1 //===------ SimplifyLibCalls.cpp - Library calls simplifier ---------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements the library calls simplifier. It does not implement
10 // any pass, but can't be used by other passes to do simplifications.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #include "llvm/Transforms/Utils/SimplifyLibCalls.h"
15 #include "llvm/ADT/APSInt.h"
16 #include "llvm/ADT/SmallString.h"
17 #include "llvm/ADT/StringMap.h"
18 #include "llvm/ADT/Triple.h"
19 #include "llvm/Analysis/BlockFrequencyInfo.h"
20 #include "llvm/Analysis/ConstantFolding.h"
21 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
22 #include "llvm/Analysis/ProfileSummaryInfo.h"
23 #include "llvm/Transforms/Utils/Local.h"
24 #include "llvm/Analysis/ValueTracking.h"
25 #include "llvm/Analysis/CaptureTracking.h"
26 #include "llvm/Analysis/Loads.h"
27 #include "llvm/IR/DataLayout.h"
28 #include "llvm/IR/Function.h"
29 #include "llvm/IR/IRBuilder.h"
30 #include "llvm/IR/IntrinsicInst.h"
31 #include "llvm/IR/Intrinsics.h"
32 #include "llvm/IR/LLVMContext.h"
33 #include "llvm/IR/Module.h"
34 #include "llvm/IR/PatternMatch.h"
35 #include "llvm/Support/CommandLine.h"
36 #include "llvm/Support/KnownBits.h"
37 #include "llvm/Support/MathExtras.h"
38 #include "llvm/Transforms/Utils/BuildLibCalls.h"
39 #include "llvm/Transforms/Utils/SizeOpts.h"
40
41 using namespace llvm;
42 using namespace PatternMatch;
43
44 static cl::opt<bool>
45 EnableUnsafeFPShrink("enable-double-float-shrink", cl::Hidden,
46 cl::init(false),
47 cl::desc("Enable unsafe double to float "
48 "shrinking for math lib calls"));
49
50 //===----------------------------------------------------------------------===//
51 // Helper Functions
52 //===----------------------------------------------------------------------===//
53
ignoreCallingConv(LibFunc Func)54 static bool ignoreCallingConv(LibFunc Func) {
55 return Func == LibFunc_abs || Func == LibFunc_labs ||
56 Func == LibFunc_llabs || Func == LibFunc_strlen;
57 }
58
isCallingConvCCompatible(CallInst * CI)59 static bool isCallingConvCCompatible(CallInst *CI) {
60 switch(CI->getCallingConv()) {
61 default:
62 return false;
63 case llvm::CallingConv::C:
64 return true;
65 case llvm::CallingConv::ARM_APCS:
66 case llvm::CallingConv::ARM_AAPCS:
67 case llvm::CallingConv::ARM_AAPCS_VFP: {
68
69 // The iOS ABI diverges from the standard in some cases, so for now don't
70 // try to simplify those calls.
71 if (Triple(CI->getModule()->getTargetTriple()).isiOS())
72 return false;
73
74 auto *FuncTy = CI->getFunctionType();
75
76 if (!FuncTy->getReturnType()->isPointerTy() &&
77 !FuncTy->getReturnType()->isIntegerTy() &&
78 !FuncTy->getReturnType()->isVoidTy())
79 return false;
80
81 for (auto Param : FuncTy->params()) {
82 if (!Param->isPointerTy() && !Param->isIntegerTy())
83 return false;
84 }
85 return true;
86 }
87 }
88 return false;
89 }
90
91 /// Return true if it is only used in equality comparisons with With.
isOnlyUsedInEqualityComparison(Value * V,Value * With)92 static bool isOnlyUsedInEqualityComparison(Value *V, Value *With) {
93 for (User *U : V->users()) {
94 if (ICmpInst *IC = dyn_cast<ICmpInst>(U))
95 if (IC->isEquality() && IC->getOperand(1) == With)
96 continue;
97 // Unknown instruction.
98 return false;
99 }
100 return true;
101 }
102
callHasFloatingPointArgument(const CallInst * CI)103 static bool callHasFloatingPointArgument(const CallInst *CI) {
104 return any_of(CI->operands(), [](const Use &OI) {
105 return OI->getType()->isFloatingPointTy();
106 });
107 }
108
callHasFP128Argument(const CallInst * CI)109 static bool callHasFP128Argument(const CallInst *CI) {
110 return any_of(CI->operands(), [](const Use &OI) {
111 return OI->getType()->isFP128Ty();
112 });
113 }
114
convertStrToNumber(CallInst * CI,StringRef & Str,int64_t Base)115 static Value *convertStrToNumber(CallInst *CI, StringRef &Str, int64_t Base) {
116 if (Base < 2 || Base > 36)
117 // handle special zero base
118 if (Base != 0)
119 return nullptr;
120
121 char *End;
122 std::string nptr = Str.str();
123 errno = 0;
124 long long int Result = strtoll(nptr.c_str(), &End, Base);
125 if (errno)
126 return nullptr;
127
128 // if we assume all possible target locales are ASCII supersets,
129 // then if strtoll successfully parses a number on the host,
130 // it will also successfully parse the same way on the target
131 if (*End != '\0')
132 return nullptr;
133
134 if (!isIntN(CI->getType()->getPrimitiveSizeInBits(), Result))
135 return nullptr;
136
137 return ConstantInt::get(CI->getType(), Result);
138 }
139
isOnlyUsedInComparisonWithZero(Value * V)140 static bool isOnlyUsedInComparisonWithZero(Value *V) {
141 for (User *U : V->users()) {
142 if (ICmpInst *IC = dyn_cast<ICmpInst>(U))
143 if (Constant *C = dyn_cast<Constant>(IC->getOperand(1)))
144 if (C->isNullValue())
145 continue;
146 // Unknown instruction.
147 return false;
148 }
149 return true;
150 }
151
canTransformToMemCmp(CallInst * CI,Value * Str,uint64_t Len,const DataLayout & DL)152 static bool canTransformToMemCmp(CallInst *CI, Value *Str, uint64_t Len,
153 const DataLayout &DL) {
154 if (!isOnlyUsedInComparisonWithZero(CI))
155 return false;
156
157 if (!isDereferenceableAndAlignedPointer(Str, Align(1), APInt(64, Len), DL))
158 return false;
159
160 if (CI->getFunction()->hasFnAttribute(Attribute::SanitizeMemory))
161 return false;
162
163 return true;
164 }
165
annotateDereferenceableBytes(CallInst * CI,ArrayRef<unsigned> ArgNos,uint64_t DereferenceableBytes)166 static void annotateDereferenceableBytes(CallInst *CI,
167 ArrayRef<unsigned> ArgNos,
168 uint64_t DereferenceableBytes) {
169 const Function *F = CI->getCaller();
170 if (!F)
171 return;
172 for (unsigned ArgNo : ArgNos) {
173 uint64_t DerefBytes = DereferenceableBytes;
174 unsigned AS = CI->getArgOperand(ArgNo)->getType()->getPointerAddressSpace();
175 if (!llvm::NullPointerIsDefined(F, AS) ||
176 CI->paramHasAttr(ArgNo, Attribute::NonNull))
177 DerefBytes = std::max(CI->getDereferenceableOrNullBytes(
178 ArgNo + AttributeList::FirstArgIndex),
179 DereferenceableBytes);
180
181 if (CI->getDereferenceableBytes(ArgNo + AttributeList::FirstArgIndex) <
182 DerefBytes) {
183 CI->removeParamAttr(ArgNo, Attribute::Dereferenceable);
184 if (!llvm::NullPointerIsDefined(F, AS) ||
185 CI->paramHasAttr(ArgNo, Attribute::NonNull))
186 CI->removeParamAttr(ArgNo, Attribute::DereferenceableOrNull);
187 CI->addParamAttr(ArgNo, Attribute::getWithDereferenceableBytes(
188 CI->getContext(), DerefBytes));
189 }
190 }
191 }
192
annotateNonNullBasedOnAccess(CallInst * CI,ArrayRef<unsigned> ArgNos)193 static void annotateNonNullBasedOnAccess(CallInst *CI,
194 ArrayRef<unsigned> ArgNos) {
195 Function *F = CI->getCaller();
196 if (!F)
197 return;
198
199 for (unsigned ArgNo : ArgNos) {
200 if (CI->paramHasAttr(ArgNo, Attribute::NonNull))
201 continue;
202 unsigned AS = CI->getArgOperand(ArgNo)->getType()->getPointerAddressSpace();
203 if (llvm::NullPointerIsDefined(F, AS))
204 continue;
205
206 CI->addParamAttr(ArgNo, Attribute::NonNull);
207 annotateDereferenceableBytes(CI, ArgNo, 1);
208 }
209 }
210
annotateNonNullAndDereferenceable(CallInst * CI,ArrayRef<unsigned> ArgNos,Value * Size,const DataLayout & DL)211 static void annotateNonNullAndDereferenceable(CallInst *CI, ArrayRef<unsigned> ArgNos,
212 Value *Size, const DataLayout &DL) {
213 if (ConstantInt *LenC = dyn_cast<ConstantInt>(Size)) {
214 annotateNonNullBasedOnAccess(CI, ArgNos);
215 annotateDereferenceableBytes(CI, ArgNos, LenC->getZExtValue());
216 } else if (isKnownNonZero(Size, DL)) {
217 annotateNonNullBasedOnAccess(CI, ArgNos);
218 const APInt *X, *Y;
219 uint64_t DerefMin = 1;
220 if (match(Size, m_Select(m_Value(), m_APInt(X), m_APInt(Y)))) {
221 DerefMin = std::min(X->getZExtValue(), Y->getZExtValue());
222 annotateDereferenceableBytes(CI, ArgNos, DerefMin);
223 }
224 }
225 }
226
227 //===----------------------------------------------------------------------===//
228 // String and Memory Library Call Optimizations
229 //===----------------------------------------------------------------------===//
230
optimizeStrCat(CallInst * CI,IRBuilderBase & B)231 Value *LibCallSimplifier::optimizeStrCat(CallInst *CI, IRBuilderBase &B) {
232 // Extract some information from the instruction
233 Value *Dst = CI->getArgOperand(0);
234 Value *Src = CI->getArgOperand(1);
235 annotateNonNullBasedOnAccess(CI, {0, 1});
236
237 // See if we can get the length of the input string.
238 uint64_t Len = GetStringLength(Src);
239 if (Len)
240 annotateDereferenceableBytes(CI, 1, Len);
241 else
242 return nullptr;
243 --Len; // Unbias length.
244
245 // Handle the simple, do-nothing case: strcat(x, "") -> x
246 if (Len == 0)
247 return Dst;
248
249 return emitStrLenMemCpy(Src, Dst, Len, B);
250 }
251
emitStrLenMemCpy(Value * Src,Value * Dst,uint64_t Len,IRBuilderBase & B)252 Value *LibCallSimplifier::emitStrLenMemCpy(Value *Src, Value *Dst, uint64_t Len,
253 IRBuilderBase &B) {
254 // We need to find the end of the destination string. That's where the
255 // memory is to be moved to. We just generate a call to strlen.
256 Value *DstLen = emitStrLen(Dst, B, DL, TLI);
257 if (!DstLen)
258 return nullptr;
259
260 // Now that we have the destination's length, we must index into the
261 // destination's pointer to get the actual memcpy destination (end of
262 // the string .. we're concatenating).
263 Value *CpyDst = B.CreateGEP(B.getInt8Ty(), Dst, DstLen, "endptr");
264
265 // We have enough information to now generate the memcpy call to do the
266 // concatenation for us. Make a memcpy to copy the nul byte with align = 1.
267 B.CreateMemCpy(
268 CpyDst, Align(1), Src, Align(1),
269 ConstantInt::get(DL.getIntPtrType(Src->getContext()), Len + 1));
270 return Dst;
271 }
272
optimizeStrNCat(CallInst * CI,IRBuilderBase & B)273 Value *LibCallSimplifier::optimizeStrNCat(CallInst *CI, IRBuilderBase &B) {
274 // Extract some information from the instruction.
275 Value *Dst = CI->getArgOperand(0);
276 Value *Src = CI->getArgOperand(1);
277 Value *Size = CI->getArgOperand(2);
278 uint64_t Len;
279 annotateNonNullBasedOnAccess(CI, 0);
280 if (isKnownNonZero(Size, DL))
281 annotateNonNullBasedOnAccess(CI, 1);
282
283 // We don't do anything if length is not constant.
284 ConstantInt *LengthArg = dyn_cast<ConstantInt>(Size);
285 if (LengthArg) {
286 Len = LengthArg->getZExtValue();
287 // strncat(x, c, 0) -> x
288 if (!Len)
289 return Dst;
290 } else {
291 return nullptr;
292 }
293
294 // See if we can get the length of the input string.
295 uint64_t SrcLen = GetStringLength(Src);
296 if (SrcLen) {
297 annotateDereferenceableBytes(CI, 1, SrcLen);
298 --SrcLen; // Unbias length.
299 } else {
300 return nullptr;
301 }
302
303 // strncat(x, "", c) -> x
304 if (SrcLen == 0)
305 return Dst;
306
307 // We don't optimize this case.
308 if (Len < SrcLen)
309 return nullptr;
310
311 // strncat(x, s, c) -> strcat(x, s)
312 // s is constant so the strcat can be optimized further.
313 return emitStrLenMemCpy(Src, Dst, SrcLen, B);
314 }
315
optimizeStrChr(CallInst * CI,IRBuilderBase & B)316 Value *LibCallSimplifier::optimizeStrChr(CallInst *CI, IRBuilderBase &B) {
317 Function *Callee = CI->getCalledFunction();
318 FunctionType *FT = Callee->getFunctionType();
319 Value *SrcStr = CI->getArgOperand(0);
320 annotateNonNullBasedOnAccess(CI, 0);
321
322 // If the second operand is non-constant, see if we can compute the length
323 // of the input string and turn this into memchr.
324 ConstantInt *CharC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
325 if (!CharC) {
326 uint64_t Len = GetStringLength(SrcStr);
327 if (Len)
328 annotateDereferenceableBytes(CI, 0, Len);
329 else
330 return nullptr;
331 if (!FT->getParamType(1)->isIntegerTy(32)) // memchr needs i32.
332 return nullptr;
333
334 return emitMemChr(SrcStr, CI->getArgOperand(1), // include nul.
335 ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len),
336 B, DL, TLI);
337 }
338
339 // Otherwise, the character is a constant, see if the first argument is
340 // a string literal. If so, we can constant fold.
341 StringRef Str;
342 if (!getConstantStringInfo(SrcStr, Str)) {
343 if (CharC->isZero()) // strchr(p, 0) -> p + strlen(p)
344 if (Value *StrLen = emitStrLen(SrcStr, B, DL, TLI))
345 return B.CreateGEP(B.getInt8Ty(), SrcStr, StrLen, "strchr");
346 return nullptr;
347 }
348
349 // Compute the offset, make sure to handle the case when we're searching for
350 // zero (a weird way to spell strlen).
351 size_t I = (0xFF & CharC->getSExtValue()) == 0
352 ? Str.size()
353 : Str.find(CharC->getSExtValue());
354 if (I == StringRef::npos) // Didn't find the char. strchr returns null.
355 return Constant::getNullValue(CI->getType());
356
357 // strchr(s+n,c) -> gep(s+n+i,c)
358 return B.CreateGEP(B.getInt8Ty(), SrcStr, B.getInt64(I), "strchr");
359 }
360
optimizeStrRChr(CallInst * CI,IRBuilderBase & B)361 Value *LibCallSimplifier::optimizeStrRChr(CallInst *CI, IRBuilderBase &B) {
362 Value *SrcStr = CI->getArgOperand(0);
363 ConstantInt *CharC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
364 annotateNonNullBasedOnAccess(CI, 0);
365
366 // Cannot fold anything if we're not looking for a constant.
367 if (!CharC)
368 return nullptr;
369
370 StringRef Str;
371 if (!getConstantStringInfo(SrcStr, Str)) {
372 // strrchr(s, 0) -> strchr(s, 0)
373 if (CharC->isZero())
374 return emitStrChr(SrcStr, '\0', B, TLI);
375 return nullptr;
376 }
377
378 // Compute the offset.
379 size_t I = (0xFF & CharC->getSExtValue()) == 0
380 ? Str.size()
381 : Str.rfind(CharC->getSExtValue());
382 if (I == StringRef::npos) // Didn't find the char. Return null.
383 return Constant::getNullValue(CI->getType());
384
385 // strrchr(s+n,c) -> gep(s+n+i,c)
386 return B.CreateGEP(B.getInt8Ty(), SrcStr, B.getInt64(I), "strrchr");
387 }
388
optimizeStrCmp(CallInst * CI,IRBuilderBase & B)389 Value *LibCallSimplifier::optimizeStrCmp(CallInst *CI, IRBuilderBase &B) {
390 Value *Str1P = CI->getArgOperand(0), *Str2P = CI->getArgOperand(1);
391 if (Str1P == Str2P) // strcmp(x,x) -> 0
392 return ConstantInt::get(CI->getType(), 0);
393
394 StringRef Str1, Str2;
395 bool HasStr1 = getConstantStringInfo(Str1P, Str1);
396 bool HasStr2 = getConstantStringInfo(Str2P, Str2);
397
398 // strcmp(x, y) -> cnst (if both x and y are constant strings)
399 if (HasStr1 && HasStr2)
400 return ConstantInt::get(CI->getType(), Str1.compare(Str2));
401
402 if (HasStr1 && Str1.empty()) // strcmp("", x) -> -*x
403 return B.CreateNeg(B.CreateZExt(
404 B.CreateLoad(B.getInt8Ty(), Str2P, "strcmpload"), CI->getType()));
405
406 if (HasStr2 && Str2.empty()) // strcmp(x,"") -> *x
407 return B.CreateZExt(B.CreateLoad(B.getInt8Ty(), Str1P, "strcmpload"),
408 CI->getType());
409
410 // strcmp(P, "x") -> memcmp(P, "x", 2)
411 uint64_t Len1 = GetStringLength(Str1P);
412 if (Len1)
413 annotateDereferenceableBytes(CI, 0, Len1);
414 uint64_t Len2 = GetStringLength(Str2P);
415 if (Len2)
416 annotateDereferenceableBytes(CI, 1, Len2);
417
418 if (Len1 && Len2) {
419 return emitMemCmp(Str1P, Str2P,
420 ConstantInt::get(DL.getIntPtrType(CI->getContext()),
421 std::min(Len1, Len2)),
422 B, DL, TLI);
423 }
424
425 // strcmp to memcmp
426 if (!HasStr1 && HasStr2) {
427 if (canTransformToMemCmp(CI, Str1P, Len2, DL))
428 return emitMemCmp(
429 Str1P, Str2P,
430 ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len2), B, DL,
431 TLI);
432 } else if (HasStr1 && !HasStr2) {
433 if (canTransformToMemCmp(CI, Str2P, Len1, DL))
434 return emitMemCmp(
435 Str1P, Str2P,
436 ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len1), B, DL,
437 TLI);
438 }
439
440 annotateNonNullBasedOnAccess(CI, {0, 1});
441 return nullptr;
442 }
443
optimizeStrNCmp(CallInst * CI,IRBuilderBase & B)444 Value *LibCallSimplifier::optimizeStrNCmp(CallInst *CI, IRBuilderBase &B) {
445 Value *Str1P = CI->getArgOperand(0);
446 Value *Str2P = CI->getArgOperand(1);
447 Value *Size = CI->getArgOperand(2);
448 if (Str1P == Str2P) // strncmp(x,x,n) -> 0
449 return ConstantInt::get(CI->getType(), 0);
450
451 if (isKnownNonZero(Size, DL))
452 annotateNonNullBasedOnAccess(CI, {0, 1});
453 // Get the length argument if it is constant.
454 uint64_t Length;
455 if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(Size))
456 Length = LengthArg->getZExtValue();
457 else
458 return nullptr;
459
460 if (Length == 0) // strncmp(x,y,0) -> 0
461 return ConstantInt::get(CI->getType(), 0);
462
463 if (Length == 1) // strncmp(x,y,1) -> memcmp(x,y,1)
464 return emitMemCmp(Str1P, Str2P, Size, B, DL, TLI);
465
466 StringRef Str1, Str2;
467 bool HasStr1 = getConstantStringInfo(Str1P, Str1);
468 bool HasStr2 = getConstantStringInfo(Str2P, Str2);
469
470 // strncmp(x, y) -> cnst (if both x and y are constant strings)
471 if (HasStr1 && HasStr2) {
472 StringRef SubStr1 = Str1.substr(0, Length);
473 StringRef SubStr2 = Str2.substr(0, Length);
474 return ConstantInt::get(CI->getType(), SubStr1.compare(SubStr2));
475 }
476
477 if (HasStr1 && Str1.empty()) // strncmp("", x, n) -> -*x
478 return B.CreateNeg(B.CreateZExt(
479 B.CreateLoad(B.getInt8Ty(), Str2P, "strcmpload"), CI->getType()));
480
481 if (HasStr2 && Str2.empty()) // strncmp(x, "", n) -> *x
482 return B.CreateZExt(B.CreateLoad(B.getInt8Ty(), Str1P, "strcmpload"),
483 CI->getType());
484
485 uint64_t Len1 = GetStringLength(Str1P);
486 if (Len1)
487 annotateDereferenceableBytes(CI, 0, Len1);
488 uint64_t Len2 = GetStringLength(Str2P);
489 if (Len2)
490 annotateDereferenceableBytes(CI, 1, Len2);
491
492 // strncmp to memcmp
493 if (!HasStr1 && HasStr2) {
494 Len2 = std::min(Len2, Length);
495 if (canTransformToMemCmp(CI, Str1P, Len2, DL))
496 return emitMemCmp(
497 Str1P, Str2P,
498 ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len2), B, DL,
499 TLI);
500 } else if (HasStr1 && !HasStr2) {
501 Len1 = std::min(Len1, Length);
502 if (canTransformToMemCmp(CI, Str2P, Len1, DL))
503 return emitMemCmp(
504 Str1P, Str2P,
505 ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len1), B, DL,
506 TLI);
507 }
508
509 return nullptr;
510 }
511
optimizeStrNDup(CallInst * CI,IRBuilderBase & B)512 Value *LibCallSimplifier::optimizeStrNDup(CallInst *CI, IRBuilderBase &B) {
513 Value *Src = CI->getArgOperand(0);
514 ConstantInt *Size = dyn_cast<ConstantInt>(CI->getArgOperand(1));
515 uint64_t SrcLen = GetStringLength(Src);
516 if (SrcLen && Size) {
517 annotateDereferenceableBytes(CI, 0, SrcLen);
518 if (SrcLen <= Size->getZExtValue() + 1)
519 return emitStrDup(Src, B, TLI);
520 }
521
522 return nullptr;
523 }
524
optimizeStrCpy(CallInst * CI,IRBuilderBase & B)525 Value *LibCallSimplifier::optimizeStrCpy(CallInst *CI, IRBuilderBase &B) {
526 Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1);
527 if (Dst == Src) // strcpy(x,x) -> x
528 return Src;
529
530 annotateNonNullBasedOnAccess(CI, {0, 1});
531 // See if we can get the length of the input string.
