llvm-6502/lib/Target/X86/X86CodeEmitter.cpp

1006 lines
35 KiB
C++
Raw Normal View History

//===-- X86/X86CodeEmitter.cpp - Convert X86 code to machine code ---------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file contains the pass that transforms the X86 machine instructions into
// relocatable machine code.
//
//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "x86-emitter"
#include "X86InstrInfo.h"
#include "X86JITInfo.h"
#include "X86Subtarget.h"
#include "X86TargetMachine.h"
#include "X86Relocations.h"
#include "X86.h"
#include "llvm/LLVMContext.h"
#include "llvm/PassManager.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/MC/MCCodeEmitter.h"
#include "llvm/MC/MCExpr.h"
#include "llvm/MC/MCInst.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Target/TargetOptions.h"
using namespace llvm;
STATISTIC(NumEmitted, "Number of machine instructions emitted");
namespace {
template<class CodeEmitter>
class Emitter : public MachineFunctionPass {
const X86InstrInfo *II;
const TargetData *TD;
X86TargetMachine &TM;
CodeEmitter &MCE;
MachineModuleInfo *MMI;
intptr_t PICBaseOffset;
bool Is64BitMode;
bool IsPIC;
public:
static char ID;
explicit Emitter(X86TargetMachine &tm, CodeEmitter &mce)
: 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),
MCE(mce), PICBaseOffset(0), Is64BitMode(is64),
IsPIC(TM.getRelocationModel() == Reloc::PIC_) {}
bool runOnMachineFunction(MachineFunction &MF);
virtual const char *getPassName() const {
return "X86 Machine Code Emitter";
}
void emitInstruction(MachineInstr &MI, const TargetInstrDesc *Desc);
void getAnalysisUsage(AnalysisUsage &AU) const {
AU.setPreservesAll();
AU.addRequired<MachineModuleInfo>();
MachineFunctionPass::getAnalysisUsage(AU);
}
private:
void emitPCRelativeBlockAddress(MachineBasicBlock *MBB);
void emitGlobalAddress(const GlobalValue *GV, unsigned Reloc,
intptr_t Disp = 0, intptr_t PCAdj = 0,
Make X86-64 in the Large model always emit 64-bit calls. The large code model is documented at http://www.x86-64.org/documentation/abi.pdf and says that calls should assume their target doesn't live within the 32-bit pc-relative offset that fits in the call instruction. To do this, we turn off the global-address->target-global-address conversion in X86TargetLowering::LowerCall(). The first attempt at this broke the lazy JIT because it can separate the movabs(imm->reg) from the actual call instruction. The lazy JIT receives the address of the movabs as a relocation and needs to record the return address from the call; and then when that call happens, it needs to patch the movabs with the newly-compiled target. We could thread the call instruction into the relocation and record the movabs<->call mapping explicitly, but that seems to require at least as much new complication in the code generator as this change. To fix this, we make lazy functions _always_ go through a call stub. You'd think we'd only have to force lazy calls through a stub on difficult platforms, but that turns out to break indirect calls through a function pointer. The right fix for that is to distinguish between calls and address-of operations on uncompiled functions, but that's complex enough to leave for someone else to do. Another attempt at this defined a new CALL64i pseudo-instruction, which expanded to a 2-instruction sequence in the assembly output and was special-cased in the X86CodeEmitter's emitInstruction() function. That broke indirect calls in the same way as above. This patch also removes a hack forcing Darwin to the small code model. Without far-call-stubs, the small code model requires things of the JITMemoryManager that the DefaultJITMemoryManager can't provide. Thanks to echristo for lots of testing! git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@88984 91177308-0d34-0410-b5e6-96231b3b80d8
2009-11-16 22:41:33 +00:00
bool Indirect = false);
void emitExternalSymbolAddress(const char *ES, unsigned Reloc);
void emitConstPoolAddress(unsigned CPI, unsigned Reloc, intptr_t Disp = 0,
intptr_t PCAdj = 0);
void emitJumpTableAddress(unsigned JTI, unsigned Reloc,
intptr_t PCAdj = 0);
void emitDisplacementField(const MachineOperand *RelocOp, int DispVal,
intptr_t Adj = 0, bool IsPCRel = true);
void emitRegModRMByte(unsigned ModRMReg, unsigned RegOpcodeField);
void emitRegModRMByte(unsigned RegOpcodeField);
void emitSIBByte(unsigned SS, unsigned Index, unsigned Base);
void emitConstant(uint64_t Val, unsigned Size);
void emitMemModRMByte(const MachineInstr &MI,
unsigned Op, unsigned RegOpcodeField,
intptr_t PCAdj = 0);
unsigned getX86RegNum(unsigned RegNo) const;
};
template<class CodeEmitter>
char Emitter<CodeEmitter>::ID = 0;
} // end anonymous namespace.
/// createX86CodeEmitterPass - Return a pass that emits the collected X86 code
/// to the specified templated MachineCodeEmitter object.
