LLVM  6.0.0svn
SimplifyLibCalls.cpp
Go to the documentation of this file.
1 //===------ SimplifyLibCalls.cpp - Library calls simplifier ---------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This is a utility pass used for testing the InstructionSimplify analysis.
11 // The analysis is applied to every instruction, and if it simplifies then the
12 // instruction is replaced by the simplification. If you are looking for a pass
13 // that performs serious instruction folding, use the instcombine pass instead.
14 //
15 //===----------------------------------------------------------------------===//
16 
18 #include "llvm/ADT/SmallString.h"
19 #include "llvm/ADT/StringMap.h"
20 #include "llvm/ADT/Triple.h"
25 #include "llvm/IR/DataLayout.h"
26 #include "llvm/IR/Function.h"
27 #include "llvm/IR/IRBuilder.h"
28 #include "llvm/IR/IntrinsicInst.h"
29 #include "llvm/IR/Intrinsics.h"
30 #include "llvm/IR/LLVMContext.h"
31 #include "llvm/IR/Module.h"
32 #include "llvm/IR/PatternMatch.h"
34 #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 /// \brief Check whether the overloaded unary floating point function
108 /// corresponding to \a Ty is available.
109 static bool hasUnaryFloatFn(const TargetLibraryInfo *TLI, Type *Ty,
110  LibFunc DoubleFn, LibFunc FloatFn,
111  LibFunc LongDoubleFn) {
112  switch (Ty->getTypeID()) {
113  case Type::FloatTyID:
114  return TLI->has(FloatFn);
115  case Type::DoubleTyID:
116  return TLI->has(DoubleFn);
117  default:
118  return TLI->has(LongDoubleFn);
119  }
120 }
121 
122 //===----------------------------------------------------------------------===//
123 // String and Memory Library Call Optimizations
124 //===----------------------------------------------------------------------===//
125 
126 Value *LibCallSimplifier::optimizeStrCat(CallInst *CI, IRBuilder<> &B) {
127  // Extract some information from the instruction
128  Value *Dst = CI->getArgOperand(0);
129  Value *Src = CI->getArgOperand(1);
130 
131  // See if we can get the length of the input string.
132  uint64_t Len = GetStringLength(Src);
133  if (Len == 0)
134  return nullptr;
135  --Len; // Unbias length.
136 
137  // Handle the simple, do-nothing case: strcat(x, "") -> x
138  if (Len == 0)
139  return Dst;
140 
141  return emitStrLenMemCpy(Src, Dst, Len, B);
142 }
143 
144 Value *LibCallSimplifier::emitStrLenMemCpy(Value *Src, Value *Dst, uint64_t Len,
145  IRBuilder<> &B) {
146  // We need to find the end of the destination string. That's where the
147  // memory is to be moved to. We just generate a call to strlen.
148  Value *DstLen = emitStrLen(Dst, B, DL, TLI);
149  if (!DstLen)
150  return nullptr;
151 
152  // Now that we have the destination's length, we must index into the
153  // destination's pointer to get the actual memcpy destination (end of
154  // the string .. we're concatenating).
155  Value *CpyDst = B.CreateGEP(B.getInt8Ty(), Dst, DstLen, "endptr");
156 
157  // We have enough information to now generate the memcpy call to do the
158  // concatenation for us. Make a memcpy to copy the nul byte with align = 1.
159  B.CreateMemCpy(CpyDst, Src,
160  ConstantInt::get(DL.getIntPtrType(Src->getContext()), Len + 1),
161  1);
162  return Dst;
163 }
164 
165 Value *LibCallSimplifier::optimizeStrNCat(CallInst *CI, IRBuilder<> &B) {
166  // Extract some information from the instruction.
167  Value *Dst = CI->getArgOperand(0);
168  Value *Src = CI->getArgOperand(1);
169  uint64_t Len;
170 
171  // We don't do anything if length is not constant.
172  if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(CI->getArgOperand(2)))
173  Len = LengthArg->getZExtValue();
174  else
175  return nullptr;
176 
177  // See if we can get the length of the input string.
178  uint64_t SrcLen = GetStringLength(Src);
179  if (SrcLen == 0)
180  return nullptr;
181  --SrcLen; // Unbias length.
182 
183  // Handle the simple, do-nothing cases:
184  // strncat(x, "", c) -> x
185  // strncat(x, c, 0) -> x
186  if (SrcLen == 0 || Len == 0)
187  return Dst;
188 
189  // We don't optimize this case.
190  if (Len < SrcLen)
191  return nullptr;
192 
193  // strncat(x, s, c) -> strcat(x, s)
194  // s is constant so the strcat can be optimized further.
195  return emitStrLenMemCpy(Src, Dst, SrcLen, B);
196 }
197 
198 Value *LibCallSimplifier::optimizeStrChr(CallInst *CI, IRBuilder<> &B) {
200  FunctionType *FT = Callee->getFunctionType();
201  Value *SrcStr = CI->getArgOperand(0);
202 
203  // If the second operand is non-constant, see if we can compute the length
204  // of the input string and turn this into memchr.
205  ConstantInt *CharC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
206  if (!CharC) {
207  uint64_t Len = GetStringLength(SrcStr);
208  if (Len == 0 || !FT->getParamType(1)->isIntegerTy(32)) // memchr needs i32.
209  return nullptr;
210 
211  return emitMemChr(SrcStr, CI->getArgOperand(1), // include nul.
212  ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len),
213  B, DL, TLI);
214  }
215 
216  // Otherwise, the character is a constant, see if the first argument is
217  // a string literal. If so, we can constant fold.
218  StringRef Str;
219  if (!getConstantStringInfo(SrcStr, Str)) {
220  if (CharC->isZero()) // strchr(p, 0) -> p + strlen(p)
221  return B.CreateGEP(B.getInt8Ty(), SrcStr, emitStrLen(SrcStr, B, DL, TLI),
222  "strchr");
223  return nullptr;
224  }
225 
226  // Compute the offset, make sure to handle the case when we're searching for
227  // zero (a weird way to spell strlen).
228  size_t I = (0xFF & CharC->getSExtValue()) == 0
229  ? Str.size()
230  : Str.find(CharC->getSExtValue());
231  if (I == StringRef::npos) // Didn't find the char. strchr returns null.
232  return Constant::getNullValue(CI->getType());
233 
234  // strchr(s+n,c) -> gep(s+n+i,c)
235  return B.CreateGEP(B.getInt8Ty(), SrcStr, B.getInt64(I), "strchr");
236 }
237 
238 Value *LibCallSimplifier::optimizeStrRChr(CallInst *CI, IRBuilder<> &B) {
239  Value *SrcStr = CI->getArgOperand(0);
240  ConstantInt *CharC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
241 
242  // Cannot fold anything if we're not looking for a constant.
243  if (!CharC)
244  return nullptr;
245 
246  StringRef Str;
247  if (!getConstantStringInfo(SrcStr, Str)) {
248  // strrchr(s, 0) -> strchr(s, 0)
249  if (CharC->isZero())
250  return emitStrChr(SrcStr, '\0', B, TLI);
251  return nullptr;
252  }
253 
254  // Compute the offset.
255  size_t I = (0xFF & CharC->getSExtValue()) == 0
256  ? Str.size()
257  : Str.rfind(CharC->getSExtValue());
258  if (I == StringRef::npos) // Didn't find the char. Return null.
259  return Constant::getNullValue(CI->getType());
260 
261  // strrchr(s+n,c) -> gep(s+n+i,c)
262  return B.CreateGEP(B.getInt8Ty(), SrcStr, B.getInt64(I), "strrchr");
263 }
264 
265 Value *LibCallSimplifier::optimizeStrCmp(CallInst *CI, IRBuilder<> &B) {
266  Value *Str1P = CI->getArgOperand(0), *Str2P = CI->getArgOperand(1);
267  if (Str1P == Str2P) // strcmp(x,x) -> 0
268  return ConstantInt::get(CI->getType(), 0);
269 
270  StringRef Str1, Str2;
271  bool HasStr1 = getConstantStringInfo(Str1P, Str1);
272  bool HasStr2 = getConstantStringInfo(Str2P, Str2);
273 
274  // strcmp(x, y) -> cnst (if both x and y are constant strings)
275  if (HasStr1 && HasStr2)
276  return ConstantInt::get(CI->getType(), Str1.compare(Str2));
277 
278  if (HasStr1 && Str1.empty()) // strcmp("", x) -> -*x
279  return B.CreateNeg(
280  B.CreateZExt(B.CreateLoad(Str2P, "strcmpload"), CI->getType()));
281 
282  if (HasStr2 && Str2.empty()) // strcmp(x,"") -> *x
283  return B.CreateZExt(B.CreateLoad(Str1P, "strcmpload"), CI->getType());
284 
285  // strcmp(P, "x") -> memcmp(P, "x", 2)
286  uint64_t Len1 = GetStringLength(Str1P);
287  uint64_t Len2 = GetStringLength(Str2P);
288  if (Len1 && Len2) {
289  return emitMemCmp(Str1P, Str2P,
290  ConstantInt::get(DL.getIntPtrType(CI->getContext()),
291  std::min(Len1, Len2)),
292  B, DL, TLI);
293  }
294 
295  return nullptr;
296 }
297 
298 Value *LibCallSimplifier::optimizeStrNCmp(CallInst *CI, IRBuilder<> &B) {
299  Value *Str1P = CI->getArgOperand(0), *Str2P = CI->getArgOperand(1);
300  if (Str1P == Str2P) // strncmp(x,x,n) -> 0
301  return ConstantInt::get(CI->getType(), 0);
302 
303  // Get the length argument if it is constant.
304  uint64_t Length;
305  if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(CI->getArgOperand(2)))
306  Length = LengthArg->getZExtValue();
307  else
308  return nullptr;
309 
310  if (Length == 0) // strncmp(x,y,0) -> 0
311  return ConstantInt::get(CI->getType(), 0);
312 
313  if (Length == 1) // strncmp(x,y,1) -> memcmp(x,y,1)
314  return emitMemCmp(Str1P, Str2P, CI->getArgOperand(2), B, DL, TLI);
315 
316  StringRef Str1, Str2;
317  bool HasStr1 = getConstantStringInfo(Str1P, Str1);
318  bool HasStr2 = getConstantStringInfo(Str2P, Str2);
319 
320  // strncmp(x, y) -> cnst (if both x and y are constant strings)
321  if (HasStr1 && HasStr2) {
322  StringRef SubStr1 = Str1.substr(0, Length);
323  StringRef SubStr2 = Str2.substr(0, Length);
324  return ConstantInt::get(CI->getType(), SubStr1.compare(SubStr2));
325  }
326 
327  if (HasStr1 && Str1.empty()) // strncmp("", x, n) -> -*x
328  return B.CreateNeg(
329  B.CreateZExt(B.CreateLoad(Str2P, "strcmpload"), CI->getType()));
330 
331  if (HasStr2 && Str2.empty()) // strncmp(x, "", n) -> *x
332  return B.CreateZExt(B.CreateLoad(Str1P, "strcmpload"), CI->getType());
333 
334  return nullptr;
335 }
336 
337 Value *LibCallSimplifier::optimizeStrCpy(CallInst *CI, IRBuilder<> &B) {
338  Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1);
339  if (Dst == Src) // strcpy(x,x) -> x
340  return Src;
341 
342  // See if we can get the length of the input string.
343  uint64_t Len = GetStringLength(Src);
344  if (Len == 0)
345  return nullptr;
346 
347  // We have enough information to now generate the memcpy call to do the
348  // copy for us. Make a memcpy to copy the nul byte with align = 1.
349  B.CreateMemCpy(Dst, Src,
350  ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len), 1);
351  return Dst;
352 }
353 
354 Value *LibCallSimplifier::optimizeStpCpy(CallInst *CI, IRBuilder<> &B) {
356  Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1);
357  if (Dst == Src) { // stpcpy(x,x) -> x+strlen(x)
358  Value *StrLen = emitStrLen(Src, B, DL, TLI);
359  return StrLen ? B.CreateInBoundsGEP(B.getInt8Ty(), Dst, StrLen) : nullptr;
360  }
361 
362  // See if we can get the length of the input string.
363  uint64_t Len = GetStringLength(Src);
364  if (Len == 0)
365  return nullptr;
366 
367  Type *PT = Callee->getFunctionType()->getParamType(0);
368  Value *LenV = ConstantInt::get(DL.getIntPtrType(PT), Len);
369  Value *DstEnd = B.CreateGEP(B.getInt8Ty(), Dst,
370  ConstantInt::get(DL.getIntPtrType(PT), Len - 1));
371 
372  // We have enough information to now generate the memcpy call to do the
373  // copy for us. Make a memcpy to copy the nul byte with align = 1.
374  B.CreateMemCpy(Dst, Src, LenV, 1);
375  return DstEnd;
376 }
377 
378 Value *LibCallSimplifier::optimizeStrNCpy(CallInst *CI, IRBuilder<> &B) {
380  Value *Dst = CI->getArgOperand(0);
381  Value *Src = CI->getArgOperand(1);
382  Value *LenOp = CI->getArgOperand(2);
383 
384  // See if we can get the length of the input string.
385  uint64_t SrcLen = GetStringLength(Src);
386  if (SrcLen == 0)
387  return nullptr;
388  --SrcLen;
389 
390  if (SrcLen == 0) {
391  // strncpy(x, "", y) -> memset(x, '\0', y, 1)
392  B.CreateMemSet(Dst, B.getInt8('\0'), LenOp, 1);
393  return Dst;
394  }
395 
396  uint64_t Len;
397  if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(LenOp))
398  Len = LengthArg->getZExtValue();
399  else
400  return nullptr;
401 
402  if (Len == 0)
403  return Dst; // strncpy(x, y, 0) -> x
404 
405  // Let strncpy handle the zero padding
406  if (Len > SrcLen + 1)
407  return nullptr;
408 
409  Type *PT = Callee->getFunctionType()->getParamType(0);
410  // strncpy(x, s, c) -> memcpy(x, s, c, 1) [s and c are constant]
411  B.CreateMemCpy(Dst, Src, ConstantInt::get(DL.getIntPtrType(PT), Len), 1);
412 
413  return Dst;
414 }
415 
416 Value *LibCallSimplifier::optimizeStringLength(CallInst *CI, IRBuilder<> &B,
417  unsigned CharSize) {
418  Value *Src = CI->getArgOperand(0);
419 
420  // Constant folding: strlen("xyz") -> 3
421  if (uint64_t Len = GetStringLength(Src, CharSize))
422  return ConstantInt::get(CI->getType(), Len - 1);
423 
424  // If s is a constant pointer pointing to a string literal, we can fold
425  // strlen(s + x) to strlen(s) - x, when x is known to be in the range
426  // [0, strlen(s)] or the string has a single null terminator '\0' at the end.
427  // We only try to simplify strlen when the pointer s points to an array
428  // of i8. Otherwise, we would need to scale the offset x before doing the
429  // subtraction. This will make the optimization more complex, and it's not
430  // very useful because calling strlen for a pointer of other types is
431  // very uncommon.
432  if (GEPOperator *GEP = dyn_cast<GEPOperator>(Src)) {
433  if (!isGEPBasedOnPointerToString(GEP, CharSize))
434  return nullptr;
435 
437  if (getConstantDataArrayInfo(GEP->getOperand(0), Slice, CharSize)) {
438  uint64_t NullTermIdx;
439  if (Slice.Array == nullptr) {
440  NullTermIdx = 0;
441  } else {
442  NullTermIdx = ~((uint64_t)0);
443  for (uint64_t I = 0, E = Slice.Length; I < E; ++I) {
444  if (Slice.Array->getElementAsInteger(I + Slice.Offset) == 0) {
445  NullTermIdx = I;
446  break;
447  }
448  }
449  // If the string does not have '\0', leave it to strlen to compute
450  // its length.
