LLVM  7.0.0svn
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 //===----------------------------------------------------------------------===//
108 // String and Memory Library Call Optimizations
109 //===----------------------------------------------------------------------===//
110 
111 Value *LibCallSimplifier::optimizeStrCat(CallInst *CI, IRBuilder<> &B) {
112  // Extract some information from the instruction
113  Value *Dst = CI->getArgOperand(0);
114  Value *Src = CI->getArgOperand(1);
115 
116  // See if we can get the length of the input string.
117  uint64_t Len = GetStringLength(Src);
118  if (Len == 0)
119  return nullptr;
120  --Len; // Unbias length.
121 
122  // Handle the simple, do-nothing case: strcat(x, "") -> x
123  if (Len == 0)
124  return Dst;
125 
126  return emitStrLenMemCpy(Src, Dst, Len, B);
127 }
128 
129 Value *LibCallSimplifier::emitStrLenMemCpy(Value *Src, Value *Dst, uint64_t Len,
130  IRBuilder<> &B) {
131  // We need to find the end of the destination string. That's where the
132  // memory is to be moved to. We just generate a call to strlen.
133  Value *DstLen = emitStrLen(Dst, B, DL, TLI);
134  if (!DstLen)
135  return nullptr;
136 
137  // Now that we have the destination's length, we must index into the
138  // destination's pointer to get the actual memcpy destination (end of
139  // the string .. we're concatenating).
140  Value *CpyDst = B.CreateGEP(B.getInt8Ty(), Dst, DstLen, "endptr");
141 
142  // We have enough information to now generate the memcpy call to do the
143  // concatenation for us. Make a memcpy to copy the nul byte with align = 1.
144  B.CreateMemCpy(CpyDst, Src,
145  ConstantInt::get(DL.getIntPtrType(Src->getContext()), Len + 1),
146  1);
147  return Dst;
148 }
149 
150 Value *LibCallSimplifier::optimizeStrNCat(CallInst *CI, IRBuilder<> &B) {
151  // Extract some information from the instruction.
152  Value *Dst = CI->getArgOperand(0);
153  Value *Src = CI->getArgOperand(1);
154  uint64_t Len;
155 
156  // We don't do anything if length is not constant.
157  if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(CI->getArgOperand(2)))
158  Len = LengthArg->getZExtValue();
159  else
160  return nullptr;
161 
162  // See if we can get the length of the input string.
163  uint64_t SrcLen = GetStringLength(Src);
164  if (SrcLen == 0)
165  return nullptr;
166  --SrcLen; // Unbias length.
167 
168  // Handle the simple, do-nothing cases:
169  // strncat(x, "", c) -> x
170  // strncat(x, c, 0) -> x
171  if (SrcLen == 0 || Len == 0)
172  return Dst;
173 
174  // We don't optimize this case.
175  if (Len < SrcLen)
176  return nullptr;
177 
178  // strncat(x, s, c) -> strcat(x, s)
179  // s is constant so the strcat can be optimized further.
180  return emitStrLenMemCpy(Src, Dst, SrcLen, B);
181 }
182 
183 Value *LibCallSimplifier::optimizeStrChr(CallInst *CI, IRBuilder<> &B) {
185  FunctionType *FT = Callee->getFunctionType();
186  Value *SrcStr = CI->getArgOperand(0);
187 
188  // If the second operand is non-constant, see if we can compute the length
189  // of the input string and turn this into memchr.
190  ConstantInt *CharC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
191  if (!CharC) {
192  uint64_t Len = GetStringLength(SrcStr);
193  if (Len == 0 || !FT->getParamType(1)->isIntegerTy(32)) // memchr needs i32.
194  return nullptr;
195 
196  return emitMemChr(SrcStr, CI->getArgOperand(1), // include nul.
197  ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len),
198  B, DL, TLI);
199  }
200 
201  // Otherwise, the character is a constant, see if the first argument is
202  // a string literal. If so, we can constant fold.
203  StringRef Str;
204  if (!getConstantStringInfo(SrcStr, Str)) {
205  if (CharC->isZero()) // strchr(p, 0) -> p + strlen(p)
206  return B.CreateGEP(B.getInt8Ty(), SrcStr, emitStrLen(SrcStr, B, DL, TLI),
207  "strchr");
208  return nullptr;
209  }
210 
211  // Compute the offset, make sure to handle the case when we're searching for
212  // zero (a weird way to spell strlen).
213  size_t I = (0xFF & CharC->getSExtValue()) == 0
214  ? Str.size()
215  : Str.find(CharC->getSExtValue());
216  if (I == StringRef::npos) // Didn't find the char. strchr returns null.
217  return Constant::getNullValue(CI->getType());
218 
219  // strchr(s+n,c) -> gep(s+n+i,c)
220  return B.CreateGEP(B.getInt8Ty(), SrcStr, B.getInt64(I), "strchr");
221 }
222 
223 Value *LibCallSimplifier::optimizeStrRChr(CallInst *CI, IRBuilder<> &B) {
224  Value *SrcStr = CI->getArgOperand(0);
225  ConstantInt *CharC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
226 
227  // Cannot fold anything if we're not looking for a constant.
228  if (!CharC)
229  return nullptr;
230 
231  StringRef Str;
232  if (!getConstantStringInfo(SrcStr, Str)) {
233  // strrchr(s, 0) -> strchr(s, 0)
234  if (CharC->isZero())
235  return emitStrChr(SrcStr, '\0', B, TLI);
236  return nullptr;
237  }
238 
239  // Compute the offset.
240  size_t I = (0xFF & CharC->getSExtValue()) == 0
241  ? Str.size()
242  : Str.rfind(CharC->getSExtValue());
243  if (I == StringRef::npos) // Didn't find the char. Return null.
244  return Constant::getNullValue(CI->getType());
245 
246  // strrchr(s+n,c) -> gep(s+n+i,c)
247  return B.CreateGEP(B.getInt8Ty(), SrcStr, B.getInt64(I), "strrchr");
248 }
249 
250 Value *LibCallSimplifier::optimizeStrCmp(CallInst *CI, IRBuilder<> &B) {
251  Value *Str1P = CI->getArgOperand(0), *Str2P = CI->getArgOperand(1);
252  if (Str1P == Str2P) // strcmp(x,x) -> 0
253  return ConstantInt::get(CI->getType(), 0);
254 
255  StringRef Str1, Str2;
256  bool HasStr1 = getConstantStringInfo(Str1P, Str1);
257  bool HasStr2 = getConstantStringInfo(Str2P, Str2);
258 
259  // strcmp(x, y) -> cnst (if both x and y are constant strings)
260  if (HasStr1 && HasStr2)
261  return ConstantInt::get(CI->getType(), Str1.compare(Str2));
262 
263  if (HasStr1 && Str1.empty()) // strcmp("", x) -> -*x
264  return B.CreateNeg(
265  B.CreateZExt(B.CreateLoad(Str2P, "strcmpload"), CI->getType()));
266 
267  if (HasStr2 && Str2.empty()) // strcmp(x,"") -> *x
268  return B.CreateZExt(B.CreateLoad(Str1P, "strcmpload"), CI->getType());
269 
270  // strcmp(P, "x") -> memcmp(P, "x", 2)
271  uint64_t Len1 = GetStringLength(Str1P);
272  uint64_t Len2 = GetStringLength(Str2P);
273  if (Len1 && Len2) {
274  return emitMemCmp(Str1P, Str2P,
275  ConstantInt::get(DL.getIntPtrType(CI->getContext()),
276  std::min(Len1, Len2)),
277  B, DL, TLI);
278  }
279 
280  return nullptr;
281 }
282 
283 Value *LibCallSimplifier::optimizeStrNCmp(CallInst *CI, IRBuilder<> &B) {
284  Value *Str1P = CI->getArgOperand(0), *Str2P = CI->getArgOperand(1);
285  if (Str1P == Str2P) // strncmp(x,x,n) -> 0
286  return ConstantInt::get(CI->getType(), 0);
287 
288  // Get the length argument if it is constant.
289  uint64_t Length;
290  if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(CI->getArgOperand(2)))
291  Length = LengthArg->getZExtValue();
292  else
293  return nullptr;
294 
295  if (Length == 0) // strncmp(x,y,0) -> 0
296  return ConstantInt::get(CI->getType(), 0);
297 
298  if (Length == 1) // strncmp(x,y,1) -> memcmp(x,y,1)
299  return emitMemCmp(Str1P, Str2P, CI->getArgOperand(2), B, DL, TLI);
300 
301  StringRef Str1, Str2;
302  bool HasStr1 = getConstantStringInfo(Str1P, Str1);
303  bool HasStr2 = getConstantStringInfo(Str2P, Str2);
304 
305  // strncmp(x, y) -> cnst (if both x and y are constant strings)
306  if (HasStr1 && HasStr2) {
307  StringRef SubStr1 = Str1.substr(0, Length);
308  StringRef SubStr2 = Str2.substr(0, Length);
309  return ConstantInt::get(CI->getType(), SubStr1.compare(SubStr2));
310  }
311 
312  if (HasStr1 && Str1.empty()) // strncmp("", x, n) -> -*x
313  return B.CreateNeg(
314  B.CreateZExt(B.CreateLoad(Str2P, "strcmpload"), CI->getType()));
315 
316  if (HasStr2 && Str2.empty()) // strncmp(x, "", n) -> *x
317  return B.CreateZExt(B.CreateLoad(Str1P, "strcmpload"), CI->getType());
318 
319  return nullptr;
320 }
321 
322 Value *LibCallSimplifier::optimizeStrCpy(CallInst *CI, IRBuilder<> &B) {
323  Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1);
324  if (Dst == Src) // strcpy(x,x) -> x
325  return Src;
326 
327  // See if we can get the length of the input string.
328  uint64_t Len = GetStringLength(Src);
329  if (Len == 0)
330  return nullptr;
331 
332  // We have enough information to now generate the memcpy call to do the
333  // copy for us. Make a memcpy to copy the nul byte with align = 1.
334  B.CreateMemCpy(Dst, Src,
335  ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len), 1);
336  return Dst;
337 }
338 
339 Value *LibCallSimplifier::optimizeStpCpy(CallInst *CI, IRBuilder<> &B) {
341  Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1);
342  if (Dst == Src) { // stpcpy(x,x) -> x+strlen(x)
343  Value *StrLen = emitStrLen(Src, B, DL, TLI);
344  return StrLen ? B.CreateInBoundsGEP(B.getInt8Ty(), Dst, StrLen) : nullptr;
345  }
346 
347  // See if we can get the length of the input string.
348  uint64_t Len = GetStringLength(Src);
349  if (Len == 0)
350  return nullptr;
351 
352  Type *PT = Callee->getFunctionType()->getParamType(0);
353  Value *LenV = ConstantInt::get(DL.getIntPtrType(PT), Len);
354  Value *DstEnd = B.CreateGEP(B.getInt8Ty(), Dst,
355  ConstantInt::get(DL.getIntPtrType(PT), Len - 1));
356 
357  // We have enough information to now generate the memcpy call to do the
358  // copy for us. Make a memcpy to copy the nul byte with align = 1.
359  B.CreateMemCpy(Dst, Src, LenV, 1);
360  return DstEnd;
361 }
362 
363 Value *LibCallSimplifier::optimizeStrNCpy(CallInst *CI, IRBuilder<> &B) {
365  Value *Dst = CI->getArgOperand(0);
366  Value *Src = CI->getArgOperand(1);
367  Value *LenOp = CI->getArgOperand(2);
368 
369  // See if we can get the length of the input string.
370  uint64_t SrcLen = GetStringLength(Src);
371  if (SrcLen == 0)
372  return nullptr;
373  --SrcLen;
374 
375  if (SrcLen == 0) {
376  // strncpy(x, "", y) -> memset(x, '\0', y, 1)
377  B.CreateMemSet(Dst, B.getInt8('\0'), LenOp, 1);
378  return Dst;
379  }
380 
381  uint64_t Len;
382  if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(LenOp))
383  Len = LengthArg->getZExtValue();
384  else
385  return nullptr;
386 
387  if (Len == 0)
388  return Dst; // strncpy(x, y, 0) -> x
389 
390  // Let strncpy handle the zero padding
391  if (Len > SrcLen + 1)
392  return nullptr;
393 
394  Type *PT = Callee->getFunctionType()->getParamType(0);
395  // strncpy(x, s, c) -> memcpy(x, s, c, 1) [s and c are constant]
396  B.CreateMemCpy(Dst, Src, ConstantInt::get(DL.getIntPtrType(PT), Len), 1);
397 
398  return Dst;
399 }
400 
401 Value *LibCallSimplifier::optimizeStringLength(CallInst *CI, IRBuilder<> &B,
402  unsigned CharSize) {
403  Value *Src = CI->getArgOperand(0);
404 
405  // Constant folding: strlen("xyz") -> 3
406  if (uint64_t Len = GetStringLength(Src, CharSize))
407  return ConstantInt::get(CI->getType(), Len - 1);
408 
409  // If s is a constant pointer pointing to a string literal, we can fold
410  // strlen(s + x) to strlen(s) - x, when x is known to be in the range
411  // [0, strlen(s)] or the string has a single null terminator '\0' at the end.
412  // We only try to simplify strlen when the pointer s points to an array
413  // of i8. Otherwise, we would need to scale the offset x before doing the
414  // subtraction. This will make the optimization more complex, and it's not
415  // very useful because calling strlen for a pointer of other types is
416  // very uncommon.
417  if (GEPOperator *GEP = dyn_cast<GEPOperator>(Src)) {
418  if (!isGEPBasedOnPointerToString(GEP, CharSize))
419  return nullptr;
420 
422  if (getConstantDataArrayInfo(GEP->getOperand(0), Slice, CharSize)) {
423  uint64_t NullTermIdx;
424  if (Slice.Array == nullptr) {
425  NullTermIdx = 0;
426  } else {
427  NullTermIdx = ~((uint64_t)0);
428  for (uint64_t I = 0, E = Slice.Length; I < E; ++I) {
429  if (Slice.Array->getElementAsInteger(I + Slice.Offset) == 0) {
430  NullTermIdx = I;
431  break;
432  }
433  }
434  // If the string does not have '\0', leave it to strlen to compute
435  // its length.