FunctionPass *llvm::createX86JITCodeEmitterPass(X86TargetMachine &TM,
JITCodeEmitter &JCE) {
return new Emitter<JITCodeEmitter>(TM, JCE);
}
template<class CodeEmitter>
bool Emitter<CodeEmitter>::runOnMachineFunction(MachineFunction &MF) {
MMI = &getAnalysis<MachineModuleInfo>();
MCE.setModuleInfo(MMI);
II = TM.getInstrInfo();
TD = TM.getTargetData();
Is64BitMode = TM.getSubtarget<X86Subtarget>().is64Bit();
IsPIC = TM.getRelocationModel() == Reloc::PIC_;
do {
DEBUG(dbgs() << "JITTing function '"
<< MF.getFunction()->getName() << "'\n");
MCE.startFunction(MF);
for (MachineFunction::iterator MBB = MF.begin(), E = MF.end();
MBB != E; ++MBB) {
MCE.StartMachineBasicBlock(MBB);
for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
I != E; ++I) {
const TargetInstrDesc &Desc = I->getDesc();
emitInstruction(*I, &Desc);
// MOVPC32r is basically a call plus a pop instruction.
if (Desc.getOpcode() == X86::MOVPC32r)
emitInstruction(*I, &II->get(X86::POP32r));
++NumEmitted; // Keep track of the # of mi's emitted
}
}
} while (MCE.finishFunction(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 TargetInstrDesc &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, TOI::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 (X86InstrInfo::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.
///
template<class CodeEmitter>
void Emitter<CodeEmitter>::emitPCRelativeBlockAddress(MachineBasicBlock *MBB) {
// Remember where this reference was and where it is to so we can
// deal with it later.
MCE.addRelocation(MachineRelocation::getBB(MCE.getCurrentPCOffset(),
X86::reloc_pcrel_word, MBB));
MCE.emitWordLE(0);
}
/// emitGlobalAddress - Emit the specified address to the code stream assuming
/// this is part of a "take the address of a global" instruction.
///
template<class CodeEmitter>
void Emitter<CodeEmitter>::emitGlobalAddress(const GlobalValue *GV,
unsigned Reloc,
intptr_t Disp /* = 0 */,
intptr_t PCAdj /* = 0 */,
bool Indirect /* = false */) {
intptr_t RelocCST = Disp;
if (Reloc == X86::reloc_picrel_word)
RelocCST = PICBaseOffset;
else if (Reloc == X86::reloc_pcrel_word)
RelocCST = PCAdj;
MachineRelocation MR = Indirect
? MachineRelocation::getIndirectSymbol(MCE.getCurrentPCOffset(), Reloc,
const_cast<GlobalValue *>(GV),
RelocCST, false)
: MachineRelocation::getGV(MCE.getCurrentPCOffset(), Reloc,
const_cast<GlobalValue *>(GV), RelocCST, false);
MCE.addRelocation(MR);
// The relocated value will be added to the displacement
if (Reloc == X86::reloc_absolute_dword)
MCE.emitDWordLE(Disp);
else
MCE.emitWordLE((int32_t)Disp);
}
/// emitExternalSymbolAddress - Arrange for the address of an external symbol to
/// be emitted to the current location in the function, and allow it to be PC
/// relative.
template<class CodeEmitter>
void Emitter<CodeEmitter>::emitExternalSymbolAddress(const char *ES,
unsigned Reloc) {
intptr_t RelocCST = (Reloc == X86::reloc_picrel_word) ? PICBaseOffset : 0;
// X86 never needs stubs because instruction selection will always pick
// an instruction sequence that is large enough to hold any address
// to a symbol.
// (see X86ISelLowering.cpp, near 2039: X86TargetLowering::LowerCall)
bool NeedStub = false;
MCE.addRelocation(MachineRelocation::getExtSym(MCE.getCurrentPCOffset(),
Reloc, ES, RelocCST,
0, NeedStub));
if (Reloc == X86::reloc_absolute_dword)
MCE.emitDWordLE(0);
else
MCE.emitWordLE(0);
}
/// emitConstPoolAddress - Arrange for the address of an constant pool
/// to be emitted to the current location in the function, and allow it to be PC
/// relative.
template<class CodeEmitter>
void Emitter<CodeEmitter>::emitConstPoolAddress(unsigned CPI, unsigned Reloc,
intptr_t Disp /* = 0 */,
intptr_t PCAdj /* = 0 */) {
intptr_t RelocCST = 0;
if (Reloc == X86::reloc_picrel_word)
RelocCST = PICBaseOffset;
else if (Reloc == X86::reloc_pcrel_word)
RelocCST = PCAdj;
MCE.addRelocation(MachineRelocation::getConstPool(MCE.getCurrentPCOffset(),
Reloc, CPI, RelocCST));
// The relocated value will be added to the displacement
if (Reloc == X86::reloc_absolute_dword)
MCE.emitDWordLE(Disp);
else
MCE.emitWordLE((int32_t)Disp);
}
/// emitJumpTableAddress - Arrange for the address of a jump table to
/// be emitted to the current location in the function, and allow it to be PC
/// relative.