532 uint64_t Len = GetStringLength(Src);
533 if (Len)
534 annotateDereferenceableBytes(CI, 1, Len);
535 else
536 return nullptr;
537
538 // We have enough information to now generate the memcpy call to do the
539 // copy for us. Make a memcpy to copy the nul byte with align = 1.
540 CallInst *NewCI =
541 B.CreateMemCpy(Dst, Align(1), Src, Align(1),
542 ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len));
543 NewCI->setAttributes(CI->getAttributes());
544 return Dst;
545 }
546
optimizeStpCpy(CallInst * CI,IRBuilderBase & B)547 Value *LibCallSimplifier::optimizeStpCpy(CallInst *CI, IRBuilderBase &B) {
548 Function *Callee = CI->getCalledFunction();
549 Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1);
550 if (Dst == Src) { // stpcpy(x,x) -> x+strlen(x)
551 Value *StrLen = emitStrLen(Src, B, DL, TLI);
552 return StrLen ? B.CreateInBoundsGEP(B.getInt8Ty(), Dst, StrLen) : nullptr;
553 }
554
555 // See if we can get the length of the input string.
556 uint64_t Len = GetStringLength(Src);
557 if (Len)
558 annotateDereferenceableBytes(CI, 1, Len);
559 else
560 return nullptr;
561
562 Type *PT = Callee->getFunctionType()->getParamType(0);
563 Value *LenV = ConstantInt::get(DL.getIntPtrType(PT), Len);
564 Value *DstEnd = B.CreateGEP(B.getInt8Ty(), Dst,
565 ConstantInt::get(DL.getIntPtrType(PT), Len - 1));
566
567 // We have enough information to now generate the memcpy call to do the
568 // copy for us. Make a memcpy to copy the nul byte with align = 1.
569 CallInst *NewCI = B.CreateMemCpy(Dst, Align(1), Src, Align(1), LenV);
570 NewCI->setAttributes(CI->getAttributes());
571 return DstEnd;
572 }
573
optimizeStrNCpy(CallInst * CI,IRBuilderBase & B)574 Value *LibCallSimplifier::optimizeStrNCpy(CallInst *CI, IRBuilderBase &B) {
575 Function *Callee = CI->getCalledFunction();
576 Value *Dst = CI->getArgOperand(0);
577 Value *Src = CI->getArgOperand(1);
578 Value *Size = CI->getArgOperand(2);
579 annotateNonNullBasedOnAccess(CI, 0);
580 if (isKnownNonZero(Size, DL))
581 annotateNonNullBasedOnAccess(CI, 1);
582
583 uint64_t Len;
584 if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(Size))
585 Len = LengthArg->getZExtValue();
586 else
587 return nullptr;
588
589 // strncpy(x, y, 0) -> x
590 if (Len == 0)
591 return Dst;
592
593 // See if we can get the length of the input string.
594 uint64_t SrcLen = GetStringLength(Src);
595 if (SrcLen) {
596 annotateDereferenceableBytes(CI, 1, SrcLen);
597 --SrcLen; // Unbias length.
598 } else {
599 return nullptr;
600 }
601
602 if (SrcLen == 0) {
603 // strncpy(x, "", y) -> memset(align 1 x, '\0', y)
604 CallInst *NewCI = B.CreateMemSet(Dst, B.getInt8('\0'), Size, Align(1));
605 AttrBuilder ArgAttrs(CI->getAttributes().getParamAttributes(0));
606 NewCI->setAttributes(NewCI->getAttributes().addParamAttributes(
607 CI->getContext(), 0, ArgAttrs));
608 return Dst;
609 }
610
611 // strncpy(a, "a", 4) - > memcpy(a, "a\0\0\0", 4)
612 if (Len > SrcLen + 1) {
613 if (Len <= 128) {
614 StringRef Str;
615 if (!getConstantStringInfo(Src, Str))
616 return nullptr;
617 std::string SrcStr = Str.str();
618 SrcStr.resize(Len, '\0');
619 Src = B.CreateGlobalString(SrcStr, "str");
620 } else {
621 return nullptr;
622 }
623 }
624
625 Type *PT = Callee->getFunctionType()->getParamType(0);
626 // strncpy(x, s, c) -> memcpy(align 1 x, align 1 s, c) [s and c are constant]
627 CallInst *NewCI = B.CreateMemCpy(Dst, Align(1), Src, Align(1),
628 ConstantInt::get(DL.getIntPtrType(PT), Len));
629 NewCI->setAttributes(CI->getAttributes());
630 return Dst;
631 }
632
optimizeStringLength(CallInst * CI,IRBuilderBase & B,unsigned CharSize)633 Value *LibCallSimplifier::optimizeStringLength(CallInst *CI, IRBuilderBase &B,
634 unsigned CharSize) {
635 Value *Src = CI->getArgOperand(0);
636
637 // Constant folding: strlen("xyz") -> 3
638 if (uint64_t Len = GetStringLength(Src, CharSize))
639 return ConstantInt::get(CI->getType(), Len - 1);
640
641 // If s is a constant pointer pointing to a string literal, we can fold
642 // strlen(s + x) to strlen(s) - x, when x is known to be in the range
643 // [0, strlen(s)] or the string has a single null terminator '\0' at the end.
644 // We only try to simplify strlen when the pointer s points to an array
645 // of i8. Otherwise, we would need to scale the offset x before doing the
646 // subtraction. This will make the optimization more complex, and it's not
647 // very useful because calling strlen for a pointer of other types is
648 // very uncommon.
649 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Src)) {
650 if (!isGEPBasedOnPointerToString(GEP, CharSize))
651 return nullptr;
652
653 ConstantDataArraySlice Slice;
654 if (getConstantDataArrayInfo(GEP->getOperand(0), Slice, CharSize)) {
655 uint64_t NullTermIdx;
656 if (Slice.Array == nullptr) {
657 NullTermIdx = 0;
658 } else {
659 NullTermIdx = ~((uint64_t)0);
660 for (uint64_t I = 0, E = Slice.Length; I < E; ++I) {
661 if (Slice.Array->getElementAsInteger(I + Slice.Offset) == 0) {
662 NullTermIdx = I;
663 break;
664 }
665 }
666 // If the string does not have '\0', leave it to strlen to compute
667 // its length.
668 if (NullTermIdx == ~((uint64_t)0))
669 return nullptr;
670 }
671
672 Value *Offset = GEP->getOperand(2);
673 KnownBits Known = computeKnownBits(Offset, DL, 0, nullptr, CI, nullptr);
674 Known.Zero.flipAllBits();
675 uint64_t ArrSize =
676 cast<ArrayType>(GEP->getSourceElementType())->getNumElements();
677
678 // KnownZero's bits are flipped, so zeros in KnownZero now represent
679 // bits known to be zeros in Offset, and ones in KnowZero represent
680 // bits unknown in Offset. Therefore, Offset is known to be in range
681 // [0, NullTermIdx] when the flipped KnownZero is non-negative and
682 // unsigned-less-than NullTermIdx.
683 //
684 // If Offset is not provably in the range [0, NullTermIdx], we can still
685 // optimize if we can prove that the program has undefined behavior when
686 // Offset is outside that range. That is the case when GEP->getOperand(0)
687 // is a pointer to an object whose memory extent is NullTermIdx+1.
688 if ((Known.Zero.isNonNegative() && Known.Zero.ule(NullTermIdx)) ||
689 (GEP->isInBounds() && isa<GlobalVariable>(GEP->getOperand(0)) &&
690 NullTermIdx == ArrSize - 1)) {
691 Offset = B.CreateSExtOrTrunc(Offset, CI->getType());
692 return B.CreateSub(ConstantInt::get(CI->getType(), NullTermIdx),
693 Offset);
694 }
695 }
696 }
697
698 // strlen(x?"foo":"bars") --> x ? 3 : 4
699 if (SelectInst *SI = dyn_cast<SelectInst>(Src)) {
700 uint64_t LenTrue = GetStringLength(SI->getTrueValue(), CharSize);
701 uint64_t LenFalse = GetStringLength(SI->getFalseValue(), CharSize);
702 if (LenTrue && LenFalse) {
703 ORE.emit([&]() {
704 return OptimizationRemark("instcombine", "simplify-libcalls", CI)
705 << "folded strlen(select) to select of constants";
706 });
707 return B.CreateSelect(SI->getCondition(),
708 ConstantInt::get(CI->getType(), LenTrue - 1),
709 ConstantInt::get(CI->getType(), LenFalse - 1));
710 }
711 }
712
713 // strlen(x) != 0 --> *x != 0
714 // strlen(x) == 0 --> *x == 0
715 if (isOnlyUsedInZeroEqualityComparison(CI))
716 return B.CreateZExt(B.CreateLoad(B.getIntNTy(CharSize), Src, "strlenfirst"),
717 CI->getType());
718
719 return nullptr;
720 }
721
optimizeStrLen(CallInst * CI,IRBuilderBase & B)722 Value *LibCallSimplifier::optimizeStrLen(CallInst *CI, IRBuilderBase &B) {
723 if (Value *V = optimizeStringLength(CI, B, 8))
724 return V;
725 annotateNonNullBasedOnAccess(CI, 0);
726 return nullptr;
727 }
728
optimizeWcslen(CallInst * CI,IRBuilderBase & B)729 Value *LibCallSimplifier::optimizeWcslen(CallInst *CI, IRBuilderBase &B) {
730 Module &M = *CI->getModule();
731 unsigned WCharSize = TLI->getWCharSize(M) * 8;
732 // We cannot perform this optimization without wchar_size metadata.
733 if (WCharSize == 0)
734 return nullptr;
735
736 return optimizeStringLength(CI, B, WCharSize);
737 }
738
optimizeStrPBrk(CallInst * CI,IRBuilderBase & B)739 Value *LibCallSimplifier::optimizeStrPBrk(CallInst *CI, IRBuilderBase &B) {
740 StringRef S1, S2;
741 bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1);
742 bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2);
743
744 // strpbrk(s, "") -> nullptr
745 // strpbrk("", s) -> nullptr
746 if ((HasS1 && S1.empty()) || (HasS2 && S2.empty()))
747 return Constant::getNullValue(CI->getType());
748
749 // Constant folding.
750 if (HasS1 && HasS2) {
751 size_t I = S1.find_first_of(S2);
752 if (I == StringRef::npos) // No match.
753 return Constant::getNullValue(CI->getType());
754
755 return B.CreateGEP(B.getInt8Ty(), CI->getArgOperand(0), B.getInt64(I),
756 "strpbrk");
757 }
758
759 // strpbrk(s, "a") -> strchr(s, 'a')
760 if (HasS2 && S2.size() == 1)
761 return emitStrChr(CI->getArgOperand(0), S2[0], B, TLI);
762
763 return nullptr;
764 }
765
optimizeStrTo(CallInst * CI,IRBuilderBase & B)766 Value *LibCallSimplifier::optimizeStrTo(CallInst *CI, IRBuilderBase &B) {
767 Value *EndPtr = CI->getArgOperand(1);
768 if (isa<ConstantPointerNull>(EndPtr)) {
769 // With a null EndPtr, this function won't capture the main argument.
770 // It would be readonly too, except that it still may write to errno.
771 CI->addParamAttr(0, Attribute::NoCapture);
772 }
773
774 return nullptr;
775 }
776
optimizeStrSpn(CallInst * CI,IRBuilderBase & B)777 Value *LibCallSimplifier::optimizeStrSpn(CallInst *CI, IRBuilderBase &B) {
778 StringRef S1, S2;
779 bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1);
780 bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2);
781
782 // strspn(s, "") -> 0
783 // strspn("", s) -> 0
784 if ((HasS1 && S1.empty()) || (HasS2 && S2.empty()))
785 return Constant::getNullValue(CI->getType());
786
787 // Constant folding.
788 if (HasS1 && HasS2) {
789 size_t Pos = S1.find_first_not_of(S2);
790 if (Pos == StringRef::npos)
791 Pos = S1.size();
792 return ConstantInt::get(CI->getType(), Pos);
793 }
794
795 return nullptr;
796 }
797
optimizeStrCSpn(CallInst * CI,IRBuilderBase & B)798 Value *LibCallSimplifier::optimizeStrCSpn(CallInst *CI, IRBuilderBase &B) {
799 StringRef S1, S2;
800 bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1);
801 bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2);
802
803 // strcspn("", s) -> 0
804 if (HasS1 && S1.empty())
805 return Constant::getNullValue(CI->getType());
806
807 // Constant folding.
808 if (HasS1 && HasS2) {
809 size_t Pos = S1.find_first_of(S2);
810 if (Pos == StringRef::npos)
811 Pos = S1.size();
812 return ConstantInt::get(CI->getType(), Pos);
813 }
814
815 // strcspn(s, "") -> strlen(s)
816 if (HasS2 && S2.empty())
817 return emitStrLen(CI->getArgOperand(0), B, DL, TLI);
818
819 return nullptr;
820 }
821
optimizeStrStr(CallInst * CI,IRBuilderBase & B)822 Value *LibCallSimplifier::optimizeStrStr(CallInst *CI, IRBuilderBase &B) {
823 // fold strstr(x, x) -> x.
824 if (CI->getArgOperand(0) == CI->getArgOperand(1))
825 return B.CreateBitCast(CI->getArgOperand(0), CI->getType());
826
827 // fold strstr(a, b) == a -> strncmp(a, b, strlen(b)) == 0
828 if (isOnlyUsedInEqualityComparison(CI, CI->getArgOperand(0))) {
829 Value *StrLen = emitStrLen(CI->getArgOperand(1), B, DL, TLI);
830 if (!StrLen)
831 return nullptr;
832 Value *StrNCmp = emitStrNCmp(CI->getArgOperand(0), CI->getArgOperand(1),
833 StrLen, B, DL, TLI);
834 if (!StrNCmp)
835 return nullptr;
836 for (auto UI = CI->user_begin(), UE = CI->user_end(); UI != UE;) {
837 ICmpInst *Old = cast<ICmpInst>(*UI++);
838 Value *Cmp =
839 B.CreateICmp(Old->getPredicate(), StrNCmp,
840 ConstantInt::getNullValue(StrNCmp->getType()), "cmp");
841 replaceAllUsesWith(Old, Cmp);
842 }
843 return CI;
844 }
845
846 // See if either input string is a constant string.
847 StringRef SearchStr, ToFindStr;
848 bool HasStr1 = getConstantStringInfo(CI->getArgOperand(0), SearchStr);
849 bool HasStr2 = getConstantStringInfo(CI->getArgOperand(1), ToFindStr);
850
851 // fold strstr(x, "") -> x.
852 if (HasStr2 && ToFindStr.empty())
853 return B.CreateBitCast(CI->getArgOperand(0), CI->getType());
854
855 // If both strings are known, constant fold it.
856 if (HasStr1 && HasStr2) {
857 size_t Offset = SearchStr.find(ToFindStr);
858
859 if (Offset == StringRef::npos) // strstr("foo", "bar") -> null
860 return Constant::getNullValue(CI->getType());
861
862 // strstr("abcd", "bc") -> gep((char*)"abcd", 1)
863 Value *Result = castToCStr(CI->getArgOperand(0), B);
864 Result =
865 B.CreateConstInBoundsGEP1_64(B.getInt8Ty(), Result, Offset, "strstr");
866 return B.CreateBitCast(Result, CI->getType());
867 }
868
869 // fold strstr(x, "y") -> strchr(x, 'y').
870 if (HasStr2 && ToFindStr.size() == 1) {
871 Value *StrChr = emitStrChr(CI->getArgOperand(0), ToFindStr[0], B, TLI);
872 return StrChr ? B.CreateBitCast(StrChr, CI->getType()) : nullptr;
873 }
874
875 annotateNonNullBasedOnAccess(CI, {0, 1});
876 return nullptr;
877 }
878
optimizeMemRChr(CallInst * CI,IRBuilderBase & B)879 Value *LibCallSimplifier::optimizeMemRChr(CallInst *CI, IRBuilderBase &B) {
880 if (isKnownNonZero(CI->getOperand(2), DL))
881 annotateNonNullBasedOnAccess(CI, 0);
882 return nullptr;
883 }
884
optimizeMemChr(CallInst * CI,IRBuilderBase & B)885 Value *LibCallSimplifier::optimizeMemChr(CallInst *CI, IRBuilderBase &B) {
886 Value *SrcStr = CI->getArgOperand(0);
887 Value *Size = CI->getArgOperand(2);
888 annotateNonNullAndDereferenceable(CI, 0, Size, DL);
889 ConstantInt *CharC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
890 ConstantInt *LenC = dyn_cast<ConstantInt>(Size);
891
892 // memchr(x, y, 0) -> null
893 if (LenC) {
894 if (LenC->isZero())
895 return Constant::getNullValue(CI->getType());
896 } else {
897 // From now on we need at least constant length and string.
898 return nullptr;
899 }
900
901 StringRef Str;
902 if (!getConstantStringInfo(SrcStr, Str, 0, /*TrimAtNul=*/false))
903 return nullptr;
904
905 // Truncate the string to LenC. If Str is smaller than LenC we will still only
906 // scan the string, as reading past the end of it is undefined and we can just
907 // return null if we don't find the char.
908 Str = Str.substr(0, LenC->getZExtValue());
909
910 // If the char is variable but the input str and length are not we can turn
911 // this memchr call into a simple bit field test. Of course this only works
912 // when the return value is only checked against null.
913 //
914 // It would be really nice to reuse switch lowering here but we can't change
915 // the CFG at this point.
916 //
917 // memchr("\r\n", C, 2) != nullptr -> (1 << C & ((1 << '\r') | (1 << '\n')))
918 // != 0
919 // after bounds check.
920 if (!CharC && !Str.empty() && isOnlyUsedInZeroEqualityComparison(CI)) {
921 unsigned char Max =
922 *std::max_element(reinterpret_cast<const unsigned char *>(Str.begin()),
923 reinterpret_cast<const unsigned char *>(Str.end()));
924
925 // Make sure the bit field we're about to create fits in a register on the
926 // target.
927 // FIXME: On a 64 bit architecture this prevents us from using the
928 // interesting range of alpha ascii chars. We could do better by emitting
929 // two bitfields or shifting the range by 64 if no lower chars are used.
930 if (!DL.fitsInLegalInteger(Max + 1))
931 return nullptr;
932
933 // For the bit field use a power-of-2 type with at least 8 bits to avoid
934 // creating unnecessary illegal types.
935 unsigned char Width = NextPowerOf2(std::max((unsigned char)7, Max));
936
937 // Now build the bit field.
938 APInt Bitfield(Width, 0);
939 for (char C : Str)
940 Bitfield.setBit((unsigned char)C);
941 Value *BitfieldC = B.getInt(Bitfield);
942
943 // Adjust width of "C" to the bitfield width, then mask off the high bits.
944 Value *C = B.CreateZExtOrTrunc(CI->getArgOperand(1), BitfieldC->getType());
945 C = B.CreateAnd(C, B.getIntN(Width, 0xFF));
946
947 // First check that the bit field access is within bounds.
948 Value *Bounds = B.CreateICmp(ICmpInst::ICMP_ULT, C, B.getIntN(Width, Width),
949 "memchr.bounds");
950
951 // Create code that checks if the given bit is set in the field.
952 Value *Shl = B.CreateShl(B.getIntN(Width, 1ULL), C);
953 Value *Bits = B.CreateIsNotNull(B.CreateAnd(Shl, BitfieldC), "memchr.bits");
954
955 // Finally merge both checks and cast to pointer type. The inttoptr
956 // implicitly zexts the i1 to intptr type.
957 return B.CreateIntToPtr(B.CreateAnd(Bounds, Bits, "memchr"), CI->getType());
958 }
959
960 // Check if all arguments are constants. If so, we can constant fold.
961 if (!CharC)
962 return nullptr;
963
964 // Compute the offset.
965 size_t I = Str.find(CharC->getSExtValue() & 0xFF);
966 if (I == StringRef::npos) // Didn't find the char. memchr returns null.
