1 //===-- X86/X86CodeEmitter.cpp - Convert X86 code to machine code ---------===//
3 // This file contains the pass that transforms the X86 machine instructions into
4 // actual executable machine code.
6 //===----------------------------------------------------------------------===//
8 #include "X86TargetMachine.h"
10 #include "llvm/PassManager.h"
11 #include "llvm/CodeGen/MachineCodeEmitter.h"
12 #include "llvm/CodeGen/MachineFunctionPass.h"
13 #include "llvm/CodeGen/MachineInstr.h"
14 #include "llvm/Value.h"
15 #include "Support/Statistic.h"
20 NumEmitted("x86-emitter", "Number of machine instructions emitted");
23 MachineCodeEmitter &MCE;
25 // LazyCodeGenMap - Keep track of call sites for functions that are to be
27 std::map<unsigned, Function*> LazyCodeGenMap;
29 // LazyResolverMap - Keep track of the lazy resolver created for a
30 // particular function so that we can reuse them if necessary.
31 std::map<Function*, unsigned> LazyResolverMap;
33 JITResolver(MachineCodeEmitter &mce) : MCE(mce) {}
34 unsigned getLazyResolver(Function *F);
35 unsigned addFunctionReference(unsigned Address, Function *F);
38 unsigned emitStubForFunction(Function *F);
39 static void CompilationCallback();
40 unsigned resolveFunctionReference(unsigned RetAddr);
43 JITResolver *TheJITResolver;
47 /// addFunctionReference - This method is called when we need to emit the
48 /// address of a function that has not yet been emitted, so we don't know the
49 /// address. Instead, we emit a call to the CompilationCallback method, and
50 /// keep track of where we are.
52 unsigned JITResolver::addFunctionReference(unsigned Address, Function *F) {
53 LazyCodeGenMap[Address] = F;
54 return (intptr_t)&JITResolver::CompilationCallback;
57 unsigned JITResolver::resolveFunctionReference(unsigned RetAddr) {
58 std::map<unsigned, Function*>::iterator I = LazyCodeGenMap.find(RetAddr);
59 assert(I != LazyCodeGenMap.end() && "Not in map!");
60 Function *F = I->second;
61 LazyCodeGenMap.erase(I);
62 return MCE.forceCompilationOf(F);
65 unsigned JITResolver::getLazyResolver(Function *F) {
66 std::map<Function*, unsigned>::iterator I = LazyResolverMap.lower_bound(F);
67 if (I != LazyResolverMap.end() && I->first == F) return I->second;
69 //std::cerr << "Getting lazy resolver for : " << ((Value*)F)->getName() << "\n";
71 unsigned Stub = emitStubForFunction(F);
72 LazyResolverMap.insert(I, std::make_pair(F, Stub));
76 void JITResolver::CompilationCallback() {
77 unsigned *StackPtr = (unsigned*)__builtin_frame_address(0);
78 unsigned RetAddr = (unsigned)(intptr_t)__builtin_return_address(0);
79 assert(StackPtr[1] == RetAddr &&
80 "Could not find return address on the stack!");
82 // It's a stub if there is an interrupt marker after the call...
83 bool isStub = ((unsigned char*)(intptr_t)RetAddr)[0] == 0xCD;
85 // FIXME FIXME FIXME FIXME: __builtin_frame_address doesn't work if frame
86 // pointer elimination has been performed. Having a variable sized alloca
87 // disables frame pointer elimination currently, even if it's dead. This is a
90 // FIXME FIXME FIXME FIXME
92 // The call instruction should have pushed the return value onto the stack...
93 RetAddr -= 4; // Backtrack to the reference itself...
96 DEBUG(std::cerr << "In callback! Addr=0x" << std::hex << RetAddr
97 << " ESP=0x" << (unsigned)StackPtr << std::dec
98 << ": Resolving call to function: "
99 << TheVM->getFunctionReferencedName((void*)RetAddr) << "\n");
102 // Sanity check to make sure this really is a call instruction...
