]>
git.proxmox.com Git - rustc.git/blob - src/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp
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 /// \brief Returns whether \p F matches the signature expected for the
120 /// string/memory copying library function \p Func.
121 /// Acceptable functions are st[rp][n]?cpy, memove, memcpy, and memset.
122 /// Their fortified (_chk) counterparts are also accepted.
123 static bool checkStringCopyLibFuncSignature(Function
*F
, LibFunc::Func Func
,
124 const DataLayout
*DL
) {
125 FunctionType
*FT
= F
->getFunctionType();
126 LLVMContext
&Context
= F
->getContext();
127 Type
*PCharTy
= Type::getInt8PtrTy(Context
);
128 Type
*SizeTTy
= DL
? DL
->getIntPtrType(Context
) : nullptr;
129 unsigned NumParams
= FT
->getNumParams();
131 // All string libfuncs return the same type as the first parameter.
132 if (FT
->getReturnType() != FT
->getParamType(0))
137 llvm_unreachable("Can't check signature for non-string-copy libfunc.");
138 case LibFunc::stpncpy_chk
:
139 case LibFunc::strncpy_chk
:
140 --NumParams
; // fallthrough
141 case LibFunc::stpncpy
:
142 case LibFunc::strncpy
: {
143 if (NumParams
!= 3 || FT
->getParamType(0) != FT
->getParamType(1) ||
144 FT
->getParamType(0) != PCharTy
|| !FT
->getParamType(2)->isIntegerTy())
148 case LibFunc::strcpy_chk
:
149 case LibFunc::stpcpy_chk
:
150 --NumParams
; // fallthrough
151 case LibFunc::stpcpy
:
152 case LibFunc::strcpy
: {
153 if (NumParams
!= 2 || FT
->getParamType(0) != FT
->getParamType(1) ||
154 FT
->getParamType(0) != PCharTy
)
158 case LibFunc::memmove_chk
:
159 case LibFunc::memcpy_chk
:
160 --NumParams
; // fallthrough
161 case LibFunc::memmove
:
162 case LibFunc::memcpy
: {
163 if (NumParams
!= 3 || !FT
->getParamType(0)->isPointerTy() ||
164 !FT
->getParamType(1)->isPointerTy() || FT
->getParamType(2) != SizeTTy
)
168 case LibFunc::memset_chk
:
169 --NumParams
; // fallthrough
170 case LibFunc::memset
: {
171 if (NumParams
!= 3 || !FT
->getParamType(0)->isPointerTy() ||
172 !FT
->getParamType(1)->isIntegerTy() || FT
->getParamType(2) != SizeTTy
)
177 // If this is a fortified libcall, the last parameter is a size_t.
178 if (NumParams
== FT
->getNumParams() - 1)
179 return FT
->getParamType(FT
->getNumParams() - 1) == SizeTTy
;
183 //===----------------------------------------------------------------------===//
184 // String and Memory Library Call Optimizations
185 //===----------------------------------------------------------------------===//
187 Value
*LibCallSimplifier::optimizeStrCat(CallInst
*CI
, IRBuilder
<> &B
) {
188 Function
*Callee
= CI
->getCalledFunction();
189 // Verify the "strcat" function prototype.
190 FunctionType
*FT
= Callee
->getFunctionType();
191 if (FT
->getNumParams() != 2||
192 FT
->getReturnType() != B
.getInt8PtrTy() ||
193 FT
->getParamType(0) != FT
->getReturnType() ||
194 FT
->getParamType(1) != FT
->getReturnType())
197 // Extract some information from the instruction
198 Value
*Dst
= CI
->getArgOperand(0);
199 Value
*Src
= CI
->getArgOperand(1);
201 // See if we can get the length of the input string.
202 uint64_t Len
= GetStringLength(Src
);
205 --Len
; // Unbias length.
207 // Handle the simple, do-nothing case: strcat(x, "") -> x
211 // These optimizations require DataLayout.
215 return emitStrLenMemCpy(Src
, Dst
, Len
, B
);
218 Value
*LibCallSimplifier::emitStrLenMemCpy(Value
*Src
, Value
*Dst
, uint64_t Len
,
220 // We need to find the end of the destination string. That's where the
221 // memory is to be moved to. We just generate a call to strlen.
222 Value
*DstLen
= EmitStrLen(Dst
, B
, DL
, TLI
);
226 // Now that we have the destination's length, we must index into the
227 // destination's pointer to get the actual memcpy destination (end of
228 // the string .. we're concatenating).
229 Value
*CpyDst
= B
.CreateGEP(Dst
, DstLen
, "endptr");
231 // We have enough information to now generate the memcpy call to do the
232 // concatenation for us. Make a memcpy to copy the nul byte with align = 1.
235 ConstantInt::get(DL
->getIntPtrType(Src
->getContext()), Len
+ 1), 1);
239 Value
*LibCallSimplifier::optimizeStrNCat(CallInst
*CI
, IRBuilder
<> &B
) {
240 Function
*Callee
= CI
->getCalledFunction();
241 // Verify the "strncat" function prototype.
242 FunctionType
*FT
= Callee
->getFunctionType();
243 if (FT
->getNumParams() != 3 || FT
->getReturnType() != B
.getInt8PtrTy() ||
244 FT
->getParamType(0) != FT
->getReturnType() ||
245 FT
->getParamType(1) != FT
->getReturnType() ||
246 !FT
->getParamType(2)->isIntegerTy())
249 // Extract some information from the instruction
250 Value
*Dst
= CI
->getArgOperand(0);
251 Value
*Src
= CI
->getArgOperand(1);
254 // We don't do anything if length is not constant
255 if (ConstantInt
*LengthArg
= dyn_cast
<ConstantInt
>(CI
->getArgOperand(2)))
256 Len
= LengthArg
->getZExtValue();
260 // See if we can get the length of the input string.
261 uint64_t SrcLen
= GetStringLength(Src
);
264 --SrcLen
; // Unbias length.
266 // Handle the simple, do-nothing cases:
267 // strncat(x, "", c) -> x
268 // strncat(x, c, 0) -> x
269 if (SrcLen
== 0 || Len
== 0)
272 // These optimizations require DataLayout.
276 // We don't optimize this case
280 // strncat(x, s, c) -> strcat(x, s)
281 // s is constant so the strcat can be optimized further
282 return emitStrLenMemCpy(Src
, Dst
, SrcLen
, B
);
285 Value
*LibCallSimplifier::optimizeStrChr(CallInst
*CI
, IRBuilder
<> &B
) {
286 Function
*Callee
= CI
->getCalledFunction();
287 // Verify the "strchr" function prototype.
288 FunctionType
*FT
= Callee
->getFunctionType();
289 if (FT
->getNumParams() != 2 || FT
->getReturnType() != B
.getInt8PtrTy() ||
290 FT
->getParamType(0) != FT
->getReturnType() ||
291 !FT
->getParamType(1)->isIntegerTy(32))
294 Value
*SrcStr
= CI
->getArgOperand(0);
296 // If the second operand is non-constant, see if we can compute the length
297 // of the input string and turn this into memchr.
298 ConstantInt
*CharC
= dyn_cast
<ConstantInt
>(CI
->getArgOperand(1));
300 // These optimizations require DataLayout.
304 uint64_t Len
= GetStringLength(SrcStr
);
305 if (Len
== 0 || !FT
->getParamType(1)->isIntegerTy(32)) // memchr needs i32.
309 SrcStr
, CI
->getArgOperand(1), // include nul.
310 ConstantInt::get(DL
->getIntPtrType(CI
->getContext()), Len
), B
, DL
, TLI
);
313 // Otherwise, the character is a constant, see if the first argument is
314 // a string literal. If so, we can constant fold.
316 if (!getConstantStringInfo(SrcStr
, Str
)) {
317 if (DL
&& CharC
->isZero()) // strchr(p, 0) -> p + strlen(p)
318 return B
.CreateGEP(SrcStr
, EmitStrLen(SrcStr
, B
, DL
, TLI
), "strchr");
322 // Compute the offset, make sure to handle the case when we're searching for
323 // zero (a weird way to spell strlen).
324 size_t I
= (0xFF & CharC
->getSExtValue()) == 0
326 : Str
.find(CharC
->getSExtValue());
327 if (I
== StringRef::npos
) // Didn't find the char. strchr returns null.
328 return Constant::getNullValue(CI
->getType());
330 // strchr(s+n,c) -> gep(s+n+i,c)
331 return B
.CreateGEP(SrcStr
, B
.getInt64(I
), "strchr");
334 Value
*LibCallSimplifier::optimizeStrRChr(CallInst
*CI
, IRBuilder
<> &B
) {
335 Function
*Callee
= CI
->getCalledFunction();
336 // Verify the "strrchr" function prototype.
337 FunctionType
*FT
= Callee
->getFunctionType();
338 if (FT
->getNumParams() != 2 || FT
->getReturnType() != B
.getInt8PtrTy() ||
339 FT
->getParamType(0) != FT
->getReturnType() ||
340 !FT
->getParamType(1)->isIntegerTy(32))
343 Value
*SrcStr
= CI
->getArgOperand(0);
344 ConstantInt
*CharC
= dyn_cast
<ConstantInt
>(CI
->getArgOperand(1));
346 // Cannot fold anything if we're not looking for a constant.
351 if (!getConstantStringInfo(SrcStr
, Str
)) {
352 // strrchr(s, 0) -> strchr(s, 0)
353 if (DL
&& CharC
->isZero())
354 return EmitStrChr(SrcStr
, '\0', B
, DL
, TLI
);
358 // Compute the offset.
359 size_t I
= (0xFF & CharC
->getSExtValue()) == 0
361 : Str
.rfind(CharC
->getSExtValue());
362 if (I
== StringRef::npos
) // Didn't find the char. Return null.
363 return Constant::getNullValue(CI
->getType());
365 // strrchr(s+n,c) -> gep(s+n+i,c)
366 return B
.CreateGEP(SrcStr
, B
.getInt64(I
), "strrchr");
369 Value
*LibCallSimplifier::optimizeStrCmp(CallInst
*CI
, IRBuilder
<> &B
) {
370 Function
*Callee
= CI
->getCalledFunction();
371 // Verify the "strcmp" function prototype.
372 FunctionType
*FT
= Callee
->getFunctionType();
373 if (FT
->getNumParams() != 2 || !FT
->getReturnType()->isIntegerTy(32) ||
374 FT
->getParamType(0) != FT
->getParamType(1) ||
375 FT
->getParamType(0) != B
.getInt8PtrTy())
378 Value
*Str1P
= CI
->getArgOperand(0), *Str2P
= CI
->getArgOperand(1);
379 if (Str1P
== Str2P
) // strcmp(x,x) -> 0
380 return ConstantInt::get(CI
->getType(), 0);
382 StringRef Str1
, Str2
;
383 bool HasStr1
= getConstantStringInfo(Str1P
, Str1
);
384 bool HasStr2
= getConstantStringInfo(Str2P
, Str2
);
386 // strcmp(x, y) -> cnst (if both x and y are constant strings)
387 if (HasStr1
&& HasStr2
)
388 return ConstantInt::get(CI
->getType(), Str1
.compare(Str2
));
390 if (HasStr1
&& Str1
.empty()) // strcmp("", x) -> -*x
392 B
.CreateZExt(B
.CreateLoad(Str2P
, "strcmpload"), CI
->getType()));
394 if (HasStr2
&& Str2
.empty()) // strcmp(x,"") -> *x
395 return B
.CreateZExt(B
.CreateLoad(Str1P
, "strcmpload"), CI
->getType());
397 // strcmp(P, "x") -> memcmp(P, "x", 2)
398 uint64_t Len1
= GetStringLength(Str1P
);
399 uint64_t Len2
= GetStringLength(Str2P
);
401 // These optimizations require DataLayout.
