diff options
Diffstat (limited to 'lib/CodeGen/SelectionDAG/FastISel.cpp')
-rw-r--r-- | lib/CodeGen/SelectionDAG/FastISel.cpp | 211 |
1 files changed, 154 insertions, 57 deletions
diff --git a/lib/CodeGen/SelectionDAG/FastISel.cpp b/lib/CodeGen/SelectionDAG/FastISel.cpp index 4bf41f28c250..b4c38332691c 100644 --- a/lib/CodeGen/SelectionDAG/FastISel.cpp +++ b/lib/CodeGen/SelectionDAG/FastISel.cpp @@ -1,4 +1,4 @@ -///===-- FastISel.cpp - Implementation of the FastISel class --------------===// +//===-- FastISel.cpp - Implementation of the FastISel class ---------------===// // // The LLVM Compiler Infrastructure // @@ -52,10 +52,11 @@ #include "llvm/Target/TargetInstrInfo.h" #include "llvm/Target/TargetLowering.h" #include "llvm/Target/TargetMachine.h" +#include "llvm/Support/ErrorHandling.h" #include "FunctionLoweringInfo.h" using namespace llvm; -unsigned FastISel::getRegForValue(Value *V) { +unsigned FastISel::getRegForValue(const Value *V) { EVT RealVT = TLI.getValueType(V->getType(), /*AllowUnknown=*/true); // Don't handle non-simple values in FastISel. if (!RealVT.isSimple()) @@ -83,7 +84,16 @@ unsigned FastISel::getRegForValue(Value *V) { if (Reg != 0) return Reg; - if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { + return materializeRegForValue(V, VT); +} + +/// materializeRegForValue - Helper for getRegForVale. This function is +/// called when the value isn't already available in a register and must +/// be materialized with new instructions. +unsigned FastISel::materializeRegForValue(const Value *V, MVT VT) { + unsigned Reg = 0; + + if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) { if (CI->getValue().getActiveBits() <= 64) Reg = FastEmit_i(VT, VT, ISD::Constant, CI->getZExtValue()); } else if (isa<AllocaInst>(V)) { @@ -93,10 +103,12 @@ unsigned FastISel::getRegForValue(Value *V) { // local-CSE'd with actual integer zeros. Reg = getRegForValue(Constant::getNullValue(TD.getIntPtrType(V->getContext()))); - } else if (ConstantFP *CF = dyn_cast<ConstantFP>(V)) { + } else if (const ConstantFP *CF = dyn_cast<ConstantFP>(V)) { + // Try to emit the constant directly. Reg = FastEmit_f(VT, VT, ISD::ConstantFP, CF); if (!Reg) { + // Try to emit the constant by using an integer constant with a cast. const APFloat &Flt = CF->getValueAPF(); EVT IntVT = TLI.getPointerTy(); @@ -114,9 +126,9 @@ unsigned FastISel::getRegForValue(Value *V) { Reg = FastEmit_r(IntVT.getSimpleVT(), VT, ISD::SINT_TO_FP, IntegerReg); } } - } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) { - if (!SelectOperator(CE, CE->getOpcode())) return 0; - Reg = LocalValueMap[CE]; + } else if (const Operator *Op = dyn_cast<Operator>(V)) { + if (!SelectOperator(Op, Op->getOpcode())) return 0; + Reg = LocalValueMap[Op]; } else if (isa<UndefValue>(V)) { Reg = createResultReg(TLI.getRegClassFor(VT)); BuildMI(MBB, DL, TII.get(TargetOpcode::IMPLICIT_DEF), Reg); @@ -134,11 +146,11 @@ unsigned FastISel::getRegForValue(Value *V) { return Reg; } -unsigned FastISel::lookUpRegForValue(Value *V) { +unsigned FastISel::lookUpRegForValue(const Value *V) { // Look up the value to see if we already have a register for it. We // cache values defined by Instructions across blocks, and other values // only locally. This is because Instructions already have the SSA - // def-dominatess-use requirement enforced. + // def-dominates-use