1 //===-- llvm/CodeGen/GlobalISel/IRTranslator.cpp - IRTranslator --*- C++ -*-==//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 /// This file implements the IRTranslator class.
11 //===----------------------------------------------------------------------===//
13 #include "llvm/CodeGen/GlobalISel/IRTranslator.h"
15 #include "llvm/ADT/SmallSet.h"
16 #include "llvm/ADT/SmallVector.h"
17 #include "llvm/CodeGen/GlobalISel/CallLowering.h"
18 #include "llvm/CodeGen/Analysis.h"
19 #include "llvm/CodeGen/MachineFunction.h"
20 #include "llvm/CodeGen/MachineFrameInfo.h"
21 #include "llvm/CodeGen/MachineModuleInfo.h"
22 #include "llvm/CodeGen/MachineRegisterInfo.h"
23 #include "llvm/CodeGen/TargetPassConfig.h"
24 #include "llvm/IR/Constant.h"
25 #include "llvm/IR/DebugInfo.h"
26 #include "llvm/IR/Function.h"
27 #include "llvm/IR/GetElementPtrTypeIterator.h"
28 #include "llvm/IR/IntrinsicInst.h"
29 #include "llvm/IR/Type.h"
30 #include "llvm/IR/Value.h"
31 #include "llvm/Target/TargetFrameLowering.h"
32 #include "llvm/Target/TargetIntrinsicInfo.h"
33 #include "llvm/Target/TargetLowering.h"
35 #define DEBUG_TYPE "irtranslator"
39 char IRTranslator::ID = 0;
40 INITIALIZE_PASS_BEGIN(IRTranslator, DEBUG_TYPE, "IRTranslator LLVM IR -> MI",
42 INITIALIZE_PASS_DEPENDENCY(TargetPassConfig)
43 INITIALIZE_PASS_END(IRTranslator, DEBUG_TYPE, "IRTranslator LLVM IR -> MI",
46 static void reportTranslationError(const Value &V, const Twine &Message) {
47 std::string ErrStorage;
48 raw_string_ostream Err(ErrStorage);
49 Err << Message << ": " << V << '\n';
50 report_fatal_error(Err.str());
53 IRTranslator::IRTranslator() : MachineFunctionPass(ID), MRI(nullptr) {
54 initializeIRTranslatorPass(*PassRegistry::getPassRegistry());
57 void IRTranslator::getAnalysisUsage(AnalysisUsage &AU) const {
58 AU.addRequired<TargetPassConfig>();
59 MachineFunctionPass::getAnalysisUsage(AU);
63 unsigned IRTranslator::getOrCreateVReg(const Value &Val) {
64 unsigned &ValReg = ValToVReg[&Val];
69 // Fill ValRegsSequence with the sequence of registers
70 // we need to concat together to produce the value.
71 assert(Val.getType()->isSized() &&
72 "Don't know how to create an empty vreg");
73 unsigned VReg = MRI->createGenericVirtualRegister(LLT{*Val.getType(), *DL});
76 if (auto CV = dyn_cast<Constant>(&Val)) {
77 bool Success = translate(*CV, VReg);
79 if (!TPC->isGlobalISelAbortEnabled()) {
80 MF->getProperties().set(
81 MachineFunctionProperties::Property::FailedISel);
84 reportTranslationError(Val, "unable to translate constant");
91 int IRTranslator::getOrCreateFrameIndex(const AllocaInst &AI) {
92 if (FrameIndices.find(&AI) != FrameIndices.end())
93 return FrameIndices[&AI];
95 unsigned ElementSize = DL->getTypeStoreSize(AI.getAllocatedType());
97 ElementSize * cast<ConstantInt>(AI.getArraySize())->getZExtValue();
99 // Always allocate at least one byte.
100 Size = std::max(Size, 1u);
102 unsigned Alignment = AI.getAlignment();
104 Alignment = DL->getABITypeAlignment(AI.getAllocatedType());
106 int &FI = FrameIndices[&AI];
107 FI = MF->getFrameInfo().CreateStackObject(Size, Alignment, false, &AI);
111 unsigned IRTranslator::getMemOpAlignment(const Instruction &I) {
112 unsigned Alignment = 0;
113 Type *ValTy = nullptr;
114 if (const StoreInst *SI = dyn_cast<StoreInst>(&I)) {
115 Alignment = SI->getAlignment();
116 ValTy = SI->getValueOperand()->getType();
117 } else if (const LoadInst *LI = dyn_cast<LoadInst>(&I)) {
118 Alignment = LI->getAlignment();
119 ValTy = LI->getType();
120 } else if (!TPC->isGlobalISelAbortEnabled()) {
121 MF->getProperties().set(
122 MachineFunctionProperties::Property::FailedISel);
125 llvm_unreachable("unhandled memory instruction");
127 return Alignment ? Alignment : DL->getABITypeAlignment(ValTy);
130 MachineBasicBlock &IRTranslator::getOrCreateBB(const BasicBlock &BB) {
131 MachineBasicBlock *&MBB = BBToMBB[&BB];
133 MBB = MF->CreateMachineBasicBlock(&BB);
136 if (BB.hasAddressTaken())
137 MBB->setHasAddressTaken();
142 void IRTranslator::addMachineCFGPred(CFGEdge Edge, MachineBasicBlock *NewPred) {
143 assert(NewPred && "new predecessor must be a real MachineBasicBlock");
144 MachinePreds[Edge].push_back(NewPred);
147 bool IRTranslator::translateBinaryOp(unsigned Opcode, const User &U,
148 MachineIRBuilder &MIRBuilder) {
149 // FIXME: handle signed/unsigned wrapping flags.
151 // Get or create a virtual register for each value.
152 // Unless the value is a Constant => loadimm cst?
153 // or inline constant each time?
154 // Creation of a virtual register needs to have a size.
