1 | //===------ SimplifyLibCalls.cpp - Library calls simplifier ---------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This file implements the library calls simplifier. It does not implement |
10 | // any pass, but can be used by other passes to do simplifications. |
11 | // |
12 | //===----------------------------------------------------------------------===// |
13 | |
14 | #include "llvm/Transforms/Utils/SimplifyLibCalls.h" |
15 | #include "llvm/ADT/APFloat.h" |
16 | #include "llvm/ADT/APSInt.h" |
17 | #include "llvm/ADT/SmallString.h" |
18 | #include "llvm/ADT/StringExtras.h" |
19 | #include "llvm/Analysis/ConstantFolding.h" |
20 | #include "llvm/Analysis/Loads.h" |
21 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" |
22 | #include "llvm/Analysis/TargetLibraryInfo.h" |
23 | #include "llvm/Analysis/ValueTracking.h" |
24 | #include "llvm/IR/AttributeMask.h" |
25 | #include "llvm/IR/DataLayout.h" |
26 | #include "llvm/IR/Function.h" |
27 | #include "llvm/IR/IRBuilder.h" |
28 | #include "llvm/IR/IntrinsicInst.h" |
29 | #include "llvm/IR/Intrinsics.h" |
30 | #include "llvm/IR/Module.h" |
31 | #include "llvm/IR/PatternMatch.h" |
32 | #include "llvm/Support/Casting.h" |
33 | #include "llvm/Support/CommandLine.h" |
34 | #include "llvm/Support/KnownBits.h" |
35 | #include "llvm/Support/KnownFPClass.h" |
36 | #include "llvm/Support/MathExtras.h" |
37 | #include "llvm/TargetParser/Triple.h" |
38 | #include "llvm/Transforms/Utils/BuildLibCalls.h" |
39 | #include "llvm/Transforms/Utils/Local.h" |
40 | #include "llvm/Transforms/Utils/SizeOpts.h" |
41 | |
42 | #include <cmath> |
43 | |
44 | using namespace llvm; |
45 | using namespace PatternMatch; |
46 | |
47 | static cl::opt<bool> |
48 | EnableUnsafeFPShrink("enable-double-float-shrink" , cl::Hidden, |
49 | cl::init(Val: false), |
50 | cl::desc("Enable unsafe double to float " |
51 | "shrinking for math lib calls" )); |
52 | |
53 | // Enable conversion of operator new calls with a MemProf hot or cold hint |
54 | // to an operator new call that takes a hot/cold hint. Off by default since |
55 | // not all allocators currently support this extension. |
56 | static cl::opt<bool> |
57 | OptimizeHotColdNew("optimize-hot-cold-new" , cl::Hidden, cl::init(Val: false), |
58 | cl::desc("Enable hot/cold operator new library calls" )); |
59 | static cl::opt<bool> OptimizeExistingHotColdNew( |
60 | "optimize-existing-hot-cold-new" , cl::Hidden, cl::init(Val: false), |
61 | cl::desc( |
62 | "Enable optimization of existing hot/cold operator new library calls" )); |
63 | |
64 | namespace { |
65 | |
66 | // Specialized parser to ensure the hint is an 8 bit value (we can't specify |
67 | // uint8_t to opt<> as that is interpreted to mean that we are passing a char |
68 | // option with a specific set of values. |
69 | struct HotColdHintParser : public cl::parser<unsigned> { |
70 | HotColdHintParser(cl::Option &O) : cl::parser<unsigned>(O) {} |
71 | |
72 | bool parse(cl::Option &O, StringRef ArgName, StringRef Arg, unsigned &Value) { |
73 | if (Arg.getAsInteger(Radix: 0, Result&: Value)) |
74 | return O.error(Message: "'" + Arg + "' value invalid for uint argument!" ); |
75 | |
76 | if (Value > 255) |
77 | return O.error(Message: "'" + Arg + "' value must be in the range [0, 255]!" ); |
78 | |
79 | return false; |
80 | } |
81 | }; |
82 | |
83 | } // end anonymous namespace |
84 | |
85 | // Hot/cold operator new takes an 8 bit hotness hint, where 0 is the coldest |
86 | // and 255 is the hottest. Default to 1 value away from the coldest and hottest |
87 | // hints, so that the compiler hinted allocations are slightly less strong than |
88 | // manually inserted hints at the two extremes. |
89 | static cl::opt<unsigned, false, HotColdHintParser> ColdNewHintValue( |
90 | "cold-new-hint-value" , cl::Hidden, cl::init(Val: 1), |
91 | cl::desc("Value to pass to hot/cold operator new for cold allocation" )); |
92 | static cl::opt<unsigned, false, HotColdHintParser> |
93 | NotColdNewHintValue("notcold-new-hint-value" , cl::Hidden, cl::init(Val: 128), |
94 | cl::desc("Value to pass to hot/cold operator new for " |
95 | "notcold (warm) allocation" )); |
96 | static cl::opt<unsigned, false, HotColdHintParser> HotNewHintValue( |
97 | "hot-new-hint-value" , cl::Hidden, cl::init(Val: 254), |
98 | cl::desc("Value to pass to hot/cold operator new for hot allocation" )); |
99 | |
100 | //===----------------------------------------------------------------------===// |
101 | // Helper Functions |
102 | //===----------------------------------------------------------------------===// |
103 | |
104 | static bool ignoreCallingConv(LibFunc Func) { |
105 | return Func == LibFunc_abs || Func == LibFunc_labs || |
106 | Func == LibFunc_llabs || Func == LibFunc_strlen; |
107 | } |
108 | |
109 | /// Return true if it is only used in equality comparisons with With. |
110 | static bool isOnlyUsedInEqualityComparison(Value *V, Value *With) { |
111 | for (User *U : V->users()) { |
112 | if (ICmpInst *IC = dyn_cast<ICmpInst>(Val: U)) |
113 | if (IC->isEquality() && IC->getOperand(i_nocapture: 1) == With) |
114 | continue; |
115 | // Unknown instruction. |
116 | return false; |
117 | } |
118 | return true; |
119 | } |
120 | |
121 | static bool callHasFloatingPointArgument(const CallInst *CI) { |
122 | return any_of(Range: CI->operands(), P: [](const Use &OI) { |
123 | return OI->getType()->isFloatingPointTy(); |
124 | }); |
125 | } |
126 | |
127 | static bool callHasFP128Argument(const CallInst *CI) { |
128 | return any_of(Range: CI->operands(), P: [](const Use &OI) { |
129 | return OI->getType()->isFP128Ty(); |
130 | }); |
131 | } |
132 | |
133 | // Convert the entire string Str representing an integer in Base, up to |
134 | // the terminating nul if present, to a constant according to the rules |
135 | // of strtoul[l] or, when AsSigned is set, of strtol[l]. On success |
136 | // return the result, otherwise null. |
137 | // The function assumes the string is encoded in ASCII and carefully |
138 | // avoids converting sequences (including "") that the corresponding |
139 | // library call might fail and set errno for. |
140 | static Value *convertStrToInt(CallInst *CI, StringRef &Str, Value *EndPtr, |
141 | uint64_t Base, bool AsSigned, IRBuilderBase &B) { |
142 | if (Base < 2 || Base > 36) |
143 | if (Base != 0) |
144 | // Fail for an invalid base (required by POSIX). |
145 | return nullptr; |
146 | |
147 | // Current offset into the original string to reflect in EndPtr. |
148 | size_t Offset = 0; |
149 | // Strip leading whitespace. |
150 | for ( ; Offset != Str.size(); ++Offset) |
151 | if (!isSpace(C: (unsigned char)Str[Offset])) { |
152 | Str = Str.substr(Start: Offset); |
153 | break; |
154 | } |
155 | |
156 | if (Str.empty()) |
157 | // Fail for empty subject sequences (POSIX allows but doesn't require |
158 | // strtol[l]/strtoul[l] to fail with EINVAL). |
159 | return nullptr; |
160 | |
161 | // Strip but remember the sign. |
162 | bool Negate = Str[0] == '-'; |
163 | if (Str[0] == '-' || Str[0] == '+') { |
164 | Str = Str.drop_front(); |
165 | if (Str.empty()) |
166 | // Fail for a sign with nothing after it. |
167 | return nullptr; |
168 | ++Offset; |
169 | } |
170 | |
171 | // Set Max to the absolute value of the minimum (for signed), or |
172 | // to the maximum (for unsigned) value representable in the type. |
173 | Type *RetTy = CI->getType(); |
174 | unsigned NBits = RetTy->getPrimitiveSizeInBits(); |
175 | uint64_t Max = AsSigned && Negate ? 1 : 0; |
176 | Max += AsSigned ? maxIntN(N: NBits) : maxUIntN(N: NBits); |
177 | |
178 | // Autodetect Base if it's zero and consume the "0x" prefix. |
179 | if (Str.size() > 1) { |
180 | if (Str[0] == '0') { |
181 | if (toUpper(x: (unsigned char)Str[1]) == 'X') { |
182 | if (Str.size() == 2 || (Base && Base != 16)) |
183 | // Fail if Base doesn't allow the "0x" prefix or for the prefix |
184 | // alone that implementations like BSD set errno to EINVAL for. |
185 | return nullptr; |
186 | |
187 | Str = Str.drop_front(N: 2); |
188 | Offset += 2; |
189 | Base = 16; |
190 | } |
191 | else if (Base == 0) |
192 | Base = 8; |
193 | } else if (Base == 0) |
194 | Base = 10; |
195 | } |
196 | else if (Base == 0) |
197 | Base = 10; |
198 | |
199 | // Convert the rest of the subject sequence, not including the sign, |
200 | // to its uint64_t representation (this assumes the source character |
201 | // set is ASCII). |
202 | uint64_t Result = 0; |
203 | for (unsigned i = 0; i != Str.size(); ++i) { |
204 | unsigned char DigVal = Str[i]; |
205 | if (isDigit(C: DigVal)) |
206 | DigVal = DigVal - '0'; |
207 | else { |
208 | DigVal = toUpper(x: DigVal); |
209 | if (isAlpha(C: DigVal)) |
210 | DigVal = DigVal - 'A' + 10; |
211 | else |
212 | return nullptr; |
213 | } |
214 | |
215 | if (DigVal >= Base) |
216 | // Fail if the digit is not valid in the Base. |
217 | return nullptr; |
218 | |
219 | // Add the digit and fail if the result is not representable in |
220 | // the (unsigned form of the) destination type. |
221 | bool VFlow; |
222 | Result = SaturatingMultiplyAdd(X: Result, Y: Base, A: (uint64_t)DigVal, ResultOverflowed: &VFlow); |
223 | if (VFlow || Result > Max) |
224 | return nullptr; |
225 | } |
226 | |
227 | if (EndPtr) { |
228 | // Store the pointer to the end. |
229 | Value *Off = B.getInt64(C: Offset + Str.size()); |
230 | Value *StrBeg = CI->getArgOperand(i: 0); |
231 | Value *StrEnd = B.CreateInBoundsGEP(Ty: B.getInt8Ty(), Ptr: StrBeg, IdxList: Off, Name: "endptr" ); |
232 | B.CreateStore(Val: StrEnd, Ptr: EndPtr); |
233 | } |
234 | |
235 | if (Negate) |
236 | // Unsigned negation doesn't overflow. |
237 | Result = -Result; |
238 | |
239 | return ConstantInt::get(Ty: RetTy, V: Result); |
240 | } |
241 | |
242 | static bool isOnlyUsedInComparisonWithZero(Value *V) { |
243 | for (User *U : V->users()) { |
244 | if (ICmpInst *IC = dyn_cast<ICmpInst>(Val: U)) |
245 | if (Constant *C = dyn_cast<Constant>(Val: IC->getOperand(i_nocapture: 1))) |
246 | if (C->isNullValue()) |
247 | continue; |
248 | // Unknown instruction. |
249 | return false; |
250 | } |
251 | return true; |
252 | } |
253 | |
254 | static bool canTransformToMemCmp(CallInst *CI, Value *Str, uint64_t Len, |
255 | const DataLayout &DL) { |
256 | if (!isOnlyUsedInComparisonWithZero(V: CI)) |
257 | return false; |
258 | |
259 | if (!isDereferenceableAndAlignedPointer(V: Str, Alignment: Align(1), Size: APInt(64, Len), DL)) |
260 | return false; |
261 | |
262 | if (CI->getFunction()->hasFnAttribute(Kind: Attribute::SanitizeMemory)) |
263 | return false; |
264 | |
265 | return true; |
266 | } |
267 | |
268 | static void annotateDereferenceableBytes(CallInst *CI, |
269 | ArrayRef<unsigned> ArgNos, |
270 | uint64_t DereferenceableBytes) { |
271 | const Function *F = CI->getCaller(); |
272 | if (!F) |
273 | return; |
274 | for (unsigned ArgNo : ArgNos) { |
275 | uint64_t DerefBytes = DereferenceableBytes; |
276 | unsigned AS = CI->getArgOperand(i: ArgNo)->getType()->getPointerAddressSpace(); |
277 | if (!llvm::NullPointerIsDefined(F, AS) || |
278 | CI->paramHasAttr(ArgNo, Kind: Attribute::NonNull)) |
279 | DerefBytes = std::max(a: CI->getParamDereferenceableOrNullBytes(i: ArgNo), |
280 | b: DereferenceableBytes); |
281 | |
282 | if (CI->getParamDereferenceableBytes(i: ArgNo) < DerefBytes) { |
283 | CI->removeParamAttr(ArgNo, Kind: Attribute::Dereferenceable); |
284 | if (!llvm::NullPointerIsDefined(F, AS) || |
285 | CI->paramHasAttr(ArgNo, Kind: Attribute::NonNull)) |
286 | CI->removeParamAttr(ArgNo, Kind: Attribute::DereferenceableOrNull); |
287 | CI->addParamAttr(ArgNo, Attr: Attribute::getWithDereferenceableBytes( |
288 | Context&: CI->getContext(), Bytes: DerefBytes)); |
289 | } |
290 | } |
291 | } |
292 | |
293 | static void annotateNonNullNoUndefBasedOnAccess(CallInst *CI, |
294 | ArrayRef<unsigned> ArgNos) { |
295 | Function *F = CI->getCaller(); |
296 | if (!F) |
297 | return; |
298 | |
299 | for (unsigned ArgNo : ArgNos) { |
300 | if (!CI->paramHasAttr(ArgNo, Kind: Attribute::NoUndef)) |
301 | CI->addParamAttr(ArgNo, Kind: Attribute::NoUndef); |
302 | |
303 | if (!CI->paramHasAttr(ArgNo, Kind: Attribute::NonNull)) { |
304 | unsigned AS = |
305 | CI->getArgOperand(i: ArgNo)->getType()->getPointerAddressSpace(); |
306 | if (llvm::NullPointerIsDefined(F, AS)) |
307 | continue; |
308 | CI->addParamAttr(ArgNo, Kind: Attribute::NonNull); |
309 | } |
310 | |
311 | annotateDereferenceableBytes(CI, ArgNos: ArgNo, DereferenceableBytes: 1); |
312 | } |
313 | } |
314 | |
315 | static void annotateNonNullAndDereferenceable(CallInst *CI, ArrayRef<unsigned> ArgNos, |
316 | Value *Size, const DataLayout &DL) { |
317 | if (ConstantInt *LenC = dyn_cast<ConstantInt>(Val: Size)) { |
318 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos); |
319 | annotateDereferenceableBytes(CI, ArgNos, DereferenceableBytes: LenC->getZExtValue()); |
320 | } else if (isKnownNonZero(V: Size, Q: DL)) { |
321 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos); |
322 | const APInt *X, *Y; |
323 | uint64_t DerefMin = 1; |
324 | if (match(V: Size, P: m_Select(C: m_Value(), L: m_APInt(Res&: X), R: m_APInt(Res&: Y)))) { |
325 | DerefMin = std::min(a: X->getZExtValue(), b: Y->getZExtValue()); |
326 | annotateDereferenceableBytes(CI, ArgNos, DereferenceableBytes: DerefMin); |
327 | } |
328 | } |
329 | } |
330 | |
331 | // Copy CallInst "flags" like musttail, notail, and tail. Return New param for |
332 | // easier chaining. Calls to emit* and B.createCall should probably be wrapped |
333 | // in this function when New is created to replace Old. Callers should take |
334 | // care to check Old.isMustTailCall() if they aren't replacing Old directly |
335 | // with New. |
336 | static Value *copyFlags(const CallInst &Old, Value *New) { |
337 | assert(!Old.isMustTailCall() && "do not copy musttail call flags" ); |
338 | assert(!Old.isNoTailCall() && "do not copy notail call flags" ); |
339 | if (auto *NewCI = dyn_cast_or_null<CallInst>(Val: New)) |
340 | NewCI->setTailCallKind(Old.getTailCallKind()); |
341 | return New; |
342 | } |
343 | |
344 | static Value *mergeAttributesAndFlags(CallInst *NewCI, const CallInst &Old) { |
345 | NewCI->setAttributes(AttributeList::get( |
346 | C&: NewCI->getContext(), Attrs: {NewCI->getAttributes(), Old.getAttributes()})); |
347 | NewCI->removeRetAttrs(AttrsToRemove: AttributeFuncs::typeIncompatible( |
348 | Ty: NewCI->getType(), AS: NewCI->getRetAttributes())); |
349 | for (unsigned I = 0; I < NewCI->arg_size(); ++I) |
350 | NewCI->removeParamAttrs( |
351 | ArgNo: I, AttrsToRemove: AttributeFuncs::typeIncompatible(Ty: NewCI->getArgOperand(i: I)->getType(), |
352 | AS: NewCI->getParamAttributes(ArgNo: I))); |
353 | |
354 | return copyFlags(Old, New: NewCI); |
355 | } |
356 | |
357 | // Helper to avoid truncating the length if size_t is 32-bits. |
358 | static StringRef substr(StringRef Str, uint64_t Len) { |
359 | return Len >= Str.size() ? Str : Str.substr(Start: 0, N: Len); |
360 | } |
361 | |
362 | //===----------------------------------------------------------------------===// |
363 | // String and Memory Library Call Optimizations |
364 | //===----------------------------------------------------------------------===// |
365 | |
366 | Value *LibCallSimplifier::optimizeStrCat(CallInst *CI, IRBuilderBase &B) { |
367 | // Extract some information from the instruction |
368 | Value *Dst = CI->getArgOperand(i: 0); |
369 | Value *Src = CI->getArgOperand(i: 1); |
370 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: {0, 1}); |
371 | |
372 | // See if we can get the length of the input string. |
373 | uint64_t Len = GetStringLength(V: Src); |
374 | if (Len) |
375 | annotateDereferenceableBytes(CI, ArgNos: 1, DereferenceableBytes: Len); |
376 | else |
377 | return nullptr; |
378 | --Len; // Unbias length. |
379 | |
380 | // Handle the simple, do-nothing case: strcat(x, "") -> x |
381 | if (Len == 0) |
382 | return Dst; |
383 | |
384 | return copyFlags(Old: *CI, New: emitStrLenMemCpy(Src, Dst, Len, B)); |
385 | } |
386 | |
387 | Value *LibCallSimplifier::emitStrLenMemCpy(Value *Src, Value *Dst, uint64_t Len, |
388 | IRBuilderBase &B) { |
389 | // We need to find the end of the destination string. That's where the |
390 | // memory is to be moved to. We just generate a call to strlen. |
391 | Value *DstLen = emitStrLen(Ptr: Dst, B, DL, TLI); |
392 | if (!DstLen) |
393 | return nullptr; |
394 | |
395 | // Now that we have the destination's length, we must index into the |
396 | // destination's pointer to get the actual memcpy destination (end of |
397 | // the string .. we're concatenating). |
398 | Value *CpyDst = B.CreateInBoundsGEP(Ty: B.getInt8Ty(), Ptr: Dst, IdxList: DstLen, Name: "endptr" ); |
399 | |
400 | // We have enough information to now generate the memcpy call to do the |
401 | // concatenation for us. Make a memcpy to copy the nul byte with align = 1. |
402 | B.CreateMemCpy(Dst: CpyDst, DstAlign: Align(1), Src, SrcAlign: Align(1), |
403 | Size: TLI->getAsSizeT(V: Len + 1, M: *B.GetInsertBlock()->getModule())); |
404 | return Dst; |
405 | } |
406 | |
407 | Value *LibCallSimplifier::optimizeStrNCat(CallInst *CI, IRBuilderBase &B) { |
408 | // Extract some information from the instruction. |
409 | Value *Dst = CI->getArgOperand(i: 0); |
410 | Value *Src = CI->getArgOperand(i: 1); |
411 | Value *Size = CI->getArgOperand(i: 2); |
412 | uint64_t Len; |
413 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: 0); |
414 | if (isKnownNonZero(V: Size, Q: DL)) |
415 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: 1); |
416 | |
417 | // We don't do anything if length is not constant. |
418 | ConstantInt *LengthArg = dyn_cast<ConstantInt>(Val: Size); |
419 | if (LengthArg) { |
420 | Len = LengthArg->getZExtValue(); |
421 | // strncat(x, c, 0) -> x |
422 | if (!Len) |
423 | return Dst; |
424 | } else { |
425 | return nullptr; |
426 | } |
427 | |
428 | // See if we can get the length of the input string. |
429 | uint64_t SrcLen = GetStringLength(V: Src); |
430 | if (SrcLen) { |
431 | annotateDereferenceableBytes(CI, ArgNos: 1, DereferenceableBytes: SrcLen); |
432 | --SrcLen; // Unbias length. |
433 | } else { |
434 | return nullptr; |
435 | } |
436 | |
437 | // strncat(x, "", c) -> x |
438 | if (SrcLen == 0) |
439 | return Dst; |
440 | |
441 | // We don't optimize this case. |
442 | if (Len < SrcLen) |
443 | return nullptr; |
444 | |
445 | // strncat(x, s, c) -> strcat(x, s) |
446 | // s is constant so the strcat can be optimized further. |
447 | return copyFlags(Old: *CI, New: emitStrLenMemCpy(Src, Dst, Len: SrcLen, B)); |
448 | } |
449 | |
450 | // Helper to transform memchr(S, C, N) == S to N && *S == C and, when |
451 | // NBytes is null, strchr(S, C) to *S == C. A precondition of the function |
452 | // is that either S is dereferenceable or the value of N is nonzero. |
453 | static Value* memChrToCharCompare(CallInst *CI, Value *NBytes, |
454 | IRBuilderBase &B, const DataLayout &DL) |
455 | { |
456 | Value *Src = CI->getArgOperand(i: 0); |
457 | Value *CharVal = CI->getArgOperand(i: 1); |
458 | |
459 | // Fold memchr(A, C, N) == A to N && *A == C. |
460 | Type *CharTy = B.getInt8Ty(); |
461 | Value *Char0 = B.CreateLoad(Ty: CharTy, Ptr: Src); |
462 | CharVal = B.CreateTrunc(V: CharVal, DestTy: CharTy); |
463 | Value *Cmp = B.CreateICmpEQ(LHS: Char0, RHS: CharVal, Name: "char0cmp" ); |
464 | |
465 | if (NBytes) { |
466 | Value *Zero = ConstantInt::get(Ty: NBytes->getType(), V: 0); |
467 | Value *And = B.CreateICmpNE(LHS: NBytes, RHS: Zero); |
468 | Cmp = B.CreateLogicalAnd(Cond1: And, Cond2: Cmp); |
469 | } |
470 | |
471 | Value *NullPtr = Constant::getNullValue(Ty: CI->getType()); |
472 | return B.CreateSelect(C: Cmp, True: Src, False: NullPtr); |
473 | } |
474 | |
475 | Value *LibCallSimplifier::optimizeStrChr(CallInst *CI, IRBuilderBase &B) { |
476 | Value *SrcStr = CI->getArgOperand(i: 0); |
477 | Value *CharVal = CI->getArgOperand(i: 1); |
478 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: 0); |
479 | |
480 | if (isOnlyUsedInEqualityComparison(V: CI, With: SrcStr)) |
481 | return memChrToCharCompare(CI, NBytes: nullptr, B, DL); |
482 | |
483 | // If the second operand is non-constant, see if we can compute the length |
484 | // of the input string and turn this into memchr. |
485 | ConstantInt *CharC = dyn_cast<ConstantInt>(Val: CharVal); |
486 | if (!CharC) { |
487 | uint64_t Len = GetStringLength(V: SrcStr); |
488 | if (Len) |
489 | annotateDereferenceableBytes(CI, ArgNos: 0, DereferenceableBytes: Len); |
490 | else |
491 | return nullptr; |
492 | |
493 | Function *Callee = CI->getCalledFunction(); |
494 | FunctionType *FT = Callee->getFunctionType(); |
495 | unsigned IntBits = TLI->getIntSize(); |
496 | if (!FT->getParamType(i: 1)->isIntegerTy(Bitwidth: IntBits)) // memchr needs 'int'. |
497 | return nullptr; |
498 | |
499 | unsigned SizeTBits = TLI->getSizeTSize(M: *CI->getModule()); |
500 | Type *SizeTTy = IntegerType::get(C&: CI->getContext(), NumBits: SizeTBits); |
501 | return copyFlags(Old: *CI, |
502 | New: emitMemChr(Ptr: SrcStr, Val: CharVal, // include nul. |
503 | Len: ConstantInt::get(Ty: SizeTTy, V: Len), B, |
504 | DL, TLI)); |
505 | } |
506 | |
507 | if (CharC->isZero()) { |
508 | Value *NullPtr = Constant::getNullValue(Ty: CI->getType()); |
509 | if (isOnlyUsedInEqualityComparison(V: CI, With: NullPtr)) |
510 | // Pre-empt the transformation to strlen below and fold |
511 | // strchr(A, '\0') == null to false. |
512 | return B.CreateIntToPtr(V: B.getTrue(), DestTy: CI->getType()); |
513 | } |
514 | |
515 | // Otherwise, the character is a constant, see if the first argument is |
516 | // a string literal. If so, we can constant fold. |
517 | StringRef Str; |
518 | if (!getConstantStringInfo(V: SrcStr, Str)) { |
519 | if (CharC->isZero()) // strchr(p, 0) -> p + strlen(p) |
520 | if (Value *StrLen = emitStrLen(Ptr: SrcStr, B, DL, TLI)) |
521 | return B.CreateInBoundsGEP(Ty: B.getInt8Ty(), Ptr: SrcStr, IdxList: StrLen, Name: "strchr" ); |
522 | return nullptr; |
523 | } |
524 | |
525 | // Compute the offset, make sure to handle the case when we're searching for |
526 | // zero (a weird way to spell strlen). |
527 | size_t I = (0xFF & CharC->getSExtValue()) == 0 |
528 | ? Str.size() |
529 | : Str.find(C: CharC->getSExtValue()); |
530 | if (I == StringRef::npos) // Didn't find the char. strchr returns null. |
531 | return Constant::getNullValue(Ty: CI->getType()); |
532 | |
533 | // strchr(s+n,c) -> gep(s+n+i,c) |
534 | return B.CreateInBoundsGEP(Ty: B.getInt8Ty(), Ptr: SrcStr, IdxList: B.getInt64(C: I), Name: "strchr" ); |
535 | } |
536 | |
537 | Value *LibCallSimplifier::optimizeStrRChr(CallInst *CI, IRBuilderBase &B) { |
538 | Value *SrcStr = CI->getArgOperand(i: 0); |
539 | Value *CharVal = CI->getArgOperand(i: 1); |
540 | ConstantInt *CharC = dyn_cast<ConstantInt>(Val: CharVal); |
541 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: 0); |
542 | |
543 | StringRef Str; |
544 | if (!getConstantStringInfo(V: SrcStr, Str)) { |
545 | // strrchr(s, 0) -> strchr(s, 0) |
546 | if (CharC && CharC->isZero()) |
547 | return copyFlags(Old: *CI, New: emitStrChr(Ptr: SrcStr, C: '\0', B, TLI)); |
548 | return nullptr; |
549 | } |
550 | |
551 | unsigned SizeTBits = TLI->getSizeTSize(M: *CI->getModule()); |
552 | Type *SizeTTy = IntegerType::get(C&: CI->getContext(), NumBits: SizeTBits); |
553 | |
554 | // Try to expand strrchr to the memrchr nonstandard extension if it's |
555 | // available, or simply fail otherwise. |
556 | uint64_t NBytes = Str.size() + 1; // Include the terminating nul. |
557 | Value *Size = ConstantInt::get(Ty: SizeTTy, V: NBytes); |
558 | return copyFlags(Old: *CI, New: emitMemRChr(Ptr: SrcStr, Val: CharVal, Len: Size, B, DL, TLI)); |
559 | } |
560 | |
561 | Value *LibCallSimplifier::optimizeStrCmp(CallInst *CI, IRBuilderBase &B) { |
562 | Value *Str1P = CI->getArgOperand(i: 0), *Str2P = CI->getArgOperand(i: 1); |
563 | if (Str1P == Str2P) // strcmp(x,x) -> 0 |
564 | return ConstantInt::get(Ty: CI->getType(), V: 0); |
565 | |
566 | StringRef Str1, Str2; |
567 | bool HasStr1 = getConstantStringInfo(V: Str1P, Str&: Str1); |
568 | bool HasStr2 = getConstantStringInfo(V: Str2P, Str&: Str2); |
569 | |
570 | // strcmp(x, y) -> cnst (if both x and y are constant strings) |
571 | if (HasStr1 && HasStr2) |
572 | return ConstantInt::get(Ty: CI->getType(), |
573 | V: std::clamp(val: Str1.compare(RHS: Str2), lo: -1, hi: 1)); |
574 | |
575 | if (HasStr1 && Str1.empty()) // strcmp("", x) -> -*x |
576 | return B.CreateNeg(V: B.CreateZExt( |
577 | V: B.CreateLoad(Ty: B.getInt8Ty(), Ptr: Str2P, Name: "strcmpload" ), DestTy: CI->getType())); |
578 | |
579 | if (HasStr2 && Str2.empty()) // strcmp(x,"") -> *x |
580 | return B.CreateZExt(V: B.CreateLoad(Ty: B.getInt8Ty(), Ptr: Str1P, Name: "strcmpload" ), |
581 | DestTy: CI->getType()); |
582 | |
583 | // strcmp(P, "x") -> memcmp(P, "x", 2) |
584 | uint64_t Len1 = GetStringLength(V: Str1P); |
585 | if (Len1) |
586 | annotateDereferenceableBytes(CI, ArgNos: 0, DereferenceableBytes: Len1); |
