1 //===------ SimplifyLibCalls.cpp - Library calls simplifier ---------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the library calls simplifier. It does not implement
11 // any pass, but can't be used by other passes to do simplifications.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Transforms/Utils/SimplifyLibCalls.h"
16 #include "llvm/ADT/APSInt.h"
17 #include "llvm/ADT/SmallString.h"
18 #include "llvm/ADT/StringMap.h"
19 #include "llvm/ADT/Triple.h"
20 #include "llvm/Analysis/ConstantFolding.h"
21 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
22 #include "llvm/Analysis/TargetLibraryInfo.h"
23 #include "llvm/Transforms/Utils/Local.h"
24 #include "llvm/Analysis/ValueTracking.h"
25 #include "llvm/Analysis/CaptureTracking.h"
26 #include "llvm/Analysis/Loads.h"
27 #include "llvm/IR/DataLayout.h"
28 #include "llvm/IR/Function.h"
29 #include "llvm/IR/IRBuilder.h"
30 #include "llvm/IR/IntrinsicInst.h"
31 #include "llvm/IR/Intrinsics.h"
32 #include "llvm/IR/LLVMContext.h"
33 #include "llvm/IR/Module.h"
34 #include "llvm/IR/PatternMatch.h"
35 #include "llvm/Support/CommandLine.h"
36 #include "llvm/Support/KnownBits.h"
37 #include "llvm/Transforms/Utils/BuildLibCalls.h"
40 using namespace PatternMatch;
43 EnableUnsafeFPShrink("enable-double-float-shrink", cl::Hidden,
45 cl::desc("Enable unsafe double to float "
46 "shrinking for math lib calls"));
49 //===----------------------------------------------------------------------===//
51 //===----------------------------------------------------------------------===//
53 static bool ignoreCallingConv(LibFunc Func) {
54 return Func == LibFunc_abs || Func == LibFunc_labs ||
55 Func == LibFunc_llabs || Func == LibFunc_strlen;
58 static bool isCallingConvCCompatible(CallInst *CI) {
59 switch(CI->getCallingConv()) {
62 case llvm::CallingConv::C:
64 case llvm::CallingConv::ARM_APCS:
65 case llvm::CallingConv::ARM_AAPCS:
66 case llvm::CallingConv::ARM_AAPCS_VFP: {
68 // The iOS ABI diverges from the standard in some cases, so for now don't
69 // try to simplify those calls.
70 if (Triple(CI->getModule()->getTargetTriple()).isiOS())
73 auto *FuncTy = CI->getFunctionType();
75 if (!FuncTy->getReturnType()->isPointerTy() &&
76 !FuncTy->getReturnType()->isIntegerTy() &&
77 !FuncTy->getReturnType()->isVoidTy())
80 for (auto Param : FuncTy->params()) {
81 if (!Param->isPointerTy() && !Param->isIntegerTy())
90 /// Return true if it is only used in equality comparisons with With.
91 static bool isOnlyUsedInEqualityComparison(Value *V, Value *With) {
92 for (User *U : V->users()) {
93 if (ICmpInst *IC = dyn_cast<ICmpInst>(U))
94 if (IC->isEquality() && IC->getOperand(1) == With)
96 // Unknown instruction.
102 static bool callHasFloatingPointArgument(const CallInst *CI) {
103 return any_of(CI->operands(), [](const Use &OI) {
104 return OI->getType()->isFloatingPointTy();
108 static Value *convertStrToNumber(CallInst *CI, StringRef &Str, int64_t Base) {
109 if (Base < 2 || Base > 36)
110 // handle special zero base
115 std::string nptr = Str.str();
117 long long int Result = strtoll(nptr.c_str(), &End, Base);
121 // if we assume all possible target locales are ASCII supersets,
122 // then if strtoll successfully parses a number on the host,
123 // it will also successfully parse the same way on the target
127 if (!isIntN(CI->getType()->getPrimitiveSizeInBits(), Result))
130 return ConstantInt::get(CI->getType(), Result);
133 static bool isLocallyOpenedFile(Value *File, CallInst *CI, IRBuilder<> &B,
134 const TargetLibraryInfo *TLI) {
135 CallInst *FOpen = dyn_cast<CallInst>(File);
139 Function *InnerCallee = FOpen->getCalledFunction();
144 if (!TLI->getLibFunc(*InnerCallee, Func) || !TLI->has(Func) ||
145 Func != LibFunc_fopen)
148 inferLibFuncAttributes(*CI->getCalledFunction(), *TLI);
149 if (PointerMayBeCaptured(File, true, true))
155 static bool isOnlyUsedInComparisonWithZero(Value *V) {
156 for (User *U : V->users()) {
157 if (ICmpInst *IC = dyn_cast<ICmpInst>(U))
158 if (Constant *C = dyn_cast<Constant>(IC->getOperand(1)))
159 if (C->isNullValue())
161 // Unknown instruction.
167 static bool canTransformToMemCmp(CallInst *CI, Value *Str, uint64_t Len,
168 const DataLayout &DL) {
169 if (!isOnlyUsedInComparisonWithZero(CI))
172 if (!isDereferenceableAndAlignedPointer(Str, 1, APInt(64, Len), DL))
175 if (CI->getFunction()->hasFnAttribute(Attribute::SanitizeMemory))
181 //===----------------------------------------------------------------------===//
182 // String and Memory Library Call Optimizations
183 //===----------------------------------------------------------------------===//
185 Value *LibCallSimplifier::optimizeStrCat(CallInst *CI, IRBuilder<> &B) {
186 // Extract some information from the instruction
187 Value *Dst = CI->getArgOperand(0);
188 Value *Src = CI->getArgOperand(1);
190 // See if we can get the length of the input string.
191 uint64_t Len = GetStringLength(Src);
194 --Len; // Unbias length.
196 // Handle the simple, do-nothing case: strcat(x, "") -> x
200 return emitStrLenMemCpy(Src, Dst, Len, B);
203 Value *LibCallSimplifier::emitStrLenMemCpy(Value *Src, Value *Dst, uint64_t Len,
205 // We need to find the end of the destination string. That's where the
206 // memory is to be moved to. We just generate a call to strlen.
207 Value *DstLen = emitStrLen(Dst, B, DL, TLI);
211 // Now that we have the destination's length, we must index into the
212 // destination's pointer to get the actual memcpy destination (end of
213 // the string .. we're concatenating).
214 Value *CpyDst = B.CreateGEP(B.getInt8Ty(), Dst, DstLen, "endptr");
216 // We have enough information to now generate the memcpy call to do the
217 // concatenation for us. Make a memcpy to copy the nul byte with align = 1.
218 B.CreateMemCpy(CpyDst, 1, Src, 1,
219 ConstantInt::get(DL.getIntPtrType(Src->getContext()), Len + 1));
223 Value *LibCallSimplifier::optimizeStrNCat(CallInst *CI, IRBuilder<> &B) {
224 // Extract some information from the instruction.
225 Value *Dst = CI->getArgOperand(0);
226 Value *Src = CI->getArgOperand(1);
229 // We don't do anything if length is not constant.
230 if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(CI->getArgOperand(2)))
231 Len = LengthArg->getZExtValue();
235 // See if we can get the length of the input string.
236 uint64_t SrcLen = GetStringLength(Src);
239 --SrcLen; // Unbias length.
241 // Handle the simple, do-nothing cases:
242 // strncat(x, "", c) -> x
243 // strncat(x, c, 0) -> x
244 if (SrcLen == 0 || Len == 0)
247 // We don't optimize this case.
251 // strncat(x, s, c) -> strcat(x, s)
252 // s is constant so the strcat can be optimized further.
253 return emitStrLenMemCpy(Src, Dst, SrcLen, B);
256 Value *LibCallSimplifier::optimizeStrChr(CallInst *CI, IRBuilder<> &B) {
257 Function *Callee = CI->getCalledFunction();
258 FunctionType *FT = Callee->getFunctionType();
259 Value *SrcStr = CI->getArgOperand(0);
261 // If the second operand is non-constant, see if we can compute the length
262 // of the input string and turn this into memchr.
263 ConstantInt *CharC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
265 uint64_t Len = GetStringLength(SrcStr);
266 if (Len == 0 || !FT->getParamType(1)->isIntegerTy(32)) // memchr needs i32.
269 return emitMemChr(SrcStr, CI->getArgOperand(1), // include nul.
270 ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len),
274 // Otherwise, the character is a constant, see if the first argument is
275 // a string literal. If so, we can constant fold.
277 if (!getConstantStringInfo(SrcStr, Str)) {
278 if (CharC->isZero()) // strchr(p, 0) -> p + strlen(p)
279 return B.CreateGEP(B.getInt8Ty(), SrcStr, emitStrLen(SrcStr, B, DL, TLI),
284 // Compute the offset, make sure to handle the case when we're searching for
285 // zero (a weird way to spell strlen).
286 size_t I = (0xFF & CharC->getSExtValue()) == 0
288 : Str.find(CharC->getSExtValue());
289 if (I == StringRef::npos) // Didn't find the char. strchr returns null.
290 return Constant::getNullValue(CI->getType());
292 // strchr(s+n,c) -> gep(s+n+i,c)
293 return B.CreateGEP(B.getInt8Ty(), SrcStr, B.getInt64(I), "strchr");
296 Value *LibCallSimplifier::optimizeStrRChr(CallInst *CI, IRBuilder<> &B) {
297 Value *SrcStr = CI->getArgOperand(0);
298 ConstantInt *CharC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
300 // Cannot fold anything if we're not looking for a constant.
305 if (!getConstantStringInfo(SrcStr, Str)) {
306 // strrchr(s, 0) -> strchr(s, 0)
308 return emitStrChr(SrcStr, '\0', B, TLI);
312 // Compute the offset.
313 size_t I = (0xFF & CharC->getSExtValue()) == 0
315 : Str.rfind(CharC->getSExtValue());
316 if (I == StringRef::npos) // Didn't find the char. Return null.
317 return Constant::getNullValue(CI->getType());
319 // strrchr(s+n,c) -> gep(s+n+i,c)
320 return B.CreateGEP(B.getInt8Ty(), SrcStr, B.getInt64(I), "strrchr");
323 Value *LibCallSimplifier::optimizeStrCmp(CallInst *CI, IRBuilder<> &B) {
324 Value *Str1P = CI->getArgOperand(0), *Str2P = CI->getArgOperand(1);
325 if (Str1P == Str2P) // strcmp(x,x) -> 0
326 return ConstantInt::get(CI->getType(), 0);
328 StringRef Str1, Str2;
329 bool HasStr1 = getConstantStringInfo(Str1P, Str1);
330 bool HasStr2 = getConstantStringInfo(Str2P, Str2);
332 // strcmp(x, y) -> cnst (if both x and y are constant strings)
333 if (HasStr1 && HasStr2)
334 return ConstantInt::get(CI->getType(), Str1.compare(Str2));
336 if (HasStr1 && Str1.empty()) // strcmp("", x) -> -*x
338 B.CreateZExt(B.CreateLoad(Str2P, "strcmpload"), CI->getType()));
340 if (HasStr2 && Str2.empty()) // strcmp(x,"") -> *x
341 return B.CreateZExt(B.CreateLoad(Str1P, "strcmpload"), CI->getType());
343 // strcmp(P, "x") -> memcmp(P, "x", 2)
344 uint64_t Len1 = GetStringLength(Str1P);
345 uint64_t Len2 = GetStringLength(Str2P);
347 return emitMemCmp(Str1P, Str2P,
348 ConstantInt::get(DL.getIntPtrType(CI->getContext()),
349 std::min(Len1, Len2)),
354 if (!HasStr1 && HasStr2) {
355 if (canTransformToMemCmp(CI, Str1P, Len2, DL))
358 ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len2), B, DL,
360 } else if (HasStr1 && !HasStr2) {
361 if (canTransformToMemCmp(CI, Str2P, Len1, DL))
364 ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len1), B, DL,
371 Value *LibCallSimplifier::optimizeStrNCmp(CallInst *CI, IRBuilder<> &B) {
372 Value *Str1P = CI->getArgOperand(0), *Str2P = CI->getArgOperand(1);
373 if (Str1P == Str2P) // strncmp(x,x,n) -> 0
374 return ConstantInt::get(CI->getType(), 0);
376 // Get the length argument if it is constant.
378 if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(CI->getArgOperand(2)))
379 Length = LengthArg->getZExtValue();
383 if (Length == 0) // strncmp(x,y,0) -> 0
384 return ConstantInt::get(CI->getType(), 0);
386 if (Length == 1) // strncmp(x,y,1) -> memcmp(x,y,1)
387 return emitMemCmp(Str1P, Str2P, CI->getArgOperand(2), B, DL, TLI);
389 StringRef Str1, Str2;
390 bool HasStr1 = getConstantStringInfo(Str1P, Str1);
391 bool HasStr2 = getConstantStringInfo(Str2P, Str2);
393 // strncmp(x, y) -> cnst (if both x and y are constant strings)
394 if (HasStr1 && HasStr2) {
395 StringRef SubStr1 = Str1.substr(0, Length);
396 StringRef SubStr2 = Str2.substr(0, Length);
397 return ConstantInt::get(CI->getType(), SubStr1.compare(SubStr2));
400 if (HasStr1 && Str1.empty()) // strncmp("", x, n) -> -*x
402 B.CreateZExt(B.CreateLoad(Str2P, "strcmpload"), CI->getType()));
404 if (HasStr2 && Str2.empty()) // strncmp(x, "", n) -> *x
405 return B.CreateZExt(B.CreateLoad(Str1P, "strcmpload"), CI->getType());
407 uint64_t Len1 = GetStringLength(Str1P);
408 uint64_t Len2 = GetStringLength(Str2P);
411 if (!HasStr1 && HasStr2) {
412 Len2 = std::min(Len2, Length);
413 if (canTransformToMemCmp(CI, Str1P, Len2, DL))
416 ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len2), B, DL,
418 } else if (HasStr1 && !HasStr2) {
419 Len1 = std::min(Len1, Length);
420 if (canTransformToMemCmp(CI, Str2P, Len1, DL))
423 ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len1), B, DL,
430 Value *LibCallSimplifier::optimizeStrCpy(CallInst *CI, IRBuilder<> &B) {
431 Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1);
432 if (Dst == Src) // strcpy(x,x) -> x
435 // See if we can get the length of the input string.
