| 1 | //===- PreISelIntrinsicLowering.cpp - Pre-ISel intrinsic lowering pass ----===// |
| 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 pass implements IR lowering for the llvm.memcpy, llvm.memmove, |
| 10 | // llvm.memset, llvm.load.relative and llvm.objc.* intrinsics. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #include "llvm/CodeGen/PreISelIntrinsicLowering.h" |
| 15 | #include "CodeGenOptions.h" |
| 16 | #include "llvm/ADT/STLExtras.h" |
| 17 | #include "llvm/Analysis/AssumptionCache.h" |
| 18 | #include "llvm/Analysis/ObjCARCInstKind.h" |
| 19 | #include "llvm/Analysis/ObjCARCUtil.h" |
| 20 | #include "llvm/Analysis/TargetLibraryInfo.h" |
| 21 | #include "llvm/Analysis/TargetTransformInfo.h" |
| 22 | #include "llvm/Analysis/ValueTracking.h" |
| 23 | #include "llvm/CodeGen/ExpandVectorPredication.h" |
| 24 | #include "llvm/CodeGen/LibcallLoweringInfo.h" |
| 25 | #include "llvm/CodeGen/Passes.h" |
| 26 | #include "llvm/CodeGen/RuntimeLibcallUtil.h" |
| 27 | #include "llvm/CodeGen/TargetLowering.h" |
| 28 | #include "llvm/CodeGen/TargetPassConfig.h" |
| 29 | #include "llvm/IR/ConstantRange.h" |
| 30 | #include "llvm/IR/Function.h" |
| 31 | #include "llvm/IR/GlobalValue.h" |
| 32 | #include "llvm/IR/IRBuilder.h" |
| 33 | #include "llvm/IR/Instructions.h" |
| 34 | #include "llvm/IR/IntrinsicInst.h" |
| 35 | #include "llvm/IR/Metadata.h" |
| 36 | #include "llvm/IR/Module.h" |
| 37 | #include "llvm/IR/ProfDataUtils.h" |
| 38 | #include "llvm/IR/RuntimeLibcalls.h" |
| 39 | #include "llvm/IR/Type.h" |
| 40 | #include "llvm/IR/Use.h" |
| 41 | #include "llvm/InitializePasses.h" |
| 42 | #include "llvm/Pass.h" |
| 43 | #include "llvm/Support/Casting.h" |
| 44 | #include "llvm/Target/TargetMachine.h" |
| 45 | #include "llvm/Transforms/Scalar/LowerConstantIntrinsics.h" |
| 46 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" |
| 47 | #include "llvm/Transforms/Utils/LowerMemIntrinsics.h" |
| 48 | #include "llvm/Transforms/Utils/LowerVectorIntrinsics.h" |
| 49 | |
| 50 | using namespace llvm; |
| 51 | |
| 52 | #define DEBUG_TYPE "pre-isel-intrinsic-lowering" |
| 53 | |
| 54 | namespace { |
| 55 | |
| 56 | struct PreISelIntrinsicLowering { |
| 57 | const CodeGenOptions &Opts; |
| 58 | const TargetMachine &TM; |
| 59 | const ModuleLibcallLoweringInfo &ModuleLibcalls; |
| 60 | const function_ref<TargetTransformInfo &(Function &)> LookupTTI; |
| 61 | const function_ref<TargetLibraryInfo &(Function &)> LookupTLI; |
| 62 | const function_ref<AssumptionCache &(Function &)> LookupAC; |
| 63 | |
| 64 | /// If this is true, assume it's preferably to leave memory intrinsic calls |
| 65 | /// for replacement with a library call later. Otherwise this depends on |
| 66 | /// TargetLoweringInfo availability of the corresponding function. |
| 67 | const bool UseMemIntrinsicLibFunc; |
| 68 | |
| 69 | explicit PreISelIntrinsicLowering( |
| 70 | const CodeGenOptions &Opts_, const TargetMachine &TM_, |
| 71 | const ModuleLibcallLoweringInfo &ModuleLibcalls_, |
| 72 | function_ref<TargetTransformInfo &(Function &)> LookupTTI_, |
| 73 | function_ref<TargetLibraryInfo &(Function &)> LookupTLI_, |
| 74 | function_ref<AssumptionCache &(Function &)> LookupAC_, |
| 75 | bool UseMemIntrinsicLibFunc_ = true) |
| 76 | : Opts(Opts_), TM(TM_), ModuleLibcalls(ModuleLibcalls_), |
| 77 | LookupTTI(LookupTTI_), LookupTLI(LookupTLI_), LookupAC(LookupAC_), |
| 78 | UseMemIntrinsicLibFunc(UseMemIntrinsicLibFunc_) {} |
| 79 | |
| 80 | bool shouldExpandMemIntrinsicWithSize(Value *Size, |
| 81 | const TargetTransformInfo &TTI) const; |
| 82 | bool tryExpandBoundedMemSet(MemSetInst *Memset, |
| 83 | const TargetTransformInfo &TTI) const; |
| 84 | bool |
| 85 | expandMemIntrinsicUses(Function &F, |
| 86 | DenseMap<Constant *, GlobalVariable *> &CMap) const; |
| 87 | bool lowerIntrinsics(Module &M) const; |
| 88 | }; |
| 89 | |
| 90 | } // namespace |
| 91 | |
| 92 | template <class T> static bool forEachCall(Function &Intrin, T Callback) { |
| 93 | // Lowering all intrinsics in a function will delete multiple uses, so we |
| 94 | // can't use an early-inc-range. In case some remain, we don't want to look |
| 95 | // at them again. Unfortunately, Value::UseList is private, so we can't use a |
| 96 | // simple Use**. If LastUse is null, the next use to consider is |
| 97 | // Intrin.use_begin(), otherwise it's LastUse->getNext(). |
| 98 | Use *LastUse = nullptr; |
| 99 | bool Changed = false; |
| 100 | while (!Intrin.use_empty() && (!LastUse || LastUse->getNext())) { |
| 101 | Use *U = LastUse ? LastUse->getNext() : &*Intrin.use_begin(); |
| 102 | bool Removed = false; |
| 103 | // An intrinsic cannot have its address taken, so it cannot be an argument |
| 104 | // operand. It might be used as operand in debug metadata, though. |
| 105 | if (auto CI = dyn_cast<CallInst>(Val: U->getUser())) |
| 106 | Changed |= Removed = Callback(CI); |
| 107 | if (!Removed) |
| 108 | LastUse = U; |
| 109 | } |
| 110 | return Changed; |
| 111 | } |
| 112 | |
| 113 | static bool lowerLoadRelative(Function &F) { |
| 114 | if (F.use_empty()) |
| 115 | return false; |
| 116 | |
| 117 | bool Changed = false; |
| 118 | Type *Int32Ty = Type::getInt32Ty(C&: F.getContext()); |
| 119 | |
| 120 | for (Use &U : llvm::make_early_inc_range(Range: F.uses())) { |
| 121 | auto CI = dyn_cast<CallInst>(Val: U.getUser()); |
| 122 | if (!CI || CI->getCalledOperand() != &F) |
| 123 | continue; |
| 124 | |
| 125 | IRBuilder<> B(CI); |
| 126 | Value *Ptr = CI->getArgOperand(i: 0); |
| 127 | Value *OffsetPtr = B.CreatePtrAddWithSExtOrTrunc(Ptr, Offset: CI->getArgOperand(i: 1)); |
| 128 | Value *OffsetI32 = B.CreateAlignedLoad(Ty: Int32Ty, Ptr: OffsetPtr, Align: Align(4)); |
