1 //===- TargetLoweringBase.cpp - Implement the TargetLoweringBase class ----===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This implements the TargetLoweringBase class.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/ADT/BitVector.h"
15 #include "llvm/ADT/STLExtras.h"
16 #include "llvm/ADT/SmallVector.h"
17 #include "llvm/ADT/StringExtras.h"
18 #include "llvm/ADT/StringRef.h"
19 #include "llvm/ADT/Triple.h"
20 #include "llvm/ADT/Twine.h"
21 #include "llvm/CodeGen/Analysis.h"
22 #include "llvm/CodeGen/ISDOpcodes.h"
23 #include "llvm/CodeGen/MachineBasicBlock.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineFunction.h"
26 #include "llvm/CodeGen/MachineInstr.h"
27 #include "llvm/CodeGen/MachineInstrBuilder.h"
28 #include "llvm/CodeGen/MachineMemOperand.h"
29 #include "llvm/CodeGen/MachineOperand.h"
30 #include "llvm/CodeGen/MachineRegisterInfo.h"
31 #include "llvm/CodeGen/RuntimeLibcalls.h"
32 #include "llvm/CodeGen/StackMaps.h"
33 #include "llvm/CodeGen/TargetLowering.h"
34 #include "llvm/CodeGen/TargetOpcodes.h"
35 #include "llvm/CodeGen/TargetRegisterInfo.h"
36 #include "llvm/CodeGen/ValueTypes.h"
37 #include "llvm/IR/Attributes.h"
38 #include "llvm/IR/CallingConv.h"
39 #include "llvm/IR/DataLayout.h"
40 #include "llvm/IR/DerivedTypes.h"
41 #include "llvm/IR/Function.h"
42 #include "llvm/IR/GlobalValue.h"
43 #include "llvm/IR/GlobalVariable.h"
44 #include "llvm/IR/IRBuilder.h"
45 #include "llvm/IR/Module.h"
46 #include "llvm/IR/Type.h"
47 #include "llvm/Support/BranchProbability.h"
48 #include "llvm/Support/Casting.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/Compiler.h"
51 #include "llvm/Support/ErrorHandling.h"
52 #include "llvm/Support/MachineValueType.h"
53 #include "llvm/Support/MathExtras.h"
54 #include "llvm/Target/TargetMachine.h"
67 static cl::opt<bool> JumpIsExpensiveOverride(
68 "jump-is-expensive", cl::init(false),
69 cl::desc("Do not create extra branches to split comparison logic."),
72 static cl::opt<unsigned> MinimumJumpTableEntries
73 ("min-jump-table-entries", cl::init(4), cl::Hidden,
74 cl::desc("Set minimum number of entries to use a jump table."));
76 static cl::opt<unsigned> MaximumJumpTableSize
77 ("max-jump-table-size", cl::init(0), cl::Hidden,
78 cl::desc("Set maximum size of jump tables; zero for no limit."));
80 /// Minimum jump table density for normal functions.
81 static cl::opt<unsigned>
82 JumpTableDensity("jump-table-density", cl::init(10), cl::Hidden,
83 cl::desc("Minimum density for building a jump table in "
84 "a normal function"));
86 /// Minimum jump table density for -Os or -Oz functions.
87 static cl::opt<unsigned> OptsizeJumpTableDensity(
88 "optsize-jump-table-density", cl::init(40), cl::Hidden,
89 cl::desc("Minimum density for building a jump table in "
90 "an optsize function"));
92 static bool darwinHasSinCos(const Triple &TT) {
93 assert(TT.isOSDarwin() && "should be called with darwin triple");
94 // Don't bother with 32 bit x86.
95 if (TT.getArch() == Triple::x86)
97 // Macos < 10.9 has no sincos_stret.
99 return !TT.isMacOSXVersionLT(10, 9) && TT.isArch64Bit();
100 // iOS < 7.0 has no sincos_stret.
102 return !TT.isOSVersionLT(7, 0);
103 // Any other darwin such as WatchOS/TvOS is new enough.
107 // Although this default value is arbitrary, it is not random. It is assumed
108 // that a condition that evaluates the same way by a higher percentage than this
109 // is best represented as control flow. Therefore, the default value N should be
110 // set such that the win from N% correct executions is greater than the loss
111 // from (100 - N)% mispredicted executions for the majority of intended targets.
112 static cl::opt<int> MinPercentageForPredictableBranch(
113 "min-predictable-branch", cl::init(99),
114 cl::desc("Minimum percentage (0-100) that a condition must be either true "
115 "or false to assume that the condition is predictable"),
118 void TargetLoweringBase::InitLibcalls(const Triple &TT) {
119 #define HANDLE_LIBCALL(code, name) \
120 setLibcallName(RTLIB::code, name);
121 #include "llvm/CodeGen/RuntimeLibcalls.def"
122 #undef HANDLE_LIBCALL
123 // Initialize calling conventions to their default.
124 for (int LC = 0; LC < RTLIB::UNKNOWN_LIBCALL; ++LC)
125 setLibcallCallingConv((RTLIB::Libcall)LC, CallingConv::C);
127 // A few names are different on particular architectures or environments.
128 if (TT.isOSDarwin()) {
129 // For f16/f32 conversions, Darwin uses the standard naming scheme, instead
130 // of the gnueabi-style __gnu_*_ieee.
131 // FIXME: What about other targets?
132 setLibcallName(RTLIB::FPEXT_F16_F32, "__extendhfsf2");
133 setLibcallName(RTLIB::FPROUND_F32_F16, "__truncsfhf2");
135 // Some darwins have an optimized __bzero/bzero function.
136 switch (TT.getArch()) {
139 if (TT.isMacOSX() && !TT.isMacOSXVersionLT(10, 6))
140 setLibcallName(RTLIB::BZERO, "__bzero");
142 case Triple::aarch64:
143 setLibcallName(RTLIB::BZERO, "bzero");
149 if (darwinHasSinCos(TT)) {
150 setLibcallName(RTLIB::SINCOS_STRET_F32, "__sincosf_stret");
151 setLibcallName(RTLIB::SINCOS_STRET_F64, "__sincos_stret");
152 if (TT.isWatchABI()) {
153 setLibcallCallingConv(RTLIB::SINCOS_STRET_F32,
154 CallingConv::ARM_AAPCS_VFP);
155 setLibcallCallingConv(RTLIB::SINCOS_STRET_F64,
156 CallingConv::ARM_AAPCS_VFP);
160 setLibcallName(RTLIB::FPEXT_F16_F32, "__gnu_h2f_ieee");
161 setLibcallName(RTLIB::FPROUND_F32_F16, "__gnu_f2h_ieee");
164 if (TT.isGNUEnvironment() || TT.isOSFuchsia()) {
165 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
166 setLibcallName(RTLIB::SINCOS_F64, "sincos");
167 setLibcallName(RTLIB::SINCOS_F80, "sincosl");
168 setLibcallName(RTLIB::SINCOS_F128, "sincosl");
169 setLibcallName(RTLIB::SINCOS_PPCF128, "sincosl");
172 if (TT.isOSOpenBSD()) {
173 setLibcallName(RTLIB::STACKPROTECTOR_CHECK_FAIL, nullptr);
177 /// getFPEXT - Return the FPEXT_*_* value for the given types, or
178 /// UNKNOWN_LIBCALL if there is none.
179 RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) {
180 if (OpVT == MVT::f16) {
181 if (RetVT == MVT::f32)
182 return FPEXT_F16_F32;
183 } else if (OpVT == MVT::f32) {
184 if (RetVT == MVT::f64)
185 return FPEXT_F32_F64;
186 if (RetVT == MVT::f128)
187 return FPEXT_F32_F128;
188 if (RetVT == MVT::ppcf128)
189 return FPEXT_F32_PPCF128;
190 } else if (OpVT == MVT::f64) {
191 if (RetVT == MVT::f128)
192 return FPEXT_F64_F128;
193 else if (RetVT == MVT::ppcf128)
194 return FPEXT_F64_PPCF128;
195 } else if (OpVT == MVT::f80) {
196 if (RetVT == MVT::f128)
197 return FPEXT_F80_F128;
200 return UNKNOWN_LIBCALL;
203 /// getFPROUND - Return the FPROUND_*_* value for the given types, or
204 /// UNKNOWN_LIBCALL if there is none.
205 RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) {
206 if (RetVT == MVT::f16) {
207 if (OpVT == MVT::f32)
208 return FPROUND_F32_F16;
209 if (OpVT == MVT::f64)
210 return FPROUND_F64_F16;
211 if (OpVT == MVT::f80)
212 return FPROUND_F80_F16;
213 if (OpVT == MVT::f128)
214 return FPROUND_F128_F16;
215 if (OpVT == MVT::ppcf128)
216 return FPROUND_PPCF128_F16;
217 } else if (RetVT == MVT::f32) {
218 if (OpVT == MVT::f64)
219 return FPROUND_F64_F32;
220 if (OpVT == MVT::f80)
221 return FPROUND_F80_F32;
222 if (OpVT == MVT::f128)
223 return FPROUND_F128_F32;
224 if (OpVT == MVT::ppcf128)
225 return FPROUND_PPCF128_F32;
226 } else if (RetVT == MVT::f64) {
227 if (OpVT == MVT::f80)
228 return FPROUND_F80_F64;
229 if (OpVT == MVT::f128)
230 return FPROUND_F128_F64;
231 if (OpVT == MVT::ppcf128)
232 return FPROUND_PPCF128_F64;
233 } else if (RetVT == MVT::f80) {
234 if (OpVT == MVT::f128)
235 return FPROUND_F128_F80;
238 return UNKNOWN_LIBCALL;
241 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or
242 /// UNKNOWN_LIBCALL if there is none.
