LLVM 19.0.0git
SimplifyLibCalls.cpp
Go to the documentation of this file.
1//===------ SimplifyLibCalls.cpp - Library calls simplifier ---------------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file implements the library calls simplifier. It does not implement
10// any pass, but can't be used by other passes to do simplifications.
11//
12//===----------------------------------------------------------------------===//
13
15#include "llvm/ADT/APSInt.h"
19#include "llvm/Analysis/Loads.h"
23#include "llvm/IR/DataLayout.h"
24#include "llvm/IR/Function.h"
25#include "llvm/IR/IRBuilder.h"
27#include "llvm/IR/Intrinsics.h"
28#include "llvm/IR/Module.h"
37
38#include <cmath>
39
40using namespace llvm;
41using namespace PatternMatch;
42
43static cl::opt<bool>
44 EnableUnsafeFPShrink("enable-double-float-shrink", cl::Hidden,
45 cl::init(false),
46 cl::desc("Enable unsafe double to float "
47 "shrinking for math lib calls"));
48
49// Enable conversion of operator new calls with a MemProf hot or cold hint
50// to an operator new call that takes a hot/cold hint. Off by default since
51// not all allocators currently support this extension.
52static cl::opt<bool>
53 OptimizeHotColdNew("optimize-hot-cold-new", cl::Hidden, cl::init(false),
54 cl::desc("Enable hot/cold operator new library calls"));
56 "optimize-existing-hot-cold-new", cl::Hidden, cl::init(false),
58 "Enable optimization of existing hot/cold operator new library calls"));
59
60namespace {
61
62// Specialized parser to ensure the hint is an 8 bit value (we can't specify
63// uint8_t to opt<> as that is interpreted to mean that we are passing a char
64// option with a specific set of values.
65struct HotColdHintParser : public cl::parser<unsigned> {
66 HotColdHintParser(cl::Option &O) : cl::parser<unsigned>(O) {}
67
68 bool parse(cl::Option &O, StringRef ArgName, StringRef Arg, unsigned &Value) {
69 if (Arg.getAsInteger(0, Value))
70 return O.error("'" + Arg + "' value invalid for uint argument!");
71
72 if (Value > 255)
73 return O.error("'" + Arg + "' value must be in the range [0, 255]!");
74
75 return false;
76 }
77};
78
79} // end anonymous namespace
80
81// Hot/cold operator new takes an 8 bit hotness hint, where 0 is the coldest
82// and 255 is the hottest. Default to 1 value away from the coldest and hottest
83// hints, so that the compiler hinted allocations are slightly less strong than
84// manually inserted hints at the two extremes.
86 "cold-new-hint-value", cl::Hidden, cl::init(1),
87 cl::desc("Value to pass to hot/cold operator new for cold allocation"));
89 NotColdNewHintValue("notcold-new-hint-value", cl::Hidden, cl::init(128),
90 cl::desc("Value to pass to hot/cold operator new for "
91 "notcold (warm) allocation"));
93 "hot-new-hint-value", cl::Hidden, cl::init(254),
94 cl::desc("Value to pass to hot/cold operator new for hot allocation"));
95
96//===----------------------------------------------------------------------===//
97// Helper Functions
98//===----------------------------------------------------------------------===//
99
100static bool ignoreCallingConv(LibFunc Func) {
101 return Func == LibFunc_abs || Func == LibFunc_labs ||
102 Func == LibFunc_llabs || Func == LibFunc_strlen;
103}
104
105/// Return true if it is only used in equality comparisons with With.
107 for (User *U : V->users()) {
108 if (ICmpInst *IC = dyn_cast<ICmpInst>(U))
109 if (IC->isEquality() && IC->getOperand(1) == With)
110 continue;
111 // Unknown instruction.
112 return false;
113 }
114 return true;
115}
116
118 return any_of(CI->operands(), [](const Use &OI) {
119 return OI->getType()->isFloatingPointTy();
120 });
121}
122
123static bool callHasFP128Argument(const CallInst *CI) {
124 return any_of(CI->operands(), [](const Use &OI) {
125 return OI->getType()->isFP128Ty();
126 });
127}
128
129// Convert the entire string Str representing an integer in Base, up to
130// the terminating nul if present, to a constant according to the rules
131// of strtoul[l] or, when AsSigned is set, of strtol[l]. On success
132// return the result, otherwise null.
133// The function assumes the string is encoded in ASCII and carefully
134// avoids converting sequences (including "") that the corresponding
135// library call might fail and set errno for.
136static Value *convertStrToInt(CallInst *CI, StringRef &Str, Value *EndPtr,
137 uint64_t Base, bool AsSigned, IRBuilderBase &B) {
138 if (Base < 2 || Base > 36)
139 if (Base != 0)
140 // Fail for an invalid base (required by POSIX).
141 return nullptr;
142
143 // Current offset into the original string to reflect in EndPtr.
144 size_t Offset = 0;
145 // Strip leading whitespace.
146 for ( ; Offset != Str.size(); ++Offset)
147 if (!isSpace((unsigned char)Str[Offset])) {
148 Str = Str.substr(Offset);
149 break;
150 }
151
152 if (Str.empty())
153 // Fail for empty subject sequences (POSIX allows but doesn't require
154 // strtol[l]/strtoul[l] to fail with EINVAL).
155 return nullptr;
156
157 // Strip but remember the sign.
158 bool Negate = Str[0] == '-';
159 if (Str[0] == '-' || Str[0] == '+') {
160 Str = Str.drop_front();
161 if (Str.empty())
162 // Fail for a sign with nothing after it.
163 return nullptr;
164 ++Offset;
165 }
166
167 // Set Max to the absolute value of the minimum (for signed), or
168 // to the maximum (for unsigned) value representable in the type.
169 Type *RetTy = CI->getType();
170 unsigned NBits = RetTy->getPrimitiveSizeInBits();
171 uint64_t Max = AsSigned && Negate ? 1 : 0;
172 Max += AsSigned ? maxIntN(NBits) : maxUIntN(NBits);
173
174 // Autodetect Base if it's zero and consume the "0x" prefix.
175 if (Str.size() > 1) {
176 if (Str[0] == '0') {
177 if (toUpper((unsigned char)Str[1]) == 'X') {
178 if (Str.size() == 2 || (Base && Base != 16))
179 // Fail if Base doesn't allow the "0x" prefix or for the prefix
180 // alone that implementations like BSD set errno to EINVAL for.
181 return nullptr;
182
183 Str = Str.drop_front(2);
184 Offset += 2;
185 Base = 16;
186 }
187 else if (Base == 0)
188 Base = 8;
189 } else if (Base == 0)
190 Base = 10;
191 }
192 else if (Base == 0)
193 Base = 10;
194
195 // Convert the rest of the subject sequence, not including the sign,
196 // to its uint64_t representation (this assumes the source character
197 // set is ASCII).
198 uint64_t Result = 0;
199 for (unsigned i = 0; i != Str.size(); ++i) {
200 unsigned char DigVal = Str[i];
201 if (isDigit(DigVal))
202 DigVal = DigVal - '0';
203 else {
204 DigVal = toUpper(DigVal);
205 if (isAlpha(DigVal))
206 DigVal = DigVal - 'A' + 10;
207 else
208 return nullptr;
209 }
210
211 if (DigVal >= Base)
212 // Fail if the digit is not valid in the Base.
213 return nullptr;
214
215 // Add the digit and fail if the result is not representable in
216 // the (unsigned form of the) destination type.
217 bool VFlow;
218 Result = SaturatingMultiplyAdd(Result, Base, (uint64_t)DigVal, &VFlow);
219 if (VFlow || Result > Max)
220 return nullptr;
221 }
222
223 if (EndPtr) {
224 // Store the pointer to the end.
225 Value *Off = B.getInt64(Offset + Str.size());
226 Value *StrBeg = CI->getArgOperand(0);
227 Value *StrEnd = B.CreateInBoundsGEP(B.getInt8Ty(), StrBeg, Off, "endptr");
228 B.CreateStore(StrEnd, EndPtr);
229 }
230
231 if (Negate)
232 // Unsigned negation doesn't overflow.
233 Result = -Result;
234
235 return ConstantInt::get(RetTy, Result);
236}
237
239 for (User *U : V->users()) {
240 if (ICmpInst *IC = dyn_cast<ICmpInst>(U))
241 if (Constant *C = dyn_cast<Constant>(IC->getOperand(1)))
242 if (C->isNullValue())
243 continue;
244 // Unknown instruction.
245 return false;
246 }
247 return true;
248}
249
250static bool canTransformToMemCmp(CallInst *CI, Value *Str, uint64_t Len,
251 const DataLayout &DL) {
253 return false;
254
255 if (!isDereferenceableAndAlignedPointer(Str, Align(1), APInt(64, Len), DL))
256 return false;
257
258 if (CI->getFunction()->hasFnAttribute(Attribute::SanitizeMemory))
259 return false;
260
261 return true;
262}
263
265 ArrayRef<unsigned> ArgNos,
266 uint64_t DereferenceableBytes) {
267 const Function *F = CI->getCaller();
268 if (!F)
269 return;
270 for (unsigned ArgNo : ArgNos) {
271 uint64_t DerefBytes = DereferenceableBytes;
272 unsigned AS = CI->getArgOperand(ArgNo)->getType()->getPointerAddressSpace();
273 if (!llvm::NullPointerIsDefined(F, AS) ||
274 CI->paramHasAttr(ArgNo, Attribute::NonNull))
275 DerefBytes = std::max(CI->getParamDereferenceableOrNullBytes(ArgNo),
276 DereferenceableBytes);
277
278 if (CI->getParamDereferenceableBytes(ArgNo) < DerefBytes) {
279 CI->removeParamAttr(ArgNo, Attribute::Dereferenceable);
280 if (!llvm::NullPointerIsDefined(F, AS) ||
281 CI->paramHasAttr(ArgNo, Attribute::NonNull))
282 CI->removeParamAttr(ArgNo, Attribute::DereferenceableOrNull);
284 CI->getContext(), DerefBytes));
285 }
286 }
287}
288
290 ArrayRef<unsigned> ArgNos) {
291 Function *F = CI->getCaller();
292 if (!F)
293 return;
294
295 for (unsigned ArgNo : ArgNos) {
296 if (!CI->paramHasAttr(ArgNo, Attribute::NoUndef))
297 CI->addParamAttr(ArgNo, Attribute::NoUndef);
298
299 if (!CI->paramHasAttr(ArgNo, Attribute::NonNull)) {
300 unsigned AS =
303 continue;
304 CI->addParamAttr(ArgNo, Attribute::NonNull);
305 }
306
307 annotateDereferenceableBytes(CI, ArgNo, 1);
308 }
309}
310
312 Value *Size, const DataLayout &DL) {
313 if (ConstantInt *LenC = dyn_cast<ConstantInt>(Size)) {
315 annotateDereferenceableBytes(CI, ArgNos, LenC->getZExtValue());
316 } else if (isKnownNonZero(Size, DL)) {
318 const APInt *X, *Y;
319 uint64_t DerefMin = 1;
320 if (match(Size, m_Select(m_Value(), m_APInt(X), m_APInt(Y)))) {
321 DerefMin = std::min(X->getZExtValue(), Y->getZExtValue());
322 annotateDereferenceableBytes(CI, ArgNos, DerefMin);
323 }
324 }
325}
326
327// Copy CallInst "flags" like musttail, notail, and tail. Return New param for
328// easier chaining. Calls to emit* and B.createCall should probably be wrapped
329// in this function when New is created to replace Old. Callers should take
330// care to check Old.isMustTailCall() if they aren't replacing Old directly
331// with New.
332static Value *copyFlags(const CallInst &Old, Value *New) {
333 assert(!Old.isMustTailCall() && "do not copy musttail call flags");
334 assert(!Old.isNoTailCall() && "do not copy notail call flags");
335 if (auto *NewCI = dyn_cast_or_null<CallInst>(New))
336 NewCI->setTailCallKind(Old.getTailCallKind());
337 return New;
338}
339
340static Value *mergeAttributesAndFlags(CallInst *NewCI, const CallInst &Old) {
342 NewCI->getContext(), {NewCI->getAttributes(), Old.getAttributes()}));
344 return copyFlags(Old, NewCI);
345}
346
347// Helper to avoid truncating the length if size_t is 32-bits.
349 return Len >= Str.size() ? Str : Str.substr(0, Len);
350}
351
352//===----------------------------------------------------------------------===//
353// String and Memory Library Call Optimizations
354//===----------------------------------------------------------------------===//
355
356Value *LibCallSimplifier::optimizeStrCat(CallInst *CI, IRBuilderBase &B) {
357 // Extract some information from the instruction
358 Value *Dst = CI->getArgOperand(0);
359 Value *Src = CI->getArgOperand(1);
361
362 // See if we can get the length of the input string.
364 if (Len)
366 else
367 return nullptr;
368 --Len; // Unbias length.
369
370 // Handle the simple, do-nothing case: strcat(x, "") -> x
371 if (Len == 0)
372 return Dst;
373
374 return copyFlags(*CI, emitStrLenMemCpy(Src, Dst, Len, B));
375}
376
377Value *LibCallSimplifier::emitStrLenMemCpy(Value *Src, Value *Dst, uint64_t Len,
378 IRBuilderBase &B) {
379 // We need to find the end of the destination string. That's where the
380 // memory is to be moved to. We just generate a call to strlen.
381 Value *DstLen = emitStrLen(Dst, B, DL, TLI);
382 if (!DstLen)
383 return nullptr;
384
385 // Now that we have the destination's length, we must index into the
386 // destination's pointer to get the actual memcpy destination (end of
387 // the string .. we're concatenating).
388 Value *CpyDst = B.CreateInBoundsGEP(B.getInt8Ty(), Dst, DstLen, "endptr");
389
390 // We have enough information to now generate the memcpy call to do the
391 // concatenation for us. Make a memcpy to copy the nul byte with align = 1.
392 B.CreateMemCpy(
393 CpyDst, Align(1), Src, Align(1),
394 ConstantInt::get(DL.getIntPtrType(Src->getContext()), Len + 1));
395 return Dst;
396}
397
398Value *LibCallSimplifier::optimizeStrNCat(CallInst *CI, IRBuilderBase &B) {
399 // Extract some information from the instruction.
400 Value *Dst = CI->getArgOperand(0);
401 Value *Src = CI->getArgOperand(1);
402 Value *Size = CI->getArgOperand(2);
405 if (isKnownNonZero(Size, DL))
407
408 // We don't do anything if length is not constant.
409 ConstantInt *LengthArg = dyn_cast<ConstantInt>(Size);
410 if (LengthArg) {
411 Len = LengthArg->getZExtValue();
412 // strncat(x, c, 0) -> x
413 if (!Len)
414 return Dst;
415 } else {
416 return nullptr;
417 }
418
419 // See if we can get the length of the input string.
420 uint64_t SrcLen = GetStringLength(Src);
421 if (SrcLen) {
422 annotateDereferenceableBytes(CI, 1, SrcLen);
423 --SrcLen; // Unbias length.
424 } else {
425 return nullptr;
426 }
427
428 // strncat(x, "", c) -> x
429 if (SrcLen == 0)
430 return Dst;
431
432 // We don't optimize this case.
433 if (Len < SrcLen)
434 return nullptr;
435
436 // strncat(x, s, c) -> strcat(x, s)
437 // s is constant so the strcat can be optimized further.
438 return copyFlags(*CI, emitStrLenMemCpy(Src, Dst, SrcLen, B));
439}
440
441// Helper to transform memchr(S, C, N) == S to N && *S == C and, when
442// NBytes is null, strchr(S, C) to *S == C. A precondition of the function
443// is that either S is dereferenceable or the value of N is nonzero.
445 IRBuilderBase &B, const DataLayout &DL)
446{
447 Value *Src = CI->getArgOperand(0);
448 Value *CharVal = CI->getArgOperand(1);
449
450 // Fold memchr(A, C, N) == A to N && *A == C.
451 Type *CharTy = B.getInt8Ty();
452 Value *Char0 = B.CreateLoad(CharTy, Src);
453 CharVal = B.CreateTrunc(CharVal, CharTy);
454 Value *Cmp = B.CreateICmpEQ(Char0, CharVal, "char0cmp");
455
456 if (NBytes) {
457 Value *Zero = ConstantInt::get(NBytes->getType(), 0);
458 Value *And = B.CreateICmpNE(NBytes, Zero);
459 Cmp = B.CreateLogicalAnd(And, Cmp);
460 }
461
462 Value *NullPtr = Constant::getNullValue(CI->getType());
463 return B.CreateSelect(Cmp, Src, NullPtr);
464}
465
466Value *LibCallSimplifier::optimizeStrChr(CallInst *CI, IRBuilderBase &B) {
467 Value *SrcStr = CI->getArgOperand(0);
468 Value *CharVal = CI->getArgOperand(1);
470
471 if (isOnlyUsedInEqualityComparison(CI, SrcStr))
472 return memChrToCharCompare(CI, nullptr, B, DL);
473
474 // If the second operand is non-constant, see if we can compute the length
475 // of the input string and turn this into memchr.
476 ConstantInt *CharC = dyn_cast<ConstantInt>(CharVal);
477 if (!CharC) {
478 uint64_t Len = GetStringLength(SrcStr);
479 if (Len)
481 else
482 return nullptr;
483
485 FunctionType *FT = Callee->getFunctionType();
486 unsigned IntBits = TLI->getIntSize();
487 if (!FT->getParamType(1)->isIntegerTy(IntBits)) // memchr needs 'int'.
488 return nullptr;
489
490 unsigned SizeTBits = TLI->getSizeTSize(*CI->getModule());
491 Type *SizeTTy = IntegerType::get(CI->getContext(), SizeTBits);
492 return copyFlags(*CI,
493 emitMemChr(SrcStr, CharVal, // include nul.
494 ConstantInt::get(SizeTTy, Len), B,
495 DL, TLI));
496 }
497
498 if (CharC->isZero()) {
499 Value *NullPtr = Constant::getNullValue(CI->getType());
500 if (isOnlyUsedInEqualityComparison(CI, NullPtr))
501 // Pre-empt the transformation to strlen below and fold
502 // strchr(A, '\0') == null to false.
503 return B.CreateIntToPtr(B.getTrue(), CI->getType());
504 }
505
506 // Otherwise, the character is a constant, see if the first argument is
507 // a string literal. If so, we can constant fold.
508 StringRef Str;
509 if (!getConstantStringInfo(SrcStr, Str)) {
510 if (CharC->isZero()) // strchr(p, 0) -> p + strlen(p)
511 if (Value *StrLen = emitStrLen(SrcStr, B, DL, TLI))
512 return B.CreateInBoundsGEP(B.getInt8Ty(), SrcStr, StrLen, "strchr");
513 return nullptr;
514 }
515
516 // Compute the offset, make sure to handle the case when we're searching for
517 // zero (a weird way to spell strlen).
518 size_t I = (0xFF & CharC->getSExtValue()) == 0
519 ? Str.size()
520 : Str.find(CharC->getSExtValue());
521 if (I == StringRef::npos) // Didn't find the char. strchr returns null.
522 return Constant::getNullValue(CI->getType());
523
524 // strchr(s+n,c) -> gep(s+n+i,c)
525 return B.CreateInBoundsGEP(B.getInt8Ty(), SrcStr, B.getInt64(I), "strchr");
526}
527
528Value *LibCallSimplifier::optimizeStrRChr(CallInst *CI, IRBuilderBase &B) {
529 Value *SrcStr = CI->getArgOperand(0);
530 Value *CharVal = CI->getArgOperand(1);
531 ConstantInt *CharC = dyn_cast<ConstantInt>(CharVal);
533
534 StringRef Str;
535 if (!getConstantStringInfo(SrcStr, Str)) {
536 // strrchr(s, 0) -> strchr(s, 0)
537 if (CharC && CharC->isZero())
538 return copyFlags(*CI, emitStrChr(SrcStr, '\0', B, TLI));
539 return nullptr;
540 }
541
542 unsigned SizeTBits = TLI->getSizeTSize(*CI->getModule());
543 Type *SizeTTy = IntegerType::get(CI->getContext(), SizeTBits);
544
545 // Try to expand strrchr to the memrchr nonstandard extension if it's
546 // available, or simply fail otherwise.
547 uint64_t NBytes = Str.size() + 1; // Include the terminating nul.
548 Value *Size = ConstantInt::get(SizeTTy, NBytes);
549 return copyFlags(*CI, emitMemRChr(SrcStr, CharVal, Size, B, DL, TLI));
550}
551
552Value *LibCallSimplifier::optimizeStrCmp(CallInst *CI, IRBuilderBase &B) {
553 Value *Str1P = CI->getArgOperand(0), *Str2P = CI->getArgOperand(1);
554 if (Str1P == Str2P) // strcmp(x,x) -> 0
555 return ConstantInt::get(CI->getType(), 0);
556
557 StringRef Str1, Str2;
558 bool HasStr1 = getConstantStringInfo(Str1P, Str1);
559 bool HasStr2 = getConstantStringInfo(Str2P, Str2);
560
561 // strcmp(x, y) -> cnst (if both x and y are constant strings)
562 if (HasStr1 && HasStr2)
563 return ConstantInt::get(CI->getType(),
564 std::clamp(Str1.compare(Str2), -1, 1));
565
566 if (HasStr1 && Str1.empty()) // strcmp("", x) -> -*x
567 return B.CreateNeg(B.CreateZExt(
568 B.CreateLoad(B.getInt8Ty(), Str2P, "strcmpload"), CI->getType()));
569
570 if (HasStr2 && Str2.empty()) // strcmp(x,"") -> *x
571 return B.CreateZExt(B.CreateLoad(B.getInt8Ty(), Str1P, "strcmpload"),
572 CI->getType());
573
574 // strcmp(P, "x") -> memcmp(P, "x", 2)
575 uint64_t Len1 = GetStringLength(Str1P);
576 if (Len1)
577 annotateDereferenceableBytes(CI, 0, Len1);
578 uint64_t Len2 = GetStringLength(Str2P);
579 if (Len2)
580 annotateDereferenceableBytes(CI, 1, Len2);
581
582 if (Len1 && Len2) {
583 return copyFlags(
584 *CI, emitMemCmp(Str1P, Str2P,
585 ConstantInt::get(DL.getIntPtrType(CI->getContext()),
586 std::min(Len1, Len2)),
587 B, DL, TLI));
588 }
589
590 // strcmp to memcmp
591 if (!HasStr1 && HasStr2) {
592 if (canTransformToMemCmp(CI, Str1P, Len2, DL))
593 return copyFlags(
594 *CI,
595 emitMemCmp(Str1P, Str2P,
596 ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len2),
597 B, DL, TLI));
598 } else if (HasStr1 && !HasStr2) {
599 if (canTransformToMemCmp(CI, Str2P, Len1, DL))
600 return copyFlags(
601 *CI,
602 emitMemCmp(Str1P, Str2P,
603 ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len1),
604 B, DL, TLI));
605 }
606
608 return nullptr;
609}
610
611// Optimize a memcmp or, when StrNCmp is true, strncmp call CI with constant
612// arrays LHS and RHS and nonconstant Size.
613static Value *optimizeMemCmpVarSize(CallInst *CI, Value *LHS, Value *RHS,
614 Value *Size, bool StrNCmp,
615 IRBuilderBase &B, const DataLayout &DL);
616
617Value *LibCallSimplifier::optimizeStrNCmp(CallInst *CI, IRBuilderBase &B) {
618 Value *Str1P = CI->getArgOperand(0);
619 Value *Str2P = CI->getArgOperand(1);
620 Value *Size = CI->getArgOperand(2);
621 if (Str1P == Str2P) // strncmp(x,x,n) -> 0
622 return ConstantInt::get(CI->getType(), 0);
623
624 if (isKnownNonZero(Size, DL))
626 // Get the length argument if it is constant.
628 if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(Size))
629 Length = LengthArg->getZExtValue();
630 else
631 return optimizeMemCmpVarSize(CI, Str1P, Str2P, Size, true, B, DL);
632
633 if (Length == 0) // strncmp(x,y,0) -> 0
634 return ConstantInt::get(CI->getType(), 0);
635
636 if (Length == 1) // strncmp(x,y,1) -> memcmp(x,y,1)
637 return copyFlags(*CI, emitMemCmp(Str1P, Str2P, Size, B, DL, TLI));
638
639 StringRef Str1, Str2;
640 bool HasStr1 = getConstantStringInfo(Str1P, Str1);
641 bool HasStr2 = getConstantStringInfo(Str2P, Str2);
642
643 // strncmp(x, y) -> cnst (if both x and y are constant strings)
644 if (HasStr1 && HasStr2) {
645 // Avoid truncating the 64-bit Length to 32 bits in ILP32.
646 StringRef SubStr1 = substr(Str1, Length);
647 StringRef SubStr2 = substr(Str2, Length);
648 return ConstantInt::get(CI->getType(),
649 std::clamp(SubStr1.compare(SubStr2), -1, 1));
650 }
651
652 if (HasStr1 && Str1.empty()) // strncmp("", x, n) -> -*x
653 return B.CreateNeg(B.CreateZExt(
654 B.CreateLoad(B.getInt8Ty(), Str2P, "strcmpload"), CI->getType()));
655
656 if (HasStr2 && Str2.empty()) // strncmp(x, "", n) -> *x
657 return B.CreateZExt(B.CreateLoad(B.getInt8Ty(), Str1P, "strcmpload"),
658 CI->getType());
659
660 uint64_t Len1 = GetStringLength(Str1P);
661 if (Len1)
662 annotateDereferenceableBytes(CI, 0, Len1);
663 uint64_t Len2 = GetStringLength(Str2P);
664 if (Len2)
665 annotateDereferenceableBytes(CI, 1, Len2);
666
667 // strncmp to memcmp
668 if (!HasStr1 && HasStr2) {
669 Len2 = std::min(Len2, Length);
670 if (canTransformToMemCmp(CI, Str1P, Len2, DL))
671 return copyFlags(
672 *CI,
673 emitMemCmp(Str1P, Str2P,
674 ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len2),
675 B, DL, TLI));
676 } else if (HasStr1 && !HasStr2) {
677 Len1 = std::min(Len1, Length);
678 if (canTransformToMemCmp(CI, Str2P, Len1, DL))
679 return copyFlags(
680 *CI,
681 emitMemCmp(Str1P, Str2P,
682 ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len1),
683 B, DL, TLI));
684 }
685
686 return nullptr;
687}
688
689Value *LibCallSimplifier::optimizeStrNDup(CallInst *CI, IRBuilderBase &B) {
690 Value *Src = CI->getArgOperand(0);
691 ConstantInt *Size = dyn_cast<ConstantInt>(CI->getArgOperand(1));
692 uint64_t SrcLen = GetStringLength(Src);
693 if (SrcLen && Size) {
694 annotateDereferenceableBytes(CI, 0, SrcLen);
695 if (SrcLen <= Size->getZExtValue() + 1)
696 return copyFlags(*CI, emitStrDup(Src, B, TLI));
697 }
698
699 return nullptr;
700}
701
702Value *LibCallSimplifier::optimizeStrCpy(CallInst *CI, IRBuilderBase &B) {
703 Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1);
704 if (Dst == Src) // strcpy(x,x) -> x
705 return Src;
706
708 // See if we can get the length of the input string.