| 129 | |
| 130 | Value *ResultPtr = B.CreatePtrAddWithSExtOrTrunc(Ptr, Offset: OffsetI32); |
| 131 | |
| 132 | CI->replaceAllUsesWith(V: ResultPtr); |
| 133 | CI->eraseFromParent(); |
| 134 | Changed = true; |
| 135 | } |
| 136 | |
| 137 | return Changed; |
| 138 | } |
| 139 | |
| 140 | /// Lower @llvm.can.load.speculatively using target-specific expansion. |
| 141 | /// Targets may provide their own expansion via |
| 142 | /// TargetLowering::emitCanLoadSpeculatively; the default expansion |
| 143 | /// conservatively returns false. |
| 144 | static bool lowerCanLoadSpeculatively(Function &F, const TargetMachine &TM) { |
| 145 | return forEachCall(Intrin&: F, Callback: [&](CallInst *CI) { |
| 146 | const TargetLowering *TLI = |
| 147 | TM.getSubtargetImpl(*CI->getFunction())->getTargetLowering(); |
| 148 | |
| 149 | IRBuilder<> Builder(CI); |
| 150 | // A null result means the target cannot answer; lower to false. |
| 151 | Value *Result = TLI->emitCanLoadSpeculatively(Builder, Ptr: CI->getArgOperand(i: 0), |
| 152 | Size: CI->getArgOperand(i: 1)); |
| 153 | if (!Result) |
| 154 | Result = Builder.getFalse(); |
| 155 | |
| 156 | CI->replaceAllUsesWith(V: Result); |
| 157 | CI->eraseFromParent(); |
| 158 | return true; |
| 159 | }); |
| 160 | } |
| 161 | |
| 162 | // ObjCARC has knowledge about whether an obj-c runtime function needs to be |
| 163 | // always tail-called or never tail-called. |
| 164 | static CallInst::TailCallKind getOverridingTailCallKind(const Function &F) { |
| 165 | objcarc::ARCInstKind Kind = objcarc::GetFunctionClass(F: &F); |
| 166 | if (objcarc::IsAlwaysTail(Class: Kind)) |
| 167 | return CallInst::TCK_Tail; |
| 168 | else if (objcarc::IsNeverTail(Class: Kind)) |
| 169 | return CallInst::TCK_NoTail; |
| 170 | return CallInst::TCK_None; |
| 171 | } |
| 172 | |
| 173 | static bool lowerObjCCall(Function &F, RTLIB::LibcallImpl NewFn, |
| 174 | bool setNonLazyBind = false) { |
| 175 | assert(IntrinsicInst::mayLowerToFunctionCall(F.getIntrinsicID()) && |
| 176 | "Pre-ISel intrinsics do lower into regular function calls" ); |
| 177 | if (F.use_empty()) |
| 178 | return false; |
| 179 | |
| 180 | // FIXME: When RuntimeLibcalls is an analysis, check if the function is really |
| 181 | // supported, and go through RTLIB::Libcall. |
| 182 | StringRef NewFnName = RTLIB::RuntimeLibcallsInfo::getLibcallImplName(CallImpl: NewFn); |
| 183 | |
| 184 | // If we haven't already looked up this function, check to see if the |
| 185 | // program already contains a function with this name. |
| 186 | Module *M = F.getParent(); |
| 187 | FunctionCallee FCache = |
| 188 | M->getOrInsertFunction(Name: NewFnName, T: F.getFunctionType()); |
| 189 | |
| 190 | if (Function *Fn = dyn_cast<Function>(Val: FCache.getCallee())) { |
| 191 | Fn->setLinkage(F.getLinkage()); |
| 192 | if (setNonLazyBind && !Fn->isWeakForLinker()) { |
| 193 | // If we have Native ARC, set nonlazybind attribute for these APIs for |
| 194 | // performance. |
| 195 | Fn->addFnAttr(Kind: Attribute::NonLazyBind); |
| 196 | } |
| 197 | } |
| 198 | |
| 199 | CallInst::TailCallKind OverridingTCK = getOverridingTailCallKind(F); |
| 200 | |
| 201 | for (Use &U : llvm::make_early_inc_range(Range: F.uses())) { |
| 202 | auto *CB = cast<CallBase>(Val: U.getUser()); |
| 203 | |
| 204 | if (CB->getCalledFunction() != &F) { |
| 205 | assert(objcarc::getAttachedARCFunction(CB) == &F && |
| 206 | "use expected to be the argument of operand bundle " |
| 207 | "\"clang.arc.attachedcall\"" ); |
| 208 | U.set(FCache.getCallee()); |
| 209 | continue; |
| 210 | } |
| 211 | |
| 212 | auto *CI = cast<CallInst>(Val: CB); |
| 213 | assert(CI->getCalledFunction() && "Cannot lower an indirect call!" ); |
| 214 | |
| 215 | IRBuilder<> Builder(CI->getIterator()); |
| 216 | SmallVector<Value *, 8> Args(CI->args()); |
| 217 | SmallVector<llvm::OperandBundleDef, 1> BundleList; |
| 218 | CI->getOperandBundlesAsDefs(Defs&: BundleList); |
| 219 | CallInst *NewCI = Builder.CreateCall(Callee: FCache, Args, OpBundles: BundleList); |
| 220 | NewCI->setName(CI->getName()); |
| 221 | |
| 222 | // Try to set the most appropriate TailCallKind based on both the current |
| 223 | // attributes and the ones that we could get from ObjCARC's special |
| 224 | // knowledge of the runtime functions. |
| 225 | // |
| 226 | // std::max respects both requirements of notail and tail here: |
| 227 | // * notail on either the call or from ObjCARC becomes notail |
| 228 | // * tail on either side is stronger than none, but not notail |
| 229 | CallInst::TailCallKind TCK = CI->getTailCallKind(); |
| 230 | NewCI->setTailCallKind(std::max(a: TCK, b: OverridingTCK)); |
| 231 | |
| 232 | // Transfer the 'returned' attribute from the intrinsic to the call site. |
| 233 | // By applying this only to intrinsic call sites, we avoid applying it to |
| 234 | // non-ARC explicit calls to things like objc_retain which have not been |
| 235 | // auto-upgraded to use the intrinsics. |
| 236 | unsigned Index; |
| 237 | if (F.getAttributes().hasAttrSomewhere(Kind: Attribute::Returned, Index: &Index) && |
| 238 | Index) |
| 239 | NewCI->addParamAttr(ArgNo: Index - AttributeList::FirstArgIndex, |
| 240 | Kind: Attribute::Returned); |
| 241 | |
| 242 | if (!CI->use_empty()) |
| 243 | CI->replaceAllUsesWith(V: NewCI); |
| 244 | CI->eraseFromParent(); |
| 245 | } |
| 246 | |
| 247 | return true; |
| 248 | } |
| 249 | |
| 250 | // TODO: Should refine based on estimated number of accesses (e.g. does it |
| 251 | // require splitting based on alignment) |
| 252 | bool PreISelIntrinsicLowering::shouldExpandMemIntrinsicWithSize( |