436 uint64_t Len = GetStringLength(Src);
440 // We have enough information to now generate the memcpy call to do the
441 // copy for us. Make a memcpy to copy the nul byte with align = 1.
442 B.CreateMemCpy(Dst, 1, Src, 1,
443 ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len));
447 Value *LibCallSimplifier::optimizeStpCpy(CallInst *CI, IRBuilder<> &B) {
448 Function *Callee = CI->getCalledFunction();
449 Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1);
450 if (Dst == Src) { // stpcpy(x,x) -> x+strlen(x)
451 Value *StrLen = emitStrLen(Src, B, DL, TLI);
452 return StrLen ? B.CreateInBoundsGEP(B.getInt8Ty(), Dst, StrLen) : nullptr;
455 // See if we can get the length of the input string.
456 uint64_t Len = GetStringLength(Src);
460 Type *PT = Callee->getFunctionType()->getParamType(0);
461 Value *LenV = ConstantInt::get(DL.getIntPtrType(PT), Len);
462 Value *DstEnd = B.CreateGEP(B.getInt8Ty(), Dst,
463 ConstantInt::get(DL.getIntPtrType(PT), Len - 1));
465 // We have enough information to now generate the memcpy call to do the
466 // copy for us. Make a memcpy to copy the nul byte with align = 1.
467 B.CreateMemCpy(Dst, 1, Src, 1, LenV);
471 Value *LibCallSimplifier::optimizeStrNCpy(CallInst *CI, IRBuilder<> &B) {
472 Function *Callee = CI->getCalledFunction();
473 Value *Dst = CI->getArgOperand(0);
474 Value *Src = CI->getArgOperand(1);
475 Value *LenOp = CI->getArgOperand(2);
477 // See if we can get the length of the input string.
478 uint64_t SrcLen = GetStringLength(Src);
484 // strncpy(x, "", y) -> memset(align 1 x, '\0', y)
485 B.CreateMemSet(Dst, B.getInt8('\0'), LenOp, 1);
490 if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(LenOp))
491 Len = LengthArg->getZExtValue();
496 return Dst; // strncpy(x, y, 0) -> x
498 // Let strncpy handle the zero padding
499 if (Len > SrcLen + 1)
502 Type *PT = Callee->getFunctionType()->getParamType(0);
503 // strncpy(x, s, c) -> memcpy(align 1 x, align 1 s, c) [s and c are constant]
504 B.CreateMemCpy(Dst, 1, Src, 1, ConstantInt::get(DL.getIntPtrType(PT), Len));
509 Value *LibCallSimplifier::optimizeStringLength(CallInst *CI, IRBuilder<> &B,
511 Value *Src = CI->getArgOperand(0);
513 // Constant folding: strlen("xyz") -> 3
514 if (uint64_t Len = GetStringLength(Src, CharSize))
515 return ConstantInt::get(CI->getType(), Len - 1);
517 // If s is a constant pointer pointing to a string literal, we can fold
518 // strlen(s + x) to strlen(s) - x, when x is known to be in the range
519 // [0, strlen(s)] or the string has a single null terminator '\0' at the end.
520 // We only try to simplify strlen when the pointer s points to an array
521 // of i8. Otherwise, we would need to scale the offset x before doing the
522 // subtraction. This will make the optimization more complex, and it's not
523 // very useful because calling strlen for a pointer of other types is
525 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Src)) {
526 if (!isGEPBasedOnPointerToString(GEP, CharSize))
529 ConstantDataArraySlice Slice;
530 if (getConstantDataArrayInfo(GEP->getOperand(0), Slice, CharSize)) {
531 uint64_t NullTermIdx;
532 if (Slice.Array == nullptr) {
535 NullTermIdx = ~((uint64_t)0);
536 for (uint64_t I = 0, E = Slice.Length; I < E; ++I) {
537 if (Slice.Array->getElementAsInteger(I + Slice.Offset) == 0) {
542 // If the string does not have '\0', leave it to strlen to compute
544 if (NullTermIdx == ~((uint64_t)0))
548 Value *Offset = GEP->getOperand(2);
549 KnownBits Known = computeKnownBits(Offset, DL, 0, nullptr, CI, nullptr);
550 Known.Zero.flipAllBits();
552 cast<ArrayType>(GEP->getSourceElementType())->getNumElements();
554 // KnownZero's bits are flipped, so zeros in KnownZero now represent
555 // bits known to be zeros in Offset, and ones in KnowZero represent
556 // bits unknown in Offset. Therefore, Offset is known to be in range
557 // [0, NullTermIdx] when the flipped KnownZero is non-negative and
558 // unsigned-less-than NullTermIdx.
560 // If Offset is not provably in the range [0, NullTermIdx], we can still
561 // optimize if we can prove that the program has undefined behavior when
562 // Offset is outside that range. That is the case when GEP->getOperand(0)
563 // is a pointer to an object whose memory extent is NullTermIdx+1.
564 if ((Known.Zero.isNonNegative() && Known.Zero.ule(NullTermIdx)) ||
565 (GEP->isInBounds() && isa<GlobalVariable>(GEP->getOperand(0)) &&
566 NullTermIdx == ArrSize - 1)) {
567 Offset = B.CreateSExtOrTrunc(Offset, CI->getType());
568 return B.CreateSub(ConstantInt::get(CI->getType(), NullTermIdx),
576 // strlen(x?"foo":"bars") --> x ? 3 : 4
577 if (SelectInst *SI = dyn_cast<SelectInst>(Src)) {
578 uint64_t LenTrue = GetStringLength(SI->getTrueValue(), CharSize);
579 uint64_t LenFalse = GetStringLength(SI->getFalseValue(), CharSize);
580 if (LenTrue && LenFalse) {
582 return OptimizationRemark("instcombine", "simplify-libcalls", CI)
583 << "folded strlen(select) to select of constants";
585 return B.CreateSelect(SI->getCondition(),
586 ConstantInt::get(CI->getType(), LenTrue - 1),
587 ConstantInt::get(CI->getType(), LenFalse - 1));
591 // strlen(x) != 0 --> *x != 0
592 // strlen(x) == 0 --> *x == 0
593 if (isOnlyUsedInZeroEqualityComparison(CI))
594 return B.CreateZExt(B.CreateLoad(Src, "strlenfirst"), CI->getType());
599 Value *LibCallSimplifier::optimizeStrLen(CallInst *CI, IRBuilder<> &B) {
600 return optimizeStringLength(CI, B, 8);
603 Value *LibCallSimplifier::optimizeWcslen(CallInst *CI, IRBuilder<> &B) {
604 Module &M = *CI->getModule();
605 unsigned WCharSize = TLI->getWCharSize(M) * 8;
606 // We cannot perform this optimization without wchar_size metadata.
610 return optimizeStringLength(CI, B, WCharSize);
613 Value *LibCallSimplifier::optimizeStrPBrk(CallInst *CI, IRBuilder<> &B) {
615 bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1);
616 bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2);
618 // strpbrk(s, "") -> nullptr
619 // strpbrk("", s) -> nullptr
620 if ((HasS1 && S1.empty()) || (HasS2 && S2.empty()))
621 return Constant::getNullValue(CI->getType());
624 if (HasS1 && HasS2) {
625 size_t I = S1.find_first_of(S2);
626 if (I == StringRef::npos) // No match.
627 return Constant::getNullValue(CI->getType());
629 return B.CreateGEP(B.getInt8Ty(), CI->getArgOperand(0), B.getInt64(I),
633 // strpbrk(s, "a") -> strchr(s, 'a')
634 if (HasS2 && S2.size() == 1)
635 return emitStrChr(CI->getArgOperand(0), S2[0], B, TLI);
640 Value *LibCallSimplifier::optimizeStrTo(CallInst *CI, IRBuilder<> &B) {
641 Value *EndPtr = CI->getArgOperand(1);
642 if (isa<ConstantPointerNull>(EndPtr)) {
643 // With a null EndPtr, this function won't capture the main argument.
644 // It would be readonly too, except that it still may write to errno.
645 CI->addParamAttr(0, Attribute::NoCapture);
651 Value *LibCallSimplifier::optimizeStrSpn(CallInst *CI, IRBuilder<> &B) {
653 bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1);
654 bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2);
656 // strspn(s, "") -> 0
657 // strspn("", s) -> 0
658 if ((HasS1 && S1.empty()) || (HasS2 && S2.empty()))
659 return Constant::getNullValue(CI->getType());
662 if (HasS1 && HasS2) {
663 size_t Pos = S1.find_first_not_of(S2);
664 if (Pos == StringRef::npos)
666 return ConstantInt::get(CI->getType(), Pos);
672 Value *LibCallSimplifier::optimizeStrCSpn(CallInst *CI, IRBuilder<> &B) {
674 bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1);
675 bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2);
677 // strcspn("", s) -> 0
678 if (HasS1 && S1.empty())
679 return Constant::getNullValue(CI->getType());
682 if (HasS1 && HasS2) {
683 size_t Pos = S1.find_first_of(S2);
684 if (Pos == StringRef::npos)
686 return ConstantInt::get(CI->getType(), Pos);
689 // strcspn(s, "") -> strlen(s)
690 if (HasS2 && S2.empty())
691 return emitStrLen(CI->getArgOperand(0), B, DL, TLI);
696 Value *LibCallSimplifier::optimizeStrStr(CallInst *CI, IRBuilder<> &B) {
697 // fold strstr(x, x) -> x.
698 if (CI->getArgOperand(0) == CI->getArgOperand(1))
699 return B.CreateBitCast(CI->getArgOperand(0), CI->getType());
701 // fold strstr(a, b) == a -> strncmp(a, b, strlen(b)) == 0
702 if (isOnlyUsedInEqualityComparison(CI, CI->getArgOperand(0))) {
703 Value *StrLen = emitStrLen(CI->getArgOperand(1), B, DL, TLI);
706 Value *StrNCmp = emitStrNCmp(CI->getArgOperand(0), CI->getArgOperand(1),
710 for (auto UI = CI->user_begin(), UE = CI->user_end(); UI != UE;) {
711 ICmpInst *Old = cast<ICmpInst>(*UI++);
713 B.CreateICmp(Old->getPredicate(), StrNCmp,
714 ConstantInt::getNullValue(StrNCmp->getType()), "cmp");
715 replaceAllUsesWith(Old, Cmp);
720 // See if either input string is a constant string.
721 StringRef SearchStr, ToFindStr;
722 bool HasStr1 = getConstantStringInfo(CI->getArgOperand(0), SearchStr);
723 bool HasStr2 = getConstantStringInfo(CI->getArgOperand(1), ToFindStr);
725 // fold strstr(x, "") -> x.
726 if (HasStr2 && ToFindStr.empty())
727 return B.CreateBitCast(CI->getArgOperand(0), CI->getType());
729 // If both strings are known, constant fold it.
730 if (HasStr1 && HasStr2) {
731 size_t Offset = SearchStr.find(ToFindStr);
733 if (Offset == StringRef::npos) // strstr("foo", "bar") -> null
734 return Constant::getNullValue(CI->getType());
736 // strstr("abcd", "bc") -> gep((char*)"abcd", 1)
737 Value *Result = castToCStr(CI->getArgOperand(0), B);
738 Result = B.CreateConstInBoundsGEP1_64(Result, Offset, "strstr");
739 return B.CreateBitCast(Result, CI->getType());
742 // fold strstr(x, "y") -> strchr(x, 'y').
743 if (HasStr2 && ToFindStr.size() == 1) {
744 Value *StrChr = emitStrChr(CI->getArgOperand(0), ToFindStr[0], B, TLI);
745 return StrChr ? B.CreateBitCast(StrChr, CI->getType()) : nullptr;
750 Value *LibCallSimplifier::optimizeMemChr(CallInst *CI, IRBuilder<> &B) {
751 Value *SrcStr = CI->getArgOperand(0);
752 ConstantInt *CharC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
753 ConstantInt *LenC = dyn_cast<ConstantInt>(CI->getArgOperand(2));
755 // memchr(x, y, 0) -> null
756 if (LenC && LenC->isZero())
757 return Constant::getNullValue(CI->getType());
759 // From now on we need at least constant length and string.
761 if (!LenC || !getConstantStringInfo(SrcStr, Str, 0, /*TrimAtNul=*/false))
764 // Truncate the string to LenC. If Str is smaller than LenC we will still only
765 // scan the string, as reading past the end of it is undefined and we can just
766 // return null if we don't find the char.
767 Str = Str.substr(0, LenC->getZExtValue());
769 // If the char is variable but the input str and length are not we can turn
770 // this memchr call into a simple bit field test. Of course this only works
771 // when the return value is only checked against null.
773 // It would be really nice to reuse switch lowering here but we can't change
774 // the CFG at this point.
776 // memchr("\r\n", C, 2) != nullptr -> (C & ((1 << '\r') | (1 << '\n'))) != 0
777 // after bounds check.
778 if (!CharC && !Str.empty() && isOnlyUsedInZeroEqualityComparison(CI)) {
780 *std::max_element(reinterpret_cast<const unsigned char *>(Str.begin()),
781 reinterpret_cast<const unsigned char *>(Str.end()));
783 // Make sure the bit field we're about to create fits in a register on the
785 // FIXME: On a 64 bit architecture this prevents us from using the
786 // interesting range of alpha ascii chars. We could do better by emitting
787 // two bitfields or shifting the range by 64 if no lower chars are used.
788 if (!DL.fitsInLegalInteger(Max + 1))
791 // For the bit field use a power-of-2 type with at least 8 bits to avoid
792 // creating unnecessary illegal types.
793 unsigned char Width = NextPowerOf2(std::max((unsigned char)7, Max));
795 // Now build the bit field.
796 APInt Bitfield(Width, 0);
798 Bitfield.setBit((unsigned char)C);
799 Value *BitfieldC = B.getInt(Bitfield);
801 // Adjust width of "C" to the bitfield width, then mask off the high bits.
802 Value *C = B.CreateZExtOrTrunc(CI->getArgOperand(1), BitfieldC->getType());
803 C = B.CreateAnd(C, B.getIntN(Width, 0xFF));
805 // First check that the bit field access is within bounds.