710 if (Len)
712 else
713 return nullptr;
714
715 // We have enough information to now generate the memcpy call to do the
716 // copy for us. Make a memcpy to copy the nul byte with align = 1.
717 CallInst *NewCI =
718 B.CreateMemCpy(Dst, Align(1), Src, Align(1),
719 ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len));
720 mergeAttributesAndFlags(NewCI, *CI);
721 return Dst;
722}
723
724Value *LibCallSimplifier::optimizeStpCpy(CallInst *CI, IRBuilderBase &B) {
726 Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1);
727
728 // stpcpy(d,s) -> strcpy(d,s) if the result is not used.
729 if (CI->use_empty())
730 return copyFlags(*CI, emitStrCpy(Dst, Src, B, TLI));
731
732 if (Dst == Src) { // stpcpy(x,x) -> x+strlen(x)
733 Value *StrLen = emitStrLen(Src, B, DL, TLI);
734 return StrLen ? B.CreateInBoundsGEP(B.getInt8Ty(), Dst, StrLen) : nullptr;
735 }
736
737 // See if we can get the length of the input string.
739 if (Len)
741 else
742 return nullptr;
743
744 Type *PT = Callee->getFunctionType()->getParamType(0);
745 Value *LenV = ConstantInt::get(DL.getIntPtrType(PT), Len);
746 Value *DstEnd = B.CreateInBoundsGEP(
747 B.getInt8Ty(), Dst, ConstantInt::get(DL.getIntPtrType(PT), Len - 1));
748
749 // We have enough information to now generate the memcpy call to do the
750 // copy for us. Make a memcpy to copy the nul byte with align = 1.
751 CallInst *NewCI = B.CreateMemCpy(Dst, Align(1), Src, Align(1), LenV);
752 mergeAttributesAndFlags(NewCI, *CI);
753 return DstEnd;
754}
755
756// Optimize a call to size_t strlcpy(char*, const char*, size_t).
757
758Value *LibCallSimplifier::optimizeStrLCpy(CallInst *CI, IRBuilderBase &B) {
759 Value *Size = CI->getArgOperand(2);
760 if (isKnownNonZero(Size, DL))
761 // Like snprintf, the function stores into the destination only when
762 // the size argument is nonzero.
764 // The function reads the source argument regardless of Size (it returns
765 // its length).
767
768 uint64_t NBytes;
769 if (ConstantInt *SizeC = dyn_cast<ConstantInt>(Size))
770 NBytes = SizeC->getZExtValue();
771 else
772 return nullptr;
773
774 Value *Dst = CI->getArgOperand(0);
775 Value *Src = CI->getArgOperand(1);
776 if (NBytes <= 1) {
777 if (NBytes == 1)
778 // For a call to strlcpy(D, S, 1) first store a nul in *D.
779 B.CreateStore(B.getInt8(0), Dst);
780
781 // Transform strlcpy(D, S, 0) to a call to strlen(S).
782 return copyFlags(*CI, emitStrLen(Src, B, DL, TLI));
783 }
784
785 // Try to determine the length of the source, substituting its size
786 // when it's not nul-terminated (as it's required to be) to avoid
787 // reading past its end.
788 StringRef Str;
789 if (!getConstantStringInfo(Src, Str, /*TrimAtNul=*/false))
790 return nullptr;
791
792 uint64_t SrcLen = Str.find('\0');
793 // Set if the terminating nul should be copied by the call to memcpy
794 // below.
795 bool NulTerm = SrcLen < NBytes;
796
797 if (NulTerm)
798 // Overwrite NBytes with the number of bytes to copy, including
799 // the terminating nul.
800 NBytes = SrcLen + 1;
801 else {
802 // Set the length of the source for the function to return to its
803 // size, and cap NBytes at the same.
804 SrcLen = std::min(SrcLen, uint64_t(Str.size()));
805 NBytes = std::min(NBytes - 1, SrcLen);
806 }
807
808 if (SrcLen == 0) {
809 // Transform strlcpy(D, "", N) to (*D = '\0, 0).
810 B.CreateStore(B.getInt8(0), Dst);
811 return ConstantInt::get(CI->getType(), 0);
812 }
813
815 Type *PT = Callee->getFunctionType()->getParamType(0);
816 // Transform strlcpy(D, S, N) to memcpy(D, S, N') where N' is the lower
817 // bound on strlen(S) + 1 and N, optionally followed by a nul store to
818 // D[N' - 1] if necessary.
819 CallInst *NewCI = B.CreateMemCpy(Dst, Align(1), Src, Align(1),
820 ConstantInt::get(DL.getIntPtrType(PT), NBytes));
821 mergeAttributesAndFlags(NewCI, *CI);
822
823 if (!NulTerm) {
824 Value *EndOff = ConstantInt::get(CI->getType(), NBytes);
825 Value *EndPtr = B.CreateInBoundsGEP(B.getInt8Ty(), Dst, EndOff);
826 B.CreateStore(B.getInt8(0), EndPtr);
827 }
828
829 // Like snprintf, strlcpy returns the number of nonzero bytes that would
830 // have been copied if the bound had been sufficiently big (which in this
831 // case is strlen(Src)).
832 return ConstantInt::get(CI->getType(), SrcLen);
833}
834
835// Optimize a call CI to either stpncpy when RetEnd is true, or to strncpy
836// otherwise.
837Value *LibCallSimplifier::optimizeStringNCpy(CallInst *CI, bool RetEnd,
838 IRBuilderBase &B) {
840 Value *Dst = CI->getArgOperand(0);
841 Value *Src = CI->getArgOperand(1);
842 Value *Size = CI->getArgOperand(2);
843
844 if (isKnownNonZero(Size, DL)) {
845 // Both st{p,r}ncpy(D, S, N) access the source and destination arrays
846 // only when N is nonzero.
849 }
850
851 // If the "bound" argument is known set N to it. Otherwise set it to
852 // UINT64_MAX and handle it later.
854 if (ConstantInt *SizeC = dyn_cast<ConstantInt>(Size))
855 N = SizeC->getZExtValue();
856
857 if (N == 0)
858 // Fold st{p,r}ncpy(D, S, 0) to D.
859 return Dst;
860
861 if (N == 1) {
862 Type *CharTy = B.getInt8Ty();
863 Value *CharVal = B.CreateLoad(CharTy, Src, "stxncpy.char0");
864 B.CreateStore(CharVal, Dst);
865 if (!RetEnd)
866 // Transform strncpy(D, S, 1) to return (*D = *S), D.
867 return Dst;
868
869 // Transform stpncpy(D, S, 1) to return (*D = *S) ? D + 1 : D.
870 Value *ZeroChar = ConstantInt::get(CharTy, 0);
871 Value *Cmp = B.CreateICmpEQ(CharVal, ZeroChar, "stpncpy.char0cmp");
872
873 Value *Off1 = B.getInt32(1);
874 Value *EndPtr = B.CreateInBoundsGEP(CharTy, Dst, Off1, "stpncpy.end");
875 return B.CreateSelect(Cmp, Dst, EndPtr, "stpncpy.sel");
876 }
877
878 // If the length of the input string is known set SrcLen to it.
879 uint64_t SrcLen = GetStringLength(Src);
880 if (SrcLen)
881 annotateDereferenceableBytes(CI, 1, SrcLen);
882 else
883 return nullptr;
884
885 --SrcLen; // Unbias length.
886
887 if (SrcLen == 0) {
888 // Transform st{p,r}ncpy(D, "", N) to memset(D, '\0', N) for any N.
889 Align MemSetAlign =
891 CallInst *NewCI = B.CreateMemSet(Dst, B.getInt8('\0'), Size, MemSetAlign);
892 AttrBuilder ArgAttrs(CI->getContext(), CI->getAttributes().getParamAttrs(0));
894 CI->getContext(), 0, ArgAttrs));
895 copyFlags(*CI, NewCI);
896 return Dst;
897 }
898
899 if (N > SrcLen + 1) {
900 if (N > 128)
901 // Bail if N is large or unknown.
902 return nullptr;
903
904 // st{p,r}ncpy(D, "a", N) -> memcpy(D, "a\0\0\0", N) for N <= 128.
905 StringRef Str;
906 if (!getConstantStringInfo(Src, Str))
907 return nullptr;
908 std::string SrcStr = Str.str();
909 // Create a bigger, nul-padded array with the same length, SrcLen,
910 // as the original string.
911 SrcStr.resize(N, '\0');
912 Src = B.CreateGlobalString(SrcStr, "str");
913 }
914
915 Type *PT = Callee->getFunctionType()->getParamType(0);
916 // st{p,r}ncpy(D, S, N) -> memcpy(align 1 D, align 1 S, N) when both
917 // S and N are constant.
918 CallInst *NewCI = B.CreateMemCpy(Dst, Align(1), Src, Align(1),
919 ConstantInt::get(DL.getIntPtrType(PT), N));
920 mergeAttributesAndFlags(NewCI, *CI);
921 if (!RetEnd)
922 return Dst;
923
924 // stpncpy(D, S, N) returns the address of the first null in D if it writes
925 // one, otherwise D + N.
926 Value *Off = B.getInt64(std::min(SrcLen, N));
927 return B.CreateInBoundsGEP(B.getInt8Ty(), Dst, Off, "endptr");
928}
929
930Value *LibCallSimplifier::optimizeStringLength(CallInst *CI, IRBuilderBase &B,
931 unsigned CharSize,
932 Value *Bound) {
933 Value *Src = CI->getArgOperand(0);
934 Type *CharTy = B.getIntNTy(CharSize);
935
937 (!Bound || isKnownNonZero(Bound, DL))) {
938 // Fold strlen:
939 // strlen(x) != 0 --> *x != 0
940 // strlen(x) == 0 --> *x == 0
941 // and likewise strnlen with constant N > 0:
942 // strnlen(x, N) != 0 --> *x != 0
943 // strnlen(x, N) == 0 --> *x == 0
944 return B.CreateZExt(B.CreateLoad(CharTy, Src, "char0"),
945 CI->getType());
946 }
947
948 if (Bound) {
949 if (ConstantInt *BoundCst = dyn_cast<ConstantInt>(Bound)) {
950 if (BoundCst->isZero())
951 // Fold strnlen(s, 0) -> 0 for any s, constant or otherwise.
952 return ConstantInt::get(CI->getType(), 0);
953
954 if (BoundCst->isOne()) {
955 // Fold strnlen(s, 1) -> *s ? 1 : 0 for any s.
956 Value *CharVal = B.CreateLoad(CharTy, Src, "strnlen.char0");
957 Value *ZeroChar = ConstantInt::get(CharTy, 0);
958 Value *Cmp = B.CreateICmpNE(CharVal, ZeroChar, "strnlen.char0cmp");
959 return B.CreateZExt(Cmp, CI->getType());
960 }
961 }
962 }
963
964 if (uint64_t Len = GetStringLength(Src, CharSize)) {
965 Value *LenC = ConstantInt::get(CI->getType(), Len - 1);
966 // Fold strlen("xyz") -> 3 and strnlen("xyz", 2) -> 2
967 // and strnlen("xyz", Bound) -> min(3, Bound) for nonconstant Bound.
968 if (Bound)
969 return B.CreateBinaryIntrinsic(Intrinsic::umin, LenC, Bound);
970 return LenC;
971 }
972
973 if (Bound)
974 // Punt for strnlen for now.
975 return nullptr;
976
977 // If s is a constant pointer pointing to a string literal, we can fold
978 // strlen(s + x) to strlen(s) - x, when x is known to be in the range
979 // [0, strlen(s)] or the string has a single null terminator '\0' at the end.
980 // We only try to simplify strlen when the pointer s points to an array
981 // of CharSize elements. Otherwise, we would need to scale the offset x before
982 // doing the subtraction. This will make the optimization more complex, and
983 // it's not very useful because calling strlen for a pointer of other types is
984 // very uncommon.
985 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Src)) {
986 // TODO: Handle subobjects.
987 if (!isGEPBasedOnPointerToString(GEP, CharSize))
988 return nullptr;
989
991 if (getConstantDataArrayInfo(GEP->getOperand(0), Slice, CharSize)) {
992 uint64_t NullTermIdx;
993 if (Slice.Array == nullptr) {
994 NullTermIdx = 0;
995 } else {
996 NullTermIdx = ~((uint64_t)0);
997 for (uint64_t I = 0, E = Slice.Length; I < E; ++I) {
998 if (Slice.Array->getElementAsInteger(I + Slice.Offset) == 0) {
999 NullTermIdx = I;
1000 break;
1001 }
1002 }
1003 // If the string does not have '\0', leave it to strlen to compute
1004 // its length.
1005 if (NullTermIdx == ~((uint64_t)0))
1006 return nullptr;
1007 }
1008
1009 Value *Offset = GEP->getOperand(2);
1010 KnownBits Known = computeKnownBits(Offset, DL, 0, nullptr, CI, nullptr);
1011 uint64_t ArrSize =
1012 cast<ArrayType>(GEP->getSourceElementType())->getNumElements();
1013
1014 // If Offset is not provably in the range [0, NullTermIdx], we can still
1015 // optimize if we can prove that the program has undefined behavior when
1016 // Offset is outside that range. That is the case when GEP->getOperand(0)
1017 // is a pointer to an object whose memory extent is NullTermIdx+1.
1018 if ((Known.isNonNegative() && Known.getMaxValue().ule(NullTermIdx)) ||
1019 (isa<GlobalVariable>(GEP->getOperand(0)) &&
1020 NullTermIdx == ArrSize - 1)) {
1021 Offset = B.CreateSExtOrTrunc(Offset, CI->getType());
1022 return B.CreateSub(ConstantInt::get(CI->getType(), NullTermIdx),
1023 Offset);
1024 }
1025 }
1026 }
1027
1028 // strlen(x?"foo":"bars") --> x ? 3 : 4
1029 if (SelectInst *SI = dyn_cast<SelectInst>(Src)) {
1030 uint64_t LenTrue = GetStringLength(SI->getTrueValue(), CharSize);
1031 uint64_t LenFalse = GetStringLength(SI->getFalseValue(), CharSize);
1032 if (LenTrue && LenFalse) {
1033 ORE.emit([&]() {
1034 return OptimizationRemark("instcombine", "simplify-libcalls", CI)
1035 << "folded strlen(select) to select of constants";
1036 });
1037 return B.CreateSelect(SI->getCondition(),
1038 ConstantInt::get(CI->getType(), LenTrue - 1),
1039 ConstantInt::get(CI->getType(), LenFalse - 1));
1040 }
1041 }
1042
1043 return nullptr;
1044}
1045
1046Value *LibCallSimplifier::optimizeStrLen(CallInst *CI, IRBuilderBase &B) {
1047 if (Value *V = optimizeStringLength(CI, B, 8))
1048 return V;
1050 return nullptr;
1051}
1052
1053Value *LibCallSimplifier::optimizeStrNLen(CallInst *CI, IRBuilderBase &B) {
1054 Value *Bound = CI->getArgOperand(1);
1055 if (Value *V = optimizeStringLength(CI, B, 8, Bound))
1056 return V;
1057
1058 if (isKnownNonZero(Bound, DL))
1060 return nullptr;
1061}
1062
1063Value *LibCallSimplifier::optimizeWcslen(CallInst *CI, IRBuilderBase &B) {
1064 Module &M = *CI->getModule();
1065 unsigned WCharSize = TLI->getWCharSize(M) * 8;
1066 // We cannot perform this optimization without wchar_size metadata.
1067 if (WCharSize == 0)
1068 return nullptr;
1069
1070 return optimizeStringLength(CI, B, WCharSize);
1071}
1072
1073Value *LibCallSimplifier::optimizeStrPBrk(CallInst *CI, IRBuilderBase &B) {
1074 StringRef S1, S2;
1075 bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1);
1076 bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2);
1077
1078 // strpbrk(s, "") -> nullptr
1079 // strpbrk("", s) -> nullptr
1080 if ((HasS1 && S1.empty()) || (HasS2 && S2.empty()))
1081 return Constant::getNullValue(CI->getType());
1082
1083 // Constant folding.
1084 if (HasS1 && HasS2) {
1085 size_t I = S1.find_first_of(S2);
1086 if (I == StringRef::npos) // No match.
1087 return Constant::getNullValue(CI->getType());
1088
1089 return B.CreateInBoundsGEP(B.getInt8Ty(), CI->getArgOperand(0),
1090 B.getInt64(I), "strpbrk");
1091 }
1092
1093 // strpbrk(s, "a") -> strchr(s, 'a')
1094 if (HasS2 && S2.size() == 1)
1095 return copyFlags(*CI, emitStrChr(CI->getArgOperand(0), S2[0], B, TLI));
1096
1097 return nullptr;
1098}
1099
1100Value *LibCallSimplifier::optimizeStrTo(CallInst *CI, IRBuilderBase &B) {
1101 Value *EndPtr = CI->getArgOperand(1);
1102 if (isa<ConstantPointerNull>(EndPtr)) {
1103 // With a null EndPtr, this function won't capture the main argument.
1104 // It would be readonly too, except that it still may write to errno.
1105 CI->addParamAttr(0, Attribute::NoCapture);
1106 }
1107
1108 return nullptr;
1109}
1110
1111Value *LibCallSimplifier::optimizeStrSpn(CallInst *CI, IRBuilderBase &B) {
1112 StringRef S1, S2;
1113 bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1);
1114 bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2);
1115
1116 // strspn(s, "") -> 0
1117 // strspn("", s) -> 0
1118 if ((HasS1 && S1.empty()) || (HasS2 && S2.empty()))
1119 return Constant::getNullValue(CI->getType());
1120
1121 // Constant folding.
1122 if (HasS1 && HasS2) {
1123 size_t Pos = S1.find_first_not_of(S2);
1124 if (Pos == StringRef::npos)
1125 Pos = S1.size();
1126 return ConstantInt::get(CI->getType(), Pos);
1127 }
1128
1129 return nullptr;
1130}
1131
1132Value *LibCallSimplifier::optimizeStrCSpn(CallInst *CI, IRBuilderBase &B) {
1133 StringRef S1, S2;
1134 bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1);
1135 bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2);
1136
1137 // strcspn("", s) -> 0
1138 if (HasS1 && S1.empty())
1139 return Constant::getNullValue(CI->getType());
1140
1141 // Constant folding.
1142 if (HasS1 && HasS2) {
1143 size_t Pos = S1.find_first_of(S2);
1144 if (Pos == StringRef::npos)
1145 Pos = S1.size();
1146 return ConstantInt::get(CI->getType(), Pos);
1147 }
1148
1149 // strcspn(s, "") -> strlen(s)
1150 if (HasS2 && S2.empty())
1151 return copyFlags(*CI, emitStrLen(CI->getArgOperand(0), B, DL, TLI));
1152
1153 return nullptr;
1154}
1155
1156Value *LibCallSimplifier::optimizeStrStr(CallInst *CI, IRBuilderBase &B) {
1157 // fold strstr(x, x) -> x.
1158 if (CI->getArgOperand(0) == CI->getArgOperand(1))
1159 return CI->getArgOperand(0);
1160
1161 // fold strstr(a, b) == a -> strncmp(a, b, strlen(b)) == 0
1163 Value *StrLen = emitStrLen(CI->getArgOperand(1), B, DL, TLI);
1164 if (!StrLen)
1165 return nullptr;
1166 Value *StrNCmp = emitStrNCmp(CI->getArgOperand(0), CI->getArgOperand(1),
1167 StrLen, B, DL, TLI);
1168 if (!StrNCmp)
1169 return nullptr;
1170 for (User *U : llvm::make_early_inc_range(CI->users())) {
1171 ICmpInst *Old = cast<ICmpInst>(U);
1172 Value *Cmp =
1173 B.CreateICmp(Old->getPredicate(), StrNCmp,
1174 ConstantInt::getNullValue(StrNCmp->getType()), "cmp");
1175 replaceAllUsesWith(Old, Cmp);
1176 }
1177 return CI;
1178 }
1179
1180 // See if either input string is a constant string.
1181 StringRef SearchStr, ToFindStr;
1182 bool HasStr1 = getConstantStringInfo(CI->getArgOperand(0), SearchStr);
1183 bool HasStr2 = getConstantStringInfo(CI->getArgOperand(1), ToFindStr);
1184
1185 // fold strstr(x, "") -> x.
1186 if (HasStr2 && ToFindStr.empty())
1187 return CI->getArgOperand(0);
1188
1189 // If both strings are known, constant fold it.
1190 if (HasStr1 && HasStr2) {
1191 size_t Offset = SearchStr.find(ToFindStr);
1192
1193 if (Offset == StringRef::npos) // strstr("foo", "bar") -> null
1194 return Constant::getNullValue(CI->getType());
1195
1196 // strstr("abcd", "bc") -> gep((char*)"abcd", 1)
1197 return B.CreateConstInBoundsGEP1_64(B.getInt8Ty(), CI->getArgOperand(0),
1198 Offset, "strstr");
1199 }
1200
1201 // fold strstr(x, "y") -> strchr(x, 'y').
1202 if (HasStr2 && ToFindStr.size() == 1) {
1203 return emitStrChr(CI->getArgOperand(0), ToFindStr[0], B, TLI);
1204 }
1205
1207 return nullptr;
1208}
1209
1210Value *LibCallSimplifier::optimizeMemRChr(CallInst *CI, IRBuilderBase &B) {
1211 Value *SrcStr = CI->getArgOperand(0);
1212 Value *Size = CI->getArgOperand(2);
1214 Value *CharVal = CI->getArgOperand(1);
1215 ConstantInt *LenC = dyn_cast<ConstantInt>(Size);
1216 Value *NullPtr = Constant::getNullValue(CI->getType());
1217
1218 if (LenC) {
1219 if (LenC->isZero())
1220 // Fold memrchr(x, y, 0) --> null.
1221 return NullPtr;
1222
1223 if (LenC->isOne()) {
1224 // Fold memrchr(x, y, 1) --> *x == y ? x : null for any x and y,
1225 // constant or otherwise.
1226 Value *Val = B.CreateLoad(B.getInt8Ty(), SrcStr, "memrchr.char0");
1227 // Slice off the character's high end bits.
1228 CharVal = B.CreateTrunc(CharVal, B.getInt8Ty());
1229 Value *Cmp = B.CreateICmpEQ(Val, CharVal, "memrchr.char0cmp");
1230 return B.CreateSelect(Cmp, SrcStr, NullPtr, "memrchr.sel");
1231 }
1232 }
1233
1234 StringRef Str;
1235 if (!getConstantStringInfo(SrcStr, Str, /*TrimAtNul=*/false))
1236 return nullptr;
1237
1238 if (Str.size() == 0)
1239 // If the array is empty fold memrchr(A, C, N) to null for any value
1240 // of C and N on the basis that the only valid value of N is zero
1241 // (otherwise the call is undefined).
1242 return NullPtr;
1243
1244 uint64_t EndOff = UINT64_MAX;
1245 if (LenC) {
1246 EndOff = LenC->getZExtValue();
1247 if (Str.size() < EndOff)
1248 // Punt out-of-bounds accesses to sanitizers and/or libc.
1249 return nullptr;
1250 }
1251
1252 if (ConstantInt *CharC = dyn_cast<ConstantInt>(CharVal)) {
1253 // Fold memrchr(S, C, N) for a constant C.
1254 size_t Pos = Str.rfind(CharC->getZExtValue(), EndOff);
1255 if (Pos == StringRef::npos)
1256 // When the character is not in the source array fold the result
1257 // to null regardless of Size.
1258 return NullPtr;
1259
1260 if (LenC)
1261 // Fold memrchr(s, c, N) --> s + Pos for constant N > Pos.
1262 return B.CreateInBoundsGEP(B.getInt8Ty(), SrcStr, B.getInt64(Pos));
1263
1264 if (Str.find(Str[Pos]) == Pos) {
1265 // When there is just a single occurrence of C in S, i.e., the one
1266 // in Str[Pos], fold
1267 // memrchr(s, c, N) --> N <= Pos ? null : s + Pos
1268 // for nonconstant N.
1269 Value *Cmp = B.CreateICmpULE(Size, ConstantInt::get(Size->getType(), Pos),
1270 "memrchr.cmp");
1271 Value *SrcPlus = B.CreateInBoundsGEP(B.getInt8Ty(), SrcStr,
1272 B.getInt64(Pos), "memrchr.ptr_plus");
1273 return B.CreateSelect(Cmp, NullPtr, SrcPlus, "memrchr.sel");
1274 }
1275 }
1276
1277 // Truncate the string to search at most EndOff characters.
1278 Str = Str.substr(0, EndOff);
1279 if (Str.find_first_not_of(Str[0]) != StringRef::npos)
1280 return nullptr;
1281
1282 // If the source array consists of all equal characters, then for any
1283 // C and N (whether in bounds or not), fold memrchr(S, C, N) to
1284 // N != 0 && *S == C ? S + N - 1 : null
1285 Type *SizeTy = Size->getType();
1286 Type *Int8Ty = B.getInt8Ty();
1287 Value *NNeZ = B.CreateICmpNE(Size, ConstantInt::get(SizeTy, 0));
1288 // Slice off the sought character's high end bits.
1289 CharVal = B.CreateTrunc(CharVal, Int8Ty);
1290 Value *CEqS0 = B.CreateICmpEQ(ConstantInt::get(Int8Ty, Str[0]), CharVal);
1291 Value *And = B.CreateLogicalAnd(NNeZ, CEqS0);
1292 Value *SizeM1 = B.CreateSub(Size, ConstantInt::get(SizeTy, 1));
1293 Value *SrcPlus =
1294 B.CreateInBoundsGEP(Int8Ty, SrcStr, SizeM1, "memrchr.ptr_plus");
1295 return B.CreateSelect(And, SrcPlus, NullPtr, "memrchr.sel");
1296}
1297
1298Value *LibCallSimplifier::optimizeMemChr(CallInst *CI, IRBuilderBase &B) {
1299 Value *SrcStr = CI->getArgOperand(0);
1300 Value *Size = CI->getArgOperand(2);
1301
1302 if (isKnownNonZero(Size, DL)) {
1304 if (isOnlyUsedInEqualityComparison(CI, SrcStr))
1305 return memChrToCharCompare(CI, Size, B, DL);
1306 }
1307
1308 Value *CharVal = CI->getArgOperand(1);
1309 ConstantInt *CharC = dyn_cast<ConstantInt>(CharVal);
1310 ConstantInt *LenC = dyn_cast<ConstantInt>(Size);
1311 Value *NullPtr = Constant::getNullValue(CI->getType());
1312
1313 // memchr(x, y, 0) -> null
1314 if (LenC) {
1315 if (LenC->isZero())
1316 return NullPtr;
1317
1318 if (LenC->isOne()) {
1319 // Fold memchr(x, y, 1) --> *x == y ? x : null for any x and y,
1320 // constant or otherwise.
