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 "InstPrinter/X86ATTInstPrinter.h"
16 #include "InstPrinter/X86InstComments.h"
17 #include "MCTargetDesc/X86BaseInfo.h"
18 #include "MCTargetDesc/X86TargetStreamer.h"
19 #include "Utils/X86ShuffleDecode.h"
20 #include "X86AsmPrinter.h"
21 #include "X86RegisterInfo.h"
22 #include "X86ShuffleDecodeConstantPool.h"
23 #include "llvm/ADT/Optional.h"
24 #include "llvm/ADT/SmallString.h"
25 #include "llvm/ADT/iterator_range.h"
26 #include "llvm/CodeGen/MachineConstantPool.h"
27 #include "llvm/CodeGen/MachineFunction.h"
28 #include "llvm/CodeGen/MachineModuleInfoImpls.h"
29 #include "llvm/CodeGen/MachineOperand.h"
30 #include "llvm/CodeGen/StackMaps.h"
31 #include "llvm/CodeGen/TargetLoweringObjectFile.h"
32 #include "llvm/IR/DataLayout.h"
33 #include "llvm/IR/GlobalValue.h"
34 #include "llvm/IR/Mangler.h"
35 #include "llvm/MC/MCAsmInfo.h"
36 #include "llvm/MC/MCCodeEmitter.h"
37 #include "llvm/MC/MCContext.h"
38 #include "llvm/MC/MCExpr.h"
39 #include "llvm/MC/MCFixup.h"
40 #include "llvm/MC/MCInst.h"
41 #include "llvm/MC/MCInstBuilder.h"
42 #include "llvm/MC/MCSection.h"
43 #include "llvm/MC/MCSectionELF.h"
44 #include "llvm/MC/MCStreamer.h"
45 #include "llvm/MC/MCSymbol.h"
46 #include "llvm/MC/MCSymbolELF.h"
52 /// X86MCInstLower - This class is used to lower an MachineInstr into an MCInst.
53 class X86MCInstLower {
55 const MachineFunction &MF;
56 const TargetMachine &TM;
58 X86AsmPrinter &AsmPrinter;
60 X86MCInstLower(const MachineFunction &MF, X86AsmPrinter &asmprinter);
62 Optional<MCOperand> LowerMachineOperand(const MachineInstr *MI,
63 const MachineOperand &MO) const;
64 void Lower(const MachineInstr *MI, MCInst &OutMI) const;
66 MCSymbol *GetSymbolFromOperand(const MachineOperand &MO) const;
67 MCOperand LowerSymbolOperand(const MachineOperand &MO, MCSymbol *Sym) const;
70 MachineModuleInfoMachO &getMachOMMI() const;
73 } // end anonymous namespace
75 // Emit a minimal sequence of nops spanning NumBytes bytes.
76 static void EmitNops(MCStreamer &OS, unsigned NumBytes, bool Is64Bit,
77 const MCSubtargetInfo &STI);
79 void X86AsmPrinter::StackMapShadowTracker::count(MCInst &Inst,
80 const MCSubtargetInfo &STI,
81 MCCodeEmitter *CodeEmitter) {
83 SmallString<256> Code;
84 SmallVector<MCFixup, 4> Fixups;
85 raw_svector_ostream VecOS(Code);
86 CodeEmitter->encodeInstruction(Inst, VecOS, Fixups, STI);
87 CurrentShadowSize += Code.size();
88 if (CurrentShadowSize >= RequiredShadowSize)
89 InShadow = false; // The shadow is big enough. Stop counting.
93 void X86AsmPrinter::StackMapShadowTracker::emitShadowPadding(
94 MCStreamer &OutStreamer, const MCSubtargetInfo &STI) {
95 if (InShadow && CurrentShadowSize < RequiredShadowSize) {
97 EmitNops(OutStreamer, RequiredShadowSize - CurrentShadowSize,
98 MF->getSubtarget<X86Subtarget>().is64Bit(), STI);
102 void X86AsmPrinter::EmitAndCountInstruction(MCInst &Inst) {
103 OutStreamer->EmitInstruction(Inst, getSubtargetInfo(),
104 EnablePrintSchedInfo &&
105 !(Inst.getFlags() & X86::NO_SCHED_INFO));
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_ABS8: RefKind = MCSymbolRefExpr::VK_X86_ABS8; break;
219 case X86II::MO_PIC_BASE_OFFSET:
220 case X86II::MO_DARWIN_NONLAZY_PIC_BASE:
221 Expr = MCSymbolRefExpr::create(Sym, Ctx);
222 // Subtract the pic base.
223 Expr = MCBinaryExpr::createSub(Expr,
224 MCSymbolRefExpr::create(MF.getPICBaseSymbol(), Ctx),
227 assert(MAI.doesSetDirectiveSuppressReloc());
228 // If .set directive is supported, use it to reduce the number of
229 // relocations the assembler will generate for differences between
230 // local labels. This is only safe when the symbols are in the same
231 // section so we are restricting it to jumptable references.
232 MCSymbol *Label = Ctx.createTempSymbol();
233 AsmPrinter.OutStreamer->EmitAssignment(Label, Expr);
234 Expr = MCSymbolRefExpr::create(Label, Ctx);
240 Expr = MCSymbolRefExpr::create(Sym, RefKind, Ctx);
242 if (!MO.isJTI() && !MO.isMBB() && MO.getOffset())
243 Expr = MCBinaryExpr::createAdd(Expr,
244 MCConstantExpr::create(MO.getOffset(), Ctx),
246 return MCOperand::createExpr(Expr);
250 /// \brief Simplify FOO $imm, %{al,ax,eax,rax} to FOO $imm, for instruction with
251 /// a short fixed-register form.
252 static void SimplifyShortImmForm(MCInst &Inst, unsigned Opcode) {
253 unsigned ImmOp = Inst.getNumOperands() - 1;
254 assert(Inst.getOperand(0).isReg() &&
255 (Inst.getOperand(ImmOp).isImm() || Inst.getOperand(ImmOp).isExpr()) &&
256 ((Inst.getNumOperands() == 3 && Inst.getOperand(1).isReg() &&
257 Inst.getOperand(0).getReg() == Inst.getOperand(1).getReg()) ||
258 Inst.getNumOperands() == 2) && "Unexpected instruction!");
260 // Check whether the destination register can be fixed.
261 unsigned Reg = Inst.getOperand(0).getReg();
262 if (Reg != X86::AL && Reg != X86::AX && Reg != X86::EAX && Reg != X86::RAX)
265 // If so, rewrite the instruction.
266 MCOperand Saved = Inst.getOperand(ImmOp);
268 Inst.setOpcode(Opcode);
269 Inst.addOperand(Saved);
272 /// \brief If a movsx instruction has a shorter encoding for the used register
273 /// simplify the instruction to use it instead.
274 static void SimplifyMOVSX(MCInst &Inst) {
275 unsigned NewOpcode = 0;
276 unsigned Op0 = Inst.getOperand(0).getReg(), Op1 = Inst.getOperand(1).getReg();
277 switch (Inst.getOpcode()) {
279 llvm_unreachable("Unexpected instruction!");
280 case X86::MOVSX16rr8: // movsbw %al, %ax --> cbtw
281 if (Op0 == X86::AX && Op1 == X86::AL)
282 NewOpcode = X86::CBW;
284 case X86::MOVSX32rr16: // movswl %ax, %eax --> cwtl
285 if (Op0 == X86::EAX && Op1 == X86::AX)
286 NewOpcode = X86::CWDE;
288 case X86::MOVSX64rr32: // movslq %eax, %rax --> cltq
289 if (Op0 == X86::RAX && Op1 == X86::EAX)
290 NewOpcode = X86::CDQE;
294 if (NewOpcode != 0) {
296 Inst.setOpcode(NewOpcode);
300 /// \brief Simplify things like MOV32rm to MOV32o32a.
301 static void SimplifyShortMoveForm(X86AsmPrinter &Printer, MCInst &Inst,
303 // Don't make these simplifications in 64-bit mode; other assemblers don't
304 // perform them because they make the code larger.
305 if (Printer.getSubtarget().is64Bit())
308 bool IsStore = Inst.getOperand(0).isReg() && Inst.getOperand(1).isReg();
309 unsigned AddrBase = IsStore;
310 unsigned RegOp = IsStore ? 0 : 5;
311 unsigned AddrOp = AddrBase + 3;
312 assert(Inst.getNumOperands() == 6 && Inst.getOperand(RegOp).isReg() &&
313 Inst.getOperand(AddrBase + X86::AddrBaseReg).isReg() &&
314 Inst.getOperand(AddrBase + X86::AddrScaleAmt).isImm() &&
315 Inst.getOperand(AddrBase + X86::AddrIndexReg).isReg() &&
316 Inst.getOperand(AddrBase + X86::AddrSegmentReg).isReg() &&
317 (Inst.getOperand(AddrOp).isExpr() ||
318 Inst.getOperand(AddrOp).isImm()) &&
319 "Unexpected instruction!");
321 // Check whether the destination register can be fixed.
322 unsigned Reg = Inst.getOperand(RegOp).getReg();
323 if (Reg != X86::AL && Reg != X86::AX && Reg != X86::EAX && Reg != X86::RAX)
326 // Check whether this is an absolute address.
327 // FIXME: We know TLVP symbol refs aren't, but there should be a better way
329 bool Absolute = true;
330 if (Inst.getOperand(AddrOp).isExpr()) {
331 const MCExpr *MCE = Inst.getOperand(AddrOp).getExpr();
332 if (const MCSymbolRefExpr *SRE = dyn_cast<MCSymbolRefExpr>(MCE))
333 if (SRE->getKind() == MCSymbolRefExpr::VK_TLVP)
338 (Inst.getOperand(AddrBase + X86::AddrBaseReg).getReg() != 0 ||
339 Inst.getOperand(AddrBase + X86::AddrScaleAmt).getImm() != 1 ||
340 Inst.getOperand(AddrBase + X86::AddrIndexReg).getReg() != 0))
343 // If so, rewrite the instruction.
344 MCOperand Saved = Inst.getOperand(AddrOp);
345 MCOperand Seg = Inst.getOperand(AddrBase + X86::AddrSegmentReg);
347 Inst.setOpcode(Opcode);
348 Inst.addOperand(Saved);
349 Inst.addOperand(Seg);
352 static unsigned getRetOpcode(const X86Subtarget &Subtarget) {
353 return Subtarget.is64Bit() ? X86::RETQ : X86::RETL;
357 X86MCInstLower::LowerMachineOperand(const MachineInstr *MI,
358 const MachineOperand &MO) const {
359 switch (MO.getType()) {
362 llvm_unreachable("unknown operand type");
363 case MachineOperand::MO_Register:
364 // Ignore all implicit register operands.