806 Value *Bounds = B.CreateICmp(ICmpInst::ICMP_ULT, C, B.getIntN(Width, Width),
809 // Create code that checks if the given bit is set in the field.
810 Value *Shl = B.CreateShl(B.getIntN(Width, 1ULL), C);
811 Value *Bits = B.CreateIsNotNull(B.CreateAnd(Shl, BitfieldC), "memchr.bits");
813 // Finally merge both checks and cast to pointer type. The inttoptr
814 // implicitly zexts the i1 to intptr type.
815 return B.CreateIntToPtr(B.CreateAnd(Bounds, Bits, "memchr"), CI->getType());
818 // Check if all arguments are constants. If so, we can constant fold.
822 // Compute the offset.
823 size_t I = Str.find(CharC->getSExtValue() & 0xFF);
824 if (I == StringRef::npos) // Didn't find the char. memchr returns null.
825 return Constant::getNullValue(CI->getType());
827 // memchr(s+n,c,l) -> gep(s+n+i,c)
828 return B.CreateGEP(B.getInt8Ty(), SrcStr, B.getInt64(I), "memchr");
831 Value *LibCallSimplifier::optimizeMemCmp(CallInst *CI, IRBuilder<> &B) {
832 Value *LHS = CI->getArgOperand(0), *RHS = CI->getArgOperand(1);
834 if (LHS == RHS) // memcmp(s,s,x) -> 0
835 return Constant::getNullValue(CI->getType());
837 // Make sure we have a constant length.
838 ConstantInt *LenC = dyn_cast<ConstantInt>(CI->getArgOperand(2));
842 uint64_t Len = LenC->getZExtValue();
843 if (Len == 0) // memcmp(s1,s2,0) -> 0
844 return Constant::getNullValue(CI->getType());
846 // memcmp(S1,S2,1) -> *(unsigned char*)LHS - *(unsigned char*)RHS
848 Value *LHSV = B.CreateZExt(B.CreateLoad(castToCStr(LHS, B), "lhsc"),
849 CI->getType(), "lhsv");
850 Value *RHSV = B.CreateZExt(B.CreateLoad(castToCStr(RHS, B), "rhsc"),
851 CI->getType(), "rhsv");
852 return B.CreateSub(LHSV, RHSV, "chardiff");
855 // memcmp(S1,S2,N/8)==0 -> (*(intN_t*)S1 != *(intN_t*)S2)==0
856 // TODO: The case where both inputs are constants does not need to be limited
857 // to legal integers or equality comparison. See block below this.
858 if (DL.isLegalInteger(Len * 8) && isOnlyUsedInZeroEqualityComparison(CI)) {
859 IntegerType *IntType = IntegerType::get(CI->getContext(), Len * 8);
860 unsigned PrefAlignment = DL.getPrefTypeAlignment(IntType);
862 // First, see if we can fold either argument to a constant.
863 Value *LHSV = nullptr;
864 if (auto *LHSC = dyn_cast<Constant>(LHS)) {
865 LHSC = ConstantExpr::getBitCast(LHSC, IntType->getPointerTo());
866 LHSV = ConstantFoldLoadFromConstPtr(LHSC, IntType, DL);
868 Value *RHSV = nullptr;
869 if (auto *RHSC = dyn_cast<Constant>(RHS)) {
870 RHSC = ConstantExpr::getBitCast(RHSC, IntType->getPointerTo());
871 RHSV = ConstantFoldLoadFromConstPtr(RHSC, IntType, DL);
874 // Don't generate unaligned loads. If either source is constant data,
875 // alignment doesn't matter for that source because there is no load.
876 if ((LHSV || getKnownAlignment(LHS, DL, CI) >= PrefAlignment) &&
877 (RHSV || getKnownAlignment(RHS, DL, CI) >= PrefAlignment)) {
880 IntType->getPointerTo(LHS->getType()->getPointerAddressSpace());
881 LHSV = B.CreateLoad(B.CreateBitCast(LHS, LHSPtrTy), "lhsv");
885 IntType->getPointerTo(RHS->getType()->getPointerAddressSpace());
886 RHSV = B.CreateLoad(B.CreateBitCast(RHS, RHSPtrTy), "rhsv");
888 return B.CreateZExt(B.CreateICmpNE(LHSV, RHSV), CI->getType(), "memcmp");
892 // Constant folding: memcmp(x, y, Len) -> constant (all arguments are const).
893 // TODO: This is limited to i8 arrays.
894 StringRef LHSStr, RHSStr;
895 if (getConstantStringInfo(LHS, LHSStr) &&
896 getConstantStringInfo(RHS, RHSStr)) {
897 // Make sure we're not reading out-of-bounds memory.
898 if (Len > LHSStr.size() || Len > RHSStr.size())
900 // Fold the memcmp and normalize the result. This way we get consistent
901 // results across multiple platforms.
903 int Cmp = memcmp(LHSStr.data(), RHSStr.data(), Len);
908 return ConstantInt::get(CI->getType(), Ret);
914 Value *LibCallSimplifier::optimizeMemCpy(CallInst *CI, IRBuilder<> &B) {
915 // memcpy(x, y, n) -> llvm.memcpy(align 1 x, align 1 y, n)
916 B.CreateMemCpy(CI->getArgOperand(0), 1, CI->getArgOperand(1), 1,
917 CI->getArgOperand(2));
918 return CI->getArgOperand(0);
921 Value *LibCallSimplifier::optimizeMemMove(CallInst *CI, IRBuilder<> &B) {
922 // memmove(x, y, n) -> llvm.memmove(align 1 x, align 1 y, n)
923 B.CreateMemMove(CI->getArgOperand(0), 1, CI->getArgOperand(1), 1,
924 CI->getArgOperand(2));
925 return CI->getArgOperand(0);
928 /// Fold memset[_chk](malloc(n), 0, n) --> calloc(1, n).
929 Value *LibCallSimplifier::foldMallocMemset(CallInst *Memset, IRBuilder<> &B) {
930 // This has to be a memset of zeros (bzero).
931 auto *FillValue = dyn_cast<ConstantInt>(Memset->getArgOperand(1));
932 if (!FillValue || FillValue->getZExtValue() != 0)
935 // TODO: We should handle the case where the malloc has more than one use.
936 // This is necessary to optimize common patterns such as when the result of
937 // the malloc is checked against null or when a memset intrinsic is used in
938 // place of a memset library call.
939 auto *Malloc = dyn_cast<CallInst>(Memset->getArgOperand(0));
940 if (!Malloc || !Malloc->hasOneUse())
943 // Is the inner call really malloc()?
944 Function *InnerCallee = Malloc->getCalledFunction();
949 if (!TLI->getLibFunc(*InnerCallee, Func) || !TLI->has(Func) ||
950 Func != LibFunc_malloc)
953 // The memset must cover the same number of bytes that are malloc'd.
954 if (Memset->getArgOperand(2) != Malloc->getArgOperand(0))
957 // Replace the malloc with a calloc. We need the data layout to know what the
958 // actual size of a 'size_t' parameter is.
959 B.SetInsertPoint(Malloc->getParent(), ++Malloc->getIterator());
960 const DataLayout &DL = Malloc->getModule()->getDataLayout();
961 IntegerType *SizeType = DL.getIntPtrType(B.GetInsertBlock()->getContext());
962 Value *Calloc = emitCalloc(ConstantInt::get(SizeType, 1),
963 Malloc->getArgOperand(0), Malloc->getAttributes(),
968 Malloc->replaceAllUsesWith(Calloc);
969 eraseFromParent(Malloc);
974 Value *LibCallSimplifier::optimizeMemSet(CallInst *CI, IRBuilder<> &B) {
975 if (auto *Calloc = foldMallocMemset(CI, B))
978 // memset(p, v, n) -> llvm.memset(align 1 p, v, n)
979 Value *Val = B.CreateIntCast(CI->getArgOperand(1), B.getInt8Ty(), false);
980 B.CreateMemSet(CI->getArgOperand(0), Val, CI->getArgOperand(2), 1);
981 return CI->getArgOperand(0);
984 Value *LibCallSimplifier::optimizeRealloc(CallInst *CI, IRBuilder<> &B) {
985 if (isa<ConstantPointerNull>(CI->getArgOperand(0)))
986 return emitMalloc(CI->getArgOperand(1), B, DL, TLI);
991 //===----------------------------------------------------------------------===//
992 // Math Library Optimizations
993 //===----------------------------------------------------------------------===//
995 // Replace a libcall \p CI with a call to intrinsic \p IID
996 static Value *replaceUnaryCall(CallInst *CI, IRBuilder<> &B, Intrinsic::ID IID) {
997 // Propagate fast-math flags from the existing call to the new call.
998 IRBuilder<>::FastMathFlagGuard Guard(B);
999 B.setFastMathFlags(CI->getFastMathFlags());
1001 Module *M = CI->getModule();
1002 Value *V = CI->getArgOperand(0);
1003 Function *F = Intrinsic::getDeclaration(M, IID, CI->getType());
1004 CallInst *NewCall = B.CreateCall(F, V);
1005 NewCall->takeName(CI);
1009 /// Return a variant of Val with float type.
1010 /// Currently this works in two cases: If Val is an FPExtension of a float
1011 /// value to something bigger, simply return the operand.
1012 /// If Val is a ConstantFP but can be converted to a float ConstantFP without
1013 /// loss of precision do so.
1014 static Value *valueHasFloatPrecision(Value *Val) {
1015 if (FPExtInst *Cast = dyn_cast<FPExtInst>(Val)) {
1016 Value *Op = Cast->getOperand(0);
1017 if (Op->getType()->isFloatTy())
1020 if (ConstantFP *Const = dyn_cast<ConstantFP>(Val)) {
1021 APFloat F = Const->getValueAPF();
1023 (void)F.convert(APFloat::IEEEsingle(), APFloat::rmNearestTiesToEven,
1026 return ConstantFP::get(Const->getContext(), F);
1031 /// Shrink double -> float functions.
1032 static Value *optimizeDoubleFP(CallInst *CI, IRBuilder<> &B,
1033 bool isBinary, bool isPrecise = false) {
1034 if (!CI->getType()->isDoubleTy())
1037 // If not all the uses of the function are converted to float, then bail out.
1038 // This matters if the precision of the result is more important than the
1039 // precision of the arguments.
1041 for (User *U : CI->users()) {
1042 FPTruncInst *Cast = dyn_cast<FPTruncInst>(U);
1043 if (!Cast || !Cast->getType()->isFloatTy())
1047 // If this is something like 'g((double) float)', convert to 'gf(float)'.
1049 V[0] = valueHasFloatPrecision(CI->getArgOperand(0));
1050 V[1] = isBinary ? valueHasFloatPrecision(CI->getArgOperand(1)) : nullptr;
1051 if (!V[0] || (isBinary && !V[1]))
1054 // If call isn't an intrinsic, check that it isn't within a function with the
1055 // same name as the float version of this call, otherwise the result is an
1056 // infinite loop. For example, from MinGW-w64:
1058 // float expf(float val) { return (float) exp((double) val); }
1059 Function *CalleeFn = CI->getCalledFunction();
1060 StringRef CalleeNm = CalleeFn->getName();
1061 AttributeList CalleeAt = CalleeFn->getAttributes();
1062 if (CalleeFn && !CalleeFn->isIntrinsic()) {
1063 const Function *Fn = CI->getFunction();
1064 StringRef FnName = Fn->getName();
1065 if (FnName.back() == 'f' &&
1066 FnName.size() == (CalleeNm.size() + 1) &&
1067 FnName.startswith(CalleeNm))
1071 // Propagate the math semantics from the current function to the new function.
1072 IRBuilder<>::FastMathFlagGuard Guard(B);
1073 B.setFastMathFlags(CI->getFastMathFlags());
1075 // g((double) float) -> (double) gf(float)
1077 if (CalleeFn->isIntrinsic()) {
1078 Module *M = CI->getModule();
1079 Intrinsic::ID IID = CalleeFn->getIntrinsicID();
1080 Function *Fn = Intrinsic::getDeclaration(M, IID, B.getFloatTy());
1081 R = isBinary ? B.CreateCall(Fn, V) : B.CreateCall(Fn, V[0]);
1084 R = isBinary ? emitBinaryFloatFnCall(V[0], V[1], CalleeNm, B, CalleeAt)
1085 : emitUnaryFloatFnCall(V[0], CalleeNm, B, CalleeAt);
1087 return B.CreateFPExt(R, B.getDoubleTy());
1090 /// Shrink double -> float for unary functions.
1091 static Value *optimizeUnaryDoubleFP(CallInst *CI, IRBuilder<> &B,
1092 bool isPrecise = false) {
1093 return optimizeDoubleFP(CI, B, false, isPrecise);
1096 /// Shrink double -> float for binary functions.
1097 static Value *optimizeBinaryDoubleFP(CallInst *CI, IRBuilder<> &B,
1098 bool isPrecise = false) {
1099 return optimizeDoubleFP(CI, B, true, isPrecise);
1102 // cabs(z) -> sqrt((creal(z)*creal(z)) + (cimag(z)*cimag(z)))
1103 Value *LibCallSimplifier::optimizeCAbs(CallInst *CI, IRBuilder<> &B) {
1107 // Propagate fast-math flags from the existing call to new instructions.
1108 IRBuilder<>::FastMathFlagGuard Guard(B);
1109 B.setFastMathFlags(CI->getFastMathFlags());
1112 if (CI->getNumArgOperands() == 1) {
1113 Value *Op = CI->getArgOperand(0);
1114 assert(Op->getType()->isArrayTy() && "Unexpected signature for cabs!");
1115 Real = B.CreateExtractValue(Op, 0, "real");
1116 Imag = B.CreateExtractValue(Op, 1, "imag");
1118 assert(CI->getNumArgOperands() == 2 && "Unexpected signature for cabs!");
1119 Real = CI->getArgOperand(0);
1120 Imag = CI->getArgOperand(1);
1123 Value *RealReal = B.CreateFMul(Real, Real);
1124 Value *ImagImag = B.CreateFMul(Imag, Imag);
1126 Function *FSqrt = Intrinsic::getDeclaration(CI->getModule(), Intrinsic::sqrt,
1128 return B.CreateCall(FSqrt, B.CreateFAdd(RealReal, ImagImag), "cabs");
1131 static Value *optimizeTrigReflections(CallInst *Call, LibFunc Func,
1133 if (!isa<FPMathOperator>(Call))
1136 IRBuilder<>::FastMathFlagGuard Guard(B);
1137 B.setFastMathFlags(Call->getFastMathFlags());
1139 // TODO: Can this be shared to also handle LLVM intrinsics?
