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/ValueTracking.h"
22 #include "llvm/IR/DataLayout.h"
23 #include "llvm/IR/DiagnosticInfo.h"
24 #include "llvm/IR/Function.h"
25 #include "llvm/IR/IRBuilder.h"
26 #include "llvm/IR/IntrinsicInst.h"
27 #include "llvm/IR/Intrinsics.h"
28 #include "llvm/IR/LLVMContext.h"
29 #include "llvm/IR/Module.h"
30 #include "llvm/IR/PatternMatch.h"
31 #include "llvm/Support/Allocator.h"
32 #include "llvm/Support/CommandLine.h"
33 #include "llvm/Target/TargetLibraryInfo.h"
34 #include "llvm/Transforms/Utils/BuildLibCalls.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) {
64 llvm_unreachable("All cases should be covered in the switch.");
67 /// isOnlyUsedInZeroEqualityComparison - Return true if it only matters that the
68 /// value is equal or not-equal to zero.
69 static bool isOnlyUsedInZeroEqualityComparison(Value *V) {
70 for (User *U : V->users()) {
71 if (ICmpInst *IC = dyn_cast<ICmpInst>(U))
73 if (Constant *C = dyn_cast<Constant>(IC->getOperand(1)))
76 // Unknown instruction.
82 /// isOnlyUsedInEqualityComparison - Return true if it is only used in equality
83 /// comparisons with With.
84 static bool isOnlyUsedInEqualityComparison(Value *V, Value *With) {
85 for (User *U : V->users()) {
86 if (ICmpInst *IC = dyn_cast<ICmpInst>(U))
87 if (IC->isEquality() && IC->getOperand(1) == With)
89 // Unknown instruction.
95 static bool callHasFloatingPointArgument(const CallInst *CI) {
96 for (CallInst::const_op_iterator it = CI->op_begin(), e = CI->op_end();
98 if ((*it)->getType()->isFloatingPointTy())
104 /// \brief Check whether the overloaded unary floating point function
105 /// corresponing to \a Ty is available.
106 static bool hasUnaryFloatFn(const TargetLibraryInfo *TLI, Type *Ty,
107 LibFunc::Func DoubleFn, LibFunc::Func FloatFn,
108 LibFunc::Func LongDoubleFn) {
109 switch (Ty->getTypeID()) {
110 case Type::FloatTyID:
111 return TLI->has(FloatFn);
112 case Type::DoubleTyID:
113 return TLI->has(DoubleFn);
115 return TLI->has(LongDoubleFn);
119 //===----------------------------------------------------------------------===//
120 // Fortified Library Call Optimizations
121 //===----------------------------------------------------------------------===//
123 static bool isFortifiedCallFoldable(CallInst *CI, unsigned SizeCIOp, unsigned SizeArgOp,
125 if (CI->getArgOperand(SizeCIOp) == CI->getArgOperand(SizeArgOp))
127 if (ConstantInt *SizeCI =
128 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp))) {
129 if (SizeCI->isAllOnesValue())
132 uint64_t Len = GetStringLength(CI->getArgOperand(SizeArgOp));
133 // If the length is 0 we don't know how long it is and so we can't
137 return SizeCI->getZExtValue() >= Len;
139 if (ConstantInt *Arg = dyn_cast<ConstantInt>(CI->getArgOperand(SizeArgOp)))
140 return SizeCI->getZExtValue() >= Arg->getZExtValue();
145 Value *LibCallSimplifier::optimizeMemCpyChk(CallInst *CI, IRBuilder<> &B) {
146 Function *Callee = CI->getCalledFunction();
147 FunctionType *FT = Callee->getFunctionType();
148 LLVMContext &Context = CI->getContext();
150 // Check if this has the right signature.
151 if (FT->getNumParams() != 4 || FT->getReturnType() != FT->getParamType(0) ||
152 !FT->getParamType(0)->isPointerTy() ||
153 !FT->getParamType(1)->isPointerTy() ||
154 FT->getParamType(2) != DL->getIntPtrType(Context) ||
155 FT->getParamType(3) != DL->getIntPtrType(Context))
158 if (isFortifiedCallFoldable(CI, 3, 2, false)) {
159 B.CreateMemCpy(CI->getArgOperand(0), CI->getArgOperand(1),
160 CI->getArgOperand(2), 1);
161 return CI->getArgOperand(0);
166 Value *LibCallSimplifier::optimizeMemMoveChk(CallInst *CI, IRBuilder<> &B) {
167 Function *Callee = CI->getCalledFunction();
168 FunctionType *FT = Callee->getFunctionType();
169 LLVMContext &Context = CI->getContext();
171 // Check if this has the right signature.
172 if (FT->getNumParams() != 4 || FT->getReturnType() != FT->getParamType(0) ||
173 !FT->getParamType(0)->isPointerTy() ||
174 !FT->getParamType(1)->isPointerTy() ||
175 FT->getParamType(2) != DL->getIntPtrType(Context) ||
176 FT->getParamType(3) != DL->getIntPtrType(Context))
179 if (isFortifiedCallFoldable(CI, 3, 2, false)) {
180 B.CreateMemMove(CI->getArgOperand(0), CI->getArgOperand(1),
181 CI->getArgOperand(2), 1);
182 return CI->getArgOperand(0);
187 Value *LibCallSimplifier::optimizeMemSetChk(CallInst *CI, IRBuilder<> &B) {
188 Function *Callee = CI->getCalledFunction();
189 FunctionType *FT = Callee->getFunctionType();
190 LLVMContext &Context = CI->getContext();
192 // Check if this has the right signature.
193 if (FT->getNumParams() != 4 || FT->getReturnType() != FT->getParamType(0) ||
194 !FT->getParamType(0)->isPointerTy() ||
195 !FT->getParamType(1)->isIntegerTy() ||
196 FT->getParamType(2) != DL->getIntPtrType(Context) ||
197 FT->getParamType(3) != DL->getIntPtrType(Context))
200 if (isFortifiedCallFoldable(CI, 3, 2, false)) {
201 Value *Val = B.CreateIntCast(CI->getArgOperand(1), B.getInt8Ty(), false);
202 B.CreateMemSet(CI->getArgOperand(0), Val, CI->getArgOperand(2), 1);
203 return CI->getArgOperand(0);
208 Value *LibCallSimplifier::optimizeStrCpyChk(CallInst *CI, IRBuilder<> &B) {
209 Function *Callee = CI->getCalledFunction();
210 StringRef Name = Callee->getName();
211 FunctionType *FT = Callee->getFunctionType();
212 LLVMContext &Context = CI->getContext();
214 // Check if this has the right signature.
215 if (FT->getNumParams() != 3 || FT->getReturnType() != FT->getParamType(0) ||
216 FT->getParamType(0) != FT->getParamType(1) ||
217 FT->getParamType(0) != Type::getInt8PtrTy(Context) ||
218 FT->getParamType(2) != DL->getIntPtrType(Context))
221 Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1);
222 if (Dst == Src) // __strcpy_chk(x,x) -> x
225 // If a) we don't have any length information, or b) we know this will
226 // fit then just lower to a plain strcpy. Otherwise we'll keep our
227 // strcpy_chk call which may fail at runtime if the size is too long.
228 // TODO: It might be nice to get a maximum length out of the possible
229 // string lengths for varying.
230 if (isFortifiedCallFoldable(CI, 2, 1, true)) {
231 Value *Ret = EmitStrCpy(Dst, Src, B, DL, TLI, Name.substr(2, 6));
234 // Maybe we can stil fold __strcpy_chk to __memcpy_chk.
235 uint64_t Len = GetStringLength(Src);
239 // This optimization require DataLayout.
243 Value *Ret = EmitMemCpyChk(
244 Dst, Src, ConstantInt::get(DL->getIntPtrType(Context), Len),
245 CI->getArgOperand(2), B, DL, TLI);
251 Value *LibCallSimplifier::optimizeStpCpyChk(CallInst *CI, IRBuilder<> &B) {
252 Function *Callee = CI->getCalledFunction();
253 StringRef Name = Callee->getName();
254 FunctionType *FT = Callee->getFunctionType();
255 LLVMContext &Context = CI->getContext();
257 // Check if this has the right signature.
258 if (FT->getNumParams() != 3 || FT->getReturnType() != FT->getParamType(0) ||
259 FT->getParamType(0) != FT->getParamType(1) ||
260 FT->getParamType(0) != Type::getInt8PtrTy(Context) ||
261 FT->getParamType(2) != DL->getIntPtrType(FT->getParamType(0)))
264 Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1);
265 if (Dst == Src) { // stpcpy(x,x) -> x+strlen(x)
266 Value *StrLen = EmitStrLen(Src, B, DL, TLI);
267 return StrLen ? B.CreateInBoundsGEP(Dst, StrLen) : nullptr;
270 // If a) we don't have any length information, or b) we know this will
271 // fit then just lower to a plain stpcpy. Otherwise we'll keep our
272 // stpcpy_chk call which may fail at runtime if the size is too long.
273 // TODO: It might be nice to get a maximum length out of the possible
274 // string lengths for varying.
275 if (isFortifiedCallFoldable(CI, 2, 1, true)) {
276 Value *Ret = EmitStrCpy(Dst, Src, B, DL, TLI, Name.substr(2, 6));
279 // Maybe we can stil fold __stpcpy_chk to __memcpy_chk.
280 uint64_t Len = GetStringLength(Src);
284 // This optimization require DataLayout.
288 Type *PT = FT->getParamType(0);
289 Value *LenV = ConstantInt::get(DL->getIntPtrType(PT), Len);
291 B.CreateGEP(Dst, ConstantInt::get(DL->getIntPtrType(PT), Len - 1));
292 if (!EmitMemCpyChk(Dst, Src, LenV, CI->getArgOperand(2), B, DL, TLI))
299 Value *LibCallSimplifier::optimizeStrNCpyChk(CallInst *CI, IRBuilder<> &B) {
300 Function *Callee = CI->getCalledFunction();
301 StringRef Name = Callee->getName();
302 FunctionType *FT = Callee->getFunctionType();
303 LLVMContext &Context = CI->getContext();
305 // Check if this has the right signature.
306 if (FT->getNumParams() != 4 || FT->getReturnType() != FT->getParamType(0) ||
307 FT->getParamType(0) != FT->getParamType(1) ||
308 FT->getParamType(0) != Type::getInt8PtrTy(Context) ||
309 !FT->getParamType(2)->isIntegerTy() ||
310 FT->getParamType(3) != DL->getIntPtrType(Context))
313 if (isFortifiedCallFoldable(CI, 3, 2, false)) {
315 EmitStrNCpy(CI->getArgOperand(0), CI->getArgOperand(1),
316 CI->getArgOperand(2), B, DL, TLI, Name.substr(2, 7));
322 //===----------------------------------------------------------------------===//
323 // String and Memory Library Call Optimizations
324 //===----------------------------------------------------------------------===//
326 Value *LibCallSimplifier::optimizeStrCat(CallInst *CI, IRBuilder<> &B) {
327 Function *Callee = CI->getCalledFunction();
328 // Verify the "strcat" function prototype.
