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 /// Return true if it only matters that the value is equal or not-equal to zero.
60 static bool isOnlyUsedInZeroEqualityComparison(Value *V) {
61 for (User *U : V->users()) {
62 if (ICmpInst *IC = dyn_cast<ICmpInst>(U))
64 if (Constant *C = dyn_cast<Constant>(IC->getOperand(1)))
67 // Unknown instruction.
73 /// Return true if it is only used in equality comparisons with With.
74 static bool isOnlyUsedInEqualityComparison(Value *V, Value *With) {
75 for (User *U : V->users()) {
76 if (ICmpInst *IC = dyn_cast<ICmpInst>(U))
77 if (IC->isEquality() && IC->getOperand(1) == With)
79 // Unknown instruction.
85 static bool callHasFloatingPointArgument(const CallInst *CI) {
86 return std::any_of(CI->op_begin(), CI->op_end(), [](const Use &OI) {
87 return OI->getType()->isFloatingPointTy();
91 /// \brief Check whether the overloaded unary floating point function
92 /// corresponding to \a Ty is available.
93 static bool hasUnaryFloatFn(const TargetLibraryInfo *TLI, Type *Ty,
94 LibFunc::Func DoubleFn, LibFunc::Func FloatFn,
95 LibFunc::Func LongDoubleFn) {
96 switch (Ty->getTypeID()) {
98 return TLI->has(FloatFn);
99 case Type::DoubleTyID:
100 return TLI->has(DoubleFn);
102 return TLI->has(LongDoubleFn);
106 //===----------------------------------------------------------------------===//
107 // String and Memory Library Call Optimizations
108 //===----------------------------------------------------------------------===//
110 Value *LibCallSimplifier::optimizeStrCat(CallInst *CI, IRBuilder<> &B) {
111 // Extract some information from the instruction
112 Value *Dst = CI->getArgOperand(0);
113 Value *Src = CI->getArgOperand(1);
115 // See if we can get the length of the input string.
116 uint64_t Len = GetStringLength(Src);
119 --Len; // Unbias length.
121 // Handle the simple, do-nothing case: strcat(x, "") -> x
125 return emitStrLenMemCpy(Src, Dst, Len, B);
128 Value *LibCallSimplifier::emitStrLenMemCpy(Value *Src, Value *Dst, uint64_t Len,
130 // We need to find the end of the destination string. That's where the
131 // memory is to be moved to. We just generate a call to strlen.
132 Value *DstLen = emitStrLen(Dst, B, DL, TLI);
136 // Now that we have the destination's length, we must index into the
137 // destination's pointer to get the actual memcpy destination (end of
138 // the string .. we're concatenating).
139 Value *CpyDst = B.CreateGEP(B.getInt8Ty(), Dst, DstLen, "endptr");
141 // We have enough information to now generate the memcpy call to do the
142 // concatenation for us. Make a memcpy to copy the nul byte with align = 1.
143 B.CreateMemCpy(CpyDst, Src,
144 ConstantInt::get(DL.getIntPtrType(Src->getContext()), Len + 1),
149 Value *LibCallSimplifier::optimizeStrNCat(CallInst *CI, IRBuilder<> &B) {
150 // Extract some information from the instruction.
151 Value *Dst = CI->getArgOperand(0);
152 Value *Src = CI->getArgOperand(1);
155 // We don't do anything if length is not constant.
156 if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(CI->getArgOperand(2)))
157 Len = LengthArg->getZExtValue();
161 // See if we can get the length of the input string.
162 uint64_t SrcLen = GetStringLength(Src);
165 --SrcLen; // Unbias length.
167 // Handle the simple, do-nothing cases:
168 // strncat(x, "", c) -> x
169 // strncat(x, c, 0) -> x
170 if (SrcLen == 0 || Len == 0)
173 // We don't optimize this case.
177 // strncat(x, s, c) -> strcat(x, s)
178 // s is constant so the strcat can be optimized further.
179 return emitStrLenMemCpy(Src, Dst, SrcLen, B);
182 Value *LibCallSimplifier::optimizeStrChr(CallInst *CI, IRBuilder<> &B) {
183 Function *Callee = CI->getCalledFunction();
184 FunctionType *FT = Callee->getFunctionType();
185 Value *SrcStr = CI->getArgOperand(0);
187 // If the second operand is non-constant, see if we can compute the length
188 // of the input string and turn this into memchr.
189 ConstantInt *CharC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
191 uint64_t Len = GetStringLength(SrcStr);
192 if (Len == 0 || !FT->getParamType(1)->isIntegerTy(32)) // memchr needs i32.
195 return emitMemChr(SrcStr, CI->getArgOperand(1), // include nul.
196 ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len),
200 // Otherwise, the character is a constant, see if the first argument is
201 // a string literal. If so, we can constant fold.
203 if (!getConstantStringInfo(SrcStr, Str)) {
204 if (CharC->isZero()) // strchr(p, 0) -> p + strlen(p)
205 return B.CreateGEP(B.getInt8Ty(), SrcStr, emitStrLen(SrcStr, B, DL, TLI),
210 // Compute the offset, make sure to handle the case when we're searching for
211 // zero (a weird way to spell strlen).
212 size_t I = (0xFF & CharC->getSExtValue()) == 0
214 : Str.find(CharC->getSExtValue());
215 if (I == StringRef::npos) // Didn't find the char. strchr returns null.
216 return Constant::getNullValue(CI->getType());
218 // strchr(s+n,c) -> gep(s+n+i,c)
219 return B.CreateGEP(B.getInt8Ty(), SrcStr, B.getInt64(I), "strchr");
222 Value *LibCallSimplifier::optimizeStrRChr(CallInst *CI, IRBuilder<> &B) {
223 Value *SrcStr = CI->getArgOperand(0);
224 ConstantInt *CharC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
226 // Cannot fold anything if we're not looking for a constant.
231 if (!getConstantStringInfo(SrcStr, Str)) {
232 // strrchr(s, 0) -> strchr(s, 0)
234 return emitStrChr(SrcStr, '\0', B, TLI);
238 // Compute the offset.
239 size_t I = (0xFF & CharC->getSExtValue()) == 0
241 : Str.rfind(CharC->getSExtValue());
242 if (I == StringRef::npos) // Didn't find the char. Return null.
243 return Constant::getNullValue(CI->getType());
245 // strrchr(s+n,c) -> gep(s+n+i,c)
246 return B.CreateGEP(B.getInt8Ty(), SrcStr, B.getInt64(I), "strrchr");
249 Value *LibCallSimplifier::optimizeStrCmp(CallInst *CI, IRBuilder<> &B) {
250 Value *Str1P = CI->getArgOperand(0), *Str2P = CI->getArgOperand(1);
251 if (Str1P == Str2P) // strcmp(x,x) -> 0
252 return ConstantInt::get(CI->getType(), 0);
254 StringRef Str1, Str2;
255 bool HasStr1 = getConstantStringInfo(Str1P, Str1);
256 bool HasStr2 = getConstantStringInfo(Str2P, Str2);
258 // strcmp(x, y) -> cnst (if both x and y are constant strings)
259 if (HasStr1 && HasStr2)
260 return ConstantInt::get(CI->getType(), Str1.compare(Str2));
262 if (HasStr1 && Str1.empty()) // strcmp("", x) -> -*x
264 B.CreateZExt(B.CreateLoad(Str2P, "strcmpload"), CI->getType()));
266 if (HasStr2 && Str2.empty()) // strcmp(x,"") -> *x
267 return B.CreateZExt(B.CreateLoad(Str1P, "strcmpload"), CI->getType());
269 // strcmp(P, "x") -> memcmp(P, "x", 2)
270 uint64_t Len1 = GetStringLength(Str1P);
271 uint64_t Len2 = GetStringLength(Str2P);
273 return emitMemCmp(Str1P, Str2P,
274 ConstantInt::get(DL.getIntPtrType(CI->getContext()),
275 std::min(Len1, Len2)),
282 Value *LibCallSimplifier::optimizeStrNCmp(CallInst *CI, IRBuilder<> &B) {
283 Value *Str1P = CI->getArgOperand(0), *Str2P = CI->getArgOperand(1);
284 if (Str1P == Str2P) // strncmp(x,x,n) -> 0
285 return ConstantInt::get(CI->getType(), 0);
287 // Get the length argument if it is constant.
289 if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(CI->getArgOperand(2)))
290 Length = LengthArg->getZExtValue();
294 if (Length == 0) // strncmp(x,y,0) -> 0
295 return ConstantInt::get(CI->getType(), 0);
297 if (Length == 1) // strncmp(x,y,1) -> memcmp(x,y,1)
298 return emitMemCmp(Str1P, Str2P, CI->getArgOperand(2), B, DL, TLI);
300 StringRef Str1, Str2;
301 bool HasStr1 = getConstantStringInfo(Str1P, Str1);
302 bool HasStr2 = getConstantStringInfo(Str2P, Str2);
304 // strncmp(x, y) -> cnst (if both x and y are constant strings)
305 if (HasStr1 && HasStr2) {
306 StringRef SubStr1 = Str1.substr(0, Length);
307 StringRef SubStr2 = Str2.substr(0, Length);
308 return ConstantInt::get(CI->getType(), SubStr1.compare(SubStr2));
311 if (HasStr1 && Str1.empty()) // strncmp("", x, n) -> -*x
313 B.CreateZExt(B.CreateLoad(Str2P, "strcmpload"), CI->getType()));
315 if (HasStr2 && Str2.empty()) // strncmp(x, "", n) -> *x
316 return B.CreateZExt(B.CreateLoad(Str1P, "strcmpload"), CI->getType());
321 Value *LibCallSimplifier::optimizeStrCpy(CallInst *CI, IRBuilder<> &B) {
322 Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1);
323 if (Dst == Src) // strcpy(x,x) -> x
326 // See if we can get the length of the input string.
327 uint64_t Len = GetStringLength(Src);
331 // We have enough information to now generate the memcpy call to do the
332 // copy for us. Make a memcpy to copy the nul byte with align = 1.
333 B.CreateMemCpy(Dst, Src,
334 ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len), 1);
338 Value *LibCallSimplifier::optimizeStpCpy(CallInst *CI, IRBuilder<> &B) {
339 Function *Callee = CI->getCalledFunction();
340 Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1);
341 if (Dst == Src) { // stpcpy(x,x) -> x+strlen(x)
342 Value *StrLen = emitStrLen(Src, B, DL, TLI);
343 return StrLen ? B.CreateInBoundsGEP(B.getInt8Ty(), Dst, StrLen) : nullptr;
346 // See if we can get the length of the input string.
347 uint64_t Len = GetStringLength(Src);
351 Type *PT = Callee->getFunctionType()->getParamType(0);
352 Value *LenV = ConstantInt::get(DL.getIntPtrType(PT), Len);
353 Value *DstEnd = B.CreateGEP(B.getInt8Ty(), Dst,
354 ConstantInt::get(DL.getIntPtrType(PT), Len - 1));
356 // We have enough information to now generate the memcpy call to do the
357 // copy for us. Make a memcpy to copy the nul byte with align = 1.
358 B.CreateMemCpy(Dst, Src, LenV, 1);
362 Value *LibCallSimplifier::optimizeStrNCpy(CallInst *CI, IRBuilder<> &B) {
363 Function *Callee = CI->getCalledFunction();
364 Value *Dst = CI->getArgOperand(0);
365 Value *Src = CI->getArgOperand(1);
366 Value *LenOp = CI->getArgOperand(2);
368 // See if we can get the length of the input string.
