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->getDereferenceableOrNullBytes(
146 ArgNo + AttributeList::FirstArgIndex),
147 DereferenceableBytes);
148
149 if (CI->getDereferenceableBytes(ArgNo + AttributeList::FirstArgIndex) <
150 DerefBytes) {
151 CI->removeParamAttr(ArgNo, Attribute::Dereferenceable);
152 if (!llvm::NullPointerIsDefined(F, AS) ||
153 CI->paramHasAttr(ArgNo, Attribute::NonNull))
154 CI->removeParamAttr(ArgNo, Attribute::DereferenceableOrNull);
155 CI->addParamAttr(ArgNo, Attribute::getWithDereferenceableBytes(
156 CI->getContext(), DerefBytes));
157 }
158 }
159 }
160
annotateNonNullNoUndefBasedOnAccess(CallInst * CI,ArrayRef<unsigned> ArgNos)161 static void annotateNonNullNoUndefBasedOnAccess(CallInst *CI,
162 ArrayRef<unsigned> ArgNos) {
163 Function *F = CI->getCaller();
164 if (!F)
165 return;
166
167 for (unsigned ArgNo : ArgNos) {
168 if (!CI->paramHasAttr(ArgNo, Attribute::NoUndef))
169 CI->addParamAttr(ArgNo, Attribute::NoUndef);
170
171 if (CI->paramHasAttr(ArgNo, Attribute::NonNull))
172 continue;
173 unsigned AS = CI->getArgOperand(ArgNo)->getType()->getPointerAddressSpace();
174 if (llvm::NullPointerIsDefined(F, AS))
175 continue;
176
177 CI->addParamAttr(ArgNo, Attribute::NonNull);
178 annotateDereferenceableBytes(CI, ArgNo, 1);
179 }
180 }
181
annotateNonNullAndDereferenceable(CallInst * CI,ArrayRef<unsigned> ArgNos,Value * Size,const DataLayout & DL)182 static void annotateNonNullAndDereferenceable(CallInst *CI, ArrayRef<unsigned> ArgNos,
183 Value *Size, const DataLayout &DL) {
184 if (ConstantInt *LenC = dyn_cast<ConstantInt>(Size)) {
185 annotateNonNullNoUndefBasedOnAccess(CI, ArgNos);
186 annotateDereferenceableBytes(CI, ArgNos, LenC->getZExtValue());
187 } else if (isKnownNonZero(Size, DL)) {
188 annotateNonNullNoUndefBasedOnAccess(CI, ArgNos);
189 const APInt *X, *Y;
190 uint64_t DerefMin = 1;
191 if (match(Size, m_Select(m_Value(), m_APInt(X), m_APInt(Y)))) {
192 DerefMin = std::min(X->getZExtValue(), Y->getZExtValue());
193 annotateDereferenceableBytes(CI, ArgNos, DerefMin);
194 }
195 }
196 }
197
198 //===----------------------------------------------------------------------===//
199 // String and Memory Library Call Optimizations
200 //===----------------------------------------------------------------------===//
201
optimizeStrCat(CallInst * CI,IRBuilderBase & B)202 Value *LibCallSimplifier::optimizeStrCat(CallInst *CI, IRBuilderBase &B) {
203 // Extract some information from the instruction
204 Value *Dst = CI->getArgOperand(0);
205 Value *Src = CI->getArgOperand(1);
206 annotateNonNullNoUndefBasedOnAccess(CI, {0, 1});
207
208 // See if we can get the length of the input string.
209 uint64_t Len = GetStringLength(Src);
210 if (Len)
211 annotateDereferenceableBytes(CI, 1, Len);
212 else
213 return nullptr;
214 --Len; // Unbias length.
215
216 // Handle the simple, do-nothing case: strcat(x, "") -> x
217 if (Len == 0)
218 return Dst;
219
220 return emitStrLenMemCpy(Src, Dst, Len, B);
221 }
222
emitStrLenMemCpy(Value * Src,Value * Dst,uint64_t Len,IRBuilderBase & B)223 Value *LibCallSimplifier::emitStrLenMemCpy(Value *Src, Value *Dst, uint64_t Len,
224 IRBuilderBase &B) {
225 // We need to find the end of the destination string. That's where the
226 // memory is to be moved to. We just generate a call to strlen.
227 Value *DstLen = emitStrLen(Dst, B, DL, TLI);
228 if (!DstLen)
229 return nullptr;
230
231 // Now that we have the destination's length, we must index into the
232 // destination's pointer to get the actual memcpy destination (end of
233 // the string .. we're concatenating).
234 Value *CpyDst = B.CreateGEP(B.getInt8Ty(), Dst, DstLen, "endptr");
235
236 // We have enough information to now generate the memcpy call to do the
237 // concatenation for us. Make a memcpy to copy the nul byte with align = 1.
238 B.CreateMemCpy(
239 CpyDst, Align(1), Src, Align(1),
240 ConstantInt::get(DL.getIntPtrType(Src->getContext()), Len + 1));
241 return Dst;
242 }
243
optimizeStrNCat(CallInst * CI,IRBuilderBase & B)244 Value *LibCallSimplifier::optimizeStrNCat(CallInst *CI, IRBuilderBase &B) {
245 // Extract some information from the instruction.
246 Value *Dst = CI->getArgOperand(0);
247 Value *Src = CI->getArgOperand(1);
248 Value *Size = CI->getArgOperand(2);
249 uint64_t Len;
250 annotateNonNullNoUndefBasedOnAccess(CI, 0);
251 if (isKnownNonZero(Size, DL))
252 annotateNonNullNoUndefBasedOnAccess(CI, 1);
253
254 // We don't do anything if length is not constant.
255 ConstantInt *LengthArg = dyn_cast<ConstantInt>(Size);
256 if (LengthArg) {
257 Len = LengthArg->getZExtValue();
258 // strncat(x, c, 0) -> x
259 if (!Len)
260 return Dst;
261 } else {
262 return nullptr;
263 }
264
265 // See if we can get the length of the input string.
266 uint64_t SrcLen = GetStringLength(Src);
267 if (SrcLen) {
268 annotateDereferenceableBytes(CI, 1, SrcLen);
269 --SrcLen; // Unbias length.
270 } else {
271 return nullptr;
272 }
273
274 // strncat(x, "", c) -> x
275 if (SrcLen == 0)
276 return Dst;
277
278 // We don't optimize this case.
279 if (Len < SrcLen)
280 return nullptr;
281
282 // strncat(x, s, c) -> strcat(x, s)
283 // s is constant so the strcat can be optimized further.
284 return emitStrLenMemCpy(Src, Dst, SrcLen, B);
285 }
286
optimizeStrChr(CallInst * CI,IRBuilderBase & B)287 Value *LibCallSimplifier::optimizeStrChr(CallInst *CI, IRBuilderBase &B) {
288 Function *Callee = CI->getCalledFunction();
289 FunctionType *FT = Callee->getFunctionType();
290 Value *SrcStr = CI->getArgOperand(0);
291 annotateNonNullNoUndefBasedOnAccess(CI, 0);
292
293 // If the second operand is non-constant, see if we can compute the length
294 // of the input string and turn this into memchr.
295 ConstantInt *CharC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
296 if (!CharC) {
297 uint64_t Len = GetStringLength(SrcStr);
298 if (Len)
299 annotateDereferenceableBytes(CI, 0, Len);
300 else
301 return nullptr;
302 if (!FT->getParamType(1)->isIntegerTy(32)) // memchr needs i32.
303 return nullptr;
304
305 return emitMemChr(SrcStr, CI->getArgOperand(1), // include nul.
306 ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len),
307 B, DL, TLI);
308 }
309
310 // Otherwise, the character is a constant, see if the first argument is
311 // a string literal. If so, we can constant fold.
312 StringRef Str;
313 if (!getConstantStringInfo(SrcStr, Str)) {
314 if (CharC->isZero()) // strchr(p, 0) -> p + strlen(p)
315 if (Value *StrLen = emitStrLen(SrcStr, B, DL, TLI))
316 return B.CreateGEP(B.getInt8Ty(), SrcStr, StrLen, "strchr");
317 return nullptr;
318 }
319
320 // Compute the offset, make sure to handle the case when we're searching for
321 // zero (a weird way to spell strlen).
322 size_t I = (0xFF & CharC->getSExtValue()) == 0
323 ? Str.size()
324 : Str.find(CharC->getSExtValue());
325 if (I == StringRef::npos) // Didn't find the char. strchr returns null.
326 return Constant::getNullValue(CI->getType());
327
328 // strchr(s+n,c) -> gep(s+n+i,c)
329 return B.CreateGEP(B.getInt8Ty(), SrcStr, B.getInt64(I), "strchr");
330 }
331
optimizeStrRChr(CallInst * CI,IRBuilderBase & B)332 Value *LibCallSimplifier::optimizeStrRChr(CallInst *CI, IRBuilderBase &B) {
333 Value *SrcStr = CI->getArgOperand(0);
334 ConstantInt *CharC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
335 annotateNonNullNoUndefBasedOnAccess(CI, 0);
336
337 // Cannot fold anything if we're not looking for a constant.
338 if (!CharC)
339 return nullptr;
340
341 StringRef Str;
342 if (!getConstantStringInfo(SrcStr, Str)) {
343 // strrchr(s, 0) -> strchr(s, 0)
344 if (CharC->isZero())
345 return emitStrChr(SrcStr, '\0', B, TLI);
346 return nullptr;
347 }
348
349 // Compute the offset.
350 size_t I = (0xFF & CharC->getSExtValue()) == 0
351 ? Str.size()
352 : Str.rfind(CharC->getSExtValue());
353 if (I == StringRef::npos) // Didn't find the char. Return null.
354 return Constant::getNullValue(CI->getType());
355
356 // strrchr(s+n,c) -> gep(s+n+i,c)
357 return B.CreateGEP(B.getInt8Ty(), SrcStr, B.getInt64(I), "strrchr");
358 }
359
optimizeStrCmp(CallInst * CI,IRBuilderBase & B)360 Value *LibCallSimplifier::optimizeStrCmp(CallInst *CI, IRBuilderBase &B) {
361 Value *Str1P = CI->getArgOperand(0), *Str2P = CI->getArgOperand(1);
362 if (Str1P == Str2P) // strcmp(x,x) -> 0
363 return ConstantInt::get(CI->getType(), 0);
364
365 StringRef Str1, Str2;
366 bool HasStr1 = getConstantStringInfo(Str1P, Str1);
367 bool HasStr2 = getConstantStringInfo(Str2P, Str2);
368
369 // strcmp(x, y) -> cnst (if both x and y are constant strings)
370 if (HasStr1 && HasStr2)
371 return ConstantInt::get(CI->getType(), Str1.compare(Str2));
372
373 if (HasStr1 && Str1.empty()) // strcmp("", x) -> -*x
374 return B.CreateNeg(B.CreateZExt(
375 B.CreateLoad(B.getInt8Ty(), Str2P, "strcmpload"), CI->getType()));
376
377 if (HasStr2 && Str2.empty()) // strcmp(x,"") -> *x
378 return B.CreateZExt(B.CreateLoad(B.getInt8Ty(), Str1P, "strcmpload"),
379 CI->getType());
380
381 // strcmp(P, "x") -> memcmp(P, "x", 2)
382 uint64_t Len1 = GetStringLength(Str1P);
383 if (Len1)
384 annotateDereferenceableBytes(CI, 0, Len1);
385 uint64_t Len2 = GetStringLength(Str2P);
386 if (Len2)
387 annotateDereferenceableBytes(CI, 1, Len2);
388
389 if (Len1 && Len2) {
390 return emitMemCmp(Str1P, Str2P,
391 ConstantInt::get(DL.getIntPtrType(CI->getContext()),
392 std::min(Len1, Len2)),
393 B, DL, TLI);
394 }
395
396 // strcmp to memcmp
397 if (!HasStr1 && HasStr2) {
398 if (canTransformToMemCmp(CI, Str1P, Len2, DL))
399 return emitMemCmp(
400 Str1P, Str2P,
401 ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len2), B, DL,
402 TLI);
403 } else if (HasStr1 && !HasStr2) {
404 if (canTransformToMemCmp(CI, Str2P, Len1, DL))
405 return emitMemCmp(
406 Str1P, Str2P,
407 ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len1), B, DL,
408 TLI);
409 }
410
411 annotateNonNullNoUndefBasedOnAccess(CI, {0, 1});
412 return nullptr;
413 }
414
optimizeStrNCmp(CallInst * CI,IRBuilderBase & B)415 Value *LibCallSimplifier::optimizeStrNCmp(CallInst *CI, IRBuilderBase &B) {
416 Value *Str1P = CI->getArgOperand(0);
417 Value *Str2P = CI->getArgOperand(1);
418 Value *Size = CI->getArgOperand(2);
419 if (Str1P == Str2P) // strncmp(x,x,n) -> 0
420 return ConstantInt::get(CI->getType(), 0);
421
422 if (isKnownNonZero(Size, DL))
423 annotateNonNullNoUndefBasedOnAccess(CI, {0, 1});
424 // Get the length argument if it is constant.
425 uint64_t Length;
426 if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(Size))
427 Length = LengthArg->getZExtValue();
428 else
429 return nullptr;
430
431 if (Length == 0) // strncmp(x,y,0) -> 0
432 return ConstantInt::get(CI->getType(), 0);
433
434 if (Length == 1) // strncmp(x,y,1) -> memcmp(x,y,1)
435 return emitMemCmp(Str1P, Str2P, Size, B, DL, TLI);
436
437 StringRef Str1, Str2;
438 bool HasStr1 = getConstantStringInfo(Str1P, Str1);
439 bool HasStr2 = getConstantStringInfo(Str2P, Str2);
440
441 // strncmp(x, y) -> cnst (if both x and y are constant strings)
442 if (HasStr1 && HasStr2) {
443 StringRef SubStr1 = Str1.substr(0, Length);
444 StringRef SubStr2 = Str2.substr(0, Length);
445 return ConstantInt::get(CI->getType(), SubStr1.compare(SubStr2));
446 }
447
448 if (HasStr1 && Str1.empty()) // strncmp("", x, n) -> -*x
449 return B.CreateNeg(B.CreateZExt(
450 B.CreateLoad(B.getInt8Ty(), Str2P, "strcmpload"), CI->getType()));
451
452 if (HasStr2 && Str2.empty()) // strncmp(x, "", n) -> *x
453 return B.CreateZExt(B.CreateLoad(B.getInt8Ty(), Str1P, "strcmpload"),
454 CI->getType());
455
456 uint64_t Len1 = GetStringLength(Str1P);
457 if (Len1)
458 annotateDereferenceableBytes(CI, 0, Len1);
459 uint64_t Len2 = GetStringLength(Str2P);
460 if (Len2)
461 annotateDereferenceableBytes(CI, 1, Len2);
462
463 // strncmp to memcmp
464 if (!HasStr1 && HasStr2) {
465 Len2 = std::min(Len2, Length);
466 if (canTransformToMemCmp(CI, Str1P, Len2, DL))
467 return emitMemCmp(
468 Str1P, Str2P,
469 ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len2), B, DL,
470 TLI);
471 } else if (HasStr1 && !HasStr2) {
472 Len1 = std::min(Len1, Length);
473 if (canTransformToMemCmp(CI, Str2P, Len1, DL))
474 return emitMemCmp(
475 Str1P, Str2P,
476 ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len1), B, DL,
477 TLI);
478 }
479
480 return nullptr;
481 }
482
optimizeStrNDup(CallInst * CI,IRBuilderBase & B)483 Value *LibCallSimplifier::optimizeStrNDup(CallInst *CI, IRBuilderBase &B) {
484 Value *Src = CI->getArgOperand(0);
485 ConstantInt *Size = dyn_cast<ConstantInt>(CI->getArgOperand(1));
486 uint64_t SrcLen = GetStringLength(Src);
487 if (SrcLen && Size) {
488 annotateDereferenceableBytes(CI, 0, SrcLen);
489 if (SrcLen <= Size->getZExtValue() + 1)
490 return emitStrDup(Src, B, TLI);
491 }
492
493 return nullptr;
494 }
495
optimizeStrCpy(CallInst * CI,IRBuilderBase & B)496 Value *LibCallSimplifier::optimizeStrCpy(CallInst *CI, IRBuilderBase &B) {
497 Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1);
498 if (Dst == Src) // strcpy(x,x) -> x
499 return Src;
500
501 annotateNonNullNoUndefBasedOnAccess(CI, {0, 1});
502 // See if we can get the length of the input string.
503 uint64_t Len = GetStringLength(Src);
504 if (Len)
505 annotateDereferenceableBytes(CI, 1, Len);
506 else
507 return nullptr;
508
509 // We have enough information to now generate the memcpy call to do the
510 // copy for us. Make a memcpy to copy the nul byte with align = 1.
511 CallInst *NewCI =
512 B.CreateMemCpy(Dst, Align(1), Src, Align(1),
513 ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len));
514 NewCI->setAttributes(CI->getAttributes());
515 NewCI->removeAttributes(AttributeList::ReturnIndex,
516 AttributeFuncs::typeIncompatible(NewCI->getType()));
517 return Dst;
518 }
519
optimizeStpCpy(CallInst * CI,IRBuilderBase & B)520 Value *LibCallSimplifier::optimizeStpCpy(CallInst *CI, IRBuilderBase &B) {
521 Function *Callee = CI->getCalledFunction();
522 Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1);
523 if (Dst == Src) { // stpcpy(x,x) -> x+strlen(x)
524 Value *StrLen = emitStrLen(Src, B, DL, TLI);
525 return StrLen ? B.CreateInBoundsGEP(B.getInt8Ty(), Dst, StrLen) : nullptr;
526 }
527
528 // See if we can get the length of the input string.
529 uint64_t Len = GetStringLength(Src);
530 if (Len)
531 annotateDereferenceableBytes(CI, 1, Len);
532 else
533 return nullptr;
534
535 Type *PT = Callee->getFunctionType()->getParamType(0);
536 Value *LenV = ConstantInt::get(DL.getIntPtrType(PT), Len);
537 Value *DstEnd = B.CreateGEP(B.getInt8Ty(), Dst,
538 ConstantInt::get(DL.getIntPtrType(PT), Len - 1));
539
540 // We have enough information to now generate the memcpy call to do the
541 // copy for us. Make a memcpy to copy the nul byte with align = 1.
542 CallInst *NewCI = B.CreateMemCpy(Dst, Align(1), Src, Align(1), LenV);
543 NewCI->setAttributes(CI->getAttributes());
544 NewCI->removeAttributes(AttributeList::ReturnIndex,
545 AttributeFuncs::typeIncompatible(NewCI->getType()));
546 return DstEnd;
547 }
548
optimizeStrNCpy(CallInst * CI,IRBuilderBase & B)549 Value *LibCallSimplifier::optimizeStrNCpy(CallInst *CI, IRBuilderBase &B) {
550 Function *Callee = CI->getCalledFunction();
551 Value *Dst = CI->getArgOperand(0);
552 Value *Src = CI->getArgOperand(1);
553 Value *Size = CI->getArgOperand(2);
554 annotateNonNullNoUndefBasedOnAccess(CI, 0);
555 if (isKnownNonZero(Size, DL))
556 annotateNonNullNoUndefBasedOnAccess(CI, 1);
557
558 uint64_t Len;
559 if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(Size))
560 Len = LengthArg->getZExtValue();
561 else
562 return nullptr;
563
564 // strncpy(x, y, 0) -> x
565 if (Len == 0)
566 return Dst;
567
568 // See if we can get the length of the input string.