405 return EmitMemCmp(Str1P
, Str2P
,
406 ConstantInt::get(DL
->getIntPtrType(CI
->getContext()),
407 std::min(Len1
, Len2
)),
414 Value
*LibCallSimplifier::optimizeStrNCmp(CallInst
*CI
, IRBuilder
<> &B
) {
415 Function
*Callee
= CI
->getCalledFunction();
416 // Verify the "strncmp" function prototype.
417 FunctionType
*FT
= Callee
->getFunctionType();
418 if (FT
->getNumParams() != 3 || !FT
->getReturnType()->isIntegerTy(32) ||
419 FT
->getParamType(0) != FT
->getParamType(1) ||
420 FT
->getParamType(0) != B
.getInt8PtrTy() ||
421 !FT
->getParamType(2)->isIntegerTy())
424 Value
*Str1P
= CI
->getArgOperand(0), *Str2P
= CI
->getArgOperand(1);
425 if (Str1P
== Str2P
) // strncmp(x,x,n) -> 0
426 return ConstantInt::get(CI
->getType(), 0);
428 // Get the length argument if it is constant.
430 if (ConstantInt
*LengthArg
= dyn_cast
<ConstantInt
>(CI
->getArgOperand(2)))
431 Length
= LengthArg
->getZExtValue();
435 if (Length
== 0) // strncmp(x,y,0) -> 0
436 return ConstantInt::get(CI
->getType(), 0);
438 if (DL
&& Length
== 1) // strncmp(x,y,1) -> memcmp(x,y,1)
439 return EmitMemCmp(Str1P
, Str2P
, CI
->getArgOperand(2), B
, DL
, TLI
);
441 StringRef Str1
, Str2
;
442 bool HasStr1
= getConstantStringInfo(Str1P
, Str1
);
443 bool HasStr2
= getConstantStringInfo(Str2P
, Str2
);
445 // strncmp(x, y) -> cnst (if both x and y are constant strings)
446 if (HasStr1
&& HasStr2
) {
447 StringRef SubStr1
= Str1
.substr(0, Length
);
448 StringRef SubStr2
= Str2
.substr(0, Length
);
449 return ConstantInt::get(CI
->getType(), SubStr1
.compare(SubStr2
));
452 if (HasStr1
&& Str1
.empty()) // strncmp("", x, n) -> -*x
454 B
.CreateZExt(B
.CreateLoad(Str2P
, "strcmpload"), CI
->getType()));
456 if (HasStr2
&& Str2
.empty()) // strncmp(x, "", n) -> *x
457 return B
.CreateZExt(B
.CreateLoad(Str1P
, "strcmpload"), CI
->getType());
462 Value
*LibCallSimplifier::optimizeStrCpy(CallInst
*CI
, IRBuilder
<> &B
) {
463 Function
*Callee
= CI
->getCalledFunction();
465 if (!checkStringCopyLibFuncSignature(Callee
, LibFunc::strcpy
, DL
))
468 Value
*Dst
= CI
->getArgOperand(0), *Src
= CI
->getArgOperand(1);
469 if (Dst
== Src
) // strcpy(x,x) -> x
472 // These optimizations require DataLayout.
476 // See if we can get the length of the input string.
477 uint64_t Len
= GetStringLength(Src
);
481 // We have enough information to now generate the memcpy call to do the
482 // copy for us. Make a memcpy to copy the nul byte with align = 1.
483 B
.CreateMemCpy(Dst
, Src
,
484 ConstantInt::get(DL
->getIntPtrType(CI
->getContext()), Len
), 1);
488 Value
*LibCallSimplifier::optimizeStpCpy(CallInst
*CI
, IRBuilder
<> &B
) {
489 Function
*Callee
= CI
->getCalledFunction();
490 // Verify the "stpcpy" function prototype.
491 FunctionType
*FT
= Callee
->getFunctionType();
493 if (!checkStringCopyLibFuncSignature(Callee
, LibFunc::stpcpy
, DL
))
496 // These optimizations require DataLayout.
500 Value
*Dst
= CI
->getArgOperand(0), *Src
= CI
->getArgOperand(1);
501 if (Dst
== Src
) { // stpcpy(x,x) -> x+strlen(x)
502 Value
*StrLen
= EmitStrLen(Src
, B
, DL
, TLI
);
503 return StrLen
? B
.CreateInBoundsGEP(Dst
, StrLen
) : nullptr;
506 // See if we can get the length of the input string.
507 uint64_t Len
= GetStringLength(Src
);
511 Type
*PT
= FT
->getParamType(0);
512 Value
*LenV
= ConstantInt::get(DL
->getIntPtrType(PT
), Len
);
514 B
.CreateGEP(Dst
, ConstantInt::get(DL
->getIntPtrType(PT
), Len
- 1));
516 // We have enough information to now generate the memcpy call to do the
517 // copy for us. Make a memcpy to copy the nul byte with align = 1.
518 B
.CreateMemCpy(Dst
, Src
, LenV
, 1);
522 Value
*LibCallSimplifier::optimizeStrNCpy(CallInst
*CI
, IRBuilder
<> &B
) {
523 Function
*Callee
= CI
->getCalledFunction();
524 FunctionType
*FT
= Callee
->getFunctionType();
526 if (!checkStringCopyLibFuncSignature(Callee
, LibFunc::strncpy
, DL
))
529 Value
*Dst
= CI
->getArgOperand(0);
530 Value
*Src
= CI
->getArgOperand(1);
531 Value
*LenOp
= CI
->getArgOperand(2);
533 // See if we can get the length of the input string.
534 uint64_t SrcLen
= GetStringLength(Src
);
540 // strncpy(x, "", y) -> memset(x, '\0', y, 1)
541 B
.CreateMemSet(Dst
, B
.getInt8('\0'), LenOp
, 1);
546 if (ConstantInt
*LengthArg
= dyn_cast
<ConstantInt
>(LenOp
))
547 Len
= LengthArg
->getZExtValue();
552 return Dst
; // strncpy(x, y, 0) -> x
554 // These optimizations require DataLayout.
558 // Let strncpy handle the zero padding
559 if (Len
> SrcLen
+ 1)
562 Type
*PT
= FT
->getParamType(0);
563 // strncpy(x, s, c) -> memcpy(x, s, c, 1) [s and c are constant]
564 B
.CreateMemCpy(Dst
, Src
, ConstantInt::get(DL
->getIntPtrType(PT
), Len
), 1);
569 Value
*LibCallSimplifier::optimizeStrLen(CallInst
*CI
, IRBuilder
<> &B
) {
570 Function
*Callee
= CI
->getCalledFunction();
571 FunctionType
*FT
= Callee
->getFunctionType();
572 if (FT
->getNumParams() != 1 || FT
->getParamType(0) != B
.getInt8PtrTy() ||
573 !FT
->getReturnType()->isIntegerTy())
576 Value
*Src
= CI
->getArgOperand(0);
578 // Constant folding: strlen("xyz") -> 3
579 if (uint64_t Len
= GetStringLength(Src
))
580 return ConstantInt::get(CI
->getType(), Len
- 1);
582 // strlen(x?"foo":"bars") --> x ? 3 : 4
583 if (SelectInst
*SI
= dyn_cast
<SelectInst
>(Src
)) {
584 uint64_t LenTrue
= GetStringLength(SI
->getTrueValue());
585 uint64_t LenFalse
= GetStringLength(SI
->getFalseValue());
586 if (LenTrue
&& LenFalse
) {
587 Function
*Caller
= CI
->getParent()->getParent();
588 emitOptimizationRemark(CI
->getContext(), "simplify-libcalls", *Caller
,
590 "folded strlen(select) to select of constants");
591 return B
.CreateSelect(SI
->getCondition(),
592 ConstantInt::get(CI
->getType(), LenTrue
- 1),
593 ConstantInt::get(CI
->getType(), LenFalse
- 1));
597 // strlen(x) != 0 --> *x != 0
598 // strlen(x) == 0 --> *x == 0
599 if (isOnlyUsedInZeroEqualityComparison(CI
))
600 return B
.CreateZExt(B
.CreateLoad(Src
, "strlenfirst"), CI
->getType());
605 Value
*LibCallSimplifier::optimizeStrPBrk(CallInst
*CI
, IRBuilder
<> &B
) {
606 Function
*Callee
= CI
->getCalledFunction();
607 FunctionType
*FT
= Callee
->getFunctionType();
608 if (FT
->getNumParams() != 2 || FT
->getParamType(0) != B
.getInt8PtrTy() ||
609 FT
->getParamType(1) != FT
->getParamType(0) ||
610 FT
->getReturnType() != FT
->getParamType(0))
614 bool HasS1
= getConstantStringInfo(CI
->getArgOperand(0), S1
);
615 bool HasS2
= getConstantStringInfo(CI
->getArgOperand(1), S2
);
617 // strpbrk(s, "") -> nullptr
618 // strpbrk("", s) -> nullptr
619 if ((HasS1
&& S1
.empty()) || (HasS2
&& S2
.empty()))
620 return Constant::getNullValue(CI
->getType());
623 if (HasS1
&& HasS2
) {
624 size_t I
= S1
.find_first_of(S2
);
625 if (I
== StringRef::npos
) // No match.
626 return Constant::getNullValue(CI
->getType());
628 return B
.CreateGEP(CI
->getArgOperand(0), B
.getInt64(I
), "strpbrk");
631 // strpbrk(s, "a") -> strchr(s, 'a')
632 if (DL
&& HasS2
&& S2
.size() == 1)
633 return EmitStrChr(CI
->getArgOperand(0), S2
[0], B
, DL
, TLI
);
638 Value
*LibCallSimplifier::optimizeStrTo(CallInst
*CI
, IRBuilder
<> &B
) {
639 Function
*Callee
= CI
->getCalledFunction();
640 FunctionType
*FT
= Callee
->getFunctionType();
641 if ((FT
->getNumParams() != 2 && FT
->getNumParams() != 3) ||
642 !FT
->getParamType(0)->isPointerTy() ||
643 !FT
->getParamType(1)->isPointerTy())
646 Value
*EndPtr
= CI
->getArgOperand(1);
647 if (isa
<ConstantPointerNull
>(EndPtr
)) {
648 // With a null EndPtr, this function won't capture the main argument.
649 // It would be readonly too, except that it still may write to errno.
650 CI
->addAttribute(1, Attribute::NoCapture
);
656 Value
*LibCallSimplifier::optimizeStrSpn(CallInst
*CI
, IRBuilder
<> &B
) {
657 Function
*Callee
= CI
->getCalledFunction();
658 FunctionType
*FT
= Callee
->getFunctionType();
659 if (FT
->getNumParams() != 2 || FT
->getParamType(0) != B
.getInt8PtrTy() ||
660 FT
->getParamType(1) != FT
->getParamType(0) ||
661 !FT
->getReturnType()->isIntegerTy())
665 bool HasS1
= getConstantStringInfo(CI
->getArgOperand(0), S1
);
666 bool HasS2
= getConstantStringInfo(CI
->getArgOperand(1), S2
);
668 // strspn(s, "") -> 0
669 // strspn("", s) -> 0
670 if ((HasS1
&& S1
.empty()) || (HasS2
&& S2
.empty()))
671 return Constant::getNullValue(CI
->getType());
674 if (HasS1
&& HasS2
) {
675 size_t Pos
= S1
.find_first_not_of(S2
);
676 if (Pos
== StringRef::npos
)
678 return ConstantInt::get(CI
->getType(), Pos
);
684 Value
*LibCallSimplifier::optimizeStrCSpn(CallInst
*CI
, IRBuilder
<> &B
) {
685 Function
*Callee
= CI
->getCalledFunction();
686 FunctionType
*FT
= Callee
->getFunctionType();
687 if (FT
->getNumParams() != 2 || FT
->getParamType(0) != B
.getInt8PtrTy() ||
688 FT
->getParamType(1) != FT
->getParamType(0) ||
689 !FT
->getReturnType()->isIntegerTy())
693 bool HasS1
= getConstantStringInfo(CI
->getArgOperand(0), S1
);
694 bool HasS2
= getConstantStringInfo(CI
->getArgOperand(1), S2
);
696 // strcspn("", s) -> 0
697 if (HasS1
&& S1
.empty())
698 return Constant::getNullValue(CI
->getType());
701 if (HasS1
&& HasS2
) {
702 size_t Pos
= S1
.find_first_of(S2
);
703 if (Pos
== StringRef::npos
)
705 return ConstantInt::get(CI
->getType(), Pos
);
708 // strcspn(s, "") -> strlen(s)
709 if (DL
&& HasS2
&& S2
.empty())
710 return EmitStrLen(CI
->getArgOperand(0), B
, DL
, TLI
);
715 Value
*LibCallSimplifier::optimizeStrStr(CallInst
*CI
, IRBuilder
<> &B
) {
716 Function
*Callee
= CI
->getCalledFunction();
717 FunctionType
*FT
= Callee
->getFunctionType();
718 if (FT
->getNumParams() != 2 || !FT
->getParamType(0)->isPointerTy() ||
719 !FT
->getParamType(1)->isPointerTy() ||
720 !FT
->getReturnType()->isPointerTy())
723 // fold strstr(x, x) -> x.