369 uint64_t SrcLen = GetStringLength(Src);
375 // strncpy(x, "", y) -> memset(x, '\0', y, 1)
376 B.CreateMemSet(Dst, B.getInt8('\0'), LenOp, 1);
381 if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(LenOp))
382 Len = LengthArg->getZExtValue();
387 return Dst; // strncpy(x, y, 0) -> x
389 // Let strncpy handle the zero padding
390 if (Len > SrcLen + 1)
393 Type *PT = Callee->getFunctionType()->getParamType(0);
394 // strncpy(x, s, c) -> memcpy(x, s, c, 1) [s and c are constant]
395 B.CreateMemCpy(Dst, Src, ConstantInt::get(DL.getIntPtrType(PT), Len), 1);
400 Value *LibCallSimplifier::optimizeStrLen(CallInst *CI, IRBuilder<> &B) {
401 Value *Src = CI->getArgOperand(0);
403 // Constant folding: strlen("xyz") -> 3
404 if (uint64_t Len = GetStringLength(Src))
405 return ConstantInt::get(CI->getType(), Len - 1);
407 // If s is a constant pointer pointing to a string literal, we can fold
408 // strlen(s + x) to strlen(s) - x, when x is known to be in the range
409 // [0, strlen(s)] or the string has a single null terminator '\0' at the end.
410 // We only try to simplify strlen when the pointer s points to an array
411 // of i8. Otherwise, we would need to scale the offset x before doing the
412 // subtraction. This will make the optimization more complex, and it's not
413 // very useful because calling strlen for a pointer of other types is
415 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Src)) {
416 if (!isGEPBasedOnPointerToString(GEP))
420 if (getConstantStringInfo(GEP->getOperand(0), Str, 0, false)) {
421 size_t NullTermIdx = Str.find('\0');
423 // If the string does not have '\0', leave it to strlen to compute
425 if (NullTermIdx == StringRef::npos)
428 Value *Offset = GEP->getOperand(2);
429 unsigned BitWidth = Offset->getType()->getIntegerBitWidth();
430 APInt KnownZero(BitWidth, 0);
431 APInt KnownOne(BitWidth, 0);
432 computeKnownBits(Offset, KnownZero, KnownOne, DL, 0, nullptr, CI,
434 KnownZero.flipAllBits();
436 cast<ArrayType>(GEP->getSourceElementType())->getNumElements();
438 // KnownZero's bits are flipped, so zeros in KnownZero now represent
439 // bits known to be zeros in Offset, and ones in KnowZero represent
440 // bits unknown in Offset. Therefore, Offset is known to be in range
441 // [0, NullTermIdx] when the flipped KnownZero is non-negative and
442 // unsigned-less-than NullTermIdx.
444 // If Offset is not provably in the range [0, NullTermIdx], we can still
445 // optimize if we can prove that the program has undefined behavior when
446 // Offset is outside that range. That is the case when GEP->getOperand(0)
447 // is a pointer to an object whose memory extent is NullTermIdx+1.
448 if ((KnownZero.isNonNegative() && KnownZero.ule(NullTermIdx)) ||
449 (GEP->isInBounds() && isa<GlobalVariable>(GEP->getOperand(0)) &&
450 NullTermIdx == ArrSize - 1))
451 return B.CreateSub(ConstantInt::get(CI->getType(), NullTermIdx),
458 // strlen(x?"foo":"bars") --> x ? 3 : 4
459 if (SelectInst *SI = dyn_cast<SelectInst>(Src)) {
460 uint64_t LenTrue = GetStringLength(SI->getTrueValue());
461 uint64_t LenFalse = GetStringLength(SI->getFalseValue());
462 if (LenTrue && LenFalse) {
463 Function *Caller = CI->getParent()->getParent();
464 emitOptimizationRemark(CI->getContext(), "simplify-libcalls", *Caller,
466 "folded strlen(select) to select of constants");
467 return B.CreateSelect(SI->getCondition(),
468 ConstantInt::get(CI->getType(), LenTrue - 1),
469 ConstantInt::get(CI->getType(), LenFalse - 1));
473 // strlen(x) != 0 --> *x != 0
474 // strlen(x) == 0 --> *x == 0
475 if (isOnlyUsedInZeroEqualityComparison(CI))
476 return B.CreateZExt(B.CreateLoad(Src, "strlenfirst"), CI->getType());
481 Value *LibCallSimplifier::optimizeStrPBrk(CallInst *CI, IRBuilder<> &B) {
483 bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1);
484 bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2);
486 // strpbrk(s, "") -> nullptr
487 // strpbrk("", s) -> nullptr
488 if ((HasS1 && S1.empty()) || (HasS2 && S2.empty()))
489 return Constant::getNullValue(CI->getType());
492 if (HasS1 && HasS2) {
493 size_t I = S1.find_first_of(S2);
494 if (I == StringRef::npos) // No match.
495 return Constant::getNullValue(CI->getType());
497 return B.CreateGEP(B.getInt8Ty(), CI->getArgOperand(0), B.getInt64(I),
501 // strpbrk(s, "a") -> strchr(s, 'a')
502 if (HasS2 && S2.size() == 1)
503 return emitStrChr(CI->getArgOperand(0), S2[0], B, TLI);
508 Value *LibCallSimplifier::optimizeStrTo(CallInst *CI, IRBuilder<> &B) {
509 Value *EndPtr = CI->getArgOperand(1);
510 if (isa<ConstantPointerNull>(EndPtr)) {
511 // With a null EndPtr, this function won't capture the main argument.
512 // It would be readonly too, except that it still may write to errno.
513 CI->addAttribute(1, Attribute::NoCapture);
519 Value *LibCallSimplifier::optimizeStrSpn(CallInst *CI, IRBuilder<> &B) {
521 bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1);
522 bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2);
524 // strspn(s, "") -> 0
525 // strspn("", s) -> 0
526 if ((HasS1 && S1.empty()) || (HasS2 && S2.empty()))
527 return Constant::getNullValue(CI->getType());
530 if (HasS1 && HasS2) {
531 size_t Pos = S1.find_first_not_of(S2);
532 if (Pos == StringRef::npos)
534 return ConstantInt::get(CI->getType(), Pos);
540 Value *LibCallSimplifier::optimizeStrCSpn(CallInst *CI, IRBuilder<> &B) {
542 bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1);
543 bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2);
545 // strcspn("", s) -> 0
546 if (HasS1 && S1.empty())
547 return Constant::getNullValue(CI->getType());
550 if (HasS1 && HasS2) {
551 size_t Pos = S1.find_first_of(S2);
552 if (Pos == StringRef::npos)
554 return ConstantInt::get(CI->getType(), Pos);
557 // strcspn(s, "") -> strlen(s)
558 if (HasS2 && S2.empty())
559 return emitStrLen(CI->getArgOperand(0), B, DL, TLI);
564 Value *LibCallSimplifier::optimizeStrStr(CallInst *CI, IRBuilder<> &B) {
565 // fold strstr(x, x) -> x.
566 if (CI->getArgOperand(0) == CI->getArgOperand(1))
567 return B.CreateBitCast(CI->getArgOperand(0), CI->getType());
569 // fold strstr(a, b) == a -> strncmp(a, b, strlen(b)) == 0
570 if (isOnlyUsedInEqualityComparison(CI, CI->getArgOperand(0))) {
571 Value *StrLen = emitStrLen(CI->getArgOperand(1), B, DL, TLI);
574 Value *StrNCmp = emitStrNCmp(CI->getArgOperand(0), CI->getArgOperand(1),
578 for (auto UI = CI->user_begin(), UE = CI->user_end(); UI != UE;) {
579 ICmpInst *Old = cast<ICmpInst>(*UI++);
581 B.CreateICmp(Old->getPredicate(), StrNCmp,
582 ConstantInt::getNullValue(StrNCmp->getType()), "cmp");
583 replaceAllUsesWith(Old, Cmp);
588 // See if either input string is a constant string.
589 StringRef SearchStr, ToFindStr;
590 bool HasStr1 = getConstantStringInfo(CI->getArgOperand(0), SearchStr);
591 bool HasStr2 = getConstantStringInfo(CI->getArgOperand(1), ToFindStr);
593 // fold strstr(x, "") -> x.
594 if (HasStr2 && ToFindStr.empty())
595 return B.CreateBitCast(CI->getArgOperand(0), CI->getType());
597 // If both strings are known, constant fold it.
598 if (HasStr1 && HasStr2) {
599 size_t Offset = SearchStr.find(ToFindStr);
601 if (Offset == StringRef::npos) // strstr("foo", "bar") -> null
602 return Constant::getNullValue(CI->getType());
604 // strstr("abcd", "bc") -> gep((char*)"abcd", 1)
605 Value *Result = castToCStr(CI->getArgOperand(0), B);
606 Result = B.CreateConstInBoundsGEP1_64(Result, Offset, "strstr");
607 return B.CreateBitCast(Result, CI->getType());
610 // fold strstr(x, "y") -> strchr(x, 'y').
611 if (HasStr2 && ToFindStr.size() == 1) {
612 Value *StrChr = emitStrChr(CI->getArgOperand(0), ToFindStr[0], B, TLI);
613 return StrChr ? B.CreateBitCast(StrChr, CI->getType()) : nullptr;
618 Value *LibCallSimplifier::optimizeMemChr(CallInst *CI, IRBuilder<> &B) {
619 Value *SrcStr = CI->getArgOperand(0);
620 ConstantInt *CharC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
621 ConstantInt *LenC = dyn_cast<ConstantInt>(CI->getArgOperand(2));
623 // memchr(x, y, 0) -> null
624 if (LenC && LenC->isNullValue())
625 return Constant::getNullValue(CI->getType());
627 // From now on we need at least constant length and string.
629 if (!LenC || !getConstantStringInfo(SrcStr, Str, 0, /*TrimAtNul=*/false))
632 // Truncate the string to LenC. If Str is smaller than LenC we will still only
633 // scan the string, as reading past the end of it is undefined and we can just
634 // return null if we don't find the char.
635 Str = Str.substr(0, LenC->getZExtValue());
637 // If the char is variable but the input str and length are not we can turn
638 // this memchr call into a simple bit field test. Of course this only works
639 // when the return value is only checked against null.
641 // It would be really nice to reuse switch lowering here but we can't change
642 // the CFG at this point.
644 // memchr("\r\n", C, 2) != nullptr -> (C & ((1 << '\r') | (1 << '\n'))) != 0
645 // after bounds check.
646 if (!CharC && !Str.empty() && isOnlyUsedInZeroEqualityComparison(CI)) {
648 *std::max_element(reinterpret_cast<const unsigned char *>(Str.begin()),
649 reinterpret_cast<const unsigned char *>(Str.end()));
651 // Make sure the bit field we're about to create fits in a register on the
653 // FIXME: On a 64 bit architecture this prevents us from using the
654 // interesting range of alpha ascii chars. We could do better by emitting
655 // two bitfields or shifting the range by 64 if no lower chars are used.
656 if (!DL.fitsInLegalInteger(Max + 1))
659 // For the bit field use a power-of-2 type with at least 8 bits to avoid
660 // creating unnecessary illegal types.
661 unsigned char Width = NextPowerOf2(std::max((unsigned char)7, Max));
663 // Now build the bit field.
664 APInt Bitfield(Width, 0);
666 Bitfield.setBit((unsigned char)C);
667 Value *BitfieldC = B.getInt(Bitfield);
669 // First check that the bit field access is within bounds.
670 Value *C = B.CreateZExtOrTrunc(CI->getArgOperand(1), BitfieldC->getType());
671 Value *Bounds = B.CreateICmp(ICmpInst::ICMP_ULT, C, B.getIntN(Width, Width),
674 // Create code that checks if the given bit is set in the field.