569 uint64_t SrcLen = GetStringLength(Src);
570 if (SrcLen) {
571 annotateDereferenceableBytes(CI, 1, SrcLen);
572 --SrcLen; // Unbias length.
573 } else {
574 return nullptr;
575 }
576
577 if (SrcLen == 0) {
578 // strncpy(x, "", y) -> memset(x, '\0', y)
579 Align MemSetAlign =
580 CI->getAttributes().getParamAttributes(0).getAlignment().valueOrOne();
581 CallInst *NewCI = B.CreateMemSet(Dst, B.getInt8('\0'), Size, MemSetAlign);
582 AttrBuilder ArgAttrs(CI->getAttributes().getParamAttributes(0));
583 NewCI->setAttributes(NewCI->getAttributes().addParamAttributes(
584 CI->getContext(), 0, ArgAttrs));
585 return Dst;
586 }
587
588 // strncpy(a, "a", 4) - > memcpy(a, "a\0\0\0", 4)
589 if (Len > SrcLen + 1) {
590 if (Len <= 128) {
591 StringRef Str;
592 if (!getConstantStringInfo(Src, Str))
593 return nullptr;
594 std::string SrcStr = Str.str();
595 SrcStr.resize(Len, '\0');
596 Src = B.CreateGlobalString(SrcStr, "str");
597 } else {
598 return nullptr;
599 }
600 }
601
602 Type *PT = Callee->getFunctionType()->getParamType(0);
603 // strncpy(x, s, c) -> memcpy(align 1 x, align 1 s, c) [s and c are constant]
604 CallInst *NewCI = B.CreateMemCpy(Dst, Align(1), Src, Align(1),
605 ConstantInt::get(DL.getIntPtrType(PT), Len));
606 NewCI->setAttributes(CI->getAttributes());
607 NewCI->removeAttributes(AttributeList::ReturnIndex,
608 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.CreateAnd(Bounds, Bits, "memchr"), CI->getType());
937 }
938
939 // Check if all arguments are constants. If so, we can constant fold.
940 if (!CharC)
941 return nullptr;
942
943 // Compute the offset.
944 size_t I = Str.find(CharC->getSExtValue() & 0xFF);
945 if (I == StringRef::npos) // Didn't find the char. memchr returns null.
946 return Constant::getNullValue(CI->getType());
947
948 // memchr(s+n,c,l) -> gep(s+n+i,c)
949 return B.CreateGEP(B.getInt8Ty(), SrcStr, B.getInt64(I), "memchr");
950 }
951
optimizeMemCmpConstantSize(CallInst * CI,Value * LHS,Value * RHS,uint64_t Len,IRBuilderBase & B,const DataLayout & DL)952 static Value *optimizeMemCmpConstantSize(CallInst *CI, Value *LHS, Value *RHS,
953 uint64_t Len, IRBuilderBase &B,
954 const DataLayout &DL) {
955 if (Len == 0) // memcmp(s1,s2,0) -> 0
956 return Constant::getNullValue(CI->getType());
957
958 // memcmp(S1,S2,1) -> *(unsigned char*)LHS - *(unsigned char*)RHS
959 if (Len == 1) {
960 Value *LHSV =
961 B.CreateZExt(B.CreateLoad(B.getInt8Ty(), castToCStr(LHS, B), "lhsc"),
962 CI->getType(), "lhsv");
963 Value *RHSV =
964 B.CreateZExt(B.CreateLoad(B.getInt8Ty(), castToCStr(RHS, B), "rhsc"),
965 CI->getType(), "rhsv");
966 return B.CreateSub(LHSV, RHSV, "chardiff");
967 }
968
969 // memcmp(S1,S2,N/8)==0 -> (*(intN_t*)S1 != *(intN_t*)S2)==0
970 // TODO: The case where both inputs are constants does not need to be limited
971 // to legal integers or equality comparison. See block below this.
972 if (DL.isLegalInteger(Len * 8) && isOnlyUsedInZeroEqualityComparison(CI)) {
973 IntegerType *IntType = IntegerType::get(CI->getContext(), Len * 8);
974 unsigned PrefAlignment = DL.getPrefTypeAlignment(IntType);
975
976 // First, see if we can fold either argument to a constant.
977 Value *LHSV = nullptr;
978 if (auto *LHSC = dyn_cast<Constant>(LHS)) {
979 LHSC = ConstantExpr::getBitCast(LHSC, IntType->getPointerTo());
980 LHSV = ConstantFoldLoadFromConstPtr(LHSC, IntType, DL);
981 }
982 Value *RHSV = nullptr;
983 if (auto *RHSC = dyn_cast<Constant>(RHS)) {
984 RHSC = ConstantExpr::getBitCast(RHSC, IntType->getPointerTo());
985 RHSV = ConstantFoldLoadFromConstPtr(RHSC, IntType, DL);
986 }
987
988 // Don't generate unaligned loads. If either source is constant data,
989 // alignment doesn't matter for that source because there is no load.
990 if ((LHSV || getKnownAlignment(LHS, DL, CI) >= PrefAlignment) &&
991 (RHSV || getKnownAlignment(RHS, DL, CI) >= PrefAlignment)) {
992 if (!LHSV) {
993 Type *LHSPtrTy =
994 IntType->getPointerTo(LHS->getType()->getPointerAddressSpace());
995 LHSV = B.CreateLoad(IntType, B.CreateBitCast(LHS, LHSPtrTy), "lhsv");
996 }
997 if (!RHSV) {
998 Type *RHSPtrTy =
999 IntType->getPointerTo(RHS->getType()->getPointerAddressSpace());
1000 RHSV = B.CreateLoad(IntType, B.CreateBitCast(RHS, RHSPtrTy), "rhsv");
1001 }
1002 return B.CreateZExt(B.CreateICmpNE(LHSV, RHSV), CI->getType(), "memcmp");
1003 }
1004 }
1005
1006 // Constant folding: memcmp(x, y, Len) -> constant (all arguments are const).
1007 // TODO: This is limited to i8 arrays.
1008 StringRef LHSStr, RHSStr;
1009 if (getConstantStringInfo(LHS, LHSStr) &&
1010 getConstantStringInfo(RHS, RHSStr)) {
1011 // Make sure we're not reading out-of-bounds memory.
1012 if (Len > LHSStr.size() || Len > RHSStr.size())
1013 return nullptr;
1014 // Fold the memcmp and normalize the result. This way we get consistent
1015 // results across multiple platforms.
1016 uint64_t Ret = 0;
1017 int Cmp = memcmp(LHSStr.data(), RHSStr.data(), Len);
1018 if (Cmp < 0)
1019 Ret = -1;
1020 else if (Cmp > 0)
1021 Ret = 1;
1022 return ConstantInt::get(CI->getType(), Ret);
1023 }
1024
1025 return nullptr;
1026 }
1027
1028 // Most simplifications for memcmp also apply to bcmp.
optimizeMemCmpBCmpCommon(CallInst * CI,IRBuilderBase & B)1029 Value *LibCallSimplifier::optimizeMemCmpBCmpCommon(CallInst *CI,
1030 IRBuilderBase &B) {
1031 Value *LHS = CI->getArgOperand(0), *RHS = CI->getArgOperand(1);
1032 Value *Size = CI->getArgOperand(2);
1033
1034 if (LHS == RHS) // memcmp(s,s,x) -> 0
1035 return Constant::getNullValue(CI->getType());
1036
1037 annotateNonNullAndDereferenceable(CI, {0, 1}, Size, DL);
1038 // Handle constant lengths.
1039 ConstantInt *LenC = dyn_cast<ConstantInt>(Size);
1040 if (!LenC)
1041 return nullptr;
1042
1043 // memcmp(d,s,0) -> 0
1044 if (LenC->getZExtValue() == 0)
1045 return Constant::getNullValue(CI->getType());
1046
1047 if (Value *Res =
1048 optimizeMemCmpConstantSize(CI, LHS, RHS, LenC->getZExtValue(), B, DL))
1049 return Res;
1050 return nullptr;
1051 }
1052
optimizeMemCmp(CallInst * CI,IRBuilderBase & B)1053 Value *LibCallSimplifier::optimizeMemCmp(CallInst *CI, IRBuilderBase &B) {
1054 if (Value *V = optimizeMemCmpBCmpCommon(CI, B))
1055 return V;
1056
1057 // memcmp(x, y, Len) == 0 -> bcmp(x, y, Len) == 0
1058 // bcmp can be more efficient than memcmp because it only has to know that
1059 // there is a difference, not how different one is to the other.
1060 if (TLI->has(LibFunc_bcmp) && isOnlyUsedInZeroEqualityComparison(CI)) {
1061 Value *LHS = CI->getArgOperand(0);
1062 Value *RHS = CI->getArgOperand(1);
1063 Value *Size = CI->getArgOperand(2);
1064 return emitBCmp(LHS, RHS, Size, B, DL, TLI);
1065 }
1066
1067 return nullptr;
1068 }
1069
optimizeBCmp(CallInst * CI,IRBuilderBase & B)1070 Value *LibCallSimplifier::optimizeBCmp(CallInst *CI, IRBuilderBase &B) {
1071 return optimizeMemCmpBCmpCommon(CI, B);
1072 }
1073
optimizeMemCpy(CallInst * CI,IRBuilderBase & B)1074 Value *LibCallSimplifier::optimizeMemCpy(CallInst *CI, IRBuilderBase &B) {
1075 Value *Size = CI->getArgOperand(2);
1076 annotateNonNullAndDereferenceable(CI, {0, 1}, Size, DL);
1077 if (isa<IntrinsicInst>(CI))
1078 return nullptr;
1079
1080 // memcpy(x, y, n) -> llvm.memcpy(align 1 x, align 1 y, n)
1081 CallInst *NewCI = B.CreateMemCpy(CI->getArgOperand(0), Align(1),
1082 CI->getArgOperand(1), Align(1), Size);
1083 NewCI->setAttributes(CI->getAttributes());
1084 NewCI->removeAttributes(AttributeList::ReturnIndex,
1085 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->removeAttributes(AttributeList::ReturnIndex,
1139 AttributeFuncs::typeIncompatible(NewCI->getType()));
1140 return B.CreateInBoundsGEP(B.getInt8Ty(), Dst, N);
1141 }
1142
optimizeMemMove(CallInst * CI,IRBuilderBase & B)1143 Value *LibCallSimplifier::optimizeMemMove(CallInst *CI, IRBuilderBase &B) {
1144 Value *Size = CI->getArgOperand(2);
1145 annotateNonNullAndDereferenceable(CI, {0, 1}, Size, DL);
1146 if (isa<IntrinsicInst>(CI))
1147 return nullptr;
1148
1149 // memmove(x, y, n) -> llvm.memmove(align 1 x, align 1 y, n)
1150 CallInst *NewCI = B.CreateMemMove(CI->getArgOperand(0), Align(1),
1151 CI->getArgOperand(1), Align(1), Size);
1152 NewCI->setAttributes(CI->getAttributes());
1153 NewCI->removeAttributes(AttributeList::ReturnIndex,
1154 AttributeFuncs::typeIncompatible(NewCI->getType()));
1155 return CI->getArgOperand(0);
1156 }
1157
1158 /// Fold memset[_chk](malloc(n), 0, n) --> calloc(1, n).
foldMallocMemset(CallInst * Memset,IRBuilderBase & B)1159 Value *LibCallSimplifier::foldMallocMemset(CallInst *Memset, IRBuilderBase &B) {
1160 // This has to be a memset of zeros (bzero).
1161 auto *FillValue = dyn_cast<ConstantInt>(Memset->getArgOperand(1));
1162 if (!FillValue || FillValue->getZExtValue() != 0)
1163 return nullptr;
1164
1165 // TODO: We should handle the case where the malloc has more than one use.
1166 // This is necessary to optimize common patterns such as when the result of
1167 // the malloc is checked against null or when a memset intrinsic is used in
1168 // place of a memset library call.
1169 auto *Malloc = dyn_cast<CallInst>(Memset->getArgOperand(0));
1170 if (!Malloc || !Malloc->hasOneUse())
1171 return nullptr;
1172
1173 // Is the inner call really malloc()?
1174 Function *InnerCallee = Malloc->getCalledFunction();
1175 if (!InnerCallee)
1176 return nullptr;
1177
1178 LibFunc Func;
1179 if (!TLI->getLibFunc(*InnerCallee, Func) || !TLI->has(Func) ||
1180 Func != LibFunc_malloc)
1181 return nullptr;
1182
1183 // The memset must cover the same number of bytes that are malloc'd.
1184 if (Memset->getArgOperand(2) != Malloc->getArgOperand(0))
1185 return nullptr;
1186
1187 // Replace the malloc with a calloc. We need the data layout to know what the
1188 // actual size of a 'size_t' parameter is.
1189 B.SetInsertPoint(Malloc->getParent(), ++Malloc->getIterator());
1190 const DataLayout &DL = Malloc->getModule()->getDataLayout();
1191 IntegerType *SizeType = DL.getIntPtrType(B.GetInsertBlock()->getContext());
1192 if (Value *Calloc = emitCalloc(ConstantInt::get(SizeType, 1),
1193 Malloc->getArgOperand(0),
1194 Malloc->getAttributes(), B, *TLI)) {
1195 substituteInParent(Malloc, Calloc);
1196 return Calloc;
1197 }
1198
1199 return nullptr;
1200 }
1201
optimizeMemSet(CallInst * CI,IRBuilderBase & B)1202 Value *LibCallSimplifier::optimizeMemSet(CallInst *CI, IRBuilderBase &B) {
1203 Value *Size = CI->getArgOperand(2);
1204 annotateNonNullAndDereferenceable(CI, 0, Size, DL);
1205 if (isa<IntrinsicInst>(CI))
1206 return nullptr;
1207
1208 if (auto *Calloc = foldMallocMemset(CI, B))
1209 return Calloc;
1210
1211 // memset(p, v, n) -> llvm.memset(align 1 p, v, n)
1212 Value *Val = B.CreateIntCast(CI->getArgOperand(1), B.getInt8Ty(), false);
1213 CallInst *NewCI = B.CreateMemSet(CI->getArgOperand(0), Val, Size, Align(1));
1214 NewCI->setAttributes(CI->getAttributes());
1215 NewCI->removeAttributes(AttributeList::ReturnIndex,
1216 AttributeFuncs::typeIncompatible(NewCI->getType()));
1217 return CI->getArgOperand(0);
1218 }
1219
optimizeRealloc(CallInst * CI,IRBuilderBase & B)1220 Value *LibCallSimplifier::optimizeRealloc(CallInst *CI, IRBuilderBase &B) {
1221 if (isa<ConstantPointerNull>(CI->getArgOperand(0)))
1222 return emitMalloc(CI->getArgOperand(1), B, DL, TLI);
1223
1224 return nullptr;
1225 }
1226
1227 //===----------------------------------------------------------------------===//
1228 // Math Library Optimizations
1229 //===----------------------------------------------------------------------===//
1230
1231 // Replace a libcall \p CI with a call to intrinsic \p IID
replaceUnaryCall(CallInst * CI,IRBuilderBase & B,Intrinsic::ID IID)1232 static Value *replaceUnaryCall(CallInst *CI, IRBuilderBase &B,
1233 Intrinsic::ID IID) {
1234 // Propagate fast-math flags from the existing call to the new call.
1235 IRBuilderBase::FastMathFlagGuard Guard(B);
1236 B.setFastMathFlags(CI->getFastMathFlags());
1237
1238 Module *M = CI->getModule();
1239 Value *V = CI->getArgOperand(0);
1240 Function *F = Intrinsic::getDeclaration(M, IID, CI->getType());
1241 CallInst *NewCall = B.CreateCall(F, V);
1242 NewCall->takeName(CI);
1243 return NewCall;
1244 }
1245
1246 /// Return a variant of Val with float type.
1247 /// Currently this works in two cases: If Val is an FPExtension of a float
1248 /// value to something bigger, simply return the operand.
1249 /// If Val is a ConstantFP but can be converted to a float ConstantFP without
1250 /// loss of precision do so.
valueHasFloatPrecision(Value * Val)1251 static Value *valueHasFloatPrecision(Value *Val) {
1252 if (FPExtInst *Cast = dyn_cast<FPExtInst>(Val)) {
1253 Value *Op = Cast->getOperand(0);
1254 if (Op->getType()->isFloatTy())
1255 return Op;
1256 }
1257 if (ConstantFP *Const = dyn_cast<ConstantFP>(Val)) {
1258 APFloat F = Const->getValueAPF();
1259 bool losesInfo;
1260 (void)F.convert(APFloat::IEEEsingle(), APFloat::rmNearestTiesToEven,
1261 &losesInfo);
1262 if (!losesInfo)
1263 return ConstantFP::get(Const->getContext(), F);
1264 }
1265 return nullptr;
1266 }
1267
1268 /// Shrink double -> float functions.
optimizeDoubleFP(CallInst * CI,IRBuilderBase & B,bool isBinary,bool isPrecise=false)1269 static Value *optimizeDoubleFP(CallInst *CI, IRBuilderBase &B,
1270 bool isBinary, bool isPrecise = false) {
1271 Function *CalleeFn = CI->getCalledFunction();
1272 if (!CI->getType()->isDoubleTy() || !CalleeFn)
1273 return nullptr;
1274
1275 // If not all the uses of the function are converted to float, then bail out.
1276 // This matters if the precision of the result is more important than the
1277 // precision of the arguments.
1278 if (isPrecise)
1279 for (User *U : CI->users()) {
1280 FPTruncInst *Cast = dyn_cast<FPTruncInst>(U);
1281 if (!Cast || !Cast->getType()->isFloatTy())
1282 return nullptr;
1283 }
1284
1285 // If this is something like 'g((double) float)', convert to 'gf(float)'.
1286 Value *V[2];
1287 V[0] = valueHasFloatPrecision(CI->getArgOperand(0));
1288 V[1] = isBinary ? valueHasFloatPrecision(CI->getArgOperand(1)) : nullptr;
1289 if (!V[0] || (isBinary && !V[1]))
1290 return nullptr;
1291
1292 // If call isn't an intrinsic, check that it isn't within a function with the
1293 // same name as the float version of this call, otherwise the result is an
1294 // infinite loop. For example, from MinGW-w64:
1295 //
1296 // float expf(float val) { return (float) exp((double) val); }
1297 StringRef CalleeName = CalleeFn->getName();
1298 bool IsIntrinsic = CalleeFn->isIntrinsic();
1299 if (!IsIntrinsic) {
1300 StringRef CallerName = CI->getFunction()->getName();
1301 if (!CallerName.empty() && CallerName.back() == 'f' &&
1302 CallerName.size() == (CalleeName.size() + 1) &&
1303 CallerName.startswith(CalleeName))
1304 return nullptr;
1305 }
1306
1307 // Propagate the math semantics from the current function to the new function.