724 if (CI
->getArgOperand(0) == CI
->getArgOperand(1))
725 return B
.CreateBitCast(CI
->getArgOperand(0), CI
->getType());
727 // fold strstr(a, b) == a -> strncmp(a, b, strlen(b)) == 0
728 if (DL
&& isOnlyUsedInEqualityComparison(CI
, CI
->getArgOperand(0))) {
729 Value
*StrLen
= EmitStrLen(CI
->getArgOperand(1), B
, DL
, TLI
);
732 Value
*StrNCmp
= EmitStrNCmp(CI
->getArgOperand(0), CI
->getArgOperand(1),
736 for (auto UI
= CI
->user_begin(), UE
= CI
->user_end(); UI
!= UE
;) {
737 ICmpInst
*Old
= cast
<ICmpInst
>(*UI
++);
739 B
.CreateICmp(Old
->getPredicate(), StrNCmp
,
740 ConstantInt::getNullValue(StrNCmp
->getType()), "cmp");
741 replaceAllUsesWith(Old
, Cmp
);
746 // See if either input string is a constant string.
747 StringRef SearchStr
, ToFindStr
;
748 bool HasStr1
= getConstantStringInfo(CI
->getArgOperand(0), SearchStr
);
749 bool HasStr2
= getConstantStringInfo(CI
->getArgOperand(1), ToFindStr
);
751 // fold strstr(x, "") -> x.
752 if (HasStr2
&& ToFindStr
.empty())
753 return B
.CreateBitCast(CI
->getArgOperand(0), CI
->getType());
755 // If both strings are known, constant fold it.
756 if (HasStr1
&& HasStr2
) {
757 size_t Offset
= SearchStr
.find(ToFindStr
);
759 if (Offset
== StringRef::npos
) // strstr("foo", "bar") -> null
760 return Constant::getNullValue(CI
->getType());
762 // strstr("abcd", "bc") -> gep((char*)"abcd", 1)
763 Value
*Result
= CastToCStr(CI
->getArgOperand(0), B
);
764 Result
= B
.CreateConstInBoundsGEP1_64(Result
, Offset
, "strstr");
765 return B
.CreateBitCast(Result
, CI
->getType());
768 // fold strstr(x, "y") -> strchr(x, 'y').
769 if (HasStr2
&& ToFindStr
.size() == 1) {
770 Value
*StrChr
= EmitStrChr(CI
->getArgOperand(0), ToFindStr
[0], B
, DL
, TLI
);
771 return StrChr
? B
.CreateBitCast(StrChr
, CI
->getType()) : nullptr;
776 Value
*LibCallSimplifier::optimizeMemCmp(CallInst
*CI
, IRBuilder
<> &B
) {
777 Function
*Callee
= CI
->getCalledFunction();
778 FunctionType
*FT
= Callee
->getFunctionType();
779 if (FT
->getNumParams() != 3 || !FT
->getParamType(0)->isPointerTy() ||
780 !FT
->getParamType(1)->isPointerTy() ||
781 !FT
->getReturnType()->isIntegerTy(32))
784 Value
*LHS
= CI
->getArgOperand(0), *RHS
= CI
->getArgOperand(1);
786 if (LHS
== RHS
) // memcmp(s,s,x) -> 0
787 return Constant::getNullValue(CI
->getType());
789 // Make sure we have a constant length.
790 ConstantInt
*LenC
= dyn_cast
<ConstantInt
>(CI
->getArgOperand(2));
793 uint64_t Len
= LenC
->getZExtValue();
795 if (Len
== 0) // memcmp(s1,s2,0) -> 0
796 return Constant::getNullValue(CI
->getType());
798 // memcmp(S1,S2,1) -> *(unsigned char*)LHS - *(unsigned char*)RHS
800 Value
*LHSV
= B
.CreateZExt(B
.CreateLoad(CastToCStr(LHS
, B
), "lhsc"),
801 CI
->getType(), "lhsv");
802 Value
*RHSV
= B
.CreateZExt(B
.CreateLoad(CastToCStr(RHS
, B
), "rhsc"),
803 CI
->getType(), "rhsv");
804 return B
.CreateSub(LHSV
, RHSV
, "chardiff");
807 // Constant folding: memcmp(x, y, l) -> cnst (all arguments are constant)
808 StringRef LHSStr
, RHSStr
;
809 if (getConstantStringInfo(LHS
, LHSStr
) &&
810 getConstantStringInfo(RHS
, RHSStr
)) {
811 // Make sure we're not reading out-of-bounds memory.
812 if (Len
> LHSStr
.size() || Len
> RHSStr
.size())
814 // Fold the memcmp and normalize the result. This way we get consistent
815 // results across multiple platforms.
817 int Cmp
= memcmp(LHSStr
.data(), RHSStr
.data(), Len
);
822 return ConstantInt::get(CI
->getType(), Ret
);
828 Value
*LibCallSimplifier::optimizeMemCpy(CallInst
*CI
, IRBuilder
<> &B
) {
829 Function
*Callee
= CI
->getCalledFunction();
830 // These optimizations require DataLayout.
834 if (!checkStringCopyLibFuncSignature(Callee
, LibFunc::memcpy
, DL
))
837 // memcpy(x, y, n) -> llvm.memcpy(x, y, n, 1)
838 B
.CreateMemCpy(CI
->getArgOperand(0), CI
->getArgOperand(1),
839 CI
->getArgOperand(2), 1);
840 return CI
->getArgOperand(0);
843 Value
*LibCallSimplifier::optimizeMemMove(CallInst
*CI
, IRBuilder
<> &B
) {
844 Function
*Callee
= CI
->getCalledFunction();
845 // These optimizations require DataLayout.
849 if (!checkStringCopyLibFuncSignature(Callee
, LibFunc::memmove
, DL
))
852 // memmove(x, y, n) -> llvm.memmove(x, y, n, 1)
853 B
.CreateMemMove(CI
->getArgOperand(0), CI
->getArgOperand(1),
854 CI
->getArgOperand(2), 1);
855 return CI
->getArgOperand(0);
858 Value
*LibCallSimplifier::optimizeMemSet(CallInst
*CI
, IRBuilder
<> &B
) {
859 Function
*Callee
= CI
->getCalledFunction();
860 // These optimizations require DataLayout.
864 if (!checkStringCopyLibFuncSignature(Callee
, LibFunc::memset
, DL
))
867 // memset(p, v, n) -> llvm.memset(p, v, n, 1)
868 Value
*Val
= B
.CreateIntCast(CI
->getArgOperand(1), B
.getInt8Ty(), false);
869 B
.CreateMemSet(CI
->getArgOperand(0), Val
, CI
->getArgOperand(2), 1);
870 return CI
->getArgOperand(0);
873 //===----------------------------------------------------------------------===//
874 // Math Library Optimizations
875 //===----------------------------------------------------------------------===//
877 /// Return a variant of Val with float type.
878 /// Currently this works in two cases: If Val is an FPExtension of a float
879 /// value to something bigger, simply return the operand.
880 /// If Val is a ConstantFP but can be converted to a float ConstantFP without
881 /// loss of precision do so.
882 static Value
*valueHasFloatPrecision(Value
*Val
) {
883 if (FPExtInst
*Cast
= dyn_cast
<FPExtInst
>(Val
)) {
884 Value
*Op
= Cast
->getOperand(0);
885 if (Op
->getType()->isFloatTy())
888 if (ConstantFP
*Const
= dyn_cast
<ConstantFP
>(Val
)) {
889 APFloat F
= Const
->getValueAPF();
891 (void)F
.convert(APFloat::IEEEsingle
, APFloat::rmNearestTiesToEven
,
894 return ConstantFP::get(Const
->getContext(), F
);
899 //===----------------------------------------------------------------------===//
900 // Double -> Float Shrinking Optimizations for Unary Functions like 'floor'
902 Value
*LibCallSimplifier::optimizeUnaryDoubleFP(CallInst
*CI
, IRBuilder
<> &B
,
904 Function
*Callee
= CI
->getCalledFunction();
905 FunctionType
*FT
= Callee
->getFunctionType();
906 if (FT
->getNumParams() != 1 || !FT
->getReturnType()->isDoubleTy() ||
907 !FT
->getParamType(0)->isDoubleTy())
911 // Check if all the uses for function like 'sin' are converted to float.
912 for (User
*U
: CI
->users()) {
913 FPTruncInst
*Cast
= dyn_cast
<FPTruncInst
>(U
);
914 if (!Cast
|| !Cast
->getType()->isFloatTy())
919 // If this is something like 'floor((double)floatval)', convert to floorf.
920 Value
*V
= valueHasFloatPrecision(CI
->getArgOperand(0));
924 // floor((double)floatval) -> (double)floorf(floatval)
925 if (Callee
->isIntrinsic()) {
926 Module
*M
= CI
->getParent()->getParent()->getParent();
927 Intrinsic::ID IID
= (Intrinsic::ID
) Callee
->getIntrinsicID();
928 Function
*F
= Intrinsic::getDeclaration(M
, IID
, B
.getFloatTy());
929 V
= B
.CreateCall(F
, V
);
931 // The call is a library call rather than an intrinsic.
932 V
= EmitUnaryFloatFnCall(V
, Callee
->getName(), B
, Callee
->getAttributes());
935 return B
.CreateFPExt(V
, B
.getDoubleTy());
938 // Double -> Float Shrinking Optimizations for Binary Functions like 'fmin/fmax'
939 Value
*LibCallSimplifier::optimizeBinaryDoubleFP(CallInst
*CI
, IRBuilder
<> &B
) {
940 Function
*Callee
= CI
->getCalledFunction();
941 FunctionType
*FT
= Callee
->getFunctionType();
942 // Just make sure this has 2 arguments of the same FP type, which match the
944 if (FT
->getNumParams() != 2 || FT
->getReturnType() != FT
->getParamType(0) ||
945 FT
->getParamType(0) != FT
->getParamType(1) ||
946 !FT
->getParamType(0)->isFloatingPointTy())
949 // If this is something like 'fmin((double)floatval1, (double)floatval2)',
950 // or fmin(1.0, (double)floatval), then we convert it to fminf.
951 Value
*V1
= valueHasFloatPrecision(CI
->getArgOperand(0));
954 Value
*V2
= valueHasFloatPrecision(CI
->getArgOperand(1));
958 // fmin((double)floatval1, (double)floatval2)
959 // -> (double)fminf(floatval1, floatval2)
960 // TODO: Handle intrinsics in the same way as in optimizeUnaryDoubleFP().
