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/IR/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 (TT.isAndroid() && !TT.isAndroidVersionLT(9))) {
166 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
167 setLibcallName(RTLIB::SINCOS_F64, "sincos");
168 setLibcallName(RTLIB::SINCOS_F80, "sincosl");
169 setLibcallName(RTLIB::SINCOS_F128, "sincosl");
170 setLibcallName(RTLIB::SINCOS_PPCF128, "sincosl");
173 if (TT.isOSOpenBSD()) {
174 setLibcallName(RTLIB::STACKPROTECTOR_CHECK_FAIL, nullptr);
178 /// getFPEXT - Return the FPEXT_*_* value for the given types, or
179 /// UNKNOWN_LIBCALL if there is none.
180 RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) {
181 if (OpVT == MVT::f16) {
182 if (RetVT == MVT::f32)
183 return FPEXT_F16_F32;
184 } else if (OpVT == MVT::f32) {
185 if (RetVT == MVT::f64)
186 return FPEXT_F32_F64;
187 if (RetVT == MVT::f128)
188 return FPEXT_F32_F128;
189 if (RetVT == MVT::ppcf128)
190 return FPEXT_F32_PPCF128;
191 } else if (OpVT == MVT::f64) {
192 if (RetVT == MVT::f128)
193 return FPEXT_F64_F128;
194 else if (RetVT == MVT::ppcf128)
195 return FPEXT_F64_PPCF128;
196 } else if (OpVT == MVT::f80) {
197 if (RetVT == MVT::f128)
198 return FPEXT_F80_F128;
201 return UNKNOWN_LIBCALL;
204 /// getFPROUND - Return the FPROUND_*_* value for the given types, or
205 /// UNKNOWN_LIBCALL if there is none.
206 RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) {
207 if (RetVT == MVT::f16) {
208 if (OpVT == MVT::f32)
209 return FPROUND_F32_F16;
210 if (OpVT == MVT::f64)
211 return FPROUND_F64_F16;
212 if (OpVT == MVT::f80)
213 return FPROUND_F80_F16;
214 if (OpVT == MVT::f128)
215 return FPROUND_F128_F16;
216 if (OpVT == MVT::ppcf128)
217 return FPROUND_PPCF128_F16;
218 } else if (RetVT == MVT::f32) {
219 if (OpVT == MVT::f64)
220 return FPROUND_F64_F32;
221 if (OpVT == MVT::f80)
222 return FPROUND_F80_F32;
223 if (OpVT == MVT::f128)
224 return FPROUND_F128_F32;
225 if (OpVT == MVT::ppcf128)
226 return FPROUND_PPCF128_F32;
227 } else if (RetVT == MVT::f64) {
228 if (OpVT == MVT::f80)
229 return FPROUND_F80_F64;
230 if (OpVT == MVT::f128)
231 return FPROUND_F128_F64;
232 if (OpVT == MVT::ppcf128)
233 return FPROUND_PPCF128_F64;
234 } else if (RetVT == MVT::f80) {
235 if (OpVT == MVT::f128)
236 return FPROUND_F128_F80;
239 return UNKNOWN_LIBCALL;
242 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or
243 /// UNKNOWN_LIBCALL if there is none.
244 RTLIB::Libcall RTLIB::getFPTOSINT(EVT OpVT, EVT RetVT) {
245 if (OpVT == MVT::f32) {
246 if (RetVT == MVT::i32)
247 return FPTOSINT_F32_I32;
248 if (RetVT == MVT::i64)
249 return FPTOSINT_F32_I64;
250 if (RetVT == MVT::i128)
251 return FPTOSINT_F32_I128;
252 } else if (OpVT == MVT::f64) {
253 if (RetVT == MVT::i32)
254 return FPTOSINT_F64_I32;
255 if (RetVT == MVT::i64)
256 return FPTOSINT_F64_I64;
257 if (RetVT == MVT::i128)
258 return FPTOSINT_F64_I128;
259 } else if (OpVT == MVT::f80) {
260 if (RetVT == MVT::i32)
261 return FPTOSINT_F80_I32;
262 if (RetVT == MVT::i64)
263 return FPTOSINT_F80_I64;
264 if (RetVT == MVT::i128)
265 return FPTOSINT_F80_I128;
266 } else if (OpVT == MVT::f128) {
267 if (RetVT == MVT::i32)
268 return FPTOSINT_F128_I32;
269 if (RetVT == MVT::i64)
270 return FPTOSINT_F128_I64;
271 if (RetVT == MVT::i128)
272 return FPTOSINT_F128_I128;
273 } else if (OpVT == MVT::ppcf128) {
274 if (RetVT == MVT::i32)
275 return FPTOSINT_PPCF128_I32;
276 if (RetVT == MVT::i64)
277 return FPTOSINT_PPCF128_I64;
278 if (RetVT == MVT::i128)
279 return FPTOSINT_PPCF128_I128;
281 return UNKNOWN_LIBCALL;
284 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or
285 /// UNKNOWN_LIBCALL if there is none.
286 RTLIB::Libcall RTLIB::getFPTOUINT(EVT OpVT, EVT RetVT) {
287 if (OpVT == MVT::f32) {
288 if (RetVT == MVT::i32)
289 return FPTOUINT_F32_I32;
290 if (RetVT == MVT::i64)
291 return FPTOUINT_F32_I64;
292 if (RetVT == MVT::i128)
293 return FPTOUINT_F32_I128;
294 } else if (OpVT == MVT::f64) {
295 if (RetVT == MVT::i32)
296 return FPTOUINT_F64_I32;
297 if (RetVT == MVT::i64)
298 return FPTOUINT_F64_I64;
299 if (RetVT == MVT::i128)
300 return FPTOUINT_F64_I128;
301 } else if (OpVT == MVT::f80) {
302 if (RetVT == MVT::i32)
303 return FPTOUINT_F80_I32;
304 if (RetVT == MVT::i64)
305 return FPTOUINT_F80_I64;
306 if (RetVT == MVT::i128)
307 return FPTOUINT_F80_I128;
308 } else if (OpVT == MVT::f128) {
309 if (RetVT == MVT::i32)
310 return FPTOUINT_F128_I32;
311 if (RetVT == MVT::i64)
312 return FPTOUINT_F128_I64;
313 if (RetVT == MVT::i128)
314 return FPTOUINT_F128_I128;
315 } else if (OpVT == MVT::ppcf128) {
316 if (RetVT == MVT::i32)
317 return FPTOUINT_PPCF128_I32;
318 if (RetVT == MVT::i64)
319 return FPTOUINT_PPCF128_I64;
320 if (RetVT == MVT::i128)
321 return FPTOUINT_PPCF128_I128;
323 return UNKNOWN_LIBCALL;
326 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or
327 /// UNKNOWN_LIBCALL if there is none.
328 RTLIB::Libcall RTLIB::getSINTTOFP(EVT OpVT, EVT RetVT) {
329 if (OpVT == MVT::i32) {
330 if (RetVT == MVT::f32)
331 return SINTTOFP_I32_F32;
332 if (RetVT == MVT::f64)
333 return SINTTOFP_I32_F64;
334 if (RetVT == MVT::f80)
335 return SINTTOFP_I32_F80;
336 if (RetVT == MVT::f128)
337 return SINTTOFP_I32_F128;
338 if (RetVT == MVT::ppcf128)
339 return SINTTOFP_I32_PPCF128;
340 } else if (OpVT == MVT::i64) {
341 if (RetVT == MVT::f32)
342 return SINTTOFP_I64_F32;
343 if (RetVT == MVT::f64)
344 return SINTTOFP_I64_F64;
345 if (RetVT == MVT::f80)
346 return SINTTOFP_I64_F80;
347 if (RetVT == MVT::f128)
348 return SINTTOFP_I64_F128;
349 if (RetVT == MVT::ppcf128)
350 return SINTTOFP_I64_PPCF128;
351 } else if (OpVT == MVT::i128) {
352 if (RetVT == MVT::f32)
353 return SINTTOFP_I128_F32;
354 if (RetVT == MVT::f64)
355 return SINTTOFP_I128_F64;
356 if (RetVT == MVT::f80)
357 return SINTTOFP_I128_F80;
358 if (RetVT == MVT::f128)
359 return SINTTOFP_I128_F128;
360 if (RetVT == MVT::ppcf128)
361 return SINTTOFP_I128_PPCF128;
363 return UNKNOWN_LIBCALL;
366 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or
367 /// UNKNOWN_LIBCALL if there is none.
