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 is a utility pass used for testing the InstructionSimplify analysis.
11 // The analysis is applied to every instruction, and if it simplifies then the
12 // instruction is replaced by the simplification. If you are looking for a pass
13 // that performs serious instruction folding, use the instcombine pass instead.
15 //===----------------------------------------------------------------------===//
17 #include "llvm/Transforms/Utils/SimplifyLibCalls.h"
18 #include "llvm/ADT/SmallString.h"
19 #include "llvm/ADT/StringMap.h"
20 #include "llvm/ADT/Triple.h"
21 #include "llvm/Analysis/TargetLibraryInfo.h"
22 #include "llvm/Analysis/ValueTracking.h"
23 #include "llvm/IR/DataLayout.h"
24 #include "llvm/IR/DiagnosticInfo.h"
25 #include "llvm/IR/Function.h"
26 #include "llvm/IR/IRBuilder.h"
27 #include "llvm/IR/IntrinsicInst.h"
28 #include "llvm/IR/Intrinsics.h"
29 #include "llvm/IR/LLVMContext.h"
30 #include "llvm/IR/Module.h"
31 #include "llvm/IR/PatternMatch.h"
32 #include "llvm/Support/CommandLine.h"
33 #include "llvm/Transforms/Utils/BuildLibCalls.h"
34 #include "llvm/Transforms/Utils/Local.h"
37 using namespace PatternMatch;
40 ColdErrorCalls("error-reporting-is-cold", cl::init(true), cl::Hidden,
41 cl::desc("Treat error-reporting calls as cold"));
44 EnableUnsafeFPShrink("enable-double-float-shrink", cl::Hidden,
46 cl::desc("Enable unsafe double to float "
47 "shrinking for math lib calls"));
50 //===----------------------------------------------------------------------===//
52 //===----------------------------------------------------------------------===//
54 static bool ignoreCallingConv(LibFunc::Func Func) {
55 return Func == LibFunc::abs || Func == LibFunc::labs ||
56 Func == LibFunc::llabs || Func == LibFunc::strlen;
59 static bool isCallingConvCCompatible(CallInst *CI) {
60 switch(CI->getCallingConv()) {
63 case llvm::CallingConv::C:
65 case llvm::CallingConv::ARM_APCS:
66 case llvm::CallingConv::ARM_AAPCS:
67 case llvm::CallingConv::ARM_AAPCS_VFP: {
69 // The iOS ABI diverges from the standard in some cases, so for now don't
70 // try to simplify those calls.
71 if (Triple(CI->getModule()->getTargetTriple()).isiOS())
74 auto *FuncTy = CI->getFunctionType();
76 if (!FuncTy->getReturnType()->isPointerTy() &&
77 !FuncTy->getReturnType()->isIntegerTy() &&
78 !FuncTy->getReturnType()->isVoidTy())
81 for (auto Param : FuncTy->params()) {
82 if (!Param->isPointerTy() && !Param->isIntegerTy())
91 /// Return true if it only matters that the value is equal or not-equal to zero.
92 static bool isOnlyUsedInZeroEqualityComparison(Value *V) {
93 for (User *U : V->users()) {
94 if (ICmpInst *IC = dyn_cast<ICmpInst>(U))
96 if (Constant *C = dyn_cast<Constant>(IC->getOperand(1)))
99 // Unknown instruction.
105 /// Return true if it is only used in equality comparisons with With.
106 static bool isOnlyUsedInEqualityComparison(Value *V, Value *With) {
107 for (User *U : V->users()) {
108 if (ICmpInst *IC = dyn_cast<ICmpInst>(U))
109 if (IC->isEquality() && IC->getOperand(1) == With)
111 // Unknown instruction.
117 static bool callHasFloatingPointArgument(const CallInst *CI) {
118 return any_of(CI->operands(), [](const Use &OI) {
119 return OI->getType()->isFloatingPointTy();
123 /// \brief Check whether the overloaded unary floating point function
124 /// corresponding to \a Ty is available.
125 static bool hasUnaryFloatFn(const TargetLibraryInfo *TLI, Type *Ty,
126 LibFunc::Func DoubleFn, LibFunc::Func FloatFn,
127 LibFunc::Func LongDoubleFn) {
128 switch (Ty->getTypeID()) {
129 case Type::FloatTyID:
130 return TLI->has(FloatFn);
131 case Type::DoubleTyID:
132 return TLI->has(DoubleFn);
134 return TLI->has(LongDoubleFn);
138 //===----------------------------------------------------------------------===//
139 // String and Memory Library Call Optimizations
140 //===----------------------------------------------------------------------===//
142 Value *LibCallSimplifier::optimizeStrCat(CallInst *CI, IRBuilder<> &B) {
143 // Extract some information from the instruction
144 Value *Dst = CI->getArgOperand(0);
145 Value *Src = CI->getArgOperand(1);
147 // See if we can get the length of the input string.
148 uint64_t Len = GetStringLength(Src);
151 --Len; // Unbias length.
153 // Handle the simple, do-nothing case: strcat(x, "") -> x
157 return emitStrLenMemCpy(Src, Dst, Len, B);
160 Value *LibCallSimplifier::emitStrLenMemCpy(Value *Src, Value *Dst, uint64_t Len,
162 // We need to find the end of the destination string. That's where the
163 // memory is to be moved to. We just generate a call to strlen.
164 Value *DstLen = emitStrLen(Dst, B, DL, TLI);
168 // Now that we have the destination's length, we must index into the
169 // destination's pointer to get the actual memcpy destination (end of
170 // the string .. we're concatenating).
171 Value *CpyDst = B.CreateGEP(B.getInt8Ty(), Dst, DstLen, "endptr");
173 // We have enough information to now generate the memcpy call to do the
174 // concatenation for us. Make a memcpy to copy the nul byte with align = 1.
175 B.CreateMemCpy(CpyDst, Src,
176 ConstantInt::get(DL.getIntPtrType(Src->getContext()), Len + 1),
181 Value *LibCallSimplifier::optimizeStrNCat(CallInst *CI, IRBuilder<> &B) {
182 // Extract some information from the instruction.
183 Value *Dst = CI->getArgOperand(0);
184 Value *Src = CI->getArgOperand(1);
187 // We don't do anything if length is not constant.
188 if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(CI->getArgOperand(2)))
189 Len = LengthArg->getZExtValue();
193 // See if we can get the length of the input string.
194 uint64_t SrcLen = GetStringLength(Src);
197 --SrcLen; // Unbias length.
199 // Handle the simple, do-nothing cases:
200 // strncat(x, "", c) -> x
201 // strncat(x, c, 0) -> x
202 if (SrcLen == 0 || Len == 0)
205 // We don't optimize this case.
209 // strncat(x, s, c) -> strcat(x, s)
210 // s is constant so the strcat can be optimized further.
211 return emitStrLenMemCpy(Src, Dst, SrcLen, B);
214 Value *LibCallSimplifier::optimizeStrChr(CallInst *CI, IRBuilder<> &B) {
215 Function *Callee = CI->getCalledFunction();
216 FunctionType *FT = Callee->getFunctionType();
217 Value *SrcStr = CI->getArgOperand(0);
219 // If the second operand is non-constant, see if we can compute the length
220 // of the input string and turn this into memchr.
221 ConstantInt *CharC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
223 uint64_t Len = GetStringLength(SrcStr);
224 if (Len == 0 || !FT->getParamType(1)->isIntegerTy(32)) // memchr needs i32.
227 return emitMemChr(SrcStr, CI->getArgOperand(1), // include nul.
228 ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len),
232 // Otherwise, the character is a constant, see if the first argument is
233 // a string literal. If so, we can constant fold.
235 if (!getConstantStringInfo(SrcStr, Str)) {
236 if (CharC->isZero()) // strchr(p, 0) -> p + strlen(p)
237 return B.CreateGEP(B.getInt8Ty(), SrcStr, emitStrLen(SrcStr, B, DL, TLI),
242 // Compute the offset, make sure to handle the case when we're searching for
243 // zero (a weird way to spell strlen).
244 size_t I = (0xFF & CharC->getSExtValue()) == 0
246 : Str.find(CharC->getSExtValue());
247 if (I == StringRef::npos) // Didn't find the char. strchr returns null.
248 return Constant::getNullValue(CI->getType());
250 // strchr(s+n,c) -> gep(s+n+i,c)
251 return B.CreateGEP(B.getInt8Ty(), SrcStr, B.getInt64(I), "strchr");
254 Value *LibCallSimplifier::optimizeStrRChr(CallInst *CI, IRBuilder<> &B) {
255 Value *SrcStr = CI->getArgOperand(0);
256 ConstantInt *CharC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
258 // Cannot fold anything if we're not looking for a constant.
263 if (!getConstantStringInfo(SrcStr, Str)) {
264 // strrchr(s, 0) -> strchr(s, 0)
266 return emitStrChr(SrcStr, '\0', B, TLI);
270 // Compute the offset.
271 size_t I = (0xFF & CharC->getSExtValue()) == 0
273 : Str.rfind(CharC->getSExtValue());
274 if (I == StringRef::npos) // Didn't find the char. Return null.
275 return Constant::getNullValue(CI->getType());
277 // strrchr(s+n,c) -> gep(s+n+i,c)
278 return B.CreateGEP(B.getInt8Ty(), SrcStr, B.getInt64(I), "strrchr");
281 Value *LibCallSimplifier::optimizeStrCmp(CallInst *CI, IRBuilder<> &B) {
282 Value *Str1P = CI->getArgOperand(0), *Str2P = CI->getArgOperand(1);
283 if (Str1P == Str2P) // strcmp(x,x) -> 0
284 return ConstantInt::get(CI->getType(), 0);
286 StringRef Str1, Str2;
287 bool HasStr1 = getConstantStringInfo(Str1P, Str1);
288 bool HasStr2 = getConstantStringInfo(Str2P, Str2);
290 // strcmp(x, y) -> cnst (if both x and y are constant strings)
291 if (HasStr1 && HasStr2)
292 return ConstantInt::get(CI->getType(), Str1.compare(Str2));
294 if (HasStr1 && Str1.empty()) // strcmp("", x) -> -*x
296 B.CreateZExt(B.CreateLoad(Str2P, "strcmpload"), CI->getType()));
298 if (HasStr2 && Str2.empty()) // strcmp(x,"") -> *x
299 return B.CreateZExt(B.CreateLoad(Str1P, "strcmpload"), CI->getType());
301 // strcmp(P, "x") -> memcmp(P, "x", 2)
302 uint64_t Len1 = GetStringLength(Str1P);
303 uint64_t Len2 = GetStringLength(Str2P);
305 return emitMemCmp(Str1P, Str2P,
306 ConstantInt::get(DL.getIntPtrType(CI->getContext()),
307 std::min(Len1, Len2)),
314 Value *LibCallSimplifier::optimizeStrNCmp(CallInst *CI, IRBuilder<> &B) {
315 Value *Str1P = CI->getArgOperand(0), *Str2P = CI->getArgOperand(1);
316 if (Str1P == Str2P) // strncmp(x,x,n) -> 0
317 return ConstantInt::get(CI->getType(), 0);
319 // Get the length argument if it is constant.
321 if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(CI->getArgOperand(2)))
322 Length = LengthArg->getZExtValue();
326 if (Length == 0) // strncmp(x,y,0) -> 0
327 return ConstantInt::get(CI->getType(), 0);
329 if (Length == 1) // strncmp(x,y,1) -> memcmp(x,y,1)
330 return emitMemCmp(Str1P, Str2P, CI->getArgOperand(2), B, DL, TLI);
332 StringRef Str1, Str2;
333 bool HasStr1 = getConstantStringInfo(Str1P, Str1);
334 bool HasStr2 = getConstantStringInfo(Str2P, Str2);
336 // strncmp(x, y) -> cnst (if both x and y are constant strings)
337 if (HasStr1 && HasStr2) {
338 StringRef SubStr1 = Str1.substr(0, Length);
339 StringRef SubStr2 = Str2.substr(0, Length);
340 return ConstantInt::get(CI->getType(), SubStr1.compare(SubStr2));
343 if (HasStr1 && Str1.empty()) // strncmp("", x, n) -> -*x
345 B.CreateZExt(B.CreateLoad(Str2P, "strcmpload"), CI->getType()));
347 if (HasStr2 && Str2.empty()) // strncmp(x, "", n) -> *x
348 return B.CreateZExt(B.CreateLoad(Str1P, "strcmpload"), CI->getType());
353 Value *LibCallSimplifier::optimizeStrCpy(CallInst *CI, IRBuilder<> &B) {
354 Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1);
355 if (Dst == Src) // strcpy(x,x) -> x
358 // See if we can get the length of the input string.
