X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FTarget%2FX86%2FX86CodeEmitter.cpp;h=8fea6edc8d780457ebbc3908174d675baa249e8e;hb=1f4b796b49d13075531ed43b35824ecc9d757467;hp=97572051470aa35d81f76f26a4f96c7a87766308;hpb=7431beaba2a01c3fe299c861b2ec85cbf1dc81c4;p=oota-llvm.git diff --git a/lib/Target/X86/X86CodeEmitter.cpp b/lib/Target/X86/X86CodeEmitter.cpp index 97572051470..8fea6edc8d7 100644 --- a/lib/Target/X86/X86CodeEmitter.cpp +++ b/lib/Target/X86/X86CodeEmitter.cpp @@ -1,4 +1,4 @@ -//===-- X86/X86CodeEmitter.cpp - Convert X86 code to machine code ---------===// +//===-- X86CodeEmitter.cpp - Convert X86 code to machine code -------------===// // // The LLVM Compiler Infrastructure // @@ -13,24 +13,23 @@ //===----------------------------------------------------------------------===// #define DEBUG_TYPE "x86-emitter" +#include "X86.h" #include "X86InstrInfo.h" #include "X86JITInfo.h" +#include "X86Relocations.h" #include "X86Subtarget.h" #include "X86TargetMachine.h" -#include "X86Relocations.h" -#include "X86.h" -#include "llvm/LLVMContext.h" -#include "llvm/PassManager.h" +#include "llvm/ADT/Statistic.h" #include "llvm/CodeGen/JITCodeEmitter.h" #include "llvm/CodeGen/MachineFunctionPass.h" #include "llvm/CodeGen/MachineInstr.h" #include "llvm/CodeGen/MachineModuleInfo.h" #include "llvm/CodeGen/Passes.h" -#include "llvm/Function.h" -#include "llvm/ADT/Statistic.h" +#include "llvm/IR/LLVMContext.h" #include "llvm/MC/MCCodeEmitter.h" #include "llvm/MC/MCExpr.h" #include "llvm/MC/MCInst.h" +#include "llvm/PassManager.h" #include "llvm/Support/Debug.h" #include "llvm/Support/ErrorHandling.h" #include "llvm/Support/raw_ostream.h" @@ -43,7 +42,7 @@ namespace { template class Emitter : public MachineFunctionPass { const X86InstrInfo *II; - const TargetData *TD; + const DataLayout *TD; X86TargetMachine &TM; CodeEmitter &MCE; MachineModuleInfo *MMI; @@ -53,12 +52,12 @@ namespace { public: static char ID; explicit Emitter(X86TargetMachine &tm, CodeEmitter &mce) - : MachineFunctionPass(&ID), II(0), TD(0), TM(tm), + : MachineFunctionPass(ID), II(0), TD(0), TM(tm), MCE(mce), PICBaseOffset(0), Is64BitMode(false), IsPIC(TM.getRelocationModel() == Reloc::PIC_) {} Emitter(X86TargetMachine &tm, CodeEmitter &mce, - const X86InstrInfo &ii, const TargetData &td, bool is64) - : MachineFunctionPass(&ID), II(&ii), TD(&td), TM(tm), + const X86InstrInfo &ii, const DataLayout &td, bool is64) + : MachineFunctionPass(ID), II(&ii), TD(&td), TM(tm), MCE(mce), PICBaseOffset(0), Is64BitMode(is64), IsPIC(TM.getRelocationModel() == Reloc::PIC_) {} @@ -68,9 +67,20 @@ namespace { return "X86 Machine Code Emitter"; } - void emitInstruction(const MachineInstr &MI, - const TargetInstrDesc *Desc); - + void emitOpcodePrefix(uint64_t TSFlags, int MemOperand, + const MachineInstr &MI, + const MCInstrDesc *Desc) const; + + void emitVEXOpcodePrefix(uint64_t TSFlags, int MemOperand, + const MachineInstr &MI, + const MCInstrDesc *Desc) const; + + void emitSegmentOverridePrefix(uint64_t TSFlags, + int MemOperand, + const MachineInstr &MI) const; + + void emitInstruction(MachineInstr &MI, const MCInstrDesc *Desc); + void getAnalysisUsage(AnalysisUsage &AU) const { AU.setPreservesAll(); AU.addRequired(); @@ -100,7 +110,13 @@ namespace { unsigned Op, unsigned RegOpcodeField, intptr_t PCAdj = 0); - unsigned getX86RegNum(unsigned RegNo) const; + unsigned getX86RegNum(unsigned RegNo) const { + const TargetRegisterInfo *TRI = TM.getRegisterInfo(); + return TRI->getEncodingValue(RegNo) & 0x7; + } + + unsigned char getVEXRegisterEncoding(const MachineInstr &MI, + unsigned OpNum) const; }; template @@ -108,7 +124,7 @@ template } // end anonymous namespace. /// createX86CodeEmitterPass - Return a pass that emits the collected X86 code -/// to the specified templated MachineCodeEmitter object. +/// to the specified JITCodeEmitter object. FunctionPass *llvm::createX86JITCodeEmitterPass(X86TargetMachine &TM, JITCodeEmitter &JCE) { return new Emitter(TM, JCE); @@ -118,22 +134,21 @@ template bool Emitter::runOnMachineFunction(MachineFunction &MF) { MMI = &getAnalysis(); MCE.setModuleInfo(MMI); - + II = TM.getInstrInfo(); - TD = TM.getTargetData(); + TD = TM.getDataLayout(); Is64BitMode = TM.getSubtarget().is64Bit(); IsPIC = TM.getRelocationModel() == Reloc::PIC_; - + do { - DEBUG(dbgs() << "JITTing function '" - << MF.getFunction()->getName() << "'\n"); + DEBUG(dbgs() << "JITTing function '" << MF.getName() << "'\n"); MCE.startFunction(MF); - for (MachineFunction::iterator MBB = MF.begin(), E = MF.end(); + for (MachineFunction::iterator MBB = MF.begin(), E = MF.end(); MBB != E; ++MBB) { MCE.StartMachineBasicBlock(MBB); - for (MachineBasicBlock::const_iterator I = MBB->begin(), E = MBB->end(); + for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); I != E; ++I) { - const TargetInstrDesc &Desc = I->getDesc(); + const MCInstrDesc &Desc = I->getDesc(); emitInstruction(*I, &Desc); // MOVPC32r is basically a call plus a pop instruction. if (Desc.getOpcode() == X86::MOVPC32r) @@ -146,6 +161,103 @@ bool Emitter::runOnMachineFunction(MachineFunction &MF) { return