675 Value *Shl = B.CreateShl(B.getIntN(Width, 1ULL), C);
676 Value *Bits = B.CreateIsNotNull(B.CreateAnd(Shl, BitfieldC), "memchr.bits");
678 // Finally merge both checks and cast to pointer type. The inttoptr
679 // implicitly zexts the i1 to intptr type.
680 return B.CreateIntToPtr(B.CreateAnd(Bounds, Bits, "memchr"), CI->getType());
683 // Check if all arguments are constants. If so, we can constant fold.
687 // Compute the offset.
688 size_t I = Str.find(CharC->getSExtValue() & 0xFF);
689 if (I == StringRef::npos) // Didn't find the char. memchr returns null.
690 return Constant::getNullValue(CI->getType());
692 // memchr(s+n,c,l) -> gep(s+n+i,c)
693 return B.CreateGEP(B.getInt8Ty(), SrcStr, B.getInt64(I), "memchr");
696 Value *LibCallSimplifier::optimizeMemCmp(CallInst *CI, IRBuilder<> &B) {
697 Value *LHS = CI->getArgOperand(0), *RHS = CI->getArgOperand(1);
699 if (LHS == RHS) // memcmp(s,s,x) -> 0
700 return Constant::getNullValue(CI->getType());
702 // Make sure we have a constant length.
703 ConstantInt *LenC = dyn_cast<ConstantInt>(CI->getArgOperand(2));
706 uint64_t Len = LenC->getZExtValue();
708 if (Len == 0) // memcmp(s1,s2,0) -> 0
709 return Constant::getNullValue(CI->getType());
711 // memcmp(S1,S2,1) -> *(unsigned char*)LHS - *(unsigned char*)RHS
713 Value *LHSV = B.CreateZExt(B.CreateLoad(castToCStr(LHS, B), "lhsc"),
714 CI->getType(), "lhsv");
715 Value *RHSV = B.CreateZExt(B.CreateLoad(castToCStr(RHS, B), "rhsc"),
716 CI->getType(), "rhsv");
717 return B.CreateSub(LHSV, RHSV, "chardiff");
720 // memcmp(S1,S2,N/8)==0 -> (*(intN_t*)S1 != *(intN_t*)S2)==0
721 if (DL.isLegalInteger(Len * 8) && isOnlyUsedInZeroEqualityComparison(CI)) {
723 IntegerType *IntType = IntegerType::get(CI->getContext(), Len * 8);
724 unsigned PrefAlignment = DL.getPrefTypeAlignment(IntType);
726 if (getKnownAlignment(LHS, DL, CI) >= PrefAlignment &&
727 getKnownAlignment(RHS, DL, CI) >= PrefAlignment) {
730 IntType->getPointerTo(LHS->getType()->getPointerAddressSpace());
732 IntType->getPointerTo(RHS->getType()->getPointerAddressSpace());
735 B.CreateLoad(B.CreateBitCast(LHS, LHSPtrTy, "lhsc"), "lhsv");
737 B.CreateLoad(B.CreateBitCast(RHS, RHSPtrTy, "rhsc"), "rhsv");
739 return B.CreateZExt(B.CreateICmpNE(LHSV, RHSV), CI->getType(), "memcmp");
743 // Constant folding: memcmp(x, y, l) -> cnst (all arguments are constant)
744 StringRef LHSStr, RHSStr;
745 if (getConstantStringInfo(LHS, LHSStr) &&
746 getConstantStringInfo(RHS, RHSStr)) {
747 // Make sure we're not reading out-of-bounds memory.
748 if (Len > LHSStr.size() || Len > RHSStr.size())
750 // Fold the memcmp and normalize the result. This way we get consistent
751 // results across multiple platforms.
753 int Cmp = memcmp(LHSStr.data(), RHSStr.data(), Len);
758 return ConstantInt::get(CI->getType(), Ret);
764 Value *LibCallSimplifier::optimizeMemCpy(CallInst *CI, IRBuilder<> &B) {
765 // memcpy(x, y, n) -> llvm.memcpy(x, y, n, 1)
766 B.CreateMemCpy(CI->getArgOperand(0), CI->getArgOperand(1),
767 CI->getArgOperand(2), 1);
768 return CI->getArgOperand(0);
771 Value *LibCallSimplifier::optimizeMemMove(CallInst *CI, IRBuilder<> &B) {
772 // memmove(x, y, n) -> llvm.memmove(x, y, n, 1)
773 B.CreateMemMove(CI->getArgOperand(0), CI->getArgOperand(1),
774 CI->getArgOperand(2), 1);
775 return CI->getArgOperand(0);
778 // TODO: Does this belong in BuildLibCalls or should all of those similar
779 // functions be moved here?
780 static Value *emitCalloc(Value *Num, Value *Size, const AttributeSet &Attrs,
781 IRBuilder<> &B, const TargetLibraryInfo &TLI) {
783 if (!TLI.getLibFunc("calloc", Func) || !TLI.has(Func))
786 Module *M = B.GetInsertBlock()->getModule();
787 const DataLayout &DL = M->getDataLayout();
788 IntegerType *PtrType = DL.getIntPtrType((B.GetInsertBlock()->getContext()));
789 Value *Calloc = M->getOrInsertFunction("calloc", Attrs, B.getInt8PtrTy(),
790 PtrType, PtrType, nullptr);
791 CallInst *CI = B.CreateCall(Calloc, { Num, Size }, "calloc");
793 if (const auto *F = dyn_cast<Function>(Calloc->stripPointerCasts()))
794 CI->setCallingConv(F->getCallingConv());
799 /// Fold memset[_chk](malloc(n), 0, n) --> calloc(1, n).
800 static Value *foldMallocMemset(CallInst *Memset, IRBuilder<> &B,
801 const TargetLibraryInfo &TLI) {
802 // This has to be a memset of zeros (bzero).
803 auto *FillValue = dyn_cast<ConstantInt>(Memset->getArgOperand(1));
804 if (!FillValue || FillValue->getZExtValue() != 0)
807 // TODO: We should handle the case where the malloc has more than one use.
808 // This is necessary to optimize common patterns such as when the result of
809 // the malloc is checked against null or when a memset intrinsic is used in
810 // place of a memset library call.
811 auto *Malloc = dyn_cast<CallInst>(Memset->getArgOperand(0));
812 if (!Malloc || !Malloc->hasOneUse())
815 // Is the inner call really malloc()?
816 Function *InnerCallee = Malloc->getCalledFunction();
818 if (!TLI.getLibFunc(*InnerCallee, Func) || !TLI.has(Func) ||
819 Func != LibFunc::malloc)
822 // The memset must cover the same number of bytes that are malloc'd.
823 if (Memset->getArgOperand(2) != Malloc->getArgOperand(0))
826 // Replace the malloc with a calloc. We need the data layout to know what the
827 // actual size of a 'size_t' parameter is.
828 B.SetInsertPoint(Malloc->getParent(), ++Malloc->getIterator());
829 const DataLayout &DL = Malloc->getModule()->getDataLayout();
830 IntegerType *SizeType = DL.getIntPtrType(B.GetInsertBlock()->getContext());
831 Value *Calloc = emitCalloc(ConstantInt::get(SizeType, 1),
832 Malloc->getArgOperand(0), Malloc->getAttributes(),
837 Malloc->replaceAllUsesWith(Calloc);
838 Malloc->eraseFromParent();
843 Value *LibCallSimplifier::optimizeMemSet(CallInst *CI, IRBuilder<> &B) {
844 if (auto *Calloc = foldMallocMemset(CI, B, *TLI))
847 // memset(p, v, n) -> llvm.memset(p, v, n, 1)
848 Value *Val = B.CreateIntCast(CI->getArgOperand(1), B.getInt8Ty(), false);
849 B.CreateMemSet(CI->getArgOperand(0), Val, CI->getArgOperand(2), 1);
850 return CI->getArgOperand(0);
853 //===----------------------------------------------------------------------===//
854 // Math Library Optimizations
855 //===----------------------------------------------------------------------===//
857 /// Return a variant of Val with float type.
858 /// Currently this works in two cases: If Val is an FPExtension of a float
859 /// value to something bigger, simply return the operand.
860 /// If Val is a ConstantFP but can be converted to a float ConstantFP without
861 /// loss of precision do so.
862 static Value *valueHasFloatPrecision(Value *Val) {
863 if (FPExtInst *Cast = dyn_cast<FPExtInst>(Val)) {
864 Value *Op = Cast->getOperand(0);
865 if (Op->getType()->isFloatTy())
868 if (ConstantFP *Const = dyn_cast<ConstantFP>(Val)) {
869 APFloat F = Const->getValueAPF();
871 (void)F.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven,
874 return ConstantFP::get(Const->getContext(), F);
879 /// Shrink double -> float for unary functions like 'floor'.
880 static Value *optimizeUnaryDoubleFP(CallInst *CI, IRBuilder<> &B,
882 Function *Callee = CI->getCalledFunction();
883 // We know this libcall has a valid prototype, but we don't know which.
884 if (!CI->getType()->isDoubleTy())
888 // Check if all the uses for function like 'sin' are converted to float.
889 for (User *U : CI->users()) {
890 FPTruncInst *Cast = dyn_cast<FPTruncInst>(U);
891 if (!Cast || !Cast->getType()->isFloatTy())
896 // If this is something like 'floor((double)floatval)', convert to floorf.
897 Value *V = valueHasFloatPrecision(CI->getArgOperand(0));
901 // Propagate fast-math flags from the existing call to the new call.
902 IRBuilder<>::FastMathFlagGuard Guard(B);
903 B.setFastMathFlags(CI->getFastMathFlags());
905 // floor((double)floatval) -> (double)floorf(floatval)
906 if (Callee->isIntrinsic()) {
907 Module *M = CI->getModule();
908 Intrinsic::ID IID = Callee->getIntrinsicID();
909 Function *F = Intrinsic::getDeclaration(M, IID, B.getFloatTy());
910 V = B.CreateCall(F, V);
912 // The call is a library call rather than an intrinsic.
913 V = emitUnaryFloatFnCall(V, Callee->getName(), B, Callee->getAttributes());
916 return B.CreateFPExt(V, B.getDoubleTy());
919 /// Shrink double -> float for binary functions like 'fmin/fmax'.
920 static Value *optimizeBinaryDoubleFP(CallInst *CI, IRBuilder<> &B) {
921 Function *Callee = CI->getCalledFunction();
922 // We know this libcall has a valid prototype, but we don't know which.
923 if (!CI->getType()->isDoubleTy())
926 // If this is something like 'fmin((double)floatval1, (double)floatval2)',
927 // or fmin(1.0, (double)floatval), then we convert it to fminf.
928 Value *V1 = valueHasFloatPrecision(CI->getArgOperand(0));
931 Value *V2 = valueHasFloatPrecision(CI->getArgOperand(1));
935 // Propagate fast-math flags from the existing call to the new call.
936 IRBuilder<>::FastMathFlagGuard Guard(B);
937 B.setFastMathFlags(CI->getFastMathFlags());
939 // fmin((double)floatval1, (double)floatval2)
940 // -> (double)fminf(floatval1, floatval2)
941 // TODO: Handle intrinsics in the same way as in optimizeUnaryDoubleFP().
942 Value *V = emitBinaryFloatFnCall(V1, V2, Callee->getName(), B,
943 Callee->getAttributes());
944 return B.CreateFPExt(V, B.getDoubleTy());
947 Value *LibCallSimplifier::optimizeCos(CallInst *CI, IRBuilder<> &B) {
948 Function *Callee = CI->getCalledFunction();
949 Value *Ret = nullptr;
950 StringRef Name = Callee->getName();
951 if (UnsafeFPShrink && Name == "cos" && hasFloatVersion(Name))
952 Ret = optimizeUnaryDoubleFP(CI, B, true);
955 Value *Op1 = CI->getArgOperand(0);
956 if (BinaryOperator::isFNeg(Op1)) {
957 BinaryOperator *BinExpr = cast<BinaryOperator>(Op1);
958 return B.CreateCall(Callee, BinExpr->getOperand(1), "cos");
963 static Value *getPow(Value *InnerChain[33], unsigned Exp, IRBuilder<> &B) {
964 // Multiplications calculated using Addition Chains.