requirement enforced. if (ValueMap.count(V)) return ValueMap[V]; return LocalValueMap[V]; @@ -150,7 +162,7 @@ unsigned FastISel::lookUpRegForValue(Value *V) { /// NOTE: This is only necessary because we might select a block that uses /// a value before we select the block that defines the value. It might be /// possible to fix this by selecting blocks in reverse postorder. -unsigned FastISel::UpdateValueMap(Value* I, unsigned Reg) { +unsigned FastISel::UpdateValueMap(const Value *I, unsigned Reg) { if (!isa<Instruction>(I)) { LocalValueMap[I] = Reg; return Reg; @@ -167,7 +179,7 @@ unsigned FastISel::UpdateValueMap(Value* I, unsigned Reg) { return AssignedReg; } -unsigned FastISel::getRegForGEPIndex(Value *Idx) { +unsigned FastISel::getRegForGEPIndex(const Value *Idx) { unsigned IdxN = getRegForValue(Idx); if (IdxN == 0) // Unhandled operand. Halt "fast" selection and bail. @@ -186,7 +198,7 @@ unsigned FastISel::getRegForGEPIndex(Value *Idx) { /// SelectBinaryOp - Select and emit code for a binary operator instruction, /// which has an opcode which directly corresponds to the given ISD opcode. /// -bool FastISel::SelectBinaryOp(User *I, unsigned ISDOpcode) { +bool FastISel::SelectBinaryOp(const User *I, unsigned ISDOpcode) { EVT VT = EVT::getEVT(I->getType(), /*HandleUnknown=*/true); if (VT == MVT::Other || !VT.isSimple()) // Unhandled type. Halt "fast" selection and bail. @@ -252,7 +264,7 @@ bool FastISel::SelectBinaryOp(User *I, unsigned ISDOpcode) { return true; } -bool FastISel::SelectGetElementPtr(User *I) { +bool FastISel::SelectGetElementPtr(const User *I) { unsigned N = getRegForValue(I->getOperand(0)); if (N == 0) // Unhandled operand. Halt "fast" selection and bail. @@ -260,9 +272,9 @@ bool FastISel::SelectGetElementPtr(User *I) { const Type *Ty = I->getOperand(0)->getType(); MVT VT = TLI.getPointerTy(); - for (GetElementPtrInst::op_iterator OI = I->op_begin()+1, E = I->op_end(); - OI != E; ++OI) { - Value *Idx = *OI; + for (GetElementPtrInst::const_op_iterator OI = I->op_begin()+1, + E = I->op_end(); OI != E; ++OI) { + const Value *Idx = *OI; if (const StructType *StTy = dyn_cast<StructType>(Ty)) { unsigned Field = cast<ConstantInt>(Idx)->getZExtValue(); if (Field) { @@ -280,7 +292,7 @@ bool FastISel::SelectGetElementPtr(User *I) { Ty = cast<SequentialType>(Ty)->getElementType(); // If this is a constant subscript, handle it quickly. - if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) { + if (const ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) { if (CI->getZExtValue() == 0) continue; uint64_t Offs = TD.getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue(); @@ -316,51 +328,56 @@ bool FastISel::SelectGetElementPtr(User *I) { return true; } -bool FastISel::SelectCall(User *I) { - Function *F = cast<CallInst>(I)->getCalledFunction(); +bool FastISel::SelectCall(const User *I) { + const Function *F = cast<CallInst>(I)->getCalledFunction(); if (!F) return false; + // Handle selected intrinsic function calls. unsigned IID = F->getIntrinsicID(); switch (IID) { default: break; case Intrinsic::dbg_declare: { - DbgDeclareInst *DI = cast<DbgDeclareInst>(I); + const DbgDeclareInst *DI = cast<DbgDeclareInst>(I); if (!DIDescriptor::ValidDebugInfo(DI->getVariable(), CodeGenOpt::None) || !MF.getMMI().hasDebugInfo()) return true; - Value *Address = DI->getAddress(); + const