Added llvmgcc version to allow tests to be xfailed by frontend version.
[llvm-complete.git] / lib / Target / X86 / X86CodeEmitter.cpp
blob51be7b82dec8522e7c7e563f0111fa6ef461a995
1 //===-- X86/X86CodeEmitter.cpp - Convert X86 code to machine code ---------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
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.
7 //
8 //===----------------------------------------------------------------------===//
9 //
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"
17 #include "X86.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"
25 #include "llvm/Target/TargetOptions.h"
26 #include <iostream>
27 using namespace llvm;
29 namespace {
30 Statistic<>
31 NumEmitted("x86-emitter", "Number of machine instructions emitted");
34 namespace {
35 class Emitter : public MachineFunctionPass {
36 const X86InstrInfo *II;
37 MachineCodeEmitter &MCE;
38 std::map<const MachineBasicBlock*, unsigned> BasicBlockAddrs;
39 std::vector<std::pair<const MachineBasicBlock *, unsigned> > BBRefs;
40 public:
41 explicit Emitter(MachineCodeEmitter &mce) : II(0), MCE(mce) {}
42 Emitter(MachineCodeEmitter &mce, const X86InstrInfo& ii)
43 : II(&ii), MCE(mce) {}
45 bool runOnMachineFunction(MachineFunction &MF);
47 virtual const char *getPassName() const {
48 return "X86 Machine Code Emitter";
51 void emitInstruction(const MachineInstr &MI);
53 private:
54 void emitBasicBlock(const MachineBasicBlock &MBB);
56 void emitPCRelativeBlockAddress(const MachineBasicBlock *BB);
57 void emitPCRelativeValue(unsigned Address);
58 void emitGlobalAddressForCall(GlobalValue *GV, bool isTailCall);
59 void emitGlobalAddressForPtr(GlobalValue *GV, int Disp = 0);
60 void emitExternalSymbolAddress(const char *ES, bool isPCRelative,
61 bool isTailCall);
63 void emitRegModRMByte(unsigned ModRMReg, unsigned RegOpcodeField);
64 void emitSIBByte(unsigned SS, unsigned Index, unsigned Base);
65 void emitConstant(unsigned Val, unsigned Size);
67 void emitMemModRMByte(const MachineInstr &MI,
68 unsigned Op, unsigned RegOpcodeField);
73 /// createX86CodeEmitterPass - Return a pass that emits the collected X86 code
74 /// to the specified MCE object.
75 FunctionPass *llvm::createX86CodeEmitterPass(MachineCodeEmitter &MCE) {
76 return new Emitter(MCE);
79 bool Emitter::runOnMachineFunction(MachineFunction &MF) {
80 assert((MF.getTarget().getRelocationModel() != Reloc::Default ||
81 MF.getTarget().getRelocationModel() != Reloc::Static) &&
82 "JIT relocation model must be set to static or default!");
83 II = ((X86TargetMachine&)MF.getTarget()).getInstrInfo();
85 MCE.startFunction(MF);
86 MCE.emitConstantPool(MF.getConstantPool());
87 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I)
88 emitBasicBlock(*I);
89 MCE.finishFunction(MF);
91 // Resolve all forward branches now...
92 for (unsigned i = 0, e = BBRefs.size(); i != e; ++i) {
93 unsigned Location = BasicBlockAddrs[BBRefs[i].first];
94 unsigned Ref = BBRefs[i].second;
95 MCE.emitWordAt(Location-Ref-4, (unsigned*)(intptr_t)Ref);
97 BBRefs.clear();
98 BasicBlockAddrs.clear();
99 return false;
102 void Emitter::emitBasicBlock(const MachineBasicBlock &MBB) {
103 if (uint64_t Addr = MCE.getCurrentPCValue())
104 BasicBlockAddrs[&MBB] = Addr;
106 for (MachineBasicBlock::const_iterator I = MBB.begin(), E = MBB.end();
107 I != E; ++I)
108 emitInstruction(*I);
111 /// emitPCRelativeValue - Emit a 32-bit PC relative address.
113 void Emitter::emitPCRelativeValue(unsigned Address) {
114 MCE.emitWord(Address-MCE.getCurrentPCValue()-4);
117 /// emitPCRelativeBlockAddress - This method emits the PC relative address of
118 /// the specified basic block, or if the basic block hasn't been emitted yet
119 /// (because this is a forward branch), it keeps track of the information
120 /// necessary to resolve this address later (and emits a dummy value).
122 void Emitter::emitPCRelativeBlockAddress(const MachineBasicBlock *MBB) {
123 // If this is a backwards branch, we already know the address of the target,
124 // so just emit the value.
