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.CreateLogicalAnd(Bounds, Bits, "memchr"),
937 CI->getType());
938 }
939
940 // Check if all arguments are constants. If so, we can constant fold.
941 if (!CharC)
942 return nullptr;
943
944 // Compute the offset.
945 size_t I = Str.find(CharC->getSExtValue() & 0xFF);
946 if (I == StringRef::npos) // Didn't find the char. memchr returns null.
947 return Constant::getNullValue(CI->getType());
948
949 // memchr(s+n,c,l) -> gep(s+n+i,c)
950 return B.CreateGEP(B.getInt8Ty(), SrcStr, B.getInt64(I), "memchr");
951 }
952
optimizeMemCmpConstantSize(CallInst * CI,Value * LHS,Value * RHS,uint64_t Len,IRBuilderBase & B,const DataLayout & DL)953 static Value *optimizeMemCmpConstantSize(CallInst *CI, Value *LHS, Value *RHS,
954 uint64_t Len, IRBuilderBase &B,
955 const DataLayout &DL) {
956 if (Len == 0) // memcmp(s1,s2,0) -> 0
957 return Constant::getNullValue(CI->getType());
958
959 // memcmp(S1,S2,1) -> *(unsigned char*)LHS - *(unsigned char*)RHS
960 if (Len == 1) {
961 Value *LHSV =
962 B.CreateZExt(B.CreateLoad(B.getInt8Ty(), castToCStr(LHS, B), "lhsc"),
963 CI->getType(), "lhsv");
964 Value *RHSV =
965 B.CreateZExt(B.CreateLoad(B.getInt8Ty(), castToCStr(RHS, B), "rhsc"),
966 CI->getType(), "rhsv");
967 return B.CreateSub(LHSV, RHSV, "chardiff");
968 }
969
970 // memcmp(S1,S2,N/8)==0 -> (*(intN_t*)S1 != *(intN_t*)S2)==0
971 // TODO: The case where both inputs are constants does not need to be limited
972 // to legal integers or equality comparison. See block below this.
973 if (DL.isLegalInteger(Len * 8) && isOnlyUsedInZeroEqualityComparison(CI)) {
974 IntegerType *IntType = IntegerType::get(CI->getContext(), Len * 8);
975 unsigned PrefAlignment = DL.getPrefTypeAlignment(IntType);
976
977 // First, see if we can fold either argument to a constant.
978 Value *LHSV = nullptr;
979 if (auto *LHSC = dyn_cast<Constant>(LHS)) {
980 LHSC = ConstantExpr::getBitCast(LHSC, IntType->getPointerTo());
981 LHSV = ConstantFoldLoadFromConstPtr(LHSC, IntType, DL);
982 }
983 Value *RHSV = nullptr;
984 if (auto *RHSC = dyn_cast<Constant>(RHS)) {
985 RHSC = ConstantExpr::getBitCast(RHSC, IntType->getPointerTo());
986 RHSV = ConstantFoldLoadFromConstPtr(RHSC, IntType, DL);
987 }
988
989 // Don't generate unaligned loads. If either source is constant data,
990 // alignment doesn't matter for that source because there is no load.
991 if ((LHSV || getKnownAlignment(LHS, DL, CI) >= PrefAlignment) &&
992 (RHSV || getKnownAlignment(RHS, DL, CI) >= PrefAlignment)) {
993 if (!LHSV) {
994 Type *LHSPtrTy =
995 IntType->getPointerTo(LHS->getType()->getPointerAddressSpace());
996 LHSV = B.CreateLoad(IntType, B.CreateBitCast(LHS, LHSPtrTy), "lhsv");
997 }
998 if (!RHSV) {
999 Type *RHSPtrTy =
1000 IntType->getPointerTo(RHS->getType()->getPointerAddressSpace());
1001 RHSV = B.CreateLoad(IntType, B.CreateBitCast(RHS, RHSPtrTy), "rhsv");
1002 }
1003 return B.CreateZExt(B.CreateICmpNE(LHSV, RHSV), CI->getType(), "memcmp");
1004 }
1005 }
1006
1007 // Constant folding: memcmp(x, y, Len) -> constant (all arguments are const).
1008 // TODO: This is limited to i8 arrays.
1009 StringRef LHSStr, RHSStr;
1010 if (getConstantStringInfo(LHS, LHSStr) &&
1011 getConstantStringInfo(RHS, RHSStr)) {
1012 // Make sure we're not reading out-of-bounds memory.
1013 if (Len > LHSStr.size() || Len > RHSStr.size())
1014 return nullptr;
1015 // Fold the memcmp and normalize the result. This way we get consistent
1016 // results across multiple platforms.
1017 uint64_t Ret = 0;
1018 int Cmp = memcmp(LHSStr.data(), RHSStr.data(), Len);
1019 if (Cmp < 0)
1020 Ret = -1;
1021 else if (Cmp > 0)
1022 Ret = 1;
1023 return ConstantInt::get(CI->getType(), Ret);
1024 }
1025
1026 return nullptr;
1027 }
1028
1029 // Most simplifications for memcmp also apply to bcmp.
optimizeMemCmpBCmpCommon(CallInst * CI,IRBuilderBase & B)1030 Value *LibCallSimplifier::optimizeMemCmpBCmpCommon(CallInst *CI,
1031 IRBuilderBase &B) {
1032 Value *LHS = CI->getArgOperand(0), *RHS = CI->getArgOperand(1);
1033 Value *Size = CI->getArgOperand(2);
1034
1035 if (LHS == RHS) // memcmp(s,s,x) -> 0
1036 return Constant::getNullValue(CI->getType());
1037
1038 annotateNonNullAndDereferenceable(CI, {0, 1}, Size, DL);
1039 // Handle constant lengths.
1040 ConstantInt *LenC = dyn_cast<ConstantInt>(Size);
1041 if (!LenC)
1042 return nullptr;
1043
1044 // memcmp(d,s,0) -> 0
1045 if (LenC->getZExtValue() == 0)
1046 return Constant::getNullValue(CI->getType());
1047
1048 if (Value *Res =
1049 optimizeMemCmpConstantSize(CI, LHS, RHS, LenC->getZExtValue(), B, DL))
1050 return Res;
1051 return nullptr;
1052 }
1053
optimizeMemCmp(CallInst * CI,IRBuilderBase & B)1054 Value *LibCallSimplifier::optimizeMemCmp(CallInst *CI, IRBuilderBase &B) {
1055 if (Value *V = optimizeMemCmpBCmpCommon(CI, B))
1056 return V;
1057
1058 // memcmp(x, y, Len) == 0 -> bcmp(x, y, Len) == 0
1059 // bcmp can be more efficient than memcmp because it only has to know that
1060 // there is a difference, not how different one is to the other.
1061 if (TLI->has(LibFunc_bcmp) && isOnlyUsedInZeroEqualityComparison(CI)) {
1062 Value *LHS = CI->getArgOperand(0);
1063 Value *RHS = CI->getArgOperand(1);
1064 Value *Size = CI->getArgOperand(2);
1065 return emitBCmp(LHS, RHS, Size, B, DL, TLI);
1066 }
1067
1068 return nullptr;
1069 }
1070
optimizeBCmp(CallInst * CI,IRBuilderBase & B)1071 Value *LibCallSimplifier::optimizeBCmp(CallInst *CI, IRBuilderBase &B) {
1072 return optimizeMemCmpBCmpCommon(CI, B);
1073 }
1074
optimizeMemCpy(CallInst * CI,IRBuilderBase & B)1075 Value *LibCallSimplifier::optimizeMemCpy(CallInst *CI, IRBuilderBase &B) {
1076 Value *Size = CI->getArgOperand(2);
1077 annotateNonNullAndDereferenceable(CI, {0, 1}, Size, DL);
1078 if (isa<IntrinsicInst>(CI))
1079 return nullptr;
1080
1081 // memcpy(x, y, n) -> llvm.memcpy(align 1 x, align 1 y, n)
1082 CallInst *NewCI = B.CreateMemCpy(CI->getArgOperand(0), Align(1),
1083 CI->getArgOperand(1), Align(1), Size);
1084 NewCI->setAttributes(CI->getAttributes());
1085 NewCI->removeAttributes(AttributeList::ReturnIndex,
1086 AttributeFuncs::typeIncompatible(NewCI->getType()));
1087 return CI->getArgOperand(0);
1088 }
1089
optimizeMemCCpy(CallInst * CI,IRBuilderBase & B)1090 Value *LibCallSimplifier::optimizeMemCCpy(CallInst *CI, IRBuilderBase &B) {
1091 Value *Dst = CI->getArgOperand(0);
1092 Value *Src = CI->getArgOperand(1);
1093 ConstantInt *StopChar = dyn_cast<ConstantInt>(CI->getArgOperand(2));
1094 ConstantInt *N = dyn_cast<ConstantInt>(CI->getArgOperand(3));
1095 StringRef SrcStr;
1096 if (CI->use_empty() && Dst == Src)
1097 return Dst;
1098 // memccpy(d, s, c, 0) -> nullptr
1099 if (N) {
1100 if (N->isNullValue())
1101 return Constant::getNullValue(CI->getType());
1102 if (!getConstantStringInfo(Src, SrcStr, /*Offset=*/0,
1103 /*TrimAtNul=*/false) ||
1104 !StopChar)
1105 return nullptr;
1106 } else {
1107 return nullptr;
1108 }
1109
1110 // Wrap arg 'c' of type int to char
1111 size_t Pos = SrcStr.find(StopChar->getSExtValue() & 0xFF);
1112 if (Pos == StringRef::npos) {
1113 if (N->getZExtValue() <= SrcStr.size()) {
1114 B.CreateMemCpy(Dst, Align(1), Src, Align(1), CI->getArgOperand(3));
1115 return Constant::getNullValue(CI->getType());
1116 }
1117 return nullptr;
1118 }
1119
1120 Value *NewN =
1121 ConstantInt::get(N->getType(), std::min(uint64_t(Pos + 1), N->getZExtValue()));
1122 // memccpy -> llvm.memcpy
1123 B.CreateMemCpy(Dst, Align(1), Src, Align(1), NewN);
1124 return Pos + 1 <= N->getZExtValue()
1125 ? B.CreateInBoundsGEP(B.getInt8Ty(), Dst, NewN)
1126 : Constant::getNullValue(CI->getType());
1127 }
1128
optimizeMemPCpy(CallInst * CI,IRBuilderBase & B)1129 Value *LibCallSimplifier::optimizeMemPCpy(CallInst *CI, IRBuilderBase &B) {
1130 Value *Dst = CI->getArgOperand(0);
1131 Value *N = CI->getArgOperand(2);
1132 // mempcpy(x, y, n) -> llvm.memcpy(align 1 x, align 1 y, n), x + n
1133 CallInst *NewCI =
1134 B.CreateMemCpy(Dst, Align(1), CI->getArgOperand(1), Align(1), N);
1135 // Propagate attributes, but memcpy has no return value, so make sure that
1136 // any return attributes are compliant.
1137 // TODO: Attach return value attributes to the 1st operand to preserve them?
1138 NewCI->setAttributes(CI->getAttributes());
1139 NewCI->removeAttributes(AttributeList::ReturnIndex,
1140 AttributeFuncs::typeIncompatible(NewCI->getType()));
1141 return B.CreateInBoundsGEP(B.getInt8Ty(), Dst, N);
1142 }
1143
optimizeMemMove(CallInst * CI,IRBuilderBase & B)1144 Value *LibCallSimplifier::optimizeMemMove(CallInst *CI, IRBuilderBase &B) {
1145 Value *Size = CI->getArgOperand(2);
1146 annotateNonNullAndDereferenceable(CI, {0, 1}, Size, DL);
1147 if (isa<IntrinsicInst>(CI))
1148 return nullptr;
1149
1150 // memmove(x, y, n) -> llvm.memmove(align 1 x, align 1 y, n)
1151 CallInst *NewCI = B.CreateMemMove(CI->getArgOperand(0), Align(1),
1152 CI->getArgOperand(1), Align(1), Size);
1153 NewCI->setAttributes(CI->getAttributes());
1154 NewCI->removeAttributes(AttributeList::ReturnIndex,
1155 AttributeFuncs::typeIncompatible(NewCI->getType()));
1156 return CI->getArgOperand(0);
1157 }
1158
1159 /// Fold memset[_chk](malloc(n), 0, n) --> calloc(1, n).
foldMallocMemset(CallInst * Memset,IRBuilderBase & B)1160 Value *LibCallSimplifier::foldMallocMemset(CallInst *Memset, IRBuilderBase &B) {
1161 // This has to be a memset of zeros (bzero).
1162 auto *FillValue = dyn_cast<ConstantInt>(Memset->getArgOperand(1));
1163 if (!FillValue || FillValue->getZExtValue() != 0)
1164 return nullptr;
1165
1166 // TODO: We should handle the case where the malloc has more than one use.
1167 // This is necessary to optimize common patterns such as when the result of
1168 // the malloc is checked against null or when a memset intrinsic is used in
1169 // place of a memset library call.
1170 auto *Malloc = dyn_cast<CallInst>(Memset->getArgOperand(0));
1171 if (!Malloc || !Malloc->hasOneUse())
1172 return nullptr;
1173
1174 // Is the inner call really malloc()?
1175 Function *InnerCallee = Malloc->getCalledFunction();
1176 if (!InnerCallee)
1177 return nullptr;
1178
1179 LibFunc Func;
1180 if (!TLI->getLibFunc(*InnerCallee, Func) || !TLI->has(Func) ||
1181 Func != LibFunc_malloc)
1182 return nullptr;
1183
1184 // The memset must cover the same number of bytes that are malloc'd.
1185 if (Memset->getArgOperand(2) != Malloc->getArgOperand(0))
1186 return nullptr;
1187
1188 // Replace the malloc with a calloc. We need the data layout to know what the
1189 // actual size of a 'size_t' parameter is.
1190 B.SetInsertPoint(Malloc->getParent(), ++Malloc->getIterator());
1191 const DataLayout &DL = Malloc->getModule()->getDataLayout();
1192 IntegerType *SizeType = DL.getIntPtrType(B.GetInsertBlock()->getContext());
1193 if (Value *Calloc = emitCalloc(ConstantInt::get(SizeType, 1),
1194 Malloc->getArgOperand(0),
1195 Malloc->getAttributes(), B, *TLI)) {
1196 substituteInParent(Malloc, Calloc);
1197 return Calloc;
1198 }
1199
1200 return nullptr;
1201 }
1202
optimizeMemSet(CallInst * CI,IRBuilderBase & B)1203 Value *LibCallSimplifier::optimizeMemSet(CallInst *CI, IRBuilderBase &B) {
1204 Value *Size = CI->getArgOperand(2);
1205 annotateNonNullAndDereferenceable(CI, 0, Size, DL);
1206 if (isa<IntrinsicInst>(CI))
1207 return nullptr;
1208
1209 if (auto *Calloc = foldMallocMemset(CI, B))
1210 return Calloc;
1211
1212 // memset(p, v, n) -> llvm.memset(align 1 p, v, n)
1213 Value *Val = B.CreateIntCast(CI->getArgOperand(1), B.getInt8Ty(), false);
1214 CallInst *NewCI = B.CreateMemSet(CI->getArgOperand(0), Val, Size, Align(1));
1215 NewCI->setAttributes(CI->getAttributes());
1216 NewCI->removeAttributes(AttributeList::ReturnIndex,
1217 AttributeFuncs::typeIncompatible(NewCI->getType()));
1218 return CI->getArgOperand(0);
1219 }
1220
optimizeRealloc(CallInst * CI,IRBuilderBase & B)1221 Value *LibCallSimplifier::optimizeRealloc(CallInst *CI, IRBuilderBase &B) {
1222 if (isa<ConstantPointerNull>(CI->getArgOperand(0)))
1223 return emitMalloc(CI->getArgOperand(1), B, DL, TLI);
1224
1225 return nullptr;
1226 }
1227
1228 //===----------------------------------------------------------------------===//
1229 // Math Library Optimizations
1230 //===----------------------------------------------------------------------===//
1231
1232 // Replace a libcall \p CI with a call to intrinsic \p IID
replaceUnaryCall(CallInst * CI,IRBuilderBase & B,Intrinsic::ID IID)1233 static Value *replaceUnaryCall(CallInst *CI, IRBuilderBase &B,
1234 Intrinsic::ID IID) {
1235 // Propagate fast-math flags from the existing call to the new call.
1236 IRBuilderBase::FastMathFlagGuard Guard(B);
1237 B.setFastMathFlags(CI->getFastMathFlags());
1238
1239 Module *M = CI->getModule();
1240 Value *V = CI->getArgOperand(0);
1241 Function *F = Intrinsic::getDeclaration(M, IID, CI->getType());
1242 CallInst *NewCall = B.CreateCall(F, V);
1243 NewCall->takeName(CI);
1244 return NewCall;
1245 }
1246
1247 /// Return a variant of Val with float type.