| 253 | Value *Size, const TargetTransformInfo &TTI) const { |
| 254 | ConstantInt *CI = dyn_cast<ConstantInt>(Val: Size); |
| 255 | if (!CI) |
| 256 | return true; |
| 257 | uint64_t Threshold = Opts.mem_intrinsic_expand_size.value_or( |
| 258 | u: TTI.getMaxMemIntrinsicInlineSizeThreshold()); |
| 259 | uint64_t SizeVal = CI->getZExtValue(); |
| 260 | |
| 261 | // Treat a threshold of 0 as a special case to force expansion of all |
| 262 | // intrinsics, including size 0. |
| 263 | return SizeVal > Threshold || Threshold == 0; |
| 264 | } |
| 265 | |
| 266 | static bool canEmitLibcall(const ModuleLibcallLoweringInfo &ModuleLowering, |
| 267 | const TargetMachine &TM, Function *F, |
| 268 | RTLIB::Libcall LC) { |
| 269 | // TODO: Should this consider the address space of the memcpy? |
| 270 | const LibcallLoweringInfo &Lowering = |
| 271 | getLibcallLowering(ModuleInfo: ModuleLowering, Subtarget: *TM.getSubtargetImpl(*F)); |
| 272 | return Lowering.getLibcallImpl(Call: LC) != RTLIB::Unsupported; |
| 273 | } |
| 274 | |
| 275 | static bool canEmitMemcpy(const ModuleLibcallLoweringInfo &ModuleLowering, |
| 276 | const TargetMachine &TM, Function *F) { |
| 277 | // TODO: Should this consider the address space of the memcpy? |
| 278 | const LibcallLoweringInfo &Lowering = |
| 279 | getLibcallLowering(ModuleInfo: ModuleLowering, Subtarget: *TM.getSubtargetImpl(*F)); |
| 280 | return Lowering.getMemcpyImpl() != RTLIB::Unsupported; |
| 281 | } |
| 282 | |
| 283 | // Expand a memset whose length is not constant but is known to be small, which |
| 284 | // is faster than calling the library. |
| 285 | bool PreISelIntrinsicLowering::tryExpandBoundedMemSet( |
| 286 | MemSetInst *Memset, const TargetTransformInfo &TTI) const { |
| 287 | // This pass runs even at -O0, so skip if e.g. llc -O0 is used. |
| 288 | if (TM.getOptLevel() == CodeGenOptLevel::None) |
| 289 | return false; |
| 290 | |
| 291 | Value *Len = Memset->getLength(); |
| 292 | Function *F = Memset->getFunction(); |
| 293 | if (isa<Constant>(Val: Len) || Memset->isVolatile() || F->hasOptNone() || |
| 294 | F->hasMinSize()) |
| 295 | return false; |
| 296 | uint64_t Threshold = Opts.bounded_memset_inline_size.value_or( |
| 297 | u: TTI.getMaxBoundedMemSetInlineSize()); |
| 298 | if (!Threshold) |
| 299 | return false; |
| 300 | |
| 301 | // Bounds from (1) assumptions, e.g. __builtin_assume(len <= 32) |
| 302 | // and (2) known bits and ranges on instructions themselves, and... |
| 303 | SimplifyQuery SQ(Memset->getDataLayout(), Memset); |
| 304 | SQ.AC = &LookupAC(*F); |
| 305 | KnownBits Known = computeKnownBits(V: Len, Q: SQ); |
| 306 | ConstantRange CR = computeConstantRangeIncludingKnownBits( |
| 307 | V: {Len, Known}, /*ForSigned=*/false, SQ); |
| 308 | |
| 309 | // ... (3) range attr on the memset. |
| 310 | Attribute RangeAttr = Memset->getParamAttr( |
| 311 | ArgNo: Memset->getLengthUse().getOperandNo(), Kind: Attribute::Range); |
| 312 | if (RangeAttr.isValid()) |
| 313 | CR = CR.intersectWith(CR: RangeAttr.getRange()); |
| 314 | |
| 315 | if (CR.isEmptySet() || CR.getUnsignedMax().ugt(RHS: Threshold)) |
| 316 | return false; |
| 317 | |
| 318 | expandBoundedMemSet(MemSet: Memset, MinLen: CR.getUnsignedMin().getZExtValue(), |
| 319 | MaxLen: CR.getUnsignedMax().getZExtValue(), LenKnown: Known); |
| 320 | return true; |
| 321 | } |
| 322 | |
| 323 | // Return a value appropriate for use with the memset_pattern16 libcall, if |
| 324 | // possible and if we know how. (Adapted from equivalent helper in |
| 325 | // LoopIdiomRecognize). |
| 326 | static Constant * |
| 327 | getMemSetPattern16Value(MemSetPatternInst *Inst, |
| 328 | const RTLIB::RuntimeLibcallsInfo &RTLCI) { |
| 329 | // TODO: This could check for UndefValue because it can be merged into any |
| 330 | // other valid pattern. |
| 331 | |
| 332 | // Don't emit libcalls if a non-default address space is being used. |
| 333 | if (Inst->getRawDest()->getType()->getPointerAddressSpace() != 0) |
| 334 | return nullptr; |
| 335 | |
| 336 | Value *V = Inst->getValue(); |
| 337 | Type *VTy = V->getType(); |
| 338 | const DataLayout &DL = Inst->getDataLayout(); |
| 339 | |
| 340 | if (!RTLCI.isAvailable(Impl: RTLIB::impl_memset_pattern16)) |
| 341 | return nullptr; |
| 342 | |
| 343 | // If the value isn't a constant, we can't promote it to being in a constant |
| 344 | // array. We could theoretically do a store to an alloca or something, but |
| 345 | // that doesn't seem worthwhile. |
| 346 | Constant *C = dyn_cast<Constant>(Val: V); |
| 347 | if (!C || isa<ConstantExpr>(Val: C)) |
| 348 | return nullptr; |
| 349 | |
| 350 | // Only handle simple values that are a power of two bytes in size. |
| 351 | uint64_t Size = DL.getTypeSizeInBits(Ty: VTy); |
| 352 | if (!DL.typeSizeEqualsStoreSize(Ty: VTy) || !isPowerOf2_64(Value: Size)) |
| 353 | return nullptr; |
| 354 | |
| 355 | // Don't care enough about darwin/ppc to implement this. |
| 356 | if (DL.isBigEndian()) |
| 357 | return nullptr; |
| 358 | |
| 359 | // Convert to size in bytes. |
| 360 | Size /= 8; |
| 361 | |
| 362 | // TODO: If CI is larger than 16-bytes, we can try slicing it in half to see |
| 363 | // if the top and bottom are the same (e.g. for vectors and large integers). |
| 364 | if (Size > 16) |
| 365 | return nullptr; |
| 366 | |
| 367 | // If the constant is exactly 16 bytes, just use it. |
| 368 | if (Size == 16) |
| 369 | return C; |
| 370 | |
| 371 | // Otherwise, we'll use an array of the constants. |
| 372 | uint64_t ArraySize = 16 / Size; |
| 373 | ArrayType *AT = ArrayType::get(ElementType: V->getType(), NumElements: ArraySize); |