1308 IRBuilderBase::FastMathFlagGuard Guard(B);
1309 B.setFastMathFlags(CI->getFastMathFlags());
1310
1311 // g((double) float) -> (double) gf(float)
1312 Value *R;
1313 if (IsIntrinsic) {
1314 Module *M = CI->getModule();
1315 Intrinsic::ID IID = CalleeFn->getIntrinsicID();
1316 Function *Fn = Intrinsic::getDeclaration(M, IID, B.getFloatTy());
1317 R = isBinary ? B.CreateCall(Fn, V) : B.CreateCall(Fn, V[0]);
1318 } else {
1319 AttributeList CalleeAttrs = CalleeFn->getAttributes();
1320 R = isBinary ? emitBinaryFloatFnCall(V[0], V[1], CalleeName, B, CalleeAttrs)
1321 : emitUnaryFloatFnCall(V[0], CalleeName, B, CalleeAttrs);
1322 }
1323 return B.CreateFPExt(R, B.getDoubleTy());
1324 }
1325
1326 /// Shrink double -> float for unary functions.
optimizeUnaryDoubleFP(CallInst * CI,IRBuilderBase & B,bool isPrecise=false)1327 static Value *optimizeUnaryDoubleFP(CallInst *CI, IRBuilderBase &B,
1328 bool isPrecise = false) {
1329 return optimizeDoubleFP(CI, B, false, isPrecise);
1330 }
1331
1332 /// Shrink double -> float for binary functions.
optimizeBinaryDoubleFP(CallInst * CI,IRBuilderBase & B,bool isPrecise=false)1333 static Value *optimizeBinaryDoubleFP(CallInst *CI, IRBuilderBase &B,
1334 bool isPrecise = false) {
1335 return optimizeDoubleFP(CI, B, true, isPrecise);
1336 }
1337
1338 // cabs(z) -> sqrt((creal(z)*creal(z)) + (cimag(z)*cimag(z)))
optimizeCAbs(CallInst * CI,IRBuilderBase & B)1339 Value *LibCallSimplifier::optimizeCAbs(CallInst *CI, IRBuilderBase &B) {
1340 if (!CI->isFast())
1341 return nullptr;
1342
1343 // Propagate fast-math flags from the existing call to new instructions.
1344 IRBuilderBase::FastMathFlagGuard Guard(B);
1345 B.setFastMathFlags(CI->getFastMathFlags());
1346
1347 Value *Real, *Imag;
1348 if (CI->getNumArgOperands() == 1) {
1349 Value *Op = CI->getArgOperand(0);
1350 assert(Op->getType()->isArrayTy() && "Unexpected signature for cabs!");
1351 Real = B.CreateExtractValue(Op, 0, "real");
1352 Imag = B.CreateExtractValue(Op, 1, "imag");
1353 } else {
1354 assert(CI->getNumArgOperands() == 2 && "Unexpected signature for cabs!");
1355 Real = CI->getArgOperand(0);
1356 Imag = CI->getArgOperand(1);
1357 }
1358
1359 Value *RealReal = B.CreateFMul(Real, Real);
1360 Value *ImagImag = B.CreateFMul(Imag, Imag);
1361
1362 Function *FSqrt = Intrinsic::getDeclaration(CI->getModule(), Intrinsic::sqrt,
1363 CI->getType());
1364 return B.CreateCall(FSqrt, B.CreateFAdd(RealReal, ImagImag), "cabs");
1365 }
1366
optimizeTrigReflections(CallInst * Call,LibFunc Func,IRBuilderBase & B)1367 static Value *optimizeTrigReflections(CallInst *Call, LibFunc Func,
1368 IRBuilderBase &B) {
1369 if (!isa<FPMathOperator>(Call))
1370 return nullptr;
1371
1372 IRBuilderBase::FastMathFlagGuard Guard(B);
1373 B.setFastMathFlags(Call->getFastMathFlags());
1374
1375 // TODO: Can this be shared to also handle LLVM intrinsics?
1376 Value *X;
1377 switch (Func) {
1378 case LibFunc_sin:
1379 case LibFunc_sinf:
1380 case LibFunc_sinl:
1381 case LibFunc_tan:
1382 case LibFunc_tanf:
1383 case LibFunc_tanl:
1384 // sin(-X) --> -sin(X)
1385 // tan(-X) --> -tan(X)
1386 if (match(Call->getArgOperand(0), m_OneUse(m_FNeg(m_Value(X)))))
1387 return B.CreateFNeg(B.CreateCall(Call->getCalledFunction(), X));
1388 break;
1389 case LibFunc_cos:
1390 case LibFunc_cosf:
1391 case LibFunc_cosl:
1392 // cos(-X) --> cos(X)
1393 if (match(Call->getArgOperand(0), m_FNeg(m_Value(X))))
1394 return B.CreateCall(Call->getCalledFunction(), X, "cos");
1395 break;
1396 default:
1397 break;
1398 }
1399 return nullptr;
1400 }
1401
getPow(Value * InnerChain[33],unsigned Exp,IRBuilderBase & B)1402 static Value *getPow(Value *InnerChain[33], unsigned Exp, IRBuilderBase &B) {
1403 // Multiplications calculated using Addition Chains.
1404 // Refer: http://wwwhomes.uni-bielefeld.de/achim/addition_chain.html
1405
1406 assert(Exp != 0 && "Incorrect exponent 0 not handled");
1407
1408 if (InnerChain[Exp])
1409 return InnerChain[Exp];
1410
1411 static const unsigned AddChain[33][2] = {
1412 {0, 0}, // Unused.
1413 {0, 0}, // Unused (base case = pow1).
1414 {1, 1}, // Unused (pre-computed).
1415 {1, 2}, {2, 2}, {2, 3}, {3, 3}, {2, 5}, {4, 4},
1416 {1, 8}, {5, 5}, {1, 10}, {6, 6}, {4, 9}, {7, 7},
1417 {3, 12}, {8, 8}, {8, 9}, {2, 16}, {1, 18}, {10, 10},
1418 {6, 15}, {11, 11}, {3, 20}, {12, 12}, {8, 17}, {13, 13},
1419 {3, 24}, {14, 14}, {4, 25}, {15, 15}, {3, 28}, {16, 16},
1420 };
1421
1422 InnerChain[Exp] = B.CreateFMul(getPow(InnerChain, AddChain[Exp][0], B),
1423 getPow(InnerChain, AddChain[Exp][1], B));
1424 return InnerChain[Exp];
1425 }
1426
1427 // Return a properly extended 32-bit integer if the operation is an itofp.
getIntToFPVal(Value * I2F,IRBuilderBase & B)1428 static Value *getIntToFPVal(Value *I2F, IRBuilderBase &B) {
1429 if (isa<SIToFPInst>(I2F) || isa<UIToFPInst>(I2F)) {
1430 Value *Op = cast<Instruction>(I2F)->getOperand(0);
1431 // Make sure that the exponent fits inside an int32_t,
1432 // thus avoiding any range issues that FP has not.
1433 unsigned BitWidth = Op->getType()->getPrimitiveSizeInBits();
1434 if (BitWidth < 32 ||
1435 (BitWidth == 32 && isa<SIToFPInst>(I2F)))
1436 return isa<SIToFPInst>(I2F) ? B.CreateSExt(Op, B.getInt32Ty())
1437 : B.CreateZExt(Op, B.getInt32Ty());
1438 }
1439
1440 return nullptr;
1441 }
1442
1443 /// Use exp{,2}(x * y) for pow(exp{,2}(x), y);
1444 /// ldexp(1.0, x) for pow(2.0, itofp(x)); exp2(n * x) for pow(2.0 ** n, x);
1445 /// exp10(x) for pow(10.0, x); exp2(log2(n) * x) for pow(n, x).
replacePowWithExp(CallInst * Pow,IRBuilderBase & B)1446 Value *LibCallSimplifier::replacePowWithExp(CallInst *Pow, IRBuilderBase &B) {
1447 Value *Base = Pow->getArgOperand(0), *Expo = Pow->getArgOperand(1);
1448 AttributeList Attrs; // Attributes are only meaningful on the original call
1449 Module *Mod = Pow->getModule();
1450 Type *Ty = Pow->getType();
1451 bool Ignored;
1452
1453 // Evaluate special cases related to a nested function as the base.
1454
1455 // pow(exp(x), y) -> exp(x * y)
1456 // pow(exp2(x), y) -> exp2(x * y)
1457 // If exp{,2}() is used only once, it is better to fold two transcendental
1458 // math functions into one. If used again, exp{,2}() would still have to be
1459 // called with the original argument, then keep both original transcendental
1460 // functions. However, this transformation is only safe with fully relaxed
1461 // math semantics, since, besides rounding differences, it changes overflow
1462 // and underflow behavior quite dramatically. For example:
1463 // pow(exp(1000), 0.001) = pow(inf, 0.001) = inf
1464 // Whereas:
1465 // exp(1000 * 0.001) = exp(1)
1466 // TODO: Loosen the requirement for fully relaxed math semantics.
1467 // TODO: Handle exp10() when more targets have it available.
1468 CallInst *BaseFn = dyn_cast<CallInst>(Base);
1469 if (BaseFn && BaseFn->hasOneUse() && BaseFn->isFast() && Pow->isFast()) {
1470 LibFunc LibFn;
1471
1472 Function *CalleeFn = BaseFn->getCalledFunction();
1473 if (CalleeFn &&
1474 TLI->getLibFunc(CalleeFn->getName(), LibFn) && TLI->has(LibFn)) {
1475 StringRef ExpName;
1476 Intrinsic::ID ID;
1477 Value *ExpFn;
1478 LibFunc LibFnFloat, LibFnDouble, LibFnLongDouble;
1479
1480 switch (LibFn) {
1481 default:
1482 return nullptr;
1483 case LibFunc_expf: case LibFunc_exp: case LibFunc_expl:
1484 ExpName = TLI->getName(LibFunc_exp);
1485 ID = Intrinsic::exp;
1486 LibFnFloat = LibFunc_expf;
1487 LibFnDouble = LibFunc_exp;
1488 LibFnLongDouble = LibFunc_expl;
1489 break;
1490 case LibFunc_exp2f: case LibFunc_exp2: case LibFunc_exp2l:
1491 ExpName = TLI->getName(LibFunc_exp2);
1492 ID = Intrinsic::exp2;
1493 LibFnFloat = LibFunc_exp2f;
1494 LibFnDouble = LibFunc_exp2;
1495 LibFnLongDouble = LibFunc_exp2l;
1496 break;
1497 }
1498
1499 // Create new exp{,2}() with the product as its argument.
1500 Value *FMul = B.CreateFMul(BaseFn->getArgOperand(0), Expo, "mul");
1501 ExpFn = BaseFn->doesNotAccessMemory()
1502 ? B.CreateCall(Intrinsic::getDeclaration(Mod, ID, Ty),
1503 FMul, ExpName)
1504 : emitUnaryFloatFnCall(FMul, TLI, LibFnDouble, LibFnFloat,
1505 LibFnLongDouble, B,
1506 BaseFn->getAttributes());
1507
1508 // Since the new exp{,2}() is different from the original one, dead code
1509 // elimination cannot be trusted to remove it, since it may have side
1510 // effects (e.g., errno). When the only consumer for the original
1511 // exp{,2}() is pow(), then it has to be explicitly erased.
1512 substituteInParent(BaseFn, ExpFn);
1513 return ExpFn;
1514 }
1515 }
1516
1517 // Evaluate special cases related to a constant base.
1518
1519 const APFloat *BaseF;
1520 if (!match(Pow->getArgOperand(0), m_APFloat(BaseF)))
1521 return nullptr;
1522
1523 // pow(2.0, itofp(x)) -> ldexp(1.0, x)
1524 if (match(Base, m_SpecificFP(2.0)) &&
1525 (isa<SIToFPInst>(Expo) || isa<UIToFPInst>(Expo)) &&
1526 hasFloatFn(TLI, Ty, LibFunc_ldexp, LibFunc_ldexpf, LibFunc_ldexpl)) {
1527 if (Value *ExpoI = getIntToFPVal(Expo, B))
1528 return emitBinaryFloatFnCall(ConstantFP::get(Ty, 1.0), ExpoI, TLI,
1529 LibFunc_ldexp, LibFunc_ldexpf, LibFunc_ldexpl,
1530 B, Attrs);
1531 }
1532
1533 // pow(2.0 ** n, x) -> exp2(n * x)
1534 if (hasFloatFn(TLI, Ty, LibFunc_exp2, LibFunc_exp2f, LibFunc_exp2l)) {
1535 APFloat BaseR = APFloat(1.0);
1536 BaseR.convert(BaseF->getSemantics(), APFloat::rmTowardZero, &Ignored);
1537 BaseR = BaseR / *BaseF;
1538 bool IsInteger = BaseF->isInteger(), IsReciprocal = BaseR.isInteger();
1539 const APFloat *NF = IsReciprocal ? &BaseR : BaseF;
1540 APSInt NI(64, false);
1541 if ((IsInteger || IsReciprocal) &&
1542 NF->convertToInteger(NI, APFloat::rmTowardZero, &Ignored) ==
1543 APFloat::opOK &&
1544 NI > 1 && NI.isPowerOf2()) {
1545 double N = NI.logBase2() * (IsReciprocal ? -1.0 : 1.0);
1546 Value *FMul = B.CreateFMul(Expo, ConstantFP::get(Ty, N), "mul");
1547 if (Pow->doesNotAccessMemory())
1548 return B.CreateCall(Intrinsic::getDeclaration(Mod, Intrinsic::exp2, Ty),
1549 FMul, "exp2");
1550 else
1551 return emitUnaryFloatFnCall(FMul, TLI, LibFunc_exp2, LibFunc_exp2f,
1552 LibFunc_exp2l, B, Attrs);
1553 }
1554 }
1555
1556 // pow(10.0, x) -> exp10(x)
1557 // TODO: There is no exp10() intrinsic yet, but some day there shall be one.
1558 if (match(Base, m_SpecificFP(10.0)) &&
1559 hasFloatFn(TLI, Ty, LibFunc_exp10, LibFunc_exp10f, LibFunc_exp10l))
1560 return emitUnaryFloatFnCall(Expo, TLI, LibFunc_exp10, LibFunc_exp10f,
1561 LibFunc_exp10l, B, Attrs);
1562
1563 // pow(x, y) -> exp2(log2(x) * y)
1564 if (Pow->hasApproxFunc() && Pow->hasNoNaNs() && BaseF->isFiniteNonZero() &&
1565 !BaseF->isNegative()) {
1566 // pow(1, inf) is defined to be 1 but exp2(log2(1) * inf) evaluates to NaN.
1567 // Luckily optimizePow has already handled the x == 1 case.
1568 assert(!match(Base, m_FPOne()) &&
1569 "pow(1.0, y) should have been simplified earlier!");
1570
1571 Value *Log = nullptr;
1572 if (Ty->isFloatTy())
1573 Log = ConstantFP::get(Ty, std::log2(BaseF->convertToFloat()));
1574 else if (Ty->isDoubleTy())
1575 Log = ConstantFP::get(Ty, std::log2(BaseF->convertToDouble()));
1576
1577 if (Log) {
1578 Value *FMul = B.CreateFMul(Log, Expo, "mul");
1579 if (Pow->doesNotAccessMemory())
1580 return B.CreateCall(Intrinsic::getDeclaration(Mod, Intrinsic::exp2, Ty),
1581 FMul, "exp2");
1582 else if (hasFloatFn(TLI, Ty, LibFunc_exp2, LibFunc_exp2f, LibFunc_exp2l))
1583 return emitUnaryFloatFnCall(FMul, TLI, LibFunc_exp2, LibFunc_exp2f,
1584 LibFunc_exp2l, B, Attrs);
1585 }
1586 }
1587
1588 return nullptr;
1589 }
1590
getSqrtCall(Value * V,AttributeList Attrs,bool NoErrno,Module * M,IRBuilderBase & B,const TargetLibraryInfo * TLI)1591 static Value *getSqrtCall(Value *V, AttributeList Attrs, bool NoErrno,
1592 Module *M, IRBuilderBase &B,
1593 const TargetLibraryInfo *TLI) {
1594 // If errno is never set, then use the intrinsic for sqrt().
1595 if (NoErrno) {
1596 Function *SqrtFn =
1597 Intrinsic::getDeclaration(M, Intrinsic::sqrt, V->getType());
1598 return B.CreateCall(SqrtFn, V, "sqrt");
1599 }
1600
1601 // Otherwise, use the libcall for sqrt().
1602 if (hasFloatFn(TLI, V->getType(), LibFunc_sqrt, LibFunc_sqrtf, LibFunc_sqrtl))
1603 // TODO: We also should check that the target can in fact lower the sqrt()
1604 // libcall. We currently have no way to ask this question, so we ask if
1605 // the target has a sqrt() libcall, which is not exactly the same.
1606 return emitUnaryFloatFnCall(V, TLI, LibFunc_sqrt, LibFunc_sqrtf,
1607 LibFunc_sqrtl, B, Attrs);
1608
1609 return nullptr;
1610 }
1611
1612 /// Use square root in place of pow(x, +/-0.5).
replacePowWithSqrt(CallInst * Pow,IRBuilderBase & B)1613 Value *LibCallSimplifier::replacePowWithSqrt(CallInst *Pow, IRBuilderBase &B) {
1614 Value *Sqrt, *Base = Pow->getArgOperand(0), *Expo = Pow->getArgOperand(1);
1615 AttributeList Attrs; // Attributes are only meaningful on the original call
1616 Module *Mod = Pow->getModule();
1617 Type *Ty = Pow->getType();
1618
1619 const APFloat *ExpoF;
1620 if (!match(Expo, m_APFloat(ExpoF)) ||
1621 (!ExpoF->isExactlyValue(0.5) && !ExpoF->isExactlyValue(-0.5)))
1622 return nullptr;
1623
1624 // Converting pow(X, -0.5) to 1/sqrt(X) may introduce an extra rounding step,
1625 // so that requires fast-math-flags (afn or reassoc).
1626 if (ExpoF->isNegative() && (!Pow->hasApproxFunc() && !Pow->hasAllowReassoc()))
1627 return nullptr;
1628
1629 // If we have a pow() library call (accesses memory) and we can't guarantee
1630 // that the base is not an infinity, give up:
1631 // pow(-Inf, 0.5) is optionally required to have a result of +Inf (not setting
1632 // errno), but sqrt(-Inf) is required by various standards to set errno.
