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 "X86ShuffleDecodeConstantPool.h"
18 #include "InstPrinter/X86ATTInstPrinter.h"
19 #include "InstPrinter/X86InstComments.h"
20 #include "MCTargetDesc/X86BaseInfo.h"
21 #include "Utils/X86ShuffleDecode.h"
22 #include "llvm/ADT/Optional.h"
23 #include "llvm/ADT/SmallString.h"
24 #include "llvm/ADT/iterator_range.h"
25 #include "llvm/CodeGen/MachineFunction.h"
26 #include "llvm/CodeGen/MachineConstantPool.h"
27 #include "llvm/CodeGen/MachineOperand.h"
28 #include "llvm/CodeGen/MachineModuleInfoImpls.h"
29 #include "llvm/CodeGen/StackMaps.h"
30 #include "llvm/IR/DataLayout.h"
31 #include "llvm/IR/GlobalValue.h"
32 #include "llvm/IR/Mangler.h"
33 #include "llvm/MC/MCAsmInfo.h"
34 #include "llvm/MC/MCCodeEmitter.h"
35 #include "llvm/MC/MCContext.h"
36 #include "llvm/MC/MCExpr.h"
37 #include "llvm/MC/MCFixup.h"
38 #include "llvm/MC/MCInst.h"
39 #include "llvm/MC/MCInstBuilder.h"
40 #include "llvm/MC/MCSection.h"
41 #include "llvm/MC/MCStreamer.h"
42 #include "llvm/MC/MCSymbol.h"
43 #include "llvm/MC/MCSymbolELF.h"
44 #include "llvm/MC/MCSectionELF.h"
45 #include "llvm/MC/MCSectionMachO.h"
46 #include "llvm/Support/TargetRegistry.h"
47 #include "llvm/Support/ELF.h"
48 #include "llvm/Target/TargetLoweringObjectFile.h"
54 /// X86MCInstLower - This class is used to lower an MachineInstr into an MCInst.
55 class X86MCInstLower {
57 const MachineFunction &MF;
58 const TargetMachine &TM;
60 X86AsmPrinter &AsmPrinter;
62 X86MCInstLower(const MachineFunction &MF, X86AsmPrinter &asmprinter);
64 Optional<MCOperand> LowerMachineOperand(const MachineInstr *MI,
65 const MachineOperand &MO) const;
66 void Lower(const MachineInstr *MI, MCInst &OutMI) const;
68 MCSymbol *GetSymbolFromOperand(const MachineOperand &MO) const;
69 MCOperand LowerSymbolOperand(const MachineOperand &MO, MCSymbol *Sym) const;
72 MachineModuleInfoMachO &getMachOMMI() const;
75 } // end anonymous namespace
77 // Emit a minimal sequence of nops spanning NumBytes bytes.
78 static void EmitNops(MCStreamer &OS, unsigned NumBytes, bool Is64Bit,
79 const MCSubtargetInfo &STI);
81 void X86AsmPrinter::StackMapShadowTracker::count(MCInst &Inst,
82 const MCSubtargetInfo &STI,
83 MCCodeEmitter *CodeEmitter) {
85 SmallString<256> Code;
86 SmallVector<MCFixup, 4> Fixups;
87 raw_svector_ostream VecOS(Code);
88 CodeEmitter->encodeInstruction(Inst, VecOS, Fixups, STI);
89 CurrentShadowSize += Code.size();
90 if (CurrentShadowSize >= RequiredShadowSize)
91 InShadow = false; // The shadow is big enough. Stop counting.
95 void X86AsmPrinter::StackMapShadowTracker::emitShadowPadding(
96 MCStreamer &OutStreamer, const MCSubtargetInfo &STI) {
97 if (InShadow && CurrentShadowSize < RequiredShadowSize) {
99 EmitNops(OutStreamer, RequiredShadowSize - CurrentShadowSize,
100 MF->getSubtarget<X86Subtarget>().is64Bit(), STI);
104 void X86AsmPrinter::EmitAndCountInstruction(MCInst &Inst) {
105 OutStreamer->EmitInstruction(Inst, getSubtargetInfo());
106 SMShadowTracker.count(Inst, getSubtargetInfo(), CodeEmitter.get());
109 X86MCInstLower::X86MCInstLower(const MachineFunction &mf,
110 X86AsmPrinter &asmprinter)
111 : Ctx(mf.getContext()), MF(mf), TM(mf.getTarget()), MAI(*TM.getMCAsmInfo()),
112 AsmPrinter(asmprinter) {}
114 MachineModuleInfoMachO &X86MCInstLower::getMachOMMI() const {
115 return MF.getMMI().getObjFileInfo<MachineModuleInfoMachO>();
119 /// GetSymbolFromOperand - Lower an MO_GlobalAddress or MO_ExternalSymbol
120 /// operand to an MCSymbol.
121 MCSymbol *X86MCInstLower::
122 GetSymbolFromOperand(const MachineOperand &MO) const {
123 const DataLayout &DL = MF.getDataLayout();
124 assert((MO.isGlobal() || MO.isSymbol() || MO.isMBB()) && "Isn't a symbol reference");
126 MCSymbol *Sym = nullptr;
127 SmallString<128> Name;
130 switch (MO.getTargetFlags()) {
131 case X86II::MO_DLLIMPORT:
132 // Handle dllimport linkage.
135 case X86II::MO_DARWIN_NONLAZY:
136 case X86II::MO_DARWIN_NONLAZY_PIC_BASE:
137 Suffix = "$non_lazy_ptr";
142 Name += DL.getPrivateGlobalPrefix();
145 const GlobalValue *GV = MO.getGlobal();
146 AsmPrinter.getNameWithPrefix(Name, GV);
147 } else if (MO.isSymbol()) {
148 Mangler::getNameWithPrefix(Name, MO.getSymbolName(), DL);
149 } else if (MO.isMBB()) {
150 assert(Suffix.empty());
151 Sym = MO.getMBB()->getSymbol();
156 Sym = Ctx.getOrCreateSymbol(Name);
158 // If the target flags on the operand changes the name of the symbol, do that
159 // before we return the symbol.
160 switch (MO.getTargetFlags()) {
162 case X86II::MO_DARWIN_NONLAZY:
163 case X86II::MO_DARWIN_NONLAZY_PIC_BASE: {
164 MachineModuleInfoImpl::StubValueTy &StubSym =
165 getMachOMMI().getGVStubEntry(Sym);
166 if (!StubSym.getPointer()) {
167 assert(MO.isGlobal() && "Extern symbol not handled yet");
169 MachineModuleInfoImpl::
170 StubValueTy(AsmPrinter.getSymbol(MO.getGlobal()),
171 !MO.getGlobal()->hasInternalLinkage());
180 MCOperand X86MCInstLower::LowerSymbolOperand(const MachineOperand &MO,
181 MCSymbol *Sym) const {
182 // FIXME: We would like an efficient form for this, so we don't have to do a
183 // lot of extra uniquing.
184 const MCExpr *Expr = nullptr;
185 MCSymbolRefExpr::VariantKind RefKind = MCSymbolRefExpr::VK_None;
187 switch (MO.getTargetFlags()) {
188 default: llvm_unreachable("Unknown target flag on GV operand");
189 case X86II::MO_NO_FLAG: // No flag.
190 // These affect the name of the symbol, not any suffix.
191 case X86II::MO_DARWIN_NONLAZY:
192 case X86II::MO_DLLIMPORT:
195 case X86II::MO_TLVP: RefKind = MCSymbolRefExpr::VK_TLVP; break;
196 case X86II::MO_TLVP_PIC_BASE:
197 Expr = MCSymbolRefExpr::create(Sym, MCSymbolRefExpr::VK_TLVP, Ctx);
198 // Subtract the pic base.
199 Expr = MCBinaryExpr::createSub(Expr,
200 MCSymbolRefExpr::create(MF.getPICBaseSymbol(),
204 case X86II::MO_SECREL: RefKind = MCSymbolRefExpr::VK_SECREL; break;
205 case X86II::MO_TLSGD: RefKind = MCSymbolRefExpr::VK_TLSGD; break;
206 case X86II::MO_TLSLD: RefKind = MCSymbolRefExpr::VK_TLSLD; break;
207 case X86II::MO_TLSLDM: RefKind = MCSymbolRefExpr::VK_TLSLDM; break;
208 case X86II::MO_GOTTPOFF: RefKind = MCSymbolRefExpr::VK_GOTTPOFF; break;
209 case X86II::MO_INDNTPOFF: RefKind = MCSymbolRefExpr::VK_INDNTPOFF; break;
210 case X86II::MO_TPOFF: RefKind = MCSymbolRefExpr::VK_TPOFF; break;
211 case X86II::MO_DTPOFF: RefKind = MCSymbolRefExpr::VK_DTPOFF; break;
212 case X86II::MO_NTPOFF: RefKind = MCSymbolRefExpr::VK_NTPOFF; break;
213 case X86II::MO_GOTNTPOFF: RefKind = MCSymbolRefExpr::VK_GOTNTPOFF; break;
214 case X86II::MO_GOTPCREL: RefKind = MCSymbolRefExpr::VK_GOTPCREL; break;
215 case X86II::MO_GOT: RefKind = MCSymbolRefExpr::VK_GOT; break;
216 case X86II::MO_GOTOFF: RefKind = MCSymbolRefExpr::VK_GOTOFF; break;
217 case X86II::MO_PLT: RefKind = MCSymbolRefExpr::VK_PLT; break;
218 case X86II::MO_PIC_BASE_OFFSET:
219 case X86II::MO_DARWIN_NONLAZY_PIC_BASE:
220 Expr = MCSymbolRefExpr::create(Sym, Ctx);
221 // Subtract the pic base.
222 Expr = MCBinaryExpr::createSub(Expr,
223 MCSymbolRefExpr::create(MF.getPICBaseSymbol(), Ctx),
226 assert(MAI.doesSetDirectiveSuppressReloc());
227 // If .set directive is supported, use it to reduce the number of
228 // relocations the assembler will generate for differences between
229 // local labels. This is only safe when the symbols are in the same
230 // section so we are restricting it to jumptable references.