243 RTLIB::Libcall RTLIB::getFPTOSINT(EVT OpVT, EVT RetVT) {
244 if (OpVT == MVT::f32) {
245 if (RetVT == MVT::i32)
246 return FPTOSINT_F32_I32;
247 if (RetVT == MVT::i64)
248 return FPTOSINT_F32_I64;
249 if (RetVT == MVT::i128)
250 return FPTOSINT_F32_I128;
251 } else if (OpVT == MVT::f64) {
252 if (RetVT == MVT::i32)
253 return FPTOSINT_F64_I32;
254 if (RetVT == MVT::i64)
255 return FPTOSINT_F64_I64;
256 if (RetVT == MVT::i128)
257 return FPTOSINT_F64_I128;
258 } else if (OpVT == MVT::f80) {
259 if (RetVT == MVT::i32)
260 return FPTOSINT_F80_I32;
261 if (RetVT == MVT::i64)
262 return FPTOSINT_F80_I64;
263 if (RetVT == MVT::i128)
264 return FPTOSINT_F80_I128;
265 } else if (OpVT == MVT::f128) {
266 if (RetVT == MVT::i32)
267 return FPTOSINT_F128_I32;
268 if (RetVT == MVT::i64)
269 return FPTOSINT_F128_I64;
270 if (RetVT == MVT::i128)
271 return FPTOSINT_F128_I128;
272 } else if (OpVT == MVT::ppcf128) {
273 if (RetVT == MVT::i32)
274 return FPTOSINT_PPCF128_I32;
275 if (RetVT == MVT::i64)
276 return FPTOSINT_PPCF128_I64;
277 if (RetVT == MVT::i128)
278 return FPTOSINT_PPCF128_I128;
280 return UNKNOWN_LIBCALL;
283 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or
284 /// UNKNOWN_LIBCALL if there is none.
285 RTLIB::Libcall RTLIB::getFPTOUINT(EVT OpVT, EVT RetVT) {
286 if (OpVT == MVT::f32) {
287 if (RetVT == MVT::i32)
288 return FPTOUINT_F32_I32;
289 if (RetVT == MVT::i64)
290 return FPTOUINT_F32_I64;
291 if (RetVT == MVT::i128)
292 return FPTOUINT_F32_I128;
293 } else if (OpVT == MVT::f64) {
294 if (RetVT == MVT::i32)
295 return FPTOUINT_F64_I32;
296 if (RetVT == MVT::i64)
297 return FPTOUINT_F64_I64;
298 if (RetVT == MVT::i128)
299 return FPTOUINT_F64_I128;
300 } else if (OpVT == MVT::f80) {
301 if (RetVT == MVT::i32)
302 return FPTOUINT_F80_I32;
303 if (RetVT == MVT::i64)
304 return FPTOUINT_F80_I64;
305 if (RetVT == MVT::i128)
306 return FPTOUINT_F80_I128;
307 } else if (OpVT == MVT::f128) {
308 if (RetVT == MVT::i32)
309 return FPTOUINT_F128_I32;
310 if (RetVT == MVT::i64)
311 return FPTOUINT_F128_I64;
312 if (RetVT == MVT::i128)
313 return FPTOUINT_F128_I128;
314 } else if (OpVT == MVT::ppcf128) {
315 if (RetVT == MVT::i32)
316 return FPTOUINT_PPCF128_I32;
317 if (RetVT == MVT::i64)
318 return FPTOUINT_PPCF128_I64;
319 if (RetVT == MVT::i128)
320 return FPTOUINT_PPCF128_I128;
322 return UNKNOWN_LIBCALL;
325 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or
326 /// UNKNOWN_LIBCALL if there is none.
327 RTLIB::Libcall RTLIB::getSINTTOFP(EVT OpVT, EVT RetVT) {
328 if (OpVT == MVT::i32) {
329 if (RetVT == MVT::f32)
330 return SINTTOFP_I32_F32;
331 if (RetVT == MVT::f64)
332 return SINTTOFP_I32_F64;
333 if (RetVT == MVT::f80)
334 return SINTTOFP_I32_F80;
335 if (RetVT == MVT::f128)
336 return SINTTOFP_I32_F128;
337 if (RetVT == MVT::ppcf128)
338 return SINTTOFP_I32_PPCF128;
339 } else if (OpVT == MVT::i64) {
340 if (RetVT == MVT::f32)
341 return SINTTOFP_I64_F32;
342 if (RetVT == MVT::f64)
343 return SINTTOFP_I64_F64;
344 if (RetVT == MVT::f80)
345 return SINTTOFP_I64_F80;
346 if (RetVT == MVT::f128)
347 return SINTTOFP_I64_F128;
348 if (RetVT == MVT::ppcf128)
349 return SINTTOFP_I64_PPCF128;
350 } else if (OpVT == MVT::i128) {
351 if (RetVT == MVT::f32)
352 return SINTTOFP_I128_F32;
353 if (RetVT == MVT::f64)
354 return SINTTOFP_I128_F64;
355 if (RetVT == MVT::f80)
356 return SINTTOFP_I128_F80;
357 if (RetVT == MVT::f128)
358 return SINTTOFP_I128_F128;
359 if (RetVT == MVT::ppcf128)
360 return SINTTOFP_I128_PPCF128;
362 return UNKNOWN_LIBCALL;
365 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or
366 /// UNKNOWN_LIBCALL if there is none.
367 RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) {
368 if (OpVT == MVT::i32) {
369 if (RetVT == MVT::f32)
370 return UINTTOFP_I32_F32;
371 if (RetVT == MVT::f64)
372 return UINTTOFP_I32_F64;
373 if (RetVT == MVT::f80)
374 return UINTTOFP_I32_F80;
375 if (RetVT == MVT::f128)
376 return UINTTOFP_I32_F128;
377 if (RetVT == MVT::ppcf128)
378 return UINTTOFP_I32_PPCF128;
379 } else if (OpVT == MVT::i64) {
380 if (RetVT == MVT::f32)
381 return UINTTOFP_I64_F32;
382 if (RetVT == MVT::f64)
383 return UINTTOFP_I64_F64;
384 if (RetVT == MVT::f80)
385 return UINTTOFP_I64_F80;
386 if (RetVT == MVT::f128)
387 return UINTTOFP_I64_F128;
388 if (RetVT == MVT::ppcf128)
389 return UINTTOFP_I64_PPCF128;
390 } else if (OpVT == MVT::i128) {
391 if (RetVT == MVT::f32)
392 return UINTTOFP_I128_F32;
393 if (RetVT == MVT::f64)
394 return UINTTOFP_I128_F64;
395 if (RetVT == MVT::f80)
396 return UINTTOFP_I128_F80;
397 if (RetVT == MVT::f128)
398 return UINTTOFP_I128_F128;
399 if (RetVT == MVT::ppcf128)
400 return UINTTOFP_I128_PPCF128;
402 return UNKNOWN_LIBCALL;
405 RTLIB::Libcall RTLIB::getSYNC(unsigned Opc, MVT VT) {
406 #define OP_TO_LIBCALL(Name, Enum) \
408 switch (VT.SimpleTy) { \
410 return UNKNOWN_LIBCALL; \
424 OP_TO_LIBCALL(ISD::ATOMIC_SWAP, SYNC_LOCK_TEST_AND_SET)
425 OP_TO_LIBCALL(ISD::ATOMIC_CMP_SWAP, SYNC_VAL_COMPARE_AND_SWAP)
426 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_ADD, SYNC_FETCH_AND_ADD)
427 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_SUB, SYNC_FETCH_AND_SUB)
428 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_AND, SYNC_FETCH_AND_AND)
429 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_OR, SYNC_FETCH_AND_OR)
430 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_XOR, SYNC_FETCH_AND_XOR)
431 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_NAND, SYNC_FETCH_AND_NAND)
432 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MAX, SYNC_FETCH_AND_MAX)
433 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMAX, SYNC_FETCH_AND_UMAX)
434 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MIN, SYNC_FETCH_AND_MIN)
435 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMIN, SYNC_FETCH_AND_UMIN)
440 return UNKNOWN_LIBCALL;
443 RTLIB::Libcall RTLIB::getMEMCPY_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) {
444 switch (ElementSize) {
446 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_1;
448 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_2;
450 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_4;
452 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_8;
454 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_16;
456 return UNKNOWN_LIBCALL;
460 RTLIB::Libcall RTLIB::getMEMMOVE_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) {
461 switch (ElementSize) {
463 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_1;
465 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_2;
467 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_4;
469 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_8;
471 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_16;
473 return UNKNOWN_LIBCALL;
477 RTLIB::Libcall RTLIB::getMEMSET_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) {
478 switch (ElementSize) {
480 return MEMSET_ELEMENT_UNORDERED_ATOMIC_1;
482 return MEMSET_ELEMENT_UNORDERED_ATOMIC_2;
484 return MEMSET_ELEMENT_UNORDERED_ATOMIC_4;
486 return MEMSET_ELEMENT_UNORDERED_ATOMIC_8;
488 return MEMSET_ELEMENT_UNORDERED_ATOMIC_16;
490 return UNKNOWN_LIBCALL;
494 /// InitCmpLibcallCCs - Set default comparison libcall CC.