451  if (NullTermIdx == ~((uint64_t)0))
452  return nullptr;
453  }
454 
455  Value *Offset = GEP->getOperand(2);
456  KnownBits Known = computeKnownBits(Offset, DL, 0, nullptr, CI, nullptr);
457  Known.Zero.flipAllBits();
458  uint64_t ArrSize =
459  cast<ArrayType>(GEP->getSourceElementType())->getNumElements();
460 
461  // KnownZero's bits are flipped, so zeros in KnownZero now represent
462  // bits known to be zeros in Offset, and ones in KnowZero represent
463  // bits unknown in Offset. Therefore, Offset is known to be in range
464  // [0, NullTermIdx] when the flipped KnownZero is non-negative and
465  // unsigned-less-than NullTermIdx.
466  //
467  // If Offset is not provably in the range [0, NullTermIdx], we can still
468  // optimize if we can prove that the program has undefined behavior when
469  // Offset is outside that range. That is the case when GEP->getOperand(0)
470  // is a pointer to an object whose memory extent is NullTermIdx+1.
471  if ((Known.Zero.isNonNegative() && Known.Zero.ule(NullTermIdx)) ||
472  (GEP->isInBounds() && isa<GlobalVariable>(GEP->getOperand(0)) &&
473  NullTermIdx == ArrSize - 1)) {
474  Offset = B.CreateSExtOrTrunc(Offset, CI->getType());
475  return B.CreateSub(ConstantInt::get(CI->getType(), NullTermIdx),
476  Offset);
477  }
478  }
479 
480  return nullptr;
481  }
482 
483  // strlen(x?"foo":"bars") --> x ? 3 : 4
484  if (SelectInst *SI = dyn_cast<SelectInst>(Src)) {
485  uint64_t LenTrue = GetStringLength(SI->getTrueValue(), CharSize);
486  uint64_t LenFalse = GetStringLength(SI->getFalseValue(), CharSize);
487  if (LenTrue && LenFalse) {
488  ORE.emit([&]() {
489  return OptimizationRemark("instcombine", "simplify-libcalls", CI)
490  << "folded strlen(select) to select of constants";
491  });
492  return B.CreateSelect(SI->getCondition(),
493  ConstantInt::get(CI->getType(), LenTrue - 1),
494  ConstantInt::get(CI->getType(), LenFalse - 1));
495  }
496  }
497 
498  // strlen(x) != 0 --> *x != 0
499  // strlen(x) == 0 --> *x == 0
501  return B.CreateZExt(B.CreateLoad(Src, "strlenfirst"), CI->getType());
502 
503  return nullptr;
504 }
505 
506 Value *LibCallSimplifier::optimizeStrLen(CallInst *CI, IRBuilder<> &B) {
507  return optimizeStringLength(CI, B, 8);
508 }
509 
510 Value *LibCallSimplifier::optimizeWcslen(CallInst *CI, IRBuilder<> &B) {
511  Module &M = *CI->getParent()->getParent()->getParent();
512  unsigned WCharSize = TLI->getWCharSize(M) * 8;
513  // We cannot perform this optimization without wchar_size metadata.
514  if (WCharSize == 0)
515  return nullptr;
516 
517  return optimizeStringLength(CI, B, WCharSize);
518 }
519 
520 Value *LibCallSimplifier::optimizeStrPBrk(CallInst *CI, IRBuilder<> &B) {
521  StringRef S1, S2;
522  bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1);
523  bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2);
524 
525  // strpbrk(s, "") -> nullptr
526  // strpbrk("", s) -> nullptr
527  if ((HasS1 && S1.empty()) || (HasS2 && S2.empty()))
528  return Constant::getNullValue(CI->getType());
529 
530  // Constant folding.
531  if (HasS1 && HasS2) {
532  size_t I = S1.find_first_of(S2);
533  if (I == StringRef::npos) // No match.
534  return Constant::getNullValue(CI->getType());
535 
536  return B.CreateGEP(B.getInt8Ty(), CI->getArgOperand(0), B.getInt64(I),
537  "strpbrk");
538  }
539 
540  // strpbrk(s, "a") -> strchr(s, 'a')
541  if (HasS2 && S2.size() == 1)
542  return emitStrChr(CI->getArgOperand(0), S2[0], B, TLI);
543 
544  return nullptr;
545 }
546 
547 Value *LibCallSimplifier::optimizeStrTo(CallInst *CI, IRBuilder<> &B) {
548  Value *EndPtr = CI->getArgOperand(1);
549  if (isa<ConstantPointerNull>(EndPtr)) {
550  // With a null EndPtr, this function won't capture the main argument.
551  // It would be readonly too, except that it still may write to errno.
552  CI->addParamAttr(0, Attribute::NoCapture);
553  }
554 
555  return nullptr;
556 }
557 
558 Value *LibCallSimplifier::optimizeStrSpn(CallInst *CI, IRBuilder<> &B) {
559  StringRef S1, S2;
560  bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1);
561  bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2);
562 
563  // strspn(s, "") -> 0
564  // strspn("", s) -> 0
565  if ((HasS1 && S1.empty()) || (HasS2 && S2.empty()))
566  return Constant::getNullValue(CI->getType());
567 
568  // Constant folding.
569  if (HasS1 && HasS2) {
570  size_t Pos = S1.find_first_not_of(S2);
571  if (Pos == StringRef::npos)
572  Pos = S1.size();
573  return ConstantInt::get(CI->getType(), Pos);
574  }
575 
576  return nullptr;
577 }
578 
579 Value *LibCallSimplifier::optimizeStrCSpn(CallInst *CI, IRBuilder<> &B) {
580  StringRef S1, S2;
581  bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1);
582  bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2);
583 
584  // strcspn("", s) -> 0
585  if (HasS1 && S1.empty())
586  return Constant::getNullValue(CI->getType());
587 
588  // Constant folding.
589  if (HasS1 && HasS2) {
590  size_t Pos = S1.find_first_of(S2);
591  if (Pos == StringRef::npos)
592  Pos = S1.size();
593  return ConstantInt::get(CI->getType(), Pos);
594  }
595 
596  // strcspn(s, "") -> strlen(s)
597  if (HasS2 && S2.empty())
598  return emitStrLen(CI->getArgOperand(0), B, DL, TLI);
599 
600  return nullptr;
601 }
602 
603 Value *LibCallSimplifier::optimizeStrStr(CallInst *CI, IRBuilder<> &B) {
604  // fold strstr(x, x) -> x.
605  if (CI->getArgOperand(0) == CI->getArgOperand(1))
606  return B.CreateBitCast(CI->getArgOperand(0), CI->getType());
607 
608  // fold strstr(a, b) == a -> strncmp(a, b, strlen(b)) == 0
610  Value *StrLen = emitStrLen(CI->getArgOperand(1), B, DL, TLI);
611  if (!StrLen)
612  return nullptr;
613  Value *StrNCmp = emitStrNCmp(CI->getArgOperand(0), CI->getArgOperand(1),
614  StrLen, B, DL, TLI);
615  if (!StrNCmp)
616  return nullptr;
617  for (auto UI = CI->user_begin(), UE = CI->user_end(); UI != UE;) {
618  ICmpInst *Old = cast<ICmpInst>(*UI++);
619  Value *Cmp =
620  B.CreateICmp(Old->getPredicate(), StrNCmp,
621  ConstantInt::getNullValue(StrNCmp->getType()), "cmp");
622  replaceAllUsesWith(Old, Cmp);
623  }
624  return CI;
625  }
626 
627  // See if either input string is a constant string.
628  StringRef SearchStr, ToFindStr;
629  bool HasStr1 = getConstantStringInfo(CI->getArgOperand(0), SearchStr);
630  bool HasStr2 = getConstantStringInfo(CI->getArgOperand(1), ToFindStr);
631 
632  // fold strstr(x, "") -> x.
633  if (HasStr2 && ToFindStr.empty())
634  return B.CreateBitCast(CI->getArgOperand(0), CI->getType());
635 
636  // If both strings are known, constant fold it.
637  if (HasStr1 && HasStr2) {
638  size_t Offset = SearchStr.find(ToFindStr);
639 
640  if (Offset == StringRef::npos) // strstr("foo", "bar") -> null
641  return Constant::getNullValue(CI->getType());
642 
643  // strstr("abcd", "bc") -> gep((char*)"abcd", 1)
644  Value *Result = castToCStr(CI->getArgOperand(0), B);
645  Result = B.CreateConstInBoundsGEP1_64(Result, Offset, "strstr");
646  return B.CreateBitCast(Result, CI->getType());
647  }
648 
649  // fold strstr(x, "y") -> strchr(x, 'y').
650  if (HasStr2 && ToFindStr.size() == 1) {
651  Value *StrChr = emitStrChr(CI->getArgOperand(0), ToFindStr[0], B, TLI);
652  return StrChr ? B.CreateBitCast(StrChr, CI->getType()) : nullptr;
653  }
654  return nullptr;
655 }
656 
657 Value *LibCallSimplifier::optimizeMemChr(CallInst *CI, IRBuilder<> &B) {
658  Value *SrcStr = CI->getArgOperand(0);
659  ConstantInt *CharC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
661 
662  // memchr(x, y, 0) -> null
663  if (LenC && LenC->isZero())
664  return Constant::getNullValue(CI->getType());
665 
666  // From now on we need at least constant length and string.
667  StringRef Str;
668  if (!LenC || !getConstantStringInfo(SrcStr, Str, 0, /*TrimAtNul=*/false))
669  return nullptr;
670 
671  // Truncate the string to LenC. If Str is smaller than LenC we will still only
672  // scan the string, as reading past the end of it is undefined and we can just
673  // return null if we don't find the char.
674  Str = Str.substr(0, LenC->getZExtValue());
675 
676  // If the char is variable but the input str and length are not we can turn
677  // this memchr call into a simple bit field test. Of course this only works
678  // when the return value is only checked against null.
679  //
680  // It would be really nice to reuse switch lowering here but we can't change
681  // the CFG at this point.
682  //
683  // memchr("\r\n", C, 2) != nullptr -> (C & ((1 << '\r') | (1 << '\n'))) != 0
684  // after bounds check.
685  if (!CharC && !Str.empty() && isOnlyUsedInZeroEqualityComparison(CI)) {
686  unsigned char Max =
687  *std::max_element(reinterpret_cast<const unsigned char *>(Str.begin()),
688  reinterpret_cast<const unsigned char *>(Str.end()));
689 
690  // Make sure the bit field we're about to create fits in a register on the
691  // target.
692  // FIXME: On a 64 bit architecture this prevents us from using the
693  // interesting range of alpha ascii chars. We could do better by emitting
694  // two bitfields or shifting the range by 64 if no lower chars are used.
695  if (!DL.fitsInLegalInteger(Max + 1))
696  return nullptr;
697 
698  // For the bit field use a power-of-2 type with at least 8 bits to avoid
699  // creating unnecessary illegal types.
700  unsigned char Width = NextPowerOf2(std::max((unsigned char)7, Max));
701 
702  // Now build the bit field.
703  APInt Bitfield(Width, 0);
704  for (char C : Str)
705  Bitfield.setBit((unsigned char)C);
706  Value *BitfieldC = B.getInt(Bitfield);
707 
708  // First check that the bit field access is within bounds.
709  Value *C = B.CreateZExtOrTrunc(CI->getArgOperand(1), BitfieldC->getType());
710  Value *Bounds = B.CreateICmp(ICmpInst::ICMP_ULT, C, B.getIntN(Width, Width),
711  "memchr.bounds");
712 
713  // Create code that checks if the given bit is set in the field.
714  Value *Shl = B.CreateShl(B.getIntN(Width, 1ULL), C);
715  Value *Bits = B.CreateIsNotNull(B.CreateAnd(Shl, BitfieldC), "memchr.bits");
716 
717  // Finally merge both checks and cast to pointer type. The inttoptr
718  // implicitly zexts the i1 to intptr type.
719  return B.CreateIntToPtr(B.CreateAnd(Bounds, Bits, "memchr"), CI->getType());
720  }
721 
722  // Check if all arguments are constants. If so, we can constant fold.
723  if (!CharC)
724  return nullptr;
725 
726  // Compute the offset.
727  size_t I = Str.find(CharC->getSExtValue() & 0xFF);
728  if (I == StringRef::npos) // Didn't find the char. memchr returns null.
729  return Constant::getNullValue(CI->getType());
730 
731  // memchr(s+n,c,l) -> gep(s+n+i,c)
732  return B.CreateGEP(B.getInt8Ty(), SrcStr, B.getInt64(I), "memchr");
733 }
734 
735 Value *LibCallSimplifier::optimizeMemCmp(CallInst *CI, IRBuilder<> &B) {
736  Value *LHS = CI->getArgOperand(0), *RHS = CI->getArgOperand(1);
737 
738  if (LHS == RHS) // memcmp(s,s,x) -> 0
739  return Constant::getNullValue(CI->getType());
740 
741  // Make sure we have a constant length.
743  if (!LenC)
744  return nullptr;
745 
746  uint64_t Len = LenC->getZExtValue();
747  if (Len == 0) // memcmp(s1,s2,0) -> 0
748  return Constant::getNullValue(CI->getType());
749 
750  // memcmp(S1,S2,1) -> *(unsigned char*)LHS - *(unsigned char*)RHS
751  if (Len == 1) {
752  Value *LHSV = B.CreateZExt(B.CreateLoad(castToCStr(LHS, B), "lhsc"),
753  CI->getType(), "lhsv");
754  Value *RHSV = B.CreateZExt(B.CreateLoad(castToCStr(RHS, B), "rhsc"),
755  CI->getType(), "rhsv");
756  return B.CreateSub(LHSV, RHSV, "chardiff");
757  }
758 
759  // memcmp(S1,S2,N/8)==0 -> (*(intN_t*)S1 != *(intN_t*)S2)==0
760  // TODO: The case where both inputs are constants does not need to be limited
761  // to legal integers or equality comparison. See block below this.
762  if (DL.isLegalInteger(Len * 8) && isOnlyUsedInZeroEqualityComparison(CI)) {
763  IntegerType *IntType = IntegerType::get(CI->getContext(), Len * 8);
764  unsigned PrefAlignment = DL.getPrefTypeAlignment(IntType);
765 
766  // First, see if we can fold either argument to a constant.
767  Value *LHSV = nullptr;
768  if (auto *LHSC = dyn_cast<Constant>(LHS)) {
769  LHSC = ConstantExpr::getBitCast(LHSC, IntType->getPointerTo());
770  LHSV = ConstantFoldLoadFromConstPtr(LHSC, IntType, DL);
771  }
772  Value *RHSV = nullptr;
773  if (auto *RHSC = dyn_cast<Constant>(RHS)) {
774  RHSC = ConstantExpr::getBitCast(RHSC, IntType->getPointerTo());
775  RHSV = ConstantFoldLoadFromConstPtr(RHSC, IntType, DL);
776  }
777 
778  // Don't generate unaligned loads. If either source is constant data,
779  // alignment doesn't matter for that source because there is no load.
780  if ((LHSV || getKnownAlignment(LHS, DL, CI) >= PrefAlignment) &&
781  (RHSV || getKnownAlignment(RHS, DL, CI) >= PrefAlignment)) {
782  if (!LHSV) {
783  Type *LHSPtrTy =
784  IntType->getPointerTo(LHS->getType()->getPointerAddressSpace());
785  LHSV = B.CreateLoad(B.CreateBitCast(LHS, LHSPtrTy), "lhsv");
786  }
787  if (!RHSV) {
788  Type *RHSPtrTy =
789  IntType->getPointerTo(RHS->getType()->getPointerAddressSpace());
790  RHSV = B.CreateLoad(B.CreateBitCast(RHS, RHSPtrTy), "rhsv");
791  }
792  return B.CreateZExt(B.CreateICmpNE(LHSV, RHSV), CI->getType(), "memcmp");
793  }
794  }
795 
796  // Constant folding: memcmp(x, y, Len) -> constant (all arguments are const).
797  // TODO: This is limited to i8 arrays.
798  StringRef LHSStr, RHSStr;
799  if (getConstantStringInfo(LHS, LHSStr) &&
800  getConstantStringInfo(RHS, RHSStr)) {
801  // Make sure we're not reading out-of-bounds memory.
802  if (Len > LHSStr.size() || Len > RHSStr.size())
803  return nullptr;
804  // Fold the memcmp and normalize the result. This way we get consistent
805  // results across multiple platforms.