1248 /// Currently this works in two cases: If Val is an FPExtension of a float
1249 /// value to something bigger, simply return the operand.
1250 /// If Val is a ConstantFP but can be converted to a float ConstantFP without
1251 /// loss of precision do so.
valueHasFloatPrecision(Value * Val)1252 static Value *valueHasFloatPrecision(Value *Val) {
1253 if (FPExtInst *Cast = dyn_cast<FPExtInst>(Val)) {
1254 Value *Op = Cast->getOperand(0);
1255 if (Op->getType()->isFloatTy())
1256 return Op;
1257 }
1258 if (ConstantFP *Const = dyn_cast<ConstantFP>(Val)) {
1259 APFloat F = Const->getValueAPF();
1260 bool losesInfo;
1261 (void)F.convert(APFloat::IEEEsingle(), APFloat::rmNearestTiesToEven,
1262 &losesInfo);
1263 if (!losesInfo)
1264 return ConstantFP::get(Const->getContext(), F);
1265 }
1266 return nullptr;
1267 }
1268
1269 /// Shrink double -> float functions.
optimizeDoubleFP(CallInst * CI,IRBuilderBase & B,bool isBinary,bool isPrecise=false)1270 static Value *optimizeDoubleFP(CallInst *CI, IRBuilderBase &B,
1271 bool isBinary, bool isPrecise = false) {
1272 Function *CalleeFn = CI->getCalledFunction();
1273 if (!CI->getType()->isDoubleTy() || !CalleeFn)
1274 return nullptr;
1275
1276 // If not all the uses of the function are converted to float, then bail out.
1277 // This matters if the precision of the result is more important than the
1278 // precision of the arguments.
1279 if (isPrecise)
1280 for (User *U : CI->users()) {
1281 FPTruncInst *Cast = dyn_cast<FPTruncInst>(U);
1282 if (!Cast || !Cast->getType()->isFloatTy())
1283 return nullptr;
1284 }
1285
1286 // If this is something like 'g((double) float)', convert to 'gf(float)'.
1287 Value *V[2];
1288 V[0] = valueHasFloatPrecision(CI->getArgOperand(0));
1289 V[1] = isBinary ? valueHasFloatPrecision(CI->getArgOperand(1)) : nullptr;
1290 if (!V[0] || (isBinary && !V[1]))
1291 return nullptr;
1292
1293 // If call isn't an intrinsic, check that it isn't within a function with the
1294 // same name as the float version of this call, otherwise the result is an
1295 // infinite loop. For example, from MinGW-w64:
1296 //
1297 // float expf(float val) { return (float) exp((double) val); }
1298 StringRef CalleeName = CalleeFn->getName();
1299 bool IsIntrinsic = CalleeFn->isIntrinsic();
1300 if (!IsIntrinsic) {
1301 StringRef CallerName = CI->getFunction()->getName();
1302 if (!CallerName.empty() && CallerName.back() == 'f' &&
1303 CallerName.size() == (CalleeName.size() + 1) &&
1304 CallerName.startswith(CalleeName))
1305 return nullptr;
1306 }
1307
1308 // Propagate the math semantics from the current function to the new function.
1309 IRBuilderBase::FastMathFlagGuard Guard(B);
1310 B.setFastMathFlags(CI->getFastMathFlags());
1311
1312 // g((double) float) -> (double) gf(float)
1313 Value *R;
1314 if (IsIntrinsic) {
1315 Module *M = CI->getModule();
1316 Intrinsic::ID IID = CalleeFn->getIntrinsicID();
1317 Function *Fn = Intrinsic::getDeclaration(M, IID, B.getFloatTy());
1318 R = isBinary ? B.CreateCall(Fn, V) : B.CreateCall(Fn, V[0]);
1319 } else {
1320 AttributeList CalleeAttrs = CalleeFn->getAttributes();
1321 R = isBinary ? emitBinaryFloatFnCall(V[0], V[1], CalleeName, B, CalleeAttrs)
1322 : emitUnaryFloatFnCall(V[0], CalleeName, B, CalleeAttrs);
1323 }
1324 return B.CreateFPExt(R, B.getDoubleTy());
1325 }
1326
1327 /// Shrink double -> float for unary functions.
optimizeUnaryDoubleFP(CallInst * CI,IRBuilderBase & B,bool isPrecise=false)1328 static Value *optimizeUnaryDoubleFP(CallInst *CI, IRBuilderBase &B,
1329 bool isPrecise = false) {
1330 return optimizeDoubleFP(CI, B, false, isPrecise);
1331 }
1332
1333 /// Shrink double -> float for binary functions.
optimizeBinaryDoubleFP(CallInst * CI,IRBuilderBase & B,bool isPrecise=false)1334 static Value *optimizeBinaryDoubleFP(CallInst *CI, IRBuilderBase &B,
1335 bool isPrecise = false) {
1336 return optimizeDoubleFP(CI, B, true, isPrecise);
1337 }
1338
1339 // cabs(z) -> sqrt((creal(z)*creal(z)) + (cimag(z)*cimag(z)))
optimizeCAbs(CallInst * CI,IRBuilderBase & B)1340 Value *LibCallSimplifier::optimizeCAbs(CallInst *CI, IRBuilderBase &B) {
1341 if (!CI->isFast())
1342 return nullptr;
1343
1344 // Propagate fast-math flags from the existing call to new instructions.
1345 IRBuilderBase::FastMathFlagGuard Guard(B);
1346 B.setFastMathFlags(CI->getFastMathFlags());
1347
1348 Value *Real, *Imag;
1349 if (CI->getNumArgOperands() == 1) {
1350 Value *Op = CI->getArgOperand(0);
1351 assert(Op->getType()->isArrayTy() && "Unexpected signature for cabs!");
1352 Real = B.CreateExtractValue(Op, 0, "real");
1353 Imag = B.CreateExtractValue(Op, 1, "imag");
1354 } else {
1355 assert(CI->getNumArgOperands() == 2 && "Unexpected signature for cabs!");
1356 Real = CI->getArgOperand(0);
1357 Imag = CI->getArgOperand(1);
1358 }
1359
1360 Value *RealReal = B.CreateFMul(Real, Real);
1361 Value *ImagImag = B.CreateFMul(Imag, Imag);
1362
1363 Function *FSqrt = Intrinsic::getDeclaration(CI->getModule(), Intrinsic::sqrt,
1364 CI->getType());
1365 return B.CreateCall(FSqrt, B.CreateFAdd(RealReal, ImagImag), "cabs");
1366 }
1367
optimizeTrigReflections(CallInst * Call,LibFunc Func,IRBuilderBase & B)1368 static Value *optimizeTrigReflections(CallInst *Call, LibFunc Func,
1369 IRBuilderBase &B) {
1370 if (!isa<FPMathOperator>(Call))
1371 return nullptr;
1372
1373 IRBuilderBase::FastMathFlagGuard Guard(B);
1374 B.setFastMathFlags(Call->getFastMathFlags());
1375
1376 // TODO: Can this be shared to also handle LLVM intrinsics?
1377 Value *X;
1378 switch (Func) {
1379 case LibFunc_sin:
1380 case LibFunc_sinf:
1381 case LibFunc_sinl:
1382 case LibFunc_tan:
1383 case LibFunc_tanf:
1384 case LibFunc_tanl:
1385 // sin(-X) --> -sin(X)
1386 // tan(-X) --> -tan(X)
1387 if (match(Call->getArgOperand(0), m_OneUse(m_FNeg(m_Value(X)))))
1388 return B.CreateFNeg(B.CreateCall(Call->getCalledFunction(), X));
1389 break;
1390 case LibFunc_cos:
1391 case LibFunc_cosf:
1392 case LibFunc_cosl:
1393 // cos(-X) --> cos(X)
1394 if (match(Call->getArgOperand(0), m_FNeg(m_Value(X))))
1395 return B.CreateCall(Call->getCalledFunction(), X, "cos");
1396 break;
1397 default:
1398 break;
1399 }
1400 return nullptr;
1401 }
1402
getPow(Value * InnerChain[33],unsigned Exp,IRBuilderBase & B)1403 static Value *getPow(Value *InnerChain[33], unsigned Exp, IRBuilderBase &B) {
1404 // Multiplications calculated using Addition Chains.
1405 // Refer: http://wwwhomes.uni-bielefeld.de/achim/addition_chain.html
1406
1407 assert(Exp != 0 && "Incorrect exponent 0 not handled");
1408
1409 if (InnerChain[Exp])
1410 return InnerChain[Exp];
1411
1412 static const unsigned AddChain[33][2] = {
1413 {0, 0}, // Unused.
1414 {0, 0}, // Unused (base case = pow1).
1415 {1, 1}, // Unused (pre-computed).
1416 {1, 2}, {2, 2}, {2, 3}, {3, 3}, {2, 5}, {4, 4},
1417 {1, 8}, {5, 5}, {1, 10}, {6, 6}, {4, 9}, {7, 7},
1418 {3, 12}, {8, 8}, {8, 9}, {2, 16}, {1, 18}, {10, 10},
1419 {6, 15}, {11, 11}, {3, 20}, {12, 12}, {8, 17}, {13, 13},
1420 {3, 24}, {14, 14}, {4, 25}, {15, 15}, {3, 28}, {16, 16},
1421 };
1422
1423 InnerChain[Exp] = B.CreateFMul(getPow(InnerChain, AddChain[Exp][0], B),
1424 getPow(InnerChain, AddChain[Exp][1], B));
1425 return InnerChain[Exp];
1426 }
1427
1428 // Return a properly extended integer (DstWidth bits wide) if the operation is
1429 // an itofp.
getIntToFPVal(Value * I2F,IRBuilderBase & B,unsigned DstWidth)1430 static Value *getIntToFPVal(Value *I2F, IRBuilderBase &B, unsigned DstWidth) {
1431 if (isa<SIToFPInst>(I2F) || isa<UIToFPInst>(I2F)) {
1432 Value *Op = cast<Instruction>(I2F)->getOperand(0);
1433 // Make sure that the exponent fits inside an "int" of size DstWidth,
1434 // thus avoiding any range issues that FP has not.
1435 unsigned BitWidth = Op->getType()->getPrimitiveSizeInBits();
1436 if (BitWidth < DstWidth ||
1437 (BitWidth == DstWidth && isa<SIToFPInst>(I2F)))
1438 return isa<SIToFPInst>(I2F) ? B.CreateSExt(Op, B.getIntNTy(DstWidth))
1439 : B.CreateZExt(Op, B.getIntNTy(DstWidth));
1440 }
1441
1442 return nullptr;
1443 }
1444
1445 /// Use exp{,2}(x * y) for pow(exp{,2}(x), y);
1446 /// ldexp(1.0, x) for pow(2.0, itofp(x)); exp2(n * x) for pow(2.0 ** n, x);
1447 /// exp10(x) for pow(10.0, x); exp2(log2(n) * x) for pow(n, x).
replacePowWithExp(CallInst * Pow,IRBuilderBase & B)1448 Value *LibCallSimplifier::replacePowWithExp(CallInst *Pow, IRBuilderBase &B) {
1449 Value *Base = Pow->getArgOperand(0), *Expo = Pow->getArgOperand(1);
1450 AttributeList Attrs; // Attributes are only meaningful on the original call
1451 Module *Mod = Pow->getModule();
1452 Type *Ty = Pow->getType();
1453 bool Ignored;
1454
1455 // Evaluate special cases related to a nested function as the base.
1456
1457 // pow(exp(x), y) -> exp(x * y)
1458 // pow(exp2(x), y) -> exp2(x * y)
1459 // If exp{,2}() is used only once, it is better to fold two transcendental
1460 // math functions into one. If used again, exp{,2}() would still have to be
1461 // called with the original argument, then keep both original transcendental
1462 // functions. However, this transformation is only safe with fully relaxed
1463 // math semantics, since, besides rounding differences, it changes overflow
1464 // and underflow behavior quite dramatically. For example:
1465 // pow(exp(1000), 0.001) = pow(inf, 0.001) = inf
1466 // Whereas:
1467 // exp(1000 * 0.001) = exp(1)
1468 // TODO: Loosen the requirement for fully relaxed math semantics.
1469 // TODO: Handle exp10() when more targets have it available.
1470 CallInst *BaseFn = dyn_cast<CallInst>(Base);
1471 if (BaseFn && BaseFn->hasOneUse() && BaseFn->isFast() && Pow->isFast()) {
1472 LibFunc LibFn;
1473
1474 Function *CalleeFn = BaseFn->getCalledFunction();
1475 if (CalleeFn &&
1476 TLI->getLibFunc(CalleeFn->getName(), LibFn) && TLI->has(LibFn)) {
1477 StringRef ExpName;
1478 Intrinsic::ID ID;
1479 Value *ExpFn;
1480 LibFunc LibFnFloat, LibFnDouble, LibFnLongDouble;
1481
1482 switch (LibFn) {
1483 default:
1484 return nullptr;
1485 case LibFunc_expf: case LibFunc_exp: case LibFunc_expl:
1486 ExpName = TLI->getName(LibFunc_exp);
1487 ID = Intrinsic::exp;
1488 LibFnFloat = LibFunc_expf;
1489 LibFnDouble = LibFunc_exp;
1490 LibFnLongDouble = LibFunc_expl;
1491 break;
1492 case LibFunc_exp2f: case LibFunc_exp2: case LibFunc_exp2l:
1493 ExpName = TLI->getName(LibFunc_exp2);
1494 ID = Intrinsic::exp2;
1495 LibFnFloat = LibFunc_exp2f;
1496 LibFnDouble = LibFunc_exp2;
1497 LibFnLongDouble = LibFunc_exp2l;
1498 break;
1499 }
1500
1501 // Create new exp{,2}() with the product as its argument.
1502 Value *FMul = B.CreateFMul(BaseFn->getArgOperand(0), Expo, "mul");
1503 ExpFn = BaseFn->doesNotAccessMemory()
1504 ? B.CreateCall(Intrinsic::getDeclaration(Mod, ID, Ty),
1505 FMul, ExpName)
1506 : emitUnaryFloatFnCall(FMul, TLI, LibFnDouble, LibFnFloat,
1507 LibFnLongDouble, B,
1508 BaseFn->getAttributes());
1509
1510 // Since the new exp{,2}() is different from the original one, dead code
1511 // elimination cannot be trusted to remove it, since it may have side
1512 // effects (e.g., errno). When the only consumer for the original
1513 // exp{,2}() is pow(), then it has to be explicitly erased.
1514 substituteInParent(BaseFn, ExpFn);
1515 return ExpFn;
1516 }
1517 }
1518
1519 // Evaluate special cases related to a constant base.
1520
1521 const APFloat *BaseF;
1522 if (!match(Pow->getArgOperand(0), m_APFloat(BaseF)))
1523 return nullptr;
1524
1525 // pow(2.0, itofp(x)) -> ldexp(1.0, x)
1526 if (match(Base, m_SpecificFP(2.0)) &&
1527 (isa<SIToFPInst>(Expo) || isa<UIToFPInst>(Expo)) &&
1528 hasFloatFn(TLI, Ty, LibFunc_ldexp, LibFunc_ldexpf, LibFunc_ldexpl)) {
1529 if (Value *ExpoI = getIntToFPVal(Expo, B, TLI->getIntSize()))
1530 return emitBinaryFloatFnCall(ConstantFP::get(Ty, 1.0), ExpoI, TLI,
1531 LibFunc_ldexp, LibFunc_ldexpf, LibFunc_ldexpl,
1532 B, Attrs);
1533 }
1534
1535 // pow(2.0 ** n, x) -> exp2(n * x)
1536 if (hasFloatFn(TLI, Ty, LibFunc_exp2, LibFunc_exp2f, LibFunc_exp2l)) {
1537 APFloat BaseR = APFloat(1.0);
1538 BaseR.convert(BaseF->getSemantics(), APFloat::rmTowardZero, &Ignored);
1539 BaseR = BaseR / *BaseF;
1540 bool IsInteger = BaseF->isInteger(), IsReciprocal = BaseR.isInteger();
1541 const APFloat *NF = IsReciprocal ? &BaseR : BaseF;
1542 APSInt NI(64, false);
1543 if ((IsInteger || IsReciprocal) &&
1544 NF->convertToInteger(NI, APFloat::rmTowardZero, &Ignored) ==
1545 APFloat::opOK &&
1546 NI > 1 && NI.isPowerOf2()) {
1547 double N = NI.logBase2() * (IsReciprocal ? -1.0 : 1.0);
1548 Value *FMul = B.CreateFMul(Expo, ConstantFP::get(Ty, N), "mul");
1549 if (Pow->doesNotAccessMemory())
1550 return B.CreateCall(Intrinsic::getDeclaration(Mod, Intrinsic::exp2, Ty),
1551 FMul, "exp2");
1552 else
1553 return emitUnaryFloatFnCall(FMul, TLI, LibFunc_exp2, LibFunc_exp2f,
1554 LibFunc_exp2l, B, Attrs);
1555 }
1556 }
1557
1558 // pow(10.0, x) -> exp10(x)
1559 // TODO: There is no exp10() intrinsic yet, but some day there shall be one.