125 std::map<const MachineBasicBlock*, unsigned>::iterator I =
126 BasicBlockAddrs.find(MBB);
127 if (I != BasicBlockAddrs.end()) {
128 emitPCRelativeValue(I->second);
129 } else {
130 // Otherwise, remember where this reference was and where it is to so we can
131 // deal with it later.
132 BBRefs.push_back(std::make_pair(MBB, MCE.getCurrentPCValue()));
133 MCE.emitWord(0);
137 /// emitGlobalAddressForCall - Emit the specified address to the code stream
138 /// assuming this is part of a function call, which is PC relative.
140 void Emitter::emitGlobalAddressForCall(GlobalValue *GV, bool isTailCall) {
141 MCE.addRelocation(MachineRelocation(MCE.getCurrentPCOffset(),
142 X86::reloc_pcrel_word, GV, 0,
143 !isTailCall /*Doesn'tNeedStub*/));
144 MCE.emitWord(0);
147 /// emitGlobalAddress - Emit the specified address to the code stream assuming
148 /// this is part of a "take the address of a global" instruction, which is not
149 /// PC relative.
151 void Emitter::emitGlobalAddressForPtr(GlobalValue *GV, int Disp /* = 0 */) {
152 MCE.addRelocation(MachineRelocation(MCE.getCurrentPCOffset(),
153 X86::reloc_absolute_word, GV));
154 MCE.emitWord(Disp); // The relocated value will be added to the displacement
157 /// emitExternalSymbolAddress - Arrange for the address of an external symbol to
158 /// be emitted to the current location in the function, and allow it to be PC
159 /// relative.
160 void Emitter::emitExternalSymbolAddress(const char *ES, bool isPCRelative,
161 bool isTailCall) {
162 MCE.addRelocation(MachineRelocation(MCE.getCurrentPCOffset(),
163 isPCRelative ? X86::reloc_pcrel_word : X86::reloc_absolute_word, ES));
164 MCE.emitWord(0);
167 /// N86 namespace - Native X86 Register numbers... used by X86 backend.
169 namespace N86 {
170 enum {
171 EAX = 0, ECX = 1, EDX = 2, EBX = 3, ESP = 4, EBP = 5, ESI = 6, EDI = 7
176 // getX86RegNum - This function maps LLVM register identifiers to their X86
177 // specific numbering, which is used in various places encoding instructions.
179 static unsigned getX86RegNum(unsigned RegNo) {
180 switch(RegNo) {
181 case X86::EAX: case X86::AX: case X86::AL: return N86::EAX;
182 case X86::ECX: case X86::CX: case X86::CL: return N86::ECX;
183 case X86::EDX: case X86::DX: case X86::DL: return N86::EDX;
184 case X86::EBX: case X86::BX: case X86::BL: return N86::EBX;
185 case X86::ESP: case X86::SP: case X86::AH: return N86::ESP;
186 case X86::EBP: case X86::BP: case X86::CH: return N86::EBP;
187 case X86::ESI: case X86::SI: case X86::DH: return N86::ESI;
188 case X86::EDI: case X86::DI: case X86::BH: return N86::EDI;
190 case X86::ST0: case X86::ST1: case X86::ST2: case X86::ST3:
191 case X86::ST4: case X86::ST5: case X86::ST6: case X86::ST7:
192 return RegNo-X86::ST0;
194 case X86::XMM0: case X86::XMM1: case X86::XMM2: case X86::XMM3:
195 case X86::XMM4: case X86::XMM5: case X86::XMM6: case X86::XMM7:
196 return RegNo-X86::XMM0;
198 default:
199 assert(MRegisterInfo::isVirtualRegister(RegNo) &&
200 "Unknown physical register!");
201 assert(0 && "Register allocator hasn't allocated reg correctly yet!");
202 return 0;
206 inline static unsigned char ModRMByte(unsigned Mod, unsigned RegOpcode,
207 unsigned RM) {
208 assert(Mod < 4 && RegOpcode < 8 && RM < 8 && "ModRM Fields out of range!");
209 return RM | (RegOpcode << 3) | (Mod << 6);
212 void Emitter::emitRegModRMByte(unsigned ModRMReg, unsigned RegOpcodeFld){
213 MCE.emitByte(ModRMByte(3, RegOpcodeFld, getX86RegNum(ModRMReg)));
216 void Emitter::emitSIBByte(unsigned SS, unsigned Index, unsigned Base) {
217 // SIB byte is in the same format as the ModRMByte...
218 MCE.emitByte(ModRMByte(SS, Index, Base));
221 void Emitter::emitConstant(unsigned Val, unsigned Size) {
222 // Output the constant in little endian byte order...