587 | uint64_t Len2 = GetStringLength(V: Str2P); |
588 | if (Len2) |
589 | annotateDereferenceableBytes(CI, ArgNos: 1, DereferenceableBytes: Len2); |
590 | |
591 | if (Len1 && Len2) { |
592 | return copyFlags( |
593 | Old: *CI, New: emitMemCmp(Ptr1: Str1P, Ptr2: Str2P, |
594 | Len: TLI->getAsSizeT(V: std::min(a: Len1, b: Len2), M: *CI->getModule()), |
595 | B, DL, TLI)); |
596 | } |
597 | |
598 | // strcmp to memcmp |
599 | if (!HasStr1 && HasStr2) { |
600 | if (canTransformToMemCmp(CI, Str: Str1P, Len: Len2, DL)) |
601 | return copyFlags(Old: *CI, New: emitMemCmp(Ptr1: Str1P, Ptr2: Str2P, |
602 | Len: TLI->getAsSizeT(V: Len2, M: *CI->getModule()), |
603 | B, DL, TLI)); |
604 | } else if (HasStr1 && !HasStr2) { |
605 | if (canTransformToMemCmp(CI, Str: Str2P, Len: Len1, DL)) |
606 | return copyFlags(Old: *CI, New: emitMemCmp(Ptr1: Str1P, Ptr2: Str2P, |
607 | Len: TLI->getAsSizeT(V: Len1, M: *CI->getModule()), |
608 | B, DL, TLI)); |
609 | } |
610 | |
611 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: {0, 1}); |
612 | return nullptr; |
613 | } |
614 | |
615 | // Optimize a memcmp or, when StrNCmp is true, strncmp call CI with constant |
616 | // arrays LHS and RHS and nonconstant Size. |
617 | static Value *optimizeMemCmpVarSize(CallInst *CI, Value *LHS, Value *RHS, |
618 | Value *Size, bool StrNCmp, |
619 | IRBuilderBase &B, const DataLayout &DL); |
620 | |
621 | Value *LibCallSimplifier::optimizeStrNCmp(CallInst *CI, IRBuilderBase &B) { |
622 | Value *Str1P = CI->getArgOperand(i: 0); |
623 | Value *Str2P = CI->getArgOperand(i: 1); |
624 | Value *Size = CI->getArgOperand(i: 2); |
625 | if (Str1P == Str2P) // strncmp(x,x,n) -> 0 |
626 | return ConstantInt::get(Ty: CI->getType(), V: 0); |
627 | |
628 | if (isKnownNonZero(V: Size, Q: DL)) |
629 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: {0, 1}); |
630 | // Get the length argument if it is constant. |
631 | uint64_t Length; |
632 | if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(Val: Size)) |
633 | Length = LengthArg->getZExtValue(); |
634 | else |
635 | return optimizeMemCmpVarSize(CI, LHS: Str1P, RHS: Str2P, Size, StrNCmp: true, B, DL); |
636 | |
637 | if (Length == 0) // strncmp(x,y,0) -> 0 |
638 | return ConstantInt::get(Ty: CI->getType(), V: 0); |
639 | |
640 | if (Length == 1) // strncmp(x,y,1) -> memcmp(x,y,1) |
641 | return copyFlags(Old: *CI, New: emitMemCmp(Ptr1: Str1P, Ptr2: Str2P, Len: Size, B, DL, TLI)); |
642 | |
643 | StringRef Str1, Str2; |
644 | bool HasStr1 = getConstantStringInfo(V: Str1P, Str&: Str1); |
645 | bool HasStr2 = getConstantStringInfo(V: Str2P, Str&: Str2); |
646 | |
647 | // strncmp(x, y) -> cnst (if both x and y are constant strings) |
648 | if (HasStr1 && HasStr2) { |
649 | // Avoid truncating the 64-bit Length to 32 bits in ILP32. |
650 | StringRef SubStr1 = substr(Str: Str1, Len: Length); |
651 | StringRef SubStr2 = substr(Str: Str2, Len: Length); |
652 | return ConstantInt::get(Ty: CI->getType(), |
653 | V: std::clamp(val: SubStr1.compare(RHS: SubStr2), lo: -1, hi: 1)); |
654 | } |
655 | |
656 | if (HasStr1 && Str1.empty()) // strncmp("", x, n) -> -*x |
657 | return B.CreateNeg(V: B.CreateZExt( |
658 | V: B.CreateLoad(Ty: B.getInt8Ty(), Ptr: Str2P, Name: "strcmpload" ), DestTy: CI->getType())); |
659 | |
660 | if (HasStr2 && Str2.empty()) // strncmp(x, "", n) -> *x |
661 | return B.CreateZExt(V: B.CreateLoad(Ty: B.getInt8Ty(), Ptr: Str1P, Name: "strcmpload" ), |
662 | DestTy: CI->getType()); |
663 | |
664 | uint64_t Len1 = GetStringLength(V: Str1P); |
665 | if (Len1) |
666 | annotateDereferenceableBytes(CI, ArgNos: 0, DereferenceableBytes: Len1); |
667 | uint64_t Len2 = GetStringLength(V: Str2P); |
668 | if (Len2) |
669 | annotateDereferenceableBytes(CI, ArgNos: 1, DereferenceableBytes: Len2); |
670 | |
671 | // strncmp to memcmp |
672 | if (!HasStr1 && HasStr2) { |
673 | Len2 = std::min(a: Len2, b: Length); |
674 | if (canTransformToMemCmp(CI, Str: Str1P, Len: Len2, DL)) |
675 | return copyFlags(Old: *CI, New: emitMemCmp(Ptr1: Str1P, Ptr2: Str2P, |
676 | Len: TLI->getAsSizeT(V: Len2, M: *CI->getModule()), |
677 | B, DL, TLI)); |
678 | } else if (HasStr1 && !HasStr2) { |
679 | Len1 = std::min(a: Len1, b: Length); |
680 | if (canTransformToMemCmp(CI, Str: Str2P, Len: Len1, DL)) |
681 | return copyFlags(Old: *CI, New: emitMemCmp(Ptr1: Str1P, Ptr2: Str2P, |
682 | Len: TLI->getAsSizeT(V: Len1, M: *CI->getModule()), |
683 | B, DL, TLI)); |
684 | } |
685 | |
686 | return nullptr; |
687 | } |
688 | |
689 | Value *LibCallSimplifier::optimizeStrNDup(CallInst *CI, IRBuilderBase &B) { |
690 | Value *Src = CI->getArgOperand(i: 0); |
691 | ConstantInt *Size = dyn_cast<ConstantInt>(Val: CI->getArgOperand(i: 1)); |
692 | uint64_t SrcLen = GetStringLength(V: Src); |
693 | if (SrcLen && Size) { |
694 | annotateDereferenceableBytes(CI, ArgNos: 0, DereferenceableBytes: SrcLen); |
695 | if (SrcLen <= Size->getZExtValue() + 1) |
696 | return copyFlags(Old: *CI, New: emitStrDup(Ptr: Src, B, TLI)); |
697 | } |
698 | |
699 | return nullptr; |
700 | } |
701 | |
702 | Value *LibCallSimplifier::optimizeStrCpy(CallInst *CI, IRBuilderBase &B) { |
703 | Value *Dst = CI->getArgOperand(i: 0), *Src = CI->getArgOperand(i: 1); |
704 | if (Dst == Src) // strcpy(x,x) -> x |
705 | return Src; |
706 | |
707 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: {0, 1}); |
708 | // See if we can get the length of the input string. |
709 | uint64_t Len = GetStringLength(V: Src); |
710 | if (Len) |
711 | annotateDereferenceableBytes(CI, ArgNos: 1, DereferenceableBytes: Len); |
712 | else |
713 | return nullptr; |
714 | |
715 | // We have enough information to now generate the memcpy call to do the |
716 | // copy for us. Make a memcpy to copy the nul byte with align = 1. |
717 | CallInst *NewCI = B.CreateMemCpy(Dst, DstAlign: Align(1), Src, SrcAlign: Align(1), |
718 | Size: TLI->getAsSizeT(V: Len, M: *CI->getModule())); |
719 | mergeAttributesAndFlags(NewCI, Old: *CI); |
720 | return Dst; |
721 | } |
722 | |
723 | Value *LibCallSimplifier::optimizeStpCpy(CallInst *CI, IRBuilderBase &B) { |
724 | Value *Dst = CI->getArgOperand(i: 0), *Src = CI->getArgOperand(i: 1); |
725 | |
726 | // stpcpy(d,s) -> strcpy(d,s) if the result is not used. |
727 | if (CI->use_empty()) |
728 | return copyFlags(Old: *CI, New: emitStrCpy(Dst, Src, B, TLI)); |
729 | |
730 | if (Dst == Src) { // stpcpy(x,x) -> x+strlen(x) |
731 | Value *StrLen = emitStrLen(Ptr: Src, B, DL, TLI); |
732 | return StrLen ? B.CreateInBoundsGEP(Ty: B.getInt8Ty(), Ptr: Dst, IdxList: StrLen) : nullptr; |
733 | } |
734 | |
735 | // See if we can get the length of the input string. |
736 | uint64_t Len = GetStringLength(V: Src); |
737 | if (Len) |
738 | annotateDereferenceableBytes(CI, ArgNos: 1, DereferenceableBytes: Len); |
739 | else |
740 | return nullptr; |
741 | |
742 | Value *LenV = TLI->getAsSizeT(V: Len, M: *CI->getModule()); |
743 | Value *DstEnd = B.CreateInBoundsGEP( |
744 | Ty: B.getInt8Ty(), Ptr: Dst, IdxList: TLI->getAsSizeT(V: Len - 1, M: *CI->getModule())); |
745 | |
746 | // We have enough information to now generate the memcpy call to do the |
747 | // copy for us. Make a memcpy to copy the nul byte with align = 1. |
748 | CallInst *NewCI = B.CreateMemCpy(Dst, DstAlign: Align(1), Src, SrcAlign: Align(1), Size: LenV); |
749 | mergeAttributesAndFlags(NewCI, Old: *CI); |
750 | return DstEnd; |
751 | } |
752 | |
753 | // Optimize a call to size_t strlcpy(char*, const char*, size_t). |
754 | |
755 | Value *LibCallSimplifier::optimizeStrLCpy(CallInst *CI, IRBuilderBase &B) { |
756 | Value *Size = CI->getArgOperand(i: 2); |
757 | if (isKnownNonZero(V: Size, Q: DL)) |
758 | // Like snprintf, the function stores into the destination only when |
759 | // the size argument is nonzero. |
760 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: 0); |
761 | // The function reads the source argument regardless of Size (it returns |
762 | // its length). |
763 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: 1); |
764 | |
765 | uint64_t NBytes; |
766 | if (ConstantInt *SizeC = dyn_cast<ConstantInt>(Val: Size)) |
767 | NBytes = SizeC->getZExtValue(); |
768 | else |
769 | return nullptr; |
770 | |
771 | Value *Dst = CI->getArgOperand(i: 0); |
772 | Value *Src = CI->getArgOperand(i: 1); |
773 | if (NBytes <= 1) { |
774 | if (NBytes == 1) |
775 | // For a call to strlcpy(D, S, 1) first store a nul in *D. |
776 | B.CreateStore(Val: B.getInt8(C: 0), Ptr: Dst); |
777 | |
778 | // Transform strlcpy(D, S, 0) to a call to strlen(S). |
779 | return copyFlags(Old: *CI, New: emitStrLen(Ptr: Src, B, DL, TLI)); |
780 | } |
781 | |
782 | // Try to determine the length of the source, substituting its size |
783 | // when it's not nul-terminated (as it's required to be) to avoid |
784 | // reading past its end. |
785 | StringRef Str; |
786 | if (!getConstantStringInfo(V: Src, Str, /*TrimAtNul=*/false)) |
787 | return nullptr; |
788 | |
789 | uint64_t SrcLen = Str.find(C: '\0'); |
790 | // Set if the terminating nul should be copied by the call to memcpy |
791 | // below. |
792 | bool NulTerm = SrcLen < NBytes; |
793 | |
794 | if (NulTerm) |
795 | // Overwrite NBytes with the number of bytes to copy, including |
796 | // the terminating nul. |
797 | NBytes = SrcLen + 1; |
798 | else { |
799 | // Set the length of the source for the function to return to its |
800 | // size, and cap NBytes at the same. |
801 | SrcLen = std::min(a: SrcLen, b: uint64_t(Str.size())); |
802 | NBytes = std::min(a: NBytes - 1, b: SrcLen); |
803 | } |
804 | |
805 | if (SrcLen == 0) { |
806 | // Transform strlcpy(D, "", N) to (*D = '\0, 0). |
807 | B.CreateStore(Val: B.getInt8(C: 0), Ptr: Dst); |
808 | return ConstantInt::get(Ty: CI->getType(), V: 0); |
809 | } |
810 | |
811 | // Transform strlcpy(D, S, N) to memcpy(D, S, N') where N' is the lower |
812 | // bound on strlen(S) + 1 and N, optionally followed by a nul store to |
813 | // D[N' - 1] if necessary. |
814 | CallInst *NewCI = B.CreateMemCpy(Dst, DstAlign: Align(1), Src, SrcAlign: Align(1), |
815 | Size: TLI->getAsSizeT(V: NBytes, M: *CI->getModule())); |
816 | mergeAttributesAndFlags(NewCI, Old: *CI); |
817 | |
818 | if (!NulTerm) { |
819 | Value *EndOff = ConstantInt::get(Ty: CI->getType(), V: NBytes); |
820 | Value *EndPtr = B.CreateInBoundsGEP(Ty: B.getInt8Ty(), Ptr: Dst, IdxList: EndOff); |
821 | B.CreateStore(Val: B.getInt8(C: 0), Ptr: EndPtr); |
822 | } |
823 | |
824 | // Like snprintf, strlcpy returns the number of nonzero bytes that would |
825 | // have been copied if the bound had been sufficiently big (which in this |
826 | // case is strlen(Src)). |
827 | return ConstantInt::get(Ty: CI->getType(), V: SrcLen); |
828 | } |
829 | |
830 | // Optimize a call CI to either stpncpy when RetEnd is true, or to strncpy |
831 | // otherwise. |
832 | Value *LibCallSimplifier::optimizeStringNCpy(CallInst *CI, bool RetEnd, |
833 | IRBuilderBase &B) { |
834 | Value *Dst = CI->getArgOperand(i: 0); |
835 | Value *Src = CI->getArgOperand(i: 1); |
836 | Value *Size = CI->getArgOperand(i: 2); |
837 | |
838 | if (isKnownNonZero(V: Size, Q: DL)) { |
839 | // Both st{p,r}ncpy(D, S, N) access the source and destination arrays |
840 | // only when N is nonzero. |
841 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: 0); |
842 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: 1); |
843 | } |
844 | |
845 | // If the "bound" argument is known set N to it. Otherwise set it to |
846 | // UINT64_MAX and handle it later. |
847 | uint64_t N = UINT64_MAX; |
848 | if (ConstantInt *SizeC = dyn_cast<ConstantInt>(Val: Size)) |
849 | N = SizeC->getZExtValue(); |
850 | |
851 | if (N == 0) |
852 | // Fold st{p,r}ncpy(D, S, 0) to D. |
853 | return Dst; |
854 | |
855 | if (N == 1) { |
856 | Type *CharTy = B.getInt8Ty(); |
857 | Value *CharVal = B.CreateLoad(Ty: CharTy, Ptr: Src, Name: "stxncpy.char0" ); |
858 | B.CreateStore(Val: CharVal, Ptr: Dst); |
859 | if (!RetEnd) |
860 | // Transform strncpy(D, S, 1) to return (*D = *S), D. |
861 | return Dst; |
862 | |
863 | // Transform stpncpy(D, S, 1) to return (*D = *S) ? D + 1 : D. |
864 | Value *ZeroChar = ConstantInt::get(Ty: CharTy, V: 0); |
865 | Value *Cmp = B.CreateICmpEQ(LHS: CharVal, RHS: ZeroChar, Name: "stpncpy.char0cmp" ); |
866 | |
867 | Value *Off1 = B.getInt32(C: 1); |
868 | Value *EndPtr = B.CreateInBoundsGEP(Ty: CharTy, Ptr: Dst, IdxList: Off1, Name: "stpncpy.end" ); |
869 | return B.CreateSelect(C: Cmp, True: Dst, False: EndPtr, Name: "stpncpy.sel" ); |
870 | } |
871 | |
872 | // If the length of the input string is known set SrcLen to it. |
873 | uint64_t SrcLen = GetStringLength(V: Src); |
874 | if (SrcLen) |
875 | annotateDereferenceableBytes(CI, ArgNos: 1, DereferenceableBytes: SrcLen); |
876 | else |
877 | return nullptr; |
878 | |
879 | --SrcLen; // Unbias length. |
880 | |
881 | if (SrcLen == 0) { |
882 | // Transform st{p,r}ncpy(D, "", N) to memset(D, '\0', N) for any N. |
883 | Align MemSetAlign = |
884 | CI->getAttributes().getParamAttrs(ArgNo: 0).getAlignment().valueOrOne(); |
885 | CallInst *NewCI = B.CreateMemSet(Ptr: Dst, Val: B.getInt8(C: '\0'), Size, Align: MemSetAlign); |
886 | AttrBuilder ArgAttrs(CI->getContext(), CI->getAttributes().getParamAttrs(ArgNo: 0)); |
887 | NewCI->setAttributes(NewCI->getAttributes().addParamAttributes( |
888 | C&: CI->getContext(), ArgNo: 0, B: ArgAttrs)); |
889 | copyFlags(Old: *CI, New: NewCI); |
890 | return Dst; |
891 | } |
892 | |
893 | if (N > SrcLen + 1) { |
894 | if (N > 128) |
895 | // Bail if N is large or unknown. |
896 | return nullptr; |
897 | |
898 | // st{p,r}ncpy(D, "a", N) -> memcpy(D, "a\0\0\0", N) for N <= 128. |
899 | StringRef Str; |
900 | if (!getConstantStringInfo(V: Src, Str)) |
901 | return nullptr; |
902 | std::string SrcStr = Str.str(); |
903 | // Create a bigger, nul-padded array with the same length, SrcLen, |
904 | // as the original string. |
905 | SrcStr.resize(n: N, c: '\0'); |
906 | Src = B.CreateGlobalString(Str: SrcStr, Name: "str" , /*AddressSpace=*/0, |
907 | /*M=*/nullptr, /*AddNull=*/false); |
908 | } |
909 | |
910 | // st{p,r}ncpy(D, S, N) -> memcpy(align 1 D, align 1 S, N) when both |
911 | // S and N are constant. |
912 | CallInst *NewCI = B.CreateMemCpy(Dst, DstAlign: Align(1), Src, SrcAlign: Align(1), |
913 | Size: TLI->getAsSizeT(V: N, M: *CI->getModule())); |
914 | mergeAttributesAndFlags(NewCI, Old: *CI); |
915 | if (!RetEnd) |
916 | return Dst; |
917 | |
918 | // stpncpy(D, S, N) returns the address of the first null in D if it writes |
919 | // one, otherwise D + N. |
920 | Value *Off = B.getInt64(C: std::min(a: SrcLen, b: N)); |
921 | return B.CreateInBoundsGEP(Ty: B.getInt8Ty(), Ptr: Dst, IdxList: Off, Name: "endptr" ); |
922 | } |
923 | |
924 | Value *LibCallSimplifier::optimizeStringLength(CallInst *CI, IRBuilderBase &B, |
925 | unsigned CharSize, |
926 | Value *Bound) { |
927 | Value *Src = CI->getArgOperand(i: 0); |
928 | Type *CharTy = B.getIntNTy(N: CharSize); |
929 | |
930 | if (isOnlyUsedInZeroEqualityComparison(CxtI: CI) && |
931 | (!Bound || isKnownNonZero(V: Bound, Q: DL))) { |
932 | // Fold strlen: |
933 | // strlen(x) != 0 --> *x != 0 |
934 | // strlen(x) == 0 --> *x == 0 |
935 | // and likewise strnlen with constant N > 0: |
936 | // strnlen(x, N) != 0 --> *x != 0 |
937 | // strnlen(x, N) == 0 --> *x == 0 |
938 | return B.CreateZExt(V: B.CreateLoad(Ty: CharTy, Ptr: Src, Name: "char0" ), |
939 | DestTy: CI->getType()); |
940 | } |
941 | |
942 | if (Bound) { |
943 | if (ConstantInt *BoundCst = dyn_cast<ConstantInt>(Val: Bound)) { |
944 | if (BoundCst->isZero()) |
945 | // Fold strnlen(s, 0) -> 0 for any s, constant or otherwise. |
946 | return ConstantInt::get(Ty: CI->getType(), V: 0); |
947 | |
948 | if (BoundCst->isOne()) { |
949 | // Fold strnlen(s, 1) -> *s ? 1 : 0 for any s. |
950 | Value *CharVal = B.CreateLoad(Ty: CharTy, Ptr: Src, Name: "strnlen.char0" ); |
951 | Value *ZeroChar = ConstantInt::get(Ty: CharTy, V: 0); |
952 | Value *Cmp = B.CreateICmpNE(LHS: CharVal, RHS: ZeroChar, Name: "strnlen.char0cmp" ); |
953 | return B.CreateZExt(V: Cmp, DestTy: CI->getType()); |
954 | } |
955 | } |
956 | } |
957 | |
958 | if (uint64_t Len = GetStringLength(V: Src, CharSize)) { |
959 | Value *LenC = ConstantInt::get(Ty: CI->getType(), V: Len - 1); |
960 | // Fold strlen("xyz") -> 3 and strnlen("xyz", 2) -> 2 |
961 | // and strnlen("xyz", Bound) -> min(3, Bound) for nonconstant Bound. |
962 | if (Bound) |
963 | return B.CreateBinaryIntrinsic(ID: Intrinsic::umin, LHS: LenC, RHS: Bound); |
964 | return LenC; |
965 | } |
966 | |
967 | if (Bound) |
968 | // Punt for strnlen for now. |
969 | return nullptr; |
970 | |
971 | // If s is a constant pointer pointing to a string literal, we can fold |
972 | // strlen(s + x) to strlen(s) - x, when x is known to be in the range |
973 | // [0, strlen(s)] or the string has a single null terminator '\0' at the end. |
974 | // We only try to simplify strlen when the pointer s points to an array |
975 | // of CharSize elements. Otherwise, we would need to scale the offset x before |
976 | // doing the subtraction. This will make the optimization more complex, and |
977 | // it's not very useful because calling strlen for a pointer of other types is |
978 | // very uncommon. |
979 | if (GEPOperator *GEP = dyn_cast<GEPOperator>(Val: Src)) { |
980 | // TODO: Handle subobjects. |
981 | if (!isGEPBasedOnPointerToString(GEP, CharSize)) |
982 | return nullptr; |
983 | |
984 | ConstantDataArraySlice Slice; |
985 | if (getConstantDataArrayInfo(V: GEP->getOperand(i_nocapture: 0), Slice, ElementSize: CharSize)) { |
986 | uint64_t NullTermIdx; |
987 | if (Slice.Array == nullptr) { |
988 | NullTermIdx = 0; |
989 | } else { |
990 | NullTermIdx = ~((uint64_t)0); |
991 | for (uint64_t I = 0, E = Slice.Length; I < E; ++I) { |
992 | if (Slice.Array->getElementAsInteger(i: I + Slice.Offset) == 0) { |
993 | NullTermIdx = I; |
994 | break; |
995 | } |
996 | } |
997 | // If the string does not have '\0', leave it to strlen to compute |
998 | // its length. |
999 | if (NullTermIdx == ~((uint64_t)0)) |
1000 | return nullptr; |
1001 | } |
1002 | |
1003 | Value *Offset = GEP->getOperand(i_nocapture: 2); |
1004 | KnownBits Known = computeKnownBits(V: Offset, DL, AC: nullptr, CxtI: CI, DT: nullptr); |
1005 | uint64_t ArrSize = |
1006 | cast<ArrayType>(Val: GEP->getSourceElementType())->getNumElements(); |
1007 | |
1008 | // If Offset is not provably in the range [0, NullTermIdx], we can still |
1009 | // optimize if we can prove that the program has undefined behavior when |
1010 | // Offset is outside that range. That is the case when GEP->getOperand(0) |
1011 | // is a pointer to an object whose memory extent is NullTermIdx+1. |
1012 | if ((Known.isNonNegative() && Known.getMaxValue().ule(RHS: NullTermIdx)) || |
1013 | (isa<GlobalVariable>(Val: GEP->getOperand(i_nocapture: 0)) && |
1014 | NullTermIdx == ArrSize - 1)) { |
1015 | Offset = B.CreateSExtOrTrunc(V: Offset, DestTy: CI->getType()); |
1016 | return B.CreateSub(LHS: ConstantInt::get(Ty: CI->getType(), V: NullTermIdx), |
1017 | RHS: Offset); |
1018 | } |
1019 | } |
1020 | } |
1021 | |
1022 | // strlen(x?"foo":"bars") --> x ? 3 : 4 |
1023 | if (SelectInst *SI = dyn_cast<SelectInst>(Val: Src)) { |
1024 | uint64_t LenTrue = GetStringLength(V: SI->getTrueValue(), CharSize); |
1025 | uint64_t LenFalse = GetStringLength(V: SI->getFalseValue(), CharSize); |
1026 | if (LenTrue && LenFalse) { |
1027 | ORE.emit(RemarkBuilder: [&]() { |
1028 | return OptimizationRemark("instcombine" , "simplify-libcalls" , CI) |
1029 | << "folded strlen(select) to select of constants" ; |
1030 | }); |
1031 | return B.CreateSelect(C: SI->getCondition(), |
1032 | True: ConstantInt::get(Ty: CI->getType(), V: LenTrue - 1), |
1033 | False: ConstantInt::get(Ty: CI->getType(), V: LenFalse - 1)); |
1034 | } |
1035 | } |
1036 | |
1037 | return nullptr; |
1038 | } |
1039 | |
1040 | Value *LibCallSimplifier::optimizeStrLen(CallInst *CI, IRBuilderBase &B) { |
1041 | if (Value *V = optimizeStringLength(CI, B, CharSize: 8)) |
1042 | return V; |
1043 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: 0); |
1044 | return nullptr; |
1045 | } |
1046 | |
1047 | Value *LibCallSimplifier::optimizeStrNLen(CallInst *CI, IRBuilderBase &B) { |
1048 | Value *Bound = CI->getArgOperand(i: 1); |
1049 | if (Value *V = optimizeStringLength(CI, B, CharSize: 8, Bound)) |
1050 | return V; |
1051 | |
1052 | if (isKnownNonZero(V: Bound, Q: DL)) |
1053 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: 0); |
1054 | return nullptr; |
1055 | } |
1056 | |
1057 | Value *LibCallSimplifier::optimizeWcslen(CallInst *CI, IRBuilderBase &B) { |
1058 | Module &M = *CI->getModule(); |
1059 | unsigned WCharSize = TLI->getWCharSize(M) * 8; |
1060 | // We cannot perform this optimization without wchar_size metadata. |
1061 | if (WCharSize == 0) |
1062 | return nullptr; |
1063 | |
1064 | return optimizeStringLength(CI, B, CharSize: WCharSize); |
1065 | } |
1066 | |
1067 | Value *LibCallSimplifier::optimizeStrPBrk(CallInst *CI, IRBuilderBase &B) { |
1068 | StringRef S1, S2; |
1069 | bool HasS1 = getConstantStringInfo(V: CI->getArgOperand(i: 0), Str&: S1); |
1070 | bool HasS2 = getConstantStringInfo(V: CI->getArgOperand(i: 1), Str&: S2); |
1071 | |
1072 | // strpbrk(s, "") -> nullptr |
1073 | // strpbrk("", s) -> nullptr |
1074 | if ((HasS1 && S1.empty()) || (HasS2 && S2.empty())) |
1075 | return Constant::getNullValue(Ty: CI->getType()); |
1076 | |
1077 | // Constant folding. |
1078 | if (HasS1 && HasS2) { |
1079 | size_t I = S1.find_first_of(Chars: S2); |
1080 | if (I == StringRef::npos) // No match. |
1081 | return Constant::getNullValue(Ty: CI->getType()); |
1082 | |
1083 | return B.CreateInBoundsGEP(Ty: B.getInt8Ty(), Ptr: CI->getArgOperand(i: 0), |
1084 | IdxList: B.getInt64(C: I), Name: "strpbrk" ); |
1085 | } |
1086 | |
1087 | // strpbrk(s, "a") -> strchr(s, 'a') |
1088 | if (HasS2 && S2.size() == 1) |
1089 | return copyFlags(Old: *CI, New: emitStrChr(Ptr: CI->getArgOperand(i: 0), C: S2[0], B, TLI)); |
1090 | |
1091 | return nullptr; |
1092 | } |
1093 | |
1094 | Value *LibCallSimplifier::optimizeStrTo(CallInst *CI, IRBuilderBase &B) { |
1095 | Value *EndPtr = CI->getArgOperand(i: 1); |
1096 | if (isa<ConstantPointerNull>(Val: EndPtr)) { |
1097 | // With a null EndPtr, this function won't capture the main argument. |
1098 | // It would be readonly too, except that it still may write to errno. |
1099 | CI->addParamAttr(ArgNo: 0, Attr: Attribute::getWithCaptureInfo(Context&: CI->getContext(), |
1100 | CI: CaptureInfo::none())); |
1101 | } |
1102 | |
1103 | return nullptr; |
1104 | } |
1105 | |
1106 | Value *LibCallSimplifier::optimizeStrSpn(CallInst *CI, IRBuilderBase &B) { |
1107 | StringRef S1, S2; |
1108 | bool HasS1 = getConstantStringInfo(V: CI->getArgOperand(i: 0), Str&: S1); |
1109 | bool HasS2 = getConstantStringInfo(V: CI->getArgOperand(i: 1), Str&: S2); |
1110 | |
1111 | // strspn(s, "") -> 0 |
1112 | // strspn("", s) -> 0 |
1113 | if ((HasS1 && S1.empty()) || (HasS2 && S2.empty())) |
1114 | return Constant::getNullValue(Ty: CI->getType()); |
1115 | |
1116 | // Constant folding. |
1117 | if (HasS1 && HasS2) { |
1118 | size_t Pos = S1.find_first_not_of(Chars: S2); |
1119 | if (Pos == StringRef::npos) |
1120 | Pos = S1.size(); |
1121 | return ConstantInt::get(Ty: CI->getType(), V: Pos); |
1122 | } |
1123 | |
1124 | return nullptr; |
1125 | } |
1126 | |
1127 | Value *LibCallSimplifier::optimizeStrCSpn(CallInst *CI, IRBuilderBase &B) { |
1128 | StringRef S1, S2; |
1129 | bool HasS1 = getConstantStringInfo(V: CI->getArgOperand(i: 0), Str&: S1); |
1130 | bool HasS2 = getConstantStringInfo(V: CI->getArgOperand(i: 1), Str&: S2); |
1131 | |
1132 | // strcspn("", s) -> 0 |
1133 | if (HasS1 && S1.empty()) |
1134 | return Constant::getNullValue(Ty: CI->getType()); |
1135 | |
1136 | // Constant folding. |
1137 | if (HasS1 && HasS2) { |
1138 | size_t Pos = S1.find_first_of(Chars: S2); |
1139 | if (Pos == StringRef::npos) |
1140 | Pos = S1.size(); |
1141 | return ConstantInt::get(Ty: CI->getType(), V: Pos); |
1142 | } |
1143 | |
1144 | // strcspn(s, "") -> strlen(s) |
1145 | if (HasS2 && S2.empty()) |
1146 | return copyFlags(Old: *CI, New: emitStrLen(Ptr: CI->getArgOperand(i: 0), B, DL, TLI)); |
1147 | |
1148 | return nullptr; |
1149 | } |
1150 | |
1151 | Value *LibCallSimplifier::optimizeStrStr(CallInst *CI, IRBuilderBase &B) { |
1152 | // fold strstr(x, x) -> x. |
1153 | if (CI->getArgOperand(i: 0) == CI->getArgOperand(i: 1)) |
1154 | return CI->getArgOperand(i: 0); |
1155 | |
1156 | // fold strstr(a, b) == a -> strncmp(a, b, strlen(b)) == 0 |
1157 | if (isOnlyUsedInEqualityComparison(V: CI, With: CI->getArgOperand(i: 0))) { |
1158 | Value *StrLen = emitStrLen(Ptr: CI->getArgOperand(i: 1), B, DL, TLI); |
1159 | if (!StrLen) |
1160 | return nullptr; |
1161 | Value *StrNCmp = emitStrNCmp(Ptr1: CI->getArgOperand(i: 0), Ptr2: CI->getArgOperand(i: 1), |
1162 | Len: StrLen, B, DL, TLI); |
1163 | if (!StrNCmp) |
1164 | return nullptr; |
1165 | for (User *U : llvm::make_early_inc_range(Range: CI->users())) { |
