1 //===- TargetLoweringBase.cpp - Implement the TargetLoweringBase class ----===//
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 //
9 // This implements the TargetLoweringBase class.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "llvm/ADT/BitVector.h"
14 #include "llvm/ADT/STLExtras.h"
15 #include "llvm/ADT/SmallVector.h"
16 #include "llvm/ADT/StringExtras.h"
17 #include "llvm/ADT/StringRef.h"
18 #include "llvm/ADT/Triple.h"
19 #include "llvm/ADT/Twine.h"
20 #include "llvm/Analysis/Loads.h"
21 #include "llvm/Analysis/TargetTransformInfo.h"
22 #include "llvm/CodeGen/Analysis.h"
23 #include "llvm/CodeGen/ISDOpcodes.h"
24 #include "llvm/CodeGen/MachineBasicBlock.h"
25 #include "llvm/CodeGen/MachineFrameInfo.h"
26 #include "llvm/CodeGen/MachineFunction.h"
27 #include "llvm/CodeGen/MachineInstr.h"
28 #include "llvm/CodeGen/MachineInstrBuilder.h"
29 #include "llvm/CodeGen/MachineMemOperand.h"
30 #include "llvm/CodeGen/MachineOperand.h"
31 #include "llvm/CodeGen/MachineRegisterInfo.h"
32 #include "llvm/CodeGen/RuntimeLibcalls.h"
33 #include "llvm/CodeGen/StackMaps.h"
34 #include "llvm/CodeGen/TargetLowering.h"
35 #include "llvm/CodeGen/TargetOpcodes.h"
36 #include "llvm/CodeGen/TargetRegisterInfo.h"
37 #include "llvm/CodeGen/ValueTypes.h"
38 #include "llvm/IR/Attributes.h"
39 #include "llvm/IR/CallingConv.h"
40 #include "llvm/IR/DataLayout.h"
41 #include "llvm/IR/DerivedTypes.h"
42 #include "llvm/IR/Function.h"
43 #include "llvm/IR/GlobalValue.h"
44 #include "llvm/IR/GlobalVariable.h"
45 #include "llvm/IR/IRBuilder.h"
46 #include "llvm/IR/Module.h"
47 #include "llvm/IR/Type.h"
48 #include "llvm/Support/BranchProbability.h"
49 #include "llvm/Support/Casting.h"
50 #include "llvm/Support/CommandLine.h"
51 #include "llvm/Support/Compiler.h"
52 #include "llvm/Support/ErrorHandling.h"
53 #include "llvm/Support/MachineValueType.h"
54 #include "llvm/Support/MathExtras.h"
55 #include "llvm/Target/TargetMachine.h"
56 #include "llvm/Transforms/Utils/SizeOpts.h"
57 #include <algorithm>
58 #include <cassert>
59 #include <cstddef>
60 #include <cstdint>
61 #include <cstring>
62 #include <iterator>
63 #include <string>
64 #include <tuple>
65 #include <utility>
66 
67 using namespace llvm;
68 
69 static cl::opt<bool> JumpIsExpensiveOverride(
70     "jump-is-expensive", cl::init(false),
71     cl::desc("Do not create extra branches to split comparison logic."),
72     cl::Hidden);
73 
74 static cl::opt<unsigned> MinimumJumpTableEntries
75   ("min-jump-table-entries", cl::init(4), cl::Hidden,
76    cl::desc("Set minimum number of entries to use a jump table."));
77 
78 static cl::opt<unsigned> MaximumJumpTableSize
79   ("max-jump-table-size", cl::init(UINT_MAX), cl::Hidden,
80    cl::desc("Set maximum size of jump tables."));
81 
82 /// Minimum jump table density for normal functions.
83 static cl::opt<unsigned>
84     JumpTableDensity("jump-table-density", cl::init(10), cl::Hidden,
85                      cl::desc("Minimum density for building a jump table in "
86                               "a normal function"));
87 
88 /// Minimum jump table density for -Os or -Oz functions.
89 static cl::opt<unsigned> OptsizeJumpTableDensity(
90     "optsize-jump-table-density", cl::init(40), cl::Hidden,
91     cl::desc("Minimum density for building a jump table in "
92              "an optsize function"));
93 
94 // FIXME: This option is only to test if the strict fp operation processed
95 // correctly by preventing mutating strict fp operation to normal fp operation
96 // during development. When the backend supports strict float operation, this
97 // option will be meaningless.
98 static cl::opt<bool> DisableStrictNodeMutation("disable-strictnode-mutation",
99        cl::desc("Don't mutate strict-float node to a legalize node"),
100        cl::init(false), cl::Hidden);
101 
102 static bool darwinHasSinCos(const Triple &TT) {
103   assert(TT.isOSDarwin() && "should be called with darwin triple");
104   // Don't bother with 32 bit x86.
105   if (TT.getArch() == Triple::x86)
106     return false;
107   // Macos < 10.9 has no sincos_stret.
108   if (TT.isMacOSX())
109     return !TT.isMacOSXVersionLT(10, 9) && TT.isArch64Bit();
110   // iOS < 7.0 has no sincos_stret.
111   if (TT.isiOS())
112     return !TT.isOSVersionLT(7, 0);
113   // Any other darwin such as WatchOS/TvOS is new enough.
114   return true;
115 }
116 
117 // Although this default value is arbitrary, it is not random. It is assumed
118 // that a condition that evaluates the same way by a higher percentage than this
119 // is best represented as control flow. Therefore, the default value N should be
120 // set such that the win from N% correct executions is greater than the loss
121 // from (100 - N)% mispredicted executions for the majority of intended targets.
122 static cl::opt<int> MinPercentageForPredictableBranch(
123     "min-predictable-branch", cl::init(99),
124     cl::desc("Minimum percentage (0-100) that a condition must be either true "
125              "or false to assume that the condition is predictable"),
126     cl::Hidden);
127 
128 void TargetLoweringBase::InitLibcalls(const Triple &TT) {
129 #define HANDLE_LIBCALL(code, name) \
130   setLibcallName(RTLIB::code, name);
131 #include "llvm/IR/RuntimeLibcalls.def"
132 #undef HANDLE_LIBCALL
133   // Initialize calling conventions to their default.
134   for (int LC = 0; LC < RTLIB::UNKNOWN_LIBCALL; ++LC)
135     setLibcallCallingConv((RTLIB::Libcall)LC, CallingConv::C);
136 
137   // For IEEE quad-precision libcall names, PPC uses "kf" instead of "tf".
138   if (TT.getArch() == Triple::ppc || TT.isPPC64()) {
139     setLibcallName(RTLIB::ADD_F128, "__addkf3");
140     setLibcallName(RTLIB::SUB_F128, "__subkf3");
141     setLibcallName(RTLIB::MUL_F128, "__mulkf3");
142     setLibcallName(RTLIB::DIV_F128, "__divkf3");
143     setLibcallName(RTLIB::FPEXT_F32_F128, "__extendsfkf2");
144     setLibcallName(RTLIB::FPEXT_F64_F128, "__extenddfkf2");
145     setLibcallName(RTLIB::FPROUND_F128_F32, "__trunckfsf2");
146     setLibcallName(RTLIB::FPROUND_F128_F64, "__trunckfdf2");
147     setLibcallName(RTLIB::FPTOSINT_F128_I32, "__fixkfsi");
148     setLibcallName(RTLIB::FPTOSINT_F128_I64, "__fixkfdi");
149     setLibcallName(RTLIB::FPTOUINT_F128_I32, "__fixunskfsi");
150     setLibcallName(RTLIB::FPTOUINT_F128_I64, "__fixunskfdi");
151     setLibcallName(RTLIB::SINTTOFP_I32_F128, "__floatsikf");
152     setLibcallName(RTLIB::SINTTOFP_I64_F128, "__floatdikf");
153     setLibcallName(RTLIB::UINTTOFP_I32_F128, "__floatunsikf");
154     setLibcallName(RTLIB::UINTTOFP_I64_F128, "__floatundikf");
155     setLibcallName(RTLIB::OEQ_F128, "__eqkf2");
156     setLibcallName(RTLIB::UNE_F128, "__nekf2");
157     setLibcallName(RTLIB::OGE_F128, "__gekf2");
158     setLibcallName(RTLIB::OLT_F128, "__ltkf2");
159     setLibcallName(RTLIB::OLE_F128, "__lekf2");
160     setLibcallName(RTLIB::OGT_F128, "__gtkf2");
161     setLibcallName(RTLIB::UO_F128, "__unordkf2");
162   }
163 
164   // A few names are different on particular architectures or environments.
165   if (TT.isOSDarwin()) {
166     // For f16/f32 conversions, Darwin uses the standard naming scheme, instead
167     // of the gnueabi-style __gnu_*_ieee.
168     // FIXME: What about other targets?
169     setLibcallName(RTLIB::FPEXT_F16_F32, "__extendhfsf2");
170     setLibcallName(RTLIB::FPROUND_F32_F16, "__truncsfhf2");
171 
172     // Some darwins have an optimized __bzero/bzero function.
173     switch (TT.getArch()) {
174     case Triple::x86:
175     case Triple::x86_64:
176       if (TT.isMacOSX() && !TT.isMacOSXVersionLT(10, 6))
177         setLibcallName(RTLIB::BZERO, "__bzero");
178       break;
179     case Triple::aarch64:
180     case Triple::aarch64_32:
181       setLibcallName(RTLIB::BZERO, "bzero");
182       break;
183     default:
184       break;
185     }
186 
187     if (darwinHasSinCos(TT)) {
188       setLibcallName(RTLIB::SINCOS_STRET_F32, "__sincosf_stret");
189       setLibcallName(RTLIB::SINCOS_STRET_F64, "__sincos_stret");
190       if (TT.isWatchABI()) {
191         setLibcallCallingConv(RTLIB::SINCOS_STRET_F32,
192                               CallingConv::ARM_AAPCS_VFP);
193         setLibcallCallingConv(RTLIB::SINCOS_STRET_F64,
194                               CallingConv::ARM_AAPCS_VFP);
195       }
196     }
197   } else {
198     setLibcallName(RTLIB::FPEXT_F16_F32, "__gnu_h2f_ieee");
199     setLibcallName(RTLIB::FPROUND_F32_F16, "__gnu_f2h_ieee");
200   }
201 
202   if (TT.isGNUEnvironment() || TT.isOSFuchsia() ||
203       (TT.isAndroid() && !TT.isAndroidVersionLT(9))) {
204     setLibcallName(RTLIB::SINCOS_F32, "sincosf");
205     setLibcallName(RTLIB::SINCOS_F64, "sincos");
206     setLibcallName(RTLIB::SINCOS_F80, "sincosl");
207     setLibcallName(RTLIB::SINCOS_F128, "sincosl");
208     setLibcallName(RTLIB::SINCOS_PPCF128, "sincosl");
209   }
210 
211   if (TT.isPS4CPU()) {
212     setLibcallName(RTLIB::SINCOS_F32, "sincosf");
213     setLibcallName(RTLIB::SINCOS_F64, "sincos");
214   }
215 
216   if (TT.isOSOpenBSD()) {
217     setLibcallName(RTLIB::STACKPROTECTOR_CHECK_FAIL, nullptr);
218   }
219 }
220 
221 /// getFPEXT - Return the FPEXT_*_* value for the given types, or
222 /// UNKNOWN_LIBCALL if there is none.
223 RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) {
224   if (OpVT == MVT::f16) {
225     if (RetVT == MVT::f32)
226       return FPEXT_F16_F32;
227   } else if (OpVT == MVT::f32) {
228     if (RetVT == MVT::f64)
229       return FPEXT_F32_F64;
230     if (RetVT == MVT::f128)
231       return FPEXT_F32_F128;
232     if (RetVT == MVT::ppcf128)
233       return FPEXT_F32_PPCF128;
234   } else if (OpVT == MVT::f64) {
235     if (RetVT == MVT::f128)
236       return FPEXT_F64_F128;
237     else if (RetVT == MVT::ppcf128)
238       return FPEXT_F64_PPCF128;
239   } else if (OpVT == MVT::f80) {
240     if (RetVT == MVT::f128)
241       return FPEXT_F80_F128;
242   }
243 
244   return UNKNOWN_LIBCALL;
245 }
246 
247 /// getFPROUND - Return the FPROUND_*_* value for the given types, or
248 /// UNKNOWN_LIBCALL if there is none.
