1 //===-- X86/X86CodeEmitter.cpp - Convert X86 code to machine code ---------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains the pass that transforms the X86 machine instructions into
11 // relocatable machine code.
13 //===----------------------------------------------------------------------===//
15 #include "X86TargetMachine.h"
16 #include "X86Relocations.h"
18 #include "llvm/PassManager.h"
19 #include "llvm/CodeGen/MachineCodeEmitter.h"
20 #include "llvm/CodeGen/MachineFunctionPass.h"
21 #include "llvm/CodeGen/MachineInstr.h"
22 #include "llvm/CodeGen/Passes.h"
23 #include "llvm/Function.h"
24 #include "llvm/ADT/Statistic.h"
29 NumEmitted("x86-emitter", "Number of machine instructions emitted");
33 class Emitter : public MachineFunctionPass {
34 const X86InstrInfo *II;
35 MachineCodeEmitter &MCE;
36 std::map<const MachineBasicBlock*, unsigned> BasicBlockAddrs;
37 std::vector<std::pair<const MachineBasicBlock *, unsigned> > BBRefs;
39 explicit Emitter(MachineCodeEmitter &mce) : II(0), MCE(mce) {}
40 Emitter(MachineCodeEmitter &mce, const X86InstrInfo& ii)
41 : II(&ii), MCE(mce) {}
43 bool runOnMachineFunction(MachineFunction &MF);
45 virtual const char *getPassName() const {
46 return "X86 Machine Code Emitter";
49 void emitInstruction(const MachineInstr &MI);
52 void emitBasicBlock(const MachineBasicBlock &MBB);
54 void emitPCRelativeBlockAddress(const MachineBasicBlock *BB);
55 void emitPCRelativeValue(unsigned Address);
56 void emitGlobalAddressForCall(GlobalValue *GV, bool isTailCall);
57 void emitGlobalAddressForPtr(GlobalValue *GV, int Disp = 0);
58 void emitExternalSymbolAddress(const char *ES, bool isPCRelative,
61 void emitRegModRMByte(unsigned ModRMReg, unsigned RegOpcodeField);
62 void emitSIBByte(unsigned SS, unsigned Index, unsigned Base);
63 void emitConstant(unsigned Val, unsigned Size);
65 void emitMemModRMByte(const MachineInstr &MI,
66 unsigned Op, unsigned RegOpcodeField);
71 /// addPassesToEmitMachineCode - Add passes to the specified pass manager to get
72 /// machine code emitted. This uses a MachineCodeEmitter object to handle
73 /// actually outputting the machine code and resolving things like the address
74 /// of functions. This method should returns true if machine code emission is
77 bool X86TargetMachine::addPassesToEmitMachineCode(FunctionPassManager &PM,
78 MachineCodeEmitter &MCE) {
79 PM.add(new Emitter(MCE));
80 // Delete machine code for this function
81 PM.add(createMachineCodeDeleter());
85 bool Emitter::runOnMachineFunction(MachineFunction &MF) {
86 II = ((X86TargetMachine&)MF.getTarget()).getInstrInfo();
88 MCE.startFunction(MF);
89 MCE.emitConstantPool(MF.getConstantPool());
90 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I)
92 MCE.finishFunction(MF);
94 // Resolve all forward branches now...
95 for (unsigned i = 0, e = BBRefs.size(); i != e; ++i) {
96 unsigned Location = BasicBlockAddrs[BBRefs[i].first];
97 unsigned Ref = BBRefs[i].second;
98 MCE.emitWordAt(Location-Ref-4, (unsigned*)(intptr_t)Ref);
101 BasicBlockAddrs.clear();
105 void Emitter::emitBasicBlock(const MachineBasicBlock &MBB) {
106 if (uint64_t Addr = MCE.getCurrentPCValue())
107 BasicBlockAddrs[&MBB] = Addr;
109 for (MachineBasicBlock::const_iterator I = MBB.begin(), E = MBB.end();
114 /// emitPCRelativeValue - Emit a 32-bit PC relative address.
