• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 
17 #include "llvm/Transforms/Utils/SimplifyLibCalls.h"
18 #include "llvm/ADT/SmallString.h"
19 #include "llvm/ADT/StringMap.h"
20 #include "llvm/ADT/Triple.h"
21 #include "llvm/Analysis/TargetLibraryInfo.h"
22 #include "llvm/Analysis/ValueTracking.h"
23 #include "llvm/IR/DataLayout.h"
24 #include "llvm/IR/DiagnosticInfo.h"
25 #include "llvm/IR/Function.h"
26 #include "llvm/IR/IRBuilder.h"
27 #include "llvm/IR/IntrinsicInst.h"
28 #include "llvm/IR/Intrinsics.h"
29 #include "llvm/IR/LLVMContext.h"
30 #include "llvm/IR/Module.h"
31 #include "llvm/IR/PatternMatch.h"
32 #include "llvm/Support/CommandLine.h"
33 #include "llvm/Transforms/Utils/BuildLibCalls.h"
34 #include "llvm/Transforms/Utils/Local.h"
35 
36 using namespace llvm;
37 using namespace PatternMatch;
38 
39 static cl::opt<bool>
40     ColdErrorCalls("error-reporting-is-cold", cl::init(true), cl::Hidden,
41                    cl::desc("Treat error-reporting calls as cold"));
42 
43 static cl::opt<bool>
44     EnableUnsafeFPShrink("enable-double-float-shrink", cl::Hidden,
45                          cl::init(false),
46                          cl::desc("Enable unsafe double to float "
47                                   "shrinking for math lib calls"));
48 
49 
50 //===----------------------------------------------------------------------===//
51 // Helper Functions
52 //===----------------------------------------------------------------------===//
53 
ignoreCallingConv(LibFunc::Func Func)54 static bool ignoreCallingConv(LibFunc::Func Func) {
55   return Func == LibFunc::abs || Func == LibFunc::labs ||
56          Func == LibFunc::llabs || Func == LibFunc::strlen;
57 }
58 
59 /// Return true if it only matters that the value is equal or not-equal to zero.
isOnlyUsedInZeroEqualityComparison(Value * V)60 static bool isOnlyUsedInZeroEqualityComparison(Value *V) {
61   for (User *U : V->users()) {
62     if (ICmpInst *IC = dyn_cast<ICmpInst>(U))
63       if (IC->isEquality())
64         if (Constant *C = dyn_cast<Constant>(IC->getOperand(1)))
65           if (C->isNullValue())
66             continue;
67     // Unknown instruction.
68     return false;
69   }
70   return true;
71 }
72 
73 /// Return true if it is only used in equality comparisons with With.
isOnlyUsedInEqualityComparison(Value * V,Value * With)74 static bool isOnlyUsedInEqualityComparison(Value *V, Value *With) {
75   for (User *U : V->users()) {
76     if (ICmpInst *IC = dyn_cast<ICmpInst>(U))
77       if (IC->isEquality() && IC->getOperand(1) == With)
78         continue;
79     // Unknown instruction.
80     return false;
81   }
82   return true;
83 }
84 
callHasFloatingPointArgument(const CallInst * CI)85 static bool callHasFloatingPointArgument(const CallInst *CI) {
86   return std::any_of(CI->op_begin(), CI->op_end(), [](const Use &OI) {
87     return OI->getType()->isFloatingPointTy();
88   });
89 }
90 
91 /// \brief Check whether the overloaded unary floating point function
92 /// corresponding to \a Ty is available.
hasUnaryFloatFn(const TargetLibraryInfo * TLI,Type * Ty,LibFunc::Func DoubleFn,LibFunc::Func FloatFn,LibFunc::Func LongDoubleFn)93 static bool hasUnaryFloatFn(const TargetLibraryInfo *TLI, Type *Ty,
94                             LibFunc::Func DoubleFn, LibFunc::Func FloatFn,
95                             LibFunc::Func LongDoubleFn) {
96   switch (Ty->getTypeID()) {
97   case Type::FloatTyID:
98     return TLI->has(FloatFn);
99   case Type::DoubleTyID:
100     return TLI->has(DoubleFn);
101   default:
102     return TLI->has(LongDoubleFn);
103   }
104 }
105 
106 //===----------------------------------------------------------------------===//
107 // String and Memory Library Call Optimizations
108 //===----------------------------------------------------------------------===//
109 
optimizeStrCat(CallInst * CI,IRBuilder<> & B)110 Value *LibCallSimplifier::optimizeStrCat(CallInst *CI, IRBuilder<> &B) {
111   // Extract some information from the instruction
112   Value *Dst = CI->getArgOperand(0);
113   Value *Src = CI->getArgOperand(1);
114 
115   // See if we can get the length of the input string.
116   uint64_t Len = GetStringLength(Src);
117   if (Len == 0)
118     return nullptr;
119   --Len; // Unbias length.
120 
121   // Handle the simple, do-nothing case: strcat(x, "") -> x
122   if (Len == 0)
123     return Dst;
124 
125   return emitStrLenMemCpy(Src, Dst, Len, B);
126 }
127 
emitStrLenMemCpy(Value * Src,Value * Dst,uint64_t Len,IRBuilder<> & B)128 Value *LibCallSimplifier::emitStrLenMemCpy(Value *Src, Value *Dst, uint64_t Len,
129                                            IRBuilder<> &B) {
130   // We need to find the end of the destination string.  That's where the
131   // memory is to be moved to. We just generate a call to strlen.
132   Value *DstLen = emitStrLen(Dst, B, DL, TLI);
133   if (!DstLen)
134     return nullptr;
135 
136   // Now that we have the destination's length, we must index into the
137   // destination's pointer to get the actual memcpy destination (end of
138   // the string .. we're concatenating).
139   Value *CpyDst = B.CreateGEP(B.getInt8Ty(), Dst, DstLen, "endptr");
140 
141   // We have enough information to now generate the memcpy call to do the
142   // concatenation for us.  Make a memcpy to copy the nul byte with align = 1.
143   B.CreateMemCpy(CpyDst, Src,
144                  ConstantInt::get(DL.getIntPtrType(Src->getContext()), Len + 1),
145                  1);
146   return Dst;
147 }
148 
optimizeStrNCat(CallInst * CI,IRBuilder<> & B)149 Value *LibCallSimplifier::optimizeStrNCat(CallInst *CI, IRBuilder<> &B) {
150   // Extract some information from the instruction.
151   Value *Dst = CI->getArgOperand(0);
152   Value *Src = CI->getArgOperand(1);
153   uint64_t Len;
154 
155   // We don't do anything if length is not constant.
156   if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(CI->getArgOperand(2)))
157     Len = LengthArg->getZExtValue();
158   else
159     return nullptr;
160 
161   // See if we can get the length of the input string.
162   uint64_t SrcLen = GetStringLength(Src);
163   if (SrcLen == 0)
164     return nullptr;
165   --SrcLen; // Unbias length.
166 
167   // Handle the simple, do-nothing cases:
168   // strncat(x, "", c) -> x
169   // strncat(x,  c, 0) -> x
170   if (SrcLen == 0 || Len == 0)
171     return Dst;
172 
173   // We don't optimize this case.
174   if (Len < SrcLen)
175     return nullptr;
176 
177   // strncat(x, s, c) -> strcat(x, s)
178   // s is constant so the strcat can be optimized further.
179   return emitStrLenMemCpy(Src, Dst, SrcLen, B);
180 }
181 
optimizeStrChr(CallInst * CI,IRBuilder<> & B)182 Value *LibCallSimplifier::optimizeStrChr(CallInst *CI, IRBuilder<> &B) {
183   Function *Callee = CI->getCalledFunction();
184   FunctionType *FT = Callee->getFunctionType();
185   Value *SrcStr = CI->getArgOperand(0);
186 
187   // If the second operand is non-constant, see if we can compute the length
188   // of the input string and turn this into memchr.
189   ConstantInt *CharC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
190   if (!CharC) {
191     uint64_t Len = GetStringLength(SrcStr);
192     if (Len == 0 || !FT->getParamType(1)->isIntegerTy(32)) // memchr needs i32.
193       return nullptr;
194 
195     return emitMemChr(SrcStr, CI->getArgOperand(1), // include nul.
196                       ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len),
197                       B, DL, TLI);
198   }
199 
200   // Otherwise, the character is a constant, see if the first argument is
201   // a string literal.  If so, we can constant fold.
202   StringRef Str;
203   if (!getConstantStringInfo(SrcStr, Str)) {
204     if (CharC->isZero()) // strchr(p, 0) -> p + strlen(p)
205       return B.CreateGEP(B.getInt8Ty(), SrcStr, emitStrLen(SrcStr, B, DL, TLI),
206                          "strchr");
207     return nullptr;
208   }
209 
210   // Compute the offset, make sure to handle the case when we're searching for
211   // zero (a weird way to spell strlen).
212   size_t I = (0xFF & CharC->getSExtValue()) == 0
213                  ? Str.size()
214                  : Str.find(CharC->getSExtValue());
215   if (I == StringRef::npos) // Didn't find the char.  strchr returns null.
216     return Constant::getNullValue(CI->getType());
217 
218   // strchr(s+n,c)  -> gep(s+n+i,c)
219   return B.CreateGEP(B.getInt8Ty(), SrcStr, B.getInt64(I), "strchr");
220 }
221 
optimizeStrRChr(CallInst * CI,IRBuilder<> & B)222 Value *LibCallSimplifier::optimizeStrRChr(CallInst *CI, IRBuilder<> &B) {
223   Value *SrcStr = CI->getArgOperand(0);
224   ConstantInt *CharC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
225 
226   // Cannot fold anything if we're not looking for a constant.
227   if (!CharC)
228     return nullptr;
229 
230   StringRef Str;
231   if (!getConstantStringInfo(SrcStr, Str)) {
232     // strrchr(s, 0) -> strchr(s, 0)
233     if (CharC->isZero())
234       return emitStrChr(SrcStr, '\0', B, TLI);
235     return nullptr;
236   }
237 
238   // Compute the offset.
239   size_t I = (0xFF & CharC->getSExtValue()) == 0
240                  ? Str.size()
241                  : Str.rfind(CharC->getSExtValue());
242   if (I == StringRef::npos) // Didn't find the char. Return null.
243     return Constant::getNullValue(CI->getType());
244 
245   // strrchr(s+n,c) -> gep(s+n+i,c)
246   return B.CreateGEP(B.getInt8Ty(), SrcStr, B.getInt64(I), "strrchr");
247 }
248 
optimizeStrCmp(CallInst * CI,IRBuilder<> & B)249 Value *LibCallSimplifier::optimizeStrCmp(CallInst *CI, IRBuilder<> &B) {
250   Value *Str1P = CI->getArgOperand(0), *Str2P = CI->getArgOperand(1);
251   if (Str1P == Str2P) // strcmp(x,x)  -> 0
252     return ConstantInt::get(CI->getType(), 0);
253 
254   StringRef Str1, Str2;
255   bool HasStr1 = getConstantStringInfo(Str1P, Str1);
256   bool HasStr2 = getConstantStringInfo(Str2P, Str2);
257 
258   // strcmp(x, y)  -> cnst  (if both x and y are constant strings)
259   if (HasStr1 && HasStr2)
260     return ConstantInt::get(CI->getType(), Str1.compare(Str2));
261 
262   if (HasStr1 && Str1.empty()) // strcmp("", x) -> -*x
263     return B.CreateNeg(
264         B.CreateZExt(B.CreateLoad(Str2P, "strcmpload"), CI->getType()));
265 
266   if (HasStr2 && Str2.empty()) // strcmp(x,"") -> *x
267     return B.CreateZExt(B.CreateLoad(Str1P, "strcmpload"), CI->getType());
268 
269   // strcmp(P, "x") -> memcmp(P, "x", 2)
270   uint64_t Len1 = GetStringLength(Str1P);
271   uint64_t Len2 = GetStringLength(Str2P);
272   if (Len1 && Len2) {
273     return emitMemCmp(Str1P, Str2P,
274                       ConstantInt::get(DL.getIntPtrType(CI->getContext()),
275                                        std::min(Len1, Len2)),
276                       B, DL, TLI);
277   }
278 
279   return nullptr;
280 }
281 
optimizeStrNCmp(CallInst * CI,IRBuilder<> & B)282 Value *LibCallSimplifier::optimizeStrNCmp(CallInst *CI, IRBuilder<> &B) {
283   Value *Str1P = CI->getArgOperand(0), *Str2P = CI->getArgOperand(1);
284   if (Str1P == Str2P) // strncmp(x,x,n)  -> 0
285     return ConstantInt::get(CI->getType(), 0);
286 
287   // Get the length argument if it is constant.
288   uint64_t Length;
289   if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(CI->getArgOperand(2)))
290     Length = LengthArg->getZExtValue();
291   else
292     return nullptr;
293 
294   if (Length == 0) // strncmp(x,y,0)   -> 0
295     return ConstantInt::get(CI->getType(), 0);
296 
297   if (Length == 1) // strncmp(x,y,1) -> memcmp(x,y,1)
298     return emitMemCmp(Str1P, Str2P, CI->getArgOperand(2), B, DL, TLI);
299 
300   StringRef Str1, Str2;
301   bool HasStr1 = getConstantStringInfo(Str1P, Str1);
302   bool HasStr2 = getConstantStringInfo(Str2P, Str2);
303 
304   // strncmp(x, y)  -> cnst  (if both x and y are constant strings)
305   if (HasStr1 && HasStr2) {
306     StringRef SubStr1 = Str1.substr(0, Length);
307     StringRef SubStr2 = Str2.substr(0, Length);
308     return ConstantInt::get(CI->getType(), SubStr1.compare(SubStr2));
309   }
310 
311   if (HasStr1 && Str1.empty()) // strncmp("", x, n) -> -*x
312     return B.CreateNeg(
313         B.CreateZExt(B.CreateLoad(Str2P, "strcmpload"), CI->getType()));
314 
315   if (HasStr2 && Str2.empty()) // strncmp(x, "", n) -> *x
316     return B.CreateZExt(B.CreateLoad(Str1P, "strcmpload"), CI->getType());
317 
318   return nullptr;
319 }
320 
optimizeStrCpy(CallInst * CI,IRBuilder<> & B)321 Value *LibCallSimplifier::optimizeStrCpy(CallInst *CI, IRBuilder<> &B) {
322   Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1);
323   if (Dst == Src) // strcpy(x,x)  -> x
324     return Src;
325 
326   // See if we can get the length of the input string.
