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/X86EncodingOptimization.h"
17#include "MCTargetDesc/X86InstComments.h"
18#include "MCTargetDesc/X86ShuffleDecode.h"
19#include "MCTargetDesc/X86TargetStreamer.h"
20#include "X86AsmPrinter.h"
21#include "X86MachineFunctionInfo.h"
22#include "X86RegisterInfo.h"
23#include "X86ShuffleDecodeConstantPool.h"
24#include "X86Subtarget.h"
25#include "llvm/ADT/STLExtras.h"
26#include "llvm/ADT/SmallString.h"
27#include "llvm/ADT/StringExtras.h"
28#include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
29#include "llvm/CodeGen/MachineConstantPool.h"
30#include "llvm/CodeGen/MachineFunction.h"
31#include "llvm/CodeGen/MachineModuleInfoImpls.h"
32#include "llvm/CodeGen/MachineOperand.h"
33#include "llvm/CodeGen/StackMaps.h"
34#include "llvm/IR/DataLayout.h"
35#include "llvm/IR/GlobalValue.h"
36#include "llvm/IR/Mangler.h"
37#include "llvm/MC/MCAsmInfo.h"
38#include "llvm/MC/MCCodeEmitter.h"
39#include "llvm/MC/MCContext.h"
40#include "llvm/MC/MCExpr.h"
41#include "llvm/MC/MCFixup.h"
42#include "llvm/MC/MCInst.h"
43#include "llvm/MC/MCInstBuilder.h"
44#include "llvm/MC/MCSection.h"
45#include "llvm/MC/MCSectionELF.h"
46#include "llvm/MC/MCStreamer.h"
47#include "llvm/MC/MCSymbol.h"
48#include "llvm/MC/MCSymbolELF.h"
49#include "llvm/MC/TargetRegistry.h"
50#include "llvm/Target/TargetLoweringObjectFile.h"
51#include "llvm/Target/TargetMachine.h"
52#include "llvm/Transforms/Instrumentation/AddressSanitizer.h"
53#include "llvm/Transforms/Instrumentation/AddressSanitizerCommon.h"
54#include <string>
55
56using namespace llvm;
57
58static cl::opt<bool> EnableBranchHint("enable-branch-hint",
59 cl::desc("Enable branch hint."),
60 cl::init(Val: false), cl::Hidden);
61static cl::opt<unsigned> BranchHintProbabilityThreshold(
62 "branch-hint-probability-threshold",
63 cl::desc("The probability threshold of enabling branch hint."),
64 cl::init(Val: 50), cl::Hidden);
65
66namespace {
67
68/// X86MCInstLower - This class is used to lower an MachineInstr into an MCInst.
69class X86MCInstLower {
70 MCContext &Ctx;
71 const MachineFunction &MF;
72 const TargetMachine &TM;
73 const MCAsmInfo &MAI;
74 X86AsmPrinter &AsmPrinter;
75
76public:
77 X86MCInstLower(const MachineFunction &MF, X86AsmPrinter &asmprinter);
78
79 MCOperand LowerMachineOperand(const MachineInstr *MI,
80 const MachineOperand &MO) const;
81 void Lower(const MachineInstr *MI, MCInst &OutMI) const;
82
83 MCSymbol *GetSymbolFromOperand(const MachineOperand &MO) const;
84 MCOperand LowerSymbolOperand(const MachineOperand &MO, MCSymbol *Sym) const;
85
86private:
87 MachineModuleInfoMachO &getMachOMMI() const;
88};
89
90} // end anonymous namespace
91
92/// A RAII helper which defines a region of instructions which can't have
93/// padding added between them for correctness.
94struct NoAutoPaddingScope {
95 MCStreamer &OS;
96 const bool OldAllowAutoPadding;
97 NoAutoPaddingScope(MCStreamer &OS)
98 : OS(OS), OldAllowAutoPadding(OS.getAllowAutoPadding()) {
99 changeAndComment(b: false);
100 }
101 ~NoAutoPaddingScope() { changeAndComment(b: OldAllowAutoPadding); }
102 void changeAndComment(bool b) {
103 if (b == OS.getAllowAutoPadding())
104 return;
105 OS.setAllowAutoPadding(b);
106 if (b)
107 OS.emitRawComment(T: "autopadding");
108 else
109 OS.emitRawComment(T: "noautopadding");
110 }
111};
112
113// Emit a minimal sequence of nops spanning NumBytes bytes.
114static void emitX86Nops(MCStreamer &OS, unsigned NumBytes,
115 const X86Subtarget *Subtarget);
116
117void X86AsmPrinter::StackMapShadowTracker::count(MCInst &Inst,
118 const MCSubtargetInfo &STI,
119 MCCodeEmitter *CodeEmitter) {
120 if (InShadow) {
121 SmallString<256> Code;
122 SmallVector<MCFixup, 4> Fixups;
123 CodeEmitter->encodeInstruction(Inst, CB&: Code, Fixups, STI);
124 CurrentShadowSize += Code.size();
125 if (CurrentShadowSize >= RequiredShadowSize)
126 InShadow = false; // The shadow is big enough. Stop counting.
127 }
128}
129
130void X86AsmPrinter::StackMapShadowTracker::emitShadowPadding(
131 MCStreamer &OutStreamer, const MCSubtargetInfo &STI) {
132 if (InShadow && CurrentShadowSize < RequiredShadowSize) {
133 InShadow = false;
134 emitX86Nops(OS&: OutStreamer, NumBytes: RequiredShadowSize - CurrentShadowSize,
135 Subtarget: &MF->getSubtarget<X86Subtarget>());
136 }
137}
138
139void X86AsmPrinter::EmitAndCountInstruction(MCInst &Inst) {
140 OutStreamer->emitInstruction(Inst, STI: getSubtargetInfo());
141 SMShadowTracker.count(Inst, STI: getSubtargetInfo(), CodeEmitter: CodeEmitter.get());
142}
143
144X86MCInstLower::X86MCInstLower(const MachineFunction &mf,
145 X86AsmPrinter &asmprinter)
146 : Ctx(mf.getContext()), MF(mf), TM(mf.getTarget()), MAI(*TM.getMCAsmInfo()),
147 AsmPrinter(asmprinter) {}
148
149MachineModuleInfoMachO &X86MCInstLower::getMachOMMI() const {
150 return AsmPrinter.MMI->getObjFileInfo<MachineModuleInfoMachO>();
151}
152
153/// GetSymbolFromOperand - Lower an MO_GlobalAddress or MO_ExternalSymbol
154/// operand to an MCSymbol.