359 uint64_t Len = GetStringLength(Src);
363 // We have enough information to now generate the memcpy call to do the
364 // copy for us. Make a memcpy to copy the nul byte with align = 1.
365 B.CreateMemCpy(Dst, Src,
366 ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len), 1);
370 Value *LibCallSimplifier::optimizeStpCpy(CallInst *CI, IRBuilder<> &B) {
371 Function *Callee = CI->getCalledFunction();
372 Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1);
373 if (Dst == Src) { // stpcpy(x,x) -> x+strlen(x)
374 Value *StrLen = emitStrLen(Src, B, DL, TLI);
375 return StrLen ? B.CreateInBoundsGEP(B.getInt8Ty(), Dst, StrLen) : nullptr;
378 // See if we can get the length of the input string.
379 uint64_t Len = GetStringLength(Src);
383 Type *PT = Callee->getFunctionType()->getParamType(0);
384 Value *LenV = ConstantInt::get(DL.getIntPtrType(PT), Len);
385 Value *DstEnd = B.CreateGEP(B.getInt8Ty(), Dst,
386 ConstantInt::get(DL.getIntPtrType(PT), Len - 1));
388 // We have enough information to now generate the memcpy call to do the
389 // copy for us. Make a memcpy to copy the nul byte with align = 1.
390 B.CreateMemCpy(Dst, Src, LenV, 1);
394 Value *LibCallSimplifier::optimizeStrNCpy(CallInst *CI, IRBuilder<> &B) {
395 Function *Callee = CI->getCalledFunction();
396 Value *Dst = CI->getArgOperand(0);
397 Value *Src = CI->getArgOperand(1);
398 Value *LenOp = CI->getArgOperand(2);
400 // See if we can get the length of the input string.
401 uint64_t SrcLen = GetStringLength(Src);
407 // strncpy(x, "", y) -> memset(x, '\0', y, 1)
408 B.CreateMemSet(Dst, B.getInt8('\0'), LenOp, 1);
413 if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(LenOp))
414 Len = LengthArg->getZExtValue();
419 return Dst; // strncpy(x, y, 0) -> x
421 // Let strncpy handle the zero padding
422 if (Len > SrcLen + 1)
425 Type *PT = Callee->getFunctionType()->getParamType(0);
426 // strncpy(x, s, c) -> memcpy(x, s, c, 1) [s and c are constant]
427 B.CreateMemCpy(Dst, Src, ConstantInt::get(DL.getIntPtrType(PT), Len), 1);
432 Value *LibCallSimplifier::optimizeStrLen(CallInst *CI, IRBuilder<> &B) {
433 Value *Src = CI->getArgOperand(0);
435 // Constant folding: strlen("xyz") -> 3
436 if (uint64_t Len = GetStringLength(Src))
437 return ConstantInt::get(CI->getType(), Len - 1);
439 // If s is a constant pointer pointing to a string literal, we can fold
440 // strlen(s + x) to strlen(s) - x, when x is known to be in the range
441 // [0, strlen(s)] or the string has a single null terminator '\0' at the end.
442 // We only try to simplify strlen when the pointer s points to an array
443 // of i8. Otherwise, we would need to scale the offset x before doing the
444 // subtraction. This will make the optimization more complex, and it's not
445 // very useful because calling strlen for a pointer of other types is
447 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Src)) {
448 if (!isGEPBasedOnPointerToString(GEP))
452 if (getConstantStringInfo(GEP->getOperand(0), Str, 0, false)) {
453 size_t NullTermIdx = Str.find('\0');
455 // If the string does not have '\0', leave it to strlen to compute
457 if (NullTermIdx == StringRef::npos)
460 Value *Offset = GEP->getOperand(2);
461 unsigned BitWidth = Offset->getType()->getIntegerBitWidth();
462 APInt KnownZero(BitWidth, 0);
463 APInt KnownOne(BitWidth, 0);
464 computeKnownBits(Offset, KnownZero, KnownOne, DL, 0, nullptr, CI,
466 KnownZero.flipAllBits();
468 cast<ArrayType>(GEP->getSourceElementType())->getNumElements();
470 // KnownZero's bits are flipped, so zeros in KnownZero now represent
471 // bits known to be zeros in Offset, and ones in KnowZero represent
472 // bits unknown in Offset. Therefore, Offset is known to be in range
473 // [0, NullTermIdx] when the flipped KnownZero is non-negative and
474 // unsigned-less-than NullTermIdx.
476 // If Offset is not provably in the range [0, NullTermIdx], we can still
477 // optimize if we can prove that the program has undefined behavior when
478 // Offset is outside that range. That is the case when GEP->getOperand(0)
479 // is a pointer to an object whose memory extent is NullTermIdx+1.
480 if ((KnownZero.isNonNegative() && KnownZero.ule(NullTermIdx)) ||
481 (GEP->isInBounds() && isa<GlobalVariable>(GEP->getOperand(0)) &&
482 NullTermIdx == ArrSize - 1))
483 return B.CreateSub(ConstantInt::get(CI->getType(), NullTermIdx),
490 // strlen(x?"foo":"bars") --> x ? 3 : 4
491 if (SelectInst *SI = dyn_cast<SelectInst>(Src)) {
492 uint64_t LenTrue = GetStringLength(SI->getTrueValue());
493 uint64_t LenFalse = GetStringLength(SI->getFalseValue());
494 if (LenTrue && LenFalse) {
495 Function *Caller = CI->getParent()->getParent();
496 emitOptimizationRemark(CI->getContext(), "simplify-libcalls", *Caller,
498 "folded strlen(select) to select of constants");
499 return B.CreateSelect(SI->getCondition(),
500 ConstantInt::get(CI->getType(), LenTrue - 1),
501 ConstantInt::get(CI->getType(), LenFalse - 1));
505 // strlen(x) != 0 --> *x != 0
506 // strlen(x) == 0 --> *x == 0
507 if (isOnlyUsedInZeroEqualityComparison(CI))
508 return B.CreateZExt(B.CreateLoad(Src, "strlenfirst"), CI->getType());
513 Value *LibCallSimplifier::optimizeStrPBrk(CallInst *CI, IRBuilder<> &B) {
515 bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1);
516 bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2);
518 // strpbrk(s, "") -> nullptr
519 // strpbrk("", s) -> nullptr
520 if ((HasS1 && S1.empty()) || (HasS2 && S2.empty()))
521 return Constant::getNullValue(CI->getType());
524 if (HasS1 && HasS2) {
525 size_t I = S1.find_first_of(S2);
526 if (I == StringRef::npos) // No match.
527 return Constant::getNullValue(CI->getType());
529 return B.CreateGEP(B.getInt8Ty(), CI->getArgOperand(0), B.getInt64(I),
533 // strpbrk(s, "a") -> strchr(s, 'a')
534 if (HasS2 && S2.size() == 1)
535 return emitStrChr(CI->getArgOperand(0), S2[0], B, TLI);
540 Value *LibCallSimplifier::optimizeStrTo(CallInst *CI, IRBuilder<> &B) {
541 Value *EndPtr = CI->getArgOperand(1);
542 if (isa<ConstantPointerNull>(EndPtr)) {
543 // With a null EndPtr, this function won't capture the main argument.
544 // It would be readonly too, except that it still may write to errno.
545 CI->addAttribute(1, Attribute::NoCapture);
551 Value *LibCallSimplifier::optimizeStrSpn(CallInst *CI, IRBuilder<> &B) {
553 bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1);
554 bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2);
556 // strspn(s, "") -> 0
557 // strspn("", s) -> 0
558 if ((HasS1 && S1.empty()) || (HasS2 && S2.empty()))
559 return Constant::getNullValue(CI->getType());
562 if (HasS1 && HasS2) {
563 size_t Pos = S1.find_first_not_of(S2);
564 if (Pos == StringRef::npos)
566 return ConstantInt::get(CI->getType(), Pos);
572 Value *LibCallSimplifier::optimizeStrCSpn(CallInst *CI, IRBuilder<> &B) {
574 bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1);
575 bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2);
577 // strcspn("", s) -> 0
578 if (HasS1 && S1.empty())
579 return Constant::getNullValue(CI->getType());
582 if (HasS1 && HasS2) {
583 size_t Pos = S1.find_first_of(S2);
584 if (Pos == StringRef::npos)
586 return ConstantInt::get(CI->getType(), Pos);
589 // strcspn(s, "") -> strlen(s)
590 if (HasS2 && S2.empty())
591 return emitStrLen(CI->getArgOperand(0), B, DL, TLI);
596 Value *LibCallSimplifier::optimizeStrStr(CallInst *CI, IRBuilder<> &B) {
597 // fold strstr(x, x) -> x.
598 if (CI->getArgOperand(0) == CI->getArgOperand(1))
599 return B.CreateBitCast(CI->getArgOperand(0), CI->getType());
601 // fold strstr(a, b) == a -> strncmp(a, b, strlen(b)) == 0
602 if (isOnlyUsedInEqualityComparison(CI, CI->getArgOperand(0))) {
603 Value *StrLen = emitStrLen(CI->getArgOperand(1), B, DL, TLI);
606 Value *StrNCmp = emitStrNCmp(CI->getArgOperand(0), CI->getArgOperand(1),
610 for (auto UI = CI->user_begin(), UE = CI->user_end(); UI != UE;) {
611 ICmpInst *Old = cast<ICmpInst>(*UI++);
613 B.CreateICmp(Old->getPredicate(), StrNCmp,
614 ConstantInt::getNullValue(StrNCmp->getType()), "cmp");
615 replaceAllUsesWith(Old, Cmp);
620 // See if either input string is a constant string.
621 StringRef SearchStr, ToFindStr;
622 bool HasStr1 = getConstantStringInfo(CI->getArgOperand(0), SearchStr);
623 bool HasStr2 = getConstantStringInfo(CI->getArgOperand(1), ToFindStr);
625 // fold strstr(x, "") -> x.
626 if (HasStr2 && ToFindStr.empty())
627 return B.CreateBitCast(CI->getArgOperand(0), CI->getType());
629 // If both strings are known, constant fold it.
630 if (HasStr1 && HasStr2) {
631 size_t Offset = SearchStr.find(ToFindStr);
633 if (Offset == StringRef::npos) // strstr("foo", "bar") -> null
634 return Constant::getNullValue(CI->getType());
636 // strstr("abcd", "bc") -> gep((char*)"abcd", 1)
637 Value *Result = castToCStr(CI->getArgOperand(0), B);
638 Result = B.CreateConstInBoundsGEP1_64(Result, Offset, "strstr");
639 return B.CreateBitCast(Result, CI->getType());
642 // fold strstr(x, "y") -> strchr(x, 'y').
643 if (HasStr2 && ToFindStr.size() == 1) {
644 Value *StrChr = emitStrChr(CI->getArgOperand(0), ToFindStr[0], B, TLI);
645 return StrChr ? B.CreateBitCast(StrChr, CI->getType()) : nullptr;
650 Value *LibCallSimplifier::optimizeMemChr(CallInst *CI, IRBuilder<> &B) {
651 Value *SrcStr = CI->getArgOperand(0);
652 ConstantInt *CharC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
653 ConstantInt *LenC = dyn_cast<ConstantInt>(CI->getArgOperand(2));
655 // memchr(x, y, 0) -> null
656 if (LenC && LenC->isNullValue())
657 return Constant::getNullValue(CI->getType());
659 // From now on we need at least constant length and string.
661 if (!LenC || !getConstantStringInfo(SrcStr, Str, 0, /*TrimAtNul=*/false))
664 // Truncate the string to LenC. If Str is smaller than LenC we will still only
665 // scan the string, as reading past the end of it is undefined and we can just
666 // return null if we don't find the char.