155 unsigned Op0 = getOrCreateVReg(*U.getOperand(0));
156 unsigned Op1 = getOrCreateVReg(*U.getOperand(1));
157 unsigned Res = getOrCreateVReg(U);
158 MIRBuilder.buildInstr(Opcode).addDef(Res).addUse(Op0).addUse(Op1);
162 bool IRTranslator::translateCompare(const User &U,
163 MachineIRBuilder &MIRBuilder) {
164 const CmpInst *CI = dyn_cast<CmpInst>(&U);
165 unsigned Op0 = getOrCreateVReg(*U.getOperand(0));
166 unsigned Op1 = getOrCreateVReg(*U.getOperand(1));
167 unsigned Res = getOrCreateVReg(U);
168 CmpInst::Predicate Pred =
169 CI ? CI->getPredicate() : static_cast<CmpInst::Predicate>(
170 cast<ConstantExpr>(U).getPredicate());
172 if (CmpInst::isIntPredicate(Pred))
173 MIRBuilder.buildICmp(Pred, Res, Op0, Op1);
175 MIRBuilder.buildFCmp(Pred, Res, Op0, Op1);
180 bool IRTranslator::translateRet(const User &U, MachineIRBuilder &MIRBuilder) {
181 const ReturnInst &RI = cast<ReturnInst>(U);
182 const Value *Ret = RI.getReturnValue();
183 // The target may mess up with the insertion point, but
184 // this is not important as a return is the last instruction
185 // of the block anyway.
186 return CLI->lowerReturn(MIRBuilder, Ret, !Ret ? 0 : getOrCreateVReg(*Ret));
189 bool IRTranslator::translateBr(const User &U, MachineIRBuilder &MIRBuilder) {
190 const BranchInst &BrInst = cast<BranchInst>(U);
192 if (!BrInst.isUnconditional()) {
193 // We want a G_BRCOND to the true BB followed by an unconditional branch.
194 unsigned Tst = getOrCreateVReg(*BrInst.getCondition());
195 const BasicBlock &TrueTgt = *cast<BasicBlock>(BrInst.getSuccessor(Succ++));
196 MachineBasicBlock &TrueBB = getOrCreateBB(TrueTgt);
197 MIRBuilder.buildBrCond(Tst, TrueBB);
200 const BasicBlock &BrTgt = *cast<BasicBlock>(BrInst.getSuccessor(Succ));
201 MachineBasicBlock &TgtBB = getOrCreateBB(BrTgt);
202 MIRBuilder.buildBr(TgtBB);
205 MachineBasicBlock &CurBB = MIRBuilder.getMBB();
206 for (const BasicBlock *Succ : BrInst.successors())
207 CurBB.addSuccessor(&getOrCreateBB(*Succ));
211 bool IRTranslator::translateSwitch(const User &U,
212 MachineIRBuilder &MIRBuilder) {
213 // For now, just translate as a chain of conditional branches.
214 // FIXME: could we share most of the logic/code in
215 // SelectionDAGBuilder::visitSwitch between SelectionDAG and GlobalISel?
216 // At first sight, it seems most of the logic in there is independent of
217 // SelectionDAG-specifics and a lot of work went in to optimize switch
218 // lowering in there.
220 const SwitchInst &SwInst = cast<SwitchInst>(U);
221 const unsigned SwCondValue = getOrCreateVReg(*SwInst.getCondition());
222 const BasicBlock *OrigBB = SwInst.getParent();
224 LLT LLTi1 = LLT(*Type::getInt1Ty(U.getContext()), *DL);
225 for (auto &CaseIt : SwInst.cases()) {
226 const unsigned CaseValueReg = getOrCreateVReg(*CaseIt.getCaseValue());
227 const unsigned Tst = MRI->createGenericVirtualRegister(LLTi1);
228 MIRBuilder.buildICmp(CmpInst::ICMP_EQ, Tst, CaseValueReg, SwCondValue);
229 MachineBasicBlock &CurMBB = MIRBuilder.getMBB();
230 const BasicBlock *TrueBB = CaseIt.getCaseSuccessor();
231 MachineBasicBlock &TrueMBB = getOrCreateBB(*TrueBB);
233 MIRBuilder.buildBrCond(Tst, TrueMBB);
234 CurMBB.addSuccessor(&TrueMBB);
235 addMachineCFGPred({OrigBB, TrueBB}, &CurMBB);
237 MachineBasicBlock *FalseMBB =
238 MF->CreateMachineBasicBlock(SwInst.getParent());
239 MF->push_back(FalseMBB);
240 MIRBuilder.buildBr(*FalseMBB);
241 CurMBB.addSuccessor(FalseMBB);
243 MIRBuilder.setMBB(*FalseMBB);
245 // handle default case
246 const BasicBlock *DefaultBB = SwInst.getDefaultDest();
247 MachineBasicBlock &DefaultMBB = getOrCreateBB(*DefaultBB);
248 MIRBuilder.buildBr(DefaultMBB);
249 MachineBasicBlock &CurMBB = MIRBuilder.getMBB();
250 CurMBB.addSuccessor(&DefaultMBB);
251 addMachineCFGPred({OrigBB, DefaultBB}, &CurMBB);
256 bool IRTranslator::translateIndirectBr(const User &U,
257 MachineIRBuilder &MIRBuilder) {
258 const IndirectBrInst &BrInst = cast<IndirectBrInst>(U);
260 const unsigned Tgt = getOrCreateVReg(*BrInst.getAddress());
261 MIRBuilder.buildBrIndirect(Tgt);
264 MachineBasicBlock &CurBB = MIRBuilder.getMBB();
265 for (const BasicBlock *Succ : BrInst.successors())
266 CurBB.addSuccessor(&getOrCreateBB(*Succ));
271 bool IRTranslator::translateLoad(const User &U, MachineIRBuilder &MIRBuilder) {
272 const LoadInst &LI = cast<LoadInst>(U);
