1 //===-- X86MCInstLower.cpp - Convert X86 MachineInstr to an MCInst --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains code to lower X86 MachineInstrs to their corresponding
13 //===----------------------------------------------------------------------===//
15 #include "X86AsmPrinter.h"
16 #include "X86RegisterInfo.h"
17 #include "InstPrinter/X86ATTInstPrinter.h"
18 #include "MCTargetDesc/X86BaseInfo.h"
19 #include "Utils/X86ShuffleDecode.h"
20 #include "llvm/ADT/SmallString.h"
21 #include "llvm/CodeGen/MachineFunction.h"
22 #include "llvm/CodeGen/MachineConstantPool.h"
23 #include "llvm/CodeGen/MachineOperand.h"
24 #include "llvm/CodeGen/MachineModuleInfoImpls.h"
25 #include "llvm/CodeGen/StackMaps.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/GlobalValue.h"
28 #include "llvm/IR/Mangler.h"
29 #include "llvm/MC/MCAsmInfo.h"
30 #include "llvm/MC/MCCodeEmitter.h"
31 #include "llvm/MC/MCContext.h"
32 #include "llvm/MC/MCExpr.h"
33 #include "llvm/MC/MCInst.h"
34 #include "llvm/MC/MCInstBuilder.h"
35 #include "llvm/MC/MCStreamer.h"
36 #include "llvm/MC/MCSymbol.h"
37 #include "llvm/Support/TargetRegistry.h"
42 /// X86MCInstLower - This class is used to lower an MachineInstr into an MCInst.
43 class X86MCInstLower {
45 const MachineFunction &MF;
46 const TargetMachine &TM;
48 X86AsmPrinter &AsmPrinter;
50 X86MCInstLower(const MachineFunction &MF, X86AsmPrinter &asmprinter);
52 void Lower(const MachineInstr *MI, MCInst &OutMI) const;
54 MCSymbol *GetSymbolFromOperand(const MachineOperand &MO) const;
55 MCOperand LowerSymbolOperand(const MachineOperand &MO, MCSymbol *Sym) const;
58 MachineModuleInfoMachO &getMachOMMI() const;
59 Mangler *getMang() const {
60 return AsmPrinter.Mang;
64 } // end anonymous namespace
66 // Emit a minimal sequence of nops spanning NumBytes bytes.
67 static void EmitNops(MCStreamer &OS, unsigned NumBytes, bool Is64Bit,
68 const MCSubtargetInfo &STI);
71 X86AsmPrinter::StackMapShadowTracker::StackMapShadowTracker(TargetMachine &TM)
72 : TM(TM), InShadow(false), RequiredShadowSize(0), CurrentShadowSize(0) {}
74 X86AsmPrinter::StackMapShadowTracker::~StackMapShadowTracker() {}
77 X86AsmPrinter::StackMapShadowTracker::startFunction(MachineFunction &MF) {
78 CodeEmitter.reset(TM.getTarget().createMCCodeEmitter(
79 *TM.getSubtargetImpl()->getInstrInfo(),
80 *TM.getSubtargetImpl()->getRegisterInfo(), *TM.getSubtargetImpl(),
84 void X86AsmPrinter::StackMapShadowTracker::count(MCInst &Inst,
85 const MCSubtargetInfo &STI) {
87 SmallString<256> Code;
88 SmallVector<MCFixup, 4> Fixups;
89 raw_svector_ostream VecOS(Code);
90 CodeEmitter->EncodeInstruction(Inst, VecOS, Fixups, STI);
92 CurrentShadowSize += Code.size();
93 if (CurrentShadowSize >= RequiredShadowSize)
94 InShadow = false; // The shadow is big enough. Stop counting.
98 void X86AsmPrinter::StackMapShadowTracker::emitShadowPadding(
99 MCStreamer &OutStreamer, const MCSubtargetInfo &STI) {
100 if (InShadow && CurrentShadowSize < RequiredShadowSize) {
102 EmitNops(OutStreamer, RequiredShadowSize - CurrentShadowSize,
103 TM.getSubtarget<X86Subtarget>().is64Bit(), STI);
107 void X86AsmPrinter::EmitAndCountInstruction(MCInst &Inst) {
108 OutStreamer.EmitInstruction(Inst, getSubtargetInfo());
109 SMShadowTracker.count(Inst, getSubtargetInfo());
111 } // end llvm namespace
113 X86MCInstLower::X86MCInstLower(const MachineFunction &mf,
114 X86AsmPrinter &asmprinter)
115 : Ctx(mf.getContext()), MF(mf), TM(mf.getTarget()),
116 MAI(*TM.getMCAsmInfo()), AsmPrinter(asmprinter) {}
118 MachineModuleInfoMachO &X86MCInstLower::getMachOMMI() const {
119 return MF.getMMI().getObjFileInfo<MachineModuleInfoMachO>();
123 /// GetSymbolFromOperand - Lower an MO_GlobalAddress or MO_ExternalSymbol
124 /// operand to an MCSymbol.
125 MCSymbol *X86MCInstLower::
126 GetSymbolFromOperand(const MachineOperand &MO) const {
127 const DataLayout *DL = TM.getSubtargetImpl()->getDataLayout();
128 assert((MO.isGlobal() || MO.isSymbol() || MO.isMBB()) && "Isn't a symbol reference");
130 SmallString<128> Name;
133 switch (MO.getTargetFlags()) {
134 case X86II::MO_DLLIMPORT:
135 // Handle dllimport linkage.
138 case X86II::MO_DARWIN_STUB:
141 case X86II::MO_DARWIN_NONLAZY:
142 case X86II::MO_DARWIN_NONLAZY_PIC_BASE:
143 case X86II::MO_DARWIN_HIDDEN_NONLAZY_PIC_BASE:
144 Suffix = "$non_lazy_ptr";
149 Name += DL->getPrivateGlobalPrefix();
151 unsigned PrefixLen = Name.size();
154 const GlobalValue *GV = MO.getGlobal();
155 AsmPrinter.getNameWithPrefix(Name, GV);
156 } else if (MO.isSymbol()) {
157 getMang()->getNameWithPrefix(Name, MO.getSymbolName());
158 } else if (MO.isMBB()) {
159 Name += MO.getMBB()->getSymbol()->getName();
161 unsigned OrigLen = Name.size() - PrefixLen;
164 MCSymbol *Sym = Ctx.GetOrCreateSymbol(Name);
166 StringRef OrigName = StringRef(Name).substr(PrefixLen, OrigLen);
168 // If the target flags on the operand changes the name of the symbol, do that
169 // before we return the symbol.