1166 | ICmpInst *Old = cast<ICmpInst>(Val: U); |
1167 | Value *Cmp = |
1168 | B.CreateICmp(P: Old->getPredicate(), LHS: StrNCmp, |
1169 | RHS: ConstantInt::getNullValue(Ty: StrNCmp->getType()), Name: "cmp" ); |
1170 | replaceAllUsesWith(I: Old, With: Cmp); |
1171 | } |
1172 | return CI; |
1173 | } |
1174 | |
1175 | // See if either input string is a constant string. |
1176 | StringRef SearchStr, ToFindStr; |
1177 | bool HasStr1 = getConstantStringInfo(V: CI->getArgOperand(i: 0), Str&: SearchStr); |
1178 | bool HasStr2 = getConstantStringInfo(V: CI->getArgOperand(i: 1), Str&: ToFindStr); |
1179 | |
1180 | // fold strstr(x, "") -> x. |
1181 | if (HasStr2 && ToFindStr.empty()) |
1182 | return CI->getArgOperand(i: 0); |
1183 | |
1184 | // If both strings are known, constant fold it. |
1185 | if (HasStr1 && HasStr2) { |
1186 | size_t Offset = SearchStr.find(Str: ToFindStr); |
1187 | |
1188 | if (Offset == StringRef::npos) // strstr("foo", "bar") -> null |
1189 | return Constant::getNullValue(Ty: CI->getType()); |
1190 | |
1191 | // strstr("abcd", "bc") -> gep((char*)"abcd", 1) |
1192 | return B.CreateConstInBoundsGEP1_64(Ty: B.getInt8Ty(), Ptr: CI->getArgOperand(i: 0), |
1193 | Idx0: Offset, Name: "strstr" ); |
1194 | } |
1195 | |
1196 | // fold strstr(x, "y") -> strchr(x, 'y'). |
1197 | if (HasStr2 && ToFindStr.size() == 1) { |
1198 | return emitStrChr(Ptr: CI->getArgOperand(i: 0), C: ToFindStr[0], B, TLI); |
1199 | } |
1200 | |
1201 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: {0, 1}); |
1202 | return nullptr; |
1203 | } |
1204 | |
1205 | Value *LibCallSimplifier::optimizeMemRChr(CallInst *CI, IRBuilderBase &B) { |
1206 | Value *SrcStr = CI->getArgOperand(i: 0); |
1207 | Value *Size = CI->getArgOperand(i: 2); |
1208 | annotateNonNullAndDereferenceable(CI, ArgNos: 0, Size, DL); |
1209 | Value *CharVal = CI->getArgOperand(i: 1); |
1210 | ConstantInt *LenC = dyn_cast<ConstantInt>(Val: Size); |
1211 | Value *NullPtr = Constant::getNullValue(Ty: CI->getType()); |
1212 | |
1213 | if (LenC) { |
1214 | if (LenC->isZero()) |
1215 | // Fold memrchr(x, y, 0) --> null. |
1216 | return NullPtr; |
1217 | |
1218 | if (LenC->isOne()) { |
1219 | // Fold memrchr(x, y, 1) --> *x == y ? x : null for any x and y, |
1220 | // constant or otherwise. |
1221 | Value *Val = B.CreateLoad(Ty: B.getInt8Ty(), Ptr: SrcStr, Name: "memrchr.char0" ); |
1222 | // Slice off the character's high end bits. |
1223 | CharVal = B.CreateTrunc(V: CharVal, DestTy: B.getInt8Ty()); |
1224 | Value *Cmp = B.CreateICmpEQ(LHS: Val, RHS: CharVal, Name: "memrchr.char0cmp" ); |
1225 | return B.CreateSelect(C: Cmp, True: SrcStr, False: NullPtr, Name: "memrchr.sel" ); |
1226 | } |
1227 | } |
1228 | |
1229 | StringRef Str; |
1230 | if (!getConstantStringInfo(V: SrcStr, Str, /*TrimAtNul=*/false)) |
1231 | return nullptr; |
1232 | |
1233 | if (Str.size() == 0) |
1234 | // If the array is empty fold memrchr(A, C, N) to null for any value |
1235 | // of C and N on the basis that the only valid value of N is zero |
1236 | // (otherwise the call is undefined). |
1237 | return NullPtr; |
1238 | |
1239 | uint64_t EndOff = UINT64_MAX; |
1240 | if (LenC) { |
1241 | EndOff = LenC->getZExtValue(); |
1242 | if (Str.size() < EndOff) |
1243 | // Punt out-of-bounds accesses to sanitizers and/or libc. |
1244 | return nullptr; |
1245 | } |
1246 | |
1247 | if (ConstantInt *CharC = dyn_cast<ConstantInt>(Val: CharVal)) { |
1248 | // Fold memrchr(S, C, N) for a constant C. |
1249 | size_t Pos = Str.rfind(C: CharC->getZExtValue(), From: EndOff); |
1250 | if (Pos == StringRef::npos) |
1251 | // When the character is not in the source array fold the result |
1252 | // to null regardless of Size. |
1253 | return NullPtr; |
1254 | |
1255 | if (LenC) |
1256 | // Fold memrchr(s, c, N) --> s + Pos for constant N > Pos. |
1257 | return B.CreateInBoundsGEP(Ty: B.getInt8Ty(), Ptr: SrcStr, IdxList: B.getInt64(C: Pos)); |
1258 | |
1259 | if (Str.find(C: Str[Pos]) == Pos) { |
1260 | // When there is just a single occurrence of C in S, i.e., the one |
1261 | // in Str[Pos], fold |
1262 | // memrchr(s, c, N) --> N <= Pos ? null : s + Pos |
1263 | // for nonconstant N. |
1264 | Value *Cmp = B.CreateICmpULE(LHS: Size, RHS: ConstantInt::get(Ty: Size->getType(), V: Pos), |
1265 | Name: "memrchr.cmp" ); |
1266 | Value *SrcPlus = B.CreateInBoundsGEP(Ty: B.getInt8Ty(), Ptr: SrcStr, |
1267 | IdxList: B.getInt64(C: Pos), Name: "memrchr.ptr_plus" ); |
1268 | return B.CreateSelect(C: Cmp, True: NullPtr, False: SrcPlus, Name: "memrchr.sel" ); |
1269 | } |
1270 | } |
1271 | |
1272 | // Truncate the string to search at most EndOff characters. |
1273 | Str = Str.substr(Start: 0, N: EndOff); |
1274 | if (Str.find_first_not_of(C: Str[0]) != StringRef::npos) |
1275 | return nullptr; |
1276 | |
1277 | // If the source array consists of all equal characters, then for any |
1278 | // C and N (whether in bounds or not), fold memrchr(S, C, N) to |
1279 | // N != 0 && *S == C ? S + N - 1 : null |
1280 | Type *SizeTy = Size->getType(); |
1281 | Type *Int8Ty = B.getInt8Ty(); |
1282 | Value *NNeZ = B.CreateICmpNE(LHS: Size, RHS: ConstantInt::get(Ty: SizeTy, V: 0)); |
1283 | // Slice off the sought character's high end bits. |
1284 | CharVal = B.CreateTrunc(V: CharVal, DestTy: Int8Ty); |
1285 | Value *CEqS0 = B.CreateICmpEQ(LHS: ConstantInt::get(Ty: Int8Ty, V: Str[0]), RHS: CharVal); |
1286 | Value *And = B.CreateLogicalAnd(Cond1: NNeZ, Cond2: CEqS0); |
1287 | Value *SizeM1 = B.CreateSub(LHS: Size, RHS: ConstantInt::get(Ty: SizeTy, V: 1)); |
1288 | Value *SrcPlus = |
1289 | B.CreateInBoundsGEP(Ty: Int8Ty, Ptr: SrcStr, IdxList: SizeM1, Name: "memrchr.ptr_plus" ); |
1290 | return B.CreateSelect(C: And, True: SrcPlus, False: NullPtr, Name: "memrchr.sel" ); |
1291 | } |
1292 | |
1293 | Value *LibCallSimplifier::optimizeMemChr(CallInst *CI, IRBuilderBase &B) { |
1294 | Value *SrcStr = CI->getArgOperand(i: 0); |
1295 | Value *Size = CI->getArgOperand(i: 2); |
1296 | |
1297 | if (isKnownNonZero(V: Size, Q: DL)) { |
1298 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: 0); |
1299 | if (isOnlyUsedInEqualityComparison(V: CI, With: SrcStr)) |
1300 | return memChrToCharCompare(CI, NBytes: Size, B, DL); |
1301 | } |
1302 | |
1303 | Value *CharVal = CI->getArgOperand(i: 1); |
1304 | ConstantInt *CharC = dyn_cast<ConstantInt>(Val: CharVal); |
1305 | ConstantInt *LenC = dyn_cast<ConstantInt>(Val: Size); |
1306 | Value *NullPtr = Constant::getNullValue(Ty: CI->getType()); |
1307 | |
1308 | // memchr(x, y, 0) -> null |
1309 | if (LenC) { |
1310 | if (LenC->isZero()) |
1311 | return NullPtr; |
1312 | |
1313 | if (LenC->isOne()) { |
1314 | // Fold memchr(x, y, 1) --> *x == y ? x : null for any x and y, |
1315 | // constant or otherwise. |
1316 | Value *Val = B.CreateLoad(Ty: B.getInt8Ty(), Ptr: SrcStr, Name: "memchr.char0" ); |
1317 | // Slice off the character's high end bits. |
1318 | CharVal = B.CreateTrunc(V: CharVal, DestTy: B.getInt8Ty()); |
1319 | Value *Cmp = B.CreateICmpEQ(LHS: Val, RHS: CharVal, Name: "memchr.char0cmp" ); |
1320 | return B.CreateSelect(C: Cmp, True: SrcStr, False: NullPtr, Name: "memchr.sel" ); |
1321 | } |
1322 | } |
1323 | |
1324 | StringRef Str; |
1325 | if (!getConstantStringInfo(V: SrcStr, Str, /*TrimAtNul=*/false)) |
1326 | return nullptr; |
1327 | |
1328 | if (CharC) { |
1329 | size_t Pos = Str.find(C: CharC->getZExtValue()); |
1330 | if (Pos == StringRef::npos) |
1331 | // When the character is not in the source array fold the result |
1332 | // to null regardless of Size. |
1333 | return NullPtr; |
1334 | |
1335 | // Fold memchr(s, c, n) -> n <= Pos ? null : s + Pos |
1336 | // When the constant Size is less than or equal to the character |
1337 | // position also fold the result to null. |
1338 | Value *Cmp = B.CreateICmpULE(LHS: Size, RHS: ConstantInt::get(Ty: Size->getType(), V: Pos), |
1339 | Name: "memchr.cmp" ); |
1340 | Value *SrcPlus = B.CreateInBoundsGEP(Ty: B.getInt8Ty(), Ptr: SrcStr, IdxList: B.getInt64(C: Pos), |
1341 | Name: "memchr.ptr" ); |
1342 | return B.CreateSelect(C: Cmp, True: NullPtr, False: SrcPlus); |
1343 | } |
1344 | |
1345 | if (Str.size() == 0) |
1346 | // If the array is empty fold memchr(A, C, N) to null for any value |
1347 | // of C and N on the basis that the only valid value of N is zero |
1348 | // (otherwise the call is undefined). |
1349 | return NullPtr; |
1350 | |
1351 | if (LenC) |
1352 | Str = substr(Str, Len: LenC->getZExtValue()); |
1353 | |
1354 | size_t Pos = Str.find_first_not_of(C: Str[0]); |
1355 | if (Pos == StringRef::npos |
1356 | || Str.find_first_not_of(C: Str[Pos], From: Pos) == StringRef::npos) { |
1357 | // If the source array consists of at most two consecutive sequences |
1358 | // of the same characters, then for any C and N (whether in bounds or |
1359 | // not), fold memchr(S, C, N) to |
1360 | // N != 0 && *S == C ? S : null |
1361 | // or for the two sequences to: |
1362 | // N != 0 && *S == C ? S : (N > Pos && S[Pos] == C ? S + Pos : null) |
1363 | // ^Sel2 ^Sel1 are denoted above. |
1364 | // The latter makes it also possible to fold strchr() calls with strings |
1365 | // of the same characters. |
1366 | Type *SizeTy = Size->getType(); |
1367 | Type *Int8Ty = B.getInt8Ty(); |
1368 | |
1369 | // Slice off the sought character's high end bits. |
1370 | CharVal = B.CreateTrunc(V: CharVal, DestTy: Int8Ty); |
1371 | |
1372 | Value *Sel1 = NullPtr; |
1373 | if (Pos != StringRef::npos) { |
1374 | // Handle two consecutive sequences of the same characters. |
1375 | Value *PosVal = ConstantInt::get(Ty: SizeTy, V: Pos); |
1376 | Value *StrPos = ConstantInt::get(Ty: Int8Ty, V: Str[Pos]); |
1377 | Value *CEqSPos = B.CreateICmpEQ(LHS: CharVal, RHS: StrPos); |
1378 | Value *NGtPos = B.CreateICmp(P: ICmpInst::ICMP_UGT, LHS: Size, RHS: PosVal); |
1379 | Value *And = B.CreateAnd(LHS: CEqSPos, RHS: NGtPos); |
1380 | Value *SrcPlus = B.CreateInBoundsGEP(Ty: B.getInt8Ty(), Ptr: SrcStr, IdxList: PosVal); |
1381 | Sel1 = B.CreateSelect(C: And, True: SrcPlus, False: NullPtr, Name: "memchr.sel1" ); |
1382 | } |
1383 | |
1384 | Value *Str0 = ConstantInt::get(Ty: Int8Ty, V: Str[0]); |
1385 | Value *CEqS0 = B.CreateICmpEQ(LHS: Str0, RHS: CharVal); |
1386 | Value *NNeZ = B.CreateICmpNE(LHS: Size, RHS: ConstantInt::get(Ty: SizeTy, V: 0)); |
1387 | Value *And = B.CreateAnd(LHS: NNeZ, RHS: CEqS0); |
1388 | return B.CreateSelect(C: And, True: SrcStr, False: Sel1, Name: "memchr.sel2" ); |
1389 | } |
1390 | |
1391 | if (!LenC) { |
1392 | if (isOnlyUsedInEqualityComparison(V: CI, With: SrcStr)) |
1393 | // S is dereferenceable so it's safe to load from it and fold |
1394 | // memchr(S, C, N) == S to N && *S == C for any C and N. |
1395 | // TODO: This is safe even for nonconstant S. |
1396 | return memChrToCharCompare(CI, NBytes: Size, B, DL); |
1397 | |
1398 | // From now on we need a constant length and constant array. |
1399 | return nullptr; |
1400 | } |
1401 | |
1402 | bool OptForSize = llvm::shouldOptimizeForSize(BB: CI->getParent(), PSI, BFI, |
1403 | QueryType: PGSOQueryType::IRPass); |
1404 | |
1405 | // If the char is variable but the input str and length are not we can turn |
1406 | // this memchr call into a simple bit field test. Of course this only works |
1407 | // when the return value is only checked against null. |
1408 | // |
1409 | // It would be really nice to reuse switch lowering here but we can't change |
1410 | // the CFG at this point. |
1411 | // |
1412 | // memchr("\r\n", C, 2) != nullptr -> (1 << C & ((1 << '\r') | (1 << '\n'))) |
1413 | // != 0 |
1414 | // after bounds check. |
1415 | if (OptForSize || Str.empty() || !isOnlyUsedInZeroEqualityComparison(CxtI: CI)) |
1416 | return nullptr; |
1417 | |
1418 | unsigned char Max = |
1419 | *std::max_element(first: reinterpret_cast<const unsigned char *>(Str.begin()), |
1420 | last: reinterpret_cast<const unsigned char *>(Str.end())); |
1421 | |
1422 | // Make sure the bit field we're about to create fits in a register on the |
1423 | // target. |
1424 | // FIXME: On a 64 bit architecture this prevents us from using the |
1425 | // interesting range of alpha ascii chars. We could do better by emitting |
1426 | // two bitfields or shifting the range by 64 if no lower chars are used. |
1427 | if (!DL.fitsInLegalInteger(Width: Max + 1)) { |
1428 | // Build chain of ORs |
1429 | // Transform: |
1430 | // memchr("abcd", C, 4) != nullptr |
1431 | // to: |
1432 | // (C == 'a' || C == 'b' || C == 'c' || C == 'd') != 0 |
1433 | std::string SortedStr = Str.str(); |
1434 | llvm::sort(C&: SortedStr); |
1435 | // Compute the number of of non-contiguous ranges. |
1436 | unsigned NonContRanges = 1; |
1437 | for (size_t i = 1; i < SortedStr.size(); ++i) { |
1438 | if (SortedStr[i] > SortedStr[i - 1] + 1) { |
1439 | NonContRanges++; |
1440 | } |
1441 | } |
1442 | |
1443 | // Restrict this optimization to profitable cases with one or two range |
1444 | // checks. |
1445 | if (NonContRanges > 2) |
1446 | return nullptr; |
1447 | |
1448 | // Slice off the character's high end bits. |
1449 | CharVal = B.CreateTrunc(V: CharVal, DestTy: B.getInt8Ty()); |
1450 | |
1451 | SmallVector<Value *> CharCompares; |
1452 | for (unsigned char C : SortedStr) |
1453 | CharCompares.push_back(Elt: B.CreateICmpEQ(LHS: CharVal, RHS: B.getInt8(C))); |
1454 | |
1455 | return B.CreateIntToPtr(V: B.CreateOr(Ops: CharCompares), DestTy: CI->getType()); |
1456 | } |
1457 | |
1458 | // For the bit field use a power-of-2 type with at least 8 bits to avoid |
1459 | // creating unnecessary illegal types. |
1460 | unsigned char Width = NextPowerOf2(A: std::max(a: (unsigned char)7, b: Max)); |
1461 | |
1462 | // Now build the bit field. |
1463 | APInt Bitfield(Width, 0); |
1464 | for (char C : Str) |
1465 | Bitfield.setBit((unsigned char)C); |
1466 | Value *BitfieldC = B.getInt(AI: Bitfield); |
1467 | |
1468 | // Adjust width of "C" to the bitfield width, then mask off the high bits. |
1469 | Value *C = B.CreateZExtOrTrunc(V: CharVal, DestTy: BitfieldC->getType()); |
1470 | C = B.CreateAnd(LHS: C, RHS: B.getIntN(N: Width, C: 0xFF)); |
1471 | |
1472 | // First check that the bit field access is within bounds. |
1473 | Value *Bounds = B.CreateICmp(P: ICmpInst::ICMP_ULT, LHS: C, RHS: B.getIntN(N: Width, C: Width), |
1474 | Name: "memchr.bounds" ); |
1475 | |
1476 | // Create code that checks if the given bit is set in the field. |
1477 | Value *Shl = B.CreateShl(LHS: B.getIntN(N: Width, C: 1ULL), RHS: C); |
1478 | Value *Bits = B.CreateIsNotNull(Arg: B.CreateAnd(LHS: Shl, RHS: BitfieldC), Name: "memchr.bits" ); |
1479 | |
1480 | // Finally merge both checks and cast to pointer type. The inttoptr |
1481 | // implicitly zexts the i1 to intptr type. |
1482 | return B.CreateIntToPtr(V: B.CreateLogicalAnd(Cond1: Bounds, Cond2: Bits, Name: "memchr" ), |
1483 | DestTy: CI->getType()); |
1484 | } |
1485 | |
1486 | // Optimize a memcmp or, when StrNCmp is true, strncmp call CI with constant |
1487 | // arrays LHS and RHS and nonconstant Size. |
1488 | static Value *optimizeMemCmpVarSize(CallInst *CI, Value *LHS, Value *RHS, |
1489 | Value *Size, bool StrNCmp, |
1490 | IRBuilderBase &B, const DataLayout &DL) { |
1491 | if (LHS == RHS) // memcmp(s,s,x) -> 0 |
1492 | return Constant::getNullValue(Ty: CI->getType()); |
1493 | |
1494 | StringRef LStr, RStr; |
1495 | if (!getConstantStringInfo(V: LHS, Str&: LStr, /*TrimAtNul=*/false) || |
1496 | !getConstantStringInfo(V: RHS, Str&: RStr, /*TrimAtNul=*/false)) |
1497 | return nullptr; |
1498 | |
1499 | // If the contents of both constant arrays are known, fold a call to |
1500 | // memcmp(A, B, N) to |
1501 | // N <= Pos ? 0 : (A < B ? -1 : B < A ? +1 : 0) |
1502 | // where Pos is the first mismatch between A and B, determined below. |
1503 | |
1504 | uint64_t Pos = 0; |
1505 | Value *Zero = ConstantInt::get(Ty: CI->getType(), V: 0); |
1506 | for (uint64_t MinSize = std::min(a: LStr.size(), b: RStr.size()); ; ++Pos) { |
1507 | if (Pos == MinSize || |
1508 | (StrNCmp && (LStr[Pos] == '\0' && RStr[Pos] == '\0'))) { |
1509 | // One array is a leading part of the other of equal or greater |
1510 | // size, or for strncmp, the arrays are equal strings. |
1511 | // Fold the result to zero. Size is assumed to be in bounds, since |
1512 | // otherwise the call would be undefined. |
1513 | return Zero; |
1514 | } |
1515 | |
1516 | if (LStr[Pos] != RStr[Pos]) |
1517 | break; |
1518 | } |
1519 | |
1520 | // Normalize the result. |
1521 | typedef unsigned char UChar; |
1522 | int IRes = UChar(LStr[Pos]) < UChar(RStr[Pos]) ? -1 : 1; |
1523 | Value *MaxSize = ConstantInt::get(Ty: Size->getType(), V: Pos); |
1524 | Value *Cmp = B.CreateICmp(P: ICmpInst::ICMP_ULE, LHS: Size, RHS: MaxSize); |
1525 | Value *Res = ConstantInt::get(Ty: CI->getType(), V: IRes); |
1526 | return B.CreateSelect(C: Cmp, True: Zero, False: Res); |
1527 | } |
1528 | |
1529 | // Optimize a memcmp call CI with constant size Len. |
1530 | static Value *optimizeMemCmpConstantSize(CallInst *CI, Value *LHS, Value *RHS, |
1531 | uint64_t Len, IRBuilderBase &B, |
1532 | const DataLayout &DL) { |
1533 | if (Len == 0) // memcmp(s1,s2,0) -> 0 |
1534 | return Constant::getNullValue(Ty: CI->getType()); |
1535 | |
1536 | // memcmp(S1,S2,1) -> *(unsigned char*)LHS - *(unsigned char*)RHS |
1537 | if (Len == 1) { |
1538 | Value *LHSV = B.CreateZExt(V: B.CreateLoad(Ty: B.getInt8Ty(), Ptr: LHS, Name: "lhsc" ), |
1539 | DestTy: CI->getType(), Name: "lhsv" ); |
1540 | Value *RHSV = B.CreateZExt(V: B.CreateLoad(Ty: B.getInt8Ty(), Ptr: RHS, Name: "rhsc" ), |
1541 | DestTy: CI->getType(), Name: "rhsv" ); |
1542 | return B.CreateSub(LHS: LHSV, RHS: RHSV, Name: "chardiff" ); |
1543 | } |
1544 | |
1545 | // memcmp(S1,S2,N/8)==0 -> (*(intN_t*)S1 != *(intN_t*)S2)==0 |
1546 | // TODO: The case where both inputs are constants does not need to be limited |
1547 | // to legal integers or equality comparison. See block below this. |
1548 | if (DL.isLegalInteger(Width: Len * 8) && isOnlyUsedInZeroEqualityComparison(CxtI: CI)) { |
1549 | IntegerType *IntType = IntegerType::get(C&: CI->getContext(), NumBits: Len * 8); |
1550 | Align PrefAlignment = DL.getPrefTypeAlign(Ty: IntType); |
1551 | |
1552 | // First, see if we can fold either argument to a constant. |
1553 | Value *LHSV = nullptr; |
1554 | if (auto *LHSC = dyn_cast<Constant>(Val: LHS)) |
1555 | LHSV = ConstantFoldLoadFromConstPtr(C: LHSC, Ty: IntType, DL); |
1556 | |
1557 | Value *RHSV = nullptr; |
1558 | if (auto *RHSC = dyn_cast<Constant>(Val: RHS)) |
1559 | RHSV = ConstantFoldLoadFromConstPtr(C: RHSC, Ty: IntType, DL); |
1560 | |
1561 | // Don't generate unaligned loads. If either source is constant data, |
1562 | // alignment doesn't matter for that source because there is no load. |
1563 | if ((LHSV || getKnownAlignment(V: LHS, DL, CxtI: CI) >= PrefAlignment) && |
1564 | (RHSV || getKnownAlignment(V: RHS, DL, CxtI: CI) >= PrefAlignment)) { |
1565 | if (!LHSV) |
1566 | LHSV = B.CreateLoad(Ty: IntType, Ptr: LHS, Name: "lhsv" ); |
1567 | if (!RHSV) |
1568 | RHSV = B.CreateLoad(Ty: IntType, Ptr: RHS, Name: "rhsv" ); |
1569 | return B.CreateZExt(V: B.CreateICmpNE(LHS: LHSV, RHS: RHSV), DestTy: CI->getType(), Name: "memcmp" ); |
1570 | } |
1571 | } |
1572 | |
1573 | return nullptr; |
1574 | } |
1575 | |
1576 | // Most simplifications for memcmp also apply to bcmp. |
1577 | Value *LibCallSimplifier::optimizeMemCmpBCmpCommon(CallInst *CI, |
1578 | IRBuilderBase &B) { |
1579 | Value *LHS = CI->getArgOperand(i: 0), *RHS = CI->getArgOperand(i: 1); |
1580 | Value *Size = CI->getArgOperand(i: 2); |
1581 | |
1582 | annotateNonNullAndDereferenceable(CI, ArgNos: {0, 1}, Size, DL); |
1583 | |
1584 | if (Value *Res = optimizeMemCmpVarSize(CI, LHS, RHS, Size, StrNCmp: false, B, DL)) |
1585 | return Res; |
1586 | |
1587 | // Handle constant Size. |
1588 | ConstantInt *LenC = dyn_cast<ConstantInt>(Val: Size); |
1589 | if (!LenC) |
1590 | return nullptr; |
1591 | |
1592 | return optimizeMemCmpConstantSize(CI, LHS, RHS, Len: LenC->getZExtValue(), B, DL); |
1593 | } |
1594 | |
1595 | Value *LibCallSimplifier::optimizeMemCmp(CallInst *CI, IRBuilderBase &B) { |
1596 | Module *M = CI->getModule(); |
1597 | if (Value *V = optimizeMemCmpBCmpCommon(CI, B)) |
1598 | return V; |
1599 | |
1600 | // memcmp(x, y, Len) == 0 -> bcmp(x, y, Len) == 0 |
1601 | // bcmp can be more efficient than memcmp because it only has to know that |
1602 | // there is a difference, not how different one is to the other. |
1603 | if (isLibFuncEmittable(M, TLI, TheLibFunc: LibFunc_bcmp) && |
1604 | isOnlyUsedInZeroEqualityComparison(CxtI: CI)) { |
1605 | Value *LHS = CI->getArgOperand(i: 0); |
1606 | Value *RHS = CI->getArgOperand(i: 1); |
1607 | Value *Size = CI->getArgOperand(i: 2); |
1608 | return copyFlags(Old: *CI, New: emitBCmp(Ptr1: LHS, Ptr2: RHS, Len: Size, B, DL, TLI)); |
1609 | } |
1610 | |
1611 | return nullptr; |
1612 | } |
1613 | |
1614 | Value *LibCallSimplifier::optimizeBCmp(CallInst *CI, IRBuilderBase &B) { |
1615 | return optimizeMemCmpBCmpCommon(CI, B); |
1616 | } |
1617 | |
1618 | Value *LibCallSimplifier::optimizeMemCpy(CallInst *CI, IRBuilderBase &B) { |
1619 | Value *Size = CI->getArgOperand(i: 2); |
1620 | annotateNonNullAndDereferenceable(CI, ArgNos: {0, 1}, Size, DL); |
1621 | if (isa<IntrinsicInst>(Val: CI)) |
1622 | return nullptr; |
1623 | |
1624 | // memcpy(x, y, n) -> llvm.memcpy(align 1 x, align 1 y, n) |
1625 | CallInst *NewCI = B.CreateMemCpy(Dst: CI->getArgOperand(i: 0), DstAlign: Align(1), |
1626 | Src: CI->getArgOperand(i: 1), SrcAlign: Align(1), Size); |
1627 | mergeAttributesAndFlags(NewCI, Old: *CI); |
1628 | return CI->getArgOperand(i: 0); |
1629 | } |
1630 | |
1631 | Value *LibCallSimplifier::optimizeMemCCpy(CallInst *CI, IRBuilderBase &B) { |
1632 | Value *Dst = CI->getArgOperand(i: 0); |
1633 | Value *Src = CI->getArgOperand(i: 1); |
1634 | ConstantInt *StopChar = dyn_cast<ConstantInt>(Val: CI->getArgOperand(i: 2)); |
1635 | ConstantInt *N = dyn_cast<ConstantInt>(Val: CI->getArgOperand(i: 3)); |
1636 | StringRef SrcStr; |
1637 | if (CI->use_empty() && Dst == Src) |
1638 | return Dst; |
1639 | // memccpy(d, s, c, 0) -> nullptr |
1640 | if (N) { |
1641 | if (N->isNullValue()) |
1642 | return Constant::getNullValue(Ty: CI->getType()); |
1643 | if (!getConstantStringInfo(V: Src, Str&: SrcStr, /*TrimAtNul=*/false) || |
1644 | // TODO: Handle zeroinitializer. |
1645 | !StopChar) |
1646 | return nullptr; |
1647 | } else { |
1648 | return nullptr; |
1649 | } |
1650 | |
1651 | // Wrap arg 'c' of type int to char |
1652 | size_t Pos = SrcStr.find(C: StopChar->getSExtValue() & 0xFF); |
1653 | if (Pos == StringRef::npos) { |
1654 | if (N->getZExtValue() <= SrcStr.size()) { |
1655 | copyFlags(Old: *CI, New: B.CreateMemCpy(Dst, DstAlign: Align(1), Src, SrcAlign: Align(1), |
1656 | Size: CI->getArgOperand(i: 3))); |
1657 | return Constant::getNullValue(Ty: CI->getType()); |
1658 | } |
1659 | return nullptr; |
1660 | } |
1661 | |
1662 | Value *NewN = |
1663 | ConstantInt::get(Ty: N->getType(), V: std::min(a: uint64_t(Pos + 1), b: N->getZExtValue())); |
1664 | // memccpy -> llvm.memcpy |
1665 | copyFlags(Old: *CI, New: B.CreateMemCpy(Dst, DstAlign: Align(1), Src, SrcAlign: Align(1), Size: NewN)); |
1666 | return Pos + 1 <= N->getZExtValue() |
1667 | ? B.CreateInBoundsGEP(Ty: B.getInt8Ty(), Ptr: Dst, IdxList: NewN) |
1668 | : Constant::getNullValue(Ty: CI->getType()); |
1669 | } |
1670 | |
1671 | Value *LibCallSimplifier::optimizeMemPCpy(CallInst *CI, IRBuilderBase &B) { |
1672 | Value *Dst = CI->getArgOperand(i: 0); |
1673 | Value *N = CI->getArgOperand(i: 2); |
1674 | // mempcpy(x, y, n) -> llvm.memcpy(align 1 x, align 1 y, n), x + n |
1675 | CallInst *NewCI = |
1676 | B.CreateMemCpy(Dst, DstAlign: Align(1), Src: CI->getArgOperand(i: 1), SrcAlign: Align(1), Size: N); |
1677 | // Propagate attributes, but memcpy has no return value, so make sure that |
1678 | // any return attributes are compliant. |
1679 | // TODO: Attach return value attributes to the 1st operand to preserve them? |
1680 | mergeAttributesAndFlags(NewCI, Old: *CI); |
1681 | return B.CreateInBoundsGEP(Ty: B.getInt8Ty(), Ptr: Dst, IdxList: N); |
1682 | } |
1683 | |
1684 | Value *LibCallSimplifier::optimizeMemMove(CallInst *CI, IRBuilderBase &B) { |
1685 | Value *Size = CI->getArgOperand(i: 2); |
1686 | annotateNonNullAndDereferenceable(CI, ArgNos: {0, 1}, Size, DL); |
1687 | if (isa<IntrinsicInst>(Val: CI)) |
1688 | return nullptr; |
1689 | |
1690 | // memmove(x, y, n) -> llvm.memmove(align 1 x, align 1 y, n) |
1691 | CallInst *NewCI = B.CreateMemMove(Dst: CI->getArgOperand(i: 0), DstAlign: Align(1), |
1692 | Src: CI->getArgOperand(i: 1), SrcAlign: Align(1), Size); |
1693 | mergeAttributesAndFlags(NewCI, Old: *CI); |
1694 | return CI->getArgOperand(i: 0); |
1695 | } |
1696 | |
1697 | Value *LibCallSimplifier::optimizeMemSet(CallInst *CI, IRBuilderBase &B) { |
1698 | Value *Size = CI->getArgOperand(i: 2); |
1699 | annotateNonNullAndDereferenceable(CI, ArgNos: 0, Size, DL); |
1700 | if (isa<IntrinsicInst>(Val: CI)) |