103 assert(((unsigned char*)(intptr_t)RetAddr)[-1] == 0xE8 &&"Not a call instr!");
105 unsigned NewVal = TheJITResolver->resolveFunctionReference(RetAddr);
107 // Rewrite the call target... so that we don't fault every time we execute
109 *(unsigned*)(intptr_t)RetAddr = NewVal-RetAddr-4;
112 // If this is a stub, rewrite the call into an unconditional branch
113 // instruction so that two return addresses are not pushed onto the stack
114 // when the requested function finally gets called. This also makes the
115 // 0xCD byte (interrupt) dead, so the marker doesn't effect anything.
116 ((unsigned char*)(intptr_t)RetAddr)[-1] = 0xE9;
119 // Change the return address to reexecute the call instruction...
123 /// emitStubForFunction - This method is used by the JIT when it needs to emit
124 /// the address of a function for a function whose code has not yet been
125 /// generated. In order to do this, it generates a stub which jumps to the lazy
126 /// function compiler, which will eventually get fixed to call the function
129 unsigned JITResolver::emitStubForFunction(Function *F) {
130 MCE.startFunctionStub(*F, 6);
131 MCE.emitByte(0xE8); // Call with 32 bit pc-rel destination...
133 unsigned Address = addFunctionReference(MCE.getCurrentPCValue(), F);
134 MCE.emitWord(Address-MCE.getCurrentPCValue()-4);
136 MCE.emitByte(0xCD); // Interrupt - Just a marker identifying the stub!
137 return (intptr_t)MCE.finishFunctionStub(*F);
143 class Emitter : public MachineFunctionPass {
144 const X86InstrInfo *II;
145 MachineCodeEmitter &MCE;
146 std::map<BasicBlock*, unsigned> BasicBlockAddrs;
147 std::vector<std::pair<BasicBlock*, unsigned> > BBRefs;
149 Emitter(MachineCodeEmitter &mce) : II(0), MCE(mce) {}
151 bool runOnMachineFunction(MachineFunction &MF);
153 virtual const char *getPassName() const {
154 return "X86 Machine Code Emitter";
158 void emitBasicBlock(MachineBasicBlock &MBB);
159 void emitInstruction(MachineInstr &MI);
161 void emitPCRelativeBlockAddress(BasicBlock *BB);
162 void emitMaybePCRelativeValue(unsigned Address, bool isPCRelative);
163 void emitGlobalAddressForCall(GlobalValue *GV);
164 void emitGlobalAddressForPtr(GlobalValue *GV);
166 void emitRegModRMByte(unsigned ModRMReg, unsigned RegOpcodeField);
167 void emitSIBByte(unsigned SS, unsigned Index, unsigned Base);
168 void emitConstant(unsigned Val, unsigned Size);
170 void emitMemModRMByte(const MachineInstr &MI,
171 unsigned Op, unsigned RegOpcodeField);
176 /// addPassesToEmitMachineCode - Add passes to the specified pass manager to get
177 /// machine code emitted. This uses a MAchineCodeEmitter object to handle
178 /// actually outputting the machine code and resolving things like the address
179 /// of functions. This method should returns true if machine code emission is
182 bool X86TargetMachine::addPassesToEmitMachineCode(PassManager &PM,
183 MachineCodeEmitter &MCE) {
184 PM.add(new Emitter(MCE));
188 bool Emitter::runOnMachineFunction(MachineFunction &MF) {
189 II = &((X86TargetMachine&)MF.getTarget()).getInstrInfo();
191 MCE.startFunction(MF);
192 MCE.emitConstantPool(MF.getConstantPool());
193 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I)
195 MCE.finishFunction(MF);
197 // Resolve all forward branches now...
