| 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/Utils/Local.h" |
| 24 | #include "llvm/Analysis/ValueTracking.h" |
| 25 | #include "llvm/IR/AttributeMask.h" |
| 26 | #include "llvm/IR/DataLayout.h" |
| 27 | #include "llvm/IR/Function.h" |
| 28 | #include "llvm/IR/IRBuilder.h" |
| 29 | #include "llvm/IR/IntrinsicInst.h" |
| 30 | #include "llvm/IR/Intrinsics.h" |
| 31 | #include "llvm/IR/Module.h" |
| 32 | #include "llvm/IR/PatternMatch.h" |
| 33 | #include "llvm/Support/Casting.h" |
| 34 | #include "llvm/Support/CommandLine.h" |
| 35 | #include "llvm/Support/KnownBits.h" |
| 36 | #include "llvm/Support/KnownFPClass.h" |
| 37 | #include "llvm/Support/MathExtras.h" |
| 38 | #include "llvm/TargetParser/Triple.h" |
| 39 | #include "llvm/Transforms/Utils/BuildLibCalls.h" |
| 40 | #include "llvm/Transforms/Utils/Local.h" |
| 41 | #include "llvm/Transforms/Utils/SizeOpts.h" |
| 42 | |
| 43 | #include <cmath> |
| 44 | |
| 45 | using namespace llvm; |
| 46 | using namespace PatternMatch; |
| 47 | |
| 48 | static cl::opt<bool> |
| 49 | EnableUnsafeFPShrink("enable-double-float-shrink" , cl::Hidden, |
| 50 | cl::init(Val: false), |
| 51 | cl::desc("Enable unsafe double to float " |
| 52 | "shrinking for math lib calls" )); |
| 53 | |
| 54 | // Enable conversion of operator new calls with a MemProf hot or cold hint |
| 55 | // to an operator new call that takes a hot/cold hint. Off by default since |
| 56 | // not all allocators currently support this extension. |
| 57 | static cl::opt<bool> |
| 58 | OptimizeHotColdNew("optimize-hot-cold-new" , cl::Hidden, cl::init(Val: false), |
| 59 | cl::desc("Enable hot/cold operator new library calls" )); |
| 60 | static cl::opt<bool> OptimizeExistingHotColdNew( |
| 61 | "optimize-existing-hot-cold-new" , cl::Hidden, cl::init(Val: false), |
| 62 | cl::desc( |
| 63 | "Enable optimization of existing hot/cold operator new library calls" )); |
| 64 | static cl::opt<bool> OptimizeNoBuiltinHotColdNew( |
| 65 | "optimize-nobuiltin-hot-cold-new-new" , cl::Hidden, cl::init(Val: false), |
| 66 | cl::desc("Enable transformation of nobuiltin operator new library calls" )); |
| 67 | |
| 68 | namespace { |
| 69 | |
| 70 | // Specialized parser to ensure the hint is an 8 bit value (we can't specify |
| 71 | // uint8_t to opt<> as that is interpreted to mean that we are passing a char |
| 72 | // option with a specific set of values. |
| 73 | struct HotColdHintParser : public cl::parser<unsigned> { |
| 74 | HotColdHintParser(cl::Option &O) : cl::parser<unsigned>(O) {} |
| 75 | |
| 76 | bool parse(cl::Option &O, StringRef ArgName, StringRef Arg, unsigned &Value) { |
| 77 | if (Arg.getAsInteger(Radix: 0, Result&: Value)) |
| 78 | return O.error(Message: "'" + Arg + "' value invalid for uint argument!" ); |
| 79 | |
| 80 | if (Value > 255) |
| 81 | return O.error(Message: "'" + Arg + "' value must be in the range [0, 255]!" ); |
| 82 | |
| 83 | return false; |
| 84 | } |
| 85 | }; |
| 86 | |
| 87 | } // end anonymous namespace |
| 88 | |
| 89 | // Hot/cold operator new takes an 8 bit hotness hint, where 0 is the coldest |
| 90 | // and 255 is the hottest. Default to 1 value away from the coldest and hottest |
| 91 | // hints, so that the compiler hinted allocations are slightly less strong than |
| 92 | // manually inserted hints at the two extremes. |
| 93 | static cl::opt<unsigned, false, HotColdHintParser> ColdNewHintValue( |
| 94 | "cold-new-hint-value" , cl::Hidden, cl::init(Val: 1), |
| 95 | cl::desc("Value to pass to hot/cold operator new for cold allocation" )); |
| 96 | static cl::opt<unsigned, false, HotColdHintParser> |
| 97 | NotColdNewHintValue("notcold-new-hint-value" , cl::Hidden, cl::init(Val: 128), |
| 98 | cl::desc("Value to pass to hot/cold operator new for " |
| 99 | "notcold (warm) allocation" )); |
| 100 | static cl::opt<unsigned, false, HotColdHintParser> HotNewHintValue( |
| 101 | "hot-new-hint-value" , cl::Hidden, cl::init(Val: 254), |
| 102 | cl::desc("Value to pass to hot/cold operator new for hot allocation" )); |
| 103 | static cl::opt<unsigned, false, HotColdHintParser> AmbiguousNewHintValue( |
| 104 | "ambiguous-new-hint-value" , cl::Hidden, cl::init(Val: 222), |
| 105 | cl::desc( |
| 106 | "Value to pass to hot/cold operator new for ambiguous allocation" )); |
| 107 | |
| 108 | //===----------------------------------------------------------------------===// |
| 109 | // Helper Functions |
| 110 | //===----------------------------------------------------------------------===// |
| 111 | |
| 112 | static bool ignoreCallingConv(LibFunc Func) { |
| 113 | return Func == LibFunc_abs || Func == LibFunc_labs || |
| 114 | Func == LibFunc_llabs || Func == LibFunc_strlen; |
| 115 | } |
| 116 | |
| 117 | /// Return true if it is only used in equality comparisons with With. |
| 118 | static bool isOnlyUsedInEqualityComparison(Value *V, Value *With) { |
| 119 | for (User *U : V->users()) { |
| 120 | if (ICmpInst *IC = dyn_cast<ICmpInst>(Val: U)) |
| 121 | if (IC->isEquality() && IC->getOperand(i_nocapture: 1) == With) |
| 122 | continue; |
| 123 | // Unknown instruction. |
| 124 | return false; |
| 125 | } |
| 126 | return true; |
| 127 | } |
| 128 | |
| 129 | static bool callHasFloatingPointArgument(const CallInst *CI) { |
| 130 | return any_of(Range: CI->operands(), P: [](const Use &OI) { |
| 131 | return OI->getType()->isFloatingPointTy(); |
| 132 | }); |
| 133 | } |
| 134 | |
| 135 | static bool callHasFP128Argument(const CallInst *CI) { |
| 136 | return any_of(Range: CI->operands(), P: [](const Use &OI) { |
| 137 | return OI->getType()->isFP128Ty(); |
| 138 | }); |
| 139 | } |
| 140 | |
| 141 | // Convert the entire string Str representing an integer in Base, up to |
| 142 | // the terminating nul if present, to a constant according to the rules |
| 143 | // of strtoul[l] or, when AsSigned is set, of strtol[l]. On success |
| 144 | // return the result, otherwise null. |
| 145 | // The function assumes the string is encoded in ASCII and carefully |
| 146 | // avoids converting sequences (including "") that the corresponding |
| 147 | // library call might fail and set errno for. |
| 148 | static Value *convertStrToInt(CallInst *CI, StringRef &Str, Value *EndPtr, |
| 149 | uint64_t Base, bool AsSigned, IRBuilderBase &B) { |
| 150 | if (Base < 2 || Base > 36) |
| 151 | if (Base != 0) |
| 152 | // Fail for an invalid base (required by POSIX). |
| 153 | return nullptr; |
| 154 | |
| 155 | // Current offset into the original string to reflect in EndPtr. |
| 156 | size_t Offset = 0; |
| 157 | // Strip leading whitespace. |
| 158 | for ( ; Offset != Str.size(); ++Offset) |
| 159 | if (!isSpace(C: (unsigned char)Str[Offset])) { |
| 160 | Str = Str.substr(Start: Offset); |
| 161 | break; |
| 162 | } |
| 163 | |
| 164 | if (Str.empty()) |
| 165 | // Fail for empty subject sequences (POSIX allows but doesn't require |
| 166 | // strtol[l]/strtoul[l] to fail with EINVAL). |
| 167 | return nullptr; |
| 168 | |
| 169 | // Strip but remember the sign. |
| 170 | bool Negate = Str[0] == '-'; |
| 171 | if (Str[0] == '-' || Str[0] == '+') { |
| 172 | Str = Str.drop_front(); |
| 173 | if (Str.empty()) |
| 174 | // Fail for a sign with nothing after it. |
| 175 | return nullptr; |
| 176 | ++Offset; |
| 177 | } |
| 178 | |
| 179 | // Set Max to the absolute value of the minimum (for signed), or |
| 180 | // to the maximum (for unsigned) value representable in the type. |
| 181 | Type *RetTy = CI->getType(); |
| 182 | unsigned NBits = RetTy->getPrimitiveSizeInBits(); |
| 183 | uint64_t Max = AsSigned && Negate ? 1 : 0; |
| 184 | Max += AsSigned ? maxIntN(N: NBits) : maxUIntN(N: NBits); |
| 185 | |
| 186 | // Autodetect Base if it's zero and consume the "0x" prefix. |
| 187 | if (Str.size() > 1) { |
| 188 | if (Str[0] == '0') { |
| 189 | if (toUpper(x: (unsigned char)Str[1]) == 'X') { |
| 190 | if (Str.size() == 2 || (Base && Base != 16)) |
| 191 | // Fail if Base doesn't allow the "0x" prefix or for the prefix |
| 192 | // alone that implementations like BSD set errno to EINVAL for. |
| 193 | return nullptr; |
| 194 | |
| 195 | Str = Str.drop_front(N: 2); |
| 196 | Offset += 2; |
| 197 | Base = 16; |
| 198 | } |
| 199 | else if (Base == 0) |
| 200 | Base = 8; |
| 201 | } else if (Base == 0) |
| 202 | Base = 10; |
| 203 | } |
| 204 | else if (Base == 0) |
| 205 | Base = 10; |
| 206 | |
| 207 | // Convert the rest of the subject sequence, not including the sign, |
| 208 | // to its uint64_t representation (this assumes the source character |
| 209 | // set is ASCII). |
| 210 | uint64_t Result = 0; |
| 211 | for (unsigned i = 0; i != Str.size(); ++i) { |
| 212 | unsigned char DigVal = Str[i]; |
| 213 | if (isDigit(C: DigVal)) |
| 214 | DigVal = DigVal - '0'; |
| 215 | else { |
| 216 | DigVal = toUpper(x: DigVal); |
| 217 | if (isAlpha(C: DigVal)) |
| 218 | DigVal = DigVal - 'A' + 10; |
| 219 | else |
| 220 | return nullptr; |
| 221 | } |
| 222 | |
| 223 | if (DigVal >= Base) |
| 224 | // Fail if the digit is not valid in the Base. |
| 225 | return nullptr; |
| 226 | |
| 227 | // Add the digit and fail if the result is not representable in |
| 228 | // the (unsigned form of the) destination type. |
| 229 | bool VFlow; |
| 230 | Result = SaturatingMultiplyAdd(X: Result, Y: Base, A: (uint64_t)DigVal, ResultOverflowed: &VFlow); |
| 231 | if (VFlow || Result > Max) |
| 232 | return nullptr; |
| 233 | } |
| 234 | |
| 235 | if (EndPtr) { |
| 236 | // Store the pointer to the end. |
| 237 | Value *Off = B.getInt64(C: Offset + Str.size()); |
| 238 | Value *StrBeg = CI->getArgOperand(i: 0); |
| 239 | Value *StrEnd = B.CreateInBoundsGEP(Ty: B.getInt8Ty(), Ptr: StrBeg, IdxList: Off, Name: "endptr" ); |
| 240 | B.CreateStore(Val: StrEnd, Ptr: EndPtr); |
| 241 | } |
| 242 | |
| 243 | if (Negate) { |
| 244 | // Unsigned negation doesn't overflow. |
| 245 | Result = -Result; |
| 246 | // For unsigned numbers, discard sign bits. |
| 247 | if (!AsSigned) |
| 248 | Result &= maxUIntN(N: NBits); |
| 249 | } |
| 250 | |
| 251 | return ConstantInt::get(Ty: RetTy, V: Result, IsSigned: AsSigned); |
| 252 | } |
| 253 | |
| 254 | static bool isOnlyUsedInComparisonWithZero(Value *V) { |
| 255 | for (User *U : V->users()) { |
| 256 | if (ICmpInst *IC = dyn_cast<ICmpInst>(Val: U)) |
| 257 | if (Constant *C = dyn_cast<Constant>(Val: IC->getOperand(i_nocapture: 1))) |
| 258 | if (C->isNullValue()) |
| 259 | continue; |
| 260 | // Unknown instruction. |
| 261 | return false; |
| 262 | } |
| 263 | return true; |
| 264 | } |
| 265 | |
| 266 | static bool canTransformToMemCmp(CallInst *CI, Value *Str, uint64_t Len, |
| 267 | const DataLayout &DL) { |
| 268 | if (!isOnlyUsedInComparisonWithZero(V: CI)) |
| 269 | return false; |
| 270 | |
| 271 | if (!isDereferenceableAndAlignedPointer(V: Str, Alignment: Align(1), Size: APInt(64, Len), DL)) |
| 272 | return false; |
| 273 | |
| 274 | if (CI->getFunction()->hasFnAttribute(Kind: Attribute::SanitizeMemory)) |
| 275 | return false; |
| 276 | |
| 277 | return true; |
| 278 | } |
| 279 | |
| 280 | static void annotateDereferenceableBytes(CallInst *CI, |
| 281 | ArrayRef<unsigned> ArgNos, |
| 282 | uint64_t DereferenceableBytes) { |
| 283 | const Function *F = CI->getCaller(); |
| 284 | if (!F) |
| 285 | return; |
| 286 | for (unsigned ArgNo : ArgNos) { |
| 287 | uint64_t DerefBytes = DereferenceableBytes; |
| 288 | unsigned AS = CI->getArgOperand(i: ArgNo)->getType()->getPointerAddressSpace(); |
| 289 | if (!llvm::NullPointerIsDefined(F, AS) || |
| 290 | CI->paramHasAttr(ArgNo, Kind: Attribute::NonNull)) |
| 291 | DerefBytes = std::max(a: CI->getParamDereferenceableOrNullBytes(i: ArgNo), |
| 292 | b: DereferenceableBytes); |
| 293 | |
| 294 | if (CI->getParamDereferenceableBytes(i: ArgNo) < DerefBytes) { |
| 295 | CI->removeParamAttr(ArgNo, Kind: Attribute::Dereferenceable); |
| 296 | if (!llvm::NullPointerIsDefined(F, AS) || |
| 297 | CI->paramHasAttr(ArgNo, Kind: Attribute::NonNull)) |
| 298 | CI->removeParamAttr(ArgNo, Kind: Attribute::DereferenceableOrNull); |
| 299 | CI->addParamAttr(ArgNo, Attr: Attribute::getWithDereferenceableBytes( |
| 300 | Context&: CI->getContext(), Bytes: DerefBytes)); |
| 301 | } |
| 302 | } |
| 303 | } |
| 304 | |
| 305 | static void annotateNonNullNoUndefBasedOnAccess(CallInst *CI, |
| 306 | ArrayRef<unsigned> ArgNos) { |
| 307 | Function *F = CI->getCaller(); |
| 308 | if (!F) |
| 309 | return; |
| 310 | |
| 311 | for (unsigned ArgNo : ArgNos) { |
| 312 | if (!CI->paramHasAttr(ArgNo, Kind: Attribute::NoUndef)) |
| 313 | CI->addParamAttr(ArgNo, Kind: Attribute::NoUndef); |
| 314 | |
| 315 | if (!CI->paramHasAttr(ArgNo, Kind: Attribute::NonNull)) { |
| 316 | unsigned AS = |
| 317 | CI->getArgOperand(i: ArgNo)->getType()->getPointerAddressSpace(); |
| 318 | if (llvm::NullPointerIsDefined(F, AS)) |
| 319 | continue; |
| 320 | CI->addParamAttr(ArgNo, Kind: Attribute::NonNull); |
| 321 | } |
| 322 | |
| 323 | annotateDereferenceableBytes(CI, ArgNos: ArgNo, DereferenceableBytes: 1); |
| 324 | } |
| 325 | } |
| 326 | |
| 327 | static void annotateNonNullAndDereferenceable(CallInst *CI, ArrayRef<unsigned> ArgNos, |
| 328 | Value *Size, const DataLayout &DL) { |
| 329 | if (ConstantInt *LenC = dyn_cast<ConstantInt>(Val: Size)) { |
| 330 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos); |
| 331 | annotateDereferenceableBytes(CI, ArgNos, DereferenceableBytes: LenC->getZExtValue()); |
| 332 | } else if (isKnownNonZero(V: Size, Q: DL)) { |
| 333 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos); |
| 334 | uint64_t X, Y; |
| 335 | uint64_t DerefMin = 1; |
| 336 | if (match(V: Size, P: m_Select(C: m_Value(), L: m_ConstantInt(V&: X), R: m_ConstantInt(V&: Y)))) { |
| 337 | DerefMin = std::min(a: X, b: Y); |
| 338 | annotateDereferenceableBytes(CI, ArgNos, DereferenceableBytes: DerefMin); |
| 339 | } |
| 340 | } |
| 341 | } |
| 342 | |
| 343 | // Copy CallInst "flags" like musttail, notail, and tail. Return New param for |
| 344 | // easier chaining. Calls to emit* and B.createCall should probably be wrapped |
| 345 | // in this function when New is created to replace Old. Callers should take |
| 346 | // care to check Old.isMustTailCall() if they aren't replacing Old directly |
| 347 | // with New. |
| 348 | static Value *copyFlags(const CallInst &Old, Value *New) { |
| 349 | assert(!Old.isMustTailCall() && "do not copy musttail call flags" ); |
| 350 | assert(!Old.isNoTailCall() && "do not copy notail call flags" ); |
| 351 | if (auto *NewCI = dyn_cast_or_null<CallInst>(Val: New)) |
| 352 | NewCI->setTailCallKind(Old.getTailCallKind()); |
| 353 | return New; |
| 354 | } |
| 355 | |
| 356 | static Value *mergeAttributesAndFlags(CallInst *NewCI, const CallInst &Old) { |
| 357 | NewCI->setAttributes(AttributeList::get( |
| 358 | C&: NewCI->getContext(), Attrs: {NewCI->getAttributes(), Old.getAttributes()})); |
| 359 | NewCI->removeRetAttrs(AttrsToRemove: AttributeFuncs::typeIncompatible( |
| 360 | Ty: NewCI->getType(), AS: NewCI->getRetAttributes())); |
| 361 | for (unsigned I = 0; I < NewCI->arg_size(); ++I) |
| 362 | NewCI->removeParamAttrs( |
| 363 | ArgNo: I, AttrsToRemove: AttributeFuncs::typeIncompatible(Ty: NewCI->getArgOperand(i: I)->getType(), |
| 364 | AS: NewCI->getParamAttributes(ArgNo: I))); |
| 365 | |
| 366 | return copyFlags(Old, New: NewCI); |
| 367 | } |
| 368 | |
| 369 | // Helper to avoid truncating the length if size_t is 32-bits. |
| 370 | static StringRef substr(StringRef Str, uint64_t Len) { |
| 371 | return Len >= Str.size() ? Str : Str.substr(Start: 0, N: Len); |
| 372 | } |
| 373 | |
| 374 | //===----------------------------------------------------------------------===// |
| 375 | // String and Memory Library Call Optimizations |
| 376 | //===----------------------------------------------------------------------===// |
| 377 | |
| 378 | Value *LibCallSimplifier::optimizeStrCat(CallInst *CI, IRBuilderBase &B) { |
| 379 | // Extract some information from the instruction |
| 380 | Value *Dst = CI->getArgOperand(i: 0); |
| 381 | Value *Src = CI->getArgOperand(i: 1); |
| 382 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: {0, 1}); |
| 383 | |
| 384 | // See if we can get the length of the input string. |
| 385 | uint64_t Len = GetStringLength(V: Src); |
| 386 | if (Len) |
| 387 | annotateDereferenceableBytes(CI, ArgNos: 1, DereferenceableBytes: Len); |
| 388 | else |
| 389 | return nullptr; |
| 390 | --Len; // Unbias length. |
| 391 | |
| 392 | // Handle the simple, do-nothing case: strcat(x, "") -> x |
| 393 | if (Len == 0) |
| 394 | return Dst; |
| 395 | |
| 396 | return copyFlags(Old: *CI, New: emitStrLenMemCpy(Src, Dst, Len, B)); |
| 397 | } |
| 398 | |
| 399 | Value *LibCallSimplifier::emitStrLenMemCpy(Value *Src, Value *Dst, uint64_t Len, |
| 400 | IRBuilderBase &B) { |
| 401 | // We need to find the end of the destination string. That's where the |
| 402 | // memory is to be moved to. We just generate a call to strlen. |
| 403 | Value *DstLen = emitStrLen(Ptr: Dst, B, DL, TLI); |
| 404 | if (!DstLen) |
| 405 | return nullptr; |
| 406 | |
| 407 | // Now that we have the destination's length, we must index into the |
| 408 | // destination's pointer to get the actual memcpy destination (end of |
| 409 | // the string .. we're concatenating). |
| 410 | Value *CpyDst = B.CreateInBoundsGEP(Ty: B.getInt8Ty(), Ptr: Dst, IdxList: DstLen, Name: "endptr" ); |
| 411 | |
| 412 | // We have enough information to now generate the memcpy call to do the |
| 413 | // concatenation for us. Make a memcpy to copy the nul byte with align = 1. |
| 414 | B.CreateMemCpy(Dst: CpyDst, DstAlign: Align(1), Src, SrcAlign: Align(1), |
| 415 | Size: TLI->getAsSizeT(V: Len + 1, M: *B.GetInsertBlock()->getModule())); |
| 416 | return Dst; |
| 417 | } |
| 418 | |
| 419 | Value *LibCallSimplifier::optimizeStrNCat(CallInst *CI, IRBuilderBase &B) { |
| 420 | // Extract some information from the instruction. |
| 421 | Value *Dst = CI->getArgOperand(i: 0); |
| 422 | Value *Src = CI->getArgOperand(i: 1); |
| 423 | Value *Size = CI->getArgOperand(i: 2); |
| 424 | uint64_t Len; |
| 425 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: 0); |
| 426 | if (isKnownNonZero(V: Size, Q: DL)) |
| 427 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: 1); |
| 428 | |
| 429 | // We don't do anything if length is not constant. |
| 430 | ConstantInt *LengthArg = dyn_cast<ConstantInt>(Val: Size); |
| 431 | if (LengthArg) { |
| 432 | Len = LengthArg->getZExtValue(); |
| 433 | // strncat(x, c, 0) -> x |
| 434 | if (!Len) |
| 435 | return Dst; |
| 436 | } else { |
| 437 | return nullptr; |
| 438 | } |
| 439 | |
| 440 | // See if we can get the length of the input string. |
| 441 | uint64_t SrcLen = GetStringLength(V: Src); |
| 442 | if (SrcLen) { |
| 443 | annotateDereferenceableBytes(CI, ArgNos: 1, DereferenceableBytes: SrcLen); |
| 444 | --SrcLen; // Unbias length. |
| 445 | } else { |
| 446 | return nullptr; |
| 447 | } |
| 448 | |
| 449 | // strncat(x, "", c) -> x |
| 450 | if (SrcLen == 0) |
| 451 | return Dst; |
| 452 | |
| 453 | // We don't optimize this case. |
| 454 | if (Len < SrcLen) |
| 455 | return nullptr; |
| 456 | |
| 457 | // strncat(x, s, c) -> strcat(x, s) |
| 458 | // s is constant so the strcat can be optimized further. |
| 459 | return copyFlags(Old: *CI, New: emitStrLenMemCpy(Src, Dst, Len: SrcLen, B)); |
| 460 | } |
| 461 | |
| 462 | // Helper to transform memchr(S, C, N) == S to N && *S == C and, when |
| 463 | // NBytes is null, strchr(S, C) to *S == C. A precondition of the function |
| 464 | // is that either S is dereferenceable or the value of N is nonzero. |
| 465 | static Value* memChrToCharCompare(CallInst *CI, Value *NBytes, |
| 466 | IRBuilderBase &B, const DataLayout &DL) |
| 467 | { |
| 468 | Value *Src = CI->getArgOperand(i: 0); |
| 469 | Value *CharVal = CI->getArgOperand(i: 1); |
| 470 | |
| 471 | // Fold memchr(A, C, N) == A to N && *A == C. |
| 472 | Type *CharTy = B.getInt8Ty(); |
| 473 | Value *Char0 = B.CreateLoad(Ty: CharTy, Ptr: Src); |
| 474 | CharVal = B.CreateTrunc(V: CharVal, DestTy: CharTy); |
| 475 | Value *Cmp = B.CreateICmpEQ(LHS: Char0, RHS: CharVal, Name: "char0cmp" ); |
| 476 | |
| 477 | if (NBytes) { |
| 478 | Value *Zero = ConstantInt::get(Ty: NBytes->getType(), V: 0); |
| 479 | Value *And = B.CreateICmpNE(LHS: NBytes, RHS: Zero); |
| 480 | Cmp = B.CreateLogicalAnd(Cond1: And, Cond2: Cmp); |
| 481 | } |
| 482 | |
| 483 | Value *NullPtr = Constant::getNullValue(Ty: CI->getType()); |
| 484 | return B.CreateSelect(C: Cmp, True: Src, False: NullPtr); |
| 485 | } |
| 486 | |
| 487 | Value *LibCallSimplifier::optimizeStrChr(CallInst *CI, IRBuilderBase &B) { |
| 488 | Value *SrcStr = CI->getArgOperand(i: 0); |
| 489 | Value *CharVal = CI->getArgOperand(i: 1); |
| 490 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: 0); |
| 491 | |
| 492 | if (isOnlyUsedInEqualityComparison(V: CI, With: SrcStr)) |
| 493 | return memChrToCharCompare(CI, NBytes: nullptr, B, DL); |
| 494 | |
| 495 | // If the second operand is non-constant, see if we can compute the length |
| 496 | // of the input string and turn this into memchr. |
| 497 | ConstantInt *CharC = dyn_cast<ConstantInt>(Val: CharVal); |
| 498 | if (!CharC) { |
| 499 | uint64_t Len = GetStringLength(V: SrcStr); |
| 500 | if (Len) |
| 501 | annotateDereferenceableBytes(CI, ArgNos: 0, DereferenceableBytes: Len); |
| 502 | else |
| 503 | return nullptr; |
| 504 | |
| 505 | Function *Callee = CI->getCalledFunction(); |
| 506 | FunctionType *FT = Callee->getFunctionType(); |
| 507 | unsigned IntBits = TLI->getIntSize(); |
| 508 | if (!FT->getParamType(i: 1)->isIntegerTy(Bitwidth: IntBits)) // memchr needs 'int'. |
| 509 | return nullptr; |
| 510 | |
| 511 | unsigned SizeTBits = TLI->getSizeTSize(M: *CI->getModule()); |
| 512 | Type *SizeTTy = IntegerType::get(C&: CI->getContext(), NumBits: SizeTBits); |
| 513 | return copyFlags(Old: *CI, |
| 514 | New: emitMemChr(Ptr: SrcStr, Val: CharVal, // include nul. |
| 515 | Len: ConstantInt::get(Ty: SizeTTy, V: Len), B, |
| 516 | DL, TLI)); |
| 517 | } |
| 518 | |
| 519 | if (CharC->isZero()) { |
| 520 | Value *NullPtr = Constant::getNullValue(Ty: CI->getType()); |
| 521 | if (isOnlyUsedInEqualityComparison(V: CI, With: NullPtr)) |
| 522 | // Pre-empt the transformation to strlen below and fold |
| 523 | // strchr(A, '\0') == null to false. |
| 524 | return B.CreateIntToPtr(V: B.getTrue(), DestTy: CI->getType()); |
| 525 | } |
| 526 | |
| 527 | // Otherwise, the character is a constant, see if the first argument is |
| 528 | // a string literal. If so, we can constant fold. |
| 529 | StringRef Str; |
| 530 | if (!getConstantStringInfo(V: SrcStr, Str)) { |
| 531 | if (CharC->isZero()) // strchr(p, 0) -> p + strlen(p) |
| 532 | if (Value *StrLen = emitStrLen(Ptr: SrcStr, B, DL, TLI)) |
| 533 | return B.CreateInBoundsGEP(Ty: B.getInt8Ty(), Ptr: SrcStr, IdxList: StrLen, Name: "strchr" ); |
| 534 | return nullptr; |
| 535 | } |
| 536 | |
| 537 | // Compute the offset, make sure to handle the case when we're searching for |
| 538 | // zero (a weird way to spell strlen). |
| 539 | size_t I = (0xFF & CharC->getSExtValue()) == 0 |
| 540 | ? Str.size() |
| 541 | : Str.find(C: CharC->getSExtValue()); |
| 542 | if (I == StringRef::npos) // Didn't find the char. strchr returns null. |
| 543 | return Constant::getNullValue(Ty: CI->getType()); |
| 544 | |
| 545 | // strchr(s+n,c) -> gep(s+n+i,c) |
| 546 | return B.CreateInBoundsGEP(Ty: B.getInt8Ty(), Ptr: SrcStr, IdxList: B.getInt64(C: I), Name: "strchr" ); |
| 547 | } |
| 548 | |
| 549 | Value *LibCallSimplifier::optimizeStrRChr(CallInst *CI, IRBuilderBase &B) { |
| 550 | Value *SrcStr = CI->getArgOperand(i: 0); |
| 551 | Value *CharVal = CI->getArgOperand(i: 1); |
| 552 | ConstantInt *CharC = dyn_cast<ConstantInt>(Val: CharVal); |
| 553 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: 0); |
| 554 | |
| 555 | StringRef Str; |
| 556 | if (!getConstantStringInfo(V: SrcStr, Str)) { |
| 557 | // strrchr(s, 0) -> strchr(s, 0) |
| 558 | if (CharC && CharC->isZero()) |
| 559 | return copyFlags(Old: *CI, New: emitStrChr(Ptr: SrcStr, C: '\0', B, TLI)); |
| 560 | return nullptr; |
| 561 | } |
| 562 | |
| 563 | unsigned SizeTBits = TLI->getSizeTSize(M: *CI->getModule()); |
| 564 | Type *SizeTTy = IntegerType::get(C&: CI->getContext(), NumBits: SizeTBits); |
| 565 | |
| 566 | // Try to expand strrchr to the memrchr nonstandard extension if it's |
| 567 | // available, or simply fail otherwise. |
| 568 | uint64_t NBytes = Str.size() + 1; // Include the terminating nul. |
| 569 | Value *Size = ConstantInt::get(Ty: SizeTTy, V: NBytes); |
| 570 | return copyFlags(Old: *CI, New: emitMemRChr(Ptr: SrcStr, Val: CharVal, Len: Size, B, DL, TLI)); |
| 571 | } |
| 572 | |
| 573 | Value *LibCallSimplifier::optimizeStrCmp(CallInst *CI, IRBuilderBase &B) { |
| 574 | Value *Str1P = CI->getArgOperand(i: 0), *Str2P = CI->getArgOperand(i: 1); |
| 575 | if (Str1P == Str2P) // strcmp(x,x) -> 0 |
| 576 | return ConstantInt::get(Ty: CI->getType(), V: 0); |
| 577 | |
| 578 | StringRef Str1, Str2; |
| 579 | bool HasStr1 = getConstantStringInfo(V: Str1P, Str&: Str1); |
| 580 | bool HasStr2 = getConstantStringInfo(V: Str2P, Str&: Str2); |
| 581 | |
| 582 | // strcmp(x, y) -> cnst (if both x and y are constant strings) |
| 583 | if (HasStr1 && HasStr2) |
| 584 | return ConstantInt::getSigned(Ty: CI->getType(), |
| 585 | V: std::clamp(val: Str1.compare(RHS: Str2), lo: -1, hi: 1)); |
| 586 | |
| 587 | if (HasStr1 && Str1.empty()) // strcmp("", x) -> -*x |
| 588 | return B.CreateNeg(V: B.CreateZExt( |
| 589 | V: B.CreateLoad(Ty: B.getInt8Ty(), Ptr: Str2P, Name: "strcmpload" ), DestTy: CI->getType())); |
| 590 | |
| 591 | if (HasStr2 && Str2.empty()) // strcmp(x,"") -> *x |
| 592 | return B.CreateZExt(V: B.CreateLoad(Ty: B.getInt8Ty(), Ptr: Str1P, Name: "strcmpload" ), |
| 593 | DestTy: CI->getType()); |
| 594 | |
| 595 | // strcmp(P, "x") -> memcmp(P, "x", 2) |
| 596 | uint64_t Len1 = GetStringLength(V: Str1P); |
| 597 | if (Len1) |
| 598 | annotateDereferenceableBytes(CI, ArgNos: 0, DereferenceableBytes: Len1); |
| 599 | uint64_t Len2 = GetStringLength(V: Str2P); |
| 600 | if (Len2) |
| 601 | annotateDereferenceableBytes(CI, ArgNos: 1, DereferenceableBytes: Len2); |
| 602 | |
| 603 | if (Len1 && Len2) { |
| 604 | return copyFlags( |
| 605 | Old: *CI, New: emitMemCmp(Ptr1: Str1P, Ptr2: Str2P, |
| 606 | Len: TLI->getAsSizeT(V: std::min(a: Len1, b: Len2), M: *CI->getModule()), |
| 607 | B, DL, TLI)); |
| 608 | } |
| 609 | |
| 610 | // strcmp to memcmp |
| 611 | if (!HasStr1 && HasStr2) { |
| 612 | if (canTransformToMemCmp(CI, Str: Str1P, Len: Len2, DL)) |
| 613 | return copyFlags(Old: *CI, New: emitMemCmp(Ptr1: Str1P, Ptr2: Str2P, |
