X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FTarget%2FX86%2FX86FastISel.cpp;h=f42465d14ce18d1da189475a38aaaadd4229b709;hb=0701e5d43bce14f30c7793f8c7f970ca8bac16ad;hp=c77e95c0ae999fd9a21a222910e628196a578a3d;hpb=0adbcf3ba9e673897165855db5fbde6e83d9b840;p=oota-llvm.git diff --git a/lib/Target/X86/X86FastISel.cpp b/lib/Target/X86/X86FastISel.cpp index c77e95c0ae9..f42465d14ce 100644 --- a/lib/Target/X86/X86FastISel.cpp +++ b/lib/Target/X86/X86FastISel.cpp @@ -16,10 +16,12 @@ #include "X86.h" #include "X86CallingConv.h" #include "X86InstrBuilder.h" +#include "X86InstrInfo.h" #include "X86MachineFunctionInfo.h" #include "X86RegisterInfo.h" #include "X86Subtarget.h" #include "X86TargetMachine.h" +#include "llvm/Analysis/BranchProbabilityInfo.h" #include "llvm/CodeGen/Analysis.h" #include "llvm/CodeGen/FastISel.h" #include "llvm/CodeGen/FunctionLoweringInfo.h" @@ -72,18 +74,22 @@ public: const LoadInst *LI) override; bool FastLowerArguments() override; + bool FastLowerCall(CallLoweringInfo &CLI) override; + bool FastLowerIntrinsicCall(const IntrinsicInst *II) override; #include "X86GenFastISel.inc" private: bool X86FastEmitCompare(const Value *LHS, const Value *RHS, EVT VT); - bool X86FastEmitLoad(EVT VT, const X86AddressMode &AM, unsigned &RR); + bool X86FastEmitLoad(EVT VT, const X86AddressMode &AM, MachineMemOperand *MMO, + unsigned &ResultReg); 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); + MachineMemOperand *MMO = nullptr, bool Aligned = false); + bool X86FastEmitStore(EVT VT, unsigned ValReg, bool ValIsKill, + const X86AddressMode &AM, + MachineMemOperand *MMO = nullptr, bool Aligned = false); bool X86FastEmitExtend(ISD::NodeType Opc, EVT DstVT, unsigned Src, EVT SrcVT, unsigned &ResultReg); @@ -107,6 +113,12 @@ private: bool X86SelectDivRem(const Instruction *I); + bool X86FastEmitCMoveSelect(MVT RetVT, const Instruction *I); + + bool X86FastEmitSSESelect(MVT RetVT, const Instruction *I); + + bool X86FastEmitPseudoSelect(MVT RetVT, const Instruction *I); + bool X86SelectSelect(const Instruction *I); bool X86SelectTrunc(const Instruction *I); @@ -114,13 +126,8 @@ private: bool X86SelectFPExt(const Instruction *I); bool X86SelectFPTrunc(const Instruction *I); - bool X86VisitIntrinsicCall(const IntrinsicInst &I); - bool X86SelectCall(const Instruction *I); - - bool DoSelectCall(const Instruction *I, const char *MemIntName); - const X86InstrInfo *getInstrInfo() const { - return getTargetMachine()->getInstrInfo(); + return getTargetMachine()->getSubtargetImpl()->getInstrInfo(); } const X86TargetMachine *getTargetMachine() const { return static_cast(&TM); @@ -128,6 +135,9 @@ private: bool handleConstantAddresses(const Value *V, X86AddressMode &AM); + unsigned X86MaterializeInt(const ConstantInt *CI, MVT VT); + unsigned X86MaterializeFP(const ConstantFP *CFP, MVT VT); + unsigned X86MaterializeGV(const GlobalValue *GV,MVT VT); unsigned TargetMaterializeConstant(const Constant *C) override; unsigned TargetMaterializeAlloca(const AllocaInst *C) override; @@ -147,10 +157,182 @@ private: bool TryEmitSmallMemcpy(X86AddressMode DestAM, X86AddressMode SrcAM, uint64_t Len); + + bool foldX86XALUIntrinsic(X86::CondCode &CC, const Instruction *I, + const Value *Cond); }; } // end anonymous namespace. +static CmpInst::Predicate optimizeCmpPredicate(const CmpInst *CI) { + // If both operands are the same, then try to optimize or fold the cmp. + CmpInst::Predicate Predicate = CI->getPredicate(); + if (CI->getOperand(0) != CI->getOperand(1)) + return Predicate; + + switch (Predicate) { + default: llvm_unreachable("Invalid predicate!"); + case CmpInst::FCMP_FALSE: Predicate = CmpInst::FCMP_FALSE; break; + case CmpInst::FCMP_OEQ: Predicate = CmpInst::FCMP_ORD; break; + case CmpInst::FCMP_OGT: Predicate = CmpInst::FCMP_FALSE; break; + case CmpInst::FCMP_OGE: Predicate = CmpInst::FCMP_ORD; break; + case CmpInst::FCMP_OLT: Predicate = CmpInst::FCMP_FALSE; break; + case CmpInst::FCMP_OLE: Predicate = CmpInst::FCMP_ORD; break; + case CmpInst::FCMP_ONE: Predicate = CmpInst::FCMP_FALSE; break; + case CmpInst::FCMP_ORD: Predicate = CmpInst::FCMP_ORD; break; + case CmpInst::FCMP_UNO: Predicate = CmpInst::FCMP_UNO; break; + case CmpInst::FCMP_UEQ: Predicate = CmpInst::FCMP_TRUE; break; + case CmpInst::FCMP_UGT: Predicate = CmpInst::FCMP_UNO; break; + case CmpInst::FCMP_UGE: Predicate = CmpInst::FCMP_TRUE; break; + case CmpInst::FCMP_ULT: Predicate = CmpInst::FCMP_UNO; break; + case CmpInst::FCMP_ULE: Predicate = CmpInst::FCMP_TRUE; break; + case CmpInst::FCMP_UNE: Predicate = CmpInst::FCMP_UNO; break; + case CmpInst::FCMP_TRUE: Predicate = CmpInst::FCMP_TRUE; break; + + case CmpInst::ICMP_EQ: Predicate = CmpInst::FCMP_TRUE; break; + case CmpInst::ICMP_NE: Predicate = CmpInst::FCMP_FALSE; break; + case CmpInst::ICMP_UGT: Predicate = CmpInst::FCMP_FALSE; break; + case CmpInst::ICMP_UGE: Predicate = CmpInst::FCMP_TRUE; break; + case CmpInst::ICMP_ULT: Predicate = CmpInst::FCMP_FALSE; break; + case CmpInst::ICMP_ULE: Predicate = CmpInst::FCMP_TRUE; break; + case CmpInst::ICMP_SGT: Predicate = CmpInst::FCMP_FALSE; break; + case CmpInst::ICMP_SGE: Predicate = CmpInst::FCMP_TRUE; break; + case CmpInst::ICMP_SLT: Predicate = CmpInst::FCMP_FALSE; break; + case CmpInst::ICMP_SLE: Predicate = CmpInst::FCMP_TRUE; break; + } + + return Predicate; +} + +static std::pair +getX86ConditionCode(CmpInst::Predicate Predicate) { + X86::CondCode CC = X86::COND_INVALID; + bool NeedSwap = false; + switch (Predicate) { + default: break; + // Floating-point Predicates + case CmpInst::FCMP_UEQ: CC = X86::COND_E; break; + case CmpInst::FCMP_OLT: NeedSwap = true; // fall-through + case CmpInst::FCMP_OGT: CC = X86::COND_A; break; + case CmpInst::FCMP_OLE: NeedSwap = true; // fall-through + case CmpInst::FCMP_OGE: CC = X86::COND_AE; break; + case CmpInst::FCMP_UGT: NeedSwap = true; // fall-through + case CmpInst::FCMP_ULT: CC = X86::COND_B; break; + case CmpInst::FCMP_UGE: NeedSwap = true; // fall-through + case CmpInst::FCMP_ULE: CC = X86::COND_BE; break; + case CmpInst::FCMP_ONE: CC = X86::COND_NE; break; + case CmpInst::FCMP_UNO: CC = X86::COND_P; break; + case CmpInst::FCMP_ORD: CC = X86::COND_NP; break; + case CmpInst::FCMP_OEQ: // fall-through + case CmpInst::FCMP_UNE: CC = X86::COND_INVALID; break; + + // Integer Predicates + case CmpInst::ICMP_EQ: CC = X86::COND_E; break; + case CmpInst::ICMP_NE: CC = X86::COND_NE; break; + case CmpInst::ICMP_UGT: CC = X86::COND_A; break; + case CmpInst::ICMP_UGE: CC = X86::COND_AE; break; + case CmpInst::ICMP_ULT: CC = X86::COND_B; break; + case CmpInst::ICMP_ULE: CC = X86::COND_BE; break; + case CmpInst::ICMP_SGT: CC = X86::COND_G; break; + case CmpInst::ICMP_SGE: CC = X86::COND_GE; break; + case CmpInst::ICMP_SLT: CC = X86::COND_L; break; + case CmpInst::ICMP_SLE: CC = X86::COND_LE; break; + } + + return std::make_pair(CC, NeedSwap); +} + +static std::pair +getX86SSEConditionCode(CmpInst::Predicate Predicate) { + unsigned CC; + bool NeedSwap = false; + + // SSE Condition code mapping: + // 0 - EQ + // 1 - LT + // 2 - LE + // 3 - UNORD + // 4 - NEQ + // 5 - NLT + // 6 - NLE + // 7 - ORD + switch (Predicate) { + default: llvm_unreachable("Unexpected predicate"); + case CmpInst::FCMP_OEQ: CC = 0; break; + case CmpInst::FCMP_OGT: NeedSwap = true; // fall-through + case CmpInst::FCMP_OLT: CC = 1; break; + case CmpInst::FCMP_OGE: NeedSwap = true; // fall-through + case CmpInst::FCMP_OLE: CC = 2; break; + case CmpInst::FCMP_UNO: CC = 3; break; + case CmpInst::FCMP_UNE: CC = 4; break; + case CmpInst::FCMP_ULE: NeedSwap = true; // fall-through + case CmpInst::FCMP_UGE: CC = 5; break; + case CmpInst::FCMP_ULT: NeedSwap = true; // fall-through + case CmpInst::FCMP_UGT: CC = 6; break; + case CmpInst::FCMP_ORD: CC = 7; break; + case CmpInst::FCMP_UEQ: + case CmpInst::FCMP_ONE: CC = 8; break; + } + + return std::make_pair(CC, NeedSwap); +} + +/// \brief Check if it is possible to fold the condition from the XALU intrinsic +/// into the user. The condition code will only be updated on success. +bool X86FastISel::foldX86XALUIntrinsic(X86::CondCode &CC, const Instruction *I, + const Value *Cond) { + if (!isa(Cond)) + return false; + + const auto *EV = cast(Cond); + if (!isa(EV->getAggregateOperand())) + return false; + + const auto *II = cast(EV->getAggregateOperand()); + MVT RetVT; + const Function *Callee = II->getCalledFunction(); + Type *RetTy = + cast(Callee->getReturnType())->getTypeAtIndex(0U); + if (!isTypeLegal(RetTy, RetVT)) + return false; + + if (RetVT != MVT::i32 && RetVT != MVT::i64) + return false; + + X86::CondCode TmpCC; + switch (II->getIntrinsicID()) { + default: return false; + case Intrinsic::sadd_with_overflow: + case Intrinsic::ssub_with_overflow: + case Intrinsic::smul_with_overflow: + case Intrinsic::umul_with_overflow: TmpCC = X86::COND_O; break; + case Intrinsic::uadd_with_overflow: + case Intrinsic::usub_with_overflow: TmpCC = X86::COND_B; break; + } + + // Check if both instructions are in the same basic block. + if (II->getParent() != I->getParent()) + return false; + + // Make sure nothing is in the way + BasicBlock::const_iterator Start = I; + BasicBlock::const_iterator End = II; + for (auto Itr = std::prev(Start); Itr != End; --Itr) { + // We only expect extractvalue instructions between the intrinsic and the + // instruction to be selected. + if (!isa(Itr)) + return false; + + // Check that the extractvalue operand comes from the intrinsic. + const auto *EVI = cast(Itr); + if (EVI->getAggregateOperand() != II) + return false; + } + + CC = TmpCC; + return true; +} + bool X86FastISel::isTypeLegal(Type *Ty, MVT &VT, bool AllowI1) { EVT evt = TLI.getValueType(Ty, /*HandleUnknown=*/true); if (evt == MVT::Other || !evt.isSimple()) @@ -180,7 +362,7 @@ bool X86FastISel::isTypeLegal(Type *Ty, MVT &VT, bool AllowI1) { /// The address is either pre-computed, i.e. Ptr, or a GlobalAddress, i.e. GV. /// Return true and the result register by reference if it is possible. bool X86FastISel::X86FastEmitLoad(EVT VT, const X86AddressMode &AM, - unsigned &ResultReg) { + MachineMemOperand *MMO, unsigned &ResultReg) { // Get opcode and regclass of the output for the given load instruction. unsigned Opc = 0; const TargetRegisterClass *RC = nullptr; @@ -228,8 +410,11 @@ bool X86FastISel::X86FastEmitLoad(EVT VT, const X86AddressMode &AM, } ResultReg = createResultReg(RC); - addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, - DbgLoc, TII.get(Opc), ResultReg), AM); + MachineInstrBuilder MIB = + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg); + addFullAddress(MIB, AM); + if (MMO) + MIB->addMemOperand(*FuncInfo.MF, MMO); return true; } @@ -237,9 +422,9 @@ bool X86FastISel::X86FastEmitLoad(EVT VT, const X86AddressMode &AM, /// type VT. The address is either pre-computed, consisted of a base ptr, Ptr /// and a displacement offset, or a GlobalAddress, /// i.e. V. Return true if it is possible. -bool -X86FastISel::X86FastEmitStore(EVT VT, unsigned ValReg, - const X86AddressMode &AM, bool Aligned) { +bool X86FastISel::X86FastEmitStore(EVT VT, unsigned ValReg, bool ValIsKill, + const X86AddressMode &AM, + MachineMemOperand *MMO, bool Aligned) { // Get opcode and regclass of the output for the given store instruction. unsigned Opc = 0; switch (VT.getSimpleVT().SimpleTy) { @@ -249,7 +434,8 @@ X86FastISel::X86FastEmitStore(EVT VT, unsigned ValReg, // Mask out all but lowest bit. unsigned AndResult = createResultReg(&X86::GR8RegClass); BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, - TII.get(X86::AND8ri), AndResult).addReg(ValReg).addImm(1); + TII.get(X86::AND8ri), AndResult) + .addReg(ValReg, getKillRegState(ValIsKill)).addImm(1); ValReg = AndResult; } // FALLTHROUGH, handling i1 as i8. @@ -288,13 +474,18 @@ X86FastISel::X86FastEmitStore(EVT VT, unsigned ValReg, break; } - addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, - DbgLoc, TII.get(Opc)), AM).addReg(ValReg); + MachineInstrBuilder MIB = + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc)); + addFullAddress(MIB, AM).addReg(ValReg, getKillRegState(ValIsKill)); + if (MMO) + MIB->addMemOperand(*FuncInfo.MF, MMO); + return true; } bool X86FastISel::X86FastEmitStore(EVT VT, const Value *Val, - const X86AddressMode &AM, bool Aligned) { + const X86AddressMode &AM, + MachineMemOperand *MMO, bool Aligned) { // Handle 'null' like i32/i64 0. if (isa(Val)) Val = Constant::getNullValue(DL.getIntPtrType(Val->getContext())); @@ -317,10 +508,12 @@ bool X86FastISel::X86FastEmitStore(EVT VT, const Value *Val, } if (Opc) { - addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, - DbgLoc, TII.get(Opc)), AM) - .addImm(Signed ? (uint64_t) CI->getSExtValue() : - CI->getZExtValue()); + MachineInstrBuilder MIB = + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc)); + addFullAddress(MIB, AM).addImm(Signed ? (uint64_t) CI->getSExtValue() + : CI->getZExtValue()); + if (MMO) + MIB->addMemOperand(*FuncInfo.MF, MMO); return true; } } @@ -329,7 +522,8 @@ bool X86FastISel::X86FastEmitStore(EVT VT, const Value *Val, if (ValReg == 0) return false; - return X86FastEmitStore(VT, ValReg, AM, Aligned); + bool ValKill = hasTrivialKill(Val); + return X86FastEmitStore(VT, ValReg, ValKill, AM, MMO, Aligned); } /// X86FastEmitExtend - Emit a machine instruction to extend a value Src of @@ -687,7 +881,7 @@ bool X86FastISel::X86SelectCallAddress(const Value *V, X86AddressMode &AM) { (AM.Base.Reg != 0 || AM.IndexReg != 0)) return false; - // Can't handle DbgLocLImport. + // Can't handle DLL Import. if (GV->hasDLLImportStorageClass()) return false; @@ -740,19 +934,24 @@ bool X86FastISel::X86SelectStore(const Instruction *I) { if (S->isAtomic()) return false; - unsigned SABIAlignment = - DL.getABITypeAlignment(S->getValueOperand()->getType()); - bool Aligned = S->getAlignment() == 0 || S->getAlignment() >= SABIAlignment; + const Value *Val = S->getValueOperand(); + const Value *Ptr = S->getPointerOperand(); MVT VT; - if (!isTypeLegal(I->getOperand(0)->getType(), VT, /*AllowI1=*/true)) + if (!isTypeLegal(Val->getType(), VT, /*AllowI1=*/true)) return false; + unsigned Alignment = S->getAlignment(); + unsigned ABIAlignment = DL.getABITypeAlignment(Val->getType()); + if (Alignment == 0) // Ensure that codegen never sees alignment 0 + Alignment = ABIAlignment; + bool Aligned = Alignment >= ABIAlignment; + X86AddressMode AM; - if (!X86SelectAddress(I->getOperand(1), AM)) + if (!X86SelectAddress(Ptr, AM)) return false; - return X86FastEmitStore(VT, I->getOperand(0), AM, Aligned); + return X86FastEmitStore(VT, Val, AM, createMachineMemOperandFor(I), Aligned); } /// X86SelectRet - Select and emit code to implement ret instructions. @@ -797,8 +996,7 @@ bool X86FastISel::X86SelectRet(const Instruction *I) { // Analyze operands of the call, assigning locations to each operand. SmallVector ValLocs; - CCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, TM, ValLocs, - I->getContext()); + CCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, ValLocs, I->getContext()); CCInfo.AnalyzeReturn(Outs, RetCC_X86); const Value *RV = Ret->getOperand(0); @@ -821,7 +1019,7 @@ bool X86FastISel::X86SelectRet(const Instruction *I) { // The calling-convention tables for x87 returns don't tell // the whole story. - if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1) + if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) return false; unsigned SrcReg = Reg + VA.getValNo(); @@ -887,25 +1085,29 @@ bool X86FastISel::X86SelectRet(const Instruction *I) { /// X86SelectLoad - Select and emit code to implement load instructions. /// -bool X86FastISel::X86SelectLoad(const Instruction *I) { +bool X86FastISel::X86SelectLoad(const Instruction *I) { + const LoadInst *LI = cast(I); + // Atomic loads need special handling. - if (cast(I)->isAtomic()) + if (LI->isAtomic()) return false; MVT VT; - if (!isTypeLegal(I->getType(), VT, /*AllowI1=*/true)) + if (!isTypeLegal(LI->getType(), VT, /*AllowI1=*/true)) return false; + const Value *Ptr = LI->getPointerOperand(); + X86AddressMode AM; - if (!X86SelectAddress(I->getOperand(0), AM)) + if (!X86SelectAddress(Ptr, AM)) return false; unsigned ResultReg = 0; - if (X86FastEmitLoad(VT, AM, ResultReg)) { - UpdateValueMap(I, ResultReg); - return true; - } - return false; + if (!X86FastEmitLoad(VT, AM, createMachineMemOperandFor(LI), ResultReg)) + return false; + + UpdateValueMap(I, ResultReg); + return true; } static unsigned X86ChooseCmpOpcode(EVT VT, const X86Subtarget *Subtarget) { @@ -985,73 +1187,89 @@ bool X86FastISel::X86SelectCmp(const Instruction *I) { if (!isTypeLegal(I->getOperand(0)->getType(), VT)) return false; - unsigned ResultReg = createResultReg(&X86::GR8RegClass); - unsigned SetCCOpc; - bool SwapArgs; // false -> compare Op0, Op1. true -> compare Op1, Op0. - switch (CI->getPredicate()) { - case CmpInst::FCMP_OEQ: { - if (!X86FastEmitCompare(CI->getOperand(0), CI->getOperand(1), VT)) + // Try to optimize or fold the cmp. + CmpInst::Predicate Predicate = optimizeCmpPredicate(CI); + unsigned ResultReg = 0; + switch (Predicate) { + default: break; + case CmpInst::FCMP_FALSE: { + ResultReg = createResultReg(&X86::GR32RegClass); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::MOV32r0), + ResultReg); + ResultReg = FastEmitInst_extractsubreg(MVT::i8, ResultReg, /*Kill=*/true, + X86::sub_8bit); + if (!ResultReg) return false; + break; + } + case CmpInst::FCMP_TRUE: { + ResultReg = createResultReg(&X86::GR8RegClass); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::MOV8ri), + ResultReg).addImm(1); + break; + } + } - unsigned EReg = createResultReg(&X86::GR8RegClass); - unsigned NPReg = createResultReg(&X86::GR8RegClass); - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::SETEr), EReg); - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, - TII.get(X86::SETNPr), NPReg); - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, - TII.get(X86::AND8rr), ResultReg).addReg(NPReg).addReg(EReg); + if (ResultReg) { UpdateValueMap(I, ResultReg); return true; } - case CmpInst::FCMP_UNE: { - if (!X86FastEmitCompare(CI->getOperand(0), CI->getOperand(1), VT)) + + const Value *LHS = CI->getOperand(0); + const Value *RHS = CI->getOperand(1); + + // The optimizer might have replaced fcmp oeq %x, %x with fcmp ord %x, 0.0. + // We don't have to materialize a zero constant for this case and can just use + // %x again on the RHS. + if (Predicate == CmpInst::FCMP_ORD || Predicate == CmpInst::FCMP_UNO) { + const auto *RHSC = dyn_cast(RHS); + if (RHSC && RHSC->isNullValue()) + RHS = LHS; + } + + // FCMP_OEQ and FCMP_UNE cannot be checked with a single instruction. + static unsigned SETFOpcTable[2][3] = { + { X86::SETEr, X86::SETNPr, X86::AND8rr }, + { X86::SETNEr, X86::SETPr, X86::OR8rr } + }; + unsigned *SETFOpc = nullptr; + switch (Predicate) { + default: break; + case CmpInst::FCMP_OEQ: SETFOpc = &SETFOpcTable[0][0]; break; + case CmpInst::FCMP_UNE: SETFOpc = &SETFOpcTable[1][0]; break; + } + + ResultReg = createResultReg(&X86::GR8RegClass); + if (SETFOpc) { + if (!X86FastEmitCompare(LHS, RHS, VT)) return false; - unsigned NEReg = createResultReg(&X86::GR8RegClass); - unsigned PReg = createResultReg(&X86::GR8RegClass); - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::SETNEr), NEReg); - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::SETPr), PReg); - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::OR8rr),ResultReg) - .addReg(PReg).addReg(NEReg); + unsigned FlagReg1 = createResultReg(&X86::GR8RegClass); + unsigned FlagReg2 = createResultReg(&X86::GR8RegClass); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(SETFOpc[0]), + FlagReg1); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(SETFOpc[1]), + FlagReg2); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(SETFOpc[2]), + ResultReg).addReg(FlagReg1).addReg(FlagReg2); UpdateValueMap(I, ResultReg); return true; } - case CmpInst::FCMP_OGT: SwapArgs = false; SetCCOpc = X86::SETAr; break; - case CmpInst::FCMP_OGE: SwapArgs = false; SetCCOpc = X86::SETAEr; break; - case CmpInst::FCMP_OLT: SwapArgs = true; SetCCOpc = X86::SETAr; break; - case CmpInst::FCMP_OLE: SwapArgs = true; SetCCOpc = X86::SETAEr; break; - case CmpInst::FCMP_ONE: SwapArgs = false; SetCCOpc = X86::SETNEr; break; - case CmpInst::FCMP_ORD: SwapArgs = false; SetCCOpc = X86::SETNPr; break; - case CmpInst::FCMP_UNO: SwapArgs = false; SetCCOpc = X86::SETPr; break; - case CmpInst::FCMP_UEQ: SwapArgs = false; SetCCOpc = X86::SETEr; break; - case CmpInst::FCMP_UGT: SwapArgs = true; SetCCOpc = X86::SETBr; break; - case CmpInst::FCMP_UGE: SwapArgs = true; SetCCOpc = X86::SETBEr; break; - case CmpInst::FCMP_ULT: SwapArgs = false; SetCCOpc = X86::SETBr; break; - case CmpInst::FCMP_ULE: SwapArgs = false; SetCCOpc = X86::SETBEr; break; - - case CmpInst::ICMP_EQ: SwapArgs = false; SetCCOpc = X86::SETEr; break; - case CmpInst::ICMP_NE: SwapArgs = false; SetCCOpc = X86::SETNEr; break; - case CmpInst::ICMP_UGT: SwapArgs = false; SetCCOpc = X86::SETAr; break; - case CmpInst::ICMP_UGE: SwapArgs = false; SetCCOpc = X86::SETAEr; break; - case CmpInst::ICMP_ULT: SwapArgs = false; SetCCOpc = X86::SETBr; break; - case CmpInst::ICMP_ULE: SwapArgs = false; SetCCOpc = X86::SETBEr; break; - case