967 return Constant::getNullValue(CI->getType());
968
969 // memchr(s+n,c,l) -> gep(s+n+i,c)
970 return B.CreateGEP(B.getInt8Ty(), SrcStr, B.getInt64(I), "memchr");
971 }
972
optimizeMemCmpConstantSize(CallInst * CI,Value * LHS,Value * RHS,uint64_t Len,IRBuilderBase & B,const DataLayout & DL)973 static Value *optimizeMemCmpConstantSize(CallInst *CI, Value *LHS, Value *RHS,
974 uint64_t Len, IRBuilderBase &B,
975 const DataLayout &DL) {
976 if (Len == 0) // memcmp(s1,s2,0) -> 0
977 return Constant::getNullValue(CI->getType());
978
979 // memcmp(S1,S2,1) -> *(unsigned char*)LHS - *(unsigned char*)RHS
980 if (Len == 1) {
981 Value *LHSV =
982 B.CreateZExt(B.CreateLoad(B.getInt8Ty(), castToCStr(LHS, B), "lhsc"),
983 CI->getType(), "lhsv");
984 Value *RHSV =
985 B.CreateZExt(B.CreateLoad(B.getInt8Ty(), castToCStr(RHS, B), "rhsc"),
986 CI->getType(), "rhsv");
987 return B.CreateSub(LHSV, RHSV, "chardiff");
988 }
989
990 // memcmp(S1,S2,N/8)==0 -> (*(intN_t*)S1 != *(intN_t*)S2)==0
991 // TODO: The case where both inputs are constants does not need to be limited
992 // to legal integers or equality comparison. See block below this.
993 if (DL.isLegalInteger(Len * 8) && isOnlyUsedInZeroEqualityComparison(CI)) {
994 IntegerType *IntType = IntegerType::get(CI->getContext(), Len * 8);
995 unsigned PrefAlignment = DL.getPrefTypeAlignment(IntType);
996
997 // First, see if we can fold either argument to a constant.
998 Value *LHSV = nullptr;
999 if (auto *LHSC = dyn_cast<Constant>(LHS)) {
1000 LHSC = ConstantExpr::getBitCast(LHSC, IntType->getPointerTo());
1001 LHSV = ConstantFoldLoadFromConstPtr(LHSC, IntType, DL);
1002 }
1003 Value *RHSV = nullptr;
1004 if (auto *RHSC = dyn_cast<Constant>(RHS)) {
1005 RHSC = ConstantExpr::getBitCast(RHSC, IntType->getPointerTo());
1006 RHSV = ConstantFoldLoadFromConstPtr(RHSC, IntType, DL);
1007 }
1008
1009 // Don't generate unaligned loads. If either source is constant data,
1010 // alignment doesn't matter for that source because there is no load.
1011 if ((LHSV || getKnownAlignment(LHS, DL, CI) >= PrefAlignment) &&
1012 (RHSV || getKnownAlignment(RHS, DL, CI) >= PrefAlignment)) {
1013 if (!LHSV) {
1014 Type *LHSPtrTy =
1015 IntType->getPointerTo(LHS->getType()->getPointerAddressSpace());
1016 LHSV = B.CreateLoad(IntType, B.CreateBitCast(LHS, LHSPtrTy), "lhsv");
1017 }
1018 if (!RHSV) {
1019 Type *RHSPtrTy =
1020 IntType->getPointerTo(RHS->getType()->getPointerAddressSpace());
1021 RHSV = B.CreateLoad(IntType, B.CreateBitCast(RHS, RHSPtrTy), "rhsv");
1022 }
1023 return B.CreateZExt(B.CreateICmpNE(LHSV, RHSV), CI->getType(), "memcmp");
1024 }
1025 }
1026
1027 // Constant folding: memcmp(x, y, Len) -> constant (all arguments are const).
1028 // TODO: This is limited to i8 arrays.
1029 StringRef LHSStr, RHSStr;
1030 if (getConstantStringInfo(LHS, LHSStr) &&
1031 getConstantStringInfo(RHS, RHSStr)) {
1032 // Make sure we're not reading out-of-bounds memory.
1033 if (Len > LHSStr.size() || Len > RHSStr.size())
1034 return nullptr;
1035 // Fold the memcmp and normalize the result. This way we get consistent
1036 // results across multiple platforms.
1037 uint64_t Ret = 0;
1038 int Cmp = memcmp(LHSStr.data(), RHSStr.data(), Len);
1039 if (Cmp < 0)
1040 Ret = -1;
1041 else if (Cmp > 0)
1042 Ret = 1;
1043 return ConstantInt::get(CI->getType(), Ret);
1044 }
1045
1046 return nullptr;
1047 }
1048
1049 // Most simplifications for memcmp also apply to bcmp.
optimizeMemCmpBCmpCommon(CallInst * CI,IRBuilderBase & B)1050 Value *LibCallSimplifier::optimizeMemCmpBCmpCommon(CallInst *CI,
1051 IRBuilderBase &B) {
1052 Value *LHS = CI->getArgOperand(0), *RHS = CI->getArgOperand(1);
1053 Value *Size = CI->getArgOperand(2);
1054
1055 if (LHS == RHS) // memcmp(s,s,x) -> 0
1056 return Constant::getNullValue(CI->getType());
1057
1058 annotateNonNullAndDereferenceable(CI, {0, 1}, Size, DL);
1059 // Handle constant lengths.
1060 ConstantInt *LenC = dyn_cast<ConstantInt>(Size);
1061 if (!LenC)
1062 return nullptr;
1063
1064 // memcmp(d,s,0) -> 0
1065 if (LenC->getZExtValue() == 0)
1066 return Constant::getNullValue(CI->getType());
1067
1068 if (Value *Res =
1069 optimizeMemCmpConstantSize(CI, LHS, RHS, LenC->getZExtValue(), B, DL))
1070 return Res;
1071 return nullptr;
1072 }
1073
optimizeMemCmp(CallInst * CI,IRBuilderBase & B)1074 Value *LibCallSimplifier::optimizeMemCmp(CallInst *CI, IRBuilderBase &B) {
1075 if (Value *V = optimizeMemCmpBCmpCommon(CI, B))
1076 return V;
1077
1078 // memcmp(x, y, Len) == 0 -> bcmp(x, y, Len) == 0
1079 // bcmp can be more efficient than memcmp because it only has to know that
1080 // there is a difference, not how different one is to the other.
1081 if (TLI->has(LibFunc_bcmp) && isOnlyUsedInZeroEqualityComparison(CI)) {
1082 Value *LHS = CI->getArgOperand(0);
1083 Value *RHS = CI->getArgOperand(1);
1084 Value *Size = CI->getArgOperand(2);
1085 return emitBCmp(LHS, RHS, Size, B, DL, TLI);
1086 }
1087
1088 return nullptr;
1089 }
1090
optimizeBCmp(CallInst * CI,IRBuilderBase & B)1091 Value *LibCallSimplifier::optimizeBCmp(CallInst *CI, IRBuilderBase &B) {
1092 return optimizeMemCmpBCmpCommon(CI, B);
1093 }
1094
optimizeMemCpy(CallInst * CI,IRBuilderBase & B)1095 Value *LibCallSimplifier::optimizeMemCpy(CallInst *CI, IRBuilderBase &B) {
1096 Value *Size = CI->getArgOperand(2);
1097 annotateNonNullAndDereferenceable(CI, {0, 1}, Size, DL);
1098 if (isa<IntrinsicInst>(CI))
1099 return nullptr;
1100
1101 // memcpy(x, y, n) -> llvm.memcpy(align 1 x, align 1 y, n)
1102 CallInst *NewCI = B.CreateMemCpy(CI->getArgOperand(0), Align(1),
1103 CI->getArgOperand(1), Align(1), Size);
1104 NewCI->setAttributes(CI->getAttributes());
1105 return CI->getArgOperand(0);
1106 }
1107
optimizeMemCCpy(CallInst * CI,IRBuilderBase & B)1108 Value *LibCallSimplifier::optimizeMemCCpy(CallInst *CI, IRBuilderBase &B) {
1109 Value *Dst = CI->getArgOperand(0);
1110 Value *Src = CI->getArgOperand(1);
1111 ConstantInt *StopChar = dyn_cast<ConstantInt>(CI->getArgOperand(2));
1112 ConstantInt *N = dyn_cast<ConstantInt>(CI->getArgOperand(3));
1113 StringRef SrcStr;
1114 if (CI->use_empty() && Dst == Src)
1115 return Dst;
1116 // memccpy(d, s, c, 0) -> nullptr
1117 if (N) {
1118 if (N->isNullValue())
1119 return Constant::getNullValue(CI->getType());
1120 if (!getConstantStringInfo(Src, SrcStr, /*Offset=*/0,
1121 /*TrimAtNul=*/false) ||
1122 !StopChar)
1123 return nullptr;
1124 } else {
1125 return nullptr;
1126 }
1127
1128 // Wrap arg 'c' of type int to char
1129 size_t Pos = SrcStr.find(StopChar->getSExtValue() & 0xFF);
1130 if (Pos == StringRef::npos) {
1131 if (N->getZExtValue() <= SrcStr.size()) {
1132 B.CreateMemCpy(Dst, Align(1), Src, Align(1), CI->getArgOperand(3));
1133 return Constant::getNullValue(CI->getType());
1134 }
1135 return nullptr;
1136 }
1137
1138 Value *NewN =
1139 ConstantInt::get(N->getType(), std::min(uint64_t(Pos + 1), N->getZExtValue()));
1140 // memccpy -> llvm.memcpy
1141 B.CreateMemCpy(Dst, Align(1), Src, Align(1), NewN);
1142 return Pos + 1 <= N->getZExtValue()
1143 ? B.CreateInBoundsGEP(B.getInt8Ty(), Dst, NewN)
1144 : Constant::getNullValue(CI->getType());
1145 }
1146
optimizeMemPCpy(CallInst * CI,IRBuilderBase & B)1147 Value *LibCallSimplifier::optimizeMemPCpy(CallInst *CI, IRBuilderBase &B) {
1148 Value *Dst = CI->getArgOperand(0);
1149 Value *N = CI->getArgOperand(2);
1150 // mempcpy(x, y, n) -> llvm.memcpy(align 1 x, align 1 y, n), x + n
1151 CallInst *NewCI =
1152 B.CreateMemCpy(Dst, Align(1), CI->getArgOperand(1), Align(1), N);
1153 NewCI->setAttributes(CI->getAttributes());
1154 return B.CreateInBoundsGEP(B.getInt8Ty(), Dst, N);
1155 }
1156
optimizeMemMove(CallInst * CI,IRBuilderBase & B)1157 Value *LibCallSimplifier::optimizeMemMove(CallInst *CI, IRBuilderBase &B) {
1158 Value *Size = CI->getArgOperand(2);
1159 annotateNonNullAndDereferenceable(CI, {0, 1}, Size, DL);
1160 if (isa<IntrinsicInst>(CI))
1161 return nullptr;
1162
1163 // memmove(x, y, n) -> llvm.memmove(align 1 x, align 1 y, n)
1164 CallInst *NewCI = B.CreateMemMove(CI->getArgOperand(0), Align(1),
1165 CI->getArgOperand(1), Align(1), Size);
1166 NewCI->setAttributes(CI->getAttributes());
1167 return CI->getArgOperand(0);
1168 }
1169
1170 /// Fold memset[_chk](malloc(n), 0, n) --> calloc(1, n).
foldMallocMemset(CallInst * Memset,IRBuilderBase & B)1171 Value *LibCallSimplifier::foldMallocMemset(CallInst *Memset, IRBuilderBase &B) {
1172 // This has to be a memset of zeros (bzero).
1173 auto *FillValue = dyn_cast<ConstantInt>(Memset->getArgOperand(1));
1174 if (!FillValue || FillValue->getZExtValue() != 0)
1175 return nullptr;
1176
1177 // TODO: We should handle the case where the malloc has more than one use.
1178 // This is necessary to optimize common patterns such as when the result of
1179 // the malloc is checked against null or when a memset intrinsic is used in
1180 // place of a memset library call.
1181 auto *Malloc = dyn_cast<CallInst>(Memset->getArgOperand(0));
1182 if (!Malloc || !Malloc->hasOneUse())
1183 return nullptr;
1184
1185 // Is the inner call really malloc()?
1186 Function *InnerCallee = Malloc->getCalledFunction();
1187 if (!InnerCallee)
1188 return nullptr;
1189
1190 LibFunc Func;
1191 if (!TLI->getLibFunc(*InnerCallee, Func) || !TLI->has(Func) ||
1192 Func != LibFunc_malloc)
1193 return nullptr;
1194
1195 // The memset must cover the same number of bytes that are malloc'd.
1196 if (Memset->getArgOperand(2) != Malloc->getArgOperand(0))
1197 return nullptr;
1198
1199 // Replace the malloc with a calloc. We need the data layout to know what the
1200 // actual size of a 'size_t' parameter is.
1201 B.SetInsertPoint(Malloc->getParent(), ++Malloc->getIterator());
1202 const DataLayout &DL = Malloc->getModule()->getDataLayout();
1203 IntegerType *SizeType = DL.getIntPtrType(B.GetInsertBlock()->getContext());
1204 if (Value *Calloc = emitCalloc(ConstantInt::get(SizeType, 1),
1205 Malloc->getArgOperand(0),
1206 Malloc->getAttributes(), B, *TLI)) {
1207 substituteInParent(Malloc, Calloc);
1208 return Calloc;
1209 }
1210
1211 return nullptr;
1212 }
1213
optimizeMemSet(CallInst * CI,IRBuilderBase & B)1214 Value *LibCallSimplifier::optimizeMemSet(CallInst *CI, IRBuilderBase &B) {
1215 Value *Size = CI->getArgOperand(2);
1216 annotateNonNullAndDereferenceable(CI, 0, Size, DL);
1217 if (isa<IntrinsicInst>(CI))
1218 return nullptr;
1219
1220 if (auto *Calloc = foldMallocMemset(CI, B))
1221 return Calloc;
1222
1223 // memset(p, v, n) -> llvm.memset(align 1 p, v, n)
1224 Value *Val = B.CreateIntCast(CI->getArgOperand(1), B.getInt8Ty(), false);
1225 CallInst *NewCI = B.CreateMemSet(CI->getArgOperand(0), Val, Size, Align(1));
1226 NewCI->setAttributes(CI->getAttributes());
1227 return CI->getArgOperand(0);
1228 }
1229
optimizeRealloc(CallInst * CI,IRBuilderBase & B)1230 Value *LibCallSimplifier::optimizeRealloc(CallInst *CI, IRBuilderBase &B) {
1231 if (isa<ConstantPointerNull>(CI->getArgOperand(0)))
1232 return emitMalloc(CI->getArgOperand(1), B, DL, TLI);
1233
1234 return nullptr;
1235 }
1236
1237 //===----------------------------------------------------------------------===//
1238 // Math Library Optimizations
1239 //===----------------------------------------------------------------------===//
1240
1241 // Replace a libcall \p CI with a call to intrinsic \p IID
replaceUnaryCall(CallInst * CI,IRBuilderBase & B,Intrinsic::ID IID)1242 static Value *replaceUnaryCall(CallInst *CI, IRBuilderBase &B,
1243 Intrinsic::ID IID) {
1244 // Propagate fast-math flags from the existing call to the new call.
1245 IRBuilderBase::FastMathFlagGuard Guard(B);
1246 B.setFastMathFlags(CI->getFastMathFlags());
1247
1248 Module *M = CI->getModule();
1249 Value *V = CI->getArgOperand(0);
1250 Function *F = Intrinsic::getDeclaration(M, IID, CI->getType());
1251 CallInst *NewCall = B.CreateCall(F, V);
1252 NewCall->takeName(CI);
1253 return NewCall;
1254 }
1255
1256 /// Return a variant of Val with float type.
1257 /// Currently this works in two cases: If Val is an FPExtension of a float
1258 /// value to something bigger, simply return the operand.
1259 /// If Val is a ConstantFP but can be converted to a float ConstantFP without
1260 /// loss of precision do so.
valueHasFloatPrecision(Value * Val)1261 static Value *valueHasFloatPrecision(Value *Val) {
1262 if (FPExtInst *Cast = dyn_cast<FPExtInst>(Val)) {
1263 Value *Op = Cast->getOperand(0);
1264 if (Op->getType()->isFloatTy())
1265 return Op;
1266 }
1267 if (ConstantFP *Const = dyn_cast<ConstantFP>(Val)) {
1268 APFloat F = Const->getValueAPF();
1269 bool losesInfo;
1270 (void)F.convert(APFloat::IEEEsingle(), APFloat::rmNearestTiesToEven,
1271 &losesInfo);
1272 if (!losesInfo)
1273 return ConstantFP::get(Const->getContext(), F);
1274 }
1275 return nullptr;
1276 }
1277
1278 /// Shrink double -> float functions.
optimizeDoubleFP(CallInst * CI,IRBuilderBase & B,bool isBinary,bool isPrecise=false)1279 static Value *optimizeDoubleFP(CallInst *CI, IRBuilderBase &B,
1280 bool isBinary, bool isPrecise = false) {
1281 Function *CalleeFn = CI->getCalledFunction();
1282 if (!CI->getType()->isDoubleTy() || !CalleeFn)
1283 return nullptr;
1284
1285 // If not all the uses of the function are converted to float, then bail out.
1286 // This matters if the precision of the result is more important than the
1287 // precision of the arguments.
1288 if (isPrecise)
1289 for (User *U : CI->users()) {
1290 FPTruncInst *Cast = dyn_cast<FPTruncInst>(U);
1291 if (!Cast || !Cast->getType()->isFloatTy())
1292 return nullptr;
1293 }
1294
1295 // If this is something like 'g((double) float)', convert to 'gf(float)'.
1296 Value *V[2];
1297 V[0] = valueHasFloatPrecision(CI->getArgOperand(0));
1298 V[1] = isBinary ? valueHasFloatPrecision(CI->getArgOperand(1)) : nullptr;
1299 if (!V[0] || (isBinary && !V[1]))
1300 return nullptr;
1301
1302 // If call isn't an intrinsic, check that it isn't within a function with the
1303 // same name as the float version of this call, otherwise the result is an
1304 // infinite loop. For example, from MinGW-w64:
1305 //
1306 // float expf(float val) { return (float) exp((double) val); }
1307 StringRef CalleeName = CalleeFn->getName();
1308 bool IsIntrinsic = CalleeFn->isIntrinsic();
1309 if (!IsIntrinsic) {
1310 StringRef CallerName = CI->getFunction()->getName();
1311 if (!CallerName.empty() && CallerName.back() == 'f' &&
1312 CallerName.size() == (CalleeName.size() + 1) &&
1313 CallerName.startswith(CalleeName))
1314 return nullptr;
1315 }
1316
1317 // Propagate the math semantics from the current function to the new function.
1318 IRBuilderBase::FastMathFlagGuard Guard(B);
1319 B.setFastMathFlags(CI->getFastMathFlags());
1320
1321 // g((double) float) -> (double) gf(float)
1322 Value *R;
1323 if (IsIntrinsic) {
1324 Module *M = CI->getModule();
1325 Intrinsic::ID IID = CalleeFn->getIntrinsicID();
1326 Function *Fn = Intrinsic::getDeclaration(M, IID, B.getFloatTy());
1327 R = isBinary ? B.CreateCall(Fn, V) : B.CreateCall(Fn, V[0]);
1328 } else {
1329 AttributeList CalleeAttrs = CalleeFn->getAttributes();
1330 R = isBinary ? emitBinaryFloatFnCall(V[0], V[1], CalleeName, B, CalleeAttrs)
1331 : emitUnaryFloatFnCall(V[0], CalleeName, B, CalleeAttrs);
1332 }
1333 return B.CreateFPExt(R, B.getDoubleTy());
1334 }
1335
1336 /// Shrink double -> float for unary functions.
optimizeUnaryDoubleFP(CallInst * CI,IRBuilderBase & B,bool isPrecise=false)1337 static Value *optimizeUnaryDoubleFP(CallInst *CI, IRBuilderBase &B,
1338 bool isPrecise = false) {
1339 return optimizeDoubleFP(CI, B, false, isPrecise);
1340 }
1341
1342 /// Shrink double -> float for binary functions.
optimizeBinaryDoubleFP(CallInst * CI,IRBuilderBase & B,bool isPrecise=false)1343 static Value *optimizeBinaryDoubleFP(CallInst *CI, IRBuilderBase &B,
1344 bool isPrecise = false) {
1345 return optimizeDoubleFP(CI, B, true, isPrecise);
1346 }
1347
1348 // cabs(z) -> sqrt((creal(z)*creal(z)) + (cimag(z)*cimag(z)))
optimizeCAbs(CallInst * CI,IRBuilderBase & B)1349 Value *LibCallSimplifier::optimizeCAbs(CallInst *CI, IRBuilderBase &B) {
1350 if (!CI->isFast())
1351 return nullptr;
1352
1353 // Propagate fast-math flags from the existing call to new instructions.
1354 IRBuilderBase::FastMathFlagGuard Guard(B);
1355 B.setFastMathFlags(CI->getFastMathFlags());
1356
1357 Value *Real, *Imag;
1358 if (CI->getNumArgOperands() == 1) {
1359 Value *Op = CI->getArgOperand(0);
1360 assert(Op->getType()->isArrayTy() && "Unexpected signature for cabs!");
1361 Real = B.CreateExtractValue(Op, 0, "real");
1362 Imag = B.CreateExtractValue(Op, 1, "imag");
1363 } else {
1364 assert(CI->getNumArgOperands() == 2 && "Unexpected signature for cabs!");
1365 Real = CI->getArgOperand(0);
1366 Imag = CI->getArgOperand(1);
1367 }
1368
1369 Value *RealReal = B.CreateFMul(Real, Real);
1370 Value *ImagImag = B.CreateFMul(Imag, Imag);
1371
1372 Function *FSqrt = Intrinsic::getDeclaration(CI->getModule(), Intrinsic::sqrt,
1373 CI->getType());
1374 return B.CreateCall(FSqrt, B.CreateFAdd(RealReal, ImagImag), "cabs");
1375 }
1376
optimizeTrigReflections(CallInst * Call,LibFunc Func,IRBuilderBase & B)1377 static Value *optimizeTrigReflections(CallInst *Call, LibFunc Func,
1378 IRBuilderBase &B) {
1379 if (!isa<FPMathOperator>(Call))
1380 return nullptr;
1381
1382 IRBuilderBase::FastMathFlagGuard Guard(B);
1383 B.setFastMathFlags(Call->getFastMathFlags());
1384
1385 // TODO: Can this be shared to also handle LLVM intrinsics?
