LLVM  9.0.0svn
SimplifyLibCalls.cpp
Go to the documentation of this file.
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 
15 #include "llvm/ADT/APSInt.h"
16 #include "llvm/ADT/SmallString.h"
17 #include "llvm/ADT/StringMap.h"
18 #include "llvm/ADT/Triple.h"
25 #include "llvm/Analysis/Loads.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/Function.h"
28 #include "llvm/IR/IRBuilder.h"
29 #include "llvm/IR/IntrinsicInst.h"
30 #include "llvm/IR/Intrinsics.h"
31 #include "llvm/IR/LLVMContext.h"
32 #include "llvm/IR/Module.h"
33 #include "llvm/IR/PatternMatch.h"
35 #include "llvm/Support/KnownBits.h"
37 
38 using namespace llvm;
39 using namespace PatternMatch;
40 
41 static cl::opt<bool>
42  EnableUnsafeFPShrink("enable-double-float-shrink", cl::Hidden,
43  cl::init(false),
44  cl::desc("Enable unsafe double to float "
45  "shrinking for math lib calls"));
46 
47 
48 //===----------------------------------------------------------------------===//
49 // Helper Functions
50 //===----------------------------------------------------------------------===//
51 
52 static bool ignoreCallingConv(LibFunc Func) {
53  return Func == LibFunc_abs || Func == LibFunc_labs ||
54  Func == LibFunc_llabs || Func == LibFunc_strlen;
55 }
56 
58  switch(CI->getCallingConv()) {
59  default:
60  return false;
62  return true;
66 
67  // The iOS ABI diverges from the standard in some cases, so for now don't
68  // try to simplify those calls.
69  if (Triple(CI->getModule()->getTargetTriple()).isiOS())
70  return false;
71 
72  auto *FuncTy = CI->getFunctionType();
73 
74  if (!FuncTy->getReturnType()->isPointerTy() &&
75  !FuncTy->getReturnType()->isIntegerTy() &&
76  !FuncTy->getReturnType()->isVoidTy())
77  return false;
78 
79  for (auto Param : FuncTy->params()) {
80  if (!Param->isPointerTy() && !Param->isIntegerTy())
81  return false;
82  }
83  return true;
84  }
85  }
86  return false;
87 }
88 
89 /// Return true if it is only used in equality comparisons with With.
90 static bool isOnlyUsedInEqualityComparison(Value *V, Value *With) {
91  for (User *U : V->users()) {
92  if (ICmpInst *IC = dyn_cast<ICmpInst>(U))
93  if (IC->isEquality() && IC->getOperand(1) == With)
94  continue;
95  // Unknown instruction.
96  return false;
97  }
98  return true;
99 }
100 
101 static bool callHasFloatingPointArgument(const CallInst *CI) {
102  return any_of(CI->operands(), [](const Use &OI) {
103  return OI->getType()->isFloatingPointTy();
104  });
105 }
106 
107 static Value *convertStrToNumber(CallInst *CI, StringRef &Str, int64_t Base) {
108  if (Base < 2 || Base > 36)
109  // handle special zero base
110  if (Base != 0)
111  return nullptr;
112 
113  char *End;
114  std::string nptr = Str.str();
115  errno = 0;
116  long long int Result = strtoll(nptr.c_str(), &End, Base);
117  if (errno)
118  return nullptr;
119 
120  // if we assume all possible target locales are ASCII supersets,
121  // then if strtoll successfully parses a number on the host,
122  // it will also successfully parse the same way on the target
123  if (*End != '\0')
124  return nullptr;
125 
126  if (!isIntN(CI->getType()->getPrimitiveSizeInBits(), Result))
127  return nullptr;
128 
129  return ConstantInt::get(CI->getType(), Result);
130 }
131 
133  const TargetLibraryInfo *TLI) {
134  CallInst *FOpen = dyn_cast<CallInst>(File);
135  if (!FOpen)
136  return false;
137 
138  Function *InnerCallee = FOpen->getCalledFunction();
139  if (!InnerCallee)
140  return false;
141 
142  LibFunc Func;
143  if (!TLI->getLibFunc(*InnerCallee, Func) || !TLI->has(Func) ||
144  Func != LibFunc_fopen)
145  return false;
146 
148  if (PointerMayBeCaptured(File, true, true))
149  return false;
150 
151  return true;
152 }
153 
155  for (User *U : V->users()) {
156  if (ICmpInst *IC = dyn_cast<ICmpInst>(U))
157  if (Constant *C = dyn_cast<Constant>(IC->getOperand(1)))
158  if (C->isNullValue())
159  continue;
160  // Unknown instruction.
161  return false;
162  }
163  return true;
164 }
165 
166 static bool canTransformToMemCmp(CallInst *CI, Value *Str, uint64_t Len,
167  const DataLayout &DL) {
169  return false;
170 
171  if (!isDereferenceableAndAlignedPointer(Str, 1, APInt(64, Len), DL))
172  return false;
173 
174  if (CI->getFunction()->hasFnAttribute(Attribute::SanitizeMemory))
175  return false;
176 
177  return true;
178 }
179 
180 //===----------------------------------------------------------------------===//
181 // String and Memory Library Call Optimizations
182 //===----------------------------------------------------------------------===//
183 
184 Value *LibCallSimplifier::optimizeStrCat(CallInst *CI, IRBuilder<> &B) {
185  // Extract some information from the instruction
186  Value *Dst = CI->getArgOperand(0);
187  Value *Src = CI->getArgOperand(1);
188 
189  // See if we can get the length of the input string.
190  uint64_t Len = GetStringLength(Src);
191  if (Len == 0)
192  return nullptr;
193  --Len; // Unbias length.
194 
195  // Handle the simple, do-nothing case: strcat(x, "") -> x
196  if (Len == 0)
197  return Dst;
198 
199  return emitStrLenMemCpy(Src, Dst, Len, B);
200 }
201 
202 Value *LibCallSimplifier::emitStrLenMemCpy(Value *Src, Value *Dst, uint64_t Len,
203  IRBuilder<> &B) {
204  // We need to find the end of the destination string. That's where the
205  // memory is to be moved to. We just generate a call to strlen.
206  Value *DstLen = emitStrLen(Dst, B, DL, TLI);
207  if (!DstLen)
208  return nullptr;
209 
210  // Now that we have the destination's length, we must index into the
211  // destination's pointer to get the actual memcpy destination (end of
212  // the string .. we're concatenating).
213  Value *CpyDst = B.CreateGEP(B.getInt8Ty(), Dst, DstLen, "endptr");
214 
215  // We have enough information to now generate the memcpy call to do the
216  // concatenation for us. Make a memcpy to copy the nul byte with align = 1.
217  B.CreateMemCpy(CpyDst, 1, Src, 1,
218  ConstantInt::get(DL.getIntPtrType(Src->getContext()), Len + 1));
219  return Dst;
220 }
221 
222 Value *LibCallSimplifier::optimizeStrNCat(CallInst *CI, IRBuilder<> &B) {
223  // Extract some information from the instruction.
224  Value *Dst = CI->getArgOperand(0);
225  Value *Src = CI->getArgOperand(1);
226  uint64_t Len;
227 
228  // We don't do anything if length is not constant.
229  if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(CI->getArgOperand(2)))
230  Len = LengthArg->getZExtValue();
231  else
232  return nullptr;
233 
234  // See if we can get the length of the input string.
235  uint64_t SrcLen = GetStringLength(Src);
236  if (SrcLen == 0)
237  return nullptr;
238  --SrcLen; // Unbias length.
239 
240  // Handle the simple, do-nothing cases:
241  // strncat(x, "", c) -> x
242  // strncat(x, c, 0) -> x
243  if (SrcLen == 0 || Len == 0)
244  return Dst;
245 
246  // We don't optimize this case.
247  if (Len < SrcLen)
248  return nullptr;
249 
250  // strncat(x, s, c) -> strcat(x, s)
251  // s is constant so the strcat can be optimized further.
252  return emitStrLenMemCpy(Src, Dst, SrcLen, B);
253 }
254 
255 Value *LibCallSimplifier::optimizeStrChr(CallInst *CI, IRBuilder<> &B) {
257  FunctionType *FT = Callee->getFunctionType();
258  Value *SrcStr = CI->getArgOperand(0);
259 
260  // If the second operand is non-constant, see if we can compute the length
261  // of the input string and turn this into memchr.
262  ConstantInt *CharC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
263  if (!CharC) {
264  uint64_t Len = GetStringLength(SrcStr);
265  if (Len == 0 || !FT->getParamType(1)->isIntegerTy(32)) // memchr needs i32.
266  return nullptr;
267 
268  return emitMemChr(SrcStr, CI->getArgOperand(1), // include nul.
269  ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len),
270  B, DL, TLI);
271  }
272 
273  // Otherwise, the character is a constant, see if the first argument is
274  // a string literal. If so, we can constant fold.
275  StringRef Str;
276  if (!getConstantStringInfo(SrcStr, Str)) {
277  if (CharC->isZero()) // strchr(p, 0) -> p + strlen(p)
278  return B.CreateGEP(B.getInt8Ty(), SrcStr, emitStrLen(SrcStr, B, DL, TLI),
279  "strchr");
280  return nullptr;
281  }
282 
283  // Compute the offset, make sure to handle the case when we're searching for
284  // zero (a weird way to spell strlen).
285  size_t I = (0xFF & CharC->getSExtValue()) == 0
286  ? Str.size()
287  : Str.find(CharC->getSExtValue());
288  if (I == StringRef::npos) // Didn't find the char. strchr returns null.
289  return Constant::getNullValue(CI->getType());
290 
291  // strchr(s+n,c) -> gep(s+n+i,c)
292  return B.CreateGEP(B.getInt8Ty(), SrcStr, B.getInt64(I), "strchr");
293 }
294 
295 Value *LibCallSimplifier::optimizeStrRChr(CallInst *CI, IRBuilder<> &B) {
296  Value *SrcStr = CI->getArgOperand(0);
297  ConstantInt *CharC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
298 
299  // Cannot fold anything if we're not looking for a constant.
300  if (!CharC)
301  return nullptr;
302 
303  StringRef Str;
304  if (!getConstantStringInfo(SrcStr, Str)) {
305  // strrchr(s, 0) -> strchr(s, 0)
306  if (CharC->isZero())
307  return emitStrChr(SrcStr, '\0', B, TLI);
308  return nullptr;
309  }
310 
311  // Compute the offset.
312  size_t I = (0xFF & CharC->getSExtValue()) == 0
313  ? Str.size()
314  : Str.rfind(CharC->getSExtValue());
315  if (I == StringRef::npos) // Didn't find the char. Return null.
316  return Constant::getNullValue(CI->getType());
317 
318  // strrchr(s+n,c) -> gep(s+n+i,c)
319  return B.CreateGEP(B.getInt8Ty(), SrcStr, B.getInt64(I), "strrchr");
320 }
321 
322 Value *LibCallSimplifier::optimizeStrCmp(CallInst *CI, IRBuilder<> &B) {
323  Value *Str1P = CI->getArgOperand(0), *Str2P = CI->getArgOperand(1);
324  if (Str1P == Str2P) // strcmp(x,x) -> 0
325  return ConstantInt::get(CI->getType(), 0);
326 
327  StringRef Str1, Str2;
328  bool HasStr1 = getConstantStringInfo(Str1P, Str1);
329  bool HasStr2 = getConstantStringInfo(Str2P, Str2);
330 
331  // strcmp(x, y) -> cnst (if both x and y are constant strings)
332  if (HasStr1 && HasStr2)
333  return ConstantInt::get(CI->getType(), Str1.compare(Str2));
334 
335  if (HasStr1 && Str1.empty()) // strcmp("", x) -> -*x
336  return B.CreateNeg(B.CreateZExt(
337  B.CreateLoad(B.getInt8Ty(), Str2P, "strcmpload"), CI->getType()));
338 
339  if (HasStr2 && Str2.empty()) // strcmp(x,"") -> *x
340  return B.CreateZExt(B.CreateLoad(B.getInt8Ty(), Str1P, "strcmpload"),
341  CI->getType());
342 
343  // strcmp(P, "x") -> memcmp(P, "x", 2)
344  uint64_t Len1 = GetStringLength(Str1P);
345  uint64_t Len2 = GetStringLength(Str2P);
346  if (Len1 && Len2) {
347  return emitMemCmp(Str1P, Str2P,
348  ConstantInt::get(DL.getIntPtrType(CI->getContext()),
349  std::min(Len1, Len2)),
350  B, DL, TLI);
351  }
352 
353  // strcmp to memcmp
354  if (!HasStr1 && HasStr2) {
355  if (canTransformToMemCmp(CI, Str1P, Len2, DL))
356  return emitMemCmp(
357  Str1P, Str2P,
358  ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len2), B, DL,
359  TLI);
360  } else if (HasStr1 && !HasStr2) {
361  if (canTransformToMemCmp(CI, Str2P, Len1, DL))
362  return emitMemCmp(
363  Str1P, Str2P,
364  ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len1), B, DL,
365  TLI);
366  }
367 
368  return nullptr;
369 }
370 
371 Value *LibCallSimplifier::optimizeStrNCmp(CallInst *CI, IRBuilder<> &B) {
372  Value *Str1P = CI->getArgOperand(0), *Str2P = CI->getArgOperand(1);
373  if (Str1P == Str2P) // strncmp(x,x,n) -> 0
374  return ConstantInt::get(CI->getType(), 0);
375 
376  // Get the length argument if it is constant.
377  uint64_t Length;
378  if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(CI->getArgOperand(2)))
379  Length = LengthArg->getZExtValue();
380  else
381  return nullptr;
382 
383  if (Length == 0) // strncmp(x,y,0) -> 0
384  return ConstantInt::get(CI->getType(), 0);
385 
386  if (Length == 1) // strncmp(x,y,1) -> memcmp(x,y,1)
387  return emitMemCmp(Str1P, Str2P, CI->getArgOperand(2), B, DL, TLI);
388 
389  StringRef Str1, Str2;
390  bool HasStr1 = getConstantStringInfo(Str1P, Str1);
391  bool HasStr2 = getConstantStringInfo(Str2P, Str2);
392 
393  // strncmp(x, y) -> cnst (if both x and y are constant strings)
394  if (HasStr1 && HasStr2) {
395  StringRef SubStr1 = Str1.substr(0, Length);
396  StringRef SubStr2 = Str2.substr(0, Length);
397  return ConstantInt::get(CI->getType(), SubStr1.compare(SubStr2));
398  }
399 
400  if (HasStr1 && Str1.empty()) // strncmp("", x, n) -> -*x
401  return B.CreateNeg(B.CreateZExt(
402  B.CreateLoad(B.getInt8Ty(), Str2P, "strcmpload"), CI->getType()));
403 
404  if (HasStr2 && Str2.empty()) // strncmp(x, "", n) -> *x
405  return B.CreateZExt(B.CreateLoad(B.getInt8Ty(), Str1P, "strcmpload"),
406  CI->getType());
407 
408  uint64_t Len1 = GetStringLength(Str1P);
409  uint64_t Len2 = GetStringLength(Str2P);
410 
411  // strncmp to memcmp
412  if (!HasStr1 && HasStr2) {
413  Len2 = std::min(Len2, Length);
414  if (canTransformToMemCmp(CI, Str1P, Len2, DL))
415  return emitMemCmp(
416  Str1P, Str2P,
417  ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len2), B, DL,
418  TLI);
419  } else if (HasStr1 && !HasStr2) {
420  Len1 = std::min(Len1, Length);
421  if (canTransformToMemCmp(CI, Str2P, Len1, DL))
422  return emitMemCmp(
423  Str1P, Str2P,
424  ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len1), B, DL,
425  TLI);
426  }
427 
428  return nullptr;
429 }
430 
431 Value *LibCallSimplifier::optimizeStrCpy(CallInst *CI, IRBuilder<> &B) {
432  Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1);
433  if (Dst == Src) // strcpy(x,x) -> x
434  return Src;
435 
436  // See if we can get the length of the input string.
437  uint64_t Len = GetStringLength(Src);
438  if (Len == 0)
439  return nullptr;
440 
441  // We have enough information to now generate the memcpy call to do the
442  // copy for us. Make a memcpy to copy the nul byte with align = 1.
443  B.CreateMemCpy(Dst, 1, Src, 1,
444  ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len));
445  return Dst;
446 }
447 
448 Value *LibCallSimplifier::optimizeStpCpy(CallInst *CI, IRBuilder<> &B) {
450  Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1);
451  if (Dst == Src) { // stpcpy(x,x) -> x+strlen(x)
452  Value *StrLen = emitStrLen(Src, B, DL, TLI);
453  return StrLen ? B.CreateInBoundsGEP(B.getInt8Ty(), Dst, StrLen) : nullptr;
454  }
455 
456  // See if we can get the length of the input string.
457  uint64_t Len = GetStringLength(Src);
458  if (Len == 0)
459  return nullptr;
460 
461  Type *PT = Callee->getFunctionType()->getParamType(0);
462  Value *LenV = ConstantInt::get(DL.getIntPtrType(PT), Len);
463  Value *DstEnd = B.CreateGEP(B.getInt8Ty(), Dst,
464  ConstantInt::get(DL.getIntPtrType(PT), Len - 1));
465 
466  // We have enough information to now generate the memcpy call to do the
467  // copy for us. Make a memcpy to copy the nul byte with align = 1.
468  B.CreateMemCpy(Dst, 1, Src, 1, LenV);
469  return DstEnd;
470 }
471 
472 Value *LibCallSimplifier::optimizeStrNCpy(CallInst *CI, IRBuilder<> &B) {
474  Value *Dst = CI->getArgOperand(0);
475  Value *Src = CI->getArgOperand(1);
476  Value *LenOp = CI->getArgOperand(2);
477 
478  // See if we can get the length of the input string.
479  uint64_t SrcLen = GetStringLength(Src);
480  if (SrcLen == 0)
481  return nullptr;
482  --SrcLen;
483 
484  if (SrcLen == 0) {
485  // strncpy(x, "", y) -> memset(align 1 x, '\0', y)
486  B.CreateMemSet(Dst, B.getInt8('\0'), LenOp, 1);
487  return Dst;
488  }
489 
490  uint64_t Len;
491  if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(LenOp))
492  Len = LengthArg->getZExtValue();
493  else
494  return nullptr;
495 
496  if (Len == 0)
497  return Dst; // strncpy(x, y, 0) -> x
498 
499  // Let strncpy handle the zero padding
500  if (Len > SrcLen + 1)
501  return nullptr;
502 
503  Type *PT = Callee->getFunctionType()->getParamType(0);
504  // strncpy(x, s, c) -> memcpy(align 1 x, align 1 s, c) [s and c are constant]
505  B.CreateMemCpy(Dst, 1, Src, 1, ConstantInt::get(DL.getIntPtrType(PT), Len));
506 
507  return Dst;
508 }
509 
510 Value *LibCallSimplifier::optimizeStringLength(CallInst *CI, IRBuilder<> &B,
511  unsigned CharSize) {
512  Value *Src = CI->getArgOperand(0);
513 
514  // Constant folding: strlen("xyz") -> 3
515  if (uint64_t Len = GetStringLength(Src, CharSize))
516  return ConstantInt::get(CI->getType(), Len - 1);
517 
518  // If s is a constant pointer pointing to a string literal, we can fold
519  // strlen(s + x) to strlen(s) - x, when x is known to be in the range
520  // [0, strlen(s)] or the string has a single null terminator '\0' at the end.
521  // We only try to simplify strlen when the pointer s points to an array
522  // of i8. Otherwise, we would need to scale the offset x before doing the
523  // subtraction. This will make the optimization more complex, and it's not
524  // very useful because calling strlen for a pointer of other types is
525  // very uncommon.
526  if (GEPOperator *GEP = dyn_cast<GEPOperator>(Src)) {
527  if (!isGEPBasedOnPointerToString(GEP, CharSize))
528  return nullptr;
529 
531  if (getConstantDataArrayInfo(GEP->getOperand(0), Slice, CharSize)) {
532  uint64_t NullTermIdx;
533  if (Slice.Array == nullptr) {
534  NullTermIdx = 0;
535  } else {
536  NullTermIdx = ~((uint64_t)0);
537  for (uint64_t I = 0, E = Slice.Length; I < E; ++I) {
538  if (Slice.Array->getElementAsInteger(I + Slice.Offset) == 0) {
539  NullTermIdx = I;
540  break;
541  }
542  }
543  // If the string does not have '\0', leave it to strlen to compute
544  // its length.