CmpInst::ICMP_SGT: SwapArgs = false; SetCCOpc = X86::SETGr; break; - case CmpInst::ICMP_SGE: SwapArgs = false; SetCCOpc = X86::SETGEr; break; - case CmpInst::ICMP_SLT: SwapArgs = false; SetCCOpc = X86::SETLr; break; - case CmpInst::ICMP_SLE: SwapArgs = false; SetCCOpc = X86::SETLEr; break; - default: - return false; - } - const Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1); + X86::CondCode CC; + bool SwapArgs; + std::tie(CC, SwapArgs) = getX86ConditionCode(Predicate); + assert(CC <= X86::LAST_VALID_COND && "Unexpected condition code."); + unsigned Opc = X86::getSETFromCond(CC); + if (SwapArgs) - std::swap(Op0, Op1); + std::swap(LHS, RHS); - // Emit a compare of Op0/Op1. - if (!X86FastEmitCompare(Op0, Op1, VT)) + // Emit a compare of LHS/RHS. + if (!X86FastEmitCompare(LHS, RHS, VT)) return false; - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(SetCCOpc), ResultReg); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg); UpdateValueMap(I, ResultReg); return true; } @@ -1117,73 +1335,88 @@ bool X86FastISel::X86SelectBranch(const Instruction *I) { // Fold the common case of a conditional branch with a comparison // in the same block (values defined on other blocks may not have // initialized registers). + X86::CondCode CC; if (const CmpInst *CI = dyn_cast(BI->getCondition())) { if (CI->hasOneUse() && CI->getParent() == I->getParent()) { EVT VT = TLI.getValueType(CI->getOperand(0)->getType()); + // Try to optimize or fold the cmp. + CmpInst::Predicate Predicate = optimizeCmpPredicate(CI); + switch (Predicate) { + default: break; + case CmpInst::FCMP_FALSE: FastEmitBranch(FalseMBB, DbgLoc); return true; + case CmpInst::FCMP_TRUE: FastEmitBranch(TrueMBB, DbgLoc); return true; + } + + const Value *CmpLHS = CI->getOperand(0); + const Value *CmpRHS = CI->getOperand(1); + + // The optimizer might have replaced fcmp oeq %x, %x with fcmp ord %x, + // 0.0. + // We don't have to materialize a zero constant for this case and can just + // use %x again on the RHS. + if (Predicate == CmpInst::FCMP_ORD || Predicate == CmpInst::FCMP_UNO) { + const auto *CmpRHSC = dyn_cast(CmpRHS); + if (CmpRHSC && CmpRHSC->isNullValue()) + CmpRHS = CmpLHS; + } + // Try to take advantage of fallthrough opportunities. - CmpInst::Predicate Predicate = CI->getPredicate(); if (FuncInfo.MBB->isLayoutSuccessor(TrueMBB)) { std::swap(TrueMBB, FalseMBB); Predicate = CmpInst::getInversePredicate(Predicate); } - bool SwapArgs; // false -> compare Op0, Op1. true -> compare Op1, Op0. - unsigned BranchOpc; // Opcode to jump on, e.g. "X86::JA" - + // FCMP_OEQ and FCMP_UNE cannot be expressed with a single flag/condition + // code check. Instead two branch instructions are required to check all + // the flags. First we change the predicate to a supported condition code, + // which will be the first branch. Later one we will emit the second + // branch. + bool NeedExtraBranch = false; switch (Predicate) { + default: break; case CmpInst::FCMP_OEQ: - std::swap(TrueMBB, FalseMBB); - Predicate = CmpInst::FCMP_UNE; - // FALL THROUGH - case CmpInst::FCMP_UNE: SwapArgs = false; BranchOpc = X86::JNE_4; break; - case CmpInst::FCMP_OGT: SwapArgs = false; BranchOpc = X86::JA_4; break; - case CmpInst::FCMP_OGE: SwapArgs = false; BranchOpc = X86::JAE_4; break; - case CmpInst::FCMP_OLT: SwapArgs = true; BranchOpc = X86::JA_4; break; - case CmpInst::FCMP_OLE: SwapArgs = true; BranchOpc = X86::JAE_4; break; - case CmpInst::FCMP_ONE: SwapArgs = false; BranchOpc = X86::JNE_4; break; - case CmpInst::FCMP_ORD: SwapArgs = false; BranchOpc = X86::JNP_4; break; - case CmpInst::FCMP_UNO: SwapArgs = false; BranchOpc = X86::JP_4; break; - case CmpInst::FCMP_UEQ: SwapArgs = false; BranchOpc = X86::JE_4; break; - case CmpInst::FCMP_UGT: SwapArgs = true; BranchOpc = X86::JB_4; break; - case CmpInst::FCMP_UGE: SwapArgs = true; BranchOpc = X86::JBE_4; break; - case CmpInst::FCMP_ULT: SwapArgs = false; BranchOpc = X86::JB_4; break; - case CmpInst::FCMP_ULE: SwapArgs = false; BranchOpc = X86::JBE_4; break; - - case CmpInst::ICMP_EQ: SwapArgs = false; BranchOpc = X86::JE_4; break; - case CmpInst::ICMP_NE: SwapArgs = false; BranchOpc = X86::JNE_4; break; - case CmpInst::ICMP_UGT: SwapArgs = false; BranchOpc = X86::JA_4; break; - case CmpInst::ICMP_UGE: SwapArgs = false; BranchOpc = X86::JAE_4; break; - case CmpInst::ICMP_ULT: SwapArgs = false; BranchOpc = X86::JB_4; break; - case CmpInst::ICMP_ULE: SwapArgs = false; BranchOpc = X86::JBE_4; break; - case CmpInst::ICMP_SGT: SwapArgs = false; BranchOpc = X86::JG_4; break; - case CmpInst::ICMP_SGE: SwapArgs = false; BranchOpc = X86::JGE_4; break; - case CmpInst::ICMP_SLT: SwapArgs = false; BranchOpc = X86::JL_4; break; - case CmpInst::ICMP_SLE: SwapArgs = false; BranchOpc = X86::JLE_4; break; - default: - return false; + std::swap(TrueMBB, FalseMBB); // fall-through + case CmpInst::FCMP_UNE: + NeedExtraBranch = true; + Predicate = CmpInst::FCMP_ONE; + break; } - const Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1); + bool SwapArgs; + unsigned BranchOpc; + std::tie(CC, SwapArgs) = getX86ConditionCode(Predicate); + assert(CC <= X86::LAST_VALID_COND && "Unexpected condition code."); + + BranchOpc = X86::GetCondBranchFromCond(CC); if (SwapArgs) - std::swap(Op0, Op1); + std::swap(CmpLHS, CmpRHS); // Emit a compare of the LHS and RHS, setting the flags. - if (!X86FastEmitCompare(Op0, Op1, VT)) + if (!X86FastEmitCompare(CmpLHS, CmpRHS, VT)) return false; BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(BranchOpc)) .addMBB(TrueMBB); - if (Predicate == CmpInst::FCMP_UNE) { - // X86 requires a second branch to handle UNE (and OEQ, - // which is mapped to UNE above). + // X86 requires a second branch to handle UNE (and OEQ, which is mapped + // to UNE above). + if (NeedExtraBranch) { BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::JP_4)) .addMBB(TrueMBB); } + // Obtain the branch weight and add the TrueBB to the successor list. + uint32_t BranchWeight = 0; + if (FuncInfo.BPI) + BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(), + TrueMBB->getBasicBlock()); + FuncInfo.MBB->addSuccessor(TrueMBB, BranchWeight); + + // Emits an unconditional branch to the FalseBB, obtains the branch + // weight, and adds it to the successor list. FastEmitBranch(FalseMBB, DbgLoc); - FuncInfo.MBB->addSuccessor(TrueMBB); + return true; } } else if (TruncInst *TI = dyn_cast(BI->getCondition())) { @@ -1215,10 +1448,32 @@ bool X86FastISel::X86SelectBranch(const Instruction *I) { BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(JmpOpc)) .addMBB(TrueMBB); FastEmitBranch(FalseMBB, DbgLoc); - FuncInfo.MBB->addSuccessor(TrueMBB); + uint32_t BranchWeight = 0; + if (FuncInfo.BPI) + BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(), + TrueMBB->getBasicBlock()); + FuncInfo.MBB->addSuccessor(TrueMBB, BranchWeight); return true; } } + } else if (foldX86XALUIntrinsic(CC, BI, BI->getCondition())) { + // Fake request the condition, otherwise the intrinsic might be completely + // optimized away. + unsigned TmpReg = getRegForValue(BI->getCondition()); + if (TmpReg == 0) + return false; + + unsigned BranchOpc = X86::GetCondBranchFromCond(CC); + + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(BranchOpc)) + .addMBB(TrueMBB); + FastEmitBranch(FalseMBB, DbgLoc); + uint32_t BranchWeight = 0; + if (FuncInfo.BPI) + BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(), + TrueMBB->getBasicBlock()); + FuncInfo.MBB->addSuccessor(TrueMBB, BranchWeight); + return true; } // Otherwise do a clumsy setcc and re-test it. @@ -1232,7 +1487,11 @@ bool X86FastISel::X86SelectBranch(const Instruction *I) { BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::JNE_4)) .addMBB(TrueMBB); FastEmitBranch(FalseMBB, DbgLoc); - FuncInfo.MBB->addSuccessor(TrueMBB); + uint32_t BranchWeight = 0; + if (FuncInfo.BPI) + BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(), + TrueMBB->getBasicBlock()); + FuncInfo.MBB->addSuccessor(TrueMBB, BranchWeight); return true; } @@ -1469,50 +1728,319 @@ bool X86FastISel::X86SelectDivRem(const Instruction *I) { return true; } -bool X86FastISel::X86SelectSelect(const Instruction *I) { - MVT VT; - if (!isTypeLegal(I->getType(), VT)) +/// \brief Emit a conditional move instruction (if the are supported) to lower +/// the select. +bool X86FastISel::X86FastEmitCMoveSelect(MVT RetVT, const Instruction *I) { + // Check if the subtarget supports these instructions. + if (!Subtarget->hasCMov()) return false; - // We only use cmov here, if we don't have a cmov instruction bail. - if (!Subtarget->hasCMov()) return false; + // FIXME: Add support for i8. + if (RetVT < MVT::i16 || RetVT > MVT::i64) + return false; - unsigned Opc = 0; - const TargetRegisterClass *RC = nullptr; - if (VT == MVT::i16) { - Opc = X86::CMOVE16rr; - RC = &X86::GR16RegClass; - } else if (VT == MVT::i32) { - Opc = X86::CMOVE32rr; - RC = &X86::GR32RegClass; - } else if (VT == MVT::i64) { - Opc = X86::CMOVE64rr; - RC = &X86::GR64RegClass; - } else { + const Value *Cond = I->getOperand(0); + const TargetRegisterClass *RC = TLI.getRegClassFor(RetVT); + bool NeedTest = true; + X86::CondCode CC = X86::COND_NE; + + // Optimize conditions coming from a compare if both instructions are in the + // same basic block (values defined in other basic blocks may not have + // initialized registers). + const auto *CI = dyn_cast(Cond); + if (CI && (CI->getParent() == I->getParent())) { + CmpInst::Predicate Predicate = optimizeCmpPredicate(CI); + + // FCMP_OEQ and FCMP_UNE cannot be checked with a single instruction. + static unsigned SETFOpcTable[2][3] = { + { X86::SETNPr, X86::SETEr , X86::TEST8rr }, + { X86::SETPr, X86::SETNEr, X86::OR8rr } + }; + unsigned *SETFOpc = nullptr; + switch (Predicate) { + default: break; + case CmpInst::FCMP_OEQ: + SETFOpc = &SETFOpcTable[0][0]; + Predicate = CmpInst::ICMP_NE; + break; + case CmpInst::FCMP_UNE: + SETFOpc = &SETFOpcTable[1][0]; + Predicate = CmpInst::ICMP_NE; + break; + } + + bool NeedSwap; + std::tie(CC, NeedSwap) = getX86ConditionCode(Predicate); + assert(CC <= X86::LAST_VALID_COND && "Unexpected condition code."); + + const Value *CmpLHS = CI->getOperand(0); + const Value *CmpRHS = CI->getOperand(1); + if (NeedSwap) + std::swap(CmpLHS, CmpRHS); + + EVT CmpVT = TLI.getValueType(CmpLHS->getType()); + // Emit a compare of the LHS and RHS, setting the flags. + if (!X86FastEmitCompare(CmpLHS, CmpRHS, CmpVT)) + return false; + + if (SETFOpc) { + unsigned FlagReg1 = createResultReg(&X86::GR8RegClass); + unsigned FlagReg2 = createResultReg(&X86::GR8RegClass); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(SETFOpc[0]), + FlagReg1); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(SETFOpc[1]), + FlagReg2); + auto const &II = TII.get(SETFOpc[2]); + if (II.getNumDefs()) { + unsigned TmpReg = createResultReg(&X86::GR8RegClass); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, TmpReg) + .addReg(FlagReg2).addReg(FlagReg1); + } else { + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II) + .addReg(FlagReg2).addReg(FlagReg1); + } + } + NeedTest = false; + } else if (foldX86XALUIntrinsic(CC, I, Cond)) { + // Fake request the condition, otherwise the intrinsic might be completely + // optimized away. + unsigned TmpReg = getRegForValue(Cond); + if (TmpReg == 0) + return false; + + NeedTest = false; + } + + if (NeedTest) { + // Selects operate on i1, however, CondReg is 8 bits width and may contain + // garbage. Indeed, only the less significant bit is supposed to be + // accurate. If we read more than the lsb, we may see non-zero values + // whereas lsb is zero. Therefore, we have to truncate Op0Reg to i1 for + // the select. This is achieved by performing TEST against 1. + unsigned CondReg = getRegForValue(Cond); + if (CondReg == 0) + return false; + bool CondIsKill = hasTrivialKill(Cond); + + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::TEST8ri)) + .addReg(CondReg, getKillRegState(CondIsKill)).addImm(1); + } + + const Value *LHS = I->getOperand(1); + const Value *RHS = I->getOperand(2); + + unsigned RHSReg = getRegForValue(RHS); + bool RHSIsKill = hasTrivialKill(RHS); + + unsigned LHSReg = getRegForValue(LHS); + bool LHSIsKill = hasTrivialKill(LHS); + + if (!LHSReg || !RHSReg) return false; + + unsigned Opc = X86::getCMovFromCond(CC, RC->getSize()); + unsigned ResultReg = FastEmitInst_rr(Opc, RC, RHSReg, RHSIsKill, + LHSReg, LHSIsKill); + UpdateValueMap(I, ResultReg); + return true; +} + +/// \brief Emit SSE instructions to lower the select. +/// +/// Try to use SSE1/SSE2 instructions to simulate a select without branches. +/// This lowers fp selects into a CMP/AND/ANDN/OR sequence when the necessary +/// SSE instructions are available. +bool X86FastISel::X86FastEmitSSESelect(MVT RetVT, const Instruction *I) { + // Optimize conditions coming from a compare if both instructions are in the + // same basic block (values defined in other basic blocks may not have + // initialized registers). + const auto *CI = dyn_cast(I->getOperand(0)); + if (!CI || (CI->getParent() != I->getParent())) + return false; + + if (I->getType() != CI->getOperand(0)->getType() || + !((Subtarget->hasSSE1() && RetVT == MVT::f32) || + (Subtarget->hasSSE2() && RetVT == MVT::f64) )) + return false; + + const Value *CmpLHS = CI->getOperand(0); + const Value *CmpRHS = CI->getOperand(1); + CmpInst::Predicate Predicate = optimizeCmpPredicate(CI); + + // The optimizer might have replaced fcmp oeq %x, %x with fcmp ord %x, 0.0. + // We don't have to materialize a zero constant for this case and can just use + // %x again on the RHS. + if (Predicate == CmpInst::FCMP_ORD || Predicate == CmpInst::FCMP_UNO) { + const auto *CmpRHSC = dyn_cast(CmpRHS); + if (CmpRHSC && CmpRHSC->isNullValue()) + CmpRHS = CmpLHS; } - unsigned Op0Reg = getRegForValue(I->getOperand(0)); - if (Op0Reg == 0) return false; - unsigned Op1Reg = getRegForValue(I->getOperand(1)); - if (Op1Reg == 0) return false; - unsigned Op2Reg = getRegForValue(I->getOperand(2)); - if (Op2Reg == 0) return false; - - // Selects operate on i1, however, Op0Reg is 8 bits width and may contain - // garbage. Indeed, only the less significant bit is supposed to be accurate. - // If we read more than the lsb, we may see non-zero values whereas lsb - // is zero. Therefore, we have to truncate Op0Reg to i1 for the select. - // This is achieved by performing TEST against 1. - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::TEST8ri)) - .addReg(Op0Reg).addImm(1); - unsigned ResultReg = createResultReg(RC); - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg) - .addReg(Op1Reg).addReg(Op2Reg); + unsigned CC; + bool NeedSwap; + std::tie(CC, NeedSwap) = getX86SSEConditionCode(Predicate); + if (CC > 7) + return false; + + if (NeedSwap) + std::swap(CmpLHS, CmpRHS); + + static unsigned OpcTable[2][2][4] = { + { { X86::CMPSSrr, X86::FsANDPSrr, X86::FsANDNPSrr, X86::FsORPSrr }, + { X86::VCMPSSrr, X86::VFsANDPSrr, X86::VFsANDNPSrr, X86::VFsORPSrr } }, + { { X86::CMPSDrr, X86::FsANDPDrr, X86::FsANDNPDrr, X86::FsORPDrr }, + { X86::VCMPSDrr, X86::VFsANDPDrr, X86::VFsANDNPDrr, X86::VFsORPDrr } } + }; + + bool HasAVX = Subtarget->hasAVX(); + unsigned *Opc = nullptr; + switch (RetVT.SimpleTy) { + default: return false; + case MVT::f32: Opc = &OpcTable[0][HasAVX][0]; break; + case MVT::f64: Opc = &OpcTable[1][HasAVX][0]; break; + } + + const Value *LHS = I->getOperand(1); + const Value *RHS = I->getOperand(2); + + unsigned LHSReg = getRegForValue(LHS); + bool LHSIsKill = hasTrivialKill(LHS); + + unsigned RHSReg = getRegForValue(RHS); + bool RHSIsKill = hasTrivialKill(RHS); + + unsigned CmpLHSReg = getRegForValue(CmpLHS); + bool CmpLHSIsKill = hasTrivialKill(CmpLHS); + + unsigned CmpRHSReg = getRegForValue(CmpRHS); + bool CmpRHSIsKill = hasTrivialKill(CmpRHS); + + if (!LHSReg || !RHSReg || !CmpLHS || !CmpRHS) + return false; + + const TargetRegisterClass *RC = TLI.getRegClassFor(RetVT); + unsigned CmpReg = FastEmitInst_rri(Opc[0], RC, CmpLHSReg, CmpLHSIsKill, + CmpRHSReg, CmpRHSIsKill, CC); + unsigned AndReg = FastEmitInst_rr(Opc[1], RC, CmpReg, /*IsKill=*/false, + LHSReg, LHSIsKill); + unsigned AndNReg = FastEmitInst_rr(Opc[2], RC, CmpReg, /*IsKill=*/true, + RHSReg, RHSIsKill); + unsigned ResultReg = FastEmitInst_rr(Opc[3], RC, AndNReg, /*IsKill=*/true, + AndReg, /*IsKill=*/true); + UpdateValueMap(I, ResultReg); + return true; +} + +bool X86FastISel::X86FastEmitPseudoSelect(MVT RetVT, const Instruction *I) { + // These are pseudo CMOV instructions and will be later expanded into control- + // flow. + unsigned Opc; + switch (RetVT.SimpleTy) { + default: return false; + case MVT::i8: Opc = X86::CMOV_GR8; break; + case MVT::i16: Opc = X86::CMOV_GR16; break; + case MVT::i32: Opc = X86::CMOV_GR32; break; + case MVT::f32: Opc = X86::CMOV_FR32; break; + case MVT::f64: Opc = X86::CMOV_FR64; break; + } + + const Value *Cond = I->getOperand(0); + X86::CondCode CC = X86::COND_NE; + + // Optimize conditions coming from a compare if both instructions are in the + // same basic block (values defined in other basic blocks may not have + // initialized registers). + const auto *CI = dyn_cast(Cond); + if (CI && (CI->getParent() == I->getParent())) { + bool NeedSwap; + std::tie(CC, NeedSwap) = getX86ConditionCode(CI->getPredicate()); + if (CC > X86::LAST_VALID_COND) + return false; + + const Value *CmpLHS = CI->getOperand(0); + const Value *CmpRHS = CI->getOperand(1); + + if (NeedSwap) + std::swap(CmpLHS, CmpRHS); + + EVT CmpVT = TLI.getValueType(CmpLHS->getType()); + if (!X86FastEmitCompare(CmpLHS, CmpRHS, CmpVT)) + return false; + } else { + unsigned CondReg = getRegForValue(Cond); + if (CondReg == 0) + return false; + bool CondIsKill = hasTrivialKill(Cond); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::TEST8ri)) + .addReg(CondReg, getKillRegState(CondIsKill)).addImm(1); + } + + const Value *LHS = I->getOperand(1); + const Value *RHS = I->getOperand(2); + + unsigned LHSReg = getRegForValue(LHS); + bool LHSIsKill = hasTrivialKill(LHS); + + unsigned RHSReg = getRegForValue(RHS); + bool RHSIsKill = hasTrivialKill(RHS); + + if (!LHSReg || !RHSReg) + return false; + + const TargetRegisterClass *RC = TLI.getRegClassFor(RetVT); + + unsigned ResultReg = + FastEmitInst_rri(Opc, RC, RHSReg, RHSIsKill, LHSReg, LHSIsKill, CC); UpdateValueMap(I, ResultReg); return true; } +bool X86FastISel::X86SelectSelect(const Instruction *I) { + MVT RetVT; + if (!isTypeLegal(I->getType(), RetVT)) + return false; + + // Check if we can fold the select. + if (const auto *CI = dyn_cast(I->getOperand(0))) { + CmpInst::Predicate Predicate = optimizeCmpPredicate(CI); + const Value *Opnd = nullptr; + switch (Predicate) { + default: break; + case CmpInst::FCMP_FALSE: Opnd = I->getOperand(2); break; + case CmpInst::FCMP_TRUE: Opnd = I->getOperand(1); break; + } + // No need for a select anymore - this is an unconditional move. + if (Opnd) { + unsigned OpReg = getRegForValue(Opnd); + if (OpReg == 0) + return false; + bool OpIsKill = hasTrivialKill(Opnd); + const TargetRegisterClass *RC = TLI.getRegClassFor(RetVT); + unsigned ResultReg = createResultReg(RC); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, + TII.get(TargetOpcode::COPY), ResultReg) + .addReg(OpReg, getKillRegState(OpIsKill)); + UpdateValueMap(I, ResultReg); + return true; + } + } + + // First try to use real conditional move instructions. + if (X86FastEmitCMoveSelect(RetVT, I)) + return true; + + // Try to use a sequence of SSE instructions to simulate a conditional move. + if (X86FastEmitSSESelect(RetVT, I)) + return true; + + // Fall-back to pseudo conditional move instructions, which will be later + // converted to control-flow. + if (X86FastEmitPseudoSelect(RetVT, I)) + return true; + + return false; +} + bool X86FastISel::X86SelectFPExt(const Instruction *I) { // fpext from float to double. if (X86ScalarSSEf64 && @@ -1624,8 +2152,8 @@ bool X86FastISel::TryEmitSmallMemcpy(X86AddressMode DestAM, } unsigned Reg; - bool RV = X86FastEmitLoad(VT, SrcAM, Reg); - RV &= X86FastEmitStore(VT, Reg, DestAM); + bool RV = X86FastEmitLoad(VT, SrcAM, nullptr, Reg); + RV &= X86FastEmitStore(VT, Reg, /*Kill=*/true, DestAM); assert(RV && "Failed to emit load or store??"); unsigned Size = VT.getSizeInBits()/8; @@ -1637,36 +2165,76 @@ bool X86FastISel::TryEmitSmallMemcpy(X86AddressMode DestAM, return true; } -static bool isCommutativeIntrinsic(IntrinsicInst const &I) { - switch (I.getIntrinsicID()) { - case Intrinsic::sadd_with_overflow: - case Intrinsic::uadd_with_overflow: - case Intrinsic::smul_with_overflow: - case Intrinsic::umul_with_overflow: - return true; - default: - return false; - } -} - -bool X86FastISel::X86VisitIntrinsicCall(const IntrinsicInst &I) { +bool X86FastISel::FastLowerIntrinsicCall(const IntrinsicInst *II) { // FIXME: Handle more intrinsics. - switch (I.getIntrinsicID()) { + switch (II->getIntrinsicID()) { default: return false; + case Intrinsic::frameaddress: { + Type *RetTy = II->getCalledFunction()->getReturnType(); + + MVT VT; + if (!isTypeLegal(RetTy, VT)) + return false; + + unsigned Opc; + const TargetRegisterClass *RC = nullptr; + + switch (VT.SimpleTy) { + default: llvm_unreachable("Invalid result type for frameaddress."); + case MVT::i32: Opc = X86::MOV32rm; RC = &X86::GR32RegClass; break; + case MVT::i64: Opc = X86::MOV64rm; RC = &X86::GR64RegClass; break; + } + + // This needs to be set before we call getFrameRegister, otherwise we get + // the wrong frame register. + MachineFrameInfo *MFI = FuncInfo.MF->getFrameInfo(); + MFI->setFrameAddressIsTaken(true); + + const X86RegisterInfo *RegInfo = static_cast( + TM.getSubtargetImpl()->getRegisterInfo()); + unsigned FrameReg = RegInfo->getFrameRegister(*(FuncInfo.MF)); + assert(((FrameReg == X86::RBP && VT == MVT::i64) || + (FrameReg == X86::EBP && VT == MVT::i32)) && + "Invalid Frame Register!"); + + // Always make a copy of the frame register to to a vreg first, so that we + // never directly reference the frame register (the TwoAddressInstruction- + // Pass doesn't like that). + unsigned SrcReg = createResultReg(RC); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, + TII.get(TargetOpcode::COPY), SrcReg).addReg(FrameReg); + + // Now recursively load from the frame address. + // movq (%rbp), %rax + // movq (%rax), %rax + // movq (%rax), %rax + // ... + unsigned DestReg; + unsigned Depth = cast(II->getOperand(0))->getZExtValue(); + while (Depth--) { + DestReg = createResultReg(RC); + addDirectMem(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, + TII.get(Opc), DestReg), SrcReg); + SrcReg = DestReg; + } + + UpdateValueMap(II, SrcReg); + return true; + } case Intrinsic::memcpy: { - const MemCpyInst &MCI = cast(I); + const MemCpyInst *MCI = cast(II); // Don't handle volatile or variable length memcpys. - if (MCI.isVolatile()) + if (MCI->isVolatile()) return false; - if (isa(MCI.getLength())) { + if (isa(MCI->getLength())) { // Small memcpy's are common enough that we want to do them // without a call if possible. - uint64_t Len = cast(MCI.getLength())->getZExtValue(); + uint64_t Len = cast(MCI->getLength())->getZExtValue(); if (IsMemcpySmall(Len)) { X86AddressMode DestAM, SrcAM; - if (!X86SelectAddress(MCI.getRawDest(), DestAM) || - !X86SelectAddress(MCI.getRawSource(), SrcAM)) + if (!X86SelectAddress(MCI->getRawDest(), DestAM) || + !X86SelectAddress(MCI->getRawSource(), SrcAM)) return false; TryEmitSmallMemcpy(DestAM, SrcAM, Len); return true; @@ -1674,35 +2242,35 @@ bool X86FastISel::X86VisitIntrinsicCall(const IntrinsicInst &I) { } unsigned SizeWidth = Subtarget->is64Bit() ? 64 : 32; - if (!MCI.getLength()->getType()->isIntegerTy(SizeWidth)) + if (!MCI->getLength()->getType()->isIntegerTy(SizeWidth)) return false; - if (MCI.getSourceAddressSpace() > 255 || MCI.getDestAddressSpace() > 255) + if (MCI->getSourceAddressSpace() > 255 || MCI->getDestAddressSpace() > 255) return false; - return DoSelectCall(&I, "memcpy"); + return LowerCallTo(II, "memcpy", II->getNumArgOperands() - 2); } case Intrinsic::memset: { - const MemSetInst &MSI = cast(I); + const MemSetInst *MSI = cast(II); - if (MSI.isVolatile()) + if (MSI->isVolatile()) return false; unsigned SizeWidth = Subtarget->is64Bit() ? 64 : 32; - if (!MSI.getLength()->getType()->isIntegerTy(SizeWidth)) + if (!MSI->getLength()->getType()->isIntegerTy(SizeWidth)) return false; - if (MSI.getDestAddressSpace() > 255) + if (MSI->getDestAddressSpace() > 255) return false; - return DoSelectCall(&I, "memset"); + return LowerCallTo(II, "memset", II->getNumArgOperands() - 2); } case Intrinsic::stackprotector: { // Emit code to store the stack guard onto the stack. EVT PtrTy = TLI.getPointerTy(); - const Value *Op1 = I.getArgOperand(0); // The guard's value. - const AllocaInst *Slot = cast(I.getArgOperand(1)); + const Value *Op1 = II->getArgOperand(0); // The guard's value. + const AllocaInst *Slot = cast(II->getArgOperand(1)); MFI.setStackProtectorIndex(FuncInfo.StaticAllocaMap[Slot]); @@ -1713,7 +2281,7 @@ bool X86FastISel::X86VisitIntrinsicCall(const IntrinsicInst &I) { return true; } case Intrinsic::dbg_declare: { - const DbgDeclareInst *DI = cast(&I); + const DbgDeclareInst *DI = cast(II); X86AddressMode AM; assert(DI->getAddress() && "Null address should be checked earlier!"); if (!X86SelectAddress(DI->getAddress(), AM)) @@ -1729,6 +2297,58 @@ bool X86FastISel::X86VisitIntrinsicCall(const IntrinsicInst &I) { BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::TRAP)); return true; } + case Intrinsic::sqrt: { + if (!Subtarget->hasSSE1()) + return false; + + Type *RetTy = II->getCalledFunction()->getReturnType(); + + MVT VT; + if (!isTypeLegal(RetTy, VT)) + return false; + + // Unfortunately we can't use FastEmit_r, because the AVX version of FSQRT + // is not generated by FastISel yet. + // FIXME: Update this code once tablegen can handle it. + static const unsigned SqrtOpc[2][2] = { + {X86::SQRTSSr, X86::VSQRTSSr}, + {X86::SQRTSDr, X86::VSQRTSDr} + }; + bool HasAVX = Subtarget->hasAVX(); + unsigned Opc; + const TargetRegisterClass *RC; + switch (VT.SimpleTy) { + default: return false; + case MVT::f32: Opc = SqrtOpc[0][HasAVX]; RC = &X86::FR32RegClass; break; + case MVT::f64: Opc = SqrtOpc[1][HasAVX]; RC = &X86::FR64RegClass; break; + } + + const Value *SrcVal = II->getArgOperand(0); + unsigned SrcReg = getRegForValue(SrcVal); + + if (SrcReg == 0) + return false; + + unsigned ImplicitDefReg = 0; + if (HasAVX) { + ImplicitDefReg = createResultReg(RC); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, + TII.get(TargetOpcode::IMPLICIT_DEF), ImplicitDefReg); + } + + unsigned ResultReg = createResultReg(RC); + MachineInstrBuilder MIB; + MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), + ResultReg); + + if (ImplicitDefReg) + MIB.addReg(ImplicitDefReg); + + MIB.addReg(SrcReg); + + UpdateValueMap(II, ResultReg); + return true; + } case Intrinsic::sadd_with_overflow: case Intrinsic::uadd_with_overflow: case Intrinsic::ssub_with_overflow: @@ -1736,8 +2356,8 @@ bool X86FastISel::X86VisitIntrinsicCall(const IntrinsicInst &I) { case Intrinsic::smul_with_overflow: case Intrinsic::umul_with_overflow: { // This implements the basic lowering of the xalu with overflow intrinsics - // into add/sub/mul folowed by either seto or setb. - const Function *Callee = I.getCalledFunction(); + // into add/sub/mul followed by either seto or setb. + const Function *Callee = II->getCalledFunction(); auto *Ty = cast(Callee->getReturnType()); Type *RetTy = Ty->getTypeAtIndex(0U); Type *CondTy = Ty->getTypeAtIndex(1); @@ -1749,27 +2369,35 @@ bool X86FastISel::X86VisitIntrinsicCall(const IntrinsicInst &I) { if (VT < MVT::i8 || VT > MVT::i64) return false; - const Value *LHS = I.getArgOperand(0); - const Value *RHS = I.getArgOperand(1); + const Value *LHS = II->getArgOperand(0); + const Value *RHS = II->getArgOperand(1); - // Canonicalize immediates to the RHS. + // Canonicalize immediate to the RHS. if (isa(LHS) && !isa(RHS) && - isCommutativeIntrinsic(I)) + isCommutativeIntrinsic(II)) std::swap(LHS, RHS); + bool UseIncDec = false; + if (isa(RHS) && cast(RHS)->isOne()) + UseIncDec = true; + unsigned BaseOpc, CondOpc; - switch (I.getIntrinsicID()) { + switch (II->getIntrinsicID()) { default: llvm_unreachable("Unexpected intrinsic!"); case Intrinsic::sadd_with_overflow: - BaseOpc = ISD::ADD; CondOpc = X86::SETOr; break; + BaseOpc = UseIncDec ? unsigned(X86ISD::INC) : unsigned(ISD::ADD); + CondOpc = X86::SETOr; + break; case Intrinsic::uadd_with_overflow: BaseOpc = ISD::ADD; CondOpc = X86::SETBr; break; case Intrinsic::ssub_with_overflow: - BaseOpc = ISD::SUB; CondOpc = X86::SETOr; break; + BaseOpc = UseIncDec ? unsigned(X86ISD::DEC) : unsigned(ISD::SUB); + CondOpc = X86::SETOr; + break; case Intrinsic::usub_with_overflow: BaseOpc = ISD::SUB; CondOpc = X86::SETBr; break; case Intrinsic::smul_with_overflow: - BaseOpc = ISD::MUL; CondOpc = X86::SETOr; break; + BaseOpc = X86ISD::SMUL; CondOpc = X86::SETOr; break; case Intrinsic::umul_with_overflow: BaseOpc = X86ISD::UMUL; CondOpc = X86::SETOr; break; } @@ -1781,9 +2409,24 @@ bool X86FastISel::X86VisitIntrinsicCall(const IntrinsicInst &I) { unsigned ResultReg = 0; // Check if we have an immediate version. - if (auto const *C = dyn_cast(RHS)) { - ResultReg = FastEmit_ri(VT, VT, BaseOpc, LHSReg, LHSIsKill, - C->getZExtValue()); + if (const auto *CI = dyn_cast(RHS)) { + static const unsigned Opc[2][2][4] = { + { { X86::INC8r, X86::INC16r, X86::INC32r, X86::INC64r }, + { X86::DEC8r, X86::DEC16r, X86::DEC32r, X86::DEC64r } }, + { { X86::INC8r, X86::INC64_16r, X86::INC64_32r, X86::INC64r }, + { X86::DEC8r, X86::DEC64_16r, X86::DEC64_32r, X86::DEC64r } } + }; + + if (BaseOpc == X86ISD::INC || BaseOpc == X86ISD::DEC) { + ResultReg = createResultReg(TLI.getRegClassFor(VT)); + bool Is64Bit = Subtarget->is64Bit(); + bool IsDec = BaseOpc == X86ISD::DEC; + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, + TII.get(Opc[Is64Bit][IsDec][VT.SimpleTy-MVT::i8]), ResultReg) + .addReg(LHSReg, getKillRegState(LHSIsKill)); + } else + ResultReg = FastEmit_ri(VT, VT, BaseOpc, LHSReg, LHSIsKill, + CI->getZExtValue()); } unsigned RHSReg; @@ -1797,10 +2440,11 @@ bool X86FastISel::X86VisitIntrinsicCall(const IntrinsicInst &I) { RHSIsKill); } - // FastISel doesn't have a pattern for X86::MUL*r. Emit it manually. + // FastISel doesn't have a pattern for all X86::MUL*r and X86::IMUL*r. Emit + // it manually. if (BaseOpc == X86ISD::UMUL && !ResultReg) { static const unsigned MULOpc[] = - { X86::MUL8r, X86::MUL16r, X86::MUL32r, X86::MUL64r }; + { X86::MUL8r, X86::MUL16r, X86::MUL32r, X86::MUL64r }; static const unsigned Reg[] = { X86::AL, X86::AX, X86::EAX, X86::RAX }; // First copy the first operand into RAX, which is an implicit input to // the X86::MUL*r instruction. @@ -1809,6 +2453,21 @@ bool X86FastISel::X86VisitIntrinsicCall(const IntrinsicInst &I) { .addReg(LHSReg, getKillRegState(LHSIsKill)); ResultReg = FastEmitInst_r(MULOpc[VT.SimpleTy-MVT::i8], TLI.getRegClassFor(VT), RHSReg, RHSIsKill); + } else if (BaseOpc == X86ISD::SMUL && !ResultReg) { + static const unsigned MULOpc[] = + { X86::IMUL8r, X86::IMUL16rr, X86::IMUL32rr, X86::IMUL64rr }; + if (VT == MVT::i8) { + // Copy the first operand into AL, which is an implicit input to the + // X86::IMUL8r instruction. + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, + TII.get(TargetOpcode::COPY), X86::AL) + .addReg(LHSReg, getKillRegState(LHSIsKill)); + ResultReg = FastEmitInst_r(MULOpc[0], TLI.getRegClassFor(VT), RHSReg, + RHSIsKill); + } else + ResultReg = FastEmitInst_rr(MULOpc[VT.SimpleTy-MVT::i8], + TLI.getRegClassFor(VT), LHSReg, LHSIsKill, + RHSReg, RHSIsKill); } if (!ResultReg) @@ -1819,7 +2478,73 @@ bool X86FastISel::X86VisitIntrinsicCall(const IntrinsicInst &I) { BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(CondOpc), ResultReg2); - UpdateValueMap(&I, ResultReg, 