329 FunctionType *FT = Callee->getFunctionType();
330 if (FT->getNumParams() != 2||
331 FT->getReturnType() != B.getInt8PtrTy() ||
332 FT->getParamType(0) != FT->getReturnType() ||
333 FT->getParamType(1) != FT->getReturnType())
336 // Extract some information from the instruction
337 Value *Dst = CI->getArgOperand(0);
338 Value *Src = CI->getArgOperand(1);
340 // See if we can get the length of the input string.
341 uint64_t Len = GetStringLength(Src);
344 --Len; // Unbias length.
346 // Handle the simple, do-nothing case: strcat(x, "") -> x
350 // These optimizations require DataLayout.
354 return emitStrLenMemCpy(Src, Dst, Len, B);
357 Value *LibCallSimplifier::emitStrLenMemCpy(Value *Src, Value *Dst, uint64_t Len,
359 // We need to find the end of the destination string. That's where the
360 // memory is to be moved to. We just generate a call to strlen.
361 Value *DstLen = EmitStrLen(Dst, B, DL, TLI);
365 // Now that we have the destination's length, we must index into the
366 // destination's pointer to get the actual memcpy destination (end of
367 // the string .. we're concatenating).
368 Value *CpyDst = B.CreateGEP(Dst, DstLen, "endptr");
370 // We have enough information to now generate the memcpy call to do the
371 // concatenation for us. Make a memcpy to copy the nul byte with align = 1.
374 ConstantInt::get(DL->getIntPtrType(Src->getContext()), Len + 1), 1);
378 Value *LibCallSimplifier::optimizeStrNCat(CallInst *CI, IRBuilder<> &B) {
379 Function *Callee = CI->getCalledFunction();
380 // Verify the "strncat" function prototype.
381 FunctionType *FT = Callee->getFunctionType();
382 if (FT->getNumParams() != 3 || FT->getReturnType() != B.getInt8PtrTy() ||
383 FT->getParamType(0) != FT->getReturnType() ||
384 FT->getParamType(1) != FT->getReturnType() ||
385 !FT->getParamType(2)->isIntegerTy())
388 // Extract some information from the instruction
389 Value *Dst = CI->getArgOperand(0);
390 Value *Src = CI->getArgOperand(1);
393 // We don't do anything if length is not constant
394 if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(CI->getArgOperand(2)))
395 Len = LengthArg->getZExtValue();
399 // See if we can get the length of the input string.
400 uint64_t SrcLen = GetStringLength(Src);
403 --SrcLen; // Unbias length.
405 // Handle the simple, do-nothing cases:
406 // strncat(x, "", c) -> x
407 // strncat(x, c, 0) -> x
408 if (SrcLen == 0 || Len == 0)
411 // These optimizations require DataLayout.
415 // We don't optimize this case
419 // strncat(x, s, c) -> strcat(x, s)
420 // s is constant so the strcat can be optimized further
421 return emitStrLenMemCpy(Src, Dst, SrcLen, B);
424 Value *LibCallSimplifier::optimizeStrChr(CallInst *CI, IRBuilder<> &B) {
425 Function *Callee = CI->getCalledFunction();
426 // Verify the "strchr" function prototype.
427 FunctionType *FT = Callee->getFunctionType();
428 if (FT->getNumParams() != 2 || FT->getReturnType() != B.getInt8PtrTy() ||
429 FT->getParamType(0) != FT->getReturnType() ||
430 !FT->getParamType(1)->isIntegerTy(32))
433 Value *SrcStr = CI->getArgOperand(0);
435 // If the second operand is non-constant, see if we can compute the length
436 // of the input string and turn this into memchr.
437 ConstantInt *CharC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
439 // These optimizations require DataLayout.
443 uint64_t Len = GetStringLength(SrcStr);
444 if (Len == 0 || !FT->getParamType(1)->isIntegerTy(32)) // memchr needs i32.
448 SrcStr, CI->getArgOperand(1), // include nul.
449 ConstantInt::get(DL->getIntPtrType(CI->getContext()), Len), B, DL, TLI);
452 // Otherwise, the character is a constant, see if the first argument is
453 // a string literal. If so, we can constant fold.
455 if (!getConstantStringInfo(SrcStr, Str)) {
456 if (DL && CharC->isZero()) // strchr(p, 0) -> p + strlen(p)
457 return B.CreateGEP(SrcStr, EmitStrLen(SrcStr, B, DL, TLI), "strchr");
461 // Compute the offset, make sure to handle the case when we're searching for
462 // zero (a weird way to spell strlen).
463 size_t I = (0xFF & CharC->getSExtValue()) == 0
465 : Str.find(CharC->getSExtValue());
466 if (I == StringRef::npos) // Didn't find the char. strchr returns null.
467 return Constant::getNullValue(CI->getType());
469 // strchr(s+n,c) -> gep(s+n+i,c)
470 return B.CreateGEP(SrcStr, B.getInt64(I), "strchr");
473 Value *LibCallSimplifier::optimizeStrRChr(CallInst *CI, IRBuilder<> &B) {
474 Function *Callee = CI->getCalledFunction();
475 // Verify the "strrchr" function prototype.
476 FunctionType *FT = Callee->getFunctionType();
477 if (FT->getNumParams() != 2 || FT->getReturnType() != B.getInt8PtrTy() ||
478 FT->getParamType(0) != FT->getReturnType() ||
479 !FT->getParamType(1)->isIntegerTy(32))
482 Value *SrcStr = CI->getArgOperand(0);
483 ConstantInt *CharC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
485 // Cannot fold anything if we're not looking for a constant.
490 if (!getConstantStringInfo(SrcStr, Str)) {
491 // strrchr(s, 0) -> strchr(s, 0)
492 if (DL && CharC->isZero())
493 return EmitStrChr(SrcStr, '\0', B, DL, TLI);
497 // Compute the offset.
498 size_t I = (0xFF & CharC->getSExtValue()) == 0
500 : Str.rfind(CharC->getSExtValue());
501 if (I == StringRef::npos) // Didn't find the char. Return null.
502 return Constant::getNullValue(CI->getType());
504 // strrchr(s+n,c) -> gep(s+n+i,c)
505 return B.CreateGEP(SrcStr, B.getInt64(I), "strrchr");
508 Value *LibCallSimplifier::optimizeStrCmp(CallInst *CI, IRBuilder<> &B) {
509 Function *Callee = CI->getCalledFunction();
510 // Verify the "strcmp" function prototype.
511 FunctionType *FT = Callee->getFunctionType();
512 if (FT->getNumParams() != 2 || !FT->getReturnType()->isIntegerTy(32) ||
513 FT->getParamType(0) != FT->getParamType(1) ||
514 FT->getParamType(0) != B.getInt8PtrTy())
517 Value *Str1P = CI->getArgOperand(0), *Str2P = CI->getArgOperand(1);
518 if (Str1P == Str2P) // strcmp(x,x) -> 0
519 return ConstantInt::get(CI->getType(), 0);
521 StringRef Str1, Str2;
522 bool HasStr1 = getConstantStringInfo(Str1P, Str1);
523 bool HasStr2 = getConstantStringInfo(Str2P, Str2);
525 // strcmp(x, y) -> cnst (if both x and y are constant strings)
526 if (HasStr1 && HasStr2)
527 return ConstantInt::get(CI->getType(), Str1.compare(Str2));
529 if (HasStr1 && Str1.empty()) // strcmp("", x) -> -*x
531 B.CreateZExt(B.CreateLoad(Str2P, "strcmpload"), CI->getType()));
533 if (HasStr2 && Str2.empty()) // strcmp(x,"") -> *x
534 return B.CreateZExt(B.CreateLoad(Str1P, "strcmpload"), CI->getType());
536 // strcmp(P, "x") -> memcmp(P, "x", 2)
537 uint64_t Len1 = GetStringLength(Str1P);
538 uint64_t Len2 = GetStringLength(Str2P);
540 // These optimizations require DataLayout.
544 return EmitMemCmp(Str1P, Str2P,
545 ConstantInt::get(DL->getIntPtrType(CI->getContext()),
546 std::min(Len1, Len2)),
553 Value *LibCallSimplifier::optimizeStrNCmp(CallInst *CI, IRBuilder<> &B) {
554 Function *Callee = CI->getCalledFunction();
555 // Verify the "strncmp" function prototype.
556 FunctionType *FT = Callee->getFunctionType();
557 if (FT->getNumParams() != 3 || !FT->getReturnType()->isIntegerTy(32) ||
558 FT->getParamType(0) != FT->getParamType(1) ||
559 FT->getParamType(0) != B.getInt8PtrTy() ||
560 !FT->getParamType(2)->isIntegerTy())
563 Value *Str1P = CI->getArgOperand(0), *Str2P = CI->getArgOperand(1);
564 if (Str1P == Str2P) // strncmp(x,x,n) -> 0
565 return ConstantInt::get(CI->getType(), 0);
567 // Get the length argument if it is constant.
569 if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(CI->getArgOperand(2)))
570 Length = LengthArg->getZExtValue();
574 if (Length == 0) // strncmp(x,y,0) -> 0
575 return ConstantInt::get(CI->getType(), 0);
577 if (DL && Length == 1) // strncmp(x,y,1) -> memcmp(x,y,1)
578 return EmitMemCmp(Str1P, Str2P, CI->getArgOperand(2), B, DL, TLI);
580 StringRef Str1, Str2;
581 bool HasStr1 = getConstantStringInfo(Str1P, Str1);
582 bool HasStr2 = getConstantStringInfo(Str2P, Str2);
584 // strncmp(x, y) -> cnst (if both x and y are constant strings)
585 if (HasStr1 && HasStr2) {
586 StringRef SubStr1 = Str1.substr(0, Length);
587 StringRef SubStr2 = Str2.substr(0, Length);
588 return ConstantInt::get(CI->getType(), SubStr1.compare(SubStr2));
591 if (HasStr1 && Str1.empty()) // strncmp("", x, n) -> -*x
593 B.CreateZExt(B.CreateLoad(Str2P, "strcmpload"), CI->getType()));
595 if (HasStr2 && Str2.empty()) // strncmp(x, "", n) -> *x
596 return B.CreateZExt(B.CreateLoad(Str1P, "strcmpload"), CI->getType());
601 Value *LibCallSimplifier::optimizeStrCpy(CallInst *CI, IRBuilder<> &B) {
602 Function *Callee = CI->getCalledFunction();
603 // Verify the "strcpy" function prototype.
604 FunctionType *FT = Callee->getFunctionType();
605 if (FT->getNumParams() != 2 || FT->getReturnType() != FT->getParamType(0) ||
606 FT->getParamType(0) != FT->getParamType(1) ||
607 FT->getParamType(0) != B.getInt8PtrTy())
610 Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1);
611 if (Dst == Src) // strcpy(x,x) -> x
614 // These optimizations require DataLayout.
618 // See if we can get the length of the input string.
619 uint64_t Len = GetStringLength(Src);
623 // We have enough information to now generate the memcpy call to do the
624 // copy for us. Make a memcpy to copy the nul byte with align = 1.