965 // Refer: http://wwwhomes.uni-bielefeld.de/achim/addition_chain.html
967 assert(Exp != 0 && "Incorrect exponent 0 not handled");
970 return InnerChain[Exp];
972 static const unsigned AddChain[33][2] = {
974 {0, 0}, // Unused (base case = pow1).
975 {1, 1}, // Unused (pre-computed).
976 {1, 2}, {2, 2}, {2, 3}, {3, 3}, {2, 5}, {4, 4},
977 {1, 8}, {5, 5}, {1, 10}, {6, 6}, {4, 9}, {7, 7},
978 {3, 12}, {8, 8}, {8, 9}, {2, 16}, {1, 18}, {10, 10},
979 {6, 15}, {11, 11}, {3, 20}, {12, 12}, {8, 17}, {13, 13},
980 {3, 24}, {14, 14}, {4, 25}, {15, 15}, {3, 28}, {16, 16},
983 InnerChain[Exp] = B.CreateFMul(getPow(InnerChain, AddChain[Exp][0], B),
984 getPow(InnerChain, AddChain[Exp][1], B));
985 return InnerChain[Exp];
988 Value *LibCallSimplifier::optimizePow(CallInst *CI, IRBuilder<> &B) {
989 Function *Callee = CI->getCalledFunction();
990 Value *Ret = nullptr;
991 StringRef Name = Callee->getName();
992 if (UnsafeFPShrink && Name == "pow" && hasFloatVersion(Name))
993 Ret = optimizeUnaryDoubleFP(CI, B, true);
995 Value *Op1 = CI->getArgOperand(0), *Op2 = CI->getArgOperand(1);
996 if (ConstantFP *Op1C = dyn_cast<ConstantFP>(Op1)) {
997 // pow(1.0, x) -> 1.0
998 if (Op1C->isExactlyValue(1.0))
1000 // pow(2.0, x) -> exp2(x)
1001 if (Op1C->isExactlyValue(2.0) &&
1002 hasUnaryFloatFn(TLI, Op1->getType(), LibFunc::exp2, LibFunc::exp2f,
1004 return emitUnaryFloatFnCall(Op2, TLI->getName(LibFunc::exp2), B,
1005 Callee->getAttributes());
1006 // pow(10.0, x) -> exp10(x)
1007 if (Op1C->isExactlyValue(10.0) &&
1008 hasUnaryFloatFn(TLI, Op1->getType(), LibFunc::exp10, LibFunc::exp10f,
1010 return emitUnaryFloatFnCall(Op2, TLI->getName(LibFunc::exp10), B,
1011 Callee->getAttributes());
1014 // pow(exp(x), y) -> exp(x * y)
1015 // pow(exp2(x), y) -> exp2(x * y)
1016 // We enable these only with fast-math. Besides rounding differences, the
1017 // transformation changes overflow and underflow behavior quite dramatically.
1018 // Example: x = 1000, y = 0.001.
1019 // pow(exp(x), y) = pow(inf, 0.001) = inf, whereas exp(x*y) = exp(1).
1020 auto *OpC = dyn_cast<CallInst>(Op1);
1021 if (OpC && OpC->hasUnsafeAlgebra() && CI->hasUnsafeAlgebra()) {
1023 Function *OpCCallee = OpC->getCalledFunction();
1024 if (OpCCallee && TLI->getLibFunc(OpCCallee->getName(), Func) &&
1025 TLI->has(Func) && (Func == LibFunc::exp || Func == LibFunc::exp2)) {
1026 IRBuilder<>::FastMathFlagGuard Guard(B);
1027 B.setFastMathFlags(CI->getFastMathFlags());
1028 Value *FMul = B.CreateFMul(OpC->getArgOperand(0), Op2, "mul");
1029 return emitUnaryFloatFnCall(FMul, OpCCallee->getName(), B,
1030 OpCCallee->getAttributes());
1034 ConstantFP *Op2C = dyn_cast<ConstantFP>(Op2);
1038 if (Op2C->getValueAPF().isZero()) // pow(x, 0.0) -> 1.0
1039 return ConstantFP::get(CI->getType(), 1.0);
1041 if (Op2C->isExactlyValue(0.5) &&
1042 hasUnaryFloatFn(TLI, Op2->getType(), LibFunc::sqrt, LibFunc::sqrtf,
1044 hasUnaryFloatFn(TLI, Op2->getType(), LibFunc::fabs, LibFunc::fabsf,
1047 // In -ffast-math, pow(x, 0.5) -> sqrt(x).
1048 if (CI->hasUnsafeAlgebra()) {
1049 IRBuilder<>::FastMathFlagGuard Guard(B);
1050 B.setFastMathFlags(CI->getFastMathFlags());
1051 return emitUnaryFloatFnCall(Op1, TLI->getName(LibFunc::sqrt), B,
1052 Callee->getAttributes());
1055 // Expand pow(x, 0.5) to (x == -infinity ? +infinity : fabs(sqrt(x))).
1056 // This is faster than calling pow, and still handles negative zero
1057 // and negative infinity correctly.
1058 // TODO: In finite-only mode, this could be just fabs(sqrt(x)).
1059 Value *Inf = ConstantFP::getInfinity(CI->getType());
1060 Value *NegInf = ConstantFP::getInfinity(CI->getType(), true);
1061 Value *Sqrt = emitUnaryFloatFnCall(Op1, "sqrt", B, Callee->getAttributes());
1063 emitUnaryFloatFnCall(Sqrt, "fabs", B, Callee->getAttributes());
1064 Value *FCmp = B.CreateFCmpOEQ(Op1, NegInf);
1065 Value *Sel = B.CreateSelect(FCmp, Inf, FAbs);
1069 if (Op2C->isExactlyValue(1.0)) // pow(x, 1.0) -> x
1071 if (Op2C->isExactlyValue(2.0)) // pow(x, 2.0) -> x*x
1072 return B.CreateFMul(Op1, Op1, "pow2");
1073 if (Op2C->isExactlyValue(-1.0)) // pow(x, -1.0) -> 1.0/x
1074 return B.CreateFDiv(ConstantFP::get(CI->getType(), 1.0), Op1, "powrecip");
1076 // In -ffast-math, generate repeated fmul instead of generating pow(x, n).
1077 if (CI->hasUnsafeAlgebra()) {
1078 APFloat V = abs(Op2C->getValueAPF());
1079 // We limit to a max of 7 fmul(s). Thus max exponent is 32.
1080 // This transformation applies to integer exponents only.
1081 if (V.compare(APFloat(V.getSemantics(), 32.0)) == APFloat::cmpGreaterThan ||
1085 // We will memoize intermediate products of the Addition Chain.
1086 Value *InnerChain[33] = {nullptr};
1087 InnerChain[1] = Op1;
1088 InnerChain[2] = B.CreateFMul(Op1, Op1);
1090 // We cannot readily convert a non-double type (like float) to a double.
1091 // So we first convert V to something which could be converted to double.
1093 V.convert(APFloat::IEEEdouble, APFloat::rmTowardZero, &ignored);
1095 // TODO: Should the new instructions propagate the 'fast' flag of the pow()?
1096 Value *FMul = getPow(InnerChain, V.convertToDouble(), B);
1097 // For negative exponents simply compute the reciprocal.
1098 if (Op2C->isNegative())
1099 FMul = B.CreateFDiv(ConstantFP::get(CI->getType(), 1.0), FMul);
1106 Value *LibCallSimplifier::optimizeExp2(CallInst *CI, IRBuilder<> &B) {
1107 Function *Callee = CI->getCalledFunction();
1108 Value *Ret = nullptr;
1109 StringRef Name = Callee->getName();
1110 if (UnsafeFPShrink && Name == "exp2" && hasFloatVersion(Name))
1111 Ret = optimizeUnaryDoubleFP(CI, B, true);
1113 Value *Op = CI->getArgOperand(0);
1114 // Turn exp2(sitofp(x)) -> ldexp(1.0, sext(x)) if sizeof(x) <= 32
1115 // Turn exp2(uitofp(x)) -> ldexp(1.0, zext(x)) if sizeof(x) < 32
1116 LibFunc::Func LdExp = LibFunc::ldexpl;
1117 if (Op->getType()->isFloatTy())
1118 LdExp = LibFunc::ldexpf;
1119 else if (Op->getType()->isDoubleTy())
1120 LdExp = LibFunc::ldexp;
1122 if (TLI->has(LdExp)) {
1123 Value *LdExpArg = nullptr;
1124 if (SIToFPInst *OpC = dyn_cast<SIToFPInst>(Op)) {
1125 if (OpC->getOperand(0)->getType()->getPrimitiveSizeInBits() <= 32)
1126 LdExpArg = B.CreateSExt(OpC->getOperand(0), B.getInt32Ty());
1127 } else if (UIToFPInst *OpC = dyn_cast<UIToFPInst>(Op)) {
1128 if (OpC->getOperand(0)->getType()->getPrimitiveSizeInBits() < 32)
1129 LdExpArg = B.CreateZExt(OpC->getOperand(0), B.getInt32Ty());
1133 Constant *One = ConstantFP::get(CI->getContext(), APFloat(1.0f));
1134 if (!Op->getType()->isFloatTy())
1135 One = ConstantExpr::getFPExtend(One, Op->getType());
1137 Module *M = CI->getModule();
1139 M->getOrInsertFunction(TLI->getName(LdExp), Op->getType(),
1140 Op->getType(), B.getInt32Ty(), nullptr);
1141 CallInst *CI = B.CreateCall(NewCallee, {One, LdExpArg});
1142 if (const Function *F = dyn_cast<Function>(Callee->stripPointerCasts()))
1143 CI->setCallingConv(F->getCallingConv());
1151 Value *LibCallSimplifier::optimizeFabs(CallInst *CI, IRBuilder<> &B) {
1152 Function *Callee = CI->getCalledFunction();
1153 Value *Ret = nullptr;
1154 StringRef Name = Callee->getName();
1155 if (Name == "fabs" && hasFloatVersion(Name))
1156 Ret = optimizeUnaryDoubleFP(CI, B, false);
1158 Value *Op = CI->getArgOperand(0);
1159 if (Instruction *I = dyn_cast<Instruction>(Op)) {
1160 // Fold fabs(x * x) -> x * x; any squared FP value must already be positive.
1161 if (I->getOpcode() == Instruction::FMul)
1162 if (I->getOperand(0) == I->getOperand(1))
1168 Value *LibCallSimplifier::optimizeFMinFMax(CallInst *CI, IRBuilder<> &B) {
1169 Function *Callee = CI->getCalledFunction();
1170 // If we can shrink the call to a float function rather than a double
1171 // function, do that first.
1172 StringRef Name = Callee->getName();
1173 if ((Name == "fmin" || Name == "fmax") && hasFloatVersion(Name))
1174 if (Value *Ret = optimizeBinaryDoubleFP(CI, B))
1177 IRBuilder<>::FastMathFlagGuard Guard(B);
1179 if (CI->hasUnsafeAlgebra()) {
1180 // Unsafe algebra sets all fast-math-flags to true.
1181 FMF.setUnsafeAlgebra();
1183 // At a minimum, no-nans-fp-math must be true.