367 return MCOperand::createReg(MO.getReg());
368 case MachineOperand::MO_Immediate:
369 return MCOperand::createImm(MO.getImm());
370 case MachineOperand::MO_MachineBasicBlock:
371 case MachineOperand::MO_GlobalAddress:
372 case MachineOperand::MO_ExternalSymbol:
373 return LowerSymbolOperand(MO, GetSymbolFromOperand(MO));
374 case MachineOperand::MO_MCSymbol:
375 return LowerSymbolOperand(MO, MO.getMCSymbol());
376 case MachineOperand::MO_JumpTableIndex:
377 return LowerSymbolOperand(MO, AsmPrinter.GetJTISymbol(MO.getIndex()));
378 case MachineOperand::MO_ConstantPoolIndex:
379 return LowerSymbolOperand(MO, AsmPrinter.GetCPISymbol(MO.getIndex()));
380 case MachineOperand::MO_BlockAddress:
381 return LowerSymbolOperand(
382 MO, AsmPrinter.GetBlockAddressSymbol(MO.getBlockAddress()));
383 case MachineOperand::MO_RegisterMask:
384 // Ignore call clobbers.
389 void X86MCInstLower::Lower(const MachineInstr *MI, MCInst &OutMI) const {
390 OutMI.setOpcode(MI->getOpcode());
392 for (const MachineOperand &MO : MI->operands())
393 if (auto MaybeMCOp = LowerMachineOperand(MI, MO))
394 OutMI.addOperand(MaybeMCOp.getValue());
396 // Handle a few special cases to eliminate operand modifiers.
398 switch (OutMI.getOpcode()) {
403 // LEA should have a segment register, but it must be empty.
404 assert(OutMI.getNumOperands() == 1+X86::AddrNumOperands &&
405 "Unexpected # of LEA operands");
406 assert(OutMI.getOperand(1+X86::AddrSegmentReg).getReg() == 0 &&
407 "LEA has segment specified!");
410 // Commute operands to get a smaller encoding by using VEX.R instead of VEX.B
411 // if one of the registers is extended, but other isn't.
412 case X86::VMOVZPQILo2PQIrr:
414 case X86::VMOVAPDYrr:
416 case X86::VMOVAPSYrr:
418 case X86::VMOVDQAYrr:
420 case X86::VMOVDQUYrr:
422 case X86::VMOVUPDYrr:
424 case X86::VMOVUPSYrr: {
425 if (!X86II::isX86_64ExtendedReg(OutMI.getOperand(0).getReg()) &&
426 X86II::isX86_64ExtendedReg(OutMI.getOperand(1).getReg())) {
428 switch (OutMI.getOpcode()) {
429 default: llvm_unreachable("Invalid opcode");
430 case X86::VMOVZPQILo2PQIrr: NewOpc = X86::VMOVPQI2QIrr; break;
431 case X86::VMOVAPDrr: NewOpc = X86::VMOVAPDrr_REV; break;
432 case X86::VMOVAPDYrr: NewOpc = X86::VMOVAPDYrr_REV; break;
433 case X86::VMOVAPSrr: NewOpc = X86::VMOVAPSrr_REV; break;
434 case X86::VMOVAPSYrr: NewOpc = X86::VMOVAPSYrr_REV; break;
435 case X86::VMOVDQArr: NewOpc = X86::VMOVDQArr_REV; break;
436 case X86::VMOVDQAYrr: NewOpc = X86::VMOVDQAYrr_REV; break;
437 case X86::VMOVDQUrr: NewOpc = X86::VMOVDQUrr_REV; break;
438 case X86::VMOVDQUYrr: NewOpc = X86::VMOVDQUYrr_REV; break;
439 case X86::VMOVUPDrr: NewOpc = X86::VMOVUPDrr_REV; break;
440 case X86::VMOVUPDYrr: NewOpc = X86::VMOVUPDYrr_REV; break;
441 case X86::VMOVUPSrr: NewOpc = X86::VMOVUPSrr_REV; break;
442 case X86::VMOVUPSYrr: NewOpc = X86::VMOVUPSYrr_REV; break;
444 OutMI.setOpcode(NewOpc);
449 case X86::VMOVSSrr: {
450 if (!X86II::isX86_64ExtendedReg(OutMI.getOperand(0).getReg()) &&
451 X86II::isX86_64ExtendedReg(OutMI.getOperand(2).getReg())) {
453 switch (OutMI.getOpcode()) {
454 default: llvm_unreachable("Invalid opcode");
455 case X86::VMOVSDrr: NewOpc = X86::VMOVSDrr_REV; break;
456 case X86::VMOVSSrr: NewOpc = X86::VMOVSSrr_REV; break;
458 OutMI.setOpcode(NewOpc);
463 // TAILJMPr64, CALL64r, CALL64pcrel32 - These instructions have register
464 // inputs modeled as normal uses instead of implicit uses. As such, truncate
465 // off all but the first operand (the callee). FIXME: Change isel.
466 case X86::TAILJMPr64:
467 case X86::TAILJMPr64_REX:
469 case X86::CALL64pcrel32: {
470 unsigned Opcode = OutMI.getOpcode();
471 MCOperand Saved = OutMI.getOperand(0);
473 OutMI.setOpcode(Opcode);
474 OutMI.addOperand(Saved);
479 case X86::EH_RETURN64: {
481 OutMI.setOpcode(getRetOpcode(AsmPrinter.getSubtarget()));
485 case X86::CLEANUPRET: {
486 // Replace CATCHRET with the appropriate RET.
488 OutMI.setOpcode(getRetOpcode(AsmPrinter.getSubtarget()));
492 case X86::CATCHRET: {
493 // Replace CATCHRET with the appropriate RET.
494 const X86Subtarget &Subtarget = AsmPrinter.getSubtarget();
495 unsigned ReturnReg = Subtarget.is64Bit() ? X86::RAX : X86::EAX;
497 OutMI.setOpcode(getRetOpcode(Subtarget));
498 OutMI.addOperand(MCOperand::createReg(ReturnReg));
502 // TAILJMPd, TAILJMPd64, TailJMPd_cc - Lower to the correct jump instruction.
504 case X86::TAILJMPr: Opcode = X86::JMP32r; goto SetTailJmpOpcode;
506 case X86::TAILJMPd64: Opcode = X86::JMP_1; goto SetTailJmpOpcode;
507 case X86::TAILJMPd_CC:
508 case X86::TAILJMPd64_CC:
509 Opcode = X86::GetCondBranchFromCond(
510 static_cast<X86::CondCode>(MI->getOperand(1).getImm()));
511 goto SetTailJmpOpcode;
514 MCOperand Saved = OutMI.getOperand(0);
516 OutMI.setOpcode(Opcode);
517 OutMI.addOperand(Saved);
525 // If we aren't in 64-bit mode we can use the 1-byte inc/dec instructions.
526 if (!AsmPrinter.getSubtarget().is64Bit()) {
528 switch (OutMI.getOpcode()) {
529 default: llvm_unreachable("Invalid opcode");
530 case X86::DEC16r: Opcode = X86::DEC16r_alt; break;
531 case X86::DEC32r: Opcode = X86::DEC32r_alt; break;
532 case X86::INC16r: Opcode = X86::INC16r_alt; break;
533 case X86::INC32r: Opcode = X86::INC32r_alt; break;
535 OutMI.setOpcode(Opcode);
539 // These are pseudo-ops for OR to help with the OR->ADD transformation. We do
540 // this with an ugly goto in case the resultant OR uses EAX and needs the
542 case X86::ADD16rr_DB: OutMI.setOpcode(X86::OR16rr); goto ReSimplify;
543 case X86::ADD32rr_DB: OutMI.setOpcode(X86::OR32rr); goto ReSimplify;
544 case X86::ADD64rr_DB: OutMI.setOpcode(X86::OR64rr); goto ReSimplify;
545 case X86::ADD16ri_DB: OutMI.setOpcode(X86::OR16ri); goto ReSimplify;
546 case X86::ADD32ri_DB: OutMI.setOpcode(X86::OR32ri); goto ReSimplify;
547 case X86::ADD64ri32_DB: OutMI.setOpcode(X86::OR64ri32); goto ReSimplify;
548 case X86::ADD16ri8_DB: OutMI.setOpcode(X86::OR16ri8); goto ReSimplify;
549 case X86::ADD32ri8_DB: OutMI.setOpcode(X86::OR32ri8); goto ReSimplify;
550 case X86::ADD64ri8_DB: OutMI.setOpcode(X86::OR64ri8); goto ReSimplify;
552 // Atomic load and store require a separate pseudo-inst because Acquire
553 // implies mayStore and Release implies mayLoad; fix these to regular MOV
555 case X86::ACQUIRE_MOV8rm: OutMI.setOpcode(X86::MOV8rm); goto ReSimplify;
556 case X86::ACQUIRE_MOV16rm: OutMI.setOpcode(X86::MOV16rm); goto ReSimplify;
557 case X86::ACQUIRE_MOV32rm: OutMI.setOpcode(X86::MOV32rm); goto ReSimplify;
558 case X86::ACQUIRE_MOV64rm: OutMI.setOpcode(X86::MOV64rm); goto ReSimplify;
559 case X86::RELEASE_MOV8mr: OutMI.setOpcode(X86::MOV8mr); goto ReSimplify;
560 case X86::RELEASE_MOV16mr: OutMI.setOpcode(X86::MOV16mr); goto ReSimplify;
561 case X86::RELEASE_MOV32mr: OutMI.setOpcode(X86::MOV32mr); goto ReSimplify;
562 case X86::RELEASE_MOV64mr: OutMI.setOpcode(X86::MOV64mr); goto ReSimplify;
563 case X86::RELEASE_MOV8mi: OutMI.setOpcode(X86::MOV8mi); goto ReSimplify;
564 case X86::RELEASE_MOV16mi: OutMI.setOpcode(X86::MOV16mi); goto ReSimplify;
565 case X86::RELEASE_MOV32mi: OutMI.setOpcode(X86::MOV32mi); goto ReSimplify;
566 case X86::RELEASE_MOV64mi32: OutMI.setOpcode(X86::MOV64mi32); goto ReSimplify;
567 case X86::RELEASE_ADD8mi: OutMI.setOpcode(X86::ADD8mi); goto ReSimplify;
568 case X86::RELEASE_ADD8mr: OutMI.setOpcode(X86::ADD8mr); goto ReSimplify;
569 case X86::RELEASE_ADD32mi: OutMI.setOpcode(X86::ADD32mi); goto ReSimplify;
570 case X86::RELEASE_ADD32mr: OutMI.setOpcode(X86::ADD32mr); goto ReSimplify;
571 case X86::RELEASE_ADD64mi32: OutMI.setOpcode(X86::ADD64mi32); goto ReSimplify;
572 case X86::RELEASE_ADD64mr: OutMI.setOpcode(X86::ADD64mr); goto ReSimplify;
573 case X86::RELEASE_AND8mi: OutMI.setOpcode(X86::AND8mi); goto ReSimplify;
574 case X86::RELEASE_AND8mr: OutMI.setOpcode(X86::AND8mr); goto ReSimplify;
575 case X86::RELEASE_AND32mi: OutMI.setOpcode(X86::AND32mi); goto ReSimplify;
576 case X86::RELEASE_AND32mr: OutMI.setOpcode(X86::AND32mr); goto ReSimplify;
577 case X86::RELEASE_AND64mi32: OutMI.setOpcode(X86::AND64mi32); goto ReSimplify;
578 case X86::RELEASE_AND64mr: OutMI.setOpcode(X86::AND64mr); goto ReSimplify;