1148 // sin(-X) --> -sin(X)
1149 // tan(-X) --> -tan(X)
1150 if (match(Call->getArgOperand(0), m_OneUse(m_FNeg(m_Value(X)))))
1151 return B.CreateFNeg(B.CreateCall(Call->getCalledFunction(), X));
1156 // cos(-X) --> cos(X)
1157 if (match(Call->getArgOperand(0), m_FNeg(m_Value(X))))
1158 return B.CreateCall(Call->getCalledFunction(), X, "cos");
1166 static Value *getPow(Value *InnerChain[33], unsigned Exp, IRBuilder<> &B) {
1167 // Multiplications calculated using Addition Chains.
1168 // Refer: http://wwwhomes.uni-bielefeld.de/achim/addition_chain.html
1170 assert(Exp != 0 && "Incorrect exponent 0 not handled");
1172 if (InnerChain[Exp])
1173 return InnerChain[Exp];
1175 static const unsigned AddChain[33][2] = {
1177 {0, 0}, // Unused (base case = pow1).
1178 {1, 1}, // Unused (pre-computed).
1179 {1, 2}, {2, 2}, {2, 3}, {3, 3}, {2, 5}, {4, 4},
1180 {1, 8}, {5, 5}, {1, 10}, {6, 6}, {4, 9}, {7, 7},
1181 {3, 12}, {8, 8}, {8, 9}, {2, 16}, {1, 18}, {10, 10},
1182 {6, 15}, {11, 11}, {3, 20}, {12, 12}, {8, 17}, {13, 13},
1183 {3, 24}, {14, 14}, {4, 25}, {15, 15}, {3, 28}, {16, 16},
1186 InnerChain[Exp] = B.CreateFMul(getPow(InnerChain, AddChain[Exp][0], B),
1187 getPow(InnerChain, AddChain[Exp][1], B));
1188 return InnerChain[Exp];
1191 /// Use exp{,2}(x * y) for pow(exp{,2}(x), y);
1192 /// exp2(n * x) for pow(2.0 ** n, x); exp10(x) for pow(10.0, x).
1193 Value *LibCallSimplifier::replacePowWithExp(CallInst *Pow, IRBuilder<> &B) {
1194 Value *Base = Pow->getArgOperand(0), *Expo = Pow->getArgOperand(1);
1195 AttributeList Attrs = Pow->getCalledFunction()->getAttributes();
1196 Module *Mod = Pow->getModule();
1197 Type *Ty = Pow->getType();
1200 // Evaluate special cases related to a nested function as the base.
1202 // pow(exp(x), y) -> exp(x * y)
1203 // pow(exp2(x), y) -> exp2(x * y)
1204 // If exp{,2}() is used only once, it is better to fold two transcendental
1205 // math functions into one. If used again, exp{,2}() would still have to be
1206 // called with the original argument, then keep both original transcendental
1207 // functions. However, this transformation is only safe with fully relaxed
1208 // math semantics, since, besides rounding differences, it changes overflow
1209 // and underflow behavior quite dramatically. For example:
1210 // pow(exp(1000), 0.001) = pow(inf, 0.001) = inf
1212 // exp(1000 * 0.001) = exp(1)
1213 // TODO: Loosen the requirement for fully relaxed math semantics.
1214 // TODO: Handle exp10() when more targets have it available.
1215 CallInst *BaseFn = dyn_cast<CallInst>(Base);
1216 if (BaseFn && BaseFn->hasOneUse() && BaseFn->isFast() && Pow->isFast()) {
1219 Function *CalleeFn = BaseFn->getCalledFunction();
1221 TLI->getLibFunc(CalleeFn->getName(), LibFn) && TLI->has(LibFn)) {
1226 LibFunc LibFnDouble;
1227 LibFunc LibFnLongDouble;
1232 case LibFunc_expf: case LibFunc_exp: case LibFunc_expl:
1233 ExpName = TLI->getName(LibFunc_exp);
1234 ID = Intrinsic::exp;
1235 LibFnFloat = LibFunc_expf;
1236 LibFnDouble = LibFunc_exp;
1237 LibFnLongDouble = LibFunc_expl;
1239 case LibFunc_exp2f: case LibFunc_exp2: case LibFunc_exp2l:
1240 ExpName = TLI->getName(LibFunc_exp2);
1241 ID = Intrinsic::exp2;
1242 LibFnFloat = LibFunc_exp2f;
1243 LibFnDouble = LibFunc_exp2;
1244 LibFnLongDouble = LibFunc_exp2l;
1248 // Create new exp{,2}() with the product as its argument.
1249 Value *FMul = B.CreateFMul(BaseFn->getArgOperand(0), Expo, "mul");
1250 ExpFn = BaseFn->doesNotAccessMemory()
1251 ? B.CreateCall(Intrinsic::getDeclaration(Mod, ID, Ty),
1253 : emitUnaryFloatFnCall(FMul, TLI, LibFnDouble, LibFnFloat,
1255 BaseFn->getAttributes());
1257 // Since the new exp{,2}() is different from the original one, dead code
1258 // elimination cannot be trusted to remove it, since it may have side
1259 // effects (e.g., errno). When the only consumer for the original
1260 // exp{,2}() is pow(), then it has to be explicitly erased.
1261 BaseFn->replaceAllUsesWith(ExpFn);
1262 eraseFromParent(BaseFn);
1268 // Evaluate special cases related to a constant base.
1270 const APFloat *BaseF;
1271 if (!match(Pow->getArgOperand(0), m_APFloat(BaseF)))
1274 // pow(2.0 ** n, x) -> exp2(n * x)
1275 if (hasUnaryFloatFn(TLI, Ty, LibFunc_exp2, LibFunc_exp2f, LibFunc_exp2l)) {
1276 APFloat BaseR = APFloat(1.0);
1277 BaseR.convert(BaseF->getSemantics(), APFloat::rmTowardZero, &Ignored);
1278 BaseR = BaseR / *BaseF;
1279 bool IsInteger = BaseF->isInteger(),
1280 IsReciprocal = BaseR.isInteger();
1281 const APFloat *NF = IsReciprocal ? &BaseR : BaseF;
1282 APSInt NI(64, false);
1283 if ((IsInteger || IsReciprocal) &&
1284 !NF->convertToInteger(NI, APFloat::rmTowardZero, &Ignored) &&
1285 NI > 1 && NI.isPowerOf2()) {
1286 double N = NI.logBase2() * (IsReciprocal ? -1.0 : 1.0);
1287 Value *FMul = B.CreateFMul(Expo, ConstantFP::get(Ty, N), "mul");
1288 if (Pow->doesNotAccessMemory())
1289 return B.CreateCall(Intrinsic::getDeclaration(Mod, Intrinsic::exp2, Ty),
1292 return emitUnaryFloatFnCall(FMul, TLI, LibFunc_exp2, LibFunc_exp2f,
1293 LibFunc_exp2l, B, Attrs);
1297 // pow(10.0, x) -> exp10(x)
1298 // TODO: There is no exp10() intrinsic yet, but some day there shall be one.
1299 if (match(Base, m_SpecificFP(10.0)) &&
1300 hasUnaryFloatFn(TLI, Ty, LibFunc_exp10, LibFunc_exp10f, LibFunc_exp10l))
1301 return emitUnaryFloatFnCall(Expo, TLI, LibFunc_exp10, LibFunc_exp10f,
1302 LibFunc_exp10l, B, Attrs);
1307 static Value *getSqrtCall(Value *V, AttributeList Attrs, bool NoErrno,
1308 Module *M, IRBuilder<> &B,
1309 const TargetLibraryInfo *TLI) {
1310 // If errno is never set, then use the intrinsic for sqrt().
1313 Intrinsic::getDeclaration(M, Intrinsic::sqrt, V->getType());
1314 return B.CreateCall(SqrtFn, V, "sqrt");
1317 // Otherwise, use the libcall for sqrt().
1318 if (hasUnaryFloatFn(TLI, V->getType(), LibFunc_sqrt, LibFunc_sqrtf,
1320 // TODO: We also should check that the target can in fact lower the sqrt()
1321 // libcall. We currently have no way to ask this question, so we ask if
1322 // the target has a sqrt() libcall, which is not exactly the same.
1323 return emitUnaryFloatFnCall(V, TLI, LibFunc_sqrt, LibFunc_sqrtf,
1324 LibFunc_sqrtl, B, Attrs);
1329 /// Use square root in place of pow(x, +/-0.5).
1330 Value *LibCallSimplifier::replacePowWithSqrt(CallInst *Pow, IRBuilder<> &B) {
1331 Value *Sqrt, *Base = Pow->getArgOperand(0), *Expo = Pow->getArgOperand(1);
1332 AttributeList Attrs = Pow->getCalledFunction()->getAttributes();
1333 Module *Mod = Pow->getModule();
1334 Type *Ty = Pow->getType();
1336 const APFloat *ExpoF;
1337 if (!match(Expo, m_APFloat(ExpoF)) ||
1338 (!ExpoF->isExactlyValue(0.5) && !ExpoF->isExactlyValue(-0.5)))
1341 Sqrt = getSqrtCall(Base, Attrs, Pow->doesNotAccessMemory(), Mod, B, TLI);
1345 // Handle signed zero base by expanding to fabs(sqrt(x)).
1346 if (!Pow->hasNoSignedZeros()) {
1347 Function *FAbsFn = Intrinsic::getDeclaration(Mod, Intrinsic::fabs, Ty);
1348 Sqrt = B.CreateCall(FAbsFn, Sqrt, "abs");
1351 // Handle non finite base by expanding to
1352 // (x == -infinity ? +infinity : sqrt(x)).
1353 if (!Pow->hasNoInfs()) {
1354 Value *PosInf = ConstantFP::getInfinity(Ty),
1355 *NegInf = ConstantFP::getInfinity(Ty, true);
1356 Value *FCmp = B.CreateFCmpOEQ(Base, NegInf, "isinf");
1357 Sqrt = B.CreateSelect(FCmp, PosInf, Sqrt);
1360 // If the exponent is negative, then get the reciprocal.
1361 if (ExpoF->isNegative())
1362 Sqrt = B.CreateFDiv(ConstantFP::get(Ty, 1.0), Sqrt, "reciprocal");
1367 Value *LibCallSimplifier::optimizePow(CallInst *Pow, IRBuilder<> &B) {
1368 Value *Base = Pow->getArgOperand(0), *Expo = Pow->getArgOperand(1);
1369 Function *Callee = Pow->getCalledFunction();
1370 StringRef Name = Callee->getName();
1371 Type *Ty = Pow->getType();
1372 Value *Shrunk = nullptr;
1375 // Bail out if simplifying libcalls to pow() is disabled.
1376 if (!hasUnaryFloatFn(TLI, Ty, LibFunc_pow, LibFunc_powf, LibFunc_powl))
1379 // Propagate the math semantics from the call to any created instructions.
1380 IRBuilder<>::FastMathFlagGuard Guard(B);
1381 B.setFastMathFlags(Pow->getFastMathFlags());
1383 // Shrink pow() to powf() if the arguments are single precision,
1384 // unless the result is expected to be double precision.
1385 if (UnsafeFPShrink &&
1386 Name == TLI->getName(LibFunc_pow) && hasFloatVersion(Name))
1387 Shrunk = optimizeBinaryDoubleFP(Pow, B, true);
1389 // Evaluate special cases related to the base.
1391 // pow(1.0, x) -> 1.0
1392 if (match(Base, m_FPOne()))
1395 if (Value *Exp = replacePowWithExp(Pow, B))
1398 // Evaluate special cases related to the exponent.
1400 // pow(x, -1.0) -> 1.0 / x
1401 if (match(Expo, m_SpecificFP(-1.0)))
1402 return B.CreateFDiv(ConstantFP::get(Ty, 1.0), Base, "reciprocal");
1404 // pow(x, 0.0) -> 1.0
1405 if (match(Expo, m_SpecificFP(0.0)))
1406 return ConstantFP::get(Ty, 1.0);
1409 if (match(Expo, m_FPOne()))
1412 // pow(x, 2.0) -> x * x
1413 if (match(Expo, m_SpecificFP(2.0)))
1414 return B.CreateFMul(Base, Base, "square");
1416 if (Value *Sqrt = replacePowWithSqrt(Pow, B))
1419 // pow(x, n) -> x * x * x * ...
1420 const APFloat *ExpoF;
1421 if (Pow->isFast() && match(Expo, m_APFloat(ExpoF))) {
1422 // We limit to a max of 7 multiplications, thus the maximum exponent is 32.
1423 // If the exponent is an integer+0.5 we generate a call to sqrt and an
1425 // TODO: This whole transformation should be backend specific (e.g. some
1426 // backends might prefer libcalls or the limit for the exponent might
1427 // be different) and it should also consider optimizing for size.
1428 APFloat LimF(ExpoF->getSemantics(), 33.0),
1430 if (ExpoA.compare(LimF) == APFloat::cmpLessThan) {
1431 // This transformation applies to integer or integer+0.5 exponents only.
1432 // For integer+0.5, we create a sqrt(Base) call.
1433 Value *Sqrt = nullptr;
1434 if (!ExpoA.isInteger()) {
1435 APFloat Expo2 = ExpoA;
1436 // To check if ExpoA is an integer + 0.5, we add it to itself. If there
1437 // is no floating point exception and the result is an integer, then
1438 // ExpoA == integer + 0.5
1439 if (Expo2.add(ExpoA, APFloat::rmNearestTiesToEven) != APFloat::opOK)
1442 if (!Expo2.isInteger())
1446 getSqrtCall(Base, Pow->getCalledFunction()->getAttributes(),
1447 Pow->doesNotAccessMemory(), Pow->getModule(), B, TLI);
1450 // We will memoize intermediate products of the Addition Chain.
1451 Value *InnerChain[33] = {nullptr};
1452 InnerChain[1] = Base;
1453 InnerChain[2] = B.CreateFMul(Base, Base, "square");
1455 // We cannot readily convert a non-double type (like float) to a double.