327   uint64_t Len = GetStringLength(Src);
328   if (Len == 0)
329     return nullptr;
330 
331   // We have enough information to now generate the memcpy call to do the
332   // copy for us.  Make a memcpy to copy the nul byte with align = 1.
333   B.CreateMemCpy(Dst, Src,
334                  ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len), 1);
335   return Dst;
336 }
337 
optimizeStpCpy(CallInst * CI,IRBuilder<> & B)338 Value *LibCallSimplifier::optimizeStpCpy(CallInst *CI, IRBuilder<> &B) {
339   Function *Callee = CI->getCalledFunction();
340   Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1);
341   if (Dst == Src) { // stpcpy(x,x)  -> x+strlen(x)
342     Value *StrLen = emitStrLen(Src, B, DL, TLI);
343     return StrLen ? B.CreateInBoundsGEP(B.getInt8Ty(), Dst, StrLen) : nullptr;
344   }
345 
346   // See if we can get the length of the input string.
347   uint64_t Len = GetStringLength(Src);
348   if (Len == 0)
349     return nullptr;
350 
351   Type *PT = Callee->getFunctionType()->getParamType(0);
352   Value *LenV = ConstantInt::get(DL.getIntPtrType(PT), Len);
353   Value *DstEnd = B.CreateGEP(B.getInt8Ty(), Dst,
354                               ConstantInt::get(DL.getIntPtrType(PT), Len - 1));
355 
356   // We have enough information to now generate the memcpy call to do the
357   // copy for us.  Make a memcpy to copy the nul byte with align = 1.
358   B.CreateMemCpy(Dst, Src, LenV, 1);
359   return DstEnd;
360 }
361 
optimizeStrNCpy(CallInst * CI,IRBuilder<> & B)362 Value *LibCallSimplifier::optimizeStrNCpy(CallInst *CI, IRBuilder<> &B) {
363   Function *Callee = CI->getCalledFunction();
364   Value *Dst = CI->getArgOperand(0);
365   Value *Src = CI->getArgOperand(1);
366   Value *LenOp = CI->getArgOperand(2);
367 
368   // See if we can get the length of the input string.
369   uint64_t SrcLen = GetStringLength(Src);
370   if (SrcLen == 0)
371     return nullptr;
372   --SrcLen;
373 
374   if (SrcLen == 0) {
375     // strncpy(x, "", y) -> memset(x, '\0', y, 1)
376     B.CreateMemSet(Dst, B.getInt8('\0'), LenOp, 1);
377     return Dst;
378   }
379 
380   uint64_t Len;
381   if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(LenOp))
382     Len = LengthArg->getZExtValue();
383   else
384     return nullptr;
385 
386   if (Len == 0)
387     return Dst; // strncpy(x, y, 0) -> x
388 
389   // Let strncpy handle the zero padding
390   if (Len > SrcLen + 1)
391     return nullptr;
392 
393   Type *PT = Callee->getFunctionType()->getParamType(0);
394   // strncpy(x, s, c) -> memcpy(x, s, c, 1) [s and c are constant]
395   B.CreateMemCpy(Dst, Src, ConstantInt::get(DL.getIntPtrType(PT), Len), 1);
396 
397   return Dst;
398 }
399 
optimizeStrLen(CallInst * CI,IRBuilder<> & B)400 Value *LibCallSimplifier::optimizeStrLen(CallInst *CI, IRBuilder<> &B) {
401   Value *Src = CI->getArgOperand(0);
402 
403   // Constant folding: strlen("xyz") -> 3
404   if (uint64_t Len = GetStringLength(Src))
405     return ConstantInt::get(CI->getType(), Len - 1);
406 
407   // If s is a constant pointer pointing to a string literal, we can fold
408   // strlen(s + x) to strlen(s) - x, when x is known to be in the range
409   // [0, strlen(s)] or the string has a single null terminator '\0' at the end.
410   // We only try to simplify strlen when the pointer s points to an array
411   // of i8. Otherwise, we would need to scale the offset x before doing the
412   // subtraction. This will make the optimization more complex, and it's not
413   // very useful because calling strlen for a pointer of other types is
414   // very uncommon.
415   if (GEPOperator *GEP = dyn_cast<GEPOperator>(Src)) {
416     if (!isGEPBasedOnPointerToString(GEP))
417       return nullptr;
418 
419     StringRef Str;
420     if (getConstantStringInfo(GEP->getOperand(0), Str, 0, false)) {
421       size_t NullTermIdx = Str.find('\0');
422 
423       // If the string does not have '\0', leave it to strlen to compute
424       // its length.
425       if (NullTermIdx == StringRef::npos)
426         return nullptr;
427 
428       Value *Offset = GEP->getOperand(2);
429       unsigned BitWidth = Offset->getType()->getIntegerBitWidth();
430       APInt KnownZero(BitWidth, 0);
431       APInt KnownOne(BitWidth, 0);
432       computeKnownBits(Offset, KnownZero, KnownOne, DL, 0, nullptr, CI,
433                        nullptr);
434       KnownZero.flipAllBits();
435       size_t ArrSize =
436              cast<ArrayType>(GEP->getSourceElementType())->getNumElements();
437 
438       // KnownZero's bits are flipped, so zeros in KnownZero now represent
439       // bits known to be zeros in Offset, and ones in KnowZero represent
440       // bits unknown in Offset. Therefore, Offset is known to be in range
441       // [0, NullTermIdx] when the flipped KnownZero is non-negative and
442       // unsigned-less-than NullTermIdx.
443       //
444       // If Offset is not provably in the range [0, NullTermIdx], we can still
445       // optimize if we can prove that the program has undefined behavior when
446       // Offset is outside that range. That is the case when GEP->getOperand(0)
447       // is a pointer to an object whose memory extent is NullTermIdx+1.
448       if ((KnownZero.isNonNegative() && KnownZero.ule(NullTermIdx)) ||
449           (GEP->isInBounds() && isa<GlobalVariable>(GEP->getOperand(0)) &&
450            NullTermIdx == ArrSize - 1))
451         return B.CreateSub(ConstantInt::get(CI->getType(), NullTermIdx),
452                            Offset);
453     }
454 
455     return nullptr;
456   }
457 
458   // strlen(x?"foo":"bars") --> x ? 3 : 4
459   if (SelectInst *SI = dyn_cast<SelectInst>(Src)) {
460     uint64_t LenTrue = GetStringLength(SI->getTrueValue());
461     uint64_t LenFalse = GetStringLength(SI->getFalseValue());
462     if (LenTrue && LenFalse) {
463       Function *Caller = CI->getParent()->getParent();
464       emitOptimizationRemark(CI->getContext(), "simplify-libcalls", *Caller,
465                              SI->getDebugLoc(),
466                              "folded strlen(select) to select of constants");
467       return B.CreateSelect(SI->getCondition(),
468                             ConstantInt::get(CI->getType(), LenTrue - 1),
469                             ConstantInt::get(CI->getType(), LenFalse - 1));
470     }
471   }
472 
473   // strlen(x) != 0 --> *x != 0
474   // strlen(x) == 0 --> *x == 0
475   if (isOnlyUsedInZeroEqualityComparison(CI))
476     return B.CreateZExt(B.CreateLoad(Src, "strlenfirst"), CI->getType());
477 
478   return nullptr;
479 }
480 
optimizeStrPBrk(CallInst * CI,IRBuilder<> & B)481 Value *LibCallSimplifier::optimizeStrPBrk(CallInst *CI, IRBuilder<> &B) {
482   StringRef S1, S2;
483   bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1);
484   bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2);
485 
486   // strpbrk(s, "") -> nullptr
487   // strpbrk("", s) -> nullptr
488   if ((HasS1 && S1.empty()) || (HasS2 && S2.empty()))
489     return Constant::getNullValue(CI->getType());
490 
491   // Constant folding.
492   if (HasS1 && HasS2) {
493     size_t I = S1.find_first_of(S2);
494     if (I == StringRef::npos) // No match.
495       return Constant::getNullValue(CI->getType());
496 
497     return B.CreateGEP(B.getInt8Ty(), CI->getArgOperand(0), B.getInt64(I),
498                        "strpbrk");
499   }
500 
501   // strpbrk(s, "a") -> strchr(s, 'a')
502   if (HasS2 && S2.size() == 1)
503     return emitStrChr(CI->getArgOperand(0), S2[0], B, TLI);
504 
505   return nullptr;
506 }
507 
optimizeStrTo(CallInst * CI,IRBuilder<> & B)508 Value *LibCallSimplifier::optimizeStrTo(CallInst *CI, IRBuilder<> &B) {
509   Value *EndPtr = CI->getArgOperand(1);
510   if (isa<ConstantPointerNull>(EndPtr)) {
511     // With a null EndPtr, this function won't capture the main argument.
512     // It would be readonly too, except that it still may write to errno.
513     CI->addAttribute(1, Attribute::NoCapture);
514   }
515 
516   return nullptr;
517 }
518 
optimizeStrSpn(CallInst * CI,IRBuilder<> & B)519 Value *LibCallSimplifier::optimizeStrSpn(CallInst *CI, IRBuilder<> &B) {
520   StringRef S1, S2;
521   bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1);
522   bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2);
523 
524   // strspn(s, "") -> 0
525   // strspn("", s) -> 0
526   if ((HasS1 && S1.empty()) || (HasS2 && S2.empty()))
527     return Constant::getNullValue(CI->getType());
528 
529   // Constant folding.
530   if (HasS1 && HasS2) {
531     size_t Pos = S1.find_first_not_of(S2);
532     if (Pos == StringRef::npos)
533       Pos = S1.size();
534     return ConstantInt::get(CI->getType(), Pos);
535   }
536 
537   return nullptr;
538 }
539 
optimizeStrCSpn(CallInst * CI,IRBuilder<> & B)540 Value *LibCallSimplifier::optimizeStrCSpn(CallInst *CI, IRBuilder<> &B) {
541   StringRef S1, S2;
542   bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1);
543   bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2);
544 
545   // strcspn("", s) -> 0
546   if (HasS1 && S1.empty())
547     return Constant::getNullValue(CI->getType());
548 
549   // Constant folding.
550   if (HasS1 && HasS2) {
551     size_t Pos = S1.find_first_of(S2);
552     if (Pos == StringRef::npos)
553       Pos = S1.size();
554     return ConstantInt::get(CI->getType(), Pos);
555   }
556 
557   // strcspn(s, "") -> strlen(s)
558   if (HasS2 && S2.empty())
559     return emitStrLen(CI->getArgOperand(0), B, DL, TLI);
560 
561   return nullptr;
562 }
563 
optimizeStrStr(CallInst * CI,IRBuilder<> & B)564 Value *LibCallSimplifier::optimizeStrStr(CallInst *CI, IRBuilder<> &B) {
565   // fold strstr(x, x) -> x.
566   if (CI->getArgOperand(0) == CI->getArgOperand(1))
567     return B.CreateBitCast(CI->getArgOperand(0), CI->getType());
568 
569   // fold strstr(a, b) == a -> strncmp(a, b, strlen(b)) == 0
570   if (isOnlyUsedInEqualityComparison(CI, CI->getArgOperand(0))) {
571     Value *StrLen = emitStrLen(CI->getArgOperand(1), B, DL, TLI);
572     if (!StrLen)
573       return nullptr;
574     Value *StrNCmp = emitStrNCmp(CI->getArgOperand(0), CI->getArgOperand(1),
575                                  StrLen, B, DL, TLI);
576     if (!StrNCmp)
577       return nullptr;
578     for (auto UI = CI->user_begin(), UE = CI->user_end(); UI != UE;) {
579       ICmpInst *Old = cast<ICmpInst>(*UI++);
580       Value *Cmp =
581           B.CreateICmp(Old->getPredicate(), StrNCmp,
582                        ConstantInt::getNullValue(StrNCmp->getType()), "cmp");
583       replaceAllUsesWith(Old, Cmp);
584     }
585     return CI;
586   }
587 
588   // See if either input string is a constant string.
589   StringRef SearchStr, ToFindStr;
590   bool HasStr1 = getConstantStringInfo(CI->getArgOperand(0), SearchStr);
591   bool HasStr2 = getConstantStringInfo(CI->getArgOperand(1), ToFindStr);
592 
593   // fold strstr(x, "") -> x.
594   if (HasStr2 && ToFindStr.empty())
595     return B.CreateBitCast(CI->getArgOperand(0), CI->getType());
596 
597   // If both strings are known, constant fold it.
598   if (HasStr1 && HasStr2) {
599     size_t Offset = SearchStr.find(ToFindStr);
600 
601     if (Offset == StringRef::npos) // strstr("foo", "bar") -> null
602       return Constant::getNullValue(CI->getType());
603 
604     // strstr("abcd", "bc") -> gep((char*)"abcd", 1)
605     Value *Result = castToCStr(CI->getArgOperand(0), B);
606     Result = B.CreateConstInBoundsGEP1_64(Result, Offset, "strstr");
607     return B.CreateBitCast(Result, CI->getType());
608   }
609 
610   // fold strstr(x, "y") -> strchr(x, 'y').
611   if (HasStr2 && ToFindStr.size() == 1) {
612     Value *StrChr = emitStrChr(CI->getArgOperand(0), ToFindStr[0], B, TLI);
613     return StrChr ? B.CreateBitCast(StrChr, CI->getType()) : nullptr;
614   }
615   return nullptr;
616 }
617 
optimizeMemChr(CallInst * CI,IRBuilder<> & B)618 Value *LibCallSimplifier::optimizeMemChr(CallInst *CI, IRBuilder<> &B) {
619   Value *SrcStr = CI->getArgOperand(0);
620   ConstantInt *CharC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
621   ConstantInt *LenC = dyn_cast<ConstantInt>(CI->getArgOperand(2));
622 
623   // memchr(x, y, 0) -> null
624   if (LenC && LenC->isNullValue())
625     return Constant::getNullValue(CI->getType());
626 
627   // From now on we need at least constant length and string.
628   StringRef Str;
629   if (!LenC || !getConstantStringInfo(SrcStr, Str, 0, /*TrimAtNul=*/false))
630     return nullptr;
631 
632   // Truncate the string to LenC. If Str is smaller than LenC we will still only
633   // scan the string, as reading past the end of it is undefined and we can just
634   // return null if we don't find the char.
635   Str = Str.substr(0, LenC->getZExtValue());
636 
637   // If the char is variable but the input str and length are not we can turn
638   // this memchr call into a simple bit field test. Of course this only works
639   // when the return value is only checked against null.
640   //
641   // It would be really nice to reuse switch lowering here but we can't change
642   // the CFG at this point.