Value *Address = DI->getAddress(); if (!Address) return true; - AllocaInst *AI = dyn_cast<AllocaInst>(Address); + if (isa<UndefValue>(Address)) + return true; + const AllocaInst *AI = dyn_cast<AllocaInst>(Address); // Don't handle byval struct arguments or VLAs, for example. - if (!AI) break; - DenseMap<const AllocaInst*, int>::iterator SI = - StaticAllocaMap.find(AI); - if (SI == StaticAllocaMap.end()) break; // VLAs. - int FI = SI->second; - if (!DI->getDebugLoc().isUnknown()) - MF.getMMI().setVariableDbgInfo(DI->getVariable(), FI, DI->getDebugLoc()); - - // Building the map above is target independent. Generating DBG_VALUE - // inline is target dependent; do this now. - (void)TargetSelectInstruction(cast<Instruction>(I)); + // Note that if we have a byval struct argument, fast ISel is turned off; + // those are handled in SelectionDAGBuilder. + if (AI) { + DenseMap<const AllocaInst*, int>::iterator SI = + StaticAllocaMap.find(AI); + if (SI == StaticAllocaMap.end()) break; // VLAs. + int FI = SI->second; + if (!DI->getDebugLoc().isUnknown()) + MF.getMMI().setVariableDbgInfo(DI->getVariable(), FI, DI->getDebugLoc()); + } else + // Building the map above is target independent. Generating DBG_VALUE + // inline is target dependent; do this now. + (void)TargetSelectInstruction(cast<Instruction>(I)); return true; } case Intrinsic::dbg_value: { - // This requires target support, but right now X86 is the only Fast target. - DbgValueInst *DI = cast<DbgValueInst>(I); + // This form of DBG_VALUE is target-independent. + const DbgValueInst *DI = cast<DbgValueInst>(I); const TargetInstrDesc &II = TII.get(TargetOpcode::DBG_VALUE); - Value *V = DI->getValue(); + const Value *V = DI->getValue(); if (!V) { // Currently the optimizer can produce this; insert an undef to // help debugging. Probably the optimizer should not do this. BuildMI(MBB, DL, II).addReg(0U).addImm(DI->getOffset()). addMetadata(DI->getVariable()); - } else if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { + } else if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) { BuildMI(MBB, DL, II).addImm(CI->getZExtValue()).addImm(DI->getOffset()). addMetadata(DI->getVariable()); - } else if (ConstantFP *CF = dyn_cast<ConstantFP>(V)) { + } else if (const ConstantFP *CF = dyn_cast<ConstantFP>(V)) { BuildMI(MBB, DL, II).addFPImm(CF).addImm(DI->getOffset()). addMetadata(DI->getVariable()); } else if (unsigned Reg = lookUpRegForValue(V)) { @@ -438,10 +455,12 @@ bool FastISel::SelectCall(User *I) { break; } } + + // An arbitrary call. Bail. return false; } -bool FastISel::SelectCast(User *I, unsigned Opcode) { +bool FastISel::SelectCast(const User *I, unsigned Opcode) { EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType()); EVT DstVT = TLI.getValueType(I->getType()); @@ -493,7 +512,7 @@ bool FastISel::SelectCast(User *I, unsigned Opcode) { return true; } -bool FastISel::SelectBitCast(User *I) { +bool FastISel::SelectBitCast(const User *I) { // If the bitcast doesn't change the type, just use the operand value. if (I->getType() == I->getOperand(0)->getType()) { unsigned Reg = getRegForValue(I->getOperand(0)); @@ -544,15 +563,28 @@ bool FastISel::SelectBitCast(User *I) { } bool -FastISel::SelectInstruction(Instruction *I) { +FastISel::SelectInstruction(const Instruction *I) { + // Just before the terminator instruction, insert