X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FTarget%2FX86%2FX86FastISel.cpp;h=928dea91b4fdcd7f1d727596630de5b4e4c0853b;hb=623d2e618f4e672c47edff9ec63ed6d733ac81d3;hp=82f5d3a50deb2ac71ca13d79674e34af6baf9011;hpb=ef41ff618f2537539b538e6c7bf471c753391f92;p=oota-llvm.git diff --git a/lib/Target/X86/X86FastISel.cpp b/lib/Target/X86/X86FastISel.cpp index 82f5d3a50de..928dea91b4f 100644 --- a/lib/Target/X86/X86FastISel.cpp +++ b/lib/Target/X86/X86FastISel.cpp @@ -14,23 +14,25 @@ //===----------------------------------------------------------------------===// #include "X86.h" -#include "X86InstrBuilder.h" +#include "X86CallingConv.h" #include "X86ISelLowering.h" +#include "X86InstrBuilder.h" #include "X86RegisterInfo.h" #include "X86Subtarget.h" #include "X86TargetMachine.h" -#include "llvm/CallingConv.h" -#include "llvm/DerivedTypes.h" -#include "llvm/GlobalVariable.h" -#include "llvm/Instructions.h" -#include "llvm/IntrinsicInst.h" -#include "llvm/Operator.h" #include "llvm/CodeGen/Analysis.h" #include "llvm/CodeGen/FastISel.h" #include "llvm/CodeGen/FunctionLoweringInfo.h" #include "llvm/CodeGen/MachineConstantPool.h" #include "llvm/CodeGen/MachineFrameInfo.h" #include "llvm/CodeGen/MachineRegisterInfo.h" +#include "llvm/IR/CallingConv.h" +#include "llvm/IR/DerivedTypes.h" +#include "llvm/IR/GlobalAlias.h" +#include "llvm/IR/GlobalVariable.h" +#include "llvm/IR/Instructions.h" +#include "llvm/IR/IntrinsicInst.h" +#include "llvm/IR/Operator.h" #include "llvm/Support/CallSite.h" #include "llvm/Support/ErrorHandling.h" #include "llvm/Support/GetElementPtrTypeIterator.h" @@ -44,10 +46,6 @@ class X86FastISel : public FastISel { /// make the right decision when generating code for different targets. const X86Subtarget *Subtarget; - /// StackPtr - Register used as the stack pointer. - /// - unsigned StackPtr; - /// X86ScalarSSEf32, X86ScalarSSEf64 - Select between SSE or x87 /// floating point ops. /// When SSE is available, use it for f32 operations. @@ -56,21 +54,24 @@ class X86FastISel : public FastISel { bool X86ScalarSSEf32; public: - explicit X86FastISel(FunctionLoweringInfo &funcInfo) : FastISel(funcInfo) { + explicit X86FastISel(FunctionLoweringInfo &funcInfo, + const TargetLibraryInfo *libInfo) + : FastISel(funcInfo, libInfo) { Subtarget = &TM.getSubtarget(); - StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP; X86ScalarSSEf64 = Subtarget->hasSSE2(); X86ScalarSSEf32 = Subtarget->hasSSE1(); } virtual bool TargetSelectInstruction(const Instruction *I); - /// TryToFoldLoad - The specified machine instr operand is a vreg, and that + /// \brief The specified machine instr operand is a vreg, and that /// vreg is being provided by the specified load instruction. If possible, /// try to fold the load as an operand to the instruction, returning true if /// possible. - virtual bool TryToFoldLoad(MachineInstr *MI, unsigned OpNo, - const LoadInst *LI); + virtual bool tryToFoldLoadIntoMI(MachineInstr *MI, unsigned OpNo, + const LoadInst *LI); + + virtual bool FastLowerArguments(); #include "X86GenFastISel.inc" @@ -79,8 +80,10 @@ private: bool X86FastEmitLoad(EVT VT, const X86AddressMode &AM, unsigned &RR); - bool X86FastEmitStore(EVT VT, const Value *Val, const X86AddressMode &AM); - bool X86FastEmitStore(EVT VT, unsigned Val, const X86AddressMode &AM); + bool X86FastEmitStore(EVT VT, const Value *Val, const X86AddressMode &AM, + bool Aligned = false); + bool X86FastEmitStore(EVT VT, unsigned ValReg, const X86AddressMode &AM, + bool Aligned = false); bool X86FastEmitExtend(ISD::NodeType Opc, EVT DstVT, unsigned Src, EVT SrcVT, unsigned &ResultReg); @@ -102,6 +105,8 @@ private: bool X86SelectShift(const Instruction *I); + bool X86SelectDivRem(const Instruction *I); + bool X86SelectSelect(const Instruction *I); bool X86SelectTrunc(const Instruction *I); @@ -121,6 +126,8 @@ private: return static_cast(&TM); } + bool handleConstantAddresses(const Value *V, X86AddressMode &AM); + unsigned TargetMaterializeConstant(const Constant *C); unsigned TargetMaterializeAlloca(const AllocaInst *C); @@ -134,7 +141,7 @@ private: (VT == MVT::f32 && X86ScalarSSEf32); // f32 is when SSE1 } - bool isTypeLegal(const Type *Ty, MVT &VT, bool AllowI1 = false); + bool isTypeLegal(Type *Ty, MVT &VT, bool AllowI1 = false); bool IsMemcpySmall(uint64_t Len); @@ -144,7 +151,7 @@ private: } // end anonymous namespace. -bool X86FastISel::isTypeLegal(const Type *Ty, MVT &VT, bool AllowI1) { +bool X86FastISel::isTypeLegal(Type *Ty, MVT &VT, bool AllowI1) { EVT evt = TLI.getValueType(Ty, /*HandleUnknown=*/true); if (evt == MVT::Other || !evt.isSimple()) // Unhandled type. Halt "fast" selection and bail. @@ -154,9 +161,9 @@ bool X86FastISel::isTypeLegal(const Type *Ty, MVT &VT, bool AllowI1) { // For now, require SSE/SSE2 for performing floating-point operations, // since x87 requires additional work. if (VT == MVT::f64 && !X86ScalarSSEf64) - return false; + return false; if (VT == MVT::f32 && !X86ScalarSSEf32) - return false; + return false; // Similarly, no f80 support yet. if (VT == MVT::f80) return false; @@ -182,37 +189,37 @@ bool X86FastISel::X86FastEmitLoad(EVT VT, const X86AddressMode &AM, case MVT::i1: case MVT::i8: Opc = X86::MOV8rm; - RC = X86::GR8RegisterClass; + RC = &X86::GR8RegClass; break; case MVT::i16: Opc = X86::MOV16rm; - RC = X86::GR16RegisterClass; + RC = &X86::GR16RegClass; break; case MVT::i32: Opc = X86::MOV32rm; - RC = X86::GR32RegisterClass; + RC = &X86::GR32RegClass; break; case MVT::i64: // Must be in x86-64 mode. Opc = X86::MOV64rm; - RC = X86::GR64RegisterClass; + RC = &X86::GR64RegClass; break; case MVT::f32: - if (Subtarget->hasSSE1()) { - Opc = X86::MOVSSrm; - RC = X86::FR32RegisterClass; + if (X86ScalarSSEf32) { + Opc = Subtarget->hasAVX() ? X86::VMOVSSrm : X86::MOVSSrm; + RC = &X86::FR32RegClass; } else { Opc = X86::LD_Fp32m; - RC = X86::RFP32RegisterClass; + RC = &X86::RFP32RegClass; } break; case MVT::f64: - if (Subtarget->hasSSE2()) { - Opc = X86::MOVSDrm; - RC = X86::FR64RegisterClass; + if (X86ScalarSSEf64) { + Opc = Subtarget->hasAVX() ? X86::VMOVSDrm : X86::MOVSDrm; + RC = &X86::FR64RegClass; } else { Opc = X86::LD_Fp64m; - RC = X86::RFP64RegisterClass; + RC = &X86::RFP64RegClass; } break; case MVT::f80: @@ -231,7 +238,8 @@ bool X86FastISel::X86FastEmitLoad(EVT VT, const X86AddressMode &AM, /// and a displacement offset, or a GlobalAddress, /// i.e. V. Return true if it is possible. bool -X86FastISel::X86FastEmitStore(EVT VT, unsigned Val, const X86AddressMode &AM) { +X86FastISel::X86FastEmitStore(EVT VT, unsigned ValReg, + const X86AddressMode &AM, bool Aligned) { // Get opcode and regclass of the output for the given store instruction. unsigned Opc = 0; switch (VT.getSimpleVT().SimpleTy) { @@ -239,10 +247,10 @@ X86FastISel::X86FastEmitStore(EVT VT, unsigned Val, const X86AddressMode &AM) { default: return false; case MVT::i1: { // Mask out all but lowest bit. - unsigned AndResult = createResultReg(X86::GR8RegisterClass); + unsigned AndResult = createResultReg(&X86::GR8RegClass); BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, - TII.get(X86::AND8ri), AndResult).addReg(Val).addImm(1); - Val = AndResult; + TII.get(X86::AND8ri), AndResult).addReg(ValReg).addImm(1); + ValReg = AndResult; } // FALLTHROUGH, handling i1 as i8. case MVT::i8: Opc = X86::MOV8mr; break; @@ -250,20 +258,43 @@ X86FastISel::X86FastEmitStore(EVT VT, unsigned Val, const X86AddressMode &AM) { case MVT::i32: Opc = X86::MOV32mr; break; case MVT::i64: Opc = X86::MOV64mr; break; // Must be in x86-64 mode. case MVT::f32: - Opc = Subtarget->hasSSE1() ? X86::MOVSSmr : X86::ST_Fp32m; + Opc = X86ScalarSSEf32 ? + (Subtarget->hasAVX() ? X86::VMOVSSmr : X86::MOVSSmr) : X86::ST_Fp32m; break; case MVT::f64: - Opc = Subtarget->hasSSE2() ? X86::MOVSDmr : X86::ST_Fp64m; + Opc = X86ScalarSSEf64 ? + (Subtarget->hasAVX() ? X86::VMOVSDmr : X86::MOVSDmr) : X86::ST_Fp64m; + break; + case MVT::v4f32: + if (Aligned) + Opc = Subtarget->hasAVX() ? X86::VMOVAPSmr : X86::MOVAPSmr; + else + Opc = Subtarget->hasAVX() ? X86::VMOVUPSmr : X86::MOVUPSmr; + break; + case MVT::v2f64: + if (Aligned) + Opc = Subtarget->hasAVX() ? X86::VMOVAPDmr : X86::MOVAPDmr; + else + Opc = Subtarget->hasAVX() ? X86::VMOVUPDmr : X86::MOVUPDmr; + break; + case MVT::v4i32: + case MVT::v2i64: + case MVT::v8i16: + case MVT::v16i8: + if (Aligned) + Opc = Subtarget->hasAVX() ? X86::VMOVDQAmr : X86::MOVDQAmr; + else + Opc = Subtarget->hasAVX() ? X86::VMOVDQUmr : X86::MOVDQUmr; break; } addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, - DL, TII.get(Opc)), AM).addReg(Val); + DL, TII.get(Opc)), AM).addReg(ValReg); return true; } bool X86FastISel::X86FastEmitStore(EVT VT, const Value *Val, - const X86AddressMode &AM) { + const X86AddressMode &AM, bool Aligned) { // Handle 'null' like i32/i64 0. if (isa(Val)) Val = Constant::getNullValue(TD.getIntPtrType(Val->getContext())); @@ -280,7 +311,7 @@ bool X86FastISel::X86FastEmitStore(EVT VT, const Value *Val, case MVT::i32: Opc = X86::MOV32mi; break; case MVT::i64: // Must be a 32-bit sign extended value. - if ((int)CI->getSExtValue() == CI->getSExtValue()) + if (isInt<32>(CI->getSExtValue())) Opc = X86::MOV64mi32; break; } @@ -298,7 +329,7 @@ bool X86FastISel::X86FastEmitStore(EVT VT, const Value *Val, if (ValReg == 0) return false; - return X86FastEmitStore(VT, ValReg, AM); + return X86FastEmitStore(VT, ValReg, AM, Aligned); } /// X86FastEmitExtend - Emit a machine instruction to extend a value Src of @@ -309,17 +340,133 @@ bool X86FastISel::X86FastEmitExtend(ISD::NodeType Opc, EVT DstVT, unsigned &ResultReg) { unsigned RR = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(), Opc, Src, /*TODO: Kill=*/false); - - if (RR != 0) { - ResultReg = RR; - return true; - } else + if (RR == 0) return false; + + ResultReg = RR; + return true; +} + +bool X86FastISel::handleConstantAddresses(const Value *V, X86AddressMode &AM) { + // Handle constant address. + if (const GlobalValue *GV = dyn_cast(V)) { + // Can't handle alternate code models yet. + if (TM.getCodeModel() != CodeModel::Small) + return false; + + // Can't handle TLS yet. + if (const GlobalVariable *GVar = dyn_cast(GV)) + if (GVar->isThreadLocal()) + return false; + + // Can't handle TLS yet, part 2 (this is slightly crazy, but this is how + // it works...). + if (const GlobalAlias *GA = dyn_cast(GV)) + if (const GlobalVariable *GVar = + dyn_cast_or_null(GA->resolveAliasedGlobal(false))) + if (GVar->isThreadLocal()) + return false; + + // RIP-relative addresses can't have additional register operands, so if + // we've already folded stuff into the addressing mode, just force the + // global value into its own register, which we can use as the basereg. + if (!Subtarget->isPICStyleRIPRel() || + (AM.Base.Reg == 0 && AM.IndexReg == 0)) { + // Okay, we've committed to selecting this global. Set up the address. + AM.GV = GV; + + // Allow the subtarget to classify the global. + unsigned char GVFlags = Subtarget->ClassifyGlobalReference(GV, TM); + + // If this reference is relative to the pic base, set it now. + if (isGlobalRelativeToPICBase(GVFlags)) { + // FIXME: How do we know Base.Reg is free?? + AM.Base.Reg = getInstrInfo()->getGlobalBaseReg(FuncInfo.MF); + } + + // Unless the ABI requires an extra load, return a direct reference to + // the global. + if (!isGlobalStubReference(GVFlags)) { + if (Subtarget->isPICStyleRIPRel()) { + // Use rip-relative addressing if we can. Above we verified that the + // base and index registers are unused. + assert(AM.Base.Reg == 0 && AM.IndexReg == 0); + AM.Base.Reg = X86::RIP; + } + AM.GVOpFlags = GVFlags; + return true; + } + + // Ok, we need to do a load from a stub. If we've already loaded from + // this stub, reuse the loaded pointer, otherwise emit the load now. + DenseMap::iterator I = LocalValueMap.find(V); + unsigned LoadReg; + if (I != LocalValueMap.end() && I->second != 0) { + LoadReg = I->second; + } else { + // Issue load from stub. + unsigned Opc = 0; + const TargetRegisterClass *RC = NULL; + X86AddressMode StubAM; + StubAM.Base.Reg = AM.Base.Reg; + StubAM.GV = GV; + StubAM.GVOpFlags = GVFlags; + + // Prepare for inserting code in the local-value area. + SavePoint SaveInsertPt = enterLocalValueArea(); + + if (TLI.getPointerTy() == MVT::i64) { + Opc = X86::MOV64rm; + RC = &X86::GR64RegClass; + + if (Subtarget->isPICStyleRIPRel()) + StubAM.Base.Reg = X86::RIP; + } else { + Opc = X86::MOV32rm; + RC = &X86::GR32RegClass; + } + + LoadReg = createResultReg(RC); + MachineInstrBuilder LoadMI = + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), LoadReg); + addFullAddress(LoadMI, StubAM); + + // Ok, back to normal mode. + leaveLocalValueArea(SaveInsertPt); + + // Prevent loading GV stub multiple times in same MBB. + LocalValueMap[V] = LoadReg; + } + + // Now construct the final address. Note that the Disp, Scale, + // and Index values may already be set here. + AM.Base.Reg = LoadReg; + AM.GV = 0; + return true; + } + } + + // If all else fails, try to materialize the value in a register. + if (!AM.GV || !Subtarget->isPICStyleRIPRel()) { + if (AM.Base.Reg == 0) { + AM.Base.Reg = getRegForValue(V); + return AM.Base.Reg != 0; + } + if (AM.IndexReg == 0) { + assert(AM.Scale == 1 && "Scale with no index!"); + AM.IndexReg = getRegForValue(V); + return AM.IndexReg != 0; + } + } + + return false; } /// X86SelectAddress - Attempt to fill in an address from the given value. /// bool X86FastISel::X86SelectAddress(const Value *V, X86AddressMode &AM) { + SmallVector GEPs; +redo_gep: const User *U = NULL; unsigned Opcode = Instruction::UserOp1; if (const Instruction *I = dyn_cast(V)) { @@ -336,7 +483,7 @@ bool X86FastISel::X86SelectAddress(const Value *V, X86AddressMode &AM) { U = C; } - if (const PointerType *Ty = dyn_cast(V->getType())) + if (PointerType *Ty = dyn_cast(V->getType())) if (Ty->getAddressSpace() > 255) // Fast instruction selection doesn't support the special // address spaces. @@ -399,7 +546,7 @@ bool X86FastISel::X86SelectAddress(const Value *V, X86AddressMode &AM) { for (User::const_op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i, ++GTI) { const Value *Op = *i; - if (const StructType *STy = dyn_cast(*GTI)) { + if (StructType *STy = dyn_cast(*GTI)) { const StructLayout *SL = TD.getStructLayout(STy); Disp += SL->getElementOffset(cast(Op)->getZExtValue()); continue; @@ -442,130 +589,43 @@ bool X86FastISel::X86SelectAddress(const Value *V, X86AddressMode &AM) { goto unsupported_gep; } } + // Check for displacement overflow. if (!isInt<32>(Disp)) break; - // Ok, the GEP indices were covered by constant-offset and scaled-index - // addressing. Update the address state and move on to examining the base. + AM.IndexReg = IndexReg; AM.Scale = Scale; AM.Disp = (uint32_t)Disp; - if (X86SelectAddress(U->getOperand(0), AM)) + GEPs.push_back(V); + + if (const GetElementPtrInst *GEP = + dyn_cast(U->getOperand(0))) { + // Ok, the GEP indices were covered by constant-offset and scaled-index + // addressing. Update the address state and move on to examining the base. + V = GEP; + goto redo_gep; + } else if (X86SelectAddress(U->getOperand(0), AM)) { return true; + } // If we couldn't merge the gep value into this addr mode, revert back to // our address and just match the value instead of completely failing. AM = SavedAM; - break; - unsupported_gep: - // Ok, the GEP indices weren't all covered. - break; - } - } - - // Handle constant address. - if (const GlobalValue *GV = dyn_cast(V)) { - // Can't handle alternate code models or TLS yet. - if (TM.getCodeModel() != CodeModel::Small) - return false; - - if (const GlobalVariable *GVar = dyn_cast(GV)) - if (GVar->isThreadLocal()) - return false; - - // RIP-relative addresses can't have additional register operands, so if - // we've already folded stuff into the addressing mode, just force the - // global value into its own register, which we can use as the basereg. - if (!Subtarget->isPICStyleRIPRel() || - (AM.Base.Reg == 0 && AM.IndexReg == 0)) { - // Okay, we've committed to selecting this global. Set up the address. - AM.GV = GV; - // Allow the subtarget to classify the global. - unsigned char GVFlags = Subtarget->ClassifyGlobalReference(GV, TM); - - // If this reference is relative to the pic base, set it now. - if (isGlobalRelativeToPICBase(GVFlags)) { - // FIXME: How do we know Base.Reg is free?? - AM.Base.Reg = getInstrInfo()->getGlobalBaseReg(FuncInfo.MF); - } - - // Unless the ABI requires an extra load, return a direct reference to - // the global. - if (!isGlobalStubReference(GVFlags)) { - if (Subtarget->isPICStyleRIPRel()) { - // Use rip-relative addressing if we can. Above we verified that the - // base and index registers are unused. - assert(AM.Base.Reg == 0 && AM.IndexReg == 0); - AM.Base.Reg = X86::RIP; - } - AM.GVOpFlags = GVFlags; + for (SmallVectorImpl::reverse_iterator + I = GEPs.rbegin(), E = GEPs.rend(); I != E; ++I) + if (handleConstantAddresses(*I, AM)) return true; - } - - // Ok, we need to do a load from a stub. If we've already loaded from - // this stub, reuse the loaded pointer, otherwise emit the load now. - DenseMap::iterator I = LocalValueMap.find(V); - unsigned LoadReg; - if (I != LocalValueMap.end() && I->second != 0) { - LoadReg = I->second; - } else { - // Issue load from stub. - unsigned Opc = 0; - const TargetRegisterClass *RC = NULL; - X86AddressMode StubAM; - StubAM.Base.Reg = AM.Base.Reg; - StubAM.GV = GV; - StubAM.GVOpFlags = GVFlags; - - // Prepare for inserting code in the local-value area. - SavePoint SaveInsertPt = enterLocalValueArea(); - - if (TLI.getPointerTy() == MVT::i64) { - Opc = X86::MOV64rm; - RC = X86::GR64RegisterClass; - - if (Subtarget->isPICStyleRIPRel()) - StubAM.Base.Reg = X86::RIP; - } else { - Opc = X86::MOV32rm; - RC = X86::GR32RegisterClass; - } - LoadReg = createResultReg(RC); - MachineInstrBuilder LoadMI = - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), LoadReg); - addFullAddress(LoadMI, StubAM); - - // Ok, back to normal mode. - leaveLocalValueArea(SaveInsertPt); - - // Prevent loading GV stub multiple times in same MBB. - LocalValueMap[V] = LoadReg; - } - - // Now construct the final address. Note that the Disp, Scale, - // and Index values may already be set here. - AM.Base.Reg = LoadReg; - AM.GV = 0; - return true; - } + return false; + unsupported_gep: + // Ok, the GEP indices weren't all covered. + break; } - - // If all else fails, try to materialize the value in a register. - if (!AM.GV || !Subtarget->isPICStyleRIPRel()) { - if (AM.Base.Reg == 0) { - AM.Base.Reg = getRegForValue(V); - return AM.Base.Reg != 0; - } - if (AM.IndexReg == 0) { - assert(AM.Scale == 1 && "Scale with no index!"); - AM.IndexReg = getRegForValue(V); - return AM.IndexReg != 0; - } } - return false; + return handleConstantAddresses(V, AM); } /// X86SelectCallAddress - Attempt to fill in an address from the given value. @@ -573,9 +633,35 @@ bool X86FastISel::X86SelectAddress(const Value *V, X86AddressMode &AM) { bool X86FastISel::X86SelectCallAddress(const Value *V, X86AddressMode &AM) { const User *U = NULL; unsigned Opcode = Instruction::UserOp1; - if (const Instruction *I = dyn_cast(V)) { + const Instruction *I = dyn_cast(V); + // Record if the value is defined in the same basic block. + // + // This information is crucial to know whether or not folding an + // operand is valid. + // Indeed, FastISel generates or reuses a virtual register for all + // operands of all instructions it selects. Obviously, the definition and + // its uses must use the same virtual register otherwise the produced + // code is incorrect. + // Before instruction selection, FunctionLoweringInfo::set sets the virtual + // registers for values that are alive across basic blocks. This ensures + // that the values are consistently set between across basic block, even + // if different instruction selection mechanisms are used (e.g., a mix of + // SDISel and FastISel). + // For values local to a basic block, the instruction selection process + // generates these virtual registers with whatever method is appropriate + // for its needs. In particular, FastISel and SDISel do not share the way + // local virtual registers are set. + // Therefore, this is impossible (or at least unsafe) to share values + // between basic blocks unless they use the same instruction selection + // method, which is not guarantee for X86. + // Moreover, things like hasOneUse could not be used accurately, if we + // allow to reference values across basic blocks whereas they are not + // alive across basic blocks initially. + bool InMBB = true; + if (I) { Opcode = I->getOpcode(); U = I; + InMBB = I->getParent() == FuncInfo.MBB->getBasicBlock(); } else if (const ConstantExpr *C = dyn_cast(V)) { Opcode = C->getOpcode(); U = C; @@ -584,18 +670,22 @@ bool X86FastISel::X86SelectCallAddress(const Value *V, X86AddressMode &AM) { switch (Opcode) { default: break; case Instruction::BitCast: - // Look past bitcasts. - return X86SelectCallAddress(U->getOperand(0), AM); + // Look past bitcasts if its operand is in the same BB. + if (InMBB) + return X86SelectCallAddress(U->getOperand(0), AM); + break; case Instruction::IntToPtr: - // Look past no-op inttoptrs. - if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy()) + // Look past no-op inttoptrs if its operand is in the same BB. + if (InMBB && + TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy()) return X86SelectCallAddress(U->getOperand(0), AM); break; case Instruction::PtrToInt: - // Look past no-op ptrtoints. - if (TLI.getValueType(U->getType()) == TLI.getPointerTy()) + // Look past no-op ptrtoints if its operand is in the same BB. + if (InMBB && + TLI.getValueType(U->getType()) == TLI.getPointerTy()) return X86SelectCallAddress(U->getOperand(0), AM); break; } @@ -658,6 +748,16 @@ bool X86FastISel::X86SelectCallAddress(const Value *V, X86AddressMode &AM) { /// X86SelectStore - Select and emit code to implement store instructions. bool X86FastISel::X86SelectStore(const Instruction *I) { + // Atomic stores need special handling. + const StoreInst *S = cast(I); + + if (S->isAtomic()) + return false; + + unsigned SABIAlignment = + TD.getABITypeAlignment(S->getValueOperand()->getType()); + bool Aligned = S->getAlignment() == 0 || S->getAlignment() >= SABIAlignment; + MVT VT; if (!isTypeLegal(I->getOperand(0)->getType(), VT, /*AllowI1=*/true)) return false; @@ -666,13 +766,15 @@ bool X86FastISel::X86SelectStore(const Instruction *I) { if (!X86SelectAddress(I->getOperand(1), AM)) return false; - return X86FastEmitStore(VT, I->getOperand(0), AM); + return X86FastEmitStore(VT, I->getOperand(0), AM, Aligned); } /// X86SelectRet - Select and emit code to implement ret instructions. bool X86FastISel::X86SelectRet(const Instruction *I) { const ReturnInst *Ret = cast(I); const Function &F = *I->getParent()->getParent(); + const X86MachineFunctionInfo *X86MFInfo = + FuncInfo.MF->getInfo(); if (!FuncInfo.CanLowerReturn) return false; @@ -680,35 +782,37 @@ bool X86FastISel::X86SelectRet(const Instruction *I) { CallingConv::ID CC = F.getCallingConv(); if (CC != CallingConv::C && CC != CallingConv::Fast && - CC != CallingConv::X86_FastCall) + CC != CallingConv::X86_FastCall && + CC != CallingConv::X86_64_SysV) return false; - if (Subtarget->isTargetWin64()) + if (Subtarget->isCallingConvWin64(CC)) return false; // Don't handle popping bytes on return for now. - if (FuncInfo.MF->getInfo() - ->getBytesToPopOnReturn() != 0) - return 0; + if (X86MFInfo->getBytesToPopOnReturn() != 0) + return false; // fastcc with -tailcallopt is intended to provide a guaranteed // tail call optimization. Fastisel doesn't know how to do that. - if (CC == CallingConv::Fast && GuaranteedTailCallOpt) + if (CC == CallingConv::Fast && TM.Options.GuaranteedTailCallOpt) return false; // Let SDISel handle vararg functions. if (F.isVarArg()) return false; + // Build a list of return value registers. + SmallVector RetRegs; + if (Ret->getNumOperands() > 0) { SmallVector Outs; - GetReturnInfo(F.getReturnType(), F.getAttributes().getRetAttributes(), - Outs, TLI); + GetReturnInfo(F.getReturnType(), F.getAttributes(), Outs, TLI); // Analyze operands of the call, assigning locations to each operand. SmallVector ValLocs; CCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, TM, ValLocs, - I->getContext()); + I->getContext()); CCInfo.AnalyzeReturn(Outs, RetCC_X86); const Value *RV = Ret->getOperand(0); @@ -768,18 +872,40 @@ bool X86FastISel::X86SelectRet(const Instruction *I) { BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY), DstReg).addReg(SrcReg); - // Mark the register as live out of the function. - MRI.addLiveOut(VA.getLocReg()); + // Add register to return instruction. + RetRegs.push_back(VA.getLocReg()); + } + + // The x86-64 ABI for returning structs by value requires that we copy + // the sret argument into %rax for the return. We saved the argument into + // a virtual register in the entry block, so now we copy the value out + // and into %rax. We also do the same with %eax for Win32. + if (F.hasStructRetAttr() && + (Subtarget->is64Bit() || Subtarget->isTargetWindows())) { + unsigned Reg = X86MFInfo->getSRetReturnReg(); + assert(Reg && + "SRetReturnReg should have been set in LowerFormalArguments()!"); + unsigned RetReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX; + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY), + RetReg).addReg(Reg); + RetRegs.push_back(RetReg); } // Now emit the RET. - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::RET)); + MachineInstrBuilder MIB = + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::RET)); + for (unsigned i = 0, e = RetRegs.size(); i != e; ++i) + MIB.addReg(RetRegs[i], RegState::Implicit); return true; } /// X86SelectLoad - Select and emit code to implement load instructions. /// bool X86FastISel::X86SelectLoad(const Instruction *I) { + // Atomic loads need special handling. + if (cast(I)->isAtomic()) + return false; + MVT VT; if (!isTypeLegal(I->getType(), VT, /*AllowI1=*/true)) return false; @@ -797,14 +923,20 @@ bool X86FastISel::X86SelectLoad(const Instruction *I) { } static unsigned X86ChooseCmpOpcode(EVT VT, const X86Subtarget *Subtarget) { + bool HasAVX = Subtarget->hasAVX(); + bool X86ScalarSSEf32 = Subtarget->hasSSE1(); + bool X86ScalarSSEf64 = Subtarget->hasSSE2(); + switch (VT.getSimpleVT().SimpleTy) { default: return 0; case MVT::i8: return X86::CMP8rr; case MVT::i16: return X86::CMP16rr; case MVT::i32: return X86::CMP32rr; case MVT::i64: return X86::CMP64rr; - case MVT::f32: return Subtarget->hasSSE1() ? X86::UCOMISSrr : 0; - case MVT::f64: return Subtarget->hasSSE2() ? X86::UCOMISDrr : 0; + case MVT::f32: + return X86ScalarSSEf32 ? (HasAVX ? X86::VUCOMISSrr : X86::UCOMISSrr) : 0; + case MVT::f64: + return X86ScalarSSEf64 ? (HasAVX ? X86::VUCOMISDrr : X86::UCOMISDrr) : 0; } } @@ -939,10 +1071,6 @@ bool X86FastISel::X86SelectCmp(const Instruction *I) { } bool X86FastISel::X86SelectZExt(const Instruction *I) { - // Handle zero-extension from i1 to i8, which is common. - if (!I->getOperand(0)->getType()->isIntegerTy(1)) - return false; - EVT DstVT = TLI.getValueType(I->getType()); if (!TLI.isTypeLegal(DstVT)) return false; @@ -951,12 +1079,37 @@ bool X86FastISel::X86SelectZExt(const Instruction *I) { if (ResultReg == 0) return false; - // Set the high bits to zero. - ResultReg = FastEmitZExtFromI1(MVT::i8, ResultReg, /*TODO: Kill=*/false); - if (ResultReg == 0) - return false; + // Handle zero-extension from i1 to i8, which is common. + MVT SrcVT = TLI.getSimpleValueType(I->getOperand(0)->getType()); + if (SrcVT.SimpleTy == MVT::i1) { + // Set the high bits to zero. + ResultReg = FastEmitZExtFromI1(MVT::i8, ResultReg, /*TODO: Kill=*/false); + SrcVT = MVT::i8; + + if (ResultReg == 0) + return false; + } - if (DstVT != MVT::i8) { + if (DstVT == MVT::i64) { + // Handle extension to 64-bits via sub-register shenanigans. + unsigned MovInst; + + switch (SrcVT.SimpleTy) { + case MVT::i8: MovInst = X86::MOVZX32rr8; break; + case MVT::i16: MovInst = X86::MOVZX32rr16; break; + case MVT::i32: MovInst = X86::MOV32rr; break; + default: llvm_unreachable("Unexpected zext to i64 source type"); + } + + unsigned Result32 = createResultReg(&X86::GR32RegClass); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(MovInst), Result32) + .addReg(ResultReg); + + ResultReg = createResultReg(&X86::GR64RegClass); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::SUBREG_TO_REG), + ResultReg) + .addImm(0).addReg(Result32).addImm(X86::sub_32bit); + } else if (DstVT != MVT::i8) { ResultReg = FastEmit_r(MVT::i8, DstVT.getSimpleVT(), ISD::ZERO_EXTEND, ResultReg, /*Kill=*/true); if (ResultReg == 0) @@ -1166,6 +1319,170 @@ bool X86FastISel::X86SelectShift(const Instruction *I) { return true; } +bool X86FastISel::X86SelectDivRem(const Instruction *I) { + const static unsigned NumTypes = 4; // i8, i16, i32, i64 + const static unsigned NumOps = 4; // SDiv, SRem, UDiv, URem + const static bool S = true; // IsSigned + const static bool U = false; // !IsSigned + const static unsigned Copy = TargetOpcode::COPY; + // For the X86 DIV/IDIV instruction, in most cases the dividend + // (numerator) must be in a specific register pair highreg:lowreg, + // producing the quotient in lowreg and the remainder in highreg. + // For most data types, to set up the instruction, the dividend is + // copied into lowreg, and lowreg is sign-extended or zero-extended + // into highreg. The exception is i8, where the dividend is defined + // as a single register rather than a register pair, and we + // therefore directly sign-extend or zero-extend the dividend into + // lowreg, instead of copying, and ignore the highreg. + const static struct DivRemEntry { + // The following portion depends only on the data type. + const TargetRegisterClass *RC; + unsigned LowInReg; // low part of the register pair + unsigned HighInReg; // high part of the register pair + // The following portion depends on both the data type and the operation. + struct DivRemResult { + unsigned OpDivRem; // The specific DIV/IDIV opcode to use. + unsigned OpSignExtend; // Opcode for sign-extending lowreg into + // highreg, or copying a zero into highreg. + unsigned OpCopy; // Opcode for copying dividend into lowreg, or + // zero/sign-extending into lowreg for i8. + unsigned DivRemResultReg; // Register containing the desired result. + bool IsOpSigned; // Whether to use signed or unsigned form. + } ResultTable[NumOps]; + } OpTable[NumTypes] = { + { &X86::GR8RegClass, X86::AX, 0, { + { X86::IDIV8r, 0, X86::MOVSX16rr8, X86::AL, S }, // SDiv + { X86::IDIV8r, 0, X86::MOVSX16rr8, X86::AH, S }, // SRem + { X86::DIV8r, 0, X86::MOVZX16rr8, X86::AL, U }, // UDiv + { X86::DIV8r, 0, X86::MOVZX16rr8, X86::AH, U }, // URem + } + }, // i8 + { &X86::GR16RegClass, X86::AX, X86::DX, { + { X86::IDIV16r, X86::CWD, Copy, X86::AX, S }, // SDiv + { X86::IDIV16r, X86::CWD, Copy, X86::DX, S }, // SRem + { X86::DIV16r, X86::MOV32r0, Copy, X86::AX, U }, // UDiv + { X86::DIV16r, X86::MOV32r0, Copy, X86::DX, U }, // URem + } + }, // i16 + { &X86::GR32RegClass, X86::EAX, X86::EDX, { + { X86::IDIV32r, X86::CDQ, Copy, X86::EAX, S }, // SDiv + { X86::IDIV32r, X86::CDQ, Copy, X86::EDX, S }, // SRem + { X86::DIV32r, X86::MOV32r0, Copy, X86::EAX, U }, // UDiv + { X86::DIV32r, X86::MOV32r0, Copy, X86::EDX, U }, // URem + } + }, // i32 + { &X86::GR64RegClass, X86::RAX, X86::RDX, { + { X86::IDIV64r, X86::CQO, Copy, X86::RAX, S }, // SDiv + { X86::IDIV64r, X86::CQO, Copy, X86::RDX, S }, // SRem + { X86::DIV64r, X86::MOV32r0, Copy, X86::RAX, U }, // UDiv + { X86::DIV64r, X86::MOV32r0, Copy, X86::RDX, U }, // URem + } + }, // i64 + }; + + MVT VT; + if (!isTypeLegal(I->getType(), VT)) + return false; + + unsigned TypeIndex, OpIndex; + switch (VT.SimpleTy) { + default: return false; + case MVT::i8: TypeIndex = 0; break; + case MVT::i16: TypeIndex = 1; break; + case MVT::i32: TypeIndex = 2; break; + case MVT::i64: TypeIndex = 3; + if (!Subtarget->is64Bit()) + return false; + break; + } + + switch (I->getOpcode()) { + default: llvm_unreachable("Unexpected div/rem opcode"); + case Instruction::SDiv: OpIndex = 0; break; + case Instruction::SRem: OpIndex = 1; break; + case Instruction::UDiv: OpIndex = 2; break; + case Instruction::URem: OpIndex = 3; break; + } + + const DivRemEntry &TypeEntry = OpTable[TypeIndex]; + const DivRemEntry::DivRemResult &OpEntry = TypeEntry.ResultTable[OpIndex]; + unsigned Op0Reg = getRegForValue(I->getOperand(0)); + if (Op0Reg == 0) + return false; + unsigned Op1Reg = getRegForValue(I->getOperand(1)); + if (Op1Reg == 0) + return false; + + // Move op0 into low-order input register. + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, + TII.get(OpEntry.OpCopy), TypeEntry.LowInReg).addReg(Op0Reg); + // Zero-extend or sign-extend into high-order input register. + if (OpEntry.OpSignExtend) { + if (OpEntry.IsOpSigned) + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, + TII.get(OpEntry.OpSignExtend)); + else { + unsigned Zero32 = createResultReg(&X86::GR32RegClass); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, + TII.get(X86::MOV32r0), Zero32); + + // Copy the zero into the appropriate sub/super/identical physical + // register. Unfortunately the operations needed are not uniform enough to + // fit neatly into the table above. + if (VT.SimpleTy == MVT::i16) { + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, + TII.get(Copy), TypeEntry.HighInReg) + .addReg(Zero32, 0, X86::sub_16bit); + } else if (VT.SimpleTy == MVT::i32) { + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, + TII.get(Copy), TypeEntry.HighInReg) + .addReg(Zero32); + } else if (VT.SimpleTy == MVT::i64) { + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, + TII.get(TargetOpcode::SUBREG_TO_REG), TypeEntry.HighInReg) + .addImm(0).addReg(Zero32).addImm(X86::sub_32bit); + } + } + } + // Generate the DIV/IDIV instruction. + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, + TII.get(OpEntry.OpDivRem)).addReg(Op1Reg); + // For i8 remainder, we can't reference AH directly, as we'll end + // up with bogus copies like %R9B = COPY %AH. Reference AX + // instead to prevent AH references in a REX instruction. + // + // The current assumption of the fast register allocator is that isel + // won't generate explicit references to the GPR8_NOREX registers. If + // the allocator and/or the backend get enhanced to be more robust in + // that regard, this can be, and should be, removed. + unsigned ResultReg = 0; + if ((I->getOpcode() == Instruction::SRem || + I->getOpcode() == Instruction::URem) && + OpEntry.DivRemResultReg == X86::AH && Subtarget->is64Bit()) { + unsigned SourceSuperReg = createResultReg(&X86::GR16RegClass); + unsigned ResultSuperReg = createResultReg(&X86::GR16RegClass); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, + TII.get(Copy), SourceSuperReg).addReg(X86::AX); + + // Shift AX right by 8 bits instead of using AH. + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::SHR16ri), + ResultSuperReg).addReg(SourceSuperReg).addImm(8); + + // Now reference the 8-bit subreg of the result. + ResultReg = FastEmitInst_extractsubreg(MVT::i8, ResultSuperReg, + /*Kill=*/true, X86::sub_8bit); + } + // Copy the result out of the physreg if we haven't already. + if (!ResultReg) { + ResultReg = createResultReg(TypeEntry.RC); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Copy), ResultReg) + .addReg(OpEntry.DivRemResultReg); + } + UpdateValueMap(I, ResultReg); + + return true; +} + bool X86FastISel::X86SelectSelect(const Instruction *I) { MVT VT; if (!isTypeLegal(I->getType(), VT)) @@ -1207,13 +1524,13 @@ bool X86FastISel::X86SelectSelect(const Instruction *I) { bool X86FastISel::X86SelectFPExt(const Instruction *I) { // fpext from float to double. - if (Subtarget->hasSSE2() && + if (X86ScalarSSEf64 && I->getType()->isDoubleTy()) { const Value *V = I->getOperand(0); if (V->getType()->isFloatTy()) { unsigned OpReg = getRegForValue(V); if (OpReg == 0) return false; - unsigned ResultReg = createResultReg(X86::FR64RegisterClass); + unsigned ResultReg = createResultReg(&X86::FR64RegClass); BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::CVTSS2SDrr), ResultReg) .addReg(OpReg); @@ -1226,13 +1543,13 @@ bool X86FastISel::X86SelectFPExt(const Instruction *I) { } bool X86FastISel::X86SelectFPTrunc(const Instruction *I) { - if (Subtarget->hasSSE2()) { + if (X86ScalarSSEf64) { if (I->getType()->isFloatTy()) { const Value *V = I->getOperand(0); if (V->getType()->isDoubleTy()) { unsigned OpReg = getRegForValue(V); if (OpReg == 0) return false; - unsigned ResultReg = createResultReg(X86::FR32RegisterClass); + unsigned ResultReg = createResultReg(&X86::FR32RegClass); BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::CVTSD2SSrr), ResultReg) .addReg(OpReg); @@ -1269,8 +1586,9 @@ bool X86FastISel::X86SelectTrunc(const Instruction *I) { if (!Subtarget->is64Bit()) { // If we're on x86-32; we can't extract an i8 from a general register. // First issue a copy to GR16_ABCD or GR32_ABCD. - const TargetRegisterClass *CopyRC = (SrcVT == MVT::i16) - ? X86::GR16_ABCDRegisterClass : X86::GR32_ABCDRegisterClass; + const TargetRegisterClass *CopyRC = (SrcVT == MVT::i16) ? + (const TargetRegisterClass*)&X86::GR16_ABCDRegClass : + (const TargetRegisterClass*)&X86::GR32_ABCDRegClass; unsigned CopyReg = createResultReg(CopyRC); BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY), CopyReg).addReg(InputReg); @@ -1311,7 +1629,6 @@ bool X86FastISel::TryEmitSmallMemcpy(X86AddressMode DestAM, else if (Len >= 2) VT = MVT::i16; else { - assert(Len == 1); VT = MVT::i8; } @@ -1365,6 +1682,9 @@ bool X86FastISel::X86VisitIntrinsicCall(const IntrinsicInst &I) { case Intrinsic::memset: { const MemSetInst &MSI = cast(I); + if (MSI.isVolatile()) + return false; + unsigned SizeWidth = Subtarget->is64Bit() ? 