1 //===-- X86MCInstLower.cpp - Convert X86 MachineInstr to an MCInst --------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file contains code to lower X86 MachineInstrs to their corresponding
10 // MCInst records.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "MCTargetDesc/X86ATTInstPrinter.h"
15 #include "MCTargetDesc/X86BaseInfo.h"
16 #include "MCTargetDesc/X86InstComments.h"
17 #include "MCTargetDesc/X86ShuffleDecode.h"
18 #include "MCTargetDesc/X86TargetStreamer.h"
19 #include "X86AsmPrinter.h"
20 #include "X86RegisterInfo.h"
21 #include "X86ShuffleDecodeConstantPool.h"
22 #include "X86Subtarget.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/IR/DataLayout.h"
32 #include "llvm/IR/GlobalValue.h"
33 #include "llvm/IR/Mangler.h"
34 #include "llvm/MC/MCAsmInfo.h"
35 #include "llvm/MC/MCCodeEmitter.h"
36 #include "llvm/MC/MCContext.h"
37 #include "llvm/MC/MCExpr.h"
38 #include "llvm/MC/MCFixup.h"
39 #include "llvm/MC/MCInst.h"
40 #include "llvm/MC/MCInstBuilder.h"
41 #include "llvm/MC/MCSection.h"
42 #include "llvm/MC/MCSectionELF.h"
43 #include "llvm/MC/MCStreamer.h"
44 #include "llvm/MC/MCSymbol.h"
45 #include "llvm/MC/MCSymbolELF.h"
46 #include "llvm/Target/TargetLoweringObjectFile.h"
47 #include "llvm/Target/TargetMachine.h"
48 
49 using namespace llvm;
50 
51 namespace {
52 
53 /// X86MCInstLower - This class is used to lower an MachineInstr into an MCInst.
54 class X86MCInstLower {
55   MCContext &Ctx;
56   const MachineFunction &MF;
57   const TargetMachine &TM;
58   const MCAsmInfo &MAI;
59   X86AsmPrinter &AsmPrinter;
60 
61 public:
62   X86MCInstLower(const MachineFunction &MF, X86AsmPrinter &asmprinter);
63 
64   Optional<MCOperand> LowerMachineOperand(const MachineInstr *MI,
65                                           const MachineOperand &MO) const;
66   void Lower(const MachineInstr *MI, MCInst &OutMI) const;
67 
68   MCSymbol *GetSymbolFromOperand(const MachineOperand &MO) const;
69   MCOperand LowerSymbolOperand(const MachineOperand &MO, MCSymbol *Sym) const;
70 
71 private:
72   MachineModuleInfoMachO &getMachOMMI() const;
73 };
74 
75 } // end anonymous namespace
76 
77 /// A RAII helper which defines a region of instructions which can't have
78 /// padding added between them for correctness.
79 struct NoAutoPaddingScope {
80   MCStreamer &OS;
81   const bool OldAllowAutoPadding;
82   NoAutoPaddingScope(MCStreamer &OS)
83       : OS(OS), OldAllowAutoPadding(OS.getAllowAutoPadding()) {
84     changeAndComment(false);
85   }
86   ~NoAutoPaddingScope() { changeAndComment(OldAllowAutoPadding); }
87   void changeAndComment(bool b) {
88     if (b == OS.getAllowAutoPadding())
89       return;
90     OS.setAllowAutoPadding(b);
91     if (b)
92       OS.emitRawComment("autopadding");
93     else
94       OS.emitRawComment("noautopadding");
95   }
96 };
97 
98 // Emit a minimal sequence of nops spanning NumBytes bytes.
99 static void EmitNops(MCStreamer &OS, unsigned NumBytes, bool Is64Bit,
100                      const MCSubtargetInfo &STI);
101 
102 void X86AsmPrinter::StackMapShadowTracker::count(MCInst &Inst,
103                                                  const MCSubtargetInfo &STI,
104                                                  MCCodeEmitter *CodeEmitter) {
105   if (InShadow) {
106     SmallString<256> Code;
107     SmallVector<MCFixup, 4> Fixups;
108     raw_svector_ostream VecOS(Code);
109     CodeEmitter->encodeInstruction(Inst, VecOS, Fixups, STI);
110     CurrentShadowSize += Code.size();
111     if (CurrentShadowSize >= RequiredShadowSize)
112       InShadow = false; // The shadow is big enough. Stop counting.
113   }
114 }
115 
116 void X86AsmPrinter::StackMapShadowTracker::emitShadowPadding(
117     MCStreamer &OutStreamer, const MCSubtargetInfo &STI) {
118   if (InShadow && CurrentShadowSize < RequiredShadowSize) {
119     InShadow = false;
120     EmitNops(OutStreamer, RequiredShadowSize - CurrentShadowSize,
121              MF->getSubtarget<X86Subtarget>().is64Bit(), STI);
122   }
123 }
124 
125 void X86AsmPrinter::EmitAndCountInstruction(MCInst &Inst) {
126   OutStreamer->emitInstruction(Inst, getSubtargetInfo());
127   SMShadowTracker.count(Inst, getSubtargetInfo(), CodeEmitter.get());
128 }
129 
130 X86MCInstLower::X86MCInstLower(const MachineFunction &mf,
131                                X86AsmPrinter &asmprinter)
132     : Ctx(mf.getContext()), MF(mf), TM(mf.getTarget()), MAI(*TM.getMCAsmInfo()),
133       AsmPrinter(asmprinter) {}
134 
135 MachineModuleInfoMachO &X86MCInstLower::getMachOMMI() const {
136   return MF.getMMI().getObjFileInfo<MachineModuleInfoMachO>();
137 }
138 
139 /// GetSymbolFromOperand - Lower an MO_GlobalAddress or MO_ExternalSymbol
140 /// operand to an MCSymbol.
141 MCSymbol *X86MCInstLower::GetSymbolFromOperand(const MachineOperand &MO) const {
142   const Triple &TT = TM.getTargetTriple();
143   if (MO.isGlobal() && TT.isOSBinFormatELF())
144     return AsmPrinter.getSymbolPreferLocal(*MO.getGlobal());
145 
146   const DataLayout &DL = MF.getDataLayout();
147   assert((MO.isGlobal() || MO.isSymbol() || MO.isMBB()) &&
148          "Isn't a symbol reference");
149 
150   MCSymbol *Sym = nullptr;
151   SmallString<128> Name;
152   StringRef Suffix;
153 
154   switch (MO.getTargetFlags()) {
155   case X86II::MO_DLLIMPORT:
156     // Handle dllimport linkage.
157     Name += "__imp_";
158     break;
159   case X86II::MO_COFFSTUB:
160     Name += ".refptr.";
161     break;
162   case X86II::MO_DARWIN_NONLAZY:
163   case X86II::MO_DARWIN_NONLAZY_PIC_BASE:
164     Suffix = "$non_lazy_ptr";
165     break;
166   }
167 
168   if (!Suffix.empty())
169     Name += DL.getPrivateGlobalPrefix();
170 
171   if (MO.isGlobal()) {
172     const GlobalValue *GV = MO.getGlobal();
173     AsmPrinter.getNameWithPrefix(Name, GV);
174   } else if (MO.isSymbol()) {
175     Mangler::getNameWithPrefix(Name, MO.getSymbolName(), DL);
176   } else if (MO.isMBB()) {
177     assert(Suffix.empty());
178     Sym = MO.getMBB()->getSymbol();
179   }
180 
181   Name += Suffix;
182   if (!Sym)
183     Sym = Ctx.getOrCreateSymbol(Name);
184 
185   // If the target flags on the operand changes the name of the symbol, do that
186   // before we return the symbol.
187   switch (MO.getTargetFlags()) {
188   default:
189     break;
190   case X86II::MO_COFFSTUB: {
191     MachineModuleInfoCOFF &MMICOFF =
192         MF.getMMI().getObjFileInfo<MachineModuleInfoCOFF>();
193     MachineModuleInfoImpl::StubValueTy &StubSym = MMICOFF.getGVStubEntry(Sym);
194     if (!StubSym.getPointer()) {
195       assert(MO.isGlobal() && "Extern symbol not handled yet");
196       StubSym = MachineModuleInfoImpl::StubValueTy(
197           AsmPrinter.getSymbol(MO.getGlobal()), true);
198     }
199     break;
200   }
201   case X86II::MO_DARWIN_NONLAZY:
202   case X86II::MO_DARWIN_NONLAZY_PIC_BASE: {
203     MachineModuleInfoImpl::StubValueTy &StubSym =
204         getMachOMMI().getGVStubEntry(Sym);
205     if (!StubSym.getPointer()) {
206       assert(MO.isGlobal() && "Extern symbol not handled yet");
207       StubSym = MachineModuleInfoImpl::StubValueTy(
208           AsmPrinter.getSymbol(MO.getGlobal()),
209           !MO.getGlobal()->hasInternalLinkage());
210     }
211     break;
212   }
213   }
214 
215   return Sym;
216 }
217 
218 MCOperand X86MCInstLower::LowerSymbolOperand(const MachineOperand &MO,
219                                              MCSymbol *Sym) const {
220   // FIXME: We would like an efficient form for this, so we don't have to do a
221   // lot of extra uniquing.
222   const MCExpr *Expr = nullptr;
223   MCSymbolRefExpr::VariantKind RefKind = MCSymbolRefExpr::VK_None;
224 
225   switch (MO.getTargetFlags()) {
226   default:
227     llvm_unreachable("Unknown target flag on GV operand");
228   case X86II::MO_NO_FLAG: // No flag.
229   // These affect the name of the symbol, not any suffix.
230   case X86II::MO_DARWIN_NONLAZY:
231   case X86II::MO_DLLIMPORT:
232   case X86II::MO_COFFSTUB:
233     break;
234 
235   case X86II::MO_TLVP:
236     RefKind = MCSymbolRefExpr::VK_TLVP;
237     break;
238   case X86II::MO_TLVP_PIC_BASE:
239     Expr = MCSymbolRefExpr::create(Sym, MCSymbolRefExpr::VK_TLVP, Ctx);
240     // Subtract the pic base.
241     Expr = MCBinaryExpr::createSub(
242         Expr, MCSymbolRefExpr::create(MF.getPICBaseSymbol(), Ctx), Ctx);
243     break;
244   case X86II::MO_SECREL:
245     RefKind = MCSymbolRefExpr::VK_SECREL;
246     break;
247   case X86II::MO_TLSGD:
248     RefKind = MCSymbolRefExpr::VK_TLSGD;
249     break;
250   case X86II::MO_TLSLD:
251     RefKind = MCSymbolRefExpr::VK_TLSLD;
252     break;
253   case X86II::MO_TLSLDM:
254     RefKind = MCSymbolRefExpr::VK_TLSLDM;
255     break;
256   case X86II::MO_GOTTPOFF:
257     RefKind = MCSymbolRefExpr::VK_GOTTPOFF;
258     break;
259   case X86II::MO_INDNTPOFF:
260     RefKind = MCSymbolRefExpr::VK_INDNTPOFF;
261     break;
262   case X86II::MO_TPOFF:
263     RefKind = MCSymbolRefExpr::VK_TPOFF;
264     break;
265   case X86II::MO_DTPOFF:
266     RefKind = MCSymbolRefExpr::VK_DTPOFF;
267     break;
268   case X86II::MO_NTPOFF:
269     RefKind = MCSymbolRefExpr::VK_NTPOFF;
270     break;
271   case X86II::MO_GOTNTPOFF:
272     RefKind = MCSymbolRefExpr::VK_GOTNTPOFF;
273     break;
274   case X86II::MO_GOTPCREL:
275     RefKind = MCSymbolRefExpr::VK_GOTPCREL;
276     break;
277   case X86II::MO_GOT:
278     RefKind = MCSymbolRefExpr::VK_GOT;
279     break;
280   case X86II::MO_GOTOFF:
281     RefKind = MCSymbolRefExpr::VK_GOTOFF;
282     break;
283   case X86II::MO_PLT:
284     RefKind = MCSymbolRefExpr::VK_PLT;
285     break;
286   case X86II::MO_ABS8:
287     RefKind = MCSymbolRefExpr::VK_X86_ABS8;
288     break;
289   case X86II::MO_PIC_BASE_OFFSET:
290   case X86II::MO_DARWIN_NONLAZY_PIC_BASE:
291     Expr = MCSymbolRefExpr::create(Sym, Ctx);
292     // Subtract the pic base.
293     Expr = MCBinaryExpr::createSub(
294         Expr, MCSymbolRefExpr::create(MF.getPICBaseSymbol(), Ctx), Ctx);
295     if (MO.isJTI()) {
296       assert(MAI.doesSetDirectiveSuppressReloc());
297       // If .set directive is supported, use it to reduce the number of
298       // relocations the assembler will generate for differences between
299       // local labels. This is only safe when the symbols are in the same
300       // section so we are restricting it to jumptable references.
301       MCSymbol *Label = Ctx.createTempSymbol();
302       AsmPrinter.OutStreamer->emitAssignment(Label, Expr);
303       Expr = MCSymbolRefExpr::create(Label, Ctx);
304     }
305     break;
306   }
307 
308   if (!Expr)
309     Expr = MCSymbolRefExpr::create(Sym, RefKind, Ctx);
310 
311   if (!MO.isJTI() && !MO.isMBB() && MO.getOffset())
312     Expr = MCBinaryExpr::createAdd(
313         Expr, MCConstantExpr::create(MO.getOffset(), Ctx), Ctx);
314   return MCOperand::createExpr(Expr);
315 }
316 
317 /// Simplify FOO $imm, %{al,ax,eax,rax} to FOO $imm, for instruction with
318 /// a short fixed-register form.
319 static void SimplifyShortImmForm(MCInst &Inst, unsigned Opcode) {
320   unsigned ImmOp = Inst.getNumOperands() - 1;
321   assert(Inst.getOperand(0).isReg() &&
322          (Inst.getOperand(ImmOp).isImm() || Inst.getOperand(ImmOp).isExpr()) &&
323          ((Inst.getNumOperands() == 3 && Inst.getOperand(1).isReg() &&
324            Inst.getOperand(0).getReg() == Inst.getOperand(1).getReg()) ||
325           Inst.getNumOperands() == 2) &&
326          "Unexpected instruction!");
327 
328   // Check whether the destination register can be fixed.
329   unsigned Reg = Inst.getOperand(0).getReg();
330   if (Reg != X86::AL && Reg != X86::AX && Reg != X86::EAX && Reg != X86::RAX)
331     return;
332 
333   // If so, rewrite the instruction.
334   MCOperand Saved = Inst.getOperand(ImmOp);
335   Inst = MCInst();
336   Inst.setOpcode(Opcode);
337   Inst.addOperand(Saved);
338 }
339 
340 /// If a movsx instruction has a shorter encoding for the used register
341 /// simplify the instruction to use it instead.
342 static void SimplifyMOVSX(MCInst &Inst) {
343   unsigned NewOpcode = 0;
344   unsigned Op0 = Inst.getOperand(0).getReg(), Op1 = Inst.getOperand(1).getReg();
345   switch (Inst.getOpcode()) {
346   default:
347     llvm_unreachable("Unexpected instruction!");
348   case X86::MOVSX16rr8: // movsbw %al, %ax   --> cbtw
349     if (Op0 == X86::AX && Op1 == X86::AL)
350       NewOpcode = X86::CBW;
351     break;
352   case X86::MOVSX32rr16: // movswl %ax, %eax  --> cwtl
353     if (Op0 == X86::EAX && Op1 == X86::AX)
354       NewOpcode = X86::CWDE;
355     break;
356   case X86::MOVSX64rr32: // movslq %eax, %rax --> cltq
357     if (Op0 == X86::RAX && Op1 == X86::EAX)
358       NewOpcode = X86::CDQE;
359     break;
360   }
361 
362   if (NewOpcode != 0) {
363     Inst = MCInst();
364     Inst.setOpcode(NewOpcode);
365   }
366 }
367 
368 /// Simplify things like MOV32rm to MOV32o32a.
369 static void SimplifyShortMoveForm(X86AsmPrinter &Printer, MCInst &Inst,
370                                   unsigned Opcode) {
371   // Don't make these simplifications in 64-bit mode; other assemblers don't
372   // perform them because they make the code larger.