231 MCSymbol *Label = Ctx.createTempSymbol();
232 AsmPrinter.OutStreamer->EmitAssignment(Label, Expr);
233 Expr = MCSymbolRefExpr::create(Label, Ctx);
239 Expr = MCSymbolRefExpr::create(Sym, RefKind, Ctx);
241 if (!MO.isJTI() && !MO.isMBB() && MO.getOffset())
242 Expr = MCBinaryExpr::createAdd(Expr,
243 MCConstantExpr::create(MO.getOffset(), Ctx),
245 return MCOperand::createExpr(Expr);
249 /// \brief Simplify FOO $imm, %{al,ax,eax,rax} to FOO $imm, for instruction with
250 /// a short fixed-register form.
251 static void SimplifyShortImmForm(MCInst &Inst, unsigned Opcode) {
252 unsigned ImmOp = Inst.getNumOperands() - 1;
253 assert(Inst.getOperand(0).isReg() &&
254 (Inst.getOperand(ImmOp).isImm() || Inst.getOperand(ImmOp).isExpr()) &&
255 ((Inst.getNumOperands() == 3 && Inst.getOperand(1).isReg() &&
256 Inst.getOperand(0).getReg() == Inst.getOperand(1).getReg()) ||
257 Inst.getNumOperands() == 2) && "Unexpected instruction!");
259 // Check whether the destination register can be fixed.
260 unsigned Reg = Inst.getOperand(0).getReg();
261 if (Reg != X86::AL && Reg != X86::AX && Reg != X86::EAX && Reg != X86::RAX)
264 // If so, rewrite the instruction.
265 MCOperand Saved = Inst.getOperand(ImmOp);
267 Inst.setOpcode(Opcode);
268 Inst.addOperand(Saved);
271 /// \brief If a movsx instruction has a shorter encoding for the used register
272 /// simplify the instruction to use it instead.
273 static void SimplifyMOVSX(MCInst &Inst) {
274 unsigned NewOpcode = 0;
275 unsigned Op0 = Inst.getOperand(0).getReg(), Op1 = Inst.getOperand(1).getReg();
276 switch (Inst.getOpcode()) {
278 llvm_unreachable("Unexpected instruction!");
279 case X86::MOVSX16rr8: // movsbw %al, %ax --> cbtw
280 if (Op0 == X86::AX && Op1 == X86::AL)
281 NewOpcode = X86::CBW;
283 case X86::MOVSX32rr16: // movswl %ax, %eax --> cwtl
284 if (Op0 == X86::EAX && Op1 == X86::AX)
285 NewOpcode = X86::CWDE;
287 case X86::MOVSX64rr32: // movslq %eax, %rax --> cltq
288 if (Op0 == X86::RAX && Op1 == X86::EAX)
289 NewOpcode = X86::CDQE;
293 if (NewOpcode != 0) {
295 Inst.setOpcode(NewOpcode);
299 /// \brief Simplify things like MOV32rm to MOV32o32a.
300 static void SimplifyShortMoveForm(X86AsmPrinter &Printer, MCInst &Inst,
302 // Don't make these simplifications in 64-bit mode; other assemblers don't
303 // perform them because they make the code larger.
304 if (Printer.getSubtarget().is64Bit())
307 bool IsStore = Inst.getOperand(0).isReg() && Inst.getOperand(1).isReg();
308 unsigned AddrBase = IsStore;
309 unsigned RegOp = IsStore ? 0 : 5;
310 unsigned AddrOp = AddrBase + 3;
311 assert(Inst.getNumOperands() == 6 && Inst.getOperand(RegOp).isReg() &&
312 Inst.getOperand(AddrBase + X86::AddrBaseReg).isReg() &&
313 Inst.getOperand(AddrBase + X86::AddrScaleAmt).isImm() &&
314 Inst.getOperand(AddrBase + X86::AddrIndexReg).isReg() &&
315 Inst.getOperand(AddrBase + X86::AddrSegmentReg).isReg() &&
316 (Inst.getOperand(AddrOp).isExpr() ||
317 Inst.getOperand(AddrOp).isImm()) &&
318 "Unexpected instruction!");
320 // Check whether the destination register can be fixed.
321 unsigned Reg = Inst.getOperand(RegOp).getReg();
322 if (Reg != X86::AL && Reg != X86::AX && Reg != X86::EAX && Reg != X86::RAX)
325 // Check whether this is an absolute address.
326 // FIXME: We know TLVP symbol refs aren't, but there should be a better way
328 bool Absolute = true;
329 if (Inst.getOperand(AddrOp).isExpr()) {
330 const MCExpr *MCE = Inst.getOperand(AddrOp).getExpr();
331 if (const MCSymbolRefExpr *SRE = dyn_cast<MCSymbolRefExpr>(MCE))
332 if (SRE->getKind() == MCSymbolRefExpr::VK_TLVP)
337 (Inst.getOperand(AddrBase + X86::AddrBaseReg).getReg() != 0 ||
338 Inst.getOperand(AddrBase + X86::AddrScaleAmt).getImm() != 1 ||
339 Inst.getOperand(AddrBase + X86::AddrIndexReg).getReg() != 0))
342 // If so, rewrite the instruction.
343 MCOperand Saved = Inst.getOperand(AddrOp);
344 MCOperand Seg = Inst.getOperand(AddrBase + X86::AddrSegmentReg);
346 Inst.setOpcode(Opcode);
347 Inst.addOperand(Saved);
348 Inst.addOperand(Seg);
351 static unsigned getRetOpcode(const X86Subtarget &Subtarget) {
352 return Subtarget.is64Bit() ? X86::RETQ : X86::RETL;
356 X86MCInstLower::LowerMachineOperand(const MachineInstr *MI,
357 const MachineOperand &MO) const {
358 switch (MO.getType()) {
361 llvm_unreachable("unknown operand type");
362 case MachineOperand::MO_Register:
363 // Ignore all implicit register operands.
366 return MCOperand::createReg(MO.getReg());
367 case MachineOperand::MO_Immediate:
368 return MCOperand::createImm(MO.getImm());
369 case MachineOperand::MO_MachineBasicBlock:
370 case MachineOperand::MO_GlobalAddress:
371 case MachineOperand::MO_ExternalSymbol:
372 return LowerSymbolOperand(MO, GetSymbolFromOperand(MO));
373 case MachineOperand::MO_MCSymbol:
374 return LowerSymbolOperand(MO, MO.getMCSymbol());
375 case MachineOperand::MO_JumpTableIndex:
376 return LowerSymbolOperand(MO, AsmPrinter.GetJTISymbol(MO.getIndex()));
377 case MachineOperand::MO_ConstantPoolIndex:
378 return LowerSymbolOperand(MO, AsmPrinter.GetCPISymbol(MO.getIndex()));
379 case MachineOperand::MO_BlockAddress:
380 return LowerSymbolOperand(
381 MO, AsmPrinter.GetBlockAddressSymbol(MO.getBlockAddress()));
382 case MachineOperand::MO_RegisterMask:
383 // Ignore call clobbers.
388 void X86MCInstLower::Lower(const MachineInstr *MI, MCInst &OutMI) const {
389 OutMI.setOpcode(MI->getOpcode());
391 for (const MachineOperand &MO : MI->operands())
392 if (auto MaybeMCOp = LowerMachineOperand(MI, MO))
393 OutMI.addOperand(MaybeMCOp.getValue());
395 // Handle a few special cases to eliminate operand modifiers.
397 switch (OutMI.getOpcode()) {
402 // LEA should have a segment register, but it must be empty.
403 assert(OutMI.getNumOperands() == 1+X86::AddrNumOperands &&
404 "Unexpected # of LEA operands");
405 assert(OutMI.getOperand(1+X86::AddrSegmentReg).getReg() == 0 &&
406 "LEA has segment specified!");
409 // Commute operands to get a smaller encoding by using VEX.R instead of VEX.B
410 // if one of the registers is extended, but other isn't.
411 case X86::VMOVZPQILo2PQIrr:
413 case X86::VMOVAPDYrr:
415 case X86::VMOVAPSYrr:
417 case X86::VMOVDQAYrr:
419 case X86::VMOVDQUYrr:
421 case X86::VMOVUPDYrr:
423 case X86::VMOVUPSYrr: {
424 if (!X86II::isX86_64ExtendedReg(OutMI.getOperand(0).getReg()) &&
425 X86II::isX86_64ExtendedReg(OutMI.getOperand(1).getReg())) {
427 switch (OutMI.getOpcode()) {
428 default: llvm_unreachable("Invalid opcode");
429 case X86::VMOVZPQILo2PQIrr: NewOpc = X86::VMOVPQI2QIrr; break;
430 case X86::VMOVAPDrr: NewOpc = X86::VMOVAPDrr_REV; break;
431 case X86::VMOVAPDYrr: NewOpc = X86::VMOVAPDYrr_REV; break;
432 case X86::VMOVAPSrr: NewOpc = X86::VMOVAPSrr_REV; break;
433 case X86::VMOVAPSYrr: NewOpc = X86::VMOVAPSYrr_REV; break;
434 case X86::VMOVDQArr: NewOpc = X86::VMOVDQArr_REV; break;
435 case X86::VMOVDQAYrr: NewOpc = X86::VMOVDQAYrr_REV; break;
436 case X86::VMOVDQUrr: NewOpc = X86::VMOVDQUrr_REV; break;
437 case X86::VMOVDQUYrr: NewOpc = X86::VMOVDQUYrr_REV; break;
438 case X86::VMOVUPDrr: NewOpc = X86::VMOVUPDrr_REV; break;
439 case X86::VMOVUPDYrr: NewOpc = X86::VMOVUPDYrr_REV; break;
440 case X86::VMOVUPSrr: NewOpc = X86::VMOVUPSrr_REV; break;
441 case X86::VMOVUPSYrr: NewOpc = X86::VMOVUPSYrr_REV; break;
443 OutMI.setOpcode(NewOpc);
448 case X86::VMOVSSrr: {
449 if (!X86II::isX86_64ExtendedReg(OutMI.getOperand(0).getReg()) &&
450 X86II::isX86_64ExtendedReg(OutMI.getOperand(2).getReg())) {
452 switch (OutMI.getOpcode()) {
453 default: llvm_unreachable("Invalid opcode");
454 case X86::VMOVSDrr: NewOpc = X86::VMOVSDrr_REV; break;
455 case X86::VMOVSSrr: NewOpc = X86::VMOVSSrr_REV; break;
457 OutMI.setOpcode(NewOpc);
462 // TAILJMPr64, CALL64r, CALL64pcrel32 - These instructions have register
463 // inputs modeled as normal uses instead of implicit uses. As such, truncate
464 // off all but the first operand (the callee). FIXME: Change isel.