instructions to + // feed PHI nodes in successor blocks. + if (isa<TerminatorInst>(I)) + if (!HandlePHINodesInSuccessorBlocks(I->getParent())) + return false; + + DL = I->getDebugLoc(); + // First, try doing target-independent selection. - if (SelectOperator(I, I->getOpcode())) + if (SelectOperator(I, I->getOpcode())) { + DL = DebugLoc(); return true; + } // Next, try calling the target to attempt to handle the instruction. - if (TargetSelectInstruction(I)) + if (TargetSelectInstruction(I)) { + DL = DebugLoc(); return true; + } + DL = DebugLoc(); return false; } @@ -573,7 +605,7 @@ FastISel::FastEmitBranch(MachineBasicBlock *MSucc) { /// SelectFNeg - Emit an FNeg operation. /// bool -FastISel::SelectFNeg(User *I) { +FastISel::SelectFNeg(const User *I) { unsigned OpReg = getRegForValue(BinaryOperator::getFNegArgument(I)); if (OpReg == 0) return false; @@ -614,7 +646,7 @@ FastISel::SelectFNeg(User *I) { } bool -FastISel::SelectOperator(User *I, unsigned Opcode) { +FastISel::SelectOperator(const User *I, unsigned Opcode) { switch (Opcode) { case Instruction::Add: return SelectBinaryOp(I, ISD::ADD); @@ -660,10 +692,10 @@ FastISel::SelectOperator(User *I, unsigned Opcode) { return SelectGetElementPtr(I); case Instruction::Br: { - BranchInst *BI = cast<BranchInst>(I); + const BranchInst *BI = cast<BranchInst>(I); if (BI->isUnconditional()) { - BasicBlock *LLVMSucc = BI->getSuccessor(0); + const BasicBlock *LLVMSucc = BI->getSuccessor(0); MachineBasicBlock *MSucc = MBBMap[LLVMSucc]; FastEmitBranch(MSucc); return true; @@ -678,10 +710,6 @@ FastISel::SelectOperator(User *I, unsigned Opcode) { // Nothing to emit. return true; - case Instruction::PHI: - // PHI nodes are already emitted. - return true; - case Instruction::Alloca: // FunctionLowering has the static-sized case covered. if (StaticAllocaMap.count(cast<AllocaInst>(I))) @@ -721,6 +749,9 @@ FastISel::SelectOperator(User *I, unsigned Opcode) { return true; } + case Instruction::PHI: + llvm_unreachable("FastISel shouldn't visit PHI nodes!"); + default: // Unhandled instruction. Halt "fast" selection and bail. return false; @@ -730,15 +761,17 @@ FastISel::SelectOperator(User *I, unsigned Opcode) { FastISel::FastISel(MachineFunction &mf, DenseMap<const Value *, unsigned> &vm, DenseMap<const BasicBlock *, MachineBasicBlock *> &bm, - DenseMap<const AllocaInst *, int> &am + DenseMap<const AllocaInst *, int> &am, + std::vector<std::pair<MachineInstr*, unsigned> > &pn #ifndef NDEBUG - , SmallSet<Instruction*, 8> &cil + , SmallSet<const Instruction *, 8> &cil #endif ) : MBB(0), ValueMap(vm), MBBMap(bm), StaticAllocaMap(am), + PHINodesToUpdate(pn), #ifndef NDEBUG CatchInfoLost(cil), #endif @@ -775,7 +808,7 @@ unsigned FastISel::FastEmit_i(MVT, MVT, unsigned, uint64_t /*Imm*/) { } unsigned FastISel::FastEmit_f(MVT, MVT, - unsigned, ConstantFP * /*FPImm*/) { + unsigned, const ConstantFP * /*FPImm*/) { return 0; } @@ -787,7 +820,7 @@ unsigned FastISel::FastEmit_ri(MVT, MVT, unsigned FastISel::FastEmit_rf(MVT, MVT, unsigned, unsigned /*Op0*/, - ConstantFP * /*FPImm*/) { + const ConstantFP * /*FPImm*/) { return 0; } @@ -820,7 +853,7 @@ unsigned FastISel::FastEmit_ri_(MVT VT, unsigned Opcode, /// FastEmit_rf. If that fails, it materializes the immediate into a register /// and try FastEmit_rr instead. unsigned FastISel::FastEmit_rf_(MVT VT, unsigned