| 374 | return ConstantArray::get(T: AT, V: std::vector<Constant *>(ArraySize, C)); |
| 375 | } |
| 376 | |
| 377 | // TODO: Handle atomic memcpy and memcpy.inline |
| 378 | // TODO: Pass ScalarEvolution |
| 379 | bool PreISelIntrinsicLowering::expandMemIntrinsicUses( |
| 380 | Function &F, DenseMap<Constant *, GlobalVariable *> &CMap) const { |
| 381 | Intrinsic::ID ID = F.getIntrinsicID(); |
| 382 | bool Changed = false; |
| 383 | |
| 384 | for (User *U : llvm::make_early_inc_range(Range: F.users())) { |
| 385 | Instruction *Inst = cast<Instruction>(Val: U); |
| 386 | |
| 387 | switch (ID) { |
| 388 | case Intrinsic::memcpy: { |
| 389 | auto *Memcpy = cast<MemCpyInst>(Val: Inst); |
| 390 | Function *ParentFunc = Memcpy->getFunction(); |
| 391 | const TargetTransformInfo &TTI = LookupTTI(*ParentFunc); |
| 392 | if (shouldExpandMemIntrinsicWithSize(Size: Memcpy->getLength(), TTI)) { |
| 393 | if (UseMemIntrinsicLibFunc && |
| 394 | canEmitMemcpy(ModuleLowering: ModuleLibcalls, TM, F: ParentFunc)) |
| 395 | break; |
| 396 | |
| 397 | // TODO: For optsize, emit the loop into a separate function |
| 398 | expandMemCpyAsLoop(MemCpy: Memcpy, TTI); |
| 399 | Changed = true; |
| 400 | Memcpy->eraseFromParent(); |
| 401 | } |
| 402 | |
| 403 | break; |
| 404 | } |
| 405 | case Intrinsic::memcpy_inline: { |
| 406 | // Only expand llvm.memcpy.inline with non-constant length in this |
| 407 | // codepath, leaving the current SelectionDAG expansion for constant |
| 408 | // length memcpy intrinsics undisturbed. |
| 409 | auto *Memcpy = cast<MemCpyInst>(Val: Inst); |
| 410 | if (isa<ConstantInt>(Val: Memcpy->getLength())) |
| 411 | break; |
| 412 | |
| 413 | Function *ParentFunc = Memcpy->getFunction(); |
| 414 | const TargetTransformInfo &TTI = LookupTTI(*ParentFunc); |
| 415 | expandMemCpyAsLoop(MemCpy: Memcpy, TTI); |
| 416 | Changed = true; |
| 417 | Memcpy->eraseFromParent(); |
| 418 | break; |
| 419 | } |
| 420 | case Intrinsic::memmove: { |
| 421 | auto *Memmove = cast<MemMoveInst>(Val: Inst); |
| 422 | Function *ParentFunc = Memmove->getFunction(); |
| 423 | const TargetTransformInfo &TTI = LookupTTI(*ParentFunc); |
| 424 | if (shouldExpandMemIntrinsicWithSize(Size: Memmove->getLength(), TTI)) { |
| 425 | if (UseMemIntrinsicLibFunc && |
| 426 | canEmitLibcall(ModuleLowering: ModuleLibcalls, TM, F: ParentFunc, LC: RTLIB::MEMMOVE)) |
| 427 | break; |
| 428 | |
| 429 | if (expandMemMoveAsLoop(MemMove: Memmove, TTI)) { |
| 430 | Changed = true; |
| 431 | Memmove->eraseFromParent(); |
| 432 | } |
| 433 | } |
| 434 | |
| 435 | break; |
| 436 | } |
| 437 | case Intrinsic::memset: { |
| 438 | auto *Memset = cast<MemSetInst>(Val: Inst); |
| 439 | Function *ParentFunc = Memset->getFunction(); |
| 440 | const TargetTransformInfo &TTI = LookupTTI(*ParentFunc); |
| 441 | if (tryExpandBoundedMemSet(Memset, TTI)) { |
| 442 | Changed = true; |
| 443 | Memset->eraseFromParent(); |
| 444 | break; |
| 445 | } |
| 446 | if (shouldExpandMemIntrinsicWithSize(Size: Memset->getLength(), TTI)) { |
| 447 | if (UseMemIntrinsicLibFunc && |
| 448 | canEmitLibcall(ModuleLowering: ModuleLibcalls, TM, F: ParentFunc, LC: RTLIB::MEMSET)) |
| 449 | break; |
| 450 | |
| 451 | expandMemSetAsLoop(MemSet: Memset, TTI); |
| 452 | Changed = true; |
| 453 | Memset->eraseFromParent(); |
| 454 | } |
| 455 | |
| 456 | break; |
| 457 | } |
| 458 | case Intrinsic::memset_inline: { |
| 459 | // Only expand llvm.memset.inline with non-constant length in this |
| 460 | // codepath, leaving the current SelectionDAG expansion for constant |
| 461 | // length memset intrinsics undisturbed. |
| 462 | auto *Memset = cast<MemSetInst>(Val: Inst); |
| 463 | if (isa<ConstantInt>(Val: Memset->getLength())) |
| 464 | break; |
| 465 | |
| 466 | Function *ParentFunc = Memset->getFunction(); |
| 467 | const TargetTransformInfo &TTI = LookupTTI(*ParentFunc); |
| 468 | expandMemSetAsLoop(MemSet: Memset, TTI); |
| 469 | Changed = true; |
| 470 | Memset->eraseFromParent(); |
| 471 | break; |
| 472 | } |
| 473 | case Intrinsic::experimental_memset_pattern: { |
| 474 | auto *Memset = cast<MemSetPatternInst>(Val: Inst); |
| 475 | Function *ParentFunc = Memset->getFunction(); |
| 476 | const RTLIB::RuntimeLibcallsInfo &RTLCI = |
| 477 | ModuleLibcalls.getRuntimeLibcallsInfo(); |
| 478 | Constant *PatternValue = getMemSetPattern16Value(Inst: Memset, RTLCI); |
| 479 | if (!PatternValue) { |
| 480 | // If it isn't possible to emit a memset_pattern16 libcall, expand to |
| 481 | // a loop instead. |
| 482 | const TargetTransformInfo &TTI = LookupTTI(*ParentFunc); |
| 483 | expandMemSetPatternAsLoop(MemSet: Memset, TTI); |
| 484 | Changed = true; |
| 485 | Memset->eraseFromParent(); |
| 486 | break; |
| 487 | } |
| 488 | // FIXME: There is currently no profitability calculation for emitting |
| 489 | // the libcall vs expanding the memset.pattern directly. |
| 490 | IRBuilder<> Builder(Inst); |
| 491 | Module *M = Memset->getModule(); |
| 492 | const DataLayout &DL = Memset->getDataLayout(); |
| 493 | |
| 494 | FunctionCallee MSP = |
| 495 | RTLCI.getOrInsertLibcallDecl(M&: *M, LibcallImpl: RTLIB::impl_memset_pattern16); |
| 496 | Type *SizeTTy = MSP.getFunctionType()->getParamType(i: 2); |
| 497 | |
| 498 | // Otherwise we should form a memset_pattern16. PatternValue is known |
| 499 | // to be an constant array of 16-bytes. Put the value into a mergable |
| 500 | // global. |
| 501 | assert(Memset->getRawDest()->getType()->getPointerAddressSpace() == 0 && |
| 502 | "Should have skipped if non-zero AS" ); |
| 503 | GlobalVariable *GV; |
| 504 | auto It = CMap.find(Val: PatternValue); |
| 505 | if (It != CMap.end()) { |
| 506 | GV = It->second; |
| 507 | } else { |