625 B.CreateMemCpy(Dst, Src,
626 ConstantInt::get(DL->getIntPtrType(CI->getContext()), Len), 1);
630 Value *LibCallSimplifier::optimizeStpCpy(CallInst *CI, IRBuilder<> &B) {
631 Function *Callee = CI->getCalledFunction();
632 // Verify the "stpcpy" function prototype.
633 FunctionType *FT = Callee->getFunctionType();
634 if (FT->getNumParams() != 2 || FT->getReturnType() != FT->getParamType(0) ||
635 FT->getParamType(0) != FT->getParamType(1) ||
636 FT->getParamType(0) != B.getInt8PtrTy())
639 // These optimizations require DataLayout.
643 Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1);
644 if (Dst == Src) { // stpcpy(x,x) -> x+strlen(x)
645 Value *StrLen = EmitStrLen(Src, B, DL, TLI);
646 return StrLen ? B.CreateInBoundsGEP(Dst, StrLen) : nullptr;
649 // See if we can get the length of the input string.
650 uint64_t Len = GetStringLength(Src);
654 Type *PT = FT->getParamType(0);
655 Value *LenV = ConstantInt::get(DL->getIntPtrType(PT), Len);
657 B.CreateGEP(Dst, ConstantInt::get(DL->getIntPtrType(PT), Len - 1));
659 // We have enough information to now generate the memcpy call to do the
660 // copy for us. Make a memcpy to copy the nul byte with align = 1.
661 B.CreateMemCpy(Dst, Src, LenV, 1);
665 Value *LibCallSimplifier::optimizeStrNCpy(CallInst *CI, IRBuilder<> &B) {
666 Function *Callee = CI->getCalledFunction();
667 FunctionType *FT = Callee->getFunctionType();
668 if (FT->getNumParams() != 3 || FT->getReturnType() != FT->getParamType(0) ||
669 FT->getParamType(0) != FT->getParamType(1) ||
670 FT->getParamType(0) != B.getInt8PtrTy() ||
671 !FT->getParamType(2)->isIntegerTy())
674 Value *Dst = CI->getArgOperand(0);
675 Value *Src = CI->getArgOperand(1);
676 Value *LenOp = CI->getArgOperand(2);
678 // See if we can get the length of the input string.
679 uint64_t SrcLen = GetStringLength(Src);
685 // strncpy(x, "", y) -> memset(x, '\0', y, 1)
686 B.CreateMemSet(Dst, B.getInt8('\0'), LenOp, 1);
691 if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(LenOp))
692 Len = LengthArg->getZExtValue();
697 return Dst; // strncpy(x, y, 0) -> x
699 // These optimizations require DataLayout.
703 // Let strncpy handle the zero padding
704 if (Len > SrcLen + 1)
707 Type *PT = FT->getParamType(0);
708 // strncpy(x, s, c) -> memcpy(x, s, c, 1) [s and c are constant]
709 B.CreateMemCpy(Dst, Src, ConstantInt::get(DL->getIntPtrType(PT), Len), 1);
714 Value *LibCallSimplifier::optimizeStrLen(CallInst *CI, IRBuilder<> &B) {
715 Function *Callee = CI->getCalledFunction();
716 FunctionType *FT = Callee->getFunctionType();
717 if (FT->getNumParams() != 1 || FT->getParamType(0) != B.getInt8PtrTy() ||
718 !FT->getReturnType()->isIntegerTy())
721 Value *Src = CI->getArgOperand(0);
723 // Constant folding: strlen("xyz") -> 3
724 if (uint64_t Len = GetStringLength(Src))
725 return ConstantInt::get(CI->getType(), Len - 1);
727 // strlen(x?"foo":"bars") --> x ? 3 : 4
728 if (SelectInst *SI = dyn_cast<SelectInst>(Src)) {
729 uint64_t LenTrue = GetStringLength(SI->getTrueValue());
730 uint64_t LenFalse = GetStringLength(SI->getFalseValue());
731 if (LenTrue && LenFalse) {
732 Function *Caller = CI->getParent()->getParent();
733 emitOptimizationRemark(CI->getContext(), "simplify-libcalls", *Caller,
735 "folded strlen(select) to select of constants");
736 return B.CreateSelect(SI->getCondition(),
737 ConstantInt::get(CI->getType(), LenTrue - 1),
738 ConstantInt::get(CI->getType(), LenFalse - 1));
742 // strlen(x) != 0 --> *x != 0
743 // strlen(x) == 0 --> *x == 0
744 if (isOnlyUsedInZeroEqualityComparison(CI))
745 return B.CreateZExt(B.CreateLoad(Src, "strlenfirst"), CI->getType());
750 Value *LibCallSimplifier::optimizeStrPBrk(CallInst *CI, IRBuilder<> &B) {
751 Function *Callee = CI->getCalledFunction();
752 FunctionType *FT = Callee->getFunctionType();
753 if (FT->getNumParams() != 2 || FT->getParamType(0) != B.getInt8PtrTy() ||
754 FT->getParamType(1) != FT->getParamType(0) ||
755 FT->getReturnType() != FT->getParamType(0))
759 bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1);
760 bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2);
762 // strpbrk(s, "") -> nullptr
763 // strpbrk("", s) -> nullptr
764 if ((HasS1 && S1.empty()) || (HasS2 && S2.empty()))
765 return Constant::getNullValue(CI->getType());
768 if (HasS1 && HasS2) {
769 size_t I = S1.find_first_of(S2);
770 if (I == StringRef::npos) // No match.
771 return Constant::getNullValue(CI->getType());
773 return B.CreateGEP(CI->getArgOperand(0), B.getInt64(I), "strpbrk");
776 // strpbrk(s, "a") -> strchr(s, 'a')
777 if (DL && HasS2 && S2.size() == 1)
778 return EmitStrChr(CI->getArgOperand(0), S2[0], B, DL, TLI);
783 Value *LibCallSimplifier::optimizeStrTo(CallInst *CI, IRBuilder<> &B) {
784 Function *Callee = CI->getCalledFunction();
785 FunctionType *FT = Callee->getFunctionType();
786 if ((FT->getNumParams() != 2 && FT->getNumParams() != 3) ||
787 !FT->getParamType(0)->isPointerTy() ||
788 !FT->getParamType(1)->isPointerTy())
791 Value *EndPtr = CI->getArgOperand(1);
792 if (isa<ConstantPointerNull>(EndPtr)) {
793 // With a null EndPtr, this function won't capture the main argument.
794 // It would be readonly too, except that it still may write to errno.
795 CI->addAttribute(1, Attribute::NoCapture);
801 Value *LibCallSimplifier::optimizeStrSpn(CallInst *CI, IRBuilder<> &B) {
802 Function *Callee = CI->getCalledFunction();
803 FunctionType *FT = Callee->getFunctionType();
804 if (FT->getNumParams() != 2 || FT->getParamType(0) != B.getInt8PtrTy() ||
805 FT->getParamType(1) != FT->getParamType(0) ||
806 !FT->getReturnType()->isIntegerTy())
810 bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1);
811 bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2);
813 // strspn(s, "") -> 0
814 // strspn("", s) -> 0
815 if ((HasS1 && S1.empty()) || (HasS2 && S2.empty()))
816 return Constant::getNullValue(CI->getType());
819 if (HasS1 && HasS2) {
820 size_t Pos = S1.find_first_not_of(S2);
821 if (Pos == StringRef::npos)
823 return ConstantInt::get(CI->getType(), Pos);
829 Value *LibCallSimplifier::optimizeStrCSpn(CallInst *CI, IRBuilder<> &B) {
830 Function *Callee = CI->getCalledFunction();
831 FunctionType *FT = Callee->getFunctionType();
832 if (FT->getNumParams() != 2 || FT->getParamType(0) != B.getInt8PtrTy() ||
833 FT->getParamType(1) != FT->getParamType(0) ||
834 !FT->getReturnType()->isIntegerTy())
838 bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1);
839 bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2);
841 // strcspn("", s) -> 0
842 if (HasS1 && S1.empty())
843 return Constant::getNullValue(CI->getType());
846 if (HasS1 && HasS2) {
847 size_t Pos = S1.find_first_of(S2);
848 if (Pos == StringRef::npos)
850 return ConstantInt::get(CI->getType(), Pos);
853 // strcspn(s, "") -> strlen(s)
854 if (DL && HasS2 && S2.empty())
855 return EmitStrLen(CI->getArgOperand(0), B, DL, TLI);
860 Value *LibCallSimplifier::optimizeStrStr(CallInst *CI, IRBuilder<> &B) {
861 Function *Callee = CI->getCalledFunction();
862 FunctionType *FT = Callee->getFunctionType();
863 if (FT->getNumParams() != 2 || !FT->getParamType(0)->isPointerTy() ||
864 !FT->getParamType(1)->isPointerTy() ||
865 !FT->getReturnType()->isPointerTy())
868 // fold strstr(x, x) -> x.
869 if (CI->getArgOperand(0) == CI->getArgOperand(1))
870 return B.CreateBitCast(CI->getArgOperand(0), CI->getType());
872 // fold strstr(a, b) == a -> strncmp(a, b, strlen(b)) == 0
873 if (DL && isOnlyUsedInEqualityComparison(CI, CI->getArgOperand(0))) {
874 Value *StrLen = EmitStrLen(CI->getArgOperand(1), B, DL, TLI);
877 Value *StrNCmp = EmitStrNCmp(CI->getArgOperand(0), CI->getArgOperand(1),
881 for (auto UI = CI->user_begin(), UE = CI->user_end(); UI != UE;) {
882 ICmpInst *Old = cast<ICmpInst>(*UI++);
884 B.CreateICmp(Old->getPredicate(), StrNCmp,
885 ConstantInt::getNullValue(StrNCmp->getType()), "cmp");
886 replaceAllUsesWith(Old, Cmp);
891 // See if either input string is a constant string.
892 StringRef SearchStr, ToFindStr;
893 bool HasStr1 = getConstantStringInfo(CI->getArgOperand(0), SearchStr);
894 bool HasStr2 = getConstantStringInfo(CI->getArgOperand(1), ToFindStr);
896 // fold strstr(x, "") -> x.
897 if (HasStr2 && ToFindStr.empty())
898 return B.CreateBitCast(CI->getArgOperand(0), CI->getType());
900 // If both strings are known, constant fold it.
901 if (HasStr1 && HasStr2) {
902 size_t Offset = SearchStr.find(ToFindStr);
904 if (Offset == StringRef::npos) // strstr("foo", "bar") -> null
905 return Constant::getNullValue(CI->getType());
907 // strstr("abcd", "bc") -> gep((char*)"abcd", 1)
908 Value *Result = CastToCStr(CI->getArgOperand(0), B);
909 Result = B.CreateConstInBoundsGEP1_64(Result, Offset, "strstr");
910 return B.CreateBitCast(Result, CI->getType());
913 // fold strstr(x, "y") -> strchr(x, 'y').