223 for (unsigned i = 0; i != Size; ++i) {
224 MCE.emitByte(Val & 255);
225 Val >>= 8;
229 static bool isDisp8(int Value) {
230 return Value == (signed char)Value;
233 void Emitter::emitMemModRMByte(const MachineInstr &MI,
234 unsigned Op, unsigned RegOpcodeField) {
235 const MachineOperand &Op3 = MI.getOperand(Op+3);
236 GlobalValue *GV = 0;
237 int DispVal = 0;
239 if (Op3.isGlobalAddress()) {
240 GV = Op3.getGlobal();
241 DispVal = Op3.getOffset();
242 } else if (Op3.isConstantPoolIndex()) {
243 DispVal += MCE.getConstantPoolEntryAddress(Op3.getConstantPoolIndex());
244 DispVal += Op3.getOffset();
245 } else {
246 DispVal = Op3.getImmedValue();
249 const MachineOperand &Base = MI.getOperand(Op);
250 const MachineOperand &Scale = MI.getOperand(Op+1);
251 const MachineOperand &IndexReg = MI.getOperand(Op+2);
253 unsigned BaseReg = Base.getReg();
255 // Is a SIB byte needed?
256 if (IndexReg.getReg() == 0 && BaseReg != X86::ESP) {
257 if (BaseReg == 0) { // Just a displacement?
258 // Emit special case [disp32] encoding
259 MCE.emitByte(ModRMByte(0, RegOpcodeField, 5));
260 if (GV)
261 emitGlobalAddressForPtr(GV, DispVal);
262 else
263 emitConstant(DispVal, 4);
264 } else {
265 unsigned BaseRegNo = getX86RegNum(BaseReg);
266 if (GV) {
267 // Emit the most general non-SIB encoding: [REG+disp32]
268 MCE.emitByte(ModRMByte(2, RegOpcodeField, BaseRegNo));
269 emitGlobalAddressForPtr(GV, DispVal);
270 } else if (DispVal == 0 && BaseRegNo != N86::EBP) {
271 // Emit simple indirect register encoding... [EAX] f.e.
272 MCE.emitByte(ModRMByte(0, RegOpcodeField, BaseRegNo));
273 } else if (isDisp8(DispVal)) {
274 // Emit the disp8 encoding... [REG+disp8]
275 MCE.emitByte(ModRMByte(1, RegOpcodeField, BaseRegNo));
276 emitConstant(DispVal, 1);
277 } else {
278 // Emit the most general non-SIB encoding: [REG+disp32]
279 MCE.emitByte(ModRMByte(2, RegOpcodeField, BaseRegNo));
280 emitConstant(DispVal, 4);
284 } else { // We need a SIB byte, so start by outputting the ModR/M byte first
285 assert(IndexReg.getReg() != X86::ESP && "Cannot use ESP as index reg!");
287 bool ForceDisp32 = false;
288 bool ForceDisp8 = false;
289 if (BaseReg == 0) {
290 // If there is no base register, we emit the special case SIB byte with
291 // MOD=0, BASE=5, to JUST get the index, scale, and displacement.
292 MCE.emitByte(ModRMByte(0, RegOpcodeField, 4));
293 ForceDisp32 = true;
294 } else if (GV) {
295 // Emit the normal disp32 encoding...
296 MCE.emitByte(ModRMByte(2, RegOpcodeField, 4));
297 ForceDisp32 = true;
298 } else if (DispVal == 0 && BaseReg != X86::EBP) {
299 // Emit no displacement ModR/M byte
300 MCE.emitByte(ModRMByte(0, RegOpcodeField, 4));
301 } else if (isDisp8(DispVal)) {
302 // Emit the disp8 encoding...
303 MCE.emitByte(ModRMByte(1, RegOpcodeField, 4));
304 ForceDisp8 = true; // Make sure to force 8 bit disp if Base=EBP
305 } else {
306 // Emit the normal disp32 encoding...
307 MCE.emitByte(ModRMByte(2, RegOpcodeField, 4));
310 // Calculate what the SS field value should be...
311 static const unsigned SSTable[] = { ~0, 0, 1, ~0, 2, ~0, ~0, ~0, 3 };
312 unsigned SS = SSTable[Scale.getImmedValue()];
314 if (BaseReg == 0) {
315 // Handle the SIB byte for the case where there is no base. The
316 // displacement has already been output.