961 Value
*V
= EmitBinaryFloatFnCall(V1
, V2
, Callee
->getName(), B
,
962 Callee
->getAttributes());
963 return B
.CreateFPExt(V
, B
.getDoubleTy());
966 Value
*LibCallSimplifier::optimizeCos(CallInst
*CI
, IRBuilder
<> &B
) {
967 Function
*Callee
= CI
->getCalledFunction();
968 Value
*Ret
= nullptr;
969 if (UnsafeFPShrink
&& Callee
->getName() == "cos" && TLI
->has(LibFunc::cosf
)) {
970 Ret
= optimizeUnaryDoubleFP(CI
, B
, true);
973 FunctionType
*FT
= Callee
->getFunctionType();
974 // Just make sure this has 1 argument of FP type, which matches the
976 if (FT
->getNumParams() != 1 || FT
->getReturnType() != FT
->getParamType(0) ||
977 !FT
->getParamType(0)->isFloatingPointTy())
981 Value
*Op1
= CI
->getArgOperand(0);
982 if (BinaryOperator::isFNeg(Op1
)) {
983 BinaryOperator
*BinExpr
= cast
<BinaryOperator
>(Op1
);
984 return B
.CreateCall(Callee
, BinExpr
->getOperand(1), "cos");
989 Value
*LibCallSimplifier::optimizePow(CallInst
*CI
, IRBuilder
<> &B
) {
990 Function
*Callee
= CI
->getCalledFunction();
992 Value
*Ret
= nullptr;
993 if (UnsafeFPShrink
&& Callee
->getName() == "pow" && TLI
->has(LibFunc::powf
)) {
994 Ret
= optimizeUnaryDoubleFP(CI
, B
, true);
997 FunctionType
*FT
= Callee
->getFunctionType();
998 // Just make sure this has 2 arguments of the same FP type, which match the
1000 if (FT
->getNumParams() != 2 || FT
->getReturnType() != FT
->getParamType(0) ||
1001 FT
->getParamType(0) != FT
->getParamType(1) ||
1002 !FT
->getParamType(0)->isFloatingPointTy())
1005 Value
*Op1
= CI
->getArgOperand(0), *Op2
= CI
->getArgOperand(1);
1006 if (ConstantFP
*Op1C
= dyn_cast
<ConstantFP
>(Op1
)) {
1007 // pow(1.0, x) -> 1.0
1008 if (Op1C
->isExactlyValue(1.0))
1010 // pow(2.0, x) -> exp2(x)
1011 if (Op1C
->isExactlyValue(2.0) &&
1012 hasUnaryFloatFn(TLI
, Op1
->getType(), LibFunc::exp2
, LibFunc::exp2f
,
1014 return EmitUnaryFloatFnCall(Op2
, "exp2", B
, Callee
->getAttributes());
1015 // pow(10.0, x) -> exp10(x)
1016 if (Op1C
->isExactlyValue(10.0) &&
1017 hasUnaryFloatFn(TLI
, Op1
->getType(), LibFunc::exp10
, LibFunc::exp10f
,
1019 return EmitUnaryFloatFnCall(Op2
, TLI
->getName(LibFunc::exp10
), B
,
1020 Callee
->getAttributes());
1023 ConstantFP
*Op2C
= dyn_cast
<ConstantFP
>(Op2
);
1027 if (Op2C
->getValueAPF().isZero()) // pow(x, 0.0) -> 1.0
1028 return ConstantFP::get(CI
->getType(), 1.0);
1030 if (Op2C
->isExactlyValue(0.5) &&
1031 hasUnaryFloatFn(TLI
, Op2
->getType(), LibFunc::sqrt
, LibFunc::sqrtf
,
1033 hasUnaryFloatFn(TLI
, Op2
->getType(), LibFunc::fabs
, LibFunc::fabsf
,
1035 // Expand pow(x, 0.5) to (x == -infinity ? +infinity : fabs(sqrt(x))).
1036 // This is faster than calling pow, and still handles negative zero
1037 // and negative infinity correctly.
1038 // TODO: In fast-math mode, this could be just sqrt(x).
1039 // TODO: In finite-only mode, this could be just fabs(sqrt(x)).
1040 Value
*Inf
= ConstantFP::getInfinity(CI
->getType());
1041 Value
*NegInf
= ConstantFP::getInfinity(CI
->getType(), true);
1042 Value
*Sqrt
= EmitUnaryFloatFnCall(Op1
, "sqrt", B
, Callee
->getAttributes());
1044 EmitUnaryFloatFnCall(Sqrt
, "fabs", B
, Callee
->getAttributes());
1045 Value
*FCmp
= B
.CreateFCmpOEQ(Op1
, NegInf
);
1046 Value
*Sel
= B
.CreateSelect(FCmp
, Inf
, FAbs
);
1050 if (Op2C
->isExactlyValue(1.0)) // pow(x, 1.0) -> x
1052 if (Op2C
->isExactlyValue(2.0)) // pow(x, 2.0) -> x*x
1053 return B
.CreateFMul(Op1
, Op1
, "pow2");
1054 if (Op2C
->isExactlyValue(-1.0)) // pow(x, -1.0) -> 1.0/x
1055 return B
.CreateFDiv(ConstantFP::get(CI
->getType(), 1.0), Op1
, "powrecip");
1059 Value
*LibCallSimplifier::optimizeExp2(CallInst
*CI
, IRBuilder
<> &B
) {
1060 Function
*Callee
= CI
->getCalledFunction();
1061 Function
*Caller
= CI
->getParent()->getParent();
1063 Value
*Ret
= nullptr;
1064 if (UnsafeFPShrink
&& Callee
->getName() == "exp2" &&
1065 TLI
->has(LibFunc::exp2f
)) {
1066 Ret
= optimizeUnaryDoubleFP(CI
, B
, true);
1069 FunctionType
*FT
= Callee
->getFunctionType();
1070 // Just make sure this has 1 argument of FP type, which matches the
1072 if (FT
->getNumParams() != 1 || FT
->getReturnType() != FT
->getParamType(0) ||
1073 !FT
->getParamType(0)->isFloatingPointTy())
1076 Value
*Op
= CI
->getArgOperand(0);
1077 // Turn exp2(sitofp(x)) -> ldexp(1.0, sext(x)) if sizeof(x) <= 32
1078 // Turn exp2(uitofp(x)) -> ldexp(1.0, zext(x)) if sizeof(x) < 32
1079 LibFunc::Func LdExp
= LibFunc::ldexpl
;
1080 if (Op
->getType()->isFloatTy())
1081 LdExp
= LibFunc::ldexpf
;
1082 else if (Op
->getType()->isDoubleTy())
1083 LdExp
= LibFunc::ldexp
;
1085 if (TLI
->has(LdExp
)) {
1086 Value
*LdExpArg
= nullptr;
1087 if (SIToFPInst
*OpC
= dyn_cast
<SIToFPInst
>(Op
)) {
1088 if (OpC
->getOperand(0)->getType()->getPrimitiveSizeInBits() <= 32)
1089 LdExpArg
= B
.CreateSExt(OpC
->getOperand(0), B
.getInt32Ty());
1090 } else if (UIToFPInst
*OpC
= dyn_cast
<UIToFPInst
>(Op
)) {
1091 if (OpC
->getOperand(0)->getType()->getPrimitiveSizeInBits() < 32)
1092 LdExpArg
= B
.CreateZExt(OpC
->getOperand(0), B
.getInt32Ty());
1096 Constant
*One
= ConstantFP::get(CI
->getContext(), APFloat(1.0f
));
1097 if (!Op
->getType()->isFloatTy())
1098 One
= ConstantExpr::getFPExtend(One
, Op
->getType());
1100 Module
*M
= Caller
->getParent();
1102 M
->getOrInsertFunction(TLI
->getName(LdExp
), Op
->getType(),
1103 Op
->getType(), B
.getInt32Ty(), nullptr);
1104 CallInst
*CI
= B
.CreateCall2(Callee
, One
, LdExpArg
);
1105 if (const Function
*F
= dyn_cast
<Function
>(Callee
->stripPointerCasts()))
1106 CI
->setCallingConv(F
->getCallingConv());
1114 Value
*LibCallSimplifier::optimizeFabs(CallInst
*CI
, IRBuilder
<> &B
) {
1115 Function
*Callee
= CI
->getCalledFunction();
1117 Value
*Ret
= nullptr;
1118 if (Callee
->getName() == "fabs" && TLI
->has(LibFunc::fabsf
)) {
1119 Ret
= optimizeUnaryDoubleFP(CI
, B
, false);
1122 FunctionType
*FT
= Callee
->getFunctionType();
1123 // Make sure this has 1 argument of FP type which matches the result type.
1124 if (FT
->getNumParams() != 1 || FT
->getReturnType() != FT
->getParamType(0) ||
1125 !FT
->getParamType(0)->isFloatingPointTy())
1128 Value
*Op
= CI
->getArgOperand(0);
1129 if (Instruction
*I
= dyn_cast
<Instruction
>(Op
)) {
1130 // Fold fabs(x * x) -> x * x; any squared FP value must already be positive.
1131 if (I
->getOpcode() == Instruction::FMul
)
1132 if (I
->getOperand(0) == I
->getOperand(1))
1138 Value
*LibCallSimplifier::optimizeSqrt(CallInst
*CI
, IRBuilder
<> &B
) {
1139 Function
*Callee
= CI
->getCalledFunction();
1141 Value
*Ret
= nullptr;
1142 if (TLI
->has(LibFunc::sqrtf
) && (Callee
->getName() == "sqrt" ||
1143 Callee
->getIntrinsicID() == Intrinsic::sqrt
))
1144 Ret
= optimizeUnaryDoubleFP(CI
, B
, true);
1146 // FIXME: For finer-grain optimization, we need intrinsics to have the same
1147 // fast-math flag decorations that are applied to FP instructions. For now,
1148 // we have to rely on the function-level unsafe-fp-math attribute to do this
1149 // optimization because there's no other way to express that the sqrt can be
1151 Function
*F
= CI
->getParent()->getParent();
1152 if (F
->hasFnAttribute("unsafe-fp-math")) {
1153 // Check for unsafe-fp-math = true.
1154 Attribute Attr
= F
->getFnAttribute("unsafe-fp-math");
1155 if (Attr
.getValueAsString() != "true")
1158 Value
*Op
= CI
->getArgOperand(0);
1159 if (Instruction
*I
= dyn_cast
<Instruction
>(Op
)) {
1160 if (I
->getOpcode() == Instruction::FMul
&& I
->hasUnsafeAlgebra()) {
1161 // We're looking for a repeated factor in a multiplication tree,
1162 // so we can do this fold: sqrt(x * x) -> fabs(x);
1163 // or this fold: sqrt(x * x * y) -> fabs(x) * sqrt(y).
1164 Value
*Op0
= I
->getOperand(0);
1165 Value
*Op1
= I
->getOperand(1);
1166 Value
*RepeatOp
= nullptr;
1167 Value
*OtherOp
= nullptr;
1169 // Simple match: the operands of the multiply are identical.
1172 // Look for a more complicated pattern: one of the operands is itself
1173 // a multiply, so search for a common factor in that multiply.
1174 // Note: We don't bother looking any deeper than this first level or for
1175 // variations of this pattern because instcombine's visitFMUL and/or the
1176 // reassociation pass should give us this form.
1177 Value
*OtherMul0
, *OtherMul1
;
1178 if (match(Op0
, m_FMul(m_Value(OtherMul0
), m_Value(OtherMul1
)))) {
1179 // Pattern: sqrt((x * y) * z)
1180 if (OtherMul0
== OtherMul1
) {
1181 // Matched: sqrt((x * x) * z)
1182 RepeatOp
= OtherMul0
;
1188 // Fast math flags for any created instructions should match the sqrt
1190 // FIXME: We're not checking the sqrt because it doesn't have
1191 // fast-math-flags (see earlier comment).
1192 IRBuilder
<true, ConstantFolder
,
1193 IRBuilderDefaultInserter
<true> >::FastMathFlagGuard
Guard(B
);
1194 B
.SetFastMathFlags(I
->getFastMathFlags());
1195 // If we found a repeated factor, hoist it out of the square root and
1196 // replace it with the fabs of that factor.