1386 Value *X;
1387 switch (Func) {
1388 case LibFunc_sin:
1389 case LibFunc_sinf:
1390 case LibFunc_sinl:
1391 case LibFunc_tan:
1392 case LibFunc_tanf:
1393 case LibFunc_tanl:
1394 // sin(-X) --> -sin(X)
1395 // tan(-X) --> -tan(X)
1396 if (match(Call->getArgOperand(0), m_OneUse(m_FNeg(m_Value(X)))))
1397 return B.CreateFNeg(B.CreateCall(Call->getCalledFunction(), X));
1398 break;
1399 case LibFunc_cos:
1400 case LibFunc_cosf:
1401 case LibFunc_cosl:
1402 // cos(-X) --> cos(X)
1403 if (match(Call->getArgOperand(0), m_FNeg(m_Value(X))))
1404 return B.CreateCall(Call->getCalledFunction(), X, "cos");
1405 break;
1406 default:
1407 break;
1408 }
1409 return nullptr;
1410 }
1411
getPow(Value * InnerChain[33],unsigned Exp,IRBuilderBase & B)1412 static Value *getPow(Value *InnerChain[33], unsigned Exp, IRBuilderBase &B) {
1413 // Multiplications calculated using Addition Chains.
1414 // Refer: http://wwwhomes.uni-bielefeld.de/achim/addition_chain.html
1415
1416 assert(Exp != 0 && "Incorrect exponent 0 not handled");
1417
1418 if (InnerChain[Exp])
1419 return InnerChain[Exp];
1420
1421 static const unsigned AddChain[33][2] = {
1422 {0, 0}, // Unused.
1423 {0, 0}, // Unused (base case = pow1).
1424 {1, 1}, // Unused (pre-computed).
1425 {1, 2}, {2, 2}, {2, 3}, {3, 3}, {2, 5}, {4, 4},
1426 {1, 8}, {5, 5}, {1, 10}, {6, 6}, {4, 9}, {7, 7},
1427 {3, 12}, {8, 8}, {8, 9}, {2, 16}, {1, 18}, {10, 10},
1428 {6, 15}, {11, 11}, {3, 20}, {12, 12}, {8, 17}, {13, 13},
1429 {3, 24}, {14, 14}, {4, 25}, {15, 15}, {3, 28}, {16, 16},
1430 };
1431
1432 InnerChain[Exp] = B.CreateFMul(getPow(InnerChain, AddChain[Exp][0], B),
1433 getPow(InnerChain, AddChain[Exp][1], B));
1434 return InnerChain[Exp];
1435 }
1436
1437 // Return a properly extended 32-bit integer if the operation is an itofp.
getIntToFPVal(Value * I2F,IRBuilderBase & B)1438 static Value *getIntToFPVal(Value *I2F, IRBuilderBase &B) {
1439 if (isa<SIToFPInst>(I2F) || isa<UIToFPInst>(I2F)) {
1440 Value *Op = cast<Instruction>(I2F)->getOperand(0);
1441 // Make sure that the exponent fits inside an int32_t,
1442 // thus avoiding any range issues that FP has not.
1443 unsigned BitWidth = Op->getType()->getPrimitiveSizeInBits();
1444 if (BitWidth < 32 ||
1445 (BitWidth == 32 && isa<SIToFPInst>(I2F)))
1446 return isa<SIToFPInst>(I2F) ? B.CreateSExt(Op, B.getInt32Ty())
1447 : B.CreateZExt(Op, B.getInt32Ty());
1448 }
1449
1450 return nullptr;
1451 }
1452
1453 /// Use exp{,2}(x * y) for pow(exp{,2}(x), y);
1454 /// ldexp(1.0, x) for pow(2.0, itofp(x)); exp2(n * x) for pow(2.0 ** n, x);
1455 /// exp10(x) for pow(10.0, x); exp2(log2(n) * x) for pow(n, x).
replacePowWithExp(CallInst * Pow,IRBuilderBase & B)1456 Value *LibCallSimplifier::replacePowWithExp(CallInst *Pow, IRBuilderBase &B) {
1457 Value *Base = Pow->getArgOperand(0), *Expo = Pow->getArgOperand(1);
1458 AttributeList Attrs; // Attributes are only meaningful on the original call
1459 Module *Mod = Pow->getModule();
1460 Type *Ty = Pow->getType();
1461 bool Ignored;
1462
1463 // Evaluate special cases related to a nested function as the base.
1464
1465 // pow(exp(x), y) -> exp(x * y)
1466 // pow(exp2(x), y) -> exp2(x * y)
1467 // If exp{,2}() is used only once, it is better to fold two transcendental
1468 // math functions into one. If used again, exp{,2}() would still have to be
1469 // called with the original argument, then keep both original transcendental
1470 // functions. However, this transformation is only safe with fully relaxed
1471 // math semantics, since, besides rounding differences, it changes overflow
1472 // and underflow behavior quite dramatically. For example:
1473 // pow(exp(1000), 0.001) = pow(inf, 0.001) = inf
1474 // Whereas:
1475 // exp(1000 * 0.001) = exp(1)
1476 // TODO: Loosen the requirement for fully relaxed math semantics.
1477 // TODO: Handle exp10() when more targets have it available.
1478 CallInst *BaseFn = dyn_cast<CallInst>(Base);
1479 if (BaseFn && BaseFn->hasOneUse() && BaseFn->isFast() && Pow->isFast()) {
1480 LibFunc LibFn;
1481
1482 Function *CalleeFn = BaseFn->getCalledFunction();
1483 if (CalleeFn &&
1484 TLI->getLibFunc(CalleeFn->getName(), LibFn) && TLI->has(LibFn)) {
1485 StringRef ExpName;
1486 Intrinsic::ID ID;
1487 Value *ExpFn;
1488 LibFunc LibFnFloat, LibFnDouble, LibFnLongDouble;
1489
1490 switch (LibFn) {
1491 default:
1492 return nullptr;
1493 case LibFunc_expf: case LibFunc_exp: case LibFunc_expl:
1494 ExpName = TLI->getName(LibFunc_exp);
1495 ID = Intrinsic::exp;
1496 LibFnFloat = LibFunc_expf;
1497 LibFnDouble = LibFunc_exp;
1498 LibFnLongDouble = LibFunc_expl;
1499 break;
1500 case LibFunc_exp2f: case LibFunc_exp2: case LibFunc_exp2l:
1501 ExpName = TLI->getName(LibFunc_exp2);
1502 ID = Intrinsic::exp2;
1503 LibFnFloat = LibFunc_exp2f;
1504 LibFnDouble = LibFunc_exp2;
1505 LibFnLongDouble = LibFunc_exp2l;
1506 break;
1507 }
1508
1509 // Create new exp{,2}() with the product as its argument.
1510 Value *FMul = B.CreateFMul(BaseFn->getArgOperand(0), Expo, "mul");
1511 ExpFn = BaseFn->doesNotAccessMemory()
1512 ? B.CreateCall(Intrinsic::getDeclaration(Mod, ID, Ty),
1513 FMul, ExpName)
1514 : emitUnaryFloatFnCall(FMul, TLI, LibFnDouble, LibFnFloat,
1515 LibFnLongDouble, B,
1516 BaseFn->getAttributes());
1517
1518 // Since the new exp{,2}() is different from the original one, dead code
1519 // elimination cannot be trusted to remove it, since it may have side
1520 // effects (e.g., errno). When the only consumer for the original
1521 // exp{,2}() is pow(), then it has to be explicitly erased.
1522 substituteInParent(BaseFn, ExpFn);
1523 return ExpFn;
1524 }
1525 }
1526
1527 // Evaluate special cases related to a constant base.
1528
1529 const APFloat *BaseF;
1530 if (!match(Pow->getArgOperand(0), m_APFloat(BaseF)))
1531 return nullptr;
1532
1533 // pow(2.0, itofp(x)) -> ldexp(1.0, x)
1534 if (match(Base, m_SpecificFP(2.0)) &&
1535 (isa<SIToFPInst>(Expo) || isa<UIToFPInst>(Expo)) &&
1536 hasFloatFn(TLI, Ty, LibFunc_ldexp, LibFunc_ldexpf, LibFunc_ldexpl)) {
1537 if (Value *ExpoI = getIntToFPVal(Expo, B))
1538 return emitBinaryFloatFnCall(ConstantFP::get(Ty, 1.0), ExpoI, TLI,
1539 LibFunc_ldexp, LibFunc_ldexpf, LibFunc_ldexpl,
1540 B, Attrs);
1541 }
1542
1543 // pow(2.0 ** n, x) -> exp2(n * x)
1544 if (hasFloatFn(TLI, Ty, LibFunc_exp2, LibFunc_exp2f, LibFunc_exp2l)) {
1545 APFloat BaseR = APFloat(1.0);
1546 BaseR.convert(BaseF->getSemantics(), APFloat::rmTowardZero, &Ignored);
1547 BaseR = BaseR / *BaseF;
1548 bool IsInteger = BaseF->isInteger(), IsReciprocal = BaseR.isInteger();
1549 const APFloat *NF = IsReciprocal ? &BaseR : BaseF;
1550 APSInt NI(64, false);
1551 if ((IsInteger || IsReciprocal) &&
1552 NF->convertToInteger(NI, APFloat::rmTowardZero, &Ignored) ==
1553 APFloat::opOK &&
1554 NI > 1 && NI.isPowerOf2()) {
1555 double N = NI.logBase2() * (IsReciprocal ? -1.0 : 1.0);
1556 Value *FMul = B.CreateFMul(Expo, ConstantFP::get(Ty, N), "mul");
1557 if (Pow->doesNotAccessMemory())
1558 return B.CreateCall(Intrinsic::getDeclaration(Mod, Intrinsic::exp2, Ty),
1559 FMul, "exp2");
1560 else
1561 return emitUnaryFloatFnCall(FMul, TLI, LibFunc_exp2, LibFunc_exp2f,
1562 LibFunc_exp2l, B, Attrs);
1563 }
1564 }
1565
1566 // pow(10.0, x) -> exp10(x)
1567 // TODO: There is no exp10() intrinsic yet, but some day there shall be one.
1568 if (match(Base, m_SpecificFP(10.0)) &&
1569 hasFloatFn(TLI, Ty, LibFunc_exp10, LibFunc_exp10f, LibFunc_exp10l))
1570 return emitUnaryFloatFnCall(Expo, TLI, LibFunc_exp10, LibFunc_exp10f,
1571 LibFunc_exp10l, B, Attrs);
1572
1573 // pow(x, y) -> exp2(log2(x) * y)
1574 if (Pow->hasApproxFunc() && Pow->hasNoNaNs() && BaseF->isFiniteNonZero() &&
1575 !BaseF->isNegative()) {
1576 // pow(1, inf) is defined to be 1 but exp2(log2(1) * inf) evaluates to NaN.
1577 // Luckily optimizePow has already handled the x == 1 case.
1578 assert(!match(Base, m_FPOne()) &&
1579 "pow(1.0, y) should have been simplified earlier!");
1580
1581 Value *Log = nullptr;
1582 if (Ty->isFloatTy())
1583 Log = ConstantFP::get(Ty, std::log2(BaseF->convertToFloat()));
1584 else if (Ty->isDoubleTy())
1585 Log = ConstantFP::get(Ty, std::log2(BaseF->convertToDouble()));
1586
1587 if (Log) {
1588 Value *FMul = B.CreateFMul(Log, Expo, "mul");
1589 if (Pow->doesNotAccessMemory())
1590 return B.CreateCall(Intrinsic::getDeclaration(Mod, Intrinsic::exp2, Ty),
1591 FMul, "exp2");
1592 else if (hasFloatFn(TLI, Ty, LibFunc_exp2, LibFunc_exp2f, LibFunc_exp2l))
1593 return emitUnaryFloatFnCall(FMul, TLI, LibFunc_exp2, LibFunc_exp2f,
1594 LibFunc_exp2l, B, Attrs);
1595 }
1596 }
1597
1598 return nullptr;
1599 }
1600
getSqrtCall(Value * V,AttributeList Attrs,bool NoErrno,Module * M,IRBuilderBase & B,const TargetLibraryInfo * TLI)1601 static Value *getSqrtCall(Value *V, AttributeList Attrs, bool NoErrno,
1602 Module *M, IRBuilderBase &B,
1603 const TargetLibraryInfo *TLI) {
1604 // If errno is never set, then use the intrinsic for sqrt().
1605 if (NoErrno) {
1606 Function *SqrtFn =
1607 Intrinsic::getDeclaration(M, Intrinsic::sqrt, V->getType());
1608 return B.CreateCall(SqrtFn, V, "sqrt");
1609 }
1610
1611 // Otherwise, use the libcall for sqrt().
1612 if (hasFloatFn(TLI, V->getType(), LibFunc_sqrt, LibFunc_sqrtf, LibFunc_sqrtl))
1613 // TODO: We also should check that the target can in fact lower the sqrt()
1614 // libcall. We currently have no way to ask this question, so we ask if
1615 // the target has a sqrt() libcall, which is not exactly the same.
1616 return emitUnaryFloatFnCall(V, TLI, LibFunc_sqrt, LibFunc_sqrtf,
1617 LibFunc_sqrtl, B, Attrs);
1618
1619 return nullptr;
1620 }
1621
1622 /// Use square root in place of pow(x, +/-0.5).
replacePowWithSqrt(CallInst * Pow,IRBuilderBase & B)1623 Value *LibCallSimplifier::replacePowWithSqrt(CallInst *Pow, IRBuilderBase &B) {
1624 Value *Sqrt, *Base = Pow->getArgOperand(0), *Expo = Pow->getArgOperand(1);
1625 AttributeList Attrs; // Attributes are only meaningful on the original call
1626 Module *Mod = Pow->getModule();
1627 Type *Ty = Pow->getType();
1628
1629 const APFloat *ExpoF;
1630 if (!match(Expo, m_APFloat(ExpoF)) ||
1631 (!ExpoF->isExactlyValue(0.5) && !ExpoF->isExactlyValue(-0.5)))
1632 return nullptr;
1633
1634 // Converting pow(X, -0.5) to 1/sqrt(X) may introduce an extra rounding step,
1635 // so that requires fast-math-flags (afn or reassoc).
1636 if (ExpoF->isNegative() && (!Pow->hasApproxFunc() && !Pow->hasAllowReassoc()))
1637 return nullptr;
1638
1639 // If we have a pow() library call (accesses memory) and we can't guarantee
1640 // that the base is not an infinity, give up:
1641 // pow(-Inf, 0.5) is optionally required to have a result of +Inf (not setting
1642 // errno), but sqrt(-Inf) is required by various standards to set errno.
1643 if (!Pow->doesNotAccessMemory() && !Pow->hasNoInfs() &&
1644 !isKnownNeverInfinity(Base, TLI))
1645 return nullptr;
1646
1647 Sqrt = getSqrtCall(Base, Attrs, Pow->doesNotAccessMemory(), Mod, B, TLI);
1648 if (!Sqrt)
1649 return nullptr;
1650
1651 // Handle signed zero base by expanding to fabs(sqrt(x)).
1652 if (!Pow->hasNoSignedZeros()) {
1653 Function *FAbsFn = Intrinsic::getDeclaration(Mod, Intrinsic::fabs, Ty);
1654 Sqrt = B.CreateCall(FAbsFn, Sqrt, "abs");
1655 }
1656
1657 // Handle non finite base by expanding to
1658 // (x == -infinity ? +infinity : sqrt(x)).
1659 if (!Pow->hasNoInfs()) {
1660 Value *PosInf = ConstantFP::getInfinity(Ty),
1661 *NegInf = ConstantFP::getInfinity(Ty, true);
1662 Value *FCmp = B.CreateFCmpOEQ(Base, NegInf, "isinf");
1663 Sqrt = B.CreateSelect(FCmp, PosInf, Sqrt);
1664 }
1665
1666 // If the exponent is negative, then get the reciprocal.
1667 if (ExpoF->isNegative())
1668 Sqrt = B.CreateFDiv(ConstantFP::get(Ty, 1.0), Sqrt, "reciprocal");
1669
1670 return Sqrt;
1671 }
1672
createPowWithIntegerExponent(Value * Base,Value * Expo,Module * M,IRBuilderBase & B)1673 static Value *createPowWithIntegerExponent(Value *Base, Value *Expo, Module *M,
1674 IRBuilderBase &B) {
1675 Value *Args[] = {Base, Expo};
1676 Function *F = Intrinsic::getDeclaration(M, Intrinsic::powi, Base->getType());
1677 return B.CreateCall(F, Args);
1678 }
1679
optimizePow(CallInst * Pow,IRBuilderBase & B)1680 Value *LibCallSimplifier::optimizePow(CallInst *Pow, IRBuilderBase &B) {
1681 Value *Base = Pow->getArgOperand(0);
1682 Value *Expo = Pow->getArgOperand(1);
1683 Function *Callee = Pow->getCalledFunction();
1684 StringRef Name = Callee->getName();
1685 Type *Ty = Pow->getType();
1686 Module *M = Pow->getModule();
1687 Value *Shrunk = nullptr;
1688 bool AllowApprox = Pow->hasApproxFunc();
1689 bool Ignored;
1690
1691 // Propagate the math semantics from the call to any created instructions.
1692 IRBuilderBase::FastMathFlagGuard Guard(B);
1693 B.setFastMathFlags(Pow->getFastMathFlags());
1694
1695 // Shrink pow() to powf() if the arguments are single precision,
1696 // unless the result is expected to be double precision.
1697 if (UnsafeFPShrink && Name == TLI->getName(LibFunc_pow) &&
1698 hasFloatVersion(Name))
1699 Shrunk = optimizeBinaryDoubleFP(Pow, B, true);
1700
1701 // Evaluate special cases related to the base.
1702
1703 // pow(1.0, x) -> 1.0
1704 if (match(Base, m_FPOne()))
1705 return Base;
1706
1707 if (Value *Exp = replacePowWithExp(Pow, B))
1708 return Exp;
1709
1710 // Evaluate special cases related to the exponent.
1711
1712 // pow(x, -1.0) -> 1.0 / x
1713 if (match(Expo, m_SpecificFP(-1.0)))
1714 return B.CreateFDiv(ConstantFP::get(Ty, 1.0), Base, "reciprocal");
1715
1716 // pow(x, +/-0.0) -> 1.0
1717 if (match(Expo, m_AnyZeroFP()))
1718 return ConstantFP::get(Ty, 1.0);
1719
1720 // pow(x, 1.0) -> x
1721 if (match(Expo, m_FPOne()))
1722 return Base;
1723
1724 // pow(x, 2.0) -> x * x
1725 if (match(Expo, m_SpecificFP(2.0)))
1726 return B.CreateFMul(Base, Base, "square");
1727
1728 if (Value *Sqrt = replacePowWithSqrt(Pow, B))
1729 return Sqrt;
1730
1731 // pow(x, n) -> x * x * x * ...
1732 const APFloat *ExpoF;
1733 if (AllowApprox && match(Expo, m_APFloat(ExpoF)) &&
1734 !ExpoF->isExactlyValue(0.5) && !ExpoF->isExactlyValue(-0.5)) {
1735 // We limit to a max of 7 multiplications, thus the maximum exponent is 32.
1736 // If the exponent is an integer+0.5 we generate a call to sqrt and an
1737 // additional fmul.
1738 // TODO: This whole transformation should be backend specific (e.g. some
1739 // backends might prefer libcalls or the limit for the exponent might
1740 // be different) and it should also consider optimizing for size.
1741 APFloat LimF(ExpoF->getSemantics(), 33),
1742 ExpoA(abs(*ExpoF));
1743 if (ExpoA < LimF) {
1744 // This transformation applies to integer or integer+0.5 exponents only.
1745 // For integer+0.5, we create a sqrt(Base) call.
1746 Value *Sqrt = nullptr;
1747 if (!ExpoA.isInteger()) {
1748 APFloat Expo2 = ExpoA;
1749 // To check if ExpoA is an integer + 0.5, we add it to itself. If there
1750 // is no floating point exception and the result is an integer, then
1751 // ExpoA == integer + 0.5
1752 if (Expo2.add(ExpoA, APFloat::rmNearestTiesToEven) != APFloat::opOK)
1753 return nullptr;
1754
1755 if (!Expo2.isInteger())
1756 return nullptr;
1757
1758 Sqrt = getSqrtCall(Base, Pow->getCalledFunction()->getAttributes(),
1759 Pow->doesNotAccessMemory(), M, B, TLI);
1760 if (!Sqrt)
1761 return nullptr;
1762 }
1763
1764 // We will memoize intermediate products of the Addition Chain.
1765 Value *InnerChain[33] = {nullptr};
1766 InnerChain[1] = Base;
1767 InnerChain[2] = B.CreateFMul(Base, Base, "square");
1768
1769 // We cannot readily convert a non-double type (like float) to a double.
1770 // So we first convert it to something which could be converted to double.
1771 ExpoA.convert(APFloat::IEEEdouble(), APFloat::rmTowardZero, &Ignored);
1772 Value *FMul = getPow(InnerChain, ExpoA.convertToDouble(), B);
1773
1774 // Expand pow(x, y+0.5) to pow(x, y) * sqrt(x).
1775 if (Sqrt)
1776 FMul = B.CreateFMul(FMul, Sqrt);
1777
1778 // If the exponent is negative, then get the reciprocal.