1633 if (!Pow->doesNotAccessMemory() && !Pow->hasNoInfs() &&
1634 !isKnownNeverInfinity(Base, TLI))
1635 return nullptr;
1636
1637 Sqrt = getSqrtCall(Base, Attrs, Pow->doesNotAccessMemory(), Mod, B, TLI);
1638 if (!Sqrt)
1639 return nullptr;
1640
1641 // Handle signed zero base by expanding to fabs(sqrt(x)).
1642 if (!Pow->hasNoSignedZeros()) {
1643 Function *FAbsFn = Intrinsic::getDeclaration(Mod, Intrinsic::fabs, Ty);
1644 Sqrt = B.CreateCall(FAbsFn, Sqrt, "abs");
1645 }
1646
1647 // Handle non finite base by expanding to
1648 // (x == -infinity ? +infinity : sqrt(x)).
1649 if (!Pow->hasNoInfs()) {
1650 Value *PosInf = ConstantFP::getInfinity(Ty),
1651 *NegInf = ConstantFP::getInfinity(Ty, true);
1652 Value *FCmp = B.CreateFCmpOEQ(Base, NegInf, "isinf");
1653 Sqrt = B.CreateSelect(FCmp, PosInf, Sqrt);
1654 }
1655
1656 // If the exponent is negative, then get the reciprocal.
1657 if (ExpoF->isNegative())
1658 Sqrt = B.CreateFDiv(ConstantFP::get(Ty, 1.0), Sqrt, "reciprocal");
1659
1660 return Sqrt;
1661 }
1662
createPowWithIntegerExponent(Value * Base,Value * Expo,Module * M,IRBuilderBase & B)1663 static Value *createPowWithIntegerExponent(Value *Base, Value *Expo, Module *M,
1664 IRBuilderBase &B) {
1665 Value *Args[] = {Base, Expo};
1666 Function *F = Intrinsic::getDeclaration(M, Intrinsic::powi, Base->getType());
1667 return B.CreateCall(F, Args);
1668 }
1669
optimizePow(CallInst * Pow,IRBuilderBase & B)1670 Value *LibCallSimplifier::optimizePow(CallInst *Pow, IRBuilderBase &B) {
1671 Value *Base = Pow->getArgOperand(0);
1672 Value *Expo = Pow->getArgOperand(1);
1673 Function *Callee = Pow->getCalledFunction();
1674 StringRef Name = Callee->getName();
1675 Type *Ty = Pow->getType();
1676 Module *M = Pow->getModule();
1677 bool AllowApprox = Pow->hasApproxFunc();
1678 bool Ignored;
1679
1680 // Propagate the math semantics from the call to any created instructions.
1681 IRBuilderBase::FastMathFlagGuard Guard(B);
1682 B.setFastMathFlags(Pow->getFastMathFlags());
1683 // Evaluate special cases related to the base.
1684
1685 // pow(1.0, x) -> 1.0
1686 if (match(Base, m_FPOne()))
1687 return Base;
1688
1689 if (Value *Exp = replacePowWithExp(Pow, B))
1690 return Exp;
1691
1692 // Evaluate special cases related to the exponent.
1693
1694 // pow(x, -1.0) -> 1.0 / x
1695 if (match(Expo, m_SpecificFP(-1.0)))
1696 return B.CreateFDiv(ConstantFP::get(Ty, 1.0), Base, "reciprocal");
1697
1698 // pow(x, +/-0.0) -> 1.0
1699 if (match(Expo, m_AnyZeroFP()))
1700 return ConstantFP::get(Ty, 1.0);
1701
1702 // pow(x, 1.0) -> x
1703 if (match(Expo, m_FPOne()))
1704 return Base;
1705
1706 // pow(x, 2.0) -> x * x
1707 if (match(Expo, m_SpecificFP(2.0)))
1708 return B.CreateFMul(Base, Base, "square");
1709
1710 if (Value *Sqrt = replacePowWithSqrt(Pow, B))
1711 return Sqrt;
1712
1713 // pow(x, n) -> x * x * x * ...
1714 const APFloat *ExpoF;
1715 if (AllowApprox && match(Expo, m_APFloat(ExpoF)) &&
1716 !ExpoF->isExactlyValue(0.5) && !ExpoF->isExactlyValue(-0.5)) {
1717 // We limit to a max of 7 multiplications, thus the maximum exponent is 32.
1718 // If the exponent is an integer+0.5 we generate a call to sqrt and an
1719 // additional fmul.
1720 // TODO: This whole transformation should be backend specific (e.g. some
1721 // backends might prefer libcalls or the limit for the exponent might
1722 // be different) and it should also consider optimizing for size.
1723 APFloat LimF(ExpoF->getSemantics(), 33),
1724 ExpoA(abs(*ExpoF));
1725 if (ExpoA < LimF) {
1726 // This transformation applies to integer or integer+0.5 exponents only.
1727 // For integer+0.5, we create a sqrt(Base) call.
1728 Value *Sqrt = nullptr;
1729 if (!ExpoA.isInteger()) {
1730 APFloat Expo2 = ExpoA;
1731 // To check if ExpoA is an integer + 0.5, we add it to itself. If there
1732 // is no floating point exception and the result is an integer, then
1733 // ExpoA == integer + 0.5
1734 if (Expo2.add(ExpoA, APFloat::rmNearestTiesToEven) != APFloat::opOK)
1735 return nullptr;
1736
1737 if (!Expo2.isInteger())
1738 return nullptr;
1739
1740 Sqrt = getSqrtCall(Base, Pow->getCalledFunction()->getAttributes(),
1741 Pow->doesNotAccessMemory(), M, B, TLI);
1742 if (!Sqrt)
1743 return nullptr;
1744 }
1745
1746 // We will memoize intermediate products of the Addition Chain.
1747 Value *InnerChain[33] = {nullptr};
1748 InnerChain[1] = Base;
1749 InnerChain[2] = B.CreateFMul(Base, Base, "square");
1750
1751 // We cannot readily convert a non-double type (like float) to a double.
1752 // So we first convert it to something which could be converted to double.
1753 ExpoA.convert(APFloat::IEEEdouble(), APFloat::rmTowardZero, &Ignored);
1754 Value *FMul = getPow(InnerChain, ExpoA.convertToDouble(), B);
1755
1756 // Expand pow(x, y+0.5) to pow(x, y) * sqrt(x).
1757 if (Sqrt)
1758 FMul = B.CreateFMul(FMul, Sqrt);
1759
1760 // If the exponent is negative, then get the reciprocal.
1761 if (ExpoF->isNegative())
1762 FMul = B.CreateFDiv(ConstantFP::get(Ty, 1.0), FMul, "reciprocal");
1763
1764 return FMul;
1765 }
1766
1767 APSInt IntExpo(32, /*isUnsigned=*/false);
1768 // powf(x, n) -> powi(x, n) if n is a constant signed integer value
1769 if (ExpoF->isInteger() &&
1770 ExpoF->convertToInteger(IntExpo, APFloat::rmTowardZero, &Ignored) ==
1771 APFloat::opOK) {
1772 return createPowWithIntegerExponent(
1773 Base, ConstantInt::get(B.getInt32Ty(), IntExpo), M, B);
1774 }
1775 }
1776
1777 // powf(x, itofp(y)) -> powi(x, y)
1778 if (AllowApprox && (isa<SIToFPInst>(Expo) || isa<UIToFPInst>(Expo))) {
1779 if (Value *ExpoI = getIntToFPVal(Expo, B))
1780 return createPowWithIntegerExponent(Base, ExpoI, M, B);
1781 }
1782
1783 // Shrink pow() to powf() if the arguments are single precision,
1784 // unless the result is expected to be double precision.
1785 if (UnsafeFPShrink && Name == TLI->getName(LibFunc_pow) &&
1786 hasFloatVersion(Name)) {
1787 if (Value *Shrunk = optimizeBinaryDoubleFP(Pow, B, true))
1788 return Shrunk;
1789 }
1790
1791 return nullptr;
1792 }
1793
optimizeExp2(CallInst * CI,IRBuilderBase & B)1794 Value *LibCallSimplifier::optimizeExp2(CallInst *CI, IRBuilderBase &B) {
1795 Function *Callee = CI->getCalledFunction();
1796 AttributeList Attrs; // Attributes are only meaningful on the original call
1797 StringRef Name = Callee->getName();
1798 Value *Ret = nullptr;
1799 if (UnsafeFPShrink && Name == TLI->getName(LibFunc_exp2) &&
1800 hasFloatVersion(Name))
1801 Ret = optimizeUnaryDoubleFP(CI, B, true);
1802
1803 Type *Ty = CI->getType();
1804 Value *Op = CI->getArgOperand(0);
1805
1806 // Turn exp2(sitofp(x)) -> ldexp(1.0, sext(x)) if sizeof(x) <= 32
1807 // Turn exp2(uitofp(x)) -> ldexp(1.0, zext(x)) if sizeof(x) < 32
1808 if ((isa<SIToFPInst>(Op) || isa<UIToFPInst>(Op)) &&
1809 hasFloatFn(TLI, Ty, LibFunc_ldexp, LibFunc_ldexpf, LibFunc_ldexpl)) {
1810 if (Value *Exp = getIntToFPVal(Op, B))
1811 return emitBinaryFloatFnCall(ConstantFP::get(Ty, 1.0), Exp, TLI,
1812 LibFunc_ldexp, LibFunc_ldexpf, LibFunc_ldexpl,
1813 B, Attrs);
1814 }
1815
1816 return Ret;
1817 }
1818
optimizeFMinFMax(CallInst * CI,IRBuilderBase & B)1819 Value *LibCallSimplifier::optimizeFMinFMax(CallInst *CI, IRBuilderBase &B) {
1820 // If we can shrink the call to a float function rather than a double
1821 // function, do that first.
1822 Function *Callee = CI->getCalledFunction();
1823 StringRef Name = Callee->getName();
1824 if ((Name == "fmin" || Name == "fmax") && hasFloatVersion(Name))
1825 if (Value *Ret = optimizeBinaryDoubleFP(CI, B))
1826 return Ret;
1827
1828 // The LLVM intrinsics minnum/maxnum correspond to fmin/fmax. Canonicalize to
1829 // the intrinsics for improved optimization (for example, vectorization).
1830 // No-signed-zeros is implied by the definitions of fmax/fmin themselves.
1831 // From the C standard draft WG14/N1256:
1832 // "Ideally, fmax would be sensitive to the sign of zero, for example
1833 // fmax(-0.0, +0.0) would return +0; however, implementation in software
1834 // might be impractical."
1835 IRBuilderBase::FastMathFlagGuard Guard(B);
1836 FastMathFlags FMF = CI->getFastMathFlags();
1837 FMF.setNoSignedZeros();
1838 B.setFastMathFlags(FMF);
1839
1840 Intrinsic::ID IID = Callee->getName().startswith("fmin") ? Intrinsic::minnum
1841 : Intrinsic::maxnum;
1842 Function *F = Intrinsic::getDeclaration(CI->getModule(), IID, CI->getType());
1843 return B.CreateCall(F, { CI->getArgOperand(0), CI->getArgOperand(1) });
1844 }
1845
optimizeLog(CallInst * Log,IRBuilderBase & B)1846 Value *LibCallSimplifier::optimizeLog(CallInst *Log, IRBuilderBase &B) {
1847 Function *LogFn = Log->getCalledFunction();
1848 AttributeList Attrs; // Attributes are only meaningful on the original call
1849 StringRef LogNm = LogFn->getName();
1850 Intrinsic::ID LogID = LogFn->getIntrinsicID();
1851 Module *Mod = Log->getModule();
1852 Type *Ty = Log->getType();
1853 Value *Ret = nullptr;
1854
1855 if (UnsafeFPShrink && hasFloatVersion(LogNm))
1856 Ret = optimizeUnaryDoubleFP(Log, B, true);
1857
1858 // The earlier call must also be 'fast' in order to do these transforms.
1859 CallInst *Arg = dyn_cast<CallInst>(Log->getArgOperand(0));
1860 if (!Log->isFast() || !Arg || !Arg->isFast() || !Arg->hasOneUse())
1861 return Ret;
1862
1863 LibFunc LogLb, ExpLb, Exp2Lb, Exp10Lb, PowLb;
1864
1865 // This is only applicable to log(), log2(), log10().
1866 if (TLI->getLibFunc(LogNm, LogLb))
1867 switch (LogLb) {
1868 case LibFunc_logf:
1869 LogID = Intrinsic::log;
1870 ExpLb = LibFunc_expf;
1871 Exp2Lb = LibFunc_exp2f;
1872 Exp10Lb = LibFunc_exp10f;
1873 PowLb = LibFunc_powf;
1874 break;
1875 case LibFunc_log:
1876 LogID = Intrinsic::log;
1877 ExpLb = LibFunc_exp;
1878 Exp2Lb = LibFunc_exp2;
1879 Exp10Lb = LibFunc_exp10;
1880 PowLb = LibFunc_pow;
1881 break;
1882 case LibFunc_logl:
1883 LogID = Intrinsic::log;
1884 ExpLb = LibFunc_expl;
1885 Exp2Lb = LibFunc_exp2l;
1886 Exp10Lb = LibFunc_exp10l;
1887 PowLb = LibFunc_powl;
1888 break;
1889 case LibFunc_log2f:
1890 LogID = Intrinsic::log2;
1891 ExpLb = LibFunc_expf;
1892 Exp2Lb = LibFunc_exp2f;
1893 Exp10Lb = LibFunc_exp10f;
1894 PowLb = LibFunc_powf;
1895 break;
1896 case LibFunc_log2:
1897 LogID = Intrinsic::log2;
1898 ExpLb = LibFunc_exp;
1899 Exp2Lb = LibFunc_exp2;
1900 Exp10Lb = LibFunc_exp10;
1901 PowLb = LibFunc_pow;
1902 break;
1903 case LibFunc_log2l:
1904 LogID = Intrinsic::log2;
1905 ExpLb = LibFunc_expl;
1906 Exp2Lb = LibFunc_exp2l;
1907 Exp10Lb = LibFunc_exp10l;
1908 PowLb = LibFunc_powl;
1909 break;
1910 case LibFunc_log10f:
1911 LogID = Intrinsic::log10;
1912 ExpLb = LibFunc_expf;
1913 Exp2Lb = LibFunc_exp2f;
1914 Exp10Lb = LibFunc_exp10f;
1915 PowLb = LibFunc_powf;
1916 break;
1917 case LibFunc_log10:
1918 LogID = Intrinsic::log10;
1919 ExpLb = LibFunc_exp;
1920 Exp2Lb = LibFunc_exp2;
1921 Exp10Lb = LibFunc_exp10;
1922 PowLb = LibFunc_pow;
1923 break;
1924 case LibFunc_log10l:
1925 LogID = Intrinsic::log10;
1926 ExpLb = LibFunc_expl;
1927 Exp2Lb = LibFunc_exp2l;
1928 Exp10Lb = LibFunc_exp10l;
1929 PowLb = LibFunc_powl;
1930 break;
1931 default:
1932 return Ret;
1933 }
1934 else if (LogID == Intrinsic::log || LogID == Intrinsic::log2 ||
1935 LogID == Intrinsic::log10) {
1936 if (Ty->getScalarType()->isFloatTy()) {
1937 ExpLb = LibFunc_expf;
1938 Exp2Lb = LibFunc_exp2f;
1939 Exp10Lb = LibFunc_exp10f;
1940 PowLb = LibFunc_powf;
1941 } else if (Ty->getScalarType()->isDoubleTy()) {
1942 ExpLb = LibFunc_exp;
1943 Exp2Lb = LibFunc_exp2;
1944 Exp10Lb = LibFunc_exp10;
1945 PowLb = LibFunc_pow;
1946 } else
1947 return Ret;
1948 } else
1949 return Ret;
1950
1951 IRBuilderBase::FastMathFlagGuard Guard(B);
1952 B.setFastMathFlags(FastMathFlags::getFast());
1953
1954 Intrinsic::ID ArgID = Arg->getIntrinsicID();
1955 LibFunc ArgLb = NotLibFunc;
1956 TLI->getLibFunc(*Arg, ArgLb);
1957
1958 // log(pow(x,y)) -> y*log(x)
1959 if (ArgLb == PowLb || ArgID == Intrinsic::pow) {
1960 Value *LogX =
1961 Log->doesNotAccessMemory()
1962 ? B.CreateCall(Intrinsic::getDeclaration(Mod, LogID, Ty),
1963 Arg->getOperand(0), "log")
1964 : emitUnaryFloatFnCall(Arg->getOperand(0), LogNm, B, Attrs);
1965 Value *MulY = B.CreateFMul(Arg->getArgOperand(1), LogX, "mul");
1966 // Since pow() may have side effects, e.g. errno,
1967 // dead code elimination may not be trusted to remove it.
1968 substituteInParent(Arg, MulY);
1969 return MulY;
1970 }
1971
1972 // log(exp{,2,10}(y)) -> y*log({e,2,10})
1973 // TODO: There is no exp10() intrinsic yet.
1974 if (ArgLb == ExpLb || ArgLb == Exp2Lb || ArgLb == Exp10Lb ||
1975 ArgID == Intrinsic::exp || ArgID == Intrinsic::exp2) {
1976 Constant *Eul;
1977 if (ArgLb == ExpLb || ArgID == Intrinsic::exp)
1978 // FIXME: Add more precise value of e for long double.
1979 Eul = ConstantFP::get(Log->getType(), numbers::e);
1980 else if (ArgLb == Exp2Lb || ArgID == Intrinsic::exp2)
1981 Eul = ConstantFP::get(Log->getType(), 2.0);
1982 else
1983 Eul = ConstantFP::get(Log->getType(), 10.0);
1984 Value *LogE = Log->doesNotAccessMemory()
1985 ? B.CreateCall(Intrinsic::getDeclaration(Mod, LogID, Ty),
1986 Eul, "log")
1987 : emitUnaryFloatFnCall(Eul, LogNm, B, Attrs);
1988 Value *MulY = B.CreateFMul(Arg->getArgOperand(0), LogE, "mul");
1989 // Since exp() may have side effects, e.g. errno,
1990 // dead code elimination may not be trusted to remove it.
1991 substituteInParent(Arg, MulY);
1992 return MulY;
1993 }
1994
1995 return Ret;
1996 }
1997
optimizeSqrt(CallInst * CI,IRBuilderBase & B)1998 Value *LibCallSimplifier::optimizeSqrt(CallInst *CI, IRBuilderBase &B) {
1999 Function *Callee = CI->getCalledFunction();
2000 Value *Ret = nullptr;
2001 // TODO: Once we have a way (other than checking for the existince of the
2002 // libcall) to tell whether our target can lower @llvm.sqrt, relax the
2003 // condition below.
2004 if (TLI->has(LibFunc_sqrtf) && (Callee->getName() == "sqrt" ||
2005 Callee->getIntrinsicID() == Intrinsic::sqrt))
2006 Ret = optimizeUnaryDoubleFP(CI, B, true);
2007
2008 if (!CI->isFast())
2009 return Ret;
2010
2011 Instruction *I = dyn_cast<Instruction>(CI->getArgOperand(0));
2012 if (!I || I->getOpcode() != Instruction::FMul || !I->isFast())
2013 return Ret;
2014
2015 // We're looking for a repeated factor in a multiplication tree,
2016 // so we can do this fold: sqrt(x * x) -> fabs(x);
2017 // or this fold: sqrt((x * x) * y) -> fabs(x) * sqrt(y).