170 switch (MO.getTargetFlags()) {
172 case X86II::MO_DARWIN_NONLAZY:
173 case X86II::MO_DARWIN_NONLAZY_PIC_BASE: {
174 MachineModuleInfoImpl::StubValueTy &StubSym =
175 getMachOMMI().getGVStubEntry(Sym);
176 if (!StubSym.getPointer()) {
177 assert(MO.isGlobal() && "Extern symbol not handled yet");
179 MachineModuleInfoImpl::
180 StubValueTy(AsmPrinter.getSymbol(MO.getGlobal()),
181 !MO.getGlobal()->hasInternalLinkage());
185 case X86II::MO_DARWIN_HIDDEN_NONLAZY_PIC_BASE: {
186 MachineModuleInfoImpl::StubValueTy &StubSym =
187 getMachOMMI().getHiddenGVStubEntry(Sym);
188 if (!StubSym.getPointer()) {
189 assert(MO.isGlobal() && "Extern symbol not handled yet");
191 MachineModuleInfoImpl::
192 StubValueTy(AsmPrinter.getSymbol(MO.getGlobal()),
193 !MO.getGlobal()->hasInternalLinkage());
197 case X86II::MO_DARWIN_STUB: {
198 MachineModuleInfoImpl::StubValueTy &StubSym =
199 getMachOMMI().getFnStubEntry(Sym);
200 if (StubSym.getPointer())
205 MachineModuleInfoImpl::
206 StubValueTy(AsmPrinter.getSymbol(MO.getGlobal()),
207 !MO.getGlobal()->hasInternalLinkage());
210 MachineModuleInfoImpl::
211 StubValueTy(Ctx.GetOrCreateSymbol(OrigName), false);
220 MCOperand X86MCInstLower::LowerSymbolOperand(const MachineOperand &MO,
221 MCSymbol *Sym) const {
222 // FIXME: We would like an efficient form for this, so we don't have to do a
223 // lot of extra uniquing.
224 const MCExpr *Expr = nullptr;
225 MCSymbolRefExpr::VariantKind RefKind = MCSymbolRefExpr::VK_None;
227 switch (MO.getTargetFlags()) {
228 default: llvm_unreachable("Unknown target flag on GV operand");
229 case X86II::MO_NO_FLAG: // No flag.
230 // These affect the name of the symbol, not any suffix.
231 case X86II::MO_DARWIN_NONLAZY:
232 case X86II::MO_DLLIMPORT:
233 case X86II::MO_DARWIN_STUB:
236 case X86II::MO_TLVP: RefKind = MCSymbolRefExpr::VK_TLVP; break;
237 case X86II::MO_TLVP_PIC_BASE:
238 Expr = MCSymbolRefExpr::Create(Sym, MCSymbolRefExpr::VK_TLVP, Ctx);
239 // Subtract the pic base.
240 Expr = MCBinaryExpr::CreateSub(Expr,
241 MCSymbolRefExpr::Create(MF.getPICBaseSymbol(),
245 case X86II::MO_SECREL: RefKind = MCSymbolRefExpr::VK_SECREL; break;
246 case X86II::MO_TLSGD: RefKind = MCSymbolRefExpr::VK_TLSGD; break;
247 case X86II::MO_TLSLD: RefKind = MCSymbolRefExpr::VK_TLSLD; break;
248 case X86II::MO_TLSLDM: RefKind = MCSymbolRefExpr::VK_TLSLDM; break;
249 case X86II::MO_GOTTPOFF: RefKind = MCSymbolRefExpr::VK_GOTTPOFF; break;
250 case X86II::MO_INDNTPOFF: RefKind = MCSymbolRefExpr::VK_INDNTPOFF; break;
251 case X86II::MO_TPOFF: RefKind = MCSymbolRefExpr::VK_TPOFF; break;
252 case X86II::MO_DTPOFF: RefKind = MCSymbolRefExpr::VK_DTPOFF; break;
253 case X86II::MO_NTPOFF: RefKind = MCSymbolRefExpr::VK_NTPOFF; break;
254 case X86II::MO_GOTNTPOFF: RefKind = MCSymbolRefExpr::VK_GOTNTPOFF; break;
255 case X86II::MO_GOTPCREL: RefKind = MCSymbolRefExpr::VK_GOTPCREL; break;
256 case X86II::MO_GOT: RefKind = MCSymbolRefExpr::VK_GOT; break;
257 case X86II::MO_GOTOFF: RefKind = MCSymbolRefExpr::VK_GOTOFF; break;
258 case X86II::MO_PLT: RefKind = MCSymbolRefExpr::VK_PLT; break;
259 case X86II::MO_PIC_BASE_OFFSET:
260 case X86II::MO_DARWIN_NONLAZY_PIC_BASE:
261 case X86II::MO_DARWIN_HIDDEN_NONLAZY_PIC_BASE:
262 Expr = MCSymbolRefExpr::Create(Sym, Ctx);
263 // Subtract the pic base.
264 Expr = MCBinaryExpr::CreateSub(Expr,
265 MCSymbolRefExpr::Create(MF.getPICBaseSymbol(), Ctx),
267 if (MO.isJTI() && MAI.hasSetDirective()) {
268 // If .set directive is supported, use it to reduce the number of
269 // relocations the assembler will generate for differences between
270 // local labels. This is only safe when the symbols are in the same
271 // section so we are restricting it to jumptable references.
272 MCSymbol *Label = Ctx.CreateTempSymbol();
273 AsmPrinter.OutStreamer.EmitAssignment(Label, Expr);
274 Expr = MCSymbolRefExpr::Create(Label, Ctx);
280 Expr = MCSymbolRefExpr::Create(Sym, RefKind, Ctx);
282 if (!MO.isJTI() && !MO.isMBB() && MO.getOffset())
283 Expr = MCBinaryExpr::CreateAdd(Expr,
284 MCConstantExpr::Create(MO.getOffset(), Ctx),
286 return MCOperand::CreateExpr(Expr);
290 /// \brief Simplify FOO $imm, %{al,ax,eax,rax} to FOO $imm, for instruction with
291 /// a short fixed-register form.