1701 | return nullptr; |
1702 | |
1703 | // memset(p, v, n) -> llvm.memset(align 1 p, v, n) |
1704 | Value *Val = B.CreateIntCast(V: CI->getArgOperand(i: 1), DestTy: B.getInt8Ty(), isSigned: false); |
1705 | CallInst *NewCI = B.CreateMemSet(Ptr: CI->getArgOperand(i: 0), Val, Size, Align: Align(1)); |
1706 | mergeAttributesAndFlags(NewCI, Old: *CI); |
1707 | return CI->getArgOperand(i: 0); |
1708 | } |
1709 | |
1710 | Value *LibCallSimplifier::optimizeRealloc(CallInst *CI, IRBuilderBase &B) { |
1711 | if (isa<ConstantPointerNull>(Val: CI->getArgOperand(i: 0))) |
1712 | return copyFlags(Old: *CI, New: emitMalloc(Num: CI->getArgOperand(i: 1), B, DL, TLI)); |
1713 | |
1714 | return nullptr; |
1715 | } |
1716 | |
1717 | // When enabled, replace operator new() calls marked with a hot or cold memprof |
1718 | // attribute with an operator new() call that takes a __hot_cold_t parameter. |
1719 | // Currently this is supported by the open source version of tcmalloc, see: |
1720 | // https://github.com/google/tcmalloc/blob/master/tcmalloc/new_extension.h |
1721 | Value *LibCallSimplifier::optimizeNew(CallInst *CI, IRBuilderBase &B, |
1722 | LibFunc &Func) { |
1723 | if (!OptimizeHotColdNew) |
1724 | return nullptr; |
1725 | |
1726 | uint8_t HotCold; |
1727 | if (CI->getAttributes().getFnAttr(Kind: "memprof" ).getValueAsString() == "cold" ) |
1728 | HotCold = ColdNewHintValue; |
1729 | else if (CI->getAttributes().getFnAttr(Kind: "memprof" ).getValueAsString() == |
1730 | "notcold" ) |
1731 | HotCold = NotColdNewHintValue; |
1732 | else if (CI->getAttributes().getFnAttr(Kind: "memprof" ).getValueAsString() == "hot" ) |
1733 | HotCold = HotNewHintValue; |
1734 | else |
1735 | return nullptr; |
1736 | |
1737 | // For calls that already pass a hot/cold hint, only update the hint if |
1738 | // directed by OptimizeExistingHotColdNew. For other calls to new, add a hint |
1739 | // if cold or hot, and leave as-is for default handling if "notcold" aka warm. |
1740 | // Note that in cases where we decide it is "notcold", it might be slightly |
1741 | // better to replace the hinted call with a non hinted call, to avoid the |
1742 | // extra parameter and the if condition check of the hint value in the |
1743 | // allocator. This can be considered in the future. |
1744 | switch (Func) { |
1745 | case LibFunc_Znwm12__hot_cold_t: |
1746 | if (OptimizeExistingHotColdNew) |
1747 | return emitHotColdNew(Num: CI->getArgOperand(i: 0), B, TLI, |
1748 | NewFunc: LibFunc_Znwm12__hot_cold_t, HotCold); |
1749 | break; |
1750 | case LibFunc_Znwm: |
1751 | if (HotCold != NotColdNewHintValue) |
1752 | return emitHotColdNew(Num: CI->getArgOperand(i: 0), B, TLI, |
1753 | NewFunc: LibFunc_Znwm12__hot_cold_t, HotCold); |
1754 | break; |
1755 | case LibFunc_Znam12__hot_cold_t: |
1756 | if (OptimizeExistingHotColdNew) |
1757 | return emitHotColdNew(Num: CI->getArgOperand(i: 0), B, TLI, |
1758 | NewFunc: LibFunc_Znam12__hot_cold_t, HotCold); |
1759 | break; |
1760 | case LibFunc_Znam: |
1761 | if (HotCold != NotColdNewHintValue) |
1762 | return emitHotColdNew(Num: CI->getArgOperand(i: 0), B, TLI, |
1763 | NewFunc: LibFunc_Znam12__hot_cold_t, HotCold); |
1764 | break; |
1765 | case LibFunc_ZnwmRKSt9nothrow_t12__hot_cold_t: |
1766 | if (OptimizeExistingHotColdNew) |
1767 | return emitHotColdNewNoThrow( |
1768 | Num: CI->getArgOperand(i: 0), NoThrow: CI->getArgOperand(i: 1), B, TLI, |
1769 | NewFunc: LibFunc_ZnwmRKSt9nothrow_t12__hot_cold_t, HotCold); |
1770 | break; |
1771 | case LibFunc_ZnwmRKSt9nothrow_t: |
1772 | if (HotCold != NotColdNewHintValue) |
1773 | return emitHotColdNewNoThrow( |
1774 | Num: CI->getArgOperand(i: 0), NoThrow: CI->getArgOperand(i: 1), B, TLI, |
1775 | NewFunc: LibFunc_ZnwmRKSt9nothrow_t12__hot_cold_t, HotCold); |
1776 | break; |
1777 | case LibFunc_ZnamRKSt9nothrow_t12__hot_cold_t: |
1778 | if (OptimizeExistingHotColdNew) |
1779 | return emitHotColdNewNoThrow( |
1780 | Num: CI->getArgOperand(i: 0), NoThrow: CI->getArgOperand(i: 1), B, TLI, |
1781 | NewFunc: LibFunc_ZnamRKSt9nothrow_t12__hot_cold_t, HotCold); |
1782 | break; |
1783 | case LibFunc_ZnamRKSt9nothrow_t: |
1784 | if (HotCold != NotColdNewHintValue) |
1785 | return emitHotColdNewNoThrow( |
1786 | Num: CI->getArgOperand(i: 0), NoThrow: CI->getArgOperand(i: 1), B, TLI, |
1787 | NewFunc: LibFunc_ZnamRKSt9nothrow_t12__hot_cold_t, HotCold); |
1788 | break; |
1789 | case LibFunc_ZnwmSt11align_val_t12__hot_cold_t: |
1790 | if (OptimizeExistingHotColdNew) |
1791 | return emitHotColdNewAligned( |
1792 | Num: CI->getArgOperand(i: 0), Align: CI->getArgOperand(i: 1), B, TLI, |
1793 | NewFunc: LibFunc_ZnwmSt11align_val_t12__hot_cold_t, HotCold); |
1794 | break; |
1795 | case LibFunc_ZnwmSt11align_val_t: |
1796 | if (HotCold != NotColdNewHintValue) |
1797 | return emitHotColdNewAligned( |
1798 | Num: CI->getArgOperand(i: 0), Align: CI->getArgOperand(i: 1), B, TLI, |
1799 | NewFunc: LibFunc_ZnwmSt11align_val_t12__hot_cold_t, HotCold); |
1800 | break; |
1801 | case LibFunc_ZnamSt11align_val_t12__hot_cold_t: |
1802 | if (OptimizeExistingHotColdNew) |
1803 | return emitHotColdNewAligned( |
1804 | Num: CI->getArgOperand(i: 0), Align: CI->getArgOperand(i: 1), B, TLI, |
1805 | NewFunc: LibFunc_ZnamSt11align_val_t12__hot_cold_t, HotCold); |
1806 | break; |
1807 | case LibFunc_ZnamSt11align_val_t: |
1808 | if (HotCold != NotColdNewHintValue) |
1809 | return emitHotColdNewAligned( |
1810 | Num: CI->getArgOperand(i: 0), Align: CI->getArgOperand(i: 1), B, TLI, |
1811 | NewFunc: LibFunc_ZnamSt11align_val_t12__hot_cold_t, HotCold); |
1812 | break; |
1813 | case LibFunc_ZnwmSt11align_val_tRKSt9nothrow_t12__hot_cold_t: |
1814 | if (OptimizeExistingHotColdNew) |
1815 | return emitHotColdNewAlignedNoThrow( |
1816 | Num: CI->getArgOperand(i: 0), Align: CI->getArgOperand(i: 1), NoThrow: CI->getArgOperand(i: 2), B, |
1817 | TLI, NewFunc: LibFunc_ZnwmSt11align_val_tRKSt9nothrow_t12__hot_cold_t, |
1818 | HotCold); |
1819 | break; |
1820 | case LibFunc_ZnwmSt11align_val_tRKSt9nothrow_t: |
1821 | if (HotCold != NotColdNewHintValue) |
1822 | return emitHotColdNewAlignedNoThrow( |
1823 | Num: CI->getArgOperand(i: 0), Align: CI->getArgOperand(i: 1), NoThrow: CI->getArgOperand(i: 2), B, |
1824 | TLI, NewFunc: LibFunc_ZnwmSt11align_val_tRKSt9nothrow_t12__hot_cold_t, |
1825 | HotCold); |
1826 | break; |
1827 | case LibFunc_ZnamSt11align_val_tRKSt9nothrow_t12__hot_cold_t: |
1828 | if (OptimizeExistingHotColdNew) |
1829 | return emitHotColdNewAlignedNoThrow( |
1830 | Num: CI->getArgOperand(i: 0), Align: CI->getArgOperand(i: 1), NoThrow: CI->getArgOperand(i: 2), B, |
1831 | TLI, NewFunc: LibFunc_ZnamSt11align_val_tRKSt9nothrow_t12__hot_cold_t, |
1832 | HotCold); |
1833 | break; |
1834 | case LibFunc_ZnamSt11align_val_tRKSt9nothrow_t: |
1835 | if (HotCold != NotColdNewHintValue) |
1836 | return emitHotColdNewAlignedNoThrow( |
1837 | Num: CI->getArgOperand(i: 0), Align: CI->getArgOperand(i: 1), NoThrow: CI->getArgOperand(i: 2), B, |
1838 | TLI, NewFunc: LibFunc_ZnamSt11align_val_tRKSt9nothrow_t12__hot_cold_t, |
1839 | HotCold); |
1840 | break; |
1841 | case LibFunc_size_returning_new: |
1842 | if (HotCold != NotColdNewHintValue) |
1843 | return emitHotColdSizeReturningNew(Num: CI->getArgOperand(i: 0), B, TLI, |
1844 | NewFunc: LibFunc_size_returning_new_hot_cold, |
1845 | HotCold); |
1846 | break; |
1847 | case LibFunc_size_returning_new_hot_cold: |
1848 | if (OptimizeExistingHotColdNew) |
1849 | return emitHotColdSizeReturningNew(Num: CI->getArgOperand(i: 0), B, TLI, |
1850 | NewFunc: LibFunc_size_returning_new_hot_cold, |
1851 | HotCold); |
1852 | break; |
1853 | case LibFunc_size_returning_new_aligned: |
1854 | if (HotCold != NotColdNewHintValue) |
1855 | return emitHotColdSizeReturningNewAligned( |
1856 | Num: CI->getArgOperand(i: 0), Align: CI->getArgOperand(i: 1), B, TLI, |
1857 | NewFunc: LibFunc_size_returning_new_aligned_hot_cold, HotCold); |
1858 | break; |
1859 | case LibFunc_size_returning_new_aligned_hot_cold: |
1860 | if (OptimizeExistingHotColdNew) |
1861 | return emitHotColdSizeReturningNewAligned( |
1862 | Num: CI->getArgOperand(i: 0), Align: CI->getArgOperand(i: 1), B, TLI, |
1863 | NewFunc: LibFunc_size_returning_new_aligned_hot_cold, HotCold); |
1864 | break; |
1865 | default: |
1866 | return nullptr; |
1867 | } |
1868 | return nullptr; |
1869 | } |
1870 | |
1871 | //===----------------------------------------------------------------------===// |
1872 | // Math Library Optimizations |
1873 | //===----------------------------------------------------------------------===// |
1874 | |
1875 | // Replace a libcall \p CI with a call to intrinsic \p IID |
1876 | static Value *replaceUnaryCall(CallInst *CI, IRBuilderBase &B, |
1877 | Intrinsic::ID IID) { |
1878 | CallInst *NewCall = B.CreateUnaryIntrinsic(ID: IID, V: CI->getArgOperand(i: 0), FMFSource: CI); |
1879 | NewCall->takeName(V: CI); |
1880 | return copyFlags(Old: *CI, New: NewCall); |
1881 | } |
1882 | |
1883 | /// Return a variant of Val with float type. |
1884 | /// Currently this works in two cases: If Val is an FPExtension of a float |
1885 | /// value to something bigger, simply return the operand. |
1886 | /// If Val is a ConstantFP but can be converted to a float ConstantFP without |
1887 | /// loss of precision do so. |
1888 | static Value *valueHasFloatPrecision(Value *Val) { |
1889 | if (FPExtInst *Cast = dyn_cast<FPExtInst>(Val)) { |
1890 | Value *Op = Cast->getOperand(i_nocapture: 0); |
1891 | if (Op->getType()->isFloatTy()) |
1892 | return Op; |
1893 | } |
1894 | if (ConstantFP *Const = dyn_cast<ConstantFP>(Val)) { |
1895 | APFloat F = Const->getValueAPF(); |
1896 | bool losesInfo; |
1897 | (void)F.convert(ToSemantics: APFloat::IEEEsingle(), RM: APFloat::rmNearestTiesToEven, |
1898 | losesInfo: &losesInfo); |
1899 | if (!losesInfo) |
1900 | return ConstantFP::get(Context&: Const->getContext(), V: F); |
1901 | } |
1902 | return nullptr; |
1903 | } |
1904 | |
1905 | /// Shrink double -> float functions. |
1906 | static Value *optimizeDoubleFP(CallInst *CI, IRBuilderBase &B, |
1907 | bool isBinary, const TargetLibraryInfo *TLI, |
1908 | bool isPrecise = false) { |
1909 | Function *CalleeFn = CI->getCalledFunction(); |
1910 | if (!CI->getType()->isDoubleTy() || !CalleeFn) |
1911 | return nullptr; |
1912 | |
1913 | // If not all the uses of the function are converted to float, then bail out. |
1914 | // This matters if the precision of the result is more important than the |
1915 | // precision of the arguments. |
1916 | if (isPrecise) |
1917 | for (User *U : CI->users()) { |
1918 | FPTruncInst *Cast = dyn_cast<FPTruncInst>(Val: U); |
1919 | if (!Cast || !Cast->getType()->isFloatTy()) |
1920 | return nullptr; |
1921 | } |
1922 | |
1923 | // If this is something like 'g((double) float)', convert to 'gf(float)'. |
1924 | Value *V[2]; |
1925 | V[0] = valueHasFloatPrecision(Val: CI->getArgOperand(i: 0)); |
1926 | V[1] = isBinary ? valueHasFloatPrecision(Val: CI->getArgOperand(i: 1)) : nullptr; |
1927 | if (!V[0] || (isBinary && !V[1])) |
1928 | return nullptr; |
1929 | |
1930 | // If call isn't an intrinsic, check that it isn't within a function with the |
1931 | // same name as the float version of this call, otherwise the result is an |
1932 | // infinite loop. For example, from MinGW-w64: |
1933 | // |
1934 | // float expf(float val) { return (float) exp((double) val); } |
1935 | StringRef CalleeName = CalleeFn->getName(); |
1936 | bool IsIntrinsic = CalleeFn->isIntrinsic(); |
1937 | if (!IsIntrinsic) { |
1938 | StringRef CallerName = CI->getFunction()->getName(); |
1939 | if (CallerName.ends_with(Suffix: 'f') && |
1940 | CallerName.size() == (CalleeName.size() + 1) && |
1941 | CallerName.starts_with(Prefix: CalleeName)) |
1942 | return nullptr; |
1943 | } |
1944 | |
1945 | // Propagate the math semantics from the current function to the new function. |
1946 | IRBuilderBase::FastMathFlagGuard Guard(B); |
1947 | B.setFastMathFlags(CI->getFastMathFlags()); |
1948 | |
1949 | // g((double) float) -> (double) gf(float) |
1950 | Value *R; |
1951 | if (IsIntrinsic) { |
1952 | Intrinsic::ID IID = CalleeFn->getIntrinsicID(); |
1953 | R = isBinary ? B.CreateIntrinsic(ID: IID, Types: B.getFloatTy(), Args: V) |
1954 | : B.CreateIntrinsic(ID: IID, Types: B.getFloatTy(), Args: V[0]); |
1955 | } else { |
1956 | AttributeList CalleeAttrs = CalleeFn->getAttributes(); |
1957 | R = isBinary ? emitBinaryFloatFnCall(Op1: V[0], Op2: V[1], TLI, Name: CalleeName, B, |
1958 | Attrs: CalleeAttrs) |
1959 | : emitUnaryFloatFnCall(Op: V[0], TLI, Name: CalleeName, B, Attrs: CalleeAttrs); |
1960 | } |
1961 | return B.CreateFPExt(V: R, DestTy: B.getDoubleTy()); |
1962 | } |
1963 | |
1964 | /// Shrink double -> float for unary functions. |
1965 | static Value *optimizeUnaryDoubleFP(CallInst *CI, IRBuilderBase &B, |
1966 | const TargetLibraryInfo *TLI, |
1967 | bool isPrecise = false) { |
1968 | return optimizeDoubleFP(CI, B, isBinary: false, TLI, isPrecise); |
1969 | } |
1970 | |
1971 | /// Shrink double -> float for binary functions. |
1972 | static Value *optimizeBinaryDoubleFP(CallInst *CI, IRBuilderBase &B, |
1973 | const TargetLibraryInfo *TLI, |
1974 | bool isPrecise = false) { |
1975 | return optimizeDoubleFP(CI, B, isBinary: true, TLI, isPrecise); |
1976 | } |
1977 | |
1978 | // cabs(z) -> sqrt((creal(z)*creal(z)) + (cimag(z)*cimag(z))) |
1979 | Value *LibCallSimplifier::optimizeCAbs(CallInst *CI, IRBuilderBase &B) { |
1980 | Value *Real, *Imag; |
1981 | |
1982 | if (CI->arg_size() == 1) { |
1983 | |
1984 | if (!CI->isFast()) |
1985 | return nullptr; |
1986 | |
1987 | Value *Op = CI->getArgOperand(i: 0); |
1988 | assert(Op->getType()->isArrayTy() && "Unexpected signature for cabs!" ); |
1989 | |
1990 | Real = B.CreateExtractValue(Agg: Op, Idxs: 0, Name: "real" ); |
1991 | Imag = B.CreateExtractValue(Agg: Op, Idxs: 1, Name: "imag" ); |
1992 | |
1993 | } else { |
1994 | assert(CI->arg_size() == 2 && "Unexpected signature for cabs!" ); |
1995 | |
1996 | Real = CI->getArgOperand(i: 0); |
1997 | Imag = CI->getArgOperand(i: 1); |
1998 | |
1999 | // if real or imaginary part is zero, simplify to abs(cimag(z)) |
2000 | // or abs(creal(z)) |
2001 | Value *AbsOp = nullptr; |
2002 | if (ConstantFP *ConstReal = dyn_cast<ConstantFP>(Val: Real)) { |
2003 | if (ConstReal->isZero()) |
2004 | AbsOp = Imag; |
2005 | |
2006 | } else if (ConstantFP *ConstImag = dyn_cast<ConstantFP>(Val: Imag)) { |
2007 | if (ConstImag->isZero()) |
2008 | AbsOp = Real; |
2009 | } |
2010 | |
2011 | if (AbsOp) |
2012 | return copyFlags( |
2013 | Old: *CI, New: B.CreateUnaryIntrinsic(ID: Intrinsic::fabs, V: AbsOp, FMFSource: CI, Name: "cabs" )); |
2014 | |
2015 | if (!CI->isFast()) |
2016 | return nullptr; |
2017 | } |
2018 | |
2019 | // Propagate fast-math flags from the existing call to new instructions. |
2020 | Value *RealReal = B.CreateFMulFMF(L: Real, R: Real, FMFSource: CI); |
2021 | Value *ImagImag = B.CreateFMulFMF(L: Imag, R: Imag, FMFSource: CI); |
2022 | return copyFlags( |
2023 | Old: *CI, New: B.CreateUnaryIntrinsic(ID: Intrinsic::sqrt, |
2024 | V: B.CreateFAddFMF(L: RealReal, R: ImagImag, FMFSource: CI), FMFSource: CI, |
2025 | Name: "cabs" )); |
2026 | } |
2027 | |
2028 | // Return a properly extended integer (DstWidth bits wide) if the operation is |
2029 | // an itofp. |
2030 | static Value *getIntToFPVal(Value *I2F, IRBuilderBase &B, unsigned DstWidth) { |
2031 | if (isa<SIToFPInst>(Val: I2F) || isa<UIToFPInst>(Val: I2F)) { |
2032 | Value *Op = cast<Instruction>(Val: I2F)->getOperand(i: 0); |
2033 | // Make sure that the exponent fits inside an "int" of size DstWidth, |
2034 | // thus avoiding any range issues that FP has not. |
2035 | unsigned BitWidth = Op->getType()->getScalarSizeInBits(); |
2036 | if (BitWidth < DstWidth || (BitWidth == DstWidth && isa<SIToFPInst>(Val: I2F))) { |
2037 | Type *IntTy = Op->getType()->getWithNewBitWidth(NewBitWidth: DstWidth); |
2038 | return isa<SIToFPInst>(Val: I2F) ? B.CreateSExt(V: Op, DestTy: IntTy) |
2039 | : B.CreateZExt(V: Op, DestTy: IntTy); |
2040 | } |
2041 | } |
2042 | |
2043 | return nullptr; |
2044 | } |
2045 | |
2046 | /// Use exp{,2}(x * y) for pow(exp{,2}(x), y); |
2047 | /// ldexp(1.0, x) for pow(2.0, itofp(x)); exp2(n * x) for pow(2.0 ** n, x); |
2048 | /// exp10(x) for pow(10.0, x); exp2(log2(n) * x) for pow(n, x). |
2049 | Value *LibCallSimplifier::replacePowWithExp(CallInst *Pow, IRBuilderBase &B) { |
2050 | Module *M = Pow->getModule(); |
2051 | Value *Base = Pow->getArgOperand(i: 0), *Expo = Pow->getArgOperand(i: 1); |
2052 | Type *Ty = Pow->getType(); |
2053 | bool Ignored; |
2054 | |
2055 | // Evaluate special cases related to a nested function as the base. |
2056 | |
2057 | // pow(exp(x), y) -> exp(x * y) |
2058 | // pow(exp2(x), y) -> exp2(x * y) |
2059 | // If exp{,2}() is used only once, it is better to fold two transcendental |
2060 | // math functions into one. If used again, exp{,2}() would still have to be |
2061 | // called with the original argument, then keep both original transcendental |
2062 | // functions. However, this transformation is only safe with fully relaxed |
2063 | // math semantics, since, besides rounding differences, it changes overflow |
2064 | // and underflow behavior quite dramatically. For example: |
2065 | // pow(exp(1000), 0.001) = pow(inf, 0.001) = inf |
2066 | // Whereas: |
2067 | // exp(1000 * 0.001) = exp(1) |
2068 | // TODO: Loosen the requirement for fully relaxed math semantics. |
2069 | // TODO: Handle exp10() when more targets have it available. |
2070 | CallInst *BaseFn = dyn_cast<CallInst>(Val: Base); |
2071 | if (BaseFn && BaseFn->hasOneUse() && BaseFn->isFast() && Pow->isFast()) { |
2072 | LibFunc LibFn; |
2073 | |
2074 | Function *CalleeFn = BaseFn->getCalledFunction(); |
2075 | if (CalleeFn && TLI->getLibFunc(funcName: CalleeFn->getName(), F&: LibFn) && |
2076 | isLibFuncEmittable(M, TLI, TheLibFunc: LibFn)) { |
2077 | StringRef ExpName; |
2078 | Intrinsic::ID ID; |
2079 | Value *ExpFn; |
2080 | LibFunc LibFnFloat, LibFnDouble, LibFnLongDouble; |
2081 | |
2082 | switch (LibFn) { |
2083 | default: |
2084 | return nullptr; |
2085 | case LibFunc_expf: |
2086 | case LibFunc_exp: |
2087 | case LibFunc_expl: |
2088 | ExpName = TLI->getName(F: LibFunc_exp); |
2089 | ID = Intrinsic::exp; |
2090 | LibFnFloat = LibFunc_expf; |
2091 | LibFnDouble = LibFunc_exp; |
2092 | LibFnLongDouble = LibFunc_expl; |
2093 | break; |
2094 | case LibFunc_exp2f: |
2095 | case LibFunc_exp2: |
2096 | case LibFunc_exp2l: |
2097 | ExpName = TLI->getName(F: LibFunc_exp2); |
2098 | ID = Intrinsic::exp2; |
2099 | LibFnFloat = LibFunc_exp2f; |
2100 | LibFnDouble = LibFunc_exp2; |
2101 | LibFnLongDouble = LibFunc_exp2l; |
2102 | break; |
2103 | } |
2104 | |
2105 | // Create new exp{,2}() with the product as its argument. |
2106 | Value *FMul = B.CreateFMul(L: BaseFn->getArgOperand(i: 0), R: Expo, Name: "mul" ); |
2107 | ExpFn = BaseFn->doesNotAccessMemory() |
2108 | ? B.CreateUnaryIntrinsic(ID, V: FMul, FMFSource: nullptr, Name: ExpName) |
2109 | : emitUnaryFloatFnCall(Op: FMul, TLI, DoubleFn: LibFnDouble, FloatFn: LibFnFloat, |
2110 | LongDoubleFn: LibFnLongDouble, B, |
2111 | Attrs: BaseFn->getAttributes()); |
2112 | |
2113 | // Since the new exp{,2}() is different from the original one, dead code |
2114 | // elimination cannot be trusted to remove it, since it may have side |
2115 | // effects (e.g., errno). When the only consumer for the original |
2116 | // exp{,2}() is pow(), then it has to be explicitly erased. |
2117 | substituteInParent(I: BaseFn, With: ExpFn); |
2118 | return ExpFn; |
2119 | } |
2120 | } |
2121 | |
2122 | // Evaluate special cases related to a constant base. |
2123 | |
2124 | const APFloat *BaseF; |
2125 | if (!match(V: Base, P: m_APFloat(Res&: BaseF))) |
2126 | return nullptr; |
2127 | |
2128 | AttributeList NoAttrs; // Attributes are only meaningful on the original call |
2129 | |
2130 | const bool UseIntrinsic = Pow->doesNotAccessMemory(); |
2131 | |
2132 | // pow(2.0, itofp(x)) -> ldexp(1.0, x) |
2133 | if ((UseIntrinsic || !Ty->isVectorTy()) && BaseF->isExactlyValue(V: 2.0) && |
2134 | (isa<SIToFPInst>(Val: Expo) || isa<UIToFPInst>(Val: Expo)) && |
2135 | (UseIntrinsic || |
2136 | hasFloatFn(M, TLI, Ty, DoubleFn: LibFunc_ldexp, FloatFn: LibFunc_ldexpf, LongDoubleFn: LibFunc_ldexpl))) { |
2137 | |
2138 | // TODO: Shouldn't really need to depend on getIntToFPVal for intrinsic. Can |
2139 | // just directly use the original integer type. |
2140 | if (Value *ExpoI = getIntToFPVal(I2F: Expo, B, DstWidth: TLI->getIntSize())) { |
2141 | Constant *One = ConstantFP::get(Ty, V: 1.0); |
2142 | |
2143 | if (UseIntrinsic) { |
2144 | return copyFlags(Old: *Pow, New: B.CreateIntrinsic(ID: Intrinsic::ldexp, |
2145 | Types: {Ty, ExpoI->getType()}, |
2146 | Args: {One, ExpoI}, FMFSource: Pow, Name: "exp2" )); |
2147 | } |
2148 | |
2149 | return copyFlags(Old: *Pow, New: emitBinaryFloatFnCall( |
2150 | Op1: One, Op2: ExpoI, TLI, DoubleFn: LibFunc_ldexp, FloatFn: LibFunc_ldexpf, |
2151 | LongDoubleFn: LibFunc_ldexpl, B, Attrs: NoAttrs)); |
2152 | } |
2153 | } |
2154 | |
2155 | // pow(2.0 ** n, x) -> exp2(n * x) |
2156 | if (hasFloatFn(M, TLI, Ty, DoubleFn: LibFunc_exp2, FloatFn: LibFunc_exp2f, LongDoubleFn: LibFunc_exp2l)) { |
2157 | APFloat BaseR = APFloat(1.0); |
2158 | BaseR.convert(ToSemantics: BaseF->getSemantics(), RM: APFloat::rmTowardZero, losesInfo: &Ignored); |
2159 | BaseR = BaseR / *BaseF; |
2160 | bool IsInteger = BaseF->isInteger(), IsReciprocal = BaseR.isInteger(); |
2161 | const APFloat *NF = IsReciprocal ? &BaseR : BaseF; |
2162 | APSInt NI(64, false); |
2163 | if ((IsInteger || IsReciprocal) && |
2164 | NF->convertToInteger(Result&: NI, RM: APFloat::rmTowardZero, IsExact: &Ignored) == |
2165 | APFloat::opOK && |
2166 | NI > 1 && NI.isPowerOf2()) { |
2167 | double N = NI.logBase2() * (IsReciprocal ? -1.0 : 1.0); |
2168 | Value *FMul = B.CreateFMul(L: Expo, R: ConstantFP::get(Ty, V: N), Name: "mul" ); |
2169 | if (Pow->doesNotAccessMemory()) |
2170 | return copyFlags(Old: *Pow, New: B.CreateUnaryIntrinsic(ID: Intrinsic::exp2, V: FMul, |
2171 | FMFSource: nullptr, Name: "exp2" )); |
2172 | else |
2173 | return copyFlags(Old: *Pow, New: emitUnaryFloatFnCall(Op: FMul, TLI, DoubleFn: LibFunc_exp2, |
2174 | FloatFn: LibFunc_exp2f, |
2175 | LongDoubleFn: LibFunc_exp2l, B, Attrs: NoAttrs)); |
2176 | } |
2177 | } |
2178 | |
2179 | // pow(10.0, x) -> exp10(x) |
2180 | if (BaseF->isExactlyValue(V: 10.0) && |
2181 | hasFloatFn(M, TLI, Ty, DoubleFn: LibFunc_exp10, FloatFn: LibFunc_exp10f, LongDoubleFn: LibFunc_exp10l)) { |
2182 | |
2183 | if (Pow->doesNotAccessMemory()) { |
2184 | CallInst *NewExp10 = |
2185 | B.CreateIntrinsic(ID: Intrinsic::exp10, Types: {Ty}, Args: {Expo}, FMFSource: Pow, Name: "exp10" ); |
2186 | return copyFlags(Old: *Pow, New: NewExp10); |
2187 | } |
2188 | |
2189 | return copyFlags(Old: *Pow, New: emitUnaryFloatFnCall(Op: Expo, TLI, DoubleFn: LibFunc_exp10, |
2190 | FloatFn: LibFunc_exp10f, LongDoubleFn: LibFunc_exp10l, |
2191 | B, Attrs: NoAttrs)); |
2192 | } |
2193 | |
2194 | // pow(x, y) -> exp2(log2(x) * y) |
2195 | if (Pow->hasApproxFunc() && Pow->hasNoNaNs() && BaseF->isFiniteNonZero() && |
2196 | !BaseF->isNegative()) { |
2197 | // pow(1, inf) is defined to be 1 but exp2(log2(1) * inf) evaluates to NaN. |
2198 | // Luckily optimizePow has already handled the x == 1 case. |
2199 | assert(!match(Base, m_FPOne()) && |
2200 | "pow(1.0, y) should have been simplified earlier!" ); |
2201 | |
2202 | Value *Log = nullptr; |
2203 | if (Ty->isFloatTy()) |
2204 | Log = ConstantFP::get(Ty, V: std::log2(x: BaseF->convertToFloat())); |
2205 | else if (Ty->isDoubleTy()) |
2206 | Log = ConstantFP::get(Ty, V: std::log2(x: BaseF->convertToDouble())); |
2207 | |
2208 | if (Log) { |
2209 | Value *FMul = B.CreateFMul(L: Log, R: Expo, Name: "mul" ); |
2210 | if (Pow->doesNotAccessMemory()) |
2211 | return copyFlags(Old: *Pow, New: B.CreateUnaryIntrinsic(ID: Intrinsic::exp2, V: FMul, |
2212 | FMFSource: nullptr, Name: "exp2" )); |
2213 | else if (hasFloatFn(M, TLI, Ty, DoubleFn: LibFunc_exp2, FloatFn: LibFunc_exp2f, |
2214 | LongDoubleFn: LibFunc_exp2l)) |
2215 | return copyFlags(Old: *Pow, New: emitUnaryFloatFnCall(Op: FMul, TLI, DoubleFn: LibFunc_exp2, |
2216 | FloatFn: LibFunc_exp2f, |
2217 | LongDoubleFn: LibFunc_exp2l, B, Attrs: NoAttrs)); |
2218 | } |
2219 | } |
2220 | |
2221 | return nullptr; |
2222 | } |
2223 | |
2224 | static Value *getSqrtCall(Value *V, AttributeList Attrs, bool NoErrno, |
2225 | Module *M, IRBuilderBase &B, |
2226 | const TargetLibraryInfo *TLI) { |
2227 | // If errno is never set, then use the intrinsic for sqrt(). |
2228 | if (NoErrno) |
2229 | return B.CreateUnaryIntrinsic(ID: Intrinsic::sqrt, V, FMFSource: nullptr, Name: "sqrt" ); |
2230 | |
2231 | // Otherwise, use the libcall for sqrt(). |
2232 | if (hasFloatFn(M, TLI, Ty: V->getType(), DoubleFn: LibFunc_sqrt, FloatFn: LibFunc_sqrtf, |
2233 | LongDoubleFn: LibFunc_sqrtl)) |
2234 | // TODO: We also should check that the target can in fact lower the sqrt() |