545  if (NullTermIdx == ~((uint64_t)0))
546  return nullptr;
547  }
548 
549  Value *Offset = GEP->getOperand(2);
550  KnownBits Known = computeKnownBits(Offset, DL, 0, nullptr, CI, nullptr);
551  Known.Zero.flipAllBits();
552  uint64_t ArrSize =
553  cast<ArrayType>(GEP->getSourceElementType())->getNumElements();
554 
555  // KnownZero's bits are flipped, so zeros in KnownZero now represent
556  // bits known to be zeros in Offset, and ones in KnowZero represent
557  // bits unknown in Offset. Therefore, Offset is known to be in range
558  // [0, NullTermIdx] when the flipped KnownZero is non-negative and
559  // unsigned-less-than NullTermIdx.
560  //
561  // If Offset is not provably in the range [0, NullTermIdx], we can still
562  // optimize if we can prove that the program has undefined behavior when
563  // Offset is outside that range. That is the case when GEP->getOperand(0)
564  // is a pointer to an object whose memory extent is NullTermIdx+1.
565  if ((Known.Zero.isNonNegative() && Known.Zero.ule(NullTermIdx)) ||
566  (GEP->isInBounds() && isa<GlobalVariable>(GEP->getOperand(0)) &&
567  NullTermIdx == ArrSize - 1)) {
568  Offset = B.CreateSExtOrTrunc(Offset, CI->getType());
569  return B.CreateSub(ConstantInt::get(CI->getType(), NullTermIdx),
570  Offset);
571  }
572  }
573 
574  return nullptr;
575  }
576 
577  // strlen(x?"foo":"bars") --> x ? 3 : 4
578  if (SelectInst *SI = dyn_cast<SelectInst>(Src)) {
579  uint64_t LenTrue = GetStringLength(SI->getTrueValue(), CharSize);
580  uint64_t LenFalse = GetStringLength(SI->getFalseValue(), CharSize);
581  if (LenTrue && LenFalse) {
582  ORE.emit([&]() {
583  return OptimizationRemark("instcombine", "simplify-libcalls", CI)
584  << "folded strlen(select) to select of constants";
585  });
586  return B.CreateSelect(SI->getCondition(),
587  ConstantInt::get(CI->getType(), LenTrue - 1),
588  ConstantInt::get(CI->getType(), LenFalse - 1));
589  }
590  }
591 
592  // strlen(x) != 0 --> *x != 0
593  // strlen(x) == 0 --> *x == 0
595  return B.CreateZExt(B.CreateLoad(B.getIntNTy(CharSize), Src, "strlenfirst"),
596  CI->getType());
597 
598  return nullptr;
599 }
600 
601 Value *LibCallSimplifier::optimizeStrLen(CallInst *CI, IRBuilder<> &B) {
602  return optimizeStringLength(CI, B, 8);
603 }
604 
605 Value *LibCallSimplifier::optimizeWcslen(CallInst *CI, IRBuilder<> &B) {
606  Module &M = *CI->getModule();
607  unsigned WCharSize = TLI->getWCharSize(M) * 8;
608  // We cannot perform this optimization without wchar_size metadata.
609  if (WCharSize == 0)
610  return nullptr;
611 
612  return optimizeStringLength(CI, B, WCharSize);
613 }
614 
615 Value *LibCallSimplifier::optimizeStrPBrk(CallInst *CI, IRBuilder<> &B) {
616  StringRef S1, S2;
617  bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1);
618  bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2);
619 
620  // strpbrk(s, "") -> nullptr
621  // strpbrk("", s) -> nullptr
622  if ((HasS1 && S1.empty()) || (HasS2 && S2.empty()))
623  return Constant::getNullValue(CI->getType());
624 
625  // Constant folding.
626  if (HasS1 && HasS2) {
627  size_t I = S1.find_first_of(S2);
628  if (I == StringRef::npos) // No match.
629  return Constant::getNullValue(CI->getType());
630 
631  return B.CreateGEP(B.getInt8Ty(), CI->getArgOperand(0), B.getInt64(I),
632  "strpbrk");
633  }
634 
635  // strpbrk(s, "a") -> strchr(s, 'a')
636  if (HasS2 && S2.size() == 1)
637  return emitStrChr(CI->getArgOperand(0), S2[0], B, TLI);
638 
639  return nullptr;
640 }
641 
642 Value *LibCallSimplifier::optimizeStrTo(CallInst *CI, IRBuilder<> &B) {
643  Value *EndPtr = CI->getArgOperand(1);
644  if (isa<ConstantPointerNull>(EndPtr)) {
645  // With a null EndPtr, this function won't capture the main argument.
646  // It would be readonly too, except that it still may write to errno.
647  CI->addParamAttr(0, Attribute::NoCapture);
648  }
649 
650  return nullptr;
651 }
652 
653 Value *LibCallSimplifier::optimizeStrSpn(CallInst *CI, IRBuilder<> &B) {
654  StringRef S1, S2;
655  bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1);
656  bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2);
657 
658  // strspn(s, "") -> 0
659  // strspn("", s) -> 0
660  if ((HasS1 && S1.empty()) || (HasS2 && S2.empty()))
661  return Constant::getNullValue(CI->getType());
662 
663  // Constant folding.
664  if (HasS1 && HasS2) {
665  size_t Pos = S1.find_first_not_of(S2);
666  if (Pos == StringRef::npos)
667  Pos = S1.size();
668  return ConstantInt::get(CI->getType(), Pos);
669  }
670 
671  return nullptr;
672 }
673 
674 Value *LibCallSimplifier::optimizeStrCSpn(CallInst *CI, IRBuilder<> &B) {
675  StringRef S1, S2;
676  bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1);
677  bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2);
678 
679  // strcspn("", s) -> 0
680  if (HasS1 && S1.empty())
681  return Constant::getNullValue(CI->getType());
682 
683  // Constant folding.
684  if (HasS1 && HasS2) {
685  size_t Pos = S1.find_first_of(S2);
686  if (Pos == StringRef::npos)
687  Pos = S1.size();
688  return ConstantInt::get(CI->getType(), Pos);
689  }
690 
691  // strcspn(s, "") -> strlen(s)
692  if (HasS2 && S2.empty())
693  return emitStrLen(CI->getArgOperand(0), B, DL, TLI);
694 
695  return nullptr;
696 }
697 
698 Value *LibCallSimplifier::optimizeStrStr(CallInst *CI, IRBuilder<> &B) {
699  // fold strstr(x, x) -> x.
700  if (CI->getArgOperand(0) == CI->getArgOperand(1))
701  return B.CreateBitCast(CI->getArgOperand(0), CI->getType());
702 
703  // fold strstr(a, b) == a -> strncmp(a, b, strlen(b)) == 0
705  Value *StrLen = emitStrLen(CI->getArgOperand(1), B, DL, TLI);
706  if (!StrLen)
707  return nullptr;
708  Value *StrNCmp = emitStrNCmp(CI->getArgOperand(0), CI->getArgOperand(1),
709  StrLen, B, DL, TLI);
710  if (!StrNCmp)
711  return nullptr;
712  for (auto UI = CI->user_begin(), UE = CI->user_end(); UI != UE;) {
713  ICmpInst *Old = cast<ICmpInst>(*UI++);
714  Value *Cmp =
715  B.CreateICmp(Old->getPredicate(), StrNCmp,
716  ConstantInt::getNullValue(StrNCmp->getType()), "cmp");
717  replaceAllUsesWith(Old, Cmp);
718  }
719  return CI;
720  }
721 
722  // See if either input string is a constant string.
723  StringRef SearchStr, ToFindStr;
724  bool HasStr1 = getConstantStringInfo(CI->getArgOperand(0), SearchStr);
725  bool HasStr2 = getConstantStringInfo(CI->getArgOperand(1), ToFindStr);
726 
727  // fold strstr(x, "") -> x.
728  if (HasStr2 && ToFindStr.empty())
729  return B.CreateBitCast(CI->getArgOperand(0), CI->getType());
730 
731  // If both strings are known, constant fold it.
732  if (HasStr1 && HasStr2) {
733  size_t Offset = SearchStr.find(ToFindStr);
734 
735  if (Offset == StringRef::npos) // strstr("foo", "bar") -> null
736  return Constant::getNullValue(CI->getType());
737 
738  // strstr("abcd", "bc") -> gep((char*)"abcd", 1)
739  Value *Result = castToCStr(CI->getArgOperand(0), B);
740  Result =
741  B.CreateConstInBoundsGEP1_64(B.getInt8Ty(), Result, Offset, "strstr");
742  return B.CreateBitCast(Result, CI->getType());
743  }
744 
745  // fold strstr(x, "y") -> strchr(x, 'y').
746  if (HasStr2 && ToFindStr.size() == 1) {
747  Value *StrChr = emitStrChr(CI->getArgOperand(0), ToFindStr[0], B, TLI);
748  return StrChr ? B.CreateBitCast(StrChr, CI->getType()) : nullptr;
749  }
750  return nullptr;
751 }
752 
753 Value *LibCallSimplifier::optimizeMemChr(CallInst *CI, IRBuilder<> &B) {
754  Value *SrcStr = CI->getArgOperand(0);
755  ConstantInt *CharC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
757 
758  // memchr(x, y, 0) -> null
759  if (LenC && LenC->isZero())
760  return Constant::getNullValue(CI->getType());
761 
762  // From now on we need at least constant length and string.
763  StringRef Str;
764  if (!LenC || !getConstantStringInfo(SrcStr, Str, 0, /*TrimAtNul=*/false))
765  return nullptr;
766 
767  // Truncate the string to LenC. If Str is smaller than LenC we will still only
768  // scan the string, as reading past the end of it is undefined and we can just
769  // return null if we don't find the char.
770  Str = Str.substr(0, LenC->getZExtValue());
771 
772  // If the char is variable but the input str and length are not we can turn
773  // this memchr call into a simple bit field test. Of course this only works
774  // when the return value is only checked against null.
775  //
776  // It would be really nice to reuse switch lowering here but we can't change
777  // the CFG at this point.
778  //
779  // memchr("\r\n", C, 2) != nullptr -> (1 << C & ((1 << '\r') | (1 << '\n')))
780  // != 0
781  // after bounds check.
782  if (!CharC && !Str.empty() && isOnlyUsedInZeroEqualityComparison(CI)) {
783  unsigned char Max =
784  *std::max_element(reinterpret_cast<const unsigned char *>(Str.begin()),
785  reinterpret_cast<const unsigned char *>(Str.end()));
786 
787  // Make sure the bit field we're about to create fits in a register on the
788  // target.
789  // FIXME: On a 64 bit architecture this prevents us from using the
790  // interesting range of alpha ascii chars. We could do better by emitting
791  // two bitfields or shifting the range by 64 if no lower chars are used.
792  if (!DL.fitsInLegalInteger(Max + 1))
793  return nullptr;
794 
795  // For the bit field use a power-of-2 type with at least 8 bits to avoid
796  // creating unnecessary illegal types.
797  unsigned char Width = NextPowerOf2(std::max((unsigned char)7, Max));
798 
799  // Now build the bit field.
800  APInt Bitfield(Width, 0);
801  for (char C : Str)
802  Bitfield.setBit((unsigned char)C);
803  Value *BitfieldC = B.getInt(Bitfield);
804 
805  // Adjust width of "C" to the bitfield width, then mask off the high bits.
806  Value *C = B.CreateZExtOrTrunc(CI->getArgOperand(1), BitfieldC->getType());
807  C = B.CreateAnd(C, B.getIntN(Width, 0xFF));
808 
809  // First check that the bit field access is within bounds.
810  Value *Bounds = B.CreateICmp(ICmpInst::ICMP_ULT, C, B.getIntN(Width, Width),
811  "memchr.bounds");
812 
813  // Create code that checks if the given bit is set in the field.
814  Value *Shl = B.CreateShl(B.getIntN(Width, 1ULL), C);
815  Value *Bits = B.CreateIsNotNull(B.CreateAnd(Shl, BitfieldC), "memchr.bits");
816 
817  // Finally merge both checks and cast to pointer type. The inttoptr
818  // implicitly zexts the i1 to intptr type.
819  return B.CreateIntToPtr(B.CreateAnd(Bounds, Bits, "memchr"), CI->getType());
820  }
821 
822  // Check if all arguments are constants. If so, we can constant fold.
823  if (!CharC)
824  return nullptr;
825 
826  // Compute the offset.
827  size_t I = Str.find(CharC->getSExtValue() & 0xFF);
828  if (I == StringRef::npos) // Didn't find the char. memchr returns null.
829  return Constant::getNullValue(CI->getType());
830 
831  // memchr(s+n,c,l) -> gep(s+n+i,c)
832  return B.CreateGEP(B.getInt8Ty(), SrcStr, B.getInt64(I), "memchr");
833 }
834 
836  uint64_t Len, IRBuilder<> &B,
837  const DataLayout &DL) {
838  if (Len == 0) // memcmp(s1,s2,0) -> 0
839  return Constant::getNullValue(CI->getType());
840 
841  // memcmp(S1,S2,1) -> *(unsigned char*)LHS - *(unsigned char*)RHS
842  if (Len == 1) {
843  Value *LHSV =
844  B.CreateZExt(B.CreateLoad(B.getInt8Ty(), castToCStr(LHS, B), "lhsc"),
845  CI->getType(), "lhsv");
846  Value *RHSV =
847  B.CreateZExt(B.CreateLoad(B.getInt8Ty(), castToCStr(RHS, B), "rhsc"),
848  CI->getType(), "rhsv");
849  return B.CreateSub(LHSV, RHSV, "chardiff");
850  }
851 
852  // memcmp(S1,S2,N/8)==0 -> (*(intN_t*)S1 != *(intN_t*)S2)==0
853  // TODO: The case where both inputs are constants does not need to be limited
854  // to legal integers or equality comparison. See block below this.
855  if (DL.isLegalInteger(Len * 8) && isOnlyUsedInZeroEqualityComparison(CI)) {
856  IntegerType *IntType = IntegerType::get(CI->getContext(), Len * 8);
857  unsigned PrefAlignment = DL.getPrefTypeAlignment(IntType);
858 
859  // First, see if we can fold either argument to a constant.
860  Value *LHSV = nullptr;
861  if (auto *LHSC = dyn_cast<Constant>(LHS)) {
862  LHSC = ConstantExpr::getBitCast(LHSC, IntType->getPointerTo());
863  LHSV = ConstantFoldLoadFromConstPtr(LHSC, IntType, DL);
864  }
865  Value *RHSV = nullptr;
866  if (auto *RHSC = dyn_cast<Constant>(RHS)) {
867  RHSC = ConstantExpr::getBitCast(RHSC, IntType->getPointerTo());
868  RHSV = ConstantFoldLoadFromConstPtr(RHSC, IntType, DL);
869  }
870 
871  // Don't generate unaligned loads. If either source is constant data,
872  // alignment doesn't matter for that source because there is no load.
873  if ((LHSV || getKnownAlignment(LHS, DL, CI) >= PrefAlignment) &&
874  (RHSV || getKnownAlignment(RHS, DL, CI) >= PrefAlignment)) {
875  if (!LHSV) {
876  Type *LHSPtrTy =
877  IntType->getPointerTo(LHS->getType()->getPointerAddressSpace());
878  LHSV = B.CreateLoad(IntType, B.CreateBitCast(LHS, LHSPtrTy), "lhsv");
879  }
880  if (!RHSV) {
881  Type *RHSPtrTy =
882  IntType->getPointerTo(RHS->getType()->getPointerAddressSpace());
883  RHSV = B.CreateLoad(IntType, B.CreateBitCast(RHS, RHSPtrTy), "rhsv");
884  }
885  return B.CreateZExt(B.CreateICmpNE(LHSV, RHSV), CI->getType(), "memcmp");
886  }
887  }
888 
889  // Constant folding: memcmp(x, y, Len) -> constant (all arguments are const).
890  // TODO: This is limited to i8 arrays.
891  StringRef LHSStr, RHSStr;
892  if (getConstantStringInfo(LHS, LHSStr) &&
893  getConstantStringInfo(RHS, RHSStr)) {
894  // Make sure we're not reading out-of-bounds memory.
895  if (Len > LHSStr.size() || Len > RHSStr.size())
896  return nullptr;
897  // Fold the memcmp and normalize the result. This way we get consistent
898  // results across multiple platforms.
899  uint64_t Ret = 0;
900  int Cmp = memcmp(LHSStr.data(), RHSStr.data(), Len);
901  if (Cmp < 0)
902  Ret = -1;
903  else if (Cmp > 0)
904  Ret = 1;
905  return ConstantInt::get(CI->getType(), Ret);
906  }
907  return nullptr;
908 }
909 
910 Value *LibCallSimplifier::optimizeMemCmp(CallInst *CI, IRBuilder<> &B) {
911  Value *LHS = CI->getArgOperand(0), *RHS = CI->getArgOperand(1);
912  Value *Size = CI->getArgOperand(2);
913 
914  if (LHS == RHS) // memcmp(s,s,x) -> 0
915  return Constant::getNullValue(CI->getType());
916 
917  // Handle constant lengths.
918  if (ConstantInt *LenC = dyn_cast<ConstantInt>(Size))
919  if (Value *Res = optimizeMemCmpConstantSize(CI, LHS, RHS,
920  LenC->getZExtValue(), B, DL))
921  return Res;
922 
923  // memcmp(x, y, Len) == 0 -> bcmp(x, y, Len) == 0
924  // `bcmp` can be more efficient than memcmp because it only has to know that
925  // there is a difference, not where it is.
926  if (isOnlyUsedInZeroEqualityComparison(CI) && TLI->has(LibFunc_bcmp)) {
927  return emitBCmp(LHS, RHS, Size, B, DL, TLI);
928  }
929 
930  return nullptr;
931 }
932 
933 Value *LibCallSimplifier::optimizeMemCpy(CallInst *CI, IRBuilder<> &B) {
934  // memcpy(x, y, n) -> llvm.memcpy(align 1 x, align 1 y, n)
935  B.CreateMemCpy(CI->getArgOperand(0), 1, CI->getArgOperand(1), 1,
936  CI->getArgOperand(2));
937  return CI->getArgOperand(0);
938 }
939 
940 Value *LibCallSimplifier::optimizeMemMove(CallInst *CI, IRBuilder<> &B) {
941  // memmove(x, y, n) -> llvm.memmove(align 1 x, align 1 y, n)
942  B.CreateMemMove(CI->getArgOperand(0), 1, CI->getArgOperand(1), 1,
943  CI->getArgOperand(2));
944  return CI->getArgOperand(0);
945 }
946 
947 /// Fold memset[_chk](malloc(n), 0, n) --> calloc(1, n).
948 Value *LibCallSimplifier::foldMallocMemset(CallInst *Memset, IRBuilder<> &B) {
949  // This has to be a memset of zeros (bzero).
950  auto *FillValue = dyn_cast<ConstantInt>(Memset->getArgOperand(1));
951  if (!FillValue || FillValue->getZExtValue() != 0)
952  return nullptr;
953 
954  // TODO: We should handle the case where the malloc has more than one use.
955  // This is necessary to optimize common patterns such as when the result of
956  // the malloc is checked against null or when a memset intrinsic is used in
957  // place of a memset library call.
958  auto *Malloc = dyn_cast<CallInst>(Memset->getArgOperand(0));
959  if (!Malloc || !Malloc->hasOneUse())
960  return nullptr;
961 
962  // Is the inner call really malloc()?
963  Function *InnerCallee = Malloc->getCalledFunction();
964  if (!InnerCallee)
965  return nullptr;
966 
967  LibFunc Func;
968  if (!TLI->getLibFunc(*InnerCallee, Func) || !TLI->has(Func) ||
969  Func != LibFunc_malloc)
970  return nullptr;
971 
972  // The memset must cover the same number of bytes that are malloc'd.
973  if (Memset->getArgOperand(2) != Malloc->getArgOperand(0))
974  return nullptr;
975 
976  // Replace the malloc with a calloc. We need the data layout to know what the
977  // actual size of a 'size_t' parameter is.
978  B.SetInsertPoint(Malloc->getParent(), ++Malloc->getIterator());
979  const DataLayout &DL = Malloc->getModule()->getDataLayout();
980  IntegerType *SizeType = DL.getIntPtrType(B.GetInsertBlock()->getContext());
981  Value *Calloc = emitCalloc(ConstantInt::get(SizeType, 1),
982  Malloc->getArgOperand(0), Malloc->getAttributes(),
983  B, *TLI);
984  if (!Calloc)
985  return nullptr;
986 
987  Malloc->replaceAllUsesWith(Calloc);
988  eraseFromParent(Malloc);
989 
990  return Calloc;
991 }
992 
993 Value *LibCallSimplifier::optimizeMemSet(CallInst *CI, IRBuilder<> &B) {
994  if (auto *Calloc = foldMallocMemset(CI, B))
995  return Calloc;
996 
997  // memset(p, v, n) -> llvm.memset(align 1 p, v, n)
998  Value *Val = B.CreateIntCast(CI->getArgOperand(1), B.getInt8Ty(), false);
999  B.CreateMemSet(CI->getArgOperand(0), Val, CI->getArgOperand(2), 1);
1000  return CI->getArgOperand(0);
1001 }
1002 
1003 Value *LibCallSimplifier::optimizeRealloc(CallInst *CI, IRBuilder<> &B) {
1004  if (isa<ConstantPointerNull>(CI->getArgOperand(0)))
1005  return emitMalloc(CI->getArgOperand(1), B, DL, TLI);
1006 
1007  return nullptr;
1008 }
1009 
1010 //===----------------------------------------------------------------------===//
1011 // Math Library Optimizations
1012 //===----------------------------------------------------------------------===//
1013 
1014 // Replace a libcall \p CI with a call to intrinsic \p IID
1016  // Propagate fast-math flags from the existing call to the new call.