495 static void InitCmpLibcallCCs(ISD::CondCode *CCs) {
496 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL);
497 CCs[RTLIB::OEQ_F32] = ISD::SETEQ;
498 CCs[RTLIB::OEQ_F64] = ISD::SETEQ;
499 CCs[RTLIB::OEQ_F128] = ISD::SETEQ;
500 CCs[RTLIB::OEQ_PPCF128] = ISD::SETEQ;
501 CCs[RTLIB::UNE_F32] = ISD::SETNE;
502 CCs[RTLIB::UNE_F64] = ISD::SETNE;
503 CCs[RTLIB::UNE_F128] = ISD::SETNE;
504 CCs[RTLIB::UNE_PPCF128] = ISD::SETNE;
505 CCs[RTLIB::OGE_F32] = ISD::SETGE;
506 CCs[RTLIB::OGE_F64] = ISD::SETGE;
507 CCs[RTLIB::OGE_F128] = ISD::SETGE;
508 CCs[RTLIB::OGE_PPCF128] = ISD::SETGE;
509 CCs[RTLIB::OLT_F32] = ISD::SETLT;
510 CCs[RTLIB::OLT_F64] = ISD::SETLT;
511 CCs[RTLIB::OLT_F128] = ISD::SETLT;
512 CCs[RTLIB::OLT_PPCF128] = ISD::SETLT;
513 CCs[RTLIB::OLE_F32] = ISD::SETLE;
514 CCs[RTLIB::OLE_F64] = ISD::SETLE;
515 CCs[RTLIB::OLE_F128] = ISD::SETLE;
516 CCs[RTLIB::OLE_PPCF128] = ISD::SETLE;
517 CCs[RTLIB::OGT_F32] = ISD::SETGT;
518 CCs[RTLIB::OGT_F64] = ISD::SETGT;
519 CCs[RTLIB::OGT_F128] = ISD::SETGT;
520 CCs[RTLIB::OGT_PPCF128] = ISD::SETGT;
521 CCs[RTLIB::UO_F32] = ISD::SETNE;
522 CCs[RTLIB::UO_F64] = ISD::SETNE;
523 CCs[RTLIB::UO_F128] = ISD::SETNE;
524 CCs[RTLIB::UO_PPCF128] = ISD::SETNE;
525 CCs[RTLIB::O_F32] = ISD::SETEQ;
526 CCs[RTLIB::O_F64] = ISD::SETEQ;
527 CCs[RTLIB::O_F128] = ISD::SETEQ;
528 CCs[RTLIB::O_PPCF128] = ISD::SETEQ;
531 /// NOTE: The TargetMachine owns TLOF.
532 TargetLoweringBase::TargetLoweringBase(const TargetMachine &tm) : TM(tm) {
535 // Perform these initializations only once.
536 MaxStoresPerMemset = MaxStoresPerMemcpy = MaxStoresPerMemmove =
537 MaxLoadsPerMemcmp = 8;
538 MaxGluedStoresPerMemcpy = 0;
539 MaxStoresPerMemsetOptSize = MaxStoresPerMemcpyOptSize =
540 MaxStoresPerMemmoveOptSize = MaxLoadsPerMemcmpOptSize = 4;
541 UseUnderscoreSetJmp = false;
542 UseUnderscoreLongJmp = false;
543 HasMultipleConditionRegisters = false;
544 HasExtractBitsInsn = false;
545 JumpIsExpensive = JumpIsExpensiveOverride;
546 PredictableSelectIsExpensive = false;
547 EnableExtLdPromotion = false;
548 HasFloatingPointExceptions = true;
549 StackPointerRegisterToSaveRestore = 0;
550 BooleanContents = UndefinedBooleanContent;
551 BooleanFloatContents = UndefinedBooleanContent;
552 BooleanVectorContents = UndefinedBooleanContent;
553 SchedPreferenceInfo = Sched::ILP;
555 JumpBufAlignment = 0;
556 MinFunctionAlignment = 0;
557 PrefFunctionAlignment = 0;
558 PrefLoopAlignment = 0;
559 GatherAllAliasesMaxDepth = 18;
560 MinStackArgumentAlignment = 1;
561 // TODO: the default will be switched to 0 in the next commit, along
562 // with the Target-specific changes necessary.
563 MaxAtomicSizeInBitsSupported = 1024;
565 MinCmpXchgSizeInBits = 0;
566 SupportsUnalignedAtomics = false;
568 std::fill(std::begin(LibcallRoutineNames), std::end(LibcallRoutineNames), nullptr);
570 InitLibcalls(TM.getTargetTriple());
571 InitCmpLibcallCCs(CmpLibcallCCs);
574 void TargetLoweringBase::initActions() {
575 // All operations default to being supported.
576 memset(OpActions, 0, sizeof(OpActions));
577 memset(LoadExtActions, 0, sizeof(LoadExtActions));
578 memset(TruncStoreActions, 0, sizeof(TruncStoreActions));
579 memset(IndexedModeActions, 0, sizeof(IndexedModeActions));
580 memset(CondCodeActions, 0, sizeof(CondCodeActions));
581 std::fill(std::begin(RegClassForVT), std::end(RegClassForVT), nullptr);
582 std::fill(std::begin(TargetDAGCombineArray),
583 std::end(TargetDAGCombineArray), 0);
585 // Set default actions for various operations.
586 for (MVT VT : MVT::all_valuetypes()) {
587 // Default all indexed load / store to expand.
588 for (unsigned IM = (unsigned)ISD::PRE_INC;
589 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) {
590 setIndexedLoadAction(IM, VT, Expand);
591 setIndexedStoreAction(IM, VT, Expand);
594 // Most backends expect to see the node which just returns the value loaded.
595 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Expand);
597 // These operations default to expand.
598 setOperationAction(ISD::FGETSIGN, VT, Expand);
599 setOperationAction(ISD::CONCAT_VECTORS, VT, Expand);
600 setOperationAction(ISD::FMINNUM, VT, Expand);
601 setOperationAction(ISD::FMAXNUM, VT, Expand);
602 setOperationAction(ISD::FMINNAN, VT, Expand);
603 setOperationAction(ISD::FMAXNAN, VT, Expand);
604 setOperationAction(ISD::FMAD, VT, Expand);
605 setOperationAction(ISD::SMIN, VT, Expand);
606 setOperationAction(ISD::SMAX, VT, Expand);
607 setOperationAction(ISD::UMIN, VT, Expand);
608 setOperationAction(ISD::UMAX, VT, Expand);
609 setOperationAction(ISD::ABS, VT, Expand);
611 // Overflow operations default to expand
612 setOperationAction(ISD::SADDO, VT, Expand);
613 setOperationAction(ISD::SSUBO, VT, Expand);
614 setOperationAction(ISD::UADDO, VT, Expand);
615 setOperationAction(ISD::USUBO, VT, Expand);
616 setOperationAction(ISD::SMULO, VT, Expand);
617 setOperationAction(ISD::UMULO, VT, Expand);
619 // ADDCARRY operations default to expand
620 setOperationAction(ISD::ADDCARRY, VT, Expand);
621 setOperationAction(ISD::SUBCARRY, VT, Expand);
622 setOperationAction(ISD::SETCCCARRY, VT, Expand);
624 // ADDC/ADDE/SUBC/SUBE default to expand.
625 setOperationAction(ISD::ADDC, VT, Expand);
626 setOperationAction(ISD::ADDE, VT, Expand);
627 setOperationAction(ISD::SUBC, VT, Expand);
628 setOperationAction(ISD::SUBE, VT, Expand);
630 // These default to Expand so they will be expanded to CTLZ/CTTZ by default.
631 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
632 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
634 setOperationAction(ISD::BITREVERSE, VT, Expand);
636 // These library functions default to expand.
637 setOperationAction(ISD::FROUND, VT, Expand);
638 setOperationAction(ISD::FPOWI, VT, Expand);
640 // These operations default to expand for vector types.
642 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
643 setOperationAction(ISD::ANY_EXTEND_VECTOR_INREG, VT, Expand);
644 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, VT, Expand);
645 setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, VT, Expand);
648 // For most targets @llvm.get.dynamic.area.offset just returns 0.
649 setOperationAction(ISD::GET_DYNAMIC_AREA_OFFSET, VT, Expand);
652 // Most targets ignore the @llvm.prefetch intrinsic.
653 setOperationAction(ISD::PREFETCH, MVT::Other, Expand);
655 // Most targets also ignore the @llvm.readcyclecounter intrinsic.
656 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Expand);
658 // ConstantFP nodes default to expand. Targets can either change this to
659 // Legal, in which case all fp constants are legal, or use isFPImmLegal()
660 // to optimize expansions for certain constants.
661 setOperationAction(ISD::ConstantFP, MVT::f16, Expand);
662 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
663 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
664 setOperationAction(ISD::ConstantFP, MVT::f80, Expand);
665 setOperationAction(ISD::ConstantFP, MVT::f128, Expand);
667 // These library functions default to expand.
