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