1321 Value *Val = B.CreateLoad(B.getInt8Ty(), SrcStr, "memchr.char0");
1322 // Slice off the character's high end bits.
1323 CharVal = B.CreateTrunc(CharVal, B.getInt8Ty());
1324 Value *Cmp = B.CreateICmpEQ(Val, CharVal, "memchr.char0cmp");
1325 return B.CreateSelect(Cmp, SrcStr, NullPtr, "memchr.sel");
1326 }
1327 }
1328
1329 StringRef Str;
1330 if (!getConstantStringInfo(SrcStr, Str, /*TrimAtNul=*/false))
1331 return nullptr;
1332
1333 if (CharC) {
1334 size_t Pos = Str.find(CharC->getZExtValue());
1335 if (Pos == StringRef::npos)
1336 // When the character is not in the source array fold the result
1337 // to null regardless of Size.
1338 return NullPtr;
1339
1340 // Fold memchr(s, c, n) -> n <= Pos ? null : s + Pos
1341 // When the constant Size is less than or equal to the character
1342 // position also fold the result to null.
1343 Value *Cmp = B.CreateICmpULE(Size, ConstantInt::get(Size->getType(), Pos),
1344 "memchr.cmp");
1345 Value *SrcPlus = B.CreateInBoundsGEP(B.getInt8Ty(), SrcStr, B.getInt64(Pos),
1346 "memchr.ptr");
1347 return B.CreateSelect(Cmp, NullPtr, SrcPlus);
1348 }
1349
1350 if (Str.size() == 0)
1351 // If the array is empty fold memchr(A, C, N) to null for any value
1352 // of C and N on the basis that the only valid value of N is zero
1353 // (otherwise the call is undefined).
1354 return NullPtr;
1355
1356 if (LenC)
1357 Str = substr(Str, LenC->getZExtValue());
1358
1359 size_t Pos = Str.find_first_not_of(Str[0]);
1360 if (Pos == StringRef::npos
1361 || Str.find_first_not_of(Str[Pos], Pos) == StringRef::npos) {
1362 // If the source array consists of at most two consecutive sequences
1363 // of the same characters, then for any C and N (whether in bounds or
1364 // not), fold memchr(S, C, N) to
1365 // N != 0 && *S == C ? S : null
1366 // or for the two sequences to:
1367 // N != 0 && *S == C ? S : (N > Pos && S[Pos] == C ? S + Pos : null)
1368 // ^Sel2 ^Sel1 are denoted above.
1369 // The latter makes it also possible to fold strchr() calls with strings
1370 // of the same characters.
1371 Type *SizeTy = Size->getType();
1372 Type *Int8Ty = B.getInt8Ty();
1373
1374 // Slice off the sought character's high end bits.
1375 CharVal = B.CreateTrunc(CharVal, Int8Ty);
1376
1377 Value *Sel1 = NullPtr;
1378 if (Pos != StringRef::npos) {
1379 // Handle two consecutive sequences of the same characters.
1380 Value *PosVal = ConstantInt::get(SizeTy, Pos);
1381 Value *StrPos = ConstantInt::get(Int8Ty, Str[Pos]);
1382 Value *CEqSPos = B.CreateICmpEQ(CharVal, StrPos);
1383 Value *NGtPos = B.CreateICmp(ICmpInst::ICMP_UGT, Size, PosVal);
1384 Value *And = B.CreateAnd(CEqSPos, NGtPos);
1385 Value *SrcPlus = B.CreateInBoundsGEP(B.getInt8Ty(), SrcStr, PosVal);
1386 Sel1 = B.CreateSelect(And, SrcPlus, NullPtr, "memchr.sel1");
1387 }
1388
1389 Value *Str0 = ConstantInt::get(Int8Ty, Str[0]);
1390 Value *CEqS0 = B.CreateICmpEQ(Str0, CharVal);
1391 Value *NNeZ = B.CreateICmpNE(Size, ConstantInt::get(SizeTy, 0));
1392 Value *And = B.CreateAnd(NNeZ, CEqS0);
1393 return B.CreateSelect(And, SrcStr, Sel1, "memchr.sel2");
1394 }
1395
1396 if (!LenC) {
1397 if (isOnlyUsedInEqualityComparison(CI, SrcStr))
1398 // S is dereferenceable so it's safe to load from it and fold
1399 // memchr(S, C, N) == S to N && *S == C for any C and N.
1400 // TODO: This is safe even for nonconstant S.
1401 return memChrToCharCompare(CI, Size, B, DL);
1402
1403 // From now on we need a constant length and constant array.
1404 return nullptr;
1405 }
1406
1407 bool OptForSize = CI->getFunction()->hasOptSize() ||
1408 llvm::shouldOptimizeForSize(CI->getParent(), PSI, BFI,
1410
1411 // If the char is variable but the input str and length are not we can turn
1412 // this memchr call into a simple bit field test. Of course this only works
1413 // when the return value is only checked against null.
1414 //
1415 // It would be really nice to reuse switch lowering here but we can't change
1416 // the CFG at this point.
1417 //
1418 // memchr("\r\n", C, 2) != nullptr -> (1 << C & ((1 << '\r') | (1 << '\n')))
1419 // != 0
1420 // after bounds check.
1421 if (OptForSize || Str.empty() || !isOnlyUsedInZeroEqualityComparison(CI))
1422 return nullptr;
1423
1424 unsigned char Max =
1425 *std::max_element(reinterpret_cast<const unsigned char *>(Str.begin()),
1426 reinterpret_cast<const unsigned char *>(Str.end()));
1427
1428 // Make sure the bit field we're about to create fits in a register on the
1429 // target.
1430 // FIXME: On a 64 bit architecture this prevents us from using the
1431 // interesting range of alpha ascii chars. We could do better by emitting
1432 // two bitfields or shifting the range by 64 if no lower chars are used.
1433 if (!DL.fitsInLegalInteger(Max + 1)) {
1434 // Build chain of ORs
1435 // Transform:
1436 // memchr("abcd", C, 4) != nullptr
1437 // to:
1438 // (C == 'a' || C == 'b' || C == 'c' || C == 'd') != 0
1439 std::string SortedStr = Str.str();
1440 llvm::sort(SortedStr);
1441 // Compute the number of of non-contiguous ranges.
1442 unsigned NonContRanges = 1;
1443 for (size_t i = 1; i < SortedStr.size(); ++i) {
1444 if (SortedStr[i] > SortedStr[i - 1] + 1) {
1445 NonContRanges++;
1446 }
1447 }
1448
1449 // Restrict this optimization to profitable cases with one or two range
1450 // checks.
1451 if (NonContRanges > 2)
1452 return nullptr;
1453
1454 SmallVector<Value *> CharCompares;
1455 for (unsigned char C : SortedStr)
1456 CharCompares.push_back(
1457 B.CreateICmpEQ(CharVal, ConstantInt::get(CharVal->getType(), C)));
1458
1459 return B.CreateIntToPtr(B.CreateOr(CharCompares), CI->getType());
1460 }
1461
1462 // For the bit field use a power-of-2 type with at least 8 bits to avoid
1463 // creating unnecessary illegal types.
1464 unsigned char Width = NextPowerOf2(std::max((unsigned char)7, Max));
1465
1466 // Now build the bit field.
1467 APInt Bitfield(Width, 0);
1468 for (char C : Str)
1469 Bitfield.setBit((unsigned char)C);
1470 Value *BitfieldC = B.getInt(Bitfield);
1471
1472 // Adjust width of "C" to the bitfield width, then mask off the high bits.
1473 Value *C = B.CreateZExtOrTrunc(CharVal, BitfieldC->getType());
1474 C = B.CreateAnd(C, B.getIntN(Width, 0xFF));
1475
1476 // First check that the bit field access is within bounds.
1477 Value *Bounds = B.CreateICmp(ICmpInst::ICMP_ULT, C, B.getIntN(Width, Width),
1478 "memchr.bounds");
1479
1480 // Create code that checks if the given bit is set in the field.
1481 Value *Shl = B.CreateShl(B.getIntN(Width, 1ULL), C);
1482 Value *Bits = B.CreateIsNotNull(B.CreateAnd(Shl, BitfieldC), "memchr.bits");
1483
1484 // Finally merge both checks and cast to pointer type. The inttoptr
1485 // implicitly zexts the i1 to intptr type.
1486 return B.CreateIntToPtr(B.CreateLogicalAnd(Bounds, Bits, "memchr"),
1487 CI->getType());
1488}
1489
1490// Optimize a memcmp or, when StrNCmp is true, strncmp call CI with constant
1491// arrays LHS and RHS and nonconstant Size.
1493 Value *Size, bool StrNCmp,
1494 IRBuilderBase &B, const DataLayout &DL) {
1495 if (LHS == RHS) // memcmp(s,s,x) -> 0
1496 return Constant::getNullValue(CI->getType());
1497
1498 StringRef LStr, RStr;
1499 if (!getConstantStringInfo(LHS, LStr, /*TrimAtNul=*/false) ||
1500 !getConstantStringInfo(RHS, RStr, /*TrimAtNul=*/false))
1501 return nullptr;
1502
1503 // If the contents of both constant arrays are known, fold a call to
1504 // memcmp(A, B, N) to
1505 // N <= Pos ? 0 : (A < B ? -1 : B < A ? +1 : 0)
1506 // where Pos is the first mismatch between A and B, determined below.
1507
1508 uint64_t Pos = 0;
1509 Value *Zero = ConstantInt::get(CI->getType(), 0);
1510 for (uint64_t MinSize = std::min(LStr.size(), RStr.size()); ; ++Pos) {
1511 if (Pos == MinSize ||
1512 (StrNCmp && (LStr[Pos] == '\0' && RStr[Pos] == '\0'))) {
1513 // One array is a leading part of the other of equal or greater
1514 // size, or for strncmp, the arrays are equal strings.
1515 // Fold the result to zero. Size is assumed to be in bounds, since
1516 // otherwise the call would be undefined.
1517 return Zero;
1518 }
1519
1520 if (LStr[Pos] != RStr[Pos])
1521 break;
1522 }
1523
1524 // Normalize the result.
1525 typedef unsigned char UChar;
1526 int IRes = UChar(LStr[Pos]) < UChar(RStr[Pos]) ? -1 : 1;
1527 Value *MaxSize = ConstantInt::get(Size->getType(), Pos);
1528 Value *Cmp = B.CreateICmp(ICmpInst::ICMP_ULE, Size, MaxSize);
1529 Value *Res = ConstantInt::get(CI->getType(), IRes);
1530 return B.CreateSelect(Cmp, Zero, Res);
1531}
1532
1533// Optimize a memcmp call CI with constant size Len.
1535 uint64_t Len, IRBuilderBase &B,
1536 const DataLayout &DL) {
1537 if (Len == 0) // memcmp(s1,s2,0) -> 0
1538 return Constant::getNullValue(CI->getType());
1539
1540 // memcmp(S1,S2,1) -> *(unsigned char*)LHS - *(unsigned char*)RHS
1541 if (Len == 1) {
1542 Value *LHSV = B.CreateZExt(B.CreateLoad(B.getInt8Ty(), LHS, "lhsc"),
1543 CI->getType(), "lhsv");
1544 Value *RHSV = B.CreateZExt(B.CreateLoad(B.getInt8Ty(), RHS, "rhsc"),
1545 CI->getType(), "rhsv");
1546 return B.CreateSub(LHSV, RHSV, "chardiff");
1547 }
1548
1549 // memcmp(S1,S2,N/8)==0 -> (*(intN_t*)S1 != *(intN_t*)S2)==0
1550 // TODO: The case where both inputs are constants does not need to be limited
1551 // to legal integers or equality comparison. See block below this.
1552 if (DL.isLegalInteger(Len * 8) && isOnlyUsedInZeroEqualityComparison(CI)) {
1553 IntegerType *IntType = IntegerType::get(CI->getContext(), Len * 8);
1554 Align PrefAlignment = DL.getPrefTypeAlign(IntType);
1555
1556 // First, see if we can fold either argument to a constant.
1557 Value *LHSV = nullptr;
1558 if (auto *LHSC = dyn_cast<Constant>(LHS))
1559 LHSV = ConstantFoldLoadFromConstPtr(LHSC, IntType, DL);
1560
1561 Value *RHSV = nullptr;
1562 if (auto *RHSC = dyn_cast<Constant>(RHS))
1563 RHSV = ConstantFoldLoadFromConstPtr(RHSC, IntType, DL);
1564
1565 // Don't generate unaligned loads. If either source is constant data,
1566 // alignment doesn't matter for that source because there is no load.
1567 if ((LHSV || getKnownAlignment(LHS, DL, CI) >= PrefAlignment) &&
1568 (RHSV || getKnownAlignment(RHS, DL, CI) >= PrefAlignment)) {
1569 if (!LHSV)
1570 LHSV = B.CreateLoad(IntType, LHS, "lhsv");
1571 if (!RHSV)
1572 RHSV = B.CreateLoad(IntType, RHS, "rhsv");
1573 return B.CreateZExt(B.CreateICmpNE(LHSV, RHSV), CI->getType(), "memcmp");
1574 }
1575 }
1576
1577 return nullptr;
1578}
1579
1580// Most simplifications for memcmp also apply to bcmp.
1581Value *LibCallSimplifier::optimizeMemCmpBCmpCommon(CallInst *CI,
1582 IRBuilderBase &B) {
1583 Value *LHS = CI->getArgOperand(0), *RHS = CI->getArgOperand(1);
1584 Value *Size = CI->getArgOperand(2);
1585
1586 annotateNonNullAndDereferenceable(CI, {0, 1}, Size, DL);
1587
1588 if (Value *Res = optimizeMemCmpVarSize(CI, LHS, RHS, Size, false, B, DL))
1589 return Res;
1590
1591 // Handle constant Size.
1592 ConstantInt *LenC = dyn_cast<ConstantInt>(Size);
1593 if (!LenC)
1594 return nullptr;
1595
1596 return optimizeMemCmpConstantSize(CI, LHS, RHS, LenC->getZExtValue(), B, DL);
1597}
1598
1599Value *LibCallSimplifier::optimizeMemCmp(CallInst *CI, IRBuilderBase &B) {
1600 Module *M = CI->getModule();
1601 if (Value *V = optimizeMemCmpBCmpCommon(CI, B))
1602 return V;
1603
1604 // memcmp(x, y, Len) == 0 -> bcmp(x, y, Len) == 0
1605 // bcmp can be more efficient than memcmp because it only has to know that
1606 // there is a difference, not how different one is to the other.
1607 if (isLibFuncEmittable(M, TLI, LibFunc_bcmp) &&
1609 Value *LHS = CI->getArgOperand(0);
1610 Value *RHS = CI->getArgOperand(1);
1611 Value *Size = CI->getArgOperand(2);
1612 return copyFlags(*CI, emitBCmp(LHS, RHS, Size, B, DL, TLI));
1613 }
1614
1615 return nullptr;
1616}
1617
1618Value *LibCallSimplifier::optimizeBCmp(CallInst *CI, IRBuilderBase &B) {
1619 return optimizeMemCmpBCmpCommon(CI, B);
1620}
1621
1622Value *LibCallSimplifier::optimizeMemCpy(CallInst *CI, IRBuilderBase &B) {
1623 Value *Size = CI->getArgOperand(2);
1624 annotateNonNullAndDereferenceable(CI, {0, 1}, Size, DL);
1625 if (isa<IntrinsicInst>(CI))
1626 return nullptr;
1627
1628 // memcpy(x, y, n) -> llvm.memcpy(align 1 x, align 1 y, n)
1629 CallInst *NewCI = B.CreateMemCpy(CI->getArgOperand(0), Align(1),
1630 CI->getArgOperand(1), Align(1), Size);
1631 mergeAttributesAndFlags(NewCI, *CI);
1632 return CI->getArgOperand(0);
1633}
1634
1635Value *LibCallSimplifier::optimizeMemCCpy(CallInst *CI, IRBuilderBase &B) {
1636 Value *Dst = CI->getArgOperand(0);
1637 Value *Src = CI->getArgOperand(1);
1638 ConstantInt *StopChar = dyn_cast<ConstantInt>(CI->getArgOperand(2));
1639 ConstantInt *N = dyn_cast<ConstantInt>(CI->getArgOperand(3));
1640 StringRef SrcStr;
1641 if (CI->use_empty() && Dst == Src)
1642 return Dst;
1643 // memccpy(d, s, c, 0) -> nullptr
1644 if (N) {
1645 if (N->isNullValue())
1646 return Constant::getNullValue(CI->getType());
1647 if (!getConstantStringInfo(Src, SrcStr, /*TrimAtNul=*/false) ||
1648 // TODO: Handle zeroinitializer.
1649 !StopChar)
1650 return nullptr;
1651 } else {
1652 return nullptr;
1653 }
1654
1655 // Wrap arg 'c' of type int to char
1656 size_t Pos = SrcStr.find(StopChar->getSExtValue() & 0xFF);
1657 if (Pos == StringRef::npos) {
1658 if (N->getZExtValue() <= SrcStr.size()) {
1659 copyFlags(*CI, B.CreateMemCpy(Dst, Align(1), Src, Align(1),
1660 CI->getArgOperand(3)));
1661 return Constant::getNullValue(CI->getType());
1662 }
1663 return nullptr;
1664 }
1665
1666 Value *NewN =
1667 ConstantInt::get(N->getType(), std::min(uint64_t(Pos + 1), N->getZExtValue()));
1668 // memccpy -> llvm.memcpy
1669 copyFlags(*CI, B.CreateMemCpy(Dst, Align(1), Src, Align(1), NewN));
1670 return Pos + 1 <= N->getZExtValue()
1671 ? B.CreateInBoundsGEP(B.getInt8Ty(), Dst, NewN)
1673}
1674
1675Value *LibCallSimplifier::optimizeMemPCpy(CallInst *CI, IRBuilderBase &B) {
1676 Value *Dst = CI->getArgOperand(0);
1677 Value *N = CI->getArgOperand(2);
1678 // mempcpy(x, y, n) -> llvm.memcpy(align 1 x, align 1 y, n), x + n
1679 CallInst *NewCI =
1680 B.CreateMemCpy(Dst, Align(1), CI->getArgOperand(1), Align(1), N);
1681 // Propagate attributes, but memcpy has no return value, so make sure that
1682 // any return attributes are compliant.
1683 // TODO: Attach return value attributes to the 1st operand to preserve them?
1684 mergeAttributesAndFlags(NewCI, *CI);
1685 return B.CreateInBoundsGEP(B.getInt8Ty(), Dst, N);
1686}
1687
1688Value *LibCallSimplifier::optimizeMemMove(CallInst *CI, IRBuilderBase &B) {
1689 Value *Size = CI->getArgOperand(2);
1690 annotateNonNullAndDereferenceable(CI, {0, 1}, Size, DL);
1691 if (isa<IntrinsicInst>(CI))
1692 return nullptr;
1693
1694 // memmove(x, y, n) -> llvm.memmove(align 1 x, align 1 y, n)
1695 CallInst *NewCI = B.CreateMemMove(CI->getArgOperand(0), Align(1),
1696 CI->getArgOperand(1), Align(1), Size);
1697 mergeAttributesAndFlags(NewCI, *CI);
1698 return CI->getArgOperand(0);
1699}
1700
1701Value *LibCallSimplifier::optimizeMemSet(CallInst *CI, IRBuilderBase &B) {
1702 Value *Size = CI->getArgOperand(2);
1704 if (isa<IntrinsicInst>(CI))
1705 return nullptr;
1706
1707 // memset(p, v, n) -> llvm.memset(align 1 p, v, n)
1708 Value *Val = B.CreateIntCast(CI->getArgOperand(1), B.getInt8Ty(), false);
1709 CallInst *NewCI = B.CreateMemSet(CI->getArgOperand(0), Val, Size, Align(1));
1710 mergeAttributesAndFlags(NewCI, *CI);
1711 return CI->getArgOperand(0);
1712}
1713
1714Value *LibCallSimplifier::optimizeRealloc(CallInst *CI, IRBuilderBase &B) {
1715 if (isa<ConstantPointerNull>(CI->getArgOperand(0)))
1716 return copyFlags(*CI, emitMalloc(CI->getArgOperand(1), B, DL, TLI));
1717
1718 return nullptr;
1719}
1720
1721// When enabled, replace operator new() calls marked with a hot or cold memprof
1722// attribute with an operator new() call that takes a __hot_cold_t parameter.
1723// Currently this is supported by the open source version of tcmalloc, see:
1724// https://github.com/google/tcmalloc/blob/master/tcmalloc/new_extension.h
1725Value *LibCallSimplifier::optimizeNew(CallInst *CI, IRBuilderBase &B,
1726 LibFunc &Func) {
1727 if (!OptimizeHotColdNew)
1728 return nullptr;
1729
1730 uint8_t HotCold;
1731 if (CI->getAttributes().getFnAttr("memprof").getValueAsString() == "cold")
1732 HotCold = ColdNewHintValue;
1733 else if (CI->getAttributes().getFnAttr("memprof").getValueAsString() ==
1734 "notcold")
1735 HotCold = NotColdNewHintValue;
1736 else if (CI->getAttributes().getFnAttr("memprof").getValueAsString() == "hot")
1737 HotCold = HotNewHintValue;
1738 else
1739 return nullptr;
1740
1741 // For calls that already pass a hot/cold hint, only update the hint if
1742 // directed by OptimizeExistingHotColdNew. For other calls to new, add a hint
1743 // if cold or hot, and leave as-is for default handling if "notcold" aka warm.
1744 // Note that in cases where we decide it is "notcold", it might be slightly
1745 // better to replace the hinted call with a non hinted call, to avoid the
1746 // extra paramter and the if condition check of the hint value in the
1747 // allocator. This can be considered in the future.
1748 switch (Func) {
1749 case LibFunc_Znwm12__hot_cold_t:
1751 return emitHotColdNew(CI->getArgOperand(0), B, TLI,
1752 LibFunc_Znwm12__hot_cold_t, HotCold);
1753 break;
1754 case LibFunc_Znwm:
1755 if (HotCold != NotColdNewHintValue)
1756 return emitHotColdNew(CI->getArgOperand(0), B, TLI,
1757 LibFunc_Znwm12__hot_cold_t, HotCold);
1758 break;
1759 case LibFunc_Znam12__hot_cold_t:
1761 return emitHotColdNew(CI->getArgOperand(0), B, TLI,
1762 LibFunc_Znam12__hot_cold_t, HotCold);
1763 break;
1764 case LibFunc_Znam:
1765 if (HotCold != NotColdNewHintValue)
1766 return emitHotColdNew(CI->getArgOperand(0), B, TLI,
1767 LibFunc_Znam12__hot_cold_t, HotCold);
1768 break;
1769 case LibFunc_ZnwmRKSt9nothrow_t12__hot_cold_t:
1771 return emitHotColdNewNoThrow(
1772 CI->getArgOperand(0), CI->getArgOperand(1), B, TLI,
1773 LibFunc_ZnwmRKSt9nothrow_t12__hot_cold_t, HotCold);
1774 break;
1775 case LibFunc_ZnwmRKSt9nothrow_t:
1776 if (HotCold != NotColdNewHintValue)
1777 return emitHotColdNewNoThrow(
1778 CI->getArgOperand(0), CI->getArgOperand(1), B, TLI,
1779 LibFunc_ZnwmRKSt9nothrow_t12__hot_cold_t, HotCold);
1780 break;
1781 case LibFunc_ZnamRKSt9nothrow_t12__hot_cold_t:
1783 return emitHotColdNewNoThrow(
1784 CI->getArgOperand(0), CI->getArgOperand(1), B, TLI,
1785 LibFunc_ZnamRKSt9nothrow_t12__hot_cold_t, HotCold);
1786 break;
1787 case LibFunc_ZnamRKSt9nothrow_t:
1788 if (HotCold != NotColdNewHintValue)
1789 return emitHotColdNewNoThrow(
1790 CI->getArgOperand(0), CI->getArgOperand(1), B, TLI,
1791 LibFunc_ZnamRKSt9nothrow_t12__hot_cold_t, HotCold);
1792 break;
1793 case LibFunc_ZnwmSt11align_val_t12__hot_cold_t:
1795 return emitHotColdNewAligned(
1796 CI->getArgOperand(0), CI->getArgOperand(1), B, TLI,
1797 LibFunc_ZnwmSt11align_val_t12__hot_cold_t, HotCold);
1798 break;
1799 case LibFunc_ZnwmSt11align_val_t:
1800 if (HotCold != NotColdNewHintValue)
1801 return emitHotColdNewAligned(
1802 CI->getArgOperand(0), CI->getArgOperand(1), B, TLI,
1803 LibFunc_ZnwmSt11align_val_t12__hot_cold_t, HotCold);
1804 break;
1805 case LibFunc_ZnamSt11align_val_t12__hot_cold_t:
1807 return emitHotColdNewAligned(
1808 CI->getArgOperand(0), CI->getArgOperand(1), B, TLI,
1809 LibFunc_ZnamSt11align_val_t12__hot_cold_t, HotCold);
1810 break;
1811 case LibFunc_ZnamSt11align_val_t:
1812 if (HotCold != NotColdNewHintValue)
1813 return emitHotColdNewAligned(
1814 CI->getArgOperand(0), CI->getArgOperand(1), B, TLI,
1815 LibFunc_ZnamSt11align_val_t12__hot_cold_t, HotCold);
1816 break;
1817 case LibFunc_ZnwmSt11align_val_tRKSt9nothrow_t12__hot_cold_t:
1820 CI->getArgOperand(0), CI->getArgOperand(1), CI->getArgOperand(2), B,
1821 TLI, LibFunc_ZnwmSt11align_val_tRKSt9nothrow_t12__hot_cold_t,
1822 HotCold);
1823 break;
1824 case LibFunc_ZnwmSt11align_val_tRKSt9nothrow_t:
1825 if (HotCold != NotColdNewHintValue)
1827 CI->getArgOperand(0), CI->getArgOperand(1), CI->getArgOperand(2), B,
1828 TLI, LibFunc_ZnwmSt11align_val_tRKSt9nothrow_t12__hot_cold_t,
1829 HotCold);
1830 break;
1831 case LibFunc_ZnamSt11align_val_tRKSt9nothrow_t12__hot_cold_t:
1834 CI->getArgOperand(0), CI->getArgOperand(1), CI->getArgOperand(2), B,
1835 TLI, LibFunc_ZnamSt11align_val_tRKSt9nothrow_t12__hot_cold_t,
1836 HotCold);
1837 break;
1838 case LibFunc_ZnamSt11align_val_tRKSt9nothrow_t:
1839 if (HotCold != NotColdNewHintValue)
1841 CI->getArgOperand(0), CI->getArgOperand(1), CI->getArgOperand(2), B,
1842 TLI, LibFunc_ZnamSt11align_val_tRKSt9nothrow_t12__hot_cold_t,
1843 HotCold);
1844 break;
1845 default:
1846 return nullptr;
1847 }
1848 return nullptr;
1849}
1850
1851//===----------------------------------------------------------------------===//
1852// Math Library Optimizations
1853//===----------------------------------------------------------------------===//
1854
1855// Replace a libcall \p CI with a call to intrinsic \p IID
1857 Intrinsic::ID IID) {
1858 CallInst *NewCall = B.CreateUnaryIntrinsic(IID, CI->getArgOperand(0), CI);
1859 NewCall->takeName(CI);
1860 return copyFlags(*CI, NewCall);
1861}
1862
1863/// Return a variant of Val with float type.