465 case X86::TAILJMPr64:
466 case X86::TAILJMPr64_REX:
468 case X86::CALL64pcrel32: {
469 unsigned Opcode = OutMI.getOpcode();
470 MCOperand Saved = OutMI.getOperand(0);
472 OutMI.setOpcode(Opcode);
473 OutMI.addOperand(Saved);
478 case X86::EH_RETURN64: {
480 OutMI.setOpcode(getRetOpcode(AsmPrinter.getSubtarget()));
484 case X86::CLEANUPRET: {
485 // Replace CATCHRET with the appropriate RET.
487 OutMI.setOpcode(getRetOpcode(AsmPrinter.getSubtarget()));
491 case X86::CATCHRET: {
492 // Replace CATCHRET with the appropriate RET.
493 const X86Subtarget &Subtarget = AsmPrinter.getSubtarget();
494 unsigned ReturnReg = Subtarget.is64Bit() ? X86::RAX : X86::EAX;
496 OutMI.setOpcode(getRetOpcode(Subtarget));
497 OutMI.addOperand(MCOperand::createReg(ReturnReg));
501 // TAILJMPd, TAILJMPd64, TailJMPd_cc - Lower to the correct jump instruction.
503 case X86::TAILJMPr: Opcode = X86::JMP32r; goto SetTailJmpOpcode;
505 case X86::TAILJMPd64: Opcode = X86::JMP_1; goto SetTailJmpOpcode;
506 case X86::TAILJMPd_CC:
507 case X86::TAILJMPd64_CC:
508 Opcode = X86::GetCondBranchFromCond(
509 static_cast<X86::CondCode>(MI->getOperand(1).getImm()));
510 goto SetTailJmpOpcode;
513 MCOperand Saved = OutMI.getOperand(0);
515 OutMI.setOpcode(Opcode);
516 OutMI.addOperand(Saved);
524 // If we aren't in 64-bit mode we can use the 1-byte inc/dec instructions.
525 if (!AsmPrinter.getSubtarget().is64Bit()) {
527 switch (OutMI.getOpcode()) {
528 default: llvm_unreachable("Invalid opcode");
529 case X86::DEC16r: Opcode = X86::DEC16r_alt; break;
530 case X86::DEC32r: Opcode = X86::DEC32r_alt; break;
531 case X86::INC16r: Opcode = X86::INC16r_alt; break;
532 case X86::INC32r: Opcode = X86::INC32r_alt; break;
534 OutMI.setOpcode(Opcode);
538 // These are pseudo-ops for OR to help with the OR->ADD transformation. We do
539 // this with an ugly goto in case the resultant OR uses EAX and needs the
541 case X86::ADD16rr_DB: OutMI.setOpcode(X86::OR16rr); goto ReSimplify;
542 case X86::ADD32rr_DB: OutMI.setOpcode(X86::OR32rr); goto ReSimplify;
543 case X86::ADD64rr_DB: OutMI.setOpcode(X86::OR64rr); goto ReSimplify;
544 case X86::ADD16ri_DB: OutMI.setOpcode(X86::OR16ri); goto ReSimplify;
545 case X86::ADD32ri_DB: OutMI.setOpcode(X86::OR32ri); goto ReSimplify;
546 case X86::ADD64ri32_DB: OutMI.setOpcode(X86::OR64ri32); goto ReSimplify;
547 case X86::ADD16ri8_DB: OutMI.setOpcode(X86::OR16ri8); goto ReSimplify;
548 case X86::ADD32ri8_DB: OutMI.setOpcode(X86::OR32ri8); goto ReSimplify;
549 case X86::ADD64ri8_DB: OutMI.setOpcode(X86::OR64ri8); goto ReSimplify;
551 // Atomic load and store require a separate pseudo-inst because Acquire
552 // implies mayStore and Release implies mayLoad; fix these to regular MOV
554 case X86::ACQUIRE_MOV8rm: OutMI.setOpcode(X86::MOV8rm); goto ReSimplify;
555 case X86::ACQUIRE_MOV16rm: OutMI.setOpcode(X86::MOV16rm); goto ReSimplify;
556 case X86::ACQUIRE_MOV32rm: OutMI.setOpcode(X86::MOV32rm); goto ReSimplify;
557 case X86::ACQUIRE_MOV64rm: OutMI.setOpcode(X86::MOV64rm); goto ReSimplify;
558 case X86::RELEASE_MOV8mr: OutMI.setOpcode(X86::MOV8mr); goto ReSimplify;
559 case X86::RELEASE_MOV16mr: OutMI.setOpcode(X86::MOV16mr); goto ReSimplify;
560 case X86::RELEASE_MOV32mr: OutMI.setOpcode(X86::MOV32mr); goto ReSimplify;
561 case X86::RELEASE_MOV64mr: OutMI.setOpcode(X86::MOV64mr); goto ReSimplify;
562 case X86::RELEASE_MOV8mi: OutMI.setOpcode(X86::MOV8mi); goto ReSimplify;
563 case X86::RELEASE_MOV16mi: OutMI.setOpcode(X86::MOV16mi); goto ReSimplify;
564 case X86::RELEASE_MOV32mi: OutMI.setOpcode(X86::MOV32mi); goto ReSimplify;
565 case X86::RELEASE_MOV64mi32: OutMI.setOpcode(X86::MOV64mi32); goto ReSimplify;
566 case X86::RELEASE_ADD8mi: OutMI.setOpcode(X86::ADD8mi); goto ReSimplify;
567 case X86::RELEASE_ADD8mr: OutMI.setOpcode(X86::ADD8mr); goto ReSimplify;
568 case X86::RELEASE_ADD32mi: OutMI.setOpcode(X86::ADD32mi); goto ReSimplify;
569 case X86::RELEASE_ADD32mr: OutMI.setOpcode(X86::ADD32mr); goto ReSimplify;
570 case X86::RELEASE_ADD64mi32: OutMI.setOpcode(X86::ADD64mi32); goto ReSimplify;
571 case X86::RELEASE_ADD64mr: OutMI.setOpcode(X86::ADD64mr); goto ReSimplify;
572 case X86::RELEASE_AND8mi: OutMI.setOpcode(X86::AND8mi); goto ReSimplify;
573 case X86::RELEASE_AND8mr: OutMI.setOpcode(X86::AND8mr); goto ReSimplify;
574 case X86::RELEASE_AND32mi: OutMI.setOpcode(X86::AND32mi); goto ReSimplify;
575 case X86::RELEASE_AND32mr: OutMI.setOpcode(X86::AND32mr); goto ReSimplify;
576 case X86::RELEASE_AND64mi32: OutMI.setOpcode(X86::AND64mi32); goto ReSimplify;
577 case X86::RELEASE_AND64mr: OutMI.setOpcode(X86::AND64mr); goto ReSimplify;
578 case X86::RELEASE_OR8mi: OutMI.setOpcode(X86::OR8mi); goto ReSimplify;
579 case X86::RELEASE_OR8mr: OutMI.setOpcode(X86::OR8mr); goto ReSimplify;
580 case X86::RELEASE_OR32mi: OutMI.setOpcode(X86::OR32mi); goto ReSimplify;
581 case X86::RELEASE_OR32mr: OutMI.setOpcode(X86::OR32mr); goto ReSimplify;
582 case X86::RELEASE_OR64mi32: OutMI.setOpcode(X86::OR64mi32); goto ReSimplify;
583 case X86::RELEASE_OR64mr: OutMI.setOpcode(X86::OR64mr); goto ReSimplify;
584 case X86::RELEASE_XOR8mi: OutMI.setOpcode(X86::XOR8mi); goto ReSimplify;
585 case X86::RELEASE_XOR8mr: OutMI.setOpcode(X86::XOR8mr); goto ReSimplify;
586 case X86::RELEASE_XOR32mi: OutMI.setOpcode(X86::XOR32mi); goto ReSimplify;
587 case X86::RELEASE_XOR32mr: OutMI.setOpcode(X86::XOR32mr); goto ReSimplify;
588 case X86::RELEASE_XOR64mi32: OutMI.setOpcode(X86::XOR64mi32); goto ReSimplify;
589 case X86::RELEASE_XOR64mr: OutMI.setOpcode(X86::XOR64mr); goto ReSimplify;
590 case X86::RELEASE_INC8m: OutMI.setOpcode(X86::INC8m); goto ReSimplify;
591 case X86::RELEASE_INC16m: OutMI.setOpcode(X86::INC16m); goto ReSimplify;
592 case X86::RELEASE_INC32m: OutMI.setOpcode(X86::INC32m); goto ReSimplify;
593 case X86::RELEASE_INC64m: OutMI.setOpcode(X86::INC64m); goto ReSimplify;
594 case X86::RELEASE_DEC8m: OutMI.setOpcode(X86::DEC8m); goto ReSimplify;
595 case X86::RELEASE_DEC16m: OutMI.setOpcode(X86::DEC16m); goto ReSimplify;
596 case X86::RELEASE_DEC32m: OutMI.setOpcode(X86::DEC32m); goto ReSimplify;
597 case X86::RELEASE_DEC64m: OutMI.setOpcode(X86::DEC64m); goto ReSimplify;
599 // We don't currently select the correct instruction form for instructions
600 // which have a short %eax, etc. form. Handle this by custom lowering, for
603 // Note, we are currently not handling the following instructions:
604 // MOV64ao8, MOV64o8a
605 // XCHG16ar, XCHG32ar, XCHG64ar
606 case X86::MOV8mr_NOREX:
608 case X86::MOV8rm_NOREX:
615 switch (OutMI.getOpcode()) {
616 default: llvm_unreachable("Invalid opcode");
617 case X86::MOV8mr_NOREX:
618 case X86::MOV8mr: NewOpc = X86::MOV8o32a; break;
619 case X86::MOV8rm_NOREX:
620 case X86::MOV8rm: NewOpc = X86::MOV8ao32; break;
621 case X86::MOV16mr: NewOpc = X86::MOV16o32a; break;
622 case X86::MOV16rm: NewOpc = X86::MOV16ao32; break;
623 case X86::MOV32mr: NewOpc = X86::MOV32o32a; break;
624 case X86::MOV32rm: NewOpc = X86::MOV32ao32; break;
626 SimplifyShortMoveForm(AsmPrinter, OutMI, NewOpc);
630 case X86::ADC8ri: case X86::ADC16ri: case X86::ADC32ri: case X86::ADC64ri32:
631 case X86::ADD8ri: case X86::ADD16ri: case X86::ADD32ri: case X86::ADD64ri32:
632 case X86::AND8ri: case X86::AND16ri: case X86::AND32ri: case X86::AND64ri32:
633 case X86::CMP8ri: case X86::CMP16ri: case X86::CMP32ri: case X86::CMP64ri32:
634 case X86::OR8ri: case X86::OR16ri: case X86::OR32ri: case X86::OR64ri32:
635 case X86::SBB8ri: case X86::SBB16ri: case X86::SBB32ri: case X86::SBB64ri32:
636 case X86::SUB8ri: case X86::SUB16ri: case X86::SUB32ri: case X86::SUB64ri32:
637 case X86::TEST8ri:case X86::TEST16ri:case X86::TEST32ri:case X86::TEST64ri32:
638 case X86::XOR8ri: case X86::XOR16ri: case X86::XOR32ri: case X86::XOR64ri32: {
640 switch (OutMI.getOpcode()) {
641 default: llvm_unreachable("Invalid opcode");
642 case X86::ADC8ri: NewOpc = X86::ADC8i8; break;
643 case X86::ADC16ri: NewOpc = X86::ADC16i16; break;
644 case X86::ADC32ri: NewOpc = X86::ADC32i32; break;
645 case X86::ADC64ri32: NewOpc = X86::ADC64i32; break;
646 case X86::ADD8ri: NewOpc = X86::ADD8i8; break;
647 case X86::ADD16ri: NewOpc = X86::ADD16i16; break;
648 case X86::ADD32ri: NewOpc = X86::ADD32i32; break;
649 case X86::ADD64ri32: NewOpc = X86::ADD64i32; break;
650 case X86::AND8ri: NewOpc = X86::AND8i8; break;
651 case X86::AND16ri: NewOpc = X86::AND16i16; break;
652 case X86::AND32ri: NewOpc = X86::AND32i32; break;
653 case X86::AND64ri32: NewOpc = X86::AND64i32; break;
654 case X86::CMP8ri: NewOpc = X86::CMP8i8; break;
655 case X86::CMP16ri: NewOpc = X86::CMP16i16; break;
656 case X86::CMP32ri: NewOpc = X86::CMP32i32; break;
657 case X86::CMP64ri32: NewOpc = X86::CMP64i32; break;
658 case X86::OR8ri: NewOpc = X86::OR8i8; break;
659 case X86::OR16ri: NewOpc = X86::OR16i16; break;
660 case X86::OR32ri: NewOpc = X86::OR32i32; break;
661 case X86::OR64ri32: NewOpc = X86::OR64i32; break;
662 case X86::SBB8ri: NewOpc = X86::SBB8i8; break;
663 case X86::SBB16ri: NewOpc = X86::SBB16i16; break;
664 case X86::SBB32ri: NewOpc = X86::SBB32i32; break;
665 case X86::SBB64ri32: NewOpc = X86::SBB64i32; break;
666 case X86::SUB8ri: NewOpc = X86::SUB8i8; break;
667 case X86::SUB16ri: NewOpc = X86::SUB16i16; break;
668 case X86::SUB32ri: NewOpc = X86::SUB32i32; break;
669 case X86::SUB64ri32: NewOpc = X86::SUB64i32; break;
670 case X86::TEST8ri: NewOpc = X86::TEST8i8; break;
671 case X86::TEST16ri: NewOpc = X86::TEST16i16; break;
672 case X86::TEST32ri: NewOpc = X86::TEST32i32; break;
673 case X86::TEST64ri32: NewOpc = X86::TEST64i32; break;
674 case X86::XOR8ri: NewOpc = X86::XOR8i8; break;
675 case X86::XOR16ri: NewOpc = X86::XOR16i16; break;
676 case X86::XOR32ri: NewOpc = X86::XOR32i32; break;
677 case X86::XOR64ri32: NewOpc = X86::XOR64i32; break;
679 SimplifyShortImmForm(OutMI, NewOpc);
683 // Try to shrink some forms of movsx.
684 case X86::MOVSX16rr8:
685 case X86::MOVSX32rr16:
686 case X86::MOVSX64rr32:
687 SimplifyMOVSX(OutMI);
692 void X86AsmPrinter::LowerTlsAddr(X86MCInstLower &MCInstLowering,
693 const MachineInstr &MI) {
695 bool is64Bits = MI.getOpcode() == X86::TLS_addr64 ||
696 MI.getOpcode() == X86::TLS_base_addr64;
698 bool needsPadding = MI.getOpcode() == X86::TLS_addr64;
700 MCContext &context = OutStreamer->getContext();
703 EmitAndCountInstruction(MCInstBuilder(X86::DATA16_PREFIX));
705 MCSymbolRefExpr::VariantKind SRVK;
706 switch (MI.getOpcode()) {
707 case X86::TLS_addr32:
708 case X86::TLS_addr64:
709 SRVK = MCSymbolRefExpr::VK_TLSGD;
711 case X86::TLS_base_addr32:
712 SRVK = MCSymbolRefExpr::VK_TLSLDM;
714 case X86::TLS_base_addr64:
715 SRVK = MCSymbolRefExpr::VK_TLSLD;
718 llvm_unreachable("unexpected opcode");
721 MCSymbol *sym = MCInstLowering.GetSymbolFromOperand(MI.getOperand(3));
722 const MCSymbolRefExpr *symRef = MCSymbolRefExpr::create(sym, SRVK, context);
726 LEA.setOpcode(X86::LEA64r);
727 LEA.addOperand(MCOperand::createReg(X86::RDI)); // dest
728 LEA.addOperand(MCOperand::createReg(X86::RIP)); // base
729 LEA.addOperand(MCOperand::createImm(1)); // scale
730 LEA.addOperand(MCOperand::createReg(0)); // index
731 LEA.addOperand(MCOperand::createExpr(symRef)); // disp
732 LEA.addOperand(MCOperand::createReg(0)); // seg
733 } else if (SRVK == MCSymbolRefExpr::VK_TLSLDM) {
734 LEA.setOpcode(X86::LEA32r);
735 LEA.addOperand(MCOperand::createReg(X86::EAX)); // dest
736 LEA.addOperand(MCOperand::createReg(X86::EBX)); // base
737 LEA.addOperand(MCOperand::createImm(1)); // scale
738 LEA.addOperand(MCOperand::createReg(0)); // index
739 LEA.addOperand(MCOperand::createExpr(symRef)); // disp
740 LEA.addOperand(MCOperand::createReg(0)); // seg
742 LEA.setOpcode(X86::LEA32r);
743 LEA.addOperand(MCOperand::createReg(X86::EAX)); // dest
744 LEA.addOperand(MCOperand::createReg(0)); // base
745 LEA.addOperand(MCOperand::createImm(1)); // scale
746 LEA.addOperand(MCOperand::createReg(X86::EBX)); // index
747 LEA.addOperand(MCOperand::createExpr(symRef)); // disp
748 LEA.addOperand(MCOperand::createReg(0)); // seg
750 EmitAndCountInstruction(LEA);
753 EmitAndCountInstruction(MCInstBuilder(X86::DATA16_PREFIX));
754 EmitAndCountInstruction(MCInstBuilder(X86::DATA16_PREFIX));
755 EmitAndCountInstruction(MCInstBuilder(X86::REX64_PREFIX));
758 StringRef name = is64Bits ? "__tls_get_addr" : "___tls_get_addr";
759 MCSymbol *tlsGetAddr = context.getOrCreateSymbol(name);
760 const MCSymbolRefExpr *tlsRef =
761 MCSymbolRefExpr::create(tlsGetAddr,
762 MCSymbolRefExpr::VK_PLT,
765 EmitAndCountInstruction(MCInstBuilder(is64Bits ? X86::CALL64pcrel32
770 /// \brief Emit the largest nop instruction smaller than or equal to \p NumBytes
771 /// bytes. Return the size of nop emitted.
772 static unsigned EmitNop(MCStreamer &OS, unsigned NumBytes, bool Is64Bit,
773 const MCSubtargetInfo &STI) {
774 // This works only for 64bit. For 32bit we have to do additional checking if
775 // the CPU supports multi-byte nops.
776 assert(Is64Bit && "EmitNops only supports X86-64");
779 unsigned Opc, BaseReg, ScaleVal, IndexReg, Displacement, SegmentReg;
780 Opc = IndexReg = Displacement = SegmentReg = 0;
784 case 0: llvm_unreachable("Zero nops?"); break;
785 case 1: NopSize = 1; Opc = X86::NOOP; break;
786 case 2: NopSize = 2; Opc = X86::XCHG16ar; break;
787 case 3: NopSize = 3; Opc = X86::NOOPL; break;
788 case 4: NopSize = 4; Opc = X86::NOOPL; Displacement = 8; break;
789 case 5: NopSize = 5; Opc = X86::NOOPL; Displacement = 8;
790 IndexReg = X86::RAX; break;
791 case 6: NopSize = 6; Opc = X86::NOOPW; Displacement = 8;
792 IndexReg = X86::RAX; break;
793 case 7: NopSize = 7; Opc = X86::NOOPL; Displacement = 512; break;
794 case 8: NopSize = 8; Opc = X86::NOOPL; Displacement = 512;
795 IndexReg = X86::RAX; break;
796 case 9: NopSize = 9; Opc = X86::NOOPW; Displacement = 512;
797 IndexReg = X86::RAX; break;
798 default: NopSize = 10; Opc = X86::NOOPW; Displacement = 512;
799 IndexReg = X86::RAX; SegmentReg = X86::CS; break;
802 unsigned NumPrefixes = std::min(NumBytes - NopSize, 5U);
803 NopSize += NumPrefixes;
804 for (unsigned i = 0; i != NumPrefixes; ++i)
805 OS.EmitBytes("\x66");
809 llvm_unreachable("Unexpected opcode");
812 OS.EmitInstruction(MCInstBuilder(Opc), STI);
815 OS.EmitInstruction(MCInstBuilder(Opc).addReg(X86::AX), STI);
819 OS.EmitInstruction(MCInstBuilder(Opc)
823 .addImm(Displacement)
828 assert(NopSize <= NumBytes && "We overemitted?");
832 /// \brief Emit the optimal amount of multi-byte nops on X86.