914 if (HasStr2 && ToFindStr.size() == 1) {
915 Value *StrChr = EmitStrChr(CI->getArgOperand(0), ToFindStr[0], B, DL, TLI);
916 return StrChr ? B.CreateBitCast(StrChr, CI->getType()) : nullptr;
921 Value *LibCallSimplifier::optimizeMemCmp(CallInst *CI, IRBuilder<> &B) {
922 Function *Callee = CI->getCalledFunction();
923 FunctionType *FT = Callee->getFunctionType();
924 if (FT->getNumParams() != 3 || !FT->getParamType(0)->isPointerTy() ||
925 !FT->getParamType(1)->isPointerTy() ||
926 !FT->getReturnType()->isIntegerTy(32))
929 Value *LHS = CI->getArgOperand(0), *RHS = CI->getArgOperand(1);
931 if (LHS == RHS) // memcmp(s,s,x) -> 0
932 return Constant::getNullValue(CI->getType());
934 // Make sure we have a constant length.
935 ConstantInt *LenC = dyn_cast<ConstantInt>(CI->getArgOperand(2));
938 uint64_t Len = LenC->getZExtValue();
940 if (Len == 0) // memcmp(s1,s2,0) -> 0
941 return Constant::getNullValue(CI->getType());
943 // memcmp(S1,S2,1) -> *(unsigned char*)LHS - *(unsigned char*)RHS
945 Value *LHSV = B.CreateZExt(B.CreateLoad(CastToCStr(LHS, B), "lhsc"),
946 CI->getType(), "lhsv");
947 Value *RHSV = B.CreateZExt(B.CreateLoad(CastToCStr(RHS, B), "rhsc"),
948 CI->getType(), "rhsv");
949 return B.CreateSub(LHSV, RHSV, "chardiff");
952 // Constant folding: memcmp(x, y, l) -> cnst (all arguments are constant)
953 StringRef LHSStr, RHSStr;
954 if (getConstantStringInfo(LHS, LHSStr) &&
955 getConstantStringInfo(RHS, RHSStr)) {
956 // Make sure we're not reading out-of-bounds memory.
957 if (Len > LHSStr.size() || Len > RHSStr.size())
959 // Fold the memcmp and normalize the result. This way we get consistent
960 // results across multiple platforms.
962 int Cmp = memcmp(LHSStr.data(), RHSStr.data(), Len);
967 return ConstantInt::get(CI->getType(), Ret);
973 Value *LibCallSimplifier::optimizeMemCpy(CallInst *CI, IRBuilder<> &B) {
974 Function *Callee = CI->getCalledFunction();
975 // These optimizations require DataLayout.
979 FunctionType *FT = Callee->getFunctionType();
980 if (FT->getNumParams() != 3 || FT->getReturnType() != FT->getParamType(0) ||
981 !FT->getParamType(0)->isPointerTy() ||
982 !FT->getParamType(1)->isPointerTy() ||
983 FT->getParamType(2) != DL->getIntPtrType(CI->getContext()))
986 // memcpy(x, y, n) -> llvm.memcpy(x, y, n, 1)
987 B.CreateMemCpy(CI->getArgOperand(0), CI->getArgOperand(1),
988 CI->getArgOperand(2), 1);
989 return CI->getArgOperand(0);
992 Value *LibCallSimplifier::optimizeMemMove(CallInst *CI, IRBuilder<> &B) {
993 Function *Callee = CI->getCalledFunction();
994 // These optimizations require DataLayout.
998 FunctionType *FT = Callee->getFunctionType();
999 if (FT->getNumParams() != 3 || FT->getReturnType() != FT->getParamType(0) ||
1000 !FT->getParamType(0)->isPointerTy() ||
1001 !FT->getParamType(1)->isPointerTy() ||
1002 FT->getParamType(2) != DL->getIntPtrType(CI->getContext()))
1005 // memmove(x, y, n) -> llvm.memmove(x, y, n, 1)
1006 B.CreateMemMove(CI->getArgOperand(0), CI->getArgOperand(1),
1007 CI->getArgOperand(2), 1);
1008 return CI->getArgOperand(0);
1011 Value *LibCallSimplifier::optimizeMemSet(CallInst *CI, IRBuilder<> &B) {
1012 Function *Callee = CI->getCalledFunction();
1013 // These optimizations require DataLayout.
1017 FunctionType *FT = Callee->getFunctionType();
1018 if (FT->getNumParams() != 3 || FT->getReturnType() != FT->getParamType(0) ||
1019 !FT->getParamType(0)->isPointerTy() ||
1020 !FT->getParamType(1)->isIntegerTy() ||
1021 FT->getParamType(2) != DL->getIntPtrType(FT->getParamType(0)))
1024 // memset(p, v, n) -> llvm.memset(p, v, n, 1)
1025 Value *Val = B.CreateIntCast(CI->getArgOperand(1), B.getInt8Ty(), false);
1026 B.CreateMemSet(CI->getArgOperand(0), Val, CI->getArgOperand(2), 1);
1027 return CI->getArgOperand(0);
1030 //===----------------------------------------------------------------------===//
1031 // Math Library Optimizations
1032 //===----------------------------------------------------------------------===//
1034 //===----------------------------------------------------------------------===//
1035 // Double -> Float Shrinking Optimizations for Unary Functions like 'floor'
1037 Value *LibCallSimplifier::optimizeUnaryDoubleFP(CallInst *CI, IRBuilder<> &B,
1038 bool CheckRetType) {
1039 Function *Callee = CI->getCalledFunction();
1040 FunctionType *FT = Callee->getFunctionType();
1041 if (FT->getNumParams() != 1 || !FT->getReturnType()->isDoubleTy() ||
1042 !FT->getParamType(0)->isDoubleTy())
1046 // Check if all the uses for function like 'sin' are converted to float.
1047 for (User *U : CI->users()) {
1048 FPTruncInst *Cast = dyn_cast<FPTruncInst>(U);
1049 if (!Cast || !Cast->getType()->isFloatTy())
1054 // If this is something like 'floor((double)floatval)', convert to floorf.
1055 FPExtInst *Cast = dyn_cast<FPExtInst>(CI->getArgOperand(0));
1056 if (!Cast || !Cast->getOperand(0)->getType()->isFloatTy())
1059 // floor((double)floatval) -> (double)floorf(floatval)
1060 Value *V = Cast->getOperand(0);
1061 if (Callee->isIntrinsic()) {
1062 Module *M = CI->getParent()->getParent()->getParent();
1063 Intrinsic::ID IID = (Intrinsic::ID) Callee->getIntrinsicID();
1064 Function *F = Intrinsic::getDeclaration(M, IID, B.getFloatTy());
1065 V = B.CreateCall(F, V);
1067 // The call is a library call rather than an intrinsic.
1068 V = EmitUnaryFloatFnCall(V, Callee->getName(), B, Callee->getAttributes());
1071 return B.CreateFPExt(V, B.getDoubleTy());
1074 // Double -> Float Shrinking Optimizations for Binary Functions like 'fmin/fmax'
1075 Value *LibCallSimplifier::optimizeBinaryDoubleFP(CallInst *CI, IRBuilder<> &B) {
1076 Function *Callee = CI->getCalledFunction();
1077 FunctionType *FT = Callee->getFunctionType();
1078 // Just make sure this has 2 arguments of the same FP type, which match the
1080 if (FT->getNumParams() != 2 || FT->getReturnType() != FT->getParamType(0) ||
1081 FT->getParamType(0) != FT->getParamType(1) ||
1082 !FT->getParamType(0)->isFloatingPointTy())
1085 // If this is something like 'fmin((double)floatval1, (double)floatval2)',
1086 // we convert it to fminf.
1087 FPExtInst *Cast1 = dyn_cast<FPExtInst>(CI->getArgOperand(0));
1088 FPExtInst *Cast2 = dyn_cast<FPExtInst>(CI->getArgOperand(1));
1089 if (!Cast1 || !Cast1->getOperand(0)->getType()->isFloatTy() || !Cast2 ||
1090 !Cast2->getOperand(0)->getType()->isFloatTy())
1093 // fmin((double)floatval1, (double)floatval2)
1094 // -> (double)fmin(floatval1, floatval2)
1096 Value *V1 = Cast1->getOperand(0);
1097 Value *V2 = Cast2->getOperand(0);
1098 // TODO: Handle intrinsics in the same way as in optimizeUnaryDoubleFP().
1099 V = EmitBinaryFloatFnCall(V1, V2, Callee->getName(), B,
1100 Callee->getAttributes());
1101 return B.CreateFPExt(V, B.getDoubleTy());
1104 Value *LibCallSimplifier::optimizeCos(CallInst *CI, IRBuilder<> &B) {
1105 Function *Callee = CI->getCalledFunction();
1106 Value *Ret = nullptr;
1107 if (UnsafeFPShrink && Callee->getName() == "cos" && TLI->has(LibFunc::cosf)) {
1108 Ret = optimizeUnaryDoubleFP(CI, B, true);
1111 FunctionType *FT = Callee->getFunctionType();
1112 // Just make sure this has 1 argument of FP type, which matches the
1114 if (FT->getNumParams() != 1 || FT->getReturnType() != FT->getParamType(0) ||
1115 !FT->getParamType(0)->isFloatingPointTy())
1118 // cos(-x) -> cos(x)
1119 Value *Op1 = CI->getArgOperand(0);
1120 if (BinaryOperator::isFNeg(Op1)) {
1121 BinaryOperator *BinExpr = cast<BinaryOperator>(Op1);
1122 return B.CreateCall(Callee, BinExpr->getOperand(1), "cos");
1127 Value *LibCallSimplifier::optimizePow(CallInst *CI, IRBuilder<> &B) {
1128 Function *Callee = CI->getCalledFunction();
1130 Value *Ret = nullptr;
1131 if (UnsafeFPShrink && Callee->getName() == "pow" && TLI->has(LibFunc::powf)) {
1132 Ret = optimizeUnaryDoubleFP(CI, B, true);
1135 FunctionType *FT = Callee->getFunctionType();
1136 // Just make sure this has 2 arguments of the same FP type, which match the
1138 if (FT->getNumParams() != 2 || FT->getReturnType() != FT->getParamType(0) ||
1139 FT->getParamType(0) != FT->getParamType(1) ||
1140 !FT->getParamType(0)->isFloatingPointTy())
1143 Value *Op1 = CI->getArgOperand(0), *Op2 = CI->getArgOperand(1);
1144 if (ConstantFP *Op1C = dyn_cast<ConstantFP>(Op1)) {
1145 // pow(1.0, x) -> 1.0
1146 if (Op1C->isExactlyValue(1.0))
1148 // pow(2.0, x) -> exp2(x)
1149 if (Op1C->isExactlyValue(2.0) &&
1150 hasUnaryFloatFn(TLI, Op1->getType(), LibFunc::exp2, LibFunc::exp2f,
1152 return EmitUnaryFloatFnCall(Op2, "exp2", B, Callee->getAttributes());
1153 // pow(10.0, x) -> exp10(x)
1154 if (Op1C->isExactlyValue(10.0) &&
1155 hasUnaryFloatFn(TLI, Op1->getType(), LibFunc::exp10, LibFunc::exp10f,
1157 return EmitUnaryFloatFnCall(Op2, TLI->getName(LibFunc::exp10), B,
1158 Callee->getAttributes());
1161 ConstantFP *Op2C = dyn_cast<ConstantFP>(Op2);
1165 if (Op2C->getValueAPF().isZero()) // pow(x, 0.0) -> 1.0
1166 return ConstantFP::get(CI->getType(), 1.0);
1168 if (Op2C->isExactlyValue(0.5) &&
1169 hasUnaryFloatFn(TLI, Op2->getType(), LibFunc::sqrt, LibFunc::sqrtf,
1171 hasUnaryFloatFn(TLI, Op2->getType(), LibFunc::fabs, LibFunc::fabsf,
1173 // Expand pow(x, 0.5) to (x == -infinity ? +infinity : fabs(sqrt(x))).