template<class CodeEmitter>
void Emitter<CodeEmitter>::emitJumpTableAddress(unsigned JTI, unsigned Reloc,
intptr_t PCAdj /* = 0 */) {
intptr_t RelocCST = 0;
if (Reloc == X86::reloc_picrel_word)
RelocCST = PICBaseOffset;
else if (Reloc == X86::reloc_pcrel_word)
RelocCST = PCAdj;
MCE.addRelocation(MachineRelocation::getJumpTable(MCE.getCurrentPCOffset(),
Reloc, JTI, RelocCST));
// The relocated value will be added to the displacement
if (Reloc == X86::reloc_absolute_dword)
MCE.emitDWordLE(0);
else
MCE.emitWordLE(0);
}
template<class CodeEmitter>
unsigned Emitter<CodeEmitter>::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!");
return RM | (RegOpcode << 3) | (Mod << 6);
}
template<class CodeEmitter>
void Emitter<CodeEmitter>::emitRegModRMByte(unsigned ModRMReg,
unsigned RegOpcodeFld){
MCE.emitByte(ModRMByte(3, RegOpcodeFld, getX86RegNum(ModRMReg)));
}
template<class CodeEmitter>
void Emitter<CodeEmitter>::emitRegModRMByte(unsigned RegOpcodeFld) {
MCE.emitByte(ModRMByte(3, RegOpcodeFld, 0));
}
template<class CodeEmitter>
void Emitter<CodeEmitter>::emitSIBByte(unsigned SS,
unsigned Index,
unsigned Base) {
// SIB byte is in the same format as the ModRMByte...
MCE.emitByte(ModRMByte(SS, Index, Base));
}
template<class CodeEmitter>
void Emitter<CodeEmitter>::emitConstant(uint64_t Val, unsigned Size) {
// Output the constant in little endian byte order...
for (unsigned i = 0; i != Size; ++i) {
MCE.emitByte(Val & 255);
Val >>= 8;
}
}
/// 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;
}
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<X86Subtarget>().is64Bit() &&
!TM.getSubtarget<X86Subtarget>().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());
}
template<class CodeEmitter>
void Emitter<CodeEmitter>::emitDisplacementField(const MachineOperand *RelocOp,
int DispVal,
intptr_t Adj /* = 0 */,
bool IsPCRel /* = true */) {
// If this is a simple integer displacement that doesn't require a relocation,
// emit it now.
if (!RelocOp) {
emitConstant(DispVal, 4);
return;
}
// Otherwise, this is something that requires a relocation. Emit it as such
// now.
unsigned RelocType = Is64BitMode ?
(IsPCRel ? X86::reloc_pcrel_word : X86::reloc_absolute_word_sext)
: (IsPIC ? X86::reloc_picrel_word : X86::reloc_absolute_word);
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).
// 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);
emitGlobalAddress(RelocOp->getGlobal(), RelocType, RelocOp->getOffset(),
Make X86-64 in the Large model always emit 64-bit calls. The large code model is documented at http://www.x86-64.org/documentation/abi.pdf and says that calls should assume their target doesn't live within the 32-bit pc-relative offset that fits in the call instruction. To do this, we turn off the global-address->target-global-address conversion in X86TargetLowering::LowerCall(). The first attempt at this broke the lazy JIT because it can separate the movabs(imm->reg) from the actual call instruction. The lazy JIT receives the address of the movabs as a relocation and needs to record the return address from the call; and then when that call happens, it needs to patch the movabs with the newly-compiled target. We could thread the call instruction into the relocation and record the movabs<->call mapping explicitly, but that seems to require at least as much new complication in the code generator as this change. To fix this, we make lazy functions _always_ go through a call stub. You'd think we'd only have to force lazy calls through a stub on difficult platforms, but that turns out to break indirect calls through a function pointer. The right fix for that is to distinguish between calls and address-of operations on uncompiled functions, but that's complex enough to leave for someone else to do. Another attempt at this defined a new CALL64i pseudo-instruction, which expanded to a 2-instruction sequence in the assembly output and was special-cased in the X86CodeEmitter's emitInstruction() function. That broke indirect calls in the same way as above. This patch also removes a hack forcing Darwin to the small code model. Without far-call-stubs, the small code model requires things of the JITMemoryManager that the DefaultJITMemoryManager can't provide. Thanks to echristo for lots of testing! git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@88984 91177308-0d34-0410-b5e6-96231b3b80d8