292 static void SimplifyShortImmForm(MCInst &Inst, unsigned Opcode) {
293 unsigned ImmOp = Inst.getNumOperands() - 1;
294 assert(Inst.getOperand(0).isReg() &&
295 (Inst.getOperand(ImmOp).isImm() || Inst.getOperand(ImmOp).isExpr()) &&
296 ((Inst.getNumOperands() == 3 && Inst.getOperand(1).isReg() &&
297 Inst.getOperand(0).getReg() == Inst.getOperand(1).getReg()) ||
298 Inst.getNumOperands() == 2) && "Unexpected instruction!");
300 // Check whether the destination register can be fixed.
301 unsigned Reg = Inst.getOperand(0).getReg();
302 if (Reg != X86::AL && Reg != X86::AX && Reg != X86::EAX && Reg != X86::RAX)
305 // If so, rewrite the instruction.
306 MCOperand Saved = Inst.getOperand(ImmOp);
308 Inst.setOpcode(Opcode);
309 Inst.addOperand(Saved);
312 /// \brief If a movsx instruction has a shorter encoding for the used register
313 /// simplify the instruction to use it instead.
314 static void SimplifyMOVSX(MCInst &Inst) {
315 unsigned NewOpcode = 0;
316 unsigned Op0 = Inst.getOperand(0).getReg(), Op1 = Inst.getOperand(1).getReg();
317 switch (Inst.getOpcode()) {
319 llvm_unreachable("Unexpected instruction!");
320 case X86::MOVSX16rr8: // movsbw %al, %ax --> cbtw
321 if (Op0 == X86::AX && Op1 == X86::AL)
322 NewOpcode = X86::CBW;
324 case X86::MOVSX32rr16: // movswl %ax, %eax --> cwtl
325 if (Op0 == X86::EAX && Op1 == X86::AX)
326 NewOpcode = X86::CWDE;
328 case X86::MOVSX64rr32: // movslq %eax, %rax --> cltq
329 if (Op0 == X86::RAX && Op1 == X86::EAX)
330 NewOpcode = X86::CDQE;
334 if (NewOpcode != 0) {
336 Inst.setOpcode(NewOpcode);
340 /// \brief Simplify things like MOV32rm to MOV32o32a.
341 static void SimplifyShortMoveForm(X86AsmPrinter &Printer, MCInst &Inst,
343 // Don't make these simplifications in 64-bit mode; other assemblers don't
344 // perform them because they make the code larger.
345 if (Printer.getSubtarget().is64Bit())
348 bool IsStore = Inst.getOperand(0).isReg() && Inst.getOperand(1).isReg();
349 unsigned AddrBase = IsStore;
350 unsigned RegOp = IsStore ? 0 : 5;
351 unsigned AddrOp = AddrBase + 3;
352 assert(Inst.getNumOperands() == 6 && Inst.getOperand(RegOp).isReg() &&
353 Inst.getOperand(AddrBase + X86::AddrBaseReg).isReg() &&
354 Inst.getOperand(AddrBase + X86::AddrScaleAmt).isImm() &&
355 Inst.getOperand(AddrBase + X86::AddrIndexReg).isReg() &&
356 Inst.getOperand(AddrBase + X86::AddrSegmentReg).isReg() &&
357 (Inst.getOperand(AddrOp).isExpr() ||
358 Inst.getOperand(AddrOp).isImm()) &&
359 "Unexpected instruction!");
361 // Check whether the destination register can be fixed.
362 unsigned Reg = Inst.getOperand(RegOp).getReg();
363 if (Reg != X86::AL && Reg != X86::AX && Reg != X86::EAX && Reg != X86::RAX)
366 // Check whether this is an absolute address.
367 // FIXME: We know TLVP symbol refs aren't, but there should be a better way
369 bool Absolute = true;
370 if (Inst.getOperand(AddrOp).isExpr()) {
371 const MCExpr *MCE = Inst.getOperand(AddrOp).getExpr();
372 if (const MCSymbolRefExpr *SRE = dyn_cast<MCSymbolRefExpr>(MCE))
373 if (SRE->getKind() == MCSymbolRefExpr::VK_TLVP)
378 (Inst.getOperand(AddrBase + X86::AddrBaseReg).getReg() != 0 ||
379 Inst.getOperand(AddrBase + X86::AddrScaleAmt).getImm() != 1 ||
380 Inst.getOperand(AddrBase + X86::AddrIndexReg).getReg() != 0))
383 // If so, rewrite the instruction.
384 MCOperand Saved = Inst.getOperand(AddrOp);
385 MCOperand Seg = Inst.getOperand(AddrBase + X86::AddrSegmentReg);
387 Inst.setOpcode(Opcode);
388 Inst.addOperand(Saved);
389 Inst.addOperand(Seg);
392 static unsigned getRetOpcode(const X86Subtarget &Subtarget)
394 return Subtarget.is64Bit() ? X86::RETQ : X86::RETL;
397 void X86MCInstLower::Lower(const MachineInstr *MI, MCInst &OutMI) const {
398 OutMI.setOpcode(MI->getOpcode());
400 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
401 const MachineOperand &MO = MI->getOperand(i);
404 switch (MO.getType()) {
407 llvm_unreachable("unknown operand type");
408 case MachineOperand::MO_Register:
409 // Ignore all implicit register operands.
410 if (MO.isImplicit()) continue;
411 MCOp = MCOperand::CreateReg(MO.getReg());
413 case MachineOperand::MO_Immediate:
414 MCOp = MCOperand::CreateImm(MO.getImm());
416 case MachineOperand::MO_MachineBasicBlock:
417 case MachineOperand::MO_GlobalAddress:
418 case MachineOperand::MO_ExternalSymbol:
419 MCOp = LowerSymbolOperand(MO, GetSymbolFromOperand(MO));
421 case MachineOperand::MO_JumpTableIndex:
422 MCOp = LowerSymbolOperand(MO, AsmPrinter.GetJTISymbol(MO.getIndex()));
424 case MachineOperand::MO_ConstantPoolIndex:
425 MCOp = LowerSymbolOperand(MO, AsmPrinter.GetCPISymbol(MO.getIndex()));
427 case MachineOperand::MO_BlockAddress:
428 MCOp = LowerSymbolOperand(MO,
429 AsmPrinter.GetBlockAddressSymbol(MO.getBlockAddress()));
431 case MachineOperand::MO_RegisterMask:
432 // Ignore call clobbers.