274 if (!TPC->isGlobalISelAbortEnabled() && LI.isAtomic())
277 assert(!LI.isAtomic() && "only non-atomic loads are supported at the moment");
278 auto Flags = LI.isVolatile() ? MachineMemOperand::MOVolatile
279 : MachineMemOperand::MONone;
280 Flags |= MachineMemOperand::MOLoad;
282 unsigned Res = getOrCreateVReg(LI);
283 unsigned Addr = getOrCreateVReg(*LI.getPointerOperand());
284 LLT VTy{*LI.getType(), *DL}, PTy{*LI.getPointerOperand()->getType(), *DL};
285 MIRBuilder.buildLoad(
287 *MF->getMachineMemOperand(MachinePointerInfo(LI.getPointerOperand()),
288 Flags, DL->getTypeStoreSize(LI.getType()),
289 getMemOpAlignment(LI)));
293 bool IRTranslator::translateStore(const User &U, MachineIRBuilder &MIRBuilder) {
294 const StoreInst &SI = cast<StoreInst>(U);
296 if (!TPC->isGlobalISelAbortEnabled() && SI.isAtomic())
299 assert(!SI.isAtomic() && "only non-atomic stores supported at the moment");
300 auto Flags = SI.isVolatile() ? MachineMemOperand::MOVolatile
301 : MachineMemOperand::MONone;
302 Flags |= MachineMemOperand::MOStore;
304 unsigned Val = getOrCreateVReg(*SI.getValueOperand());
305 unsigned Addr = getOrCreateVReg(*SI.getPointerOperand());
306 LLT VTy{*SI.getValueOperand()->getType(), *DL},
307 PTy{*SI.getPointerOperand()->getType(), *DL};
309 MIRBuilder.buildStore(
311 *MF->getMachineMemOperand(
312 MachinePointerInfo(SI.getPointerOperand()), Flags,
313 DL->getTypeStoreSize(SI.getValueOperand()->getType()),
314 getMemOpAlignment(SI)));
318 bool IRTranslator::translateExtractValue(const User &U,
319 MachineIRBuilder &MIRBuilder) {
320 const Value *Src = U.getOperand(0);
321 Type *Int32Ty = Type::getInt32Ty(U.getContext());
322 SmallVector<Value *, 1> Indices;
324 // getIndexedOffsetInType is designed for GEPs, so the first index is the
325 // usual array element rather than looking into the actual aggregate.
326 Indices.push_back(ConstantInt::get(Int32Ty, 0));
328 if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(&U)) {
329 for (auto Idx : EVI->indices())
330 Indices.push_back(ConstantInt::get(Int32Ty, Idx));
332 for (unsigned i = 1; i < U.getNumOperands(); ++i)
333 Indices.push_back(U.getOperand(i));
336 uint64_t Offset = 8 * DL->getIndexedOffsetInType(Src->getType(), Indices);
338 unsigned Res = getOrCreateVReg(U);
339 MIRBuilder.buildExtract(Res, Offset, getOrCreateVReg(*Src));
344 bool IRTranslator::translateInsertValue(const User &U,
345 MachineIRBuilder &MIRBuilder) {
346 const Value *Src = U.getOperand(0);
347 Type *Int32Ty = Type::getInt32Ty(U.getContext());
348 SmallVector<Value *, 1> Indices;
350 // getIndexedOffsetInType is designed for GEPs, so the first index is the
351 // usual array element rather than looking into the actual aggregate.
352 Indices.push_back(ConstantInt::get(Int32Ty, 0));
354 if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(&U)) {
355 for (auto Idx : IVI->indices())
356 Indices.push_back(ConstantInt::get(Int32Ty, Idx));
358 for (unsigned i = 2; i < U.getNumOperands(); ++i)
359 Indices.push_back(U.getOperand(i));
362 uint64_t Offset = 8 * DL->getIndexedOffsetInType(Src->getType(), Indices);
364 unsigned Res = getOrCreateVReg(U);
365 const Value &Inserted = *U.getOperand(1);
366 MIRBuilder.buildInsert(Res, getOrCreateVReg(*Src), getOrCreateVReg(Inserted),
372 bool IRTranslator::translateSelect(const User &U,
373 MachineIRBuilder &MIRBuilder) {
374 MIRBuilder.buildSelect(getOrCreateVReg(U), getOrCreateVReg(*U.getOperand(0)),
375 getOrCreateVReg(*U.getOperand(1)),
376 getOrCreateVReg(*U.getOperand(2)));
380 bool IRTranslator::translateBitCast(const User &U,
381 MachineIRBuilder &MIRBuilder) {
382 if (LLT{*U.getOperand(0)->getType(), *DL} == LLT{*U.getType(), *DL}) {
383 unsigned &Reg = ValToVReg[&U];
385 MIRBuilder.buildCopy(Reg, getOrCreateVReg(*U.getOperand(0)));
387 Reg = getOrCreateVReg(*U.getOperand(0));
390 return translateCast(TargetOpcode::G_BITCAST, U, MIRBuilder);
393 bool IRTranslator::translateCast(unsigned Opcode, const User &U,
394 MachineIRBuilder &MIRBuilder) {
395 unsigned Op = getOrCreateVReg(*U.getOperand(0));
396 unsigned Res = getOrCreateVReg(U);
397 MIRBuilder.buildInstr(Opcode).addDef(Res).addUse(Op);
401 bool IRTranslator::translateGetElementPtr(const User &U,
402 MachineIRBuilder &MIRBuilder) {
403 // FIXME: support vector GEPs.