436  if (NullTermIdx == ~((uint64_t)0))
437  return nullptr;
438  }
439 
440  Value *Offset = GEP->getOperand(2);
441  KnownBits Known = computeKnownBits(Offset, DL, 0, nullptr, CI, nullptr);
442  Known.Zero.flipAllBits();
443  uint64_t ArrSize =
444  cast<ArrayType>(GEP->getSourceElementType())->getNumElements();
445 
446  // KnownZero's bits are flipped, so zeros in KnownZero now represent
447  // bits known to be zeros in Offset, and ones in KnowZero represent
448  // bits unknown in Offset. Therefore, Offset is known to be in range
449  // [0, NullTermIdx] when the flipped KnownZero is non-negative and
450  // unsigned-less-than NullTermIdx.
451  //
452  // If Offset is not provably in the range [0, NullTermIdx], we can still
453  // optimize if we can prove that the program has undefined behavior when
454  // Offset is outside that range. That is the case when GEP->getOperand(0)
455  // is a pointer to an object whose memory extent is NullTermIdx+1.
456  if ((Known.Zero.isNonNegative() && Known.Zero.ule(NullTermIdx)) ||
457  (GEP->isInBounds() && isa<GlobalVariable>(GEP->getOperand(0)) &&
458  NullTermIdx == ArrSize - 1)) {
459  Offset = B.CreateSExtOrTrunc(Offset, CI->getType());
460  return B.CreateSub(ConstantInt::get(CI->getType(), NullTermIdx),
461  Offset);
462  }
463  }
464 
465  return nullptr;
466  }
467 
468  // strlen(x?"foo":"bars") --> x ? 3 : 4
469  if (SelectInst *SI = dyn_cast<SelectInst>(Src)) {
470  uint64_t LenTrue = GetStringLength(SI->getTrueValue(), CharSize);
471  uint64_t LenFalse = GetStringLength(SI->getFalseValue(), CharSize);
472  if (LenTrue && LenFalse) {
473  ORE.emit([&]() {
474  return OptimizationRemark("instcombine", "simplify-libcalls", CI)
475  << "folded strlen(select) to select of constants";
476  });
477  return B.CreateSelect(SI->getCondition(),
478  ConstantInt::get(CI->getType(), LenTrue - 1),
479  ConstantInt::get(CI->getType(), LenFalse - 1));
480  }
481  }
482 
483  // strlen(x) != 0 --> *x != 0
484  // strlen(x) == 0 --> *x == 0
486  return B.CreateZExt(B.CreateLoad(Src, "strlenfirst"), CI->getType());
487 
488  return nullptr;
489 }
490 
491 Value *LibCallSimplifier::optimizeStrLen(CallInst *CI, IRBuilder<> &B) {
492  return optimizeStringLength(CI, B, 8);
493 }
494 
495 Value *LibCallSimplifier::optimizeWcslen(CallInst *CI, IRBuilder<> &B) {
496  Module &M = *CI->getParent()->getParent()->getParent();
497  unsigned WCharSize = TLI->getWCharSize(M) * 8;
498  // We cannot perform this optimization without wchar_size metadata.
499  if (WCharSize == 0)
500  return nullptr;
501 
502  return optimizeStringLength(CI, B, WCharSize);
503 }
504 
505 Value *LibCallSimplifier::optimizeStrPBrk(CallInst *CI, IRBuilder<> &B) {
506  StringRef S1, S2;
507  bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1);
508  bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2);
509 
510  // strpbrk(s, "") -> nullptr
511  // strpbrk("", s) -> nullptr
512  if ((HasS1 && S1.empty()) || (HasS2 && S2.empty()))
513  return Constant::getNullValue(CI->getType());
514 
515  // Constant folding.
516  if (HasS1 && HasS2) {
517  size_t I = S1.find_first_of(S2);
518  if (I == StringRef::npos) // No match.
519  return Constant::getNullValue(CI->getType());
520 
521  return B.CreateGEP(B.getInt8Ty(), CI->getArgOperand(0), B.getInt64(I),
522  "strpbrk");
523  }
524 
525  // strpbrk(s, "a") -> strchr(s, 'a')
526  if (HasS2 && S2.size() == 1)
527  return emitStrChr(CI->getArgOperand(0), S2[0], B, TLI);
528 
529  return nullptr;
530 }
531 
532 Value *LibCallSimplifier::optimizeStrTo(CallInst *CI, IRBuilder<> &B) {
533  Value *EndPtr = CI->getArgOperand(1);
534  if (isa<ConstantPointerNull>(EndPtr)) {
535  // With a null EndPtr, this function won't capture the main argument.
536  // It would be readonly too, except that it still may write to errno.
537  CI->addParamAttr(0, Attribute::NoCapture);
538  }
539 
540  return nullptr;
541 }
542 
543 Value *LibCallSimplifier::optimizeStrSpn(CallInst *CI, IRBuilder<> &B) {
544  StringRef S1, S2;
545  bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1);
546  bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2);
547 
548  // strspn(s, "") -> 0
549  // strspn("", s) -> 0
550  if ((HasS1 && S1.empty()) || (HasS2 && S2.empty()))
551  return Constant::getNullValue(CI->getType());
552 
553  // Constant folding.
554  if (HasS1 && HasS2) {
555  size_t Pos = S1.find_first_not_of(S2);
556  if (Pos == StringRef::npos)
557  Pos = S1.size();
558  return ConstantInt::get(CI->getType(), Pos);
559  }
560 
561  return nullptr;
562 }
563 
564 Value *LibCallSimplifier::optimizeStrCSpn(CallInst *CI, IRBuilder<> &B) {
565  StringRef S1, S2;
566  bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1);
567  bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2);
568 
569  // strcspn("", s) -> 0
570  if (HasS1 && S1.empty())
571  return Constant::getNullValue(CI->getType());
572 
573  // Constant folding.
574  if (HasS1 && HasS2) {
575  size_t Pos = S1.find_first_of(S2);
576  if (Pos == StringRef::npos)
577  Pos = S1.size();
578  return ConstantInt::get(CI->getType(), Pos);
579  }
580 
581  // strcspn(s, "") -> strlen(s)
582  if (HasS2 && S2.empty())
583  return emitStrLen(CI->getArgOperand(0), B, DL, TLI);
584 
585  return nullptr;
586 }
587 
588 Value *LibCallSimplifier::optimizeStrStr(CallInst *CI, IRBuilder<> &B) {
589  // fold strstr(x, x) -> x.
590  if (CI->getArgOperand(0) == CI->getArgOperand(1))
591  return B.CreateBitCast(CI->getArgOperand(0), CI->getType());
592 
593  // fold strstr(a, b) == a -> strncmp(a, b, strlen(b)) == 0
595  Value *StrLen = emitStrLen(CI->getArgOperand(1), B, DL, TLI);
596  if (!StrLen)
597  return nullptr;
598  Value *StrNCmp = emitStrNCmp(CI->getArgOperand(0), CI->getArgOperand(1),
599  StrLen, B, DL, TLI);
600  if (!StrNCmp)
601  return nullptr;
602  for (auto UI = CI->user_begin(), UE = CI->user_end(); UI != UE;) {
603  ICmpInst *Old = cast<ICmpInst>(*UI++);
604  Value *Cmp =
605  B.CreateICmp(Old->getPredicate(), StrNCmp,
606  ConstantInt::getNullValue(StrNCmp->getType()), "cmp");
607  replaceAllUsesWith(Old, Cmp);
608  }
609  return CI;
610  }
611 
612  // See if either input string is a constant string.
613  StringRef SearchStr, ToFindStr;
614  bool HasStr1 = getConstantStringInfo(CI->getArgOperand(0), SearchStr);
615  bool HasStr2 = getConstantStringInfo(CI->getArgOperand(1), ToFindStr);
616 
617  // fold strstr(x, "") -> x.
618  if (HasStr2 && ToFindStr.empty())
619  return B.CreateBitCast(CI->getArgOperand(0), CI->getType());
620 
621  // If both strings are known, constant fold it.
622  if (HasStr1 && HasStr2) {
623  size_t Offset = SearchStr.find(ToFindStr);
624 
625  if (Offset == StringRef::npos) // strstr("foo", "bar") -> null
626  return Constant::getNullValue(CI->getType());
627 
628  // strstr("abcd", "bc") -> gep((char*)"abcd", 1)
629  Value *Result = castToCStr(CI->getArgOperand(0), B);
630  Result = B.CreateConstInBoundsGEP1_64(Result, Offset, "strstr");
631  return B.CreateBitCast(Result, CI->getType());
632  }
633 
634  // fold strstr(x, "y") -> strchr(x, 'y').
635  if (HasStr2 && ToFindStr.size() == 1) {
636  Value *StrChr = emitStrChr(CI->getArgOperand(0), ToFindStr[0], B, TLI);
637  return StrChr ? B.CreateBitCast(StrChr, CI->getType()) : nullptr;
638  }
639  return nullptr;
640 }
641 
642 Value *LibCallSimplifier::optimizeMemChr(CallInst *CI, IRBuilder<> &B) {
643  Value *SrcStr = CI->getArgOperand(0);
644  ConstantInt *CharC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
646 
647  // memchr(x, y, 0) -> null
648  if (LenC && LenC->isZero())
649  return Constant::getNullValue(CI->getType());
650 
651  // From now on we need at least constant length and string.
652  StringRef Str;
653  if (!LenC || !getConstantStringInfo(SrcStr, Str, 0, /*TrimAtNul=*/false))
654  return nullptr;
655 
656  // Truncate the string to LenC. If Str is smaller than LenC we will still only
657  // scan the string, as reading past the end of it is undefined and we can just
658  // return null if we don't find the char.
659  Str = Str.substr(0, LenC->getZExtValue());
660 
661  // If the char is variable but the input str and length are not we can turn
662  // this memchr call into a simple bit field test. Of course this only works
663  // when the return value is only checked against null.
664  //
665  // It would be really nice to reuse switch lowering here but we can't change
666  // the CFG at this point.
667  //
668  // memchr("\r\n", C, 2) != nullptr -> (C & ((1 << '\r') | (1 << '\n'))) != 0
669  // after bounds check.
670  if (!CharC && !Str.empty() && isOnlyUsedInZeroEqualityComparison(CI)) {
671  unsigned char Max =
672  *std::max_element(reinterpret_cast<const unsigned char *>(Str.begin()),
673  reinterpret_cast<const unsigned char *>(Str.end()));
674 
675  // Make sure the bit field we're about to create fits in a register on the
676  // target.
677  // FIXME: On a 64 bit architecture this prevents us from using the
678  // interesting range of alpha ascii chars. We could do better by emitting
679  // two bitfields or shifting the range by 64 if no lower chars are used.
680  if (!DL.fitsInLegalInteger(Max + 1))
681  return nullptr;
682 
683  // For the bit field use a power-of-2 type with at least 8 bits to avoid
684  // creating unnecessary illegal types.
685  unsigned char Width = NextPowerOf2(std::max((unsigned char)7, Max));
686 
687  // Now build the bit field.
688  APInt Bitfield(Width, 0);
689  for (char C : Str)
690  Bitfield.setBit((unsigned char)C);
691  Value *BitfieldC = B.getInt(Bitfield);
692 
693  // First check that the bit field access is within bounds.
694  Value *C = B.CreateZExtOrTrunc(CI->getArgOperand(1), BitfieldC->getType());
695  Value *Bounds = B.CreateICmp(ICmpInst::ICMP_ULT, C, B.getIntN(Width, Width),
696  "memchr.bounds");
697 
698  // Create code that checks if the given bit is set in the field.
699  Value *Shl = B.CreateShl(B.getIntN(Width, 1ULL), C);
700  Value *Bits = B.CreateIsNotNull(B.CreateAnd(Shl, BitfieldC), "memchr.bits");
701 
702  // Finally merge both checks and cast to pointer type. The inttoptr
703  // implicitly zexts the i1 to intptr type.
704  return B.CreateIntToPtr(B.CreateAnd(Bounds, Bits, "memchr"), CI->getType());
705  }
706 
707  // Check if all arguments are constants. If so, we can constant fold.
708  if (!CharC)
709  return nullptr;
710 
711  // Compute the offset.
712  size_t I = Str.find(CharC->getSExtValue() & 0xFF);
713  if (I == StringRef::npos) // Didn't find the char. memchr returns null.
714  return Constant::getNullValue(CI->getType());
715 
716  // memchr(s+n,c,l) -> gep(s+n+i,c)
717  return B.CreateGEP(B.getInt8Ty(), SrcStr, B.getInt64(I), "memchr");
718 }
719 
720 Value *LibCallSimplifier::optimizeMemCmp(CallInst *CI, IRBuilder<> &B) {
721  Value *LHS = CI->getArgOperand(0), *RHS = CI->getArgOperand(1);
722 
723  if (LHS == RHS) // memcmp(s,s,x) -> 0
724  return Constant::getNullValue(CI->getType());
725 
726  // Make sure we have a constant length.
728  if (!LenC)
729  return nullptr;
730 
731  uint64_t Len = LenC->getZExtValue();
732  if (Len == 0) // memcmp(s1,s2,0) -> 0
733  return Constant::getNullValue(CI->getType());
734 
735  // memcmp(S1,S2,1) -> *(unsigned char*)LHS - *(unsigned char*)RHS
736  if (Len == 1) {
737  Value *LHSV = B.CreateZExt(B.CreateLoad(castToCStr(LHS, B), "lhsc"),
738  CI->getType(), "lhsv");
739  Value *RHSV = B.CreateZExt(B.CreateLoad(castToCStr(RHS, B), "rhsc"),
740  CI->getType(), "rhsv");
741  return B.CreateSub(LHSV, RHSV, "chardiff");
742  }
743 
744  // memcmp(S1,S2,N/8)==0 -> (*(intN_t*)S1 != *(intN_t*)S2)==0
745  // TODO: The case where both inputs are constants does not need to be limited
746  // to legal integers or equality comparison. See block below this.