368 RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) {
369 if (OpVT == MVT::i32) {
370 if (RetVT == MVT::f32)
371 return UINTTOFP_I32_F32;
372 if (RetVT == MVT::f64)
373 return UINTTOFP_I32_F64;
374 if (RetVT == MVT::f80)
375 return UINTTOFP_I32_F80;
376 if (RetVT == MVT::f128)
377 return UINTTOFP_I32_F128;
378 if (RetVT == MVT::ppcf128)
379 return UINTTOFP_I32_PPCF128;
380 } else if (OpVT == MVT::i64) {
381 if (RetVT == MVT::f32)
382 return UINTTOFP_I64_F32;
383 if (RetVT == MVT::f64)
384 return UINTTOFP_I64_F64;
385 if (RetVT == MVT::f80)
386 return UINTTOFP_I64_F80;
387 if (RetVT == MVT::f128)
388 return UINTTOFP_I64_F128;
389 if (RetVT == MVT::ppcf128)
390 return UINTTOFP_I64_PPCF128;
391 } else if (OpVT == MVT::i128) {
392 if (RetVT == MVT::f32)
393 return UINTTOFP_I128_F32;
394 if (RetVT == MVT::f64)
395 return UINTTOFP_I128_F64;
396 if (RetVT == MVT::f80)
397 return UINTTOFP_I128_F80;
398 if (RetVT == MVT::f128)
399 return UINTTOFP_I128_F128;
400 if (RetVT == MVT::ppcf128)
401 return UINTTOFP_I128_PPCF128;
403 return UNKNOWN_LIBCALL;
406 RTLIB::Libcall RTLIB::getSYNC(unsigned Opc, MVT VT) {
407 #define OP_TO_LIBCALL(Name, Enum) \
409 switch (VT.SimpleTy) { \
411 return UNKNOWN_LIBCALL; \
425 OP_TO_LIBCALL(ISD::ATOMIC_SWAP, SYNC_LOCK_TEST_AND_SET)
426 OP_TO_LIBCALL(ISD::ATOMIC_CMP_SWAP, SYNC_VAL_COMPARE_AND_SWAP)
427 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_ADD, SYNC_FETCH_AND_ADD)
428 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_SUB, SYNC_FETCH_AND_SUB)
429 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_AND, SYNC_FETCH_AND_AND)
430 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_OR, SYNC_FETCH_AND_OR)
431 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_XOR, SYNC_FETCH_AND_XOR)
432 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_NAND, SYNC_FETCH_AND_NAND)
433 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MAX, SYNC_FETCH_AND_MAX)
434 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMAX, SYNC_FETCH_AND_UMAX)
435 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MIN, SYNC_FETCH_AND_MIN)
436 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMIN, SYNC_FETCH_AND_UMIN)
441 return UNKNOWN_LIBCALL;
444 RTLIB::Libcall RTLIB::getMEMCPY_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) {
445 switch (ElementSize) {
447 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_1;
449 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_2;
451 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_4;
453 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_8;
455 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_16;
457 return UNKNOWN_LIBCALL;
461 RTLIB::Libcall RTLIB::getMEMMOVE_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) {
462 switch (ElementSize) {
464 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_1;
466 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_2;
468 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_4;
470 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_8;
472 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_16;
474 return UNKNOWN_LIBCALL;
478 RTLIB::Libcall RTLIB::getMEMSET_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) {
479 switch (ElementSize) {
481 return MEMSET_ELEMENT_UNORDERED_ATOMIC_1;
483 return MEMSET_ELEMENT_UNORDERED_ATOMIC_2;
485 return MEMSET_ELEMENT_UNORDERED_ATOMIC_4;
487 return MEMSET_ELEMENT_UNORDERED_ATOMIC_8;
489 return MEMSET_ELEMENT_UNORDERED_ATOMIC_16;
491 return UNKNOWN_LIBCALL;
495 /// InitCmpLibcallCCs - Set default comparison libcall CC.
496 static void InitCmpLibcallCCs(ISD::CondCode *CCs) {
497 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL);
498 CCs[RTLIB::OEQ_F32] = ISD::SETEQ;
499 CCs[RTLIB::OEQ_F64] = ISD::SETEQ;
500 CCs[RTLIB::OEQ_F128] = ISD::SETEQ;
501 CCs[RTLIB::OEQ_PPCF128] = ISD::SETEQ;
502 CCs[RTLIB::UNE_F32] = ISD::SETNE;
503 CCs[RTLIB::UNE_F64] = ISD::SETNE;
504 CCs[RTLIB::UNE_F128] = ISD::SETNE;
505 CCs[RTLIB::UNE_PPCF128] = ISD::SETNE;
506 CCs[RTLIB::OGE_F32] = ISD::SETGE;
507 CCs[RTLIB::OGE_F64] = ISD::SETGE;
508 CCs[RTLIB::OGE_F128] = ISD::SETGE;
509 CCs[RTLIB::OGE_PPCF128] = ISD::SETGE;
510 CCs[RTLIB::OLT_F32] = ISD::SETLT;
511 CCs[RTLIB::OLT_F64] = ISD::SETLT;
512 CCs[RTLIB::OLT_F128] = ISD::SETLT;
513 CCs[RTLIB::OLT_PPCF128] = ISD::SETLT;
514 CCs[RTLIB::OLE_F32] = ISD::SETLE;
515 CCs[RTLIB::OLE_F64] = ISD::SETLE;
516 CCs[RTLIB::OLE_F128] = ISD::SETLE;
517 CCs[RTLIB::OLE_PPCF128] = ISD::SETLE;
518 CCs[RTLIB::OGT_F32] = ISD::SETGT;
519 CCs[RTLIB::OGT_F64] = ISD::SETGT;
520 CCs[RTLIB::OGT_F128] = ISD::SETGT;
521 CCs[RTLIB::OGT_PPCF128] = ISD::SETGT;
522 CCs[RTLIB::UO_F32] = ISD::SETNE;
523 CCs[RTLIB::UO_F64] = ISD::SETNE;
524 CCs[RTLIB::UO_F128] = ISD::SETNE;
525 CCs[RTLIB::UO_PPCF128] = ISD::SETNE;
526 CCs[RTLIB::O_F32] = ISD::SETEQ;
527 CCs[RTLIB::O_F64] = ISD::SETEQ;
528 CCs[RTLIB::O_F128] = ISD::SETEQ;
529 CCs[RTLIB::O_PPCF128] = ISD::SETEQ;
532 /// NOTE: The TargetMachine owns TLOF.
533 TargetLoweringBase::TargetLoweringBase(const TargetMachine &tm) : TM(tm) {
536 // Perform these initializations only once.
537 MaxStoresPerMemset = MaxStoresPerMemcpy = MaxStoresPerMemmove =
538 MaxLoadsPerMemcmp = 8;
539 MaxGluedStoresPerMemcpy = 0;
540 MaxStoresPerMemsetOptSize = MaxStoresPerMemcpyOptSize =
541 MaxStoresPerMemmoveOptSize = MaxLoadsPerMemcmpOptSize = 4;
542 UseUnderscoreSetJmp = false;
543 UseUnderscoreLongJmp = false;
544 HasMultipleConditionRegisters = false;
545 HasExtractBitsInsn = false;
546 JumpIsExpensive = JumpIsExpensiveOverride;
547 PredictableSelectIsExpensive = false;
548 EnableExtLdPromotion = false;
549 HasFloatingPointExceptions = true;
550 StackPointerRegisterToSaveRestore = 0;
551 BooleanContents = UndefinedBooleanContent;
552 BooleanFloatContents = UndefinedBooleanContent;
553 BooleanVectorContents = UndefinedBooleanContent;
554 SchedPreferenceInfo = Sched::ILP;
556 JumpBufAlignment = 0;
557 MinFunctionAlignment = 0;
558 PrefFunctionAlignment = 0;
559 PrefLoopAlignment = 0;
560 GatherAllAliasesMaxDepth = 18;
561 MinStackArgumentAlignment = 1;
562 // TODO: the default will be switched to 0 in the next commit, along
563 // with the Target-specific changes necessary.
564 MaxAtomicSizeInBitsSupported = 1024;
566 MinCmpXchgSizeInBits = 0;
567 SupportsUnalignedAtomics = false;
569 std::fill(std::begin(LibcallRoutineNames), std::end(LibcallRoutineNames), nullptr);
571 InitLibcalls(TM.getTargetTriple());
572 InitCmpLibcallCCs(CmpLibcallCCs);
575 void TargetLoweringBase::initActions() {
576 // All operations default to being supported.