116 void Emitter::emitPCRelativeValue(unsigned Address) {
117 MCE.emitWord(Address-MCE.getCurrentPCValue()-4);
120 /// emitPCRelativeBlockAddress - This method emits the PC relative address of
121 /// the specified basic block, or if the basic block hasn't been emitted yet
122 /// (because this is a forward branch), it keeps track of the information
123 /// necessary to resolve this address later (and emits a dummy value).
125 void Emitter::emitPCRelativeBlockAddress(const MachineBasicBlock *MBB) {
126 // If this is a backwards branch, we already know the address of the target,
127 // so just emit the value.
128 std::map<const MachineBasicBlock*, unsigned>::iterator I =
129 BasicBlockAddrs.find(MBB);
130 if (I != BasicBlockAddrs.end()) {
131 emitPCRelativeValue(I->second);
133 // Otherwise, remember where this reference was and where it is to so we can
134 // deal with it later.
135 BBRefs.push_back(std::make_pair(MBB, MCE.getCurrentPCValue()));
140 /// emitGlobalAddressForCall - Emit the specified address to the code stream
141 /// assuming this is part of a function call, which is PC relative.
143 void Emitter::emitGlobalAddressForCall(GlobalValue *GV, bool isTailCall) {
144 MCE.addRelocation(MachineRelocation(MCE.getCurrentPCOffset(),
145 X86::reloc_pcrel_word, GV, 0,
146 !isTailCall /*Doesn'tNeedStub*/));
150 /// emitGlobalAddress - Emit the specified address to the code stream assuming
151 /// this is part of a "take the address of a global" instruction, which is not
154 void Emitter::emitGlobalAddressForPtr(GlobalValue *GV, int Disp /* = 0 */) {
155 MCE.addRelocation(MachineRelocation(MCE.getCurrentPCOffset(),
156 X86::reloc_absolute_word, GV));
157 MCE.emitWord(Disp); // The relocated value will be added to the displacement
160 /// emitExternalSymbolAddress - Arrange for the address of an external symbol to
161 /// be emitted to the current location in the function, and allow it to be PC
163 void Emitter::emitExternalSymbolAddress(const char *ES, bool isPCRelative,
165 MCE.addRelocation(MachineRelocation(MCE.getCurrentPCOffset(),
166 isPCRelative ? X86::reloc_pcrel_word : X86::reloc_absolute_word, ES));
170 /// N86 namespace - Native X86 Register numbers... used by X86 backend.
174 EAX = 0, ECX = 1, EDX = 2, EBX = 3, ESP = 4, EBP = 5, ESI = 6, EDI = 7
179 // getX86RegNum - This function maps LLVM register identifiers to their X86
180 // specific numbering, which is used in various places encoding instructions.
182 static unsigned getX86RegNum(unsigned RegNo) {
184 case X86::EAX: case X86::AX: case X86::AL: return N86::EAX;
185 case X86::ECX: case X86::CX: case X86::CL: return N86::ECX;
186 case X86::EDX: case X86::DX: case X86::DL: return N86::EDX;
187 case X86::EBX: case X86::BX: case X86::BL: return N86::EBX;
188 case X86::ESP: case X86::SP: case X86::AH: return N86::ESP;
189 case X86::EBP: case X86::BP: case X86::CH: return N86::EBP;
190 case X86::ESI: case X86::SI: case X86::DH: return N86::ESI;
191 case X86::EDI: case X86::DI: case X86::BH: return N86::EDI;
193 case X86::ST0: case X86::ST1: case X86::ST2: case X86::ST3:
194 case X86::ST4: case X86::ST5: case X86::ST6: case X86::ST7:
195 return RegNo-X86::ST0;
197 assert(MRegisterInfo::isVirtualRegister(RegNo) &&
198 "Unknown physical register!");
199 assert(0 && "Register allocator hasn't allocated reg correctly yet!");
204 inline static unsigned char ModRMByte(unsigned Mod, unsigned RegOpcode,
206 assert(Mod < 4 && RegOpcode < 8 && RM < 8 && "ModRM Fields out of range!");
207 return RM | (RegOpcode << 3) | (Mod << 6);
210 void Emitter::emitRegModRMByte(unsigned ModRMReg, unsigned RegOpcodeFld){
211 MCE.emitByte(ModRMByte(3, RegOpcodeFld, getX86RegNum(ModRMReg)));
214 void Emitter::emitSIBByte(unsigned SS, unsigned Index, unsigned Base) {
215 // SIB byte is in the same format as the ModRMByte...
