT_Real_AllAddr_gfx10<0x025>;
1512
1513let SubtargetPredicate = HasAtomicFaddInsts in {
1514
1515defm GLOBAL_ATOMIC_ADD_F32    : FLAT_Real_AllAddr_vi <0x04d>;
1516defm GLOBAL_ATOMIC_PK_ADD_F16 : FLAT_Real_AllAddr_vi <0x04e>;
1517
1518} // End SubtargetPredicate = HasAtomicFaddInsts
1519