249 RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) {
250   if (RetVT == MVT::f16) {
251     if (OpVT == MVT::f32)
252       return FPROUND_F32_F16;
253     if (OpVT == MVT::f64)
254       return FPROUND_F64_F16;
255     if (OpVT == MVT::f80)
256       return FPROUND_F80_F16;
257     if (OpVT == MVT::f128)
258       return FPROUND_F128_F16;
259     if (OpVT == MVT::ppcf128)
260       return FPROUND_PPCF128_F16;
261   } else if (RetVT == MVT::f32) {
262     if (OpVT == MVT::f64)
263       return FPROUND_F64_F32;
264     if (OpVT == MVT::f80)
265       return FPROUND_F80_F32;
266     if (OpVT == MVT::f128)
267       return FPROUND_F128_F32;
268     if (OpVT == MVT::ppcf128)
269       return FPROUND_PPCF128_F32;
270   } else if (RetVT == MVT::f64) {
271     if (OpVT == MVT::f80)
272       return FPROUND_F80_F64;
273     if (OpVT == MVT::f128)
274       return FPROUND_F128_F64;
275     if (OpVT == MVT::ppcf128)
276       return FPROUND_PPCF128_F64;
277   } else if (RetVT == MVT::f80) {
278     if (OpVT == MVT::f128)
279       return FPROUND_F128_F80;
280   }
281 
282   return UNKNOWN_LIBCALL;
283 }
284 
285 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or
286 /// UNKNOWN_LIBCALL if there is none.
287 RTLIB::Libcall RTLIB::getFPTOSINT(EVT OpVT, EVT RetVT) {
288   if (OpVT == MVT::f32) {
289     if (RetVT == MVT::i32)
290       return FPTOSINT_F32_I32;
291     if (RetVT == MVT::i64)
292       return FPTOSINT_F32_I64;
293     if (RetVT == MVT::i128)
294       return FPTOSINT_F32_I128;
295   } else if (OpVT == MVT::f64) {
296     if (RetVT == MVT::i32)
297       return FPTOSINT_F64_I32;
298     if (RetVT == MVT::i64)
299       return FPTOSINT_F64_I64;
300     if (RetVT == MVT::i128)
301       return FPTOSINT_F64_I128;
302   } else if (OpVT == MVT::f80) {
303     if (RetVT == MVT::i32)
304       return FPTOSINT_F80_I32;
305     if (RetVT == MVT::i64)
306       return FPTOSINT_F80_I64;
307     if (RetVT == MVT::i128)
308       return FPTOSINT_F80_I128;
309   } else if (OpVT == MVT::f128) {
310     if (RetVT == MVT::i32)
311       return FPTOSINT_F128_I32;
312     if (RetVT == MVT::i64)
313       return FPTOSINT_F128_I64;
314     if (RetVT == MVT::i128)
315       return FPTOSINT_F128_I128;
316   } else if (OpVT == MVT::ppcf128) {
317     if (RetVT == MVT::i32)
318       return FPTOSINT_PPCF128_I32;
319     if (RetVT == MVT::i64)
320       return FPTOSINT_PPCF128_I64;
321     if (RetVT == MVT::i128)
322       return FPTOSINT_PPCF128_I128;
323   }
324   return UNKNOWN_LIBCALL;
325 }
326 
327 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or
328 /// UNKNOWN_LIBCALL if there is none.
329 RTLIB::Libcall RTLIB::getFPTOUINT(EVT OpVT, EVT RetVT) {
330   if (OpVT == MVT::f32) {
331     if (RetVT == MVT::i32)
332       return FPTOUINT_F32_I32;
333     if (RetVT == MVT::i64)
334       return FPTOUINT_F32_I64;
335     if (RetVT == MVT::i128)
336       return FPTOUINT_F32_I128;
337   } else if (OpVT == MVT::f64) {
338     if (RetVT == MVT::i32)
339       return FPTOUINT_F64_I32;
340     if (RetVT == MVT::i64)
341       return FPTOUINT_F64_I64;
342     if (RetVT == MVT::i128)
343       return FPTOUINT_F64_I128;
344   } else if (OpVT == MVT::f80) {
345     if (RetVT == MVT::i32)
346       return FPTOUINT_F80_I32;
347     if (RetVT == MVT::i64)
348       return FPTOUINT_F80_I64;
349     if (RetVT == MVT::i128)
350       return FPTOUINT_F80_I128;
351   } else if (OpVT == MVT::f128) {
352     if (RetVT == MVT::i32)
353       return FPTOUINT_F128_I32;
354     if (RetVT == MVT::i64)
355       return FPTOUINT_F128_I64;
356     if (RetVT == MVT::i128)
357       return FPTOUINT_F128_I128;
358   } else if (OpVT == MVT::ppcf128) {
359     if (RetVT == MVT::i32)
360       return FPTOUINT_PPCF128_I32;
361     if (RetVT == MVT::i64)
362       return FPTOUINT_PPCF128_I64;
363     if (RetVT == MVT::i128)
364       return FPTOUINT_PPCF128_I128;
365   }
366   return UNKNOWN_LIBCALL;
367 }
368 
369 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or
370 /// UNKNOWN_LIBCALL if there is none.
371 RTLIB::Libcall RTLIB::getSINTTOFP(EVT OpVT, EVT RetVT) {
372   if (OpVT == MVT::i32) {
373     if (RetVT == MVT::f32)
374       return SINTTOFP_I32_F32;
375     if (RetVT == MVT::f64)
376       return SINTTOFP_I32_F64;
377     if (RetVT == MVT::f80)
378       return SINTTOFP_I32_F80;
379     if (RetVT == MVT::f128)
380       return SINTTOFP_I32_F128;
381     if (RetVT == MVT::ppcf128)
382       return SINTTOFP_I32_PPCF128;
383   } else if (OpVT == MVT::i64) {
384     if (RetVT == MVT::f32)
385       return SINTTOFP_I64_F32;
386     if (RetVT == MVT::f64)
387       return SINTTOFP_I64_F64;
388     if (RetVT == MVT::f80)
389       return SINTTOFP_I64_F80;
390     if (RetVT == MVT::f128)
391       return SINTTOFP_I64_F128;
392     if (RetVT == MVT::ppcf128)
393       return SINTTOFP_I64_PPCF128;
394   } else if (OpVT == MVT::i128) {
395     if (RetVT == MVT::f32)
396       return SINTTOFP_I128_F32;
397     if (RetVT == MVT::f64)
398       return SINTTOFP_I128_F64;
399     if (RetVT == MVT::f80)
400       return SINTTOFP_I128_F80;
401     if (RetVT == MVT::f128)
402       return SINTTOFP_I128_F128;
403     if (RetVT == MVT::ppcf128)
404       return SINTTOFP_I128_PPCF128;
405   }
406   return UNKNOWN_LIBCALL;
407 }
408 
409 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or
410 /// UNKNOWN_LIBCALL if there is none.
411 RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) {
412   if (OpVT == MVT::i32) {
413     if (RetVT == MVT::f32)
414       return UINTTOFP_I32_F32;
415     if (RetVT == MVT::f64)
416       return UINTTOFP_I32_F64;
417     if (RetVT == MVT::f80)
418       return UINTTOFP_I32_F80;
419     if (RetVT == MVT::f128)
420       return UINTTOFP_I32_F128;
421     if (RetVT == MVT::ppcf128)
422       return UINTTOFP_I32_PPCF128;
423   } else if (OpVT == MVT::i64) {
424     if (RetVT == MVT::f32)
425       return UINTTOFP_I64_F32;
426     if (RetVT == MVT::f64)
427       return UINTTOFP_I64_F64;
428     if (RetVT == MVT::f80)
429       return UINTTOFP_I64_F80;
430     if (RetVT == MVT::f128)
431       return UINTTOFP_I64_F128;
432     if (RetVT == MVT::ppcf128)
433       return UINTTOFP_I64_PPCF128;
434   } else if (OpVT == MVT::i128) {
435     if (RetVT == MVT::f32)
436       return UINTTOFP_I128_F32;
437     if (RetVT == MVT::f64)
438       return UINTTOFP_I128_F64;
439     if (RetVT == MVT::f80)
440       return UINTTOFP_I128_F80;
441     if (RetVT == MVT::f128)
442       return UINTTOFP_I128_F128;
443     if (RetVT == MVT::ppcf128)
444       return UINTTOFP_I128_PPCF128;
445   }
446   return UNKNOWN_LIBCALL;
447 }
448 
449 RTLIB::Libcall RTLIB::getSYNC(unsigned Opc, MVT VT) {
450 #define OP_TO_LIBCALL(Name, Enum)                                              \
451   case Name:                                                                   \
452     switch (VT.SimpleTy) {                                                     \
453     default:                                                                   \
454       return UNKNOWN_LIBCALL;                                                  \
455     case MVT::i8:                                                              \
456       return Enum##_1;                                                         \
457     case MVT::i16:                                                             \
458       return Enum##_2;                                                         \
459     case MVT::i32:                                                             \
460       return Enum##_4;                                                         \
461     case MVT::i64:                                                             \
462       return Enum##_8;                                                         \
463     case MVT::i128:                                                            \
464       return Enum##_16;                                                        \
465     }
466 
467   switch (Opc) {
468     OP_TO_LIBCALL(ISD::ATOMIC_SWAP, SYNC_LOCK_TEST_AND_SET)
469     OP_TO_LIBCALL(ISD::ATOMIC_CMP_SWAP, SYNC_VAL_COMPARE_AND_SWAP)
470     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_ADD, SYNC_FETCH_AND_ADD)
471     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_SUB, SYNC_FETCH_AND_SUB)
472     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_AND, SYNC_FETCH_AND_AND)
473     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_OR, SYNC_FETCH_AND_OR)
474     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_XOR, SYNC_FETCH_AND_XOR)
475     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_NAND, SYNC_FETCH_AND_NAND)
476     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MAX, SYNC_FETCH_AND_MAX)
477     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMAX, SYNC_FETCH_AND_UMAX)
478     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MIN, SYNC_FETCH_AND_MIN)
479     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMIN, SYNC_FETCH_AND_UMIN)
480   }
481 
482 #undef OP_TO_LIBCALL
483 
484   return UNKNOWN_LIBCALL;
485 }
486 
487 RTLIB::Libcall RTLIB::getMEMCPY_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) {
488   switch (ElementSize) {
489   case 1:
490     return MEMCPY_ELEMENT_UNORDERED_ATOMIC_1;
491   case 2:
492     return MEMCPY_ELEMENT_UNORDERED_ATOMIC_2;
493   case 4:
494     return MEMCPY_ELEMENT_UNORDERED_ATOMIC_4;
495   case 8:
496     return MEMCPY_ELEMENT_UNORDERED_ATOMIC_8;
497   case 16:
498     return MEMCPY_ELEMENT_UNORDERED_ATOMIC_16;
499   default:
500     return UNKNOWN_LIBCALL;
501   }
502 }
503 
504 RTLIB::Libcall RTLIB::getMEMMOVE_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) {
505   switch (ElementSize) {
506   case 1:
507     return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_1;
508   case 2:
509     return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_2;
510   case 4:
511     return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_4;
512   case 8:
513     return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_8;
514   case 16:
515     return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_16;
516   default:
517     return UNKNOWN_LIBCALL;
518   }
519 }
520 
521 RTLIB::Libcall RTLIB::getMEMSET_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) {
522   switch (ElementSize) {
523   case 1:
524     return MEMSET_ELEMENT_UNORDERED_ATOMIC_1;
525   case 2:
526     return MEMSET_ELEMENT_UNORDERED_ATOMIC_2;
527   case 4:
528     return MEMSET_ELEMENT_UNORDERED_ATOMIC_4;
529   case 8:
530     return MEMSET_ELEMENT_UNORDERED_ATOMIC_8;
531   case 16:
532     return MEMSET_ELEMENT_UNORDERED_ATOMIC_16;
533   default:
534     return UNKNOWN_LIBCALL;
535   }
536 }
537 
538 /// InitCmpLibcallCCs - Set default comparison libcall CC.
539 static void InitCmpLibcallCCs(ISD::CondCode *CCs) {
540   memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL);
541   CCs[RTLIB::OEQ_F32] = ISD::SETEQ;
542   CCs[RTLIB::OEQ_F64] = ISD::SETEQ;
543   CCs[RTLIB::OEQ_F128] = ISD::SETEQ;
544   CCs[RTLIB::OEQ_PPCF128] = ISD::SETEQ;
545   CCs[RTLIB::UNE_F32] = ISD::SETNE;
546   CCs[RTLIB::UNE_F64] = ISD::SETNE;
547   CCs[RTLIB::UNE_F128] = ISD::SETNE;
548   CCs[RTLIB::UNE_PPCF128] = ISD::SETNE;
549   CCs[RTLIB::OGE_F32] = ISD::SETGE;
550   CCs[RTLIB::OGE_F64] = ISD::SETGE;
551   CCs[RTLIB::OGE_F128] = ISD::SETGE;
552   CCs[RTLIB::OGE_PPCF128] = ISD::SETGE;
553   CCs[RTLIB::OLT_F32] = ISD::SETLT;
554   CCs[RTLIB::OLT_F64] = ISD::SETLT;
555   CCs[RTLIB::OLT_F128] = ISD::SETLT;
556   CCs[RTLIB::OLT_PPCF128] = ISD::SETLT;
557   CCs[RTLIB::OLE_F32] = ISD::SETLE;
558   CCs[RTLIB::OLE_F64] = ISD::SETLE;
559   CCs[RTLIB::OLE_F128] = ISD::SETLE;
560   CCs[RTLIB::OLE_PPCF128] = ISD::SETLE;
561   CCs[RTLIB::OGT_F32] = ISD::SETGT;
562   CCs[RTLIB::OGT_F64] = ISD::SETGT;
563   CCs[RTLIB::OGT_F128] = ISD::SETGT;
564   CCs[RTLIB::OGT_PPCF128] = ISD::SETGT;
565   CCs[RTLIB::UO_F32] = ISD::SETNE;
566   CCs[RTLIB::UO_F64] = ISD::SETNE;
567   CCs[RTLIB::UO_F128] = ISD::SETNE;
568   CCs[RTLIB::UO_PPCF128] = ISD::SETNE;
569 }
570 
571 /// NOTE: The TargetMachine owns TLOF.