2009-11-16 22:41:33 +00:00
Adj, Indirect);
} else if (RelocOp->isSymbol()) {
emitExternalSymbolAddress(RelocOp->getSymbolName(), RelocType);
} else if (RelocOp->isCPI()) {
emitConstPoolAddress(RelocOp->getIndex(), RelocType,
RelocOp->getOffset(), Adj);
} else {
assert(RelocOp->isJTI() && "Unexpected machine operand!");
emitJumpTableAddress(RelocOp->getIndex(), RelocType, Adj);
}
}
template<class CodeEmitter>
void Emitter<CodeEmitter>::emitMemModRMByte(const MachineInstr &MI,
unsigned Op,unsigned RegOpcodeField,
intptr_t PCAdj) {
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;
} else if (Op3.isSymbol()) {
DispForReloc = &Op3;
} else if (Op3.isCPI()) {
if (!MCE.earlyResolveAddresses() || Is64BitMode || IsPIC) {
DispForReloc = &Op3;
} else {
DispVal += MCE.getConstantPoolEntryAddress(Op3.getIndex());
DispVal += Op3.getOffset();
}
} else if (Op3.isJTI()) {
if (!MCE.earlyResolveAddresses() || Is64BitMode || IsPIC) {
DispForReloc = &Op3;
} else {
DispVal += MCE.getJumpTableEntryAddress(Op3.getIndex());
}
} else {
DispVal = Op3.getImm();
}
const MachineOperand &Base = MI.getOperand(Op);
const MachineOperand &Scale = MI.getOperand(Op+1);
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
assert(IndexReg.getReg() == 0 && Is64BitMode &&
"Invalid rip-relative address");
MCE.emitByte(ModRMByte(0, RegOpcodeField, 5));
emitDisplacementField(DispForReloc, DispVal, PCAdj, true);
return;
}
// Indicate that the displacement will use an pcrel or absolute reference
// by default. MCEs able to resolve addresses on-the-fly use pcrel by default
// while others, unless explicit asked to use RIP, use absolute references.
bool IsPCRel = MCE.earlyResolveAddresses() ? true : false;
// Is a SIB byte needed?
// 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;
if (BaseReg != 0 && BaseReg != X86::RIP)
BaseRegNo = getX86RegNum(BaseReg);
if (// The SIB byte must be used if there is an index register.
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.
BaseRegNo != N86::ESP &&
// If there is no base register and we're in 64-bit mode, we need a SIB
// byte to emit an addr that is just 'disp32' (the non-RIP relative form).
(!Is64BitMode || BaseReg != 0)) {
if (BaseReg == 0 || // [disp32] in X86-32 mode
BaseReg == X86::RIP) { // [disp32+RIP] in X86-64 mode
MCE.emitByte(ModRMByte(0, RegOpcodeField, 5));
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
// by emitting a displacement of 0 below.
if (!DispForReloc && DispVal == 0 && BaseRegNo != N86::EBP) {
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!");
bool ForceDisp32 = false;
bool ForceDisp8 = false;
if (BaseReg == 0) {
// If there is no base register, we emit the special case SIB byte with
// MOD=0, BASE=4, to JUST get the index, scale, and displacement.
MCE.emitByte(ModRMByte(0, RegOpcodeField, 4));
ForceDisp32 = true;
} else if (DispForReloc) {
// Emit the normal disp32 encoding.
MCE.emitByte(ModRMByte(2, RegOpcodeField, 4));
ForceDisp32 = true;
} else if (DispVal == 0 && BaseRegNo != N86::EBP) {
// Emit no displacement ModR/M byte
MCE.emitByte(ModRMByte(0, RegOpcodeField, 4));
} else if (isDisp8(DispVal)) {
// Emit the disp8 encoding...
MCE.emitByte(ModRMByte(1, RegOpcodeField, 4));
ForceDisp8 = true; // Make sure to force 8 bit disp if Base=EBP
} else {
// Emit the normal disp32 encoding...
MCE.emitByte(ModRMByte(2, RegOpcodeField, 4));
}
// Calculate what the SS field value should be...
static const unsigned SSTable[] = { ~0, 0, 1, ~0, 2, ~0, ~0, ~0, 3 };
unsigned SS = SSTable[Scale.getImm()];
if (BaseReg == 0) {
// 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())
IndexRegNo = getX86RegNum(IndexReg.getReg());
else // Examples: [ESP+1*<noreg>+4] or [scaled idx]+disp32 (MOD=0,BASE=5)
IndexRegNo = 4;
emitSIBByte(SS, IndexRegNo, 5);
} else {
unsigned BaseRegNo = getX86RegNum(BaseReg);
unsigned IndexRegNo;
if (IndexReg.getReg())
IndexRegNo = getX86RegNum(IndexReg.getReg());
else
IndexRegNo = 4; // For example [ESP+1*<noreg>+4]
emitSIBByte(SS, IndexRegNo, BaseRegNo);
}
// Do we need to output a displacement?