| 614 | Len: TLI->getAsSizeT(V: Len2, M: *CI->getModule()), |
| 615 | B, DL, TLI)); |
| 616 | } else if (HasStr1 && !HasStr2) { |
| 617 | if (canTransformToMemCmp(CI, Str: Str2P, Len: Len1, DL)) |
| 618 | return copyFlags(Old: *CI, New: emitMemCmp(Ptr1: Str1P, Ptr2: Str2P, |
| 619 | Len: TLI->getAsSizeT(V: Len1, M: *CI->getModule()), |
| 620 | B, DL, TLI)); |
| 621 | } |
| 622 | |
| 623 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: {0, 1}); |
| 624 | return nullptr; |
| 625 | } |
| 626 | |
| 627 | // Optimize a memcmp or, when StrNCmp is true, strncmp call CI with constant |
| 628 | // arrays LHS and RHS and nonconstant Size. |
| 629 | static Value *optimizeMemCmpVarSize(CallInst *CI, Value *LHS, Value *RHS, |
| 630 | Value *Size, bool StrNCmp, |
| 631 | IRBuilderBase &B, const DataLayout &DL); |
| 632 | |
| 633 | Value *LibCallSimplifier::optimizeStrNCmp(CallInst *CI, IRBuilderBase &B) { |
| 634 | Value *Str1P = CI->getArgOperand(i: 0); |
| 635 | Value *Str2P = CI->getArgOperand(i: 1); |
| 636 | Value *Size = CI->getArgOperand(i: 2); |
| 637 | if (Str1P == Str2P) // strncmp(x,x,n) -> 0 |
| 638 | return ConstantInt::get(Ty: CI->getType(), V: 0); |
| 639 | |
| 640 | if (isKnownNonZero(V: Size, Q: DL)) |
| 641 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: {0, 1}); |
| 642 | // Get the length argument if it is constant. |
| 643 | uint64_t Length; |
| 644 | if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(Val: Size)) |
| 645 | Length = LengthArg->getZExtValue(); |
| 646 | else |
| 647 | return optimizeMemCmpVarSize(CI, LHS: Str1P, RHS: Str2P, Size, StrNCmp: true, B, DL); |
| 648 | |
| 649 | if (Length == 0) // strncmp(x,y,0) -> 0 |
| 650 | return ConstantInt::get(Ty: CI->getType(), V: 0); |
| 651 | |
| 652 | if (Length == 1) // strncmp(x,y,1) -> memcmp(x,y,1) |
| 653 | return copyFlags(Old: *CI, New: emitMemCmp(Ptr1: Str1P, Ptr2: Str2P, Len: Size, B, DL, TLI)); |
| 654 | |
| 655 | StringRef Str1, Str2; |
| 656 | bool HasStr1 = getConstantStringInfo(V: Str1P, Str&: Str1); |
| 657 | bool HasStr2 = getConstantStringInfo(V: Str2P, Str&: Str2); |
| 658 | |
| 659 | // strncmp(x, y) -> cnst (if both x and y are constant strings) |
| 660 | if (HasStr1 && HasStr2) { |
| 661 | // Avoid truncating the 64-bit Length to 32 bits in ILP32. |
| 662 | StringRef SubStr1 = substr(Str: Str1, Len: Length); |
| 663 | StringRef SubStr2 = substr(Str: Str2, Len: Length); |
| 664 | return ConstantInt::getSigned(Ty: CI->getType(), |
| 665 | V: std::clamp(val: SubStr1.compare(RHS: SubStr2), lo: -1, hi: 1)); |
| 666 | } |
| 667 | |
| 668 | if (HasStr1 && Str1.empty()) // strncmp("", x, n) -> -*x |
| 669 | return B.CreateNeg(V: B.CreateZExt( |
| 670 | V: B.CreateLoad(Ty: B.getInt8Ty(), Ptr: Str2P, Name: "strcmpload" ), DestTy: CI->getType())); |
| 671 | |
| 672 | if (HasStr2 && Str2.empty()) // strncmp(x, "", n) -> *x |
| 673 | return B.CreateZExt(V: B.CreateLoad(Ty: B.getInt8Ty(), Ptr: Str1P, Name: "strcmpload" ), |
| 674 | DestTy: CI->getType()); |
| 675 | |
| 676 | uint64_t Len1 = GetStringLength(V: Str1P); |
| 677 | if (Len1) |
| 678 | annotateDereferenceableBytes(CI, ArgNos: 0, DereferenceableBytes: Len1); |
| 679 | uint64_t Len2 = GetStringLength(V: Str2P); |
| 680 | if (Len2) |
| 681 | annotateDereferenceableBytes(CI, ArgNos: 1, DereferenceableBytes: Len2); |
| 682 | |
| 683 | // strncmp to memcmp |
| 684 | if (!HasStr1 && HasStr2) { |
| 685 | Len2 = std::min(a: Len2, b: Length); |
| 686 | if (canTransformToMemCmp(CI, Str: Str1P, Len: Len2, DL)) |
| 687 | return copyFlags(Old: *CI, New: emitMemCmp(Ptr1: Str1P, Ptr2: Str2P, |
| 688 | Len: TLI->getAsSizeT(V: Len2, M: *CI->getModule()), |
| 689 | B, DL, TLI)); |
| 690 | } else if (HasStr1 && !HasStr2) { |
| 691 | Len1 = std::min(a: Len1, b: Length); |
| 692 | if (canTransformToMemCmp(CI, Str: Str2P, Len: Len1, DL)) |
| 693 | return copyFlags(Old: *CI, New: emitMemCmp(Ptr1: Str1P, Ptr2: Str2P, |
| 694 | Len: TLI->getAsSizeT(V: Len1, M: *CI->getModule()), |
| 695 | B, DL, TLI)); |
| 696 | } |
| 697 | |
| 698 | return nullptr; |
| 699 | } |
| 700 | |
| 701 | Value *LibCallSimplifier::optimizeStrNDup(CallInst *CI, IRBuilderBase &B) { |
| 702 | Value *Src = CI->getArgOperand(i: 0); |
| 703 | ConstantInt *Size = dyn_cast<ConstantInt>(Val: CI->getArgOperand(i: 1)); |
| 704 | uint64_t SrcLen = GetStringLength(V: Src); |
| 705 | if (SrcLen && Size) { |
| 706 | annotateDereferenceableBytes(CI, ArgNos: 0, DereferenceableBytes: SrcLen); |
| 707 | if (SrcLen <= Size->getZExtValue() + 1) |
| 708 | return copyFlags(Old: *CI, New: emitStrDup(Ptr: Src, B, TLI)); |
| 709 | } |
| 710 | |
| 711 | return nullptr; |
| 712 | } |
| 713 | |
| 714 | Value *LibCallSimplifier::optimizeStrCpy(CallInst *CI, IRBuilderBase &B) { |
| 715 | Value *Dst = CI->getArgOperand(i: 0), *Src = CI->getArgOperand(i: 1); |
| 716 | if (Dst == Src) // strcpy(x,x) -> x |
| 717 | return Src; |
| 718 | |
| 719 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: {0, 1}); |
| 720 | // See if we can get the length of the input string. |
| 721 | uint64_t Len = GetStringLength(V: Src); |
| 722 | if (Len) |
| 723 | annotateDereferenceableBytes(CI, ArgNos: 1, DereferenceableBytes: Len); |
| 724 | else |
| 725 | return nullptr; |
| 726 | |
| 727 | // We have enough information to now generate the memcpy call to do the |
| 728 | // copy for us. Make a memcpy to copy the nul byte with align = 1. |
| 729 | CallInst *NewCI = B.CreateMemCpy(Dst, DstAlign: Align(1), Src, SrcAlign: Align(1), |
| 730 | Size: TLI->getAsSizeT(V: Len, M: *CI->getModule())); |
| 731 | mergeAttributesAndFlags(NewCI, Old: *CI); |
| 732 | return Dst; |
| 733 | } |
| 734 | |
| 735 | Value *LibCallSimplifier::optimizeStpCpy(CallInst *CI, IRBuilderBase &B) { |
| 736 | Value *Dst = CI->getArgOperand(i: 0), *Src = CI->getArgOperand(i: 1); |
| 737 | |
| 738 | // stpcpy(d,s) -> strcpy(d,s) if the result is not used. |
| 739 | if (CI->use_empty()) |
| 740 | return copyFlags(Old: *CI, New: emitStrCpy(Dst, Src, B, TLI)); |
| 741 | |
| 742 | if (Dst == Src) { // stpcpy(x,x) -> x+strlen(x) |
| 743 | Value *StrLen = emitStrLen(Ptr: Src, B, DL, TLI); |
| 744 | return StrLen ? B.CreateInBoundsGEP(Ty: B.getInt8Ty(), Ptr: Dst, IdxList: StrLen) : nullptr; |
| 745 | } |
| 746 | |
| 747 | // See if we can get the length of the input string. |
| 748 | uint64_t Len = GetStringLength(V: Src); |
| 749 | if (Len) |
| 750 | annotateDereferenceableBytes(CI, ArgNos: 1, DereferenceableBytes: Len); |
| 751 | else |
| 752 | return nullptr; |
| 753 | |
| 754 | Value *LenV = TLI->getAsSizeT(V: Len, M: *CI->getModule()); |
| 755 | Value *DstEnd = B.CreateInBoundsGEP( |
| 756 | Ty: B.getInt8Ty(), Ptr: Dst, IdxList: TLI->getAsSizeT(V: Len - 1, M: *CI->getModule())); |
| 757 | |
| 758 | // We have enough information to now generate the memcpy call to do the |
| 759 | // copy for us. Make a memcpy to copy the nul byte with align = 1. |
| 760 | CallInst *NewCI = B.CreateMemCpy(Dst, DstAlign: Align(1), Src, SrcAlign: Align(1), Size: LenV); |
| 761 | mergeAttributesAndFlags(NewCI, Old: *CI); |
| 762 | return DstEnd; |
| 763 | } |
| 764 | |
| 765 | // Optimize a call to size_t strlcpy(char*, const char*, size_t). |
| 766 | |
| 767 | Value *LibCallSimplifier::optimizeStrLCpy(CallInst *CI, IRBuilderBase &B) { |
| 768 | Value *Size = CI->getArgOperand(i: 2); |
| 769 | if (isKnownNonZero(V: Size, Q: DL)) |
| 770 | // Like snprintf, the function stores into the destination only when |
| 771 | // the size argument is nonzero. |
| 772 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: 0); |
| 773 | // The function reads the source argument regardless of Size (it returns |
| 774 | // its length). |
| 775 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: 1); |
| 776 | |
| 777 | uint64_t NBytes; |
| 778 | if (ConstantInt *SizeC = dyn_cast<ConstantInt>(Val: Size)) |
| 779 | NBytes = SizeC->getZExtValue(); |
| 780 | else |
| 781 | return nullptr; |
| 782 | |
| 783 | Value *Dst = CI->getArgOperand(i: 0); |
| 784 | Value *Src = CI->getArgOperand(i: 1); |
| 785 | if (NBytes <= 1) { |
| 786 | if (NBytes == 1) |
| 787 | // For a call to strlcpy(D, S, 1) first store a nul in *D. |
| 788 | B.CreateStore(Val: B.getInt8(C: 0), Ptr: Dst); |
| 789 | |
| 790 | // Transform strlcpy(D, S, 0) to a call to strlen(S). |
| 791 | return copyFlags(Old: *CI, New: emitStrLen(Ptr: Src, B, DL, TLI)); |
| 792 | } |
| 793 | |
| 794 | // Try to determine the length of the source, substituting its size |
| 795 | // when it's not nul-terminated (as it's required to be) to avoid |
| 796 | // reading past its end. |
| 797 | StringRef Str; |
| 798 | if (!getConstantStringInfo(V: Src, Str, /*TrimAtNul=*/false)) |
| 799 | return nullptr; |
| 800 | |
| 801 | uint64_t SrcLen = Str.find(C: '\0'); |
| 802 | // Set if the terminating nul should be copied by the call to memcpy |
| 803 | // below. |
| 804 | bool NulTerm = SrcLen < NBytes; |
| 805 | |
| 806 | if (NulTerm) |
| 807 | // Overwrite NBytes with the number of bytes to copy, including |
| 808 | // the terminating nul. |
| 809 | NBytes = SrcLen + 1; |
| 810 | else { |
| 811 | // Set the length of the source for the function to return to its |
| 812 | // size, and cap NBytes at the same. |
| 813 | SrcLen = std::min(a: SrcLen, b: uint64_t(Str.size())); |
| 814 | NBytes = std::min(a: NBytes - 1, b: SrcLen); |
| 815 | } |
| 816 | |
| 817 | if (SrcLen == 0) { |
| 818 | // Transform strlcpy(D, "", N) to (*D = '\0, 0). |
| 819 | B.CreateStore(Val: B.getInt8(C: 0), Ptr: Dst); |
| 820 | return ConstantInt::get(Ty: CI->getType(), V: 0); |
| 821 | } |
| 822 | |
| 823 | // Transform strlcpy(D, S, N) to memcpy(D, S, N') where N' is the lower |
| 824 | // bound on strlen(S) + 1 and N, optionally followed by a nul store to |
| 825 | // D[N' - 1] if necessary. |
| 826 | CallInst *NewCI = B.CreateMemCpy(Dst, DstAlign: Align(1), Src, SrcAlign: Align(1), |
| 827 | Size: TLI->getAsSizeT(V: NBytes, M: *CI->getModule())); |
| 828 | mergeAttributesAndFlags(NewCI, Old: *CI); |
| 829 | |
| 830 | if (!NulTerm) { |
| 831 | Value *EndOff = ConstantInt::get(Ty: CI->getType(), V: NBytes); |
| 832 | Value *EndPtr = B.CreateInBoundsGEP(Ty: B.getInt8Ty(), Ptr: Dst, IdxList: EndOff); |
| 833 | B.CreateStore(Val: B.getInt8(C: 0), Ptr: EndPtr); |
| 834 | } |
| 835 | |
| 836 | // Like snprintf, strlcpy returns the number of nonzero bytes that would |
| 837 | // have been copied if the bound had been sufficiently big (which in this |
| 838 | // case is strlen(Src)). |
| 839 | return ConstantInt::get(Ty: CI->getType(), V: SrcLen); |
| 840 | } |
| 841 | |
| 842 | // Optimize a call CI to either stpncpy when RetEnd is true, or to strncpy |
| 843 | // otherwise. |
| 844 | Value *LibCallSimplifier::optimizeStringNCpy(CallInst *CI, bool RetEnd, |
| 845 | IRBuilderBase &B) { |
| 846 | Value *Dst = CI->getArgOperand(i: 0); |
| 847 | Value *Src = CI->getArgOperand(i: 1); |
| 848 | Value *Size = CI->getArgOperand(i: 2); |
| 849 | |
| 850 | if (isKnownNonZero(V: Size, Q: DL)) { |
| 851 | // Both st{p,r}ncpy(D, S, N) access the source and destination arrays |
| 852 | // only when N is nonzero. |
| 853 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: 0); |
| 854 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: 1); |
| 855 | } |
| 856 | |
| 857 | // If the "bound" argument is known set N to it. Otherwise set it to |
| 858 | // UINT64_MAX and handle it later. |
| 859 | uint64_t N = UINT64_MAX; |
| 860 | if (ConstantInt *SizeC = dyn_cast<ConstantInt>(Val: Size)) |
| 861 | N = SizeC->getZExtValue(); |
| 862 | |
| 863 | if (N == 0) |
| 864 | // Fold st{p,r}ncpy(D, S, 0) to D. |
| 865 | return Dst; |
| 866 | |
| 867 | if (N == 1) { |
| 868 | Type *CharTy = B.getInt8Ty(); |
| 869 | Value *CharVal = B.CreateLoad(Ty: CharTy, Ptr: Src, Name: "stxncpy.char0" ); |
| 870 | B.CreateStore(Val: CharVal, Ptr: Dst); |
| 871 | if (!RetEnd) |
| 872 | // Transform strncpy(D, S, 1) to return (*D = *S), D. |
| 873 | return Dst; |
| 874 | |
| 875 | // Transform stpncpy(D, S, 1) to return (*D = *S) ? D + 1 : D. |
| 876 | Value *ZeroChar = ConstantInt::get(Ty: CharTy, V: 0); |
| 877 | Value *Cmp = B.CreateICmpEQ(LHS: CharVal, RHS: ZeroChar, Name: "stpncpy.char0cmp" ); |
| 878 | |
| 879 | Value *Off1 = B.getInt32(C: 1); |
| 880 | Value *EndPtr = B.CreateInBoundsGEP(Ty: CharTy, Ptr: Dst, IdxList: Off1, Name: "stpncpy.end" ); |
| 881 | return B.CreateSelect(C: Cmp, True: Dst, False: EndPtr, Name: "stpncpy.sel" ); |
| 882 | } |
| 883 | |
| 884 | // If the length of the input string is known set SrcLen to it. |
| 885 | uint64_t SrcLen = GetStringLength(V: Src); |
| 886 | if (SrcLen) |
| 887 | annotateDereferenceableBytes(CI, ArgNos: 1, DereferenceableBytes: SrcLen); |
| 888 | else |
| 889 | return nullptr; |
| 890 | |
| 891 | --SrcLen; // Unbias length. |
| 892 | |
| 893 | if (SrcLen == 0) { |
| 894 | // Transform st{p,r}ncpy(D, "", N) to memset(D, '\0', N) for any N. |
| 895 | Align MemSetAlign = |
| 896 | CI->getAttributes().getParamAttrs(ArgNo: 0).getAlignment().valueOrOne(); |
| 897 | CallInst *NewCI = B.CreateMemSet(Ptr: Dst, Val: B.getInt8(C: '\0'), Size, Align: MemSetAlign); |
| 898 | AttrBuilder ArgAttrs(CI->getContext(), CI->getAttributes().getParamAttrs(ArgNo: 0)); |
| 899 | NewCI->setAttributes(NewCI->getAttributes().addParamAttributes( |
| 900 | C&: CI->getContext(), ArgNo: 0, B: ArgAttrs)); |
| 901 | copyFlags(Old: *CI, New: NewCI); |
| 902 | return Dst; |
| 903 | } |
| 904 | |
| 905 | if (N > SrcLen + 1) { |
| 906 | if (N > 128) |
| 907 | // Bail if N is large or unknown. |
| 908 | return nullptr; |
| 909 | |
| 910 | // st{p,r}ncpy(D, "a", N) -> memcpy(D, "a\0\0\0", N) for N <= 128. |
| 911 | StringRef Str; |
| 912 | if (!getConstantStringInfo(V: Src, Str)) |
| 913 | return nullptr; |
| 914 | std::string SrcStr = Str.str(); |
| 915 | // Create a bigger, nul-padded array with the same length, SrcLen, |
| 916 | // as the original string. |
| 917 | SrcStr.resize(n: N, c: '\0'); |
| 918 | Src = B.CreateGlobalString(Str: SrcStr, Name: "str" , /*AddressSpace=*/0, |
| 919 | /*M=*/nullptr, /*AddNull=*/false); |
| 920 | } |
| 921 | |
| 922 | // st{p,r}ncpy(D, S, N) -> memcpy(align 1 D, align 1 S, N) when both |
| 923 | // S and N are constant. |
| 924 | CallInst *NewCI = B.CreateMemCpy(Dst, DstAlign: Align(1), Src, SrcAlign: Align(1), |
| 925 | Size: TLI->getAsSizeT(V: N, M: *CI->getModule())); |
| 926 | mergeAttributesAndFlags(NewCI, Old: *CI); |
| 927 | if (!RetEnd) |
| 928 | return Dst; |
| 929 | |
| 930 | // stpncpy(D, S, N) returns the address of the first null in D if it writes |
| 931 | // one, otherwise D + N. |
| 932 | Value *Off = B.getInt64(C: std::min(a: SrcLen, b: N)); |
| 933 | return B.CreateInBoundsGEP(Ty: B.getInt8Ty(), Ptr: Dst, IdxList: Off, Name: "endptr" ); |
| 934 | } |
| 935 | |
| 936 | Value *LibCallSimplifier::optimizeStringLength(CallInst *CI, IRBuilderBase &B, |
| 937 | unsigned CharSize, |
| 938 | Value *Bound) { |
| 939 | Value *Src = CI->getArgOperand(i: 0); |
| 940 | Type *CharTy = B.getIntNTy(N: CharSize); |
| 941 | |
| 942 | if (isOnlyUsedInZeroEqualityComparison(CxtI: CI) && |
| 943 | (!Bound || isKnownNonZero(V: Bound, Q: DL))) { |
| 944 | // Fold strlen: |
| 945 | // strlen(x) != 0 --> *x != 0 |
| 946 | // strlen(x) == 0 --> *x == 0 |
| 947 | // and likewise strnlen with constant N > 0: |
| 948 | // strnlen(x, N) != 0 --> *x != 0 |
| 949 | // strnlen(x, N) == 0 --> *x == 0 |
| 950 | return B.CreateZExt(V: B.CreateLoad(Ty: CharTy, Ptr: Src, Name: "char0" ), |
| 951 | DestTy: CI->getType()); |
| 952 | } |
| 953 | |
| 954 | if (Bound) { |
| 955 | if (ConstantInt *BoundCst = dyn_cast<ConstantInt>(Val: Bound)) { |
| 956 | if (BoundCst->isZero()) |
| 957 | // Fold strnlen(s, 0) -> 0 for any s, constant or otherwise. |
| 958 | return ConstantInt::get(Ty: CI->getType(), V: 0); |
| 959 | |
| 960 | if (BoundCst->isOne()) { |
| 961 | // Fold strnlen(s, 1) -> *s ? 1 : 0 for any s. |
| 962 | Value *CharVal = B.CreateLoad(Ty: CharTy, Ptr: Src, Name: "strnlen.char0" ); |
| 963 | Value *ZeroChar = ConstantInt::get(Ty: CharTy, V: 0); |
| 964 | Value *Cmp = B.CreateICmpNE(LHS: CharVal, RHS: ZeroChar, Name: "strnlen.char0cmp" ); |
| 965 | return B.CreateZExt(V: Cmp, DestTy: CI->getType()); |
| 966 | } |
| 967 | } |
| 968 | } |
| 969 | |
| 970 | if (uint64_t Len = GetStringLength(V: Src, CharSize)) { |
| 971 | Value *LenC = ConstantInt::get(Ty: CI->getType(), V: Len - 1); |
| 972 | // Fold strlen("xyz") -> 3 and strnlen("xyz", 2) -> 2 |
| 973 | // and strnlen("xyz", Bound) -> min(3, Bound) for nonconstant Bound. |
| 974 | if (Bound) |
| 975 | return B.CreateBinaryIntrinsic(ID: Intrinsic::umin, LHS: LenC, RHS: Bound); |
| 976 | return LenC; |
| 977 | } |
| 978 | |
| 979 | if (Bound) |
| 980 | // Punt for strnlen for now. |
| 981 | return nullptr; |
| 982 | |
| 983 | // If s is a constant pointer pointing to a string literal, we can fold |
| 984 | // strlen(s + x) to strlen(s) - x, when x is known to be in the range |
| 985 | // [0, strlen(s)] or the string has a single null terminator '\0' at the end. |
| 986 | // We only try to simplify strlen when the pointer s points to an array |
| 987 | // of CharSize elements. Otherwise, we would need to scale the offset x before |
| 988 | // doing the subtraction. This will make the optimization more complex, and |
| 989 | // it's not very useful because calling strlen for a pointer of other types is |
| 990 | // very uncommon. |
| 991 | if (GEPOperator *GEP = dyn_cast<GEPOperator>(Val: Src)) { |
| 992 | unsigned BW = DL.getIndexTypeSizeInBits(Ty: GEP->getType()); |
| 993 | SmallMapVector<Value *, APInt, 4> VarOffsets; |
| 994 | APInt ConstOffset(BW, 0); |
| 995 | assert(CharSize % 8 == 0 && "Expected a multiple of 8 sized CharSize" ); |
| 996 | // Check the gep is a single variable offset. |
| 997 | if (!GEP->collectOffset(DL, BitWidth: BW, VariableOffsets&: VarOffsets, ConstantOffset&: ConstOffset) || |
| 998 | VarOffsets.size() != 1 || ConstOffset != 0 || |
| 999 | VarOffsets.begin()->second != CharSize / 8) |
| 1000 | return nullptr; |
| 1001 | |
| 1002 | ConstantDataArraySlice Slice; |
| 1003 | if (getConstantDataArrayInfo(V: GEP->getOperand(i_nocapture: 0), Slice, ElementSize: CharSize)) { |
| 1004 | uint64_t NullTermIdx; |
| 1005 | if (Slice.Array == nullptr) { |
| 1006 | NullTermIdx = 0; |
| 1007 | } else { |
| 1008 | NullTermIdx = ~((uint64_t)0); |
| 1009 | for (uint64_t I = 0, E = Slice.Length; I < E; ++I) { |
| 1010 | if (Slice.Array->getElementAsInteger(i: I + Slice.Offset) == 0) { |
| 1011 | NullTermIdx = I; |
| 1012 | break; |
| 1013 | } |
| 1014 | } |
| 1015 | // If the string does not have '\0', leave it to strlen to compute |
| 1016 | // its length. |
| 1017 | if (NullTermIdx == ~((uint64_t)0)) |
| 1018 | return nullptr; |
| 1019 | } |
| 1020 | |
| 1021 | Value *Offset = VarOffsets.begin()->first; |
| 1022 | KnownBits Known = computeKnownBits(V: Offset, DL, AC: nullptr, CxtI: CI, DT: nullptr); |
| 1023 | |
| 1024 | // If Offset is not provably in the range [0, NullTermIdx], we can still |
| 1025 | // optimize if we can prove that the program has undefined behavior when |
| 1026 | // Offset is outside that range. That is the case when GEP->getOperand(0) |
| 1027 | // is a pointer to an object whose memory extent is NullTermIdx+1. |
| 1028 | if ((Known.isNonNegative() && Known.getMaxValue().ule(RHS: NullTermIdx)) || |
| 1029 | (isa<GlobalVariable>(Val: GEP->getOperand(i_nocapture: 0)) && |
| 1030 | NullTermIdx == Slice.Length - 1)) { |
| 1031 | Offset = B.CreateSExtOrTrunc(V: Offset, DestTy: CI->getType()); |
| 1032 | return B.CreateSub(LHS: ConstantInt::get(Ty: CI->getType(), V: NullTermIdx), |
| 1033 | RHS: Offset); |
| 1034 | } |
| 1035 | } |
| 1036 | } |
| 1037 | |
| 1038 | // strlen(x?"foo":"bars") --> x ? 3 : 4 |
| 1039 | if (SelectInst *SI = dyn_cast<SelectInst>(Val: Src)) { |
| 1040 | uint64_t LenTrue = GetStringLength(V: SI->getTrueValue(), CharSize); |
| 1041 | uint64_t LenFalse = GetStringLength(V: SI->getFalseValue(), CharSize); |
| 1042 | if (LenTrue && LenFalse) { |
| 1043 | ORE.emit(RemarkBuilder: [&]() { |
| 1044 | return OptimizationRemark("instcombine" , "simplify-libcalls" , CI) |
| 1045 | << "folded strlen(select) to select of constants" ; |
| 1046 | }); |
| 1047 | return B.CreateSelect(C: SI->getCondition(), |
| 1048 | True: ConstantInt::get(Ty: CI->getType(), V: LenTrue - 1), |
| 1049 | False: ConstantInt::get(Ty: CI->getType(), V: LenFalse - 1)); |
| 1050 | } |
| 1051 | } |
| 1052 | |
| 1053 | return nullptr; |
| 1054 | } |
| 1055 | |
| 1056 | Value *LibCallSimplifier::optimizeStrLen(CallInst *CI, IRBuilderBase &B) { |
| 1057 | if (Value *V = optimizeStringLength(CI, B, CharSize: 8)) |
| 1058 | return V; |
| 1059 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: 0); |
| 1060 | return nullptr; |
| 1061 | } |
| 1062 | |
| 1063 | Value *LibCallSimplifier::optimizeStrNLen(CallInst *CI, IRBuilderBase &B) { |
| 1064 | Value *Bound = CI->getArgOperand(i: 1); |
| 1065 | if (Value *V = optimizeStringLength(CI, B, CharSize: 8, Bound)) |
| 1066 | return V; |
| 1067 | |
| 1068 | if (isKnownNonZero(V: Bound, Q: DL)) |
| 1069 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: 0); |
| 1070 | return nullptr; |
| 1071 | } |
| 1072 | |
| 1073 | Value *LibCallSimplifier::optimizeWcslen(CallInst *CI, IRBuilderBase &B) { |
| 1074 | Module &M = *CI->getModule(); |
| 1075 | unsigned WCharSize = TLI->getWCharSize(M) * 8; |
| 1076 | // We cannot perform this optimization without wchar_size metadata. |
| 1077 | if (WCharSize == 0) |
| 1078 | return nullptr; |
| 1079 | |
| 1080 | return optimizeStringLength(CI, B, CharSize: WCharSize); |
| 1081 | } |
| 1082 | |
| 1083 | Value *LibCallSimplifier::optimizeStrPBrk(CallInst *CI, IRBuilderBase &B) { |
| 1084 | StringRef S1, S2; |
| 1085 | bool HasS1 = getConstantStringInfo(V: CI->getArgOperand(i: 0), Str&: S1); |
| 1086 | bool HasS2 = getConstantStringInfo(V: CI->getArgOperand(i: 1), Str&: S2); |
| 1087 | |
| 1088 | // strpbrk(s, "") -> nullptr |
| 1089 | // strpbrk("", s) -> nullptr |
| 1090 | if ((HasS1 && S1.empty()) || (HasS2 && S2.empty())) |
| 1091 | return Constant::getNullValue(Ty: CI->getType()); |
| 1092 | |
| 1093 | // Constant folding. |
| 1094 | if (HasS1 && HasS2) { |
| 1095 | size_t I = S1.find_first_of(Chars: S2); |
| 1096 | if (I == StringRef::npos) // No match. |
| 1097 | return Constant::getNullValue(Ty: CI->getType()); |
| 1098 | |
| 1099 | return B.CreateInBoundsGEP(Ty: B.getInt8Ty(), Ptr: CI->getArgOperand(i: 0), |
| 1100 | IdxList: B.getInt64(C: I), Name: "strpbrk" ); |
| 1101 | } |
| 1102 | |
| 1103 | // strpbrk(s, "a") -> strchr(s, 'a') |
| 1104 | if (HasS2 && S2.size() == 1) |
| 1105 | return copyFlags(Old: *CI, New: emitStrChr(Ptr: CI->getArgOperand(i: 0), C: S2[0], B, TLI)); |
| 1106 | |
| 1107 | return nullptr; |
| 1108 | } |
| 1109 | |
| 1110 | Value *LibCallSimplifier::optimizeStrTo(CallInst *CI, IRBuilderBase &B) { |
| 1111 | Value *EndPtr = CI->getArgOperand(i: 1); |
| 1112 | if (isa<ConstantPointerNull>(Val: EndPtr)) { |
| 1113 | // With a null EndPtr, this function won't capture the main argument. |
| 1114 | // It would be readonly too, except that it still may write to errno. |
| 1115 | CI->addParamAttr(ArgNo: 0, Attr: Attribute::getWithCaptureInfo(Context&: CI->getContext(), |
| 1116 | CI: CaptureInfo::none())); |
| 1117 | } |
| 1118 | |
| 1119 | return nullptr; |
| 1120 | } |
| 1121 | |
| 1122 | Value *LibCallSimplifier::optimizeStrSpn(CallInst *CI, IRBuilderBase &B) { |
| 1123 | StringRef S1, S2; |
| 1124 | bool HasS1 = getConstantStringInfo(V: CI->getArgOperand(i: 0), Str&: S1); |
| 1125 | bool HasS2 = getConstantStringInfo(V: CI->getArgOperand(i: 1), Str&: S2); |
| 1126 | |
| 1127 | // strspn(s, "") -> 0 |
| 1128 | // strspn("", s) -> 0 |
| 1129 | if ((HasS1 && S1.empty()) || (HasS2 && S2.empty())) |
| 1130 | return Constant::getNullValue(Ty: CI->getType()); |
| 1131 | |
| 1132 | // Constant folding. |
| 1133 | if (HasS1 && HasS2) { |
| 1134 | size_t Pos = S1.find_first_not_of(Chars: S2); |
| 1135 | if (Pos == StringRef::npos) |
| 1136 | Pos = S1.size(); |
| 1137 | return ConstantInt::get(Ty: CI->getType(), V: Pos); |
| 1138 | } |
| 1139 | |
| 1140 | return nullptr; |
| 1141 | } |
| 1142 | |
| 1143 | Value *LibCallSimplifier::optimizeStrCSpn(CallInst *CI, IRBuilderBase &B) { |
| 1144 | StringRef S1, S2; |
| 1145 | bool HasS1 = getConstantStringInfo(V: CI->getArgOperand(i: 0), Str&: S1); |
| 1146 | bool HasS2 = getConstantStringInfo(V: CI->getArgOperand(i: 1), Str&: S2); |
| 1147 | |
| 1148 | // strcspn("", s) -> 0 |
| 1149 | if (HasS1 && S1.empty()) |
| 1150 | return Constant::getNullValue(Ty: CI->getType()); |
| 1151 | |
| 1152 | // Constant folding. |
| 1153 | if (HasS1 && HasS2) { |
| 1154 | size_t Pos = S1.find_first_of(Chars: S2); |
| 1155 | if (Pos == StringRef::npos) |
| 1156 | Pos = S1.size(); |
| 1157 | return ConstantInt::get(Ty: CI->getType(), V: Pos); |
| 1158 | } |
| 1159 | |
| 1160 | // strcspn(s, "") -> strlen(s) |
| 1161 | if (HasS2 && S2.empty()) |
| 1162 | return copyFlags(Old: *CI, New: emitStrLen(Ptr: CI->getArgOperand(i: 0), B, DL, TLI)); |
| 1163 | |
| 1164 | return nullptr; |
| 1165 | } |
| 1166 | |
| 1167 | Value *LibCallSimplifier::optimizeStrStr(CallInst *CI, IRBuilderBase &B) { |
| 1168 | // fold strstr(x, x) -> x. |
| 1169 | if (CI->getArgOperand(i: 0) == CI->getArgOperand(i: 1)) |
| 1170 | return CI->getArgOperand(i: 0); |
| 1171 | |
| 1172 | // fold strstr(a, b) == a -> strncmp(a, b, strlen(b)) == 0 |
| 1173 | if (isOnlyUsedInEqualityComparison(V: CI, With: CI->getArgOperand(i: 0))) { |
| 1174 | Value *StrLen = emitStrLen(Ptr: CI->getArgOperand(i: 1), B, DL, TLI); |
| 1175 | if (!StrLen) |
| 1176 | return nullptr; |
| 1177 | Value *StrNCmp = emitStrNCmp(Ptr1: CI->getArgOperand(i: 0), Ptr2: CI->getArgOperand(i: 1), |
| 1178 | Len: StrLen, B, DL, TLI); |
| 1179 | if (!StrNCmp) |
| 1180 | return nullptr; |
| 1181 | for (User *U : llvm::make_early_inc_range(Range: CI->users())) { |
| 1182 | ICmpInst *Old = cast<ICmpInst>(Val: U); |
| 1183 | Value *Cmp = |
| 1184 | B.CreateICmp(P: Old->getPredicate(), LHS: StrNCmp, |
| 1185 | RHS: ConstantInt::getNullValue(Ty: StrNCmp->getType()), Name: "cmp" ); |
| 1186 | replaceAllUsesWith(I: Old, With: Cmp); |
| 1187 | } |
| 1188 | return CI; |
| 1189 | } |
| 1190 | |
| 1191 | // See if either input string is a constant string. |
| 1192 | StringRef SearchStr, ToFindStr; |
| 1193 | bool HasStr1 = getConstantStringInfo(V: CI->getArgOperand(i: 0), Str&: SearchStr); |
| 1194 | bool HasStr2 = getConstantStringInfo(V: CI->getArgOperand(i: 1), Str&: ToFindStr); |
| 1195 | |
| 1196 | // fold strstr(x, "") -> x. |
| 1197 | if (HasStr2 && ToFindStr.empty()) |