1184 if (!CI->hasNoNaNs())
1186 // No-signed-zeros is implied by the definitions of fmax/fmin themselves:
1187 // "Ideally, fmax would be sensitive to the sign of zero, for example
1188 // fmax(-0. 0, +0. 0) would return +0; however, implementation in software
1189 // might be impractical."
1190 FMF.setNoSignedZeros();
1193 B.setFastMathFlags(FMF);
1195 // We have a relaxed floating-point environment. We can ignore NaN-handling
1196 // and transform to a compare and select. We do not have to consider errno or
1197 // exceptions, because fmin/fmax do not have those.
1198 Value *Op0 = CI->getArgOperand(0);
1199 Value *Op1 = CI->getArgOperand(1);
1200 Value *Cmp = Callee->getName().startswith("fmin") ?
1201 B.CreateFCmpOLT(Op0, Op1) : B.CreateFCmpOGT(Op0, Op1);
1202 return B.CreateSelect(Cmp, Op0, Op1);
1205 Value *LibCallSimplifier::optimizeLog(CallInst *CI, IRBuilder<> &B) {
1206 Function *Callee = CI->getCalledFunction();
1207 Value *Ret = nullptr;
1208 StringRef Name = Callee->getName();
1209 if (UnsafeFPShrink && hasFloatVersion(Name))
1210 Ret = optimizeUnaryDoubleFP(CI, B, true);
1212 if (!CI->hasUnsafeAlgebra())
1214 Value *Op1 = CI->getArgOperand(0);
1215 auto *OpC = dyn_cast<CallInst>(Op1);
1217 // The earlier call must also be unsafe in order to do these transforms.
1218 if (!OpC || !OpC->hasUnsafeAlgebra())
1221 // log(pow(x,y)) -> y*log(x)
1222 // This is only applicable to log, log2, log10.
1223 if (Name != "log" && Name != "log2" && Name != "log10")
1226 IRBuilder<>::FastMathFlagGuard Guard(B);
1228 FMF.setUnsafeAlgebra();
1229 B.setFastMathFlags(FMF);
1232 Function *F = OpC->getCalledFunction();
1233 if (F && ((TLI->getLibFunc(F->getName(), Func) && TLI->has(Func) &&
1234 Func == LibFunc::pow) || F->getIntrinsicID() == Intrinsic::pow))
1235 return B.CreateFMul(OpC->getArgOperand(1),
1236 emitUnaryFloatFnCall(OpC->getOperand(0), Callee->getName(), B,
1237 Callee->getAttributes()), "mul");
1239 // log(exp2(y)) -> y*log(2)
1240 if (F && Name == "log" && TLI->getLibFunc(F->getName(), Func) &&
1241 TLI->has(Func) && Func == LibFunc::exp2)
1242 return B.CreateFMul(
1243 OpC->getArgOperand(0),
1244 emitUnaryFloatFnCall(ConstantFP::get(CI->getType(), 2.0),
1245 Callee->getName(), B, Callee->getAttributes()),
1250 Value *LibCallSimplifier::optimizeSqrt(CallInst *CI, IRBuilder<> &B) {
1251 Function *Callee = CI->getCalledFunction();
1252 Value *Ret = nullptr;
1253 if (TLI->has(LibFunc::sqrtf) && (Callee->getName() == "sqrt" ||
1254 Callee->getIntrinsicID() == Intrinsic::sqrt))
1255 Ret = optimizeUnaryDoubleFP(CI, B, true);
1257 if (!CI->hasUnsafeAlgebra())
1260 Instruction *I = dyn_cast<Instruction>(CI->getArgOperand(0));
1261 if (!I || I->getOpcode() != Instruction::FMul || !I->hasUnsafeAlgebra())
1264 // We're looking for a repeated factor in a multiplication tree,
1265 // so we can do this fold: sqrt(x * x) -> fabs(x);
1266 // or this fold: sqrt((x * x) * y) -> fabs(x) * sqrt(y).
1267 Value *Op0 = I->getOperand(0);
1268 Value *Op1 = I->getOperand(1);
1269 Value *RepeatOp = nullptr;
1270 Value *OtherOp = nullptr;
1272 // Simple match: the operands of the multiply are identical.
1275 // Look for a more complicated pattern: one of the operands is itself
1276 // a multiply, so search for a common factor in that multiply.
1277 // Note: We don't bother looking any deeper than this first level or for
1278 // variations of this pattern because instcombine's visitFMUL and/or the
1279 // reassociation pass should give us this form.
1280 Value *OtherMul0, *OtherMul1;
1281 if (match(Op0, m_FMul(m_Value(OtherMul0), m_Value(OtherMul1)))) {
1282 // Pattern: sqrt((x * y) * z)
1283 if (OtherMul0 == OtherMul1 &&
1284 cast<Instruction>(Op0)->hasUnsafeAlgebra()) {
1285 // Matched: sqrt((x * x) * z)
1286 RepeatOp = OtherMul0;
1294 // Fast math flags for any created instructions should match the sqrt
1296 IRBuilder<>::FastMathFlagGuard Guard(B);
1297 B.setFastMathFlags(I->getFastMathFlags());
1299 // If we found a repeated factor, hoist it out of the square root and
1300 // replace it with the fabs of that factor.
1301 Module *M = Callee->getParent();
1302 Type *ArgType = I->getType();
1303 Value *Fabs = Intrinsic::getDeclaration(M, Intrinsic::fabs, ArgType);
1304 Value *FabsCall = B.CreateCall(Fabs, RepeatOp, "fabs");
1306 // If we found a non-repeated factor, we still need to get its square
1307 // root. We then multiply that by the value that was simplified out
1308 // of the square root calculation.
1309 Value *Sqrt = Intrinsic::getDeclaration(M, Intrinsic::sqrt, ArgType);
1310 Value *SqrtCall = B.CreateCall(Sqrt, OtherOp, "sqrt");
1311 return B.CreateFMul(FabsCall, SqrtCall);
1316 // TODO: Generalize to handle any trig function and its inverse.
1317 Value *LibCallSimplifier::optimizeTan(CallInst *CI, IRBuilder<> &B) {
1318 Function *Callee = CI->getCalledFunction();
1319 Value *Ret = nullptr;
1320 StringRef Name = Callee->getName();
1321 if (UnsafeFPShrink && Name == "tan" && hasFloatVersion(Name))
1322 Ret = optimizeUnaryDoubleFP(CI, B, true);
1324 Value *Op1 = CI->getArgOperand(0);
1325 auto *OpC = dyn_cast<CallInst>(Op1);
1329 // Both calls must allow unsafe optimizations in order to remove them.
1330 if (!CI->hasUnsafeAlgebra() || !OpC->hasUnsafeAlgebra())
1333 // tan(atan(x)) -> x
1334 // tanf(atanf(x)) -> x
1335 // tanl(atanl(x)) -> x
1337 Function *F = OpC->getCalledFunction();
1338 if (F && TLI->getLibFunc(F->getName(), Func) && TLI->has(Func) &&
1339 ((Func == LibFunc::atan && Callee->getName() == "tan") ||
1340 (Func == LibFunc::atanf && Callee->getName() == "tanf") ||
1341 (Func == LibFunc::atanl && Callee->getName() == "tanl")))
1342 Ret = OpC->getArgOperand(0);
1346 static bool isTrigLibCall(CallInst *CI) {
1347 // We can only hope to do anything useful if we can ignore things like errno
1348 // and floating-point exceptions.
1349 // We already checked the prototype.
1350 return CI->hasFnAttr(Attribute::NoUnwind) &&
1351 CI->hasFnAttr(Attribute::ReadNone);
1354 static void insertSinCosCall(IRBuilder<> &B, Function *OrigCallee, Value *Arg,
1355 bool UseFloat, Value *&Sin, Value *&Cos,
1357 Type *ArgTy = Arg->getType();
1361 Triple T(OrigCallee->getParent()->getTargetTriple());
1363 Name = "__sincospif_stret";
1365 assert(T.getArch() != Triple::x86 && "x86 messy and unsupported for now");
1366 // x86_64 can't use {float, float} since that would be returned in both
1367 // xmm0 and xmm1, which isn't what a real struct would do.
1368 ResTy = T.getArch() == Triple::x86_64
1369 ? static_cast<Type *>(VectorType::get(ArgTy, 2))
1370 : static_cast<Type *>(StructType::get(ArgTy, ArgTy, nullptr));
1372 Name = "__sincospi_stret";
1373 ResTy = StructType::get(ArgTy, ArgTy, nullptr);
1376 Module *M = OrigCallee->getParent();
1377 Value *Callee = M->getOrInsertFunction(Name, OrigCallee->getAttributes(),
1378 ResTy, ArgTy, nullptr);
1380 if (Instruction *ArgInst = dyn_cast<Instruction>(Arg)) {
1381 // If the argument is an instruction, it must dominate all uses so put our
1382 // sincos call there.
1383 B.SetInsertPoint(ArgInst->getParent(), ++ArgInst->getIterator());
1385 // Otherwise (e.g. for a constant) the beginning of the function is as
1386 // good a place as any.
1387 BasicBlock &EntryBB = B.GetInsertBlock()->getParent()->getEntryBlock();
1388 B.SetInsertPoint(&EntryBB, EntryBB.begin());
1391 SinCos = B.CreateCall(Callee, Arg, "sincospi");
1393 if (SinCos->getType()->isStructTy()) {
1394 Sin = B.CreateExtractValue(SinCos, 0, "sinpi");
1395 Cos = B.CreateExtractValue(SinCos, 1, "cospi");
1397 Sin = B.CreateExtractElement(SinCos, ConstantInt::get(B.getInt32Ty(), 0),
1399 Cos = B.CreateExtractElement(SinCos, ConstantInt::get(B.getInt32Ty(), 1),
1404 Value *LibCallSimplifier::optimizeSinCosPi(CallInst *CI, IRBuilder<> &B) {
1405 // Make sure the prototype is as expected, otherwise the rest of the
1406 // function is probably invalid and likely to abort.
1407 if (!isTrigLibCall(CI))
1410 Value *Arg = CI->getArgOperand(0);
1411 SmallVector<CallInst *, 1> SinCalls;
1412 SmallVector<CallInst *, 1> CosCalls;
1413 SmallVector<CallInst *, 1> SinCosCalls;
1415 bool IsFloat = Arg->getType()->isFloatTy();
1417 // Look for all compatible sinpi, cospi and sincospi calls with the same
1418 // argument. If there are enough (in some sense) we can make the
1420 Function *F = CI->getFunction();
1421 for (User *U : Arg->users())
1422 classifyArgUse(U, F, IsFloat, SinCalls, CosCalls, SinCosCalls);
1424 // It's only worthwhile if both sinpi and cospi are actually used.
1425 if (SinCosCalls.empty() && (SinCalls.empty() || CosCalls.empty()))
1428 Value *Sin, *Cos, *SinCos;
1429 insertSinCosCall(B, CI->getCalledFunction(), Arg, IsFloat, Sin, Cos, SinCos);
1431 replaceTrigInsts(SinCalls, Sin);
1432 replaceTrigInsts(CosCalls, Cos);
1433 replaceTrigInsts(SinCosCalls, SinCos);
1438 void LibCallSimplifier::classifyArgUse(
1439 Value *Val, Function *F, bool IsFloat,
1440 SmallVectorImpl<CallInst *> &SinCalls,
1441 SmallVectorImpl<CallInst *> &CosCalls,
1442 SmallVectorImpl<CallInst *> &SinCosCalls) {
1443 CallInst *CI = dyn_cast<CallInst>(Val);
1448 // Don't consider calls in other functions.