572 TargetLoweringBase::TargetLoweringBase(const TargetMachine &tm) : TM(tm) {
573   initActions();
574 
575   // Perform these initializations only once.
576   MaxStoresPerMemset = MaxStoresPerMemcpy = MaxStoresPerMemmove =
577       MaxLoadsPerMemcmp = 8;
578   MaxGluedStoresPerMemcpy = 0;
579   MaxStoresPerMemsetOptSize = MaxStoresPerMemcpyOptSize =
580       MaxStoresPerMemmoveOptSize = MaxLoadsPerMemcmpOptSize = 4;
581   HasMultipleConditionRegisters = false;
582   HasExtractBitsInsn = false;
583   JumpIsExpensive = JumpIsExpensiveOverride;
584   PredictableSelectIsExpensive = false;
585   EnableExtLdPromotion = false;
586   StackPointerRegisterToSaveRestore = 0;
587   BooleanContents = UndefinedBooleanContent;
588   BooleanFloatContents = UndefinedBooleanContent;
589   BooleanVectorContents = UndefinedBooleanContent;
590   SchedPreferenceInfo = Sched::ILP;
591   GatherAllAliasesMaxDepth = 18;
592   IsStrictFPEnabled = DisableStrictNodeMutation;
593   // TODO: the default will be switched to 0 in the next commit, along
594   // with the Target-specific changes necessary.
595   MaxAtomicSizeInBitsSupported = 1024;
596 
597   MinCmpXchgSizeInBits = 0;
598   SupportsUnalignedAtomics = false;
599 
600   std::fill(std::begin(LibcallRoutineNames), std::end(LibcallRoutineNames), nullptr);
601 
602   InitLibcalls(TM.getTargetTriple());
603   InitCmpLibcallCCs(CmpLibcallCCs);
604 }
605 
606 void TargetLoweringBase::initActions() {
607   // All operations default to being supported.
608   memset(OpActions, 0, sizeof(OpActions));
609   memset(LoadExtActions, 0, sizeof(LoadExtActions));
610   memset(TruncStoreActions, 0, sizeof(TruncStoreActions));
611   memset(IndexedModeActions, 0, sizeof(IndexedModeActions));
612   memset(CondCodeActions, 0, sizeof(CondCodeActions));
613   std::fill(std::begin(RegClassForVT), std::end(RegClassForVT), nullptr);
614   std::fill(std::begin(TargetDAGCombineArray),
615             std::end(TargetDAGCombineArray), 0);
616 
617   for (MVT VT : MVT::fp_valuetypes()) {
618     MVT IntVT = MVT::getIntegerVT(VT.getSizeInBits().getFixedSize());
619     if (IntVT.isValid()) {
620       setOperationAction(ISD::ATOMIC_SWAP, VT, Promote);
621       AddPromotedToType(ISD::ATOMIC_SWAP, VT, IntVT);
622     }
623   }
624 
625   // Set default actions for various operations.
626   for (MVT VT : MVT::all_valuetypes()) {
627     // Default all indexed load / store to expand.
628     for (unsigned IM = (unsigned)ISD::PRE_INC;
629          IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) {
630       setIndexedLoadAction(IM, VT, Expand);
631       setIndexedStoreAction(IM, VT, Expand);
632       setIndexedMaskedLoadAction(IM, VT, Expand);
633       setIndexedMaskedStoreAction(IM, VT, Expand);
634     }
635 
636     // Most backends expect to see the node which just returns the value loaded.
637     setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Expand);
638 
639     // These operations default to expand.
640     setOperationAction(ISD::FGETSIGN, VT, Expand);
641     setOperationAction(ISD::CONCAT_VECTORS, VT, Expand);
642     setOperationAction(ISD::FMINNUM, VT, Expand);
643     setOperationAction(ISD::FMAXNUM, VT, Expand);
644     setOperationAction(ISD::FMINNUM_IEEE, VT, Expand);
645     setOperationAction(ISD::FMAXNUM_IEEE, VT, Expand);
646     setOperationAction(ISD::FMINIMUM, VT, Expand);
647     setOperationAction(ISD::FMAXIMUM, VT, Expand);
648     setOperationAction(ISD::FMAD, VT, Expand);
649     setOperationAction(ISD::SMIN, VT, Expand);
650     setOperationAction(ISD::SMAX, VT, Expand);
651     setOperationAction(ISD::UMIN, VT, Expand);
652     setOperationAction(ISD::UMAX, VT, Expand);
653     setOperationAction(ISD::ABS, VT, Expand);
654     setOperationAction(ISD::FSHL, VT, Expand);
655     setOperationAction(ISD::FSHR, VT, Expand);
656     setOperationAction(ISD::SADDSAT, VT, Expand);
657     setOperationAction(ISD::UADDSAT, VT, Expand);
658     setOperationAction(ISD::SSUBSAT, VT, Expand);
659     setOperationAction(ISD::USUBSAT, VT, Expand);
660     setOperationAction(ISD::SSHLSAT, VT, Expand);
661     setOperationAction(ISD::USHLSAT, VT, Expand);
662     setOperationAction(ISD::SMULFIX, VT, Expand);
663     setOperationAction(ISD::SMULFIXSAT, VT, Expand);
664     setOperationAction(ISD::UMULFIX, VT, Expand);
665     setOperationAction(ISD::UMULFIXSAT, VT, Expand);
666     setOperationAction(ISD::SDIVFIX, VT, Expand);
667     setOperationAction(ISD::SDIVFIXSAT, VT, Expand);
668     setOperationAction(ISD::UDIVFIX, VT, Expand);
669     setOperationAction(ISD::UDIVFIXSAT, VT, Expand);
670 
671     // Overflow operations default to expand
672     setOperationAction(ISD::SADDO, VT, Expand);
673     setOperationAction(ISD::SSUBO, VT, Expand);
674     setOperationAction(ISD::UADDO, VT, Expand);
675     setOperationAction(ISD::USUBO, VT, Expand);
676     setOperationAction(ISD::SMULO, VT, Expand);
677     setOperationAction(ISD::UMULO, VT, Expand);
678 
679     // ADDCARRY operations default to expand
680     setOperationAction(ISD::ADDCARRY, VT, Expand);
681     setOperationAction(ISD::SUBCARRY, VT, Expand);
682     setOperationAction(ISD::SETCCCARRY, VT, Expand);
683 
684     // ADDC/ADDE/SUBC/SUBE default to expand.
685     setOperationAction(ISD::ADDC, VT, Expand);
686     setOperationAction(ISD::ADDE, VT, Expand);
687     setOperationAction(ISD::SUBC, VT, Expand);
688     setOperationAction(ISD::SUBE, VT, Expand);
689 
690     // These default to Expand so they will be expanded to CTLZ/CTTZ by default.
691     setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
692     setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
693 
694     setOperationAction(ISD::BITREVERSE, VT, Expand);
695 
696     // These library functions default to expand.
697     setOperationAction(ISD::FROUND, VT, Expand);
698     setOperationAction(ISD::FROUNDEVEN, VT, Expand);
699     setOperationAction(ISD::FPOWI, VT, Expand);
700 
701     // These operations default to expand for vector types.
702     if (VT.isVector()) {
703       setOperationAction(ISD::FCOPYSIGN, VT, Expand);
704       setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
705       setOperationAction(ISD::ANY_EXTEND_VECTOR_INREG, VT, Expand);
706       setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, VT, Expand);
707       setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, VT, Expand);
708       setOperationAction(ISD::SPLAT_VECTOR, VT, Expand);
709     }
710 
711     // Constrained floating-point operations default to expand.
712 #define DAG_INSTRUCTION(NAME, NARG, ROUND_MODE, INTRINSIC, DAGN)               \
713     setOperationAction(ISD::STRICT_##DAGN, VT, Expand);
714 #include "llvm/IR/ConstrainedOps.def"
715 
716     // For most targets @llvm.get.dynamic.area.offset just returns 0.
717     setOperationAction(ISD::GET_DYNAMIC_AREA_OFFSET, VT, Expand);
718 
719     // Vector reduction default to expand.
720     setOperationAction(ISD::VECREDUCE_FADD, VT, Expand);
721     setOperationAction(ISD::VECREDUCE_FMUL, VT, Expand);
722     setOperationAction(ISD::VECREDUCE_ADD, VT, Expand);
723     setOperationAction(ISD::VECREDUCE_MUL, VT, Expand);
724     setOperationAction(ISD::VECREDUCE_AND, VT, Expand);
725     setOperationAction(ISD::VECREDUCE_OR, VT, Expand);
726     setOperationAction(ISD::VECREDUCE_XOR, VT, Expand);
727     setOperationAction(ISD::VECREDUCE_SMAX, VT, Expand);
728     setOperationAction(ISD::VECREDUCE_SMIN, VT, Expand);
729     setOperationAction(ISD::VECREDUCE_UMAX, VT, Expand);
730     setOperationAction(ISD::VECREDUCE_UMIN, VT, Expand);
731     setOperationAction(ISD::VECREDUCE_FMAX, VT, Expand);
732     setOperationAction(ISD::VECREDUCE_FMIN, VT, Expand);
733   }
734 
735   // Most targets ignore the @llvm.prefetch intrinsic.
736   setOperationAction(ISD::PREFETCH, MVT::Other, Expand);
737 
738   // Most targets also ignore the @llvm.readcyclecounter intrinsic.
739   setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Expand);
740 
741   // ConstantFP nodes default to expand.  Targets can either change this to
742   // Legal, in which case all fp constants are legal, or use isFPImmLegal()
743   // to optimize expansions for certain constants.
744   setOperationAction(ISD::ConstantFP, MVT::f16, Expand);
745   setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
746   setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
747   setOperationAction(ISD::ConstantFP, MVT::f80, Expand);
748   setOperationAction(ISD::ConstantFP, MVT::f128, Expand);
749 
750   // These library functions default to expand.
751   for (MVT VT : {MVT::f32, MVT::f64, MVT::f128}) {
752     setOperationAction(ISD::FCBRT,      VT, Expand);
753     setOperationAction(ISD::FLOG ,      VT, Expand);
754     setOperationAction(ISD::FLOG2,      VT, Expand);
755     setOperationAction(ISD::FLOG10,     VT, Expand);
756     setOperationAction(ISD::FEXP ,      VT, Expand);
757     setOperationAction(ISD::FEXP2,      VT, Expand);
758     setOperationAction(ISD::FFLOOR,     VT, Expand);
759     setOperationAction(ISD::FNEARBYINT, VT, Expand);
760     setOperationAction(ISD::FCEIL,      VT, Expand);
761     setOperationAction(ISD::FRINT,      VT, Expand);
762     setOperationAction(ISD::FTRUNC,     VT, Expand);
763     setOperationAction(ISD::FROUND,     VT, Expand);
764     setOperationAction(ISD::FROUNDEVEN, VT, Expand);
765     setOperationAction(ISD::LROUND,     VT, Expand);
766     setOperationAction(ISD::LLROUND,    VT, Expand);
767     setOperationAction(ISD::LRINT,      VT, Expand);
768     setOperationAction(ISD::LLRINT,     VT, Expand);
769   }
770 
771   // Default ISD::TRAP to expand (which turns it into abort).
772   setOperationAction(ISD::TRAP, MVT::Other, Expand);
773 
774   // On most systems, DEBUGTRAP and TRAP have no difference. The "Expand"
775   // here is to inform DAG Legalizer to replace DEBUGTRAP with TRAP.