747  if (DL.isLegalInteger(Len * 8) && isOnlyUsedInZeroEqualityComparison(CI)) {
748  IntegerType *IntType = IntegerType::get(CI->getContext(), Len * 8);
749  unsigned PrefAlignment = DL.getPrefTypeAlignment(IntType);
750 
751  // First, see if we can fold either argument to a constant.
752  Value *LHSV = nullptr;
753  if (auto *LHSC = dyn_cast<Constant>(LHS)) {
754  LHSC = ConstantExpr::getBitCast(LHSC, IntType->getPointerTo());
755  LHSV = ConstantFoldLoadFromConstPtr(LHSC, IntType, DL);
756  }
757  Value *RHSV = nullptr;
758  if (auto *RHSC = dyn_cast<Constant>(RHS)) {
759  RHSC = ConstantExpr::getBitCast(RHSC, IntType->getPointerTo());
760  RHSV = ConstantFoldLoadFromConstPtr(RHSC, IntType, DL);
761  }
762 
763  // Don't generate unaligned loads. If either source is constant data,
764  // alignment doesn't matter for that source because there is no load.
765  if ((LHSV || getKnownAlignment(LHS, DL, CI) >= PrefAlignment) &&
766  (RHSV || getKnownAlignment(RHS, DL, CI) >= PrefAlignment)) {
767  if (!LHSV) {
768  Type *LHSPtrTy =
769  IntType->getPointerTo(LHS->getType()->getPointerAddressSpace());
770  LHSV = B.CreateLoad(B.CreateBitCast(LHS, LHSPtrTy), "lhsv");
771  }
772  if (!RHSV) {
773  Type *RHSPtrTy =
774  IntType->getPointerTo(RHS->getType()->getPointerAddressSpace());
775  RHSV = B.CreateLoad(B.CreateBitCast(RHS, RHSPtrTy), "rhsv");
776  }
777  return B.CreateZExt(B.CreateICmpNE(LHSV, RHSV), CI->getType(), "memcmp");
778  }
779  }
780 
781  // Constant folding: memcmp(x, y, Len) -> constant (all arguments are const).
782  // TODO: This is limited to i8 arrays.
783  StringRef LHSStr, RHSStr;
784  if (getConstantStringInfo(LHS, LHSStr) &&
785  getConstantStringInfo(RHS, RHSStr)) {
786  // Make sure we're not reading out-of-bounds memory.
787  if (Len > LHSStr.size() || Len > RHSStr.size())
788  return nullptr;
789  // Fold the memcmp and normalize the result. This way we get consistent
790  // results across multiple platforms.
791  uint64_t Ret = 0;
792  int Cmp = memcmp(LHSStr.data(), RHSStr.data(), Len);
793  if (Cmp < 0)
794  Ret = -1;
795  else if (Cmp > 0)
796  Ret = 1;
797  return ConstantInt::get(CI->getType(), Ret);
798  }
799 
800  return nullptr;
801 }
802 
803 Value *LibCallSimplifier::optimizeMemCpy(CallInst *CI, IRBuilder<> &B) {
804  // memcpy(x, y, n) -> llvm.memcpy(x, y, n, 1)
805  B.CreateMemCpy(CI->getArgOperand(0), CI->getArgOperand(1),
806  CI->getArgOperand(2), 1);
807  return CI->getArgOperand(0);
808 }
809 
810 Value *LibCallSimplifier::optimizeMemMove(CallInst *CI, IRBuilder<> &B) {
811  // memmove(x, y, n) -> llvm.memmove(x, y, n, 1)
812  B.CreateMemMove(CI->getArgOperand(0), CI->getArgOperand(1),
813  CI->getArgOperand(2), 1);
814  return CI->getArgOperand(0);
815 }
816 
817 // TODO: Does this belong in BuildLibCalls or should all of those similar
818 // functions be moved here?
820  IRBuilder<> &B, const TargetLibraryInfo &TLI) {
821  LibFunc Func;
822  if (!TLI.getLibFunc("calloc", Func) || !TLI.has(Func))
823  return nullptr;
824 
825  Module *M = B.GetInsertBlock()->getModule();
826  const DataLayout &DL = M->getDataLayout();
827  IntegerType *PtrType = DL.getIntPtrType((B.GetInsertBlock()->getContext()));
828  Value *Calloc = M->getOrInsertFunction("calloc", Attrs, B.getInt8PtrTy(),
829  PtrType, PtrType);
830  CallInst *CI = B.CreateCall(Calloc, { Num, Size }, "calloc");
831 
832  if (const auto *F = dyn_cast<Function>(Calloc->stripPointerCasts()))
833  CI->setCallingConv(F->getCallingConv());
834 
835  return CI;
836 }
837 
838 /// Fold memset[_chk](malloc(n), 0, n) --> calloc(1, n).
840  const TargetLibraryInfo &TLI) {
841  // This has to be a memset of zeros (bzero).
842  auto *FillValue = dyn_cast<ConstantInt>(Memset->getArgOperand(1));
843  if (!FillValue || FillValue->getZExtValue() != 0)
844  return nullptr;
845 
846  // TODO: We should handle the case where the malloc has more than one use.
847  // This is necessary to optimize common patterns such as when the result of
848  // the malloc is checked against null or when a memset intrinsic is used in
849  // place of a memset library call.
850  auto *Malloc = dyn_cast<CallInst>(Memset->getArgOperand(0));
851  if (!Malloc || !Malloc->hasOneUse())
852  return nullptr;
853 
854  // Is the inner call really malloc()?
855  Function *InnerCallee = Malloc->getCalledFunction();
856  if (!InnerCallee)
857  return nullptr;
858 
859  LibFunc Func;
860  if (!TLI.getLibFunc(*InnerCallee, Func) || !TLI.has(Func) ||
861  Func != LibFunc_malloc)
862  return nullptr;
863 
864  // The memset must cover the same number of bytes that are malloc'd.
865  if (Memset->getArgOperand(2) != Malloc->getArgOperand(0))
866  return nullptr;
867 
868  // Replace the malloc with a calloc. We need the data layout to know what the
869  // actual size of a 'size_t' parameter is.
870  B.SetInsertPoint(Malloc->getParent(), ++Malloc->getIterator());
871  const DataLayout &DL = Malloc->getModule()->getDataLayout();
872  IntegerType *SizeType = DL.getIntPtrType(B.GetInsertBlock()->getContext());
873  Value *Calloc = emitCalloc(ConstantInt::get(SizeType, 1),
874  Malloc->getArgOperand(0), Malloc->getAttributes(),
875  B, TLI);
876  if (!Calloc)
877  return nullptr;
878 
879  Malloc->replaceAllUsesWith(Calloc);
880  Malloc->eraseFromParent();
881 
882  return Calloc;
883 }
884 
885 Value *LibCallSimplifier::optimizeMemSet(CallInst *CI, IRBuilder<> &B) {
886  if (auto *Calloc = foldMallocMemset(CI, B, *TLI))
887  return Calloc;
888 
889  // memset(p, v, n) -> llvm.memset(p, v, n, 1)
890  Value *Val = B.CreateIntCast(CI->getArgOperand(1), B.getInt8Ty(), false);
891  B.CreateMemSet(CI->getArgOperand(0), Val, CI->getArgOperand(2), 1);
892  return CI->getArgOperand(0);
893 }
894 
895 //===----------------------------------------------------------------------===//
896 // Math Library Optimizations
897 //===----------------------------------------------------------------------===//
898 
899 /// Return a variant of Val with float type.
900 /// Currently this works in two cases: If Val is an FPExtension of a float
901 /// value to something bigger, simply return the operand.
902 /// If Val is a ConstantFP but can be converted to a float ConstantFP without
903 /// loss of precision do so.
905  if (FPExtInst *Cast = dyn_cast<FPExtInst>(Val)) {
906  Value *Op = Cast->getOperand(0);
907  if (Op->getType()->isFloatTy())
908  return Op;
909  }
910  if (ConstantFP *Const = dyn_cast<ConstantFP>(Val)) {
911  APFloat F = Const->getValueAPF();
912  bool losesInfo;
914  &losesInfo);
915  if (!losesInfo)
916  return ConstantFP::get(Const->getContext(), F);
917  }
918  return nullptr;
919 }
920 
921 /// Shrink double -> float for unary functions like 'floor'.
923  bool CheckRetType) {
925  // We know this libcall has a valid prototype, but we don't know which.
926  if (!CI->getType()->isDoubleTy())
927  return nullptr;
928 
929  if (CheckRetType) {
930  // Check if all the uses for function like 'sin' are converted to float.
931  for (User *U : CI->users()) {
932  FPTruncInst *Cast = dyn_cast<FPTruncInst>(U);
933  if (!Cast || !Cast->getType()->isFloatTy())
934  return nullptr;
935  }
936  }
937 
938  // If this is something like 'floor((double)floatval)', convert to floorf.
940  if (V == nullptr)
941  return nullptr;
942 
943  // If call isn't an intrinsic, check that it isn't within a function with the
944  // same name as the float version of this call.
945  //
946  // e.g. inline float expf(float val) { return (float) exp((double) val); }
947  //
948  // A similar such definition exists in the MinGW-w64 math.h header file which
949  // when compiled with -O2 -ffast-math causes the generation of infinite loops
950  // where expf is called.
951  if (!Callee->isIntrinsic()) {
952  const Function *F = CI->getFunction();
953  StringRef FName = F->getName();
954  StringRef CalleeName = Callee->getName();
955  if ((FName.size() == (CalleeName.size() + 1)) &&
956  (FName.back() == 'f') &&
957  FName.startswith(CalleeName))
958  return nullptr;
959  }
960 
961  // Propagate fast-math flags from the existing call to the new call.
964 
965  // floor((double)floatval) -> (double)floorf(floatval)
966  if (Callee->isIntrinsic()) {
967  Module *M = CI->getModule();
968  Intrinsic::ID IID = Callee->getIntrinsicID();
970  V = B.CreateCall(F, V);
971  } else {
972  // The call is a library call rather than an intrinsic.
973  V = emitUnaryFloatFnCall(V, Callee->getName(), B, Callee->getAttributes());
974  }
975 
976  return B.CreateFPExt(V, B.getDoubleTy());
977 }
978 
979 // Replace a libcall \p CI with a call to intrinsic \p IID
981  // Propagate fast-math flags from the existing call to the new call.
984 
985  Module *M = CI->getModule();
986  Value *V = CI->getArgOperand(0);
987  Function *F = Intrinsic::getDeclaration(M, IID, CI->getType());
988  CallInst *NewCall = B.CreateCall(F, V);
989  NewCall->takeName(CI);
990  return NewCall;
991 }
992 
993 /// Shrink double -> float for binary functions like 'fmin/fmax'.
996  // We know this libcall has a valid prototype, but we don't know which.
997  if (!CI->getType()->isDoubleTy())
998  return nullptr;
999 
1000  // If this is something like 'fmin((double)floatval1, (double)floatval2)',
1001  // or fmin(1.0, (double)floatval), then we convert it to fminf.
1003  if (V1 == nullptr)
1004  return nullptr;
1006  if (V2 == nullptr)
1007  return nullptr;
1008 
1009  // Propagate fast-math flags from the existing call to the new call.
1012 
1013  // fmin((double)floatval1, (double)floatval2)
1014  // -> (double)fminf(floatval1, floatval2)
1015  // TODO: Handle intrinsics in the same way as in optimizeUnaryDoubleFP().
1016  Value *V = emitBinaryFloatFnCall(V1, V2, Callee->getName(), B,
1017  Callee->getAttributes());
1018  return B.CreateFPExt(V, B.getDoubleTy());
1019 }
1020 
1021 // cabs(z) -> sqrt((creal(z)*creal(z)) + (cimag(z)*cimag(z)))
1022 Value *LibCallSimplifier::optimizeCAbs(CallInst *CI, IRBuilder<> &B) {
1023  if (!CI->isFast())
1024  return nullptr;
1025 
1026  // Propagate fast-math flags from the existing call to new instructions.
1029 
1030  Value *Real, *Imag;
1031  if (CI->getNumArgOperands() == 1) {
1032  Value *Op = CI->getArgOperand(0);
1033  assert(Op->getType()->isArrayTy() && "Unexpected signature for cabs!");
1034  Real = B.CreateExtractValue(Op, 0, "real");
1035  Imag = B.CreateExtractValue(Op, 1, "imag");
1036  } else {
1037  assert(CI->getNumArgOperands() == 2 && "Unexpected signature for cabs!");
1038  Real = CI->getArgOperand(0);
1039  Imag = CI->getArgOperand(1);
1040  }
1041 
1042  Value *RealReal = B.CreateFMul(Real, Real);
1043  Value *ImagImag = B.CreateFMul(Imag, Imag);
1044 
1045  Function *FSqrt = Intrinsic::getDeclaration(CI->getModule(), Intrinsic::sqrt,
1046  CI->getType());
1047  return B.CreateCall(FSqrt, B.CreateFAdd(RealReal, ImagImag), "cabs");
1048 }
1049 
1050 Value *LibCallSimplifier::optimizeCos(CallInst *CI, IRBuilder<> &B) {
1052  Value *Ret = nullptr;
1053  StringRef Name = Callee->getName();
1054  if (UnsafeFPShrink && Name == "cos" && hasFloatVersion(Name))
1055  Ret = optimizeUnaryDoubleFP(CI, B, true);
1056 
1057  // cos(-x) -> cos(x)
1058  Value *Op1 = CI->getArgOperand(0);
1059  if (BinaryOperator::isFNeg(Op1)) {
1060  BinaryOperator *BinExpr = cast<BinaryOperator>(Op1);
1061  return B.CreateCall(Callee, BinExpr->getOperand(1), "cos");
1062  }
1063  return Ret;
1064 }
1065 
1066 static Value *getPow(Value *InnerChain[33], unsigned Exp, IRBuilder<> &B) {
1067  // Multiplications calculated using Addition Chains.