1197 Module
*M
= Callee
->getParent();
1198 Type
*ArgType
= Op
->getType();
1199 Value
*Fabs
= Intrinsic::getDeclaration(M
, Intrinsic::fabs
, ArgType
);
1200 Value
*FabsCall
= B
.CreateCall(Fabs
, RepeatOp
, "fabs");
1202 // If we found a non-repeated factor, we still need to get its square
1203 // root. We then multiply that by the value that was simplified out
1204 // of the square root calculation.
1205 Value
*Sqrt
= Intrinsic::getDeclaration(M
, Intrinsic::sqrt
, ArgType
);
1206 Value
*SqrtCall
= B
.CreateCall(Sqrt
, OtherOp
, "sqrt");
1207 return B
.CreateFMul(FabsCall
, SqrtCall
);
1216 static bool isTrigLibCall(CallInst
*CI
);
1217 static void insertSinCosCall(IRBuilder
<> &B
, Function
*OrigCallee
, Value
*Arg
,
1218 bool UseFloat
, Value
*&Sin
, Value
*&Cos
,
1221 Value
*LibCallSimplifier::optimizeSinCosPi(CallInst
*CI
, IRBuilder
<> &B
) {
1223 // Make sure the prototype is as expected, otherwise the rest of the
1224 // function is probably invalid and likely to abort.
1225 if (!isTrigLibCall(CI
))
1228 Value
*Arg
= CI
->getArgOperand(0);
1229 SmallVector
<CallInst
*, 1> SinCalls
;
1230 SmallVector
<CallInst
*, 1> CosCalls
;
1231 SmallVector
<CallInst
*, 1> SinCosCalls
;
1233 bool IsFloat
= Arg
->getType()->isFloatTy();
1235 // Look for all compatible sinpi, cospi and sincospi calls with the same
1236 // argument. If there are enough (in some sense) we can make the
1238 for (User
*U
: Arg
->users())
1239 classifyArgUse(U
, CI
->getParent(), IsFloat
, SinCalls
, CosCalls
,
1242 // It's only worthwhile if both sinpi and cospi are actually used.
1243 if (SinCosCalls
.empty() && (SinCalls
.empty() || CosCalls
.empty()))
1246 Value
*Sin
, *Cos
, *SinCos
;
1247 insertSinCosCall(B
, CI
->getCalledFunction(), Arg
, IsFloat
, Sin
, Cos
, SinCos
);
1249 replaceTrigInsts(SinCalls
, Sin
);
1250 replaceTrigInsts(CosCalls
, Cos
);
1251 replaceTrigInsts(SinCosCalls
, SinCos
);
1256 static bool isTrigLibCall(CallInst
*CI
) {
1257 Function
*Callee
= CI
->getCalledFunction();
1258 FunctionType
*FT
= Callee
->getFunctionType();
1260 // We can only hope to do anything useful if we can ignore things like errno
1261 // and floating-point exceptions.
1262 bool AttributesSafe
=
1263 CI
->hasFnAttr(Attribute::NoUnwind
) && CI
->hasFnAttr(Attribute::ReadNone
);
1265 // Other than that we need float(float) or double(double)
1266 return AttributesSafe
&& FT
->getNumParams() == 1 &&
1267 FT
->getReturnType() == FT
->getParamType(0) &&
1268 (FT
->getParamType(0)->isFloatTy() ||
1269 FT
->getParamType(0)->isDoubleTy());
1273 LibCallSimplifier::classifyArgUse(Value
*Val
, BasicBlock
*BB
, bool IsFloat
,
1274 SmallVectorImpl
<CallInst
*> &SinCalls
,
1275 SmallVectorImpl
<CallInst
*> &CosCalls
,
1276 SmallVectorImpl
<CallInst
*> &SinCosCalls
) {
1277 CallInst
*CI
= dyn_cast
<CallInst
>(Val
);
1282 Function
*Callee
= CI
->getCalledFunction();
1283 StringRef FuncName
= Callee
->getName();
1285 if (!TLI
->getLibFunc(FuncName
, Func
) || !TLI
->has(Func
) || !isTrigLibCall(CI
))
1289 if (Func
== LibFunc::sinpif
)
1290 SinCalls
.push_back(CI
);
1291 else if (Func
== LibFunc::cospif
)
1292 CosCalls
.push_back(CI
);
1293 else if (Func
== LibFunc::sincospif_stret
)
1294 SinCosCalls
.push_back(CI
);
1296 if (Func
== LibFunc::sinpi
)
1297 SinCalls
.push_back(CI
);
1298 else if (Func
== LibFunc::cospi
)
1299 CosCalls
.push_back(CI
);
1300 else if (Func
== LibFunc::sincospi_stret
)
1301 SinCosCalls
.push_back(CI
);
1305 void LibCallSimplifier::replaceTrigInsts(SmallVectorImpl
<CallInst
*> &Calls
,
1307 for (SmallVectorImpl
<CallInst
*>::iterator I
= Calls
.begin(), E
= Calls
.end();
1309 replaceAllUsesWith(*I
, Res
);
1313 void insertSinCosCall(IRBuilder
<> &B
, Function
*OrigCallee
, Value
*Arg
,
1314 bool UseFloat
, Value
*&Sin
, Value
*&Cos
, Value
*&SinCos
) {
1315 Type
*ArgTy
= Arg
->getType();
1319 Triple
T(OrigCallee
->getParent()->getTargetTriple());
1321 Name
= "__sincospif_stret";
1323 assert(T
.getArch() != Triple::x86
&& "x86 messy and unsupported for now");
1324 // x86_64 can't use {float, float} since that would be returned in both
1325 // xmm0 and xmm1, which isn't what a real struct would do.
1326 ResTy
= T
.getArch() == Triple::x86_64
1327 ? static_cast<Type
*>(VectorType::get(ArgTy
, 2))
1328 : static_cast<Type
*>(StructType::get(ArgTy
, ArgTy
, nullptr));
1330 Name
= "__sincospi_stret";
1331 ResTy
= StructType::get(ArgTy
, ArgTy
, nullptr);
1334 Module
*M
= OrigCallee
->getParent();
1335 Value
*Callee
= M
->getOrInsertFunction(Name
, OrigCallee
->getAttributes(),
1336 ResTy
, ArgTy
, nullptr);
1338 if (Instruction
*ArgInst
= dyn_cast
<Instruction
>(Arg
)) {
1339 // If the argument is an instruction, it must dominate all uses so put our
1340 // sincos call there.
1341 BasicBlock::iterator Loc
= ArgInst
;
1342 B
.SetInsertPoint(ArgInst
->getParent(), ++Loc
);
1344 // Otherwise (e.g. for a constant) the beginning of the function is as
1345 // good a place as any.
1346 BasicBlock
&EntryBB
= B
.GetInsertBlock()->getParent()->getEntryBlock();
1347 B
.SetInsertPoint(&EntryBB
, EntryBB
.begin());
1350 SinCos
= B
.CreateCall(Callee
, Arg
, "sincospi");
1352 if (SinCos
->getType()->isStructTy()) {
1353 Sin
= B
.CreateExtractValue(SinCos
, 0, "sinpi");
1354 Cos
= B
.CreateExtractValue(SinCos
, 1, "cospi");
1356 Sin
= B
.CreateExtractElement(SinCos
, ConstantInt::get(B
.getInt32Ty(), 0),
1358 Cos
= B
.CreateExtractElement(SinCos
, ConstantInt::get(B
.getInt32Ty(), 1),
1363 //===----------------------------------------------------------------------===//
1364 // Integer Library Call Optimizations
1365 //===----------------------------------------------------------------------===//
1367 Value
*LibCallSimplifier::optimizeFFS(CallInst
*CI
, IRBuilder
<> &B
) {
1368 Function
*Callee
= CI
->getCalledFunction();
1369 FunctionType
*FT
= Callee
->getFunctionType();
1370 // Just make sure this has 2 arguments of the same FP type, which match the
1372 if (FT
->getNumParams() != 1 || !FT
->getReturnType()->isIntegerTy(32) ||
1373 !FT
->getParamType(0)->isIntegerTy())
1376 Value
*Op
= CI
->getArgOperand(0);
1379 if (ConstantInt
*CI
= dyn_cast
<ConstantInt
>(Op
)) {
1380 if (CI
->isZero()) // ffs(0) -> 0.
1381 return B
.getInt32(0);
1382 // ffs(c) -> cttz(c)+1
1383 return B
.getInt32(CI
->getValue().countTrailingZeros() + 1);
1386 // ffs(x) -> x != 0 ? (i32)llvm.cttz(x)+1 : 0
1387 Type
*ArgType
= Op
->getType();
1389 Intrinsic::getDeclaration(Callee
->getParent(), Intrinsic::cttz
, ArgType
);
1390 Value
*V
= B
.CreateCall2(F
, Op
, B
.getFalse(), "cttz");
1391 V
= B
.CreateAdd(V
, ConstantInt::get(V
->getType(), 1));
1392 V
= B
.CreateIntCast(V
, B
.getInt32Ty(), false);
1394 Value
*Cond
= B
.CreateICmpNE(Op
, Constant::getNullValue(ArgType
));
1395 return B
.CreateSelect(Cond
, V
, B
.getInt32(0));
1398 Value
*LibCallSimplifier::optimizeAbs(CallInst
*CI
, IRBuilder
<> &B
) {
1399 Function
*Callee
= CI
->getCalledFunction();
1400 FunctionType
*FT
= Callee
->getFunctionType();
1401 // We require integer(integer) where the types agree.
1402 if (FT
->getNumParams() != 1 || !FT
->getReturnType()->isIntegerTy() ||
1403 FT
->getParamType(0) != FT
->getReturnType())
1406 // abs(x) -> x >s -1 ? x : -x
1407 Value
*Op
= CI
->getArgOperand(0);
1409 B
.CreateICmpSGT(Op
, Constant::getAllOnesValue(Op
->getType()), "ispos");
1410 Value
*Neg
= B
.CreateNeg(Op
, "neg");
1411 return B
.CreateSelect(Pos
, Op
, Neg
);
1414 Value
*LibCallSimplifier::optimizeIsDigit(CallInst
*CI
, IRBuilder
<> &B
) {
1415 Function
*Callee
= CI
->getCalledFunction();
1416 FunctionType
*FT
= Callee
->getFunctionType();
1417 // We require integer(i32)
1418 if (FT
->getNumParams() != 1 || !FT
->getReturnType()->isIntegerTy() ||
1419 !FT
->getParamType(0)->isIntegerTy(32))
1422 // isdigit(c) -> (c-'0') <u 10
1423 Value
*Op
= CI
->getArgOperand(0);
1424 Op
= B
.CreateSub(Op
, B
.getInt32('0'), "isdigittmp");
1425 Op
= B
.CreateICmpULT(Op
, B
.getInt32(10), "isdigit");
1426 return B
.CreateZExt(Op
, CI
->getType());
1429 Value
*LibCallSimplifier::optimizeIsAscii(CallInst
*CI
, IRBuilder
<> &B
) {
1430 Function
*Callee
= CI
->getCalledFunction();
1431 FunctionType
*FT
= Callee
->getFunctionType();
1432 // We require integer(i32)
1433 if (FT
->getNumParams() != 1 || !FT
->getReturnType()->isIntegerTy() ||
1434 !FT
->getParamType(0)->isIntegerTy(32))
1437 // isascii(c) -> c <u 128
1438 Value
*Op
= CI
->getArgOperand(0);
1439 Op
= B
.CreateICmpULT(Op
, B
.getInt32(128), "isascii");
1440 return B
.CreateZExt(Op
, CI
->getType());
1443 Value
*LibCallSimplifier::optimizeToAscii(CallInst
*CI
, IRBuilder
<> &B
) {
1444 Function
*Callee
= CI
->getCalledFunction();
1445 FunctionType
*FT
= Callee
->getFunctionType();
1446 // We require i32(i32)
1447 if (FT
->getNumParams() != 1 || FT
->getReturnType() != FT
->getParamType(0) ||
1448 !FT
->getParamType(0)->isIntegerTy(32))
1451 // toascii(c) -> c & 0x7f
1452 return B
.CreateAnd(CI
->getArgOperand(0),
1453 ConstantInt::get(CI
->getType(), 0x7F));
1456 //===----------------------------------------------------------------------===//
1457 // Formatting and IO Library Call Optimizations
1458 //===----------------------------------------------------------------------===//
1460 static bool isReportingError(Function
*Callee
, CallInst
*CI
, int StreamArg
);
1462 Value
*LibCallSimplifier::optimizeErrorReporting(CallInst
*CI
, IRBuilder
<> &B
,
1464 // Error reporting calls should be cold, mark them as such.