if (ForceDisp8) {
emitConstant(DispVal, 1);
} else if (DispVal != 0 || ForceDisp32) {
emitDisplacementField(DispForReloc, DispVal, PCAdj, IsPCRel);
}
}
template<class CodeEmitter>
void Emitter<CodeEmitter>::emitInstruction(MachineInstr &MI,
const TargetInstrDesc *Desc) {
DEBUG(dbgs() << MI);
// If this is a pseudo instruction, lower it.
switch (Desc->getOpcode()) {
case X86::ADD16rr_DB: Desc = &II->get(X86::OR16rr); MI.setDesc(*Desc);break;
case X86::ADD32rr_DB: Desc = &II->get(X86::OR32rr); MI.setDesc(*Desc);break;
case X86::ADD64rr_DB: Desc = &II->get(X86::OR64rr); MI.setDesc(*Desc);break;
case X86::ADD16ri_DB: Desc = &II->get(X86::OR16ri); MI.setDesc(*Desc);break;
case X86::ADD32ri_DB: Desc = &II->get(X86::OR32ri); MI.setDesc(*Desc);break;
case X86::ADD64ri32_DB:Desc = &II->get(X86::OR64ri32);MI.setDesc(*Desc);break;
case X86::ADD16ri8_DB: Desc = &II->get(X86::OR16ri8);MI.setDesc(*Desc);break;
case X86::ADD32ri8_DB: Desc = &II->get(X86::OR32ri8);MI.setDesc(*Desc);break;
case X86::ADD64ri8_DB: Desc = &II->get(X86::OR64ri8);MI.setDesc(*Desc);break;
}
MCE.processDebugLoc(MI.getDebugLoc(), true);
Add line numbers to OProfile. To do this, I added a processDebugLoc() call to the MachineCodeEmitter interface and made copying the start line of a function not conditional on whether we're emitting Dwarf debug information. I'll propagate the processDebugLoc() calls to the non-X86 targets in a followup patch. In the long run, it'll probably be better to gather this information through the DwarfWriter, but the DwarfWriter currently depends on the AsmPrinter and TargetAsmInfo, and fixing that would be out of the way for this patch. There's a bug in OProfile 0.9.4 that makes it ignore line numbers for addresses above 4G, and a patch fixing it at http://thread.gmane.org/gmane.linux.oprofile/7634 Sample output: $ sudo opcontrol --reset; sudo opcontrol --start-daemon; sudo opcontrol --start; `pwd`/Debug/bin/lli fib.bc; sudo opcontrol --stop Signalling daemon... done Profiler running. fib(40) == 165580141 Stopping profiling. $ opreport -g -d -l `pwd`/Debug/bin/lli|head -60 Overflow stats not available CPU: Core 2, speed 1998 MHz (estimated) Counted CPU_CLK_UNHALTED events (Clock cycles when not halted) with a unit mask of 0x00 (Unhalted core cycles) count 100000 vma samples % linenr info image name symbol name 00007f67a30370b0 25489 61.2554 fib.c:24 10946.jo fib_left 00007f67a30370b0 1634 6.4106 fib.c:24 00007f67a30370b1 83 0.3256 fib.c:24 00007f67a30370b9 1997 7.8348 fib.c:24 00007f67a30370c6 2080 8.1604 fib.c:27 00007f67a30370c8 988 3.8762 fib.c:27 00007f67a30370cd 1315 5.1591 fib.c:27 00007f67a30370cf 251 0.9847 fib.c:27 00007f67a30370d3 1191 4.6726 fib.c:27 00007f67a30370d6 975 3.8252 fib.c:27 00007f67a30370db 1010 3.9625 fib.c:27 00007f67a30370dd 242 0.9494 fib.c:27 00007f67a30370e1 2782 10.9145 fib.c:28 00007f67a30370e5 3768 14.7828 fib.c:28 00007f67a30370eb 615 2.4128 (no location information) 00007f67a30370f3 6558 25.7287 (no location information) 00007f67a3037100 15603 37.4973 fib.c:29 10946.jo fib_right 00007f67a3037100 1646 10.5493 fib.c:29 00007f67a3037101 45 0.2884 fib.c:29 00007f67a3037109 2372 15.2022 fib.c:29 00007f67a3037116 2234 14.3178 fib.c:32 00007f67a3037118 612 3.9223 fib.c:32 00007f67a303711d 622 3.9864 fib.c:32 00007f67a303711f 385 2.4675 fib.c:32 00007f67a3037123 404 2.5892 fib.c:32 00007f67a3037126 634 4.0633 fib.c:32 00007f67a303712b 870 5.5759 fib.c:32 00007f67a303712d 62 0.3974 fib.c:32 00007f67a3037131 1848 11.8439 fib.c:33 00007f67a3037135 2840 18.2016 fib.c:33 00007f67a303713a 1 0.0064 fib.c:33 00007f67a303713b 1023 6.5564 (no location information) 00007f67a3037143 5 0.0320 (no location information) 000000000080c1e4 15 0.0360 MachineOperand.h:150 lli llvm::MachineOperand::isReg() const 000000000080c1e4 6 40.0000 MachineOperand.h:150 000000000080c1ec 2 13.3333 MachineOperand.h:150 ... git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@76102 91177308-0d34-0410-b5e6-96231b3b80d8
2009-07-16 21:07:26 +00:00
unsigned Opcode = Desc->Opcode;
// 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!