| 508 | GV = new GlobalVariable( |
| 509 | *M, PatternValue->getType(), /*isConstant=*/true, |
| 510 | GlobalValue::PrivateLinkage, PatternValue, ".memset_pattern" ); |
| 511 | GV->setUnnamedAddr( |
| 512 | GlobalValue::UnnamedAddr::Global); // Ok to merge these. |
| 513 | // TODO: Consider relaxing alignment requirement. |
| 514 | GV->setAlignment(Align(16)); |
| 515 | CMap[PatternValue] = GV; |
| 516 | } |
| 517 | Value *PatternPtr = GV; |
| 518 | Value *NumBytes = Builder.CreateMul( |
| 519 | LHS: ConstantInt::get(Ty: SizeTTy, |
| 520 | V: DL.getTypeAllocSize(Ty: Memset->getValue()->getType())), |
| 521 | RHS: Builder.CreateZExtOrTrunc(V: Memset->getLength(), DestTy: SizeTTy)); |
| 522 | CallInst *MemsetPattern16Call = |
| 523 | Builder.CreateCall(Callee: MSP, Args: {Memset->getRawDest(), PatternPtr, NumBytes}); |
| 524 | MemsetPattern16Call->setCallingConv( |
| 525 | RTLCI.getLibcallImplCallingConv(Call: RTLIB::impl_memset_pattern16)); |
| 526 | MemsetPattern16Call->setAAMetadata(Memset->getAAMetadata()); |
| 527 | // Preserve any call site attributes on the destination pointer |
| 528 | // argument (e.g. alignment). |
| 529 | AttrBuilder ArgAttrs(Memset->getContext(), |
| 530 | Memset->getAttributes().getParamAttrs(ArgNo: 0)); |
| 531 | MemsetPattern16Call->setAttributes( |
| 532 | MemsetPattern16Call->getAttributes().addParamAttributes( |
| 533 | C&: Memset->getContext(), ArgNo: 0, B: ArgAttrs)); |
| 534 | Changed = true; |
| 535 | Memset->eraseFromParent(); |
| 536 | break; |
| 537 | } |
| 538 | default: |
| 539 | llvm_unreachable("unhandled intrinsic" ); |
| 540 | } |
| 541 | } |
| 542 | |
| 543 | return Changed; |
| 544 | } |
| 545 | |
| 546 | static GlobalValue *getDeactivationSymbol(CallInst *Call) { |
| 547 | if (auto Bundle = Call->getOperandBundle(ID: LLVMContext::OB_deactivation_symbol)) |
| 548 | return cast<GlobalValue>(Val: Bundle->Inputs[0]); |
| 549 | return nullptr; |
| 550 | } |
| 551 | |
| 552 | static bool expandPtrauthForEmuPAC(Function &Intr) { |
| 553 | Module &M = *Intr.getParent(); |
| 554 | if (Triple(M.getTargetTriple()).isArm64e()) |
| 555 | return false; |
| 556 | |
| 557 | Type *Int64Ty = Type::getInt64Ty(C&: M.getContext()); |
| 558 | |
| 559 | assert(Intr.getIntrinsicID() == Intrinsic::ptrauth_sign || |
| 560 | Intr.getIntrinsicID() == Intrinsic::ptrauth_auth); |
| 561 | auto *EmuFnTy = FunctionType::get(Result: Int64Ty, Params: {Int64Ty, Int64Ty}, isVarArg: false); |
| 562 | FunctionCallee EmuIntr = M.getOrInsertFunction( |
| 563 | Name: Intr.getIntrinsicID() == Intrinsic::ptrauth_auth ? "__emupac_autda" |
| 564 | : "__emupac_pacda" , |
| 565 | T: EmuFnTy); |
| 566 | |
| 567 | for (User *U : llvm::make_early_inc_range(Range: Intr.users())) { |
| 568 | auto *Call = cast<CallInst>(Val: U); |
| 569 | // We only support the DA key for now. |
| 570 | if (auto *Key = dyn_cast<ConstantInt>(Val: Call->getArgOperand(i: 1)); |
| 571 | !Key || Key->getZExtValue() != /*AArch64PACKey::DA*/ 2) |
| 572 | continue; |
| 573 | |
| 574 | Function *F = Call->getParent()->getParent(); |
| 575 | Attribute FSAttr = F->getFnAttribute(Kind: "target-features" ); |
| 576 | if (FSAttr.isValid() && FSAttr.getValueAsString().contains(Other: "+pauth" )) |
| 577 | continue; |
| 578 | |
| 579 | std::vector<OperandBundleDef> DSBundle; |
| 580 | if (auto *DS = getDeactivationSymbol(Call)) |
| 581 | DSBundle.push_back(x: OperandBundleDef("deactivation-symbol" , DS)); |
| 582 | |
| 583 | IRBuilder<> B(Call); |
| 584 | auto *EmuCall = B.CreateCall( |
| 585 | Callee: EmuIntr, Args: {Call->getArgOperand(i: 0), Call->getArgOperand(i: 2)}, OpBundles: DSBundle); |
| 586 | Call->replaceAllUsesWith(V: EmuCall); |
| 587 | Call->eraseFromParent(); |
| 588 | } |
| 589 | return true; |
| 590 | } |
| 591 | |
| 592 | static bool expandProtectedFieldPtr(Function &Intr) { |
| 593 | Module &M = *Intr.getParent(); |
| 594 | |
| 595 | SmallPtrSet<GlobalValue *, 2> DSsToDeactivate; |
| 596 | |
| 597 | Type *Int8Ty = Type::getInt8Ty(C&: M.getContext()); |
| 598 | Type *Int64Ty = Type::getInt64Ty(C&: M.getContext()); |
| 599 | PointerType *PtrTy = PointerType::get(C&: M.getContext(), AddressSpace: 0); |
| 600 | |
| 601 | for (User *U : llvm::make_early_inc_range(Range: Intr.users())) { |
| 602 | auto *Call = cast<CallInst>(Val: U); |
| 603 | |
| 604 | auto *Pointer = Call->getArgOperand(i: 0); |
| 605 | bool UseHWEncoding = |
| 606 | cast<ConstantInt>(Val: Call->getArgOperand(i: 2))->getZExtValue(); |
| 607 | if (!UseHWEncoding) |
| 608 | reportFatalUsageError(reason: "software encoding currently unsupported" ); |
| 609 | |
| 610 | auto *DS = getDeactivationSymbol(Call); |
| 611 | |
| 612 | for (Use &U : llvm::make_early_inc_range(Range: Call->uses())) { |
| 613 | // Comparisons against null cannot be used to recover the original |
| 614 | // pointer so we replace them with comparisons against the original |
| 615 | // pointer. |
| 616 | if (auto *CI = dyn_cast<ICmpInst>(Val: U.getUser())) { |
| 617 | if (auto *Op = dyn_cast<Constant>(Val: CI->getOperand(i_nocapture: 0))) { |
| 618 | if (Op->isNullValue()) { |
| 619 | CI->setOperand(i_nocapture: 1, Val_nocapture: Pointer); |
| 620 | continue; |
| 621 | } |
| 622 | } |
| 623 | if (auto *Op = dyn_cast<Constant>(Val: CI->getOperand(i_nocapture: 1))) { |
| 624 | if (Op->isNullValue()) { |
| 625 | CI->setOperand(i_nocapture: 0, Val_nocapture: Pointer); |
| 626 | continue; |
| 627 | } |
| 628 | } |
| 629 | } |
| 630 | |
| 631 | // If we are here, this means that we couldn't rewrite away this use of |
| 632 | // the intrinsic. Any load or store uses were removed by InstCombine, and |
| 633 | // in general, we can't rewrite away non-load/store uses of |