436 OutMI.addOperand(MCOp);
439 // Handle a few special cases to eliminate operand modifiers.
441 switch (OutMI.getOpcode()) {
446 // LEA should have a segment register, but it must be empty.
447 assert(OutMI.getNumOperands() == 1+X86::AddrNumOperands &&
448 "Unexpected # of LEA operands");
449 assert(OutMI.getOperand(1+X86::AddrSegmentReg).getReg() == 0 &&
450 "LEA has segment specified!");
454 OutMI.setOpcode(X86::MOV32ri);
457 // Commute operands to get a smaller encoding by using VEX.R instead of VEX.B
458 // if one of the registers is extended, but other isn't.
460 case X86::VMOVAPDYrr:
462 case X86::VMOVAPSYrr:
464 case X86::VMOVDQAYrr:
466 case X86::VMOVDQUYrr:
468 case X86::VMOVUPDYrr:
470 case X86::VMOVUPSYrr: {
471 if (!X86II::isX86_64ExtendedReg(OutMI.getOperand(0).getReg()) &&
472 X86II::isX86_64ExtendedReg(OutMI.getOperand(1).getReg())) {
474 switch (OutMI.getOpcode()) {
475 default: llvm_unreachable("Invalid opcode");
476 case X86::VMOVAPDrr: NewOpc = X86::VMOVAPDrr_REV; break;
477 case X86::VMOVAPDYrr: NewOpc = X86::VMOVAPDYrr_REV; break;
478 case X86::VMOVAPSrr: NewOpc = X86::VMOVAPSrr_REV; break;
479 case X86::VMOVAPSYrr: NewOpc = X86::VMOVAPSYrr_REV; break;
480 case X86::VMOVDQArr: NewOpc = X86::VMOVDQArr_REV; break;
481 case X86::VMOVDQAYrr: NewOpc = X86::VMOVDQAYrr_REV; break;
482 case X86::VMOVDQUrr: NewOpc = X86::VMOVDQUrr_REV; break;
483 case X86::VMOVDQUYrr: NewOpc = X86::VMOVDQUYrr_REV; break;
484 case X86::VMOVUPDrr: NewOpc = X86::VMOVUPDrr_REV; break;
485 case X86::VMOVUPDYrr: NewOpc = X86::VMOVUPDYrr_REV; break;
486 case X86::VMOVUPSrr: NewOpc = X86::VMOVUPSrr_REV; break;
487 case X86::VMOVUPSYrr: NewOpc = X86::VMOVUPSYrr_REV; break;
489 OutMI.setOpcode(NewOpc);
494 case X86::VMOVSSrr: {
495 if (!X86II::isX86_64ExtendedReg(OutMI.getOperand(0).getReg()) &&
496 X86II::isX86_64ExtendedReg(OutMI.getOperand(2).getReg())) {
498 switch (OutMI.getOpcode()) {
499 default: llvm_unreachable("Invalid opcode");
500 case X86::VMOVSDrr: NewOpc = X86::VMOVSDrr_REV; break;
501 case X86::VMOVSSrr: NewOpc = X86::VMOVSSrr_REV; break;
503 OutMI.setOpcode(NewOpc);
508 // TAILJMPr64, CALL64r, CALL64pcrel32 - These instructions have register
509 // inputs modeled as normal uses instead of implicit uses. As such, truncate
510 // off all but the first operand (the callee). FIXME: Change isel.
511 case X86::TAILJMPr64:
513 case X86::CALL64pcrel32: {
514 unsigned Opcode = OutMI.getOpcode();
515 MCOperand Saved = OutMI.getOperand(0);
517 OutMI.setOpcode(Opcode);
518 OutMI.addOperand(Saved);
523 case X86::EH_RETURN64: {
525 OutMI.setOpcode(getRetOpcode(AsmPrinter.getSubtarget()));
529 // TAILJMPd, TAILJMPd64 - Lower to the correct jump instructions.
532 case X86::TAILJMPd64: {
534 switch (OutMI.getOpcode()) {
535 default: llvm_unreachable("Invalid opcode");
536 case X86::TAILJMPr: Opcode = X86::JMP32r; break;
538 case X86::TAILJMPd64: Opcode = X86::JMP_1; break;
541 MCOperand Saved = OutMI.getOperand(0);
543 OutMI.setOpcode(Opcode);
544 OutMI.addOperand(Saved);
548 // These are pseudo-ops for OR to help with the OR->ADD transformation. We do
549 // this with an ugly goto in case the resultant OR uses EAX and needs the
551 case X86::ADD16rr_DB: OutMI.setOpcode(X86::OR16rr); goto ReSimplify;
552 case X86::ADD32rr_DB: OutMI.setOpcode(X86::OR32rr); goto ReSimplify;
553 case X86::ADD64rr_DB: OutMI.setOpcode(X86::OR64rr); goto ReSimplify;
554 case X86::ADD16ri_DB: OutMI.setOpcode(X86::OR16ri); goto ReSimplify;
555 case X86::ADD32ri_DB: OutMI.setOpcode(X86::OR32ri); goto ReSimplify;
556 case X86::ADD64ri32_DB: OutMI.setOpcode(X86::OR64ri32); goto ReSimplify;
557 case X86::ADD16ri8_DB: OutMI.setOpcode(X86::OR16ri8); goto ReSimplify;
558 case X86::ADD32ri8_DB: OutMI.setOpcode(X86::OR32ri8); goto ReSimplify;
559 case X86::ADD64ri8_DB: OutMI.setOpcode(X86::OR64ri8); goto ReSimplify;
561 // The assembler backend wants to see branches in their small form and relax
562 // them to their large form. The JIT can only handle the large form because
563 // it does not do relaxation. For now, translate the large form to the
565 case X86::JMP_4: OutMI.setOpcode(X86::JMP_1); break;
566 case X86::JO_4: OutMI.setOpcode(X86::JO_1); break;
567 case X86::JNO_4: OutMI.setOpcode(X86::JNO_1); break;
568 case X86::JB_4: OutMI.setOpcode(X86::JB_1); break;
569 case X86::JAE_4: OutMI.setOpcode(X86::JAE_1); break;
570 case X86::JE_4: OutMI.setOpcode(X86::JE_1); break;
571 case X86::JNE_4: OutMI.setOpcode(X86::JNE_1); break;
572 case X86::JBE_4: OutMI.setOpcode(X86::JBE_1); break;
573 case X86::JA_4: OutMI.setOpcode(X86::JA_1); break;
574 case X86::JS_4: OutMI.setOpcode(X86::JS_1); break;
575 case X86::JNS_4: OutMI.setOpcode(X86::JNS_1); break;