776   setOperationAction(ISD::DEBUGTRAP, MVT::Other, Expand);
777 }
778 
779 MVT TargetLoweringBase::getScalarShiftAmountTy(const DataLayout &DL,
780                                                EVT) const {
781   return MVT::getIntegerVT(DL.getPointerSizeInBits(0));
782 }
783 
784 EVT TargetLoweringBase::getShiftAmountTy(EVT LHSTy, const DataLayout &DL,
785                                          bool LegalTypes) const {
786   assert(LHSTy.isInteger() && "Shift amount is not an integer type!");
787   if (LHSTy.isVector())
788     return LHSTy;
789   return LegalTypes ? getScalarShiftAmountTy(DL, LHSTy)
790                     : getPointerTy(DL);
791 }
792 
793 bool TargetLoweringBase::canOpTrap(unsigned Op, EVT VT) const {
794   assert(isTypeLegal(VT));
795   switch (Op) {
796   default:
797     return false;
798   case ISD::SDIV:
799   case ISD::UDIV:
800   case ISD::SREM:
801   case ISD::UREM:
802     return true;
803   }
804 }
805 
806 bool TargetLoweringBase::isFreeAddrSpaceCast(unsigned SrcAS,
807                                              unsigned DestAS) const {
808   return TM.isNoopAddrSpaceCast(SrcAS, DestAS);
809 }
810 
811 void TargetLoweringBase::setJumpIsExpensive(bool isExpensive) {
812   // If the command-line option was specified, ignore this request.
813   if (!JumpIsExpensiveOverride.getNumOccurrences())
814     JumpIsExpensive = isExpensive;
815 }
816 
817 TargetLoweringBase::LegalizeKind
818 TargetLoweringBase::getTypeConversion(LLVMContext &Context, EVT VT) const {
819   // If this is a simple type, use the ComputeRegisterProp mechanism.
820   if (VT.isSimple()) {
821     MVT SVT = VT.getSimpleVT();
822     assert((unsigned)SVT.SimpleTy < array_lengthof(TransformToType));
823     MVT NVT = TransformToType[SVT.SimpleTy];
824     LegalizeTypeAction LA = ValueTypeActions.getTypeAction(SVT);
825 
826     assert((LA == TypeLegal || LA == TypeSoftenFloat ||
827             LA == TypeSoftPromoteHalf ||
828             (NVT.isVector() ||
829              ValueTypeActions.getTypeAction(NVT) != TypePromoteInteger)) &&
830            "Promote may not follow Expand or Promote");
831 
832     if (LA == TypeSplitVector)
833       return LegalizeKind(LA,
834                           EVT::getVectorVT(Context, SVT.getVectorElementType(),
835                                            SVT.getVectorElementCount() / 2));
836     if (LA == TypeScalarizeVector)
837       return LegalizeKind(LA, SVT.getVectorElementType());
838     return LegalizeKind(LA, NVT);
839   }
840 
841   // Handle Extended Scalar Types.
842   if (!VT.isVector()) {
843     assert(VT.isInteger() && "Float types must be simple");
844     unsigned BitSize = VT.getSizeInBits();
845     // First promote to a power-of-two size, then expand if necessary.
846     if (BitSize < 8 || !isPowerOf2_32(BitSize)) {
847       EVT NVT = VT.getRoundIntegerType(Context);
848       assert(NVT != VT && "Unable to round integer VT");
849       LegalizeKind NextStep = getTypeConversion(Context, NVT);
850       // Avoid multi-step promotion.
851       if (NextStep.first == TypePromoteInteger)
852         return NextStep;
853       // Return rounded integer type.
854       return LegalizeKind(TypePromoteInteger, NVT);
855     }
856 
857     return LegalizeKind(TypeExpandInteger,
858                         EVT::getIntegerVT(Context, VT.getSizeInBits() / 2));
859   }
860 
861   // Handle vector types.
862   ElementCount NumElts = VT.getVectorElementCount();
863   EVT EltVT = VT.getVectorElementType();
864 
865   // Vectors with only one element are always scalarized.
866   if (NumElts == 1)
867     return LegalizeKind(TypeScalarizeVector, EltVT);
868 
869   if (VT.getVectorElementCount() == ElementCount::getScalable(1))
870     report_fatal_error("Cannot legalize this vector");
871 
872   // Try to widen vector elements until the element type is a power of two and
873   // promote it to a legal type later on, for example:
874   // <3 x i8> -> <4 x i8> -> <4 x i32>
875   if (EltVT.isInteger()) {
876     // Vectors with a number of elements that is not a power of two are always
877     // widened, for example <3 x i8> -> <4 x i8>.
878     if (!VT.isPow2VectorType()) {
879       NumElts = NumElts.NextPowerOf2();
880       EVT NVT = EVT::getVectorVT(Context, EltVT, NumElts);
881       return LegalizeKind(TypeWidenVector, NVT);
882     }
883 
884     // Examine the element type.
885     LegalizeKind LK = getTypeConversion(Context, EltVT);
886 
887     // If type is to be expanded, split the vector.
888     //  <4 x i140> -> <2 x i140>
889     if (LK.first == TypeExpandInteger)
890       return LegalizeKind(TypeSplitVector,
891                           EVT::getVectorVT(Context, EltVT, NumElts / 2));
892 
893     // Promote the integer element types until a legal vector type is found
894     // or until the element integer type is too big. If a legal type was not
895     // found, fallback to the usual mechanism of widening/splitting the
896     // vector.
897     EVT OldEltVT = EltVT;
898     while (true) {
899       // Increase the bitwidth of the element to the next pow-of-two
900       // (which is greater than 8 bits).
901       EltVT = EVT::getIntegerVT(Context, 1 + EltVT.getSizeInBits())
902                   .getRoundIntegerType(Context);
903 
904       // Stop trying when getting a non-simple element type.
905       // Note that vector elements may be greater than legal vector element
906       // types. Example: X86 XMM registers hold 64bit element on 32bit
907       // systems.
908       if (!EltVT.isSimple())
909         break;
910 
911       // Build a new vector type and check if it is legal.
912       MVT NVT = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts);
913       // Found a legal promoted vector type.
914       if (NVT != MVT() && ValueTypeActions.getTypeAction(NVT) == TypeLegal)
915         return LegalizeKind(TypePromoteInteger,
916                             EVT::getVectorVT(Context, EltVT, NumElts));
917     }
918 
919     // Reset the type to the unexpanded type if we did not find a legal vector
920     // type with a promoted vector element type.
921     EltVT = OldEltVT;
922   }
923 
924   // Try to widen the vector until a legal type is found.
925   // If there is no wider legal type, split the vector.
926   while (true) {
927     // Round up to the next power of 2.
928     NumElts = NumElts.NextPowerOf2();
929 
930     // If there is no simple vector type with this many elements then there
931     // cannot be a larger legal vector type.  Note that this assumes that
932     // there are no skipped intermediate vector types in the simple types.
933     if (!EltVT.isSimple())
934       break;
935     MVT LargerVector = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts);
936     if (LargerVector == MVT())
937       break;
938 
939     // If this type is legal then widen the vector.
940     if (ValueTypeActions.getTypeAction(LargerVector) == TypeLegal)
941       return LegalizeKind(TypeWidenVector, LargerVector);
942   }
943 
944   // Widen odd vectors to next power of two.
945   if (!VT.isPow2VectorType()) {
946     EVT NVT = VT.getPow2VectorType(Context);
947     return LegalizeKind(TypeWidenVector, NVT);
948   }
949 
950   // Vectors with illegal element types are expanded.
951   EVT NVT = EVT::getVectorVT(Context, EltVT, VT.getVectorElementCount() / 2);
952   return LegalizeKind(TypeSplitVector, NVT);
953 }
954 
955 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT,
956                                           unsigned &NumIntermediates,
957                                           MVT &RegisterVT,
958                                           TargetLoweringBase *TLI) {
959   // Figure out the right, legal destination reg to copy into.
960   ElementCount EC = VT.getVectorElementCount();
961   MVT EltTy = VT.getVectorElementType();
962 
963   unsigned NumVectorRegs = 1;
964 
965   // Scalable vectors cannot be scalarized, so splitting or widening is
966   // required.
967   if (VT.isScalableVector() && !isPowerOf2_32(EC.Min))
968     llvm_unreachable(
969         "Splitting or widening of non-power-of-2 MVTs is not implemented.");
970 
971   // FIXME: We don't support non-power-of-2-sized vectors for now.
972   // Ideally we could break down into LHS/RHS like LegalizeDAG does.
973   if (!isPowerOf2_32(EC.Min)) {
974     // Split EC to unit size (scalable property is preserved).
975     NumVectorRegs = EC.Min;
976     EC = EC / NumVectorRegs;
977   }
978 
979   // Divide the input until we get to a supported size. This will
980   // always end up with an EC that represent a scalar or a scalable
981   // scalar.
982   while (EC.Min > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, EC))) {
983     EC.Min >>= 1;
984     NumVectorRegs <<= 1;
985   }
986 
987   NumIntermediates = NumVectorRegs;
988 
989   MVT NewVT = MVT::getVectorVT(EltTy, EC);
990   if (!TLI->isTypeLegal(NewVT))
991     NewVT = EltTy;
992   IntermediateVT = NewVT;
993 
994   unsigned LaneSizeInBits = NewVT.getScalarSizeInBits().getFixedSize();
995 
996   // Convert sizes such as i33 to i64.
997   if (!isPowerOf2_32(LaneSizeInBits))
998     LaneSizeInBits = NextPowerOf2(LaneSizeInBits);
999 
1000   MVT DestVT = TLI->getRegisterType(NewVT);
1001   RegisterVT = DestVT;
1002   if (EVT(DestVT).bitsLT(NewVT))    // Value is expanded, e.g. i64 -> i16.
1003     return NumVectorRegs *
1004            (LaneSizeInBits / DestVT.getScalarSizeInBits().getFixedSize());
1005 
1006   // Otherwise, promotion or legal types use the same number of registers as
1007   // the vector decimated to the appropriate level.
1008   return NumVectorRegs;
1009 }
1010 
1011 /// isLegalRC - Return true if the value types that can be represented by the
1012 /// specified register class are all legal.
1013 bool TargetLoweringBase::isLegalRC(const TargetRegisterInfo &TRI,
1014                                    const TargetRegisterClass &RC) const {
1015   for (auto I = TRI.legalclasstypes_begin(RC); *I != MVT::Other; ++I)
1016     if (isTypeLegal(*I))
1017       return true;
1018   return false;
1019 }
1020 
1021 /// Replace/modify any TargetFrameIndex operands with a targte-dependent
1022 /// sequence of memory operands that is recognized by PrologEpilogInserter.
1023 MachineBasicBlock *
1024 TargetLoweringBase::emitPatchPoint(MachineInstr &InitialMI,
1025                                    MachineBasicBlock *MBB) const {
1026   MachineInstr *MI = &InitialMI;
1027   MachineFunction &MF = *MI->getMF();
1028   MachineFrameInfo &MFI = MF.getFrameInfo();
1029 
1030   // We're handling multiple types of operands here:
1031   // PATCHPOINT MetaArgs - live-in, read only, direct
1032   // STATEPOINT Deopt Spill - live-through, read only, indirect
1033   // STATEPOINT Deopt Alloca - live-through, read only, direct
1034   // (We're currently conservative and mark the deopt slots read/write in
1035   // practice.)
1036   // STATEPOINT GC Spill - live-through, read/write, indirect
1037   // STATEPOINT GC Alloca - live-through, read/write, direct
1038   // The live-in vs live-through is handled already (the live through ones are
1039   // all stack slots), but we need to handle the different type of stackmap
1040   // operands and memory effects here.
1041 
1042   if (!llvm::any_of(MI->operands(),
1043                     [](MachineOperand &Operand) { return Operand.isFI(); }))
1044     return MBB;
1045 
1046   MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), MI->getDesc());
1047 
1048   // Inherit previous memory operands.
1049   MIB.cloneMemRefs(*MI);
1050 
1051   for (unsigned i = 0; i < MI->getNumOperands(); ++i) {
1052     MachineOperand &MO = MI->getOperand(i);
1053     if (!MO.isFI()) {
1054       // Index of Def operand this Use it tied to.
1055       // Since Defs are coming before Uses, if Use is tied, then
1056       // index of Def must be smaller that index of that Use.
1057       // Also, Defs preserve their position in new MI.
1058       unsigned TiedTo = i;
1059       if (MO.isReg() && MO.isTied())
1060         TiedTo = MI->findTiedOperandIdx(i);
1061       MIB.add(MO);
1062       if (TiedTo < i)
1063         MIB->tieOperands(TiedTo, MIB->getNumOperands() - 1);
1064       continue;
1065     }
1066 
1067     // foldMemoryOperand builds a new MI after replacing a single FI operand
1068     // with the canonical set of five x86 addressing-mode operands.
1069     int FI = MO.getIndex();
1070 
1071     // Add frame index operands recognized by stackmaps.cpp
1072     if (MFI.isStatepointSpillSlotObjectIndex(FI)) {
1073       // indirect-mem-ref tag, size, #FI, offset.
1074       // Used for spills inserted by StatepointLowering.  This codepath is not
1075       // used for patchpoints/stackmaps at all, for these spilling is done via
1076       // foldMemoryOperand callback only.