2235 | // libcall. We currently have no way to ask this question, so we ask if |
2236 | // the target has a sqrt() libcall, which is not exactly the same. |
2237 | return emitUnaryFloatFnCall(Op: V, TLI, DoubleFn: LibFunc_sqrt, FloatFn: LibFunc_sqrtf, |
2238 | LongDoubleFn: LibFunc_sqrtl, B, Attrs); |
2239 | |
2240 | return nullptr; |
2241 | } |
2242 | |
2243 | /// Use square root in place of pow(x, +/-0.5). |
2244 | Value *LibCallSimplifier::replacePowWithSqrt(CallInst *Pow, IRBuilderBase &B) { |
2245 | Value *Sqrt, *Base = Pow->getArgOperand(i: 0), *Expo = Pow->getArgOperand(i: 1); |
2246 | Module *Mod = Pow->getModule(); |
2247 | Type *Ty = Pow->getType(); |
2248 | |
2249 | const APFloat *ExpoF; |
2250 | if (!match(V: Expo, P: m_APFloat(Res&: ExpoF)) || |
2251 | (!ExpoF->isExactlyValue(V: 0.5) && !ExpoF->isExactlyValue(V: -0.5))) |
2252 | return nullptr; |
2253 | |
2254 | // Converting pow(X, -0.5) to 1/sqrt(X) may introduce an extra rounding step, |
2255 | // so that requires fast-math-flags (afn or reassoc). |
2256 | if (ExpoF->isNegative() && (!Pow->hasApproxFunc() && !Pow->hasAllowReassoc())) |
2257 | return nullptr; |
2258 | |
2259 | // If we have a pow() library call (accesses memory) and we can't guarantee |
2260 | // that the base is not an infinity, give up: |
2261 | // pow(-Inf, 0.5) is optionally required to have a result of +Inf (not setting |
2262 | // errno), but sqrt(-Inf) is required by various standards to set errno. |
2263 | if (!Pow->doesNotAccessMemory() && !Pow->hasNoInfs() && |
2264 | !isKnownNeverInfinity( |
2265 | V: Base, SQ: SimplifyQuery(DL, TLI, DT, AC, Pow, true, true, DC))) |
2266 | return nullptr; |
2267 | |
2268 | Sqrt = getSqrtCall(V: Base, Attrs: AttributeList(), NoErrno: Pow->doesNotAccessMemory(), M: Mod, B, |
2269 | TLI); |
2270 | if (!Sqrt) |
2271 | return nullptr; |
2272 | |
2273 | // Handle signed zero base by expanding to fabs(sqrt(x)). |
2274 | if (!Pow->hasNoSignedZeros()) |
2275 | Sqrt = B.CreateUnaryIntrinsic(ID: Intrinsic::fabs, V: Sqrt, FMFSource: nullptr, Name: "abs" ); |
2276 | |
2277 | Sqrt = copyFlags(Old: *Pow, New: Sqrt); |
2278 | |
2279 | // Handle non finite base by expanding to |
2280 | // (x == -infinity ? +infinity : sqrt(x)). |
2281 | if (!Pow->hasNoInfs()) { |
2282 | Value *PosInf = ConstantFP::getInfinity(Ty), |
2283 | *NegInf = ConstantFP::getInfinity(Ty, Negative: true); |
2284 | Value *FCmp = B.CreateFCmpOEQ(LHS: Base, RHS: NegInf, Name: "isinf" ); |
2285 | Sqrt = B.CreateSelect(C: FCmp, True: PosInf, False: Sqrt); |
2286 | } |
2287 | |
2288 | // If the exponent is negative, then get the reciprocal. |
2289 | if (ExpoF->isNegative()) |
2290 | Sqrt = B.CreateFDiv(L: ConstantFP::get(Ty, V: 1.0), R: Sqrt, Name: "reciprocal" ); |
2291 | |
2292 | return Sqrt; |
2293 | } |
2294 | |
2295 | static Value *createPowWithIntegerExponent(Value *Base, Value *Expo, Module *M, |
2296 | IRBuilderBase &B) { |
2297 | Value *Args[] = {Base, Expo}; |
2298 | Type *Types[] = {Base->getType(), Expo->getType()}; |
2299 | return B.CreateIntrinsic(ID: Intrinsic::powi, Types, Args); |
2300 | } |
2301 | |
2302 | Value *LibCallSimplifier::optimizePow(CallInst *Pow, IRBuilderBase &B) { |
2303 | Value *Base = Pow->getArgOperand(i: 0); |
2304 | Value *Expo = Pow->getArgOperand(i: 1); |
2305 | Function *Callee = Pow->getCalledFunction(); |
2306 | StringRef Name = Callee->getName(); |
2307 | Type *Ty = Pow->getType(); |
2308 | Module *M = Pow->getModule(); |
2309 | bool AllowApprox = Pow->hasApproxFunc(); |
2310 | bool Ignored; |
2311 | |
2312 | // Propagate the math semantics from the call to any created instructions. |
2313 | IRBuilderBase::FastMathFlagGuard Guard(B); |
2314 | B.setFastMathFlags(Pow->getFastMathFlags()); |
2315 | // Evaluate special cases related to the base. |
2316 | |
2317 | // pow(1.0, x) -> 1.0 |
2318 | if (match(V: Base, P: m_FPOne())) |
2319 | return Base; |
2320 | |
2321 | if (Value *Exp = replacePowWithExp(Pow, B)) |
2322 | return Exp; |
2323 | |
2324 | // Evaluate special cases related to the exponent. |
2325 | |
2326 | // pow(x, -1.0) -> 1.0 / x |
2327 | if (match(V: Expo, P: m_SpecificFP(V: -1.0))) |
2328 | return B.CreateFDiv(L: ConstantFP::get(Ty, V: 1.0), R: Base, Name: "reciprocal" ); |
2329 | |
2330 | // pow(x, +/-0.0) -> 1.0 |
2331 | if (match(V: Expo, P: m_AnyZeroFP())) |
2332 | return ConstantFP::get(Ty, V: 1.0); |
2333 | |
2334 | // pow(x, 1.0) -> x |
2335 | if (match(V: Expo, P: m_FPOne())) |
2336 | return Base; |
2337 | |
2338 | // pow(x, 2.0) -> x * x |
2339 | if (match(V: Expo, P: m_SpecificFP(V: 2.0))) |
2340 | return B.CreateFMul(L: Base, R: Base, Name: "square" ); |
2341 | |
2342 | if (Value *Sqrt = replacePowWithSqrt(Pow, B)) |
2343 | return Sqrt; |
2344 | |
2345 | // If we can approximate pow: |
2346 | // pow(x, n) -> powi(x, n) * sqrt(x) if n has exactly a 0.5 fraction |
2347 | // pow(x, n) -> powi(x, n) if n is a constant signed integer value |
2348 | const APFloat *ExpoF; |
2349 | if (AllowApprox && match(V: Expo, P: m_APFloat(Res&: ExpoF)) && |
2350 | !ExpoF->isExactlyValue(V: 0.5) && !ExpoF->isExactlyValue(V: -0.5)) { |
2351 | APFloat ExpoA(abs(X: *ExpoF)); |
2352 | APFloat ExpoI(*ExpoF); |
2353 | Value *Sqrt = nullptr; |
2354 | if (!ExpoA.isInteger()) { |
2355 | APFloat Expo2 = ExpoA; |
2356 | // To check if ExpoA is an integer + 0.5, we add it to itself. If there |
2357 | // is no floating point exception and the result is an integer, then |
2358 | // ExpoA == integer + 0.5 |
2359 | if (Expo2.add(RHS: ExpoA, RM: APFloat::rmNearestTiesToEven) != APFloat::opOK) |
2360 | return nullptr; |
2361 | |
2362 | if (!Expo2.isInteger()) |
2363 | return nullptr; |
2364 | |
2365 | if (ExpoI.roundToIntegral(RM: APFloat::rmTowardNegative) != |
2366 | APFloat::opInexact) |
2367 | return nullptr; |
2368 | if (!ExpoI.isInteger()) |
2369 | return nullptr; |
2370 | ExpoF = &ExpoI; |
2371 | |
2372 | Sqrt = getSqrtCall(V: Base, Attrs: AttributeList(), NoErrno: Pow->doesNotAccessMemory(), M, |
2373 | B, TLI); |
2374 | if (!Sqrt) |
2375 | return nullptr; |
2376 | } |
2377 | |
2378 | // 0.5 fraction is now optionally handled. |
2379 | // Do pow -> powi for remaining integer exponent |
2380 | APSInt IntExpo(TLI->getIntSize(), /*isUnsigned=*/false); |
2381 | if (ExpoF->isInteger() && |
2382 | ExpoF->convertToInteger(Result&: IntExpo, RM: APFloat::rmTowardZero, IsExact: &Ignored) == |
2383 | APFloat::opOK) { |
2384 | Value *PowI = copyFlags( |
2385 | Old: *Pow, |
2386 | New: createPowWithIntegerExponent( |
2387 | Base, Expo: ConstantInt::get(Ty: B.getIntNTy(N: TLI->getIntSize()), V: IntExpo), |
2388 | M, B)); |
2389 | |
2390 | if (PowI && Sqrt) |
2391 | return B.CreateFMul(L: PowI, R: Sqrt); |
2392 | |
2393 | return PowI; |
2394 | } |
2395 | } |
2396 | |
2397 | // powf(x, itofp(y)) -> powi(x, y) |
2398 | if (AllowApprox && (isa<SIToFPInst>(Val: Expo) || isa<UIToFPInst>(Val: Expo))) { |
2399 | if (Value *ExpoI = getIntToFPVal(I2F: Expo, B, DstWidth: TLI->getIntSize())) |
2400 | return copyFlags(Old: *Pow, New: createPowWithIntegerExponent(Base, Expo: ExpoI, M, B)); |
2401 | } |
2402 | |
2403 | // Shrink pow() to powf() if the arguments are single precision, |
2404 | // unless the result is expected to be double precision. |
2405 | if (UnsafeFPShrink && Name == TLI->getName(F: LibFunc_pow) && |
2406 | hasFloatVersion(M, FuncName: Name)) { |
2407 | if (Value *Shrunk = optimizeBinaryDoubleFP(CI: Pow, B, TLI, isPrecise: true)) |
2408 | return Shrunk; |
2409 | } |
2410 | |
2411 | return nullptr; |
2412 | } |
2413 | |
2414 | Value *LibCallSimplifier::optimizeExp2(CallInst *CI, IRBuilderBase &B) { |
2415 | Module *M = CI->getModule(); |
2416 | Function *Callee = CI->getCalledFunction(); |
2417 | StringRef Name = Callee->getName(); |
2418 | Value *Ret = nullptr; |
2419 | if (UnsafeFPShrink && Name == TLI->getName(F: LibFunc_exp2) && |
2420 | hasFloatVersion(M, FuncName: Name)) |
2421 | Ret = optimizeUnaryDoubleFP(CI, B, TLI, isPrecise: true); |
2422 | |
2423 | // If we have an llvm.exp2 intrinsic, emit the llvm.ldexp intrinsic. If we |
2424 | // have the libcall, emit the libcall. |
2425 | // |
2426 | // TODO: In principle we should be able to just always use the intrinsic for |
2427 | // any doesNotAccessMemory callsite. |
2428 | |
2429 | const bool UseIntrinsic = Callee->isIntrinsic(); |
2430 | // Bail out for vectors because the code below only expects scalars. |
2431 | Type *Ty = CI->getType(); |
2432 | if (!UseIntrinsic && Ty->isVectorTy()) |
2433 | return Ret; |
2434 | |
2435 | // exp2(sitofp(x)) -> ldexp(1.0, sext(x)) if sizeof(x) <= IntSize |
2436 | // exp2(uitofp(x)) -> ldexp(1.0, zext(x)) if sizeof(x) < IntSize |
2437 | Value *Op = CI->getArgOperand(i: 0); |
2438 | if ((isa<SIToFPInst>(Val: Op) || isa<UIToFPInst>(Val: Op)) && |
2439 | (UseIntrinsic || |
2440 | hasFloatFn(M, TLI, Ty, DoubleFn: LibFunc_ldexp, FloatFn: LibFunc_ldexpf, LongDoubleFn: LibFunc_ldexpl))) { |
2441 | if (Value *Exp = getIntToFPVal(I2F: Op, B, DstWidth: TLI->getIntSize())) { |
2442 | Constant *One = ConstantFP::get(Ty, V: 1.0); |
2443 | |
2444 | if (UseIntrinsic) { |
2445 | return copyFlags(Old: *CI, New: B.CreateIntrinsic(ID: Intrinsic::ldexp, |
2446 | Types: {Ty, Exp->getType()}, |
2447 | Args: {One, Exp}, FMFSource: CI)); |
2448 | } |
2449 | |
2450 | IRBuilderBase::FastMathFlagGuard Guard(B); |
2451 | B.setFastMathFlags(CI->getFastMathFlags()); |
2452 | return copyFlags(Old: *CI, New: emitBinaryFloatFnCall( |
2453 | Op1: One, Op2: Exp, TLI, DoubleFn: LibFunc_ldexp, FloatFn: LibFunc_ldexpf, |
2454 | LongDoubleFn: LibFunc_ldexpl, B, Attrs: AttributeList())); |
2455 | } |
2456 | } |
2457 | |
2458 | return Ret; |
2459 | } |
2460 | |
2461 | Value *LibCallSimplifier::optimizeFMinFMax(CallInst *CI, IRBuilderBase &B) { |
2462 | Module *M = CI->getModule(); |
2463 | |
2464 | // If we can shrink the call to a float function rather than a double |
2465 | // function, do that first. |
2466 | Function *Callee = CI->getCalledFunction(); |
2467 | StringRef Name = Callee->getName(); |
2468 | if ((Name == "fmin" || Name == "fmax" ) && hasFloatVersion(M, FuncName: Name)) |
2469 | if (Value *Ret = optimizeBinaryDoubleFP(CI, B, TLI)) |
2470 | return Ret; |
2471 | |
2472 | // The LLVM intrinsics minnum/maxnum correspond to fmin/fmax. Canonicalize to |
2473 | // the intrinsics for improved optimization (for example, vectorization). |
2474 | // No-signed-zeros is implied by the definitions of fmax/fmin themselves. |
2475 | // From the C standard draft WG14/N1256: |
2476 | // "Ideally, fmax would be sensitive to the sign of zero, for example |
2477 | // fmax(-0.0, +0.0) would return +0; however, implementation in software |
2478 | // might be impractical." |
2479 | FastMathFlags FMF = CI->getFastMathFlags(); |
2480 | FMF.setNoSignedZeros(); |
2481 | |
2482 | Intrinsic::ID IID = Callee->getName().starts_with(Prefix: "fmin" ) ? Intrinsic::minnum |
2483 | : Intrinsic::maxnum; |
2484 | return copyFlags(Old: *CI, New: B.CreateBinaryIntrinsic(ID: IID, LHS: CI->getArgOperand(i: 0), |
2485 | RHS: CI->getArgOperand(i: 1), FMFSource: FMF)); |
2486 | } |
2487 | |
2488 | Value *LibCallSimplifier::optimizeLog(CallInst *Log, IRBuilderBase &B) { |
2489 | Function *LogFn = Log->getCalledFunction(); |
2490 | StringRef LogNm = LogFn->getName(); |
2491 | Intrinsic::ID LogID = LogFn->getIntrinsicID(); |
2492 | Module *Mod = Log->getModule(); |
2493 | Type *Ty = Log->getType(); |
2494 | |
2495 | if (UnsafeFPShrink && hasFloatVersion(M: Mod, FuncName: LogNm)) |
2496 | if (Value *Ret = optimizeUnaryDoubleFP(CI: Log, B, TLI, isPrecise: true)) |
2497 | return Ret; |
2498 | |
2499 | LibFunc LogLb, ExpLb, Exp2Lb, Exp10Lb, PowLb; |
2500 | |
2501 | // This is only applicable to log(), log2(), log10(). |
2502 | if (TLI->getLibFunc(funcName: LogNm, F&: LogLb)) { |
2503 | switch (LogLb) { |
2504 | case LibFunc_logf: |
2505 | LogID = Intrinsic::log; |
2506 | ExpLb = LibFunc_expf; |
2507 | Exp2Lb = LibFunc_exp2f; |
2508 | Exp10Lb = LibFunc_exp10f; |
2509 | PowLb = LibFunc_powf; |
2510 | break; |
2511 | case LibFunc_log: |
2512 | LogID = Intrinsic::log; |
2513 | ExpLb = LibFunc_exp; |
2514 | Exp2Lb = LibFunc_exp2; |
2515 | Exp10Lb = LibFunc_exp10; |
2516 | PowLb = LibFunc_pow; |
2517 | break; |
2518 | case LibFunc_logl: |
2519 | LogID = Intrinsic::log; |
2520 | ExpLb = LibFunc_expl; |
2521 | Exp2Lb = LibFunc_exp2l; |
2522 | Exp10Lb = LibFunc_exp10l; |
2523 | PowLb = LibFunc_powl; |
2524 | break; |
2525 | case LibFunc_log2f: |
2526 | LogID = Intrinsic::log2; |
2527 | ExpLb = LibFunc_expf; |
2528 | Exp2Lb = LibFunc_exp2f; |
2529 | Exp10Lb = LibFunc_exp10f; |
2530 | PowLb = LibFunc_powf; |
2531 | break; |
2532 | case LibFunc_log2: |
2533 | LogID = Intrinsic::log2; |
2534 | ExpLb = LibFunc_exp; |
2535 | Exp2Lb = LibFunc_exp2; |
2536 | Exp10Lb = LibFunc_exp10; |
2537 | PowLb = LibFunc_pow; |
2538 | break; |
2539 | case LibFunc_log2l: |
2540 | LogID = Intrinsic::log2; |
2541 | ExpLb = LibFunc_expl; |
2542 | Exp2Lb = LibFunc_exp2l; |
2543 | Exp10Lb = LibFunc_exp10l; |
2544 | PowLb = LibFunc_powl; |
2545 | break; |
2546 | case LibFunc_log10f: |
2547 | LogID = Intrinsic::log10; |
2548 | ExpLb = LibFunc_expf; |
2549 | Exp2Lb = LibFunc_exp2f; |
2550 | Exp10Lb = LibFunc_exp10f; |
2551 | PowLb = LibFunc_powf; |
2552 | break; |
2553 | case LibFunc_log10: |
2554 | LogID = Intrinsic::log10; |
2555 | ExpLb = LibFunc_exp; |
2556 | Exp2Lb = LibFunc_exp2; |
2557 | Exp10Lb = LibFunc_exp10; |
2558 | PowLb = LibFunc_pow; |
2559 | break; |
2560 | case LibFunc_log10l: |
2561 | LogID = Intrinsic::log10; |
2562 | ExpLb = LibFunc_expl; |
2563 | Exp2Lb = LibFunc_exp2l; |
2564 | Exp10Lb = LibFunc_exp10l; |
2565 | PowLb = LibFunc_powl; |
2566 | break; |
2567 | default: |
2568 | return nullptr; |
2569 | } |
2570 | |
2571 | // Convert libcall to intrinsic if the value is known > 0. |
2572 | bool IsKnownNoErrno = Log->hasNoNaNs() && Log->hasNoInfs(); |
2573 | if (!IsKnownNoErrno) { |
2574 | SimplifyQuery SQ(DL, TLI, DT, AC, Log, true, true, DC); |
2575 | KnownFPClass Known = computeKnownFPClass( |
2576 | V: Log->getOperand(i_nocapture: 0), |
2577 | InterestedClasses: KnownFPClass::OrderedLessThanZeroMask | fcSubnormal, SQ); |
2578 | Function *F = Log->getParent()->getParent(); |
2579 | const fltSemantics &FltSem = Ty->getScalarType()->getFltSemantics(); |
2580 | IsKnownNoErrno = |
2581 | Known.cannotBeOrderedLessThanZero() && |
2582 | Known.isKnownNeverLogicalZero(Mode: F->getDenormalMode(FPType: FltSem)); |
2583 | } |
2584 | if (IsKnownNoErrno) { |
2585 | auto *NewLog = B.CreateUnaryIntrinsic(ID: LogID, V: Log->getArgOperand(i: 0), FMFSource: Log); |
2586 | NewLog->copyMetadata(SrcInst: *Log); |
2587 | return copyFlags(Old: *Log, New: NewLog); |
2588 | } |
2589 | } else if (LogID == Intrinsic::log || LogID == Intrinsic::log2 || |
2590 | LogID == Intrinsic::log10) { |
2591 | if (Ty->getScalarType()->isFloatTy()) { |
2592 | ExpLb = LibFunc_expf; |
2593 | Exp2Lb = LibFunc_exp2f; |
2594 | Exp10Lb = LibFunc_exp10f; |
2595 | PowLb = LibFunc_powf; |
2596 | } else if (Ty->getScalarType()->isDoubleTy()) { |
2597 | ExpLb = LibFunc_exp; |
2598 | Exp2Lb = LibFunc_exp2; |
2599 | Exp10Lb = LibFunc_exp10; |
2600 | PowLb = LibFunc_pow; |
2601 | } else |
2602 | return nullptr; |
2603 | } else |
2604 | return nullptr; |
2605 | |
2606 | // The earlier call must also be 'fast' in order to do these transforms. |
2607 | CallInst *Arg = dyn_cast<CallInst>(Val: Log->getArgOperand(i: 0)); |
2608 | if (!Log->isFast() || !Arg || !Arg->isFast() || !Arg->hasOneUse()) |
2609 | return nullptr; |
2610 | |
2611 | IRBuilderBase::FastMathFlagGuard Guard(B); |
2612 | B.setFastMathFlags(FastMathFlags::getFast()); |
2613 | |
2614 | Intrinsic::ID ArgID = Arg->getIntrinsicID(); |
2615 | LibFunc ArgLb = NotLibFunc; |
2616 | TLI->getLibFunc(CB: *Arg, F&: ArgLb); |
2617 | |
2618 | // log(pow(x,y)) -> y*log(x) |
2619 | AttributeList NoAttrs; |
2620 | if (ArgLb == PowLb || ArgID == Intrinsic::pow || ArgID == Intrinsic::powi) { |
2621 | Value *LogX = |
2622 | Log->doesNotAccessMemory() |
2623 | ? B.CreateUnaryIntrinsic(ID: LogID, V: Arg->getOperand(i_nocapture: 0), FMFSource: nullptr, Name: "log" ) |
2624 | : emitUnaryFloatFnCall(Op: Arg->getOperand(i_nocapture: 0), TLI, Name: LogNm, B, Attrs: NoAttrs); |
2625 | Value *Y = Arg->getArgOperand(i: 1); |
2626 | // Cast exponent to FP if integer. |
2627 | if (ArgID == Intrinsic::powi) |
2628 | Y = B.CreateSIToFP(V: Y, DestTy: Ty, Name: "cast" ); |
2629 | Value *MulY = B.CreateFMul(L: Y, R: LogX, Name: "mul" ); |
2630 | // Since pow() may have side effects, e.g. errno, |
2631 | // dead code elimination may not be trusted to remove it. |
2632 | substituteInParent(I: Arg, With: MulY); |
2633 | return MulY; |
2634 | } |
2635 | |
2636 | // log(exp{,2,10}(y)) -> y*log({e,2,10}) |
2637 | // TODO: There is no exp10() intrinsic yet. |
2638 | if (ArgLb == ExpLb || ArgLb == Exp2Lb || ArgLb == Exp10Lb || |
2639 | ArgID == Intrinsic::exp || ArgID == Intrinsic::exp2) { |
2640 | Constant *Eul; |
2641 | if (ArgLb == ExpLb || ArgID == Intrinsic::exp) |
2642 | // FIXME: Add more precise value of e for long double. |
2643 | Eul = ConstantFP::get(Ty: Log->getType(), V: numbers::e); |
2644 | else if (ArgLb == Exp2Lb || ArgID == Intrinsic::exp2) |
2645 | Eul = ConstantFP::get(Ty: Log->getType(), V: 2.0); |
2646 | else |
2647 | Eul = ConstantFP::get(Ty: Log->getType(), V: 10.0); |
2648 | Value *LogE = Log->doesNotAccessMemory() |
2649 | ? B.CreateUnaryIntrinsic(ID: LogID, V: Eul, FMFSource: nullptr, Name: "log" ) |
2650 | : emitUnaryFloatFnCall(Op: Eul, TLI, Name: LogNm, B, Attrs: NoAttrs); |
2651 | Value *MulY = B.CreateFMul(L: Arg->getArgOperand(i: 0), R: LogE, Name: "mul" ); |
2652 | // Since exp() may have side effects, e.g. errno, |
2653 | // dead code elimination may not be trusted to remove it. |
2654 | substituteInParent(I: Arg, With: MulY); |
2655 | return MulY; |
2656 | } |
2657 | |
2658 | return nullptr; |
2659 | } |
2660 | |
2661 | // sqrt(exp(X)) -> exp(X * 0.5) |
2662 | Value *LibCallSimplifier::mergeSqrtToExp(CallInst *CI, IRBuilderBase &B) { |
2663 | if (!CI->hasAllowReassoc()) |
2664 | return nullptr; |
2665 | |
2666 | Function *SqrtFn = CI->getCalledFunction(); |
2667 | CallInst *Arg = dyn_cast<CallInst>(Val: CI->getArgOperand(i: 0)); |
2668 | if (!Arg || !Arg->hasAllowReassoc() || !Arg->hasOneUse()) |
2669 | return nullptr; |
2670 | Intrinsic::ID ArgID = Arg->getIntrinsicID(); |
2671 | LibFunc ArgLb = NotLibFunc; |
2672 | TLI->getLibFunc(CB: *Arg, F&: ArgLb); |
2673 | |
2674 | LibFunc SqrtLb, ExpLb, Exp2Lb, Exp10Lb; |
2675 | |
2676 | if (TLI->getLibFunc(funcName: SqrtFn->getName(), F&: SqrtLb)) |
2677 | switch (SqrtLb) { |
2678 | case LibFunc_sqrtf: |
2679 | ExpLb = LibFunc_expf; |
2680 | Exp2Lb = LibFunc_exp2f; |
2681 | Exp10Lb = LibFunc_exp10f; |
2682 | break; |
2683 | case LibFunc_sqrt: |
2684 | ExpLb = LibFunc_exp; |
2685 | Exp2Lb = LibFunc_exp2; |
2686 | Exp10Lb = LibFunc_exp10; |
2687 | break; |
2688 | case LibFunc_sqrtl: |
2689 | ExpLb = LibFunc_expl; |
2690 | Exp2Lb = LibFunc_exp2l; |
2691 | Exp10Lb = LibFunc_exp10l; |
2692 | break; |
2693 | default: |
2694 | return nullptr; |
2695 | } |
2696 | else if (SqrtFn->getIntrinsicID() == Intrinsic::sqrt) { |
2697 | if (CI->getType()->getScalarType()->isFloatTy()) { |
2698 | ExpLb = LibFunc_expf; |
2699 | Exp2Lb = LibFunc_exp2f; |
2700 | Exp10Lb = LibFunc_exp10f; |
2701 | } else if (CI->getType()->getScalarType()->isDoubleTy()) { |
2702 | ExpLb = LibFunc_exp; |
2703 | Exp2Lb = LibFunc_exp2; |
2704 | Exp10Lb = LibFunc_exp10; |
2705 | } else |
2706 | return nullptr; |
2707 | } else |
2708 | return nullptr; |
2709 | |
2710 | if (ArgLb != ExpLb && ArgLb != Exp2Lb && ArgLb != Exp10Lb && |
2711 | ArgID != Intrinsic::exp && ArgID != Intrinsic::exp2) |
2712 | return nullptr; |
2713 | |
2714 | IRBuilderBase::InsertPointGuard Guard(B); |
2715 | B.SetInsertPoint(Arg); |
2716 | auto *ExpOperand = Arg->getOperand(i_nocapture: 0); |
2717 | auto *FMul = |
2718 | B.CreateFMulFMF(L: ExpOperand, R: ConstantFP::get(Ty: ExpOperand->getType(), V: 0.5), |
2719 | FMFSource: CI, Name: "merged.sqrt" ); |
2720 | |
2721 | Arg->setOperand(i_nocapture: 0, Val_nocapture: FMul); |
2722 | return Arg; |
2723 | } |
2724 | |
2725 | Value *LibCallSimplifier::optimizeSqrt(CallInst *CI, IRBuilderBase &B) { |
2726 | Module *M = CI->getModule(); |
2727 | Function *Callee = CI->getCalledFunction(); |
2728 | Value *Ret = nullptr; |
2729 | // TODO: Once we have a way (other than checking for the existince of the |
2730 | // libcall) to tell whether our target can lower @llvm.sqrt, relax the |
2731 | // condition below. |
2732 | if (isLibFuncEmittable(M, TLI, TheLibFunc: LibFunc_sqrtf) && |
2733 | (Callee->getName() == "sqrt" || |
2734 | Callee->getIntrinsicID() == Intrinsic::sqrt)) |
2735 | Ret = optimizeUnaryDoubleFP(CI, B, TLI, isPrecise: true); |
2736 | |
2737 | if (Value *Opt = mergeSqrtToExp(CI, B)) |
2738 | return Opt; |
2739 | |
2740 | if (!CI->isFast()) |
2741 | return Ret; |
2742 | |
2743 | Instruction *I = dyn_cast<Instruction>(Val: CI->getArgOperand(i: 0)); |
2744 | if (!I || I->getOpcode() != Instruction::FMul || !I->isFast()) |
2745 | return Ret; |
2746 | |
2747 | // We're looking for a repeated factor in a multiplication tree, |
2748 | // so we can do this fold: sqrt(x * x) -> fabs(x); |
2749 | // or this fold: sqrt((x * x) * y) -> fabs(x) * sqrt(y). |
2750 | Value *Op0 = I->getOperand(i: 0); |
2751 | Value *Op1 = I->getOperand(i: 1); |
2752 | Value *RepeatOp = nullptr; |
2753 | Value *OtherOp = nullptr; |
2754 | if (Op0 == Op1) { |
2755 | // Simple match: the operands of the multiply are identical. |
2756 | RepeatOp = Op0; |
2757 | } else { |
2758 | // Look for a more complicated pattern: one of the operands is itself |
2759 | // a multiply, so search for a common factor in that multiply. |
2760 | // Note: We don't bother looking any deeper than this first level or for |
2761 | // variations of this pattern because instcombine's visitFMUL and/or the |
2762 | // reassociation pass should give us this form. |
2763 | Value *MulOp; |
2764 | if (match(V: Op0, P: m_FMul(L: m_Value(V&: MulOp), R: m_Deferred(V: MulOp))) && |
2765 | cast<Instruction>(Val: Op0)->isFast()) { |
2766 | // Pattern: sqrt((x * x) * z) |
2767 | RepeatOp = MulOp; |
2768 | OtherOp = Op1; |
2769 | } else if (match(V: Op1, P: m_FMul(L: m_Value(V&: MulOp), R: m_Deferred(V: MulOp))) && |
2770 | cast<Instruction>(Val: Op1)->isFast()) { |
2771 | // Pattern: sqrt(z * (x * x)) |
2772 | RepeatOp = MulOp; |
2773 | OtherOp = Op0; |
2774 | } |
2775 | } |
2776 | if (!RepeatOp) |
2777 | return Ret; |
2778 | |
2779 | // Fast math flags for any created instructions should match the sqrt |
2780 | // and multiply. |
2781 | |
2782 | // If we found a repeated factor, hoist it out of the square root and |
2783 | // replace it with the fabs of that factor. |
2784 | Value *FabsCall = |
2785 | B.CreateUnaryIntrinsic(ID: Intrinsic::fabs, V: RepeatOp, FMFSource: I, Name: "fabs" ); |
2786 | if (OtherOp) { |
2787 | // If we found a non-repeated factor, we still need to get its square |
2788 | // root. We then multiply that by the value that was simplified out |
2789 | // of the square root calculation. |
2790 | Value *SqrtCall = |
2791 | B.CreateUnaryIntrinsic(ID: Intrinsic::sqrt, V: OtherOp, FMFSource: I, Name: "sqrt" ); |
2792 | return copyFlags(Old: *CI, New: B.CreateFMulFMF(L: FabsCall, R: SqrtCall, FMFSource: I)); |
2793 | } |
2794 | return copyFlags(Old: *CI, New: FabsCall); |
2795 | } |
2796 | |
2797 | Value *LibCallSimplifier::optimizeFMod(CallInst *CI, IRBuilderBase &B) { |
2798 | |
2799 | // fmod(x,y) can set errno if y == 0 or x == +/-inf, and returns Nan in those |
2800 | // case. If we know those do not happen, then we can convert the fmod into |
2801 | // frem. |
2802 | bool IsNoNan = CI->hasNoNaNs(); |
2803 | if (!IsNoNan) { |
2804 | SimplifyQuery SQ(DL, TLI, DT, AC, CI, true, true, DC); |
2805 | KnownFPClass Known0 = computeKnownFPClass(V: CI->getOperand(i_nocapture: 0), InterestedClasses: fcInf, SQ); |
2806 | if (Known0.isKnownNeverInfinity()) { |
2807 | KnownFPClass Known1 = |
2808 | computeKnownFPClass(V: CI->getOperand(i_nocapture: 1), InterestedClasses: fcZero | fcSubnormal, SQ); |
2809 | Function *F = CI->getParent()->getParent(); |
2810 | const fltSemantics &FltSem = |
2811 | CI->getType()->getScalarType()->getFltSemantics(); |
2812 | IsNoNan = Known1.isKnownNeverLogicalZero(Mode: F->getDenormalMode(FPType: FltSem)); |
2813 | } |
2814 | } |
2815 | |
2816 | if (IsNoNan) { |
2817 | Value *FRem = B.CreateFRemFMF(L: CI->getOperand(i_nocapture: 0), R: CI->getOperand(i_nocapture: 1), FMFSource: CI); |
2818 | if (auto *FRemI = dyn_cast<Instruction>(Val: FRem)) |
2819 | FRemI->setHasNoNaNs(true); |
2820 | return FRem; |
2821 | } |
2822 | return nullptr; |
2823 | } |
2824 | |
2825 | Value *LibCallSimplifier::optimizeTrigInversionPairs(CallInst *CI, |
2826 | IRBuilderBase &B) { |
2827 | Module *M = CI->getModule(); |
2828 | Function *Callee = CI->getCalledFunction(); |
2829 | Value *Ret = nullptr; |
2830 | StringRef Name = Callee->getName(); |
2831 | if (UnsafeFPShrink && |