2); + UpdateValueMap(II, ResultReg, 2); + return true; + } + case Intrinsic::x86_sse_cvttss2si: + case Intrinsic::x86_sse_cvttss2si64: + case Intrinsic::x86_sse2_cvttsd2si: + case Intrinsic::x86_sse2_cvttsd2si64: { + bool IsInputDouble; + switch (II->getIntrinsicID()) { + default: llvm_unreachable("Unexpected intrinsic."); + case Intrinsic::x86_sse_cvttss2si: + case Intrinsic::x86_sse_cvttss2si64: + if (!Subtarget->hasSSE1()) + return false; + IsInputDouble = false; + break; + case Intrinsic::x86_sse2_cvttsd2si: + case Intrinsic::x86_sse2_cvttsd2si64: + if (!Subtarget->hasSSE2()) + return false; + IsInputDouble = true; + break; + } + + Type *RetTy = II->getCalledFunction()->getReturnType(); + MVT VT; + if (!isTypeLegal(RetTy, VT)) + return false; + + static const unsigned CvtOpc[2][2][2] = { + { { X86::CVTTSS2SIrr, X86::VCVTTSS2SIrr }, + { X86::CVTTSS2SI64rr, X86::VCVTTSS2SI64rr } }, + { { X86::CVTTSD2SIrr, X86::VCVTTSD2SIrr }, + { X86::CVTTSD2SI64rr, X86::VCVTTSD2SI64rr } } + }; + bool HasAVX = Subtarget->hasAVX(); + unsigned Opc; + switch (VT.SimpleTy) { + default: llvm_unreachable("Unexpected result type."); + case MVT::i32: Opc = CvtOpc[IsInputDouble][0][HasAVX]; break; + case MVT::i64: Opc = CvtOpc[IsInputDouble][1][HasAVX]; break; + } + + // Check if we can fold insertelement instructions into the convert. + const Value *Op = II->getArgOperand(0); + while (auto *IE = dyn_cast(Op)) { + const Value *Index = IE->getOperand(2); + if (!isa(Index)) + break; + unsigned Idx = cast(Index)->getZExtValue(); + + if (Idx == 0) { + Op = IE->getOperand(1); + break; + } + Op = IE->getOperand(0); + } + + unsigned Reg = getRegForValue(Op); + if (Reg == 0) + return false; + + unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT)); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg) + .addReg(Reg); + + UpdateValueMap(II, ResultReg); return true; } } @@ -1844,31 +2569,43 @@ bool X86FastISel::FastLowerArguments() { 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; - + unsigned GPRCnt = 0; + unsigned FPRCnt = 0; + unsigned Idx = 0; + for (auto const &Arg : F->args()) { + // The first argument is at index 1. + ++Idx; 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(); + Type *ArgTy = Arg.getType(); if (ArgTy->isStructTy() || ArgTy->isArrayTy() || ArgTy->isVectorTy()) return false; EVT ArgVT = TLI.getValueType(ArgTy); if (!ArgVT.isSimple()) return false; switch (ArgVT.getSimpleVT().SimpleTy) { + default: return false; case MVT::i32: case MVT::i64: + ++GPRCnt; + break; + case MVT::f32: + case MVT::f64: + if (!Subtarget->hasSSE1()) + return false; + ++FPRCnt; break; - default: - return false; } + + if (GPRCnt > 6) + return false; + + if (FPRCnt > 8) + return false; } static const MCPhysReg GPR32ArgRegs[] = { @@ -1877,75 +2614,83 @@ bool X86FastISel::FastLowerArguments() { static const MCPhysReg GPR64ArgRegs[] = { X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8 , X86::R9 }; + static const MCPhysReg XMMArgRegs[] = { + X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3, + X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7 + }; - 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 GPRIdx = 0; + unsigned FPRIdx = 0; + for (auto const &Arg : F->args()) { + MVT VT = TLI.getSimpleValueType(Arg.getType()); + const TargetRegisterClass *RC = TLI.getRegClassFor(VT); + unsigned SrcReg; + switch (VT.SimpleTy) { + default: llvm_unreachable("Unexpected value type."); + case MVT::i32: SrcReg = GPR32ArgRegs[GPRIdx++]; break; + case MVT::i64: SrcReg = GPR64ArgRegs[GPRIdx++]; break; + case MVT::f32: // fall-through + case MVT::f64: SrcReg = XMMArgRegs[FPRIdx++]; break; + } 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, DbgLoc, - TII.get(TargetOpcode::COPY), - ResultReg).addReg(DstReg, getKillRegState(true)); - UpdateValueMap(I, ResultReg); + TII.get(TargetOpcode::COPY), ResultReg) + .addReg(DstReg, getKillRegState(true)); + UpdateValueMap(&Arg, ResultReg); } return true; } -bool X86FastISel::X86SelectCall(const Instruction *I) { - const CallInst *CI = cast(I); - const Value *Callee = CI->getCalledValue(); - - // Can't handle inline asm yet. - if (isa(Callee)) - return false; - - // Handle intrinsic calls. - 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, nullptr); -} - -static unsigned computeBytesPoppedByCallee(const X86Subtarget &Subtarget, - const ImmutableCallSite &CS) { - if (Subtarget.is64Bit()) +static unsigned computeBytesPoppedByCallee(const X86Subtarget *Subtarget, + CallingConv::ID CC, + ImmutableCallSite *CS) { + if (Subtarget->is64Bit()) return 0; - if (Subtarget.getTargetTriple().isOSMSVCRT()) + if (Subtarget->getTargetTriple().isOSMSVCRT()) return 0; - CallingConv::ID CC = CS.getCallingConv(); - if (CC == CallingConv::Fast || CC == CallingConv::GHC) + if (CC == CallingConv::Fast || CC == CallingConv::GHC || + CC == CallingConv::HiPE) return 0; - if (!CS.paramHasAttr(1, Attribute::StructRet)) + if (CS && !CS->paramHasAttr(1, Attribute::StructRet)) return 0; - if (CS.paramHasAttr(1, Attribute::InReg)) + if (CS && 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); - const Value *Callee = CI->getCalledValue(); - - // 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_64_Win64 && - CC != CallingConv::X86_64_SysV) +bool X86FastISel::FastLowerCall(CallLoweringInfo &CLI) { + auto &OutVals = CLI.OutVals; + auto &OutFlags = CLI.OutFlags; + auto &OutRegs = CLI.OutRegs; + auto &Ins = CLI.Ins; + auto &InRegs = CLI.InRegs; + CallingConv::ID CC = CLI.CallConv; + bool &IsTailCall = CLI.IsTailCall; + bool IsVarArg = CLI.IsVarArg; + const Value *Callee = CLI.Callee; + const char *SymName = CLI.SymName; + + bool Is64Bit = Subtarget->is64Bit(); + bool IsWin64 = Subtarget->isCallingConvWin64(CC); + + // Handle only C, fastcc, and webkit_js calling conventions for now. + switch (CC) { + default: return false; + case CallingConv::C: + case CallingConv::Fast: + case CallingConv::WebKit_JS: + case CallingConv::X86_FastCall: + case CallingConv::X86_64_Win64: + case CallingConv::X86_64_SysV: + break; + } + + // Allow SelectionDAG isel to handle tail calls. + if (IsTailCall) return false; // fastcc with -tailcallopt is intended to provide a guaranteed @@ -1953,150 +2698,77 @@ bool X86FastISel::DoSelectCall(const Instruction *I, const char *MemIntName) { if (CC == CallingConv::Fast && TM.Options.GuaranteedTailCallOpt) return false; - 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 && isWin64) + // x86-32. Special handling for x86-64 is implemented. + if (IsVarArg && IsWin64) return false; // Don't know about inalloca yet. - if (CS.hasInAllocaArgument()) + if (CLI.CS && CLI.CS->hasInAllocaArgument()) return false; // Fast-isel doesn't know about callee-pop yet. - if (X86::isCalleePop(CC, Subtarget->is64Bit(), isVarArg, + if (X86::isCalleePop(CC, Subtarget->is64Bit(), IsVarArg, TM.Options.GuaranteedTailCallOpt)) return false; - // Check whether the function can return without sret-demotion. - SmallVector Outs; - GetReturnInfo(I->getType(), CS.getAttributes(), Outs, TLI); - bool CanLowerReturn = TLI.CanLowerReturn(CS.getCallingConv(), - *FuncInfo.MF, FTy->isVarArg(), - Outs, FTy->getContext()); - if (!CanLowerReturn) - return false; - - // Materialize callee address in a register. FIXME: GV address can be - // handled with a CALLpcrel32 instead. - X86AddressMode CalleeAM; - if (!X86SelectCallAddress(Callee, CalleeAM)) - return false; - unsigned CalleeOp = 0; - const GlobalValue *GV = nullptr; - if (CalleeAM.GV != nullptr) { - GV = CalleeAM.GV; - } else if (CalleeAM.Base.Reg != 0) { - CalleeOp = CalleeAM.Base.Reg; - } else - return false; - - // Deal with call operands first. - SmallVector ArgVals; - SmallVector Args; - SmallVector ArgVTs; - SmallVector ArgFlags; - 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 - // last two arguments, which should not passed to the underlying functions. - if (MemIntName && e-i <= 2) - break; - Value *ArgVal = *i; - ISD::ArgFlagsTy Flags; - unsigned AttrInd = i - CS.arg_begin() + 1; - if (CS.paramHasAttr(AttrInd, Attribute::SExt)) - Flags.setSExt(); - if (CS.paramHasAttr(AttrInd, Attribute::ZExt)) - Flags.setZExt(); - - if (CS.paramHasAttr(AttrInd, Attribute::ByVal)) { - PointerType *Ty = cast(ArgVal->getType()); - Type *ElementTy = Ty->getElementType(); - unsigned FrameSize = DL.getTypeAllocSize(ElementTy); - unsigned FrameAlign = CS.getParamAlignment(AttrInd); - if (!FrameAlign) - FrameAlign = TLI.getByValTypeAlignment(ElementTy); - Flags.setByVal(); - Flags.setByValSize(FrameSize); - Flags.setByValAlign(FrameAlign); - if (!IsMemcpySmall(FrameSize)) - return false; - } - - if (CS.paramHasAttr(AttrInd, Attribute::InReg)) - Flags.setInReg(); - if (CS.paramHasAttr(AttrInd, Attribute::Nest)) - Flags.setNest(); - - // If this is an i1/i8/i16 argument, promote to i32 to avoid an extra - // instruction. This is safe because it is common to all fastisel supported - // calling conventions on x86. - if (ConstantInt *CI = dyn_cast(ArgVal)) { - if (CI->getBitWidth() == 1 || CI->getBitWidth() == 8 || - CI->getBitWidth() == 16) { + // If this is a constant i1/i8/i16 argument, promote to i32 to avoid an extra + // instruction. This is safe because it is common to all FastISel supported + // calling conventions on x86. + for (int i = 0, e = OutVals.size(); i != e; ++i) { + Value *&Val = OutVals[i]; + ISD::ArgFlagsTy Flags = OutFlags[i]; + if (auto *CI = dyn_cast(Val)) { + if (CI->getBitWidth() < 32) { if (Flags.isSExt()) - ArgVal = ConstantExpr::getSExt(CI,Type::getInt32Ty(CI->getContext())); + Val = ConstantExpr::getSExt(CI, Type::getInt32Ty(CI->getContext())); else - ArgVal = ConstantExpr::getZExt(CI,Type::getInt32Ty(CI->getContext())); + Val = ConstantExpr::getZExt(CI, Type::getInt32Ty(CI->getContext())); } } - unsigned ArgReg; - // Passing bools around ends up doing a trunc to i1 and passing it. // Codegen this as an argument + "and 1". - if (ArgVal->getType()->isIntegerTy(1) && isa(ArgVal) && - cast(ArgVal)->getParent() == I->getParent() && - ArgVal->hasOneUse()) { - ArgVal = cast(ArgVal)->getOperand(0); - ArgReg = getRegForValue(ArgVal); - if (ArgReg == 0) return false; - - MVT ArgVT; - if (!isTypeLegal(ArgVal->getType(), ArgVT)) return false; - - ArgReg = FastEmit_ri(ArgVT, ArgVT, ISD::AND, ArgReg, - ArgVal->hasOneUse(), 1); - } else { - ArgReg = getRegForValue(ArgVal); - } + if (auto *TI = dyn_cast(Val)) { + if (TI->getType()->isIntegerTy(1) && CLI.CS && + (TI->getParent() == CLI.CS->getInstruction()->getParent()) && + TI->hasOneUse()) { + Val = cast(Val)->getOperand(0); + unsigned ResultReg = getRegForValue(Val); + + if (!ResultReg) + return false; - if (ArgReg == 0) return false; + MVT ArgVT; + if (!isTypeLegal(Val->getType(), ArgVT)) + return false; - Type *ArgTy = ArgVal->getType(); - MVT ArgVT; - if (!isTypeLegal(ArgTy, ArgVT)) - return false; - if (ArgVT == MVT::x86mmx) - return false; - unsigned OriginalAlignment = DL.getABITypeAlignment(ArgTy); - Flags.setOrigAlign(OriginalAlignment); + ResultReg = + FastEmit_ri(ArgVT, ArgVT, ISD::AND, ResultReg, Val->hasOneUse(), 1); - Args.push_back(ArgReg); - ArgVals.push_back(ArgVal); - ArgVTs.push_back(ArgVT); - ArgFlags.push_back(Flags); + if (!ResultReg) + return false; + UpdateValueMap(Val, ResultReg); + } + } } // Analyze operands of the call, assigning locations to each operand. SmallVector ArgLocs; - CCState CCInfo(CC, isVarArg, *FuncInfo.MF, TM, ArgLocs, - I->getParent()->getContext()); + CCState CCInfo(CC, IsVarArg, *FuncInfo.MF, ArgLocs, CLI.RetTy->getContext()); // Allocate shadow area for Win64 - if (isWin64) + if (IsWin64) CCInfo.AllocateStack(32, 8); - CCInfo.AnalyzeCallOperands(ArgVTs, ArgFlags, CC_X86); + SmallVector OutVTs; + for (auto *Val : OutVals) { + MVT VT; + if (!isTypeLegal(Val->getType(), VT)) + return false; + OutVTs.push_back(VT); + } + CCInfo.AnalyzeCallOperands(OutVTs, OutFlags, CC_X86); // Get a count of how many bytes are to be pushed on the stack. unsigned NumBytes = CCInfo.getNextStackOffset(); @@ -2106,13 +2778,20 @@ bool X86FastISel::DoSelectCall(const Instruction *I, const char *MemIntName) { BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AdjStackDown)) .addImm(NumBytes); - // Process argument: walk the register/memloc assignments, inserting - // copies / loads. - SmallVector RegArgs; + // Walk the register/memloc assignments, inserting copies/loads. + const X86RegisterInfo *RegInfo = static_cast( + TM.getSubtargetImpl()->getRegisterInfo()); for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { - CCValAssign &VA = ArgLocs[i]; - unsigned Arg = Args[VA.getValNo()]; - EVT ArgVT = ArgVTs[VA.getValNo()]; + CCValAssign const &VA = ArgLocs[i]; + const Value *ArgVal = OutVals[VA.getValNo()]; + MVT ArgVT = OutVTs[VA.getValNo()]; + + if (ArgVT == MVT::x86mmx) + return false; + + unsigned ArgReg = getRegForValue(ArgVal); + if (!ArgReg) + return false; // Promote the value if needed. switch (VA.getLocInfo()) { @@ -2120,8 +2799,8 @@ bool X86FastISel::DoSelectCall(const Instruction *I, const char *MemIntName) { case CCValAssign::SExt: { assert(VA.getLocVT().isInteger() && !VA.getLocVT().isVector() && "Unexpected extend"); - bool Emitted = X86FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(), - Arg, ArgVT, Arg); + bool Emitted = X86FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(), ArgReg, + ArgVT, ArgReg); assert(Emitted && "Failed to emit a sext!"); (void)Emitted; ArgVT = VA.getLocVT(); break; @@ -2129,8 +2808,8 @@ bool X86FastISel::DoSelectCall(const Instruction *I, const char *MemIntName) { case CCValAssign::ZExt: { assert(VA.getLocVT().isInteger() && !VA.getLocVT().isVector() && "Unexpected extend"); - bool Emitted = X86FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(), - Arg, ArgVT, Arg); + bool Emitted = X86FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(), ArgReg, + ArgVT, ArgReg); assert(Emitted && "Failed to emit a zext!"); (void)Emitted; ArgVT = VA.getLocVT(); break; @@ -2138,66 +2817,72 @@ bool X86FastISel::DoSelectCall(const Instruction *I, const char *MemIntName) { case CCValAssign::AExt: { assert(VA.getLocVT().isInteger() && !VA.getLocVT().isVector() && "Unexpected extend"); - bool Emitted = X86FastEmitExtend(ISD::ANY_EXTEND, VA.getLocVT(), - Arg, ArgVT, Arg); + bool Emitted = X86FastEmitExtend(ISD::ANY_EXTEND, VA.getLocVT(), ArgReg, + ArgVT, ArgReg); if (!Emitted) - Emitted = X86FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(), - Arg, ArgVT, Arg); + Emitted = X86FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(), ArgReg, + ArgVT, ArgReg); if (!Emitted) - Emitted = X86FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(), - Arg, ArgVT, Arg); + Emitted = X86FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(), ArgReg, + ArgVT, ArgReg); assert(Emitted && "Failed to emit a aext!"); (void)Emitted; ArgVT = VA.getLocVT(); break; } case CCValAssign::BCvt: { - unsigned BC = FastEmit_r(ArgVT.getSimpleVT(), VA.getLocVT(), - ISD::BITCAST, Arg, /*TODO: Kill=*/false); - assert(BC != 0 && "Failed to emit a bitcast!"); - Arg = BC; + ArgReg = FastEmit_r(ArgVT, VA.getLocVT(), ISD::BITCAST, ArgReg, + /*TODO: Kill=*/false); + assert(ArgReg && "Failed to emit a bitcast!"); ArgVT = VA.getLocVT(); break; } - case CCValAssign::VExt: + 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::FPExt: + llvm_unreachable("Unexpected loc info!"); case CCValAssign::Indirect: // FIXME: Indirect doesn't need extending, but fast-isel doesn't fully // support this. return false; - case CCValAssign::FPExt: - llvm_unreachable("Unexpected loc info!"); } if (VA.isRegLoc()) { BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, - TII.get(TargetOpcode::COPY), VA.getLocReg()).addReg(Arg); - RegArgs.push_back(VA.getLocReg()); + TII.get(TargetOpcode::COPY), VA.getLocReg()).addReg(ArgReg); + OutRegs.push_back(VA.getLocReg()); } else { + assert(VA.isMemLoc()); + + // Don't emit stores for undef values. + if (isa(ArgVal)) + continue; + unsigned LocMemOffset = VA.getLocMemOffset(); X86AddressMode AM; - 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()]; - + ISD::ArgFlagsTy Flags = OutFlags[VA.getValNo()]; + unsigned Alignment = DL.getABITypeAlignment(ArgVal->getType()); + MachineMemOperand *MMO = FuncInfo.MF->getMachineMemOperand( + MachinePointerInfo::getStack(LocMemOffset), MachineMemOperand::MOStore, + ArgVT.getStoreSize(), Alignment); if (Flags.isByVal()) { X86AddressMode SrcAM; - SrcAM.Base.Reg = Arg; - bool Res = TryEmitSmallMemcpy(AM, SrcAM, Flags.getByValSize()); - assert(Res && "memcpy length already checked!"); (void)Res; + SrcAM.Base.Reg = ArgReg; + if (!TryEmitSmallMemcpy(AM, SrcAM, Flags.getByValSize())) + return false; } 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, // as it can cause us to reevaluate the argument. - if (!X86FastEmitStore(ArgVT, ArgVal, AM)) + if (!X86FastEmitStore(ArgVT, ArgVal, AM, MMO)) return false; } else { - if (!X86FastEmitStore(ArgVT, Arg, AM)) + bool ValIsKill = hasTrivialKill(ArgVal); + if (!X86FastEmitStore(ArgVT, ArgReg, ValIsKill, AM, MMO)) return false; } } @@ -2211,37 +2896,53 @@ bool X86FastISel::DoSelectCall(const Instruction *I, const char *MemIntName) { TII.get(TargetOpcode::COPY), X86::EBX).addReg(Base); } - if (Subtarget->is64Bit() && isVarArg && !isWin64) { + if (Is64Bit && IsVarArg && !IsWin64) { + // From AMD64 ABI document: + // For calls that may call functions that use varargs or stdargs + // (prototype-less calls or calls to functions containing ellipsis (...) in + // the declaration) %al is used as hidden argument to specify the number + // of SSE registers used. The contents of %al do not need to match exactly + // the number of registers, but must be an ubound on the number of SSE + // registers used and is in the range 0 - 8 inclusive. + // Count the number of XMM registers allocated. static const MCPhysReg XMMArgRegs[] = { X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3, X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7 }; unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8); + assert((Subtarget->hasSSE1() || !NumXMMRegs) + && "SSE registers cannot be used when SSE is disabled"); BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::MOV8ri), X86::AL).addImm(NumXMMRegs); } + // Materialize callee address in a register. FIXME: GV address can be + // handled with a CALLpcrel32 instead. + X86AddressMode CalleeAM; + if (!X86SelectCallAddress(Callee, CalleeAM)) + return false; + + unsigned CalleeOp = 0; + const GlobalValue *GV = nullptr; + if (CalleeAM.GV != nullptr) { + GV = CalleeAM.GV; + } else if (CalleeAM.Base.Reg != 0) { + CalleeOp = CalleeAM.Base.Reg; + } else + return false; + // Issue the call. MachineInstrBuilder MIB; if (CalleeOp) { // Register-indirect call. - unsigned CallOpc; - if (Subtarget->is64Bit()) - CallOpc = X86::CALL64r; - else - CallOpc = X86::CALL32r; + unsigned CallOpc = Is64Bit ? X86::CALL64r : X86::CALL32r; MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(CallOpc)) .addReg(CalleeOp); - } else { // Direct call. assert(GV && "Not a direct call"); - unsigned CallOpc; - if (Subtarget->is64Bit()) - CallOpc = X86::CALL64pcrel32; - else - CallOpc = X86::CALLpcrel32; + unsigned CallOpc = Is64Bit ? X86::CALL64pcrel32 : X86::CALLpcrel32; // See if we need any target-specific flags on the GV operand. unsigned char OpFlags = 0; @@ -2264,92 +2965,72 @@ bool X86FastISel::DoSelectCall(const Instruction *I, const char *MemIntName) { OpFlags = X86II::MO_DARWIN_STUB; } - MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(CallOpc)); - if (MemIntName) - MIB.addExternalSymbol(MemIntName, OpFlags); + if (SymName) + MIB.addExternalSymbol(SymName, OpFlags); else MIB.addGlobalAddress(GV, 0, OpFlags); } - // Add a register mask with the call-preserved registers. + // Add a register mask operand representing the call-preserved registers. // Proper defs for return values will be added by setPhysRegsDeadExcept(). - MIB.addRegMask(TRI.getCallPreservedMask(CS.getCallingConv())); + MIB.addRegMask(TRI.getCallPreservedMask(CC)); // Add an implicit use GOT pointer in EBX. if (Subtarget->isPICStyleGOT()) MIB.addReg(X86::EBX, RegState::Implicit); - if (Subtarget->is64Bit() && isVarArg && !isWin64) + if (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], RegState::Implicit); + for (auto Reg : OutRegs) + MIB.addReg(Reg, RegState::Implicit); // Issue CALLSEQ_END + unsigned NumBytesForCalleeToPop = + computeBytesPoppedByCallee(Subtarget, CC, CLI.CS); unsigned AdjStackUp = TII.getCallFrameDestroyOpcode(); - const unsigned NumBytesCallee = computeBytesPoppedByCallee(*Subtarget, CS); BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AdjStackUp)) - .addImm(NumBytes).addImm(NumBytesCallee); - - // Build info for return calling conv lowering code. - // FIXME: This is practically a copy-paste from TargetLowering::LowerCallTo. - SmallVector Ins; - SmallVector RetTys; - ComputeValueVTs(TLI, I->getType(), RetTys); - for (unsigned i = 0, e = RetTys.size(); i != e; ++i) { - EVT VT = RetTys[i]; - 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; - MyFlags.Used = !CS.getInstruction()->use_empty(); - if (CS.paramHasAttr(0, Attribute::SExt)) - MyFlags.Flags.setSExt(); - if (CS.paramHasAttr(0, Attribute::ZExt)) - MyFlags.Flags.setZExt(); - if (CS.paramHasAttr(0, Attribute::InReg)) - MyFlags.Flags.setInReg(); - Ins.push_back(MyFlags); - } - } + .addImm(NumBytes).addImm(NumBytesForCalleeToPop); // Now handle call return values. - SmallVector UsedRegs; SmallVector RVLocs; - CCState CCRetInfo(CC, false, *FuncInfo.MF, TM, RVLocs, - I->getParent()->getContext()); - unsigned ResultReg = FuncInfo.CreateRegs(I->getType()); + CCState CCRetInfo(CC, IsVarArg, *FuncInfo.MF, RVLocs, + CLI.RetTy->getContext()); CCRetInfo.AnalyzeCallResult(Ins, RetCC_X86); + + // Copy all of the result registers out of their specified physreg. + unsigned ResultReg = FuncInfo.CreateRegs(CLI.RetTy); for (unsigned i = 0; i != RVLocs.size(); ++i) { - EVT CopyVT = RVLocs[i].getValVT(); + CCValAssign &VA = RVLocs[i]; + EVT CopyVT = VA.getValVT(); unsigned CopyReg = ResultReg + i; - // If this is a call to a function that returns an fp value on the x87 fp - // 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)) { - if (isScalarFPTypeInSSEReg(RVLocs[i].getValVT())) { - CopyVT = MVT::f80; - CopyReg = createResultReg(&X86::RFP80RegClass); - } - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, - TII.get(X86::FpPOP_RETVAL), CopyReg); - } else { - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, - TII.get(TargetOpcode::COPY), - CopyReg).addReg(RVLocs[i].getLocReg()); - UsedRegs.push_back(RVLocs[i].getLocReg()); + // If this is x86-64, and we disabled SSE, we can't return FP values + if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) && + ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) { + report_fatal_error("SSE register return with SSE disabled"); } - 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 - // then loading it back. Ewww... - EVT ResVT = RVLocs[i].getValVT(); + // If 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 ((VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) && + isScalarFPTypeInSSEReg(VA.getValVT())) { + CopyVT = MVT::f80; + CopyReg = createResultReg(&X86::RFP80RegClass); + } + + // Copy out the result. + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, + TII.get(TargetOpcode::COPY), CopyReg).addReg(VA.getLocReg()); + InRegs.push_back(VA.getLocReg()); + + // Round the f80 to the right size, which also moves it to the appropriate + // xmm register. This is accomplished by storing the f80 value in memory + // and then loading it back. + if (CopyVT != VA.getValVT()) { + EVT ResVT = VA.getValVT(); unsigned Opc = ResVT == MVT::f32 ? X86::ST_Fp80m32 : X86::ST_Fp80m64; unsigned MemSize = ResVT.getSizeInBits()/8; int FI = MFI.CreateStackObject(MemSize, MemSize, false); @@ -2362,16 +3043,13 @@ bool X86FastISel::DoSelectCall(const Instruction *I, const char *MemIntName) { } } - if (RVLocs.size()) - UpdateValueMap(I, ResultReg, RVLocs.size()); - - // Set all unused physreg defs as dead. - static_cast(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI); + CLI.ResultReg = ResultReg; + CLI.NumResultRegs = RVLocs.size(); + CLI.Call = MIB; return true; } - bool X86FastISel::TargetSelectInstruction(const Instruction *I) { switch (I->getOpcode()) { @@ -2389,8 +3067,6 @@ X86FastISel::TargetSelectInstruction(const Instruction *I) { return X86SelectZExt(I); case Instruction::Br: return X86SelectBranch(I); - case Instruction::Call: - return X86SelectCall(I); case Instruction::LShr: case Instruction::AShr: case Instruction::Shl: @@ -2426,11 +3102,66 @@ X86FastISel::TargetSelectInstruction(const Instruction *I) { return false; } -unsigned X86FastISel::TargetMaterializeConstant(const Constant *C) { - MVT VT; - if (!isTypeLegal(C->getType(), VT)) +unsigned X86FastISel::X86MaterializeInt(const ConstantInt *CI, MVT VT) { + if (VT > MVT::i64) return 0; + uint64_t Imm = CI->getZExtValue(); + if (Imm == 0) { + unsigned SrcReg = FastEmitInst_(X86::MOV32r0, &X86::GR32RegClass); + switch (VT.SimpleTy) { + default: llvm_unreachable("Unexpected value type"); + case MVT::i1: + case MVT::i8: + return FastEmitInst_extractsubreg(MVT::i8, SrcReg, /*Kill=*/true, + X86::sub_8bit); + case MVT::i16: + return FastEmitInst_extractsubreg(MVT::i16, SrcReg, /*Kill=*/true, + X86::sub_16bit); + case MVT::i32: + return SrcReg; + case MVT::i64: { + unsigned ResultReg = createResultReg(&X86::GR64RegClass); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, + TII.get(TargetOpcode::SUBREG_TO_REG), ResultReg) + .addImm(0).addReg(SrcReg).addImm(X86::sub_32bit); + return ResultReg; + } + } + } + + unsigned Opc = 0; + switch (VT.SimpleTy) { + default: llvm_unreachable("Unexpected value type"); + case MVT::i1: VT = MVT::i8; // fall-through + case MVT::i8: Opc = X86::MOV8ri; break; + case MVT::i16: Opc = X86::MOV16ri; break; + case MVT::i32: Opc = X86::MOV32ri; break; + case MVT::i64: { + if (isUInt<32>(Imm)) + Opc = X86::MOV32ri; + else if (isInt<32>(Imm)) + Opc = X86::MOV64ri32; + else + Opc = X86::MOV64ri; + break; + } + } + if (VT == MVT::i64 && Opc == X86::MOV32ri) { + unsigned SrcReg = FastEmitInst_i(Opc, &X86::GR32RegClass, Imm); + unsigned ResultReg = createResultReg(&X86::GR64RegClass); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, + TII.get(TargetOpcode::SUBREG_TO_REG), ResultReg) + .addImm(0).addReg(SrcReg).addImm(X86::sub_32bit); + return ResultReg; + } + return FastEmitInst_i(Opc, TLI.getRegClassFor(VT), Imm); +} + +unsigned X86FastISel::X86MaterializeFP(const ConstantFP *CFP, MVT VT) { + if (CFP->isNullValue()) + return TargetMaterializeFloatZero(CFP); + // Can't handle alternate code models yet. if (TM.getCodeModel() != CodeModel::Small) return 0; @@ -2440,23 +3171,6 @@ unsigned X86FastISel::TargetMaterializeConstant(const Constant *C) { const TargetRegisterClass *RC = nullptr; switch (VT.SimpleTy) { default: return 0; - case MVT::i8: - Opc = X86::MOV8rm; - RC = &X86::GR8RegClass; - break; - case MVT::i16: - Opc = X86::MOV16rm; - RC = &X86::GR16RegClass; - break; - case MVT::i32: - Opc = X86::MOV32rm; - RC = &X86::GR32RegClass; - break; - case MVT::i64: - // Must be in x86-64 mode. - Opc = X86::MOV64rm; - RC = &X86::GR64RegClass; - break; case MVT::f32: if (X86ScalarSSEf32) { Opc = Subtarget->hasAVX() ? X86::VMOVSSrm : X86::MOVSSrm; @@ -2480,30 +3194,11 @@ unsigned X86FastISel::TargetMaterializeConstant(const Constant *C) { return 0; } - // Materialize addresses with LEA instructions. - if (isa(C)) { - X86AddressMode AM; - if (X86SelectAddress(C, AM)) { - // If the expression is just a basereg, then we're done, otherwise we need - // to emit an LEA. - if (AM.BaseType == X86AddressMode::RegBase && - AM.IndexReg == 0 && AM.Disp == 0 && AM.GV == nullptr) - return AM.Base.Reg; - - Opc = TLI.getPointerTy() == MVT::i32 ? X86::LEA32r : X86::LEA64r; - unsigned ResultReg = createResultReg(RC); - addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, - TII.get(Opc), ResultReg), AM); - return ResultReg; - } - return 0; - } - // MachineConstantPool wants an explicit alignment. - unsigned Align = DL.getPrefTypeAlignment(C->getType()); + unsigned Align = DL.getPrefTypeAlignment(CFP->getType()); if (Align == 0) { - // Alignment of vector types. FIXME! - Align = DL.getTypeAllocSize(C->getType()); + // Alignment of vector types. FIXME! + Align = DL.getTypeAllocSize(CFP->getType()); } // x86-32 PIC requires a PIC base register for constant pools. @@ -2521,15 +3216,65 @@ unsigned X86FastISel::TargetMaterializeConstant(const Constant *C) { } // Create the load from the constant pool. - unsigned MCPOffset = MCP.getConstantPoolIndex(C, Align); + unsigned CPI = MCP.getConstantPoolIndex(CFP, Align); unsigned ResultReg = createResultReg(RC); + addConstantPoolReference(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg), - MCPOffset, PICBase, OpFlag); - + CPI, PICBase, OpFlag); return ResultReg; } +unsigned X86FastISel::X86MaterializeGV(const GlobalValue *GV, MVT VT) { + // Can't handle alternate code models yet. + if (TM.getCodeModel() != CodeModel::Small) + return 0; + + // Materialize addresses with LEA/MOV instructions. + X86AddressMode AM; + if (X86SelectAddress(GV, AM)) { + // If the expression is just a basereg, then we're done, otherwise we need + // to emit an LEA. + if (AM.BaseType == X86AddressMode::RegBase && + AM.IndexReg == 0 && AM.Disp == 0 && AM.GV == nullptr) + return AM.Base.Reg; + + unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT)); + if (TM.getRelocationModel() == Reloc::Static && + TLI.getPointerTy() == MVT::i64) { + // The displacement code could be more than 32 bits away so we need to use + // an instruction with a 64 bit immediate + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::MOV64ri), + ResultReg) + .addGlobalAddress(GV); + } else { + unsigned Opc = TLI.getPointerTy() == MVT::i32 ? X86::LEA32r : X86::LEA64r; + addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, + TII.get(Opc), ResultReg), AM); + } + return ResultReg; + } + return 0; +} + +unsigned X86FastISel::TargetMaterializeConstant(const Constant *C) { + EVT CEVT = TLI.getValueType(C->getType(), true); + + // Only handle simple types. + if (!CEVT.isSimple()) + return 0; + MVT VT = CEVT.getSimpleVT(); + + if (const auto *CI = dyn_cast(C)) + return X86MaterializeInt(CI, VT); + else if (const ConstantFP *CFP = dyn_cast(C)) + return X86MaterializeFP(CFP, VT); + else if (const GlobalValue *GV = dyn_cast(C)) + return X86MaterializeGV(GV, VT); + + return 0; +} + unsigned X86FastISel::TargetMaterializeAlloca(const AllocaInst *C) { // Fail on dynamic allocas. At this point, getRegForValue has already // checked its CSE maps, so if we're here trying to handle a dynamic @@ -2594,8 +3339,9 @@ unsigned X86FastISel::TargetMaterializeFloatZero(const ConstantFP *CF) { bool X86FastISel::tryToFoldLoadIntoMI(MachineInstr *MI, unsigned OpNo, const LoadInst *LI) { + const Value *Ptr = LI->getPointerOperand(); X86AddressMode AM; - if (!X86SelectAddress(LI->getOperand(0), AM)) + if (!X86SelectAddress(Ptr, AM)) return false; const X86InstrInfo &XII = (const X86InstrInfo&)TII; @@ -2603,13 +3349,18 @@ bool X86FastISel::tryToFoldLoadIntoMI(MachineInstr *MI, unsigned OpNo, unsigned Size = DL.getTypeAllocSize(LI->getType()); unsigned Alignment = LI->getAlignment(); + if (Alignment == 0) // Ensure that codegen never sees alignment 0 + Alignment = DL.getABITypeAlignment(LI->getType()); + SmallVector AddrOps; AM.getFullAddress(AddrOps); MachineInstr *Result = XII.foldMemoryOperandImpl(*FuncInfo.MF, MI, OpNo, AddrOps, Size, Alignment); - if (!Result) return false; + if (!Result) + return false; + Result->addMemOperand(*FuncInfo.MF, createMachineMemOperandFor(LI)); FuncInfo.MBB->insert(FuncInfo.InsertPt, Result); MI->eraseFromParent(); return true;