1077       assert(MI->getOpcode() == TargetOpcode::STATEPOINT && "sanity");
1078       MIB.addImm(StackMaps::IndirectMemRefOp);
1079       MIB.addImm(MFI.getObjectSize(FI));
1080       MIB.add(MO);
1081       MIB.addImm(0);
1082     } else {
1083       // direct-mem-ref tag, #FI, offset.
1084       // Used by patchpoint, and direct alloca arguments to statepoints
1085       MIB.addImm(StackMaps::DirectMemRefOp);
1086       MIB.add(MO);
1087       MIB.addImm(0);
1088     }
1089 
1090     assert(MIB->mayLoad() && "Folded a stackmap use to a non-load!");
1091 
1092     // Add a new memory operand for this FI.
1093     assert(MFI.getObjectOffset(FI) != -1);
1094 
1095     // Note: STATEPOINT MMOs are added during SelectionDAG.  STACKMAP, and
1096     // PATCHPOINT should be updated to do the same. (TODO)
1097     if (MI->getOpcode() != TargetOpcode::STATEPOINT) {
1098       auto Flags = MachineMemOperand::MOLoad;
1099       MachineMemOperand *MMO = MF.getMachineMemOperand(
1100           MachinePointerInfo::getFixedStack(MF, FI), Flags,
1101           MF.getDataLayout().getPointerSize(), MFI.getObjectAlign(FI));
1102       MIB->addMemOperand(MF, MMO);
1103     }
1104   }
1105   MBB->insert(MachineBasicBlock::iterator(MI), MIB);
1106   MI->eraseFromParent();
1107   return MBB;
1108 }
1109 
1110 MachineBasicBlock *
1111 TargetLoweringBase::emitXRayCustomEvent(MachineInstr &MI,
1112                                         MachineBasicBlock *MBB) const {
1113   assert(MI.getOpcode() == TargetOpcode::PATCHABLE_EVENT_CALL &&
1114          "Called emitXRayCustomEvent on the wrong MI!");
1115   auto &MF = *MI.getMF();
1116   auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc());
1117   for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx)
1118     MIB.add(MI.getOperand(OpIdx));
1119 
1120   MBB->insert(MachineBasicBlock::iterator(MI), MIB);
1121   MI.eraseFromParent();
1122   return MBB;
1123 }
1124 
1125 MachineBasicBlock *
1126 TargetLoweringBase::emitXRayTypedEvent(MachineInstr &MI,
1127                                        MachineBasicBlock *MBB) const {
1128   assert(MI.getOpcode() == TargetOpcode::PATCHABLE_TYPED_EVENT_CALL &&
1129          "Called emitXRayTypedEvent on the wrong MI!");
1130   auto &MF = *MI.getMF();
1131   auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc());
1132   for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx)
1133     MIB.add(MI.getOperand(OpIdx));
1134 
1135   MBB->insert(MachineBasicBlock::iterator(MI), MIB);
1136   MI.eraseFromParent();
1137   return MBB;
1138 }
1139 
1140 /// findRepresentativeClass - Return the largest legal super-reg register class
1141 /// of the register class for the specified type and its associated "cost".
1142 // This function is in TargetLowering because it uses RegClassForVT which would
1143 // need to be moved to TargetRegisterInfo and would necessitate moving
1144 // isTypeLegal over as well - a massive change that would just require
1145 // TargetLowering having a TargetRegisterInfo class member that it would use.
1146 std::pair<const TargetRegisterClass *, uint8_t>
1147 TargetLoweringBase::findRepresentativeClass(const TargetRegisterInfo *TRI,
1148                                             MVT VT) const {
1149   const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy];
1150   if (!RC)
1151     return std::make_pair(RC, 0);
1152 
1153   // Compute the set of all super-register classes.
1154   BitVector SuperRegRC(TRI->getNumRegClasses());
1155   for (SuperRegClassIterator RCI(RC, TRI); RCI.isValid(); ++RCI)
1156     SuperRegRC.setBitsInMask(RCI.getMask());
1157 
1158   // Find the first legal register class with the largest spill size.
1159   const TargetRegisterClass *BestRC = RC;
1160   for (unsigned i : SuperRegRC.set_bits()) {
1161     const TargetRegisterClass *SuperRC = TRI->getRegClass(i);
1162     // We want the largest possible spill size.
1163     if (TRI->getSpillSize(*SuperRC) <= TRI->getSpillSize(*BestRC))
1164       continue;
1165     if (!isLegalRC(*TRI, *SuperRC))
1166       continue;
1167     BestRC = SuperRC;
1168   }
1169   return std::make_pair(BestRC, 1);
1170 }
1171 
1172 /// computeRegisterProperties - Once all of the register classes are added,
1173 /// this allows us to compute derived properties we expose.
1174 void TargetLoweringBase::computeRegisterProperties(
1175     const TargetRegisterInfo *TRI) {
1176   static_assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE,
1177                 "Too many value types for ValueTypeActions to hold!");
1178 
1179   // Everything defaults to needing one register.
1180   for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
1181     NumRegistersForVT[i] = 1;
1182     RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i;
1183   }
1184   // ...except isVoid, which doesn't need any registers.
1185   NumRegistersForVT[MVT::isVoid] = 0;
1186 
1187   // Find the largest integer register class.
1188   unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE;
1189   for (; RegClassForVT[LargestIntReg] == nullptr; --LargestIntReg)
1190     assert(LargestIntReg != MVT::i1 && "No integer registers defined!");
1191 
1192   // Every integer value type larger than this largest register takes twice as
1193   // many registers to represent as the previous ValueType.
1194   for (unsigned ExpandedReg = LargestIntReg + 1;
1195        ExpandedReg <= MVT::LAST_INTEGER_VALUETYPE; ++ExpandedReg) {
1196     NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1];
1197     RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg;
1198     TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1);
1199     ValueTypeActions.setTypeAction((MVT::SimpleValueType)ExpandedReg,
1200                                    TypeExpandInteger);
1201   }
1202 
1203   // Inspect all of the ValueType's smaller than the largest integer
1204   // register to see which ones need promotion.
1205   unsigned LegalIntReg = LargestIntReg;
1206   for (unsigned IntReg = LargestIntReg - 1;
1207        IntReg >= (unsigned)MVT::i1; --IntReg) {
1208     MVT IVT = (MVT::SimpleValueType)IntReg;
1209     if (isTypeLegal(IVT)) {
1210       LegalIntReg = IntReg;
1211     } else {
1212       RegisterTypeForVT[IntReg] = TransformToType[IntReg] =
1213         (MVT::SimpleValueType)LegalIntReg;
1214       ValueTypeActions.setTypeAction(IVT, TypePromoteInteger);
1215     }
1216   }
1217 
1218   // ppcf128 type is really two f64's.
1219   if (!isTypeLegal(MVT::ppcf128)) {
1220     if (isTypeLegal(MVT::f64)) {
1221       NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64];
1222       RegisterTypeForVT[MVT::ppcf128] = MVT::f64;
1223       TransformToType[MVT::ppcf128] = MVT::f64;
1224       ValueTypeActions.setTypeAction(MVT::ppcf128, TypeExpandFloat);
1225     } else {
1226       NumRegistersForVT[MVT::ppcf128] = NumRegistersForVT[MVT::i128];
1227       RegisterTypeForVT[MVT::ppcf128] = RegisterTypeForVT[MVT::i128];
1228       TransformToType[MVT::ppcf128] = MVT::i128;
1229       ValueTypeActions.setTypeAction(MVT::ppcf128, TypeSoftenFloat);
1230     }
1231   }
1232 
1233   // Decide how to handle f128. If the target does not have native f128 support,
1234   // expand it to i128 and we will be generating soft float library calls.
1235   if (!isTypeLegal(MVT::f128)) {
1236     NumRegistersForVT[MVT::f128] = NumRegistersForVT[MVT::i128];
1237     RegisterTypeForVT[MVT::f128] = RegisterTypeForVT[MVT::i128];
1238     TransformToType[MVT::f128] = MVT::i128;
1239     ValueTypeActions.setTypeAction(MVT::f128, TypeSoftenFloat);
1240   }
1241 
1242   // Decide how to handle f64. If the target does not have native f64 support,
1243   // expand it to i64 and we will be generating soft float library calls.
1244   if (!isTypeLegal(MVT::f64)) {
1245     NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64];
1246     RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64];
1247     TransformToType[MVT::f64] = MVT::i64;
1248     ValueTypeActions.setTypeAction(MVT::f64, TypeSoftenFloat);
1249   }
1250 
1251   // Decide how to handle f32. If the target does not have native f32 support,
1252   // expand it to i32 and we will be generating soft float library calls.
1253   if (!isTypeLegal(MVT::f32)) {
1254     NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32];
1255     RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32];
1256     TransformToType[MVT::f32] = MVT::i32;
1257     ValueTypeActions.setTypeAction(MVT::f32, TypeSoftenFloat);
1258   }
1259 
1260   // Decide how to handle f16. If the target does not have native f16 support,
1261   // promote it to f32, because there are no f16 library calls (except for
1262   // conversions).
1263   if (!isTypeLegal(MVT::f16)) {
1264     // Allow targets to control how we legalize half.
1265     if (softPromoteHalfType()) {
1266       NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::i16];
1267       RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::i16];
1268       TransformToType[MVT::f16] = MVT::f32;
1269       ValueTypeActions.setTypeAction(MVT::f16, TypeSoftPromoteHalf);
1270     } else {
1271       NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::f32];
1272       RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::f32];
1273       TransformToType[MVT::f16] = MVT::f32;
1274       ValueTypeActions.setTypeAction(MVT::f16, TypePromoteFloat);
1275     }
1276   }
1277 
1278   // Loop over all of the vector value types to see which need transformations.
1279   for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE;
1280        i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
1281     MVT VT = (MVT::SimpleValueType) i;
1282     if (isTypeLegal(VT))
1283       continue;
1284 
1285     MVT EltVT = VT.getVectorElementType();
1286     ElementCount EC = VT.getVectorElementCount();
1287     bool IsLegalWiderType = false;
1288     bool IsScalable = VT.isScalableVector();
1289     LegalizeTypeAction PreferredAction = getPreferredVectorAction(VT);
1290     switch (PreferredAction) {
1291     case TypePromoteInteger: {
1292       MVT::SimpleValueType EndVT = IsScalable ?
1293                                    MVT::LAST_INTEGER_SCALABLE_VECTOR_VALUETYPE :
1294                                    MVT::LAST_INTEGER_FIXEDLEN_VECTOR_VALUETYPE;
1295       // Try to promote the elements of integer vectors. If no legal
1296       // promotion was found, fall through to the widen-vector method.
1297       for (unsigned nVT = i + 1;
1298            (MVT::SimpleValueType)nVT <= EndVT; ++nVT) {
1299         MVT SVT = (MVT::SimpleValueType) nVT;
1300         // Promote vectors of integers to vectors with the same number
1301         // of elements, with a wider element type.
1302         if (SVT.getScalarSizeInBits() > EltVT.getSizeInBits() &&
1303             SVT.getVectorElementCount() == EC && isTypeLegal(SVT)) {
1304           TransformToType[i] = SVT;
1305           RegisterTypeForVT[i] = SVT;
1306           NumRegistersForVT[i] = 1;
1307           ValueTypeActions.setTypeAction(VT, TypePromoteInteger);
1308           IsLegalWiderType = true;
1309           break;
1310         }
1311       }
1312       if (IsLegalWiderType)
1313         break;
1314       LLVM_FALLTHROUGH;
1315     }
1316 
1317     case TypeWidenVector:
1318       if (isPowerOf2_32(EC.Min)) {
1319         // Try to widen the vector.
1320         for (unsigned nVT = i + 1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
1321           MVT SVT = (MVT::SimpleValueType) nVT;
1322           if (SVT.getVectorElementType() == EltVT &&
1323               SVT.isScalableVector() == IsScalable &&
1324               SVT.getVectorElementCount().Min > EC.Min && isTypeLegal(SVT)) {
1325             TransformToType[i] = SVT;
1326             RegisterTypeForVT[i] = SVT;
1327             NumRegistersForVT[i] = 1;
1328             ValueTypeActions.setTypeAction(VT, TypeWidenVector);
1329             IsLegalWiderType = true;
1330             break;
1331           }
1332         }
1333         if (IsLegalWiderType)
1334           break;
1335       } else {
1336         // Only widen to the next power of 2 to keep consistency with EVT.