| 1198 | return CI->getArgOperand(i: 0); |
| 1199 | |
| 1200 | // If both strings are known, constant fold it. |
| 1201 | if (HasStr1 && HasStr2) { |
| 1202 | size_t Offset = SearchStr.find(Str: ToFindStr); |
| 1203 | |
| 1204 | if (Offset == StringRef::npos) // strstr("foo", "bar") -> null |
| 1205 | return Constant::getNullValue(Ty: CI->getType()); |
| 1206 | |
| 1207 | // strstr("abcd", "bc") -> gep((char*)"abcd", 1) |
| 1208 | return B.CreateConstInBoundsGEP1_64(Ty: B.getInt8Ty(), Ptr: CI->getArgOperand(i: 0), |
| 1209 | Idx0: Offset, Name: "strstr" ); |
| 1210 | } |
| 1211 | |
| 1212 | // fold strstr(x, "y") -> strchr(x, 'y'). |
| 1213 | if (HasStr2 && ToFindStr.size() == 1) { |
| 1214 | return emitStrChr(Ptr: CI->getArgOperand(i: 0), C: ToFindStr[0], B, TLI); |
| 1215 | } |
| 1216 | |
| 1217 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: {0, 1}); |
| 1218 | return nullptr; |
| 1219 | } |
| 1220 | |
| 1221 | Value *LibCallSimplifier::optimizeMemRChr(CallInst *CI, IRBuilderBase &B) { |
| 1222 | Value *SrcStr = CI->getArgOperand(i: 0); |
| 1223 | Value *Size = CI->getArgOperand(i: 2); |
| 1224 | annotateNonNullAndDereferenceable(CI, ArgNos: 0, Size, DL); |
| 1225 | Value *CharVal = CI->getArgOperand(i: 1); |
| 1226 | ConstantInt *LenC = dyn_cast<ConstantInt>(Val: Size); |
| 1227 | Value *NullPtr = Constant::getNullValue(Ty: CI->getType()); |
| 1228 | |
| 1229 | if (LenC) { |
| 1230 | if (LenC->isZero()) |
| 1231 | // Fold memrchr(x, y, 0) --> null. |
| 1232 | return NullPtr; |
| 1233 | |
| 1234 | if (LenC->isOne()) { |
| 1235 | // Fold memrchr(x, y, 1) --> *x == y ? x : null for any x and y, |
| 1236 | // constant or otherwise. |
| 1237 | Value *Val = B.CreateLoad(Ty: B.getInt8Ty(), Ptr: SrcStr, Name: "memrchr.char0" ); |
| 1238 | // Slice off the character's high end bits. |
| 1239 | CharVal = B.CreateTrunc(V: CharVal, DestTy: B.getInt8Ty()); |
| 1240 | Value *Cmp = B.CreateICmpEQ(LHS: Val, RHS: CharVal, Name: "memrchr.char0cmp" ); |
| 1241 | return B.CreateSelect(C: Cmp, True: SrcStr, False: NullPtr, Name: "memrchr.sel" ); |
| 1242 | } |
| 1243 | } |
| 1244 | |
| 1245 | StringRef Str; |
| 1246 | if (!getConstantStringInfo(V: SrcStr, Str, /*TrimAtNul=*/false)) |
| 1247 | return nullptr; |
| 1248 | |
| 1249 | if (Str.size() == 0) |
| 1250 | // If the array is empty fold memrchr(A, C, N) to null for any value |
| 1251 | // of C and N on the basis that the only valid value of N is zero |
| 1252 | // (otherwise the call is undefined). |
| 1253 | return NullPtr; |
| 1254 | |
| 1255 | uint64_t EndOff = UINT64_MAX; |
| 1256 | if (LenC) { |
| 1257 | EndOff = LenC->getZExtValue(); |
| 1258 | if (Str.size() < EndOff) |
| 1259 | // Punt out-of-bounds accesses to sanitizers and/or libc. |
| 1260 | return nullptr; |
| 1261 | } |
| 1262 | |
| 1263 | if (ConstantInt *CharC = dyn_cast<ConstantInt>(Val: CharVal)) { |
| 1264 | // Fold memrchr(S, C, N) for a constant C. |
| 1265 | size_t Pos = Str.rfind(C: CharC->getZExtValue(), From: EndOff); |
| 1266 | if (Pos == StringRef::npos) |
| 1267 | // When the character is not in the source array fold the result |
| 1268 | // to null regardless of Size. |
| 1269 | return NullPtr; |
| 1270 | |
| 1271 | if (LenC) |
| 1272 | // Fold memrchr(s, c, N) --> s + Pos for constant N > Pos. |
| 1273 | return B.CreateInBoundsGEP(Ty: B.getInt8Ty(), Ptr: SrcStr, IdxList: B.getInt64(C: Pos)); |
| 1274 | |
| 1275 | if (Str.find(C: Str[Pos]) == Pos) { |
| 1276 | // When there is just a single occurrence of C in S, i.e., the one |
| 1277 | // in Str[Pos], fold |
| 1278 | // memrchr(s, c, N) --> N <= Pos ? null : s + Pos |
| 1279 | // for nonconstant N. |
| 1280 | Value *Cmp = B.CreateICmpULE(LHS: Size, RHS: ConstantInt::get(Ty: Size->getType(), V: Pos), |
| 1281 | Name: "memrchr.cmp" ); |
| 1282 | Value *SrcPlus = B.CreateInBoundsGEP(Ty: B.getInt8Ty(), Ptr: SrcStr, |
| 1283 | IdxList: B.getInt64(C: Pos), Name: "memrchr.ptr_plus" ); |
| 1284 | return B.CreateSelect(C: Cmp, True: NullPtr, False: SrcPlus, Name: "memrchr.sel" ); |
| 1285 | } |
| 1286 | } |
| 1287 | |
| 1288 | // Truncate the string to search at most EndOff characters. |
| 1289 | Str = Str.substr(Start: 0, N: EndOff); |
| 1290 | if (Str.find_first_not_of(C: Str[0]) != StringRef::npos) |
| 1291 | return nullptr; |
| 1292 | |
| 1293 | // If the source array consists of all equal characters, then for any |
| 1294 | // C and N (whether in bounds or not), fold memrchr(S, C, N) to |
| 1295 | // N != 0 && *S == C ? S + N - 1 : null |
| 1296 | Type *SizeTy = Size->getType(); |
| 1297 | Type *Int8Ty = B.getInt8Ty(); |
| 1298 | Value *NNeZ = B.CreateICmpNE(LHS: Size, RHS: ConstantInt::get(Ty: SizeTy, V: 0)); |
| 1299 | // Slice off the sought character's high end bits. |
| 1300 | CharVal = B.CreateTrunc(V: CharVal, DestTy: Int8Ty); |
| 1301 | Value *CEqS0 = B.CreateICmpEQ(LHS: ConstantInt::get(Ty: Int8Ty, V: Str[0]), RHS: CharVal); |
| 1302 | Value *And = B.CreateLogicalAnd(Cond1: NNeZ, Cond2: CEqS0); |
| 1303 | Value *SizeM1 = B.CreateSub(LHS: Size, RHS: ConstantInt::get(Ty: SizeTy, V: 1)); |
| 1304 | Value *SrcPlus = |
| 1305 | B.CreateInBoundsGEP(Ty: Int8Ty, Ptr: SrcStr, IdxList: SizeM1, Name: "memrchr.ptr_plus" ); |
| 1306 | return B.CreateSelect(C: And, True: SrcPlus, False: NullPtr, Name: "memrchr.sel" ); |
| 1307 | } |
| 1308 | |
| 1309 | Value *LibCallSimplifier::optimizeMemChr(CallInst *CI, IRBuilderBase &B) { |
| 1310 | Value *SrcStr = CI->getArgOperand(i: 0); |
| 1311 | Value *Size = CI->getArgOperand(i: 2); |
| 1312 | |
| 1313 | if (isKnownNonZero(V: Size, Q: DL)) { |
| 1314 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: 0); |
| 1315 | if (isOnlyUsedInEqualityComparison(V: CI, With: SrcStr)) |
| 1316 | return memChrToCharCompare(CI, NBytes: Size, B, DL); |
| 1317 | } |
| 1318 | |
| 1319 | Value *CharVal = CI->getArgOperand(i: 1); |
| 1320 | ConstantInt *CharC = dyn_cast<ConstantInt>(Val: CharVal); |
| 1321 | ConstantInt *LenC = dyn_cast<ConstantInt>(Val: Size); |
| 1322 | Value *NullPtr = Constant::getNullValue(Ty: CI->getType()); |
| 1323 | |
| 1324 | // memchr(x, y, 0) -> null |
| 1325 | if (LenC) { |
| 1326 | if (LenC->isZero()) |
| 1327 | return NullPtr; |
| 1328 | |
| 1329 | if (LenC->isOne()) { |
| 1330 | // Fold memchr(x, y, 1) --> *x == y ? x : null for any x and y, |
| 1331 | // constant or otherwise. |
| 1332 | Value *Val = B.CreateLoad(Ty: B.getInt8Ty(), Ptr: SrcStr, Name: "memchr.char0" ); |
| 1333 | // Slice off the character's high end bits. |
| 1334 | CharVal = B.CreateTrunc(V: CharVal, DestTy: B.getInt8Ty()); |
| 1335 | Value *Cmp = B.CreateICmpEQ(LHS: Val, RHS: CharVal, Name: "memchr.char0cmp" ); |
| 1336 | return B.CreateSelect(C: Cmp, True: SrcStr, False: NullPtr, Name: "memchr.sel" ); |
| 1337 | } |
| 1338 | } |
| 1339 | |
| 1340 | StringRef Str; |
| 1341 | if (!getConstantStringInfo(V: SrcStr, Str, /*TrimAtNul=*/false)) |
| 1342 | return nullptr; |
| 1343 | |
| 1344 | if (CharC) { |
| 1345 | size_t Pos = Str.find(C: CharC->getZExtValue()); |
| 1346 | if (Pos == StringRef::npos) |
| 1347 | // When the character is not in the source array fold the result |
| 1348 | // to null regardless of Size. |
| 1349 | return NullPtr; |
| 1350 | |
| 1351 | // Fold memchr(s, c, n) -> n <= Pos ? null : s + Pos |
| 1352 | // When the constant Size is less than or equal to the character |
| 1353 | // position also fold the result to null. |
| 1354 | Value *Cmp = B.CreateICmpULE(LHS: Size, RHS: ConstantInt::get(Ty: Size->getType(), V: Pos), |
| 1355 | Name: "memchr.cmp" ); |
| 1356 | Value *SrcPlus = B.CreateInBoundsGEP(Ty: B.getInt8Ty(), Ptr: SrcStr, IdxList: B.getInt64(C: Pos), |
| 1357 | Name: "memchr.ptr" ); |
| 1358 | return B.CreateSelect(C: Cmp, True: NullPtr, False: SrcPlus); |
| 1359 | } |
| 1360 | |
| 1361 | if (Str.size() == 0) |
| 1362 | // If the array is empty fold memchr(A, C, N) to null for any value |
| 1363 | // of C and N on the basis that the only valid value of N is zero |
| 1364 | // (otherwise the call is undefined). |
| 1365 | return NullPtr; |
| 1366 | |
| 1367 | if (LenC) |
| 1368 | Str = substr(Str, Len: LenC->getZExtValue()); |
| 1369 | |
| 1370 | size_t Pos = Str.find_first_not_of(C: Str[0]); |
| 1371 | if (Pos == StringRef::npos |
| 1372 | || Str.find_first_not_of(C: Str[Pos], From: Pos) == StringRef::npos) { |
| 1373 | // If the source array consists of at most two consecutive sequences |
| 1374 | // of the same characters, then for any C and N (whether in bounds or |
| 1375 | // not), fold memchr(S, C, N) to |
| 1376 | // N != 0 && *S == C ? S : null |
| 1377 | // or for the two sequences to: |
| 1378 | // N != 0 && *S == C ? S : (N > Pos && S[Pos] == C ? S + Pos : null) |
| 1379 | // ^Sel2 ^Sel1 are denoted above. |
| 1380 | // The latter makes it also possible to fold strchr() calls with strings |
| 1381 | // of the same characters. |
| 1382 | Type *SizeTy = Size->getType(); |
| 1383 | Type *Int8Ty = B.getInt8Ty(); |
| 1384 | |
| 1385 | // Slice off the sought character's high end bits. |
| 1386 | CharVal = B.CreateTrunc(V: CharVal, DestTy: Int8Ty); |
| 1387 | |
| 1388 | Value *Sel1 = NullPtr; |
| 1389 | if (Pos != StringRef::npos) { |
| 1390 | // Handle two consecutive sequences of the same characters. |
| 1391 | Value *PosVal = ConstantInt::get(Ty: SizeTy, V: Pos); |
| 1392 | Value *StrPos = ConstantInt::get(Ty: Int8Ty, V: Str[Pos]); |
| 1393 | Value *CEqSPos = B.CreateICmpEQ(LHS: CharVal, RHS: StrPos); |
| 1394 | Value *NGtPos = B.CreateICmp(P: ICmpInst::ICMP_UGT, LHS: Size, RHS: PosVal); |
| 1395 | Value *And = B.CreateAnd(LHS: CEqSPos, RHS: NGtPos); |
| 1396 | Value *SrcPlus = B.CreateInBoundsGEP(Ty: B.getInt8Ty(), Ptr: SrcStr, IdxList: PosVal); |
| 1397 | Sel1 = B.CreateSelect(C: And, True: SrcPlus, False: NullPtr, Name: "memchr.sel1" ); |
| 1398 | } |
| 1399 | |
| 1400 | Value *Str0 = ConstantInt::get(Ty: Int8Ty, V: Str[0]); |
| 1401 | Value *CEqS0 = B.CreateICmpEQ(LHS: Str0, RHS: CharVal); |
| 1402 | Value *NNeZ = B.CreateICmpNE(LHS: Size, RHS: ConstantInt::get(Ty: SizeTy, V: 0)); |
| 1403 | Value *And = B.CreateAnd(LHS: NNeZ, RHS: CEqS0); |
| 1404 | return B.CreateSelect(C: And, True: SrcStr, False: Sel1, Name: "memchr.sel2" ); |
| 1405 | } |
| 1406 | |
| 1407 | if (!LenC) { |
| 1408 | if (isOnlyUsedInEqualityComparison(V: CI, With: SrcStr)) |
| 1409 | // S is dereferenceable so it's safe to load from it and fold |
| 1410 | // memchr(S, C, N) == S to N && *S == C for any C and N. |
| 1411 | // TODO: This is safe even for nonconstant S. |
| 1412 | return memChrToCharCompare(CI, NBytes: Size, B, DL); |
| 1413 | |
| 1414 | // From now on we need a constant length and constant array. |
| 1415 | return nullptr; |
| 1416 | } |
| 1417 | |
| 1418 | bool OptForSize = llvm::shouldOptimizeForSize(BB: CI->getParent(), PSI, BFI, |
| 1419 | QueryType: PGSOQueryType::IRPass); |
| 1420 | |
| 1421 | // If the char is variable but the input str and length are not we can turn |
| 1422 | // this memchr call into a simple bit field test. Of course this only works |
| 1423 | // when the return value is only checked against null. |
| 1424 | // |
| 1425 | // It would be really nice to reuse switch lowering here but we can't change |
| 1426 | // the CFG at this point. |
| 1427 | // |
| 1428 | // memchr("\r\n", C, 2) != nullptr -> (1 << C & ((1 << '\r') | (1 << '\n'))) |
| 1429 | // != 0 |
| 1430 | // after bounds check. |
| 1431 | if (OptForSize || Str.empty() || !isOnlyUsedInZeroEqualityComparison(CxtI: CI)) |
| 1432 | return nullptr; |
| 1433 | |
| 1434 | unsigned char Max = |
| 1435 | *std::max_element(first: reinterpret_cast<const unsigned char *>(Str.begin()), |
| 1436 | last: reinterpret_cast<const unsigned char *>(Str.end())); |
| 1437 | |
| 1438 | // Make sure the bit field we're about to create fits in a register on the |
| 1439 | // target. |
| 1440 | // FIXME: On a 64 bit architecture this prevents us from using the |
| 1441 | // interesting range of alpha ascii chars. We could do better by emitting |
| 1442 | // two bitfields or shifting the range by 64 if no lower chars are used. |
| 1443 | if (!DL.fitsInLegalInteger(Width: Max + 1)) { |
| 1444 | // Build chain of ORs |
| 1445 | // Transform: |
| 1446 | // memchr("abcd", C, 4) != nullptr |
| 1447 | // to: |
| 1448 | // (C == 'a' || C == 'b' || C == 'c' || C == 'd') != 0 |
| 1449 | std::string SortedStr = Str.str(); |
| 1450 | llvm::sort(C&: SortedStr); |
| 1451 | // Compute the number of of non-contiguous ranges. |
| 1452 | unsigned NonContRanges = 1; |
| 1453 | for (size_t i = 1; i < SortedStr.size(); ++i) { |
| 1454 | if (SortedStr[i] > SortedStr[i - 1] + 1) { |
| 1455 | NonContRanges++; |
| 1456 | } |
| 1457 | } |
| 1458 | |
| 1459 | // Restrict this optimization to profitable cases with one or two range |
| 1460 | // checks. |
| 1461 | if (NonContRanges > 2) |
| 1462 | return nullptr; |
| 1463 | |
| 1464 | // Slice off the character's high end bits. |
| 1465 | CharVal = B.CreateTrunc(V: CharVal, DestTy: B.getInt8Ty()); |
| 1466 | |
| 1467 | SmallVector<Value *> CharCompares; |
| 1468 | for (unsigned char C : SortedStr) |
| 1469 | CharCompares.push_back(Elt: B.CreateICmpEQ(LHS: CharVal, RHS: B.getInt8(C))); |
| 1470 | |
| 1471 | return B.CreateIntToPtr(V: B.CreateOr(Ops: CharCompares), DestTy: CI->getType()); |
| 1472 | } |
| 1473 | |
| 1474 | // For the bit field use a power-of-2 type with at least 8 bits to avoid |
| 1475 | // creating unnecessary illegal types. |
| 1476 | unsigned char Width = NextPowerOf2(A: std::max(a: (unsigned char)7, b: Max)); |
| 1477 | |
| 1478 | // Now build the bit field. |
| 1479 | APInt Bitfield(Width, 0); |
| 1480 | for (char C : Str) |
| 1481 | Bitfield.setBit((unsigned char)C); |
| 1482 | Value *BitfieldC = B.getInt(AI: Bitfield); |
| 1483 | |
| 1484 | // Adjust width of "C" to the bitfield width, then mask off the high bits. |
| 1485 | Value *C = B.CreateZExtOrTrunc(V: CharVal, DestTy: BitfieldC->getType()); |
| 1486 | C = B.CreateAnd(LHS: C, RHS: B.getIntN(N: Width, C: 0xFF)); |
| 1487 | |
| 1488 | // First check that the bit field access is within bounds. |
| 1489 | Value *Bounds = B.CreateICmp(P: ICmpInst::ICMP_ULT, LHS: C, RHS: B.getIntN(N: Width, C: Width), |
| 1490 | Name: "memchr.bounds" ); |
| 1491 | |
| 1492 | // Create code that checks if the given bit is set in the field. |
| 1493 | Value *Shl = B.CreateShl(LHS: B.getIntN(N: Width, C: 1ULL), RHS: C); |
| 1494 | Value *Bits = B.CreateIsNotNull(Arg: B.CreateAnd(LHS: Shl, RHS: BitfieldC), Name: "memchr.bits" ); |
| 1495 | |
| 1496 | // Finally merge both checks and cast to pointer type. The inttoptr |
| 1497 | // implicitly zexts the i1 to intptr type. |
| 1498 | return B.CreateIntToPtr(V: B.CreateLogicalAnd(Cond1: Bounds, Cond2: Bits, Name: "memchr" ), |
| 1499 | DestTy: CI->getType()); |
| 1500 | } |
| 1501 | |
| 1502 | // Optimize a memcmp or, when StrNCmp is true, strncmp call CI with constant |
| 1503 | // arrays LHS and RHS and nonconstant Size. |
| 1504 | static Value *optimizeMemCmpVarSize(CallInst *CI, Value *LHS, Value *RHS, |
| 1505 | Value *Size, bool StrNCmp, |
| 1506 | IRBuilderBase &B, const DataLayout &DL) { |
| 1507 | if (LHS == RHS) // memcmp(s,s,x) -> 0 |
| 1508 | return Constant::getNullValue(Ty: CI->getType()); |
| 1509 | |
| 1510 | StringRef LStr, RStr; |
| 1511 | if (!getConstantStringInfo(V: LHS, Str&: LStr, /*TrimAtNul=*/false) || |
| 1512 | !getConstantStringInfo(V: RHS, Str&: RStr, /*TrimAtNul=*/false)) |
| 1513 | return nullptr; |
| 1514 | |
| 1515 | // If the contents of both constant arrays are known, fold a call to |
| 1516 | // memcmp(A, B, N) to |
| 1517 | // N <= Pos ? 0 : (A < B ? -1 : B < A ? +1 : 0) |
| 1518 | // where Pos is the first mismatch between A and B, determined below. |
| 1519 | |
| 1520 | uint64_t Pos = 0; |
| 1521 | Value *Zero = ConstantInt::get(Ty: CI->getType(), V: 0); |
| 1522 | for (uint64_t MinSize = std::min(a: LStr.size(), b: RStr.size()); ; ++Pos) { |
| 1523 | if (Pos == MinSize || |
| 1524 | (StrNCmp && (LStr[Pos] == '\0' && RStr[Pos] == '\0'))) { |
| 1525 | // One array is a leading part of the other of equal or greater |
| 1526 | // size, or for strncmp, the arrays are equal strings. |
| 1527 | // Fold the result to zero. Size is assumed to be in bounds, since |
| 1528 | // otherwise the call would be undefined. |
| 1529 | return Zero; |
| 1530 | } |
| 1531 | |
| 1532 | if (LStr[Pos] != RStr[Pos]) |
| 1533 | break; |
| 1534 | } |
| 1535 | |
| 1536 | // Normalize the result. |
| 1537 | typedef unsigned char UChar; |
| 1538 | int IRes = UChar(LStr[Pos]) < UChar(RStr[Pos]) ? -1 : 1; |
| 1539 | Value *MaxSize = ConstantInt::get(Ty: Size->getType(), V: Pos); |
| 1540 | Value *Cmp = B.CreateICmp(P: ICmpInst::ICMP_ULE, LHS: Size, RHS: MaxSize); |
| 1541 | Value *Res = ConstantInt::getSigned(Ty: CI->getType(), V: IRes); |
| 1542 | return B.CreateSelect(C: Cmp, True: Zero, False: Res); |
| 1543 | } |
| 1544 | |
| 1545 | // Optimize a memcmp call CI with constant size Len. |
| 1546 | static Value *optimizeMemCmpConstantSize(CallInst *CI, Value *LHS, Value *RHS, |
| 1547 | uint64_t Len, IRBuilderBase &B, |
| 1548 | const DataLayout &DL) { |
| 1549 | if (Len == 0) // memcmp(s1,s2,0) -> 0 |
| 1550 | return Constant::getNullValue(Ty: CI->getType()); |
| 1551 | |
| 1552 | // memcmp(S1,S2,1) -> *(unsigned char*)LHS - *(unsigned char*)RHS |
| 1553 | if (Len == 1) { |
| 1554 | Value *LHSV = B.CreateZExt(V: B.CreateLoad(Ty: B.getInt8Ty(), Ptr: LHS, Name: "lhsc" ), |
| 1555 | DestTy: CI->getType(), Name: "lhsv" ); |
| 1556 | Value *RHSV = B.CreateZExt(V: B.CreateLoad(Ty: B.getInt8Ty(), Ptr: RHS, Name: "rhsc" ), |
| 1557 | DestTy: CI->getType(), Name: "rhsv" ); |
| 1558 | return B.CreateSub(LHS: LHSV, RHS: RHSV, Name: "chardiff" ); |
| 1559 | } |
| 1560 | |
| 1561 | // memcmp(S1,S2,N/8)==0 -> (*(intN_t*)S1 != *(intN_t*)S2)==0 |
| 1562 | // TODO: The case where both inputs are constants does not need to be limited |
| 1563 | // to legal integers or equality comparison. See block below this. |
| 1564 | if (DL.isLegalInteger(Width: Len * 8) && isOnlyUsedInZeroEqualityComparison(CxtI: CI)) { |
| 1565 | IntegerType *IntType = IntegerType::get(C&: CI->getContext(), NumBits: Len * 8); |
| 1566 | Align PrefAlignment = DL.getPrefTypeAlign(Ty: IntType); |
| 1567 | |
| 1568 | // First, see if we can fold either argument to a constant. |
| 1569 | Value *LHSV = nullptr; |
| 1570 | if (auto *LHSC = dyn_cast<Constant>(Val: LHS)) |
| 1571 | LHSV = ConstantFoldLoadFromConstPtr(C: LHSC, Ty: IntType, DL); |
| 1572 | |
| 1573 | Value *RHSV = nullptr; |
| 1574 | if (auto *RHSC = dyn_cast<Constant>(Val: RHS)) |
| 1575 | RHSV = ConstantFoldLoadFromConstPtr(C: RHSC, Ty: IntType, DL); |
| 1576 | |
| 1577 | // Don't generate unaligned loads. If either source is constant data, |
| 1578 | // alignment doesn't matter for that source because there is no load. |
| 1579 | if ((LHSV || getKnownAlignment(V: LHS, DL, CxtI: CI) >= PrefAlignment) && |
| 1580 | (RHSV || getKnownAlignment(V: RHS, DL, CxtI: CI) >= PrefAlignment)) { |
| 1581 | if (!LHSV) |
| 1582 | LHSV = B.CreateLoad(Ty: IntType, Ptr: LHS, Name: "lhsv" ); |
| 1583 | if (!RHSV) |
| 1584 | RHSV = B.CreateLoad(Ty: IntType, Ptr: RHS, Name: "rhsv" ); |
| 1585 | return B.CreateZExt(V: B.CreateICmpNE(LHS: LHSV, RHS: RHSV), DestTy: CI->getType(), Name: "memcmp" ); |
| 1586 | } |
| 1587 | } |
| 1588 | |
| 1589 | return nullptr; |
| 1590 | } |
| 1591 | |
| 1592 | // Most simplifications for memcmp also apply to bcmp. |
| 1593 | Value *LibCallSimplifier::optimizeMemCmpBCmpCommon(CallInst *CI, |
| 1594 | IRBuilderBase &B) { |
| 1595 | Value *LHS = CI->getArgOperand(i: 0), *RHS = CI->getArgOperand(i: 1); |
| 1596 | Value *Size = CI->getArgOperand(i: 2); |
| 1597 | |
| 1598 | annotateNonNullAndDereferenceable(CI, ArgNos: {0, 1}, Size, DL); |
| 1599 | |
| 1600 | if (Value *Res = optimizeMemCmpVarSize(CI, LHS, RHS, Size, StrNCmp: false, B, DL)) |
| 1601 | return Res; |
| 1602 | |
| 1603 | // Handle constant Size. |
| 1604 | ConstantInt *LenC = dyn_cast<ConstantInt>(Val: Size); |
| 1605 | if (!LenC) |
| 1606 | return nullptr; |
| 1607 | |
| 1608 | return optimizeMemCmpConstantSize(CI, LHS, RHS, Len: LenC->getZExtValue(), B, DL); |
| 1609 | } |
| 1610 | |
| 1611 | Value *LibCallSimplifier::optimizeMemCmp(CallInst *CI, IRBuilderBase &B) { |
| 1612 | Module *M = CI->getModule(); |
| 1613 | if (Value *V = optimizeMemCmpBCmpCommon(CI, B)) |
| 1614 | return V; |
| 1615 | |
| 1616 | // memcmp(x, y, Len) == 0 -> bcmp(x, y, Len) == 0 |
| 1617 | // bcmp can be more efficient than memcmp because it only has to know that |
| 1618 | // there is a difference, not how different one is to the other. |
| 1619 | if (isLibFuncEmittable(M, TLI, TheLibFunc: LibFunc_bcmp) && |
| 1620 | isOnlyUsedInZeroEqualityComparison(CxtI: CI)) { |
| 1621 | Value *LHS = CI->getArgOperand(i: 0); |
| 1622 | Value *RHS = CI->getArgOperand(i: 1); |
| 1623 | Value *Size = CI->getArgOperand(i: 2); |
| 1624 | return copyFlags(Old: *CI, New: emitBCmp(Ptr1: LHS, Ptr2: RHS, Len: Size, B, DL, TLI)); |
| 1625 | } |
| 1626 | |
| 1627 | return nullptr; |
| 1628 | } |
| 1629 | |
| 1630 | Value *LibCallSimplifier::optimizeBCmp(CallInst *CI, IRBuilderBase &B) { |
| 1631 | return optimizeMemCmpBCmpCommon(CI, B); |
| 1632 | } |
| 1633 | |
| 1634 | Value *LibCallSimplifier::optimizeMemCpy(CallInst *CI, IRBuilderBase &B) { |
| 1635 | Value *Size = CI->getArgOperand(i: 2); |
| 1636 | annotateNonNullAndDereferenceable(CI, ArgNos: {0, 1}, Size, DL); |
| 1637 | if (isa<IntrinsicInst>(Val: CI)) |
| 1638 | return nullptr; |
| 1639 | |
| 1640 | // memcpy(x, y, n) -> llvm.memcpy(align 1 x, align 1 y, n) |
| 1641 | CallInst *NewCI = B.CreateMemCpy(Dst: CI->getArgOperand(i: 0), DstAlign: Align(1), |
| 1642 | Src: CI->getArgOperand(i: 1), SrcAlign: Align(1), Size); |
| 1643 | mergeAttributesAndFlags(NewCI, Old: *CI); |
| 1644 | return CI->getArgOperand(i: 0); |
| 1645 | } |
| 1646 | |
| 1647 | Value *LibCallSimplifier::optimizeMemCCpy(CallInst *CI, IRBuilderBase &B) { |
| 1648 | Value *Dst = CI->getArgOperand(i: 0); |
| 1649 | Value *Src = CI->getArgOperand(i: 1); |
| 1650 | ConstantInt *StopChar = dyn_cast<ConstantInt>(Val: CI->getArgOperand(i: 2)); |
| 1651 | ConstantInt *N = dyn_cast<ConstantInt>(Val: CI->getArgOperand(i: 3)); |
| 1652 | StringRef SrcStr; |
| 1653 | if (CI->use_empty() && Dst == Src) |
| 1654 | return Dst; |
| 1655 | // memccpy(d, s, c, 0) -> nullptr |
| 1656 | if (N) { |
| 1657 | if (N->isNullValue()) |
| 1658 | return Constant::getNullValue(Ty: CI->getType()); |
| 1659 | if (!getConstantStringInfo(V: Src, Str&: SrcStr, /*TrimAtNul=*/false) || |
| 1660 | // TODO: Handle zeroinitializer. |
| 1661 | !StopChar) |
| 1662 | return nullptr; |
| 1663 | } else { |
| 1664 | return nullptr; |
| 1665 | } |
| 1666 | |
| 1667 | // Wrap arg 'c' of type int to char |
| 1668 | size_t Pos = SrcStr.find(C: StopChar->getSExtValue() & 0xFF); |
| 1669 | if (Pos == StringRef::npos) { |
| 1670 | if (N->getZExtValue() <= SrcStr.size()) { |
| 1671 | copyFlags(Old: *CI, New: B.CreateMemCpy(Dst, DstAlign: Align(1), Src, SrcAlign: Align(1), |
| 1672 | Size: CI->getArgOperand(i: 3))); |
| 1673 | return Constant::getNullValue(Ty: CI->getType()); |
| 1674 | } |
| 1675 | return nullptr; |
| 1676 | } |
| 1677 | |
| 1678 | Value *NewN = |
| 1679 | ConstantInt::get(Ty: N->getType(), V: std::min(a: uint64_t(Pos + 1), b: N->getZExtValue())); |
| 1680 | // memccpy -> llvm.memcpy |
| 1681 | copyFlags(Old: *CI, New: B.CreateMemCpy(Dst, DstAlign: Align(1), Src, SrcAlign: Align(1), Size: NewN)); |
| 1682 | return Pos + 1 <= N->getZExtValue() |
| 1683 | ? B.CreateInBoundsGEP(Ty: B.getInt8Ty(), Ptr: Dst, IdxList: NewN) |
| 1684 | : Constant::getNullValue(Ty: CI->getType()); |
| 1685 | } |
| 1686 | |
| 1687 | Value *LibCallSimplifier::optimizeMemPCpy(CallInst *CI, IRBuilderBase &B) { |
| 1688 | Value *Dst = CI->getArgOperand(i: 0); |
| 1689 | Value *N = CI->getArgOperand(i: 2); |
| 1690 | // mempcpy(x, y, n) -> llvm.memcpy(align 1 x, align 1 y, n), x + n |
| 1691 | CallInst *NewCI = |
| 1692 | B.CreateMemCpy(Dst, DstAlign: Align(1), Src: CI->getArgOperand(i: 1), SrcAlign: Align(1), Size: N); |
| 1693 | // Propagate attributes, but memcpy has no return value, so make sure that |
| 1694 | // any return attributes are compliant. |
| 1695 | // TODO: Attach return value attributes to the 1st operand to preserve them? |
| 1696 | mergeAttributesAndFlags(NewCI, Old: *CI); |
| 1697 | return B.CreateInBoundsGEP(Ty: B.getInt8Ty(), Ptr: Dst, IdxList: N); |
| 1698 | } |
| 1699 | |
| 1700 | Value *LibCallSimplifier::optimizeMemMove(CallInst *CI, IRBuilderBase &B) { |
| 1701 | Value *Size = CI->getArgOperand(i: 2); |
| 1702 | annotateNonNullAndDereferenceable(CI, ArgNos: {0, 1}, Size, DL); |
| 1703 | if (isa<IntrinsicInst>(Val: CI)) |
| 1704 | return nullptr; |
| 1705 | |
| 1706 | // memmove(x, y, n) -> llvm.memmove(align 1 x, align 1 y, n) |
| 1707 | CallInst *NewCI = B.CreateMemMove(Dst: CI->getArgOperand(i: 0), DstAlign: Align(1), |
| 1708 | Src: CI->getArgOperand(i: 1), SrcAlign: Align(1), Size); |
| 1709 | mergeAttributesAndFlags(NewCI, Old: *CI); |
| 1710 | return CI->getArgOperand(i: 0); |
| 1711 | } |
| 1712 | |
| 1713 | Value *LibCallSimplifier::optimizeMemSet(CallInst *CI, IRBuilderBase &B) { |
| 1714 | Value *Size = CI->getArgOperand(i: 2); |
| 1715 | annotateNonNullAndDereferenceable(CI, ArgNos: 0, Size, DL); |
| 1716 | if (isa<IntrinsicInst>(Val: CI)) |
| 1717 | return nullptr; |
| 1718 | |
| 1719 | // memset(p, v, n) -> llvm.memset(align 1 p, v, n) |
| 1720 | Value *Val = B.CreateIntCast(V: CI->getArgOperand(i: 1), DestTy: B.getInt8Ty(), isSigned: false); |
| 1721 | CallInst *NewCI = B.CreateMemSet(Ptr: CI->getArgOperand(i: 0), Val, Size, Align: Align(1)); |
| 1722 | mergeAttributesAndFlags(NewCI, Old: *CI); |
| 1723 | return CI->getArgOperand(i: 0); |
| 1724 | } |
| 1725 | |
| 1726 | Value *LibCallSimplifier::optimizeRealloc(CallInst *CI, IRBuilderBase &B) { |
| 1727 | if (isa<ConstantPointerNull>(Val: CI->getArgOperand(i: 0))) |
| 1728 | return copyFlags(Old: *CI, New: emitMalloc(Num: CI->getArgOperand(i: 1), B, DL, TLI)); |
| 1729 | |
| 1730 | return nullptr; |
| 1731 | } |
| 1732 | |
| 1733 | // Optionally allow optimization of nobuiltin calls to operator new and its |
| 1734 | // variants. |
| 1735 | Value *LibCallSimplifier::maybeOptimizeNoBuiltinOperatorNew(CallInst *CI, |
| 1736 | IRBuilderBase &B) { |
| 1737 | if (!OptimizeHotColdNew) |
| 1738 | return nullptr; |
| 1739 | Function *Callee = CI->getCalledFunction(); |
| 1740 | if (!Callee) |
| 1741 | return nullptr; |
| 1742 | LibFunc Func; |
| 1743 | if (!TLI->getLibFunc(FDecl: *Callee, F&: Func)) |
| 1744 | return nullptr; |
| 1745 | switch (Func) { |
| 1746 | case LibFunc_Znwm: |