1449 if (CI->getFunction() != F)
1452 Function *Callee = CI->getCalledFunction();
1454 if (!Callee || !TLI->getLibFunc(*Callee, Func) || !TLI->has(Func) ||
1459 if (Func == LibFunc::sinpif)
1460 SinCalls.push_back(CI);
1461 else if (Func == LibFunc::cospif)
1462 CosCalls.push_back(CI);
1463 else if (Func == LibFunc::sincospif_stret)
1464 SinCosCalls.push_back(CI);
1466 if (Func == LibFunc::sinpi)
1467 SinCalls.push_back(CI);
1468 else if (Func == LibFunc::cospi)
1469 CosCalls.push_back(CI);
1470 else if (Func == LibFunc::sincospi_stret)
1471 SinCosCalls.push_back(CI);
1475 void LibCallSimplifier::replaceTrigInsts(SmallVectorImpl<CallInst *> &Calls,
1477 for (CallInst *C : Calls)
1478 replaceAllUsesWith(C, Res);
1481 //===----------------------------------------------------------------------===//
1482 // Integer Library Call Optimizations
1483 //===----------------------------------------------------------------------===//
1485 Value *LibCallSimplifier::optimizeFFS(CallInst *CI, IRBuilder<> &B) {
1486 Function *Callee = CI->getCalledFunction();
1487 Value *Op = CI->getArgOperand(0);
1490 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
1491 if (CI->isZero()) // ffs(0) -> 0.
1492 return B.getInt32(0);
1493 // ffs(c) -> cttz(c)+1
1494 return B.getInt32(CI->getValue().countTrailingZeros() + 1);
1497 // ffs(x) -> x != 0 ? (i32)llvm.cttz(x)+1 : 0
1498 Type *ArgType = Op->getType();
1500 Intrinsic::getDeclaration(Callee->getParent(), Intrinsic::cttz, ArgType);
1501 Value *V = B.CreateCall(F, {Op, B.getTrue()}, "cttz");
1502 V = B.CreateAdd(V, ConstantInt::get(V->getType(), 1));
1503 V = B.CreateIntCast(V, B.getInt32Ty(), false);
1505 Value *Cond = B.CreateICmpNE(Op, Constant::getNullValue(ArgType));
1506 return B.CreateSelect(Cond, V, B.getInt32(0));
1509 Value *LibCallSimplifier::optimizeAbs(CallInst *CI, IRBuilder<> &B) {
1510 // abs(x) -> x >s -1 ? x : -x
1511 Value *Op = CI->getArgOperand(0);
1513 B.CreateICmpSGT(Op, Constant::getAllOnesValue(Op->getType()), "ispos");
1514 Value *Neg = B.CreateNeg(Op, "neg");
1515 return B.CreateSelect(Pos, Op, Neg);
1518 Value *LibCallSimplifier::optimizeIsDigit(CallInst *CI, IRBuilder<> &B) {
1519 // isdigit(c) -> (c-'0') <u 10
1520 Value *Op = CI->getArgOperand(0);
1521 Op = B.CreateSub(Op, B.getInt32('0'), "isdigittmp");
1522 Op = B.CreateICmpULT(Op, B.getInt32(10), "isdigit");
1523 return B.CreateZExt(Op, CI->getType());
1526 Value *LibCallSimplifier::optimizeIsAscii(CallInst *CI, IRBuilder<> &B) {
1527 // isascii(c) -> c <u 128
1528 Value *Op = CI->getArgOperand(0);
1529 Op = B.CreateICmpULT(Op, B.getInt32(128), "isascii");
1530 return B.CreateZExt(Op, CI->getType());
1533 Value *LibCallSimplifier::optimizeToAscii(CallInst *CI, IRBuilder<> &B) {
1534 // toascii(c) -> c & 0x7f
1535 return B.CreateAnd(CI->getArgOperand(0),
1536 ConstantInt::get(CI->getType(), 0x7F));
1539 //===----------------------------------------------------------------------===//
1540 // Formatting and IO Library Call Optimizations
1541 //===----------------------------------------------------------------------===//
1543 static bool isReportingError(Function *Callee, CallInst *CI, int StreamArg);
1545 Value *LibCallSimplifier::optimizeErrorReporting(CallInst *CI, IRBuilder<> &B,
1547 Function *Callee = CI->getCalledFunction();
1548 // Error reporting calls should be cold, mark them as such.
1549 // This applies even to non-builtin calls: it is only a hint and applies to
1550 // functions that the frontend might not understand as builtins.
1552 // This heuristic was suggested in:
1553 // Improving Static Branch Prediction in a Compiler
1554 // Brian L. Deitrich, Ben-Chung Cheng, Wen-mei W. Hwu
1555 // Proceedings of PACT'98, Oct. 1998, IEEE
1556 if (!CI->hasFnAttr(Attribute::Cold) &&
1557 isReportingError(Callee, CI, StreamArg)) {
1558 CI->addAttribute(AttributeSet::FunctionIndex, Attribute::Cold);
1564 static bool isReportingError(Function *Callee, CallInst *CI, int StreamArg) {
1565 if (!ColdErrorCalls || !Callee || !Callee->isDeclaration())
1571 // These functions might be considered cold, but only if their stream
1572 // argument is stderr.
1574 if (StreamArg >= (int)CI->getNumArgOperands())
1576 LoadInst *LI = dyn_cast<LoadInst>(CI->getArgOperand(StreamArg));
1579 GlobalVariable *GV = dyn_cast<GlobalVariable>(LI->getPointerOperand());
1580 if (!GV || !GV->isDeclaration())
1582 return GV->getName() == "stderr";
1585 Value *LibCallSimplifier::optimizePrintFString(CallInst *CI, IRBuilder<> &B) {
1586 // Check for a fixed format string.
1587 StringRef FormatStr;
1588 if (!getConstantStringInfo(CI->getArgOperand(0), FormatStr))
1591 // Empty format string -> noop.
1592 if (FormatStr.empty()) // Tolerate printf's declared void.
1593 return CI->use_empty() ? (Value *)CI : ConstantInt::get(CI->getType(), 0);
1595 // Do not do any of the following transformations if the printf return value
1596 // is used, in general the printf return value is not compatible with either
1597 // putchar() or puts().
1598 if (!CI->use_empty())
1601 // printf("x") -> putchar('x'), even for "%" and "%%".
1602 if (FormatStr.size() == 1 || FormatStr == "%%")
1603 return emitPutChar(B.getInt32(FormatStr[0]), B, TLI);
1605 // printf("%s", "a") --> putchar('a')
1606 if (FormatStr == "%s" && CI->getNumArgOperands() > 1) {
1608 if (!getConstantStringInfo(CI->getOperand(1), ChrStr))
1610 if (ChrStr.size() != 1)
1612 return emitPutChar(B.getInt32(ChrStr[0]), B, TLI);
1615 // printf("foo\n") --> puts("foo")
1616 if (FormatStr[FormatStr.size() - 1] == '\n' &&
1617 FormatStr.find('%') == StringRef::npos) { // No format characters.
1618 // Create a string literal with no \n on it. We expect the constant merge
1619 // pass to be run after this pass, to merge duplicate strings.
1620 FormatStr = FormatStr.drop_back();
1621 Value *GV = B.CreateGlobalString(FormatStr, "str");
1622 return emitPutS(GV, B, TLI);
1625 // Optimize specific format strings.
1626 // printf("%c", chr) --> putchar(chr)
1627 if (FormatStr == "%c" && CI->getNumArgOperands() > 1 &&
1628 CI->getArgOperand(1)->getType()->isIntegerTy())
1629 return emitPutChar(CI->getArgOperand(1), B, TLI);
1631 // printf("%s\n", str) --> puts(str)
1632 if (FormatStr == "%s\n" && CI->getNumArgOperands() > 1 &&
1633 CI->getArgOperand(1)->getType()->isPointerTy())
1634 return emitPutS(CI->getArgOperand(1), B, TLI);
1638 Value *LibCallSimplifier::optimizePrintF(CallInst *CI, IRBuilder<> &B) {
1640 Function *Callee = CI->getCalledFunction();
1641 FunctionType *FT = Callee->getFunctionType();
1642 if (Value *V = optimizePrintFString(CI, B)) {
1646 // printf(format, ...) -> iprintf(format, ...) if no floating point
1648 if (TLI->has(LibFunc::iprintf) && !callHasFloatingPointArgument(CI)) {
1649 Module *M = B.GetInsertBlock()->getParent()->getParent();
1650 Constant *IPrintFFn =
1651 M->getOrInsertFunction("iprintf", FT, Callee->getAttributes());
1652 CallInst *New = cast<CallInst>(CI->clone());
1653 New->setCalledFunction(IPrintFFn);
1660 Value *LibCallSimplifier::optimizeSPrintFString(CallInst *CI, IRBuilder<> &B) {
1661 // Check for a fixed format string.
1662 StringRef FormatStr;
1663 if (!getConstantStringInfo(CI->getArgOperand(1), FormatStr))
1666 // If we just have a format string (nothing else crazy) transform it.
1667 if (CI->getNumArgOperands() == 2) {
1668 // Make sure there's no % in the constant array. We could try to handle
1669 // %% -> % in the future if we cared.
1670 for (unsigned i = 0, e = FormatStr.size(); i != e; ++i)
1671 if (FormatStr[i] == '%')
1672 return nullptr; // we found a format specifier, bail out.
1674 // sprintf(str, fmt) -> llvm.memcpy(str, fmt, strlen(fmt)+1, 1)
1675 B.CreateMemCpy(CI->getArgOperand(0), CI->getArgOperand(1),
1676 ConstantInt::get(DL.getIntPtrType(CI->getContext()),
1677 FormatStr.size() + 1),
1678 1); // Copy the null byte.
1679 return ConstantInt::get(CI->getType(), FormatStr.size());
1682 // The remaining optimizations require the format string to be "%s" or "%c"
1683 // and have an extra operand.
1684 if (FormatStr.size() != 2 || FormatStr[0] != '%' ||
1685 CI->getNumArgOperands() < 3)
1688 // Decode the second character of the format string.
1689 if (FormatStr[1] == 'c') {
1690 // sprintf(dst, "%c", chr) --> *(i8*)dst = chr; *((i8*)dst+1) = 0
1691 if (!CI->getArgOperand(2)->getType()->isIntegerTy())
1693 Value *V = B.CreateTrunc(CI->getArgOperand(2), B.getInt8Ty(), "char");
1694 Value *Ptr = castToCStr(CI->getArgOperand(0), B);
1695 B.CreateStore(V, Ptr);
1696 Ptr = B.CreateGEP(B.getInt8Ty(), Ptr, B.getInt32(1), "nul");
1697 B.CreateStore(B.getInt8(0), Ptr);
1699 return ConstantInt::get(CI->getType(), 1);
1702 if (FormatStr[1] == 's') {
1703 // sprintf(dest, "%s", str) -> llvm.memcpy(dest, str, strlen(str)+1, 1)
1704 if (!CI->getArgOperand(2)->getType()->isPointerTy())
1707 Value *Len = emitStrLen(CI->getArgOperand(2), B, DL, TLI);
1711 B.CreateAdd(Len, ConstantInt::get(Len->getType(), 1), "leninc");
1712 B.CreateMemCpy(CI->getArgOperand(0), CI->getArgOperand(2), IncLen, 1);
1714 // The sprintf result is the unincremented number of bytes in the string.
1715 return B.CreateIntCast(Len, CI->getType(), false);
1720 Value *LibCallSimplifier::optimizeSPrintF(CallInst *CI, IRBuilder<> &B) {
1721 Function *Callee = CI->getCalledFunction();
1722 FunctionType *FT = Callee->getFunctionType();
1723 if (Value *V = optimizeSPrintFString(CI, B)) {
1727 // sprintf(str, format, ...) -> siprintf(str, format, ...) if no floating
1729 if (TLI->has(LibFunc::siprintf) && !callHasFloatingPointArgument(CI)) {
1730 Module *M = B.GetInsertBlock()->getParent()->getParent();
1731 Constant *SIPrintFFn =
1732 M->getOrInsertFunction("siprintf", FT, Callee->getAttributes());
1733 CallInst *New = cast<CallInst>(CI->clone());
1734 New->setCalledFunction(SIPrintFFn);
1741 Value *LibCallSimplifier::optimizeFPrintFString(CallInst *CI, IRBuilder<> &B) {
1742 optimizeErrorReporting(CI, B, 0);
1744 // All the optimizations depend on the format string.