155MCSymbol *X86MCInstLower::GetSymbolFromOperand(const MachineOperand &MO) const {
156 const Triple &TT = TM.getTargetTriple();
157 if (MO.isGlobal() && TT.isOSBinFormatELF())
158 return AsmPrinter.getSymbolPreferLocal(GV: *MO.getGlobal());
159
160 const DataLayout &DL = MF.getDataLayout();
161 assert((MO.isGlobal() || MO.isSymbol() || MO.isMBB()) &&
162 "Isn't a symbol reference");
163
164 MCSymbol *Sym = nullptr;
165 SmallString<128> Name;
166 StringRef Suffix;
167
168 switch (MO.getTargetFlags()) {
169 case X86II::MO_DLLIMPORT:
170 // Handle dllimport linkage.
171 Name += "__imp_";
172 break;
173 case X86II::MO_COFFSTUB:
174 Name += ".refptr.";
175 break;
176 case X86II::MO_DARWIN_NONLAZY:
177 case X86II::MO_DARWIN_NONLAZY_PIC_BASE:
178 Suffix = "$non_lazy_ptr";
179 break;
180 }
181
182 if (!Suffix.empty())
183 Name += DL.getPrivateGlobalPrefix();
184
185 if (MO.isGlobal()) {
186 const GlobalValue *GV = MO.getGlobal();
187 AsmPrinter.getNameWithPrefix(Name, GV);
188 } else if (MO.isSymbol()) {
189 Mangler::getNameWithPrefix(OutName&: Name, GVName: MO.getSymbolName(), DL);
190 } else if (MO.isMBB()) {
191 assert(Suffix.empty());
192 Sym = MO.getMBB()->getSymbol();
193 }
194
195 Name += Suffix;
196 if (!Sym)
197 Sym = Ctx.getOrCreateSymbol(Name);
198
199 // If the target flags on the operand changes the name of the symbol, do that
200 // before we return the symbol.
201 switch (MO.getTargetFlags()) {
202 default:
203 break;
204 case X86II::MO_COFFSTUB: {
205 MachineModuleInfoCOFF &MMICOFF =
206 AsmPrinter.MMI->getObjFileInfo<MachineModuleInfoCOFF>();
207 MachineModuleInfoImpl::StubValueTy &StubSym = MMICOFF.getGVStubEntry(Sym);
208 if (!StubSym.getPointer()) {
209 assert(MO.isGlobal() && "Extern symbol not handled yet");
210 StubSym = MachineModuleInfoImpl::StubValueTy(
211 AsmPrinter.getSymbol(GV: MO.getGlobal()), true);
212 }
213 break;
214 }
215 case X86II::MO_DARWIN_NONLAZY:
216 case X86II::MO_DARWIN_NONLAZY_PIC_BASE: {
217 MachineModuleInfoImpl::StubValueTy &StubSym =
218 getMachOMMI().getGVStubEntry(Sym);
219 if (!StubSym.getPointer()) {
220 assert(MO.isGlobal() && "Extern symbol not handled yet");
221 StubSym = MachineModuleInfoImpl::StubValueTy(
222 AsmPrinter.getSymbol(GV: MO.getGlobal()),
223 !MO.getGlobal()->hasInternalLinkage());
224 }
225 break;
226 }
227 }
228
229 return Sym;
230}
231
232MCOperand X86MCInstLower::LowerSymbolOperand(const MachineOperand &MO,
233 MCSymbol *Sym) const {
234 // FIXME: We would like an efficient form for this, so we don't have to do a
235 // lot of extra uniquing.
236 const MCExpr *Expr = nullptr;
237 MCSymbolRefExpr::VariantKind RefKind = MCSymbolRefExpr::VK_None;
238
239 switch (MO.getTargetFlags()) {
240 default:
241 llvm_unreachable("Unknown target flag on GV operand");
242 case X86II::MO_NO_FLAG: // No flag.
243 // These affect the name of the symbol, not any suffix.
244 case X86II::MO_DARWIN_NONLAZY:
245 case X86II::MO_DLLIMPORT:
246 case X86II::MO_COFFSTUB:
247 break;
248
249 case X86II::MO_TLVP:
250 RefKind = MCSymbolRefExpr::VK_TLVP;
251 break;
252 case X86II::MO_TLVP_PIC_BASE:
253 Expr = MCSymbolRefExpr::create(Symbol: Sym, Kind: MCSymbolRefExpr::VK_TLVP, Ctx);
254 // Subtract the pic base.
255 Expr = MCBinaryExpr::createSub(
256 LHS: Expr, RHS: MCSymbolRefExpr::create(Symbol: MF.getPICBaseSymbol(), Ctx), Ctx);
257 break;
258 case X86II::MO_SECREL:
259 RefKind = MCSymbolRefExpr::VK_SECREL;
260 break;
261 case X86II::MO_TLSGD:
262 RefKind = MCSymbolRefExpr::VK_TLSGD;
263 break;
264 case X86II::MO_TLSLD:
265 RefKind = MCSymbolRefExpr::VK_TLSLD;
266 break;
267 case X86II::MO_TLSLDM:
268 RefKind = MCSymbolRefExpr::VK_TLSLDM;
269 break;
270 case X86II::MO_GOTTPOFF:
271 RefKind = MCSymbolRefExpr::VK_GOTTPOFF;
272 break;
273 case X86II::MO_INDNTPOFF:
274 RefKind = MCSymbolRefExpr::VK_INDNTPOFF;
275 break;
276 case X86II::MO_TPOFF:
277 RefKind = MCSymbolRefExpr::VK_TPOFF;
278 break;
279 case X86II::MO_DTPOFF:
280 RefKind = MCSymbolRefExpr::VK_DTPOFF;
281 break;
282 case X86II::MO_NTPOFF:
283 RefKind = MCSymbolRefExpr::VK_NTPOFF;
284 break;
285 case X86II::MO_GOTNTPOFF:
286 RefKind = MCSymbolRefExpr::VK_GOTNTPOFF;
287 break;
288 case X86II::MO_GOTPCREL:
289 RefKind = MCSymbolRefExpr::VK_GOTPCREL;
290 break;
291 case X86II::MO_GOTPCREL_NORELAX:
292 RefKind = MCSymbolRefExpr::VK_GOTPCREL_NORELAX;
293 break;
294 case X86II::MO_GOT:
295 RefKind = MCSymbolRefExpr::VK_GOT;
296 break;
297 case X86II::MO_GOTOFF:
298 RefKind = MCSymbolRefExpr::VK_GOTOFF;
299 break;
300 case X86II::MO_PLT:
301 RefKind = MCSymbolRefExpr::VK_PLT;
302 break;
303 case X86II::MO_ABS8:
304 RefKind = MCSymbolRefExpr::VK_X86_ABS8;
305 break;
306 case X86II::MO_PIC_BASE_OFFSET:
307 case X86II::MO_DARWIN_NONLAZY_PIC_BASE:
308 Expr = MCSymbolRefExpr::create(Symbol: Sym, Ctx);
309 // Subtract the pic base.