404 if (U.getType()->isVectorTy())
407 Value &Op0 = *U.getOperand(0);
408 unsigned BaseReg = getOrCreateVReg(Op0);
409 LLT PtrTy{*Op0.getType(), *DL};
410 unsigned PtrSize = DL->getPointerSizeInBits(PtrTy.getAddressSpace());
411 LLT OffsetTy = LLT::scalar(PtrSize);
414 for (gep_type_iterator GTI = gep_type_begin(&U), E = gep_type_end(&U);
416 const Value *Idx = GTI.getOperand();
417 if (StructType *StTy = GTI.getStructTypeOrNull()) {
418 unsigned Field = cast<Constant>(Idx)->getUniqueInteger().getZExtValue();
419 Offset += DL->getStructLayout(StTy)->getElementOffset(Field);
422 uint64_t ElementSize = DL->getTypeAllocSize(GTI.getIndexedType());
424 // If this is a scalar constant or a splat vector of constants,
425 // handle it quickly.
426 if (const auto *CI = dyn_cast<ConstantInt>(Idx)) {
427 Offset += ElementSize * CI->getSExtValue();
432 unsigned NewBaseReg = MRI->createGenericVirtualRegister(PtrTy);
433 unsigned OffsetReg = MRI->createGenericVirtualRegister(OffsetTy);
434 MIRBuilder.buildConstant(OffsetReg, Offset);
435 MIRBuilder.buildGEP(NewBaseReg, BaseReg, OffsetReg);
437 BaseReg = NewBaseReg;
441 // N = N + Idx * ElementSize;
442 unsigned ElementSizeReg = MRI->createGenericVirtualRegister(OffsetTy);
443 MIRBuilder.buildConstant(ElementSizeReg, ElementSize);
445 unsigned IdxReg = getOrCreateVReg(*Idx);
446 if (MRI->getType(IdxReg) != OffsetTy) {
447 unsigned NewIdxReg = MRI->createGenericVirtualRegister(OffsetTy);
448 MIRBuilder.buildSExtOrTrunc(NewIdxReg, IdxReg);
452 unsigned OffsetReg = MRI->createGenericVirtualRegister(OffsetTy);
453 MIRBuilder.buildMul(OffsetReg, ElementSizeReg, IdxReg);
455 unsigned NewBaseReg = MRI->createGenericVirtualRegister(PtrTy);
456 MIRBuilder.buildGEP(NewBaseReg, BaseReg, OffsetReg);
457 BaseReg = NewBaseReg;
462 unsigned OffsetReg = MRI->createGenericVirtualRegister(OffsetTy);
463 MIRBuilder.buildConstant(OffsetReg, Offset);
464 MIRBuilder.buildGEP(getOrCreateVReg(U), BaseReg, OffsetReg);
468 MIRBuilder.buildCopy(getOrCreateVReg(U), BaseReg);
472 bool IRTranslator::translateMemfunc(const CallInst &CI,
473 MachineIRBuilder &MIRBuilder,
475 LLT SizeTy{*CI.getArgOperand(2)->getType(), *DL};
476 Type *DstTy = CI.getArgOperand(0)->getType();
477 if (cast<PointerType>(DstTy)->getAddressSpace() != 0 ||
478 SizeTy.getSizeInBits() != DL->getPointerSizeInBits(0))
481 SmallVector<CallLowering::ArgInfo, 8> Args;
482 for (int i = 0; i < 3; ++i) {
483 const auto &Arg = CI.getArgOperand(i);
484 Args.emplace_back(getOrCreateVReg(*Arg), Arg->getType());
489 case Intrinsic::memmove:
490 case Intrinsic::memcpy: {
491 Type *SrcTy = CI.getArgOperand(1)->getType();
492 if(cast<PointerType>(SrcTy)->getAddressSpace() != 0)
494 Callee = ID == Intrinsic::memcpy ? "memcpy" : "memmove";
497 case Intrinsic::memset:
504 return CLI->lowerCall(MIRBuilder, MachineOperand::CreateES(Callee),
505 CallLowering::ArgInfo(0, CI.getType()), Args);
508 void IRTranslator::getStackGuard(unsigned DstReg,
509 MachineIRBuilder &MIRBuilder) {
510 const TargetRegisterInfo *TRI = MF->getSubtarget().getRegisterInfo();
511 MRI->setRegClass(DstReg, TRI->getPointerRegClass(*MF));
512 auto MIB = MIRBuilder.buildInstr(TargetOpcode::LOAD_STACK_GUARD);
515 auto &TLI = *MF->getSubtarget().getTargetLowering();
516 Value *Global = TLI.getSDagStackGuard(*MF->getFunction()->getParent());
520 MachinePointerInfo MPInfo(Global);
521 MachineInstr::mmo_iterator MemRefs = MF->allocateMemRefsArray(1);
522 auto Flags = MachineMemOperand::MOLoad | MachineMemOperand::MOInvariant |
523 MachineMemOperand::MODereferenceable;
525 MF->getMachineMemOperand(MPInfo, Flags, DL->getPointerSizeInBits() / 8,
526 DL->getPointerABIAlignment());
527 MIB.setMemRefs(MemRefs, MemRefs + 1);
530 bool IRTranslator::translateOverflowIntrinsic(const CallInst &CI, unsigned Op,
531 MachineIRBuilder &MIRBuilder) {
532 LLT Ty{*CI.getOperand(0)->getType(), *DL};
533 LLT s1 = LLT::scalar(1);
534 unsigned Width = Ty.getSizeInBits();
535 unsigned Res = MRI->createGenericVirtualRegister(Ty);
536 unsigned Overflow = MRI->createGenericVirtualRegister(s1);
537 auto MIB = MIRBuilder.buildInstr(Op)
540 .addUse(getOrCreateVReg(*CI.getOperand(0)))
541 .addUse(getOrCreateVReg(*CI.getOperand(1)));
543 if (Op == TargetOpcode::G_UADDE || Op == TargetOpcode::G_USUBE) {
544 unsigned Zero = MRI->createGenericVirtualRegister(s1);
545 EntryBuilder.buildConstant(Zero, 0);
549 MIRBuilder.buildSequence(getOrCreateVReg(CI), Res, 0, Overflow, Width);
553 bool IRTranslator::translateKnownIntrinsic(const CallInst &CI, Intrinsic::ID ID,
554 MachineIRBuilder &MIRBuilder) {
558 case Intrinsic::dbg_declare: {
559 const DbgDeclareInst &DI = cast<DbgDeclareInst>(CI);
560 assert(DI.getVariable() && "Missing variable");
562 const Value *Address = DI.getAddress();
563 if (!Address || isa<UndefValue>(Address)) {
564 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
568 unsigned Reg = getOrCreateVReg(*Address);
569 auto RegDef = MRI->def_instr_begin(Reg);
570 assert(DI.getVariable()->isValidLocationForIntrinsic(
571 MIRBuilder.getDebugLoc()) &&
572 "Expected inlined-at fields to agree");
574 if (RegDef != MRI->def_instr_end() &&
575 RegDef->getOpcode() == TargetOpcode::G_FRAME_INDEX) {
576 MIRBuilder.buildFIDbgValue(RegDef->getOperand(1).getIndex(),
577 DI.getVariable(), DI.getExpression());
579 MIRBuilder.buildDirectDbgValue(Reg, DI.getVariable(), DI.getExpression());
582 case Intrinsic::vaend:
583 // No target I know of cares about va_end. Certainly no in-tree target
584 // does. Simplest intrinsic ever!