1337         MVT NVT = VT.getPow2VectorType();
1338         if (isTypeLegal(NVT)) {
1339           TransformToType[i] = NVT;
1340           ValueTypeActions.setTypeAction(VT, TypeWidenVector);
1341           RegisterTypeForVT[i] = NVT;
1342           NumRegistersForVT[i] = 1;
1343           break;
1344         }
1345       }
1346       LLVM_FALLTHROUGH;
1347 
1348     case TypeSplitVector:
1349     case TypeScalarizeVector: {
1350       MVT IntermediateVT;
1351       MVT RegisterVT;
1352       unsigned NumIntermediates;
1353       unsigned NumRegisters = getVectorTypeBreakdownMVT(VT, IntermediateVT,
1354           NumIntermediates, RegisterVT, this);
1355       NumRegistersForVT[i] = NumRegisters;
1356       assert(NumRegistersForVT[i] == NumRegisters &&
1357              "NumRegistersForVT size cannot represent NumRegisters!");
1358       RegisterTypeForVT[i] = RegisterVT;
1359 
1360       MVT NVT = VT.getPow2VectorType();
1361       if (NVT == VT) {
1362         // Type is already a power of 2.  The default action is to split.
1363         TransformToType[i] = MVT::Other;
1364         if (PreferredAction == TypeScalarizeVector)
1365           ValueTypeActions.setTypeAction(VT, TypeScalarizeVector);
1366         else if (PreferredAction == TypeSplitVector)
1367           ValueTypeActions.setTypeAction(VT, TypeSplitVector);
1368         else if (EC.Min > 1)
1369           ValueTypeActions.setTypeAction(VT, TypeSplitVector);
1370         else
1371           ValueTypeActions.setTypeAction(VT, EC.Scalable
1372                                                  ? TypeScalarizeScalableVector
1373                                                  : TypeScalarizeVector);
1374       } else {
1375         TransformToType[i] = NVT;
1376         ValueTypeActions.setTypeAction(VT, TypeWidenVector);
1377       }
1378       break;
1379     }
1380     default:
1381       llvm_unreachable("Unknown vector legalization action!");
1382     }
1383   }
1384 
1385   // Determine the 'representative' register class for each value type.
1386   // An representative register class is the largest (meaning one which is
1387   // not a sub-register class / subreg register class) legal register class for
1388   // a group of value types. For example, on i386, i8, i16, and i32
1389   // representative would be GR32; while on x86_64 it's GR64.
1390   for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
1391     const TargetRegisterClass* RRC;
1392     uint8_t Cost;
1393     std::tie(RRC, Cost) = findRepresentativeClass(TRI, (MVT::SimpleValueType)i);
1394     RepRegClassForVT[i] = RRC;
1395     RepRegClassCostForVT[i] = Cost;
1396   }
1397 }
1398 
1399 EVT TargetLoweringBase::getSetCCResultType(const DataLayout &DL, LLVMContext &,
1400                                            EVT VT) const {
1401   assert(!VT.isVector() && "No default SetCC type for vectors!");
1402   return getPointerTy(DL).SimpleTy;
1403 }
1404 
1405 MVT::SimpleValueType TargetLoweringBase::getCmpLibcallReturnType() const {
1406   return MVT::i32; // return the default value
1407 }
1408 
1409 /// getVectorTypeBreakdown - Vector types are broken down into some number of
1410 /// legal first class types.  For example, MVT::v8f32 maps to 2 MVT::v4f32
1411 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
1412 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
1413 ///
1414 /// This method returns the number of registers needed, and the VT for each
1415 /// register.  It also returns the VT and quantity of the intermediate values
1416 /// before they are promoted/expanded.
1417 unsigned TargetLoweringBase::getVectorTypeBreakdown(LLVMContext &Context, EVT VT,
1418                                                 EVT &IntermediateVT,
1419                                                 unsigned &NumIntermediates,
1420                                                 MVT &RegisterVT) const {
1421   ElementCount EltCnt = VT.getVectorElementCount();
1422 
1423   // If there is a wider vector type with the same element type as this one,
1424   // or a promoted vector type that has the same number of elements which
1425   // are wider, then we should convert to that legal vector type.
1426   // This handles things like <2 x float> -> <4 x float> and
1427   // <4 x i1> -> <4 x i32>.
1428   LegalizeTypeAction TA = getTypeAction(Context, VT);
1429   if (EltCnt.Min != 1 && (TA == TypeWidenVector || TA == TypePromoteInteger)) {
1430     EVT RegisterEVT = getTypeToTransformTo(Context, VT);
1431     if (isTypeLegal(RegisterEVT)) {
1432       IntermediateVT = RegisterEVT;
1433       RegisterVT = RegisterEVT.getSimpleVT();
1434       NumIntermediates = 1;
1435       return 1;
1436     }
1437   }
1438 
1439   // Figure out the right, legal destination reg to copy into.
1440   EVT EltTy = VT.getVectorElementType();
1441 
1442   unsigned NumVectorRegs = 1;
1443 
1444   // Scalable vectors cannot be scalarized, so handle the legalisation of the
1445   // types like done elsewhere in SelectionDAG.
1446   if (VT.isScalableVector() && !isPowerOf2_32(EltCnt.Min)) {
1447     LegalizeKind LK;
1448     EVT PartVT = VT;
1449     do {
1450       // Iterate until we've found a legal (part) type to hold VT.
1451       LK = getTypeConversion(Context, PartVT);
1452       PartVT = LK.second;
1453     } while (LK.first != TypeLegal);
1454 
1455     NumIntermediates =
1456         VT.getVectorElementCount().Min / PartVT.getVectorElementCount().Min;
1457 
1458     // FIXME: This code needs to be extended to handle more complex vector
1459     // breakdowns, like nxv7i64 -> nxv8i64 -> 4 x nxv2i64. Currently the only
1460     // supported cases are vectors that are broken down into equal parts
1461     // such as nxv6i64 -> 3 x nxv2i64.
1462     assert(NumIntermediates * PartVT.getVectorElementCount().Min ==
1463                VT.getVectorElementCount().Min &&
1464            "Expected an integer multiple of PartVT");
1465     IntermediateVT = PartVT;
1466     RegisterVT = getRegisterType(Context, IntermediateVT);
1467     return NumIntermediates;
1468   }
1469 
1470   // FIXME: We don't support non-power-of-2-sized vectors for now.  Ideally
1471   // we could break down into LHS/RHS like LegalizeDAG does.
1472   if (!isPowerOf2_32(EltCnt.Min)) {
1473     NumVectorRegs = EltCnt.Min;
1474     EltCnt.Min = 1;
1475   }
1476 
1477   // Divide the input until we get to a supported size.  This will always
1478   // end with a scalar if the target doesn't support vectors.
1479   while (EltCnt.Min > 1 &&
1480          !isTypeLegal(EVT::getVectorVT(Context, EltTy, EltCnt))) {
1481     EltCnt.Min >>= 1;
1482     NumVectorRegs <<= 1;
1483   }
1484 
1485   NumIntermediates = NumVectorRegs;
1486 
1487   EVT NewVT = EVT::getVectorVT(Context, EltTy, EltCnt);
1488   if (!isTypeLegal(NewVT))
1489     NewVT = EltTy;
1490   IntermediateVT = NewVT;
1491 
1492   MVT DestVT = getRegisterType(Context, NewVT);
1493   RegisterVT = DestVT;
1494 
1495   if (EVT(DestVT).bitsLT(NewVT)) {  // Value is expanded, e.g. i64 -> i16.
1496     TypeSize NewVTSize = NewVT.getSizeInBits();
1497     // Convert sizes such as i33 to i64.
1498     if (!isPowerOf2_32(NewVTSize.getKnownMinSize()))
1499       NewVTSize = NewVTSize.NextPowerOf2();
1500     return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits());
1501   }
1502 
1503   // Otherwise, promotion or legal types use the same number of registers as
1504   // the vector decimated to the appropriate level.
1505   return NumVectorRegs;
1506 }
1507 
1508 bool TargetLoweringBase::isSuitableForJumpTable(const SwitchInst *SI,
1509                                                 uint64_t NumCases,
1510                                                 uint64_t Range,
1511                                                 ProfileSummaryInfo *PSI,
1512                                                 BlockFrequencyInfo *BFI) const {
1513   // FIXME: This function check the maximum table size and density, but the
1514   // minimum size is not checked. It would be nice if the minimum size is
1515   // also combined within this function. Currently, the minimum size check is
1516   // performed in findJumpTable() in SelectionDAGBuiler and
1517   // getEstimatedNumberOfCaseClusters() in BasicTTIImpl.
1518   const bool OptForSize =
1519       SI->getParent()->getParent()->hasOptSize() ||
1520       llvm::shouldOptimizeForSize(SI->getParent(), PSI, BFI);
1521   const unsigned MinDensity = getMinimumJumpTableDensity(OptForSize);
1522   const unsigned MaxJumpTableSize = getMaximumJumpTableSize();
1523 
1524   // Check whether the number of cases is small enough and
1525   // the range is dense enough for a jump table.
1526   return (OptForSize || Range <= MaxJumpTableSize) &&
1527          (NumCases * 100 >= Range * MinDensity);
1528 }
1529 
1530 /// Get the EVTs and ArgFlags collections that represent the legalized return
1531 /// type of the given function.  This does not require a DAG or a return value,
1532 /// and is suitable for use before any DAGs for the function are constructed.
1533 /// TODO: Move this out of TargetLowering.cpp.
1534 void llvm::GetReturnInfo(CallingConv::ID CC, Type *ReturnType,
1535                          AttributeList attr,
1536                          SmallVectorImpl<ISD::OutputArg> &Outs,
1537                          const TargetLowering &TLI, const DataLayout &DL) {
1538   SmallVector<EVT, 4> ValueVTs;
1539   ComputeValueVTs(TLI, DL, ReturnType, ValueVTs);
1540   unsigned NumValues = ValueVTs.size();
1541   if (NumValues == 0) return;
1542 
1543   for (unsigned j = 0, f = NumValues; j != f; ++j) {
1544     EVT VT = ValueVTs[j];
1545     ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
1546 
1547     if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt))
1548       ExtendKind = ISD::SIGN_EXTEND;
1549     else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt))
1550       ExtendKind = ISD::ZERO_EXTEND;
1551 
1552     // FIXME: C calling convention requires the return type to be promoted to
1553     // at least 32-bit. But this is not necessary for non-C calling
1554     // conventions. The frontend should mark functions whose return values
1555     // require promoting with signext or zeroext attributes.
1556     if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) {
1557       MVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32);
1558       if (VT.bitsLT(MinVT))
1559         VT = MinVT;
1560     }
1561 
1562     unsigned NumParts =
1563         TLI.getNumRegistersForCallingConv(ReturnType->getContext(), CC, VT);
1564     MVT PartVT =
1565         TLI.getRegisterTypeForCallingConv(ReturnType->getContext(), CC, VT);
1566 
1567     // 'inreg' on function refers to return value
1568     ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
1569     if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::InReg))
1570       Flags.setInReg();
1571 
1572     // Propagate extension type if any
1573     if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt))
1574       Flags.setSExt();
1575     else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt))
1576       Flags.setZExt();
1577 
1578     for (unsigned i = 0; i < NumParts; ++i)
1579       Outs.push_back(ISD::OutputArg(Flags, PartVT, VT, /*isfixed=*/true, 0, 0));
1580   }
1581 }
1582 
1583 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1584 /// function arguments in the caller parameter area.  This is the actual
1585 /// alignment, not its logarithm.
1586 unsigned TargetLoweringBase::getByValTypeAlignment(Type *Ty,
1587                                                    const DataLayout &DL) const {
1588   return DL.getABITypeAlign(Ty).value();
1589 }
1590 
1591 bool TargetLoweringBase::allowsMemoryAccessForAlignment(
1592     LLVMContext &Context, const DataLayout &DL, EVT VT, unsigned AddrSpace,
1593     Align Alignment, MachineMemOperand::Flags Flags, bool *Fast) const {
1594   // Check if the specified alignment is sufficient based on the data layout.
1595   // TODO: While using the data layout works in practice, a better solution
1596   // would be to implement this check directly (make this a virtual function).
1597   // For example, the ABI alignment may change based on software platform while
1598   // this function should only be affected by hardware implementation.
1599   Type *Ty = VT.getTypeForEVT(Context);
1600   if (Alignment >= DL.getABITypeAlign(Ty)) {
1601     // Assume that an access that meets the ABI-specified alignment is fast.
1602     if (Fast != nullptr)
1603       *Fast = true;
1604     return true;
1605   }
1606 
1607   // This is a misaligned access.