1019 
1020  Module *M = CI->getModule();
1021  Value *V = CI->getArgOperand(0);
1022  Function *F = Intrinsic::getDeclaration(M, IID, CI->getType());
1023  CallInst *NewCall = B.CreateCall(F, V);
1024  NewCall->takeName(CI);
1025  return NewCall;
1026 }
1027 
1028 /// Return a variant of Val with float type.
1029 /// Currently this works in two cases: If Val is an FPExtension of a float
1030 /// value to something bigger, simply return the operand.
1031 /// If Val is a ConstantFP but can be converted to a float ConstantFP without
1032 /// loss of precision do so.
1034  if (FPExtInst *Cast = dyn_cast<FPExtInst>(Val)) {
1035  Value *Op = Cast->getOperand(0);
1036  if (Op->getType()->isFloatTy())
1037  return Op;
1038  }
1039  if (ConstantFP *Const = dyn_cast<ConstantFP>(Val)) {
1040  APFloat F = Const->getValueAPF();
1041  bool losesInfo;
1043  &losesInfo);
1044  if (!losesInfo)
1045  return ConstantFP::get(Const->getContext(), F);
1046  }
1047  return nullptr;
1048 }
1049 
1050 /// Shrink double -> float functions.
1052  bool isBinary, bool isPrecise = false) {
1053  if (!CI->getType()->isDoubleTy())
1054  return nullptr;
1055 
1056  // If not all the uses of the function are converted to float, then bail out.
1057  // This matters if the precision of the result is more important than the
1058  // precision of the arguments.
1059  if (isPrecise)
1060  for (User *U : CI->users()) {
1061  FPTruncInst *Cast = dyn_cast<FPTruncInst>(U);
1062  if (!Cast || !Cast->getType()->isFloatTy())
1063  return nullptr;
1064  }
1065 
1066  // If this is something like 'g((double) float)', convert to 'gf(float)'.
1067  Value *V[2];
1068  V[0] = valueHasFloatPrecision(CI->getArgOperand(0));
1069  V[1] = isBinary ? valueHasFloatPrecision(CI->getArgOperand(1)) : nullptr;
1070  if (!V[0] || (isBinary && !V[1]))
1071  return nullptr;
1072 
1073  // If call isn't an intrinsic, check that it isn't within a function with the
1074  // same name as the float version of this call, otherwise the result is an
1075  // infinite loop. For example, from MinGW-w64:
1076  //
1077  // float expf(float val) { return (float) exp((double) val); }
1078  Function *CalleeFn = CI->getCalledFunction();
1079  StringRef CalleeNm = CalleeFn->getName();
1080  AttributeList CalleeAt = CalleeFn->getAttributes();
1081  if (CalleeFn && !CalleeFn->isIntrinsic()) {
1082  const Function *Fn = CI->getFunction();
1083  StringRef FnName = Fn->getName();
1084  if (FnName.back() == 'f' &&
1085  FnName.size() == (CalleeNm.size() + 1) &&
1086  FnName.startswith(CalleeNm))
1087  return nullptr;
1088  }
1089 
1090  // Propagate the math semantics from the current function to the new function.
1093 
1094  // g((double) float) -> (double) gf(float)
1095  Value *R;
1096  if (CalleeFn->isIntrinsic()) {
1097  Module *M = CI->getModule();
1098  Intrinsic::ID IID = CalleeFn->getIntrinsicID();
1099  Function *Fn = Intrinsic::getDeclaration(M, IID, B.getFloatTy());
1100  R = isBinary ? B.CreateCall(Fn, V) : B.CreateCall(Fn, V[0]);
1101  }
1102  else
1103  R = isBinary ? emitBinaryFloatFnCall(V[0], V[1], CalleeNm, B, CalleeAt)
1104  : emitUnaryFloatFnCall(V[0], CalleeNm, B, CalleeAt);
1105 
1106  return B.CreateFPExt(R, B.getDoubleTy());
1107 }
1108 
1109 /// Shrink double -> float for unary functions.
1111  bool isPrecise = false) {
1112  return optimizeDoubleFP(CI, B, false, isPrecise);
1113 }
1114 
1115 /// Shrink double -> float for binary functions.
1117  bool isPrecise = false) {
1118  return optimizeDoubleFP(CI, B, true, isPrecise);
1119 }
1120 
1121 // cabs(z) -> sqrt((creal(z)*creal(z)) + (cimag(z)*cimag(z)))
1122 Value *LibCallSimplifier::optimizeCAbs(CallInst *CI, IRBuilder<> &B) {
1123  if (!CI->isFast())
1124  return nullptr;
1125 
1126  // Propagate fast-math flags from the existing call to new instructions.
1129 
1130  Value *Real, *Imag;
1131  if (CI->getNumArgOperands() == 1) {
1132  Value *Op = CI->getArgOperand(0);
1133  assert(Op->getType()->isArrayTy() && "Unexpected signature for cabs!");
1134  Real = B.CreateExtractValue(Op, 0, "real");
1135  Imag = B.CreateExtractValue(Op, 1, "imag");
1136  } else {
1137  assert(CI->getNumArgOperands() == 2 && "Unexpected signature for cabs!");
1138  Real = CI->getArgOperand(0);
1139  Imag = CI->getArgOperand(1);
1140  }
1141 
1142  Value *RealReal = B.CreateFMul(Real, Real);
1143  Value *ImagImag = B.CreateFMul(Imag, Imag);
1144 
1145  Function *FSqrt = Intrinsic::getDeclaration(CI->getModule(), Intrinsic::sqrt,
1146  CI->getType());
1147  return B.CreateCall(FSqrt, B.CreateFAdd(RealReal, ImagImag), "cabs");
1148 }
1149 
1151  IRBuilder<> &B) {
1152  if (!isa<FPMathOperator>(Call))
1153  return nullptr;
1154 
1157 
1158  // TODO: Can this be shared to also handle LLVM intrinsics?
1159  Value *X;
1160  switch (Func) {
1161  case LibFunc_sin:
1162  case LibFunc_sinf:
1163  case LibFunc_sinl:
1164  case LibFunc_tan:
1165  case LibFunc_tanf:
1166  case LibFunc_tanl:
1167  // sin(-X) --> -sin(X)
1168  // tan(-X) --> -tan(X)
1169  if (match(Call->getArgOperand(0), m_OneUse(m_FNeg(m_Value(X)))))
1170  return B.CreateFNeg(B.CreateCall(Call->getCalledFunction(), X));
1171  break;
1172  case LibFunc_cos:
1173  case LibFunc_cosf:
1174  case LibFunc_cosl:
1175  // cos(-X) --> cos(X)
1176  if (match(Call->getArgOperand(0), m_FNeg(m_Value(X))))
1177  return B.CreateCall(Call->getCalledFunction(), X, "cos");
1178  break;
1179  default:
1180  break;
1181  }
1182  return nullptr;
1183 }
1184 
1185 static Value *getPow(Value *InnerChain[33], unsigned Exp, IRBuilder<> &B) {
1186  // Multiplications calculated using Addition Chains.
1187  // Refer: http://wwwhomes.uni-bielefeld.de/achim/addition_chain.html
1188 
1189  assert(Exp != 0 && "Incorrect exponent 0 not handled");
1190 
1191  if (InnerChain[Exp])
1192  return InnerChain[Exp];
1193 
1194  static const unsigned AddChain[33][2] = {
1195  {0, 0}, // Unused.
1196  {0, 0}, // Unused (base case = pow1).
1197  {1, 1}, // Unused (pre-computed).
1198  {1, 2}, {2, 2}, {2, 3}, {3, 3}, {2, 5}, {4, 4},
1199  {1, 8}, {5, 5}, {1, 10}, {6, 6}, {4, 9}, {7, 7},
1200  {3, 12}, {8, 8}, {8, 9}, {2, 16}, {1, 18}, {10, 10},
1201  {6, 15}, {11, 11}, {3, 20}, {12, 12}, {8, 17}, {13, 13},
1202  {3, 24}, {14, 14}, {4, 25}, {15, 15}, {3, 28}, {16, 16},
1203  };
1204 
1205  InnerChain[Exp] = B.CreateFMul(getPow(InnerChain, AddChain[Exp][0], B),
1206  getPow(InnerChain, AddChain[Exp][1], B));
1207  return InnerChain[Exp];
1208 }
1209 
1210 /// Use exp{,2}(x * y) for pow(exp{,2}(x), y);
1211 /// exp2(n * x) for pow(2.0 ** n, x); exp10(x) for pow(10.0, x).
1212 Value *LibCallSimplifier::replacePowWithExp(CallInst *Pow, IRBuilder<> &B) {
1213  Value *Base = Pow->getArgOperand(0), *Expo = Pow->getArgOperand(1);
1215  Module *Mod = Pow->getModule();
1216  Type *Ty = Pow->getType();
1217  bool Ignored;
1218 
1219  // Evaluate special cases related to a nested function as the base.
1220 
1221  // pow(exp(x), y) -> exp(x * y)
1222  // pow(exp2(x), y) -> exp2(x * y)
1223  // If exp{,2}() is used only once, it is better to fold two transcendental
1224  // math functions into one. If used again, exp{,2}() would still have to be
1225  // called with the original argument, then keep both original transcendental
1226  // functions. However, this transformation is only safe with fully relaxed
1227  // math semantics, since, besides rounding differences, it changes overflow
1228  // and underflow behavior quite dramatically. For example:
1229  // pow(exp(1000), 0.001) = pow(inf, 0.001) = inf
1230  // Whereas:
1231  // exp(1000 * 0.001) = exp(1)
1232  // TODO: Loosen the requirement for fully relaxed math semantics.
1233  // TODO: Handle exp10() when more targets have it available.
1234  CallInst *BaseFn = dyn_cast<CallInst>(Base);
1235  if (BaseFn && BaseFn->hasOneUse() && BaseFn->isFast() && Pow->isFast()) {
1236  LibFunc LibFn;
1237 
1238  Function *CalleeFn = BaseFn->getCalledFunction();
1239  if (CalleeFn &&
1240  TLI->getLibFunc(CalleeFn->getName(), LibFn) && TLI->has(LibFn)) {
1241  StringRef ExpName;
1242  Intrinsic::ID ID;
1243  Value *ExpFn;
1244  LibFunc LibFnFloat;
1245  LibFunc LibFnDouble;
1246  LibFunc LibFnLongDouble;
1247 
1248  switch (LibFn) {
1249  default:
1250  return nullptr;
1251  case LibFunc_expf: case LibFunc_exp: case LibFunc_expl:
1252  ExpName = TLI->getName(LibFunc_exp);
1253  ID = Intrinsic::exp;
1254  LibFnFloat = LibFunc_expf;
1255  LibFnDouble = LibFunc_exp;
1256  LibFnLongDouble = LibFunc_expl;
1257  break;
1258  case LibFunc_exp2f: case LibFunc_exp2: case LibFunc_exp2l:
1259  ExpName = TLI->getName(LibFunc_exp2);
1260  ID = Intrinsic::exp2;
1261  LibFnFloat = LibFunc_exp2f;
1262  LibFnDouble = LibFunc_exp2;
1263  LibFnLongDouble = LibFunc_exp2l;
1264  break;
1265  }
1266 
1267  // Create new exp{,2}() with the product as its argument.
1268  Value *FMul = B.CreateFMul(BaseFn->getArgOperand(0), Expo, "mul");
1269  ExpFn = BaseFn->doesNotAccessMemory()
1270  ? B.CreateCall(Intrinsic::getDeclaration(Mod, ID, Ty),
1271  FMul, ExpName)
1272  : emitUnaryFloatFnCall(FMul, TLI, LibFnDouble, LibFnFloat,
1273  LibFnLongDouble, B,
1274  BaseFn->getAttributes());
1275 
1276  // Since the new exp{,2}() is different from the original one, dead code
1277  // elimination cannot be trusted to remove it, since it may have side
1278  // effects (e.g., errno). When the only consumer for the original
1279  // exp{,2}() is pow(), then it has to be explicitly erased.
1280  BaseFn->replaceAllUsesWith(ExpFn);
1281  eraseFromParent(BaseFn);
1282 
1283  return ExpFn;
1284  }
1285  }
1286 
1287  // Evaluate special cases related to a constant base.
1288 
1289  const APFloat *BaseF;
1290  if (!match(Pow->getArgOperand(0), m_APFloat(BaseF)))
1291  return nullptr;
1292 
1293  // pow(2.0 ** n, x) -> exp2(n * x)
1294  if (hasUnaryFloatFn(TLI, Ty, LibFunc_exp2, LibFunc_exp2f, LibFunc_exp2l)) {
1295  APFloat BaseR = APFloat(1.0);
1296  BaseR.convert(BaseF->getSemantics(), APFloat::rmTowardZero, &Ignored);
1297  BaseR = BaseR / *BaseF;
1298  bool IsInteger = BaseF->isInteger(),
1299  IsReciprocal = BaseR.isInteger();
1300  const APFloat *NF = IsReciprocal ? &BaseR : BaseF;
1301  APSInt NI(64, false);
1302  if ((IsInteger || IsReciprocal) &&
1303  !NF->convertToInteger(NI, APFloat::rmTowardZero, &Ignored) &&
1304  NI > 1 && NI.isPowerOf2()) {
1305  double N = NI.logBase2() * (IsReciprocal ? -1.0 : 1.0);
1306  Value *FMul = B.CreateFMul(Expo, ConstantFP::get(Ty, N), "mul");
1307  if (Pow->doesNotAccessMemory())
1308  return B.CreateCall(Intrinsic::getDeclaration(Mod, Intrinsic::exp2, Ty),
1309  FMul, "exp2");
1310  else
1311  return emitUnaryFloatFnCall(FMul, TLI, LibFunc_exp2, LibFunc_exp2f,
1312  LibFunc_exp2l, B, Attrs);
1313  }
1314  }
1315 
1316  // pow(10.0, x) -> exp10(x)
1317  // TODO: There is no exp10() intrinsic yet, but some day there shall be one.
1318  if (match(Base, m_SpecificFP(10.0)) &&
1319  hasUnaryFloatFn(TLI, Ty, LibFunc_exp10, LibFunc_exp10f, LibFunc_exp10l))
1320  return emitUnaryFloatFnCall(Expo, TLI, LibFunc_exp10, LibFunc_exp10f,
1321  LibFunc_exp10l, B, Attrs);
1322 
1323  return nullptr;
1324 }
1325 
1326 static Value *getSqrtCall(Value *V, AttributeList Attrs, bool NoErrno,
1327  Module *M, IRBuilder<> &B,
1328  const TargetLibraryInfo *TLI) {
1329  // If errno is never set, then use the intrinsic for sqrt().
1330  if (NoErrno) {
1331  Function *SqrtFn =
1332  Intrinsic::getDeclaration(M, Intrinsic::sqrt, V->getType());
1333  return B.CreateCall(SqrtFn, V, "sqrt");
1334  }
1335 
1336  // Otherwise, use the libcall for sqrt().
1337  if (hasUnaryFloatFn(TLI, V->getType(), LibFunc_sqrt, LibFunc_sqrtf,
1338  LibFunc_sqrtl))
1339  // TODO: We also should check that the target can in fact lower the sqrt()
1340  // libcall. We currently have no way to ask this question, so we ask if
1341  // the target has a sqrt() libcall, which is not exactly the same.
1342  return emitUnaryFloatFnCall(V, TLI, LibFunc_sqrt, LibFunc_sqrtf,
1343  LibFunc_sqrtl, B, Attrs);
1344 
1345  return nullptr;
1346 }
1347 
1348 /// Use square root in place of pow(x, +/-0.5).
1349 Value *LibCallSimplifier::replacePowWithSqrt(CallInst *Pow, IRBuilder<> &B) {
1350  Value *Sqrt, *Base = Pow->getArgOperand(0), *Expo = Pow->getArgOperand(1);
1352  Module *Mod = Pow->getModule();
1353  Type *Ty = Pow->getType();
1354 
1355  const APFloat *ExpoF;
1356  if (!match(Expo, m_APFloat(ExpoF)) ||
1357  (!ExpoF->isExactlyValue(0.5) && !ExpoF->isExactlyValue(-0.5)))
1358  return nullptr;
1359 
1360  Sqrt = getSqrtCall(Base, Attrs, Pow->doesNotAccessMemory(), Mod, B, TLI);
1361  if (!Sqrt)
1362  return nullptr;
1363 
1364  // Handle signed zero base by expanding to fabs(sqrt(x)).
1365  if (!Pow->hasNoSignedZeros()) {
1366  Function *FAbsFn = Intrinsic::getDeclaration(Mod, Intrinsic::fabs, Ty);
1367  Sqrt = B.CreateCall(FAbsFn, Sqrt, "abs");
1368  }
1369 
1370  // Handle non finite base by expanding to
1371  // (x == -infinity ? +infinity : sqrt(x)).
1372  if (!Pow->hasNoInfs()) {
1373  Value *PosInf = ConstantFP::getInfinity(Ty),
1374  *NegInf = ConstantFP::getInfinity(Ty, true);
1375  Value *FCmp = B.CreateFCmpOEQ(Base, NegInf, "isinf");
1376  Sqrt = B.CreateSelect(FCmp, PosInf, Sqrt);
1377  }
1378 
1379  // If the exponent is negative, then get the reciprocal.
1380  if (ExpoF->isNegative())
1381  Sqrt = B.CreateFDiv(ConstantFP::get(Ty, 1.0), Sqrt, "reciprocal");
1382 
1383  return Sqrt;
1384 }
1385 
1386 Value *LibCallSimplifier::optimizePow(CallInst *Pow, IRBuilder<> &B) {
1387  Value *Base = Pow->getArgOperand(0), *Expo = Pow->getArgOperand(1);
1388  Function *Callee = Pow->getCalledFunction();
1389  StringRef Name = Callee->getName();
1390  Type *Ty = Pow->getType();
1391  Value *Shrunk = nullptr;
1392  bool Ignored;
1393 
1394  // Bail out if simplifying libcalls to pow() is disabled.
1395  if (!hasUnaryFloatFn(TLI, Ty, LibFunc_pow, LibFunc_powf, LibFunc_powl))
1396  return nullptr;
1397 
1398  // Propagate the math semantics from the call to any created instructions.
1401 
1402  // Shrink pow() to powf() if the arguments are single precision,
1403  // unless the result is expected to be double precision.
1404  if (UnsafeFPShrink &&
1405  Name == TLI->getName(LibFunc_pow) && hasFloatVersion(Name))
1406  Shrunk = optimizeBinaryDoubleFP(Pow, B, true);
1407 
1408  // Evaluate special cases related to the base.
1409 
1410  // pow(1.0, x) -> 1.0
1411  if (match(Base, m_FPOne()))
1412  return Base;
1413 
1414  if (Value *Exp = replacePowWithExp(Pow, B))
1415  return Exp;
1416 
1417  // Evaluate special cases related to the exponent.
1418 
1419  // pow(x, -1.0) -> 1.0 / x
1420  if (match(Expo, m_SpecificFP(-1.0)))
1421  return B.CreateFDiv(ConstantFP::get(Ty, 1.0), Base, "reciprocal");
1422 
1423  // pow(x, 0.0) -> 1.0
1424  if (match(Expo, m_SpecificFP(0.0)))
1425  return ConstantFP::get(Ty, 1.0);
1426 
1427  // pow(x, 1.0) -> x
1428  if (match(Expo, m_FPOne()))
1429  return Base;
1430 
1431  // pow(x, 2.0) -> x * x
1432  if (match(Expo, m_SpecificFP(2.0)))
1433  return B.CreateFMul(Base, Base, "square");
1434 
1435  if (Value *Sqrt = replacePowWithSqrt(Pow, B))
1436  return Sqrt;
1437 
1438  // pow(x, n) -> x * x * x * ...
1439  const APFloat *ExpoF;
1440  if (Pow->isFast() && match(Expo, m_APFloat(ExpoF))) {
1441  // We limit to a max of 7 multiplications, thus the maximum exponent is 32.
1442  // If the exponent is an integer+0.5 we generate a call to sqrt and an
1443  // additional fmul.