1456 // So we first convert it to something which could be converted to double.
1457 ExpoA.convert(APFloat::IEEEdouble(), APFloat::rmTowardZero, &Ignored);
1458 Value *FMul = getPow(InnerChain, ExpoA.convertToDouble(), B);
1460 // Expand pow(x, y+0.5) to pow(x, y) * sqrt(x).
1462 FMul = B.CreateFMul(FMul, Sqrt);
1464 // If the exponent is negative, then get the reciprocal.
1465 if (ExpoF->isNegative())
1466 FMul = B.CreateFDiv(ConstantFP::get(Ty, 1.0), FMul, "reciprocal");
1475 Value *LibCallSimplifier::optimizeExp2(CallInst *CI, IRBuilder<> &B) {
1476 Function *Callee = CI->getCalledFunction();
1477 Value *Ret = nullptr;
1478 StringRef Name = Callee->getName();
1479 if (UnsafeFPShrink && Name == "exp2" && hasFloatVersion(Name))
1480 Ret = optimizeUnaryDoubleFP(CI, B, true);
1482 Value *Op = CI->getArgOperand(0);
1483 // Turn exp2(sitofp(x)) -> ldexp(1.0, sext(x)) if sizeof(x) <= 32
1484 // Turn exp2(uitofp(x)) -> ldexp(1.0, zext(x)) if sizeof(x) < 32
1485 LibFunc LdExp = LibFunc_ldexpl;
1486 if (Op->getType()->isFloatTy())
1487 LdExp = LibFunc_ldexpf;
1488 else if (Op->getType()->isDoubleTy())
1489 LdExp = LibFunc_ldexp;
1491 if (TLI->has(LdExp)) {
1492 Value *LdExpArg = nullptr;
1493 if (SIToFPInst *OpC = dyn_cast<SIToFPInst>(Op)) {
1494 if (OpC->getOperand(0)->getType()->getPrimitiveSizeInBits() <= 32)
1495 LdExpArg = B.CreateSExt(OpC->getOperand(0), B.getInt32Ty());
1496 } else if (UIToFPInst *OpC = dyn_cast<UIToFPInst>(Op)) {
1497 if (OpC->getOperand(0)->getType()->getPrimitiveSizeInBits() < 32)
1498 LdExpArg = B.CreateZExt(OpC->getOperand(0), B.getInt32Ty());
1502 Constant *One = ConstantFP::get(CI->getContext(), APFloat(1.0f));
1503 if (!Op->getType()->isFloatTy())
1504 One = ConstantExpr::getFPExtend(One, Op->getType());
1506 Module *M = CI->getModule();
1508 M->getOrInsertFunction(TLI->getName(LdExp), Op->getType(),
1509 Op->getType(), B.getInt32Ty());
1510 CallInst *CI = B.CreateCall(NewCallee, {One, LdExpArg});
1511 if (const Function *F = dyn_cast<Function>(Callee->stripPointerCasts()))
1512 CI->setCallingConv(F->getCallingConv());
1520 Value *LibCallSimplifier::optimizeFMinFMax(CallInst *CI, IRBuilder<> &B) {
1521 Function *Callee = CI->getCalledFunction();
1522 // If we can shrink the call to a float function rather than a double
1523 // function, do that first.
1524 StringRef Name = Callee->getName();
1525 if ((Name == "fmin" || Name == "fmax") && hasFloatVersion(Name))
1526 if (Value *Ret = optimizeBinaryDoubleFP(CI, B))
1529 IRBuilder<>::FastMathFlagGuard Guard(B);
1532 // If the call is 'fast', then anything we create here will also be 'fast'.
1535 // At a minimum, no-nans-fp-math must be true.
1536 if (!CI->hasNoNaNs())
1538 // No-signed-zeros is implied by the definitions of fmax/fmin themselves:
1539 // "Ideally, fmax would be sensitive to the sign of zero, for example
1540 // fmax(-0. 0, +0. 0) would return +0; however, implementation in software
1541 // might be impractical."
1542 FMF.setNoSignedZeros();
1545 B.setFastMathFlags(FMF);
1547 // We have a relaxed floating-point environment. We can ignore NaN-handling
1548 // and transform to a compare and select. We do not have to consider errno or
1549 // exceptions, because fmin/fmax do not have those.
1550 Value *Op0 = CI->getArgOperand(0);
1551 Value *Op1 = CI->getArgOperand(1);
1552 Value *Cmp = Callee->getName().startswith("fmin") ?
1553 B.CreateFCmpOLT(Op0, Op1) : B.CreateFCmpOGT(Op0, Op1);
1554 return B.CreateSelect(Cmp, Op0, Op1);
1557 Value *LibCallSimplifier::optimizeLog(CallInst *CI, IRBuilder<> &B) {
1558 Function *Callee = CI->getCalledFunction();
1559 Value *Ret = nullptr;
1560 StringRef Name = Callee->getName();
1561 if (UnsafeFPShrink && hasFloatVersion(Name))
1562 Ret = optimizeUnaryDoubleFP(CI, B, true);
1566 Value *Op1 = CI->getArgOperand(0);
1567 auto *OpC = dyn_cast<CallInst>(Op1);
1569 // The earlier call must also be 'fast' in order to do these transforms.
1570 if (!OpC || !OpC->isFast())
1573 // log(pow(x,y)) -> y*log(x)
1574 // This is only applicable to log, log2, log10.
1575 if (Name != "log" && Name != "log2" && Name != "log10")
1578 IRBuilder<>::FastMathFlagGuard Guard(B);
1581 B.setFastMathFlags(FMF);
1584 Function *F = OpC->getCalledFunction();
1585 if (F && ((TLI->getLibFunc(F->getName(), Func) && TLI->has(Func) &&
1586 Func == LibFunc_pow) || F->getIntrinsicID() == Intrinsic::pow))
1587 return B.CreateFMul(OpC->getArgOperand(1),
1588 emitUnaryFloatFnCall(OpC->getOperand(0), Callee->getName(), B,
1589 Callee->getAttributes()), "mul");
1591 // log(exp2(y)) -> y*log(2)
1592 if (F && Name == "log" && TLI->getLibFunc(F->getName(), Func) &&
1593 TLI->has(Func) && Func == LibFunc_exp2)
1594 return B.CreateFMul(
1595 OpC->getArgOperand(0),
1596 emitUnaryFloatFnCall(ConstantFP::get(CI->getType(), 2.0),
1597 Callee->getName(), B, Callee->getAttributes()),
1602 Value *LibCallSimplifier::optimizeSqrt(CallInst *CI, IRBuilder<> &B) {
1603 Function *Callee = CI->getCalledFunction();
1604 Value *Ret = nullptr;
1605 // TODO: Once we have a way (other than checking for the existince of the
1606 // libcall) to tell whether our target can lower @llvm.sqrt, relax the
1608 if (TLI->has(LibFunc_sqrtf) && (Callee->getName() == "sqrt" ||
1609 Callee->getIntrinsicID() == Intrinsic::sqrt))
1610 Ret = optimizeUnaryDoubleFP(CI, B, true);
1615 Instruction *I = dyn_cast<Instruction>(CI->getArgOperand(0));
1616 if (!I || I->getOpcode() != Instruction::FMul || !I->isFast())
1619 // We're looking for a repeated factor in a multiplication tree,
1620 // so we can do this fold: sqrt(x * x) -> fabs(x);
1621 // or this fold: sqrt((x * x) * y) -> fabs(x) * sqrt(y).
1622 Value *Op0 = I->getOperand(0);
1623 Value *Op1 = I->getOperand(1);
1624 Value *RepeatOp = nullptr;
1625 Value *OtherOp = nullptr;
1627 // Simple match: the operands of the multiply are identical.
1630 // Look for a more complicated pattern: one of the operands is itself
1631 // a multiply, so search for a common factor in that multiply.
1632 // Note: We don't bother looking any deeper than this first level or for
1633 // variations of this pattern because instcombine's visitFMUL and/or the
1634 // reassociation pass should give us this form.
1635 Value *OtherMul0, *OtherMul1;
1636 if (match(Op0, m_FMul(m_Value(OtherMul0), m_Value(OtherMul1)))) {
1637 // Pattern: sqrt((x * y) * z)
1638 if (OtherMul0 == OtherMul1 && cast<Instruction>(Op0)->isFast()) {
1639 // Matched: sqrt((x * x) * z)
1640 RepeatOp = OtherMul0;
1648 // Fast math flags for any created instructions should match the sqrt
1650 IRBuilder<>::FastMathFlagGuard Guard(B);
1651 B.setFastMathFlags(I->getFastMathFlags());
1653 // If we found a repeated factor, hoist it out of the square root and
1654 // replace it with the fabs of that factor.
1655 Module *M = Callee->getParent();
1656 Type *ArgType = I->getType();
1657 Value *Fabs = Intrinsic::getDeclaration(M, Intrinsic::fabs, ArgType);
1658 Value *FabsCall = B.CreateCall(Fabs, RepeatOp, "fabs");
1660 // If we found a non-repeated factor, we still need to get its square
1661 // root. We then multiply that by the value that was simplified out
1662 // of the square root calculation.
1663 Value *Sqrt = Intrinsic::getDeclaration(M, Intrinsic::sqrt, ArgType);
1664 Value *SqrtCall = B.CreateCall(Sqrt, OtherOp, "sqrt");
1665 return B.CreateFMul(FabsCall, SqrtCall);
1670 // TODO: Generalize to handle any trig function and its inverse.
1671 Value *LibCallSimplifier::optimizeTan(CallInst *CI, IRBuilder<> &B) {
1672 Function *Callee = CI->getCalledFunction();
1673 Value *Ret = nullptr;
1674 StringRef Name = Callee->getName();
1675 if (UnsafeFPShrink && Name == "tan" && hasFloatVersion(Name))
1676 Ret = optimizeUnaryDoubleFP(CI, B, true);
1678 Value *Op1 = CI->getArgOperand(0);
1679 auto *OpC = dyn_cast<CallInst>(Op1);
1683 // Both calls must be 'fast' in order to remove them.
1684 if (!CI->isFast() || !OpC->isFast())
1687 // tan(atan(x)) -> x
1688 // tanf(atanf(x)) -> x
1689 // tanl(atanl(x)) -> x
1691 Function *F = OpC->getCalledFunction();
1692 if (F && TLI->getLibFunc(F->getName(), Func) && TLI->has(Func) &&
1693 ((Func == LibFunc_atan && Callee->getName() == "tan") ||
1694 (Func == LibFunc_atanf && Callee->getName() == "tanf") ||
1695 (Func == LibFunc_atanl && Callee->getName() == "tanl")))
1696 Ret = OpC->getArgOperand(0);
1700 static bool isTrigLibCall(CallInst *CI) {
1701 // We can only hope to do anything useful if we can ignore things like errno
1702 // and floating-point exceptions.
1703 // We already checked the prototype.
1704 return CI->hasFnAttr(Attribute::NoUnwind) &&
1705 CI->hasFnAttr(Attribute::ReadNone);
1708 static void insertSinCosCall(IRBuilder<> &B, Function *OrigCallee, Value *Arg,
1709 bool UseFloat, Value *&Sin, Value *&Cos,
1711 Type *ArgTy = Arg->getType();
1715 Triple T(OrigCallee->getParent()->getTargetTriple());
1717 Name = "__sincospif_stret";
1719 assert(T.getArch() != Triple::x86 && "x86 messy and unsupported for now");
1720 // x86_64 can't use {float, float} since that would be returned in both
1721 // xmm0 and xmm1, which isn't what a real struct would do.
1722 ResTy = T.getArch() == Triple::x86_64
1723 ? static_cast<Type *>(VectorType::get(ArgTy, 2))
1724 : static_cast<Type *>(StructType::get(ArgTy, ArgTy));
1726 Name = "__sincospi_stret";
1727 ResTy = StructType::get(ArgTy, ArgTy);
1730 Module *M = OrigCallee->getParent();
1731 Value *Callee = M->getOrInsertFunction(Name, OrigCallee->getAttributes(),
1734 if (Instruction *ArgInst = dyn_cast<Instruction>(Arg)) {
1735 // If the argument is an instruction, it must dominate all uses so put our
1736 // sincos call there.
1737 B.SetInsertPoint(ArgInst->getParent(), ++ArgInst->getIterator());
1739 // Otherwise (e.g. for a constant) the beginning of the function is as
1740 // good a place as any.
1741 BasicBlock &EntryBB = B.GetInsertBlock()->getParent()->getEntryBlock();
1742 B.SetInsertPoint(&EntryBB, EntryBB.begin());
1745 SinCos = B.CreateCall(Callee, Arg, "sincospi");
1747 if (SinCos->getType()->isStructTy()) {
1748 Sin = B.CreateExtractValue(SinCos, 0, "sinpi");
1749 Cos = B.CreateExtractValue(SinCos, 1, "cospi");
1751 Sin = B.CreateExtractElement(SinCos, ConstantInt::get(B.getInt32Ty(), 0),
1753 Cos = B.CreateExtractElement(SinCos, ConstantInt::get(B.getInt32Ty(), 1),
1758 Value *LibCallSimplifier::optimizeSinCosPi(CallInst *CI, IRBuilder<> &B) {
1759 // Make sure the prototype is as expected, otherwise the rest of the
1760 // function is probably invalid and likely to abort.
1761 if (!isTrigLibCall(CI))
1764 Value *Arg = CI->getArgOperand(0);
1765 SmallVector<CallInst *, 1> SinCalls;
1766 SmallVector<CallInst *, 1> CosCalls;
1767 SmallVector<CallInst *, 1> SinCosCalls;
1769 bool IsFloat = Arg->getType()->isFloatTy();
1771 // Look for all compatible sinpi, cospi and sincospi calls with the same
1772 // argument. If there are enough (in some sense) we can make the
1774 Function *F = CI->getFunction();
1775 for (User *U : Arg->users())
1776 classifyArgUse(U, F, IsFloat, SinCalls, CosCalls, SinCosCalls);
1778 // It's only worthwhile if both sinpi and cospi are actually used.