Opcode, - unsigned Op0, ConstantFP *FPImm, + unsigned Op0, const ConstantFP *FPImm, MVT ImmType) { // First check if immediate type is legal. If not, we can't use the rf form. unsigned ResultReg = FastEmit_rf(VT, VT, Opcode, Op0, FPImm); @@ -930,7 +963,7 @@ unsigned FastISel::FastEmitInst_ri(unsigned MachineInstOpcode, unsigned FastISel::FastEmitInst_rf(unsigned MachineInstOpcode, const TargetRegisterClass *RC, - unsigned Op0, ConstantFP *FPImm) { + unsigned Op0, const ConstantFP *FPImm) { unsigned ResultReg = createResultReg(RC); const TargetInstrDesc &II = TII.get(MachineInstOpcode); @@ -1006,3 +1039,67 @@ unsigned FastISel::FastEmitInst_extractsubreg(MVT RetVT, unsigned FastISel::FastEmitZExtFromI1(MVT VT, unsigned Op) { return FastEmit_ri(VT, VT, ISD::AND, Op, 1); } + +/// HandlePHINodesInSuccessorBlocks - Handle PHI nodes in successor blocks. +/// Emit code to ensure constants are copied into registers when needed. +/// Remember the virtual registers that need to be added to the Machine PHI +/// nodes as input. We cannot just directly add them, because expansion +/// might result in multiple MBB's for one BB. As such, the start of the +/// BB might correspond to a different MBB than the end. +bool FastISel::HandlePHINodesInSuccessorBlocks(const BasicBlock *LLVMBB) { + const TerminatorInst *TI = LLVMBB->getTerminator(); + + SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled; + unsigned OrigNumPHINodesToUpdate = PHINodesToUpdate.size(); + + // Check successor nodes' PHI nodes that expect a constant to be available + // from this block. + for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) { + const BasicBlock *SuccBB = TI->getSuccessor(succ); + if (!isa<PHINode>(SuccBB->begin())) continue; + MachineBasicBlock *SuccMBB = MBBMap[SuccBB]; + + // If this terminator has multiple identical successors (common for + // switches), only handle each succ once. + if (!SuccsHandled.insert(SuccMBB)) continue; + + MachineBasicBlock::iterator MBBI = SuccMBB->begin(); + + // At this point we know that there is a 1-1 correspondence between LLVM PHI + // nodes and Machine PHI nodes, but the incoming operands have not been + // emitted yet. + for (BasicBlock::const_iterator I = SuccBB->begin(); + const PHINode *PN = dyn_cast<PHINode>(I); ++I) { + // Ignore dead phi's. + if (PN->use_empty()) continue; + + // Only handle legal types. Two interesting things to note here. First, + // by bailing out early, we may leave behind some dead instructions, + // since SelectionDAG's HandlePHINodesInSuccessorBlocks will insert its + // own moves. Second, this check is necessary becuase FastISel doesn't + // use CreateRegForValue to create registers, so it always creates + // exactly one register for each non-void instruction. + EVT VT = TLI.getValueType(PN->getType(), /*AllowUnknown=*/true); + if (VT == MVT::Other || !TLI.isTypeLegal(VT)) { + // Promote MVT::i1. + if (VT == MVT::i1) + VT = TLI.getTypeToTransformTo(LLVMBB->getContext(), VT); + else { + PHINodesToUpdate.resize(OrigNumPHINodesToUpdate); + return false; + } + } + + const Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB); + + unsigned Reg = getRegForValue(PHIOp); + if (Reg == 0) { + PHINodesToUpdate.resize(OrigNumPHINodesToUpdate); + return false; + } + PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg)); + } + } + + return true; +} |