577 memset(OpActions, 0, sizeof(OpActions));
578 memset(LoadExtActions, 0, sizeof(LoadExtActions));
579 memset(TruncStoreActions, 0, sizeof(TruncStoreActions));
580 memset(IndexedModeActions, 0, sizeof(IndexedModeActions));
581 memset(CondCodeActions, 0, sizeof(CondCodeActions));
582 std::fill(std::begin(RegClassForVT), std::end(RegClassForVT), nullptr);
583 std::fill(std::begin(TargetDAGCombineArray),
584 std::end(TargetDAGCombineArray), 0);
586 // Set default actions for various operations.
587 for (MVT VT : MVT::all_valuetypes()) {
588 // Default all indexed load / store to expand.
589 for (unsigned IM = (unsigned)ISD::PRE_INC;
590 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) {
591 setIndexedLoadAction(IM, VT, Expand);
592 setIndexedStoreAction(IM, VT, Expand);
595 // Most backends expect to see the node which just returns the value loaded.
596 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Expand);
598 // These operations default to expand.
599 setOperationAction(ISD::FGETSIGN, VT, Expand);
600 setOperationAction(ISD::CONCAT_VECTORS, VT, Expand);
601 setOperationAction(ISD::FMINNUM, VT, Expand);
602 setOperationAction(ISD::FMAXNUM, VT, Expand);
603 setOperationAction(ISD::FMINNUM_IEEE, VT, Expand);
604 setOperationAction(ISD::FMAXNUM_IEEE, VT, Expand);
605 setOperationAction(ISD::FMINIMUM, VT, Expand);
606 setOperationAction(ISD::FMAXIMUM, VT, Expand);
607 setOperationAction(ISD::FMAD, VT, Expand);
608 setOperationAction(ISD::SMIN, VT, Expand);
609 setOperationAction(ISD::SMAX, VT, Expand);
610 setOperationAction(ISD::UMIN, VT, Expand);
611 setOperationAction(ISD::UMAX, VT, Expand);
612 setOperationAction(ISD::ABS, VT, Expand);
613 setOperationAction(ISD::FSHL, VT, Expand);
614 setOperationAction(ISD::FSHR, VT, Expand);
615 setOperationAction(ISD::SADDSAT, VT, Expand);
616 setOperationAction(ISD::UADDSAT, VT, Expand);
617 setOperationAction(ISD::SSUBSAT, VT, Expand);
618 setOperationAction(ISD::USUBSAT, VT, Expand);
620 // Overflow operations default to expand
621 setOperationAction(ISD::SADDO, VT, Expand);
622 setOperationAction(ISD::SSUBO, VT, Expand);
623 setOperationAction(ISD::UADDO, VT, Expand);
624 setOperationAction(ISD::USUBO, VT, Expand);
625 setOperationAction(ISD::SMULO, VT, Expand);
626 setOperationAction(ISD::UMULO, VT, Expand);
628 // ADDCARRY operations default to expand
629 setOperationAction(ISD::ADDCARRY, VT, Expand);
630 setOperationAction(ISD::SUBCARRY, VT, Expand);
631 setOperationAction(ISD::SETCCCARRY, VT, Expand);
633 // ADDC/ADDE/SUBC/SUBE default to expand.
634 setOperationAction(ISD::ADDC, VT, Expand);
635 setOperationAction(ISD::ADDE, VT, Expand);
636 setOperationAction(ISD::SUBC, VT, Expand);
637 setOperationAction(ISD::SUBE, VT, Expand);
639 // These default to Expand so they will be expanded to CTLZ/CTTZ by default.
640 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
641 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
643 setOperationAction(ISD::BITREVERSE, VT, Expand);
645 // These library functions default to expand.
646 setOperationAction(ISD::FROUND, VT, Expand);
647 setOperationAction(ISD::FPOWI, VT, Expand);
649 // These operations default to expand for vector types.
651 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
652 setOperationAction(ISD::ANY_EXTEND_VECTOR_INREG, VT, Expand);
653 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, VT, Expand);
654 setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, VT, Expand);
657 // For most targets @llvm.get.dynamic.area.offset just returns 0.
658 setOperationAction(ISD::GET_DYNAMIC_AREA_OFFSET, VT, Expand);
661 // Most targets ignore the @llvm.prefetch intrinsic.
662 setOperationAction(ISD::PREFETCH, MVT::Other, Expand);
664 // Most targets also ignore the @llvm.readcyclecounter intrinsic.
665 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Expand);
667 // ConstantFP nodes default to expand. Targets can either change this to
668 // Legal, in which case all fp constants are legal, or use isFPImmLegal()
669 // to optimize expansions for certain constants.
670 setOperationAction(ISD::ConstantFP, MVT::f16, Expand);
671 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
672 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
673 setOperationAction(ISD::ConstantFP, MVT::f80, Expand);
674 setOperationAction(ISD::ConstantFP, MVT::f128, Expand);
676 // These library functions default to expand.
677 for (MVT VT : {MVT::f32, MVT::f64, MVT::f128}) {
678 setOperationAction(ISD::FCBRT, VT, Expand);
679 setOperationAction(ISD::FLOG , VT, Expand);
680 setOperationAction(ISD::FLOG2, VT, Expand);
681 setOperationAction(ISD::FLOG10, VT, Expand);
682 setOperationAction(ISD::FEXP , VT, Expand);
683 setOperationAction(ISD::FEXP2, VT, Expand);
684 setOperationAction(ISD::FFLOOR, VT, Expand);
685 setOperationAction(ISD::FNEARBYINT, VT, Expand);
686 setOperationAction(ISD::FCEIL, VT, Expand);
687 setOperationAction(ISD::FRINT, VT, Expand);
688 setOperationAction(ISD::FTRUNC, VT, Expand);
689 setOperationAction(ISD::FROUND, VT, Expand);
692 // Default ISD::TRAP to expand (which turns it into abort).
693 setOperationAction(ISD::TRAP, MVT::Other, Expand);
695 // On most systems, DEBUGTRAP and TRAP have no difference. The "Expand"
696 // here is to inform DAG Legalizer to replace DEBUGTRAP with TRAP.
697 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Expand);
700 MVT TargetLoweringBase::getScalarShiftAmountTy(const DataLayout &DL,
702 return MVT::getIntegerVT(8 * DL.getPointerSize(0));
705 EVT TargetLoweringBase::getShiftAmountTy(EVT LHSTy, const DataLayout &DL,
706 bool LegalTypes) const {
707 assert(LHSTy.isInteger() && "Shift amount is not an integer type!");
708 if (LHSTy.isVector())
710 return LegalTypes ? getScalarShiftAmountTy(DL, LHSTy)
714 bool TargetLoweringBase::canOpTrap(unsigned Op, EVT VT) const {
715 assert(isTypeLegal(VT));
727 void TargetLoweringBase::setJumpIsExpensive(bool isExpensive) {
728 // If the command-line option was specified, ignore this request.
729 if (!JumpIsExpensiveOverride.getNumOccurrences())
730 JumpIsExpensive = isExpensive;
733 TargetLoweringBase::LegalizeKind
734 TargetLoweringBase::getTypeConversion(LLVMContext &Context, EVT VT) const {
735 // If this is a simple type, use the ComputeRegisterProp mechanism.
737 MVT SVT = VT.getSimpleVT();
738 assert((unsigned)SVT.SimpleTy < array_lengthof(TransformToType));
739 MVT NVT = TransformToType[SVT.SimpleTy];
740 LegalizeTypeAction LA = ValueTypeActions.getTypeAction(SVT);
742 assert((LA == TypeLegal || LA == TypeSoftenFloat ||
743 ValueTypeActions.getTypeAction(NVT) != TypePromoteInteger) &&
744 "Promote may not follow Expand or Promote");
746 if (LA == TypeSplitVector)
747 return LegalizeKind(LA,
748 EVT::getVectorVT(Context, SVT.getVectorElementType(),
749 SVT.getVectorNumElements() / 2));
750 if (LA == TypeScalarizeVector)
751 return LegalizeKind(LA, SVT.getVectorElementType());
752 return LegalizeKind(LA, NVT);
755 // Handle Extended Scalar Types.
756 if (!VT.isVector()) {
757 assert(VT.isInteger() && "Float types must be simple");
758 unsigned BitSize = VT.getSizeInBits();
759 // First promote to a power-of-two size, then expand if necessary.
760 if (BitSize < 8 || !isPowerOf2_32(BitSize)) {
761 EVT NVT = VT.getRoundIntegerType(Context);
762 assert(NVT != VT && "Unable to round integer VT");
763 LegalizeKind NextStep = getTypeConversion(Context, NVT);
764 // Avoid multi-step promotion.