667 Str = Str.substr(0, LenC->getZExtValue());
669 // If the char is variable but the input str and length are not we can turn
670 // this memchr call into a simple bit field test. Of course this only works
671 // when the return value is only checked against null.
673 // It would be really nice to reuse switch lowering here but we can't change
674 // the CFG at this point.
676 // memchr("\r\n", C, 2) != nullptr -> (C & ((1 << '\r') | (1 << '\n'))) != 0
677 // after bounds check.
678 if (!CharC && !Str.empty() && isOnlyUsedInZeroEqualityComparison(CI)) {
680 *std::max_element(reinterpret_cast<const unsigned char *>(Str.begin()),
681 reinterpret_cast<const unsigned char *>(Str.end()));
683 // Make sure the bit field we're about to create fits in a register on the
685 // FIXME: On a 64 bit architecture this prevents us from using the
686 // interesting range of alpha ascii chars. We could do better by emitting
687 // two bitfields or shifting the range by 64 if no lower chars are used.
688 if (!DL.fitsInLegalInteger(Max + 1))
691 // For the bit field use a power-of-2 type with at least 8 bits to avoid
692 // creating unnecessary illegal types.
693 unsigned char Width = NextPowerOf2(std::max((unsigned char)7, Max));
695 // Now build the bit field.
696 APInt Bitfield(Width, 0);
698 Bitfield.setBit((unsigned char)C);
699 Value *BitfieldC = B.getInt(Bitfield);
701 // First check that the bit field access is within bounds.
702 Value *C = B.CreateZExtOrTrunc(CI->getArgOperand(1), BitfieldC->getType());
703 Value *Bounds = B.CreateICmp(ICmpInst::ICMP_ULT, C, B.getIntN(Width, Width),
706 // Create code that checks if the given bit is set in the field.
707 Value *Shl = B.CreateShl(B.getIntN(Width, 1ULL), C);
708 Value *Bits = B.CreateIsNotNull(B.CreateAnd(Shl, BitfieldC), "memchr.bits");
710 // Finally merge both checks and cast to pointer type. The inttoptr
711 // implicitly zexts the i1 to intptr type.
712 return B.CreateIntToPtr(B.CreateAnd(Bounds, Bits, "memchr"), CI->getType());
715 // Check if all arguments are constants. If so, we can constant fold.
719 // Compute the offset.
720 size_t I = Str.find(CharC->getSExtValue() & 0xFF);
721 if (I == StringRef::npos) // Didn't find the char. memchr returns null.
722 return Constant::getNullValue(CI->getType());
724 // memchr(s+n,c,l) -> gep(s+n+i,c)
725 return B.CreateGEP(B.getInt8Ty(), SrcStr, B.getInt64(I), "memchr");
728 Value *LibCallSimplifier::optimizeMemCmp(CallInst *CI, IRBuilder<> &B) {
729 Value *LHS = CI->getArgOperand(0), *RHS = CI->getArgOperand(1);
731 if (LHS == RHS) // memcmp(s,s,x) -> 0
732 return Constant::getNullValue(CI->getType());
734 // Make sure we have a constant length.
735 ConstantInt *LenC = dyn_cast<ConstantInt>(CI->getArgOperand(2));
738 uint64_t Len = LenC->getZExtValue();
740 if (Len == 0) // memcmp(s1,s2,0) -> 0
741 return Constant::getNullValue(CI->getType());
743 // memcmp(S1,S2,1) -> *(unsigned char*)LHS - *(unsigned char*)RHS
745 Value *LHSV = B.CreateZExt(B.CreateLoad(castToCStr(LHS, B), "lhsc"),
746 CI->getType(), "lhsv");
747 Value *RHSV = B.CreateZExt(B.CreateLoad(castToCStr(RHS, B), "rhsc"),
748 CI->getType(), "rhsv");
749 return B.CreateSub(LHSV, RHSV, "chardiff");
752 // memcmp(S1,S2,N/8)==0 -> (*(intN_t*)S1 != *(intN_t*)S2)==0
753 if (DL.isLegalInteger(Len * 8) && isOnlyUsedInZeroEqualityComparison(CI)) {
755 IntegerType *IntType = IntegerType::get(CI->getContext(), Len * 8);
756 unsigned PrefAlignment = DL.getPrefTypeAlignment(IntType);
758 if (getKnownAlignment(LHS, DL, CI) >= PrefAlignment &&
759 getKnownAlignment(RHS, DL, CI) >= PrefAlignment) {
762 IntType->getPointerTo(LHS->getType()->getPointerAddressSpace());
764 IntType->getPointerTo(RHS->getType()->getPointerAddressSpace());
767 B.CreateLoad(B.CreateBitCast(LHS, LHSPtrTy, "lhsc"), "lhsv");
769 B.CreateLoad(B.CreateBitCast(RHS, RHSPtrTy, "rhsc"), "rhsv");
771 return B.CreateZExt(B.CreateICmpNE(LHSV, RHSV), CI->getType(), "memcmp");
775 // Constant folding: memcmp(x, y, l) -> cnst (all arguments are constant)
776 StringRef LHSStr, RHSStr;
777 if (getConstantStringInfo(LHS, LHSStr) &&
778 getConstantStringInfo(RHS, RHSStr)) {
779 // Make sure we're not reading out-of-bounds memory.
780 if (Len > LHSStr.size() || Len > RHSStr.size())
782 // Fold the memcmp and normalize the result. This way we get consistent
783 // results across multiple platforms.
785 int Cmp = memcmp(LHSStr.data(), RHSStr.data(), Len);
790 return ConstantInt::get(CI->getType(), Ret);
796 Value *LibCallSimplifier::optimizeMemCpy(CallInst *CI, IRBuilder<> &B) {
797 // memcpy(x, y, n) -> llvm.memcpy(x, y, n, 1)
798 B.CreateMemCpy(CI->getArgOperand(0), CI->getArgOperand(1),
799 CI->getArgOperand(2), 1);
800 return CI->getArgOperand(0);
803 Value *LibCallSimplifier::optimizeMemMove(CallInst *CI, IRBuilder<> &B) {
804 // memmove(x, y, n) -> llvm.memmove(x, y, n, 1)
805 B.CreateMemMove(CI->getArgOperand(0), CI->getArgOperand(1),
806 CI->getArgOperand(2), 1);
807 return CI->getArgOperand(0);
810 // TODO: Does this belong in BuildLibCalls or should all of those similar
811 // functions be moved here?
812 static Value *emitCalloc(Value *Num, Value *Size, const AttributeSet &Attrs,
813 IRBuilder<> &B, const TargetLibraryInfo &TLI) {
815 if (!TLI.getLibFunc("calloc", Func) || !TLI.has(Func))
818 Module *M = B.GetInsertBlock()->getModule();
819 const DataLayout &DL = M->getDataLayout();
820 IntegerType *PtrType = DL.getIntPtrType((B.GetInsertBlock()->getContext()));
821 Value *Calloc = M->getOrInsertFunction("calloc", Attrs, B.getInt8PtrTy(),
822 PtrType, PtrType, nullptr);
823 CallInst *CI = B.CreateCall(Calloc, { Num, Size }, "calloc");
825 if (const auto *F = dyn_cast<Function>(Calloc->stripPointerCasts()))
826 CI->setCallingConv(F->getCallingConv());
831 /// Fold memset[_chk](malloc(n), 0, n) --> calloc(1, n).
832 static Value *foldMallocMemset(CallInst *Memset, IRBuilder<> &B,
833 const TargetLibraryInfo &TLI) {
834 // This has to be a memset of zeros (bzero).
835 auto *FillValue = dyn_cast<ConstantInt>(Memset->getArgOperand(1));
836 if (!FillValue || FillValue->getZExtValue() != 0)
839 // TODO: We should handle the case where the malloc has more than one use.
840 // This is necessary to optimize common patterns such as when the result of
841 // the malloc is checked against null or when a memset intrinsic is used in
842 // place of a memset library call.
843 auto *Malloc = dyn_cast<CallInst>(Memset->getArgOperand(0));
844 if (!Malloc || !Malloc->hasOneUse())
847 // Is the inner call really malloc()?
848 Function *InnerCallee = Malloc->getCalledFunction();
850 if (!TLI.getLibFunc(*InnerCallee, Func) || !TLI.has(Func) ||
851 Func != LibFunc::malloc)
854 // The memset must cover the same number of bytes that are malloc'd.
855 if (Memset->getArgOperand(2) != Malloc->getArgOperand(0))
858 // Replace the malloc with a calloc. We need the data layout to know what the
859 // actual size of a 'size_t' parameter is.
860 B.SetInsertPoint(Malloc->getParent(), ++Malloc->getIterator());
861 const DataLayout &DL = Malloc->getModule()->getDataLayout();
862 IntegerType *SizeType = DL.getIntPtrType(B.GetInsertBlock()->getContext());
863 Value *Calloc = emitCalloc(ConstantInt::get(SizeType, 1),
864 Malloc->getArgOperand(0), Malloc->getAttributes(),
869 Malloc->replaceAllUsesWith(Calloc);
870 Malloc->eraseFromParent();
875 Value *LibCallSimplifier::optimizeMemSet(CallInst *CI, IRBuilder<> &B) {
876 if (auto *Calloc = foldMallocMemset(CI, B, *TLI))
879 // memset(p, v, n) -> llvm.memset(p, v, n, 1)
880 Value *Val = B.CreateIntCast(CI->getArgOperand(1), B.getInt8Ty(), false);
881 B.CreateMemSet(CI->getArgOperand(0), Val, CI->getArgOperand(2), 1);
882 return CI->getArgOperand(0);
885 //===----------------------------------------------------------------------===//
886 // Math Library Optimizations
887 //===----------------------------------------------------------------------===//
889 /// Return a variant of Val with float type.
890 /// Currently this works in two cases: If Val is an FPExtension of a float
891 /// value to something bigger, simply return the operand.
892 /// If Val is a ConstantFP but can be converted to a float ConstantFP without
893 /// loss of precision do so.
894 static Value *valueHasFloatPrecision(Value *Val) {
895 if (FPExtInst *Cast = dyn_cast<FPExtInst>(Val)) {
896 Value *Op = Cast->getOperand(0);
897 if (Op->getType()->isFloatTy())
900 if (ConstantFP *Const = dyn_cast<ConstantFP>(Val)) {
901 APFloat F = Const->getValueAPF();
903 (void)F.convert(APFloat::IEEEsingle(), APFloat::rmNearestTiesToEven,
906 return ConstantFP::get(Const->getContext(), F);
911 /// Shrink double -> float for unary functions like 'floor'.
912 static Value *optimizeUnaryDoubleFP(CallInst *CI, IRBuilder<> &B,
914 Function *Callee = CI->getCalledFunction();
915 // We know this libcall has a valid prototype, but we don't know which.
916 if (!CI->getType()->isDoubleTy())
920 // Check if all the uses for function like 'sin' are converted to float.
921 for (User *U : CI->users()) {
922 FPTruncInst *Cast = dyn_cast<FPTruncInst>(U);
923 if (!Cast || !Cast->getType()->isFloatTy())
928 // If this is something like 'floor((double)floatval)', convert to floorf.
929 Value *V = valueHasFloatPrecision(CI->getArgOperand(0));
933 // Propagate fast-math flags from the existing call to the new call.
934 IRBuilder<>::FastMathFlagGuard Guard(B);
935 B.setFastMathFlags(CI->getFastMathFlags());
937 // floor((double)floatval) -> (double)floorf(floatval)
938 if (Callee->isIntrinsic()) {
939 Module *M = CI->getModule();
940 Intrinsic::ID IID = Callee->getIntrinsicID();
941 Function *F = Intrinsic::getDeclaration(M, IID, B.getFloatTy());
942 V = B.CreateCall(F, V);
944 // The call is a library call rather than an intrinsic.
945 V = emitUnaryFloatFnCall(V, Callee->getName(), B, Callee->getAttributes());
948 return B.CreateFPExt(V, B.getDoubleTy());
951 /// Shrink double -> float for binary functions like 'fmin/fmax'.