310 Expr = MCBinaryExpr::createSub(
311 LHS: Expr, RHS: MCSymbolRefExpr::create(Symbol: MF.getPICBaseSymbol(), Ctx), Ctx);
312 if (MO.isJTI()) {
313 assert(MAI.doesSetDirectiveSuppressReloc());
314 // If .set directive is supported, use it to reduce the number of
315 // relocations the assembler will generate for differences between
316 // local labels. This is only safe when the symbols are in the same
317 // section so we are restricting it to jumptable references.
318 MCSymbol *Label = Ctx.createTempSymbol();
319 AsmPrinter.OutStreamer->emitAssignment(Symbol: Label, Value: Expr);
320 Expr = MCSymbolRefExpr::create(Symbol: Label, Ctx);
321 }
322 break;
323 }
324
325 if (!Expr)
326 Expr = MCSymbolRefExpr::create(Symbol: Sym, Kind: RefKind, Ctx);
327
328 if (!MO.isJTI() && !MO.isMBB() && MO.getOffset())
329 Expr = MCBinaryExpr::createAdd(
330 LHS: Expr, RHS: MCConstantExpr::create(Value: MO.getOffset(), Ctx), Ctx);
331 return MCOperand::createExpr(Val: Expr);
332}
333
334static unsigned getRetOpcode(const X86Subtarget &Subtarget) {
335 return Subtarget.is64Bit() ? X86::RET64 : X86::RET32;
336}
337
338MCOperand X86MCInstLower::LowerMachineOperand(const MachineInstr *MI,
339 const MachineOperand &MO) const {
340 switch (MO.getType()) {
341 default:
342 MI->print(OS&: errs());
343 llvm_unreachable("unknown operand type");
344 case MachineOperand::MO_Register:
345 // Ignore all implicit register operands.
346 if (MO.isImplicit())
347 return MCOperand();
348 return MCOperand::createReg(Reg: MO.getReg());
349 case MachineOperand::MO_Immediate:
350 return MCOperand::createImm(Val: MO.getImm());
351 case MachineOperand::MO_MachineBasicBlock:
352 case MachineOperand::MO_GlobalAddress:
353 case MachineOperand::MO_ExternalSymbol:
354 return LowerSymbolOperand(MO, Sym: GetSymbolFromOperand(MO));
355 case MachineOperand::MO_MCSymbol:
356 return LowerSymbolOperand(MO, Sym: MO.getMCSymbol());
357 case MachineOperand::MO_JumpTableIndex:
358 return LowerSymbolOperand(MO, Sym: AsmPrinter.GetJTISymbol(JTID: MO.getIndex()));
359 case MachineOperand::MO_ConstantPoolIndex:
360 return LowerSymbolOperand(MO, Sym: AsmPrinter.GetCPISymbol(CPID: MO.getIndex()));
361 case MachineOperand::MO_BlockAddress:
362 return LowerSymbolOperand(
363 MO, Sym: AsmPrinter.GetBlockAddressSymbol(BA: MO.getBlockAddress()));
364 case MachineOperand::MO_RegisterMask:
365 // Ignore call clobbers.
366 return MCOperand();
367 }
368}
369
370// Replace TAILJMP opcodes with their equivalent opcodes that have encoding
371// information.
372static unsigned convertTailJumpOpcode(unsigned Opcode) {
373 switch (Opcode) {
374 case X86::TAILJMPr:
375 Opcode = X86::JMP32r;
376 break;
377 case X86::TAILJMPm:
378 Opcode = X86::JMP32m;
379 break;
380 case X86::TAILJMPr64:
381 Opcode = X86::JMP64r;
382 break;
383 case X86::TAILJMPm64:
384 Opcode = X86::JMP64m;
385 break;
386 case X86::TAILJMPr64_REX:
387 Opcode = X86::JMP64r_REX;
388 break;
389 case X86::TAILJMPm64_REX:
390 Opcode = X86::JMP64m_REX;
391 break;
392 case X86::TAILJMPd:
393 case X86::TAILJMPd64:
394 Opcode = X86::JMP_1;
395 break;
396 case X86::TAILJMPd_CC:
397 case X86::TAILJMPd64_CC:
398 Opcode = X86::JCC_1;
399 break;
400 }
401
402 return Opcode;
403}
404
405void X86MCInstLower::Lower(const MachineInstr *MI, MCInst &OutMI) const {
406 OutMI.setOpcode(MI->getOpcode());
407
408 for (const MachineOperand &MO : MI->operands())
409 if (auto Op = LowerMachineOperand(MI, MO); Op.isValid())
410 OutMI.addOperand(Op);
411
412 bool In64BitMode = AsmPrinter.getSubtarget().is64Bit();
413 if (X86::optimizeInstFromVEX3ToVEX2(MI&: OutMI, Desc: MI->getDesc()) ||
414 X86::optimizeShiftRotateWithImmediateOne(MI&: OutMI) ||
415 X86::optimizeVPCMPWithImmediateOneOrSix(MI&: OutMI) ||
416 X86::optimizeMOVSX(MI&: OutMI) || X86::optimizeINCDEC(MI&: OutMI, In64BitMode) ||
417 X86::optimizeMOV(MI&: OutMI, In64BitMode) ||
418 X86::optimizeToFixedRegisterOrShortImmediateForm(MI&: OutMI))
419 return;
420
421 // Handle a few special cases to eliminate operand modifiers.