1779 if (ExpoF->isNegative())
1780 FMul = B.CreateFDiv(ConstantFP::get(Ty, 1.0), FMul, "reciprocal");
1781
1782 return FMul;
1783 }
1784
1785 APSInt IntExpo(32, /*isUnsigned=*/false);
1786 // powf(x, n) -> powi(x, n) if n is a constant signed integer value
1787 if (ExpoF->isInteger() &&
1788 ExpoF->convertToInteger(IntExpo, APFloat::rmTowardZero, &Ignored) ==
1789 APFloat::opOK) {
1790 return createPowWithIntegerExponent(
1791 Base, ConstantInt::get(B.getInt32Ty(), IntExpo), M, B);
1792 }
1793 }
1794
1795 // powf(x, itofp(y)) -> powi(x, y)
1796 if (AllowApprox && (isa<SIToFPInst>(Expo) || isa<UIToFPInst>(Expo))) {
1797 if (Value *ExpoI = getIntToFPVal(Expo, B))
1798 return createPowWithIntegerExponent(Base, ExpoI, M, B);
1799 }
1800
1801 return Shrunk;
1802 }
1803
optimizeExp2(CallInst * CI,IRBuilderBase & B)1804 Value *LibCallSimplifier::optimizeExp2(CallInst *CI, IRBuilderBase &B) {
1805 Function *Callee = CI->getCalledFunction();
1806 AttributeList Attrs; // Attributes are only meaningful on the original call
1807 StringRef Name = Callee->getName();
1808 Value *Ret = nullptr;
1809 if (UnsafeFPShrink && Name == TLI->getName(LibFunc_exp2) &&
1810 hasFloatVersion(Name))
1811 Ret = optimizeUnaryDoubleFP(CI, B, true);
1812
1813 Type *Ty = CI->getType();
1814 Value *Op = CI->getArgOperand(0);
1815
1816 // Turn exp2(sitofp(x)) -> ldexp(1.0, sext(x)) if sizeof(x) <= 32
1817 // Turn exp2(uitofp(x)) -> ldexp(1.0, zext(x)) if sizeof(x) < 32
1818 if ((isa<SIToFPInst>(Op) || isa<UIToFPInst>(Op)) &&
1819 hasFloatFn(TLI, Ty, LibFunc_ldexp, LibFunc_ldexpf, LibFunc_ldexpl)) {
1820 if (Value *Exp = getIntToFPVal(Op, B))
1821 return emitBinaryFloatFnCall(ConstantFP::get(Ty, 1.0), Exp, TLI,
1822 LibFunc_ldexp, LibFunc_ldexpf, LibFunc_ldexpl,
1823 B, Attrs);
1824 }
1825
1826 return Ret;
1827 }
1828
optimizeFMinFMax(CallInst * CI,IRBuilderBase & B)1829 Value *LibCallSimplifier::optimizeFMinFMax(CallInst *CI, IRBuilderBase &B) {
1830 // If we can shrink the call to a float function rather than a double
1831 // function, do that first.
1832 Function *Callee = CI->getCalledFunction();
1833 StringRef Name = Callee->getName();
1834 if ((Name == "fmin" || Name == "fmax") && hasFloatVersion(Name))
1835 if (Value *Ret = optimizeBinaryDoubleFP(CI, B))
1836 return Ret;
1837
1838 // The LLVM intrinsics minnum/maxnum correspond to fmin/fmax. Canonicalize to
1839 // the intrinsics for improved optimization (for example, vectorization).
1840 // No-signed-zeros is implied by the definitions of fmax/fmin themselves.
1841 // From the C standard draft WG14/N1256:
1842 // "Ideally, fmax would be sensitive to the sign of zero, for example
1843 // fmax(-0.0, +0.0) would return +0; however, implementation in software
1844 // might be impractical."
1845 IRBuilderBase::FastMathFlagGuard Guard(B);
1846 FastMathFlags FMF = CI->getFastMathFlags();
1847 FMF.setNoSignedZeros();
1848 B.setFastMathFlags(FMF);
1849
1850 Intrinsic::ID IID = Callee->getName().startswith("fmin") ? Intrinsic::minnum
1851 : Intrinsic::maxnum;
1852 Function *F = Intrinsic::getDeclaration(CI->getModule(), IID, CI->getType());
1853 return B.CreateCall(F, { CI->getArgOperand(0), CI->getArgOperand(1) });
1854 }
1855
optimizeLog(CallInst * Log,IRBuilderBase & B)1856 Value *LibCallSimplifier::optimizeLog(CallInst *Log, IRBuilderBase &B) {
1857 Function *LogFn = Log->getCalledFunction();
1858 AttributeList Attrs; // Attributes are only meaningful on the original call
1859 StringRef LogNm = LogFn->getName();
1860 Intrinsic::ID LogID = LogFn->getIntrinsicID();
1861 Module *Mod = Log->getModule();
1862 Type *Ty = Log->getType();
1863 Value *Ret = nullptr;
1864
1865 if (UnsafeFPShrink && hasFloatVersion(LogNm))
1866 Ret = optimizeUnaryDoubleFP(Log, B, true);
1867
1868 // The earlier call must also be 'fast' in order to do these transforms.
1869 CallInst *Arg = dyn_cast<CallInst>(Log->getArgOperand(0));
1870 if (!Log->isFast() || !Arg || !Arg->isFast() || !Arg->hasOneUse())
1871 return Ret;
1872
1873 LibFunc LogLb, ExpLb, Exp2Lb, Exp10Lb, PowLb;
1874
1875 // This is only applicable to log(), log2(), log10().
1876 if (TLI->getLibFunc(LogNm, LogLb))
1877 switch (LogLb) {
1878 case LibFunc_logf:
1879 LogID = Intrinsic::log;
1880 ExpLb = LibFunc_expf;
1881 Exp2Lb = LibFunc_exp2f;
1882 Exp10Lb = LibFunc_exp10f;
1883 PowLb = LibFunc_powf;
1884 break;
1885 case LibFunc_log:
1886 LogID = Intrinsic::log;
1887 ExpLb = LibFunc_exp;
1888 Exp2Lb = LibFunc_exp2;
1889 Exp10Lb = LibFunc_exp10;
1890 PowLb = LibFunc_pow;
1891 break;
1892 case LibFunc_logl:
1893 LogID = Intrinsic::log;
1894 ExpLb = LibFunc_expl;
1895 Exp2Lb = LibFunc_exp2l;
1896 Exp10Lb = LibFunc_exp10l;
1897 PowLb = LibFunc_powl;
1898 break;
1899 case LibFunc_log2f:
1900 LogID = Intrinsic::log2;
1901 ExpLb = LibFunc_expf;
1902 Exp2Lb = LibFunc_exp2f;
1903 Exp10Lb = LibFunc_exp10f;
1904 PowLb = LibFunc_powf;
1905 break;
1906 case LibFunc_log2:
1907 LogID = Intrinsic::log2;
1908 ExpLb = LibFunc_exp;
1909 Exp2Lb = LibFunc_exp2;
1910 Exp10Lb = LibFunc_exp10;
1911 PowLb = LibFunc_pow;
1912 break;
1913 case LibFunc_log2l:
1914 LogID = Intrinsic::log2;
1915 ExpLb = LibFunc_expl;
1916 Exp2Lb = LibFunc_exp2l;
1917 Exp10Lb = LibFunc_exp10l;
1918 PowLb = LibFunc_powl;
1919 break;
1920 case LibFunc_log10f:
1921 LogID = Intrinsic::log10;
1922 ExpLb = LibFunc_expf;
1923 Exp2Lb = LibFunc_exp2f;
1924 Exp10Lb = LibFunc_exp10f;
1925 PowLb = LibFunc_powf;
1926 break;
1927 case LibFunc_log10:
1928 LogID = Intrinsic::log10;
1929 ExpLb = LibFunc_exp;
1930 Exp2Lb = LibFunc_exp2;
1931 Exp10Lb = LibFunc_exp10;
1932 PowLb = LibFunc_pow;
1933 break;
1934 case LibFunc_log10l:
1935 LogID = Intrinsic::log10;
1936 ExpLb = LibFunc_expl;
1937 Exp2Lb = LibFunc_exp2l;
1938 Exp10Lb = LibFunc_exp10l;
1939 PowLb = LibFunc_powl;
1940 break;
1941 default:
1942 return Ret;
1943 }
1944 else if (LogID == Intrinsic::log || LogID == Intrinsic::log2 ||
1945 LogID == Intrinsic::log10) {
1946 if (Ty->getScalarType()->isFloatTy()) {
1947 ExpLb = LibFunc_expf;
1948 Exp2Lb = LibFunc_exp2f;
1949 Exp10Lb = LibFunc_exp10f;
1950 PowLb = LibFunc_powf;
1951 } else if (Ty->getScalarType()->isDoubleTy()) {
1952 ExpLb = LibFunc_exp;
1953 Exp2Lb = LibFunc_exp2;
1954 Exp10Lb = LibFunc_exp10;
1955 PowLb = LibFunc_pow;
1956 } else
1957 return Ret;
1958 } else
1959 return Ret;
1960
1961 IRBuilderBase::FastMathFlagGuard Guard(B);
1962 B.setFastMathFlags(FastMathFlags::getFast());
1963
1964 Intrinsic::ID ArgID = Arg->getIntrinsicID();
1965 LibFunc ArgLb = NotLibFunc;
1966 TLI->getLibFunc(*Arg, ArgLb);
1967
1968 // log(pow(x,y)) -> y*log(x)
1969 if (ArgLb == PowLb || ArgID == Intrinsic::pow) {
1970 Value *LogX =
1971 Log->doesNotAccessMemory()
1972 ? B.CreateCall(Intrinsic::getDeclaration(Mod, LogID, Ty),
1973 Arg->getOperand(0), "log")
1974 : emitUnaryFloatFnCall(Arg->getOperand(0), LogNm, B, Attrs);
1975 Value *MulY = B.CreateFMul(Arg->getArgOperand(1), LogX, "mul");
1976 // Since pow() may have side effects, e.g. errno,
1977 // dead code elimination may not be trusted to remove it.
1978 substituteInParent(Arg, MulY);
1979 return MulY;
1980 }
1981
1982 // log(exp{,2,10}(y)) -> y*log({e,2,10})
1983 // TODO: There is no exp10() intrinsic yet.
1984 if (ArgLb == ExpLb || ArgLb == Exp2Lb || ArgLb == Exp10Lb ||
1985 ArgID == Intrinsic::exp || ArgID == Intrinsic::exp2) {
1986 Constant *Eul;
1987 if (ArgLb == ExpLb || ArgID == Intrinsic::exp)
1988 // FIXME: Add more precise value of e for long double.
1989 Eul = ConstantFP::get(Log->getType(), numbers::e);
1990 else if (ArgLb == Exp2Lb || ArgID == Intrinsic::exp2)
1991 Eul = ConstantFP::get(Log->getType(), 2.0);
1992 else
1993 Eul = ConstantFP::get(Log->getType(), 10.0);
1994 Value *LogE = Log->doesNotAccessMemory()
1995 ? B.CreateCall(Intrinsic::getDeclaration(Mod, LogID, Ty),
1996 Eul, "log")
1997 : emitUnaryFloatFnCall(Eul, LogNm, B, Attrs);
1998 Value *MulY = B.CreateFMul(Arg->getArgOperand(0), LogE, "mul");
1999 // Since exp() may have side effects, e.g. errno,
2000 // dead code elimination may not be trusted to remove it.
2001 substituteInParent(Arg, MulY);
2002 return MulY;
2003 }
2004
2005 return Ret;
2006 }
2007
optimizeSqrt(CallInst * CI,IRBuilderBase & B)2008 Value *LibCallSimplifier::optimizeSqrt(CallInst *CI, IRBuilderBase &B) {
2009 Function *Callee = CI->getCalledFunction();
2010 Value *Ret = nullptr;
2011 // TODO: Once we have a way (other than checking for the existince of the
2012 // libcall) to tell whether our target can lower @llvm.sqrt, relax the
2013 // condition below.
2014 if (TLI->has(LibFunc_sqrtf) && (Callee->getName() == "sqrt" ||
2015 Callee->getIntrinsicID() == Intrinsic::sqrt))
2016 Ret = optimizeUnaryDoubleFP(CI, B, true);
2017
2018 if (!CI->isFast())
2019 return Ret;
2020
2021 Instruction *I = dyn_cast<Instruction>(CI->getArgOperand(0));
2022 if (!I || I->getOpcode() != Instruction::FMul || !I->isFast())
2023 return Ret;
2024
2025 // We're looking for a repeated factor in a multiplication tree,
2026 // so we can do this fold: sqrt(x * x) -> fabs(x);
2027 // or this fold: sqrt((x * x) * y) -> fabs(x) * sqrt(y).
2028 Value *Op0 = I->getOperand(0);
2029 Value *Op1 = I->getOperand(1);
2030 Value *RepeatOp = nullptr;
2031 Value *OtherOp = nullptr;
2032 if (Op0 == Op1) {
2033 // Simple match: the operands of the multiply are identical.
2034 RepeatOp = Op0;
2035 } else {
2036 // Look for a more complicated pattern: one of the operands is itself
2037 // a multiply, so search for a common factor in that multiply.
2038 // Note: We don't bother looking any deeper than this first level or for
2039 // variations of this pattern because instcombine's visitFMUL and/or the
2040 // reassociation pass should give us this form.
2041 Value *OtherMul0, *OtherMul1;
2042 if (match(Op0, m_FMul(m_Value(OtherMul0), m_Value(OtherMul1)))) {
2043 // Pattern: sqrt((x * y) * z)
2044 if (OtherMul0 == OtherMul1 && cast<Instruction>(Op0)->isFast()) {
2045 // Matched: sqrt((x * x) * z)
2046 RepeatOp = OtherMul0;
2047 OtherOp = Op1;
2048 }
2049 }
2050 }
2051 if (!RepeatOp)
2052 return Ret;
2053
2054 // Fast math flags for any created instructions should match the sqrt
2055 // and multiply.
2056 IRBuilderBase::FastMathFlagGuard Guard(B);
2057 B.setFastMathFlags(I->getFastMathFlags());
2058
2059 // If we found a repeated factor, hoist it out of the square root and
2060 // replace it with the fabs of that factor.
2061 Module *M = Callee->getParent();
2062 Type *ArgType = I->getType();
2063 Function *Fabs = Intrinsic::getDeclaration(M, Intrinsic::fabs, ArgType);
2064 Value *FabsCall = B.CreateCall(Fabs, RepeatOp, "fabs");
2065 if (OtherOp) {
2066 // If we found a non-repeated factor, we still need to get its square
2067 // root. We then multiply that by the value that was simplified out
2068 // of the square root calculation.
2069 Function *Sqrt = Intrinsic::getDeclaration(M, Intrinsic::sqrt, ArgType);
2070 Value *SqrtCall = B.CreateCall(Sqrt, OtherOp, "sqrt");
2071 return B.CreateFMul(FabsCall, SqrtCall);
2072 }
2073 return FabsCall;
2074 }
2075
2076 // TODO: Generalize to handle any trig function and its inverse.
optimizeTan(CallInst * CI,IRBuilderBase & B)2077 Value *LibCallSimplifier::optimizeTan(CallInst *CI, IRBuilderBase &B) {
2078 Function *Callee = CI->getCalledFunction();
2079 Value *Ret = nullptr;
2080 StringRef Name = Callee->getName();
2081 if (UnsafeFPShrink && Name == "tan" && hasFloatVersion(Name))
2082 Ret = optimizeUnaryDoubleFP(CI, B, true);
2083
2084 Value *Op1 = CI->getArgOperand(0);
2085 auto *OpC = dyn_cast<CallInst>(Op1);
2086 if (!OpC)
2087 return Ret;
2088
2089 // Both calls must be 'fast' in order to remove them.
2090 if (!CI->isFast() || !OpC->isFast())
2091 return Ret;
2092
2093 // tan(atan(x)) -> x
2094 // tanf(atanf(x)) -> x
2095 // tanl(atanl(x)) -> x
2096 LibFunc Func;
2097 Function *F = OpC->getCalledFunction();
2098 if (F && TLI->getLibFunc(F->getName(), Func) && TLI->has(Func) &&
2099 ((Func == LibFunc_atan && Callee->getName() == "tan") ||
2100 (Func == LibFunc_atanf && Callee->getName() == "tanf") ||
2101 (Func == LibFunc_atanl && Callee->getName() == "tanl")))
2102 Ret = OpC->getArgOperand(0);
2103 return Ret;
2104 }
2105
isTrigLibCall(CallInst * CI)2106 static bool isTrigLibCall(CallInst *CI) {
2107 // We can only hope to do anything useful if we can ignore things like errno
2108 // and floating-point exceptions.
2109 // We already checked the prototype.
2110 return CI->hasFnAttr(Attribute::NoUnwind) &&
2111 CI->hasFnAttr(Attribute::ReadNone);
2112 }
2113
insertSinCosCall(IRBuilderBase & B,Function * OrigCallee,Value * Arg,bool UseFloat,Value * & Sin,Value * & Cos,Value * & SinCos)2114 static void insertSinCosCall(IRBuilderBase &B, Function *OrigCallee, Value *Arg,
2115 bool UseFloat, Value *&Sin, Value *&Cos,
2116 Value *&SinCos) {
2117 Type *ArgTy = Arg->getType();
2118 Type *ResTy;
2119 StringRef Name;
2120
2121 Triple T(OrigCallee->getParent()->getTargetTriple());
2122 if (UseFloat) {
2123 Name = "__sincospif_stret";
2124
2125 assert(T.getArch() != Triple::x86 && "x86 messy and unsupported for now");
2126 // x86_64 can't use {float, float} since that would be returned in both
2127 // xmm0 and xmm1, which isn't what a real struct would do.
2128 ResTy = T.getArch() == Triple::x86_64
2129 ? static_cast<Type *>(FixedVectorType::get(ArgTy, 2))
2130 : static_cast<Type *>(StructType::get(ArgTy, ArgTy));
2131 } else {
2132 Name = "__sincospi_stret";
2133 ResTy = StructType::get(ArgTy, ArgTy);
2134 }
2135
2136 Module *M = OrigCallee->getParent();
2137 FunctionCallee Callee =
2138 M->getOrInsertFunction(Name, OrigCallee->getAttributes(), ResTy, ArgTy);
2139
2140 if (Instruction *ArgInst = dyn_cast<Instruction>(Arg)) {
2141 // If the argument is an instruction, it must dominate all uses so put our
2142 // sincos call there.
2143 B.SetInsertPoint(ArgInst->getParent(), ++ArgInst->getIterator());
2144 } else {
2145 // Otherwise (e.g. for a constant) the beginning of the function is as
2146 // good a place as any.
2147 BasicBlock &EntryBB = B.GetInsertBlock()->getParent()->getEntryBlock();
2148 B.SetInsertPoint(&EntryBB, EntryBB.begin());
2149 }
2150
2151 SinCos = B.CreateCall(Callee, Arg, "sincospi");
2152
2153 if (SinCos->getType()->isStructTy()) {
2154 Sin = B.CreateExtractValue(SinCos, 0, "sinpi");
2155 Cos = B.CreateExtractValue(SinCos, 1, "cospi");
2156 } else {
2157 Sin = B.CreateExtractElement(SinCos, ConstantInt::get(B.getInt32Ty(), 0),
2158 "sinpi");
2159 Cos = B.CreateExtractElement(SinCos, ConstantInt::get(B.getInt32Ty(), 1),
2160 "cospi");
2161 }
2162 }
2163
optimizeSinCosPi(CallInst * CI,IRBuilderBase & B)2164 Value *LibCallSimplifier::optimizeSinCosPi(CallInst *CI, IRBuilderBase &B) {
2165 // Make sure the prototype is as expected, otherwise the rest of the
2166 // function is probably invalid and likely to abort.
2167 if (!isTrigLibCall(CI))
2168 return nullptr;
2169
2170 Value *Arg = CI->getArgOperand(0);
2171 SmallVector<CallInst *, 1> SinCalls;
2172 SmallVector<CallInst *, 1> CosCalls;
2173 SmallVector<CallInst *, 1> SinCosCalls;
2174
2175 bool IsFloat = Arg->getType()->isFloatTy();
2176
2177 // Look for all compatible sinpi, cospi and sincospi calls with the same
2178 // argument. If there are enough (in some sense) we can make the
2179 // substitution.
2180 Function *F = CI->getFunction();
2181 for (User *U : Arg->users())
2182 classifyArgUse(U, F, IsFloat, SinCalls, CosCalls, SinCosCalls);
2183
2184 // It's only worthwhile if both sinpi and cospi are actually used.
2185 if (SinCosCalls.empty() && (SinCalls.empty() || CosCalls.empty()))
2186 return nullptr;
2187
2188 Value *Sin, *Cos, *SinCos;
2189 insertSinCosCall(B, CI->getCalledFunction(), Arg, IsFloat, Sin, Cos, SinCos);
2190
2191 auto replaceTrigInsts = [this](SmallVectorImpl<CallInst *> &Calls,
2192 Value *Res) {
2193 for (CallInst *C : Calls)
2194 replaceAllUsesWith(C, Res);
2195 };
2196
2197 replaceTrigInsts(SinCalls, Sin);
2198 replaceTrigInsts(CosCalls, Cos);
2199 replaceTrigInsts(SinCosCalls, SinCos);
2200
2201 return nullptr;
2202 }
2203
classifyArgUse(Value * Val,Function * F,bool IsFloat,SmallVectorImpl<CallInst * > & SinCalls,SmallVectorImpl<CallInst * > & CosCalls,SmallVectorImpl<CallInst * > & SinCosCalls)2204 void LibCallSimplifier::classifyArgUse(
2205 Value *Val, Function *F, bool IsFloat,
2206 SmallVectorImpl<CallInst *> &SinCalls,
2207 SmallVectorImpl<CallInst *> &CosCalls,
2208 SmallVectorImpl<CallInst *> &SinCosCalls) {
2209 CallInst *CI = dyn_cast<CallInst>(Val);
2210
2211 if (!CI)
2212 return;
2213
2214 // Don't consider calls in other functions.