1560 if (match(Base, m_SpecificFP(10.0)) &&
1561 hasFloatFn(TLI, Ty, LibFunc_exp10, LibFunc_exp10f, LibFunc_exp10l))
1562 return emitUnaryFloatFnCall(Expo, TLI, LibFunc_exp10, LibFunc_exp10f,
1563 LibFunc_exp10l, B, Attrs);
1564
1565 // pow(x, y) -> exp2(log2(x) * y)
1566 if (Pow->hasApproxFunc() && Pow->hasNoNaNs() && BaseF->isFiniteNonZero() &&
1567 !BaseF->isNegative()) {
1568 // pow(1, inf) is defined to be 1 but exp2(log2(1) * inf) evaluates to NaN.
1569 // Luckily optimizePow has already handled the x == 1 case.
1570 assert(!match(Base, m_FPOne()) &&
1571 "pow(1.0, y) should have been simplified earlier!");
1572
1573 Value *Log = nullptr;
1574 if (Ty->isFloatTy())
1575 Log = ConstantFP::get(Ty, std::log2(BaseF->convertToFloat()));
1576 else if (Ty->isDoubleTy())
1577 Log = ConstantFP::get(Ty, std::log2(BaseF->convertToDouble()));
1578
1579 if (Log) {
1580 Value *FMul = B.CreateFMul(Log, Expo, "mul");
1581 if (Pow->doesNotAccessMemory())
1582 return B.CreateCall(Intrinsic::getDeclaration(Mod, Intrinsic::exp2, Ty),
1583 FMul, "exp2");
1584 else if (hasFloatFn(TLI, Ty, LibFunc_exp2, LibFunc_exp2f, LibFunc_exp2l))
1585 return emitUnaryFloatFnCall(FMul, TLI, LibFunc_exp2, LibFunc_exp2f,
1586 LibFunc_exp2l, B, Attrs);
1587 }
1588 }
1589
1590 return nullptr;
1591 }
1592
getSqrtCall(Value * V,AttributeList Attrs,bool NoErrno,Module * M,IRBuilderBase & B,const TargetLibraryInfo * TLI)1593 static Value *getSqrtCall(Value *V, AttributeList Attrs, bool NoErrno,
1594 Module *M, IRBuilderBase &B,
1595 const TargetLibraryInfo *TLI) {
1596 // If errno is never set, then use the intrinsic for sqrt().
1597 if (NoErrno) {
1598 Function *SqrtFn =
1599 Intrinsic::getDeclaration(M, Intrinsic::sqrt, V->getType());
1600 return B.CreateCall(SqrtFn, V, "sqrt");
1601 }
1602
1603 // Otherwise, use the libcall for sqrt().
1604 if (hasFloatFn(TLI, V->getType(), LibFunc_sqrt, LibFunc_sqrtf, LibFunc_sqrtl))
1605 // TODO: We also should check that the target can in fact lower the sqrt()
1606 // libcall. We currently have no way to ask this question, so we ask if
1607 // the target has a sqrt() libcall, which is not exactly the same.
1608 return emitUnaryFloatFnCall(V, TLI, LibFunc_sqrt, LibFunc_sqrtf,
1609 LibFunc_sqrtl, B, Attrs);
1610
1611 return nullptr;
1612 }
1613
1614 /// Use square root in place of pow(x, +/-0.5).
replacePowWithSqrt(CallInst * Pow,IRBuilderBase & B)1615 Value *LibCallSimplifier::replacePowWithSqrt(CallInst *Pow, IRBuilderBase &B) {
1616 Value *Sqrt, *Base = Pow->getArgOperand(0), *Expo = Pow->getArgOperand(1);
1617 AttributeList Attrs; // Attributes are only meaningful on the original call
1618 Module *Mod = Pow->getModule();
1619 Type *Ty = Pow->getType();
1620
1621 const APFloat *ExpoF;
1622 if (!match(Expo, m_APFloat(ExpoF)) ||
1623 (!ExpoF->isExactlyValue(0.5) && !ExpoF->isExactlyValue(-0.5)))
1624 return nullptr;
1625
1626 // Converting pow(X, -0.5) to 1/sqrt(X) may introduce an extra rounding step,
1627 // so that requires fast-math-flags (afn or reassoc).
1628 if (ExpoF->isNegative() && (!Pow->hasApproxFunc() && !Pow->hasAllowReassoc()))
1629 return nullptr;
1630
1631 // If we have a pow() library call (accesses memory) and we can't guarantee
1632 // that the base is not an infinity, give up:
1633 // pow(-Inf, 0.5) is optionally required to have a result of +Inf (not setting
1634 // errno), but sqrt(-Inf) is required by various standards to set errno.
1635 if (!Pow->doesNotAccessMemory() && !Pow->hasNoInfs() &&
1636 !isKnownNeverInfinity(Base, TLI))
1637 return nullptr;
1638
1639 Sqrt = getSqrtCall(Base, Attrs, Pow->doesNotAccessMemory(), Mod, B, TLI);
1640 if (!Sqrt)
1641 return nullptr;
1642
1643 // Handle signed zero base by expanding to fabs(sqrt(x)).
1644 if (!Pow->hasNoSignedZeros()) {
1645 Function *FAbsFn = Intrinsic::getDeclaration(Mod, Intrinsic::fabs, Ty);
1646 Sqrt = B.CreateCall(FAbsFn, Sqrt, "abs");
1647 }
1648
1649 // Handle non finite base by expanding to
1650 // (x == -infinity ? +infinity : sqrt(x)).
1651 if (!Pow->hasNoInfs()) {
1652 Value *PosInf = ConstantFP::getInfinity(Ty),
1653 *NegInf = ConstantFP::getInfinity(Ty, true);
1654 Value *FCmp = B.CreateFCmpOEQ(Base, NegInf, "isinf");
1655 Sqrt = B.CreateSelect(FCmp, PosInf, Sqrt);
1656 }
1657
1658 // If the exponent is negative, then get the reciprocal.
1659 if (ExpoF->isNegative())
1660 Sqrt = B.CreateFDiv(ConstantFP::get(Ty, 1.0), Sqrt, "reciprocal");
1661
1662 return Sqrt;
1663 }
1664
createPowWithIntegerExponent(Value * Base,Value * Expo,Module * M,IRBuilderBase & B)1665 static Value *createPowWithIntegerExponent(Value *Base, Value *Expo, Module *M,
1666 IRBuilderBase &B) {
1667 Value *Args[] = {Base, Expo};
1668 Type *Types[] = {Base->getType(), Expo->getType()};
1669 Function *F = Intrinsic::getDeclaration(M, Intrinsic::powi, Types);
1670 return B.CreateCall(F, Args);
1671 }
1672
optimizePow(CallInst * Pow,IRBuilderBase & B)1673 Value *LibCallSimplifier::optimizePow(CallInst *Pow, IRBuilderBase &B) {
1674 Value *Base = Pow->getArgOperand(0);
1675 Value *Expo = Pow->getArgOperand(1);
1676 Function *Callee = Pow->getCalledFunction();
1677 StringRef Name = Callee->getName();
1678 Type *Ty = Pow->getType();
1679 Module *M = Pow->getModule();
1680 bool AllowApprox = Pow->hasApproxFunc();
1681 bool Ignored;
1682
1683 // Propagate the math semantics from the call to any created instructions.
1684 IRBuilderBase::FastMathFlagGuard Guard(B);
1685 B.setFastMathFlags(Pow->getFastMathFlags());
1686 // Evaluate special cases related to the base.
1687
1688 // pow(1.0, x) -> 1.0
1689 if (match(Base, m_FPOne()))
1690 return Base;
1691
1692 if (Value *Exp = replacePowWithExp(Pow, B))
1693 return Exp;
1694
1695 // Evaluate special cases related to the exponent.
1696
1697 // pow(x, -1.0) -> 1.0 / x
1698 if (match(Expo, m_SpecificFP(-1.0)))
1699 return B.CreateFDiv(ConstantFP::get(Ty, 1.0), Base, "reciprocal");
1700
1701 // pow(x, +/-0.0) -> 1.0
1702 if (match(Expo, m_AnyZeroFP()))
1703 return ConstantFP::get(Ty, 1.0);
1704
1705 // pow(x, 1.0) -> x
1706 if (match(Expo, m_FPOne()))
1707 return Base;
1708
1709 // pow(x, 2.0) -> x * x
1710 if (match(Expo, m_SpecificFP(2.0)))
1711 return B.CreateFMul(Base, Base, "square");
1712
1713 if (Value *Sqrt = replacePowWithSqrt(Pow, B))
1714 return Sqrt;
1715
1716 // pow(x, n) -> x * x * x * ...
1717 const APFloat *ExpoF;
1718 if (AllowApprox && match(Expo, m_APFloat(ExpoF)) &&
1719 !ExpoF->isExactlyValue(0.5) && !ExpoF->isExactlyValue(-0.5)) {
1720 // We limit to a max of 7 multiplications, thus the maximum exponent is 32.
1721 // If the exponent is an integer+0.5 we generate a call to sqrt and an
1722 // additional fmul.
1723 // TODO: This whole transformation should be backend specific (e.g. some
1724 // backends might prefer libcalls or the limit for the exponent might
1725 // be different) and it should also consider optimizing for size.
1726 APFloat LimF(ExpoF->getSemantics(), 33),
1727 ExpoA(abs(*ExpoF));
1728 if (ExpoA < LimF) {
1729 // This transformation applies to integer or integer+0.5 exponents only.
1730 // For integer+0.5, we create a sqrt(Base) call.
1731 Value *Sqrt = nullptr;
1732 if (!ExpoA.isInteger()) {
1733 APFloat Expo2 = ExpoA;
1734 // To check if ExpoA is an integer + 0.5, we add it to itself. If there
1735 // is no floating point exception and the result is an integer, then
1736 // ExpoA == integer + 0.5
1737 if (Expo2.add(ExpoA, APFloat::rmNearestTiesToEven) != APFloat::opOK)
1738 return nullptr;
1739
1740 if (!Expo2.isInteger())
1741 return nullptr;
1742
1743 Sqrt = getSqrtCall(Base, Pow->getCalledFunction()->getAttributes(),
1744 Pow->doesNotAccessMemory(), M, B, TLI);
1745 if (!Sqrt)
1746 return nullptr;
1747 }
1748
1749 // We will memoize intermediate products of the Addition Chain.
1750 Value *InnerChain[33] = {nullptr};
1751 InnerChain[1] = Base;
1752 InnerChain[2] = B.CreateFMul(Base, Base, "square");
1753
1754 // We cannot readily convert a non-double type (like float) to a double.
1755 // So we first convert it to something which could be converted to double.
1756 ExpoA.convert(APFloat::IEEEdouble(), APFloat::rmTowardZero, &Ignored);
1757 Value *FMul = getPow(InnerChain, ExpoA.convertToDouble(), B);
1758
1759 // Expand pow(x, y+0.5) to pow(x, y) * sqrt(x).
1760 if (Sqrt)
1761 FMul = B.CreateFMul(FMul, Sqrt);
1762
1763 // If the exponent is negative, then get the reciprocal.
1764 if (ExpoF->isNegative())
1765 FMul = B.CreateFDiv(ConstantFP::get(Ty, 1.0), FMul, "reciprocal");
1766
1767 return FMul;
1768 }
1769
1770 APSInt IntExpo(TLI->getIntSize(), /*isUnsigned=*/false);
1771 // powf(x, n) -> powi(x, n) if n is a constant signed integer value
1772 if (ExpoF->isInteger() &&
1773 ExpoF->convertToInteger(IntExpo, APFloat::rmTowardZero, &Ignored) ==
1774 APFloat::opOK) {
1775 return createPowWithIntegerExponent(
1776 Base, ConstantInt::get(B.getIntNTy(TLI->getIntSize()), IntExpo), M, B);
1777 }
1778 }
1779
1780 // powf(x, itofp(y)) -> powi(x, y)
1781 if (AllowApprox && (isa<SIToFPInst>(Expo) || isa<UIToFPInst>(Expo))) {
1782 if (Value *ExpoI = getIntToFPVal(Expo, B, TLI->getIntSize()))
1783 return createPowWithIntegerExponent(Base, ExpoI, M, B);
1784 }
1785
1786 // Shrink pow() to powf() if the arguments are single precision,
1787 // unless the result is expected to be double precision.
1788 if (UnsafeFPShrink && Name == TLI->getName(LibFunc_pow) &&
1789 hasFloatVersion(Name)) {
1790 if (Value *Shrunk = optimizeBinaryDoubleFP(Pow, B, true))
1791 return Shrunk;
1792 }
1793
1794 return nullptr;
1795 }
1796
optimizeExp2(CallInst * CI,IRBuilderBase & B)1797 Value *LibCallSimplifier::optimizeExp2(CallInst *CI, IRBuilderBase &B) {
1798 Function *Callee = CI->getCalledFunction();
1799 AttributeList Attrs; // Attributes are only meaningful on the original call
1800 StringRef Name = Callee->getName();
1801 Value *Ret = nullptr;
1802 if (UnsafeFPShrink && Name == TLI->getName(LibFunc_exp2) &&
1803 hasFloatVersion(Name))
1804 Ret = optimizeUnaryDoubleFP(CI, B, true);
1805
1806 Type *Ty = CI->getType();
1807 Value *Op = CI->getArgOperand(0);
1808
1809 // Turn exp2(sitofp(x)) -> ldexp(1.0, sext(x)) if sizeof(x) <= IntSize
1810 // Turn exp2(uitofp(x)) -> ldexp(1.0, zext(x)) if sizeof(x) < IntSize
1811 if ((isa<SIToFPInst>(Op) || isa<UIToFPInst>(Op)) &&
1812 hasFloatFn(TLI, Ty, LibFunc_ldexp, LibFunc_ldexpf, LibFunc_ldexpl)) {
1813 if (Value *Exp = getIntToFPVal(Op, B, TLI->getIntSize()))
1814 return emitBinaryFloatFnCall(ConstantFP::get(Ty, 1.0), Exp, TLI,
1815 LibFunc_ldexp, LibFunc_ldexpf, LibFunc_ldexpl,
1816 B, Attrs);
1817 }
1818
1819 return Ret;
1820 }
1821
optimizeFMinFMax(CallInst * CI,IRBuilderBase & B)1822 Value *LibCallSimplifier::optimizeFMinFMax(CallInst *CI, IRBuilderBase &B) {
1823 // If we can shrink the call to a float function rather than a double
1824 // function, do that first.
1825 Function *Callee = CI->getCalledFunction();
1826 StringRef Name = Callee->getName();
1827 if ((Name == "fmin" || Name == "fmax") && hasFloatVersion(Name))
1828 if (Value *Ret = optimizeBinaryDoubleFP(CI, B))
1829 return Ret;
1830
1831 // The LLVM intrinsics minnum/maxnum correspond to fmin/fmax. Canonicalize to
1832 // the intrinsics for improved optimization (for example, vectorization).
1833 // No-signed-zeros is implied by the definitions of fmax/fmin themselves.
1834 // From the C standard draft WG14/N1256:
1835 // "Ideally, fmax would be sensitive to the sign of zero, for example
1836 // fmax(-0.0, +0.0) would return +0; however, implementation in software
1837 // might be impractical."
1838 IRBuilderBase::FastMathFlagGuard Guard(B);
1839 FastMathFlags FMF = CI->getFastMathFlags();
1840 FMF.setNoSignedZeros();
1841 B.setFastMathFlags(FMF);
1842
1843 Intrinsic::ID IID = Callee->getName().startswith("fmin") ? Intrinsic::minnum
1844 : Intrinsic::maxnum;
1845 Function *F = Intrinsic::getDeclaration(CI->getModule(), IID, CI->getType());
1846 return B.CreateCall(F, { CI->getArgOperand(0), CI->getArgOperand(1) });
1847 }
1848
optimizeLog(CallInst * Log,IRBuilderBase & B)1849 Value *LibCallSimplifier::optimizeLog(CallInst *Log, IRBuilderBase &B) {
1850 Function *LogFn = Log->getCalledFunction();
1851 AttributeList Attrs; // Attributes are only meaningful on the original call
1852 StringRef LogNm = LogFn->getName();
1853 Intrinsic::ID LogID = LogFn->getIntrinsicID();
1854 Module *Mod = Log->getModule();
1855 Type *Ty = Log->getType();
1856 Value *Ret = nullptr;
1857
1858 if (UnsafeFPShrink && hasFloatVersion(LogNm))
1859 Ret = optimizeUnaryDoubleFP(Log, B, true);
1860
1861 // The earlier call must also be 'fast' in order to do these transforms.