198 for (unsigned i = 0, e = BBRefs.size(); i != e; ++i) {
199 unsigned Location = BasicBlockAddrs[BBRefs[i].first];
200 unsigned Ref = BBRefs[i].second;
201 *(unsigned*)(intptr_t)Ref = Location-Ref-4;
204 BasicBlockAddrs.clear();
208 void Emitter::emitBasicBlock(MachineBasicBlock &MBB) {
209 if (uint64_t Addr = MCE.getCurrentPCValue())
210 BasicBlockAddrs[MBB.getBasicBlock()] = Addr;
212 for (MachineBasicBlock::iterator I = MBB.begin(), E = MBB.end(); I != E; ++I)
213 emitInstruction(**I);
217 /// emitPCRelativeBlockAddress - This method emits the PC relative address of
218 /// the specified basic block, or if the basic block hasn't been emitted yet
219 /// (because this is a forward branch), it keeps track of the information
220 /// necessary to resolve this address later (and emits a dummy value).
222 void Emitter::emitPCRelativeBlockAddress(BasicBlock *BB) {
223 // FIXME: Emit backward branches directly
224 BBRefs.push_back(std::make_pair(BB, MCE.getCurrentPCValue()));
225 MCE.emitWord(0); // Emit a dummy value
228 /// emitMaybePCRelativeValue - Emit a 32-bit address which may be PC relative.
230 void Emitter::emitMaybePCRelativeValue(unsigned Address, bool isPCRelative) {
232 MCE.emitWord(Address-MCE.getCurrentPCValue()-4);
234 MCE.emitWord(Address);
237 /// emitGlobalAddressForCall - Emit the specified address to the code stream
238 /// assuming this is part of a function call, which is PC relative.
240 void Emitter::emitGlobalAddressForCall(GlobalValue *GV) {
241 // Get the address from the backend...
242 unsigned Address = MCE.getGlobalValueAddress(GV);
244 // If the machine code emitter doesn't know what the address IS yet, we have
245 // to take special measures.
248 // FIXME: this is JIT specific!
249 if (TheJITResolver == 0)
250 TheJITResolver = new JITResolver(MCE);
251 Address = TheJITResolver->addFunctionReference(MCE.getCurrentPCValue(),
254 emitMaybePCRelativeValue(Address, true);
257 /// emitGlobalAddress - Emit the specified address to the code stream assuming
258 /// this is part of a "take the address of a global" instruction, which is not
261 void Emitter::emitGlobalAddressForPtr(GlobalValue *GV) {
262 // Get the address from the backend...
263 unsigned Address = MCE.getGlobalValueAddress(GV);
265 // If the machine code emitter doesn't know what the address IS yet, we have
266 // to take special measures.
269 // FIXME: this is JIT specific!
270 if (TheJITResolver == 0)
271 TheJITResolver = new JITResolver(MCE);
272 Address = TheJITResolver->getLazyResolver((Function*)GV);
275 emitMaybePCRelativeValue(Address, false);
280 /// N86 namespace - Native X86 Register numbers... used by X86 backend.
284 EAX = 0, ECX = 1, EDX = 2, EBX = 3, ESP = 4, EBP = 5, ESI = 6, EDI = 7
289 // getX86RegNum - This function maps LLVM register identifiers to their X86
290 // specific numbering, which is used in various places encoding instructions.
292 static unsigned getX86RegNum(unsigned RegNo) {
294 case X86::EAX: case X86::AX: case X86::AL: return N86::EAX;
295 case X86::ECX: case X86::CX: case X86::CL: return N86::ECX;
296 case X86::EDX: case X86::DX: case X86::DL: return N86::EDX;
297 case X86::EBX: case X86::BX: case X86::BL: return N86::EBX;
298 case X86::ESP: case X86::SP: case X86::AH: return N86::ESP;
299 case X86::EBP: case X86::BP: case X86::CH: return N86::EBP;
300 case X86::ESI: case X86::SI: case X86::DH: return N86::ESI;
301 case X86::EDI: case X86::DI: case X86::BH: return N86::EDI;
303 case X86::ST0: case X86::ST1: case X86::ST2: case X86::ST3:
304 case X86::ST4: case X86::ST5: case X86::ST6: case X86::ST7:
305 return RegNo-X86::ST0;
307 assert(RegNo >= MRegisterInfo::FirstVirtualRegister &&
308 "Unknown physical register!");
309 assert(0 && "Register allocator hasn't allocated reg correctly yet!");
314 inline static unsigned char ModRMByte(unsigned Mod, unsigned RegOpcode,
316 assert(Mod < 4 && RegOpcode < 8 && RM < 8 && "ModRM Fields out of range!");
317 return RM | (RegOpcode << 3) | (Mod << 6);
320 void Emitter::emitRegModRMByte(unsigned ModRMReg, unsigned RegOpcodeFld){
321 MCE.emitByte(ModRMByte(3, RegOpcodeFld, getX86RegNum(ModRMReg)));
324 void Emitter::emitSIBByte(unsigned SS, unsigned Index, unsigned Base) {
325 // SIB byte is in the same format as the ModRMByte...