2215 if (CI->getFunction() != F)
2216 return;
2217
2218 Function *Callee = CI->getCalledFunction();
2219 LibFunc Func;
2220 if (!Callee || !TLI->getLibFunc(*Callee, Func) || !TLI->has(Func) ||
2221 !isTrigLibCall(CI))
2222 return;
2223
2224 if (IsFloat) {
2225 if (Func == LibFunc_sinpif)
2226 SinCalls.push_back(CI);
2227 else if (Func == LibFunc_cospif)
2228 CosCalls.push_back(CI);
2229 else if (Func == LibFunc_sincospif_stret)
2230 SinCosCalls.push_back(CI);
2231 } else {
2232 if (Func == LibFunc_sinpi)
2233 SinCalls.push_back(CI);
2234 else if (Func == LibFunc_cospi)
2235 CosCalls.push_back(CI);
2236 else if (Func == LibFunc_sincospi_stret)
2237 SinCosCalls.push_back(CI);
2238 }
2239 }
2240
2241 //===----------------------------------------------------------------------===//
2242 // Integer Library Call Optimizations
2243 //===----------------------------------------------------------------------===//
2244
optimizeFFS(CallInst * CI,IRBuilderBase & B)2245 Value *LibCallSimplifier::optimizeFFS(CallInst *CI, IRBuilderBase &B) {
2246 // ffs(x) -> x != 0 ? (i32)llvm.cttz(x)+1 : 0
2247 Value *Op = CI->getArgOperand(0);
2248 Type *ArgType = Op->getType();
2249 Function *F = Intrinsic::getDeclaration(CI->getCalledFunction()->getParent(),
2250 Intrinsic::cttz, ArgType);
2251 Value *V = B.CreateCall(F, {Op, B.getTrue()}, "cttz");
2252 V = B.CreateAdd(V, ConstantInt::get(V->getType(), 1));
2253 V = B.CreateIntCast(V, B.getInt32Ty(), false);
2254
2255 Value *Cond = B.CreateICmpNE(Op, Constant::getNullValue(ArgType));
2256 return B.CreateSelect(Cond, V, B.getInt32(0));
2257 }
2258
optimizeFls(CallInst * CI,IRBuilderBase & B)2259 Value *LibCallSimplifier::optimizeFls(CallInst *CI, IRBuilderBase &B) {
2260 // fls(x) -> (i32)(sizeInBits(x) - llvm.ctlz(x, false))
2261 Value *Op = CI->getArgOperand(0);
2262 Type *ArgType = Op->getType();
2263 Function *F = Intrinsic::getDeclaration(CI->getCalledFunction()->getParent(),
2264 Intrinsic::ctlz, ArgType);
2265 Value *V = B.CreateCall(F, {Op, B.getFalse()}, "ctlz");
2266 V = B.CreateSub(ConstantInt::get(V->getType(), ArgType->getIntegerBitWidth()),
2267 V);
2268 return B.CreateIntCast(V, CI->getType(), false);
2269 }
2270
optimizeAbs(CallInst * CI,IRBuilderBase & B)2271 Value *LibCallSimplifier::optimizeAbs(CallInst *CI, IRBuilderBase &B) {
2272 // abs(x) -> x <s 0 ? -x : x
2273 // The negation has 'nsw' because abs of INT_MIN is undefined.
2274 Value *X = CI->getArgOperand(0);
2275 Value *IsNeg = B.CreateICmpSLT(X, Constant::getNullValue(X->getType()));
2276 Value *NegX = B.CreateNSWNeg(X, "neg");
2277 return B.CreateSelect(IsNeg, NegX, X);
2278 }
2279
optimizeIsDigit(CallInst * CI,IRBuilderBase & B)2280 Value *LibCallSimplifier::optimizeIsDigit(CallInst *CI, IRBuilderBase &B) {
2281 // isdigit(c) -> (c-'0') <u 10
2282 Value *Op = CI->getArgOperand(0);
2283 Op = B.CreateSub(Op, B.getInt32('0'), "isdigittmp");
2284 Op = B.CreateICmpULT(Op, B.getInt32(10), "isdigit");
2285 return B.CreateZExt(Op, CI->getType());
2286 }
2287
optimizeIsAscii(CallInst * CI,IRBuilderBase & B)2288 Value *LibCallSimplifier::optimizeIsAscii(CallInst *CI, IRBuilderBase &B) {
2289 // isascii(c) -> c <u 128
2290 Value *Op = CI->getArgOperand(0);
2291 Op = B.CreateICmpULT(Op, B.getInt32(128), "isascii");
2292 return B.CreateZExt(Op, CI->getType());
2293 }
2294
optimizeToAscii(CallInst * CI,IRBuilderBase & B)2295 Value *LibCallSimplifier::optimizeToAscii(CallInst *CI, IRBuilderBase &B) {
2296 // toascii(c) -> c & 0x7f
2297 return B.CreateAnd(CI->getArgOperand(0),
2298 ConstantInt::get(CI->getType(), 0x7F));
2299 }
2300
optimizeAtoi(CallInst * CI,IRBuilderBase & B)2301 Value *LibCallSimplifier::optimizeAtoi(CallInst *CI, IRBuilderBase &B) {
2302 StringRef Str;
2303 if (!getConstantStringInfo(CI->getArgOperand(0), Str))
2304 return nullptr;
2305
2306 return convertStrToNumber(CI, Str, 10);
2307 }
2308
optimizeStrtol(CallInst * CI,IRBuilderBase & B)2309 Value *LibCallSimplifier::optimizeStrtol(CallInst *CI, IRBuilderBase &B) {
2310 StringRef Str;
2311 if (!getConstantStringInfo(CI->getArgOperand(0), Str))
2312 return nullptr;
2313
2314 if (!isa<ConstantPointerNull>(CI->getArgOperand(1)))
2315 return nullptr;
2316
2317 if (ConstantInt *CInt = dyn_cast<ConstantInt>(CI->getArgOperand(2))) {
2318 return convertStrToNumber(CI, Str, CInt->getSExtValue());
2319 }
2320
2321 return nullptr;
2322 }
2323
2324 //===----------------------------------------------------------------------===//
2325 // Formatting and IO Library Call Optimizations
2326 //===----------------------------------------------------------------------===//
2327
2328 static bool isReportingError(Function *Callee, CallInst *CI, int StreamArg);
2329
optimizeErrorReporting(CallInst * CI,IRBuilderBase & B,int StreamArg)2330 Value *LibCallSimplifier::optimizeErrorReporting(CallInst *CI, IRBuilderBase &B,
2331 int StreamArg) {
2332 Function *Callee = CI->getCalledFunction();
2333 // Error reporting calls should be cold, mark them as such.
2334 // This applies even to non-builtin calls: it is only a hint and applies to
2335 // functions that the frontend might not understand as builtins.
2336
2337 // This heuristic was suggested in:
2338 // Improving Static Branch Prediction in a Compiler
2339 // Brian L. Deitrich, Ben-Chung Cheng, Wen-mei W. Hwu
2340 // Proceedings of PACT'98, Oct. 1998, IEEE
2341 if (!CI->hasFnAttr(Attribute::Cold) &&
2342 isReportingError(Callee, CI, StreamArg)) {
2343 CI->addAttribute(AttributeList::FunctionIndex, Attribute::Cold);
2344 }
2345
2346 return nullptr;
2347 }
2348
isReportingError(Function * Callee,CallInst * CI,int StreamArg)2349 static bool isReportingError(Function *Callee, CallInst *CI, int StreamArg) {
2350 if (!Callee || !Callee->isDeclaration())
2351 return false;
2352
2353 if (StreamArg < 0)
2354 return true;
2355
2356 // These functions might be considered cold, but only if their stream
2357 // argument is stderr.
2358
2359 if (StreamArg >= (int)CI->getNumArgOperands())
2360 return false;
2361 LoadInst *LI = dyn_cast<LoadInst>(CI->getArgOperand(StreamArg));
2362 if (!LI)
2363 return false;
2364 GlobalVariable *GV = dyn_cast<GlobalVariable>(LI->getPointerOperand());
2365 if (!GV || !GV->isDeclaration())
2366 return false;
2367 return GV->getName() == "stderr";
2368 }
2369
optimizePrintFString(CallInst * CI,IRBuilderBase & B)2370 Value *LibCallSimplifier::optimizePrintFString(CallInst *CI, IRBuilderBase &B) {
2371 // Check for a fixed format string.
2372 StringRef FormatStr;
2373 if (!getConstantStringInfo(CI->getArgOperand(0), FormatStr))
2374 return nullptr;
2375
2376 // Empty format string -> noop.
2377 if (FormatStr.empty()) // Tolerate printf's declared void.
2378 return CI->use_empty() ? (Value *)CI : ConstantInt::get(CI->getType(), 0);
2379
2380 // Do not do any of the following transformations if the printf return value
2381 // is used, in general the printf return value is not compatible with either
2382 // putchar() or puts().
2383 if (!CI->use_empty())
2384 return nullptr;
2385
2386 // printf("x") -> putchar('x'), even for "%" and "%%".
2387 if (FormatStr.size() == 1 || FormatStr == "%%")
2388 return emitPutChar(B.getInt32(FormatStr[0]), B, TLI);
2389
2390 // printf("%s", "a") --> putchar('a')
2391 if (FormatStr == "%s" && CI->getNumArgOperands() > 1) {
2392 StringRef ChrStr;
2393 if (!getConstantStringInfo(CI->getOperand(1), ChrStr))
2394 return nullptr;
2395 if (ChrStr.size() != 1)
2396 return nullptr;
2397 return emitPutChar(B.getInt32(ChrStr[0]), B, TLI);
2398 }
2399
2400 // printf("foo\n") --> puts("foo")
2401 if (FormatStr[FormatStr.size() - 1] == '\n' &&
2402 FormatStr.find('%') == StringRef::npos) { // No format characters.
2403 // Create a string literal with no \n on it. We expect the constant merge
2404 // pass to be run after this pass, to merge duplicate strings.
2405 FormatStr = FormatStr.drop_back();
2406 Value *GV = B.CreateGlobalString(FormatStr, "str");
2407 return emitPutS(GV, B, TLI);
2408 }
2409
2410 // Optimize specific format strings.
2411 // printf("%c", chr) --> putchar(chr)
2412 if (FormatStr == "%c" && CI->getNumArgOperands() > 1 &&
2413 CI->getArgOperand(1)->getType()->isIntegerTy())
2414 return emitPutChar(CI->getArgOperand(1), B, TLI);
2415
2416 // printf("%s\n", str) --> puts(str)
2417 if (FormatStr == "%s\n" && CI->getNumArgOperands() > 1 &&
2418 CI->getArgOperand(1)->getType()->isPointerTy())
2419 return emitPutS(CI->getArgOperand(1), B, TLI);
2420 return nullptr;
2421 }
2422
optimizePrintF(CallInst * CI,IRBuilderBase & B)2423 Value *LibCallSimplifier::optimizePrintF(CallInst *CI, IRBuilderBase &B) {
2424
2425 Function *Callee = CI->getCalledFunction();
2426 FunctionType *FT = Callee->getFunctionType();
2427 if (Value *V = optimizePrintFString(CI, B)) {
2428 return V;
2429 }
2430
2431 // printf(format, ...) -> iprintf(format, ...) if no floating point
2432 // arguments.
2433 if (TLI->has(LibFunc_iprintf) && !callHasFloatingPointArgument(CI)) {
2434 Module *M = B.GetInsertBlock()->getParent()->getParent();
2435 FunctionCallee IPrintFFn =
2436 M->getOrInsertFunction("iprintf", FT, Callee->getAttributes());
2437 CallInst *New = cast<CallInst>(CI->clone());
2438 New->setCalledFunction(IPrintFFn);
2439 B.Insert(New);
2440 return New;
2441 }
2442
2443 // printf(format, ...) -> __small_printf(format, ...) if no 128-bit floating point
2444 // arguments.
2445 if (TLI->has(LibFunc_small_printf) && !callHasFP128Argument(CI)) {
2446 Module *M = B.GetInsertBlock()->getParent()->getParent();
2447 auto SmallPrintFFn =
2448 M->getOrInsertFunction(TLI->getName(LibFunc_small_printf),
2449 FT, Callee->getAttributes());
2450 CallInst *New = cast<CallInst>(CI->clone());
2451 New->setCalledFunction(SmallPrintFFn);
2452 B.Insert(New);
2453 return New;
2454 }
2455
2456 annotateNonNullBasedOnAccess(CI, 0);
2457 return nullptr;
2458 }
2459
optimizeSPrintFString(CallInst * CI,IRBuilderBase & B)2460 Value *LibCallSimplifier::optimizeSPrintFString(CallInst *CI,
2461 IRBuilderBase &B) {
2462 // Check for a fixed format string.
2463 StringRef FormatStr;
2464 if (!getConstantStringInfo(CI->getArgOperand(1), FormatStr))
2465 return nullptr;
2466
2467 // If we just have a format string (nothing else crazy) transform it.
2468 if (CI->getNumArgOperands() == 2) {
2469 // Make sure there's no % in the constant array. We could try to handle
2470 // %% -> % in the future if we cared.
2471 if (FormatStr.find('%') != StringRef::npos)
2472 return nullptr; // we found a format specifier, bail out.
2473
2474 // sprintf(str, fmt) -> llvm.memcpy(align 1 str, align 1 fmt, strlen(fmt)+1)
2475 B.CreateMemCpy(
2476 CI->getArgOperand(0), Align(1), CI->getArgOperand(1), Align(1),
2477 ConstantInt::get(DL.getIntPtrType(CI->getContext()),
2478 FormatStr.size() + 1)); // Copy the null byte.
2479 return ConstantInt::get(CI->getType(), FormatStr.size());
2480 }
2481
2482 // The remaining optimizations require the format string to be "%s" or "%c"
2483 // and have an extra operand.
2484 if (FormatStr.size() != 2 || FormatStr[0] != '%' ||
2485 CI->getNumArgOperands() < 3)
2486 return nullptr;
2487
2488 // Decode the second character of the format string.
2489 if (FormatStr[1] == 'c') {
2490 // sprintf(dst, "%c", chr) --> *(i8*)dst = chr; *((i8*)dst+1) = 0
2491 if (!CI->getArgOperand(2)->getType()->isIntegerTy())
2492 return nullptr;
2493 Value *V = B.CreateTrunc(CI->getArgOperand(2), B.getInt8Ty(), "char");
2494 Value *Ptr = castToCStr(CI->getArgOperand(0), B);
2495 B.CreateStore(V, Ptr);
2496 Ptr = B.CreateGEP(B.getInt8Ty(), Ptr, B.getInt32(1), "nul");
2497 B.CreateStore(B.getInt8(0), Ptr);
2498
2499 return ConstantInt::get(CI->getType(), 1);
2500 }
2501
2502 if (FormatStr[1] == 's') {
2503 // sprintf(dest, "%s", str) -> llvm.memcpy(align 1 dest, align 1 str,
2504 // strlen(str)+1)
2505 if (!CI->getArgOperand(2)->getType()->isPointerTy())
2506 return nullptr;
2507
2508 if (CI->use_empty())
2509 // sprintf(dest, "%s", str) -> strcpy(dest, str)
2510 return emitStrCpy(CI->getArgOperand(0), CI->getArgOperand(2), B, TLI);
2511
2512 uint64_t SrcLen = GetStringLength(CI->getArgOperand(2));
2513 if (SrcLen) {
2514 B.CreateMemCpy(
2515 CI->getArgOperand(0), Align(1), CI->getArgOperand(2), Align(1),
2516 ConstantInt::get(DL.getIntPtrType(CI->getContext()), SrcLen));
2517 // Returns total number of characters written without null-character.
2518 return ConstantInt::get(CI->getType(), SrcLen - 1);
2519 } else if (Value *V = emitStpCpy(CI->getArgOperand(0), CI->getArgOperand(2),
2520 B, TLI)) {
2521 // sprintf(dest, "%s", str) -> stpcpy(dest, str) - dest
2522 Value *PtrDiff = B.CreatePtrDiff(V, CI->getArgOperand(0));
2523 return B.CreateIntCast(PtrDiff, CI->getType(), false);
2524 }
2525
2526 bool OptForSize = CI->getFunction()->hasOptSize() ||
2527 llvm::shouldOptimizeForSize(CI->getParent(), PSI, BFI,
2528 PGSOQueryType::IRPass);
2529 if (OptForSize)
2530 return nullptr;
2531
2532 Value *Len = emitStrLen(CI->getArgOperand(2), B, DL, TLI);
2533 if (!Len)
2534 return nullptr;
2535 Value *IncLen =
2536 B.CreateAdd(Len, ConstantInt::get(Len->getType(), 1), "leninc");
2537 B.CreateMemCpy(CI->getArgOperand(0), Align(1), CI->getArgOperand(2),
2538 Align(1), IncLen);
2539
2540 // The sprintf result is the unincremented number of bytes in the string.
2541 return B.CreateIntCast(Len, CI->getType(), false);
2542 }
2543 return nullptr;
2544 }
2545
optimizeSPrintF(CallInst * CI,IRBuilderBase & B)2546 Value *LibCallSimplifier::optimizeSPrintF(CallInst *CI, IRBuilderBase &B) {
2547 Function *Callee = CI->getCalledFunction();
2548 FunctionType *FT = Callee->getFunctionType();
2549 if (Value *V = optimizeSPrintFString(CI, B)) {
2550 return V;
2551 }
2552
2553 // sprintf(str, format, ...) -> siprintf(str, format, ...) if no floating
2554 // point arguments.
2555 if (TLI->has(LibFunc_siprintf) && !callHasFloatingPointArgument(CI)) {
2556 Module *M = B.GetInsertBlock()->getParent()->getParent();
2557 FunctionCallee SIPrintFFn =
2558 M->getOrInsertFunction("siprintf", FT, Callee->getAttributes());
2559 CallInst *New = cast<CallInst>(CI->clone());
2560 New->setCalledFunction(SIPrintFFn);
2561 B.Insert(New);
2562 return New;
2563 }
2564
2565 // sprintf(str, format, ...) -> __small_sprintf(str, format, ...) if no 128-bit
2566 // floating point arguments.
2567 if (TLI->has(LibFunc_small_sprintf) && !callHasFP128Argument(CI)) {
2568 Module *M = B.GetInsertBlock()->getParent()->getParent();
2569 auto SmallSPrintFFn =
2570 M->getOrInsertFunction(TLI->getName(LibFunc_small_sprintf),
2571 FT, Callee->getAttributes());
2572 CallInst *New = cast<CallInst>(CI->clone());
2573 New->setCalledFunction(SmallSPrintFFn);
2574 B.Insert(New);
2575 return New;
2576 }
2577
2578 annotateNonNullBasedOnAccess(CI, {0, 1});
2579 return nullptr;
2580 }
2581
optimizeSnPrintFString(CallInst * CI,IRBuilderBase & B)2582 Value *LibCallSimplifier::optimizeSnPrintFString(CallInst *CI,
2583 IRBuilderBase &B) {
2584 // Check for size
2585 ConstantInt *Size = dyn_cast<ConstantInt>(CI->getArgOperand(1));
2586 if (!Size)
2587 return nullptr;
2588
2589 uint64_t N = Size->getZExtValue();
2590 // Check for a fixed format string.
2591 StringRef FormatStr;
2592 if (!getConstantStringInfo(CI->getArgOperand(2), FormatStr))
2593 return nullptr;
2594
2595 // If we just have a format string (nothing else crazy) transform it.
2596 if (CI->getNumArgOperands() == 3) {
2597 // Make sure there's no % in the constant array. We could try to handle
2598 // %% -> % in the future if we cared.
2599 if (FormatStr.find('%') != StringRef::npos)
2600 return nullptr; // we found a format specifier, bail out.
2601
2602 if (N == 0)
2603 return ConstantInt::get(CI->getType(), FormatStr.size());
2604 else if (N < FormatStr.size() + 1)
2605 return nullptr;
2606
2607 // snprintf(dst, size, fmt) -> llvm.memcpy(align 1 dst, align 1 fmt,
2608 // strlen(fmt)+1)
2609 B.CreateMemCpy(
2610 CI->getArgOperand(0), Align(1), CI->getArgOperand(2), Align(1),
2611 ConstantInt::get(DL.getIntPtrType(CI->getContext()),
2612 FormatStr.size() + 1)); // Copy the null byte.
2613 return ConstantInt::get(CI->getType(), FormatStr.size());
2614 }
2615
2616 // The remaining optimizations require the format string to be "%s" or "%c"
2617 // and have an extra operand.
2618 if (FormatStr.size() == 2 && FormatStr[0] == '%' &&
2619 CI->getNumArgOperands() == 4) {
2620
2621 // Decode the second character of the format string.