765 if (NextStep.first == TypePromoteInteger)
767 // Return rounded integer type.
768 return LegalizeKind(TypePromoteInteger, NVT);
771 return LegalizeKind(TypeExpandInteger,
772 EVT::getIntegerVT(Context, VT.getSizeInBits() / 2));
775 // Handle vector types.
776 unsigned NumElts = VT.getVectorNumElements();
777 EVT EltVT = VT.getVectorElementType();
779 // Vectors with only one element are always scalarized.
781 return LegalizeKind(TypeScalarizeVector, EltVT);
783 // Try to widen vector elements until the element type is a power of two and
784 // promote it to a legal type later on, for example:
785 // <3 x i8> -> <4 x i8> -> <4 x i32>
786 if (EltVT.isInteger()) {
787 // Vectors with a number of elements that is not a power of two are always
788 // widened, for example <3 x i8> -> <4 x i8>.
789 if (!VT.isPow2VectorType()) {
790 NumElts = (unsigned)NextPowerOf2(NumElts);
791 EVT NVT = EVT::getVectorVT(Context, EltVT, NumElts);
792 return LegalizeKind(TypeWidenVector, NVT);
795 // Examine the element type.
796 LegalizeKind LK = getTypeConversion(Context, EltVT);
798 // If type is to be expanded, split the vector.
799 // <4 x i140> -> <2 x i140>
800 if (LK.first == TypeExpandInteger)
801 return LegalizeKind(TypeSplitVector,
802 EVT::getVectorVT(Context, EltVT, NumElts / 2));
804 // Promote the integer element types until a legal vector type is found
805 // or until the element integer type is too big. If a legal type was not
806 // found, fallback to the usual mechanism of widening/splitting the
808 EVT OldEltVT = EltVT;
810 // Increase the bitwidth of the element to the next pow-of-two
811 // (which is greater than 8 bits).
812 EltVT = EVT::getIntegerVT(Context, 1 + EltVT.getSizeInBits())
813 .getRoundIntegerType(Context);
815 // Stop trying when getting a non-simple element type.
816 // Note that vector elements may be greater than legal vector element
817 // types. Example: X86 XMM registers hold 64bit element on 32bit
819 if (!EltVT.isSimple())
822 // Build a new vector type and check if it is legal.
823 MVT NVT = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts);
824 // Found a legal promoted vector type.
825 if (NVT != MVT() && ValueTypeActions.getTypeAction(NVT) == TypeLegal)
826 return LegalizeKind(TypePromoteInteger,
827 EVT::getVectorVT(Context, EltVT, NumElts));
830 // Reset the type to the unexpanded type if we did not find a legal vector
831 // type with a promoted vector element type.
835 // Try to widen the vector until a legal type is found.
836 // If there is no wider legal type, split the vector.
838 // Round up to the next power of 2.
839 NumElts = (unsigned)NextPowerOf2(NumElts);
841 // If there is no simple vector type with this many elements then there
842 // cannot be a larger legal vector type. Note that this assumes that
843 // there are no skipped intermediate vector types in the simple types.
844 if (!EltVT.isSimple())
846 MVT LargerVector = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts);
847 if (LargerVector == MVT())
850 // If this type is legal then widen the vector.
851 if (ValueTypeActions.getTypeAction(LargerVector) == TypeLegal)
852 return LegalizeKind(TypeWidenVector, LargerVector);
855 // Widen odd vectors to next power of two.
856 if (!VT.isPow2VectorType()) {
857 EVT NVT = VT.getPow2VectorType(Context);
858 return LegalizeKind(TypeWidenVector, NVT);
861 // Vectors with illegal element types are expanded.
862 EVT NVT = EVT::getVectorVT(Context, EltVT, VT.getVectorNumElements() / 2);
863 return LegalizeKind(TypeSplitVector, NVT);
866 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT,
867 unsigned &NumIntermediates,
869 TargetLoweringBase *TLI) {
870 // Figure out the right, legal destination reg to copy into.
871 unsigned NumElts = VT.getVectorNumElements();
872 MVT EltTy = VT.getVectorElementType();
874 unsigned NumVectorRegs = 1;
876 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
877 // could break down into LHS/RHS like LegalizeDAG does.
878 if (!isPowerOf2_32(NumElts)) {
879 NumVectorRegs = NumElts;
883 // Divide the input until we get to a supported size. This will always
884 // end with a scalar if the target doesn't support vectors.
885 while (NumElts > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) {
890 NumIntermediates = NumVectorRegs;
892 MVT NewVT = MVT::getVectorVT(EltTy, NumElts);
893 if (!TLI->isTypeLegal(NewVT))
895 IntermediateVT = NewVT;
897 unsigned NewVTSize = NewVT.getSizeInBits();
899 // Convert sizes such as i33 to i64.
900 if (!isPowerOf2_32(NewVTSize))
901 NewVTSize = NextPowerOf2(NewVTSize);
903 MVT DestVT = TLI->getRegisterType(NewVT);
905 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16.
906 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits());
908 // Otherwise, promotion or legal types use the same number of registers as
909 // the vector decimated to the appropriate level.
910 return NumVectorRegs;
913 /// isLegalRC - Return true if the value types that can be represented by the
914 /// specified register class are all legal.
915 bool TargetLoweringBase::isLegalRC(const TargetRegisterInfo &TRI,
916 const TargetRegisterClass &RC) const {
917 for (auto I = TRI.legalclasstypes_begin(RC); *I != MVT::Other; ++I)
923 /// Replace/modify any TargetFrameIndex operands with a targte-dependent
924 /// sequence of memory operands that is recognized by PrologEpilogInserter.
926 TargetLoweringBase::emitPatchPoint(MachineInstr &InitialMI,
927 MachineBasicBlock *MBB) const {
928 MachineInstr *MI = &InitialMI;
929 MachineFunction &MF = *MI->getMF();
930 MachineFrameInfo &MFI = MF.getFrameInfo();
932 // We're handling multiple types of operands here:
933 // PATCHPOINT MetaArgs - live-in, read only, direct
934 // STATEPOINT Deopt Spill - live-through, read only, indirect
935 // STATEPOINT Deopt Alloca - live-through, read only, direct
936 // (We're currently conservative and mark the deopt slots read/write in
938 // STATEPOINT GC Spill - live-through, read/write, indirect
939 // STATEPOINT GC Alloca - live-through, read/write, direct
940 // The live-in vs live-through is handled already (the live through ones are
941 // all stack slots), but we need to handle the different type of stackmap
942 // operands and memory effects here.
944 // MI changes inside this loop as we grow operands.
945 for(unsigned OperIdx = 0; OperIdx != MI->getNumOperands(); ++OperIdx) {
946 MachineOperand &MO = MI->getOperand(OperIdx);
950 // foldMemoryOperand builds a new MI after replacing a single FI operand
951 // with the canonical set of five x86 addressing-mode operands.
952 int FI = MO.getIndex();
953 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), MI->getDesc());
955 // Copy operands before the frame-index.
956 for (unsigned i = 0; i < OperIdx; ++i)
957 MIB.add(MI->getOperand(i));
958 // Add frame index operands recognized by stackmaps.cpp
959 if (MFI.isStatepointSpillSlotObjectIndex(FI)) {
960 // indirect-mem-ref tag, size, #FI, offset.
961 // Used for spills inserted by StatepointLowering. This codepath is not
962 // used for patchpoints/stackmaps at all, for these spilling is done via
963 // foldMemoryOperand callback only.
964 assert(MI->getOpcode() == TargetOpcode::STATEPOINT && "sanity");
965 MIB.addImm(StackMaps::IndirectMemRefOp);
966 MIB.addImm(MFI.getObjectSize(FI));
967 MIB.add(MI->getOperand(OperIdx));
970 // direct-mem-ref tag, #FI, offset.
971 // Used by patchpoint, and direct alloca arguments to statepoints
972 MIB.addImm(StackMaps::DirectMemRefOp);
973 MIB.add(MI->getOperand(OperIdx));
976 // Copy the operands after the frame index.
977 for (unsigned i = OperIdx + 1; i != MI->getNumOperands(); ++i)
978 MIB.add(MI->getOperand(i));
980 // Inherit previous memory operands.
981 MIB.cloneMemRefs(*MI);
982 assert(MIB->mayLoad() && "Folded a stackmap use to a non-load!");
984 // Add a new memory operand for this FI.