64 : 32; if (!MSI.getLength()->getType()->isIntegerTy(SizeWidth)) return false; @@ -1375,7 +1695,7 @@ bool X86FastISel::X86VisitIntrinsicCall(const IntrinsicInst &I) { return DoSelectCall(&I, "memset"); } case Intrinsic::stackprotector: { - // Emit code inline code to store the stack guard onto the stack. + // Emit code to store the stack guard onto the stack. EVT PtrTy = TLI.getPointerTy(); const Value *Op1 = I.getArgOperand(0); // The guard's value. @@ -1393,7 +1713,7 @@ bool X86FastISel::X86VisitIntrinsicCall(const IntrinsicInst &I) { assert(DI->getAddress() && "Null address should be checked earlier!"); if (!X86SelectAddress(DI->getAddress(), AM)) return false; - const TargetInstrDesc &II = TII.get(TargetOpcode::DBG_VALUE); + const MCInstrDesc &II = TII.get(TargetOpcode::DBG_VALUE); // FIXME may need to add RegState::Debug to any registers produced, // although ESP/EBP should be the only ones at the moment. addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II), AM). @@ -1411,7 +1731,7 @@ bool X86FastISel::X86VisitIntrinsicCall(const IntrinsicInst &I) { // Replace "add with overflow" intrinsics with an "add" instruction followed // by a seto/setc instruction. const Function *Callee = I.getCalledFunction(); - const Type *RetTy = + Type *RetTy = cast(Callee->getReturnType())->getTypeAtIndex(unsigned(0)); MVT VT; @@ -1436,7 +1756,7 @@ bool X86FastISel::X86VisitIntrinsicCall(const IntrinsicInst &I) { return false; // The call to CreateRegs builds two sequential registers, to store the - // both the the returned values. + // both the returned values. unsigned ResultReg = FuncInfo.CreateRegs(I.getType()); BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(OpC), ResultReg) .addReg(Reg1).addReg(Reg2); @@ -1452,6 +1772,79 @@ bool X86FastISel::X86VisitIntrinsicCall(const IntrinsicInst &I) { } } +bool X86FastISel::FastLowerArguments() { + if (!FuncInfo.CanLowerReturn) + return false; + + const Function *F = FuncInfo.Fn; + if (F->isVarArg()) + return false; + + CallingConv::ID CC = F->getCallingConv(); + if (CC != CallingConv::C) + return false; + + if (Subtarget->isCallingConvWin64(CC)) + return false; + + if (!Subtarget->is64Bit()) + return false; + + // Only handle simple cases. i.e. Up to 6 i32/i64 scalar arguments. + unsigned Idx = 1; + for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end(); + I != E; ++I, ++Idx) { + if (Idx > 6) + return false; + + if (F->getAttributes().hasAttribute(Idx, Attribute::ByVal) || + F->getAttributes().hasAttribute(Idx, Attribute::InReg) || + F->getAttributes().hasAttribute(Idx, Attribute::StructRet) || + F->getAttributes().hasAttribute(Idx, Attribute::Nest)) + return false; + + Type *ArgTy = I->getType(); + if (ArgTy->isStructTy() || ArgTy->isArrayTy() || ArgTy->isVectorTy()) + return false; + + EVT ArgVT = TLI.getValueType(ArgTy); + if (!ArgVT.isSimple()) return false; + switch (ArgVT.getSimpleVT().SimpleTy) { + case MVT::i32: + case MVT::i64: + break; + default: + return false; + } + } + + static const uint16_t GPR32ArgRegs[] = { + X86::EDI, X86::ESI, X86::EDX, X86::ECX, X86::R8D, X86::R9D + }; + static const uint16_t GPR64ArgRegs[] = { + X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8 , X86::R9 + }; + + Idx = 0; + const TargetRegisterClass *RC32 = TLI.getRegClassFor(MVT::i32); + const TargetRegisterClass *RC64 = TLI.getRegClassFor(MVT::i64); + for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end(); + I != E; ++I, ++Idx) { + bool is32Bit = TLI.getValueType(I->getType()) == MVT::i32; + const TargetRegisterClass *RC = is32Bit ? RC32 : RC64; + unsigned SrcReg = is32Bit ? GPR32ArgRegs[Idx] : GPR64ArgRegs[Idx]; + unsigned DstReg = FuncInfo.MF->addLiveIn(SrcReg, RC); + // FIXME: Unfortunately it's necessary to emit a copy from the livein copy. + // Without this, EmitLiveInCopies may eliminate the livein if its only + // use is a bitcast (which isn't turned into an instruction). + unsigned ResultReg = createResultReg(RC); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY), + ResultReg).addReg(DstReg, getKillRegState(true)); + UpdateValueMap(I, ResultReg); + } + return true; +} + bool X86FastISel::X86SelectCall(const Instruction *I) { const CallInst *CI = cast(I); const Value *Callee = CI->getCalledValue(); @@ -1464,9 +1857,29 @@ bool X86FastISel::X86SelectCall(const Instruction *I) { if (const IntrinsicInst *II = dyn_cast(CI)) return X86VisitIntrinsicCall(*II); + // Allow SelectionDAG isel to handle tail calls. + if (cast(I)->isTailCall()) + return false; + return DoSelectCall(I, 0); } +static unsigned computeBytesPoppedByCallee(const X86Subtarget &Subtarget, + const ImmutableCallSite &CS) { + if (Subtarget.is64Bit()) + return 0; + if (Subtarget.isTargetWindows()) + return 0; + CallingConv::ID CC = CS.getCallingConv(); + if (CC == CallingConv::Fast || CC == CallingConv::GHC) + return 0; + if (!CS.paramHasAttr(1, Attribute::StructRet)) + return 0; + if (CS.paramHasAttr(1, Attribute::InReg)) + return 0; + return 4; +} + // Select either a call, or an llvm.memcpy/memmove/memset intrinsic bool X86FastISel::DoSelectCall(const Instruction *I, const char *MemIntName) { const CallInst *CI = cast(I); @@ -1475,37 +1888,37 @@ bool X86FastISel::DoSelectCall(const Instruction *I, const char *MemIntName) { // Handle only C and fastcc calling conventions for now. ImmutableCallSite CS(CI); CallingConv::ID CC = CS.getCallingConv(); + bool isWin64 = Subtarget->isCallingConvWin64(CC); if (CC != CallingConv::C && CC != CallingConv::Fast && - CC != CallingConv::X86_FastCall) + CC != CallingConv::X86_FastCall && CC != CallingConv::X86_64_Win64 && + CC != CallingConv::X86_64_SysV) return false; // fastcc with -tailcallopt is intended to provide a guaranteed // tail call optimization. Fastisel doesn't know how to do that. - if (CC == CallingConv::Fast && GuaranteedTailCallOpt) + if (CC == CallingConv::Fast && TM.Options.GuaranteedTailCallOpt) return false; - const PointerType *PT = cast(CS.getCalledValue()->getType()); - const FunctionType *FTy = cast(PT->getElementType()); + PointerType *PT = cast(CS.getCalledValue()->getType()); + FunctionType *FTy = cast(PT->getElementType()); bool isVarArg = FTy->isVarArg(); // Don't know how to handle Win64 varargs yet. Nothing special needed for // x86-32. Special handling for x86-64 