952 static Value *optimizeBinaryDoubleFP(CallInst *CI, IRBuilder<> &B) {
953 Function *Callee = CI->getCalledFunction();
954 // We know this libcall has a valid prototype, but we don't know which.
955 if (!CI->getType()->isDoubleTy())
958 // If this is something like 'fmin((double)floatval1, (double)floatval2)',
959 // or fmin(1.0, (double)floatval), then we convert it to fminf.
960 Value *V1 = valueHasFloatPrecision(CI->getArgOperand(0));
963 Value *V2 = valueHasFloatPrecision(CI->getArgOperand(1));
967 // Propagate fast-math flags from the existing call to the new call.
968 IRBuilder<>::FastMathFlagGuard Guard(B);
969 B.setFastMathFlags(CI->getFastMathFlags());
971 // fmin((double)floatval1, (double)floatval2)
972 // -> (double)fminf(floatval1, floatval2)
973 // TODO: Handle intrinsics in the same way as in optimizeUnaryDoubleFP().
974 Value *V = emitBinaryFloatFnCall(V1, V2, Callee->getName(), B,
975 Callee->getAttributes());
976 return B.CreateFPExt(V, B.getDoubleTy());
979 Value *LibCallSimplifier::optimizeCos(CallInst *CI, IRBuilder<> &B) {
980 Function *Callee = CI->getCalledFunction();
981 Value *Ret = nullptr;
982 StringRef Name = Callee->getName();
983 if (UnsafeFPShrink && Name == "cos" && hasFloatVersion(Name))
984 Ret = optimizeUnaryDoubleFP(CI, B, true);
987 Value *Op1 = CI->getArgOperand(0);
988 if (BinaryOperator::isFNeg(Op1)) {
989 BinaryOperator *BinExpr = cast<BinaryOperator>(Op1);
990 return B.CreateCall(Callee, BinExpr->getOperand(1), "cos");
995 static Value *getPow(Value *InnerChain[33], unsigned Exp, IRBuilder<> &B) {
996 // Multiplications calculated using Addition Chains.
997 // Refer: http://wwwhomes.uni-bielefeld.de/achim/addition_chain.html
999 assert(Exp != 0 && "Incorrect exponent 0 not handled");
1001 if (InnerChain[Exp])
1002 return InnerChain[Exp];
1004 static const unsigned AddChain[33][2] = {
1006 {0, 0}, // Unused (base case = pow1).
1007 {1, 1}, // Unused (pre-computed).
1008 {1, 2}, {2, 2}, {2, 3}, {3, 3}, {2, 5}, {4, 4},
1009 {1, 8}, {5, 5}, {1, 10}, {6, 6}, {4, 9}, {7, 7},
1010 {3, 12}, {8, 8}, {8, 9}, {2, 16}, {1, 18}, {10, 10},
1011 {6, 15}, {11, 11}, {3, 20}, {12, 12}, {8, 17}, {13, 13},
1012 {3, 24}, {14, 14}, {4, 25}, {15, 15}, {3, 28}, {16, 16},
1015 InnerChain[Exp] = B.CreateFMul(getPow(InnerChain, AddChain[Exp][0], B),
1016 getPow(InnerChain, AddChain[Exp][1], B));
1017 return InnerChain[Exp];
1020 Value *LibCallSimplifier::optimizePow(CallInst *CI, IRBuilder<> &B) {
1021 Function *Callee = CI->getCalledFunction();
1022 Value *Ret = nullptr;
1023 StringRef Name = Callee->getName();
1024 if (UnsafeFPShrink && Name == "pow" && hasFloatVersion(Name))
1025 Ret = optimizeUnaryDoubleFP(CI, B, true);
1027 Value *Op1 = CI->getArgOperand(0), *Op2 = CI->getArgOperand(1);
1029 // pow(1.0, x) -> 1.0
1030 if (match(Op1, m_SpecificFP(1.0)))
1032 // pow(2.0, x) -> llvm.exp2(x)
1033 if (match(Op1, m_SpecificFP(2.0))) {
1034 Value *Exp2 = Intrinsic::getDeclaration(CI->getModule(), Intrinsic::exp2,
1036 return B.CreateCall(Exp2, Op2, "exp2");
1039 // There's no llvm.exp10 intrinsic yet, but, maybe, some day there will
1041 if (ConstantFP *Op1C = dyn_cast<ConstantFP>(Op1)) {
1042 // pow(10.0, x) -> exp10(x)
1043 if (Op1C->isExactlyValue(10.0) &&
1044 hasUnaryFloatFn(TLI, Op1->getType(), LibFunc::exp10, LibFunc::exp10f,
1046 return emitUnaryFloatFnCall(Op2, TLI->getName(LibFunc::exp10), B,
1047 Callee->getAttributes());
1050 // pow(exp(x), y) -> exp(x * y)
1051 // pow(exp2(x), y) -> exp2(x * y)
1052 // We enable these only with fast-math. Besides rounding differences, the
1053 // transformation changes overflow and underflow behavior quite dramatically.
1054 // Example: x = 1000, y = 0.001.
1055 // pow(exp(x), y) = pow(inf, 0.001) = inf, whereas exp(x*y) = exp(1).
1056 auto *OpC = dyn_cast<CallInst>(Op1);
1057 if (OpC && OpC->hasUnsafeAlgebra() && CI->hasUnsafeAlgebra()) {
1059 Function *OpCCallee = OpC->getCalledFunction();
1060 if (OpCCallee && TLI->getLibFunc(OpCCallee->getName(), Func) &&
1061 TLI->has(Func) && (Func == LibFunc::exp || Func == LibFunc::exp2)) {
1062 IRBuilder<>::FastMathFlagGuard Guard(B);
1063 B.setFastMathFlags(CI->getFastMathFlags());
1064 Value *FMul = B.CreateFMul(OpC->getArgOperand(0), Op2, "mul");
1065 return emitUnaryFloatFnCall(FMul, OpCCallee->getName(), B,
1066 OpCCallee->getAttributes());
1070 ConstantFP *Op2C = dyn_cast<ConstantFP>(Op2);
1074 if (Op2C->getValueAPF().isZero()) // pow(x, 0.0) -> 1.0
1075 return ConstantFP::get(CI->getType(), 1.0);
1077 if (Op2C->isExactlyValue(0.5) &&
1078 hasUnaryFloatFn(TLI, Op2->getType(), LibFunc::sqrt, LibFunc::sqrtf,
1080 hasUnaryFloatFn(TLI, Op2->getType(), LibFunc::fabs, LibFunc::fabsf,
1083 // In -ffast-math, pow(x, 0.5) -> sqrt(x).
1084 if (CI->hasUnsafeAlgebra()) {
1085 IRBuilder<>::FastMathFlagGuard Guard(B);
1086 B.setFastMathFlags(CI->getFastMathFlags());
1088 // Unlike other math intrinsics, sqrt has differerent semantics
1089 // from the libc function. See LangRef for details.
1090 return emitUnaryFloatFnCall(Op1, TLI->getName(LibFunc::sqrt), B,
1091 Callee->getAttributes());
1094 // Expand pow(x, 0.5) to (x == -infinity ? +infinity : fabs(sqrt(x))).
1095 // This is faster than calling pow, and still handles negative zero
1096 // and negative infinity correctly.
1097 // TODO: In finite-only mode, this could be just fabs(sqrt(x)).
1098 Value *Inf = ConstantFP::getInfinity(CI->getType());
1099 Value *NegInf = ConstantFP::getInfinity(CI->getType(), true);
1100 Value *Sqrt = emitUnaryFloatFnCall(Op1, "sqrt", B, Callee->getAttributes());
1102 emitUnaryFloatFnCall(Sqrt, "fabs", B, Callee->getAttributes());
1103 Value *FCmp = B.CreateFCmpOEQ(Op1, NegInf);
1104 Value *Sel = B.CreateSelect(FCmp, Inf, FAbs);
1108 if (Op2C->isExactlyValue(1.0)) // pow(x, 1.0) -> x
1110 if (Op2C->isExactlyValue(2.0)) // pow(x, 2.0) -> x*x
1111 return B.CreateFMul(Op1, Op1, "pow2");
1112 if (Op2C->isExactlyValue(-1.0)) // pow(x, -1.0) -> 1.0/x
1113 return B.CreateFDiv(ConstantFP::get(CI->getType(), 1.0), Op1, "powrecip");
1115 // In -ffast-math, generate repeated fmul instead of generating pow(x, n).
1116 if (CI->hasUnsafeAlgebra()) {
1117 APFloat V = abs(Op2C->getValueAPF());
1118 // We limit to a max of 7 fmul(s). Thus max exponent is 32.
1119 // This transformation applies to integer exponents only.
1120 if (V.compare(APFloat(V.getSemantics(), 32.0)) == APFloat::cmpGreaterThan ||
1124 // We will memoize intermediate products of the Addition Chain.
1125 Value *InnerChain[33] = {nullptr};
1126 InnerChain[1] = Op1;
1127 InnerChain[2] = B.CreateFMul(Op1, Op1);
1129 // We cannot readily convert a non-double type (like float) to a double.
1130 // So we first convert V to something which could be converted to double.
1132 V.convert(APFloat::IEEEdouble(), APFloat::rmTowardZero, &ignored);
1134 // TODO: Should the new instructions propagate the 'fast' flag of the pow()?
1135 Value *FMul = getPow(InnerChain, V.convertToDouble(), B);
1136 // For negative exponents simply compute the reciprocal.
1137 if (Op2C->isNegative())
1138 FMul = B.CreateFDiv(ConstantFP::get(CI->getType(), 1.0), FMul);
1145 Value *LibCallSimplifier::optimizeExp2(CallInst *CI, IRBuilder<> &B) {
1146 Function *Callee = CI->getCalledFunction();
1147 Value *Ret = nullptr;
1148 StringRef Name = Callee->getName();
1149 if (UnsafeFPShrink && Name == "exp2" && hasFloatVersion(Name))
1150 Ret = optimizeUnaryDoubleFP(CI, B, true);
1152 Value *Op = CI->getArgOperand(0);
1153 // Turn exp2(sitofp(x)) -> ldexp(1.0, sext(x)) if sizeof(x) <= 32
1154 // Turn exp2(uitofp(x)) -> ldexp(1.0, zext(x)) if sizeof(x) < 32
1155 LibFunc::Func LdExp = LibFunc::ldexpl;
1156 if (Op->getType()->isFloatTy())
1157 LdExp = LibFunc::ldexpf;
1158 else if (Op->getType()->isDoubleTy())
1159 LdExp = LibFunc::ldexp;
1161 if (TLI->has(LdExp)) {
1162 Value *LdExpArg = nullptr;
1163 if (SIToFPInst *OpC = dyn_cast<SIToFPInst>(Op)) {
1164 if (OpC->getOperand(0)->getType()->getPrimitiveSizeInBits() <= 32)
1165 LdExpArg = B.CreateSExt(OpC->getOperand(0), B.getInt32Ty());
1166 } else if (UIToFPInst *OpC = dyn_cast<UIToFPInst>(Op)) {
1167 if (OpC->getOperand(0)->getType()->getPrimitiveSizeInBits() < 32)
1168 LdExpArg = B.CreateZExt(OpC->getOperand(0), B.getInt32Ty());
1172 Constant *One = ConstantFP::get(CI->getContext(), APFloat(1.0f));
1173 if (!Op->getType()->isFloatTy())
1174 One = ConstantExpr::getFPExtend(One, Op->getType());
1176 Module *M = CI->getModule();
1178 M->getOrInsertFunction(TLI->getName(LdExp), Op->getType(),
1179 Op->getType(), B.getInt32Ty(), nullptr);
1180 CallInst *CI = B.CreateCall(NewCallee, {One, LdExpArg});
1181 if (const Function *F = dyn_cast<Function>(Callee->stripPointerCasts()))
1182 CI->setCallingConv(F->getCallingConv());
1190 Value *LibCallSimplifier::optimizeFabs(CallInst *CI, IRBuilder<> &B) {
1191 Function *Callee = CI->getCalledFunction();
1192 Value *Ret = nullptr;
1193 StringRef Name = Callee->getName();
1194 if (Name == "fabs" && hasFloatVersion(Name))
1195 Ret = optimizeUnaryDoubleFP(CI, B, false);
1197 Value *Op = CI->getArgOperand(0);
1198 if (Instruction *I = dyn_cast<Instruction>(Op)) {
1199 // Fold fabs(x * x) -> x * x; any squared FP value must already be positive.