833 static void EmitNops(MCStreamer &OS, unsigned NumBytes, bool Is64Bit,
834 const MCSubtargetInfo &STI) {
835 unsigned NopsToEmit = NumBytes;
838 NumBytes -= EmitNop(OS, NumBytes, Is64Bit, STI);
839 assert(NopsToEmit >= NumBytes && "Emitted more than I asked for!");
843 void X86AsmPrinter::LowerSTATEPOINT(const MachineInstr &MI,
844 X86MCInstLower &MCIL) {
845 assert(Subtarget->is64Bit() && "Statepoint currently only supports X86-64");
847 StatepointOpers SOpers(&MI);
848 if (unsigned PatchBytes = SOpers.getNumPatchBytes()) {
849 EmitNops(*OutStreamer, PatchBytes, Subtarget->is64Bit(),
852 // Lower call target and choose correct opcode
853 const MachineOperand &CallTarget = SOpers.getCallTarget();
854 MCOperand CallTargetMCOp;
856 switch (CallTarget.getType()) {
857 case MachineOperand::MO_GlobalAddress:
858 case MachineOperand::MO_ExternalSymbol:
859 CallTargetMCOp = MCIL.LowerSymbolOperand(
860 CallTarget, MCIL.GetSymbolFromOperand(CallTarget));
861 CallOpcode = X86::CALL64pcrel32;
862 // Currently, we only support relative addressing with statepoints.
863 // Otherwise, we'll need a scratch register to hold the target
864 // address. You'll fail asserts during load & relocation if this
865 // symbol is to far away. (TODO: support non-relative addressing)
867 case MachineOperand::MO_Immediate:
868 CallTargetMCOp = MCOperand::createImm(CallTarget.getImm());
869 CallOpcode = X86::CALL64pcrel32;
870 // Currently, we only support relative addressing with statepoints.
871 // Otherwise, we'll need a scratch register to hold the target
872 // immediate. You'll fail asserts during load & relocation if this
873 // address is to far away. (TODO: support non-relative addressing)
875 case MachineOperand::MO_Register:
876 CallTargetMCOp = MCOperand::createReg(CallTarget.getReg());
877 CallOpcode = X86::CALL64r;
880 llvm_unreachable("Unsupported operand type in statepoint call target");
886 CallInst.setOpcode(CallOpcode);
887 CallInst.addOperand(CallTargetMCOp);
888 OutStreamer->EmitInstruction(CallInst, getSubtargetInfo());
891 // Record our statepoint node in the same section used by STACKMAP
893 SM.recordStatepoint(MI);
896 void X86AsmPrinter::LowerFAULTING_LOAD_OP(const MachineInstr &MI,
897 X86MCInstLower &MCIL) {
898 // FAULTING_LOAD_OP <def>, <MBB handler>, <load opcode>, <load operands>
900 unsigned LoadDefRegister = MI.getOperand(0).getReg();
901 MCSymbol *HandlerLabel = MI.getOperand(1).getMBB()->getSymbol();
902 unsigned LoadOpcode = MI.getOperand(2).getImm();
903 unsigned LoadOperandsBeginIdx = 3;
905 FM.recordFaultingOp(FaultMaps::FaultingLoad, HandlerLabel);
908 LoadMI.setOpcode(LoadOpcode);
910 if (LoadDefRegister != X86::NoRegister)
911 LoadMI.addOperand(MCOperand::createReg(LoadDefRegister));
913 for (auto I = MI.operands_begin() + LoadOperandsBeginIdx,
914 E = MI.operands_end();
916 if (auto MaybeOperand = MCIL.LowerMachineOperand(&MI, *I))
917 LoadMI.addOperand(MaybeOperand.getValue());
919 OutStreamer->EmitInstruction(LoadMI, getSubtargetInfo());
922 void X86AsmPrinter::LowerPATCHABLE_OP(const MachineInstr &MI,
923 X86MCInstLower &MCIL) {
924 // PATCHABLE_OP minsize, opcode, operands
926 unsigned MinSize = MI.getOperand(0).getImm();
927 unsigned Opcode = MI.getOperand(1).getImm();
930 MCI.setOpcode(Opcode);
931 for (auto &MO : make_range(MI.operands_begin() + 2, MI.operands_end()))
932 if (auto MaybeOperand = MCIL.LowerMachineOperand(&MI, MO))
933 MCI.addOperand(MaybeOperand.getValue());
935 SmallString<256> Code;
936 SmallVector<MCFixup, 4> Fixups;
937 raw_svector_ostream VecOS(Code);
938 CodeEmitter->encodeInstruction(MCI, VecOS, Fixups, getSubtargetInfo());
940 if (Code.size() < MinSize) {
941 if (MinSize == 2 && Opcode == X86::PUSH64r) {
942 // This is an optimization that lets us get away without emitting a nop in
945 // NB! In some cases the encoding for PUSH64r (e.g. PUSH64r %R9) takes two
946 // bytes too, so the check on MinSize is important.
947 MCI.setOpcode(X86::PUSH64rmr);
949 unsigned NopSize = EmitNop(*OutStreamer, MinSize, Subtarget->is64Bit(),
951 assert(NopSize == MinSize && "Could not implement MinSize!");
956 OutStreamer->EmitInstruction(MCI, getSubtargetInfo());
959 // Lower a stackmap of the form:
960 // <id>, <shadowBytes>, ...
961 void X86AsmPrinter::LowerSTACKMAP(const MachineInstr &MI) {
962 SMShadowTracker.emitShadowPadding(*OutStreamer, getSubtargetInfo());
963 SM.recordStackMap(MI);
964 unsigned NumShadowBytes = MI.getOperand(1).getImm();
965 SMShadowTracker.reset(NumShadowBytes);
968 // Lower a patchpoint of the form:
969 // [<def>], <id>, <numBytes>, <target>, <numArgs>, <cc>, ...
970 void X86AsmPrinter::LowerPATCHPOINT(const MachineInstr &MI,
971 X86MCInstLower &MCIL) {
972 assert(Subtarget->is64Bit() && "Patchpoint currently only supports X86-64");
974 SMShadowTracker.emitShadowPadding(*OutStreamer, getSubtargetInfo());
976 SM.recordPatchPoint(MI);
978 PatchPointOpers opers(&MI);
979 unsigned ScratchIdx = opers.getNextScratchIdx();
980 unsigned EncodedBytes = 0;
981 const MachineOperand &CalleeMO = opers.getCallTarget();
983 // Check for null target. If target is non-null (i.e. is non-zero or is
984 // symbolic) then emit a call.
985 if (!(CalleeMO.isImm() && !CalleeMO.getImm())) {
986 MCOperand CalleeMCOp;
987 switch (CalleeMO.getType()) {
989 /// FIXME: Add a verifier check for bad callee types.
990 llvm_unreachable("Unrecognized callee operand type.");
991 case MachineOperand::MO_Immediate:
992 if (CalleeMO.getImm())
993 CalleeMCOp = MCOperand::createImm(CalleeMO.getImm());
995 case MachineOperand::MO_ExternalSymbol:
996 case MachineOperand::MO_GlobalAddress:
998 MCIL.LowerSymbolOperand(CalleeMO,
999 MCIL.GetSymbolFromOperand(CalleeMO));
1003 // Emit MOV to materialize the target address and the CALL to target.
1004 // This is encoded with 12-13 bytes, depending on which register is used.
1005 unsigned ScratchReg = MI.getOperand(ScratchIdx).getReg();
1006 if (X86II::isX86_64ExtendedReg(ScratchReg))
1011 EmitAndCountInstruction(
1012 MCInstBuilder(X86::MOV64ri).addReg(ScratchReg).addOperand(CalleeMCOp));
1013 EmitAndCountInstruction(MCInstBuilder(X86::CALL64r).addReg(ScratchReg));
1017 unsigned NumBytes = opers.getNumPatchBytes();
1018 assert(NumBytes >= EncodedBytes &&
1019 "Patchpoint can't request size less than the length of a call.");
1021 EmitNops(*OutStreamer, NumBytes - EncodedBytes, Subtarget->is64Bit(),
1022 getSubtargetInfo());
1025 void X86AsmPrinter::LowerPATCHABLE_FUNCTION_ENTER(const MachineInstr &MI,
1026 X86MCInstLower &MCIL) {
1027 // We want to emit the following pattern:
1032 // # 9 bytes worth of noops
1035 // We need the 9 bytes because at runtime, we'd be patching over the full 11
1036 // bytes with the following pattern:
1038 // mov %r10, <function id, 32-bit> // 6 bytes
1039 // call <relative offset, 32-bits> // 5 bytes
1041 auto CurSled = OutContext.createTempSymbol("xray_sled_", true);
1042 OutStreamer->EmitCodeAlignment(2);
1043 OutStreamer->EmitLabel(CurSled);
1044 auto Target = OutContext.createTempSymbol();
1046 // Use a two-byte `jmp`. This version of JMP takes an 8-bit relative offset as
1047 // an operand (computed as an offset from the jmp instruction).
1048 // FIXME: Find another less hacky way do force the relative jump.
1049 OutStreamer->EmitBytes("\xeb\x09");
1050 EmitNops(*OutStreamer, 9, Subtarget->is64Bit(), getSubtargetInfo());
1051 OutStreamer->EmitLabel(Target);
1052 recordSled(CurSled, MI, SledKind::FUNCTION_ENTER);
1055 void X86AsmPrinter::LowerPATCHABLE_RET(const MachineInstr &MI,
1056 X86MCInstLower &MCIL) {
1057 // Since PATCHABLE_RET takes the opcode of the return statement as an
1058 // argument, we use that to emit the correct form of the RET that we want.
1059 // i.e. when we see this:
1061 // PATCHABLE_RET X86::RET ...
1063 // We should emit the RET followed by sleds.