806  uint64_t Ret = 0;
807  int Cmp = memcmp(LHSStr.data(), RHSStr.data(), Len);
808  if (Cmp < 0)
809  Ret = -1;
810  else if (Cmp > 0)
811  Ret = 1;
812  return ConstantInt::get(CI->getType(), Ret);
813  }
814 
815  return nullptr;
816 }
817 
818 Value *LibCallSimplifier::optimizeMemCpy(CallInst *CI, IRBuilder<> &B) {
819  // memcpy(x, y, n) -> llvm.memcpy(x, y, n, 1)
820  B.CreateMemCpy(CI->getArgOperand(0), CI->getArgOperand(1),
821  CI->getArgOperand(2), 1);
822  return CI->getArgOperand(0);
823 }
824 
825 Value *LibCallSimplifier::optimizeMemMove(CallInst *CI, IRBuilder<> &B) {
826  // memmove(x, y, n) -> llvm.memmove(x, y, n, 1)
827  B.CreateMemMove(CI->getArgOperand(0), CI->getArgOperand(1),
828  CI->getArgOperand(2), 1);
829  return CI->getArgOperand(0);
830 }
831 
832 // TODO: Does this belong in BuildLibCalls or should all of those similar
833 // functions be moved here?
835  IRBuilder<> &B, const TargetLibraryInfo &TLI) {
836  LibFunc Func;
837  if (!TLI.getLibFunc("calloc", Func) || !TLI.has(Func))
838  return nullptr;
839 
840  Module *M = B.GetInsertBlock()->getModule();
841  const DataLayout &DL = M->getDataLayout();
842  IntegerType *PtrType = DL.getIntPtrType((B.GetInsertBlock()->getContext()));
843  Value *Calloc = M->getOrInsertFunction("calloc", Attrs, B.getInt8PtrTy(),
844  PtrType, PtrType);
845  CallInst *CI = B.CreateCall(Calloc, { Num, Size }, "calloc");
846 
847  if (const auto *F = dyn_cast<Function>(Calloc->stripPointerCasts()))
848  CI->setCallingConv(F->getCallingConv());
849 
850  return CI;
851 }
852 
853 /// Fold memset[_chk](malloc(n), 0, n) --> calloc(1, n).
855  const TargetLibraryInfo &TLI) {
856  // This has to be a memset of zeros (bzero).
857  auto *FillValue = dyn_cast<ConstantInt>(Memset->getArgOperand(1));
858  if (!FillValue || FillValue->getZExtValue() != 0)
859  return nullptr;
860 
861  // TODO: We should handle the case where the malloc has more than one use.
862  // This is necessary to optimize common patterns such as when the result of
863  // the malloc is checked against null or when a memset intrinsic is used in
864  // place of a memset library call.
865  auto *Malloc = dyn_cast<CallInst>(Memset->getArgOperand(0));
866  if (!Malloc || !Malloc->hasOneUse())
867  return nullptr;
868 
869  // Is the inner call really malloc()?
870  Function *InnerCallee = Malloc->getCalledFunction();
871  if (!InnerCallee)
872  return nullptr;
873 
874  LibFunc Func;
875  if (!TLI.getLibFunc(*InnerCallee, Func) || !TLI.has(Func) ||
876  Func != LibFunc_malloc)
877  return nullptr;
878 
879  // The memset must cover the same number of bytes that are malloc'd.
880  if (Memset->getArgOperand(2) != Malloc->getArgOperand(0))
881  return nullptr;
882 
883  // Replace the malloc with a calloc. We need the data layout to know what the
884  // actual size of a 'size_t' parameter is.
885  B.SetInsertPoint(Malloc->getParent(), ++Malloc->getIterator());
886  const DataLayout &DL = Malloc->getModule()->getDataLayout();
887  IntegerType *SizeType = DL.getIntPtrType(B.GetInsertBlock()->getContext());
888  Value *Calloc = emitCalloc(ConstantInt::get(SizeType, 1),
889  Malloc->getArgOperand(0), Malloc->getAttributes(),
890  B, TLI);
891  if (!Calloc)
892  return nullptr;
893 
894  Malloc->replaceAllUsesWith(Calloc);
895  Malloc->eraseFromParent();
896 
897  return Calloc;
898 }
899 
900 Value *LibCallSimplifier::optimizeMemSet(CallInst *CI, IRBuilder<> &B) {
901  if (auto *Calloc = foldMallocMemset(CI, B, *TLI))
902  return Calloc;
903 
904  // memset(p, v, n) -> llvm.memset(p, v, n, 1)
905  Value *Val = B.CreateIntCast(CI->getArgOperand(1), B.getInt8Ty(), false);
906  B.CreateMemSet(CI->getArgOperand(0), Val, CI->getArgOperand(2), 1);
907  return CI->getArgOperand(0);
908 }
909 
910 //===----------------------------------------------------------------------===//
911 // Math Library Optimizations
912 //===----------------------------------------------------------------------===//
913 
914 /// Return a variant of Val with float type.
915 /// Currently this works in two cases: If Val is an FPExtension of a float
916 /// value to something bigger, simply return the operand.
917 /// If Val is a ConstantFP but can be converted to a float ConstantFP without
918 /// loss of precision do so.
920  if (FPExtInst *Cast = dyn_cast<FPExtInst>(Val)) {
921  Value *Op = Cast->getOperand(0);
922  if (Op->getType()->isFloatTy())
923  return Op;
924  }
925  if (ConstantFP *Const = dyn_cast<ConstantFP>(Val)) {
926  APFloat F = Const->getValueAPF();
927  bool losesInfo;
929  &losesInfo);
930  if (!losesInfo)
931  return ConstantFP::get(Const->getContext(), F);
932  }
933  return nullptr;
934 }
935 
936 /// Shrink double -> float for unary functions like 'floor'.
938  bool CheckRetType) {
940  // We know this libcall has a valid prototype, but we don't know which.
941  if (!CI->getType()->isDoubleTy())
942  return nullptr;
943 
944  if (CheckRetType) {
945  // Check if all the uses for function like 'sin' are converted to float.
946  for (User *U : CI->users()) {
947  FPTruncInst *Cast = dyn_cast<FPTruncInst>(U);
948  if (!Cast || !Cast->getType()->isFloatTy())
949  return nullptr;
950  }
951  }
952 
953  // If this is something like 'floor((double)floatval)', convert to floorf.
955  if (V == nullptr)
956  return nullptr;
957 
958  // If call isn't an intrinsic, check that it isn't within a function with the
959  // same name as the float version of this call.
960  //
961  // e.g. inline float expf(float val) { return (float) exp((double) val); }
962  //
963  // A similar such definition exists in the MinGW-w64 math.h header file which
964  // when compiled with -O2 -ffast-math causes the generation of infinite loops
965  // where expf is called.
966  if (!Callee->isIntrinsic()) {
967  const Function *F = CI->getFunction();
968  StringRef FName = F->getName();
969  StringRef CalleeName = Callee->getName();
970  if ((FName.size() == (CalleeName.size() + 1)) &&
971  (FName.back() == 'f') &&
972  FName.startswith(CalleeName))
973  return nullptr;
974  }
975 
976  // Propagate fast-math flags from the existing call to the new call.
979 
980  // floor((double)floatval) -> (double)floorf(floatval)
981  if (Callee->isIntrinsic()) {
982  Module *M = CI->getModule();
983  Intrinsic::ID IID = Callee->getIntrinsicID();
985  V = B.CreateCall(F, V);
986  } else {
987  // The call is a library call rather than an intrinsic.
988  V = emitUnaryFloatFnCall(V, Callee->getName(), B, Callee->getAttributes());
989  }
990 
991  return B.CreateFPExt(V, B.getDoubleTy());
992 }
993 
994 // Replace a libcall \p CI with a call to intrinsic \p IID
996  // Propagate fast-math flags from the existing call to the new call.
999 
1000  Module *M = CI->getModule();
1001  Value *V = CI->getArgOperand(0);
1002  Function *F = Intrinsic::getDeclaration(M, IID, CI->getType());
1003  CallInst *NewCall = B.CreateCall(F, V);
1004  NewCall->takeName(CI);
1005  return NewCall;
1006 }
1007 
1008 /// Shrink double -> float for binary functions like 'fmin/fmax'.
1011  // We know this libcall has a valid prototype, but we don't know which.
1012  if (!CI->getType()->isDoubleTy())
1013  return nullptr;
1014 
1015  // If this is something like 'fmin((double)floatval1, (double)floatval2)',
1016  // or fmin(1.0, (double)floatval), then we convert it to fminf.
1018  if (V1 == nullptr)
1019  return nullptr;
1021  if (V2 == nullptr)
1022  return nullptr;
1023 
1024  // Propagate fast-math flags from the existing call to the new call.
1027 
1028  // fmin((double)floatval1, (double)floatval2)
1029  // -> (double)fminf(floatval1, floatval2)
1030  // TODO: Handle intrinsics in the same way as in optimizeUnaryDoubleFP().
1031  Value *V = emitBinaryFloatFnCall(V1, V2, Callee->getName(), B,
1032  Callee->getAttributes());
1033  return B.CreateFPExt(V, B.getDoubleTy());
1034 }
1035 
1036 Value *LibCallSimplifier::optimizeCos(CallInst *CI, IRBuilder<> &B) {
1038  Value *Ret = nullptr;
1039  StringRef Name = Callee->getName();
1040  if (UnsafeFPShrink && Name == "cos" && hasFloatVersion(Name))
1041  Ret = optimizeUnaryDoubleFP(CI, B, true);
1042 
1043  // cos(-x) -> cos(x)
1044  Value *Op1 = CI->getArgOperand(0);
1045  if (BinaryOperator::isFNeg(Op1)) {
1046  BinaryOperator *BinExpr = cast<BinaryOperator>(Op1);
1047  return B.CreateCall(Callee, BinExpr->getOperand(1), "cos");
1048  }
1049  return Ret;
1050 }
1051 
1052 static Value *getPow(Value *InnerChain[33], unsigned Exp, IRBuilder<> &B) {
1053  // Multiplications calculated using Addition Chains.
1054  // Refer: http://wwwhomes.uni-bielefeld.de/achim/addition_chain.html
1055 
1056  assert(Exp != 0 && "Incorrect exponent 0 not handled");
1057 
1058  if (InnerChain[Exp])
1059  return InnerChain[Exp];
1060 
1061  static const unsigned AddChain[33][2] = {
1062  {0, 0}, // Unused.
1063  {0, 0}, // Unused (base case = pow1).
1064  {1, 1}, // Unused (pre-computed).
1065  {1, 2}, {2, 2}, {2, 3}, {3, 3}, {2, 5}, {4, 4},
1066  {1, 8}, {5, 5}, {1, 10}, {6, 6}, {4, 9}, {7, 7},
1067  {3, 12}, {8, 8}, {8, 9}, {2, 16}, {1, 18}, {10, 10},
1068  {6, 15}, {11, 11}, {3, 20}, {12, 12}, {8, 17}, {13, 13},
1069  {3, 24}, {14, 14}, {4, 25}, {15, 15}, {3, 28}, {16, 16},
1070  };
1071 
1072  InnerChain[Exp] = B.CreateFMul(getPow(InnerChain, AddChain[Exp][0], B),
1073  getPow(InnerChain, AddChain[Exp][1], B));
1074  return InnerChain[Exp];
1075 }
1076 
1077 Value *LibCallSimplifier::optimizePow(CallInst *CI, IRBuilder<> &B) {
1079  Value *Ret = nullptr;
1080  StringRef Name = Callee->getName();
1081  if (UnsafeFPShrink && Name == "pow" && hasFloatVersion(Name))
1082  Ret = optimizeUnaryDoubleFP(CI, B, true);
1083 
1084  Value *Op1 = CI->getArgOperand(0), *Op2 = CI->getArgOperand(1);
1085 
1086  // pow(1.0, x) -> 1.0
1087  if (match(Op1, m_SpecificFP(1.0)))
1088  return Op1;
1089  // pow(2.0, x) -> llvm.exp2(x)
1090  if (match(Op1, m_SpecificFP(2.0))) {
1091  Value *Exp2 = Intrinsic::getDeclaration(CI->getModule(), Intrinsic::exp2,
1092  CI->getType());
1093  return B.CreateCall(Exp2, Op2, "exp2");
1094  }
1095 
1096  // There's no llvm.exp10 intrinsic yet, but, maybe, some day there will
1097  // be one.
1098  if (ConstantFP *Op1C = dyn_cast<ConstantFP>(Op1)) {
1099  // pow(10.0, x) -> exp10(x)
1100  if (Op1C->isExactlyValue(10.0) &&
1101  hasUnaryFloatFn(TLI, Op1->getType(), LibFunc_exp10, LibFunc_exp10f,
1102  LibFunc_exp10l))
1103  return emitUnaryFloatFnCall(Op2, TLI->getName(LibFunc_exp10), B,
1104  Callee->getAttributes());
1105  }
1106 
1107  // pow(exp(x), y) -> exp(x * y)
1108  // pow(exp2(x), y) -> exp2(x * y)
1109  // We enable these only with fast-math. Besides rounding differences, the
1110  // transformation changes overflow and underflow behavior quite dramatically.
1111  // Example: x = 1000, y = 0.001.
1112  // pow(exp(x), y) = pow(inf, 0.001) = inf, whereas exp(x*y) = exp(1).
1113  auto *OpC = dyn_cast<CallInst>(Op1);
1114  if (OpC && OpC->isFast() && CI->isFast()) {
1115  LibFunc Func;
1116  Function *OpCCallee = OpC->getCalledFunction();
1117  if (OpCCallee && TLI->getLibFunc(OpCCallee->getName(), Func) &&
1118  TLI->has(Func) && (Func == LibFunc_exp || Func == LibFunc_exp2)) {
1121  Value *FMul = B.CreateFMul(OpC->getArgOperand(0), Op2, "mul");
1122  return emitUnaryFloatFnCall(FMul, OpCCallee->getName(), B,
1123  OpCCallee->getAttributes());
1124  }
1125  }
1126 
1127  ConstantFP *Op2C = dyn_cast<ConstantFP>(Op2);
1128  if (!Op2C)
1129  return Ret;
1130 
1131  if (Op2C->getValueAPF().isZero()) // pow(x, 0.0) -> 1.0
1132  return ConstantFP::get(CI->getType(), 1.0);
1133 
1134  if (Op2C->isExactlyValue(-0.5) &&
1135  hasUnaryFloatFn(TLI, Op2->getType(), LibFunc_sqrt, LibFunc_sqrtf,
1136  LibFunc_sqrtl)) {
1137  // If -ffast-math:
1138  // pow(x, -0.5) -> 1.0 / sqrt(x)
1139  if (CI->isFast()) {
1142 
1143  // TODO: If the pow call is an intrinsic, we should lower to the sqrt
1144  // intrinsic, so we match errno semantics. We also should check that the
1145  // target can in fact lower the sqrt intrinsic -- we currently have no way
1146  // to ask this question other than asking whether the target has a sqrt
1147  // libcall, which is a sufficient but not necessary condition.
1148  Value *Sqrt = emitUnaryFloatFnCall(Op1, TLI->getName(LibFunc_sqrt), B,
1149  Callee->getAttributes());
1150 
1151  return B.CreateFDiv(ConstantFP::get(CI->getType(), 1.0), Sqrt, "sqrtrecip");
1152  }
1153  }
1154 
1155  if (Op2C->isExactlyValue(0.5) &&
1156  hasUnaryFloatFn(TLI, Op2->getType(), LibFunc_sqrt, LibFunc_sqrtf,
1157  LibFunc_sqrtl)) {
1158 
1159  // In -ffast-math, pow(x, 0.5) -> sqrt(x).
1160  if (CI->isFast()) {
1163 
1164  // TODO: As above, we should lower to the sqrt intrinsic if the pow is an
1165  // intrinsic, to match errno semantics.
1166  return emitUnaryFloatFnCall(Op1, TLI->getName(LibFunc_sqrt), B,
1167  Callee->getAttributes());
1168  }
1169 
1170  // Expand pow(x, 0.5) to (x == -infinity ? +infinity : fabs(sqrt(x))).