1200 if (I->getOpcode() == Instruction::FMul)
1201 if (I->getOperand(0) == I->getOperand(1))
1207 Value *LibCallSimplifier::optimizeFMinFMax(CallInst *CI, IRBuilder<> &B) {
1208 Function *Callee = CI->getCalledFunction();
1209 // If we can shrink the call to a float function rather than a double
1210 // function, do that first.
1211 StringRef Name = Callee->getName();
1212 if ((Name == "fmin" || Name == "fmax") && hasFloatVersion(Name))
1213 if (Value *Ret = optimizeBinaryDoubleFP(CI, B))
1216 IRBuilder<>::FastMathFlagGuard Guard(B);
1218 if (CI->hasUnsafeAlgebra()) {
1219 // Unsafe algebra sets all fast-math-flags to true.
1220 FMF.setUnsafeAlgebra();
1222 // At a minimum, no-nans-fp-math must be true.
1223 if (!CI->hasNoNaNs())
1225 // No-signed-zeros is implied by the definitions of fmax/fmin themselves:
1226 // "Ideally, fmax would be sensitive to the sign of zero, for example
1227 // fmax(-0. 0, +0. 0) would return +0; however, implementation in software
1228 // might be impractical."
1229 FMF.setNoSignedZeros();
1232 B.setFastMathFlags(FMF);
1234 // We have a relaxed floating-point environment. We can ignore NaN-handling
1235 // and transform to a compare and select. We do not have to consider errno or
1236 // exceptions, because fmin/fmax do not have those.
1237 Value *Op0 = CI->getArgOperand(0);
1238 Value *Op1 = CI->getArgOperand(1);
1239 Value *Cmp = Callee->getName().startswith("fmin") ?
1240 B.CreateFCmpOLT(Op0, Op1) : B.CreateFCmpOGT(Op0, Op1);
1241 return B.CreateSelect(Cmp, Op0, Op1);
1244 Value *LibCallSimplifier::optimizeLog(CallInst *CI, IRBuilder<> &B) {
1245 Function *Callee = CI->getCalledFunction();
1246 Value *Ret = nullptr;
1247 StringRef Name = Callee->getName();
1248 if (UnsafeFPShrink && hasFloatVersion(Name))
1249 Ret = optimizeUnaryDoubleFP(CI, B, true);
1251 if (!CI->hasUnsafeAlgebra())
1253 Value *Op1 = CI->getArgOperand(0);
1254 auto *OpC = dyn_cast<CallInst>(Op1);
1256 // The earlier call must also be unsafe in order to do these transforms.
1257 if (!OpC || !OpC->hasUnsafeAlgebra())
1260 // log(pow(x,y)) -> y*log(x)
1261 // This is only applicable to log, log2, log10.
1262 if (Name != "log" && Name != "log2" && Name != "log10")
1265 IRBuilder<>::FastMathFlagGuard Guard(B);
1267 FMF.setUnsafeAlgebra();
1268 B.setFastMathFlags(FMF);
1271 Function *F = OpC->getCalledFunction();
1272 if (F && ((TLI->getLibFunc(F->getName(), Func) && TLI->has(Func) &&
1273 Func == LibFunc::pow) || F->getIntrinsicID() == Intrinsic::pow))
1274 return B.CreateFMul(OpC->getArgOperand(1),
1275 emitUnaryFloatFnCall(OpC->getOperand(0), Callee->getName(), B,
1276 Callee->getAttributes()), "mul");
1278 // log(exp2(y)) -> y*log(2)
1279 if (F && Name == "log" && TLI->getLibFunc(F->getName(), Func) &&
1280 TLI->has(Func) && Func == LibFunc::exp2)
1281 return B.CreateFMul(
1282 OpC->getArgOperand(0),
1283 emitUnaryFloatFnCall(ConstantFP::get(CI->getType(), 2.0),
1284 Callee->getName(), B, Callee->getAttributes()),
1289 Value *LibCallSimplifier::optimizeSqrt(CallInst *CI, IRBuilder<> &B) {
1290 Function *Callee = CI->getCalledFunction();
1291 Value *Ret = nullptr;
1292 if (TLI->has(LibFunc::sqrtf) && (Callee->getName() == "sqrt" ||
1293 Callee->getIntrinsicID() == Intrinsic::sqrt))
1294 Ret = optimizeUnaryDoubleFP(CI, B, true);
1296 if (!CI->hasUnsafeAlgebra())
1299 Instruction *I = dyn_cast<Instruction>(CI->getArgOperand(0));
1300 if (!I || I->getOpcode() != Instruction::FMul || !I->hasUnsafeAlgebra())
1303 // We're looking for a repeated factor in a multiplication tree,
1304 // so we can do this fold: sqrt(x * x) -> fabs(x);
1305 // or this fold: sqrt((x * x) * y) -> fabs(x) * sqrt(y).
1306 Value *Op0 = I->getOperand(0);
1307 Value *Op1 = I->getOperand(1);
1308 Value *RepeatOp = nullptr;
1309 Value *OtherOp = nullptr;
1311 // Simple match: the operands of the multiply are identical.
1314 // Look for a more complicated pattern: one of the operands is itself
1315 // a multiply, so search for a common factor in that multiply.
1316 // Note: We don't bother looking any deeper than this first level or for
1317 // variations of this pattern because instcombine's visitFMUL and/or the
1318 // reassociation pass should give us this form.
1319 Value *OtherMul0, *OtherMul1;
1320 if (match(Op0, m_FMul(m_Value(OtherMul0), m_Value(OtherMul1)))) {
1321 // Pattern: sqrt((x * y) * z)
1322 if (OtherMul0 == OtherMul1 &&
1323 cast<Instruction>(Op0)->hasUnsafeAlgebra()) {
1324 // Matched: sqrt((x * x) * z)
1325 RepeatOp = OtherMul0;
1333 // Fast math flags for any created instructions should match the sqrt
1335 IRBuilder<>::FastMathFlagGuard Guard(B);
1336 B.setFastMathFlags(I->getFastMathFlags());
1338 // If we found a repeated factor, hoist it out of the square root and
1339 // replace it with the fabs of that factor.
1340 Module *M = Callee->getParent();
1341 Type *ArgType = I->getType();
1342 Value *Fabs = Intrinsic::getDeclaration(M, Intrinsic::fabs, ArgType);
1343 Value *FabsCall = B.CreateCall(Fabs, RepeatOp, "fabs");
1345 // If we found a non-repeated factor, we still need to get its square
1346 // root. We then multiply that by the value that was simplified out
1347 // of the square root calculation.
1348 Value *Sqrt = Intrinsic::getDeclaration(M, Intrinsic::sqrt, ArgType);
1349 Value *SqrtCall = B.CreateCall(Sqrt, OtherOp, "sqrt");
1350 return B.CreateFMul(FabsCall, SqrtCall);
1355 // TODO: Generalize to handle any trig function and its inverse.
1356 Value *LibCallSimplifier::optimizeTan(CallInst *CI, IRBuilder<> &B) {
1357 Function *Callee = CI->getCalledFunction();
1358 Value *Ret = nullptr;
1359 StringRef Name = Callee->getName();
1360 if (UnsafeFPShrink && Name == "tan" && hasFloatVersion(Name))
1361 Ret = optimizeUnaryDoubleFP(CI, B, true);
1363 Value *Op1 = CI->getArgOperand(0);
1364 auto *OpC = dyn_cast<CallInst>(Op1);
1368 // Both calls must allow unsafe optimizations in order to remove them.
1369 if (!CI->hasUnsafeAlgebra() || !OpC->hasUnsafeAlgebra())
1372 // tan(atan(x)) -> x
1373 // tanf(atanf(x)) -> x
1374 // tanl(atanl(x)) -> x
1376 Function *F = OpC->getCalledFunction();
1377 if (F && TLI->getLibFunc(F->getName(), Func) && TLI->has(Func) &&
1378 ((Func == LibFunc::atan && Callee->getName() == "tan") ||
1379 (Func == LibFunc::atanf && Callee->getName() == "tanf") ||
1380 (Func == LibFunc::atanl && Callee->getName() == "tanl")))
1381 Ret = OpC->getArgOperand(0);
1385 static bool isTrigLibCall(CallInst *CI) {
1386 // We can only hope to do anything useful if we can ignore things like errno
1387 // and floating-point exceptions.
1388 // We already checked the prototype.
1389 return CI->hasFnAttr(Attribute::NoUnwind) &&
1390 CI->hasFnAttr(Attribute::ReadNone);
1393 static void insertSinCosCall(IRBuilder<> &B, Function *OrigCallee, Value *Arg,
1394 bool UseFloat, Value *&Sin, Value *&Cos,
1396 Type *ArgTy = Arg->getType();
1400 Triple T(OrigCallee->getParent()->getTargetTriple());
1402 Name = "__sincospif_stret";
1404 assert(T.getArch() != Triple::x86 && "x86 messy and unsupported for now");
1405 // x86_64 can't use {float, float} since that would be returned in both
1406 // xmm0 and xmm1, which isn't what a real struct would do.
1407 ResTy = T.getArch() == Triple::x86_64
1408 ? static_cast<Type *>(VectorType::get(ArgTy, 2))
1409 : static_cast<Type *>(StructType::get(ArgTy, ArgTy, nullptr));
1411 Name = "__sincospi_stret";
1412 ResTy = StructType::get(ArgTy, ArgTy, nullptr);
1415 Module *M = OrigCallee->getParent();
1416 Value *Callee = M->getOrInsertFunction(Name, OrigCallee->getAttributes(),
1417 ResTy, ArgTy, nullptr);
1419 if (Instruction *ArgInst = dyn_cast<Instruction>(Arg)) {
1420 // If the argument is an instruction, it must dominate all uses so put our
1421 // sincos call there.
1422 B.SetInsertPoint(ArgInst->getParent(), ++ArgInst->getIterator());
1424 // Otherwise (e.g. for a constant) the beginning of the function is as
1425 // good a place as any.
1426 BasicBlock &EntryBB = B.GetInsertBlock()->getParent()->getEntryBlock();
1427 B.SetInsertPoint(&EntryBB, EntryBB.begin());
1430 SinCos = B.CreateCall(Callee, Arg, "sincospi");
1432 if (SinCos->getType()->isStructTy()) {
1433 Sin = B.CreateExtractValue(SinCos, 0, "sinpi");
1434 Cos = B.CreateExtractValue(SinCos, 1, "cospi");
1436 Sin = B.CreateExtractElement(SinCos, ConstantInt::get(B.getInt32Ty(), 0),
1438 Cos = B.CreateExtractElement(SinCos, ConstantInt::get(B.getInt32Ty(), 1),
1443 Value *LibCallSimplifier::optimizeSinCosPi(CallInst *CI, IRBuilder<> &B) {
1444 // Make sure the prototype is as expected, otherwise the rest of the
1445 // function is probably invalid and likely to abort.
1446 if (!isTrigLibCall(CI))
1449 Value *Arg = CI->getArgOperand(0);
1450 SmallVector<CallInst *, 1> SinCalls;
1451 SmallVector<CallInst *, 1> CosCalls;
1452 SmallVector<CallInst *, 1> SinCosCalls;
1454 bool IsFloat = Arg->getType()->isFloatTy();
1456 // Look for all compatible sinpi, cospi and sincospi calls with the same
1457 // argument. If there are enough (in some sense) we can make the
1459 Function *F = CI->getFunction();
1460 for (User *U : Arg->users())
1461 classifyArgUse(U, F, IsFloat, SinCalls, CosCalls, SinCosCalls);
1463 // It's only worthwhile if both sinpi and cospi are actually used.