1067 // ret # or equivalent instruction
1068 // # 10 bytes worth of noops
1070 // This just makes sure that the alignment for the next instruction is 2.
1071 auto CurSled = OutContext.createTempSymbol("xray_sled_", true);
1072 OutStreamer->EmitCodeAlignment(2);
1073 OutStreamer->EmitLabel(CurSled);
1074 unsigned OpCode = MI.getOperand(0).getImm();
1076 Ret.setOpcode(OpCode);
1077 for (auto &MO : make_range(MI.operands_begin() + 1, MI.operands_end()))
1078 if (auto MaybeOperand = MCIL.LowerMachineOperand(&MI, MO))
1079 Ret.addOperand(MaybeOperand.getValue());
1080 OutStreamer->EmitInstruction(Ret, getSubtargetInfo());
1081 EmitNops(*OutStreamer, 10, Subtarget->is64Bit(), getSubtargetInfo());
1082 recordSled(CurSled, MI, SledKind::FUNCTION_EXIT);
1085 void X86AsmPrinter::LowerPATCHABLE_TAIL_CALL(const MachineInstr &MI, X86MCInstLower &MCIL) {
1086 // Like PATCHABLE_RET, we have the actual instruction in the operands to this
1087 // instruction so we lower that particular instruction and its operands.
1088 // Unlike PATCHABLE_RET though, we put the sled before the JMP, much like how
1089 // we do it for PATCHABLE_FUNCTION_ENTER. The sled should be very similar to
1090 // the PATCHABLE_FUNCTION_ENTER case, followed by the lowering of the actual
1091 // tail call much like how we have it in PATCHABLE_RET.
1092 auto CurSled = OutContext.createTempSymbol("xray_sled_", true);
1093 OutStreamer->EmitCodeAlignment(2);
1094 OutStreamer->EmitLabel(CurSled);
1095 auto Target = OutContext.createTempSymbol();
1097 // Use a two-byte `jmp`. This version of JMP takes an 8-bit relative offset as
1098 // an operand (computed as an offset from the jmp instruction).
1099 // FIXME: Find another less hacky way do force the relative jump.
1100 OutStreamer->EmitBytes("\xeb\x09");
1101 EmitNops(*OutStreamer, 9, Subtarget->is64Bit(), getSubtargetInfo());
1102 OutStreamer->EmitLabel(Target);
1103 recordSled(CurSled, MI, SledKind::TAIL_CALL);
1105 unsigned OpCode = MI.getOperand(0).getImm();
1107 TC.setOpcode(OpCode);
1109 // Before emitting the instruction, add a comment to indicate that this is
1110 // indeed a tail call.
1111 OutStreamer->AddComment("TAILCALL");
1112 for (auto &MO : make_range(MI.operands_begin() + 1, MI.operands_end()))
1113 if (auto MaybeOperand = MCIL.LowerMachineOperand(&MI, MO))
1114 TC.addOperand(MaybeOperand.getValue());
1115 OutStreamer->EmitInstruction(TC, getSubtargetInfo());
1118 void X86AsmPrinter::EmitXRayTable() {
1122 auto PrevSection = OutStreamer->getCurrentSectionOnly();
1123 auto Fn = MF->getFunction();
1124 MCSection *Section = nullptr;
1125 if (Subtarget->isTargetELF()) {
1126 if (Fn->hasComdat()) {
1127 Section = OutContext.getELFSection("xray_instr_map", ELF::SHT_PROGBITS,
1128 ELF::SHF_ALLOC | ELF::SHF_GROUP, 0,
1129 Fn->getComdat()->getName());
1131 Section = OutContext.getELFSection("xray_instr_map", ELF::SHT_PROGBITS,
1134 } else if (Subtarget->isTargetMachO()) {
1135 Section = OutContext.getMachOSection("__DATA", "xray_instr_map", 0,
1136 SectionKind::getReadOnlyWithRel());
1138 llvm_unreachable("Unsupported target");
1141 // Before we switch over, we force a reference to a label inside the
1142 // xray_instr_map section. Since EmitXRayTable() is always called just
1143 // before the function's end, we assume that this is happening after the
1144 // last return instruction.
1146 // We then align the reference to 16 byte boundaries, which we determined
1147 // experimentally to be beneficial to avoid causing decoder stalls.
1148 MCSymbol *Tmp = OutContext.createTempSymbol("xray_synthetic_", true);
1149 OutStreamer->EmitCodeAlignment(16);
1150 OutStreamer->EmitSymbolValue(Tmp, 8, false);
1151 OutStreamer->SwitchSection(Section);
1152 OutStreamer->EmitLabel(Tmp);
1153 for (const auto &Sled : Sleds) {
1154 OutStreamer->EmitSymbolValue(Sled.Sled, 8);
1155 OutStreamer->EmitSymbolValue(CurrentFnSym, 8);
1156 auto Kind = static_cast<uint8_t>(Sled.Kind);
1157 OutStreamer->EmitBytes(
1158 StringRef(reinterpret_cast<const char *>(&Kind), 1));
1159 OutStreamer->EmitBytes(
1160 StringRef(reinterpret_cast<const char *>(&Sled.AlwaysInstrument), 1));
1161 OutStreamer->EmitZeros(14);
1163 OutStreamer->SwitchSection(PrevSection);
1168 // Returns instruction preceding MBBI in MachineFunction.
1169 // If MBBI is the first instruction of the first basic block, returns null.
1170 static MachineBasicBlock::const_iterator
1171 PrevCrossBBInst(MachineBasicBlock::const_iterator MBBI) {
1172 const MachineBasicBlock *MBB = MBBI->getParent();
1173 while (MBBI == MBB->begin()) {
1174 if (MBB == &MBB->getParent()->front())
1175 return MachineBasicBlock::const_iterator();
1176 MBB = MBB->getPrevNode();
1182 static const Constant *getConstantFromPool(const MachineInstr &MI,
1183 const MachineOperand &Op) {
1187 ArrayRef<MachineConstantPoolEntry> Constants =
1188 MI.getParent()->getParent()->getConstantPool()->getConstants();
1189 const MachineConstantPoolEntry &ConstantEntry =
1190 Constants[Op.getIndex()];
1192 // Bail if this is a machine constant pool entry, we won't be able to dig out
1194 if (ConstantEntry.isMachineConstantPoolEntry())
1197 auto *C = dyn_cast<Constant>(ConstantEntry.Val.ConstVal);
1198 assert((!C || ConstantEntry.getType() == C->getType()) &&
1199 "Expected a constant of the same type!");
1203 static std::string getShuffleComment(const MachineInstr *MI,
1206 ArrayRef<int> Mask) {
1207 std::string Comment;
1209 // Compute the name for a register. This is really goofy because we have
1210 // multiple instruction printers that could (in theory) use different
1211 // names. Fortunately most people use the ATT style (outside of Windows)
1212 // and they actually agree on register naming here. Ultimately, this is
1213 // a comment, and so its OK if it isn't perfect.
1214 auto GetRegisterName = [](unsigned RegNum) -> StringRef {
1215 return X86ATTInstPrinter::getRegisterName(RegNum);
1218 const MachineOperand &DstOp = MI->getOperand(0);
1219 const MachineOperand &SrcOp1 = MI->getOperand(SrcOp1Idx);
1220 const MachineOperand &SrcOp2 = MI->getOperand(SrcOp2Idx);
1222 StringRef DstName = DstOp.isReg() ? GetRegisterName(DstOp.getReg()) : "mem";
1223 StringRef Src1Name =
1224 SrcOp1.isReg() ? GetRegisterName(SrcOp1.getReg()) : "mem";
1225 StringRef Src2Name =
1226 SrcOp2.isReg() ? GetRegisterName(SrcOp2.getReg()) : "mem";
1228 // One source operand, fix the mask to print all elements in one span.
1229 SmallVector<int, 8> ShuffleMask(Mask.begin(), Mask.end());
1230 if (Src1Name == Src2Name)
1231 for (int i = 0, e = ShuffleMask.size(); i != e; ++i)
1232 if (ShuffleMask[i] >= e)
1233 ShuffleMask[i] -= e;
1235 raw_string_ostream CS(Comment);
1238 // Handle AVX512 MASK/MASXZ write mask comments.
1240 // MASKZ: zmmX {%kY} {z}
1241 if (SrcOp1Idx > 1) {
1242 assert((SrcOp1Idx == 2 || SrcOp1Idx == 3) && "Unexpected writemask");
1244 const MachineOperand &WriteMaskOp = MI->getOperand(SrcOp1Idx - 1);
1245 if (WriteMaskOp.isReg()) {
1246 CS << " {%" << GetRegisterName(WriteMaskOp.getReg()) << "}";
1248 if (SrcOp1Idx == 2) {
1256 for (int i = 0, e = ShuffleMask.size(); i != e; ++i) {
1259 if (ShuffleMask[i] == SM_SentinelZero) {
1264 // Otherwise, it must come from src1 or src2. Print the span of elements
1265 // that comes from this src.
1266 bool isSrc1 = ShuffleMask[i] < (int)e;
1267 CS << (isSrc1 ? Src1Name : Src2Name) << '[';
1269 bool IsFirst = true;
1270 while (i != e && ShuffleMask[i] != SM_SentinelZero &&
1271 (ShuffleMask[i] < (int)e) == isSrc1) {
1276 if (ShuffleMask[i] == SM_SentinelUndef)
1279 CS << ShuffleMask[i] % (int)e;
1283 --i; // For loop increments element #.
1290 void X86AsmPrinter::EmitInstruction(const MachineInstr *MI) {
1291 X86MCInstLower MCInstLowering(*MF, *this);
1292 const X86RegisterInfo *RI = MF->getSubtarget<X86Subtarget>().getRegisterInfo();
1294 // Add a comment about EVEX-2-VEX compression for AVX-512 instrs that
1295 // are compressed from EVEX encoding to VEX encoding.
1296 if (TM.Options.MCOptions.ShowMCEncoding) {
1297 if (MI->getAsmPrinterFlags() & AC_EVEX_2_VEX)
1298 OutStreamer->AddComment("EVEX TO VEX Compression ", false);
1301 switch (MI->getOpcode()) {
1302 case TargetOpcode::DBG_VALUE:
1303 llvm_unreachable("Should be handled target independently");
1305 // Emit nothing here but a comment if we can.