is implemented. - if (isVarArg && Subtarget->isTargetWin64()) + if (isVarArg && isWin64) return false; // Fast-isel doesn't know about callee-pop yet. if (X86::isCalleePop(CC, Subtarget->is64Bit(), isVarArg, - GuaranteedTailCallOpt)) + TM.Options.GuaranteedTailCallOpt)) return false; // Check whether the function can return without sret-demotion. SmallVector Outs; - SmallVector Offsets; - GetReturnInfo(I->getType(), CS.getAttributes().getRetAttributes(), - Outs, TLI, &Offsets); + GetReturnInfo(I->getType(), CS.getAttributes(), Outs, TLI); bool CanLowerReturn = TLI.CanLowerReturn(CS.getCallingConv(), - *FuncInfo.MF, FTy->isVarArg(), - Outs, FTy->getContext()); + *FuncInfo.MF, FTy->isVarArg(), + Outs, FTy->getContext()); if (!CanLowerReturn) return false; @@ -1528,10 +1941,11 @@ bool X86FastISel::DoSelectCall(const Instruction *I, const char *MemIntName) { SmallVector Args; SmallVector ArgVTs; SmallVector ArgFlags; - Args.reserve(CS.arg_size()); - ArgVals.reserve(CS.arg_size()); - ArgVTs.reserve(CS.arg_size()); - ArgFlags.reserve(CS.arg_size()); + unsigned arg_size = CS.arg_size(); + Args.reserve(arg_size); + ArgVals.reserve(arg_size); + ArgVTs.reserve(arg_size); + ArgFlags.reserve(arg_size); for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end(); i != e; ++i) { // If we're lowering a mem intrinsic instead of a regular call, skip the @@ -1547,8 +1961,8 @@ bool X86FastISel::DoSelectCall(const Instruction *I, const char *MemIntName) { Flags.setZExt(); if (CS.paramHasAttr(AttrInd, Attribute::ByVal)) { - const PointerType *Ty = cast(ArgVal->getType()); - const Type *ElementTy = Ty->getElementType(); + PointerType *Ty = cast(ArgVal->getType()); + Type *ElementTy = Ty->getElementType(); unsigned FrameSize = TD.getTypeAllocSize(ElementTy); unsigned FrameAlign = CS.getParamAlignment(AttrInd); if (!FrameAlign) @@ -1600,7 +2014,7 @@ bool X86FastISel::DoSelectCall(const Instruction *I, const char *MemIntName) { if (ArgReg == 0) return false; - const Type *ArgTy = ArgVal->getType(); + Type *ArgTy = ArgVal->getType(); MVT ArgVT; if (!isTypeLegal(ArgTy, ArgVT)) return false; @@ -1618,10 +2032,10 @@ bool X86FastISel::DoSelectCall(const Instruction *I, const char *MemIntName) { // Analyze operands of the call, assigning locations to each operand. SmallVector ArgLocs; CCState CCInfo(CC, isVarArg, *FuncInfo.MF, TM, ArgLocs, - I->getParent()->getContext()); + I->getParent()->getContext()); // Allocate shadow area for Win64 - if (Subtarget->isTargetWin64()) + if (isWin64) CCInfo.AllocateStack(32, 8); CCInfo.AnalyzeCallOperands(ArgVTs, ArgFlags, CC_X86); @@ -1630,7 +2044,7 @@ bool X86FastISel::DoSelectCall(const Instruction *I, const char *MemIntName) { unsigned NumBytes = CCInfo.getNextStackOffset(); // Issue CALLSEQ_START - unsigned AdjStackDown = TM.getRegisterInfo()->getCallFrameSetupOpcode(); + unsigned AdjStackDown = TII.getCallFrameSetupOpcode(); BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(AdjStackDown)) .addImm(NumBytes); @@ -1644,7 +2058,6 @@ bool X86FastISel::DoSelectCall(const Instruction *I, const char *MemIntName) { // Promote the value if needed. switch (VA.getLocInfo()) { - default: llvm_unreachable("Unknown loc info!"); case CCValAssign::Full: break; case CCValAssign::SExt: { assert(VA.getLocVT().isInteger() && !VA.getLocVT().isVector() && @@ -1688,6 +2101,14 @@ bool X86FastISel::DoSelectCall(const Instruction *I, const char *MemIntName) { ArgVT = VA.getLocVT(); break; } + case CCValAssign::VExt: + // VExt has not been implemented, so this should be impossible to reach + // for now. However, fallback to Selection DAG isel once implemented. + return false; + case CCValAssign::Indirect: + // FIXME: Indirect doesn't need extending, but fast-isel doesn't fully + // support this. + return false; } if (VA.isRegLoc()) { @@ -1697,7 +2118,9 @@ bool X86FastISel::DoSelectCall(const Instruction *I, const char *MemIntName) { } else { unsigned LocMemOffset = VA.getLocMemOffset(); X86AddressMode AM; - AM.Base.Reg = StackPtr; + const X86RegisterInfo *RegInfo = static_cast( + getTargetMachine()->getRegisterInfo()); + AM.Base.Reg = RegInfo->getStackRegister(); AM.Disp = LocMemOffset; const Value *ArgVal = ArgVals[VA.getValNo()]; ISD::ArgFlagsTy Flags = ArgFlags[VA.getValNo()]; @@ -1709,11 +2132,13 @@ bool X86FastISel::DoSelectCall(const Instruction *I, const char *MemIntName) { assert(Res && "memcpy length already checked!"); (void)Res; } else if (isa(ArgVal) || isa(ArgVal)) { // If this is a really simple value, emit this with the Value* version - //of X86FastEmitStore. If it isn't simple, we don't want to do this, + // of X86FastEmitStore. If it isn't simple, we don't want to do this, // as it can cause us to reevaluate the argument. - X86FastEmitStore(ArgVT, ArgVal, AM); + if (!X86FastEmitStore(ArgVT, ArgVal, AM)) + return false; } else { - X86FastEmitStore(ArgVT, Arg, AM); + if (!X86FastEmitStore(ArgVT, Arg, AM)) + return false; } } } @@ -1726,9 +2151,9 @@ bool X86FastISel::DoSelectCall(const Instruction *I, const char *MemIntName) { X86::EBX).addReg(Base); } - if (Subtarget->is64Bit() && isVarArg && !Subtarget->isTargetWin64()) { + if (Subtarget->is64Bit() && isVarArg && !isWin64) { // Count the number of XMM registers allocated. - static const unsigned XMMArgRegs[] = { + static const uint16_t XMMArgRegs[] = { X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3, X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7 }; @@ -1742,9 +2167,7 @@ bool X86FastISel::DoSelectCall(const Instruction *I, const char *MemIntName) { if (CalleeOp) { // Register-indirect call. unsigned CallOpc; - if (Subtarget->isTargetWin64()) - CallOpc = X86::WINCALL64r; - else if (Subtarget->is64Bit()) + if (Subtarget->is64Bit()) CallOpc = X86::CALL64r; else CallOpc = X86::CALL32r; @@ -1755,9 +2178,7 @@ bool X86FastISel::DoSelectCall(const Instruction *I, const char *MemIntName) { // Direct call. assert(GV && "Not a direct call"); unsigned CallOpc; - if (Subtarget->isTargetWin64()) - CallOpc = X86::WINCALL64pcrel32; - else if (Subtarget->is64Bit()) + if (Subtarget->is64Bit()) CallOpc = X86::CALL64pcrel32; else CallOpc = X86::CALLpcrel32; @@ -1791,22 +2212,24 @@ bool X86FastISel::DoSelectCall(const Instruction *I, const char *MemIntName) { MIB.addGlobalAddress(GV, 0, OpFlags); } + // Add a register mask with the call-preserved registers. + // Proper defs for return values will be added by setPhysRegsDeadExcept(). + MIB.addRegMask(TRI.getCallPreservedMask(CS.getCallingConv())); + // Add an implicit use GOT pointer in EBX. if (Subtarget->isPICStyleGOT()) - MIB.addReg(X86::EBX); + MIB.addReg(X86::EBX, RegState::Implicit); - if (Subtarget->is64Bit() && isVarArg && !Subtarget->isTargetWin64()) - MIB.addReg(X86::AL); + if (Subtarget->is64Bit() && isVarArg && !isWin64) + MIB.addReg(X86::AL, RegState::Implicit); // Add implicit physical register uses to the call. for (unsigned i = 0, e = RegArgs.size(); i != e; ++i) - MIB.addReg(RegArgs[i]); + MIB.addReg(RegArgs[i], RegState::Implicit); // Issue CALLSEQ_END - unsigned AdjStackUp = TM.getRegisterInfo()->getCallFrameDestroyOpcode(); - unsigned NumBytesCallee = 0; - if (!Subtarget->is64Bit() && CS.paramHasAttr(1, Attribute::StructRet)) - NumBytesCallee = 4; + unsigned AdjStackUp = TII.getCallFrameDestroyOpcode(); + const unsigned NumBytesCallee = computeBytesPoppedByCallee(*Subtarget, CS); BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(AdjStackUp)) .addImm(NumBytes).addImm(NumBytesCallee); @@ -1817,11 +2240,11 @@ bool X86FastISel::DoSelectCall(const Instruction *I, const char *MemIntName) { ComputeValueVTs(TLI, I->getType(), RetTys); for (unsigned i = 0, e = RetTys.size(); i != e; ++i) { EVT VT = RetTys[i]; - EVT RegisterVT = TLI.getRegisterType(I->getParent()->getContext(), VT); + MVT RegisterVT = TLI.getRegisterType(I->getParent()->getContext(), VT); unsigned NumRegs = TLI.getNumRegisters(I->getParent()->getContext(), VT); for (unsigned j = 0; j != NumRegs; ++j) { ISD::InputArg MyFlags; - MyFlags.VT = RegisterVT.getSimpleVT(); + MyFlags.VT = RegisterVT; MyFlags.Used = !CS.getInstruction()->use_empty(); if (CS.paramHasAttr(0, Attribute::SExt)) MyFlags.Flags.setSExt(); @@ -1837,7 +2260,7 @@ bool X86FastISel::DoSelectCall(const Instruction *I, const char *MemIntName) { SmallVector UsedRegs; SmallVector RVLocs; CCState CCRetInfo(CC, false, *FuncInfo.MF, TM, RVLocs, - I->getParent()->getContext()); + I->getParent()->getContext()); unsigned ResultReg = FuncInfo.CreateRegs(I->getType()); CCRetInfo.AnalyzeCallResult(Ins, RetCC_X86); for (unsigned i = 0; i != RVLocs.size(); ++i) { @@ -1848,16 +2271,19 @@ bool X86FastISel::DoSelectCall(const Instruction *I, const char *MemIntName) { // stack, but where we prefer to use the value in xmm registers, copy it // out as F80 and use a truncate to move it from fp stack reg to xmm reg. if ((RVLocs[i].getLocReg() == X86::ST0 || - RVLocs[i].getLocReg() == X86::ST1) && - isScalarFPTypeInSSEReg(RVLocs[0].getValVT())) { - CopyVT = MVT::f80; - CopyReg = createResultReg(X86::RFP80RegisterClass); + RVLocs[i].getLocReg() == X86::ST1)) { + if (isScalarFPTypeInSSEReg(RVLocs[i].getValVT())) { + CopyVT = MVT::f80; + CopyReg = createResultReg(&X86::RFP80RegClass); + } + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::FpPOP_RETVAL), + CopyReg); + } else { + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY), + CopyReg).addReg(RVLocs[i].getLocReg()); + UsedRegs.push_back(RVLocs[i].getLocReg()); } - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY), - CopyReg).addReg(RVLocs[i].getLocReg()); - UsedRegs.push_back(RVLocs[i].getLocReg()); - if (CopyVT != RVLocs[i].getValVT()) { // Round the F80 the right size, which also moves to the appropriate xmm // register. This is accomplished by storing the F80 value in memory and @@ -1908,6 +2334,11 @@ X86FastISel::TargetSelectInstruction(const Instruction *I) { case Instruction::AShr: case Instruction::Shl: return X86SelectShift(I); + case Instruction::SDiv: + case Instruction::UDiv: + case Instruction::SRem: + case Instruction::URem: + return X86SelectDivRem(I); case Instruction::Select: return X86SelectSelect(I); case Instruction::Trunc: @@ -1937,51 +2368,55 @@ X86FastISel::TargetSelectInstruction(const Instruction *I) { unsigned X86FastISel::TargetMaterializeConstant(const Constant *C) { MVT VT; if (!isTypeLegal(C->getType(), VT)) - return false; + return 0; + + // Can't handle alternate code models yet. + if (TM.getCodeModel() != CodeModel::Small) + return 0; // Get opcode and regclass of the output for the given load instruction. unsigned Opc = 0; const TargetRegisterClass *RC = NULL; switch (VT.SimpleTy) { - default: return false; + default: return 0; case MVT::i8: Opc = X86::MOV8rm; - RC = X86::GR8RegisterClass; + RC = &X86::GR8RegClass; break; case MVT::i16: Opc = X86::MOV16rm; - RC = X86::GR16RegisterClass; + RC = &X86::GR16RegClass; break; case MVT::i32: Opc = X86::MOV32rm; - RC = X86::GR32RegisterClass; + RC = &X86::GR32RegClass; break; case MVT::i64: // Must be in x86-64 mode. Opc = X86::MOV64rm; - RC = X86::GR64RegisterClass; + RC = &X86::GR64RegClass; break; case MVT::f32: - if (Subtarget->hasSSE1()) { - Opc = X86::MOVSSrm; - RC = X86::FR32RegisterClass; + if (X86ScalarSSEf32) { + Opc = Subtarget->hasAVX() ? X86::VMOVSSrm : X86::MOVSSrm; + RC = &X86::FR32RegClass; } else { Opc = X86::LD_Fp32m; - RC = X86::RFP32RegisterClass; + RC = &X86::RFP32RegClass; } break; case MVT::f64: - if (Subtarget->hasSSE2()) { - Opc = X86::MOVSDrm; - RC = X86::FR64RegisterClass; + if (X86ScalarSSEf64) { + Opc = Subtarget->hasAVX() ? X86::VMOVSDrm : X86::MOVSDrm; + RC = &X86::FR64RegClass; } else { Opc = X86::LD_Fp64m; - RC = X86::RFP64RegisterClass; + RC = &X86::RFP64RegClass; } break; case MVT::f80: // No f80 support yet. - return false; + return 0; } // Materialize addresses with LEA instructions. @@ -2049,7 +2484,7 @@ unsigned X86FastISel::TargetMaterializeAlloca(const AllocaInst *C) { if (!X86SelectAddress(C, AM)) return 0; unsigned Opc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r; - TargetRegisterClass* RC = TLI.getRegClassFor(TLI.getPointerTy()); + const TargetRegisterClass* RC = TLI.getRegClassFor(TLI.getPointerTy()); unsigned ResultReg = createResultReg(RC); addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), ResultReg), AM); @@ -2059,34 +2494,34 @@ unsigned X86FastISel::TargetMaterializeAlloca(const AllocaInst *C) { unsigned X86FastISel::TargetMaterializeFloatZero(const ConstantFP *CF) { MVT VT; if (!isTypeLegal(CF->getType(), VT)) - return false; + return 0; // Get opcode and regclass for the given zero. unsigned Opc = 0; const TargetRegisterClass *RC = NULL; switch (VT.SimpleTy) { - default: return false; - case MVT::f32: - if (Subtarget->hasSSE1()) { - Opc = X86::FsFLD0SS; - RC = X86::FR32RegisterClass; - } else { - Opc = X86::LD_Fp032; - RC = X86::RFP32RegisterClass; - } - break; - case MVT::f64: - if (Subtarget->hasSSE2()) { - Opc = X86::FsFLD0SD; - RC = X86::FR64RegisterClass; - } else { - Opc = X86::LD_Fp064; - RC = X86::RFP64RegisterClass; - } - break; - case MVT::f80: - // No f80 support yet. - return false; + default: return 0; + case MVT::f32: + if (X86ScalarSSEf32) { + Opc = X86::FsFLD0SS; + RC = &X86::FR32RegClass; + } else { + Opc = X86::LD_Fp032; + RC = &X86::RFP32RegClass; + } + break; + case MVT::f64: + if (X86ScalarSSEf64) { + Opc = X86::FsFLD0SD; + RC = &X86::FR64RegClass; + } else { + Opc = X86::LD_Fp064; + RC = &X86::RFP64RegClass; + } + break; + case MVT::f80: + // No f80 support yet. + return 0; } unsigned ResultReg = createResultReg(RC); @@ -2095,17 +2530,13 @@ unsigned X86FastISel::TargetMaterializeFloatZero(const ConstantFP *CF) { } -/// TryToFoldLoad - The specified machine instr operand is a vreg, and that -/// vreg is being provided by the specified load instruction. If possible, -/// try to fold the load as an operand to the instruction, returning true if -/// possible. -bool X86FastISel::TryToFoldLoad(MachineInstr *MI, unsigned OpNo, - const LoadInst *LI) { +bool X86FastISel::tryToFoldLoadIntoMI(MachineInstr *MI, unsigned OpNo, + const LoadInst *LI) { X86AddressMode AM; if (!X86SelectAddress(LI->getOperand(0), AM)) return false; - X86InstrInfo &XII = (X86InstrInfo&)TII; + const X86InstrInfo &XII = (const X86InstrInfo&)TII; unsigned Size = TD.getTypeAllocSize(LI->getType()); unsigned Alignment = LI->getAlignment(); @@ -2124,7 +2555,8 @@ bool X86FastISel::TryToFoldLoad(MachineInstr *MI, unsigned OpNo, namespace llvm { - llvm::FastISel *X86::createFastISel(FunctionLoweringInfo &funcInfo) { - return new X86FastISel(funcInfo); + FastISel *X86::createFastISel(FunctionLoweringInfo &funcInfo, + const TargetLibraryInfo *libInfo) { + return new X86FastISel(funcInfo, libInfo); } }