1 //===- llvm/CodeGen/GlobalISel/IRTranslator.cpp - IRTranslator ---*- C++ -*-==// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 /// \file 10 /// This file implements the IRTranslator class. 11 //===----------------------------------------------------------------------===// 12 13 #include "llvm/CodeGen/GlobalISel/IRTranslator.h" 14 #include "llvm/ADT/PostOrderIterator.h" 15 #include "llvm/ADT/STLExtras.h" 16 #include "llvm/ADT/ScopeExit.h" 17 #include "llvm/ADT/SmallSet.h" 18 #include "llvm/ADT/SmallVector.h" 19 #include "llvm/Analysis/OptimizationRemarkEmitter.h" 20 #include "llvm/CodeGen/Analysis.h" 21 #include "llvm/CodeGen/GlobalISel/CallLowering.h" 22 #include "llvm/CodeGen/LowLevelType.h" 23 #include "llvm/CodeGen/MachineBasicBlock.h" 24 #include "llvm/CodeGen/MachineFrameInfo.h" 25 #include "llvm/CodeGen/MachineFunction.h" 26 #include "llvm/CodeGen/MachineInstrBuilder.h" 27 #include "llvm/CodeGen/MachineMemOperand.h" 28 #include "llvm/CodeGen/MachineOperand.h" 29 #include "llvm/CodeGen/MachineRegisterInfo.h" 30 #include "llvm/CodeGen/StackProtector.h" 31 #include "llvm/CodeGen/TargetFrameLowering.h" 32 #include "llvm/CodeGen/TargetLowering.h" 33 #include "llvm/CodeGen/TargetPassConfig.h" 34 #include "llvm/CodeGen/TargetRegisterInfo.h" 35 #include "llvm/CodeGen/TargetSubtargetInfo.h" 36 #include "llvm/IR/BasicBlock.h" 37 #include "llvm/IR/CFG.h" 38 #include "llvm/IR/Constant.h" 39 #include "llvm/IR/Constants.h" 40 #include "llvm/IR/DataLayout.h" 41 #include "llvm/IR/DebugInfo.h" 42 #include "llvm/IR/DerivedTypes.h" 43 #include "llvm/IR/Function.h" 44 #include "llvm/IR/GetElementPtrTypeIterator.h" 45 #include "llvm/IR/InlineAsm.h" 46 #include "llvm/IR/InstrTypes.h" 47 #include "llvm/IR/Instructions.h" 48 #include "llvm/IR/IntrinsicInst.h" 49 #include "llvm/IR/Intrinsics.h" 50 #include "llvm/IR/LLVMContext.h" 51 #include "llvm/IR/Metadata.h" 52 #include "llvm/IR/Type.h" 53 #include "llvm/IR/User.h" 54 #include "llvm/IR/Value.h" 55 #include "llvm/MC/MCContext.h" 56 #include "llvm/Pass.h" 57 #include "llvm/Support/Casting.h" 58 #include "llvm/Support/CodeGen.h" 59 #include "llvm/Support/Debug.h" 60 #include "llvm/Support/ErrorHandling.h" 61 #include "llvm/Support/LowLevelTypeImpl.h" 62 #include "llvm/Support/MathExtras.h" 63 #include "llvm/Support/raw_ostream.h" 64 #include "llvm/Target/TargetIntrinsicInfo.h" 65 #include "llvm/Target/TargetMachine.h" 66 #include <algorithm> 67 #include <cassert> 68 #include <cstdint> 69 #include <iterator> 70 #include <string> 71 #include <utility> 72 #include <vector> 73 74 #define DEBUG_TYPE "irtranslator" 75 76 using namespace llvm; 77 78 char IRTranslator::ID = 0; 79 80 INITIALIZE_PASS_BEGIN(IRTranslator, DEBUG_TYPE, "IRTranslator LLVM IR -> MI", 81 false, false) 82 INITIALIZE_PASS_DEPENDENCY(TargetPassConfig) 83 INITIALIZE_PASS_END(IRTranslator, DEBUG_TYPE, "IRTranslator LLVM IR -> MI", 84 false, false) 85 86 static void reportTranslationError(MachineFunction &MF, 87 const TargetPassConfig &TPC, 88 OptimizationRemarkEmitter &ORE, 89 OptimizationRemarkMissed &R) { 90 MF.getProperties().set(MachineFunctionProperties::Property::FailedISel); 91 92 // Print the function name explicitly if we don't have a debug location (which 93 // makes the diagnostic less useful) or if we're going to emit a raw error. 94 if (!R.getLocation().isValid() || TPC.isGlobalISelAbortEnabled()) 95 R << (" (in function: " + MF.getName() + ")").str(); 96 97 if (TPC.isGlobalISelAbortEnabled()) 98 report_fatal_error(R.getMsg()); 99 else 100 ORE.emit(R); 101 } 102 103 IRTranslator::IRTranslator() : MachineFunctionPass(ID) { 104 initializeIRTranslatorPass(*PassRegistry::getPassRegistry()); 105 } 106 107 void IRTranslator::getAnalysisUsage(AnalysisUsage &AU) const { 108 AU.addRequired<StackProtector>(); 109 AU.addRequired<TargetPassConfig>(); 110 getSelectionDAGFallbackAnalysisUsage(AU); 111 MachineFunctionPass::getAnalysisUsage(AU); 112 } 113 114 static void computeValueLLTs(const DataLayout &DL, Type &Ty, 115 SmallVectorImpl<LLT> &ValueTys, 116 SmallVectorImpl<uint64_t> *Offsets = nullptr, 117 uint64_t StartingOffset = 0) { 118 // Given a struct type, recursively traverse the elements. 119 if (StructType *STy = dyn_cast<StructType>(&Ty)) { 120 const StructLayout *SL = DL.getStructLayout(STy); 121 for (unsigned I = 0, E = STy->getNumElements(); I != E; ++I) 122 computeValueLLTs(DL, *STy->getElementType(I), ValueTys, Offsets, 123 StartingOffset + SL->getElementOffset(I)); 124 return; 125 } 126 // Given an array type, recursively traverse the elements. 127 if (ArrayType *ATy = dyn_cast<ArrayType>(&Ty)) { 128 Type *EltTy = ATy->getElementType(); 129 uint64_t EltSize = DL.getTypeAllocSize(EltTy); 130 for (unsigned i = 0, e = ATy->getNumElements(); i != e; ++i) 131 computeValueLLTs(DL, *EltTy, ValueTys, Offsets, 132 StartingOffset + i * EltSize); 133 return; 134 } 135 // Interpret void as zero return values. 136 if (Ty.isVoidTy()) 137 return; 138 // Base case: we can get an LLT for this LLVM IR type. 139 ValueTys.push_back(getLLTForType(Ty, DL)); 140 if (Offsets != nullptr) 141 Offsets->push_back(StartingOffset * 8); 142 } 143 144 IRTranslator::ValueToVRegInfo::VRegListT & 145 IRTranslator::allocateVRegs(const Value &Val) { 146 assert(!VMap.contains(Val) && "Value already allocated in VMap"); 147 auto *Regs = VMap.getVRegs(Val); 148 auto *Offsets = VMap.getOffsets(Val); 149 SmallVector<LLT, 4> SplitTys; 150 computeValueLLTs(*DL, *Val.getType(), SplitTys, 151 Offsets->empty() ? Offsets : nullptr); 152 for (unsigned i = 0; i < SplitTys.size(); ++i) 153 Regs->push_back(0); 154 return *Regs; 155 } 156 157 ArrayRef<unsigned> IRTranslator::getOrCreateVRegs(const Value &Val) { 158 auto VRegsIt = VMap.findVRegs(Val); 159 if (VRegsIt != VMap.vregs_end()) 160 return *VRegsIt->second; 161 162 if (Val.getType()->isVoidTy()) 163 return *VMap.getVRegs(Val); 164 165 // Create entry for this type. 166 auto *VRegs = VMap.getVRegs(Val); 167 auto *Offsets = VMap.getOffsets(Val); 168 169 assert(Val.getType()->isSized() && 170 "Don't know how to create an empty vreg"); 171 172 SmallVector<LLT, 4> SplitTys; 173 computeValueLLTs(*DL, *Val.getType(), SplitTys, 174 Offsets->empty() ? Offsets : nullptr); 175 176 if (!isa<Constant>(Val)) { 177 for (auto Ty : SplitTys) 178 VRegs->push_back(MRI->createGenericVirtualRegister(Ty)); 179 return *VRegs; 180 } 181 182 if (Val.getType()->isAggregateType()) { 183 // UndefValue, ConstantAggregateZero 184 auto &C = cast<Constant>(Val); 185 unsigned Idx = 0; 186 while (auto Elt = C.getAggregateElement(Idx++)) { 187 auto EltRegs = getOrCreateVRegs(*Elt); 188 std::copy(EltRegs.begin(), EltRegs.end(), std::back_inserter(*VRegs)); 189 } 190 } else { 191 assert(SplitTys.size() == 1 && "unexpectedly split LLT"); 192 VRegs->push_back(MRI->createGenericVirtualRegister(SplitTys[0])); 193 bool Success = translate(cast<Constant>(Val), VRegs->front()); 194 if (!Success) { 195 OptimizationRemarkMissed R("gisel-irtranslator", "GISelFailure", 196 MF->getFunction().getSubprogram(), 197 &MF->getFunction().getEntryBlock()); 198 R << "unable to translate constant: " << ore::NV("Type", Val.getType()); 199 reportTranslationError(*MF, *TPC, *ORE, R); 200 return *VRegs; 201 } 202 } 203 204 return *VRegs; 205 } 206 207 int IRTranslator::getOrCreateFrameIndex(const AllocaInst &AI) { 208 if (FrameIndices.find(&AI) != FrameIndices.end()) 209 return FrameIndices[&AI]; 210 211 unsigned ElementSize = DL->getTypeStoreSize(AI.getAllocatedType()); 212 unsigned Size = 213 ElementSize * cast<ConstantInt>(AI.getArraySize())->getZExtValue(); 214 215 // Always allocate at least one byte. 216 Size = std::max(Size, 1u); 217 218 unsigned Alignment = AI.getAlignment(); 219 if (!Alignment) 220 Alignment = DL->getABITypeAlignment(AI.getAllocatedType()); 221 222 int &FI = FrameIndices[&AI]; 223 FI = MF->getFrameInfo().CreateStackObject(Size, Alignment, false, &AI); 224 return FI; 225 } 226 227 unsigned IRTranslator::getMemOpAlignment(const Instruction &I) { 228 unsigned Alignment = 0; 229 Type *ValTy = nullptr; 230 if (const StoreInst *SI = dyn_cast<StoreInst>(&I)) { 231 Alignment = SI->getAlignment(); 232 ValTy = SI->getValueOperand()->getType(); 233 } else if (const LoadInst *LI = dyn_cast<LoadInst>(&I)) { 234 Alignment = LI->getAlignment(); 235 ValTy = LI->getType(); 236 } else if (const AtomicCmpXchgInst *AI = dyn_cast<AtomicCmpXchgInst>(&I)) { 237 // TODO(PR27168): This instruction has no alignment attribute, but unlike 238 // the default alignment for load/store, the default here is to assume 239 // it has NATURAL alignment, not DataLayout-specified alignment. 