2018 Value *Op0 = I->getOperand(0);
2019 Value *Op1 = I->getOperand(1);
2020 Value *RepeatOp = nullptr;
2021 Value *OtherOp = nullptr;
2022 if (Op0 == Op1) {
2023 // Simple match: the operands of the multiply are identical.
2024 RepeatOp = Op0;
2025 } else {
2026 // Look for a more complicated pattern: one of the operands is itself
2027 // a multiply, so search for a common factor in that multiply.
2028 // Note: We don't bother looking any deeper than this first level or for
2029 // variations of this pattern because instcombine's visitFMUL and/or the
2030 // reassociation pass should give us this form.
2031 Value *OtherMul0, *OtherMul1;
2032 if (match(Op0, m_FMul(m_Value(OtherMul0), m_Value(OtherMul1)))) {
2033 // Pattern: sqrt((x * y) * z)
2034 if (OtherMul0 == OtherMul1 && cast<Instruction>(Op0)->isFast()) {
2035 // Matched: sqrt((x * x) * z)
2036 RepeatOp = OtherMul0;
2037 OtherOp = Op1;
2038 }
2039 }
2040 }
2041 if (!RepeatOp)
2042 return Ret;
2043
2044 // Fast math flags for any created instructions should match the sqrt
2045 // and multiply.
2046 IRBuilderBase::FastMathFlagGuard Guard(B);
2047 B.setFastMathFlags(I->getFastMathFlags());
2048
2049 // If we found a repeated factor, hoist it out of the square root and
2050 // replace it with the fabs of that factor.
2051 Module *M = Callee->getParent();
2052 Type *ArgType = I->getType();
2053 Function *Fabs = Intrinsic::getDeclaration(M, Intrinsic::fabs, ArgType);
2054 Value *FabsCall = B.CreateCall(Fabs, RepeatOp, "fabs");
2055 if (OtherOp) {
2056 // If we found a non-repeated factor, we still need to get its square
2057 // root. We then multiply that by the value that was simplified out
2058 // of the square root calculation.
2059 Function *Sqrt = Intrinsic::getDeclaration(M, Intrinsic::sqrt, ArgType);
2060 Value *SqrtCall = B.CreateCall(Sqrt, OtherOp, "sqrt");
2061 return B.CreateFMul(FabsCall, SqrtCall);
2062 }
2063 return FabsCall;
2064 }
2065
2066 // TODO: Generalize to handle any trig function and its inverse.
optimizeTan(CallInst * CI,IRBuilderBase & B)2067 Value *LibCallSimplifier::optimizeTan(CallInst *CI, IRBuilderBase &B) {
2068 Function *Callee = CI->getCalledFunction();
2069 Value *Ret = nullptr;
2070 StringRef Name = Callee->getName();
2071 if (UnsafeFPShrink && Name == "tan" && hasFloatVersion(Name))
2072 Ret = optimizeUnaryDoubleFP(CI, B, true);
2073
2074 Value *Op1 = CI->getArgOperand(0);
2075 auto *OpC = dyn_cast<CallInst>(Op1);
2076 if (!OpC)
2077 return Ret;
2078
2079 // Both calls must be 'fast' in order to remove them.
2080 if (!CI->isFast() || !OpC->isFast())
2081 return Ret;
2082
2083 // tan(atan(x)) -> x
2084 // tanf(atanf(x)) -> x
2085 // tanl(atanl(x)) -> x
2086 LibFunc Func;
2087 Function *F = OpC->getCalledFunction();
2088 if (F && TLI->getLibFunc(F->getName(), Func) && TLI->has(Func) &&
2089 ((Func == LibFunc_atan && Callee->getName() == "tan") ||
2090 (Func == LibFunc_atanf && Callee->getName() == "tanf") ||
2091 (Func == LibFunc_atanl && Callee->getName() == "tanl")))
2092 Ret = OpC->getArgOperand(0);
2093 return Ret;
2094 }
2095
isTrigLibCall(CallInst * CI)2096 static bool isTrigLibCall(CallInst *CI) {
2097 // We can only hope to do anything useful if we can ignore things like errno
2098 // and floating-point exceptions.
2099 // We already checked the prototype.
2100 return CI->hasFnAttr(Attribute::NoUnwind) &&
2101 CI->hasFnAttr(Attribute::ReadNone);
2102 }
2103
insertSinCosCall(IRBuilderBase & B,Function * OrigCallee,Value * Arg,bool UseFloat,Value * & Sin,Value * & Cos,Value * & SinCos)2104 static void insertSinCosCall(IRBuilderBase &B, Function *OrigCallee, Value *Arg,
2105 bool UseFloat, Value *&Sin, Value *&Cos,
2106 Value *&SinCos) {
2107 Type *ArgTy = Arg->getType();
2108 Type *ResTy;
2109 StringRef Name;
2110
2111 Triple T(OrigCallee->getParent()->getTargetTriple());
2112 if (UseFloat) {
2113 Name = "__sincospif_stret";
2114
2115 assert(T.getArch() != Triple::x86 && "x86 messy and unsupported for now");
2116 // x86_64 can't use {float, float} since that would be returned in both
2117 // xmm0 and xmm1, which isn't what a real struct would do.
2118 ResTy = T.getArch() == Triple::x86_64
2119 ? static_cast<Type *>(FixedVectorType::get(ArgTy, 2))
2120 : static_cast<Type *>(StructType::get(ArgTy, ArgTy));
2121 } else {
2122 Name = "__sincospi_stret";
2123 ResTy = StructType::get(ArgTy, ArgTy);
2124 }
2125
2126 Module *M = OrigCallee->getParent();
2127 FunctionCallee Callee =
2128 M->getOrInsertFunction(Name, OrigCallee->getAttributes(), ResTy, ArgTy);
2129
2130 if (Instruction *ArgInst = dyn_cast<Instruction>(Arg)) {
2131 // If the argument is an instruction, it must dominate all uses so put our
2132 // sincos call there.
2133 B.SetInsertPoint(ArgInst->getParent(), ++ArgInst->getIterator());
2134 } else {
2135 // Otherwise (e.g. for a constant) the beginning of the function is as
2136 // good a place as any.
2137 BasicBlock &EntryBB = B.GetInsertBlock()->getParent()->getEntryBlock();
2138 B.SetInsertPoint(&EntryBB, EntryBB.begin());
2139 }
2140
2141 SinCos = B.CreateCall(Callee, Arg, "sincospi");
2142
2143 if (SinCos->getType()->isStructTy()) {
2144 Sin = B.CreateExtractValue(SinCos, 0, "sinpi");
2145 Cos = B.CreateExtractValue(SinCos, 1, "cospi");
2146 } else {
2147 Sin = B.CreateExtractElement(SinCos, ConstantInt::get(B.getInt32Ty(), 0),
2148 "sinpi");
2149 Cos = B.CreateExtractElement(SinCos, ConstantInt::get(B.getInt32Ty(), 1),
2150 "cospi");
2151 }
2152 }
2153
optimizeSinCosPi(CallInst * CI,IRBuilderBase & B)2154 Value *LibCallSimplifier::optimizeSinCosPi(CallInst *CI, IRBuilderBase &B) {
2155 // Make sure the prototype is as expected, otherwise the rest of the
2156 // function is probably invalid and likely to abort.
2157 if (!isTrigLibCall(CI))
2158 return nullptr;
2159
2160 Value *Arg = CI->getArgOperand(0);
2161 SmallVector<CallInst *, 1> SinCalls;
2162 SmallVector<CallInst *, 1> CosCalls;
2163 SmallVector<CallInst *, 1> SinCosCalls;
2164
2165 bool IsFloat = Arg->getType()->isFloatTy();
2166
2167 // Look for all compatible sinpi, cospi and sincospi calls with the same
2168 // argument. If there are enough (in some sense) we can make the
2169 // substitution.
2170 Function *F = CI->getFunction();
2171 for (User *U : Arg->users())
2172 classifyArgUse(U, F, IsFloat, SinCalls, CosCalls, SinCosCalls);
2173
2174 // It's only worthwhile if both sinpi and cospi are actually used.
2175 if (SinCalls.empty() || CosCalls.empty())
2176 return nullptr;
2177
2178 Value *Sin, *Cos, *SinCos;
2179 insertSinCosCall(B, CI->getCalledFunction(), Arg, IsFloat, Sin, Cos, SinCos);
2180
2181 auto replaceTrigInsts = [this](SmallVectorImpl<CallInst *> &Calls,
2182 Value *Res) {
2183 for (CallInst *C : Calls)
2184 replaceAllUsesWith(C, Res);
2185 };
2186
2187 replaceTrigInsts(SinCalls, Sin);
2188 replaceTrigInsts(CosCalls, Cos);
2189 replaceTrigInsts(SinCosCalls, SinCos);
2190
2191 return nullptr;
2192 }
2193
classifyArgUse(Value * Val,Function * F,bool IsFloat,SmallVectorImpl<CallInst * > & SinCalls,SmallVectorImpl<CallInst * > & CosCalls,SmallVectorImpl<CallInst * > & SinCosCalls)2194 void LibCallSimplifier::classifyArgUse(
2195 Value *Val, Function *F, bool IsFloat,
2196 SmallVectorImpl<CallInst *> &SinCalls,
2197 SmallVectorImpl<CallInst *> &CosCalls,
2198 SmallVectorImpl<CallInst *> &SinCosCalls) {
2199 CallInst *CI = dyn_cast<CallInst>(Val);
2200
2201 if (!CI || CI->use_empty())
2202 return;
2203
2204 // Don't consider calls in other functions.
2205 if (CI->getFunction() != F)
2206 return;
2207
2208 Function *Callee = CI->getCalledFunction();
2209 LibFunc Func;
2210 if (!Callee || !TLI->getLibFunc(*Callee, Func) || !TLI->has(Func) ||
2211 !isTrigLibCall(CI))
2212 return;
2213
2214 if (IsFloat) {
2215 if (Func == LibFunc_sinpif)
2216 SinCalls.push_back(CI);
2217 else if (Func == LibFunc_cospif)
2218 CosCalls.push_back(CI);
2219 else if (Func == LibFunc_sincospif_stret)
2220 SinCosCalls.push_back(CI);
2221 } else {
2222 if (Func == LibFunc_sinpi)
2223 SinCalls.push_back(CI);
2224 else if (Func == LibFunc_cospi)
2225 CosCalls.push_back(CI);
2226 else if (Func == LibFunc_sincospi_stret)
2227 SinCosCalls.push_back(CI);
2228 }
2229 }
2230
2231 //===----------------------------------------------------------------------===//
2232 // Integer Library Call Optimizations
2233 //===----------------------------------------------------------------------===//
2234
optimizeFFS(CallInst * CI,IRBuilderBase & B)2235 Value *LibCallSimplifier::optimizeFFS(CallInst *CI, IRBuilderBase &B) {
2236 // ffs(x) -> x != 0 ? (i32)llvm.cttz(x)+1 : 0
2237 Value *Op = CI->getArgOperand(0);
2238 Type *ArgType = Op->getType();
2239 Function *F = Intrinsic::getDeclaration(CI->getCalledFunction()->getParent(),
2240 Intrinsic::cttz, ArgType);
2241 Value *V = B.CreateCall(F, {Op, B.getTrue()}, "cttz");
2242 V = B.CreateAdd(V, ConstantInt::get(V->getType(), 1));
2243 V = B.CreateIntCast(V, B.getInt32Ty(), false);
2244
2245 Value *Cond = B.CreateICmpNE(Op, Constant::getNullValue(ArgType));
2246 return B.CreateSelect(Cond, V, B.getInt32(0));
2247 }
2248
optimizeFls(CallInst * CI,IRBuilderBase & B)2249 Value *LibCallSimplifier::optimizeFls(CallInst *CI, IRBuilderBase &B) {
2250 // fls(x) -> (i32)(sizeInBits(x) - llvm.ctlz(x, false))
2251 Value *Op = CI->getArgOperand(0);
2252 Type *ArgType = Op->getType();
2253 Function *F = Intrinsic::getDeclaration(CI->getCalledFunction()->getParent(),
2254 Intrinsic::ctlz, ArgType);
2255 Value *V = B.CreateCall(F, {Op, B.getFalse()}, "ctlz");
2256 V = B.CreateSub(ConstantInt::get(V->getType(), ArgType->getIntegerBitWidth()),
2257 V);
2258 return B.CreateIntCast(V, CI->getType(), false);
2259 }
2260
optimizeAbs(CallInst * CI,IRBuilderBase & B)2261 Value *LibCallSimplifier::optimizeAbs(CallInst *CI, IRBuilderBase &B) {
2262 // abs(x) -> x <s 0 ? -x : x
2263 // The negation has 'nsw' because abs of INT_MIN is undefined.
2264 Value *X = CI->getArgOperand(0);
2265 Value *IsNeg = B.CreateICmpSLT(X, Constant::getNullValue(X->getType()));
2266 Value *NegX = B.CreateNSWNeg(X, "neg");
2267 return B.CreateSelect(IsNeg, NegX, X);
2268 }
2269
optimizeIsDigit(CallInst * CI,IRBuilderBase & B)2270 Value *LibCallSimplifier::optimizeIsDigit(CallInst *CI, IRBuilderBase &B) {
2271 // isdigit(c) -> (c-'0') <u 10
2272 Value *Op = CI->getArgOperand(0);
2273 Op = B.CreateSub(Op, B.getInt32('0'), "isdigittmp");
2274 Op = B.CreateICmpULT(Op, B.getInt32(10), "isdigit");
2275 return B.CreateZExt(Op, CI->getType());
2276 }
2277
optimizeIsAscii(CallInst * CI,IRBuilderBase & B)2278 Value *LibCallSimplifier::optimizeIsAscii(CallInst *CI, IRBuilderBase &B) {
2279 // isascii(c) -> c <u 128
2280 Value *Op = CI->getArgOperand(0);
2281 Op = B.CreateICmpULT(Op, B.getInt32(128), "isascii");
2282 return B.CreateZExt(Op, CI->getType());
2283 }
2284
optimizeToAscii(CallInst * CI,IRBuilderBase & B)2285 Value *LibCallSimplifier::optimizeToAscii(CallInst *CI, IRBuilderBase &B) {
2286 // toascii(c) -> c & 0x7f
2287 return B.CreateAnd(CI->getArgOperand(0),
2288 ConstantInt::get(CI->getType(), 0x7F));
2289 }
2290
optimizeAtoi(CallInst * CI,IRBuilderBase & B)2291 Value *LibCallSimplifier::optimizeAtoi(CallInst *CI, IRBuilderBase &B) {
2292 StringRef Str;
2293 if (!getConstantStringInfo(CI->getArgOperand(0), Str))
2294 return nullptr;
2295
2296 return convertStrToNumber(CI, Str, 10);
2297 }
2298
optimizeStrtol(CallInst * CI,IRBuilderBase & B)2299 Value *LibCallSimplifier::optimizeStrtol(CallInst *CI, IRBuilderBase &B) {
2300 StringRef Str;
2301 if (!getConstantStringInfo(CI->getArgOperand(0), Str))
2302 return nullptr;
2303
2304 if (!isa<ConstantPointerNull>(CI->getArgOperand(1)))
2305 return nullptr;
2306
2307 if (ConstantInt *CInt = dyn_cast<ConstantInt>(CI->getArgOperand(2))) {
2308 return convertStrToNumber(CI, Str, CInt->getSExtValue());
2309 }
2310
2311 return nullptr;
2312 }
2313
2314 //===----------------------------------------------------------------------===//
2315 // Formatting and IO Library Call Optimizations
2316 //===----------------------------------------------------------------------===//
2317
2318 static bool isReportingError(Function *Callee, CallInst *CI, int StreamArg);
2319
optimizeErrorReporting(CallInst * CI,IRBuilderBase & B,int StreamArg)2320 Value *LibCallSimplifier::optimizeErrorReporting(CallInst *CI, IRBuilderBase &B,
2321 int StreamArg) {
2322 Function *Callee = CI->getCalledFunction();
2323 // Error reporting calls should be cold, mark them as such.
2324 // This applies even to non-builtin calls: it is only a hint and applies to
2325 // functions that the frontend might not understand as builtins.
2326
2327 // This heuristic was suggested in:
2328 // Improving Static Branch Prediction in a Compiler
2329 // Brian L. Deitrich, Ben-Chung Cheng, Wen-mei W. Hwu
2330 // Proceedings of PACT'98, Oct. 1998, IEEE
2331 if (!CI->hasFnAttr(Attribute::Cold) &&
2332 isReportingError(Callee, CI, StreamArg)) {
2333 CI->addAttribute(AttributeList::FunctionIndex, Attribute::Cold);
2334 }
2335
2336 return nullptr;
2337 }
2338
isReportingError(Function * Callee,CallInst * CI,int StreamArg)2339 static bool isReportingError(Function *Callee, CallInst *CI, int StreamArg) {
2340 if (!Callee || !Callee->isDeclaration())
2341 return false;
2342
2343 if (StreamArg < 0)
2344 return true;
2345
2346 // These functions might be considered cold, but only if their stream
2347 // argument is stderr.
2348
2349 if (StreamArg >= (int)CI->getNumArgOperands())
2350 return false;
2351 LoadInst *LI = dyn_cast<LoadInst>(CI->getArgOperand(StreamArg));
2352 if (!LI)
2353 return false;
2354 GlobalVariable *GV = dyn_cast<GlobalVariable>(LI->getPointerOperand());
2355 if (!GV || !GV->isDeclaration())
2356 return false;
2357 return GV->getName() == "stderr";
2358 }
2359
optimizePrintFString(CallInst * CI,IRBuilderBase & B)2360 Value *LibCallSimplifier::optimizePrintFString(CallInst *CI, IRBuilderBase &B) {
2361 // Check for a fixed format string.
2362 StringRef FormatStr;
2363 if (!getConstantStringInfo(CI->getArgOperand(0), FormatStr))
2364 return nullptr;
2365
2366 // Empty format string -> noop.
2367 if (FormatStr.empty()) // Tolerate printf's declared void.
2368 return CI->use_empty() ? (Value *)CI : ConstantInt::get(CI->getType(), 0);
2369
2370 // Do not do any of the following transformations if the printf return value
2371 // is used, in general the printf return value is not compatible with either
2372 // putchar() or puts().
2373 if (!CI->use_empty())
2374 return nullptr;
2375
2376 // printf("x") -> putchar('x'), even for "%" and "%%".
2377 if (FormatStr.size() == 1 || FormatStr == "%%")
2378 return emitPutChar(B.getInt32(FormatStr[0]), B, TLI);
2379
2380 // Try to remove call or emit putchar/puts.