643   //
644   // memchr("\r\n", C, 2) != nullptr -> (C & ((1 << '\r') | (1 << '\n'))) != 0
645   //   after bounds check.
646   if (!CharC && !Str.empty() && isOnlyUsedInZeroEqualityComparison(CI)) {
647     unsigned char Max =
648         *std::max_element(reinterpret_cast<const unsigned char *>(Str.begin()),
649                           reinterpret_cast<const unsigned char *>(Str.end()));
650 
651     // Make sure the bit field we're about to create fits in a register on the
652     // target.
653     // FIXME: On a 64 bit architecture this prevents us from using the
654     // interesting range of alpha ascii chars. We could do better by emitting
655     // two bitfields or shifting the range by 64 if no lower chars are used.
656     if (!DL.fitsInLegalInteger(Max + 1))
657       return nullptr;
658 
659     // For the bit field use a power-of-2 type with at least 8 bits to avoid
660     // creating unnecessary illegal types.
661     unsigned char Width = NextPowerOf2(std::max((unsigned char)7, Max));
662 
663     // Now build the bit field.
664     APInt Bitfield(Width, 0);
665     for (char C : Str)
666       Bitfield.setBit((unsigned char)C);
667     Value *BitfieldC = B.getInt(Bitfield);
668 
669     // First check that the bit field access is within bounds.
670     Value *C = B.CreateZExtOrTrunc(CI->getArgOperand(1), BitfieldC->getType());
671     Value *Bounds = B.CreateICmp(ICmpInst::ICMP_ULT, C, B.getIntN(Width, Width),
672                                  "memchr.bounds");
673 
674     // Create code that checks if the given bit is set in the field.
675     Value *Shl = B.CreateShl(B.getIntN(Width, 1ULL), C);
676     Value *Bits = B.CreateIsNotNull(B.CreateAnd(Shl, BitfieldC), "memchr.bits");
677 
678     // Finally merge both checks and cast to pointer type. The inttoptr
679     // implicitly zexts the i1 to intptr type.
680     return B.CreateIntToPtr(B.CreateAnd(Bounds, Bits, "memchr"), CI->getType());
681   }
682 
683   // Check if all arguments are constants.  If so, we can constant fold.
684   if (!CharC)
685     return nullptr;
686 
687   // Compute the offset.
688   size_t I = Str.find(CharC->getSExtValue() & 0xFF);
689   if (I == StringRef::npos) // Didn't find the char.  memchr returns null.
690     return Constant::getNullValue(CI->getType());
691 
692   // memchr(s+n,c,l) -> gep(s+n+i,c)
693   return B.CreateGEP(B.getInt8Ty(), SrcStr, B.getInt64(I), "memchr");
694 }
695 
optimizeMemCmp(CallInst * CI,IRBuilder<> & B)696 Value *LibCallSimplifier::optimizeMemCmp(CallInst *CI, IRBuilder<> &B) {
697   Value *LHS = CI->getArgOperand(0), *RHS = CI->getArgOperand(1);
698 
699   if (LHS == RHS) // memcmp(s,s,x) -> 0
700     return Constant::getNullValue(CI->getType());
701 
702   // Make sure we have a constant length.
703   ConstantInt *LenC = dyn_cast<ConstantInt>(CI->getArgOperand(2));
704   if (!LenC)
705     return nullptr;
706   uint64_t Len = LenC->getZExtValue();
707 
708   if (Len == 0) // memcmp(s1,s2,0) -> 0
709     return Constant::getNullValue(CI->getType());
710 
711   // memcmp(S1,S2,1) -> *(unsigned char*)LHS - *(unsigned char*)RHS
712   if (Len == 1) {
713     Value *LHSV = B.CreateZExt(B.CreateLoad(castToCStr(LHS, B), "lhsc"),
714                                CI->getType(), "lhsv");
715     Value *RHSV = B.CreateZExt(B.CreateLoad(castToCStr(RHS, B), "rhsc"),
716                                CI->getType(), "rhsv");
717     return B.CreateSub(LHSV, RHSV, "chardiff");
718   }
719 
720   // memcmp(S1,S2,N/8)==0 -> (*(intN_t*)S1 != *(intN_t*)S2)==0
721   if (DL.isLegalInteger(Len * 8) && isOnlyUsedInZeroEqualityComparison(CI)) {
722 
723     IntegerType *IntType = IntegerType::get(CI->getContext(), Len * 8);
724     unsigned PrefAlignment = DL.getPrefTypeAlignment(IntType);
725 
726     if (getKnownAlignment(LHS, DL, CI) >= PrefAlignment &&
727         getKnownAlignment(RHS, DL, CI) >= PrefAlignment) {
728 
729       Type *LHSPtrTy =
730           IntType->getPointerTo(LHS->getType()->getPointerAddressSpace());
731       Type *RHSPtrTy =
732           IntType->getPointerTo(RHS->getType()->getPointerAddressSpace());
733 
734       Value *LHSV =
735           B.CreateLoad(B.CreateBitCast(LHS, LHSPtrTy, "lhsc"), "lhsv");
736       Value *RHSV =
737           B.CreateLoad(B.CreateBitCast(RHS, RHSPtrTy, "rhsc"), "rhsv");
738 
739       return B.CreateZExt(B.CreateICmpNE(LHSV, RHSV), CI->getType(), "memcmp");
740     }
741   }
742 
743   // Constant folding: memcmp(x, y, l) -> cnst (all arguments are constant)
744   StringRef LHSStr, RHSStr;
745   if (getConstantStringInfo(LHS, LHSStr) &&
746       getConstantStringInfo(RHS, RHSStr)) {
747     // Make sure we're not reading out-of-bounds memory.
748     if (Len > LHSStr.size() || Len > RHSStr.size())
749       return nullptr;
750     // Fold the memcmp and normalize the result.  This way we get consistent
751     // results across multiple platforms.
752     uint64_t Ret = 0;
753     int Cmp = memcmp(LHSStr.data(), RHSStr.data(), Len);
754     if (Cmp < 0)
755       Ret = -1;
756     else if (Cmp > 0)
757       Ret = 1;
758     return ConstantInt::get(CI->getType(), Ret);
759   }
760 
761   return nullptr;
762 }
763 
optimizeMemCpy(CallInst * CI,IRBuilder<> & B)764 Value *LibCallSimplifier::optimizeMemCpy(CallInst *CI, IRBuilder<> &B) {
765   // memcpy(x, y, n) -> llvm.memcpy(x, y, n, 1)
766   B.CreateMemCpy(CI->getArgOperand(0), CI->getArgOperand(1),
767                  CI->getArgOperand(2), 1);
768   return CI->getArgOperand(0);
769 }
770 
optimizeMemMove(CallInst * CI,IRBuilder<> & B)771 Value *LibCallSimplifier::optimizeMemMove(CallInst *CI, IRBuilder<> &B) {
772   // memmove(x, y, n) -> llvm.memmove(x, y, n, 1)
773   B.CreateMemMove(CI->getArgOperand(0), CI->getArgOperand(1),
774                   CI->getArgOperand(2), 1);
775   return CI->getArgOperand(0);
776 }
777 
778 // TODO: Does this belong in BuildLibCalls or should all of those similar
779 // functions be moved here?
emitCalloc(Value * Num,Value * Size,const AttributeSet & Attrs,IRBuilder<> & B,const TargetLibraryInfo & TLI)780 static Value *emitCalloc(Value *Num, Value *Size, const AttributeSet &Attrs,
781                          IRBuilder<> &B, const TargetLibraryInfo &TLI) {
782   LibFunc::Func Func;
783   if (!TLI.getLibFunc("calloc", Func) || !TLI.has(Func))
784     return nullptr;
785 
786   Module *M = B.GetInsertBlock()->getModule();
787   const DataLayout &DL = M->getDataLayout();
788   IntegerType *PtrType = DL.getIntPtrType((B.GetInsertBlock()->getContext()));
789   Value *Calloc = M->getOrInsertFunction("calloc", Attrs, B.getInt8PtrTy(),
790                                          PtrType, PtrType, nullptr);
791   CallInst *CI = B.CreateCall(Calloc, { Num, Size }, "calloc");
792 
793   if (const auto *F = dyn_cast<Function>(Calloc->stripPointerCasts()))
794     CI->setCallingConv(F->getCallingConv());
795 
796   return CI;
797 }
798 
799 /// Fold memset[_chk](malloc(n), 0, n) --> calloc(1, n).
foldMallocMemset(CallInst * Memset,IRBuilder<> & B,const TargetLibraryInfo & TLI)800 static Value *foldMallocMemset(CallInst *Memset, IRBuilder<> &B,
801                                const TargetLibraryInfo &TLI) {
802   // This has to be a memset of zeros (bzero).
803   auto *FillValue = dyn_cast<ConstantInt>(Memset->getArgOperand(1));
804   if (!FillValue || FillValue->getZExtValue() != 0)
805     return nullptr;
806 
807   // TODO: We should handle the case where the malloc has more than one use.
808   // This is necessary to optimize common patterns such as when the result of
809   // the malloc is checked against null or when a memset intrinsic is used in
810   // place of a memset library call.
811   auto *Malloc = dyn_cast<CallInst>(Memset->getArgOperand(0));
812   if (!Malloc || !Malloc->hasOneUse())
813     return nullptr;
814 
815   // Is the inner call really malloc()?
816   Function *InnerCallee = Malloc->getCalledFunction();
817   LibFunc::Func Func;
818   if (!TLI.getLibFunc(*InnerCallee, Func) || !TLI.has(Func) ||
819       Func != LibFunc::malloc)
820     return nullptr;
821 
822   // The memset must cover the same number of bytes that are malloc'd.
823   if (Memset->getArgOperand(2) != Malloc->getArgOperand(0))
824     return nullptr;
825 
826   // Replace the malloc with a calloc. We need the data layout to know what the
827   // actual size of a 'size_t' parameter is.
828   B.SetInsertPoint(Malloc->getParent(), ++Malloc->getIterator());
829   const DataLayout &DL = Malloc->getModule()->getDataLayout();
830   IntegerType *SizeType = DL.getIntPtrType(B.GetInsertBlock()->getContext());
831   Value *Calloc = emitCalloc(ConstantInt::get(SizeType, 1),
832                              Malloc->getArgOperand(0), Malloc->getAttributes(),
833                              B, TLI);
834   if (!Calloc)
835     return nullptr;
836 
837   Malloc->replaceAllUsesWith(Calloc);
838   Malloc->eraseFromParent();
839 
840   return Calloc;
841 }
842 
optimizeMemSet(CallInst * CI,IRBuilder<> & B)843 Value *LibCallSimplifier::optimizeMemSet(CallInst *CI, IRBuilder<> &B) {
844   if (auto *Calloc = foldMallocMemset(CI, B, *TLI))
845     return Calloc;
846 
847   // memset(p, v, n) -> llvm.memset(p, v, n, 1)
848   Value *Val = B.CreateIntCast(CI->getArgOperand(1), B.getInt8Ty(), false);
849   B.CreateMemSet(CI->getArgOperand(0), Val, CI->getArgOperand(2), 1);
850   return CI->getArgOperand(0);
851 }
852 
853 //===----------------------------------------------------------------------===//
854 // Math Library Optimizations
855 //===----------------------------------------------------------------------===//
856 
857 /// Return a variant of Val with float type.
858 /// Currently this works in two cases: If Val is an FPExtension of a float
859 /// value to something bigger, simply return the operand.
860 /// If Val is a ConstantFP but can be converted to a float ConstantFP without
861 /// loss of precision do so.
valueHasFloatPrecision(Value * Val)862 static Value *valueHasFloatPrecision(Value *Val) {
863   if (FPExtInst *Cast = dyn_cast<FPExtInst>(Val)) {
864     Value *Op = Cast->getOperand(0);
865     if (Op->getType()->isFloatTy())
866       return Op;
867   }
868   if (ConstantFP *Const = dyn_cast<ConstantFP>(Val)) {
869     APFloat F = Const->getValueAPF();
870     bool losesInfo;
871     (void)F.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven,
872                     &losesInfo);
873     if (!losesInfo)
874       return ConstantFP::get(Const->getContext(), F);
875   }
876   return nullptr;
877 }
878 
879 /// Shrink double -> float for unary functions like 'floor'.
optimizeUnaryDoubleFP(CallInst * CI,IRBuilder<> & B,bool CheckRetType)880 static Value *optimizeUnaryDoubleFP(CallInst *CI, IRBuilder<> &B,
881                                     bool CheckRetType) {
882   Function *Callee = CI->getCalledFunction();
883   // We know this libcall has a valid prototype, but we don't know which.
884   if (!CI->getType()->isDoubleTy())
885     return nullptr;
886 
887   if (CheckRetType) {
888     // Check if all the uses for function like 'sin' are converted to float.
889     for (User *U : CI->users()) {
890       FPTruncInst *Cast = dyn_cast<FPTruncInst>(U);
891       if (!Cast || !Cast->getType()->isFloatTy())
892         return nullptr;
893     }
894   }
895 
896   // If this is something like 'floor((double)floatval)', convert to floorf.
897   Value *V = valueHasFloatPrecision(CI->getArgOperand(0));
898   if (V == nullptr)
899     return nullptr;
900 
901   // Propagate fast-math flags from the existing call to the new call.
902   IRBuilder<>::FastMathFlagGuard Guard(B);
903   B.setFastMathFlags(CI->getFastMathFlags());
904 
905   // floor((double)floatval) -> (double)floorf(floatval)
906   if (Callee->isIntrinsic()) {
907     Module *M = CI->getModule();
908     Intrinsic::ID IID = Callee->getIntrinsicID();
909     Function *F = Intrinsic::getDeclaration(M, IID, B.getFloatTy());
910     V = B.CreateCall(F, V);
911   } else {
912     // The call is a library call rather than an intrinsic.
913     V = emitUnaryFloatFnCall(V, Callee->getName(), B, Callee->getAttributes());
914   }
915 
916   return B.CreateFPExt(V, B.getDoubleTy());
917 }
918 
919 /// Shrink double -> float for binary functions like 'fmin/fmax'.
optimizeBinaryDoubleFP(CallInst * CI,IRBuilder<> & B)920 static Value *optimizeBinaryDoubleFP(CallInst *CI, IRBuilder<> &B) {
921   Function *Callee = CI->getCalledFunction();
922   // We know this libcall has a valid prototype, but we don't know which.
923   if (!CI->getType()->isDoubleTy())
924     return nullptr;
925 
926   // If this is something like 'fmin((double)floatval1, (double)floatval2)',
927   // or fmin(1.0, (double)floatval), then we convert it to fminf.