240 const DataLayout &DL = AI->getModule()->getDataLayout(); 241 Alignment = DL.getTypeStoreSize(AI->getCompareOperand()->getType()); 242 ValTy = AI->getCompareOperand()->getType(); 243 } else if (const AtomicRMWInst *AI = dyn_cast<AtomicRMWInst>(&I)) { 244 // TODO(PR27168): This instruction has no alignment attribute, but unlike 245 // the default alignment for load/store, the default here is to assume 246 // it has NATURAL alignment, not DataLayout-specified alignment. 247 const DataLayout &DL = AI->getModule()->getDataLayout(); 248 Alignment = DL.getTypeStoreSize(AI->getValOperand()->getType()); 249 ValTy = AI->getType(); 250 } else { 251 OptimizationRemarkMissed R("gisel-irtranslator", "", &I); 252 R << "unable to translate memop: " << ore::NV("Opcode", &I); 253 reportTranslationError(*MF, *TPC, *ORE, R); 254 return 1; 255 } 256 257 return Alignment ? Alignment : DL->getABITypeAlignment(ValTy); 258 } 259 260 MachineBasicBlock &IRTranslator::getMBB(const BasicBlock &BB) { 261 MachineBasicBlock *&MBB = BBToMBB[&BB]; 262 assert(MBB && "BasicBlock was not encountered before"); 263 return *MBB; 264 } 265 266 void IRTranslator::addMachineCFGPred(CFGEdge Edge, MachineBasicBlock *NewPred) { 267 assert(NewPred && "new predecessor must be a real MachineBasicBlock"); 268 MachinePreds[Edge].push_back(NewPred); 269 } 270 271 bool IRTranslator::translateBinaryOp(unsigned Opcode, const User &U, 272 MachineIRBuilder &MIRBuilder) { 273 // FIXME: handle signed/unsigned wrapping flags. 274 275 // Get or create a virtual register for each value. 276 // Unless the value is a Constant => loadimm cst? 277 // or inline constant each time? 278 // Creation of a virtual register needs to have a size. 279 unsigned Op0 = getOrCreateVReg(*U.getOperand(0)); 280 unsigned Op1 = getOrCreateVReg(*U.getOperand(1)); 281 unsigned Res = getOrCreateVReg(U); 282 MIRBuilder.buildInstr(Opcode).addDef(Res).addUse(Op0).addUse(Op1); 283 return true; 284 } 285 286 bool IRTranslator::translateFSub(const User &U, MachineIRBuilder &MIRBuilder) { 287 // -0.0 - X --> G_FNEG 288 if (isa<Constant>(U.getOperand(0)) && 289 U.getOperand(0) == ConstantFP::getZeroValueForNegation(U.getType())) { 290 MIRBuilder.buildInstr(TargetOpcode::G_FNEG) 291 .addDef(getOrCreateVReg(U)) 292 .addUse(getOrCreateVReg(*U.getOperand(1))); 293 return true; 294 } 295 return translateBinaryOp(TargetOpcode::G_FSUB, U, MIRBuilder); 296 } 297 298 bool IRTranslator::translateCompare(const User &U, 299 MachineIRBuilder &MIRBuilder) { 300 const CmpInst *CI = dyn_cast<CmpInst>(&U); 301 unsigned Op0 = getOrCreateVReg(*U.getOperand(0)); 302 unsigned Op1 = getOrCreateVReg(*U.getOperand(1)); 303 unsigned Res = getOrCreateVReg(U); 304 CmpInst::Predicate Pred = 305 CI ? CI->getPredicate() : static_cast<CmpInst::Predicate>( 306 cast<ConstantExpr>(U).getPredicate()); 307 if (CmpInst::isIntPredicate(Pred)) 308 MIRBuilder.buildICmp(Pred, Res, Op0, Op1); 309 else if (Pred == CmpInst::FCMP_FALSE) 310 MIRBuilder.buildCopy( 311 Res, getOrCreateVReg(*Constant::getNullValue(CI->getType()))); 312 else if (Pred == CmpInst::FCMP_TRUE) 313 MIRBuilder.buildCopy( 314 Res, getOrCreateVReg(*Constant::getAllOnesValue(CI->getType()))); 315 else 316 MIRBuilder.buildFCmp(Pred, Res, Op0, Op1); 317 318 return true; 319 } 320 321 bool IRTranslator::translateRet(const User &U, MachineIRBuilder &MIRBuilder) { 322 const ReturnInst &RI = cast<ReturnInst>(U); 323 const Value *Ret = RI.getReturnValue(); 324 if (Ret && DL->getTypeStoreSize(Ret->getType()) == 0) 325 Ret = nullptr; 326 327 ArrayRef<unsigned> VRegs; 328 if (Ret) 329 VRegs = getOrCreateVRegs(*Ret); 330 331 // The target may mess up with the insertion point, but 332 // this is not important as a return is the last instruction 333 // of the block anyway. 334 335 return CLI->lowerReturn(MIRBuilder, Ret, VRegs); 336 } 337 338 bool IRTranslator::translateBr(const User &U, MachineIRBuilder &MIRBuilder) { 339 const BranchInst &BrInst = cast<BranchInst>(U); 340 unsigned Succ = 0; 341 if (!BrInst.isUnconditional()) { 342 // We want a G_BRCOND to the true BB followed by an unconditional branch. 343 unsigned Tst = getOrCreateVReg(*BrInst.getCondition()); 344 const BasicBlock &TrueTgt = *cast<BasicBlock>(BrInst.getSuccessor(Succ++)); 345 MachineBasicBlock &TrueBB = getMBB(TrueTgt); 346 MIRBuilder.buildBrCond(Tst, TrueBB); 347 } 348 349 const BasicBlock &BrTgt = *cast<BasicBlock>(BrInst.getSuccessor(Succ)); 350 MachineBasicBlock &TgtBB = getMBB(BrTgt); 351 MachineBasicBlock &CurBB = MIRBuilder.getMBB(); 352 353 // If the unconditional target is the layout successor, fallthrough. 354 if (!CurBB.isLayoutSuccessor(&TgtBB)) 355 MIRBuilder.buildBr(TgtBB); 356 357 // Link successors. 358 for (const BasicBlock *Succ : BrInst.successors()) 359 CurBB.addSuccessor(&getMBB(*Succ)); 360 return true; 361 } 362 363 bool IRTranslator::translateSwitch(const User &U, 364 MachineIRBuilder &MIRBuilder) { 365 // For now, just translate as a chain of conditional branches. 366 // FIXME: could we share most of the logic/code in 367 // SelectionDAGBuilder::visitSwitch between SelectionDAG and GlobalISel? 368 // At first sight, it seems most of the logic in there is independent of 369 // SelectionDAG-specifics and a lot of work went in to optimize switch 370 // lowering in there. 371 372 const SwitchInst &SwInst = cast<SwitchInst>(U); 373 const unsigned SwCondValue = getOrCreateVReg(*SwInst.getCondition()); 374 const BasicBlock *OrigBB = SwInst.getParent(); 375 376 LLT LLTi1 = getLLTForType(*Type::getInt1Ty(U.getContext()), *DL); 377 for (auto &CaseIt : SwInst.cases()) { 378 const unsigned CaseValueReg = getOrCreateVReg(*CaseIt.getCaseValue()); 379 const unsigned Tst = MRI->createGenericVirtualRegister(LLTi1); 380 MIRBuilder.buildICmp(CmpInst::ICMP_EQ, Tst, CaseValueReg, SwCondValue); 381 MachineBasicBlock &CurMBB = MIRBuilder.getMBB(); 382 const BasicBlock *TrueBB = CaseIt.getCaseSuccessor(); 383 MachineBasicBlock &TrueMBB = getMBB(*TrueBB); 384 385 MIRBuilder.buildBrCond(Tst, TrueMBB); 386 CurMBB.addSuccessor(&TrueMBB); 387 addMachineCFGPred({OrigBB, TrueBB}, &CurMBB); 388 389 MachineBasicBlock *FalseMBB = 390 MF->CreateMachineBasicBlock(SwInst.getParent()); 391 // Insert the comparison blocks one after the other. 392 MF->insert(std::next(CurMBB.getIterator()), FalseMBB); 393 MIRBuilder.buildBr(*FalseMBB); 394 CurMBB.addSuccessor(FalseMBB); 395 396 MIRBuilder.setMBB(*FalseMBB); 397 } 398 // handle default case 399 const BasicBlock *DefaultBB = SwInst.getDefaultDest(); 400 MachineBasicBlock &DefaultMBB = getMBB(*DefaultBB); 401 MIRBuilder.buildBr(DefaultMBB); 402 MachineBasicBlock &CurMBB = MIRBuilder.getMBB(); 403 CurMBB.addSuccessor(&DefaultMBB); 404 addMachineCFGPred({OrigBB, DefaultBB}, &CurMBB); 405 406 return true; 407 } 408 409 bool IRTranslator::translateIndirectBr(const User &U, 410 MachineIRBuilder &MIRBuilder) { 411 const IndirectBrInst &BrInst = cast<IndirectBrInst>(U); 412 413 const unsigned Tgt = getOrCreateVReg(*BrInst.getAddress()); 414 MIRBuilder.buildBrIndirect(Tgt); 415 416 // Link successors. 