1608   return allowsMisalignedMemoryAccesses(VT, AddrSpace, Alignment.value(), Flags,
1609                                         Fast);
1610 }
1611 
1612 bool TargetLoweringBase::allowsMemoryAccessForAlignment(
1613     LLVMContext &Context, const DataLayout &DL, EVT VT,
1614     const MachineMemOperand &MMO, bool *Fast) const {
1615   return allowsMemoryAccessForAlignment(Context, DL, VT, MMO.getAddrSpace(),
1616                                         MMO.getAlign(), MMO.getFlags(), Fast);
1617 }
1618 
1619 bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context,
1620                                             const DataLayout &DL, EVT VT,
1621                                             unsigned AddrSpace, Align Alignment,
1622                                             MachineMemOperand::Flags Flags,
1623                                             bool *Fast) const {
1624   return allowsMemoryAccessForAlignment(Context, DL, VT, AddrSpace, Alignment,
1625                                         Flags, Fast);
1626 }
1627 
1628 bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context,
1629                                             const DataLayout &DL, EVT VT,
1630                                             const MachineMemOperand &MMO,
1631                                             bool *Fast) const {
1632   return allowsMemoryAccess(Context, DL, VT, MMO.getAddrSpace(), MMO.getAlign(),
1633                             MMO.getFlags(), Fast);
1634 }
1635 
1636 BranchProbability TargetLoweringBase::getPredictableBranchThreshold() const {
1637   return BranchProbability(MinPercentageForPredictableBranch, 100);
1638 }
1639 
1640 //===----------------------------------------------------------------------===//
1641 //  TargetTransformInfo Helpers
1642 //===----------------------------------------------------------------------===//
1643 
1644 int TargetLoweringBase::InstructionOpcodeToISD(unsigned Opcode) const {
1645   enum InstructionOpcodes {
1646 #define HANDLE_INST(NUM, OPCODE, CLASS) OPCODE = NUM,
1647 #define LAST_OTHER_INST(NUM) InstructionOpcodesCount = NUM
1648 #include "llvm/IR/Instruction.def"
1649   };
1650   switch (static_cast<InstructionOpcodes>(Opcode)) {
1651   case Ret:            return 0;
1652   case Br:             return 0;
1653   case Switch:         return 0;
1654   case IndirectBr:     return 0;
1655   case Invoke:         return 0;
1656   case CallBr:         return 0;
1657   case Resume:         return 0;
1658   case Unreachable:    return 0;
1659   case CleanupRet:     return 0;
1660   case CatchRet:       return 0;
1661   case CatchPad:       return 0;
1662   case CatchSwitch:    return 0;
1663   case CleanupPad:     return 0;
1664   case FNeg:           return ISD::FNEG;
1665   case Add:            return ISD::ADD;
1666   case FAdd:           return ISD::FADD;
1667   case Sub:            return ISD::SUB;
1668   case FSub:           return ISD::FSUB;
1669   case Mul:            return ISD::MUL;
1670   case FMul:           return ISD::FMUL;
1671   case UDiv:           return ISD::UDIV;
1672   case SDiv:           return ISD::SDIV;
1673   case FDiv:           return ISD::FDIV;
1674   case URem:           return ISD::UREM;
1675   case SRem:           return ISD::SREM;
1676   case FRem:           return ISD::FREM;
1677   case Shl:            return ISD::SHL;
1678   case LShr:           return ISD::SRL;
1679   case AShr:           return ISD::SRA;
1680   case And:            return ISD::AND;
1681   case Or:             return ISD::OR;
1682   case Xor:            return ISD::XOR;
1683   case Alloca:         return 0;
1684   case Load:           return ISD::LOAD;
1685   case Store:          return ISD::STORE;
1686   case GetElementPtr:  return 0;
1687   case Fence:          return 0;
1688   case AtomicCmpXchg:  return 0;
1689   case AtomicRMW:      return 0;
1690   case Trunc:          return ISD::TRUNCATE;
1691   case ZExt:           return ISD::ZERO_EXTEND;
1692   case SExt:           return ISD::SIGN_EXTEND;
1693   case FPToUI:         return ISD::FP_TO_UINT;
1694   case FPToSI:         return ISD::FP_TO_SINT;
1695   case UIToFP:         return ISD::UINT_TO_FP;
1696   case SIToFP:         return ISD::SINT_TO_FP;
1697   case FPTrunc:        return ISD::FP_ROUND;
1698   case FPExt:          return ISD::FP_EXTEND;
1699   case PtrToInt:       return ISD::BITCAST;
1700   case IntToPtr:       return ISD::BITCAST;
1701   case BitCast:        return ISD::BITCAST;
1702   case AddrSpaceCast:  return ISD::ADDRSPACECAST;
1703   case ICmp:           return ISD::SETCC;
1704   case FCmp:           return ISD::SETCC;
1705   case PHI:            return 0;
1706   case Call:           return 0;
1707   case Select:         return ISD::SELECT;
1708   case UserOp1:        return 0;
1709   case UserOp2:        return 0;
1710   case VAArg:          return 0;
1711   case ExtractElement: return ISD::EXTRACT_VECTOR_ELT;
1712   case InsertElement:  return ISD::INSERT_VECTOR_ELT;
1713   case ShuffleVector:  return ISD::VECTOR_SHUFFLE;
1714   case ExtractValue:   return ISD::MERGE_VALUES;
1715   case InsertValue:    return ISD::MERGE_VALUES;
1716   case LandingPad:     return 0;
1717   case Freeze:         return ISD::FREEZE;
1718   }
1719 
1720   llvm_unreachable("Unknown instruction type encountered!");
1721 }
1722 
1723 std::pair<int, MVT>
1724 TargetLoweringBase::getTypeLegalizationCost(const DataLayout &DL,
1725                                             Type *Ty) const {
1726   LLVMContext &C = Ty->getContext();
1727   EVT MTy = getValueType(DL, Ty);
1728 
1729   int Cost = 1;
1730   // We keep legalizing the type until we find a legal kind. We assume that
1731   // the only operation that costs anything is the split. After splitting
1732   // we need to handle two types.
1733   while (true) {
1734     LegalizeKind LK = getTypeConversion(C, MTy);
1735 
1736     if (LK.first == TypeLegal)
1737       return std::make_pair(Cost, MTy.getSimpleVT());
1738 
1739     if (LK.first == TypeSplitVector || LK.first == TypeExpandInteger)
1740       Cost *= 2;
1741 
1742     // Do not loop with f128 type.
1743     if (MTy == LK.second)
1744       return std::make_pair(Cost, MTy.getSimpleVT());
1745 
1746     // Keep legalizing the type.
1747     MTy = LK.second;
1748   }
1749 }
1750 
1751 Value *TargetLoweringBase::getDefaultSafeStackPointerLocation(IRBuilder<> &IRB,
1752                                                               bool UseTLS) const {
1753   // compiler-rt provides a variable with a magic name.  Targets that do not
1754   // link with compiler-rt may also provide such a variable.
1755   Module *M = IRB.GetInsertBlock()->getParent()->getParent();
1756   const char *UnsafeStackPtrVar = "__safestack_unsafe_stack_ptr";
1757   auto UnsafeStackPtr =
1758       dyn_cast_or_null<GlobalVariable>(M->getNamedValue(UnsafeStackPtrVar));
1759 
1760   Type *StackPtrTy = Type::getInt8PtrTy(M->getContext());
1761 
1762   if (!UnsafeStackPtr) {
1763     auto TLSModel = UseTLS ?
1764         GlobalValue::InitialExecTLSModel :
1765         GlobalValue::NotThreadLocal;
1766     // The global variable is not defined yet, define it ourselves.
1767     // We use the initial-exec TLS model because we do not support the
1768     // variable living anywhere other than in the main executable.
1769     UnsafeStackPtr = new GlobalVariable(
1770         *M, StackPtrTy, false, GlobalValue::ExternalLinkage, nullptr,
1771         UnsafeStackPtrVar, nullptr, TLSModel);
1772   } else {
1773     // The variable exists, check its type and attributes.
1774     if (UnsafeStackPtr->getValueType() != StackPtrTy)
1775       report_fatal_error(Twine(UnsafeStackPtrVar) + " must have void* type");
1776     if (UseTLS != UnsafeStackPtr->isThreadLocal())
1777       report_fatal_error(Twine(UnsafeStackPtrVar) + " must " +
1778                          (UseTLS ? "" : "not ") + "be thread-local");
1779   }
1780   return UnsafeStackPtr;
1781 }
1782 
1783 Value *TargetLoweringBase::getSafeStackPointerLocation(IRBuilder<> &IRB) const {
1784   if (!TM.getTargetTriple().isAndroid())
1785     return getDefaultSafeStackPointerLocation(IRB, true);
1786 
1787   // Android provides a libc function to retrieve the address of the current
1788   // thread's unsafe stack pointer.
1789   Module *M = IRB.GetInsertBlock()->getParent()->getParent();
1790   Type *StackPtrTy = Type::getInt8PtrTy(M->getContext());
1791   FunctionCallee Fn = M->getOrInsertFunction("__safestack_pointer_address",
1792                                              StackPtrTy->getPointerTo(0));
1793   return IRB.CreateCall(Fn);
1794 }
1795 
1796 //===----------------------------------------------------------------------===//
1797 //  Loop Strength Reduction hooks
1798 //===----------------------------------------------------------------------===//
1799 
1800 /// isLegalAddressingMode - Return true if the addressing mode represented
1801 /// by AM is legal for this target, for a load/store of the specified type.
1802 bool TargetLoweringBase::isLegalAddressingMode(const DataLayout &DL,
1803                                                const AddrMode &AM, Type *Ty,
1804                                                unsigned AS, Instruction *I) const {
1805   // The default implementation of this implements a conservative RISCy, r+r and
1806   // r+i addr mode.
1807 
1808   // Allows a sign-extended 16-bit immediate field.
1809   if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
1810     return false;
1811 
1812   // No global is ever allowed as a base.
1813   if (AM.BaseGV)
1814     return false;
1815 
1816   // Only support r+r,
1817   switch (AM.Scale) {
1818   case 0:  // "r+i" or just "i", depending on HasBaseReg.
1819     break;
1820   case 1:
1821     if (AM.HasBaseReg && AM.BaseOffs)  // "r+r+i" is not allowed.
1822       return false;
1823     // Otherwise we have r+r or r+i.
1824     break;
1825   case 2:
1826     if (AM.HasBaseReg || AM.BaseOffs)  // 2*r+r  or  2*r+i is not allowed.
1827       return false;
1828     // Allow 2*r as r+r.
1829     break;
1830   default: // Don't allow n * r
1831     return false;
1832   }
1833 
1834   return true;
1835 }
1836 
1837 //===----------------------------------------------------------------------===//
1838 //  Stack Protector
1839 //===----------------------------------------------------------------------===//
1840 
1841 // For OpenBSD return its special guard variable. Otherwise return nullptr,
1842 // so that SelectionDAG handle SSP.
1843 Value *TargetLoweringBase::getIRStackGuard(IRBuilder<> &IRB) const {
1844   if (getTargetMachine().getTargetTriple().isOSOpenBSD()) {
1845     Module &M = *IRB.GetInsertBlock()->getParent()->getParent();
1846     PointerType *PtrTy = Type::getInt8PtrTy(M.getContext());
1847     return M.getOrInsertGlobal("__guard_local", PtrTy);
1848   }
1849   return nullptr;
1850 }
1851 
1852 // Currently only support "standard" __stack_chk_guard.
1853 // TODO: add LOAD_STACK_GUARD support.
1854 void TargetLoweringBase::insertSSPDeclarations(Module &M) const {
1855   if (!M.getNamedValue("__stack_chk_guard"))
1856     new GlobalVariable(M, Type::getInt8PtrTy(M.getContext()), false,
1857                        GlobalVariable::ExternalLinkage,
1858                        nullptr, "__stack_chk_guard");
1859 }
1860 
1861 // Currently only support "standard" __stack_chk_guard.
1862 // TODO: add LOAD_STACK_GUARD support.
1863 Value *TargetLoweringBase::getSDagStackGuard(const Module &M) const {
1864   return M.getNamedValue("__stack_chk_guard");
1865 }
1866 
1867 Function *TargetLoweringBase::getSSPStackGuardCheck(const Module &M) const {
1868   return nullptr;
1869 }
1870 
1871 unsigned TargetLoweringBase::getMinimumJumpTableEntries() const {
1872   return MinimumJumpTableEntries;
1873 }
1874 
1875 void TargetLoweringBase::setMinimumJumpTableEntries(unsigned Val) {
1876   MinimumJumpTableEntries = Val;
1877 }
1878 
1879 unsigned TargetLoweringBase::getMinimumJumpTableDensity(bool OptForSize) const {
1880   return OptForSize ? OptsizeJumpTableDensity : JumpTableDensity;
1881 }
1882 
1883 unsigned TargetLoweringBase::getMaximumJumpTableSize() const {
1884   return MaximumJumpTableSize;
1885 }
1886 
1887 void TargetLoweringBase::setMaximumJumpTableSize(unsigned Val) {
1888   MaximumJumpTableSize = Val;
1889 }
1890 
1891 bool TargetLoweringBase::isJumpTableRelative() const {
1892   return getTargetMachine().isPositionIndependent();
1893 }
1894 
1895 //===----------------------------------------------------------------------===//
1896 //  Reciprocal Estimates
1897 //===----------------------------------------------------------------------===//
1898 
1899 /// Get the reciprocal estimate attribute string for a function that will
1900 /// override the target defaults.