}
// 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)
MCE.emitByte(0x67);
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
case X86II::A6: // 0F A6
case X86II::A7: // 0F A7
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!
}
// Handle REX prefix.
if (Is64BitMode) {
if (unsigned REX = determineREX(MI))
MCE.emitByte(0x40 | REX);
}
// 0x0F escape code must be emitted just before the opcode.
if (Need0FPrefix)
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::A6: // 0F A6
MCE.emitByte(0xA6);
break;
case X86II::A7: // 0F A7
MCE.emitByte(0xA7);
break;
}
// 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)
++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;
unsigned char BaseOpcode = X86II::getBaseOpcodeFor(Desc->TSFlags);
switch (Desc->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("pseudo instructions should be removed before code"
" emission");
break;
// 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.
if (MI.getOperand(0).getSymbolName()[0])
report_fatal_error("JIT does not support inline asm!");
break;
case TargetOpcode::PROLOG_LABEL:
case TargetOpcode::GC_LABEL:
case TargetOpcode::EH_LABEL:
MCE.emitLabel(MI.getOperand(0).getMCSymbol());
break;
case TargetOpcode::IMPLICIT_DEF:
case TargetOpcode::KILL:
break;
case X86::MOVPC32r: {
// This emits the "call" portion of this pseudo instruction.
MCE.emitByte(BaseOpcode);
emitConstant(0, X86II::getSizeOfImm(Desc->TSFlags));
// Remember PIC base.
PICBaseOffset = (intptr_t) MCE.getCurrentPCOffset();
X86JITInfo *JTI = TM.getJITInfo();
JTI->setPICBase(MCE.getCurrentPCValue());
break;
}
}
CurOp = NumOps;
break;
case X86II::RawFrm: {
MCE.emitByte(BaseOpcode);
if (CurOp == NumOps)
break;
const MachineOperand &MO = MI.getOperand(CurOp++);
DEBUG(dbgs() << "RawFrm CurOp " << CurOp << "\n");
DEBUG(dbgs() << "isMBB " << MO.isMBB() << "\n");
DEBUG(dbgs() << "isGlobal " << MO.isGlobal() << "\n");
DEBUG(dbgs() << "isSymbol " << MO.isSymbol() << "\n");
DEBUG(dbgs() << "isImm " << MO.isImm() << "\n");
if (MO.isMBB()) {
emitPCRelativeBlockAddress(MO.getMBB());
break;
}
if (MO.isGlobal()) {
emitGlobalAddress(MO.getGlobal(), X86::reloc_pcrel_word,
Make X86-64 in the Large model always emit 64-bit calls. The large code model is documented at http://www.x86-64.org/documentation/abi.pdf and says that calls should assume their target doesn't live within the 32-bit pc-relative offset that fits in the call instruction. To do this, we turn off the global-address->target-global-address conversion in X86TargetLowering::LowerCall(). The first attempt at this broke the lazy JIT because it can separate the movabs(imm->reg) from the actual call instruction. The lazy JIT receives the address of the movabs as a relocation and needs to record the return address from the call; and then when that call happens, it needs to patch the movabs with the newly-compiled target. We could thread the call instruction into the relocation and record the movabs<->call mapping explicitly, but that seems to require at least as much new complication in the code generator as this change. To fix this, we make lazy functions _always_ go through a call stub. You'd think we'd only have to force lazy calls through a stub on difficult platforms, but that turns out to break indirect calls through a function pointer. The right fix for that is to distinguish between calls and address-of operations on uncompiled functions, but that's complex enough to leave for someone else to do. Another attempt at this defined a new CALL64i pseudo-instruction, which expanded to a 2-instruction sequence in the assembly output and was special-cased in the X86CodeEmitter's emitInstruction() function. That broke indirect calls in the same way as above. This patch also removes a hack forcing Darwin to the small code model. Without far-call-stubs, the small code model requires things of the JITMemoryManager that the DefaultJITMemoryManager can't provide. Thanks to echristo for lots of testing! git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@88984 91177308-0d34-0410-b5e6-96231b3b80d8
2009-11-16 22:41:33 +00:00
MO.getOffset(), 0);
break;
}
if (MO.isSymbol()) {
emitExternalSymbolAddress(MO.getSymbolName(), X86::reloc_pcrel_word);
break;
}
// FIXME: Only used by hackish MCCodeEmitter, remove when dead.
if (MO.isJTI()) {
emitJumpTableAddress(MO.getIndex(), X86::reloc_pcrel_word);
break;
}
assert(MO.isImm() && "Unknown RawFrm operand!");
if (Opcode == X86::CALLpcrel32 || Opcode == X86::CALL64pcrel32 ||
Opcode == X86::WINCALL64pcrel32) {
// Fix up immediate operand for pc relative calls.
intptr_t Imm = (intptr_t)MO.getImm();
Imm = Imm - MCE.getCurrentPCValue() - 4;
emitConstant(Imm, X86II::getSizeOfImm(Desc->TSFlags));
} else
emitConstant(MO.getImm(), X86II::getSizeOfImm(Desc->TSFlags));
break;
}
case X86II::AddRegFrm: {
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)
rt = X86::reloc_absolute_word; // FIXME: add X86II flag?