1174 // This is faster than calling pow, and still handles negative zero
1175 // and negative infinity correctly.
1176 // TODO: In fast-math mode, this could be just sqrt(x).
1177 // TODO: In finite-only mode, this could be just fabs(sqrt(x)).
1178 Value *Inf = ConstantFP::getInfinity(CI->getType());
1179 Value *NegInf = ConstantFP::getInfinity(CI->getType(), true);
1180 Value *Sqrt = EmitUnaryFloatFnCall(Op1, "sqrt", B, Callee->getAttributes());
1182 EmitUnaryFloatFnCall(Sqrt, "fabs", B, Callee->getAttributes());
1183 Value *FCmp = B.CreateFCmpOEQ(Op1, NegInf);
1184 Value *Sel = B.CreateSelect(FCmp, Inf, FAbs);
1188 if (Op2C->isExactlyValue(1.0)) // pow(x, 1.0) -> x
1190 if (Op2C->isExactlyValue(2.0)) // pow(x, 2.0) -> x*x
1191 return B.CreateFMul(Op1, Op1, "pow2");
1192 if (Op2C->isExactlyValue(-1.0)) // pow(x, -1.0) -> 1.0/x
1193 return B.CreateFDiv(ConstantFP::get(CI->getType(), 1.0), Op1, "powrecip");
1197 Value *LibCallSimplifier::optimizeExp2(CallInst *CI, IRBuilder<> &B) {
1198 Function *Callee = CI->getCalledFunction();
1199 Function *Caller = CI->getParent()->getParent();
1201 Value *Ret = nullptr;
1202 if (UnsafeFPShrink && Callee->getName() == "exp2" &&
1203 TLI->has(LibFunc::exp2f)) {
1204 Ret = optimizeUnaryDoubleFP(CI, B, true);
1207 FunctionType *FT = Callee->getFunctionType();
1208 // Just make sure this has 1 argument of FP type, which matches the
1210 if (FT->getNumParams() != 1 || FT->getReturnType() != FT->getParamType(0) ||
1211 !FT->getParamType(0)->isFloatingPointTy())
1214 Value *Op = CI->getArgOperand(0);
1215 // Turn exp2(sitofp(x)) -> ldexp(1.0, sext(x)) if sizeof(x) <= 32
1216 // Turn exp2(uitofp(x)) -> ldexp(1.0, zext(x)) if sizeof(x) < 32
1217 LibFunc::Func LdExp = LibFunc::ldexpl;
1218 if (Op->getType()->isFloatTy())
1219 LdExp = LibFunc::ldexpf;
1220 else if (Op->getType()->isDoubleTy())
1221 LdExp = LibFunc::ldexp;
1223 if (TLI->has(LdExp)) {
1224 Value *LdExpArg = nullptr;
1225 if (SIToFPInst *OpC = dyn_cast<SIToFPInst>(Op)) {
1226 if (OpC->getOperand(0)->getType()->getPrimitiveSizeInBits() <= 32)
1227 LdExpArg = B.CreateSExt(OpC->getOperand(0), B.getInt32Ty());
1228 } else if (UIToFPInst *OpC = dyn_cast<UIToFPInst>(Op)) {
1229 if (OpC->getOperand(0)->getType()->getPrimitiveSizeInBits() < 32)
1230 LdExpArg = B.CreateZExt(OpC->getOperand(0), B.getInt32Ty());
1234 Constant *One = ConstantFP::get(CI->getContext(), APFloat(1.0f));
1235 if (!Op->getType()->isFloatTy())
1236 One = ConstantExpr::getFPExtend(One, Op->getType());
1238 Module *M = Caller->getParent();
1240 M->getOrInsertFunction(TLI->getName(LdExp), Op->getType(),
1241 Op->getType(), B.getInt32Ty(), nullptr);
1242 CallInst *CI = B.CreateCall2(Callee, One, LdExpArg);
1243 if (const Function *F = dyn_cast<Function>(Callee->stripPointerCasts()))
1244 CI->setCallingConv(F->getCallingConv());
1252 Value *LibCallSimplifier::optimizeFabs(CallInst *CI, IRBuilder<> &B) {
1253 Function *Callee = CI->getCalledFunction();
1255 Value *Ret = nullptr;
1256 if (Callee->getName() == "fabs" && TLI->has(LibFunc::fabsf)) {
1257 Ret = optimizeUnaryDoubleFP(CI, B, false);
1260 FunctionType *FT = Callee->getFunctionType();
1261 // Make sure this has 1 argument of FP type which matches the result type.
1262 if (FT->getNumParams() != 1 || FT->getReturnType() != FT->getParamType(0) ||
1263 !FT->getParamType(0)->isFloatingPointTy())
1266 Value *Op = CI->getArgOperand(0);
1267 if (Instruction *I = dyn_cast<Instruction>(Op)) {
1268 // Fold fabs(x * x) -> x * x; any squared FP value must already be positive.
1269 if (I->getOpcode() == Instruction::FMul)
1270 if (I->getOperand(0) == I->getOperand(1))
1276 Value *LibCallSimplifier::optimizeSqrt(CallInst *CI, IRBuilder<> &B) {
1277 Function *Callee = CI->getCalledFunction();
1279 Value *Ret = nullptr;
1280 if (TLI->has(LibFunc::sqrtf) && (Callee->getName() == "sqrt" ||
1281 Callee->getIntrinsicID() == Intrinsic::sqrt))
1282 Ret = optimizeUnaryDoubleFP(CI, B, true);
1284 // FIXME: For finer-grain optimization, we need intrinsics to have the same
1285 // fast-math flag decorations that are applied to FP instructions. For now,
1286 // we have to rely on the function-level unsafe-fp-math attribute to do this
1287 // optimization because there's no other way to express that the sqrt can be
1289 Function *F = CI->getParent()->getParent();
1290 if (F->hasFnAttribute("unsafe-fp-math")) {
1291 // Check for unsafe-fp-math = true.
1292 Attribute Attr = F->getFnAttribute("unsafe-fp-math");
1293 if (Attr.getValueAsString() != "true")
1296 Value *Op = CI->getArgOperand(0);
1297 if (Instruction *I = dyn_cast<Instruction>(Op)) {
1298 if (I->getOpcode() == Instruction::FMul && I->hasUnsafeAlgebra()) {
1299 // We're looking for a repeated factor in a multiplication tree,
1300 // so we can do this fold: sqrt(x * x) -> fabs(x);
1301 // or this fold: sqrt(x * x * y) -> fabs(x) * sqrt(y).
1302 Value *Op0 = I->getOperand(0);
1303 Value *Op1 = I->getOperand(1);
1304 Value *RepeatOp = nullptr;
1305 Value *OtherOp = nullptr;
1307 // Simple match: the operands of the multiply are identical.
1310 // Look for a more complicated pattern: one of the operands is itself
1311 // a multiply, so search for a common factor in that multiply.
1312 // Note: We don't bother looking any deeper than this first level or for
1313 // variations of this pattern because instcombine's visitFMUL and/or the
1314 // reassociation pass should give us this form.
1315 Value *OtherMul0, *OtherMul1;
1316 if (match(Op0, m_FMul(m_Value(OtherMul0), m_Value(OtherMul1)))) {
1317 // Pattern: sqrt((x * y) * z)
1318 if (OtherMul0 == OtherMul1) {
1319 // Matched: sqrt((x * x) * z)
1320 RepeatOp = OtherMul0;
1326 // Fast math flags for any created instructions should match the sqrt
1328 // FIXME: We're not checking the sqrt because it doesn't have
1329 // fast-math-flags (see earlier comment).
1330 IRBuilder<true, ConstantFolder,
1331 IRBuilderDefaultInserter<true> >::FastMathFlagGuard Guard(B);
1332 B.SetFastMathFlags(I->getFastMathFlags());
1333 // If we found a repeated factor, hoist it out of the square root and
1334 // replace it with the fabs of that factor.
1335 Module *M = Callee->getParent();
1336 Type *ArgType = Op->getType();
1337 Value *Fabs = Intrinsic::getDeclaration(M, Intrinsic::fabs, ArgType);
1338 Value *FabsCall = B.CreateCall(Fabs, RepeatOp, "fabs");
1340 // If we found a non-repeated factor, we still need to get its square
1341 // root. We then multiply that by the value that was simplified out
1342 // of the square root calculation.
1343 Value *Sqrt = Intrinsic::getDeclaration(M, Intrinsic::sqrt, ArgType);
1344 Value *SqrtCall = B.CreateCall(Sqrt, OtherOp, "sqrt");
1345 return B.CreateFMul(FabsCall, SqrtCall);
1354 static bool isTrigLibCall(CallInst *CI);
1355 static void insertSinCosCall(IRBuilder<> &B, Function *OrigCallee, Value *Arg,
1356 bool UseFloat, Value *&Sin, Value *&Cos,
1359 Value *LibCallSimplifier::optimizeSinCosPi(CallInst *CI, IRBuilder<> &B) {
1361 // Make sure the prototype is as expected, otherwise the rest of the
1362 // function is probably invalid and likely to abort.
1363 if (!isTrigLibCall(CI))
1366 Value *Arg = CI->getArgOperand(0);
1367 SmallVector<CallInst *, 1> SinCalls;
1368 SmallVector<CallInst *, 1> CosCalls;
1369 SmallVector<CallInst *, 1> SinCosCalls;
1371 bool IsFloat = Arg->getType()->isFloatTy();
1373 // Look for all compatible sinpi, cospi and sincospi calls with the same
1374 // argument. If there are enough (in some sense) we can make the
1376 for (User *U : Arg->users())
1377 classifyArgUse(U, CI->getParent(), IsFloat, SinCalls, CosCalls,
1380 // It's only worthwhile if both sinpi and cospi are actually used.
1381 if (SinCosCalls.empty() && (SinCalls.empty() || CosCalls.empty()))
1384 Value *Sin, *Cos, *SinCos;
1385 insertSinCosCall(B, CI->getCalledFunction(), Arg, IsFloat, Sin, Cos, SinCos);
1387 replaceTrigInsts(SinCalls, Sin);
1388 replaceTrigInsts(CosCalls, Cos);
1389 replaceTrigInsts(SinCosCalls, SinCos);
1394 static bool isTrigLibCall(CallInst *CI) {
1395 Function *Callee = CI->getCalledFunction();
1396 FunctionType *FT = Callee->getFunctionType();
1398 // We can only hope to do anything useful if we can ignore things like errno
1399 // and floating-point exceptions.