579 case X86::RELEASE_OR8mi: OutMI.setOpcode(X86::OR8mi); goto ReSimplify;
580 case X86::RELEASE_OR8mr: OutMI.setOpcode(X86::OR8mr); goto ReSimplify;
581 case X86::RELEASE_OR32mi: OutMI.setOpcode(X86::OR32mi); goto ReSimplify;
582 case X86::RELEASE_OR32mr: OutMI.setOpcode(X86::OR32mr); goto ReSimplify;
583 case X86::RELEASE_OR64mi32: OutMI.setOpcode(X86::OR64mi32); goto ReSimplify;
584 case X86::RELEASE_OR64mr: OutMI.setOpcode(X86::OR64mr); goto ReSimplify;
585 case X86::RELEASE_XOR8mi: OutMI.setOpcode(X86::XOR8mi); goto ReSimplify;
586 case X86::RELEASE_XOR8mr: OutMI.setOpcode(X86::XOR8mr); goto ReSimplify;
587 case X86::RELEASE_XOR32mi: OutMI.setOpcode(X86::XOR32mi); goto ReSimplify;
588 case X86::RELEASE_XOR32mr: OutMI.setOpcode(X86::XOR32mr); goto ReSimplify;
589 case X86::RELEASE_XOR64mi32: OutMI.setOpcode(X86::XOR64mi32); goto ReSimplify;
590 case X86::RELEASE_XOR64mr: OutMI.setOpcode(X86::XOR64mr); goto ReSimplify;
591 case X86::RELEASE_INC8m: OutMI.setOpcode(X86::INC8m); goto ReSimplify;
592 case X86::RELEASE_INC16m: OutMI.setOpcode(X86::INC16m); goto ReSimplify;
593 case X86::RELEASE_INC32m: OutMI.setOpcode(X86::INC32m); goto ReSimplify;
594 case X86::RELEASE_INC64m: OutMI.setOpcode(X86::INC64m); goto ReSimplify;
595 case X86::RELEASE_DEC8m: OutMI.setOpcode(X86::DEC8m); goto ReSimplify;
596 case X86::RELEASE_DEC16m: OutMI.setOpcode(X86::DEC16m); goto ReSimplify;
597 case X86::RELEASE_DEC32m: OutMI.setOpcode(X86::DEC32m); goto ReSimplify;
598 case X86::RELEASE_DEC64m: OutMI.setOpcode(X86::DEC64m); goto ReSimplify;
600 // We don't currently select the correct instruction form for instructions
601 // which have a short %eax, etc. form. Handle this by custom lowering, for
604 // Note, we are currently not handling the following instructions:
605 // MOV64ao8, MOV64o8a
606 // XCHG16ar, XCHG32ar, XCHG64ar
607 case X86::MOV8mr_NOREX:
609 case X86::MOV8rm_NOREX:
616 switch (OutMI.getOpcode()) {
617 default: llvm_unreachable("Invalid opcode");
618 case X86::MOV8mr_NOREX:
619 case X86::MOV8mr: NewOpc = X86::MOV8o32a; break;
620 case X86::MOV8rm_NOREX:
621 case X86::MOV8rm: NewOpc = X86::MOV8ao32; break;
622 case X86::MOV16mr: NewOpc = X86::MOV16o32a; break;
623 case X86::MOV16rm: NewOpc = X86::MOV16ao32; break;
624 case X86::MOV32mr: NewOpc = X86::MOV32o32a; break;
625 case X86::MOV32rm: NewOpc = X86::MOV32ao32; break;
627 SimplifyShortMoveForm(AsmPrinter, OutMI, NewOpc);
631 case X86::ADC8ri: case X86::ADC16ri: case X86::ADC32ri: case X86::ADC64ri32:
632 case X86::ADD8ri: case X86::ADD16ri: case X86::ADD32ri: case X86::ADD64ri32:
633 case X86::AND8ri: case X86::AND16ri: case X86::AND32ri: case X86::AND64ri32:
634 case X86::CMP8ri: case X86::CMP16ri: case X86::CMP32ri: case X86::CMP64ri32:
635 case X86::OR8ri: case X86::OR16ri: case X86::OR32ri: case X86::OR64ri32:
636 case X86::SBB8ri: case X86::SBB16ri: case X86::SBB32ri: case X86::SBB64ri32:
637 case X86::SUB8ri: case X86::SUB16ri: case X86::SUB32ri: case X86::SUB64ri32:
638 case X86::TEST8ri:case X86::TEST16ri:case X86::TEST32ri:case X86::TEST64ri32:
639 case X86::XOR8ri: case X86::XOR16ri: case X86::XOR32ri: case X86::XOR64ri32: {
641 switch (OutMI.getOpcode()) {
642 default: llvm_unreachable("Invalid opcode");
643 case X86::ADC8ri: NewOpc = X86::ADC8i8; break;
644 case X86::ADC16ri: NewOpc = X86::ADC16i16; break;
645 case X86::ADC32ri: NewOpc = X86::ADC32i32; break;
646 case X86::ADC64ri32: NewOpc = X86::ADC64i32; break;
647 case X86::ADD8ri: NewOpc = X86::ADD8i8; break;
648 case X86::ADD16ri: NewOpc = X86::ADD16i16; break;
649 case X86::ADD32ri: NewOpc = X86::ADD32i32; break;
650 case X86::ADD64ri32: NewOpc = X86::ADD64i32; break;
651 case X86::AND8ri: NewOpc = X86::AND8i8; break;
652 case X86::AND16ri: NewOpc = X86::AND16i16; break;
653 case X86::AND32ri: NewOpc = X86::AND32i32; break;
654 case X86::AND64ri32: NewOpc = X86::AND64i32; break;
655 case X86::CMP8ri: NewOpc = X86::CMP8i8; break;
656 case X86::CMP16ri: NewOpc = X86::CMP16i16; break;
657 case X86::CMP32ri: NewOpc = X86::CMP32i32; break;
658 case X86::CMP64ri32: NewOpc = X86::CMP64i32; break;
659 case X86::OR8ri: NewOpc = X86::OR8i8; break;
660 case X86::OR16ri: NewOpc = X86::OR16i16; break;
661 case X86::OR32ri: NewOpc = X86::OR32i32; break;
662 case X86::OR64ri32: NewOpc = X86::OR64i32; break;
663 case X86::SBB8ri: NewOpc = X86::SBB8i8; break;
664 case X86::SBB16ri: NewOpc = X86::SBB16i16; break;
665 case X86::SBB32ri: NewOpc = X86::SBB32i32; break;
666 case X86::SBB64ri32: NewOpc = X86::SBB64i32; break;
667 case X86::SUB8ri: NewOpc = X86::SUB8i8; break;
668 case X86::SUB16ri: NewOpc = X86::SUB16i16; break;
669 case X86::SUB32ri: NewOpc = X86::SUB32i32; break;
670 case X86::SUB64ri32: NewOpc = X86::SUB64i32; break;
671 case X86::TEST8ri: NewOpc = X86::TEST8i8; break;
672 case X86::TEST16ri: NewOpc = X86::TEST16i16; break;
673 case X86::TEST32ri: NewOpc = X86::TEST32i32; break;
674 case X86::TEST64ri32: NewOpc = X86::TEST64i32; break;
675 case X86::XOR8ri: NewOpc = X86::XOR8i8; break;
676 case X86::XOR16ri: NewOpc = X86::XOR16i16; break;
677 case X86::XOR32ri: NewOpc = X86::XOR32i32; break;
678 case X86::XOR64ri32: NewOpc = X86::XOR64i32; break;
680 SimplifyShortImmForm(OutMI, NewOpc);
684 // Try to shrink some forms of movsx.
685 case X86::MOVSX16rr8:
686 case X86::MOVSX32rr16:
687 case X86::MOVSX64rr32:
688 SimplifyMOVSX(OutMI);
693 void X86AsmPrinter::LowerTlsAddr(X86MCInstLower &MCInstLowering,
694 const MachineInstr &MI) {
696 bool is64Bits = MI.getOpcode() == X86::TLS_addr64 ||
697 MI.getOpcode() == X86::TLS_base_addr64;
699 bool needsPadding = MI.getOpcode() == X86::TLS_addr64;
701 MCContext &context = OutStreamer->getContext();
704 EmitAndCountInstruction(MCInstBuilder(X86::DATA16_PREFIX));
706 MCSymbolRefExpr::VariantKind SRVK;
707 switch (MI.getOpcode()) {
708 case X86::TLS_addr32:
709 case X86::TLS_addr64:
710 SRVK = MCSymbolRefExpr::VK_TLSGD;
712 case X86::TLS_base_addr32:
713 SRVK = MCSymbolRefExpr::VK_TLSLDM;
715 case X86::TLS_base_addr64:
716 SRVK = MCSymbolRefExpr::VK_TLSLD;
719 llvm_unreachable("unexpected opcode");
722 MCSymbol *sym = MCInstLowering.GetSymbolFromOperand(MI.getOperand(3));
723 const MCSymbolRefExpr *symRef = MCSymbolRefExpr::create(sym, SRVK, context);
727 LEA.setOpcode(X86::LEA64r);
728 LEA.addOperand(MCOperand::createReg(X86::RDI)); // dest
729 LEA.addOperand(MCOperand::createReg(X86::RIP)); // base
730 LEA.addOperand(MCOperand::createImm(1)); // scale
731 LEA.addOperand(MCOperand::createReg(0)); // index
732 LEA.addOperand(MCOperand::createExpr(symRef)); // disp
733 LEA.addOperand(MCOperand::createReg(0)); // seg
734 } else if (SRVK == MCSymbolRefExpr::VK_TLSLDM) {
735 LEA.setOpcode(X86::LEA32r);
736 LEA.addOperand(MCOperand::createReg(X86::EAX)); // dest
737 LEA.addOperand(MCOperand::createReg(X86::EBX)); // base
738 LEA.addOperand(MCOperand::createImm(1)); // scale
739 LEA.addOperand(MCOperand::createReg(0)); // index
740 LEA.addOperand(MCOperand::createExpr(symRef)); // disp
741 LEA.addOperand(MCOperand::createReg(0)); // seg
743 LEA.setOpcode(X86::LEA32r);
744 LEA.addOperand(MCOperand::createReg(X86::EAX)); // dest
745 LEA.addOperand(MCOperand::createReg(0)); // base
746 LEA.addOperand(MCOperand::createImm(1)); // scale
747 LEA.addOperand(MCOperand::createReg(X86::EBX)); // index
748 LEA.addOperand(MCOperand::createExpr(symRef)); // disp
749 LEA.addOperand(MCOperand::createReg(0)); // seg
751 EmitAndCountInstruction(LEA);
754 EmitAndCountInstruction(MCInstBuilder(X86::DATA16_PREFIX));
755 EmitAndCountInstruction(MCInstBuilder(X86::DATA16_PREFIX));
756 EmitAndCountInstruction(MCInstBuilder(X86::REX64_PREFIX));
759 StringRef name = is64Bits ? "__tls_get_addr" : "___tls_get_addr";
760 MCSymbol *tlsGetAddr = context.getOrCreateSymbol(name);
761 const MCSymbolRefExpr *tlsRef =
762 MCSymbolRefExpr::create(tlsGetAddr,
763 MCSymbolRefExpr::VK_PLT,
766 EmitAndCountInstruction(MCInstBuilder(is64Bits ? X86::CALL64pcrel32
771 /// \brief Emit the largest nop instruction smaller than or equal to \p NumBytes
772 /// bytes. Return the size of nop emitted.