1779 if (SinCosCalls.empty() && (SinCalls.empty() || CosCalls.empty()))
1782 Value *Sin, *Cos, *SinCos;
1783 insertSinCosCall(B, CI->getCalledFunction(), Arg, IsFloat, Sin, Cos, SinCos);
1785 auto replaceTrigInsts = [this](SmallVectorImpl<CallInst *> &Calls,
1787 for (CallInst *C : Calls)
1788 replaceAllUsesWith(C, Res);
1791 replaceTrigInsts(SinCalls, Sin);
1792 replaceTrigInsts(CosCalls, Cos);
1793 replaceTrigInsts(SinCosCalls, SinCos);
1798 void LibCallSimplifier::classifyArgUse(
1799 Value *Val, Function *F, bool IsFloat,
1800 SmallVectorImpl<CallInst *> &SinCalls,
1801 SmallVectorImpl<CallInst *> &CosCalls,
1802 SmallVectorImpl<CallInst *> &SinCosCalls) {
1803 CallInst *CI = dyn_cast<CallInst>(Val);
1808 // Don't consider calls in other functions.
1809 if (CI->getFunction() != F)
1812 Function *Callee = CI->getCalledFunction();
1814 if (!Callee || !TLI->getLibFunc(*Callee, Func) || !TLI->has(Func) ||
1819 if (Func == LibFunc_sinpif)
1820 SinCalls.push_back(CI);
1821 else if (Func == LibFunc_cospif)
1822 CosCalls.push_back(CI);
1823 else if (Func == LibFunc_sincospif_stret)
1824 SinCosCalls.push_back(CI);
1826 if (Func == LibFunc_sinpi)
1827 SinCalls.push_back(CI);
1828 else if (Func == LibFunc_cospi)
1829 CosCalls.push_back(CI);
1830 else if (Func == LibFunc_sincospi_stret)
1831 SinCosCalls.push_back(CI);
1835 //===----------------------------------------------------------------------===//
1836 // Integer Library Call Optimizations
1837 //===----------------------------------------------------------------------===//
1839 Value *LibCallSimplifier::optimizeFFS(CallInst *CI, IRBuilder<> &B) {
1840 // ffs(x) -> x != 0 ? (i32)llvm.cttz(x)+1 : 0
1841 Value *Op = CI->getArgOperand(0);
1842 Type *ArgType = Op->getType();
1843 Value *F = Intrinsic::getDeclaration(CI->getCalledFunction()->getParent(),
1844 Intrinsic::cttz, ArgType);
1845 Value *V = B.CreateCall(F, {Op, B.getTrue()}, "cttz");
1846 V = B.CreateAdd(V, ConstantInt::get(V->getType(), 1));
1847 V = B.CreateIntCast(V, B.getInt32Ty(), false);
1849 Value *Cond = B.CreateICmpNE(Op, Constant::getNullValue(ArgType));
1850 return B.CreateSelect(Cond, V, B.getInt32(0));
1853 Value *LibCallSimplifier::optimizeFls(CallInst *CI, IRBuilder<> &B) {
1854 // fls(x) -> (i32)(sizeInBits(x) - llvm.ctlz(x, false))
1855 Value *Op = CI->getArgOperand(0);
1856 Type *ArgType = Op->getType();
1857 Value *F = Intrinsic::getDeclaration(CI->getCalledFunction()->getParent(),
1858 Intrinsic::ctlz, ArgType);
1859 Value *V = B.CreateCall(F, {Op, B.getFalse()}, "ctlz");
1860 V = B.CreateSub(ConstantInt::get(V->getType(), ArgType->getIntegerBitWidth()),
1862 return B.CreateIntCast(V, CI->getType(), false);
1865 Value *LibCallSimplifier::optimizeAbs(CallInst *CI, IRBuilder<> &B) {
1866 // abs(x) -> x <s 0 ? -x : x
1867 // The negation has 'nsw' because abs of INT_MIN is undefined.
1868 Value *X = CI->getArgOperand(0);
1869 Value *IsNeg = B.CreateICmpSLT(X, Constant::getNullValue(X->getType()));
1870 Value *NegX = B.CreateNSWNeg(X, "neg");
1871 return B.CreateSelect(IsNeg, NegX, X);
1874 Value *LibCallSimplifier::optimizeIsDigit(CallInst *CI, IRBuilder<> &B) {
1875 // isdigit(c) -> (c-'0') <u 10
1876 Value *Op = CI->getArgOperand(0);
1877 Op = B.CreateSub(Op, B.getInt32('0'), "isdigittmp");
1878 Op = B.CreateICmpULT(Op, B.getInt32(10), "isdigit");
1879 return B.CreateZExt(Op, CI->getType());
1882 Value *LibCallSimplifier::optimizeIsAscii(CallInst *CI, IRBuilder<> &B) {
1883 // isascii(c) -> c <u 128
1884 Value *Op = CI->getArgOperand(0);
1885 Op = B.CreateICmpULT(Op, B.getInt32(128), "isascii");
1886 return B.CreateZExt(Op, CI->getType());
1889 Value *LibCallSimplifier::optimizeToAscii(CallInst *CI, IRBuilder<> &B) {
1890 // toascii(c) -> c & 0x7f
1891 return B.CreateAnd(CI->getArgOperand(0),
1892 ConstantInt::get(CI->getType(), 0x7F));
1895 Value *LibCallSimplifier::optimizeAtoi(CallInst *CI, IRBuilder<> &B) {
1897 if (!getConstantStringInfo(CI->getArgOperand(0), Str))
1900 return convertStrToNumber(CI, Str, 10);
1903 Value *LibCallSimplifier::optimizeStrtol(CallInst *CI, IRBuilder<> &B) {
1905 if (!getConstantStringInfo(CI->getArgOperand(0), Str))
1908 if (!isa<ConstantPointerNull>(CI->getArgOperand(1)))
1911 if (ConstantInt *CInt = dyn_cast<ConstantInt>(CI->getArgOperand(2))) {
1912 return convertStrToNumber(CI, Str, CInt->getSExtValue());
1918 //===----------------------------------------------------------------------===//
1919 // Formatting and IO Library Call Optimizations
1920 //===----------------------------------------------------------------------===//
1922 static bool isReportingError(Function *Callee, CallInst *CI, int StreamArg);
1924 Value *LibCallSimplifier::optimizeErrorReporting(CallInst *CI, IRBuilder<> &B,
1926 Function *Callee = CI->getCalledFunction();
1927 // Error reporting calls should be cold, mark them as such.
1928 // This applies even to non-builtin calls: it is only a hint and applies to
1929 // functions that the frontend might not understand as builtins.
1931 // This heuristic was suggested in:
1932 // Improving Static Branch Prediction in a Compiler
1933 // Brian L. Deitrich, Ben-Chung Cheng, Wen-mei W. Hwu
1934 // Proceedings of PACT'98, Oct. 1998, IEEE
1935 if (!CI->hasFnAttr(Attribute::Cold) &&
1936 isReportingError(Callee, CI, StreamArg)) {
1937 CI->addAttribute(AttributeList::FunctionIndex, Attribute::Cold);
1943 static bool isReportingError(Function *Callee, CallInst *CI, int StreamArg) {
1944 if (!Callee || !Callee->isDeclaration())
1950 // These functions might be considered cold, but only if their stream
1951 // argument is stderr.
1953 if (StreamArg >= (int)CI->getNumArgOperands())
1955 LoadInst *LI = dyn_cast<LoadInst>(CI->getArgOperand(StreamArg));
1958 GlobalVariable *GV = dyn_cast<GlobalVariable>(LI->getPointerOperand());
1959 if (!GV || !GV->isDeclaration())
1961 return GV->getName() == "stderr";
1964 Value *LibCallSimplifier::optimizePrintFString(CallInst *CI, IRBuilder<> &B) {
1965 // Check for a fixed format string.
1966 StringRef FormatStr;
1967 if (!getConstantStringInfo(CI->getArgOperand(0), FormatStr))
1970 // Empty format string -> noop.
1971 if (FormatStr.empty()) // Tolerate printf's declared void.
1972 return CI->use_empty() ? (Value *)CI : ConstantInt::get(CI->getType(), 0);
1974 // Do not do any of the following transformations if the printf return value
1975 // is used, in general the printf return value is not compatible with either
1976 // putchar() or puts().
1977 if (!CI->use_empty())
1980 // printf("x") -> putchar('x'), even for "%" and "%%".
1981 if (FormatStr.size() == 1 || FormatStr == "%%")
1982 return emitPutChar(B.getInt32(FormatStr[0]), B, TLI);
1984 // printf("%s", "a") --> putchar('a')
1985 if (FormatStr == "%s" && CI->getNumArgOperands() > 1) {
1987 if (!getConstantStringInfo(CI->getOperand(1), ChrStr))
1989 if (ChrStr.size() != 1)
1991 return emitPutChar(B.getInt32(ChrStr[0]), B, TLI);
1994 // printf("foo\n") --> puts("foo")
1995 if (FormatStr[FormatStr.size() - 1] == '\n' &&
1996 FormatStr.find('%') == StringRef::npos) { // No format characters.
1997 // Create a string literal with no \n on it. We expect the constant merge
1998 // pass to be run after this pass, to merge duplicate strings.
1999 FormatStr = FormatStr.drop_back();
2000 Value *GV = B.CreateGlobalString(FormatStr, "str");
2001 return emitPutS(GV, B, TLI);
2004 // Optimize specific format strings.
2005 // printf("%c", chr) --> putchar(chr)
2006 if (FormatStr == "%c" && CI->getNumArgOperands() > 1 &&
2007 CI->getArgOperand(1)->getType()->isIntegerTy())
2008 return emitPutChar(CI->getArgOperand(1), B, TLI);
2010 // printf("%s\n", str) --> puts(str)
2011 if (FormatStr == "%s\n" && CI->getNumArgOperands() > 1 &&
2012 CI->getArgOperand(1)->getType()->isPointerTy())
2013 return emitPutS(CI->getArgOperand(1), B, TLI);
2017 Value *LibCallSimplifier::optimizePrintF(CallInst *CI, IRBuilder<> &B) {
2019 Function *Callee = CI->getCalledFunction();
2020 FunctionType *FT = Callee->getFunctionType();
2021 if (Value *V = optimizePrintFString(CI, B)) {
2025 // printf(format, ...) -> iprintf(format, ...) if no floating point
2027 if (TLI->has(LibFunc_iprintf) && !callHasFloatingPointArgument(CI)) {
2028 Module *M = B.GetInsertBlock()->getParent()->getParent();
2029 Constant *IPrintFFn =
2030 M->getOrInsertFunction("iprintf", FT, Callee->getAttributes());
2031 CallInst *New = cast<CallInst>(CI->clone());
2032 New->setCalledFunction(IPrintFFn);
2039 Value *LibCallSimplifier::optimizeSPrintFString(CallInst *CI, IRBuilder<> &B) {
2040 // Check for a fixed format string.
2041 StringRef FormatStr;
2042 if (!getConstantStringInfo(CI->getArgOperand(1), FormatStr))
2045 // If we just have a format string (nothing else crazy) transform it.
2046 if (CI->getNumArgOperands() == 2) {
2047 // Make sure there's no % in the constant array. We could try to handle
2048 // %% -> % in the future if we cared.
2049 if (FormatStr.find('%') != StringRef::npos)
2050 return nullptr; // we found a format specifier, bail out.
2052 // sprintf(str, fmt) -> llvm.memcpy(align 1 str, align 1 fmt, strlen(fmt)+1)
2053 B.CreateMemCpy(CI->getArgOperand(0), 1, CI->getArgOperand(1), 1,
2054 ConstantInt::get(DL.getIntPtrType(CI->getContext()),
2055 FormatStr.size() + 1)); // Copy the null byte.
2056 return ConstantInt::get(CI->getType(), FormatStr.size());
2059 // The remaining optimizations require the format string to be "%s" or "%c"
2060 // and have an extra operand.
2061 if (FormatStr.size() != 2 || FormatStr[0] != '%' ||
2062 CI->getNumArgOperands() < 3)
2065 // Decode the second character of the format string.
2066 if (FormatStr[1] == 'c') {
2067 // sprintf(dst, "%c", chr) --> *(i8*)dst = chr; *((i8*)dst+1) = 0
2068 if (!CI->getArgOperand(2)->getType()->isIntegerTy())
2070 Value *V = B.CreateTrunc(CI->getArgOperand(2), B.getInt8Ty(), "char");
2071 Value *Ptr = castToCStr(CI->getArgOperand(0), B);
2072 B.CreateStore(V, Ptr);
2073 Ptr = B.CreateGEP(B.getInt8Ty(), Ptr, B.getInt32(1), "nul");
2074 B.CreateStore(B.getInt8(0), Ptr);
2076 return ConstantInt::get(CI->getType(), 1);
2079 if (FormatStr[1] == 's') {
2080 // sprintf(dest, "%s", str) -> llvm.memcpy(dest, str, strlen(str)+1, 1)
2081 if (!CI->getArgOperand(2)->getType()->isPointerTy())
2084 Value *Len = emitStrLen(CI->getArgOperand(2), B, DL, TLI);
2088 B.CreateAdd(Len, ConstantInt::get(Len->getType(), 1), "leninc");
2089 B.CreateMemCpy(CI->getArgOperand(0), 1, CI->getArgOperand(2), 1, IncLen);
2091 // The sprintf result is the unincremented number of bytes in the string.
2092 return B.CreateIntCast(Len, CI->getType(), false);
2097 Value *LibCallSimplifier::optimizeSPrintF(CallInst *CI, IRBuilder<> &B) {
2098 Function *Callee = CI->getCalledFunction();
2099 FunctionType *FT = Callee->getFunctionType();
2100 if (Value *V = optimizeSPrintFString(CI, B)) {
2104 // sprintf(str, format, ...) -> siprintf(str, format, ...) if no floating
2106 if (TLI->has(LibFunc_siprintf) && !callHasFloatingPointArgument(CI)) {
2107 Module *M = B.GetInsertBlock()->getParent()->getParent();
2108 Constant *SIPrintFFn =
2109 M->getOrInsertFunction("siprintf", FT, Callee->getAttributes());
2110 CallInst *New = cast<CallInst>(CI->clone());
2111 New->setCalledFunction(SIPrintFFn);
2118 Value *LibCallSimplifier::optimizeSnPrintFString(CallInst *CI, IRBuilder<> &B) {
2119 // Check for a fixed format string.