1465 // This applies even to non-builtin calls: it is only a hint and applies to
1466 // functions that the frontend might not understand as builtins.
1468 // This heuristic was suggested in:
1469 // Improving Static Branch Prediction in a Compiler
1470 // Brian L. Deitrich, Ben-Chung Cheng, Wen-mei W. Hwu
1471 // Proceedings of PACT'98, Oct. 1998, IEEE
1472 Function
*Callee
= CI
->getCalledFunction();
1474 if (!CI
->hasFnAttr(Attribute::Cold
) &&
1475 isReportingError(Callee
, CI
, StreamArg
)) {
1476 CI
->addAttribute(AttributeSet::FunctionIndex
, Attribute::Cold
);
1482 static bool isReportingError(Function
*Callee
, CallInst
*CI
, int StreamArg
) {
1483 if (!ColdErrorCalls
)
1486 if (!Callee
|| !Callee
->isDeclaration())
1492 // These functions might be considered cold, but only if their stream
1493 // argument is stderr.
1495 if (StreamArg
>= (int)CI
->getNumArgOperands())
1497 LoadInst
*LI
= dyn_cast
<LoadInst
>(CI
->getArgOperand(StreamArg
));
1500 GlobalVariable
*GV
= dyn_cast
<GlobalVariable
>(LI
->getPointerOperand());
1501 if (!GV
|| !GV
->isDeclaration())
1503 return GV
->getName() == "stderr";
1506 Value
*LibCallSimplifier::optimizePrintFString(CallInst
*CI
, IRBuilder
<> &B
) {
1507 // Check for a fixed format string.
1508 StringRef FormatStr
;
1509 if (!getConstantStringInfo(CI
->getArgOperand(0), FormatStr
))
1512 // Empty format string -> noop.
1513 if (FormatStr
.empty()) // Tolerate printf's declared void.
1514 return CI
->use_empty() ? (Value
*)CI
: ConstantInt::get(CI
->getType(), 0);
1516 // Do not do any of the following transformations if the printf return value
1517 // is used, in general the printf return value is not compatible with either
1518 // putchar() or puts().
1519 if (!CI
->use_empty())
1522 // printf("x") -> putchar('x'), even for '%'.
1523 if (FormatStr
.size() == 1) {
1524 Value
*Res
= EmitPutChar(B
.getInt32(FormatStr
[0]), B
, DL
, TLI
);
1525 if (CI
->use_empty() || !Res
)
1527 return B
.CreateIntCast(Res
, CI
->getType(), true);
1530 // printf("foo\n") --> puts("foo")
1531 if (FormatStr
[FormatStr
.size() - 1] == '\n' &&
1532 FormatStr
.find('%') == StringRef::npos
) { // No format characters.
1533 // Create a string literal with no \n on it. We expect the constant merge
1534 // pass to be run after this pass, to merge duplicate strings.
1535 FormatStr
= FormatStr
.drop_back();
1536 Value
*GV
= B
.CreateGlobalString(FormatStr
, "str");
1537 Value
*NewCI
= EmitPutS(GV
, B
, DL
, TLI
);
1538 return (CI
->use_empty() || !NewCI
)
1540 : ConstantInt::get(CI
->getType(), FormatStr
.size() + 1);
1543 // Optimize specific format strings.
1544 // printf("%c", chr) --> putchar(chr)
1545 if (FormatStr
== "%c" && CI
->getNumArgOperands() > 1 &&
1546 CI
->getArgOperand(1)->getType()->isIntegerTy()) {
1547 Value
*Res
= EmitPutChar(CI
->getArgOperand(1), B
, DL
, TLI
);
1549 if (CI
->use_empty() || !Res
)
1551 return B
.CreateIntCast(Res
, CI
->getType(), true);
1554 // printf("%s\n", str) --> puts(str)
1555 if (FormatStr
== "%s\n" && CI
->getNumArgOperands() > 1 &&
1556 CI
->getArgOperand(1)->getType()->isPointerTy()) {
1557 return EmitPutS(CI
->getArgOperand(1), B
, DL
, TLI
);
1562 Value
*LibCallSimplifier::optimizePrintF(CallInst
*CI
, IRBuilder
<> &B
) {
1564 Function
*Callee
= CI
->getCalledFunction();
1565 // Require one fixed pointer argument and an integer/void result.
1566 FunctionType
*FT
= Callee
->getFunctionType();
1567 if (FT
->getNumParams() < 1 || !FT
->getParamType(0)->isPointerTy() ||
1568 !(FT
->getReturnType()->isIntegerTy() || FT
->getReturnType()->isVoidTy()))
1571 if (Value
*V
= optimizePrintFString(CI
, B
)) {
1575 // printf(format, ...) -> iprintf(format, ...) if no floating point
1577 if (TLI
->has(LibFunc::iprintf
) && !callHasFloatingPointArgument(CI
)) {
1578 Module
*M
= B
.GetInsertBlock()->getParent()->getParent();
1579 Constant
*IPrintFFn
=
1580 M
->getOrInsertFunction("iprintf", FT
, Callee
->getAttributes());
1581 CallInst
*New
= cast
<CallInst
>(CI
->clone());
1582 New
->setCalledFunction(IPrintFFn
);
1589 Value
*LibCallSimplifier::optimizeSPrintFString(CallInst
*CI
, IRBuilder
<> &B
) {
1590 // Check for a fixed format string.
1591 StringRef FormatStr
;
1592 if (!getConstantStringInfo(CI
->getArgOperand(1), FormatStr
))
1595 // If we just have a format string (nothing else crazy) transform it.
1596 if (CI
->getNumArgOperands() == 2) {
1597 // Make sure there's no % in the constant array. We could try to handle
1598 // %% -> % in the future if we cared.
1599 for (unsigned i
= 0, e
= FormatStr
.size(); i
!= e
; ++i
)
1600 if (FormatStr
[i
] == '%')
1601 return nullptr; // we found a format specifier, bail out.
1603 // These optimizations require DataLayout.
1607 // sprintf(str, fmt) -> llvm.memcpy(str, fmt, strlen(fmt)+1, 1)
1609 CI
->getArgOperand(0), CI
->getArgOperand(1),
1610 ConstantInt::get(DL
->getIntPtrType(CI
->getContext()),
1611 FormatStr
.size() + 1),
1612 1); // Copy the null byte.
1613 return ConstantInt::get(CI
->getType(), FormatStr
.size());
1616 // The remaining optimizations require the format string to be "%s" or "%c"
1617 // and have an extra operand.
1618 if (FormatStr
.size() != 2 || FormatStr
[0] != '%' ||
1619 CI
->getNumArgOperands() < 3)
1622 // Decode the second character of the format string.
1623 if (FormatStr
[1] == 'c') {
1624 // sprintf(dst, "%c", chr) --> *(i8*)dst = chr; *((i8*)dst+1) = 0
1625 if (!CI
->getArgOperand(2)->getType()->isIntegerTy())
1627 Value
*V
= B
.CreateTrunc(CI
->getArgOperand(2), B
.getInt8Ty(), "char");
1628 Value
*Ptr
= CastToCStr(CI
->getArgOperand(0), B
);
1629 B
.CreateStore(V
, Ptr
);
1630 Ptr
= B
.CreateGEP(Ptr
, B
.getInt32(1), "nul");
1631 B
.CreateStore(B
.getInt8(0), Ptr
);
1633 return ConstantInt::get(CI
->getType(), 1);
1636 if (FormatStr
[1] == 's') {
1637 // These optimizations require DataLayout.
1641 // sprintf(dest, "%s", str) -> llvm.memcpy(dest, str, strlen(str)+1, 1)
1642 if (!CI
->getArgOperand(2)->getType()->isPointerTy())
1645 Value
*Len
= EmitStrLen(CI
->getArgOperand(2), B
, DL
, TLI
);
1649 B
.CreateAdd(Len
, ConstantInt::get(Len
->getType(), 1), "leninc");
1650 B
.CreateMemCpy(CI
->getArgOperand(0), CI
->getArgOperand(2), IncLen
, 1);
1652 // The sprintf result is the unincremented number of bytes in the string.
1653 return B
.CreateIntCast(Len
, CI
->getType(), false);
1658 Value
*LibCallSimplifier::optimizeSPrintF(CallInst
*CI
, IRBuilder
<> &B
) {
1659 Function
*Callee
= CI
->getCalledFunction();
1660 // Require two fixed pointer arguments and an integer result.
1661 FunctionType
*FT
= Callee
->getFunctionType();
1662 if (FT
->getNumParams() != 2 || !FT
->getParamType(0)->isPointerTy() ||
1663 !FT
->getParamType(1)->isPointerTy() ||
1664 !FT
->getReturnType()->isIntegerTy())
1667 if (Value
*V
= optimizeSPrintFString(CI
, B
)) {
1671 // sprintf(str, format, ...) -> siprintf(str, format, ...) if no floating
1673 if (TLI
->has(LibFunc::siprintf
) && !callHasFloatingPointArgument(CI
)) {
1674 Module
*M
= B
.GetInsertBlock()->getParent()->getParent();
1675 Constant
*SIPrintFFn
=
1676 M
->getOrInsertFunction("siprintf", FT
, Callee
->getAttributes());
1677 CallInst
*New
= cast
<CallInst
>(CI
->clone());
1678 New
->setCalledFunction(SIPrintFFn
);
1685 Value
*LibCallSimplifier::optimizeFPrintFString(CallInst
*CI
, IRBuilder
<> &B
) {
1686 optimizeErrorReporting(CI
, B
, 0);
1688 // All the optimizations depend on the format string.
1689 StringRef FormatStr
;
1690 if (!getConstantStringInfo(CI
->getArgOperand(1), FormatStr
))
1693 // Do not do any of the following transformations if the fprintf return
1694 // value is used, in general the fprintf return value is not compatible
1695 // with fwrite(), fputc() or fputs().
1696 if (!CI
->use_empty())
1699 // fprintf(F, "foo") --> fwrite("foo", 3, 1, F)
1700 if (CI
->getNumArgOperands() == 2) {
1701 for (unsigned i
= 0, e
= FormatStr
.size(); i
!= e
; ++i
)
1702 if (FormatStr
[i
] == '%') // Could handle %% -> % if we cared.
1703 return nullptr; // We found a format specifier.
1705 // These optimizations require DataLayout.
1710 CI
->getArgOperand(1),
1711 ConstantInt::get(DL
->getIntPtrType(CI
->getContext()), FormatStr
.size()),
1712 CI
->getArgOperand(0), B
, DL
, TLI
);
1715 // The remaining optimizations require the format string to be "%s" or "%c"
1716 // and have an extra operand.
1717 if (FormatStr
.size() != 2 || FormatStr
[0] != '%' ||
1718 CI
->getNumArgOperands() < 3)
1721 // Decode the second character of the format string.
1722 if (FormatStr
[1] == 'c') {
1723 // fprintf(F, "%c", chr) --> fputc(chr, F)
1724 if (!CI
->getArgOperand(2)->getType()->isIntegerTy())
1726 return EmitFPutC(CI
->getArgOperand(2), CI
->getArgOperand(0), B
, DL
, TLI
);
1729 if (FormatStr
[1] == 's') {
1730 // fprintf(F, "%s", str) --> fputs(str, F)
1731 if (!CI
->getArgOperand(2)->getType()->isPointerTy())
1733 return EmitFPutS(CI
->getArgOperand(2), CI
->getArgOperand(0), B
, DL
, TLI
);
1738 Value
*LibCallSimplifier::optimizeFPrintF(CallInst
*CI
, IRBuilder
<> &B
) {
1739 Function
*Callee
= CI
->getCalledFunction();
1740 // Require two fixed paramters as pointers and integer result.