773 static unsigned EmitNop(MCStreamer &OS, unsigned NumBytes, bool Is64Bit,
774 const MCSubtargetInfo &STI) {
775 // This works only for 64bit. For 32bit we have to do additional checking if
776 // the CPU supports multi-byte nops.
777 assert(Is64Bit && "EmitNops only supports X86-64");
780 unsigned Opc, BaseReg, ScaleVal, IndexReg, Displacement, SegmentReg;
781 Opc = IndexReg = Displacement = SegmentReg = 0;
785 case 0: llvm_unreachable("Zero nops?"); break;
786 case 1: NopSize = 1; Opc = X86::NOOP; break;
787 case 2: NopSize = 2; Opc = X86::XCHG16ar; break;
788 case 3: NopSize = 3; Opc = X86::NOOPL; break;
789 case 4: NopSize = 4; Opc = X86::NOOPL; Displacement = 8; break;
790 case 5: NopSize = 5; Opc = X86::NOOPL; Displacement = 8;
791 IndexReg = X86::RAX; break;
792 case 6: NopSize = 6; Opc = X86::NOOPW; Displacement = 8;
793 IndexReg = X86::RAX; break;
794 case 7: NopSize = 7; Opc = X86::NOOPL; Displacement = 512; break;
795 case 8: NopSize = 8; Opc = X86::NOOPL; Displacement = 512;
796 IndexReg = X86::RAX; break;
797 case 9: NopSize = 9; Opc = X86::NOOPW; Displacement = 512;
798 IndexReg = X86::RAX; break;
799 default: NopSize = 10; Opc = X86::NOOPW; Displacement = 512;
800 IndexReg = X86::RAX; SegmentReg = X86::CS; break;
803 unsigned NumPrefixes = std::min(NumBytes - NopSize, 5U);
804 NopSize += NumPrefixes;
805 for (unsigned i = 0; i != NumPrefixes; ++i)
806 OS.EmitBytes("\x66");
810 llvm_unreachable("Unexpected opcode");
813 OS.EmitInstruction(MCInstBuilder(Opc), STI);
816 OS.EmitInstruction(MCInstBuilder(Opc).addReg(X86::AX), STI);
820 OS.EmitInstruction(MCInstBuilder(Opc)
824 .addImm(Displacement)
829 assert(NopSize <= NumBytes && "We overemitted?");
833 /// \brief Emit the optimal amount of multi-byte nops on X86.
834 static void EmitNops(MCStreamer &OS, unsigned NumBytes, bool Is64Bit,
835 const MCSubtargetInfo &STI) {
836 unsigned NopsToEmit = NumBytes;
839 NumBytes -= EmitNop(OS, NumBytes, Is64Bit, STI);
840 assert(NopsToEmit >= NumBytes && "Emitted more than I asked for!");
844 void X86AsmPrinter::LowerSTATEPOINT(const MachineInstr &MI,
845 X86MCInstLower &MCIL) {
846 assert(Subtarget->is64Bit() && "Statepoint currently only supports X86-64");
848 StatepointOpers SOpers(&MI);
849 if (unsigned PatchBytes = SOpers.getNumPatchBytes()) {
850 EmitNops(*OutStreamer, PatchBytes, Subtarget->is64Bit(),
853 // Lower call target and choose correct opcode
854 const MachineOperand &CallTarget = SOpers.getCallTarget();
855 MCOperand CallTargetMCOp;
857 switch (CallTarget.getType()) {
858 case MachineOperand::MO_GlobalAddress:
859 case MachineOperand::MO_ExternalSymbol:
860 CallTargetMCOp = MCIL.LowerSymbolOperand(
861 CallTarget, MCIL.GetSymbolFromOperand(CallTarget));
862 CallOpcode = X86::CALL64pcrel32;
863 // Currently, we only support relative addressing with statepoints.
864 // Otherwise, we'll need a scratch register to hold the target
865 // address. You'll fail asserts during load & relocation if this
866 // symbol is to far away. (TODO: support non-relative addressing)
868 case MachineOperand::MO_Immediate:
869 CallTargetMCOp = MCOperand::createImm(CallTarget.getImm());
870 CallOpcode = X86::CALL64pcrel32;
871 // Currently, we only support relative addressing with statepoints.
872 // Otherwise, we'll need a scratch register to hold the target
873 // immediate. You'll fail asserts during load & relocation if this
874 // address is to far away. (TODO: support non-relative addressing)
876 case MachineOperand::MO_Register:
877 // FIXME: Add retpoline support and remove this.
878 if (Subtarget->useRetpoline())
879 report_fatal_error("Lowering register statepoints with retpoline not "
881 CallTargetMCOp = MCOperand::createReg(CallTarget.getReg());
882 CallOpcode = X86::CALL64r;
885 llvm_unreachable("Unsupported operand type in statepoint call target");
891 CallInst.setOpcode(CallOpcode);
892 CallInst.addOperand(CallTargetMCOp);
893 OutStreamer->EmitInstruction(CallInst, getSubtargetInfo());
896 // Record our statepoint node in the same section used by STACKMAP
898 SM.recordStatepoint(MI);
901 void X86AsmPrinter::LowerFAULTING_OP(const MachineInstr &FaultingMI,
902 X86MCInstLower &MCIL) {
903 // FAULTING_LOAD_OP <def>, <faltinf type>, <MBB handler>,
904 // <opcode>, <operands>
906 unsigned DefRegister = FaultingMI.getOperand(0).getReg();
907 FaultMaps::FaultKind FK =
908 static_cast<FaultMaps::FaultKind>(FaultingMI.getOperand(1).getImm());
909 MCSymbol *HandlerLabel = FaultingMI.getOperand(2).getMBB()->getSymbol();
910 unsigned Opcode = FaultingMI.getOperand(3).getImm();
911 unsigned OperandsBeginIdx = 4;
913 assert(FK < FaultMaps::FaultKindMax && "Invalid Faulting Kind!");
914 FM.recordFaultingOp(FK, HandlerLabel);
917 MI.setOpcode(Opcode);
919 if (DefRegister != X86::NoRegister)
920 MI.addOperand(MCOperand::createReg(DefRegister));
922 for (auto I = FaultingMI.operands_begin() + OperandsBeginIdx,
923 E = FaultingMI.operands_end();
925 if (auto MaybeOperand = MCIL.LowerMachineOperand(&FaultingMI, *I))
926 MI.addOperand(MaybeOperand.getValue());
928 OutStreamer->EmitInstruction(MI, getSubtargetInfo());
931 void X86AsmPrinter::LowerFENTRY_CALL(const MachineInstr &MI,
932 X86MCInstLower &MCIL) {
933 bool Is64Bits = Subtarget->is64Bit();
934 MCContext &Ctx = OutStreamer->getContext();
935 MCSymbol *fentry = Ctx.getOrCreateSymbol("__fentry__");
936 const MCSymbolRefExpr *Op =
937 MCSymbolRefExpr::create(fentry, MCSymbolRefExpr::VK_None, Ctx);
939 EmitAndCountInstruction(
940 MCInstBuilder(Is64Bits ? X86::CALL64pcrel32 : X86::CALLpcrel32)
944 void X86AsmPrinter::LowerPATCHABLE_OP(const MachineInstr &MI,
945 X86MCInstLower &MCIL) {
946 // PATCHABLE_OP minsize, opcode, operands
948 unsigned MinSize = MI.getOperand(0).getImm();
949 unsigned Opcode = MI.getOperand(1).getImm();
952 MCI.setOpcode(Opcode);
953 for (auto &MO : make_range(MI.operands_begin() + 2, MI.operands_end()))
954 if (auto MaybeOperand = MCIL.LowerMachineOperand(&MI, MO))
955 MCI.addOperand(MaybeOperand.getValue());
957 SmallString<256> Code;
958 SmallVector<MCFixup, 4> Fixups;
959 raw_svector_ostream VecOS(Code);
960 CodeEmitter->encodeInstruction(MCI, VecOS, Fixups, getSubtargetInfo());
962 if (Code.size() < MinSize) {
963 if (MinSize == 2 && Opcode == X86::PUSH64r) {
964 // This is an optimization that lets us get away without emitting a nop in
967 // NB! In some cases the encoding for PUSH64r (e.g. PUSH64r %r9) takes two
968 // bytes too, so the check on MinSize is important.
969 MCI.setOpcode(X86::PUSH64rmr);
971 unsigned NopSize = EmitNop(*OutStreamer, MinSize, Subtarget->is64Bit(),
973 assert(NopSize == MinSize && "Could not implement MinSize!");
978 OutStreamer->EmitInstruction(MCI, getSubtargetInfo());
981 // Lower a stackmap of the form:
982 // <id>, <shadowBytes>, ...
983 void X86AsmPrinter::LowerSTACKMAP(const MachineInstr &MI) {
984 SMShadowTracker.emitShadowPadding(*OutStreamer, getSubtargetInfo());
985 SM.recordStackMap(MI);
986 unsigned NumShadowBytes = MI.getOperand(1).getImm();
987 SMShadowTracker.reset(NumShadowBytes);
990 // Lower a patchpoint of the form:
991 // [<def>], <id>, <numBytes>, <target>, <numArgs>, <cc>, ...
992 void X86AsmPrinter::LowerPATCHPOINT(const MachineInstr &MI,
993 X86MCInstLower &MCIL) {
994 assert(Subtarget->is64Bit() && "Patchpoint currently only supports X86-64");
996 SMShadowTracker.emitShadowPadding(*OutStreamer, getSubtargetInfo());
998 SM.recordPatchPoint(MI);
1000 PatchPointOpers opers(&MI);
1001 unsigned ScratchIdx = opers.getNextScratchIdx();
1002 unsigned EncodedBytes = 0;
1003 const MachineOperand &CalleeMO = opers.getCallTarget();
1005 // Check for null target. If target is non-null (i.e. is non-zero or is
1006 // symbolic) then emit a call.
1007 if (!(CalleeMO.isImm() && !CalleeMO.getImm())) {
1008 MCOperand CalleeMCOp;
1009 switch (CalleeMO.getType()) {
1011 /// FIXME: Add a verifier check for bad callee types.