1862 CallInst *Arg = dyn_cast<CallInst>(Log->getArgOperand(0));
1863 if (!Log->isFast() || !Arg || !Arg->isFast() || !Arg->hasOneUse())
1864 return Ret;
1865
1866 LibFunc LogLb, ExpLb, Exp2Lb, Exp10Lb, PowLb;
1867
1868 // This is only applicable to log(), log2(), log10().
1869 if (TLI->getLibFunc(LogNm, LogLb))
1870 switch (LogLb) {
1871 case LibFunc_logf:
1872 LogID = Intrinsic::log;
1873 ExpLb = LibFunc_expf;
1874 Exp2Lb = LibFunc_exp2f;
1875 Exp10Lb = LibFunc_exp10f;
1876 PowLb = LibFunc_powf;
1877 break;
1878 case LibFunc_log:
1879 LogID = Intrinsic::log;
1880 ExpLb = LibFunc_exp;
1881 Exp2Lb = LibFunc_exp2;
1882 Exp10Lb = LibFunc_exp10;
1883 PowLb = LibFunc_pow;
1884 break;
1885 case LibFunc_logl:
1886 LogID = Intrinsic::log;
1887 ExpLb = LibFunc_expl;
1888 Exp2Lb = LibFunc_exp2l;
1889 Exp10Lb = LibFunc_exp10l;
1890 PowLb = LibFunc_powl;
1891 break;
1892 case LibFunc_log2f:
1893 LogID = Intrinsic::log2;
1894 ExpLb = LibFunc_expf;
1895 Exp2Lb = LibFunc_exp2f;
1896 Exp10Lb = LibFunc_exp10f;
1897 PowLb = LibFunc_powf;
1898 break;
1899 case LibFunc_log2:
1900 LogID = Intrinsic::log2;
1901 ExpLb = LibFunc_exp;
1902 Exp2Lb = LibFunc_exp2;
1903 Exp10Lb = LibFunc_exp10;
1904 PowLb = LibFunc_pow;
1905 break;
1906 case LibFunc_log2l:
1907 LogID = Intrinsic::log2;
1908 ExpLb = LibFunc_expl;
1909 Exp2Lb = LibFunc_exp2l;
1910 Exp10Lb = LibFunc_exp10l;
1911 PowLb = LibFunc_powl;
1912 break;
1913 case LibFunc_log10f:
1914 LogID = Intrinsic::log10;
1915 ExpLb = LibFunc_expf;
1916 Exp2Lb = LibFunc_exp2f;
1917 Exp10Lb = LibFunc_exp10f;
1918 PowLb = LibFunc_powf;
1919 break;
1920 case LibFunc_log10:
1921 LogID = Intrinsic::log10;
1922 ExpLb = LibFunc_exp;
1923 Exp2Lb = LibFunc_exp2;
1924 Exp10Lb = LibFunc_exp10;
1925 PowLb = LibFunc_pow;
1926 break;
1927 case LibFunc_log10l:
1928 LogID = Intrinsic::log10;
1929 ExpLb = LibFunc_expl;
1930 Exp2Lb = LibFunc_exp2l;
1931 Exp10Lb = LibFunc_exp10l;
1932 PowLb = LibFunc_powl;
1933 break;
1934 default:
1935 return Ret;
1936 }
1937 else if (LogID == Intrinsic::log || LogID == Intrinsic::log2 ||
1938 LogID == Intrinsic::log10) {
1939 if (Ty->getScalarType()->isFloatTy()) {
1940 ExpLb = LibFunc_expf;
1941 Exp2Lb = LibFunc_exp2f;
1942 Exp10Lb = LibFunc_exp10f;
1943 PowLb = LibFunc_powf;
1944 } else if (Ty->getScalarType()->isDoubleTy()) {
1945 ExpLb = LibFunc_exp;
1946 Exp2Lb = LibFunc_exp2;
1947 Exp10Lb = LibFunc_exp10;
1948 PowLb = LibFunc_pow;
1949 } else
1950 return Ret;
1951 } else
1952 return Ret;
1953
1954 IRBuilderBase::FastMathFlagGuard Guard(B);
1955 B.setFastMathFlags(FastMathFlags::getFast());
1956
1957 Intrinsic::ID ArgID = Arg->getIntrinsicID();
1958 LibFunc ArgLb = NotLibFunc;
1959 TLI->getLibFunc(*Arg, ArgLb);
1960
1961 // log(pow(x,y)) -> y*log(x)
1962 if (ArgLb == PowLb || ArgID == Intrinsic::pow) {
1963 Value *LogX =
1964 Log->doesNotAccessMemory()
1965 ? B.CreateCall(Intrinsic::getDeclaration(Mod, LogID, Ty),
1966 Arg->getOperand(0), "log")
1967 : emitUnaryFloatFnCall(Arg->getOperand(0), LogNm, B, Attrs);
1968 Value *MulY = B.CreateFMul(Arg->getArgOperand(1), LogX, "mul");
1969 // Since pow() may have side effects, e.g. errno,
1970 // dead code elimination may not be trusted to remove it.
1971 substituteInParent(Arg, MulY);
1972 return MulY;
1973 }
1974
1975 // log(exp{,2,10}(y)) -> y*log({e,2,10})
1976 // TODO: There is no exp10() intrinsic yet.
1977 if (ArgLb == ExpLb || ArgLb == Exp2Lb || ArgLb == Exp10Lb ||
1978 ArgID == Intrinsic::exp || ArgID == Intrinsic::exp2) {
1979 Constant *Eul;
1980 if (ArgLb == ExpLb || ArgID == Intrinsic::exp)
1981 // FIXME: Add more precise value of e for long double.
1982 Eul = ConstantFP::get(Log->getType(), numbers::e);
1983 else if (ArgLb == Exp2Lb || ArgID == Intrinsic::exp2)
1984 Eul = ConstantFP::get(Log->getType(), 2.0);
1985 else
1986 Eul = ConstantFP::get(Log->getType(), 10.0);
1987 Value *LogE = Log->doesNotAccessMemory()
1988 ? B.CreateCall(Intrinsic::getDeclaration(Mod, LogID, Ty),
1989 Eul, "log")
1990 : emitUnaryFloatFnCall(Eul, LogNm, B, Attrs);
1991 Value *MulY = B.CreateFMul(Arg->getArgOperand(0), LogE, "mul");
1992 // Since exp() may have side effects, e.g. errno,
1993 // dead code elimination may not be trusted to remove it.
1994 substituteInParent(Arg, MulY);
1995 return MulY;
1996 }
1997
1998 return Ret;
1999 }
2000
optimizeSqrt(CallInst * CI,IRBuilderBase & B)2001 Value *LibCallSimplifier::optimizeSqrt(CallInst *CI, IRBuilderBase &B) {
2002 Function *Callee = CI->getCalledFunction();
2003 Value *Ret = nullptr;
2004 // TODO: Once we have a way (other than checking for the existince of the
2005 // libcall) to tell whether our target can lower @llvm.sqrt, relax the
2006 // condition below.
2007 if (TLI->has(LibFunc_sqrtf) && (Callee->getName() == "sqrt" ||
2008 Callee->getIntrinsicID() == Intrinsic::sqrt))
2009 Ret = optimizeUnaryDoubleFP(CI, B, true);
2010
2011 if (!CI->isFast())
2012 return Ret;
2013
2014 Instruction *I = dyn_cast<Instruction>(CI->getArgOperand(0));
2015 if (!I || I->getOpcode() != Instruction::FMul || !I->isFast())
2016 return Ret;
2017
2018 // We're looking for a repeated factor in a multiplication tree,
2019 // so we can do this fold: sqrt(x * x) -> fabs(x);
2020 // or this fold: sqrt((x * x) * y) -> fabs(x) * sqrt(y).
2021 Value *Op0 = I->getOperand(0);
2022 Value *Op1 = I->getOperand(1);
2023 Value *RepeatOp = nullptr;
2024 Value *OtherOp = nullptr;
2025 if (Op0 == Op1) {
2026 // Simple match: the operands of the multiply are identical.
2027 RepeatOp = Op0;
2028 } else {
2029 // Look for a more complicated pattern: one of the operands is itself
2030 // a multiply, so search for a common factor in that multiply.
2031 // Note: We don't bother looking any deeper than this first level or for
2032 // variations of this pattern because instcombine's visitFMUL and/or the
2033 // reassociation pass should give us this form.
2034 Value *OtherMul0, *OtherMul1;
2035 if (match(Op0, m_FMul(m_Value(OtherMul0), m_Value(OtherMul1)))) {
2036 // Pattern: sqrt((x * y) * z)
2037 if (OtherMul0 == OtherMul1 && cast<Instruction>(Op0)->isFast()) {
2038 // Matched: sqrt((x * x) * z)
2039 RepeatOp = OtherMul0;
2040 OtherOp = Op1;
2041 }
2042 }
2043 }
2044 if (!RepeatOp)
2045 return Ret;
2046
2047 // Fast math flags for any created instructions should match the sqrt
2048 // and multiply.
2049 IRBuilderBase::FastMathFlagGuard Guard(B);
2050 B.setFastMathFlags(I->getFastMathFlags());
2051
2052 // If we found a repeated factor, hoist it out of the square root and
2053 // replace it with the fabs of that factor.
2054 Module *M = Callee->getParent();
2055 Type *ArgType = I->getType();
2056 Function *Fabs = Intrinsic::getDeclaration(M, Intrinsic::fabs, ArgType);
2057 Value *FabsCall = B.CreateCall(Fabs, RepeatOp, "fabs");
2058 if (OtherOp) {
2059 // If we found a non-repeated factor, we still need to get its square
2060 // root. We then multiply that by the value that was simplified out
2061 // of the square root calculation.
2062 Function *Sqrt = Intrinsic::getDeclaration(M, Intrinsic::sqrt, ArgType);
2063 Value *SqrtCall = B.CreateCall(Sqrt, OtherOp, "sqrt");
2064 return B.CreateFMul(FabsCall, SqrtCall);
2065 }
2066 return FabsCall;
2067 }
2068
2069 // TODO: Generalize to handle any trig function and its inverse.
optimizeTan(CallInst * CI,IRBuilderBase & B)2070 Value *LibCallSimplifier::optimizeTan(CallInst *CI, IRBuilderBase &B) {
2071 Function *Callee = CI->getCalledFunction();
2072 Value *Ret = nullptr;
2073 StringRef Name = Callee->getName();
2074 if (UnsafeFPShrink && Name == "tan" && hasFloatVersion(Name))
2075 Ret = optimizeUnaryDoubleFP(CI, B, true);
2076
2077 Value *Op1 = CI->getArgOperand(0);
2078 auto *OpC = dyn_cast<CallInst>(Op1);
2079 if (!OpC)
2080 return Ret;
2081
2082 // Both calls must be 'fast' in order to remove them.
2083 if (!CI->isFast() || !OpC->isFast())
2084 return Ret;
2085
2086 // tan(atan(x)) -> x
2087 // tanf(atanf(x)) -> x
2088 // tanl(atanl(x)) -> x
2089 LibFunc Func;
2090 Function *F = OpC->getCalledFunction();
2091 if (F && TLI->getLibFunc(F->getName(), Func) && TLI->has(Func) &&
2092 ((Func == LibFunc_atan && Callee->getName() == "tan") ||
2093 (Func == LibFunc_atanf && Callee->getName() == "tanf") ||
2094 (Func == LibFunc_atanl && Callee->getName() == "tanl")))
2095 Ret = OpC->getArgOperand(0);
2096 return Ret;
2097 }
2098
isTrigLibCall(CallInst * CI)2099 static bool isTrigLibCall(CallInst *CI) {
2100 // We can only hope to do anything useful if we can ignore things like errno
2101 // and floating-point exceptions.
2102 // We already checked the prototype.
2103 return CI->hasFnAttr(Attribute::NoUnwind) &&
2104 CI->hasFnAttr(Attribute::ReadNone);
2105 }
2106
insertSinCosCall(IRBuilderBase & B,Function * OrigCallee,Value * Arg,bool UseFloat,Value * & Sin,Value * & Cos,Value * & SinCos)2107 static void insertSinCosCall(IRBuilderBase &B, Function *OrigCallee, Value *Arg,
2108 bool UseFloat, Value *&Sin, Value *&Cos,
2109 Value *&SinCos) {
2110 Type *ArgTy = Arg->getType();
2111 Type *ResTy;
2112 StringRef Name;
2113
2114 Triple T(OrigCallee->getParent()->getTargetTriple());
2115 if (UseFloat) {
2116 Name = "__sincospif_stret";
2117
2118 assert(T.getArch() != Triple::x86 && "x86 messy and unsupported for now");
2119 // x86_64 can't use {float, float} since that would be returned in both
2120 // xmm0 and xmm1, which isn't what a real struct would do.
2121 ResTy = T.getArch() == Triple::x86_64
2122 ? static_cast<Type *>(FixedVectorType::get(ArgTy, 2))
2123 : static_cast<Type *>(StructType::get(ArgTy, ArgTy));
2124 } else {
2125 Name = "__sincospi_stret";
2126 ResTy = StructType::get(ArgTy, ArgTy);
2127 }
2128
2129 Module *M = OrigCallee->getParent();
2130 FunctionCallee Callee =
2131 M->getOrInsertFunction(Name, OrigCallee->getAttributes(), ResTy, ArgTy);
2132
2133 if (Instruction *ArgInst = dyn_cast<Instruction>(Arg)) {
2134 // If the argument is an instruction, it must dominate all uses so put our
2135 // sincos call there.
2136 B.SetInsertPoint(ArgInst->getParent(), ++ArgInst->getIterator());
2137 } else {
2138 // Otherwise (e.g. for a constant) the beginning of the function is as
2139 // good a place as any.
2140 BasicBlock &EntryBB = B.GetInsertBlock()->getParent()->getEntryBlock();
2141 B.SetInsertPoint(&EntryBB, EntryBB.begin());
2142 }
2143
2144 SinCos = B.CreateCall(Callee, Arg, "sincospi");
2145
2146 if (SinCos->getType()->isStructTy()) {
2147 Sin = B.CreateExtractValue(SinCos, 0, "sinpi");
2148 Cos = B.CreateExtractValue(SinCos, 1, "cospi");
2149 } else {
2150 Sin = B.CreateExtractElement(SinCos, ConstantInt::get(B.getInt32Ty(), 0),
2151 "sinpi");
2152 Cos = B.CreateExtractElement(SinCos, ConstantInt::get(B.getInt32Ty(), 1),
2153 "cospi");
2154 }
2155 }
2156
optimizeSinCosPi(CallInst * CI,IRBuilderBase & B)2157 Value *LibCallSimplifier::optimizeSinCosPi(CallInst *CI, IRBuilderBase &B) {
2158 // Make sure the prototype is as expected, otherwise the rest of the
2159 // function is probably invalid and likely to abort.
2160 if (!isTrigLibCall(CI))
2161 return nullptr;
2162
2163 Value *Arg = CI->getArgOperand(0);
2164 SmallVector<CallInst *, 1> SinCalls;
2165 SmallVector<CallInst *, 1> CosCalls;
2166 SmallVector<CallInst *, 1> SinCosCalls;
2167
2168 bool IsFloat = Arg->getType()->isFloatTy();
2169
2170 // Look for all compatible sinpi, cospi and sincospi calls with the same
2171 // argument. If there are enough (in some sense) we can make the
2172 // substitution.
2173 Function *F = CI->getFunction();
2174 for (User *U : Arg->users())
2175 classifyArgUse(U, F, IsFloat, SinCalls, CosCalls, SinCosCalls);
2176
2177 // It's only worthwhile if both sinpi and cospi are actually used.