1400 bool AttributesSafe =
1401 CI->hasFnAttr(Attribute::NoUnwind) && CI->hasFnAttr(Attribute::ReadNone);
1403 // Other than that we need float(float) or double(double)
1404 return AttributesSafe && FT->getNumParams() == 1 &&
1405 FT->getReturnType() == FT->getParamType(0) &&
1406 (FT->getParamType(0)->isFloatTy() ||
1407 FT->getParamType(0)->isDoubleTy());
1411 LibCallSimplifier::classifyArgUse(Value *Val, BasicBlock *BB, bool IsFloat,
1412 SmallVectorImpl<CallInst *> &SinCalls,
1413 SmallVectorImpl<CallInst *> &CosCalls,
1414 SmallVectorImpl<CallInst *> &SinCosCalls) {
1415 CallInst *CI = dyn_cast<CallInst>(Val);
1420 Function *Callee = CI->getCalledFunction();
1421 StringRef FuncName = Callee->getName();
1423 if (!TLI->getLibFunc(FuncName, Func) || !TLI->has(Func) || !isTrigLibCall(CI))
1427 if (Func == LibFunc::sinpif)
1428 SinCalls.push_back(CI);
1429 else if (Func == LibFunc::cospif)
1430 CosCalls.push_back(CI);
1431 else if (Func == LibFunc::sincospif_stret)
1432 SinCosCalls.push_back(CI);
1434 if (Func == LibFunc::sinpi)
1435 SinCalls.push_back(CI);
1436 else if (Func == LibFunc::cospi)
1437 CosCalls.push_back(CI);
1438 else if (Func == LibFunc::sincospi_stret)
1439 SinCosCalls.push_back(CI);
1443 void LibCallSimplifier::replaceTrigInsts(SmallVectorImpl<CallInst *> &Calls,
1445 for (SmallVectorImpl<CallInst *>::iterator I = Calls.begin(), E = Calls.end();
1447 replaceAllUsesWith(*I, Res);
1451 void insertSinCosCall(IRBuilder<> &B, Function *OrigCallee, Value *Arg,
1452 bool UseFloat, Value *&Sin, Value *&Cos, Value *&SinCos) {
1453 Type *ArgTy = Arg->getType();
1457 Triple T(OrigCallee->getParent()->getTargetTriple());
1459 Name = "__sincospif_stret";
1461 assert(T.getArch() != Triple::x86 && "x86 messy and unsupported for now");
1462 // x86_64 can't use {float, float} since that would be returned in both
1463 // xmm0 and xmm1, which isn't what a real struct would do.
1464 ResTy = T.getArch() == Triple::x86_64
1465 ? static_cast<Type *>(VectorType::get(ArgTy, 2))
1466 : static_cast<Type *>(StructType::get(ArgTy, ArgTy, nullptr));
1468 Name = "__sincospi_stret";
1469 ResTy = StructType::get(ArgTy, ArgTy, nullptr);
1472 Module *M = OrigCallee->getParent();
1473 Value *Callee = M->getOrInsertFunction(Name, OrigCallee->getAttributes(),
1474 ResTy, ArgTy, nullptr);
1476 if (Instruction *ArgInst = dyn_cast<Instruction>(Arg)) {
1477 // If the argument is an instruction, it must dominate all uses so put our
1478 // sincos call there.
1479 BasicBlock::iterator Loc = ArgInst;
1480 B.SetInsertPoint(ArgInst->getParent(), ++Loc);
1482 // Otherwise (e.g. for a constant) the beginning of the function is as
1483 // good a place as any.
1484 BasicBlock &EntryBB = B.GetInsertBlock()->getParent()->getEntryBlock();
1485 B.SetInsertPoint(&EntryBB, EntryBB.begin());
1488 SinCos = B.CreateCall(Callee, Arg, "sincospi");
1490 if (SinCos->getType()->isStructTy()) {
1491 Sin = B.CreateExtractValue(SinCos, 0, "sinpi");
1492 Cos = B.CreateExtractValue(SinCos, 1, "cospi");
1494 Sin = B.CreateExtractElement(SinCos, ConstantInt::get(B.getInt32Ty(), 0),
1496 Cos = B.CreateExtractElement(SinCos, ConstantInt::get(B.getInt32Ty(), 1),
1501 //===----------------------------------------------------------------------===//
1502 // Integer Library Call Optimizations
1503 //===----------------------------------------------------------------------===//
1505 Value *LibCallSimplifier::optimizeFFS(CallInst *CI, IRBuilder<> &B) {
1506 Function *Callee = CI->getCalledFunction();
1507 FunctionType *FT = Callee->getFunctionType();
1508 // Just make sure this has 2 arguments of the same FP type, which match the
1510 if (FT->getNumParams() != 1 || !FT->getReturnType()->isIntegerTy(32) ||
1511 !FT->getParamType(0)->isIntegerTy())
1514 Value *Op = CI->getArgOperand(0);
1517 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
1518 if (CI->isZero()) // ffs(0) -> 0.
1519 return B.getInt32(0);
1520 // ffs(c) -> cttz(c)+1
1521 return B.getInt32(CI->getValue().countTrailingZeros() + 1);
1524 // ffs(x) -> x != 0 ? (i32)llvm.cttz(x)+1 : 0
1525 Type *ArgType = Op->getType();
1527 Intrinsic::getDeclaration(Callee->getParent(), Intrinsic::cttz, ArgType);
1528 Value *V = B.CreateCall2(F, Op, B.getFalse(), "cttz");
1529 V = B.CreateAdd(V, ConstantInt::get(V->getType(), 1));
1530 V = B.CreateIntCast(V, B.getInt32Ty(), false);
1532 Value *Cond = B.CreateICmpNE(Op, Constant::getNullValue(ArgType));
1533 return B.CreateSelect(Cond, V, B.getInt32(0));
1536 Value *LibCallSimplifier::optimizeAbs(CallInst *CI, IRBuilder<> &B) {
1537 Function *Callee = CI->getCalledFunction();
1538 FunctionType *FT = Callee->getFunctionType();
1539 // We require integer(integer) where the types agree.
1540 if (FT->getNumParams() != 1 || !FT->getReturnType()->isIntegerTy() ||
1541 FT->getParamType(0) != FT->getReturnType())
1544 // abs(x) -> x >s -1 ? x : -x
1545 Value *Op = CI->getArgOperand(0);
1547 B.CreateICmpSGT(Op, Constant::getAllOnesValue(Op->getType()), "ispos");
1548 Value *Neg = B.CreateNeg(Op, "neg");
1549 return B.CreateSelect(Pos, Op, Neg);
1552 Value *LibCallSimplifier::optimizeIsDigit(CallInst *CI, IRBuilder<> &B) {
1553 Function *Callee = CI->getCalledFunction();
1554 FunctionType *FT = Callee->getFunctionType();
1555 // We require integer(i32)
1556 if (FT->getNumParams() != 1 || !FT->getReturnType()->isIntegerTy() ||
1557 !FT->getParamType(0)->isIntegerTy(32))
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 Function *Callee = CI->getCalledFunction();
1569 FunctionType *FT = Callee->getFunctionType();
1570 // We require integer(i32)
1571 if (FT->getNumParams() != 1 || !FT->getReturnType()->isIntegerTy() ||
1572 !FT->getParamType(0)->isIntegerTy(32))
1575 // isascii(c) -> c <u 128
1576 Value *Op = CI->getArgOperand(0);
1577 Op = B.CreateICmpULT(Op, B.getInt32(128), "isascii");
1578 return B.CreateZExt(Op, CI->getType());
1581 Value *LibCallSimplifier::optimizeToAscii(CallInst *CI, IRBuilder<> &B) {
1582 Function *Callee = CI->getCalledFunction();
1583 FunctionType *FT = Callee->getFunctionType();
1584 // We require i32(i32)
1585 if (FT->getNumParams() != 1 || FT->getReturnType() != FT->getParamType(0) ||
1586 !FT->getParamType(0)->isIntegerTy(32))
1589 // toascii(c) -> c & 0x7f
1590 return B.CreateAnd(CI->getArgOperand(0),
1591 ConstantInt::get(CI->getType(), 0x7F));
1594 //===----------------------------------------------------------------------===//
1595 // Formatting and IO Library Call Optimizations
1596 //===----------------------------------------------------------------------===//
1598 static bool isReportingError(Function *Callee, CallInst *CI, int StreamArg);
1600 Value *LibCallSimplifier::optimizeErrorReporting(CallInst *CI, IRBuilder<> &B,
1602 // Error reporting calls should be cold, mark them as such.
1603 // This applies even to non-builtin calls: it is only a hint and applies to
1604 // functions that the frontend might not understand as builtins.
1606 // This heuristic was suggested in:
1607 // Improving Static Branch Prediction in a Compiler
1608 // Brian L. Deitrich, Ben-Chung Cheng, Wen-mei W. Hwu
1609 // Proceedings of PACT'98, Oct. 1998, IEEE
1610 Function *Callee = CI->getCalledFunction();
1612 if (!CI->hasFnAttr(Attribute::Cold) &&
1613 isReportingError(Callee, CI, StreamArg)) {
1614 CI->addAttribute(AttributeSet::FunctionIndex, Attribute::Cold);
1620 static bool isReportingError(Function *Callee, CallInst *CI, int StreamArg) {
1621 if (!ColdErrorCalls)
1624 if (!Callee || !Callee->isDeclaration())
1630 // These functions might be considered cold, but only if their stream
1631 // argument is stderr.
1633 if (StreamArg >= (int)CI->getNumArgOperands())
1635 LoadInst *LI = dyn_cast<LoadInst>(CI->getArgOperand(StreamArg));
1638 GlobalVariable *GV = dyn_cast<GlobalVariable>(LI->getPointerOperand());
1639 if (!GV || !GV->isDeclaration())
1641 return GV->getName() == "stderr";
1644 Value *LibCallSimplifier::optimizePrintFString(CallInst *CI, IRBuilder<> &B) {
1645 // Check for a fixed format string.
1646 StringRef FormatStr;
1647 if (!getConstantStringInfo(CI->getArgOperand(0), FormatStr))
1650 // Empty format string -> noop.
1651 if (FormatStr.empty()) // Tolerate printf's declared void.
1652 return CI->use_empty() ? (Value *)CI : ConstantInt::get(CI->getType(), 0);
1654 // Do not do any of the following transformations if the printf return value
1655 // is used, in general the printf return value is not compatible with either
1656 // putchar() or puts().
1657 if (!CI->use_empty())
1660 // printf("x") -> putchar('x'), even for '%'.
1661 if (FormatStr.size() == 1) {
1662 Value *Res = EmitPutChar(B.getInt32(FormatStr[0]), B, DL, TLI);
1663 if (CI->use_empty() || !Res)
1665 return B.CreateIntCast(Res, CI->getType(), true);
1668 // printf("foo\n") --> puts("foo")
1669 if (FormatStr[FormatStr.size() - 1] == '\n' &&
1670 FormatStr.find('%') == StringRef::npos) { // No format characters.
1671 // Create a string literal with no \n on it. We expect the constant merge
1672 // pass to be run after this pass, to merge duplicate strings.