1171  // This is faster than calling pow, and still handles negative zero
1172  // and negative infinity correctly.
1173  // TODO: In finite-only mode, this could be just fabs(sqrt(x)).
1174  Value *Inf = ConstantFP::getInfinity(CI->getType());
1175  Value *NegInf = ConstantFP::getInfinity(CI->getType(), true);
1176 
1177  // TODO: As above, we should lower to the sqrt intrinsic if the pow is an
1178  // intrinsic, to match errno semantics.
1179  Value *Sqrt = emitUnaryFloatFnCall(Op1, "sqrt", B, Callee->getAttributes());
1180 
1181  Module *M = Callee->getParent();
1182  Function *FabsF = Intrinsic::getDeclaration(M, Intrinsic::fabs,
1183  CI->getType());
1184  Value *FAbs = B.CreateCall(FabsF, Sqrt);
1185 
1186  Value *FCmp = B.CreateFCmpOEQ(Op1, NegInf);
1187  Value *Sel = B.CreateSelect(FCmp, Inf, FAbs);
1188  return Sel;
1189  }
1190 
1191  if (Op2C->isExactlyValue(1.0)) // pow(x, 1.0) -> x
1192  return Op1;
1193  if (Op2C->isExactlyValue(2.0)) // pow(x, 2.0) -> x*x
1194  return B.CreateFMul(Op1, Op1, "pow2");
1195  if (Op2C->isExactlyValue(-1.0)) // pow(x, -1.0) -> 1.0/x
1196  return B.CreateFDiv(ConstantFP::get(CI->getType(), 1.0), Op1, "powrecip");
1197 
1198  // In -ffast-math, generate repeated fmul instead of generating pow(x, n).
1199  if (CI->isFast()) {
1200  APFloat V = abs(Op2C->getValueAPF());
1201  // We limit to a max of 7 fmul(s). Thus max exponent is 32.
1202  // This transformation applies to integer exponents only.
1203  if (V.compare(APFloat(V.getSemantics(), 32.0)) == APFloat::cmpGreaterThan ||
1204  !V.isInteger())
1205  return nullptr;
1206 
1207  // Propagate fast math flags.
1210 
1211  // We will memoize intermediate products of the Addition Chain.
1212  Value *InnerChain[33] = {nullptr};
1213  InnerChain[1] = Op1;
1214  InnerChain[2] = B.CreateFMul(Op1, Op1);
1215 
1216  // We cannot readily convert a non-double type (like float) to a double.
1217  // So we first convert V to something which could be converted to double.
1218  bool ignored;
1220 
1221  Value *FMul = getPow(InnerChain, V.convertToDouble(), B);
1222  // For negative exponents simply compute the reciprocal.
1223  if (Op2C->isNegative())
1224  FMul = B.CreateFDiv(ConstantFP::get(CI->getType(), 1.0), FMul);
1225  return FMul;
1226  }
1227 
1228  return nullptr;
1229 }
1230 
1231 Value *LibCallSimplifier::optimizeExp2(CallInst *CI, IRBuilder<> &B) {
1233  Value *Ret = nullptr;
1234  StringRef Name = Callee->getName();
1235  if (UnsafeFPShrink && Name == "exp2" && hasFloatVersion(Name))
1236  Ret = optimizeUnaryDoubleFP(CI, B, true);
1237 
1238  Value *Op = CI->getArgOperand(0);
1239  // Turn exp2(sitofp(x)) -> ldexp(1.0, sext(x)) if sizeof(x) <= 32
1240  // Turn exp2(uitofp(x)) -> ldexp(1.0, zext(x)) if sizeof(x) < 32
1241  LibFunc LdExp = LibFunc_ldexpl;
1242  if (Op->getType()->isFloatTy())
1243  LdExp = LibFunc_ldexpf;
1244  else if (Op->getType()->isDoubleTy())
1245  LdExp = LibFunc_ldexp;
1246 
1247  if (TLI->has(LdExp)) {
1248  Value *LdExpArg = nullptr;
1249  if (SIToFPInst *OpC = dyn_cast<SIToFPInst>(Op)) {
1250  if (OpC->getOperand(0)->getType()->getPrimitiveSizeInBits() <= 32)
1251  LdExpArg = B.CreateSExt(OpC->getOperand(0), B.getInt32Ty());
1252  } else if (UIToFPInst *OpC = dyn_cast<UIToFPInst>(Op)) {
1253  if (OpC->getOperand(0)->getType()->getPrimitiveSizeInBits() < 32)
1254  LdExpArg = B.CreateZExt(OpC->getOperand(0), B.getInt32Ty());
1255  }
1256 
1257  if (LdExpArg) {
1258  Constant *One = ConstantFP::get(CI->getContext(), APFloat(1.0f));
1259  if (!Op->getType()->isFloatTy())
1260  One = ConstantExpr::getFPExtend(One, Op->getType());
1261 
1262  Module *M = CI->getModule();
1263  Value *NewCallee =
1264  M->getOrInsertFunction(TLI->getName(LdExp), Op->getType(),
1265  Op->getType(), B.getInt32Ty());
1266  CallInst *CI = B.CreateCall(NewCallee, {One, LdExpArg});
1267  if (const Function *F = dyn_cast<Function>(Callee->stripPointerCasts()))
1268  CI->setCallingConv(F->getCallingConv());
1269 
1270  return CI;
1271  }
1272  }
1273  return Ret;
1274 }
1275 
1276 Value *LibCallSimplifier::optimizeFMinFMax(CallInst *CI, IRBuilder<> &B) {
1278  // If we can shrink the call to a float function rather than a double
1279  // function, do that first.
1280  StringRef Name = Callee->getName();
1281  if ((Name == "fmin" || Name == "fmax") && hasFloatVersion(Name))
1282  if (Value *Ret = optimizeBinaryDoubleFP(CI, B))
1283  return Ret;
1284 
1286  FastMathFlags FMF;
1287  if (CI->isFast()) {
1288  // If the call is 'fast', then anything we create here will also be 'fast'.
1289  FMF.setFast();
1290  } else {
1291  // At a minimum, no-nans-fp-math must be true.
1292  if (!CI->hasNoNaNs())
1293  return nullptr;
1294  // No-signed-zeros is implied by the definitions of fmax/fmin themselves:
1295  // "Ideally, fmax would be sensitive to the sign of zero, for example
1296  // fmax(-0. 0, +0. 0) would return +0; however, implementation in software
1297  // might be impractical."
1298  FMF.setNoSignedZeros();
1299  FMF.setNoNaNs();
1300  }
1301  B.setFastMathFlags(FMF);
1302 
1303  // We have a relaxed floating-point environment. We can ignore NaN-handling
1304  // and transform to a compare and select. We do not have to consider errno or
1305  // exceptions, because fmin/fmax do not have those.
1306  Value *Op0 = CI->getArgOperand(0);
1307  Value *Op1 = CI->getArgOperand(1);
1308  Value *Cmp = Callee->getName().startswith("fmin") ?
1309  B.CreateFCmpOLT(Op0, Op1) : B.CreateFCmpOGT(Op0, Op1);
1310  return B.CreateSelect(Cmp, Op0, Op1);
1311 }
1312 
1313 Value *LibCallSimplifier::optimizeLog(CallInst *CI, IRBuilder<> &B) {
1315  Value *Ret = nullptr;
1316  StringRef Name = Callee->getName();
1317  if (UnsafeFPShrink && hasFloatVersion(Name))
1318  Ret = optimizeUnaryDoubleFP(CI, B, true);
1319 
1320  if (!CI->isFast())
1321  return Ret;
1322  Value *Op1 = CI->getArgOperand(0);
1323  auto *OpC = dyn_cast<CallInst>(Op1);
1324 
1325  // The earlier call must also be 'fast' in order to do these transforms.
1326  if (!OpC || !OpC->isFast())
1327  return Ret;
1328 
1329  // log(pow(x,y)) -> y*log(x)
1330  // This is only applicable to log, log2, log10.
1331  if (Name != "log" && Name != "log2" && Name != "log10")
1332  return Ret;
1333 
1335  FastMathFlags FMF;
1336  FMF.setFast();
1337  B.setFastMathFlags(FMF);
1338 
1339  LibFunc Func;
1340  Function *F = OpC->getCalledFunction();
1341  if (F && ((TLI->getLibFunc(F->getName(), Func) && TLI->has(Func) &&
1342  Func == LibFunc_pow) || F->getIntrinsicID() == Intrinsic::pow))
1343  return B.CreateFMul(OpC->getArgOperand(1),
1344  emitUnaryFloatFnCall(OpC->getOperand(0), Callee->getName(), B,
1345  Callee->getAttributes()), "mul");
1346 
1347  // log(exp2(y)) -> y*log(2)
1348  if (F && Name == "log" && TLI->getLibFunc(F->getName(), Func) &&
1349  TLI->has(Func) && Func == LibFunc_exp2)
1350  return B.CreateFMul(
1351  OpC->getArgOperand(0),
1353  Callee->getName(), B, Callee->getAttributes()),
1354  "logmul");
1355  return Ret;
1356 }
1357 
1358 Value *LibCallSimplifier::optimizeSqrt(CallInst *CI, IRBuilder<> &B) {
1360  Value *Ret = nullptr;
1361  // TODO: Once we have a way (other than checking for the existince of the
1362  // libcall) to tell whether our target can lower @llvm.sqrt, relax the
1363  // condition below.
1364  if (TLI->has(LibFunc_sqrtf) && (Callee->getName() == "sqrt" ||
1365  Callee->getIntrinsicID() == Intrinsic::sqrt))
1366  Ret = optimizeUnaryDoubleFP(CI, B, true);
1367 
1368  if (!CI->isFast())
1369  return Ret;
1370 
1372  if (!I || I->getOpcode() != Instruction::FMul || !I->isFast())
1373  return Ret;
1374 
1375  // We're looking for a repeated factor in a multiplication tree,
1376  // so we can do this fold: sqrt(x * x) -> fabs(x);
1377  // or this fold: sqrt((x * x) * y) -> fabs(x) * sqrt(y).
1378  Value *Op0 = I->getOperand(0);
1379  Value *Op1 = I->getOperand(1);
1380  Value *RepeatOp = nullptr;
1381  Value *OtherOp = nullptr;
1382  if (Op0 == Op1) {
1383  // Simple match: the operands of the multiply are identical.
1384  RepeatOp = Op0;
1385  } else {
1386  // Look for a more complicated pattern: one of the operands is itself
1387  // a multiply, so search for a common factor in that multiply.
1388  // Note: We don't bother looking any deeper than this first level or for
1389  // variations of this pattern because instcombine's visitFMUL and/or the
1390  // reassociation pass should give us this form.
1391  Value *OtherMul0, *OtherMul1;
1392  if (match(Op0, m_FMul(m_Value(OtherMul0), m_Value(OtherMul1)))) {
1393  // Pattern: sqrt((x * y) * z)
1394  if (OtherMul0 == OtherMul1 && cast<Instruction>(Op0)->isFast()) {
1395  // Matched: sqrt((x * x) * z)
1396  RepeatOp = OtherMul0;
1397  OtherOp = Op1;
1398  }
1399  }
1400  }
1401  if (!RepeatOp)
1402  return Ret;
1403 
1404  // Fast math flags for any created instructions should match the sqrt
1405  // and multiply.
1408 
1409  // If we found a repeated factor, hoist it out of the square root and
1410  // replace it with the fabs of that factor.
1411  Module *M = Callee->getParent();
1412  Type *ArgType = I->getType();
1413  Value *Fabs = Intrinsic::getDeclaration(M, Intrinsic::fabs, ArgType);
1414  Value *FabsCall = B.CreateCall(Fabs, RepeatOp, "fabs");
1415  if (OtherOp) {
1416  // If we found a non-repeated factor, we still need to get its square
1417  // root. We then multiply that by the value that was simplified out
1418  // of the square root calculation.
1419  Value *Sqrt = Intrinsic::getDeclaration(M, Intrinsic::sqrt, ArgType);
1420  Value *SqrtCall = B.CreateCall(Sqrt, OtherOp, "sqrt");
1421  return B.CreateFMul(FabsCall, SqrtCall);
1422  }
1423  return FabsCall;
1424 }
1425 
1426 // TODO: Generalize to handle any trig function and its inverse.
1427 Value *LibCallSimplifier::optimizeTan(CallInst *CI, IRBuilder<> &B) {
1429  Value *Ret = nullptr;
1430  StringRef Name = Callee->getName();
1431  if (UnsafeFPShrink && Name == "tan" && hasFloatVersion(Name))
1432  Ret = optimizeUnaryDoubleFP(CI, B, true);
1433 
1434  Value *Op1 = CI->getArgOperand(0);
1435  auto *OpC = dyn_cast<CallInst>(Op1);
1436  if (!OpC)
1437  return Ret;
1438 
1439  // Both calls must be 'fast' in order to remove them.
1440  if (!CI->isFast() || !OpC->isFast())
1441  return Ret;
1442 
1443  // tan(atan(x)) -> x
1444  // tanf(atanf(x)) -> x
1445  // tanl(atanl(x)) -> x
1446  LibFunc Func;
1447  Function *F = OpC->getCalledFunction();
1448  if (F && TLI->getLibFunc(F->getName(), Func) && TLI->has(Func) &&
1449  ((Func == LibFunc_atan && Callee->getName() == "tan") ||
1450  (Func == LibFunc_atanf && Callee->getName() == "tanf") ||
1451  (Func == LibFunc_atanl && Callee->getName() == "tanl")))
1452  Ret = OpC->getArgOperand(0);
1453  return Ret;
1454 }
1455 
1456 static bool isTrigLibCall(CallInst *CI) {
1457  // We can only hope to do anything useful if we can ignore things like errno
1458  // and floating-point exceptions.
1459  // We already checked the prototype.
1460  return CI->hasFnAttr(Attribute::NoUnwind) &&
1461  CI->hasFnAttr(Attribute::ReadNone);
1462 }
1463 
1464 static void insertSinCosCall(IRBuilder<> &B, Function *OrigCallee, Value *Arg,
1465  bool UseFloat, Value *&Sin, Value *&Cos,
1466  Value *&SinCos) {
1467  Type *ArgTy = Arg->getType();
1468  Type *ResTy;
1469  StringRef Name;
1470 
1471  Triple T(OrigCallee->getParent()->getTargetTriple());
1472  if (UseFloat) {
1473  Name = "__sincospif_stret";
1474 
1475  assert(T.getArch() != Triple::x86 && "x86 messy and unsupported for now");
1476  // x86_64 can't use {float, float} since that would be returned in both
1477  // xmm0 and xmm1, which isn't what a real struct would do.
1478  ResTy = T.getArch() == Triple::x86_64
1479  ? static_cast<Type *>(VectorType::get(ArgTy, 2))
1480  : static_cast<Type *>(StructType::get(ArgTy, ArgTy));
1481  } else {
1482  Name = "__sincospi_stret";
1483  ResTy = StructType::get(ArgTy, ArgTy);
1484  }
1485 
1486  Module *M = OrigCallee->getParent();
1487  Value *Callee = M->getOrInsertFunction(Name, OrigCallee->getAttributes(),
1488  ResTy, ArgTy);
1489 
1490  if (Instruction *ArgInst = dyn_cast<Instruction>(Arg)) {
1491  // If the argument is an instruction, it must dominate all uses so put our
1492  // sincos call there.
1493  B.SetInsertPoint(ArgInst->getParent(), ++ArgInst->getIterator());
1494  } else {
1495  // Otherwise (e.g. for a constant) the beginning of the function is as
1496  // good a place as any.
1497  BasicBlock &EntryBB = B.GetInsertBlock()->getParent()->getEntryBlock();
1498  B.SetInsertPoint(&EntryBB, EntryBB.begin());
1499  }
1500 
1501  SinCos = B.CreateCall(Callee, Arg, "sincospi");
1502 
1503  if (SinCos->getType()->isStructTy()) {
1504  Sin = B.CreateExtractValue(SinCos, 0, "sinpi");
1505  Cos = B.CreateExtractValue(SinCos, 1, "cospi");
1506  } else {
1507  Sin = B.CreateExtractElement(SinCos, ConstantInt::get(B.getInt32Ty(), 0),
1508  "sinpi");
1509  Cos = B.CreateExtractElement(SinCos, ConstantInt::get(B.getInt32Ty(), 1),
1510  "cospi");
1511  }
1512 }
1513 
1514 Value *LibCallSimplifier::optimizeSinCosPi(CallInst *CI, IRBuilder<> &B) {
1515  // Make sure the prototype is as expected, otherwise the rest of the
1516  // function is probably invalid and likely to abort.