1068  // Refer: http://wwwhomes.uni-bielefeld.de/achim/addition_chain.html
1069 
1070  assert(Exp != 0 && "Incorrect exponent 0 not handled");
1071 
1072  if (InnerChain[Exp])
1073  return InnerChain[Exp];
1074 
1075  static const unsigned AddChain[33][2] = {
1076  {0, 0}, // Unused.
1077  {0, 0}, // Unused (base case = pow1).
1078  {1, 1}, // Unused (pre-computed).
1079  {1, 2}, {2, 2}, {2, 3}, {3, 3}, {2, 5}, {4, 4},
1080  {1, 8}, {5, 5}, {1, 10}, {6, 6}, {4, 9}, {7, 7},
1081  {3, 12}, {8, 8}, {8, 9}, {2, 16}, {1, 18}, {10, 10},
1082  {6, 15}, {11, 11}, {3, 20}, {12, 12}, {8, 17}, {13, 13},
1083  {3, 24}, {14, 14}, {4, 25}, {15, 15}, {3, 28}, {16, 16},
1084  };
1085 
1086  InnerChain[Exp] = B.CreateFMul(getPow(InnerChain, AddChain[Exp][0], B),
1087  getPow(InnerChain, AddChain[Exp][1], B));
1088  return InnerChain[Exp];
1089 }
1090 
1091 /// Use square root in place of pow(x, +/-0.5).
1092 Value *LibCallSimplifier::replacePowWithSqrt(CallInst *Pow, IRBuilder<> &B) {
1093  // TODO: There is some subset of 'fast' under which these transforms should
1094  // be allowed.
1095  if (!Pow->isFast())
1096  return nullptr;
1097 
1098  const APFloat *Arg1C;
1099  if (!match(Pow->getArgOperand(1), m_APFloat(Arg1C)))
1100  return nullptr;
1101  if (!Arg1C->isExactlyValue(0.5) && !Arg1C->isExactlyValue(-0.5))
1102  return nullptr;
1103 
1104  // Fast-math flags from the pow() are propagated to all replacement ops.
1107  Type *Ty = Pow->getType();
1108  Value *Sqrt;
1109  if (Pow->hasFnAttr(Attribute::ReadNone)) {
1110  // We know that errno is never set, so replace with an intrinsic:
1111  // pow(x, 0.5) --> llvm.sqrt(x)
1112  // llvm.pow(x, 0.5) --> llvm.sqrt(x)
1113  auto *F = Intrinsic::getDeclaration(Pow->getModule(), Intrinsic::sqrt, Ty);
1114  Sqrt = B.CreateCall(F, Pow->getArgOperand(0));
1115  } else if (hasUnaryFloatFn(TLI, Ty, LibFunc_sqrt, LibFunc_sqrtf,
1116  LibFunc_sqrtl)) {
1117  // Errno could be set, so we must use a sqrt libcall.
1118  // TODO: We also should check that the target can in fact lower the sqrt
1119  // libcall. We currently have no way to ask this question, so we ask
1120  // whether the target has a sqrt libcall which is not exactly the same.
1121  Sqrt = emitUnaryFloatFnCall(Pow->getArgOperand(0),
1122  TLI->getName(LibFunc_sqrt), B,
1123  Pow->getCalledFunction()->getAttributes());
1124  } else {
1125  // We can't replace with an intrinsic or a libcall.
1126  return nullptr;
1127  }
1128 
1129  // If this is pow(x, -0.5), get the reciprocal.
1130  if (Arg1C->isExactlyValue(-0.5))
1131  Sqrt = B.CreateFDiv(ConstantFP::get(Ty, 1.0), Sqrt);
1132 
1133  return Sqrt;
1134 }
1135 
1136 Value *LibCallSimplifier::optimizePow(CallInst *CI, IRBuilder<> &B) {
1138  Value *Ret = nullptr;
1139  StringRef Name = Callee->getName();
1140  if (UnsafeFPShrink && Name == "pow" && hasFloatVersion(Name))
1141  Ret = optimizeUnaryDoubleFP(CI, B, true);
1142 
1143  Value *Op1 = CI->getArgOperand(0), *Op2 = CI->getArgOperand(1);
1144 
1145  // pow(1.0, x) -> 1.0
1146  if (match(Op1, m_SpecificFP(1.0)))
1147  return Op1;
1148  // pow(2.0, x) -> llvm.exp2(x)
1149  if (match(Op1, m_SpecificFP(2.0))) {
1150  Value *Exp2 = Intrinsic::getDeclaration(CI->getModule(), Intrinsic::exp2,
1151  CI->getType());
1152  return B.CreateCall(Exp2, Op2, "exp2");
1153  }
1154 
1155  // There's no llvm.exp10 intrinsic yet, but, maybe, some day there will
1156  // be one.
1157  if (ConstantFP *Op1C = dyn_cast<ConstantFP>(Op1)) {
1158  // pow(10.0, x) -> exp10(x)
1159  if (Op1C->isExactlyValue(10.0) &&
1160  hasUnaryFloatFn(TLI, Op1->getType(), LibFunc_exp10, LibFunc_exp10f,
1161  LibFunc_exp10l))
1162  return emitUnaryFloatFnCall(Op2, TLI->getName(LibFunc_exp10), B,
1163  Callee->getAttributes());
1164  }
1165 
1166  // pow(exp(x), y) -> exp(x * y)
1167  // pow(exp2(x), y) -> exp2(x * y)
1168  // We enable these only with fast-math. Besides rounding differences, the
1169  // transformation changes overflow and underflow behavior quite dramatically.
1170  // Example: x = 1000, y = 0.001.
1171  // pow(exp(x), y) = pow(inf, 0.001) = inf, whereas exp(x*y) = exp(1).
1172  auto *OpC = dyn_cast<CallInst>(Op1);
1173  if (OpC && OpC->isFast() && CI->isFast()) {
1174  LibFunc Func;
1175  Function *OpCCallee = OpC->getCalledFunction();
1176  if (OpCCallee && TLI->getLibFunc(OpCCallee->getName(), Func) &&
1177  TLI->has(Func) && (Func == LibFunc_exp || Func == LibFunc_exp2)) {
1180  Value *FMul = B.CreateFMul(OpC->getArgOperand(0), Op2, "mul");
1181  return emitUnaryFloatFnCall(FMul, OpCCallee->getName(), B,
1182  OpCCallee->getAttributes());
1183  }
1184  }
1185 
1186  if (Value *Sqrt = replacePowWithSqrt(CI, B))
1187  return Sqrt;
1188 
1189  ConstantFP *Op2C = dyn_cast<ConstantFP>(Op2);
1190  if (!Op2C)
1191  return Ret;
1192 
1193  if (Op2C->getValueAPF().isZero()) // pow(x, 0.0) -> 1.0
1194  return ConstantFP::get(CI->getType(), 1.0);
1195 
1196  // FIXME: Correct the transforms and pull this into replacePowWithSqrt().
1197  if (Op2C->isExactlyValue(0.5) &&
1198  hasUnaryFloatFn(TLI, Op2->getType(), LibFunc_sqrt, LibFunc_sqrtf,
1199  LibFunc_sqrtl)) {
1200  // Expand pow(x, 0.5) to (x == -infinity ? +infinity : fabs(sqrt(x))).
1201  // This is faster than calling pow, and still handles negative zero
1202  // and negative infinity correctly.
1203  // TODO: In finite-only mode, this could be just fabs(sqrt(x)).
1204  Value *Inf = ConstantFP::getInfinity(CI->getType());
1205  Value *NegInf = ConstantFP::getInfinity(CI->getType(), true);
1206 
1207  // TODO: As above, we should lower to the sqrt intrinsic if the pow is an
1208  // intrinsic, to match errno semantics.
1209  Value *Sqrt = emitUnaryFloatFnCall(Op1, "sqrt", B, Callee->getAttributes());
1210 
1211  Module *M = Callee->getParent();
1212  Function *FabsF = Intrinsic::getDeclaration(M, Intrinsic::fabs,
1213  CI->getType());
1214  Value *FAbs = B.CreateCall(FabsF, Sqrt);
1215 
1216  Value *FCmp = B.CreateFCmpOEQ(Op1, NegInf);
1217  Value *Sel = B.CreateSelect(FCmp, Inf, FAbs);
1218  return Sel;
1219  }
1220 
1221  // Propagate fast-math-flags from the call to any created instructions.
1224  // pow(x, 1.0) --> x
1225  if (Op2C->isExactlyValue(1.0))
1226  return Op1;
1227  // pow(x, 2.0) --> x * x
1228  if (Op2C->isExactlyValue(2.0))
1229  return B.CreateFMul(Op1, Op1, "pow2");
1230  // pow(x, -1.0) --> 1.0 / x
1231  if (Op2C->isExactlyValue(-1.0))
1232  return B.CreateFDiv(ConstantFP::get(CI->getType(), 1.0), Op1, "powrecip");
1233 
1234  // In -ffast-math, generate repeated fmul instead of generating pow(x, n).
1235  if (CI->isFast()) {
1236  APFloat V = abs(Op2C->getValueAPF());
1237  // We limit to a max of 7 fmul(s). Thus max exponent is 32.
1238  // This transformation applies to integer exponents only.
1239  if (V.compare(APFloat(V.getSemantics(), 32.0)) == APFloat::cmpGreaterThan ||
1240  !V.isInteger())
1241  return nullptr;
1242 
1243  // We will memoize intermediate products of the Addition Chain.
1244  Value *InnerChain[33] = {nullptr};
1245  InnerChain[1] = Op1;
1246  InnerChain[2] = B.CreateFMul(Op1, Op1);
1247 
1248  // We cannot readily convert a non-double type (like float) to a double.
1249  // So we first convert V to something which could be converted to double.
1250  bool Ignored;
1252 
1253  Value *FMul = getPow(InnerChain, V.convertToDouble(), B);
1254  // For negative exponents simply compute the reciprocal.
1255  if (Op2C->isNegative())
1256  FMul = B.CreateFDiv(ConstantFP::get(CI->getType(), 1.0), FMul);
1257  return FMul;
1258  }
1259 
1260  return nullptr;
1261 }
1262 
1263 Value *LibCallSimplifier::optimizeExp2(CallInst *CI, IRBuilder<> &B) {
1265  Value *Ret = nullptr;
1266  StringRef Name = Callee->getName();
1267  if (UnsafeFPShrink && Name == "exp2" && hasFloatVersion(Name))
1268  Ret = optimizeUnaryDoubleFP(CI, B, true);
1269 
1270  Value *Op = CI->getArgOperand(0);
1271  // Turn exp2(sitofp(x)) -> ldexp(1.0, sext(x)) if sizeof(x) <= 32
1272  // Turn exp2(uitofp(x)) -> ldexp(1.0, zext(x)) if sizeof(x) < 32
1273  LibFunc LdExp = LibFunc_ldexpl;
1274  if (Op->getType()->isFloatTy())
1275  LdExp = LibFunc_ldexpf;
1276  else if (Op->getType()->isDoubleTy())
1277  LdExp = LibFunc_ldexp;
1278 
1279  if (TLI->has(LdExp)) {
1280  Value *LdExpArg = nullptr;
1281  if (SIToFPInst *OpC = dyn_cast<SIToFPInst>(Op)) {
1282  if (OpC->getOperand(0)->getType()->getPrimitiveSizeInBits() <= 32)
1283  LdExpArg = B.CreateSExt(OpC->getOperand(0), B.getInt32Ty());
1284  } else if (UIToFPInst *OpC = dyn_cast<UIToFPInst>(Op)) {
1285  if (OpC->getOperand(0)->getType()->getPrimitiveSizeInBits() < 32)
1286  LdExpArg = B.CreateZExt(OpC->getOperand(0), B.getInt32Ty());
1287  }
1288 
1289  if (LdExpArg) {
1290  Constant *One = ConstantFP::get(CI->getContext(), APFloat(1.0f));
1291  if (!Op->getType()->isFloatTy())
1292  One = ConstantExpr::getFPExtend(One, Op->getType());
1293 
1294  Module *M = CI->getModule();
1295  Value *NewCallee =
1296  M->getOrInsertFunction(TLI->getName(LdExp), Op->getType(),
1297  Op->getType(), B.getInt32Ty());
1298  CallInst *CI = B.CreateCall(NewCallee, {One, LdExpArg});
1299  if (const Function *F = dyn_cast<Function>(Callee->stripPointerCasts()))
1300  CI->setCallingConv(F->getCallingConv());
1301 
1302  return CI;
1303  }
1304  }
1305  return Ret;
1306 }
1307 
1308 Value *LibCallSimplifier::optimizeFMinFMax(CallInst *CI, IRBuilder<> &B) {
1310  // If we can shrink the call to a float function rather than a double
1311  // function, do that first.
1312  StringRef Name = Callee->getName();
1313  if ((Name == "fmin" || Name == "fmax") && hasFloatVersion(Name))
1314  if (Value *Ret = optimizeBinaryDoubleFP(CI, B))
1315  return Ret;
1316 
1318  FastMathFlags FMF;
1319  if (CI->isFast()) {
1320  // If the call is 'fast', then anything we create here will also be 'fast'.
1321  FMF.setFast();
1322  } else {
1323  // At a minimum, no-nans-fp-math must be true.
1324  if (!CI->hasNoNaNs())
1325  return nullptr;
1326  // No-signed-zeros is implied by the definitions of fmax/fmin themselves:
1327  // "Ideally, fmax would be sensitive to the sign of zero, for example
1328  // fmax(-0. 0, +0. 0) would return +0; however, implementation in software
1329  // might be impractical."
1330  FMF.setNoSignedZeros();
1331  FMF.setNoNaNs();
1332  }
1333  B.setFastMathFlags(FMF);
1334 
1335  // We have a relaxed floating-point environment. We can ignore NaN-handling
1336  // and transform to a compare and select. We do not have to consider errno or
1337  // exceptions, because fmin/fmax do not have those.