1864/// Currently this works in two cases: If Val is an FPExtension of a float
1865/// value to something bigger, simply return the operand.
1866/// If Val is a ConstantFP but can be converted to a float ConstantFP without
1867/// loss of precision do so.
1869 if (FPExtInst *Cast = dyn_cast<FPExtInst>(Val)) {
1870 Value *Op = Cast->getOperand(0);
1871 if (Op->getType()->isFloatTy())
1872 return Op;
1873 }
1874 if (ConstantFP *Const = dyn_cast<ConstantFP>(Val)) {
1875 APFloat F = Const->getValueAPF();
1876 bool losesInfo;
1878 &losesInfo);
1879 if (!losesInfo)
1880 return ConstantFP::get(Const->getContext(), F);
1881 }
1882 return nullptr;
1883}
1884
1885/// Shrink double -> float functions.
1887 bool isBinary, const TargetLibraryInfo *TLI,
1888 bool isPrecise = false) {
1889 Function *CalleeFn = CI->getCalledFunction();
1890 if (!CI->getType()->isDoubleTy() || !CalleeFn)
1891 return nullptr;
1892
1893 // If not all the uses of the function are converted to float, then bail out.
1894 // This matters if the precision of the result is more important than the
1895 // precision of the arguments.
1896 if (isPrecise)
1897 for (User *U : CI->users()) {
1898 FPTruncInst *Cast = dyn_cast<FPTruncInst>(U);
1899 if (!Cast || !Cast->getType()->isFloatTy())
1900 return nullptr;
1901 }
1902
1903 // If this is something like 'g((double) float)', convert to 'gf(float)'.
1904 Value *V[2];
1906 V[1] = isBinary ? valueHasFloatPrecision(CI->getArgOperand(1)) : nullptr;
1907 if (!V[0] || (isBinary && !V[1]))
1908 return nullptr;
1909
1910 // If call isn't an intrinsic, check that it isn't within a function with the
1911 // same name as the float version of this call, otherwise the result is an
1912 // infinite loop. For example, from MinGW-w64:
1913 //
1914 // float expf(float val) { return (float) exp((double) val); }
1915 StringRef CalleeName = CalleeFn->getName();
1916 bool IsIntrinsic = CalleeFn->isIntrinsic();
1917 if (!IsIntrinsic) {
1918 StringRef CallerName = CI->getFunction()->getName();
1919 if (!CallerName.empty() && CallerName.back() == 'f' &&
1920 CallerName.size() == (CalleeName.size() + 1) &&
1921 CallerName.starts_with(CalleeName))
1922 return nullptr;
1923 }
1924
1925 // Propagate the math semantics from the current function to the new function.
1927 B.setFastMathFlags(CI->getFastMathFlags());
1928
1929 // g((double) float) -> (double) gf(float)
1930 Value *R;
1931 if (IsIntrinsic) {
1932 Module *M = CI->getModule();
1933 Intrinsic::ID IID = CalleeFn->getIntrinsicID();
1934 Function *Fn = Intrinsic::getDeclaration(M, IID, B.getFloatTy());
1935 R = isBinary ? B.CreateCall(Fn, V) : B.CreateCall(Fn, V[0]);
1936 } else {
1937 AttributeList CalleeAttrs = CalleeFn->getAttributes();
1938 R = isBinary ? emitBinaryFloatFnCall(V[0], V[1], TLI, CalleeName, B,
1939 CalleeAttrs)
1940 : emitUnaryFloatFnCall(V[0], TLI, CalleeName, B, CalleeAttrs);
1941 }
1942 return B.CreateFPExt(R, B.getDoubleTy());
1943}
1944
1945/// Shrink double -> float for unary functions.
1947 const TargetLibraryInfo *TLI,
1948 bool isPrecise = false) {
1949 return optimizeDoubleFP(CI, B, false, TLI, isPrecise);
1950}
1951
1952/// Shrink double -> float for binary functions.
1954 const TargetLibraryInfo *TLI,
1955 bool isPrecise = false) {
1956 return optimizeDoubleFP(CI, B, true, TLI, isPrecise);
1957}
1958
1959// cabs(z) -> sqrt((creal(z)*creal(z)) + (cimag(z)*cimag(z)))
1960Value *LibCallSimplifier::optimizeCAbs(CallInst *CI, IRBuilderBase &B) {
1961 if (!CI->isFast())
1962 return nullptr;
1963
1964 // Propagate fast-math flags from the existing call to new instructions.
1966 B.setFastMathFlags(CI->getFastMathFlags());
1967
1968 Value *Real, *Imag;
1969 if (CI->arg_size() == 1) {
1970 Value *Op = CI->getArgOperand(0);
1971 assert(Op->getType()->isArrayTy() && "Unexpected signature for cabs!");
1972 Real = B.CreateExtractValue(Op, 0, "real");
1973 Imag = B.CreateExtractValue(Op, 1, "imag");
1974 } else {
1975 assert(CI->arg_size() == 2 && "Unexpected signature for cabs!");
1976 Real = CI->getArgOperand(0);
1977 Imag = CI->getArgOperand(1);
1978 }
1979
1980 Value *RealReal = B.CreateFMul(Real, Real);
1981 Value *ImagImag = B.CreateFMul(Imag, Imag);
1982
1983 return copyFlags(*CI, B.CreateUnaryIntrinsic(Intrinsic::sqrt,
1984 B.CreateFAdd(RealReal, ImagImag),
1985 nullptr, "cabs"));
1986}
1987
1988// Return a properly extended integer (DstWidth bits wide) if the operation is
1989// an itofp.
1990static Value *getIntToFPVal(Value *I2F, IRBuilderBase &B, unsigned DstWidth) {
1991 if (isa<SIToFPInst>(I2F) || isa<UIToFPInst>(I2F)) {
1992 Value *Op = cast<Instruction>(I2F)->getOperand(0);
1993 // Make sure that the exponent fits inside an "int" of size DstWidth,
1994 // thus avoiding any range issues that FP has not.
1995 unsigned BitWidth = Op->getType()->getScalarSizeInBits();
1996 if (BitWidth < DstWidth || (BitWidth == DstWidth && isa<SIToFPInst>(I2F))) {
1997 Type *IntTy = Op->getType()->getWithNewBitWidth(DstWidth);
1998 return isa<SIToFPInst>(I2F) ? B.CreateSExt(Op, IntTy)
1999 : B.CreateZExt(Op, IntTy);
2000 }
2001 }
2002
2003 return nullptr;
2004}
2005
2006/// Use exp{,2}(x * y) for pow(exp{,2}(x), y);
2007/// ldexp(1.0, x) for pow(2.0, itofp(x)); exp2(n * x) for pow(2.0 ** n, x);
2008/// exp10(x) for pow(10.0, x); exp2(log2(n) * x) for pow(n, x).
2009Value *LibCallSimplifier::replacePowWithExp(CallInst *Pow, IRBuilderBase &B) {
2010 Module *M = Pow->getModule();
2011 Value *Base = Pow->getArgOperand(0), *Expo = Pow->getArgOperand(1);
2012 Type *Ty = Pow->getType();
2013 bool Ignored;
2014
2015 // Evaluate special cases related to a nested function as the base.
2016
2017 // pow(exp(x), y) -> exp(x * y)
2018 // pow(exp2(x), y) -> exp2(x * y)
2019 // If exp{,2}() is used only once, it is better to fold two transcendental
2020 // math functions into one. If used again, exp{,2}() would still have to be
2021 // called with the original argument, then keep both original transcendental
2022 // functions. However, this transformation is only safe with fully relaxed
2023 // math semantics, since, besides rounding differences, it changes overflow
2024 // and underflow behavior quite dramatically. For example:
2025 // pow(exp(1000), 0.001) = pow(inf, 0.001) = inf
2026 // Whereas:
2027 // exp(1000 * 0.001) = exp(1)
2028 // TODO: Loosen the requirement for fully relaxed math semantics.
2029 // TODO: Handle exp10() when more targets have it available.
2030 CallInst *BaseFn = dyn_cast<CallInst>(Base);
2031 if (BaseFn && BaseFn->hasOneUse() && BaseFn->isFast() && Pow->isFast()) {
2032 LibFunc LibFn;
2033
2034 Function *CalleeFn = BaseFn->getCalledFunction();
2035 if (CalleeFn && TLI->getLibFunc(CalleeFn->getName(), LibFn) &&
2036 isLibFuncEmittable(M, TLI, LibFn)) {
2037 StringRef ExpName;
2039 Value *ExpFn;
2040 LibFunc LibFnFloat, LibFnDouble, LibFnLongDouble;
2041
2042 switch (LibFn) {
2043 default:
2044 return nullptr;
2045 case LibFunc_expf:
2046 case LibFunc_exp:
2047 case LibFunc_expl:
2048 ExpName = TLI->getName(LibFunc_exp);
2049 ID = Intrinsic::exp;
2050 LibFnFloat = LibFunc_expf;
2051 LibFnDouble = LibFunc_exp;
2052 LibFnLongDouble = LibFunc_expl;
2053 break;
2054 case LibFunc_exp2f:
2055 case LibFunc_exp2:
2056 case LibFunc_exp2l:
2057 ExpName = TLI->getName(LibFunc_exp2);
2058 ID = Intrinsic::exp2;
2059 LibFnFloat = LibFunc_exp2f;
2060 LibFnDouble = LibFunc_exp2;
2061 LibFnLongDouble = LibFunc_exp2l;
2062 break;
2063 }
2064
2065 // Create new exp{,2}() with the product as its argument.
2066 Value *FMul = B.CreateFMul(BaseFn->getArgOperand(0), Expo, "mul");
2067 ExpFn = BaseFn->doesNotAccessMemory()
2068 ? B.CreateUnaryIntrinsic(ID, FMul, nullptr, ExpName)
2069 : emitUnaryFloatFnCall(FMul, TLI, LibFnDouble, LibFnFloat,
2070 LibFnLongDouble, B,
2071 BaseFn->getAttributes());
2072
2073 // Since the new exp{,2}() is different from the original one, dead code
2074 // elimination cannot be trusted to remove it, since it may have side
2075 // effects (e.g., errno). When the only consumer for the original
2076 // exp{,2}() is pow(), then it has to be explicitly erased.
2077 substituteInParent(BaseFn, ExpFn);
2078 return ExpFn;
2079 }
2080 }
2081
2082 // Evaluate special cases related to a constant base.
2083
2084 const APFloat *BaseF;
2085 if (!match(Base, m_APFloat(BaseF)))
2086 return nullptr;
2087
2088 AttributeList NoAttrs; // Attributes are only meaningful on the original call
2089
2090 const bool UseIntrinsic = Pow->doesNotAccessMemory();
2091
2092 // pow(2.0, itofp(x)) -> ldexp(1.0, x)
2093 if ((UseIntrinsic || !Ty->isVectorTy()) && BaseF->isExactlyValue(2.0) &&
2094 (isa<SIToFPInst>(Expo) || isa<UIToFPInst>(Expo)) &&
2095 (UseIntrinsic ||
2096 hasFloatFn(M, TLI, Ty, LibFunc_ldexp, LibFunc_ldexpf, LibFunc_ldexpl))) {
2097
2098 // TODO: Shouldn't really need to depend on getIntToFPVal for intrinsic. Can
2099 // just directly use the original integer type.
2100 if (Value *ExpoI = getIntToFPVal(Expo, B, TLI->getIntSize())) {
2101 Constant *One = ConstantFP::get(Ty, 1.0);
2102
2103 if (UseIntrinsic) {
2104 return copyFlags(*Pow, B.CreateIntrinsic(Intrinsic::ldexp,
2105 {Ty, ExpoI->getType()},
2106 {One, ExpoI}, Pow, "exp2"));
2107 }
2108
2109 return copyFlags(*Pow, emitBinaryFloatFnCall(
2110 One, ExpoI, TLI, LibFunc_ldexp, LibFunc_ldexpf,
2111 LibFunc_ldexpl, B, NoAttrs));
2112 }
2113 }
2114
2115 // pow(2.0 ** n, x) -> exp2(n * x)
2116 if (hasFloatFn(M, TLI, Ty, LibFunc_exp2, LibFunc_exp2f, LibFunc_exp2l)) {
2117 APFloat BaseR = APFloat(1.0);
2118 BaseR.convert(BaseF->getSemantics(), APFloat::rmTowardZero, &Ignored);
2119 BaseR = BaseR / *BaseF;
2120 bool IsInteger = BaseF->isInteger(), IsReciprocal = BaseR.isInteger();
2121 const APFloat *NF = IsReciprocal ? &BaseR : BaseF;
2122 APSInt NI(64, false);
2123 if ((IsInteger || IsReciprocal) &&
2124 NF->convertToInteger(NI, APFloat::rmTowardZero, &Ignored) ==
2125 APFloat::opOK &&
2126 NI > 1 && NI.isPowerOf2()) {
2127 double N = NI.logBase2() * (IsReciprocal ? -1.0 : 1.0);
2128 Value *FMul = B.CreateFMul(Expo, ConstantFP::get(Ty, N), "mul");
2129 if (Pow->doesNotAccessMemory())
2130 return copyFlags(*Pow, B.CreateUnaryIntrinsic(Intrinsic::exp2, FMul,
2131 nullptr, "exp2"));
2132 else
2133 return copyFlags(*Pow, emitUnaryFloatFnCall(FMul, TLI, LibFunc_exp2,
2134 LibFunc_exp2f,
2135 LibFunc_exp2l, B, NoAttrs));
2136 }
2137 }
2138
2139 // pow(10.0, x) -> exp10(x)
2140 if (BaseF->isExactlyValue(10.0) &&
2141 hasFloatFn(M, TLI, Ty, LibFunc_exp10, LibFunc_exp10f, LibFunc_exp10l)) {
2142
2143 if (Pow->doesNotAccessMemory()) {
2144 CallInst *NewExp10 =
2145 B.CreateIntrinsic(Intrinsic::exp10, {Ty}, {Expo}, Pow, "exp10");
2146 return copyFlags(*Pow, NewExp10);
2147 }
2148
2149 return copyFlags(*Pow, emitUnaryFloatFnCall(Expo, TLI, LibFunc_exp10,
2150 LibFunc_exp10f, LibFunc_exp10l,
2151 B, NoAttrs));
2152 }
2153
2154 // pow(x, y) -> exp2(log2(x) * y)
2155 if (Pow->hasApproxFunc() && Pow->hasNoNaNs() && BaseF->isFiniteNonZero() &&
2156 !BaseF->isNegative()) {
2157 // pow(1, inf) is defined to be 1 but exp2(log2(1) * inf) evaluates to NaN.
2158 // Luckily optimizePow has already handled the x == 1 case.
2159 assert(!match(Base, m_FPOne()) &&
2160 "pow(1.0, y) should have been simplified earlier!");
2161
2162 Value *Log = nullptr;
2163 if (Ty->isFloatTy())
2164 Log = ConstantFP::get(Ty, std::log2(BaseF->convertToFloat()));
2165 else if (Ty->isDoubleTy())
2166 Log = ConstantFP::get(Ty, std::log2(BaseF->convertToDouble()));
2167
2168 if (Log) {
2169 Value *FMul = B.CreateFMul(Log, Expo, "mul");
2170 if (Pow->doesNotAccessMemory())
2171 return copyFlags(*Pow, B.CreateUnaryIntrinsic(Intrinsic::exp2, FMul,
2172 nullptr, "exp2"));
2173 else if (hasFloatFn(M, TLI, Ty, LibFunc_exp2, LibFunc_exp2f,
2174 LibFunc_exp2l))
2175 return copyFlags(*Pow, emitUnaryFloatFnCall(FMul, TLI, LibFunc_exp2,
2176 LibFunc_exp2f,
2177 LibFunc_exp2l, B, NoAttrs));
2178 }
2179 }
2180
2181 return nullptr;
2182}
2183
2184static Value *getSqrtCall(Value *V, AttributeList Attrs, bool NoErrno,
2185 Module *M, IRBuilderBase &B,
2186 const TargetLibraryInfo *TLI) {
2187 // If errno is never set, then use the intrinsic for sqrt().
2188 if (NoErrno)
2189 return B.CreateUnaryIntrinsic(Intrinsic::sqrt, V, nullptr, "sqrt");
2190
2191 // Otherwise, use the libcall for sqrt().
2192 if (hasFloatFn(M, TLI, V->getType(), LibFunc_sqrt, LibFunc_sqrtf,
2193 LibFunc_sqrtl))
2194 // TODO: We also should check that the target can in fact lower the sqrt()
2195 // libcall. We currently have no way to ask this question, so we ask if
2196 // the target has a sqrt() libcall, which is not exactly the same.
2197 return emitUnaryFloatFnCall(V, TLI, LibFunc_sqrt, LibFunc_sqrtf,
2198 LibFunc_sqrtl, B, Attrs);
2199
2200 return nullptr;
2201}
2202
2203/// Use square root in place of pow(x, +/-0.5).
2204Value *LibCallSimplifier::replacePowWithSqrt(CallInst *Pow, IRBuilderBase &B) {
2205 Value *Sqrt, *Base = Pow->getArgOperand(0), *Expo = Pow->getArgOperand(1);
2206 Module *Mod = Pow->getModule();
2207 Type *Ty = Pow->getType();
2208
2209 const APFloat *ExpoF;
2210 if (!match(Expo, m_APFloat(ExpoF)) ||
2211 (!ExpoF->isExactlyValue(0.5) && !ExpoF->isExactlyValue(-0.5)))
2212 return nullptr;
2213
2214 // Converting pow(X, -0.5) to 1/sqrt(X) may introduce an extra rounding step,
2215 // so that requires fast-math-flags (afn or reassoc).
2216 if (ExpoF->isNegative() && (!Pow->hasApproxFunc() && !Pow->hasAllowReassoc()))
2217 return nullptr;
2218
2219 // If we have a pow() library call (accesses memory) and we can't guarantee
2220 // that the base is not an infinity, give up:
2221 // pow(-Inf, 0.5) is optionally required to have a result of +Inf (not setting
2222 // errno), but sqrt(-Inf) is required by various standards to set errno.
2223 if (!Pow->doesNotAccessMemory() && !Pow->hasNoInfs() &&
2225 SimplifyQuery(DL, TLI, /*DT=*/nullptr, AC, Pow)))
2226 return nullptr;
2227
2229 TLI);
2230 if (!Sqrt)
2231 return nullptr;
2232
2233 // Handle signed zero base by expanding to fabs(sqrt(x)).
2234 if (!Pow->hasNoSignedZeros())
2235 Sqrt = B.CreateUnaryIntrinsic(Intrinsic::fabs, Sqrt, nullptr, "abs");
2236
2237 Sqrt = copyFlags(*Pow, Sqrt);
2238
2239 // Handle non finite base by expanding to
2240 // (x == -infinity ? +infinity : sqrt(x)).
2241 if (!Pow->hasNoInfs()) {
2242 Value *PosInf = ConstantFP::getInfinity(Ty),
2243 *NegInf = ConstantFP::getInfinity(Ty, true);
2244 Value *FCmp = B.CreateFCmpOEQ(Base, NegInf, "isinf");
2245 Sqrt = B.CreateSelect(FCmp, PosInf, Sqrt);
2246 }
2247
2248 // If the exponent is negative, then get the reciprocal.
2249 if (ExpoF->isNegative())
2250 Sqrt = B.CreateFDiv(ConstantFP::get(Ty, 1.0), Sqrt, "reciprocal");
2251
2252 return Sqrt;
2253}
2254
2256 IRBuilderBase &B) {
2257 Value *Args[] = {Base, Expo};
2258 Type *Types[] = {Base->getType(), Expo->getType()};
2259 return B.CreateIntrinsic(Intrinsic::powi, Types, Args);
2260}
2261
2262Value *LibCallSimplifier::optimizePow(CallInst *Pow, IRBuilderBase &B) {
2263 Value *Base = Pow->getArgOperand(0);
2264 Value *Expo = Pow->getArgOperand(1);
2266 StringRef Name = Callee->getName();
2267 Type *Ty = Pow->getType();
2268 Module *M = Pow->getModule();
2269 bool AllowApprox = Pow->hasApproxFunc();
2270 bool Ignored;
2271
2272 // Propagate the math semantics from the call to any created instructions.
2274 B.setFastMathFlags(Pow->getFastMathFlags());
2275 // Evaluate special cases related to the base.
2276
2277 // pow(1.0, x) -> 1.0
2278 if (match(Base, m_FPOne()))
2279 return Base;
2280
2281 if (Value *Exp = replacePowWithExp(Pow, B))
2282 return Exp;
2283
2284 // Evaluate special cases related to the exponent.
2285
2286 // pow(x, -1.0) -> 1.0 / x
2287 if (match(Expo, m_SpecificFP(-1.0)))
2288 return B.CreateFDiv(ConstantFP::get(Ty, 1.0), Base, "reciprocal");
2289
2290 // pow(x, +/-0.0) -> 1.0
2291 if (match(Expo, m_AnyZeroFP()))
2292 return ConstantFP::get(Ty, 1.0);
2293
2294 // pow(x, 1.0) -> x
2295 if (match(Expo, m_FPOne()))
2296 return Base;
2297
2298 // pow(x, 2.0) -> x * x
2299 if (match(Expo, m_SpecificFP(2.0)))
2300 return B.CreateFMul(Base, Base, "square");
2301
2302 if (Value *Sqrt = replacePowWithSqrt(Pow, B))
2303 return Sqrt;
2304
2305 // If we can approximate pow:
2306 // pow(x, n) -> powi(x, n) * sqrt(x) if n has exactly a 0.5 fraction
2307 // pow(x, n) -> powi(x, n) if n is a constant signed integer value
2308 const APFloat *ExpoF;
2309 if (AllowApprox && match(Expo, m_APFloat(ExpoF)) &&
2310 !ExpoF->isExactlyValue(0.5) && !ExpoF->isExactlyValue(-0.5)) {
2311 APFloat ExpoA(abs(*ExpoF));
2312 APFloat ExpoI(*ExpoF);
2313 Value *Sqrt = nullptr;
2314 if (!ExpoA.isInteger()) {
2315 APFloat Expo2 = ExpoA;
2316 // To check if ExpoA is an integer + 0.5, we add it to itself. If there
2317 // is no floating point exception and the result is an integer, then
2318 // ExpoA == integer + 0.5
2319 if (Expo2.add(ExpoA, APFloat::rmNearestTiesToEven) != APFloat::opOK)
2320 return nullptr;
2321
2322 if (!Expo2.isInteger())
2323 return nullptr;
2324
2325 if (ExpoI.roundToIntegral(APFloat::rmTowardNegative) !=
2327 return nullptr;
2328 if (!ExpoI.isInteger())
2329 return nullptr;
2330 ExpoF = &ExpoI;
2331
2333 B, TLI);
2334 if (!Sqrt)
2335 return nullptr;
2336 }
2337
2338 // 0.5 fraction is now optionally handled.
2339 // Do pow -> powi for remaining integer exponent
2340 APSInt IntExpo(TLI->getIntSize(), /*isUnsigned=*/false);
2341 if (ExpoF->isInteger() &&
2342 ExpoF->convertToInteger(IntExpo, APFloat::rmTowardZero, &Ignored) ==
2343 APFloat::opOK) {
2344 Value *PowI = copyFlags(
2345 *Pow,
2347 Base, ConstantInt::get(B.getIntNTy(TLI->getIntSize()), IntExpo),
2348 M, B));
2349
2350 if (PowI && Sqrt)
2351 return B.CreateFMul(PowI, Sqrt);
2352
2353 return PowI;
2354 }
2355 }
2356
2357 // powf(x, itofp(y)) -> powi(x, y)
2358 if (AllowApprox && (isa<SIToFPInst>(Expo) || isa<UIToFPInst>(Expo))) {
2359 if (Value *ExpoI = getIntToFPVal(Expo, B, TLI->getIntSize()))
2360 return copyFlags(*Pow, createPowWithIntegerExponent(Base, ExpoI, M, B));
2361 }
2362
2363 // Shrink pow() to powf() if the arguments are single precision,
2364 // unless the result is expected to be double precision.
2365 if (UnsafeFPShrink && Name == TLI->getName(LibFunc_pow) &&
2366 hasFloatVersion(M, Name)) {
2367 if (Value *Shrunk = optimizeBinaryDoubleFP(Pow, B, TLI, true))
2368 return Shrunk;
2369 }
2370
2371 return nullptr;
2372}
2373
2374Value *LibCallSimplifier::optimizeExp2(CallInst *CI, IRBuilderBase &B) {
2375 Module *M = CI->getModule();
2377 StringRef Name = Callee->getName();
2378 Value *Ret = nullptr;
2379 if (UnsafeFPShrink && Name == TLI->getName(LibFunc_exp2) &&
2380 hasFloatVersion(M, Name))
2381 Ret = optimizeUnaryDoubleFP(CI, B, TLI, true);
2382
2383 // If we have an llvm.exp2 intrinsic, emit the llvm.ldexp intrinsic. If we
2384 // have the libcall, emit the libcall.
2385 //
2386 // TODO: In principle we should be able to just always use the intrinsic for
2387 // any doesNotAccessMemory callsite.