668 for (MVT VT : {MVT::f32, MVT::f64, MVT::f128}) {
669 setOperationAction(ISD::FLOG , VT, Expand);
670 setOperationAction(ISD::FLOG2, VT, Expand);
671 setOperationAction(ISD::FLOG10, VT, Expand);
672 setOperationAction(ISD::FEXP , VT, Expand);
673 setOperationAction(ISD::FEXP2, VT, Expand);
674 setOperationAction(ISD::FFLOOR, VT, Expand);
675 setOperationAction(ISD::FNEARBYINT, VT, Expand);
676 setOperationAction(ISD::FCEIL, VT, Expand);
677 setOperationAction(ISD::FRINT, VT, Expand);
678 setOperationAction(ISD::FTRUNC, VT, Expand);
679 setOperationAction(ISD::FROUND, VT, Expand);
682 // Default ISD::TRAP to expand (which turns it into abort).
683 setOperationAction(ISD::TRAP, MVT::Other, Expand);
685 // On most systems, DEBUGTRAP and TRAP have no difference. The "Expand"
686 // here is to inform DAG Legalizer to replace DEBUGTRAP with TRAP.
687 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Expand);
690 MVT TargetLoweringBase::getScalarShiftAmountTy(const DataLayout &DL,
692 return MVT::getIntegerVT(8 * DL.getPointerSize(0));
695 EVT TargetLoweringBase::getShiftAmountTy(EVT LHSTy, const DataLayout &DL,
696 bool LegalTypes) const {
697 assert(LHSTy.isInteger() && "Shift amount is not an integer type!");
698 if (LHSTy.isVector())
700 return LegalTypes ? getScalarShiftAmountTy(DL, LHSTy)
704 bool TargetLoweringBase::canOpTrap(unsigned Op, EVT VT) const {
705 assert(isTypeLegal(VT));
717 void TargetLoweringBase::setJumpIsExpensive(bool isExpensive) {
718 // If the command-line option was specified, ignore this request.
719 if (!JumpIsExpensiveOverride.getNumOccurrences())
720 JumpIsExpensive = isExpensive;
723 TargetLoweringBase::LegalizeKind
724 TargetLoweringBase::getTypeConversion(LLVMContext &Context, EVT VT) const {
725 // If this is a simple type, use the ComputeRegisterProp mechanism.
727 MVT SVT = VT.getSimpleVT();
728 assert((unsigned)SVT.SimpleTy < array_lengthof(TransformToType));
729 MVT NVT = TransformToType[SVT.SimpleTy];
730 LegalizeTypeAction LA = ValueTypeActions.getTypeAction(SVT);
732 assert((LA == TypeLegal || LA == TypeSoftenFloat ||
733 ValueTypeActions.getTypeAction(NVT) != TypePromoteInteger) &&
734 "Promote may not follow Expand or Promote");
736 if (LA == TypeSplitVector)
737 return LegalizeKind(LA,
738 EVT::getVectorVT(Context, SVT.getVectorElementType(),
739 SVT.getVectorNumElements() / 2));
740 if (LA == TypeScalarizeVector)
741 return LegalizeKind(LA, SVT.getVectorElementType());
742 return LegalizeKind(LA, NVT);
745 // Handle Extended Scalar Types.
746 if (!VT.isVector()) {
747 assert(VT.isInteger() && "Float types must be simple");
748 unsigned BitSize = VT.getSizeInBits();
749 // First promote to a power-of-two size, then expand if necessary.
750 if (BitSize < 8 || !isPowerOf2_32(BitSize)) {
751 EVT NVT = VT.getRoundIntegerType(Context);
752 assert(NVT != VT && "Unable to round integer VT");
753 LegalizeKind NextStep = getTypeConversion(Context, NVT);
754 // Avoid multi-step promotion.
755 if (NextStep.first == TypePromoteInteger)
757 // Return rounded integer type.
758 return LegalizeKind(TypePromoteInteger, NVT);
761 return LegalizeKind(TypeExpandInteger,
762 EVT::getIntegerVT(Context, VT.getSizeInBits() / 2));
765 // Handle vector types.
766 unsigned NumElts = VT.getVectorNumElements();
767 EVT EltVT = VT.getVectorElementType();
769 // Vectors with only one element are always scalarized.
771 return LegalizeKind(TypeScalarizeVector, EltVT);
773 // Try to widen vector elements until the element type is a power of two and
774 // promote it to a legal type later on, for example:
775 // <3 x i8> -> <4 x i8> -> <4 x i32>
776 if (EltVT.isInteger()) {
777 // Vectors with a number of elements that is not a power of two are always
778 // widened, for example <3 x i8> -> <4 x i8>.
779 if (!VT.isPow2VectorType()) {
780 NumElts = (unsigned)NextPowerOf2(NumElts);
781 EVT NVT = EVT::getVectorVT(Context, EltVT, NumElts);
782 return LegalizeKind(TypeWidenVector, NVT);
785 // Examine the element type.
786 LegalizeKind LK = getTypeConversion(Context, EltVT);
788 // If type is to be expanded, split the vector.
789 // <4 x i140> -> <2 x i140>
790 if (LK.first == TypeExpandInteger)
791 return LegalizeKind(TypeSplitVector,
792 EVT::getVectorVT(Context, EltVT, NumElts / 2));
794 // Promote the integer element types until a legal vector type is found
795 // or until the element integer type is too big. If a legal type was not
796 // found, fallback to the usual mechanism of widening/splitting the
798 EVT OldEltVT = EltVT;
800 // Increase the bitwidth of the element to the next pow-of-two
801 // (which is greater than 8 bits).
802 EltVT = EVT::getIntegerVT(Context, 1 + EltVT.getSizeInBits())
803 .getRoundIntegerType(Context);
805 // Stop trying when getting a non-simple element type.
806 // Note that vector elements may be greater than legal vector element
807 // types. Example: X86 XMM registers hold 64bit element on 32bit
809 if (!EltVT.isSimple())
812 // Build a new vector type and check if it is legal.
813 MVT NVT = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts);
814 // Found a legal promoted vector type.
815 if (NVT != MVT() && ValueTypeActions.getTypeAction(NVT) == TypeLegal)
816 return LegalizeKind(TypePromoteInteger,
817 EVT::getVectorVT(Context, EltVT, NumElts));
820 // Reset the type to the unexpanded type if we did not find a legal vector
821 // type with a promoted vector element type.
825 // Try to widen the vector until a legal type is found.
826 // If there is no wider legal type, split the vector.
828 // Round up to the next power of 2.
829 NumElts = (unsigned)NextPowerOf2(NumElts);
831 // If there is no simple vector type with this many elements then there
832 // cannot be a larger legal vector type. Note that this assumes that
833 // there are no skipped intermediate vector types in the simple types.
834 if (!EltVT.isSimple())
836 MVT LargerVector = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts);
837 if (LargerVector == MVT())
840 // If this type is legal then widen the vector.
841 if (ValueTypeActions.getTypeAction(LargerVector) == TypeLegal)
842 return LegalizeKind(TypeWidenVector, LargerVector);
845 // Widen odd vectors to next power of two.
846 if (!VT.isPow2VectorType()) {
847 EVT NVT = VT.getPow2VectorType(Context);
848 return LegalizeKind(TypeWidenVector, NVT);
851 // Vectors with illegal element types are expanded.
852 EVT NVT = EVT::getVectorVT(Context, EltVT, VT.getVectorNumElements() / 2);
853 return LegalizeKind(TypeSplitVector, NVT);
856 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT,
857 unsigned &NumIntermediates,
859 TargetLoweringBase *TLI) {
860 // Figure out the right, legal destination reg to copy into.
861 unsigned NumElts = VT.getVectorNumElements();
862 MVT EltTy = VT.getVectorElementType();
864 unsigned NumVectorRegs = 1;
866 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
867 // could break down into LHS/RHS like LegalizeDAG does.
868 if (!isPowerOf2_32(NumElts)) {
869 NumVectorRegs = NumElts;
873 // Divide the input until we get to a supported size. This will always
874 // end with a scalar if the target doesn't support vectors.
875 while (NumElts > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) {
880 NumIntermediates = NumVectorRegs;
882 MVT NewVT = MVT::getVectorVT(EltTy, NumElts);
883 if (!TLI->isTypeLegal(NewVT))
885 IntermediateVT = NewVT;
887 unsigned NewVTSize = NewVT.getSizeInBits();
889 // Convert sizes such as i33 to i64.
890 if (!isPowerOf2_32(NewVTSize))
891 NewVTSize = NextPowerOf2(NewVTSize);
893 MVT DestVT = TLI->getRegisterType(NewVT);
895 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16.
896 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits());
898 // Otherwise, promotion or legal types use the same number of registers as
899 // the vector decimated to the appropriate level.
900 return NumVectorRegs;
903 /// isLegalRC - Return true if the value types that can be represented by the
904 /// specified register class are all legal.
905 bool TargetLoweringBase::isLegalRC(const TargetRegisterInfo &TRI,
906 const TargetRegisterClass &RC) const {
907 for (auto I = TRI.legalclasstypes_begin(RC); *I != MVT::Other; ++I)
913 /// Replace/modify any TargetFrameIndex operands with a targte-dependent
914 /// sequence of memory operands that is recognized by PrologEpilogInserter.