1012 llvm_unreachable("Unrecognized callee operand type.");
1013 case MachineOperand::MO_Immediate:
1014 if (CalleeMO.getImm())
1015 CalleeMCOp = MCOperand::createImm(CalleeMO.getImm());
1017 case MachineOperand::MO_ExternalSymbol:
1018 case MachineOperand::MO_GlobalAddress:
1020 MCIL.LowerSymbolOperand(CalleeMO,
1021 MCIL.GetSymbolFromOperand(CalleeMO));
1025 // Emit MOV to materialize the target address and the CALL to target.
1026 // This is encoded with 12-13 bytes, depending on which register is used.
1027 unsigned ScratchReg = MI.getOperand(ScratchIdx).getReg();
1028 if (X86II::isX86_64ExtendedReg(ScratchReg))
1033 EmitAndCountInstruction(
1034 MCInstBuilder(X86::MOV64ri).addReg(ScratchReg).addOperand(CalleeMCOp));
1035 // FIXME: Add retpoline support and remove this.
1036 if (Subtarget->useRetpoline())
1038 "Lowering patchpoint with retpoline not yet implemented.");
1039 EmitAndCountInstruction(MCInstBuilder(X86::CALL64r).addReg(ScratchReg));
1043 unsigned NumBytes = opers.getNumPatchBytes();
1044 assert(NumBytes >= EncodedBytes &&
1045 "Patchpoint can't request size less than the length of a call.");
1047 EmitNops(*OutStreamer, NumBytes - EncodedBytes, Subtarget->is64Bit(),
1048 getSubtargetInfo());
1051 void X86AsmPrinter::LowerPATCHABLE_EVENT_CALL(const MachineInstr &MI,
1052 X86MCInstLower &MCIL) {
1053 assert(Subtarget->is64Bit() && "XRay custom events only supports X86-64");
1055 // We want to emit the following pattern, which follows the x86 calling
1056 // convention to prepare for the trampoline call to be patched in.
1059 // .Lxray_event_sled_N:
1060 // jmp +N // jump across the instrumentation sled
1061 // ... // set up arguments in register
1062 // callq __xray_CustomEvent@plt // force dependency to symbol
1066 // After patching, it would look something like:
1068 // nopw (2-byte nop)
1070 // callq __xrayCustomEvent // already lowered
1074 // First we emit the label and the jump.
1075 auto CurSled = OutContext.createTempSymbol("xray_event_sled_", true);
1076 OutStreamer->AddComment("# XRay Custom Event Log");
1077 OutStreamer->EmitCodeAlignment(2);
1078 OutStreamer->EmitLabel(CurSled);
1080 // Use a two-byte `jmp`. This version of JMP takes an 8-bit relative offset as
1081 // an operand (computed as an offset from the jmp instruction).
1082 // FIXME: Find another less hacky way do force the relative jump.
1083 OutStreamer->EmitBinaryData("\xeb\x0f");
1085 // The default C calling convention will place two arguments into %rcx and
1086 // %rdx -- so we only work with those.
1087 unsigned UsedRegs[] = {X86::RDI, X86::RSI};
1088 bool UsedMask[] = {false, false};
1090 // Then we put the operands in the %rdi and %rsi registers. We spill the
1091 // values in the register before we clobber them, and mark them as used in
1092 // UsedMask. In case the arguments are already in the correct register, we use
1093 // emit nops appropriately sized to keep the sled the same size in every
1095 for (unsigned I = 0; I < MI.getNumOperands(); ++I)
1096 if (auto Op = MCIL.LowerMachineOperand(&MI, MI.getOperand(I))) {
1097 assert(Op->isReg() && "Only support arguments in registers");
1098 if (Op->getReg() != UsedRegs[I]) {
1100 EmitAndCountInstruction(
1101 MCInstBuilder(X86::PUSH64r).addReg(UsedRegs[I]));
1102 EmitAndCountInstruction(MCInstBuilder(X86::MOV64rr)
1103 .addReg(UsedRegs[I])
1104 .addReg(Op->getReg()));
1106 EmitNops(*OutStreamer, 4, Subtarget->is64Bit(), getSubtargetInfo());
1110 // We emit a hard dependency on the __xray_CustomEvent symbol, which is the
1111 // name of the trampoline to be implemented by the XRay runtime.
1112 auto TSym = OutContext.getOrCreateSymbol("__xray_CustomEvent");
1113 MachineOperand TOp = MachineOperand::CreateMCSymbol(TSym);
1114 if (isPositionIndependent())
1115 TOp.setTargetFlags(X86II::MO_PLT);
1117 // Emit the call instruction.
1118 EmitAndCountInstruction(MCInstBuilder(X86::CALL64pcrel32)
1119 .addOperand(MCIL.LowerSymbolOperand(TOp, TSym)));
1121 // Restore caller-saved and used registers.
1122 for (unsigned I = sizeof UsedMask; I-- > 0;)
1124 EmitAndCountInstruction(MCInstBuilder(X86::POP64r).addReg(UsedRegs[I]));
1126 EmitNops(*OutStreamer, 1, Subtarget->is64Bit(), getSubtargetInfo());
1128 OutStreamer->AddComment("xray custom event end.");
1130 // Record the sled version. Older versions of this sled were spelled
1131 // differently, so we let the runtime handle the different offsets we're
1133 recordSled(CurSled, MI, SledKind::CUSTOM_EVENT, 1);
1136 void X86AsmPrinter::LowerPATCHABLE_FUNCTION_ENTER(const MachineInstr &MI,
1137 X86MCInstLower &MCIL) {
1138 // We want to emit the following pattern:
1143 // # 9 bytes worth of noops
1145 // We need the 9 bytes because at runtime, we'd be patching over the full 11
1146 // bytes with the following pattern:
1148 // mov %r10, <function id, 32-bit> // 6 bytes
1149 // call <relative offset, 32-bits> // 5 bytes
1151 auto CurSled = OutContext.createTempSymbol("xray_sled_", true);
1152 OutStreamer->EmitCodeAlignment(2);
1153 OutStreamer->EmitLabel(CurSled);
1155 // Use a two-byte `jmp`. This version of JMP takes an 8-bit relative offset as
1156 // an operand (computed as an offset from the jmp instruction).
1157 // FIXME: Find another less hacky way do force the relative jump.
1158 OutStreamer->EmitBytes("\xeb\x09");
1159 EmitNops(*OutStreamer, 9, Subtarget->is64Bit(), getSubtargetInfo());
1160 recordSled(CurSled, MI, SledKind::FUNCTION_ENTER);
1163 void X86AsmPrinter::LowerPATCHABLE_RET(const MachineInstr &MI,
1164 X86MCInstLower &MCIL) {
1165 // Since PATCHABLE_RET takes the opcode of the return statement as an
1166 // argument, we use that to emit the correct form of the RET that we want.
1167 // i.e. when we see this:
1169 // PATCHABLE_RET X86::RET ...
1171 // We should emit the RET followed by sleds.
1175 // ret # or equivalent instruction
1176 // # 10 bytes worth of noops
1178 // This just makes sure that the alignment for the next instruction is 2.
1179 auto CurSled = OutContext.createTempSymbol("xray_sled_", true);
1180 OutStreamer->EmitCodeAlignment(2);
1181 OutStreamer->EmitLabel(CurSled);
1182 unsigned OpCode = MI.getOperand(0).getImm();
1184 Ret.setOpcode(OpCode);
1185 for (auto &MO : make_range(MI.operands_begin() + 1, MI.operands_end()))
1186 if (auto MaybeOperand = MCIL.LowerMachineOperand(&MI, MO))
1187 Ret.addOperand(MaybeOperand.getValue());
1188 OutStreamer->EmitInstruction(Ret, getSubtargetInfo());
1189 EmitNops(*OutStreamer, 10, Subtarget->is64Bit(), getSubtargetInfo());
1190 recordSled(CurSled, MI, SledKind::FUNCTION_EXIT);
1193 void X86AsmPrinter::LowerPATCHABLE_TAIL_CALL(const MachineInstr &MI, X86MCInstLower &MCIL) {
1194 // Like PATCHABLE_RET, we have the actual instruction in the operands to this
1195 // instruction so we lower that particular instruction and its operands.
1196 // Unlike PATCHABLE_RET though, we put the sled before the JMP, much like how
1197 // we do it for PATCHABLE_FUNCTION_ENTER. The sled should be very similar to
1198 // the PATCHABLE_FUNCTION_ENTER case, followed by the lowering of the actual
1199 // tail call much like how we have it in PATCHABLE_RET.
1200 auto CurSled = OutContext.createTempSymbol("xray_sled_", true);
1201 OutStreamer->EmitCodeAlignment(2);
1202 OutStreamer->EmitLabel(CurSled);
1203 auto Target = OutContext.createTempSymbol();
1205 // Use a two-byte `jmp`. This version of JMP takes an 8-bit relative offset as
1206 // an operand (computed as an offset from the jmp instruction).
1207 // FIXME: Find another less hacky way do force the relative jump.
1208 OutStreamer->EmitBytes("\xeb\x09");
1209 EmitNops(*OutStreamer, 9, Subtarget->is64Bit(), getSubtargetInfo());
1210 OutStreamer->EmitLabel(Target);
1211 recordSled(CurSled, MI, SledKind::TAIL_CALL);
1213 unsigned OpCode = MI.getOperand(0).getImm();
1215 TC.setOpcode(OpCode);
1217 // Before emitting the instruction, add a comment to indicate that this is
1218 // indeed a tail call.
1219 OutStreamer->AddComment("TAILCALL");
1220 for (auto &MO : make_range(MI.operands_begin() + 1, MI.operands_end()))
1221 if (auto MaybeOperand = MCIL.LowerMachineOperand(&MI, MO))
1222 TC.addOperand(MaybeOperand.getValue());
1223 OutStreamer->EmitInstruction(TC, getSubtargetInfo());
1226 // Returns instruction preceding MBBI in MachineFunction.
1227 // If MBBI is the first instruction of the first basic block, returns null.
1228 static MachineBasicBlock::const_iterator
1229 PrevCrossBBInst(MachineBasicBlock::const_iterator MBBI) {
1230 const MachineBasicBlock *MBB = MBBI->getParent();
1231 while (MBBI == MBB->begin()) {
1232 if (MBB == &MBB->getParent()->front())
1233 return MachineBasicBlock::const_iterator();
1234 MBB = MBB->getPrevNode();
1240 static const Constant *getConstantFromPool(const MachineInstr &MI,
1241 const MachineOperand &Op) {
1245 ArrayRef<MachineConstantPoolEntry> Constants =
1246 MI.getParent()->getParent()->getConstantPool()->getConstants();
1247 const MachineConstantPoolEntry &ConstantEntry =
1248 Constants[Op.getIndex()];
1250 // Bail if this is a machine constant pool entry, we won't be able to dig out
1252 if (ConstantEntry.isMachineConstantPoolEntry())
1255 auto *C = dyn_cast<Constant>(ConstantEntry.Val.ConstVal);
1256 assert((!C || ConstantEntry.getType() == C->getType()) &&
1257 "Expected a constant of the same type!");
1261 static std::string getShuffleComment(const MachineInstr *MI,
1264 ArrayRef<int> Mask) {
1265 std::string Comment;
1267 // Compute the name for a register. This is really goofy because we have
1268 // multiple instruction printers that could (in theory) use different
1269 // names. Fortunately most people use the ATT style (outside of Windows)
1270 // and they actually agree on register naming here. Ultimately, this is
1271 // a comment, and so its OK if it isn't perfect.