2120 StringRef FormatStr;
2121 if (!getConstantStringInfo(CI->getArgOperand(2), FormatStr))
2125 ConstantInt *Size = dyn_cast<ConstantInt>(CI->getArgOperand(1));
2129 uint64_t N = Size->getZExtValue();
2131 // If we just have a format string (nothing else crazy) transform it.
2132 if (CI->getNumArgOperands() == 3) {
2133 // Make sure there's no % in the constant array. We could try to handle
2134 // %% -> % in the future if we cared.
2135 if (FormatStr.find('%') != StringRef::npos)
2136 return nullptr; // we found a format specifier, bail out.
2139 return ConstantInt::get(CI->getType(), FormatStr.size());
2140 else if (N < FormatStr.size() + 1)
2143 // sprintf(str, size, fmt) -> llvm.memcpy(align 1 str, align 1 fmt,
2146 CI->getArgOperand(0), 1, CI->getArgOperand(2), 1,
2147 ConstantInt::get(DL.getIntPtrType(CI->getContext()),
2148 FormatStr.size() + 1)); // Copy the null byte.
2149 return ConstantInt::get(CI->getType(), FormatStr.size());
2152 // The remaining optimizations require the format string to be "%s" or "%c"
2153 // and have an extra operand.
2154 if (FormatStr.size() == 2 && FormatStr[0] == '%' &&
2155 CI->getNumArgOperands() == 4) {
2157 // Decode the second character of the format string.
2158 if (FormatStr[1] == 'c') {
2160 return ConstantInt::get(CI->getType(), 1);
2164 // snprintf(dst, size, "%c", chr) --> *(i8*)dst = chr; *((i8*)dst+1) = 0
2165 if (!CI->getArgOperand(3)->getType()->isIntegerTy())
2167 Value *V = B.CreateTrunc(CI->getArgOperand(3), B.getInt8Ty(), "char");
2168 Value *Ptr = castToCStr(CI->getArgOperand(0), B);
2169 B.CreateStore(V, Ptr);
2170 Ptr = B.CreateGEP(B.getInt8Ty(), Ptr, B.getInt32(1), "nul");
2171 B.CreateStore(B.getInt8(0), Ptr);
2173 return ConstantInt::get(CI->getType(), 1);
2176 if (FormatStr[1] == 's') {
2177 // snprintf(dest, size, "%s", str) to llvm.memcpy(dest, str, len+1, 1)
2179 if (!getConstantStringInfo(CI->getArgOperand(3), Str))
2183 return ConstantInt::get(CI->getType(), Str.size());
2184 else if (N < Str.size() + 1)
2187 B.CreateMemCpy(CI->getArgOperand(0), 1, CI->getArgOperand(3), 1,
2188 ConstantInt::get(CI->getType(), Str.size() + 1));
2190 // The snprintf result is the unincremented number of bytes in the string.
2191 return ConstantInt::get(CI->getType(), Str.size());
2197 Value *LibCallSimplifier::optimizeSnPrintF(CallInst *CI, IRBuilder<> &B) {
2198 if (Value *V = optimizeSnPrintFString(CI, B)) {
2205 Value *LibCallSimplifier::optimizeFPrintFString(CallInst *CI, IRBuilder<> &B) {
2206 optimizeErrorReporting(CI, B, 0);
2208 // All the optimizations depend on the format string.
2209 StringRef FormatStr;
2210 if (!getConstantStringInfo(CI->getArgOperand(1), FormatStr))
2213 // Do not do any of the following transformations if the fprintf return
2214 // value is used, in general the fprintf return value is not compatible
2215 // with fwrite(), fputc() or fputs().
2216 if (!CI->use_empty())
2219 // fprintf(F, "foo") --> fwrite("foo", 3, 1, F)
2220 if (CI->getNumArgOperands() == 2) {
2221 // Could handle %% -> % if we cared.
2222 if (FormatStr.find('%') != StringRef::npos)
2223 return nullptr; // We found a format specifier.
2226 CI->getArgOperand(1),
2227 ConstantInt::get(DL.getIntPtrType(CI->getContext()), FormatStr.size()),
2228 CI->getArgOperand(0), B, DL, TLI);
2231 // The remaining optimizations require the format string to be "%s" or "%c"
2232 // and have an extra operand.
2233 if (FormatStr.size() != 2 || FormatStr[0] != '%' ||
2234 CI->getNumArgOperands() < 3)
2237 // Decode the second character of the format string.
2238 if (FormatStr[1] == 'c') {
2239 // fprintf(F, "%c", chr) --> fputc(chr, F)
2240 if (!CI->getArgOperand(2)->getType()->isIntegerTy())
2242 return emitFPutC(CI->getArgOperand(2), CI->getArgOperand(0), B, TLI);
2245 if (FormatStr[1] == 's') {
2246 // fprintf(F, "%s", str) --> fputs(str, F)
2247 if (!CI->getArgOperand(2)->getType()->isPointerTy())
2249 return emitFPutS(CI->getArgOperand(2), CI->getArgOperand(0), B, TLI);
2254 Value *LibCallSimplifier::optimizeFPrintF(CallInst *CI, IRBuilder<> &B) {
2255 Function *Callee = CI->getCalledFunction();
2256 FunctionType *FT = Callee->getFunctionType();
2257 if (Value *V = optimizeFPrintFString(CI, B)) {
2261 // fprintf(stream, format, ...) -> fiprintf(stream, format, ...) if no
2262 // floating point arguments.
2263 if (TLI->has(LibFunc_fiprintf) && !callHasFloatingPointArgument(CI)) {
2264 Module *M = B.GetInsertBlock()->getParent()->getParent();
2265 Constant *FIPrintFFn =
2266 M->getOrInsertFunction("fiprintf", FT, Callee->getAttributes());
2267 CallInst *New = cast<CallInst>(CI->clone());
2268 New->setCalledFunction(FIPrintFFn);
2275 Value *LibCallSimplifier::optimizeFWrite(CallInst *CI, IRBuilder<> &B) {
2276 optimizeErrorReporting(CI, B, 3);
2278 // Get the element size and count.
2279 ConstantInt *SizeC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
2280 ConstantInt *CountC = dyn_cast<ConstantInt>(CI->getArgOperand(2));
2281 if (SizeC && CountC) {
2282 uint64_t Bytes = SizeC->getZExtValue() * CountC->getZExtValue();
2284 // If this is writing zero records, remove the call (it's a noop).
2286 return ConstantInt::get(CI->getType(), 0);
2288 // If this is writing one byte, turn it into fputc.
2289 // This optimisation is only valid, if the return value is unused.
2290 if (Bytes == 1 && CI->use_empty()) { // fwrite(S,1,1,F) -> fputc(S[0],F)
2291 Value *Char = B.CreateLoad(castToCStr(CI->getArgOperand(0), B), "char");
2292 Value *NewCI = emitFPutC(Char, CI->getArgOperand(3), B, TLI);
2293 return NewCI ? ConstantInt::get(CI->getType(), 1) : nullptr;
2297 if (isLocallyOpenedFile(CI->getArgOperand(3), CI, B, TLI))
2298 return emitFWriteUnlocked(CI->getArgOperand(0), CI->getArgOperand(1),
2299 CI->getArgOperand(2), CI->getArgOperand(3), B, DL,
2305 Value *LibCallSimplifier::optimizeFPuts(CallInst *CI, IRBuilder<> &B) {
2306 optimizeErrorReporting(CI, B, 1);
2308 // Don't rewrite fputs to fwrite when optimising for size because fwrite
2309 // requires more arguments and thus extra MOVs are required.
2310 if (CI->getFunction()->optForSize())
2313 // Check if has any use
2314 if (!CI->use_empty()) {
2315 if (isLocallyOpenedFile(CI->getArgOperand(1), CI, B, TLI))
2316 return emitFPutSUnlocked(CI->getArgOperand(0), CI->getArgOperand(1), B,
2319 // We can't optimize if return value is used.
2323 // fputs(s,F) --> fwrite(s,1,strlen(s),F)
2324 uint64_t Len = GetStringLength(CI->getArgOperand(0));
2328 // Known to have no uses (see above).
2330 CI->getArgOperand(0),
2331 ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len - 1),
2332 CI->getArgOperand(1), B, DL, TLI);
2335 Value *LibCallSimplifier::optimizeFPutc(CallInst *CI, IRBuilder<> &B) {
2336 optimizeErrorReporting(CI, B, 1);
2338 if (isLocallyOpenedFile(CI->getArgOperand(1), CI, B, TLI))
2339 return emitFPutCUnlocked(CI->getArgOperand(0), CI->getArgOperand(1), B,
2345 Value *LibCallSimplifier::optimizeFGetc(CallInst *CI, IRBuilder<> &B) {
2346 if (isLocallyOpenedFile(CI->getArgOperand(0), CI, B, TLI))
2347 return emitFGetCUnlocked(CI->getArgOperand(0), B, TLI);
2352 Value *LibCallSimplifier::optimizeFGets(CallInst *CI, IRBuilder<> &B) {
2353 if (isLocallyOpenedFile(CI->getArgOperand(2), CI, B, TLI))
2354 return emitFGetSUnlocked(CI->getArgOperand(0), CI->getArgOperand(1),
2355 CI->getArgOperand(2), B, TLI);
2360 Value *LibCallSimplifier::optimizeFRead(CallInst *CI, IRBuilder<> &B) {
2361 if (isLocallyOpenedFile(CI->getArgOperand(3), CI, B, TLI))
2362 return emitFReadUnlocked(CI->getArgOperand(0), CI->getArgOperand(1),
2363 CI->getArgOperand(2), CI->getArgOperand(3), B, DL,
2369 Value *LibCallSimplifier::optimizePuts(CallInst *CI, IRBuilder<> &B) {
2370 // Check for a constant string.
2372 if (!getConstantStringInfo(CI->getArgOperand(0), Str))
2375 if (Str.empty() && CI->use_empty()) {
2376 // puts("") -> putchar('\n')
2377 Value *Res = emitPutChar(B.getInt32('\n'), B, TLI);
2378 if (CI->use_empty() || !Res)
2380 return B.CreateIntCast(Res, CI->getType(), true);
2386 bool LibCallSimplifier::hasFloatVersion(StringRef FuncName) {
2388 SmallString<20> FloatFuncName = FuncName;
2389 FloatFuncName += 'f';
2390 if (TLI->getLibFunc(FloatFuncName, Func))
2391 return TLI->has(Func);
2395 Value *LibCallSimplifier::optimizeStringMemoryLibCall(CallInst *CI,
2396 IRBuilder<> &Builder) {
2398 Function *Callee = CI->getCalledFunction();
2399 // Check for string/memory library functions.
2400 if (TLI->getLibFunc(*Callee, Func) && TLI->has(Func)) {
2401 // Make sure we never change the calling convention.
2402 assert((ignoreCallingConv(Func) ||
2403 isCallingConvCCompatible(CI)) &&
2404 "Optimizing string/memory libcall would change the calling convention");
2406 case LibFunc_strcat:
2407 return optimizeStrCat(CI, Builder);
2408 case LibFunc_strncat:
2409 return optimizeStrNCat(CI, Builder);
2410 case LibFunc_strchr:
2411 return optimizeStrChr(CI, Builder);
2412 case LibFunc_strrchr:
2413 return optimizeStrRChr(CI, Builder);
2414 case LibFunc_strcmp:
2415 return optimizeStrCmp(CI, Builder);
2416 case LibFunc_strncmp:
2417 return optimizeStrNCmp(CI, Builder);
2418 case LibFunc_strcpy:
2419 return optimizeStrCpy(CI, Builder);
2420 case LibFunc_stpcpy:
2421 return optimizeStpCpy(CI, Builder);
2422 case LibFunc_strncpy:
2423 return optimizeStrNCpy(CI, Builder);
2424 case LibFunc_strlen:
2425 return optimizeStrLen(CI, Builder);
2426 case LibFunc_strpbrk:
2427 return optimizeStrPBrk(CI, Builder);
2428 case LibFunc_strtol:
2429 case LibFunc_strtod:
2430 case LibFunc_strtof:
2431 case LibFunc_strtoul:
2432 case LibFunc_strtoll:
2433 case LibFunc_strtold:
2434 case LibFunc_strtoull:
2435 return optimizeStrTo(CI, Builder);
2436 case LibFunc_strspn:
2437 return optimizeStrSpn(CI, Builder);
2438 case LibFunc_strcspn:
2439 return optimizeStrCSpn(CI, Builder);
2440 case LibFunc_strstr:
2441 return optimizeStrStr(CI, Builder);
2442 case LibFunc_memchr:
2443 return optimizeMemChr(CI, Builder);
2444 case LibFunc_memcmp:
2445 return optimizeMemCmp(CI, Builder);
2446 case LibFunc_memcpy:
2447 return optimizeMemCpy(CI, Builder);
2448 case LibFunc_memmove:
2449 return optimizeMemMove(CI, Builder);
2450 case LibFunc_memset:
2451 return optimizeMemSet(CI, Builder);
2452 case LibFunc_realloc:
2453 return optimizeRealloc(CI, Builder);
2454 case LibFunc_wcslen:
2455 return optimizeWcslen(CI, Builder);
2463 Value *LibCallSimplifier::optimizeFloatingPointLibCall(CallInst *CI,
2465 IRBuilder<> &Builder) {
2466 // Don't optimize calls that require strict floating point semantics.