417 MachineBasicBlock &CurBB = MIRBuilder.getMBB(); 418 for (const BasicBlock *Succ : BrInst.successors()) 419 CurBB.addSuccessor(&getMBB(*Succ)); 420 421 return true; 422 } 423 424 bool IRTranslator::translateLoad(const User &U, MachineIRBuilder &MIRBuilder) { 425 const LoadInst &LI = cast<LoadInst>(U); 426 427 auto Flags = LI.isVolatile() ? MachineMemOperand::MOVolatile 428 : MachineMemOperand::MONone; 429 Flags |= MachineMemOperand::MOLoad; 430 431 if (DL->getTypeStoreSize(LI.getType()) == 0) 432 return true; 433 434 ArrayRef<unsigned> Regs = getOrCreateVRegs(LI); 435 ArrayRef<uint64_t> Offsets = *VMap.getOffsets(LI); 436 unsigned Base = getOrCreateVReg(*LI.getPointerOperand()); 437 438 for (unsigned i = 0; i < Regs.size(); ++i) { 439 unsigned Addr = 0; 440 MIRBuilder.materializeGEP(Addr, Base, LLT::scalar(64), Offsets[i] / 8); 441 442 MachinePointerInfo Ptr(LI.getPointerOperand(), Offsets[i] / 8); 443 unsigned BaseAlign = getMemOpAlignment(LI); 444 auto MMO = MF->getMachineMemOperand( 445 Ptr, Flags, (MRI->getType(Regs[i]).getSizeInBits() + 7) / 8, 446 MinAlign(BaseAlign, Offsets[i] / 8), AAMDNodes(), nullptr, 447 LI.getSyncScopeID(), LI.getOrdering()); 448 MIRBuilder.buildLoad(Regs[i], Addr, *MMO); 449 } 450 451 return true; 452 } 453 454 bool IRTranslator::translateStore(const User &U, MachineIRBuilder &MIRBuilder) { 455 const StoreInst &SI = cast<StoreInst>(U); 456 auto Flags = SI.isVolatile() ? MachineMemOperand::MOVolatile 457 : MachineMemOperand::MONone; 458 Flags |= MachineMemOperand::MOStore; 459 460 if (DL->getTypeStoreSize(SI.getValueOperand()->getType()) == 0) 461 return true; 462 463 ArrayRef<unsigned> Vals = getOrCreateVRegs(*SI.getValueOperand()); 464 ArrayRef<uint64_t> Offsets = *VMap.getOffsets(*SI.getValueOperand()); 465 unsigned Base = getOrCreateVReg(*SI.getPointerOperand()); 466 467 for (unsigned i = 0; i < Vals.size(); ++i) { 468 unsigned Addr = 0; 469 MIRBuilder.materializeGEP(Addr, Base, LLT::scalar(64), Offsets[i] / 8); 470 471 MachinePointerInfo Ptr(SI.getPointerOperand(), Offsets[i] / 8); 472 unsigned BaseAlign = getMemOpAlignment(SI); 473 auto MMO = MF->getMachineMemOperand( 474 Ptr, Flags, (MRI->getType(Vals[i]).getSizeInBits() + 7) / 8, 475 MinAlign(BaseAlign, Offsets[i] / 8), AAMDNodes(), nullptr, 476 SI.getSyncScopeID(), SI.getOrdering()); 477 MIRBuilder.buildStore(Vals[i], Addr, *MMO); 478 } 479 return true; 480 } 481 482 static uint64_t getOffsetFromIndices(const User &U, const DataLayout &DL) { 483 const Value *Src = U.getOperand(0); 484 Type *Int32Ty = Type::getInt32Ty(U.getContext()); 485 486 // getIndexedOffsetInType is designed for GEPs, so the first index is the 487 // usual array element rather than looking into the actual aggregate. 488 SmallVector<Value *, 1> Indices; 489 Indices.push_back(ConstantInt::get(Int32Ty, 0)); 490 491 if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(&U)) { 492 for (auto Idx : EVI->indices()) 493 Indices.push_back(ConstantInt::get(Int32Ty, Idx)); 494 } else if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(&U)) { 495 for (auto Idx : IVI->indices()) 496 Indices.push_back(ConstantInt::get(Int32Ty, Idx)); 497 } else { 498 for (unsigned i = 1; i < U.getNumOperands(); ++i) 499 Indices.push_back(U.getOperand(i)); 500 } 501 502 return 8 * static_cast<uint64_t>( 503 DL.getIndexedOffsetInType(Src->getType(), Indices)); 504 } 505 506 bool IRTranslator::translateExtractValue(const User &U, 507 MachineIRBuilder &MIRBuilder) { 508 const Value *Src = U.getOperand(0); 509 uint64_t Offset = getOffsetFromIndices(U, *DL); 510 ArrayRef<unsigned> SrcRegs = getOrCreateVRegs(*Src); 511 ArrayRef<uint64_t> Offsets = *VMap.getOffsets(*Src); 512 unsigned Idx = std::lower_bound(Offsets.begin(), Offsets.end(), Offset) - 513 Offsets.begin(); 514 auto &DstRegs = allocateVRegs(U); 515 516 for (unsigned i = 0; i < DstRegs.size(); ++i) 517 DstRegs[i] = SrcRegs[Idx++]; 518 519 return true; 520 } 521 522 bool IRTranslator::translateInsertValue(const User &U, 523 MachineIRBuilder &MIRBuilder) { 524 const Value *Src = U.getOperand(0); 525 uint64_t Offset = getOffsetFromIndices(U, *DL); 526 auto &DstRegs = allocateVRegs(U); 527 ArrayRef<uint64_t> DstOffsets = *VMap.getOffsets(U); 528 ArrayRef<unsigned> SrcRegs = getOrCreateVRegs(*Src); 529 ArrayRef<unsigned> InsertedRegs = getOrCreateVRegs(*U.getOperand(1)); 530 auto InsertedIt = InsertedRegs.begin(); 531 532 for (unsigned i = 0; i < DstRegs.size(); ++i) { 533 if (DstOffsets[i] >= Offset && InsertedIt != InsertedRegs.end()) 534 DstRegs[i] = *InsertedIt++; 535 else 536 DstRegs[i] = SrcRegs[i]; 537 } 538 539 return true; 540 } 541 542 bool IRTranslator::translateSelect(const User &U, 543 MachineIRBuilder &MIRBuilder) { 544 unsigned Tst = getOrCreateVReg(*U.getOperand(0)); 545 ArrayRef<unsigned> ResRegs = getOrCreateVRegs(U); 546 ArrayRef<unsigned> Op0Regs = getOrCreateVRegs(*U.getOperand(1)); 547 ArrayRef<unsigned> Op1Regs = getOrCreateVRegs(*U.getOperand(2)); 548 549 for (unsigned i = 0; i < ResRegs.size(); ++i) 550 MIRBuilder.buildSelect(ResRegs[i], Tst, Op0Regs[i], Op1Regs[i]); 551 552 return true; 553 } 554 555 bool IRTranslator::translateBitCast(const User &U, 556 MachineIRBuilder &MIRBuilder) { 557 // If we're bitcasting to the source type, we can reuse the source vreg. 558 if (getLLTForType(*U.getOperand(0)->getType(), *DL) == 559 getLLTForType(*U.getType(), *DL)) { 560 unsigned SrcReg = getOrCreateVReg(*U.getOperand(0)); 561 auto &Regs = *VMap.getVRegs(U); 562 // If we already assigned a vreg for this bitcast, we can't change that. 563 // Emit a copy to satisfy the users we already emitted. 564 if (!Regs.empty()) 565 MIRBuilder.buildCopy(Regs[0], SrcReg); 566 else { 567 Regs.push_back(SrcReg); 568 VMap.getOffsets(U)->push_back(0); 569 } 570 return true; 571 } 572 return translateCast(TargetOpcode::G_BITCAST, U, MIRBuilder); 573 } 574 575 bool IRTranslator::translateCast(unsigned Opcode, const User &U, 576 MachineIRBuilder &MIRBuilder) { 577 unsigned Op = getOrCreateVReg(*U.getOperand(0)); 578 unsigned Res = getOrCreateVReg(U); 579 MIRBuilder.buildInstr(Opcode).addDef(Res).addUse(Op); 580 return true; 581 } 582 583 bool IRTranslator::translateGetElementPtr(const User &U, 584 MachineIRBuilder &MIRBuilder) { 585 // FIXME: support vector GEPs. 586 if (U.getType()->isVectorTy()) 587 return false; 588 589 Value &Op0 = *U.getOperand(0); 590 unsigned BaseReg = getOrCreateVReg(Op0); 591 Type *PtrIRTy = Op0.getType(); 592 LLT PtrTy = getLLTForType(*PtrIRTy, *DL); 593 Type *OffsetIRTy = DL->getIntPtrType(PtrIRTy); 594 LLT OffsetTy = getLLTForType(*OffsetIRTy, *DL); 595 596 int64_t Offset = 0; 597 for (gep_type_iterator GTI = gep_type_begin(&U), E = gep_type_end(&U); 598 GTI != E; ++GTI) { 599 const Value *Idx = GTI.getOperand(); 600 if (StructType *StTy = GTI.getStructTypeOrNull()) { 601 unsigned Field = cast<Constant>(Idx)->getUniqueInteger().getZExtValue(); 602 Offset += DL->getStructLayout(StTy)->getElementOffset(Field); 603 continue; 604 } else { 605 uint64_t ElementSize = DL->getTypeAllocSize(GTI.getIndexedType()); 606 607 // If this is a scalar constant or a splat vector of constants, 608 // handle it quickly. 609 if (const auto *CI = dyn_cast<ConstantInt>(Idx)) { 610 Offset += ElementSize * CI->getSExtValue(); 611 continue; 612 } 613 614 if (Offset != 0) { 615 unsigned NewBaseReg = MRI->createGenericVirtualRegister(PtrTy); 616 unsigned OffsetReg = 617 getOrCreateVReg(*ConstantInt::get(OffsetIRTy, Offset)); 618 MIRBuilder.buildGEP(NewBaseReg, BaseReg, OffsetReg); 619 620 BaseReg = NewBaseReg; 621 Offset = 0; 622 } 623 624 unsigned IdxReg = getOrCreateVReg(*Idx); 625 if (MRI->getType(IdxReg) != OffsetTy) { 626 unsigned NewIdxReg = MRI->createGenericVirtualRegister(OffsetTy); 627 MIRBuilder.buildSExtOrTrunc(NewIdxReg, IdxReg); 628 IdxReg = NewIdxReg; 629 } 630 631 // N = N + Idx * ElementSize; 632 // Avoid doing it for ElementSize of 1. 