1272 auto GetRegisterName = [](unsigned RegNum) -> StringRef {
1273 return X86ATTInstPrinter::getRegisterName(RegNum);
1276 const MachineOperand &DstOp = MI->getOperand(0);
1277 const MachineOperand &SrcOp1 = MI->getOperand(SrcOp1Idx);
1278 const MachineOperand &SrcOp2 = MI->getOperand(SrcOp2Idx);
1280 StringRef DstName = DstOp.isReg() ? GetRegisterName(DstOp.getReg()) : "mem";
1281 StringRef Src1Name =
1282 SrcOp1.isReg() ? GetRegisterName(SrcOp1.getReg()) : "mem";
1283 StringRef Src2Name =
1284 SrcOp2.isReg() ? GetRegisterName(SrcOp2.getReg()) : "mem";
1286 // One source operand, fix the mask to print all elements in one span.
1287 SmallVector<int, 8> ShuffleMask(Mask.begin(), Mask.end());
1288 if (Src1Name == Src2Name)
1289 for (int i = 0, e = ShuffleMask.size(); i != e; ++i)
1290 if (ShuffleMask[i] >= e)
1291 ShuffleMask[i] -= e;
1293 raw_string_ostream CS(Comment);
1296 // Handle AVX512 MASK/MASXZ write mask comments.
1298 // MASKZ: zmmX {%kY} {z}
1299 if (SrcOp1Idx > 1) {
1300 assert((SrcOp1Idx == 2 || SrcOp1Idx == 3) && "Unexpected writemask");
1302 const MachineOperand &WriteMaskOp = MI->getOperand(SrcOp1Idx - 1);
1303 if (WriteMaskOp.isReg()) {
1304 CS << " {%" << GetRegisterName(WriteMaskOp.getReg()) << "}";
1306 if (SrcOp1Idx == 2) {
1314 for (int i = 0, e = ShuffleMask.size(); i != e; ++i) {
1317 if (ShuffleMask[i] == SM_SentinelZero) {
1322 // Otherwise, it must come from src1 or src2. Print the span of elements
1323 // that comes from this src.
1324 bool isSrc1 = ShuffleMask[i] < (int)e;
1325 CS << (isSrc1 ? Src1Name : Src2Name) << '[';
1327 bool IsFirst = true;
1328 while (i != e && ShuffleMask[i] != SM_SentinelZero &&
1329 (ShuffleMask[i] < (int)e) == isSrc1) {
1334 if (ShuffleMask[i] == SM_SentinelUndef)
1337 CS << ShuffleMask[i] % (int)e;
1341 --i; // For loop increments element #.
1348 static void printConstant(const Constant *COp, raw_ostream &CS) {
1349 if (isa<UndefValue>(COp)) {
1351 } else if (auto *CI = dyn_cast<ConstantInt>(COp)) {
1352 if (CI->getBitWidth() <= 64) {
1353 CS << CI->getZExtValue();
1355 // print multi-word constant as (w0,w1)
1356 const auto &Val = CI->getValue();
1358 for (int i = 0, N = Val.getNumWords(); i < N; ++i) {
1361 CS << Val.getRawData()[i];
1365 } else if (auto *CF = dyn_cast<ConstantFP>(COp)) {
1366 SmallString<32> Str;
1367 CF->getValueAPF().toString(Str);
1374 void X86AsmPrinter::EmitSEHInstruction(const MachineInstr *MI) {
1375 assert(MF->hasWinCFI() && "SEH_ instruction in function without WinCFI?");
1376 assert(getSubtarget().isOSWindows() && "SEH_ instruction Windows only");
1377 const X86RegisterInfo *RI =
1378 MF->getSubtarget<X86Subtarget>().getRegisterInfo();
1380 // Use the .cv_fpo directives if we're emitting CodeView on 32-bit x86.
1382 X86TargetStreamer *XTS =
1383 static_cast<X86TargetStreamer *>(OutStreamer->getTargetStreamer());
1384 switch (MI->getOpcode()) {
1385 case X86::SEH_PushReg:
1386 XTS->emitFPOPushReg(MI->getOperand(0).getImm());
1388 case X86::SEH_StackAlloc:
1389 XTS->emitFPOStackAlloc(MI->getOperand(0).getImm());
1391 case X86::SEH_SetFrame:
1392 assert(MI->getOperand(1).getImm() == 0 &&
1393 ".cv_fpo_setframe takes no offset");
1394 XTS->emitFPOSetFrame(MI->getOperand(0).getImm());
1396 case X86::SEH_EndPrologue:
1397 XTS->emitFPOEndPrologue();
1399 case X86::SEH_SaveReg:
1400 case X86::SEH_SaveXMM:
1401 case X86::SEH_PushFrame:
1402 llvm_unreachable("SEH_ directive incompatible with FPO");
1405 llvm_unreachable("expected SEH_ instruction");
1410 // Otherwise, use the .seh_ directives for all other Windows platforms.
1411 switch (MI->getOpcode()) {
1412 case X86::SEH_PushReg:
1413 OutStreamer->EmitWinCFIPushReg(
1414 RI->getSEHRegNum(MI->getOperand(0).getImm()));
1417 case X86::SEH_SaveReg:
1418 OutStreamer->EmitWinCFISaveReg(RI->getSEHRegNum(MI->getOperand(0).getImm()),
1419 MI->getOperand(1).getImm());
1422 case X86::SEH_SaveXMM:
1423 OutStreamer->EmitWinCFISaveXMM(RI->getSEHRegNum(MI->getOperand(0).getImm()),
1424 MI->getOperand(1).getImm());
1427 case X86::SEH_StackAlloc:
1428 OutStreamer->EmitWinCFIAllocStack(MI->getOperand(0).getImm());
1431 case X86::SEH_SetFrame:
1432 OutStreamer->EmitWinCFISetFrame(
1433 RI->getSEHRegNum(MI->getOperand(0).getImm()),
1434 MI->getOperand(1).getImm());
1437 case X86::SEH_PushFrame:
1438 OutStreamer->EmitWinCFIPushFrame(MI->getOperand(0).getImm());
1441 case X86::SEH_EndPrologue:
1442 OutStreamer->EmitWinCFIEndProlog();
1446 llvm_unreachable("expected SEH_ instruction");
1450 void X86AsmPrinter::EmitInstruction(const MachineInstr *MI) {
1451 X86MCInstLower MCInstLowering(*MF, *this);
1452 const X86RegisterInfo *RI = MF->getSubtarget<X86Subtarget>().getRegisterInfo();
1454 // Add a comment about EVEX-2-VEX compression for AVX-512 instrs that
1455 // are compressed from EVEX encoding to VEX encoding.
1456 if (TM.Options.MCOptions.ShowMCEncoding) {
1457 if (MI->getAsmPrinterFlags() & AC_EVEX_2_VEX)
1458 OutStreamer->AddComment("EVEX TO VEX Compression ", false);
1461 switch (MI->getOpcode()) {
1462 case TargetOpcode::DBG_VALUE:
1463 llvm_unreachable("Should be handled target independently");
1465 // Emit nothing here but a comment if we can.
1466 case X86::Int_MemBarrier:
1467 OutStreamer->emitRawComment("MEMBARRIER");
1471 case X86::EH_RETURN:
1472 case X86::EH_RETURN64: {
1473 // Lower these as normal, but add some comments.
1474 unsigned Reg = MI->getOperand(0).getReg();
1475 OutStreamer->AddComment(StringRef("eh_return, addr: %") +
1476 X86ATTInstPrinter::getRegisterName(Reg));
1479 case X86::CLEANUPRET: {
1480 // Lower these as normal, but add some comments.
1481 OutStreamer->AddComment("CLEANUPRET");
1485 case X86::CATCHRET: {
1486 // Lower these as normal, but add some comments.
1487 OutStreamer->AddComment("CATCHRET");
1494 case X86::TAILJMPd_CC:
1495 case X86::TAILJMPr64:
1496 case X86::TAILJMPm64:
1497 case X86::TAILJMPd64:
1498 case X86::TAILJMPd64_CC:
1499 case X86::TAILJMPr64_REX:
1500 case X86::TAILJMPm64_REX:
1501 // Lower these as normal, but add some comments.
1502 OutStreamer->AddComment("TAILCALL");
1505 case X86::TLS_addr32:
1506 case X86::TLS_addr64:
1507 case X86::TLS_base_addr32:
1508 case X86::TLS_base_addr64:
1509 return LowerTlsAddr(MCInstLowering, *MI);
1511 case X86::MOVPC32r: {
1512 // This is a pseudo op for a two instruction sequence with a label, which
1519 MCSymbol *PICBase = MF->getPICBaseSymbol();
1520 // FIXME: We would like an efficient form for this, so we don't have to do a
1521 // lot of extra uniquing.
1522 EmitAndCountInstruction(MCInstBuilder(X86::CALLpcrel32)
1523 .addExpr(MCSymbolRefExpr::create(PICBase, OutContext)));
1525 const X86FrameLowering* FrameLowering =
1526 MF->getSubtarget<X86Subtarget>().getFrameLowering();
1527 bool hasFP = FrameLowering->hasFP(*MF);
1529 // TODO: This is needed only if we require precise CFA.
1530 bool HasActiveDwarfFrame = OutStreamer->getNumFrameInfos() &&
1531 !OutStreamer->getDwarfFrameInfos().back().End;
1533 int stackGrowth = -RI->getSlotSize();
1535 if (HasActiveDwarfFrame && !hasFP) {
1536 OutStreamer->EmitCFIAdjustCfaOffset(-stackGrowth);
1540 OutStreamer->EmitLabel(PICBase);
1543 EmitAndCountInstruction(MCInstBuilder(X86::POP32r)
1544 .addReg(MI->getOperand(0).getReg()));
1546 if (HasActiveDwarfFrame && !hasFP) {
1547 OutStreamer->EmitCFIAdjustCfaOffset(stackGrowth);
1552 case X86::ADD32ri: {
1553 // Lower the MO_GOT_ABSOLUTE_ADDRESS form of ADD32ri.
1554 if (MI->getOperand(2).getTargetFlags() != X86II::MO_GOT_ABSOLUTE_ADDRESS)
1557 // Okay, we have something like:
1558 // EAX = ADD32ri EAX, MO_GOT_ABSOLUTE_ADDRESS(@MYGLOBAL)
1560 // For this, we want to print something like:
1561 // MYGLOBAL + (. - PICBASE)
1562 // However, we can't generate a ".", so just emit a new label here and refer
1564 MCSymbol *DotSym = OutContext.createTempSymbol();
1565 OutStreamer->EmitLabel(DotSym);
1567 // Now that we have emitted the label, lower the complex operand expression.