326 MCE.emitByte(ModRMByte(SS, Index, Base));
329 void Emitter::emitConstant(unsigned Val, unsigned Size) {
330 // Output the constant in little endian byte order...
331 for (unsigned i = 0; i != Size; ++i) {
332 MCE.emitByte(Val & 255);
337 static bool isDisp8(int Value) {
338 return Value == (signed char)Value;
341 void Emitter::emitMemModRMByte(const MachineInstr &MI,
342 unsigned Op, unsigned RegOpcodeField) {
343 const MachineOperand &Disp = MI.getOperand(Op+3);
344 if (MI.getOperand(Op).isConstantPoolIndex()) {
345 // Emit a direct address reference [disp32] where the displacement of the
346 // constant pool entry is controlled by the MCE.
347 MCE.emitByte(ModRMByte(0, RegOpcodeField, 5));
348 unsigned Index = MI.getOperand(Op).getConstantPoolIndex();
349 unsigned Address = MCE.getConstantPoolEntryAddress(Index);
350 MCE.emitWord(Address+Disp.getImmedValue());
354 const MachineOperand &BaseReg = MI.getOperand(Op);
355 const MachineOperand &Scale = MI.getOperand(Op+1);
356 const MachineOperand &IndexReg = MI.getOperand(Op+2);
358 // Is a SIB byte needed?
359 if (IndexReg.getReg() == 0 && BaseReg.getReg() != X86::ESP) {
360 if (BaseReg.getReg() == 0) { // Just a displacement?
361 // Emit special case [disp32] encoding
362 MCE.emitByte(ModRMByte(0, RegOpcodeField, 5));
363 emitConstant(Disp.getImmedValue(), 4);
365 unsigned BaseRegNo = getX86RegNum(BaseReg.getReg());
366 if (Disp.getImmedValue() == 0 && BaseRegNo != N86::EBP) {
367 // Emit simple indirect register encoding... [EAX] f.e.
368 MCE.emitByte(ModRMByte(0, RegOpcodeField, BaseRegNo));
369 } else if (isDisp8(Disp.getImmedValue())) {
370 // Emit the disp8 encoding... [REG+disp8]
371 MCE.emitByte(ModRMByte(1, RegOpcodeField, BaseRegNo));
372 emitConstant(Disp.getImmedValue(), 1);
374 // Emit the most general non-SIB encoding: [REG+disp32]
375 MCE.emitByte(ModRMByte(2, RegOpcodeField, BaseRegNo));
376 emitConstant(Disp.getImmedValue(), 4);
380 } else { // We need a SIB byte, so start by outputting the ModR/M byte first
381 assert(IndexReg.getReg() != X86::ESP && "Cannot use ESP as index reg!");
383 bool ForceDisp32 = false;
384 bool ForceDisp8 = false;
385 if (BaseReg.getReg() == 0) {
386 // If there is no base register, we emit the special case SIB byte with
387 // MOD=0, BASE=5, to JUST get the index, scale, and displacement.
388 MCE.emitByte(ModRMByte(0, RegOpcodeField, 4));
390 } else if (Disp.getImmedValue() == 0 && BaseReg.getReg() != X86::EBP) {
391 // Emit no displacement ModR/M byte
392 MCE.emitByte(ModRMByte(0, RegOpcodeField, 4));
393 } else if (isDisp8(Disp.getImmedValue())) {
394 // Emit the disp8 encoding...