633 unsigned GepOffsetReg; 634 if (ElementSize != 1) { 635 unsigned ElementSizeReg = 636 getOrCreateVReg(*ConstantInt::get(OffsetIRTy, ElementSize)); 637 638 GepOffsetReg = MRI->createGenericVirtualRegister(OffsetTy); 639 MIRBuilder.buildMul(GepOffsetReg, ElementSizeReg, IdxReg); 640 } else 641 GepOffsetReg = IdxReg; 642 643 unsigned NewBaseReg = MRI->createGenericVirtualRegister(PtrTy); 644 MIRBuilder.buildGEP(NewBaseReg, BaseReg, GepOffsetReg); 645 BaseReg = NewBaseReg; 646 } 647 } 648 649 if (Offset != 0) { 650 unsigned OffsetReg = getOrCreateVReg(*ConstantInt::get(OffsetIRTy, Offset)); 651 MIRBuilder.buildGEP(getOrCreateVReg(U), BaseReg, OffsetReg); 652 return true; 653 } 654 655 MIRBuilder.buildCopy(getOrCreateVReg(U), BaseReg); 656 return true; 657 } 658 659 bool IRTranslator::translateMemfunc(const CallInst &CI, 660 MachineIRBuilder &MIRBuilder, 661 unsigned ID) { 662 LLT SizeTy = getLLTForType(*CI.getArgOperand(2)->getType(), *DL); 663 Type *DstTy = CI.getArgOperand(0)->getType(); 664 if (cast<PointerType>(DstTy)->getAddressSpace() != 0 || 665 SizeTy.getSizeInBits() != DL->getPointerSizeInBits(0)) 666 return false; 667 668 SmallVector<CallLowering::ArgInfo, 8> Args; 669 for (int i = 0; i < 3; ++i) { 670 const auto &Arg = CI.getArgOperand(i); 671 Args.emplace_back(getOrCreateVReg(*Arg), Arg->getType()); 672 } 673 674 const char *Callee; 675 switch (ID) { 676 case Intrinsic::memmove: 677 case Intrinsic::memcpy: { 678 Type *SrcTy = CI.getArgOperand(1)->getType(); 679 if(cast<PointerType>(SrcTy)->getAddressSpace() != 0) 680 return false; 681 Callee = ID == Intrinsic::memcpy ? "memcpy" : "memmove"; 682 break; 683 } 684 case Intrinsic::memset: 685 Callee = "memset"; 686 break; 687 default: 688 return false; 689 } 690 691 return CLI->lowerCall(MIRBuilder, CI.getCallingConv(), 692 MachineOperand::CreateES(Callee), 693 CallLowering::ArgInfo(0, CI.getType()), Args); 694 } 695 696 void IRTranslator::getStackGuard(unsigned DstReg, 697 MachineIRBuilder &MIRBuilder) { 698 const TargetRegisterInfo *TRI = MF->getSubtarget().getRegisterInfo(); 699 MRI->setRegClass(DstReg, TRI->getPointerRegClass(*MF)); 700 auto MIB = MIRBuilder.buildInstr(TargetOpcode::LOAD_STACK_GUARD); 701 MIB.addDef(DstReg); 702 703 auto &TLI = *MF->getSubtarget().getTargetLowering(); 704 Value *Global = TLI.getSDagStackGuard(*MF->getFunction().getParent()); 705 if (!Global) 706 return; 707 708 MachinePointerInfo MPInfo(Global); 709 MachineInstr::mmo_iterator MemRefs = MF->allocateMemRefsArray(1); 710 auto Flags = MachineMemOperand::MOLoad | MachineMemOperand::MOInvariant | 711 MachineMemOperand::MODereferenceable; 712 *MemRefs = 713 MF->getMachineMemOperand(MPInfo, Flags, DL->getPointerSizeInBits() / 8, 714 DL->getPointerABIAlignment(0)); 715 MIB.setMemRefs(MemRefs, MemRefs + 1); 716 } 717 718 bool IRTranslator::translateOverflowIntrinsic(const CallInst &CI, unsigned Op, 719 MachineIRBuilder &MIRBuilder) { 720 ArrayRef<unsigned> ResRegs = getOrCreateVRegs(CI); 721 auto MIB = MIRBuilder.buildInstr(Op) 722 .addDef(ResRegs[0]) 723 .addDef(ResRegs[1]) 724 .addUse(getOrCreateVReg(*CI.getOperand(0))) 725 .addUse(getOrCreateVReg(*CI.getOperand(1))); 726 727 if (Op == TargetOpcode::G_UADDE || Op == TargetOpcode::G_USUBE) { 728 unsigned Zero = getOrCreateVReg( 729 *Constant::getNullValue(Type::getInt1Ty(CI.getContext()))); 730 MIB.addUse(Zero); 731 } 732 733 return true; 734 } 735 736 bool IRTranslator::translateKnownIntrinsic(const CallInst &CI, Intrinsic::ID ID, 737 MachineIRBuilder &MIRBuilder) { 738 switch (ID) { 739 default: 740 break; 741 case Intrinsic::lifetime_start: 742 case Intrinsic::lifetime_end: 743 // Stack coloring is not enabled in O0 (which we care about now) so we can 744 // drop these. Make sure someone notices when we start compiling at higher 745 // opts though. 746 if (MF->getTarget().getOptLevel() != CodeGenOpt::None) 747 return false; 748 return true; 749 case Intrinsic::dbg_declare: { 750 const DbgDeclareInst &DI = cast<DbgDeclareInst>(CI); 751 assert(DI.getVariable() && "Missing variable"); 752 753 const Value *Address = DI.getAddress(); 754 if (!Address || isa<UndefValue>(Address)) { 755 LLVM_DEBUG(dbgs() << "Dropping debug info for " << DI << "\n"); 756 return true; 757 } 758 759 assert(DI.getVariable()->isValidLocationForIntrinsic( 760 MIRBuilder.getDebugLoc()) && 761 "Expected inlined-at fields to agree"); 762 auto AI = dyn_cast<AllocaInst>(Address); 763 if (AI && AI->isStaticAlloca()) { 764 // Static allocas are tracked at the MF level, no need for DBG_VALUE 765 // instructions (in fact, they get ignored if they *do* exist). 766 MF->setVariableDbgInfo(DI.getVariable(), DI.getExpression(), 767 getOrCreateFrameIndex(*AI), DI.getDebugLoc()); 768 } else 769 MIRBuilder.buildDirectDbgValue(getOrCreateVReg(*Address), 770 DI.getVariable(), DI.getExpression()); 771 return true; 772 } 773 case Intrinsic::vaend: 774 // No target I know of cares about va_end. Certainly no in-tree target 775 // does. Simplest intrinsic ever! 776 return true; 777 case Intrinsic::vastart: { 778 auto &TLI = *MF->getSubtarget().getTargetLowering(); 779 Value *Ptr = CI.getArgOperand(0); 780 unsigned ListSize = TLI.getVaListSizeInBits(*DL) / 8; 781 782 MIRBuilder.buildInstr(TargetOpcode::G_VASTART) 783 .addUse(getOrCreateVReg(*Ptr)) 784 .addMemOperand(MF->getMachineMemOperand( 785 MachinePointerInfo(Ptr), MachineMemOperand::MOStore, ListSize, 0)); 786 return true; 787 } 788 case Intrinsic::dbg_value: { 789 // This form of DBG_VALUE is target-independent. 790 const DbgValueInst &DI = cast<DbgValueInst>(CI); 791 const Value *V = DI.getValue(); 792 assert(DI.getVariable()->isValidLocationForIntrinsic( 793 MIRBuilder.getDebugLoc()) && 794 "Expected inlined-at fields to agree"); 795 if (!V) { 796 // Currently the optimizer can produce this; insert an undef to 797 // help debugging. Probably the optimizer should not do this. 798 MIRBuilder.buildIndirectDbgValue(0, DI.getVariable(), DI.getExpression()); 799 } else if (const auto *CI = dyn_cast<Constant>(V)) { 800 MIRBuilder.buildConstDbgValue(*CI, DI.getVariable(), DI.getExpression()); 801 } else { 802 unsigned Reg = getOrCreateVReg(*V); 803 // FIXME: This does not handle register-indirect values at offset 0. The 804 // direct/indirect thing shouldn't really be handled by something as 805 // implicit as reg+noreg vs reg+imm in the first palce, but it seems 806 // pretty baked in right now. 807 MIRBuilder.buildDirectDbgValue(Reg, DI.getVariable(), DI.getExpression()); 808 } 809 return true; 810 } 811 case Intrinsic::uadd_with_overflow: 812 return translateOverflowIntrinsic(CI, TargetOpcode::G_UADDE, MIRBuilder); 813 case Intrinsic::sadd_with_overflow: 814 return translateOverflowIntrinsic(CI, TargetOpcode::G_SADDO, MIRBuilder); 815 case Intrinsic::usub_with_overflow: 816 return translateOverflowIntrinsic(CI, TargetOpcode::G_USUBE, MIRBuilder); 817 case Intrinsic::ssub_with_overflow: 818 return translateOverflowIntrinsic(CI, TargetOpcode::G_SSUBO, MIRBuilder); 819 case Intrinsic::umul_with_overflow: 820 return translateOverflowIntrinsic(CI, TargetOpcode::G_UMULO, MIRBuilder); 821 case Intrinsic::smul_with_overflow: 822 return translateOverflowIntrinsic(CI, TargetOpcode::G_SMULO, MIRBuilder); 823 case Intrinsic::pow: 824 MIRBuilder.buildInstr(TargetOpcode::G_FPOW) 825 .addDef(getOrCreateVReg(CI)) 826 .addUse(getOrCreateVReg(*CI.getArgOperand(0))) 827 .addUse(getOrCreateVReg(*CI.getArgOperand(1))); 828 return true; 829 case Intrinsic::exp: 830 MIRBuilder.buildInstr(TargetOpcode::G_FEXP) 831 .addDef(getOrCreateVReg(CI)) 832 .addUse(getOrCreateVReg(*CI.getArgOperand(0))); 833 return true; 834 case Intrinsic::exp2: 835 MIRBuilder.buildInstr(TargetOpcode::G_FEXP2) 836 .addDef(getOrCreateVReg(CI)) 837 .addUse(getOrCreateVReg(*CI.getArgOperand(0))); 838 return true; 839 case Intrinsic::log: 840 MIRBuilder.buildInstr(TargetOpcode::G_FLOG) 