2388
2389 const bool UseIntrinsic = Callee->isIntrinsic();
2390 // Bail out for vectors because the code below only expects scalars.
2391 Type *Ty = CI->getType();
2392 if (!UseIntrinsic && Ty->isVectorTy())
2393 return Ret;
2394
2395 // exp2(sitofp(x)) -> ldexp(1.0, sext(x)) if sizeof(x) <= IntSize
2396 // exp2(uitofp(x)) -> ldexp(1.0, zext(x)) if sizeof(x) < IntSize
2397 Value *Op = CI->getArgOperand(0);
2398 if ((isa<SIToFPInst>(Op) || isa<UIToFPInst>(Op)) &&
2399 (UseIntrinsic ||
2400 hasFloatFn(M, TLI, Ty, LibFunc_ldexp, LibFunc_ldexpf, LibFunc_ldexpl))) {
2401 if (Value *Exp = getIntToFPVal(Op, B, TLI->getIntSize())) {
2402 Constant *One = ConstantFP::get(Ty, 1.0);
2403
2404 if (UseIntrinsic) {
2405 return copyFlags(*CI, B.CreateIntrinsic(Intrinsic::ldexp,
2406 {Ty, Exp->getType()},
2407 {One, Exp}, CI));
2408 }
2409
2411 B.setFastMathFlags(CI->getFastMathFlags());
2412 return copyFlags(*CI, emitBinaryFloatFnCall(
2413 One, Exp, TLI, LibFunc_ldexp, LibFunc_ldexpf,
2414 LibFunc_ldexpl, B, AttributeList()));
2415 }
2416 }
2417
2418 return Ret;
2419}
2420
2421Value *LibCallSimplifier::optimizeFMinFMax(CallInst *CI, IRBuilderBase &B) {
2422 Module *M = CI->getModule();
2423
2424 // If we can shrink the call to a float function rather than a double
2425 // function, do that first.
2427 StringRef Name = Callee->getName();
2428 if ((Name == "fmin" || Name == "fmax") && hasFloatVersion(M, Name))
2429 if (Value *Ret = optimizeBinaryDoubleFP(CI, B, TLI))
2430 return Ret;
2431
2432 // The LLVM intrinsics minnum/maxnum correspond to fmin/fmax. Canonicalize to
2433 // the intrinsics for improved optimization (for example, vectorization).
2434 // No-signed-zeros is implied by the definitions of fmax/fmin themselves.
2435 // From the C standard draft WG14/N1256:
2436 // "Ideally, fmax would be sensitive to the sign of zero, for example
2437 // fmax(-0.0, +0.0) would return +0; however, implementation in software
2438 // might be impractical."
2440 FastMathFlags FMF = CI->getFastMathFlags();
2441 FMF.setNoSignedZeros();
2442 B.setFastMathFlags(FMF);
2443
2444 Intrinsic::ID IID = Callee->getName().starts_with("fmin") ? Intrinsic::minnum
2445 : Intrinsic::maxnum;
2446 return copyFlags(*CI, B.CreateBinaryIntrinsic(IID, CI->getArgOperand(0),
2447 CI->getArgOperand(1)));
2448}
2449
2450Value *LibCallSimplifier::optimizeLog(CallInst *Log, IRBuilderBase &B) {
2451 Function *LogFn = Log->getCalledFunction();
2452 StringRef LogNm = LogFn->getName();
2453 Intrinsic::ID LogID = LogFn->getIntrinsicID();
2454 Module *Mod = Log->getModule();
2455 Type *Ty = Log->getType();
2456 Value *Ret = nullptr;
2457
2458 if (UnsafeFPShrink && hasFloatVersion(Mod, LogNm))
2459 Ret = optimizeUnaryDoubleFP(Log, B, TLI, true);
2460
2461 // The earlier call must also be 'fast' in order to do these transforms.
2462 CallInst *Arg = dyn_cast<CallInst>(Log->getArgOperand(0));
2463 if (!Log->isFast() || !Arg || !Arg->isFast() || !Arg->hasOneUse())
2464 return Ret;
2465
2466 LibFunc LogLb, ExpLb, Exp2Lb, Exp10Lb, PowLb;
2467
2468 // This is only applicable to log(), log2(), log10().
2469 if (TLI->getLibFunc(LogNm, LogLb))
2470 switch (LogLb) {
2471 case LibFunc_logf:
2472 LogID = Intrinsic::log;
2473 ExpLb = LibFunc_expf;
2474 Exp2Lb = LibFunc_exp2f;
2475 Exp10Lb = LibFunc_exp10f;
2476 PowLb = LibFunc_powf;
2477 break;
2478 case LibFunc_log:
2479 LogID = Intrinsic::log;
2480 ExpLb = LibFunc_exp;
2481 Exp2Lb = LibFunc_exp2;
2482 Exp10Lb = LibFunc_exp10;
2483 PowLb = LibFunc_pow;
2484 break;
2485 case LibFunc_logl:
2486 LogID = Intrinsic::log;
2487 ExpLb = LibFunc_expl;
2488 Exp2Lb = LibFunc_exp2l;
2489 Exp10Lb = LibFunc_exp10l;
2490 PowLb = LibFunc_powl;
2491 break;
2492 case LibFunc_log2f:
2493 LogID = Intrinsic::log2;
2494 ExpLb = LibFunc_expf;
2495 Exp2Lb = LibFunc_exp2f;
2496 Exp10Lb = LibFunc_exp10f;
2497 PowLb = LibFunc_powf;
2498 break;
2499 case LibFunc_log2:
2500 LogID = Intrinsic::log2;
2501 ExpLb = LibFunc_exp;
2502 Exp2Lb = LibFunc_exp2;
2503 Exp10Lb = LibFunc_exp10;
2504 PowLb = LibFunc_pow;
2505 break;
2506 case LibFunc_log2l:
2507 LogID = Intrinsic::log2;
2508 ExpLb = LibFunc_expl;
2509 Exp2Lb = LibFunc_exp2l;
2510 Exp10Lb = LibFunc_exp10l;
2511 PowLb = LibFunc_powl;
2512 break;
2513 case LibFunc_log10f:
2514 LogID = Intrinsic::log10;
2515 ExpLb = LibFunc_expf;
2516 Exp2Lb = LibFunc_exp2f;
2517 Exp10Lb = LibFunc_exp10f;
2518 PowLb = LibFunc_powf;
2519 break;
2520 case LibFunc_log10:
2521 LogID = Intrinsic::log10;
2522 ExpLb = LibFunc_exp;
2523 Exp2Lb = LibFunc_exp2;
2524 Exp10Lb = LibFunc_exp10;
2525 PowLb = LibFunc_pow;
2526 break;
2527 case LibFunc_log10l:
2528 LogID = Intrinsic::log10;
2529 ExpLb = LibFunc_expl;
2530 Exp2Lb = LibFunc_exp2l;
2531 Exp10Lb = LibFunc_exp10l;
2532 PowLb = LibFunc_powl;
2533 break;
2534 default:
2535 return Ret;
2536 }
2537 else if (LogID == Intrinsic::log || LogID == Intrinsic::log2 ||
2538 LogID == Intrinsic::log10) {
2539 if (Ty->getScalarType()->isFloatTy()) {
2540 ExpLb = LibFunc_expf;
2541 Exp2Lb = LibFunc_exp2f;
2542 Exp10Lb = LibFunc_exp10f;
2543 PowLb = LibFunc_powf;
2544 } else if (Ty->getScalarType()->isDoubleTy()) {
2545 ExpLb = LibFunc_exp;
2546 Exp2Lb = LibFunc_exp2;
2547 Exp10Lb = LibFunc_exp10;
2548 PowLb = LibFunc_pow;
2549 } else
2550 return Ret;
2551 } else
2552 return Ret;
2553
2555 B.setFastMathFlags(FastMathFlags::getFast());
2556
2557 Intrinsic::ID ArgID = Arg->getIntrinsicID();
2558 LibFunc ArgLb = NotLibFunc;
2559 TLI->getLibFunc(*Arg, ArgLb);
2560
2561 // log(pow(x,y)) -> y*log(x)
2562 AttributeList NoAttrs;
2563 if (ArgLb == PowLb || ArgID == Intrinsic::pow || ArgID == Intrinsic::powi) {
2564 Value *LogX =
2565 Log->doesNotAccessMemory()
2566 ? B.CreateUnaryIntrinsic(LogID, Arg->getOperand(0), nullptr, "log")
2567 : emitUnaryFloatFnCall(Arg->getOperand(0), TLI, LogNm, B, NoAttrs);
2568 Value *Y = Arg->getArgOperand(1);
2569 // Cast exponent to FP if integer.
2570 if (ArgID == Intrinsic::powi)
2571 Y = B.CreateSIToFP(Y, Ty, "cast");
2572 Value *MulY = B.CreateFMul(Y, LogX, "mul");
2573 // Since pow() may have side effects, e.g. errno,
2574 // dead code elimination may not be trusted to remove it.
2575 substituteInParent(Arg, MulY);
2576 return MulY;
2577 }
2578
2579 // log(exp{,2,10}(y)) -> y*log({e,2,10})
2580 // TODO: There is no exp10() intrinsic yet.
2581 if (ArgLb == ExpLb || ArgLb == Exp2Lb || ArgLb == Exp10Lb ||
2582 ArgID == Intrinsic::exp || ArgID == Intrinsic::exp2) {
2583 Constant *Eul;
2584 if (ArgLb == ExpLb || ArgID == Intrinsic::exp)
2585 // FIXME: Add more precise value of e for long double.
2586 Eul = ConstantFP::get(Log->getType(), numbers::e);
2587 else if (ArgLb == Exp2Lb || ArgID == Intrinsic::exp2)
2588 Eul = ConstantFP::get(Log->getType(), 2.0);
2589 else
2590 Eul = ConstantFP::get(Log->getType(), 10.0);
2591 Value *LogE = Log->doesNotAccessMemory()
2592 ? B.CreateUnaryIntrinsic(LogID, Eul, nullptr, "log")
2593 : emitUnaryFloatFnCall(Eul, TLI, LogNm, B, NoAttrs);
2594 Value *MulY = B.CreateFMul(Arg->getArgOperand(0), LogE, "mul");
2595 // Since exp() may have side effects, e.g. errno,
2596 // dead code elimination may not be trusted to remove it.
2597 substituteInParent(Arg, MulY);
2598 return MulY;
2599 }
2600
2601 return Ret;
2602}
2603
2604// sqrt(exp(X)) -> exp(X * 0.5)
2605Value *LibCallSimplifier::mergeSqrtToExp(CallInst *CI, IRBuilderBase &B) {
2606 if (!CI->hasAllowReassoc())
2607 return nullptr;
2608
2609 Function *SqrtFn = CI->getCalledFunction();
2610 CallInst *Arg = dyn_cast<CallInst>(CI->getArgOperand(0));
2611 if (!Arg || !Arg->hasAllowReassoc() || !Arg->hasOneUse())
2612 return nullptr;
2613 Intrinsic::ID ArgID = Arg->getIntrinsicID();
2614 LibFunc ArgLb = NotLibFunc;
2615 TLI->getLibFunc(*Arg, ArgLb);
2616
2617 LibFunc SqrtLb, ExpLb, Exp2Lb, Exp10Lb;
2618
2619 if (TLI->getLibFunc(SqrtFn->getName(), SqrtLb))
2620 switch (SqrtLb) {
2621 case LibFunc_sqrtf:
2622 ExpLb = LibFunc_expf;
2623 Exp2Lb = LibFunc_exp2f;
2624 Exp10Lb = LibFunc_exp10f;
2625 break;
2626 case LibFunc_sqrt:
2627 ExpLb = LibFunc_exp;
2628 Exp2Lb = LibFunc_exp2;
2629 Exp10Lb = LibFunc_exp10;
2630 break;
2631 case LibFunc_sqrtl:
2632 ExpLb = LibFunc_expl;
2633 Exp2Lb = LibFunc_exp2l;
2634 Exp10Lb = LibFunc_exp10l;
2635 break;
2636 default:
2637 return nullptr;
2638 }
2639 else if (SqrtFn->getIntrinsicID() == Intrinsic::sqrt) {
2640 if (CI->getType()->getScalarType()->isFloatTy()) {
2641 ExpLb = LibFunc_expf;
2642 Exp2Lb = LibFunc_exp2f;
2643 Exp10Lb = LibFunc_exp10f;
2644 } else if (CI->getType()->getScalarType()->isDoubleTy()) {
2645 ExpLb = LibFunc_exp;
2646 Exp2Lb = LibFunc_exp2;
2647 Exp10Lb = LibFunc_exp10;
2648 } else
2649 return nullptr;
2650 } else
2651 return nullptr;
2652
2653 if (ArgLb != ExpLb && ArgLb != Exp2Lb && ArgLb != Exp10Lb &&
2654 ArgID != Intrinsic::exp && ArgID != Intrinsic::exp2)
2655 return nullptr;
2656
2658 B.SetInsertPoint(Arg);
2659 auto *ExpOperand = Arg->getOperand(0);
2660 auto *FMul =
2661 B.CreateFMulFMF(ExpOperand, ConstantFP::get(ExpOperand->getType(), 0.5),
2662 CI, "merged.sqrt");
2663
2664 Arg->setOperand(0, FMul);
2665 return Arg;
2666}
2667
2668Value *LibCallSimplifier::optimizeSqrt(CallInst *CI, IRBuilderBase &B) {
2669 Module *M = CI->getModule();
2671 Value *Ret = nullptr;
2672 // TODO: Once we have a way (other than checking for the existince of the
2673 // libcall) to tell whether our target can lower @llvm.sqrt, relax the
2674 // condition below.
2675 if (isLibFuncEmittable(M, TLI, LibFunc_sqrtf) &&
2676 (Callee->getName() == "sqrt" ||
2677 Callee->getIntrinsicID() == Intrinsic::sqrt))
2678 Ret = optimizeUnaryDoubleFP(CI, B, TLI, true);
2679
2680 if (Value *Opt = mergeSqrtToExp(CI, B))
2681 return Opt;
2682
2683 if (!CI->isFast())
2684 return Ret;
2685
2686 Instruction *I = dyn_cast<Instruction>(CI->getArgOperand(0));
2687 if (!I || I->getOpcode() != Instruction::FMul || !I->isFast())
2688 return Ret;
2689
2690 // We're looking for a repeated factor in a multiplication tree,
2691 // so we can do this fold: sqrt(x * x) -> fabs(x);
2692 // or this fold: sqrt((x * x) * y) -> fabs(x) * sqrt(y).
2693 Value *Op0 = I->getOperand(0);
2694 Value *Op1 = I->getOperand(1);
2695 Value *RepeatOp = nullptr;
2696 Value *OtherOp = nullptr;
2697 if (Op0 == Op1) {
2698 // Simple match: the operands of the multiply are identical.
2699 RepeatOp = Op0;
2700 } else {
2701 // Look for a more complicated pattern: one of the operands is itself
2702 // a multiply, so search for a common factor in that multiply.
2703 // Note: We don't bother looking any deeper than this first level or for
2704 // variations of this pattern because instcombine's visitFMUL and/or the
2705 // reassociation pass should give us this form.
2706 Value *OtherMul0, *OtherMul1;
2707 if (match(Op0, m_FMul(m_Value(OtherMul0), m_Value(OtherMul1)))) {
2708 // Pattern: sqrt((x * y) * z)
2709 if (OtherMul0 == OtherMul1 && cast<Instruction>(Op0)->isFast()) {
2710 // Matched: sqrt((x * x) * z)
2711 RepeatOp = OtherMul0;
2712 OtherOp = Op1;
2713 }
2714 }
2715 }
2716 if (!RepeatOp)
2717 return Ret;
2718
2719 // Fast math flags for any created instructions should match the sqrt
2720 // and multiply.
2722 B.setFastMathFlags(I->getFastMathFlags());
2723
2724 // If we found a repeated factor, hoist it out of the square root and
2725 // replace it with the fabs of that factor.
2726 Value *FabsCall =
2727 B.CreateUnaryIntrinsic(Intrinsic::fabs, RepeatOp, nullptr, "fabs");
2728 if (OtherOp) {
2729 // If we found a non-repeated factor, we still need to get its square
2730 // root. We then multiply that by the value that was simplified out
2731 // of the square root calculation.
2732 Value *SqrtCall =
2733 B.CreateUnaryIntrinsic(Intrinsic::sqrt, OtherOp, nullptr, "sqrt");
2734 return copyFlags(*CI, B.CreateFMul(FabsCall, SqrtCall));
2735 }
2736 return copyFlags(*CI, FabsCall);
2737}
2738
2739Value *LibCallSimplifier::optimizeTrigInversionPairs(CallInst *CI,
2740 IRBuilderBase &B) {
2741 Module *M = CI->getModule();
2743 Value *Ret = nullptr;
2744 StringRef Name = Callee->getName();
2745 if (UnsafeFPShrink &&
2746 (Name == "tan" || Name == "atanh" || Name == "sinh" || Name == "cosh" ||
2747 Name == "asinh") &&
2748 hasFloatVersion(M, Name))
2749 Ret = optimizeUnaryDoubleFP(CI, B, TLI, true);
2750
2751 Value *Op1 = CI->getArgOperand(0);
2752 auto *OpC = dyn_cast<CallInst>(Op1);
2753 if (!OpC)
2754 return Ret;
2755
2756 // Both calls must be 'fast' in order to remove them.
2757 if (!CI->isFast() || !OpC->isFast())
2758 return Ret;
2759
2760 // tan(atan(x)) -> x
2761 // atanh(tanh(x)) -> x
2762 // sinh(asinh(x)) -> x
2763 // asinh(sinh(x)) -> x
2764 // cosh(acosh(x)) -> x
2765 LibFunc Func;
2766 Function *F = OpC->getCalledFunction();
2767 if (F && TLI->getLibFunc(F->getName(), Func) &&
2768 isLibFuncEmittable(M, TLI, Func)) {
2769 LibFunc inverseFunc = llvm::StringSwitch<LibFunc>(Callee->getName())
2770 .Case("tan", LibFunc_atan)
2771 .Case("atanh", LibFunc_tanh)
2772 .Case("sinh", LibFunc_asinh)
2773 .Case("cosh", LibFunc_acosh)
2774 .Case("tanf", LibFunc_atanf)
2775 .Case("atanhf", LibFunc_tanhf)
2776 .Case("sinhf", LibFunc_asinhf)
2777 .Case("coshf", LibFunc_acoshf)
2778 .Case("tanl", LibFunc_atanl)
2779 .Case("atanhl", LibFunc_tanhl)
2780 .Case("sinhl", LibFunc_asinhl)
2781 .Case("coshl", LibFunc_acoshl)
2782 .Case("asinh", LibFunc_sinh)
2783 .Case("asinhf", LibFunc_sinhf)
2784 .Case("asinhl", LibFunc_sinhl)
2785 .Default(NumLibFuncs); // Used as error value
2786 if (Func == inverseFunc)
2787 Ret = OpC->getArgOperand(0);
2788 }
2789 return Ret;
2790}
2791
2792static bool isTrigLibCall(CallInst *CI) {
2793 // We can only hope to do anything useful if we can ignore things like errno
2794 // and floating-point exceptions.
2795 // We already checked the prototype.
2796 return CI->doesNotThrow() && CI->doesNotAccessMemory();
2797}
2798
2799static bool insertSinCosCall(IRBuilderBase &B, Function *OrigCallee, Value *Arg,
2800 bool UseFloat, Value *&Sin, Value *&Cos,
2801 Value *&SinCos, const TargetLibraryInfo *TLI) {
2802 Module *M = OrigCallee->getParent();
2803 Type *ArgTy = Arg->getType();
2804 Type *ResTy;
2806
2807 Triple T(OrigCallee->getParent()->getTargetTriple());
2808 if (UseFloat) {
2809 Name = "__sincospif_stret";
2810
2811 assert(T.getArch() != Triple::x86 && "x86 messy and unsupported for now");
2812 // x86_64 can't use {float, float} since that would be returned in both
2813 // xmm0 and xmm1, which isn't what a real struct would do.
2814 ResTy = T.getArch() == Triple::x86_64
2815 ? static_cast<Type *>(FixedVectorType::get(ArgTy, 2))
2816 : static_cast<Type *>(StructType::get(ArgTy, ArgTy));
2817 } else {
2818 Name = "__sincospi_stret";
2819 ResTy = StructType::get(ArgTy, ArgTy);
2820 }
2821
2822 if (!isLibFuncEmittable(M, TLI, Name))
2823 return false;
2824 LibFunc TheLibFunc;
2825 TLI->getLibFunc(Name, TheLibFunc);
2827 M, *TLI, TheLibFunc, OrigCallee->getAttributes(), ResTy, ArgTy);
2828
2829 if (Instruction *ArgInst = dyn_cast<Instruction>(Arg)) {
2830 // If the argument is an instruction, it must dominate all uses so put our
2831 // sincos call there.
2832 B.SetInsertPoint(ArgInst->getParent(), ++ArgInst->getIterator());
2833 } else {
2834 // Otherwise (e.g. for a constant) the beginning of the function is as
2835 // good a place as any.
2836 BasicBlock &EntryBB = B.GetInsertBlock()->getParent()->getEntryBlock();
2837 B.SetInsertPoint(&EntryBB, EntryBB.begin());
2838 }
2839
2840 SinCos = B.CreateCall(Callee, Arg, "sincospi");
2841
2842 if (SinCos->getType()->isStructTy()) {
2843 Sin = B.CreateExtractValue(SinCos, 0, "sinpi");
2844 Cos = B.CreateExtractValue(SinCos, 1, "cospi");
2845 } else {
2846 Sin = B.CreateExtractElement(SinCos, ConstantInt::get(B.getInt32Ty(), 0),
2847 "sinpi");
2848 Cos = B.CreateExtractElement(SinCos, ConstantInt::get(B.getInt32Ty(), 1),
2849 "cospi");
2850 }
2851
2852 return true;
2853}
2854
2855static Value *optimizeSymmetricCall(CallInst *CI, bool IsEven,
2856 IRBuilderBase &B) {
2857 Value *X;
2858 Value *Src = CI->getArgOperand(0);
2859
2860 if (match(Src, m_OneUse(m_FNeg(m_Value(X))))) {
2862 B.setFastMathFlags(CI->getFastMathFlags());
2863
2864 auto *CallInst = copyFlags(*CI, B.CreateCall(CI->getCalledFunction(), {X}));
2865 if (IsEven) {
2866 // Even function: f(-x) = f(x)
2867 return CallInst;
2868 }
2869 // Odd function: f(-x) = -f(x)
2870 return B.CreateFNeg(CallInst);
2871 }
2872
2873 // Even function: f(abs(x)) = f(x), f(copysign(x, y)) = f(x)
2874 if (IsEven && (match(Src, m_FAbs(m_Value(X))) ||
2875 match(Src, m_CopySign(m_Value(X), m_Value())))) {
2877 B.setFastMathFlags(CI->getFastMathFlags());
2878
2879 auto *CallInst = copyFlags(*CI, B.CreateCall(CI->getCalledFunction(), {X}));
2880 return CallInst;
2881 }
2882
2883 return nullptr;
2884}
2885
2886Value *LibCallSimplifier::optimizeSymmetric(CallInst *CI, LibFunc Func,
2887 IRBuilderBase &B) {
2888 switch (Func) {
2889 case LibFunc_cos:
2890 case LibFunc_cosf:
2891 case LibFunc_cosl:
2892 return optimizeSymmetricCall(CI, /*IsEven*/ true, B);
2893
2894 case LibFunc_sin:
2895 case LibFunc_sinf:
2896 case LibFunc_sinl:
2897
2898 case LibFunc_tan:
2899 case LibFunc_tanf:
2900 case LibFunc_tanl:
2901
2902 case LibFunc_erf:
2903 case LibFunc_erff:
2904 case LibFunc_erfl:
2905 return optimizeSymmetricCall(CI, /*IsEven*/ false, B);
2906
2907 default:
2908 return nullptr;
2909 }
2910}
2911
2912Value *LibCallSimplifier::optimizeSinCosPi(CallInst *CI, bool IsSin, IRBuilderBase &B) {
2913 // Make sure the prototype is as expected, otherwise the rest of the
2914 // function is probably invalid and likely to abort.
2915 if (!isTrigLibCall(CI))
2916 return nullptr;
2917
2918 Value *Arg = CI->getArgOperand(0);
2921 SmallVector<CallInst *, 1> SinCosCalls;
2922
2923 bool IsFloat = Arg->getType()->isFloatTy();
2924
2925 // Look for all compatible sinpi, cospi and sincospi calls with the same
2926 // argument. If there are enough (in some sense) we can make the
2927 // substitution.
2928 Function *F = CI->getFunction();
2929 for (User *U : Arg->users())
2930 classifyArgUse(U, F, IsFloat, SinCalls, CosCalls, SinCosCalls);
2931
2932 // It's only worthwhile if both sinpi and cospi are actually used.
2933 if (SinCalls.empty() || CosCalls.empty())
2934 return nullptr;
2935
2936 Value *Sin, *Cos, *SinCos;
2937 if (!insertSinCosCall(B, CI->getCalledFunction(), Arg, IsFloat, Sin, Cos,
2938 SinCos, TLI))
2939 return nullptr;
2940
2941 auto replaceTrigInsts = [this](SmallVectorImpl<CallInst *> &Calls,
2942 Value *Res) {
2943 for (CallInst *C : Calls)
2944 replaceAllUsesWith(C, Res);
2945 };
2946
2947 replaceTrigInsts(SinCalls, Sin);
2948 replaceTrigInsts(CosCalls, Cos);
2949 replaceTrigInsts(SinCosCalls, SinCos);
2950
2951 return IsSin ? Sin : Cos;
2952}
2953
2954void LibCallSimplifier::classifyArgUse(
2955 Value *Val, Function *F, bool IsFloat,
2958 SmallVectorImpl<CallInst *> &SinCosCalls) {
2959 auto *CI = dyn_cast<CallInst>(Val);
2960 if (!CI || CI->use_empty())
2961 return;
2962
2963 // Don't consider calls in other functions.
