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