2178 if (SinCalls.empty() || CosCalls.empty())
2179 return nullptr;
2180
2181 Value *Sin, *Cos, *SinCos;
2182 insertSinCosCall(B, CI->getCalledFunction(), Arg, IsFloat, Sin, Cos, SinCos);
2183
2184 auto replaceTrigInsts = [this](SmallVectorImpl<CallInst *> &Calls,
2185 Value *Res) {
2186 for (CallInst *C : Calls)
2187 replaceAllUsesWith(C, Res);
2188 };
2189
2190 replaceTrigInsts(SinCalls, Sin);
2191 replaceTrigInsts(CosCalls, Cos);
2192 replaceTrigInsts(SinCosCalls, SinCos);
2193
2194 return nullptr;
2195 }
2196
classifyArgUse(Value * Val,Function * F,bool IsFloat,SmallVectorImpl<CallInst * > & SinCalls,SmallVectorImpl<CallInst * > & CosCalls,SmallVectorImpl<CallInst * > & SinCosCalls)2197 void LibCallSimplifier::classifyArgUse(
2198 Value *Val, Function *F, bool IsFloat,
2199 SmallVectorImpl<CallInst *> &SinCalls,
2200 SmallVectorImpl<CallInst *> &CosCalls,
2201 SmallVectorImpl<CallInst *> &SinCosCalls) {
2202 CallInst *CI = dyn_cast<CallInst>(Val);
2203
2204 if (!CI || CI->use_empty())
2205 return;
2206
2207 // Don't consider calls in other functions.
2208 if (CI->getFunction() != F)
2209 return;
2210
2211 Function *Callee = CI->getCalledFunction();
2212 LibFunc Func;
2213 if (!Callee || !TLI->getLibFunc(*Callee, Func) || !TLI->has(Func) ||
2214 !isTrigLibCall(CI))
2215 return;
2216
2217 if (IsFloat) {
2218 if (Func == LibFunc_sinpif)
2219 SinCalls.push_back(CI);
2220 else if (Func == LibFunc_cospif)
2221 CosCalls.push_back(CI);
2222 else if (Func == LibFunc_sincospif_stret)
2223 SinCosCalls.push_back(CI);
2224 } else {
2225 if (Func == LibFunc_sinpi)
2226 SinCalls.push_back(CI);
2227 else if (Func == LibFunc_cospi)
2228 CosCalls.push_back(CI);
2229 else if (Func == LibFunc_sincospi_stret)
2230 SinCosCalls.push_back(CI);
2231 }
2232 }
2233
2234 //===----------------------------------------------------------------------===//
2235 // Integer Library Call Optimizations
2236 //===----------------------------------------------------------------------===//
2237
optimizeFFS(CallInst * CI,IRBuilderBase & B)2238 Value *LibCallSimplifier::optimizeFFS(CallInst *CI, IRBuilderBase &B) {
2239 // ffs(x) -> x != 0 ? (i32)llvm.cttz(x)+1 : 0
2240 Value *Op = CI->getArgOperand(0);
2241 Type *ArgType = Op->getType();
2242 Function *F = Intrinsic::getDeclaration(CI->getCalledFunction()->getParent(),
2243 Intrinsic::cttz, ArgType);
2244 Value *V = B.CreateCall(F, {Op, B.getTrue()}, "cttz");
2245 V = B.CreateAdd(V, ConstantInt::get(V->getType(), 1));
2246 V = B.CreateIntCast(V, B.getInt32Ty(), false);
2247
2248 Value *Cond = B.CreateICmpNE(Op, Constant::getNullValue(ArgType));
2249 return B.CreateSelect(Cond, V, B.getInt32(0));
2250 }
2251
optimizeFls(CallInst * CI,IRBuilderBase & B)2252 Value *LibCallSimplifier::optimizeFls(CallInst *CI, IRBuilderBase &B) {
2253 // fls(x) -> (i32)(sizeInBits(x) - llvm.ctlz(x, false))
2254 Value *Op = CI->getArgOperand(0);
2255 Type *ArgType = Op->getType();
2256 Function *F = Intrinsic::getDeclaration(CI->getCalledFunction()->getParent(),
2257 Intrinsic::ctlz, ArgType);
2258 Value *V = B.CreateCall(F, {Op, B.getFalse()}, "ctlz");
2259 V = B.CreateSub(ConstantInt::get(V->getType(), ArgType->getIntegerBitWidth()),
2260 V);
2261 return B.CreateIntCast(V, CI->getType(), false);
2262 }
2263
optimizeAbs(CallInst * CI,IRBuilderBase & B)2264 Value *LibCallSimplifier::optimizeAbs(CallInst *CI, IRBuilderBase &B) {
2265 // abs(x) -> x <s 0 ? -x : x
2266 // The negation has 'nsw' because abs of INT_MIN is undefined.
2267 Value *X = CI->getArgOperand(0);
2268 Value *IsNeg = B.CreateICmpSLT(X, Constant::getNullValue(X->getType()));
2269 Value *NegX = B.CreateNSWNeg(X, "neg");
2270 return B.CreateSelect(IsNeg, NegX, X);
2271 }
2272
optimizeIsDigit(CallInst * CI,IRBuilderBase & B)2273 Value *LibCallSimplifier::optimizeIsDigit(CallInst *CI, IRBuilderBase &B) {
2274 // isdigit(c) -> (c-'0') <u 10
2275 Value *Op = CI->getArgOperand(0);
2276 Op = B.CreateSub(Op, B.getInt32('0'), "isdigittmp");
2277 Op = B.CreateICmpULT(Op, B.getInt32(10), "isdigit");
2278 return B.CreateZExt(Op, CI->getType());
2279 }
2280
optimizeIsAscii(CallInst * CI,IRBuilderBase & B)2281 Value *LibCallSimplifier::optimizeIsAscii(CallInst *CI, IRBuilderBase &B) {
2282 // isascii(c) -> c <u 128
2283 Value *Op = CI->getArgOperand(0);
2284 Op = B.CreateICmpULT(Op, B.getInt32(128), "isascii");
2285 return B.CreateZExt(Op, CI->getType());
2286 }
2287
optimizeToAscii(CallInst * CI,IRBuilderBase & B)2288 Value *LibCallSimplifier::optimizeToAscii(CallInst *CI, IRBuilderBase &B) {
2289 // toascii(c) -> c & 0x7f
2290 return B.CreateAnd(CI->getArgOperand(0),
2291 ConstantInt::get(CI->getType(), 0x7F));
2292 }
2293
optimizeAtoi(CallInst * CI,IRBuilderBase & B)2294 Value *LibCallSimplifier::optimizeAtoi(CallInst *CI, IRBuilderBase &B) {
2295 StringRef Str;
2296 if (!getConstantStringInfo(CI->getArgOperand(0), Str))
2297 return nullptr;
2298
2299 return convertStrToNumber(CI, Str, 10);
2300 }
2301
optimizeStrtol(CallInst * CI,IRBuilderBase & B)2302 Value *LibCallSimplifier::optimizeStrtol(CallInst *CI, IRBuilderBase &B) {
2303 StringRef Str;
2304 if (!getConstantStringInfo(CI->getArgOperand(0), Str))
2305 return nullptr;
2306
2307 if (!isa<ConstantPointerNull>(CI->getArgOperand(1)))
2308 return nullptr;
2309
2310 if (ConstantInt *CInt = dyn_cast<ConstantInt>(CI->getArgOperand(2))) {
2311 return convertStrToNumber(CI, Str, CInt->getSExtValue());
2312 }
2313
2314 return nullptr;
2315 }
2316
2317 //===----------------------------------------------------------------------===//
2318 // Formatting and IO Library Call Optimizations
2319 //===----------------------------------------------------------------------===//
2320
2321 static bool isReportingError(Function *Callee, CallInst *CI, int StreamArg);
2322
optimizeErrorReporting(CallInst * CI,IRBuilderBase & B,int StreamArg)2323 Value *LibCallSimplifier::optimizeErrorReporting(CallInst *CI, IRBuilderBase &B,
2324 int StreamArg) {
2325 Function *Callee = CI->getCalledFunction();
2326 // Error reporting calls should be cold, mark them as such.
2327 // This applies even to non-builtin calls: it is only a hint and applies to
2328 // functions that the frontend might not understand as builtins.
2329
2330 // This heuristic was suggested in:
2331 // Improving Static Branch Prediction in a Compiler
2332 // Brian L. Deitrich, Ben-Chung Cheng, Wen-mei W. Hwu
2333 // Proceedings of PACT'98, Oct. 1998, IEEE
2334 if (!CI->hasFnAttr(Attribute::Cold) &&
2335 isReportingError(Callee, CI, StreamArg)) {
2336 CI->addAttribute(AttributeList::FunctionIndex, Attribute::Cold);
2337 }
2338
2339 return nullptr;
2340 }
2341
isReportingError(Function * Callee,CallInst * CI,int StreamArg)2342 static bool isReportingError(Function *Callee, CallInst *CI, int StreamArg) {
2343 if (!Callee || !Callee->isDeclaration())
2344 return false;
2345
2346 if (StreamArg < 0)
2347 return true;
2348
2349 // These functions might be considered cold, but only if their stream
2350 // argument is stderr.
2351
2352 if (StreamArg >= (int)CI->getNumArgOperands())
2353 return false;
2354 LoadInst *LI = dyn_cast<LoadInst>(CI->getArgOperand(StreamArg));
2355 if (!LI)
2356 return false;
2357 GlobalVariable *GV = dyn_cast<GlobalVariable>(LI->getPointerOperand());
2358 if (!GV || !GV->isDeclaration())
2359 return false;
2360 return GV->getName() == "stderr";
2361 }
2362
optimizePrintFString(CallInst * CI,IRBuilderBase & B)2363 Value *LibCallSimplifier::optimizePrintFString(CallInst *CI, IRBuilderBase &B) {
2364 // Check for a fixed format string.
2365 StringRef FormatStr;
2366 if (!getConstantStringInfo(CI->getArgOperand(0), FormatStr))
2367 return nullptr;
2368
2369 // Empty format string -> noop.
2370 if (FormatStr.empty()) // Tolerate printf's declared void.
2371 return CI->use_empty() ? (Value *)CI : ConstantInt::get(CI->getType(), 0);
2372
2373 // Do not do any of the following transformations if the printf return value
2374 // is used, in general the printf return value is not compatible with either
2375 // putchar() or puts().
2376 if (!CI->use_empty())
2377 return nullptr;
2378
2379 // printf("x") -> putchar('x'), even for "%" and "%%".
2380 if (FormatStr.size() == 1 || FormatStr == "%%")
2381 return emitPutChar(B.getInt32(FormatStr[0]), B, TLI);
2382
2383 // Try to remove call or emit putchar/puts.
2384 if (FormatStr == "%s" && CI->getNumArgOperands() > 1) {
2385 StringRef OperandStr;
2386 if (!getConstantStringInfo(CI->getOperand(1), OperandStr))
2387 return nullptr;
2388 // printf("%s", "") --> NOP
2389 if (OperandStr.empty())
2390 return (Value *)CI;
2391 // printf("%s", "a") --> putchar('a')
2392 if (OperandStr.size() == 1)
2393 return emitPutChar(B.getInt32(OperandStr[0]), B, TLI);
2394 // printf("%s", str"\n") --> puts(str)
2395 if (OperandStr.back() == '\n') {
2396 OperandStr = OperandStr.drop_back();
2397 Value *GV = B.CreateGlobalString(OperandStr, "str");
2398 return emitPutS(GV, B, TLI);
2399 }
2400 return nullptr;
2401 }
2402
2403 // printf("foo\n") --> puts("foo")
2404 if (FormatStr.back() == '\n' &&
2405 FormatStr.find('%') == StringRef::npos) { // No format characters.
2406 // Create a string literal with no \n on it. We expect the constant merge
2407 // pass to be run after this pass, to merge duplicate strings.
2408 FormatStr = FormatStr.drop_back();
2409 Value *GV = B.CreateGlobalString(FormatStr, "str");
2410 return emitPutS(GV, B, TLI);
2411 }
2412
2413 // Optimize specific format strings.
2414 // printf("%c", chr) --> putchar(chr)
2415 if (FormatStr == "%c" && CI->getNumArgOperands() > 1 &&
2416 CI->getArgOperand(1)->getType()->isIntegerTy())
2417 return emitPutChar(CI->getArgOperand(1), B, TLI);
2418
2419 // printf("%s\n", str) --> puts(str)
2420 if (FormatStr == "%s\n" && CI->getNumArgOperands() > 1 &&
2421 CI->getArgOperand(1)->getType()->isPointerTy())
2422 return emitPutS(CI->getArgOperand(1), B, TLI);
2423 return nullptr;
2424 }
2425
optimizePrintF(CallInst * CI,IRBuilderBase & B)2426 Value *LibCallSimplifier::optimizePrintF(CallInst *CI, IRBuilderBase &B) {
2427
2428 Function *Callee = CI->getCalledFunction();
2429 FunctionType *FT = Callee->getFunctionType();
2430 if (Value *V = optimizePrintFString(CI, B)) {
2431 return V;
2432 }
2433
2434 // printf(format, ...) -> iprintf(format, ...) if no floating point
2435 // arguments.
2436 if (TLI->has(LibFunc_iprintf) && !callHasFloatingPointArgument(CI)) {
2437 Module *M = B.GetInsertBlock()->getParent()->getParent();
2438 FunctionCallee IPrintFFn =
2439 M->getOrInsertFunction("iprintf", FT, Callee->getAttributes());
2440 CallInst *New = cast<CallInst>(CI->clone());
2441 New->setCalledFunction(IPrintFFn);
2442 B.Insert(New);
2443 return New;
2444 }
2445
2446 // printf(format, ...) -> __small_printf(format, ...) if no 128-bit floating point
2447 // arguments.
2448 if (TLI->has(LibFunc_small_printf) && !callHasFP128Argument(CI)) {
2449 Module *M = B.GetInsertBlock()->getParent()->getParent();
2450 auto SmallPrintFFn =
2451 M->getOrInsertFunction(TLI->getName(LibFunc_small_printf),
2452 FT, Callee->getAttributes());
2453 CallInst *New = cast<CallInst>(CI->clone());
2454 New->setCalledFunction(SmallPrintFFn);
2455 B.Insert(New);
2456 return New;
2457 }
2458
2459 annotateNonNullNoUndefBasedOnAccess(CI, 0);
2460 return nullptr;
2461 }
2462
optimizeSPrintFString(CallInst * CI,IRBuilderBase & B)2463 Value *LibCallSimplifier::optimizeSPrintFString(CallInst *CI,
2464 IRBuilderBase &B) {
2465 // Check for a fixed format string.
2466 StringRef FormatStr;
2467 if (!getConstantStringInfo(CI->getArgOperand(1), FormatStr))
2468 return nullptr;
2469
2470 // If we just have a format string (nothing else crazy) transform it.
2471 Value *Dest = CI->getArgOperand(0);
2472 if (CI->getNumArgOperands() == 2) {
2473 // Make sure there's no % in the constant array. We could try to handle
2474 // %% -> % in the future if we cared.
2475 if (FormatStr.find('%') != StringRef::npos)
2476 return nullptr; // we found a format specifier, bail out.
2477
2478 // sprintf(str, fmt) -> llvm.memcpy(align 1 str, align 1 fmt, strlen(fmt)+1)
2479 B.CreateMemCpy(
2480 Dest, Align(1), CI->getArgOperand(1), Align(1),
2481 ConstantInt::get(DL.getIntPtrType(CI->getContext()),
2482 FormatStr.size() + 1)); // Copy the null byte.
2483 return ConstantInt::get(CI->getType(), FormatStr.size());
2484 }
2485
2486 // The remaining optimizations require the format string to be "%s" or "%c"
2487 // and have an extra operand.
2488 if (FormatStr.size() != 2 || FormatStr[0] != '%' ||
2489 CI->getNumArgOperands() < 3)
2490 return nullptr;
2491
2492 // Decode the second character of the format string.
2493 if (FormatStr[1] == 'c') {
2494 // sprintf(dst, "%c", chr) --> *(i8*)dst = chr; *((i8*)dst+1) = 0
2495 if (!CI->getArgOperand(2)->getType()->isIntegerTy())
2496 return nullptr;
2497 Value *V = B.CreateTrunc(CI->getArgOperand(2), B.getInt8Ty(), "char");
2498 Value *Ptr = castToCStr(Dest, B);
2499 B.CreateStore(V, Ptr);
2500 Ptr = B.CreateGEP(B.getInt8Ty(), Ptr, B.getInt32(1), "nul");
2501 B.CreateStore(B.getInt8(0), Ptr);
2502
2503 return ConstantInt::get(CI->getType(), 1);
2504 }
2505
2506 if (FormatStr[1] == 's') {
2507 // sprintf(dest, "%s", str) -> llvm.memcpy(align 1 dest, align 1 str,
2508 // strlen(str)+1)
2509 if (!CI->getArgOperand(2)->getType()->isPointerTy())
2510 return nullptr;
2511
2512 if (CI->use_empty())
2513 // sprintf(dest, "%s", str) -> strcpy(dest, str)
2514 return emitStrCpy(Dest, CI->getArgOperand(2), B, TLI);
2515
2516 uint64_t SrcLen = GetStringLength(CI->getArgOperand(2));
2517 if (SrcLen) {
2518 B.CreateMemCpy(
2519 Dest, Align(1), CI->getArgOperand(2), Align(1),
2520 ConstantInt::get(DL.getIntPtrType(CI->getContext()), SrcLen));
2521 // Returns total number of characters written without null-character.