1745 StringRef FormatStr;
1746 if (!getConstantStringInfo(CI->getArgOperand(1), FormatStr))
1749 // Do not do any of the following transformations if the fprintf return
1750 // value is used, in general the fprintf return value is not compatible
1751 // with fwrite(), fputc() or fputs().
1752 if (!CI->use_empty())
1755 // fprintf(F, "foo") --> fwrite("foo", 3, 1, F)
1756 if (CI->getNumArgOperands() == 2) {
1757 for (unsigned i = 0, e = FormatStr.size(); i != e; ++i)
1758 if (FormatStr[i] == '%') // Could handle %% -> % if we cared.
1759 return nullptr; // We found a format specifier.
1762 CI->getArgOperand(1),
1763 ConstantInt::get(DL.getIntPtrType(CI->getContext()), FormatStr.size()),
1764 CI->getArgOperand(0), B, DL, TLI);
1767 // The remaining optimizations require the format string to be "%s" or "%c"
1768 // and have an extra operand.
1769 if (FormatStr.size() != 2 || FormatStr[0] != '%' ||
1770 CI->getNumArgOperands() < 3)
1773 // Decode the second character of the format string.
1774 if (FormatStr[1] == 'c') {
1775 // fprintf(F, "%c", chr) --> fputc(chr, F)
1776 if (!CI->getArgOperand(2)->getType()->isIntegerTy())
1778 return emitFPutC(CI->getArgOperand(2), CI->getArgOperand(0), B, TLI);
1781 if (FormatStr[1] == 's') {
1782 // fprintf(F, "%s", str) --> fputs(str, F)
1783 if (!CI->getArgOperand(2)->getType()->isPointerTy())
1785 return emitFPutS(CI->getArgOperand(2), CI->getArgOperand(0), B, TLI);
1790 Value *LibCallSimplifier::optimizeFPrintF(CallInst *CI, IRBuilder<> &B) {
1791 Function *Callee = CI->getCalledFunction();
1792 FunctionType *FT = Callee->getFunctionType();
1793 if (Value *V = optimizeFPrintFString(CI, B)) {
1797 // fprintf(stream, format, ...) -> fiprintf(stream, format, ...) if no
1798 // floating point arguments.
1799 if (TLI->has(LibFunc::fiprintf) && !callHasFloatingPointArgument(CI)) {
1800 Module *M = B.GetInsertBlock()->getParent()->getParent();
1801 Constant *FIPrintFFn =
1802 M->getOrInsertFunction("fiprintf", FT, Callee->getAttributes());
1803 CallInst *New = cast<CallInst>(CI->clone());
1804 New->setCalledFunction(FIPrintFFn);
1811 Value *LibCallSimplifier::optimizeFWrite(CallInst *CI, IRBuilder<> &B) {
1812 optimizeErrorReporting(CI, B, 3);
1814 // Get the element size and count.
1815 ConstantInt *SizeC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
1816 ConstantInt *CountC = dyn_cast<ConstantInt>(CI->getArgOperand(2));
1817 if (!SizeC || !CountC)
1819 uint64_t Bytes = SizeC->getZExtValue() * CountC->getZExtValue();
1821 // If this is writing zero records, remove the call (it's a noop).
1823 return ConstantInt::get(CI->getType(), 0);
1825 // If this is writing one byte, turn it into fputc.
1826 // This optimisation is only valid, if the return value is unused.
1827 if (Bytes == 1 && CI->use_empty()) { // fwrite(S,1,1,F) -> fputc(S[0],F)
1828 Value *Char = B.CreateLoad(castToCStr(CI->getArgOperand(0), B), "char");
1829 Value *NewCI = emitFPutC(Char, CI->getArgOperand(3), B, TLI);
1830 return NewCI ? ConstantInt::get(CI->getType(), 1) : nullptr;
1836 Value *LibCallSimplifier::optimizeFPuts(CallInst *CI, IRBuilder<> &B) {
1837 optimizeErrorReporting(CI, B, 1);
1839 // Don't rewrite fputs to fwrite when optimising for size because fwrite
1840 // requires more arguments and thus extra MOVs are required.
1841 if (CI->getParent()->getParent()->optForSize())
1844 // We can't optimize if return value is used.
1845 if (!CI->use_empty())
1848 // fputs(s,F) --> fwrite(s,1,strlen(s),F)
1849 uint64_t Len = GetStringLength(CI->getArgOperand(0));
1853 // Known to have no uses (see above).
1855 CI->getArgOperand(0),
1856 ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len - 1),
1857 CI->getArgOperand(1), B, DL, TLI);
1860 Value *LibCallSimplifier::optimizePuts(CallInst *CI, IRBuilder<> &B) {
1861 // Check for a constant string.
1863 if (!getConstantStringInfo(CI->getArgOperand(0), Str))
1866 if (Str.empty() && CI->use_empty()) {
1867 // puts("") -> putchar('\n')
1868 Value *Res = emitPutChar(B.getInt32('\n'), B, TLI);
1869 if (CI->use_empty() || !Res)
1871 return B.CreateIntCast(Res, CI->getType(), true);
1877 bool LibCallSimplifier::hasFloatVersion(StringRef FuncName) {
1879 SmallString<20> FloatFuncName = FuncName;
1880 FloatFuncName += 'f';
1881 if (TLI->getLibFunc(FloatFuncName, Func))
1882 return TLI->has(Func);
1886 Value *LibCallSimplifier::optimizeStringMemoryLibCall(CallInst *CI,
1887 IRBuilder<> &Builder) {
1889 Function *Callee = CI->getCalledFunction();
1890 // Check for string/memory library functions.
1891 if (TLI->getLibFunc(*Callee, Func) && TLI->has(Func)) {
1892 // Make sure we never change the calling convention.
1893 assert((ignoreCallingConv(Func) ||
1894 CI->getCallingConv() == llvm::CallingConv::C) &&
1895 "Optimizing string/memory libcall would change the calling convention");
1897 case LibFunc::strcat:
1898 return optimizeStrCat(CI, Builder);
1899 case LibFunc::strncat:
1900 return optimizeStrNCat(CI, Builder);
1901 case LibFunc::strchr:
1902 return optimizeStrChr(CI, Builder);
1903 case LibFunc::strrchr:
1904 return optimizeStrRChr(CI, Builder);
1905 case LibFunc::strcmp:
1906 return optimizeStrCmp(CI, Builder);
1907 case LibFunc::strncmp:
1908 return optimizeStrNCmp(CI, Builder);
1909 case LibFunc::strcpy:
1910 return optimizeStrCpy(CI, Builder);
1911 case LibFunc::stpcpy:
1912 return optimizeStpCpy(CI, Builder);
1913 case LibFunc::strncpy:
1914 return optimizeStrNCpy(CI, Builder);
1915 case LibFunc::strlen:
1916 return optimizeStrLen(CI, Builder);
1917 case LibFunc::strpbrk:
1918 return optimizeStrPBrk(CI, Builder);
1919 case LibFunc::strtol:
1920 case LibFunc::strtod:
1921 case LibFunc::strtof:
1922 case LibFunc::strtoul:
1923 case LibFunc::strtoll:
1924 case LibFunc::strtold:
1925 case LibFunc::strtoull:
1926 return optimizeStrTo(CI, Builder);
1927 case LibFunc::strspn:
1928 return optimizeStrSpn(CI, Builder);
1929 case LibFunc::strcspn:
1930 return optimizeStrCSpn(CI, Builder);
1931 case LibFunc::strstr:
1932 return optimizeStrStr(CI, Builder);
1933 case LibFunc::memchr:
1934 return optimizeMemChr(CI, Builder);
1935 case LibFunc::memcmp:
1936 return optimizeMemCmp(CI, Builder);
1937 case LibFunc::memcpy:
1938 return optimizeMemCpy(CI, Builder);
1939 case LibFunc::memmove:
1940 return optimizeMemMove(CI, Builder);
1941 case LibFunc::memset:
1942 return optimizeMemSet(CI, Builder);
1950 Value *LibCallSimplifier::optimizeCall(CallInst *CI) {
1951 if (CI->isNoBuiltin())
1955 Function *Callee = CI->getCalledFunction();
1956 StringRef FuncName = Callee->getName();
1958 SmallVector<OperandBundleDef, 2> OpBundles;
1959 CI->getOperandBundlesAsDefs(OpBundles);
1960 IRBuilder<> Builder(CI, /*FPMathTag=*/nullptr, OpBundles);
1961 bool isCallingConvC = CI->getCallingConv() == llvm::CallingConv::C;
1963 // Command-line parameter overrides instruction attribute.
1964 if (EnableUnsafeFPShrink.getNumOccurrences() > 0)
1965 UnsafeFPShrink = EnableUnsafeFPShrink;
1966 else if (isa<FPMathOperator>(CI) && CI->hasUnsafeAlgebra())
1967 UnsafeFPShrink = true;
1969 // First, check for intrinsics.
1970 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
1971 if (!isCallingConvC)
1973 switch (II->getIntrinsicID()) {
1974 case Intrinsic::pow:
1975 return optimizePow(CI, Builder);
1976 case Intrinsic::exp2:
1977 return optimizeExp2(CI, Builder);
1978 case Intrinsic::fabs:
1979 return optimizeFabs(CI, Builder);
1980 case Intrinsic::log:
1981 return optimizeLog(CI, Builder);
1982 case Intrinsic::sqrt:
1983 return optimizeSqrt(CI, Builder);
1984 // TODO: Use foldMallocMemset() with memset intrinsic.
1990 // Also try to simplify calls to fortified library functions.
1991 if (Value *SimplifiedFortifiedCI = FortifiedSimplifier.optimizeCall(CI)) {
1992 // Try to further simplify the result.
1993 CallInst *SimplifiedCI = dyn_cast<CallInst>(SimplifiedFortifiedCI);
1994 if (SimplifiedCI && SimplifiedCI->getCalledFunction()) {
1995 // Use an IR Builder from SimplifiedCI if available instead of CI
1996 // to guarantee we reach all uses we might replace later on.
1997 IRBuilder<> TmpBuilder(SimplifiedCI);
1998 if (Value *V = optimizeStringMemoryLibCall(SimplifiedCI, TmpBuilder)) {
1999 // If we were able to further simplify, remove the now redundant call.
2000 SimplifiedCI->replaceAllUsesWith(V);
2001 SimplifiedCI->eraseFromParent();
2005 return SimplifiedFortifiedCI;
2008 // Then check for known library functions.
2009 if (TLI->getLibFunc(*Callee, Func) && TLI->has(Func)) {
2010 // We never change the calling convention.