576 case X86::JP_4: OutMI.setOpcode(X86::JP_1); break;
577 case X86::JNP_4: OutMI.setOpcode(X86::JNP_1); break;
578 case X86::JL_4: OutMI.setOpcode(X86::JL_1); break;
579 case X86::JGE_4: OutMI.setOpcode(X86::JGE_1); break;
580 case X86::JLE_4: OutMI.setOpcode(X86::JLE_1); break;
581 case X86::JG_4: OutMI.setOpcode(X86::JG_1); break;
583 // Atomic load and store require a separate pseudo-inst because Acquire
584 // implies mayStore and Release implies mayLoad; fix these to regular MOV
586 case X86::ACQUIRE_MOV8rm: OutMI.setOpcode(X86::MOV8rm); goto ReSimplify;
587 case X86::ACQUIRE_MOV16rm: OutMI.setOpcode(X86::MOV16rm); goto ReSimplify;
588 case X86::ACQUIRE_MOV32rm: OutMI.setOpcode(X86::MOV32rm); goto ReSimplify;
589 case X86::ACQUIRE_MOV64rm: OutMI.setOpcode(X86::MOV64rm); goto ReSimplify;
590 case X86::RELEASE_MOV8mr: OutMI.setOpcode(X86::MOV8mr); goto ReSimplify;
591 case X86::RELEASE_MOV16mr: OutMI.setOpcode(X86::MOV16mr); goto ReSimplify;
592 case X86::RELEASE_MOV32mr: OutMI.setOpcode(X86::MOV32mr); goto ReSimplify;
593 case X86::RELEASE_MOV64mr: OutMI.setOpcode(X86::MOV64mr); goto ReSimplify;
595 // We don't currently select the correct instruction form for instructions
596 // which have a short %eax, etc. form. Handle this by custom lowering, for
599 // Note, we are currently not handling the following instructions:
600 // MOV64ao8, MOV64o8a
601 // XCHG16ar, XCHG32ar, XCHG64ar
602 case X86::MOV8mr_NOREX:
603 case X86::MOV8mr: SimplifyShortMoveForm(AsmPrinter, OutMI, X86::MOV8ao8); break;
604 case X86::MOV8rm_NOREX:
605 case X86::MOV8rm: SimplifyShortMoveForm(AsmPrinter, OutMI, X86::MOV8o8a); break;
606 case X86::MOV16mr: SimplifyShortMoveForm(AsmPrinter, OutMI, X86::MOV16ao16); break;
607 case X86::MOV16rm: SimplifyShortMoveForm(AsmPrinter, OutMI, X86::MOV16o16a); break;
608 case X86::MOV32mr: SimplifyShortMoveForm(AsmPrinter, OutMI, X86::MOV32ao32); break;
609 case X86::MOV32rm: SimplifyShortMoveForm(AsmPrinter, OutMI, X86::MOV32o32a); break;
611 case X86::ADC8ri: SimplifyShortImmForm(OutMI, X86::ADC8i8); break;
612 case X86::ADC16ri: SimplifyShortImmForm(OutMI, X86::ADC16i16); break;
613 case X86::ADC32ri: SimplifyShortImmForm(OutMI, X86::ADC32i32); break;
614 case X86::ADC64ri32: SimplifyShortImmForm(OutMI, X86::ADC64i32); break;
615 case X86::ADD8ri: SimplifyShortImmForm(OutMI, X86::ADD8i8); break;
616 case X86::ADD16ri: SimplifyShortImmForm(OutMI, X86::ADD16i16); break;
617 case X86::ADD32ri: SimplifyShortImmForm(OutMI, X86::ADD32i32); break;
618 case X86::ADD64ri32: SimplifyShortImmForm(OutMI, X86::ADD64i32); break;
619 case X86::AND8ri: SimplifyShortImmForm(OutMI, X86::AND8i8); break;
620 case X86::AND16ri: SimplifyShortImmForm(OutMI, X86::AND16i16); break;
621 case X86::AND32ri: SimplifyShortImmForm(OutMI, X86::AND32i32); break;
622 case X86::AND64ri32: SimplifyShortImmForm(OutMI, X86::AND64i32); break;
623 case X86::CMP8ri: SimplifyShortImmForm(OutMI, X86::CMP8i8); break;
624 case X86::CMP16ri: SimplifyShortImmForm(OutMI, X86::CMP16i16); break;
625 case X86::CMP32ri: SimplifyShortImmForm(OutMI, X86::CMP32i32); break;
626 case X86::CMP64ri32: SimplifyShortImmForm(OutMI, X86::CMP64i32); break;
627 case X86::OR8ri: SimplifyShortImmForm(OutMI, X86::OR8i8); break;
628 case X86::OR16ri: SimplifyShortImmForm(OutMI, X86::OR16i16); break;
629 case X86::OR32ri: SimplifyShortImmForm(OutMI, X86::OR32i32); break;
630 case X86::OR64ri32: SimplifyShortImmForm(OutMI, X86::OR64i32); break;
631 case X86::SBB8ri: SimplifyShortImmForm(OutMI, X86::SBB8i8); break;
632 case X86::SBB16ri: SimplifyShortImmForm(OutMI, X86::SBB16i16); break;
633 case X86::SBB32ri: SimplifyShortImmForm(OutMI, X86::SBB32i32); break;
634 case X86::SBB64ri32: SimplifyShortImmForm(OutMI, X86::SBB64i32); break;
635 case X86::SUB8ri: SimplifyShortImmForm(OutMI, X86::SUB8i8); break;
636 case X86::SUB16ri: SimplifyShortImmForm(OutMI, X86::SUB16i16); break;
637 case X86::SUB32ri: SimplifyShortImmForm(OutMI, X86::SUB32i32); break;
638 case X86::SUB64ri32: SimplifyShortImmForm(OutMI, X86::SUB64i32); break;
639 case X86::TEST8ri: SimplifyShortImmForm(OutMI, X86::TEST8i8); break;
640 case X86::TEST16ri: SimplifyShortImmForm(OutMI, X86::TEST16i16); break;
641 case X86::TEST32ri: SimplifyShortImmForm(OutMI, X86::TEST32i32); break;
642 case X86::TEST64ri32: SimplifyShortImmForm(OutMI, X86::TEST64i32); break;
643 case X86::XOR8ri: SimplifyShortImmForm(OutMI, X86::XOR8i8); break;
644 case X86::XOR16ri: SimplifyShortImmForm(OutMI, X86::XOR16i16); break;
645 case X86::XOR32ri: SimplifyShortImmForm(OutMI, X86::XOR32i32); break;
646 case X86::XOR64ri32: SimplifyShortImmForm(OutMI, X86::XOR64i32); break;
648 // Try to shrink some forms of movsx.