2522 return ConstantInt::get(CI->getType(), SrcLen - 1);
2523 } else if (Value *V = emitStpCpy(Dest, CI->getArgOperand(2), B, TLI)) {
2524 // sprintf(dest, "%s", str) -> stpcpy(dest, str) - dest
2525 // Handle mismatched pointer types (goes away with typeless pointers?).
2526 V = B.CreatePointerCast(V, Dest->getType());
2527 Value *PtrDiff = B.CreatePtrDiff(V, Dest);
2528 return B.CreateIntCast(PtrDiff, CI->getType(), false);
2529 }
2530
2531 bool OptForSize = CI->getFunction()->hasOptSize() ||
2532 llvm::shouldOptimizeForSize(CI->getParent(), PSI, BFI,
2533 PGSOQueryType::IRPass);
2534 if (OptForSize)
2535 return nullptr;
2536
2537 Value *Len = emitStrLen(CI->getArgOperand(2), B, DL, TLI);
2538 if (!Len)
2539 return nullptr;
2540 Value *IncLen =
2541 B.CreateAdd(Len, ConstantInt::get(Len->getType(), 1), "leninc");
2542 B.CreateMemCpy(Dest, Align(1), CI->getArgOperand(2), Align(1), IncLen);
2543
2544 // The sprintf result is the unincremented number of bytes in the string.
2545 return B.CreateIntCast(Len, CI->getType(), false);
2546 }
2547 return nullptr;
2548 }
2549
optimizeSPrintF(CallInst * CI,IRBuilderBase & B)2550 Value *LibCallSimplifier::optimizeSPrintF(CallInst *CI, IRBuilderBase &B) {
2551 Function *Callee = CI->getCalledFunction();
2552 FunctionType *FT = Callee->getFunctionType();
2553 if (Value *V = optimizeSPrintFString(CI, B)) {
2554 return V;
2555 }
2556
2557 // sprintf(str, format, ...) -> siprintf(str, format, ...) if no floating
2558 // point arguments.
2559 if (TLI->has(LibFunc_siprintf) && !callHasFloatingPointArgument(CI)) {
2560 Module *M = B.GetInsertBlock()->getParent()->getParent();
2561 FunctionCallee SIPrintFFn =
2562 M->getOrInsertFunction("siprintf", FT, Callee->getAttributes());
2563 CallInst *New = cast<CallInst>(CI->clone());
2564 New->setCalledFunction(SIPrintFFn);
2565 B.Insert(New);
2566 return New;
2567 }
2568
2569 // sprintf(str, format, ...) -> __small_sprintf(str, format, ...) if no 128-bit
2570 // floating point arguments.
2571 if (TLI->has(LibFunc_small_sprintf) && !callHasFP128Argument(CI)) {
2572 Module *M = B.GetInsertBlock()->getParent()->getParent();
2573 auto SmallSPrintFFn =
2574 M->getOrInsertFunction(TLI->getName(LibFunc_small_sprintf),
2575 FT, Callee->getAttributes());
2576 CallInst *New = cast<CallInst>(CI->clone());
2577 New->setCalledFunction(SmallSPrintFFn);
2578 B.Insert(New);
2579 return New;
2580 }
2581
2582 annotateNonNullNoUndefBasedOnAccess(CI, {0, 1});
2583 return nullptr;
2584 }
2585
optimizeSnPrintFString(CallInst * CI,IRBuilderBase & B)2586 Value *LibCallSimplifier::optimizeSnPrintFString(CallInst *CI,
2587 IRBuilderBase &B) {
2588 // Check for size
2589 ConstantInt *Size = dyn_cast<ConstantInt>(CI->getArgOperand(1));
2590 if (!Size)
2591 return nullptr;
2592
2593 uint64_t N = Size->getZExtValue();
2594 // Check for a fixed format string.
2595 StringRef FormatStr;
2596 if (!getConstantStringInfo(CI->getArgOperand(2), FormatStr))
2597 return nullptr;
2598
2599 // If we just have a format string (nothing else crazy) transform it.
2600 if (CI->getNumArgOperands() == 3) {
2601 // Make sure there's no % in the constant array. We could try to handle
2602 // %% -> % in the future if we cared.
2603 if (FormatStr.find('%') != StringRef::npos)
2604 return nullptr; // we found a format specifier, bail out.
2605
2606 if (N == 0)
2607 return ConstantInt::get(CI->getType(), FormatStr.size());
2608 else if (N < FormatStr.size() + 1)
2609 return nullptr;
2610
2611 // snprintf(dst, size, fmt) -> llvm.memcpy(align 1 dst, align 1 fmt,
2612 // strlen(fmt)+1)
2613 B.CreateMemCpy(
2614 CI->getArgOperand(0), Align(1), CI->getArgOperand(2), Align(1),
2615 ConstantInt::get(DL.getIntPtrType(CI->getContext()),
2616 FormatStr.size() + 1)); // Copy the null byte.
2617 return ConstantInt::get(CI->getType(), FormatStr.size());
2618 }
2619
2620 // The remaining optimizations require the format string to be "%s" or "%c"
2621 // and have an extra operand.
2622 if (FormatStr.size() == 2 && FormatStr[0] == '%' &&
2623 CI->getNumArgOperands() == 4) {
2624
2625 // Decode the second character of the format string.
2626 if (FormatStr[1] == 'c') {
2627 if (N == 0)
2628 return ConstantInt::get(CI->getType(), 1);
2629 else if (N == 1)
2630 return nullptr;
2631
2632 // snprintf(dst, size, "%c", chr) --> *(i8*)dst = chr; *((i8*)dst+1) = 0
2633 if (!CI->getArgOperand(3)->getType()->isIntegerTy())
2634 return nullptr;
2635 Value *V = B.CreateTrunc(CI->getArgOperand(3), B.getInt8Ty(), "char");
2636 Value *Ptr = castToCStr(CI->getArgOperand(0), B);
2637 B.CreateStore(V, Ptr);
2638 Ptr = B.CreateGEP(B.getInt8Ty(), Ptr, B.getInt32(1), "nul");
2639 B.CreateStore(B.getInt8(0), Ptr);
2640
2641 return ConstantInt::get(CI->getType(), 1);
2642 }
2643
2644 if (FormatStr[1] == 's') {
2645 // snprintf(dest, size, "%s", str) to llvm.memcpy(dest, str, len+1, 1)
2646 StringRef Str;
2647 if (!getConstantStringInfo(CI->getArgOperand(3), Str))
2648 return nullptr;
2649
2650 if (N == 0)
2651 return ConstantInt::get(CI->getType(), Str.size());
2652 else if (N < Str.size() + 1)
2653 return nullptr;
2654
2655 B.CreateMemCpy(CI->getArgOperand(0), Align(1), CI->getArgOperand(3),
2656 Align(1), ConstantInt::get(CI->getType(), Str.size() + 1));
2657
2658 // The snprintf result is the unincremented number of bytes in the string.
2659 return ConstantInt::get(CI->getType(), Str.size());
2660 }
2661 }
2662 return nullptr;
2663 }
2664
optimizeSnPrintF(CallInst * CI,IRBuilderBase & B)2665 Value *LibCallSimplifier::optimizeSnPrintF(CallInst *CI, IRBuilderBase &B) {
2666 if (Value *V = optimizeSnPrintFString(CI, B)) {
2667 return V;
2668 }
2669
2670 if (isKnownNonZero(CI->getOperand(1), DL))
2671 annotateNonNullNoUndefBasedOnAccess(CI, 0);
2672 return nullptr;
2673 }
2674
optimizeFPrintFString(CallInst * CI,IRBuilderBase & B)2675 Value *LibCallSimplifier::optimizeFPrintFString(CallInst *CI,
2676 IRBuilderBase &B) {
2677 optimizeErrorReporting(CI, B, 0);
2678
2679 // All the optimizations depend on the format string.
2680 StringRef FormatStr;
2681 if (!getConstantStringInfo(CI->getArgOperand(1), FormatStr))
2682 return nullptr;
2683
2684 // Do not do any of the following transformations if the fprintf return
2685 // value is used, in general the fprintf return value is not compatible
2686 // with fwrite(), fputc() or fputs().
2687 if (!CI->use_empty())
2688 return nullptr;
2689
2690 // fprintf(F, "foo") --> fwrite("foo", 3, 1, F)
2691 if (CI->getNumArgOperands() == 2) {
2692 // Could handle %% -> % if we cared.
2693 if (FormatStr.find('%') != StringRef::npos)
2694 return nullptr; // We found a format specifier.
2695
2696 return emitFWrite(
2697 CI->getArgOperand(1),
2698 ConstantInt::get(DL.getIntPtrType(CI->getContext()), FormatStr.size()),
2699 CI->getArgOperand(0), B, DL, TLI);
2700 }
2701
2702 // The remaining optimizations require the format string to be "%s" or "%c"
2703 // and have an extra operand.
2704 if (FormatStr.size() != 2 || FormatStr[0] != '%' ||
2705 CI->getNumArgOperands() < 3)
2706 return nullptr;
2707
2708 // Decode the second character of the format string.
2709 if (FormatStr[1] == 'c') {
2710 // fprintf(F, "%c", chr) --> fputc(chr, F)
2711 if (!CI->getArgOperand(2)->getType()->isIntegerTy())
2712 return nullptr;
2713 return emitFPutC(CI->getArgOperand(2), CI->getArgOperand(0), B, TLI);
2714 }
2715
2716 if (FormatStr[1] == 's') {
2717 // fprintf(F, "%s", str) --> fputs(str, F)
2718 if (!CI->getArgOperand(2)->getType()->isPointerTy())
2719 return nullptr;
2720 return emitFPutS(CI->getArgOperand(2), CI->getArgOperand(0), B, TLI);
2721 }
2722 return nullptr;
2723 }
2724
optimizeFPrintF(CallInst * CI,IRBuilderBase & B)2725 Value *LibCallSimplifier::optimizeFPrintF(CallInst *CI, IRBuilderBase &B) {
2726 Function *Callee = CI->getCalledFunction();
2727 FunctionType *FT = Callee->getFunctionType();
2728 if (Value *V = optimizeFPrintFString(CI, B)) {
2729 return V;
2730 }
2731
2732 // fprintf(stream, format, ...) -> fiprintf(stream, format, ...) if no
2733 // floating point arguments.
2734 if (TLI->has(LibFunc_fiprintf) && !callHasFloatingPointArgument(CI)) {
2735 Module *M = B.GetInsertBlock()->getParent()->getParent();
2736 FunctionCallee FIPrintFFn =
2737 M->getOrInsertFunction("fiprintf", FT, Callee->getAttributes());
2738 CallInst *New = cast<CallInst>(CI->clone());
2739 New->setCalledFunction(FIPrintFFn);
2740 B.Insert(New);
2741 return New;
2742 }
2743
2744 // fprintf(stream, format, ...) -> __small_fprintf(stream, format, ...) if no
2745 // 128-bit floating point arguments.
2746 if (TLI->has(LibFunc_small_fprintf) && !callHasFP128Argument(CI)) {
2747 Module *M = B.GetInsertBlock()->getParent()->getParent();
2748 auto SmallFPrintFFn =
2749 M->getOrInsertFunction(TLI->getName(LibFunc_small_fprintf),
2750 FT, Callee->getAttributes());
2751 CallInst *New = cast<CallInst>(CI->clone());
2752 New->setCalledFunction(SmallFPrintFFn);
2753 B.Insert(New);
2754 return New;
2755 }
2756
2757 return nullptr;
2758 }
2759
optimizeFWrite(CallInst * CI,IRBuilderBase & B)2760 Value *LibCallSimplifier::optimizeFWrite(CallInst *CI, IRBuilderBase &B) {
2761 optimizeErrorReporting(CI, B, 3);
2762
2763 // Get the element size and count.
2764 ConstantInt *SizeC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
2765 ConstantInt *CountC = dyn_cast<ConstantInt>(CI->getArgOperand(2));
2766 if (SizeC && CountC) {
2767 uint64_t Bytes = SizeC->getZExtValue() * CountC->getZExtValue();
2768
2769 // If this is writing zero records, remove the call (it's a noop).
2770 if (Bytes == 0)
2771 return ConstantInt::get(CI->getType(), 0);
2772
2773 // If this is writing one byte, turn it into fputc.
2774 // This optimisation is only valid, if the return value is unused.
2775 if (Bytes == 1 && CI->use_empty()) { // fwrite(S,1,1,F) -> fputc(S[0],F)
2776 Value *Char = B.CreateLoad(B.getInt8Ty(),
2777 castToCStr(CI->getArgOperand(0), B), "char");
2778 Value *NewCI = emitFPutC(Char, CI->getArgOperand(3), B, TLI);
2779 return NewCI ? ConstantInt::get(CI->getType(), 1) : nullptr;
2780 }
2781 }
2782
2783 return nullptr;
2784 }
2785
optimizeFPuts(CallInst * CI,IRBuilderBase & B)2786 Value *LibCallSimplifier::optimizeFPuts(CallInst *CI, IRBuilderBase &B) {
2787 optimizeErrorReporting(CI, B, 1);
2788
2789 // Don't rewrite fputs to fwrite when optimising for size because fwrite
2790 // requires more arguments and thus extra MOVs are required.
2791 bool OptForSize = CI->getFunction()->hasOptSize() ||
2792 llvm::shouldOptimizeForSize(CI->getParent(), PSI, BFI,
2793 PGSOQueryType::IRPass);
2794 if (OptForSize)
2795 return nullptr;
2796
2797 // We can't optimize if return value is used.
2798 if (!CI->use_empty())
2799 return nullptr;
2800
2801 // fputs(s,F) --> fwrite(s,strlen(s),1,F)
2802 uint64_t Len = GetStringLength(CI->getArgOperand(0));
2803 if (!Len)
2804 return nullptr;
2805
2806 // Known to have no uses (see above).
2807 return emitFWrite(
2808 CI->getArgOperand(0),
2809 ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len - 1),
2810 CI->getArgOperand(1), B, DL, TLI);
2811 }
2812
optimizePuts(CallInst * CI,IRBuilderBase & B)2813 Value *LibCallSimplifier::optimizePuts(CallInst *CI, IRBuilderBase &B) {
2814 annotateNonNullNoUndefBasedOnAccess(CI, 0);
2815 if (!CI->use_empty())
2816 return nullptr;
2817
2818 // Check for a constant string.
2819 // puts("") -> putchar('\n')
2820 StringRef Str;
2821 if (getConstantStringInfo(CI->getArgOperand(0), Str) && Str.empty())
2822 return emitPutChar(B.getInt32('\n'), B, TLI);
2823
2824 return nullptr;
2825 }
2826
optimizeBCopy(CallInst * CI,IRBuilderBase & B)2827 Value *LibCallSimplifier::optimizeBCopy(CallInst *CI, IRBuilderBase &B) {
2828 // bcopy(src, dst, n) -> llvm.memmove(dst, src, n)
2829 return B.CreateMemMove(CI->getArgOperand(1), Align(1), CI->getArgOperand(0),
2830 Align(1), CI->getArgOperand(2));
2831 }
2832
hasFloatVersion(StringRef FuncName)2833 bool LibCallSimplifier::hasFloatVersion(StringRef FuncName) {
2834 LibFunc Func;
2835 SmallString<20> FloatFuncName = FuncName;
2836 FloatFuncName += 'f';
2837 if (TLI->getLibFunc(FloatFuncName, Func))
2838 return TLI->has(Func);
2839 return false;
2840 }
2841
optimizeStringMemoryLibCall(CallInst * CI,IRBuilderBase & Builder)2842 Value *LibCallSimplifier::optimizeStringMemoryLibCall(CallInst *CI,
2843 IRBuilderBase &Builder) {
2844 LibFunc Func;
2845 Function *Callee = CI->getCalledFunction();
2846 // Check for string/memory library functions.
2847 if (TLI->getLibFunc(*Callee, Func) && TLI->has(Func)) {
2848 // Make sure we never change the calling convention.