| 1747 | case LibFunc_ZnwmRKSt9nothrow_t: |
| 1748 | case LibFunc_ZnwmSt11align_val_t: |
| 1749 | case LibFunc_ZnwmSt11align_val_tRKSt9nothrow_t: |
| 1750 | case LibFunc_Znam: |
| 1751 | case LibFunc_ZnamRKSt9nothrow_t: |
| 1752 | case LibFunc_ZnamSt11align_val_t: |
| 1753 | case LibFunc_ZnamSt11align_val_tRKSt9nothrow_t: |
| 1754 | case LibFunc_size_returning_new: |
| 1755 | case LibFunc_size_returning_new_aligned: |
| 1756 | // By default normal operator new calls (not already passing a hot_cold_t |
| 1757 | // parameter) are not mutated if the call is not marked builtin. Optionally |
| 1758 | // enable that in cases where it is known to be safe. |
| 1759 | if (!OptimizeNoBuiltinHotColdNew) |
| 1760 | return nullptr; |
| 1761 | break; |
| 1762 | case LibFunc_Znwm12__hot_cold_t: |
| 1763 | case LibFunc_ZnwmRKSt9nothrow_t12__hot_cold_t: |
| 1764 | case LibFunc_ZnwmSt11align_val_t12__hot_cold_t: |
| 1765 | case LibFunc_ZnwmSt11align_val_tRKSt9nothrow_t12__hot_cold_t: |
| 1766 | case LibFunc_Znam12__hot_cold_t: |
| 1767 | case LibFunc_ZnamRKSt9nothrow_t12__hot_cold_t: |
| 1768 | case LibFunc_ZnamSt11align_val_t12__hot_cold_t: |
| 1769 | case LibFunc_ZnamSt11align_val_tRKSt9nothrow_t12__hot_cold_t: |
| 1770 | case LibFunc_size_returning_new_hot_cold: |
| 1771 | case LibFunc_size_returning_new_aligned_hot_cold: |
| 1772 | // If the nobuiltin call already passes a hot_cold_t parameter, allow update |
| 1773 | // of that parameter when enabled. |
| 1774 | if (!OptimizeExistingHotColdNew) |
| 1775 | return nullptr; |
| 1776 | break; |
| 1777 | default: |
| 1778 | return nullptr; |
| 1779 | } |
| 1780 | return optimizeNew(CI, B, Func); |
| 1781 | } |
| 1782 | |
| 1783 | // When enabled, replace operator new() calls marked with a hot or cold memprof |
| 1784 | // attribute with an operator new() call that takes a __hot_cold_t parameter. |
| 1785 | // Currently this is supported by the open source version of tcmalloc, see: |
| 1786 | // https://github.com/google/tcmalloc/blob/master/tcmalloc/new_extension.h |
| 1787 | Value *LibCallSimplifier::optimizeNew(CallInst *CI, IRBuilderBase &B, |
| 1788 | LibFunc &Func) { |
| 1789 | if (!OptimizeHotColdNew) |
| 1790 | return nullptr; |
| 1791 | |
| 1792 | uint8_t HotCold; |
| 1793 | if (CI->getAttributes().getFnAttr(Kind: "memprof" ).getValueAsString() == "cold" ) |
| 1794 | HotCold = ColdNewHintValue; |
| 1795 | else if (CI->getAttributes().getFnAttr(Kind: "memprof" ).getValueAsString() == |
| 1796 | "notcold" ) |
| 1797 | HotCold = NotColdNewHintValue; |
| 1798 | else if (CI->getAttributes().getFnAttr(Kind: "memprof" ).getValueAsString() == "hot" ) |
| 1799 | HotCold = HotNewHintValue; |
| 1800 | else if (CI->getAttributes().getFnAttr(Kind: "memprof" ).getValueAsString() == |
| 1801 | "ambiguous" ) |
| 1802 | HotCold = AmbiguousNewHintValue; |
| 1803 | else |
| 1804 | return nullptr; |
| 1805 | |
| 1806 | // For calls that already pass a hot/cold hint, only update the hint if |
| 1807 | // directed by OptimizeExistingHotColdNew. For other calls to new, add a hint |
| 1808 | // if cold or hot, and leave as-is for default handling if "notcold" aka warm. |
| 1809 | // Note that in cases where we decide it is "notcold", it might be slightly |
| 1810 | // better to replace the hinted call with a non hinted call, to avoid the |
| 1811 | // extra parameter and the if condition check of the hint value in the |
| 1812 | // allocator. This can be considered in the future. |
| 1813 | Value *NewCall = nullptr; |
| 1814 | switch (Func) { |
| 1815 | case LibFunc_Znwm12__hot_cold_t: |
| 1816 | if (OptimizeExistingHotColdNew) |
| 1817 | NewCall = emitHotColdNew(Num: CI->getArgOperand(i: 0), B, TLI, |
| 1818 | NewFunc: LibFunc_Znwm12__hot_cold_t, HotCold); |
| 1819 | break; |
| 1820 | case LibFunc_Znwm: |
| 1821 | NewCall = emitHotColdNew(Num: CI->getArgOperand(i: 0), B, TLI, |
| 1822 | NewFunc: LibFunc_Znwm12__hot_cold_t, HotCold); |
| 1823 | break; |
| 1824 | case LibFunc_Znam12__hot_cold_t: |
| 1825 | if (OptimizeExistingHotColdNew) |
| 1826 | NewCall = emitHotColdNew(Num: CI->getArgOperand(i: 0), B, TLI, |
| 1827 | NewFunc: LibFunc_Znam12__hot_cold_t, HotCold); |
| 1828 | break; |
| 1829 | case LibFunc_Znam: |
| 1830 | NewCall = emitHotColdNew(Num: CI->getArgOperand(i: 0), B, TLI, |
| 1831 | NewFunc: LibFunc_Znam12__hot_cold_t, HotCold); |
| 1832 | break; |
| 1833 | case LibFunc_ZnwmRKSt9nothrow_t12__hot_cold_t: |
| 1834 | if (OptimizeExistingHotColdNew) |
| 1835 | NewCall = emitHotColdNewNoThrow( |
| 1836 | Num: CI->getArgOperand(i: 0), NoThrow: CI->getArgOperand(i: 1), B, TLI, |
| 1837 | NewFunc: LibFunc_ZnwmRKSt9nothrow_t12__hot_cold_t, HotCold); |
| 1838 | break; |
| 1839 | case LibFunc_ZnwmRKSt9nothrow_t: |
| 1840 | NewCall = emitHotColdNewNoThrow( |
| 1841 | Num: CI->getArgOperand(i: 0), NoThrow: CI->getArgOperand(i: 1), B, TLI, |
| 1842 | NewFunc: LibFunc_ZnwmRKSt9nothrow_t12__hot_cold_t, HotCold); |
| 1843 | break; |
| 1844 | case LibFunc_ZnamRKSt9nothrow_t12__hot_cold_t: |
| 1845 | if (OptimizeExistingHotColdNew) |
| 1846 | NewCall = emitHotColdNewNoThrow( |
| 1847 | Num: CI->getArgOperand(i: 0), NoThrow: CI->getArgOperand(i: 1), B, TLI, |
| 1848 | NewFunc: LibFunc_ZnamRKSt9nothrow_t12__hot_cold_t, HotCold); |
| 1849 | break; |
| 1850 | case LibFunc_ZnamRKSt9nothrow_t: |
| 1851 | NewCall = emitHotColdNewNoThrow( |
| 1852 | Num: CI->getArgOperand(i: 0), NoThrow: CI->getArgOperand(i: 1), B, TLI, |
| 1853 | NewFunc: LibFunc_ZnamRKSt9nothrow_t12__hot_cold_t, HotCold); |
| 1854 | break; |
| 1855 | case LibFunc_ZnwmSt11align_val_t12__hot_cold_t: |
| 1856 | if (OptimizeExistingHotColdNew) |
| 1857 | NewCall = emitHotColdNewAligned( |
| 1858 | Num: CI->getArgOperand(i: 0), Align: CI->getArgOperand(i: 1), B, TLI, |
| 1859 | NewFunc: LibFunc_ZnwmSt11align_val_t12__hot_cold_t, HotCold); |
| 1860 | break; |
| 1861 | case LibFunc_ZnwmSt11align_val_t: |
| 1862 | NewCall = emitHotColdNewAligned( |
| 1863 | Num: CI->getArgOperand(i: 0), Align: CI->getArgOperand(i: 1), B, TLI, |
| 1864 | NewFunc: LibFunc_ZnwmSt11align_val_t12__hot_cold_t, HotCold); |
| 1865 | break; |
| 1866 | case LibFunc_ZnamSt11align_val_t12__hot_cold_t: |
| 1867 | if (OptimizeExistingHotColdNew) |
| 1868 | NewCall = emitHotColdNewAligned( |
| 1869 | Num: CI->getArgOperand(i: 0), Align: CI->getArgOperand(i: 1), B, TLI, |
| 1870 | NewFunc: LibFunc_ZnamSt11align_val_t12__hot_cold_t, HotCold); |
| 1871 | break; |
| 1872 | case LibFunc_ZnamSt11align_val_t: |
| 1873 | NewCall = emitHotColdNewAligned( |
| 1874 | Num: CI->getArgOperand(i: 0), Align: CI->getArgOperand(i: 1), B, TLI, |
| 1875 | NewFunc: LibFunc_ZnamSt11align_val_t12__hot_cold_t, HotCold); |
| 1876 | break; |
| 1877 | case LibFunc_ZnwmSt11align_val_tRKSt9nothrow_t12__hot_cold_t: |
| 1878 | if (OptimizeExistingHotColdNew) |
| 1879 | NewCall = emitHotColdNewAlignedNoThrow( |
| 1880 | Num: CI->getArgOperand(i: 0), Align: CI->getArgOperand(i: 1), NoThrow: CI->getArgOperand(i: 2), B, |
| 1881 | TLI, NewFunc: LibFunc_ZnwmSt11align_val_tRKSt9nothrow_t12__hot_cold_t, |
| 1882 | HotCold); |
| 1883 | break; |
| 1884 | case LibFunc_ZnwmSt11align_val_tRKSt9nothrow_t: |
| 1885 | NewCall = emitHotColdNewAlignedNoThrow( |
| 1886 | Num: CI->getArgOperand(i: 0), Align: CI->getArgOperand(i: 1), NoThrow: CI->getArgOperand(i: 2), B, |
| 1887 | TLI, NewFunc: LibFunc_ZnwmSt11align_val_tRKSt9nothrow_t12__hot_cold_t, HotCold); |
| 1888 | break; |
| 1889 | case LibFunc_ZnamSt11align_val_tRKSt9nothrow_t12__hot_cold_t: |
| 1890 | if (OptimizeExistingHotColdNew) |
| 1891 | NewCall = emitHotColdNewAlignedNoThrow( |
| 1892 | Num: CI->getArgOperand(i: 0), Align: CI->getArgOperand(i: 1), NoThrow: CI->getArgOperand(i: 2), B, |
| 1893 | TLI, NewFunc: LibFunc_ZnamSt11align_val_tRKSt9nothrow_t12__hot_cold_t, |
| 1894 | HotCold); |
| 1895 | break; |
| 1896 | case LibFunc_ZnamSt11align_val_tRKSt9nothrow_t: |
| 1897 | NewCall = emitHotColdNewAlignedNoThrow( |
| 1898 | Num: CI->getArgOperand(i: 0), Align: CI->getArgOperand(i: 1), NoThrow: CI->getArgOperand(i: 2), B, |
| 1899 | TLI, NewFunc: LibFunc_ZnamSt11align_val_tRKSt9nothrow_t12__hot_cold_t, HotCold); |
| 1900 | break; |
| 1901 | case LibFunc_size_returning_new: |
| 1902 | NewCall = emitHotColdSizeReturningNew(Num: CI->getArgOperand(i: 0), B, TLI, |
| 1903 | NewFunc: LibFunc_size_returning_new_hot_cold, |
| 1904 | HotCold); |
| 1905 | break; |
| 1906 | case LibFunc_size_returning_new_hot_cold: |
| 1907 | if (OptimizeExistingHotColdNew) |
| 1908 | NewCall = emitHotColdSizeReturningNew(Num: CI->getArgOperand(i: 0), B, TLI, |
| 1909 | NewFunc: LibFunc_size_returning_new_hot_cold, |
| 1910 | HotCold); |
| 1911 | break; |
| 1912 | case LibFunc_size_returning_new_aligned: |
| 1913 | NewCall = emitHotColdSizeReturningNewAligned( |
| 1914 | Num: CI->getArgOperand(i: 0), Align: CI->getArgOperand(i: 1), B, TLI, |
| 1915 | NewFunc: LibFunc_size_returning_new_aligned_hot_cold, HotCold); |
| 1916 | break; |
| 1917 | case LibFunc_size_returning_new_aligned_hot_cold: |
| 1918 | if (OptimizeExistingHotColdNew) |
| 1919 | NewCall = emitHotColdSizeReturningNewAligned( |
| 1920 | Num: CI->getArgOperand(i: 0), Align: CI->getArgOperand(i: 1), B, TLI, |
| 1921 | NewFunc: LibFunc_size_returning_new_aligned_hot_cold, HotCold); |
| 1922 | break; |
| 1923 | default: |
| 1924 | return nullptr; |
| 1925 | } |
| 1926 | |
| 1927 | if (auto *NewCI = dyn_cast_or_null<Instruction>(Val: NewCall)) |
| 1928 | NewCI->copyMetadata(SrcInst: *CI); |
| 1929 | |
| 1930 | return NewCall; |
| 1931 | } |
| 1932 | |
| 1933 | //===----------------------------------------------------------------------===// |
| 1934 | // Math Library Optimizations |
| 1935 | //===----------------------------------------------------------------------===// |
| 1936 | |
| 1937 | // Replace a libcall \p CI with a call to intrinsic \p IID |
| 1938 | static Value *replaceUnaryCall(CallInst *CI, IRBuilderBase &B, |
| 1939 | Intrinsic::ID IID) { |
| 1940 | CallInst *NewCall = B.CreateUnaryIntrinsic(ID: IID, V: CI->getArgOperand(i: 0), FMFSource: CI); |
| 1941 | NewCall->takeName(V: CI); |
| 1942 | return copyFlags(Old: *CI, New: NewCall); |
| 1943 | } |
| 1944 | |
| 1945 | /// Return a variant of Val with float type. |
| 1946 | /// Currently this works in two cases: If Val is an FPExtension of a float |
| 1947 | /// value to something bigger, simply return the operand. |
| 1948 | /// If Val is a ConstantFP but can be converted to a float ConstantFP without |
| 1949 | /// loss of precision do so. |
| 1950 | static Value *valueHasFloatPrecision(Value *Val) { |
| 1951 | if (FPExtInst *Cast = dyn_cast<FPExtInst>(Val)) { |
| 1952 | Value *Op = Cast->getOperand(i_nocapture: 0); |
| 1953 | if (Op->getType()->isFloatTy()) |
| 1954 | return Op; |
| 1955 | } |
| 1956 | if (ConstantFP *Const = dyn_cast<ConstantFP>(Val)) { |
| 1957 | APFloat F = Const->getValueAPF(); |
| 1958 | bool losesInfo; |
| 1959 | (void)F.convert(ToSemantics: APFloat::IEEEsingle(), RM: APFloat::rmNearestTiesToEven, |
| 1960 | losesInfo: &losesInfo); |
| 1961 | if (!losesInfo) |
| 1962 | return ConstantFP::get(Context&: Const->getContext(), V: F); |
| 1963 | } |
| 1964 | return nullptr; |
| 1965 | } |
| 1966 | |
| 1967 | /// Shrink double -> float functions. |
| 1968 | static Value *optimizeDoubleFP(CallInst *CI, IRBuilderBase &B, |
| 1969 | bool isBinary, const TargetLibraryInfo *TLI, |
| 1970 | bool isPrecise = false) { |
| 1971 | Function *CalleeFn = CI->getCalledFunction(); |
| 1972 | if (!CI->getType()->isDoubleTy() || !CalleeFn) |
| 1973 | return nullptr; |
| 1974 | |
| 1975 | // If not all the uses of the function are converted to float, then bail out. |
| 1976 | // This matters if the precision of the result is more important than the |
| 1977 | // precision of the arguments. |
| 1978 | if (isPrecise) |
| 1979 | for (User *U : CI->users()) { |
| 1980 | FPTruncInst *Cast = dyn_cast<FPTruncInst>(Val: U); |
| 1981 | if (!Cast || !Cast->getType()->isFloatTy()) |
| 1982 | return nullptr; |
| 1983 | } |
| 1984 | |
| 1985 | // If this is something like 'g((double) float)', convert to 'gf(float)'. |
| 1986 | Value *V[2]; |
| 1987 | V[0] = valueHasFloatPrecision(Val: CI->getArgOperand(i: 0)); |
| 1988 | V[1] = isBinary ? valueHasFloatPrecision(Val: CI->getArgOperand(i: 1)) : nullptr; |
| 1989 | if (!V[0] || (isBinary && !V[1])) |
| 1990 | return nullptr; |
| 1991 | |
| 1992 | // If call isn't an intrinsic, check that it isn't within a function with the |
| 1993 | // same name as the float version of this call, otherwise the result is an |
| 1994 | // infinite loop. For example, from MinGW-w64: |
| 1995 | // |
| 1996 | // float expf(float val) { return (float) exp((double) val); } |
| 1997 | StringRef CalleeName = CalleeFn->getName(); |
| 1998 | bool IsIntrinsic = CalleeFn->isIntrinsic(); |
| 1999 | if (!IsIntrinsic) { |
| 2000 | StringRef CallerName = CI->getFunction()->getName(); |
| 2001 | if (CallerName.ends_with(Suffix: 'f') && |
| 2002 | CallerName.size() == (CalleeName.size() + 1) && |
| 2003 | CallerName.starts_with(Prefix: CalleeName)) |
| 2004 | return nullptr; |
| 2005 | } |
| 2006 | |
| 2007 | // Propagate the math semantics from the current function to the new function. |
| 2008 | IRBuilderBase::FastMathFlagGuard Guard(B); |
| 2009 | B.setFastMathFlags(CI->getFastMathFlags()); |
| 2010 | |
| 2011 | // g((double) float) -> (double) gf(float) |
| 2012 | Value *R; |
| 2013 | if (IsIntrinsic) { |
| 2014 | Intrinsic::ID IID = CalleeFn->getIntrinsicID(); |
| 2015 | R = isBinary ? B.CreateIntrinsic(ID: IID, Types: B.getFloatTy(), Args: V) |
| 2016 | : B.CreateIntrinsic(ID: IID, Types: B.getFloatTy(), Args: V[0]); |
| 2017 | } else { |
| 2018 | AttributeList CalleeAttrs = CalleeFn->getAttributes(); |
| 2019 | R = isBinary ? emitBinaryFloatFnCall(Op1: V[0], Op2: V[1], TLI, Name: CalleeName, B, |
| 2020 | Attrs: CalleeAttrs) |
| 2021 | : emitUnaryFloatFnCall(Op: V[0], TLI, Name: CalleeName, B, Attrs: CalleeAttrs); |
| 2022 | } |
| 2023 | return B.CreateFPExt(V: R, DestTy: B.getDoubleTy()); |
| 2024 | } |
| 2025 | |
| 2026 | /// Shrink double -> float for unary functions. |
| 2027 | static Value *optimizeUnaryDoubleFP(CallInst *CI, IRBuilderBase &B, |
| 2028 | const TargetLibraryInfo *TLI, |
| 2029 | bool isPrecise = false) { |
| 2030 | return optimizeDoubleFP(CI, B, isBinary: false, TLI, isPrecise); |
| 2031 | } |
| 2032 | |
| 2033 | /// Shrink double -> float for binary functions. |
| 2034 | static Value *optimizeBinaryDoubleFP(CallInst *CI, IRBuilderBase &B, |
| 2035 | const TargetLibraryInfo *TLI, |
| 2036 | bool isPrecise = false) { |
| 2037 | return optimizeDoubleFP(CI, B, isBinary: true, TLI, isPrecise); |
| 2038 | } |
| 2039 | |
| 2040 | // cabs(z) -> sqrt((creal(z)*creal(z)) + (cimag(z)*cimag(z))) |
| 2041 | Value *LibCallSimplifier::optimizeCAbs(CallInst *CI, IRBuilderBase &B) { |
| 2042 | Value *Real, *Imag; |
| 2043 | |
| 2044 | if (CI->arg_size() == 1) { |
| 2045 | |
| 2046 | if (!CI->isFast()) |
| 2047 | return nullptr; |
| 2048 | |
| 2049 | Value *Op = CI->getArgOperand(i: 0); |
| 2050 | assert(Op->getType()->isArrayTy() && "Unexpected signature for cabs!" ); |
| 2051 | |
| 2052 | Real = B.CreateExtractValue(Agg: Op, Idxs: 0, Name: "real" ); |
| 2053 | Imag = B.CreateExtractValue(Agg: Op, Idxs: 1, Name: "imag" ); |
| 2054 | |
| 2055 | } else { |
| 2056 | assert(CI->arg_size() == 2 && "Unexpected signature for cabs!" ); |
| 2057 | |
| 2058 | Real = CI->getArgOperand(i: 0); |
| 2059 | Imag = CI->getArgOperand(i: 1); |
| 2060 | |
| 2061 | // if real or imaginary part is zero, simplify to abs(cimag(z)) |
| 2062 | // or abs(creal(z)) |
| 2063 | Value *AbsOp = nullptr; |
| 2064 | if (ConstantFP *ConstReal = dyn_cast<ConstantFP>(Val: Real)) { |
| 2065 | if (ConstReal->isZero()) |
| 2066 | AbsOp = Imag; |
| 2067 | |
| 2068 | } else if (ConstantFP *ConstImag = dyn_cast<ConstantFP>(Val: Imag)) { |
| 2069 | if (ConstImag->isZero()) |
| 2070 | AbsOp = Real; |
| 2071 | } |
| 2072 | |
| 2073 | if (AbsOp) |
| 2074 | return copyFlags( |
| 2075 | Old: *CI, New: B.CreateUnaryIntrinsic(ID: Intrinsic::fabs, V: AbsOp, FMFSource: CI, Name: "cabs" )); |
| 2076 | |
| 2077 | if (!CI->isFast()) |
| 2078 | return nullptr; |
| 2079 | } |
| 2080 | |
| 2081 | // Propagate fast-math flags from the existing call to new instructions. |
| 2082 | Value *RealReal = B.CreateFMulFMF(L: Real, R: Real, FMFSource: CI); |
| 2083 | Value *ImagImag = B.CreateFMulFMF(L: Imag, R: Imag, FMFSource: CI); |
| 2084 | return copyFlags( |
| 2085 | Old: *CI, New: B.CreateUnaryIntrinsic(ID: Intrinsic::sqrt, |
| 2086 | V: B.CreateFAddFMF(L: RealReal, R: ImagImag, FMFSource: CI), FMFSource: CI, |
| 2087 | Name: "cabs" )); |
| 2088 | } |
| 2089 | |
| 2090 | // Return a properly extended integer (DstWidth bits wide) if the operation is |
| 2091 | // an itofp. |
| 2092 | static Value *getIntToFPVal(Value *I2F, IRBuilderBase &B, unsigned DstWidth) { |
| 2093 | if (isa<SIToFPInst>(Val: I2F) || isa<UIToFPInst>(Val: I2F)) { |
| 2094 | Value *Op = cast<Instruction>(Val: I2F)->getOperand(i: 0); |
| 2095 | // Make sure that the exponent fits inside an "int" of size DstWidth, |
| 2096 | // thus avoiding any range issues that FP has not. |
| 2097 | unsigned BitWidth = Op->getType()->getScalarSizeInBits(); |
| 2098 | if (BitWidth < DstWidth || (BitWidth == DstWidth && isa<SIToFPInst>(Val: I2F))) { |
| 2099 | Type *IntTy = Op->getType()->getWithNewBitWidth(NewBitWidth: DstWidth); |
| 2100 | return isa<SIToFPInst>(Val: I2F) ? B.CreateSExt(V: Op, DestTy: IntTy) |
| 2101 | : B.CreateZExt(V: Op, DestTy: IntTy); |
| 2102 | } |
| 2103 | } |
| 2104 | |
| 2105 | return nullptr; |
| 2106 | } |
| 2107 | |
| 2108 | /// Use exp{,2}(x * y) for pow(exp{,2}(x), y); |
| 2109 | /// ldexp(1.0, x) for pow(2.0, itofp(x)); exp2(n * x) for pow(2.0 ** n, x); |
| 2110 | /// exp10(x) for pow(10.0, x); exp2(log2(n) * x) for pow(n, x). |
| 2111 | Value *LibCallSimplifier::replacePowWithExp(CallInst *Pow, IRBuilderBase &B) { |
| 2112 | Module *M = Pow->getModule(); |
| 2113 | Value *Base = Pow->getArgOperand(i: 0), *Expo = Pow->getArgOperand(i: 1); |
| 2114 | Type *Ty = Pow->getType(); |
| 2115 | bool Ignored; |
| 2116 | |
| 2117 | // Evaluate special cases related to a nested function as the base. |
| 2118 | |
| 2119 | // pow(exp(x), y) -> exp(x * y) |
| 2120 | // pow(exp2(x), y) -> exp2(x * y) |
| 2121 | // If exp{,2}() is used only once, it is better to fold two transcendental |
| 2122 | // math functions into one. If used again, exp{,2}() would still have to be |
| 2123 | // called with the original argument, then keep both original transcendental |
| 2124 | // functions. However, this transformation is only safe with fully relaxed |
| 2125 | // math semantics, since, besides rounding differences, it changes overflow |
| 2126 | // and underflow behavior quite dramatically. For example: |
| 2127 | // pow(exp(1000), 0.001) = pow(inf, 0.001) = inf |
| 2128 | // Whereas: |
| 2129 | // exp(1000 * 0.001) = exp(1) |
| 2130 | // TODO: Loosen the requirement for fully relaxed math semantics. |
| 2131 | // TODO: Handle exp10() when more targets have it available. |
| 2132 | CallInst *BaseFn = dyn_cast<CallInst>(Val: Base); |
| 2133 | if (BaseFn && BaseFn->hasOneUse() && BaseFn->isFast() && Pow->isFast()) { |
| 2134 | LibFunc LibFn; |
| 2135 | |
| 2136 | Function *CalleeFn = BaseFn->getCalledFunction(); |
| 2137 | if (CalleeFn && TLI->getLibFunc(funcName: CalleeFn->getName(), F&: LibFn) && |
| 2138 | isLibFuncEmittable(M, TLI, TheLibFunc: LibFn)) { |
| 2139 | StringRef ExpName; |
| 2140 | Intrinsic::ID ID; |
| 2141 | Value *ExpFn; |
| 2142 | LibFunc LibFnFloat, LibFnDouble, LibFnLongDouble; |
| 2143 | |
| 2144 | switch (LibFn) { |
| 2145 | default: |
| 2146 | return nullptr; |
| 2147 | case LibFunc_expf: |
| 2148 | case LibFunc_exp: |
| 2149 | case LibFunc_expl: |
| 2150 | ExpName = TLI->getName(F: LibFunc_exp); |
| 2151 | ID = Intrinsic::exp; |
| 2152 | LibFnFloat = LibFunc_expf; |
| 2153 | LibFnDouble = LibFunc_exp; |
| 2154 | LibFnLongDouble = LibFunc_expl; |
| 2155 | break; |
| 2156 | case LibFunc_exp2f: |
| 2157 | case LibFunc_exp2: |
| 2158 | case LibFunc_exp2l: |
| 2159 | ExpName = TLI->getName(F: LibFunc_exp2); |
| 2160 | ID = Intrinsic::exp2; |
| 2161 | LibFnFloat = LibFunc_exp2f; |
| 2162 | LibFnDouble = LibFunc_exp2; |
| 2163 | LibFnLongDouble = LibFunc_exp2l; |
| 2164 | break; |
| 2165 | } |
| 2166 | |
| 2167 | // Create new exp{,2}() with the product as its argument. |
| 2168 | Value *FMul = B.CreateFMul(L: BaseFn->getArgOperand(i: 0), R: Expo, Name: "mul" ); |
| 2169 | ExpFn = BaseFn->doesNotAccessMemory() |
| 2170 | ? B.CreateUnaryIntrinsic(ID, V: FMul, FMFSource: nullptr, Name: ExpName) |
| 2171 | : emitUnaryFloatFnCall(Op: FMul, TLI, DoubleFn: LibFnDouble, FloatFn: LibFnFloat, |
| 2172 | LongDoubleFn: LibFnLongDouble, B, |
| 2173 | Attrs: BaseFn->getAttributes()); |
| 2174 | |
| 2175 | // Since the new exp{,2}() is different from the original one, dead code |
| 2176 | // elimination cannot be trusted to remove it, since it may have side |
| 2177 | // effects (e.g., errno). When the only consumer for the original |
| 2178 | // exp{,2}() is pow(), then it has to be explicitly erased. |
| 2179 | substituteInParent(I: BaseFn, With: ExpFn); |
| 2180 | return ExpFn; |
| 2181 | } |
| 2182 | } |
| 2183 | |
| 2184 | // Evaluate special cases related to a constant base. |
| 2185 | |
| 2186 | const APFloat *BaseF; |
| 2187 | if (!match(V: Base, P: m_APFloat(Res&: BaseF))) |
| 2188 | return nullptr; |
| 2189 | |
| 2190 | AttributeList NoAttrs; // Attributes are only meaningful on the original call |
| 2191 | |
| 2192 | const bool UseIntrinsic = Pow->doesNotAccessMemory(); |
| 2193 | |
| 2194 | // pow(2.0, itofp(x)) -> ldexp(1.0, x) |
| 2195 | if ((UseIntrinsic || !Ty->isVectorTy()) && BaseF->isExactlyValue(V: 2.0) && |
| 2196 | (isa<SIToFPInst>(Val: Expo) || isa<UIToFPInst>(Val: Expo)) && |
| 2197 | (UseIntrinsic || |
| 2198 | hasFloatFn(M, TLI, Ty, DoubleFn: LibFunc_ldexp, FloatFn: LibFunc_ldexpf, LongDoubleFn: LibFunc_ldexpl))) { |
| 2199 | |
| 2200 | // TODO: Shouldn't really need to depend on getIntToFPVal for intrinsic. Can |
| 2201 | // just directly use the original integer type. |
| 2202 | if (Value *ExpoI = getIntToFPVal(I2F: Expo, B, DstWidth: TLI->getIntSize())) { |
| 2203 | Constant *One = ConstantFP::get(Ty, V: 1.0); |
| 2204 | |
| 2205 | if (UseIntrinsic) { |
| 2206 | return copyFlags(Old: *Pow, New: B.CreateIntrinsic(ID: Intrinsic::ldexp, |
| 2207 | Types: {Ty, ExpoI->getType()}, |
| 2208 | Args: {One, ExpoI}, FMFSource: Pow, Name: "exp2" )); |
| 2209 | } |
| 2210 | |
| 2211 | return copyFlags(Old: *Pow, New: emitBinaryFloatFnCall( |
| 2212 | Op1: One, Op2: ExpoI, TLI, DoubleFn: LibFunc_ldexp, FloatFn: LibFunc_ldexpf, |
| 2213 | LongDoubleFn: LibFunc_ldexpl, B, Attrs: NoAttrs)); |
| 2214 | } |
| 2215 | } |
| 2216 | |
| 2217 | // pow(2.0 ** n, x) -> exp2(n * x) |
| 2218 | if (hasFloatFn(M, TLI, Ty, DoubleFn: LibFunc_exp2, FloatFn: LibFunc_exp2f, LongDoubleFn: LibFunc_exp2l)) { |
| 2219 | APFloat BaseR = APFloat(1.0); |
| 2220 | BaseR.convert(ToSemantics: BaseF->getSemantics(), RM: APFloat::rmTowardZero, losesInfo: &Ignored); |
| 2221 | BaseR = BaseR / *BaseF; |
| 2222 | bool IsInteger = BaseF->isInteger(), IsReciprocal = BaseR.isInteger(); |
| 2223 | const APFloat *NF = IsReciprocal ? &BaseR : BaseF; |
| 2224 | APSInt NI(64, false); |
| 2225 | if ((IsInteger || IsReciprocal) && |
| 2226 | NF->convertToInteger(Result&: NI, RM: APFloat::rmTowardZero, IsExact: &Ignored) == |
| 2227 | APFloat::opOK && |
| 2228 | NI > 1 && NI.isPowerOf2()) { |
| 2229 | double N = NI.logBase2() * (IsReciprocal ? -1.0 : 1.0); |
| 2230 | Value *FMul = B.CreateFMul(L: Expo, R: ConstantFP::get(Ty, V: N), Name: "mul" ); |
| 2231 | if (Pow->doesNotAccessMemory()) |
| 2232 | return copyFlags(Old: *Pow, New: B.CreateUnaryIntrinsic(ID: Intrinsic::exp2, V: FMul, |
| 2233 | FMFSource: nullptr, Name: "exp2" )); |
| 2234 | else |
| 2235 | return copyFlags(Old: *Pow, New: emitUnaryFloatFnCall(Op: FMul, TLI, DoubleFn: LibFunc_exp2, |
| 2236 | FloatFn: LibFunc_exp2f, |
| 2237 | LongDoubleFn: LibFunc_exp2l, B, Attrs: NoAttrs)); |
| 2238 | } |
| 2239 | } |
| 2240 | |
| 2241 | // pow(10.0, x) -> exp10(x) |
| 2242 | if (BaseF->isExactlyValue(V: 10.0) && |
| 2243 | hasFloatFn(M, TLI, Ty, DoubleFn: LibFunc_exp10, FloatFn: LibFunc_exp10f, LongDoubleFn: LibFunc_exp10l)) { |
| 2244 | |
| 2245 | if (Pow->doesNotAccessMemory()) { |
| 2246 | CallInst *NewExp10 = |
| 2247 | B.CreateIntrinsic(ID: Intrinsic::exp10, Types: {Ty}, Args: {Expo}, FMFSource: Pow, Name: "exp10" ); |
| 2248 | return copyFlags(Old: *Pow, New: NewExp10); |
| 2249 | } |
| 2250 | |
| 2251 | return copyFlags(Old: *Pow, New: emitUnaryFloatFnCall(Op: Expo, TLI, DoubleFn: LibFunc_exp10, |
| 2252 | FloatFn: LibFunc_exp10f, LongDoubleFn: LibFunc_exp10l, |
| 2253 | B, Attrs: NoAttrs)); |
| 2254 | } |
| 2255 | |
| 2256 | // pow(x, y) -> exp2(log2(x) * y) |
| 2257 | if (Pow->hasApproxFunc() && Pow->hasNoNaNs() && BaseF->isFiniteNonZero() && |
| 2258 | !BaseF->isNegative()) { |
| 2259 | // pow(1, inf) is defined to be 1 but exp2(log2(1) * inf) evaluates to NaN. |
| 2260 | // Luckily optimizePow has already handled the x == 1 case. |
| 2261 | assert(!match(Base, m_FPOne()) && |
| 2262 | "pow(1.0, y) should have been simplified earlier!" ); |
| 2263 | |
| 2264 | Value *Log = nullptr; |
| 2265 | if (Ty->isFloatTy()) |
| 2266 | Log = ConstantFP::get(Ty, V: std::log2(x: BaseF->convertToFloat())); |
| 2267 | else if (Ty->isDoubleTy()) |
| 2268 | Log = ConstantFP::get(Ty, V: std::log2(x: BaseF->convertToDouble())); |
| 2269 | |
| 2270 | if (Log) { |
| 2271 | Value *FMul = B.CreateFMul(L: Log, R: Expo, Name: "mul" ); |
| 2272 | if (Pow->doesNotAccessMemory()) |
| 2273 | return copyFlags(Old: *Pow, New: B.CreateUnaryIntrinsic(ID: Intrinsic::exp2, V: FMul, |
| 2274 | FMFSource: nullptr, Name: "exp2" )); |
| 2275 | else if (hasFloatFn(M, TLI, Ty, DoubleFn: LibFunc_exp2, FloatFn: LibFunc_exp2f, |
| 2276 | LongDoubleFn: LibFunc_exp2l)) |
| 2277 | return copyFlags(Old: *Pow, New: emitUnaryFloatFnCall(Op: FMul, TLI, DoubleFn: LibFunc_exp2, |
| 2278 | FloatFn: LibFunc_exp2f, |
| 2279 | LongDoubleFn: LibFunc_exp2l, B, Attrs: NoAttrs)); |
| 2280 | } |
| 2281 | } |
| 2282 | |
| 2283 | return nullptr; |
| 2284 | } |
| 2285 | |
| 2286 | static Value *getSqrtCall(Value *V, AttributeList Attrs, bool NoErrno, |
| 2287 | Module *M, IRBuilderBase &B, |
| 2288 | const TargetLibraryInfo *TLI) { |
| 2289 | // If errno is never set, then use the intrinsic for sqrt(). |
| 2290 | if (NoErrno) |
| 2291 | return B.CreateUnaryIntrinsic(ID: Intrinsic::sqrt, V, FMFSource: nullptr, Name: "sqrt" ); |
| 2292 | |
| 2293 | // Otherwise, use the libcall for sqrt(). |