216 MCE.emitByte(ModRMByte(SS, Index, Base));
219 void Emitter::emitConstant(unsigned Val, unsigned Size) {
220 // Output the constant in little endian byte order...
221 for (unsigned i = 0; i != Size; ++i) {
222 MCE.emitByte(Val & 255);
227 static bool isDisp8(int Value) {
228 return Value == (signed char)Value;
231 void Emitter::emitMemModRMByte(const MachineInstr &MI,
232 unsigned Op, unsigned RegOpcodeField) {
233 const MachineOperand &Op3 = MI.getOperand(Op+3);
237 if (Op3.isGlobalAddress()) {
238 GV = Op3.getGlobal();
239 DispVal = Op3.getOffset();
241 DispVal = Op3.getImmedValue();
244 const MachineOperand &Base = MI.getOperand(Op);
245 const MachineOperand &Scale = MI.getOperand(Op+1);
246 const MachineOperand &IndexReg = MI.getOperand(Op+2);
248 unsigned BaseReg = 0;
250 if (Base.isConstantPoolIndex()) {
251 // Emit a direct address reference [disp32] where the displacement of the
252 // constant pool entry is controlled by the MCE.
253 assert(!GV && "Constant Pool reference cannot be relative to global!");
254 DispVal += MCE.getConstantPoolEntryAddress(Base.getConstantPoolIndex());
256 BaseReg = Base.getReg();
259 // Is a SIB byte needed?
260 if (IndexReg.getReg() == 0 && BaseReg != X86::ESP) {
261 if (BaseReg == 0) { // Just a displacement?
262 // Emit special case [disp32] encoding
263 MCE.emitByte(ModRMByte(0, RegOpcodeField, 5));
265 emitGlobalAddressForPtr(GV, DispVal);
267 emitConstant(DispVal, 4);
269 unsigned BaseRegNo = getX86RegNum(BaseReg);
271 // Emit the most general non-SIB encoding: [REG+disp32]
272 MCE.emitByte(ModRMByte(2, RegOpcodeField, BaseRegNo));
273 emitGlobalAddressForPtr(GV, DispVal);
274 } else if (DispVal == 0 && BaseRegNo != N86::EBP) {
275 // Emit simple indirect register encoding... [EAX] f.e.
276 MCE.emitByte(ModRMByte(0, RegOpcodeField, BaseRegNo));
277 } else if (isDisp8(DispVal)) {
278 // Emit the disp8 encoding... [REG+disp8]
279 MCE.emitByte(ModRMByte(1, RegOpcodeField, BaseRegNo));
280 emitConstant(DispVal, 1);
282 // Emit the most general non-SIB encoding: [REG+disp32]
283 MCE.emitByte(ModRMByte(2, RegOpcodeField, BaseRegNo));
284 emitConstant(DispVal, 4);
288 } else { // We need a SIB byte, so start by outputting the ModR/M byte first
289 assert(IndexReg.getReg() != X86::ESP && "Cannot use ESP as index reg!");
291 bool ForceDisp32 = false;
292 bool ForceDisp8 = false;
294 // If there is no base register, we emit the special case SIB byte with
295 // MOD=0, BASE=5, to JUST get the index, scale, and displacement.
296 MCE.emitByte(ModRMByte(0, RegOpcodeField, 4));
299 // Emit the normal disp32 encoding...
300 MCE.emitByte(ModRMByte(2, RegOpcodeField, 4));
302 } else if (DispVal == 0 && BaseReg != X86::EBP) {
303 // Emit no displacement ModR/M byte
304 MCE.emitByte(ModRMByte(0, RegOpcodeField, 4));
305 } else if (isDisp8(DispVal)) {
306 // Emit the disp8 encoding...
307 MCE.emitByte(ModRMByte(1, RegOpcodeField, 4));
308 ForceDisp8 = true; // Make sure to force 8 bit disp if Base=EBP
310 // Emit the normal disp32 encoding...
311 MCE.emitByte(ModRMByte(2, RegOpcodeField, 4));
314 // Calculate what the SS field value should be...