1901 static StringRef getRecipEstimateForFunc(MachineFunction &MF) {
1902   const Function &F = MF.getFunction();
1903   return F.getFnAttribute("reciprocal-estimates").getValueAsString();
1904 }
1905 
1906 /// Construct a string for the given reciprocal operation of the given type.
1907 /// This string should match the corresponding option to the front-end's
1908 /// "-mrecip" flag assuming those strings have been passed through in an
1909 /// attribute string. For example, "vec-divf" for a division of a vXf32.
1910 static std::string getReciprocalOpName(bool IsSqrt, EVT VT) {
1911   std::string Name = VT.isVector() ? "vec-" : "";
1912 
1913   Name += IsSqrt ? "sqrt" : "div";
1914 
1915   // TODO: Handle "half" or other float types?
1916   if (VT.getScalarType() == MVT::f64) {
1917     Name += "d";
1918   } else {
1919     assert(VT.getScalarType() == MVT::f32 &&
1920            "Unexpected FP type for reciprocal estimate");
1921     Name += "f";
1922   }
1923 
1924   return Name;
1925 }
1926 
1927 /// Return the character position and value (a single numeric character) of a
1928 /// customized refinement operation in the input string if it exists. Return
1929 /// false if there is no customized refinement step count.
1930 static bool parseRefinementStep(StringRef In, size_t &Position,
1931                                 uint8_t &Value) {
1932   const char RefStepToken = ':';
1933   Position = In.find(RefStepToken);
1934   if (Position == StringRef::npos)
1935     return false;
1936 
1937   StringRef RefStepString = In.substr(Position + 1);
1938   // Allow exactly one numeric character for the additional refinement
1939   // step parameter.
1940   if (RefStepString.size() == 1) {
1941     char RefStepChar = RefStepString[0];
1942     if (RefStepChar >= '0' && RefStepChar <= '9') {
1943       Value = RefStepChar - '0';
1944       return true;
1945     }
1946   }
1947   report_fatal_error("Invalid refinement step for -recip.");
1948 }
1949 
1950 /// For the input attribute string, return one of the ReciprocalEstimate enum
1951 /// status values (enabled, disabled, or not specified) for this operation on
1952 /// the specified data type.
1953 static int getOpEnabled(bool IsSqrt, EVT VT, StringRef Override) {
1954   if (Override.empty())
1955     return TargetLoweringBase::ReciprocalEstimate::Unspecified;
1956 
1957   SmallVector<StringRef, 4> OverrideVector;
1958   Override.split(OverrideVector, ',');
1959   unsigned NumArgs = OverrideVector.size();
1960 
1961   // Check if "all", "none", or "default" was specified.
1962   if (NumArgs == 1) {
1963     // Look for an optional setting of the number of refinement steps needed
1964     // for this type of reciprocal operation.
1965     size_t RefPos;
1966     uint8_t RefSteps;
1967     if (parseRefinementStep(Override, RefPos, RefSteps)) {
1968       // Split the string for further processing.
1969       Override = Override.substr(0, RefPos);
1970     }
1971 
1972     // All reciprocal types are enabled.
1973     if (Override == "all")
1974       return TargetLoweringBase::ReciprocalEstimate::Enabled;
1975 
1976     // All reciprocal types are disabled.
1977     if (Override == "none")
1978       return TargetLoweringBase::ReciprocalEstimate::Disabled;
1979 
1980     // Target defaults for enablement are used.
1981     if (Override == "default")
1982       return TargetLoweringBase::ReciprocalEstimate::Unspecified;
1983   }
1984 
1985   // The attribute string may omit the size suffix ('f'/'d').
1986   std::string VTName = getReciprocalOpName(IsSqrt, VT);
1987   std::string VTNameNoSize = VTName;
1988   VTNameNoSize.pop_back();
1989   static const char DisabledPrefix = '!';
1990 
1991   for (StringRef RecipType : OverrideVector) {
1992     size_t RefPos;
1993     uint8_t RefSteps;
1994     if (parseRefinementStep(RecipType, RefPos, RefSteps))
1995       RecipType = RecipType.substr(0, RefPos);
1996 
1997     // Ignore the disablement token for string matching.
1998     bool IsDisabled = RecipType[0] == DisabledPrefix;
1999     if (IsDisabled)
2000       RecipType = RecipType.substr(1);
2001 
2002     if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize))
2003       return IsDisabled ? TargetLoweringBase::ReciprocalEstimate::Disabled
2004                         : TargetLoweringBase::ReciprocalEstimate::Enabled;
2005   }
2006 
2007   return TargetLoweringBase::ReciprocalEstimate::Unspecified;
2008 }
2009 
2010 /// For the input attribute string, return the customized refinement step count
2011 /// for this operation on the specified data type. If the step count does not
2012 /// exist, return the ReciprocalEstimate enum value for unspecified.
2013 static int getOpRefinementSteps(bool IsSqrt, EVT VT, StringRef Override) {
2014   if (Override.empty())
2015     return TargetLoweringBase::ReciprocalEstimate::Unspecified;
2016 
2017   SmallVector<StringRef, 4> OverrideVector;
2018   Override.split(OverrideVector, ',');
2019   unsigned NumArgs = OverrideVector.size();
2020 
2021   // Check if "all", "default", or "none" was specified.
2022   if (NumArgs == 1) {
2023     // Look for an optional setting of the number of refinement steps needed
2024     // for this type of reciprocal operation.
2025     size_t RefPos;
2026     uint8_t RefSteps;
2027     if (!parseRefinementStep(Override, RefPos, RefSteps))
2028       return TargetLoweringBase::ReciprocalEstimate::Unspecified;
2029 
2030     // Split the string for further processing.
2031     Override = Override.substr(0, RefPos);
2032     assert(Override != "none" &&
2033            "Disabled reciprocals, but specifed refinement steps?");
2034 
2035     // If this is a general override, return the specified number of steps.
2036     if (Override == "all" || Override == "default")
2037       return RefSteps;
2038   }
2039 
2040   // The attribute string may omit the size suffix ('f'/'d').
2041   std::string VTName = getReciprocalOpName(IsSqrt, VT);
2042   std::string VTNameNoSize = VTName;
2043   VTNameNoSize.pop_back();
2044 
2045   for (StringRef RecipType : OverrideVector) {
2046     size_t RefPos;
2047     uint8_t RefSteps;
2048     if (!parseRefinementStep(RecipType, RefPos, RefSteps))
2049       continue;
2050 
2051     RecipType = RecipType.substr(0, RefPos);
2052     if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize))
2053       return RefSteps;
2054   }
2055 
2056   return TargetLoweringBase::ReciprocalEstimate::Unspecified;
2057 }
2058 
2059 int TargetLoweringBase::getRecipEstimateSqrtEnabled(EVT VT,
2060                                                     MachineFunction &MF) const {
2061   return getOpEnabled(true, VT, getRecipEstimateForFunc(MF));
2062 }
2063 
2064 int TargetLoweringBase::getRecipEstimateDivEnabled(EVT VT,
2065                                                    MachineFunction &MF) const {
2066   return getOpEnabled(false, VT, getRecipEstimateForFunc(MF));
2067 }
2068 
2069 int TargetLoweringBase::getSqrtRefinementSteps(EVT VT,
2070                                                MachineFunction &MF) const {
2071   return getOpRefinementSteps(true, VT, getRecipEstimateForFunc(MF));
2072 }
2073 
2074 int TargetLoweringBase::getDivRefinementSteps(EVT VT,
2075                                               MachineFunction &MF) const {
2076   return getOpRefinementSteps(false, VT, getRecipEstimateForFunc(MF));
2077 }
2078 
2079 void TargetLoweringBase::finalizeLowering(MachineFunction &MF) const {
2080   MF.getRegInfo().freezeReservedRegs(MF);
2081 }
2082 
2083 MachineMemOperand::Flags
2084 TargetLoweringBase::getLoadMemOperandFlags(const LoadInst &LI,
2085                                            const DataLayout &DL) const {
2086   MachineMemOperand::Flags Flags = MachineMemOperand::MOLoad;
2087   if (LI.isVolatile())
2088     Flags |= MachineMemOperand::MOVolatile;
2089 
2090   if (LI.hasMetadata(LLVMContext::MD_nontemporal))
2091     Flags |= MachineMemOperand::MONonTemporal;
2092 
2093   if (LI.hasMetadata(LLVMContext::MD_invariant_load))
2094     Flags |= MachineMemOperand::MOInvariant;
2095 
2096   if (isDereferenceablePointer(LI.getPointerOperand(), LI.getType(), DL))
2097     Flags |= MachineMemOperand::MODereferenceable;
2098 
2099   Flags |= getTargetMMOFlags(LI);
2100   return Flags;
2101 }
2102 
2103 MachineMemOperand::Flags
2104 TargetLoweringBase::getStoreMemOperandFlags(const StoreInst &SI,
2105                                             const DataLayout &DL) const {
2106   MachineMemOperand::Flags Flags = MachineMemOperand::MOStore;
2107 
2108   if (SI.isVolatile())
2109     Flags |= MachineMemOperand::MOVolatile;
2110 
2111   if (SI.hasMetadata(LLVMContext::MD_nontemporal))
2112     Flags |= MachineMemOperand::MONonTemporal;
2113 
2114   // FIXME: Not preserving dereferenceable
2115   Flags |= getTargetMMOFlags(SI);
2116   return Flags;
2117 }
2118 
2119 MachineMemOperand::Flags
2120 TargetLoweringBase::getAtomicMemOperandFlags(const Instruction &AI,
2121                                              const DataLayout &DL) const {
2122   auto Flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore;
2123 
2124   if (const AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(&AI)) {
2125     if (RMW->isVolatile())
2126       Flags |= MachineMemOperand::MOVolatile;
2127   } else if (const AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(&AI)) {
2128     if (CmpX->isVolatile())
2129       Flags |= MachineMemOperand::MOVolatile;
2130   } else
2131     llvm_unreachable("not an atomic instruction");
2132 
2133   // FIXME: Not preserving dereferenceable
2134   Flags |= getTargetMMOFlags(AI);
2135   return Flags;
2136 }
2137 
2138 //===----------------------------------------------------------------------===//
2139 //  GlobalISel Hooks
2140 //===----------------------------------------------------------------------===//
2141 
2142 bool TargetLoweringBase::shouldLocalize(const MachineInstr &MI,
2143                                         const TargetTransformInfo *TTI) const {
2144   auto &MF = *MI.getMF();
2145   auto &MRI = MF.getRegInfo();
2146   // Assuming a spill and reload of a value has a cost of 1 instruction each,
2147   // this helper function computes the maximum number of uses we should consider
2148   // for remat. E.g. on arm64 global addresses take 2 insts to materialize. We
2149   // break even in terms of code size when the original MI has 2 users vs
2150   // choosing to potentially spill. Any more than 2 users we we have a net code
2151   // size increase. This doesn't take into account register pressure though.
2152   auto maxUses = [](unsigned RematCost) {
2153     // A cost of 1 means remats are basically free.
2154     if (RematCost == 1)
2155       return UINT_MAX;
2156     if (RematCost == 2)
2157       return 2U;
2158 
2159     // Remat is too expensive, only sink if there's one user.
2160     if (RematCost > 2)
2161       return 1U;
2162     llvm_unreachable("Unexpected remat cost");
2163   };
2164 
2165   // Helper to walk through uses and terminate if we've reached a limit. Saves
2166   // us spending time traversing uses if all we want to know is if it's >= min.
2167   auto isUsesAtMost = [&](unsigned Reg, unsigned MaxUses) {
2168     unsigned NumUses = 0;
2169     auto UI = MRI.use_instr_nodbg_begin(Reg), UE = MRI.use_instr_nodbg_end();
2170     for (; UI != UE && NumUses < MaxUses; ++UI) {
2171       NumUses++;
2172     }
2173     // If we haven't reached the end yet then there are more than MaxUses users.
2174     return UI == UE;
2175   };
2176 
2177   switch (MI.getOpcode()) {
2178   default:
2179     return false;
2180   // Constants-like instructions should be close to their users.
2181   // We don't want long live-ranges for them.
2182   case TargetOpcode::G_CONSTANT:
2183   case TargetOpcode::G_FCONSTANT:
2184   case TargetOpcode::G_FRAME_INDEX:
2185   case TargetOpcode::G_INTTOPTR:
2186     return true;
2187   case TargetOpcode::G_GLOBAL_VALUE: {
2188     unsigned RematCost = TTI->getGISelRematGlobalCost();
2189     Register Reg = MI.getOperand(0).getReg();
2190     unsigned MaxUses = maxUses(RematCost);
2191     if (MaxUses == UINT_MAX)
2192       return true; // Remats are "free" so always localize.
2193     bool B = isUsesAtMost(Reg, MaxUses);
2194     return B;
2195   }
2196   }
2197 }
2198