// This should not occur on Darwin for relocatable objects.
if (Opcode == X86::MOV64ri)
rt = X86::reloc_absolute_dword; // FIXME: add X86II flag?
if (MO1.isGlobal()) {
bool Indirect = gvNeedsNonLazyPtr(MO1, TM);
emitGlobalAddress(MO1.getGlobal(), rt, MO1.getOffset(), 0,
Make X86-64 in the Large model always emit 64-bit calls. The large code model is documented at http://www.x86-64.org/documentation/abi.pdf and says that calls should assume their target doesn't live within the 32-bit pc-relative offset that fits in the call instruction. To do this, we turn off the global-address->target-global-address conversion in X86TargetLowering::LowerCall(). The first attempt at this broke the lazy JIT because it can separate the movabs(imm->reg) from the actual call instruction. The lazy JIT receives the address of the movabs as a relocation and needs to record the return address from the call; and then when that call happens, it needs to patch the movabs with the newly-compiled target. We could thread the call instruction into the relocation and record the movabs<->call mapping explicitly, but that seems to require at least as much new complication in the code generator as this change. To fix this, we make lazy functions _always_ go through a call stub. You'd think we'd only have to force lazy calls through a stub on difficult platforms, but that turns out to break indirect calls through a function pointer. The right fix for that is to distinguish between calls and address-of operations on uncompiled functions, but that's complex enough to leave for someone else to do. Another attempt at this defined a new CALL64i pseudo-instruction, which expanded to a 2-instruction sequence in the assembly output and was special-cased in the X86CodeEmitter's emitInstruction() function. That broke indirect calls in the same way as above. This patch also removes a hack forcing Darwin to the small code model. Without far-call-stubs, the small code model requires things of the JITMemoryManager that the DefaultJITMemoryManager can't provide. Thanks to echristo for lots of testing! git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@88984 91177308-0d34-0410-b5e6-96231b3b80d8
2009-11-16 22:41:33 +00:00
Indirect);
} else if (MO1.isSymbol())
emitExternalSymbolAddress(MO1.getSymbolName(), rt);
else if (MO1.isCPI())
emitConstPoolAddress(MO1.getIndex(), rt);
else if (MO1.isJTI())
emitJumpTableAddress(MO1.getIndex(), rt);
break;
}
case X86II::MRMDestReg: {
MCE.emitByte(BaseOpcode);
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));
break;
}
case X86II::MRMDestMem: {
MCE.emitByte(BaseOpcode);
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));
break;
}
case X86II::MRMSrcReg:
MCE.emitByte(BaseOpcode);
emitRegModRMByte(MI.getOperand(CurOp+1).getReg(),
getX86RegNum(MI.getOperand(CurOp).getReg()));
CurOp += 2;
if (CurOp != NumOps)
emitConstant(MI.getOperand(CurOp++).getImm(),
X86II::getSizeOfImm(Desc->TSFlags));
break;
case X86II::MRMSrcMem: {
int AddrOperands = X86::AddrNumOperands;
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);
CurOp += AddrOperands + 1;
if (CurOp != NumOps)
emitConstant(MI.getOperand(CurOp++).getImm(),
X86II::getSizeOfImm(Desc->TSFlags));
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: {
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)
rt = X86::reloc_absolute_word_sext; // FIXME: add X86II flag?