false; } +/// determineREX - Determine if the MachineInstr has to be encoded with a X86-64 +/// REX prefix which specifies 1) 64-bit instructions, 2) non-default operand +/// size, and 3) use of X86-64 extended registers. +static unsigned determineREX(const MachineInstr &MI) { + unsigned REX = 0; + const MCInstrDesc &Desc = MI.getDesc(); + + // Pseudo instructions do not need REX prefix byte. + if ((Desc.TSFlags & X86II::FormMask) == X86II::Pseudo) + return 0; + if (Desc.TSFlags & X86II::REX_W) + REX |= 1 << 3; + + unsigned NumOps = Desc.getNumOperands(); + if (NumOps) { + bool isTwoAddr = NumOps > 1 && + Desc.getOperandConstraint(1, MCOI::TIED_TO) != -1; + + // If it accesses SPL, BPL, SIL, or DIL, then it requires a 0x40 REX prefix. + unsigned i = isTwoAddr ? 1 : 0; + for (unsigned e = NumOps; i != e; ++i) { + const MachineOperand& MO = MI.getOperand(i); + if (MO.isReg()) { + unsigned Reg = MO.getReg(); + if (X86II::isX86_64NonExtLowByteReg(Reg)) + REX |= 0x40; + } + } + + switch (Desc.TSFlags & X86II::FormMask) { + case X86II::MRMInitReg: + if (X86InstrInfo::isX86_64ExtendedReg(MI.getOperand(0))) + REX |= (1 << 0) | (1 << 2); + break; + case X86II::MRMSrcReg: { + if (X86InstrInfo::isX86_64ExtendedReg(MI.getOperand(0))) + REX |= 1 << 2; + i = isTwoAddr ? 2 : 1; + for (unsigned e = NumOps; i != e; ++i) { + const MachineOperand& MO = MI.getOperand(i); + if (X86InstrInfo::isX86_64ExtendedReg(MO)) + REX |= 1 << 0; + } + break; + } + case X86II::MRMSrcMem: { + if (X86InstrInfo::isX86_64ExtendedReg(MI.getOperand(0))) + REX |= 1 << 2; + unsigned Bit = 0; + i = isTwoAddr ? 2 : 1; + for (; i != NumOps; ++i) { + const MachineOperand& MO = MI.getOperand(i); + if (MO.isReg()) { + if (X86InstrInfo::isX86_64ExtendedReg(MO)) + REX |= 1 << Bit; + Bit++; + } + } + break; + } + case X86II::MRM0m: case X86II::MRM1m: + case X86II::MRM2m: case X86II::MRM3m: + case X86II::MRM4m: case X86II::MRM5m: + case X86II::MRM6m: case X86II::MRM7m: + case X86II::MRMDestMem: { + unsigned e = (isTwoAddr ? X86::AddrNumOperands+1 : X86::AddrNumOperands); + i = isTwoAddr ? 1 : 0; + if (NumOps > e && X86InstrInfo::isX86_64ExtendedReg(MI.getOperand(e))) + REX |= 1 << 2; + unsigned Bit = 0; + for (; i != e; ++i) { + const MachineOperand& MO = MI.getOperand(i); + if (MO.isReg()) { + if (X86InstrInfo::isX86_64ExtendedReg(MO)) + REX |= 1 << Bit; + Bit++; + } + } + break; + } + default: { + if (X86InstrInfo::isX86_64ExtendedReg(MI.getOperand(0))) + REX |= 1 << 0; + i = isTwoAddr ? 2 : 1; + for (unsigned e = NumOps; i != e; ++i) { + const MachineOperand& MO = MI.getOperand(i); + if (X86InstrInfo::isX86_64ExtendedReg(MO)) + REX |= 1 << 2; + } + break; + } + } + } + return REX; +} + + /// emitPCRelativeBlockAddress - This method keeps track of the information /// necessary to resolve the address of this block later and emits a dummy /// value. @@ -250,11 +362,6 @@ void Emitter::emitJumpTableAddress(unsigned JTI, unsigned Reloc, MCE.emitWordLE(0); } -template -unsigned Emitter::getX86RegNum(unsigned RegNo) const { - return X86RegisterInfo::getX86RegNum(RegNo); -} - inline static unsigned char ModRMByte(unsigned Mod, unsigned RegOpcode, unsigned RM) { assert(Mod < 4 && RegOpcode < 8 && RM < 8 && "ModRM Fields out of range!"); @@ -273,7 +380,7 @@ void Emitter::emitRegModRMByte(unsigned RegOpcodeFld) { } template -void Emitter::emitSIBByte(unsigned SS, +void Emitter::emitSIBByte(unsigned SS, unsigned Index, unsigned Base) { // SIB byte is in the same format as the ModRMByte... @@ -289,8 +396,8 @@ void Emitter::emitConstant(uint64_t Val, unsigned Size) { } } -/// isDisp8 - Return true if this signed displacement fits in a 8-bit -/// sign-extended field. +/// isDisp8 - Return true if this signed displacement fits in a 8-bit +/// sign-extended field. static bool isDisp8(int Value) { return Value == (signed char)Value; } @@ -299,10 +406,10 @@ static bool gvNeedsNonLazyPtr(const MachineOperand &GVOp, const TargetMachine &TM) { // For Darwin-64, simulate the linktime GOT by using the same non-lazy-pointer // mechanism as 32-bit mode. - if (TM.getSubtarget().is64Bit() && + if (TM.getSubtarget().is64Bit() && !TM.getSubtarget().isTargetDarwin()) return false; - + // Return true if this is a reference to a stub containing the address of the // global, not the global itself. return isGlobalStubReference(GVOp.getTargetFlags()); @@ -328,7 +435,7 @@ void Emitter::emitDisplacementField(const MachineOperand *RelocOp, if (RelocOp->isGlobal()) { // In 64-bit static small code model, we could potentially emit absolute. // But it's probably not beneficial. If the MCE supports using RIP directly - // do it, otherwise fallback to absolute (this is determined by IsPCRel). + // do it, otherwise fallback to absolute (this is determined by IsPCRel). // 89 05 00 00 00 00 mov %eax,0(%rip) # PC-relative // 89 04 25 00 00 00 00 mov %eax,0x0 # Absolute bool Indirect = gvNeedsNonLazyPtr(*RelocOp, TM); @@ -352,7 +459,7 @@ void Emitter::emitMemModRMByte(const MachineInstr &MI, const MachineOperand &Op3 = MI.getOperand(Op+3); int DispVal = 0; const MachineOperand *DispForReloc = 0; - + // Figure out