2849 assert(
2850 (ignoreCallingConv(Func) ||
2851 TargetLibraryInfoImpl::isCallingConvCCompatible(CI)) &&
2852 "Optimizing string/memory libcall would change the calling convention");
2853 switch (Func) {
2854 case LibFunc_strcat:
2855 return optimizeStrCat(CI, Builder);
2856 case LibFunc_strncat:
2857 return optimizeStrNCat(CI, Builder);
2858 case LibFunc_strchr:
2859 return optimizeStrChr(CI, Builder);
2860 case LibFunc_strrchr:
2861 return optimizeStrRChr(CI, Builder);
2862 case LibFunc_strcmp:
2863 return optimizeStrCmp(CI, Builder);
2864 case LibFunc_strncmp:
2865 return optimizeStrNCmp(CI, Builder);
2866 case LibFunc_strcpy:
2867 return optimizeStrCpy(CI, Builder);
2868 case LibFunc_stpcpy:
2869 return optimizeStpCpy(CI, Builder);
2870 case LibFunc_strncpy:
2871 return optimizeStrNCpy(CI, Builder);
2872 case LibFunc_strlen:
2873 return optimizeStrLen(CI, Builder);
2874 case LibFunc_strpbrk:
2875 return optimizeStrPBrk(CI, Builder);
2876 case LibFunc_strndup:
2877 return optimizeStrNDup(CI, Builder);
2878 case LibFunc_strtol:
2879 case LibFunc_strtod:
2880 case LibFunc_strtof:
2881 case LibFunc_strtoul:
2882 case LibFunc_strtoll:
2883 case LibFunc_strtold:
2884 case LibFunc_strtoull:
2885 return optimizeStrTo(CI, Builder);
2886 case LibFunc_strspn:
2887 return optimizeStrSpn(CI, Builder);
2888 case LibFunc_strcspn:
2889 return optimizeStrCSpn(CI, Builder);
2890 case LibFunc_strstr:
2891 return optimizeStrStr(CI, Builder);
2892 case LibFunc_memchr:
2893 return optimizeMemChr(CI, Builder);
2894 case LibFunc_memrchr:
2895 return optimizeMemRChr(CI, Builder);
2896 case LibFunc_bcmp:
2897 return optimizeBCmp(CI, Builder);
2898 case LibFunc_memcmp:
2899 return optimizeMemCmp(CI, Builder);
2900 case LibFunc_memcpy:
2901 return optimizeMemCpy(CI, Builder);
2902 case LibFunc_memccpy:
2903 return optimizeMemCCpy(CI, Builder);
2904 case LibFunc_mempcpy:
2905 return optimizeMemPCpy(CI, Builder);
2906 case LibFunc_memmove:
2907 return optimizeMemMove(CI, Builder);
2908 case LibFunc_memset:
2909 return optimizeMemSet(CI, Builder);
2910 case LibFunc_realloc:
2911 return optimizeRealloc(CI, Builder);
2912 case LibFunc_wcslen:
2913 return optimizeWcslen(CI, Builder);
2914 case LibFunc_bcopy:
2915 return optimizeBCopy(CI, Builder);
2916 default:
2917 break;
2918 }
2919 }
2920 return nullptr;
2921 }
2922
optimizeFloatingPointLibCall(CallInst * CI,LibFunc Func,IRBuilderBase & Builder)2923 Value *LibCallSimplifier::optimizeFloatingPointLibCall(CallInst *CI,
2924 LibFunc Func,
2925 IRBuilderBase &Builder) {
2926 // Don't optimize calls that require strict floating point semantics.
2927 if (CI->isStrictFP())
2928 return nullptr;
2929
2930 if (Value *V = optimizeTrigReflections(CI, Func, Builder))
2931 return V;
2932
2933 switch (Func) {
2934 case LibFunc_sinpif:
2935 case LibFunc_sinpi:
2936 case LibFunc_cospif:
2937 case LibFunc_cospi:
2938 return optimizeSinCosPi(CI, Builder);
2939 case LibFunc_powf:
2940 case LibFunc_pow:
2941 case LibFunc_powl:
2942 return optimizePow(CI, Builder);
2943 case LibFunc_exp2l:
2944 case LibFunc_exp2:
2945 case LibFunc_exp2f:
2946 return optimizeExp2(CI, Builder);
2947 case LibFunc_fabsf:
2948 case LibFunc_fabs:
2949 case LibFunc_fabsl:
2950 return replaceUnaryCall(CI, Builder, Intrinsic::fabs);
2951 case LibFunc_sqrtf:
2952 case LibFunc_sqrt:
2953 case LibFunc_sqrtl:
2954 return optimizeSqrt(CI, Builder);
2955 case LibFunc_logf:
2956 case LibFunc_log:
2957 case LibFunc_logl:
2958 case LibFunc_log10f:
2959 case LibFunc_log10:
2960 case LibFunc_log10l:
2961 case LibFunc_log1pf:
2962 case LibFunc_log1p:
2963 case LibFunc_log1pl:
2964 case LibFunc_log2f:
2965 case LibFunc_log2:
2966 case LibFunc_log2l:
2967 case LibFunc_logbf:
2968 case LibFunc_logb:
2969 case LibFunc_logbl:
2970 return optimizeLog(CI, Builder);
2971 case LibFunc_tan:
2972 case LibFunc_tanf:
2973 case LibFunc_tanl:
2974 return optimizeTan(CI, Builder);
2975 case LibFunc_ceil:
2976 return replaceUnaryCall(CI, Builder, Intrinsic::ceil);
2977 case LibFunc_floor:
2978 return replaceUnaryCall(CI, Builder, Intrinsic::floor);
2979 case LibFunc_round:
2980 return replaceUnaryCall(CI, Builder, Intrinsic::round);
2981 case LibFunc_roundeven:
2982 return replaceUnaryCall(CI, Builder, Intrinsic::roundeven);
2983 case LibFunc_nearbyint:
2984 return replaceUnaryCall(CI, Builder, Intrinsic::nearbyint);
2985 case LibFunc_rint:
2986 return replaceUnaryCall(CI, Builder, Intrinsic::rint);
2987 case LibFunc_trunc:
2988 return replaceUnaryCall(CI, Builder, Intrinsic::trunc);
2989 case LibFunc_acos:
2990 case LibFunc_acosh:
2991 case LibFunc_asin:
2992 case LibFunc_asinh:
2993 case LibFunc_atan:
2994 case LibFunc_atanh:
2995 case LibFunc_cbrt:
2996 case LibFunc_cosh:
2997 case LibFunc_exp:
2998 case LibFunc_exp10:
2999 case LibFunc_expm1:
3000 case LibFunc_cos:
3001 case LibFunc_sin:
3002 case LibFunc_sinh:
3003 case LibFunc_tanh:
3004 if (UnsafeFPShrink && hasFloatVersion(CI->getCalledFunction()->getName()))
3005 return optimizeUnaryDoubleFP(CI, Builder, true);
3006 return nullptr;
3007 case LibFunc_copysign:
3008 if (hasFloatVersion(CI->getCalledFunction()->getName()))
3009 return optimizeBinaryDoubleFP(CI, Builder);
3010 return nullptr;
3011 case LibFunc_fminf:
3012 case LibFunc_fmin:
3013 case LibFunc_fminl:
3014 case LibFunc_fmaxf:
3015 case LibFunc_fmax:
3016 case LibFunc_fmaxl:
3017 return optimizeFMinFMax(CI, Builder);
3018 case LibFunc_cabs:
3019 case LibFunc_cabsf:
3020 case LibFunc_cabsl:
3021 return optimizeCAbs(CI, Builder);
3022 default:
3023 return nullptr;
3024 }
3025 }
3026
optimizeCall(CallInst * CI,IRBuilderBase & Builder)3027 Value *LibCallSimplifier::optimizeCall(CallInst *CI, IRBuilderBase &Builder) {
3028 // TODO: Split out the code below that operates on FP calls so that
3029 // we can all non-FP calls with the StrictFP attribute to be
3030 // optimized.
3031 if (CI->isNoBuiltin())
3032 return nullptr;
3033
3034 LibFunc Func;
3035 Function *Callee = CI->getCalledFunction();
3036 bool IsCallingConvC = TargetLibraryInfoImpl::isCallingConvCCompatible(CI);
3037
3038 SmallVector<OperandBundleDef, 2> OpBundles;
3039 CI->getOperandBundlesAsDefs(OpBundles);
3040
3041 IRBuilderBase::OperandBundlesGuard Guard(Builder);
3042 Builder.setDefaultOperandBundles(OpBundles);
3043
3044 // Command-line parameter overrides instruction attribute.
3045 // This can't be moved to optimizeFloatingPointLibCall() because it may be
3046 // used by the intrinsic optimizations.
3047 if (EnableUnsafeFPShrink.getNumOccurrences() > 0)
3048 UnsafeFPShrink = EnableUnsafeFPShrink;
3049 else if (isa<FPMathOperator>(CI) && CI->isFast())
3050 UnsafeFPShrink = true;
3051
3052 // First, check for intrinsics.
3053 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
3054 if (!IsCallingConvC)
3055 return nullptr;
3056 // The FP intrinsics have corresponding constrained versions so we don't
3057 // need to check for the StrictFP attribute here.
3058 switch (II->getIntrinsicID()) {
3059 case Intrinsic::pow:
3060 return optimizePow(CI, Builder);
3061 case Intrinsic::exp2:
3062 return optimizeExp2(CI, Builder);
3063 case Intrinsic::log:
3064 case Intrinsic::log2:
3065 case Intrinsic::log10:
3066 return optimizeLog(CI, Builder);
3067 case Intrinsic::sqrt:
3068 return optimizeSqrt(CI, Builder);
3069 // TODO: Use foldMallocMemset() with memset intrinsic.
3070 case Intrinsic::memset:
3071 return optimizeMemSet(CI, Builder);
3072 case Intrinsic::memcpy:
3073 return optimizeMemCpy(CI, Builder);
3074 case Intrinsic::memmove:
3075 return optimizeMemMove(CI, Builder);
3076 default:
3077 return nullptr;
3078 }
3079 }
3080
3081 // Also try to simplify calls to fortified library functions.
3082 if (Value *SimplifiedFortifiedCI =
3083 FortifiedSimplifier.optimizeCall(CI, Builder)) {
3084 // Try to further simplify the result.
3085 CallInst *SimplifiedCI = dyn_cast<CallInst>(SimplifiedFortifiedCI);
3086 if (SimplifiedCI && SimplifiedCI->getCalledFunction()) {
3087 // Ensure that SimplifiedCI's uses are complete, since some calls have
3088 // their uses analyzed.
3089 replaceAllUsesWith(CI, SimplifiedCI);
3090
3091 // Set insertion point to SimplifiedCI to guarantee we reach all uses
3092 // we might replace later on.
3093 IRBuilderBase::InsertPointGuard Guard(Builder);
3094 Builder.SetInsertPoint(SimplifiedCI);
3095 if (Value *V = optimizeStringMemoryLibCall(SimplifiedCI, Builder)) {
3096 // If we were able to further simplify, remove the now redundant call.
3097 substituteInParent(SimplifiedCI, V);
3098 return V;
3099 }
3100 }
3101 return SimplifiedFortifiedCI;
3102 }
3103
3104 // Then check for known library functions.
3105 if (TLI->getLibFunc(*Callee, Func) && TLI->has(Func)) {
3106 // We never change the calling convention.
3107 if (!ignoreCallingConv(Func) && !IsCallingConvC)
3108 return nullptr;
3109 if (Value *V = optimizeStringMemoryLibCall(CI, Builder))
3110 return V;
3111 if (Value *V = optimizeFloatingPointLibCall(CI, Func, Builder))
3112 return V;
3113 switch (Func) {
3114 case LibFunc_ffs:
3115 case LibFunc_ffsl:
3116 case LibFunc_ffsll:
3117 return optimizeFFS(CI, Builder);
3118 case LibFunc_fls:
3119 case LibFunc_flsl:
3120 case LibFunc_flsll:
3121 return optimizeFls(CI, Builder);
3122 case LibFunc_abs:
3123 case LibFunc_labs:
3124 case LibFunc_llabs:
3125 return optimizeAbs(CI, Builder);
3126 case LibFunc_isdigit:
3127 return optimizeIsDigit(CI, Builder);
3128 case LibFunc_isascii:
3129 return optimizeIsAscii(CI, Builder);
3130 case LibFunc_toascii:
3131 return optimizeToAscii(CI, Builder);
3132 case LibFunc_atoi:
3133 case LibFunc_atol:
3134 case LibFunc_atoll:
3135 return optimizeAtoi(CI, Builder);
3136 case LibFunc_strtol:
3137 case LibFunc_strtoll:
3138 return optimizeStrtol(CI, Builder);
3139 case LibFunc_printf:
3140 return optimizePrintF(CI, Builder);
3141 case LibFunc_sprintf:
3142 return optimizeSPrintF(CI, Builder);
3143 case LibFunc_snprintf:
3144 return optimizeSnPrintF(CI, Builder);
3145 case LibFunc_fprintf:
3146 return optimizeFPrintF(CI, Builder);
3147 case LibFunc_fwrite:
3148 return optimizeFWrite(CI, Builder);
3149 case LibFunc_fputs:
3150 return optimizeFPuts(CI, Builder);
3151 case LibFunc_puts:
3152 return optimizePuts(CI, Builder);
3153 case LibFunc_perror:
3154 return optimizeErrorReporting(CI, Builder);
3155 case LibFunc_vfprintf:
3156 case LibFunc_fiprintf:
3157 return optimizeErrorReporting(CI, Builder, 0);
3158 default:
3159 return nullptr;
3160 }
3161 }
3162 return nullptr;
3163 }
3164
LibCallSimplifier(const DataLayout & DL,const TargetLibraryInfo * TLI,OptimizationRemarkEmitter & ORE,BlockFrequencyInfo * BFI,ProfileSummaryInfo * PSI,function_ref<void (Instruction *,Value *)> Replacer,function_ref<void (Instruction *)> Eraser)3165 LibCallSimplifier::LibCallSimplifier(
3166 const DataLayout &DL, const TargetLibraryInfo *TLI,
3167 OptimizationRemarkEmitter &ORE,
3168 BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI,
3169 function_ref<void(Instruction *, Value *)> Replacer,
3170 function_ref<void(Instruction *)> Eraser)
3171 : FortifiedSimplifier(TLI), DL(DL), TLI(TLI), ORE(ORE), BFI(BFI), PSI(PSI),
3172 UnsafeFPShrink(false), Replacer(Replacer), Eraser(Eraser) {}
3173
replaceAllUsesWith(Instruction * I,Value * With)3174 void LibCallSimplifier::replaceAllUsesWith(Instruction *I, Value *With) {
3175 // Indirect through the replacer used in this instance.
3176 Replacer(I, With);
3177 }
3178
eraseFromParent(Instruction * I)3179 void LibCallSimplifier::eraseFromParent(Instruction *I) {
3180 Eraser(I);
3181 }
3182
3183 // TODO:
3184 // Additional cases that we need to add to this file:
3185 //
3186 // cbrt:
3187 // * cbrt(expN(X)) -> expN(x/3)
3188 // * cbrt(sqrt(x)) -> pow(x,1/6)
3189 // * cbrt(cbrt(x)) -> pow(x,1/9)
3190 //
3191 // exp, expf, expl:
3192 // * exp(log(x)) -> x
3193 //
3194 // log, logf, logl:
3195 // * log(exp(x)) -> x
3196 // * log(exp(y)) -> y*log(e)
3197 // * log(exp10(y)) -> y*log(10)
3198 // * log(sqrt(x)) -> 0.5*log(x)
3199 //
3200 // pow, powf, powl:
3201 // * pow(sqrt(x),y) -> pow(x,y*0.5)
3202 // * pow(pow(x,y),z)-> pow(x,y*z)
3203 //
3204 // signbit:
3205 // * signbit(cnst) -> cnst'
3206 // * signbit(nncst) -> 0 (if pstv is a non-negative constant)
3207 //
3208 // sqrt, sqrtf, sqrtl:
3209 // * sqrt(expN(x)) -> expN(x*0.5)
3210 // * sqrt(Nroot(x)) -> pow(x,1/(2*N))
3211 // * sqrt(pow(x,y)) -> pow(|x|,y*0.5)
3212 //
3213
3214 //===----------------------------------------------------------------------===//
3215 // Fortified Library Call Optimizations
3216 //===----------------------------------------------------------------------===//
3217
3218 bool
isFortifiedCallFoldable(CallInst * CI,unsigned ObjSizeOp,Optional<unsigned> SizeOp,Optional<unsigned> StrOp,Optional<unsigned> FlagOp)3219 FortifiedLibCallSimplifier::isFortifiedCallFoldable(CallInst *CI,
3220 unsigned ObjSizeOp,
3221 Optional<unsigned> SizeOp,
3222 Optional<unsigned> StrOp,
3223 Optional<unsigned> FlagOp) {
3224 // If this function takes a flag argument, the implementation may use it to
3225 // perform extra checks. Don't fold into the non-checking variant.
3226 if (FlagOp) {
3227 ConstantInt *Flag = dyn_cast<ConstantInt>(CI->getArgOperand(*FlagOp));
3228 if (!Flag || !Flag->isZero())
3229 return false;
3230 }
3231
3232 if (SizeOp && CI->getArgOperand(ObjSizeOp) == CI->getArgOperand(*SizeOp))
3233 return true;
3234
3235 if (ConstantInt *ObjSizeCI =
3236 dyn_cast<ConstantInt>(CI->getArgOperand(ObjSizeOp))) {
3237 if (ObjSizeCI->isMinusOne())
3238 return true;
3239 // If the object size wasn't -1 (unknown), bail out if we were asked to.