916 TargetLoweringBase::emitPatchPoint(MachineInstr &InitialMI,
917 MachineBasicBlock *MBB) const {
918 MachineInstr *MI = &InitialMI;
919 MachineFunction &MF = *MI->getMF();
920 MachineFrameInfo &MFI = MF.getFrameInfo();
922 // We're handling multiple types of operands here:
923 // PATCHPOINT MetaArgs - live-in, read only, direct
924 // STATEPOINT Deopt Spill - live-through, read only, indirect
925 // STATEPOINT Deopt Alloca - live-through, read only, direct
926 // (We're currently conservative and mark the deopt slots read/write in
928 // STATEPOINT GC Spill - live-through, read/write, indirect
929 // STATEPOINT GC Alloca - live-through, read/write, direct
930 // The live-in vs live-through is handled already (the live through ones are
931 // all stack slots), but we need to handle the different type of stackmap
932 // operands and memory effects here.
934 // MI changes inside this loop as we grow operands.
935 for(unsigned OperIdx = 0; OperIdx != MI->getNumOperands(); ++OperIdx) {
936 MachineOperand &MO = MI->getOperand(OperIdx);
940 // foldMemoryOperand builds a new MI after replacing a single FI operand
941 // with the canonical set of five x86 addressing-mode operands.
942 int FI = MO.getIndex();
943 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), MI->getDesc());
945 // Copy operands before the frame-index.
946 for (unsigned i = 0; i < OperIdx; ++i)
947 MIB.add(MI->getOperand(i));
948 // Add frame index operands recognized by stackmaps.cpp
949 if (MFI.isStatepointSpillSlotObjectIndex(FI)) {
950 // indirect-mem-ref tag, size, #FI, offset.
951 // Used for spills inserted by StatepointLowering. This codepath is not
952 // used for patchpoints/stackmaps at all, for these spilling is done via
953 // foldMemoryOperand callback only.
954 assert(MI->getOpcode() == TargetOpcode::STATEPOINT && "sanity");
955 MIB.addImm(StackMaps::IndirectMemRefOp);
956 MIB.addImm(MFI.getObjectSize(FI));
957 MIB.add(MI->getOperand(OperIdx));
960 // direct-mem-ref tag, #FI, offset.
961 // Used by patchpoint, and direct alloca arguments to statepoints
962 MIB.addImm(StackMaps::DirectMemRefOp);
963 MIB.add(MI->getOperand(OperIdx));
966 // Copy the operands after the frame index.
967 for (unsigned i = OperIdx + 1; i != MI->getNumOperands(); ++i)
968 MIB.add(MI->getOperand(i));
970 // Inherit previous memory operands.
971 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
972 assert(MIB->mayLoad() && "Folded a stackmap use to a non-load!");
974 // Add a new memory operand for this FI.
975 assert(MFI.getObjectOffset(FI) != -1);
977 auto Flags = MachineMemOperand::MOLoad;
978 if (MI->getOpcode() == TargetOpcode::STATEPOINT) {
979 Flags |= MachineMemOperand::MOStore;
980 Flags |= MachineMemOperand::MOVolatile;
982 MachineMemOperand *MMO = MF.getMachineMemOperand(
983 MachinePointerInfo::getFixedStack(MF, FI), Flags,
984 MF.getDataLayout().getPointerSize(), MFI.getObjectAlignment(FI));
985 MIB->addMemOperand(MF, MMO);
987 // Replace the instruction and update the operand index.
988 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
989 OperIdx += (MIB->getNumOperands() - MI->getNumOperands()) - 1;
990 MI->eraseFromParent();
997 TargetLoweringBase::emitXRayCustomEvent(MachineInstr &MI,
998 MachineBasicBlock *MBB) const {
999 assert(MI.getOpcode() == TargetOpcode::PATCHABLE_EVENT_CALL &&
1000 "Called emitXRayCustomEvent on the wrong MI!");
1001 auto &MF = *MI.getMF();
1002 auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc());
1003 for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx)
1004 MIB.add(MI.getOperand(OpIdx));
1006 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
1007 MI.eraseFromParent();
1012 TargetLoweringBase::emitXRayTypedEvent(MachineInstr &MI,
1013 MachineBasicBlock *MBB) const {
1014 assert(MI.getOpcode() == TargetOpcode::PATCHABLE_TYPED_EVENT_CALL &&
1015 "Called emitXRayTypedEvent on the wrong MI!");
1016 auto &MF = *MI.getMF();
1017 auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc());
1018 for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx)
1019 MIB.add(MI.getOperand(OpIdx));
1021 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
1022 MI.eraseFromParent();
1026 /// findRepresentativeClass - Return the largest legal super-reg register class
1027 /// of the register class for the specified type and its associated "cost".
1028 // This function is in TargetLowering because it uses RegClassForVT which would
1029 // need to be moved to TargetRegisterInfo and would necessitate moving
1030 // isTypeLegal over as well - a massive change that would just require
1031 // TargetLowering having a TargetRegisterInfo class member that it would use.
1032 std::pair<const TargetRegisterClass *, uint8_t>
1033 TargetLoweringBase::findRepresentativeClass(const TargetRegisterInfo *TRI,
1035 const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy];
1037 return std::make_pair(RC, 0);
1039 // Compute the set of all super-register classes.
1040 BitVector SuperRegRC(TRI->getNumRegClasses());
1041 for (SuperRegClassIterator RCI(RC, TRI); RCI.isValid(); ++RCI)
1042 SuperRegRC.setBitsInMask(RCI.getMask());
1044 // Find the first legal register class with the largest spill size.
1045 const TargetRegisterClass *BestRC = RC;
1046 for (unsigned i : SuperRegRC.set_bits()) {
1047 const TargetRegisterClass *SuperRC = TRI->getRegClass(i);
1048 // We want the largest possible spill size.
1049 if (TRI->getSpillSize(*SuperRC) <= TRI->getSpillSize(*BestRC))
1051 if (!isLegalRC(*TRI, *SuperRC))
1055 return std::make_pair(BestRC, 1);
1058 /// computeRegisterProperties - Once all of the register classes are added,
1059 /// this allows us to compute derived properties we expose.
1060 void TargetLoweringBase::computeRegisterProperties(
1061 const TargetRegisterInfo *TRI) {
1062 static_assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE,
1063 "Too many value types for ValueTypeActions to hold!");
1065 // Everything defaults to needing one register.
1066 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
1067 NumRegistersForVT[i] = 1;
1068 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i;
1070 // ...except isVoid, which doesn't need any registers.
1071 NumRegistersForVT[MVT::isVoid] = 0;
1073 // Find the largest integer register class.
1074 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE;
1075 for (; RegClassForVT[LargestIntReg] == nullptr; --LargestIntReg)
1076 assert(LargestIntReg != MVT::i1 && "No integer registers defined!");
1078 // Every integer value type larger than this largest register takes twice as
1079 // many registers to represent as the previous ValueType.
1080 for (unsigned ExpandedReg = LargestIntReg + 1;
1081 ExpandedReg <= MVT::LAST_INTEGER_VALUETYPE; ++ExpandedReg) {
1082 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1];
1083 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg;
1084 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1);
1085 ValueTypeActions.setTypeAction((MVT::SimpleValueType)ExpandedReg,
1089 // Inspect all of the ValueType's smaller than the largest integer
1090 // register to see which ones need promotion.
1091 unsigned LegalIntReg = LargestIntReg;
1092 for (unsigned IntReg = LargestIntReg - 1;
1093 IntReg >= (unsigned)MVT::i1; --IntReg) {
1094 MVT IVT = (MVT::SimpleValueType)IntReg;
1095 if (isTypeLegal(IVT)) {
1096 LegalIntReg = IntReg;
1098 RegisterTypeForVT[IntReg] = TransformToType[IntReg] =
1099 (const MVT::SimpleValueType)LegalIntReg;
1100 ValueTypeActions.setTypeAction(IVT, TypePromoteInteger);
1104 // ppcf128 type is really two f64's.
1105 if (!isTypeLegal(MVT::ppcf128)) {
1106 if (isTypeLegal(MVT::f64)) {
1107 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64];
1108 RegisterTypeForVT[MVT::ppcf128] = MVT::f64;
1109 TransformToType[MVT::ppcf128] = MVT::f64;
1110 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeExpandFloat);
1112 NumRegistersForVT[MVT::ppcf128] = NumRegistersForVT[MVT::i128];
1113 RegisterTypeForVT[MVT::ppcf128] = RegisterTypeForVT[MVT::i128];
1114 TransformToType[MVT::ppcf128] = MVT::i128;
1115 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeSoftenFloat);
1119 // Decide how to handle f128. If the target does not have native f128 support,
1120 // expand it to i128 and we will be generating soft float library calls.
1121 if (!isTypeLegal(MVT::f128)) {
1122 NumRegistersForVT[MVT::f128] = NumRegistersForVT[MVT::i128];
1123 RegisterTypeForVT[MVT::f128] = RegisterTypeForVT[MVT::i128];
1124 TransformToType[MVT::f128] = MVT::i128;
1125 ValueTypeActions.setTypeAction(MVT::f128, TypeSoftenFloat);
1128 // Decide how to handle f64. If the target does not have native f64 support,
1129 // expand it to i64 and we will be generating soft float library calls.