422 switch (OutMI.getOpcode()) {
423 case X86::LEA64_32r:
424 case X86::LEA64r:
425 case X86::LEA16r:
426 case X86::LEA32r:
427 // LEA should have a segment register, but it must be empty.
428 assert(OutMI.getNumOperands() == 1 + X86::AddrNumOperands &&
429 "Unexpected # of LEA operands");
430 assert(OutMI.getOperand(1 + X86::AddrSegmentReg).getReg() == 0 &&
431 "LEA has segment specified!");
432 break;
433 case X86::MULX32Hrr:
434 case X86::MULX32Hrm:
435 case X86::MULX64Hrr:
436 case X86::MULX64Hrm: {
437 // Turn into regular MULX by duplicating the destination.
438 unsigned NewOpc;
439 switch (OutMI.getOpcode()) {
440 default: llvm_unreachable("Invalid opcode");
441 case X86::MULX32Hrr: NewOpc = X86::MULX32rr; break;
442 case X86::MULX32Hrm: NewOpc = X86::MULX32rm; break;
443 case X86::MULX64Hrr: NewOpc = X86::MULX64rr; break;
444 case X86::MULX64Hrm: NewOpc = X86::MULX64rm; break;
445 }
446 OutMI.setOpcode(NewOpc);
447 // Duplicate the destination.
448 unsigned DestReg = OutMI.getOperand(i: 0).getReg();
449 OutMI.insert(I: OutMI.begin(), Op: MCOperand::createReg(Reg: DestReg));
450 break;
451 }
452 // CALL64r, CALL64pcrel32 - These instructions used to have
453 // register inputs modeled as normal uses instead of implicit uses. As such,
454 // they we used to truncate off all but the first operand (the callee). This
455 // issue seems to have been fixed at some point. This assert verifies that.
456 case X86::CALL64r:
457 case X86::CALL64pcrel32:
458 assert(OutMI.getNumOperands() == 1 && "Unexpected number of operands!");
459 break;
460 case X86::EH_RETURN:
461 case X86::EH_RETURN64: {
462 OutMI = MCInst();
463 OutMI.setOpcode(getRetOpcode(Subtarget: AsmPrinter.getSubtarget()));
464 break;
465 }
466 case X86::CLEANUPRET: {
467 // Replace CLEANUPRET with the appropriate RET.
468 OutMI = MCInst();
469 OutMI.setOpcode(getRetOpcode(Subtarget: AsmPrinter.getSubtarget()));
470 break;
471 }
472 case X86::CATCHRET: {
473 // Replace CATCHRET with the appropriate RET.
474 const X86Subtarget &Subtarget = AsmPrinter.getSubtarget();
475 unsigned ReturnReg = In64BitMode ? X86::RAX : X86::EAX;
476 OutMI = MCInst();
477 OutMI.setOpcode(getRetOpcode(Subtarget));
478 OutMI.addOperand(Op: MCOperand::createReg(Reg: ReturnReg));
479 break;
480 }
481 // TAILJMPd, TAILJMPd64, TailJMPd_cc - Lower to the correct jump
482 // instruction.
483 case X86::TAILJMPr:
484 case X86::TAILJMPr64:
485 case X86::TAILJMPr64_REX:
486 case X86::TAILJMPd:
487 case X86::TAILJMPd64:
488 assert(OutMI.getNumOperands() == 1 && "Unexpected number of operands!");
489 OutMI.setOpcode(convertTailJumpOpcode(Opcode: OutMI.getOpcode()));
490 break;
491 case X86::TAILJMPd_CC:
492 case X86::TAILJMPd64_CC:
493 assert(OutMI.getNumOperands() == 2 && "Unexpected number of operands!");
494 OutMI.setOpcode(convertTailJumpOpcode(Opcode: OutMI.getOpcode()));
495 break;
496 case X86::TAILJMPm:
497 case X86::TAILJMPm64:
498 case X86::TAILJMPm64_REX:
499 assert(OutMI.getNumOperands() == X86::AddrNumOperands &&
500 "Unexpected number of operands!");
501 OutMI.setOpcode(convertTailJumpOpcode(Opcode: OutMI.getOpcode()));
502 break;
503 case X86::MASKMOVDQU:
504 case X86::VMASKMOVDQU:
505 if (In64BitMode)
506 OutMI.setFlags(X86::IP_HAS_AD_SIZE);
507 break;
508 case X86::BSF16rm:
509 case X86::BSF16rr:
510 case X86::BSF32rm:
511 case X86::BSF32rr:
512 case X86::BSF64rm:
513 case X86::BSF64rr: {
514 // Add an REP prefix to BSF instructions so that new processors can
515 // recognize as TZCNT, which has better performance than BSF.
516 // BSF and TZCNT have different interpretations on ZF bit. So make sure
517 // it won't be used later.
518 const MachineOperand *FlagDef =
519 MI->findRegisterDefOperand(Reg: X86::EFLAGS, /*TRI=*/nullptr);
520 if (!MF.getFunction().hasOptSize() && FlagDef && FlagDef->isDead())
521 OutMI.setFlags(X86::IP_HAS_REPEAT);
522 break;
523 }
524 default:
525 break;
526 }
527}
528
529void X86AsmPrinter::LowerTlsAddr(X86MCInstLower &MCInstLowering,
530 const MachineInstr &MI) {
531 NoAutoPaddingScope NoPadScope(*OutStreamer);
532 bool Is64Bits = getSubtarget().is64Bit();
533 bool Is64BitsLP64 = getSubtarget().isTarget64BitLP64();
534 MCContext &Ctx = OutStreamer->getContext();
535
536 MCSymbolRefExpr::VariantKind SRVK;
537 switch (MI.getOpcode()) {
538 case X86::TLS_addr32:
539 case X86::TLS_addr64:
540 case X86::TLS_addrX32:
541 SRVK = MCSymbolRefExpr::VK_TLSGD;
542 break;
543 case X86::TLS_base_addr32:
544 SRVK = MCSymbolRefExpr::VK_TLSLDM;
545 break;
546 case X86::TLS_base_addr64:
547 case X86::TLS_base_addrX32:
548 SRVK = MCSymbolRefExpr::VK_TLSLD;
549 break;
550 case X86::TLS_desc32:
551 case X86::TLS_desc64:
552 SRVK = MCSymbolRefExpr::VK_TLSDESC;
553 break;
554 default:
555 llvm_unreachable("unexpected opcode");
556 }
557
558 const MCSymbolRefExpr *Sym = MCSymbolRefExpr::create(
559 Symbol: MCInstLowering.GetSymbolFromOperand(MO: MI.getOperand(i: 3)), Kind: SRVK, Ctx);
560
561 // Before binutils 2.41, ld has a bogus TLS relaxation error when the GD/LD
562 // code sequence using R_X86_64_GOTPCREL (instead of R_X86_64_GOTPCRELX) is
563 // attempted to be relaxed to IE/LE (binutils PR24784). Work around the bug by
564 // only using GOT when GOTPCRELX is enabled.