373   if (Printer.getSubtarget().is64Bit())
374     return;
375 
376   bool IsStore = Inst.getOperand(0).isReg() && Inst.getOperand(1).isReg();
377   unsigned AddrBase = IsStore;
378   unsigned RegOp = IsStore ? 0 : 5;
379   unsigned AddrOp = AddrBase + 3;
380   assert(
381       Inst.getNumOperands() == 6 && Inst.getOperand(RegOp).isReg() &&
382       Inst.getOperand(AddrBase + X86::AddrBaseReg).isReg() &&
383       Inst.getOperand(AddrBase + X86::AddrScaleAmt).isImm() &&
384       Inst.getOperand(AddrBase + X86::AddrIndexReg).isReg() &&
385       Inst.getOperand(AddrBase + X86::AddrSegmentReg).isReg() &&
386       (Inst.getOperand(AddrOp).isExpr() || Inst.getOperand(AddrOp).isImm()) &&
387       "Unexpected instruction!");
388 
389   // Check whether the destination register can be fixed.
390   unsigned Reg = Inst.getOperand(RegOp).getReg();
391   if (Reg != X86::AL && Reg != X86::AX && Reg != X86::EAX && Reg != X86::RAX)
392     return;
393 
394   // Check whether this is an absolute address.
395   // FIXME: We know TLVP symbol refs aren't, but there should be a better way
396   // to do this here.
397   bool Absolute = true;
398   if (Inst.getOperand(AddrOp).isExpr()) {
399     const MCExpr *MCE = Inst.getOperand(AddrOp).getExpr();
400     if (const MCSymbolRefExpr *SRE = dyn_cast<MCSymbolRefExpr>(MCE))
401       if (SRE->getKind() == MCSymbolRefExpr::VK_TLVP)
402         Absolute = false;
403   }
404 
405   if (Absolute &&
406       (Inst.getOperand(AddrBase + X86::AddrBaseReg).getReg() != 0 ||
407        Inst.getOperand(AddrBase + X86::AddrScaleAmt).getImm() != 1 ||
408        Inst.getOperand(AddrBase + X86::AddrIndexReg).getReg() != 0))
409     return;
410 
411   // If so, rewrite the instruction.
412   MCOperand Saved = Inst.getOperand(AddrOp);
413   MCOperand Seg = Inst.getOperand(AddrBase + X86::AddrSegmentReg);
414   Inst = MCInst();
415   Inst.setOpcode(Opcode);
416   Inst.addOperand(Saved);
417   Inst.addOperand(Seg);
418 }
419 
420 static unsigned getRetOpcode(const X86Subtarget &Subtarget) {
421   return Subtarget.is64Bit() ? X86::RETQ : X86::RETL;
422 }
423 
424 Optional<MCOperand>
425 X86MCInstLower::LowerMachineOperand(const MachineInstr *MI,
426                                     const MachineOperand &MO) const {
427   switch (MO.getType()) {
428   default:
429     MI->print(errs());
430     llvm_unreachable("unknown operand type");
431   case MachineOperand::MO_Register:
432     // Ignore all implicit register operands.
433     if (MO.isImplicit())
434       return None;
435     return MCOperand::createReg(MO.getReg());
436   case MachineOperand::MO_Immediate:
437     return MCOperand::createImm(MO.getImm());
438   case MachineOperand::MO_MachineBasicBlock:
439   case MachineOperand::MO_GlobalAddress:
440   case MachineOperand::MO_ExternalSymbol:
441     return LowerSymbolOperand(MO, GetSymbolFromOperand(MO));
442   case MachineOperand::MO_MCSymbol:
443     return LowerSymbolOperand(MO, MO.getMCSymbol());
444   case MachineOperand::MO_JumpTableIndex:
445     return LowerSymbolOperand(MO, AsmPrinter.GetJTISymbol(MO.getIndex()));
446   case MachineOperand::MO_ConstantPoolIndex:
447     return LowerSymbolOperand(MO, AsmPrinter.GetCPISymbol(MO.getIndex()));
448   case MachineOperand::MO_BlockAddress:
449     return LowerSymbolOperand(
450         MO, AsmPrinter.GetBlockAddressSymbol(MO.getBlockAddress()));
451   case MachineOperand::MO_RegisterMask:
452     // Ignore call clobbers.
453     return None;
454   }
455 }
456 
457 // Replace TAILJMP opcodes with their equivalent opcodes that have encoding
458 // information.
459 static unsigned convertTailJumpOpcode(unsigned Opcode) {
460   switch (Opcode) {
461   case X86::TAILJMPr:
462     Opcode = X86::JMP32r;
463     break;
464   case X86::TAILJMPm:
465     Opcode = X86::JMP32m;
466     break;
467   case X86::TAILJMPr64:
468     Opcode = X86::JMP64r;
469     break;
470   case X86::TAILJMPm64:
471     Opcode = X86::JMP64m;
472     break;
473   case X86::TAILJMPr64_REX:
474     Opcode = X86::JMP64r_REX;
475     break;
476   case X86::TAILJMPm64_REX:
477     Opcode = X86::JMP64m_REX;
478     break;
479   case X86::TAILJMPd:
480   case X86::TAILJMPd64:
481     Opcode = X86::JMP_1;
482     break;
483   case X86::TAILJMPd_CC:
484   case X86::TAILJMPd64_CC:
485     Opcode = X86::JCC_1;
486     break;
487   }
488 
489   return Opcode;
490 }
491 
492 void X86MCInstLower::Lower(const MachineInstr *MI, MCInst &OutMI) const {
493   OutMI.setOpcode(MI->getOpcode());
494 
495   for (const MachineOperand &MO : MI->operands())
496     if (auto MaybeMCOp = LowerMachineOperand(MI, MO))
497       OutMI.addOperand(MaybeMCOp.getValue());
498 
499   // Handle a few special cases to eliminate operand modifiers.
500   switch (OutMI.getOpcode()) {
501   case X86::LEA64_32r:
502   case X86::LEA64r:
503   case X86::LEA16r:
504   case X86::LEA32r:
505     // LEA should have a segment register, but it must be empty.
506     assert(OutMI.getNumOperands() == 1 + X86::AddrNumOperands &&
507            "Unexpected # of LEA operands");
508     assert(OutMI.getOperand(1 + X86::AddrSegmentReg).getReg() == 0 &&
509            "LEA has segment specified!");
510     break;
511 
512   // Commute operands to get a smaller encoding by using VEX.R instead of VEX.B
513   // if one of the registers is extended, but other isn't.
514   case X86::VMOVZPQILo2PQIrr:
515   case X86::VMOVAPDrr:
516   case X86::VMOVAPDYrr:
517   case X86::VMOVAPSrr:
518   case X86::VMOVAPSYrr:
519   case X86::VMOVDQArr:
520   case X86::VMOVDQAYrr:
521   case X86::VMOVDQUrr:
522   case X86::VMOVDQUYrr:
523   case X86::VMOVUPDrr:
524   case X86::VMOVUPDYrr:
525   case X86::VMOVUPSrr:
526   case X86::VMOVUPSYrr: {
527     if (!X86II::isX86_64ExtendedReg(OutMI.getOperand(0).getReg()) &&
528         X86II::isX86_64ExtendedReg(OutMI.getOperand(1).getReg())) {
529       unsigned NewOpc;
530       switch (OutMI.getOpcode()) {
531       default: llvm_unreachable("Invalid opcode");
532       case X86::VMOVZPQILo2PQIrr: NewOpc = X86::VMOVPQI2QIrr;   break;
533       case X86::VMOVAPDrr:        NewOpc = X86::VMOVAPDrr_REV;  break;
534       case X86::VMOVAPDYrr:       NewOpc = X86::VMOVAPDYrr_REV; break;
535       case X86::VMOVAPSrr:        NewOpc = X86::VMOVAPSrr_REV;  break;
536       case X86::VMOVAPSYrr:       NewOpc = X86::VMOVAPSYrr_REV; break;
537       case X86::VMOVDQArr:        NewOpc = X86::VMOVDQArr_REV;  break;
538       case X86::VMOVDQAYrr:       NewOpc = X86::VMOVDQAYrr_REV; break;
539       case X86::VMOVDQUrr:        NewOpc = X86::VMOVDQUrr_REV;  break;
540       case X86::VMOVDQUYrr:       NewOpc = X86::VMOVDQUYrr_REV; break;
541       case X86::VMOVUPDrr:        NewOpc = X86::VMOVUPDrr_REV;  break;
542       case X86::VMOVUPDYrr:       NewOpc = X86::VMOVUPDYrr_REV; break;
543       case X86::VMOVUPSrr:        NewOpc = X86::VMOVUPSrr_REV;  break;
544       case X86::VMOVUPSYrr:       NewOpc = X86::VMOVUPSYrr_REV; break;
545       }
546       OutMI.setOpcode(NewOpc);
547     }
548     break;
549   }
550   case X86::VMOVSDrr:
551   case X86::VMOVSSrr: {
552     if (!X86II::isX86_64ExtendedReg(OutMI.getOperand(0).getReg()) &&
553         X86II::isX86_64ExtendedReg(OutMI.getOperand(2).getReg())) {
554       unsigned NewOpc;
555       switch (OutMI.getOpcode()) {
556       default: llvm_unreachable("Invalid opcode");
557       case X86::VMOVSDrr: NewOpc = X86::VMOVSDrr_REV; break;
558       case X86::VMOVSSrr: NewOpc = X86::VMOVSSrr_REV; break;
559       }
560       OutMI.setOpcode(NewOpc);
561     }
562     break;
563   }
564 
565   case X86::VPCMPBZ128rmi:  case X86::VPCMPBZ128rmik:
566   case X86::VPCMPBZ128rri:  case X86::VPCMPBZ128rrik:
567   case X86::VPCMPBZ256rmi:  case X86::VPCMPBZ256rmik:
568   case X86::VPCMPBZ256rri:  case X86::VPCMPBZ256rrik:
569   case X86::VPCMPBZrmi:     case X86::VPCMPBZrmik:
570   case X86::VPCMPBZrri:     case X86::VPCMPBZrrik:
571   case X86::VPCMPDZ128rmi:  case X86::VPCMPDZ128rmik:
572   case X86::VPCMPDZ128rmib: case X86::VPCMPDZ128rmibk:
573   case X86::VPCMPDZ128rri:  case X86::VPCMPDZ128rrik:
574   case X86::VPCMPDZ256rmi:  case X86::VPCMPDZ256rmik:
575   case X86::VPCMPDZ256rmib: case X86::VPCMPDZ256rmibk:
576   case X86::VPCMPDZ256rri:  case X86::VPCMPDZ256rrik:
577   case X86::VPCMPDZrmi:     case X86::VPCMPDZrmik:
578   case X86::VPCMPDZrmib:    case X86::VPCMPDZrmibk:
579   case X86::VPCMPDZrri:     case X86::VPCMPDZrrik:
580   case X86::VPCMPQZ128rmi:  case X86::VPCMPQZ128rmik:
581   case X86::VPCMPQZ128rmib: case X86::VPCMPQZ128rmibk:
582   case X86::VPCMPQZ128rri:  case X86::VPCMPQZ128rrik:
583   case X86::VPCMPQZ256rmi:  case X86::VPCMPQZ256rmik:
584   case X86::VPCMPQZ256rmib: case X86::VPCMPQZ256rmibk:
585   case X86::VPCMPQZ256rri:  case X86::VPCMPQZ256rrik:
586   case X86::VPCMPQZrmi:     case X86::VPCMPQZrmik:
587   case X86::VPCMPQZrmib:    case X86::VPCMPQZrmibk:
588   case X86::VPCMPQZrri:     case X86::VPCMPQZrrik:
589   case X86::VPCMPWZ128rmi:  case X86::VPCMPWZ128rmik:
590   case X86::VPCMPWZ128rri:  case X86::VPCMPWZ128rrik:
591   case X86::VPCMPWZ256rmi:  case X86::VPCMPWZ256rmik:
592   case X86::VPCMPWZ256rri:  case X86::VPCMPWZ256rrik:
593   case X86::VPCMPWZrmi:     case X86::VPCMPWZrmik:
594   case X86::VPCMPWZrri:     case X86::VPCMPWZrrik: {
595     // Turn immediate 0 into the VPCMPEQ instruction.
596     if (OutMI.getOperand(OutMI.getNumOperands() - 1).getImm() == 0) {
597       unsigned NewOpc;
598       switch (OutMI.getOpcode()) {
599       default: llvm_unreachable("Invalid opcode");
600       case X86::VPCMPBZ128rmi:   NewOpc = X86::VPCMPEQBZ128rm;   break;
601       case X86::VPCMPBZ128rmik:  NewOpc = X86::VPCMPEQBZ128rmk;  break;
602       case X86::VPCMPBZ128rri:   NewOpc = X86::VPCMPEQBZ128rr;   break;
603       case X86::VPCMPBZ128rrik:  NewOpc = X86::VPCMPEQBZ128rrk;  break;
604       case X86::VPCMPBZ256rmi:   NewOpc = X86::VPCMPEQBZ256rm;   break;
605       case X86::VPCMPBZ256rmik:  NewOpc = X86::VPCMPEQBZ256rmk;  break;
606       case X86::VPCMPBZ256rri:   NewOpc = X86::VPCMPEQBZ256rr;   break;
607       case X86::VPCMPBZ256rrik:  NewOpc = X86::VPCMPEQBZ256rrk;  break;
608       case X86::VPCMPBZrmi:      NewOpc = X86::VPCMPEQBZrm;      break;
609       case X86::VPCMPBZrmik:     NewOpc = X86::VPCMPEQBZrmk;     break;
610       case X86::VPCMPBZrri:      NewOpc = X86::VPCMPEQBZrr;      break;
611       case X86::VPCMPBZrrik:     NewOpc = X86::VPCMPEQBZrrk;     break;
612       case X86::VPCMPDZ128rmi:   NewOpc = X86::VPCMPEQDZ128rm;   break;
613       case X86::VPCMPDZ128rmib:  NewOpc = X86::VPCMPEQDZ128rmb;  break;
614       case X86::VPCMPDZ128rmibk: NewOpc = X86::VPCMPEQDZ128rmbk; break;
615       case X86::VPCMPDZ128rmik:  NewOpc = X86::VPCMPEQDZ128rmk;  break;
616       case X86::VPCMPDZ128rri:   NewOpc = X86::VPCMPEQDZ128rr;   break;
617       case X86::VPCMPDZ128rrik:  NewOpc = X86::VPCMPEQDZ128rrk;  break;
618       case X86::VPCMPDZ256rmi:   NewOpc = X86::VPCMPEQDZ256rm;   break;
619       case X86::VPCMPDZ256rmib:  NewOpc = X86::VPCMPEQDZ256rmb;  break;
620       case X86::VPCMPDZ256rmibk: NewOpc = X86::VPCMPEQDZ256rmbk; break;
621       case X86::VPCMPDZ256rmik:  NewOpc = X86::VPCMPEQDZ256rmk;  break;
622       case X86::VPCMPDZ256rri:   NewOpc = X86::VPCMPEQDZ256rr;   break;
623       case X86::VPCMPDZ256rrik:  NewOpc = X86::VPCMPEQDZ256rrk;  break;
624       case X86::VPCMPDZrmi:      NewOpc = X86::VPCMPEQDZrm;      break;
625       case X86::VPCMPDZrmib:     NewOpc = X86::VPCMPEQDZrmb;     break;
626       case X86::VPCMPDZrmibk:    NewOpc = X86::VPCMPEQDZrmbk;    break;
627       case X86::VPCMPDZrmik:     NewOpc = X86::VPCMPEQDZrmk;     break;
628       case X86::VPCMPDZrri:      NewOpc = X86::VPCMPEQDZrr;      break;
629       case X86::VPCMPDZrrik:     NewOpc = X86::VPCMPEQDZrrk;     break;
630       case X86::VPCMPQZ128rmi:   NewOpc = X86::VPCMPEQQZ128rm;   break;
631       case X86::VPCMPQZ128rmib:  NewOpc = X86::VPCMPEQQZ128rmb;  break;
632       case X86::VPCMPQZ128rmibk: NewOpc = X86::VPCMPEQQZ128rmbk; break;
633       case X86::VPCMPQZ128rmik:  NewOpc = X86::VPCMPEQQZ128rmk;  break;
634       case X86::VPCMPQZ128rri:   NewOpc = X86::VPCMPEQQZ128rr;   break;
635       case X86::VPCMPQZ128rrik:  NewOpc = X86::VPCMPEQQZ128rrk;  break;
636       case X86::VPCMPQZ256rmi:   NewOpc = X86::VPCMPEQQZ256rm;   break;
637       case X86::VPCMPQZ256rmib:  NewOpc = X86::VPCMPEQQZ256rmb;  break;
638       case X86::VPCMPQZ256rmibk: NewOpc = X86::VPCMPEQQZ256rmbk; break;
639       case X86::VPCMPQZ256rmik:  NewOpc = X86::VPCMPEQQZ256rmk;  break;
640       case X86::VPCMPQZ256rri:   NewOpc = X86::VPCMPEQQZ256rr;   break;
641       case X86::VPCMPQZ256rrik:  NewOpc = X86::VPCMPEQQZ256rrk;  break;
642       case X86::VPCMPQZrmi:      NewOpc = X86::VPCMPEQQZrm;      break;
643       case X86::VPCMPQZrmib:     NewOpc = X86::VPCMPEQQZrmb;     break;
644       case X86::VPCMPQZrmibk:    NewOpc = X86::VPCMPEQQZrmbk;    break;
645       case X86::VPCMPQZrmik:     NewOpc = X86::VPCMPEQQZrmk;     break;
646       case X86::VPCMPQZrri:      NewOpc = X86::VPCMPEQQZrr;      break;
647       case X86::VPCMPQZrrik:     NewOpc = X86::VPCMPEQQZrrk;     break;
648       case X86::VPCMPWZ128rmi:   NewOpc = X86::VPCMPEQWZ128rm;   break;
649       case X86::VPCMPWZ128rmik:  NewOpc = X86::VPCMPEQWZ128rmk;  break;
650       case X86::VPCMPWZ128rri:   NewOpc = X86::VPCMPEQWZ128rr;   break;
651       case X86::VPCMPWZ128rrik:  NewOpc = X86::VPCMPEQWZ128rrk;  break;
652       case X86::VPCMPWZ256rmi:   NewOpc = X86::VPCMPEQWZ256rm;   break;
653       case X86::VPCMPWZ256rmik:  NewOpc = X86::VPCMPEQWZ256rmk;  break;
654       case X86::VPCMPWZ256rri:   NewOpc = X86::VPCMPEQWZ256rr;   break;
655       case X86::VPCMPWZ256rrik:  NewOpc = X86::VPCMPEQWZ256rrk;  break;
656       case X86::VPCMPWZrmi:      NewOpc = X86::VPCMPEQWZrm;      break;
657       case X86::VPCMPWZrmik:     NewOpc = X86::VPCMPEQWZrmk;     break;
658       case X86::VPCMPWZrri:      NewOpc = X86::VPCMPEQWZrr;      break;
659       case X86::VPCMPWZrrik:     NewOpc = X86::VPCMPEQWZrrk;     break;
660       }
661 
662       OutMI.setOpcode(NewOpc);
663       OutMI.erase(&OutMI.getOperand(OutMI.getNumOperands() - 1));
664       break;
665     }
666 
667     // Turn immediate 6 into the VPCMPGT instruction.