1741 FunctionType
*FT
= Callee
->getFunctionType();
1742 if (FT
->getNumParams() != 2 || !FT
->getParamType(0)->isPointerTy() ||
1743 !FT
->getParamType(1)->isPointerTy() ||
1744 !FT
->getReturnType()->isIntegerTy())
1747 if (Value
*V
= optimizeFPrintFString(CI
, B
)) {
1751 // fprintf(stream, format, ...) -> fiprintf(stream, format, ...) if no
1752 // floating point arguments.
1753 if (TLI
->has(LibFunc::fiprintf
) && !callHasFloatingPointArgument(CI
)) {
1754 Module
*M
= B
.GetInsertBlock()->getParent()->getParent();
1755 Constant
*FIPrintFFn
=
1756 M
->getOrInsertFunction("fiprintf", FT
, Callee
->getAttributes());
1757 CallInst
*New
= cast
<CallInst
>(CI
->clone());
1758 New
->setCalledFunction(FIPrintFFn
);
1765 Value
*LibCallSimplifier::optimizeFWrite(CallInst
*CI
, IRBuilder
<> &B
) {
1766 optimizeErrorReporting(CI
, B
, 3);
1768 Function
*Callee
= CI
->getCalledFunction();
1769 // Require a pointer, an integer, an integer, a pointer, returning integer.
1770 FunctionType
*FT
= Callee
->getFunctionType();
1771 if (FT
->getNumParams() != 4 || !FT
->getParamType(0)->isPointerTy() ||
1772 !FT
->getParamType(1)->isIntegerTy() ||
1773 !FT
->getParamType(2)->isIntegerTy() ||
1774 !FT
->getParamType(3)->isPointerTy() ||
1775 !FT
->getReturnType()->isIntegerTy())
1778 // Get the element size and count.
1779 ConstantInt
*SizeC
= dyn_cast
<ConstantInt
>(CI
->getArgOperand(1));
1780 ConstantInt
*CountC
= dyn_cast
<ConstantInt
>(CI
->getArgOperand(2));
1781 if (!SizeC
|| !CountC
)
1783 uint64_t Bytes
= SizeC
->getZExtValue() * CountC
->getZExtValue();
1785 // If this is writing zero records, remove the call (it's a noop).
1787 return ConstantInt::get(CI
->getType(), 0);
1789 // If this is writing one byte, turn it into fputc.
1790 // This optimisation is only valid, if the return value is unused.
1791 if (Bytes
== 1 && CI
->use_empty()) { // fwrite(S,1,1,F) -> fputc(S[0],F)
1792 Value
*Char
= B
.CreateLoad(CastToCStr(CI
->getArgOperand(0), B
), "char");
1793 Value
*NewCI
= EmitFPutC(Char
, CI
->getArgOperand(3), B
, DL
, TLI
);
1794 return NewCI
? ConstantInt::get(CI
->getType(), 1) : nullptr;
1800 Value
*LibCallSimplifier::optimizeFPuts(CallInst
*CI
, IRBuilder
<> &B
) {
1801 optimizeErrorReporting(CI
, B
, 1);
1803 Function
*Callee
= CI
->getCalledFunction();
1805 // These optimizations require DataLayout.
1809 // Require two pointers. Also, we can't optimize if return value is used.
1810 FunctionType
*FT
= Callee
->getFunctionType();
1811 if (FT
->getNumParams() != 2 || !FT
->getParamType(0)->isPointerTy() ||
1812 !FT
->getParamType(1)->isPointerTy() || !CI
->use_empty())
1815 // fputs(s,F) --> fwrite(s,1,strlen(s),F)
1816 uint64_t Len
= GetStringLength(CI
->getArgOperand(0));
1820 // Known to have no uses (see above).
1822 CI
->getArgOperand(0),
1823 ConstantInt::get(DL
->getIntPtrType(CI
->getContext()), Len
- 1),
1824 CI
->getArgOperand(1), B
, DL
, TLI
);
1827 Value
*LibCallSimplifier::optimizePuts(CallInst
*CI
, IRBuilder
<> &B
) {
1828 Function
*Callee
= CI
->getCalledFunction();
1829 // Require one fixed pointer argument and an integer/void result.
1830 FunctionType
*FT
= Callee
->getFunctionType();
1831 if (FT
->getNumParams() < 1 || !FT
->getParamType(0)->isPointerTy() ||
1832 !(FT
->getReturnType()->isIntegerTy() || FT
->getReturnType()->isVoidTy()))
1835 // Check for a constant string.
1837 if (!getConstantStringInfo(CI
->getArgOperand(0), Str
))
1840 if (Str
.empty() && CI
->use_empty()) {
1841 // puts("") -> putchar('\n')
1842 Value
*Res
= EmitPutChar(B
.getInt32('\n'), B
, DL
, TLI
);
1843 if (CI
->use_empty() || !Res
)
1845 return B
.CreateIntCast(Res
, CI
->getType(), true);
1851 bool LibCallSimplifier::hasFloatVersion(StringRef FuncName
) {
1853 SmallString
<20> FloatFuncName
= FuncName
;
1854 FloatFuncName
+= 'f';
1855 if (TLI
->getLibFunc(FloatFuncName
, Func
))
1856 return TLI
->has(Func
);
1860 Value
*LibCallSimplifier::optimizeStringMemoryLibCall(CallInst
*CI
,
1861 IRBuilder
<> &Builder
) {
1863 Function
*Callee
= CI
->getCalledFunction();
1864 StringRef FuncName
= Callee
->getName();
1866 // Check for string/memory library functions.
1867 if (TLI
->getLibFunc(FuncName
, Func
) && TLI
->has(Func
)) {
1868 // Make sure we never change the calling convention.
1869 assert((ignoreCallingConv(Func
) ||
1870 CI
->getCallingConv() == llvm::CallingConv::C
) &&
1871 "Optimizing string/memory libcall would change the calling convention");
1873 case LibFunc::strcat
:
1874 return optimizeStrCat(CI
, Builder
);
1875 case LibFunc::strncat
:
1876 return optimizeStrNCat(CI
, Builder
);
1877 case LibFunc::strchr
:
1878 return optimizeStrChr(CI
, Builder
);
1879 case LibFunc::strrchr
:
1880 return optimizeStrRChr(CI
, Builder
);
1881 case LibFunc::strcmp
:
1882 return optimizeStrCmp(CI
, Builder
);
1883 case LibFunc::strncmp
:
1884 return optimizeStrNCmp(CI
, Builder
);
1885 case LibFunc::strcpy
:
1886 return optimizeStrCpy(CI
, Builder
);
1887 case LibFunc::stpcpy
:
1888 return optimizeStpCpy(CI
, Builder
);
1889 case LibFunc::strncpy
:
1890 return optimizeStrNCpy(CI
, Builder
);
1891 case LibFunc::strlen
:
1892 return optimizeStrLen(CI
, Builder
);
1893 case LibFunc::strpbrk
:
1894 return optimizeStrPBrk(CI
, Builder
);
1895 case LibFunc::strtol
:
1896 case LibFunc::strtod
:
1897 case LibFunc::strtof
:
1898 case LibFunc::strtoul
:
1899 case LibFunc::strtoll
:
1900 case LibFunc::strtold
:
1901 case LibFunc::strtoull
:
1902 return optimizeStrTo(CI
, Builder
);
1903 case LibFunc::strspn
:
1904 return optimizeStrSpn(CI
, Builder
);
1905 case LibFunc::strcspn
:
1906 return optimizeStrCSpn(CI
, Builder
);
1907 case LibFunc::strstr
:
1908 return optimizeStrStr(CI
, Builder
);
1909 case LibFunc::memcmp
:
1910 return optimizeMemCmp(CI
, Builder
);
1911 case LibFunc::memcpy
:
1912 return optimizeMemCpy(CI
, Builder
);
1913 case LibFunc::memmove
:
1914 return optimizeMemMove(CI
, Builder
);
1915 case LibFunc::memset
:
1916 return optimizeMemSet(CI
, Builder
);
1924 Value
*LibCallSimplifier::optimizeCall(CallInst
*CI
) {
1925 if (CI
->isNoBuiltin())
1929 Function
*Callee
= CI
->getCalledFunction();
1930 StringRef FuncName
= Callee
->getName();
1931 IRBuilder
<> Builder(CI
);
1932 bool isCallingConvC
= CI
->getCallingConv() == llvm::CallingConv::C
;
1934 // Command-line parameter overrides function attribute.
1935 if (EnableUnsafeFPShrink
.getNumOccurrences() > 0)
1936 UnsafeFPShrink
= EnableUnsafeFPShrink
;
1937 else if (Callee
->hasFnAttribute("unsafe-fp-math")) {
1938 // FIXME: This is the same problem as described in optimizeSqrt().
1939 // If calls gain access to IR-level FMF, then use that instead of a
1940 // function attribute.
1942 // Check for unsafe-fp-math = true.
1943 Attribute Attr
= Callee
->getFnAttribute("unsafe-fp-math");
1944 if (Attr
.getValueAsString() == "true")
1945 UnsafeFPShrink
= true;
1948 // First, check for intrinsics.
1949 if (IntrinsicInst
*II
= dyn_cast
<IntrinsicInst
>(CI
)) {
1950 if (!isCallingConvC
)
1952 switch (II
->getIntrinsicID()) {
1953 case Intrinsic::pow
:
1954 return optimizePow(CI
, Builder
);
1955 case Intrinsic::exp2
:
1956 return optimizeExp2(CI
, Builder
);
1957 case Intrinsic::fabs
:
1958 return optimizeFabs(CI
, Builder
);
1959 case Intrinsic::sqrt
:
1960 return optimizeSqrt(CI
, Builder
);
1966 // Also try to simplify calls to fortified library functions.
1967 if (Value
*SimplifiedFortifiedCI
= FortifiedSimplifier
.optimizeCall(CI
)) {
1968 // Try to further simplify the result.
1969 CallInst
*SimplifiedCI
= dyn_cast
<CallInst
>(SimplifiedFortifiedCI
);
1970 if (SimplifiedCI
&& SimplifiedCI
->getCalledFunction())
1971 if (Value
*V
= optimizeStringMemoryLibCall(SimplifiedCI
, Builder
)) {
1972 // If we were able to further simplify, remove the now redundant call.
1973 SimplifiedCI
->replaceAllUsesWith(V
);
1974 SimplifiedCI
->eraseFromParent();
1977 return SimplifiedFortifiedCI
;
1980 // Then check for known library functions.
1981 if (TLI
->getLibFunc(FuncName
, Func
) && TLI
->has(Func
)) {
1982 // We never change the calling convention.