what sort of displacement we have to handle here. if (Op3.isGlobal()) { DispForReloc = &Op3; @@ -380,7 +487,7 @@ void Emitter::emitMemModRMByte(const MachineInstr &MI, const MachineOperand &IndexReg = MI.getOperand(Op+2); unsigned BaseReg = Base.getReg(); - + // Handle %rip relative addressing. if (BaseReg == X86::RIP || (Is64BitMode && DispForReloc)) { // [disp32+RIP] in X86-64 mode @@ -397,7 +504,7 @@ void Emitter::emitMemModRMByte(const MachineInstr &MI, bool IsPCRel = MCE.earlyResolveAddresses() ? true : false; // Is a SIB byte needed? - // If no BaseReg, issue a RIP relative instruction only if the MCE can + // If no BaseReg, issue a RIP relative instruction only if the MCE can // resolve addresses on-the-fly, otherwise use SIB (Intel Manual 2A, table // 2-7) and absolute references. unsigned BaseRegNo = -1U; @@ -405,7 +512,7 @@ void Emitter::emitMemModRMByte(const MachineInstr &MI, BaseRegNo = getX86RegNum(BaseReg); if (// The SIB byte must be used if there is an index register. - IndexReg.getReg() == 0 && + IndexReg.getReg() == 0 && // The SIB byte must be used if the base is ESP/RSP/R12, all of which // encode to an R/M value of 4, which indicates that a SIB byte is // present. @@ -419,7 +526,7 @@ void Emitter::emitMemModRMByte(const MachineInstr &MI, emitDisplacementField(DispForReloc, DispVal, PCAdj, true); return; } - + // If the base is not EBP/ESP and there is no displacement, use simple // indirect register encoding, this handles addresses like [EAX]. The // encoding for [EBP] with no displacement means [disp32] so we handle it @@ -428,20 +535,20 @@ void Emitter::emitMemModRMByte(const MachineInstr &MI, MCE.emitByte(ModRMByte(0, RegOpcodeField, BaseRegNo)); return; } - + // Otherwise, if the displacement fits in a byte, encode as [REG+disp8]. if (!DispForReloc && isDisp8(DispVal)) { MCE.emitByte(ModRMByte(1, RegOpcodeField, BaseRegNo)); emitConstant(DispVal, 1); return; } - + // Otherwise, emit the most general non-SIB encoding: [REG+disp32] MCE.emitByte(ModRMByte(2, RegOpcodeField, BaseRegNo)); emitDisplacementField(DispForReloc, DispVal, PCAdj, IsPCRel); return; } - + // Otherwise we need a SIB byte, so start by outputting the ModR/M byte first. assert(IndexReg.getReg() != X86::ESP && IndexReg.getReg() != X86::RSP && "Cannot use ESP as index reg!"); @@ -470,11 +577,11 @@ void Emitter::emitMemModRMByte(const MachineInstr &MI, } // Calculate what the SS field value should be... - static const unsigned SSTable[] = { ~0, 0, 1, ~0, 2, ~0, ~0, ~0, 3 }; + static const unsigned SSTable[] = { ~0U, 0, 1, ~0U, 2, ~0U, ~0U, ~0U, 3 }; unsigned SS = SSTable[Scale.getImm()]; if (BaseReg == 0) { - // Handle the SIB byte for the case where there is no base, see Intel + // Handle the SIB byte for the case where there is no base, see Intel // Manual 2A, table 2-7. The displacement has already been output. unsigned IndexRegNo; if (IndexReg.getReg()) @@ -500,76 +607,128 @@ void Emitter::emitMemModRMByte(const MachineInstr &MI, } } -template -void Emitter::emitInstruction(const MachineInstr &MI, - const TargetInstrDesc *Desc) { - DEBUG(dbgs() << MI); +static const MCInstrDesc *UpdateOp(MachineInstr &MI, const X86InstrInfo *II, + unsigned Opcode) { + const MCInstrDesc *Desc = &II->get(Opcode); + MI.setDesc(*Desc); + return Desc; +} - MCE.processDebugLoc(MI.getDebugLoc(), true); +/// Is16BitMemOperand - Return true if the specified instruction has +/// a 16-bit memory operand. Op specifies the operand # of the memoperand. +static bool Is16BitMemOperand(const MachineInstr &MI, unsigned Op) { + const MachineOperand &BaseReg = MI.getOperand(Op+X86::AddrBaseReg); + const MachineOperand &IndexReg = MI.getOperand(Op+X86::AddrIndexReg); + + if ((BaseReg.getReg() != 0 && + X86MCRegisterClasses[X86::GR16RegClassID].contains(BaseReg.getReg())) || + (IndexReg.getReg() != 0 && + X86MCRegisterClasses[X86::GR16RegClassID].contains(IndexReg.getReg()))) + return true; + return false; +} - unsigned Opcode = Desc->Opcode; +/// Is32BitMemOperand - Return true if the specified instruction has +/// a 32-bit memory operand. Op specifies the operand # of the memoperand. +static bool Is32BitMemOperand(const MachineInstr &MI, unsigned Op) { + const MachineOperand &BaseReg = MI.getOperand(Op+X86::AddrBaseReg); + const MachineOperand &IndexReg = MI.getOperand(Op+X86::AddrIndexReg); + + if ((BaseReg.getReg() != 0 && + X86MCRegisterClasses[X86::GR32RegClassID].contains(BaseReg.getReg())) || + (IndexReg.getReg() != 0 && + X86MCRegisterClasses[X86::GR32RegClassID].contains(IndexReg.getReg()))) + return true; + return false; +} +/// Is64BitMemOperand - Return true if the specified instruction has +/// a 64-bit memory operand. Op specifies the operand # of the memoperand. +#ifndef NDEBUG +static bool Is64BitMemOperand(const MachineInstr &MI, unsigned Op) { + const MachineOperand &BaseReg = MI.getOperand(Op+X86::AddrBaseReg); + const MachineOperand &IndexReg = MI.getOperand(Op+X86::AddrIndexReg); + + if ((BaseReg.getReg() != 0 && + X86MCRegisterClasses[X86::GR64RegClassID].contains(BaseReg.getReg())) || + (IndexReg.getReg() != 0 && + X86MCRegisterClasses[X86::GR64RegClassID].contains(IndexReg.getReg()))) + return true; + return false; +} +#endif + +template +void Emitter::emitOpcodePrefix(uint64_t TSFlags, + int MemOperand, + const MachineInstr &MI, + const MCInstrDesc *Desc) const { // Emit the lock opcode prefix as needed. if (Desc->TSFlags & X86II::LOCK) MCE.emitByte(0xF0); // Emit segment override opcode prefix as needed. - switch (Desc->TSFlags & X86II::SegOvrMask) { - case X86II::FS: - MCE.emitByte(0x64); - break; - case X86II::GS: - MCE.emitByte(0x65); - break; - default: llvm_unreachable("Invalid segment!"); - case 0: break; // No segment override! - } + emitSegmentOverridePrefix(TSFlags, MemOperand, MI); // Emit the repeat opcode prefix as needed. if ((Desc->TSFlags & X86II::Op0Mask) == X86II::REP) MCE.emitByte(0xF3); - // Emit the operand size opcode prefix as needed. - if (Desc->TSFlags & X86II::OpSize) - MCE.emitByte(0x66); - // Emit the address size opcode prefix as needed. - if (Desc->TSFlags & X86II::AdSize) + bool need_address_override; + if (TSFlags & X86II::AdSize) { + need_address_override = true; + } else if (MemOperand == -1) { + need_address_override = false; + } else if (Is64BitMode) { + assert(!Is16BitMemOperand(MI, MemOperand)); + need_address_override = Is32BitMemOperand(MI, MemOperand); + } else { + assert(!Is64BitMemOperand(MI, MemOperand)); + need_address_override = Is16BitMemOperand(MI, MemOperand); + } + + if (need_address_override) MCE.emitByte(0x67); + // Emit the operand size opcode prefix as needed. + if (TSFlags & X86II::OpSize) + MCE.emitByte(0x66); + bool Need0FPrefix = false; switch (Desc->TSFlags & X86II::Op0Mask) { - case X86II::TB: // Two-byte opcode prefix - case X86II::T8: // 0F 38 - case X86II::TA: // 0F 3A - Need0FPrefix = true; - break; - case X86II::TF: // F2 0F 38 - MCE.emitByte(0xF2); - Need0FPrefix = true; - break; - case X86II::REP: break; // already handled. - case X86II::XS: // F3 0F - MCE.emitByte(0xF3); - Need0FPrefix = true; - break; - case X86II::XD: // F2 0F - MCE.emitByte(0xF2); - Need0FPrefix = true; - break; - case X86II::D8: case X86II::D9: case X86II::DA: case X86II::DB: - case X86II::DC: case X86II::DD: case X86II::DE: case X86II::DF: - MCE.emitByte(0xD8+ - (((Desc->TSFlags & X86II::Op0Mask)-X86II::D8) - >> X86II::Op0Shift)); - break; // Two-byte opcode prefix - default: llvm_unreachable("Invalid prefix!"); - case 0: break; // No prefix! + case X86II::TB: // Two-byte opcode prefix + case X86II::T8: // 0F 38 + case X86II::TA: // 0F 3A + case X86II::A6: // 0F A6 + case X86II::A7: // 0F A7 + Need0FPrefix = true; + break; + case X86II::REP: break; // already handled. + case X86II::T8XS: // F3 0F 38 + case X86II::XS: // F3 0F + MCE.emitByte(0xF3); + Need0FPrefix = true; + break; + case X86II::T8XD: // F2 0F 38 + case X86II::TAXD: // F2 0F 3A + case X86II::XD: // F2 0F + MCE.emitByte(0xF2); + Need0FPrefix = true; + break; + case X86II::D8: case X86II::D9: case X86II::DA: case X86II::DB: + case X86II::DC: case X86II::DD: case X86II::DE: case X86II::DF: + MCE.emitByte(0xD8+ + (((Desc->TSFlags & X86II::Op0Mask)-X86II::D8) + >> X86II::Op0Shift)); + break; // Two-byte opcode prefix + default: llvm_unreachable("Invalid prefix!"); + case 0: break; // No prefix! } // Handle REX prefix. if (Is64BitMode) { - if (unsigned REX = X86InstrInfo::determineREX(MI)) + if (unsigned REX = determineREX(MI)) MCE.emitByte(0x40 | REX); } @@ -578,36 +737,448 @@ void Emitter::emitInstruction(const MachineInstr &MI, MCE.emitByte(0x0F); switch (Desc->TSFlags & X86II::Op0Mask) { - case X86II::TF: // F2 0F 38 - case X86II::T8: // 0F 38 - MCE.emitByte(0x38); - break; - case X86II::TA: // 0F 3A - MCE.emitByte(0x3A); - break; + case X86II::T8XD: // F2 0F 38 + case X86II::T8XS: // F3 0F 38 + case X86II::T8: // 0F 38 + MCE.emitByte(0x38); + break; + case X86II::TAXD: // F2 0F 38 + case X86II::TA: // 0F 3A + MCE.emitByte(0x3A); + break; + case X86II::A6: // 0F A6 + MCE.emitByte(0xA6); + break; + case X86II::A7: // 0F A7 + MCE.emitByte(0xA7); + break; + } +} + +// On regular x86, both XMM0-XMM7 and XMM8-XMM15 are encoded in the range +// 0-7 and the difference between the 2 groups is given by the REX prefix. +// In the VEX prefix, registers are seen sequencially from 0-15 and encoded +// in 1's complement form, example: +// +// ModRM field => XMM9 => 1 +// VEX.VVVV => XMM9 => ~9 +// +// See table 4-35 of Intel AVX Programming Reference for details. +template +unsigned char +Emitter::getVEXRegisterEncoding(const MachineInstr &MI, + unsigned OpNum) const { + unsigned SrcReg = MI.getOperand(OpNum).getReg(); + unsigned SrcRegNum = getX86RegNum(MI.getOperand(OpNum).getReg()); + if (X86II::isX86_64ExtendedReg(SrcReg)) + SrcRegNum |= 8; + + // The registers represented through VEX_VVVV should + // be encoded in 1's complement form. + return (~SrcRegNum) & 0xf; +} + +/// EmitSegmentOverridePrefix - Emit segment override opcode prefix as needed +template +void Emitter::emitSegmentOverridePrefix(uint64_t TSFlags, + int MemOperand, + const MachineInstr &MI) const { + switch (TSFlags & X86II::SegOvrMask) { + default: llvm_unreachable("Invalid