586 case Intrinsic::dbg_value: {
587 // This form of DBG_VALUE is target-independent.
588 const DbgValueInst &DI = cast<DbgValueInst>(CI);
589 const Value *V = DI.getValue();
590 assert(DI.getVariable()->isValidLocationForIntrinsic(
591 MIRBuilder.getDebugLoc()) &&
592 "Expected inlined-at fields to agree");
594 // Currently the optimizer can produce this; insert an undef to
595 // help debugging. Probably the optimizer should not do this.
596 MIRBuilder.buildIndirectDbgValue(0, DI.getOffset(), DI.getVariable(),
598 } else if (const auto *CI = dyn_cast<Constant>(V)) {
599 MIRBuilder.buildConstDbgValue(*CI, DI.getOffset(), DI.getVariable(),
602 unsigned Reg = getOrCreateVReg(*V);
603 // FIXME: This does not handle register-indirect values at offset 0. The
604 // direct/indirect thing shouldn't really be handled by something as
605 // implicit as reg+noreg vs reg+imm in the first palce, but it seems
606 // pretty baked in right now.
607 if (DI.getOffset() != 0)
608 MIRBuilder.buildIndirectDbgValue(Reg, DI.getOffset(), DI.getVariable(),
611 MIRBuilder.buildDirectDbgValue(Reg, DI.getVariable(),
616 case Intrinsic::uadd_with_overflow:
617 return translateOverflowIntrinsic(CI, TargetOpcode::G_UADDE, MIRBuilder);
618 case Intrinsic::sadd_with_overflow:
619 return translateOverflowIntrinsic(CI, TargetOpcode::G_SADDO, MIRBuilder);
620 case Intrinsic::usub_with_overflow:
621 return translateOverflowIntrinsic(CI, TargetOpcode::G_USUBE, MIRBuilder);
622 case Intrinsic::ssub_with_overflow:
623 return translateOverflowIntrinsic(CI, TargetOpcode::G_SSUBO, MIRBuilder);
624 case Intrinsic::umul_with_overflow:
625 return translateOverflowIntrinsic(CI, TargetOpcode::G_UMULO, MIRBuilder);
626 case Intrinsic::smul_with_overflow:
627 return translateOverflowIntrinsic(CI, TargetOpcode::G_SMULO, MIRBuilder);
628 case Intrinsic::memcpy:
629 case Intrinsic::memmove:
630 case Intrinsic::memset:
631 return translateMemfunc(CI, MIRBuilder, ID);
632 case Intrinsic::eh_typeid_for: {
633 GlobalValue *GV = ExtractTypeInfo(CI.getArgOperand(0));
634 unsigned Reg = getOrCreateVReg(CI);
635 unsigned TypeID = MF->getTypeIDFor(GV);
636 MIRBuilder.buildConstant(Reg, TypeID);
639 case Intrinsic::objectsize: {
640 // If we don't know by now, we're never going to know.