| 2294 | if (hasFloatFn(M, TLI, Ty: V->getType(), DoubleFn: LibFunc_sqrt, FloatFn: LibFunc_sqrtf, |
| 2295 | LongDoubleFn: LibFunc_sqrtl)) |
| 2296 | // TODO: We also should check that the target can in fact lower the sqrt() |
| 2297 | // libcall. We currently have no way to ask this question, so we ask if |
| 2298 | // the target has a sqrt() libcall, which is not exactly the same. |
| 2299 | return emitUnaryFloatFnCall(Op: V, TLI, DoubleFn: LibFunc_sqrt, FloatFn: LibFunc_sqrtf, |
| 2300 | LongDoubleFn: LibFunc_sqrtl, B, Attrs); |
| 2301 | |
| 2302 | return nullptr; |
| 2303 | } |
| 2304 | |
| 2305 | /// Use square root in place of pow(x, +/-0.5). |
| 2306 | Value *LibCallSimplifier::replacePowWithSqrt(CallInst *Pow, IRBuilderBase &B) { |
| 2307 | Value *Sqrt, *Base = Pow->getArgOperand(i: 0), *Expo = Pow->getArgOperand(i: 1); |
| 2308 | Module *Mod = Pow->getModule(); |
| 2309 | Type *Ty = Pow->getType(); |
| 2310 | |
| 2311 | const APFloat *ExpoF; |
| 2312 | if (!match(V: Expo, P: m_APFloat(Res&: ExpoF)) || |
| 2313 | (!ExpoF->isExactlyValue(V: 0.5) && !ExpoF->isExactlyValue(V: -0.5))) |
| 2314 | return nullptr; |
| 2315 | |
| 2316 | // Converting pow(X, -0.5) to 1/sqrt(X) may introduce an extra rounding step, |
| 2317 | // so that requires fast-math-flags (afn or reassoc). |
| 2318 | if (ExpoF->isNegative() && (!Pow->hasApproxFunc() && !Pow->hasAllowReassoc())) |
| 2319 | return nullptr; |
| 2320 | |
| 2321 | // If we have a pow() library call (accesses memory) and we can't guarantee |
| 2322 | // that the base is not an infinity, give up: |
| 2323 | // pow(-Inf, 0.5) is optionally required to have a result of +Inf (not setting |
| 2324 | // errno), but sqrt(-Inf) is required by various standards to set errno. |
| 2325 | if (!Pow->doesNotAccessMemory() && !Pow->hasNoInfs() && |
| 2326 | !isKnownNeverInfinity( |
| 2327 | V: Base, SQ: SimplifyQuery(DL, TLI, DT, AC, Pow, true, true, DC))) |
| 2328 | return nullptr; |
| 2329 | |
| 2330 | Sqrt = getSqrtCall(V: Base, Attrs: AttributeList(), NoErrno: Pow->doesNotAccessMemory(), M: Mod, B, |
| 2331 | TLI); |
| 2332 | if (!Sqrt) |
| 2333 | return nullptr; |
| 2334 | |
| 2335 | // Handle signed zero base by expanding to fabs(sqrt(x)). |
| 2336 | if (!Pow->hasNoSignedZeros()) |
| 2337 | Sqrt = B.CreateUnaryIntrinsic(ID: Intrinsic::fabs, V: Sqrt, FMFSource: nullptr, Name: "abs" ); |
| 2338 | |
| 2339 | Sqrt = copyFlags(Old: *Pow, New: Sqrt); |
| 2340 | |
| 2341 | // Handle non finite base by expanding to |
| 2342 | // (x == -infinity ? +infinity : sqrt(x)). |
| 2343 | if (!Pow->hasNoInfs()) { |
| 2344 | Value *PosInf = ConstantFP::getInfinity(Ty), |
| 2345 | *NegInf = ConstantFP::getInfinity(Ty, Negative: true); |
| 2346 | Value *FCmp = B.CreateFCmpOEQ(LHS: Base, RHS: NegInf, Name: "isinf" ); |
| 2347 | Sqrt = B.CreateSelect(C: FCmp, True: PosInf, False: Sqrt); |
| 2348 | } |
| 2349 | |
| 2350 | // If the exponent is negative, then get the reciprocal. |
| 2351 | if (ExpoF->isNegative()) |
| 2352 | Sqrt = B.CreateFDiv(L: ConstantFP::get(Ty, V: 1.0), R: Sqrt, Name: "reciprocal" ); |
| 2353 | |
| 2354 | return Sqrt; |
| 2355 | } |
| 2356 | |
| 2357 | static Value *createPowWithIntegerExponent(Value *Base, Value *Expo, Module *M, |
| 2358 | IRBuilderBase &B) { |
| 2359 | Value *Args[] = {Base, Expo}; |
| 2360 | Type *Types[] = {Base->getType(), Expo->getType()}; |
| 2361 | return B.CreateIntrinsic(ID: Intrinsic::powi, Types, Args); |
| 2362 | } |
| 2363 | |
| 2364 | Value *LibCallSimplifier::optimizePow(CallInst *Pow, IRBuilderBase &B) { |
| 2365 | Value *Base = Pow->getArgOperand(i: 0); |
| 2366 | Value *Expo = Pow->getArgOperand(i: 1); |
| 2367 | Function *Callee = Pow->getCalledFunction(); |
| 2368 | StringRef Name = Callee->getName(); |
| 2369 | Type *Ty = Pow->getType(); |
| 2370 | Module *M = Pow->getModule(); |
| 2371 | bool AllowApprox = Pow->hasApproxFunc(); |
| 2372 | bool Ignored; |
| 2373 | |
| 2374 | // Propagate the math semantics from the call to any created instructions. |
| 2375 | IRBuilderBase::FastMathFlagGuard Guard(B); |
| 2376 | B.setFastMathFlags(Pow->getFastMathFlags()); |
| 2377 | // Evaluate special cases related to the base. |
| 2378 | |
| 2379 | // pow(1.0, x) -> 1.0 |
| 2380 | if (match(V: Base, P: m_FPOne())) |
| 2381 | return Base; |
| 2382 | |
| 2383 | if (Value *Exp = replacePowWithExp(Pow, B)) |
| 2384 | return Exp; |
| 2385 | |
| 2386 | // Evaluate special cases related to the exponent. |
| 2387 | |
| 2388 | // pow(x, -1.0) -> 1.0 / x |
| 2389 | if (match(V: Expo, P: m_SpecificFP(V: -1.0))) |
| 2390 | return B.CreateFDiv(L: ConstantFP::get(Ty, V: 1.0), R: Base, Name: "reciprocal" ); |
| 2391 | |
| 2392 | // pow(x, +/-0.0) -> 1.0 |
| 2393 | if (match(V: Expo, P: m_AnyZeroFP())) |
| 2394 | return ConstantFP::get(Ty, V: 1.0); |
| 2395 | |
| 2396 | // pow(x, 1.0) -> x |
| 2397 | if (match(V: Expo, P: m_FPOne())) |
| 2398 | return Base; |
| 2399 | |
| 2400 | // pow(x, 2.0) -> x * x |
| 2401 | if (match(V: Expo, P: m_SpecificFP(V: 2.0))) |
| 2402 | return B.CreateFMul(L: Base, R: Base, Name: "square" ); |
| 2403 | |
| 2404 | if (Value *Sqrt = replacePowWithSqrt(Pow, B)) |
| 2405 | return Sqrt; |
| 2406 | |
| 2407 | // If we can approximate pow: |
| 2408 | // pow(x, n) -> powi(x, n) * sqrt(x) if n has exactly a 0.5 fraction |
| 2409 | // pow(x, n) -> powi(x, n) if n is a constant signed integer value |
| 2410 | const APFloat *ExpoF; |
| 2411 | if (AllowApprox && match(V: Expo, P: m_APFloat(Res&: ExpoF)) && |
| 2412 | !ExpoF->isExactlyValue(V: 0.5) && !ExpoF->isExactlyValue(V: -0.5)) { |
| 2413 | APFloat ExpoA(abs(X: *ExpoF)); |
| 2414 | APFloat ExpoI(*ExpoF); |
| 2415 | Value *Sqrt = nullptr; |
| 2416 | if (!ExpoA.isInteger()) { |
| 2417 | APFloat Expo2 = ExpoA; |
| 2418 | // To check if ExpoA is an integer + 0.5, we add it to itself. If there |
| 2419 | // is no floating point exception and the result is an integer, then |
| 2420 | // ExpoA == integer + 0.5 |
| 2421 | if (Expo2.add(RHS: ExpoA, RM: APFloat::rmNearestTiesToEven) != APFloat::opOK) |
| 2422 | return nullptr; |
| 2423 | |
| 2424 | if (!Expo2.isInteger()) |
| 2425 | return nullptr; |
| 2426 | |
| 2427 | if (ExpoI.roundToIntegral(RM: APFloat::rmTowardNegative) != |
| 2428 | APFloat::opInexact) |
| 2429 | return nullptr; |
| 2430 | if (!ExpoI.isInteger()) |
| 2431 | return nullptr; |
| 2432 | ExpoF = &ExpoI; |
| 2433 | |
| 2434 | Sqrt = getSqrtCall(V: Base, Attrs: AttributeList(), NoErrno: Pow->doesNotAccessMemory(), M, |
| 2435 | B, TLI); |
| 2436 | if (!Sqrt) |
| 2437 | return nullptr; |
| 2438 | } |
| 2439 | |
| 2440 | // 0.5 fraction is now optionally handled. |
| 2441 | // Do pow -> powi for remaining integer exponent |
| 2442 | APSInt IntExpo(TLI->getIntSize(), /*isUnsigned=*/false); |
| 2443 | if (ExpoF->isInteger() && |
| 2444 | ExpoF->convertToInteger(Result&: IntExpo, RM: APFloat::rmTowardZero, IsExact: &Ignored) == |
| 2445 | APFloat::opOK) { |
| 2446 | Value *PowI = copyFlags( |
| 2447 | Old: *Pow, |
| 2448 | New: createPowWithIntegerExponent( |
| 2449 | Base, Expo: ConstantInt::get(Ty: B.getIntNTy(N: TLI->getIntSize()), V: IntExpo), |
| 2450 | M, B)); |
| 2451 | |
| 2452 | if (PowI && Sqrt) |
| 2453 | return B.CreateFMul(L: PowI, R: Sqrt); |
| 2454 | |
| 2455 | return PowI; |
| 2456 | } |
| 2457 | } |
| 2458 | |
| 2459 | // powf(x, itofp(y)) -> powi(x, y) |
| 2460 | if (AllowApprox && (isa<SIToFPInst>(Val: Expo) || isa<UIToFPInst>(Val: Expo))) { |
| 2461 | if (Value *ExpoI = getIntToFPVal(I2F: Expo, B, DstWidth: TLI->getIntSize())) |
| 2462 | return copyFlags(Old: *Pow, New: createPowWithIntegerExponent(Base, Expo: ExpoI, M, B)); |
| 2463 | } |
| 2464 | |
| 2465 | // Shrink pow() to powf() if the arguments are single precision, |
| 2466 | // unless the result is expected to be double precision. |
| 2467 | if (UnsafeFPShrink && Name == TLI->getName(F: LibFunc_pow) && |
| 2468 | hasFloatVersion(M, FuncName: Name)) { |
| 2469 | if (Value *Shrunk = optimizeBinaryDoubleFP(CI: Pow, B, TLI, isPrecise: true)) |
| 2470 | return Shrunk; |
| 2471 | } |
| 2472 | |
| 2473 | return nullptr; |
| 2474 | } |
| 2475 | |
| 2476 | Value *LibCallSimplifier::optimizeExp2(CallInst *CI, IRBuilderBase &B) { |
| 2477 | Module *M = CI->getModule(); |
| 2478 | Function *Callee = CI->getCalledFunction(); |
| 2479 | StringRef Name = Callee->getName(); |
| 2480 | Value *Ret = nullptr; |
| 2481 | if (UnsafeFPShrink && Name == TLI->getName(F: LibFunc_exp2) && |
| 2482 | hasFloatVersion(M, FuncName: Name)) |
| 2483 | Ret = optimizeUnaryDoubleFP(CI, B, TLI, isPrecise: true); |
| 2484 | |
| 2485 | // If we have an llvm.exp2 intrinsic, emit the llvm.ldexp intrinsic. If we |
| 2486 | // have the libcall, emit the libcall. |
| 2487 | // |
| 2488 | // TODO: In principle we should be able to just always use the intrinsic for |
| 2489 | // any doesNotAccessMemory callsite. |
| 2490 | |
| 2491 | const bool UseIntrinsic = Callee->isIntrinsic(); |
| 2492 | // Bail out for vectors because the code below only expects scalars. |
| 2493 | Type *Ty = CI->getType(); |
| 2494 | if (!UseIntrinsic && Ty->isVectorTy()) |
| 2495 | return Ret; |
| 2496 | |
| 2497 | // exp2(sitofp(x)) -> ldexp(1.0, sext(x)) if sizeof(x) <= IntSize |
| 2498 | // exp2(uitofp(x)) -> ldexp(1.0, zext(x)) if sizeof(x) < IntSize |
| 2499 | Value *Op = CI->getArgOperand(i: 0); |
| 2500 | if ((isa<SIToFPInst>(Val: Op) || isa<UIToFPInst>(Val: Op)) && |
| 2501 | (UseIntrinsic || |
| 2502 | hasFloatFn(M, TLI, Ty, DoubleFn: LibFunc_ldexp, FloatFn: LibFunc_ldexpf, LongDoubleFn: LibFunc_ldexpl))) { |
| 2503 | if (Value *Exp = getIntToFPVal(I2F: Op, B, DstWidth: TLI->getIntSize())) { |
| 2504 | Constant *One = ConstantFP::get(Ty, V: 1.0); |
| 2505 | |
| 2506 | if (UseIntrinsic) { |
| 2507 | return copyFlags(Old: *CI, New: B.CreateIntrinsic(ID: Intrinsic::ldexp, |
| 2508 | Types: {Ty, Exp->getType()}, |
| 2509 | Args: {One, Exp}, FMFSource: CI)); |
| 2510 | } |
| 2511 | |
| 2512 | IRBuilderBase::FastMathFlagGuard Guard(B); |
| 2513 | B.setFastMathFlags(CI->getFastMathFlags()); |
| 2514 | return copyFlags(Old: *CI, New: emitBinaryFloatFnCall( |
| 2515 | Op1: One, Op2: Exp, TLI, DoubleFn: LibFunc_ldexp, FloatFn: LibFunc_ldexpf, |
| 2516 | LongDoubleFn: LibFunc_ldexpl, B, Attrs: AttributeList())); |
| 2517 | } |
| 2518 | } |
| 2519 | |
| 2520 | return Ret; |
| 2521 | } |
| 2522 | |
| 2523 | Value *LibCallSimplifier::optimizeFMinFMax(CallInst *CI, IRBuilderBase &B) { |
| 2524 | Module *M = CI->getModule(); |
| 2525 | |
| 2526 | // If we can shrink the call to a float function rather than a double |
| 2527 | // function, do that first. |
| 2528 | Function *Callee = CI->getCalledFunction(); |
| 2529 | StringRef Name = Callee->getName(); |
| 2530 | if ((Name == "fmin" || Name == "fmax" ) && hasFloatVersion(M, FuncName: Name)) |
| 2531 | if (Value *Ret = optimizeBinaryDoubleFP(CI, B, TLI)) |
| 2532 | return Ret; |
| 2533 | |
| 2534 | // The LLVM intrinsics minnum/maxnum correspond to fmin/fmax. Canonicalize to |
| 2535 | // the intrinsics for improved optimization (for example, vectorization). |
| 2536 | // No-signed-zeros is implied by the definitions of fmax/fmin themselves. |
| 2537 | // From the C standard draft WG14/N1256: |
| 2538 | // "Ideally, fmax would be sensitive to the sign of zero, for example |
| 2539 | // fmax(-0.0, +0.0) would return +0; however, implementation in software |
| 2540 | // might be impractical." |
| 2541 | FastMathFlags FMF = CI->getFastMathFlags(); |
| 2542 | FMF.setNoSignedZeros(); |
| 2543 | |
| 2544 | Intrinsic::ID IID = Callee->getName().starts_with(Prefix: "fmin" ) ? Intrinsic::minnum |
| 2545 | : Intrinsic::maxnum; |
| 2546 | return copyFlags(Old: *CI, New: B.CreateBinaryIntrinsic(ID: IID, LHS: CI->getArgOperand(i: 0), |
| 2547 | RHS: CI->getArgOperand(i: 1), FMFSource: FMF)); |
| 2548 | } |
| 2549 | |
| 2550 | Value *LibCallSimplifier::optimizeFMinimumnumFMaximumnum(CallInst *CI, |
| 2551 | IRBuilderBase &B) { |
| 2552 | Module *M = CI->getModule(); |
| 2553 | |
| 2554 | // If we can shrink the call to a float function rather than a double |
| 2555 | // function, do that first. |
| 2556 | Function *Callee = CI->getCalledFunction(); |
| 2557 | StringRef Name = Callee->getName(); |
| 2558 | if ((Name == "fminimum_num" || Name == "fmaximum_num" ) && |
| 2559 | hasFloatVersion(M, FuncName: Name)) |
| 2560 | if (Value *Ret = optimizeBinaryDoubleFP(CI, B, TLI)) |
| 2561 | return Ret; |
| 2562 | |
| 2563 | // The new fminimum_num/fmaximum_num functions, unlike fmin/fmax, *are* |
| 2564 | // sensitive to the sign of zero, so we don't change the fast-math flags like |
| 2565 | // we did for those. |
| 2566 | |
| 2567 | Intrinsic::ID IID = Callee->getName().starts_with(Prefix: "fminimum_num" ) |
| 2568 | ? Intrinsic::minimumnum |
| 2569 | : Intrinsic::maximumnum; |
| 2570 | return copyFlags(Old: *CI, New: B.CreateBinaryIntrinsic(ID: IID, LHS: CI->getArgOperand(i: 0), |
| 2571 | RHS: CI->getArgOperand(i: 1), FMFSource: CI)); |
| 2572 | } |
| 2573 | |
| 2574 | Value *LibCallSimplifier::optimizeLog(CallInst *Log, IRBuilderBase &B) { |
| 2575 | Function *LogFn = Log->getCalledFunction(); |
| 2576 | StringRef LogNm = LogFn->getName(); |
| 2577 | Intrinsic::ID LogID = LogFn->getIntrinsicID(); |
| 2578 | Module *Mod = Log->getModule(); |
| 2579 | Type *Ty = Log->getType(); |
| 2580 | |
| 2581 | if (UnsafeFPShrink && hasFloatVersion(M: Mod, FuncName: LogNm)) |
| 2582 | if (Value *Ret = optimizeUnaryDoubleFP(CI: Log, B, TLI, isPrecise: true)) |
| 2583 | return Ret; |
| 2584 | |
| 2585 | LibFunc LogLb, ExpLb, Exp2Lb, Exp10Lb, PowLb; |
| 2586 | |
| 2587 | // This is only applicable to log(), log2(), log10(). |
| 2588 | if (TLI->getLibFunc(funcName: LogNm, F&: LogLb)) { |
| 2589 | switch (LogLb) { |
| 2590 | case LibFunc_logf: |
| 2591 | LogID = Intrinsic::log; |
| 2592 | ExpLb = LibFunc_expf; |
| 2593 | Exp2Lb = LibFunc_exp2f; |
| 2594 | Exp10Lb = LibFunc_exp10f; |
| 2595 | PowLb = LibFunc_powf; |
| 2596 | break; |
| 2597 | case LibFunc_log: |
| 2598 | LogID = Intrinsic::log; |
| 2599 | ExpLb = LibFunc_exp; |
| 2600 | Exp2Lb = LibFunc_exp2; |
| 2601 | Exp10Lb = LibFunc_exp10; |
| 2602 | PowLb = LibFunc_pow; |
| 2603 | break; |
| 2604 | case LibFunc_logl: |
| 2605 | LogID = Intrinsic::log; |
| 2606 | ExpLb = LibFunc_expl; |
| 2607 | Exp2Lb = LibFunc_exp2l; |
| 2608 | Exp10Lb = LibFunc_exp10l; |
| 2609 | PowLb = LibFunc_powl; |
| 2610 | break; |
| 2611 | case LibFunc_log2f: |
| 2612 | LogID = Intrinsic::log2; |
| 2613 | ExpLb = LibFunc_expf; |
| 2614 | Exp2Lb = LibFunc_exp2f; |
| 2615 | Exp10Lb = LibFunc_exp10f; |
| 2616 | PowLb = LibFunc_powf; |
| 2617 | break; |
| 2618 | case LibFunc_log2: |
| 2619 | LogID = Intrinsic::log2; |
| 2620 | ExpLb = LibFunc_exp; |
| 2621 | Exp2Lb = LibFunc_exp2; |
| 2622 | Exp10Lb = LibFunc_exp10; |
| 2623 | PowLb = LibFunc_pow; |
| 2624 | break; |
| 2625 | case LibFunc_log2l: |
| 2626 | LogID = Intrinsic::log2; |
| 2627 | ExpLb = LibFunc_expl; |
| 2628 | Exp2Lb = LibFunc_exp2l; |
| 2629 | Exp10Lb = LibFunc_exp10l; |
| 2630 | PowLb = LibFunc_powl; |
| 2631 | break; |
| 2632 | case LibFunc_log10f: |
| 2633 | LogID = Intrinsic::log10; |
| 2634 | ExpLb = LibFunc_expf; |
| 2635 | Exp2Lb = LibFunc_exp2f; |
| 2636 | Exp10Lb = LibFunc_exp10f; |
| 2637 | PowLb = LibFunc_powf; |
| 2638 | break; |
| 2639 | case LibFunc_log10: |
| 2640 | LogID = Intrinsic::log10; |
| 2641 | ExpLb = LibFunc_exp; |
| 2642 | Exp2Lb = LibFunc_exp2; |
| 2643 | Exp10Lb = LibFunc_exp10; |
| 2644 | PowLb = LibFunc_pow; |
| 2645 | break; |
| 2646 | case LibFunc_log10l: |
| 2647 | LogID = Intrinsic::log10; |
| 2648 | ExpLb = LibFunc_expl; |
| 2649 | Exp2Lb = LibFunc_exp2l; |
| 2650 | Exp10Lb = LibFunc_exp10l; |
| 2651 | PowLb = LibFunc_powl; |
| 2652 | break; |
| 2653 | default: |
| 2654 | return nullptr; |
| 2655 | } |
| 2656 | |
| 2657 | // Convert libcall to intrinsic if the value is known > 0. |
| 2658 | bool IsKnownNoErrno = Log->hasNoNaNs() && Log->hasNoInfs(); |
| 2659 | if (!IsKnownNoErrno) { |
| 2660 | SimplifyQuery SQ(DL, TLI, DT, AC, Log, true, true, DC); |
| 2661 | KnownFPClass Known = computeKnownFPClass( |
| 2662 | V: Log->getOperand(i_nocapture: 0), |
| 2663 | InterestedClasses: KnownFPClass::OrderedLessThanZeroMask | fcSubnormal, SQ); |
| 2664 | Function *F = Log->getParent()->getParent(); |
| 2665 | const fltSemantics &FltSem = Ty->getScalarType()->getFltSemantics(); |
| 2666 | IsKnownNoErrno = |
| 2667 | Known.cannotBeOrderedLessThanZero() && |
| 2668 | Known.isKnownNeverLogicalZero(Mode: F->getDenormalMode(FPType: FltSem)); |
| 2669 | } |
| 2670 | if (IsKnownNoErrno) { |
| 2671 | auto *NewLog = B.CreateUnaryIntrinsic(ID: LogID, V: Log->getArgOperand(i: 0), FMFSource: Log); |
| 2672 | NewLog->copyMetadata(SrcInst: *Log); |
| 2673 | return copyFlags(Old: *Log, New: NewLog); |
| 2674 | } |
| 2675 | } else if (LogID == Intrinsic::log || LogID == Intrinsic::log2 || |
| 2676 | LogID == Intrinsic::log10) { |
| 2677 | if (Ty->getScalarType()->isFloatTy()) { |
| 2678 | ExpLb = LibFunc_expf; |
| 2679 | Exp2Lb = LibFunc_exp2f; |
| 2680 | Exp10Lb = LibFunc_exp10f; |
| 2681 | PowLb = LibFunc_powf; |
| 2682 | } else if (Ty->getScalarType()->isDoubleTy()) { |
| 2683 | ExpLb = LibFunc_exp; |
| 2684 | Exp2Lb = LibFunc_exp2; |
| 2685 | Exp10Lb = LibFunc_exp10; |
| 2686 | PowLb = LibFunc_pow; |
| 2687 | } else |
| 2688 | return nullptr; |
| 2689 | } else |
| 2690 | return nullptr; |
| 2691 | |
| 2692 | // The earlier call must also be 'fast' in order to do these transforms. |
| 2693 | CallInst *Arg = dyn_cast<CallInst>(Val: Log->getArgOperand(i: 0)); |
| 2694 | if (!Log->isFast() || !Arg || !Arg->isFast() || !Arg->hasOneUse()) |
| 2695 | return nullptr; |
| 2696 | |
| 2697 | IRBuilderBase::FastMathFlagGuard Guard(B); |
| 2698 | B.setFastMathFlags(FastMathFlags::getFast()); |
| 2699 | |
| 2700 | Intrinsic::ID ArgID = Arg->getIntrinsicID(); |
| 2701 | LibFunc ArgLb = NotLibFunc; |
| 2702 | TLI->getLibFunc(CB: *Arg, F&: ArgLb); |
| 2703 | |
| 2704 | // log(pow(x,y)) -> y*log(x) |
| 2705 | AttributeList NoAttrs; |
| 2706 | if (ArgLb == PowLb || ArgID == Intrinsic::pow || ArgID == Intrinsic::powi) { |
| 2707 | Value *LogX = |
| 2708 | Log->doesNotAccessMemory() |
| 2709 | ? B.CreateUnaryIntrinsic(ID: LogID, V: Arg->getOperand(i_nocapture: 0), FMFSource: nullptr, Name: "log" ) |
| 2710 | : emitUnaryFloatFnCall(Op: Arg->getOperand(i_nocapture: 0), TLI, Name: LogNm, B, Attrs: NoAttrs); |
| 2711 | Value *Y = Arg->getArgOperand(i: 1); |
| 2712 | // Cast exponent to FP if integer. |
| 2713 | if (ArgID == Intrinsic::powi) |
| 2714 | Y = B.CreateSIToFP(V: Y, DestTy: Ty, Name: "cast" ); |
| 2715 | Value *MulY = B.CreateFMul(L: Y, R: LogX, Name: "mul" ); |
| 2716 | // Since pow() may have side effects, e.g. errno, |
| 2717 | // dead code elimination may not be trusted to remove it. |
| 2718 | substituteInParent(I: Arg, With: MulY); |
| 2719 | return MulY; |
| 2720 | } |
| 2721 | |
| 2722 | // log(exp{,2,10}(y)) -> y*log({e,2,10}) |
| 2723 | // TODO: There is no exp10() intrinsic yet. |
| 2724 | if (ArgLb == ExpLb || ArgLb == Exp2Lb || ArgLb == Exp10Lb || |
| 2725 | ArgID == Intrinsic::exp || ArgID == Intrinsic::exp2) { |
| 2726 | Constant *Eul; |
| 2727 | if (ArgLb == ExpLb || ArgID == Intrinsic::exp) |
| 2728 | // FIXME: Add more precise value of e for long double. |
| 2729 | Eul = ConstantFP::get(Ty: Log->getType(), V: numbers::e); |
| 2730 | else if (ArgLb == Exp2Lb || ArgID == Intrinsic::exp2) |
| 2731 | Eul = ConstantFP::get(Ty: Log->getType(), V: 2.0); |
| 2732 | else |
| 2733 | Eul = ConstantFP::get(Ty: Log->getType(), V: 10.0); |
| 2734 | Value *LogE = Log->doesNotAccessMemory() |
| 2735 | ? B.CreateUnaryIntrinsic(ID: LogID, V: Eul, FMFSource: nullptr, Name: "log" ) |
| 2736 | : emitUnaryFloatFnCall(Op: Eul, TLI, Name: LogNm, B, Attrs: NoAttrs); |
| 2737 | Value *MulY = B.CreateFMul(L: Arg->getArgOperand(i: 0), R: LogE, Name: "mul" ); |
| 2738 | // Since exp() may have side effects, e.g. errno, |
| 2739 | // dead code elimination may not be trusted to remove it. |
| 2740 | substituteInParent(I: Arg, With: MulY); |
| 2741 | return MulY; |
| 2742 | } |
| 2743 | |
| 2744 | return nullptr; |
| 2745 | } |
| 2746 | |
| 2747 | // sqrt(exp(X)) -> exp(X * 0.5) |
| 2748 | Value *LibCallSimplifier::mergeSqrtToExp(CallInst *CI, IRBuilderBase &B) { |
| 2749 | if (!CI->hasAllowReassoc()) |
| 2750 | return nullptr; |
| 2751 | |
| 2752 | Function *SqrtFn = CI->getCalledFunction(); |
| 2753 | CallInst *Arg = dyn_cast<CallInst>(Val: CI->getArgOperand(i: 0)); |
| 2754 | if (!Arg || !Arg->hasAllowReassoc() || !Arg->hasOneUse()) |
| 2755 | return nullptr; |
| 2756 | Intrinsic::ID ArgID = Arg->getIntrinsicID(); |
| 2757 | LibFunc ArgLb = NotLibFunc; |
| 2758 | TLI->getLibFunc(CB: *Arg, F&: ArgLb); |
| 2759 | |
| 2760 | LibFunc SqrtLb, ExpLb, Exp2Lb, Exp10Lb; |
| 2761 | |
| 2762 | if (TLI->getLibFunc(funcName: SqrtFn->getName(), F&: SqrtLb)) |
| 2763 | switch (SqrtLb) { |
| 2764 | case LibFunc_sqrtf: |
| 2765 | ExpLb = LibFunc_expf; |
| 2766 | Exp2Lb = LibFunc_exp2f; |
| 2767 | Exp10Lb = LibFunc_exp10f; |
| 2768 | break; |
| 2769 | case LibFunc_sqrt: |
| 2770 | ExpLb = LibFunc_exp; |
| 2771 | Exp2Lb = LibFunc_exp2; |
| 2772 | Exp10Lb = LibFunc_exp10; |
| 2773 | break; |
| 2774 | case LibFunc_sqrtl: |
| 2775 | ExpLb = LibFunc_expl; |
| 2776 | Exp2Lb = LibFunc_exp2l; |
| 2777 | Exp10Lb = LibFunc_exp10l; |
| 2778 | break; |
| 2779 | default: |
| 2780 | return nullptr; |
| 2781 | } |
| 2782 | else if (SqrtFn->getIntrinsicID() == Intrinsic::sqrt) { |
| 2783 | if (CI->getType()->getScalarType()->isFloatTy()) { |
| 2784 | ExpLb = LibFunc_expf; |
| 2785 | Exp2Lb = LibFunc_exp2f; |
| 2786 | Exp10Lb = LibFunc_exp10f; |
| 2787 | } else if (CI->getType()->getScalarType()->isDoubleTy()) { |
| 2788 | ExpLb = LibFunc_exp; |
| 2789 | Exp2Lb = LibFunc_exp2; |
| 2790 | Exp10Lb = LibFunc_exp10; |
| 2791 | } else |
| 2792 | return nullptr; |
| 2793 | } else |
| 2794 | return nullptr; |
| 2795 | |
| 2796 | if (ArgLb != ExpLb && ArgLb != Exp2Lb && ArgLb != Exp10Lb && |
| 2797 | ArgID != Intrinsic::exp && ArgID != Intrinsic::exp2) |
| 2798 | return nullptr; |
| 2799 | |
| 2800 | IRBuilderBase::InsertPointGuard Guard(B); |
| 2801 | B.SetInsertPoint(Arg); |
| 2802 | auto *ExpOperand = Arg->getOperand(i_nocapture: 0); |
| 2803 | auto *FMul = |
| 2804 | B.CreateFMulFMF(L: ExpOperand, R: ConstantFP::get(Ty: ExpOperand->getType(), V: 0.5), |
| 2805 | FMFSource: CI, Name: "merged.sqrt" ); |
| 2806 | |
| 2807 | Arg->setOperand(i_nocapture: 0, Val_nocapture: FMul); |
| 2808 | return Arg; |
| 2809 | } |
| 2810 | |
| 2811 | Value *LibCallSimplifier::optimizeSqrt(CallInst *CI, IRBuilderBase &B) { |
| 2812 | Module *M = CI->getModule(); |
| 2813 | Function *Callee = CI->getCalledFunction(); |
| 2814 | Value *Ret = nullptr; |
| 2815 | // TODO: Once we have a way (other than checking for the existince of the |
| 2816 | // libcall) to tell whether our target can lower @llvm.sqrt, relax the |
| 2817 | // condition below. |
| 2818 | if (isLibFuncEmittable(M, TLI, TheLibFunc: LibFunc_sqrtf) && |
| 2819 | (Callee->getName() == "sqrt" || |
| 2820 | Callee->getIntrinsicID() == Intrinsic::sqrt)) |
| 2821 | Ret = optimizeUnaryDoubleFP(CI, B, TLI, isPrecise: true); |
| 2822 | |
| 2823 | if (Value *Opt = mergeSqrtToExp(CI, B)) |
| 2824 | return Opt; |
| 2825 | |
| 2826 | if (!CI->isFast()) |
| 2827 | return Ret; |
| 2828 | |
| 2829 | Instruction *I = dyn_cast<Instruction>(Val: CI->getArgOperand(i: 0)); |
| 2830 | if (!I || I->getOpcode() != Instruction::FMul || !I->isFast()) |
| 2831 | return Ret; |
| 2832 | |
| 2833 | // We're looking for a repeated factor in a multiplication tree, |
| 2834 | // so we can do this fold: sqrt(x * x) -> fabs(x); |
| 2835 | // or this fold: sqrt((x * x) * y) -> fabs(x) * sqrt(y). |
| 2836 | Value *Op0 = I->getOperand(i: 0); |
| 2837 | Value *Op1 = I->getOperand(i: 1); |
| 2838 | Value *RepeatOp = nullptr; |
| 2839 | Value *OtherOp = nullptr; |
| 2840 | if (Op0 == Op1) { |
| 2841 | // Simple match: the operands of the multiply are identical. |
| 2842 | RepeatOp = Op0; |
| 2843 | } else { |
| 2844 | // Look for a more complicated pattern: one of the operands is itself |
| 2845 | // a multiply, so search for a common factor in that multiply. |
| 2846 | // Note: We don't bother looking any deeper than this first level or for |
| 2847 | // variations of this pattern because instcombine's visitFMUL and/or the |
| 2848 | // reassociation pass should give us this form. |
| 2849 | Value *MulOp; |
| 2850 | if (match(V: Op0, P: m_FMul(L: m_Value(V&: MulOp), R: m_Deferred(V: MulOp))) && |
| 2851 | cast<Instruction>(Val: Op0)->isFast()) { |
| 2852 | // Pattern: sqrt((x * x) * z) |
| 2853 | RepeatOp = MulOp; |
| 2854 | OtherOp = Op1; |
| 2855 | } else if (match(V: Op1, P: m_FMul(L: m_Value(V&: MulOp), R: m_Deferred(V: MulOp))) && |
| 2856 | cast<Instruction>(Val: Op1)->isFast()) { |
| 2857 | // Pattern: sqrt(z * (x * x)) |
| 2858 | RepeatOp = MulOp; |
| 2859 | OtherOp = Op0; |
| 2860 | } |
| 2861 | } |
| 2862 | if (!RepeatOp) |
| 2863 | return Ret; |
| 2864 | |
| 2865 | // Fast math flags for any created instructions should match the sqrt |
| 2866 | // and multiply. |
| 2867 | |
| 2868 | // If we found a repeated factor, hoist it out of the square root and |
| 2869 | // replace it with the fabs of that factor. |
| 2870 | Value *FabsCall = |
| 2871 | B.CreateUnaryIntrinsic(ID: Intrinsic::fabs, V: RepeatOp, FMFSource: I, Name: "fabs" ); |
| 2872 | if (OtherOp) { |
| 2873 | // If we found a non-repeated factor, we still need to get its square |
| 2874 | // root. We then multiply that by the value that was simplified out |
| 2875 | // of the square root calculation. |
| 2876 | Value *SqrtCall = |
| 2877 | B.CreateUnaryIntrinsic(ID: Intrinsic::sqrt, V: OtherOp, FMFSource: I, Name: "sqrt" ); |
| 2878 | return copyFlags(Old: *CI, New: B.CreateFMulFMF(L: FabsCall, R: SqrtCall, FMFSource: I)); |
| 2879 | } |
| 2880 | return copyFlags(Old: *CI, New: FabsCall); |
| 2881 | } |
| 2882 | |
| 2883 | Value *LibCallSimplifier::optimizeFMod(CallInst *CI, IRBuilderBase &B) { |
| 2884 | |
| 2885 | // fmod(x,y) can set errno if y == 0 or x == +/-inf, and returns Nan in those |
| 2886 | // case. If we know those do not happen, then we can convert the fmod into |
| 2887 | // frem. |
| 2888 | bool IsNoNan = CI->hasNoNaNs(); |
| 2889 | if (!IsNoNan) { |
| 2890 | SimplifyQuery SQ(DL, TLI, DT, AC, CI, true, true, DC); |
| 2891 | KnownFPClass Known0 = computeKnownFPClass(V: CI->getOperand(i_nocapture: 0), InterestedClasses: fcInf, SQ); |
| 2892 | if (Known0.isKnownNeverInfinity()) { |
| 2893 | KnownFPClass Known1 = |
| 2894 | computeKnownFPClass(V: CI->getOperand(i_nocapture: 1), InterestedClasses: fcZero | fcSubnormal, SQ); |