1673 FormatStr = FormatStr.drop_back();
1674 Value *GV = B.CreateGlobalString(FormatStr, "str");
1675 Value *NewCI = EmitPutS(GV, B, DL, TLI);
1676 return (CI->use_empty() || !NewCI)
1678 : ConstantInt::get(CI->getType(), FormatStr.size() + 1);
1681 // Optimize specific format strings.
1682 // printf("%c", chr) --> putchar(chr)
1683 if (FormatStr == "%c" && CI->getNumArgOperands() > 1 &&
1684 CI->getArgOperand(1)->getType()->isIntegerTy()) {
1685 Value *Res = EmitPutChar(CI->getArgOperand(1), B, DL, TLI);
1687 if (CI->use_empty() || !Res)
1689 return B.CreateIntCast(Res, CI->getType(), true);
1692 // printf("%s\n", str) --> puts(str)
1693 if (FormatStr == "%s\n" && CI->getNumArgOperands() > 1 &&
1694 CI->getArgOperand(1)->getType()->isPointerTy()) {
1695 return EmitPutS(CI->getArgOperand(1), B, DL, TLI);
1700 Value *LibCallSimplifier::optimizePrintF(CallInst *CI, IRBuilder<> &B) {
1702 Function *Callee = CI->getCalledFunction();
1703 // Require one fixed pointer argument and an integer/void result.
1704 FunctionType *FT = Callee->getFunctionType();
1705 if (FT->getNumParams() < 1 || !FT->getParamType(0)->isPointerTy() ||
1706 !(FT->getReturnType()->isIntegerTy() || FT->getReturnType()->isVoidTy()))
1709 if (Value *V = optimizePrintFString(CI, B)) {
1713 // printf(format, ...) -> iprintf(format, ...) if no floating point
1715 if (TLI->has(LibFunc::iprintf) && !callHasFloatingPointArgument(CI)) {
1716 Module *M = B.GetInsertBlock()->getParent()->getParent();
1717 Constant *IPrintFFn =
1718 M->getOrInsertFunction("iprintf", FT, Callee->getAttributes());
1719 CallInst *New = cast<CallInst>(CI->clone());
1720 New->setCalledFunction(IPrintFFn);
1727 Value *LibCallSimplifier::optimizeSPrintFString(CallInst *CI, IRBuilder<> &B) {
1728 // Check for a fixed format string.
1729 StringRef FormatStr;
1730 if (!getConstantStringInfo(CI->getArgOperand(1), FormatStr))
1733 // If we just have a format string (nothing else crazy) transform it.
1734 if (CI->getNumArgOperands() == 2) {
1735 // Make sure there's no % in the constant array. We could try to handle
1736 // %% -> % in the future if we cared.
1737 for (unsigned i = 0, e = FormatStr.size(); i != e; ++i)
1738 if (FormatStr[i] == '%')
1739 return nullptr; // we found a format specifier, bail out.
1741 // These optimizations require DataLayout.
1745 // sprintf(str, fmt) -> llvm.memcpy(str, fmt, strlen(fmt)+1, 1)
1747 CI->getArgOperand(0), CI->getArgOperand(1),
1748 ConstantInt::get(DL->getIntPtrType(CI->getContext()),
1749 FormatStr.size() + 1),
1750 1); // Copy the null byte.
1751 return ConstantInt::get(CI->getType(), FormatStr.size());
1754 // The remaining optimizations require the format string to be "%s" or "%c"
1755 // and have an extra operand.
1756 if (FormatStr.size() != 2 || FormatStr[0] != '%' ||
1757 CI->getNumArgOperands() < 3)
1760 // Decode the second character of the format string.
1761 if (FormatStr[1] == 'c') {
1762 // sprintf(dst, "%c", chr) --> *(i8*)dst = chr; *((i8*)dst+1) = 0
1763 if (!CI->getArgOperand(2)->getType()->isIntegerTy())
1765 Value *V = B.CreateTrunc(CI->getArgOperand(2), B.getInt8Ty(), "char");
1766 Value *Ptr = CastToCStr(CI->getArgOperand(0), B);
1767 B.CreateStore(V, Ptr);
1768 Ptr = B.CreateGEP(Ptr, B.getInt32(1), "nul");
1769 B.CreateStore(B.getInt8(0), Ptr);
1771 return ConstantInt::get(CI->getType(), 1);
1774 if (FormatStr[1] == 's') {
1775 // These optimizations require DataLayout.
1779 // sprintf(dest, "%s", str) -> llvm.memcpy(dest, str, strlen(str)+1, 1)
1780 if (!CI->getArgOperand(2)->getType()->isPointerTy())
1783 Value *Len = EmitStrLen(CI->getArgOperand(2), B, DL, TLI);
1787 B.CreateAdd(Len, ConstantInt::get(Len->getType(), 1), "leninc");
1788 B.CreateMemCpy(CI->getArgOperand(0), CI->getArgOperand(2), IncLen, 1);
1790 // The sprintf result is the unincremented number of bytes in the string.
1791 return B.CreateIntCast(Len, CI->getType(), false);
1796 Value *LibCallSimplifier::optimizeSPrintF(CallInst *CI, IRBuilder<> &B) {
1797 Function *Callee = CI->getCalledFunction();
1798 // Require two fixed pointer arguments and an integer result.
1799 FunctionType *FT = Callee->getFunctionType();
1800 if (FT->getNumParams() != 2 || !FT->getParamType(0)->isPointerTy() ||
1801 !FT->getParamType(1)->isPointerTy() ||
1802 !FT->getReturnType()->isIntegerTy())
1805 if (Value *V = optimizeSPrintFString(CI, B)) {
1809 // sprintf(str, format, ...) -> siprintf(str, format, ...) if no floating
1811 if (TLI->has(LibFunc::siprintf) && !callHasFloatingPointArgument(CI)) {
1812 Module *M = B.GetInsertBlock()->getParent()->getParent();
1813 Constant *SIPrintFFn =
1814 M->getOrInsertFunction("siprintf", FT, Callee->getAttributes());
1815 CallInst *New = cast<CallInst>(CI->clone());
1816 New->setCalledFunction(SIPrintFFn);
1823 Value *LibCallSimplifier::optimizeFPrintFString(CallInst *CI, IRBuilder<> &B) {
1824 optimizeErrorReporting(CI, B, 0);
1826 // All the optimizations depend on the format string.
1827 StringRef FormatStr;
1828 if (!getConstantStringInfo(CI->getArgOperand(1), FormatStr))
1831 // Do not do any of the following transformations if the fprintf return
1832 // value is used, in general the fprintf return value is not compatible
1833 // with fwrite(), fputc() or fputs().
1834 if (!CI->use_empty())
1837 // fprintf(F, "foo") --> fwrite("foo", 3, 1, F)
1838 if (CI->getNumArgOperands() == 2) {
1839 for (unsigned i = 0, e = FormatStr.size(); i != e; ++i)
1840 if (FormatStr[i] == '%') // Could handle %% -> % if we cared.
1841 return nullptr; // We found a format specifier.
1843 // These optimizations require DataLayout.
1848 CI->getArgOperand(1),
1849 ConstantInt::get(DL->getIntPtrType(CI->getContext()), FormatStr.size()),
1850 CI->getArgOperand(0), B, DL, TLI);
1853 // The remaining optimizations require the format string to be "%s" or "%c"
1854 // and have an extra operand.
1855 if (FormatStr.size() != 2 || FormatStr[0] != '%' ||
1856 CI->getNumArgOperands() < 3)
1859 // Decode the second character of the format string.
1860 if (FormatStr[1] == 'c') {
1861 // fprintf(F, "%c", chr) --> fputc(chr, F)
1862 if (!CI->getArgOperand(2)->getType()->isIntegerTy())
1864 return EmitFPutC(CI->getArgOperand(2), CI->getArgOperand(0), B, DL, TLI);
1867 if (FormatStr[1] == 's') {
1868 // fprintf(F, "%s", str) --> fputs(str, F)
1869 if (!CI->getArgOperand(2)->getType()->isPointerTy())
1871 return EmitFPutS(CI->getArgOperand(2), CI->getArgOperand(0), B, DL, TLI);
1876 Value *LibCallSimplifier::optimizeFPrintF(CallInst *CI, IRBuilder<> &B) {
1877 Function *Callee = CI->getCalledFunction();
1878 // Require two fixed paramters as pointers and integer result.
1879 FunctionType *FT = Callee->getFunctionType();
1880 if (FT->getNumParams() != 2 || !FT->getParamType(0)->isPointerTy() ||
1881 !FT->getParamType(1)->isPointerTy() ||
1882 !FT->getReturnType()->isIntegerTy())
1885 if (Value *V = optimizeFPrintFString(CI, B)) {
1889 // fprintf(stream, format, ...) -> fiprintf(stream, format, ...) if no
1890 // floating point arguments.
1891 if (TLI->has(LibFunc::fiprintf) && !callHasFloatingPointArgument(CI)) {
1892 Module *M = B.GetInsertBlock()->getParent()->getParent();
1893 Constant *FIPrintFFn =
1894 M->getOrInsertFunction("fiprintf", FT, Callee->getAttributes());
1895 CallInst *New = cast<CallInst>(CI->clone());
1896 New->setCalledFunction(FIPrintFFn);
1903 Value *LibCallSimplifier::optimizeFWrite(CallInst *CI, IRBuilder<> &B) {
1904 optimizeErrorReporting(CI, B, 3);
1906 Function *Callee = CI->getCalledFunction();
1907 // Require a pointer, an integer, an integer, a pointer, returning integer.
1908 FunctionType *FT = Callee->getFunctionType();
1909 if (FT->getNumParams() != 4 || !FT->getParamType(0)->isPointerTy() ||
1910 !FT->getParamType(1)->isIntegerTy() ||
1911 !FT->getParamType(2)->isIntegerTy() ||
1912 !FT->getParamType(3)->isPointerTy() ||
1913 !FT->getReturnType()->isIntegerTy())
1916 // Get the element size and count.
1917 ConstantInt *SizeC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
1918 ConstantInt *CountC = dyn_cast<ConstantInt>(CI->getArgOperand(2));
1919 if (!SizeC || !CountC)
1921 uint64_t Bytes = SizeC->getZExtValue() * CountC->getZExtValue();
1923 // If this is writing zero records, remove the call (it's a noop).
1925 return ConstantInt::get(CI->getType(), 0);
1927 // If this is writing one byte, turn it into fputc.
1928 // This optimisation is only valid, if the return value is unused.
1929 if (Bytes == 1 && CI->use_empty()) { // fwrite(S,1,1,F) -> fputc(S[0],F)
1930 Value *Char = B.CreateLoad(CastToCStr(CI->getArgOperand(0), B), "char");
1931 Value *NewCI = EmitFPutC(Char, CI->getArgOperand(3), B, DL, TLI);
1932 return NewCI ? ConstantInt::get(CI->getType(), 1) : nullptr;
1938 Value *LibCallSimplifier::optimizeFPuts(CallInst *CI, IRBuilder<> &B) {
1939 optimizeErrorReporting(CI, B, 1);
1941 Function *Callee = CI->getCalledFunction();
1943 // These optimizations require DataLayout.