1338  Value *Op0 = CI->getArgOperand(0);
1339  Value *Op1 = CI->getArgOperand(1);
1340  Value *Cmp = Callee->getName().startswith("fmin") ?
1341  B.CreateFCmpOLT(Op0, Op1) : B.CreateFCmpOGT(Op0, Op1);
1342  return B.CreateSelect(Cmp, Op0, Op1);
1343 }
1344 
1345 Value *LibCallSimplifier::optimizeLog(CallInst *CI, IRBuilder<> &B) {
1347  Value *Ret = nullptr;
1348  StringRef Name = Callee->getName();
1349  if (UnsafeFPShrink && hasFloatVersion(Name))
1350  Ret = optimizeUnaryDoubleFP(CI, B, true);
1351 
1352  if (!CI->isFast())
1353  return Ret;
1354  Value *Op1 = CI->getArgOperand(0);
1355  auto *OpC = dyn_cast<CallInst>(Op1);
1356 
1357  // The earlier call must also be 'fast' in order to do these transforms.
1358  if (!OpC || !OpC->isFast())
1359  return Ret;
1360 
1361  // log(pow(x,y)) -> y*log(x)
1362  // This is only applicable to log, log2, log10.
1363  if (Name != "log" && Name != "log2" && Name != "log10")
1364  return Ret;
1365 
1367  FastMathFlags FMF;
1368  FMF.setFast();
1369  B.setFastMathFlags(FMF);
1370 
1371  LibFunc Func;
1372  Function *F = OpC->getCalledFunction();
1373  if (F && ((TLI->getLibFunc(F->getName(), Func) && TLI->has(Func) &&
1374  Func == LibFunc_pow) || F->getIntrinsicID() == Intrinsic::pow))
1375  return B.CreateFMul(OpC->getArgOperand(1),
1376  emitUnaryFloatFnCall(OpC->getOperand(0), Callee->getName(), B,
1377  Callee->getAttributes()), "mul");
1378 
1379  // log(exp2(y)) -> y*log(2)
1380  if (F && Name == "log" && TLI->getLibFunc(F->getName(), Func) &&
1381  TLI->has(Func) && Func == LibFunc_exp2)
1382  return B.CreateFMul(
1383  OpC->getArgOperand(0),
1385  Callee->getName(), B, Callee->getAttributes()),
1386  "logmul");
1387  return Ret;
1388 }
1389 
1390 Value *LibCallSimplifier::optimizeSqrt(CallInst *CI, IRBuilder<> &B) {
1392  Value *Ret = nullptr;
1393  // TODO: Once we have a way (other than checking for the existince of the
1394  // libcall) to tell whether our target can lower @llvm.sqrt, relax the
1395  // condition below.
1396  if (TLI->has(LibFunc_sqrtf) && (Callee->getName() == "sqrt" ||
1397  Callee->getIntrinsicID() == Intrinsic::sqrt))
1398  Ret = optimizeUnaryDoubleFP(CI, B, true);
1399 
1400  if (!CI->isFast())
1401  return Ret;
1402 
1404  if (!I || I->getOpcode() != Instruction::FMul || !I->isFast())
1405  return Ret;
1406 
1407  // We're looking for a repeated factor in a multiplication tree,
1408  // so we can do this fold: sqrt(x * x) -> fabs(x);
1409  // or this fold: sqrt((x * x) * y) -> fabs(x) * sqrt(y).
1410  Value *Op0 = I->getOperand(0);
1411  Value *Op1 = I->getOperand(1);
1412  Value *RepeatOp = nullptr;
1413  Value *OtherOp = nullptr;
1414  if (Op0 == Op1) {
1415  // Simple match: the operands of the multiply are identical.
1416  RepeatOp = Op0;
1417  } else {
1418  // Look for a more complicated pattern: one of the operands is itself
1419  // a multiply, so search for a common factor in that multiply.
1420  // Note: We don't bother looking any deeper than this first level or for
1421  // variations of this pattern because instcombine's visitFMUL and/or the
1422  // reassociation pass should give us this form.
1423  Value *OtherMul0, *OtherMul1;
1424  if (match(Op0, m_FMul(m_Value(OtherMul0), m_Value(OtherMul1)))) {
1425  // Pattern: sqrt((x * y) * z)
1426  if (OtherMul0 == OtherMul1 && cast<Instruction>(Op0)->isFast()) {
1427  // Matched: sqrt((x * x) * z)
1428  RepeatOp = OtherMul0;
1429  OtherOp = Op1;
1430  }
1431  }
1432  }
1433  if (!RepeatOp)
1434  return Ret;
1435 
1436  // Fast math flags for any created instructions should match the sqrt
1437  // and multiply.
1440 
1441  // If we found a repeated factor, hoist it out of the square root and
1442  // replace it with the fabs of that factor.
1443  Module *M = Callee->getParent();
1444  Type *ArgType = I->getType();
1445  Value *Fabs = Intrinsic::getDeclaration(M, Intrinsic::fabs, ArgType);
1446  Value *FabsCall = B.CreateCall(Fabs, RepeatOp, "fabs");
1447  if (OtherOp) {
1448  // If we found a non-repeated factor, we still need to get its square
1449  // root. We then multiply that by the value that was simplified out
1450  // of the square root calculation.
1451  Value *Sqrt = Intrinsic::getDeclaration(M, Intrinsic::sqrt, ArgType);
1452  Value *SqrtCall = B.CreateCall(Sqrt, OtherOp, "sqrt");
1453  return B.CreateFMul(FabsCall, SqrtCall);
1454  }
1455  return FabsCall;
1456 }
1457 
1458 // TODO: Generalize to handle any trig function and its inverse.
1459 Value *LibCallSimplifier::optimizeTan(CallInst *CI, IRBuilder<> &B) {
1461  Value *Ret = nullptr;
1462  StringRef Name = Callee->getName();
1463  if (UnsafeFPShrink && Name == "tan" && hasFloatVersion(Name))
1464  Ret = optimizeUnaryDoubleFP(CI, B, true);
1465 
1466  Value *Op1 = CI->getArgOperand(0);
1467  auto *OpC = dyn_cast<CallInst>(Op1);
1468  if (!OpC)
1469  return Ret;
1470 
1471  // Both calls must be 'fast' in order to remove them.
1472  if (!CI->isFast() || !OpC->isFast())
1473  return Ret;
1474 
1475  // tan(atan(x)) -> x
1476  // tanf(atanf(x)) -> x
1477  // tanl(atanl(x)) -> x
1478  LibFunc Func;
1479  Function *F = OpC->getCalledFunction();
1480  if (F && TLI->getLibFunc(F->getName(), Func) && TLI->has(Func) &&
1481  ((Func == LibFunc_atan && Callee->getName() == "tan") ||
1482  (Func == LibFunc_atanf && Callee->getName() == "tanf") ||
1483  (Func == LibFunc_atanl && Callee->getName() == "tanl")))
1484  Ret = OpC->getArgOperand(0);
1485  return Ret;
1486 }
1487 
1488 static bool isTrigLibCall(CallInst *CI) {
1489  // We can only hope to do anything useful if we can ignore things like errno
1490  // and floating-point exceptions.
1491  // We already checked the prototype.
1492  return CI->hasFnAttr(Attribute::NoUnwind) &&
1493  CI->hasFnAttr(Attribute::ReadNone);
1494 }
1495 
1496 static void insertSinCosCall(IRBuilder<> &B, Function *OrigCallee, Value *Arg,
1497  bool UseFloat, Value *&Sin, Value *&Cos,
1498  Value *&SinCos) {
1499  Type *ArgTy = Arg->getType();
1500  Type *ResTy;
1501  StringRef Name;
1502 
1503  Triple T(OrigCallee->getParent()->getTargetTriple());
1504  if (UseFloat) {
1505  Name = "__sincospif_stret";
1506 
1507  assert(T.getArch() != Triple::x86 && "x86 messy and unsupported for now");
1508  // x86_64 can't use {float, float} since that would be returned in both
1509  // xmm0 and xmm1, which isn't what a real struct would do.
1510  ResTy = T.getArch() == Triple::x86_64
1511  ? static_cast<Type *>(VectorType::get(ArgTy, 2))
1512  : static_cast<Type *>(StructType::get(ArgTy, ArgTy));
1513  } else {
1514  Name = "__sincospi_stret";
1515  ResTy = StructType::get(ArgTy, ArgTy);
1516  }
1517 
1518  Module *M = OrigCallee->getParent();
1519  Value *Callee = M->getOrInsertFunction(Name, OrigCallee->getAttributes(),
1520  ResTy, ArgTy);
1521 
1522  if (Instruction *ArgInst = dyn_cast<Instruction>(Arg)) {
1523  // If the argument is an instruction, it must dominate all uses so put our
1524  // sincos call there.
1525  B.SetInsertPoint(ArgInst->getParent(), ++ArgInst->getIterator());
1526  } else {
1527  // Otherwise (e.g. for a constant) the beginning of the function is as
1528  // good a place as any.
1529  BasicBlock &EntryBB = B.GetInsertBlock()->getParent()->getEntryBlock();
1530  B.SetInsertPoint(&EntryBB, EntryBB.begin());
1531  }
1532 
1533  SinCos = B.CreateCall(Callee, Arg, "sincospi");
1534 
1535  if (SinCos->getType()->isStructTy()) {
1536  Sin = B.CreateExtractValue(SinCos, 0, "sinpi");
1537  Cos = B.CreateExtractValue(SinCos, 1, "cospi");
1538  } else {
1539  Sin = B.CreateExtractElement(SinCos, ConstantInt::get(B.getInt32Ty(), 0),
1540  "sinpi");
1541  Cos = B.CreateExtractElement(SinCos, ConstantInt::get(B.getInt32Ty(), 1),
1542  "cospi");
1543  }
1544 }
1545 
1546 Value *LibCallSimplifier::optimizeSinCosPi(CallInst *CI, IRBuilder<> &B) {
1547  // Make sure the prototype is as expected, otherwise the rest of the
1548  // function is probably invalid and likely to abort.
1549  if (!isTrigLibCall(CI))
1550  return nullptr;
1551 
1552  Value *Arg = CI->getArgOperand(0);
1553  SmallVector<CallInst *, 1> SinCalls;
1554  SmallVector<CallInst *, 1> CosCalls;
1555  SmallVector<CallInst *, 1> SinCosCalls;
1556 
1557  bool IsFloat = Arg->getType()->isFloatTy();
1558 
1559  // Look for all compatible sinpi, cospi and sincospi calls with the same
1560  // argument. If there are enough (in some sense) we can make the
1561  // substitution.
1562  Function *F = CI->getFunction();
1563  for (User *U : Arg->users())
1564  classifyArgUse(U, F, IsFloat, SinCalls, CosCalls, SinCosCalls);
1565 
1566  // It's only worthwhile if both sinpi and cospi are actually used.
1567  if (SinCosCalls.empty() && (SinCalls.empty() || CosCalls.empty()))
1568  return nullptr;
1569 
1570  Value *Sin, *Cos, *SinCos;
1571  insertSinCosCall(B, CI->getCalledFunction(), Arg, IsFloat, Sin, Cos, SinCos);
1572 
1573  auto replaceTrigInsts = [this](SmallVectorImpl<CallInst *> &Calls,
1574  Value *Res) {
1575  for (CallInst *C : Calls)
1576  replaceAllUsesWith(C, Res);
1577  };
1578 
1579  replaceTrigInsts(SinCalls, Sin);
1580  replaceTrigInsts(CosCalls, Cos);
1581  replaceTrigInsts(SinCosCalls, SinCos);
1582 
1583  return nullptr;
1584 }
1585 
1586 void LibCallSimplifier::classifyArgUse(
1587  Value *Val, Function *F, bool IsFloat,
1588  SmallVectorImpl<CallInst *> &SinCalls,
1589  SmallVectorImpl<CallInst *> &CosCalls,
1590  SmallVectorImpl<CallInst *> &SinCosCalls) {
1591  CallInst *CI = dyn_cast<CallInst>(Val);
1592 
1593  if (!CI)
1594  return;
1595 
1596  // Don't consider calls in other functions.