317 assert(IndexReg.getReg() && "Index register must be specified!");
318 emitSIBByte(SS, getX86RegNum(IndexReg.getReg()), 5);
319 } else {
320 unsigned BaseRegNo = getX86RegNum(BaseReg);
321 unsigned IndexRegNo;
322 if (IndexReg.getReg())
323 IndexRegNo = getX86RegNum(IndexReg.getReg());
324 else
325 IndexRegNo = 4; // For example [ESP+1*<noreg>+4]
326 emitSIBByte(SS, IndexRegNo, BaseRegNo);
329 // Do we need to output a displacement?
330 if (DispVal != 0 || ForceDisp32 || ForceDisp8) {
331 if (!ForceDisp32 && isDisp8(DispVal))
332 emitConstant(DispVal, 1);
333 else if (GV)
334 emitGlobalAddressForPtr(GV, DispVal);
335 else
336 emitConstant(DispVal, 4);
341 static unsigned sizeOfImm(const TargetInstrDescriptor &Desc) {
342 switch (Desc.TSFlags & X86II::ImmMask) {
343 case X86II::Imm8: return 1;
344 case X86II::Imm16: return 2;
345 case X86II::Imm32: return 4;
346 default: assert(0 && "Immediate size not set!");
347 return 0;
351 void Emitter::emitInstruction(const MachineInstr &MI) {
352 NumEmitted++; // Keep track of the # of mi's emitted
354 unsigned Opcode = MI.getOpcode();
355 const TargetInstrDescriptor &Desc = II->get(Opcode);
357 // Emit the repeat opcode prefix as needed.
358 if ((Desc.TSFlags & X86II::Op0Mask) == X86II::REP) MCE.emitByte(0xF3);
360 // Emit the operand size opcode prefix as needed.
361 if (Desc.TSFlags & X86II::OpSize) MCE.emitByte(0x66);
363 switch (Desc.TSFlags & X86II::Op0Mask) {
364 case X86II::TB:
365 MCE.emitByte(0x0F); // Two-byte opcode prefix
366 break;
367 case X86II::REP: break; // already handled.
368 case X86II::XS: // F3 0F
369 MCE.emitByte(0xF3);
370 MCE.emitByte(0x0F);
371 break;
372 case X86II::XD: // F2 0F
373 MCE.emitByte(0xF2);
374 MCE.emitByte(0x0F);
375 break;
376 case X86II::D8: case X86II::D9: case X86II::DA: case X86II::DB:
377 case X86II::DC: case X86II::DD: case X86II::DE: case X86II::DF:
378 MCE.emitByte(0xD8+
379 (((Desc.TSFlags & X86II::Op0Mask)-X86II::D8)
380 >> X86II::Op0Shift));
381 break; // Two-byte opcode prefix
382 default: assert(0 && "Invalid prefix!");
383 case 0: break; // No prefix!
386 unsigned char BaseOpcode = II->getBaseOpcodeFor(Opcode);
387 switch (Desc.TSFlags & X86II::FormMask) {
388 default: assert(0 && "Unknown FormMask value in X86 MachineCodeEmitter!");
389 case X86II::Pseudo:
390 #ifndef NDEBUG
391 switch (Opcode) {
392 default:
393 assert(0 && "psuedo instructions should be removed before code emission");
394 case X86::IMPLICIT_USE:
395 case X86::IMPLICIT_DEF:
396 case X86::IMPLICIT_DEF_R8:
397 case X86::IMPLICIT_DEF_R16:
398 case X86::IMPLICIT_DEF_R32:
399 case X86::IMPLICIT_DEF_FR32:
400 case X86::IMPLICIT_DEF_FR64:
401 case X86::IMPLICIT_DEF_VR64:
402 case X86::IMPLICIT_DEF_VR128:
403 case X86::FP_REG_KILL:
404 break;
406 #endif
407 break;
409 case X86II::RawFrm:
410 MCE.emitByte(BaseOpcode);
411 if (MI.getNumOperands() == 1) {
412 const MachineOperand &MO = MI.getOperand(0);
413 if (MO.isMachineBasicBlock()) {
414 emitPCRelativeBlockAddress(MO.getMachineBasicBlock());
415 } else if (MO.isGlobalAddress()) {
416 bool isTailCall = Opcode == X86::TAILJMPd ||
417 Opcode == X86::TAILJMPr || Opcode == X86::TAILJMPm;
418 emitGlobalAddressForCall(MO.getGlobal(), isTailCall);
419 } else if (MO.isExternalSymbol()) {
420 bool isTailCall = Opcode == X86::TAILJMPd ||
421 Opcode == X86::TAILJMPr || Opcode == X86::TAILJMPm;