1130 if (!isTypeLegal(MVT::f64)) {
1131 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64];
1132 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64];
1133 TransformToType[MVT::f64] = MVT::i64;
1134 ValueTypeActions.setTypeAction(MVT::f64, TypeSoftenFloat);
1137 // Decide how to handle f32. If the target does not have native f32 support,
1138 // expand it to i32 and we will be generating soft float library calls.
1139 if (!isTypeLegal(MVT::f32)) {
1140 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32];
1141 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32];
1142 TransformToType[MVT::f32] = MVT::i32;
1143 ValueTypeActions.setTypeAction(MVT::f32, TypeSoftenFloat);
1146 // Decide how to handle f16. If the target does not have native f16 support,
1147 // promote it to f32, because there are no f16 library calls (except for
1149 if (!isTypeLegal(MVT::f16)) {
1150 NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::f32];
1151 RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::f32];
1152 TransformToType[MVT::f16] = MVT::f32;
1153 ValueTypeActions.setTypeAction(MVT::f16, TypePromoteFloat);
1156 // Loop over all of the vector value types to see which need transformations.
1157 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE;
1158 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
1159 MVT VT = (MVT::SimpleValueType) i;
1160 if (isTypeLegal(VT))
1163 MVT EltVT = VT.getVectorElementType();
1164 unsigned NElts = VT.getVectorNumElements();
1165 bool IsLegalWiderType = false;
1166 LegalizeTypeAction PreferredAction = getPreferredVectorAction(VT);
1167 switch (PreferredAction) {
1168 case TypePromoteInteger:
1169 // Try to promote the elements of integer vectors. If no legal
1170 // promotion was found, fall through to the widen-vector method.
1171 for (unsigned nVT = i + 1; nVT <= MVT::LAST_INTEGER_VECTOR_VALUETYPE; ++nVT) {
1172 MVT SVT = (MVT::SimpleValueType) nVT;
1173 // Promote vectors of integers to vectors with the same number
1174 // of elements, with a wider element type.
1175 if (SVT.getScalarSizeInBits() > EltVT.getSizeInBits() &&
1176 SVT.getVectorNumElements() == NElts && isTypeLegal(SVT)) {
1177 TransformToType[i] = SVT;
1178 RegisterTypeForVT[i] = SVT;
1179 NumRegistersForVT[i] = 1;
1180 ValueTypeActions.setTypeAction(VT, TypePromoteInteger);
1181 IsLegalWiderType = true;
1185 if (IsLegalWiderType)
1189 case TypeWidenVector:
1190 // Try to widen the vector.
1191 for (unsigned nVT = i + 1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
1192 MVT SVT = (MVT::SimpleValueType) nVT;
1193 if (SVT.getVectorElementType() == EltVT
1194 && SVT.getVectorNumElements() > NElts && isTypeLegal(SVT)) {
1195 TransformToType[i] = SVT;
1196 RegisterTypeForVT[i] = SVT;
1197 NumRegistersForVT[i] = 1;
1198 ValueTypeActions.setTypeAction(VT, TypeWidenVector);
1199 IsLegalWiderType = true;
1203 if (IsLegalWiderType)
1207 case TypeSplitVector:
1208 case TypeScalarizeVector: {
1211 unsigned NumIntermediates;
1212 NumRegistersForVT[i] = getVectorTypeBreakdownMVT(VT, IntermediateVT,
1213 NumIntermediates, RegisterVT, this);
1214 RegisterTypeForVT[i] = RegisterVT;
1216 MVT NVT = VT.getPow2VectorType();
1218 // Type is already a power of 2. The default action is to split.
1219 TransformToType[i] = MVT::Other;
1220 if (PreferredAction == TypeScalarizeVector)
1221 ValueTypeActions.setTypeAction(VT, TypeScalarizeVector);
1222 else if (PreferredAction == TypeSplitVector)
1223 ValueTypeActions.setTypeAction(VT, TypeSplitVector);
1225 // Set type action according to the number of elements.
1226 ValueTypeActions.setTypeAction(VT, NElts == 1 ? TypeScalarizeVector
1229 TransformToType[i] = NVT;
1230 ValueTypeActions.setTypeAction(VT, TypeWidenVector);
1235 llvm_unreachable("Unknown vector legalization action!");
1239 // Determine the 'representative' register class for each value type.
1240 // An representative register class is the largest (meaning one which is
1241 // not a sub-register class / subreg register class) legal register class for
1242 // a group of value types. For example, on i386, i8, i16, and i32
1243 // representative would be GR32; while on x86_64 it's GR64.
1244 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
1245 const TargetRegisterClass* RRC;
1247 std::tie(RRC, Cost) = findRepresentativeClass(TRI, (MVT::SimpleValueType)i);
1248 RepRegClassForVT[i] = RRC;
1249 RepRegClassCostForVT[i] = Cost;
1253 EVT TargetLoweringBase::getSetCCResultType(const DataLayout &DL, LLVMContext &,
1255 assert(!VT.isVector() && "No default SetCC type for vectors!");
1256 return getPointerTy(DL).SimpleTy;
1259 MVT::SimpleValueType TargetLoweringBase::getCmpLibcallReturnType() const {
1260 return MVT::i32; // return the default value
1263 /// getVectorTypeBreakdown - Vector types are broken down into some number of
1264 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32
1265 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
1266 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
1268 /// This method returns the number of registers needed, and the VT for each
1269 /// register. It also returns the VT and quantity of the intermediate values
1270 /// before they are promoted/expanded.
1271 unsigned TargetLoweringBase::getVectorTypeBreakdown(LLVMContext &Context, EVT VT,
1272 EVT &IntermediateVT,
1273 unsigned &NumIntermediates,
1274 MVT &RegisterVT) const {
1275 unsigned NumElts = VT.getVectorNumElements();
1277 // If there is a wider vector type with the same element type as this one,
1278 // or a promoted vector type that has the same number of elements which
1279 // are wider, then we should convert to that legal vector type.
1280 // This handles things like <2 x float> -> <4 x float> and
1281 // <4 x i1> -> <4 x i32>.
1282 LegalizeTypeAction TA = getTypeAction(Context, VT);
1283 if (NumElts != 1 && (TA == TypeWidenVector || TA == TypePromoteInteger)) {
1284 EVT RegisterEVT = getTypeToTransformTo(Context, VT);
1285 if (isTypeLegal(RegisterEVT)) {
1286 IntermediateVT = RegisterEVT;
1287 RegisterVT = RegisterEVT.getSimpleVT();
1288 NumIntermediates = 1;
1293 // Figure out the right, legal destination reg to copy into.
1294 EVT EltTy = VT.getVectorElementType();
1296 unsigned NumVectorRegs = 1;
1298 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
1299 // could break down into LHS/RHS like LegalizeDAG does.
1300 if (!isPowerOf2_32(NumElts)) {
1301 NumVectorRegs = NumElts;
1305 // Divide the input until we get to a supported size. This will always
1306 // end with a scalar if the target doesn't support vectors.
1307 while (NumElts > 1 && !isTypeLegal(
1308 EVT::getVectorVT(Context, EltTy, NumElts))) {
1310 NumVectorRegs <<= 1;
1313 NumIntermediates = NumVectorRegs;
1315 EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts);
1316 if (!isTypeLegal(NewVT))
1318 IntermediateVT = NewVT;
1320 MVT DestVT = getRegisterType(Context, NewVT);
1321 RegisterVT = DestVT;
1322 unsigned NewVTSize = NewVT.getSizeInBits();
1324 // Convert sizes such as i33 to i64.
1325 if (!isPowerOf2_32(NewVTSize))
1326 NewVTSize = NextPowerOf2(NewVTSize);
1328 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16.
1329 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits());
1331 // Otherwise, promotion or legal types use the same number of registers as
1332 // the vector decimated to the appropriate level.
1333 return NumVectorRegs;
1336 /// Get the EVTs and ArgFlags collections that represent the legalized return
1337 /// type of the given function. This does not require a DAG or a return value,
1338 /// and is suitable for use before any DAGs for the function are constructed.
1339 /// TODO: Move this out of TargetLowering.cpp.
1340 void llvm::GetReturnInfo(Type *ReturnType, AttributeList attr,
1341 SmallVectorImpl<ISD::OutputArg> &Outs,
1342 const TargetLowering &TLI, const DataLayout &DL) {
1343 SmallVector<EVT, 4> ValueVTs;
1344 ComputeValueVTs(TLI, DL, ReturnType, ValueVTs);
1345 unsigned NumValues = ValueVTs.size();
1346 if (NumValues == 0) return;
1348 for (unsigned j = 0, f = NumValues; j != f; ++j) {
1349 EVT VT = ValueVTs[j];
1350 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
1352 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt))
1353 ExtendKind = ISD::SIGN_EXTEND;
1354 else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt))
1355 ExtendKind = ISD::ZERO_EXTEND;
1357 // FIXME: C calling convention requires the return type to be promoted to
1358 // at least 32-bit. But this is not necessary for non-C calling
1359 // conventions. The frontend should mark functions whose return values
1360 // require promoting with signext or zeroext attributes.