928   Value *V1 = valueHasFloatPrecision(CI->getArgOperand(0));
929   if (V1 == nullptr)
930     return nullptr;
931   Value *V2 = valueHasFloatPrecision(CI->getArgOperand(1));
932   if (V2 == nullptr)
933     return nullptr;
934 
935   // Propagate fast-math flags from the existing call to the new call.
936   IRBuilder<>::FastMathFlagGuard Guard(B);
937   B.setFastMathFlags(CI->getFastMathFlags());
938 
939   // fmin((double)floatval1, (double)floatval2)
940   //                      -> (double)fminf(floatval1, floatval2)
941   // TODO: Handle intrinsics in the same way as in optimizeUnaryDoubleFP().
942   Value *V = emitBinaryFloatFnCall(V1, V2, Callee->getName(), B,
943                                    Callee->getAttributes());
944   return B.CreateFPExt(V, B.getDoubleTy());
945 }
946 
optimizeCos(CallInst * CI,IRBuilder<> & B)947 Value *LibCallSimplifier::optimizeCos(CallInst *CI, IRBuilder<> &B) {
948   Function *Callee = CI->getCalledFunction();
949   Value *Ret = nullptr;
950   StringRef Name = Callee->getName();
951   if (UnsafeFPShrink && Name == "cos" && hasFloatVersion(Name))
952     Ret = optimizeUnaryDoubleFP(CI, B, true);
953 
954   // cos(-x) -> cos(x)
955   Value *Op1 = CI->getArgOperand(0);
956   if (BinaryOperator::isFNeg(Op1)) {
957     BinaryOperator *BinExpr = cast<BinaryOperator>(Op1);
958     return B.CreateCall(Callee, BinExpr->getOperand(1), "cos");
959   }
960   return Ret;
961 }
962 
getPow(Value * InnerChain[33],unsigned Exp,IRBuilder<> & B)963 static Value *getPow(Value *InnerChain[33], unsigned Exp, IRBuilder<> &B) {
964   // Multiplications calculated using Addition Chains.
965   // Refer: http://wwwhomes.uni-bielefeld.de/achim/addition_chain.html
966 
967   assert(Exp != 0 && "Incorrect exponent 0 not handled");
968 
969   if (InnerChain[Exp])
970     return InnerChain[Exp];
971 
972   static const unsigned AddChain[33][2] = {
973       {0, 0}, // Unused.
974       {0, 0}, // Unused (base case = pow1).
975       {1, 1}, // Unused (pre-computed).
976       {1, 2},  {2, 2},   {2, 3},  {3, 3},   {2, 5},  {4, 4},
977       {1, 8},  {5, 5},   {1, 10}, {6, 6},   {4, 9},  {7, 7},
978       {3, 12}, {8, 8},   {8, 9},  {2, 16},  {1, 18}, {10, 10},
979       {6, 15}, {11, 11}, {3, 20}, {12, 12}, {8, 17}, {13, 13},
980       {3, 24}, {14, 14}, {4, 25}, {15, 15}, {3, 28}, {16, 16},
981   };
982 
983   InnerChain[Exp] = B.CreateFMul(getPow(InnerChain, AddChain[Exp][0], B),
984                                  getPow(InnerChain, AddChain[Exp][1], B));
985   return InnerChain[Exp];
986 }
987 
optimizePow(CallInst * CI,IRBuilder<> & B)988 Value *LibCallSimplifier::optimizePow(CallInst *CI, IRBuilder<> &B) {
989   Function *Callee = CI->getCalledFunction();
990   Value *Ret = nullptr;
991   StringRef Name = Callee->getName();
992   if (UnsafeFPShrink && Name == "pow" && hasFloatVersion(Name))
993     Ret = optimizeUnaryDoubleFP(CI, B, true);
994 
995   Value *Op1 = CI->getArgOperand(0), *Op2 = CI->getArgOperand(1);
996   if (ConstantFP *Op1C = dyn_cast<ConstantFP>(Op1)) {
997     // pow(1.0, x) -> 1.0
998     if (Op1C->isExactlyValue(1.0))
999       return Op1C;
1000     // pow(2.0, x) -> exp2(x)
1001     if (Op1C->isExactlyValue(2.0) &&
1002         hasUnaryFloatFn(TLI, Op1->getType(), LibFunc::exp2, LibFunc::exp2f,
1003                         LibFunc::exp2l))
1004       return emitUnaryFloatFnCall(Op2, TLI->getName(LibFunc::exp2), B,
1005                                   Callee->getAttributes());
1006     // pow(10.0, x) -> exp10(x)
1007     if (Op1C->isExactlyValue(10.0) &&
1008         hasUnaryFloatFn(TLI, Op1->getType(), LibFunc::exp10, LibFunc::exp10f,
1009                         LibFunc::exp10l))
1010       return emitUnaryFloatFnCall(Op2, TLI->getName(LibFunc::exp10), B,
1011                                   Callee->getAttributes());
1012   }
1013 
1014   // pow(exp(x), y) -> exp(x * y)
1015   // pow(exp2(x), y) -> exp2(x * y)
1016   // We enable these only with fast-math. Besides rounding differences, the
1017   // transformation changes overflow and underflow behavior quite dramatically.
1018   // Example: x = 1000, y = 0.001.
1019   // pow(exp(x), y) = pow(inf, 0.001) = inf, whereas exp(x*y) = exp(1).
1020   auto *OpC = dyn_cast<CallInst>(Op1);
1021   if (OpC && OpC->hasUnsafeAlgebra() && CI->hasUnsafeAlgebra()) {
1022     LibFunc::Func Func;
1023     Function *OpCCallee = OpC->getCalledFunction();
1024     if (OpCCallee && TLI->getLibFunc(OpCCallee->getName(), Func) &&
1025         TLI->has(Func) && (Func == LibFunc::exp || Func == LibFunc::exp2)) {
1026       IRBuilder<>::FastMathFlagGuard Guard(B);
1027       B.setFastMathFlags(CI->getFastMathFlags());
1028       Value *FMul = B.CreateFMul(OpC->getArgOperand(0), Op2, "mul");
1029       return emitUnaryFloatFnCall(FMul, OpCCallee->getName(), B,
1030                                   OpCCallee->getAttributes());
1031     }
1032   }
1033 
1034   ConstantFP *Op2C = dyn_cast<ConstantFP>(Op2);
1035   if (!Op2C)
1036     return Ret;
1037 
1038   if (Op2C->getValueAPF().isZero()) // pow(x, 0.0) -> 1.0
1039     return ConstantFP::get(CI->getType(), 1.0);
1040 
1041   if (Op2C->isExactlyValue(0.5) &&
1042       hasUnaryFloatFn(TLI, Op2->getType(), LibFunc::sqrt, LibFunc::sqrtf,
1043                       LibFunc::sqrtl) &&
1044       hasUnaryFloatFn(TLI, Op2->getType(), LibFunc::fabs, LibFunc::fabsf,
1045                       LibFunc::fabsl)) {
1046 
1047     // In -ffast-math, pow(x, 0.5) -> sqrt(x).
1048     if (CI->hasUnsafeAlgebra()) {
1049       IRBuilder<>::FastMathFlagGuard Guard(B);
1050       B.setFastMathFlags(CI->getFastMathFlags());
1051       return emitUnaryFloatFnCall(Op1, TLI->getName(LibFunc::sqrt), B,
1052                                   Callee->getAttributes());
1053     }
1054 
1055     // Expand pow(x, 0.5) to (x == -infinity ? +infinity : fabs(sqrt(x))).
1056     // This is faster than calling pow, and still handles negative zero
1057     // and negative infinity correctly.
1058     // TODO: In finite-only mode, this could be just fabs(sqrt(x)).
1059     Value *Inf = ConstantFP::getInfinity(CI->getType());
1060     Value *NegInf = ConstantFP::getInfinity(CI->getType(), true);
1061     Value *Sqrt = emitUnaryFloatFnCall(Op1, "sqrt", B, Callee->getAttributes());
1062     Value *FAbs =
1063         emitUnaryFloatFnCall(Sqrt, "fabs", B, Callee->getAttributes());
1064     Value *FCmp = B.CreateFCmpOEQ(Op1, NegInf);
1065     Value *Sel = B.CreateSelect(FCmp, Inf, FAbs);
1066     return Sel;
1067   }
1068 
1069   if (Op2C->isExactlyValue(1.0)) // pow(x, 1.0) -> x
1070     return Op1;
1071   if (Op2C->isExactlyValue(2.0)) // pow(x, 2.0) -> x*x
1072     return B.CreateFMul(Op1, Op1, "pow2");
1073   if (Op2C->isExactlyValue(-1.0)) // pow(x, -1.0) -> 1.0/x
1074     return B.CreateFDiv(ConstantFP::get(CI->getType(), 1.0), Op1, "powrecip");
1075 
1076   // In -ffast-math, generate repeated fmul instead of generating pow(x, n).
1077   if (CI->hasUnsafeAlgebra()) {
1078     APFloat V = abs(Op2C->getValueAPF());
1079     // We limit to a max of 7 fmul(s). Thus max exponent is 32.
1080     // This transformation applies to integer exponents only.
1081     if (V.compare(APFloat(V.getSemantics(), 32.0)) == APFloat::cmpGreaterThan ||
1082         !V.isInteger())
1083       return nullptr;
1084 
1085     // We will memoize intermediate products of the Addition Chain.
1086     Value *InnerChain[33] = {nullptr};
1087     InnerChain[1] = Op1;
1088     InnerChain[2] = B.CreateFMul(Op1, Op1);
1089 
1090     // We cannot readily convert a non-double type (like float) to a double.
1091     // So we first convert V to something which could be converted to double.
1092     bool ignored;
1093     V.convert(APFloat::IEEEdouble, APFloat::rmTowardZero, &ignored);
1094 
1095     // TODO: Should the new instructions propagate the 'fast' flag of the pow()?
1096     Value *FMul = getPow(InnerChain, V.convertToDouble(), B);
1097     // For negative exponents simply compute the reciprocal.
1098     if (Op2C->isNegative())
1099       FMul = B.CreateFDiv(ConstantFP::get(CI->getType(), 1.0), FMul);
1100     return FMul;
1101   }
1102 
1103   return nullptr;
1104 }
1105 
optimizeExp2(CallInst * CI,IRBuilder<> & B)1106 Value *LibCallSimplifier::optimizeExp2(CallInst *CI, IRBuilder<> &B) {
1107   Function *Callee = CI->getCalledFunction();
1108   Value *Ret = nullptr;
1109   StringRef Name = Callee->getName();
1110   if (UnsafeFPShrink && Name == "exp2" && hasFloatVersion(Name))
1111     Ret = optimizeUnaryDoubleFP(CI, B, true);
1112 
1113   Value *Op = CI->getArgOperand(0);
1114   // Turn exp2(sitofp(x)) -> ldexp(1.0, sext(x))  if sizeof(x) <= 32
1115   // Turn exp2(uitofp(x)) -> ldexp(1.0, zext(x))  if sizeof(x) < 32
1116   LibFunc::Func LdExp = LibFunc::ldexpl;
1117   if (Op->getType()->isFloatTy())
1118     LdExp = LibFunc::ldexpf;
1119   else if (Op->getType()->isDoubleTy())
1120     LdExp = LibFunc::ldexp;
1121 
1122   if (TLI->has(LdExp)) {
1123     Value *LdExpArg = nullptr;
1124     if (SIToFPInst *OpC = dyn_cast<SIToFPInst>(Op)) {
1125       if (OpC->getOperand(0)->getType()->getPrimitiveSizeInBits() <= 32)
1126         LdExpArg = B.CreateSExt(OpC->getOperand(0), B.getInt32Ty());
1127     } else if (UIToFPInst *OpC = dyn_cast<UIToFPInst>(Op)) {
1128       if (OpC->getOperand(0)->getType()->getPrimitiveSizeInBits() < 32)
1129         LdExpArg = B.CreateZExt(OpC->getOperand(0), B.getInt32Ty());
1130     }
1131 
1132     if (LdExpArg) {
1133       Constant *One = ConstantFP::get(CI->getContext(), APFloat(1.0f));
1134       if (!Op->getType()->isFloatTy())
1135         One = ConstantExpr::getFPExtend(One, Op->getType());
1136 
1137       Module *M = CI->getModule();
1138       Value *NewCallee =
1139           M->getOrInsertFunction(TLI->getName(LdExp), Op->getType(),
1140                                  Op->getType(), B.getInt32Ty(), nullptr);
1141       CallInst *CI = B.CreateCall(NewCallee, {One, LdExpArg});
1142       if (const Function *F = dyn_cast<Function>(Callee->stripPointerCasts()))
1143         CI->setCallingConv(F->getCallingConv());
1144 
1145       return CI;
1146     }
1147   }
1148   return Ret;
1149 }
1150 
optimizeFabs(CallInst * CI,IRBuilder<> & B)1151 Value *LibCallSimplifier::optimizeFabs(CallInst *CI, IRBuilder<> &B) {
1152   Function *Callee = CI->getCalledFunction();
1153   Value *Ret = nullptr;
1154   StringRef Name = Callee->getName();
1155   if (Name == "fabs" && hasFloatVersion(Name))
1156     Ret = optimizeUnaryDoubleFP(CI, B, false);
1157 
1158   Value *Op = CI->getArgOperand(0);
1159   if (Instruction *I = dyn_cast<Instruction>(Op)) {
1160     // Fold fabs(x * x) -> x * x; any squared FP value must already be positive.
1161     if (I->getOpcode() == Instruction::FMul)
1162       if (I->getOperand(0) == I->getOperand(1))
1163         return Op;
1164   }
1165   return Ret;
1166 }
1167 
optimizeFMinFMax(CallInst * CI,IRBuilder<> & B)1168 Value *LibCallSimplifier::optimizeFMinFMax(CallInst *CI, IRBuilder<> &B) {
1169   Function *Callee = CI->getCalledFunction();
1170   // If we can shrink the call to a float function rather than a double
1171   // function, do that first.
1172   StringRef Name = Callee->getName();
1173   if ((Name == "fmin" || Name == "fmax") && hasFloatVersion(Name))
1174     if (Value *Ret = optimizeBinaryDoubleFP(CI, B))
1175       return Ret;
1176 
1177   IRBuilder<>::FastMathFlagGuard Guard(B);
1178   FastMathFlags FMF;
1179   if (CI->hasUnsafeAlgebra()) {
1180     // Unsafe algebra sets all fast-math-flags to true.