641 const ConstantInt *Min = cast<ConstantInt>(CI.getArgOperand(1));
643 MIRBuilder.buildConstant(getOrCreateVReg(CI), Min->isZero() ? -1ULL : 0);
646 case Intrinsic::stackguard:
647 getStackGuard(getOrCreateVReg(CI), MIRBuilder);
649 case Intrinsic::stackprotector: {
650 LLT PtrTy{*CI.getArgOperand(0)->getType(), *DL};
651 unsigned GuardVal = MRI->createGenericVirtualRegister(PtrTy);
652 getStackGuard(GuardVal, MIRBuilder);
654 AllocaInst *Slot = cast<AllocaInst>(CI.getArgOperand(1));
655 MIRBuilder.buildStore(
656 GuardVal, getOrCreateVReg(*Slot),
657 *MF->getMachineMemOperand(
658 MachinePointerInfo::getFixedStack(*MF,
659 getOrCreateFrameIndex(*Slot)),
660 MachineMemOperand::MOStore | MachineMemOperand::MOVolatile,
661 PtrTy.getSizeInBits() / 8, 8));
668 bool IRTranslator::translateCall(const User &U, MachineIRBuilder &MIRBuilder) {
669 const CallInst &CI = cast<CallInst>(U);
670 auto TII = MF->getTarget().getIntrinsicInfo();
671 const Function *F = CI.getCalledFunction();
673 if (CI.isInlineAsm())
676 if (!F || !F->isIntrinsic()) {
677 unsigned Res = CI.getType()->isVoidTy() ? 0 : getOrCreateVReg(CI);
678 SmallVector<unsigned, 8> Args;
679 for (auto &Arg: CI.arg_operands())
680 Args.push_back(getOrCreateVReg(*Arg));
682 return CLI->lowerCall(MIRBuilder, CI, Res, Args, [&]() {
683 return getOrCreateVReg(*CI.getCalledValue());
687 Intrinsic::ID ID = F->getIntrinsicID();
688 if (TII && ID == Intrinsic::not_intrinsic)
689 ID = static_cast<Intrinsic::ID>(TII->getIntrinsicID(F));
691 assert(ID != Intrinsic::not_intrinsic && "unknown intrinsic");
693 if (translateKnownIntrinsic(CI, ID, MIRBuilder))
696 unsigned Res = CI.getType()->isVoidTy() ? 0 : getOrCreateVReg(CI);
697 MachineInstrBuilder MIB =
698 MIRBuilder.buildIntrinsic(ID, Res, !CI.doesNotAccessMemory());
700 for (auto &Arg : CI.arg_operands()) {
701 if (ConstantInt *CI = dyn_cast<ConstantInt>(Arg))
702 MIB.addImm(CI->getSExtValue());
704 MIB.addUse(getOrCreateVReg(*Arg));
709 bool IRTranslator::translateInvoke(const User &U,
710 MachineIRBuilder &MIRBuilder) {
711 const InvokeInst &I = cast<InvokeInst>(U);
712 MCContext &Context = MF->getContext();
714 const BasicBlock *ReturnBB = I.getSuccessor(0);
715 const BasicBlock *EHPadBB = I.getSuccessor(1);
717 const Value *Callee(I.getCalledValue());
718 const Function *Fn = dyn_cast<Function>(Callee);
719 if (isa<InlineAsm>(Callee))
722 // FIXME: support invoking patchpoint and statepoint intrinsics.
723 if (Fn && Fn->isIntrinsic())
726 // FIXME: support whatever these are.
727 if (I.countOperandBundlesOfType(LLVMContext::OB_deopt))
730 // FIXME: support Windows exception handling.
731 if (!isa<LandingPadInst>(EHPadBB->front()))
735 // Emit the actual call, bracketed by EH_LABELs so that the MF knows about
736 // the region covered by the try.
737 MCSymbol *BeginSymbol = Context.createTempSymbol();
738 MIRBuilder.buildInstr(TargetOpcode::EH_LABEL).addSym(BeginSymbol);
740 unsigned Res = I.getType()->isVoidTy() ? 0 : getOrCreateVReg(I);
741 SmallVector<unsigned, 8> Args;
742 for (auto &Arg: I.arg_operands())
743 Args.push_back(getOrCreateVReg(*Arg));
745 CLI->lowerCall(MIRBuilder, I, Res, Args,
746 [&]() { return getOrCreateVReg(*I.getCalledValue()); });
748 MCSymbol *EndSymbol = Context.createTempSymbol();
749 MIRBuilder.buildInstr(TargetOpcode::EH_LABEL).addSym(EndSymbol);
751 // FIXME: track probabilities.
752 MachineBasicBlock &EHPadMBB = getOrCreateBB(*EHPadBB),
753 &ReturnMBB = getOrCreateBB(*ReturnBB);
754 MF->addInvoke(&EHPadMBB, BeginSymbol, EndSymbol);
755 MIRBuilder.getMBB().addSuccessor(&ReturnMBB);
756 MIRBuilder.getMBB().addSuccessor(&EHPadMBB);
757 MIRBuilder.buildBr(ReturnMBB);
762 bool IRTranslator::translateLandingPad(const User &U,
763 MachineIRBuilder &MIRBuilder) {
764 const LandingPadInst &LP = cast<LandingPadInst>(U);
766 MachineBasicBlock &MBB = MIRBuilder.getMBB();
767 addLandingPadInfo(LP, MBB);
771 // If there aren't registers to copy the values into (e.g., during SjLj
772 // exceptions), then don't bother.
773 auto &TLI = *MF->getSubtarget().getTargetLowering();
774 const Constant *PersonalityFn = MF->getFunction()->getPersonalityFn();
775 if (TLI.getExceptionPointerRegister(PersonalityFn) == 0 &&
776 TLI.getExceptionSelectorRegister(PersonalityFn) == 0)
779 // If landingpad's return type is token type, we don't create DAG nodes
780 // for its exception pointer and selector value. The extraction of exception
781 // pointer or selector value from token type landingpads is not currently
783 if (LP.getType()->isTokenTy())
786 // Add a label to mark the beginning of the landing pad. Deletion of the
787 // landing pad can thus be detected via the MachineModuleInfo.
788 MIRBuilder.buildInstr(TargetOpcode::EH_LABEL)
789 .addSym(MF->addLandingPad(&MBB));
791 SmallVector<LLT, 2> Tys;
792 for (Type *Ty : cast<StructType>(LP.getType())->elements())
793 Tys.push_back(LLT{*Ty, *DL});
794 assert(Tys.size() == 2 && "Only two-valued landingpads are supported");
796 // Mark exception register as live in.