| 2895 | Function *F = CI->getParent()->getParent(); |
| 2896 | const fltSemantics &FltSem = |
| 2897 | CI->getType()->getScalarType()->getFltSemantics(); |
| 2898 | IsNoNan = Known1.isKnownNeverLogicalZero(Mode: F->getDenormalMode(FPType: FltSem)); |
| 2899 | } |
| 2900 | } |
| 2901 | |
| 2902 | if (IsNoNan) { |
| 2903 | Value *FRem = B.CreateFRemFMF(L: CI->getOperand(i_nocapture: 0), R: CI->getOperand(i_nocapture: 1), FMFSource: CI); |
| 2904 | if (auto *FRemI = dyn_cast<Instruction>(Val: FRem)) |
| 2905 | FRemI->setHasNoNaNs(true); |
| 2906 | return FRem; |
| 2907 | } |
| 2908 | return nullptr; |
| 2909 | } |
| 2910 | |
| 2911 | Value *LibCallSimplifier::optimizeTrigInversionPairs(CallInst *CI, |
| 2912 | IRBuilderBase &B) { |
| 2913 | Module *M = CI->getModule(); |
| 2914 | Function *Callee = CI->getCalledFunction(); |
| 2915 | Value *Ret = nullptr; |
| 2916 | StringRef Name = Callee->getName(); |
| 2917 | if (UnsafeFPShrink && |
| 2918 | (Name == "tan" || Name == "atanh" || Name == "sinh" || Name == "cosh" || |
| 2919 | Name == "asinh" ) && |
| 2920 | hasFloatVersion(M, FuncName: Name)) |
| 2921 | Ret = optimizeUnaryDoubleFP(CI, B, TLI, isPrecise: true); |
| 2922 | |
| 2923 | Value *Op1 = CI->getArgOperand(i: 0); |
| 2924 | auto *OpC = dyn_cast<CallInst>(Val: Op1); |
| 2925 | if (!OpC) |
| 2926 | return Ret; |
| 2927 | |
| 2928 | // Both calls must be 'fast' in order to remove them. |
| 2929 | if (!CI->isFast() || !OpC->isFast()) |
| 2930 | return Ret; |
| 2931 | |
| 2932 | // tan(atan(x)) -> x |
| 2933 | // atanh(tanh(x)) -> x |
| 2934 | // sinh(asinh(x)) -> x |
| 2935 | // asinh(sinh(x)) -> x |
| 2936 | // cosh(acosh(x)) -> x |
| 2937 | LibFunc Func; |
| 2938 | Function *F = OpC->getCalledFunction(); |
| 2939 | if (F && TLI->getLibFunc(funcName: F->getName(), F&: Func) && |
| 2940 | isLibFuncEmittable(M, TLI, TheLibFunc: Func)) { |
| 2941 | LibFunc inverseFunc = llvm::StringSwitch<LibFunc>(Callee->getName()) |
| 2942 | .Case(S: "tan" , Value: LibFunc_atan) |
| 2943 | .Case(S: "atanh" , Value: LibFunc_tanh) |
| 2944 | .Case(S: "sinh" , Value: LibFunc_asinh) |
| 2945 | .Case(S: "cosh" , Value: LibFunc_acosh) |
| 2946 | .Case(S: "tanf" , Value: LibFunc_atanf) |
| 2947 | .Case(S: "atanhf" , Value: LibFunc_tanhf) |
| 2948 | .Case(S: "sinhf" , Value: LibFunc_asinhf) |
| 2949 | .Case(S: "coshf" , Value: LibFunc_acoshf) |
| 2950 | .Case(S: "tanl" , Value: LibFunc_atanl) |
| 2951 | .Case(S: "atanhl" , Value: LibFunc_tanhl) |
| 2952 | .Case(S: "sinhl" , Value: LibFunc_asinhl) |
| 2953 | .Case(S: "coshl" , Value: LibFunc_acoshl) |
| 2954 | .Case(S: "asinh" , Value: LibFunc_sinh) |
| 2955 | .Case(S: "asinhf" , Value: LibFunc_sinhf) |
| 2956 | .Case(S: "asinhl" , Value: LibFunc_sinhl) |
| 2957 | .Default(Value: NotLibFunc); // Used as error value |
| 2958 | if (Func == inverseFunc) |
| 2959 | Ret = OpC->getArgOperand(i: 0); |
| 2960 | } |
| 2961 | return Ret; |
| 2962 | } |
| 2963 | |
| 2964 | static bool isTrigLibCall(CallInst *CI) { |
| 2965 | // We can only hope to do anything useful if we can ignore things like errno |
| 2966 | // and floating-point exceptions. |
| 2967 | // We already checked the prototype. |
| 2968 | return CI->doesNotThrow() && CI->doesNotAccessMemory(); |
| 2969 | } |
| 2970 | |
| 2971 | static bool insertSinCosCall(IRBuilderBase &B, Function *OrigCallee, Value *Arg, |
| 2972 | bool UseFloat, Value *&Sin, Value *&Cos, |
| 2973 | Value *&SinCos, const TargetLibraryInfo *TLI) { |
| 2974 | Module *M = OrigCallee->getParent(); |
| 2975 | Type *ArgTy = Arg->getType(); |
| 2976 | Type *ResTy; |
| 2977 | StringRef Name; |
| 2978 | |
| 2979 | Triple T(OrigCallee->getParent()->getTargetTriple()); |
| 2980 | if (UseFloat) { |
| 2981 | Name = "__sincospif_stret" ; |
| 2982 | |
| 2983 | assert(T.getArch() != Triple::x86 && "x86 messy and unsupported for now" ); |
| 2984 | // x86_64 can't use {float, float} since that would be returned in both |
| 2985 | // xmm0 and xmm1, which isn't what a real struct would do. |
| 2986 | ResTy = T.getArch() == Triple::x86_64 |
| 2987 | ? static_cast<Type *>(FixedVectorType::get(ElementType: ArgTy, NumElts: 2)) |
| 2988 | : static_cast<Type *>(StructType::get(elt1: ArgTy, elts: ArgTy)); |
| 2989 | } else { |
| 2990 | Name = "__sincospi_stret" ; |
| 2991 | ResTy = StructType::get(elt1: ArgTy, elts: ArgTy); |
| 2992 | } |
| 2993 | |
| 2994 | if (!isLibFuncEmittable(M, TLI, Name)) |
| 2995 | return false; |
| 2996 | LibFunc TheLibFunc; |
| 2997 | TLI->getLibFunc(funcName: Name, F&: TheLibFunc); |
| 2998 | FunctionCallee Callee = getOrInsertLibFunc( |
| 2999 | M, TLI: *TLI, TheLibFunc, AttributeList: OrigCallee->getAttributes(), RetTy: ResTy, Args: ArgTy); |
| 3000 | |
| 3001 | if (Instruction *ArgInst = dyn_cast<Instruction>(Val: Arg)) { |
| 3002 | // If the argument is an instruction, it must dominate all uses so put our |
| 3003 | // sincos call there. |
| 3004 | B.SetInsertPoint(TheBB: ArgInst->getParent(), IP: ++ArgInst->getIterator()); |
| 3005 | } else { |
| 3006 | // Otherwise (e.g. for a constant) the beginning of the function is as |
| 3007 | // good a place as any. |
| 3008 | BasicBlock &EntryBB = B.GetInsertBlock()->getParent()->getEntryBlock(); |
| 3009 | B.SetInsertPoint(TheBB: &EntryBB, IP: EntryBB.begin()); |
| 3010 | } |
| 3011 | |
| 3012 | SinCos = B.CreateCall(Callee, Args: Arg, Name: "sincospi" ); |
| 3013 | |
| 3014 | if (SinCos->getType()->isStructTy()) { |
| 3015 | Sin = B.CreateExtractValue(Agg: SinCos, Idxs: 0, Name: "sinpi" ); |
| 3016 | Cos = B.CreateExtractValue(Agg: SinCos, Idxs: 1, Name: "cospi" ); |
| 3017 | } else { |
| 3018 | Sin = B.CreateExtractElement(Vec: SinCos, Idx: ConstantInt::get(Ty: B.getInt32Ty(), V: 0), |
| 3019 | Name: "sinpi" ); |
| 3020 | Cos = B.CreateExtractElement(Vec: SinCos, Idx: ConstantInt::get(Ty: B.getInt32Ty(), V: 1), |
| 3021 | Name: "cospi" ); |
| 3022 | } |
| 3023 | |
| 3024 | return true; |
| 3025 | } |
| 3026 | |
| 3027 | static Value *optimizeSymmetricCall(CallInst *CI, bool IsEven, |
| 3028 | IRBuilderBase &B) { |
| 3029 | Value *X; |
| 3030 | Value *Src = CI->getArgOperand(i: 0); |
| 3031 | |
| 3032 | if (match(V: Src, P: m_OneUse(SubPattern: m_FNeg(X: m_Value(V&: X))))) { |
| 3033 | auto *Call = B.CreateCall(Callee: CI->getCalledFunction(), Args: {X}); |
| 3034 | Call->copyFastMathFlags(I: CI); |
| 3035 | auto *CallInst = copyFlags(Old: *CI, New: Call); |
| 3036 | if (IsEven) { |
| 3037 | // Even function: f(-x) = f(x) |
| 3038 | return CallInst; |
| 3039 | } |
| 3040 | // Odd function: f(-x) = -f(x) |
| 3041 | return B.CreateFNegFMF(V: CallInst, FMFSource: CI); |
| 3042 | } |
| 3043 | |
| 3044 | // Even function: f(abs(x)) = f(x), f(copysign(x, y)) = f(x) |
| 3045 | if (IsEven && (match(V: Src, P: m_FAbs(Op0: m_Value(V&: X))) || |
| 3046 | match(V: Src, P: m_CopySign(Op0: m_Value(V&: X), Op1: m_Value())))) { |
| 3047 | auto *Call = B.CreateCall(Callee: CI->getCalledFunction(), Args: {X}); |
| 3048 | Call->copyFastMathFlags(I: CI); |
| 3049 | return copyFlags(Old: *CI, New: Call); |
| 3050 | } |
| 3051 | |
| 3052 | return nullptr; |
| 3053 | } |
| 3054 | |
| 3055 | Value *LibCallSimplifier::optimizeSymmetric(CallInst *CI, LibFunc Func, |
| 3056 | IRBuilderBase &B) { |
| 3057 | switch (Func) { |
| 3058 | case LibFunc_cos: |
| 3059 | case LibFunc_cosf: |
| 3060 | case LibFunc_cosl: |
| 3061 | return optimizeSymmetricCall(CI, /*IsEven*/ true, B); |
| 3062 | |
| 3063 | case LibFunc_sin: |
| 3064 | case LibFunc_sinf: |
| 3065 | case LibFunc_sinl: |
| 3066 | |
| 3067 | case LibFunc_tan: |
| 3068 | case LibFunc_tanf: |
| 3069 | case LibFunc_tanl: |
| 3070 | |
| 3071 | case LibFunc_erf: |
| 3072 | case LibFunc_erff: |
| 3073 | case LibFunc_erfl: |
| 3074 | return optimizeSymmetricCall(CI, /*IsEven*/ false, B); |
| 3075 | |
| 3076 | default: |
| 3077 | return nullptr; |
| 3078 | } |
| 3079 | } |
| 3080 | |
| 3081 | Value *LibCallSimplifier::optimizeSinCosPi(CallInst *CI, bool IsSin, IRBuilderBase &B) { |
| 3082 | // Make sure the prototype is as expected, otherwise the rest of the |
| 3083 | // function is probably invalid and likely to abort. |
| 3084 | if (!isTrigLibCall(CI)) |
| 3085 | return nullptr; |
| 3086 | |
| 3087 | Value *Arg = CI->getArgOperand(i: 0); |
| 3088 | if (isa<ConstantData>(Val: Arg)) |
| 3089 | return nullptr; |
| 3090 | |
| 3091 | SmallVector<CallInst *, 1> SinCalls; |
| 3092 | SmallVector<CallInst *, 1> CosCalls; |
| 3093 | SmallVector<CallInst *, 1> SinCosCalls; |
| 3094 | |
| 3095 | bool IsFloat = Arg->getType()->isFloatTy(); |
| 3096 | |
| 3097 | // Look for all compatible sinpi, cospi and sincospi calls with the same |
| 3098 | // argument. If there are enough (in some sense) we can make the |
| 3099 | // substitution. |
| 3100 | Function *F = CI->getFunction(); |
| 3101 | for (User *U : Arg->users()) |
| 3102 | classifyArgUse(Val: U, F, IsFloat, SinCalls, CosCalls, SinCosCalls); |
| 3103 | |
| 3104 | // It's only worthwhile if both sinpi and cospi are actually used. |
| 3105 | if (SinCalls.empty() || CosCalls.empty()) |
| 3106 | return nullptr; |
| 3107 | |
| 3108 | Value *Sin, *Cos, *SinCos; |
| 3109 | if (!insertSinCosCall(B, OrigCallee: CI->getCalledFunction(), Arg, UseFloat: IsFloat, Sin, Cos, |
| 3110 | SinCos, TLI)) |
| 3111 | return nullptr; |
| 3112 | |
| 3113 | auto replaceTrigInsts = [this](SmallVectorImpl<CallInst *> &Calls, |
| 3114 | Value *Res) { |
| 3115 | for (CallInst *C : Calls) |
| 3116 | replaceAllUsesWith(I: C, With: Res); |
| 3117 | }; |
| 3118 | |
| 3119 | replaceTrigInsts(SinCalls, Sin); |
| 3120 | replaceTrigInsts(CosCalls, Cos); |
| 3121 | replaceTrigInsts(SinCosCalls, SinCos); |
| 3122 | |
| 3123 | return IsSin ? Sin : Cos; |
| 3124 | } |
| 3125 | |
| 3126 | void LibCallSimplifier::classifyArgUse( |
| 3127 | Value *Val, Function *F, bool IsFloat, |
| 3128 | SmallVectorImpl<CallInst *> &SinCalls, |
| 3129 | SmallVectorImpl<CallInst *> &CosCalls, |
| 3130 | SmallVectorImpl<CallInst *> &SinCosCalls) { |
| 3131 | auto *CI = dyn_cast<CallInst>(Val); |
| 3132 | if (!CI || CI->use_empty()) |
| 3133 | return; |
| 3134 | |
| 3135 | // Don't consider calls in other functions. |
| 3136 | if (CI->getFunction() != F) |
| 3137 | return; |
| 3138 | |
| 3139 | Module *M = CI->getModule(); |
| 3140 | Function *Callee = CI->getCalledFunction(); |
| 3141 | LibFunc Func; |
| 3142 | if (!Callee || !TLI->getLibFunc(FDecl: *Callee, F&: Func) || |
| 3143 | !isLibFuncEmittable(M, TLI, TheLibFunc: Func) || |
| 3144 | !isTrigLibCall(CI)) |
| 3145 | return; |
| 3146 | |
| 3147 | if (IsFloat) { |
| 3148 | if (Func == LibFunc_sinpif) |
| 3149 | SinCalls.push_back(Elt: CI); |
| 3150 | else if (Func == LibFunc_cospif) |
| 3151 | CosCalls.push_back(Elt: CI); |
| 3152 | else if (Func == LibFunc_sincospif_stret) |
| 3153 | SinCosCalls.push_back(Elt: CI); |
| 3154 | } else { |
| 3155 | if (Func == LibFunc_sinpi) |
| 3156 | SinCalls.push_back(Elt: CI); |
| 3157 | else if (Func == LibFunc_cospi) |
| 3158 | CosCalls.push_back(Elt: CI); |
| 3159 | else if (Func == LibFunc_sincospi_stret) |
| 3160 | SinCosCalls.push_back(Elt: CI); |
| 3161 | } |
| 3162 | } |
| 3163 | |
| 3164 | /// Constant folds remquo |
| 3165 | Value *LibCallSimplifier::optimizeRemquo(CallInst *CI, IRBuilderBase &B) { |
| 3166 | const APFloat *X, *Y; |
| 3167 | if (!match(V: CI->getArgOperand(i: 0), P: m_APFloat(Res&: X)) || |
| 3168 | !match(V: CI->getArgOperand(i: 1), P: m_APFloat(Res&: Y))) |
| 3169 | return nullptr; |
| 3170 | |
| 3171 | APFloat::opStatus Status; |
| 3172 | APFloat Quot = *X; |
| 3173 | Status = Quot.divide(RHS: *Y, RM: APFloat::rmNearestTiesToEven); |
| 3174 | if (Status != APFloat::opOK && Status != APFloat::opInexact) |
| 3175 | return nullptr; |
| 3176 | APFloat Rem = *X; |
| 3177 | if (Rem.remainder(RHS: *Y) != APFloat::opOK) |
| 3178 | return nullptr; |
| 3179 | |
| 3180 | // TODO: We can only keep at least the three of the last bits of x/y |
| 3181 | unsigned IntBW = TLI->getIntSize(); |
| 3182 | APSInt QuotInt(IntBW, /*isUnsigned=*/false); |
| 3183 | bool IsExact; |
| 3184 | Status = |
| 3185 | Quot.convertToInteger(Result&: QuotInt, RM: APFloat::rmNearestTiesToEven, IsExact: &IsExact); |
| 3186 | if (Status != APFloat::opOK && Status != APFloat::opInexact) |
| 3187 | return nullptr; |
| 3188 | |
| 3189 | B.CreateAlignedStore( |
| 3190 | Val: ConstantInt::getSigned(Ty: B.getIntNTy(N: IntBW), V: QuotInt.getExtValue()), |
| 3191 | Ptr: CI->getArgOperand(i: 2), Align: CI->getParamAlign(ArgNo: 2)); |
| 3192 | return ConstantFP::get(Ty: CI->getType(), V: Rem); |
| 3193 | } |
| 3194 | |
| 3195 | /// Constant folds fdim |
| 3196 | Value *LibCallSimplifier::optimizeFdim(CallInst *CI, IRBuilderBase &B) { |
| 3197 | // Cannot perform the fold unless the call has attribute memory(none) |
| 3198 | if (!CI->doesNotAccessMemory()) |
| 3199 | return nullptr; |
| 3200 | |
| 3201 | // TODO : Handle undef values |
| 3202 | // Propagate poison if any |
| 3203 | if (isa<PoisonValue>(Val: CI->getArgOperand(i: 0))) |
| 3204 | return CI->getArgOperand(i: 0); |
| 3205 | if (isa<PoisonValue>(Val: CI->getArgOperand(i: 1))) |
| 3206 | return CI->getArgOperand(i: 1); |
| 3207 | |
| 3208 | const APFloat *X, *Y; |
| 3209 | // Check if both values are constants |
| 3210 | if (!match(V: CI->getArgOperand(i: 0), P: m_APFloat(Res&: X)) || |
| 3211 | !match(V: CI->getArgOperand(i: 1), P: m_APFloat(Res&: Y))) |
| 3212 | return nullptr; |
| 3213 | |
| 3214 | APFloat Difference = *X; |
| 3215 | Difference.subtract(RHS: *Y, RM: RoundingMode::NearestTiesToEven); |
| 3216 | |
| 3217 | APFloat MaxVal = |
| 3218 | maximum(A: Difference, B: APFloat::getZero(Sem: CI->getType()->getFltSemantics())); |
| 3219 | return ConstantFP::get(Ty: CI->getType(), V: MaxVal); |
| 3220 | } |
| 3221 | |
| 3222 | //===----------------------------------------------------------------------===// |
| 3223 | // Integer Library Call Optimizations |
| 3224 | //===----------------------------------------------------------------------===// |
| 3225 | |
| 3226 | Value *LibCallSimplifier::optimizeFFS(CallInst *CI, IRBuilderBase &B) { |
| 3227 | // All variants of ffs return int which need not be 32 bits wide. |
| 3228 | // ffs{,l,ll}(x) -> x != 0 ? (int)llvm.cttz(x)+1 : 0 |
| 3229 | Type *RetType = CI->getType(); |
| 3230 | Value *Op = CI->getArgOperand(i: 0); |
| 3231 | Type *ArgType = Op->getType(); |
| 3232 | Value *V = B.CreateIntrinsic(ID: Intrinsic::cttz, Types: {ArgType}, Args: {Op, B.getTrue()}, |
| 3233 | FMFSource: nullptr, Name: "cttz" ); |
| 3234 | V = B.CreateAdd(LHS: V, RHS: ConstantInt::get(Ty: V->getType(), V: 1)); |
| 3235 | V = B.CreateIntCast(V, DestTy: RetType, isSigned: false); |
| 3236 | |
| 3237 | Value *Cond = B.CreateICmpNE(LHS: Op, RHS: Constant::getNullValue(Ty: ArgType)); |
| 3238 | return B.CreateSelect(C: Cond, True: V, False: ConstantInt::get(Ty: RetType, V: 0)); |
| 3239 | } |
| 3240 | |
| 3241 | Value *LibCallSimplifier::optimizeFls(CallInst *CI, IRBuilderBase &B) { |
| 3242 | // All variants of fls return int which need not be 32 bits wide. |
| 3243 | // fls{,l,ll}(x) -> (int)(sizeInBits(x) - llvm.ctlz(x, false)) |
| 3244 | Value *Op = CI->getArgOperand(i: 0); |
| 3245 | Type *ArgType = Op->getType(); |
| 3246 | Value *V = B.CreateIntrinsic(ID: Intrinsic::ctlz, Types: {ArgType}, Args: {Op, B.getFalse()}, |
| 3247 | FMFSource: nullptr, Name: "ctlz" ); |
| 3248 | V = B.CreateSub(LHS: ConstantInt::get(Ty: V->getType(), V: ArgType->getIntegerBitWidth()), |
| 3249 | RHS: V); |
| 3250 | return B.CreateIntCast(V, DestTy: CI->getType(), isSigned: false); |
| 3251 | } |
| 3252 | |
| 3253 | Value *LibCallSimplifier::optimizeAbs(CallInst *CI, IRBuilderBase &B) { |
| 3254 | // abs(x) -> x <s 0 ? -x : x |
| 3255 | // The negation has 'nsw' because abs of INT_MIN is undefined. |
| 3256 | Value *X = CI->getArgOperand(i: 0); |
| 3257 | Value *IsNeg = B.CreateIsNeg(Arg: X); |
| 3258 | Value *NegX = B.CreateNSWNeg(V: X, Name: "neg" ); |
| 3259 | return B.CreateSelect(C: IsNeg, True: NegX, False: X); |
| 3260 | } |
| 3261 | |
| 3262 | Value *LibCallSimplifier::optimizeIsDigit(CallInst *CI, IRBuilderBase &B) { |
| 3263 | // isdigit(c) -> (c-'0') <u 10 |
| 3264 | Value *Op = CI->getArgOperand(i: 0); |
| 3265 | Type *ArgType = Op->getType(); |
| 3266 | Op = B.CreateSub(LHS: Op, RHS: ConstantInt::get(Ty: ArgType, V: '0'), Name: "isdigittmp" ); |
| 3267 | Op = B.CreateICmpULT(LHS: Op, RHS: ConstantInt::get(Ty: ArgType, V: 10), Name: "isdigit" ); |
| 3268 | return B.CreateZExt(V: Op, DestTy: CI->getType()); |
| 3269 | } |
| 3270 | |
| 3271 | Value *LibCallSimplifier::optimizeIsAscii(CallInst *CI, IRBuilderBase &B) { |
| 3272 | // isascii(c) -> c <u 128 |
| 3273 | Value *Op = CI->getArgOperand(i: 0); |
| 3274 | Type *ArgType = Op->getType(); |
| 3275 | Op = B.CreateICmpULT(LHS: Op, RHS: ConstantInt::get(Ty: ArgType, V: 128), Name: "isascii" ); |
| 3276 | return B.CreateZExt(V: Op, DestTy: CI->getType()); |
| 3277 | } |
| 3278 | |
| 3279 | Value *LibCallSimplifier::optimizeToAscii(CallInst *CI, IRBuilderBase &B) { |
| 3280 | // toascii(c) -> c & 0x7f |
| 3281 | return B.CreateAnd(LHS: CI->getArgOperand(i: 0), |
| 3282 | RHS: ConstantInt::get(Ty: CI->getType(), V: 0x7F)); |
| 3283 | } |
| 3284 | |
| 3285 | // Fold calls to atoi, atol, and atoll. |
| 3286 | Value *LibCallSimplifier::optimizeAtoi(CallInst *CI, IRBuilderBase &B) { |
| 3287 | StringRef Str; |
| 3288 | if (!getConstantStringInfo(V: CI->getArgOperand(i: 0), Str)) |
| 3289 | return nullptr; |
| 3290 | |
| 3291 | return convertStrToInt(CI, Str, EndPtr: nullptr, Base: 10, /*AsSigned=*/true, B); |
| 3292 | } |
| 3293 | |
| 3294 | // Fold calls to strtol, strtoll, strtoul, and strtoull. |
| 3295 | Value *LibCallSimplifier::optimizeStrToInt(CallInst *CI, IRBuilderBase &B, |
| 3296 | bool AsSigned) { |
| 3297 | Value *EndPtr = CI->getArgOperand(i: 1); |
| 3298 | if (isa<ConstantPointerNull>(Val: EndPtr)) { |
| 3299 | // With a null EndPtr, this function won't capture the main argument. |
| 3300 | // It would be readonly too, except that it still may write to errno. |
| 3301 | CI->addParamAttr(ArgNo: 0, Attr: Attribute::getWithCaptureInfo(Context&: CI->getContext(), |
| 3302 | CI: CaptureInfo::none())); |
| 3303 | EndPtr = nullptr; |
| 3304 | } else if (!isKnownNonZero(V: EndPtr, Q: DL)) |
| 3305 | return nullptr; |
| 3306 | |
| 3307 | StringRef Str; |
| 3308 | if (!getConstantStringInfo(V: CI->getArgOperand(i: 0), Str)) |
| 3309 | return nullptr; |
| 3310 | |
| 3311 | if (ConstantInt *CInt = dyn_cast<ConstantInt>(Val: CI->getArgOperand(i: 2))) { |
| 3312 | return convertStrToInt(CI, Str, EndPtr, Base: CInt->getSExtValue(), AsSigned, B); |
| 3313 | } |
| 3314 | |
| 3315 | return nullptr; |
| 3316 | } |
| 3317 | |
| 3318 | //===----------------------------------------------------------------------===// |
| 3319 | // Formatting and IO Library Call Optimizations |
| 3320 | //===----------------------------------------------------------------------===// |
| 3321 | |
| 3322 | static bool isReportingError(Function *Callee, CallInst *CI, int StreamArg); |
| 3323 | |
| 3324 | Value *LibCallSimplifier::optimizeErrorReporting(CallInst *CI, IRBuilderBase &B, |
| 3325 | int StreamArg) { |
| 3326 | Function *Callee = CI->getCalledFunction(); |
| 3327 | // Error reporting calls should be cold, mark them as such. |
| 3328 | // This applies even to non-builtin calls: it is only a hint and applies to |
| 3329 | // functions that the frontend might not understand as builtins. |
| 3330 | |
| 3331 | // This heuristic was suggested in: |
| 3332 | // Improving Static Branch Prediction in a Compiler |
| 3333 | // Brian L. Deitrich, Ben-Chung Cheng, Wen-mei W. Hwu |
| 3334 | // Proceedings of PACT'98, Oct. 1998, IEEE |
| 3335 | if (!CI->hasFnAttr(Kind: Attribute::Cold) && |
| 3336 | isReportingError(Callee, CI, StreamArg)) { |
| 3337 | CI->addFnAttr(Kind: Attribute::Cold); |
| 3338 | } |
| 3339 | |
| 3340 | return nullptr; |
| 3341 | } |
| 3342 | |
| 3343 | static bool isReportingError(Function *Callee, CallInst *CI, int StreamArg) { |
| 3344 | if (!Callee || !Callee->isDeclaration()) |
| 3345 | return false; |
| 3346 | |
| 3347 | if (StreamArg < 0) |
| 3348 | return true; |
| 3349 | |
| 3350 | // These functions might be considered cold, but only if their stream |
| 3351 | // argument is stderr. |
| 3352 | |
| 3353 | if (StreamArg >= (int)CI->arg_size()) |
| 3354 | return false; |
| 3355 | LoadInst *LI = dyn_cast<LoadInst>(Val: CI->getArgOperand(i: StreamArg)); |
| 3356 | if (!LI) |
| 3357 | return false; |
| 3358 | GlobalVariable *GV = dyn_cast<GlobalVariable>(Val: LI->getPointerOperand()); |
| 3359 | if (!GV || !GV->isDeclaration()) |
| 3360 | return false; |
| 3361 | return GV->getName() == "stderr" ; |
| 3362 | } |
| 3363 | |
| 3364 | Value *LibCallSimplifier::optimizePrintFString(CallInst *CI, IRBuilderBase &B) { |
| 3365 | // Check for a fixed format string. |
| 3366 | StringRef FormatStr; |
| 3367 | if (!getConstantStringInfo(V: CI->getArgOperand(i: 0), Str&: FormatStr)) |
| 3368 | return nullptr; |
| 3369 | |
| 3370 | // Empty format string -> noop. |
| 3371 | if (FormatStr.empty()) // Tolerate printf's declared void. |
| 3372 | return CI->use_empty() ? (Value *)CI : ConstantInt::get(Ty: CI->getType(), V: 0); |
| 3373 | |
| 3374 | // Do not do any of the following transformations if the printf return value |
| 3375 | // is used, in general the printf return value is not compatible with either |
| 3376 | // putchar() or puts(). |
| 3377 | if (!CI->use_empty()) |
| 3378 | return nullptr; |
| 3379 | |
| 3380 | Type *IntTy = CI->getType(); |
| 3381 | // printf("x") -> putchar('x'), even for "%" and "%%". |
| 3382 | if (FormatStr.size() == 1 || FormatStr == "%%" ) { |
| 3383 | // Convert the character to unsigned char before passing it to putchar |
| 3384 | // to avoid host-specific sign extension in the IR. Putchar converts |
| 3385 | // it to unsigned char regardless. |
| 3386 | Value *IntChar = ConstantInt::get(Ty: IntTy, V: (unsigned char)FormatStr[0]); |
| 3387 | return copyFlags(Old: *CI, New: emitPutChar(Char: IntChar, B, TLI)); |
| 3388 | } |
| 3389 | |
| 3390 | // Try to remove call or emit putchar/puts. |
| 3391 | if (FormatStr == "%s" && CI->arg_size() > 1) { |
| 3392 | StringRef OperandStr; |
| 3393 | if (!getConstantStringInfo(V: CI->getOperand(i_nocapture: 1), Str&: OperandStr)) |
| 3394 | return nullptr; |
| 3395 | // printf("%s", "") --> NOP |
| 3396 | if (OperandStr.empty()) |
| 3397 | return (Value *)CI; |
| 3398 | // printf("%s", "a") --> putchar('a') |
| 3399 | if (OperandStr.size() == 1) { |
| 3400 | // Convert the character to unsigned char before passing it to putchar |
| 3401 | // to avoid host-specific sign extension in the IR. Putchar converts |
| 3402 | // it to unsigned char regardless. |
| 3403 | Value *IntChar = ConstantInt::get(Ty: IntTy, V: (unsigned char)OperandStr[0]); |
| 3404 | return copyFlags(Old: *CI, New: emitPutChar(Char: IntChar, B, TLI)); |
| 3405 | } |
| 3406 | // printf("%s", str"\n") --> puts(str) |
| 3407 | if (OperandStr.back() == '\n') { |
| 3408 | OperandStr = OperandStr.drop_back(); |
| 3409 | Value *GV = B.CreateGlobalString(Str: OperandStr, Name: "str" ); |
| 3410 | return copyFlags(Old: *CI, New: emitPutS(Str: GV, B, TLI)); |
| 3411 | } |
| 3412 | return nullptr; |
| 3413 | } |
| 3414 | |
| 3415 | // printf("foo\n") --> puts("foo") |
| 3416 | if (FormatStr.back() == '\n' && |
| 3417 | !FormatStr.contains(C: '%')) { // No format characters. |
| 3418 | // Create a string literal with no \n on it. We expect the constant merge |
| 3419 | // pass to be run after this pass, to merge duplicate strings. |
| 3420 | FormatStr = FormatStr.drop_back(); |
| 3421 | Value *GV = B.CreateGlobalString(Str: FormatStr, Name: "str" ); |
| 3422 | return copyFlags(Old: *CI, New: emitPutS(Str: GV, B, TLI)); |
| 3423 | } |
| 3424 | |
| 3425 | // Optimize specific format strings. |
| 3426 | // printf("%c", chr) --> putchar(chr) |
| 3427 | if (FormatStr == "%c" && CI->arg_size() > 1 && |
| 3428 | CI->getArgOperand(i: 1)->getType()->isIntegerTy()) { |
| 3429 | // Convert the argument to the type expected by putchar, i.e., int, which |
| 3430 | // need not be 32 bits wide but which is the same as printf's return type. |
| 3431 | Value *IntChar = B.CreateIntCast(V: CI->getArgOperand(i: 1), DestTy: IntTy, isSigned: false); |
| 3432 | return copyFlags(Old: *CI, New: emitPutChar(Char: IntChar, B, TLI)); |
| 3433 | } |
| 3434 | |
| 3435 | // printf("%s\n", str) --> puts(str) |
| 3436 | if (FormatStr == "%s\n" && CI->arg_size() > 1 && |
| 3437 | CI->getArgOperand(i: 1)->getType()->isPointerTy()) |
| 3438 | return copyFlags(Old: *CI, New: emitPutS(Str: CI->getArgOperand(i: 1), B, TLI)); |
| 3439 | return nullptr; |
| 3440 | } |
| 3441 | |
| 3442 | Value *LibCallSimplifier::optimizePrintF(CallInst *CI, IRBuilderBase &B) { |
| 3443 | |
| 3444 | Module *M = CI->getModule(); |
| 3445 | Function *Callee = CI->getCalledFunction(); |
| 3446 | FunctionType *FT = Callee->getFunctionType(); |
| 3447 | if (Value *V = optimizePrintFString(CI, B)) { |
| 3448 | return V; |
| 3449 | } |
| 3450 | |
| 3451 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: 0); |
| 3452 | |
| 3453 | // printf(format, ...) -> iprintf(format, ...) if no floating point |
| 3454 | // arguments. |
| 3455 | if (isLibFuncEmittable(M, TLI, TheLibFunc: LibFunc_iprintf) && |
| 3456 | !callHasFloatingPointArgument(CI)) { |
| 3457 | FunctionCallee IPrintFFn = getOrInsertLibFunc(M, TLI: *TLI, TheLibFunc: LibFunc_iprintf, T: FT, |
| 3458 | AttributeList: Callee->getAttributes()); |
| 3459 | CallInst *New = cast<CallInst>(Val: CI->clone()); |
| 3460 | New->setCalledFunction(IPrintFFn); |
| 3461 | B.Insert(I: New); |
| 3462 | return New; |
| 3463 | } |
| 3464 | |
| 3465 | // printf(format, ...) -> __small_printf(format, ...) if no 128-bit floating point |
| 3466 | // arguments. |
| 3467 | if (isLibFuncEmittable(M, TLI, TheLibFunc: LibFunc_small_printf) && |
| 3468 | !callHasFP128Argument(CI)) { |
| 3469 | auto SmallPrintFFn = getOrInsertLibFunc(M, TLI: *TLI, TheLibFunc: LibFunc_small_printf, T: FT, |
| 3470 | AttributeList: Callee->getAttributes()); |
| 3471 | CallInst *New = cast<CallInst>(Val: CI->clone()); |
| 3472 | New->setCalledFunction(SmallPrintFFn); |
| 3473 | B.Insert(I: New); |
| 3474 | return New; |
| 3475 | } |
| 3476 | |
| 3477 | return nullptr; |
| 3478 | } |
| 3479 | |
| 3480 | Value *LibCallSimplifier::optimizeSPrintFString(CallInst *CI, |
| 3481 | IRBuilderBase &B) { |
| 3482 | // Check for a fixed format string. |
| 3483 | StringRef FormatStr; |
| 3484 | if (!getConstantStringInfo(V: CI->getArgOperand(i: 1), Str&: FormatStr)) |
| 3485 | return nullptr; |
| 3486 | |