segment!"); + case 0: + // No segment override, check for explicit one on memory operand. + if (MemOperand != -1) { // If the instruction has a memory operand. + switch (MI.getOperand(MemOperand+X86::AddrSegmentReg).getReg()) { + default: llvm_unreachable("Unknown segment register!"); + case 0: break; + case X86::CS: MCE.emitByte(0x2E); break; + case X86::SS: MCE.emitByte(0x36); break; + case X86::DS: MCE.emitByte(0x3E); break; + case X86::ES: MCE.emitByte(0x26); break; + case X86::FS: MCE.emitByte(0x64); break; + case X86::GS: MCE.emitByte(0x65); break; + } + } + break; + case X86II::FS: + MCE.emitByte(0x64); + break; + case X86II::GS: + MCE.emitByte(0x65); + break; } +} + +template +void Emitter::emitVEXOpcodePrefix(uint64_t TSFlags, + int MemOperand, + const MachineInstr &MI, + const MCInstrDesc *Desc) const { + bool HasVEX_4V = (TSFlags >> X86II::VEXShift) & X86II::VEX_4V; + bool HasVEX_4VOp3 = (TSFlags >> X86II::VEXShift) & X86II::VEX_4VOp3; + bool HasMemOp4 = (TSFlags >> X86II::VEXShift) & X86II::MemOp4; + + // VEX_R: opcode externsion equivalent to REX.R in + // 1's complement (inverted) form + // + // 1: Same as REX_R=0 (must be 1 in 32-bit mode) + // 0: Same as REX_R=1 (64 bit mode only) + // + unsigned char VEX_R = 0x1; + + // VEX_X: equivalent to REX.X, only used when a + // register is used for index in SIB Byte. + // + // 1: Same as REX.X=0 (must be 1 in 32-bit mode) + // 0: Same as REX.X=1 (64-bit mode only) + unsigned char VEX_X = 0x1; + + // VEX_B: + // + // 1: Same as REX_B=0 (ignored in 32-bit mode) + // 0: Same as REX_B=1 (64 bit mode only) + // + unsigned char VEX_B = 0x1; + + // VEX_W: opcode specific (use like REX.W, or used for + // opcode extension, or ignored, depending on the opcode byte) + unsigned char VEX_W = 0; + + // XOP: Use XOP prefix byte 0x8f instead of VEX. + unsigned char XOP = 0; + + // VEX_5M (VEX m-mmmmm field): + // + // 0b00000: Reserved for future use + // 0b00001: implied 0F leading opcode + // 0b00010: implied 0F 38 leading opcode bytes + // 0b00011: implied 0F 3A leading opcode bytes + // 0b00100-0b11111: Reserved for future use + // 0b01000: XOP map select - 08h instructions with imm byte + // 0b10001: XOP map select - 09h instructions with no imm byte + unsigned char VEX_5M = 0x1; + + // VEX_4V (VEX vvvv field): a register specifier + // (in 1's complement form) or 1111 if unused. + unsigned char VEX_4V = 0xf; + + // VEX_L (Vector Length): + // + // 0: scalar or 128-bit vector + // 1: 256-bit vector + // + unsigned char VEX_L = 0; + + // VEX_PP: opcode extension providing equivalent + // functionality of a SIMD prefix + // + // 0b00: None + // 0b01: 66 + // 0b10: F3 + // 0b11: F2 + // + unsigned char VEX_PP = 0; + + // Encode the operand size opcode prefix as needed. + if (TSFlags & X86II::OpSize) + VEX_PP = 0x01; + + if ((TSFlags >> X86II::VEXShift) & X86II::VEX_W) + VEX_W = 1; + + if ((TSFlags >> X86II::VEXShift) & X86II::XOP) + XOP = 1; + + if ((TSFlags >> X86II::VEXShift) & X86II::VEX_L) + VEX_L = 1; + + switch (TSFlags & X86II::Op0Mask) { + default: llvm_unreachable("Invalid prefix!"); + case X86II::T8: // 0F 38 + VEX_5M = 0x2; + break; + case X86II::TA: // 0F 3A + VEX_5M = 0x3; + break; + case X86II::T8XS: // F3 0F 38 + VEX_PP = 0x2; + VEX_5M = 0x2; + break; + case X86II::T8XD: // F2 0F 38 + VEX_PP = 0x3; + VEX_5M = 0x2; + break; + case X86II::TAXD: // F2 0F 3A + VEX_PP = 0x3; + VEX_5M = 0x3; + break; + case X86II::XS: // F3 0F + VEX_PP = 0x2; + break; + case X86II::XD: // F2 0F + VEX_PP = 0x3; + break; + case X86II::XOP8: + VEX_5M = 0x8; + break; + case X86II::XOP9: + VEX_5M = 0x9; + break; + case X86II::A6: // Bypass: Not used by VEX + case X86II::A7: // Bypass: Not used by VEX + case X86II::TB: // Bypass: Not used by VEX + case 0: + break; // No prefix! + } + + + // Classify VEX_B, VEX_4V, VEX_R, VEX_X + unsigned NumOps = Desc->getNumOperands(); + unsigned CurOp = 0; + if (NumOps > 1 && Desc->getOperandConstraint(1, MCOI::TIED_TO) == 0) + ++CurOp; + else if (NumOps > 3 && Desc->getOperandConstraint(2, MCOI::TIED_TO) == 0) { + assert(Desc->getOperandConstraint(NumOps - 1, MCOI::TIED_TO) == 1); + // Special case for GATHER with 2 TIED_TO operands + // Skip the first 2 operands: dst, mask_wb + CurOp += 2; + } + + switch (TSFlags & X86II::FormMask) { + case X86II::MRMInitReg: + // Duplicate register. + if (X86II::isX86_64ExtendedReg(MI.getOperand(CurOp).getReg())) + VEX_R = 0x0; + + if (HasVEX_4V) + VEX_4V = getVEXRegisterEncoding(MI, CurOp); + if (X86II::isX86_64ExtendedReg(MI.getOperand(CurOp).getReg())) + VEX_B = 0x0; + if (HasVEX_4VOp3) + VEX_4V = getVEXRegisterEncoding(MI, CurOp); + break; + case X86II::MRMDestMem: { + // MRMDestMem instructions forms: + // MemAddr, src1(ModR/M) + // MemAddr, src1(VEX_4V), src2(ModR/M) + // MemAddr, src1(ModR/M), imm8 + // + if (X86II::isX86_64ExtendedReg(MI.getOperand(X86::AddrBaseReg).getReg())) + VEX_B = 0x0; + if (X86II::isX86_64ExtendedReg(MI.getOperand(X86::AddrIndexReg).getReg())) + VEX_X = 0x0; + + CurOp = X86::AddrNumOperands; + if (HasVEX_4V) + VEX_4V = getVEXRegisterEncoding(MI, CurOp++); + + const MachineOperand &MO = MI.getOperand(CurOp); + if (MO.isReg() && X86II::isX86_64ExtendedReg(MO.getReg())) + VEX_R = 0x0; + break; + } + case X86II::MRMSrcMem: + // MRMSrcMem instructions forms: + // src1(ModR/M), MemAddr + // src1(ModR/M), src2(VEX_4V), MemAddr + // src1(ModR/M), MemAddr, imm8 + // src1(ModR/M), MemAddr, src2(VEX_I8IMM) + // + // FMA4: + // dst(ModR/M.reg), src1(VEX_4V), src2(ModR/M), src3(VEX_I8IMM) + // dst(ModR/M.reg), src1(VEX_4V), src2(VEX_I8IMM), src3(ModR/M), + if (X86II::isX86_64ExtendedReg(MI.getOperand(0).getReg())) + VEX_R = 0x0; + + if (HasVEX_4V) + VEX_4V = getVEXRegisterEncoding(MI, 1); + + if (X86II::isX86_64ExtendedReg( + MI.getOperand(MemOperand+X86::AddrBaseReg).getReg())) + VEX_B = 0x0; + if (X86II::isX86_64ExtendedReg( + MI.getOperand(MemOperand+X86::AddrIndexReg).getReg())) + VEX_X = 0x0; + + if (HasVEX_4VOp3) + VEX_4V = getVEXRegisterEncoding(MI, X86::AddrNumOperands+1); + break; + case X86II::MRM0m: case X86II::MRM1m: + case X86II::MRM2m: case X86II::MRM3m: + case X86II::MRM4m: case X86II::MRM5m: + case X86II::MRM6m: case X86II::MRM7m: { + // MRM[0-9]m instructions forms: + // MemAddr + // src1(VEX_4V), MemAddr + if (HasVEX_4V) + VEX_4V = getVEXRegisterEncoding(MI, 0); + + if (X86II::isX86_64ExtendedReg( + MI.getOperand(MemOperand+X86::AddrBaseReg).getReg())) + VEX_B = 0x0; + if (X86II::isX86_64ExtendedReg( + MI.getOperand(MemOperand+X86::AddrIndexReg).getReg())) + VEX_X = 0x0; + break; + } + case X86II::MRMSrcReg: + // MRMSrcReg instructions forms: + // dst(ModR/M), src1(VEX_4V), src2(ModR/M), src3(VEX_I8IMM) + // dst(ModR/M), src1(ModR/M) + // dst(ModR/M), src1(ModR/M), imm8 + // + if (X86II::isX86_64ExtendedReg(MI.getOperand(CurOp).getReg())) + VEX_R = 0x0; + CurOp++; + + if (HasVEX_4V) + VEX_4V = getVEXRegisterEncoding(MI, CurOp++); + + if (HasMemOp4) // Skip second register source (encoded in I8IMM) + CurOp++; + + if (X86II::isX86_64ExtendedReg(MI.getOperand(CurOp).getReg())) + VEX_B = 0x0; + CurOp++; + if (HasVEX_4VOp3) + VEX_4V = getVEXRegisterEncoding(MI, CurOp); + break; + case X86II::MRMDestReg: + // MRMDestReg instructions forms: + // dst(ModR/M), src(ModR/M) + // dst(ModR/M), src(ModR/M), imm8 + // dst(ModR/M), src1(VEX_4V), src2(ModR/M) + if (X86II::isX86_64ExtendedReg(MI.getOperand(CurOp).getReg())) + VEX_B = 0x0; + CurOp++; + + if (HasVEX_4V) + VEX_4V = getVEXRegisterEncoding(MI, CurOp++); + + if (X86II::isX86_64ExtendedReg(MI.getOperand(CurOp).getReg())) + VEX_R = 0x0; + break; + case X86II::MRM0r: case X86II::MRM1r: + case X86II::MRM2r: case X86II::MRM3r: + case X86II::MRM4r: case X86II::MRM5r: + case X86II::MRM6r: case X86II::MRM7r: + // MRM0r-MRM7r instructions forms: + // dst(VEX_4V), src(ModR/M), imm8 + VEX_4V = getVEXRegisterEncoding(MI, 0); + if (X86II::isX86_64ExtendedReg(MI.getOperand(1).getReg())) + VEX_B = 0x0; + break; + default: // RawFrm + break; + } + + // Emit segment override opcode prefix as needed. + emitSegmentOverridePrefix(TSFlags, MemOperand, MI); + + // VEX opcode prefix can have 2 or 3 bytes + // + // 3 bytes: + // +-----+ +--------------+ +-------------------+ + // | C4h | | RXB | m-mmmm | | W | vvvv | L | pp | + // +-----+ +--------------+ +-------------------+ + // 2 bytes: + // +-----+ +-------------------+ + // | C5h | | R | vvvv | L | pp | + // +-----+ +-------------------+ + // + unsigned char LastByte = VEX_PP | (VEX_L << 2) | (VEX_4V << 3); + + if (VEX_B && VEX_X && !VEX_W && !XOP && (VEX_5M == 1)) { // 2 byte VEX prefix + MCE.emitByte(0xC5); + MCE.emitByte(LastByte | (VEX_R << 7)); + return; + } + + // 3 byte VEX prefix + MCE.emitByte(XOP ? 0x8F : 0xC4); + MCE.emitByte(VEX_R << 7 | VEX_X << 6 | VEX_B << 5 | VEX_5M); + MCE.emitByte(LastByte | (VEX_W << 7)); +} + +template +void Emitter::emitInstruction(MachineInstr &MI, + const MCInstrDesc *Desc) { + DEBUG(dbgs() << MI); + + // If this is a pseudo instruction, lower it. + switch (Desc->getOpcode()) { + case X86::ADD16rr_DB: Desc = UpdateOp(MI, II, X86::OR16rr); break; + case X86::ADD32rr_DB: Desc = UpdateOp(MI, II, X86::OR32rr); break; + case X86::ADD64rr_DB: Desc = UpdateOp(MI, II, X86::OR64rr); break; + case X86::ADD16ri_DB: Desc = UpdateOp(MI, II, X86::OR16ri); break; + case X86::ADD32ri_DB: Desc = UpdateOp(MI, II, X86::OR32ri); break; + case X86::ADD64ri32_DB: Desc = UpdateOp(MI, II, X86::OR64ri32); break; + case X86::ADD16ri8_DB: Desc = UpdateOp(MI, II, X86::OR16ri8); break; + case X86::ADD32ri8_DB: Desc = UpdateOp(MI, II, X86::OR32ri8); break; + case X86::ADD64ri8_DB: Desc = UpdateOp(MI, II, X86::OR64ri8); break; + case X86::ACQUIRE_MOV8rm: Desc = UpdateOp(MI, II, X86::MOV8rm); break; + case X86::ACQUIRE_MOV16rm: Desc = UpdateOp(MI, II, X86::MOV16rm); break; + case X86::ACQUIRE_MOV32rm: Desc = UpdateOp(MI, II, X86::MOV32rm); break; + case X86::ACQUIRE_MOV64rm: Desc = UpdateOp(MI, II, X86::MOV64rm); break; + case X86::RELEASE_MOV8mr: Desc = UpdateOp(MI, II, X86::MOV8mr); break; + case X86::RELEASE_MOV16mr: Desc = UpdateOp(MI, II, X86::MOV16mr); break; + case X86::RELEASE_MOV32mr: Desc = UpdateOp(MI, II, X86::MOV32mr); break; + case X86::RELEASE_MOV64mr: Desc = UpdateOp(MI, II, X86::MOV64mr); break; + } + + + MCE.processDebugLoc(MI.getDebugLoc(), true); + + unsigned Opcode = Desc->Opcode; // If this is a two-address instruction, skip one of the register operands. unsigned NumOps = Desc->getNumOperands(); unsigned CurOp = 0; - if (NumOps > 1 && Desc->getOperandConstraint(1, TOI::TIED_TO) != -1) + if (NumOps > 1 && Desc->getOperandConstraint(1, MCOI::TIED_TO) == 0) ++CurOp; - else if (NumOps > 2 && Desc->getOperandConstraint(NumOps-1, TOI::TIED_TO)== 0) - // Skip the last source operand that is tied_to the dest reg. e.g. LXADD32 - --NumOps; + else if (NumOps > 3 && Desc->getOperandConstraint(2, MCOI::TIED_TO) == 0) { + assert(Desc->getOperandConstraint(NumOps - 1, MCOI::TIED_TO) == 1); + // Special case for GATHER with 2 TIED_TO operands + // Skip the first 2 operands: dst, mask_wb + CurOp += 2; + } + + uint64_t TSFlags = Desc->TSFlags; + + // Is this instruction encoded using the AVX VEX prefix? + bool HasVEXPrefix = (TSFlags >> X86II::VEXShift) & X86II::VEX; + // It uses the VEX.VVVV field? + bool HasVEX_4V = (TSFlags >> X86II::VEXShift) & X86II::VEX_4V; + bool HasVEX_4VOp3 = (TSFlags >> X86II::VEXShift) & X86II::VEX_4VOp3; + bool HasMemOp4 = (TSFlags >> X86II::VEXShift) & X86II::MemOp4; + const unsigned MemOp4_I8IMMOperand = 2; + + // Determine where the memory operand starts, if present. + int MemoryOperand = X86II::getMemoryOperandNo(TSFlags, Opcode); + if (MemoryOperand != -1) MemoryOperand += CurOp; + + if (!HasVEXPrefix) + emitOpcodePrefix(TSFlags, MemoryOperand, MI, Desc); + else + emitVEXOpcodePrefix(TSFlags, MemoryOperand, MI, Desc); unsigned char BaseOpcode = X86II::getBaseOpcodeFor(Desc->TSFlags); - switch (Desc->TSFlags & X86II::FormMask) { + switch (TSFlags & X86II::FormMask) { default: llvm_unreachable("Unknown FormMask value in X86 MachineCodeEmitter!"); case X86II::Pseudo: // Remember the current PC offset, this is the PIC relocation // base address. switch (Opcode) { - default: - llvm_unreachable("psuedo instructions should be removed before code" + default: + llvm_unreachable("pseudo instructions should be removed before code" " emission"); + // Do nothing for Int_MemBarrier - it's just a comment. Add a debug + // to make it slightly easier to see. + case X86::Int_MemBarrier: + DEBUG(dbgs() << "#MEMBARRIER\n"); break; + case TargetOpcode::INLINEASM: // We allow inline assembler nodes with empty bodies - they can // implicitly define registers, which is ok for JIT. @@ -619,7 +1190,7 @@ void Emitter::emitInstruction(const MachineInstr &MI, case TargetOpcode::EH_LABEL: MCE.emitLabel(MI.getOperand(0).getMCSymbol()); break; - + case TargetOpcode::IMPLICIT_DEF: case TargetOpcode::KILL: break; @@ -641,7 +1212,7 @@ void Emitter::emitInstruction(const MachineInstr &MI, if (CurOp == NumOps) break; - + const MachineOperand &MO = MI.getOperand(CurOp++); DEBUG(dbgs() << "RawFrm CurOp " << CurOp << "\n"); @@ -654,13 +1225,13 @@ void Emitter::emitInstruction(const MachineInstr &MI, emitPCRelativeBlockAddress(MO.getMBB()); break; } - + if (MO.isGlobal()) { emitGlobalAddress(MO.getGlobal(), X86::reloc_pcrel_word, MO.getOffset(), 0); break; } - + if (MO.isSymbol()) { emitExternalSymbolAddress(MO.getSymbolName(), X86::reloc_pcrel_word); break; @@ -671,7 +1242,7 @@ void Emitter::emitInstruction(const MachineInstr &MI, emitJumpTableAddress(MO.getIndex(), X86::reloc_pcrel_word); break; } - + assert(MO.isImm() && "Unknown RawFrm operand!"); if (Opcode == X86::CALLpcrel32 || Opcode == X86::CALL64pcrel32) { // Fix up immediate operand for pc relative calls. @@ -682,20 +1253,21 @@ void Emitter::emitInstruction(const MachineInstr &MI, emitConstant(MO.getImm(), X86II::getSizeOfImm(Desc->TSFlags)); break; } - + case X86II::AddRegFrm: { - MCE.emitByte(BaseOpcode + getX86RegNum(MI.getOperand(CurOp++).getReg())); - + MCE.emitByte(BaseOpcode + + getX86RegNum(MI.getOperand(CurOp++).getReg())); + if (CurOp == NumOps) break; - + const MachineOperand &MO1 = MI.getOperand(CurOp++); unsigned Size = X86II::getSizeOfImm(Desc->TSFlags); if (MO1.isImm()) { emitConstant(MO1.getImm(), Size); break; } - + unsigned rt = Is64BitMode ? X86::reloc_pcrel_word : (IsPIC ? X86::reloc_picrel_word : X86::reloc_absolute_word); if (Opcode == X86::MOV64ri64i32) @@ -718,49 +1290,65 @@ void Emitter::emitInstruction(const MachineInstr &MI, case X86II::MRMDestReg: { MCE.emitByte(BaseOpcode); + + unsigned SrcRegNum = CurOp+1; + if (HasVEX_4V) // Skip 1st src (which is encoded in VEX_VVVV) + SrcRegNum++; + emitRegModRMByte(MI.getOperand(CurOp).getReg(), - getX86RegNum(MI.getOperand(CurOp+1).getReg())); - CurOp += 2; - if (CurOp != NumOps) - emitConstant(MI.getOperand(CurOp++).getImm(), - X86II::getSizeOfImm(Desc->TSFlags)); + getX86RegNum(MI.getOperand(SrcRegNum).getReg())); + CurOp = SrcRegNum + 1; break; } case X86II::MRMDestMem: { MCE.emitByte(BaseOpcode); + + unsigned SrcRegNum = CurOp + X86::AddrNumOperands; + if (HasVEX_4V) // Skip 1st src (which is encoded in VEX_VVVV) + SrcRegNum++; emitMemModRMByte(MI, CurOp, - getX86RegNum(MI.getOperand(CurOp + X86::AddrNumOperands) - .getReg())); - CurOp += X86::AddrNumOperands + 1; - if (CurOp != NumOps) - emitConstant(MI.getOperand(CurOp++).getImm(), - X86II::getSizeOfImm(Desc->TSFlags)); + getX86RegNum(MI.getOperand(SrcRegNum).getReg())); + CurOp = SrcRegNum + 1; break; } - case X86II::MRMSrcReg: + case X86II::MRMSrcReg: { MCE.emitByte(BaseOpcode); - emitRegModRMByte(MI.getOperand(CurOp+1).getReg(), + + unsigned SrcRegNum = CurOp+1; + if (HasVEX_4V) // Skip 1st src (which is encoded in VEX_VVVV) + ++SrcRegNum; + + if (HasMemOp4) // Skip 2nd src (which is encoded in I8IMM) + ++SrcRegNum; + + emitRegModRMByte(MI.getOperand(SrcRegNum).getReg(), getX86RegNum(MI.getOperand(CurOp).getReg())); - CurOp += 2; - if (CurOp != NumOps) - emitConstant(MI.getOperand(CurOp++).getImm(), - X86II::getSizeOfImm(Desc->TSFlags)); + // 2 operands skipped with HasMemOp4, compensate accordingly + CurOp = HasMemOp4 ? SrcRegNum : SrcRegNum + 1; + if (HasVEX_4VOp3) + ++CurOp; break; - + } case X86II::MRMSrcMem: { int AddrOperands = X86::AddrNumOperands; + unsigned FirstMemOp = CurOp+1; + if (HasVEX_4V) { + ++AddrOperands; + ++FirstMemOp; // Skip the register source (which is encoded in VEX_VVVV). + } + if (HasMemOp4) // Skip second register source (encoded in I8IMM) + ++FirstMemOp; + + MCE.emitByte(BaseOpcode); intptr_t PCAdj = (CurOp + AddrOperands + 1 != NumOps) ? X86II::getSizeOfImm(Desc->TSFlags) : 0; - - MCE.emitByte(BaseOpcode); - emitMemModRMByte(MI, CurOp+1, getX86RegNum(MI.getOperand(CurOp).getReg()), - PCAdj); + emitMemModRMByte(MI, FirstMemOp, + getX86RegNum(MI.getOperand(CurOp).getReg()),PCAdj); CurOp += AddrOperands + 1; - if (CurOp != NumOps) - emitConstant(MI.getOperand(CurOp++).getImm(), - X86II::getSizeOfImm(Desc->TSFlags)); + if (HasVEX_4VOp3) + ++CurOp; break; } @@ -768,20 +1356,22 @@ void Emitter::emitInstruction(const MachineInstr &MI, case X86II::MRM2r: case X86II::MRM3r: case X86II::MRM4r: case X86II::MRM5r: case X86II::MRM6r: case X86II::MRM7r: { + if (HasVEX_4V) // Skip the register dst (which is encoded in VEX_VVVV). + ++CurOp; MCE.emitByte(BaseOpcode); emitRegModRMByte(MI.getOperand(CurOp++).getReg(), (Desc->TSFlags & X86II::FormMask)-X86II::MRM0r); if (CurOp == NumOps) break; - + const MachineOperand &MO1 = MI.getOperand(CurOp++); unsigned Size = X86II::getSizeOfImm(Desc->TSFlags); if (MO1.isImm()) { emitConstant(MO1.getImm(), Size); break; } - + unsigned rt = Is64BitMode ? X86::reloc_pcrel_word : (IsPIC ? X86::reloc_picrel_word : X86::reloc_absolute_word); if (Opcode == X86::MOV64ri32) @@ -803,8 +1393,10 @@ void Emitter::emitInstruction(const MachineInstr &MI, case X86II::MRM2m: case X86II::MRM3m: case X86II::MRM4m: case X86II::MRM5m: case X86II::MRM6m: case X86II::MRM7m: { + if (HasVEX_4V) // Skip the register dst (which is encoded in VEX_VVVV). + ++CurOp; intptr_t PCAdj = (CurOp + X86::AddrNumOperands != NumOps) ? - (MI.getOperand(CurOp+X86::AddrNumOperands).isImm() ? + (MI.getOperand(CurOp+X86::AddrNumOperands).isImm() ? X86II::getSizeOfImm(Desc->TSFlags) : 4) : 0; MCE.emitByte(BaseOpcode); @@ -814,14 +1406,14 @@ void Emitter::emitInstruction(const MachineInstr &MI, if (CurOp == NumOps) break; - + const MachineOperand &MO = MI.getOperand(CurOp++); unsigned Size = X86II::getSizeOfImm(Desc->TSFlags); if (MO.isImm()) { emitConstant(MO.getImm(), Size); break; } - + unsigned rt = Is64BitMode ? X86::reloc_pcrel_word : (IsPIC ? X86::reloc_picrel_word : X86::reloc_absolute_word); if (Opcode == X86::MOV64mi32) @@ -846,7 +1438,7 @@ void Emitter::emitInstruction(const MachineInstr &MI, getX86RegNum(MI.getOperand(CurOp).getReg())); ++CurOp; break; - + case X86II::MRM_C1: MCE.emitByte(BaseOpcode); MCE.emitByte(0xC1); @@ -859,6 +1451,14 @@ void Emitter::emitInstruction(const MachineInstr &MI, MCE.emitByte(BaseOpcode); MCE.emitByte(0xC9); break; + case X86II::MRM_CA: + MCE.emitByte(BaseOpcode); + MCE.emitByte(0xCA); + break; + case X86II::MRM_CB: + MCE.emitByte(BaseOpcode); + MCE.emitByte(0xCB); + break; case X86II::MRM_E8: MCE.emitByte(BaseOpcode); MCE.emitByte(0xE8); @@ -869,7 +1469,34 @@ void Emitter::emitInstruction(const MachineInstr &MI, break; } - if (!Desc->isVariadic() && CurOp != NumOps) { + while (CurOp != NumOps && NumOps - CurOp <= 2) { + // The last source register of a 4 operand instruction in AVX is encoded + // in bits[7:4] of a immediate byte. + if ((TSFlags >> X86II::VEXShift) & X86II::VEX_I8IMM) { + const MachineOperand &MO = MI.getOperand(HasMemOp4 ? MemOp4_I8IMMOperand + : CurOp); + ++CurOp; + unsigned RegNum = getX86RegNum(MO.getReg()) << 4; + if (X86II::isX86_64ExtendedReg(MO.getReg())) + RegNum |= 1 << 7; + // If there is an additional 5th operand it must be an immediate, which + // is encoded in bits[3:0] + if (CurOp != NumOps) { + const MachineOperand &MIMM = MI.getOperand(CurOp++); + if (MIMM.isImm()) { + unsigned Val = MIMM.getImm(); + assert(Val < 16 && "Immediate operand value out of range"); + RegNum |= Val; + } + } + emitConstant(RegNum, 1); + } else { + emitConstant(MI.getOperand(CurOp++).getImm(), + X86II::getSizeOfImm(Desc->TSFlags)); + } + } + + if (!MI.isVariadic() && CurOp != NumOps) { #ifndef NDEBUG dbgs() << "Cannot encode all operands of: " << MI << "\n"; #endif