985 assert(MFI.getObjectOffset(FI) != -1);
987 auto Flags = MachineMemOperand::MOLoad;
988 if (MI->getOpcode() == TargetOpcode::STATEPOINT) {
989 Flags |= MachineMemOperand::MOStore;
990 Flags |= MachineMemOperand::MOVolatile;
992 MachineMemOperand *MMO = MF.getMachineMemOperand(
993 MachinePointerInfo::getFixedStack(MF, FI), Flags,
994 MF.getDataLayout().getPointerSize(), MFI.getObjectAlignment(FI));
995 MIB->addMemOperand(MF, MMO);
997 // Replace the instruction and update the operand index.
998 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
999 OperIdx += (MIB->getNumOperands() - MI->getNumOperands()) - 1;
1000 MI->eraseFromParent();
1007 TargetLoweringBase::emitXRayCustomEvent(MachineInstr &MI,
1008 MachineBasicBlock *MBB) const {
1009 assert(MI.getOpcode() == TargetOpcode::PATCHABLE_EVENT_CALL &&
1010 "Called emitXRayCustomEvent on the wrong MI!");
1011 auto &MF = *MI.getMF();
1012 auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc());
1013 for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx)
1014 MIB.add(MI.getOperand(OpIdx));
1016 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
1017 MI.eraseFromParent();
1022 TargetLoweringBase::emitXRayTypedEvent(MachineInstr &MI,
1023 MachineBasicBlock *MBB) const {
1024 assert(MI.getOpcode() == TargetOpcode::PATCHABLE_TYPED_EVENT_CALL &&
1025 "Called emitXRayTypedEvent on the wrong MI!");
1026 auto &MF = *MI.getMF();
1027 auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc());
1028 for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx)
1029 MIB.add(MI.getOperand(OpIdx));
1031 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
1032 MI.eraseFromParent();
1036 /// findRepresentativeClass - Return the largest legal super-reg register class
1037 /// of the register class for the specified type and its associated "cost".
1038 // This function is in TargetLowering because it uses RegClassForVT which would
1039 // need to be moved to TargetRegisterInfo and would necessitate moving
1040 // isTypeLegal over as well - a massive change that would just require
1041 // TargetLowering having a TargetRegisterInfo class member that it would use.
1042 std::pair<const TargetRegisterClass *, uint8_t>
1043 TargetLoweringBase::findRepresentativeClass(const TargetRegisterInfo *TRI,
1045 const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy];
1047 return std::make_pair(RC, 0);
1049 // Compute the set of all super-register classes.
1050 BitVector SuperRegRC(TRI->getNumRegClasses());
1051 for (SuperRegClassIterator RCI(RC, TRI); RCI.isValid(); ++RCI)
1052 SuperRegRC.setBitsInMask(RCI.getMask());
1054 // Find the first legal register class with the largest spill size.
1055 const TargetRegisterClass *BestRC = RC;
1056 for (unsigned i : SuperRegRC.set_bits()) {
1057 const TargetRegisterClass *SuperRC = TRI->getRegClass(i);
1058 // We want the largest possible spill size.
1059 if (TRI->getSpillSize(*SuperRC) <= TRI->getSpillSize(*BestRC))
1061 if (!isLegalRC(*TRI, *SuperRC))
1065 return std::make_pair(BestRC, 1);
1068 /// computeRegisterProperties - Once all of the register classes are added,
1069 /// this allows us to compute derived properties we expose.
1070 void TargetLoweringBase::computeRegisterProperties(
1071 const TargetRegisterInfo *TRI) {
1072 static_assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE,
1073 "Too many value types for ValueTypeActions to hold!");
1075 // Everything defaults to needing one register.
1076 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
1077 NumRegistersForVT[i] = 1;
1078 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i;
1080 // ...except isVoid, which doesn't need any registers.
1081 NumRegistersForVT[MVT::isVoid] = 0;
1083 // Find the largest integer register class.
1084 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE;
1085 for (; RegClassForVT[LargestIntReg] == nullptr; --LargestIntReg)
1086 assert(LargestIntReg != MVT::i1 && "No integer registers defined!");
1088 // Every integer value type larger than this largest register takes twice as
1089 // many registers to represent as the previous ValueType.
1090 for (unsigned ExpandedReg = LargestIntReg + 1;
1091 ExpandedReg <= MVT::LAST_INTEGER_VALUETYPE; ++ExpandedReg) {
1092 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1];
1093 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg;
1094 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1);
1095 ValueTypeActions.setTypeAction((MVT::SimpleValueType)ExpandedReg,
1099 // Inspect all of the ValueType's smaller than the largest integer
1100 // register to see which ones need promotion.
1101 unsigned LegalIntReg = LargestIntReg;
1102 for (unsigned IntReg = LargestIntReg - 1;
1103 IntReg >= (unsigned)MVT::i1; --IntReg) {
1104 MVT IVT = (MVT::SimpleValueType)IntReg;
1105 if (isTypeLegal(IVT)) {
1106 LegalIntReg = IntReg;
1108 RegisterTypeForVT[IntReg] = TransformToType[IntReg] =
1109 (MVT::SimpleValueType)LegalIntReg;
1110 ValueTypeActions.setTypeAction(IVT, TypePromoteInteger);
1114 // ppcf128 type is really two f64's.
1115 if (!isTypeLegal(MVT::ppcf128)) {
1116 if (isTypeLegal(MVT::f64)) {
1117 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64];
1118 RegisterTypeForVT[MVT::ppcf128] = MVT::f64;
1119 TransformToType[MVT::ppcf128] = MVT::f64;
1120 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeExpandFloat);
1122 NumRegistersForVT[MVT::ppcf128] = NumRegistersForVT[MVT::i128];
1123 RegisterTypeForVT[MVT::ppcf128] = RegisterTypeForVT[MVT::i128];
1124 TransformToType[MVT::ppcf128] = MVT::i128;
1125 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeSoftenFloat);
1129 // Decide how to handle f128. If the target does not have native f128 support,
1130 // expand it to i128 and we will be generating soft float library calls.
1131 if (!isTypeLegal(MVT::f128)) {
1132 NumRegistersForVT[MVT::f128] = NumRegistersForVT[MVT::i128];
1133 RegisterTypeForVT[MVT::f128] = RegisterTypeForVT[MVT::i128];
1134 TransformToType[MVT::f128] = MVT::i128;
1135 ValueTypeActions.setTypeAction(MVT::f128, TypeSoftenFloat);
1138 // Decide how to handle f64. If the target does not have native f64 support,
1139 // expand it to i64 and we will be generating soft float library calls.
1140 if (!isTypeLegal(MVT::f64)) {
1141 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64];
1142 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64];
1143 TransformToType[MVT::f64] = MVT::i64;
1144 ValueTypeActions.setTypeAction(MVT::f64, TypeSoftenFloat);
1147 // Decide how to handle f32. If the target does not have native f32 support,
1148 // expand it to i32 and we will be generating soft float library calls.
1149 if (!isTypeLegal(MVT::f32)) {
1150 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32];
1151 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32];
1152 TransformToType[MVT::f32] = MVT::i32;
1153 ValueTypeActions.setTypeAction(MVT::f32, TypeSoftenFloat);
1156 // Decide how to handle f16. If the target does not have native f16 support,
1157 // promote it to f32, because there are no f16 library calls (except for
1159 if (!isTypeLegal(MVT::f16)) {
1160 NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::f32];
1161 RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::f32];
1162 TransformToType[MVT::f16] = MVT::f32;
1163 ValueTypeActions.setTypeAction(MVT::f16, TypePromoteFloat);
1166 // Loop over all of the vector value types to see which need transformations.
1167 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE;
1168 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
1169 MVT VT = (MVT::SimpleValueType) i;
1170 if (isTypeLegal(VT))
1173 MVT EltVT = VT.getVectorElementType();
1174 unsigned NElts = VT.getVectorNumElements();
1175 bool IsLegalWiderType = false;
1176 LegalizeTypeAction PreferredAction = getPreferredVectorAction(VT);
1177 switch (PreferredAction) {
1178 case TypePromoteInteger:
1179 // Try to promote the elements of integer vectors. If no legal
1180 // promotion was found, fall through to the widen-vector method.
1181 for (unsigned nVT = i + 1; nVT <= MVT::LAST_INTEGER_VECTOR_VALUETYPE; ++nVT) {
1182 MVT SVT = (MVT::SimpleValueType) nVT;
1183 // Promote vectors of integers to vectors with the same number
1184 // of elements, with a wider element type.