315 static const unsigned SSTable[] = { ~0, 0, 1, ~0, 2, ~0, ~0, ~0, 3 };
316 unsigned SS = SSTable[Scale.getImmedValue()];
319 // Handle the SIB byte for the case where there is no base. The
320 // displacement has already been output.
321 assert(IndexReg.getReg() && "Index register must be specified!");
322 emitSIBByte(SS, getX86RegNum(IndexReg.getReg()), 5);
324 unsigned BaseRegNo = getX86RegNum(BaseReg);
326 if (IndexReg.getReg())
327 IndexRegNo = getX86RegNum(IndexReg.getReg());
329 IndexRegNo = 4; // For example [ESP+1*<noreg>+4]
330 emitSIBByte(SS, IndexRegNo, BaseRegNo);
333 // Do we need to output a displacement?
334 if (DispVal != 0 || ForceDisp32 || ForceDisp8) {
335 if (!ForceDisp32 && isDisp8(DispVal))
336 emitConstant(DispVal, 1);
338 emitGlobalAddressForPtr(GV, DispVal);
340 emitConstant(DispVal, 4);
345 static unsigned sizeOfImm(const TargetInstrDescriptor &Desc) {
346 switch (Desc.TSFlags & X86II::ImmMask) {
347 case X86II::Imm8: return 1;
348 case X86II::Imm16: return 2;
349 case X86II::Imm32: return 4;
350 default: assert(0 && "Immediate size not set!");
355 void Emitter::emitInstruction(const MachineInstr &MI) {
356 NumEmitted++; // Keep track of the # of mi's emitted
358 unsigned Opcode = MI.getOpcode();
359 const TargetInstrDescriptor &Desc = II->get(Opcode);
361 // Emit the repeat opcode prefix as needed.
362 if ((Desc.TSFlags & X86II::Op0Mask) == X86II::REP) MCE.emitByte(0xF3);
364 // Emit instruction prefixes if necessary
365 if (Desc.TSFlags & X86II::OpSize) MCE.emitByte(0x66);// Operand size...
367 switch (Desc.TSFlags & X86II::Op0Mask) {
369 MCE.emitByte(0x0F); // Two-byte opcode prefix
371 case X86II::REP: break; // already handled.
372 case X86II::D8: case X86II::D9: case X86II::DA: case X86II::DB:
373 case X86II::DC: case X86II::DD: case X86II::DE: case X86II::DF:
375 (((Desc.TSFlags & X86II::Op0Mask)-X86II::D8)
376 >> X86II::Op0Shift));
377 break; // Two-byte opcode prefix
378 default: assert(0 && "Invalid prefix!");
379 case 0: break; // No prefix!
382 unsigned char BaseOpcode = II->getBaseOpcodeFor(Opcode);
383 switch (Desc.TSFlags & X86II::FormMask) {
384 default: assert(0 && "Unknown FormMask value in X86 MachineCodeEmitter!");
386 if (Opcode != X86::IMPLICIT_USE &&
387 Opcode != X86::IMPLICIT_DEF &&
388 Opcode != X86::FP_REG_KILL)
389 std::cerr << "X86 Machine Code Emitter: No 'form', not emitting: " << MI;
393 MCE.emitByte(BaseOpcode);
394 if (MI.getNumOperands() == 1) {
395 const MachineOperand &MO = MI.getOperand(0);
396 if (MO.isMachineBasicBlock()) {
397 emitPCRelativeBlockAddress(MO.getMachineBasicBlock());
398 } else if (MO.isGlobalAddress()) {
399 assert(MO.isPCRelative() && "Call target is not PC Relative?");
400 bool isTailCall = Opcode == X86::TAILJMPd ||
401 Opcode == X86::TAILJMPr || Opcode == X86::TAILJMPm;
402 emitGlobalAddressForCall(MO.getGlobal(), isTailCall);
403 } else if (MO.isExternalSymbol()) {
404 bool isTailCall = Opcode == X86::TAILJMPd ||
405 Opcode == X86::TAILJMPr || Opcode == X86::TAILJMPm;
406 emitExternalSymbolAddress(MO.getSymbolName(), true, isTailCall);