1464 if (SinCosCalls.empty() && (SinCalls.empty() || CosCalls.empty()))
1467 Value *Sin, *Cos, *SinCos;
1468 insertSinCosCall(B, CI->getCalledFunction(), Arg, IsFloat, Sin, Cos, SinCos);
1470 auto replaceTrigInsts = [this](SmallVectorImpl<CallInst *> &Calls,
1472 for (CallInst *C : Calls)
1473 replaceAllUsesWith(C, Res);
1476 replaceTrigInsts(SinCalls, Sin);
1477 replaceTrigInsts(CosCalls, Cos);
1478 replaceTrigInsts(SinCosCalls, SinCos);
1483 void LibCallSimplifier::classifyArgUse(
1484 Value *Val, Function *F, bool IsFloat,
1485 SmallVectorImpl<CallInst *> &SinCalls,
1486 SmallVectorImpl<CallInst *> &CosCalls,
1487 SmallVectorImpl<CallInst *> &SinCosCalls) {
1488 CallInst *CI = dyn_cast<CallInst>(Val);
1493 // Don't consider calls in other functions.
1494 if (CI->getFunction() != F)
1497 Function *Callee = CI->getCalledFunction();
1499 if (!Callee || !TLI->getLibFunc(*Callee, Func) || !TLI->has(Func) ||
1504 if (Func == LibFunc::sinpif)
1505 SinCalls.push_back(CI);
1506 else if (Func == LibFunc::cospif)
1507 CosCalls.push_back(CI);
1508 else if (Func == LibFunc::sincospif_stret)
1509 SinCosCalls.push_back(CI);
1511 if (Func == LibFunc::sinpi)
1512 SinCalls.push_back(CI);
1513 else if (Func == LibFunc::cospi)
1514 CosCalls.push_back(CI);
1515 else if (Func == LibFunc::sincospi_stret)
1516 SinCosCalls.push_back(CI);
1520 //===----------------------------------------------------------------------===//
1521 // Integer Library Call Optimizations
1522 //===----------------------------------------------------------------------===//
1524 Value *LibCallSimplifier::optimizeFFS(CallInst *CI, IRBuilder<> &B) {
1525 // ffs(x) -> x != 0 ? (i32)llvm.cttz(x)+1 : 0
1526 Value *Op = CI->getArgOperand(0);
1527 Type *ArgType = Op->getType();
1528 Value *F = Intrinsic::getDeclaration(CI->getCalledFunction()->getParent(),
1529 Intrinsic::cttz, ArgType);
1530 Value *V = B.CreateCall(F, {Op, B.getTrue()}, "cttz");
1531 V = B.CreateAdd(V, ConstantInt::get(V->getType(), 1));
1532 V = B.CreateIntCast(V, B.getInt32Ty(), false);
1534 Value *Cond = B.CreateICmpNE(Op, Constant::getNullValue(ArgType));
1535 return B.CreateSelect(Cond, V, B.getInt32(0));
1538 Value *LibCallSimplifier::optimizeFls(CallInst *CI, IRBuilder<> &B) {
1539 // fls(x) -> (i32)(sizeInBits(x) - llvm.ctlz(x, false))
1540 Value *Op = CI->getArgOperand(0);
1541 Type *ArgType = Op->getType();
1542 Value *F = Intrinsic::getDeclaration(CI->getCalledFunction()->getParent(),
1543 Intrinsic::ctlz, ArgType);
1544 Value *V = B.CreateCall(F, {Op, B.getFalse()}, "ctlz");
1545 V = B.CreateSub(ConstantInt::get(V->getType(), ArgType->getIntegerBitWidth()),
1547 return B.CreateIntCast(V, CI->getType(), false);
1550 Value *LibCallSimplifier::optimizeAbs(CallInst *CI, IRBuilder<> &B) {
1551 // abs(x) -> x >s -1 ? x : -x
1552 Value *Op = CI->getArgOperand(0);
1554 B.CreateICmpSGT(Op, Constant::getAllOnesValue(Op->getType()), "ispos");
1555 Value *Neg = B.CreateNeg(Op, "neg");
1556 return B.CreateSelect(Pos, Op, Neg);
1559 Value *LibCallSimplifier::optimizeIsDigit(CallInst *CI, IRBuilder<> &B) {
1560 // isdigit(c) -> (c-'0') <u 10
1561 Value *Op = CI->getArgOperand(0);
1562 Op = B.CreateSub(Op, B.getInt32('0'), "isdigittmp");
1563 Op = B.CreateICmpULT(Op, B.getInt32(10), "isdigit");
1564 return B.CreateZExt(Op, CI->getType());
1567 Value *LibCallSimplifier::optimizeIsAscii(CallInst *CI, IRBuilder<> &B) {
1568 // isascii(c) -> c <u 128
1569 Value *Op = CI->getArgOperand(0);
1570 Op = B.CreateICmpULT(Op, B.getInt32(128), "isascii");
1571 return B.CreateZExt(Op, CI->getType());
1574 Value *LibCallSimplifier::optimizeToAscii(CallInst *CI, IRBuilder<> &B) {
1575 // toascii(c) -> c & 0x7f
1576 return B.CreateAnd(CI->getArgOperand(0),
1577 ConstantInt::get(CI->getType(), 0x7F));
1580 //===----------------------------------------------------------------------===//
1581 // Formatting and IO Library Call Optimizations
1582 //===----------------------------------------------------------------------===//
1584 static bool isReportingError(Function *Callee, CallInst *CI, int StreamArg);
1586 Value *LibCallSimplifier::optimizeErrorReporting(CallInst *CI, IRBuilder<> &B,
1588 Function *Callee = CI->getCalledFunction();
1589 // Error reporting calls should be cold, mark them as such.
1590 // This applies even to non-builtin calls: it is only a hint and applies to
1591 // functions that the frontend might not understand as builtins.
1593 // This heuristic was suggested in:
1594 // Improving Static Branch Prediction in a Compiler
1595 // Brian L. Deitrich, Ben-Chung Cheng, Wen-mei W. Hwu
1596 // Proceedings of PACT'98, Oct. 1998, IEEE
1597 if (!CI->hasFnAttr(Attribute::Cold) &&
1598 isReportingError(Callee, CI, StreamArg)) {
1599 CI->addAttribute(AttributeSet::FunctionIndex, Attribute::Cold);
1605 static bool isReportingError(Function *Callee, CallInst *CI, int StreamArg) {
1606 if (!ColdErrorCalls || !Callee || !Callee->isDeclaration())
1612 // These functions might be considered cold, but only if their stream
1613 // argument is stderr.
1615 if (StreamArg >= (int)CI->getNumArgOperands())
1617 LoadInst *LI = dyn_cast<LoadInst>(CI->getArgOperand(StreamArg));
1620 GlobalVariable *GV = dyn_cast<GlobalVariable>(LI->getPointerOperand());
1621 if (!GV || !GV->isDeclaration())
1623 return GV->getName() == "stderr";
1626 Value *LibCallSimplifier::optimizePrintFString(CallInst *CI, IRBuilder<> &B) {
1627 // Check for a fixed format string.
1628 StringRef FormatStr;
1629 if (!getConstantStringInfo(CI->getArgOperand(0), FormatStr))
1632 // Empty format string -> noop.
1633 if (FormatStr.empty()) // Tolerate printf's declared void.
1634 return CI->use_empty() ? (Value *)CI : ConstantInt::get(CI->getType(), 0);
1636 // Do not do any of the following transformations if the printf return value
1637 // is used, in general the printf return value is not compatible with either
1638 // putchar() or puts().
1639 if (!CI->use_empty())
1642 // printf("x") -> putchar('x'), even for "%" and "%%".
1643 if (FormatStr.size() == 1 || FormatStr == "%%")
1644 return emitPutChar(B.getInt32(FormatStr[0]), B, TLI);
1646 // printf("%s", "a") --> putchar('a')
1647 if (FormatStr == "%s" && CI->getNumArgOperands() > 1) {
1649 if (!getConstantStringInfo(CI->getOperand(1), ChrStr))
1651 if (ChrStr.size() != 1)
1653 return emitPutChar(B.getInt32(ChrStr[0]), B, TLI);
1656 // printf("foo\n") --> puts("foo")
1657 if (FormatStr[FormatStr.size() - 1] == '\n' &&
1658 FormatStr.find('%') == StringRef::npos) { // No format characters.
1659 // Create a string literal with no \n on it. We expect the constant merge
1660 // pass to be run after this pass, to merge duplicate strings.
1661 FormatStr = FormatStr.drop_back();
1662 Value *GV = B.CreateGlobalString(FormatStr, "str");
1663 return emitPutS(GV, B, TLI);
1666 // Optimize specific format strings.
1667 // printf("%c", chr) --> putchar(chr)
1668 if (FormatStr == "%c" && CI->getNumArgOperands() > 1 &&
1669 CI->getArgOperand(1)->getType()->isIntegerTy())
1670 return emitPutChar(CI->getArgOperand(1), B, TLI);
1672 // printf("%s\n", str) --> puts(str)
1673 if (FormatStr == "%s\n" && CI->getNumArgOperands() > 1 &&
1674 CI->getArgOperand(1)->getType()->isPointerTy())
1675 return emitPutS(CI->getArgOperand(1), B, TLI);
1679 Value *LibCallSimplifier::optimizePrintF(CallInst *CI, IRBuilder<> &B) {
1681 Function *Callee = CI->getCalledFunction();
1682 FunctionType *FT = Callee->getFunctionType();
1683 if (Value *V = optimizePrintFString(CI, B)) {
1687 // printf(format, ...) -> iprintf(format, ...) if no floating point
1689 if (TLI->has(LibFunc::iprintf) && !callHasFloatingPointArgument(CI)) {
1690 Module *M = B.GetInsertBlock()->getParent()->getParent();
1691 Constant *IPrintFFn =
1692 M->getOrInsertFunction("iprintf", FT, Callee->getAttributes());
1693 CallInst *New = cast<CallInst>(CI->clone());
1694 New->setCalledFunction(IPrintFFn);
1701 Value *LibCallSimplifier::optimizeSPrintFString(CallInst *CI, IRBuilder<> &B) {
1702 // Check for a fixed format string.
1703 StringRef FormatStr;
1704 if (!getConstantStringInfo(CI->getArgOperand(1), FormatStr))
1707 // If we just have a format string (nothing else crazy) transform it.
1708 if (CI->getNumArgOperands() == 2) {
1709 // Make sure there's no % in the constant array. We could try to handle
1710 // %% -> % in the future if we cared.
1711 for (unsigned i = 0, e = FormatStr.size(); i != e; ++i)
1712 if (FormatStr[i] == '%')
1713 return nullptr; // we found a format specifier, bail out.
1715 // sprintf(str, fmt) -> llvm.memcpy(str, fmt, strlen(fmt)+1, 1)
1716 B.CreateMemCpy(CI->getArgOperand(0), CI->getArgOperand(1),
1717 ConstantInt::get(DL.getIntPtrType(CI->getContext()),
1718 FormatStr.size() + 1),
1719 1); // Copy the null byte.
1720 return ConstantInt::get(CI->getType(), FormatStr.size());
1723 // The remaining optimizations require the format string to be "%s" or "%c"
1724 // and have an extra operand.
1725 if (FormatStr.size() != 2 || FormatStr[0] != '%' ||
1726 CI->getNumArgOperands() < 3)
1729 // Decode the second character of the format string.
1730 if (FormatStr[1] == 'c') {
1731 // sprintf(dst, "%c", chr) --> *(i8*)dst = chr; *((i8*)dst+1) = 0
1732 if (!CI->getArgOperand(2)->getType()->isIntegerTy())
1734 Value *V = B.CreateTrunc(CI->getArgOperand(2), B.getInt8Ty(), "char");
1735 Value *Ptr = castToCStr(CI->getArgOperand(0), B);
1736 B.CreateStore(V, Ptr);
1737 Ptr = B.CreateGEP(B.getInt8Ty(), Ptr, B.getInt32(1), "nul");
1738 B.CreateStore(B.getInt8(0), Ptr);
1740 return ConstantInt::get(CI->getType(), 1);
1743 if (FormatStr[1] == 's') {
1744 // sprintf(dest, "%s", str) -> llvm.memcpy(dest, str, strlen(str)+1, 1)
1745 if (!CI->getArgOperand(2)->getType()->isPointerTy())
1748 Value *Len = emitStrLen(CI->getArgOperand(2), B, DL, TLI);
1752 B.CreateAdd(Len, ConstantInt::get(Len->getType(), 1), "leninc");
1753 B.CreateMemCpy(CI->getArgOperand(0), CI->getArgOperand(2), IncLen, 1);
1755 // The sprintf result is the unincremented number of bytes in the string.