1517  if (!isTrigLibCall(CI))
1518  return nullptr;
1519 
1520  Value *Arg = CI->getArgOperand(0);
1521  SmallVector<CallInst *, 1> SinCalls;
1522  SmallVector<CallInst *, 1> CosCalls;
1523  SmallVector<CallInst *, 1> SinCosCalls;
1524 
1525  bool IsFloat = Arg->getType()->isFloatTy();
1526 
1527  // Look for all compatible sinpi, cospi and sincospi calls with the same
1528  // argument. If there are enough (in some sense) we can make the
1529  // substitution.
1530  Function *F = CI->getFunction();
1531  for (User *U : Arg->users())
1532  classifyArgUse(U, F, IsFloat, SinCalls, CosCalls, SinCosCalls);
1533 
1534  // It's only worthwhile if both sinpi and cospi are actually used.
1535  if (SinCosCalls.empty() && (SinCalls.empty() || CosCalls.empty()))
1536  return nullptr;
1537 
1538  Value *Sin, *Cos, *SinCos;
1539  insertSinCosCall(B, CI->getCalledFunction(), Arg, IsFloat, Sin, Cos, SinCos);
1540 
1541  auto replaceTrigInsts = [this](SmallVectorImpl<CallInst *> &Calls,
1542  Value *Res) {
1543  for (CallInst *C : Calls)
1544  replaceAllUsesWith(C, Res);
1545  };
1546 
1547  replaceTrigInsts(SinCalls, Sin);
1548  replaceTrigInsts(CosCalls, Cos);
1549  replaceTrigInsts(SinCosCalls, SinCos);
1550 
1551  return nullptr;
1552 }
1553 
1554 void LibCallSimplifier::classifyArgUse(
1555  Value *Val, Function *F, bool IsFloat,
1556  SmallVectorImpl<CallInst *> &SinCalls,
1557  SmallVectorImpl<CallInst *> &CosCalls,
1558  SmallVectorImpl<CallInst *> &SinCosCalls) {
1559  CallInst *CI = dyn_cast<CallInst>(Val);
1560 
1561  if (!CI)
1562  return;
1563 
1564  // Don't consider calls in other functions.
1565  if (CI->getFunction() != F)
1566  return;
1567 
1569  LibFunc Func;
1570  if (!Callee || !TLI->getLibFunc(*Callee, Func) || !TLI->has(Func) ||
1571  !isTrigLibCall(CI))
1572  return;
1573 
1574  if (IsFloat) {
1575  if (Func == LibFunc_sinpif)
1576  SinCalls.push_back(CI);
1577  else if (Func == LibFunc_cospif)
1578  CosCalls.push_back(CI);
1579  else if (Func == LibFunc_sincospif_stret)
1580  SinCosCalls.push_back(CI);
1581  } else {
1582  if (Func == LibFunc_sinpi)
1583  SinCalls.push_back(CI);
1584  else if (Func == LibFunc_cospi)
1585  CosCalls.push_back(CI);
1586  else if (Func == LibFunc_sincospi_stret)
1587  SinCosCalls.push_back(CI);
1588  }
1589 }
1590 
1591 //===----------------------------------------------------------------------===//
1592 // Integer Library Call Optimizations
1593 //===----------------------------------------------------------------------===//
1594 
1595 Value *LibCallSimplifier::optimizeFFS(CallInst *CI, IRBuilder<> &B) {
1596  // ffs(x) -> x != 0 ? (i32)llvm.cttz(x)+1 : 0
1597  Value *Op = CI->getArgOperand(0);
1598  Type *ArgType = Op->getType();
1600  Intrinsic::cttz, ArgType);
1601  Value *V = B.CreateCall(F, {Op, B.getTrue()}, "cttz");
1602  V = B.CreateAdd(V, ConstantInt::get(V->getType(), 1));
1603  V = B.CreateIntCast(V, B.getInt32Ty(), false);
1604 
1605  Value *Cond = B.CreateICmpNE(Op, Constant::getNullValue(ArgType));
1606  return B.CreateSelect(Cond, V, B.getInt32(0));
1607 }
1608 
1609 Value *LibCallSimplifier::optimizeFls(CallInst *CI, IRBuilder<> &B) {
1610  // fls(x) -> (i32)(sizeInBits(x) - llvm.ctlz(x, false))
1611  Value *Op = CI->getArgOperand(0);
1612  Type *ArgType = Op->getType();
1614  Intrinsic::ctlz, ArgType);
1615  Value *V = B.CreateCall(F, {Op, B.getFalse()}, "ctlz");
1616  V = B.CreateSub(ConstantInt::get(V->getType(), ArgType->getIntegerBitWidth()),
1617  V);
1618  return B.CreateIntCast(V, CI->getType(), false);
1619 }
1620 
1621 Value *LibCallSimplifier::optimizeAbs(CallInst *CI, IRBuilder<> &B) {
1622  // abs(x) -> x >s -1 ? x : -x
1623  Value *Op = CI->getArgOperand(0);
1624  Value *Pos =
1625  B.CreateICmpSGT(Op, Constant::getAllOnesValue(Op->getType()), "ispos");
1626  Value *Neg = B.CreateNeg(Op, "neg");
1627  return B.CreateSelect(Pos, Op, Neg);
1628 }
1629 
1630 Value *LibCallSimplifier::optimizeIsDigit(CallInst *CI, IRBuilder<> &B) {
1631  // isdigit(c) -> (c-'0') <u 10
1632  Value *Op = CI->getArgOperand(0);
1633  Op = B.CreateSub(Op, B.getInt32('0'), "isdigittmp");
1634  Op = B.CreateICmpULT(Op, B.getInt32(10), "isdigit");
1635  return B.CreateZExt(Op, CI->getType());
1636 }
1637 
1638 Value *LibCallSimplifier::optimizeIsAscii(CallInst *CI, IRBuilder<> &B) {
1639  // isascii(c) -> c <u 128
1640  Value *Op = CI->getArgOperand(0);
1641  Op = B.CreateICmpULT(Op, B.getInt32(128), "isascii");
1642  return B.CreateZExt(Op, CI->getType());
1643 }
1644 
1645 Value *LibCallSimplifier::optimizeToAscii(CallInst *CI, IRBuilder<> &B) {
1646  // toascii(c) -> c & 0x7f
1647  return B.CreateAnd(CI->getArgOperand(0),
1648  ConstantInt::get(CI->getType(), 0x7F));
1649 }
1650 
1651 //===----------------------------------------------------------------------===//
1652 // Formatting and IO Library Call Optimizations
1653 //===----------------------------------------------------------------------===//
1654 
1655 static bool isReportingError(Function *Callee, CallInst *CI, int StreamArg);
1656 
1657 Value *LibCallSimplifier::optimizeErrorReporting(CallInst *CI, IRBuilder<> &B,
1658  int StreamArg) {
1659  Function *Callee = CI->getCalledFunction();
1660  // Error reporting calls should be cold, mark them as such.
1661  // This applies even to non-builtin calls: it is only a hint and applies to
1662  // functions that the frontend might not understand as builtins.
1663 
1664  // This heuristic was suggested in:
1665  // Improving Static Branch Prediction in a Compiler
1666  // Brian L. Deitrich, Ben-Chung Cheng, Wen-mei W. Hwu
1667  // Proceedings of PACT'98, Oct. 1998, IEEE
1668  if (!CI->hasFnAttr(Attribute::Cold) &&
1669  isReportingError(Callee, CI, StreamArg)) {
1671  }
1672 
1673  return nullptr;
1674 }
1675 
1676 static bool isReportingError(Function *Callee, CallInst *CI, int StreamArg) {
1677  if (!Callee || !Callee->isDeclaration())
1678  return false;
1679 
1680  if (StreamArg < 0)
1681  return true;
1682 
1683  // These functions might be considered cold, but only if their stream
1684  // argument is stderr.
1685 
1686  if (StreamArg >= (int)CI->getNumArgOperands())
1687  return false;
1688  LoadInst *LI = dyn_cast<LoadInst>(CI->getArgOperand(StreamArg));
1689  if (!LI)
1690  return false;
1692  if (!GV || !GV->isDeclaration())
1693  return false;
1694  return GV->getName() == "stderr";
1695 }
1696 
1697 Value *LibCallSimplifier::optimizePrintFString(CallInst *CI, IRBuilder<> &B) {
1698  // Check for a fixed format string.
1699  StringRef FormatStr;
1700  if (!getConstantStringInfo(CI->getArgOperand(0), FormatStr))
1701  return nullptr;
1702 
1703  // Empty format string -> noop.
1704  if (FormatStr.empty()) // Tolerate printf's declared void.
1705  return CI->use_empty() ? (Value *)CI : ConstantInt::get(CI->getType(), 0);
1706 
1707  // Do not do any of the following transformations if the printf return value
1708  // is used, in general the printf return value is not compatible with either
1709  // putchar() or puts().
1710  if (!CI->use_empty())
1711  return nullptr;
1712 
1713  // printf("x") -> putchar('x'), even for "%" and "%%".
1714  if (FormatStr.size() == 1 || FormatStr == "%%")
1715  return emitPutChar(B.getInt32(FormatStr[0]), B, TLI);
1716 
1717  // printf("%s", "a") --> putchar('a')
1718  if (FormatStr == "%s" && CI->getNumArgOperands() > 1) {
1719  StringRef ChrStr;
1720  if (!getConstantStringInfo(CI->getOperand(1), ChrStr))
1721  return nullptr;
1722  if (ChrStr.size() != 1)
1723  return nullptr;
1724  return emitPutChar(B.getInt32(ChrStr[0]), B, TLI);
1725  }
1726 
1727  // printf("foo\n") --> puts("foo")
1728  if (FormatStr[FormatStr.size() - 1] == '\n' &&
1729  FormatStr.find('%') == StringRef::npos) { // No format characters.
1730  // Create a string literal with no \n on it. We expect the constant merge
1731  // pass to be run after this pass, to merge duplicate strings.
1732  FormatStr = FormatStr.drop_back();
1733  Value *GV = B.CreateGlobalString(FormatStr, "str");
1734  return emitPutS(GV, B, TLI);
1735  }
1736 
1737  // Optimize specific format strings.
1738  // printf("%c", chr) --> putchar(chr)
1739  if (FormatStr == "%c" && CI->getNumArgOperands() > 1 &&
1740  CI->getArgOperand(1)->getType()->isIntegerTy())
1741  return emitPutChar(CI->getArgOperand(1), B, TLI);
1742 
1743  // printf("%s\n", str) --> puts(str)
1744  if (FormatStr == "%s\n" && CI->getNumArgOperands() > 1 &&
1745  CI->getArgOperand(1)->getType()->isPointerTy())
1746  return emitPutS(CI->getArgOperand(1), B, TLI);
1747  return nullptr;
1748 }
1749 
1750 Value *LibCallSimplifier::optimizePrintF(CallInst *CI, IRBuilder<> &B) {
1751 
1752  Function *Callee = CI->getCalledFunction();
1753  FunctionType *FT = Callee->getFunctionType();
1754  if (Value *V = optimizePrintFString(CI, B)) {
1755  return V;
1756  }
1757 
1758  // printf(format, ...) -> iprintf(format, ...) if no floating point
1759  // arguments.
1760  if (TLI->has(LibFunc_iprintf) && !callHasFloatingPointArgument(CI)) {
1761  Module *M = B.GetInsertBlock()->getParent()->getParent();
1762  Constant *IPrintFFn =
1763  M->getOrInsertFunction("iprintf", FT, Callee->getAttributes());
1764  CallInst *New = cast<CallInst>(CI->clone());
1765  New->setCalledFunction(IPrintFFn);
1766  B.Insert(New);
1767  return New;
1768  }
1769  return nullptr;
1770 }
1771 
1772 Value *LibCallSimplifier::optimizeSPrintFString(CallInst *CI, IRBuilder<> &B) {
1773  // Check for a fixed format string.
1774  StringRef FormatStr;
1775  if (!getConstantStringInfo(CI->getArgOperand(1), FormatStr))
1776  return nullptr;
1777 
1778  // If we just have a format string (nothing else crazy) transform it.
1779  if (CI->getNumArgOperands() == 2) {
1780  // Make sure there's no % in the constant array. We could try to handle
1781  // %% -> % in the future if we cared.
1782  for (unsigned i = 0, e = FormatStr.size(); i != e; ++i)
1783  if (FormatStr[i] == '%')
1784  return nullptr; // we found a format specifier, bail out.
1785 
1786  // sprintf(str, fmt) -> llvm.memcpy(str, fmt, strlen(fmt)+1, 1)
1787  B.CreateMemCpy(CI->getArgOperand(0), CI->getArgOperand(1),
1788  ConstantInt::get(DL.getIntPtrType(CI->getContext()),
1789  FormatStr.size() + 1),
1790  1); // Copy the null byte.
1791  return ConstantInt::get(CI->getType(), FormatStr.size());
1792  }
1793 
1794  // The remaining optimizations require the format string to be "%s" or "%c"
1795  // and have an extra operand.
1796  if (FormatStr.size() != 2 || FormatStr[0] != '%' ||
1797  CI->getNumArgOperands() < 3)
1798  return nullptr;
1799 
1800  // Decode the second character of the format string.
1801  if (FormatStr[1] == 'c') {
1802  // sprintf(dst, "%c", chr) --> *(i8*)dst = chr; *((i8*)dst+1) = 0
1803  if (!CI->getArgOperand(2)->getType()->isIntegerTy())
1804  return nullptr;
1805  Value *V = B.CreateTrunc(CI->getArgOperand(2), B.getInt8Ty(), "char");
1806  Value *Ptr = castToCStr(CI->getArgOperand(0), B);
1807  B.CreateStore(V, Ptr);
1808  Ptr = B.CreateGEP(B.getInt8Ty(), Ptr, B.getInt32(1), "nul");
1809  B.CreateStore(B.getInt8(0), Ptr);
1810 
1811  return ConstantInt::get(CI->getType(), 1);
1812  }
1813 
1814  if (FormatStr[1] == 's') {
1815  // sprintf(dest, "%s", str) -> llvm.memcpy(dest, str, strlen(str)+1, 1)
1816  if (!CI->getArgOperand(2)->getType()->isPointerTy())
1817  return nullptr;
1818 
1819  Value *Len = emitStrLen(CI->getArgOperand(2), B, DL, TLI);
1820  if (!Len)
1821  return nullptr;
1822  Value *IncLen =
1823  B.CreateAdd(Len, ConstantInt::get(Len->getType(), 1), "leninc");
1824  B.CreateMemCpy(CI->getArgOperand(0), CI->getArgOperand(2), IncLen, 1);
1825 
1826  // The sprintf result is the unincremented number of bytes in the string.
1827  return B.CreateIntCast(Len, CI->getType(), false);
1828  }
1829  return nullptr;
1830 }
1831 
1832 Value *LibCallSimplifier::optimizeSPrintF(CallInst *CI, IRBuilder<> &B) {
1833  Function *Callee = CI->getCalledFunction();
1834  FunctionType *FT = Callee->getFunctionType();
1835  if (Value *V = optimizeSPrintFString(CI, B)) {
1836  return V;
1837  }
1838 
1839  // sprintf(str, format, ...) -> siprintf(str, format, ...) if no floating
1840  // point arguments.
1841  if (TLI->has(LibFunc_siprintf) && !callHasFloatingPointArgument(CI)) {
1842  Module *M = B.GetInsertBlock()->getParent()->getParent();
1843  Constant *SIPrintFFn =
1844  M->getOrInsertFunction("siprintf", FT, Callee->getAttributes());
1845  CallInst *New = cast<CallInst>(CI->clone());
1846  New->setCalledFunction(SIPrintFFn);
1847  B.Insert(New);
1848  return New;
1849  }
1850  return nullptr;
1851 }
1852 
1853 Value *LibCallSimplifier::optimizeFPrintFString(CallInst *CI, IRBuilder<> &B) {
1854  optimizeErrorReporting(CI, B, 0);
1855 
1856  // All the optimizations depend on the format string.