422 emitExternalSymbolAddress(MO.getSymbolName(), true, isTailCall);
423 } else if (MO.isImmediate()) {
424 emitConstant(MO.getImmedValue(), sizeOfImm(Desc));
425 } else {
426 assert(0 && "Unknown RawFrm operand!");
429 break;
431 case X86II::AddRegFrm:
432 MCE.emitByte(BaseOpcode + getX86RegNum(MI.getOperand(0).getReg()));
433 if (MI.getNumOperands() == 2) {
434 const MachineOperand &MO1 = MI.getOperand(1);
435 if (Value *V = MO1.getVRegValueOrNull()) {
436 assert(sizeOfImm(Desc) == 4 &&
437 "Don't know how to emit non-pointer values!");
438 emitGlobalAddressForPtr(cast<GlobalValue>(V));
439 } else if (MO1.isGlobalAddress()) {
440 assert(sizeOfImm(Desc) == 4 &&
441 "Don't know how to emit non-pointer values!");
442 assert(!MO1.isPCRelative() && "Function pointer ref is PC relative?");
443 emitGlobalAddressForPtr(MO1.getGlobal(), MO1.getOffset());
444 } else if (MO1.isExternalSymbol()) {
445 assert(sizeOfImm(Desc) == 4 &&
446 "Don't know how to emit non-pointer values!");
447 emitExternalSymbolAddress(MO1.getSymbolName(), false, false);
448 } else {
449 emitConstant(MO1.getImmedValue(), sizeOfImm(Desc));
452 break;
454 case X86II::MRMDestReg: {
455 MCE.emitByte(BaseOpcode);
456 emitRegModRMByte(MI.getOperand(0).getReg(),
457 getX86RegNum(MI.getOperand(1).getReg()));
458 if (MI.getNumOperands() == 3)
459 emitConstant(MI.getOperand(2).getImmedValue(), sizeOfImm(Desc));
460 break;
462 case X86II::MRMDestMem:
463 MCE.emitByte(BaseOpcode);
464 emitMemModRMByte(MI, 0, getX86RegNum(MI.getOperand(4).getReg()));
465 if (MI.getNumOperands() == 6)
466 emitConstant(MI.getOperand(5).getImmedValue(), sizeOfImm(Desc));
467 break;
469 case X86II::MRMSrcReg:
470 MCE.emitByte(BaseOpcode);
471 emitRegModRMByte(MI.getOperand(1).getReg(),
472 getX86RegNum(MI.getOperand(0).getReg()));
473 if (MI.getNumOperands() == 3)
474 emitConstant(MI.getOperand(2).getImmedValue(), sizeOfImm(Desc));
475 break;
477 case X86II::MRMSrcMem:
478 MCE.emitByte(BaseOpcode);
479 emitMemModRMByte(MI, 1, getX86RegNum(MI.getOperand(0).getReg()));
480 if (MI.getNumOperands() == 2+4)
481 emitConstant(MI.getOperand(5).getImmedValue(), sizeOfImm(Desc));
482 break;
484 case X86II::MRM0r: case X86II::MRM1r:
485 case X86II::MRM2r: case X86II::MRM3r:
486 case X86II::MRM4r: case X86II::MRM5r:
487 case X86II::MRM6r: case X86II::MRM7r:
488 MCE.emitByte(BaseOpcode);
489 emitRegModRMByte(MI.getOperand(0).getReg(),
490 (Desc.TSFlags & X86II::FormMask)-X86II::MRM0r);
492 if (MI.getOperand(MI.getNumOperands()-1).isImmediate()) {
493 emitConstant(MI.getOperand(MI.getNumOperands()-1).getImmedValue(),
494 sizeOfImm(Desc));
496 break;
498 case X86II::MRM0m: case X86II::MRM1m:
499 case X86II::MRM2m: case X86II::MRM3m:
500 case X86II::MRM4m: case X86II::MRM5m:
501 case X86II::MRM6m: case X86II::MRM7m:
502 MCE.emitByte(BaseOpcode);
503 emitMemModRMByte(MI, 0, (Desc.TSFlags & X86II::FormMask)-X86II::MRM0m);
505 if (MI.getNumOperands() == 5) {
506 if (MI.getOperand(4).isImmediate())
507 emitConstant(MI.getOperand(4).getImmedValue(), sizeOfImm(Desc));
508 else if (MI.getOperand(4).isGlobalAddress())
509 emitGlobalAddressForPtr(MI.getOperand(4).getGlobal(),
510 MI.getOperand(4).getOffset());
511 else
512 assert(0 && "Unknown operand!");
514 break;
516 case X86II::MRMInitReg:
517 MCE.emitByte(BaseOpcode);
518 emitRegModRMByte(MI.getOperand(0).getReg(),
519 getX86RegNum(MI.getOperand(0).getReg()));
520 break;