1444  // TODO: This whole transformation should be backend specific (e.g. some
1445  // backends might prefer libcalls or the limit for the exponent might
1446  // be different) and it should also consider optimizing for size.
1447  APFloat LimF(ExpoF->getSemantics(), 33.0),
1448  ExpoA(abs(*ExpoF));
1449  if (ExpoA.compare(LimF) == APFloat::cmpLessThan) {
1450  // This transformation applies to integer or integer+0.5 exponents only.
1451  // For integer+0.5, we create a sqrt(Base) call.
1452  Value *Sqrt = nullptr;
1453  if (!ExpoA.isInteger()) {
1454  APFloat Expo2 = ExpoA;
1455  // To check if ExpoA is an integer + 0.5, we add it to itself. If there
1456  // is no floating point exception and the result is an integer, then
1457  // ExpoA == integer + 0.5
1458  if (Expo2.add(ExpoA, APFloat::rmNearestTiesToEven) != APFloat::opOK)
1459  return nullptr;
1460 
1461  if (!Expo2.isInteger())
1462  return nullptr;
1463 
1464  Sqrt =
1466  Pow->doesNotAccessMemory(), Pow->getModule(), B, TLI);
1467  }
1468 
1469  // We will memoize intermediate products of the Addition Chain.
1470  Value *InnerChain[33] = {nullptr};
1471  InnerChain[1] = Base;
1472  InnerChain[2] = B.CreateFMul(Base, Base, "square");
1473 
1474  // We cannot readily convert a non-double type (like float) to a double.
1475  // So we first convert it to something which could be converted to double.
1477  Value *FMul = getPow(InnerChain, ExpoA.convertToDouble(), B);
1478 
1479  // Expand pow(x, y+0.5) to pow(x, y) * sqrt(x).
1480  if (Sqrt)
1481  FMul = B.CreateFMul(FMul, Sqrt);
1482 
1483  // If the exponent is negative, then get the reciprocal.
1484  if (ExpoF->isNegative())
1485  FMul = B.CreateFDiv(ConstantFP::get(Ty, 1.0), FMul, "reciprocal");
1486 
1487  return FMul;
1488  }
1489  }
1490 
1491  return Shrunk;
1492 }
1493 
1494 Value *LibCallSimplifier::optimizeExp2(CallInst *CI, IRBuilder<> &B) {
1496  Value *Ret = nullptr;
1497  StringRef Name = Callee->getName();
1498  if (UnsafeFPShrink && Name == "exp2" && hasFloatVersion(Name))
1499  Ret = optimizeUnaryDoubleFP(CI, B, true);
1500 
1501  Value *Op = CI->getArgOperand(0);
1502  // Turn exp2(sitofp(x)) -> ldexp(1.0, sext(x)) if sizeof(x) <= 32
1503  // Turn exp2(uitofp(x)) -> ldexp(1.0, zext(x)) if sizeof(x) < 32
1504  LibFunc LdExp = LibFunc_ldexpl;
1505  if (Op->getType()->isFloatTy())
1506  LdExp = LibFunc_ldexpf;
1507  else if (Op->getType()->isDoubleTy())
1508  LdExp = LibFunc_ldexp;
1509 
1510  if (TLI->has(LdExp)) {
1511  Value *LdExpArg = nullptr;
1512  if (SIToFPInst *OpC = dyn_cast<SIToFPInst>(Op)) {
1513  if (OpC->getOperand(0)->getType()->getPrimitiveSizeInBits() <= 32)
1514  LdExpArg = B.CreateSExt(OpC->getOperand(0), B.getInt32Ty());
1515  } else if (UIToFPInst *OpC = dyn_cast<UIToFPInst>(Op)) {
1516  if (OpC->getOperand(0)->getType()->getPrimitiveSizeInBits() < 32)
1517  LdExpArg = B.CreateZExt(OpC->getOperand(0), B.getInt32Ty());
1518  }
1519 
1520  if (LdExpArg) {
1521  Constant *One = ConstantFP::get(CI->getContext(), APFloat(1.0f));
1522  if (!Op->getType()->isFloatTy())
1523  One = ConstantExpr::getFPExtend(One, Op->getType());
1524 
1525  Module *M = CI->getModule();
1526  FunctionCallee NewCallee = M->getOrInsertFunction(
1527  TLI->getName(LdExp), Op->getType(), Op->getType(), B.getInt32Ty());
1528  CallInst *CI = B.CreateCall(NewCallee, {One, LdExpArg});
1529  if (const Function *F = dyn_cast<Function>(Callee->stripPointerCasts()))
1530  CI->setCallingConv(F->getCallingConv());
1531 
1532  return CI;
1533  }
1534  }
1535  return Ret;
1536 }
1537 
1538 Value *LibCallSimplifier::optimizeFMinFMax(CallInst *CI, IRBuilder<> &B) {
1540  // If we can shrink the call to a float function rather than a double
1541  // function, do that first.
1542  StringRef Name = Callee->getName();
1543  if ((Name == "fmin" || Name == "fmax") && hasFloatVersion(Name))
1544  if (Value *Ret = optimizeBinaryDoubleFP(CI, B))
1545  return Ret;
1546 
1548  FastMathFlags FMF;
1549  if (CI->isFast()) {
1550  // If the call is 'fast', then anything we create here will also be 'fast'.
1551  FMF.setFast();
1552  } else {
1553  // At a minimum, no-nans-fp-math must be true.
1554  if (!CI->hasNoNaNs())
1555  return nullptr;
1556  // No-signed-zeros is implied by the definitions of fmax/fmin themselves:
1557  // "Ideally, fmax would be sensitive to the sign of zero, for example
1558  // fmax(-0. 0, +0. 0) would return +0; however, implementation in software
1559  // might be impractical."
1560  FMF.setNoSignedZeros();
1561  FMF.setNoNaNs();
1562  }
1563  B.setFastMathFlags(FMF);
1564 
1565  // We have a relaxed floating-point environment. We can ignore NaN-handling
1566  // and transform to a compare and select. We do not have to consider errno or
1567  // exceptions, because fmin/fmax do not have those.
1568  Value *Op0 = CI->getArgOperand(0);
1569  Value *Op1 = CI->getArgOperand(1);
1570  Value *Cmp = Callee->getName().startswith("fmin") ?
1571  B.CreateFCmpOLT(Op0, Op1) : B.CreateFCmpOGT(Op0, Op1);
1572  return B.CreateSelect(Cmp, Op0, Op1);
1573 }
1574 
1575 Value *LibCallSimplifier::optimizeLog(CallInst *CI, IRBuilder<> &B) {
1577  Value *Ret = nullptr;
1578  StringRef Name = Callee->getName();
1579  if (UnsafeFPShrink && hasFloatVersion(Name))
1580  Ret = optimizeUnaryDoubleFP(CI, B, true);
1581 
1582  if (!CI->isFast())
1583  return Ret;
1584  Value *Op1 = CI->getArgOperand(0);
1585  auto *OpC = dyn_cast<CallInst>(Op1);
1586 
1587  // The earlier call must also be 'fast' in order to do these transforms.
1588  if (!OpC || !OpC->isFast())
1589  return Ret;
1590 
1591  // log(pow(x,y)) -> y*log(x)
1592  // This is only applicable to log, log2, log10.
1593  if (Name != "log" && Name != "log2" && Name != "log10")
1594  return Ret;
1595 
1597  FastMathFlags FMF;
1598  FMF.setFast();
1599  B.setFastMathFlags(FMF);
1600 
1601  LibFunc Func;
1602  Function *F = OpC->getCalledFunction();
1603  if (F && ((TLI->getLibFunc(F->getName(), Func) && TLI->has(Func) &&
1604  Func == LibFunc_pow) || F->getIntrinsicID() == Intrinsic::pow))
1605  return B.CreateFMul(OpC->getArgOperand(1),
1606  emitUnaryFloatFnCall(OpC->getOperand(0), Callee->getName(), B,
1607  Callee->getAttributes()), "mul");
1608 
1609  // log(exp2(y)) -> y*log(2)
1610  if (F && Name == "log" && TLI->getLibFunc(F->getName(), Func) &&
1611  TLI->has(Func) && Func == LibFunc_exp2)
1612  return B.CreateFMul(
1613  OpC->getArgOperand(0),
1615  Callee->getName(), B, Callee->getAttributes()),
1616  "logmul");
1617  return Ret;
1618 }
1619 
1620 Value *LibCallSimplifier::optimizeSqrt(CallInst *CI, IRBuilder<> &B) {
1622  Value *Ret = nullptr;
1623  // TODO: Once we have a way (other than checking for the existince of the
1624  // libcall) to tell whether our target can lower @llvm.sqrt, relax the
1625  // condition below.
1626  if (TLI->has(LibFunc_sqrtf) && (Callee->getName() == "sqrt" ||
1627  Callee->getIntrinsicID() == Intrinsic::sqrt))
1628  Ret = optimizeUnaryDoubleFP(CI, B, true);
1629 
1630  if (!CI->isFast())
1631  return Ret;
1632 
1634  if (!I || I->getOpcode() != Instruction::FMul || !I->isFast())
1635  return Ret;
1636 
1637  // We're looking for a repeated factor in a multiplication tree,
1638  // so we can do this fold: sqrt(x * x) -> fabs(x);
1639  // or this fold: sqrt((x * x) * y) -> fabs(x) * sqrt(y).
1640  Value *Op0 = I->getOperand(0);
1641  Value *Op1 = I->getOperand(1);
1642  Value *RepeatOp = nullptr;
1643  Value *OtherOp = nullptr;
1644  if (Op0 == Op1) {
1645  // Simple match: the operands of the multiply are identical.
1646  RepeatOp = Op0;
1647  } else {
1648  // Look for a more complicated pattern: one of the operands is itself
1649  // a multiply, so search for a common factor in that multiply.
1650  // Note: We don't bother looking any deeper than this first level or for
1651  // variations of this pattern because instcombine's visitFMUL and/or the
1652  // reassociation pass should give us this form.
1653  Value *OtherMul0, *OtherMul1;
1654  if (match(Op0, m_FMul(m_Value(OtherMul0), m_Value(OtherMul1)))) {
1655  // Pattern: sqrt((x * y) * z)
1656  if (OtherMul0 == OtherMul1 && cast<Instruction>(Op0)->isFast()) {
1657  // Matched: sqrt((x * x) * z)
1658  RepeatOp = OtherMul0;
1659  OtherOp = Op1;
1660  }
1661  }
1662  }
1663  if (!RepeatOp)
1664  return Ret;
1665 
1666  // Fast math flags for any created instructions should match the sqrt
1667  // and multiply.
1670 
1671  // If we found a repeated factor, hoist it out of the square root and
1672  // replace it with the fabs of that factor.
1673  Module *M = Callee->getParent();
1674  Type *ArgType = I->getType();
1675  Function *Fabs = Intrinsic::getDeclaration(M, Intrinsic::fabs, ArgType);
1676  Value *FabsCall = B.CreateCall(Fabs, RepeatOp, "fabs");
1677  if (OtherOp) {
1678  // If we found a non-repeated factor, we still need to get its square
1679  // root. We then multiply that by the value that was simplified out
1680  // of the square root calculation.
1681  Function *Sqrt = Intrinsic::getDeclaration(M, Intrinsic::sqrt, ArgType);
1682  Value *SqrtCall = B.CreateCall(Sqrt, OtherOp, "sqrt");
1683  return B.CreateFMul(FabsCall, SqrtCall);
1684  }
1685  return FabsCall;
1686 }
1687 
1688 // TODO: Generalize to handle any trig function and its inverse.
1689 Value *LibCallSimplifier::optimizeTan(CallInst *CI, IRBuilder<> &B) {
1691  Value *Ret = nullptr;
1692  StringRef Name = Callee->getName();
1693  if (UnsafeFPShrink && Name == "tan" && hasFloatVersion(Name))
1694  Ret = optimizeUnaryDoubleFP(CI, B, true);
1695 
1696  Value *Op1 = CI->getArgOperand(0);
1697  auto *OpC = dyn_cast<CallInst>(Op1);
1698  if (!OpC)
1699  return Ret;
1700 
1701  // Both calls must be 'fast' in order to remove them.
1702  if (!CI->isFast() || !OpC->isFast())
1703  return Ret;
1704 
1705  // tan(atan(x)) -> x
1706  // tanf(atanf(x)) -> x
1707  // tanl(atanl(x)) -> x
1708  LibFunc Func;
1709  Function *F = OpC->getCalledFunction();
1710  if (F && TLI->getLibFunc(F->getName(), Func) && TLI->has(Func) &&
1711  ((Func == LibFunc_atan && Callee->getName() == "tan") ||
1712  (Func == LibFunc_atanf && Callee->getName() == "tanf") ||
1713  (Func == LibFunc_atanl && Callee->getName() == "tanl")))
1714  Ret = OpC->getArgOperand(0);
1715  return Ret;
1716 }
1717 
1718 static bool isTrigLibCall(CallInst *CI) {
1719  // We can only hope to do anything useful if we can ignore things like errno
1720  // and floating-point exceptions.
1721  // We already checked the prototype.
1722  return CI->hasFnAttr(Attribute::NoUnwind) &&
1723  CI->hasFnAttr(Attribute::ReadNone);
1724 }
1725 
1726 static void insertSinCosCall(IRBuilder<> &B, Function *OrigCallee, Value *Arg,
1727  bool UseFloat, Value *&Sin, Value *&Cos,
1728  Value *&SinCos) {
1729  Type *ArgTy = Arg->getType();
1730  Type *ResTy;
1731  StringRef Name;
1732 
1733  Triple T(OrigCallee->getParent()->getTargetTriple());
1734  if (UseFloat) {
1735  Name = "__sincospif_stret";
1736 
1737  assert(T.getArch() != Triple::x86 && "x86 messy and unsupported for now");
1738  // x86_64 can't use {float, float} since that would be returned in both
1739  // xmm0 and xmm1, which isn't what a real struct would do.
1740  ResTy = T.getArch() == Triple::x86_64
1741  ? static_cast<Type *>(VectorType::get(ArgTy, 2))
1742  : static_cast<Type *>(StructType::get(ArgTy, ArgTy));
1743  } else {
1744  Name = "__sincospi_stret";
1745  ResTy = StructType::get(ArgTy, ArgTy);
1746  }
1747 
1748  Module *M = OrigCallee->getParent();
1750  M->getOrInsertFunction(Name, OrigCallee->getAttributes(), ResTy, ArgTy);
1751 
1752  if (Instruction *ArgInst = dyn_cast<Instruction>(Arg)) {
1753  // If the argument is an instruction, it must dominate all uses so put our
1754  // sincos call there.
1755  B.SetInsertPoint(ArgInst->getParent(), ++ArgInst->getIterator());
1756  } else {
1757  // Otherwise (e.g. for a constant) the beginning of the function is as
1758  // good a place as any.
1759  BasicBlock &EntryBB = B.GetInsertBlock()->getParent()->getEntryBlock();
1760  B.SetInsertPoint(&EntryBB, EntryBB.begin());
1761  }
1762 
1763  SinCos = B.CreateCall(Callee, Arg, "sincospi");
1764 
1765  if (SinCos->getType()->isStructTy()) {
1766  Sin = B.CreateExtractValue(SinCos, 0, "sinpi");
1767  Cos = B.CreateExtractValue(SinCos, 1, "cospi");
1768  } else {
1769  Sin = B.CreateExtractElement(SinCos, ConstantInt::get(B.getInt32Ty(), 0),
1770  "sinpi");
1771  Cos = B.CreateExtractElement(SinCos, ConstantInt::get(B.getInt32Ty(), 1),
1772  "cospi");
1773  }
1774 }
1775 
1776 Value *LibCallSimplifier::optimizeSinCosPi(CallInst *CI, IRBuilder<> &B) {
1777  // Make sure the prototype is as expected, otherwise the rest of the
1778  // function is probably invalid and likely to abort.
1779  if (!isTrigLibCall(CI))
1780  return nullptr;
1781 
1782  Value *Arg = CI->getArgOperand(0);
1783  SmallVector<CallInst *, 1> SinCalls;
1784  SmallVector<CallInst *, 1> CosCalls;
1785  SmallVector<CallInst *, 1> SinCosCalls;
1786 
1787  bool IsFloat = Arg->getType()->isFloatTy();
1788 
1789  // Look for all compatible sinpi, cospi and sincospi calls with the same
1790  // argument. If there are enough (in some sense) we can make the
1791  // substitution.
1792  Function *F = CI->getFunction();
1793  for (User *U : Arg->users())
1794  classifyArgUse(U, F, IsFloat, SinCalls, CosCalls, SinCosCalls);
1795 
1796  // It's only worthwhile if both sinpi and cospi are actually used.
1797  if (SinCosCalls.empty() && (SinCalls.empty() || CosCalls.empty()))
1798  return nullptr;
1799 
1800  Value *Sin, *Cos, *SinCos;
1801  insertSinCosCall(B, CI->getCalledFunction(), Arg, IsFloat, Sin, Cos, SinCos);
1802 
1803  auto replaceTrigInsts = [this](SmallVectorImpl<CallInst *> &Calls,
1804  Value *Res) {
1805  for (CallInst *C : Calls)
1806  replaceAllUsesWith(C, Res);
1807  };
1808 
1809  replaceTrigInsts(SinCalls, Sin);
1810  replaceTrigInsts(CosCalls, Cos);
1811  replaceTrigInsts(SinCosCalls, SinCos);
1812 
1813  return nullptr;
1814 }
1815 
1816 void LibCallSimplifier::classifyArgUse(
1817  Value *Val, Function *F, bool IsFloat,
1818  SmallVectorImpl<CallInst *> &SinCalls,
1819  SmallVectorImpl<CallInst *> &CosCalls,
1820  SmallVectorImpl<CallInst *> &SinCosCalls) {
1821  CallInst *CI = dyn_cast<CallInst>(Val);
1822 
1823  if (!CI)
1824  return;
1825 
1826  // Don't consider calls in other functions.
1827  if (CI->getFunction() != F)
1828  return;
1829 
1831  LibFunc Func;
1832  if (!Callee || !TLI->getLibFunc(*Callee, Func) || !TLI->has(Func) ||
1833  !isTrigLibCall(CI))
1834  return;
1835 
1836  if (IsFloat) {
1837  if (Func == LibFunc_sinpif)
1838  SinCalls.push_back(CI);
1839  else if (Func == LibFunc_cospif)
1840  CosCalls.push_back(CI);
1841  else if (Func == LibFunc_sincospif_stret)
1842  SinCosCalls.push_back(CI);
1843  } else {
1844  if (Func == LibFunc_sinpi)
1845  SinCalls.push_back(CI);
1846  else if (Func == LibFunc_cospi)
1847  CosCalls.push_back(CI);
1848  else if (Func == LibFunc_sincospi_stret)
1849  SinCosCalls.push_back(CI);
1850  }
1851 }
1852 
1853 //===----------------------------------------------------------------------===//
1854 // Integer Library Call Optimizations
1855 //===----------------------------------------------------------------------===//
1856 
1857 Value *LibCallSimplifier::optimizeFFS(CallInst *CI, IRBuilder<> &B) {
1858  // ffs(x) -> x != 0 ? (i32)llvm.cttz(x)+1 : 0
1859  Value *Op = CI->getArgOperand(0);
1860  Type *ArgType = Op->getType();
1862  Intrinsic::cttz, ArgType);
1863  Value *V = B.CreateCall(F, {Op, B.getTrue()}, "cttz");
1864  V = B.CreateAdd(V, ConstantInt::get(V->getType(), 1));
1865  V = B.CreateIntCast(V, B.getInt32Ty(), false);
1866 
1867  Value *Cond = B.CreateICmpNE(Op, Constant::getNullValue(ArgType));
1868  return B.CreateSelect(Cond, V, B.getInt32(0));
1869 }
1870 
1871 Value *LibCallSimplifier::optimizeFls(CallInst *CI, IRBuilder<> &B) {
1872  // fls(x) -> (i32)(sizeInBits(x) - llvm.ctlz(x, false))
1873  Value *Op = CI->getArgOperand(0);
1874  Type *ArgType = Op->getType();
1876  Intrinsic::ctlz, ArgType);
1877  Value *V = B.CreateCall(F, {Op, B.getFalse()}, "ctlz");
1878  V = B.CreateSub(ConstantInt::get(V->getType(), ArgType->getIntegerBitWidth()),
1879  V);
1880  return B.CreateIntCast(V, CI->getType(), false);
1881 }
1882 
1883 Value *LibCallSimplifier::optimizeAbs(CallInst *CI, IRBuilder<> &B) {
1884  // abs(x) -> x <s 0 ? -x : x
1885  // The negation has 'nsw' because abs of INT_MIN is undefined.