565 // TODO Delete the workaround when rustc no longer relies on the hack
566 bool UseGot = MMI->getModule()->getRtLibUseGOT() &&
567 Ctx.getTargetOptions()->X86RelaxRelocations;
568
569 if (SRVK == MCSymbolRefExpr::VK_TLSDESC) {
570 const MCSymbolRefExpr *Expr = MCSymbolRefExpr::create(
571 Symbol: MCInstLowering.GetSymbolFromOperand(MO: MI.getOperand(i: 3)),
572 Kind: MCSymbolRefExpr::VK_TLSCALL, Ctx);
573 EmitAndCountInstruction(
574 Inst&: MCInstBuilder(Is64BitsLP64 ? X86::LEA64r : X86::LEA32r)
575 .addReg(Reg: Is64BitsLP64 ? X86::RAX : X86::EAX)
576 .addReg(Reg: Is64Bits ? X86::RIP : X86::EBX)
577 .addImm(Val: 1)
578 .addReg(Reg: 0)
579 .addExpr(Val: Sym)
580 .addReg(Reg: 0));
581 EmitAndCountInstruction(
582 Inst&: MCInstBuilder(Is64Bits ? X86::CALL64m : X86::CALL32m)
583 .addReg(Reg: Is64BitsLP64 ? X86::RAX : X86::EAX)
584 .addImm(Val: 1)
585 .addReg(Reg: 0)
586 .addExpr(Val: Expr)
587 .addReg(Reg: 0));
588 } else if (Is64Bits) {
589 bool NeedsPadding = SRVK == MCSymbolRefExpr::VK_TLSGD;
590 if (NeedsPadding && Is64BitsLP64)
591 EmitAndCountInstruction(Inst&: MCInstBuilder(X86::DATA16_PREFIX));
592 EmitAndCountInstruction(Inst&: MCInstBuilder(X86::LEA64r)
593 .addReg(Reg: X86::RDI)
594 .addReg(Reg: X86::RIP)
595 .addImm(Val: 1)
596 .addReg(Reg: 0)
597 .addExpr(Val: Sym)
598 .addReg(Reg: 0));
599 const MCSymbol *TlsGetAddr = Ctx.getOrCreateSymbol(Name: "__tls_get_addr");
600 if (NeedsPadding) {
601 if (!UseGot)
602 EmitAndCountInstruction(Inst&: MCInstBuilder(X86::DATA16_PREFIX));
603 EmitAndCountInstruction(Inst&: MCInstBuilder(X86::DATA16_PREFIX));
604 EmitAndCountInstruction(Inst&: MCInstBuilder(X86::REX64_PREFIX));
605 }
606 if (UseGot) {
607 const MCExpr *Expr = MCSymbolRefExpr::create(
608 Symbol: TlsGetAddr, Kind: MCSymbolRefExpr::VK_GOTPCREL, Ctx);
609 EmitAndCountInstruction(Inst&: MCInstBuilder(X86::CALL64m)
610 .addReg(Reg: X86::RIP)
611 .addImm(Val: 1)
612 .addReg(Reg: 0)
613 .addExpr(Val: Expr)
614 .addReg(Reg: 0));
615 } else {
616 EmitAndCountInstruction(
617 Inst&: MCInstBuilder(X86::CALL64pcrel32)
618 .addExpr(Val: MCSymbolRefExpr::create(Symbol: TlsGetAddr,
619 Kind: MCSymbolRefExpr::VK_PLT, Ctx)));
620 }
621 } else {
622 if (SRVK == MCSymbolRefExpr::VK_TLSGD && !UseGot) {
623 EmitAndCountInstruction(Inst&: MCInstBuilder(X86::LEA32r)
624 .addReg(Reg: X86::EAX)
625 .addReg(Reg: 0)
626 .addImm(Val: 1)
627 .addReg(Reg: X86::EBX)
628 .addExpr(Val: Sym)
629 .addReg(Reg: 0));
630 } else {
631 EmitAndCountInstruction(Inst&: MCInstBuilder(X86::LEA32r)
632 .addReg(Reg: X86::EAX)
633 .addReg(Reg: X86::EBX)
634 .addImm(Val: 1)
635 .addReg(Reg: 0)
636 .addExpr(Val: Sym)
637 .addReg(Reg: 0));
638 }
639
640 const MCSymbol *TlsGetAddr = Ctx.getOrCreateSymbol(Name: "___tls_get_addr");
641 if (UseGot) {
642 const MCExpr *Expr =
643 MCSymbolRefExpr::create(Symbol: TlsGetAddr, Kind: MCSymbolRefExpr::VK_GOT, Ctx);
644 EmitAndCountInstruction(Inst&: MCInstBuilder(X86::CALL32m)
645 .addReg(Reg: X86::EBX)
646 .addImm(Val: 1)
647 .addReg(Reg: 0)
648 .addExpr(Val: Expr)
649 .addReg(Reg: 0));
650 } else {
651 EmitAndCountInstruction(
652 Inst&: MCInstBuilder(X86::CALLpcrel32)
653 .addExpr(Val: MCSymbolRefExpr::create(Symbol: TlsGetAddr,
654 Kind: MCSymbolRefExpr::VK_PLT, Ctx)));
655 }
656 }
657}
658
659/// Emit the largest nop instruction smaller than or equal to \p NumBytes
660/// bytes. Return the size of nop emitted.