1181     FMF.setUnsafeAlgebra();
1182   } else {
1183     // At a minimum, no-nans-fp-math must be true.
1184     if (!CI->hasNoNaNs())
1185       return nullptr;
1186     // No-signed-zeros is implied by the definitions of fmax/fmin themselves:
1187     // "Ideally, fmax would be sensitive to the sign of zero, for example
1188     // fmax(-0. 0, +0. 0) would return +0; however, implementation in software
1189     // might be impractical."
1190     FMF.setNoSignedZeros();
1191     FMF.setNoNaNs();
1192   }
1193   B.setFastMathFlags(FMF);
1194 
1195   // We have a relaxed floating-point environment. We can ignore NaN-handling
1196   // and transform to a compare and select. We do not have to consider errno or
1197   // exceptions, because fmin/fmax do not have those.
1198   Value *Op0 = CI->getArgOperand(0);
1199   Value *Op1 = CI->getArgOperand(1);
1200   Value *Cmp = Callee->getName().startswith("fmin") ?
1201     B.CreateFCmpOLT(Op0, Op1) : B.CreateFCmpOGT(Op0, Op1);
1202   return B.CreateSelect(Cmp, Op0, Op1);
1203 }
1204 
optimizeLog(CallInst * CI,IRBuilder<> & B)1205 Value *LibCallSimplifier::optimizeLog(CallInst *CI, IRBuilder<> &B) {
1206   Function *Callee = CI->getCalledFunction();
1207   Value *Ret = nullptr;
1208   StringRef Name = Callee->getName();
1209   if (UnsafeFPShrink && hasFloatVersion(Name))
1210     Ret = optimizeUnaryDoubleFP(CI, B, true);
1211 
1212   if (!CI->hasUnsafeAlgebra())
1213     return Ret;
1214   Value *Op1 = CI->getArgOperand(0);
1215   auto *OpC = dyn_cast<CallInst>(Op1);
1216 
1217   // The earlier call must also be unsafe in order to do these transforms.
1218   if (!OpC || !OpC->hasUnsafeAlgebra())
1219     return Ret;
1220 
1221   // log(pow(x,y)) -> y*log(x)
1222   // This is only applicable to log, log2, log10.
1223   if (Name != "log" && Name != "log2" && Name != "log10")
1224     return Ret;
1225 
1226   IRBuilder<>::FastMathFlagGuard Guard(B);
1227   FastMathFlags FMF;
1228   FMF.setUnsafeAlgebra();
1229   B.setFastMathFlags(FMF);
1230 
1231   LibFunc::Func Func;
1232   Function *F = OpC->getCalledFunction();
1233   if (F && ((TLI->getLibFunc(F->getName(), Func) && TLI->has(Func) &&
1234       Func == LibFunc::pow) || F->getIntrinsicID() == Intrinsic::pow))
1235     return B.CreateFMul(OpC->getArgOperand(1),
1236       emitUnaryFloatFnCall(OpC->getOperand(0), Callee->getName(), B,
1237                            Callee->getAttributes()), "mul");
1238 
1239   // log(exp2(y)) -> y*log(2)
1240   if (F && Name == "log" && TLI->getLibFunc(F->getName(), Func) &&
1241       TLI->has(Func) && Func == LibFunc::exp2)
1242     return B.CreateFMul(
1243         OpC->getArgOperand(0),
1244         emitUnaryFloatFnCall(ConstantFP::get(CI->getType(), 2.0),
1245                              Callee->getName(), B, Callee->getAttributes()),
1246         "logmul");
1247   return Ret;
1248 }
1249 
optimizeSqrt(CallInst * CI,IRBuilder<> & B)1250 Value *LibCallSimplifier::optimizeSqrt(CallInst *CI, IRBuilder<> &B) {
1251   Function *Callee = CI->getCalledFunction();
1252   Value *Ret = nullptr;
1253   if (TLI->has(LibFunc::sqrtf) && (Callee->getName() == "sqrt" ||
1254                                    Callee->getIntrinsicID() == Intrinsic::sqrt))
1255     Ret = optimizeUnaryDoubleFP(CI, B, true);
1256 
1257   if (!CI->hasUnsafeAlgebra())
1258     return Ret;
1259 
1260   Instruction *I = dyn_cast<Instruction>(CI->getArgOperand(0));
1261   if (!I || I->getOpcode() != Instruction::FMul || !I->hasUnsafeAlgebra())
1262     return Ret;
1263 
1264   // We're looking for a repeated factor in a multiplication tree,
1265   // so we can do this fold: sqrt(x * x) -> fabs(x);
1266   // or this fold: sqrt((x * x) * y) -> fabs(x) * sqrt(y).
1267   Value *Op0 = I->getOperand(0);
1268   Value *Op1 = I->getOperand(1);
1269   Value *RepeatOp = nullptr;
1270   Value *OtherOp = nullptr;
1271   if (Op0 == Op1) {
1272     // Simple match: the operands of the multiply are identical.
1273     RepeatOp = Op0;
1274   } else {
1275     // Look for a more complicated pattern: one of the operands is itself
1276     // a multiply, so search for a common factor in that multiply.
1277     // Note: We don't bother looking any deeper than this first level or for
1278     // variations of this pattern because instcombine's visitFMUL and/or the
1279     // reassociation pass should give us this form.
1280     Value *OtherMul0, *OtherMul1;
1281     if (match(Op0, m_FMul(m_Value(OtherMul0), m_Value(OtherMul1)))) {
1282       // Pattern: sqrt((x * y) * z)
1283       if (OtherMul0 == OtherMul1 &&
1284           cast<Instruction>(Op0)->hasUnsafeAlgebra()) {
1285         // Matched: sqrt((x * x) * z)
1286         RepeatOp = OtherMul0;
1287         OtherOp = Op1;
1288       }
1289     }
1290   }
1291   if (!RepeatOp)
1292     return Ret;
1293 
1294   // Fast math flags for any created instructions should match the sqrt
1295   // and multiply.
1296   IRBuilder<>::FastMathFlagGuard Guard(B);
1297   B.setFastMathFlags(I->getFastMathFlags());
1298 
1299   // If we found a repeated factor, hoist it out of the square root and
1300   // replace it with the fabs of that factor.
1301   Module *M = Callee->getParent();
1302   Type *ArgType = I->getType();
1303   Value *Fabs = Intrinsic::getDeclaration(M, Intrinsic::fabs, ArgType);
1304   Value *FabsCall = B.CreateCall(Fabs, RepeatOp, "fabs");
1305   if (OtherOp) {
1306     // If we found a non-repeated factor, we still need to get its square
1307     // root. We then multiply that by the value that was simplified out
1308     // of the square root calculation.
1309     Value *Sqrt = Intrinsic::getDeclaration(M, Intrinsic::sqrt, ArgType);
1310     Value *SqrtCall = B.CreateCall(Sqrt, OtherOp, "sqrt");
1311     return B.CreateFMul(FabsCall, SqrtCall);
1312   }
1313   return FabsCall;
1314 }
1315 
1316 // TODO: Generalize to handle any trig function and its inverse.
optimizeTan(CallInst * CI,IRBuilder<> & B)1317 Value *LibCallSimplifier::optimizeTan(CallInst *CI, IRBuilder<> &B) {
1318   Function *Callee = CI->getCalledFunction();
1319   Value *Ret = nullptr;
1320   StringRef Name = Callee->getName();
1321   if (UnsafeFPShrink && Name == "tan" && hasFloatVersion(Name))
1322     Ret = optimizeUnaryDoubleFP(CI, B, true);
1323 
1324   Value *Op1 = CI->getArgOperand(0);
1325   auto *OpC = dyn_cast<CallInst>(Op1);
1326   if (!OpC)
1327     return Ret;
1328 
1329   // Both calls must allow unsafe optimizations in order to remove them.
1330   if (!CI->hasUnsafeAlgebra() || !OpC->hasUnsafeAlgebra())
1331     return Ret;
1332 
1333   // tan(atan(x)) -> x
1334   // tanf(atanf(x)) -> x
1335   // tanl(atanl(x)) -> x
1336   LibFunc::Func Func;
1337   Function *F = OpC->getCalledFunction();
1338   if (F && TLI->getLibFunc(F->getName(), Func) && TLI->has(Func) &&
1339       ((Func == LibFunc::atan && Callee->getName() == "tan") ||
1340        (Func == LibFunc::atanf && Callee->getName() == "tanf") ||
1341        (Func == LibFunc::atanl && Callee->getName() == "tanl")))
1342     Ret = OpC->getArgOperand(0);
1343   return Ret;
1344 }
1345 
isTrigLibCall(CallInst * CI)1346 static bool isTrigLibCall(CallInst *CI) {
1347   // We can only hope to do anything useful if we can ignore things like errno
1348   // and floating-point exceptions.
1349   // We already checked the prototype.
1350   return CI->hasFnAttr(Attribute::NoUnwind) &&
1351          CI->hasFnAttr(Attribute::ReadNone);
1352 }
1353 
insertSinCosCall(IRBuilder<> & B,Function * OrigCallee,Value * Arg,bool UseFloat,Value * & Sin,Value * & Cos,Value * & SinCos)1354 static void insertSinCosCall(IRBuilder<> &B, Function *OrigCallee, Value *Arg,
1355                              bool UseFloat, Value *&Sin, Value *&Cos,
1356                              Value *&SinCos) {
1357   Type *ArgTy = Arg->getType();
1358   Type *ResTy;
1359   StringRef Name;
1360 
1361   Triple T(OrigCallee->getParent()->getTargetTriple());
1362   if (UseFloat) {
1363     Name = "__sincospif_stret";
1364 
1365     assert(T.getArch() != Triple::x86 && "x86 messy and unsupported for now");
1366     // x86_64 can't use {float, float} since that would be returned in both
1367     // xmm0 and xmm1, which isn't what a real struct would do.
1368     ResTy = T.getArch() == Triple::x86_64
1369     ? static_cast<Type *>(VectorType::get(ArgTy, 2))
1370     : static_cast<Type *>(StructType::get(ArgTy, ArgTy, nullptr));
1371   } else {
1372     Name = "__sincospi_stret";
1373     ResTy = StructType::get(ArgTy, ArgTy, nullptr);
1374   }
1375 
1376   Module *M = OrigCallee->getParent();
1377   Value *Callee = M->getOrInsertFunction(Name, OrigCallee->getAttributes(),
1378                                          ResTy, ArgTy, nullptr);
1379 
1380   if (Instruction *ArgInst = dyn_cast<Instruction>(Arg)) {
1381     // If the argument is an instruction, it must dominate all uses so put our
1382     // sincos call there.
1383     B.SetInsertPoint(ArgInst->getParent(), ++ArgInst->getIterator());
1384   } else {
1385     // Otherwise (e.g. for a constant) the beginning of the function is as
1386     // good a place as any.
1387     BasicBlock &EntryBB = B.GetInsertBlock()->getParent()->getEntryBlock();
1388     B.SetInsertPoint(&EntryBB, EntryBB.begin());
1389   }
1390 
1391   SinCos = B.CreateCall(Callee, Arg, "sincospi");
1392 
1393   if (SinCos->getType()->isStructTy()) {
1394     Sin = B.CreateExtractValue(SinCos, 0, "sinpi");
1395     Cos = B.CreateExtractValue(SinCos, 1, "cospi");
1396   } else {
1397     Sin = B.CreateExtractElement(SinCos, ConstantInt::get(B.getInt32Ty(), 0),
1398                                  "sinpi");
1399     Cos = B.CreateExtractElement(SinCos, ConstantInt::get(B.getInt32Ty(), 1),
1400                                  "cospi");
1401   }
1402 }
1403 
optimizeSinCosPi(CallInst * CI,IRBuilder<> & B)1404 Value *LibCallSimplifier::optimizeSinCosPi(CallInst *CI, IRBuilder<> &B) {
1405   // Make sure the prototype is as expected, otherwise the rest of the
1406   // function is probably invalid and likely to abort.
1407   if (!isTrigLibCall(CI))
1408     return nullptr;
1409 
1410   Value *Arg = CI->getArgOperand(0);
1411   SmallVector<CallInst *, 1> SinCalls;
1412   SmallVector<CallInst *, 1> CosCalls;
1413   SmallVector<CallInst *, 1> SinCosCalls;
1414 
1415   bool IsFloat = Arg->getType()->isFloatTy();
1416 
1417   // Look for all compatible sinpi, cospi and sincospi calls with the same
1418   // argument. If there are enough (in some sense) we can make the
1419   // substitution.
1420   Function *F = CI->getFunction();
1421   for (User *U : Arg->users())
1422     classifyArgUse(U, F, IsFloat, SinCalls, CosCalls, SinCosCalls);
1423 
1424   // It's only worthwhile if both sinpi and cospi are actually used.
1425   if (SinCosCalls.empty() && (SinCalls.empty() || CosCalls.empty()))
1426     return nullptr;
1427 
1428   Value *Sin, *Cos, *SinCos;
1429   insertSinCosCall(B, CI->getCalledFunction(), Arg, IsFloat, Sin, Cos, SinCos);
1430 
1431   replaceTrigInsts(SinCalls, Sin);
1432   replaceTrigInsts(CosCalls, Cos);
1433   replaceTrigInsts(SinCosCalls, SinCos);
1434 
1435   return nullptr;
1436 }
1437 
classifyArgUse(Value * Val,Function * F,bool IsFloat,SmallVectorImpl<CallInst * > & SinCalls,SmallVectorImpl<CallInst * > & CosCalls,SmallVectorImpl<CallInst * > & SinCosCalls)1438 void LibCallSimplifier::classifyArgUse(
1439     Value *Val, Function *F, bool IsFloat,
1440     SmallVectorImpl<CallInst *> &SinCalls,
1441     SmallVectorImpl<CallInst *> &CosCalls,
1442     SmallVectorImpl<CallInst *> &SinCosCalls) {
1443   CallInst *CI = dyn_cast<CallInst>(Val);
1444 
1445   if (!CI)
1446     return;
1447 
1448   // Don't consider calls in other functions.