1857  StringRef FormatStr;
1858  if (!getConstantStringInfo(CI->getArgOperand(1), FormatStr))
1859  return nullptr;
1860 
1861  // Do not do any of the following transformations if the fprintf return
1862  // value is used, in general the fprintf return value is not compatible
1863  // with fwrite(), fputc() or fputs().
1864  if (!CI->use_empty())
1865  return nullptr;
1866 
1867  // fprintf(F, "foo") --> fwrite("foo", 3, 1, F)
1868  if (CI->getNumArgOperands() == 2) {
1869  for (unsigned i = 0, e = FormatStr.size(); i != e; ++i)
1870  if (FormatStr[i] == '%') // Could handle %% -> % if we cared.
1871  return nullptr; // We found a format specifier.
1872 
1873  return emitFWrite(
1874  CI->getArgOperand(1),
1875  ConstantInt::get(DL.getIntPtrType(CI->getContext()), FormatStr.size()),
1876  CI->getArgOperand(0), B, DL, TLI);
1877  }
1878 
1879  // The remaining optimizations require the format string to be "%s" or "%c"
1880  // and have an extra operand.
1881  if (FormatStr.size() != 2 || FormatStr[0] != '%' ||
1882  CI->getNumArgOperands() < 3)
1883  return nullptr;
1884 
1885  // Decode the second character of the format string.
1886  if (FormatStr[1] == 'c') {
1887  // fprintf(F, "%c", chr) --> fputc(chr, F)
1888  if (!CI->getArgOperand(2)->getType()->isIntegerTy())
1889  return nullptr;
1890  return emitFPutC(CI->getArgOperand(2), CI->getArgOperand(0), B, TLI);
1891  }
1892 
1893  if (FormatStr[1] == 's') {
1894  // fprintf(F, "%s", str) --> fputs(str, F)
1895  if (!CI->getArgOperand(2)->getType()->isPointerTy())
1896  return nullptr;
1897  return emitFPutS(CI->getArgOperand(2), CI->getArgOperand(0), B, TLI);
1898  }
1899  return nullptr;
1900 }
1901 
1902 Value *LibCallSimplifier::optimizeFPrintF(CallInst *CI, IRBuilder<> &B) {
1903  Function *Callee = CI->getCalledFunction();
1904  FunctionType *FT = Callee->getFunctionType();
1905  if (Value *V = optimizeFPrintFString(CI, B)) {
1906  return V;
1907  }
1908 
1909  // fprintf(stream, format, ...) -> fiprintf(stream, format, ...) if no
1910  // floating point arguments.
1911  if (TLI->has(LibFunc_fiprintf) && !callHasFloatingPointArgument(CI)) {
1912  Module *M = B.GetInsertBlock()->getParent()->getParent();
1913  Constant *FIPrintFFn =
1914  M->getOrInsertFunction("fiprintf", FT, Callee->getAttributes());
1915  CallInst *New = cast<CallInst>(CI->clone());
1916  New->setCalledFunction(FIPrintFFn);
1917  B.Insert(New);
1918  return New;
1919  }
1920  return nullptr;
1921 }
1922 
1923 Value *LibCallSimplifier::optimizeFWrite(CallInst *CI, IRBuilder<> &B) {
1924  optimizeErrorReporting(CI, B, 3);
1925 
1926  // Get the element size and count.
1927  ConstantInt *SizeC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
1928  ConstantInt *CountC = dyn_cast<ConstantInt>(CI->getArgOperand(2));
1929  if (!SizeC || !CountC)
1930  return nullptr;
1931  uint64_t Bytes = SizeC->getZExtValue() * CountC->getZExtValue();
1932 
1933  // If this is writing zero records, remove the call (it's a noop).
1934  if (Bytes == 0)
1935  return ConstantInt::get(CI->getType(), 0);
1936 
1937  // If this is writing one byte, turn it into fputc.
1938  // This optimisation is only valid, if the return value is unused.
1939  if (Bytes == 1 && CI->use_empty()) { // fwrite(S,1,1,F) -> fputc(S[0],F)
1940  Value *Char = B.CreateLoad(castToCStr(CI->getArgOperand(0), B), "char");
1941  Value *NewCI = emitFPutC(Char, CI->getArgOperand(3), B, TLI);
1942  return NewCI ? ConstantInt::get(CI->getType(), 1) : nullptr;
1943  }
1944 
1945  return nullptr;
1946 }
1947 
1948 Value *LibCallSimplifier::optimizeFPuts(CallInst *CI, IRBuilder<> &B) {
1949  optimizeErrorReporting(CI, B, 1);
1950 
1951  // Don't rewrite fputs to fwrite when optimising for size because fwrite
1952  // requires more arguments and thus extra MOVs are required.
1953  if (CI->getParent()->getParent()->optForSize())
1954  return nullptr;
1955 
1956  // We can't optimize if return value is used.
1957  if (!CI->use_empty())
1958  return nullptr;
1959 
1960  // fputs(s,F) --> fwrite(s,1,strlen(s),F)
1961  uint64_t Len = GetStringLength(CI->getArgOperand(0));
1962  if (!Len)
1963  return nullptr;
1964 
1965  // Known to have no uses (see above).
1966  return emitFWrite(
1967  CI->getArgOperand(0),
1968  ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len - 1),
1969  CI->getArgOperand(1), B, DL, TLI);
1970 }
1971 
1972 Value *LibCallSimplifier::optimizePuts(CallInst *CI, IRBuilder<> &B) {
1973  // Check for a constant string.
1974  StringRef Str;
1975  if (!getConstantStringInfo(CI->getArgOperand(0), Str))
1976  return nullptr;
1977 
1978  if (Str.empty() && CI->use_empty()) {
1979  // puts("") -> putchar('\n')
1980  Value *Res = emitPutChar(B.getInt32('\n'), B, TLI);
1981  if (CI->use_empty() || !Res)
1982  return Res;
1983  return B.CreateIntCast(Res, CI->getType(), true);
1984  }
1985 
1986  return nullptr;
1987 }
1988 
1989 bool LibCallSimplifier::hasFloatVersion(StringRef FuncName) {
1990  LibFunc Func;
1991  SmallString<20> FloatFuncName = FuncName;
1992  FloatFuncName += 'f';
1993  if (TLI->getLibFunc(FloatFuncName, Func))
1994  return TLI->has(Func);
1995  return false;
1996 }
1997 
1998 Value *LibCallSimplifier::optimizeStringMemoryLibCall(CallInst *CI,
1999  IRBuilder<> &Builder) {
2000  LibFunc Func;
2001  Function *Callee = CI->getCalledFunction();
2002  // Check for string/memory library functions.
2003  if (TLI->getLibFunc(*Callee, Func) && TLI->has(Func)) {
2004  // Make sure we never change the calling convention.
2005  assert((ignoreCallingConv(Func) ||
2006  isCallingConvCCompatible(CI)) &&
2007  "Optimizing string/memory libcall would change the calling convention");
2008  switch (Func) {
2009  case LibFunc_strcat:
2010  return optimizeStrCat(CI, Builder);
2011  case LibFunc_strncat:
2012  return optimizeStrNCat(CI, Builder);
2013  case LibFunc_strchr:
2014  return optimizeStrChr(CI, Builder);
2015  case LibFunc_strrchr:
2016  return optimizeStrRChr(CI, Builder);
2017  case LibFunc_strcmp:
2018  return optimizeStrCmp(CI, Builder);
2019  case LibFunc_strncmp:
2020  return optimizeStrNCmp(CI, Builder);
2021  case LibFunc_strcpy:
2022  return optimizeStrCpy(CI, Builder);
2023  case LibFunc_stpcpy:
2024  return optimizeStpCpy(CI, Builder);
2025  case LibFunc_strncpy:
2026  return optimizeStrNCpy(CI, Builder);
2027  case LibFunc_strlen:
2028  return optimizeStrLen(CI, Builder);
2029  case LibFunc_strpbrk:
2030  return optimizeStrPBrk(CI, Builder);
2031  case LibFunc_strtol:
2032  case LibFunc_strtod:
2033  case LibFunc_strtof:
2034  case LibFunc_strtoul:
2035  case LibFunc_strtoll:
2036  case LibFunc_strtold:
2037  case LibFunc_strtoull:
2038  return optimizeStrTo(CI, Builder);
2039  case LibFunc_strspn:
2040  return optimizeStrSpn(CI, Builder);
2041  case LibFunc_strcspn:
2042  return optimizeStrCSpn(CI, Builder);
2043  case LibFunc_strstr:
2044  return optimizeStrStr(CI, Builder);
2045  case LibFunc_memchr:
2046  return optimizeMemChr(CI, Builder);
2047  case LibFunc_memcmp:
2048  return optimizeMemCmp(CI, Builder);
2049  case LibFunc_memcpy:
2050  return optimizeMemCpy(CI, Builder);
2051  case LibFunc_memmove:
2052  return optimizeMemMove(CI, Builder);
2053  case LibFunc_memset:
2054  return optimizeMemSet(CI, Builder);
2055  case LibFunc_wcslen:
2056  return optimizeWcslen(CI, Builder);
2057  default:
2058  break;
2059  }
2060  }
2061  return nullptr;
2062 }
2063 
2064 Value *LibCallSimplifier::optimizeFloatingPointLibCall(CallInst *CI,
2065  LibFunc Func,
2066  IRBuilder<> &Builder) {
2067  // Don't optimize calls that require strict floating point semantics.
2068  if (CI->isStrictFP())
2069  return nullptr;
2070 
2071  switch (Func) {
2072  case LibFunc_cosf:
2073  case LibFunc_cos:
2074  case LibFunc_cosl:
2075  return optimizeCos(CI, Builder);
2076  case LibFunc_sinpif:
2077  case LibFunc_sinpi:
2078  case LibFunc_cospif:
2079  case LibFunc_cospi:
2080  return optimizeSinCosPi(CI, Builder);
2081  case LibFunc_powf:
2082  case LibFunc_pow:
2083  case LibFunc_powl:
2084  return optimizePow(CI, Builder);
2085  case LibFunc_exp2l:
2086  case LibFunc_exp2:
2087  case LibFunc_exp2f:
2088  return optimizeExp2(CI, Builder);
2089  case LibFunc_fabsf:
2090  case LibFunc_fabs:
2091  case LibFunc_fabsl:
2092  return replaceUnaryCall(CI, Builder, Intrinsic::fabs);
2093  case LibFunc_sqrtf:
2094  case LibFunc_sqrt:
2095  case LibFunc_sqrtl:
2096  return optimizeSqrt(CI, Builder);
2097  case LibFunc_log:
2098  case LibFunc_log10:
2099  case LibFunc_log1p:
2100  case LibFunc_log2:
2101  case LibFunc_logb:
2102  return optimizeLog(CI, Builder);
2103  case LibFunc_tan:
2104  case LibFunc_tanf:
2105  case LibFunc_tanl:
2106  return optimizeTan(CI, Builder);
2107  case LibFunc_ceil:
2108  return replaceUnaryCall(CI, Builder, Intrinsic::ceil);
2109  case LibFunc_floor:
2110  return replaceUnaryCall(CI, Builder, Intrinsic::floor);
2111  case LibFunc_round:
2112  return replaceUnaryCall(CI, Builder, Intrinsic::round);
2113  case LibFunc_nearbyint:
2114  return replaceUnaryCall(CI, Builder, Intrinsic::nearbyint);
2115  case LibFunc_rint:
2116  return replaceUnaryCall(CI, Builder, Intrinsic::rint);
2117  case LibFunc_trunc:
2118  return replaceUnaryCall(CI, Builder, Intrinsic::trunc);
2119  case LibFunc_acos:
2120  case LibFunc_acosh:
2121  case LibFunc_asin:
2122  case LibFunc_asinh:
2123  case LibFunc_atan:
2124  case LibFunc_atanh:
2125  case LibFunc_cbrt:
2126  case LibFunc_cosh:
2127  case LibFunc_exp:
2128  case LibFunc_exp10:
2129  case LibFunc_expm1:
2130  case LibFunc_sin:
2131  case LibFunc_sinh:
2132  case LibFunc_tanh:
2133  if (UnsafeFPShrink && hasFloatVersion(CI->getCalledFunction()->getName()))
2134  return optimizeUnaryDoubleFP(CI, Builder, true);
2135  return nullptr;
2136  case LibFunc_copysign:
2137  if (hasFloatVersion(CI->getCalledFunction()->getName()))
2138  return optimizeBinaryDoubleFP(CI, Builder);
2139  return nullptr;
2140  case LibFunc_fminf:
2141  case LibFunc_fmin:
2142  case LibFunc_fminl:
2143  case LibFunc_fmaxf:
2144  case LibFunc_fmax:
2145  case LibFunc_fmaxl:
2146  return optimizeFMinFMax(CI, Builder);
2147  default:
2148  return nullptr;
2149  }
2150 }
2151 
2153  // TODO: Split out the code below that operates on FP calls so that
2154  // we can all non-FP calls with the StrictFP attribute to be
2155  // optimized.
2156  if (CI->isNoBuiltin())
2157  return nullptr;
2158 
2159  LibFunc Func;
2160  Function *Callee = CI->getCalledFunction();
2161 
2163  CI->getOperandBundlesAsDefs(OpBundles);
2164  IRBuilder<> Builder(CI, /*FPMathTag=*/nullptr, OpBundles);
2165  bool isCallingConvC = isCallingConvCCompatible(CI);
2166 
2167  // Command-line parameter overrides instruction attribute.
2168  // This can't be moved to optimizeFloatingPointLibCall() because it may be
2169  // used by the intrinsic optimizations.
2170  if (EnableUnsafeFPShrink.getNumOccurrences() > 0)
2171  UnsafeFPShrink = EnableUnsafeFPShrink;
2172  else if (isa<FPMathOperator>(CI) && CI->isFast())
2173  UnsafeFPShrink = true;
2174 
2175  // First, check for intrinsics.
2176  if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
2177  if (!isCallingConvC)
2178  return nullptr;
2179  // The FP intrinsics have corresponding constrained versions so we don't
2180  // need to check for the StrictFP attribute here.
2181  switch (II->getIntrinsicID()) {
2182  case Intrinsic::pow:
2183  return optimizePow(CI, Builder);
2184  case Intrinsic::exp2:
2185  return optimizeExp2(CI, Builder);
2186  case Intrinsic::log:
2187  return optimizeLog(CI, Builder);
2188  case Intrinsic::sqrt:
2189  return optimizeSqrt(CI, Builder);
2190  // TODO: Use foldMallocMemset() with memset intrinsic.
2191  default:
2192  return nullptr;
2193  }
2194  }
2195 
2196  // Also try to simplify calls to fortified library functions.
2197  if (Value *SimplifiedFortifiedCI = FortifiedSimplifier.optimizeCall(CI)) {
2198  // Try to further simplify the result.
2199  CallInst *SimplifiedCI = dyn_cast<CallInst>(SimplifiedFortifiedCI);
2200  if (SimplifiedCI && SimplifiedCI->getCalledFunction()) {
2201  // Use an IR Builder from SimplifiedCI if available instead of CI
2202  // to guarantee we reach all uses we might replace later on.
2203  IRBuilder<> TmpBuilder(SimplifiedCI);
2204  if (Value *V = optimizeStringMemoryLibCall(SimplifiedCI, TmpBuilder)) {
2205  // If we were able to further simplify, remove the now redundant call.
2206  SimplifiedCI->replaceAllUsesWith(V);
2207  SimplifiedCI->eraseFromParent();
2208  return V;
2209  }
2210  }
2211  return SimplifiedFortifiedCI;
2212  }
2213 
2214  // Then check for known library functions.
2215  if (TLI->getLibFunc(*Callee, Func) && TLI->has(Func)) {
2216  // We never change the calling convention.