if (MO1.isGlobal()) {
bool Indirect = gvNeedsNonLazyPtr(MO1, TM);
emitGlobalAddress(MO1.getGlobal(), rt, MO1.getOffset(), 0,
Make X86-64 in the Large model always emit 64-bit calls. The large code model is documented at http://www.x86-64.org/documentation/abi.pdf and says that calls should assume their target doesn't live within the 32-bit pc-relative offset that fits in the call instruction. To do this, we turn off the global-address->target-global-address conversion in X86TargetLowering::LowerCall(). The first attempt at this broke the lazy JIT because it can separate the movabs(imm->reg) from the actual call instruction. The lazy JIT receives the address of the movabs as a relocation and needs to record the return address from the call; and then when that call happens, it needs to patch the movabs with the newly-compiled target. We could thread the call instruction into the relocation and record the movabs<->call mapping explicitly, but that seems to require at least as much new complication in the code generator as this change. To fix this, we make lazy functions _always_ go through a call stub. You'd think we'd only have to force lazy calls through a stub on difficult platforms, but that turns out to break indirect calls through a function pointer. The right fix for that is to distinguish between calls and address-of operations on uncompiled functions, but that's complex enough to leave for someone else to do. Another attempt at this defined a new CALL64i pseudo-instruction, which expanded to a 2-instruction sequence in the assembly output and was special-cased in the X86CodeEmitter's emitInstruction() function. That broke indirect calls in the same way as above. This patch also removes a hack forcing Darwin to the small code model. Without far-call-stubs, the small code model requires things of the JITMemoryManager that the DefaultJITMemoryManager can't provide. Thanks to echristo for lots of testing! git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@88984 91177308-0d34-0410-b5e6-96231b3b80d8
2009-11-16 22:41:33 +00:00
Indirect);
} else if (MO1.isSymbol())
emitExternalSymbolAddress(MO1.getSymbolName(), rt);
else if (MO1.isCPI())
emitConstPoolAddress(MO1.getIndex(), rt);
else if (MO1.isJTI())
emitJumpTableAddress(MO1.getIndex(), rt);
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: {
intptr_t PCAdj = (CurOp + X86::AddrNumOperands != NumOps) ?
(MI.getOperand(CurOp+X86::AddrNumOperands).isImm() ?
X86II::getSizeOfImm(Desc->TSFlags) : 4) : 0;
MCE.emitByte(BaseOpcode);
emitMemModRMByte(MI, CurOp, (Desc->TSFlags & X86II::FormMask)-X86II::MRM0m,
PCAdj);
CurOp += X86::AddrNumOperands;
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)
rt = X86::reloc_absolute_word_sext; // FIXME: add X86II flag?
if (MO.isGlobal()) {
bool Indirect = gvNeedsNonLazyPtr(MO, TM);
emitGlobalAddress(MO.getGlobal(), rt, MO.getOffset(), 0,
Make X86-64 in the Large model always emit 64-bit calls. The large code model is documented at http://www.x86-64.org/documentation/abi.pdf and says that calls should assume their target doesn't live within the 32-bit pc-relative offset that fits in the call instruction. To do this, we turn off the global-address->target-global-address conversion in X86TargetLowering::LowerCall(). The first attempt at this broke the lazy JIT because it can separate the movabs(imm->reg) from the actual call instruction. The lazy JIT receives the address of the movabs as a relocation and needs to record the return address from the call; and then when that call happens, it needs to patch the movabs with the newly-compiled target. We could thread the call instruction into the relocation and record the movabs<->call mapping explicitly, but that seems to require at least as much new complication in the code generator as this change. To fix this, we make lazy functions _always_ go through a call stub. You'd think we'd only have to force lazy calls through a stub on difficult platforms, but that turns out to break indirect calls through a function pointer. The right fix for that is to distinguish between calls and address-of operations on uncompiled functions, but that's complex enough to leave for someone else to do. Another attempt at this defined a new CALL64i pseudo-instruction, which expanded to a 2-instruction sequence in the assembly output and was special-cased in the X86CodeEmitter's emitInstruction() function. That broke indirect calls in the same way as above. This patch also removes a hack forcing Darwin to the small code model. Without far-call-stubs, the small code model requires things of the JITMemoryManager that the DefaultJITMemoryManager can't provide. Thanks to echristo for lots of testing! git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@88984 91177308-0d34-0410-b5e6-96231b3b80d8
2009-11-16 22:41:33 +00:00
Indirect);
} else if (MO.isSymbol())
emitExternalSymbolAddress(MO.getSymbolName(), rt);
else if (MO.isCPI())
emitConstPoolAddress(MO.getIndex(), rt);
else if (MO.isJTI())
emitJumpTableAddress(MO.getIndex(), rt);
break;
}
case X86II::MRMInitReg:
MCE.emitByte(BaseOpcode);
// Duplicate register, used by things like MOV8r0 (aka xor reg,reg).
emitRegModRMByte(MI.getOperand(CurOp).getReg(),
getX86RegNum(MI.getOperand(CurOp).getReg()));
++CurOp;
break;
case X86II::MRM_C1:
MCE.emitByte(BaseOpcode);
MCE.emitByte(0xC1);
break;
case X86II::MRM_C8:
MCE.emitByte(BaseOpcode);
MCE.emitByte(0xC8);
break;
case X86II::MRM_C9:
MCE.emitByte(BaseOpcode);
MCE.emitByte(0xC9);
break;
case X86II::MRM_E8:
MCE.emitByte(BaseOpcode);
MCE.emitByte(0xE8);
break;
case X86II::MRM_F0:
MCE.emitByte(BaseOpcode);
MCE.emitByte(0xF0);
break;
}
if (!Desc->isVariadic() && CurOp != NumOps) {
#ifndef NDEBUG
dbgs() << "Cannot encode all operands of: " << MI << "\n";
#endif
llvm_unreachable(0);
}
MCE.processDebugLoc(MI.getDebugLoc(), false);
}