1449   if (CI->getFunction() != F)
1450     return;
1451 
1452   Function *Callee = CI->getCalledFunction();
1453   LibFunc::Func Func;
1454   if (!Callee || !TLI->getLibFunc(*Callee, Func) || !TLI->has(Func) ||
1455       !isTrigLibCall(CI))
1456     return;
1457 
1458   if (IsFloat) {
1459     if (Func == LibFunc::sinpif)
1460       SinCalls.push_back(CI);
1461     else if (Func == LibFunc::cospif)
1462       CosCalls.push_back(CI);
1463     else if (Func == LibFunc::sincospif_stret)
1464       SinCosCalls.push_back(CI);
1465   } else {
1466     if (Func == LibFunc::sinpi)
1467       SinCalls.push_back(CI);
1468     else if (Func == LibFunc::cospi)
1469       CosCalls.push_back(CI);
1470     else if (Func == LibFunc::sincospi_stret)
1471       SinCosCalls.push_back(CI);
1472   }
1473 }
1474 
replaceTrigInsts(SmallVectorImpl<CallInst * > & Calls,Value * Res)1475 void LibCallSimplifier::replaceTrigInsts(SmallVectorImpl<CallInst *> &Calls,
1476                                          Value *Res) {
1477   for (CallInst *C : Calls)
1478     replaceAllUsesWith(C, Res);
1479 }
1480 
1481 //===----------------------------------------------------------------------===//
1482 // Integer Library Call Optimizations
1483 //===----------------------------------------------------------------------===//
1484 
optimizeFFS(CallInst * CI,IRBuilder<> & B)1485 Value *LibCallSimplifier::optimizeFFS(CallInst *CI, IRBuilder<> &B) {
1486   Function *Callee = CI->getCalledFunction();
1487   Value *Op = CI->getArgOperand(0);
1488 
1489   // Constant fold.
1490   if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
1491     if (CI->isZero()) // ffs(0) -> 0.
1492       return B.getInt32(0);
1493     // ffs(c) -> cttz(c)+1
1494     return B.getInt32(CI->getValue().countTrailingZeros() + 1);
1495   }
1496 
1497   // ffs(x) -> x != 0 ? (i32)llvm.cttz(x)+1 : 0
1498   Type *ArgType = Op->getType();
1499   Value *F =
1500       Intrinsic::getDeclaration(Callee->getParent(), Intrinsic::cttz, ArgType);
1501   Value *V = B.CreateCall(F, {Op, B.getTrue()}, "cttz");
1502   V = B.CreateAdd(V, ConstantInt::get(V->getType(), 1));
1503   V = B.CreateIntCast(V, B.getInt32Ty(), false);
1504 
1505   Value *Cond = B.CreateICmpNE(Op, Constant::getNullValue(ArgType));
1506   return B.CreateSelect(Cond, V, B.getInt32(0));
1507 }
1508 
optimizeAbs(CallInst * CI,IRBuilder<> & B)1509 Value *LibCallSimplifier::optimizeAbs(CallInst *CI, IRBuilder<> &B) {
1510   // abs(x) -> x >s -1 ? x : -x
1511   Value *Op = CI->getArgOperand(0);
1512   Value *Pos =
1513       B.CreateICmpSGT(Op, Constant::getAllOnesValue(Op->getType()), "ispos");
1514   Value *Neg = B.CreateNeg(Op, "neg");
1515   return B.CreateSelect(Pos, Op, Neg);
1516 }
1517 
optimizeIsDigit(CallInst * CI,IRBuilder<> & B)1518 Value *LibCallSimplifier::optimizeIsDigit(CallInst *CI, IRBuilder<> &B) {
1519   // isdigit(c) -> (c-'0') <u 10
1520   Value *Op = CI->getArgOperand(0);
1521   Op = B.CreateSub(Op, B.getInt32('0'), "isdigittmp");
1522   Op = B.CreateICmpULT(Op, B.getInt32(10), "isdigit");
1523   return B.CreateZExt(Op, CI->getType());
1524 }
1525 
optimizeIsAscii(CallInst * CI,IRBuilder<> & B)1526 Value *LibCallSimplifier::optimizeIsAscii(CallInst *CI, IRBuilder<> &B) {
1527   // isascii(c) -> c <u 128
1528   Value *Op = CI->getArgOperand(0);
1529   Op = B.CreateICmpULT(Op, B.getInt32(128), "isascii");
1530   return B.CreateZExt(Op, CI->getType());
1531 }
1532 
optimizeToAscii(CallInst * CI,IRBuilder<> & B)1533 Value *LibCallSimplifier::optimizeToAscii(CallInst *CI, IRBuilder<> &B) {
1534   // toascii(c) -> c & 0x7f
1535   return B.CreateAnd(CI->getArgOperand(0),
1536                      ConstantInt::get(CI->getType(), 0x7F));
1537 }
1538 
1539 //===----------------------------------------------------------------------===//
1540 // Formatting and IO Library Call Optimizations
1541 //===----------------------------------------------------------------------===//
1542 
1543 static bool isReportingError(Function *Callee, CallInst *CI, int StreamArg);
1544 
optimizeErrorReporting(CallInst * CI,IRBuilder<> & B,int StreamArg)1545 Value *LibCallSimplifier::optimizeErrorReporting(CallInst *CI, IRBuilder<> &B,
1546                                                  int StreamArg) {
1547   Function *Callee = CI->getCalledFunction();
1548   // Error reporting calls should be cold, mark them as such.
1549   // This applies even to non-builtin calls: it is only a hint and applies to
1550   // functions that the frontend might not understand as builtins.
1551 
1552   // This heuristic was suggested in:
1553   // Improving Static Branch Prediction in a Compiler
1554   // Brian L. Deitrich, Ben-Chung Cheng, Wen-mei W. Hwu
1555   // Proceedings of PACT'98, Oct. 1998, IEEE
1556   if (!CI->hasFnAttr(Attribute::Cold) &&
1557       isReportingError(Callee, CI, StreamArg)) {
1558     CI->addAttribute(AttributeSet::FunctionIndex, Attribute::Cold);
1559   }
1560 
1561   return nullptr;
1562 }
1563 
isReportingError(Function * Callee,CallInst * CI,int StreamArg)1564 static bool isReportingError(Function *Callee, CallInst *CI, int StreamArg) {
1565   if (!ColdErrorCalls || !Callee || !Callee->isDeclaration())
1566     return false;
1567 
1568   if (StreamArg < 0)
1569     return true;
1570 
1571   // These functions might be considered cold, but only if their stream
1572   // argument is stderr.
1573 
1574   if (StreamArg >= (int)CI->getNumArgOperands())
1575     return false;
1576   LoadInst *LI = dyn_cast<LoadInst>(CI->getArgOperand(StreamArg));
1577   if (!LI)
1578     return false;
1579   GlobalVariable *GV = dyn_cast<GlobalVariable>(LI->getPointerOperand());
1580   if (!GV || !GV->isDeclaration())
1581     return false;
1582   return GV->getName() == "stderr";
1583 }
1584 
optimizePrintFString(CallInst * CI,IRBuilder<> & B)1585 Value *LibCallSimplifier::optimizePrintFString(CallInst *CI, IRBuilder<> &B) {
1586   // Check for a fixed format string.
1587   StringRef FormatStr;
1588   if (!getConstantStringInfo(CI->getArgOperand(0), FormatStr))
1589     return nullptr;
1590 
1591   // Empty format string -> noop.
1592   if (FormatStr.empty()) // Tolerate printf's declared void.
1593     return CI->use_empty() ? (Value *)CI : ConstantInt::get(CI->getType(), 0);
1594 
1595   // Do not do any of the following transformations if the printf return value
1596   // is used, in general the printf return value is not compatible with either
1597   // putchar() or puts().
1598   if (!CI->use_empty())
1599     return nullptr;
1600 
1601   // printf("x") -> putchar('x'), even for "%" and "%%".
1602   if (FormatStr.size() == 1 || FormatStr == "%%")
1603     return emitPutChar(B.getInt32(FormatStr[0]), B, TLI);
1604 
1605   // printf("%s", "a") --> putchar('a')
1606   if (FormatStr == "%s" && CI->getNumArgOperands() > 1) {
1607     StringRef ChrStr;
1608     if (!getConstantStringInfo(CI->getOperand(1), ChrStr))
1609       return nullptr;
1610     if (ChrStr.size() != 1)
1611       return nullptr;
1612     return emitPutChar(B.getInt32(ChrStr[0]), B, TLI);
1613   }
1614 
1615   // printf("foo\n") --> puts("foo")
1616   if (FormatStr[FormatStr.size() - 1] == '\n' &&
1617       FormatStr.find('%') == StringRef::npos) { // No format characters.
1618     // Create a string literal with no \n on it.  We expect the constant merge
1619     // pass to be run after this pass, to merge duplicate strings.
1620     FormatStr = FormatStr.drop_back();
1621     Value *GV = B.CreateGlobalString(FormatStr, "str");
1622     return emitPutS(GV, B, TLI);
1623   }
1624 
1625   // Optimize specific format strings.
1626   // printf("%c", chr) --> putchar(chr)
1627   if (FormatStr == "%c" && CI->getNumArgOperands() > 1 &&
1628       CI->getArgOperand(1)->getType()->isIntegerTy())
1629     return emitPutChar(CI->getArgOperand(1), B, TLI);
1630 
1631   // printf("%s\n", str) --> puts(str)
1632   if (FormatStr == "%s\n" && CI->getNumArgOperands() > 1 &&
1633       CI->getArgOperand(1)->getType()->isPointerTy())
1634     return emitPutS(CI->getArgOperand(1), B, TLI);
1635   return nullptr;
1636 }
1637 
optimizePrintF(CallInst * CI,IRBuilder<> & B)1638 Value *LibCallSimplifier::optimizePrintF(CallInst *CI, IRBuilder<> &B) {
1639 
1640   Function *Callee = CI->getCalledFunction();
1641   FunctionType *FT = Callee->getFunctionType();
1642   if (Value *V = optimizePrintFString(CI, B)) {
1643     return V;
1644   }
1645 
1646   // printf(format, ...) -> iprintf(format, ...) if no floating point
1647   // arguments.
1648   if (TLI->has(LibFunc::iprintf) && !callHasFloatingPointArgument(CI)) {
1649     Module *M = B.GetInsertBlock()->getParent()->getParent();
1650     Constant *IPrintFFn =
1651         M->getOrInsertFunction("iprintf", FT, Callee->getAttributes());
1652     CallInst *New = cast<CallInst>(CI->clone());
1653     New->setCalledFunction(IPrintFFn);
1654     B.Insert(New);
1655     return New;
1656   }
1657   return nullptr;
1658 }
1659 
optimizeSPrintFString(CallInst * CI,IRBuilder<> & B)1660 Value *LibCallSimplifier::optimizeSPrintFString(CallInst *CI, IRBuilder<> &B) {
1661   // Check for a fixed format string.
1662   StringRef FormatStr;
1663   if (!getConstantStringInfo(CI->getArgOperand(1), FormatStr))
1664     return nullptr;
1665 
1666   // If we just have a format string (nothing else crazy) transform it.
1667   if (CI->getNumArgOperands() == 2) {
1668     // Make sure there's no % in the constant array.  We could try to handle
1669     // %% -> % in the future if we cared.
1670     for (unsigned i = 0, e = FormatStr.size(); i != e; ++i)
1671       if (FormatStr[i] == '%')
1672         return nullptr; // we found a format specifier, bail out.
1673 
1674     // sprintf(str, fmt) -> llvm.memcpy(str, fmt, strlen(fmt)+1, 1)
1675     B.CreateMemCpy(CI->getArgOperand(0), CI->getArgOperand(1),
1676                    ConstantInt::get(DL.getIntPtrType(CI->getContext()),
1677                                     FormatStr.size() + 1),
1678                    1); // Copy the null byte.
1679     return ConstantInt::get(CI->getType(), FormatStr.size());
1680   }
1681 
1682   // The remaining optimizations require the format string to be "%s" or "%c"
1683   // and have an extra operand.
1684   if (FormatStr.size() != 2 || FormatStr[0] != '%' ||
1685       CI->getNumArgOperands() < 3)
1686     return nullptr;
1687 
1688   // Decode the second character of the format string.
1689   if (FormatStr[1] == 'c') {
1690     // sprintf(dst, "%c", chr) --> *(i8*)dst = chr; *((i8*)dst+1) = 0
1691     if (!CI->getArgOperand(2)->getType()->isIntegerTy())
1692       return nullptr;
1693     Value *V = B.CreateTrunc(CI->getArgOperand(2), B.getInt8Ty(), "char");
1694     Value *Ptr = castToCStr(CI->getArgOperand(0), B);
1695     B.CreateStore(V, Ptr);
1696     Ptr = B.CreateGEP(B.getInt8Ty(), Ptr, B.getInt32(1), "nul");
1697     B.CreateStore(B.getInt8(0), Ptr);
1698 
1699     return ConstantInt::get(CI->getType(), 1);
1700   }
1701 
1702   if (FormatStr[1] == 's') {
1703     // sprintf(dest, "%s", str) -> llvm.memcpy(dest, str, strlen(str)+1, 1)
1704     if (!CI->getArgOperand(2)->getType()->isPointerTy())
1705       return nullptr;
1706 
1707     Value *Len = emitStrLen(CI->getArgOperand(2), B, DL, TLI);
1708     if (!Len)
1709       return nullptr;
1710     Value *IncLen =
1711         B.CreateAdd(Len, ConstantInt::get(Len->getType(), 1), "leninc");
1712     B.CreateMemCpy(CI->getArgOperand(0), CI->getArgOperand(2), IncLen, 1);
1713 
1714     // The sprintf result is the unincremented number of bytes in the string.
1715     return B.CreateIntCast(Len, CI->getType(), false);
1716   }
1717   return nullptr;
1718 }
1719 
optimizeSPrintF(CallInst * CI,IRBuilder<> & B)1720 Value *LibCallSimplifier::optimizeSPrintF(CallInst *CI, IRBuilder<> &B) {
1721   Function *Callee = CI->getCalledFunction();
1722   FunctionType *FT = Callee->getFunctionType();
1723   if (Value *V = optimizeSPrintFString(CI, B)) {
1724     return V;
1725   }
1726 
1727   // sprintf(str, format, ...) -> siprintf(str, format, ...) if no floating
1728   // point arguments.
1729   if (TLI->has(LibFunc::siprintf) && !callHasFloatingPointArgument(CI)) {
1730     Module *M = B.GetInsertBlock()->getParent()->getParent();
1731     Constant *SIPrintFFn =
1732         M->getOrInsertFunction("siprintf", FT, Callee->getAttributes());
1733     CallInst *New = cast<CallInst>(CI->clone());
1734     New->setCalledFunction(SIPrintFFn);
1735     B.Insert(New);
1736     return New;
1737   }
1738   return nullptr;
1739 }
1740 
optimizeFPrintFString(CallInst * CI,IRBuilder<> & B)1741 Value *LibCallSimplifier::optimizeFPrintFString(CallInst *CI, IRBuilder<> &B) {
1742   optimizeErrorReporting(CI, B, 0);
1743 
1744   // All the optimizations depend on the format string.