407 } else if (MO.isImmediate()) {
408 emitConstant(MO.getImmedValue(), sizeOfImm(Desc));
410 assert(0 && "Unknown RawFrm operand!");
415 case X86II::AddRegFrm:
416 MCE.emitByte(BaseOpcode + getX86RegNum(MI.getOperand(0).getReg()));
417 if (MI.getNumOperands() == 2) {
418 const MachineOperand &MO1 = MI.getOperand(1);
419 if (Value *V = MO1.getVRegValueOrNull()) {
420 assert(sizeOfImm(Desc) == 4 &&
421 "Don't know how to emit non-pointer values!");
422 emitGlobalAddressForPtr(cast<GlobalValue>(V));
423 } else if (MO1.isGlobalAddress()) {
424 assert(sizeOfImm(Desc) == 4 &&
425 "Don't know how to emit non-pointer values!");
426 assert(!MO1.isPCRelative() && "Function pointer ref is PC relative?");
427 emitGlobalAddressForPtr(MO1.getGlobal(), MO1.getOffset());
428 } else if (MO1.isExternalSymbol()) {
429 assert(sizeOfImm(Desc) == 4 &&
430 "Don't know how to emit non-pointer values!");
431 emitExternalSymbolAddress(MO1.getSymbolName(), false, false);
433 emitConstant(MO1.getImmedValue(), sizeOfImm(Desc));
438 case X86II::MRMDestReg: {
439 MCE.emitByte(BaseOpcode);
440 emitRegModRMByte(MI.getOperand(0).getReg(),
441 getX86RegNum(MI.getOperand(1).getReg()));
442 if (MI.getNumOperands() == 3)
443 emitConstant(MI.getOperand(2).getImmedValue(), sizeOfImm(Desc));
446 case X86II::MRMDestMem:
447 MCE.emitByte(BaseOpcode);
448 emitMemModRMByte(MI, 0, getX86RegNum(MI.getOperand(4).getReg()));
449 if (MI.getNumOperands() == 6)
450 emitConstant(MI.getOperand(5).getImmedValue(), sizeOfImm(Desc));
453 case X86II::MRMSrcReg:
454 MCE.emitByte(BaseOpcode);
456 emitRegModRMByte(MI.getOperand(1).getReg(),
457 getX86RegNum(MI.getOperand(0).getReg()));
458 if (MI.getNumOperands() == 3)
459 emitConstant(MI.getOperand(2).getImmedValue(), sizeOfImm(Desc));
462 case X86II::MRMSrcMem:
463 MCE.emitByte(BaseOpcode);
464 emitMemModRMByte(MI, 1, getX86RegNum(MI.getOperand(0).getReg()));
465 if (MI.getNumOperands() == 2+4)
466 emitConstant(MI.getOperand(5).getImmedValue(), sizeOfImm(Desc));
469 case X86II::MRM0r: case X86II::MRM1r:
470 case X86II::MRM2r: case X86II::MRM3r:
471 case X86II::MRM4r: case X86II::MRM5r:
472 case X86II::MRM6r: case X86II::MRM7r:
473 MCE.emitByte(BaseOpcode);
474 emitRegModRMByte(MI.getOperand(0).getReg(),
475 (Desc.TSFlags & X86II::FormMask)-X86II::MRM0r);
477 if (MI.getOperand(MI.getNumOperands()-1).isImmediate()) {
478 emitConstant(MI.getOperand(MI.getNumOperands()-1).getImmedValue(),
483 case X86II::MRM0m: case X86II::MRM1m:
484 case X86II::MRM2m: case X86II::MRM3m:
485 case X86II::MRM4m: case X86II::MRM5m:
486 case X86II::MRM6m: case X86II::MRM7m:
487 MCE.emitByte(BaseOpcode);
488 emitMemModRMByte(MI, 0, (Desc.TSFlags & X86II::FormMask)-X86II::MRM0m);
490 if (MI.getNumOperands() == 5) {
491 if (MI.getOperand(4).isImmediate())
492 emitConstant(MI.getOperand(4).getImmedValue(), sizeOfImm(Desc));
493 else if (MI.getOperand(4).isGlobalAddress())
494 emitGlobalAddressForPtr(MI.getOperand(4).getGlobal(),
495 MI.getOperand(4).getOffset());
497 assert(0 && "Unknown operand!");