2381 if (FormatStr == "%s" && CI->getNumArgOperands() > 1) {
2382 StringRef OperandStr;
2383 if (!getConstantStringInfo(CI->getOperand(1), OperandStr))
2384 return nullptr;
2385 // printf("%s", "") --> NOP
2386 if (OperandStr.empty())
2387 return (Value *)CI;
2388 // printf("%s", "a") --> putchar('a')
2389 if (OperandStr.size() == 1)
2390 return emitPutChar(B.getInt32(OperandStr[0]), B, TLI);
2391 // printf("%s", str"\n") --> puts(str)
2392 if (OperandStr.back() == '\n') {
2393 OperandStr = OperandStr.drop_back();
2394 Value *GV = B.CreateGlobalString(OperandStr, "str");
2395 return emitPutS(GV, B, TLI);
2396 }
2397 return nullptr;
2398 }
2399
2400 // printf("foo\n") --> puts("foo")
2401 if (FormatStr.back() == '\n' &&
2402 FormatStr.find('%') == StringRef::npos) { // No format characters.
2403 // Create a string literal with no \n on it. We expect the constant merge
2404 // pass to be run after this pass, to merge duplicate strings.
2405 FormatStr = FormatStr.drop_back();
2406 Value *GV = B.CreateGlobalString(FormatStr, "str");
2407 return emitPutS(GV, B, TLI);
2408 }
2409
2410 // Optimize specific format strings.
2411 // printf("%c", chr) --> putchar(chr)
2412 if (FormatStr == "%c" && CI->getNumArgOperands() > 1 &&
2413 CI->getArgOperand(1)->getType()->isIntegerTy())
2414 return emitPutChar(CI->getArgOperand(1), B, TLI);
2415
2416 // printf("%s\n", str) --> puts(str)
2417 if (FormatStr == "%s\n" && CI->getNumArgOperands() > 1 &&
2418 CI->getArgOperand(1)->getType()->isPointerTy())
2419 return emitPutS(CI->getArgOperand(1), B, TLI);
2420 return nullptr;
2421 }
2422
optimizePrintF(CallInst * CI,IRBuilderBase & B)2423 Value *LibCallSimplifier::optimizePrintF(CallInst *CI, IRBuilderBase &B) {
2424
2425 Function *Callee = CI->getCalledFunction();
2426 FunctionType *FT = Callee->getFunctionType();
2427 if (Value *V = optimizePrintFString(CI, B)) {
2428 return V;
2429 }
2430
2431 // printf(format, ...) -> iprintf(format, ...) if no floating point
2432 // arguments.
2433 if (TLI->has(LibFunc_iprintf) && !callHasFloatingPointArgument(CI)) {
2434 Module *M = B.GetInsertBlock()->getParent()->getParent();
2435 FunctionCallee IPrintFFn =
2436 M->getOrInsertFunction("iprintf", FT, Callee->getAttributes());
2437 CallInst *New = cast<CallInst>(CI->clone());
2438 New->setCalledFunction(IPrintFFn);
2439 B.Insert(New);
2440 return New;
2441 }
2442
2443 // printf(format, ...) -> __small_printf(format, ...) if no 128-bit floating point
2444 // arguments.
2445 if (TLI->has(LibFunc_small_printf) && !callHasFP128Argument(CI)) {
2446 Module *M = B.GetInsertBlock()->getParent()->getParent();
2447 auto SmallPrintFFn =
2448 M->getOrInsertFunction(TLI->getName(LibFunc_small_printf),
2449 FT, Callee->getAttributes());
2450 CallInst *New = cast<CallInst>(CI->clone());
2451 New->setCalledFunction(SmallPrintFFn);
2452 B.Insert(New);
2453 return New;
2454 }
2455
2456 annotateNonNullNoUndefBasedOnAccess(CI, 0);
2457 return nullptr;
2458 }
2459
optimizeSPrintFString(CallInst * CI,IRBuilderBase & B)2460 Value *LibCallSimplifier::optimizeSPrintFString(CallInst *CI,
2461 IRBuilderBase &B) {
2462 // Check for a fixed format string.
2463 StringRef FormatStr;
2464 if (!getConstantStringInfo(CI->getArgOperand(1), FormatStr))
2465 return nullptr;
2466
2467 // If we just have a format string (nothing else crazy) transform it.
2468 if (CI->getNumArgOperands() == 2) {
2469 // Make sure there's no % in the constant array. We could try to handle
2470 // %% -> % in the future if we cared.
2471 if (FormatStr.find('%') != StringRef::npos)
2472 return nullptr; // we found a format specifier, bail out.
2473
2474 // sprintf(str, fmt) -> llvm.memcpy(align 1 str, align 1 fmt, strlen(fmt)+1)
2475 B.CreateMemCpy(
2476 CI->getArgOperand(0), Align(1), CI->getArgOperand(1), Align(1),
2477 ConstantInt::get(DL.getIntPtrType(CI->getContext()),
2478 FormatStr.size() + 1)); // Copy the null byte.
2479 return ConstantInt::get(CI->getType(), FormatStr.size());
2480 }
2481
2482 // The remaining optimizations require the format string to be "%s" or "%c"
2483 // and have an extra operand.
2484 if (FormatStr.size() != 2 || FormatStr[0] != '%' ||
2485 CI->getNumArgOperands() < 3)
2486 return nullptr;
2487
2488 // Decode the second character of the format string.
2489 if (FormatStr[1] == 'c') {
2490 // sprintf(dst, "%c", chr) --> *(i8*)dst = chr; *((i8*)dst+1) = 0
2491 if (!CI->getArgOperand(2)->getType()->isIntegerTy())
2492 return nullptr;
2493 Value *V = B.CreateTrunc(CI->getArgOperand(2), B.getInt8Ty(), "char");
2494 Value *Ptr = castToCStr(CI->getArgOperand(0), B);
2495 B.CreateStore(V, Ptr);
2496 Ptr = B.CreateGEP(B.getInt8Ty(), Ptr, B.getInt32(1), "nul");
2497 B.CreateStore(B.getInt8(0), Ptr);
2498
2499 return ConstantInt::get(CI->getType(), 1);
2500 }
2501
2502 if (FormatStr[1] == 's') {
2503 // sprintf(dest, "%s", str) -> llvm.memcpy(align 1 dest, align 1 str,
2504 // strlen(str)+1)
2505 if (!CI->getArgOperand(2)->getType()->isPointerTy())
2506 return nullptr;
2507
2508 if (CI->use_empty())
2509 // sprintf(dest, "%s", str) -> strcpy(dest, str)
2510 return emitStrCpy(CI->getArgOperand(0), CI->getArgOperand(2), B, TLI);
2511
2512 uint64_t SrcLen = GetStringLength(CI->getArgOperand(2));
2513 if (SrcLen) {
2514 B.CreateMemCpy(
2515 CI->getArgOperand(0), Align(1), CI->getArgOperand(2), Align(1),
2516 ConstantInt::get(DL.getIntPtrType(CI->getContext()), SrcLen));
2517 // Returns total number of characters written without null-character.
2518 return ConstantInt::get(CI->getType(), SrcLen - 1);
2519 } else if (Value *V = emitStpCpy(CI->getArgOperand(0), CI->getArgOperand(2),
2520 B, TLI)) {
2521 // sprintf(dest, "%s", str) -> stpcpy(dest, str) - dest
2522 Value *PtrDiff = B.CreatePtrDiff(V, CI->getArgOperand(0));
2523 return B.CreateIntCast(PtrDiff, CI->getType(), false);
2524 }
2525
2526 bool OptForSize = CI->getFunction()->hasOptSize() ||
2527 llvm::shouldOptimizeForSize(CI->getParent(), PSI, BFI,
2528 PGSOQueryType::IRPass);
2529 if (OptForSize)
2530 return nullptr;
2531
2532 Value *Len = emitStrLen(CI->getArgOperand(2), B, DL, TLI);
2533 if (!Len)
2534 return nullptr;
2535 Value *IncLen =
2536 B.CreateAdd(Len, ConstantInt::get(Len->getType(), 1), "leninc");
2537 B.CreateMemCpy(CI->getArgOperand(0), Align(1), CI->getArgOperand(2),
2538 Align(1), IncLen);
2539
2540 // The sprintf result is the unincremented number of bytes in the string.
2541 return B.CreateIntCast(Len, CI->getType(), false);
2542 }
2543 return nullptr;
2544 }
2545
optimizeSPrintF(CallInst * CI,IRBuilderBase & B)2546 Value *LibCallSimplifier::optimizeSPrintF(CallInst *CI, IRBuilderBase &B) {
2547 Function *Callee = CI->getCalledFunction();
2548 FunctionType *FT = Callee->getFunctionType();
2549 if (Value *V = optimizeSPrintFString(CI, B)) {
2550 return V;
2551 }
2552
2553 // sprintf(str, format, ...) -> siprintf(str, format, ...) if no floating
2554 // point arguments.
2555 if (TLI->has(LibFunc_siprintf) && !callHasFloatingPointArgument(CI)) {
2556 Module *M = B.GetInsertBlock()->getParent()->getParent();
2557 FunctionCallee SIPrintFFn =
2558 M->getOrInsertFunction("siprintf", FT, Callee->getAttributes());
2559 CallInst *New = cast<CallInst>(CI->clone());
2560 New->setCalledFunction(SIPrintFFn);
2561 B.Insert(New);
2562 return New;
2563 }
2564
2565 // sprintf(str, format, ...) -> __small_sprintf(str, format, ...) if no 128-bit
2566 // floating point arguments.
2567 if (TLI->has(LibFunc_small_sprintf) && !callHasFP128Argument(CI)) {
2568 Module *M = B.GetInsertBlock()->getParent()->getParent();
2569 auto SmallSPrintFFn =
2570 M->getOrInsertFunction(TLI->getName(LibFunc_small_sprintf),
2571 FT, Callee->getAttributes());
2572 CallInst *New = cast<CallInst>(CI->clone());
2573 New->setCalledFunction(SmallSPrintFFn);
2574 B.Insert(New);
2575 return New;
2576 }
2577
2578 annotateNonNullNoUndefBasedOnAccess(CI, {0, 1});
2579 return nullptr;
2580 }
2581
optimizeSnPrintFString(CallInst * CI,IRBuilderBase & B)2582 Value *LibCallSimplifier::optimizeSnPrintFString(CallInst *CI,
2583 IRBuilderBase &B) {
2584 // Check for size
2585 ConstantInt *Size = dyn_cast<ConstantInt>(CI->getArgOperand(1));
2586 if (!Size)
2587 return nullptr;
2588
2589 uint64_t N = Size->getZExtValue();
2590 // Check for a fixed format string.
2591 StringRef FormatStr;
2592 if (!getConstantStringInfo(CI->getArgOperand(2), FormatStr))
2593 return nullptr;
2594
2595 // If we just have a format string (nothing else crazy) transform it.
2596 if (CI->getNumArgOperands() == 3) {
2597 // Make sure there's no % in the constant array. We could try to handle
2598 // %% -> % in the future if we cared.
2599 if (FormatStr.find('%') != StringRef::npos)
2600 return nullptr; // we found a format specifier, bail out.
2601
2602 if (N == 0)
2603 return ConstantInt::get(CI->getType(), FormatStr.size());
2604 else if (N < FormatStr.size() + 1)
2605 return nullptr;
2606
2607 // snprintf(dst, size, fmt) -> llvm.memcpy(align 1 dst, align 1 fmt,
2608 // strlen(fmt)+1)
2609 B.CreateMemCpy(
2610 CI->getArgOperand(0), Align(1), CI->getArgOperand(2), Align(1),
2611 ConstantInt::get(DL.getIntPtrType(CI->getContext()),
2612 FormatStr.size() + 1)); // Copy the null byte.
2613 return ConstantInt::get(CI->getType(), FormatStr.size());
2614 }
2615
2616 // The remaining optimizations require the format string to be "%s" or "%c"
2617 // and have an extra operand.
2618 if (FormatStr.size() == 2 && FormatStr[0] == '%' &&
2619 CI->getNumArgOperands() == 4) {
2620
2621 // Decode the second character of the format string.
2622 if (FormatStr[1] == 'c') {
2623 if (N == 0)
2624 return ConstantInt::get(CI->getType(), 1);
2625 else if (N == 1)
2626 return nullptr;
2627
2628 // snprintf(dst, size, "%c", chr) --> *(i8*)dst = chr; *((i8*)dst+1) = 0
2629 if (!CI->getArgOperand(3)->getType()->isIntegerTy())
2630 return nullptr;
2631 Value *V = B.CreateTrunc(CI->getArgOperand(3), B.getInt8Ty(), "char");
2632 Value *Ptr = castToCStr(CI->getArgOperand(0), B);
2633 B.CreateStore(V, Ptr);
2634 Ptr = B.CreateGEP(B.getInt8Ty(), Ptr, B.getInt32(1), "nul");
2635 B.CreateStore(B.getInt8(0), Ptr);
2636
2637 return ConstantInt::get(CI->getType(), 1);
2638 }
2639
2640 if (FormatStr[1] == 's') {
2641 // snprintf(dest, size, "%s", str) to llvm.memcpy(dest, str, len+1, 1)
2642 StringRef Str;
2643 if (!getConstantStringInfo(CI->getArgOperand(3), Str))
2644 return nullptr;
2645
2646 if (N == 0)
2647 return ConstantInt::get(CI->getType(), Str.size());
2648 else if (N < Str.size() + 1)
2649 return nullptr;
2650
2651 B.CreateMemCpy(CI->getArgOperand(0), Align(1), CI->getArgOperand(3),
2652 Align(1), ConstantInt::get(CI->getType(), Str.size() + 1));
2653
2654 // The snprintf result is the unincremented number of bytes in the string.
2655 return ConstantInt::get(CI->getType(), Str.size());
2656 }
2657 }
2658 return nullptr;
2659 }
2660
optimizeSnPrintF(CallInst * CI,IRBuilderBase & B)2661 Value *LibCallSimplifier::optimizeSnPrintF(CallInst *CI, IRBuilderBase &B) {
2662 if (Value *V = optimizeSnPrintFString(CI, B)) {
2663 return V;
2664 }
2665
2666 if (isKnownNonZero(CI->getOperand(1), DL))
2667 annotateNonNullNoUndefBasedOnAccess(CI, 0);
2668 return nullptr;
2669 }
2670
optimizeFPrintFString(CallInst * CI,IRBuilderBase & B)2671 Value *LibCallSimplifier::optimizeFPrintFString(CallInst *CI,
2672 IRBuilderBase &B) {
2673 optimizeErrorReporting(CI, B, 0);
2674
2675 // All the optimizations depend on the format string.
2676 StringRef FormatStr;
2677 if (!getConstantStringInfo(CI->getArgOperand(1), FormatStr))
2678 return nullptr;
2679
2680 // Do not do any of the following transformations if the fprintf return
2681 // value is used, in general the fprintf return value is not compatible
2682 // with fwrite(), fputc() or fputs().
2683 if (!CI->use_empty())
2684 return nullptr;
2685
2686 // fprintf(F, "foo") --> fwrite("foo", 3, 1, F)
2687 if (CI->getNumArgOperands() == 2) {
2688 // Could handle %% -> % if we cared.
2689 if (FormatStr.find('%') != StringRef::npos)
2690 return nullptr; // We found a format specifier.
2691
2692 return emitFWrite(
2693 CI->getArgOperand(1),
2694 ConstantInt::get(DL.getIntPtrType(CI->getContext()), FormatStr.size()),
2695 CI->getArgOperand(0), B, DL, TLI);
2696 }
2697
2698 // The remaining optimizations require the format string to be "%s" or "%c"
2699 // and have an extra operand.
2700 if (FormatStr.size() != 2 || FormatStr[0] != '%' ||
2701 CI->getNumArgOperands() < 3)
2702 return nullptr;
2703
2704 // Decode the second character of the format string.
2705 if (FormatStr[1] == 'c') {
2706 // fprintf(F, "%c", chr) --> fputc(chr, F)
2707 if (!CI->getArgOperand(2)->getType()->isIntegerTy())
2708 return nullptr;
2709 return emitFPutC(CI->getArgOperand(2), CI->getArgOperand(0), B, TLI);
2710 }
2711
2712 if (FormatStr[1] == 's') {
2713 // fprintf(F, "%s", str) --> fputs(str, F)
2714 if (!CI->getArgOperand(2)->getType()->isPointerTy())
2715 return nullptr;
2716 return emitFPutS(CI->getArgOperand(2), CI->getArgOperand(0), B, TLI);
2717 }
2718 return nullptr;
2719 }
2720
optimizeFPrintF(CallInst * CI,IRBuilderBase & B)2721 Value *LibCallSimplifier::optimizeFPrintF(CallInst *CI, IRBuilderBase &B) {
2722 Function *Callee = CI->getCalledFunction();
2723 FunctionType *FT = Callee->getFunctionType();
2724 if (Value *V = optimizeFPrintFString(CI, B)) {
2725 return V;
2726 }
2727
2728 // fprintf(stream, format, ...) -> fiprintf(stream, format, ...) if no
2729 // floating point arguments.
2730 if (TLI->has(LibFunc_fiprintf) && !callHasFloatingPointArgument(CI)) {
2731 Module *M = B.GetInsertBlock()->getParent()->getParent();
2732 FunctionCallee FIPrintFFn =
2733 M->getOrInsertFunction("fiprintf", FT, Callee->getAttributes());
2734 CallInst *New = cast<CallInst>(CI->clone());
2735 New->setCalledFunction(FIPrintFFn);
2736 B.Insert(New);
2737 return New;
2738 }
2739
2740 // fprintf(stream, format, ...) -> __small_fprintf(stream, format, ...) if no
2741 // 128-bit floating point arguments.
2742 if (TLI->has(LibFunc_small_fprintf) && !callHasFP128Argument(CI)) {
2743 Module *M = B.GetInsertBlock()->getParent()->getParent();
2744 auto SmallFPrintFFn =
2745 M->getOrInsertFunction(TLI->getName(LibFunc_small_fprintf),
2746 FT, Callee->getAttributes());
2747 CallInst *New = cast<CallInst>(CI->clone());
2748 New->setCalledFunction(SmallFPrintFFn);
2749 B.Insert(New);
2750 return New;
2751 }
2752
2753 return nullptr;
2754 }
2755
optimizeFWrite(CallInst * CI,IRBuilderBase & B)2756 Value *LibCallSimplifier::optimizeFWrite(CallInst *CI, IRBuilderBase &B) {
2757 optimizeErrorReporting(CI, B, 3);
2758
2759 // Get the element size and count.
2760 ConstantInt *SizeC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
2761 ConstantInt *CountC = dyn_cast<ConstantInt>(CI->getArgOperand(2));
2762 if (SizeC && CountC) {
2763 uint64_t Bytes = SizeC->getZExtValue() * CountC->getZExtValue();
2764
2765 // If this is writing zero records, remove the call (it's a noop).