2011 if (!ignoreCallingConv(Func) && !isCallingConvC)
2013 if (Value *V = optimizeStringMemoryLibCall(CI, Builder))
2019 return optimizeCos(CI, Builder);
2020 case LibFunc::sinpif:
2021 case LibFunc::sinpi:
2022 case LibFunc::cospif:
2023 case LibFunc::cospi:
2024 return optimizeSinCosPi(CI, Builder);
2028 return optimizePow(CI, Builder);
2029 case LibFunc::exp2l:
2031 case LibFunc::exp2f:
2032 return optimizeExp2(CI, Builder);
2033 case LibFunc::fabsf:
2035 case LibFunc::fabsl:
2036 return optimizeFabs(CI, Builder);
2037 case LibFunc::sqrtf:
2039 case LibFunc::sqrtl:
2040 return optimizeSqrt(CI, Builder);
2043 case LibFunc::ffsll:
2044 return optimizeFFS(CI, Builder);
2047 case LibFunc::llabs:
2048 return optimizeAbs(CI, Builder);
2049 case LibFunc::isdigit:
2050 return optimizeIsDigit(CI, Builder);
2051 case LibFunc::isascii:
2052 return optimizeIsAscii(CI, Builder);
2053 case LibFunc::toascii:
2054 return optimizeToAscii(CI, Builder);
2055 case LibFunc::printf:
2056 return optimizePrintF(CI, Builder);
2057 case LibFunc::sprintf:
2058 return optimizeSPrintF(CI, Builder);
2059 case LibFunc::fprintf:
2060 return optimizeFPrintF(CI, Builder);
2061 case LibFunc::fwrite:
2062 return optimizeFWrite(CI, Builder);
2063 case LibFunc::fputs:
2064 return optimizeFPuts(CI, Builder);
2066 case LibFunc::log10:
2067 case LibFunc::log1p:
2070 return optimizeLog(CI, Builder);
2072 return optimizePuts(CI, Builder);
2076 return optimizeTan(CI, Builder);
2077 case LibFunc::perror:
2078 return optimizeErrorReporting(CI, Builder);
2079 case LibFunc::vfprintf:
2080 case LibFunc::fiprintf:
2081 return optimizeErrorReporting(CI, Builder, 0);
2082 case LibFunc::fputc:
2083 return optimizeErrorReporting(CI, Builder, 1);
2085 case LibFunc::floor:
2087 case LibFunc::round:
2088 case LibFunc::nearbyint:
2089 case LibFunc::trunc:
2090 if (hasFloatVersion(FuncName))
2091 return optimizeUnaryDoubleFP(CI, Builder, false);
2094 case LibFunc::acosh:
2096 case LibFunc::asinh:
2098 case LibFunc::atanh:
2102 case LibFunc::exp10:
2103 case LibFunc::expm1:
2107 if (UnsafeFPShrink && hasFloatVersion(FuncName))
2108 return optimizeUnaryDoubleFP(CI, Builder, true);
2110 case LibFunc::copysign:
2111 if (hasFloatVersion(FuncName))
2112 return optimizeBinaryDoubleFP(CI, Builder);
2114 case LibFunc::fminf:
2116 case LibFunc::fminl:
2117 case LibFunc::fmaxf:
2119 case LibFunc::fmaxl:
2120 return optimizeFMinFMax(CI, Builder);
2128 LibCallSimplifier::LibCallSimplifier(
2129 const DataLayout &DL, const TargetLibraryInfo *TLI,
2130 function_ref<void(Instruction *, Value *)> Replacer)
2131 : FortifiedSimplifier(TLI), DL(DL), TLI(TLI), UnsafeFPShrink(false),
2132 Replacer(Replacer) {}
2134 void LibCallSimplifier::replaceAllUsesWith(Instruction *I, Value *With) {
2135 // Indirect through the replacer used in this instance.
2140 // Additional cases that we need to add to this file:
2143 // * cbrt(expN(X)) -> expN(x/3)
2144 // * cbrt(sqrt(x)) -> pow(x,1/6)
2145 // * cbrt(cbrt(x)) -> pow(x,1/9)
2148 // * exp(log(x)) -> x
2151 // * log(exp(x)) -> x
2152 // * log(exp(y)) -> y*log(e)
2153 // * log(exp10(y)) -> y*log(10)
2154 // * log(sqrt(x)) -> 0.5*log(x)
2156 // lround, lroundf, lroundl:
2157 // * lround(cnst) -> cnst'
2160 // * pow(sqrt(x),y) -> pow(x,y*0.5)
2161 // * pow(pow(x,y),z)-> pow(x,y*z)
2163 // round, roundf, roundl:
2164 // * round(cnst) -> cnst'
2167 // * signbit(cnst) -> cnst'
2168 // * signbit(nncst) -> 0 (if pstv is a non-negative constant)
2170 // sqrt, sqrtf, sqrtl:
2171 // * sqrt(expN(x)) -> expN(x*0.5)
2172 // * sqrt(Nroot(x)) -> pow(x,1/(2*N))
2173 // * sqrt(pow(x,y)) -> pow(|x|,y*0.5)
2175 // trunc, truncf, truncl:
2176 // * trunc(cnst) -> cnst'
2180 //===----------------------------------------------------------------------===//
2181 // Fortified Library Call Optimizations
2182 //===----------------------------------------------------------------------===//
2184 bool FortifiedLibCallSimplifier::isFortifiedCallFoldable(CallInst *CI,
2188 if (CI->getArgOperand(ObjSizeOp) == CI->getArgOperand(SizeOp))
2190 if (ConstantInt *ObjSizeCI =
2191 dyn_cast<ConstantInt>(CI->getArgOperand(ObjSizeOp))) {
2192 if (ObjSizeCI->isAllOnesValue())
2194 // If the object size wasn't -1 (unknown), bail out if we were asked to.
2195 if (OnlyLowerUnknownSize)
2198 uint64_t Len = GetStringLength(CI->getArgOperand(SizeOp));
2199 // If the length is 0 we don't know how long it is and so we can't
2200 // remove the check.
2203 return ObjSizeCI->getZExtValue() >= Len;
2205 if (ConstantInt *SizeCI = dyn_cast<ConstantInt>(CI->getArgOperand(SizeOp)))
2206 return ObjSizeCI->getZExtValue() >= SizeCI->getZExtValue();
2211 Value *FortifiedLibCallSimplifier::optimizeMemCpyChk(CallInst *CI,
2213 if (isFortifiedCallFoldable(CI, 3, 2, false)) {
2214 B.CreateMemCpy(CI->getArgOperand(0), CI->getArgOperand(1),
2215 CI->getArgOperand(2), 1);
2216 return CI->getArgOperand(0);
2221 Value *FortifiedLibCallSimplifier::optimizeMemMoveChk(CallInst *CI,
2223 if (isFortifiedCallFoldable(CI, 3, 2, false)) {
2224 B.CreateMemMove(CI->getArgOperand(0), CI->getArgOperand(1),
2225 CI->getArgOperand(2), 1);
2226 return CI->getArgOperand(0);
2231 Value *FortifiedLibCallSimplifier::optimizeMemSetChk(CallInst *CI,
2233 // TODO: Try foldMallocMemset() here.
2235 if (isFortifiedCallFoldable(CI, 3, 2, false)) {
2236 Value *Val = B.CreateIntCast(CI->getArgOperand(1), B.getInt8Ty(), false);
2237 B.CreateMemSet(CI->getArgOperand(0), Val, CI->getArgOperand(2), 1);
2238 return CI->getArgOperand(0);
2243 Value *FortifiedLibCallSimplifier::optimizeStrpCpyChk(CallInst *CI,
2245 LibFunc::Func Func) {
2246 Function *Callee = CI->getCalledFunction();
2247 StringRef Name = Callee->getName();
2248 const DataLayout &DL = CI->getModule()->getDataLayout();
2249 Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1),
2250 *ObjSize = CI->getArgOperand(2);
2252 // __stpcpy_chk(x,x,...) -> x+strlen(x)
2253 if (Func == LibFunc::stpcpy_chk && !OnlyLowerUnknownSize && Dst == Src) {
2254 Value *StrLen = emitStrLen(Src, B, DL, TLI);
2255 return StrLen ? B.CreateInBoundsGEP(B.getInt8Ty(), Dst, StrLen) : nullptr;
2258 // If a) we don't have any length information, or b) we know this will
2259 // fit then just lower to a plain st[rp]cpy. Otherwise we'll keep our
2260 // st[rp]cpy_chk call which may fail at runtime if the size is too long.
2261 // TODO: It might be nice to get a maximum length out of the possible
2262 // string lengths for varying.
2263 if (isFortifiedCallFoldable(CI, 2, 1, true))
2264 return emitStrCpy(Dst, Src, B, TLI, Name.substr(2, 6));
2266 if (OnlyLowerUnknownSize)
2269 // Maybe we can stil fold __st[rp]cpy_chk to __memcpy_chk.
2270 uint64_t Len = GetStringLength(Src);
2274 Type *SizeTTy = DL.getIntPtrType(CI->getContext());
2275 Value *LenV = ConstantInt::get(SizeTTy, Len);
2276 Value *Ret = emitMemCpyChk(Dst, Src, LenV, ObjSize, B, DL, TLI);
2277 // If the function was an __stpcpy_chk, and we were able to fold it into
2278 // a __memcpy_chk, we still need to return the correct end pointer.
2279 if (Ret && Func == LibFunc::stpcpy_chk)
2280 return B.CreateGEP(B.getInt8Ty(), Dst, ConstantInt::get(SizeTTy, Len - 1));
2284 Value *FortifiedLibCallSimplifier::optimizeStrpNCpyChk(CallInst *CI,
2286 LibFunc::Func Func) {
2287 Function *Callee = CI->getCalledFunction();
2288 StringRef Name = Callee->getName();
2289 if (isFortifiedCallFoldable(CI, 3, 2, false)) {
2290 Value *Ret = emitStrNCpy(CI->getArgOperand(0), CI->getArgOperand(1),
2291 CI->getArgOperand(2), B, TLI, Name.substr(2, 7));
2297 Value *FortifiedLibCallSimplifier::optimizeCall(CallInst *CI) {
2298 // FIXME: We shouldn't be changing "nobuiltin" or TLI unavailable calls here.
2299 // Some clang users checked for _chk libcall availability using:
2300 // __has_builtin(__builtin___memcpy_chk)
2301 // When compiling with -fno-builtin, this is always true.
2302 // When passing -ffreestanding/-mkernel, which both imply -fno-builtin, we
2303 // end up with fortified libcalls, which isn't acceptable in a freestanding
2304 // environment which only provides their non-fortified counterparts.
2306 // Until we change clang and/or teach external users to check for availability
2307 // differently, disregard the "nobuiltin" attribute and TLI::has.
2312 Function *Callee = CI->getCalledFunction();
2314 SmallVector<OperandBundleDef, 2> OpBundles;
2315 CI->getOperandBundlesAsDefs(OpBundles);
2316 IRBuilder<> Builder(CI, /*FPMathTag=*/nullptr, OpBundles);
2317 bool isCallingConvC = CI->getCallingConv() == llvm::CallingConv::C;
2319 // First, check that this is a known library functions and that the prototype
2321 if (!TLI->getLibFunc(*Callee, Func))
2324 // We never change the calling convention.
2325 if (!ignoreCallingConv(Func) && !isCallingConvC)
2329 case LibFunc::memcpy_chk:
2330 return optimizeMemCpyChk(CI, Builder);
2331 case LibFunc::memmove_chk:
2332 return optimizeMemMoveChk(CI, Builder);
2333 case LibFunc::memset_chk:
2334 return optimizeMemSetChk(CI, Builder);
2335 case LibFunc::stpcpy_chk:
2336 case LibFunc::strcpy_chk:
2337 return optimizeStrpCpyChk(CI, Builder, Func);
2338 case LibFunc::stpncpy_chk:
2339 case LibFunc::strncpy_chk:
2340 return optimizeStrpNCpyChk(CI, Builder, Func);
2347 FortifiedLibCallSimplifier::FortifiedLibCallSimplifier(
2348 const TargetLibraryInfo *TLI, bool OnlyLowerUnknownSize)
2349 : TLI(TLI), OnlyLowerUnknownSize(OnlyLowerUnknownSize) {}