1361 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) {
1362 MVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32);
1363 if (VT.bitsLT(MinVT))
1368 TLI.getNumRegistersForCallingConv(ReturnType->getContext(), VT);
1370 TLI.getRegisterTypeForCallingConv(ReturnType->getContext(), VT);
1372 // 'inreg' on function refers to return value
1373 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
1374 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::InReg))
1377 // Propagate extension type if any
1378 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt))
1380 else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt))
1383 for (unsigned i = 0; i < NumParts; ++i)
1384 Outs.push_back(ISD::OutputArg(Flags, PartVT, VT, /*isFixed=*/true, 0, 0));
1388 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1389 /// function arguments in the caller parameter area. This is the actual
1390 /// alignment, not its logarithm.
1391 unsigned TargetLoweringBase::getByValTypeAlignment(Type *Ty,
1392 const DataLayout &DL) const {
1393 return DL.getABITypeAlignment(Ty);
1396 bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context,
1397 const DataLayout &DL, EVT VT,
1401 // Check if the specified alignment is sufficient based on the data layout.
1402 // TODO: While using the data layout works in practice, a better solution
1403 // would be to implement this check directly (make this a virtual function).
1404 // For example, the ABI alignment may change based on software platform while
1405 // this function should only be affected by hardware implementation.
1406 Type *Ty = VT.getTypeForEVT(Context);
1407 if (Alignment >= DL.getABITypeAlignment(Ty)) {
1408 // Assume that an access that meets the ABI-specified alignment is fast.
1409 if (Fast != nullptr)
1414 // This is a misaligned access.
1415 return allowsMisalignedMemoryAccesses(VT, AddrSpace, Alignment, Fast);
1418 BranchProbability TargetLoweringBase::getPredictableBranchThreshold() const {
1419 return BranchProbability(MinPercentageForPredictableBranch, 100);
1422 //===----------------------------------------------------------------------===//
1423 // TargetTransformInfo Helpers
1424 //===----------------------------------------------------------------------===//
1426 int TargetLoweringBase::InstructionOpcodeToISD(unsigned Opcode) const {
1427 enum InstructionOpcodes {
1428 #define HANDLE_INST(NUM, OPCODE, CLASS) OPCODE = NUM,
1429 #define LAST_OTHER_INST(NUM) InstructionOpcodesCount = NUM
1430 #include "llvm/IR/Instruction.def"
1432 switch (static_cast<InstructionOpcodes>(Opcode)) {
1435 case Switch: return 0;
1436 case IndirectBr: return 0;
1437 case Invoke: return 0;
1438 case Resume: return 0;
1439 case Unreachable: return 0;
1440 case CleanupRet: return 0;
1441 case CatchRet: return 0;
1442 case CatchPad: return 0;
1443 case CatchSwitch: return 0;
1444 case CleanupPad: return 0;
1445 case Add: return ISD::ADD;
1446 case FAdd: return ISD::FADD;
1447 case Sub: return ISD::SUB;
1448 case FSub: return ISD::FSUB;
1449 case Mul: return ISD::MUL;
1450 case FMul: return ISD::FMUL;
1451 case UDiv: return ISD::UDIV;
1452 case SDiv: return ISD::SDIV;
1453 case FDiv: return ISD::FDIV;
1454 case URem: return ISD::UREM;
1455 case SRem: return ISD::SREM;
1456 case FRem: return ISD::FREM;
1457 case Shl: return ISD::SHL;
1458 case LShr: return ISD::SRL;
1459 case AShr: return ISD::SRA;
1460 case And: return ISD::AND;
1461 case Or: return ISD::OR;
1462 case Xor: return ISD::XOR;
1463 case Alloca: return 0;
1464 case Load: return ISD::LOAD;
1465 case Store: return ISD::STORE;
1466 case GetElementPtr: return 0;
1467 case Fence: return 0;
1468 case AtomicCmpXchg: return 0;
1469 case AtomicRMW: return 0;
1470 case Trunc: return ISD::TRUNCATE;
1471 case ZExt: return ISD::ZERO_EXTEND;
1472 case SExt: return ISD::SIGN_EXTEND;
1473 case FPToUI: return ISD::FP_TO_UINT;
1474 case FPToSI: return ISD::FP_TO_SINT;
1475 case UIToFP: return ISD::UINT_TO_FP;
1476 case SIToFP: return ISD::SINT_TO_FP;
1477 case FPTrunc: return ISD::FP_ROUND;
1478 case FPExt: return ISD::FP_EXTEND;
1479 case PtrToInt: return ISD::BITCAST;
1480 case IntToPtr: return ISD::BITCAST;
1481 case BitCast: return ISD::BITCAST;
1482 case AddrSpaceCast: return ISD::ADDRSPACECAST;
1483 case ICmp: return ISD::SETCC;
1484 case FCmp: return ISD::SETCC;
1486 case Call: return 0;
1487 case Select: return ISD::SELECT;
1488 case UserOp1: return 0;
1489 case UserOp2: return 0;
1490 case VAArg: return 0;
1491 case ExtractElement: return ISD::EXTRACT_VECTOR_ELT;
1492 case InsertElement: return ISD::INSERT_VECTOR_ELT;
1493 case ShuffleVector: return ISD::VECTOR_SHUFFLE;
1494 case ExtractValue: return ISD::MERGE_VALUES;
1495 case InsertValue: return ISD::MERGE_VALUES;
1496 case LandingPad: return 0;
1499 llvm_unreachable("Unknown instruction type encountered!");
1503 TargetLoweringBase::getTypeLegalizationCost(const DataLayout &DL,
1505 LLVMContext &C = Ty->getContext();
1506 EVT MTy = getValueType(DL, Ty);
1509 // We keep legalizing the type until we find a legal kind. We assume that
1510 // the only operation that costs anything is the split. After splitting
1511 // we need to handle two types.
1513 LegalizeKind LK = getTypeConversion(C, MTy);
1515 if (LK.first == TypeLegal)
1516 return std::make_pair(Cost, MTy.getSimpleVT());
1518 if (LK.first == TypeSplitVector || LK.first == TypeExpandInteger)
1521 // Do not loop with f128 type.
1522 if (MTy == LK.second)
1523 return std::make_pair(Cost, MTy.getSimpleVT());
1525 // Keep legalizing the type.
1530 Value *TargetLoweringBase::getDefaultSafeStackPointerLocation(IRBuilder<> &IRB,
1531 bool UseTLS) const {
1532 // compiler-rt provides a variable with a magic name. Targets that do not
1533 // link with compiler-rt may also provide such a variable.
1534 Module *M = IRB.GetInsertBlock()->getParent()->getParent();
1535 const char *UnsafeStackPtrVar = "__safestack_unsafe_stack_ptr";
1536 auto UnsafeStackPtr =
1537 dyn_cast_or_null<GlobalVariable>(M->getNamedValue(UnsafeStackPtrVar));
1539 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext());
1541 if (!UnsafeStackPtr) {
1542 auto TLSModel = UseTLS ?
1543 GlobalValue::InitialExecTLSModel :
1544 GlobalValue::NotThreadLocal;
1545 // The global variable is not defined yet, define it ourselves.
1546 // We use the initial-exec TLS model because we do not support the
1547 // variable living anywhere other than in the main executable.
1548 UnsafeStackPtr = new GlobalVariable(
1549 *M, StackPtrTy, false, GlobalValue::ExternalLinkage, nullptr,
1550 UnsafeStackPtrVar, nullptr, TLSModel);
1552 // The variable exists, check its type and attributes.
1553 if (UnsafeStackPtr->getValueType() != StackPtrTy)
1554 report_fatal_error(Twine(UnsafeStackPtrVar) + " must have void* type");
1555 if (UseTLS != UnsafeStackPtr->isThreadLocal())
1556 report_fatal_error(Twine(UnsafeStackPtrVar) + " must " +
1557 (UseTLS ? "" : "not ") + "be thread-local");
1559 return UnsafeStackPtr;
1562 Value *TargetLoweringBase::getSafeStackPointerLocation(IRBuilder<> &IRB) const {
1563 if (!TM.getTargetTriple().isAndroid())
1564 return getDefaultSafeStackPointerLocation(IRB, true);
1566 // Android provides a libc function to retrieve the address of the current
1567 // thread's unsafe stack pointer.
1568 Module *M = IRB.GetInsertBlock()->getParent()->getParent();
1569 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext());
1570 Value *Fn = M->getOrInsertFunction("__safestack_pointer_address",
1571 StackPtrTy->getPointerTo(0));
1572 return IRB.CreateCall(Fn);
1575 //===----------------------------------------------------------------------===//
1576 // Loop Strength Reduction hooks
1577 //===----------------------------------------------------------------------===//
1579 /// isLegalAddressingMode - Return true if the addressing mode represented
1580 /// by AM is legal for this target, for a load/store of the specified type.
1581 bool TargetLoweringBase::isLegalAddressingMode(const DataLayout &DL,
1582 const AddrMode &AM, Type *Ty,
1583 unsigned AS, Instruction *I) const {
1584 // The default implementation of this implements a conservative RISCy, r+r and
1587 // Allows a sign-extended 16-bit immediate field.
1588 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
1591 // No global is ever allowed as a base.