797 SmallVector<unsigned, 2> Regs;
798 SmallVector<uint64_t, 2> Offsets;
799 if (unsigned Reg = TLI.getExceptionPointerRegister(PersonalityFn)) {
801 unsigned VReg = MRI->createGenericVirtualRegister(Tys[0]);
802 MIRBuilder.buildCopy(VReg, Reg);
803 Regs.push_back(VReg);
804 Offsets.push_back(0);
807 if (unsigned Reg = TLI.getExceptionSelectorRegister(PersonalityFn)) {
810 // N.b. the exception selector register always has pointer type and may not
811 // match the actual IR-level type in the landingpad so an extra cast is
813 unsigned PtrVReg = MRI->createGenericVirtualRegister(Tys[0]);
814 MIRBuilder.buildCopy(PtrVReg, Reg);
816 unsigned VReg = MRI->createGenericVirtualRegister(Tys[1]);
817 MIRBuilder.buildInstr(TargetOpcode::G_PTRTOINT)
820 Regs.push_back(VReg);
821 Offsets.push_back(Tys[0].getSizeInBits());
824 MIRBuilder.buildSequence(getOrCreateVReg(LP), Regs, Offsets);
828 bool IRTranslator::translateAlloca(const User &U,
829 MachineIRBuilder &MIRBuilder) {
830 auto &AI = cast<AllocaInst>(U);
832 if (AI.isStaticAlloca()) {
833 unsigned Res = getOrCreateVReg(AI);
834 int FI = getOrCreateFrameIndex(AI);
835 MIRBuilder.buildFrameIndex(Res, FI);
839 // Now we're in the harder dynamic case.
840 Type *Ty = AI.getAllocatedType();
842 std::max((unsigned)DL->getPrefTypeAlignment(Ty), AI.getAlignment());
844 unsigned NumElts = getOrCreateVReg(*AI.getArraySize());
846 LLT IntPtrTy = LLT::scalar(DL->getPointerSizeInBits());
847 if (MRI->getType(NumElts) != IntPtrTy) {
848 unsigned ExtElts = MRI->createGenericVirtualRegister(IntPtrTy);
849 MIRBuilder.buildZExtOrTrunc(ExtElts, NumElts);
853 unsigned AllocSize = MRI->createGenericVirtualRegister(IntPtrTy);
854 unsigned TySize = MRI->createGenericVirtualRegister(IntPtrTy);
855 MIRBuilder.buildConstant(TySize, DL->getTypeAllocSize(Ty));
856 MIRBuilder.buildMul(AllocSize, NumElts, TySize);
858 LLT PtrTy = LLT{*AI.getType(), *DL};
859 auto &TLI = *MF->getSubtarget().getTargetLowering();
860 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
862 unsigned SPTmp = MRI->createGenericVirtualRegister(PtrTy);
863 MIRBuilder.buildCopy(SPTmp, SPReg);
865 unsigned SPInt = MRI->createGenericVirtualRegister(IntPtrTy);
866 MIRBuilder.buildInstr(TargetOpcode::G_PTRTOINT).addDef(SPInt).addUse(SPTmp);
868 unsigned AllocInt = MRI->createGenericVirtualRegister(IntPtrTy);
869 MIRBuilder.buildSub(AllocInt, SPInt, AllocSize);
871 // Handle alignment. We have to realign if the allocation granule was smaller
872 // than stack alignment, or the specific alloca requires more than stack
874 unsigned StackAlign =
875 MF->getSubtarget().getFrameLowering()->getStackAlignment();
876 Align = std::max(Align, StackAlign);
877 if (Align > StackAlign || DL->getTypeAllocSize(Ty) % StackAlign != 0) {
878 // Round the size of the allocation up to the stack alignment size
879 // by add SA-1 to the size. This doesn't overflow because we're computing
880 // an address inside an alloca.
881 unsigned TmpSize = MRI->createGenericVirtualRegister(IntPtrTy);
882 unsigned AlignMinus1 = MRI->createGenericVirtualRegister(IntPtrTy);
883 MIRBuilder.buildConstant(AlignMinus1, Align - 1);
884 MIRBuilder.buildSub(TmpSize, AllocInt, AlignMinus1);
886 unsigned AlignedAlloc = MRI->createGenericVirtualRegister(IntPtrTy);
887 unsigned AlignMask = MRI->createGenericVirtualRegister(IntPtrTy);
888 MIRBuilder.buildConstant(AlignMask, -(uint64_t)Align);
889 MIRBuilder.buildAnd(AlignedAlloc, TmpSize, AlignMask);
891 AllocInt = AlignedAlloc;
894 unsigned DstReg = getOrCreateVReg(AI);
895 MIRBuilder.buildInstr(TargetOpcode::G_INTTOPTR)
899 MIRBuilder.buildCopy(SPReg, DstReg);
901 MF->getFrameInfo().CreateVariableSizedObject(Align ? Align : 1, &AI);
902 assert(MF->getFrameInfo().hasVarSizedObjects());
906 bool IRTranslator::translatePHI(const User &U, MachineIRBuilder &MIRBuilder) {
907 const PHINode &PI = cast<PHINode>(U);
908 auto MIB = MIRBuilder.buildInstr(TargetOpcode::PHI);
909 MIB.addDef(getOrCreateVReg(PI));
911 PendingPHIs.emplace_back(&PI, MIB.getInstr());
915 void IRTranslator::finishPendingPhis() {
916 for (std::pair<const PHINode *, MachineInstr *> &Phi : PendingPHIs) {
917 const PHINode *PI = Phi.first;
918 MachineInstrBuilder MIB(*MF, Phi.second);
920 // All MachineBasicBlocks exist, add them to the PHI. We assume IRTranslator
921 // won't create extra control flow here, otherwise we need to find the
922 // dominating predecessor here (or perhaps force the weirder IRTranslators
923 // to provide a simple boundary).