841 .addDef(getOrCreateVReg(CI)) 842 .addUse(getOrCreateVReg(*CI.getArgOperand(0))); 843 return true; 844 case Intrinsic::log2: 845 MIRBuilder.buildInstr(TargetOpcode::G_FLOG2) 846 .addDef(getOrCreateVReg(CI)) 847 .addUse(getOrCreateVReg(*CI.getArgOperand(0))); 848 return true; 849 case Intrinsic::fabs: 850 MIRBuilder.buildInstr(TargetOpcode::G_FABS) 851 .addDef(getOrCreateVReg(CI)) 852 .addUse(getOrCreateVReg(*CI.getArgOperand(0))); 853 return true; 854 case Intrinsic::fma: 855 MIRBuilder.buildInstr(TargetOpcode::G_FMA) 856 .addDef(getOrCreateVReg(CI)) 857 .addUse(getOrCreateVReg(*CI.getArgOperand(0))) 858 .addUse(getOrCreateVReg(*CI.getArgOperand(1))) 859 .addUse(getOrCreateVReg(*CI.getArgOperand(2))); 860 return true; 861 case Intrinsic::fmuladd: { 862 const TargetMachine &TM = MF->getTarget(); 863 const TargetLowering &TLI = *MF->getSubtarget().getTargetLowering(); 864 unsigned Dst = getOrCreateVReg(CI); 865 unsigned Op0 = getOrCreateVReg(*CI.getArgOperand(0)); 866 unsigned Op1 = getOrCreateVReg(*CI.getArgOperand(1)); 867 unsigned Op2 = getOrCreateVReg(*CI.getArgOperand(2)); 868 if (TM.Options.AllowFPOpFusion != FPOpFusion::Strict && 869 TLI.isFMAFasterThanFMulAndFAdd(TLI.getValueType(*DL, CI.getType()))) { 870 // TODO: Revisit this to see if we should move this part of the 871 // lowering to the combiner. 872 MIRBuilder.buildInstr(TargetOpcode::G_FMA, Dst, Op0, Op1, Op2); 873 } else { 874 LLT Ty = getLLTForType(*CI.getType(), *DL); 875 auto FMul = MIRBuilder.buildInstr(TargetOpcode::G_FMUL, Ty, Op0, Op1); 876 MIRBuilder.buildInstr(TargetOpcode::G_FADD, Dst, FMul, Op2); 877 } 878 return true; 879 } 880 case Intrinsic::memcpy: 881 case Intrinsic::memmove: 882 case Intrinsic::memset: 883 return translateMemfunc(CI, MIRBuilder, ID); 884 case Intrinsic::eh_typeid_for: { 885 GlobalValue *GV = ExtractTypeInfo(CI.getArgOperand(0)); 886 unsigned Reg = getOrCreateVReg(CI); 887 unsigned TypeID = MF->getTypeIDFor(GV); 888 MIRBuilder.buildConstant(Reg, TypeID); 889 return true; 890 } 891 case Intrinsic::objectsize: { 892 // If we don't know by now, we're never going to know. 893 const ConstantInt *Min = cast<ConstantInt>(CI.getArgOperand(1)); 894 895 MIRBuilder.buildConstant(getOrCreateVReg(CI), Min->isZero() ? -1ULL : 0); 896 return true; 897 } 898 case Intrinsic::stackguard: 899 getStackGuard(getOrCreateVReg(CI), MIRBuilder); 900 return true; 901 case Intrinsic::stackprotector: { 902 LLT PtrTy = getLLTForType(*CI.getArgOperand(0)->getType(), *DL); 903 unsigned GuardVal = MRI->createGenericVirtualRegister(PtrTy); 904 getStackGuard(GuardVal, MIRBuilder); 905 906 AllocaInst *Slot = cast<AllocaInst>(CI.getArgOperand(1)); 907 MIRBuilder.buildStore( 908 GuardVal, getOrCreateVReg(*Slot), 909 *MF->getMachineMemOperand( 910 MachinePointerInfo::getFixedStack(*MF, 911 getOrCreateFrameIndex(*Slot)), 912 MachineMemOperand::MOStore | MachineMemOperand::MOVolatile, 913 PtrTy.getSizeInBits() / 8, 8)); 914 return true; 915 } 916 case Intrinsic::cttz: 917 case Intrinsic::ctlz: { 918 ConstantInt *Cst = cast<ConstantInt>(CI.getArgOperand(1)); 919 bool isTrailing = ID == Intrinsic::cttz; 920 unsigned Opcode = isTrailing 921 ? Cst->isZero() ? TargetOpcode::G_CTTZ 922 : TargetOpcode::G_CTTZ_ZERO_UNDEF 923 : Cst->isZero() ? TargetOpcode::G_CTLZ 924 : TargetOpcode::G_CTLZ_ZERO_UNDEF; 925 MIRBuilder.buildInstr(Opcode) 926 .addDef(getOrCreateVReg(CI)) 927 .addUse(getOrCreateVReg(*CI.getArgOperand(0))); 928 return true; 929 } 930 case Intrinsic::ctpop: { 931 MIRBuilder.buildInstr(TargetOpcode::G_CTPOP) 932 .addDef(getOrCreateVReg(CI)) 933 .addUse(getOrCreateVReg(*CI.getArgOperand(0))); 934 return true; 935 } 936 } 937 return false; 938 } 939 940 bool IRTranslator::translateInlineAsm(const CallInst &CI, 941 MachineIRBuilder &MIRBuilder) { 942 const InlineAsm &IA = cast<InlineAsm>(*CI.getCalledValue()); 943 if (!IA.getConstraintString().empty()) 944 return false; 945 946 unsigned ExtraInfo = 0; 947 if (IA.hasSideEffects()) 948 ExtraInfo |= InlineAsm::Extra_HasSideEffects; 949 if (IA.getDialect() == InlineAsm::AD_Intel) 950 ExtraInfo |= InlineAsm::Extra_AsmDialect; 951 952 MIRBuilder.buildInstr(TargetOpcode::INLINEASM) 953 .addExternalSymbol(IA.getAsmString().c_str()) 954 .addImm(ExtraInfo); 955 956 return true; 957 } 958 959 unsigned IRTranslator::packRegs(const Value &V, 960 MachineIRBuilder &MIRBuilder) { 961 ArrayRef<unsigned> Regs = getOrCreateVRegs(V); 962 ArrayRef<uint64_t> Offsets = *VMap.getOffsets(V); 963 LLT BigTy = getLLTForType(*V.getType(), *DL); 964 965 if (Regs.size() == 1) 966 return Regs[0]; 967 968 unsigned Dst = MRI->createGenericVirtualRegister(BigTy); 969 MIRBuilder.buildUndef(Dst); 970 for (unsigned i = 0; i < Regs.size(); ++i) { 971 unsigned NewDst = MRI->createGenericVirtualRegister(BigTy); 972 MIRBuilder.buildInsert(NewDst, Dst, Regs[i], Offsets[i]); 973 Dst = NewDst; 974 } 975 return Dst; 976 } 977 978 void IRTranslator::unpackRegs(const Value &V, unsigned Src, 979 MachineIRBuilder &MIRBuilder) { 980 ArrayRef<unsigned> Regs = getOrCreateVRegs(V); 981 ArrayRef<uint64_t> Offsets = *VMap.getOffsets(V); 982 983 for (unsigned i = 0; i < Regs.size(); ++i) 984 MIRBuilder.buildExtract(Regs[i], Src, Offsets[i]); 985 } 986 987 bool IRTranslator::translateCall(const User &U, MachineIRBuilder &MIRBuilder) { 988 const CallInst &CI = cast<CallInst>(U); 989 auto TII = MF->getTarget().getIntrinsicInfo(); 990 const Function *F = CI.getCalledFunction(); 991 992 // FIXME: support Windows dllimport function calls. 993 if (F && F->hasDLLImportStorageClass()) 994 return false; 995 996 if (CI.isInlineAsm()) 997 return translateInlineAsm(CI, MIRBuilder); 998 999 Intrinsic::ID ID = Intrinsic::not_intrinsic; 1000 if (F && F->isIntrinsic()) { 1001 ID = F->getIntrinsicID(); 1002 if (TII && ID == Intrinsic::not_intrinsic) 1003 ID = static_cast<Intrinsic::ID>(TII->getIntrinsicID(F)); 1004 } 1005 1006 bool IsSplitType = valueIsSplit(CI); 1007 if (!F || !F->isIntrinsic() || ID == Intrinsic::not_intrinsic) { 1008 unsigned Res = IsSplitType ? MRI->createGenericVirtualRegister( 1009 getLLTForType(*CI.getType(), *DL)) 1010 : getOrCreateVReg(CI); 1011 1012 SmallVector<unsigned, 8> Args; 1013 for (auto &Arg: CI.arg_operands()) 1014 Args.push_back(packRegs(*Arg, MIRBuilder)); 1015 1016 MF->getFrameInfo().setHasCalls(true); 1017 bool Success = CLI->lowerCall(MIRBuilder, &CI, Res, Args, [&]() { 1018 return getOrCreateVReg(*CI.getCalledValue()); 1019 }); 1020 1021 if (IsSplitType) 1022 unpackRegs(CI, Res, MIRBuilder); 1023 return Success; 1024 } 1025 1026 assert(ID != Intrinsic::not_intrinsic && "unknown intrinsic"); 1027 1028 if (translateKnownIntrinsic(CI, ID, MIRBuilder)) 1029 return true; 1030 1031 unsigned Res = 0; 1032 if (!CI.getType()->isVoidTy()) { 1033 if (IsSplitType) 1034 Res = 1035 MRI->createGenericVirtualRegister(getLLTForType(*CI.getType(), *DL)); 1036 else 1037 Res = getOrCreateVReg(CI); 1038 } 1039 MachineInstrBuilder MIB = 1040 MIRBuilder.buildIntrinsic(ID, Res, !CI.doesNotAccessMemory()); 1041 1042 for (auto &Arg : CI.arg_operands()) { 1043 // Some intrinsics take metadata parameters. Reject them. 1044 if (isa<MetadataAsValue>(Arg)) 1045 return false; 1046 MIB.addUse(packRegs(*Arg, MIRBuilder)); 1047 } 1048 1049 if (IsSplitType) 1050 unpackRegs(CI, Res, MIRBuilder); 1051 1052 // Add a MachineMemOperand if it is a target mem intrinsic. 1053 const TargetLowering &TLI = *MF->getSubtarget().getTargetLowering(); 1054 TargetLowering::IntrinsicInfo Info; 1055 // TODO: Add a GlobalISel version of getTgtMemIntrinsic. 