1886  Value *X = CI->getArgOperand(0);
1887  Value *IsNeg = B.CreateICmpSLT(X, Constant::getNullValue(X->getType()));
1888  Value *NegX = B.CreateNSWNeg(X, "neg");
1889  return B.CreateSelect(IsNeg, NegX, X);
1890 }
1891 
1892 Value *LibCallSimplifier::optimizeIsDigit(CallInst *CI, IRBuilder<> &B) {
1893  // isdigit(c) -> (c-'0') <u 10
1894  Value *Op = CI->getArgOperand(0);
1895  Op = B.CreateSub(Op, B.getInt32('0'), "isdigittmp");
1896  Op = B.CreateICmpULT(Op, B.getInt32(10), "isdigit");
1897  return B.CreateZExt(Op, CI->getType());
1898 }
1899 
1900 Value *LibCallSimplifier::optimizeIsAscii(CallInst *CI, IRBuilder<> &B) {
1901  // isascii(c) -> c <u 128
1902  Value *Op = CI->getArgOperand(0);
1903  Op = B.CreateICmpULT(Op, B.getInt32(128), "isascii");
1904  return B.CreateZExt(Op, CI->getType());
1905 }
1906 
1907 Value *LibCallSimplifier::optimizeToAscii(CallInst *CI, IRBuilder<> &B) {
1908  // toascii(c) -> c & 0x7f
1909  return B.CreateAnd(CI->getArgOperand(0),
1910  ConstantInt::get(CI->getType(), 0x7F));
1911 }
1912 
1913 Value *LibCallSimplifier::optimizeAtoi(CallInst *CI, IRBuilder<> &B) {
1914  StringRef Str;
1915  if (!getConstantStringInfo(CI->getArgOperand(0), Str))
1916  return nullptr;
1917 
1918  return convertStrToNumber(CI, Str, 10);
1919 }
1920 
1921 Value *LibCallSimplifier::optimizeStrtol(CallInst *CI, IRBuilder<> &B) {
1922  StringRef Str;
1923  if (!getConstantStringInfo(CI->getArgOperand(0), Str))
1924  return nullptr;
1925 
1926  if (!isa<ConstantPointerNull>(CI->getArgOperand(1)))
1927  return nullptr;
1928 
1929  if (ConstantInt *CInt = dyn_cast<ConstantInt>(CI->getArgOperand(2))) {
1930  return convertStrToNumber(CI, Str, CInt->getSExtValue());
1931  }
1932 
1933  return nullptr;
1934 }
1935 
1936 //===----------------------------------------------------------------------===//
1937 // Formatting and IO Library Call Optimizations
1938 //===----------------------------------------------------------------------===//
1939 
1940 static bool isReportingError(Function *Callee, CallInst *CI, int StreamArg);
1941 
1942 Value *LibCallSimplifier::optimizeErrorReporting(CallInst *CI, IRBuilder<> &B,
1943  int StreamArg) {
1944  Function *Callee = CI->getCalledFunction();
1945  // Error reporting calls should be cold, mark them as such.
1946  // This applies even to non-builtin calls: it is only a hint and applies to
1947  // functions that the frontend might not understand as builtins.
1948 
1949  // This heuristic was suggested in:
1950  // Improving Static Branch Prediction in a Compiler
1951  // Brian L. Deitrich, Ben-Chung Cheng, Wen-mei W. Hwu
1952  // Proceedings of PACT'98, Oct. 1998, IEEE
1953  if (!CI->hasFnAttr(Attribute::Cold) &&
1954  isReportingError(Callee, CI, StreamArg)) {
1956  }
1957 
1958  return nullptr;
1959 }
1960 
1961 static bool isReportingError(Function *Callee, CallInst *CI, int StreamArg) {
1962  if (!Callee || !Callee->isDeclaration())
1963  return false;
1964 
1965  if (StreamArg < 0)
1966  return true;
1967 
1968  // These functions might be considered cold, but only if their stream
1969  // argument is stderr.
1970 
1971  if (StreamArg >= (int)CI->getNumArgOperands())
1972  return false;
1973  LoadInst *LI = dyn_cast<LoadInst>(CI->getArgOperand(StreamArg));
1974  if (!LI)
1975  return false;
1977  if (!GV || !GV->isDeclaration())
1978  return false;
1979  return GV->getName() == "stderr";
1980 }
1981 
1982 Value *LibCallSimplifier::optimizePrintFString(CallInst *CI, IRBuilder<> &B) {
1983  // Check for a fixed format string.
1984  StringRef FormatStr;
1985  if (!getConstantStringInfo(CI->getArgOperand(0), FormatStr))
1986  return nullptr;
1987 
1988  // Empty format string -> noop.
1989  if (FormatStr.empty()) // Tolerate printf's declared void.
1990  return CI->use_empty() ? (Value *)CI : ConstantInt::get(CI->getType(), 0);
1991 
1992  // Do not do any of the following transformations if the printf return value
1993  // is used, in general the printf return value is not compatible with either
1994  // putchar() or puts().
1995  if (!CI->use_empty())
1996  return nullptr;
1997 
1998  // printf("x") -> putchar('x'), even for "%" and "%%".
1999  if (FormatStr.size() == 1 || FormatStr == "%%")
2000  return emitPutChar(B.getInt32(FormatStr[0]), B, TLI);
2001 
2002  // printf("%s", "a") --> putchar('a')
2003  if (FormatStr == "%s" && CI->getNumArgOperands() > 1) {
2004  StringRef ChrStr;
2005  if (!getConstantStringInfo(CI->getOperand(1), ChrStr))
2006  return nullptr;
2007  if (ChrStr.size() != 1)
2008  return nullptr;
2009  return emitPutChar(B.getInt32(ChrStr[0]), B, TLI);
2010  }
2011 
2012  // printf("foo\n") --> puts("foo")
2013  if (FormatStr[FormatStr.size() - 1] == '\n' &&
2014  FormatStr.find('%') == StringRef::npos) { // No format characters.
2015  // Create a string literal with no \n on it. We expect the constant merge
2016  // pass to be run after this pass, to merge duplicate strings.
2017  FormatStr = FormatStr.drop_back();
2018  Value *GV = B.CreateGlobalString(FormatStr, "str");
2019  return emitPutS(GV, B, TLI);
2020  }
2021 
2022  // Optimize specific format strings.
2023  // printf("%c", chr) --> putchar(chr)
2024  if (FormatStr == "%c" && CI->getNumArgOperands() > 1 &&
2025  CI->getArgOperand(1)->getType()->isIntegerTy())
2026  return emitPutChar(CI->getArgOperand(1), B, TLI);
2027 
2028  // printf("%s\n", str) --> puts(str)
2029  if (FormatStr == "%s\n" && CI->getNumArgOperands() > 1 &&
2030  CI->getArgOperand(1)->getType()->isPointerTy())
2031  return emitPutS(CI->getArgOperand(1), B, TLI);
2032  return nullptr;
2033 }
2034 
2035 Value *LibCallSimplifier::optimizePrintF(CallInst *CI, IRBuilder<> &B) {
2036 
2037  Function *Callee = CI->getCalledFunction();
2038  FunctionType *FT = Callee->getFunctionType();
2039  if (Value *V = optimizePrintFString(CI, B)) {
2040  return V;
2041  }
2042 
2043  // printf(format, ...) -> iprintf(format, ...) if no floating point
2044  // arguments.
2045  if (TLI->has(LibFunc_iprintf) && !callHasFloatingPointArgument(CI)) {
2046  Module *M = B.GetInsertBlock()->getParent()->getParent();
2047  FunctionCallee IPrintFFn =
2048  M->getOrInsertFunction("iprintf", FT, Callee->getAttributes());
2049  CallInst *New = cast<CallInst>(CI->clone());
2050  New->setCalledFunction(IPrintFFn);
2051  B.Insert(New);
2052  return New;
2053  }
2054  return nullptr;
2055 }
2056 
2057 Value *LibCallSimplifier::optimizeSPrintFString(CallInst *CI, IRBuilder<> &B) {
2058  // Check for a fixed format string.
2059  StringRef FormatStr;
2060  if (!getConstantStringInfo(CI->getArgOperand(1), FormatStr))
2061  return nullptr;
2062 
2063  // If we just have a format string (nothing else crazy) transform it.
2064  if (CI->getNumArgOperands() == 2) {
2065  // Make sure there's no % in the constant array. We could try to handle
2066  // %% -> % in the future if we cared.
2067  if (FormatStr.find('%') != StringRef::npos)
2068  return nullptr; // we found a format specifier, bail out.
2069 
2070  // sprintf(str, fmt) -> llvm.memcpy(align 1 str, align 1 fmt, strlen(fmt)+1)
2071  B.CreateMemCpy(CI->getArgOperand(0), 1, CI->getArgOperand(1), 1,
2072  ConstantInt::get(DL.getIntPtrType(CI->getContext()),
2073  FormatStr.size() + 1)); // Copy the null byte.
2074  return ConstantInt::get(CI->getType(), FormatStr.size());
2075  }
2076 
2077  // The remaining optimizations require the format string to be "%s" or "%c"
2078  // and have an extra operand.
2079  if (FormatStr.size() != 2 || FormatStr[0] != '%' ||
2080  CI->getNumArgOperands() < 3)
2081  return nullptr;
2082 
2083  // Decode the second character of the format string.
2084  if (FormatStr[1] == 'c') {
2085  // sprintf(dst, "%c", chr) --> *(i8*)dst = chr; *((i8*)dst+1) = 0
2086  if (!CI->getArgOperand(2)->getType()->isIntegerTy())
2087  return nullptr;
2088  Value *V = B.CreateTrunc(CI->getArgOperand(2), B.getInt8Ty(), "char");
2089  Value *Ptr = castToCStr(CI->getArgOperand(0), B);
2090  B.CreateStore(V, Ptr);
2091  Ptr = B.CreateGEP(B.getInt8Ty(), Ptr, B.getInt32(1), "nul");
2092  B.CreateStore(B.getInt8(0), Ptr);
2093 
2094  return ConstantInt::get(CI->getType(), 1);
2095  }
2096 
2097  if (FormatStr[1] == 's') {
2098  // sprintf(dest, "%s", str) -> llvm.memcpy(align 1 dest, align 1 str,
2099  // strlen(str)+1)
2100  if (!CI->getArgOperand(2)->getType()->isPointerTy())
2101  return nullptr;
2102 
2103  Value *Len = emitStrLen(CI->getArgOperand(2), B, DL, TLI);
2104  if (!Len)
2105  return nullptr;
2106  Value *IncLen =
2107  B.CreateAdd(Len, ConstantInt::get(Len->getType(), 1), "leninc");
2108  B.CreateMemCpy(CI->getArgOperand(0), 1, CI->getArgOperand(2), 1, IncLen);
2109 
2110  // The sprintf result is the unincremented number of bytes in the string.
2111  return B.CreateIntCast(Len, CI->getType(), false);
2112  }
2113  return nullptr;
2114 }
2115 
2116 Value *LibCallSimplifier::optimizeSPrintF(CallInst *CI, IRBuilder<> &B) {
2117  Function *Callee = CI->getCalledFunction();
2118  FunctionType *FT = Callee->getFunctionType();
2119  if (Value *V = optimizeSPrintFString(CI, B)) {
2120  return V;
2121  }
2122 
2123  // sprintf(str, format, ...) -> siprintf(str, format, ...) if no floating
2124  // point arguments.
2125  if (TLI->has(LibFunc_siprintf) && !callHasFloatingPointArgument(CI)) {
2126  Module *M = B.GetInsertBlock()->getParent()->getParent();
2127  FunctionCallee SIPrintFFn =
2128  M->getOrInsertFunction("siprintf", FT, Callee->getAttributes());
2129  CallInst *New = cast<CallInst>(CI->clone());
2130  New->setCalledFunction(SIPrintFFn);
2131  B.Insert(New);
2132  return New;
2133  }
2134  return nullptr;
2135 }
2136 
2137 Value *LibCallSimplifier::optimizeSnPrintFString(CallInst *CI, IRBuilder<> &B) {
2138  // Check for a fixed format string.
2139  StringRef FormatStr;
2140  if (!getConstantStringInfo(CI->getArgOperand(2), FormatStr))
2141  return nullptr;
2142 
2143  // Check for size
2145  if (!Size)
2146  return nullptr;
2147 
2148  uint64_t N = Size->getZExtValue();
2149 
2150  // If we just have a format string (nothing else crazy) transform it.
2151  if (CI->getNumArgOperands() == 3) {
2152  // Make sure there's no % in the constant array. We could try to handle
2153  // %% -> % in the future if we cared.
2154  if (FormatStr.find('%') != StringRef::npos)
2155  return nullptr; // we found a format specifier, bail out.
2156 
2157  if (N == 0)
2158  return ConstantInt::get(CI->getType(), FormatStr.size());
2159  else if (N < FormatStr.size() + 1)
2160  return nullptr;
2161 
2162  // snprintf(dst, size, fmt) -> llvm.memcpy(align 1 dst, align 1 fmt,
2163  // strlen(fmt)+1)
2164  B.CreateMemCpy(
2165  CI->getArgOperand(0), 1, CI->getArgOperand(2), 1,
2166  ConstantInt::get(DL.getIntPtrType(CI->getContext()),
2167  FormatStr.size() + 1)); // Copy the null byte.
2168  return ConstantInt::get(CI->getType(), FormatStr.size());
2169  }
2170 
2171  // The remaining optimizations require the format string to be "%s" or "%c"
2172  // and have an extra operand.
2173  if (FormatStr.size() == 2 && FormatStr[0] == '%' &&
2174  CI->getNumArgOperands() == 4) {
2175 
2176  // Decode the second character of the format string.
2177  if (FormatStr[1] == 'c') {
2178  if (N == 0)
2179  return ConstantInt::get(CI->getType(), 1);
2180  else if (N == 1)
2181  return nullptr;
2182 
2183  // snprintf(dst, size, "%c", chr) --> *(i8*)dst = chr; *((i8*)dst+1) = 0
2184  if (!CI->getArgOperand(3)->getType()->isIntegerTy())
2185  return nullptr;
2186  Value *V = B.CreateTrunc(CI->getArgOperand(3), B.getInt8Ty(), "char");
2187  Value *Ptr = castToCStr(CI->getArgOperand(0), B);
2188  B.CreateStore(V, Ptr);
2189  Ptr = B.CreateGEP(B.getInt8Ty(), Ptr, B.getInt32(1), "nul");
2190  B.CreateStore(B.getInt8(0), Ptr);
2191 
2192  return ConstantInt::get(CI->getType(), 1);
2193  }
2194 
2195  if (FormatStr[1] == 's') {
2196  // snprintf(dest, size, "%s", str) to llvm.memcpy(dest, str, len+1, 1)
2197  StringRef Str;
2198  if (!getConstantStringInfo(CI->getArgOperand(3), Str))
2199  return nullptr;
2200 
2201  if (N == 0)
2202  return ConstantInt::get(CI->getType(), Str.size());
2203  else if (N < Str.size() + 1)
2204  return nullptr;
2205 
2206  B.CreateMemCpy(CI->getArgOperand(0), 1, CI->getArgOperand(3), 1,
2207  ConstantInt::get(CI->getType(), Str.size() + 1));
2208 
2209  // The snprintf result is the unincremented number of bytes in the string.
2210  return ConstantInt::get(CI->getType(), Str.size());
2211  }
2212  }
2213  return nullptr;
2214 }
2215 
2216 Value *LibCallSimplifier::optimizeSnPrintF(CallInst *CI, IRBuilder<> &B) {
2217  if (Value *V = optimizeSnPrintFString(CI, B)) {
2218  return V;
2219  }
2220 
2221  return nullptr;
2222 }
2223 
2224 Value *LibCallSimplifier::optimizeFPrintFString(CallInst *CI, IRBuilder<> &B) {
2225  optimizeErrorReporting(CI, B, 0);
2226 
2227  // All the optimizations depend on the format string.
2228  StringRef FormatStr;
2229  if (!getConstantStringInfo(CI->getArgOperand(1), FormatStr))
2230  return nullptr;
2231 
2232  // Do not do any of the following transformations if the fprintf return
2233  // value is used, in general the fprintf return value is not compatible
2234  // with fwrite(), fputc() or fputs().
2235  if (!CI->use_empty())
2236  return nullptr;
2237 
2238  // fprintf(F, "foo") --> fwrite("foo", 3, 1, F)
2239  if (CI->getNumArgOperands() == 2) {
2240  // Could handle %% -> % if we cared.
2241  if (FormatStr.find('%') != StringRef::npos)
2242  return nullptr; // We found a format specifier.
2243 
2244  return emitFWrite(
2245  CI->getArgOperand(1),
2246  ConstantInt::get(DL.getIntPtrType(CI->getContext()), FormatStr.size()),
2247  CI->getArgOperand(0), B, DL, TLI);
2248  }
2249 
2250  // The remaining optimizations require the format string to be "%s" or "%c"
2251  // and have an extra operand.
2252  if (FormatStr.size() != 2 || FormatStr[0] != '%' ||
2253  CI->getNumArgOperands() < 3)
2254  return nullptr;
2255 
2256  // Decode the second character of the format string.
2257  if (FormatStr[1] == 'c') {
2258  // fprintf(F, "%c", chr) --> fputc(chr, F)
2259  if (!CI->getArgOperand(2)->getType()->isIntegerTy())
2260  return nullptr;
2261  return emitFPutC(CI->getArgOperand(2), CI->getArgOperand(0), B, TLI);
2262  }
2263 
2264  if (FormatStr[1] == 's') {
2265  // fprintf(F, "%s", str) --> fputs(str, F)
2266  if (!CI->getArgOperand(2)->getType()->isPointerTy())
2267  return nullptr;
2268  return emitFPutS(CI->getArgOperand(2), CI->getArgOperand(0), B, TLI);
2269  }
2270  return nullptr;
2271 }
2272 
2273 Value *LibCallSimplifier::optimizeFPrintF(CallInst *CI, IRBuilder<> &B) {
2274  Function *Callee = CI->getCalledFunction();
2275  FunctionType *FT = Callee->getFunctionType();
2276  if (Value *V = optimizeFPrintFString(CI, B)) {
2277  return V;
2278  }
2279 
2280  // fprintf(stream, format, ...) -> fiprintf(stream, format, ...) if no
2281  // floating point arguments.
2282  if (TLI->has(LibFunc_fiprintf) && !callHasFloatingPointArgument(CI)) {
2283  Module *M = B.GetInsertBlock()->getParent()->getParent();
2284  FunctionCallee FIPrintFFn =
2285  M->getOrInsertFunction("fiprintf", FT, Callee->getAttributes());
2286  CallInst *New = cast<CallInst>(CI->clone());
2287  New->setCalledFunction(FIPrintFFn);
2288  B.Insert(New);
2289  return New;
2290  }
2291  return nullptr;
2292 }
2293 
2294 Value *LibCallSimplifier::optimizeFWrite(CallInst *CI, IRBuilder<> &B) {
2295  optimizeErrorReporting(CI, B, 3);
2296 
2297  // Get the element size and count.
2298  ConstantInt *SizeC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
2299  ConstantInt *CountC = dyn_cast<ConstantInt>(CI->getArgOperand(2));
2300  if (SizeC && CountC) {
2301  uint64_t Bytes = SizeC->getZExtValue() * CountC->getZExtValue();
2302 
2303  // If this is writing zero records, remove the call (it's a noop).
2304  if (Bytes == 0)
2305  return ConstantInt::get(CI->getType(), 0);
2306 
2307  // If this is writing one byte, turn it into fputc.
2308  // This optimisation is only valid, if the return value is unused.
2309  if (Bytes == 1 && CI->use_empty()) { // fwrite(S,1,1,F) -> fputc(S[0],F)
2310  Value *Char = B.CreateLoad(B.getInt8Ty(),
2311  castToCStr(CI->getArgOperand(0), B), "char");
2312  Value *NewCI = emitFPutC(Char, CI->getArgOperand(3), B, TLI);
2313  return NewCI ? ConstantInt::get(CI->getType(), 1) : nullptr;
2314  }
2315  }
2316 
2317  if (isLocallyOpenedFile(CI->getArgOperand(3), CI, B, TLI))
2318  return emitFWriteUnlocked(CI->getArgOperand(0), CI->getArgOperand(1),
2319  CI->getArgOperand(2), CI->getArgOperand(3), B, DL,
2320  TLI);
2321 
2322  return nullptr;
2323 }
2324 
2325 Value *LibCallSimplifier::optimizeFPuts(CallInst *CI, IRBuilder<> &B) {
2326  optimizeErrorReporting(CI, B, 1);
2327 
2328  // Don't rewrite fputs to fwrite when optimising for size because fwrite
2329  // requires more arguments and thus extra MOVs are required.
2330  if (CI->getFunction()->optForSize())
2331  return nullptr;
2332 
2333  // Check if has any use
2334  if (!CI->use_empty()) {
2335  if (isLocallyOpenedFile(CI->getArgOperand(1), CI, B, TLI))
2336  return emitFPutSUnlocked(CI->getArgOperand(0), CI->getArgOperand(1), B,
2337  TLI);
2338  else
2339  // We can't optimize if return value is used.