3240 if (OnlyLowerUnknownSize)
3241 return false;
3242 if (StrOp) {
3243 uint64_t Len = GetStringLength(CI->getArgOperand(*StrOp));
3244 // If the length is 0 we don't know how long it is and so we can't
3245 // remove the check.
3246 if (Len)
3247 annotateDereferenceableBytes(CI, *StrOp, Len);
3248 else
3249 return false;
3250 return ObjSizeCI->getZExtValue() >= Len;
3251 }
3252
3253 if (SizeOp) {
3254 if (ConstantInt *SizeCI =
3255 dyn_cast<ConstantInt>(CI->getArgOperand(*SizeOp)))
3256 return ObjSizeCI->getZExtValue() >= SizeCI->getZExtValue();
3257 }
3258 }
3259 return false;
3260 }
3261
optimizeMemCpyChk(CallInst * CI,IRBuilderBase & B)3262 Value *FortifiedLibCallSimplifier::optimizeMemCpyChk(CallInst *CI,
3263 IRBuilderBase &B) {
3264 if (isFortifiedCallFoldable(CI, 3, 2)) {
3265 CallInst *NewCI =
3266 B.CreateMemCpy(CI->getArgOperand(0), Align(1), CI->getArgOperand(1),
3267 Align(1), CI->getArgOperand(2));
3268 NewCI->setAttributes(CI->getAttributes());
3269 NewCI->removeAttributes(AttributeList::ReturnIndex,
3270 AttributeFuncs::typeIncompatible(NewCI->getType()));
3271 return CI->getArgOperand(0);
3272 }
3273 return nullptr;
3274 }
3275
optimizeMemMoveChk(CallInst * CI,IRBuilderBase & B)3276 Value *FortifiedLibCallSimplifier::optimizeMemMoveChk(CallInst *CI,
3277 IRBuilderBase &B) {
3278 if (isFortifiedCallFoldable(CI, 3, 2)) {
3279 CallInst *NewCI =
3280 B.CreateMemMove(CI->getArgOperand(0), Align(1), CI->getArgOperand(1),
3281 Align(1), CI->getArgOperand(2));
3282 NewCI->setAttributes(CI->getAttributes());
3283 NewCI->removeAttributes(AttributeList::ReturnIndex,
3284 AttributeFuncs::typeIncompatible(NewCI->getType()));
3285 return CI->getArgOperand(0);
3286 }
3287 return nullptr;
3288 }
3289
optimizeMemSetChk(CallInst * CI,IRBuilderBase & B)3290 Value *FortifiedLibCallSimplifier::optimizeMemSetChk(CallInst *CI,
3291 IRBuilderBase &B) {
3292 // TODO: Try foldMallocMemset() here.
3293
3294 if (isFortifiedCallFoldable(CI, 3, 2)) {
3295 Value *Val = B.CreateIntCast(CI->getArgOperand(1), B.getInt8Ty(), false);
3296 CallInst *NewCI = B.CreateMemSet(CI->getArgOperand(0), Val,
3297 CI->getArgOperand(2), Align(1));
3298 NewCI->setAttributes(CI->getAttributes());
3299 NewCI->removeAttributes(AttributeList::ReturnIndex,
3300 AttributeFuncs::typeIncompatible(NewCI->getType()));
3301 return CI->getArgOperand(0);
3302 }
3303 return nullptr;
3304 }
3305
optimizeMemPCpyChk(CallInst * CI,IRBuilderBase & B)3306 Value *FortifiedLibCallSimplifier::optimizeMemPCpyChk(CallInst *CI,
3307 IRBuilderBase &B) {
3308 const DataLayout &DL = CI->getModule()->getDataLayout();
3309 if (isFortifiedCallFoldable(CI, 3, 2))
3310 if (Value *Call = emitMemPCpy(CI->getArgOperand(0), CI->getArgOperand(1),
3311 CI->getArgOperand(2), B, DL, TLI)) {
3312 CallInst *NewCI = cast<CallInst>(Call);
3313 NewCI->setAttributes(CI->getAttributes());
3314 NewCI->removeAttributes(
3315 AttributeList::ReturnIndex,
3316 AttributeFuncs::typeIncompatible(NewCI->getType()));
3317 return NewCI;
3318 }
3319 return nullptr;
3320 }
3321
optimizeStrpCpyChk(CallInst * CI,IRBuilderBase & B,LibFunc Func)3322 Value *FortifiedLibCallSimplifier::optimizeStrpCpyChk(CallInst *CI,
3323 IRBuilderBase &B,
3324 LibFunc Func) {
3325 const DataLayout &DL = CI->getModule()->getDataLayout();
3326 Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1),
3327 *ObjSize = CI->getArgOperand(2);
3328
3329 // __stpcpy_chk(x,x,...) -> x+strlen(x)
3330 if (Func == LibFunc_stpcpy_chk && !OnlyLowerUnknownSize && Dst == Src) {
3331 Value *StrLen = emitStrLen(Src, B, DL, TLI);
3332 return StrLen ? B.CreateInBoundsGEP(B.getInt8Ty(), Dst, StrLen) : nullptr;
3333 }
3334
3335 // If a) we don't have any length information, or b) we know this will
3336 // fit then just lower to a plain st[rp]cpy. Otherwise we'll keep our
3337 // st[rp]cpy_chk call which may fail at runtime if the size is too long.
3338 // TODO: It might be nice to get a maximum length out of the possible
3339 // string lengths for varying.
3340 if (isFortifiedCallFoldable(CI, 2, None, 1)) {
3341 if (Func == LibFunc_strcpy_chk)
3342 return emitStrCpy(Dst, Src, B, TLI);
3343 else
3344 return emitStpCpy(Dst, Src, B, TLI);
3345 }
3346
3347 if (OnlyLowerUnknownSize)
3348 return nullptr;
3349
3350 // Maybe we can stil fold __st[rp]cpy_chk to __memcpy_chk.
3351 uint64_t Len = GetStringLength(Src);
3352 if (Len)
3353 annotateDereferenceableBytes(CI, 1, Len);
3354 else
3355 return nullptr;
3356
3357 Type *SizeTTy = DL.getIntPtrType(CI->getContext());
3358 Value *LenV = ConstantInt::get(SizeTTy, Len);
3359 Value *Ret = emitMemCpyChk(Dst, Src, LenV, ObjSize, B, DL, TLI);
3360 // If the function was an __stpcpy_chk, and we were able to fold it into
3361 // a __memcpy_chk, we still need to return the correct end pointer.
3362 if (Ret && Func == LibFunc_stpcpy_chk)
3363 return B.CreateGEP(B.getInt8Ty(), Dst, ConstantInt::get(SizeTTy, Len - 1));
3364 return Ret;
3365 }
3366
optimizeStrLenChk(CallInst * CI,IRBuilderBase & B)3367 Value *FortifiedLibCallSimplifier::optimizeStrLenChk(CallInst *CI,
3368 IRBuilderBase &B) {
3369 if (isFortifiedCallFoldable(CI, 1, None, 0))
3370 return emitStrLen(CI->getArgOperand(0), B, CI->getModule()->getDataLayout(),
3371 TLI);
3372 return nullptr;
3373 }
3374
optimizeStrpNCpyChk(CallInst * CI,IRBuilderBase & B,LibFunc Func)3375 Value *FortifiedLibCallSimplifier::optimizeStrpNCpyChk(CallInst *CI,
3376 IRBuilderBase &B,
3377 LibFunc Func) {
3378 if (isFortifiedCallFoldable(CI, 3, 2)) {
3379 if (Func == LibFunc_strncpy_chk)
3380 return emitStrNCpy(CI->getArgOperand(0), CI->getArgOperand(1),
3381 CI->getArgOperand(2), B, TLI);
3382 else
3383 return emitStpNCpy(CI->getArgOperand(0), CI->getArgOperand(1),
3384 CI->getArgOperand(2), B, TLI);
3385 }
3386
3387 return nullptr;
3388 }
3389
optimizeMemCCpyChk(CallInst * CI,IRBuilderBase & B)3390 Value *FortifiedLibCallSimplifier::optimizeMemCCpyChk(CallInst *CI,
3391 IRBuilderBase &B) {
3392 if (isFortifiedCallFoldable(CI, 4, 3))
3393 return emitMemCCpy(CI->getArgOperand(0), CI->getArgOperand(1),
3394 CI->getArgOperand(2), CI->getArgOperand(3), B, TLI);
3395
3396 return nullptr;
3397 }
3398
optimizeSNPrintfChk(CallInst * CI,IRBuilderBase & B)3399 Value *FortifiedLibCallSimplifier::optimizeSNPrintfChk(CallInst *CI,
3400 IRBuilderBase &B) {
3401 if (isFortifiedCallFoldable(CI, 3, 1, None, 2)) {
3402 SmallVector<Value *, 8> VariadicArgs(drop_begin(CI->args(), 5));
3403 return emitSNPrintf(CI->getArgOperand(0), CI->getArgOperand(1),
3404 CI->getArgOperand(4), VariadicArgs, B, TLI);
3405 }
3406
3407 return nullptr;
3408 }
3409
optimizeSPrintfChk(CallInst * CI,IRBuilderBase & B)3410 Value *FortifiedLibCallSimplifier::optimizeSPrintfChk(CallInst *CI,
3411 IRBuilderBase &B) {
3412 if (isFortifiedCallFoldable(CI, 2, None, None, 1)) {
3413 SmallVector<Value *, 8> VariadicArgs(drop_begin(CI->args(), 4));
3414 return emitSPrintf(CI->getArgOperand(0), CI->getArgOperand(3), VariadicArgs,
3415 B, TLI);
3416 }
3417
3418 return nullptr;
3419 }
3420
optimizeStrCatChk(CallInst * CI,IRBuilderBase & B)3421 Value *FortifiedLibCallSimplifier::optimizeStrCatChk(CallInst *CI,
3422 IRBuilderBase &B) {
3423 if (isFortifiedCallFoldable(CI, 2))
3424 return emitStrCat(CI->getArgOperand(0), CI->getArgOperand(1), B, TLI);
3425
3426 return nullptr;
3427 }
3428
optimizeStrLCat(CallInst * CI,IRBuilderBase & B)3429 Value *FortifiedLibCallSimplifier::optimizeStrLCat(CallInst *CI,
3430 IRBuilderBase &B) {
3431 if (isFortifiedCallFoldable(CI, 3))
3432 return emitStrLCat(CI->getArgOperand(0), CI->getArgOperand(1),
3433 CI->getArgOperand(2), B, TLI);
3434
3435 return nullptr;
3436 }
3437
optimizeStrNCatChk(CallInst * CI,IRBuilderBase & B)3438 Value *FortifiedLibCallSimplifier::optimizeStrNCatChk(CallInst *CI,
3439 IRBuilderBase &B) {
3440 if (isFortifiedCallFoldable(CI, 3))
3441 return emitStrNCat(CI->getArgOperand(0), CI->getArgOperand(1),
3442 CI->getArgOperand(2), B, TLI);
3443
3444 return nullptr;
3445 }
3446
optimizeStrLCpyChk(CallInst * CI,IRBuilderBase & B)3447 Value *FortifiedLibCallSimplifier::optimizeStrLCpyChk(CallInst *CI,
3448 IRBuilderBase &B) {
3449 if (isFortifiedCallFoldable(CI, 3))
3450 return emitStrLCpy(CI->getArgOperand(0), CI->getArgOperand(1),
3451 CI->getArgOperand(2), B, TLI);
3452
3453 return nullptr;
3454 }
3455
optimizeVSNPrintfChk(CallInst * CI,IRBuilderBase & B)3456 Value *FortifiedLibCallSimplifier::optimizeVSNPrintfChk(CallInst *CI,
3457 IRBuilderBase &B) {
3458 if (isFortifiedCallFoldable(CI, 3, 1, None, 2))
3459 return emitVSNPrintf(CI->getArgOperand(0), CI->getArgOperand(1),
3460 CI->getArgOperand(4), CI->getArgOperand(5), B, TLI);
3461
3462 return nullptr;
3463 }
3464
optimizeVSPrintfChk(CallInst * CI,IRBuilderBase & B)3465 Value *FortifiedLibCallSimplifier::optimizeVSPrintfChk(CallInst *CI,
3466 IRBuilderBase &B) {
3467 if (isFortifiedCallFoldable(CI, 2, None, None, 1))
3468 return emitVSPrintf(CI->getArgOperand(0), CI->getArgOperand(3),
3469 CI->getArgOperand(4), B, TLI);
3470
3471 return nullptr;
3472 }
3473
optimizeCall(CallInst * CI,IRBuilderBase & Builder)3474 Value *FortifiedLibCallSimplifier::optimizeCall(CallInst *CI,
3475 IRBuilderBase &Builder) {
3476 // FIXME: We shouldn't be changing "nobuiltin" or TLI unavailable calls here.
3477 // Some clang users checked for _chk libcall availability using:
3478 // __has_builtin(__builtin___memcpy_chk)
3479 // When compiling with -fno-builtin, this is always true.
3480 // When passing -ffreestanding/-mkernel, which both imply -fno-builtin, we
3481 // end up with fortified libcalls, which isn't acceptable in a freestanding
3482 // environment which only provides their non-fortified counterparts.
3483 //
3484 // Until we change clang and/or teach external users to check for availability
3485 // differently, disregard the "nobuiltin" attribute and TLI::has.
3486 //
3487 // PR23093.
3488
3489 LibFunc Func;
3490 Function *Callee = CI->getCalledFunction();
3491 bool IsCallingConvC = TargetLibraryInfoImpl::isCallingConvCCompatible(CI);
3492
3493 SmallVector<OperandBundleDef, 2> OpBundles;
3494 CI->getOperandBundlesAsDefs(OpBundles);
3495
3496 IRBuilderBase::OperandBundlesGuard Guard(Builder);
3497 Builder.setDefaultOperandBundles(OpBundles);
3498
3499 // First, check that this is a known library functions and that the prototype
3500 // is correct.
3501 if (!TLI->getLibFunc(*Callee, Func))
3502 return nullptr;
3503
3504 // We never change the calling convention.
3505 if (!ignoreCallingConv(Func) && !IsCallingConvC)
3506 return nullptr;
3507
3508 switch (Func) {
3509 case LibFunc_memcpy_chk:
3510 return optimizeMemCpyChk(CI, Builder);
3511 case LibFunc_mempcpy_chk:
3512 return optimizeMemPCpyChk(CI, Builder);
3513 case LibFunc_memmove_chk:
3514 return optimizeMemMoveChk(CI, Builder);
3515 case LibFunc_memset_chk:
3516 return optimizeMemSetChk(CI, Builder);
3517 case LibFunc_stpcpy_chk:
3518 case LibFunc_strcpy_chk:
3519 return optimizeStrpCpyChk(CI, Builder, Func);
3520 case LibFunc_strlen_chk:
3521 return optimizeStrLenChk(CI, Builder);
3522 case LibFunc_stpncpy_chk:
3523 case LibFunc_strncpy_chk:
3524 return optimizeStrpNCpyChk(CI, Builder, Func);
3525 case LibFunc_memccpy_chk:
3526 return optimizeMemCCpyChk(CI, Builder);
3527 case LibFunc_snprintf_chk:
3528 return optimizeSNPrintfChk(CI, Builder);
3529 case LibFunc_sprintf_chk:
3530 return optimizeSPrintfChk(CI, Builder);
3531 case LibFunc_strcat_chk:
3532 return optimizeStrCatChk(CI, Builder);
3533 case LibFunc_strlcat_chk:
3534 return optimizeStrLCat(CI, Builder);
3535 case LibFunc_strncat_chk:
3536 return optimizeStrNCatChk(CI, Builder);
3537 case LibFunc_strlcpy_chk:
3538 return optimizeStrLCpyChk(CI, Builder);
3539 case LibFunc_vsnprintf_chk:
3540 return optimizeVSNPrintfChk(CI, Builder);
3541 case LibFunc_vsprintf_chk:
3542 return optimizeVSPrintfChk(CI, Builder);
3543 default:
3544 break;
3545 }
3546 return nullptr;
3547 }
3548
FortifiedLibCallSimplifier(const TargetLibraryInfo * TLI,bool OnlyLowerUnknownSize)3549 FortifiedLibCallSimplifier::FortifiedLibCallSimplifier(
3550 const TargetLibraryInfo *TLI, bool OnlyLowerUnknownSize)
3551 : TLI(TLI), OnlyLowerUnknownSize(OnlyLowerUnknownSize) {}
3552