1745   StringRef FormatStr;
1746   if (!getConstantStringInfo(CI->getArgOperand(1), FormatStr))
1747     return nullptr;
1748 
1749   // Do not do any of the following transformations if the fprintf return
1750   // value is used, in general the fprintf return value is not compatible
1751   // with fwrite(), fputc() or fputs().
1752   if (!CI->use_empty())
1753     return nullptr;
1754 
1755   // fprintf(F, "foo") --> fwrite("foo", 3, 1, F)
1756   if (CI->getNumArgOperands() == 2) {
1757     for (unsigned i = 0, e = FormatStr.size(); i != e; ++i)
1758       if (FormatStr[i] == '%') // Could handle %% -> % if we cared.
1759         return nullptr;        // We found a format specifier.
1760 
1761     return emitFWrite(
1762         CI->getArgOperand(1),
1763         ConstantInt::get(DL.getIntPtrType(CI->getContext()), FormatStr.size()),
1764         CI->getArgOperand(0), B, DL, TLI);
1765   }
1766 
1767   // The remaining optimizations require the format string to be "%s" or "%c"
1768   // and have an extra operand.
1769   if (FormatStr.size() != 2 || FormatStr[0] != '%' ||
1770       CI->getNumArgOperands() < 3)
1771     return nullptr;
1772 
1773   // Decode the second character of the format string.
1774   if (FormatStr[1] == 'c') {
1775     // fprintf(F, "%c", chr) --> fputc(chr, F)
1776     if (!CI->getArgOperand(2)->getType()->isIntegerTy())
1777       return nullptr;
1778     return emitFPutC(CI->getArgOperand(2), CI->getArgOperand(0), B, TLI);
1779   }
1780 
1781   if (FormatStr[1] == 's') {
1782     // fprintf(F, "%s", str) --> fputs(str, F)
1783     if (!CI->getArgOperand(2)->getType()->isPointerTy())
1784       return nullptr;
1785     return emitFPutS(CI->getArgOperand(2), CI->getArgOperand(0), B, TLI);
1786   }
1787   return nullptr;
1788 }
1789 
optimizeFPrintF(CallInst * CI,IRBuilder<> & B)1790 Value *LibCallSimplifier::optimizeFPrintF(CallInst *CI, IRBuilder<> &B) {
1791   Function *Callee = CI->getCalledFunction();
1792   FunctionType *FT = Callee->getFunctionType();
1793   if (Value *V = optimizeFPrintFString(CI, B)) {
1794     return V;
1795   }
1796 
1797   // fprintf(stream, format, ...) -> fiprintf(stream, format, ...) if no
1798   // floating point arguments.
1799   if (TLI->has(LibFunc::fiprintf) && !callHasFloatingPointArgument(CI)) {
1800     Module *M = B.GetInsertBlock()->getParent()->getParent();
1801     Constant *FIPrintFFn =
1802         M->getOrInsertFunction("fiprintf", FT, Callee->getAttributes());
1803     CallInst *New = cast<CallInst>(CI->clone());
1804     New->setCalledFunction(FIPrintFFn);
1805     B.Insert(New);
1806     return New;
1807   }
1808   return nullptr;
1809 }
1810 
optimizeFWrite(CallInst * CI,IRBuilder<> & B)1811 Value *LibCallSimplifier::optimizeFWrite(CallInst *CI, IRBuilder<> &B) {
1812   optimizeErrorReporting(CI, B, 3);
1813 
1814   // Get the element size and count.
1815   ConstantInt *SizeC = dyn_cast<ConstantInt>(CI->getArgOperand(1));
1816   ConstantInt *CountC = dyn_cast<ConstantInt>(CI->getArgOperand(2));
1817   if (!SizeC || !CountC)
1818     return nullptr;
1819   uint64_t Bytes = SizeC->getZExtValue() * CountC->getZExtValue();
1820 
1821   // If this is writing zero records, remove the call (it's a noop).
1822   if (Bytes == 0)
1823     return ConstantInt::get(CI->getType(), 0);
1824 
1825   // If this is writing one byte, turn it into fputc.
1826   // This optimisation is only valid, if the return value is unused.
1827   if (Bytes == 1 && CI->use_empty()) { // fwrite(S,1,1,F) -> fputc(S[0],F)
1828     Value *Char = B.CreateLoad(castToCStr(CI->getArgOperand(0), B), "char");
1829     Value *NewCI = emitFPutC(Char, CI->getArgOperand(3), B, TLI);
1830     return NewCI ? ConstantInt::get(CI->getType(), 1) : nullptr;
1831   }
1832 
1833   return nullptr;
1834 }
1835 
optimizeFPuts(CallInst * CI,IRBuilder<> & B)1836 Value *LibCallSimplifier::optimizeFPuts(CallInst *CI, IRBuilder<> &B) {
1837   optimizeErrorReporting(CI, B, 1);
1838 
1839   // Don't rewrite fputs to fwrite when optimising for size because fwrite
1840   // requires more arguments and thus extra MOVs are required.
1841   if (CI->getParent()->getParent()->optForSize())
1842     return nullptr;
1843 
1844   // We can't optimize if return value is used.
1845   if (!CI->use_empty())
1846     return nullptr;
1847 
1848   // fputs(s,F) --> fwrite(s,1,strlen(s),F)
1849   uint64_t Len = GetStringLength(CI->getArgOperand(0));
1850   if (!Len)
1851     return nullptr;
1852 
1853   // Known to have no uses (see above).
1854   return emitFWrite(
1855       CI->getArgOperand(0),
1856       ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len - 1),
1857       CI->getArgOperand(1), B, DL, TLI);
1858 }
1859 
optimizePuts(CallInst * CI,IRBuilder<> & B)1860 Value *LibCallSimplifier::optimizePuts(CallInst *CI, IRBuilder<> &B) {
1861   // Check for a constant string.
1862   StringRef Str;
1863   if (!getConstantStringInfo(CI->getArgOperand(0), Str))
1864     return nullptr;
1865 
1866   if (Str.empty() && CI->use_empty()) {
1867     // puts("") -> putchar('\n')
1868     Value *Res = emitPutChar(B.getInt32('\n'), B, TLI);
1869     if (CI->use_empty() || !Res)
1870       return Res;
1871     return B.CreateIntCast(Res, CI->getType(), true);
1872   }
1873 
1874   return nullptr;
1875 }
1876 
hasFloatVersion(StringRef FuncName)1877 bool LibCallSimplifier::hasFloatVersion(StringRef FuncName) {
1878   LibFunc::Func Func;
1879   SmallString<20> FloatFuncName = FuncName;
1880   FloatFuncName += 'f';
1881   if (TLI->getLibFunc(FloatFuncName, Func))
1882     return TLI->has(Func);
1883   return false;
1884 }
1885 
optimizeStringMemoryLibCall(CallInst * CI,IRBuilder<> & Builder)1886 Value *LibCallSimplifier::optimizeStringMemoryLibCall(CallInst *CI,
1887                                                       IRBuilder<> &Builder) {
1888   LibFunc::Func Func;
1889   Function *Callee = CI->getCalledFunction();
1890   // Check for string/memory library functions.
1891   if (TLI->getLibFunc(*Callee, Func) && TLI->has(Func)) {
1892     // Make sure we never change the calling convention.
1893     assert((ignoreCallingConv(Func) ||
1894             CI->getCallingConv() == llvm::CallingConv::C) &&
1895       "Optimizing string/memory libcall would change the calling convention");
1896     switch (Func) {
1897     case LibFunc::strcat:
1898       return optimizeStrCat(CI, Builder);
1899     case LibFunc::strncat:
1900       return optimizeStrNCat(CI, Builder);
1901     case LibFunc::strchr:
1902       return optimizeStrChr(CI, Builder);
1903     case LibFunc::strrchr:
1904       return optimizeStrRChr(CI, Builder);
1905     case LibFunc::strcmp:
1906       return optimizeStrCmp(CI, Builder);
1907     case LibFunc::strncmp:
1908       return optimizeStrNCmp(CI, Builder);
1909     case LibFunc::strcpy:
1910       return optimizeStrCpy(CI, Builder);
1911     case LibFunc::stpcpy:
1912       return optimizeStpCpy(CI, Builder);
1913     case LibFunc::strncpy:
1914       return optimizeStrNCpy(CI, Builder);
1915     case LibFunc::strlen:
1916       return optimizeStrLen(CI, Builder);
1917     case LibFunc::strpbrk:
1918       return optimizeStrPBrk(CI, Builder);
1919     case LibFunc::strtol:
1920     case LibFunc::strtod:
1921     case LibFunc::strtof:
1922     case LibFunc::strtoul:
1923     case LibFunc::strtoll:
1924     case LibFunc::strtold:
1925     case LibFunc::strtoull:
1926       return optimizeStrTo(CI, Builder);
1927     case LibFunc::strspn:
1928       return optimizeStrSpn(CI, Builder);
1929     case LibFunc::strcspn:
1930       return optimizeStrCSpn(CI, Builder);
1931     case LibFunc::strstr:
1932       return optimizeStrStr(CI, Builder);
1933     case LibFunc::memchr:
1934       return optimizeMemChr(CI, Builder);
1935     case LibFunc::memcmp:
1936       return optimizeMemCmp(CI, Builder);
1937     case LibFunc::memcpy:
1938       return optimizeMemCpy(CI, Builder);
1939     case LibFunc::memmove:
1940       return optimizeMemMove(CI, Builder);
1941     case LibFunc::memset:
1942       return optimizeMemSet(CI, Builder);
1943     default:
1944       break;
1945     }
1946   }
1947   return nullptr;
1948 }
1949 
optimizeCall(CallInst * CI)1950 Value *LibCallSimplifier::optimizeCall(CallInst *CI) {
1951   if (CI->isNoBuiltin())
1952     return nullptr;
1953 
1954   LibFunc::Func Func;
1955   Function *Callee = CI->getCalledFunction();
1956   StringRef FuncName = Callee->getName();
1957 
1958   SmallVector<OperandBundleDef, 2> OpBundles;
1959   CI->getOperandBundlesAsDefs(OpBundles);
1960   IRBuilder<> Builder(CI, /*FPMathTag=*/nullptr, OpBundles);
1961   bool isCallingConvC = CI->getCallingConv() == llvm::CallingConv::C;
1962 
1963   // Command-line parameter overrides instruction attribute.
1964   if (EnableUnsafeFPShrink.getNumOccurrences() > 0)
1965     UnsafeFPShrink = EnableUnsafeFPShrink;
1966   else if (isa<FPMathOperator>(CI) && CI->hasUnsafeAlgebra())
1967     UnsafeFPShrink = true;
1968 
1969   // First, check for intrinsics.
1970   if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
1971     if (!isCallingConvC)
1972       return nullptr;
1973     switch (II->getIntrinsicID()) {
1974     case Intrinsic::pow:
1975       return optimizePow(CI, Builder);
1976     case Intrinsic::exp2:
1977       return optimizeExp2(CI, Builder);
1978     case Intrinsic::fabs:
1979       return optimizeFabs(CI, Builder);
1980     case Intrinsic::log:
1981       return optimizeLog(CI, Builder);
1982     case Intrinsic::sqrt:
1983       return optimizeSqrt(CI, Builder);
1984     // TODO: Use foldMallocMemset() with memset intrinsic.
1985     default:
1986       return nullptr;
1987     }
1988   }
1989 
1990   // Also try to simplify calls to fortified library functions.
1991   if (Value *SimplifiedFortifiedCI = FortifiedSimplifier.optimizeCall(CI)) {
1992     // Try to further simplify the result.
1993     CallInst *SimplifiedCI = dyn_cast<CallInst>(SimplifiedFortifiedCI);
1994     if (SimplifiedCI && SimplifiedCI->getCalledFunction()) {
1995       // Use an IR Builder from SimplifiedCI if available instead of CI
1996       // to guarantee we reach all uses we might replace later on.
1997       IRBuilder<> TmpBuilder(SimplifiedCI);
1998       if (Value *V = optimizeStringMemoryLibCall(SimplifiedCI, TmpBuilder)) {
1999         // If we were able to further simplify, remove the now redundant call.
2000         SimplifiedCI->replaceAllUsesWith(V);
2001         SimplifiedCI->eraseFromParent();
2002         return V;
2003       }
2004     }
2005     return SimplifiedFortifiedCI;
2006   }
2007 
2008   // Then check for known library functions.
2009   if (TLI->getLibFunc(*Callee, Func) && TLI->has(Func)) {
2010     // We never change the calling convention.