| 634 | // llvm.protected.field.ptr because doing so could expose the encoded |
| 635 | // pointer value to the program. Replace it with the pointer operand, and |
| 636 | // arrange to define a deactivation symbol. |
| 637 | U.set(Pointer); |
| 638 | if (DS) |
| 639 | DSsToDeactivate.insert(Ptr: DS); |
| 640 | } |
| 641 | |
| 642 | Call->eraseFromParent(); |
| 643 | } |
| 644 | |
| 645 | if (!DSsToDeactivate.empty()) { |
| 646 | // This is an AArch64 NOP instruction. When the deactivation symbol support |
| 647 | // is expanded to more architectures, there will likely need to be an API |
| 648 | // for retrieving this constant. |
| 649 | Constant *Nop = |
| 650 | ConstantExpr::getIntToPtr(C: ConstantInt::get(Ty: Int64Ty, V: 0xd503201f), Ty: PtrTy); |
| 651 | for (GlobalValue *OldDS : DSsToDeactivate) { |
| 652 | GlobalValue *DS = GlobalAlias::create( |
| 653 | Ty: Int8Ty, AddressSpace: 0, Linkage: GlobalValue::ExternalLinkage, Name: OldDS->getName(), Aliasee: Nop, Parent: &M); |
| 654 | DS->setVisibility(GlobalValue::HiddenVisibility); |
| 655 | DS->takeName(V: OldDS); |
| 656 | OldDS->replaceAllUsesWith(V: DS); |
| 657 | OldDS->eraseFromParent(); |
| 658 | } |
| 659 | } |
| 660 | return true; |
| 661 | } |
| 662 | |
| 663 | static bool expandCondLoop(Function &Intr) { |
| 664 | for (User *U : llvm::make_early_inc_range(Range: Intr.users())) { |
| 665 | auto *Call = cast<CallInst>(Val: U); |
| 666 | |
| 667 | auto *Br = cast<UncondBrInst>( |
| 668 | Val: SplitBlockAndInsertIfThen(Cond: Call->getArgOperand(i: 0), SplitBefore: Call, Unreachable: false, |
| 669 | BranchWeights: getExplicitlyUnknownBranchWeightsIfProfiled( |
| 670 | F&: *Call->getFunction(), DEBUG_TYPE))); |
| 671 | Br->setSuccessor(Br->getParent()); |
| 672 | Call->eraseFromParent(); |
| 673 | } |
| 674 | return true; |
| 675 | } |
| 676 | |
| 677 | static bool expandLoopTrap(Function &Intr) { |
| 678 | for (User *U : make_early_inc_range(Range: Intr.users())) { |
| 679 | auto *Call = cast<CallInst>(Val: U); |
| 680 | if (!Call->getParent()->isEntryBlock() && |
| 681 | std::all_of(first: Call->getParent()->begin(), last: BasicBlock::iterator(Call), |
| 682 | pred: [](Instruction &I) { return !I.mayHaveSideEffects(); })) { |
| 683 | for (auto *BB : predecessors(BB: Call->getParent())) { |
| 684 | auto *BI = dyn_cast<CondBrInst>(Val: BB->getTerminator()); |
| 685 | if (!BI) |
| 686 | continue; |
| 687 | IRBuilder<> B(BI); |
| 688 | Value *Cond; |
| 689 | // The looptrap can either be on the true branch or the false branch. |
| 690 | // We insert the cond loop before the branch, which uses the branch's |
| 691 | // original condition for going to the looptrap as its condition, and |
| 692 | // force the branch to take whichever path does not lead to the |
| 693 | // looptrap, as the original path to the looptrap is now unreachable |
| 694 | // thanks to the cond loop. The codegenprepare pass will clean up our |
| 695 | // "unconditional conditional branch" by combining the two basic blocks |
| 696 | // if possible, or replacing it with an unconditional branch. |
| 697 | if (BI->getSuccessor(i: 0) == Call->getParent()) { |
| 698 | // The looptrap is on the true branch. |
| 699 | Cond = BI->getCondition(); |
| 700 | BI->setCondition(ConstantInt::getFalse(Context&: BI->getContext())); |
| 701 | } else { |
| 702 | // The looptrap is on the false branch, which means that we need to |
| 703 | // invert the condition. |
| 704 | Cond = B.CreateNot(V: BI->getCondition()); |
| 705 | BI->setCondition(ConstantInt::getTrue(Context&: BI->getContext())); |
| 706 | } |
| 707 | B.CreateIntrinsic(ID: Intrinsic::cond_loop, Args: Cond); |
| 708 | } |
| 709 | } |
| 710 | IRBuilder<> B(Call); |
| 711 | B.CreateIntrinsic(ID: Intrinsic::cond_loop, |
| 712 | Args: ConstantInt::getTrue(Context&: Call->getContext())); |
| 713 | Call->eraseFromParent(); |
| 714 | } |
| 715 | return true; |
| 716 | } |
| 717 | |
| 718 | bool PreISelIntrinsicLowering::lowerIntrinsics(Module &M) const { |
| 719 | // Map unique constants to globals. |
| 720 | DenseMap<Constant *, GlobalVariable *> CMap; |
| 721 | bool Changed = false; |
| 722 | for (Function &F : M) { |
| 723 | switch (F.getIntrinsicID()) { |
| 724 | default: |
| 725 | break; |
| 726 | case Intrinsic::memcpy: |
| 727 | case Intrinsic::memcpy_inline: |
| 728 | case Intrinsic::memmove: |
| 729 | case Intrinsic::memset: |
| 730 | case Intrinsic::memset_inline: |
| 731 | case Intrinsic::experimental_memset_pattern: |
| 732 | Changed |= expandMemIntrinsicUses(F, CMap); |
| 733 | break; |
| 734 | case Intrinsic::load_relative: |
| 735 | Changed |= lowerLoadRelative(F); |
| 736 | break; |
| 737 | case Intrinsic::can_load_speculatively: |
| 738 | Changed |= lowerCanLoadSpeculatively(F, TM); |
| 739 | break; |
| 740 | case Intrinsic::is_constant: |
| 741 | case Intrinsic::objectsize: |
| 742 | Changed |= forEachCall(Intrin&: F, Callback: [&](CallInst *CI) { |
| 743 | Function *Parent = CI->getParent()->getParent(); |
| 744 | TargetLibraryInfo &TLI = LookupTLI(*Parent); |
| 745 | // Intrinsics in unreachable code are not lowered. |
| 746 | bool Changed = lowerConstantIntrinsics(F&: *Parent, TLI, /*DT=*/nullptr); |
| 747 | return Changed; |
| 748 | }); |
| 749 | break; |
| 750 | #define BEGIN_REGISTER_VP_INTRINSIC(VPID, MASKPOS, VLENPOS) \ |
| 751 | case Intrinsic::VPID: |
| 752 | #include "llvm/IR/VPIntrinsics.def" |
| 753 | forEachCall(Intrin&: F, Callback: [&](CallInst *CI) { |
| 754 | Function *Parent = CI->getParent()->getParent(); |
| 755 | const TargetTransformInfo &TTI = LookupTTI(*Parent); |
| 756 | auto *VPI = cast<VPIntrinsic>(Val: CI); |