649 case X86::MOVSX16rr8:
650 case X86::MOVSX32rr16:
651 case X86::MOVSX64rr32:
652 SimplifyMOVSX(OutMI);
657 void X86AsmPrinter::LowerTlsAddr(X86MCInstLower &MCInstLowering,
658 const MachineInstr &MI) {
660 bool is64Bits = MI.getOpcode() == X86::TLS_addr64 ||
661 MI.getOpcode() == X86::TLS_base_addr64;
663 bool needsPadding = MI.getOpcode() == X86::TLS_addr64;
665 MCContext &context = OutStreamer.getContext();
668 EmitAndCountInstruction(MCInstBuilder(X86::DATA16_PREFIX));
670 MCSymbolRefExpr::VariantKind SRVK;
671 switch (MI.getOpcode()) {
672 case X86::TLS_addr32:
673 case X86::TLS_addr64:
674 SRVK = MCSymbolRefExpr::VK_TLSGD;
676 case X86::TLS_base_addr32:
677 SRVK = MCSymbolRefExpr::VK_TLSLDM;
679 case X86::TLS_base_addr64:
680 SRVK = MCSymbolRefExpr::VK_TLSLD;
683 llvm_unreachable("unexpected opcode");
686 MCSymbol *sym = MCInstLowering.GetSymbolFromOperand(MI.getOperand(3));
687 const MCSymbolRefExpr *symRef = MCSymbolRefExpr::Create(sym, SRVK, context);
691 LEA.setOpcode(X86::LEA64r);
692 LEA.addOperand(MCOperand::CreateReg(X86::RDI)); // dest
693 LEA.addOperand(MCOperand::CreateReg(X86::RIP)); // base
694 LEA.addOperand(MCOperand::CreateImm(1)); // scale
695 LEA.addOperand(MCOperand::CreateReg(0)); // index
696 LEA.addOperand(MCOperand::CreateExpr(symRef)); // disp
697 LEA.addOperand(MCOperand::CreateReg(0)); // seg
698 } else if (SRVK == MCSymbolRefExpr::VK_TLSLDM) {
699 LEA.setOpcode(X86::LEA32r);
700 LEA.addOperand(MCOperand::CreateReg(X86::EAX)); // dest
701 LEA.addOperand(MCOperand::CreateReg(X86::EBX)); // base
702 LEA.addOperand(MCOperand::CreateImm(1)); // scale
703 LEA.addOperand(MCOperand::CreateReg(0)); // index
704 LEA.addOperand(MCOperand::CreateExpr(symRef)); // disp
705 LEA.addOperand(MCOperand::CreateReg(0)); // seg
707 LEA.setOpcode(X86::LEA32r);
708 LEA.addOperand(MCOperand::CreateReg(X86::EAX)); // dest
709 LEA.addOperand(MCOperand::CreateReg(0)); // base
710 LEA.addOperand(MCOperand::CreateImm(1)); // scale
711 LEA.addOperand(MCOperand::CreateReg(X86::EBX)); // index
712 LEA.addOperand(MCOperand::CreateExpr(symRef)); // disp
713 LEA.addOperand(MCOperand::CreateReg(0)); // seg
715 EmitAndCountInstruction(LEA);
718 EmitAndCountInstruction(MCInstBuilder(X86::DATA16_PREFIX));
719 EmitAndCountInstruction(MCInstBuilder(X86::DATA16_PREFIX));
720 EmitAndCountInstruction(MCInstBuilder(X86::REX64_PREFIX));
723 StringRef name = is64Bits ? "__tls_get_addr" : "___tls_get_addr";
724 MCSymbol *tlsGetAddr = context.GetOrCreateSymbol(name);
725 const MCSymbolRefExpr *tlsRef =
726 MCSymbolRefExpr::Create(tlsGetAddr,
727 MCSymbolRefExpr::VK_PLT,
730 EmitAndCountInstruction(MCInstBuilder(is64Bits ? X86::CALL64pcrel32
735 /// \brief Emit the optimal amount of multi-byte nops on X86.
736 static void EmitNops(MCStreamer &OS, unsigned NumBytes, bool Is64Bit, const MCSubtargetInfo &STI) {
737 // This works only for 64bit. For 32bit we have to do additional checking if
738 // the CPU supports multi-byte nops.