2217  if (!ignoreCallingConv(Func) && !isCallingConvC)
2218  return nullptr;
2219  if (Value *V = optimizeStringMemoryLibCall(CI, Builder))
2220  return V;
2221  if (Value *V = optimizeFloatingPointLibCall(CI, Func, Builder))
2222  return V;
2223  switch (Func) {
2224  case LibFunc_ffs:
2225  case LibFunc_ffsl:
2226  case LibFunc_ffsll:
2227  return optimizeFFS(CI, Builder);
2228  case LibFunc_fls:
2229  case LibFunc_flsl:
2230  case LibFunc_flsll:
2231  return optimizeFls(CI, Builder);
2232  case LibFunc_abs:
2233  case LibFunc_labs:
2234  case LibFunc_llabs:
2235  return optimizeAbs(CI, Builder);
2236  case LibFunc_isdigit:
2237  return optimizeIsDigit(CI, Builder);
2238  case LibFunc_isascii:
2239  return optimizeIsAscii(CI, Builder);
2240  case LibFunc_toascii:
2241  return optimizeToAscii(CI, Builder);
2242  case LibFunc_printf:
2243  return optimizePrintF(CI, Builder);
2244  case LibFunc_sprintf:
2245  return optimizeSPrintF(CI, Builder);
2246  case LibFunc_fprintf:
2247  return optimizeFPrintF(CI, Builder);
2248  case LibFunc_fwrite:
2249  return optimizeFWrite(CI, Builder);
2250  case LibFunc_fputs:
2251  return optimizeFPuts(CI, Builder);
2252  case LibFunc_puts:
2253  return optimizePuts(CI, Builder);
2254  case LibFunc_perror:
2255  return optimizeErrorReporting(CI, Builder);
2256  case LibFunc_vfprintf:
2257  case LibFunc_fiprintf:
2258  return optimizeErrorReporting(CI, Builder, 0);
2259  case LibFunc_fputc:
2260  return optimizeErrorReporting(CI, Builder, 1);
2261  default:
2262  return nullptr;
2263  }
2264  }
2265  return nullptr;
2266 }
2267 
2269  const DataLayout &DL, const TargetLibraryInfo *TLI,
2271  function_ref<void(Instruction *, Value *)> Replacer)
2272  : FortifiedSimplifier(TLI), DL(DL), TLI(TLI), ORE(ORE),
2273  UnsafeFPShrink(false), Replacer(Replacer) {}
2274 
2275 void LibCallSimplifier::replaceAllUsesWith(Instruction *I, Value *With) {
2276  // Indirect through the replacer used in this instance.
2277  Replacer(I, With);
2278 }
2279 
2280 // TODO:
2281 // Additional cases that we need to add to this file:
2282 //
2283 // cbrt:
2284 // * cbrt(expN(X)) -> expN(x/3)
2285 // * cbrt(sqrt(x)) -> pow(x,1/6)
2286 // * cbrt(cbrt(x)) -> pow(x,1/9)
2287 //
2288 // exp, expf, expl:
2289 // * exp(log(x)) -> x
2290 //
2291 // log, logf, logl:
2292 // * log(exp(x)) -> x
2293 // * log(exp(y)) -> y*log(e)
2294 // * log(exp10(y)) -> y*log(10)
2295 // * log(sqrt(x)) -> 0.5*log(x)
2296 //
2297 // pow, powf, powl:
2298 // * pow(sqrt(x),y) -> pow(x,y*0.5)
2299 // * pow(pow(x,y),z)-> pow(x,y*z)
2300 //
2301 // signbit:
2302 // * signbit(cnst) -> cnst'
2303 // * signbit(nncst) -> 0 (if pstv is a non-negative constant)
2304 //
2305 // sqrt, sqrtf, sqrtl:
2306 // * sqrt(expN(x)) -> expN(x*0.5)
2307 // * sqrt(Nroot(x)) -> pow(x,1/(2*N))
2308 // * sqrt(pow(x,y)) -> pow(|x|,y*0.5)
2309 //
2310 
2311 //===----------------------------------------------------------------------===//
2312 // Fortified Library Call Optimizations
2313 //===----------------------------------------------------------------------===//
2314 
2315 bool FortifiedLibCallSimplifier::isFortifiedCallFoldable(CallInst *CI,
2316  unsigned ObjSizeOp,
2317  unsigned SizeOp,
2318  bool isString) {
2319  if (CI->getArgOperand(ObjSizeOp) == CI->getArgOperand(SizeOp))
2320  return true;
2321  if (ConstantInt *ObjSizeCI =
2322  dyn_cast<ConstantInt>(CI->getArgOperand(ObjSizeOp))) {
2323  if (ObjSizeCI->isMinusOne())
2324  return true;
2325  // If the object size wasn't -1 (unknown), bail out if we were asked to.
2326  if (OnlyLowerUnknownSize)
2327  return false;
2328  if (isString) {
2329  uint64_t Len = GetStringLength(CI->getArgOperand(SizeOp));
2330  // If the length is 0 we don't know how long it is and so we can't
2331  // remove the check.
2332  if (Len == 0)
2333  return false;
2334  return ObjSizeCI->getZExtValue() >= Len;
2335  }
2336  if (ConstantInt *SizeCI = dyn_cast<ConstantInt>(CI->getArgOperand(SizeOp)))
2337  return ObjSizeCI->getZExtValue() >= SizeCI->getZExtValue();
2338  }
2339  return false;
2340 }
2341 
2342 Value *FortifiedLibCallSimplifier::optimizeMemCpyChk(CallInst *CI,
2343  IRBuilder<> &B) {
2344  if (isFortifiedCallFoldable(CI, 3, 2, false)) {
2345  B.CreateMemCpy(CI->getArgOperand(0), CI->getArgOperand(1),
2346  CI->getArgOperand(2), 1);
2347  return CI->getArgOperand(0);
2348  }
2349  return nullptr;
2350 }
2351 
2352 Value *FortifiedLibCallSimplifier::optimizeMemMoveChk(CallInst *CI,
2353  IRBuilder<> &B) {
2354  if (isFortifiedCallFoldable(CI, 3, 2, false)) {
2355  B.CreateMemMove(CI->getArgOperand(0), CI->getArgOperand(1),
2356  CI->getArgOperand(2), 1);
2357  return CI->getArgOperand(0);
2358  }
2359  return nullptr;
2360 }
2361 
2362 Value *FortifiedLibCallSimplifier::optimizeMemSetChk(CallInst *CI,
2363  IRBuilder<> &B) {
2364  // TODO: Try foldMallocMemset() here.
2365 
2366  if (isFortifiedCallFoldable(CI, 3, 2, false)) {
2367  Value *Val = B.CreateIntCast(CI->getArgOperand(1), B.getInt8Ty(), false);
2368  B.CreateMemSet(CI->getArgOperand(0), Val, CI->getArgOperand(2), 1);
2369  return CI->getArgOperand(0);
2370  }
2371  return nullptr;
2372 }
2373 
2374 Value *FortifiedLibCallSimplifier::optimizeStrpCpyChk(CallInst *CI,
2375  IRBuilder<> &B,
2376  LibFunc Func) {
2377  Function *Callee = CI->getCalledFunction();
2378  StringRef Name = Callee->getName();
2379  const DataLayout &DL = CI->getModule()->getDataLayout();
2380  Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1),
2381  *ObjSize = CI->getArgOperand(2);
2382 
2383  // __stpcpy_chk(x,x,...) -> x+strlen(x)
2384  if (Func == LibFunc_stpcpy_chk && !OnlyLowerUnknownSize && Dst == Src) {
2385  Value *StrLen = emitStrLen(Src, B, DL, TLI);
2386  return StrLen ? B.CreateInBoundsGEP(B.getInt8Ty(), Dst, StrLen) : nullptr;
2387  }
2388 
2389  // If a) we don't have any length information, or b) we know this will
2390  // fit then just lower to a plain st[rp]cpy. Otherwise we'll keep our
2391  // st[rp]cpy_chk call which may fail at runtime if the size is too long.
2392  // TODO: It might be nice to get a maximum length out of the possible
2393  // string lengths for varying.
2394  if (isFortifiedCallFoldable(CI, 2, 1, true))
2395  return emitStrCpy(Dst, Src, B, TLI, Name.substr(2, 6));
2396 
2397  if (OnlyLowerUnknownSize)
2398  return nullptr;
2399 
2400  // Maybe we can stil fold __st[rp]cpy_chk to __memcpy_chk.
2401  uint64_t Len = GetStringLength(Src);
2402  if (Len == 0)
2403  return nullptr;
2404 
2405  Type *SizeTTy = DL.getIntPtrType(CI->getContext());
2406  Value *LenV = ConstantInt::get(SizeTTy, Len);
2407  Value *Ret = emitMemCpyChk(Dst, Src, LenV, ObjSize, B, DL, TLI);
2408  // If the function was an __stpcpy_chk, and we were able to fold it into
2409  // a __memcpy_chk, we still need to return the correct end pointer.
2410  if (Ret && Func == LibFunc_stpcpy_chk)
2411  return B.CreateGEP(B.getInt8Ty(), Dst, ConstantInt::get(SizeTTy, Len - 1));
2412  return Ret;
2413 }
2414 
2415 Value *FortifiedLibCallSimplifier::optimizeStrpNCpyChk(CallInst *CI,
2416  IRBuilder<> &B,
2417  LibFunc Func) {
2418  Function *Callee = CI->getCalledFunction();
2419  StringRef Name = Callee->getName();
2420  if (isFortifiedCallFoldable(CI, 3, 2, false)) {
2421  Value *Ret = emitStrNCpy(CI->getArgOperand(0), CI->getArgOperand(1),
2422  CI->getArgOperand(2), B, TLI, Name.substr(2, 7));
2423  return Ret;
2424  }
2425  return nullptr;
2426 }
2427 
2429  // FIXME: We shouldn't be changing "nobuiltin" or TLI unavailable calls here.
2430  // Some clang users checked for _chk libcall availability using:
2431  // __has_builtin(__builtin___memcpy_chk)
2432  // When compiling with -fno-builtin, this is always true.
2433  // When passing -ffreestanding/-mkernel, which both imply -fno-builtin, we
2434  // end up with fortified libcalls, which isn't acceptable in a freestanding
2435  // environment which only provides their non-fortified counterparts.
2436  //
2437  // Until we change clang and/or teach external users to check for availability
2438  // differently, disregard the "nobuiltin" attribute and TLI::has.
2439  //
2440  // PR23093.
2441 
2442  LibFunc Func;
2443  Function *Callee = CI->getCalledFunction();
2444 
2446  CI->getOperandBundlesAsDefs(OpBundles);
2447  IRBuilder<> Builder(CI, /*FPMathTag=*/nullptr, OpBundles);
2448  bool isCallingConvC = isCallingConvCCompatible(CI);
2449 
2450  // First, check that this is a known library functions and that the prototype
2451  // is correct.
2452  if (!TLI->getLibFunc(*Callee, Func))
2453  return nullptr;
2454 
2455  // We never change the calling convention.
2456  if (!ignoreCallingConv(Func) && !isCallingConvC)
2457  return nullptr;
2458 
2459  switch (Func) {
2460  case LibFunc_memcpy_chk:
2461  return optimizeMemCpyChk(CI, Builder);
2462  case LibFunc_memmove_chk:
2463  return optimizeMemMoveChk(CI, Builder);
2464  case LibFunc_memset_chk:
2465  return optimizeMemSetChk(CI, Builder);
2466  case LibFunc_stpcpy_chk:
2467  case LibFunc_strcpy_chk:
2468  return optimizeStrpCpyChk(CI, Builder, Func);
2469  case LibFunc_stpncpy_chk:
2470  case LibFunc_strncpy_chk:
2471  return optimizeStrpNCpyChk(CI, Builder, Func);
2472  default:
2473  break;
2474  }
2475  return nullptr;
2476 }
2477 
2479  const TargetLibraryInfo *TLI, bool OnlyLowerUnknownSize)
2480  : TLI(TLI), OnlyLowerUnknownSize(OnlyLowerUnknownSize) {}
static Value * optimizeUnaryDoubleFP(CallInst *CI, IRBuilder<> &B, bool CheckRetType)
Shrink double -> float for unary functions like &#39;floor&#39;.
Value * CreateInBoundsGEP(Value *Ptr, ArrayRef< Value *> IdxList, const Twine &Name="")
Definition: IRBuilder.h:1244
uint64_t CallInst * C
bool isIntrinsic() const
isIntrinsic - Returns true if the function&#39;s name starts with "llvm.".
Definition: Function.h:180
SymbolTableList< Instruction >::iterator eraseFromParent()
This method unlinks &#39;this&#39; from the containing basic block and deletes it.
Definition: Instruction.cpp:69
void push_back(const T &Elt)
Definition: SmallVector.h:212
A parsed version of the target data layout string in and methods for querying it. ...
Definition: DataLayout.h:109
const std::string & getTargetTriple() const
Get the target triple which is a string describing the target host.
Definition: Module.h:233
class_match< Value > m_Value()
Match an arbitrary value and ignore it.
Definition: PatternMatch.h:72
Value * CreateICmp(CmpInst::Predicate P, Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1638
static Value * foldMallocMemset(CallInst *Memset, IRBuilder<> &B, const TargetLibraryInfo &TLI)
Fold memset[_chk](malloc(n), 0, n) –> calloc(1, n).
void flipAllBits()
Toggle every bit to its opposite value.
Definition: APInt.h:1456
bool isOnlyUsedInZeroEqualityComparison(const Instruction *CxtI)
GCNRegPressure max(const GCNRegPressure &P1, const GCNRegPressure &P2)
void setNoSignedZeros()
Definition: Operator.h:213
Value * CreateICmpNE(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1553
Value * CreateIsNotNull(Value *Arg, const Twine &Name="")
Return an i1 value testing if Arg is not null.
Definition: IRBuilder.h:1788
Value * CreateZExtOrTrunc(Value *V, Type *DestTy, const Twine &Name="")
Create a ZExt or Trunc from the integer value V to DestTy.
Definition: IRBuilder.h:1395
Compute iterated dominance frontiers using a linear time algorithm.
Definition: AllocatorList.h:24
static bool callHasFloatingPointArgument(const CallInst *CI)
Type * getParamType(unsigned i) const
Parameter type accessors.
Definition: DerivedTypes.h:135
Value * CreateICmpULT(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1562
Constant * getOrInsertFunction(StringRef Name, FunctionType *T, AttributeList AttributeList)
Look up the specified function in the module symbol table.
Definition: Module.cpp:142
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:63
static Constant * getInfinity(Type *Ty, bool Negative=false)
Definition: Constants.cpp:713
2: 32-bit floating point type
Definition: Type.h:59
LLVM_NODISCARD size_t rfind(char C, size_t From=npos) const
Search for the last character C in the string.
Definition: StringRef.h:360
LLVM_NODISCARD LLVM_ATTRIBUTE_ALWAYS_INLINE size_t size() const
size - Get the string size.
Definition: StringRef.h:138
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:878
An efficient, type-erasing, non-owning reference to a callable.
Definition: STLExtras.h:89
bool isStrictFP() const
Determine if the call requires strict floating point semantics.
ARM_AAPCS - ARM Architecture Procedure Calling Standard calling convention (aka EABI).
Definition: CallingConv.h:100
LLVMContext & getContext() const
All values hold a context through their type.
Definition: Value.cpp:728
Value * CreateSExt(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:1390
F(f)
const fltSemantics & getSemantics() const
Definition: APFloat.h:1155
unsigned getPointerAddressSpace() const
Get the address space of this pointer or pointer vector type.
Definition: DerivedTypes.h:503
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:405
An instruction for reading from memory.
Definition: Instructions.h:164
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.
uint64_t Offset
Slice starts at this Offset.
static Value * emitCalloc(Value *Num, Value *Size, const AttributeList &Attrs, IRBuilder<> &B, const TargetLibraryInfo &TLI)
CallingConv::ID getCallingConv() const
getCallingConv/setCallingConv - Get or set the calling convention of this function call...
LLVMContext & getContext() const
Get the context in which this basic block lives.