661static unsigned emitNop(MCStreamer &OS, unsigned NumBytes,
662 const X86Subtarget *Subtarget) {
663 // Determine the longest nop which can be efficiently decoded for the given
664 // target cpu. 15-bytes is the longest single NOP instruction, but some
665 // platforms can't decode the longest forms efficiently.
666 unsigned MaxNopLength = 1;
667 if (Subtarget->is64Bit()) {
668 // FIXME: We can use NOOPL on 32-bit targets with FeatureNOPL, but the
669 // IndexReg/BaseReg below need to be updated.
670 if (Subtarget->hasFeature(Feature: X86::TuningFast7ByteNOP))
671 MaxNopLength = 7;
672 else if (Subtarget->hasFeature(Feature: X86::TuningFast15ByteNOP))
673 MaxNopLength = 15;
674 else if (Subtarget->hasFeature(Feature: X86::TuningFast11ByteNOP))
675 MaxNopLength = 11;
676 else
677 MaxNopLength = 10;
678 } if (Subtarget->is32Bit())
679 MaxNopLength = 2;
680
681 // Cap a single nop emission at the profitable value for the target
682 NumBytes = std::min(a: NumBytes, b: MaxNopLength);
683
684 unsigned NopSize;
685 unsigned Opc, BaseReg, ScaleVal, IndexReg, Displacement, SegmentReg;
686 IndexReg = Displacement = SegmentReg = 0;
687 BaseReg = X86::RAX;
688 ScaleVal = 1;
689 switch (NumBytes) {
690 case 0:
691 llvm_unreachable("Zero nops?");
692 break;
693 case 1:
694 NopSize = 1;
695 Opc = X86::NOOP;
696 break;
697 case 2:
698 NopSize = 2;
699 Opc = X86::XCHG16ar;
700 break;
701 case 3:
702 NopSize = 3;
703 Opc = X86::NOOPL;
704 break;
705 case 4:
706 NopSize = 4;
707 Opc = X86::NOOPL;
708 Displacement = 8;
709 break;
710 case 5:
711 NopSize = 5;
712 Opc = X86::NOOPL;
713 Displacement = 8;
714 IndexReg = X86::RAX;
715 break;
716 case 6:
717 NopSize = 6;
718 Opc = X86::NOOPW;
719 Displacement = 8;
720 IndexReg = X86::RAX;
721 break;
722 case 7:
723 NopSize = 7;
724 Opc = X86::NOOPL;
725 Displacement = 512;
726 break;
727 case 8:
728 NopSize = 8;
729 Opc = X86::NOOPL;
730 Displacement = 512;
731 IndexReg = X86::RAX;
732 break;
733 case 9:
734 NopSize = 9;
735 Opc = X86::NOOPW;
736 Displacement = 512;
737 IndexReg = X86::RAX;
738 break;
739 default:
740 NopSize = 10;
741 Opc = X86::NOOPW;
742 Displacement = 512;
743 IndexReg = X86::RAX;
744 SegmentReg = X86::CS;
745 break;
746 }
747
748 unsigned NumPrefixes = std::min(a: NumBytes - NopSize, b: 5U);
749 NopSize += NumPrefixes;
750 for (unsigned i = 0; i != NumPrefixes; ++i)
751 OS.emitBytes(Data: "\x66");
752
753 switch (Opc) {
754 default: llvm_unreachable("Unexpected opcode");
755 case X86::NOOP:
756 OS.emitInstruction(Inst: MCInstBuilder(Opc), STI: *Subtarget);
757 break;
758 case X86::XCHG16ar:
759 OS.emitInstruction(Inst: MCInstBuilder(Opc).addReg(Reg: X86::AX).addReg(Reg: X86::AX),
760 STI: *Subtarget);
761 break;
762 case X86::NOOPL:
763 case X86::NOOPW:
764 OS.emitInstruction(Inst: MCInstBuilder(Opc)
765 .addReg(Reg: BaseReg)
766 .addImm(Val: ScaleVal)
767 .addReg(Reg: IndexReg)
768 .addImm(Val: Displacement)
769 .addReg(Reg: SegmentReg),
770 STI: *Subtarget);
771 break;
772 }
773 assert(NopSize <= NumBytes && "We overemitted?");
774 return NopSize;
775}
776
777/// Emit the optimal amount of multi-byte nops on X86.
778static void emitX86Nops(MCStreamer &OS, unsigned NumBytes,
779 const X86Subtarget *Subtarget) {
780 unsigned NopsToEmit = NumBytes;
781 (void)NopsToEmit;
782 while (NumBytes) {
783 NumBytes -= emitNop(OS, NumBytes, Subtarget);
784 assert(NopsToEmit >= NumBytes && "Emitted more than I asked for!");
785 }
786}
787
788void X86AsmPrinter::LowerSTATEPOINT(const MachineInstr &MI,
789 X86MCInstLower &MCIL) {
790 assert(Subtarget->is64Bit() && "Statepoint currently only supports X86-64");
791
792 NoAutoPaddingScope NoPadScope(*OutStreamer);
793
794 StatepointOpers SOpers(&MI);
795 if (unsigned PatchBytes = SOpers.getNumPatchBytes()) {
796 emitX86Nops(OS&: *OutStreamer, NumBytes: PatchBytes, Subtarget);
797 } else {
798 // Lower call target and choose correct opcode
799 const MachineOperand &CallTarget = SOpers.getCallTarget();
800 MCOperand CallTargetMCOp;
801 unsigned CallOpcode;
802 switch (CallTarget.getType()) {
803 case MachineOperand::MO_GlobalAddress:
804 case MachineOperand::MO_ExternalSymbol:
805 CallTargetMCOp = MCIL.LowerSymbolOperand(
806 MO: CallTarget, Sym: MCIL.GetSymbolFromOperand(MO: CallTarget));
807 CallOpcode = X86::CALL64pcrel32;
808 // Currently, we only support relative addressing with statepoints.
809 // Otherwise, we'll need a scratch register to hold the target
810 // address. You'll fail asserts during load & relocation if this
811 // symbol is to far away. (TODO: support non-relative addressing)
812 break;
813 case MachineOperand::MO_Immediate:
814 CallTargetMCOp = MCOperand::createImm(Val: CallTarget.getImm());
815 CallOpcode = X86::CALL64pcrel32;
816 // Currently, we only support relative addressing with statepoints.