1595 // Only support r+r,
1597 case 0: // "r+i" or just "i", depending on HasBaseReg.
1600 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
1602 // Otherwise we have r+r or r+i.
1605 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
1607 // Allow 2*r as r+r.
1609 default: // Don't allow n * r
1616 //===----------------------------------------------------------------------===//
1618 //===----------------------------------------------------------------------===//
1620 // For OpenBSD return its special guard variable. Otherwise return nullptr,
1621 // so that SelectionDAG handle SSP.
1622 Value *TargetLoweringBase::getIRStackGuard(IRBuilder<> &IRB) const {
1623 if (getTargetMachine().getTargetTriple().isOSOpenBSD()) {
1624 Module &M = *IRB.GetInsertBlock()->getParent()->getParent();
1625 PointerType *PtrTy = Type::getInt8PtrTy(M.getContext());
1626 return M.getOrInsertGlobal("__guard_local", PtrTy);
1631 // Currently only support "standard" __stack_chk_guard.
1632 // TODO: add LOAD_STACK_GUARD support.
1633 void TargetLoweringBase::insertSSPDeclarations(Module &M) const {
1634 if (!M.getNamedValue("__stack_chk_guard"))
1635 new GlobalVariable(M, Type::getInt8PtrTy(M.getContext()), false,
1636 GlobalVariable::ExternalLinkage,
1637 nullptr, "__stack_chk_guard");
1640 // Currently only support "standard" __stack_chk_guard.
1641 // TODO: add LOAD_STACK_GUARD support.
1642 Value *TargetLoweringBase::getSDagStackGuard(const Module &M) const {
1643 return M.getNamedValue("__stack_chk_guard");
1646 Value *TargetLoweringBase::getSSPStackGuardCheck(const Module &M) const {
1650 unsigned TargetLoweringBase::getMinimumJumpTableEntries() const {
1651 return MinimumJumpTableEntries;
1654 void TargetLoweringBase::setMinimumJumpTableEntries(unsigned Val) {
1655 MinimumJumpTableEntries = Val;
1658 unsigned TargetLoweringBase::getMinimumJumpTableDensity(bool OptForSize) const {
1659 return OptForSize ? OptsizeJumpTableDensity : JumpTableDensity;
1662 unsigned TargetLoweringBase::getMaximumJumpTableSize() const {
1663 return MaximumJumpTableSize;
1666 void TargetLoweringBase::setMaximumJumpTableSize(unsigned Val) {
1667 MaximumJumpTableSize = Val;
1670 //===----------------------------------------------------------------------===//
1671 // Reciprocal Estimates
1672 //===----------------------------------------------------------------------===//
1674 /// Get the reciprocal estimate attribute string for a function that will
1675 /// override the target defaults.
1676 static StringRef getRecipEstimateForFunc(MachineFunction &MF) {
1677 const Function &F = MF.getFunction();
1678 return F.getFnAttribute("reciprocal-estimates").getValueAsString();
1681 /// Construct a string for the given reciprocal operation of the given type.
1682 /// This string should match the corresponding option to the front-end's
1683 /// "-mrecip" flag assuming those strings have been passed through in an
1684 /// attribute string. For example, "vec-divf" for a division of a vXf32.
1685 static std::string getReciprocalOpName(bool IsSqrt, EVT VT) {
1686 std::string Name = VT.isVector() ? "vec-" : "";
1688 Name += IsSqrt ? "sqrt" : "div";
1690 // TODO: Handle "half" or other float types?
1691 if (VT.getScalarType() == MVT::f64) {
1694 assert(VT.getScalarType() == MVT::f32 &&
1695 "Unexpected FP type for reciprocal estimate");
1702 /// Return the character position and value (a single numeric character) of a
1703 /// customized refinement operation in the input string if it exists. Return
1704 /// false if there is no customized refinement step count.
1705 static bool parseRefinementStep(StringRef In, size_t &Position,
1707 const char RefStepToken = ':';
1708 Position = In.find(RefStepToken);
1709 if (Position == StringRef::npos)
1712 StringRef RefStepString = In.substr(Position + 1);
1713 // Allow exactly one numeric character for the additional refinement
1715 if (RefStepString.size() == 1) {
1716 char RefStepChar = RefStepString[0];
1717 if (RefStepChar >= '0' && RefStepChar <= '9') {
1718 Value = RefStepChar - '0';
1722 report_fatal_error("Invalid refinement step for -recip.");
1725 /// For the input attribute string, return one of the ReciprocalEstimate enum
1726 /// status values (enabled, disabled, or not specified) for this operation on
1727 /// the specified data type.
1728 static int getOpEnabled(bool IsSqrt, EVT VT, StringRef Override) {
1729 if (Override.empty())
1730 return TargetLoweringBase::ReciprocalEstimate::Unspecified;
1732 SmallVector<StringRef, 4> OverrideVector;
1733 Override.split(OverrideVector, ',');
1734 unsigned NumArgs = OverrideVector.size();
1736 // Check if "all", "none", or "default" was specified.
1738 // Look for an optional setting of the number of refinement steps needed
1739 // for this type of reciprocal operation.
1742 if (parseRefinementStep(Override, RefPos, RefSteps)) {
1743 // Split the string for further processing.
1744 Override = Override.substr(0, RefPos);
1747 // All reciprocal types are enabled.
1748 if (Override == "all")
1749 return TargetLoweringBase::ReciprocalEstimate::Enabled;
1751 // All reciprocal types are disabled.
1752 if (Override == "none")
1753 return TargetLoweringBase::ReciprocalEstimate::Disabled;
1755 // Target defaults for enablement are used.
1756 if (Override == "default")
1757 return TargetLoweringBase::ReciprocalEstimate::Unspecified;
1760 // The attribute string may omit the size suffix ('f'/'d').
1761 std::string VTName = getReciprocalOpName(IsSqrt, VT);
1762 std::string VTNameNoSize = VTName;
1763 VTNameNoSize.pop_back();
1764 static const char DisabledPrefix = '!';
1766 for (StringRef RecipType : OverrideVector) {
1769 if (parseRefinementStep(RecipType, RefPos, RefSteps))
1770 RecipType = RecipType.substr(0, RefPos);
1772 // Ignore the disablement token for string matching.
1773 bool IsDisabled = RecipType[0] == DisabledPrefix;
1775 RecipType = RecipType.substr(1);
1777 if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize))
1778 return IsDisabled ? TargetLoweringBase::ReciprocalEstimate::Disabled
1779 : TargetLoweringBase::ReciprocalEstimate::Enabled;
1782 return TargetLoweringBase::ReciprocalEstimate::Unspecified;
1785 /// For the input attribute string, return the customized refinement step count
1786 /// for this operation on the specified data type. If the step count does not
1787 /// exist, return the ReciprocalEstimate enum value for unspecified.
1788 static int getOpRefinementSteps(bool IsSqrt, EVT VT, StringRef Override) {
1789 if (Override.empty())
1790 return TargetLoweringBase::ReciprocalEstimate::Unspecified;
1792 SmallVector<StringRef, 4> OverrideVector;
1793 Override.split(OverrideVector, ',');
1794 unsigned NumArgs = OverrideVector.size();
1796 // Check if "all", "default", or "none" was specified.
1798 // Look for an optional setting of the number of refinement steps needed
1799 // for this type of reciprocal operation.
1802 if (!parseRefinementStep(Override, RefPos, RefSteps))
1803 return TargetLoweringBase::ReciprocalEstimate::Unspecified;
1805 // Split the string for further processing.
1806 Override = Override.substr(0, RefPos);
1807 assert(Override != "none" &&
1808 "Disabled reciprocals, but specifed refinement steps?");
1810 // If this is a general override, return the specified number of steps.
1811 if (Override == "all" || Override == "default")
1815 // The attribute string may omit the size suffix ('f'/'d').
1816 std::string VTName = getReciprocalOpName(IsSqrt, VT);
1817 std::string VTNameNoSize = VTName;
1818 VTNameNoSize.pop_back();
1820 for (StringRef RecipType : OverrideVector) {
1823 if (!parseRefinementStep(RecipType, RefPos, RefSteps))
1826 RecipType = RecipType.substr(0, RefPos);
1827 if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize))
1831 return TargetLoweringBase::ReciprocalEstimate::Unspecified;
1834 int TargetLoweringBase::getRecipEstimateSqrtEnabled(EVT VT,
1835 MachineFunction &MF) const {
1836 return getOpEnabled(true, VT, getRecipEstimateForFunc(MF));
1839 int TargetLoweringBase::getRecipEstimateDivEnabled(EVT VT,
1840 MachineFunction &MF) const {
1841 return getOpEnabled(false, VT, getRecipEstimateForFunc(MF));
1844 int TargetLoweringBase::getSqrtRefinementSteps(EVT VT,
1845 MachineFunction &MF) const {
1846 return getOpRefinementSteps(true, VT, getRecipEstimateForFunc(MF));
1849 int TargetLoweringBase::getDivRefinementSteps(EVT VT,
1850 MachineFunction &MF) const {
1851 return getOpRefinementSteps(false, VT, getRecipEstimateForFunc(MF));
1854 void TargetLoweringBase::finalizeLowering(MachineFunction &MF) const {
1855 MF.getRegInfo().freezeReservedRegs(MF);