395 MCE.emitByte(ModRMByte(1, RegOpcodeField, 4));
396 ForceDisp8 = true; // Make sure to force 8 bit disp if Base=EBP
398 // Emit the normal disp32 encoding...
399 MCE.emitByte(ModRMByte(2, RegOpcodeField, 4));
402 // Calculate what the SS field value should be...
403 static const unsigned SSTable[] = { ~0, 0, 1, ~0, 2, ~0, ~0, ~0, 3 };
404 unsigned SS = SSTable[Scale.getImmedValue()];
406 if (BaseReg.getReg() == 0) {
407 // Handle the SIB byte for the case where there is no base. The
408 // displacement has already been output.
409 assert(IndexReg.getReg() && "Index register must be specified!");
410 emitSIBByte(SS, getX86RegNum(IndexReg.getReg()), 5);
412 unsigned BaseRegNo = getX86RegNum(BaseReg.getReg());
414 if (IndexReg.getReg())
415 IndexRegNo = getX86RegNum(IndexReg.getReg());
417 IndexRegNo = 4; // For example [ESP+1*<noreg>+4]
418 emitSIBByte(SS, IndexRegNo, BaseRegNo);
421 // Do we need to output a displacement?
422 if (Disp.getImmedValue() != 0 || ForceDisp32 || ForceDisp8) {
423 if (!ForceDisp32 && isDisp8(Disp.getImmedValue()))
424 emitConstant(Disp.getImmedValue(), 1);
426 emitConstant(Disp.getImmedValue(), 4);
431 static unsigned sizeOfPtr(const TargetInstrDescriptor &Desc) {
432 switch (Desc.TSFlags & X86II::ArgMask) {
433 case X86II::Arg8: return 1;
434 case X86II::Arg16: return 2;
435 case X86II::Arg32: return 4;
436 case X86II::ArgF32: return 4;
437 case X86II::ArgF64: return 8;
438 case X86II::ArgF80: return 10;
439 default: assert(0 && "Memory size not set!");
444 void Emitter::emitInstruction(MachineInstr &MI) {
445 NumEmitted++; // Keep track of the # of mi's emitted
447 unsigned Opcode = MI.getOpcode();
448 const TargetInstrDescriptor &Desc = II->get(Opcode);
450 // Emit instruction prefixes if neccesary
451 if (Desc.TSFlags & X86II::OpSize) MCE.emitByte(0x66);// Operand size...
453 switch (Desc.TSFlags & X86II::Op0Mask) {
455 MCE.emitByte(0x0F); // Two-byte opcode prefix
457 case X86II::D8: case X86II::D9: case X86II::DA: case X86II::DB:
458 case X86II::DC: case X86II::DD: case X86II::DE: case X86II::DF:
460 (((Desc.TSFlags & X86II::Op0Mask)-X86II::D8)
461 >> X86II::Op0Shift));
462 break; // Two-byte opcode prefix
463 default: assert(0 && "Invalid prefix!");
464 case 0: break; // No prefix!
467 unsigned char BaseOpcode = II->getBaseOpcodeFor(Opcode);
468 switch (Desc.TSFlags & X86II::FormMask) {
469 default: assert(0 && "Unknown FormMask value in X86 MachineCodeEmitter!");
471 if (Opcode != X86::IMPLICIT_USE)
472 std::cerr << "X86 Machine Code Emitter: No 'form', not emitting: " << MI;
476 MCE.emitByte(BaseOpcode);
477 if (MI.getNumOperands() == 1) {
478 MachineOperand &MO = MI.getOperand(0);
479 if (MO.isPCRelativeDisp()) {
480 // Conditional branch... FIXME: this should use an MBB destination!