2622 if (FormatStr[1] == 'c') {
2623 if (N == 0)
2624 return ConstantInt::get(CI->getType(), 1);
2625 else if (N == 1)
2626 return nullptr;
2627
2628 // snprintf(dst, size, "%c", chr) --> *(i8*)dst = chr; *((i8*)dst+1) = 0
2629 if (!CI->getArgOperand(3)->getType()->isIntegerTy())
2630 return nullptr;
2631 Value *V = B.CreateTrunc(CI->getArgOperand(3), B.getInt8Ty(), "char");
2632 Value *Ptr = castToCStr(CI->getArgOperand(0), B);
2633 B.CreateStore(V, Ptr);
2634 Ptr = B.CreateGEP(B.getInt8Ty(), Ptr, B.getInt32(1), "nul");
2635 B.CreateStore(B.getInt8(0), Ptr);
2636
2637 return ConstantInt::get(CI->getType(), 1);
2638 }
2639
2640 if (FormatStr[1] == 's') {
2641 // snprintf(dest, size, "%s", str) to llvm.memcpy(dest, str, len+1, 1)
2642 StringRef Str;
2643 if (!getConstantStringInfo(CI->getArgOperand(3), Str))
2644 return nullptr;
2645
2646 if (N == 0)
2647 return ConstantInt::get(CI->getType(), Str.size());
2648 else if (N < Str.size() + 1)
2649 return nullptr;
2650
2651 B.CreateMemCpy(CI->getArgOperand(0), Align(1), CI->getArgOperand(3),
2652 Align(1), ConstantInt::get(CI->getType(), Str.size() + 1));
2653
2654 // The snprintf result is the unincremented number of bytes in the string.
2655 return ConstantInt::get(CI->getType(), Str.size());
2656 }
2657 }
2658 return nullptr;
2659 }
2660
optimizeSnPrintF(CallInst * CI,IRBuilderBase & B)2661 Value *LibCallSimplifier::optimizeSnPrintF(CallInst *CI, IRBuilderBase &B) {
2662 if (Value *V = optimizeSnPrintFString(CI, B)) {
2663 return V;
2664 }
2665
2666 if (isKnownNonZero(CI->getOperand(1), DL))
2667 annotateNonNullBasedOnAccess(CI, 0);
2668 return nullptr;
2669 }
2670
optimizeFPrintFString(CallInst * CI,IRBuilderBase & B)2671 Value *LibCallSimplifier::optimizeFPrintFString(CallInst *CI,
2672 IRBuilderBase &B) {
2673 optimizeErrorReporting(CI, B, 0);
2674
2675 // All the optimizations depend on the format string.
2676 StringRef FormatStr;
2677 if (!getConstantStringInfo(CI->getArgOperand(1), FormatStr))
2678 return nullptr;
2679
2680 // Do not do any of the following transformations if the fprintf return
2681 // value is used, in general the fprintf return value is not compatible
2682 // with fwrite(), fputc() or fputs().
2683 if (!CI->use_empty())
2684 return nullptr;
2685
2686 // fprintf(F, "foo") --> fwrite("foo", 3, 1, F)
2687 if (CI->getNumArgOperands() == 2) {
2688 // Could handle %% -> % if we cared.
2689 if (FormatStr.find('%') != StringRef::npos)
2690 return nullptr; // We found a format specifier.
2691
2692 return emitFWrite(
2693 CI->getArgOperand(1),
2694 ConstantInt::get(DL.getIntPtrType(CI->getContext()), FormatStr.size()),
2695 CI->getArgOperand(0), B, DL, TLI);
2696 }
2697
2698 // The remaining optimizations require the format string to be "%s" or "%c"
2699 // and have an extra operand.
2700 if (FormatStr.size() != 2 || FormatStr[0] != '%' ||
2701 CI->getNumArgOperands() < 3)
2702 return nullptr;
2703
2704 // Decode the second character of the format string.
2705 if (FormatStr[1] == 'c') {
2706 // fprintf(F, "%c", chr) --> fputc(chr, F)
2707 if (!CI->getArgOperand(2)->getType()->isIntegerTy())
2708 return nullptr;
2709 return emitFPutC(CI->getArgOperand(2), CI->getArgOperand(0), B, TLI);
2710 }
2711
2712 if (FormatStr[1] == 's') {
2713 // fprintf(F, "%s", str) --> fputs(str, F)
2714 if (!CI->getArgOperand(2)->getType()->isPointerTy())
2715 return nullptr;
2716 return emitFPutS(CI->getArgOperand(2), CI->getArgOperand(0), B, TLI);
2717 }
2718 return nullptr;
2719 }
2720
optimizeFPrintF(CallInst * CI,IRBuilderBase & B)2721 Value *LibCallSimplifier::optimizeFPrintF(CallInst *CI, IRBuilderBase &B) {
2722 Function *Callee = CI->getCalledFunction();
2723 FunctionType *FT = Callee->getFunctionType();
2724 if (Value *V = optimizeFPrintFString(CI, B)) {
2725 return V;
2726 }
2727
2728 // fprintf(stream, format, ...) -> fiprintf(stream, format, ...) if no
2729 // floating point arguments.
2730 if (TLI->has(LibFunc_fiprintf) && !callHasFloatingPointArgument(CI)) {
2731 Module *M = B.GetInsertBlock()->getParent()->getParent();
2732 FunctionCallee FIPrintFFn =
2733 M->getOrInsertFunction("fiprintf", FT, Callee->getAttributes());
2734 CallInst *New = cast<CallInst>(CI->clone());
2735 New->setCalledFunction(FIPrintFFn);
2736 B.Insert(New);
2737 return New;
2738 }
2739
2740 // fprintf(stream, format, ...) -> __small_fprintf(stream, format, ...) if no
2741 // 128-bit floating point arguments.
2742 if (TLI->has(LibFunc_small_fprintf) && !callHasFP128Argument(CI)) {
2743 Module *M = B.GetInsertBlock()->getParent()->getParent();
2744 auto SmallFPrintFFn =
2745 M->getOrInsertFunction(TLI->getName(LibFunc_small_fprintf),
2746 FT, Callee->getAttributes());
2747 CallInst *New = cast<CallInst>(CI->clone());
2748 New->setCalledFunction(SmallFPrintFFn);
2749 B.Insert(New);
2750 return New;
2751 }
2752
2753 return nullptr;
2754 }
2755
optimizeFWrite(CallInst * CI,IRBuilderBase & B)2756 Value *LibCallSimplifier::optimizeFWrite(CallInst *CI, IRBuilderBase &B) {
2757 optimizeErrorReporting(CI, B, 3);
2758
2759 // Get the element size and count.
2760 ConstantInt *SizeC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
2761 ConstantInt *CountC = dyn_cast<ConstantInt>(CI->getArgOperand(2));
2762 if (SizeC && CountC) {
2763 uint64_t Bytes = SizeC->getZExtValue() * CountC->getZExtValue();
2764
2765 // If this is writing zero records, remove the call (it's a noop).
2766 if (Bytes == 0)
2767 return ConstantInt::get(CI->getType(), 0);
2768
2769 // If this is writing one byte, turn it into fputc.
2770 // This optimisation is only valid, if the return value is unused.
2771 if (Bytes == 1 && CI->use_empty()) { // fwrite(S,1,1,F) -> fputc(S[0],F)
2772 Value *Char = B.CreateLoad(B.getInt8Ty(),
2773 castToCStr(CI->getArgOperand(0), B), "char");
2774 Value *NewCI = emitFPutC(Char, CI->getArgOperand(3), B, TLI);
2775 return NewCI ? ConstantInt::get(CI->getType(), 1) : nullptr;
2776 }
2777 }
2778
2779 return nullptr;
2780 }
2781
optimizeFPuts(CallInst * CI,IRBuilderBase & B)2782 Value *LibCallSimplifier::optimizeFPuts(CallInst *CI, IRBuilderBase &B) {
2783 optimizeErrorReporting(CI, B, 1);
2784
2785 // Don't rewrite fputs to fwrite when optimising for size because fwrite
2786 // requires more arguments and thus extra MOVs are required.
2787 bool OptForSize = CI->getFunction()->hasOptSize() ||
2788 llvm::shouldOptimizeForSize(CI->getParent(), PSI, BFI,
2789 PGSOQueryType::IRPass);
2790 if (OptForSize)
2791 return nullptr;
2792
2793 // We can't optimize if return value is used.
2794 if (!CI->use_empty())
2795 return nullptr;
2796
2797 // fputs(s,F) --> fwrite(s,strlen(s),1,F)
2798 uint64_t Len = GetStringLength(CI->getArgOperand(0));
2799 if (!Len)
2800 return nullptr;
2801
2802 // Known to have no uses (see above).
2803 return emitFWrite(
2804 CI->getArgOperand(0),
2805 ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len - 1),
2806 CI->getArgOperand(1), B, DL, TLI);
2807 }
2808
optimizePuts(CallInst * CI,IRBuilderBase & B)2809 Value *LibCallSimplifier::optimizePuts(CallInst *CI, IRBuilderBase &B) {
2810 annotateNonNullBasedOnAccess(CI, 0);
2811 if (!CI->use_empty())
2812 return nullptr;
2813
2814 // Check for a constant string.
2815 // puts("") -> putchar('\n')
2816 StringRef Str;
2817 if (getConstantStringInfo(CI->getArgOperand(0), Str) && Str.empty())
2818 return emitPutChar(B.getInt32('\n'), B, TLI);
2819
2820 return nullptr;
2821 }
2822
optimizeBCopy(CallInst * CI,IRBuilderBase & B)2823 Value *LibCallSimplifier::optimizeBCopy(CallInst *CI, IRBuilderBase &B) {
2824 // bcopy(src, dst, n) -> llvm.memmove(dst, src, n)
2825 return B.CreateMemMove(CI->getArgOperand(1), Align(1), CI->getArgOperand(0),
2826 Align(1), CI->getArgOperand(2));
2827 }
2828
hasFloatVersion(StringRef FuncName)2829 bool LibCallSimplifier::hasFloatVersion(StringRef FuncName) {
2830 LibFunc Func;
2831 SmallString<20> FloatFuncName = FuncName;
2832 FloatFuncName += 'f';
2833 if (TLI->getLibFunc(FloatFuncName, Func))
2834 return TLI->has(Func);
2835 return false;
2836 }
2837
optimizeStringMemoryLibCall(CallInst * CI,IRBuilderBase & Builder)2838 Value *LibCallSimplifier::optimizeStringMemoryLibCall(CallInst *CI,
2839 IRBuilderBase &Builder) {
2840 LibFunc Func;
2841 Function *Callee = CI->getCalledFunction();
2842 // Check for string/memory library functions.
2843 if (TLI->getLibFunc(*Callee, Func) && TLI->has(Func)) {
2844 // Make sure we never change the calling convention.
2845 assert((ignoreCallingConv(Func) ||
2846 isCallingConvCCompatible(CI)) &&
2847 "Optimizing string/memory libcall would change the calling convention");
2848 switch (Func) {
2849 case LibFunc_strcat:
2850 return optimizeStrCat(CI, Builder);
2851 case LibFunc_strncat:
2852 return optimizeStrNCat(CI, Builder);
2853 case LibFunc_strchr:
2854 return optimizeStrChr(CI, Builder);
2855 case LibFunc_strrchr:
2856 return optimizeStrRChr(CI, Builder);
2857 case LibFunc_strcmp:
2858 return optimizeStrCmp(CI, Builder);
2859 case LibFunc_strncmp:
2860 return optimizeStrNCmp(CI, Builder);
2861 case LibFunc_strcpy:
2862 return optimizeStrCpy(CI, Builder);
2863 case LibFunc_stpcpy:
2864 return optimizeStpCpy(CI, Builder);
2865 case LibFunc_strncpy:
2866 return optimizeStrNCpy(CI, Builder);
2867 case LibFunc_strlen:
2868 return optimizeStrLen(CI, Builder);
2869 case LibFunc_strpbrk:
2870 return optimizeStrPBrk(CI, Builder);
2871 case LibFunc_strndup:
2872 return optimizeStrNDup(CI, Builder);
2873 case LibFunc_strtol:
2874 case LibFunc_strtod:
2875 case LibFunc_strtof:
2876 case LibFunc_strtoul:
2877 case LibFunc_strtoll:
2878 case LibFunc_strtold:
2879 case LibFunc_strtoull:
2880 return optimizeStrTo(CI, Builder);
2881 case LibFunc_strspn:
2882 return optimizeStrSpn(CI, Builder);
2883 case LibFunc_strcspn:
2884 return optimizeStrCSpn(CI, Builder);
2885 case LibFunc_strstr:
2886 return optimizeStrStr(CI, Builder);
2887 case LibFunc_memchr:
2888 return optimizeMemChr(CI, Builder);
2889 case LibFunc_memrchr:
2890 return optimizeMemRChr(CI, Builder);
2891 case LibFunc_bcmp:
2892 return optimizeBCmp(CI, Builder);
2893 case LibFunc_memcmp:
2894 return optimizeMemCmp(CI, Builder);
2895 case LibFunc_memcpy:
2896 return optimizeMemCpy(CI, Builder);
2897 case LibFunc_memccpy:
2898 return optimizeMemCCpy(CI, Builder);
2899 case LibFunc_mempcpy:
2900 return optimizeMemPCpy(CI, Builder);
2901 case LibFunc_memmove:
2902 return optimizeMemMove(CI, Builder);
2903 case LibFunc_memset:
2904 return optimizeMemSet(CI, Builder);
2905 case LibFunc_realloc:
2906 return optimizeRealloc(CI, Builder);
2907 case LibFunc_wcslen:
2908 return optimizeWcslen(CI, Builder);
2909 case LibFunc_bcopy:
2910 return optimizeBCopy(CI, Builder);
2911 default:
2912 break;
2913 }
2914 }
2915 return nullptr;
2916 }
2917
optimizeFloatingPointLibCall(CallInst * CI,LibFunc Func,IRBuilderBase & Builder)2918 Value *LibCallSimplifier::optimizeFloatingPointLibCall(CallInst *CI,
2919 LibFunc Func,
2920 IRBuilderBase &Builder) {
2921 // Don't optimize calls that require strict floating point semantics.
2922 if (CI->isStrictFP())
2923 return nullptr;
2924
2925 if (Value *V = optimizeTrigReflections(CI, Func, Builder))
2926 return V;
2927
2928 switch (Func) {
2929 case LibFunc_sinpif:
2930 case LibFunc_sinpi:
2931 case LibFunc_cospif:
2932 case LibFunc_cospi:
2933 return optimizeSinCosPi(CI, Builder);
2934 case LibFunc_powf:
2935 case LibFunc_pow:
2936 case LibFunc_powl:
2937 return optimizePow(CI, Builder);
2938 case LibFunc_exp2l:
2939 case LibFunc_exp2:
2940 case LibFunc_exp2f:
2941 return optimizeExp2(CI, Builder);
2942 case LibFunc_fabsf:
2943 case LibFunc_fabs:
2944 case LibFunc_fabsl:
2945 return replaceUnaryCall(CI, Builder, Intrinsic::fabs);
2946 case LibFunc_sqrtf:
2947 case LibFunc_sqrt:
2948 case LibFunc_sqrtl:
2949 return optimizeSqrt(CI, Builder);
2950 case LibFunc_logf:
2951 case LibFunc_log:
2952 case LibFunc_logl:
2953 case LibFunc_log10f:
2954 case LibFunc_log10:
2955 case LibFunc_log10l:
2956 case LibFunc_log1pf:
2957 case LibFunc_log1p:
2958 case LibFunc_log1pl:
2959 case LibFunc_log2f:
2960 case LibFunc_log2:
2961 case LibFunc_log2l:
2962 case LibFunc_logbf:
2963 case LibFunc_logb:
2964 case LibFunc_logbl:
2965 return optimizeLog(CI, Builder);
2966 case LibFunc_tan:
2967 case LibFunc_tanf:
2968 case LibFunc_tanl:
2969 return optimizeTan(CI, Builder);
2970 case LibFunc_ceil:
2971 return replaceUnaryCall(CI, Builder, Intrinsic::ceil);
2972 case LibFunc_floor:
2973 return replaceUnaryCall(CI, Builder, Intrinsic::floor);
2974 case LibFunc_round:
2975 return replaceUnaryCall(CI, Builder, Intrinsic::round);
2976 case LibFunc_roundeven:
2977 return replaceUnaryCall(CI, Builder, Intrinsic::roundeven);
2978 case LibFunc_nearbyint:
2979 return replaceUnaryCall(CI, Builder, Intrinsic::nearbyint);
2980 case LibFunc_rint:
2981 return replaceUnaryCall(CI, Builder, Intrinsic::rint);
2982 case LibFunc_trunc:
2983 return replaceUnaryCall(CI, Builder, Intrinsic::trunc);
2984 case LibFunc_acos:
2985 case LibFunc_acosh:
2986 case LibFunc_asin:
2987 case LibFunc_asinh:
2988 case LibFunc_atan:
2989 case LibFunc_atanh:
2990 case LibFunc_cbrt:
2991 case LibFunc_cosh:
2992 case LibFunc_exp:
2993 case LibFunc_exp10:
2994 case LibFunc_expm1:
2995 case LibFunc_cos:
2996 case LibFunc_sin:
2997 case LibFunc_sinh:
2998 case LibFunc_tanh:
2999 if (UnsafeFPShrink && hasFloatVersion(CI->getCalledFunction()->getName()))
3000 return optimizeUnaryDoubleFP(CI, Builder, true);
3001 return nullptr;
3002 case LibFunc_copysign:
3003 if (hasFloatVersion(CI->getCalledFunction()->getName()))
3004 return optimizeBinaryDoubleFP(CI, Builder);
3005 return nullptr;
3006 case LibFunc_fminf:
3007 case LibFunc_fmin:
3008 case LibFunc_fminl:
3009 case LibFunc_fmaxf:
3010 case LibFunc_fmax:
3011 case LibFunc_fmaxl:
3012 return optimizeFMinFMax(CI, Builder);
3013 case LibFunc_cabs:
3014 case LibFunc_cabsf:
3015 case LibFunc_cabsl:
3016 return optimizeCAbs(CI, Builder);
3017 default:
3018 return nullptr;
3019 }
3020 }
3021
optimizeCall(CallInst * CI,IRBuilderBase & Builder)3022 Value *LibCallSimplifier::optimizeCall(CallInst *CI, IRBuilderBase &Builder) {
3023 // TODO: Split out the code below that operates on FP calls so that
3024 // we can all non-FP calls with the StrictFP attribute to be
3025 // optimized.
3026 if (CI->isNoBuiltin())
3027 return nullptr;
3028
3029 LibFunc Func;
3030 Function *Callee = CI->getCalledFunction();
3031 bool isCallingConvC = isCallingConvCCompatible(CI);
3032
3033 SmallVector<OperandBundleDef, 2> OpBundles;
3034 CI->getOperandBundlesAsDefs(OpBundles);
3035
3036 IRBuilderBase::OperandBundlesGuard Guard(Builder);
3037 Builder.setDefaultOperandBundles(OpBundles);
3038
3039 // Command-line parameter overrides instruction attribute.
3040 // This can't be moved to optimizeFloatingPointLibCall() because it may be
3041 // used by the intrinsic optimizations.
3042 if (EnableUnsafeFPShrink.getNumOccurrences() > 0)
3043 UnsafeFPShrink = EnableUnsafeFPShrink;
3044 else if (isa<FPMathOperator>(CI) && CI->isFast())
3045 UnsafeFPShrink = true;
3046
3047 // First, check for intrinsics.
3048 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
3049 if (!isCallingConvC)
3050 return nullptr;
3051 // The FP intrinsics have corresponding constrained versions so we don't
3052 // need to check for the StrictFP attribute here.
3053 switch (II->getIntrinsicID()) {
3054 case Intrinsic::pow:
3055 return optimizePow(CI, Builder);
3056 case Intrinsic::exp2:
3057 return optimizeExp2(CI, Builder);
3058 case Intrinsic::log:
3059 case Intrinsic::log2:
3060 case Intrinsic::log10:
3061 return optimizeLog(CI, Builder);
3062 case Intrinsic::sqrt:
3063 return optimizeSqrt(CI, Builder);
3064 // TODO: Use foldMallocMemset() with memset intrinsic.
3065 case Intrinsic::memset:
3066 return optimizeMemSet(CI, Builder);
3067 case Intrinsic::memcpy:
3068 return optimizeMemCpy(CI, Builder);
3069 case Intrinsic::memmove:
3070 return optimizeMemMove(CI, Builder);
3071 default:
3072 return nullptr;
3073 }
3074 }
3075
3076 // Also try to simplify calls to fortified library functions.
3077 if (Value *SimplifiedFortifiedCI =
3078 FortifiedSimplifier.optimizeCall(CI, Builder)) {
3079 // Try to further simplify the result.
3080 CallInst *SimplifiedCI = dyn_cast<CallInst>(SimplifiedFortifiedCI);
3081 if (SimplifiedCI && SimplifiedCI->getCalledFunction()) {
3082 // Ensure that SimplifiedCI's uses are complete, since some calls have
3083 // their uses analyzed.
3084 replaceAllUsesWith(CI, SimplifiedCI);
3085
3086 // Set insertion point to SimplifiedCI to guarantee we reach all uses
3087 // we might replace later on.
3088 IRBuilderBase::InsertPointGuard Guard(Builder);
3089 Builder.SetInsertPoint(SimplifiedCI);
3090 if (Value *V = optimizeStringMemoryLibCall(SimplifiedCI, Builder)) {
3091 // If we were able to further simplify, remove the now redundant call.
3092 substituteInParent(SimplifiedCI, V);
3093 return V;
3094 }
3095 }
3096 return SimplifiedFortifiedCI;
3097 }
3098
3099 // Then check for known library functions.
3100 if (TLI->getLibFunc(*Callee, Func) && TLI->has(Func)) {
3101 // We never change the calling convention.