1983 if (!ignoreCallingConv(Func
) && !isCallingConvC
)
1985 if (Value
*V
= optimizeStringMemoryLibCall(CI
, Builder
))
1991 return optimizeCos(CI
, Builder
);
1992 case LibFunc::sinpif
:
1993 case LibFunc::sinpi
:
1994 case LibFunc::cospif
:
1995 case LibFunc::cospi
:
1996 return optimizeSinCosPi(CI
, Builder
);
2000 return optimizePow(CI
, Builder
);
2001 case LibFunc::exp2l
:
2003 case LibFunc::exp2f
:
2004 return optimizeExp2(CI
, Builder
);
2005 case LibFunc::fabsf
:
2007 case LibFunc::fabsl
:
2008 return optimizeFabs(CI
, Builder
);
2009 case LibFunc::sqrtf
:
2011 case LibFunc::sqrtl
:
2012 return optimizeSqrt(CI
, Builder
);
2015 case LibFunc::ffsll
:
2016 return optimizeFFS(CI
, Builder
);
2019 case LibFunc::llabs
:
2020 return optimizeAbs(CI
, Builder
);
2021 case LibFunc::isdigit
:
2022 return optimizeIsDigit(CI
, Builder
);
2023 case LibFunc::isascii
:
2024 return optimizeIsAscii(CI
, Builder
);
2025 case LibFunc::toascii
:
2026 return optimizeToAscii(CI
, Builder
);
2027 case LibFunc::printf
:
2028 return optimizePrintF(CI
, Builder
);
2029 case LibFunc::sprintf
:
2030 return optimizeSPrintF(CI
, Builder
);
2031 case LibFunc::fprintf
:
2032 return optimizeFPrintF(CI
, Builder
);
2033 case LibFunc::fwrite
:
2034 return optimizeFWrite(CI
, Builder
);
2035 case LibFunc::fputs
:
2036 return optimizeFPuts(CI
, Builder
);
2038 return optimizePuts(CI
, Builder
);
2039 case LibFunc::perror
:
2040 return optimizeErrorReporting(CI
, Builder
);
2041 case LibFunc::vfprintf
:
2042 case LibFunc::fiprintf
:
2043 return optimizeErrorReporting(CI
, Builder
, 0);
2044 case LibFunc::fputc
:
2045 return optimizeErrorReporting(CI
, Builder
, 1);
2047 case LibFunc::floor
:
2049 case LibFunc::round
:
2050 case LibFunc::nearbyint
:
2051 case LibFunc::trunc
:
2052 if (hasFloatVersion(FuncName
))
2053 return optimizeUnaryDoubleFP(CI
, Builder
, false);
2056 case LibFunc::acosh
:
2058 case LibFunc::asinh
:
2060 case LibFunc::atanh
:
2064 case LibFunc::exp10
:
2065 case LibFunc::expm1
:
2067 case LibFunc::log10
:
2068 case LibFunc::log1p
:
2075 if (UnsafeFPShrink
&& hasFloatVersion(FuncName
))
2076 return optimizeUnaryDoubleFP(CI
, Builder
, true);
2078 case LibFunc::copysign
:
2081 if (hasFloatVersion(FuncName
))
2082 return optimizeBinaryDoubleFP(CI
, Builder
);
2091 LibCallSimplifier::LibCallSimplifier(const DataLayout
*DL
,
2092 const TargetLibraryInfo
*TLI
) :
2093 FortifiedSimplifier(DL
, TLI
),
2096 UnsafeFPShrink(false) {
2099 void LibCallSimplifier::replaceAllUsesWith(Instruction
*I
, Value
*With
) const {
2100 I
->replaceAllUsesWith(With
);
2101 I
->eraseFromParent();
2105 // Additional cases that we need to add to this file:
2108 // * cbrt(expN(X)) -> expN(x/3)
2109 // * cbrt(sqrt(x)) -> pow(x,1/6)
2110 // * cbrt(sqrt(x)) -> pow(x,1/9)
2113 // * exp(log(x)) -> x
2116 // * log(exp(x)) -> x
2117 // * log(x**y) -> y*log(x)
2118 // * log(exp(y)) -> y*log(e)
2119 // * log(exp2(y)) -> y*log(2)
2120 // * log(exp10(y)) -> y*log(10)
2121 // * log(sqrt(x)) -> 0.5*log(x)
2122 // * log(pow(x,y)) -> y*log(x)
2124 // lround, lroundf, lroundl:
2125 // * lround(cnst) -> cnst'
2128 // * pow(exp(x),y) -> exp(x*y)
2129 // * pow(sqrt(x),y) -> pow(x,y*0.5)
2130 // * pow(pow(x,y),z)-> pow(x,y*z)
2132 // round, roundf, roundl:
2133 // * round(cnst) -> cnst'
2136 // * signbit(cnst) -> cnst'
2137 // * signbit(nncst) -> 0 (if pstv is a non-negative constant)
2139 // sqrt, sqrtf, sqrtl:
2140 // * sqrt(expN(x)) -> expN(x*0.5)
2141 // * sqrt(Nroot(x)) -> pow(x,1/(2*N))
2142 // * sqrt(pow(x,y)) -> pow(|x|,y*0.5)
2145 // * tan(atan(x)) -> x
2147 // trunc, truncf, truncl:
2148 // * trunc(cnst) -> cnst'
2152 //===----------------------------------------------------------------------===//
2153 // Fortified Library Call Optimizations
2154 //===----------------------------------------------------------------------===//
2156 bool FortifiedLibCallSimplifier::isFortifiedCallFoldable(CallInst
*CI
,
2160 if (CI
->getArgOperand(ObjSizeOp
) == CI
->getArgOperand(SizeOp
))
2162 if (ConstantInt
*ObjSizeCI
=
2163 dyn_cast
<ConstantInt
>(CI
->getArgOperand(ObjSizeOp
))) {
2164 if (ObjSizeCI
->isAllOnesValue())
2166 // If the object size wasn't -1 (unknown), bail out if we were asked to.
2167 if (OnlyLowerUnknownSize
)
2170 uint64_t Len
= GetStringLength(CI
->getArgOperand(SizeOp
));
2171 // If the length is 0 we don't know how long it is and so we can't
2172 // remove the check.
2175 return ObjSizeCI
->getZExtValue() >= Len
;
2177 if (ConstantInt
*SizeCI
= dyn_cast
<ConstantInt
>(CI
->getArgOperand(SizeOp
)))
2178 return ObjSizeCI
->getZExtValue() >= SizeCI
->getZExtValue();
2183 Value
*FortifiedLibCallSimplifier::optimizeMemCpyChk(CallInst
*CI
, IRBuilder
<> &B
) {
2184 Function
*Callee
= CI
->getCalledFunction();
2186 if (!checkStringCopyLibFuncSignature(Callee
, LibFunc::memcpy_chk
, DL
))
2189 if (isFortifiedCallFoldable(CI
, 3, 2, false)) {
2190 B
.CreateMemCpy(CI
->getArgOperand(0), CI
->getArgOperand(1),
2191 CI
->getArgOperand(2), 1);
2192 return CI
->getArgOperand(0);
2197 Value
*FortifiedLibCallSimplifier::optimizeMemMoveChk(CallInst
*CI
, IRBuilder
<> &B
) {
2198 Function
*Callee
= CI
->getCalledFunction();
2200 if (!checkStringCopyLibFuncSignature(Callee
, LibFunc::memmove_chk
, DL
))
2203 if (isFortifiedCallFoldable(CI
, 3, 2, false)) {
2204 B
.CreateMemMove(CI
->getArgOperand(0), CI
->getArgOperand(1),
2205 CI
->getArgOperand(2), 1);
2206 return CI
->getArgOperand(0);
2211 Value
*FortifiedLibCallSimplifier::optimizeMemSetChk(CallInst
*CI
, IRBuilder
<> &B
) {
2212 Function
*Callee
= CI
->getCalledFunction();
2214 if (!checkStringCopyLibFuncSignature(Callee
, LibFunc::memset_chk
, DL
))
2217 if (isFortifiedCallFoldable(CI
, 3, 2, false)) {
2218 Value
*Val
= B
.CreateIntCast(CI
->getArgOperand(1), B
.getInt8Ty(), false);
2219 B
.CreateMemSet(CI
->getArgOperand(0), Val
, CI
->getArgOperand(2), 1);
2220 return CI
->getArgOperand(0);
2225 Value
*FortifiedLibCallSimplifier::optimizeStrpCpyChk(CallInst
*CI
,
2227 LibFunc::Func Func
) {
2228 Function
*Callee
= CI
->getCalledFunction();
2229 StringRef Name
= Callee
->getName();
2231 if (!checkStringCopyLibFuncSignature(Callee
, Func
, DL
))
2234 Value
*Dst
= CI
->getArgOperand(0), *Src
= CI
->getArgOperand(1),
2235 *ObjSize
= CI
->getArgOperand(2);
2237 // __stpcpy_chk(x,x,...) -> x+strlen(x)
2238 if (Func
== LibFunc::stpcpy_chk
&& !OnlyLowerUnknownSize
&& Dst
== Src
) {
2239 Value
*StrLen
= EmitStrLen(Src
, B
, DL
, TLI
);
2240 return StrLen
? B
.CreateInBoundsGEP(Dst
, StrLen
) : nullptr;
2243 // If a) we don't have any length information, or b) we know this will
2244 // fit then just lower to a plain st[rp]cpy. Otherwise we'll keep our
2245 // st[rp]cpy_chk call which may fail at runtime if the size is too long.
2246 // TODO: It might be nice to get a maximum length out of the possible
2247 // string lengths for varying.
2248 if (isFortifiedCallFoldable(CI
, 2, 1, true)) {
2249 Value
*Ret
= EmitStrCpy(Dst
, Src
, B
, DL
, TLI
, Name
.substr(2, 6));
2251 } else if (!OnlyLowerUnknownSize
) {
2252 // Maybe we can stil fold __st[rp]cpy_chk to __memcpy_chk.
2253 uint64_t Len
= GetStringLength(Src
);
2257 // This optimization requires DataLayout.
2261 Type
*SizeTTy
= DL
->getIntPtrType(CI
->getContext());
2262 Value
*LenV
= ConstantInt::get(SizeTTy
, Len
);
2263 Value
*Ret
= EmitMemCpyChk(Dst
, Src
, LenV
, ObjSize
, B
, DL
, TLI
);
2264 // If the function was an __stpcpy_chk, and we were able to fold it into
2265 // a __memcpy_chk, we still need to return the correct end pointer.
2266 if (Ret
&& Func
== LibFunc::stpcpy_chk
)
2267 return B
.CreateGEP(Dst
, ConstantInt::get(SizeTTy
, Len
- 1));
2273 Value
*FortifiedLibCallSimplifier::optimizeStrpNCpyChk(CallInst
*CI
,
2275 LibFunc::Func Func
) {
2276 Function
*Callee
= CI
->getCalledFunction();
2277 StringRef Name
= Callee
->getName();
2279 if (!checkStringCopyLibFuncSignature(Callee
, Func
, DL
))
2281 if (isFortifiedCallFoldable(CI
, 3, 2, false)) {
2283 EmitStrNCpy(CI
->getArgOperand(0), CI
->getArgOperand(1),
2284 CI
->getArgOperand(2), B
, DL
, TLI
, Name
.substr(2, 7));
2290 Value
*FortifiedLibCallSimplifier::optimizeCall(CallInst
*CI
) {
2291 if (CI
->isNoBuiltin())
2295 Function
*Callee
= CI
->getCalledFunction();
2296 StringRef FuncName
= Callee
->getName();
2297 IRBuilder
<> Builder(CI
);
2298 bool isCallingConvC
= CI
->getCallingConv() == llvm::CallingConv::C
;
2300 // First, check that this is a known library functions.
2301 if (!TLI
->getLibFunc(FuncName
, Func
) || !TLI
->has(Func
))
2304 // We never change the calling convention.
2305 if (!ignoreCallingConv(Func
) && !isCallingConvC
)
2309 case LibFunc::memcpy_chk
:
2310 return optimizeMemCpyChk(CI
, Builder
);
2311 case LibFunc::memmove_chk
:
2312 return optimizeMemMoveChk(CI
, Builder
);
2313 case LibFunc::memset_chk
:
2314 return optimizeMemSetChk(CI
, Builder
);
2315 case LibFunc::stpcpy_chk
:
2316 case LibFunc::strcpy_chk
:
2317 return optimizeStrpCpyChk(CI
, Builder
, Func
);
2318 case LibFunc::stpncpy_chk
:
2319 case LibFunc::strncpy_chk
:
2320 return optimizeStrpNCpyChk(CI
, Builder
, Func
);
2327 FortifiedLibCallSimplifier::
2328 FortifiedLibCallSimplifier(const DataLayout
*DL
, const TargetLibraryInfo
*TLI
,
2329 bool OnlyLowerUnknownSize
)
2330 : DL(DL
), TLI(TLI
), OnlyLowerUnknownSize(OnlyLowerUnknownSize
) {