2832 | (Name == "tan" || Name == "atanh" || Name == "sinh" || Name == "cosh" || |
2833 | Name == "asinh" ) && |
2834 | hasFloatVersion(M, FuncName: Name)) |
2835 | Ret = optimizeUnaryDoubleFP(CI, B, TLI, isPrecise: true); |
2836 | |
2837 | Value *Op1 = CI->getArgOperand(i: 0); |
2838 | auto *OpC = dyn_cast<CallInst>(Val: Op1); |
2839 | if (!OpC) |
2840 | return Ret; |
2841 | |
2842 | // Both calls must be 'fast' in order to remove them. |
2843 | if (!CI->isFast() || !OpC->isFast()) |
2844 | return Ret; |
2845 | |
2846 | // tan(atan(x)) -> x |
2847 | // atanh(tanh(x)) -> x |
2848 | // sinh(asinh(x)) -> x |
2849 | // asinh(sinh(x)) -> x |
2850 | // cosh(acosh(x)) -> x |
2851 | LibFunc Func; |
2852 | Function *F = OpC->getCalledFunction(); |
2853 | if (F && TLI->getLibFunc(funcName: F->getName(), F&: Func) && |
2854 | isLibFuncEmittable(M, TLI, TheLibFunc: Func)) { |
2855 | LibFunc inverseFunc = llvm::StringSwitch<LibFunc>(Callee->getName()) |
2856 | .Case(S: "tan" , Value: LibFunc_atan) |
2857 | .Case(S: "atanh" , Value: LibFunc_tanh) |
2858 | .Case(S: "sinh" , Value: LibFunc_asinh) |
2859 | .Case(S: "cosh" , Value: LibFunc_acosh) |
2860 | .Case(S: "tanf" , Value: LibFunc_atanf) |
2861 | .Case(S: "atanhf" , Value: LibFunc_tanhf) |
2862 | .Case(S: "sinhf" , Value: LibFunc_asinhf) |
2863 | .Case(S: "coshf" , Value: LibFunc_acoshf) |
2864 | .Case(S: "tanl" , Value: LibFunc_atanl) |
2865 | .Case(S: "atanhl" , Value: LibFunc_tanhl) |
2866 | .Case(S: "sinhl" , Value: LibFunc_asinhl) |
2867 | .Case(S: "coshl" , Value: LibFunc_acoshl) |
2868 | .Case(S: "asinh" , Value: LibFunc_sinh) |
2869 | .Case(S: "asinhf" , Value: LibFunc_sinhf) |
2870 | .Case(S: "asinhl" , Value: LibFunc_sinhl) |
2871 | .Default(Value: NumLibFuncs); // Used as error value |
2872 | if (Func == inverseFunc) |
2873 | Ret = OpC->getArgOperand(i: 0); |
2874 | } |
2875 | return Ret; |
2876 | } |
2877 | |
2878 | static bool isTrigLibCall(CallInst *CI) { |
2879 | // We can only hope to do anything useful if we can ignore things like errno |
2880 | // and floating-point exceptions. |
2881 | // We already checked the prototype. |
2882 | return CI->doesNotThrow() && CI->doesNotAccessMemory(); |
2883 | } |
2884 | |
2885 | static bool insertSinCosCall(IRBuilderBase &B, Function *OrigCallee, Value *Arg, |
2886 | bool UseFloat, Value *&Sin, Value *&Cos, |
2887 | Value *&SinCos, const TargetLibraryInfo *TLI) { |
2888 | Module *M = OrigCallee->getParent(); |
2889 | Type *ArgTy = Arg->getType(); |
2890 | Type *ResTy; |
2891 | StringRef Name; |
2892 | |
2893 | Triple T(OrigCallee->getParent()->getTargetTriple()); |
2894 | if (UseFloat) { |
2895 | Name = "__sincospif_stret" ; |
2896 | |
2897 | assert(T.getArch() != Triple::x86 && "x86 messy and unsupported for now" ); |
2898 | // x86_64 can't use {float, float} since that would be returned in both |
2899 | // xmm0 and xmm1, which isn't what a real struct would do. |
2900 | ResTy = T.getArch() == Triple::x86_64 |
2901 | ? static_cast<Type *>(FixedVectorType::get(ElementType: ArgTy, NumElts: 2)) |
2902 | : static_cast<Type *>(StructType::get(elt1: ArgTy, elts: ArgTy)); |
2903 | } else { |
2904 | Name = "__sincospi_stret" ; |
2905 | ResTy = StructType::get(elt1: ArgTy, elts: ArgTy); |
2906 | } |
2907 | |
2908 | if (!isLibFuncEmittable(M, TLI, Name)) |
2909 | return false; |
2910 | LibFunc TheLibFunc; |
2911 | TLI->getLibFunc(funcName: Name, F&: TheLibFunc); |
2912 | FunctionCallee Callee = getOrInsertLibFunc( |
2913 | M, TLI: *TLI, TheLibFunc, AttributeList: OrigCallee->getAttributes(), RetTy: ResTy, Args: ArgTy); |
2914 | |
2915 | if (Instruction *ArgInst = dyn_cast<Instruction>(Val: Arg)) { |
2916 | // If the argument is an instruction, it must dominate all uses so put our |
2917 | // sincos call there. |
2918 | B.SetInsertPoint(TheBB: ArgInst->getParent(), IP: ++ArgInst->getIterator()); |
2919 | } else { |
2920 | // Otherwise (e.g. for a constant) the beginning of the function is as |
2921 | // good a place as any. |
2922 | BasicBlock &EntryBB = B.GetInsertBlock()->getParent()->getEntryBlock(); |
2923 | B.SetInsertPoint(TheBB: &EntryBB, IP: EntryBB.begin()); |
2924 | } |
2925 | |
2926 | SinCos = B.CreateCall(Callee, Args: Arg, Name: "sincospi" ); |
2927 | |
2928 | if (SinCos->getType()->isStructTy()) { |
2929 | Sin = B.CreateExtractValue(Agg: SinCos, Idxs: 0, Name: "sinpi" ); |
2930 | Cos = B.CreateExtractValue(Agg: SinCos, Idxs: 1, Name: "cospi" ); |
2931 | } else { |
2932 | Sin = B.CreateExtractElement(Vec: SinCos, Idx: ConstantInt::get(Ty: B.getInt32Ty(), V: 0), |
2933 | Name: "sinpi" ); |
2934 | Cos = B.CreateExtractElement(Vec: SinCos, Idx: ConstantInt::get(Ty: B.getInt32Ty(), V: 1), |
2935 | Name: "cospi" ); |
2936 | } |
2937 | |
2938 | return true; |
2939 | } |
2940 | |
2941 | static Value *optimizeSymmetricCall(CallInst *CI, bool IsEven, |
2942 | IRBuilderBase &B) { |
2943 | Value *X; |
2944 | Value *Src = CI->getArgOperand(i: 0); |
2945 | |
2946 | if (match(V: Src, P: m_OneUse(SubPattern: m_FNeg(X: m_Value(V&: X))))) { |
2947 | auto *Call = B.CreateCall(Callee: CI->getCalledFunction(), Args: {X}); |
2948 | Call->copyFastMathFlags(I: CI); |
2949 | auto *CallInst = copyFlags(Old: *CI, New: Call); |
2950 | if (IsEven) { |
2951 | // Even function: f(-x) = f(x) |
2952 | return CallInst; |
2953 | } |
2954 | // Odd function: f(-x) = -f(x) |
2955 | return B.CreateFNegFMF(V: CallInst, FMFSource: CI); |
2956 | } |
2957 | |
2958 | // Even function: f(abs(x)) = f(x), f(copysign(x, y)) = f(x) |
2959 | if (IsEven && (match(V: Src, P: m_FAbs(Op0: m_Value(V&: X))) || |
2960 | match(V: Src, P: m_CopySign(Op0: m_Value(V&: X), Op1: m_Value())))) { |
2961 | auto *Call = B.CreateCall(Callee: CI->getCalledFunction(), Args: {X}); |
2962 | Call->copyFastMathFlags(I: CI); |
2963 | return copyFlags(Old: *CI, New: Call); |
2964 | } |
2965 | |
2966 | return nullptr; |
2967 | } |
2968 | |
2969 | Value *LibCallSimplifier::optimizeSymmetric(CallInst *CI, LibFunc Func, |
2970 | IRBuilderBase &B) { |
2971 | switch (Func) { |
2972 | case LibFunc_cos: |
2973 | case LibFunc_cosf: |
2974 | case LibFunc_cosl: |
2975 | return optimizeSymmetricCall(CI, /*IsEven*/ true, B); |
2976 | |
2977 | case LibFunc_sin: |
2978 | case LibFunc_sinf: |
2979 | case LibFunc_sinl: |
2980 | |
2981 | case LibFunc_tan: |
2982 | case LibFunc_tanf: |
2983 | case LibFunc_tanl: |
2984 | |
2985 | case LibFunc_erf: |
2986 | case LibFunc_erff: |
2987 | case LibFunc_erfl: |
2988 | return optimizeSymmetricCall(CI, /*IsEven*/ false, B); |
2989 | |
2990 | default: |
2991 | return nullptr; |
2992 | } |
2993 | } |
2994 | |
2995 | Value *LibCallSimplifier::optimizeSinCosPi(CallInst *CI, bool IsSin, IRBuilderBase &B) { |
2996 | // Make sure the prototype is as expected, otherwise the rest of the |
2997 | // function is probably invalid and likely to abort. |
2998 | if (!isTrigLibCall(CI)) |
2999 | return nullptr; |
3000 | |
3001 | Value *Arg = CI->getArgOperand(i: 0); |
3002 | if (isa<ConstantData>(Val: Arg)) |
3003 | return nullptr; |
3004 | |
3005 | SmallVector<CallInst *, 1> SinCalls; |
3006 | SmallVector<CallInst *, 1> CosCalls; |
3007 | SmallVector<CallInst *, 1> SinCosCalls; |
3008 | |
3009 | bool IsFloat = Arg->getType()->isFloatTy(); |
3010 | |
3011 | // Look for all compatible sinpi, cospi and sincospi calls with the same |
3012 | // argument. If there are enough (in some sense) we can make the |
3013 | // substitution. |
3014 | Function *F = CI->getFunction(); |
3015 | for (User *U : Arg->users()) |
3016 | classifyArgUse(Val: U, F, IsFloat, SinCalls, CosCalls, SinCosCalls); |
3017 | |
3018 | // It's only worthwhile if both sinpi and cospi are actually used. |
3019 | if (SinCalls.empty() || CosCalls.empty()) |
3020 | return nullptr; |
3021 | |
3022 | Value *Sin, *Cos, *SinCos; |
3023 | if (!insertSinCosCall(B, OrigCallee: CI->getCalledFunction(), Arg, UseFloat: IsFloat, Sin, Cos, |
3024 | SinCos, TLI)) |
3025 | return nullptr; |
3026 | |
3027 | auto replaceTrigInsts = [this](SmallVectorImpl<CallInst *> &Calls, |
3028 | Value *Res) { |
3029 | for (CallInst *C : Calls) |
3030 | replaceAllUsesWith(I: C, With: Res); |
3031 | }; |
3032 | |
3033 | replaceTrigInsts(SinCalls, Sin); |
3034 | replaceTrigInsts(CosCalls, Cos); |
3035 | replaceTrigInsts(SinCosCalls, SinCos); |
3036 | |
3037 | return IsSin ? Sin : Cos; |
3038 | } |
3039 | |
3040 | void LibCallSimplifier::classifyArgUse( |
3041 | Value *Val, Function *F, bool IsFloat, |
3042 | SmallVectorImpl<CallInst *> &SinCalls, |
3043 | SmallVectorImpl<CallInst *> &CosCalls, |
3044 | SmallVectorImpl<CallInst *> &SinCosCalls) { |
3045 | auto *CI = dyn_cast<CallInst>(Val); |
3046 | if (!CI || CI->use_empty()) |
3047 | return; |
3048 | |
3049 | // Don't consider calls in other functions. |
3050 | if (CI->getFunction() != F) |
3051 | return; |
3052 | |
3053 | Module *M = CI->getModule(); |
3054 | Function *Callee = CI->getCalledFunction(); |
3055 | LibFunc Func; |
3056 | if (!Callee || !TLI->getLibFunc(FDecl: *Callee, F&: Func) || |
3057 | !isLibFuncEmittable(M, TLI, TheLibFunc: Func) || |
3058 | !isTrigLibCall(CI)) |
3059 | return; |
3060 | |
3061 | if (IsFloat) { |
3062 | if (Func == LibFunc_sinpif) |
3063 | SinCalls.push_back(Elt: CI); |
3064 | else if (Func == LibFunc_cospif) |
3065 | CosCalls.push_back(Elt: CI); |
3066 | else if (Func == LibFunc_sincospif_stret) |
3067 | SinCosCalls.push_back(Elt: CI); |
3068 | } else { |
3069 | if (Func == LibFunc_sinpi) |
3070 | SinCalls.push_back(Elt: CI); |
3071 | else if (Func == LibFunc_cospi) |
3072 | CosCalls.push_back(Elt: CI); |
3073 | else if (Func == LibFunc_sincospi_stret) |
3074 | SinCosCalls.push_back(Elt: CI); |
3075 | } |
3076 | } |
3077 | |
3078 | /// Constant folds remquo |
3079 | Value *LibCallSimplifier::optimizeRemquo(CallInst *CI, IRBuilderBase &B) { |
3080 | const APFloat *X, *Y; |
3081 | if (!match(V: CI->getArgOperand(i: 0), P: m_APFloat(Res&: X)) || |
3082 | !match(V: CI->getArgOperand(i: 1), P: m_APFloat(Res&: Y))) |
3083 | return nullptr; |
3084 | |
3085 | APFloat::opStatus Status; |
3086 | APFloat Quot = *X; |
3087 | Status = Quot.divide(RHS: *Y, RM: APFloat::rmNearestTiesToEven); |
3088 | if (Status != APFloat::opOK && Status != APFloat::opInexact) |
3089 | return nullptr; |
3090 | APFloat Rem = *X; |
3091 | if (Rem.remainder(RHS: *Y) != APFloat::opOK) |
3092 | return nullptr; |
3093 | |
3094 | // TODO: We can only keep at least the three of the last bits of x/y |
3095 | unsigned IntBW = TLI->getIntSize(); |
3096 | APSInt QuotInt(IntBW, /*isUnsigned=*/false); |
3097 | bool IsExact; |
3098 | Status = |
3099 | Quot.convertToInteger(Result&: QuotInt, RM: APFloat::rmNearestTiesToEven, IsExact: &IsExact); |
3100 | if (Status != APFloat::opOK && Status != APFloat::opInexact) |
3101 | return nullptr; |
3102 | |
3103 | B.CreateAlignedStore( |
3104 | Val: ConstantInt::get(Ty: B.getIntNTy(N: IntBW), V: QuotInt.getExtValue()), |
3105 | Ptr: CI->getArgOperand(i: 2), Align: CI->getParamAlign(ArgNo: 2)); |
3106 | return ConstantFP::get(Ty: CI->getType(), V: Rem); |
3107 | } |
3108 | |
3109 | /// Constant folds fdim |
3110 | Value *LibCallSimplifier::optimizeFdim(CallInst *CI, IRBuilderBase &B) { |
3111 | // Cannot perform the fold unless the call has attribute memory(none) |
3112 | if (!CI->doesNotAccessMemory()) |
3113 | return nullptr; |
3114 | |
3115 | // TODO : Handle undef values |
3116 | // Propagate poison if any |
3117 | if (isa<PoisonValue>(Val: CI->getArgOperand(i: 0))) |
3118 | return CI->getArgOperand(i: 0); |
3119 | if (isa<PoisonValue>(Val: CI->getArgOperand(i: 1))) |
3120 | return CI->getArgOperand(i: 1); |
3121 | |
3122 | const APFloat *X, *Y; |
3123 | // Check if both values are constants |
3124 | if (!match(V: CI->getArgOperand(i: 0), P: m_APFloat(Res&: X)) || |
3125 | !match(V: CI->getArgOperand(i: 1), P: m_APFloat(Res&: Y))) |
3126 | return nullptr; |
3127 | |
3128 | APFloat Difference = *X; |
3129 | Difference.subtract(RHS: *Y, RM: RoundingMode::NearestTiesToEven); |
3130 | |
3131 | APFloat MaxVal = |
3132 | maximum(A: Difference, B: APFloat::getZero(Sem: CI->getType()->getFltSemantics())); |
3133 | return ConstantFP::get(Ty: CI->getType(), V: MaxVal); |
3134 | } |
3135 | |
3136 | //===----------------------------------------------------------------------===// |
3137 | // Integer Library Call Optimizations |
3138 | //===----------------------------------------------------------------------===// |
3139 | |
3140 | Value *LibCallSimplifier::optimizeFFS(CallInst *CI, IRBuilderBase &B) { |
3141 | // All variants of ffs return int which need not be 32 bits wide. |
3142 | // ffs{,l,ll}(x) -> x != 0 ? (int)llvm.cttz(x)+1 : 0 |
3143 | Type *RetType = CI->getType(); |
3144 | Value *Op = CI->getArgOperand(i: 0); |
3145 | Type *ArgType = Op->getType(); |
3146 | Value *V = B.CreateIntrinsic(ID: Intrinsic::cttz, Types: {ArgType}, Args: {Op, B.getTrue()}, |
3147 | FMFSource: nullptr, Name: "cttz" ); |
3148 | V = B.CreateAdd(LHS: V, RHS: ConstantInt::get(Ty: V->getType(), V: 1)); |
3149 | V = B.CreateIntCast(V, DestTy: RetType, isSigned: false); |
3150 | |
3151 | Value *Cond = B.CreateICmpNE(LHS: Op, RHS: Constant::getNullValue(Ty: ArgType)); |
3152 | return B.CreateSelect(C: Cond, True: V, False: ConstantInt::get(Ty: RetType, V: 0)); |
3153 | } |
3154 | |
3155 | Value *LibCallSimplifier::optimizeFls(CallInst *CI, IRBuilderBase &B) { |
3156 | // All variants of fls return int which need not be 32 bits wide. |
3157 | // fls{,l,ll}(x) -> (int)(sizeInBits(x) - llvm.ctlz(x, false)) |
3158 | Value *Op = CI->getArgOperand(i: 0); |
3159 | Type *ArgType = Op->getType(); |
3160 | Value *V = B.CreateIntrinsic(ID: Intrinsic::ctlz, Types: {ArgType}, Args: {Op, B.getFalse()}, |
3161 | FMFSource: nullptr, Name: "ctlz" ); |
3162 | V = B.CreateSub(LHS: ConstantInt::get(Ty: V->getType(), V: ArgType->getIntegerBitWidth()), |
3163 | RHS: V); |
3164 | return B.CreateIntCast(V, DestTy: CI->getType(), isSigned: false); |
3165 | } |
3166 | |
3167 | Value *LibCallSimplifier::optimizeAbs(CallInst *CI, IRBuilderBase &B) { |
3168 | // abs(x) -> x <s 0 ? -x : x |
3169 | // The negation has 'nsw' because abs of INT_MIN is undefined. |
3170 | Value *X = CI->getArgOperand(i: 0); |
3171 | Value *IsNeg = B.CreateIsNeg(Arg: X); |
3172 | Value *NegX = B.CreateNSWNeg(V: X, Name: "neg" ); |
3173 | return B.CreateSelect(C: IsNeg, True: NegX, False: X); |
3174 | } |
3175 | |
3176 | Value *LibCallSimplifier::optimizeIsDigit(CallInst *CI, IRBuilderBase &B) { |
3177 | // isdigit(c) -> (c-'0') <u 10 |
3178 | Value *Op = CI->getArgOperand(i: 0); |
3179 | Type *ArgType = Op->getType(); |
3180 | Op = B.CreateSub(LHS: Op, RHS: ConstantInt::get(Ty: ArgType, V: '0'), Name: "isdigittmp" ); |
3181 | Op = B.CreateICmpULT(LHS: Op, RHS: ConstantInt::get(Ty: ArgType, V: 10), Name: "isdigit" ); |
3182 | return B.CreateZExt(V: Op, DestTy: CI->getType()); |
3183 | } |
3184 | |
3185 | Value *LibCallSimplifier::optimizeIsAscii(CallInst *CI, IRBuilderBase &B) { |
3186 | // isascii(c) -> c <u 128 |
3187 | Value *Op = CI->getArgOperand(i: 0); |
3188 | Type *ArgType = Op->getType(); |
3189 | Op = B.CreateICmpULT(LHS: Op, RHS: ConstantInt::get(Ty: ArgType, V: 128), Name: "isascii" ); |
3190 | return B.CreateZExt(V: Op, DestTy: CI->getType()); |
3191 | } |
3192 | |
3193 | Value *LibCallSimplifier::optimizeToAscii(CallInst *CI, IRBuilderBase &B) { |
3194 | // toascii(c) -> c & 0x7f |
3195 | return B.CreateAnd(LHS: CI->getArgOperand(i: 0), |
3196 | RHS: ConstantInt::get(Ty: CI->getType(), V: 0x7F)); |
3197 | } |
3198 | |
3199 | // Fold calls to atoi, atol, and atoll. |
3200 | Value *LibCallSimplifier::optimizeAtoi(CallInst *CI, IRBuilderBase &B) { |
3201 | StringRef Str; |
3202 | if (!getConstantStringInfo(V: CI->getArgOperand(i: 0), Str)) |
3203 | return nullptr; |
3204 | |
3205 | return convertStrToInt(CI, Str, EndPtr: nullptr, Base: 10, /*AsSigned=*/true, B); |
3206 | } |
3207 | |
3208 | // Fold calls to strtol, strtoll, strtoul, and strtoull. |
3209 | Value *LibCallSimplifier::optimizeStrToInt(CallInst *CI, IRBuilderBase &B, |
3210 | bool AsSigned) { |
3211 | Value *EndPtr = CI->getArgOperand(i: 1); |
3212 | if (isa<ConstantPointerNull>(Val: EndPtr)) { |
3213 | // With a null EndPtr, this function won't capture the main argument. |
3214 | // It would be readonly too, except that it still may write to errno. |
3215 | CI->addParamAttr(ArgNo: 0, Attr: Attribute::getWithCaptureInfo(Context&: CI->getContext(), |
3216 | CI: CaptureInfo::none())); |
3217 | EndPtr = nullptr; |
3218 | } else if (!isKnownNonZero(V: EndPtr, Q: DL)) |
3219 | return nullptr; |
3220 | |
3221 | StringRef Str; |
3222 | if (!getConstantStringInfo(V: CI->getArgOperand(i: 0), Str)) |
3223 | return nullptr; |
3224 | |
3225 | if (ConstantInt *CInt = dyn_cast<ConstantInt>(Val: CI->getArgOperand(i: 2))) { |
3226 | return convertStrToInt(CI, Str, EndPtr, Base: CInt->getSExtValue(), AsSigned, B); |
3227 | } |
3228 | |
3229 | return nullptr; |
3230 | } |
3231 | |
3232 | //===----------------------------------------------------------------------===// |
3233 | // Formatting and IO Library Call Optimizations |
3234 | //===----------------------------------------------------------------------===// |
3235 | |
3236 | static bool isReportingError(Function *Callee, CallInst *CI, int StreamArg); |
3237 | |
3238 | Value *LibCallSimplifier::optimizeErrorReporting(CallInst *CI, IRBuilderBase &B, |
3239 | int StreamArg) { |
3240 | Function *Callee = CI->getCalledFunction(); |
3241 | // Error reporting calls should be cold, mark them as such. |
3242 | // This applies even to non-builtin calls: it is only a hint and applies to |
3243 | // functions that the frontend might not understand as builtins. |
3244 | |
3245 | // This heuristic was suggested in: |
3246 | // Improving Static Branch Prediction in a Compiler |
3247 | // Brian L. Deitrich, Ben-Chung Cheng, Wen-mei W. Hwu |
3248 | // Proceedings of PACT'98, Oct. 1998, IEEE |
3249 | if (!CI->hasFnAttr(Kind: Attribute::Cold) && |
3250 | isReportingError(Callee, CI, StreamArg)) { |
3251 | CI->addFnAttr(Kind: Attribute::Cold); |
3252 | } |
3253 | |
3254 | return nullptr; |
3255 | } |
3256 | |
3257 | static bool isReportingError(Function *Callee, CallInst *CI, int StreamArg) { |
3258 | if (!Callee || !Callee->isDeclaration()) |
3259 | return false; |
3260 | |
3261 | if (StreamArg < 0) |
3262 | return true; |
3263 | |
3264 | // These functions might be considered cold, but only if their stream |
3265 | // argument is stderr. |
3266 | |
3267 | if (StreamArg >= (int)CI->arg_size()) |
3268 | return false; |
3269 | LoadInst *LI = dyn_cast<LoadInst>(Val: CI->getArgOperand(i: StreamArg)); |
3270 | if (!LI) |
3271 | return false; |
3272 | GlobalVariable *GV = dyn_cast<GlobalVariable>(Val: LI->getPointerOperand()); |
3273 | if (!GV || !GV->isDeclaration()) |
3274 | return false; |
3275 | return GV->getName() == "stderr" ; |
3276 | } |
3277 | |
3278 | Value *LibCallSimplifier::optimizePrintFString(CallInst *CI, IRBuilderBase &B) { |
3279 | // Check for a fixed format string. |
3280 | StringRef FormatStr; |
3281 | if (!getConstantStringInfo(V: CI->getArgOperand(i: 0), Str&: FormatStr)) |
3282 | return nullptr; |
3283 | |
3284 | // Empty format string -> noop. |
3285 | if (FormatStr.empty()) // Tolerate printf's declared void. |
3286 | return CI->use_empty() ? (Value *)CI : ConstantInt::get(Ty: CI->getType(), V: 0); |
3287 | |
3288 | // Do not do any of the following transformations if the printf return value |
3289 | // is used, in general the printf return value is not compatible with either |
3290 | // putchar() or puts(). |
3291 | if (!CI->use_empty()) |
3292 | return nullptr; |
3293 | |
3294 | Type *IntTy = CI->getType(); |
3295 | // printf("x") -> putchar('x'), even for "%" and "%%". |
3296 | if (FormatStr.size() == 1 || FormatStr == "%%" ) { |
3297 | // Convert the character to unsigned char before passing it to putchar |
3298 | // to avoid host-specific sign extension in the IR. Putchar converts |
3299 | // it to unsigned char regardless. |
3300 | Value *IntChar = ConstantInt::get(Ty: IntTy, V: (unsigned char)FormatStr[0]); |
3301 | return copyFlags(Old: *CI, New: emitPutChar(Char: IntChar, B, TLI)); |
3302 | } |
3303 | |
3304 | // Try to remove call or emit putchar/puts. |
3305 | if (FormatStr == "%s" && CI->arg_size() > 1) { |
3306 | StringRef OperandStr; |
3307 | if (!getConstantStringInfo(V: CI->getOperand(i_nocapture: 1), Str&: OperandStr)) |
3308 | return nullptr; |
3309 | // printf("%s", "") --> NOP |
3310 | if (OperandStr.empty()) |
3311 | return (Value *)CI; |
3312 | // printf("%s", "a") --> putchar('a') |
3313 | if (OperandStr.size() == 1) { |
3314 | // Convert the character to unsigned char before passing it to putchar |
3315 | // to avoid host-specific sign extension in the IR. Putchar converts |
3316 | // it to unsigned char regardless. |
3317 | Value *IntChar = ConstantInt::get(Ty: IntTy, V: (unsigned char)OperandStr[0]); |
3318 | return copyFlags(Old: *CI, New: emitPutChar(Char: IntChar, B, TLI)); |
3319 | } |
3320 | // printf("%s", str"\n") --> puts(str) |
3321 | if (OperandStr.back() == '\n') { |
3322 | OperandStr = OperandStr.drop_back(); |
3323 | Value *GV = B.CreateGlobalString(Str: OperandStr, Name: "str" ); |
3324 | return copyFlags(Old: *CI, New: emitPutS(Str: GV, B, TLI)); |
3325 | } |
3326 | return nullptr; |
3327 | } |
3328 | |
3329 | // printf("foo\n") --> puts("foo") |
3330 | if (FormatStr.back() == '\n' && |
3331 | !FormatStr.contains(C: '%')) { // No format characters. |
3332 | // Create a string literal with no \n on it. We expect the constant merge |
3333 | // pass to be run after this pass, to merge duplicate strings. |
3334 | FormatStr = FormatStr.drop_back(); |
3335 | Value *GV = B.CreateGlobalString(Str: FormatStr, Name: "str" ); |
3336 | return copyFlags(Old: *CI, New: emitPutS(Str: GV, B, TLI)); |
3337 | } |
3338 | |
3339 | // Optimize specific format strings. |
3340 | // printf("%c", chr) --> putchar(chr) |
3341 | if (FormatStr == "%c" && CI->arg_size() > 1 && |
3342 | CI->getArgOperand(i: 1)->getType()->isIntegerTy()) { |
3343 | // Convert the argument to the type expected by putchar, i.e., int, which |
3344 | // need not be 32 bits wide but which is the same as printf's return type. |
3345 | Value *IntChar = B.CreateIntCast(V: CI->getArgOperand(i: 1), DestTy: IntTy, isSigned: false); |
3346 | return copyFlags(Old: *CI, New: emitPutChar(Char: IntChar, B, TLI)); |
3347 | } |
3348 | |
3349 | // printf("%s\n", str) --> puts(str) |
3350 | if (FormatStr == "%s\n" && CI->arg_size() > 1 && |
3351 | CI->getArgOperand(i: 1)->getType()->isPointerTy()) |
3352 | return copyFlags(Old: *CI, New: emitPutS(Str: CI->getArgOperand(i: 1), B, TLI)); |
3353 | return nullptr; |
3354 | } |
3355 | |
3356 | Value *LibCallSimplifier::optimizePrintF(CallInst *CI, IRBuilderBase &B) { |
3357 | |
3358 | Module *M = CI->getModule(); |
3359 | Function *Callee = CI->getCalledFunction(); |
3360 | FunctionType *FT = Callee->getFunctionType(); |
3361 | if (Value *V = optimizePrintFString(CI, B)) { |
3362 | return V; |
3363 | } |
3364 | |
3365 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: 0); |
3366 | |
3367 | // printf(format, ...) -> iprintf(format, ...) if no floating point |
3368 | // arguments. |
3369 | if (isLibFuncEmittable(M, TLI, TheLibFunc: LibFunc_iprintf) && |
3370 | !callHasFloatingPointArgument(CI)) { |
3371 | FunctionCallee IPrintFFn = getOrInsertLibFunc(M, TLI: *TLI, TheLibFunc: LibFunc_iprintf, T: FT, |
3372 | AttributeList: Callee->getAttributes()); |
3373 | CallInst *New = cast<CallInst>(Val: CI->clone()); |
3374 | New->setCalledFunction(IPrintFFn); |
3375 | B.Insert(I: New); |
3376 | return New; |
3377 | } |
3378 | |
3379 | // printf(format, ...) -> __small_printf(format, ...) if no 128-bit floating point |
3380 | // arguments. |
3381 | if (isLibFuncEmittable(M, TLI, TheLibFunc: LibFunc_small_printf) && |
3382 | !callHasFP128Argument(CI)) { |
3383 | auto SmallPrintFFn = getOrInsertLibFunc(M, TLI: *TLI, TheLibFunc: LibFunc_small_printf, T: FT, |
3384 | AttributeList: Callee->getAttributes()); |
3385 | CallInst *New = cast<CallInst>(Val: CI->clone()); |
3386 | New->setCalledFunction(SmallPrintFFn); |
3387 | B.Insert(I: New); |
3388 | return New; |
3389 | } |
3390 | |
3391 | return nullptr; |
3392 | } |
3393 | |
3394 | Value *LibCallSimplifier::optimizeSPrintFString(CallInst *CI, |
3395 | IRBuilderBase &B) { |
3396 | // Check for a fixed format string. |
3397 | StringRef FormatStr; |
3398 | if (!getConstantStringInfo(V: CI->getArgOperand(i: 1), Str&: FormatStr)) |
3399 | return nullptr; |
3400 | |
3401 | // If we just have a format string (nothing else crazy) transform it. |
3402 | Value *Dest = CI->getArgOperand(i: 0); |
3403 | if (CI->arg_size() == 2) { |
3404 | // Make sure there's no % in the constant array. We could try to handle |
3405 | // %% -> % in the future if we cared. |
3406 | if (FormatStr.contains(C: '%')) |
3407 | return nullptr; // we found a format specifier, bail out. |
3408 | |
3409 | // sprintf(str, fmt) -> llvm.memcpy(align 1 str, align 1 fmt, strlen(fmt)+1) |