1185 if (SVT.getScalarSizeInBits() > EltVT.getSizeInBits() &&
1186 SVT.getVectorNumElements() == NElts && isTypeLegal(SVT)) {
1187 TransformToType[i] = SVT;
1188 RegisterTypeForVT[i] = SVT;
1189 NumRegistersForVT[i] = 1;
1190 ValueTypeActions.setTypeAction(VT, TypePromoteInteger);
1191 IsLegalWiderType = true;
1195 if (IsLegalWiderType)
1199 case TypeWidenVector:
1200 // Try to widen the vector.
1201 for (unsigned nVT = i + 1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
1202 MVT SVT = (MVT::SimpleValueType) nVT;
1203 if (SVT.getVectorElementType() == EltVT
1204 && SVT.getVectorNumElements() > NElts && isTypeLegal(SVT)) {
1205 TransformToType[i] = SVT;
1206 RegisterTypeForVT[i] = SVT;
1207 NumRegistersForVT[i] = 1;
1208 ValueTypeActions.setTypeAction(VT, TypeWidenVector);
1209 IsLegalWiderType = true;
1213 if (IsLegalWiderType)
1217 case TypeSplitVector:
1218 case TypeScalarizeVector: {
1221 unsigned NumIntermediates;
1222 NumRegistersForVT[i] = getVectorTypeBreakdownMVT(VT, IntermediateVT,
1223 NumIntermediates, RegisterVT, this);
1224 RegisterTypeForVT[i] = RegisterVT;
1226 MVT NVT = VT.getPow2VectorType();
1228 // Type is already a power of 2. The default action is to split.
1229 TransformToType[i] = MVT::Other;
1230 if (PreferredAction == TypeScalarizeVector)
1231 ValueTypeActions.setTypeAction(VT, TypeScalarizeVector);
1232 else if (PreferredAction == TypeSplitVector)
1233 ValueTypeActions.setTypeAction(VT, TypeSplitVector);
1235 // Set type action according to the number of elements.
1236 ValueTypeActions.setTypeAction(VT, NElts == 1 ? TypeScalarizeVector
1239 TransformToType[i] = NVT;
1240 ValueTypeActions.setTypeAction(VT, TypeWidenVector);
1245 llvm_unreachable("Unknown vector legalization action!");
1249 // Determine the 'representative' register class for each value type.
1250 // An representative register class is the largest (meaning one which is
1251 // not a sub-register class / subreg register class) legal register class for
1252 // a group of value types. For example, on i386, i8, i16, and i32
1253 // representative would be GR32; while on x86_64 it's GR64.
1254 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
1255 const TargetRegisterClass* RRC;
1257 std::tie(RRC, Cost) = findRepresentativeClass(TRI, (MVT::SimpleValueType)i);
1258 RepRegClassForVT[i] = RRC;
1259 RepRegClassCostForVT[i] = Cost;
1263 EVT TargetLoweringBase::getSetCCResultType(const DataLayout &DL, LLVMContext &,
1265 assert(!VT.isVector() && "No default SetCC type for vectors!");
1266 return getPointerTy(DL).SimpleTy;
1269 MVT::SimpleValueType TargetLoweringBase::getCmpLibcallReturnType() const {
1270 return MVT::i32; // return the default value
1273 /// getVectorTypeBreakdown - Vector types are broken down into some number of
1274 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32
1275 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
1276 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
1278 /// This method returns the number of registers needed, and the VT for each
1279 /// register. It also returns the VT and quantity of the intermediate values
1280 /// before they are promoted/expanded.
1281 unsigned TargetLoweringBase::getVectorTypeBreakdown(LLVMContext &Context, EVT VT,
1282 EVT &IntermediateVT,
1283 unsigned &NumIntermediates,
1284 MVT &RegisterVT) const {
1285 unsigned NumElts = VT.getVectorNumElements();
1287 // If there is a wider vector type with the same element type as this one,
1288 // or a promoted vector type that has the same number of elements which
1289 // are wider, then we should convert to that legal vector type.
1290 // This handles things like <2 x float> -> <4 x float> and
1291 // <4 x i1> -> <4 x i32>.
1292 LegalizeTypeAction TA = getTypeAction(Context, VT);
1293 if (NumElts != 1 && (TA == TypeWidenVector || TA == TypePromoteInteger)) {
1294 EVT RegisterEVT = getTypeToTransformTo(Context, VT);
1295 if (isTypeLegal(RegisterEVT)) {
1296 IntermediateVT = RegisterEVT;
1297 RegisterVT = RegisterEVT.getSimpleVT();
1298 NumIntermediates = 1;
1303 // Figure out the right, legal destination reg to copy into.
1304 EVT EltTy = VT.getVectorElementType();
1306 unsigned NumVectorRegs = 1;
1308 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
1309 // could break down into LHS/RHS like LegalizeDAG does.
1310 if (!isPowerOf2_32(NumElts)) {
1311 NumVectorRegs = NumElts;
1315 // Divide the input until we get to a supported size. This will always
1316 // end with a scalar if the target doesn't support vectors.
1317 while (NumElts > 1 && !isTypeLegal(
1318 EVT::getVectorVT(Context, EltTy, NumElts))) {
1320 NumVectorRegs <<= 1;
1323 NumIntermediates = NumVectorRegs;
1325 EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts);
1326 if (!isTypeLegal(NewVT))
1328 IntermediateVT = NewVT;
1330 MVT DestVT = getRegisterType(Context, NewVT);
1331 RegisterVT = DestVT;
1332 unsigned NewVTSize = NewVT.getSizeInBits();
1334 // Convert sizes such as i33 to i64.
1335 if (!isPowerOf2_32(NewVTSize))
1336 NewVTSize = NextPowerOf2(NewVTSize);
1338 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16.
1339 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits());
1341 // Otherwise, promotion or legal types use the same number of registers as
1342 // the vector decimated to the appropriate level.
1343 return NumVectorRegs;
1346 /// Get the EVTs and ArgFlags collections that represent the legalized return
1347 /// type of the given function. This does not require a DAG or a return value,
1348 /// and is suitable for use before any DAGs for the function are constructed.
1349 /// TODO: Move this out of TargetLowering.cpp.
1350 void llvm::GetReturnInfo(CallingConv::ID CC, Type *ReturnType,
1352 SmallVectorImpl<ISD::OutputArg> &Outs,
1353 const TargetLowering &TLI, const DataLayout &DL) {
1354 SmallVector<EVT, 4> ValueVTs;
1355 ComputeValueVTs(TLI, DL, ReturnType, ValueVTs);
1356 unsigned NumValues = ValueVTs.size();
1357 if (NumValues == 0) return;
1359 for (unsigned j = 0, f = NumValues; j != f; ++j) {
1360 EVT VT = ValueVTs[j];
1361 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
1363 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt))
1364 ExtendKind = ISD::SIGN_EXTEND;
1365 else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt))
1366 ExtendKind = ISD::ZERO_EXTEND;
1368 // FIXME: C calling convention requires the return type to be promoted to
1369 // at least 32-bit. But this is not necessary for non-C calling
1370 // conventions. The frontend should mark functions whose return values
1371 // require promoting with signext or zeroext attributes.
1372 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) {
1373 MVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32);
1374 if (VT.bitsLT(MinVT))
1379 TLI.getNumRegistersForCallingConv(ReturnType->getContext(), CC, VT);
1381 TLI.getRegisterTypeForCallingConv(ReturnType->getContext(), CC, VT);
1383 // 'inreg' on function refers to return value
1384 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
1385 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::InReg))
1388 // Propagate extension type if any
1389 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt))
1391 else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt))
1394 for (unsigned i = 0; i < NumParts; ++i)
1395 Outs.push_back(ISD::OutputArg(Flags, PartVT, VT, /*isFixed=*/true, 0, 0));
1399 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1400 /// function arguments in the caller parameter area. This is the actual
1401 /// alignment, not its logarithm.
1402 unsigned TargetLoweringBase::getByValTypeAlignment(Type *Ty,
1403 const DataLayout &DL) const {
1404 return DL.getABITypeAlignment(Ty);
1407 bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context,
1408 const DataLayout &DL, EVT VT,
1412 // Check if the specified alignment is sufficient based on the data layout.
1413 // TODO: While using the data layout works in practice, a better solution
1414 // would be to implement this check directly (make this a virtual function).
1415 // For example, the ABI alignment may change based on software platform while
1416 // this function should only be affected by hardware implementation.
1417 Type *Ty = VT.getTypeForEVT(Context);
1418 if (Alignment >= DL.getABITypeAlignment(Ty)) {
1419 // Assume that an access that meets the ABI-specified alignment is fast.