1756 return B.CreateIntCast(Len, CI->getType(), false);
1761 Value *LibCallSimplifier::optimizeSPrintF(CallInst *CI, IRBuilder<> &B) {
1762 Function *Callee = CI->getCalledFunction();
1763 FunctionType *FT = Callee->getFunctionType();
1764 if (Value *V = optimizeSPrintFString(CI, B)) {
1768 // sprintf(str, format, ...) -> siprintf(str, format, ...) if no floating
1770 if (TLI->has(LibFunc::siprintf) && !callHasFloatingPointArgument(CI)) {
1771 Module *M = B.GetInsertBlock()->getParent()->getParent();
1772 Constant *SIPrintFFn =
1773 M->getOrInsertFunction("siprintf", FT, Callee->getAttributes());
1774 CallInst *New = cast<CallInst>(CI->clone());
1775 New->setCalledFunction(SIPrintFFn);
1782 Value *LibCallSimplifier::optimizeFPrintFString(CallInst *CI, IRBuilder<> &B) {
1783 optimizeErrorReporting(CI, B, 0);
1785 // All the optimizations depend on the format string.
1786 StringRef FormatStr;
1787 if (!getConstantStringInfo(CI->getArgOperand(1), FormatStr))
1790 // Do not do any of the following transformations if the fprintf return
1791 // value is used, in general the fprintf return value is not compatible
1792 // with fwrite(), fputc() or fputs().
1793 if (!CI->use_empty())
1796 // fprintf(F, "foo") --> fwrite("foo", 3, 1, F)
1797 if (CI->getNumArgOperands() == 2) {
1798 for (unsigned i = 0, e = FormatStr.size(); i != e; ++i)
1799 if (FormatStr[i] == '%') // Could handle %% -> % if we cared.
1800 return nullptr; // We found a format specifier.
1803 CI->getArgOperand(1),
1804 ConstantInt::get(DL.getIntPtrType(CI->getContext()), FormatStr.size()),
1805 CI->getArgOperand(0), B, DL, TLI);
1808 // The remaining optimizations require the format string to be "%s" or "%c"
1809 // and have an extra operand.
1810 if (FormatStr.size() != 2 || FormatStr[0] != '%' ||
1811 CI->getNumArgOperands() < 3)
1814 // Decode the second character of the format string.
1815 if (FormatStr[1] == 'c') {
1816 // fprintf(F, "%c", chr) --> fputc(chr, F)
1817 if (!CI->getArgOperand(2)->getType()->isIntegerTy())
1819 return emitFPutC(CI->getArgOperand(2), CI->getArgOperand(0), B, TLI);
1822 if (FormatStr[1] == 's') {
1823 // fprintf(F, "%s", str) --> fputs(str, F)
1824 if (!CI->getArgOperand(2)->getType()->isPointerTy())
1826 return emitFPutS(CI->getArgOperand(2), CI->getArgOperand(0), B, TLI);
1831 Value *LibCallSimplifier::optimizeFPrintF(CallInst *CI, IRBuilder<> &B) {
1832 Function *Callee = CI->getCalledFunction();
1833 FunctionType *FT = Callee->getFunctionType();
1834 if (Value *V = optimizeFPrintFString(CI, B)) {
1838 // fprintf(stream, format, ...) -> fiprintf(stream, format, ...) if no
1839 // floating point arguments.
1840 if (TLI->has(LibFunc::fiprintf) && !callHasFloatingPointArgument(CI)) {
1841 Module *M = B.GetInsertBlock()->getParent()->getParent();
1842 Constant *FIPrintFFn =
1843 M->getOrInsertFunction("fiprintf", FT, Callee->getAttributes());
1844 CallInst *New = cast<CallInst>(CI->clone());
1845 New->setCalledFunction(FIPrintFFn);
1852 Value *LibCallSimplifier::optimizeFWrite(CallInst *CI, IRBuilder<> &B) {
1853 optimizeErrorReporting(CI, B, 3);
1855 // Get the element size and count.
1856 ConstantInt *SizeC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
1857 ConstantInt *CountC = dyn_cast<ConstantInt>(CI->getArgOperand(2));
1858 if (!SizeC || !CountC)
1860 uint64_t Bytes = SizeC->getZExtValue() * CountC->getZExtValue();
1862 // If this is writing zero records, remove the call (it's a noop).
1864 return ConstantInt::get(CI->getType(), 0);
1866 // If this is writing one byte, turn it into fputc.
1867 // This optimisation is only valid, if the return value is unused.
1868 if (Bytes == 1 && CI->use_empty()) { // fwrite(S,1,1,F) -> fputc(S[0],F)
1869 Value *Char = B.CreateLoad(castToCStr(CI->getArgOperand(0), B), "char");
1870 Value *NewCI = emitFPutC(Char, CI->getArgOperand(3), B, TLI);
1871 return NewCI ? ConstantInt::get(CI->getType(), 1) : nullptr;
1877 Value *LibCallSimplifier::optimizeFPuts(CallInst *CI, IRBuilder<> &B) {
1878 optimizeErrorReporting(CI, B, 1);
1880 // Don't rewrite fputs to fwrite when optimising for size because fwrite
1881 // requires more arguments and thus extra MOVs are required.
1882 if (CI->getParent()->getParent()->optForSize())
1885 // We can't optimize if return value is used.
1886 if (!CI->use_empty())
1889 // fputs(s,F) --> fwrite(s,1,strlen(s),F)
1890 uint64_t Len = GetStringLength(CI->getArgOperand(0));
1894 // Known to have no uses (see above).
1896 CI->getArgOperand(0),
1897 ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len - 1),
1898 CI->getArgOperand(1), B, DL, TLI);
1901 Value *LibCallSimplifier::optimizePuts(CallInst *CI, IRBuilder<> &B) {
1902 // Check for a constant string.
1904 if (!getConstantStringInfo(CI->getArgOperand(0), Str))
1907 if (Str.empty() && CI->use_empty()) {
1908 // puts("") -> putchar('\n')
1909 Value *Res = emitPutChar(B.getInt32('\n'), B, TLI);
1910 if (CI->use_empty() || !Res)
1912 return B.CreateIntCast(Res, CI->getType(), true);
1918 bool LibCallSimplifier::hasFloatVersion(StringRef FuncName) {
1920 SmallString<20> FloatFuncName = FuncName;
1921 FloatFuncName += 'f';
1922 if (TLI->getLibFunc(FloatFuncName, Func))
1923 return TLI->has(Func);
1927 Value *LibCallSimplifier::optimizeStringMemoryLibCall(CallInst *CI,
1928 IRBuilder<> &Builder) {
1930 Function *Callee = CI->getCalledFunction();
1931 // Check for string/memory library functions.
1932 if (TLI->getLibFunc(*Callee, Func) && TLI->has(Func)) {
1933 // Make sure we never change the calling convention.
1934 assert((ignoreCallingConv(Func) ||
1935 isCallingConvCCompatible(CI)) &&
1936 "Optimizing string/memory libcall would change the calling convention");
1938 case LibFunc::strcat:
1939 return optimizeStrCat(CI, Builder);
1940 case LibFunc::strncat:
1941 return optimizeStrNCat(CI, Builder);
1942 case LibFunc::strchr:
1943 return optimizeStrChr(CI, Builder);
1944 case LibFunc::strrchr:
1945 return optimizeStrRChr(CI, Builder);
1946 case LibFunc::strcmp:
1947 return optimizeStrCmp(CI, Builder);
1948 case LibFunc::strncmp:
1949 return optimizeStrNCmp(CI, Builder);
1950 case LibFunc::strcpy:
1951 return optimizeStrCpy(CI, Builder);
1952 case LibFunc::stpcpy:
1953 return optimizeStpCpy(CI, Builder);
1954 case LibFunc::strncpy:
1955 return optimizeStrNCpy(CI, Builder);
1956 case LibFunc::strlen:
1957 return optimizeStrLen(CI, Builder);
1958 case LibFunc::strpbrk:
1959 return optimizeStrPBrk(CI, Builder);
1960 case LibFunc::strtol:
1961 case LibFunc::strtod:
1962 case LibFunc::strtof:
1963 case LibFunc::strtoul:
1964 case LibFunc::strtoll:
1965 case LibFunc::strtold:
1966 case LibFunc::strtoull:
1967 return optimizeStrTo(CI, Builder);
1968 case LibFunc::strspn:
1969 return optimizeStrSpn(CI, Builder);
1970 case LibFunc::strcspn:
1971 return optimizeStrCSpn(CI, Builder);
1972 case LibFunc::strstr:
1973 return optimizeStrStr(CI, Builder);
1974 case LibFunc::memchr:
1975 return optimizeMemChr(CI, Builder);
1976 case LibFunc::memcmp:
1977 return optimizeMemCmp(CI, Builder);
1978 case LibFunc::memcpy:
1979 return optimizeMemCpy(CI, Builder);
1980 case LibFunc::memmove:
1981 return optimizeMemMove(CI, Builder);
1982 case LibFunc::memset:
1983 return optimizeMemSet(CI, Builder);
1991 Value *LibCallSimplifier::optimizeCall(CallInst *CI) {
1992 if (CI->isNoBuiltin())
1996 Function *Callee = CI->getCalledFunction();
1997 StringRef FuncName = Callee->getName();
1999 SmallVector<OperandBundleDef, 2> OpBundles;
2000 CI->getOperandBundlesAsDefs(OpBundles);
2001 IRBuilder<> Builder(CI, /*FPMathTag=*/nullptr, OpBundles);
2002 bool isCallingConvC = isCallingConvCCompatible(CI);
2004 // Command-line parameter overrides instruction attribute.
2005 if (EnableUnsafeFPShrink.getNumOccurrences() > 0)
2006 UnsafeFPShrink = EnableUnsafeFPShrink;
2007 else if (isa<FPMathOperator>(CI) && CI->hasUnsafeAlgebra())
2008 UnsafeFPShrink = true;
2010 // First, check for intrinsics.
2011 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
2012 if (!isCallingConvC)
2014 switch (II->getIntrinsicID()) {
2015 case Intrinsic::pow:
2016 return optimizePow(CI, Builder);
2017 case Intrinsic::exp2:
2018 return optimizeExp2(CI, Builder);
2019 case Intrinsic::fabs:
2020 return optimizeFabs(CI, Builder);
2021 case Intrinsic::log:
2022 return optimizeLog(CI, Builder);
2023 case Intrinsic::sqrt:
2024 return optimizeSqrt(CI, Builder);
2025 // TODO: Use foldMallocMemset() with memset intrinsic.
2031 // Also try to simplify calls to fortified library functions.
2032 if (Value *SimplifiedFortifiedCI = FortifiedSimplifier.optimizeCall(CI)) {
2033 // Try to further simplify the result.
2034 CallInst *SimplifiedCI = dyn_cast<CallInst>(SimplifiedFortifiedCI);
2035 if (SimplifiedCI && SimplifiedCI->getCalledFunction()) {
2036 // Use an IR Builder from SimplifiedCI if available instead of CI
2037 // to guarantee we reach all uses we might replace later on.
2038 IRBuilder<> TmpBuilder(SimplifiedCI);
2039 if (Value *V = optimizeStringMemoryLibCall(SimplifiedCI, TmpBuilder)) {
2040 // If we were able to further simplify, remove the now redundant call.
2041 SimplifiedCI->replaceAllUsesWith(V);
2042 SimplifiedCI->eraseFromParent();
2046 return SimplifiedFortifiedCI;
2049 // Then check for known library functions.
2050 if (TLI->getLibFunc(*Callee, Func) && TLI->has(Func)) {
2051 // We never change the calling convention.