| 3487 | // If we just have a format string (nothing else crazy) transform it. |
| 3488 | Value *Dest = CI->getArgOperand(i: 0); |
| 3489 | if (CI->arg_size() == 2) { |
| 3490 | // Make sure there's no % in the constant array. We could try to handle |
| 3491 | // %% -> % in the future if we cared. |
| 3492 | if (FormatStr.contains(C: '%')) |
| 3493 | return nullptr; // we found a format specifier, bail out. |
| 3494 | |
| 3495 | // sprintf(str, fmt) -> llvm.memcpy(align 1 str, align 1 fmt, strlen(fmt)+1) |
| 3496 | B.CreateMemCpy(Dst: Dest, DstAlign: Align(1), Src: CI->getArgOperand(i: 1), SrcAlign: Align(1), |
| 3497 | // Copy the null byte. |
| 3498 | Size: TLI->getAsSizeT(V: FormatStr.size() + 1, M: *CI->getModule())); |
| 3499 | return ConstantInt::get(Ty: CI->getType(), V: FormatStr.size()); |
| 3500 | } |
| 3501 | |
| 3502 | // The remaining optimizations require the format string to be "%s" or "%c" |
| 3503 | // and have an extra operand. |
| 3504 | if (FormatStr.size() != 2 || FormatStr[0] != '%' || CI->arg_size() < 3) |
| 3505 | return nullptr; |
| 3506 | |
| 3507 | // Decode the second character of the format string. |
| 3508 | if (FormatStr[1] == 'c') { |
| 3509 | // sprintf(dst, "%c", chr) --> *(i8*)dst = chr; *((i8*)dst+1) = 0 |
| 3510 | if (!CI->getArgOperand(i: 2)->getType()->isIntegerTy()) |
| 3511 | return nullptr; |
| 3512 | Value *V = B.CreateTrunc(V: CI->getArgOperand(i: 2), DestTy: B.getInt8Ty(), Name: "char" ); |
| 3513 | Value *Ptr = Dest; |
| 3514 | B.CreateStore(Val: V, Ptr); |
| 3515 | Ptr = B.CreateInBoundsGEP(Ty: B.getInt8Ty(), Ptr, IdxList: B.getInt32(C: 1), Name: "nul" ); |
| 3516 | B.CreateStore(Val: B.getInt8(C: 0), Ptr); |
| 3517 | |
| 3518 | return ConstantInt::get(Ty: CI->getType(), V: 1); |
| 3519 | } |
| 3520 | |
| 3521 | if (FormatStr[1] == 's') { |
| 3522 | // sprintf(dest, "%s", str) -> llvm.memcpy(align 1 dest, align 1 str, |
| 3523 | // strlen(str)+1) |
| 3524 | if (!CI->getArgOperand(i: 2)->getType()->isPointerTy()) |
| 3525 | return nullptr; |
| 3526 | |
| 3527 | if (CI->use_empty()) |
| 3528 | // sprintf(dest, "%s", str) -> strcpy(dest, str) |
| 3529 | return copyFlags(Old: *CI, New: emitStrCpy(Dst: Dest, Src: CI->getArgOperand(i: 2), B, TLI)); |
| 3530 | |
| 3531 | uint64_t SrcLen = GetStringLength(V: CI->getArgOperand(i: 2)); |
| 3532 | if (SrcLen) { |
| 3533 | B.CreateMemCpy(Dst: Dest, DstAlign: Align(1), Src: CI->getArgOperand(i: 2), SrcAlign: Align(1), |
| 3534 | Size: TLI->getAsSizeT(V: SrcLen, M: *CI->getModule())); |
| 3535 | // Returns total number of characters written without null-character. |
| 3536 | return ConstantInt::get(Ty: CI->getType(), V: SrcLen - 1); |
| 3537 | } else if (Value *V = emitStpCpy(Dst: Dest, Src: CI->getArgOperand(i: 2), B, TLI)) { |
| 3538 | // sprintf(dest, "%s", str) -> stpcpy(dest, str) - dest |
| 3539 | Value *PtrDiff = B.CreatePtrDiff(ElemTy: B.getInt8Ty(), LHS: V, RHS: Dest); |
| 3540 | return B.CreateIntCast(V: PtrDiff, DestTy: CI->getType(), isSigned: false); |
| 3541 | } |
| 3542 | |
| 3543 | if (llvm::shouldOptimizeForSize(BB: CI->getParent(), PSI, BFI, |
| 3544 | QueryType: PGSOQueryType::IRPass)) |
| 3545 | return nullptr; |
| 3546 | |
| 3547 | Value *Len = emitStrLen(Ptr: CI->getArgOperand(i: 2), B, DL, TLI); |
| 3548 | if (!Len) |
| 3549 | return nullptr; |
| 3550 | Value *IncLen = |
| 3551 | B.CreateAdd(LHS: Len, RHS: ConstantInt::get(Ty: Len->getType(), V: 1), Name: "leninc" ); |
| 3552 | B.CreateMemCpy(Dst: Dest, DstAlign: Align(1), Src: CI->getArgOperand(i: 2), SrcAlign: Align(1), Size: IncLen); |
| 3553 | |
| 3554 | // The sprintf result is the unincremented number of bytes in the string. |
| 3555 | return B.CreateIntCast(V: Len, DestTy: CI->getType(), isSigned: false); |
| 3556 | } |
| 3557 | return nullptr; |
| 3558 | } |
| 3559 | |
| 3560 | Value *LibCallSimplifier::optimizeSPrintF(CallInst *CI, IRBuilderBase &B) { |
| 3561 | Module *M = CI->getModule(); |
| 3562 | Function *Callee = CI->getCalledFunction(); |
| 3563 | FunctionType *FT = Callee->getFunctionType(); |
| 3564 | if (Value *V = optimizeSPrintFString(CI, B)) { |
| 3565 | return V; |
| 3566 | } |
| 3567 | |
| 3568 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: {0, 1}); |
| 3569 | |
| 3570 | // sprintf(str, format, ...) -> siprintf(str, format, ...) if no floating |
| 3571 | // point arguments. |
| 3572 | if (isLibFuncEmittable(M, TLI, TheLibFunc: LibFunc_siprintf) && |
| 3573 | !callHasFloatingPointArgument(CI)) { |
| 3574 | FunctionCallee SIPrintFFn = getOrInsertLibFunc(M, TLI: *TLI, TheLibFunc: LibFunc_siprintf, |
| 3575 | T: FT, AttributeList: Callee->getAttributes()); |
| 3576 | CallInst *New = cast<CallInst>(Val: CI->clone()); |
| 3577 | New->setCalledFunction(SIPrintFFn); |
| 3578 | B.Insert(I: New); |
| 3579 | return New; |
| 3580 | } |
| 3581 | |
| 3582 | // sprintf(str, format, ...) -> __small_sprintf(str, format, ...) if no 128-bit |
| 3583 | // floating point arguments. |
| 3584 | if (isLibFuncEmittable(M, TLI, TheLibFunc: LibFunc_small_sprintf) && |
| 3585 | !callHasFP128Argument(CI)) { |
| 3586 | auto SmallSPrintFFn = getOrInsertLibFunc(M, TLI: *TLI, TheLibFunc: LibFunc_small_sprintf, T: FT, |
| 3587 | AttributeList: Callee->getAttributes()); |
| 3588 | CallInst *New = cast<CallInst>(Val: CI->clone()); |
| 3589 | New->setCalledFunction(SmallSPrintFFn); |
| 3590 | B.Insert(I: New); |
| 3591 | return New; |
| 3592 | } |
| 3593 | |
| 3594 | return nullptr; |
| 3595 | } |
| 3596 | |
| 3597 | // Transform an snprintf call CI with the bound N to format the string Str |
| 3598 | // either to a call to memcpy, or to single character a store, or to nothing, |
| 3599 | // and fold the result to a constant. A nonnull StrArg refers to the string |
| 3600 | // argument being formatted. Otherwise the call is one with N < 2 and |
| 3601 | // the "%c" directive to format a single character. |
| 3602 | Value *LibCallSimplifier::emitSnPrintfMemCpy(CallInst *CI, Value *StrArg, |
| 3603 | StringRef Str, uint64_t N, |
| 3604 | IRBuilderBase &B) { |
| 3605 | assert(StrArg || (N < 2 && Str.size() == 1)); |
| 3606 | |
| 3607 | unsigned IntBits = TLI->getIntSize(); |
| 3608 | uint64_t IntMax = maxIntN(N: IntBits); |
| 3609 | if (Str.size() > IntMax) |
| 3610 | // Bail if the string is longer than INT_MAX. POSIX requires |
| 3611 | // implementations to set errno to EOVERFLOW in this case, in |
| 3612 | // addition to when N is larger than that (checked by the caller). |
| 3613 | return nullptr; |
| 3614 | |
| 3615 | Value *StrLen = ConstantInt::get(Ty: CI->getType(), V: Str.size()); |
| 3616 | if (N == 0) |
| 3617 | return StrLen; |
| 3618 | |
| 3619 | // Set to the number of bytes to copy fron StrArg which is also |
| 3620 | // the offset of the terinating nul. |
| 3621 | uint64_t NCopy; |
| 3622 | if (N > Str.size()) |
| 3623 | // Copy the full string, including the terminating nul (which must |
| 3624 | // be present regardless of the bound). |
| 3625 | NCopy = Str.size() + 1; |
| 3626 | else |
| 3627 | NCopy = N - 1; |
| 3628 | |
| 3629 | Value *DstArg = CI->getArgOperand(i: 0); |
| 3630 | if (NCopy && StrArg) |
| 3631 | // Transform the call to lvm.memcpy(dst, fmt, N). |
| 3632 | copyFlags(Old: *CI, New: B.CreateMemCpy(Dst: DstArg, DstAlign: Align(1), Src: StrArg, SrcAlign: Align(1), |
| 3633 | Size: TLI->getAsSizeT(V: NCopy, M: *CI->getModule()))); |
| 3634 | |
| 3635 | if (N > Str.size()) |
| 3636 | // Return early when the whole format string, including the final nul, |
| 3637 | // has been copied. |
| 3638 | return StrLen; |
| 3639 | |
| 3640 | // Otherwise, when truncating the string append a terminating nul. |
| 3641 | Type *Int8Ty = B.getInt8Ty(); |
| 3642 | Value *NulOff = B.getIntN(N: IntBits, C: NCopy); |
| 3643 | Value *DstEnd = B.CreateInBoundsGEP(Ty: Int8Ty, Ptr: DstArg, IdxList: NulOff, Name: "endptr" ); |
| 3644 | B.CreateStore(Val: ConstantInt::get(Ty: Int8Ty, V: 0), Ptr: DstEnd); |
| 3645 | return StrLen; |
| 3646 | } |
| 3647 | |
| 3648 | Value *LibCallSimplifier::optimizeSnPrintFString(CallInst *CI, |
| 3649 | IRBuilderBase &B) { |
| 3650 | // Check for size |
| 3651 | ConstantInt *Size = dyn_cast<ConstantInt>(Val: CI->getArgOperand(i: 1)); |
| 3652 | if (!Size) |
| 3653 | return nullptr; |
| 3654 | |
| 3655 | uint64_t N = Size->getZExtValue(); |
| 3656 | uint64_t IntMax = maxIntN(N: TLI->getIntSize()); |
| 3657 | if (N > IntMax) |
| 3658 | // Bail if the bound exceeds INT_MAX. POSIX requires implementations |
| 3659 | // to set errno to EOVERFLOW in this case. |
| 3660 | return nullptr; |
| 3661 | |
| 3662 | Value *DstArg = CI->getArgOperand(i: 0); |
| 3663 | Value *FmtArg = CI->getArgOperand(i: 2); |
| 3664 | |
| 3665 | // Check for a fixed format string. |
| 3666 | StringRef FormatStr; |
| 3667 | if (!getConstantStringInfo(V: FmtArg, Str&: FormatStr)) |
| 3668 | return nullptr; |
| 3669 | |
| 3670 | // If we just have a format string (nothing else crazy) transform it. |
| 3671 | if (CI->arg_size() == 3) { |
| 3672 | if (FormatStr.contains(C: '%')) |
| 3673 | // Bail if the format string contains a directive and there are |
| 3674 | // no arguments. We could handle "%%" in the future. |
| 3675 | return nullptr; |
| 3676 | |
| 3677 | return emitSnPrintfMemCpy(CI, StrArg: FmtArg, Str: FormatStr, N, B); |
| 3678 | } |
| 3679 | |
| 3680 | // The remaining optimizations require the format string to be "%s" or "%c" |
| 3681 | // and have an extra operand. |
| 3682 | if (FormatStr.size() != 2 || FormatStr[0] != '%' || CI->arg_size() != 4) |
| 3683 | return nullptr; |
| 3684 | |
| 3685 | // Decode the second character of the format string. |
| 3686 | if (FormatStr[1] == 'c') { |
| 3687 | if (N <= 1) { |
| 3688 | // Use an arbitary string of length 1 to transform the call into |
| 3689 | // either a nul store (N == 1) or a no-op (N == 0) and fold it |
| 3690 | // to one. |
| 3691 | StringRef CharStr("*" ); |
| 3692 | return emitSnPrintfMemCpy(CI, StrArg: nullptr, Str: CharStr, N, B); |
| 3693 | } |
| 3694 | |
| 3695 | // snprintf(dst, size, "%c", chr) --> *(i8*)dst = chr; *((i8*)dst+1) = 0 |
| 3696 | if (!CI->getArgOperand(i: 3)->getType()->isIntegerTy()) |
| 3697 | return nullptr; |
| 3698 | Value *V = B.CreateTrunc(V: CI->getArgOperand(i: 3), DestTy: B.getInt8Ty(), Name: "char" ); |
| 3699 | Value *Ptr = DstArg; |
| 3700 | B.CreateStore(Val: V, Ptr); |
| 3701 | Ptr = B.CreateInBoundsGEP(Ty: B.getInt8Ty(), Ptr, IdxList: B.getInt32(C: 1), Name: "nul" ); |
| 3702 | B.CreateStore(Val: B.getInt8(C: 0), Ptr); |
| 3703 | return ConstantInt::get(Ty: CI->getType(), V: 1); |
| 3704 | } |
| 3705 | |
| 3706 | if (FormatStr[1] != 's') |
| 3707 | return nullptr; |
| 3708 | |
| 3709 | Value *StrArg = CI->getArgOperand(i: 3); |
| 3710 | // snprintf(dest, size, "%s", str) to llvm.memcpy(dest, str, len+1, 1) |
| 3711 | StringRef Str; |
| 3712 | if (!getConstantStringInfo(V: StrArg, Str)) |
| 3713 | return nullptr; |
| 3714 | |
| 3715 | return emitSnPrintfMemCpy(CI, StrArg, Str, N, B); |
| 3716 | } |
| 3717 | |
| 3718 | Value *LibCallSimplifier::optimizeSnPrintF(CallInst *CI, IRBuilderBase &B) { |
| 3719 | if (Value *V = optimizeSnPrintFString(CI, B)) { |
| 3720 | return V; |
| 3721 | } |
| 3722 | |
| 3723 | if (isKnownNonZero(V: CI->getOperand(i_nocapture: 1), Q: DL)) |
| 3724 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: 0); |
| 3725 | return nullptr; |
| 3726 | } |
| 3727 | |
| 3728 | Value *LibCallSimplifier::optimizeFPrintFString(CallInst *CI, |
| 3729 | IRBuilderBase &B) { |
| 3730 | optimizeErrorReporting(CI, B, StreamArg: 0); |
| 3731 | |
| 3732 | // All the optimizations depend on the format string. |
| 3733 | StringRef FormatStr; |
| 3734 | if (!getConstantStringInfo(V: CI->getArgOperand(i: 1), Str&: FormatStr)) |
| 3735 | return nullptr; |
| 3736 | |
| 3737 | // Do not do any of the following transformations if the fprintf return |
| 3738 | // value is used, in general the fprintf return value is not compatible |
| 3739 | // with fwrite(), fputc() or fputs(). |
| 3740 | if (!CI->use_empty()) |
| 3741 | return nullptr; |
| 3742 | |
| 3743 | // fprintf(F, "foo") --> fwrite("foo", 3, 1, F) |
| 3744 | if (CI->arg_size() == 2) { |
| 3745 | // Could handle %% -> % if we cared. |
| 3746 | if (FormatStr.contains(C: '%')) |
| 3747 | return nullptr; // We found a format specifier. |
| 3748 | |
| 3749 | return copyFlags( |
| 3750 | Old: *CI, New: emitFWrite(Ptr: CI->getArgOperand(i: 1), |
| 3751 | Size: TLI->getAsSizeT(V: FormatStr.size(), M: *CI->getModule()), |
| 3752 | File: CI->getArgOperand(i: 0), B, DL, TLI)); |
| 3753 | } |
| 3754 | |
| 3755 | // The remaining optimizations require the format string to be "%s" or "%c" |
| 3756 | // and have an extra operand. |
| 3757 | if (FormatStr.size() != 2 || FormatStr[0] != '%' || CI->arg_size() < 3) |
| 3758 | return nullptr; |
| 3759 | |
| 3760 | // Decode the second character of the format string. |
| 3761 | if (FormatStr[1] == 'c') { |
| 3762 | // fprintf(F, "%c", chr) --> fputc((int)chr, F) |
| 3763 | if (!CI->getArgOperand(i: 2)->getType()->isIntegerTy()) |
| 3764 | return nullptr; |
| 3765 | Type *IntTy = B.getIntNTy(N: TLI->getIntSize()); |
| 3766 | Value *V = B.CreateIntCast(V: CI->getArgOperand(i: 2), DestTy: IntTy, /*isSigned*/ true, |
| 3767 | Name: "chari" ); |
| 3768 | return copyFlags(Old: *CI, New: emitFPutC(Char: V, File: CI->getArgOperand(i: 0), B, TLI)); |
| 3769 | } |
| 3770 | |
| 3771 | if (FormatStr[1] == 's') { |
| 3772 | // fprintf(F, "%s", str) --> fputs(str, F) |
| 3773 | if (!CI->getArgOperand(i: 2)->getType()->isPointerTy()) |
| 3774 | return nullptr; |
| 3775 | return copyFlags( |
| 3776 | Old: *CI, New: emitFPutS(Str: CI->getArgOperand(i: 2), File: CI->getArgOperand(i: 0), B, TLI)); |
| 3777 | } |
| 3778 | return nullptr; |
| 3779 | } |
| 3780 | |
| 3781 | Value *LibCallSimplifier::optimizeFPrintF(CallInst *CI, IRBuilderBase &B) { |
| 3782 | Module *M = CI->getModule(); |
| 3783 | Function *Callee = CI->getCalledFunction(); |
| 3784 | FunctionType *FT = Callee->getFunctionType(); |
| 3785 | if (Value *V = optimizeFPrintFString(CI, B)) { |
| 3786 | return V; |
| 3787 | } |
| 3788 | |
| 3789 | // fprintf(stream, format, ...) -> fiprintf(stream, format, ...) if no |
| 3790 | // floating point arguments. |
| 3791 | if (isLibFuncEmittable(M, TLI, TheLibFunc: LibFunc_fiprintf) && |
| 3792 | !callHasFloatingPointArgument(CI)) { |
| 3793 | FunctionCallee FIPrintFFn = getOrInsertLibFunc(M, TLI: *TLI, TheLibFunc: LibFunc_fiprintf, |
| 3794 | T: FT, AttributeList: Callee->getAttributes()); |
| 3795 | CallInst *New = cast<CallInst>(Val: CI->clone()); |
| 3796 | New->setCalledFunction(FIPrintFFn); |
| 3797 | B.Insert(I: New); |
| 3798 | return New; |
| 3799 | } |
| 3800 | |
| 3801 | // fprintf(stream, format, ...) -> __small_fprintf(stream, format, ...) if no |
| 3802 | // 128-bit floating point arguments. |
| 3803 | if (isLibFuncEmittable(M, TLI, TheLibFunc: LibFunc_small_fprintf) && |
| 3804 | !callHasFP128Argument(CI)) { |
| 3805 | auto SmallFPrintFFn = |
| 3806 | getOrInsertLibFunc(M, TLI: *TLI, TheLibFunc: LibFunc_small_fprintf, T: FT, |
| 3807 | AttributeList: Callee->getAttributes()); |
| 3808 | CallInst *New = cast<CallInst>(Val: CI->clone()); |
| 3809 | New->setCalledFunction(SmallFPrintFFn); |
| 3810 | B.Insert(I: New); |
| 3811 | return New; |
| 3812 | } |
| 3813 | |
| 3814 | return nullptr; |
| 3815 | } |
| 3816 | |
| 3817 | Value *LibCallSimplifier::optimizeFWrite(CallInst *CI, IRBuilderBase &B) { |
| 3818 | optimizeErrorReporting(CI, B, StreamArg: 3); |
| 3819 | |
| 3820 | // Get the element size and count. |
| 3821 | ConstantInt *SizeC = dyn_cast<ConstantInt>(Val: CI->getArgOperand(i: 1)); |
| 3822 | ConstantInt *CountC = dyn_cast<ConstantInt>(Val: CI->getArgOperand(i: 2)); |
| 3823 | if (SizeC && CountC) { |
| 3824 | uint64_t Bytes = SizeC->getZExtValue() * CountC->getZExtValue(); |
| 3825 | |
| 3826 | // If this is writing zero records, remove the call (it's a noop). |
| 3827 | if (Bytes == 0) |
| 3828 | return ConstantInt::get(Ty: CI->getType(), V: 0); |
| 3829 | |
| 3830 | // If this is writing one byte, turn it into fputc. |
| 3831 | // This optimisation is only valid, if the return value is unused. |
| 3832 | if (Bytes == 1 && CI->use_empty()) { // fwrite(S,1,1,F) -> fputc(S[0],F) |
| 3833 | Value *Char = B.CreateLoad(Ty: B.getInt8Ty(), Ptr: CI->getArgOperand(i: 0), Name: "char" ); |
| 3834 | Type *IntTy = B.getIntNTy(N: TLI->getIntSize()); |
| 3835 | Value *Cast = B.CreateIntCast(V: Char, DestTy: IntTy, /*isSigned*/ true, Name: "chari" ); |
| 3836 | Value *NewCI = emitFPutC(Char: Cast, File: CI->getArgOperand(i: 3), B, TLI); |
| 3837 | return NewCI ? ConstantInt::get(Ty: CI->getType(), V: 1) : nullptr; |
| 3838 | } |
| 3839 | } |
| 3840 | |
| 3841 | return nullptr; |
| 3842 | } |
| 3843 | |
| 3844 | Value *LibCallSimplifier::optimizeFPuts(CallInst *CI, IRBuilderBase &B) { |
| 3845 | optimizeErrorReporting(CI, B, StreamArg: 1); |
| 3846 | |
| 3847 | // Don't rewrite fputs to fwrite when optimising for size because fwrite |
| 3848 | // requires more arguments and thus extra MOVs are required. |
| 3849 | if (llvm::shouldOptimizeForSize(BB: CI->getParent(), PSI, BFI, |
| 3850 | QueryType: PGSOQueryType::IRPass)) |
| 3851 | return nullptr; |
| 3852 | |
| 3853 | // We can't optimize if return value is used. |
| 3854 | if (!CI->use_empty()) |
| 3855 | return nullptr; |
| 3856 | |
| 3857 | // fputs(s,F) --> fwrite(s,strlen(s),1,F) |
| 3858 | uint64_t Len = GetStringLength(V: CI->getArgOperand(i: 0)); |
| 3859 | if (!Len) |
| 3860 | return nullptr; |
| 3861 | |
| 3862 | // Known to have no uses (see above). |
| 3863 | unsigned SizeTBits = TLI->getSizeTSize(M: *CI->getModule()); |
| 3864 | Type *SizeTTy = IntegerType::get(C&: CI->getContext(), NumBits: SizeTBits); |
| 3865 | return copyFlags( |
| 3866 | Old: *CI, |
| 3867 | New: emitFWrite(Ptr: CI->getArgOperand(i: 0), |
| 3868 | Size: ConstantInt::get(Ty: SizeTTy, V: Len - 1), |
| 3869 | File: CI->getArgOperand(i: 1), B, DL, TLI)); |
| 3870 | } |
| 3871 | |
| 3872 | Value *LibCallSimplifier::optimizePuts(CallInst *CI, IRBuilderBase &B) { |
| 3873 | annotateNonNullNoUndefBasedOnAccess(CI, ArgNos: 0); |
| 3874 | if (!CI->use_empty()) |
| 3875 | return nullptr; |
| 3876 | |
| 3877 | // Check for a constant string. |
| 3878 | // puts("") -> putchar('\n') |
| 3879 | StringRef Str; |
| 3880 | if (getConstantStringInfo(V: CI->getArgOperand(i: 0), Str) && Str.empty()) { |
| 3881 | // putchar takes an argument of the same type as puts returns, i.e., |
| 3882 | // int, which need not be 32 bits wide. |
| 3883 | Type *IntTy = CI->getType(); |
| 3884 | return copyFlags(Old: *CI, New: emitPutChar(Char: ConstantInt::get(Ty: IntTy, V: '\n'), B, TLI)); |
| 3885 | } |
| 3886 | |
| 3887 | return nullptr; |
| 3888 | } |
| 3889 | |
| 3890 | Value *LibCallSimplifier::optimizeExit(CallInst *CI) { |
| 3891 | |
| 3892 | // Mark 'exit' as cold if its not exit(0) (success). |
| 3893 | const APInt *C; |
| 3894 | if (!CI->hasFnAttr(Kind: Attribute::Cold) && |
| 3895 | match(V: CI->getArgOperand(i: 0), P: m_APInt(Res&: C)) && !C->isZero()) { |
| 3896 | CI->addFnAttr(Kind: Attribute::Cold); |
| 3897 | } |
| 3898 | return nullptr; |
| 3899 | } |
| 3900 | |
| 3901 | Value *LibCallSimplifier::optimizeBCopy(CallInst *CI, IRBuilderBase &B) { |
| 3902 | // bcopy(src, dst, n) -> llvm.memmove(dst, src, n) |
| 3903 | return copyFlags(Old: *CI, New: B.CreateMemMove(Dst: CI->getArgOperand(i: 1), DstAlign: Align(1), |
| 3904 | Src: CI->getArgOperand(i: 0), SrcAlign: Align(1), |
| 3905 | Size: CI->getArgOperand(i: 2))); |
| 3906 | } |
| 3907 | |
| 3908 | bool LibCallSimplifier::hasFloatVersion(const Module *M, StringRef FuncName) { |
| 3909 | SmallString<20> FloatFuncName = FuncName; |
| 3910 | FloatFuncName += 'f'; |
| 3911 | return isLibFuncEmittable(M, TLI, Name: FloatFuncName); |
| 3912 | } |
| 3913 | |
| 3914 | Value *LibCallSimplifier::optimizeStringMemoryLibCall(CallInst *CI, |
| 3915 | IRBuilderBase &Builder) { |
| 3916 | Module *M = CI->getModule(); |
| 3917 | LibFunc Func; |
| 3918 | Function *Callee = CI->getCalledFunction(); |
| 3919 | |
| 3920 | // Check for string/memory library functions. |
| 3921 | if (TLI->getLibFunc(FDecl: *Callee, F&: Func) && isLibFuncEmittable(M, TLI, TheLibFunc: Func)) { |
| 3922 | // Make sure we never change the calling convention. |
| 3923 | assert( |
| 3924 | (ignoreCallingConv(Func) || |
| 3925 | TargetLibraryInfoImpl::isCallingConvCCompatible(CI)) && |
| 3926 | "Optimizing string/memory libcall would change the calling convention" ); |
| 3927 | switch (Func) { |
| 3928 | case LibFunc_strcat: |
| 3929 | return optimizeStrCat(CI, B&: Builder); |
| 3930 | case LibFunc_strncat: |
| 3931 | return optimizeStrNCat(CI, B&: Builder); |
| 3932 | case LibFunc_strchr: |
| 3933 | return optimizeStrChr(CI, B&: Builder); |
| 3934 | case LibFunc_strrchr: |
| 3935 | return optimizeStrRChr(CI, B&: Builder); |
| 3936 | case LibFunc_strcmp: |
| 3937 | return optimizeStrCmp(CI, B&: Builder); |
| 3938 | case LibFunc_strncmp: |
| 3939 | return optimizeStrNCmp(CI, B&: Builder); |
| 3940 | case LibFunc_strcpy: |
| 3941 | return optimizeStrCpy(CI, B&: Builder); |
| 3942 | case LibFunc_stpcpy: |
| 3943 | return optimizeStpCpy(CI, B&: Builder); |
| 3944 | case LibFunc_strlcpy: |
| 3945 | return optimizeStrLCpy(CI, B&: Builder); |
| 3946 | case LibFunc_stpncpy: |
| 3947 | return optimizeStringNCpy(CI, /*RetEnd=*/true, B&: Builder); |
| 3948 | case LibFunc_strncpy: |
| 3949 | return optimizeStringNCpy(CI, /*RetEnd=*/false, B&: Builder); |
| 3950 | case LibFunc_strlen: |
| 3951 | return optimizeStrLen(CI, B&: Builder); |
| 3952 | case LibFunc_strnlen: |
| 3953 | return optimizeStrNLen(CI, B&: Builder); |
| 3954 | case LibFunc_strpbrk: |
| 3955 | return optimizeStrPBrk(CI, B&: Builder); |
| 3956 | case LibFunc_strndup: |
| 3957 | return optimizeStrNDup(CI, B&: Builder); |
| 3958 | case LibFunc_strtol: |
| 3959 | case LibFunc_strtod: |
| 3960 | case LibFunc_strtof: |
| 3961 | case LibFunc_strtoul: |
| 3962 | case LibFunc_strtoll: |
| 3963 | case LibFunc_strtold: |
| 3964 | case LibFunc_strtoull: |
| 3965 | return optimizeStrTo(CI, B&: Builder); |
| 3966 | case LibFunc_strspn: |
| 3967 | return optimizeStrSpn(CI, B&: Builder); |
| 3968 | case LibFunc_strcspn: |
| 3969 | return optimizeStrCSpn(CI, B&: Builder); |
| 3970 | case LibFunc_strstr: |
| 3971 | return optimizeStrStr(CI, B&: Builder); |
| 3972 | case LibFunc_memchr: |
| 3973 | return optimizeMemChr(CI, B&: Builder); |
| 3974 | case LibFunc_memrchr: |
| 3975 | return optimizeMemRChr(CI, B&: Builder); |
| 3976 | case LibFunc_bcmp: |
| 3977 | return optimizeBCmp(CI, B&: Builder); |
| 3978 | case LibFunc_memcmp: |
| 3979 | return optimizeMemCmp(CI, B&: Builder); |
| 3980 | case LibFunc_memcpy: |
| 3981 | return optimizeMemCpy(CI, B&: Builder); |
| 3982 | case LibFunc_memccpy: |
| 3983 | return optimizeMemCCpy(CI, B&: Builder); |
| 3984 | case LibFunc_mempcpy: |
| 3985 | return optimizeMemPCpy(CI, B&: Builder); |
| 3986 | case LibFunc_memmove: |
| 3987 | return optimizeMemMove(CI, B&: Builder); |
| 3988 | case LibFunc_memset: |
| 3989 | return optimizeMemSet(CI, B&: Builder); |
| 3990 | case LibFunc_realloc: |
| 3991 | return optimizeRealloc(CI, B&: Builder); |
| 3992 | case LibFunc_wcslen: |
| 3993 | return optimizeWcslen(CI, B&: Builder); |
| 3994 | case LibFunc_bcopy: |
| 3995 | return optimizeBCopy(CI, B&: Builder); |
| 3996 | case LibFunc_Znwm: |
| 3997 | case LibFunc_ZnwmRKSt9nothrow_t: |
| 3998 | case LibFunc_ZnwmSt11align_val_t: |
| 3999 | case LibFunc_ZnwmSt11align_val_tRKSt9nothrow_t: |
| 4000 | case LibFunc_Znam: |
| 4001 | case LibFunc_ZnamRKSt9nothrow_t: |
| 4002 | case LibFunc_ZnamSt11align_val_t: |
| 4003 | case LibFunc_ZnamSt11align_val_tRKSt9nothrow_t: |
| 4004 | case LibFunc_Znwm12__hot_cold_t: |
| 4005 | case LibFunc_ZnwmRKSt9nothrow_t12__hot_cold_t: |
| 4006 | case LibFunc_ZnwmSt11align_val_t12__hot_cold_t: |
| 4007 | case LibFunc_ZnwmSt11align_val_tRKSt9nothrow_t12__hot_cold_t: |
| 4008 | case LibFunc_Znam12__hot_cold_t: |
| 4009 | case LibFunc_ZnamRKSt9nothrow_t12__hot_cold_t: |
| 4010 | case LibFunc_ZnamSt11align_val_t12__hot_cold_t: |
| 4011 | case LibFunc_ZnamSt11align_val_tRKSt9nothrow_t12__hot_cold_t: |
| 4012 | case LibFunc_size_returning_new: |
| 4013 | case LibFunc_size_returning_new_hot_cold: |
| 4014 | case LibFunc_size_returning_new_aligned: |
| 4015 | case LibFunc_size_returning_new_aligned_hot_cold: |
| 4016 | return optimizeNew(CI, B&: Builder, Func); |
| 4017 | default: |
| 4018 | break; |
| 4019 | } |
| 4020 | } |
| 4021 | return nullptr; |
| 4022 | } |
| 4023 | |
| 4024 | /// Constant folding nan/nanf/nanl. |
| 4025 | static Value *optimizeNaN(CallInst *CI) { |
| 4026 | StringRef CharSeq; |
| 4027 | if (!getConstantStringInfo(V: CI->getArgOperand(i: 0), Str&: CharSeq)) |
| 4028 | return nullptr; |
| 4029 | |
| 4030 | APInt Fill; |
| 4031 | // Treat empty strings as if they were zero. |
| 4032 | if (CharSeq.empty()) |
| 4033 | Fill = APInt(32, 0); |
| 4034 | else if (CharSeq.getAsInteger(Radix: 0, Result&: Fill)) |
| 4035 | return nullptr; |
| 4036 | |
| 4037 | return ConstantFP::getQNaN(Ty: CI->getType(), /*Negative=*/false, Payload: &Fill); |
| 4038 | } |
| 4039 | |
| 4040 | Value *LibCallSimplifier::optimizeFloatingPointLibCall(CallInst *CI, |
| 4041 | LibFunc Func, |
| 4042 | IRBuilderBase &Builder) { |
| 4043 | const Module *M = CI->getModule(); |
| 4044 | |
| 4045 | // Don't optimize calls that require strict floating point semantics. |
| 4046 | if (CI->isStrictFP()) |
| 4047 | return nullptr; |
| 4048 | |
| 4049 | if (Value *V = optimizeSymmetric(CI, Func, B&: Builder)) |
| 4050 | return V; |
| 4051 | |
| 4052 | switch (Func) { |
| 4053 | case LibFunc_sinpif: |
| 4054 | case LibFunc_sinpi: |
| 4055 | return optimizeSinCosPi(CI, /*IsSin*/true, B&: Builder); |
| 4056 | case LibFunc_cospif: |
| 4057 | case LibFunc_cospi: |
| 4058 | return optimizeSinCosPi(CI, /*IsSin*/false, B&: Builder); |
| 4059 | case LibFunc_powf: |
| 4060 | case LibFunc_pow: |
| 4061 | case LibFunc_powl: |
| 4062 | return optimizePow(Pow: CI, B&: Builder); |
| 4063 | case LibFunc_exp2l: |
| 4064 | case LibFunc_exp2: |
| 4065 | case LibFunc_exp2f: |
| 4066 | return optimizeExp2(CI, B&: Builder); |
| 4067 | case LibFunc_fabsf: |
| 4068 | case LibFunc_fabs: |
| 4069 | case LibFunc_fabsl: |
| 4070 | return replaceUnaryCall(CI, B&: Builder, IID: Intrinsic::fabs); |
| 4071 | case LibFunc_sqrtf: |
| 4072 | case LibFunc_sqrt: |
| 4073 | case LibFunc_sqrtl: |
| 4074 | return optimizeSqrt(CI, B&: Builder); |
| 4075 | case LibFunc_fmod: |
| 4076 | case LibFunc_fmodf: |
| 4077 | case LibFunc_fmodl: |
| 4078 | return optimizeFMod(CI, B&: Builder); |
| 4079 | case LibFunc_logf: |