1568 MCSymbol *OpSym = MCInstLowering.GetSymbolFromOperand(MI->getOperand(2));
1570 const MCExpr *DotExpr = MCSymbolRefExpr::create(DotSym, OutContext);
1571 const MCExpr *PICBase =
1572 MCSymbolRefExpr::create(MF->getPICBaseSymbol(), OutContext);
1573 DotExpr = MCBinaryExpr::createSub(DotExpr, PICBase, OutContext);
1575 DotExpr = MCBinaryExpr::createAdd(MCSymbolRefExpr::create(OpSym,OutContext),
1576 DotExpr, OutContext);
1578 EmitAndCountInstruction(MCInstBuilder(X86::ADD32ri)
1579 .addReg(MI->getOperand(0).getReg())
1580 .addReg(MI->getOperand(1).getReg())
1584 case TargetOpcode::STATEPOINT:
1585 return LowerSTATEPOINT(*MI, MCInstLowering);
1587 case TargetOpcode::FAULTING_OP:
1588 return LowerFAULTING_OP(*MI, MCInstLowering);
1590 case TargetOpcode::FENTRY_CALL:
1591 return LowerFENTRY_CALL(*MI, MCInstLowering);
1593 case TargetOpcode::PATCHABLE_OP:
1594 return LowerPATCHABLE_OP(*MI, MCInstLowering);
1596 case TargetOpcode::STACKMAP:
1597 return LowerSTACKMAP(*MI);
1599 case TargetOpcode::PATCHPOINT:
1600 return LowerPATCHPOINT(*MI, MCInstLowering);
1602 case TargetOpcode::PATCHABLE_FUNCTION_ENTER:
1603 return LowerPATCHABLE_FUNCTION_ENTER(*MI, MCInstLowering);
1605 case TargetOpcode::PATCHABLE_RET:
1606 return LowerPATCHABLE_RET(*MI, MCInstLowering);
1608 case TargetOpcode::PATCHABLE_TAIL_CALL:
1609 return LowerPATCHABLE_TAIL_CALL(*MI, MCInstLowering);
1611 case TargetOpcode::PATCHABLE_EVENT_CALL:
1612 return LowerPATCHABLE_EVENT_CALL(*MI, MCInstLowering);
1614 case X86::MORESTACK_RET:
1615 EmitAndCountInstruction(MCInstBuilder(getRetOpcode(*Subtarget)));
1618 case X86::MORESTACK_RET_RESTORE_R10:
1619 // Return, then restore R10.
1620 EmitAndCountInstruction(MCInstBuilder(getRetOpcode(*Subtarget)));
1621 EmitAndCountInstruction(MCInstBuilder(X86::MOV64rr)
1626 case X86::SEH_PushReg:
1627 case X86::SEH_SaveReg:
1628 case X86::SEH_SaveXMM:
1629 case X86::SEH_StackAlloc:
1630 case X86::SEH_SetFrame:
1631 case X86::SEH_PushFrame:
1632 case X86::SEH_EndPrologue:
1633 EmitSEHInstruction(MI);
1636 case X86::SEH_Epilogue: {
1637 assert(MF->hasWinCFI() && "SEH_ instruction in function without WinCFI?");
1638 MachineBasicBlock::const_iterator MBBI(MI);
1639 // Check if preceded by a call and emit nop if so.
1640 for (MBBI = PrevCrossBBInst(MBBI);
1641 MBBI != MachineBasicBlock::const_iterator();
1642 MBBI = PrevCrossBBInst(MBBI)) {
1643 // Conservatively assume that pseudo instructions don't emit code and keep
1644 // looking for a call. We may emit an unnecessary nop in some cases.
1645 if (!MBBI->isPseudo()) {
1647 EmitAndCountInstruction(MCInstBuilder(X86::NOOP));
1654 // Lower PSHUFB and VPERMILP normally but add a comment if we can find
1655 // a constant shuffle mask. We won't be able to do this at the MC layer
1656 // because the mask isn't an immediate.
1658 case X86::VPSHUFBrm:
1659 case X86::VPSHUFBYrm:
1660 case X86::VPSHUFBZ128rm:
1661 case X86::VPSHUFBZ128rmk:
1662 case X86::VPSHUFBZ128rmkz:
1663 case X86::VPSHUFBZ256rm:
1664 case X86::VPSHUFBZ256rmk:
1665 case X86::VPSHUFBZ256rmkz:
1666 case X86::VPSHUFBZrm:
1667 case X86::VPSHUFBZrmk:
1668 case X86::VPSHUFBZrmkz: {
1669 if (!OutStreamer->isVerboseAsm())
1671 unsigned SrcIdx, MaskIdx;
1672 switch (MI->getOpcode()) {
1673 default: llvm_unreachable("Invalid opcode");
1675 case X86::VPSHUFBrm:
1676 case X86::VPSHUFBYrm:
1677 case X86::VPSHUFBZ128rm:
1678 case X86::VPSHUFBZ256rm:
1679 case X86::VPSHUFBZrm:
1680 SrcIdx = 1; MaskIdx = 5; break;
1681 case X86::VPSHUFBZ128rmkz:
1682 case X86::VPSHUFBZ256rmkz:
1683 case X86::VPSHUFBZrmkz:
1684 SrcIdx = 2; MaskIdx = 6; break;
1685 case X86::VPSHUFBZ128rmk:
1686 case X86::VPSHUFBZ256rmk:
1687 case X86::VPSHUFBZrmk:
1688 SrcIdx = 3; MaskIdx = 7; break;
1691 assert(MI->getNumOperands() >= 6 &&
1692 "We should always have at least 6 operands!");
1694 const MachineOperand &MaskOp = MI->getOperand(MaskIdx);
1695 if (auto *C = getConstantFromPool(*MI, MaskOp)) {
1696 SmallVector<int, 64> Mask;
1697 DecodePSHUFBMask(C, Mask);
1699 OutStreamer->AddComment(getShuffleComment(MI, SrcIdx, SrcIdx, Mask),
1700 !EnablePrintSchedInfo);
1705 case X86::VPERMILPSrm:
1706 case X86::VPERMILPSYrm:
1707 case X86::VPERMILPSZ128rm:
1708 case X86::VPERMILPSZ128rmk:
1709 case X86::VPERMILPSZ128rmkz:
1710 case X86::VPERMILPSZ256rm:
1711 case X86::VPERMILPSZ256rmk:
1712 case X86::VPERMILPSZ256rmkz:
1713 case X86::VPERMILPSZrm:
1714 case X86::VPERMILPSZrmk:
1715 case X86::VPERMILPSZrmkz:
1716 case X86::VPERMILPDrm:
1717 case X86::VPERMILPDYrm:
1718 case X86::VPERMILPDZ128rm:
1719 case X86::VPERMILPDZ128rmk:
1720 case X86::VPERMILPDZ128rmkz:
1721 case X86::VPERMILPDZ256rm:
1722 case X86::VPERMILPDZ256rmk:
1723 case X86::VPERMILPDZ256rmkz:
1724 case X86::VPERMILPDZrm:
1725 case X86::VPERMILPDZrmk:
1726 case X86::VPERMILPDZrmkz: {
1727 if (!OutStreamer->isVerboseAsm())
1729 unsigned SrcIdx, MaskIdx;
1731 switch (MI->getOpcode()) {
1732 default: llvm_unreachable("Invalid opcode");
1733 case X86::VPERMILPSrm:
1734 case X86::VPERMILPSYrm:
1735 case X86::VPERMILPSZ128rm:
1736 case X86::VPERMILPSZ256rm:
1737 case X86::VPERMILPSZrm:
1738 SrcIdx = 1; MaskIdx = 5; ElSize = 32; break;
1739 case X86::VPERMILPSZ128rmkz:
1740 case X86::VPERMILPSZ256rmkz:
1741 case X86::VPERMILPSZrmkz:
1742 SrcIdx = 2; MaskIdx = 6; ElSize = 32; break;
1743 case X86::VPERMILPSZ128rmk:
1744 case X86::VPERMILPSZ256rmk:
1745 case X86::VPERMILPSZrmk:
1746 SrcIdx = 3; MaskIdx = 7; ElSize = 32; break;
1747 case X86::VPERMILPDrm:
1748 case X86::VPERMILPDYrm:
1749 case X86::VPERMILPDZ128rm:
1750 case X86::VPERMILPDZ256rm:
1751 case X86::VPERMILPDZrm:
1752 SrcIdx = 1; MaskIdx = 5; ElSize = 64; break;
1753 case X86::VPERMILPDZ128rmkz:
1754 case X86::VPERMILPDZ256rmkz:
1755 case X86::VPERMILPDZrmkz:
1756 SrcIdx = 2; MaskIdx = 6; ElSize = 64; break;
1757 case X86::VPERMILPDZ128rmk:
1758 case X86::VPERMILPDZ256rmk:
1759 case X86::VPERMILPDZrmk:
1760 SrcIdx = 3; MaskIdx = 7; ElSize = 64; break;
1763 assert(MI->getNumOperands() >= 6 &&
1764 "We should always have at least 6 operands!");
1766 const MachineOperand &MaskOp = MI->getOperand(MaskIdx);
1767 if (auto *C = getConstantFromPool(*MI, MaskOp)) {
1768 SmallVector<int, 16> Mask;
1769 DecodeVPERMILPMask(C, ElSize, Mask);
1771 OutStreamer->AddComment(getShuffleComment(MI, SrcIdx, SrcIdx, Mask),
1772 !EnablePrintSchedInfo);
1777 case X86::VPERMIL2PDrm:
1778 case X86::VPERMIL2PSrm:
1779 case X86::VPERMIL2PDYrm:
1780 case X86::VPERMIL2PSYrm: {
1781 if (!OutStreamer->isVerboseAsm())
1783 assert(MI->getNumOperands() >= 8 &&
1784 "We should always have at least 8 operands!");
1786 const MachineOperand &CtrlOp = MI->getOperand(MI->getNumOperands() - 1);
1787 if (!CtrlOp.isImm())
1791 switch (MI->getOpcode()) {
1792 default: llvm_unreachable("Invalid opcode");
1793 case X86::VPERMIL2PSrm: case X86::VPERMIL2PSYrm: ElSize = 32; break;
1794 case X86::VPERMIL2PDrm: case X86::VPERMIL2PDYrm: ElSize = 64; break;
1797 const MachineOperand &MaskOp = MI->getOperand(6);
1798 if (auto *C = getConstantFromPool(*MI, MaskOp)) {
1799 SmallVector<int, 16> Mask;
1800 DecodeVPERMIL2PMask(C, (unsigned)CtrlOp.getImm(), ElSize, Mask);
1802 OutStreamer->AddComment(getShuffleComment(MI, 1, 2, Mask),
1803 !EnablePrintSchedInfo);
1808 case X86::VPPERMrrm: {
1809 if (!OutStreamer->isVerboseAsm())
1811 assert(MI->getNumOperands() >= 7 &&
1812 "We should always have at least 7 operands!");
1814 const MachineOperand &MaskOp = MI->getOperand(6);
1815 if (auto *C = getConstantFromPool(*MI, MaskOp)) {
1816 SmallVector<int, 16> Mask;
1817 DecodeVPPERMMask(C, Mask);
1819 OutStreamer->AddComment(getShuffleComment(MI, 1, 2, Mask),
1820 !EnablePrintSchedInfo);
1825 #define MOV_CASE(Prefix, Suffix) \
1826 case X86::Prefix##MOVAPD##Suffix##rm: \
1827 case X86::Prefix##MOVAPS##Suffix##rm: \
1828 case X86::Prefix##MOVUPD##Suffix##rm: \
1829 case X86::Prefix##MOVUPS##Suffix##rm: \
1830 case X86::Prefix##MOVDQA##Suffix##rm: \
1831 case X86::Prefix##MOVDQU##Suffix##rm:
1833 #define MOV_AVX512_CASE(Suffix) \
1834 case X86::VMOVDQA64##Suffix##rm: \
1835 case X86::VMOVDQA32##Suffix##rm: \
1836 case X86::VMOVDQU64##Suffix##rm: \
1837 case X86::VMOVDQU32##Suffix##rm: \
1838 case X86::VMOVDQU16##Suffix##rm: \
1839 case X86::VMOVDQU8##Suffix##rm: \
1840 case X86::VMOVAPS##Suffix##rm: \
1841 case X86::VMOVAPD##Suffix##rm: \
1842 case X86::VMOVUPS##Suffix##rm: \
1843 case X86::VMOVUPD##Suffix##rm:
1845 #define CASE_ALL_MOV_RM() \
1846 MOV_CASE(, ) /* SSE */ \
1847 MOV_CASE(V, ) /* AVX-128 */ \
1848 MOV_CASE(V, Y) /* AVX-256 */ \
1849 MOV_AVX512_CASE(Z) \
1850 MOV_AVX512_CASE(Z256) \
1851 MOV_AVX512_CASE(Z128)
1853 // For loads from a constant pool to a vector register, print the constant
1856 case X86::VBROADCASTF128:
1857 case X86::VBROADCASTI128:
1858 case X86::VBROADCASTF32X4Z256rm:
1859 case X86::VBROADCASTF32X4rm:
1860 case X86::VBROADCASTF32X8rm:
1861 case X86::VBROADCASTF64X2Z128rm:
1862 case X86::VBROADCASTF64X2rm:
1863 case X86::VBROADCASTF64X4rm:
1864 case X86::VBROADCASTI32X4Z256rm:
1865 case X86::VBROADCASTI32X4rm:
1866 case X86::VBROADCASTI32X8rm:
1867 case X86::VBROADCASTI64X2Z128rm:
1868 case X86::VBROADCASTI64X2rm:
1869 case X86::VBROADCASTI64X4rm:
1870 if (!OutStreamer->isVerboseAsm())
1872 if (MI->getNumOperands() <= 4)
1874 if (auto *C = getConstantFromPool(*MI, MI->getOperand(4))) {
1876 // Override NumLanes for the broadcast instructions.