2964 if (CI->getFunction() != F)
2965 return;
2966
2967 Module *M = CI->getModule();
2969 LibFunc Func;
2970 if (!Callee || !TLI->getLibFunc(*Callee, Func) ||
2971 !isLibFuncEmittable(M, TLI, Func) ||
2972 !isTrigLibCall(CI))
2973 return;
2974
2975 if (IsFloat) {
2976 if (Func == LibFunc_sinpif)
2977 SinCalls.push_back(CI);
2978 else if (Func == LibFunc_cospif)
2979 CosCalls.push_back(CI);
2980 else if (Func == LibFunc_sincospif_stret)
2981 SinCosCalls.push_back(CI);
2982 } else {
2983 if (Func == LibFunc_sinpi)
2984 SinCalls.push_back(CI);
2985 else if (Func == LibFunc_cospi)
2986 CosCalls.push_back(CI);
2987 else if (Func == LibFunc_sincospi_stret)
2988 SinCosCalls.push_back(CI);
2989 }
2990}
2991
2992//===----------------------------------------------------------------------===//
2993// Integer Library Call Optimizations
2994//===----------------------------------------------------------------------===//
2995
2996Value *LibCallSimplifier::optimizeFFS(CallInst *CI, IRBuilderBase &B) {
2997 // All variants of ffs return int which need not be 32 bits wide.
2998 // ffs{,l,ll}(x) -> x != 0 ? (int)llvm.cttz(x)+1 : 0
2999 Type *RetType = CI->getType();
3000 Value *Op = CI->getArgOperand(0);
3001 Type *ArgType = Op->getType();
3002 Value *V = B.CreateIntrinsic(Intrinsic::cttz, {ArgType}, {Op, B.getTrue()},
3003 nullptr, "cttz");
3004 V = B.CreateAdd(V, ConstantInt::get(V->getType(), 1));
3005 V = B.CreateIntCast(V, RetType, false);
3006
3007 Value *Cond = B.CreateICmpNE(Op, Constant::getNullValue(ArgType));
3008 return B.CreateSelect(Cond, V, ConstantInt::get(RetType, 0));
3009}
3010
3011Value *LibCallSimplifier::optimizeFls(CallInst *CI, IRBuilderBase &B) {
3012 // All variants of fls return int which need not be 32 bits wide.
3013 // fls{,l,ll}(x) -> (int)(sizeInBits(x) - llvm.ctlz(x, false))
3014 Value *Op = CI->getArgOperand(0);
3015 Type *ArgType = Op->getType();
3016 Value *V = B.CreateIntrinsic(Intrinsic::ctlz, {ArgType}, {Op, B.getFalse()},
3017 nullptr, "ctlz");
3018 V = B.CreateSub(ConstantInt::get(V->getType(), ArgType->getIntegerBitWidth()),
3019 V);
3020 return B.CreateIntCast(V, CI->getType(), false);
3021}
3022
3023Value *LibCallSimplifier::optimizeAbs(CallInst *CI, IRBuilderBase &B) {
3024 // abs(x) -> x <s 0 ? -x : x
3025 // The negation has 'nsw' because abs of INT_MIN is undefined.
3026 Value *X = CI->getArgOperand(0);
3027 Value *IsNeg = B.CreateIsNeg(X);
3028 Value *NegX = B.CreateNSWNeg(X, "neg");
3029 return B.CreateSelect(IsNeg, NegX, X);
3030}
3031
3032Value *LibCallSimplifier::optimizeIsDigit(CallInst *CI, IRBuilderBase &B) {
3033 // isdigit(c) -> (c-'0') <u 10
3034 Value *Op = CI->getArgOperand(0);
3035 Type *ArgType = Op->getType();
3036 Op = B.CreateSub(Op, ConstantInt::get(ArgType, '0'), "isdigittmp");
3037 Op = B.CreateICmpULT(Op, ConstantInt::get(ArgType, 10), "isdigit");
3038 return B.CreateZExt(Op, CI->getType());
3039}
3040
3041Value *LibCallSimplifier::optimizeIsAscii(CallInst *CI, IRBuilderBase &B) {
3042 // isascii(c) -> c <u 128
3043 Value *Op = CI->getArgOperand(0);
3044 Type *ArgType = Op->getType();
3045 Op = B.CreateICmpULT(Op, ConstantInt::get(ArgType, 128), "isascii");
3046 return B.CreateZExt(Op, CI->getType());
3047}
3048
3049Value *LibCallSimplifier::optimizeToAscii(CallInst *CI, IRBuilderBase &B) {
3050 // toascii(c) -> c & 0x7f
3051 return B.CreateAnd(CI->getArgOperand(0),
3052 ConstantInt::get(CI->getType(), 0x7F));
3053}
3054
3055// Fold calls to atoi, atol, and atoll.
3056Value *LibCallSimplifier::optimizeAtoi(CallInst *CI, IRBuilderBase &B) {
3057 CI->addParamAttr(0, Attribute::NoCapture);
3058
3059 StringRef Str;
3060 if (!getConstantStringInfo(CI->getArgOperand(0), Str))
3061 return nullptr;
3062
3063 return convertStrToInt(CI, Str, nullptr, 10, /*AsSigned=*/true, B);
3064}
3065
3066// Fold calls to strtol, strtoll, strtoul, and strtoull.
3067Value *LibCallSimplifier::optimizeStrToInt(CallInst *CI, IRBuilderBase &B,
3068 bool AsSigned) {
3069 Value *EndPtr = CI->getArgOperand(1);
3070 if (isa<ConstantPointerNull>(EndPtr)) {
3071 // With a null EndPtr, this function won't capture the main argument.
3072 // It would be readonly too, except that it still may write to errno.
3073 CI->addParamAttr(0, Attribute::NoCapture);
3074 EndPtr = nullptr;
3075 } else if (!isKnownNonZero(EndPtr, DL))
3076 return nullptr;
3077
3078 StringRef Str;
3079 if (!getConstantStringInfo(CI->getArgOperand(0), Str))
3080 return nullptr;
3081
3082 if (ConstantInt *CInt = dyn_cast<ConstantInt>(CI->getArgOperand(2))) {
3083 return convertStrToInt(CI, Str, EndPtr, CInt->getSExtValue(), AsSigned, B);
3084 }
3085
3086 return nullptr;
3087}
3088
3089//===----------------------------------------------------------------------===//
3090// Formatting and IO Library Call Optimizations
3091//===----------------------------------------------------------------------===//
3092
3093static bool isReportingError(Function *Callee, CallInst *CI, int StreamArg);
3094
3095Value *LibCallSimplifier::optimizeErrorReporting(CallInst *CI, IRBuilderBase &B,
3096 int StreamArg) {
3098 // Error reporting calls should be cold, mark them as such.
3099 // This applies even to non-builtin calls: it is only a hint and applies to
3100 // functions that the frontend might not understand as builtins.
3101
3102 // This heuristic was suggested in:
3103 // Improving Static Branch Prediction in a Compiler
3104 // Brian L. Deitrich, Ben-Chung Cheng, Wen-mei W. Hwu
3105 // Proceedings of PACT'98, Oct. 1998, IEEE
3106 if (!CI->hasFnAttr(Attribute::Cold) &&
3107 isReportingError(Callee, CI, StreamArg)) {
3108 CI->addFnAttr(Attribute::Cold);
3109 }
3110
3111 return nullptr;
3112}
3113
3114static bool isReportingError(Function *Callee, CallInst *CI, int StreamArg) {
3115 if (!Callee || !Callee->isDeclaration())
3116 return false;
3117
3118 if (StreamArg < 0)
3119 return true;
3120
3121 // These functions might be considered cold, but only if their stream
3122 // argument is stderr.
3123
3124 if (StreamArg >= (int)CI->arg_size())
3125 return false;
3126 LoadInst *LI = dyn_cast<LoadInst>(CI->getArgOperand(StreamArg));
3127 if (!LI)
3128 return false;
3129 GlobalVariable *GV = dyn_cast<GlobalVariable>(LI->getPointerOperand());
3130 if (!GV || !GV->isDeclaration())
3131 return false;
3132 return GV->getName() == "stderr";
3133}
3134
3135Value *LibCallSimplifier::optimizePrintFString(CallInst *CI, IRBuilderBase &B) {
3136 // Check for a fixed format string.
3137 StringRef FormatStr;
3138 if (!getConstantStringInfo(CI->getArgOperand(0), FormatStr))
3139 return nullptr;
3140
3141 // Empty format string -> noop.
3142 if (FormatStr.empty()) // Tolerate printf's declared void.
3143 return CI->use_empty() ? (Value *)CI : ConstantInt::get(CI->getType(), 0);
3144
3145 // Do not do any of the following transformations if the printf return value
3146 // is used, in general the printf return value is not compatible with either
3147 // putchar() or puts().
3148 if (!CI->use_empty())
3149 return nullptr;
3150
3151 Type *IntTy = CI->getType();
3152 // printf("x") -> putchar('x'), even for "%" and "%%".
3153 if (FormatStr.size() == 1 || FormatStr == "%%") {
3154 // Convert the character to unsigned char before passing it to putchar
3155 // to avoid host-specific sign extension in the IR. Putchar converts
3156 // it to unsigned char regardless.
3157 Value *IntChar = ConstantInt::get(IntTy, (unsigned char)FormatStr[0]);
3158 return copyFlags(*CI, emitPutChar(IntChar, B, TLI));
3159 }
3160
3161 // Try to remove call or emit putchar/puts.
3162 if (FormatStr == "%s" && CI->arg_size() > 1) {
3163 StringRef OperandStr;
3164 if (!getConstantStringInfo(CI->getOperand(1), OperandStr))
3165 return nullptr;
3166 // printf("%s", "") --> NOP
3167 if (OperandStr.empty())
3168 return (Value *)CI;
3169 // printf("%s", "a") --> putchar('a')
3170 if (OperandStr.size() == 1) {
3171 // Convert the character to unsigned char before passing it to putchar
3172 // to avoid host-specific sign extension in the IR. Putchar converts
3173 // it to unsigned char regardless.
3174 Value *IntChar = ConstantInt::get(IntTy, (unsigned char)OperandStr[0]);
3175 return copyFlags(*CI, emitPutChar(IntChar, B, TLI));
3176 }
3177 // printf("%s", str"\n") --> puts(str)
3178 if (OperandStr.back() == '\n') {
3179 OperandStr = OperandStr.drop_back();
3180 Value *GV = B.CreateGlobalString(OperandStr, "str");
3181 return copyFlags(*CI, emitPutS(GV, B, TLI));
3182 }
3183 return nullptr;
3184 }
3185
3186 // printf("foo\n") --> puts("foo")
3187 if (FormatStr.back() == '\n' &&
3188 !FormatStr.contains('%')) { // No format characters.
3189 // Create a string literal with no \n on it. We expect the constant merge
3190 // pass to be run after this pass, to merge duplicate strings.
3191 FormatStr = FormatStr.drop_back();
3192 Value *GV = B.CreateGlobalString(FormatStr, "str");
3193 return copyFlags(*CI, emitPutS(GV, B, TLI));
3194 }
3195
3196 // Optimize specific format strings.
3197 // printf("%c", chr) --> putchar(chr)
3198 if (FormatStr == "%c" && CI->arg_size() > 1 &&
3199 CI->getArgOperand(1)->getType()->isIntegerTy()) {
3200 // Convert the argument to the type expected by putchar, i.e., int, which
3201 // need not be 32 bits wide but which is the same as printf's return type.
3202 Value *IntChar = B.CreateIntCast(CI->getArgOperand(1), IntTy, false);
3203 return copyFlags(*CI, emitPutChar(IntChar, B, TLI));
3204 }
3205
3206 // printf("%s\n", str) --> puts(str)
3207 if (FormatStr == "%s\n" && CI->arg_size() > 1 &&
3208 CI->getArgOperand(1)->getType()->isPointerTy())
3209 return copyFlags(*CI, emitPutS(CI->getArgOperand(1), B, TLI));
3210 return nullptr;
3211}
3212
3213Value *LibCallSimplifier::optimizePrintF(CallInst *CI, IRBuilderBase &B) {
3214
3215 Module *M = CI->getModule();
3217 FunctionType *FT = Callee->getFunctionType();
3218 if (Value *V = optimizePrintFString(CI, B)) {
3219 return V;
3220 }
3221
3223
3224 // printf(format, ...) -> iprintf(format, ...) if no floating point
3225 // arguments.
3226 if (isLibFuncEmittable(M, TLI, LibFunc_iprintf) &&
3228 FunctionCallee IPrintFFn = getOrInsertLibFunc(M, *TLI, LibFunc_iprintf, FT,
3229 Callee->getAttributes());
3230 CallInst *New = cast<CallInst>(CI->clone());
3231 New->setCalledFunction(IPrintFFn);
3232 B.Insert(New);
3233 return New;
3234 }
3235
3236 // printf(format, ...) -> __small_printf(format, ...) if no 128-bit floating point
3237 // arguments.
3238 if (isLibFuncEmittable(M, TLI, LibFunc_small_printf) &&
3239 !callHasFP128Argument(CI)) {
3240 auto SmallPrintFFn = getOrInsertLibFunc(M, *TLI, LibFunc_small_printf, FT,
3241 Callee->getAttributes());
3242 CallInst *New = cast<CallInst>(CI->clone());
3243 New->setCalledFunction(SmallPrintFFn);
3244 B.Insert(New);
3245 return New;
3246 }
3247
3248 return nullptr;
3249}
3250
3251Value *LibCallSimplifier::optimizeSPrintFString(CallInst *CI,
3252 IRBuilderBase &B) {
3253 // Check for a fixed format string.
3254 StringRef FormatStr;
3255 if (!getConstantStringInfo(CI->getArgOperand(1), FormatStr))
3256 return nullptr;
3257
3258 // If we just have a format string (nothing else crazy) transform it.
3259 Value *Dest = CI->getArgOperand(0);
3260 if (CI->arg_size() == 2) {
3261 // Make sure there's no % in the constant array. We could try to handle
3262 // %% -> % in the future if we cared.
3263 if (FormatStr.contains('%'))
3264 return nullptr; // we found a format specifier, bail out.
3265
3266 // sprintf(str, fmt) -> llvm.memcpy(align 1 str, align 1 fmt, strlen(fmt)+1)
3267 B.CreateMemCpy(
3268 Dest, Align(1), CI->getArgOperand(1), Align(1),
3269 ConstantInt::get(DL.getIntPtrType(CI->getContext()),
3270 FormatStr.size() + 1)); // Copy the null byte.
3271 return ConstantInt::get(CI->getType(), FormatStr.size());
3272 }
3273
3274 // The remaining optimizations require the format string to be "%s" or "%c"
3275 // and have an extra operand.
3276 if (FormatStr.size() != 2 || FormatStr[0] != '%' || CI->arg_size() < 3)
3277 return nullptr;
3278
3279 // Decode the second character of the format string.
3280 if (FormatStr[1] == 'c') {
3281 // sprintf(dst, "%c", chr) --> *(i8*)dst = chr; *((i8*)dst+1) = 0
3282 if (!CI->getArgOperand(2)->getType()->isIntegerTy())
3283 return nullptr;
3284 Value *V = B.CreateTrunc(CI->getArgOperand(2), B.getInt8Ty(), "char");
3285 Value *Ptr = Dest;
3286 B.CreateStore(V, Ptr);
3287 Ptr = B.CreateInBoundsGEP(B.getInt8Ty(), Ptr, B.getInt32(1), "nul");
3288 B.CreateStore(B.getInt8(0), Ptr);
3289
3290 return ConstantInt::get(CI->getType(), 1);
3291 }
3292
3293 if (FormatStr[1] == 's') {
3294 // sprintf(dest, "%s", str) -> llvm.memcpy(align 1 dest, align 1 str,
3295 // strlen(str)+1)
3296 if (!CI->getArgOperand(2)->getType()->isPointerTy())
3297 return nullptr;
3298
3299 if (CI->use_empty())
3300 // sprintf(dest, "%s", str) -> strcpy(dest, str)
3301 return copyFlags(*CI, emitStrCpy(Dest, CI->getArgOperand(2), B, TLI));
3302
3303 uint64_t SrcLen = GetStringLength(CI->getArgOperand(2));
3304 if (SrcLen) {
3305 B.CreateMemCpy(
3306 Dest, Align(1), CI->getArgOperand(2), Align(1),
3307 ConstantInt::get(DL.getIntPtrType(CI->getContext()), SrcLen));
3308 // Returns total number of characters written without null-character.
3309 return ConstantInt::get(CI->getType(), SrcLen - 1);
3310 } else if (Value *V = emitStpCpy(Dest, CI->getArgOperand(2), B, TLI)) {
3311 // sprintf(dest, "%s", str) -> stpcpy(dest, str) - dest
3312 Value *PtrDiff = B.CreatePtrDiff(B.getInt8Ty(), V, Dest);
3313 return B.CreateIntCast(PtrDiff, CI->getType(), false);
3314 }
3315
3316 bool OptForSize = CI->getFunction()->hasOptSize() ||
3317 llvm::shouldOptimizeForSize(CI->getParent(), PSI, BFI,
3319 if (OptForSize)
3320 return nullptr;
3321
3322 Value *Len = emitStrLen(CI->getArgOperand(2), B, DL, TLI);
3323 if (!Len)
3324 return nullptr;
3325 Value *IncLen =
3326 B.CreateAdd(Len, ConstantInt::get(Len->getType(), 1), "leninc");
3327 B.CreateMemCpy(Dest, Align(1), CI->getArgOperand(2), Align(1), IncLen);
3328
3329 // The sprintf result is the unincremented number of bytes in the string.
3330 return B.CreateIntCast(Len, CI->getType(), false);
3331 }
3332 return nullptr;
3333}
3334
3335Value *LibCallSimplifier::optimizeSPrintF(CallInst *CI, IRBuilderBase &B) {
3336 Module *M = CI->getModule();
3338 FunctionType *FT = Callee->getFunctionType();
3339 if (Value *V = optimizeSPrintFString(CI, B)) {
3340 return V;
3341 }
3342
3344
3345 // sprintf(str, format, ...) -> siprintf(str, format, ...) if no floating
3346 // point arguments.
3347 if (isLibFuncEmittable(M, TLI, LibFunc_siprintf) &&
3349 FunctionCallee SIPrintFFn = getOrInsertLibFunc(M, *TLI, LibFunc_siprintf,
3350 FT, Callee->getAttributes());
3351 CallInst *New = cast<CallInst>(CI->clone());
3352 New->setCalledFunction(SIPrintFFn);
3353 B.Insert(New);
3354 return New;
3355 }
3356
3357 // sprintf(str, format, ...) -> __small_sprintf(str, format, ...) if no 128-bit
3358 // floating point arguments.
3359 if (isLibFuncEmittable(M, TLI, LibFunc_small_sprintf) &&
3360 !callHasFP128Argument(CI)) {
3361 auto SmallSPrintFFn = getOrInsertLibFunc(M, *TLI, LibFunc_small_sprintf, FT,
3362 Callee->getAttributes());
3363 CallInst *New = cast<CallInst>(CI->clone());
3364 New->setCalledFunction(SmallSPrintFFn);
3365 B.Insert(New);
3366 return New;
3367 }
3368
3369 return nullptr;
3370}
3371
3372// Transform an snprintf call CI with the bound N to format the string Str
3373// either to a call to memcpy, or to single character a store, or to nothing,
3374// and fold the result to a constant. A nonnull StrArg refers to the string
3375// argument being formatted. Otherwise the call is one with N < 2 and
3376// the "%c" directive to format a single character.
3377Value *LibCallSimplifier::emitSnPrintfMemCpy(CallInst *CI, Value *StrArg,
3378 StringRef Str, uint64_t N,
3379 IRBuilderBase &B) {
3380 assert(StrArg || (N < 2 && Str.size() == 1));
3381
3382 unsigned IntBits = TLI->getIntSize();
3383 uint64_t IntMax = maxIntN(IntBits);
3384 if (Str.size() > IntMax)
3385 // Bail if the string is longer than INT_MAX. POSIX requires
3386 // implementations to set errno to EOVERFLOW in this case, in
3387 // addition to when N is larger than that (checked by the caller).
3388 return nullptr;
3389
3390 Value *StrLen = ConstantInt::get(CI->getType(), Str.size());
3391 if (N == 0)
3392 return StrLen;
3393
3394 // Set to the number of bytes to copy fron StrArg which is also
3395 // the offset of the terinating nul.
3396 uint64_t NCopy;
3397 if (N > Str.size())
3398 // Copy the full string, including the terminating nul (which must
3399 // be present regardless of the bound).
3400 NCopy = Str.size() + 1;
3401 else
3402 NCopy = N - 1;
3403
3404 Value *DstArg = CI->getArgOperand(0);
3405 if (NCopy && StrArg)
3406 // Transform the call to lvm.memcpy(dst, fmt, N).
3407 copyFlags(
3408 *CI,
3409 B.CreateMemCpy(
3410 DstArg, Align(1), StrArg, Align(1),
3411 ConstantInt::get(DL.getIntPtrType(CI->getContext()), NCopy)));
3412
3413 if (N > Str.size())
3414 // Return early when the whole format string, including the final nul,
3415 // has been copied.
3416 return StrLen;
3417
3418 // Otherwise, when truncating the string append a terminating nul.
3419 Type *Int8Ty = B.getInt8Ty();
3420 Value *NulOff = B.getIntN(IntBits, NCopy);
3421 Value *DstEnd = B.CreateInBoundsGEP(Int8Ty, DstArg, NulOff, "endptr");
3422 B.CreateStore(ConstantInt::get(Int8Ty, 0), DstEnd);
3423 return StrLen;
3424}
3425
3426Value *LibCallSimplifier::optimizeSnPrintFString(CallInst *CI,
3427 IRBuilderBase &B) {
3428 // Check for size
3429 ConstantInt *Size = dyn_cast<ConstantInt>(CI->getArgOperand(1));
3430 if (!Size)
3431 return nullptr;
3432
3433 uint64_t N = Size->getZExtValue();
3434 uint64_t IntMax = maxIntN(TLI->getIntSize());
3435 if (N > IntMax)
3436 // Bail if the bound exceeds INT_MAX. POSIX requires implementations
3437 // to set errno to EOVERFLOW in this case.
3438 return nullptr;
3439
3440 Value *DstArg = CI->getArgOperand(0);
3441 Value *FmtArg = CI->getArgOperand(2);
3442
3443 // Check for a fixed format string.
3444 StringRef FormatStr;
3445 if (!getConstantStringInfo(FmtArg, FormatStr))
3446 return nullptr;
3447
3448 // If we just have a format string (nothing else crazy) transform it.
3449 if (CI->arg_size() == 3) {
3450 if (FormatStr.contains('%'))
3451 // Bail if the format string contains a directive and there are
3452 // no arguments. We could handle "%%" in the future.
3453 return nullptr;
3454
3455 return emitSnPrintfMemCpy(CI, FmtArg, FormatStr, N, B);
3456 }
3457
3458 // The remaining optimizations require the format string to be "%s" or "%c"
3459 // and have an extra operand.
3460 if (FormatStr.size() != 2 || FormatStr[0] != '%' || CI->arg_size() != 4)
3461 return nullptr;
3462
3463 // Decode the second character of the format string.
3464 if (FormatStr[1] == 'c') {
3465 if (N <= 1) {
3466 // Use an arbitary string of length 1 to transform the call into
3467 // either a nul store (N == 1) or a no-op (N == 0) and fold it
3468 // to one.
3469 StringRef CharStr("*");
3470 return emitSnPrintfMemCpy(CI, nullptr, CharStr, N, B);
3471 }
3472
3473 // snprintf(dst, size, "%c", chr) --> *(i8*)dst = chr; *((i8*)dst+1) = 0
3474 if (!CI->getArgOperand(3)->getType()->isIntegerTy())
3475 return nullptr;
3476 Value *V = B.CreateTrunc(CI->getArgOperand(3), B.getInt8Ty(), "char");
3477 Value *Ptr = DstArg;
3478 B.CreateStore(V, Ptr);
3479 Ptr = B.CreateInBoundsGEP(B.getInt8Ty(), Ptr, B.getInt32(1), "nul");
3480 B.CreateStore(B.getInt8(0), Ptr);
3481 return ConstantInt::get(CI->getType(), 1);
3482 }
3483
3484 if (FormatStr[1] != 's')
3485 return nullptr;
3486
3487 Value *StrArg = CI->getArgOperand(3);
3488 // snprintf(dest, size, "%s", str) to llvm.memcpy(dest, str, len+1, 1)
3489 StringRef Str;
3490 if (!getConstantStringInfo(StrArg, Str))
3491 return nullptr;
3492
3493 return emitSnPrintfMemCpy(CI, StrArg, Str, N, B);
3494}
3495
3496Value *LibCallSimplifier::optimizeSnPrintF(CallInst *CI, IRBuilderBase &B) {
3497 if (Value *V = optimizeSnPrintFString(CI, B)) {
3498 return V;
3499 }
3500
3501 if (isKnownNonZero(CI->getOperand(1), DL))
3503 return nullptr;
3504}
3505
3506Value *LibCallSimplifier::optimizeFPrintFString(CallInst *CI,
3507 IRBuilderBase &B) {
3508 optimizeErrorReporting(CI, B, 0);
3509
3510 // All the optimizations depend on the format string.
3511 StringRef FormatStr;
3512 if (!getConstantStringInfo(CI->getArgOperand(1), FormatStr))
3513 return nullptr;
3514
3515 // Do not do any of the following transformations if the fprintf return
3516 // value is used, in general the fprintf return value is not compatible
3517 // with fwrite(), fputc() or fputs().
3518 if (!CI->use_empty())
3519 return nullptr;
3520
3521 // fprintf(F, "foo") --> fwrite("foo", 3, 1, F)
3522 if (CI->arg_size() == 2) {
3523 // Could handle %% -> % if we cared.
3524 if (FormatStr.contains('%'))
3525 return nullptr; // We found a format specifier.
3526
3527 unsigned SizeTBits = TLI->getSizeTSize(*CI->getModule());
3528 Type *SizeTTy = IntegerType::get(CI->getContext(), SizeTBits);
3529 return copyFlags(
3530 *CI, emitFWrite(CI->getArgOperand(1),
3531 ConstantInt::get(SizeTTy, FormatStr.size()),
3532 CI->getArgOperand(0), B, DL, TLI));
3533 }
3534
3535 // The remaining optimizations require the format string to be "%s" or "%c"
3536 // and have an extra operand.
3537 if (FormatStr.size() != 2 || FormatStr[0] != '%' || CI->arg_size() < 3)
3538 return nullptr;
3539
3540 // Decode the second character of the format string.