668     if (OutMI.getOperand(OutMI.getNumOperands() - 1).getImm() == 6) {
669       unsigned NewOpc;
670       switch (OutMI.getOpcode()) {
671       default: llvm_unreachable("Invalid opcode");
672       case X86::VPCMPBZ128rmi:   NewOpc = X86::VPCMPGTBZ128rm;   break;
673       case X86::VPCMPBZ128rmik:  NewOpc = X86::VPCMPGTBZ128rmk;  break;
674       case X86::VPCMPBZ128rri:   NewOpc = X86::VPCMPGTBZ128rr;   break;
675       case X86::VPCMPBZ128rrik:  NewOpc = X86::VPCMPGTBZ128rrk;  break;
676       case X86::VPCMPBZ256rmi:   NewOpc = X86::VPCMPGTBZ256rm;   break;
677       case X86::VPCMPBZ256rmik:  NewOpc = X86::VPCMPGTBZ256rmk;  break;
678       case X86::VPCMPBZ256rri:   NewOpc = X86::VPCMPGTBZ256rr;   break;
679       case X86::VPCMPBZ256rrik:  NewOpc = X86::VPCMPGTBZ256rrk;  break;
680       case X86::VPCMPBZrmi:      NewOpc = X86::VPCMPGTBZrm;      break;
681       case X86::VPCMPBZrmik:     NewOpc = X86::VPCMPGTBZrmk;     break;
682       case X86::VPCMPBZrri:      NewOpc = X86::VPCMPGTBZrr;      break;
683       case X86::VPCMPBZrrik:     NewOpc = X86::VPCMPGTBZrrk;     break;
684       case X86::VPCMPDZ128rmi:   NewOpc = X86::VPCMPGTDZ128rm;   break;
685       case X86::VPCMPDZ128rmib:  NewOpc = X86::VPCMPGTDZ128rmb;  break;
686       case X86::VPCMPDZ128rmibk: NewOpc = X86::VPCMPGTDZ128rmbk; break;
687       case X86::VPCMPDZ128rmik:  NewOpc = X86::VPCMPGTDZ128rmk;  break;
688       case X86::VPCMPDZ128rri:   NewOpc = X86::VPCMPGTDZ128rr;   break;
689       case X86::VPCMPDZ128rrik:  NewOpc = X86::VPCMPGTDZ128rrk;  break;
690       case X86::VPCMPDZ256rmi:   NewOpc = X86::VPCMPGTDZ256rm;   break;
691       case X86::VPCMPDZ256rmib:  NewOpc = X86::VPCMPGTDZ256rmb;  break;
692       case X86::VPCMPDZ256rmibk: NewOpc = X86::VPCMPGTDZ256rmbk; break;
693       case X86::VPCMPDZ256rmik:  NewOpc = X86::VPCMPGTDZ256rmk;  break;
694       case X86::VPCMPDZ256rri:   NewOpc = X86::VPCMPGTDZ256rr;   break;
695       case X86::VPCMPDZ256rrik:  NewOpc = X86::VPCMPGTDZ256rrk;  break;
696       case X86::VPCMPDZrmi:      NewOpc = X86::VPCMPGTDZrm;      break;
697       case X86::VPCMPDZrmib:     NewOpc = X86::VPCMPGTDZrmb;     break;
698       case X86::VPCMPDZrmibk:    NewOpc = X86::VPCMPGTDZrmbk;    break;
699       case X86::VPCMPDZrmik:     NewOpc = X86::VPCMPGTDZrmk;     break;
700       case X86::VPCMPDZrri:      NewOpc = X86::VPCMPGTDZrr;      break;
701       case X86::VPCMPDZrrik:     NewOpc = X86::VPCMPGTDZrrk;     break;
702       case X86::VPCMPQZ128rmi:   NewOpc = X86::VPCMPGTQZ128rm;   break;
703       case X86::VPCMPQZ128rmib:  NewOpc = X86::VPCMPGTQZ128rmb;  break;
704       case X86::VPCMPQZ128rmibk: NewOpc = X86::VPCMPGTQZ128rmbk; break;
705       case X86::VPCMPQZ128rmik:  NewOpc = X86::VPCMPGTQZ128rmk;  break;
706       case X86::VPCMPQZ128rri:   NewOpc = X86::VPCMPGTQZ128rr;   break;
707       case X86::VPCMPQZ128rrik:  NewOpc = X86::VPCMPGTQZ128rrk;  break;
708       case X86::VPCMPQZ256rmi:   NewOpc = X86::VPCMPGTQZ256rm;   break;
709       case X86::VPCMPQZ256rmib:  NewOpc = X86::VPCMPGTQZ256rmb;  break;
710       case X86::VPCMPQZ256rmibk: NewOpc = X86::VPCMPGTQZ256rmbk; break;
711       case X86::VPCMPQZ256rmik:  NewOpc = X86::VPCMPGTQZ256rmk;  break;
712       case X86::VPCMPQZ256rri:   NewOpc = X86::VPCMPGTQZ256rr;   break;
713       case X86::VPCMPQZ256rrik:  NewOpc = X86::VPCMPGTQZ256rrk;  break;
714       case X86::VPCMPQZrmi:      NewOpc = X86::VPCMPGTQZrm;      break;
715       case X86::VPCMPQZrmib:     NewOpc = X86::VPCMPGTQZrmb;     break;
716       case X86::VPCMPQZrmibk:    NewOpc = X86::VPCMPGTQZrmbk;    break;
717       case X86::VPCMPQZrmik:     NewOpc = X86::VPCMPGTQZrmk;     break;
718       case X86::VPCMPQZrri:      NewOpc = X86::VPCMPGTQZrr;      break;
719       case X86::VPCMPQZrrik:     NewOpc = X86::VPCMPGTQZrrk;     break;
720       case X86::VPCMPWZ128rmi:   NewOpc = X86::VPCMPGTWZ128rm;   break;
721       case X86::VPCMPWZ128rmik:  NewOpc = X86::VPCMPGTWZ128rmk;  break;
722       case X86::VPCMPWZ128rri:   NewOpc = X86::VPCMPGTWZ128rr;   break;
723       case X86::VPCMPWZ128rrik:  NewOpc = X86::VPCMPGTWZ128rrk;  break;
724       case X86::VPCMPWZ256rmi:   NewOpc = X86::VPCMPGTWZ256rm;   break;
725       case X86::VPCMPWZ256rmik:  NewOpc = X86::VPCMPGTWZ256rmk;  break;
726       case X86::VPCMPWZ256rri:   NewOpc = X86::VPCMPGTWZ256rr;   break;
727       case X86::VPCMPWZ256rrik:  NewOpc = X86::VPCMPGTWZ256rrk;  break;
728       case X86::VPCMPWZrmi:      NewOpc = X86::VPCMPGTWZrm;      break;
729       case X86::VPCMPWZrmik:     NewOpc = X86::VPCMPGTWZrmk;     break;
730       case X86::VPCMPWZrri:      NewOpc = X86::VPCMPGTWZrr;      break;
731       case X86::VPCMPWZrrik:     NewOpc = X86::VPCMPGTWZrrk;     break;
732       }
733 
734       OutMI.setOpcode(NewOpc);
735       OutMI.erase(&OutMI.getOperand(OutMI.getNumOperands() - 1));
736       break;
737     }
738 
739     break;
740   }
741 
742   // CALL64r, CALL64pcrel32 - These instructions used to have
743   // register inputs modeled as normal uses instead of implicit uses.  As such,
744   // they we used to truncate off all but the first operand (the callee). This
745   // issue seems to have been fixed at some point. This assert verifies that.
746   case X86::CALL64r:
747   case X86::CALL64pcrel32:
748     assert(OutMI.getNumOperands() == 1 && "Unexpected number of operands!");
749     break;
750 
751   case X86::EH_RETURN:
752   case X86::EH_RETURN64: {
753     OutMI = MCInst();
754     OutMI.setOpcode(getRetOpcode(AsmPrinter.getSubtarget()));
755     break;
756   }
757 
758   case X86::CLEANUPRET: {
759     // Replace CLEANUPRET with the appropriate RET.
760     OutMI = MCInst();
761     OutMI.setOpcode(getRetOpcode(AsmPrinter.getSubtarget()));
762     break;
763   }
764 
765   case X86::CATCHRET: {
766     // Replace CATCHRET with the appropriate RET.
767     const X86Subtarget &Subtarget = AsmPrinter.getSubtarget();
768     unsigned ReturnReg = Subtarget.is64Bit() ? X86::RAX : X86::EAX;
769     OutMI = MCInst();
770     OutMI.setOpcode(getRetOpcode(Subtarget));
771     OutMI.addOperand(MCOperand::createReg(ReturnReg));
772     break;
773   }
774 
775   // TAILJMPd, TAILJMPd64, TailJMPd_cc - Lower to the correct jump
776   // instruction.
777   case X86::TAILJMPr:
778   case X86::TAILJMPr64:
779   case X86::TAILJMPr64_REX:
780   case X86::TAILJMPd:
781   case X86::TAILJMPd64:
782     assert(OutMI.getNumOperands() == 1 && "Unexpected number of operands!");
783     OutMI.setOpcode(convertTailJumpOpcode(OutMI.getOpcode()));
784     break;
785 
786   case X86::TAILJMPd_CC:
787   case X86::TAILJMPd64_CC:
788     assert(OutMI.getNumOperands() == 2 && "Unexpected number of operands!");
789     OutMI.setOpcode(convertTailJumpOpcode(OutMI.getOpcode()));
790     break;
791 
792   case X86::TAILJMPm:
793   case X86::TAILJMPm64:
794   case X86::TAILJMPm64_REX:
795     assert(OutMI.getNumOperands() == X86::AddrNumOperands &&
796            "Unexpected number of operands!");
797     OutMI.setOpcode(convertTailJumpOpcode(OutMI.getOpcode()));
798     break;
799 
800   case X86::DEC16r:
801   case X86::DEC32r:
802   case X86::INC16r:
803   case X86::INC32r:
804     // If we aren't in 64-bit mode we can use the 1-byte inc/dec instructions.
805     if (!AsmPrinter.getSubtarget().is64Bit()) {
806       unsigned Opcode;
807       switch (OutMI.getOpcode()) {
808       default: llvm_unreachable("Invalid opcode");
809       case X86::DEC16r: Opcode = X86::DEC16r_alt; break;
810       case X86::DEC32r: Opcode = X86::DEC32r_alt; break;
811       case X86::INC16r: Opcode = X86::INC16r_alt; break;
812       case X86::INC32r: Opcode = X86::INC32r_alt; break;
813       }
814       OutMI.setOpcode(Opcode);
815     }
816     break;
817 
818   // We don't currently select the correct instruction form for instructions
819   // which have a short %eax, etc. form. Handle this by custom lowering, for
820   // now.
821   //
822   // Note, we are currently not handling the following instructions:
823   // MOV64ao8, MOV64o8a
824   // XCHG16ar, XCHG32ar, XCHG64ar
825   case X86::MOV8mr_NOREX:
826   case X86::MOV8mr:
827   case X86::MOV8rm_NOREX:
828   case X86::MOV8rm:
829   case X86::MOV16mr:
830   case X86::MOV16rm:
831   case X86::MOV32mr:
832   case X86::MOV32rm: {
833     unsigned NewOpc;
834     switch (OutMI.getOpcode()) {
835     default: llvm_unreachable("Invalid opcode");
836     case X86::MOV8mr_NOREX:
837     case X86::MOV8mr:  NewOpc = X86::MOV8o32a; break;
838     case X86::MOV8rm_NOREX:
839     case X86::MOV8rm:  NewOpc = X86::MOV8ao32; break;
840     case X86::MOV16mr: NewOpc = X86::MOV16o32a; break;
841     case X86::MOV16rm: NewOpc = X86::MOV16ao32; break;
842     case X86::MOV32mr: NewOpc = X86::MOV32o32a; break;
843     case X86::MOV32rm: NewOpc = X86::MOV32ao32; break;
844     }
845     SimplifyShortMoveForm(AsmPrinter, OutMI, NewOpc);
846     break;
847   }
848 
849   case X86::ADC8ri: case X86::ADC16ri: case X86::ADC32ri: case X86::ADC64ri32:
850   case X86::ADD8ri: case X86::ADD16ri: case X86::ADD32ri: case X86::ADD64ri32:
851   case X86::AND8ri: case X86::AND16ri: case X86::AND32ri: case X86::AND64ri32:
852   case X86::CMP8ri: case X86::CMP16ri: case X86::CMP32ri: case X86::CMP64ri32:
853   case X86::OR8ri:  case X86::OR16ri:  case X86::OR32ri:  case X86::OR64ri32:
854   case X86::SBB8ri: case X86::SBB16ri: case X86::SBB32ri: case X86::SBB64ri32:
855   case X86::SUB8ri: case X86::SUB16ri: case X86::SUB32ri: case X86::SUB64ri32:
856   case X86::TEST8ri:case X86::TEST16ri:case X86::TEST32ri:case X86::TEST64ri32:
857   case X86::XOR8ri: case X86::XOR16ri: case X86::XOR32ri: case X86::XOR64ri32: {
858     unsigned NewOpc;
859     switch (OutMI.getOpcode()) {
860     default: llvm_unreachable("Invalid opcode");
861     case X86::ADC8ri:     NewOpc = X86::ADC8i8;    break;
862     case X86::ADC16ri:    NewOpc = X86::ADC16i16;  break;
863     case X86::ADC32ri:    NewOpc = X86::ADC32i32;  break;
864     case X86::ADC64ri32:  NewOpc = X86::ADC64i32;  break;
865     case X86::ADD8ri:     NewOpc = X86::ADD8i8;    break;
866     case X86::ADD16ri:    NewOpc = X86::ADD16i16;  break;
867     case X86::ADD32ri:    NewOpc = X86::ADD32i32;  break;
868     case X86::ADD64ri32:  NewOpc = X86::ADD64i32;  break;
869     case X86::AND8ri:     NewOpc = X86::AND8i8;    break;
870     case X86::AND16ri:    NewOpc = X86::AND16i16;  break;
871     case X86::AND32ri:    NewOpc = X86::AND32i32;  break;
872     case X86::AND64ri32:  NewOpc = X86::AND64i32;  break;
873     case X86::CMP8ri:     NewOpc = X86::CMP8i8;    break;
874     case X86::CMP16ri:    NewOpc = X86::CMP16i16;  break;
875     case X86::CMP32ri:    NewOpc = X86::CMP32i32;  break;
876     case X86::CMP64ri32:  NewOpc = X86::CMP64i32;  break;
877     case X86::OR8ri:      NewOpc = X86::OR8i8;     break;
878     case X86::OR16ri:     NewOpc = X86::OR16i16;   break;
879     case X86::OR32ri:     NewOpc = X86::OR32i32;   break;
880     case X86::OR64ri32:   NewOpc = X86::OR64i32;   break;
881     case X86::SBB8ri:     NewOpc = X86::SBB8i8;    break;
882     case X86::SBB16ri:    NewOpc = X86::SBB16i16;  break;
883     case X86::SBB32ri:    NewOpc = X86::SBB32i32;  break;
884     case X86::SBB64ri32:  NewOpc = X86::SBB64i32;  break;
885     case X86::SUB8ri:     NewOpc = X86::SUB8i8;    break;
886     case X86::SUB16ri:    NewOpc = X86::SUB16i16;  break;
887     case X86::SUB32ri:    NewOpc = X86::SUB32i32;  break;
888     case X86::SUB64ri32:  NewOpc = X86::SUB64i32;  break;
889     case X86::TEST8ri:    NewOpc = X86::TEST8i8;   break;
890     case X86::TEST16ri:   NewOpc = X86::TEST16i16; break;
891     case X86::TEST32ri:   NewOpc = X86::TEST32i32; break;
892     case X86::TEST64ri32: NewOpc = X86::TEST64i32; break;
893     case X86::XOR8ri:     NewOpc = X86::XOR8i8;    break;
894     case X86::XOR16ri:    NewOpc = X86::XOR16i16;  break;
895     case X86::XOR32ri:    NewOpc = X86::XOR32i32;  break;
896     case X86::XOR64ri32:  NewOpc = X86::XOR64i32;  break;
897     }
898     SimplifyShortImmForm(OutMI, NewOpc);
899     break;
900   }
901 
902   // Try to shrink some forms of movsx.