739 assert(Is64Bit && "EmitNops only supports X86-64");
741 unsigned Opc, BaseReg, ScaleVal, IndexReg, Displacement, SegmentReg;
742 Opc = IndexReg = Displacement = SegmentReg = 0;
743 BaseReg = X86::RAX; ScaleVal = 1;
745 case 0: llvm_unreachable("Zero nops?"); break;
746 case 1: NumBytes -= 1; Opc = X86::NOOP; break;
747 case 2: NumBytes -= 2; Opc = X86::XCHG16ar; break;
748 case 3: NumBytes -= 3; Opc = X86::NOOPL; break;
749 case 4: NumBytes -= 4; Opc = X86::NOOPL; Displacement = 8; break;
750 case 5: NumBytes -= 5; Opc = X86::NOOPL; Displacement = 8;
751 IndexReg = X86::RAX; break;
752 case 6: NumBytes -= 6; Opc = X86::NOOPW; Displacement = 8;
753 IndexReg = X86::RAX; break;
754 case 7: NumBytes -= 7; Opc = X86::NOOPL; Displacement = 512; break;
755 case 8: NumBytes -= 8; Opc = X86::NOOPL; Displacement = 512;
756 IndexReg = X86::RAX; break;
757 case 9: NumBytes -= 9; Opc = X86::NOOPW; Displacement = 512;
758 IndexReg = X86::RAX; break;
759 default: NumBytes -= 10; Opc = X86::NOOPW; Displacement = 512;
760 IndexReg = X86::RAX; SegmentReg = X86::CS; break;
763 unsigned NumPrefixes = std::min(NumBytes, 5U);
764 NumBytes -= NumPrefixes;
765 for (unsigned i = 0; i != NumPrefixes; ++i)
766 OS.EmitBytes("\x66");
769 default: llvm_unreachable("Unexpected opcode"); break;
771 OS.EmitInstruction(MCInstBuilder(Opc), STI);
774 OS.EmitInstruction(MCInstBuilder(Opc).addReg(X86::AX), STI);
778 OS.EmitInstruction(MCInstBuilder(Opc).addReg(BaseReg)
779 .addImm(ScaleVal).addReg(IndexReg)
780 .addImm(Displacement).addReg(SegmentReg), STI);
783 } // while (NumBytes)
786 // Lower a stackmap of the form:
787 // <id>, <shadowBytes>, ...
788 void X86AsmPrinter::LowerSTACKMAP(const MachineInstr &MI) {
789 SMShadowTracker.emitShadowPadding(OutStreamer, getSubtargetInfo());
790 SM.recordStackMap(MI);
791 unsigned NumShadowBytes = MI.getOperand(1).getImm();
792 SMShadowTracker.reset(NumShadowBytes);
795 // Lower a patchpoint of the form:
796 // [<def>], <id>, <numBytes>, <target>, <numArgs>, <cc>, ...
797 void X86AsmPrinter::LowerPATCHPOINT(const MachineInstr &MI) {
798 assert(Subtarget->is64Bit() && "Patchpoint currently only supports X86-64");
800 SMShadowTracker.emitShadowPadding(OutStreamer, getSubtargetInfo());
802 SM.recordPatchPoint(MI);
804 PatchPointOpers opers(&MI);
805 unsigned ScratchIdx = opers.getNextScratchIdx();
806 unsigned EncodedBytes = 0;
807 int64_t CallTarget = opers.getMetaOper(PatchPointOpers::TargetPos).getImm();
809 // Emit MOV to materialize the target address and the CALL to target.
810 // This is encoded with 12-13 bytes, depending on which register is used.
811 unsigned ScratchReg = MI.getOperand(ScratchIdx).getReg();
812 if (X86II::isX86_64ExtendedReg(ScratchReg))
816 EmitAndCountInstruction(MCInstBuilder(X86::MOV64ri).addReg(ScratchReg)
817 .addImm(CallTarget));
818 EmitAndCountInstruction(MCInstBuilder(X86::CALL64r).addReg(ScratchReg));
821 unsigned NumBytes = opers.getMetaOper(PatchPointOpers::NBytesPos).getImm();
822 assert(NumBytes >= EncodedBytes &&
823 "Patchpoint can't request size less than the length of a call.");
825 EmitNops(OutStreamer, NumBytes - EncodedBytes, Subtarget->is64Bit(),
829 // Returns instruction preceding MBBI in MachineFunction.
830 // If MBBI is the first instruction of the first basic block, returns null.
831 static MachineBasicBlock::const_iterator
832 PrevCrossBBInst(MachineBasicBlock::const_iterator MBBI) {
833 const MachineBasicBlock *MBB = MBBI->getParent();
834 while (MBBI == MBB->begin()) {
835 if (MBB == MBB->getParent()->begin())
837 MBB = MBB->getPrevNode();
843 void X86AsmPrinter::EmitInstruction(const MachineInstr *MI) {
844 X86MCInstLower MCInstLowering(*MF, *this);
845 const X86RegisterInfo *RI = static_cast<const X86RegisterInfo *>(
846 TM.getSubtargetImpl()->getRegisterInfo());
848 switch (MI->getOpcode()) {
849 case TargetOpcode::DBG_VALUE:
850 llvm_unreachable("Should be handled target independently");
852 // Emit nothing here but a comment if we can.
853 case X86::Int_MemBarrier:
854 OutStreamer.emitRawComment("MEMBARRIER");
859 case X86::EH_RETURN64: {
860 // Lower these as normal, but add some comments.
861 unsigned Reg = MI->getOperand(0).getReg();
862 OutStreamer.AddComment(StringRef("eh_return, addr: %") +
863 X86ATTInstPrinter::getRegisterName(Reg));
868 case X86::TAILJMPd64:
869 // Lower these as normal, but add some comments.
870 OutStreamer.AddComment("TAILCALL");
873 case X86::TLS_addr32:
874 case X86::TLS_addr64:
875 case X86::TLS_base_addr32:
876 case X86::TLS_base_addr64:
877 return LowerTlsAddr(MCInstLowering, *MI);
879 case X86::MOVPC32r: {
880 // This is a pseudo op for a two instruction sequence with a label, which
887 MCSymbol *PICBase = MF->getPICBaseSymbol();
888 // FIXME: We would like an efficient form for this, so we don't have to do a
889 // lot of extra uniquing.
890 EmitAndCountInstruction(MCInstBuilder(X86::CALLpcrel32)
891 .addExpr(MCSymbolRefExpr::Create(PICBase, OutContext)));
894 OutStreamer.EmitLabel(PICBase);
897 EmitAndCountInstruction(MCInstBuilder(X86::POP32r)
898 .addReg(MI->getOperand(0).getReg()));
903 // Lower the MO_GOT_ABSOLUTE_ADDRESS form of ADD32ri.
904 if (MI->getOperand(2).getTargetFlags() != X86II::MO_GOT_ABSOLUTE_ADDRESS)
907 // Okay, we have something like:
908 // EAX = ADD32ri EAX, MO_GOT_ABSOLUTE_ADDRESS(@MYGLOBAL)
910 // For this, we want to print something like:
911 // MYGLOBAL + (. - PICBASE)
912 // However, we can't generate a ".", so just emit a new label here and refer
914 MCSymbol *DotSym = OutContext.CreateTempSymbol();
915 OutStreamer.EmitLabel(DotSym);
917 // Now that we have emitted the label, lower the complex operand expression.