1056 if (TLI.getTgtMemIntrinsic(Info, CI, *MF, ID)) { 1057 uint64_t Size = Info.memVT.getStoreSize(); 1058 MIB.addMemOperand(MF->getMachineMemOperand(MachinePointerInfo(Info.ptrVal), 1059 Info.flags, Size, Info.align)); 1060 } 1061 1062 return true; 1063 } 1064 1065 bool IRTranslator::translateInvoke(const User &U, 1066 MachineIRBuilder &MIRBuilder) { 1067 const InvokeInst &I = cast<InvokeInst>(U); 1068 MCContext &Context = MF->getContext(); 1069 1070 const BasicBlock *ReturnBB = I.getSuccessor(0); 1071 const BasicBlock *EHPadBB = I.getSuccessor(1); 1072 1073 const Value *Callee = I.getCalledValue(); 1074 const Function *Fn = dyn_cast<Function>(Callee); 1075 if (isa<InlineAsm>(Callee)) 1076 return false; 1077 1078 // FIXME: support invoking patchpoint and statepoint intrinsics. 1079 if (Fn && Fn->isIntrinsic()) 1080 return false; 1081 1082 // FIXME: support whatever these are. 1083 if (I.countOperandBundlesOfType(LLVMContext::OB_deopt)) 1084 return false; 1085 1086 // FIXME: support Windows exception handling. 1087 if (!isa<LandingPadInst>(EHPadBB->front())) 1088 return false; 1089 1090 // Emit the actual call, bracketed by EH_LABELs so that the MF knows about 1091 // the region covered by the try. 1092 MCSymbol *BeginSymbol = Context.createTempSymbol(); 1093 MIRBuilder.buildInstr(TargetOpcode::EH_LABEL).addSym(BeginSymbol); 1094 1095 unsigned Res = 1096 MRI->createGenericVirtualRegister(getLLTForType(*I.getType(), *DL)); 1097 SmallVector<unsigned, 8> Args; 1098 for (auto &Arg: I.arg_operands()) 1099 Args.push_back(packRegs(*Arg, MIRBuilder)); 1100 1101 if (!CLI->lowerCall(MIRBuilder, &I, Res, Args, 1102 [&]() { return getOrCreateVReg(*I.getCalledValue()); })) 1103 return false; 1104 1105 unpackRegs(I, Res, MIRBuilder); 1106 1107 MCSymbol *EndSymbol = Context.createTempSymbol(); 1108 MIRBuilder.buildInstr(TargetOpcode::EH_LABEL).addSym(EndSymbol); 1109 1110 // FIXME: track probabilities. 1111 MachineBasicBlock &EHPadMBB = getMBB(*EHPadBB), 1112 &ReturnMBB = getMBB(*ReturnBB); 1113 MF->addInvoke(&EHPadMBB, BeginSymbol, EndSymbol); 1114 MIRBuilder.getMBB().addSuccessor(&ReturnMBB); 1115 MIRBuilder.getMBB().addSuccessor(&EHPadMBB); 1116 MIRBuilder.buildBr(ReturnMBB); 1117 1118 return true; 1119 } 1120 1121 bool IRTranslator::translateLandingPad(const User &U, 1122 MachineIRBuilder &MIRBuilder) { 1123 const LandingPadInst &LP = cast<LandingPadInst>(U); 1124 1125 MachineBasicBlock &MBB = MIRBuilder.getMBB(); 1126 addLandingPadInfo(LP, MBB); 1127 1128 MBB.setIsEHPad(); 1129 1130 // If there aren't registers to copy the values into (e.g., during SjLj 1131 // exceptions), then don't bother. 1132 auto &TLI = *MF->getSubtarget().getTargetLowering(); 1133 const Constant *PersonalityFn = MF->getFunction().getPersonalityFn(); 1134 if (TLI.getExceptionPointerRegister(PersonalityFn) == 0 && 1135 TLI.getExceptionSelectorRegister(PersonalityFn) == 0) 1136 return true; 1137 1138 // If landingpad's return type is token type, we don't create DAG nodes 1139 // for its exception pointer and selector value. The extraction of exception 1140 // pointer or selector value from token type landingpads is not currently 1141 // supported. 1142 if (LP.getType()->isTokenTy()) 1143 return true; 1144 1145 // Add a label to mark the beginning of the landing pad. Deletion of the 1146 // landing pad can thus be detected via the MachineModuleInfo. 1147 MIRBuilder.buildInstr(TargetOpcode::EH_LABEL) 1148 .addSym(MF->addLandingPad(&MBB)); 1149 1150 LLT Ty = getLLTForType(*LP.getType(), *DL); 1151 unsigned Undef = MRI->createGenericVirtualRegister(Ty); 1152 MIRBuilder.buildUndef(Undef); 1153 1154 SmallVector<LLT, 2> Tys; 1155 for (Type *Ty : cast<StructType>(LP.getType())->elements()) 1156 Tys.push_back(getLLTForType(*Ty, *DL)); 1157 assert(Tys.size() == 2 && "Only two-valued landingpads are supported"); 1158 1159 // Mark exception register as live in. 1160 unsigned ExceptionReg = TLI.getExceptionPointerRegister(PersonalityFn); 1161 if (!ExceptionReg) 1162 return false; 1163 1164 MBB.addLiveIn(ExceptionReg); 1165 ArrayRef<unsigned> ResRegs = getOrCreateVRegs(LP); 1166 MIRBuilder.buildCopy(ResRegs[0], ExceptionReg); 1167 1168 unsigned SelectorReg = TLI.getExceptionSelectorRegister(PersonalityFn); 1169 if (!SelectorReg) 1170 return false; 1171 1172 MBB.addLiveIn(SelectorReg); 1173 unsigned PtrVReg = MRI->createGenericVirtualRegister(Tys[0]); 1174 MIRBuilder.buildCopy(PtrVReg, SelectorReg); 1175 MIRBuilder.buildCast(ResRegs[1], PtrVReg); 1176 1177 return true; 1178 } 1179 1180 bool IRTranslator::translateAlloca(const User &U, 1181 MachineIRBuilder &MIRBuilder) { 1182 auto &AI = cast<AllocaInst>(U); 1183 1184 if (AI.isSwiftError()) 1185 return false; 1186 1187 if (AI.isStaticAlloca()) { 1188 unsigned Res = getOrCreateVReg(AI); 1189 int FI = getOrCreateFrameIndex(AI); 1190 MIRBuilder.buildFrameIndex(Res, FI); 1191 return true; 1192 } 1193 1194 // FIXME: support stack probing for Windows. 1195 if (MF->getTarget().getTargetTriple().isOSWindows()) 1196 return false; 1197 1198 // Now we're in the harder dynamic case. 1199 Type *Ty = AI.getAllocatedType(); 1200 unsigned Align = 1201 std::max((unsigned)DL->getPrefTypeAlignment(Ty), AI.getAlignment()); 1202 1203 unsigned NumElts = getOrCreateVReg(*AI.getArraySize()); 1204 1205 Type *IntPtrIRTy = DL->getIntPtrType(AI.getType()); 1206 LLT IntPtrTy = getLLTForType(*IntPtrIRTy, *DL); 1207 if (MRI->getType(NumElts) != IntPtrTy) { 1208 unsigned ExtElts = MRI->createGenericVirtualRegister(IntPtrTy); 1209 MIRBuilder.buildZExtOrTrunc(ExtElts, NumElts); 1210 NumElts = ExtElts; 1211 } 1212 1213 unsigned AllocSize = MRI->createGenericVirtualRegister(IntPtrTy); 1214 unsigned TySize = 1215 getOrCreateVReg(*ConstantInt::get(IntPtrIRTy, -DL->getTypeAllocSize(Ty))); 1216 MIRBuilder.buildMul(AllocSize, NumElts, TySize); 1217 1218 LLT PtrTy = getLLTForType(*AI.getType(), *DL); 1219 auto &TLI = *MF->getSubtarget().getTargetLowering(); 1220 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore(); 1221 1222 unsigned SPTmp = MRI->createGenericVirtualRegister(PtrTy); 1223 MIRBuilder.buildCopy(SPTmp, SPReg); 1224 1225 unsigned AllocTmp = MRI->createGenericVirtualRegister(PtrTy); 1226 MIRBuilder.buildGEP(AllocTmp, SPTmp, AllocSize); 1227 1228 // Handle alignment. We have to realign if the allocation granule was smaller 1229 // than stack alignment, or the specific alloca requires more than stack 1230 // alignment. 1231 unsigned StackAlign = 1232 MF->getSubtarget().getFrameLowering()->getStackAlignment(); 1233 Align = std::max(Align, StackAlign); 1234 if (Align > StackAlign || DL->getTypeAllocSize(Ty) % StackAlign != 0) { 1235 // Round the size of the allocation up to the stack alignment size 1236 // by add SA-1 to the size. This doesn't overflow because we're computing 1237 // an address inside an alloca. 1238 unsigned AlignedAlloc = MRI->createGenericVirtualRegister(PtrTy); 1239 MIRBuilder.buildPtrMask(AlignedAlloc, AllocTmp, Log2_32(Align)); 1240 AllocTmp = AlignedAlloc; 1241 } 1242 1243 MIRBuilder.buildCopy(SPReg, AllocTmp); 1244 MIRBuilder.buildCopy(getOrCreateVReg(AI), AllocTmp); 1245 1246 MF->getFrameInfo().CreateVariableSizedObject(Align ? Align : 1, &AI); 1247 assert(MF->getFrameInfo().hasVarSizedObjects()); 1248 return true; 1249 } 1250 1251 bool IRTranslator::translateVAArg(const User &U, MachineIRBuilder &MIRBuilder) { 1252 // FIXME: We may need more info about the type. Because of how LLT works, 1253 // we're completely discarding the i64/double distinction here (amongst 1254 // others). Fortunately the ABIs I know of where that matters don't use va_arg 1255 // anyway but that's not guaranteed. 