903   case X86::MOVSX16rr8:
904   case X86::MOVSX32rr16:
905   case X86::MOVSX64rr32:
906     SimplifyMOVSX(OutMI);
907     break;
908 
909   case X86::VCMPPDrri:
910   case X86::VCMPPDYrri:
911   case X86::VCMPPSrri:
912   case X86::VCMPPSYrri:
913   case X86::VCMPSDrr:
914   case X86::VCMPSSrr: {
915     // Swap the operands if it will enable a 2 byte VEX encoding.
916     // FIXME: Change the immediate to improve opportunities?
917     if (!X86II::isX86_64ExtendedReg(OutMI.getOperand(1).getReg()) &&
918         X86II::isX86_64ExtendedReg(OutMI.getOperand(2).getReg())) {
919       unsigned Imm = MI->getOperand(3).getImm() & 0x7;
920       switch (Imm) {
921       default: break;
922       case 0x00: // EQUAL
923       case 0x03: // UNORDERED
924       case 0x04: // NOT EQUAL
925       case 0x07: // ORDERED
926         std::swap(OutMI.getOperand(1), OutMI.getOperand(2));
927         break;
928       }
929     }
930     break;
931   }
932 
933   case X86::VMOVHLPSrr:
934   case X86::VUNPCKHPDrr:
935     // These are not truly commutable so hide them from the default case.
936     break;
937 
938   default: {
939     // If the instruction is a commutable arithmetic instruction we might be
940     // able to commute the operands to get a 2 byte VEX prefix.
941     uint64_t TSFlags = MI->getDesc().TSFlags;
942     if (MI->getDesc().isCommutable() &&
943         (TSFlags & X86II::EncodingMask) == X86II::VEX &&
944         (TSFlags & X86II::OpMapMask) == X86II::TB &&
945         (TSFlags & X86II::FormMask) == X86II::MRMSrcReg &&
946         !(TSFlags & X86II::VEX_W) && (TSFlags & X86II::VEX_4V) &&
947         OutMI.getNumOperands() == 3) {
948       if (!X86II::isX86_64ExtendedReg(OutMI.getOperand(1).getReg()) &&
949           X86II::isX86_64ExtendedReg(OutMI.getOperand(2).getReg()))
950         std::swap(OutMI.getOperand(1), OutMI.getOperand(2));
951     }
952     break;
953   }
954   }
955 }
956 
957 void X86AsmPrinter::LowerTlsAddr(X86MCInstLower &MCInstLowering,
958                                  const MachineInstr &MI) {
959   NoAutoPaddingScope NoPadScope(*OutStreamer);
960   bool Is64Bits = MI.getOpcode() == X86::TLS_addr64 ||
961                   MI.getOpcode() == X86::TLS_base_addr64;
962   MCContext &Ctx = OutStreamer->getContext();
963 
964   MCSymbolRefExpr::VariantKind SRVK;
965   switch (MI.getOpcode()) {
966   case X86::TLS_addr32:
967   case X86::TLS_addr64:
968     SRVK = MCSymbolRefExpr::VK_TLSGD;
969     break;
970   case X86::TLS_base_addr32:
971     SRVK = MCSymbolRefExpr::VK_TLSLDM;
972     break;
973   case X86::TLS_base_addr64:
974     SRVK = MCSymbolRefExpr::VK_TLSLD;
975     break;
976   default:
977     llvm_unreachable("unexpected opcode");
978   }
979 
980   const MCSymbolRefExpr *Sym = MCSymbolRefExpr::create(
981       MCInstLowering.GetSymbolFromOperand(MI.getOperand(3)), SRVK, Ctx);
982 
983   // As of binutils 2.32, ld has a bogus TLS relaxation error when the GD/LD
984   // code sequence using R_X86_64_GOTPCREL (instead of R_X86_64_GOTPCRELX) is
985   // attempted to be relaxed to IE/LE (binutils PR24784). Work around the bug by
986   // only using GOT when GOTPCRELX is enabled.
987   // TODO Delete the workaround when GOTPCRELX becomes commonplace.
988   bool UseGot = MMI->getModule()->getRtLibUseGOT() &&
989                 Ctx.getAsmInfo()->canRelaxRelocations();
990 
991   if (Is64Bits) {
992     bool NeedsPadding = SRVK == MCSymbolRefExpr::VK_TLSGD;
993     if (NeedsPadding)
994       EmitAndCountInstruction(MCInstBuilder(X86::DATA16_PREFIX));
995     EmitAndCountInstruction(MCInstBuilder(X86::LEA64r)
996                                 .addReg(X86::RDI)
997                                 .addReg(X86::RIP)
998                                 .addImm(1)
999                                 .addReg(0)
1000                                 .addExpr(Sym)
1001                                 .addReg(0));
1002     const MCSymbol *TlsGetAddr = Ctx.getOrCreateSymbol("__tls_get_addr");
1003     if (NeedsPadding) {
1004       if (!UseGot)
1005         EmitAndCountInstruction(MCInstBuilder(X86::DATA16_PREFIX));
1006       EmitAndCountInstruction(MCInstBuilder(X86::DATA16_PREFIX));
1007       EmitAndCountInstruction(MCInstBuilder(X86::REX64_PREFIX));
1008     }
1009     if (UseGot) {
1010       const MCExpr *Expr = MCSymbolRefExpr::create(
1011           TlsGetAddr, MCSymbolRefExpr::VK_GOTPCREL, Ctx);
1012       EmitAndCountInstruction(MCInstBuilder(X86::CALL64m)
1013                                   .addReg(X86::RIP)
1014                                   .addImm(1)
1015                                   .addReg(0)
1016                                   .addExpr(Expr)
1017                                   .addReg(0));
1018     } else {
1019       EmitAndCountInstruction(
1020           MCInstBuilder(X86::CALL64pcrel32)
1021               .addExpr(MCSymbolRefExpr::create(TlsGetAddr,
1022                                                MCSymbolRefExpr::VK_PLT, Ctx)));
1023     }
1024   } else {
1025     if (SRVK == MCSymbolRefExpr::VK_TLSGD && !UseGot) {
1026       EmitAndCountInstruction(MCInstBuilder(X86::LEA32r)
1027                                   .addReg(X86::EAX)
1028                                   .addReg(0)
1029                                   .addImm(1)
1030                                   .addReg(X86::EBX)
1031                                   .addExpr(Sym)
1032                                   .addReg(0));
1033     } else {
1034       EmitAndCountInstruction(MCInstBuilder(X86::LEA32r)
1035                                   .addReg(X86::EAX)
1036                                   .addReg(X86::EBX)
1037                                   .addImm(1)
1038                                   .addReg(0)
1039                                   .addExpr(Sym)
1040                                   .addReg(0));
1041     }
1042 
1043     const MCSymbol *TlsGetAddr = Ctx.getOrCreateSymbol("___tls_get_addr");
1044     if (UseGot) {
1045       const MCExpr *Expr =
1046           MCSymbolRefExpr::create(TlsGetAddr, MCSymbolRefExpr::VK_GOT, Ctx);
1047       EmitAndCountInstruction(MCInstBuilder(X86::CALL32m)
1048                                   .addReg(X86::EBX)
1049                                   .addImm(1)
1050                                   .addReg(0)
1051                                   .addExpr(Expr)
1052                                   .addReg(0));
1053     } else {
1054       EmitAndCountInstruction(
1055           MCInstBuilder(X86::CALLpcrel32)
1056               .addExpr(MCSymbolRefExpr::create(TlsGetAddr,
1057                                                MCSymbolRefExpr::VK_PLT, Ctx)));
1058     }
1059   }
1060 }
1061 
1062 /// Return the longest nop which can be efficiently decoded for the given
1063 /// target cpu.  15-bytes is the longest single NOP instruction, but some
1064 /// platforms can't decode the longest forms efficiently.
1065 static unsigned MaxLongNopLength(const MCSubtargetInfo &STI) {
1066   uint64_t MaxNopLength = 10;
1067   if (STI.getFeatureBits()[X86::ProcIntelSLM])
1068     MaxNopLength = 7;
1069   else if (STI.getFeatureBits()[X86::FeatureFast15ByteNOP])
1070     MaxNopLength = 15;
1071   else if (STI.getFeatureBits()[X86::FeatureFast11ByteNOP])
1072     MaxNopLength = 11;
1073   return MaxNopLength;
1074 }
1075 
1076 /// Emit the largest nop instruction smaller than or equal to \p NumBytes
1077 /// bytes.  Return the size of nop emitted.
1078 static unsigned EmitNop(MCStreamer &OS, unsigned NumBytes, bool Is64Bit,
1079                         const MCSubtargetInfo &STI) {
1080   if (!Is64Bit) {
1081     // TODO Do additional checking if the CPU supports multi-byte nops.
1082     OS.emitInstruction(MCInstBuilder(X86::NOOP), STI);
1083     return 1;
1084   }
1085 
1086   // Cap a single nop emission at the profitable value for the target
1087   NumBytes = std::min(NumBytes, MaxLongNopLength(STI));
1088 
1089   unsigned NopSize;
1090   unsigned Opc, BaseReg, ScaleVal, IndexReg, Displacement, SegmentReg;
1091   IndexReg = Displacement = SegmentReg = 0;
1092   BaseReg = X86::RAX;
1093   ScaleVal = 1;
1094   switch (NumBytes) {
1095   case 0:
1096     llvm_unreachable("Zero nops?");
1097     break;
1098   case 1:
1099     NopSize = 1;
1100     Opc = X86::NOOP;
1101     break;
1102   case 2:
1103     NopSize = 2;
1104     Opc = X86::XCHG16ar;
1105     break;
1106   case 3:
1107     NopSize = 3;
1108     Opc = X86::NOOPL;
1109     break;
1110   case 4:
1111     NopSize = 4;
1112     Opc = X86::NOOPL;
1113     Displacement = 8;
1114     break;
1115   case 5:
1116     NopSize = 5;
1117     Opc = X86::NOOPL;
1118     Displacement = 8;
1119     IndexReg = X86::RAX;
1120     break;
1121   case 6:
1122     NopSize = 6;
1123     Opc = X86::NOOPW;
1124     Displacement = 8;
1125     IndexReg = X86::RAX;
1126     break;
1127   case 7:
1128     NopSize = 7;
1129     Opc = X86::NOOPL;
1130     Displacement = 512;
1131     break;
1132   case 8:
1133     NopSize = 8;
1134     Opc = X86::NOOPL;
1135     Displacement = 512;
1136     IndexReg = X86::RAX;
1137     break;
1138   case 9:
1139     NopSize = 9;
1140     Opc = X86::NOOPW;
1141     Displacement = 512;
1142     IndexReg = X86::RAX;
1143     break;
1144   default:
1145     NopSize = 10;
1146     Opc = X86::NOOPW;
1147     Displacement = 512;
1148     IndexReg = X86::RAX;
1149     SegmentReg = X86::CS;
1150     break;
1151   }
1152 
1153   unsigned NumPrefixes = std::min(NumBytes - NopSize, 5U);
1154   NopSize += NumPrefixes;
1155   for (unsigned i = 0; i != NumPrefixes; ++i)
1156     OS.emitBytes("\x66");
1157 
1158   switch (Opc) {
1159   default: llvm_unreachable("Unexpected opcode");
1160   case X86::NOOP:
1161     OS.emitInstruction(MCInstBuilder(Opc), STI);
1162     break;
1163   case X86::XCHG16ar:
1164     OS.emitInstruction(MCInstBuilder(Opc).addReg(X86::AX).addReg(X86::AX), STI);
1165     break;
1166   case X86::NOOPL:
1167   case X86::NOOPW:
1168     OS.emitInstruction(MCInstBuilder(Opc)
1169                            .addReg(BaseReg)
1170                            .addImm(ScaleVal)
1171                            .addReg(IndexReg)
1172                            .addImm(Displacement)
1173                            .addReg(SegmentReg),
1174                        STI);
1175     break;
1176   }
1177   assert(NopSize <= NumBytes && "We overemitted?");
1178   return NopSize;
1179 }
1180 
1181 /// Emit the optimal amount of multi-byte nops on X86.
1182 static void EmitNops(MCStreamer &OS, unsigned NumBytes, bool Is64Bit,
1183                      const MCSubtargetInfo &STI) {
1184   unsigned NopsToEmit = NumBytes;
1185   (void)NopsToEmit;
1186   while (NumBytes) {
1187     NumBytes -= EmitNop(OS, NumBytes, Is64Bit, STI);
1188     assert(NopsToEmit >= NumBytes && "Emitted more than I asked for!");
1189   }
1190 }
1191 
1192 void X86AsmPrinter::LowerSTATEPOINT(const MachineInstr &MI,
1193                                     X86MCInstLower &MCIL) {
1194   assert(Subtarget->is64Bit() && "Statepoint currently only supports X86-64");
1195 
1196   NoAutoPaddingScope NoPadScope(*OutStreamer);
1197 
1198   StatepointOpers SOpers(&MI);
1199   if (unsigned PatchBytes = SOpers.getNumPatchBytes()) {
1200     EmitNops(*OutStreamer, PatchBytes, Subtarget->is64Bit(),
1201              getSubtargetInfo());
1202   } else {
1203     // Lower call target and choose correct opcode
1204     const MachineOperand &CallTarget = SOpers.getCallTarget();
1205     MCOperand CallTargetMCOp;
1206     unsigned CallOpcode;
1207     switch (CallTarget.getType()) {
1208     case MachineOperand::MO_GlobalAddress:
1209     case MachineOperand::MO_ExternalSymbol:
1210       CallTargetMCOp = MCIL.LowerSymbolOperand(
1211           CallTarget, MCIL.GetSymbolFromOperand(CallTarget));
1212       CallOpcode = X86::CALL64pcrel32;
1213       // Currently, we only support relative addressing with statepoints.
1214       // Otherwise, we'll need a scratch register to hold the target
1215       // address.  You'll fail asserts during load & relocation if this
1216       // symbol is to far away. (TODO: support non-relative addressing)
1217       break;
1218     case MachineOperand::MO_Immediate:
1219       CallTargetMCOp = MCOperand::createImm(CallTarget.getImm());
1220       CallOpcode = X86::CALL64pcrel32;
1221       // Currently, we only support relative addressing with statepoints.
1222       // Otherwise, we'll need a scratch register to hold the target
1223       // immediate.  You'll fail asserts during load & relocation if this
1224       // address is to far away. (TODO: support non-relative addressing)
1225       break;
1226     case MachineOperand::MO_Register:
1227       // FIXME: Add retpoline support and remove this.