918 MCSymbol *OpSym = MCInstLowering.GetSymbolFromOperand(MI->getOperand(2));
920 const MCExpr *DotExpr = MCSymbolRefExpr::Create(DotSym, OutContext);
921 const MCExpr *PICBase =
922 MCSymbolRefExpr::Create(MF->getPICBaseSymbol(), OutContext);
923 DotExpr = MCBinaryExpr::CreateSub(DotExpr, PICBase, OutContext);
925 DotExpr = MCBinaryExpr::CreateAdd(MCSymbolRefExpr::Create(OpSym,OutContext),
926 DotExpr, OutContext);
928 EmitAndCountInstruction(MCInstBuilder(X86::ADD32ri)
929 .addReg(MI->getOperand(0).getReg())
930 .addReg(MI->getOperand(1).getReg())
935 case TargetOpcode::STACKMAP:
936 return LowerSTACKMAP(*MI);
938 case TargetOpcode::PATCHPOINT:
939 return LowerPATCHPOINT(*MI);
941 case X86::MORESTACK_RET:
942 EmitAndCountInstruction(MCInstBuilder(getRetOpcode(*Subtarget)));
945 case X86::MORESTACK_RET_RESTORE_R10:
946 // Return, then restore R10.
947 EmitAndCountInstruction(MCInstBuilder(getRetOpcode(*Subtarget)));
948 EmitAndCountInstruction(MCInstBuilder(X86::MOV64rr)
953 case X86::SEH_PushReg:
954 OutStreamer.EmitWinCFIPushReg(RI->getSEHRegNum(MI->getOperand(0).getImm()));
957 case X86::SEH_SaveReg:
958 OutStreamer.EmitWinCFISaveReg(RI->getSEHRegNum(MI->getOperand(0).getImm()),
959 MI->getOperand(1).getImm());
962 case X86::SEH_SaveXMM:
963 OutStreamer.EmitWinCFISaveXMM(RI->getSEHRegNum(MI->getOperand(0).getImm()),
964 MI->getOperand(1).getImm());
967 case X86::SEH_StackAlloc:
968 OutStreamer.EmitWinCFIAllocStack(MI->getOperand(0).getImm());
971 case X86::SEH_SetFrame:
972 OutStreamer.EmitWinCFISetFrame(RI->getSEHRegNum(MI->getOperand(0).getImm()),
973 MI->getOperand(1).getImm());
976 case X86::SEH_PushFrame:
977 OutStreamer.EmitWinCFIPushFrame(MI->getOperand(0).getImm());
980 case X86::SEH_EndPrologue:
981 OutStreamer.EmitWinCFIEndProlog();
984 case X86::SEH_Epilogue: {
985 MachineBasicBlock::const_iterator MBBI(MI);
986 // Check if preceded by a call and emit nop if so.
987 for (MBBI = PrevCrossBBInst(MBBI); MBBI; MBBI = PrevCrossBBInst(MBBI)) {
988 // Conservatively assume that pseudo instructions don't emit code and keep
989 // looking for a call. We may emit an unnecessary nop in some cases.
990 if (!MBBI->isPseudo()) {
992 EmitAndCountInstruction(MCInstBuilder(X86::NOOP));
1000 case X86::VPSHUFBrm:
1001 // Lower PSHUFB normally but add a comment if we can find a constant
1002 // shuffle mask. We won't be able to do this at the MC layer because the
1003 // mask isn't an immediate.
1004 std::string Comment;
1005 raw_string_ostream CS(Comment);
1006 SmallVector<int, 16> Mask;
1008 assert(MI->getNumOperands() >= 6 &&
1009 "Wrong number of operands for PSHUFBrm or VPSHUFBrm");
1010 const MachineOperand &DstOp = MI->getOperand(0);
1011 const MachineOperand &SrcOp = MI->getOperand(1);
1012 const MachineOperand &MaskOp = MI->getOperand(5);
1014 // Compute the name for a register. This is really goofy because we have
1015 // multiple instruction printers that could (in theory) use different
1016 // names. Fortunately most people use the ATT style (outside of Windows)
1017 // and they actually agree on register naming here. Ultimately, this is
1018 // a comment, and so its OK if it isn't perfect.
1019 auto GetRegisterName = [](unsigned RegNum) -> StringRef {
1020 return X86ATTInstPrinter::getRegisterName(RegNum);
1023 StringRef DstName = DstOp.isReg() ? GetRegisterName(DstOp.getReg()) : "mem";
1024 StringRef SrcName = SrcOp.isReg() ? GetRegisterName(SrcOp.getReg()) : "mem";
1025 CS << DstName << " = ";
1027 if (MaskOp.isCPI()) {
1028 ArrayRef<MachineConstantPoolEntry> Constants =
1029 MI->getParent()->getParent()->getConstantPool()->getConstants();
1030 const MachineConstantPoolEntry &MaskConstantEntry =
1031 Constants[MaskOp.getIndex()];
1032 Type *MaskTy = MaskConstantEntry.getType();
1034 if (!MaskConstantEntry.isMachineConstantPoolEntry())
1035 if (auto *C = dyn_cast<ConstantDataSequential>(
1036 MaskConstantEntry.Val.ConstVal)) {
1037 assert(MaskTy == C->getType() &&
1038 "Expected a constant of the same type!");
1040 DecodePSHUFBMask(C, Mask);
1041 assert(Mask.size() == MaskTy->getVectorNumElements() &&
1042 "Shuffle mask has a different size than its type!");
1046 if (!Mask.empty()) {
1047 bool NeedComma = false;
1049 for (int M : Mask) {
1050 // Wrap up any prior entry...
1051 if (M == SM_SentinelZero && InSrc) {
1060 // Print this shuffle...
1061 if (M == SM_SentinelZero) {
1066 CS << SrcName << "[";
1074 OutStreamer.AddComment(CS.str());
1080 MCInstLowering.Lower(MI, TmpInst);
1081 EmitAndCountInstruction(TmpInst);