1877 switch (MI->getOpcode()) {
1878 case X86::VBROADCASTF128: NumLanes = 2; break;
1879 case X86::VBROADCASTI128: NumLanes = 2; break;
1880 case X86::VBROADCASTF32X4Z256rm: NumLanes = 2; break;
1881 case X86::VBROADCASTF32X4rm: NumLanes = 4; break;
1882 case X86::VBROADCASTF32X8rm: NumLanes = 2; break;
1883 case X86::VBROADCASTF64X2Z128rm: NumLanes = 2; break;
1884 case X86::VBROADCASTF64X2rm: NumLanes = 4; break;
1885 case X86::VBROADCASTF64X4rm: NumLanes = 2; break;
1886 case X86::VBROADCASTI32X4Z256rm: NumLanes = 2; break;
1887 case X86::VBROADCASTI32X4rm: NumLanes = 4; break;
1888 case X86::VBROADCASTI32X8rm: NumLanes = 2; break;
1889 case X86::VBROADCASTI64X2Z128rm: NumLanes = 2; break;
1890 case X86::VBROADCASTI64X2rm: NumLanes = 4; break;
1891 case X86::VBROADCASTI64X4rm: NumLanes = 2; break;
1894 std::string Comment;
1895 raw_string_ostream CS(Comment);
1896 const MachineOperand &DstOp = MI->getOperand(0);
1897 CS << X86ATTInstPrinter::getRegisterName(DstOp.getReg()) << " = ";
1898 if (auto *CDS = dyn_cast<ConstantDataSequential>(C)) {
1900 for (int l = 0; l != NumLanes; ++l) {
1901 for (int i = 0, NumElements = CDS->getNumElements(); i < NumElements; ++i) {
1902 if (i != 0 || l != 0)
1904 if (CDS->getElementType()->isIntegerTy())
1905 CS << CDS->getElementAsInteger(i);
1906 else if (CDS->getElementType()->isFloatTy())
1907 CS << CDS->getElementAsFloat(i);
1908 else if (CDS->getElementType()->isDoubleTy())
1909 CS << CDS->getElementAsDouble(i);
1915 OutStreamer->AddComment(CS.str(), !EnablePrintSchedInfo);
1916 } else if (auto *CV = dyn_cast<ConstantVector>(C)) {
1918 for (int l = 0; l != NumLanes; ++l) {
1919 for (int i = 0, NumOperands = CV->getNumOperands(); i < NumOperands; ++i) {
1920 if (i != 0 || l != 0)
1922 printConstant(CV->getOperand(i), CS);
1926 OutStreamer->AddComment(CS.str(), !EnablePrintSchedInfo);
1930 case X86::VBROADCASTSSrm:
1931 case X86::VBROADCASTSSYrm:
1932 case X86::VBROADCASTSSZ128m:
1933 case X86::VBROADCASTSSZ256m:
1934 case X86::VBROADCASTSSZm:
1935 case X86::VBROADCASTSDYrm:
1936 case X86::VBROADCASTSDZ256m:
1937 case X86::VBROADCASTSDZm:
1938 case X86::VPBROADCASTBrm:
1939 case X86::VPBROADCASTBYrm:
1940 case X86::VPBROADCASTBZ128m:
1941 case X86::VPBROADCASTBZ256m:
1942 case X86::VPBROADCASTBZm:
1943 case X86::VPBROADCASTDrm:
1944 case X86::VPBROADCASTDYrm:
1945 case X86::VPBROADCASTDZ128m:
1946 case X86::VPBROADCASTDZ256m:
1947 case X86::VPBROADCASTDZm:
1948 case X86::VPBROADCASTQrm:
1949 case X86::VPBROADCASTQYrm:
1950 case X86::VPBROADCASTQZ128m:
1951 case X86::VPBROADCASTQZ256m:
1952 case X86::VPBROADCASTQZm:
1953 case X86::VPBROADCASTWrm:
1954 case X86::VPBROADCASTWYrm:
1955 case X86::VPBROADCASTWZ128m:
1956 case X86::VPBROADCASTWZ256m:
1957 case X86::VPBROADCASTWZm:
1958 if (!OutStreamer->isVerboseAsm())
1960 if (MI->getNumOperands() <= 4)
1962 if (auto *C = getConstantFromPool(*MI, MI->getOperand(4))) {
1964 switch (MI->getOpcode()) {
1965 default: llvm_unreachable("Invalid opcode");
1966 case X86::VBROADCASTSSrm: NumElts = 4; break;
1967 case X86::VBROADCASTSSYrm: NumElts = 8; break;
1968 case X86::VBROADCASTSSZ128m: NumElts = 4; break;
1969 case X86::VBROADCASTSSZ256m: NumElts = 8; break;
1970 case X86::VBROADCASTSSZm: NumElts = 16; break;
1971 case X86::VBROADCASTSDYrm: NumElts = 4; break;
1972 case X86::VBROADCASTSDZ256m: NumElts = 4; break;
1973 case X86::VBROADCASTSDZm: NumElts = 8; break;
1974 case X86::VPBROADCASTBrm: NumElts = 16; break;
1975 case X86::VPBROADCASTBYrm: NumElts = 32; break;
1976 case X86::VPBROADCASTBZ128m: NumElts = 16; break;
1977 case X86::VPBROADCASTBZ256m: NumElts = 32; break;
1978 case X86::VPBROADCASTBZm: NumElts = 64; break;
1979 case X86::VPBROADCASTDrm: NumElts = 4; break;
1980 case X86::VPBROADCASTDYrm: NumElts = 8; break;
1981 case X86::VPBROADCASTDZ128m: NumElts = 4; break;
1982 case X86::VPBROADCASTDZ256m: NumElts = 8; break;
1983 case X86::VPBROADCASTDZm: NumElts = 16; break;
1984 case X86::VPBROADCASTQrm: NumElts = 2; break;
1985 case X86::VPBROADCASTQYrm: NumElts = 4; break;
1986 case X86::VPBROADCASTQZ128m: NumElts = 2; break;
1987 case X86::VPBROADCASTQZ256m: NumElts = 4; break;
1988 case X86::VPBROADCASTQZm: NumElts = 8; break;
1989 case X86::VPBROADCASTWrm: NumElts = 8; break;
1990 case X86::VPBROADCASTWYrm: NumElts = 16; break;
1991 case X86::VPBROADCASTWZ128m: NumElts = 8; break;
1992 case X86::VPBROADCASTWZ256m: NumElts = 16; break;
1993 case X86::VPBROADCASTWZm: NumElts = 32; break;
1996 std::string Comment;
1997 raw_string_ostream CS(Comment);
1998 const MachineOperand &DstOp = MI->getOperand(0);
1999 CS << X86ATTInstPrinter::getRegisterName(DstOp.getReg()) << " = ";
2001 for (int i = 0; i != NumElts; ++i) {
2004 printConstant(C, CS);
2007 OutStreamer->AddComment(CS.str(), !EnablePrintSchedInfo);
2012 MCInstLowering.Lower(MI, TmpInst);
2013 if (MI->getAsmPrinterFlag(MachineInstr::NoSchedComment))
2014 TmpInst.setFlags(TmpInst.getFlags() | X86::NO_SCHED_INFO);
2016 // Stackmap shadows cannot include branch targets, so we can count the bytes
2017 // in a call towards the shadow, but must ensure that the no thread returns
2018 // in to the stackmap shadow. The only way to achieve this is if the call
2019 // is at the end of the shadow.
2021 // Count then size of the call towards the shadow
2022 SMShadowTracker.count(TmpInst, getSubtargetInfo(), CodeEmitter.get());
2023 // Then flush the shadow so that we fill with nops before the call, not
2025 SMShadowTracker.emitShadowPadding(*OutStreamer, getSubtargetInfo());
2026 // Then emit the call
2027 OutStreamer->EmitInstruction(TmpInst, getSubtargetInfo());
2031 EmitAndCountInstruction(TmpInst);