924 SmallSet<const BasicBlock *, 4> HandledPreds;
926 for (unsigned i = 0; i < PI->getNumIncomingValues(); ++i) {
927 auto IRPred = PI->getIncomingBlock(i);
928 if (HandledPreds.count(IRPred))
931 HandledPreds.insert(IRPred);
932 unsigned ValReg = getOrCreateVReg(*PI->getIncomingValue(i));
933 for (auto Pred : getMachinePredBBs({IRPred, PI->getParent()})) {
934 assert(Pred->isSuccessor(MIB->getParent()) &&
935 "incorrect CFG at MachineBasicBlock level");
943 bool IRTranslator::translate(const Instruction &Inst) {
944 CurBuilder.setDebugLoc(Inst.getDebugLoc());
945 switch(Inst.getOpcode()) {
946 #define HANDLE_INST(NUM, OPCODE, CLASS) \
947 case Instruction::OPCODE: return translate##OPCODE(Inst, CurBuilder);
948 #include "llvm/IR/Instruction.def"
950 if (!TPC->isGlobalISelAbortEnabled())
952 llvm_unreachable("unknown opcode");
956 bool IRTranslator::translate(const Constant &C, unsigned Reg) {
957 if (auto CI = dyn_cast<ConstantInt>(&C))
958 EntryBuilder.buildConstant(Reg, *CI);
959 else if (auto CF = dyn_cast<ConstantFP>(&C))
960 EntryBuilder.buildFConstant(Reg, *CF);
961 else if (isa<UndefValue>(C))
962 EntryBuilder.buildInstr(TargetOpcode::IMPLICIT_DEF).addDef(Reg);
963 else if (isa<ConstantPointerNull>(C))
964 EntryBuilder.buildConstant(Reg, 0);
965 else if (auto GV = dyn_cast<GlobalValue>(&C))
966 EntryBuilder.buildGlobalValue(Reg, GV);
967 else if (auto CE = dyn_cast<ConstantExpr>(&C)) {
968 switch(CE->getOpcode()) {
969 #define HANDLE_INST(NUM, OPCODE, CLASS) \
970 case Instruction::OPCODE: return translate##OPCODE(*CE, EntryBuilder);
971 #include "llvm/IR/Instruction.def"
973 if (!TPC->isGlobalISelAbortEnabled())
975 llvm_unreachable("unknown opcode");
977 } else if (!TPC->isGlobalISelAbortEnabled())
980 llvm_unreachable("unhandled constant kind");
985 void IRTranslator::finalizeFunction() {
986 // Release the memory used by the different maps we
987 // needed during the translation.
990 FrameIndices.clear();
992 MachinePreds.clear();
995 bool IRTranslator::runOnMachineFunction(MachineFunction &CurMF) {
997 const Function &F = *MF->getFunction();
1000 CLI = MF->getSubtarget().getCallLowering();
1001 CurBuilder.setMF(*MF);
1002 EntryBuilder.setMF(*MF);
1003 MRI = &MF->getRegInfo();
1004 DL = &F.getParent()->getDataLayout();
1005 TPC = &getAnalysis<TargetPassConfig>();
1007 assert(PendingPHIs.empty() && "stale PHIs");
1009 // Setup a separate basic-block for the arguments and constants, falling
1010 // through to the IR-level Function's entry block.
1011 MachineBasicBlock *EntryBB = MF->CreateMachineBasicBlock();
1012 MF->push_back(EntryBB);
1013 EntryBB->addSuccessor(&getOrCreateBB(F.front()));
1014 EntryBuilder.setMBB(*EntryBB);
1016 // Lower the actual args into this basic block.
1017 SmallVector<unsigned, 8> VRegArgs;
1018 for (const Argument &Arg: F.args())
1019 VRegArgs.push_back(getOrCreateVReg(Arg));
1020 bool Succeeded = CLI->lowerFormalArguments(EntryBuilder, F, VRegArgs);
1022 if (!TPC->isGlobalISelAbortEnabled()) {
1023 MF->getProperties().set(
1024 MachineFunctionProperties::Property::FailedISel);
1028 report_fatal_error("Unable to lower arguments");
1031 // And translate the function!
1032 for (const BasicBlock &BB: F) {
1033 MachineBasicBlock &MBB = getOrCreateBB(BB);
1034 // Set the insertion point of all the following translations to
1035 // the end of this basic block.
1036 CurBuilder.setMBB(MBB);
1038 for (const Instruction &Inst: BB) {
1039 Succeeded &= translate(Inst);
1041 if (TPC->isGlobalISelAbortEnabled())
1042 reportTranslationError(Inst, "unable to translate instruction");
1043 MF->getProperties().set(
1044 MachineFunctionProperties::Property::FailedISel);
1051 finishPendingPhis();
1053 // Now that the MachineFrameInfo has been configured, no further changes to
1054 // the reserved registers are possible.
1055 MRI->freezeReservedRegs(*MF);
1057 // Merge the argument lowering and constants block with its single
1058 // successor, the LLVM-IR entry block. We want the basic block to
1060 assert(EntryBB->succ_size() == 1 &&
1061 "Custom BB used for lowering should have only one successor");
1062 // Get the successor of the current entry block.
1063 MachineBasicBlock &NewEntryBB = **EntryBB->succ_begin();
1064 assert(NewEntryBB.pred_size() == 1 &&
1065 "LLVM-IR entry block has a predecessor!?");
1066 // Move all the instruction from the current entry block to the
1068 NewEntryBB.splice(NewEntryBB.begin(), EntryBB, EntryBB->begin(),
1071 // Update the live-in information for the new entry block.
1072 for (const MachineBasicBlock::RegisterMaskPair &LiveIn : EntryBB->liveins())
1073 NewEntryBB.addLiveIn(LiveIn);
1074 NewEntryBB.sortUniqueLiveIns();
1076 // Get rid of the now empty basic block.
1077 EntryBB->removeSuccessor(&NewEntryBB);
1078 MF->remove(EntryBB);
1079 MF->DeleteMachineBasicBlock(EntryBB);
1081 assert(&MF->front() == &NewEntryBB &&
1082 "New entry wasn't next in the list of basic block!");