817 // Otherwise, we'll need a scratch register to hold the target
818 // immediate. You'll fail asserts during load & relocation if this
819 // address is to far away. (TODO: support non-relative addressing)
820 break;
821 case MachineOperand::MO_Register:
822 // FIXME: Add retpoline support and remove this.
823 if (Subtarget->useIndirectThunkCalls())
824 report_fatal_error(reason: "Lowering register statepoints with thunks not "
825 "yet implemented.");
826 CallTargetMCOp = MCOperand::createReg(Reg: CallTarget.getReg());
827 CallOpcode = X86::CALL64r;
828 break;
829 default:
830 llvm_unreachable("Unsupported operand type in statepoint call target");
831 break;
832 }
833
834 // Emit call
835 MCInst CallInst;
836 CallInst.setOpcode(CallOpcode);
837 CallInst.addOperand(Op: CallTargetMCOp);
838 OutStreamer->emitInstruction(Inst: CallInst, STI: getSubtargetInfo());
839 }
840
841 // Record our statepoint node in the same section used by STACKMAP
842 // and PATCHPOINT
843 auto &Ctx = OutStreamer->getContext();
844 MCSymbol *MILabel = Ctx.createTempSymbol();
845 OutStreamer->emitLabel(Symbol: MILabel);
846 SM.recordStatepoint(L: *MILabel, MI);
847}
848
849void X86AsmPrinter::LowerFAULTING_OP(const MachineInstr &FaultingMI,
850 X86MCInstLower &MCIL) {
851 // FAULTING_LOAD_OP <def>, <faltinf type>, <MBB handler>,
852 // <opcode>, <operands>
853
854 NoAutoPaddingScope NoPadScope(*OutStreamer);
855
856 Register DefRegister = FaultingMI.getOperand(i: 0).getReg();
857 FaultMaps::FaultKind FK =
858 static_cast<FaultMaps::FaultKind>(FaultingMI.getOperand(i: 1).getImm());
859 MCSymbol *HandlerLabel = FaultingMI.getOperand(i: 2).getMBB()->getSymbol();
860 unsigned Opcode = FaultingMI.getOperand(i: 3).getImm();
861 unsigned OperandsBeginIdx = 4;
862
863 auto &Ctx = OutStreamer->getContext();
864 MCSymbol *FaultingLabel = Ctx.createTempSymbol();
865 OutStreamer->emitLabel(Symbol: FaultingLabel);
866
867 assert(FK < FaultMaps::FaultKindMax && "Invalid Faulting Kind!");
868 FM.recordFaultingOp(FaultTy: FK, FaultingLabel, HandlerLabel);
869
870 MCInst MI;
871 MI.setOpcode(Opcode);
872
873 if (DefRegister != X86::NoRegister)
874 MI.addOperand(Op: MCOperand::createReg(Reg: DefRegister));
875
876 for (const MachineOperand &MO :
877 llvm::drop_begin(RangeOrContainer: FaultingMI.operands(), N: OperandsBeginIdx))
878 if (auto Op = MCIL.LowerMachineOperand(MI: &FaultingMI, MO); Op.isValid())
879 MI.addOperand(Op);
880
881 OutStreamer->AddComment(T: "on-fault: " + HandlerLabel->getName());
882 OutStreamer->emitInstruction(Inst: MI, STI: getSubtargetInfo());
883}
884
885void X86AsmPrinter::LowerFENTRY_CALL(const MachineInstr &MI,
886 X86MCInstLower &MCIL) {
887 bool Is64Bits = Subtarget->is64Bit();
888 MCContext &Ctx = OutStreamer->getContext();
889 MCSymbol *fentry = Ctx.getOrCreateSymbol(Name: "__fentry__");
890 const MCSymbolRefExpr *Op =
891 MCSymbolRefExpr::create(Symbol: fentry, Kind: MCSymbolRefExpr::VK_None, Ctx);
892
893 EmitAndCountInstruction(
894 Inst&: MCInstBuilder(Is64Bits ? X86::CALL64pcrel32 : X86::CALLpcrel32)
895 .addExpr(Val: Op));
896}
897
898void X86AsmPrinter::LowerKCFI_CHECK(const MachineInstr &MI) {
899 assert(std::next(MI.getIterator())->isCall() &&
900 "KCFI_CHECK not followed by a call instruction");
901
902 // Adjust the offset for patchable-function-prefix. X86InstrInfo::getNop()
903 // returns a 1-byte X86::NOOP, which means the offset is the same in
904 // bytes. This assumes that patchable-function-prefix is the same for all
905 // functions.
906 const MachineFunction &MF = *MI.getMF();
907 int64_t PrefixNops = 0;
908 (void)MF.getFunction()
909 .getFnAttribute(Kind: "patchable-function-prefix")
910 .getValueAsString()
911 .getAsInteger(Radix: 10, Result&: PrefixNops);
912
913 // KCFI allows indirect calls to any location that's preceded by a valid
914 // type identifier. To avoid encoding the full constant into an instruction,
915 // and thus emitting potential call target gadgets at each indirect call
916 // site, load a negated constant to a register and compare that to the
917 // expected value at the call target.
918 const Register AddrReg = MI.getOperand(i: 0).getReg();
919 const uint32_t Type = MI.getOperand(i: 1).getImm();
920 // The check is immediately before the call. If the call target is in R10,
921 // we can clobber R11 for the check instead.