| 4080 | case LibFunc_log: |
| 4081 | case LibFunc_logl: |
| 4082 | case LibFunc_log10f: |
| 4083 | case LibFunc_log10: |
| 4084 | case LibFunc_log10l: |
| 4085 | case LibFunc_log1pf: |
| 4086 | case LibFunc_log1p: |
| 4087 | case LibFunc_log1pl: |
| 4088 | case LibFunc_log2f: |
| 4089 | case LibFunc_log2: |
| 4090 | case LibFunc_log2l: |
| 4091 | case LibFunc_logbf: |
| 4092 | case LibFunc_logb: |
| 4093 | case LibFunc_logbl: |
| 4094 | return optimizeLog(Log: CI, B&: Builder); |
| 4095 | case LibFunc_tan: |
| 4096 | case LibFunc_tanf: |
| 4097 | case LibFunc_tanl: |
| 4098 | case LibFunc_sinh: |
| 4099 | case LibFunc_sinhf: |
| 4100 | case LibFunc_sinhl: |
| 4101 | case LibFunc_asinh: |
| 4102 | case LibFunc_asinhf: |
| 4103 | case LibFunc_asinhl: |
| 4104 | case LibFunc_cosh: |
| 4105 | case LibFunc_coshf: |
| 4106 | case LibFunc_coshl: |
| 4107 | case LibFunc_atanh: |
| 4108 | case LibFunc_atanhf: |
| 4109 | case LibFunc_atanhl: |
| 4110 | return optimizeTrigInversionPairs(CI, B&: Builder); |
| 4111 | case LibFunc_ceil: |
| 4112 | return replaceUnaryCall(CI, B&: Builder, IID: Intrinsic::ceil); |
| 4113 | case LibFunc_floor: |
| 4114 | return replaceUnaryCall(CI, B&: Builder, IID: Intrinsic::floor); |
| 4115 | case LibFunc_round: |
| 4116 | return replaceUnaryCall(CI, B&: Builder, IID: Intrinsic::round); |
| 4117 | case LibFunc_roundeven: |
| 4118 | return replaceUnaryCall(CI, B&: Builder, IID: Intrinsic::roundeven); |
| 4119 | case LibFunc_nearbyint: |
| 4120 | return replaceUnaryCall(CI, B&: Builder, IID: Intrinsic::nearbyint); |
| 4121 | case LibFunc_rint: |
| 4122 | return replaceUnaryCall(CI, B&: Builder, IID: Intrinsic::rint); |
| 4123 | case LibFunc_trunc: |
| 4124 | return replaceUnaryCall(CI, B&: Builder, IID: Intrinsic::trunc); |
| 4125 | case LibFunc_acos: |
| 4126 | case LibFunc_acosh: |
| 4127 | case LibFunc_asin: |
| 4128 | case LibFunc_atan: |
| 4129 | case LibFunc_cbrt: |
| 4130 | case LibFunc_exp: |
| 4131 | case LibFunc_exp10: |
| 4132 | case LibFunc_expm1: |
| 4133 | case LibFunc_cos: |
| 4134 | case LibFunc_sin: |
| 4135 | case LibFunc_tanh: |
| 4136 | if (UnsafeFPShrink && hasFloatVersion(M, FuncName: CI->getCalledFunction()->getName())) |
| 4137 | return optimizeUnaryDoubleFP(CI, B&: Builder, TLI, isPrecise: true); |
| 4138 | return nullptr; |
| 4139 | case LibFunc_copysign: |
| 4140 | if (hasFloatVersion(M, FuncName: CI->getCalledFunction()->getName())) |
| 4141 | return optimizeBinaryDoubleFP(CI, B&: Builder, TLI); |
| 4142 | return nullptr; |
| 4143 | case LibFunc_fdim: |
| 4144 | case LibFunc_fdimf: |
| 4145 | case LibFunc_fdiml: |
| 4146 | return optimizeFdim(CI, B&: Builder); |
| 4147 | case LibFunc_fminf: |
| 4148 | case LibFunc_fmin: |
| 4149 | case LibFunc_fminl: |
| 4150 | case LibFunc_fmaxf: |
| 4151 | case LibFunc_fmax: |
| 4152 | case LibFunc_fmaxl: |
| 4153 | return optimizeFMinFMax(CI, B&: Builder); |
| 4154 | case LibFunc_fminimum_numf: |
| 4155 | case LibFunc_fminimum_num: |
| 4156 | case LibFunc_fminimum_numl: |
| 4157 | case LibFunc_fmaximum_numf: |
| 4158 | case LibFunc_fmaximum_num: |
| 4159 | case LibFunc_fmaximum_numl: |
| 4160 | return optimizeFMinimumnumFMaximumnum(CI, B&: Builder); |
| 4161 | case LibFunc_cabs: |
| 4162 | case LibFunc_cabsf: |
| 4163 | case LibFunc_cabsl: |
| 4164 | return optimizeCAbs(CI, B&: Builder); |
| 4165 | case LibFunc_remquo: |
| 4166 | case LibFunc_remquof: |
| 4167 | case LibFunc_remquol: |
| 4168 | return optimizeRemquo(CI, B&: Builder); |
| 4169 | case LibFunc_nan: |
| 4170 | case LibFunc_nanf: |
| 4171 | case LibFunc_nanl: |
| 4172 | return optimizeNaN(CI); |
| 4173 | default: |
| 4174 | return nullptr; |
| 4175 | } |
| 4176 | } |
| 4177 | |
| 4178 | Value *LibCallSimplifier::optimizeCall(CallInst *CI, IRBuilderBase &Builder) { |
| 4179 | Module *M = CI->getModule(); |
| 4180 | assert(!CI->isMustTailCall() && "These transforms aren't musttail safe." ); |
| 4181 | |
| 4182 | // TODO: Split out the code below that operates on FP calls so that |
| 4183 | // we can all non-FP calls with the StrictFP attribute to be |
| 4184 | // optimized. |
| 4185 | if (CI->isNoBuiltin()) { |
| 4186 | // Optionally update operator new calls. |
| 4187 | return maybeOptimizeNoBuiltinOperatorNew(CI, B&: Builder); |
| 4188 | } |
| 4189 | |
| 4190 | LibFunc Func; |
| 4191 | Function *Callee = CI->getCalledFunction(); |
| 4192 | bool IsCallingConvC = TargetLibraryInfoImpl::isCallingConvCCompatible(CI); |
| 4193 | |
| 4194 | SmallVector<OperandBundleDef, 2> OpBundles; |
| 4195 | CI->getOperandBundlesAsDefs(Defs&: OpBundles); |
| 4196 | |
| 4197 | IRBuilderBase::OperandBundlesGuard Guard(Builder); |
| 4198 | Builder.setDefaultOperandBundles(OpBundles); |
| 4199 | |
| 4200 | // Command-line parameter overrides instruction attribute. |
| 4201 | // This can't be moved to optimizeFloatingPointLibCall() because it may be |
| 4202 | // used by the intrinsic optimizations. |
| 4203 | if (EnableUnsafeFPShrink.getNumOccurrences() > 0) |
| 4204 | UnsafeFPShrink = EnableUnsafeFPShrink; |
| 4205 | else if (isa<FPMathOperator>(Val: CI) && CI->isFast()) |
| 4206 | UnsafeFPShrink = true; |
| 4207 | |
| 4208 | // First, check for intrinsics. |
| 4209 | if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: CI)) { |
| 4210 | if (!IsCallingConvC) |
| 4211 | return nullptr; |
| 4212 | // The FP intrinsics have corresponding constrained versions so we don't |
| 4213 | // need to check for the StrictFP attribute here. |
| 4214 | switch (II->getIntrinsicID()) { |
| 4215 | case Intrinsic::pow: |
| 4216 | return optimizePow(Pow: CI, B&: Builder); |
| 4217 | case Intrinsic::exp2: |
| 4218 | return optimizeExp2(CI, B&: Builder); |
| 4219 | case Intrinsic::log: |
| 4220 | case Intrinsic::log2: |
| 4221 | case Intrinsic::log10: |
| 4222 | return optimizeLog(Log: CI, B&: Builder); |
| 4223 | case Intrinsic::sqrt: |
| 4224 | return optimizeSqrt(CI, B&: Builder); |
| 4225 | case Intrinsic::memset: |
| 4226 | return optimizeMemSet(CI, B&: Builder); |
| 4227 | case Intrinsic::memcpy: |
| 4228 | return optimizeMemCpy(CI, B&: Builder); |
| 4229 | case Intrinsic::memmove: |
| 4230 | return optimizeMemMove(CI, B&: Builder); |
| 4231 | case Intrinsic::sin: |
| 4232 | case Intrinsic::cos: |
| 4233 | if (UnsafeFPShrink) |
| 4234 | return optimizeUnaryDoubleFP(CI, B&: Builder, TLI, /*isPrecise=*/true); |
| 4235 | return nullptr; |
| 4236 | default: |
| 4237 | return nullptr; |
| 4238 | } |
| 4239 | } |
| 4240 | |
| 4241 | // Also try to simplify calls to fortified library functions. |
| 4242 | if (Value *SimplifiedFortifiedCI = |
| 4243 | FortifiedSimplifier.optimizeCall(CI, B&: Builder)) |
| 4244 | return SimplifiedFortifiedCI; |
| 4245 | |
| 4246 | // Then check for known library functions. |
| 4247 | if (TLI->getLibFunc(FDecl: *Callee, F&: Func) && isLibFuncEmittable(M, TLI, TheLibFunc: Func)) { |
| 4248 | // We never change the calling convention. |
| 4249 | if (!ignoreCallingConv(Func) && !IsCallingConvC) |
| 4250 | return nullptr; |
| 4251 | if (Value *V = optimizeStringMemoryLibCall(CI, Builder)) |
| 4252 | return V; |
| 4253 | if (Value *V = optimizeFloatingPointLibCall(CI, Func, Builder)) |
| 4254 | return V; |
| 4255 | switch (Func) { |
| 4256 | case LibFunc_ffs: |
| 4257 | case LibFunc_ffsl: |
| 4258 | case LibFunc_ffsll: |
| 4259 | return optimizeFFS(CI, B&: Builder); |
| 4260 | case LibFunc_fls: |
| 4261 | case LibFunc_flsl: |
| 4262 | case LibFunc_flsll: |
| 4263 | return optimizeFls(CI, B&: Builder); |
| 4264 | case LibFunc_abs: |
| 4265 | case LibFunc_labs: |
| 4266 | case LibFunc_llabs: |
| 4267 | return optimizeAbs(CI, B&: Builder); |
| 4268 | case LibFunc_isdigit: |
| 4269 | return optimizeIsDigit(CI, B&: Builder); |
| 4270 | case LibFunc_isascii: |
| 4271 | return optimizeIsAscii(CI, B&: Builder); |
| 4272 | case LibFunc_toascii: |
| 4273 | return optimizeToAscii(CI, B&: Builder); |
| 4274 | case LibFunc_atoi: |
| 4275 | case LibFunc_atol: |
| 4276 | case LibFunc_atoll: |
| 4277 | return optimizeAtoi(CI, B&: Builder); |
| 4278 | case LibFunc_strtol: |
| 4279 | case LibFunc_strtoll: |
| 4280 | return optimizeStrToInt(CI, B&: Builder, /*AsSigned=*/true); |
| 4281 | case LibFunc_strtoul: |
| 4282 | case LibFunc_strtoull: |
| 4283 | return optimizeStrToInt(CI, B&: Builder, /*AsSigned=*/false); |
| 4284 | case LibFunc_printf: |
| 4285 | return optimizePrintF(CI, B&: Builder); |
| 4286 | case LibFunc_sprintf: |
| 4287 | return optimizeSPrintF(CI, B&: Builder); |
| 4288 | case LibFunc_snprintf: |
| 4289 | return optimizeSnPrintF(CI, B&: Builder); |
| 4290 | case LibFunc_fprintf: |
| 4291 | return optimizeFPrintF(CI, B&: Builder); |
| 4292 | case LibFunc_fwrite: |
| 4293 | return optimizeFWrite(CI, B&: Builder); |
| 4294 | case LibFunc_fputs: |
| 4295 | return optimizeFPuts(CI, B&: Builder); |
| 4296 | case LibFunc_puts: |
| 4297 | return optimizePuts(CI, B&: Builder); |
| 4298 | case LibFunc_perror: |
| 4299 | return optimizeErrorReporting(CI, B&: Builder); |
| 4300 | case LibFunc_vfprintf: |
| 4301 | case LibFunc_fiprintf: |
| 4302 | return optimizeErrorReporting(CI, B&: Builder, StreamArg: 0); |
| 4303 | case LibFunc_exit: |
| 4304 | case LibFunc_Exit: |
| 4305 | return optimizeExit(CI); |
| 4306 | default: |
| 4307 | return nullptr; |
| 4308 | } |
| 4309 | } |
| 4310 | return nullptr; |
| 4311 | } |
| 4312 | |
| 4313 | LibCallSimplifier::LibCallSimplifier( |
| 4314 | const DataLayout &DL, const TargetLibraryInfo *TLI, DominatorTree *DT, |
| 4315 | DomConditionCache *DC, AssumptionCache *AC, OptimizationRemarkEmitter &ORE, |
| 4316 | BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, |
| 4317 | function_ref<void(Instruction *, Value *)> Replacer, |
| 4318 | function_ref<void(Instruction *)> Eraser) |
| 4319 | : FortifiedSimplifier(TLI), DL(DL), TLI(TLI), DT(DT), DC(DC), AC(AC), |
| 4320 | ORE(ORE), BFI(BFI), PSI(PSI), Replacer(Replacer), Eraser(Eraser) {} |
| 4321 | |
| 4322 | void LibCallSimplifier::replaceAllUsesWith(Instruction *I, Value *With) { |
| 4323 | // Indirect through the replacer used in this instance. |
| 4324 | Replacer(I, With); |
| 4325 | } |
| 4326 | |
| 4327 | void LibCallSimplifier::eraseFromParent(Instruction *I) { |
| 4328 | Eraser(I); |
| 4329 | } |
| 4330 | |
| 4331 | // TODO: |
| 4332 | // Additional cases that we need to add to this file: |
| 4333 | // |
| 4334 | // cbrt: |
| 4335 | // * cbrt(expN(X)) -> expN(x/3) |
| 4336 | // * cbrt(sqrt(x)) -> pow(x,1/6) |
| 4337 | // * cbrt(cbrt(x)) -> pow(x,1/9) |
| 4338 | // |
| 4339 | // exp, expf, expl: |
| 4340 | // * exp(log(x)) -> x |
| 4341 | // |
| 4342 | // log, logf, logl: |
| 4343 | // * log(exp(x)) -> x |
| 4344 | // * log(exp(y)) -> y*log(e) |
| 4345 | // * log(exp10(y)) -> y*log(10) |
| 4346 | // * log(sqrt(x)) -> 0.5*log(x) |
| 4347 | // |
| 4348 | // pow, powf, powl: |
| 4349 | // * pow(sqrt(x),y) -> pow(x,y*0.5) |
| 4350 | // * pow(pow(x,y),z)-> pow(x,y*z) |
| 4351 | // |
| 4352 | // signbit: |
| 4353 | // * signbit(cnst) -> cnst' |
| 4354 | // * signbit(nncst) -> 0 (if pstv is a non-negative constant) |
| 4355 | // |
| 4356 | // sqrt, sqrtf, sqrtl: |
| 4357 | // * sqrt(expN(x)) -> expN(x*0.5) |
| 4358 | // * sqrt(Nroot(x)) -> pow(x,1/(2*N)) |
| 4359 | // * sqrt(pow(x,y)) -> pow(|x|,y*0.5) |
| 4360 | // |
| 4361 | |
| 4362 | //===----------------------------------------------------------------------===// |
| 4363 | // Fortified Library Call Optimizations |
| 4364 | //===----------------------------------------------------------------------===// |
| 4365 | |
| 4366 | bool FortifiedLibCallSimplifier::isFortifiedCallFoldable( |
| 4367 | CallInst *CI, unsigned ObjSizeOp, std::optional<unsigned> SizeOp, |
| 4368 | std::optional<unsigned> StrOp, std::optional<unsigned> FlagOp) { |
| 4369 | // If this function takes a flag argument, the implementation may use it to |
| 4370 | // perform extra checks. Don't fold into the non-checking variant. |
| 4371 | if (FlagOp) { |
| 4372 | ConstantInt *Flag = dyn_cast<ConstantInt>(Val: CI->getArgOperand(i: *FlagOp)); |
| 4373 | if (!Flag || !Flag->isZero()) |
| 4374 | return false; |
| 4375 | } |
| 4376 | |
| 4377 | if (SizeOp && CI->getArgOperand(i: ObjSizeOp) == CI->getArgOperand(i: *SizeOp)) |
| 4378 | return true; |
| 4379 | |
| 4380 | if (ConstantInt *ObjSizeCI = |
| 4381 | dyn_cast<ConstantInt>(Val: CI->getArgOperand(i: ObjSizeOp))) { |
| 4382 | if (ObjSizeCI->isMinusOne()) |
| 4383 | return true; |
| 4384 | // If the object size wasn't -1 (unknown), bail out if we were asked to. |
| 4385 | if (OnlyLowerUnknownSize) |
| 4386 | return false; |
| 4387 | if (StrOp) { |
| 4388 | uint64_t Len = GetStringLength(V: CI->getArgOperand(i: *StrOp)); |
| 4389 | // If the length is 0 we don't know how long it is and so we can't |
| 4390 | // remove the check. |
| 4391 | if (Len) |
| 4392 | annotateDereferenceableBytes(CI, ArgNos: *StrOp, DereferenceableBytes: Len); |
| 4393 | else |
| 4394 | return false; |
| 4395 | return ObjSizeCI->getZExtValue() >= Len; |
| 4396 | } |
| 4397 | |
| 4398 | if (SizeOp) { |
| 4399 | if (ConstantInt *SizeCI = |
| 4400 | dyn_cast<ConstantInt>(Val: CI->getArgOperand(i: *SizeOp))) |
| 4401 | return ObjSizeCI->getZExtValue() >= SizeCI->getZExtValue(); |
| 4402 | } |
| 4403 | } |
| 4404 | return false; |
| 4405 | } |
| 4406 | |
| 4407 | Value *FortifiedLibCallSimplifier::optimizeMemCpyChk(CallInst *CI, |
| 4408 | IRBuilderBase &B) { |
| 4409 | if (isFortifiedCallFoldable(CI, ObjSizeOp: 3, SizeOp: 2)) { |
| 4410 | CallInst *NewCI = |
| 4411 | B.CreateMemCpy(Dst: CI->getArgOperand(i: 0), DstAlign: Align(1), Src: CI->getArgOperand(i: 1), |
| 4412 | SrcAlign: Align(1), Size: CI->getArgOperand(i: 2)); |
| 4413 | mergeAttributesAndFlags(NewCI, Old: *CI); |
| 4414 | return CI->getArgOperand(i: 0); |
| 4415 | } |
| 4416 | return nullptr; |
| 4417 | } |
| 4418 | |
| 4419 | Value *FortifiedLibCallSimplifier::optimizeMemMoveChk(CallInst *CI, |
| 4420 | IRBuilderBase &B) { |
| 4421 | if (isFortifiedCallFoldable(CI, ObjSizeOp: 3, SizeOp: 2)) { |
| 4422 | CallInst *NewCI = |
| 4423 | B.CreateMemMove(Dst: CI->getArgOperand(i: 0), DstAlign: Align(1), Src: CI->getArgOperand(i: 1), |
| 4424 | SrcAlign: Align(1), Size: CI->getArgOperand(i: 2)); |
| 4425 | mergeAttributesAndFlags(NewCI, Old: *CI); |
| 4426 | return CI->getArgOperand(i: 0); |
| 4427 | } |
| 4428 | return nullptr; |
| 4429 | } |
| 4430 | |
| 4431 | Value *FortifiedLibCallSimplifier::optimizeMemSetChk(CallInst *CI, |
| 4432 | IRBuilderBase &B) { |
| 4433 | if (isFortifiedCallFoldable(CI, ObjSizeOp: 3, SizeOp: 2)) { |
| 4434 | Value *Val = B.CreateIntCast(V: CI->getArgOperand(i: 1), DestTy: B.getInt8Ty(), isSigned: false); |
| 4435 | CallInst *NewCI = B.CreateMemSet(Ptr: CI->getArgOperand(i: 0), Val, |
| 4436 | Size: CI->getArgOperand(i: 2), Align: Align(1)); |
| 4437 | mergeAttributesAndFlags(NewCI, Old: *CI); |
| 4438 | return CI->getArgOperand(i: 0); |
| 4439 | } |
| 4440 | return nullptr; |
| 4441 | } |
| 4442 | |
| 4443 | Value *FortifiedLibCallSimplifier::optimizeMemPCpyChk(CallInst *CI, |
| 4444 | IRBuilderBase &B) { |
| 4445 | const DataLayout &DL = CI->getDataLayout(); |
| 4446 | if (isFortifiedCallFoldable(CI, ObjSizeOp: 3, SizeOp: 2)) |
| 4447 | if (Value *Call = emitMemPCpy(Dst: CI->getArgOperand(i: 0), Src: CI->getArgOperand(i: 1), |
| 4448 | Len: CI->getArgOperand(i: 2), B, DL, TLI)) { |
| 4449 | return mergeAttributesAndFlags(NewCI: cast<CallInst>(Val: Call), Old: *CI); |
| 4450 | } |
| 4451 | return nullptr; |
| 4452 | } |
| 4453 | |
| 4454 | Value *FortifiedLibCallSimplifier::optimizeStrpCpyChk(CallInst *CI, |
| 4455 | IRBuilderBase &B, |
| 4456 | LibFunc Func) { |
| 4457 | const DataLayout &DL = CI->getDataLayout(); |
| 4458 | Value *Dst = CI->getArgOperand(i: 0), *Src = CI->getArgOperand(i: 1), |
| 4459 | *ObjSize = CI->getArgOperand(i: 2); |
| 4460 | |
| 4461 | // __stpcpy_chk(x,x,...) -> x+strlen(x) |
| 4462 | if (Func == LibFunc_stpcpy_chk && !OnlyLowerUnknownSize && Dst == Src) { |
| 4463 | Value *StrLen = emitStrLen(Ptr: Src, B, DL, TLI); |
| 4464 | return StrLen ? B.CreateInBoundsGEP(Ty: B.getInt8Ty(), Ptr: Dst, IdxList: StrLen) : nullptr; |
| 4465 | } |
| 4466 | |
| 4467 | // If a) we don't have any length information, or b) we know this will |
| 4468 | // fit then just lower to a plain st[rp]cpy. Otherwise we'll keep our |
| 4469 | // st[rp]cpy_chk call which may fail at runtime if the size is too long. |
| 4470 | // TODO: It might be nice to get a maximum length out of the possible |
| 4471 | // string lengths for varying. |
| 4472 | if (isFortifiedCallFoldable(CI, ObjSizeOp: 2, SizeOp: std::nullopt, StrOp: 1)) { |
| 4473 | if (Func == LibFunc_strcpy_chk) |
| 4474 | return copyFlags(Old: *CI, New: emitStrCpy(Dst, Src, B, TLI)); |
| 4475 | else |
| 4476 | return copyFlags(Old: *CI, New: emitStpCpy(Dst, Src, B, TLI)); |
| 4477 | } |
| 4478 | |
| 4479 | if (OnlyLowerUnknownSize) |
| 4480 | return nullptr; |
| 4481 | |
| 4482 | // Maybe we can stil fold __st[rp]cpy_chk to __memcpy_chk. |
| 4483 | uint64_t Len = GetStringLength(V: Src); |
| 4484 | if (Len) |
| 4485 | annotateDereferenceableBytes(CI, ArgNos: 1, DereferenceableBytes: Len); |
| 4486 | else |
| 4487 | return nullptr; |
| 4488 | |
| 4489 | unsigned SizeTBits = TLI->getSizeTSize(M: *CI->getModule()); |
| 4490 | Type *SizeTTy = IntegerType::get(C&: CI->getContext(), NumBits: SizeTBits); |
| 4491 | Value *LenV = ConstantInt::get(Ty: SizeTTy, V: Len); |
| 4492 | Value *Ret = emitMemCpyChk(Dst, Src, Len: LenV, ObjSize, B, DL, TLI); |
| 4493 | // If the function was an __stpcpy_chk, and we were able to fold it into |
| 4494 | // a __memcpy_chk, we still need to return the correct end pointer. |
| 4495 | if (Ret && Func == LibFunc_stpcpy_chk) |
| 4496 | return B.CreateInBoundsGEP(Ty: B.getInt8Ty(), Ptr: Dst, |
| 4497 | IdxList: ConstantInt::get(Ty: SizeTTy, V: Len - 1)); |
| 4498 | return copyFlags(Old: *CI, New: cast<CallInst>(Val: Ret)); |
| 4499 | } |
| 4500 | |
| 4501 | Value *FortifiedLibCallSimplifier::optimizeStrLenChk(CallInst *CI, |
| 4502 | IRBuilderBase &B) { |
| 4503 | if (isFortifiedCallFoldable(CI, ObjSizeOp: 1, SizeOp: std::nullopt, StrOp: 0)) |
| 4504 | return copyFlags(Old: *CI, New: emitStrLen(Ptr: CI->getArgOperand(i: 0), B, |
| 4505 | DL: CI->getDataLayout(), TLI)); |
| 4506 | return nullptr; |
| 4507 | } |
| 4508 | |
| 4509 | Value *FortifiedLibCallSimplifier::optimizeStrpNCpyChk(CallInst *CI, |
| 4510 | IRBuilderBase &B, |
| 4511 | LibFunc Func) { |
| 4512 | if (isFortifiedCallFoldable(CI, ObjSizeOp: 3, SizeOp: 2)) { |
| 4513 | if (Func == LibFunc_strncpy_chk) |
| 4514 | return copyFlags(Old: *CI, |
| 4515 | New: emitStrNCpy(Dst: CI->getArgOperand(i: 0), Src: CI->getArgOperand(i: 1), |
| 4516 | Len: CI->getArgOperand(i: 2), B, TLI)); |
| 4517 | else |
| 4518 | return copyFlags(Old: *CI, |
| 4519 | New: emitStpNCpy(Dst: CI->getArgOperand(i: 0), Src: CI->getArgOperand(i: 1), |
| 4520 | Len: CI->getArgOperand(i: 2), B, TLI)); |
| 4521 | } |
| 4522 | |
| 4523 | return nullptr; |
| 4524 | } |
| 4525 | |
| 4526 | Value *FortifiedLibCallSimplifier::optimizeMemCCpyChk(CallInst *CI, |
| 4527 | IRBuilderBase &B) { |
| 4528 | if (isFortifiedCallFoldable(CI, ObjSizeOp: 4, SizeOp: 3)) |
| 4529 | return copyFlags( |
| 4530 | Old: *CI, New: emitMemCCpy(Ptr1: CI->getArgOperand(i: 0), Ptr2: CI->getArgOperand(i: 1), |
| 4531 | Val: CI->getArgOperand(i: 2), Len: CI->getArgOperand(i: 3), B, TLI)); |
| 4532 | |
| 4533 | return nullptr; |
| 4534 | } |
| 4535 | |
| 4536 | Value *FortifiedLibCallSimplifier::optimizeSNPrintfChk(CallInst *CI, |
| 4537 | IRBuilderBase &B) { |
| 4538 | if (isFortifiedCallFoldable(CI, ObjSizeOp: 3, SizeOp: 1, StrOp: std::nullopt, FlagOp: 2)) { |
| 4539 | SmallVector<Value *, 8> VariadicArgs(drop_begin(RangeOrContainer: CI->args(), N: 5)); |
| 4540 | return copyFlags(Old: *CI, |
| 4541 | New: emitSNPrintf(Dest: CI->getArgOperand(i: 0), Size: CI->getArgOperand(i: 1), |
| 4542 | Fmt: CI->getArgOperand(i: 4), Args: VariadicArgs, B, TLI)); |
| 4543 | } |
| 4544 | |
| 4545 | return nullptr; |
| 4546 | } |
| 4547 | |
| 4548 | Value *FortifiedLibCallSimplifier::optimizeSPrintfChk(CallInst *CI, |
| 4549 | IRBuilderBase &B) { |
| 4550 | if (isFortifiedCallFoldable(CI, ObjSizeOp: 2, SizeOp: std::nullopt, StrOp: std::nullopt, FlagOp: 1)) { |
| 4551 | SmallVector<Value *, 8> VariadicArgs(drop_begin(RangeOrContainer: CI->args(), N: 4)); |
| 4552 | return copyFlags(Old: *CI, |
| 4553 | New: emitSPrintf(Dest: CI->getArgOperand(i: 0), Fmt: CI->getArgOperand(i: 3), |
| 4554 | VariadicArgs, B, TLI)); |
| 4555 | } |
| 4556 | |
| 4557 | return nullptr; |
| 4558 | } |
| 4559 | |
| 4560 | Value *FortifiedLibCallSimplifier::optimizeStrCatChk(CallInst *CI, |
| 4561 | IRBuilderBase &B) { |
| 4562 | if (isFortifiedCallFoldable(CI, ObjSizeOp: 2)) |
| 4563 | return copyFlags( |
| 4564 | Old: *CI, New: emitStrCat(Dest: CI->getArgOperand(i: 0), Src: CI->getArgOperand(i: 1), B, TLI)); |
| 4565 | |
| 4566 | return nullptr; |
| 4567 | } |
| 4568 | |
| 4569 | Value *FortifiedLibCallSimplifier::optimizeStrLCat(CallInst *CI, |
| 4570 | IRBuilderBase &B) { |
| 4571 | if (isFortifiedCallFoldable(CI, ObjSizeOp: 3)) |
| 4572 | return copyFlags(Old: *CI, |
| 4573 | New: emitStrLCat(Dest: CI->getArgOperand(i: 0), Src: CI->getArgOperand(i: 1), |
| 4574 | Size: CI->getArgOperand(i: 2), B, TLI)); |
| 4575 | |
| 4576 | return nullptr; |
| 4577 | } |
| 4578 | |
| 4579 | Value *FortifiedLibCallSimplifier::optimizeStrNCatChk(CallInst *CI, |
| 4580 | IRBuilderBase &B) { |
| 4581 | if (isFortifiedCallFoldable(CI, ObjSizeOp: 3)) |
| 4582 | return copyFlags(Old: *CI, |
| 4583 | New: emitStrNCat(Dest: CI->getArgOperand(i: 0), Src: CI->getArgOperand(i: 1), |
| 4584 | Size: CI->getArgOperand(i: 2), B, TLI)); |
| 4585 | |
| 4586 | return nullptr; |
| 4587 | } |
| 4588 | |
| 4589 | Value *FortifiedLibCallSimplifier::optimizeStrLCpyChk(CallInst *CI, |
| 4590 | IRBuilderBase &B) { |
| 4591 | if (isFortifiedCallFoldable(CI, ObjSizeOp: 3)) |
| 4592 | return copyFlags(Old: *CI, |
| 4593 | New: emitStrLCpy(Dest: CI->getArgOperand(i: 0), Src: CI->getArgOperand(i: 1), |
| 4594 | Size: CI->getArgOperand(i: 2), B, TLI)); |
| 4595 | |
| 4596 | return nullptr; |
| 4597 | } |
| 4598 | |
| 4599 | Value *FortifiedLibCallSimplifier::optimizeVSNPrintfChk(CallInst *CI, |
| 4600 | IRBuilderBase &B) { |
| 4601 | if (isFortifiedCallFoldable(CI, ObjSizeOp: 3, SizeOp: 1, StrOp: std::nullopt, FlagOp: 2)) |
| 4602 | return copyFlags( |
| 4603 | Old: *CI, New: emitVSNPrintf(Dest: CI->getArgOperand(i: 0), Size: CI->getArgOperand(i: 1), |
| 4604 | Fmt: CI->getArgOperand(i: 4), VAList: CI->getArgOperand(i: 5), B, TLI)); |
| 4605 | |
| 4606 | return nullptr; |
| 4607 | } |
| 4608 | |
| 4609 | Value *FortifiedLibCallSimplifier::optimizeVSPrintfChk(CallInst *CI, |
| 4610 | IRBuilderBase &B) { |
| 4611 | if (isFortifiedCallFoldable(CI, ObjSizeOp: 2, SizeOp: std::nullopt, StrOp: std::nullopt, FlagOp: 1)) |
| 4612 | return copyFlags(Old: *CI, |
| 4613 | New: emitVSPrintf(Dest: CI->getArgOperand(i: 0), Fmt: CI->getArgOperand(i: 3), |
| 4614 | VAList: CI->getArgOperand(i: 4), B, TLI)); |
| 4615 | |
| 4616 | return nullptr; |
| 4617 | } |
| 4618 | |
| 4619 | Value *FortifiedLibCallSimplifier::optimizeCall(CallInst *CI, |
| 4620 | IRBuilderBase &Builder) { |
| 4621 | // FIXME: We shouldn't be changing "nobuiltin" or TLI unavailable calls here. |
| 4622 | // Some clang users checked for _chk libcall availability using: |
| 4623 | // __has_builtin(__builtin___memcpy_chk) |
| 4624 | // When compiling with -fno-builtin, this is always true. |
| 4625 | // When passing -ffreestanding/-mkernel, which both imply -fno-builtin, we |
| 4626 | // end up with fortified libcalls, which isn't acceptable in a freestanding |
| 4627 | // environment which only provides their non-fortified counterparts. |
| 4628 | // |
| 4629 | // Until we change clang and/or teach external users to check for availability |
| 4630 | // differently, disregard the "nobuiltin" attribute and TLI::has. |
| 4631 | // |
| 4632 | // PR23093. |
| 4633 | |
| 4634 | LibFunc Func; |
| 4635 | Function *Callee = CI->getCalledFunction(); |
| 4636 | bool IsCallingConvC = TargetLibraryInfoImpl::isCallingConvCCompatible(CI); |
| 4637 | |
| 4638 | SmallVector<OperandBundleDef, 2> OpBundles; |
| 4639 | CI->getOperandBundlesAsDefs(Defs&: OpBundles); |
| 4640 | |
| 4641 | IRBuilderBase::OperandBundlesGuard Guard(Builder); |
| 4642 | Builder.setDefaultOperandBundles(OpBundles); |
| 4643 | |
| 4644 | // First, check that this is a known library functions and that the prototype |
| 4645 | // is correct. |
| 4646 | if (!TLI->getLibFunc(FDecl: *Callee, F&: Func)) |
| 4647 | return nullptr; |
| 4648 | |
| 4649 | // We never change the calling convention. |
| 4650 | if (!ignoreCallingConv(Func) && !IsCallingConvC) |
| 4651 | return nullptr; |
| 4652 | |
| 4653 | switch (Func) { |
| 4654 | case LibFunc_memcpy_chk: |
| 4655 | return optimizeMemCpyChk(CI, B&: Builder); |
| 4656 | case LibFunc_mempcpy_chk: |
| 4657 | return optimizeMemPCpyChk(CI, B&: Builder); |
| 4658 | case LibFunc_memmove_chk: |
| 4659 | return optimizeMemMoveChk(CI, B&: Builder); |
| 4660 | case LibFunc_memset_chk: |
| 4661 | return optimizeMemSetChk(CI, B&: Builder); |
| 4662 | case LibFunc_stpcpy_chk: |
| 4663 | case LibFunc_strcpy_chk: |
| 4664 | return optimizeStrpCpyChk(CI, B&: Builder, Func); |
| 4665 | case LibFunc_strlen_chk: |
| 4666 | return optimizeStrLenChk(CI, B&: Builder); |
| 4667 | case LibFunc_stpncpy_chk: |
| 4668 | case LibFunc_strncpy_chk: |
| 4669 | return optimizeStrpNCpyChk(CI, B&: Builder, Func); |
| 4670 | case LibFunc_memccpy_chk: |
| 4671 | return optimizeMemCCpyChk(CI, B&: Builder); |
| 4672 | case LibFunc_snprintf_chk: |
| 4673 | return optimizeSNPrintfChk(CI, B&: Builder); |
| 4674 | case LibFunc_sprintf_chk: |
| 4675 | return optimizeSPrintfChk(CI, B&: Builder); |
| 4676 | case LibFunc_strcat_chk: |
| 4677 | return optimizeStrCatChk(CI, B&: Builder); |
| 4678 | case LibFunc_strlcat_chk: |
| 4679 | return optimizeStrLCat(CI, B&: Builder); |
| 4680 | case LibFunc_strncat_chk: |
| 4681 | return optimizeStrNCatChk(CI, B&: Builder); |
| 4682 | case LibFunc_strlcpy_chk: |
| 4683 | return optimizeStrLCpyChk(CI, B&: Builder); |
| 4684 | case LibFunc_vsnprintf_chk: |
| 4685 | return optimizeVSNPrintfChk(CI, B&: Builder); |
| 4686 | case LibFunc_vsprintf_chk: |
| 4687 | return optimizeVSPrintfChk(CI, B&: Builder); |
| 4688 | default: |
| 4689 | break; |
| 4690 | } |
| 4691 | return nullptr; |
| 4692 | } |
| 4693 | |
| 4694 | FortifiedLibCallSimplifier::FortifiedLibCallSimplifier( |
| 4695 | const TargetLibraryInfo *TLI, bool OnlyLowerUnknownSize) |
| 4696 | : TLI(TLI), OnlyLowerUnknownSize(OnlyLowerUnknownSize) {} |
| 4697 | |