1420 if (Fast != nullptr)
1425 // This is a misaligned access.
1426 return allowsMisalignedMemoryAccesses(VT, AddrSpace, Alignment, Fast);
1429 BranchProbability TargetLoweringBase::getPredictableBranchThreshold() const {
1430 return BranchProbability(MinPercentageForPredictableBranch, 100);
1433 //===----------------------------------------------------------------------===//
1434 // TargetTransformInfo Helpers
1435 //===----------------------------------------------------------------------===//
1437 int TargetLoweringBase::InstructionOpcodeToISD(unsigned Opcode) const {
1438 enum InstructionOpcodes {
1439 #define HANDLE_INST(NUM, OPCODE, CLASS) OPCODE = NUM,
1440 #define LAST_OTHER_INST(NUM) InstructionOpcodesCount = NUM
1441 #include "llvm/IR/Instruction.def"
1443 switch (static_cast<InstructionOpcodes>(Opcode)) {
1446 case Switch: return 0;
1447 case IndirectBr: return 0;
1448 case Invoke: return 0;
1449 case Resume: return 0;
1450 case Unreachable: return 0;
1451 case CleanupRet: return 0;
1452 case CatchRet: return 0;
1453 case CatchPad: return 0;
1454 case CatchSwitch: return 0;
1455 case CleanupPad: return 0;
1456 case FNeg: return ISD::FNEG;
1457 case Add: return ISD::ADD;
1458 case FAdd: return ISD::FADD;
1459 case Sub: return ISD::SUB;
1460 case FSub: return ISD::FSUB;
1461 case Mul: return ISD::MUL;
1462 case FMul: return ISD::FMUL;
1463 case UDiv: return ISD::UDIV;
1464 case SDiv: return ISD::SDIV;
1465 case FDiv: return ISD::FDIV;
1466 case URem: return ISD::UREM;
1467 case SRem: return ISD::SREM;
1468 case FRem: return ISD::FREM;
1469 case Shl: return ISD::SHL;
1470 case LShr: return ISD::SRL;
1471 case AShr: return ISD::SRA;
1472 case And: return ISD::AND;
1473 case Or: return ISD::OR;
1474 case Xor: return ISD::XOR;
1475 case Alloca: return 0;
1476 case Load: return ISD::LOAD;
1477 case Store: return ISD::STORE;
1478 case GetElementPtr: return 0;
1479 case Fence: return 0;
1480 case AtomicCmpXchg: return 0;
1481 case AtomicRMW: return 0;
1482 case Trunc: return ISD::TRUNCATE;
1483 case ZExt: return ISD::ZERO_EXTEND;
1484 case SExt: return ISD::SIGN_EXTEND;
1485 case FPToUI: return ISD::FP_TO_UINT;
1486 case FPToSI: return ISD::FP_TO_SINT;
1487 case UIToFP: return ISD::UINT_TO_FP;
1488 case SIToFP: return ISD::SINT_TO_FP;
1489 case FPTrunc: return ISD::FP_ROUND;
1490 case FPExt: return ISD::FP_EXTEND;
1491 case PtrToInt: return ISD::BITCAST;
1492 case IntToPtr: return ISD::BITCAST;
1493 case BitCast: return ISD::BITCAST;
1494 case AddrSpaceCast: return ISD::ADDRSPACECAST;
1495 case ICmp: return ISD::SETCC;
1496 case FCmp: return ISD::SETCC;
1498 case Call: return 0;
1499 case Select: return ISD::SELECT;
1500 case UserOp1: return 0;
1501 case UserOp2: return 0;
1502 case VAArg: return 0;
1503 case ExtractElement: return ISD::EXTRACT_VECTOR_ELT;
1504 case InsertElement: return ISD::INSERT_VECTOR_ELT;
1505 case ShuffleVector: return ISD::VECTOR_SHUFFLE;
1506 case ExtractValue: return ISD::MERGE_VALUES;
1507 case InsertValue: return ISD::MERGE_VALUES;
1508 case LandingPad: return 0;
1511 llvm_unreachable("Unknown instruction type encountered!");
1515 TargetLoweringBase::getTypeLegalizationCost(const DataLayout &DL,
1517 LLVMContext &C = Ty->getContext();
1518 EVT MTy = getValueType(DL, Ty);
1521 // We keep legalizing the type until we find a legal kind. We assume that
1522 // the only operation that costs anything is the split. After splitting
1523 // we need to handle two types.
1525 LegalizeKind LK = getTypeConversion(C, MTy);
1527 if (LK.first == TypeLegal)
1528 return std::make_pair(Cost, MTy.getSimpleVT());
1530 if (LK.first == TypeSplitVector || LK.first == TypeExpandInteger)
1533 // Do not loop with f128 type.
1534 if (MTy == LK.second)
1535 return std::make_pair(Cost, MTy.getSimpleVT());
1537 // Keep legalizing the type.
1542 Value *TargetLoweringBase::getDefaultSafeStackPointerLocation(IRBuilder<> &IRB,
1543 bool UseTLS) const {
1544 // compiler-rt provides a variable with a magic name. Targets that do not
1545 // link with compiler-rt may also provide such a variable.
1546 Module *M = IRB.GetInsertBlock()->getParent()->getParent();
1547 const char *UnsafeStackPtrVar = "__safestack_unsafe_stack_ptr";
1548 auto UnsafeStackPtr =
1549 dyn_cast_or_null<GlobalVariable>(M->getNamedValue(UnsafeStackPtrVar));
1551 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext());
1553 if (!UnsafeStackPtr) {
1554 auto TLSModel = UseTLS ?
1555 GlobalValue::InitialExecTLSModel :
1556 GlobalValue::NotThreadLocal;
1557 // The global variable is not defined yet, define it ourselves.
1558 // We use the initial-exec TLS model because we do not support the
1559 // variable living anywhere other than in the main executable.
1560 UnsafeStackPtr = new GlobalVariable(
1561 *M, StackPtrTy, false, GlobalValue::ExternalLinkage, nullptr,
1562 UnsafeStackPtrVar, nullptr, TLSModel);
1564 // The variable exists, check its type and attributes.
1565 if (UnsafeStackPtr->getValueType() != StackPtrTy)
1566 report_fatal_error(Twine(UnsafeStackPtrVar) + " must have void* type");
1567 if (UseTLS != UnsafeStackPtr->isThreadLocal())
1568 report_fatal_error(Twine(UnsafeStackPtrVar) + " must " +
1569 (UseTLS ? "" : "not ") + "be thread-local");
1571 return UnsafeStackPtr;
1574 Value *TargetLoweringBase::getSafeStackPointerLocation(IRBuilder<> &IRB) const {
1575 if (!TM.getTargetTriple().isAndroid())
1576 return getDefaultSafeStackPointerLocation(IRB, true);
1578 // Android provides a libc function to retrieve the address of the current
1579 // thread's unsafe stack pointer.
1580 Module *M = IRB.GetInsertBlock()->getParent()->getParent();
1581 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext());
1582 Value *Fn = M->getOrInsertFunction("__safestack_pointer_address",
1583 StackPtrTy->getPointerTo(0));
1584 return IRB.CreateCall(Fn);
1587 //===----------------------------------------------------------------------===//
1588 // Loop Strength Reduction hooks
1589 //===----------------------------------------------------------------------===//
1591 /// isLegalAddressingMode - Return true if the addressing mode represented
1592 /// by AM is legal for this target, for a load/store of the specified type.
1593 bool TargetLoweringBase::isLegalAddressingMode(const DataLayout &DL,
1594 const AddrMode &AM, Type *Ty,
1595 unsigned AS, Instruction *I) const {
1596 // The default implementation of this implements a conservative RISCy, r+r and
1599 // Allows a sign-extended 16-bit immediate field.
1600 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
1603 // No global is ever allowed as a base.
1607 // Only support r+r,
1609 case 0: // "r+i" or just "i", depending on HasBaseReg.
1612 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
1614 // Otherwise we have r+r or r+i.
1617 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
1619 // Allow 2*r as r+r.
1621 default: // Don't allow n * r
1628 //===----------------------------------------------------------------------===//
1630 //===----------------------------------------------------------------------===//
1632 // For OpenBSD return its special guard variable. Otherwise return nullptr,
1633 // so that SelectionDAG handle SSP.
1634 Value *TargetLoweringBase::getIRStackGuard(IRBuilder<> &IRB) const {
1635 if (getTargetMachine().getTargetTriple().isOSOpenBSD()) {
1636 Module &M = *IRB.GetInsertBlock()->getParent()->getParent();
1637 PointerType *PtrTy = Type::getInt8PtrTy(M.getContext());
1638 return M.getOrInsertGlobal("__guard_local", PtrTy);
1643 // Currently only support "standard" __stack_chk_guard.