481 emitPCRelativeBlockAddress(cast<BasicBlock>(MO.getVRegValue()));
482 } else if (MO.isGlobalAddress()) {
483 assert(MO.isPCRelative() && "Call target is not PC Relative?");
484 emitGlobalAddressForCall(MO.getGlobal());
485 } else if (MO.isExternalSymbol()) {
486 unsigned Address = MCE.getGlobalValueAddress(MO.getSymbolName());
487 assert(Address && "Unknown external symbol!");
488 emitMaybePCRelativeValue(Address, MO.isPCRelative());
490 assert(0 && "Unknown RawFrm operand!");
495 case X86II::AddRegFrm:
496 MCE.emitByte(BaseOpcode + getX86RegNum(MI.getOperand(0).getReg()));
497 if (MI.getNumOperands() == 2) {
498 MachineOperand &MO1 = MI.getOperand(1);
499 if (MO1.isImmediate() || MO1.getVRegValueOrNull() ||
500 MO1.isGlobalAddress() || MO1.isExternalSymbol()) {
501 unsigned Size = sizeOfPtr(Desc);
502 if (Value *V = MO1.getVRegValueOrNull()) {
503 assert(Size == 4 && "Don't know how to emit non-pointer values!");
504 emitGlobalAddressForPtr(cast<GlobalValue>(V));
505 } else if (MO1.isGlobalAddress()) {
506 assert(Size == 4 && "Don't know how to emit non-pointer values!");
507 assert(!MO1.isPCRelative() && "Function pointer ref is PC relative?");
508 emitGlobalAddressForPtr(MO1.getGlobal());
509 } else if (MO1.isExternalSymbol()) {
510 assert(Size == 4 && "Don't know how to emit non-pointer values!");
512 unsigned Address = MCE.getGlobalValueAddress(MO1.getSymbolName());
513 assert(Address && "Unknown external symbol!");
514 emitMaybePCRelativeValue(Address, MO1.isPCRelative());
516 emitConstant(MO1.getImmedValue(), Size);
522 case X86II::MRMDestReg: {
523 MCE.emitByte(BaseOpcode);
524 MachineOperand &SrcOp = MI.getOperand(1+II->isTwoAddrInstr(Opcode));
525 emitRegModRMByte(MI.getOperand(0).getReg(), getX86RegNum(SrcOp.getReg()));
526 if (MI.getNumOperands() == 4)
527 emitConstant(MI.getOperand(3).getImmedValue(), sizeOfPtr(Desc));
530 case X86II::MRMDestMem:
531 MCE.emitByte(BaseOpcode);
532 emitMemModRMByte(MI, 0, getX86RegNum(MI.getOperand(4).getReg()));
535 case X86II::MRMSrcReg:
536 MCE.emitByte(BaseOpcode);
537 emitRegModRMByte(MI.getOperand(MI.getNumOperands()-1).getReg(),
538 getX86RegNum(MI.getOperand(0).getReg()));
541 case X86II::MRMSrcMem:
542 MCE.emitByte(BaseOpcode);
543 emitMemModRMByte(MI, MI.getNumOperands()-4,
544 getX86RegNum(MI.getOperand(0).getReg()));
547 case X86II::MRMS0r: case X86II::MRMS1r:
548 case X86II::MRMS2r: case X86II::MRMS3r:
549 case X86II::MRMS4r: case X86II::MRMS5r:
550 case X86II::MRMS6r: case X86II::MRMS7r:
551 MCE.emitByte(BaseOpcode);
552 emitRegModRMByte(MI.getOperand(0).getReg(),
553 (Desc.TSFlags & X86II::FormMask)-X86II::MRMS0r);
555 if (MI.getOperand(MI.getNumOperands()-1).isImmediate()) {
556 unsigned Size = sizeOfPtr(Desc);
557 emitConstant(MI.getOperand(MI.getNumOperands()-1).getImmedValue(), Size);
561 case X86II::MRMS0m: case X86II::MRMS1m:
562 case X86II::MRMS2m: case X86II::MRMS3m:
563 case X86II::MRMS4m: case X86II::MRMS5m:
564 case X86II::MRMS6m: case X86II::MRMS7m:
565 MCE.emitByte(BaseOpcode);
566 emitMemModRMByte(MI, 0, (Desc.TSFlags & X86II::FormMask)-X86II::MRMS0m);
568 if (MI.getNumOperands() == 5) {
569 unsigned Size = sizeOfPtr(Desc);
570 emitConstant(MI.getOperand(4).getImmedValue(), Size);