1228       if (Subtarget->useIndirectThunkCalls())
1229         report_fatal_error("Lowering register statepoints with thunks not "
1230                            "yet implemented.");
1231       CallTargetMCOp = MCOperand::createReg(CallTarget.getReg());
1232       CallOpcode = X86::CALL64r;
1233       break;
1234     default:
1235       llvm_unreachable("Unsupported operand type in statepoint call target");
1236       break;
1237     }
1238 
1239     // Emit call
1240     MCInst CallInst;
1241     CallInst.setOpcode(CallOpcode);
1242     CallInst.addOperand(CallTargetMCOp);
1243     OutStreamer->emitInstruction(CallInst, getSubtargetInfo());
1244   }
1245 
1246   // Record our statepoint node in the same section used by STACKMAP
1247   // and PATCHPOINT
1248   auto &Ctx = OutStreamer->getContext();
1249   MCSymbol *MILabel = Ctx.createTempSymbol();
1250   OutStreamer->emitLabel(MILabel);
1251   SM.recordStatepoint(*MILabel, MI);
1252 }
1253 
1254 void X86AsmPrinter::LowerFAULTING_OP(const MachineInstr &FaultingMI,
1255                                      X86MCInstLower &MCIL) {
1256   // FAULTING_LOAD_OP <def>, <faltinf type>, <MBB handler>,
1257   //                  <opcode>, <operands>
1258 
1259   NoAutoPaddingScope NoPadScope(*OutStreamer);
1260 
1261   Register DefRegister = FaultingMI.getOperand(0).getReg();
1262   FaultMaps::FaultKind FK =
1263       static_cast<FaultMaps::FaultKind>(FaultingMI.getOperand(1).getImm());
1264   MCSymbol *HandlerLabel = FaultingMI.getOperand(2).getMBB()->getSymbol();
1265   unsigned Opcode = FaultingMI.getOperand(3).getImm();
1266   unsigned OperandsBeginIdx = 4;
1267 
1268   auto &Ctx = OutStreamer->getContext();
1269   MCSymbol *FaultingLabel = Ctx.createTempSymbol();
1270   OutStreamer->emitLabel(FaultingLabel);
1271 
1272   assert(FK < FaultMaps::FaultKindMax && "Invalid Faulting Kind!");
1273   FM.recordFaultingOp(FK, FaultingLabel, HandlerLabel);
1274 
1275   MCInst MI;
1276   MI.setOpcode(Opcode);
1277 
1278   if (DefRegister != X86::NoRegister)
1279     MI.addOperand(MCOperand::createReg(DefRegister));
1280 
1281   for (auto I = FaultingMI.operands_begin() + OperandsBeginIdx,
1282             E = FaultingMI.operands_end();
1283        I != E; ++I)
1284     if (auto MaybeOperand = MCIL.LowerMachineOperand(&FaultingMI, *I))
1285       MI.addOperand(MaybeOperand.getValue());
1286 
1287   OutStreamer->AddComment("on-fault: " + HandlerLabel->getName());
1288   OutStreamer->emitInstruction(MI, getSubtargetInfo());
1289 }
1290 
1291 void X86AsmPrinter::LowerFENTRY_CALL(const MachineInstr &MI,
1292                                      X86MCInstLower &MCIL) {
1293   bool Is64Bits = Subtarget->is64Bit();
1294   MCContext &Ctx = OutStreamer->getContext();
1295   MCSymbol *fentry = Ctx.getOrCreateSymbol("__fentry__");
1296   const MCSymbolRefExpr *Op =
1297       MCSymbolRefExpr::create(fentry, MCSymbolRefExpr::VK_None, Ctx);
1298 
1299   EmitAndCountInstruction(
1300       MCInstBuilder(Is64Bits ? X86::CALL64pcrel32 : X86::CALLpcrel32)
1301           .addExpr(Op));
1302 }
1303 
1304 void X86AsmPrinter::LowerPATCHABLE_OP(const MachineInstr &MI,
1305                                       X86MCInstLower &MCIL) {
1306   // PATCHABLE_OP minsize, opcode, operands
1307 
1308   NoAutoPaddingScope NoPadScope(*OutStreamer);
1309 
1310   unsigned MinSize = MI.getOperand(0).getImm();
1311   unsigned Opcode = MI.getOperand(1).getImm();
1312 
1313   MCInst MCI;
1314   MCI.setOpcode(Opcode);
1315   for (auto &MO : make_range(MI.operands_begin() + 2, MI.operands_end()))
1316     if (auto MaybeOperand = MCIL.LowerMachineOperand(&MI, MO))
1317       MCI.addOperand(MaybeOperand.getValue());
1318 
1319   SmallString<256> Code;
1320   SmallVector<MCFixup, 4> Fixups;
1321   raw_svector_ostream VecOS(Code);
1322   CodeEmitter->encodeInstruction(MCI, VecOS, Fixups, getSubtargetInfo());
1323 
1324   if (Code.size() < MinSize) {
1325     if (MinSize == 2 && Opcode == X86::PUSH64r) {
1326       // This is an optimization that lets us get away without emitting a nop in
1327       // many cases.
1328       //
1329       // NB! In some cases the encoding for PUSH64r (e.g. PUSH64r %r9) takes two
1330       // bytes too, so the check on MinSize is important.
1331       MCI.setOpcode(X86::PUSH64rmr);
1332     } else {
1333       unsigned NopSize = EmitNop(*OutStreamer, MinSize, Subtarget->is64Bit(),
1334                                  getSubtargetInfo());
1335       assert(NopSize == MinSize && "Could not implement MinSize!");
1336       (void)NopSize;
1337     }
1338   }
1339 
1340   OutStreamer->emitInstruction(MCI, getSubtargetInfo());
1341 }
1342 
1343 // Lower a stackmap of the form:
1344 // <id>, <shadowBytes>, ...
1345 void X86AsmPrinter::LowerSTACKMAP(const MachineInstr &MI) {
1346   SMShadowTracker.emitShadowPadding(*OutStreamer, getSubtargetInfo());
1347 
1348   auto &Ctx = OutStreamer->getContext();
1349   MCSymbol *MILabel = Ctx.createTempSymbol();
1350   OutStreamer->emitLabel(MILabel);
1351 
1352   SM.recordStackMap(*MILabel, MI);
1353   unsigned NumShadowBytes = MI.getOperand(1).getImm();
1354   SMShadowTracker.reset(NumShadowBytes);
1355 }
1356 
1357 // Lower a patchpoint of the form:
1358 // [<def>], <id>, <numBytes>, <target>, <numArgs>, <cc>, ...
1359 void X86AsmPrinter::LowerPATCHPOINT(const MachineInstr &MI,
1360                                     X86MCInstLower &MCIL) {
1361   assert(Subtarget->is64Bit() && "Patchpoint currently only supports X86-64");
1362 
1363   SMShadowTracker.emitShadowPadding(*OutStreamer, getSubtargetInfo());
1364 
1365   NoAutoPaddingScope NoPadScope(*OutStreamer);
1366 
1367   auto &Ctx = OutStreamer->getContext();
1368   MCSymbol *MILabel = Ctx.createTempSymbol();
1369   OutStreamer->emitLabel(MILabel);
1370   SM.recordPatchPoint(*MILabel, MI);
1371 
1372   PatchPointOpers opers(&MI);
1373   unsigned ScratchIdx = opers.getNextScratchIdx();
1374   unsigned EncodedBytes = 0;
1375   const MachineOperand &CalleeMO = opers.getCallTarget();
1376 
1377   // Check for null target. If target is non-null (i.e. is non-zero or is
1378   // symbolic) then emit a call.
1379   if (!(CalleeMO.isImm() && !CalleeMO.getImm())) {
1380     MCOperand CalleeMCOp;
1381     switch (CalleeMO.getType()) {
1382     default:
1383       /// FIXME: Add a verifier check for bad callee types.
1384       llvm_unreachable("Unrecognized callee operand type.");
1385     case MachineOperand::MO_Immediate:
1386       if (CalleeMO.getImm())
1387         CalleeMCOp = MCOperand::createImm(CalleeMO.getImm());
1388       break;
1389     case MachineOperand::MO_ExternalSymbol:
1390     case MachineOperand::MO_GlobalAddress:
1391       CalleeMCOp = MCIL.LowerSymbolOperand(CalleeMO,
1392                                            MCIL.GetSymbolFromOperand(CalleeMO));
1393       break;
1394     }
1395 
1396     // Emit MOV to materialize the target address and the CALL to target.
1397     // This is encoded with 12-13 bytes, depending on which register is used.
1398     Register ScratchReg = MI.getOperand(ScratchIdx).getReg();
1399     if (X86II::isX86_64ExtendedReg(ScratchReg))
1400       EncodedBytes = 13;
1401     else
1402       EncodedBytes = 12;
1403 
1404     EmitAndCountInstruction(
1405         MCInstBuilder(X86::MOV64ri).addReg(ScratchReg).addOperand(CalleeMCOp));
1406     // FIXME: Add retpoline support and remove this.
1407     if (Subtarget->useIndirectThunkCalls())
1408       report_fatal_error(
1409           "Lowering patchpoint with thunks not yet implemented.");
1410     EmitAndCountInstruction(MCInstBuilder(X86::CALL64r).addReg(ScratchReg));
1411   }
1412 
1413   // Emit padding.
1414   unsigned NumBytes = opers.getNumPatchBytes();
1415   assert(NumBytes >= EncodedBytes &&
1416          "Patchpoint can't request size less than the length of a call.");
1417 
1418   EmitNops(*OutStreamer, NumBytes - EncodedBytes, Subtarget->is64Bit(),
1419            getSubtargetInfo());
1420 }
1421 
1422 void X86AsmPrinter::LowerPATCHABLE_EVENT_CALL(const MachineInstr &MI,
1423                                               X86MCInstLower &MCIL) {
1424   assert(Subtarget->is64Bit() && "XRay custom events only supports X86-64");
1425 
1426   NoAutoPaddingScope NoPadScope(*OutStreamer);
1427 
1428   // We want to emit the following pattern, which follows the x86 calling
1429   // convention to prepare for the trampoline call to be patched in.
1430   //
1431   //   .p2align 1, ...
1432   // .Lxray_event_sled_N:
1433   //   jmp +N                        // jump across the instrumentation sled
1434   //   ...                           // set up arguments in register
1435   //   callq __xray_CustomEvent@plt  // force dependency to symbol
1436   //   ...
1437   //   <jump here>
1438   //
1439   // After patching, it would look something like:
1440   //
1441   //   nopw (2-byte nop)
1442   //   ...
1443   //   callq __xrayCustomEvent  // already lowered
1444   //   ...
1445   //
1446   // ---
1447   // First we emit the label and the jump.
1448   auto CurSled = OutContext.createTempSymbol("xray_event_sled_", true);
1449   OutStreamer->AddComment("# XRay Custom Event Log");
1450   OutStreamer->emitCodeAlignment(2);
1451   OutStreamer->emitLabel(CurSled);
1452 
1453   // Use a two-byte `jmp`. This version of JMP takes an 8-bit relative offset as
1454   // an operand (computed as an offset from the jmp instruction).
1455   // FIXME: Find another less hacky way do force the relative jump.
1456   OutStreamer->emitBinaryData("\xeb\x0f");
1457 
1458   // The default C calling convention will place two arguments into %rcx and
1459   // %rdx -- so we only work with those.
1460   const Register DestRegs[] = {X86::RDI, X86::RSI};
1461   bool UsedMask[] = {false, false};
1462   // Filled out in loop.
1463   Register SrcRegs[] = {0, 0};
1464 
1465   // Then we put the operands in the %rdi and %rsi registers. We spill the
1466   // values in the register before we clobber them, and mark them as used in
1467   // UsedMask. In case the arguments are already in the correct register, we use
1468   // emit nops appropriately sized to keep the sled the same size in every
1469   // situation.
1470   for (unsigned I = 0; I < MI.getNumOperands(); ++I)
1471     if (auto Op = MCIL.LowerMachineOperand(&MI, MI.getOperand(I))) {
1472       assert(Op->isReg() && "Only support arguments in registers");
1473       SrcRegs[I] = getX86SubSuperRegister(Op->getReg(), 64);
1474       if (SrcRegs[I] != DestRegs[I]) {
1475         UsedMask[I] = true;
1476         EmitAndCountInstruction(
1477             MCInstBuilder(X86::PUSH64r).addReg(DestRegs[I]));
1478       } else {
1479         EmitNops(*OutStreamer, 4, Subtarget->is64Bit(), getSubtargetInfo());
1480       }
1481     }
1482 
1483   // Now that the register values are stashed, mov arguments into place.
1484   // FIXME: This doesn't work if one of the later SrcRegs is equal to an
1485   // earlier DestReg. We will have already overwritten over the register before
1486   // we can copy from it.
1487   for (unsigned I = 0; I < MI.getNumOperands(); ++I)
1488     if (SrcRegs[I] != DestRegs[I])
1489       EmitAndCountInstruction(
1490           MCInstBuilder(X86::MOV64rr).addReg(DestRegs[I]).addReg(SrcRegs[I]));
1491 
1492   // We emit a hard dependency on the __xray_CustomEvent symbol, which is the
1493   // name of the trampoline to be implemented by the XRay runtime.
1494   auto TSym = OutContext.getOrCreateSymbol("__xray_CustomEvent");
1495   MachineOperand TOp = MachineOperand::CreateMCSymbol(TSym);
1496   if (isPositionIndependent())
1497     TOp.setTargetFlags(X86II::MO_PLT);
1498 
1499   // Emit the call instruction.
1500   EmitAndCountInstruction(MCInstBuilder(X86::CALL64pcrel32)
1501                               .addOperand(MCIL.LowerSymbolOperand(TOp, TSym)));
1502 
1503   // Restore caller-saved and used registers.
1504   for (unsigned I = sizeof UsedMask; I-- > 0;)
1505     if (UsedMask[I])
1506       EmitAndCountInstruction(MCInstBuilder(X86::POP64r).addReg(DestRegs[I]));
1507     else
1508       EmitNops(*OutStreamer, 1, Subtarget->is64Bit(), getSubtargetInfo());
1509 
1510   OutStreamer->AddComment("xray custom event end.");
1511 
1512   // Record the sled version. Version 0 of this sled was spelled differently, so
1513   // we let the runtime handle the different offsets we're using. Version 2
1514   // changed the absolute address to a PC-relative address.
1515   recordSled(CurSled, MI, SledKind::CUSTOM_EVENT, 2);
1516 }
1517 
1518 void X86AsmPrinter::LowerPATCHABLE_TYPED_EVENT_CALL(const MachineInstr &MI,
1519                                                     X86MCInstLower &MCIL) {
1520   assert(Subtarget->is64Bit() && "XRay typed events only supports X86-64");
1521 
1522   NoAutoPaddingScope NoPadScope(*OutStreamer);
1523 
1524   // We want to emit the following pattern, which follows the x86 calling
1525   // convention to prepare for the trampoline call to be patched in.
1526   //
1527   //   .p2align 1, ...
1528   // .Lxray_event_sled_N:
1529   //   jmp +N                        // jump across the instrumentation sled
1530   //   ...                           // set up arguments in register
1531   //   callq __xray_TypedEvent@plt  // force dependency to symbol
1532   //   ...
1533   //   <jump here>
1534   //
1535   // After patching, it would look something like:
1536   //
1537   //   nopw (2-byte nop)
1538   //   ...
1539   //   callq __xrayTypedEvent  // already lowered
1540   //   ...
1541   //
1542   // ---
1543   // First we emit the label and the jump.
1544   auto CurSled = OutContext.createTempSymbol("xray_typed_event_sled_", true);
1545   OutStreamer->AddComment("# XRay Typed Event Log");
1546   OutStreamer->emitCodeAlignment(2);
1547   OutStreamer->emitLabel(CurSled);
1548 
1549   // Use a two-byte `jmp`. This version of JMP takes an 8-bit relative offset as
1550   // an operand (computed as an offset from the jmp instruction).
1551   // FIXME: Find another less hacky way do force the relative jump.
1552   OutStreamer->emitBinaryData("\xeb\x14");
1553 
1554   // An x86-64 convention may place three arguments into %rcx, %rdx, and R8,
1555   // so we'll work with those. Or we may be called via SystemV, in which case
1556   // we don't have to do any translation.
1557   const Register DestRegs[] = {X86::RDI, X86::RSI, X86::RDX};
1558   bool UsedMask[] = {false, false, false};
1559 
1560   // Will fill out src regs in the loop.
1561   Register SrcRegs[] = {0, 0, 0};
1562 
1563   // Then we put the operands in the SystemV registers. We spill the values in
1564   // the registers before we clobber them, and mark them as used in UsedMask.
1565   // In case the arguments are already in the correct register, we emit nops
1566   // appropriately sized to keep the sled the same size in every situation.
1567   for (unsigned I = 0; I < MI.getNumOperands(); ++I)
1568     if (auto Op = MCIL.LowerMachineOperand(&MI, MI.getOperand(I))) {
1569       // TODO: Is register only support adequate?