| 757 | VPExpansionDetails ED = expandVectorPredicationIntrinsic(VPI&: *VPI, TTI); |
| 758 | // Expansion of VP intrinsics may change the IR but not actually |
| 759 | // replace the intrinsic, so update Changed for the pass |
| 760 | // and compute Removed for forEachCall. |
| 761 | Changed |= ED != VPExpansionDetails::IntrinsicUnchanged; |
| 762 | bool Removed = ED == VPExpansionDetails::IntrinsicReplaced; |
| 763 | return Removed; |
| 764 | }); |
| 765 | break; |
| 766 | case Intrinsic::objc_autorelease: |
| 767 | Changed |= lowerObjCCall(F, NewFn: RTLIB::impl_objc_autorelease); |
| 768 | break; |
| 769 | case Intrinsic::objc_autoreleasePoolPop: |
| 770 | Changed |= lowerObjCCall(F, NewFn: RTLIB::impl_objc_autoreleasePoolPop); |
| 771 | break; |
| 772 | case Intrinsic::objc_autoreleasePoolPush: |
| 773 | Changed |= lowerObjCCall(F, NewFn: RTLIB::impl_objc_autoreleasePoolPush); |
| 774 | break; |
| 775 | case Intrinsic::objc_autoreleaseReturnValue: |
| 776 | Changed |= lowerObjCCall(F, NewFn: RTLIB::impl_objc_autoreleaseReturnValue); |
| 777 | break; |
| 778 | case Intrinsic::objc_copyWeak: |
| 779 | Changed |= lowerObjCCall(F, NewFn: RTLIB::impl_objc_copyWeak); |
| 780 | break; |
| 781 | case Intrinsic::objc_destroyWeak: |
| 782 | Changed |= lowerObjCCall(F, NewFn: RTLIB::impl_objc_destroyWeak); |
| 783 | break; |
| 784 | case Intrinsic::objc_initWeak: |
| 785 | Changed |= lowerObjCCall(F, NewFn: RTLIB::impl_objc_initWeak); |
| 786 | break; |
| 787 | case Intrinsic::objc_loadWeak: |
| 788 | Changed |= lowerObjCCall(F, NewFn: RTLIB::impl_objc_loadWeak); |
| 789 | break; |
| 790 | case Intrinsic::objc_loadWeakRetained: |
| 791 | Changed |= lowerObjCCall(F, NewFn: RTLIB::impl_objc_loadWeakRetained); |
| 792 | break; |
| 793 | case Intrinsic::objc_moveWeak: |
| 794 | Changed |= lowerObjCCall(F, NewFn: RTLIB::impl_objc_moveWeak); |
| 795 | break; |
| 796 | case Intrinsic::objc_release: |
| 797 | Changed |= lowerObjCCall(F, NewFn: RTLIB::impl_objc_release, setNonLazyBind: true); |
| 798 | break; |
| 799 | case Intrinsic::objc_retain: |
| 800 | Changed |= lowerObjCCall(F, NewFn: RTLIB::impl_objc_retain, setNonLazyBind: true); |
| 801 | break; |
| 802 | case Intrinsic::objc_retainAutorelease: |
| 803 | Changed |= lowerObjCCall(F, NewFn: RTLIB::impl_objc_retainAutorelease); |
| 804 | break; |
| 805 | case Intrinsic::objc_retainAutoreleaseReturnValue: |
| 806 | Changed |= |
| 807 | lowerObjCCall(F, NewFn: RTLIB::impl_objc_retainAutoreleaseReturnValue); |
| 808 | break; |
| 809 | case Intrinsic::objc_retainAutoreleasedReturnValue: |
| 810 | Changed |= |
| 811 | lowerObjCCall(F, NewFn: RTLIB::impl_objc_retainAutoreleasedReturnValue); |
| 812 | break; |
| 813 | case Intrinsic::objc_claimAutoreleasedReturnValue: |
| 814 | Changed |= |
| 815 | lowerObjCCall(F, NewFn: RTLIB::impl_objc_claimAutoreleasedReturnValue); |
| 816 | break; |
| 817 | case Intrinsic::objc_retainBlock: |
| 818 | Changed |= lowerObjCCall(F, NewFn: RTLIB::impl_objc_retainBlock); |
| 819 | break; |
| 820 | case Intrinsic::objc_storeStrong: |
| 821 | Changed |= lowerObjCCall(F, NewFn: RTLIB::impl_objc_storeStrong); |
| 822 | break; |
| 823 | case Intrinsic::objc_storeWeak: |
| 824 | Changed |= lowerObjCCall(F, NewFn: RTLIB::impl_objc_storeWeak); |
| 825 | break; |
| 826 | case Intrinsic::objc_unsafeClaimAutoreleasedReturnValue: |
| 827 | Changed |= |
| 828 | lowerObjCCall(F, NewFn: RTLIB::impl_objc_unsafeClaimAutoreleasedReturnValue); |
| 829 | break; |
| 830 | case Intrinsic::objc_retainedObject: |
| 831 | Changed |= lowerObjCCall(F, NewFn: RTLIB::impl_objc_retainedObject); |
| 832 | break; |
| 833 | case Intrinsic::objc_unretainedObject: |
| 834 | Changed |= lowerObjCCall(F, NewFn: RTLIB::impl_objc_unretainedObject); |
| 835 | break; |
| 836 | case Intrinsic::objc_unretainedPointer: |
| 837 | Changed |= lowerObjCCall(F, NewFn: RTLIB::impl_objc_unretainedPointer); |
| 838 | break; |
| 839 | case Intrinsic::objc_retain_autorelease: |
| 840 | Changed |= lowerObjCCall(F, NewFn: RTLIB::impl_objc_retain_autorelease); |
| 841 | break; |
| 842 | case Intrinsic::objc_sync_enter: |
| 843 | Changed |= lowerObjCCall(F, NewFn: RTLIB::impl_objc_sync_enter); |
| 844 | break; |
| 845 | case Intrinsic::objc_sync_exit: |
| 846 | Changed |= lowerObjCCall(F, NewFn: RTLIB::impl_objc_sync_exit); |
| 847 | break; |
| 848 | case Intrinsic::acos: |
| 849 | case Intrinsic::asin: |
| 850 | case Intrinsic::atan: |
| 851 | case Intrinsic::cos: |
| 852 | case Intrinsic::cosh: |
| 853 | case Intrinsic::exp: |
| 854 | case Intrinsic::exp2: |
| 855 | case Intrinsic::exp10: |
| 856 | case Intrinsic::log: |
| 857 | case Intrinsic::log2: |
| 858 | case Intrinsic::log10: |
| 859 | case Intrinsic::sin: |
| 860 | case Intrinsic::sinh: |
| 861 | case Intrinsic::tan: |
| 862 | case Intrinsic::tanh: |
| 863 | Changed |= forEachCall(Intrin&: F, Callback: [&](CallInst *CI) { |
| 864 | Type *Ty = CI->getArgOperand(i: 0)->getType(); |
| 865 | if (!isa<ScalableVectorType>(Val: Ty)) |
| 866 | return false; |
| 867 | const TargetLowering *TL = TM.getSubtargetImpl(F)->getTargetLowering(); |
| 868 | unsigned Op = TL->IntrinsicIDToISD(ID: F.getIntrinsicID()); |
| 869 | assert(Op != ISD::DELETED_NODE && "unsupported intrinsic" ); |
| 870 | if (!TL->isOperationExpand(Op, VT: EVT::getEVT(Ty))) |
| 871 | return false; |
| 872 | return lowerUnaryVectorIntrinsicAsLoop(M, CI); |
| 873 | }); |
| 874 | break; |
| 875 | case Intrinsic::modf: |
| 876 | case Intrinsic::sincos: |
| 877 | case Intrinsic::sincospi: |
| 878 | Changed |= forEachCall(Intrin&: F, Callback: [&](CallInst *CI) { |
| 879 | Type *Ty = CI->getArgOperand(i: 0)->getType(); |
| 880 | if (!isa<ScalableVectorType>(Val: Ty)) |