2467 if (CI->isStrictFP())
2470 if (Value *V = optimizeTrigReflections(CI, Func, Builder))
2474 case LibFunc_sinpif:
2476 case LibFunc_cospif:
2478 return optimizeSinCosPi(CI, Builder);
2482 return optimizePow(CI, Builder);
2486 return optimizeExp2(CI, Builder);
2490 return replaceUnaryCall(CI, Builder, Intrinsic::fabs);
2494 return optimizeSqrt(CI, Builder);
2500 return optimizeLog(CI, Builder);
2504 return optimizeTan(CI, Builder);
2506 return replaceUnaryCall(CI, Builder, Intrinsic::ceil);
2508 return replaceUnaryCall(CI, Builder, Intrinsic::floor);
2510 return replaceUnaryCall(CI, Builder, Intrinsic::round);
2511 case LibFunc_nearbyint:
2512 return replaceUnaryCall(CI, Builder, Intrinsic::nearbyint);
2514 return replaceUnaryCall(CI, Builder, Intrinsic::rint);
2516 return replaceUnaryCall(CI, Builder, Intrinsic::trunc);
2532 if (UnsafeFPShrink && hasFloatVersion(CI->getCalledFunction()->getName()))
2533 return optimizeUnaryDoubleFP(CI, Builder, true);
2535 case LibFunc_copysign:
2536 if (hasFloatVersion(CI->getCalledFunction()->getName()))
2537 return optimizeBinaryDoubleFP(CI, Builder);
2545 return optimizeFMinFMax(CI, Builder);
2549 return optimizeCAbs(CI, Builder);
2555 Value *LibCallSimplifier::optimizeCall(CallInst *CI) {
2556 // TODO: Split out the code below that operates on FP calls so that
2557 // we can all non-FP calls with the StrictFP attribute to be
2559 if (CI->isNoBuiltin())
2563 Function *Callee = CI->getCalledFunction();
2565 SmallVector<OperandBundleDef, 2> OpBundles;
2566 CI->getOperandBundlesAsDefs(OpBundles);
2567 IRBuilder<> Builder(CI, /*FPMathTag=*/nullptr, OpBundles);
2568 bool isCallingConvC = isCallingConvCCompatible(CI);
2570 // Command-line parameter overrides instruction attribute.
2571 // This can't be moved to optimizeFloatingPointLibCall() because it may be
2572 // used by the intrinsic optimizations.
2573 if (EnableUnsafeFPShrink.getNumOccurrences() > 0)
2574 UnsafeFPShrink = EnableUnsafeFPShrink;
2575 else if (isa<FPMathOperator>(CI) && CI->isFast())
2576 UnsafeFPShrink = true;
2578 // First, check for intrinsics.
2579 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
2580 if (!isCallingConvC)
2582 // The FP intrinsics have corresponding constrained versions so we don't
2583 // need to check for the StrictFP attribute here.
2584 switch (II->getIntrinsicID()) {
2585 case Intrinsic::pow:
2586 return optimizePow(CI, Builder);
2587 case Intrinsic::exp2:
2588 return optimizeExp2(CI, Builder);
2589 case Intrinsic::log:
2590 return optimizeLog(CI, Builder);
2591 case Intrinsic::sqrt:
2592 return optimizeSqrt(CI, Builder);
2593 // TODO: Use foldMallocMemset() with memset intrinsic.
2599 // Also try to simplify calls to fortified library functions.
2600 if (Value *SimplifiedFortifiedCI = FortifiedSimplifier.optimizeCall(CI)) {
2601 // Try to further simplify the result.
2602 CallInst *SimplifiedCI = dyn_cast<CallInst>(SimplifiedFortifiedCI);
2603 if (SimplifiedCI && SimplifiedCI->getCalledFunction()) {
2604 // Use an IR Builder from SimplifiedCI if available instead of CI
2605 // to guarantee we reach all uses we might replace later on.
2606 IRBuilder<> TmpBuilder(SimplifiedCI);
2607 if (Value *V = optimizeStringMemoryLibCall(SimplifiedCI, TmpBuilder)) {
2608 // If we were able to further simplify, remove the now redundant call.
2609 SimplifiedCI->replaceAllUsesWith(V);
2610 eraseFromParent(SimplifiedCI);
2614 return SimplifiedFortifiedCI;
2617 // Then check for known library functions.
2618 if (TLI->getLibFunc(*Callee, Func) && TLI->has(Func)) {
2619 // We never change the calling convention.
2620 if (!ignoreCallingConv(Func) && !isCallingConvC)
2622 if (Value *V = optimizeStringMemoryLibCall(CI, Builder))
2624 if (Value *V = optimizeFloatingPointLibCall(CI, Func, Builder))
2630 return optimizeFFS(CI, Builder);
2634 return optimizeFls(CI, Builder);
2638 return optimizeAbs(CI, Builder);
2639 case LibFunc_isdigit:
2640 return optimizeIsDigit(CI, Builder);
2641 case LibFunc_isascii:
2642 return optimizeIsAscii(CI, Builder);
2643 case LibFunc_toascii:
2644 return optimizeToAscii(CI, Builder);
2648 return optimizeAtoi(CI, Builder);
2649 case LibFunc_strtol:
2650 case LibFunc_strtoll:
2651 return optimizeStrtol(CI, Builder);
2652 case LibFunc_printf:
2653 return optimizePrintF(CI, Builder);
2654 case LibFunc_sprintf:
2655 return optimizeSPrintF(CI, Builder);
2656 case LibFunc_snprintf:
2657 return optimizeSnPrintF(CI, Builder);
2658 case LibFunc_fprintf:
2659 return optimizeFPrintF(CI, Builder);
2660 case LibFunc_fwrite:
2661 return optimizeFWrite(CI, Builder);
2663 return optimizeFRead(CI, Builder);
2665 return optimizeFPuts(CI, Builder);
2667 return optimizeFGets(CI, Builder);
2669 return optimizeFPutc(CI, Builder);
2671 return optimizeFGetc(CI, Builder);
2673 return optimizePuts(CI, Builder);
2674 case LibFunc_perror:
2675 return optimizeErrorReporting(CI, Builder);
2676 case LibFunc_vfprintf:
2677 case LibFunc_fiprintf:
2678 return optimizeErrorReporting(CI, Builder, 0);
2686 LibCallSimplifier::LibCallSimplifier(
2687 const DataLayout &DL, const TargetLibraryInfo *TLI,
2688 OptimizationRemarkEmitter &ORE,
2689 function_ref<void(Instruction *, Value *)> Replacer,
2690 function_ref<void(Instruction *)> Eraser)
2691 : FortifiedSimplifier(TLI), DL(DL), TLI(TLI), ORE(ORE),
2692 UnsafeFPShrink(false), Replacer(Replacer), Eraser(Eraser) {}
2694 void LibCallSimplifier::replaceAllUsesWith(Instruction *I, Value *With) {
2695 // Indirect through the replacer used in this instance.
2699 void LibCallSimplifier::eraseFromParent(Instruction *I) {
2704 // Additional cases that we need to add to this file:
2707 // * cbrt(expN(X)) -> expN(x/3)
2708 // * cbrt(sqrt(x)) -> pow(x,1/6)
2709 // * cbrt(cbrt(x)) -> pow(x,1/9)
2712 // * exp(log(x)) -> x
2715 // * log(exp(x)) -> x
2716 // * log(exp(y)) -> y*log(e)
2717 // * log(exp10(y)) -> y*log(10)
2718 // * log(sqrt(x)) -> 0.5*log(x)
2721 // * pow(sqrt(x),y) -> pow(x,y*0.5)
2722 // * pow(pow(x,y),z)-> pow(x,y*z)
2725 // * signbit(cnst) -> cnst'
2726 // * signbit(nncst) -> 0 (if pstv is a non-negative constant)
2728 // sqrt, sqrtf, sqrtl:
2729 // * sqrt(expN(x)) -> expN(x*0.5)
2730 // * sqrt(Nroot(x)) -> pow(x,1/(2*N))
2731 // * sqrt(pow(x,y)) -> pow(|x|,y*0.5)
2734 //===----------------------------------------------------------------------===//
2735 // Fortified Library Call Optimizations
2736 //===----------------------------------------------------------------------===//
2738 bool FortifiedLibCallSimplifier::isFortifiedCallFoldable(CallInst *CI,
2742 if (CI->getArgOperand(ObjSizeOp) == CI->getArgOperand(SizeOp))
2744 if (ConstantInt *ObjSizeCI =
2745 dyn_cast<ConstantInt>(CI->getArgOperand(ObjSizeOp))) {
2746 if (ObjSizeCI->isMinusOne())
2748 // If the object size wasn't -1 (unknown), bail out if we were asked to.
2749 if (OnlyLowerUnknownSize)
2752 uint64_t Len = GetStringLength(CI->getArgOperand(SizeOp));
2753 // If the length is 0 we don't know how long it is and so we can't
2754 // remove the check.
2757 return ObjSizeCI->getZExtValue() >= Len;
2759 if (ConstantInt *SizeCI = dyn_cast<ConstantInt>(CI->getArgOperand(SizeOp)))
2760 return ObjSizeCI->getZExtValue() >= SizeCI->getZExtValue();
2765 Value *FortifiedLibCallSimplifier::optimizeMemCpyChk(CallInst *CI,
2767 if (isFortifiedCallFoldable(CI, 3, 2, false)) {
2768 B.CreateMemCpy(CI->getArgOperand(0), 1, CI->getArgOperand(1), 1,
2769 CI->getArgOperand(2));
2770 return CI->getArgOperand(0);
2775 Value *FortifiedLibCallSimplifier::optimizeMemMoveChk(CallInst *CI,
2777 if (isFortifiedCallFoldable(CI, 3, 2, false)) {
2778 B.CreateMemMove(CI->getArgOperand(0), 1, CI->getArgOperand(1), 1,
2779 CI->getArgOperand(2));
2780 return CI->getArgOperand(0);
2785 Value *FortifiedLibCallSimplifier::optimizeMemSetChk(CallInst *CI,
2787 // TODO: Try foldMallocMemset() here.
2789 if (isFortifiedCallFoldable(CI, 3, 2, false)) {
2790 Value *Val = B.CreateIntCast(CI->getArgOperand(1), B.getInt8Ty(), false);
2791 B.CreateMemSet(CI->getArgOperand(0), Val, CI->getArgOperand(2), 1);
2792 return CI->getArgOperand(0);
2797 Value *FortifiedLibCallSimplifier::optimizeStrpCpyChk(CallInst *CI,
2800 Function *Callee = CI->getCalledFunction();
2801 StringRef Name = Callee->getName();
2802 const DataLayout &DL = CI->getModule()->getDataLayout();
2803 Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1),
2804 *ObjSize = CI->getArgOperand(2);
2806 // __stpcpy_chk(x,x,...) -> x+strlen(x)
2807 if (Func == LibFunc_stpcpy_chk && !OnlyLowerUnknownSize && Dst == Src) {
2808 Value *StrLen = emitStrLen(Src, B, DL, TLI);
2809 return StrLen ? B.CreateInBoundsGEP(B.getInt8Ty(), Dst, StrLen) : nullptr;
2812 // If a) we don't have any length information, or b) we know this will
2813 // fit then just lower to a plain st[rp]cpy. Otherwise we'll keep our
2814 // st[rp]cpy_chk call which may fail at runtime if the size is too long.
2815 // TODO: It might be nice to get a maximum length out of the possible
2816 // string lengths for varying.
2817 if (isFortifiedCallFoldable(CI, 2, 1, true))
2818 return emitStrCpy(Dst, Src, B, TLI, Name.substr(2, 6));
2820 if (OnlyLowerUnknownSize)
2823 // Maybe we can stil fold __st[rp]cpy_chk to __memcpy_chk.
2824 uint64_t Len = GetStringLength(Src);
2828 Type *SizeTTy = DL.getIntPtrType(CI->getContext());
2829 Value *LenV = ConstantInt::get(SizeTTy, Len);
2830 Value *Ret = emitMemCpyChk(Dst, Src, LenV, ObjSize, B, DL, TLI);
2831 // If the function was an __stpcpy_chk, and we were able to fold it into
2832 // a __memcpy_chk, we still need to return the correct end pointer.
2833 if (Ret && Func == LibFunc_stpcpy_chk)
2834 return B.CreateGEP(B.getInt8Ty(), Dst, ConstantInt::get(SizeTTy, Len - 1));
2838 Value *FortifiedLibCallSimplifier::optimizeStrpNCpyChk(CallInst *CI,
2841 Function *Callee = CI->getCalledFunction();
2842 StringRef Name = Callee->getName();
2843 if (isFortifiedCallFoldable(CI, 3, 2, false)) {
2844 Value *Ret = emitStrNCpy(CI->getArgOperand(0), CI->getArgOperand(1),
2845 CI->getArgOperand(2), B, TLI, Name.substr(2, 7));
2851 Value *FortifiedLibCallSimplifier::optimizeCall(CallInst *CI) {
2852 // FIXME: We shouldn't be changing "nobuiltin" or TLI unavailable calls here.
2853 // Some clang users checked for _chk libcall availability using:
2854 // __has_builtin(__builtin___memcpy_chk)
2855 // When compiling with -fno-builtin, this is always true.
2856 // When passing -ffreestanding/-mkernel, which both imply -fno-builtin, we
2857 // end up with fortified libcalls, which isn't acceptable in a freestanding
2858 // environment which only provides their non-fortified counterparts.
2860 // Until we change clang and/or teach external users to check for availability
2861 // differently, disregard the "nobuiltin" attribute and TLI::has.
2866 Function *Callee = CI->getCalledFunction();
2868 SmallVector<OperandBundleDef, 2> OpBundles;
2869 CI->getOperandBundlesAsDefs(OpBundles);
2870 IRBuilder<> Builder(CI, /*FPMathTag=*/nullptr, OpBundles);
2871 bool isCallingConvC = isCallingConvCCompatible(CI);
2873 // First, check that this is a known library functions and that the prototype
2875 if (!TLI->getLibFunc(*Callee, Func))
2878 // We never change the calling convention.
2879 if (!ignoreCallingConv(Func) && !isCallingConvC)
2883 case LibFunc_memcpy_chk:
2884 return optimizeMemCpyChk(CI, Builder);
2885 case LibFunc_memmove_chk:
2886 return optimizeMemMoveChk(CI, Builder);
2887 case LibFunc_memset_chk:
2888 return optimizeMemSetChk(CI, Builder);
2889 case LibFunc_stpcpy_chk:
2890 case LibFunc_strcpy_chk:
2891 return optimizeStrpCpyChk(CI, Builder, Func);
2892 case LibFunc_stpncpy_chk:
2893 case LibFunc_strncpy_chk:
2894 return optimizeStrpNCpyChk(CI, Builder, Func);
2901 FortifiedLibCallSimplifier::FortifiedLibCallSimplifier(
2902 const TargetLibraryInfo *TLI, bool OnlyLowerUnknownSize)
2903 : TLI(TLI), OnlyLowerUnknownSize(OnlyLowerUnknownSize) {}