1597  if (CI->getFunction() != F)
1598  return;
1599 
1601  LibFunc Func;
1602  if (!Callee || !TLI->getLibFunc(*Callee, Func) || !TLI->has(Func) ||
1603  !isTrigLibCall(CI))
1604  return;
1605 
1606  if (IsFloat) {
1607  if (Func == LibFunc_sinpif)
1608  SinCalls.push_back(CI);
1609  else if (Func == LibFunc_cospif)
1610  CosCalls.push_back(CI);
1611  else if (Func == LibFunc_sincospif_stret)
1612  SinCosCalls.push_back(CI);
1613  } else {
1614  if (Func == LibFunc_sinpi)
1615  SinCalls.push_back(CI);
1616  else if (Func == LibFunc_cospi)
1617  CosCalls.push_back(CI);
1618  else if (Func == LibFunc_sincospi_stret)
1619  SinCosCalls.push_back(CI);
1620  }
1621 }
1622 
1623 //===----------------------------------------------------------------------===//
1624 // Integer Library Call Optimizations
1625 //===----------------------------------------------------------------------===//
1626 
1627 Value *LibCallSimplifier::optimizeFFS(CallInst *CI, IRBuilder<> &B) {
1628  // ffs(x) -> x != 0 ? (i32)llvm.cttz(x)+1 : 0
1629  Value *Op = CI->getArgOperand(0);
1630  Type *ArgType = Op->getType();
1632  Intrinsic::cttz, ArgType);
1633  Value *V = B.CreateCall(F, {Op, B.getTrue()}, "cttz");
1634  V = B.CreateAdd(V, ConstantInt::get(V->getType(), 1));
1635  V = B.CreateIntCast(V, B.getInt32Ty(), false);
1636 
1637  Value *Cond = B.CreateICmpNE(Op, Constant::getNullValue(ArgType));
1638  return B.CreateSelect(Cond, V, B.getInt32(0));
1639 }
1640 
1641 Value *LibCallSimplifier::optimizeFls(CallInst *CI, IRBuilder<> &B) {
1642  // fls(x) -> (i32)(sizeInBits(x) - llvm.ctlz(x, false))
1643  Value *Op = CI->getArgOperand(0);
1644  Type *ArgType = Op->getType();
1646  Intrinsic::ctlz, ArgType);
1647  Value *V = B.CreateCall(F, {Op, B.getFalse()}, "ctlz");
1648  V = B.CreateSub(ConstantInt::get(V->getType(), ArgType->getIntegerBitWidth()),
1649  V);
1650  return B.CreateIntCast(V, CI->getType(), false);
1651 }
1652 
1653 Value *LibCallSimplifier::optimizeAbs(CallInst *CI, IRBuilder<> &B) {
1654  // abs(x) -> x >s -1 ? x : -x
1655  Value *Op = CI->getArgOperand(0);
1656  Value *Pos =
1657  B.CreateICmpSGT(Op, Constant::getAllOnesValue(Op->getType()), "ispos");
1658  Value *Neg = B.CreateNeg(Op, "neg");
1659  return B.CreateSelect(Pos, Op, Neg);
1660 }
1661 
1662 Value *LibCallSimplifier::optimizeIsDigit(CallInst *CI, IRBuilder<> &B) {
1663  // isdigit(c) -> (c-'0') <u 10
1664  Value *Op = CI->getArgOperand(0);
1665  Op = B.CreateSub(Op, B.getInt32('0'), "isdigittmp");
1666  Op = B.CreateICmpULT(Op, B.getInt32(10), "isdigit");
1667  return B.CreateZExt(Op, CI->getType());
1668 }
1669 
1670 Value *LibCallSimplifier::optimizeIsAscii(CallInst *CI, IRBuilder<> &B) {
1671  // isascii(c) -> c <u 128
1672  Value *Op = CI->getArgOperand(0);
1673  Op = B.CreateICmpULT(Op, B.getInt32(128), "isascii");
1674  return B.CreateZExt(Op, CI->getType());
1675 }
1676 
1677 Value *LibCallSimplifier::optimizeToAscii(CallInst *CI, IRBuilder<> &B) {
1678  // toascii(c) -> c & 0x7f
1679  return B.CreateAnd(CI->getArgOperand(0),
1680  ConstantInt::get(CI->getType(), 0x7F));
1681 }
1682 
1683 //===----------------------------------------------------------------------===//
1684 // Formatting and IO Library Call Optimizations
1685 //===----------------------------------------------------------------------===//
1686 
1687 static bool isReportingError(Function *Callee, CallInst *CI, int StreamArg);
1688 
1689 Value *LibCallSimplifier::optimizeErrorReporting(CallInst *CI, IRBuilder<> &B,
1690  int StreamArg) {
1691  Function *Callee = CI->getCalledFunction();
1692  // Error reporting calls should be cold, mark them as such.
1693  // This applies even to non-builtin calls: it is only a hint and applies to
1694  // functions that the frontend might not understand as builtins.
1695 
1696  // This heuristic was suggested in:
1697  // Improving Static Branch Prediction in a Compiler
1698  // Brian L. Deitrich, Ben-Chung Cheng, Wen-mei W. Hwu
1699  // Proceedings of PACT'98, Oct. 1998, IEEE
1700  if (!CI->hasFnAttr(Attribute::Cold) &&
1701  isReportingError(Callee, CI, StreamArg)) {
1703  }
1704 
1705  return nullptr;
1706 }
1707 
1708 static bool isReportingError(Function *Callee, CallInst *CI, int StreamArg) {
1709  if (!Callee || !Callee->isDeclaration())
1710  return false;
1711 
1712  if (StreamArg < 0)
1713  return true;
1714 
1715  // These functions might be considered cold, but only if their stream
1716  // argument is stderr.
1717 
1718  if (StreamArg >= (int)CI->getNumArgOperands())
1719  return false;
1720  LoadInst *LI = dyn_cast<LoadInst>(CI->getArgOperand(StreamArg));
1721  if (!LI)
1722  return false;
1724  if (!GV || !GV->isDeclaration())
1725  return false;
1726  return GV->getName() == "stderr";
1727 }
1728 
1729 Value *LibCallSimplifier::optimizePrintFString(CallInst *CI, IRBuilder<> &B) {
1730  // Check for a fixed format string.
1731  StringRef FormatStr;
1732  if (!getConstantStringInfo(CI->getArgOperand(0), FormatStr))
1733  return nullptr;
1734 
1735  // Empty format string -> noop.
1736  if (FormatStr.empty()) // Tolerate printf's declared void.
1737  return CI->use_empty() ? (Value *)CI : ConstantInt::get(CI->getType(), 0);
1738 
1739  // Do not do any of the following transformations if the printf return value
1740  // is used, in general the printf return value is not compatible with either
1741  // putchar() or puts().
1742  if (!CI->use_empty())
1743  return nullptr;
1744 
1745  // printf("x") -> putchar('x'), even for "%" and "%%".
1746  if (FormatStr.size() == 1 || FormatStr == "%%")
1747  return emitPutChar(B.getInt32(FormatStr[0]), B, TLI);
1748 
1749  // printf("%s", "a") --> putchar('a')
1750  if (FormatStr == "%s" && CI->getNumArgOperands() > 1) {
1751  StringRef ChrStr;
1752  if (!getConstantStringInfo(CI->getOperand(1), ChrStr))
1753  return nullptr;
1754  if (ChrStr.size() != 1)
1755  return nullptr;
1756  return emitPutChar(B.getInt32(ChrStr[0]), B, TLI);
1757  }
1758 
1759  // printf("foo\n") --> puts("foo")
1760  if (FormatStr[FormatStr.size() - 1] == '\n' &&
1761  FormatStr.find('%') == StringRef::npos) { // No format characters.
1762  // Create a string literal with no \n on it. We expect the constant merge
1763  // pass to be run after this pass, to merge duplicate strings.
1764  FormatStr = FormatStr.drop_back();
1765  Value *GV = B.CreateGlobalString(FormatStr, "str");
1766  return emitPutS(GV, B, TLI);
1767  }
1768 
1769  // Optimize specific format strings.
1770  // printf("%c", chr) --> putchar(chr)
1771  if (FormatStr == "%c" && CI->getNumArgOperands() > 1 &&
1772  CI->getArgOperand(1)->getType()->isIntegerTy())
1773  return emitPutChar(CI->getArgOperand(1), B, TLI);
1774 
1775  // printf("%s\n", str) --> puts(str)
1776  if (FormatStr == "%s\n" && CI->getNumArgOperands() > 1 &&
1777  CI->getArgOperand(1)->getType()->isPointerTy())
1778  return emitPutS(CI->getArgOperand(1), B, TLI);
1779  return nullptr;
1780 }
1781 
1782 Value *LibCallSimplifier::optimizePrintF(CallInst *CI, IRBuilder<> &B) {
1783 
1784  Function *Callee = CI->getCalledFunction();
1785  FunctionType *FT = Callee->getFunctionType();
1786  if (Value *V = optimizePrintFString(CI, B)) {
1787  return V;
1788  }
1789 
1790  // printf(format, ...) -> iprintf(format, ...) if no floating point
1791  // arguments.
1792  if (TLI->has(LibFunc_iprintf) && !callHasFloatingPointArgument(CI)) {
1793  Module *M = B.GetInsertBlock()->getParent()->getParent();
1794  Constant *IPrintFFn =
1795  M->getOrInsertFunction("iprintf", FT, Callee->getAttributes());
1796  CallInst *New = cast<CallInst>(CI->clone());
1797  New->setCalledFunction(IPrintFFn);
1798  B.Insert(New);
1799  return New;
1800  }
1801  return nullptr;
1802 }
1803 
1804 Value *LibCallSimplifier::optimizeSPrintFString(CallInst *CI, IRBuilder<> &B) {
1805  // Check for a fixed format string.
1806  StringRef FormatStr;
1807  if (!getConstantStringInfo(CI->getArgOperand(1), FormatStr))
1808  return nullptr;
1809 
1810  // If we just have a format string (nothing else crazy) transform it.
1811  if (CI->getNumArgOperands() == 2) {
1812  // Make sure there's no % in the constant array. We could try to handle
1813  // %% -> % in the future if we cared.
1814  for (unsigned i = 0, e = FormatStr.size(); i != e; ++i)
1815  if (FormatStr[i] == '%')
1816  return nullptr; // we found a format specifier, bail out.
1817 
1818  // sprintf(str, fmt) -> llvm.memcpy(str, fmt, strlen(fmt)+1, 1)
1819  B.CreateMemCpy(CI->getArgOperand(0), CI->getArgOperand(1),
1820  ConstantInt::get(DL.getIntPtrType(CI->getContext()),
1821  FormatStr.size() + 1),
1822  1); // Copy the null byte.
1823  return ConstantInt::get(CI->getType(), FormatStr.size());
1824  }
1825 
1826  // The remaining optimizations require the format string to be "%s" or "%c"
1827  // and have an extra operand.
1828  if (FormatStr.size() != 2 || FormatStr[0] != '%' ||
1829  CI->getNumArgOperands() < 3)
1830  return nullptr;
1831 
1832  // Decode the second character of the format string.
1833  if (FormatStr[1] == 'c') {
1834  // sprintf(dst, "%c", chr) --> *(i8*)dst = chr; *((i8*)dst+1) = 0
1835  if (!CI->getArgOperand(2)->getType()->isIntegerTy())
1836  return nullptr;
1837  Value *V = B.CreateTrunc(CI->getArgOperand(2), B.getInt8Ty(), "char");
1838  Value *Ptr = castToCStr(CI->getArgOperand(0), B);
1839  B.CreateStore(V, Ptr);
1840  Ptr = B.CreateGEP(B.getInt8Ty(), Ptr, B.getInt32(1), "nul");
1841  B.CreateStore(B.getInt8(0), Ptr);
1842 
1843  return ConstantInt::get(CI->getType(), 1);
1844  }
1845 
1846  if (FormatStr[1] == 's') {
1847  // sprintf(dest, "%s", str) -> llvm.memcpy(dest, str, strlen(str)+1, 1)
1848  if (!CI->getArgOperand(2)->getType()->isPointerTy())
1849  return nullptr;
1850 
1851  Value *Len = emitStrLen(CI->getArgOperand(2), B, DL, TLI);
1852  if (!Len)
1853  return nullptr;
1854  Value *IncLen =
1855  B.CreateAdd(Len, ConstantInt::get(Len->getType(), 1), "leninc");
1856  B.CreateMemCpy(CI->getArgOperand(0), CI->getArgOperand(2), IncLen, 1);
1857 
1858  // The sprintf result is the unincremented number of bytes in the string.
1859  return B.CreateIntCast(Len, CI->getType(), false);
1860  }
1861  return nullptr;
1862 }
1863 
1864 Value *LibCallSimplifier::optimizeSPrintF(CallInst *CI, IRBuilder<> &B) {
1865  Function *Callee = CI->getCalledFunction();
1866  FunctionType *FT = Callee->getFunctionType();
1867  if (Value *V = optimizeSPrintFString(CI, B)) {
1868  return V;
1869  }
1870 
1871  // sprintf(str, format, ...) -> siprintf(str, format, ...) if no floating
1872  // point arguments.
1873  if (TLI->has(LibFunc_siprintf) && !callHasFloatingPointArgument(CI)) {
1874  Module *M = B.GetInsertBlock()->getParent()->getParent();
1875  Constant *SIPrintFFn =
1876  M->getOrInsertFunction("siprintf", FT, Callee->getAttributes());
1877  CallInst *New = cast<CallInst>(CI->clone());
1878  New->setCalledFunction(SIPrintFFn);
1879  B.Insert(New);
1880  return New;
1881  }
1882  return nullptr;
1883 }
1884 
1885 Value *LibCallSimplifier::optimizeFPrintFString(CallInst *CI, IRBuilder<> &B) {
1886  optimizeErrorReporting(CI, B, 0);
1887 
1888  // All the optimizations depend on the format string.
1889  StringRef FormatStr;
1890  if (!getConstantStringInfo(CI->getArgOperand(1), FormatStr))
1891  return nullptr;
1892 
1893  // Do not do any of the following transformations if the fprintf return
1894  // value is used, in general the fprintf return value is not compatible
1895  // with fwrite(), fputc() or fputs().
1896  if (!CI->use_empty())
1897  return nullptr;
1898 
1899  // fprintf(F, "foo") --> fwrite("foo", 3, 1, F)
1900  if (CI->getNumArgOperands() == 2) {
1901  for (unsigned i = 0, e = FormatStr.size(); i != e; ++i)
1902  if (FormatStr[i] == '%') // Could handle %% -> % if we cared.
1903  return nullptr; // We found a format specifier.
1904 
1905  return emitFWrite(
1906  CI->getArgOperand(1),
1907  ConstantInt::get(DL.getIntPtrType(CI->getContext()), FormatStr.size()),
1908  CI->getArgOperand(0), B, DL, TLI);
1909  }
1910 
1911  // The remaining optimizations require the format string to be "%s" or "%c"
1912  // and have an extra operand.
1913  if (FormatStr.size() != 2 || FormatStr[0] != '%' ||
1914  CI->getNumArgOperands() < 3)
1915  return nullptr;
1916 
1917  // Decode the second character of the format string.
1918  if (FormatStr[1] == 'c') {
1919  // fprintf(F, "%c", chr) --> fputc(chr, F)
1920  if (!CI->getArgOperand(2)->getType()->isIntegerTy())
1921  return nullptr;
1922  return emitFPutC(CI->getArgOperand(2), CI->getArgOperand(0), B, TLI);
1923  }
1924 
1925  if (FormatStr[1] == 's') {
1926  // fprintf(F, "%s", str) --> fputs(str, F)
1927  if (!CI->getArgOperand(2)->getType()->isPointerTy())
1928  return nullptr;
1929  return emitFPutS(CI->getArgOperand(2), CI->getArgOperand(0), B, TLI);
1930  }
1931  return nullptr;
1932 }
1933 
1934 Value *LibCallSimplifier::optimizeFPrintF(CallInst *CI, IRBuilder<> &B) {
1935  Function *Callee = CI->getCalledFunction();
1936  FunctionType *FT = Callee->getFunctionType();
1937  if (Value *V = optimizeFPrintFString(CI, B)) {
1938  return V;
1939  }
1940 
1941  // fprintf(stream, format, ...) -> fiprintf(stream, format, ...) if no
1942  // floating point arguments.