1256 MIRBuilder.buildInstr(TargetOpcode::G_VAARG) 1257 .addDef(getOrCreateVReg(U)) 1258 .addUse(getOrCreateVReg(*U.getOperand(0))) 1259 .addImm(DL->getABITypeAlignment(U.getType())); 1260 return true; 1261 } 1262 1263 bool IRTranslator::translateInsertElement(const User &U, 1264 MachineIRBuilder &MIRBuilder) { 1265 // If it is a <1 x Ty> vector, use the scalar as it is 1266 // not a legal vector type in LLT. 1267 if (U.getType()->getVectorNumElements() == 1) { 1268 unsigned Elt = getOrCreateVReg(*U.getOperand(1)); 1269 auto &Regs = *VMap.getVRegs(U); 1270 if (Regs.empty()) { 1271 Regs.push_back(Elt); 1272 VMap.getOffsets(U)->push_back(0); 1273 } else { 1274 MIRBuilder.buildCopy(Regs[0], Elt); 1275 } 1276 return true; 1277 } 1278 1279 unsigned Res = getOrCreateVReg(U); 1280 unsigned Val = getOrCreateVReg(*U.getOperand(0)); 1281 unsigned Elt = getOrCreateVReg(*U.getOperand(1)); 1282 unsigned Idx = getOrCreateVReg(*U.getOperand(2)); 1283 MIRBuilder.buildInsertVectorElement(Res, Val, Elt, Idx); 1284 return true; 1285 } 1286 1287 bool IRTranslator::translateExtractElement(const User &U, 1288 MachineIRBuilder &MIRBuilder) { 1289 // If it is a <1 x Ty> vector, use the scalar as it is 1290 // not a legal vector type in LLT. 1291 if (U.getOperand(0)->getType()->getVectorNumElements() == 1) { 1292 unsigned Elt = getOrCreateVReg(*U.getOperand(0)); 1293 auto &Regs = *VMap.getVRegs(U); 1294 if (Regs.empty()) { 1295 Regs.push_back(Elt); 1296 VMap.getOffsets(U)->push_back(0); 1297 } else { 1298 MIRBuilder.buildCopy(Regs[0], Elt); 1299 } 1300 return true; 1301 } 1302 unsigned Res = getOrCreateVReg(U); 1303 unsigned Val = getOrCreateVReg(*U.getOperand(0)); 1304 unsigned Idx = getOrCreateVReg(*U.getOperand(1)); 1305 MIRBuilder.buildExtractVectorElement(Res, Val, Idx); 1306 return true; 1307 } 1308 1309 bool IRTranslator::translateShuffleVector(const User &U, 1310 MachineIRBuilder &MIRBuilder) { 1311 MIRBuilder.buildInstr(TargetOpcode::G_SHUFFLE_VECTOR) 1312 .addDef(getOrCreateVReg(U)) 1313 .addUse(getOrCreateVReg(*U.getOperand(0))) 1314 .addUse(getOrCreateVReg(*U.getOperand(1))) 1315 .addUse(getOrCreateVReg(*U.getOperand(2))); 1316 return true; 1317 } 1318 1319 bool IRTranslator::translatePHI(const User &U, MachineIRBuilder &MIRBuilder) { 1320 const PHINode &PI = cast<PHINode>(U); 1321 1322 SmallVector<MachineInstr *, 4> Insts; 1323 for (auto Reg : getOrCreateVRegs(PI)) { 1324 auto MIB = MIRBuilder.buildInstr(TargetOpcode::G_PHI, Reg); 1325 Insts.push_back(MIB.getInstr()); 1326 } 1327 1328 PendingPHIs.emplace_back(&PI, std::move(Insts)); 1329 return true; 1330 } 1331 1332 bool IRTranslator::translateAtomicCmpXchg(const User &U, 1333 MachineIRBuilder &MIRBuilder) { 1334 const AtomicCmpXchgInst &I = cast<AtomicCmpXchgInst>(U); 1335 1336 if (I.isWeak()) 1337 return false; 1338 1339 auto Flags = I.isVolatile() ? MachineMemOperand::MOVolatile 1340 : MachineMemOperand::MONone; 1341 Flags |= MachineMemOperand::MOLoad | MachineMemOperand::MOStore; 1342 1343 Type *ResType = I.getType(); 1344 Type *ValType = ResType->Type::getStructElementType(0); 1345 1346 auto Res = getOrCreateVRegs(I); 1347 unsigned OldValRes = Res[0]; 1348 unsigned SuccessRes = Res[1]; 1349 unsigned Addr = getOrCreateVReg(*I.getPointerOperand()); 1350 unsigned Cmp = getOrCreateVReg(*I.getCompareOperand()); 1351 unsigned NewVal = getOrCreateVReg(*I.getNewValOperand()); 1352 1353 MIRBuilder.buildAtomicCmpXchgWithSuccess( 1354 OldValRes, SuccessRes, Addr, Cmp, NewVal, 1355 *MF->getMachineMemOperand(MachinePointerInfo(I.getPointerOperand()), 1356 Flags, DL->getTypeStoreSize(ValType), 1357 getMemOpAlignment(I), AAMDNodes(), nullptr, 1358 I.getSyncScopeID(), I.getSuccessOrdering(), 1359 I.getFailureOrdering())); 1360 return true; 1361 } 1362 1363 bool IRTranslator::translateAtomicRMW(const User &U, 1364 MachineIRBuilder &MIRBuilder) { 1365 const AtomicRMWInst &I = cast<AtomicRMWInst>(U); 1366 1367 auto Flags = I.isVolatile() ? MachineMemOperand::MOVolatile 1368 : MachineMemOperand::MONone; 1369 Flags |= MachineMemOperand::MOLoad | MachineMemOperand::MOStore; 1370 1371 Type *ResType = I.getType(); 1372 1373 unsigned Res = getOrCreateVReg(I); 1374 unsigned Addr = getOrCreateVReg(*I.getPointerOperand()); 1375 unsigned Val = getOrCreateVReg(*I.getValOperand()); 1376 1377 unsigned Opcode = 0; 1378 switch (I.getOperation()) { 1379 default: 1380 llvm_unreachable("Unknown atomicrmw op"); 1381 return false; 1382 case AtomicRMWInst::Xchg: 1383 Opcode = TargetOpcode::G_ATOMICRMW_XCHG; 1384 break; 1385 case AtomicRMWInst::Add: 1386 Opcode = TargetOpcode::G_ATOMICRMW_ADD; 1387 break; 1388 case AtomicRMWInst::Sub: 1389 Opcode = TargetOpcode::G_ATOMICRMW_SUB; 1390 break; 1391 case AtomicRMWInst::And: 1392 Opcode = TargetOpcode::G_ATOMICRMW_AND; 1393 break; 1394 case AtomicRMWInst::Nand: 1395 Opcode = TargetOpcode::G_ATOMICRMW_NAND; 1396 break; 1397 case AtomicRMWInst::Or: 1398 Opcode = TargetOpcode::G_ATOMICRMW_OR; 1399 break; 1400 case AtomicRMWInst::Xor: 1401 Opcode = TargetOpcode::G_ATOMICRMW_XOR; 1402 break; 1403 case AtomicRMWInst::Max: 1404 Opcode = TargetOpcode::G_ATOMICRMW_MAX; 1405 break; 1406 case AtomicRMWInst::Min: 1407 Opcode = TargetOpcode::G_ATOMICRMW_MIN; 1408 break; 1409 case AtomicRMWInst::UMax: 1410 Opcode = TargetOpcode::G_ATOMICRMW_UMAX; 1411 break; 1412 case AtomicRMWInst::UMin: 1413 Opcode = TargetOpcode::G_ATOMICRMW_UMIN; 1414 break; 1415 } 1416 1417 MIRBuilder.buildAtomicRMW( 1418 Opcode, Res, Addr, Val, 1419 *MF->getMachineMemOperand(MachinePointerInfo(I.getPointerOperand()), 1420 Flags, DL->getTypeStoreSize(ResType), 1421 getMemOpAlignment(I), AAMDNodes(), nullptr, 1422 I.getSyncScopeID(), I.getOrdering())); 1423 return true; 1424 } 1425 1426 void IRTranslator::finishPendingPhis() { 1427 for (auto &Phi : PendingPHIs) { 1428 const PHINode *PI = Phi.first; 1429 ArrayRef<MachineInstr *> ComponentPHIs = Phi.second; 1430 1431 // All MachineBasicBlocks exist, add them to the PHI. We assume IRTranslator 1432 // won't create extra control flow here, otherwise we need to find the 1433 // dominating predecessor here (or perhaps force the weirder IRTranslators 1434 // to provide a simple boundary). 1435 SmallSet<const BasicBlock *, 4> HandledPreds; 1436 1437 for (unsigned i = 0; i < PI->getNumIncomingValues(); ++i) { 1438 auto IRPred = PI->getIncomingBlock(i); 1439 if (HandledPreds.count(IRPred)) 1440 continue; 1441 1442 HandledPreds.insert(IRPred); 1443 ArrayRef<unsigned> ValRegs = getOrCreateVRegs(*PI->getIncomingValue(i)); 1444 for (auto Pred : getMachinePredBBs({IRPred, PI->getParent()})) { 1445 assert(Pred->isSuccessor(ComponentPHIs[0]->getParent()) && 1446 "incorrect CFG at MachineBasicBlock level"); 1447 for (unsigned j = 0; j < ValRegs.size(); ++j) { 1448 MachineInstrBuilder MIB(*MF, ComponentPHIs[j]); 1449 MIB.addUse(ValRegs[j]); 1450 MIB.addMBB(Pred); 1451 } 1452 } 1453 } 1454 } 1455 } 1456 1457 bool IRTranslator::valueIsSplit(const Value &V, 1458 SmallVectorImpl<uint64_t> *Offsets) { 1459 SmallVector<LLT, 4> SplitTys; 1460 computeValueLLTs(*DL, *V.getType(), SplitTys, Offsets); 1461 return SplitTys.size() > 1; 1462 } 1463 1464 bool IRTranslator::translate(const Instruction &Inst) { 1465 CurBuilder.setDebugLoc(Inst.getDebugLoc()); 1466 switch(Inst.getOpcode()) { 1467 #define HANDLE_INST(NUM, OPCODE, CLASS) \ 1468 case Instruction::OPCODE: return translate##OPCODE(Inst, CurBuilder); 1469 #include "llvm/IR/Instruction.def" 1470 default: 1471 return false; 1472 } 1473 } 1474 1475 bool IRTranslator::translate(const Constant &C, unsigned Reg) { 1476 if (auto CI = dyn_cast<ConstantInt>(&C)) 1477 EntryBuilder.buildConstant(Reg, *CI); 1478 else if (auto CF = dyn_cast<ConstantFP>(&C)) 1479 EntryBuilder.buildFConstant(Reg, *CF); 1480 else if (isa<UndefValue>(C)) 1481 EntryBuilder.buildUndef(Reg); 1482 else if (isa<ConstantPointerNull>(C)) { 1483 // As we are trying to build a constant val of 0 into a pointer, 1484 // insert a cast to make them correct with respect to types. 1485 unsigned NullSize = DL->getTypeSizeInBits(C.getType()); 1486 auto *ZeroTy = Type::getIntNTy(C.getContext(), NullSize); 1487 auto *ZeroVal = ConstantInt::get(ZeroTy, 0); 1488 unsigned ZeroReg = getOrCreateVReg(*ZeroVal); 1489 EntryBuilder.buildCast(Reg, ZeroReg); 1490 } else if (auto GV = dyn_cast<GlobalValue>(&C)) 1491 EntryBuilder.buildGlobalValue(Reg, GV); 1492 else if (auto CAZ = dyn_cast<ConstantAggregateZero>(&C)) { 1493 if (!CAZ->getType()->isVectorTy()) 1494 return false; 1495 // Return the scalar if it is a <1 x Ty> vector. 