2052 if (!ignoreCallingConv(Func) && !isCallingConvC)
2054 if (Value *V = optimizeStringMemoryLibCall(CI, Builder))
2060 return optimizeCos(CI, Builder);
2061 case LibFunc::sinpif:
2062 case LibFunc::sinpi:
2063 case LibFunc::cospif:
2064 case LibFunc::cospi:
2065 return optimizeSinCosPi(CI, Builder);
2069 return optimizePow(CI, Builder);
2070 case LibFunc::exp2l:
2072 case LibFunc::exp2f:
2073 return optimizeExp2(CI, Builder);
2074 case LibFunc::fabsf:
2076 case LibFunc::fabsl:
2077 return optimizeFabs(CI, Builder);
2078 case LibFunc::sqrtf:
2080 case LibFunc::sqrtl:
2081 return optimizeSqrt(CI, Builder);
2084 case LibFunc::ffsll:
2085 return optimizeFFS(CI, Builder);
2088 case LibFunc::flsll:
2089 return optimizeFls(CI, Builder);
2092 case LibFunc::llabs:
2093 return optimizeAbs(CI, Builder);
2094 case LibFunc::isdigit:
2095 return optimizeIsDigit(CI, Builder);
2096 case LibFunc::isascii:
2097 return optimizeIsAscii(CI, Builder);
2098 case LibFunc::toascii:
2099 return optimizeToAscii(CI, Builder);
2100 case LibFunc::printf:
2101 return optimizePrintF(CI, Builder);
2102 case LibFunc::sprintf:
2103 return optimizeSPrintF(CI, Builder);
2104 case LibFunc::fprintf:
2105 return optimizeFPrintF(CI, Builder);
2106 case LibFunc::fwrite:
2107 return optimizeFWrite(CI, Builder);
2108 case LibFunc::fputs:
2109 return optimizeFPuts(CI, Builder);
2111 case LibFunc::log10:
2112 case LibFunc::log1p:
2115 return optimizeLog(CI, Builder);
2117 return optimizePuts(CI, Builder);
2121 return optimizeTan(CI, Builder);
2122 case LibFunc::perror:
2123 return optimizeErrorReporting(CI, Builder);
2124 case LibFunc::vfprintf:
2125 case LibFunc::fiprintf:
2126 return optimizeErrorReporting(CI, Builder, 0);
2127 case LibFunc::fputc:
2128 return optimizeErrorReporting(CI, Builder, 1);
2130 case LibFunc::floor:
2132 case LibFunc::round:
2133 case LibFunc::nearbyint:
2134 case LibFunc::trunc:
2135 if (hasFloatVersion(FuncName))
2136 return optimizeUnaryDoubleFP(CI, Builder, false);
2139 case LibFunc::acosh:
2141 case LibFunc::asinh:
2143 case LibFunc::atanh:
2147 case LibFunc::exp10:
2148 case LibFunc::expm1:
2152 if (UnsafeFPShrink && hasFloatVersion(FuncName))
2153 return optimizeUnaryDoubleFP(CI, Builder, true);
2155 case LibFunc::copysign:
2156 if (hasFloatVersion(FuncName))
2157 return optimizeBinaryDoubleFP(CI, Builder);
2159 case LibFunc::fminf:
2161 case LibFunc::fminl:
2162 case LibFunc::fmaxf:
2164 case LibFunc::fmaxl:
2165 return optimizeFMinFMax(CI, Builder);
2173 LibCallSimplifier::LibCallSimplifier(
2174 const DataLayout &DL, const TargetLibraryInfo *TLI,
2175 function_ref<void(Instruction *, Value *)> Replacer)
2176 : FortifiedSimplifier(TLI), DL(DL), TLI(TLI), UnsafeFPShrink(false),
2177 Replacer(Replacer) {}
2179 void LibCallSimplifier::replaceAllUsesWith(Instruction *I, Value *With) {
2180 // Indirect through the replacer used in this instance.
2185 // Additional cases that we need to add to this file:
2188 // * cbrt(expN(X)) -> expN(x/3)
2189 // * cbrt(sqrt(x)) -> pow(x,1/6)
2190 // * cbrt(cbrt(x)) -> pow(x,1/9)
2193 // * exp(log(x)) -> x
2196 // * log(exp(x)) -> x
2197 // * log(exp(y)) -> y*log(e)
2198 // * log(exp10(y)) -> y*log(10)
2199 // * log(sqrt(x)) -> 0.5*log(x)
2201 // lround, lroundf, lroundl:
2202 // * lround(cnst) -> cnst'
2205 // * pow(sqrt(x),y) -> pow(x,y*0.5)
2206 // * pow(pow(x,y),z)-> pow(x,y*z)
2208 // round, roundf, roundl:
2209 // * round(cnst) -> cnst'
2212 // * signbit(cnst) -> cnst'
2213 // * signbit(nncst) -> 0 (if pstv is a non-negative constant)
2215 // sqrt, sqrtf, sqrtl:
2216 // * sqrt(expN(x)) -> expN(x*0.5)
2217 // * sqrt(Nroot(x)) -> pow(x,1/(2*N))
2218 // * sqrt(pow(x,y)) -> pow(|x|,y*0.5)
2220 // trunc, truncf, truncl:
2221 // * trunc(cnst) -> cnst'
2225 //===----------------------------------------------------------------------===//
2226 // Fortified Library Call Optimizations
2227 //===----------------------------------------------------------------------===//
2229 bool FortifiedLibCallSimplifier::isFortifiedCallFoldable(CallInst *CI,
2233 if (CI->getArgOperand(ObjSizeOp) == CI->getArgOperand(SizeOp))
2235 if (ConstantInt *ObjSizeCI =
2236 dyn_cast<ConstantInt>(CI->getArgOperand(ObjSizeOp))) {
2237 if (ObjSizeCI->isAllOnesValue())
2239 // If the object size wasn't -1 (unknown), bail out if we were asked to.
2240 if (OnlyLowerUnknownSize)
2243 uint64_t Len = GetStringLength(CI->getArgOperand(SizeOp));
2244 // If the length is 0 we don't know how long it is and so we can't
2245 // remove the check.
2248 return ObjSizeCI->getZExtValue() >= Len;
2250 if (ConstantInt *SizeCI = dyn_cast<ConstantInt>(CI->getArgOperand(SizeOp)))
2251 return ObjSizeCI->getZExtValue() >= SizeCI->getZExtValue();
2256 Value *FortifiedLibCallSimplifier::optimizeMemCpyChk(CallInst *CI,
2258 if (isFortifiedCallFoldable(CI, 3, 2, false)) {
2259 B.CreateMemCpy(CI->getArgOperand(0), CI->getArgOperand(1),
2260 CI->getArgOperand(2), 1);
2261 return CI->getArgOperand(0);
2266 Value *FortifiedLibCallSimplifier::optimizeMemMoveChk(CallInst *CI,
2268 if (isFortifiedCallFoldable(CI, 3, 2, false)) {
2269 B.CreateMemMove(CI->getArgOperand(0), CI->getArgOperand(1),
2270 CI->getArgOperand(2), 1);
2271 return CI->getArgOperand(0);
2276 Value *FortifiedLibCallSimplifier::optimizeMemSetChk(CallInst *CI,
2278 // TODO: Try foldMallocMemset() here.
2280 if (isFortifiedCallFoldable(CI, 3, 2, false)) {
2281 Value *Val = B.CreateIntCast(CI->getArgOperand(1), B.getInt8Ty(), false);
2282 B.CreateMemSet(CI->getArgOperand(0), Val, CI->getArgOperand(2), 1);
2283 return CI->getArgOperand(0);
2288 Value *FortifiedLibCallSimplifier::optimizeStrpCpyChk(CallInst *CI,
2290 LibFunc::Func Func) {
2291 Function *Callee = CI->getCalledFunction();
2292 StringRef Name = Callee->getName();
2293 const DataLayout &DL = CI->getModule()->getDataLayout();
2294 Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1),
2295 *ObjSize = CI->getArgOperand(2);
2297 // __stpcpy_chk(x,x,...) -> x+strlen(x)
2298 if (Func == LibFunc::stpcpy_chk && !OnlyLowerUnknownSize && Dst == Src) {
2299 Value *StrLen = emitStrLen(Src, B, DL, TLI);
2300 return StrLen ? B.CreateInBoundsGEP(B.getInt8Ty(), Dst, StrLen) : nullptr;
2303 // If a) we don't have any length information, or b) we know this will
2304 // fit then just lower to a plain st[rp]cpy. Otherwise we'll keep our
2305 // st[rp]cpy_chk call which may fail at runtime if the size is too long.
2306 // TODO: It might be nice to get a maximum length out of the possible
2307 // string lengths for varying.
2308 if (isFortifiedCallFoldable(CI, 2, 1, true))
2309 return emitStrCpy(Dst, Src, B, TLI, Name.substr(2, 6));
2311 if (OnlyLowerUnknownSize)
2314 // Maybe we can stil fold __st[rp]cpy_chk to __memcpy_chk.
2315 uint64_t Len = GetStringLength(Src);
2319 Type *SizeTTy = DL.getIntPtrType(CI->getContext());
2320 Value *LenV = ConstantInt::get(SizeTTy, Len);
2321 Value *Ret = emitMemCpyChk(Dst, Src, LenV, ObjSize, B, DL, TLI);
2322 // If the function was an __stpcpy_chk, and we were able to fold it into
2323 // a __memcpy_chk, we still need to return the correct end pointer.
2324 if (Ret && Func == LibFunc::stpcpy_chk)
2325 return B.CreateGEP(B.getInt8Ty(), Dst, ConstantInt::get(SizeTTy, Len - 1));
2329 Value *FortifiedLibCallSimplifier::optimizeStrpNCpyChk(CallInst *CI,
2331 LibFunc::Func Func) {
2332 Function *Callee = CI->getCalledFunction();
2333 StringRef Name = Callee->getName();
2334 if (isFortifiedCallFoldable(CI, 3, 2, false)) {
2335 Value *Ret = emitStrNCpy(CI->getArgOperand(0), CI->getArgOperand(1),
2336 CI->getArgOperand(2), B, TLI, Name.substr(2, 7));
2342 Value *FortifiedLibCallSimplifier::optimizeCall(CallInst *CI) {
2343 // FIXME: We shouldn't be changing "nobuiltin" or TLI unavailable calls here.
2344 // Some clang users checked for _chk libcall availability using:
2345 // __has_builtin(__builtin___memcpy_chk)
2346 // When compiling with -fno-builtin, this is always true.
2347 // When passing -ffreestanding/-mkernel, which both imply -fno-builtin, we
2348 // end up with fortified libcalls, which isn't acceptable in a freestanding
2349 // environment which only provides their non-fortified counterparts.
2351 // Until we change clang and/or teach external users to check for availability
2352 // differently, disregard the "nobuiltin" attribute and TLI::has.
2357 Function *Callee = CI->getCalledFunction();
2359 SmallVector<OperandBundleDef, 2> OpBundles;
2360 CI->getOperandBundlesAsDefs(OpBundles);
2361 IRBuilder<> Builder(CI, /*FPMathTag=*/nullptr, OpBundles);
2362 bool isCallingConvC = isCallingConvCCompatible(CI);
2364 // First, check that this is a known library functions and that the prototype
2366 if (!TLI->getLibFunc(*Callee, Func))
2369 // We never change the calling convention.
2370 if (!ignoreCallingConv(Func) && !isCallingConvC)
2374 case LibFunc::memcpy_chk:
2375 return optimizeMemCpyChk(CI, Builder);
2376 case LibFunc::memmove_chk:
2377 return optimizeMemMoveChk(CI, Builder);
2378 case LibFunc::memset_chk:
2379 return optimizeMemSetChk(CI, Builder);
2380 case LibFunc::stpcpy_chk:
2381 case LibFunc::strcpy_chk:
2382 return optimizeStrpCpyChk(CI, Builder, Func);
2383 case LibFunc::stpncpy_chk:
2384 case LibFunc::strncpy_chk:
2385 return optimizeStrpNCpyChk(CI, Builder, Func);
2392 FortifiedLibCallSimplifier::FortifiedLibCallSimplifier(
2393 const TargetLibraryInfo *TLI, bool OnlyLowerUnknownSize)
2394 : TLI(TLI), OnlyLowerUnknownSize(OnlyLowerUnknownSize) {}