3410 | B.CreateMemCpy(Dst: Dest, DstAlign: Align(1), Src: CI->getArgOperand(i: 1), SrcAlign: Align(1), |
3411 | // Copy the null byte. |
3412 | Size: TLI->getAsSizeT(V: FormatStr.size() + 1, M: *CI->getModule())); |
3413 | return ConstantInt::get(Ty: CI->getType(), V: FormatStr.size()); |
3414 | } |
3415 | |
3416 | // The remaining optimizations require the format string to be "%s" or "%c" |
3417 | // and have an extra operand. |
3418 | if (FormatStr.size() != 2 || FormatStr[0] != '%' || CI->arg_size() < 3) |
3419 | return nullptr; |
3420 | |
3421 | // Decode the second character of the format string. |
3422 | if (FormatStr[1] == 'c') { |
3423 | // sprintf(dst, "%c", chr) --> *(i8*)dst = chr; *((i8*)dst+1) = 0 |
3424 | if (!CI->getArgOperand(i: 2)->getType()->isIntegerTy()) |
3425 | return nullptr; |
3426 | Value *V = B.CreateTrunc(V: CI->getArgOperand(i: 2), DestTy: B.getInt8Ty(), Name: "char" ); |
3427 | Value *Ptr = Dest; |
3428 | B.CreateStore(Val: V, Ptr); |
3429 | Ptr = B.CreateInBoundsGEP(Ty: B.getInt8Ty(), Ptr, IdxList: B.getInt32(C: 1), Name: "nul" ); |
3430 | B.CreateStore(Val: B.getInt8(C: 0), Ptr); |
3431 | |
3432 | return ConstantInt::get(Ty: CI->getType(), V: 1); |
3433 | } |
3434 | |
3435 | if (FormatStr[1] == 's') { |
3436 | // sprintf(dest, "%s", str) -> llvm.memcpy(align 1 dest, align 1 str, |
3437 | // strlen(str)+1) |
3438 | if (!CI->getArgOperand(i: 2)->getType()->isPointerTy()) |
3439 | return nullptr; |
3440 | |
3441 | if (CI->use_empty()) |
3442 | // sprintf(dest, "%s", str) -> strcpy(dest, str) |
3443 | return copyFlags(Old: *CI, New: emitStrCpy(Dst: Dest, Src: CI->getArgOperand(i: 2), B, TLI)); |
3444 | |
3445 | uint64_t SrcLen = GetStringLength(V: CI->getArgOperand(i: 2)); |
3446 | if (SrcLen) { |
3447 | B.CreateMemCpy(Dst: Dest, DstAlign: Align(1), Src: CI->getArgOperand(i: 2), SrcAlign: Align(1), |
3448 | Size: TLI->getAsSizeT(V: SrcLen, M: *CI->getModule())); |
3449 | // Returns total number of characters written without null-character. |
3450 | return ConstantInt::get(Ty: CI->getType(), V: SrcLen - 1); |
3451 | } else if (Value *V = emitStpCpy(Dst: Dest, Src: CI->getArgOperand(i: 2), B, TLI)) { |
3452 | // sprintf(dest, "%s", str) -> stpcpy(dest, str) - dest |
3453 | Value *PtrDiff = B.CreatePtrDiff(ElemTy: B.getInt8Ty(), LHS: V, RHS: Dest); |
3454 | return B.CreateIntCast(V: PtrDiff, DestTy: CI->getType(), isSigned: false); |
3455 | } |
3456 | |
3457 | if (llvm::shouldOptimizeForSize(BB: CI->getParent(), PSI, BFI, |
3458 | QueryType: PGSOQueryType::IRPass)) |
3459 | return nullptr; |
3460 | |
3461 | Value *Len = emitStrLen(Ptr: CI->getArgOperand(i: 2), B, DL, TLI); |
3462 | if (!Len) |
3463 | return nullptr; |
3464 | Value *IncLen = |
3465 | B.CreateAdd(LHS: Len, RHS: ConstantInt::get(Ty: Len->getType(), V: 1), Name: "leninc" ); |
3466 | B.CreateMemCpy(Dst: Dest, DstAlign: Align(1), Src: CI->getArgOperand(i: 2), SrcAlign: Align(1), Size: IncLen); |
3467 | |
3468 | // The sprintf result is the unincremented number of bytes in the string. |
3469 | return B.CreateIntCast(V: Len, DestTy: CI->getType(), isSigned: false); |
3470 | } |
3471 | return nullptr; |
3472 | } |
3473 | |
3474 | Value *LibCallSimplifier::optimizeSPrintF(CallInst *CI, IRBuilderBase &B) { |
3475 | Module *M = CI->getModule(); |
3476 | Function *Callee = CI->getCalledFunction(); |
3477 | FunctionType *FT = Callee->getFunctionType(); |
3478 | if (Value *V = optimizeSPrintFString(CI, B)) { |
3479 | return V; |
3480 | } |
3481 | |
3482 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: {0, 1}); |
3483 | |
3484 | // sprintf(str, format, ...) -> siprintf(str, format, ...) if no floating |
3485 | // point arguments. |
3486 | if (isLibFuncEmittable(M, TLI, TheLibFunc: LibFunc_siprintf) && |
3487 | !callHasFloatingPointArgument(CI)) { |
3488 | FunctionCallee SIPrintFFn = getOrInsertLibFunc(M, TLI: *TLI, TheLibFunc: LibFunc_siprintf, |
3489 | T: FT, AttributeList: Callee->getAttributes()); |
3490 | CallInst *New = cast<CallInst>(Val: CI->clone()); |
3491 | New->setCalledFunction(SIPrintFFn); |
3492 | B.Insert(I: New); |
3493 | return New; |
3494 | } |
3495 | |
3496 | // sprintf(str, format, ...) -> __small_sprintf(str, format, ...) if no 128-bit |
3497 | // floating point arguments. |
3498 | if (isLibFuncEmittable(M, TLI, TheLibFunc: LibFunc_small_sprintf) && |
3499 | !callHasFP128Argument(CI)) { |
3500 | auto SmallSPrintFFn = getOrInsertLibFunc(M, TLI: *TLI, TheLibFunc: LibFunc_small_sprintf, T: FT, |
3501 | AttributeList: Callee->getAttributes()); |
3502 | CallInst *New = cast<CallInst>(Val: CI->clone()); |
3503 | New->setCalledFunction(SmallSPrintFFn); |
3504 | B.Insert(I: New); |
3505 | return New; |
3506 | } |
3507 | |
3508 | return nullptr; |
3509 | } |
3510 | |
3511 | // Transform an snprintf call CI with the bound N to format the string Str |
3512 | // either to a call to memcpy, or to single character a store, or to nothing, |
3513 | // and fold the result to a constant. A nonnull StrArg refers to the string |
3514 | // argument being formatted. Otherwise the call is one with N < 2 and |
3515 | // the "%c" directive to format a single character. |
3516 | Value *LibCallSimplifier::emitSnPrintfMemCpy(CallInst *CI, Value *StrArg, |
3517 | StringRef Str, uint64_t N, |
3518 | IRBuilderBase &B) { |
3519 | assert(StrArg || (N < 2 && Str.size() == 1)); |
3520 | |
3521 | unsigned IntBits = TLI->getIntSize(); |
3522 | uint64_t IntMax = maxIntN(N: IntBits); |
3523 | if (Str.size() > IntMax) |
3524 | // Bail if the string is longer than INT_MAX. POSIX requires |
3525 | // implementations to set errno to EOVERFLOW in this case, in |
3526 | // addition to when N is larger than that (checked by the caller). |
3527 | return nullptr; |
3528 | |
3529 | Value *StrLen = ConstantInt::get(Ty: CI->getType(), V: Str.size()); |
3530 | if (N == 0) |
3531 | return StrLen; |
3532 | |
3533 | // Set to the number of bytes to copy fron StrArg which is also |
3534 | // the offset of the terinating nul. |
3535 | uint64_t NCopy; |
3536 | if (N > Str.size()) |
3537 | // Copy the full string, including the terminating nul (which must |
3538 | // be present regardless of the bound). |
3539 | NCopy = Str.size() + 1; |
3540 | else |
3541 | NCopy = N - 1; |
3542 | |
3543 | Value *DstArg = CI->getArgOperand(i: 0); |
3544 | if (NCopy && StrArg) |
3545 | // Transform the call to lvm.memcpy(dst, fmt, N). |
3546 | copyFlags(Old: *CI, New: B.CreateMemCpy(Dst: DstArg, DstAlign: Align(1), Src: StrArg, SrcAlign: Align(1), |
3547 | Size: TLI->getAsSizeT(V: NCopy, M: *CI->getModule()))); |
3548 | |
3549 | if (N > Str.size()) |
3550 | // Return early when the whole format string, including the final nul, |
3551 | // has been copied. |
3552 | return StrLen; |
3553 | |
3554 | // Otherwise, when truncating the string append a terminating nul. |
3555 | Type *Int8Ty = B.getInt8Ty(); |
3556 | Value *NulOff = B.getIntN(N: IntBits, C: NCopy); |
3557 | Value *DstEnd = B.CreateInBoundsGEP(Ty: Int8Ty, Ptr: DstArg, IdxList: NulOff, Name: "endptr" ); |
3558 | B.CreateStore(Val: ConstantInt::get(Ty: Int8Ty, V: 0), Ptr: DstEnd); |
3559 | return StrLen; |
3560 | } |
3561 | |
3562 | Value *LibCallSimplifier::optimizeSnPrintFString(CallInst *CI, |
3563 | IRBuilderBase &B) { |
3564 | // Check for size |
3565 | ConstantInt *Size = dyn_cast<ConstantInt>(Val: CI->getArgOperand(i: 1)); |
3566 | if (!Size) |
3567 | return nullptr; |
3568 | |
3569 | uint64_t N = Size->getZExtValue(); |
3570 | uint64_t IntMax = maxIntN(N: TLI->getIntSize()); |
3571 | if (N > IntMax) |
3572 | // Bail if the bound exceeds INT_MAX. POSIX requires implementations |
3573 | // to set errno to EOVERFLOW in this case. |
3574 | return nullptr; |
3575 | |
3576 | Value *DstArg = CI->getArgOperand(i: 0); |
3577 | Value *FmtArg = CI->getArgOperand(i: 2); |
3578 | |
3579 | // Check for a fixed format string. |
3580 | StringRef FormatStr; |
3581 | if (!getConstantStringInfo(V: FmtArg, Str&: FormatStr)) |
3582 | return nullptr; |
3583 | |
3584 | // If we just have a format string (nothing else crazy) transform it. |
3585 | if (CI->arg_size() == 3) { |
3586 | if (FormatStr.contains(C: '%')) |
3587 | // Bail if the format string contains a directive and there are |
3588 | // no arguments. We could handle "%%" in the future. |
3589 | return nullptr; |
3590 | |
3591 | return emitSnPrintfMemCpy(CI, StrArg: FmtArg, Str: FormatStr, N, B); |
3592 | } |
3593 | |
3594 | // The remaining optimizations require the format string to be "%s" or "%c" |
3595 | // and have an extra operand. |
3596 | if (FormatStr.size() != 2 || FormatStr[0] != '%' || CI->arg_size() != 4) |
3597 | return nullptr; |
3598 | |
3599 | // Decode the second character of the format string. |
3600 | if (FormatStr[1] == 'c') { |
3601 | if (N <= 1) { |
3602 | // Use an arbitary string of length 1 to transform the call into |
3603 | // either a nul store (N == 1) or a no-op (N == 0) and fold it |
3604 | // to one. |
3605 | StringRef CharStr("*" ); |
3606 | return emitSnPrintfMemCpy(CI, StrArg: nullptr, Str: CharStr, N, B); |
3607 | } |
3608 | |
3609 | // snprintf(dst, size, "%c", chr) --> *(i8*)dst = chr; *((i8*)dst+1) = 0 |
3610 | if (!CI->getArgOperand(i: 3)->getType()->isIntegerTy()) |
3611 | return nullptr; |
3612 | Value *V = B.CreateTrunc(V: CI->getArgOperand(i: 3), DestTy: B.getInt8Ty(), Name: "char" ); |
3613 | Value *Ptr = DstArg; |
3614 | B.CreateStore(Val: V, Ptr); |
3615 | Ptr = B.CreateInBoundsGEP(Ty: B.getInt8Ty(), Ptr, IdxList: B.getInt32(C: 1), Name: "nul" ); |
3616 | B.CreateStore(Val: B.getInt8(C: 0), Ptr); |
3617 | return ConstantInt::get(Ty: CI->getType(), V: 1); |
3618 | } |
3619 | |
3620 | if (FormatStr[1] != 's') |
3621 | return nullptr; |
3622 | |
3623 | Value *StrArg = CI->getArgOperand(i: 3); |
3624 | // snprintf(dest, size, "%s", str) to llvm.memcpy(dest, str, len+1, 1) |
3625 | StringRef Str; |
3626 | if (!getConstantStringInfo(V: StrArg, Str)) |
3627 | return nullptr; |
3628 | |
3629 | return emitSnPrintfMemCpy(CI, StrArg, Str, N, B); |
3630 | } |
3631 | |
3632 | Value *LibCallSimplifier::optimizeSnPrintF(CallInst *CI, IRBuilderBase &B) { |
3633 | if (Value *V = optimizeSnPrintFString(CI, B)) { |
3634 | return V; |
3635 | } |
3636 | |
3637 | if (isKnownNonZero(V: CI->getOperand(i_nocapture: 1), Q: DL)) |
3638 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: 0); |
3639 | return nullptr; |
3640 | } |
3641 | |
3642 | Value *LibCallSimplifier::optimizeFPrintFString(CallInst *CI, |
3643 | IRBuilderBase &B) { |
3644 | optimizeErrorReporting(CI, B, StreamArg: 0); |
3645 | |
3646 | // All the optimizations depend on the format string. |
3647 | StringRef FormatStr; |
3648 | if (!getConstantStringInfo(V: CI->getArgOperand(i: 1), Str&: FormatStr)) |
3649 | return nullptr; |
3650 | |
3651 | // Do not do any of the following transformations if the fprintf return |
3652 | // value is used, in general the fprintf return value is not compatible |
3653 | // with fwrite(), fputc() or fputs(). |
3654 | if (!CI->use_empty()) |
3655 | return nullptr; |
3656 | |
3657 | // fprintf(F, "foo") --> fwrite("foo", 3, 1, F) |
3658 | if (CI->arg_size() == 2) { |
3659 | // Could handle %% -> % if we cared. |
3660 | if (FormatStr.contains(C: '%')) |
3661 | return nullptr; // We found a format specifier. |
3662 | |
3663 | return copyFlags( |
3664 | Old: *CI, New: emitFWrite(Ptr: CI->getArgOperand(i: 1), |
3665 | Size: TLI->getAsSizeT(V: FormatStr.size(), M: *CI->getModule()), |
3666 | File: CI->getArgOperand(i: 0), B, DL, TLI)); |
3667 | } |
3668 | |
3669 | // The remaining optimizations require the format string to be "%s" or "%c" |
3670 | // and have an extra operand. |
3671 | if (FormatStr.size() != 2 || FormatStr[0] != '%' || CI->arg_size() < 3) |
3672 | return nullptr; |
3673 | |
3674 | // Decode the second character of the format string. |
3675 | if (FormatStr[1] == 'c') { |
3676 | // fprintf(F, "%c", chr) --> fputc((int)chr, F) |
3677 | if (!CI->getArgOperand(i: 2)->getType()->isIntegerTy()) |
3678 | return nullptr; |
3679 | Type *IntTy = B.getIntNTy(N: TLI->getIntSize()); |
3680 | Value *V = B.CreateIntCast(V: CI->getArgOperand(i: 2), DestTy: IntTy, /*isSigned*/ true, |
3681 | Name: "chari" ); |
3682 | return copyFlags(Old: *CI, New: emitFPutC(Char: V, File: CI->getArgOperand(i: 0), B, TLI)); |
3683 | } |
3684 | |
3685 | if (FormatStr[1] == 's') { |
3686 | // fprintf(F, "%s", str) --> fputs(str, F) |
3687 | if (!CI->getArgOperand(i: 2)->getType()->isPointerTy()) |
3688 | return nullptr; |
3689 | return copyFlags( |
3690 | Old: *CI, New: emitFPutS(Str: CI->getArgOperand(i: 2), File: CI->getArgOperand(i: 0), B, TLI)); |
3691 | } |
3692 | return nullptr; |
3693 | } |
3694 | |
3695 | Value *LibCallSimplifier::optimizeFPrintF(CallInst *CI, IRBuilderBase &B) { |
3696 | Module *M = CI->getModule(); |
3697 | Function *Callee = CI->getCalledFunction(); |
3698 | FunctionType *FT = Callee->getFunctionType(); |
3699 | if (Value *V = optimizeFPrintFString(CI, B)) { |
3700 | return V; |
3701 | } |
3702 | |
3703 | // fprintf(stream, format, ...) -> fiprintf(stream, format, ...) if no |
3704 | // floating point arguments. |
3705 | if (isLibFuncEmittable(M, TLI, TheLibFunc: LibFunc_fiprintf) && |
3706 | !callHasFloatingPointArgument(CI)) { |
3707 | FunctionCallee FIPrintFFn = getOrInsertLibFunc(M, TLI: *TLI, TheLibFunc: LibFunc_fiprintf, |
3708 | T: FT, AttributeList: Callee->getAttributes()); |
3709 | CallInst *New = cast<CallInst>(Val: CI->clone()); |
3710 | New->setCalledFunction(FIPrintFFn); |
3711 | B.Insert(I: New); |
3712 | return New; |
3713 | } |
3714 | |
3715 | // fprintf(stream, format, ...) -> __small_fprintf(stream, format, ...) if no |
3716 | // 128-bit floating point arguments. |
3717 | if (isLibFuncEmittable(M, TLI, TheLibFunc: LibFunc_small_fprintf) && |
3718 | !callHasFP128Argument(CI)) { |
3719 | auto SmallFPrintFFn = |
3720 | getOrInsertLibFunc(M, TLI: *TLI, TheLibFunc: LibFunc_small_fprintf, T: FT, |
3721 | AttributeList: Callee->getAttributes()); |
3722 | CallInst *New = cast<CallInst>(Val: CI->clone()); |
3723 | New->setCalledFunction(SmallFPrintFFn); |
3724 | B.Insert(I: New); |
3725 | return New; |
3726 | } |
3727 | |
3728 | return nullptr; |
3729 | } |
3730 | |
3731 | Value *LibCallSimplifier::optimizeFWrite(CallInst *CI, IRBuilderBase &B) { |
3732 | optimizeErrorReporting(CI, B, StreamArg: 3); |
3733 | |
3734 | // Get the element size and count. |
3735 | ConstantInt *SizeC = dyn_cast<ConstantInt>(Val: CI->getArgOperand(i: 1)); |
3736 | ConstantInt *CountC = dyn_cast<ConstantInt>(Val: CI->getArgOperand(i: 2)); |
3737 | if (SizeC && CountC) { |
3738 | uint64_t Bytes = SizeC->getZExtValue() * CountC->getZExtValue(); |
3739 | |
3740 | // If this is writing zero records, remove the call (it's a noop). |
3741 | if (Bytes == 0) |
3742 | return ConstantInt::get(Ty: CI->getType(), V: 0); |
3743 | |
3744 | // If this is writing one byte, turn it into fputc. |
3745 | // This optimisation is only valid, if the return value is unused. |
3746 | if (Bytes == 1 && CI->use_empty()) { // fwrite(S,1,1,F) -> fputc(S[0],F) |
3747 | Value *Char = B.CreateLoad(Ty: B.getInt8Ty(), Ptr: CI->getArgOperand(i: 0), Name: "char" ); |
3748 | Type *IntTy = B.getIntNTy(N: TLI->getIntSize()); |
3749 | Value *Cast = B.CreateIntCast(V: Char, DestTy: IntTy, /*isSigned*/ true, Name: "chari" ); |
3750 | Value *NewCI = emitFPutC(Char: Cast, File: CI->getArgOperand(i: 3), B, TLI); |
3751 | return NewCI ? ConstantInt::get(Ty: CI->getType(), V: 1) : nullptr; |
3752 | } |
3753 | } |
3754 | |
3755 | return nullptr; |
3756 | } |
3757 | |
3758 | Value *LibCallSimplifier::optimizeFPuts(CallInst *CI, IRBuilderBase &B) { |
3759 | optimizeErrorReporting(CI, B, StreamArg: 1); |
3760 | |
3761 | // Don't rewrite fputs to fwrite when optimising for size because fwrite |
3762 | // requires more arguments and thus extra MOVs are required. |
3763 | if (llvm::shouldOptimizeForSize(BB: CI->getParent(), PSI, BFI, |
3764 | QueryType: PGSOQueryType::IRPass)) |
3765 | return nullptr; |
3766 | |
3767 | // We can't optimize if return value is used. |
3768 | if (!CI->use_empty()) |
3769 | return nullptr; |
3770 | |
3771 | // fputs(s,F) --> fwrite(s,strlen(s),1,F) |
3772 | uint64_t Len = GetStringLength(V: CI->getArgOperand(i: 0)); |
3773 | if (!Len) |
3774 | return nullptr; |
3775 | |
3776 | // Known to have no uses (see above). |
3777 | unsigned SizeTBits = TLI->getSizeTSize(M: *CI->getModule()); |
3778 | Type *SizeTTy = IntegerType::get(C&: CI->getContext(), NumBits: SizeTBits); |
3779 | return copyFlags( |
3780 | Old: *CI, |
3781 | New: emitFWrite(Ptr: CI->getArgOperand(i: 0), |
3782 | Size: ConstantInt::get(Ty: SizeTTy, V: Len - 1), |
3783 | File: CI->getArgOperand(i: 1), B, DL, TLI)); |
3784 | } |
3785 | |
3786 | Value *LibCallSimplifier::optimizePuts(CallInst *CI, IRBuilderBase &B) { |
3787 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: 0); |
3788 | if (!CI->use_empty()) |
3789 | return nullptr; |
3790 | |
3791 | // Check for a constant string. |
3792 | // puts("") -> putchar('\n') |
3793 | StringRef Str; |
3794 | if (getConstantStringInfo(V: CI->getArgOperand(i: 0), Str) && Str.empty()) { |
3795 | // putchar takes an argument of the same type as puts returns, i.e., |
3796 | // int, which need not be 32 bits wide. |
3797 | Type *IntTy = CI->getType(); |
3798 | return copyFlags(Old: *CI, New: emitPutChar(Char: ConstantInt::get(Ty: IntTy, V: '\n'), B, TLI)); |
3799 | } |
3800 | |
3801 | return nullptr; |
3802 | } |
3803 | |
3804 | Value *LibCallSimplifier::optimizeExit(CallInst *CI) { |
3805 | |
3806 | // Mark 'exit' as cold if its not exit(0) (success). |
3807 | const APInt *C; |
3808 | if (!CI->hasFnAttr(Kind: Attribute::Cold) && |
3809 | match(V: CI->getArgOperand(i: 0), P: m_APInt(Res&: C)) && !C->isZero()) { |
3810 | CI->addFnAttr(Kind: Attribute::Cold); |
3811 | } |
3812 | return nullptr; |
3813 | } |
3814 | |
3815 | Value *LibCallSimplifier::optimizeBCopy(CallInst *CI, IRBuilderBase &B) { |
3816 | // bcopy(src, dst, n) -> llvm.memmove(dst, src, n) |
3817 | return copyFlags(Old: *CI, New: B.CreateMemMove(Dst: CI->getArgOperand(i: 1), DstAlign: Align(1), |
3818 | Src: CI->getArgOperand(i: 0), SrcAlign: Align(1), |
3819 | Size: CI->getArgOperand(i: 2))); |
3820 | } |
3821 | |
3822 | bool LibCallSimplifier::hasFloatVersion(const Module *M, StringRef FuncName) { |
3823 | SmallString<20> FloatFuncName = FuncName; |
3824 | FloatFuncName += 'f'; |
3825 | return isLibFuncEmittable(M, TLI, Name: FloatFuncName); |
3826 | } |
3827 | |
3828 | Value *LibCallSimplifier::optimizeStringMemoryLibCall(CallInst *CI, |
3829 | IRBuilderBase &Builder) { |
3830 | Module *M = CI->getModule(); |
3831 | LibFunc Func; |
3832 | Function *Callee = CI->getCalledFunction(); |
3833 | |
3834 | // Check for string/memory library functions. |
3835 | if (TLI->getLibFunc(FDecl: *Callee, F&: Func) && isLibFuncEmittable(M, TLI, TheLibFunc: Func)) { |
3836 | // Make sure we never change the calling convention. |
3837 | assert( |
3838 | (ignoreCallingConv(Func) || |
3839 | TargetLibraryInfoImpl::isCallingConvCCompatible(CI)) && |
3840 | "Optimizing string/memory libcall would change the calling convention" ); |
3841 | switch (Func) { |
3842 | case LibFunc_strcat: |
3843 | return optimizeStrCat(CI, B&: Builder); |
3844 | case LibFunc_strncat: |
3845 | return optimizeStrNCat(CI, B&: Builder); |
3846 | case LibFunc_strchr: |
3847 | return optimizeStrChr(CI, B&: Builder); |
3848 | case LibFunc_strrchr: |
3849 | return optimizeStrRChr(CI, B&: Builder); |
3850 | case LibFunc_strcmp: |
3851 | return optimizeStrCmp(CI, B&: Builder); |
3852 | case LibFunc_strncmp: |
3853 | return optimizeStrNCmp(CI, B&: Builder); |
3854 | case LibFunc_strcpy: |
3855 | return optimizeStrCpy(CI, B&: Builder); |
3856 | case LibFunc_stpcpy: |
3857 | return optimizeStpCpy(CI, B&: Builder); |
3858 | case LibFunc_strlcpy: |
3859 | return optimizeStrLCpy(CI, B&: Builder); |
3860 | case LibFunc_stpncpy: |
3861 | return optimizeStringNCpy(CI, /*RetEnd=*/true, B&: Builder); |
3862 | case LibFunc_strncpy: |
3863 | return optimizeStringNCpy(CI, /*RetEnd=*/false, B&: Builder); |
3864 | case LibFunc_strlen: |
3865 | return optimizeStrLen(CI, B&: Builder); |
3866 | case LibFunc_strnlen: |
3867 | return optimizeStrNLen(CI, B&: Builder); |
3868 | case LibFunc_strpbrk: |
3869 | return optimizeStrPBrk(CI, B&: Builder); |
3870 | case LibFunc_strndup: |
3871 | return optimizeStrNDup(CI, B&: Builder); |
3872 | case LibFunc_strtol: |
3873 | case LibFunc_strtod: |
3874 | case LibFunc_strtof: |
3875 | case LibFunc_strtoul: |
3876 | case LibFunc_strtoll: |
3877 | case LibFunc_strtold: |
3878 | case LibFunc_strtoull: |
3879 | return optimizeStrTo(CI, B&: Builder); |
3880 | case LibFunc_strspn: |
3881 | return optimizeStrSpn(CI, B&: Builder); |
3882 | case LibFunc_strcspn: |
3883 | return optimizeStrCSpn(CI, B&: Builder); |
3884 | case LibFunc_strstr: |
3885 | return optimizeStrStr(CI, B&: Builder); |
3886 | case LibFunc_memchr: |
3887 | return optimizeMemChr(CI, B&: Builder); |
3888 | case LibFunc_memrchr: |
3889 | return optimizeMemRChr(CI, B&: Builder); |
3890 | case LibFunc_bcmp: |
3891 | return optimizeBCmp(CI, B&: Builder); |
3892 | case LibFunc_memcmp: |
3893 | return optimizeMemCmp(CI, B&: Builder); |
3894 | case LibFunc_memcpy: |
3895 | return optimizeMemCpy(CI, B&: Builder); |
3896 | case LibFunc_memccpy: |
3897 | return optimizeMemCCpy(CI, B&: Builder); |
3898 | case LibFunc_mempcpy: |
3899 | return optimizeMemPCpy(CI, B&: Builder); |
3900 | case LibFunc_memmove: |
3901 | return optimizeMemMove(CI, B&: Builder); |
3902 | case LibFunc_memset: |
3903 | return optimizeMemSet(CI, B&: Builder); |
3904 | case LibFunc_realloc: |
3905 | return optimizeRealloc(CI, B&: Builder); |
3906 | case LibFunc_wcslen: |
3907 | return optimizeWcslen(CI, B&: Builder); |
3908 | case LibFunc_bcopy: |
3909 | return optimizeBCopy(CI, B&: Builder); |
3910 | case LibFunc_Znwm: |
3911 | case LibFunc_ZnwmRKSt9nothrow_t: |
3912 | case LibFunc_ZnwmSt11align_val_t: |
3913 | case LibFunc_ZnwmSt11align_val_tRKSt9nothrow_t: |
3914 | case LibFunc_Znam: |
3915 | case LibFunc_ZnamRKSt9nothrow_t: |
3916 | case LibFunc_ZnamSt11align_val_t: |
3917 | case LibFunc_ZnamSt11align_val_tRKSt9nothrow_t: |
3918 | case LibFunc_Znwm12__hot_cold_t: |
3919 | case LibFunc_ZnwmRKSt9nothrow_t12__hot_cold_t: |
3920 | case LibFunc_ZnwmSt11align_val_t12__hot_cold_t: |
3921 | case LibFunc_ZnwmSt11align_val_tRKSt9nothrow_t12__hot_cold_t: |
3922 | case LibFunc_Znam12__hot_cold_t: |
3923 | case LibFunc_ZnamRKSt9nothrow_t12__hot_cold_t: |
3924 | case LibFunc_ZnamSt11align_val_t12__hot_cold_t: |
3925 | case LibFunc_ZnamSt11align_val_tRKSt9nothrow_t12__hot_cold_t: |
3926 | case LibFunc_size_returning_new: |
3927 | case LibFunc_size_returning_new_hot_cold: |
3928 | case LibFunc_size_returning_new_aligned: |
3929 | case LibFunc_size_returning_new_aligned_hot_cold: |
3930 | return optimizeNew(CI, B&: Builder, Func); |
3931 | default: |
3932 | break; |
3933 | } |
3934 | } |
3935 | return nullptr; |
3936 | } |
3937 | |
3938 | /// Constant folding nan/nanf/nanl. |
3939 | static Value *optimizeNaN(CallInst *CI) { |
3940 | StringRef CharSeq; |
3941 | if (!getConstantStringInfo(V: CI->getArgOperand(i: 0), Str&: CharSeq)) |
3942 | return nullptr; |
3943 | |
3944 | APInt Fill; |
3945 | // Treat empty strings as if they were zero. |
3946 | if (CharSeq.empty()) |
3947 | Fill = APInt(32, 0); |
3948 | else if (CharSeq.getAsInteger(Radix: 0, Result&: Fill)) |
3949 | return nullptr; |
3950 | |
3951 | return ConstantFP::getQNaN(Ty: CI->getType(), /*Negative=*/false, Payload: &Fill); |
3952 | } |
3953 | |
3954 | Value *LibCallSimplifier::optimizeFloatingPointLibCall(CallInst *CI, |
3955 | LibFunc Func, |
3956 | IRBuilderBase &Builder) { |
3957 | const Module *M = CI->getModule(); |
3958 | |
3959 | // Don't optimize calls that require strict floating point semantics. |
3960 | if (CI->isStrictFP()) |
3961 | return nullptr; |
3962 | |
3963 | if (Value *V = optimizeSymmetric(CI, Func, B&: Builder)) |
3964 | return V; |
3965 | |
3966 | switch (Func) { |
3967 | case LibFunc_sinpif: |
3968 | case LibFunc_sinpi: |
3969 | return optimizeSinCosPi(CI, /*IsSin*/true, B&: Builder); |
3970 | case LibFunc_cospif: |
3971 | case LibFunc_cospi: |
3972 | return optimizeSinCosPi(CI, /*IsSin*/false, B&: Builder); |
3973 | case LibFunc_powf: |
3974 | case LibFunc_pow: |
3975 | case LibFunc_powl: |
3976 | return optimizePow(Pow: CI, B&: Builder); |
3977 | case LibFunc_exp2l: |
3978 | case LibFunc_exp2: |
3979 | case LibFunc_exp2f: |
3980 | return optimizeExp2(CI, B&: Builder); |
3981 | case LibFunc_fabsf: |
3982 | case LibFunc_fabs: |
3983 | case LibFunc_fabsl: |
3984 | return replaceUnaryCall(CI, B&: Builder, IID: Intrinsic::fabs); |
3985 | case LibFunc_sqrtf: |
3986 | case LibFunc_sqrt: |
3987 | case LibFunc_sqrtl: |
3988 | return optimizeSqrt(CI, B&: Builder); |
3989 | case LibFunc_fmod: |
3990 | case LibFunc_fmodf: |
3991 | case LibFunc_fmodl: |
3992 | return optimizeFMod(CI, B&: Builder); |
3993 | case LibFunc_logf: |