2766 if (Bytes == 0)
2767 return ConstantInt::get(CI->getType(), 0);
2768
2769 // If this is writing one byte, turn it into fputc.
2770 // This optimisation is only valid, if the return value is unused.
2771 if (Bytes == 1 && CI->use_empty()) { // fwrite(S,1,1,F) -> fputc(S[0],F)
2772 Value *Char = B.CreateLoad(B.getInt8Ty(),
2773 castToCStr(CI->getArgOperand(0), B), "char");
2774 Value *NewCI = emitFPutC(Char, CI->getArgOperand(3), B, TLI);
2775 return NewCI ? ConstantInt::get(CI->getType(), 1) : nullptr;
2776 }
2777 }
2778
2779 return nullptr;
2780 }
2781
optimizeFPuts(CallInst * CI,IRBuilderBase & B)2782 Value *LibCallSimplifier::optimizeFPuts(CallInst *CI, IRBuilderBase &B) {
2783 optimizeErrorReporting(CI, B, 1);
2784
2785 // Don't rewrite fputs to fwrite when optimising for size because fwrite
2786 // requires more arguments and thus extra MOVs are required.
2787 bool OptForSize = CI->getFunction()->hasOptSize() ||
2788 llvm::shouldOptimizeForSize(CI->getParent(), PSI, BFI,
2789 PGSOQueryType::IRPass);
2790 if (OptForSize)
2791 return nullptr;
2792
2793 // We can't optimize if return value is used.
2794 if (!CI->use_empty())
2795 return nullptr;
2796
2797 // fputs(s,F) --> fwrite(s,strlen(s),1,F)
2798 uint64_t Len = GetStringLength(CI->getArgOperand(0));
2799 if (!Len)
2800 return nullptr;
2801
2802 // Known to have no uses (see above).
2803 return emitFWrite(
2804 CI->getArgOperand(0),
2805 ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len - 1),
2806 CI->getArgOperand(1), B, DL, TLI);
2807 }
2808
optimizePuts(CallInst * CI,IRBuilderBase & B)2809 Value *LibCallSimplifier::optimizePuts(CallInst *CI, IRBuilderBase &B) {
2810 annotateNonNullNoUndefBasedOnAccess(CI, 0);
2811 if (!CI->use_empty())
2812 return nullptr;
2813
2814 // Check for a constant string.
2815 // puts("") -> putchar('\n')
2816 StringRef Str;
2817 if (getConstantStringInfo(CI->getArgOperand(0), Str) && Str.empty())
2818 return emitPutChar(B.getInt32('\n'), B, TLI);
2819
2820 return nullptr;
2821 }
2822
optimizeBCopy(CallInst * CI,IRBuilderBase & B)2823 Value *LibCallSimplifier::optimizeBCopy(CallInst *CI, IRBuilderBase &B) {
2824 // bcopy(src, dst, n) -> llvm.memmove(dst, src, n)
2825 return B.CreateMemMove(CI->getArgOperand(1), Align(1), CI->getArgOperand(0),
2826 Align(1), CI->getArgOperand(2));
2827 }
2828
hasFloatVersion(StringRef FuncName)2829 bool LibCallSimplifier::hasFloatVersion(StringRef FuncName) {
2830 LibFunc Func;
2831 SmallString<20> FloatFuncName = FuncName;
2832 FloatFuncName += 'f';
2833 if (TLI->getLibFunc(FloatFuncName, Func))
2834 return TLI->has(Func);
2835 return false;
2836 }
2837
optimizeStringMemoryLibCall(CallInst * CI,IRBuilderBase & Builder)2838 Value *LibCallSimplifier::optimizeStringMemoryLibCall(CallInst *CI,
2839 IRBuilderBase &Builder) {
2840 LibFunc Func;
2841 Function *Callee = CI->getCalledFunction();
2842 // Check for string/memory library functions.
2843 if (TLI->getLibFunc(*Callee, Func) && TLI->has(Func)) {
2844 // Make sure we never change the calling convention.
2845 assert(
2846 (ignoreCallingConv(Func) ||
2847 TargetLibraryInfoImpl::isCallingConvCCompatible(CI)) &&
2848 "Optimizing string/memory libcall would change the calling convention");
2849 switch (Func) {
2850 case LibFunc_strcat:
2851 return optimizeStrCat(CI, Builder);
2852 case LibFunc_strncat:
2853 return optimizeStrNCat(CI, Builder);
2854 case LibFunc_strchr:
2855 return optimizeStrChr(CI, Builder);
2856 case LibFunc_strrchr:
2857 return optimizeStrRChr(CI, Builder);
2858 case LibFunc_strcmp:
2859 return optimizeStrCmp(CI, Builder);
2860 case LibFunc_strncmp:
2861 return optimizeStrNCmp(CI, Builder);
2862 case LibFunc_strcpy:
2863 return optimizeStrCpy(CI, Builder);
2864 case LibFunc_stpcpy:
2865 return optimizeStpCpy(CI, Builder);
2866 case LibFunc_strncpy:
2867 return optimizeStrNCpy(CI, Builder);
2868 case LibFunc_strlen:
2869 return optimizeStrLen(CI, Builder);
2870 case LibFunc_strpbrk:
2871 return optimizeStrPBrk(CI, Builder);
2872 case LibFunc_strndup:
2873 return optimizeStrNDup(CI, Builder);
2874 case LibFunc_strtol:
2875 case LibFunc_strtod:
2876 case LibFunc_strtof:
2877 case LibFunc_strtoul:
2878 case LibFunc_strtoll:
2879 case LibFunc_strtold:
2880 case LibFunc_strtoull:
2881 return optimizeStrTo(CI, Builder);
2882 case LibFunc_strspn:
2883 return optimizeStrSpn(CI, Builder);
2884 case LibFunc_strcspn:
2885 return optimizeStrCSpn(CI, Builder);
2886 case LibFunc_strstr:
2887 return optimizeStrStr(CI, Builder);
2888 case LibFunc_memchr:
2889 return optimizeMemChr(CI, Builder);
2890 case LibFunc_memrchr:
2891 return optimizeMemRChr(CI, Builder);
2892 case LibFunc_bcmp:
2893 return optimizeBCmp(CI, Builder);
2894 case LibFunc_memcmp:
2895 return optimizeMemCmp(CI, Builder);
2896 case LibFunc_memcpy:
2897 return optimizeMemCpy(CI, Builder);
2898 case LibFunc_memccpy:
2899 return optimizeMemCCpy(CI, Builder);
2900 case LibFunc_mempcpy:
2901 return optimizeMemPCpy(CI, Builder);
2902 case LibFunc_memmove:
2903 return optimizeMemMove(CI, Builder);
2904 case LibFunc_memset:
2905 return optimizeMemSet(CI, Builder);
2906 case LibFunc_realloc:
2907 return optimizeRealloc(CI, Builder);
2908 case LibFunc_wcslen:
2909 return optimizeWcslen(CI, Builder);
2910 case LibFunc_bcopy:
2911 return optimizeBCopy(CI, Builder);
2912 default:
2913 break;
2914 }
2915 }
2916 return nullptr;
2917 }
2918
optimizeFloatingPointLibCall(CallInst * CI,LibFunc Func,IRBuilderBase & Builder)2919 Value *LibCallSimplifier::optimizeFloatingPointLibCall(CallInst *CI,
2920 LibFunc Func,
2921 IRBuilderBase &Builder) {
2922 // Don't optimize calls that require strict floating point semantics.
2923 if (CI->isStrictFP())
2924 return nullptr;
2925
2926 if (Value *V = optimizeTrigReflections(CI, Func, Builder))
2927 return V;
2928
2929 switch (Func) {
2930 case LibFunc_sinpif:
2931 case LibFunc_sinpi:
2932 case LibFunc_cospif:
2933 case LibFunc_cospi:
2934 return optimizeSinCosPi(CI, Builder);
2935 case LibFunc_powf:
2936 case LibFunc_pow:
2937 case LibFunc_powl:
2938 return optimizePow(CI, Builder);
2939 case LibFunc_exp2l:
2940 case LibFunc_exp2:
2941 case LibFunc_exp2f:
2942 return optimizeExp2(CI, Builder);
2943 case LibFunc_fabsf:
2944 case LibFunc_fabs:
2945 case LibFunc_fabsl:
2946 return replaceUnaryCall(CI, Builder, Intrinsic::fabs);
2947 case LibFunc_sqrtf:
2948 case LibFunc_sqrt:
2949 case LibFunc_sqrtl:
2950 return optimizeSqrt(CI, Builder);
2951 case LibFunc_logf:
2952 case LibFunc_log:
2953 case LibFunc_logl:
2954 case LibFunc_log10f:
2955 case LibFunc_log10:
2956 case LibFunc_log10l:
2957 case LibFunc_log1pf:
2958 case LibFunc_log1p:
2959 case LibFunc_log1pl:
2960 case LibFunc_log2f:
2961 case LibFunc_log2:
2962 case LibFunc_log2l:
2963 case LibFunc_logbf:
2964 case LibFunc_logb:
2965 case LibFunc_logbl:
2966 return optimizeLog(CI, Builder);
2967 case LibFunc_tan:
2968 case LibFunc_tanf:
2969 case LibFunc_tanl:
2970 return optimizeTan(CI, Builder);
2971 case LibFunc_ceil:
2972 return replaceUnaryCall(CI, Builder, Intrinsic::ceil);
2973 case LibFunc_floor:
2974 return replaceUnaryCall(CI, Builder, Intrinsic::floor);
2975 case LibFunc_round:
2976 return replaceUnaryCall(CI, Builder, Intrinsic::round);
2977 case LibFunc_roundeven:
2978 return replaceUnaryCall(CI, Builder, Intrinsic::roundeven);
2979 case LibFunc_nearbyint:
2980 return replaceUnaryCall(CI, Builder, Intrinsic::nearbyint);
2981 case LibFunc_rint:
2982 return replaceUnaryCall(CI, Builder, Intrinsic::rint);
2983 case LibFunc_trunc:
2984 return replaceUnaryCall(CI, Builder, Intrinsic::trunc);
2985 case LibFunc_acos:
2986 case LibFunc_acosh:
2987 case LibFunc_asin:
2988 case LibFunc_asinh:
2989 case LibFunc_atan:
2990 case LibFunc_atanh:
2991 case LibFunc_cbrt:
2992 case LibFunc_cosh:
2993 case LibFunc_exp:
2994 case LibFunc_exp10:
2995 case LibFunc_expm1:
2996 case LibFunc_cos:
2997 case LibFunc_sin:
2998 case LibFunc_sinh:
2999 case LibFunc_tanh:
3000 if (UnsafeFPShrink && hasFloatVersion(CI->getCalledFunction()->getName()))
3001 return optimizeUnaryDoubleFP(CI, Builder, true);
3002 return nullptr;
3003 case LibFunc_copysign:
3004 if (hasFloatVersion(CI->getCalledFunction()->getName()))
3005 return optimizeBinaryDoubleFP(CI, Builder);
3006 return nullptr;
3007 case LibFunc_fminf:
3008 case LibFunc_fmin:
3009 case LibFunc_fminl:
3010 case LibFunc_fmaxf:
3011 case LibFunc_fmax:
3012 case LibFunc_fmaxl:
3013 return optimizeFMinFMax(CI, Builder);
3014 case LibFunc_cabs:
3015 case LibFunc_cabsf:
3016 case LibFunc_cabsl:
3017 return optimizeCAbs(CI, Builder);
3018 default:
3019 return nullptr;
3020 }
3021 }
3022
optimizeCall(CallInst * CI,IRBuilderBase & Builder)3023 Value *LibCallSimplifier::optimizeCall(CallInst *CI, IRBuilderBase &Builder) {
3024 // TODO: Split out the code below that operates on FP calls so that
3025 // we can all non-FP calls with the StrictFP attribute to be
3026 // optimized.
3027 if (CI->isNoBuiltin())
3028 return nullptr;
3029
3030 LibFunc Func;
3031 Function *Callee = CI->getCalledFunction();
3032 bool IsCallingConvC = TargetLibraryInfoImpl::isCallingConvCCompatible(CI);
3033
3034 SmallVector<OperandBundleDef, 2> OpBundles;
3035 CI->getOperandBundlesAsDefs(OpBundles);
3036
3037 IRBuilderBase::OperandBundlesGuard Guard(Builder);
3038 Builder.setDefaultOperandBundles(OpBundles);
3039
3040 // Command-line parameter overrides instruction attribute.
3041 // This can't be moved to optimizeFloatingPointLibCall() because it may be
3042 // used by the intrinsic optimizations.
3043 if (EnableUnsafeFPShrink.getNumOccurrences() > 0)
3044 UnsafeFPShrink = EnableUnsafeFPShrink;
3045 else if (isa<FPMathOperator>(CI) && CI->isFast())
3046 UnsafeFPShrink = true;
3047
3048 // First, check for intrinsics.
3049 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
3050 if (!IsCallingConvC)
3051 return nullptr;
3052 // The FP intrinsics have corresponding constrained versions so we don't
3053 // need to check for the StrictFP attribute here.
3054 switch (II->getIntrinsicID()) {
3055 case Intrinsic::pow:
3056 return optimizePow(CI, Builder);
3057 case Intrinsic::exp2:
3058 return optimizeExp2(CI, Builder);
3059 case Intrinsic::log:
3060 case Intrinsic::log2:
3061 case Intrinsic::log10:
3062 return optimizeLog(CI, Builder);
3063 case Intrinsic::sqrt:
3064 return optimizeSqrt(CI, Builder);
3065 // TODO: Use foldMallocMemset() with memset intrinsic.
3066 case Intrinsic::memset:
3067 return optimizeMemSet(CI, Builder);
3068 case Intrinsic::memcpy:
3069 return optimizeMemCpy(CI, Builder);
3070 case Intrinsic::memmove:
3071 return optimizeMemMove(CI, Builder);
3072 default:
3073 return nullptr;
3074 }
3075 }
3076
3077 // Also try to simplify calls to fortified library functions.
3078 if (Value *SimplifiedFortifiedCI =
3079 FortifiedSimplifier.optimizeCall(CI, Builder)) {
3080 // Try to further simplify the result.
3081 CallInst *SimplifiedCI = dyn_cast<CallInst>(SimplifiedFortifiedCI);
3082 if (SimplifiedCI && SimplifiedCI->getCalledFunction()) {
3083 // Ensure that SimplifiedCI's uses are complete, since some calls have
3084 // their uses analyzed.
3085 replaceAllUsesWith(CI, SimplifiedCI);
3086
3087 // Set insertion point to SimplifiedCI to guarantee we reach all uses
3088 // we might replace later on.
3089 IRBuilderBase::InsertPointGuard Guard(Builder);
3090 Builder.SetInsertPoint(SimplifiedCI);
3091 if (Value *V = optimizeStringMemoryLibCall(SimplifiedCI, Builder)) {
3092 // If we were able to further simplify, remove the now redundant call.
3093 substituteInParent(SimplifiedCI, V);
3094 return V;
3095 }
3096 }
3097 return SimplifiedFortifiedCI;
3098 }
3099
3100 // Then check for known library functions.
3101 if (TLI->getLibFunc(*Callee, Func) && TLI->has(Func)) {
3102 // We never change the calling convention.
3103 if (!ignoreCallingConv(Func) && !IsCallingConvC)
3104 return nullptr;
3105 if (Value *V = optimizeStringMemoryLibCall(CI, Builder))
3106 return V;
3107 if (Value *V = optimizeFloatingPointLibCall(CI, Func, Builder))
3108 return V;
3109 switch (Func) {
3110 case LibFunc_ffs:
3111 case LibFunc_ffsl:
3112 case LibFunc_ffsll:
3113 return optimizeFFS(CI, Builder);
3114 case LibFunc_fls:
3115 case LibFunc_flsl:
3116 case LibFunc_flsll:
3117 return optimizeFls(CI, Builder);
3118 case LibFunc_abs:
3119 case LibFunc_labs:
3120 case LibFunc_llabs:
3121 return optimizeAbs(CI, Builder);
3122 case LibFunc_isdigit:
3123 return optimizeIsDigit(CI, Builder);
3124 case LibFunc_isascii:
3125 return optimizeIsAscii(CI, Builder);
3126 case LibFunc_toascii:
3127 return optimizeToAscii(CI, Builder);
3128 case LibFunc_atoi:
3129 case LibFunc_atol:
3130 case LibFunc_atoll:
3131 return optimizeAtoi(CI, Builder);
3132 case LibFunc_strtol:
3133 case LibFunc_strtoll:
3134 return optimizeStrtol(CI, Builder);
3135 case LibFunc_printf:
3136 return optimizePrintF(CI, Builder);
3137 case LibFunc_sprintf:
3138 return optimizeSPrintF(CI, Builder);
3139 case LibFunc_snprintf:
3140 return optimizeSnPrintF(CI, Builder);
3141 case LibFunc_fprintf:
3142 return optimizeFPrintF(CI, Builder);
3143 case LibFunc_fwrite:
3144 return optimizeFWrite(CI, Builder);
3145 case LibFunc_fputs:
3146 return optimizeFPuts(CI, Builder);
3147 case LibFunc_puts:
3148 return optimizePuts(CI, Builder);
3149 case LibFunc_perror:
3150 return optimizeErrorReporting(CI, Builder);
3151 case LibFunc_vfprintf:
3152 case LibFunc_fiprintf:
3153 return optimizeErrorReporting(CI, Builder, 0);
3154 default:
3155 return nullptr;
3156 }
3157 }
3158 return nullptr;
3159 }
3160
LibCallSimplifier(const DataLayout & DL,const TargetLibraryInfo * TLI,OptimizationRemarkEmitter & ORE,BlockFrequencyInfo * BFI,ProfileSummaryInfo * PSI,function_ref<void (Instruction *,Value *)> Replacer,function_ref<void (Instruction *)> Eraser)3161 LibCallSimplifier::LibCallSimplifier(
3162 const DataLayout &DL, const TargetLibraryInfo *TLI,
3163 OptimizationRemarkEmitter &ORE,
3164 BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI,
3165 function_ref<void(Instruction *, Value *)> Replacer,
3166 function_ref<void(Instruction *)> Eraser)
3167 : FortifiedSimplifier(TLI), DL(DL), TLI(TLI), ORE(ORE), BFI(BFI), PSI(PSI),
3168 UnsafeFPShrink(false), Replacer(Replacer), Eraser(Eraser) {}
3169
replaceAllUsesWith(Instruction * I,Value * With)3170 void LibCallSimplifier::replaceAllUsesWith(Instruction *I, Value *With) {
3171 // Indirect through the replacer used in this instance.