1496 if (CAZ->getNumElements() == 1) 1497 return translate(*CAZ->getElementValue(0u), Reg); 1498 std::vector<unsigned> Ops; 1499 for (unsigned i = 0; i < CAZ->getNumElements(); ++i) { 1500 Constant &Elt = *CAZ->getElementValue(i); 1501 Ops.push_back(getOrCreateVReg(Elt)); 1502 } 1503 EntryBuilder.buildMerge(Reg, Ops); 1504 } else if (auto CV = dyn_cast<ConstantDataVector>(&C)) { 1505 // Return the scalar if it is a <1 x Ty> vector. 1506 if (CV->getNumElements() == 1) 1507 return translate(*CV->getElementAsConstant(0), Reg); 1508 std::vector<unsigned> Ops; 1509 for (unsigned i = 0; i < CV->getNumElements(); ++i) { 1510 Constant &Elt = *CV->getElementAsConstant(i); 1511 Ops.push_back(getOrCreateVReg(Elt)); 1512 } 1513 EntryBuilder.buildMerge(Reg, Ops); 1514 } else if (auto CE = dyn_cast<ConstantExpr>(&C)) { 1515 switch(CE->getOpcode()) { 1516 #define HANDLE_INST(NUM, OPCODE, CLASS) \ 1517 case Instruction::OPCODE: return translate##OPCODE(*CE, EntryBuilder); 1518 #include "llvm/IR/Instruction.def" 1519 default: 1520 return false; 1521 } 1522 } else if (auto CV = dyn_cast<ConstantVector>(&C)) { 1523 if (CV->getNumOperands() == 1) 1524 return translate(*CV->getOperand(0), Reg); 1525 SmallVector<unsigned, 4> Ops; 1526 for (unsigned i = 0; i < CV->getNumOperands(); ++i) { 1527 Ops.push_back(getOrCreateVReg(*CV->getOperand(i))); 1528 } 1529 EntryBuilder.buildMerge(Reg, Ops); 1530 } else if (auto *BA = dyn_cast<BlockAddress>(&C)) { 1531 EntryBuilder.buildBlockAddress(Reg, BA); 1532 } else 1533 return false; 1534 1535 return true; 1536 } 1537 1538 void IRTranslator::finalizeFunction() { 1539 // Release the memory used by the different maps we 1540 // needed during the translation. 1541 PendingPHIs.clear(); 1542 VMap.reset(); 1543 FrameIndices.clear(); 1544 MachinePreds.clear(); 1545 // MachineIRBuilder::DebugLoc can outlive the DILocation it holds. Clear it 1546 // to avoid accessing free’d memory (in runOnMachineFunction) and to avoid 1547 // destroying it twice (in ~IRTranslator() and ~LLVMContext()) 1548 EntryBuilder = MachineIRBuilder(); 1549 CurBuilder = MachineIRBuilder(); 1550 } 1551 1552 bool IRTranslator::runOnMachineFunction(MachineFunction &CurMF) { 1553 MF = &CurMF; 1554 const Function &F = MF->getFunction(); 1555 if (F.empty()) 1556 return false; 1557 CLI = MF->getSubtarget().getCallLowering(); 1558 CurBuilder.setMF(*MF); 1559 EntryBuilder.setMF(*MF); 1560 MRI = &MF->getRegInfo(); 1561 DL = &F.getParent()->getDataLayout(); 1562 TPC = &getAnalysis<TargetPassConfig>(); 1563 ORE = llvm::make_unique<OptimizationRemarkEmitter>(&F); 1564 1565 assert(PendingPHIs.empty() && "stale PHIs"); 1566 1567 if (!DL->isLittleEndian()) { 1568 // Currently we don't properly handle big endian code. 1569 OptimizationRemarkMissed R("gisel-irtranslator", "GISelFailure", 1570 F.getSubprogram(), &F.getEntryBlock()); 1571 R << "unable to translate in big endian mode"; 1572 reportTranslationError(*MF, *TPC, *ORE, R); 1573 } 1574 1575 // Release the per-function state when we return, whether we succeeded or not. 1576 auto FinalizeOnReturn = make_scope_exit([this]() { finalizeFunction(); }); 1577 1578 // Setup a separate basic-block for the arguments and constants 1579 MachineBasicBlock *EntryBB = MF->CreateMachineBasicBlock(); 1580 MF->push_back(EntryBB); 1581 EntryBuilder.setMBB(*EntryBB); 1582 1583 // Create all blocks, in IR order, to preserve the layout. 1584 for (const BasicBlock &BB: F) { 1585 auto *&MBB = BBToMBB[&BB]; 1586 1587 MBB = MF->CreateMachineBasicBlock(&BB); 1588 MF->push_back(MBB); 1589 1590 if (BB.hasAddressTaken()) 1591 MBB->setHasAddressTaken(); 1592 } 1593 1594 // Make our arguments/constants entry block fallthrough to the IR entry block. 1595 EntryBB->addSuccessor(&getMBB(F.front())); 1596 1597 // Lower the actual args into this basic block. 1598 SmallVector<unsigned, 8> VRegArgs; 1599 for (const Argument &Arg: F.args()) { 1600 if (DL->getTypeStoreSize(Arg.getType()) == 0) 1601 continue; // Don't handle zero sized types. 1602 VRegArgs.push_back( 1603 MRI->createGenericVirtualRegister(getLLTForType(*Arg.getType(), *DL))); 1604 } 1605 1606 // We don't currently support translating swifterror or swiftself functions. 1607 for (auto &Arg : F.args()) { 1608 if (Arg.hasSwiftErrorAttr() || Arg.hasSwiftSelfAttr()) { 1609 OptimizationRemarkMissed R("gisel-irtranslator", "GISelFailure", 1610 F.getSubprogram(), &F.getEntryBlock()); 1611 R << "unable to lower arguments due to swifterror/swiftself: " 1612 << ore::NV("Prototype", F.getType()); 1613 reportTranslationError(*MF, *TPC, *ORE, R); 1614 return false; 1615 } 1616 } 1617 1618 if (!CLI->lowerFormalArguments(EntryBuilder, F, VRegArgs)) { 1619 OptimizationRemarkMissed R("gisel-irtranslator", "GISelFailure", 1620 F.getSubprogram(), &F.getEntryBlock()); 1621 R << "unable to lower arguments: " << ore::NV("Prototype", F.getType()); 1622 reportTranslationError(*MF, *TPC, *ORE, R); 1623 return false; 1624 } 1625 1626 auto ArgIt = F.arg_begin(); 1627 for (auto &VArg : VRegArgs) { 1628 // If the argument is an unsplit scalar then don't use unpackRegs to avoid 1629 // creating redundant copies. 1630 if (!valueIsSplit(*ArgIt, VMap.getOffsets(*ArgIt))) { 1631 auto &VRegs = *VMap.getVRegs(cast<Value>(*ArgIt)); 1632 assert(VRegs.empty() && "VRegs already populated?"); 1633 VRegs.push_back(VArg); 1634 } else { 1635 unpackRegs(*ArgIt, VArg, EntryBuilder); 1636 } 1637 ArgIt++; 1638 } 1639 1640 // Need to visit defs before uses when translating instructions. 1641 ReversePostOrderTraversal<const Function *> RPOT(&F); 1642 for (const BasicBlock *BB : RPOT) { 1643 MachineBasicBlock &MBB = getMBB(*BB); 1644 // Set the insertion point of all the following translations to 1645 // the end of this basic block. 1646 CurBuilder.setMBB(MBB); 1647 1648 for (const Instruction &Inst : *BB) { 1649 if (translate(Inst)) 1650 continue; 1651 1652 OptimizationRemarkMissed R("gisel-irtranslator", "GISelFailure", 1653 Inst.getDebugLoc(), BB); 1654 R << "unable to translate instruction: " << ore::NV("Opcode", &Inst); 1655 1656 if (ORE->allowExtraAnalysis("gisel-irtranslator")) { 1657 std::string InstStrStorage; 1658 raw_string_ostream InstStr(InstStrStorage); 1659 InstStr << Inst; 1660 1661 R << ": '" << InstStr.str() << "'"; 1662 } 1663 1664 reportTranslationError(*MF, *TPC, *ORE, R); 1665 return false; 1666 } 1667 } 1668 1669 finishPendingPhis(); 1670 1671 // Merge the argument lowering and constants block with its single 1672 // successor, the LLVM-IR entry block. We want the basic block to 1673 // be maximal. 1674 assert(EntryBB->succ_size() == 1 && 1675 "Custom BB used for lowering should have only one successor"); 1676 // Get the successor of the current entry block. 1677 MachineBasicBlock &NewEntryBB = **EntryBB->succ_begin(); 1678 assert(NewEntryBB.pred_size() == 1 && 1679 "LLVM-IR entry block has a predecessor!?"); 1680 // Move all the instruction from the current entry block to the 1681 // new entry block. 1682 NewEntryBB.splice(NewEntryBB.begin(), EntryBB, EntryBB->begin(), 1683 EntryBB->end()); 1684 1685 // Update the live-in information for the new entry block. 1686 for (const MachineBasicBlock::RegisterMaskPair &LiveIn : EntryBB->liveins()) 1687 NewEntryBB.addLiveIn(LiveIn); 1688 NewEntryBB.sortUniqueLiveIns(); 1689 1690 // Get rid of the now empty basic block. 1691 EntryBB->removeSuccessor(&NewEntryBB); 1692 MF->remove(EntryBB); 1693 MF->DeleteMachineBasicBlock(EntryBB); 1694 1695 assert(&MF->front() == &NewEntryBB && 1696 "New entry wasn't next in the list of basic block!"); 1697 1698 // Initialize stack protector information. 1699 StackProtector &SP = getAnalysis<StackProtector>(); 1700 SP.copyToMachineFrameInfo(MF->getFrameInfo()); 1701 1702 return false; 1703 } 1704