| 881 | return false; |
| 882 | const TargetSubtargetInfo *Subtarget = TM.getSubtargetImpl(F); |
| 883 | const TargetLowering *TL = Subtarget->getTargetLowering(); |
| 884 | unsigned Op = TL->IntrinsicIDToISD(ID: F.getIntrinsicID()); |
| 885 | assert(Op != ISD::DELETED_NODE && "unsupported intrinsic" ); |
| 886 | EVT VT = EVT::getEVT(Ty); |
| 887 | if (!TL->isOperationExpand(Op, VT)) |
| 888 | return false; |
| 889 | // The vector legalizer can expand these to a vector math library call. |
| 890 | RTLIB::Libcall LC; |
| 891 | switch (Op) { |
| 892 | case ISD::FMODF: |
| 893 | LC = RTLIB::getMODF(VT); |
| 894 | break; |
| 895 | case ISD::FSINCOS: |
| 896 | LC = RTLIB::getSINCOS(VT); |
| 897 | break; |
| 898 | case ISD::FSINCOSPI: |
| 899 | LC = RTLIB::getSINCOSPI(VT); |
| 900 | break; |
| 901 | default: |
| 902 | llvm_unreachable("unexpected intrinsic" ); |
| 903 | } |
| 904 | |
| 905 | const LibcallLoweringInfo &Libcalls = |
| 906 | getLibcallLowering(ModuleInfo: ModuleLibcalls, Subtarget: *Subtarget); |
| 907 | if (Libcalls.getVectorLibcallImplOrMasked(Call: LC) != RTLIB::Unsupported) |
| 908 | return false; |
| 909 | return lowerUnaryVectorIntrinsicAsLoop(M, CI); |
| 910 | }); |
| 911 | break; |
| 912 | case Intrinsic::ptrauth_sign: |
| 913 | case Intrinsic::ptrauth_auth: |
| 914 | Changed |= expandPtrauthForEmuPAC(Intr&: F); |
| 915 | break; |
| 916 | case Intrinsic::protected_field_ptr: |
| 917 | Changed |= expandProtectedFieldPtr(Intr&: F); |
| 918 | break; |
| 919 | case Intrinsic::cond_loop: |
| 920 | if (!TM.canLowerCondLoop()) |
| 921 | Changed |= expandCondLoop(Intr&: F); |
| 922 | break; |
| 923 | case Intrinsic::looptrap: |
| 924 | Changed |= expandLoopTrap(Intr&: F); |
| 925 | if (!TM.canLowerCondLoop()) |
| 926 | if (auto *CondLoop = M.getFunction(Name: "llvm.cond.loop" )) |
| 927 | Changed |= expandCondLoop(Intr&: *CondLoop); |
| 928 | break; |
| 929 | } |
| 930 | } |
| 931 | return Changed; |
| 932 | } |
| 933 | |
| 934 | namespace { |
| 935 | |
| 936 | class PreISelIntrinsicLoweringLegacyPass : public ModulePass { |
| 937 | public: |
| 938 | static char ID; |
| 939 | |
| 940 | PreISelIntrinsicLoweringLegacyPass() : ModulePass(ID) {} |
| 941 | |
| 942 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
| 943 | AU.addRequired<TargetTransformInfoWrapperPass>(); |
| 944 | AU.addRequired<TargetLibraryInfoWrapperPass>(); |
| 945 | AU.addRequired<LibcallLoweringInfoWrapper>(); |
| 946 | AU.addRequired<TargetPassConfig>(); |
| 947 | AU.addRequired<AssumptionCacheTracker>(); |
| 948 | } |
| 949 | |
| 950 | bool runOnModule(Module &M) override { |
| 951 | const ModuleLibcallLoweringInfo &ModuleLibcalls = |
| 952 | getAnalysis<LibcallLoweringInfoWrapper>().getResult(M); |
| 953 | |
| 954 | auto LookupTTI = [this](Function &F) -> TargetTransformInfo & { |
| 955 | return this->getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); |
| 956 | }; |
| 957 | auto LookupTLI = [this](Function &F) -> TargetLibraryInfo & { |
| 958 | return this->getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F); |
| 959 | }; |
| 960 | auto LookupAC = [this](Function &F) -> AssumptionCache & { |
| 961 | return this->getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); |
| 962 | }; |
| 963 | |
| 964 | const auto &TM = getAnalysis<TargetPassConfig>().getTM<TargetMachine>(); |
| 965 | PreISelIntrinsicLowering Lowering(getCodeGenOptions(M), TM, ModuleLibcalls, |
| 966 | LookupTTI, LookupTLI, LookupAC); |
| 967 | return Lowering.lowerIntrinsics(M); |
| 968 | } |
| 969 | }; |
| 970 | |
| 971 | } // end anonymous namespace |
| 972 | |
| 973 | char PreISelIntrinsicLoweringLegacyPass::ID; |
| 974 | |
| 975 | INITIALIZE_PASS_BEGIN(PreISelIntrinsicLoweringLegacyPass, |
| 976 | "pre-isel-intrinsic-lowering" , |
| 977 | "Pre-ISel Intrinsic Lowering" , false, false) |
| 978 | INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) |
| 979 | INITIALIZE_PASS_DEPENDENCY(LibcallLoweringInfoWrapper) |
| 980 | INITIALIZE_PASS_DEPENDENCY(RuntimeLibraryInfoWrapper) |
| 981 | INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) |
| 982 | INITIALIZE_PASS_DEPENDENCY(TargetPassConfig) |
| 983 | INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) |
| 984 | INITIALIZE_PASS_END(PreISelIntrinsicLoweringLegacyPass, |
| 985 | "pre-isel-intrinsic-lowering" , |
| 986 | "Pre-ISel Intrinsic Lowering" , false, false) |
| 987 | |
| 988 | ModulePass *llvm::createPreISelIntrinsicLoweringPass() { |
| 989 | return new PreISelIntrinsicLoweringLegacyPass(); |
| 990 | } |
| 991 | |
| 992 | PreservedAnalyses |
| 993 | PreISelIntrinsicLoweringPass::run(Module &M, ModuleAnalysisManager &MAM) { |
| 994 | const ModuleLibcallLoweringInfo &LibcallLowering = |
| 995 | MAM.getResult<LibcallLoweringModuleAnalysis>(IR&: M); |
| 996 | |
| 997 | auto &FAM = MAM.getResult<FunctionAnalysisManagerModuleProxy>(IR&: M).getManager(); |
| 998 | |
| 999 | auto LookupTTI = [&FAM](Function &F) -> TargetTransformInfo & { |
| 1000 | return FAM.getResult<TargetIRAnalysis>(IR&: F); |
| 1001 | }; |
| 1002 | auto LookupTLI = [&FAM](Function &F) -> TargetLibraryInfo & { |
| 1003 | return FAM.getResult<TargetLibraryAnalysis>(IR&: F); |
| 1004 | }; |
| 1005 | auto LookupAC = [&FAM](Function &F) -> AssumptionCache & { |
| 1006 | return FAM.getResult<AssumptionAnalysis>(IR&: F); |
| 1007 | }; |
| 1008 | |
| 1009 | PreISelIntrinsicLowering Lowering(getCodeGenOptions(M), TM, LibcallLowering, |
| 1010 | LookupTTI, LookupTLI, LookupAC); |
| 1011 | if (!Lowering.lowerIntrinsics(M)) |
| 1012 | return PreservedAnalyses::all(); |
| 1013 | else |
| 1014 | return PreservedAnalyses::none(); |
| 1015 | } |
| 1016 | |