1570       assert(Op->isReg() && "Only supports arguments in registers");
1571       SrcRegs[I] = getX86SubSuperRegister(Op->getReg(), 64);
1572       if (SrcRegs[I] != DestRegs[I]) {
1573         UsedMask[I] = true;
1574         EmitAndCountInstruction(
1575             MCInstBuilder(X86::PUSH64r).addReg(DestRegs[I]));
1576       } else {
1577         EmitNops(*OutStreamer, 4, Subtarget->is64Bit(), getSubtargetInfo());
1578       }
1579     }
1580 
1581   // In the above loop we only stash all of the destination registers or emit
1582   // nops if the arguments are already in the right place. Doing the actually
1583   // moving is postponed until after all the registers are stashed so nothing
1584   // is clobbers. We've already added nops to account for the size of mov and
1585   // push if the register is in the right place, so we only have to worry about
1586   // emitting movs.
1587   // FIXME: This doesn't work if one of the later SrcRegs is equal to an
1588   // earlier DestReg. We will have already overwritten over the register before
1589   // we can copy from it.
1590   for (unsigned I = 0; I < MI.getNumOperands(); ++I)
1591     if (UsedMask[I])
1592       EmitAndCountInstruction(
1593           MCInstBuilder(X86::MOV64rr).addReg(DestRegs[I]).addReg(SrcRegs[I]));
1594 
1595   // We emit a hard dependency on the __xray_TypedEvent symbol, which is the
1596   // name of the trampoline to be implemented by the XRay runtime.
1597   auto TSym = OutContext.getOrCreateSymbol("__xray_TypedEvent");
1598   MachineOperand TOp = MachineOperand::CreateMCSymbol(TSym);
1599   if (isPositionIndependent())
1600     TOp.setTargetFlags(X86II::MO_PLT);
1601 
1602   // Emit the call instruction.
1603   EmitAndCountInstruction(MCInstBuilder(X86::CALL64pcrel32)
1604                               .addOperand(MCIL.LowerSymbolOperand(TOp, TSym)));
1605 
1606   // Restore caller-saved and used registers.
1607   for (unsigned I = sizeof UsedMask; I-- > 0;)
1608     if (UsedMask[I])
1609       EmitAndCountInstruction(MCInstBuilder(X86::POP64r).addReg(DestRegs[I]));
1610     else
1611       EmitNops(*OutStreamer, 1, Subtarget->is64Bit(), getSubtargetInfo());
1612 
1613   OutStreamer->AddComment("xray typed event end.");
1614 
1615   // Record the sled version.
1616   recordSled(CurSled, MI, SledKind::TYPED_EVENT, 2);
1617 }
1618 
1619 void X86AsmPrinter::LowerPATCHABLE_FUNCTION_ENTER(const MachineInstr &MI,
1620                                                   X86MCInstLower &MCIL) {
1621 
1622   NoAutoPaddingScope NoPadScope(*OutStreamer);
1623 
1624   const Function &F = MF->getFunction();
1625   if (F.hasFnAttribute("patchable-function-entry")) {
1626     unsigned Num;
1627     if (F.getFnAttribute("patchable-function-entry")
1628             .getValueAsString()
1629             .getAsInteger(10, Num))
1630       return;
1631     EmitNops(*OutStreamer, Num, Subtarget->is64Bit(), getSubtargetInfo());
1632     return;
1633   }
1634   // We want to emit the following pattern:
1635   //
1636   //   .p2align 1, ...
1637   // .Lxray_sled_N:
1638   //   jmp .tmpN
1639   //   # 9 bytes worth of noops
1640   //
1641   // We need the 9 bytes because at runtime, we'd be patching over the full 11
1642   // bytes with the following pattern:
1643   //
1644   //   mov %r10, <function id, 32-bit>   // 6 bytes
1645   //   call <relative offset, 32-bits>   // 5 bytes
1646   //
1647   auto CurSled = OutContext.createTempSymbol("xray_sled_", true);
1648   OutStreamer->emitCodeAlignment(2);
1649   OutStreamer->emitLabel(CurSled);
1650 
1651   // Use a two-byte `jmp`. This version of JMP takes an 8-bit relative offset as
1652   // an operand (computed as an offset from the jmp instruction).
1653   // FIXME: Find another less hacky way do force the relative jump.
1654   OutStreamer->emitBytes("\xeb\x09");
1655   EmitNops(*OutStreamer, 9, Subtarget->is64Bit(), getSubtargetInfo());
1656   recordSled(CurSled, MI, SledKind::FUNCTION_ENTER, 2);
1657 }
1658 
1659 void X86AsmPrinter::LowerPATCHABLE_RET(const MachineInstr &MI,
1660                                        X86MCInstLower &MCIL) {
1661   NoAutoPaddingScope NoPadScope(*OutStreamer);
1662 
1663   // Since PATCHABLE_RET takes the opcode of the return statement as an
1664   // argument, we use that to emit the correct form of the RET that we want.
1665   // i.e. when we see this:
1666   //
1667   //   PATCHABLE_RET X86::RET ...
1668   //
1669   // We should emit the RET followed by sleds.
1670   //
1671   //   .p2align 1, ...
1672   // .Lxray_sled_N:
1673   //   ret  # or equivalent instruction
1674   //   # 10 bytes worth of noops
1675   //
1676   // This just makes sure that the alignment for the next instruction is 2.
1677   auto CurSled = OutContext.createTempSymbol("xray_sled_", true);
1678   OutStreamer->emitCodeAlignment(2);
1679   OutStreamer->emitLabel(CurSled);
1680   unsigned OpCode = MI.getOperand(0).getImm();
1681   MCInst Ret;
1682   Ret.setOpcode(OpCode);
1683   for (auto &MO : make_range(MI.operands_begin() + 1, MI.operands_end()))
1684     if (auto MaybeOperand = MCIL.LowerMachineOperand(&MI, MO))
1685       Ret.addOperand(MaybeOperand.getValue());
1686   OutStreamer->emitInstruction(Ret, getSubtargetInfo());
1687   EmitNops(*OutStreamer, 10, Subtarget->is64Bit(), getSubtargetInfo());
1688   recordSled(CurSled, MI, SledKind::FUNCTION_EXIT, 2);
1689 }
1690 
1691 void X86AsmPrinter::LowerPATCHABLE_TAIL_CALL(const MachineInstr &MI,
1692                                              X86MCInstLower &MCIL) {
1693   NoAutoPaddingScope NoPadScope(*OutStreamer);
1694 
1695   // Like PATCHABLE_RET, we have the actual instruction in the operands to this
1696   // instruction so we lower that particular instruction and its operands.
1697   // Unlike PATCHABLE_RET though, we put the sled before the JMP, much like how
1698   // we do it for PATCHABLE_FUNCTION_ENTER. The sled should be very similar to
1699   // the PATCHABLE_FUNCTION_ENTER case, followed by the lowering of the actual
1700   // tail call much like how we have it in PATCHABLE_RET.
1701   auto CurSled = OutContext.createTempSymbol("xray_sled_", true);
1702   OutStreamer->emitCodeAlignment(2);
1703   OutStreamer->emitLabel(CurSled);
1704   auto Target = OutContext.createTempSymbol();
1705 
1706   // Use a two-byte `jmp`. This version of JMP takes an 8-bit relative offset as
1707   // an operand (computed as an offset from the jmp instruction).
1708   // FIXME: Find another less hacky way do force the relative jump.
1709   OutStreamer->emitBytes("\xeb\x09");
1710   EmitNops(*OutStreamer, 9, Subtarget->is64Bit(), getSubtargetInfo());
1711   OutStreamer->emitLabel(Target);
1712   recordSled(CurSled, MI, SledKind::TAIL_CALL, 2);
1713 
1714   unsigned OpCode = MI.getOperand(0).getImm();
1715   OpCode = convertTailJumpOpcode(OpCode);
1716   MCInst TC;
1717   TC.setOpcode(OpCode);
1718 
1719   // Before emitting the instruction, add a comment to indicate that this is
1720   // indeed a tail call.
1721   OutStreamer->AddComment("TAILCALL");
1722   for (auto &MO : make_range(MI.operands_begin() + 1, MI.operands_end()))
1723     if (auto MaybeOperand = MCIL.LowerMachineOperand(&MI, MO))
1724       TC.addOperand(MaybeOperand.getValue());
1725   OutStreamer->emitInstruction(TC, getSubtargetInfo());
1726 }
1727 
1728 // Returns instruction preceding MBBI in MachineFunction.
1729 // If MBBI is the first instruction of the first basic block, returns null.
1730 static MachineBasicBlock::const_iterator
1731 PrevCrossBBInst(MachineBasicBlock::const_iterator MBBI) {
1732   const MachineBasicBlock *MBB = MBBI->getParent();
1733   while (MBBI == MBB->begin()) {
1734     if (MBB == &MBB->getParent()->front())
1735       return MachineBasicBlock::const_iterator();
1736     MBB = MBB->getPrevNode();
1737     MBBI = MBB->end();
1738   }
1739   --MBBI;
1740   return MBBI;
1741 }
1742 
1743 static const Constant *getConstantFromPool(const MachineInstr &MI,
1744                                            const MachineOperand &Op) {
1745   if (!Op.isCPI() || Op.getOffset() != 0)
1746     return nullptr;
1747 
1748   ArrayRef<MachineConstantPoolEntry> Constants =
1749       MI.getParent()->getParent()->getConstantPool()->getConstants();
1750   const MachineConstantPoolEntry &ConstantEntry = Constants[Op.getIndex()];
1751 
1752   // Bail if this is a machine constant pool entry, we won't be able to dig out
1753   // anything useful.
1754   if (ConstantEntry.isMachineConstantPoolEntry())
1755     return nullptr;
1756 
1757   const Constant *C = ConstantEntry.Val.ConstVal;
1758   assert((!C || ConstantEntry.getType() == C->getType()) &&
1759          "Expected a constant of the same type!");
1760   return C;
1761 }
1762 
1763 static std::string getShuffleComment(const MachineInstr *MI, unsigned SrcOp1Idx,
1764                                      unsigned SrcOp2Idx, ArrayRef<int> Mask) {
1765   std::string Comment;
1766 
1767   // Compute the name for a register. This is really goofy because we have
1768   // multiple instruction printers that could (in theory) use different
1769   // names. Fortunately most people use the ATT style (outside of Windows)
1770   // and they actually agree on register naming here. Ultimately, this is
1771   // a comment, and so its OK if it isn't perfect.
1772   auto GetRegisterName = [](unsigned RegNum) -> StringRef {
1773     return X86ATTInstPrinter::getRegisterName(RegNum);
1774   };
1775 
1776   const MachineOperand &DstOp = MI->getOperand(0);
1777   const MachineOperand &SrcOp1 = MI->getOperand(SrcOp1Idx);
1778   const MachineOperand &SrcOp2 = MI->getOperand(SrcOp2Idx);
1779 
1780   StringRef DstName = DstOp.isReg() ? GetRegisterName(DstOp.getReg()) : "mem";
1781   StringRef Src1Name =
1782       SrcOp1.isReg() ? GetRegisterName(SrcOp1.getReg()) : "mem";
1783   StringRef Src2Name =
1784       SrcOp2.isReg() ? GetRegisterName(SrcOp2.getReg()) : "mem";
1785 
1786   // One source operand, fix the mask to print all elements in one span.
1787   SmallVector<int, 8> ShuffleMask(Mask.begin(), Mask.end());
1788   if (Src1Name == Src2Name)
1789     for (int i = 0, e = ShuffleMask.size(); i != e; ++i)
1790       if (ShuffleMask[i] >= e)
1791         ShuffleMask[i] -= e;
1792 
1793   raw_string_ostream CS(Comment);
1794   CS << DstName;
1795 
1796   // Handle AVX512 MASK/MASXZ write mask comments.
1797   // MASK: zmmX {%kY}
1798   // MASKZ: zmmX {%kY} {z}
1799   if (SrcOp1Idx > 1) {
1800     assert((SrcOp1Idx == 2 || SrcOp1Idx == 3) && "Unexpected writemask");
1801 
1802     const MachineOperand &WriteMaskOp = MI->getOperand(SrcOp1Idx - 1);
1803     if (WriteMaskOp.isReg()) {
1804       CS << " {%" << GetRegisterName(WriteMaskOp.getReg()) << "}";
1805 
1806       if (SrcOp1Idx == 2) {
1807         CS << " {z}";
1808       }
1809     }
1810   }
1811 
1812   CS << " = ";
1813 
1814   for (int i = 0, e = ShuffleMask.size(); i != e; ++i) {
1815     if (i != 0)
1816       CS << ",";
1817     if (ShuffleMask[i] == SM_SentinelZero) {
1818       CS << "zero";
1819       continue;
1820     }
1821 
1822     // Otherwise, it must come from src1 or src2.  Print the span of elements
1823     // that comes from this src.
1824     bool isSrc1 = ShuffleMask[i] < (int)e;
1825     CS << (isSrc1 ? Src1Name : Src2Name) << '[';
1826 
1827     bool IsFirst = true;
1828     while (i != e && ShuffleMask[i] != SM_SentinelZero &&
1829            (ShuffleMask[i] < (int)e) == isSrc1) {
1830       if (!IsFirst)
1831         CS << ',';
1832       else
1833         IsFirst = false;
1834       if (ShuffleMask[i] == SM_SentinelUndef)
1835         CS << "u";
1836       else
1837         CS << ShuffleMask[i] % (int)e;
1838       ++i;
1839     }
1840     CS << ']';
1841     --i; // For loop increments element #.
1842   }
1843   CS.flush();
1844 
1845   return Comment;
1846 }
1847 
1848 static void printConstant(const APInt &Val, raw_ostream &CS) {
1849   if (Val.getBitWidth() <= 64) {
1850     CS << Val.getZExtValue();
1851   } else {
1852     // print multi-word constant as (w0,w1)
1853     CS << "(";
1854     for (int i = 0, N = Val.getNumWords(); i < N; ++i) {
1855       if (i > 0)
1856         CS << ",";
1857       CS << Val.getRawData()[i];
1858     }
1859     CS << ")";
1860   }
1861 }
1862 
1863 static void printConstant(const APFloat &Flt, raw_ostream &CS) {
1864   SmallString<32> Str;
1865   // Force scientific notation to distinquish from integers.
1866   Flt.toString(Str, 0, 0);
1867   CS << Str;
1868 }
1869 
1870 static void printConstant(const Constant *COp, raw_ostream &CS) {
1871   if (isa<UndefValue>(COp)) {
1872     CS << "u";
1873   } else if (auto *CI = dyn_cast<ConstantInt>(COp)) {
1874     printConstant(CI->getValue(), CS);
1875   } else if (auto *CF = dyn_cast<ConstantFP>(COp)) {
1876     printConstant(CF->getValueAPF(), CS);
1877   } else {
1878     CS << "?";
1879   }
1880 }
1881 
1882 void X86AsmPrinter::EmitSEHInstruction(const MachineInstr *MI) {
1883   assert(MF->hasWinCFI() && "SEH_ instruction in function without WinCFI?");
1884   assert(getSubtarget().isOSWindows() && "SEH_ instruction Windows only");
1885 
1886   // Use the .cv_fpo directives if we're emitting CodeView on 32-bit x86.
1887   if (EmitFPOData) {
1888     X86TargetStreamer *XTS =
1889         static_cast<X86TargetStreamer *>(OutStreamer->getTargetStreamer());
1890     switch (MI->getOpcode()) {
1891     case X86::SEH_PushReg:
1892       XTS->emitFPOPushReg(MI->getOperand(0).getImm());
1893       break;
1894     case X86::SEH_StackAlloc:
1895       XTS->emitFPOStackAlloc(MI->getOperand(0).getImm());
1896       break;
1897     case X86::SEH_StackAlign:
1898       XTS->emitFPOStackAlign(MI->getOperand(0).getImm());
1899       break;
1900     case X86::SEH_SetFrame:
1901       assert(MI->getOperand(1).getImm() == 0 &&
1902              ".cv_fpo_setframe takes no offset");
1903       XTS->emitFPOSetFrame(MI->getOperand(0).getImm());
1904       break;
1905     case X86::SEH_EndPrologue:
1906       XTS->emitFPOEndPrologue();
1907       break;
1908     case X86::SEH_SaveReg:
1909     case X86::SEH_SaveXMM:
1910     case X86::SEH_PushFrame:
1911       llvm_unreachable("SEH_ directive incompatible with FPO");
1912       break;
1913     default:
1914       llvm_unreachable("expected SEH_ instruction");
1915     }
1916     return;
1917   }
1918 
1919   // Otherwise, use the .seh_ directives for all other Windows platforms.