1943  if (TLI->has(LibFunc_fiprintf) && !callHasFloatingPointArgument(CI)) {
1944  Module *M = B.GetInsertBlock()->getParent()->getParent();
1945  Constant *FIPrintFFn =
1946  M->getOrInsertFunction("fiprintf", FT, Callee->getAttributes());
1947  CallInst *New = cast<CallInst>(CI->clone());
1948  New->setCalledFunction(FIPrintFFn);
1949  B.Insert(New);
1950  return New;
1951  }
1952  return nullptr;
1953 }
1954 
1955 Value *LibCallSimplifier::optimizeFWrite(CallInst *CI, IRBuilder<> &B) {
1956  optimizeErrorReporting(CI, B, 3);
1957 
1958  // Get the element size and count.
1959  ConstantInt *SizeC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
1960  ConstantInt *CountC = dyn_cast<ConstantInt>(CI->getArgOperand(2));
1961  if (!SizeC || !CountC)
1962  return nullptr;
1963  uint64_t Bytes = SizeC->getZExtValue() * CountC->getZExtValue();
1964 
1965  // If this is writing zero records, remove the call (it's a noop).
1966  if (Bytes == 0)
1967  return ConstantInt::get(CI->getType(), 0);
1968 
1969  // If this is writing one byte, turn it into fputc.
1970  // This optimisation is only valid, if the return value is unused.
1971  if (Bytes == 1 && CI->use_empty()) { // fwrite(S,1,1,F) -> fputc(S[0],F)
1972  Value *Char = B.CreateLoad(castToCStr(CI->getArgOperand(0), B), "char");
1973  Value *NewCI = emitFPutC(Char, CI->getArgOperand(3), B, TLI);
1974  return NewCI ? ConstantInt::get(CI->getType(), 1) : nullptr;
1975  }
1976 
1977  return nullptr;
1978 }
1979 
1980 Value *LibCallSimplifier::optimizeFPuts(CallInst *CI, IRBuilder<> &B) {
1981  optimizeErrorReporting(CI, B, 1);
1982 
1983  // Don't rewrite fputs to fwrite when optimising for size because fwrite
1984  // requires more arguments and thus extra MOVs are required.
1985  if (CI->getParent()->getParent()->optForSize())
1986  return nullptr;
1987 
1988  // We can't optimize if return value is used.
1989  if (!CI->use_empty())
1990  return nullptr;
1991 
1992  // fputs(s,F) --> fwrite(s,1,strlen(s),F)
1993  uint64_t Len = GetStringLength(CI->getArgOperand(0));
1994  if (!Len)
1995  return nullptr;
1996 
1997  // Known to have no uses (see above).
1998  return emitFWrite(
1999  CI->getArgOperand(0),
2000  ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len - 1),
2001  CI->getArgOperand(1), B, DL, TLI);
2002 }
2003 
2004 Value *LibCallSimplifier::optimizePuts(CallInst *CI, IRBuilder<> &B) {
2005  // Check for a constant string.
2006  StringRef Str;
2007  if (!getConstantStringInfo(CI->getArgOperand(0), Str))
2008  return nullptr;
2009 
2010  if (Str.empty() && CI->use_empty()) {
2011  // puts("") -> putchar('\n')
2012  Value *Res = emitPutChar(B.getInt32('\n'), B, TLI);
2013  if (CI->use_empty() || !Res)
2014  return Res;
2015  return B.CreateIntCast(Res, CI->getType(), true);
2016  }
2017 
2018  return nullptr;
2019 }
2020 
2021 bool LibCallSimplifier::hasFloatVersion(StringRef FuncName) {
2022  LibFunc Func;
2023  SmallString<20> FloatFuncName = FuncName;
2024  FloatFuncName += 'f';
2025  if (TLI->getLibFunc(FloatFuncName, Func))
2026  return TLI->has(Func);
2027  return false;
2028 }
2029 
2030 Value *LibCallSimplifier::optimizeStringMemoryLibCall(CallInst *CI,
2031  IRBuilder<> &Builder) {
2032  LibFunc Func;
2033  Function *Callee = CI->getCalledFunction();
2034  // Check for string/memory library functions.
2035  if (TLI->getLibFunc(*Callee, Func) && TLI->has(Func)) {
2036  // Make sure we never change the calling convention.
2037  assert((ignoreCallingConv(Func) ||
2038  isCallingConvCCompatible(CI)) &&
2039  "Optimizing string/memory libcall would change the calling convention");
2040  switch (Func) {
2041  case LibFunc_strcat:
2042  return optimizeStrCat(CI, Builder);
2043  case LibFunc_strncat:
2044  return optimizeStrNCat(CI, Builder);
2045  case LibFunc_strchr:
2046  return optimizeStrChr(CI, Builder);
2047  case LibFunc_strrchr:
2048  return optimizeStrRChr(CI, Builder);
2049  case LibFunc_strcmp:
2050  return optimizeStrCmp(CI, Builder);
2051  case LibFunc_strncmp:
2052  return optimizeStrNCmp(CI, Builder);
2053  case LibFunc_strcpy:
2054  return optimizeStrCpy(CI, Builder);
2055  case LibFunc_stpcpy:
2056  return optimizeStpCpy(CI, Builder);
2057  case LibFunc_strncpy:
2058  return optimizeStrNCpy(CI, Builder);
2059  case LibFunc_strlen:
2060  return optimizeStrLen(CI, Builder);
2061  case LibFunc_strpbrk:
2062  return optimizeStrPBrk(CI, Builder);
2063  case LibFunc_strtol:
2064  case LibFunc_strtod:
2065  case LibFunc_strtof:
2066  case LibFunc_strtoul:
2067  case LibFunc_strtoll:
2068  case LibFunc_strtold:
2069  case LibFunc_strtoull:
2070  return optimizeStrTo(CI, Builder);
2071  case LibFunc_strspn:
2072  return optimizeStrSpn(CI, Builder);
2073  case LibFunc_strcspn:
2074  return optimizeStrCSpn(CI, Builder);
2075  case LibFunc_strstr:
2076  return optimizeStrStr(CI, Builder);
2077  case LibFunc_memchr:
2078  return optimizeMemChr(CI, Builder);
2079  case LibFunc_memcmp:
2080  return optimizeMemCmp(CI, Builder);
2081  case LibFunc_memcpy:
2082  return optimizeMemCpy(CI, Builder);
2083  case LibFunc_memmove:
2084  return optimizeMemMove(CI, Builder);
2085  case LibFunc_memset:
2086  return optimizeMemSet(CI, Builder);
2087  case LibFunc_wcslen:
2088  return optimizeWcslen(CI, Builder);
2089  default:
2090  break;
2091  }
2092  }
2093  return nullptr;
2094 }
2095 
2096 Value *LibCallSimplifier::optimizeFloatingPointLibCall(CallInst *CI,
2097  LibFunc Func,
2098  IRBuilder<> &Builder) {
2099  // Don't optimize calls that require strict floating point semantics.
2100  if (CI->isStrictFP())
2101  return nullptr;
2102 
2103  switch (Func) {
2104  case LibFunc_cosf:
2105  case LibFunc_cos:
2106  case LibFunc_cosl:
2107  return optimizeCos(CI, Builder);
2108  case LibFunc_sinpif:
2109  case LibFunc_sinpi:
2110  case LibFunc_cospif:
2111  case LibFunc_cospi:
2112  return optimizeSinCosPi(CI, Builder);
2113  case LibFunc_powf:
2114  case LibFunc_pow:
2115  case LibFunc_powl:
2116  return optimizePow(CI, Builder);
2117  case LibFunc_exp2l:
2118  case LibFunc_exp2:
2119  case LibFunc_exp2f:
2120  return optimizeExp2(CI, Builder);
2121  case LibFunc_fabsf:
2122  case LibFunc_fabs:
2123  case LibFunc_fabsl:
2124  return replaceUnaryCall(CI, Builder, Intrinsic::fabs);
2125  case LibFunc_sqrtf:
2126  case LibFunc_sqrt:
2127  case LibFunc_sqrtl:
2128  return optimizeSqrt(CI, Builder);
2129  case LibFunc_log:
2130  case LibFunc_log10:
2131  case LibFunc_log1p:
2132  case LibFunc_log2:
2133  case LibFunc_logb:
2134  return optimizeLog(CI, Builder);
2135  case LibFunc_tan:
2136  case LibFunc_tanf:
2137  case LibFunc_tanl:
2138  return optimizeTan(CI, Builder);
2139  case LibFunc_ceil:
2140  return replaceUnaryCall(CI, Builder, Intrinsic::ceil);
2141  case LibFunc_floor:
2142  return replaceUnaryCall(CI, Builder, Intrinsic::floor);
2143  case LibFunc_round:
2144  return replaceUnaryCall(CI, Builder, Intrinsic::round);
2145  case LibFunc_nearbyint:
2146  return replaceUnaryCall(CI, Builder, Intrinsic::nearbyint);
2147  case LibFunc_rint:
2148  return replaceUnaryCall(CI, Builder, Intrinsic::rint);
2149  case LibFunc_trunc:
2150  return replaceUnaryCall(CI, Builder, Intrinsic::trunc);
2151  case LibFunc_acos:
2152  case LibFunc_acosh:
2153  case LibFunc_asin:
2154  case LibFunc_asinh:
2155  case LibFunc_atan:
2156  case LibFunc_atanh:
2157  case LibFunc_cbrt:
2158  case LibFunc_cosh:
2159  case LibFunc_exp:
2160  case LibFunc_exp10:
2161  case LibFunc_expm1:
2162  case LibFunc_sin:
2163  case LibFunc_sinh:
2164  case LibFunc_tanh:
2165  if (UnsafeFPShrink && hasFloatVersion(CI->getCalledFunction()->getName()))
2166  return optimizeUnaryDoubleFP(CI, Builder, true);
2167  return nullptr;
2168  case LibFunc_copysign:
2169  if (hasFloatVersion(CI->getCalledFunction()->getName()))
2170  return optimizeBinaryDoubleFP(CI, Builder);
2171  return nullptr;
2172  case LibFunc_fminf:
2173  case LibFunc_fmin:
2174  case LibFunc_fminl:
2175  case LibFunc_fmaxf:
2176  case LibFunc_fmax:
2177  case LibFunc_fmaxl:
2178  return optimizeFMinFMax(CI, Builder);
2179  case LibFunc_cabs:
2180  case LibFunc_cabsf:
2181  case LibFunc_cabsl:
2182  return optimizeCAbs(CI, Builder);
2183  default:
2184  return nullptr;
2185  }
2186 }
2187 
2189  // TODO: Split out the code below that operates on FP calls so that
2190  // we can all non-FP calls with the StrictFP attribute to be
2191  // optimized.
2192  if (CI->isNoBuiltin())
2193  return nullptr;
2194 
2195  LibFunc Func;
2196  Function *Callee = CI->getCalledFunction();
2197 
2199  CI->getOperandBundlesAsDefs(OpBundles);
2200  IRBuilder<> Builder(CI, /*FPMathTag=*/nullptr, OpBundles);
2201  bool isCallingConvC = isCallingConvCCompatible(CI);
2202 
2203  // Command-line parameter overrides instruction attribute.
2204  // This can't be moved to optimizeFloatingPointLibCall() because it may be
2205  // used by the intrinsic optimizations.
2206  if (EnableUnsafeFPShrink.getNumOccurrences() > 0)
2207  UnsafeFPShrink = EnableUnsafeFPShrink;
2208  else if (isa<FPMathOperator>(CI) && CI->isFast())
2209  UnsafeFPShrink = true;
2210 
2211  // First, check for intrinsics.
2212  if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
2213  if (!isCallingConvC)
2214  return nullptr;
2215  // The FP intrinsics have corresponding constrained versions so we don't
2216  // need to check for the StrictFP attribute here.
2217  switch (II->getIntrinsicID()) {
2218  case Intrinsic::pow:
2219  return optimizePow(CI, Builder);
2220  case Intrinsic::exp2:
2221  return optimizeExp2(CI, Builder);
2222  case Intrinsic::log:
2223  return optimizeLog(CI, Builder);
2224  case Intrinsic::sqrt:
2225  return optimizeSqrt(CI, Builder);
2226  // TODO: Use foldMallocMemset() with memset intrinsic.
2227  default:
2228  return nullptr;
2229  }
2230  }
2231 
2232  // Also try to simplify calls to fortified library functions.
2233  if (Value *SimplifiedFortifiedCI = FortifiedSimplifier.optimizeCall(CI)) {
2234  // Try to further simplify the result.
2235  CallInst *SimplifiedCI = dyn_cast<CallInst>(SimplifiedFortifiedCI);
2236  if (SimplifiedCI && SimplifiedCI->getCalledFunction()) {
2237  // Use an IR Builder from SimplifiedCI if available instead of CI
2238  // to guarantee we reach all uses we might replace later on.
2239  IRBuilder<> TmpBuilder(SimplifiedCI);
2240  if (Value *V = optimizeStringMemoryLibCall(SimplifiedCI, TmpBuilder)) {
2241  // If we were able to further simplify, remove the now redundant call.
2242  SimplifiedCI->replaceAllUsesWith(V);
2243  SimplifiedCI->eraseFromParent();
2244  return V;
2245  }
2246  }
2247  return SimplifiedFortifiedCI;
2248  }
2249 
2250  // Then check for known library functions.
2251  if (TLI->getLibFunc(*Callee, Func) && TLI->has(Func)) {
2252  // We never change the calling convention.