1947 // Require two pointers. Also, we can't optimize if return value is used.
1948 FunctionType *FT = Callee->getFunctionType();
1949 if (FT->getNumParams() != 2 || !FT->getParamType(0)->isPointerTy() ||
1950 !FT->getParamType(1)->isPointerTy() || !CI->use_empty())
1953 // fputs(s,F) --> fwrite(s,1,strlen(s),F)
1954 uint64_t Len = GetStringLength(CI->getArgOperand(0));
1958 // Known to have no uses (see above).
1960 CI->getArgOperand(0),
1961 ConstantInt::get(DL->getIntPtrType(CI->getContext()), Len - 1),
1962 CI->getArgOperand(1), B, DL, TLI);
1965 Value *LibCallSimplifier::optimizePuts(CallInst *CI, IRBuilder<> &B) {
1966 Function *Callee = CI->getCalledFunction();
1967 // Require one fixed pointer argument and an integer/void result.
1968 FunctionType *FT = Callee->getFunctionType();
1969 if (FT->getNumParams() < 1 || !FT->getParamType(0)->isPointerTy() ||
1970 !(FT->getReturnType()->isIntegerTy() || FT->getReturnType()->isVoidTy()))
1973 // Check for a constant string.
1975 if (!getConstantStringInfo(CI->getArgOperand(0), Str))
1978 if (Str.empty() && CI->use_empty()) {
1979 // puts("") -> putchar('\n')
1980 Value *Res = EmitPutChar(B.getInt32('\n'), B, DL, TLI);
1981 if (CI->use_empty() || !Res)
1983 return B.CreateIntCast(Res, CI->getType(), true);
1989 bool LibCallSimplifier::hasFloatVersion(StringRef FuncName) {
1991 SmallString<20> FloatFuncName = FuncName;
1992 FloatFuncName += 'f';
1993 if (TLI->getLibFunc(FloatFuncName, Func))
1994 return TLI->has(Func);
1998 Value *LibCallSimplifier::optimizeCall(CallInst *CI) {
1999 if (CI->isNoBuiltin())
2003 Function *Callee = CI->getCalledFunction();
2004 StringRef FuncName = Callee->getName();
2005 IRBuilder<> Builder(CI);
2006 bool isCallingConvC = CI->getCallingConv() == llvm::CallingConv::C;
2008 // Command-line parameter overrides function attribute.
2009 if (EnableUnsafeFPShrink.getNumOccurrences() > 0)
2010 UnsafeFPShrink = EnableUnsafeFPShrink;
2011 else if (Callee->hasFnAttribute("unsafe-fp-math")) {
2012 // FIXME: This is the same problem as described in optimizeSqrt().
2013 // If calls gain access to IR-level FMF, then use that instead of a
2014 // function attribute.
2016 // Check for unsafe-fp-math = true.
2017 Attribute Attr = Callee->getFnAttribute("unsafe-fp-math");
2018 if (Attr.getValueAsString() == "true")
2019 UnsafeFPShrink = true;
2022 // First, check for intrinsics.
2023 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
2024 if (!isCallingConvC)
2026 switch (II->getIntrinsicID()) {
2027 case Intrinsic::pow:
2028 return optimizePow(CI, Builder);
2029 case Intrinsic::exp2:
2030 return optimizeExp2(CI, Builder);
2031 case Intrinsic::fabs:
2032 return optimizeFabs(CI, Builder);
2033 case Intrinsic::sqrt:
2034 return optimizeSqrt(CI, Builder);
2040 // Then check for known library functions.
2041 if (TLI->getLibFunc(FuncName, Func) && TLI->has(Func)) {
2042 // We never change the calling convention.
2043 if (!ignoreCallingConv(Func) && !isCallingConvC)
2046 case LibFunc::strcat:
2047 return optimizeStrCat(CI, Builder);
2048 case LibFunc::strncat:
2049 return optimizeStrNCat(CI, Builder);
2050 case LibFunc::strchr:
2051 return optimizeStrChr(CI, Builder);
2052 case LibFunc::strrchr:
2053 return optimizeStrRChr(CI, Builder);
2054 case LibFunc::strcmp:
2055 return optimizeStrCmp(CI, Builder);
2056 case LibFunc::strncmp:
2057 return optimizeStrNCmp(CI, Builder);
2058 case LibFunc::strcpy:
2059 return optimizeStrCpy(CI, Builder);
2060 case LibFunc::stpcpy:
2061 return optimizeStpCpy(CI, Builder);
2062 case LibFunc::strncpy:
2063 return optimizeStrNCpy(CI, Builder);
2064 case LibFunc::strlen:
2065 return optimizeStrLen(CI, Builder);
2066 case LibFunc::strpbrk:
2067 return optimizeStrPBrk(CI, Builder);
2068 case LibFunc::strtol:
2069 case LibFunc::strtod:
2070 case LibFunc::strtof:
2071 case LibFunc::strtoul:
2072 case LibFunc::strtoll:
2073 case LibFunc::strtold:
2074 case LibFunc::strtoull:
2075 return optimizeStrTo(CI, Builder);
2076 case LibFunc::strspn:
2077 return optimizeStrSpn(CI, Builder);
2078 case LibFunc::strcspn:
2079 return optimizeStrCSpn(CI, Builder);
2080 case LibFunc::strstr:
2081 return optimizeStrStr(CI, Builder);
2082 case LibFunc::memcmp:
2083 return optimizeMemCmp(CI, Builder);
2084 case LibFunc::memcpy:
2085 return optimizeMemCpy(CI, Builder);
2086 case LibFunc::memmove:
2087 return optimizeMemMove(CI, Builder);
2088 case LibFunc::memset:
2089 return optimizeMemSet(CI, Builder);
2093 return optimizeCos(CI, Builder);
2094 case LibFunc::sinpif:
2095 case LibFunc::sinpi:
2096 case LibFunc::cospif:
2097 case LibFunc::cospi:
2098 return optimizeSinCosPi(CI, Builder);
2102 return optimizePow(CI, Builder);
2103 case LibFunc::exp2l:
2105 case LibFunc::exp2f:
2106 return optimizeExp2(CI, Builder);
2107 case LibFunc::fabsf:
2109 case LibFunc::fabsl:
2110 return optimizeFabs(CI, Builder);
2111 case LibFunc::sqrtf:
2113 case LibFunc::sqrtl:
2114 return optimizeSqrt(CI, Builder);
2117 case LibFunc::ffsll:
2118 return optimizeFFS(CI, Builder);
2121 case LibFunc::llabs:
2122 return optimizeAbs(CI, Builder);
2123 case LibFunc::isdigit:
2124 return optimizeIsDigit(CI, Builder);
2125 case LibFunc::isascii:
2126 return optimizeIsAscii(CI, Builder);
2127 case LibFunc::toascii:
2128 return optimizeToAscii(CI, Builder);
2129 case LibFunc::printf:
2130 return optimizePrintF(CI, Builder);
2131 case LibFunc::sprintf:
2132 return optimizeSPrintF(CI, Builder);
2133 case LibFunc::fprintf:
2134 return optimizeFPrintF(CI, Builder);
2135 case LibFunc::fwrite:
2136 return optimizeFWrite(CI, Builder);
2137 case LibFunc::fputs:
2138 return optimizeFPuts(CI, Builder);
2140 return optimizePuts(CI, Builder);
2141 case LibFunc::perror:
2142 return optimizeErrorReporting(CI, Builder);
2143 case LibFunc::vfprintf:
2144 case LibFunc::fiprintf:
2145 return optimizeErrorReporting(CI, Builder, 0);
2146 case LibFunc::fputc:
2147 return optimizeErrorReporting(CI, Builder, 1);
2149 case LibFunc::floor:
2151 case LibFunc::round:
2152 case LibFunc::nearbyint:
2153 case LibFunc::trunc:
2154 if (hasFloatVersion(FuncName))
2155 return optimizeUnaryDoubleFP(CI, Builder, false);
2158 case LibFunc::acosh:
2160 case LibFunc::asinh:
2162 case LibFunc::atanh:
2166 case LibFunc::exp10:
2167 case LibFunc::expm1:
2169 case LibFunc::log10:
2170 case LibFunc::log1p:
2177 if (UnsafeFPShrink && hasFloatVersion(FuncName))
2178 return optimizeUnaryDoubleFP(CI, Builder, true);
2182 if (hasFloatVersion(FuncName))
2183 return optimizeBinaryDoubleFP(CI, Builder);
2185 case LibFunc::memcpy_chk:
2186 return optimizeMemCpyChk(CI, Builder);
2187 case LibFunc::memmove_chk:
2188 return optimizeMemMoveChk(CI, Builder);
2189 case LibFunc::memset_chk:
2190 return optimizeMemSetChk(CI, Builder);
2191 case LibFunc::strcpy_chk:
2192 return optimizeStrCpyChk(CI, Builder);
2193 case LibFunc::stpcpy_chk:
2194 return optimizeStpCpyChk(CI, Builder);
2195 case LibFunc::stpncpy_chk:
2196 case LibFunc::strncpy_chk:
2197 return optimizeStrNCpyChk(CI, Builder);
2206 LibCallSimplifier::LibCallSimplifier(const DataLayout *DL,
2207 const TargetLibraryInfo *TLI) :
2210 UnsafeFPShrink(false) {
2213 void LibCallSimplifier::replaceAllUsesWith(Instruction *I, Value *With) const {
2214 I->replaceAllUsesWith(With);
2215 I->eraseFromParent();
2219 // Additional cases that we need to add to this file:
2222 // * cbrt(expN(X)) -> expN(x/3)
2223 // * cbrt(sqrt(x)) -> pow(x,1/6)
2224 // * cbrt(sqrt(x)) -> pow(x,1/9)
2227 // * exp(log(x)) -> x
2230 // * log(exp(x)) -> x
2231 // * log(x**y) -> y*log(x)
2232 // * log(exp(y)) -> y*log(e)
2233 // * log(exp2(y)) -> y*log(2)
2234 // * log(exp10(y)) -> y*log(10)
2235 // * log(sqrt(x)) -> 0.5*log(x)
2236 // * log(pow(x,y)) -> y*log(x)
2238 // lround, lroundf, lroundl:
2239 // * lround(cnst) -> cnst'
2242 // * pow(exp(x),y) -> exp(x*y)
2243 // * pow(sqrt(x),y) -> pow(x,y*0.5)
2244 // * pow(pow(x,y),z)-> pow(x,y*z)
2246 // round, roundf, roundl:
2247 // * round(cnst) -> cnst'
2250 // * signbit(cnst) -> cnst'
2251 // * signbit(nncst) -> 0 (if pstv is a non-negative constant)
2253 // sqrt, sqrtf, sqrtl:
2254 // * sqrt(expN(x)) -> expN(x*0.5)
2255 // * sqrt(Nroot(x)) -> pow(x,1/(2*N))
2256 // * sqrt(pow(x,y)) -> pow(|x|,y*0.5)
2259 // * tan(atan(x)) -> x
2261 // trunc, truncf, truncl:
2262 // * trunc(cnst) -> cnst'