3102 if (!ignoreCallingConv(Func) && !isCallingConvC)
3103 return nullptr;
3104 if (Value *V = optimizeStringMemoryLibCall(CI, Builder))
3105 return V;
3106 if (Value *V = optimizeFloatingPointLibCall(CI, Func, Builder))
3107 return V;
3108 switch (Func) {
3109 case LibFunc_ffs:
3110 case LibFunc_ffsl:
3111 case LibFunc_ffsll:
3112 return optimizeFFS(CI, Builder);
3113 case LibFunc_fls:
3114 case LibFunc_flsl:
3115 case LibFunc_flsll:
3116 return optimizeFls(CI, Builder);
3117 case LibFunc_abs:
3118 case LibFunc_labs:
3119 case LibFunc_llabs:
3120 return optimizeAbs(CI, Builder);
3121 case LibFunc_isdigit:
3122 return optimizeIsDigit(CI, Builder);
3123 case LibFunc_isascii:
3124 return optimizeIsAscii(CI, Builder);
3125 case LibFunc_toascii:
3126 return optimizeToAscii(CI, Builder);
3127 case LibFunc_atoi:
3128 case LibFunc_atol:
3129 case LibFunc_atoll:
3130 return optimizeAtoi(CI, Builder);
3131 case LibFunc_strtol:
3132 case LibFunc_strtoll:
3133 return optimizeStrtol(CI, Builder);
3134 case LibFunc_printf:
3135 return optimizePrintF(CI, Builder);
3136 case LibFunc_sprintf:
3137 return optimizeSPrintF(CI, Builder);
3138 case LibFunc_snprintf:
3139 return optimizeSnPrintF(CI, Builder);
3140 case LibFunc_fprintf:
3141 return optimizeFPrintF(CI, Builder);
3142 case LibFunc_fwrite:
3143 return optimizeFWrite(CI, Builder);
3144 case LibFunc_fputs:
3145 return optimizeFPuts(CI, Builder);
3146 case LibFunc_puts:
3147 return optimizePuts(CI, Builder);
3148 case LibFunc_perror:
3149 return optimizeErrorReporting(CI, Builder);
3150 case LibFunc_vfprintf:
3151 case LibFunc_fiprintf:
3152 return optimizeErrorReporting(CI, Builder, 0);
3153 default:
3154 return nullptr;
3155 }
3156 }
3157 return nullptr;
3158 }
3159
LibCallSimplifier(const DataLayout & DL,const TargetLibraryInfo * TLI,OptimizationRemarkEmitter & ORE,BlockFrequencyInfo * BFI,ProfileSummaryInfo * PSI,function_ref<void (Instruction *,Value *)> Replacer,function_ref<void (Instruction *)> Eraser)3160 LibCallSimplifier::LibCallSimplifier(
3161 const DataLayout &DL, const TargetLibraryInfo *TLI,
3162 OptimizationRemarkEmitter &ORE,
3163 BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI,
3164 function_ref<void(Instruction *, Value *)> Replacer,
3165 function_ref<void(Instruction *)> Eraser)
3166 : FortifiedSimplifier(TLI), DL(DL), TLI(TLI), ORE(ORE), BFI(BFI), PSI(PSI),
3167 UnsafeFPShrink(false), Replacer(Replacer), Eraser(Eraser) {}
3168
replaceAllUsesWith(Instruction * I,Value * With)3169 void LibCallSimplifier::replaceAllUsesWith(Instruction *I, Value *With) {
3170 // Indirect through the replacer used in this instance.
3171 Replacer(I, With);
3172 }
3173
eraseFromParent(Instruction * I)3174 void LibCallSimplifier::eraseFromParent(Instruction *I) {
3175 Eraser(I);
3176 }
3177
3178 // TODO:
3179 // Additional cases that we need to add to this file:
3180 //
3181 // cbrt:
3182 // * cbrt(expN(X)) -> expN(x/3)
3183 // * cbrt(sqrt(x)) -> pow(x,1/6)
3184 // * cbrt(cbrt(x)) -> pow(x,1/9)
3185 //
3186 // exp, expf, expl:
3187 // * exp(log(x)) -> x
3188 //
3189 // log, logf, logl:
3190 // * log(exp(x)) -> x
3191 // * log(exp(y)) -> y*log(e)
3192 // * log(exp10(y)) -> y*log(10)
3193 // * log(sqrt(x)) -> 0.5*log(x)
3194 //
3195 // pow, powf, powl:
3196 // * pow(sqrt(x),y) -> pow(x,y*0.5)
3197 // * pow(pow(x,y),z)-> pow(x,y*z)
3198 //
3199 // signbit:
3200 // * signbit(cnst) -> cnst'
3201 // * signbit(nncst) -> 0 (if pstv is a non-negative constant)
3202 //
3203 // sqrt, sqrtf, sqrtl:
3204 // * sqrt(expN(x)) -> expN(x*0.5)
3205 // * sqrt(Nroot(x)) -> pow(x,1/(2*N))
3206 // * sqrt(pow(x,y)) -> pow(|x|,y*0.5)
3207 //
3208
3209 //===----------------------------------------------------------------------===//
3210 // Fortified Library Call Optimizations
3211 //===----------------------------------------------------------------------===//
3212
3213 bool
isFortifiedCallFoldable(CallInst * CI,unsigned ObjSizeOp,Optional<unsigned> SizeOp,Optional<unsigned> StrOp,Optional<unsigned> FlagOp)3214 FortifiedLibCallSimplifier::isFortifiedCallFoldable(CallInst *CI,
3215 unsigned ObjSizeOp,
3216 Optional<unsigned> SizeOp,
3217 Optional<unsigned> StrOp,
3218 Optional<unsigned> FlagOp) {
3219 // If this function takes a flag argument, the implementation may use it to
3220 // perform extra checks. Don't fold into the non-checking variant.
3221 if (FlagOp) {
3222 ConstantInt *Flag = dyn_cast<ConstantInt>(CI->getArgOperand(*FlagOp));
3223 if (!Flag || !Flag->isZero())
3224 return false;
3225 }
3226
3227 if (SizeOp && CI->getArgOperand(ObjSizeOp) == CI->getArgOperand(*SizeOp))
3228 return true;
3229
3230 if (ConstantInt *ObjSizeCI =
3231 dyn_cast<ConstantInt>(CI->getArgOperand(ObjSizeOp))) {
3232 if (ObjSizeCI->isMinusOne())
3233 return true;
3234 // If the object size wasn't -1 (unknown), bail out if we were asked to.
3235 if (OnlyLowerUnknownSize)
3236 return false;
3237 if (StrOp) {
3238 uint64_t Len = GetStringLength(CI->getArgOperand(*StrOp));
3239 // If the length is 0 we don't know how long it is and so we can't
3240 // remove the check.
3241 if (Len)
3242 annotateDereferenceableBytes(CI, *StrOp, Len);
3243 else
3244 return false;
3245 return ObjSizeCI->getZExtValue() >= Len;
3246 }
3247
3248 if (SizeOp) {
3249 if (ConstantInt *SizeCI =
3250 dyn_cast<ConstantInt>(CI->getArgOperand(*SizeOp)))
3251 return ObjSizeCI->getZExtValue() >= SizeCI->getZExtValue();
3252 }
3253 }
3254 return false;
3255 }
3256
optimizeMemCpyChk(CallInst * CI,IRBuilderBase & B)3257 Value *FortifiedLibCallSimplifier::optimizeMemCpyChk(CallInst *CI,
3258 IRBuilderBase &B) {
3259 if (isFortifiedCallFoldable(CI, 3, 2)) {
3260 CallInst *NewCI =
3261 B.CreateMemCpy(CI->getArgOperand(0), Align(1), CI->getArgOperand(1),
3262 Align(1), CI->getArgOperand(2));
3263 NewCI->setAttributes(CI->getAttributes());
3264 return CI->getArgOperand(0);
3265 }
3266 return nullptr;
3267 }
3268
optimizeMemMoveChk(CallInst * CI,IRBuilderBase & B)3269 Value *FortifiedLibCallSimplifier::optimizeMemMoveChk(CallInst *CI,
3270 IRBuilderBase &B) {
3271 if (isFortifiedCallFoldable(CI, 3, 2)) {
3272 CallInst *NewCI =
3273 B.CreateMemMove(CI->getArgOperand(0), Align(1), CI->getArgOperand(1),
3274 Align(1), CI->getArgOperand(2));
3275 NewCI->setAttributes(CI->getAttributes());
3276 return CI->getArgOperand(0);
3277 }
3278 return nullptr;
3279 }
3280
optimizeMemSetChk(CallInst * CI,IRBuilderBase & B)3281 Value *FortifiedLibCallSimplifier::optimizeMemSetChk(CallInst *CI,
3282 IRBuilderBase &B) {
3283 // TODO: Try foldMallocMemset() here.
3284
3285 if (isFortifiedCallFoldable(CI, 3, 2)) {
3286 Value *Val = B.CreateIntCast(CI->getArgOperand(1), B.getInt8Ty(), false);
3287 CallInst *NewCI = B.CreateMemSet(CI->getArgOperand(0), Val,
3288 CI->getArgOperand(2), Align(1));
3289 NewCI->setAttributes(CI->getAttributes());
3290 return CI->getArgOperand(0);
3291 }
3292 return nullptr;
3293 }
3294
optimizeMemPCpyChk(CallInst * CI,IRBuilderBase & B)3295 Value *FortifiedLibCallSimplifier::optimizeMemPCpyChk(CallInst *CI,
3296 IRBuilderBase &B) {
3297 const DataLayout &DL = CI->getModule()->getDataLayout();
3298 if (isFortifiedCallFoldable(CI, 3, 2))
3299 if (Value *Call = emitMemPCpy(CI->getArgOperand(0), CI->getArgOperand(1),
3300 CI->getArgOperand(2), B, DL, TLI)) {
3301 CallInst *NewCI = cast<CallInst>(Call);
3302 NewCI->setAttributes(CI->getAttributes());
3303 return NewCI;
3304 }
3305 return nullptr;
3306 }
3307
optimizeStrpCpyChk(CallInst * CI,IRBuilderBase & B,LibFunc Func)3308 Value *FortifiedLibCallSimplifier::optimizeStrpCpyChk(CallInst *CI,
3309 IRBuilderBase &B,
3310 LibFunc Func) {
3311 const DataLayout &DL = CI->getModule()->getDataLayout();
3312 Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1),
3313 *ObjSize = CI->getArgOperand(2);
3314
3315 // __stpcpy_chk(x,x,...) -> x+strlen(x)
3316 if (Func == LibFunc_stpcpy_chk && !OnlyLowerUnknownSize && Dst == Src) {
3317 Value *StrLen = emitStrLen(Src, B, DL, TLI);
3318 return StrLen ? B.CreateInBoundsGEP(B.getInt8Ty(), Dst, StrLen) : nullptr;
3319 }
3320
3321 // If a) we don't have any length information, or b) we know this will
3322 // fit then just lower to a plain st[rp]cpy. Otherwise we'll keep our
3323 // st[rp]cpy_chk call which may fail at runtime if the size is too long.
3324 // TODO: It might be nice to get a maximum length out of the possible
3325 // string lengths for varying.
3326 if (isFortifiedCallFoldable(CI, 2, None, 1)) {
3327 if (Func == LibFunc_strcpy_chk)
3328 return emitStrCpy(Dst, Src, B, TLI);
3329 else
3330 return emitStpCpy(Dst, Src, B, TLI);
3331 }
3332
3333 if (OnlyLowerUnknownSize)
3334 return nullptr;
3335
3336 // Maybe we can stil fold __st[rp]cpy_chk to __memcpy_chk.
3337 uint64_t Len = GetStringLength(Src);
3338 if (Len)
3339 annotateDereferenceableBytes(CI, 1, Len);
3340 else
3341 return nullptr;
3342
3343 Type *SizeTTy = DL.getIntPtrType(CI->getContext());
3344 Value *LenV = ConstantInt::get(SizeTTy, Len);
3345 Value *Ret = emitMemCpyChk(Dst, Src, LenV, ObjSize, B, DL, TLI);
3346 // If the function was an __stpcpy_chk, and we were able to fold it into
3347 // a __memcpy_chk, we still need to return the correct end pointer.
3348 if (Ret && Func == LibFunc_stpcpy_chk)
3349 return B.CreateGEP(B.getInt8Ty(), Dst, ConstantInt::get(SizeTTy, Len - 1));
3350 return Ret;
3351 }
3352
optimizeStrLenChk(CallInst * CI,IRBuilderBase & B)3353 Value *FortifiedLibCallSimplifier::optimizeStrLenChk(CallInst *CI,
3354 IRBuilderBase &B) {
3355 if (isFortifiedCallFoldable(CI, 1, None, 0))
3356 return emitStrLen(CI->getArgOperand(0), B, CI->getModule()->getDataLayout(),
3357 TLI);
3358 return nullptr;
3359 }
3360
optimizeStrpNCpyChk(CallInst * CI,IRBuilderBase & B,LibFunc Func)3361 Value *FortifiedLibCallSimplifier::optimizeStrpNCpyChk(CallInst *CI,
3362 IRBuilderBase &B,
3363 LibFunc Func) {
3364 if (isFortifiedCallFoldable(CI, 3, 2)) {
3365 if (Func == LibFunc_strncpy_chk)
3366 return emitStrNCpy(CI->getArgOperand(0), CI->getArgOperand(1),
3367 CI->getArgOperand(2), B, TLI);
3368 else
3369 return emitStpNCpy(CI->getArgOperand(0), CI->getArgOperand(1),
3370 CI->getArgOperand(2), B, TLI);
3371 }
3372
3373 return nullptr;
3374 }
3375
optimizeMemCCpyChk(CallInst * CI,IRBuilderBase & B)3376 Value *FortifiedLibCallSimplifier::optimizeMemCCpyChk(CallInst *CI,
3377 IRBuilderBase &B) {
3378 if (isFortifiedCallFoldable(CI, 4, 3))
3379 return emitMemCCpy(CI->getArgOperand(0), CI->getArgOperand(1),
3380 CI->getArgOperand(2), CI->getArgOperand(3), B, TLI);
3381
3382 return nullptr;
3383 }
3384
optimizeSNPrintfChk(CallInst * CI,IRBuilderBase & B)3385 Value *FortifiedLibCallSimplifier::optimizeSNPrintfChk(CallInst *CI,
3386 IRBuilderBase &B) {
3387 if (isFortifiedCallFoldable(CI, 3, 1, None, 2)) {
3388 SmallVector<Value *, 8> VariadicArgs(CI->arg_begin() + 5, CI->arg_end());
3389 return emitSNPrintf(CI->getArgOperand(0), CI->getArgOperand(1),
3390 CI->getArgOperand(4), VariadicArgs, B, TLI);
3391 }
3392
3393 return nullptr;
3394 }
3395
optimizeSPrintfChk(CallInst * CI,IRBuilderBase & B)3396 Value *FortifiedLibCallSimplifier::optimizeSPrintfChk(CallInst *CI,
3397 IRBuilderBase &B) {
3398 if (isFortifiedCallFoldable(CI, 2, None, None, 1)) {
3399 SmallVector<Value *, 8> VariadicArgs(CI->arg_begin() + 4, CI->arg_end());
3400 return emitSPrintf(CI->getArgOperand(0), CI->getArgOperand(3), VariadicArgs,
3401 B, TLI);
3402 }
3403
3404 return nullptr;
3405 }
3406
optimizeStrCatChk(CallInst * CI,IRBuilderBase & B)3407 Value *FortifiedLibCallSimplifier::optimizeStrCatChk(CallInst *CI,
3408 IRBuilderBase &B) {
3409 if (isFortifiedCallFoldable(CI, 2))
3410 return emitStrCat(CI->getArgOperand(0), CI->getArgOperand(1), B, TLI);
3411
3412 return nullptr;
3413 }
3414
optimizeStrLCat(CallInst * CI,IRBuilderBase & B)3415 Value *FortifiedLibCallSimplifier::optimizeStrLCat(CallInst *CI,
3416 IRBuilderBase &B) {
3417 if (isFortifiedCallFoldable(CI, 3))
3418 return emitStrLCat(CI->getArgOperand(0), CI->getArgOperand(1),
3419 CI->getArgOperand(2), B, TLI);
3420
3421 return nullptr;
3422 }
3423
optimizeStrNCatChk(CallInst * CI,IRBuilderBase & B)3424 Value *FortifiedLibCallSimplifier::optimizeStrNCatChk(CallInst *CI,
3425 IRBuilderBase &B) {
3426 if (isFortifiedCallFoldable(CI, 3))
3427 return emitStrNCat(CI->getArgOperand(0), CI->getArgOperand(1),
3428 CI->getArgOperand(2), B, TLI);
3429
3430 return nullptr;
3431 }
3432
optimizeStrLCpyChk(CallInst * CI,IRBuilderBase & B)3433 Value *FortifiedLibCallSimplifier::optimizeStrLCpyChk(CallInst *CI,
3434 IRBuilderBase &B) {
3435 if (isFortifiedCallFoldable(CI, 3))
3436 return emitStrLCpy(CI->getArgOperand(0), CI->getArgOperand(1),
3437 CI->getArgOperand(2), B, TLI);
3438
3439 return nullptr;
3440 }
3441
optimizeVSNPrintfChk(CallInst * CI,IRBuilderBase & B)3442 Value *FortifiedLibCallSimplifier::optimizeVSNPrintfChk(CallInst *CI,
3443 IRBuilderBase &B) {
3444 if (isFortifiedCallFoldable(CI, 3, 1, None, 2))
3445 return emitVSNPrintf(CI->getArgOperand(0), CI->getArgOperand(1),
3446 CI->getArgOperand(4), CI->getArgOperand(5), B, TLI);
3447
3448 return nullptr;
3449 }
3450
optimizeVSPrintfChk(CallInst * CI,IRBuilderBase & B)3451 Value *FortifiedLibCallSimplifier::optimizeVSPrintfChk(CallInst *CI,
3452 IRBuilderBase &B) {
3453 if (isFortifiedCallFoldable(CI, 2, None, None, 1))
3454 return emitVSPrintf(CI->getArgOperand(0), CI->getArgOperand(3),
3455 CI->getArgOperand(4), B, TLI);
3456
3457 return nullptr;
3458 }
3459
optimizeCall(CallInst * CI,IRBuilderBase & Builder)3460 Value *FortifiedLibCallSimplifier::optimizeCall(CallInst *CI,
3461 IRBuilderBase &Builder) {
3462 // FIXME: We shouldn't be changing "nobuiltin" or TLI unavailable calls here.
3463 // Some clang users checked for _chk libcall availability using:
3464 // __has_builtin(__builtin___memcpy_chk)
3465 // When compiling with -fno-builtin, this is always true.
3466 // When passing -ffreestanding/-mkernel, which both imply -fno-builtin, we
3467 // end up with fortified libcalls, which isn't acceptable in a freestanding
3468 // environment which only provides their non-fortified counterparts.
3469 //
3470 // Until we change clang and/or teach external users to check for availability
3471 // differently, disregard the "nobuiltin" attribute and TLI::has.
3472 //
3473 // PR23093.
3474
3475 LibFunc Func;
3476 Function *Callee = CI->getCalledFunction();
3477 bool isCallingConvC = isCallingConvCCompatible(CI);
3478
3479 SmallVector<OperandBundleDef, 2> OpBundles;
3480 CI->getOperandBundlesAsDefs(OpBundles);
3481
3482 IRBuilderBase::OperandBundlesGuard Guard(Builder);
3483 Builder.setDefaultOperandBundles(OpBundles);
3484
3485 // First, check that this is a known library functions and that the prototype
3486 // is correct.
3487 if (!TLI->getLibFunc(*Callee, Func))
3488 return nullptr;
3489
3490 // We never change the calling convention.
3491 if (!ignoreCallingConv(Func) && !isCallingConvC)
3492 return nullptr;
3493
3494 switch (Func) {
3495 case LibFunc_memcpy_chk:
3496 return optimizeMemCpyChk(CI, Builder);
3497 case LibFunc_mempcpy_chk:
3498 return optimizeMemPCpyChk(CI, Builder);
3499 case LibFunc_memmove_chk:
3500 return optimizeMemMoveChk(CI, Builder);
3501 case LibFunc_memset_chk:
3502 return optimizeMemSetChk(CI, Builder);
3503 case LibFunc_stpcpy_chk:
3504 case LibFunc_strcpy_chk:
3505 return optimizeStrpCpyChk(CI, Builder, Func);
3506 case LibFunc_strlen_chk:
3507 return optimizeStrLenChk(CI, Builder);
3508 case LibFunc_stpncpy_chk:
3509 case LibFunc_strncpy_chk:
3510 return optimizeStrpNCpyChk(CI, Builder, Func);
3511 case LibFunc_memccpy_chk:
3512 return optimizeMemCCpyChk(CI, Builder);
3513 case LibFunc_snprintf_chk:
3514 return optimizeSNPrintfChk(CI, Builder);
3515 case LibFunc_sprintf_chk:
3516 return optimizeSPrintfChk(CI, Builder);
3517 case LibFunc_strcat_chk:
3518 return optimizeStrCatChk(CI, Builder);
3519 case LibFunc_strlcat_chk:
3520 return optimizeStrLCat(CI, Builder);
3521 case LibFunc_strncat_chk:
3522 return optimizeStrNCatChk(CI, Builder);
3523 case LibFunc_strlcpy_chk:
3524 return optimizeStrLCpyChk(CI, Builder);
3525 case LibFunc_vsnprintf_chk:
3526 return optimizeVSNPrintfChk(CI, Builder);
3527 case LibFunc_vsprintf_chk:
3528 return optimizeVSPrintfChk(CI, Builder);
3529 default:
3530 break;
3531 }
3532 return nullptr;
3533 }
3534
FortifiedLibCallSimplifier(const TargetLibraryInfo * TLI,bool OnlyLowerUnknownSize)3535 FortifiedLibCallSimplifier::FortifiedLibCallSimplifier(
3536 const TargetLibraryInfo *TLI, bool OnlyLowerUnknownSize)
3537 : TLI(TLI), OnlyLowerUnknownSize(OnlyLowerUnknownSize) {}
3538