2340  return nullptr;
2341  }
2342 
2343  // fputs(s,F) --> fwrite(s,strlen(s),1,F)
2344  uint64_t Len = GetStringLength(CI->getArgOperand(0));
2345  if (!Len)
2346  return nullptr;
2347 
2348  // Known to have no uses (see above).
2349  return emitFWrite(
2350  CI->getArgOperand(0),
2351  ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len - 1),
2352  CI->getArgOperand(1), B, DL, TLI);
2353 }
2354 
2355 Value *LibCallSimplifier::optimizeFPutc(CallInst *CI, IRBuilder<> &B) {
2356  optimizeErrorReporting(CI, B, 1);
2357 
2358  if (isLocallyOpenedFile(CI->getArgOperand(1), CI, B, TLI))
2359  return emitFPutCUnlocked(CI->getArgOperand(0), CI->getArgOperand(1), B,
2360  TLI);
2361 
2362  return nullptr;
2363 }
2364 
2365 Value *LibCallSimplifier::optimizeFGetc(CallInst *CI, IRBuilder<> &B) {
2366  if (isLocallyOpenedFile(CI->getArgOperand(0), CI, B, TLI))
2367  return emitFGetCUnlocked(CI->getArgOperand(0), B, TLI);
2368 
2369  return nullptr;
2370 }
2371 
2372 Value *LibCallSimplifier::optimizeFGets(CallInst *CI, IRBuilder<> &B) {
2373  if (isLocallyOpenedFile(CI->getArgOperand(2), CI, B, TLI))
2374  return emitFGetSUnlocked(CI->getArgOperand(0), CI->getArgOperand(1),
2375  CI->getArgOperand(2), B, TLI);
2376 
2377  return nullptr;
2378 }
2379 
2380 Value *LibCallSimplifier::optimizeFRead(CallInst *CI, IRBuilder<> &B) {
2381  if (isLocallyOpenedFile(CI->getArgOperand(3), CI, B, TLI))
2382  return emitFReadUnlocked(CI->getArgOperand(0), CI->getArgOperand(1),
2383  CI->getArgOperand(2), CI->getArgOperand(3), B, DL,
2384  TLI);
2385 
2386  return nullptr;
2387 }
2388 
2389 Value *LibCallSimplifier::optimizePuts(CallInst *CI, IRBuilder<> &B) {
2390  if (!CI->use_empty())
2391  return nullptr;
2392 
2393  // Check for a constant string.
2394  // puts("") -> putchar('\n')
2395  StringRef Str;
2396  if (getConstantStringInfo(CI->getArgOperand(0), Str) && Str.empty())
2397  return emitPutChar(B.getInt32('\n'), B, TLI);
2398 
2399  return nullptr;
2400 }
2401 
2402 bool LibCallSimplifier::hasFloatVersion(StringRef FuncName) {
2403  LibFunc Func;
2404  SmallString<20> FloatFuncName = FuncName;
2405  FloatFuncName += 'f';
2406  if (TLI->getLibFunc(FloatFuncName, Func))
2407  return TLI->has(Func);
2408  return false;
2409 }
2410 
2411 Value *LibCallSimplifier::optimizeStringMemoryLibCall(CallInst *CI,
2412  IRBuilder<> &Builder) {
2413  LibFunc Func;
2414  Function *Callee = CI->getCalledFunction();
2415  // Check for string/memory library functions.
2416  if (TLI->getLibFunc(*Callee, Func) && TLI->has(Func)) {
2417  // Make sure we never change the calling convention.
2418  assert((ignoreCallingConv(Func) ||
2419  isCallingConvCCompatible(CI)) &&
2420  "Optimizing string/memory libcall would change the calling convention");
2421  switch (Func) {
2422  case LibFunc_strcat:
2423  return optimizeStrCat(CI, Builder);
2424  case LibFunc_strncat:
2425  return optimizeStrNCat(CI, Builder);
2426  case LibFunc_strchr:
2427  return optimizeStrChr(CI, Builder);
2428  case LibFunc_strrchr:
2429  return optimizeStrRChr(CI, Builder);
2430  case LibFunc_strcmp:
2431  return optimizeStrCmp(CI, Builder);
2432  case LibFunc_strncmp:
2433  return optimizeStrNCmp(CI, Builder);
2434  case LibFunc_strcpy:
2435  return optimizeStrCpy(CI, Builder);
2436  case LibFunc_stpcpy:
2437  return optimizeStpCpy(CI, Builder);
2438  case LibFunc_strncpy:
2439  return optimizeStrNCpy(CI, Builder);
2440  case LibFunc_strlen:
2441  return optimizeStrLen(CI, Builder);
2442  case LibFunc_strpbrk:
2443  return optimizeStrPBrk(CI, Builder);
2444  case LibFunc_strtol:
2445  case LibFunc_strtod:
2446  case LibFunc_strtof:
2447  case LibFunc_strtoul:
2448  case LibFunc_strtoll:
2449  case LibFunc_strtold:
2450  case LibFunc_strtoull:
2451  return optimizeStrTo(CI, Builder);
2452  case LibFunc_strspn:
2453  return optimizeStrSpn(CI, Builder);
2454  case LibFunc_strcspn:
2455  return optimizeStrCSpn(CI, Builder);
2456  case LibFunc_strstr:
2457  return optimizeStrStr(CI, Builder);
2458  case LibFunc_memchr:
2459  return optimizeMemChr(CI, Builder);
2460  case LibFunc_memcmp:
2461  return optimizeMemCmp(CI, Builder);
2462  case LibFunc_memcpy:
2463  return optimizeMemCpy(CI, Builder);
2464  case LibFunc_memmove:
2465  return optimizeMemMove(CI, Builder);
2466  case LibFunc_memset:
2467  return optimizeMemSet(CI, Builder);
2468  case LibFunc_realloc:
2469  return optimizeRealloc(CI, Builder);
2470  case LibFunc_wcslen:
2471  return optimizeWcslen(CI, Builder);
2472  default:
2473  break;
2474  }
2475  }
2476  return nullptr;
2477 }
2478 
2479 Value *LibCallSimplifier::optimizeFloatingPointLibCall(CallInst *CI,
2480  LibFunc Func,
2481  IRBuilder<> &Builder) {
2482  // Don't optimize calls that require strict floating point semantics.
2483  if (CI->isStrictFP())
2484  return nullptr;
2485 
2486  if (Value *V = optimizeTrigReflections(CI, Func, Builder))
2487  return V;
2488 
2489  switch (Func) {
2490  case LibFunc_sinpif:
2491  case LibFunc_sinpi:
2492  case LibFunc_cospif:
2493  case LibFunc_cospi:
2494  return optimizeSinCosPi(CI, Builder);
2495  case LibFunc_powf:
2496  case LibFunc_pow:
2497  case LibFunc_powl:
2498  return optimizePow(CI, Builder);
2499  case LibFunc_exp2l:
2500  case LibFunc_exp2:
2501  case LibFunc_exp2f:
2502  return optimizeExp2(CI, Builder);
2503  case LibFunc_fabsf:
2504  case LibFunc_fabs:
2505  case LibFunc_fabsl:
2506  return replaceUnaryCall(CI, Builder, Intrinsic::fabs);
2507  case LibFunc_sqrtf:
2508  case LibFunc_sqrt:
2509  case LibFunc_sqrtl:
2510  return optimizeSqrt(CI, Builder);
2511  case LibFunc_log:
2512  case LibFunc_log10:
2513  case LibFunc_log1p:
2514  case LibFunc_log2:
2515  case LibFunc_logb:
2516  return optimizeLog(CI, Builder);
2517  case LibFunc_tan:
2518  case LibFunc_tanf:
2519  case LibFunc_tanl:
2520  return optimizeTan(CI, Builder);
2521  case LibFunc_ceil:
2522  return replaceUnaryCall(CI, Builder, Intrinsic::ceil);
2523  case LibFunc_floor:
2524  return replaceUnaryCall(CI, Builder, Intrinsic::floor);
2525  case LibFunc_round:
2526  return replaceUnaryCall(CI, Builder, Intrinsic::round);
2527  case LibFunc_nearbyint:
2528  return replaceUnaryCall(CI, Builder, Intrinsic::nearbyint);
2529  case LibFunc_rint:
2530  return replaceUnaryCall(CI, Builder, Intrinsic::rint);
2531  case LibFunc_trunc:
2532  return replaceUnaryCall(CI, Builder, Intrinsic::trunc);
2533  case LibFunc_acos:
2534  case LibFunc_acosh:
2535  case LibFunc_asin:
2536  case LibFunc_asinh:
2537  case LibFunc_atan:
2538  case LibFunc_atanh:
2539  case LibFunc_cbrt:
2540  case LibFunc_cosh:
2541  case LibFunc_exp:
2542  case LibFunc_exp10:
2543  case LibFunc_expm1:
2544  case LibFunc_cos:
2545  case LibFunc_sin:
2546  case LibFunc_sinh:
2547  case LibFunc_tanh:
2548  if (UnsafeFPShrink && hasFloatVersion(CI->getCalledFunction()->getName()))
2549  return optimizeUnaryDoubleFP(CI, Builder, true);
2550  return nullptr;
2551  case LibFunc_copysign:
2552  if (hasFloatVersion(CI->getCalledFunction()->getName()))
2553  return optimizeBinaryDoubleFP(CI, Builder);
2554  return nullptr;
2555  case LibFunc_fminf:
2556  case LibFunc_fmin:
2557  case LibFunc_fminl:
2558  case LibFunc_fmaxf:
2559  case LibFunc_fmax:
2560  case LibFunc_fmaxl:
2561  return optimizeFMinFMax(CI, Builder);
2562  case LibFunc_cabs:
2563  case LibFunc_cabsf:
2564  case LibFunc_cabsl:
2565  return optimizeCAbs(CI, Builder);
2566  default:
2567  return nullptr;
2568  }
2569 }
2570 
2572  // TODO: Split out the code below that operates on FP calls so that
2573  // we can all non-FP calls with the StrictFP attribute to be
2574  // optimized.
2575  if (CI->isNoBuiltin())
2576  return nullptr;
2577 
2578  LibFunc Func;
2579  Function *Callee = CI->getCalledFunction();
2580 
2582  CI->getOperandBundlesAsDefs(OpBundles);
2583  IRBuilder<> Builder(CI, /*FPMathTag=*/nullptr, OpBundles);
2584  bool isCallingConvC = isCallingConvCCompatible(CI);
2585 
2586  // Command-line parameter overrides instruction attribute.
2587  // This can't be moved to optimizeFloatingPointLibCall() because it may be
2588  // used by the intrinsic optimizations.
2589  if (EnableUnsafeFPShrink.getNumOccurrences() > 0)
2590  UnsafeFPShrink = EnableUnsafeFPShrink;
2591  else if (isa<FPMathOperator>(CI) && CI->isFast())
2592  UnsafeFPShrink = true;
2593 
2594  // First, check for intrinsics.
2595  if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
2596  if (!isCallingConvC)
2597  return nullptr;
2598  // The FP intrinsics have corresponding constrained versions so we don't
2599  // need to check for the StrictFP attribute here.
2600  switch (II->getIntrinsicID()) {
2601  case Intrinsic::pow:
2602  return optimizePow(CI, Builder);
2603  case Intrinsic::exp2:
2604  return optimizeExp2(CI, Builder);
2605  case Intrinsic::log:
2606  return optimizeLog(CI, Builder);
2607  case Intrinsic::sqrt:
2608  return optimizeSqrt(CI, Builder);
2609  // TODO: Use foldMallocMemset() with memset intrinsic.
2610  default:
2611  return nullptr;
2612  }
2613  }
2614 
2615  // Also try to simplify calls to fortified library functions.
2616  if (Value *SimplifiedFortifiedCI = FortifiedSimplifier.optimizeCall(CI)) {
2617  // Try to further simplify the result.
2618  CallInst *SimplifiedCI = dyn_cast<CallInst>(SimplifiedFortifiedCI);
2619  if (SimplifiedCI && SimplifiedCI->getCalledFunction()) {
2620  // Use an IR Builder from SimplifiedCI if available instead of CI
2621  // to guarantee we reach all uses we might replace later on.
2622  IRBuilder<> TmpBuilder(SimplifiedCI);
2623  if (Value *V = optimizeStringMemoryLibCall(SimplifiedCI, TmpBuilder)) {
2624  // If we were able to further simplify, remove the now redundant call.
2625  SimplifiedCI->replaceAllUsesWith(V);
2626  eraseFromParent(SimplifiedCI);
2627  return V;
2628  }
2629  }
2630  return SimplifiedFortifiedCI;
2631  }
2632 
2633  // Then check for known library functions.
2634  if (TLI->getLibFunc(*Callee, Func) && TLI->has(Func)) {
2635  // We never change the calling convention.
2636  if (!ignoreCallingConv(Func) && !isCallingConvC)
2637  return nullptr;
2638  if (Value *V = optimizeStringMemoryLibCall(CI, Builder))
2639  return V;
2640  if (Value *V = optimizeFloatingPointLibCall(CI, Func, Builder))
2641  return V;
2642  switch (Func) {
2643  case LibFunc_ffs:
2644  case LibFunc_ffsl:
2645  case LibFunc_ffsll:
2646  return optimizeFFS(CI, Builder);
2647  case LibFunc_fls:
2648  case LibFunc_flsl:
2649  case LibFunc_flsll:
2650  return optimizeFls(CI, Builder);
2651  case LibFunc_abs:
2652  case LibFunc_labs:
2653  case LibFunc_llabs:
2654  return optimizeAbs(CI, Builder);
2655  case LibFunc_isdigit:
2656  return optimizeIsDigit(CI, Builder);
2657  case LibFunc_isascii:
2658  return optimizeIsAscii(CI, Builder);
2659  case LibFunc_toascii:
2660  return optimizeToAscii(CI, Builder);
2661  case LibFunc_atoi:
2662  case LibFunc_atol:
2663  case LibFunc_atoll:
2664  return optimizeAtoi(CI, Builder);
2665  case LibFunc_strtol:
2666  case LibFunc_strtoll:
2667  return optimizeStrtol(CI, Builder);
2668  case LibFunc_printf:
2669  return optimizePrintF(CI, Builder);
2670  case LibFunc_sprintf:
2671  return optimizeSPrintF(CI, Builder);
2672  case LibFunc_snprintf:
2673  return optimizeSnPrintF(CI, Builder);
2674  case LibFunc_fprintf:
2675  return optimizeFPrintF(CI, Builder);
2676  case LibFunc_fwrite:
2677  return optimizeFWrite(CI, Builder);
2678  case LibFunc_fread:
2679  return optimizeFRead(CI, Builder);
2680  case LibFunc_fputs:
2681  return optimizeFPuts(CI, Builder);
2682  case LibFunc_fgets:
2683  return optimizeFGets(CI, Builder);
2684  case LibFunc_fputc:
2685  return optimizeFPutc(CI, Builder);
2686  case LibFunc_fgetc:
2687  return optimizeFGetc(CI, Builder);
2688  case LibFunc_puts:
2689  return optimizePuts(CI, Builder);
2690  case LibFunc_perror:
2691  return optimizeErrorReporting(CI, Builder);
2692  case LibFunc_vfprintf:
2693  case LibFunc_fiprintf:
2694  return optimizeErrorReporting(CI, Builder, 0);
2695  default:
2696  return nullptr;
2697  }
2698  }
2699  return nullptr;
2700 }
2701 
2703  const DataLayout &DL, const TargetLibraryInfo *TLI,
2705  function_ref<void(Instruction *, Value *)> Replacer,
2706  function_ref<void(Instruction *)> Eraser)
2707  : FortifiedSimplifier(TLI), DL(DL), TLI(TLI), ORE(ORE),
2708  UnsafeFPShrink(false), Replacer(Replacer), Eraser(Eraser) {}
2709 
2710 void LibCallSimplifier::replaceAllUsesWith(Instruction *I, Value *With) {
2711  // Indirect through the replacer used in this instance.
2712  Replacer(I, With);
2713 }
2714 
2715 void LibCallSimplifier::eraseFromParent(Instruction *I) {
2716  Eraser(I);
2717 }
2718 
2719 // TODO:
2720 // Additional cases that we need to add to this file:
2721 //
2722 // cbrt:
2723 // * cbrt(expN(X)) -> expN(x/3)
2724 // * cbrt(sqrt(x)) -> pow(x,1/6)
2725 // * cbrt(cbrt(x)) -> pow(x,1/9)
2726 //
2727 // exp, expf, expl:
2728 // * exp(log(x)) -> x
2729 //
2730 // log, logf, logl:
2731 // * log(exp(x)) -> x
2732 // * log(exp(y)) -> y*log(e)
2733 // * log(exp10(y)) -> y*log(10)
2734 // * log(sqrt(x)) -> 0.5*log(x)
2735 //
2736 // pow, powf, powl:
2737 // * pow(sqrt(x),y) -> pow(x,y*0.5)
2738 // * pow(pow(x,y),z)-> pow(x,y*z)
2739 //
2740 // signbit:
2741 // * signbit(cnst) -> cnst'
2742 // * signbit(nncst) -> 0 (if pstv is a non-negative constant)
2743 //
2744 // sqrt, sqrtf, sqrtl:
2745 // * sqrt(expN(x)) -> expN(x*0.5)
2746 // * sqrt(Nroot(x)) -> pow(x,1/(2*N))
2747 // * sqrt(pow(x,y)) -> pow(|x|,y*0.5)
2748 //
2749 
2750 //===----------------------------------------------------------------------===//
2751 // Fortified Library Call Optimizations
2752 //===----------------------------------------------------------------------===//
2753 
2754 bool FortifiedLibCallSimplifier::isFortifiedCallFoldable(CallInst *CI,
2755  unsigned ObjSizeOp,
2756  unsigned SizeOp,
2757  bool isString) {
2758  if (CI->getArgOperand(ObjSizeOp) == CI->getArgOperand(SizeOp))
2759  return true;
2760  if (ConstantInt *ObjSizeCI =
2761  dyn_cast<ConstantInt>(CI->getArgOperand(ObjSizeOp))) {
2762  if (ObjSizeCI->isMinusOne())
2763  return true;
2764  // If the object size wasn't -1 (unknown), bail out if we were asked to.
2765  if (OnlyLowerUnknownSize)
2766  return false;
2767  if (isString) {
2768  uint64_t Len = GetStringLength(CI->getArgOperand(SizeOp));
2769  // If the length is 0 we don't know how long it is and so we can't
2770  // remove the check.
2771  if (Len == 0)
2772  return false;
2773  return ObjSizeCI->getZExtValue() >= Len;
2774  }
2775  if (ConstantInt *SizeCI = dyn_cast<ConstantInt>(CI->getArgOperand(SizeOp)))
2776  return ObjSizeCI->getZExtValue() >= SizeCI->getZExtValue();
2777  }
2778  return false;
2779 }
2780 
2781 Value *FortifiedLibCallSimplifier::optimizeMemCpyChk(CallInst *CI,
2782  IRBuilder<> &B) {
2783  if (isFortifiedCallFoldable(CI, 3, 2, false)) {
2784  B.CreateMemCpy(CI->getArgOperand(0), 1, CI->getArgOperand(1), 1,
2785  CI->getArgOperand(2));
2786  return CI->getArgOperand(0);
2787  }
2788  return nullptr;
2789 }
2790 
2791 Value *FortifiedLibCallSimplifier::optimizeMemMoveChk(CallInst *CI,
2792  IRBuilder<> &B) {
2793  if (isFortifiedCallFoldable(CI, 3, 2, false)) {
2794  B.CreateMemMove(CI->getArgOperand(0), 1, CI->getArgOperand(1), 1,
2795  CI->getArgOperand(2));
2796  return CI->getArgOperand(0);
2797  }
2798  return nullptr;
2799 }
2800 
2801 Value *FortifiedLibCallSimplifier::optimizeMemSetChk(CallInst *CI,
2802  IRBuilder<> &B) {
2803  // TODO: Try foldMallocMemset() here.