1306 case X86::Int_MemBarrier:
1307 OutStreamer->emitRawComment("MEMBARRIER");
1311 case X86::EH_RETURN:
1312 case X86::EH_RETURN64: {
1313 // Lower these as normal, but add some comments.
1314 unsigned Reg = MI->getOperand(0).getReg();
1315 OutStreamer->AddComment(StringRef("eh_return, addr: %") +
1316 X86ATTInstPrinter::getRegisterName(Reg));
1319 case X86::CLEANUPRET: {
1320 // Lower these as normal, but add some comments.
1321 OutStreamer->AddComment("CLEANUPRET");
1325 case X86::CATCHRET: {
1326 // Lower these as normal, but add some comments.
1327 OutStreamer->AddComment("CATCHRET");
1334 case X86::TAILJMPd_CC:
1335 case X86::TAILJMPr64:
1336 case X86::TAILJMPm64:
1337 case X86::TAILJMPd64:
1338 case X86::TAILJMPd64_CC:
1339 case X86::TAILJMPr64_REX:
1340 case X86::TAILJMPm64_REX:
1341 // Lower these as normal, but add some comments.
1342 OutStreamer->AddComment("TAILCALL");
1345 case X86::TLS_addr32:
1346 case X86::TLS_addr64:
1347 case X86::TLS_base_addr32:
1348 case X86::TLS_base_addr64:
1349 return LowerTlsAddr(MCInstLowering, *MI);
1351 case X86::MOVPC32r: {
1352 // This is a pseudo op for a two instruction sequence with a label, which
1359 MCSymbol *PICBase = MF->getPICBaseSymbol();
1360 // FIXME: We would like an efficient form for this, so we don't have to do a
1361 // lot of extra uniquing.
1362 EmitAndCountInstruction(MCInstBuilder(X86::CALLpcrel32)
1363 .addExpr(MCSymbolRefExpr::create(PICBase, OutContext)));
1365 const X86FrameLowering* FrameLowering =
1366 MF->getSubtarget<X86Subtarget>().getFrameLowering();
1367 bool hasFP = FrameLowering->hasFP(*MF);
1369 // TODO: This is needed only if we require precise CFA.
1370 bool HasActiveDwarfFrame = OutStreamer->getNumFrameInfos() &&
1371 !OutStreamer->getDwarfFrameInfos().back().End;
1373 int stackGrowth = -RI->getSlotSize();
1375 if (HasActiveDwarfFrame && !hasFP) {
1376 OutStreamer->EmitCFIAdjustCfaOffset(-stackGrowth);
1380 OutStreamer->EmitLabel(PICBase);
1383 EmitAndCountInstruction(MCInstBuilder(X86::POP32r)
1384 .addReg(MI->getOperand(0).getReg()));
1386 if (HasActiveDwarfFrame && !hasFP) {
1387 OutStreamer->EmitCFIAdjustCfaOffset(stackGrowth);
1392 case X86::ADD32ri: {
1393 // Lower the MO_GOT_ABSOLUTE_ADDRESS form of ADD32ri.
1394 if (MI->getOperand(2).getTargetFlags() != X86II::MO_GOT_ABSOLUTE_ADDRESS)
1397 // Okay, we have something like:
1398 // EAX = ADD32ri EAX, MO_GOT_ABSOLUTE_ADDRESS(@MYGLOBAL)
1400 // For this, we want to print something like:
1401 // MYGLOBAL + (. - PICBASE)
1402 // However, we can't generate a ".", so just emit a new label here and refer
1404 MCSymbol *DotSym = OutContext.createTempSymbol();
1405 OutStreamer->EmitLabel(DotSym);
1407 // Now that we have emitted the label, lower the complex operand expression.
1408 MCSymbol *OpSym = MCInstLowering.GetSymbolFromOperand(MI->getOperand(2));
1410 const MCExpr *DotExpr = MCSymbolRefExpr::create(DotSym, OutContext);
1411 const MCExpr *PICBase =
1412 MCSymbolRefExpr::create(MF->getPICBaseSymbol(), OutContext);
1413 DotExpr = MCBinaryExpr::createSub(DotExpr, PICBase, OutContext);
1415 DotExpr = MCBinaryExpr::createAdd(MCSymbolRefExpr::create(OpSym,OutContext),
1416 DotExpr, OutContext);
1418 EmitAndCountInstruction(MCInstBuilder(X86::ADD32ri)
1419 .addReg(MI->getOperand(0).getReg())
1420 .addReg(MI->getOperand(1).getReg())
1424 case TargetOpcode::STATEPOINT:
1425 return LowerSTATEPOINT(*MI, MCInstLowering);
1427 case TargetOpcode::FAULTING_LOAD_OP:
1428 return LowerFAULTING_LOAD_OP(*MI, MCInstLowering);
1430 case TargetOpcode::PATCHABLE_OP:
1431 return LowerPATCHABLE_OP(*MI, MCInstLowering);
1433 case TargetOpcode::STACKMAP:
1434 return LowerSTACKMAP(*MI);
1436 case TargetOpcode::PATCHPOINT:
1437 return LowerPATCHPOINT(*MI, MCInstLowering);
1439 case TargetOpcode::PATCHABLE_FUNCTION_ENTER:
1440 return LowerPATCHABLE_FUNCTION_ENTER(*MI, MCInstLowering);
1442 case TargetOpcode::PATCHABLE_RET:
1443 return LowerPATCHABLE_RET(*MI, MCInstLowering);
1445 case TargetOpcode::PATCHABLE_TAIL_CALL:
1446 return LowerPATCHABLE_TAIL_CALL(*MI, MCInstLowering);
1448 case X86::MORESTACK_RET:
1449 EmitAndCountInstruction(MCInstBuilder(getRetOpcode(*Subtarget)));
1452 case X86::MORESTACK_RET_RESTORE_R10:
1453 // Return, then restore R10.
1454 EmitAndCountInstruction(MCInstBuilder(getRetOpcode(*Subtarget)));
1455 EmitAndCountInstruction(MCInstBuilder(X86::MOV64rr)
1460 case X86::SEH_PushReg:
1461 assert(MF->hasWinCFI() && "SEH_ instruction in function without WinCFI?");
1462 OutStreamer->EmitWinCFIPushReg(RI->getSEHRegNum(MI->getOperand(0).getImm()));
1465 case X86::SEH_SaveReg:
1466 assert(MF->hasWinCFI() && "SEH_ instruction in function without WinCFI?");
1467 OutStreamer->EmitWinCFISaveReg(RI->getSEHRegNum(MI->getOperand(0).getImm()),
1468 MI->getOperand(1).getImm());
1471 case X86::SEH_SaveXMM:
1472 assert(MF->hasWinCFI() && "SEH_ instruction in function without WinCFI?");
1473 OutStreamer->EmitWinCFISaveXMM(RI->getSEHRegNum(MI->getOperand(0).getImm()),
1474 MI->getOperand(1).getImm());
1477 case X86::SEH_StackAlloc:
1478 assert(MF->hasWinCFI() && "SEH_ instruction in function without WinCFI?");
1479 OutStreamer->EmitWinCFIAllocStack(MI->getOperand(0).getImm());
1482 case X86::SEH_SetFrame:
1483 assert(MF->hasWinCFI() && "SEH_ instruction in function without WinCFI?");
1484 OutStreamer->EmitWinCFISetFrame(RI->getSEHRegNum(MI->getOperand(0).getImm()),
1485 MI->getOperand(1).getImm());
1488 case X86::SEH_PushFrame:
1489 assert(MF->hasWinCFI() && "SEH_ instruction in function without WinCFI?");
1490 OutStreamer->EmitWinCFIPushFrame(MI->getOperand(0).getImm());
1493 case X86::SEH_EndPrologue:
1494 assert(MF->hasWinCFI() && "SEH_ instruction in function without WinCFI?");
1495 OutStreamer->EmitWinCFIEndProlog();
1498 case X86::SEH_Epilogue: {
1499 assert(MF->hasWinCFI() && "SEH_ instruction in function without WinCFI?");
1500 MachineBasicBlock::const_iterator MBBI(MI);
1501 // Check if preceded by a call and emit nop if so.
1502 for (MBBI = PrevCrossBBInst(MBBI);
1503 MBBI != MachineBasicBlock::const_iterator();
1504 MBBI = PrevCrossBBInst(MBBI)) {
1505 // Conservatively assume that pseudo instructions don't emit code and keep
1506 // looking for a call. We may emit an unnecessary nop in some cases.
1507 if (!MBBI->isPseudo()) {
1509 EmitAndCountInstruction(MCInstBuilder(X86::NOOP));
1516 // Lower PSHUFB and VPERMILP normally but add a comment if we can find
1517 // a constant shuffle mask. We won't be able to do this at the MC layer
1518 // because the mask isn't an immediate.