1920   switch (MI->getOpcode()) {
1921   case X86::SEH_PushReg:
1922     OutStreamer->EmitWinCFIPushReg(MI->getOperand(0).getImm());
1923     break;
1924 
1925   case X86::SEH_SaveReg:
1926     OutStreamer->EmitWinCFISaveReg(MI->getOperand(0).getImm(),
1927                                    MI->getOperand(1).getImm());
1928     break;
1929 
1930   case X86::SEH_SaveXMM:
1931     OutStreamer->EmitWinCFISaveXMM(MI->getOperand(0).getImm(),
1932                                    MI->getOperand(1).getImm());
1933     break;
1934 
1935   case X86::SEH_StackAlloc:
1936     OutStreamer->EmitWinCFIAllocStack(MI->getOperand(0).getImm());
1937     break;
1938 
1939   case X86::SEH_SetFrame:
1940     OutStreamer->EmitWinCFISetFrame(MI->getOperand(0).getImm(),
1941                                     MI->getOperand(1).getImm());
1942     break;
1943 
1944   case X86::SEH_PushFrame:
1945     OutStreamer->EmitWinCFIPushFrame(MI->getOperand(0).getImm());
1946     break;
1947 
1948   case X86::SEH_EndPrologue:
1949     OutStreamer->EmitWinCFIEndProlog();
1950     break;
1951 
1952   default:
1953     llvm_unreachable("expected SEH_ instruction");
1954   }
1955 }
1956 
1957 static unsigned getRegisterWidth(const MCOperandInfo &Info) {
1958   if (Info.RegClass == X86::VR128RegClassID ||
1959       Info.RegClass == X86::VR128XRegClassID)
1960     return 128;
1961   if (Info.RegClass == X86::VR256RegClassID ||
1962       Info.RegClass == X86::VR256XRegClassID)
1963     return 256;
1964   if (Info.RegClass == X86::VR512RegClassID)
1965     return 512;
1966   llvm_unreachable("Unknown register class!");
1967 }
1968 
1969 void X86AsmPrinter::emitInstruction(const MachineInstr *MI) {
1970   X86MCInstLower MCInstLowering(*MF, *this);
1971   const X86RegisterInfo *RI =
1972       MF->getSubtarget<X86Subtarget>().getRegisterInfo();
1973 
1974   // Add a comment about EVEX-2-VEX compression for AVX-512 instrs that
1975   // are compressed from EVEX encoding to VEX encoding.
1976   if (TM.Options.MCOptions.ShowMCEncoding) {
1977     if (MI->getAsmPrinterFlags() & X86::AC_EVEX_2_VEX)
1978       OutStreamer->AddComment("EVEX TO VEX Compression ", false);
1979   }
1980 
1981   switch (MI->getOpcode()) {
1982   case TargetOpcode::DBG_VALUE:
1983     llvm_unreachable("Should be handled target independently");
1984 
1985   // Emit nothing here but a comment if we can.
1986   case X86::Int_MemBarrier:
1987     OutStreamer->emitRawComment("MEMBARRIER");
1988     return;
1989 
1990   case X86::EH_RETURN:
1991   case X86::EH_RETURN64: {
1992     // Lower these as normal, but add some comments.
1993     Register Reg = MI->getOperand(0).getReg();
1994     OutStreamer->AddComment(StringRef("eh_return, addr: %") +
1995                             X86ATTInstPrinter::getRegisterName(Reg));
1996     break;
1997   }
1998   case X86::CLEANUPRET: {
1999     // Lower these as normal, but add some comments.
2000     OutStreamer->AddComment("CLEANUPRET");
2001     break;
2002   }
2003 
2004   case X86::CATCHRET: {
2005     // Lower these as normal, but add some comments.
2006     OutStreamer->AddComment("CATCHRET");
2007     break;
2008   }
2009 
2010   case X86::ENDBR32:
2011   case X86::ENDBR64: {
2012     // CurrentPatchableFunctionEntrySym can be CurrentFnBegin only for
2013     // -fpatchable-function-entry=N,0. The entry MBB is guaranteed to be
2014     // non-empty. If MI is the initial ENDBR, place the
2015     // __patchable_function_entries label after ENDBR.
2016     if (CurrentPatchableFunctionEntrySym &&
2017         CurrentPatchableFunctionEntrySym == CurrentFnBegin &&
2018         MI == &MF->front().front()) {
2019       MCInst Inst;
2020       MCInstLowering.Lower(MI, Inst);
2021       EmitAndCountInstruction(Inst);
2022       CurrentPatchableFunctionEntrySym = createTempSymbol("patch");
2023       OutStreamer->emitLabel(CurrentPatchableFunctionEntrySym);
2024       return;
2025     }
2026     break;
2027   }
2028 
2029   case X86::TAILJMPr:
2030   case X86::TAILJMPm:
2031   case X86::TAILJMPd:
2032   case X86::TAILJMPd_CC:
2033   case X86::TAILJMPr64:
2034   case X86::TAILJMPm64:
2035   case X86::TAILJMPd64:
2036   case X86::TAILJMPd64_CC:
2037   case X86::TAILJMPr64_REX:
2038   case X86::TAILJMPm64_REX:
2039     // Lower these as normal, but add some comments.
2040     OutStreamer->AddComment("TAILCALL");
2041     break;
2042 
2043   case X86::TLS_addr32:
2044   case X86::TLS_addr64:
2045   case X86::TLS_base_addr32:
2046   case X86::TLS_base_addr64:
2047     return LowerTlsAddr(MCInstLowering, *MI);
2048 
2049   case X86::MOVPC32r: {
2050     // This is a pseudo op for a two instruction sequence with a label, which
2051     // looks like:
2052     //     call "L1$pb"
2053     // "L1$pb":
2054     //     popl %esi
2055 
2056     // Emit the call.
2057     MCSymbol *PICBase = MF->getPICBaseSymbol();
2058     // FIXME: We would like an efficient form for this, so we don't have to do a
2059     // lot of extra uniquing.
2060     EmitAndCountInstruction(
2061         MCInstBuilder(X86::CALLpcrel32)
2062             .addExpr(MCSymbolRefExpr::create(PICBase, OutContext)));
2063 
2064     const X86FrameLowering *FrameLowering =
2065         MF->getSubtarget<X86Subtarget>().getFrameLowering();
2066     bool hasFP = FrameLowering->hasFP(*MF);
2067 
2068     // TODO: This is needed only if we require precise CFA.
2069     bool HasActiveDwarfFrame = OutStreamer->getNumFrameInfos() &&
2070                                !OutStreamer->getDwarfFrameInfos().back().End;
2071 
2072     int stackGrowth = -RI->getSlotSize();
2073 
2074     if (HasActiveDwarfFrame && !hasFP) {
2075       OutStreamer->emitCFIAdjustCfaOffset(-stackGrowth);
2076     }
2077 
2078     // Emit the label.
2079     OutStreamer->emitLabel(PICBase);
2080 
2081     // popl $reg
2082     EmitAndCountInstruction(
2083         MCInstBuilder(X86::POP32r).addReg(MI->getOperand(0).getReg()));
2084 
2085     if (HasActiveDwarfFrame && !hasFP) {
2086       OutStreamer->emitCFIAdjustCfaOffset(stackGrowth);
2087     }
2088     return;
2089   }
2090 
2091   case X86::ADD32ri: {
2092     // Lower the MO_GOT_ABSOLUTE_ADDRESS form of ADD32ri.
2093     if (MI->getOperand(2).getTargetFlags() != X86II::MO_GOT_ABSOLUTE_ADDRESS)
2094       break;
2095 
2096     // Okay, we have something like:
2097     //  EAX = ADD32ri EAX, MO_GOT_ABSOLUTE_ADDRESS(@MYGLOBAL)
2098 
2099     // For this, we want to print something like:
2100     //   MYGLOBAL + (. - PICBASE)
2101     // However, we can't generate a ".", so just emit a new label here and refer
2102     // to it.
2103     MCSymbol *DotSym = OutContext.createTempSymbol();
2104     OutStreamer->emitLabel(DotSym);
2105 
2106     // Now that we have emitted the label, lower the complex operand expression.
2107     MCSymbol *OpSym = MCInstLowering.GetSymbolFromOperand(MI->getOperand(2));
2108 
2109     const MCExpr *DotExpr = MCSymbolRefExpr::create(DotSym, OutContext);
2110     const MCExpr *PICBase =
2111         MCSymbolRefExpr::create(MF->getPICBaseSymbol(), OutContext);
2112     DotExpr = MCBinaryExpr::createSub(DotExpr, PICBase, OutContext);
2113 
2114     DotExpr = MCBinaryExpr::createAdd(
2115         MCSymbolRefExpr::create(OpSym, OutContext), DotExpr, OutContext);
2116 
2117     EmitAndCountInstruction(MCInstBuilder(X86::ADD32ri)
2118                                 .addReg(MI->getOperand(0).getReg())
2119                                 .addReg(MI->getOperand(1).getReg())
2120                                 .addExpr(DotExpr));
2121     return;
2122   }
2123   case TargetOpcode::STATEPOINT:
2124     return LowerSTATEPOINT(*MI, MCInstLowering);
2125 
2126   case TargetOpcode::FAULTING_OP:
2127     return LowerFAULTING_OP(*MI, MCInstLowering);
2128 
2129   case TargetOpcode::FENTRY_CALL:
2130     return LowerFENTRY_CALL(*MI, MCInstLowering);
2131 
2132   case TargetOpcode::PATCHABLE_OP:
2133     return LowerPATCHABLE_OP(*MI, MCInstLowering);
2134 
2135   case TargetOpcode::STACKMAP:
2136     return LowerSTACKMAP(*MI);
2137 
2138   case TargetOpcode::PATCHPOINT:
2139     return LowerPATCHPOINT(*MI, MCInstLowering);
2140 
2141   case TargetOpcode::PATCHABLE_FUNCTION_ENTER:
2142     return LowerPATCHABLE_FUNCTION_ENTER(*MI, MCInstLowering);
2143 
2144   case TargetOpcode::PATCHABLE_RET:
2145     return LowerPATCHABLE_RET(*MI, MCInstLowering);
2146 
2147   case TargetOpcode::PATCHABLE_TAIL_CALL:
2148     return LowerPATCHABLE_TAIL_CALL(*MI, MCInstLowering);
2149 
2150   case TargetOpcode::PATCHABLE_EVENT_CALL:
2151     return LowerPATCHABLE_EVENT_CALL(*MI, MCInstLowering);
2152 
2153   case TargetOpcode::PATCHABLE_TYPED_EVENT_CALL:
2154     return LowerPATCHABLE_TYPED_EVENT_CALL(*MI, MCInstLowering);
2155 
2156   case X86::MORESTACK_RET:
2157     EmitAndCountInstruction(MCInstBuilder(getRetOpcode(*Subtarget)));
2158     return;
2159 
2160   case X86::MORESTACK_RET_RESTORE_R10:
2161     // Return, then restore R10.
2162     EmitAndCountInstruction(MCInstBuilder(getRetOpcode(*Subtarget)));
2163     EmitAndCountInstruction(
2164         MCInstBuilder(X86::MOV64rr).addReg(X86::R10).addReg(X86::RAX));
2165     return;
2166 
2167   case X86::SEH_PushReg:
2168   case X86::SEH_SaveReg:
2169   case X86::SEH_SaveXMM:
2170   case X86::SEH_StackAlloc:
2171   case X86::SEH_StackAlign:
2172   case X86::SEH_SetFrame:
2173   case X86::SEH_PushFrame:
2174   case X86::SEH_EndPrologue:
2175     EmitSEHInstruction(MI);
2176     return;
2177 
2178   case X86::SEH_Epilogue: {
2179     assert(MF->hasWinCFI() && "SEH_ instruction in function without WinCFI?");
2180     MachineBasicBlock::const_iterator MBBI(MI);
2181     // Check if preceded by a call and emit nop if so.
2182     for (MBBI = PrevCrossBBInst(MBBI);
2183          MBBI != MachineBasicBlock::const_iterator();
2184          MBBI = PrevCrossBBInst(MBBI)) {
2185       // Conservatively assume that pseudo instructions don't emit code and keep
2186       // looking for a call. We may emit an unnecessary nop in some cases.
2187       if (!MBBI->isPseudo()) {
2188         if (MBBI->isCall())
2189           EmitAndCountInstruction(MCInstBuilder(X86::NOOP));
2190         break;
2191       }
2192     }
2193     return;
2194   }
2195 
2196   // Lower PSHUFB and VPERMILP normally but add a comment if we can find
2197   // a constant shuffle mask. We won't be able to do this at the MC layer
2198   // because the mask isn't an immediate.