3994 | case LibFunc_log: |
3995 | case LibFunc_logl: |
3996 | case LibFunc_log10f: |
3997 | case LibFunc_log10: |
3998 | case LibFunc_log10l: |
3999 | case LibFunc_log1pf: |
4000 | case LibFunc_log1p: |
4001 | case LibFunc_log1pl: |
4002 | case LibFunc_log2f: |
4003 | case LibFunc_log2: |
4004 | case LibFunc_log2l: |
4005 | case LibFunc_logbf: |
4006 | case LibFunc_logb: |
4007 | case LibFunc_logbl: |
4008 | return optimizeLog(Log: CI, B&: Builder); |
4009 | case LibFunc_tan: |
4010 | case LibFunc_tanf: |
4011 | case LibFunc_tanl: |
4012 | case LibFunc_sinh: |
4013 | case LibFunc_sinhf: |
4014 | case LibFunc_sinhl: |
4015 | case LibFunc_asinh: |
4016 | case LibFunc_asinhf: |
4017 | case LibFunc_asinhl: |
4018 | case LibFunc_cosh: |
4019 | case LibFunc_coshf: |
4020 | case LibFunc_coshl: |
4021 | case LibFunc_atanh: |
4022 | case LibFunc_atanhf: |
4023 | case LibFunc_atanhl: |
4024 | return optimizeTrigInversionPairs(CI, B&: Builder); |
4025 | case LibFunc_ceil: |
4026 | return replaceUnaryCall(CI, B&: Builder, IID: Intrinsic::ceil); |
4027 | case LibFunc_floor: |
4028 | return replaceUnaryCall(CI, B&: Builder, IID: Intrinsic::floor); |
4029 | case LibFunc_round: |
4030 | return replaceUnaryCall(CI, B&: Builder, IID: Intrinsic::round); |
4031 | case LibFunc_roundeven: |
4032 | return replaceUnaryCall(CI, B&: Builder, IID: Intrinsic::roundeven); |
4033 | case LibFunc_nearbyint: |
4034 | return replaceUnaryCall(CI, B&: Builder, IID: Intrinsic::nearbyint); |
4035 | case LibFunc_rint: |
4036 | return replaceUnaryCall(CI, B&: Builder, IID: Intrinsic::rint); |
4037 | case LibFunc_trunc: |
4038 | return replaceUnaryCall(CI, B&: Builder, IID: Intrinsic::trunc); |
4039 | case LibFunc_acos: |
4040 | case LibFunc_acosh: |
4041 | case LibFunc_asin: |
4042 | case LibFunc_atan: |
4043 | case LibFunc_cbrt: |
4044 | case LibFunc_exp: |
4045 | case LibFunc_exp10: |
4046 | case LibFunc_expm1: |
4047 | case LibFunc_cos: |
4048 | case LibFunc_sin: |
4049 | case LibFunc_tanh: |
4050 | if (UnsafeFPShrink && hasFloatVersion(M, FuncName: CI->getCalledFunction()->getName())) |
4051 | return optimizeUnaryDoubleFP(CI, B&: Builder, TLI, isPrecise: true); |
4052 | return nullptr; |
4053 | case LibFunc_copysign: |
4054 | if (hasFloatVersion(M, FuncName: CI->getCalledFunction()->getName())) |
4055 | return optimizeBinaryDoubleFP(CI, B&: Builder, TLI); |
4056 | return nullptr; |
4057 | case LibFunc_fdim: |
4058 | case LibFunc_fdimf: |
4059 | case LibFunc_fdiml: |
4060 | return optimizeFdim(CI, B&: Builder); |
4061 | case LibFunc_fminf: |
4062 | case LibFunc_fmin: |
4063 | case LibFunc_fminl: |
4064 | case LibFunc_fmaxf: |
4065 | case LibFunc_fmax: |
4066 | case LibFunc_fmaxl: |
4067 | return optimizeFMinFMax(CI, B&: Builder); |
4068 | case LibFunc_cabs: |
4069 | case LibFunc_cabsf: |
4070 | case LibFunc_cabsl: |
4071 | return optimizeCAbs(CI, B&: Builder); |
4072 | case LibFunc_remquo: |
4073 | case LibFunc_remquof: |
4074 | case LibFunc_remquol: |
4075 | return optimizeRemquo(CI, B&: Builder); |
4076 | case LibFunc_nan: |
4077 | case LibFunc_nanf: |
4078 | case LibFunc_nanl: |
4079 | return optimizeNaN(CI); |
4080 | default: |
4081 | return nullptr; |
4082 | } |
4083 | } |
4084 | |
4085 | Value *LibCallSimplifier::optimizeCall(CallInst *CI, IRBuilderBase &Builder) { |
4086 | Module *M = CI->getModule(); |
4087 | assert(!CI->isMustTailCall() && "These transforms aren't musttail safe." ); |
4088 | |
4089 | // TODO: Split out the code below that operates on FP calls so that |
4090 | // we can all non-FP calls with the StrictFP attribute to be |
4091 | // optimized. |
4092 | if (CI->isNoBuiltin()) |
4093 | return nullptr; |
4094 | |
4095 | LibFunc Func; |
4096 | Function *Callee = CI->getCalledFunction(); |
4097 | bool IsCallingConvC = TargetLibraryInfoImpl::isCallingConvCCompatible(CI); |
4098 | |
4099 | SmallVector<OperandBundleDef, 2> OpBundles; |
4100 | CI->getOperandBundlesAsDefs(Defs&: OpBundles); |
4101 | |
4102 | IRBuilderBase::OperandBundlesGuard Guard(Builder); |
4103 | Builder.setDefaultOperandBundles(OpBundles); |
4104 | |
4105 | // Command-line parameter overrides instruction attribute. |
4106 | // This can't be moved to optimizeFloatingPointLibCall() because it may be |
4107 | // used by the intrinsic optimizations. |
4108 | if (EnableUnsafeFPShrink.getNumOccurrences() > 0) |
4109 | UnsafeFPShrink = EnableUnsafeFPShrink; |
4110 | else if (isa<FPMathOperator>(Val: CI) && CI->isFast()) |
4111 | UnsafeFPShrink = true; |
4112 | |
4113 | // First, check for intrinsics. |
4114 | if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: CI)) { |
4115 | if (!IsCallingConvC) |
4116 | return nullptr; |
4117 | // The FP intrinsics have corresponding constrained versions so we don't |
4118 | // need to check for the StrictFP attribute here. |
4119 | switch (II->getIntrinsicID()) { |
4120 | case Intrinsic::pow: |
4121 | return optimizePow(Pow: CI, B&: Builder); |
4122 | case Intrinsic::exp2: |
4123 | return optimizeExp2(CI, B&: Builder); |
4124 | case Intrinsic::log: |
4125 | case Intrinsic::log2: |
4126 | case Intrinsic::log10: |
4127 | return optimizeLog(Log: CI, B&: Builder); |
4128 | case Intrinsic::sqrt: |
4129 | return optimizeSqrt(CI, B&: Builder); |
4130 | case Intrinsic::memset: |
4131 | return optimizeMemSet(CI, B&: Builder); |
4132 | case Intrinsic::memcpy: |
4133 | return optimizeMemCpy(CI, B&: Builder); |
4134 | case Intrinsic::memmove: |
4135 | return optimizeMemMove(CI, B&: Builder); |
4136 | case Intrinsic::sin: |
4137 | case Intrinsic::cos: |
4138 | if (UnsafeFPShrink) |
4139 | return optimizeUnaryDoubleFP(CI, B&: Builder, TLI, /*isPrecise=*/true); |
4140 | return nullptr; |
4141 | default: |
4142 | return nullptr; |
4143 | } |
4144 | } |
4145 | |
4146 | // Also try to simplify calls to fortified library functions. |
4147 | if (Value *SimplifiedFortifiedCI = |
4148 | FortifiedSimplifier.optimizeCall(CI, B&: Builder)) |
4149 | return SimplifiedFortifiedCI; |
4150 | |
4151 | // Then check for known library functions. |
4152 | if (TLI->getLibFunc(FDecl: *Callee, F&: Func) && isLibFuncEmittable(M, TLI, TheLibFunc: Func)) { |
4153 | // We never change the calling convention. |
4154 | if (!ignoreCallingConv(Func) && !IsCallingConvC) |
4155 | return nullptr; |
4156 | if (Value *V = optimizeStringMemoryLibCall(CI, Builder)) |
4157 | return V; |
4158 | if (Value *V = optimizeFloatingPointLibCall(CI, Func, Builder)) |
4159 | return V; |
4160 | switch (Func) { |
4161 | case LibFunc_ffs: |
4162 | case LibFunc_ffsl: |
4163 | case LibFunc_ffsll: |
4164 | return optimizeFFS(CI, B&: Builder); |
4165 | case LibFunc_fls: |
4166 | case LibFunc_flsl: |
4167 | case LibFunc_flsll: |
4168 | return optimizeFls(CI, B&: Builder); |
4169 | case LibFunc_abs: |
4170 | case LibFunc_labs: |
4171 | case LibFunc_llabs: |
4172 | return optimizeAbs(CI, B&: Builder); |
4173 | case LibFunc_isdigit: |
4174 | return optimizeIsDigit(CI, B&: Builder); |
4175 | case LibFunc_isascii: |
4176 | return optimizeIsAscii(CI, B&: Builder); |
4177 | case LibFunc_toascii: |
4178 | return optimizeToAscii(CI, B&: Builder); |
4179 | case LibFunc_atoi: |
4180 | case LibFunc_atol: |
4181 | case LibFunc_atoll: |
4182 | return optimizeAtoi(CI, B&: Builder); |
4183 | case LibFunc_strtol: |
4184 | case LibFunc_strtoll: |
4185 | return optimizeStrToInt(CI, B&: Builder, /*AsSigned=*/true); |
4186 | case LibFunc_strtoul: |
4187 | case LibFunc_strtoull: |
4188 | return optimizeStrToInt(CI, B&: Builder, /*AsSigned=*/false); |
4189 | case LibFunc_printf: |
4190 | return optimizePrintF(CI, B&: Builder); |
4191 | case LibFunc_sprintf: |
4192 | return optimizeSPrintF(CI, B&: Builder); |
4193 | case LibFunc_snprintf: |
4194 | return optimizeSnPrintF(CI, B&: Builder); |
4195 | case LibFunc_fprintf: |
4196 | return optimizeFPrintF(CI, B&: Builder); |
4197 | case LibFunc_fwrite: |
4198 | return optimizeFWrite(CI, B&: Builder); |
4199 | case LibFunc_fputs: |
4200 | return optimizeFPuts(CI, B&: Builder); |
4201 | case LibFunc_puts: |
4202 | return optimizePuts(CI, B&: Builder); |
4203 | case LibFunc_perror: |
4204 | return optimizeErrorReporting(CI, B&: Builder); |
4205 | case LibFunc_vfprintf: |
4206 | case LibFunc_fiprintf: |
4207 | return optimizeErrorReporting(CI, B&: Builder, StreamArg: 0); |
4208 | case LibFunc_exit: |
4209 | case LibFunc_Exit: |
4210 | return optimizeExit(CI); |
4211 | default: |
4212 | return nullptr; |
4213 | } |
4214 | } |
4215 | return nullptr; |
4216 | } |
4217 | |
4218 | LibCallSimplifier::LibCallSimplifier( |
4219 | const DataLayout &DL, const TargetLibraryInfo *TLI, DominatorTree *DT, |
4220 | DomConditionCache *DC, AssumptionCache *AC, OptimizationRemarkEmitter &ORE, |
4221 | BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, |
4222 | function_ref<void(Instruction *, Value *)> Replacer, |
4223 | function_ref<void(Instruction *)> Eraser) |
4224 | : FortifiedSimplifier(TLI), DL(DL), TLI(TLI), DT(DT), DC(DC), AC(AC), |
4225 | ORE(ORE), BFI(BFI), PSI(PSI), Replacer(Replacer), Eraser(Eraser) {} |
4226 | |
4227 | void LibCallSimplifier::replaceAllUsesWith(Instruction *I, Value *With) { |
4228 | // Indirect through the replacer used in this instance. |
4229 | Replacer(I, With); |
4230 | } |
4231 | |
4232 | void LibCallSimplifier::eraseFromParent(Instruction *I) { |
4233 | Eraser(I); |
4234 | } |
4235 | |
4236 | // TODO: |
4237 | // Additional cases that we need to add to this file: |
4238 | // |
4239 | // cbrt: |
4240 | // * cbrt(expN(X)) -> expN(x/3) |
4241 | // * cbrt(sqrt(x)) -> pow(x,1/6) |
4242 | // * cbrt(cbrt(x)) -> pow(x,1/9) |
4243 | // |
4244 | // exp, expf, expl: |
4245 | // * exp(log(x)) -> x |
4246 | // |
4247 | // log, logf, logl: |
4248 | // * log(exp(x)) -> x |
4249 | // * log(exp(y)) -> y*log(e) |
4250 | // * log(exp10(y)) -> y*log(10) |
4251 | // * log(sqrt(x)) -> 0.5*log(x) |
4252 | // |
4253 | // pow, powf, powl: |
4254 | // * pow(sqrt(x),y) -> pow(x,y*0.5) |
4255 | // * pow(pow(x,y),z)-> pow(x,y*z) |
4256 | // |
4257 | // signbit: |
4258 | // * signbit(cnst) -> cnst' |
4259 | // * signbit(nncst) -> 0 (if pstv is a non-negative constant) |
4260 | // |
4261 | // sqrt, sqrtf, sqrtl: |
4262 | // * sqrt(expN(x)) -> expN(x*0.5) |
4263 | // * sqrt(Nroot(x)) -> pow(x,1/(2*N)) |
4264 | // * sqrt(pow(x,y)) -> pow(|x|,y*0.5) |
4265 | // |
4266 | |
4267 | //===----------------------------------------------------------------------===// |
4268 | // Fortified Library Call Optimizations |
4269 | //===----------------------------------------------------------------------===// |
4270 | |
4271 | bool FortifiedLibCallSimplifier::isFortifiedCallFoldable( |
4272 | CallInst *CI, unsigned ObjSizeOp, std::optional<unsigned> SizeOp, |
4273 | std::optional<unsigned> StrOp, std::optional<unsigned> FlagOp) { |
4274 | // If this function takes a flag argument, the implementation may use it to |
4275 | // perform extra checks. Don't fold into the non-checking variant. |
4276 | if (FlagOp) { |
4277 | ConstantInt *Flag = dyn_cast<ConstantInt>(Val: CI->getArgOperand(i: *FlagOp)); |
4278 | if (!Flag || !Flag->isZero()) |
4279 | return false; |
4280 | } |
4281 | |
4282 | if (SizeOp && CI->getArgOperand(i: ObjSizeOp) == CI->getArgOperand(i: *SizeOp)) |
4283 | return true; |
4284 | |
4285 | if (ConstantInt *ObjSizeCI = |
4286 | dyn_cast<ConstantInt>(Val: CI->getArgOperand(i: ObjSizeOp))) { |
4287 | if (ObjSizeCI->isMinusOne()) |
4288 | return true; |
4289 | // If the object size wasn't -1 (unknown), bail out if we were asked to. |
4290 | if (OnlyLowerUnknownSize) |
4291 | return false; |
4292 | if (StrOp) { |
4293 | uint64_t Len = GetStringLength(V: CI->getArgOperand(i: *StrOp)); |
4294 | // If the length is 0 we don't know how long it is and so we can't |
4295 | // remove the check. |
4296 | if (Len) |
4297 | annotateDereferenceableBytes(CI, ArgNos: *StrOp, DereferenceableBytes: Len); |
4298 | else |
4299 | return false; |
4300 | return ObjSizeCI->getZExtValue() >= Len; |
4301 | } |
4302 | |
4303 | if (SizeOp) { |
4304 | if (ConstantInt *SizeCI = |
4305 | dyn_cast<ConstantInt>(Val: CI->getArgOperand(i: *SizeOp))) |
4306 | return ObjSizeCI->getZExtValue() >= SizeCI->getZExtValue(); |
4307 | } |
4308 | } |
4309 | return false; |
4310 | } |
4311 | |
4312 | Value *FortifiedLibCallSimplifier::optimizeMemCpyChk(CallInst *CI, |
4313 | IRBuilderBase &B) { |
4314 | if (isFortifiedCallFoldable(CI, ObjSizeOp: 3, SizeOp: 2)) { |
4315 | CallInst *NewCI = |
4316 | B.CreateMemCpy(Dst: CI->getArgOperand(i: 0), DstAlign: Align(1), Src: CI->getArgOperand(i: 1), |
4317 | SrcAlign: Align(1), Size: CI->getArgOperand(i: 2)); |
4318 | mergeAttributesAndFlags(NewCI, Old: *CI); |
4319 | return CI->getArgOperand(i: 0); |
4320 | } |
4321 | return nullptr; |
4322 | } |
4323 | |
4324 | Value *FortifiedLibCallSimplifier::optimizeMemMoveChk(CallInst *CI, |
4325 | IRBuilderBase &B) { |
4326 | if (isFortifiedCallFoldable(CI, ObjSizeOp: 3, SizeOp: 2)) { |
4327 | CallInst *NewCI = |
4328 | B.CreateMemMove(Dst: CI->getArgOperand(i: 0), DstAlign: Align(1), Src: CI->getArgOperand(i: 1), |
4329 | SrcAlign: Align(1), Size: CI->getArgOperand(i: 2)); |
4330 | mergeAttributesAndFlags(NewCI, Old: *CI); |
4331 | return CI->getArgOperand(i: 0); |
4332 | } |
4333 | return nullptr; |
4334 | } |
4335 | |
4336 | Value *FortifiedLibCallSimplifier::optimizeMemSetChk(CallInst *CI, |
4337 | IRBuilderBase &B) { |
4338 | if (isFortifiedCallFoldable(CI, ObjSizeOp: 3, SizeOp: 2)) { |
4339 | Value *Val = B.CreateIntCast(V: CI->getArgOperand(i: 1), DestTy: B.getInt8Ty(), isSigned: false); |
4340 | CallInst *NewCI = B.CreateMemSet(Ptr: CI->getArgOperand(i: 0), Val, |
4341 | Size: CI->getArgOperand(i: 2), Align: Align(1)); |
4342 | mergeAttributesAndFlags(NewCI, Old: *CI); |
4343 | return CI->getArgOperand(i: 0); |
4344 | } |
4345 | return nullptr; |
4346 | } |
4347 | |
4348 | Value *FortifiedLibCallSimplifier::optimizeMemPCpyChk(CallInst *CI, |
4349 | IRBuilderBase &B) { |
4350 | const DataLayout &DL = CI->getDataLayout(); |
4351 | if (isFortifiedCallFoldable(CI, ObjSizeOp: 3, SizeOp: 2)) |
4352 | if (Value *Call = emitMemPCpy(Dst: CI->getArgOperand(i: 0), Src: CI->getArgOperand(i: 1), |
4353 | Len: CI->getArgOperand(i: 2), B, DL, TLI)) { |
4354 | return mergeAttributesAndFlags(NewCI: cast<CallInst>(Val: Call), Old: *CI); |
4355 | } |
4356 | return nullptr; |
4357 | } |
4358 | |
4359 | Value *FortifiedLibCallSimplifier::optimizeStrpCpyChk(CallInst *CI, |
4360 | IRBuilderBase &B, |
4361 | LibFunc Func) { |
4362 | const DataLayout &DL = CI->getDataLayout(); |
4363 | Value *Dst = CI->getArgOperand(i: 0), *Src = CI->getArgOperand(i: 1), |
4364 | *ObjSize = CI->getArgOperand(i: 2); |
4365 | |
4366 | // __stpcpy_chk(x,x,...) -> x+strlen(x) |
4367 | if (Func == LibFunc_stpcpy_chk && !OnlyLowerUnknownSize && Dst == Src) { |
4368 | Value *StrLen = emitStrLen(Ptr: Src, B, DL, TLI); |
4369 | return StrLen ? B.CreateInBoundsGEP(Ty: B.getInt8Ty(), Ptr: Dst, IdxList: StrLen) : nullptr; |
4370 | } |
4371 | |
4372 | // If a) we don't have any length information, or b) we know this will |
4373 | // fit then just lower to a plain st[rp]cpy. Otherwise we'll keep our |
4374 | // st[rp]cpy_chk call which may fail at runtime if the size is too long. |
4375 | // TODO: It might be nice to get a maximum length out of the possible |
4376 | // string lengths for varying. |
4377 | if (isFortifiedCallFoldable(CI, ObjSizeOp: 2, SizeOp: std::nullopt, StrOp: 1)) { |
4378 | if (Func == LibFunc_strcpy_chk) |
4379 | return copyFlags(Old: *CI, New: emitStrCpy(Dst, Src, B, TLI)); |
4380 | else |
4381 | return copyFlags(Old: *CI, New: emitStpCpy(Dst, Src, B, TLI)); |
4382 | } |
4383 | |
4384 | if (OnlyLowerUnknownSize) |
4385 | return nullptr; |
4386 | |
4387 | // Maybe we can stil fold __st[rp]cpy_chk to __memcpy_chk. |
4388 | uint64_t Len = GetStringLength(V: Src); |
4389 | if (Len) |
4390 | annotateDereferenceableBytes(CI, ArgNos: 1, DereferenceableBytes: Len); |
4391 | else |
4392 | return nullptr; |
4393 | |
4394 | unsigned SizeTBits = TLI->getSizeTSize(M: *CI->getModule()); |
4395 | Type *SizeTTy = IntegerType::get(C&: CI->getContext(), NumBits: SizeTBits); |
4396 | Value *LenV = ConstantInt::get(Ty: SizeTTy, V: Len); |
4397 | Value *Ret = emitMemCpyChk(Dst, Src, Len: LenV, ObjSize, B, DL, TLI); |
4398 | // If the function was an __stpcpy_chk, and we were able to fold it into |
4399 | // a __memcpy_chk, we still need to return the correct end pointer. |
4400 | if (Ret && Func == LibFunc_stpcpy_chk) |
4401 | return B.CreateInBoundsGEP(Ty: B.getInt8Ty(), Ptr: Dst, |
4402 | IdxList: ConstantInt::get(Ty: SizeTTy, V: Len - 1)); |
4403 | return copyFlags(Old: *CI, New: cast<CallInst>(Val: Ret)); |
4404 | } |
4405 | |
4406 | Value *FortifiedLibCallSimplifier::optimizeStrLenChk(CallInst *CI, |
4407 | IRBuilderBase &B) { |
4408 | if (isFortifiedCallFoldable(CI, ObjSizeOp: 1, SizeOp: std::nullopt, StrOp: 0)) |
4409 | return copyFlags(Old: *CI, New: emitStrLen(Ptr: CI->getArgOperand(i: 0), B, |
4410 | DL: CI->getDataLayout(), TLI)); |
4411 | return nullptr; |
4412 | } |
4413 | |
4414 | Value *FortifiedLibCallSimplifier::optimizeStrpNCpyChk(CallInst *CI, |
4415 | IRBuilderBase &B, |
4416 | LibFunc Func) { |
4417 | if (isFortifiedCallFoldable(CI, ObjSizeOp: 3, SizeOp: 2)) { |
4418 | if (Func == LibFunc_strncpy_chk) |
4419 | return copyFlags(Old: *CI, |
4420 | New: emitStrNCpy(Dst: CI->getArgOperand(i: 0), Src: CI->getArgOperand(i: 1), |
4421 | Len: CI->getArgOperand(i: 2), B, TLI)); |
4422 | else |
4423 | return copyFlags(Old: *CI, |
4424 | New: emitStpNCpy(Dst: CI->getArgOperand(i: 0), Src: CI->getArgOperand(i: 1), |
4425 | Len: CI->getArgOperand(i: 2), B, TLI)); |
4426 | } |
4427 | |
4428 | return nullptr; |
4429 | } |
4430 | |
4431 | Value *FortifiedLibCallSimplifier::optimizeMemCCpyChk(CallInst *CI, |
4432 | IRBuilderBase &B) { |
4433 | if (isFortifiedCallFoldable(CI, ObjSizeOp: 4, SizeOp: 3)) |
4434 | return copyFlags( |
4435 | Old: *CI, New: emitMemCCpy(Ptr1: CI->getArgOperand(i: 0), Ptr2: CI->getArgOperand(i: 1), |
4436 | Val: CI->getArgOperand(i: 2), Len: CI->getArgOperand(i: 3), B, TLI)); |
4437 | |
4438 | return nullptr; |
4439 | } |
4440 | |
4441 | Value *FortifiedLibCallSimplifier::optimizeSNPrintfChk(CallInst *CI, |
4442 | IRBuilderBase &B) { |
4443 | if (isFortifiedCallFoldable(CI, ObjSizeOp: 3, SizeOp: 1, StrOp: std::nullopt, FlagOp: 2)) { |
4444 | SmallVector<Value *, 8> VariadicArgs(drop_begin(RangeOrContainer: CI->args(), N: 5)); |
4445 | return copyFlags(Old: *CI, |
4446 | New: emitSNPrintf(Dest: CI->getArgOperand(i: 0), Size: CI->getArgOperand(i: 1), |
4447 | Fmt: CI->getArgOperand(i: 4), Args: VariadicArgs, B, TLI)); |
4448 | } |
4449 | |
4450 | return nullptr; |
4451 | } |
4452 | |
4453 | Value *FortifiedLibCallSimplifier::optimizeSPrintfChk(CallInst *CI, |
4454 | IRBuilderBase &B) { |
4455 | if (isFortifiedCallFoldable(CI, ObjSizeOp: 2, SizeOp: std::nullopt, StrOp: std::nullopt, FlagOp: 1)) { |
4456 | SmallVector<Value *, 8> VariadicArgs(drop_begin(RangeOrContainer: CI->args(), N: 4)); |
4457 | return copyFlags(Old: *CI, |
4458 | New: emitSPrintf(Dest: CI->getArgOperand(i: 0), Fmt: CI->getArgOperand(i: 3), |
4459 | VariadicArgs, B, TLI)); |
4460 | } |
4461 | |
4462 | return nullptr; |
4463 | } |
4464 | |
4465 | Value *FortifiedLibCallSimplifier::optimizeStrCatChk(CallInst *CI, |
4466 | IRBuilderBase &B) { |
4467 | if (isFortifiedCallFoldable(CI, ObjSizeOp: 2)) |
4468 | return copyFlags( |
4469 | Old: *CI, New: emitStrCat(Dest: CI->getArgOperand(i: 0), Src: CI->getArgOperand(i: 1), B, TLI)); |
4470 | |
4471 | return nullptr; |
4472 | } |
4473 | |
4474 | Value *FortifiedLibCallSimplifier::optimizeStrLCat(CallInst *CI, |
4475 | IRBuilderBase &B) { |
4476 | if (isFortifiedCallFoldable(CI, ObjSizeOp: 3)) |
4477 | return copyFlags(Old: *CI, |
4478 | New: emitStrLCat(Dest: CI->getArgOperand(i: 0), Src: CI->getArgOperand(i: 1), |
4479 | Size: CI->getArgOperand(i: 2), B, TLI)); |
4480 | |
4481 | return nullptr; |
4482 | } |
4483 | |
4484 | Value *FortifiedLibCallSimplifier::optimizeStrNCatChk(CallInst *CI, |
4485 | IRBuilderBase &B) { |
4486 | if (isFortifiedCallFoldable(CI, ObjSizeOp: 3)) |
4487 | return copyFlags(Old: *CI, |
4488 | New: emitStrNCat(Dest: CI->getArgOperand(i: 0), Src: CI->getArgOperand(i: 1), |
4489 | Size: CI->getArgOperand(i: 2), B, TLI)); |
4490 | |
4491 | return nullptr; |
4492 | } |
4493 | |
4494 | Value *FortifiedLibCallSimplifier::optimizeStrLCpyChk(CallInst *CI, |
4495 | IRBuilderBase &B) { |
4496 | if (isFortifiedCallFoldable(CI, ObjSizeOp: 3)) |
4497 | return copyFlags(Old: *CI, |
4498 | New: emitStrLCpy(Dest: CI->getArgOperand(i: 0), Src: CI->getArgOperand(i: 1), |
4499 | Size: CI->getArgOperand(i: 2), B, TLI)); |
4500 | |
4501 | return nullptr; |
4502 | } |
4503 | |
4504 | Value *FortifiedLibCallSimplifier::optimizeVSNPrintfChk(CallInst *CI, |
4505 | IRBuilderBase &B) { |
4506 | if (isFortifiedCallFoldable(CI, ObjSizeOp: 3, SizeOp: 1, StrOp: std::nullopt, FlagOp: 2)) |
4507 | return copyFlags( |
4508 | Old: *CI, New: emitVSNPrintf(Dest: CI->getArgOperand(i: 0), Size: CI->getArgOperand(i: 1), |
4509 | Fmt: CI->getArgOperand(i: 4), VAList: CI->getArgOperand(i: 5), B, TLI)); |
4510 | |
4511 | return nullptr; |
4512 | } |
4513 | |
4514 | Value *FortifiedLibCallSimplifier::optimizeVSPrintfChk(CallInst *CI, |
4515 | IRBuilderBase &B) { |
4516 | if (isFortifiedCallFoldable(CI, ObjSizeOp: 2, SizeOp: std::nullopt, StrOp: std::nullopt, FlagOp: 1)) |
4517 | return copyFlags(Old: *CI, |
4518 | New: emitVSPrintf(Dest: CI->getArgOperand(i: 0), Fmt: CI->getArgOperand(i: 3), |
4519 | VAList: CI->getArgOperand(i: 4), B, TLI)); |
4520 | |
4521 | return nullptr; |
4522 | } |
4523 | |
4524 | Value *FortifiedLibCallSimplifier::optimizeCall(CallInst *CI, |
4525 | IRBuilderBase &Builder) { |
4526 | // FIXME: We shouldn't be changing "nobuiltin" or TLI unavailable calls here. |
4527 | // Some clang users checked for _chk libcall availability using: |
4528 | // __has_builtin(__builtin___memcpy_chk) |
4529 | // When compiling with -fno-builtin, this is always true. |
4530 | // When passing -ffreestanding/-mkernel, which both imply -fno-builtin, we |
4531 | // end up with fortified libcalls, which isn't acceptable in a freestanding |
4532 | // environment which only provides their non-fortified counterparts. |
4533 | // |
4534 | // Until we change clang and/or teach external users to check for availability |
4535 | // differently, disregard the "nobuiltin" attribute and TLI::has. |
4536 | // |
4537 | // PR23093. |
4538 | |
4539 | LibFunc Func; |
4540 | Function *Callee = CI->getCalledFunction(); |
4541 | bool IsCallingConvC = TargetLibraryInfoImpl::isCallingConvCCompatible(CI); |
4542 | |
4543 | SmallVector<OperandBundleDef, 2> OpBundles; |
4544 | CI->getOperandBundlesAsDefs(Defs&: OpBundles); |
4545 | |
4546 | IRBuilderBase::OperandBundlesGuard Guard(Builder); |
4547 | Builder.setDefaultOperandBundles(OpBundles); |
4548 | |
4549 | // First, check that this is a known library functions and that the prototype |
4550 | // is correct. |
4551 | if (!TLI->getLibFunc(FDecl: *Callee, F&: Func)) |
4552 | return nullptr; |
4553 | |
4554 | // We never change the calling convention. |
4555 | if (!ignoreCallingConv(Func) && !IsCallingConvC) |
4556 | return nullptr; |
4557 | |
4558 | switch (Func) { |
4559 | case LibFunc_memcpy_chk: |
4560 | return optimizeMemCpyChk(CI, B&: Builder); |
4561 | case LibFunc_mempcpy_chk: |
4562 | return optimizeMemPCpyChk(CI, B&: Builder); |
4563 | case LibFunc_memmove_chk: |
4564 | return optimizeMemMoveChk(CI, B&: Builder); |
4565 | case LibFunc_memset_chk: |
4566 | return optimizeMemSetChk(CI, B&: Builder); |
4567 | case LibFunc_stpcpy_chk: |
4568 | case LibFunc_strcpy_chk: |
4569 | return optimizeStrpCpyChk(CI, B&: Builder, Func); |
4570 | case LibFunc_strlen_chk: |
4571 | return optimizeStrLenChk(CI, B&: Builder); |
4572 | case LibFunc_stpncpy_chk: |
4573 | case LibFunc_strncpy_chk: |
4574 | return optimizeStrpNCpyChk(CI, B&: Builder, Func); |
4575 | case LibFunc_memccpy_chk: |
4576 | return optimizeMemCCpyChk(CI, B&: Builder); |
4577 | case LibFunc_snprintf_chk: |
4578 | return optimizeSNPrintfChk(CI, B&: Builder); |
4579 | case LibFunc_sprintf_chk: |
4580 | return optimizeSPrintfChk(CI, B&: Builder); |
4581 | case LibFunc_strcat_chk: |
4582 | return optimizeStrCatChk(CI, B&: Builder); |
4583 | case LibFunc_strlcat_chk: |
4584 | return optimizeStrLCat(CI, B&: Builder); |
4585 | case LibFunc_strncat_chk: |
4586 | return optimizeStrNCatChk(CI, B&: Builder); |
4587 | case LibFunc_strlcpy_chk: |
4588 | return optimizeStrLCpyChk(CI, B&: Builder); |
4589 | case LibFunc_vsnprintf_chk: |
4590 | return optimizeVSNPrintfChk(CI, B&: Builder); |
4591 | case LibFunc_vsprintf_chk: |
4592 | return optimizeVSPrintfChk(CI, B&: Builder); |
4593 | default: |
4594 | break; |
4595 | } |
4596 | return nullptr; |
4597 | } |
4598 | |
4599 | FortifiedLibCallSimplifier::FortifiedLibCallSimplifier( |
4600 | const TargetLibraryInfo *TLI, bool OnlyLowerUnknownSize) |
4601 | : TLI(TLI), OnlyLowerUnknownSize(OnlyLowerUnknownSize) {} |
4602 | |