1520 case X86::VPSHUFBrm:
1521 case X86::VPSHUFBYrm:
1522 case X86::VPSHUFBZ128rm:
1523 case X86::VPSHUFBZ128rmk:
1524 case X86::VPSHUFBZ128rmkz:
1525 case X86::VPSHUFBZ256rm:
1526 case X86::VPSHUFBZ256rmk:
1527 case X86::VPSHUFBZ256rmkz:
1528 case X86::VPSHUFBZrm:
1529 case X86::VPSHUFBZrmk:
1530 case X86::VPSHUFBZrmkz: {
1531 if (!OutStreamer->isVerboseAsm())
1533 unsigned SrcIdx, MaskIdx;
1534 switch (MI->getOpcode()) {
1535 default: llvm_unreachable("Invalid opcode");
1537 case X86::VPSHUFBrm:
1538 case X86::VPSHUFBYrm:
1539 case X86::VPSHUFBZ128rm:
1540 case X86::VPSHUFBZ256rm:
1541 case X86::VPSHUFBZrm:
1542 SrcIdx = 1; MaskIdx = 5; break;
1543 case X86::VPSHUFBZ128rmkz:
1544 case X86::VPSHUFBZ256rmkz:
1545 case X86::VPSHUFBZrmkz:
1546 SrcIdx = 2; MaskIdx = 6; break;
1547 case X86::VPSHUFBZ128rmk:
1548 case X86::VPSHUFBZ256rmk:
1549 case X86::VPSHUFBZrmk:
1550 SrcIdx = 3; MaskIdx = 7; break;
1553 assert(MI->getNumOperands() >= 6 &&
1554 "We should always have at least 6 operands!");
1556 const MachineOperand &MaskOp = MI->getOperand(MaskIdx);
1557 if (auto *C = getConstantFromPool(*MI, MaskOp)) {
1558 SmallVector<int, 64> Mask;
1559 DecodePSHUFBMask(C, Mask);
1561 OutStreamer->AddComment(getShuffleComment(MI, SrcIdx, SrcIdx, Mask));
1566 case X86::VPERMILPSrm:
1567 case X86::VPERMILPSYrm:
1568 case X86::VPERMILPSZ128rm:
1569 case X86::VPERMILPSZ128rmk:
1570 case X86::VPERMILPSZ128rmkz:
1571 case X86::VPERMILPSZ256rm:
1572 case X86::VPERMILPSZ256rmk:
1573 case X86::VPERMILPSZ256rmkz:
1574 case X86::VPERMILPSZrm:
1575 case X86::VPERMILPSZrmk:
1576 case X86::VPERMILPSZrmkz:
1577 case X86::VPERMILPDrm:
1578 case X86::VPERMILPDYrm:
1579 case X86::VPERMILPDZ128rm:
1580 case X86::VPERMILPDZ128rmk:
1581 case X86::VPERMILPDZ128rmkz:
1582 case X86::VPERMILPDZ256rm:
1583 case X86::VPERMILPDZ256rmk:
1584 case X86::VPERMILPDZ256rmkz:
1585 case X86::VPERMILPDZrm:
1586 case X86::VPERMILPDZrmk:
1587 case X86::VPERMILPDZrmkz: {
1588 if (!OutStreamer->isVerboseAsm())
1590 unsigned SrcIdx, MaskIdx;
1592 switch (MI->getOpcode()) {
1593 default: llvm_unreachable("Invalid opcode");
1594 case X86::VPERMILPSrm:
1595 case X86::VPERMILPSYrm:
1596 case X86::VPERMILPSZ128rm:
1597 case X86::VPERMILPSZ256rm:
1598 case X86::VPERMILPSZrm:
1599 SrcIdx = 1; MaskIdx = 5; ElSize = 32; break;
1600 case X86::VPERMILPSZ128rmkz:
1601 case X86::VPERMILPSZ256rmkz:
1602 case X86::VPERMILPSZrmkz:
1603 SrcIdx = 2; MaskIdx = 6; ElSize = 32; break;
1604 case X86::VPERMILPSZ128rmk:
1605 case X86::VPERMILPSZ256rmk:
1606 case X86::VPERMILPSZrmk:
1607 SrcIdx = 3; MaskIdx = 7; ElSize = 32; break;
1608 case X86::VPERMILPDrm:
1609 case X86::VPERMILPDYrm:
1610 case X86::VPERMILPDZ128rm:
1611 case X86::VPERMILPDZ256rm:
1612 case X86::VPERMILPDZrm:
1613 SrcIdx = 1; MaskIdx = 5; ElSize = 64; break;
1614 case X86::VPERMILPDZ128rmkz:
1615 case X86::VPERMILPDZ256rmkz:
1616 case X86::VPERMILPDZrmkz:
1617 SrcIdx = 2; MaskIdx = 6; ElSize = 64; break;
1618 case X86::VPERMILPDZ128rmk:
1619 case X86::VPERMILPDZ256rmk:
1620 case X86::VPERMILPDZrmk:
1621 SrcIdx = 3; MaskIdx = 7; ElSize = 64; break;
1624 assert(MI->getNumOperands() >= 6 &&
1625 "We should always have at least 6 operands!");
1627 const MachineOperand &MaskOp = MI->getOperand(MaskIdx);
1628 if (auto *C = getConstantFromPool(*MI, MaskOp)) {
1629 SmallVector<int, 16> Mask;
1630 DecodeVPERMILPMask(C, ElSize, Mask);
1632 OutStreamer->AddComment(getShuffleComment(MI, SrcIdx, SrcIdx, Mask));
1637 case X86::VPERMIL2PDrm:
1638 case X86::VPERMIL2PSrm:
1639 case X86::VPERMIL2PDrmY:
1640 case X86::VPERMIL2PSrmY: {
1641 if (!OutStreamer->isVerboseAsm())
1643 assert(MI->getNumOperands() >= 8 &&
1644 "We should always have at least 8 operands!");
1646 const MachineOperand &CtrlOp = MI->getOperand(MI->getNumOperands() - 1);
1647 if (!CtrlOp.isImm())
1651 switch (MI->getOpcode()) {
1652 default: llvm_unreachable("Invalid opcode");
1653 case X86::VPERMIL2PSrm: case X86::VPERMIL2PSrmY: ElSize = 32; break;
1654 case X86::VPERMIL2PDrm: case X86::VPERMIL2PDrmY: ElSize = 64; break;
1657 const MachineOperand &MaskOp = MI->getOperand(6);
1658 if (auto *C = getConstantFromPool(*MI, MaskOp)) {
1659 SmallVector<int, 16> Mask;
1660 DecodeVPERMIL2PMask(C, (unsigned)CtrlOp.getImm(), ElSize, Mask);
1662 OutStreamer->AddComment(getShuffleComment(MI, 1, 2, Mask));
1667 case X86::VPPERMrrm: {
1668 if (!OutStreamer->isVerboseAsm())
1670 assert(MI->getNumOperands() >= 7 &&
1671 "We should always have at least 7 operands!");
1673 const MachineOperand &MaskOp = MI->getOperand(6);
1674 if (auto *C = getConstantFromPool(*MI, MaskOp)) {
1675 SmallVector<int, 16> Mask;
1676 DecodeVPPERMMask(C, Mask);
1678 OutStreamer->AddComment(getShuffleComment(MI, 1, 2, Mask));
1683 #define MOV_CASE(Prefix, Suffix) \
1684 case X86::Prefix##MOVAPD##Suffix##rm: \
1685 case X86::Prefix##MOVAPS##Suffix##rm: \
1686 case X86::Prefix##MOVUPD##Suffix##rm: \
1687 case X86::Prefix##MOVUPS##Suffix##rm: \
1688 case X86::Prefix##MOVDQA##Suffix##rm: \
1689 case X86::Prefix##MOVDQU##Suffix##rm:
1691 #define MOV_AVX512_CASE(Suffix) \
1692 case X86::VMOVDQA64##Suffix##rm: \
1693 case X86::VMOVDQA32##Suffix##rm: \
1694 case X86::VMOVDQU64##Suffix##rm: \
1695 case X86::VMOVDQU32##Suffix##rm: \
1696 case X86::VMOVDQU16##Suffix##rm: \
1697 case X86::VMOVDQU8##Suffix##rm: \
1698 case X86::VMOVAPS##Suffix##rm: \
1699 case X86::VMOVAPD##Suffix##rm: \
1700 case X86::VMOVUPS##Suffix##rm: \
1701 case X86::VMOVUPD##Suffix##rm:
1703 #define CASE_ALL_MOV_RM() \
1704 MOV_CASE(, ) /* SSE */ \
1705 MOV_CASE(V, ) /* AVX-128 */ \
1706 MOV_CASE(V, Y) /* AVX-256 */ \
1707 MOV_AVX512_CASE(Z) \
1708 MOV_AVX512_CASE(Z256) \
1709 MOV_AVX512_CASE(Z128)
1711 // For loads from a constant pool to a vector register, print the constant
1714 if (!OutStreamer->isVerboseAsm())
1716 if (MI->getNumOperands() <= 4)
1718 if (auto *C = getConstantFromPool(*MI, MI->getOperand(4))) {
1719 std::string Comment;
1720 raw_string_ostream CS(Comment);
1721 const MachineOperand &DstOp = MI->getOperand(0);
1722 CS << X86ATTInstPrinter::getRegisterName(DstOp.getReg()) << " = ";
1723 if (auto *CDS = dyn_cast<ConstantDataSequential>(C)) {
1725 for (int i = 0, NumElements = CDS->getNumElements(); i < NumElements; ++i) {
1728 if (CDS->getElementType()->isIntegerTy())
1729 CS << CDS->getElementAsInteger(i);
1730 else if (CDS->getElementType()->isFloatTy())
1731 CS << CDS->getElementAsFloat(i);
1732 else if (CDS->getElementType()->isDoubleTy())
1733 CS << CDS->getElementAsDouble(i);
1738 OutStreamer->AddComment(CS.str());
1739 } else if (auto *CV = dyn_cast<ConstantVector>(C)) {
1741 for (int i = 0, NumOperands = CV->getNumOperands(); i < NumOperands; ++i) {
1744 Constant *COp = CV->getOperand(i);
1745 if (isa<UndefValue>(COp)) {
1747 } else if (auto *CI = dyn_cast<ConstantInt>(COp)) {
1748 if (CI->getBitWidth() <= 64) {
1749 CS << CI->getZExtValue();
1751 // print multi-word constant as (w0,w1)
1752 const auto &Val = CI->getValue();
1754 for (int i = 0, N = Val.getNumWords(); i < N; ++i) {
1757 CS << Val.getRawData()[i];
1761 } else if (auto *CF = dyn_cast<ConstantFP>(COp)) {
1762 SmallString<32> Str;
1763 CF->getValueAPF().toString(Str);
1770 OutStreamer->AddComment(CS.str());
1777 MCInstLowering.Lower(MI, TmpInst);
1779 // Stackmap shadows cannot include branch targets, so we can count the bytes
1780 // in a call towards the shadow, but must ensure that the no thread returns
1781 // in to the stackmap shadow. The only way to achieve this is if the call
1782 // is at the end of the shadow.
1784 // Count then size of the call towards the shadow
1785 SMShadowTracker.count(TmpInst, getSubtargetInfo(), CodeEmitter.get());
1786 // Then flush the shadow so that we fill with nops before the call, not
1788 SMShadowTracker.emitShadowPadding(*OutStreamer, getSubtargetInfo());
1789 // Then emit the call
1790 OutStreamer->EmitInstruction(TmpInst, getSubtargetInfo());
1794 EmitAndCountInstruction(TmpInst);