2199   case X86::PSHUFBrm:
2200   case X86::VPSHUFBrm:
2201   case X86::VPSHUFBYrm:
2202   case X86::VPSHUFBZ128rm:
2203   case X86::VPSHUFBZ128rmk:
2204   case X86::VPSHUFBZ128rmkz:
2205   case X86::VPSHUFBZ256rm:
2206   case X86::VPSHUFBZ256rmk:
2207   case X86::VPSHUFBZ256rmkz:
2208   case X86::VPSHUFBZrm:
2209   case X86::VPSHUFBZrmk:
2210   case X86::VPSHUFBZrmkz: {
2211     if (!OutStreamer->isVerboseAsm())
2212       break;
2213     unsigned SrcIdx, MaskIdx;
2214     switch (MI->getOpcode()) {
2215     default: llvm_unreachable("Invalid opcode");
2216     case X86::PSHUFBrm:
2217     case X86::VPSHUFBrm:
2218     case X86::VPSHUFBYrm:
2219     case X86::VPSHUFBZ128rm:
2220     case X86::VPSHUFBZ256rm:
2221     case X86::VPSHUFBZrm:
2222       SrcIdx = 1; MaskIdx = 5; break;
2223     case X86::VPSHUFBZ128rmkz:
2224     case X86::VPSHUFBZ256rmkz:
2225     case X86::VPSHUFBZrmkz:
2226       SrcIdx = 2; MaskIdx = 6; break;
2227     case X86::VPSHUFBZ128rmk:
2228     case X86::VPSHUFBZ256rmk:
2229     case X86::VPSHUFBZrmk:
2230       SrcIdx = 3; MaskIdx = 7; break;
2231     }
2232 
2233     assert(MI->getNumOperands() >= 6 &&
2234            "We should always have at least 6 operands!");
2235 
2236     const MachineOperand &MaskOp = MI->getOperand(MaskIdx);
2237     if (auto *C = getConstantFromPool(*MI, MaskOp)) {
2238       unsigned Width = getRegisterWidth(MI->getDesc().OpInfo[0]);
2239       SmallVector<int, 64> Mask;
2240       DecodePSHUFBMask(C, Width, Mask);
2241       if (!Mask.empty())
2242         OutStreamer->AddComment(getShuffleComment(MI, SrcIdx, SrcIdx, Mask));
2243     }
2244     break;
2245   }
2246 
2247   case X86::VPERMILPSrm:
2248   case X86::VPERMILPSYrm:
2249   case X86::VPERMILPSZ128rm:
2250   case X86::VPERMILPSZ128rmk:
2251   case X86::VPERMILPSZ128rmkz:
2252   case X86::VPERMILPSZ256rm:
2253   case X86::VPERMILPSZ256rmk:
2254   case X86::VPERMILPSZ256rmkz:
2255   case X86::VPERMILPSZrm:
2256   case X86::VPERMILPSZrmk:
2257   case X86::VPERMILPSZrmkz:
2258   case X86::VPERMILPDrm:
2259   case X86::VPERMILPDYrm:
2260   case X86::VPERMILPDZ128rm:
2261   case X86::VPERMILPDZ128rmk:
2262   case X86::VPERMILPDZ128rmkz:
2263   case X86::VPERMILPDZ256rm:
2264   case X86::VPERMILPDZ256rmk:
2265   case X86::VPERMILPDZ256rmkz:
2266   case X86::VPERMILPDZrm:
2267   case X86::VPERMILPDZrmk:
2268   case X86::VPERMILPDZrmkz: {
2269     if (!OutStreamer->isVerboseAsm())
2270       break;
2271     unsigned SrcIdx, MaskIdx;
2272     unsigned ElSize;
2273     switch (MI->getOpcode()) {
2274     default: llvm_unreachable("Invalid opcode");
2275     case X86::VPERMILPSrm:
2276     case X86::VPERMILPSYrm:
2277     case X86::VPERMILPSZ128rm:
2278     case X86::VPERMILPSZ256rm:
2279     case X86::VPERMILPSZrm:
2280       SrcIdx = 1; MaskIdx = 5; ElSize = 32; break;
2281     case X86::VPERMILPSZ128rmkz:
2282     case X86::VPERMILPSZ256rmkz:
2283     case X86::VPERMILPSZrmkz:
2284       SrcIdx = 2; MaskIdx = 6; ElSize = 32; break;
2285     case X86::VPERMILPSZ128rmk:
2286     case X86::VPERMILPSZ256rmk:
2287     case X86::VPERMILPSZrmk:
2288       SrcIdx = 3; MaskIdx = 7; ElSize = 32; break;
2289     case X86::VPERMILPDrm:
2290     case X86::VPERMILPDYrm:
2291     case X86::VPERMILPDZ128rm:
2292     case X86::VPERMILPDZ256rm:
2293     case X86::VPERMILPDZrm:
2294       SrcIdx = 1; MaskIdx = 5; ElSize = 64; break;
2295     case X86::VPERMILPDZ128rmkz:
2296     case X86::VPERMILPDZ256rmkz:
2297     case X86::VPERMILPDZrmkz:
2298       SrcIdx = 2; MaskIdx = 6; ElSize = 64; break;
2299     case X86::VPERMILPDZ128rmk:
2300     case X86::VPERMILPDZ256rmk:
2301     case X86::VPERMILPDZrmk:
2302       SrcIdx = 3; MaskIdx = 7; ElSize = 64; break;
2303     }
2304 
2305     assert(MI->getNumOperands() >= 6 &&
2306            "We should always have at least 6 operands!");
2307 
2308     const MachineOperand &MaskOp = MI->getOperand(MaskIdx);
2309     if (auto *C = getConstantFromPool(*MI, MaskOp)) {
2310       unsigned Width = getRegisterWidth(MI->getDesc().OpInfo[0]);
2311       SmallVector<int, 16> Mask;
2312       DecodeVPERMILPMask(C, ElSize, Width, Mask);
2313       if (!Mask.empty())
2314         OutStreamer->AddComment(getShuffleComment(MI, SrcIdx, SrcIdx, Mask));
2315     }
2316     break;
2317   }
2318 
2319   case X86::VPERMIL2PDrm:
2320   case X86::VPERMIL2PSrm:
2321   case X86::VPERMIL2PDYrm:
2322   case X86::VPERMIL2PSYrm: {
2323     if (!OutStreamer->isVerboseAsm())
2324       break;
2325     assert(MI->getNumOperands() >= 8 &&
2326            "We should always have at least 8 operands!");
2327 
2328     const MachineOperand &CtrlOp = MI->getOperand(MI->getNumOperands() - 1);
2329     if (!CtrlOp.isImm())
2330       break;
2331 
2332     unsigned ElSize;
2333     switch (MI->getOpcode()) {
2334     default: llvm_unreachable("Invalid opcode");
2335     case X86::VPERMIL2PSrm: case X86::VPERMIL2PSYrm: ElSize = 32; break;
2336     case X86::VPERMIL2PDrm: case X86::VPERMIL2PDYrm: ElSize = 64; break;
2337     }
2338 
2339     const MachineOperand &MaskOp = MI->getOperand(6);
2340     if (auto *C = getConstantFromPool(*MI, MaskOp)) {
2341       unsigned Width = getRegisterWidth(MI->getDesc().OpInfo[0]);
2342       SmallVector<int, 16> Mask;
2343       DecodeVPERMIL2PMask(C, (unsigned)CtrlOp.getImm(), ElSize, Width, Mask);
2344       if (!Mask.empty())
2345         OutStreamer->AddComment(getShuffleComment(MI, 1, 2, Mask));
2346     }
2347     break;
2348   }
2349 
2350   case X86::VPPERMrrm: {
2351     if (!OutStreamer->isVerboseAsm())
2352       break;
2353     assert(MI->getNumOperands() >= 7 &&
2354            "We should always have at least 7 operands!");
2355 
2356     const MachineOperand &MaskOp = MI->getOperand(6);
2357     if (auto *C = getConstantFromPool(*MI, MaskOp)) {
2358       unsigned Width = getRegisterWidth(MI->getDesc().OpInfo[0]);
2359       SmallVector<int, 16> Mask;
2360       DecodeVPPERMMask(C, Width, Mask);
2361       if (!Mask.empty())
2362         OutStreamer->AddComment(getShuffleComment(MI, 1, 2, Mask));
2363     }
2364     break;
2365   }
2366 
2367   case X86::MMX_MOVQ64rm: {
2368     if (!OutStreamer->isVerboseAsm())
2369       break;
2370     if (MI->getNumOperands() <= 4)
2371       break;
2372     if (auto *C = getConstantFromPool(*MI, MI->getOperand(4))) {
2373       std::string Comment;
2374       raw_string_ostream CS(Comment);
2375       const MachineOperand &DstOp = MI->getOperand(0);
2376       CS << X86ATTInstPrinter::getRegisterName(DstOp.getReg()) << " = ";
2377       if (auto *CF = dyn_cast<ConstantFP>(C)) {
2378         CS << "0x" << CF->getValueAPF().bitcastToAPInt().toString(16, false);
2379         OutStreamer->AddComment(CS.str());
2380       }
2381     }
2382     break;
2383   }
2384 
2385 #define MOV_CASE(Prefix, Suffix)                                               \
2386   case X86::Prefix##MOVAPD##Suffix##rm:                                        \
2387   case X86::Prefix##MOVAPS##Suffix##rm:                                        \
2388   case X86::Prefix##MOVUPD##Suffix##rm:                                        \
2389   case X86::Prefix##MOVUPS##Suffix##rm:                                        \
2390   case X86::Prefix##MOVDQA##Suffix##rm:                                        \
2391   case X86::Prefix##MOVDQU##Suffix##rm:
2392 
2393 #define MOV_AVX512_CASE(Suffix)                                                \
2394   case X86::VMOVDQA64##Suffix##rm:                                             \
2395   case X86::VMOVDQA32##Suffix##rm:                                             \
2396   case X86::VMOVDQU64##Suffix##rm:                                             \
2397   case X86::VMOVDQU32##Suffix##rm:                                             \
2398   case X86::VMOVDQU16##Suffix##rm:                                             \
2399   case X86::VMOVDQU8##Suffix##rm:                                              \
2400   case X86::VMOVAPS##Suffix##rm:                                               \
2401   case X86::VMOVAPD##Suffix##rm:                                               \
2402   case X86::VMOVUPS##Suffix##rm:                                               \
2403   case X86::VMOVUPD##Suffix##rm:
2404 
2405 #define CASE_ALL_MOV_RM()                                                      \
2406   MOV_CASE(, )   /* SSE */                                                     \
2407   MOV_CASE(V, )  /* AVX-128 */                                                 \
2408   MOV_CASE(V, Y) /* AVX-256 */                                                 \
2409   MOV_AVX512_CASE(Z)                                                           \
2410   MOV_AVX512_CASE(Z256)                                                        \
2411   MOV_AVX512_CASE(Z128)
2412 
2413     // For loads from a constant pool to a vector register, print the constant
2414     // loaded.
2415     CASE_ALL_MOV_RM()
2416   case X86::VBROADCASTF128:
2417   case X86::VBROADCASTI128:
2418   case X86::VBROADCASTF32X4Z256rm:
2419   case X86::VBROADCASTF32X4rm:
2420   case X86::VBROADCASTF32X8rm:
2421   case X86::VBROADCASTF64X2Z128rm:
2422   case X86::VBROADCASTF64X2rm:
2423   case X86::VBROADCASTF64X4rm:
2424   case X86::VBROADCASTI32X4Z256rm:
2425   case X86::VBROADCASTI32X4rm:
2426   case X86::VBROADCASTI32X8rm:
2427   case X86::VBROADCASTI64X2Z128rm:
2428   case X86::VBROADCASTI64X2rm:
2429   case X86::VBROADCASTI64X4rm:
2430     if (!OutStreamer->isVerboseAsm())
2431       break;
2432     if (MI->getNumOperands() <= 4)
2433       break;
2434     if (auto *C = getConstantFromPool(*MI, MI->getOperand(4))) {
2435       int NumLanes = 1;
2436       // Override NumLanes for the broadcast instructions.
2437       switch (MI->getOpcode()) {
2438       case X86::VBROADCASTF128:        NumLanes = 2; break;
2439       case X86::VBROADCASTI128:        NumLanes = 2; break;
2440       case X86::VBROADCASTF32X4Z256rm: NumLanes = 2; break;
2441       case X86::VBROADCASTF32X4rm:     NumLanes = 4; break;
2442       case X86::VBROADCASTF32X8rm:     NumLanes = 2; break;
2443       case X86::VBROADCASTF64X2Z128rm: NumLanes = 2; break;
2444       case X86::VBROADCASTF64X2rm:     NumLanes = 4; break;
2445       case X86::VBROADCASTF64X4rm:     NumLanes = 2; break;
2446       case X86::VBROADCASTI32X4Z256rm: NumLanes = 2; break;
2447       case X86::VBROADCASTI32X4rm:     NumLanes = 4; break;
2448       case X86::VBROADCASTI32X8rm:     NumLanes = 2; break;
2449       case X86::VBROADCASTI64X2Z128rm: NumLanes = 2; break;
2450       case X86::VBROADCASTI64X2rm:     NumLanes = 4; break;
2451       case X86::VBROADCASTI64X4rm:     NumLanes = 2; break;
2452       }
2453 
2454       std::string Comment;
2455       raw_string_ostream CS(Comment);
2456       const MachineOperand &DstOp = MI->getOperand(0);
2457       CS << X86ATTInstPrinter::getRegisterName(DstOp.getReg()) << " = ";
2458       if (auto *CDS = dyn_cast<ConstantDataSequential>(C)) {
2459         CS << "[";
2460         for (int l = 0; l != NumLanes; ++l) {
2461           for (int i = 0, NumElements = CDS->getNumElements(); i < NumElements;
2462                ++i) {
2463             if (i != 0 || l != 0)
2464               CS << ",";
2465             if (CDS->getElementType()->isIntegerTy())
2466               printConstant(CDS->getElementAsAPInt(i), CS);
2467             else if (CDS->getElementType()->isHalfTy() ||
2468                      CDS->getElementType()->isFloatTy() ||
2469                      CDS->getElementType()->isDoubleTy())
2470               printConstant(CDS->getElementAsAPFloat(i), CS);
2471             else
2472               CS << "?";
2473           }
2474         }
2475         CS << "]";
2476         OutStreamer->AddComment(CS.str());
2477       } else if (auto *CV = dyn_cast<ConstantVector>(C)) {
2478         CS << "<";
2479         for (int l = 0; l != NumLanes; ++l) {
2480           for (int i = 0, NumOperands = CV->getNumOperands(); i < NumOperands;
2481                ++i) {
2482             if (i != 0 || l != 0)
2483               CS << ",";
2484             printConstant(CV->getOperand(i), CS);
2485           }
2486         }
2487         CS << ">";
2488         OutStreamer->AddComment(CS.str());
2489       }
2490     }
2491     break;
2492   case X86::MOVDDUPrm:
2493   case X86::VMOVDDUPrm:
2494   case X86::VMOVDDUPZ128rm:
2495   case X86::VBROADCASTSSrm:
2496   case X86::VBROADCASTSSYrm:
2497   case X86::VBROADCASTSSZ128rm:
2498   case X86::VBROADCASTSSZ256rm:
2499   case X86::VBROADCASTSSZrm:
2500   case X86::VBROADCASTSDYrm:
2501   case X86::VBROADCASTSDZ256rm:
2502   case X86::VBROADCASTSDZrm:
2503   case X86::VPBROADCASTBrm:
2504   case X86::VPBROADCASTBYrm:
2505   case X86::VPBROADCASTBZ128rm:
2506   case X86::VPBROADCASTBZ256rm:
2507   case X86::VPBROADCASTBZrm:
2508   case X86::VPBROADCASTDrm:
2509   case X86::VPBROADCASTDYrm:
2510   case X86::VPBROADCASTDZ128rm:
2511   case X86::VPBROADCASTDZ256rm:
2512   case X86::VPBROADCASTDZrm:
2513   case X86::VPBROADCASTQrm:
2514   case X86::VPBROADCASTQYrm:
2515   case X86::VPBROADCASTQZ128rm:
2516   case X86::VPBROADCASTQZ256rm:
2517   case X86::VPBROADCASTQZrm:
2518   case X86::VPBROADCASTWrm:
2519   case X86::VPBROADCASTWYrm:
2520   case X86::VPBROADCASTWZ128rm:
2521   case X86::VPBROADCASTWZ256rm:
2522   case X86::VPBROADCASTWZrm:
2523     if (!OutStreamer->isVerboseAsm())
2524       break;
2525     if (MI->getNumOperands() <= 4)
2526       break;
2527     if (auto *C = getConstantFromPool(*MI, MI->getOperand(4))) {
2528       int NumElts;
2529       switch (MI->getOpcode()) {
2530       default: llvm_unreachable("Invalid opcode");
2531       case X86::MOVDDUPrm:          NumElts = 2;  break;
2532       case X86::VMOVDDUPrm:         NumElts = 2;  break;
2533       case X86::VMOVDDUPZ128rm:     NumElts = 2;  break;
2534       case X86::VBROADCASTSSrm:     NumElts = 4;  break;
2535       case X86::VBROADCASTSSYrm:    NumElts = 8;  break;
2536       case X86::VBROADCASTSSZ128rm: NumElts = 4;  break;
2537       case X86::VBROADCASTSSZ256rm: NumElts = 8;  break;
2538       case X86::VBROADCASTSSZrm:    NumElts = 16; break;
2539       case X86::VBROADCASTSDYrm:    NumElts = 4;  break;
2540       case X86::VBROADCASTSDZ256rm: NumElts = 4;  break;
2541       case X86::VBROADCASTSDZrm:    NumElts = 8;  break;
2542       case X86::VPBROADCASTBrm:     NumElts = 16; break;
2543       case X86::VPBROADCASTBYrm:    NumElts = 32; break;
2544       case X86::VPBROADCASTBZ128rm: NumElts = 16; break;
2545       case X86::VPBROADCASTBZ256rm: NumElts = 32; break;
2546       case X86::VPBROADCASTBZrm:    NumElts = 64; break;
2547       case X86::VPBROADCASTDrm:     NumElts = 4;  break;
2548       case X86::VPBROADCASTDYrm:    NumElts = 8;  break;
2549       case X86::VPBROADCASTDZ128rm: NumElts = 4;  break;
2550       case X86::VPBROADCASTDZ256rm: NumElts = 8;  break;
2551       case X86::VPBROADCASTDZrm:    NumElts = 16; break;
2552       case X86::VPBROADCASTQrm:     NumElts = 2;  break;
2553       case X86::VPBROADCASTQYrm:    NumElts = 4;  break;
2554       case X86::VPBROADCASTQZ128rm: NumElts = 2;  break;
2555       case X86::VPBROADCASTQZ256rm: NumElts = 4;  break;
2556       case X86::VPBROADCASTQZrm:    NumElts = 8;  break;
2557       case X86::VPBROADCASTWrm:     NumElts = 8;  break;
2558       case X86::VPBROADCASTWYrm:    NumElts = 16; break;
2559       case X86::VPBROADCASTWZ128rm: NumElts = 8;  break;
2560       case X86::VPBROADCASTWZ256rm: NumElts = 16; break;
2561       case X86::VPBROADCASTWZrm:    NumElts = 32; break;
2562       }
2563 
2564       std::string Comment;
2565       raw_string_ostream CS(Comment);
2566       const MachineOperand &DstOp = MI->getOperand(0);
2567       CS << X86ATTInstPrinter::getRegisterName(DstOp.getReg()) << " = ";
2568       CS << "[";
2569       for (int i = 0; i != NumElts; ++i) {
2570         if (i != 0)
2571           CS << ",";
2572         printConstant(C, CS);
2573       }
2574       CS << "]";
2575       OutStreamer->AddComment(CS.str());
2576     }
2577   }
2578 
2579   MCInst TmpInst;
2580   MCInstLowering.Lower(MI, TmpInst);
2581 
2582   // Stackmap shadows cannot include branch targets, so we can count the bytes
2583   // in a call towards the shadow, but must ensure that the no thread returns
2584   // in to the stackmap shadow.  The only way to achieve this is if the call
2585   // is at the end of the shadow.
2586   if (MI->isCall()) {
2587     // Count then size of the call towards the shadow
2588     SMShadowTracker.count(TmpInst, getSubtargetInfo(), CodeEmitter.get());
2589     // Then flush the shadow so that we fill with nops before the call, not
2590     // after it.
2591     SMShadowTracker.emitShadowPadding(*OutStreamer, getSubtargetInfo());
2592     // Then emit the call
2593     OutStreamer->emitInstruction(TmpInst, getSubtargetInfo());
2594     return;
2595   }
2596 
2597   EmitAndCountInstruction(TmpInst);
2598 }
2599