3541 if (FormatStr[1] == 'c') {
3542 // fprintf(F, "%c", chr) --> fputc((int)chr, F)
3543 if (!CI->getArgOperand(2)->getType()->isIntegerTy())
3544 return nullptr;
3545 Type *IntTy = B.getIntNTy(TLI->getIntSize());
3546 Value *V = B.CreateIntCast(CI->getArgOperand(2), IntTy, /*isSigned*/ true,
3547 "chari");
3548 return copyFlags(*CI, emitFPutC(V, CI->getArgOperand(0), B, TLI));
3549 }
3550
3551 if (FormatStr[1] == 's') {
3552 // fprintf(F, "%s", str) --> fputs(str, F)
3553 if (!CI->getArgOperand(2)->getType()->isPointerTy())
3554 return nullptr;
3555 return copyFlags(
3556 *CI, emitFPutS(CI->getArgOperand(2), CI->getArgOperand(0), B, TLI));
3557 }
3558 return nullptr;
3559}
3560
3561Value *LibCallSimplifier::optimizeFPrintF(CallInst *CI, IRBuilderBase &B) {
3562 Module *M = CI->getModule();
3564 FunctionType *FT = Callee->getFunctionType();
3565 if (Value *V = optimizeFPrintFString(CI, B)) {
3566 return V;
3567 }
3568
3569 // fprintf(stream, format, ...) -> fiprintf(stream, format, ...) if no
3570 // floating point arguments.
3571 if (isLibFuncEmittable(M, TLI, LibFunc_fiprintf) &&
3573 FunctionCallee FIPrintFFn = getOrInsertLibFunc(M, *TLI, LibFunc_fiprintf,
3574 FT, Callee->getAttributes());
3575 CallInst *New = cast<CallInst>(CI->clone());
3576 New->setCalledFunction(FIPrintFFn);
3577 B.Insert(New);
3578 return New;
3579 }
3580
3581 // fprintf(stream, format, ...) -> __small_fprintf(stream, format, ...) if no
3582 // 128-bit floating point arguments.
3583 if (isLibFuncEmittable(M, TLI, LibFunc_small_fprintf) &&
3584 !callHasFP128Argument(CI)) {
3585 auto SmallFPrintFFn =
3586 getOrInsertLibFunc(M, *TLI, LibFunc_small_fprintf, FT,
3587 Callee->getAttributes());
3588 CallInst *New = cast<CallInst>(CI->clone());
3589 New->setCalledFunction(SmallFPrintFFn);
3590 B.Insert(New);
3591 return New;
3592 }
3593
3594 return nullptr;
3595}
3596
3597Value *LibCallSimplifier::optimizeFWrite(CallInst *CI, IRBuilderBase &B) {
3598 optimizeErrorReporting(CI, B, 3);
3599
3600 // Get the element size and count.
3601 ConstantInt *SizeC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
3602 ConstantInt *CountC = dyn_cast<ConstantInt>(CI->getArgOperand(2));
3603 if (SizeC && CountC) {
3604 uint64_t Bytes = SizeC->getZExtValue() * CountC->getZExtValue();
3605
3606 // If this is writing zero records, remove the call (it's a noop).
3607 if (Bytes == 0)
3608 return ConstantInt::get(CI->getType(), 0);
3609
3610 // If this is writing one byte, turn it into fputc.
3611 // This optimisation is only valid, if the return value is unused.
3612 if (Bytes == 1 && CI->use_empty()) { // fwrite(S,1,1,F) -> fputc(S[0],F)
3613 Value *Char = B.CreateLoad(B.getInt8Ty(), CI->getArgOperand(0), "char");
3614 Type *IntTy = B.getIntNTy(TLI->getIntSize());
3615 Value *Cast = B.CreateIntCast(Char, IntTy, /*isSigned*/ true, "chari");
3616 Value *NewCI = emitFPutC(Cast, CI->getArgOperand(3), B, TLI);
3617 return NewCI ? ConstantInt::get(CI->getType(), 1) : nullptr;
3618 }
3619 }
3620
3621 return nullptr;
3622}
3623
3624Value *LibCallSimplifier::optimizeFPuts(CallInst *CI, IRBuilderBase &B) {
3625 optimizeErrorReporting(CI, B, 1);
3626
3627 // Don't rewrite fputs to fwrite when optimising for size because fwrite
3628 // requires more arguments and thus extra MOVs are required.
3629 bool OptForSize = CI->getFunction()->hasOptSize() ||
3630 llvm::shouldOptimizeForSize(CI->getParent(), PSI, BFI,
3632 if (OptForSize)
3633 return nullptr;
3634
3635 // We can't optimize if return value is used.
3636 if (!CI->use_empty())
3637 return nullptr;
3638
3639 // fputs(s,F) --> fwrite(s,strlen(s),1,F)
3641 if (!Len)
3642 return nullptr;
3643
3644 // Known to have no uses (see above).
3645 unsigned SizeTBits = TLI->getSizeTSize(*CI->getModule());
3646 Type *SizeTTy = IntegerType::get(CI->getContext(), SizeTBits);
3647 return copyFlags(
3648 *CI,
3650 ConstantInt::get(SizeTTy, Len - 1),
3651 CI->getArgOperand(1), B, DL, TLI));
3652}
3653
3654Value *LibCallSimplifier::optimizePuts(CallInst *CI, IRBuilderBase &B) {
3656 if (!CI->use_empty())
3657 return nullptr;
3658
3659 // Check for a constant string.
3660 // puts("") -> putchar('\n')
3661 StringRef Str;
3662 if (getConstantStringInfo(CI->getArgOperand(0), Str) && Str.empty()) {
3663 // putchar takes an argument of the same type as puts returns, i.e.,
3664 // int, which need not be 32 bits wide.
3665 Type *IntTy = CI->getType();
3666 return copyFlags(*CI, emitPutChar(ConstantInt::get(IntTy, '\n'), B, TLI));
3667 }
3668
3669 return nullptr;
3670}
3671
3672Value *LibCallSimplifier::optimizeBCopy(CallInst *CI, IRBuilderBase &B) {
3673 // bcopy(src, dst, n) -> llvm.memmove(dst, src, n)
3674 return copyFlags(*CI, B.CreateMemMove(CI->getArgOperand(1), Align(1),
3675 CI->getArgOperand(0), Align(1),
3676 CI->getArgOperand(2)));
3677}
3678
3679bool LibCallSimplifier::hasFloatVersion(const Module *M, StringRef FuncName) {
3680 SmallString<20> FloatFuncName = FuncName;
3681 FloatFuncName += 'f';
3682 return isLibFuncEmittable(M, TLI, FloatFuncName);
3683}
3684
3685Value *LibCallSimplifier::optimizeStringMemoryLibCall(CallInst *CI,
3686 IRBuilderBase &Builder) {
3687 Module *M = CI->getModule();
3688 LibFunc Func;
3690 // Check for string/memory library functions.
3691 if (TLI->getLibFunc(*Callee, Func) && isLibFuncEmittable(M, TLI, Func)) {
3692 // Make sure we never change the calling convention.
3693 assert(
3694 (ignoreCallingConv(Func) ||
3696 "Optimizing string/memory libcall would change the calling convention");
3697 switch (Func) {
3698 case LibFunc_strcat:
3699 return optimizeStrCat(CI, Builder);
3700 case LibFunc_strncat:
3701 return optimizeStrNCat(CI, Builder);
3702 case LibFunc_strchr:
3703 return optimizeStrChr(CI, Builder);
3704 case LibFunc_strrchr:
3705 return optimizeStrRChr(CI, Builder);
3706 case LibFunc_strcmp:
3707 return optimizeStrCmp(CI, Builder);
3708 case LibFunc_strncmp:
3709 return optimizeStrNCmp(CI, Builder);
3710 case LibFunc_strcpy:
3711 return optimizeStrCpy(CI, Builder);
3712 case LibFunc_stpcpy:
3713 return optimizeStpCpy(CI, Builder);
3714 case LibFunc_strlcpy:
3715 return optimizeStrLCpy(CI, Builder);
3716 case LibFunc_stpncpy:
3717 return optimizeStringNCpy(CI, /*RetEnd=*/true, Builder);
3718 case LibFunc_strncpy:
3719 return optimizeStringNCpy(CI, /*RetEnd=*/false, Builder);
3720 case LibFunc_strlen:
3721 return optimizeStrLen(CI, Builder);
3722 case LibFunc_strnlen:
3723 return optimizeStrNLen(CI, Builder);
3724 case LibFunc_strpbrk:
3725 return optimizeStrPBrk(CI, Builder);
3726 case LibFunc_strndup:
3727 return optimizeStrNDup(CI, Builder);
3728 case LibFunc_strtol:
3729 case LibFunc_strtod:
3730 case LibFunc_strtof:
3731 case LibFunc_strtoul:
3732 case LibFunc_strtoll:
3733 case LibFunc_strtold:
3734 case LibFunc_strtoull:
3735 return optimizeStrTo(CI, Builder);
3736 case LibFunc_strspn:
3737 return optimizeStrSpn(CI, Builder);
3738 case LibFunc_strcspn:
3739 return optimizeStrCSpn(CI, Builder);
3740 case LibFunc_strstr:
3741 return optimizeStrStr(CI, Builder);
3742 case LibFunc_memchr:
3743 return optimizeMemChr(CI, Builder);
3744 case LibFunc_memrchr:
3745 return optimizeMemRChr(CI, Builder);
3746 case LibFunc_bcmp:
3747 return optimizeBCmp(CI, Builder);
3748 case LibFunc_memcmp:
3749 return optimizeMemCmp(CI, Builder);
3750 case LibFunc_memcpy:
3751 return optimizeMemCpy(CI, Builder);
3752 case LibFunc_memccpy:
3753 return optimizeMemCCpy(CI, Builder);
3754 case LibFunc_mempcpy:
3755 return optimizeMemPCpy(CI, Builder);
3756 case LibFunc_memmove:
3757 return optimizeMemMove(CI, Builder);
3758 case LibFunc_memset:
3759 return optimizeMemSet(CI, Builder);
3760 case LibFunc_realloc:
3761 return optimizeRealloc(CI, Builder);
3762 case LibFunc_wcslen:
3763 return optimizeWcslen(CI, Builder);
3764 case LibFunc_bcopy:
3765 return optimizeBCopy(CI, Builder);
3766 case LibFunc_Znwm:
3767 case LibFunc_ZnwmRKSt9nothrow_t:
3768 case LibFunc_ZnwmSt11align_val_t:
3769 case LibFunc_ZnwmSt11align_val_tRKSt9nothrow_t:
3770 case LibFunc_Znam:
3771 case LibFunc_ZnamRKSt9nothrow_t:
3772 case LibFunc_ZnamSt11align_val_t:
3773 case LibFunc_ZnamSt11align_val_tRKSt9nothrow_t:
3774 case LibFunc_Znwm12__hot_cold_t:
3775 case LibFunc_ZnwmRKSt9nothrow_t12__hot_cold_t:
3776 case LibFunc_ZnwmSt11align_val_t12__hot_cold_t:
3777 case LibFunc_ZnwmSt11align_val_tRKSt9nothrow_t12__hot_cold_t:
3778 case LibFunc_Znam12__hot_cold_t:
3779 case LibFunc_ZnamRKSt9nothrow_t12__hot_cold_t:
3780 case LibFunc_ZnamSt11align_val_t12__hot_cold_t:
3781 case LibFunc_ZnamSt11align_val_tRKSt9nothrow_t12__hot_cold_t:
3782 return optimizeNew(CI, Builder, Func);
3783 default:
3784 break;
3785 }
3786 }
3787 return nullptr;
3788}
3789
3790Value *LibCallSimplifier::optimizeFloatingPointLibCall(CallInst *CI,
3791 LibFunc Func,
3792 IRBuilderBase &Builder) {
3793 const Module *M = CI->getModule();
3794
3795 // Don't optimize calls that require strict floating point semantics.
3796 if (CI->isStrictFP())
3797 return nullptr;
3798
3799 if (Value *V = optimizeSymmetric(CI, Func, Builder))
3800 return V;
3801
3802 switch (Func) {
3803 case LibFunc_sinpif:
3804 case LibFunc_sinpi:
3805 return optimizeSinCosPi(CI, /*IsSin*/true, Builder);
3806 case LibFunc_cospif:
3807 case LibFunc_cospi:
3808 return optimizeSinCosPi(CI, /*IsSin*/false, Builder);
3809 case LibFunc_powf:
3810 case LibFunc_pow:
3811 case LibFunc_powl:
3812 return optimizePow(CI, Builder);
3813 case LibFunc_exp2l:
3814 case LibFunc_exp2:
3815 case LibFunc_exp2f:
3816 return optimizeExp2(CI, Builder);
3817 case LibFunc_fabsf:
3818 case LibFunc_fabs:
3819 case LibFunc_fabsl:
3820 return replaceUnaryCall(CI, Builder, Intrinsic::fabs);
3821 case LibFunc_sqrtf:
3822 case LibFunc_sqrt:
3823 case LibFunc_sqrtl:
3824 return optimizeSqrt(CI, Builder);
3825 case LibFunc_logf:
3826 case LibFunc_log:
3827 case LibFunc_logl:
3828 case LibFunc_log10f:
3829 case LibFunc_log10:
3830 case LibFunc_log10l:
3831 case LibFunc_log1pf:
3832 case LibFunc_log1p:
3833 case LibFunc_log1pl:
3834 case LibFunc_log2f:
3835 case LibFunc_log2:
3836 case LibFunc_log2l:
3837 case LibFunc_logbf:
3838 case LibFunc_logb:
3839 case LibFunc_logbl:
3840 return optimizeLog(CI, Builder);
3841 case LibFunc_tan:
3842 case LibFunc_tanf:
3843 case LibFunc_tanl:
3844 case LibFunc_sinh:
3845 case LibFunc_sinhf:
3846 case LibFunc_sinhl:
3847 case LibFunc_asinh:
3848 case LibFunc_asinhf:
3849 case LibFunc_asinhl:
3850 case LibFunc_cosh:
3851 case LibFunc_coshf:
3852 case LibFunc_coshl:
3853 case LibFunc_atanh:
3854 case LibFunc_atanhf:
3855 case LibFunc_atanhl:
3856 return optimizeTrigInversionPairs(CI, Builder);
3857 case LibFunc_ceil:
3858 return replaceUnaryCall(CI, Builder, Intrinsic::ceil);
3859 case LibFunc_floor:
3860 return replaceUnaryCall(CI, Builder, Intrinsic::floor);
3861 case LibFunc_round:
3862 return replaceUnaryCall(CI, Builder, Intrinsic::round);
3863 case LibFunc_roundeven:
3864 return replaceUnaryCall(CI, Builder, Intrinsic::roundeven);
3865 case LibFunc_nearbyint:
3866 return replaceUnaryCall(CI, Builder, Intrinsic::nearbyint);
3867 case LibFunc_rint:
3868 return replaceUnaryCall(CI, Builder, Intrinsic::rint);
3869 case LibFunc_trunc:
3870 return replaceUnaryCall(CI, Builder, Intrinsic::trunc);
3871 case LibFunc_acos:
3872 case LibFunc_acosh:
3873 case LibFunc_asin:
3874 case LibFunc_atan:
3875 case LibFunc_cbrt:
3876 case LibFunc_exp:
3877 case LibFunc_exp10:
3878 case LibFunc_expm1:
3879 case LibFunc_cos:
3880 case LibFunc_sin:
3881 case LibFunc_tanh:
3882 if (UnsafeFPShrink && hasFloatVersion(M, CI->getCalledFunction()->getName()))
3883 return optimizeUnaryDoubleFP(CI, Builder, TLI, true);
3884 return nullptr;
3885 case LibFunc_copysign:
3886 if (hasFloatVersion(M, CI->getCalledFunction()->getName()))
3887 return optimizeBinaryDoubleFP(CI, Builder, TLI);
3888 return nullptr;
3889 case LibFunc_fminf:
3890 case LibFunc_fmin:
3891 case LibFunc_fminl:
3892 case LibFunc_fmaxf:
3893 case LibFunc_fmax:
3894 case LibFunc_fmaxl:
3895 return optimizeFMinFMax(CI, Builder);
3896 case LibFunc_cabs:
3897 case LibFunc_cabsf:
3898 case LibFunc_cabsl:
3899 return optimizeCAbs(CI, Builder);
3900 default:
3901 return nullptr;
3902 }
3903}
3904
3906 Module *M = CI->getModule();
3907 assert(!CI->isMustTailCall() && "These transforms aren't musttail safe.");
3908
3909 // TODO: Split out the code below that operates on FP calls so that
3910 // we can all non-FP calls with the StrictFP attribute to be
3911 // optimized.
3912 if (CI->isNoBuiltin())
3913 return nullptr;
3914
3915 LibFunc Func;
3916 Function *Callee = CI->getCalledFunction();
3917 bool IsCallingConvC = TargetLibraryInfoImpl::isCallingConvCCompatible(CI);
3918
3920 CI->getOperandBundlesAsDefs(OpBundles);
3921
3923 Builder.setDefaultOperandBundles(OpBundles);
3924
3925 // Command-line parameter overrides instruction attribute.
3926 // This can't be moved to optimizeFloatingPointLibCall() because it may be
3927 // used by the intrinsic optimizations.
3928 if (EnableUnsafeFPShrink.getNumOccurrences() > 0)
3929 UnsafeFPShrink = EnableUnsafeFPShrink;
3930 else if (isa<FPMathOperator>(CI) && CI->isFast())
3931 UnsafeFPShrink = true;
3932
3933 // First, check for intrinsics.
3934 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
3935 if (!IsCallingConvC)
3936 return nullptr;
3937 // The FP intrinsics have corresponding constrained versions so we don't
3938 // need to check for the StrictFP attribute here.
3939 switch (II->getIntrinsicID()) {
3940 case Intrinsic::pow:
3941 return optimizePow(CI, Builder);
3942 case Intrinsic::exp2:
3943 return optimizeExp2(CI, Builder);
3944 case Intrinsic::log:
3945 case Intrinsic::log2:
3946 case Intrinsic::log10:
3947 return optimizeLog(CI, Builder);
3948 case Intrinsic::sqrt:
3949 return optimizeSqrt(CI, Builder);
3950 case Intrinsic::memset:
3951 return optimizeMemSet(CI, Builder);
3952 case Intrinsic::memcpy:
3953 return optimizeMemCpy(CI, Builder);
3954 case Intrinsic::memmove:
3955 return optimizeMemMove(CI, Builder);
3956 default:
3957 return nullptr;
3958 }
3959 }
3960
3961 // Also try to simplify calls to fortified library functions.
3962 if (Value *SimplifiedFortifiedCI =
3963 FortifiedSimplifier.optimizeCall(CI, Builder))
3964 return SimplifiedFortifiedCI;
3965
3966 // Then check for known library functions.
3967 if (TLI->getLibFunc(*Callee, Func) && isLibFuncEmittable(M, TLI, Func)) {
3968 // We never change the calling convention.
3969 if (!ignoreCallingConv(Func) && !IsCallingConvC)
3970 return nullptr;
3971 if (Value *V = optimizeStringMemoryLibCall(CI, Builder))
3972 return V;
3973 if (Value *V = optimizeFloatingPointLibCall(CI, Func, Builder))
3974 return V;
3975 switch (Func) {
3976 case LibFunc_ffs:
3977 case LibFunc_ffsl:
3978 case LibFunc_ffsll:
3979 return optimizeFFS(CI, Builder);
3980 case LibFunc_fls:
3981 case LibFunc_flsl:
3982 case LibFunc_flsll:
3983 return optimizeFls(CI, Builder);
3984 case LibFunc_abs:
3985 case LibFunc_labs:
3986 case LibFunc_llabs:
3987 return optimizeAbs(CI, Builder);
3988 case LibFunc_isdigit:
3989 return optimizeIsDigit(CI, Builder);
3990 case LibFunc_isascii:
3991 return optimizeIsAscii(CI, Builder);
3992 case LibFunc_toascii:
3993 return optimizeToAscii(CI, Builder);
3994 case LibFunc_atoi:
3995 case LibFunc_atol:
3996 case LibFunc_atoll:
3997 return optimizeAtoi(CI, Builder);
3998 case LibFunc_strtol:
3999 case LibFunc_strtoll:
4000 return optimizeStrToInt(CI, Builder, /*AsSigned=*/true);
4001 case LibFunc_strtoul:
4002 case LibFunc_strtoull:
4003 return optimizeStrToInt(CI, Builder, /*AsSigned=*/false);
4004 case LibFunc_printf:
4005 return optimizePrintF(CI, Builder);
4006 case LibFunc_sprintf:
4007 return optimizeSPrintF(CI, Builder);
4008 case LibFunc_snprintf:
4009 return optimizeSnPrintF(CI, Builder);
4010 case LibFunc_fprintf:
4011 return optimizeFPrintF(CI, Builder);
4012 case LibFunc_fwrite:
4013 return optimizeFWrite(CI, Builder);
4014 case LibFunc_fputs:
4015 return optimizeFPuts(CI, Builder);
4016 case LibFunc_puts:
4017 return optimizePuts(CI, Builder);
4018 case LibFunc_perror:
4019 return optimizeErrorReporting(CI, Builder);
4020 case LibFunc_vfprintf:
4021 case LibFunc_fiprintf:
4022 return optimizeErrorReporting(CI, Builder, 0);
4023 default:
4024 return nullptr;
4025 }
4026 }
4027 return nullptr;
4028}
4029
4031 const DataLayout &DL, const TargetLibraryInfo *TLI, AssumptionCache *AC,
4033 ProfileSummaryInfo *PSI,
4034 function_ref<void(Instruction *, Value *)> Replacer,
4035 function_ref<void(Instruction *)> Eraser)
4036 : FortifiedSimplifier(TLI), DL(DL), TLI(TLI), AC(AC), ORE(ORE), BFI(BFI),
4037 PSI(PSI), Replacer(Replacer), Eraser(Eraser) {}
4038
4039void LibCallSimplifier::replaceAllUsesWith(Instruction *I, Value *With) {
4040 // Indirect through the replacer used in this instance.
4041 Replacer(I, With);
4042}
4043
4044void LibCallSimplifier::eraseFromParent(Instruction *I) {
4045 Eraser(I);
4046}
4047
4048// TODO:
4049// Additional cases that we need to add to this file:
4050//
4051// cbrt:
4052// * cbrt(expN(X)) -> expN(x/3)
4053// * cbrt(sqrt(x)) -> pow(x,1/6)
4054// * cbrt(cbrt(x)) -> pow(x,1/9)
4055//
4056// exp, expf, expl:
4057// * exp(log(x)) -> x
4058//
4059// log, logf, logl:
4060// * log(exp(x)) -> x
4061// * log(exp(y)) -> y*log(e)
4062// * log(exp10(y)) -> y*log(10)
4063// * log(sqrt(x)) -> 0.5*log(x)
4064//
4065// pow, powf, powl:
4066// * pow(sqrt(x),y) -> pow(x,y*0.5)
4067// * pow(pow(x,y),z)-> pow(x,y*z)
4068//
4069// signbit:
4070// * signbit(cnst) -> cnst'
4071// * signbit(nncst) -> 0 (if pstv is a non-negative constant)
4072//
4073// sqrt, sqrtf, sqrtl:
4074// * sqrt(expN(x)) -> expN(x*0.5)
4075// * sqrt(Nroot(x)) -> pow(x,1/(2*N))
4076// * sqrt(pow(x,y)) -> pow(|x|,y*0.5)
4077//
4078
4079//===----------------------------------------------------------------------===//
4080// Fortified Library Call Optimizations
4081//===----------------------------------------------------------------------===//
4082
4083bool FortifiedLibCallSimplifier::isFortifiedCallFoldable(
4084 CallInst *CI, unsigned ObjSizeOp, std::optional<unsigned> SizeOp,
4085 std::optional<unsigned> StrOp, std::optional<unsigned> FlagOp) {
4086 // If this function takes a flag argument, the implementation may use it to
4087 // perform extra checks. Don't fold into the non-checking variant.
4088 if (FlagOp) {
4089 ConstantInt *Flag = dyn_cast<ConstantInt>(CI->getArgOperand(*FlagOp));
4090 if (!Flag || !Flag->isZero())
4091 return false;
4092 }
4093
4094 if (SizeOp && CI->getArgOperand(ObjSizeOp) == CI->getArgOperand(*SizeOp))
4095 return true;
4096
4097 if (ConstantInt *ObjSizeCI =
4098 dyn_cast<ConstantInt>(CI->getArgOperand(ObjSizeOp))) {
4099 if (ObjSizeCI->isMinusOne())
4100 return true;
4101 // If the object size wasn't -1 (unknown), bail out if we were asked to.
4102 if (OnlyLowerUnknownSize)
4103 return false;
4104 if (StrOp) {
4106 // If the length is 0 we don't know how long it is and so we can't
4107 // remove the check.
4108 if (Len)
4109 annotateDereferenceableBytes(CI, *StrOp, Len);
4110 else
4111 return false;
4112 return ObjSizeCI->getZExtValue() >= Len;
4113 }
4114
4115 if (SizeOp) {
4116 if (ConstantInt *SizeCI =
4117 dyn_cast<ConstantInt>(CI->getArgOperand(*SizeOp)))
4118 return ObjSizeCI->getZExtValue() >= SizeCI->getZExtValue();
4119 }
4120 }
4121 return false;
4122}
4123
4124Value *FortifiedLibCallSimplifier::optimizeMemCpyChk(CallInst *CI,
4125 IRBuilderBase &B) {
4126 if (isFortifiedCallFoldable(CI, 3, 2)) {
4127 CallInst *NewCI =
4128 B.CreateMemCpy(CI->getArgOperand(0), Align(1), CI->getArgOperand(1),
4129 Align(1), CI->getArgOperand(2));
4130 mergeAttributesAndFlags(NewCI, *CI);
4131 return CI->getArgOperand(0);
4132 }
4133 return nullptr;
4134}
4135
4136Value *FortifiedLibCallSimplifier::optimizeMemMoveChk(CallInst *CI,
4137 IRBuilderBase &B) {
4138 if (isFortifiedCallFoldable(CI, 3, 2)) {
4139 CallInst *NewCI =
4140 B.CreateMemMove(CI->getArgOperand(0), Align(1), CI->getArgOperand(1),
4141 Align(1), CI->getArgOperand(2));
4142 mergeAttributesAndFlags(NewCI, *CI);
4143 return CI->getArgOperand(0);
4144 }
4145 return nullptr;
4146}
4147
4148Value *FortifiedLibCallSimplifier::optimizeMemSetChk(CallInst *CI,
4149 IRBuilderBase &B) {
4150 if (isFortifiedCallFoldable(CI, 3, 2)) {
4151 Value *Val = B.CreateIntCast(CI->getArgOperand(1), B.getInt8Ty(), false);
4152 CallInst *NewCI = B.CreateMemSet(CI->getArgOperand(0), Val,
4153 CI->getArgOperand(2), Align(1));
4154 mergeAttributesAndFlags(NewCI, *CI);
4155 return CI->getArgOperand(0);
4156 }
4157 return nullptr;
4158}
4159
4160Value *FortifiedLibCallSimplifier::optimizeMemPCpyChk(CallInst *CI,
4161 IRBuilderBase &B) {
4162 const DataLayout &DL = CI->getModule()->getDataLayout();
4163 if (isFortifiedCallFoldable(CI, 3, 2))
4164 if (Value *Call = emitMemPCpy(CI->getArgOperand(0), CI->getArgOperand(1),
4165 CI->getArgOperand(2), B, DL, TLI)) {
4166 return mergeAttributesAndFlags(cast<CallInst>(Call), *CI);
4167 }
4168 return nullptr;
4169}
4170
4171Value *FortifiedLibCallSimplifier::optimizeStrpCpyChk(CallInst *CI,
4173 LibFunc Func) {
4174 const DataLayout &DL = CI->getModule()->getDataLayout();
4175 Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1),
4176 *ObjSize = CI->getArgOperand(2);
4177
4178 // __stpcpy_chk(x,x,...) -> x+strlen(x)
4179 if (Func == LibFunc_stpcpy_chk && !OnlyLowerUnknownSize && Dst == Src) {
4180 Value *StrLen = emitStrLen(Src, B, DL, TLI);
4181 return StrLen ? B.CreateInBoundsGEP(B.getInt8Ty(), Dst, StrLen) : nullptr;
4182 }
4183
4184 // If a) we don't have any length information, or b) we know this will
4185 // fit then just lower to a plain st[rp]cpy. Otherwise we'll keep our
4186 // st[rp]cpy_chk call which may fail at runtime if the size is too long.