2804 
2805  if (isFortifiedCallFoldable(CI, 3, 2, false)) {
2806  Value *Val = B.CreateIntCast(CI->getArgOperand(1), B.getInt8Ty(), false);
2807  B.CreateMemSet(CI->getArgOperand(0), Val, CI->getArgOperand(2), 1);
2808  return CI->getArgOperand(0);
2809  }
2810  return nullptr;
2811 }
2812 
2813 Value *FortifiedLibCallSimplifier::optimizeStrpCpyChk(CallInst *CI,
2814  IRBuilder<> &B,
2815  LibFunc Func) {
2816  Function *Callee = CI->getCalledFunction();
2817  StringRef Name = Callee->getName();
2818  const DataLayout &DL = CI->getModule()->getDataLayout();
2819  Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1),
2820  *ObjSize = CI->getArgOperand(2);
2821 
2822  // __stpcpy_chk(x,x,...) -> x+strlen(x)
2823  if (Func == LibFunc_stpcpy_chk && !OnlyLowerUnknownSize && Dst == Src) {
2824  Value *StrLen = emitStrLen(Src, B, DL, TLI);
2825  return StrLen ? B.CreateInBoundsGEP(B.getInt8Ty(), Dst, StrLen) : nullptr;
2826  }
2827 
2828  // If a) we don't have any length information, or b) we know this will
2829  // fit then just lower to a plain st[rp]cpy. Otherwise we'll keep our
2830  // st[rp]cpy_chk call which may fail at runtime if the size is too long.
2831  // TODO: It might be nice to get a maximum length out of the possible
2832  // string lengths for varying.
2833  if (isFortifiedCallFoldable(CI, 2, 1, true))
2834  return emitStrCpy(Dst, Src, B, TLI, Name.substr(2, 6));
2835 
2836  if (OnlyLowerUnknownSize)
2837  return nullptr;
2838 
2839  // Maybe we can stil fold __st[rp]cpy_chk to __memcpy_chk.
2840  uint64_t Len = GetStringLength(Src);
2841  if (Len == 0)
2842  return nullptr;
2843 
2844  Type *SizeTTy = DL.getIntPtrType(CI->getContext());
2845  Value *LenV = ConstantInt::get(SizeTTy, Len);
2846  Value *Ret = emitMemCpyChk(Dst, Src, LenV, ObjSize, B, DL, TLI);
2847  // If the function was an __stpcpy_chk, and we were able to fold it into
2848  // a __memcpy_chk, we still need to return the correct end pointer.
2849  if (Ret && Func == LibFunc_stpcpy_chk)
2850  return B.CreateGEP(B.getInt8Ty(), Dst, ConstantInt::get(SizeTTy, Len - 1));
2851  return Ret;
2852 }
2853 
2854 Value *FortifiedLibCallSimplifier::optimizeStrpNCpyChk(CallInst *CI,
2855  IRBuilder<> &B,
2856  LibFunc Func) {
2857  Function *Callee = CI->getCalledFunction();
2858  StringRef Name = Callee->getName();
2859  if (isFortifiedCallFoldable(CI, 3, 2, false)) {
2860  Value *Ret = emitStrNCpy(CI->getArgOperand(0), CI->getArgOperand(1),
2861  CI->getArgOperand(2), B, TLI, Name.substr(2, 7));
2862  return Ret;
2863  }
2864  return nullptr;
2865 }
2866 
2868  // FIXME: We shouldn't be changing "nobuiltin" or TLI unavailable calls here.
2869  // Some clang users checked for _chk libcall availability using:
2870  // __has_builtin(__builtin___memcpy_chk)
2871  // When compiling with -fno-builtin, this is always true.
2872  // When passing -ffreestanding/-mkernel, which both imply -fno-builtin, we
2873  // end up with fortified libcalls, which isn't acceptable in a freestanding
2874  // environment which only provides their non-fortified counterparts.
2875  //
2876  // Until we change clang and/or teach external users to check for availability
2877  // differently, disregard the "nobuiltin" attribute and TLI::has.
2878  //
2879  // PR23093.
2880 
2881  LibFunc Func;
2882  Function *Callee = CI->getCalledFunction();
2883 
2885  CI->getOperandBundlesAsDefs(OpBundles);
2886  IRBuilder<> Builder(CI, /*FPMathTag=*/nullptr, OpBundles);
2887  bool isCallingConvC = isCallingConvCCompatible(CI);
2888 
2889  // First, check that this is a known library functions and that the prototype
2890  // is correct.
2891  if (!TLI->getLibFunc(*Callee, Func))
2892  return nullptr;
2893 
2894  // We never change the calling convention.
2895  if (!ignoreCallingConv(Func) && !isCallingConvC)
2896  return nullptr;
2897 
2898  switch (Func) {
2899  case LibFunc_memcpy_chk:
2900  return optimizeMemCpyChk(CI, Builder);
2901  case LibFunc_memmove_chk:
2902  return optimizeMemMoveChk(CI, Builder);
2903  case LibFunc_memset_chk:
2904  return optimizeMemSetChk(CI, Builder);
2905  case LibFunc_stpcpy_chk:
2906  case LibFunc_strcpy_chk:
2907  return optimizeStrpCpyChk(CI, Builder, Func);
2908  case LibFunc_stpncpy_chk:
2909  case LibFunc_strncpy_chk:
2910  return optimizeStrpNCpyChk(CI, Builder, Func);
2911  default:
2912  break;
2913  }
2914  return nullptr;
2915 }
2916 
2918  const TargetLibraryInfo *TLI, bool OnlyLowerUnknownSize)
2919  : TLI(TLI), OnlyLowerUnknownSize(OnlyLowerUnknownSize) {}
Value * CreateNSWNeg(Value *V, const Twine &Name="")
Definition: IRBuilder.h:1353
Value * CreateInBoundsGEP(Value *Ptr, ArrayRef< Value *> IdxList, const Twine &Name="")
Definition: IRBuilder.h:1512
uint64_t CallInst * C
void computeKnownBits(const Value *V, KnownBits &Known, const DataLayout &DL, unsigned Depth=0, AssumptionCache *AC=nullptr, const Instruction *CxtI=nullptr, const DominatorTree *DT=nullptr, OptimizationRemarkEmitter *ORE=nullptr, bool UseInstrInfo=true)
Determine which bits of V are known to be either zero or one and return them in the KnownZero/KnownOn...
bool isIntrinsic() const
isIntrinsic - Returns true if the function&#39;s name starts with "llvm.".
Definition: Function.h:198
bool hasNoInfs() const
Determine whether the no-infs flag is set.
A parsed version of the target data layout string in and methods for querying it. ...
Definition: DataLayout.h:110
const std::string & getTargetTriple() const
Get the target triple which is a string describing the target host.
Definition: Module.h:240
class_match< Value > m_Value()
Match an arbitrary value and ignore it.
Definition: PatternMatch.h:70
Value * CreateICmp(CmpInst::Predicate P, Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1984
static GCMetadataPrinterRegistry::Add< ErlangGCPrinter > X("erlang", "erlang-compatible garbage collector")
bool hasNoSignedZeros() const
Determine whether the no-signed-zeros flag is set.
void flipAllBits()
Toggle every bit to its opposite value.
Definition: APInt.h:1476
bool isOnlyUsedInZeroEqualityComparison(const Instruction *CxtI)
GCNRegPressure max(const GCNRegPressure &P1, const GCNRegPressure &P2)
LLVM_NODISCARD std::string str() const
str - Get the contents as an std::string.
Definition: StringRef.h:218
bool doesNotAccessMemory(unsigned OpNo) const
Definition: InstrTypes.h:1465
void setFast(bool B=true)
Definition: Operator.h:236
Value * CreateICmpNE(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1878
Value * CreateIsNotNull(Value *Arg, const Twine &Name="")
Return an i1 value testing if Arg is not null.
Definition: IRBuilder.h:2152
Value * CreateZExtOrTrunc(Value *V, Type *DestTy, const Twine &Name="")
Create a ZExt or Trunc from the integer value V to DestTy.
Definition: IRBuilder.h:1704
This class represents lattice values for constants.
Definition: AllocatorList.h:23
static bool callHasFloatingPointArgument(const CallInst *CI)
Type * getParamType(unsigned i) const
Parameter type accessors.
Definition: DerivedTypes.h:134
Value * CreateICmpULT(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1890
LoadInst * CreateLoad(Type *Ty, Value *Ptr, const char *Name)
Provided to resolve &#39;CreateLoad(Ty, Ptr, "...")&#39; correctly, instead of converting the string to &#39;bool...
Definition: IRBuilder.h:1392
Value * optimizeCall(CallInst *CI)
Take the given call instruction and return a more optimal value to replace the instruction with or 0 ...
A Module instance is used to store all the information related to an LLVM module. ...
Definition: Module.h:65
static Constant * getInfinity(Type *Ty, bool Negative=false)
Definition: Constants.cpp:817
amdgpu Simplify well known AMD library false FunctionCallee Value const Twine & Name
A handy container for a FunctionType+Callee-pointer pair, which can be passed around as a single enti...
Definition: DerivedTypes.h:164
LLVM_NODISCARD bool startswith(StringRef Prefix) const
Check if this string starts with the given Prefix.
Definition: StringRef.h:256
LLVM_NODISCARD size_t rfind(char C, size_t From=npos) const
Search for the last character C in the string.
Definition: StringRef.h:345
void push_back(const T &Elt)
Definition: SmallVector.h:211
LLVM_NODISCARD int compare(StringRef RHS) const
compare - Compare two strings; the result is -1, 0, or 1 if this string is lexicographically less tha...
Definition: StringRef.h:174
Value * CreateICmpSLT(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1906
This class represents a function call, abstracting a target machine&#39;s calling convention.
static uint64_t round(uint64_t Acc, uint64_t Input)
Definition: xxhash.cpp:57
unsigned less than
Definition: InstrTypes.h:671
An efficient, type-erasing, non-owning reference to a callable.
Definition: STLExtras.h:116
LLVMContext & getContext() const
All values hold a context through their type.
Definition: Value.cpp:709
bool hasFnAttribute(Attribute::AttrKind Kind) const
Return true if the function has the attribute.
Definition: Function.h:320
Value * CreateSExt(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:1698
F(f)
const fltSemantics & getSemantics() const
Definition: APFloat.h:1154
unsigned getPointerAddressSpace() const
Get the address space of this pointer or pointer vector type.
Definition: DerivedTypes.h:534
CallInst * CreateMemSet(Value *Ptr, Value *Val, uint64_t Size, unsigned Align, bool isVolatile=false, MDNode *TBAATag=nullptr, MDNode *ScopeTag=nullptr, MDNode *NoAliasTag=nullptr)
Create and insert a memset to the specified pointer and the specified value.
Definition: IRBuilder.h:403
An instruction for reading from memory.
Definition: Instructions.h:167
Hexagon Common GEP
static bool isTrigLibCall(CallInst *CI)
Value * emitStrNCpy(Value *Dst, Value *Src, Value *Len, IRBuilder<> &B, const TargetLibraryInfo *TLI, StringRef Name="strncpy")
Emit a call to the strncpy function to the builder, for the specified pointer arguments and length...
bool isGEPBasedOnPointerToString(const GEPOperator *GEP, unsigned CharSize=8)
Returns true if the GEP is based on a pointer to a string (array of.
Value * emitFPutSUnlocked(Value *Str, Value *File, IRBuilder<> &B, const TargetLibraryInfo *TLI)
Emit a call to the fputs_unlocked function.
uint64_t Offset
Slice starts at this Offset.
void addAttribute(unsigned i, Attribute::AttrKind Kind)
adds the attribute to the list of attributes.
Definition: InstrTypes.h:1297
LLVMContext & getContext() const
Get the context in which this basic block lives.
Definition: BasicBlock.cpp:32
Value * emitMalloc(Value *Num, IRBuilder<> &B, const DataLayout &DL, const TargetLibraryInfo *TLI)
Emit a call to the malloc function.
static bool isReportingError(Function *Callee, CallInst *CI, int StreamArg)
static Constant * getNullValue(Type *Ty)
Constructor to create a &#39;0&#39; constant of arbitrary type.
Definition: Constants.cpp:274
iterator begin()
Instruction iterator methods.
Definition: BasicBlock.h:268
IntegerType * getInt32Ty()
Fetch the type representing a 32-bit integer.
Definition: IRBuilder.h:346
static Value * replaceUnaryCall(CallInst *CI, IRBuilder<> &B, Intrinsic::ID IID)
Value * getArgOperand(unsigned i) const
Definition: InstrTypes.h:1155
opStatus convertToInteger(MutableArrayRef< integerPart > Input, unsigned int Width, bool IsSigned, roundingMode RM, bool *IsExact) const
Definition: APFloat.h:1068
specific_fpval m_SpecificFP(double V)
Match a specific floating point value or vector with all elements equal to the value.
Definition: PatternMatch.h:539
bool match(Val *V, const Pattern &P)
Definition: PatternMatch.h:47
LibCallSimplifier(const DataLayout &DL, const TargetLibraryInfo *TLI, OptimizationRemarkEmitter &ORE, function_ref< void(Instruction *, Value *)> Replacer=&replaceAllUsesWithDefault, function_ref< void(Instruction *)> Eraser=&eraseFromParentDefault)
Value * CreateFPExt(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:1753
void setBit(unsigned BitPosition)
Set a given bit to 1.
Definition: APInt.h:1402
This class represents the LLVM &#39;select&#39; instruction.
const DataLayout & getDataLayout() const
Get the data layout for the module&#39;s target platform.
Definition: Module.cpp:369
bool isNonNegative() const
Determine if this APInt Value is non-negative (>= 0)
Definition: APInt.h:368
FortifiedLibCallSimplifier(const TargetLibraryInfo *TLI, bool OnlyLowerUnknownSize=false)
A Use represents the edge between a Value definition and its users.
Definition: Use.h:55
PointerType * getPointerTo(unsigned AddrSpace=0) const
Return a pointer to the current type.
Definition: Type.cpp:651
Value * emitPutChar(Value *Char, IRBuilder<> &B, const TargetLibraryInfo *TLI)
Emit a call to the putchar function. This assumes that Char is an integer.
bool isIntegerTy() const
True if this is an instance of IntegerType.
Definition: Type.h:196
This provides a uniform API for creating instructions and inserting them into a basic block: either a...
Definition: IRBuilder.h:742
Value * CreateAdd(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1049
CallInst * CreateMemMove(Value *Dst, unsigned DstAlign, Value *Src, unsigned SrcAlign, uint64_t Size, bool isVolatile=false, MDNode *TBAATag=nullptr, MDNode *ScopeTag=nullptr, MDNode *NoAliasTag=nullptr)
Create and insert a memmove between the specified pointers.
Definition: IRBuilder.h:493
Value * emitUnaryFloatFnCall(Value *Op, StringRef Name, IRBuilder<> &B, const AttributeList &Attrs)
Emit a call to the unary function named &#39;Name&#39; (e.g.
static cl::opt< bool > EnableUnsafeFPShrink("enable-double-float-shrink", cl::Hidden, cl::init(false), cl::desc("Enable unsafe double to float " "shrinking for math lib calls"))
Value * emitFPutCUnlocked(Value *Char, Value *File, IRBuilder<> &B, const TargetLibraryInfo *TLI)
Emit a call to the fputc_unlocked function.
LLVM_NODISCARD StringRef substr(size_t Start, size_t N=npos) const
Return a reference to the substring from [Start, Start + N).
Definition: StringRef.h:578
LLVM_NODISCARD bool empty() const
empty - Check if the string is empty.
Definition: StringRef.h:126
static StructType * get(LLVMContext &Context, ArrayRef< Type *> Elements, bool isPacked=false)
This static method is the primary way to create a literal StructType.
Definition: Type.cpp:341
StoreInst * CreateStore(Value *Val, Value *Ptr, bool isVolatile=false)
Definition: IRBuilder.h:1421
static Value * convertStrToNumber(CallInst *CI, StringRef &Str, int64_t Base)
Value * CreateIntToPtr(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:1762
Instruction * clone() const
Create a copy of &#39;this&#39; instruction that is identical in all ways except the following: ...
Class to represent function types.
Definition: DerivedTypes.h:102
Value * CreateBitCast(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:1767
FastMathFlags getFastMathFlags() const
Convenience function for getting all the fast-math flags, which must be an operator which supports th...
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:244
Value * CreateSExtOrTrunc(Value *V, Type *DestTy, const Twine &Name="")
Create a SExt or Trunc from the integer value V to DestTy.
Definition: IRBuilder.h:1719
apfloat_match m_APFloat(const APFloat *&Res)
Match a ConstantFP or splatted ConstantVector, binding the specified pointer to the contained APFloat...
Definition: PatternMatch.h:179
const ConstantDataArray * Array
ConstantDataArray pointer.
opStatus convert(const fltSemantics &ToSemantics, roundingMode RM, bool *losesInfo)
Definition: APFloat.cpp:4443
void setCalledFunction(Function *Fn)
Sets the function called, including updating the function type.
Definition: InstrTypes.h:1237
#define T
BasicBlock * GetInsertBlock() const
Definition: IRBuilder.h:120
Value * emitMemCpyChk(Value *Dst, Value *Src, Value *Len, Value *ObjSize, IRBuilder<> &B, const DataLayout &DL, const TargetLibraryInfo *TLI)
Emit a call to the __memcpy_chk function to the builder.
bool has(LibFunc F) const
Tests whether a library function is available.
LLVM_NODISCARD size_t size() const
size - Get the string size.
Definition: StringRef.h:130
SmallString - A SmallString is just a SmallVector with methods and accessors that make it work better...
Definition: SmallString.h:25
unsigned getOpcode() const
Returns a member of one of the enums like Instruction::Add.
Definition: Instruction.h:125
Value * emitStrCpy(Value *Dst, Value *Src, IRBuilder<> &B, const TargetLibraryInfo *TLI, StringRef Name="strcpy")
Emit a call to the strcpy function to the builder, for the specified pointer arguments.
Value * CreateSub(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1066
AttributeList getAttributes() const
Return the attribute list for this Function.
Definition: Function.h:223
void replaceAllUsesWith(Value *V)
Change all uses of this to point to a new Value.
Definition: Value.cpp:429
static const fltSemantics & IEEEdouble() LLVM_READNONE
Definition: APFloat.cpp:122
static Value * valueHasFloatPrecision(Value *Val)
Return a variant of Val with float type.
Value * CreateZExt(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:1694
constexpr char Attrs[]
Key for Kernel::Metadata::mAttrs.
uint64_t GetStringLength(const Value *V, unsigned CharSize=8)
If we can compute the length of the string pointed to by the specified pointer, return &#39;len+1&#39;...
void takeName(Value *V)
Transfer the name from V to this value.
Definition: Value.cpp:291
Function * getDeclaration(Module *M, ID id, ArrayRef< Type *> Tys=None)
Create or insert an LLVM Function declaration for an intrinsic, and return it.
Definition: Function.cpp:1018
void SetInsertPoint(BasicBlock *TheBB)
This specifies that created instructions should be appended to the end of the specified block...
Definition: IRBuilder.h:126
Value * getOperand(unsigned i) const
Definition: User.h:169
ConstantInt * getIntN(unsigned N, uint64_t C)
Get a constant N-bit value, zero extended or truncated from a 64-bit value.
Definition: IRBuilder.h:317
static bool isOnlyUsedInComparisonWithZero(Value *V)
bool getConstantDataArrayInfo(const Value *V, ConstantDataArraySlice &Slice, unsigned ElementSize, uint64_t Offset=0)
Returns true if the value V is a pointer into a ConstantDataArray.
static Constant * getBitCast(Constant *C, Type *Ty, bool OnlyIfReduced=false)
Definition: Constants.cpp:1782
Value * emitFPutS(Value *Str, Value *File, IRBuilder<> &B, const TargetLibraryInfo *TLI)
Emit a call to the fputs function.
bool isFloatTy() const
Return true if this is &#39;float&#39;, a 32-bit IEEE fp type.
Definition: Type.h:146
const BasicBlock & getEntryBlock() const
Definition: Function.h:639
IntegerType * getIntPtrType(LLVMContext &C, unsigned AddressSpace=0) const
Returns an integer type with size at least as big as that of a pointer in the given address space...
Definition: DataLayout.cpp:769
bool inferLibFuncAttributes(Function &F, const TargetLibraryInfo &TLI)
Analyze the name and prototype of the given function and set any applicable attributes.
ARM_AAPCS_VFP - Same as ARM_AAPCS, but uses hard floating point ABI.
Definition: CallingConv.h:102
Type * getDoubleTy()
Fetch the type representing a 64-bit floating point value.
Definition: IRBuilder.h:374
OneUse_match< T > m_OneUse(const T &SubPattern)
Definition: PatternMatch.h:61
initializer< Ty > init(const Ty &Val)
Definition: CommandLine.h:423
static bool isLocallyOpenedFile(Value *File, CallInst *CI, IRBuilder<> &B, const TargetLibraryInfo *TLI)
uint64_t getZExtValue() const
Return the constant as a 64-bit unsigned integer value after it has been zero extended as appropriate...
Definition: Constants.h:148
static GCRegistry::Add< OcamlGC > B("ocaml", "ocaml 3.10-compatible GC")
Value * CreateFMul(Value *L, Value *R, const Twine &Name="", MDNode *FPMD=nullptr)
Definition: IRBuilder.h:1282
void setNoSignedZeros(bool B=true)
Definition: Operator.h:224
bool isNegative() const
Definition: APFloat.h:1146
LLVM Basic Block Representation.