Definition: BasicBlock.cpp:33
LLVM_NODISCARD LLVM_ATTRIBUTE_ALWAYS_INLINE const char * data() const
data - Get a pointer to the start of the string (which may not be null terminated).
Definition: StringRef.h:128
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:207
iterator begin()
Instruction iterator methods.
Definition: BasicBlock.h:252
void setCallingConv(CallingConv::ID CC)
IntegerType * getInt32Ty()
Fetch the type representing a 32-bit integer.
Definition: IRBuilder.h:348
static Value * replaceUnaryCall(CallInst *CI, IRBuilder<> &B, Intrinsic::ID IID)
specific_fpval m_SpecificFP(double V)
Match a specific floating point value or vector with all elements equal to the value.
Definition: PatternMatch.h:404
bool match(Val *V, const Pattern &P)
Definition: PatternMatch.h:49
const Module * getModule() const
Return the module owning the function this basic block belongs to, or nullptr it the function does no...
Definition: BasicBlock.cpp:116
Value * CreateFPExt(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:1437
void setBit(unsigned BitPosition)
Set a given bit to 1.
Definition: APInt.h:1382
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:361
unsigned getNumArgOperands() const
Return the number of call arguments.
bool isNonNegative() const
Determine if this APInt Value is non-negative (>= 0)
Definition: APInt.h:362
TypeID getTypeID() const
Return the type id for the type.
Definition: Type.h:138
FortifiedLibCallSimplifier(const TargetLibraryInfo *TLI, bool OnlyLowerUnknownSize=false)
A Use represents the edge between a Value definition and its users.
Definition: Use.h:56
PointerType * getPointerTo(unsigned AddrSpace=0) const
Return a pointer to the current type.
Definition: Type.cpp:639
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:197
This provides a uniform API for creating instructions and inserting them into a basic block: either a...
Definition: IRBuilder.h:668
LibCallSimplifier(const DataLayout &DL, const TargetLibraryInfo *TLI, OptimizationRemarkEmitter &ORE, function_ref< void(Instruction *, Value *)> Replacer=&replaceAllUsesWithDefault)
Value * CreateAdd(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:893
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 * CreateICmpSGT(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1568
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:336
StoreInst * CreateStore(Value *Val, Value *Ptr, bool isVolatile=false)
Definition: IRBuilder.h:1180
Value * CreateIntToPtr(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:1444
Instruction * clone() const
Create a copy of &#39;this&#39; instruction that is identical in all ways except the following: ...
LLVM_NODISCARD LLVM_ATTRIBUTE_ALWAYS_INLINE bool startswith(StringRef Prefix) const
Check if this string starts with the given Prefix.
Definition: StringRef.h:267
Class to represent function types.
Definition: DerivedTypes.h:103
Value * CreateBitCast(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:1448
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:245
Value * CreateSExtOrTrunc(Value *V, Type *DestTy, const Twine &Name="")
Create a SExt or Trunc from the integer value V to DestTy.
Definition: IRBuilder.h:1409
const ConstantDataArray * Array
ConstantDataArray pointer.
opStatus convert(const fltSemantics &ToSemantics, roundingMode RM, bool *losesInfo)
Definition: APFloat.cpp:4441
C - The default llvm calling convention, compatible with C.
Definition: CallingConv.h:35
#define T
BasicBlock * GetInsertBlock() const
Definition: IRBuilder.h:122
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 LLVM_ATTRIBUTE_ALWAYS_INLINE bool empty() const
empty - Check if the string is empty.
Definition: StringRef.h:133
CallInst * CreateMemMove(Value *Dst, Value *Src, uint64_t Size, unsigned Align, bool isVolatile=false, MDNode *TBAATag=nullptr, MDNode *ScopeTag=nullptr, MDNode *NoAliasTag=nullptr)
Create and insert a memmove between the specified pointers.
Definition: IRBuilder.h:468
SmallString - A SmallString is just a SmallVector with methods and accessors that make it work better...
Definition: SmallString.h:26
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:915
AttributeList getAttributes() const
Return the attribute list for this Function.
Definition: Function.h:205
void replaceAllUsesWith(Value *V)
Change all uses of this to point to a new Value.
Definition: Value.cpp:430
static const fltSemantics & IEEEdouble() LLVM_READNONE
Definition: APFloat.cpp:122
LLVM_NODISCARD LLVM_ATTRIBUTE_ALWAYS_INLINE StringRef substr(size_t Start, size_t N=npos) const
Return a reference to the substring from [Start, Start + N).
Definition: StringRef.h:598
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:1387
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:292
Value * CreateFCmpOEQ(Value *LHS, Value *RHS, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:1581
amdgpu Simplify well known AMD library false Value * Callee
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:980
void SetInsertPoint(BasicBlock *TheBB)
This specifies that created instructions should be appended to the end of the specified block...
Definition: IRBuilder.h:128
Value * getOperand(unsigned i) const
Definition: User.h:154
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:319
bool isNoBuiltin() const
Return true if the call should not be treated as a call to a builtin.
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:1678
Value * emitFPutS(Value *Str, Value *File, IRBuilder<> &B, const TargetLibraryInfo *TLI)
Emit a call to the puts function.
bool isFloatTy() const
Return true if this is &#39;float&#39;, a 32-bit IEEE fp type.
Definition: Type.h:147
const BasicBlock & getEntryBlock() const
Definition: Function.h:572
LoadInst * CreateLoad(Value *Ptr, const char *Name)
Definition: IRBuilder.h:1168
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:702
static 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...
Type * getDoubleTy()
Fetch the type representing a 64-bit floating point value.
Definition: IRBuilder.h:376
initializer< Ty > init(const Ty &Val)
Definition: CommandLine.h:406
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:149
static GCRegistry::Add< OcamlGC > B("ocaml", "ocaml 3.10-compatible GC")
LLVM Basic Block Representation.
Definition: BasicBlock.h:59
The instances of the Type class are immutable: once they are created, they are never changed...
Definition: Type.h:46
ConstantInt * getTrue()
Get the constant value for i1 true.
Definition: IRBuilder.h:288
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
This is an important base class in LLVM.
Definition: Constant.h:42
Value * CreateSelect(Value *C, Value *True, Value *False, const Twine &Name="", Instruction *MDFrom=nullptr)
Definition: IRBuilder.h:1693
bool isPointerTy() const
True if this is an instance of PointerType.
Definition: Type.h:221
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:265
ConstantFP - Floating Point Values [float, double].
Definition: Constants.h:264
Value * CreateFCmpOGT(Value *LHS, Value *RHS, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:1585
double convertToDouble() const
Definition: APFloat.h:1097
Diagnostic information for applied optimization remarks.
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:774
Value * CreateNeg(Value *V, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1124
bool optForSize() const
Optimize this function for size (-Os) or minimum size (-Oz).
Definition: Function.h:530
This instruction compares its operands according to the predicate given to the constructor.
op_range operands()
Definition: User.h:222
Value * getPointerOperand()
Definition: Instructions.h:270
LLVM_NODISCARD LLVM_ATTRIBUTE_ALWAYS_INLINE 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:184
static bool isCallingConvCCompatible(CallInst *CI)
Class to represent integer types.
Definition: DerivedTypes.h:40
ConstantInt * getInt64(uint64_t C)
Get a constant 64-bit value.
Definition: IRBuilder.h:313
Value * castToCStr(Value *V, IRBuilder<> &B)
Return V if it is an i8*, otherwise cast it to i8*.
Represents offset+length into a ConstantDataArray.
Value * CreateExtractElement(Value *Vec, Value *Idx, const Twine &Name="")
Definition: IRBuilder.h:1713
const Function * getFunction() const
Return the function this instruction belongs to.
Definition: Instruction.cpp:61
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 Constant * getAllOnesValue(Type *Ty)
Get the all ones value.
Definition: Constants.cpp:261
uint64_t NextPowerOf2(uint64_t A)
Returns the next power of two (in 64-bits) that is strictly greater than A.
Definition: MathExtras.h:632
const Constant * stripPointerCasts() const
Definition: Constant.h:153
const Value * stripPointerCasts() const
Strip off pointer casts, all-zero GEPs, and aliases.
Definition: Value.cpp:558
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:1755
LLVM_NODISCARD char back() const
back - Get the last character in the string.
Definition: StringRef.h:149
PointerType * getInt8PtrTy(unsigned AddrSpace=0)
Fetch the type representing a pointer to an 8-bit integer value.
Definition: IRBuilder.h:386
Value * CreateTrunc(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:1384
Triple - Helper class for working with autoconf configuration names.
Definition: Triple.h:44
FunctionType * getFunctionType() const
LLVM_NODISCARD LLVM_ATTRIBUTE_ALWAYS_INLINE 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:654
Value * CreateGEP(Value *Ptr, ArrayRef< Value *> IdxList, const Twine &Name="")
Definition: IRBuilder.h:1227
Value * CreateFCmpOLT(Value *LHS, Value *RHS, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:1593
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:240
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
bool isInteger() const
Definition: APFloat.h:1162
This is the shared class of boolean and integer constants.
Definition: Constants.h:84
GlobalVariable * CreateGlobalString(StringRef Str, const Twine &Name="", unsigned AddressSpace=0)
Make a new global variable with initializer type i8*.
Definition: IRBuilder.cpp:27
Value * CreateIntCast(Value *V, Type *DestTy, bool isSigned, const Twine &Name="")
Definition: IRBuilder.h:1511
This is a &#39;vector&#39; (really, a variable-sized array), optimized for the case when the array is small...
Definition: SmallVector.h:864
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:2564
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.
ARM_APCS - ARM Procedure Calling Standard calling convention (obsolete, but still used on some target...
Definition: CallingConv.h:96
ConstantInt * getInt32(uint32_t C)
Get a constant 32-bit value.
Definition: IRBuilder.h:308
bool hasFnAttr(Attribute::AttrKind Kind) const
Determine whether this call has the given attribute.
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:560
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:623
Value * CreateConstInBoundsGEP1_64(Value *Ptr, uint64_t Idx0, const Twine &Name="")
Definition: IRBuilder.h:1331
BinaryOp_match< LHS, RHS, Instruction::FMul > m_FMul(const LHS &L, const RHS &R)
Definition: PatternMatch.h:532
Intrinsic::ID getIntrinsicID() const LLVM_READONLY
getIntrinsicID - This method returns the ID number of the specified function, or Intrinsic::not_intri...
Definition: Function.h:175
FunctionType * getFunctionType() const
Returns the FunctionType for me.
Definition: Function.h:145
const Module * getModule() const
Return the module owning the function this instruction belongs to or nullptr it the function does not...
Definition: Instruction.cpp:57
Class for arbitrary precision integers.
Definition: APInt.h:69
bool ule(const APInt &RHS) const
Unsigned less or equal comparison.
Definition: APInt.h:1202
iterator_range< user_iterator > users()
Definition: Value.h:401
IntegerType * getInt8Ty()
Fetch the type representing an 8-bit integer.
Definition: IRBuilder.h:338
Value * CreateShl(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1012
bool getLibFunc(StringRef funcName, LibFunc &F) const
Searches for a particular function name.
iterator begin() const
Definition: StringRef.h:106
static Value * optimizeBinaryDoubleFP(CallInst *CI, IRBuilder<> &B)
Shrink double -> float for binary functions like &#39;fmin/fmax&#39;.
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:293
amdgpu Simplify well known AMD library false Value Value * Arg
Function * getCalledFunction() const
Return the function called, or null if this is an indirect function invocation.
Predicate getPredicate() const
Return the predicate for this instruction.
Definition: InstrTypes.h:927
void setCalledFunction(Value *Fn)
Set the function called.
CallInst * CreateMemCpy(Value *Dst, Value *Src, uint64_t Size, unsigned Align, 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:423
Merge contiguous icmps into a memcmp
Definition: MergeICmps.cpp:649
static const size_t npos
Definition: StringRef.h:51
unsigned getIntegerBitWidth() const
Definition: DerivedTypes.h:97
LLVM_NODISCARD bool empty() const
Definition: SmallVector.h:61
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.
static bool isFNeg(const Value *V, bool IgnoreZeroSign=false)
Value * getArgOperand(unsigned i) const
getArgOperand/setArgOperand - Return/set the i-th call argument.
StringRef getName() const
Return a constant reference to the value&#39;s name.
Definition: Value.cpp:220
const Function * getParent() const
Return the enclosing method, or null if none.
Definition: BasicBlock.h:108
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:395
#define I(x, y, z)
Definition: MD5.cpp:58
APFloat abs(APFloat X)
Returns the absolute value of the argument.
Definition: APFloat.h:1213
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:193
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:234
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:323
static void insertSinCosCall(IRBuilder<> &B, Function *OrigCallee, Value *Arg, bool UseFloat, Value *&Sin, Value *&Cos, Value *&SinCos)
Type * getFloatTy()
Fetch the type representing a 32-bit floating point value.
Definition: IRBuilder.h:371
Value * CreateAnd(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1067
void getOperandBundlesAsDefs(SmallVectorImpl< OperandBundleDef > &Defs) const
Return the list of operand bundles attached to this instruction as a vector of OperandBundleDefs.
Definition: InstrTypes.h:1398
bool isDeclaration() const
Return true if the primary definition of this global value is outside of the current translation unit...
Definition: Globals.cpp:201
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:719
3: 64-bit floating point type
Definition: Type.h:60
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.
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:377
This class represents a truncation of floating point types.
Module * getParent()
Get the module that this global value is contained inside of...
Definition: GlobalValue.h:556
LLVM Value Representation.
Definition: Value.h:73
Value * emitPutS(Value *Str, IRBuilder<> &B, const TargetLibraryInfo *TLI)
Emit a call to the puts function. This assumes that Str is some pointer.
constexpr char Size[]
Key for Kernel::Arg::Metadata::mSize.
static VectorType * get(Type *ElementType, unsigned NumElements)
This static method is the primary way to construct an VectorType.
Definition: Type.cpp:593
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:221
Value * CreateFDiv(Value *LHS, Value *RHS, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:983
ConstantInt * getInt8(uint8_t C)
Get a constant 8-bit value.
Definition: IRBuilder.h:298
bool hasNoNaNs() const
Determine whether the no-NaNs flag is set.
Convenience struct for specifying and reasoning about fast-math flags.
Definition: Operator.h:160
StringRef - Represent a constant reference to a string, i.e.
Definition: StringRef.h:49
void addAttribute(unsigned i, Attribute::AttrKind Kind)
adds the attribute to the list of attributes.
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:157
ConstantInt * getInt(const APInt &AI)
Get a constant integer value.
Definition: IRBuilder.h:324
LLVM_NODISCARD LLVM_ATTRIBUTE_ALWAYS_INLINE size_t find(char C, size_t From=0) const
Search for the first character C in the string.
Definition: StringRef.h:298
static bool ignoreCallingConv(LibFunc Func)
This class represents an extension of floating point types.
Value * CreateFMul(Value *LHS, Value *RHS, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:951
iterator end() const
Definition: StringRef.h:108
static Constant * getFPExtend(Constant *C, Type *Ty, bool OnlyIfReduced=false)
Definition: Constants.cpp:1596
bool isDoubleTy() const
Return true if this is &#39;double&#39;, a 64-bit IEEE fp type.
Definition: Type.h:150
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)
Determine which bits of V are known to be either zero or one and return them in the KnownZero/KnownOn...
The optimization diagnostic interface.
bool use_empty() const
Definition: Value.h:328
static Value * getPow(Value *InnerChain[33], unsigned Exp, IRBuilder<> &B)
ARM_AAPCS_VFP - Same as ARM_AAPCS, but uses hard floating point ABI.
Definition: CallingConv.h:103
bool isStructTy() const
True if this is an instance of StructType.
Definition: Type.h:215
cmpResult compare(const APFloat &RHS) const
Definition: APFloat.h:1102
A wrapper class for inspecting calls to intrinsic functions.
Definition: IntrinsicInst.h:44
const BasicBlock * getParent() const
Definition: Instruction.h:66
CallInst * CreateCall(Value *Callee, ArrayRef< Value *> Args=None, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:1663
void addParamAttr(unsigned ArgNo, Attribute::AttrKind Kind)
Adds the attribute to the indicated argument.
user_iterator user_end()
Definition: Value.h:385