1644 // TODO: add LOAD_STACK_GUARD support.
1645 void TargetLoweringBase::insertSSPDeclarations(Module &M) const {
1646 if (!M.getNamedValue("__stack_chk_guard"))
1647 new GlobalVariable(M, Type::getInt8PtrTy(M.getContext()), false,
1648 GlobalVariable::ExternalLinkage,
1649 nullptr, "__stack_chk_guard");
1652 // Currently only support "standard" __stack_chk_guard.
1653 // TODO: add LOAD_STACK_GUARD support.
1654 Value *TargetLoweringBase::getSDagStackGuard(const Module &M) const {
1655 return M.getNamedValue("__stack_chk_guard");
1658 Value *TargetLoweringBase::getSSPStackGuardCheck(const Module &M) const {
1662 unsigned TargetLoweringBase::getMinimumJumpTableEntries() const {
1663 return MinimumJumpTableEntries;
1666 void TargetLoweringBase::setMinimumJumpTableEntries(unsigned Val) {
1667 MinimumJumpTableEntries = Val;
1670 unsigned TargetLoweringBase::getMinimumJumpTableDensity(bool OptForSize) const {
1671 return OptForSize ? OptsizeJumpTableDensity : JumpTableDensity;
1674 unsigned TargetLoweringBase::getMaximumJumpTableSize() const {
1675 return MaximumJumpTableSize;
1678 void TargetLoweringBase::setMaximumJumpTableSize(unsigned Val) {
1679 MaximumJumpTableSize = Val;
1682 //===----------------------------------------------------------------------===//
1683 // Reciprocal Estimates
1684 //===----------------------------------------------------------------------===//
1686 /// Get the reciprocal estimate attribute string for a function that will
1687 /// override the target defaults.
1688 static StringRef getRecipEstimateForFunc(MachineFunction &MF) {
1689 const Function &F = MF.getFunction();
1690 return F.getFnAttribute("reciprocal-estimates").getValueAsString();
1693 /// Construct a string for the given reciprocal operation of the given type.
1694 /// This string should match the corresponding option to the front-end's
1695 /// "-mrecip" flag assuming those strings have been passed through in an
1696 /// attribute string. For example, "vec-divf" for a division of a vXf32.
1697 static std::string getReciprocalOpName(bool IsSqrt, EVT VT) {
1698 std::string Name = VT.isVector() ? "vec-" : "";
1700 Name += IsSqrt ? "sqrt" : "div";
1702 // TODO: Handle "half" or other float types?
1703 if (VT.getScalarType() == MVT::f64) {
1706 assert(VT.getScalarType() == MVT::f32 &&
1707 "Unexpected FP type for reciprocal estimate");
1714 /// Return the character position and value (a single numeric character) of a
1715 /// customized refinement operation in the input string if it exists. Return
1716 /// false if there is no customized refinement step count.
1717 static bool parseRefinementStep(StringRef In, size_t &Position,
1719 const char RefStepToken = ':';
1720 Position = In.find(RefStepToken);
1721 if (Position == StringRef::npos)
1724 StringRef RefStepString = In.substr(Position + 1);
1725 // Allow exactly one numeric character for the additional refinement
1727 if (RefStepString.size() == 1) {
1728 char RefStepChar = RefStepString[0];
1729 if (RefStepChar >= '0' && RefStepChar <= '9') {
1730 Value = RefStepChar - '0';
1734 report_fatal_error("Invalid refinement step for -recip.");
1737 /// For the input attribute string, return one of the ReciprocalEstimate enum
1738 /// status values (enabled, disabled, or not specified) for this operation on
1739 /// the specified data type.
1740 static int getOpEnabled(bool IsSqrt, EVT VT, StringRef Override) {
1741 if (Override.empty())
1742 return TargetLoweringBase::ReciprocalEstimate::Unspecified;
1744 SmallVector<StringRef, 4> OverrideVector;
1745 Override.split(OverrideVector, ',');
1746 unsigned NumArgs = OverrideVector.size();
1748 // Check if "all", "none", or "default" was specified.
1750 // Look for an optional setting of the number of refinement steps needed
1751 // for this type of reciprocal operation.
1754 if (parseRefinementStep(Override, RefPos, RefSteps)) {
1755 // Split the string for further processing.
1756 Override = Override.substr(0, RefPos);
1759 // All reciprocal types are enabled.
1760 if (Override == "all")
1761 return TargetLoweringBase::ReciprocalEstimate::Enabled;
1763 // All reciprocal types are disabled.
1764 if (Override == "none")
1765 return TargetLoweringBase::ReciprocalEstimate::Disabled;
1767 // Target defaults for enablement are used.
1768 if (Override == "default")
1769 return TargetLoweringBase::ReciprocalEstimate::Unspecified;
1772 // The attribute string may omit the size suffix ('f'/'d').
1773 std::string VTName = getReciprocalOpName(IsSqrt, VT);
1774 std::string VTNameNoSize = VTName;
1775 VTNameNoSize.pop_back();
1776 static const char DisabledPrefix = '!';
1778 for (StringRef RecipType : OverrideVector) {
1781 if (parseRefinementStep(RecipType, RefPos, RefSteps))
1782 RecipType = RecipType.substr(0, RefPos);
1784 // Ignore the disablement token for string matching.
1785 bool IsDisabled = RecipType[0] == DisabledPrefix;
1787 RecipType = RecipType.substr(1);
1789 if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize))
1790 return IsDisabled ? TargetLoweringBase::ReciprocalEstimate::Disabled
1791 : TargetLoweringBase::ReciprocalEstimate::Enabled;
1794 return TargetLoweringBase::ReciprocalEstimate::Unspecified;
1797 /// For the input attribute string, return the customized refinement step count
1798 /// for this operation on the specified data type. If the step count does not
1799 /// exist, return the ReciprocalEstimate enum value for unspecified.
1800 static int getOpRefinementSteps(bool IsSqrt, EVT VT, StringRef Override) {
1801 if (Override.empty())
1802 return TargetLoweringBase::ReciprocalEstimate::Unspecified;
1804 SmallVector<StringRef, 4> OverrideVector;
1805 Override.split(OverrideVector, ',');
1806 unsigned NumArgs = OverrideVector.size();
1808 // Check if "all", "default", or "none" was specified.
1810 // Look for an optional setting of the number of refinement steps needed
1811 // for this type of reciprocal operation.
1814 if (!parseRefinementStep(Override, RefPos, RefSteps))
1815 return TargetLoweringBase::ReciprocalEstimate::Unspecified;
1817 // Split the string for further processing.
1818 Override = Override.substr(0, RefPos);
1819 assert(Override != "none" &&
1820 "Disabled reciprocals, but specifed refinement steps?");
1822 // If this is a general override, return the specified number of steps.
1823 if (Override == "all" || Override == "default")
1827 // The attribute string may omit the size suffix ('f'/'d').
1828 std::string VTName = getReciprocalOpName(IsSqrt, VT);
1829 std::string VTNameNoSize = VTName;
1830 VTNameNoSize.pop_back();
1832 for (StringRef RecipType : OverrideVector) {
1835 if (!parseRefinementStep(RecipType, RefPos, RefSteps))
1838 RecipType = RecipType.substr(0, RefPos);
1839 if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize))
1843 return TargetLoweringBase::ReciprocalEstimate::Unspecified;
1846 int TargetLoweringBase::getRecipEstimateSqrtEnabled(EVT VT,
1847 MachineFunction &MF) const {
1848 return getOpEnabled(true, VT, getRecipEstimateForFunc(MF));
1851 int TargetLoweringBase::getRecipEstimateDivEnabled(EVT VT,
1852 MachineFunction &MF) const {
1853 return getOpEnabled(false, VT, getRecipEstimateForFunc(MF));
1856 int TargetLoweringBase::getSqrtRefinementSteps(EVT VT,
1857 MachineFunction &MF) const {
1858 return getOpRefinementSteps(true, VT, getRecipEstimateForFunc(MF));
1861 int TargetLoweringBase::getDivRefinementSteps(EVT VT,
1862 MachineFunction &MF) const {
1863 return getOpRefinementSteps(false, VT, getRecipEstimateForFunc(MF));
1866 void TargetLoweringBase::finalizeLowering(MachineFunction &MF) const {
1867 MF.getRegInfo().freezeReservedRegs(MF);