3172 Replacer(I, With);
3173 }
3174
eraseFromParent(Instruction * I)3175 void LibCallSimplifier::eraseFromParent(Instruction *I) {
3176 Eraser(I);
3177 }
3178
3179 // TODO:
3180 // Additional cases that we need to add to this file:
3181 //
3182 // cbrt:
3183 // * cbrt(expN(X)) -> expN(x/3)
3184 // * cbrt(sqrt(x)) -> pow(x,1/6)
3185 // * cbrt(cbrt(x)) -> pow(x,1/9)
3186 //
3187 // exp, expf, expl:
3188 // * exp(log(x)) -> x
3189 //
3190 // log, logf, logl:
3191 // * log(exp(x)) -> x
3192 // * log(exp(y)) -> y*log(e)
3193 // * log(exp10(y)) -> y*log(10)
3194 // * log(sqrt(x)) -> 0.5*log(x)
3195 //
3196 // pow, powf, powl:
3197 // * pow(sqrt(x),y) -> pow(x,y*0.5)
3198 // * pow(pow(x,y),z)-> pow(x,y*z)
3199 //
3200 // signbit:
3201 // * signbit(cnst) -> cnst'
3202 // * signbit(nncst) -> 0 (if pstv is a non-negative constant)
3203 //
3204 // sqrt, sqrtf, sqrtl:
3205 // * sqrt(expN(x)) -> expN(x*0.5)
3206 // * sqrt(Nroot(x)) -> pow(x,1/(2*N))
3207 // * sqrt(pow(x,y)) -> pow(|x|,y*0.5)
3208 //
3209
3210 //===----------------------------------------------------------------------===//
3211 // Fortified Library Call Optimizations
3212 //===----------------------------------------------------------------------===//
3213
3214 bool
isFortifiedCallFoldable(CallInst * CI,unsigned ObjSizeOp,Optional<unsigned> SizeOp,Optional<unsigned> StrOp,Optional<unsigned> FlagOp)3215 FortifiedLibCallSimplifier::isFortifiedCallFoldable(CallInst *CI,
3216 unsigned ObjSizeOp,
3217 Optional<unsigned> SizeOp,
3218 Optional<unsigned> StrOp,
3219 Optional<unsigned> FlagOp) {
3220 // If this function takes a flag argument, the implementation may use it to
3221 // perform extra checks. Don't fold into the non-checking variant.
3222 if (FlagOp) {
3223 ConstantInt *Flag = dyn_cast<ConstantInt>(CI->getArgOperand(*FlagOp));
3224 if (!Flag || !Flag->isZero())
3225 return false;
3226 }
3227
3228 if (SizeOp && CI->getArgOperand(ObjSizeOp) == CI->getArgOperand(*SizeOp))
3229 return true;
3230
3231 if (ConstantInt *ObjSizeCI =
3232 dyn_cast<ConstantInt>(CI->getArgOperand(ObjSizeOp))) {
3233 if (ObjSizeCI->isMinusOne())
3234 return true;
3235 // If the object size wasn't -1 (unknown), bail out if we were asked to.
3236 if (OnlyLowerUnknownSize)
3237 return false;
3238 if (StrOp) {
3239 uint64_t Len = GetStringLength(CI->getArgOperand(*StrOp));
3240 // If the length is 0 we don't know how long it is and so we can't
3241 // remove the check.
3242 if (Len)
3243 annotateDereferenceableBytes(CI, *StrOp, Len);
3244 else
3245 return false;
3246 return ObjSizeCI->getZExtValue() >= Len;
3247 }
3248
3249 if (SizeOp) {
3250 if (ConstantInt *SizeCI =
3251 dyn_cast<ConstantInt>(CI->getArgOperand(*SizeOp)))
3252 return ObjSizeCI->getZExtValue() >= SizeCI->getZExtValue();
3253 }
3254 }
3255 return false;
3256 }
3257
optimizeMemCpyChk(CallInst * CI,IRBuilderBase & B)3258 Value *FortifiedLibCallSimplifier::optimizeMemCpyChk(CallInst *CI,
3259 IRBuilderBase &B) {
3260 if (isFortifiedCallFoldable(CI, 3, 2)) {
3261 CallInst *NewCI =
3262 B.CreateMemCpy(CI->getArgOperand(0), Align(1), CI->getArgOperand(1),
3263 Align(1), CI->getArgOperand(2));
3264 NewCI->setAttributes(CI->getAttributes());
3265 NewCI->removeAttributes(AttributeList::ReturnIndex,
3266 AttributeFuncs::typeIncompatible(NewCI->getType()));
3267 return CI->getArgOperand(0);
3268 }
3269 return nullptr;
3270 }
3271
optimizeMemMoveChk(CallInst * CI,IRBuilderBase & B)3272 Value *FortifiedLibCallSimplifier::optimizeMemMoveChk(CallInst *CI,
3273 IRBuilderBase &B) {
3274 if (isFortifiedCallFoldable(CI, 3, 2)) {
3275 CallInst *NewCI =
3276 B.CreateMemMove(CI->getArgOperand(0), Align(1), CI->getArgOperand(1),
3277 Align(1), CI->getArgOperand(2));
3278 NewCI->setAttributes(CI->getAttributes());
3279 NewCI->removeAttributes(AttributeList::ReturnIndex,
3280 AttributeFuncs::typeIncompatible(NewCI->getType()));
3281 return CI->getArgOperand(0);
3282 }
3283 return nullptr;
3284 }
3285
optimizeMemSetChk(CallInst * CI,IRBuilderBase & B)3286 Value *FortifiedLibCallSimplifier::optimizeMemSetChk(CallInst *CI,
3287 IRBuilderBase &B) {
3288 // TODO: Try foldMallocMemset() here.
3289
3290 if (isFortifiedCallFoldable(CI, 3, 2)) {
3291 Value *Val = B.CreateIntCast(CI->getArgOperand(1), B.getInt8Ty(), false);
3292 CallInst *NewCI = B.CreateMemSet(CI->getArgOperand(0), Val,
3293 CI->getArgOperand(2), Align(1));
3294 NewCI->setAttributes(CI->getAttributes());
3295 NewCI->removeAttributes(AttributeList::ReturnIndex,
3296 AttributeFuncs::typeIncompatible(NewCI->getType()));
3297 return CI->getArgOperand(0);
3298 }
3299 return nullptr;
3300 }
3301
optimizeMemPCpyChk(CallInst * CI,IRBuilderBase & B)3302 Value *FortifiedLibCallSimplifier::optimizeMemPCpyChk(CallInst *CI,
3303 IRBuilderBase &B) {
3304 const DataLayout &DL = CI->getModule()->getDataLayout();
3305 if (isFortifiedCallFoldable(CI, 3, 2))
3306 if (Value *Call = emitMemPCpy(CI->getArgOperand(0), CI->getArgOperand(1),
3307 CI->getArgOperand(2), B, DL, TLI)) {
3308 CallInst *NewCI = cast<CallInst>(Call);
3309 NewCI->setAttributes(CI->getAttributes());
3310 NewCI->removeAttributes(
3311 AttributeList::ReturnIndex,
3312 AttributeFuncs::typeIncompatible(NewCI->getType()));
3313 return NewCI;
3314 }
3315 return nullptr;
3316 }
3317
optimizeStrpCpyChk(CallInst * CI,IRBuilderBase & B,LibFunc Func)3318 Value *FortifiedLibCallSimplifier::optimizeStrpCpyChk(CallInst *CI,
3319 IRBuilderBase &B,
3320 LibFunc Func) {
3321 const DataLayout &DL = CI->getModule()->getDataLayout();
3322 Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1),
3323 *ObjSize = CI->getArgOperand(2);
3324
3325 // __stpcpy_chk(x,x,...) -> x+strlen(x)
3326 if (Func == LibFunc_stpcpy_chk && !OnlyLowerUnknownSize && Dst == Src) {
3327 Value *StrLen = emitStrLen(Src, B, DL, TLI);
3328 return StrLen ? B.CreateInBoundsGEP(B.getInt8Ty(), Dst, StrLen) : nullptr;
3329 }
3330
3331 // If a) we don't have any length information, or b) we know this will
3332 // fit then just lower to a plain st[rp]cpy. Otherwise we'll keep our
3333 // st[rp]cpy_chk call which may fail at runtime if the size is too long.
3334 // TODO: It might be nice to get a maximum length out of the possible
3335 // string lengths for varying.
3336 if (isFortifiedCallFoldable(CI, 2, None, 1)) {
3337 if (Func == LibFunc_strcpy_chk)
3338 return emitStrCpy(Dst, Src, B, TLI);
3339 else
3340 return emitStpCpy(Dst, Src, B, TLI);
3341 }
3342
3343 if (OnlyLowerUnknownSize)
3344 return nullptr;
3345
3346 // Maybe we can stil fold __st[rp]cpy_chk to __memcpy_chk.
3347 uint64_t Len = GetStringLength(Src);
3348 if (Len)
3349 annotateDereferenceableBytes(CI, 1, Len);
3350 else
3351 return nullptr;
3352
3353 Type *SizeTTy = DL.getIntPtrType(CI->getContext());
3354 Value *LenV = ConstantInt::get(SizeTTy, Len);
3355 Value *Ret = emitMemCpyChk(Dst, Src, LenV, ObjSize, B, DL, TLI);
3356 // If the function was an __stpcpy_chk, and we were able to fold it into
3357 // a __memcpy_chk, we still need to return the correct end pointer.
3358 if (Ret && Func == LibFunc_stpcpy_chk)
3359 return B.CreateGEP(B.getInt8Ty(), Dst, ConstantInt::get(SizeTTy, Len - 1));
3360 return Ret;
3361 }
3362
optimizeStrLenChk(CallInst * CI,IRBuilderBase & B)3363 Value *FortifiedLibCallSimplifier::optimizeStrLenChk(CallInst *CI,
3364 IRBuilderBase &B) {
3365 if (isFortifiedCallFoldable(CI, 1, None, 0))
3366 return emitStrLen(CI->getArgOperand(0), B, CI->getModule()->getDataLayout(),
3367 TLI);
3368 return nullptr;
3369 }
3370
optimizeStrpNCpyChk(CallInst * CI,IRBuilderBase & B,LibFunc Func)3371 Value *FortifiedLibCallSimplifier::optimizeStrpNCpyChk(CallInst *CI,
3372 IRBuilderBase &B,
3373 LibFunc Func) {
3374 if (isFortifiedCallFoldable(CI, 3, 2)) {
3375 if (Func == LibFunc_strncpy_chk)
3376 return emitStrNCpy(CI->getArgOperand(0), CI->getArgOperand(1),
3377 CI->getArgOperand(2), B, TLI);
3378 else
3379 return emitStpNCpy(CI->getArgOperand(0), CI->getArgOperand(1),
3380 CI->getArgOperand(2), B, TLI);
3381 }
3382
3383 return nullptr;
3384 }
3385
optimizeMemCCpyChk(CallInst * CI,IRBuilderBase & B)3386 Value *FortifiedLibCallSimplifier::optimizeMemCCpyChk(CallInst *CI,
3387 IRBuilderBase &B) {
3388 if (isFortifiedCallFoldable(CI, 4, 3))
3389 return emitMemCCpy(CI->getArgOperand(0), CI->getArgOperand(1),
3390 CI->getArgOperand(2), CI->getArgOperand(3), B, TLI);
3391
3392 return nullptr;
3393 }
3394
optimizeSNPrintfChk(CallInst * CI,IRBuilderBase & B)3395 Value *FortifiedLibCallSimplifier::optimizeSNPrintfChk(CallInst *CI,
3396 IRBuilderBase &B) {
3397 if (isFortifiedCallFoldable(CI, 3, 1, None, 2)) {
3398 SmallVector<Value *, 8> VariadicArgs(drop_begin(CI->args(), 5));
3399 return emitSNPrintf(CI->getArgOperand(0), CI->getArgOperand(1),
3400 CI->getArgOperand(4), VariadicArgs, B, TLI);
3401 }
3402
3403 return nullptr;
3404 }
3405
optimizeSPrintfChk(CallInst * CI,IRBuilderBase & B)3406 Value *FortifiedLibCallSimplifier::optimizeSPrintfChk(CallInst *CI,
3407 IRBuilderBase &B) {
3408 if (isFortifiedCallFoldable(CI, 2, None, None, 1)) {
3409 SmallVector<Value *, 8> VariadicArgs(drop_begin(CI->args(), 4));
3410 return emitSPrintf(CI->getArgOperand(0), CI->getArgOperand(3), VariadicArgs,
3411 B, TLI);
3412 }
3413
3414 return nullptr;
3415 }
3416
optimizeStrCatChk(CallInst * CI,IRBuilderBase & B)3417 Value *FortifiedLibCallSimplifier::optimizeStrCatChk(CallInst *CI,
3418 IRBuilderBase &B) {
3419 if (isFortifiedCallFoldable(CI, 2))
3420 return emitStrCat(CI->getArgOperand(0), CI->getArgOperand(1), B, TLI);
3421
3422 return nullptr;
3423 }
3424
optimizeStrLCat(CallInst * CI,IRBuilderBase & B)3425 Value *FortifiedLibCallSimplifier::optimizeStrLCat(CallInst *CI,
3426 IRBuilderBase &B) {
3427 if (isFortifiedCallFoldable(CI, 3))
3428 return emitStrLCat(CI->getArgOperand(0), CI->getArgOperand(1),
3429 CI->getArgOperand(2), B, TLI);
3430
3431 return nullptr;
3432 }
3433
optimizeStrNCatChk(CallInst * CI,IRBuilderBase & B)3434 Value *FortifiedLibCallSimplifier::optimizeStrNCatChk(CallInst *CI,
3435 IRBuilderBase &B) {
3436 if (isFortifiedCallFoldable(CI, 3))
3437 return emitStrNCat(CI->getArgOperand(0), CI->getArgOperand(1),
3438 CI->getArgOperand(2), B, TLI);
3439
3440 return nullptr;
3441 }
3442
optimizeStrLCpyChk(CallInst * CI,IRBuilderBase & B)3443 Value *FortifiedLibCallSimplifier::optimizeStrLCpyChk(CallInst *CI,
3444 IRBuilderBase &B) {
3445 if (isFortifiedCallFoldable(CI, 3))
3446 return emitStrLCpy(CI->getArgOperand(0), CI->getArgOperand(1),
3447 CI->getArgOperand(2), B, TLI);
3448
3449 return nullptr;
3450 }
3451
optimizeVSNPrintfChk(CallInst * CI,IRBuilderBase & B)3452 Value *FortifiedLibCallSimplifier::optimizeVSNPrintfChk(CallInst *CI,
3453 IRBuilderBase &B) {
3454 if (isFortifiedCallFoldable(CI, 3, 1, None, 2))
3455 return emitVSNPrintf(CI->getArgOperand(0), CI->getArgOperand(1),
3456 CI->getArgOperand(4), CI->getArgOperand(5), B, TLI);
3457
3458 return nullptr;
3459 }
3460
optimizeVSPrintfChk(CallInst * CI,IRBuilderBase & B)3461 Value *FortifiedLibCallSimplifier::optimizeVSPrintfChk(CallInst *CI,
3462 IRBuilderBase &B) {
3463 if (isFortifiedCallFoldable(CI, 2, None, None, 1))
3464 return emitVSPrintf(CI->getArgOperand(0), CI->getArgOperand(3),
3465 CI->getArgOperand(4), B, TLI);
3466
3467 return nullptr;
3468 }
3469
optimizeCall(CallInst * CI,IRBuilderBase & Builder)3470 Value *FortifiedLibCallSimplifier::optimizeCall(CallInst *CI,
3471 IRBuilderBase &Builder) {
3472 // FIXME: We shouldn't be changing "nobuiltin" or TLI unavailable calls here.
3473 // Some clang users checked for _chk libcall availability using:
3474 // __has_builtin(__builtin___memcpy_chk)
3475 // When compiling with -fno-builtin, this is always true.
3476 // When passing -ffreestanding/-mkernel, which both imply -fno-builtin, we
3477 // end up with fortified libcalls, which isn't acceptable in a freestanding
3478 // environment which only provides their non-fortified counterparts.
3479 //
3480 // Until we change clang and/or teach external users to check for availability
3481 // differently, disregard the "nobuiltin" attribute and TLI::has.
3482 //
3483 // PR23093.
3484
3485 LibFunc Func;
3486 Function *Callee = CI->getCalledFunction();
3487 bool IsCallingConvC = TargetLibraryInfoImpl::isCallingConvCCompatible(CI);
3488
3489 SmallVector<OperandBundleDef, 2> OpBundles;
3490 CI->getOperandBundlesAsDefs(OpBundles);
3491
3492 IRBuilderBase::OperandBundlesGuard Guard(Builder);
3493 Builder.setDefaultOperandBundles(OpBundles);
3494
3495 // First, check that this is a known library functions and that the prototype
3496 // is correct.
3497 if (!TLI->getLibFunc(*Callee, Func))
3498 return nullptr;
3499
3500 // We never change the calling convention.
3501 if (!ignoreCallingConv(Func) && !IsCallingConvC)
3502 return nullptr;
3503
3504 switch (Func) {
3505 case LibFunc_memcpy_chk:
3506 return optimizeMemCpyChk(CI, Builder);
3507 case LibFunc_mempcpy_chk:
3508 return optimizeMemPCpyChk(CI, Builder);
3509 case LibFunc_memmove_chk:
3510 return optimizeMemMoveChk(CI, Builder);
3511 case LibFunc_memset_chk:
3512 return optimizeMemSetChk(CI, Builder);
3513 case LibFunc_stpcpy_chk:
3514 case LibFunc_strcpy_chk:
3515 return optimizeStrpCpyChk(CI, Builder, Func);
3516 case LibFunc_strlen_chk:
3517 return optimizeStrLenChk(CI, Builder);
3518 case LibFunc_stpncpy_chk:
3519 case LibFunc_strncpy_chk:
3520 return optimizeStrpNCpyChk(CI, Builder, Func);
3521 case LibFunc_memccpy_chk:
3522 return optimizeMemCCpyChk(CI, Builder);
3523 case LibFunc_snprintf_chk:
3524 return optimizeSNPrintfChk(CI, Builder);
3525 case LibFunc_sprintf_chk:
3526 return optimizeSPrintfChk(CI, Builder);
3527 case LibFunc_strcat_chk:
3528 return optimizeStrCatChk(CI, Builder);
3529 case LibFunc_strlcat_chk:
3530 return optimizeStrLCat(CI, Builder);
3531 case LibFunc_strncat_chk:
3532 return optimizeStrNCatChk(CI, Builder);
3533 case LibFunc_strlcpy_chk:
3534 return optimizeStrLCpyChk(CI, Builder);
3535 case LibFunc_vsnprintf_chk:
3536 return optimizeVSNPrintfChk(CI, Builder);
3537 case LibFunc_vsprintf_chk:
3538 return optimizeVSPrintfChk(CI, Builder);
3539 default:
3540 break;
3541 }
3542 return nullptr;
3543 }
3544
FortifiedLibCallSimplifier(const TargetLibraryInfo * TLI,bool OnlyLowerUnknownSize)3545 FortifiedLibCallSimplifier::FortifiedLibCallSimplifier(
3546 const TargetLibraryInfo *TLI, bool OnlyLowerUnknownSize)
3547 : TLI(TLI), OnlyLowerUnknownSize(OnlyLowerUnknownSize) {}
3548