2253  if (!ignoreCallingConv(Func) && !isCallingConvC)
2254  return nullptr;
2255  if (Value *V = optimizeStringMemoryLibCall(CI, Builder))
2256  return V;
2257  if (Value *V = optimizeFloatingPointLibCall(CI, Func, Builder))
2258  return V;
2259  switch (Func) {
2260  case LibFunc_ffs:
2261  case LibFunc_ffsl:
2262  case LibFunc_ffsll:
2263  return optimizeFFS(CI, Builder);
2264  case LibFunc_fls:
2265  case LibFunc_flsl:
2266  case LibFunc_flsll:
2267  return optimizeFls(CI, Builder);
2268  case LibFunc_abs:
2269  case LibFunc_labs:
2270  case LibFunc_llabs:
2271  return optimizeAbs(CI, Builder);
2272  case LibFunc_isdigit:
2273  return optimizeIsDigit(CI, Builder);
2274  case LibFunc_isascii:
2275  return optimizeIsAscii(CI, Builder);
2276  case LibFunc_toascii:
2277  return optimizeToAscii(CI, Builder);
2278  case LibFunc_printf:
2279  return optimizePrintF(CI, Builder);
2280  case LibFunc_sprintf:
2281  return optimizeSPrintF(CI, Builder);
2282  case LibFunc_fprintf:
2283  return optimizeFPrintF(CI, Builder);
2284  case LibFunc_fwrite:
2285  return optimizeFWrite(CI, Builder);
2286  case LibFunc_fputs:
2287  return optimizeFPuts(CI, Builder);
2288  case LibFunc_puts:
2289  return optimizePuts(CI, Builder);
2290  case LibFunc_perror:
2291  return optimizeErrorReporting(CI, Builder);
2292  case LibFunc_vfprintf:
2293  case LibFunc_fiprintf:
2294  return optimizeErrorReporting(CI, Builder, 0);
2295  case LibFunc_fputc:
2296  return optimizeErrorReporting(CI, Builder, 1);
2297  default:
2298  return nullptr;
2299  }
2300  }
2301  return nullptr;
2302 }
2303 
2305  const DataLayout &DL, const TargetLibraryInfo *TLI,
2307  function_ref<void(Instruction *, Value *)> Replacer)
2308  : FortifiedSimplifier(TLI), DL(DL), TLI(TLI), ORE(ORE),
2309  UnsafeFPShrink(false), Replacer(Replacer) {}
2310 
2311 void LibCallSimplifier::replaceAllUsesWith(Instruction *I, Value *With) {
2312  // Indirect through the replacer used in this instance.
2313  Replacer(I, With);
2314 }
2315 
2316 // TODO:
2317 // Additional cases that we need to add to this file:
2318 //
2319 // cbrt:
2320 // * cbrt(expN(X)) -> expN(x/3)
2321 // * cbrt(sqrt(x)) -> pow(x,1/6)
2322 // * cbrt(cbrt(x)) -> pow(x,1/9)
2323 //
2324 // exp, expf, expl:
2325 // * exp(log(x)) -> x
2326 //
2327 // log, logf, logl:
2328 // * log(exp(x)) -> x
2329 // * log(exp(y)) -> y*log(e)
2330 // * log(exp10(y)) -> y*log(10)
2331 // * log(sqrt(x)) -> 0.5*log(x)
2332 //
2333 // pow, powf, powl:
2334 // * pow(sqrt(x),y) -> pow(x,y*0.5)
2335 // * pow(pow(x,y),z)-> pow(x,y*z)
2336 //
2337 // signbit:
2338 // * signbit(cnst) -> cnst'
2339 // * signbit(nncst) -> 0 (if pstv is a non-negative constant)
2340 //
2341 // sqrt, sqrtf, sqrtl:
2342 // * sqrt(expN(x)) -> expN(x*0.5)
2343 // * sqrt(Nroot(x)) -> pow(x,1/(2*N))
2344 // * sqrt(pow(x,y)) -> pow(|x|,y*0.5)
2345 //
2346 
2347 //===----------------------------------------------------------------------===//
2348 // Fortified Library Call Optimizations
2349 //===----------------------------------------------------------------------===//
2350 
2351 bool FortifiedLibCallSimplifier::isFortifiedCallFoldable(CallInst *CI,
2352  unsigned ObjSizeOp,
2353  unsigned SizeOp,
2354  bool isString) {
2355  if (CI->getArgOperand(ObjSizeOp) == CI->getArgOperand(SizeOp))
2356  return true;
2357  if (ConstantInt *ObjSizeCI =
2358  dyn_cast<ConstantInt>(CI->getArgOperand(ObjSizeOp))) {
2359  if (ObjSizeCI->isMinusOne())
2360  return true;
2361  // If the object size wasn't -1 (unknown), bail out if we were asked to.
2362  if (OnlyLowerUnknownSize)
2363  return false;
2364  if (isString) {
2365  uint64_t Len = GetStringLength(CI->getArgOperand(SizeOp));
2366  // If the length is 0 we don't know how long it is and so we can't
2367  // remove the check.
2368  if (Len == 0)
2369  return false;
2370  return ObjSizeCI->getZExtValue() >= Len;
2371  }
2372  if (ConstantInt *SizeCI = dyn_cast<ConstantInt>(CI->getArgOperand(SizeOp)))
2373  return ObjSizeCI->getZExtValue() >= SizeCI->getZExtValue();
2374  }
2375  return false;
2376 }
2377 
2378 Value *FortifiedLibCallSimplifier::optimizeMemCpyChk(CallInst *CI,
2379  IRBuilder<> &B) {
2380  if (isFortifiedCallFoldable(CI, 3, 2, false)) {
2381  B.CreateMemCpy(CI->getArgOperand(0), CI->getArgOperand(1),
2382  CI->getArgOperand(2), 1);
2383  return CI->getArgOperand(0);
2384  }
2385  return nullptr;
2386 }
2387 
2388 Value *FortifiedLibCallSimplifier::optimizeMemMoveChk(CallInst *CI,
2389  IRBuilder<> &B) {
2390  if (isFortifiedCallFoldable(CI, 3, 2, false)) {
2391  B.CreateMemMove(CI->getArgOperand(0), CI->getArgOperand(1),
2392  CI->getArgOperand(2), 1);
2393  return CI->getArgOperand(0);
2394  }
2395  return nullptr;
2396 }
2397 
2398 Value *FortifiedLibCallSimplifier::optimizeMemSetChk(CallInst *CI,
2399  IRBuilder<> &B) {
2400  // TODO: Try foldMallocMemset() here.
2401 
2402  if (isFortifiedCallFoldable(CI, 3, 2, false)) {
2403  Value *Val = B.CreateIntCast(CI->getArgOperand(1), B.getInt8Ty(), false);
2404  B.CreateMemSet(CI->getArgOperand(0), Val, CI->getArgOperand(2), 1);
2405  return CI->getArgOperand(0);
2406  }
2407  return nullptr;
2408 }
2409 
2410 Value *FortifiedLibCallSimplifier::optimizeStrpCpyChk(CallInst *CI,
2411  IRBuilder<> &B,
2412  LibFunc Func) {
2413  Function *Callee = CI->getCalledFunction();
2414  StringRef Name = Callee->getName();
2415  const DataLayout &DL = CI->getModule()->getDataLayout();
2416  Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1),
2417  *ObjSize = CI->getArgOperand(2);
2418 
2419  // __stpcpy_chk(x,x,...) -> x+strlen(x)
2420  if (Func == LibFunc_stpcpy_chk && !OnlyLowerUnknownSize && Dst == Src) {
2421  Value *StrLen = emitStrLen(Src, B, DL, TLI);
2422  return StrLen ? B.CreateInBoundsGEP(B.getInt8Ty(), Dst, StrLen) : nullptr;
2423  }
2424 
2425  // If a) we don't have any length information, or b) we know this will
2426  // fit then just lower to a plain st[rp]cpy. Otherwise we'll keep our
2427  // st[rp]cpy_chk call which may fail at runtime if the size is too long.
2428  // TODO: It might be nice to get a maximum length out of the possible
2429  // string lengths for varying.
2430  if (isFortifiedCallFoldable(CI, 2, 1, true))
2431  return emitStrCpy(Dst, Src, B, TLI, Name.substr(2, 6));
2432 
2433  if (OnlyLowerUnknownSize)
2434  return nullptr;
2435 
2436  // Maybe we can stil fold __st[rp]cpy_chk to __memcpy_chk.
2437  uint64_t Len = GetStringLength(Src);
2438  if (Len == 0)
2439  return nullptr;
2440 
2441  Type *SizeTTy = DL.getIntPtrType(CI->getContext());
2442  Value *LenV = ConstantInt::get(SizeTTy, Len);
2443  Value *Ret = emitMemCpyChk(Dst, Src, LenV, ObjSize, B, DL, TLI);
2444  // If the function was an __stpcpy_chk, and we were able to fold it into
2445  // a __memcpy_chk, we still need to return the correct end pointer.
2446  if (Ret && Func == LibFunc_stpcpy_chk)
2447  return B.CreateGEP(B.getInt8Ty(), Dst, ConstantInt::get(SizeTTy, Len - 1));
2448  return Ret;
2449 }
2450 
2451 Value *FortifiedLibCallSimplifier::optimizeStrpNCpyChk(CallInst *CI,
2452  IRBuilder<> &B,
2453  LibFunc Func) {
2454  Function *Callee = CI->getCalledFunction();
2455  StringRef Name = Callee->getName();
2456  if (isFortifiedCallFoldable(CI, 3, 2, false)) {
2457  Value *Ret = emitStrNCpy(CI->getArgOperand(0), CI->getArgOperand(1),
2458  CI->getArgOperand(2), B, TLI, Name.substr(2, 7));
2459  return Ret;
2460  }
2461  return nullptr;
2462 }
2463 
2465  // FIXME: We shouldn't be changing "nobuiltin" or TLI unavailable calls here.
2466  // Some clang users checked for _chk libcall availability using:
2467  // __has_builtin(__builtin___memcpy_chk)
2468  // When compiling with -fno-builtin, this is always true.
2469  // When passing -ffreestanding/-mkernel, which both imply -fno-builtin, we
2470  // end up with fortified libcalls, which isn't acceptable in a freestanding
2471  // environment which only provides their non-fortified counterparts.
2472  //
2473  // Until we change clang and/or teach external users to check for availability
2474  // differently, disregard the "nobuiltin" attribute and TLI::has.
2475  //
2476  // PR23093.
2477 
2478  LibFunc Func;
2479  Function *Callee = CI->getCalledFunction();
2480 
2482  CI->getOperandBundlesAsDefs(OpBundles);
2483  IRBuilder<> Builder(CI, /*FPMathTag=*/nullptr, OpBundles);
2484  bool isCallingConvC = isCallingConvCCompatible(CI);
2485 
2486  // First, check that this is a known library functions and that the prototype
2487  // is correct.
2488  if (!TLI->getLibFunc(*Callee, Func))
2489  return nullptr;
2490 
2491  // We never change the calling convention.
2492  if (!ignoreCallingConv(Func) && !isCallingConvC)
2493  return nullptr;
2494 
2495  switch (Func) {
2496  case LibFunc_memcpy_chk:
2497  return optimizeMemCpyChk(CI, Builder);
2498  case LibFunc_memmove_chk:
2499  return optimizeMemMoveChk(CI, Builder);
2500  case LibFunc_memset_chk:
2501  return optimizeMemSetChk(CI, Builder);
2502  case LibFunc_stpcpy_chk:
2503  case LibFunc_strcpy_chk:
2504  return optimizeStrpCpyChk(CI, Builder, Func);
2505  case LibFunc_stpncpy_chk:
2506  case LibFunc_strncpy_chk:
2507  return optimizeStrpNCpyChk(CI, Builder, Func);
2508  default:
2509  break;
2510  }
2511  return nullptr;
2512 }
2513 
2515  const TargetLibraryInfo *TLI, bool OnlyLowerUnknownSize)
2516  : 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:67
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
bool isZero() const
Definition: APFloat.h:1143
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:712
ARM_AAPCS - ARM Architecture Procedure Calling Standard calling convention (aka EABI).
Definition: CallingConv.h:100
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.
LLVMContext & getContext() const
All values hold a context through their type.
Definition: Value.cpp:738
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:206
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
ARM_APCS - ARM Procedure Calling Standard calling convention (obsolete, but still used on some target...
Definition: CallingConv.h:96
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
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
apfloat_match m_APFloat(const APFloat *&Res)
Match a ConstantFP or splatted ConstantVector, binding the specified pointer to the contained APFloat...
Definition: PatternMatch.h:264
const ConstantDataArray * Array
ConstantDataArray pointer.
opStatus convert(const fltSemantics &ToSemantics, roundingMode RM, bool *losesInfo)
Definition: APFloat.cpp:4441
#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 isNegative() const
Return true if the sign bit is set.
Definition: Constants.h:300
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:126
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:439
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:301
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:979
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:1677
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:586
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
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
ARM_AAPCS_VFP - Same as ARM_AAPCS, but uses hard floating point ABI.
Definition: CallingConv.h:103
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:250
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 isExactlyValue(double V) const
We don&#39;t rely on operator== working on double values, as it returns true for things that are clearly ...
Definition: APFloat.h:1130
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:821
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:544
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:59
Constant * ConstantFoldLoadFromConstPtr(Constant *C, Type *Ty, const DataLayout &DL)
ConstantFoldLoadFromConstPtr - Return the value that a load from C would produce if it is constant an...
static Constant * getAllOnesValue(Type *Ty)
Get the all ones value.
Definition: Constants.cpp:260
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:567
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
const APFloat & getValueAPF() const
Definition: Constants.h:294
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
Value * CreateFAdd(Value *LHS, Value *RHS, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:907
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:862
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:2563
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.
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:559
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:622
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:55
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:405
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
C - The default llvm calling convention, compatible with C.
Definition: CallingConv.h:35
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:647
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:224
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:246
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
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:381
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:559
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
bool isExactlyValue(const APFloat &V) const
We don&#39;t rely on operator== working on double values, as it returns true for things that are clearly ...
Definition: Constants.cpp:728
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:1595
bool isDoubleTy() const
Return true if this is &#39;double&#39;, a 64-bit IEEE fp type.
Definition: Type.h:150
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...
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)
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
bool isArrayTy() const
True if this is an instance of ArrayType.
Definition: Type.h:218
A wrapper class for inspecting calls to intrinsic functions.
Definition: IntrinsicInst.h:44
const BasicBlock * getParent() const
Definition: Instruction.h:67
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:389