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