4187 // TODO: It might be nice to get a maximum length out of the possible
4188 // string lengths for varying.
4189 if (isFortifiedCallFoldable(CI, 2, std::nullopt, 1)) {
4190 if (Func == LibFunc_strcpy_chk)
4191 return copyFlags(*CI, emitStrCpy(Dst, Src, B, TLI));
4192 else
4193 return copyFlags(*CI, emitStpCpy(Dst, Src, B, TLI));
4194 }
4195
4196 if (OnlyLowerUnknownSize)
4197 return nullptr;
4198
4199 // Maybe we can stil fold __st[rp]cpy_chk to __memcpy_chk.
4201 if (Len)
4202 annotateDereferenceableBytes(CI, 1, Len);
4203 else
4204 return nullptr;
4205
4206 unsigned SizeTBits = TLI->getSizeTSize(*CI->getModule());
4207 Type *SizeTTy = IntegerType::get(CI->getContext(), SizeTBits);
4208 Value *LenV = ConstantInt::get(SizeTTy, Len);
4209 Value *Ret = emitMemCpyChk(Dst, Src, LenV, ObjSize, B, DL, TLI);
4210 // If the function was an __stpcpy_chk, and we were able to fold it into
4211 // a __memcpy_chk, we still need to return the correct end pointer.
4212 if (Ret && Func == LibFunc_stpcpy_chk)
4213 return B.CreateInBoundsGEP(B.getInt8Ty(), Dst,
4214 ConstantInt::get(SizeTTy, Len - 1));
4215 return copyFlags(*CI, cast<CallInst>(Ret));
4216}
4217
4218Value *FortifiedLibCallSimplifier::optimizeStrLenChk(CallInst *CI,
4219 IRBuilderBase &B) {
4220 if (isFortifiedCallFoldable(CI, 1, std::nullopt, 0))
4221 return copyFlags(*CI, emitStrLen(CI->getArgOperand(0), B,
4222 CI->getModule()->getDataLayout(), TLI));
4223 return nullptr;
4224}
4225
4226Value *FortifiedLibCallSimplifier::optimizeStrpNCpyChk(CallInst *CI,
4228 LibFunc Func) {
4229 if (isFortifiedCallFoldable(CI, 3, 2)) {
4230 if (Func == LibFunc_strncpy_chk)
4231 return copyFlags(*CI,
4233 CI->getArgOperand(2), B, TLI));
4234 else
4235 return copyFlags(*CI,
4237 CI->getArgOperand(2), B, TLI));
4238 }
4239
4240 return nullptr;
4241}
4242
4243Value *FortifiedLibCallSimplifier::optimizeMemCCpyChk(CallInst *CI,
4244 IRBuilderBase &B) {
4245 if (isFortifiedCallFoldable(CI, 4, 3))
4246 return copyFlags(
4247 *CI, emitMemCCpy(CI->getArgOperand(0), CI->getArgOperand(1),
4248 CI->getArgOperand(2), CI->getArgOperand(3), B, TLI));
4249
4250 return nullptr;
4251}
4252
4253Value *FortifiedLibCallSimplifier::optimizeSNPrintfChk(CallInst *CI,
4254 IRBuilderBase &B) {
4255 if (isFortifiedCallFoldable(CI, 3, 1, std::nullopt, 2)) {
4256 SmallVector<Value *, 8> VariadicArgs(drop_begin(CI->args(), 5));
4257 return copyFlags(*CI,
4259 CI->getArgOperand(4), VariadicArgs, B, TLI));
4260 }
4261
4262 return nullptr;
4263}
4264
4265Value *FortifiedLibCallSimplifier::optimizeSPrintfChk(CallInst *CI,
4266 IRBuilderBase &B) {
4267 if (isFortifiedCallFoldable(CI, 2, std::nullopt, std::nullopt, 1)) {
4268 SmallVector<Value *, 8> VariadicArgs(drop_begin(CI->args(), 4));
4269 return copyFlags(*CI,
4271 VariadicArgs, B, TLI));
4272 }
4273
4274 return nullptr;
4275}
4276
4277Value *FortifiedLibCallSimplifier::optimizeStrCatChk(CallInst *CI,
4278 IRBuilderBase &B) {
4279 if (isFortifiedCallFoldable(CI, 2))
4280 return copyFlags(
4281 *CI, emitStrCat(CI->getArgOperand(0), CI->getArgOperand(1), B, TLI));
4282
4283 return nullptr;
4284}
4285
4286Value *FortifiedLibCallSimplifier::optimizeStrLCat(CallInst *CI,
4287 IRBuilderBase &B) {
4288 if (isFortifiedCallFoldable(CI, 3))
4289 return copyFlags(*CI,
4291 CI->getArgOperand(2), B, TLI));
4292
4293 return nullptr;
4294}
4295
4296Value *FortifiedLibCallSimplifier::optimizeStrNCatChk(CallInst *CI,
4297 IRBuilderBase &B) {
4298 if (isFortifiedCallFoldable(CI, 3))
4299 return copyFlags(*CI,
4301 CI->getArgOperand(2), B, TLI));
4302
4303 return nullptr;
4304}
4305
4306Value *FortifiedLibCallSimplifier::optimizeStrLCpyChk(CallInst *CI,
4307 IRBuilderBase &B) {
4308 if (isFortifiedCallFoldable(CI, 3))
4309 return copyFlags(*CI,
4311 CI->getArgOperand(2), B, TLI));
4312
4313 return nullptr;
4314}
4315
4316Value *FortifiedLibCallSimplifier::optimizeVSNPrintfChk(CallInst *CI,
4317 IRBuilderBase &B) {
4318 if (isFortifiedCallFoldable(CI, 3, 1, std::nullopt, 2))
4319 return copyFlags(
4320 *CI, emitVSNPrintf(CI->getArgOperand(0), CI->getArgOperand(1),
4321 CI->getArgOperand(4), CI->getArgOperand(5), B, TLI));
4322
4323 return nullptr;
4324}
4325
4326Value *FortifiedLibCallSimplifier::optimizeVSPrintfChk(CallInst *CI,
4327 IRBuilderBase &B) {
4328 if (isFortifiedCallFoldable(CI, 2, std::nullopt, std::nullopt, 1))
4329 return copyFlags(*CI,
4331 CI->getArgOperand(4), B, TLI));
4332
4333 return nullptr;
4334}
4335
4337 IRBuilderBase &Builder) {
4338 // FIXME: We shouldn't be changing "nobuiltin" or TLI unavailable calls here.
4339 // Some clang users checked for _chk libcall availability using:
4340 // __has_builtin(__builtin___memcpy_chk)
4341 // When compiling with -fno-builtin, this is always true.
4342 // When passing -ffreestanding/-mkernel, which both imply -fno-builtin, we
4343 // end up with fortified libcalls, which isn't acceptable in a freestanding
4344 // environment which only provides their non-fortified counterparts.
4345 //
4346 // Until we change clang and/or teach external users to check for availability
4347 // differently, disregard the "nobuiltin" attribute and TLI::has.
4348 //
4349 // PR23093.
4350
4351 LibFunc Func;
4352 Function *Callee = CI->getCalledFunction();
4353 bool IsCallingConvC = TargetLibraryInfoImpl::isCallingConvCCompatible(CI);
4354
4356 CI->getOperandBundlesAsDefs(OpBundles);
4357
4359 Builder.setDefaultOperandBundles(OpBundles);
4360
4361 // First, check that this is a known library functions and that the prototype
4362 // is correct.
4363 if (!TLI->getLibFunc(*Callee, Func))
4364 return nullptr;
4365
4366 // We never change the calling convention.
4367 if (!ignoreCallingConv(Func) && !IsCallingConvC)
4368 return nullptr;
4369
4370 switch (Func) {
4371 case LibFunc_memcpy_chk:
4372 return optimizeMemCpyChk(CI, Builder);
4373 case LibFunc_mempcpy_chk:
4374 return optimizeMemPCpyChk(CI, Builder);
4375 case LibFunc_memmove_chk:
4376 return optimizeMemMoveChk(CI, Builder);
4377 case LibFunc_memset_chk:
4378 return optimizeMemSetChk(CI, Builder);
4379 case LibFunc_stpcpy_chk:
4380 case LibFunc_strcpy_chk:
4381 return optimizeStrpCpyChk(CI, Builder, Func);
4382 case LibFunc_strlen_chk:
4383 return optimizeStrLenChk(CI, Builder);
4384 case LibFunc_stpncpy_chk:
4385 case LibFunc_strncpy_chk:
4386 return optimizeStrpNCpyChk(CI, Builder, Func);
4387 case LibFunc_memccpy_chk:
4388 return optimizeMemCCpyChk(CI, Builder);
4389 case LibFunc_snprintf_chk:
4390 return optimizeSNPrintfChk(CI, Builder);
4391 case LibFunc_sprintf_chk:
4392 return optimizeSPrintfChk(CI, Builder);
4393 case LibFunc_strcat_chk:
4394 return optimizeStrCatChk(CI, Builder);
4395 case LibFunc_strlcat_chk:
4396 return optimizeStrLCat(CI, Builder);
4397 case LibFunc_strncat_chk:
4398 return optimizeStrNCatChk(CI, Builder);
4399 case LibFunc_strlcpy_chk:
4400 return optimizeStrLCpyChk(CI, Builder);
4401 case LibFunc_vsnprintf_chk:
4402 return optimizeVSNPrintfChk(CI, Builder);
4403 case LibFunc_vsprintf_chk:
4404 return optimizeVSPrintfChk(CI, Builder);
4405 default:
4406 break;
4407 }
4408 return nullptr;
4409}
4410
4412 const TargetLibraryInfo *TLI, bool OnlyLowerUnknownSize)
4413 : TLI(TLI), OnlyLowerUnknownSize(OnlyLowerUnknownSize) {}
MachineBasicBlock MachineBasicBlock::iterator DebugLoc DL
static const LLT S1
This file implements the APSInt class, which is a simple class that represents an arbitrary sized int...
static GCRegistry::Add< OcamlGC > B("ocaml", "ocaml 3.10-compatible GC")
return RetTy
std::string Name
uint64_t Size
static GCMetadataPrinterRegistry::Add< ErlangGCPrinter > X("erlang", "erlang-compatible garbage collector")
Hexagon Common GEP
#define F(x, y, z)
Definition: MD5.cpp:55
#define I(x, y, z)
Definition: MD5.cpp:58
Module.h This file contains the declarations for the Module class.
uint64_t IntrinsicInst * II
static GCMetadataPrinterRegistry::Add< OcamlGCMetadataPrinter > Y("ocaml", "ocaml 3.10-compatible collector")
static bool isBinary(MachineInstr &MI)
const SmallVectorImpl< MachineOperand > & Cond
static bool isDigit(const char C)
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
static bool isOnlyUsedInEqualityComparison(Value *V, Value *With)
Return true if it is only used in equality comparisons with With.
static void annotateNonNullAndDereferenceable(CallInst *CI, ArrayRef< unsigned > ArgNos, Value *Size, const DataLayout &DL)
static cl::opt< unsigned, false, HotColdHintParser > ColdNewHintValue("cold-new-hint-value", cl::Hidden, cl::init(1), cl::desc("Value to pass to hot/cold operator new for cold allocation"))
static bool insertSinCosCall(IRBuilderBase &B, Function *OrigCallee, Value *Arg, bool UseFloat, Value *&Sin, Value *&Cos, Value *&SinCos, const TargetLibraryInfo *TLI)
static bool canTransformToMemCmp(CallInst *CI, Value *Str, uint64_t Len, const DataLayout &DL)
static Value * mergeAttributesAndFlags(CallInst *NewCI, const CallInst &Old)
static cl::opt< bool > OptimizeHotColdNew("optimize-hot-cold-new", cl::Hidden, cl::init(false), cl::desc("Enable hot/cold operator new library calls"))
static Value * optimizeBinaryDoubleFP(CallInst *CI, IRBuilderBase &B, const TargetLibraryInfo *TLI, bool isPrecise=false)
Shrink double -> float for binary functions.
static bool ignoreCallingConv(LibFunc Func)
static cl::opt< bool > OptimizeExistingHotColdNew("optimize-existing-hot-cold-new", cl::Hidden, cl::init(false), cl::desc("Enable optimization of existing hot/cold operator new library calls"))
static void annotateDereferenceableBytes(CallInst *CI, ArrayRef< unsigned > ArgNos, uint64_t DereferenceableBytes)
static bool isReportingError(Function *Callee, CallInst *CI, int StreamArg)
static Value * optimizeDoubleFP(CallInst *CI, IRBuilderBase &B, bool isBinary, const TargetLibraryInfo *TLI, bool isPrecise=false)
Shrink double -> float functions.
static Value * optimizeSymmetricCall(CallInst *CI, bool IsEven, IRBuilderBase &B)
static Value * getSqrtCall(Value *V, AttributeList Attrs, bool NoErrno, Module *M, IRBuilderBase &B, const TargetLibraryInfo *TLI)
static Value * valueHasFloatPrecision(Value *Val)
Return a variant of Val with float type.
static Value * optimizeMemCmpConstantSize(CallInst *CI, Value *LHS, Value *RHS, uint64_t Len, IRBuilderBase &B, const DataLayout &DL)
static Value * createPowWithIntegerExponent(Value *Base, Value *Expo, Module *M, IRBuilderBase &B)
static Value * convertStrToInt(CallInst *CI, StringRef &Str, Value *EndPtr, uint64_t Base, bool AsSigned, IRBuilderBase &B)
static Value * memChrToCharCompare(CallInst *CI, Value *NBytes, IRBuilderBase &B, const DataLayout &DL)
static Value * copyFlags(const CallInst &Old, Value *New)
static StringRef substr(StringRef Str, uint64_t Len)
static cl::opt< unsigned, false, HotColdHintParser > HotNewHintValue("hot-new-hint-value", cl::Hidden, cl::init(254), cl::desc("Value to pass to hot/cold operator new for hot allocation"))
static bool isTrigLibCall(CallInst *CI)
static bool isOnlyUsedInComparisonWithZero(Value *V)
static Value * replaceUnaryCall(CallInst *CI, IRBuilderBase &B, Intrinsic::ID IID)
static bool callHasFloatingPointArgument(const CallInst *CI)
static Value * optimizeUnaryDoubleFP(CallInst *CI, IRBuilderBase &B, const TargetLibraryInfo *TLI, bool isPrecise=false)
Shrink double -> float for unary functions.
static bool callHasFP128Argument(const CallInst *CI)
static void annotateNonNullNoUndefBasedOnAccess(CallInst *CI, ArrayRef< unsigned > ArgNos)
static Value * optimizeMemCmpVarSize(CallInst *CI, Value *LHS, Value *RHS, Value *Size, bool StrNCmp, IRBuilderBase &B, const DataLayout &DL)
static Value * getIntToFPVal(Value *I2F, IRBuilderBase &B, unsigned DstWidth)
static cl::opt< bool > EnableUnsafeFPShrink("enable-double-float-shrink", cl::Hidden, cl::init(false), cl::desc("Enable unsafe double to float " "shrinking for math lib calls"))
static cl::opt< unsigned, false, HotColdHintParser > NotColdNewHintValue("notcold-new-hint-value", cl::Hidden, cl::init(128), cl::desc("Value to pass to hot/cold operator new for " "notcold (warm) allocation"))
This file defines the SmallString class.
This file contains some functions that are useful when dealing with strings.
Value * RHS
Value * LHS
bool isFiniteNonZero() const
Definition: APFloat.h:1358
opStatus convert(const fltSemantics &ToSemantics, roundingMode RM, bool *losesInfo)
Definition: APFloat.cpp:5282
bool isNegative() const
Definition: APFloat.h:1348
double convertToDouble() const
Converts this APFloat to host double value.
Definition: APFloat.cpp:5341
bool isExactlyValue(double V) const
We don't rely on operator== working on double values, as it returns true for things that are clearly ...
Definition: APFloat.h:1331
opStatus add(const APFloat &RHS, roundingMode RM)
Definition: APFloat.h:1086
const fltSemantics & getSemantics() const
Definition: APFloat.h:1356
float convertToFloat() const
Converts this APFloat to host float value.
Definition: APFloat.cpp:5369
opStatus convertToInteger(MutableArrayRef< integerPart > Input, unsigned int Width, bool IsSigned, roundingMode RM, bool *IsExact) const
Definition: APFloat.h:1229
bool isInteger() const
Definition: APFloat.h:1365
Class for arbitrary precision integers.
Definition: APInt.h:77
bool ule(const APInt &RHS) const
Unsigned less or equal comparison.
Definition: APInt.h:1129
An arbitrary precision integer that knows its signedness.
Definition: APSInt.h:23
ArrayRef - Represent a constant reference to an array (0 or more elements consecutively in memory),...
Definition: ArrayRef.h:41
A cache of @llvm.assume calls within a function.
static AttributeList get(LLVMContext &C, ArrayRef< std::pair< unsigned, Attribute > > Attrs)
Create an AttributeList with the specified parameters in it.
Attribute getFnAttr(Attribute::AttrKind Kind) const
Return the attribute object that exists for the function.
Definition: Attributes.h:847
AttributeSet getParamAttrs(unsigned ArgNo) const
The attributes for the argument or parameter at the given index are returned.
AttributeList addParamAttributes(LLVMContext &C, unsigned ArgNo, const AttrBuilder &B) const
Add an argument attribute to the list.
Definition: Attributes.h:610
MaybeAlign getAlignment() const
Definition: Attributes.cpp:852
static Attribute getWithDereferenceableBytes(LLVMContext &Context, uint64_t Bytes)
Definition: Attributes.cpp:204
StringRef getValueAsString() const
Return the attribute's value as a string.
Definition: Attributes.cpp:349
LLVM Basic Block Representation.
Definition: BasicBlock.h:60
iterator begin()
Instruction iterator methods.
Definition: BasicBlock.h:430
BlockFrequencyInfo pass uses BlockFrequencyInfoImpl implementation to estimate IR basic block frequen...
void addFnAttr(Attribute::AttrKind Kind)
Adds the attribute to the function.
Definition: InstrTypes.h:1859
void getOperandBundlesAsDefs(SmallVectorImpl< OperandBundleDef > &Defs) const
Return the list of operand bundles attached to this instruction as a vector of OperandBundleDefs.
bool isNoBuiltin() const
Return true if the call should not be treated as a call to a builtin.
Definition: InstrTypes.h:2250
void removeParamAttr(unsigned ArgNo, Attribute::AttrKind Kind)
Removes the attribute from the given argument.
Definition: InstrTypes.h:1926
Function * getCalledFunction() const
Returns the function called, or null if this is an indirect function invocation or the function signa...
Definition: InstrTypes.h:1750
bool doesNotAccessMemory(unsigned OpNo) const
Definition: InstrTypes.h:2095
void removeRetAttrs(const AttributeMask &AttrsToRemove)
Removes the attributes from the return value.
Definition: InstrTypes.h:1921
bool hasFnAttr(Attribute::AttrKind Kind) const
Determine whether this call has the given attribute.
Definition: InstrTypes.h:1836
bool isStrictFP() const
Determine if the call requires strict floating point semantics.
Definition: InstrTypes.h:2256
uint64_t getParamDereferenceableBytes(unsigned i) const
Extract the number of dereferenceable bytes for a call or parameter (0=unknown).
Definition: InstrTypes.h:2194
bool paramHasAttr(unsigned ArgNo, Attribute::AttrKind Kind) const
Determine whether the argument or parameter has the given attribute.
void setAttributes(AttributeList A)
Set the parameter attributes for this call.
Definition: InstrTypes.h:1831
bool doesNotThrow() const
Determine if the call cannot unwind.
Definition: InstrTypes.h:2300
Value * getArgOperand(unsigned i) const
Definition: InstrTypes.h:1695
uint64_t getParamDereferenceableOrNullBytes(unsigned i) const
Extract the number of dereferenceable_or_null bytes for a parameter (0=unknown).
Definition: InstrTypes.h:2212
Intrinsic::ID getIntrinsicID() const
Returns the intrinsic ID of the intrinsic called or Intrinsic::not_intrinsic if the called function i...
iterator_range< User::op_iterator > args()
Iteration adapter for range-for loops.
Definition: InstrTypes.h:1686
unsigned arg_size() const
Definition: InstrTypes.h:1693
AttributeList getAttributes() const
Return the parameter attributes for this call.
Definition: InstrTypes.h:1827
void addParamAttr(unsigned ArgNo, Attribute::AttrKind Kind)
Adds the attribute to the indicated argument.
Definition: InstrTypes.h:1879
Function * getCaller()
Helper to get the caller (the parent function).
This class represents a function call, abstracting a target machine's calling convention.
bool isNoTailCall() const
TailCallKind getTailCallKind() const
bool isMustTailCall() const
@ ICMP_UGT
unsigned greater than
Definition: InstrTypes.h:1016
@ ICMP_ULT
unsigned less than
Definition: InstrTypes.h:1018
@ ICMP_ULE
unsigned less or equal
Definition: InstrTypes.h:1019
Predicate getPredicate() const
Return the predicate for this instruction.
Definition: InstrTypes.h:1105
uint64_t getElementAsInteger(unsigned i) const
If this is a sequential container of integers (of any size), return the specified element in the low ...
Definition: Constants.cpp:3037
ConstantFP - Floating Point Values [float, double].
Definition: Constants.h:269
static Constant * getInfinity(Type *Ty, bool Negative=false)
Definition: Constants.cpp:1084
This is the shared class of boolean and integer constants.
Definition: Constants.h:81
bool isOne() const
This is just a convenience method to make client code smaller for a common case.
Definition: Constants.h:212
bool isZero() const
This is just a convenience method to make client code smaller for a common code.
Definition: Constants.h:206
int64_t getSExtValue() const
Return the constant as a 64-bit integer value after it has been sign extended as appropriate for the ...
Definition: Constants.h:161
uint64_t getZExtValue() const
Return the constant as a 64-bit unsigned integer value after it has been zero extended as appropriate...
Definition: Constants.h:155
This is an important base class in LLVM.
Definition: Constant.h:41
static Constant * getNullValue(Type *Ty)
Constructor to create a '0' constant of arbitrary type.
Definition: Constants.cpp:370
This class represents an Operation in the Expression.
A parsed version of the target data layout string in and methods for querying it.
Definition: DataLayout.h:110
IntegerType * getIntPtrType(LLVMContext &C, unsigned AddressSpace=0) const
Returns an integer type with size at least as big as that of a pointer in the given address space.
Definition: DataLayout.cpp:878
bool fitsInLegalInteger(unsigned Width) const
Returns true if the specified type fits in a native integer type supported by the CPU.
Definition: DataLayout.h:359
This class represents an extension of floating point types.
This class represents a truncation of floating point types.
Convenience struct for specifying and reasoning about fast-math flags.
Definition: FMF.h:20
void setNoSignedZeros(bool B=true)
Definition: FMF.h:85
static FastMathFlags getFast()
Definition: FMF.h:51
static FixedVectorType * get(Type *ElementType, unsigned NumElts)
Definition: Type.cpp:692
FortifiedLibCallSimplifier(const TargetLibraryInfo *TLI, bool OnlyLowerUnknownSize=false)
Value * optimizeCall(CallInst *CI, IRBuilderBase &B)
Take the given call instruction and return a more optimal value to replace the instruction with or 0 ...
A handy container for a FunctionType+Callee-pointer pair, which can be passed around as a single enti...
Definition: DerivedTypes.h:168
bool hasOptSize() const
Optimize this function for size (-Os) or minimum size (-Oz).
Definition: Function.h:688
Intrinsic::ID getIntrinsicID() const LLVM_READONLY
getIntrinsicID - This method returns the ID number of the specified function, or Intrinsic::not_intri...
Definition: Function.h:232
AttributeList getAttributes() const
Return the attribute list for this Function.
Definition: Function.h:340
bool isIntrinsic() const
isIntrinsic - Returns true if the function's name starts with "llvm.".
Definition: Function.h:237
bool hasFnAttribute(Attribute::AttrKind Kind) const
Return true if the function has the attribute.
Definition: Function.cpp:680
bool isDeclaration() const
Return true if the primary definition of this global value is outside of the current translation unit...
Definition: Globals.cpp:286
Module * getParent()
Get the module that this global value is contained inside of...
Definition: GlobalValue.h:655
This instruction compares its operands according to the predicate given to the constructor.
Common base class shared among various IRBuilders.
Definition: IRBuilder.h:94
void setDefaultOperandBundles(ArrayRef< OperandBundleDef > OpBundles)
Definition: IRBuilder.h:365
Instruction * clone() const
Create a copy of 'this' instruction that is identical in all ways except the following:
bool hasNoNaNs() const LLVM_READONLY
Determine whether the no-NaNs flag is set.
bool hasNoInfs() const LLVM_READONLY
Determine whether the no-infs flag is set.
bool hasNoSignedZeros() const LLVM_READONLY
Determine whether the no-signed-zeros flag is set.
const Module * getModule() const
Return the module owning the function this instruction belongs to or nullptr it the function does not...
Definition: Instruction.cpp:83
const BasicBlock * getParent() const
Definition: Instruction.h:152
bool isFast() const LLVM_READONLY
Determine whether all fast-math-flags are set.
const Function * getFunction() const
Return the function this instruction belongs to.
Definition: Instruction.cpp:87