Definition: BasicBlock.h:57
The instances of the Type class are immutable: once they are created, they are never changed...
Definition: Type.h:45
Value * CreateConstInBoundsGEP1_64(Type *Ty, Value *Ptr, uint64_t Idx0, const Twine &Name="")
Definition: IRBuilder.h:1615
ConstantInt * getTrue()
Get the constant value for i1 true.
Definition: IRBuilder.h:286
bool isNoBuiltin() const
Return true if the call should not be treated as a call to a builtin.
Definition: InstrTypes.h:1519
Instrumentation for Order File
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
This is an important base class in LLVM.
Definition: Constant.h:41
Value * CreateSelect(Value *C, Value *True, Value *False, const Twine &Name="", Instruction *MDFrom=nullptr)
Definition: IRBuilder.h:2057
bool isPointerTy() const
True if this is an instance of PointerType.
Definition: Type.h:223
LLVM_NODISCARD size_t find_first_not_of(char C, size_t From=0) const
Find the first character in the string that is not C or npos if not found.
Definition: StringRef.cpp:249
FunctionType * getFunctionType() const
Definition: InstrTypes.h:1058
ConstantFP - Floating Point Values [float, double].
Definition: Constants.h:263
Value * CreateFCmpOGT(Value *LHS, Value *RHS, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:1919
double convertToDouble() const
Definition: APFloat.h:1096
Diagnostic information for applied optimization remarks.
ARM_AAPCS - ARM Architecture Procedure Calling Standard calling convention (aka EABI).
Definition: CallingConv.h:99
LLVM_NODISCARD size_t find(char C, size_t From=0) const
Search for the first character C in the string.
Definition: StringRef.h:285
unsigned getPrefTypeAlignment(Type *Ty) const
Returns the preferred stack/global alignment for the specified type.
Definition: DataLayout.cpp:759
bool isExactlyValue(double V) const
We don&#39;t rely on operator== working on double values, as it returns true for things that are clearly ...
Definition: APFloat.h:1129
bool isFast() const
Determine whether all fast-math-flags are set.
bool any_of(R &&range, UnaryPredicate P)
Provide wrappers to std::any_of which take ranges instead of having to pass begin/end explicitly...
Definition: STLExtras.h:1192
Value * CreateNeg(Value *V, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1343
bool optForSize() const
Optimize this function for size (-Os) or minimum size (-Oz).
Definition: Function.h:597
This instruction compares its operands according to the predicate given to the constructor.
void setCallingConv(CallingConv::ID CC)
Definition: InstrTypes.h:1262
amdgpu Simplify well known AMD library false FunctionCallee Value * Arg
op_range operands()
Definition: User.h:237
Value * getPointerOperand()
Definition: Instructions.h:284
static bool isCallingConvCCompatible(CallInst *CI)
Class to represent integer types.
Definition: DerivedTypes.h:39
ConstantInt * getInt64(uint64_t C)
Get a constant 64-bit value.
Definition: IRBuilder.h:311
C - The default llvm calling convention, compatible with C.
Definition: CallingConv.h:34
Value * castToCStr(Value *V, IRBuilder<> &B)
Return V if it is an i8*, otherwise cast it to i8*.
IntegerType * getIntNTy(unsigned N)
Fetch the type representing an N-bit integer.
Definition: IRBuilder.h:359
Represents offset+length into a ConstantDataArray.
Value * CreateExtractElement(Value *Vec, Value *Idx, const Twine &Name="")
Definition: IRBuilder.h:2077
const Function * getFunction() const
Return the function this instruction belongs to.
Definition: Instruction.cpp:59
Constant * ConstantFoldLoadFromConstPtr(Constant *C, Type *Ty, const DataLayout &DL)
ConstantFoldLoadFromConstPtr - Return the value that a load from C would produce if it is constant an...
static Value * optimizeMemCmpConstantSize(CallInst *CI, Value *LHS, Value *RHS, uint64_t Len, IRBuilder<> &B, const DataLayout &DL)
bool isIntN(unsigned N, int64_t x)
Checks if an signed integer fits into the given (dynamic) bit width.
Definition: MathExtras.h:397
uint64_t NextPowerOf2(uint64_t A)
Returns the next power of two (in 64-bits) that is strictly greater than A.
Definition: MathExtras.h:639
const Constant * stripPointerCasts() const
Definition: Constant.h:177
Value * emitStrChr(Value *Ptr, char C, IRBuilder<> &B, const TargetLibraryInfo *TLI)
Emit a call to the strchr function to the builder, for the specified pointer and character.
Value * CreateExtractValue(Value *Agg, ArrayRef< unsigned > Idxs, const Twine &Name="")
Definition: IRBuilder.h:2119
LLVM_NODISCARD char back() const
back - Get the last character in the string.
Definition: StringRef.h:141
bool hasFnAttr(Attribute::AttrKind Kind) const
Determine whether this call has the given attribute.
Definition: InstrTypes.h:1287
static bool isBinary(MachineInstr &MI)
Value * CreateTrunc(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:1690
Triple - Helper class for working with autoconf configuration names.
Definition: Triple.h:43
Value * emitFWriteUnlocked(Value *Ptr, Value *Size, Value *N, Value *File, IRBuilder<> &B, const DataLayout &DL, const TargetLibraryInfo *TLI)
Emit a call to the fwrite_unlocked function.
Value * CreateGEP(Value *Ptr, ArrayRef< Value *> IdxList, const Twine &Name="")
Definition: IRBuilder.h:1493
Value * CreateFCmpOLT(Value *LHS, Value *RHS, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:1929
Value * emitStrLen(Value *Ptr, IRBuilder<> &B, const DataLayout &DL, const TargetLibraryInfo *TLI)
Emit a call to the strlen function to the builder, for the specified pointer.
static IntegerType * get(LLVMContext &C, unsigned NumBits)
This static method is the primary way of constructing an IntegerType.
Definition: Type.cpp:239
Value * emitBinaryFloatFnCall(Value *Op1, Value *Op2, StringRef Name, IRBuilder<> &B, const AttributeList &Attrs)
Emit a call to the binary function named &#39;Name&#39; (e.g.
static const fltSemantics & IEEEsingle() LLVM_READNONE
Definition: APFloat.cpp:119
void addParamAttr(unsigned ArgNo, Attribute::AttrKind Kind)
Adds the attribute to the indicated argument.
Definition: InstrTypes.h:1311
static Value * optimizeDoubleFP(CallInst *CI, IRBuilder<> &B, bool isBinary, bool isPrecise=false)
Shrink double -> float functions.
bool isInteger() const
Definition: APFloat.h:1161
Value * emitFGetCUnlocked(Value *File, IRBuilder<> &B, const TargetLibraryInfo *TLI)
Emit a call to the fgetc_unlocked function. File is a pointer to FILE.
This is the shared class of boolean and integer constants.
Definition: Constants.h:83
GlobalVariable * CreateGlobalString(StringRef Str, const Twine &Name="", unsigned AddressSpace=0)
Make a new global variable with initializer type i8*.
Definition: IRBuilder.cpp:42
Value * CreateIntCast(Value *V, Type *DestTy, bool isSigned, const Twine &Name="")
Definition: IRBuilder.h:1836
This is a &#39;vector&#39; (really, a variable-sized array), optimized for the case when the array is small...
Definition: SmallVector.h:841
uint64_t getElementAsInteger(unsigned i) const
If this is a sequential container of integers (of any size), return the specified element in the low ...
Definition: Constants.cpp:2689
Module.h This file contains the declarations for the Module class.
uint64_t Length
Length of the slice.
Provides information about what library functions are available for the current target.
bool isLegalInteger(uint64_t Width) const
Returns true if the specified type is known to be a native integer type supported by the CPU...
Definition: DataLayout.h:254
ConstantInt * getInt32(uint32_t C)
Get a constant 32-bit value.
Definition: IRBuilder.h:306
void getOperandBundlesAsDefs(SmallVectorImpl< OperandBundleDef > &Defs) const
Return the list of operand bundles attached to this instruction as a vector of OperandBundleDefs.
Definition: InstrTypes.h:1751
CallInst * CreateMemCpy(Value *Dst, unsigned DstAlign, Value *Src, unsigned SrcAlign, uint64_t Size, bool isVolatile=false, MDNode *TBAATag=nullptr, MDNode *TBAAStructTag=nullptr, MDNode *ScopeTag=nullptr, MDNode *NoAliasTag=nullptr)
Create and insert a memcpy between the specified pointers.
Definition: IRBuilder.h:445
static Constant * get(Type *Ty, uint64_t V, bool isSigned=false)
If Ty is a vector type, return a Constant with a splat of the given value.
Definition: Constants.cpp:631
FunctionCallee getOrInsertFunction(StringRef Name, FunctionType *T, AttributeList AttributeList)
Look up the specified function in the module symbol table.
Definition: Module.cpp:143
static Constant * get(Type *Ty, double V)
This returns a ConstantFP, or a vector containing a splat of a ConstantFP, for the specified value in...
Definition: Constants.cpp:694
AttributeList getAttributes() const
Return the parameter attributes for this call.
Definition: InstrTypes.h:1280
BinaryOp_match< LHS, RHS, Instruction::FMul > m_FMul(const LHS &L, const RHS &R)
Definition: PatternMatch.h:706
Intrinsic::ID getIntrinsicID() const LLVM_READONLY
getIntrinsicID - This method returns the ID number of the specified function, or Intrinsic::not_intri...
Definition: Function.h:193
void setNoNaNs(bool B=true)
Definition: Operator.h:218
unsigned logBase2() const
Definition: APInt.h:1747
The access may modify the value stored in memory.
FunctionType * getFunctionType() const
Returns the FunctionType for me.
Definition: Function.h:163
const Module * getModule() const
Return the module owning the function this instruction belongs to or nullptr it the function does not...
Definition: Instruction.cpp:55
Class for arbitrary precision integers.
Definition: APInt.h:69
bool ule(const APInt &RHS) const
Unsigned less or equal comparison.
Definition: APInt.h:1222
amdgpu Simplify well known AMD library false FunctionCallee Callee
bool isPowerOf2() const
Check if this APInt&#39;s value is a power of two greater than zero.
Definition: APInt.h:463
iterator_range< user_iterator > users()
Definition: Value.h:399
IntegerType * getInt8Ty()
Fetch the type representing an 8-bit integer.
Definition: IRBuilder.h:336
Value * CreateShl(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1138
bool getLibFunc(StringRef funcName, LibFunc &F) const
Searches for a particular function name.
iterator begin() const
Definition: StringRef.h:101
Value * emitFPutC(Value *Char, Value *File, IRBuilder<> &B, const TargetLibraryInfo *TLI)
Emit a call to the fputc function.
ConstantInt * getFalse()
Get the constant value for i1 false.
Definition: IRBuilder.h:291
Value * emitFGetSUnlocked(Value *Str, Value *Size, Value *File, IRBuilder<> &B, const TargetLibraryInfo *TLI)
Emit a call to the fgets_unlocked function.
specific_fpval m_FPOne()
Match a float 1.0 or vector with all elements equal to 1.0.
Definition: PatternMatch.h:542
opStatus add(const APFloat &RHS, roundingMode RM)
Definition: APFloat.h:940
Predicate getPredicate() const
Return the predicate for this instruction.
Definition: InstrTypes.h:721
FNeg_match< OpTy > m_FNeg(const OpTy &X)
Match &#39;fneg X&#39; as &#39;fsub -0.0, X&#39;.
Definition: PatternMatch.h:688
unsigned getNumArgOperands() const
Definition: InstrTypes.h:1153
Merge contiguous icmps into a memcmp
Definition: MergeICmps.cpp:866
static const size_t npos
Definition: StringRef.h:50
unsigned getIntegerBitWidth() const
Definition: DerivedTypes.h:96
LLVM_NODISCARD bool empty() const
Definition: SmallVector.h:55
static Value * optimizeUnaryDoubleFP(CallInst *CI, IRBuilder<> &B, bool isPrecise=false)
Shrink double -> float for unary functions.
Value * emitStrNCmp(Value *Ptr1, Value *Ptr2, Value *Len, IRBuilder<> &B, const DataLayout &DL, const TargetLibraryInfo *TLI)
Emit a call to the strncmp function to the builder.
CallingConv::ID getCallingConv() const
Definition: InstrTypes.h:1258
StringRef getName() const
Return a constant reference to the value&#39;s name.
Definition: Value.cpp:214
Function * getCalledFunction() const
Returns the function called, or null if this is an indirect function invocation.
Definition: InstrTypes.h:1201
const Function * getParent() const
Return the enclosing method, or null if none.
Definition: BasicBlock.h:106
LLVM_NODISCARD size_t find_first_of(char C, size_t From=0) const
Find the first character in the string that is C, or npos if not found.
Definition: StringRef.h:380
#define I(x, y, z)
Definition: MD5.cpp:58
#define N
APFloat abs(APFloat X)
Returns the absolute value of the argument.
Definition: APFloat.h:1212
This class represents a cast unsigned integer to floating point.
bool isZero() const
This is just a convenience method to make client code smaller for a common code.
Definition: Constants.h:192
ARM_APCS - ARM Procedure Calling Standard calling convention (obsolete, but still used on some target...
Definition: CallingConv.h:95
static Value * optimizeBinaryDoubleFP(CallInst *CI, IRBuilder<> &B, bool isPrecise=false)
Shrink double -> float for binary functions.
Value * optimizeCall(CallInst *CI)
optimizeCall - Take the given call instruction and return a more optimal value to replace the instruc...
unsigned getKnownAlignment(Value *V, const DataLayout &DL, const Instruction *CxtI=nullptr, AssumptionCache *AC=nullptr, const DominatorTree *DT=nullptr)
Try to infer an alignment for the specified pointer.
Definition: Local.h:266
LLVM_NODISCARD std::enable_if<!is_simple_type< Y >::value, typename cast_retty< X, const Y >::ret_type >::type dyn_cast(const Y &Val)
Definition: Casting.h:322
uint32_t Size
Definition: Profile.cpp:46
CallInst * CreateCall(FunctionType *FTy, Value *Callee, ArrayRef< Value *> Args=None, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:2009
static void insertSinCosCall(IRBuilder<> &B, Function *OrigCallee, Value *Arg, bool UseFloat, Value *&Sin, Value *&Cos, Value *&SinCos)
Value * emitBCmp(Value *Ptr1, Value *Ptr2, Value *Len, IRBuilder<> &B, const DataLayout &DL, const TargetLibraryInfo *TLI)
Emit a call to the bcmp function.
Type * getFloatTy()
Fetch the type representing a 32-bit floating point value.
Definition: IRBuilder.h:369
Value * CreateFAdd(Value *L, Value *R, const Twine &Name="", MDNode *FPMD=nullptr)
Definition: IRBuilder.h:1248
bool isStrictFP() const
Determine if the call requires strict floating point semantics.
Definition: InstrTypes.h:1525
Value * CreateAnd(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1199
Value * CreateFDiv(Value *L, Value *R, const Twine &Name="", MDNode *FPMD=nullptr)
Definition: IRBuilder.h:1299
bool isDeclaration() const
Return true if the primary definition of this global value is outside of the current translation unit...
Definition: Globals.cpp:205
bool getConstantStringInfo(const Value *V, StringRef &Str, uint64_t Offset=0, bool TrimAtNul=true)
This function computes the length of a null-terminated C string pointed to by V.
Value * emitFWrite(Value *Ptr, Value *Size, Value *File, IRBuilder<> &B, const DataLayout &DL, const TargetLibraryInfo *TLI)
Emit a call to the fwrite function.
InstTy * Insert(InstTy *I, const Twine &Name="") const
Insert and return the specified instruction.
Definition: IRBuilder.h:793
Value * emitMemCmp(Value *Ptr1, Value *Ptr2, Value *Len, IRBuilder<> &B, const DataLayout &DL, const TargetLibraryInfo *TLI)
Emit a call to the memcmp function.
This class represents a cast from signed integer to floating point.
LLVM_NODISCARD const char * data() const
data - Get a pointer to the start of the string (which may not be null terminated).
Definition: StringRef.h:122
static bool isOnlyUsedInEqualityComparison(Value *V, Value *With)
Return true if it is only used in equality comparisons with With.
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
user_iterator user_begin()
Definition: Value.h:375
This class represents a truncation of floating point types.
Value * emitFReadUnlocked(Value *Ptr, Value *Size, Value *N, Value *File, IRBuilder<> &B, const DataLayout &DL, const TargetLibraryInfo *TLI)
Emit a call to the fread_unlocked function.
unsigned getPrimitiveSizeInBits() const LLVM_READONLY
Return the basic size of this type if it is a primitive type.
Definition: Type.cpp:114
Module * getParent()
Get the module that this global value is contained inside of...
Definition: GlobalValue.h:565
LLVM Value Representation.
Definition: Value.h:72
Value * emitPutS(Value *Str, IRBuilder<> &B, const TargetLibraryInfo *TLI)
Emit a call to the puts function. This assumes that Str is some pointer.
LLVM_NODISCARD StringRef drop_back(size_t N=1) const
Return a StringRef equal to &#39;this&#39; but with the last N elements dropped.
Definition: StringRef.h:628
static VectorType * get(Type *ElementType, unsigned NumElements)
This static method is the primary way to construct an VectorType.
Definition: Type.cpp:605
Value * emitMemChr(Value *Ptr, Value *Val, Value *Len, IRBuilder<> &B, const DataLayout &DL, const TargetLibraryInfo *TLI)
Emit a call to the memchr function.
void setFastMathFlags(FastMathFlags NewFMF)
Set the fast-math flags to be used with generated fp-math operators.
Definition: IRBuilder.h:219
static bool canTransformToMemCmp(CallInst *CI, Value *Str, uint64_t Len, const DataLayout &DL)
Value * emitCalloc(Value *Num, Value *Size, const AttributeList &Attrs, IRBuilder<> &B, const TargetLibraryInfo &TLI)
Emit a call to the calloc function.
ConstantInt * getInt8(uint8_t C)
Get a constant 8-bit value.
Definition: IRBuilder.h:296
static Value * optimizeTrigReflections(CallInst *Call, LibFunc Func, IRBuilder<> &B)
bool hasNoNaNs() const
Determine whether the no-NaNs flag is set.
bool hasOneUse() const
Return true if there is exactly one user of this value.
Definition: Value.h:412
static Value * getSqrtCall(Value *V, AttributeList Attrs, bool NoErrno, Module *M, IRBuilder<> &B, const TargetLibraryInfo *TLI)
Convenience struct for specifying and reasoning about fast-math flags.
Definition: Operator.h:159
StringRef - Represent a constant reference to a string, i.e.
Definition: StringRef.h:48
bool isDereferenceableAndAlignedPointer(const Value *V, unsigned Align, const DataLayout &DL, const Instruction *CtxI=nullptr, const DominatorTree *DT=nullptr)
Returns true if V is always a dereferenceable pointer with alignment greater or equal than requested...
Definition: Loads.cpp:128
int64_t getSExtValue() const
Return the constant as a 64-bit integer value after it has been sign extended as appropriate for the ...
Definition: Constants.h:156
ConstantInt * getInt(const APInt &AI)
Get a constant integer value.
Definition: IRBuilder.h:322
static bool ignoreCallingConv(LibFunc Func)
This class represents an extension of floating point types.
iterator end() const
Definition: StringRef.h:103
static Constant * getFPExtend(Constant *C, Type *Ty, bool OnlyIfReduced=false)
Definition: Constants.cpp:1700
bool isDoubleTy() const
Return true if this is &#39;double&#39;, a 64-bit IEEE fp type.
Definition: Type.h:149
bool hasUnaryFloatFn(const TargetLibraryInfo *TLI, Type *Ty, LibFunc DoubleFn, LibFunc FloatFn, LibFunc LongDoubleFn)
Check whether the overloaded unary floating point function corresponding to Ty is available...
The optimization diagnostic interface.
bool use_empty() const
Definition: Value.h:322
static Value * getPow(Value *InnerChain[33], unsigned Exp, IRBuilder<> &B)
Value * CreateFNeg(Value *V, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:1361
bool isStructTy() const
True if this is an instance of StructType.
Definition: Type.h:217
cmpResult compare(const APFloat &RHS) const
Definition: APFloat.h:1101
bool isArrayTy() const
True if this is an instance of ArrayType.
Definition: Type.h:220
A wrapper class for inspecting calls to intrinsic functions.
Definition: IntrinsicInst.h:43
bool PointerMayBeCaptured(const Value *V, bool ReturnCaptures, bool StoreCaptures, unsigned MaxUsesToExplore=DefaultMaxUsesToExplore)
PointerMayBeCaptured - Return true if this pointer value may be captured by the enclosing function (w...
user_iterator user_end()
Definition: Value.h:383