2011     if (!ignoreCallingConv(Func) && !isCallingConvC)
2012       return nullptr;
2013     if (Value *V = optimizeStringMemoryLibCall(CI, Builder))
2014       return V;
2015     switch (Func) {
2016     case LibFunc::cosf:
2017     case LibFunc::cos:
2018     case LibFunc::cosl:
2019       return optimizeCos(CI, Builder);
2020     case LibFunc::sinpif:
2021     case LibFunc::sinpi:
2022     case LibFunc::cospif:
2023     case LibFunc::cospi:
2024       return optimizeSinCosPi(CI, Builder);
2025     case LibFunc::powf:
2026     case LibFunc::pow:
2027     case LibFunc::powl:
2028       return optimizePow(CI, Builder);
2029     case LibFunc::exp2l:
2030     case LibFunc::exp2:
2031     case LibFunc::exp2f:
2032       return optimizeExp2(CI, Builder);
2033     case LibFunc::fabsf:
2034     case LibFunc::fabs:
2035     case LibFunc::fabsl:
2036       return optimizeFabs(CI, Builder);
2037     case LibFunc::sqrtf:
2038     case LibFunc::sqrt:
2039     case LibFunc::sqrtl:
2040       return optimizeSqrt(CI, Builder);
2041     case LibFunc::ffs:
2042     case LibFunc::ffsl:
2043     case LibFunc::ffsll:
2044       return optimizeFFS(CI, Builder);
2045     case LibFunc::abs:
2046     case LibFunc::labs:
2047     case LibFunc::llabs:
2048       return optimizeAbs(CI, Builder);
2049     case LibFunc::isdigit:
2050       return optimizeIsDigit(CI, Builder);
2051     case LibFunc::isascii:
2052       return optimizeIsAscii(CI, Builder);
2053     case LibFunc::toascii:
2054       return optimizeToAscii(CI, Builder);
2055     case LibFunc::printf:
2056       return optimizePrintF(CI, Builder);
2057     case LibFunc::sprintf:
2058       return optimizeSPrintF(CI, Builder);
2059     case LibFunc::fprintf:
2060       return optimizeFPrintF(CI, Builder);
2061     case LibFunc::fwrite:
2062       return optimizeFWrite(CI, Builder);
2063     case LibFunc::fputs:
2064       return optimizeFPuts(CI, Builder);
2065     case LibFunc::log:
2066     case LibFunc::log10:
2067     case LibFunc::log1p:
2068     case LibFunc::log2:
2069     case LibFunc::logb:
2070       return optimizeLog(CI, Builder);
2071     case LibFunc::puts:
2072       return optimizePuts(CI, Builder);
2073     case LibFunc::tan:
2074     case LibFunc::tanf:
2075     case LibFunc::tanl:
2076       return optimizeTan(CI, Builder);
2077     case LibFunc::perror:
2078       return optimizeErrorReporting(CI, Builder);
2079     case LibFunc::vfprintf:
2080     case LibFunc::fiprintf:
2081       return optimizeErrorReporting(CI, Builder, 0);
2082     case LibFunc::fputc:
2083       return optimizeErrorReporting(CI, Builder, 1);
2084     case LibFunc::ceil:
2085     case LibFunc::floor:
2086     case LibFunc::rint:
2087     case LibFunc::round:
2088     case LibFunc::nearbyint:
2089     case LibFunc::trunc:
2090       if (hasFloatVersion(FuncName))
2091         return optimizeUnaryDoubleFP(CI, Builder, false);
2092       return nullptr;
2093     case LibFunc::acos:
2094     case LibFunc::acosh:
2095     case LibFunc::asin:
2096     case LibFunc::asinh:
2097     case LibFunc::atan:
2098     case LibFunc::atanh:
2099     case LibFunc::cbrt:
2100     case LibFunc::cosh:
2101     case LibFunc::exp:
2102     case LibFunc::exp10:
2103     case LibFunc::expm1:
2104     case LibFunc::sin:
2105     case LibFunc::sinh:
2106     case LibFunc::tanh:
2107       if (UnsafeFPShrink && hasFloatVersion(FuncName))
2108         return optimizeUnaryDoubleFP(CI, Builder, true);
2109       return nullptr;
2110     case LibFunc::copysign:
2111       if (hasFloatVersion(FuncName))
2112         return optimizeBinaryDoubleFP(CI, Builder);
2113       return nullptr;
2114     case LibFunc::fminf:
2115     case LibFunc::fmin:
2116     case LibFunc::fminl:
2117     case LibFunc::fmaxf:
2118     case LibFunc::fmax:
2119     case LibFunc::fmaxl:
2120       return optimizeFMinFMax(CI, Builder);
2121     default:
2122       return nullptr;
2123     }
2124   }
2125   return nullptr;
2126 }
2127 
LibCallSimplifier(const DataLayout & DL,const TargetLibraryInfo * TLI,function_ref<void (Instruction *,Value *)> Replacer)2128 LibCallSimplifier::LibCallSimplifier(
2129     const DataLayout &DL, const TargetLibraryInfo *TLI,
2130     function_ref<void(Instruction *, Value *)> Replacer)
2131     : FortifiedSimplifier(TLI), DL(DL), TLI(TLI), UnsafeFPShrink(false),
2132       Replacer(Replacer) {}
2133 
replaceAllUsesWith(Instruction * I,Value * With)2134 void LibCallSimplifier::replaceAllUsesWith(Instruction *I, Value *With) {
2135   // Indirect through the replacer used in this instance.
2136   Replacer(I, With);
2137 }
2138 
2139 // TODO:
2140 //   Additional cases that we need to add to this file:
2141 //
2142 // cbrt:
2143 //   * cbrt(expN(X))  -> expN(x/3)
2144 //   * cbrt(sqrt(x))  -> pow(x,1/6)
2145 //   * cbrt(cbrt(x))  -> pow(x,1/9)
2146 //
2147 // exp, expf, expl:
2148 //   * exp(log(x))  -> x
2149 //
2150 // log, logf, logl:
2151 //   * log(exp(x))   -> x
2152 //   * log(exp(y))   -> y*log(e)
2153 //   * log(exp10(y)) -> y*log(10)
2154 //   * log(sqrt(x))  -> 0.5*log(x)
2155 //
2156 // lround, lroundf, lroundl:
2157 //   * lround(cnst) -> cnst'
2158 //
2159 // pow, powf, powl:
2160 //   * pow(sqrt(x),y) -> pow(x,y*0.5)
2161 //   * pow(pow(x,y),z)-> pow(x,y*z)
2162 //
2163 // round, roundf, roundl:
2164 //   * round(cnst) -> cnst'
2165 //
2166 // signbit:
2167 //   * signbit(cnst) -> cnst'
2168 //   * signbit(nncst) -> 0 (if pstv is a non-negative constant)
2169 //
2170 // sqrt, sqrtf, sqrtl:
2171 //   * sqrt(expN(x))  -> expN(x*0.5)
2172 //   * sqrt(Nroot(x)) -> pow(x,1/(2*N))
2173 //   * sqrt(pow(x,y)) -> pow(|x|,y*0.5)
2174 //
2175 // trunc, truncf, truncl:
2176 //   * trunc(cnst) -> cnst'
2177 //
2178 //
2179 
2180 //===----------------------------------------------------------------------===//
2181 // Fortified Library Call Optimizations
2182 //===----------------------------------------------------------------------===//
2183 
isFortifiedCallFoldable(CallInst * CI,unsigned ObjSizeOp,unsigned SizeOp,bool isString)2184 bool FortifiedLibCallSimplifier::isFortifiedCallFoldable(CallInst *CI,
2185                                                          unsigned ObjSizeOp,
2186                                                          unsigned SizeOp,
2187                                                          bool isString) {
2188   if (CI->getArgOperand(ObjSizeOp) == CI->getArgOperand(SizeOp))
2189     return true;
2190   if (ConstantInt *ObjSizeCI =
2191           dyn_cast<ConstantInt>(CI->getArgOperand(ObjSizeOp))) {
2192     if (ObjSizeCI->isAllOnesValue())
2193       return true;
2194     // If the object size wasn't -1 (unknown), bail out if we were asked to.
2195     if (OnlyLowerUnknownSize)
2196       return false;
2197     if (isString) {
2198       uint64_t Len = GetStringLength(CI->getArgOperand(SizeOp));
2199       // If the length is 0 we don't know how long it is and so we can't
2200       // remove the check.
2201       if (Len == 0)
2202         return false;
2203       return ObjSizeCI->getZExtValue() >= Len;
2204     }
2205     if (ConstantInt *SizeCI = dyn_cast<ConstantInt>(CI->getArgOperand(SizeOp)))
2206       return ObjSizeCI->getZExtValue() >= SizeCI->getZExtValue();
2207   }
2208   return false;
2209 }
2210 
optimizeMemCpyChk(CallInst * CI,IRBuilder<> & B)2211 Value *FortifiedLibCallSimplifier::optimizeMemCpyChk(CallInst *CI,
2212                                                      IRBuilder<> &B) {
2213   if (isFortifiedCallFoldable(CI, 3, 2, false)) {
2214     B.CreateMemCpy(CI->getArgOperand(0), CI->getArgOperand(1),
2215                    CI->getArgOperand(2), 1);
2216     return CI->getArgOperand(0);
2217   }
2218   return nullptr;
2219 }
2220 
optimizeMemMoveChk(CallInst * CI,IRBuilder<> & B)2221 Value *FortifiedLibCallSimplifier::optimizeMemMoveChk(CallInst *CI,
2222                                                       IRBuilder<> &B) {
2223   if (isFortifiedCallFoldable(CI, 3, 2, false)) {
2224     B.CreateMemMove(CI->getArgOperand(0), CI->getArgOperand(1),
2225                     CI->getArgOperand(2), 1);
2226     return CI->getArgOperand(0);
2227   }
2228   return nullptr;
2229 }
2230 
optimizeMemSetChk(CallInst * CI,IRBuilder<> & B)2231 Value *FortifiedLibCallSimplifier::optimizeMemSetChk(CallInst *CI,
2232                                                      IRBuilder<> &B) {
2233   // TODO: Try foldMallocMemset() here.
2234 
2235   if (isFortifiedCallFoldable(CI, 3, 2, false)) {
2236     Value *Val = B.CreateIntCast(CI->getArgOperand(1), B.getInt8Ty(), false);
2237     B.CreateMemSet(CI->getArgOperand(0), Val, CI->getArgOperand(2), 1);
2238     return CI->getArgOperand(0);
2239   }
2240   return nullptr;
2241 }
2242 
optimizeStrpCpyChk(CallInst * CI,IRBuilder<> & B,LibFunc::Func Func)2243 Value *FortifiedLibCallSimplifier::optimizeStrpCpyChk(CallInst *CI,
2244                                                       IRBuilder<> &B,
2245                                                       LibFunc::Func Func) {
2246   Function *Callee = CI->getCalledFunction();
2247   StringRef Name = Callee->getName();
2248   const DataLayout &DL = CI->getModule()->getDataLayout();
2249   Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1),
2250         *ObjSize = CI->getArgOperand(2);
2251 
2252   // __stpcpy_chk(x,x,...)  -> x+strlen(x)
2253   if (Func == LibFunc::stpcpy_chk && !OnlyLowerUnknownSize && Dst == Src) {
2254     Value *StrLen = emitStrLen(Src, B, DL, TLI);
2255     return StrLen ? B.CreateInBoundsGEP(B.getInt8Ty(), Dst, StrLen) : nullptr;
2256   }
2257 
2258   // If a) we don't have any length information, or b) we know this will
2259   // fit then just lower to a plain st[rp]cpy. Otherwise we'll keep our
2260   // st[rp]cpy_chk call which may fail at runtime if the size is too long.
2261   // TODO: It might be nice to get a maximum length out of the possible
2262   // string lengths for varying.
2263   if (isFortifiedCallFoldable(CI, 2, 1, true))
2264     return emitStrCpy(Dst, Src, B, TLI, Name.substr(2, 6));
2265 
2266   if (OnlyLowerUnknownSize)
2267     return nullptr;
2268 
2269   // Maybe we can stil fold __st[rp]cpy_chk to __memcpy_chk.
2270   uint64_t Len = GetStringLength(Src);
2271   if (Len == 0)
2272     return nullptr;
2273 
2274   Type *SizeTTy = DL.getIntPtrType(CI->getContext());
2275   Value *LenV = ConstantInt::get(SizeTTy, Len);
2276   Value *Ret = emitMemCpyChk(Dst, Src, LenV, ObjSize, B, DL, TLI);
2277   // If the function was an __stpcpy_chk, and we were able to fold it into
2278   // a __memcpy_chk, we still need to return the correct end pointer.
2279   if (Ret && Func == LibFunc::stpcpy_chk)
2280     return B.CreateGEP(B.getInt8Ty(), Dst, ConstantInt::get(SizeTTy, Len - 1));
2281   return Ret;
2282 }
2283 
optimizeStrpNCpyChk(CallInst * CI,IRBuilder<> & B,LibFunc::Func Func)2284 Value *FortifiedLibCallSimplifier::optimizeStrpNCpyChk(CallInst *CI,
2285                                                        IRBuilder<> &B,
2286                                                        LibFunc::Func Func) {
2287   Function *Callee = CI->getCalledFunction();
2288   StringRef Name = Callee->getName();
2289   if (isFortifiedCallFoldable(CI, 3, 2, false)) {
2290     Value *Ret = emitStrNCpy(CI->getArgOperand(0), CI->getArgOperand(1),
2291                              CI->getArgOperand(2), B, TLI, Name.substr(2, 7));
2292     return Ret;
2293   }
2294   return nullptr;
2295 }
2296 
optimizeCall(CallInst * CI)2297 Value *FortifiedLibCallSimplifier::optimizeCall(CallInst *CI) {
2298   // FIXME: We shouldn't be changing "nobuiltin" or TLI unavailable calls here.
2299   // Some clang users checked for _chk libcall availability using:
2300   //   __has_builtin(__builtin___memcpy_chk)
2301   // When compiling with -fno-builtin, this is always true.
2302   // When passing -ffreestanding/-mkernel, which both imply -fno-builtin, we
2303   // end up with fortified libcalls, which isn't acceptable in a freestanding
2304   // environment which only provides their non-fortified counterparts.
2305   //
2306   // Until we change clang and/or teach external users to check for availability
2307   // differently, disregard the "nobuiltin" attribute and TLI::has.
2308   //
2309   // PR23093.
2310 
2311   LibFunc::Func Func;
2312   Function *Callee = CI->getCalledFunction();
2313 
2314   SmallVector<OperandBundleDef, 2> OpBundles;
2315   CI->getOperandBundlesAsDefs(OpBundles);
2316   IRBuilder<> Builder(CI, /*FPMathTag=*/nullptr, OpBundles);
2317   bool isCallingConvC = CI->getCallingConv() == llvm::CallingConv::C;
2318 
2319   // First, check that this is a known library functions and that the prototype
2320   // is correct.
2321   if (!TLI->getLibFunc(*Callee, Func))
2322     return nullptr;
2323 
2324   // We never change the calling convention.
2325   if (!ignoreCallingConv(Func) && !isCallingConvC)
2326     return nullptr;
2327 
2328   switch (Func) {
2329   case LibFunc::memcpy_chk:
2330     return optimizeMemCpyChk(CI, Builder);
2331   case LibFunc::memmove_chk:
2332     return optimizeMemMoveChk(CI, Builder);
2333   case LibFunc::memset_chk:
2334     return optimizeMemSetChk(CI, Builder);
2335   case LibFunc::stpcpy_chk:
2336   case LibFunc::strcpy_chk:
2337     return optimizeStrpCpyChk(CI, Builder, Func);
2338   case LibFunc::stpncpy_chk:
2339   case LibFunc::strncpy_chk:
2340     return optimizeStrpNCpyChk(CI, Builder, Func);
2341   default:
2342     break;
2343   }
2344   return nullptr;
2345 }
2346 
FortifiedLibCallSimplifier(const TargetLibraryInfo * TLI,bool OnlyLowerUnknownSize)2347 FortifiedLibCallSimplifier::FortifiedLibCallSimplifier(
2348     const TargetLibraryInfo *TLI, bool OnlyLowerUnknownSize)
2349     : TLI(TLI), OnlyLowerUnknownSize(OnlyLowerUnknownSize) {}
2350