922 unsigned TempReg = AddrReg == X86::R10 ? X86::R11D : X86::R10D;
923 EmitAndCountInstruction(
924 Inst&: MCInstBuilder(X86::MOV32ri).addReg(Reg: TempReg).addImm(Val: -MaskKCFIType(Value: Type)));
925 EmitAndCountInstruction(Inst&: MCInstBuilder(X86::ADD32rm)
926 .addReg(Reg: X86::NoRegister)
927 .addReg(Reg: TempReg)
928 .addReg(Reg: AddrReg)
929 .addImm(Val: 1)
930 .addReg(Reg: X86::NoRegister)
931 .addImm(Val: -(PrefixNops + 4))
932 .addReg(Reg: X86::NoRegister));
933
934 MCSymbol *Pass = OutContext.createTempSymbol();
935 EmitAndCountInstruction(
936 Inst&: MCInstBuilder(X86::JCC_1)
937 .addExpr(Val: MCSymbolRefExpr::create(Symbol: Pass, Ctx&: OutContext))
938 .addImm(Val: X86::COND_E));
939
940 MCSymbol *Trap = OutContext.createTempSymbol();
941 OutStreamer->emitLabel(Symbol: Trap);
942 EmitAndCountInstruction(Inst&: MCInstBuilder(X86::TRAP));
943 emitKCFITrapEntry(MF, Symbol: Trap);
944 OutStreamer->emitLabel(Symbol: Pass);
945}
946
947void X86AsmPrinter::LowerASAN_CHECK_MEMACCESS(const MachineInstr &MI) {
948 // FIXME: Make this work on non-ELF.
949 if (!TM.getTargetTriple().isOSBinFormatELF()) {
950 report_fatal_error(reason: "llvm.asan.check.memaccess only supported on ELF");
951 return;
952 }
953
954 const auto &Reg = MI.getOperand(i: 0).getReg();
955 ASanAccessInfo AccessInfo(MI.getOperand(i: 1).getImm());
956
957 uint64_t ShadowBase;
958 int MappingScale;
959 bool OrShadowOffset;
960 getAddressSanitizerParams(TargetTriple: Triple(TM.getTargetTriple()), LongSize: 64,
961 IsKasan: AccessInfo.CompileKernel, ShadowBase: &ShadowBase,
962 MappingScale: &MappingScale, OrShadowOffset: &OrShadowOffset);
963
964 StringRef Name = AccessInfo.IsWrite ? "store" : "load";
965 StringRef Op = OrShadowOffset ? "or" : "add";
966 std::string SymName = ("__asan_check_" + Name + "_" + Op + "_" +
967 Twine(1ULL << AccessInfo.AccessSizeIndex) + "_" +
968 TM.getMCRegisterInfo()->getName(RegNo: Reg.asMCReg()))
969 .str();
970 if (OrShadowOffset)
971 report_fatal_error(
972 reason: "OrShadowOffset is not supported with optimized callbacks");
973
974 EmitAndCountInstruction(
975 Inst&: MCInstBuilder(X86::CALL64pcrel32)
976 .addExpr(Val: MCSymbolRefExpr::create(
977 Symbol: OutContext.getOrCreateSymbol(Name: SymName), Ctx&: OutContext)));
978}
979
980void X86AsmPrinter::LowerPATCHABLE_OP(const MachineInstr &MI,
981 X86MCInstLower &MCIL) {
982 // PATCHABLE_OP minsize
983
984 NoAutoPaddingScope NoPadScope(*OutStreamer);
985
986 auto NextMI = std::find_if(first: std::next(x: MI.getIterator()),
987 last: MI.getParent()->end().getInstrIterator(),
988 pred: [](auto &II) { return !II.isMetaInstruction(); });
989
990 SmallString<256> Code;
991 unsigned MinSize = MI.getOperand(i: 0).getImm();
992
993 if (NextMI != MI.getParent()->end() && !NextMI->isInlineAsm()) {
994 // Lower the next MachineInstr to find its byte size.
995 // If the next instruction is inline assembly, we skip lowering it for now,
996 // and assume we should always generate NOPs.
997 MCInst MCI;
998 MCIL.Lower(MI: &*NextMI, OutMI&: MCI);
999
1000 SmallVector<MCFixup, 4> Fixups;
1001 CodeEmitter->encodeInstruction(Inst: MCI, CB&: Code, Fixups, STI: getSubtargetInfo());
1002 }
1003
1004 if (Code.size() < MinSize) {
1005 if (MinSize == 2 && Subtarget->is32Bit() &&
1006 Subtarget->isTargetWindowsMSVC() &&
1007 (Subtarget->getCPU().empty() || Subtarget->getCPU() == "pentium3")) {
1008 // For compatibility reasons, when targetting MSVC, it is important to
1009 // generate a 'legacy' NOP in the form of a 8B FF MOV EDI, EDI. Some tools
1010 // rely specifically on this pattern to be able to patch a function.
1011 // This is only for 32-bit targets, when using /arch:IA32 or /arch:SSE.
1012 OutStreamer->emitInstruction(
1013 Inst: MCInstBuilder(X86::MOV32rr_REV).addReg(Reg: X86::EDI).addReg(Reg: X86::EDI),
1014 STI: *Subtarget);
1015 } else {
1016 unsigned NopSize = emitNop(OS&: *OutStreamer, NumBytes: MinSize, Subtarget);
1017 assert(NopSize == MinSize && "Could not implement MinSize!");
1018 (void)NopSize;
1019 }
1020 }
1021}
1022
1023// Lower a stackmap of the form:
1024// <id>, <shadowBytes>, ...
1025void X86AsmPrinter::LowerSTACKMAP(const MachineInstr &MI) {
1026 SMShadowTracker.emitShadowPadding(OutStreamer&: *OutStreamer, STI: getSubtargetInfo());
1027
1028 auto &Ctx = OutStreamer->getContext();
1029 MCSymbol *MILabel = Ctx.createTempSymbol();
1030 OutStreamer->emitLabel(Symbol: MILabel);
1031
1032 SM.recordStackMap(L: *MILabel, MI);
1033 unsigned NumShadowBytes = MI.getOperand(i: 1).getImm();
1034 SMShadowTracker.reset(RequiredSize: NumShadowBytes);
1035}
1036
1037// Lower a patchpoint of the form:
1038// [<def>], <id>, <numBytes>, <target>, <numArgs>, <cc>, ...
1039void X86AsmPrinter::LowerPATCHPOINT(const MachineInstr &MI,
1040 X86MCInstLower &MCIL) {
1041 assert(Subtarget->is64Bit() && "Patchpoint currently only supports X86-64");
1042
1043 SMShadowTracker.emitShadowPadding(OutStreamer&: *OutStreamer, STI: getSubtargetInfo());
1044
1045 NoAutoPaddingScope NoPadScope(*OutStreamer);
1046
1047 auto &Ctx = OutStreamer->getContext();
1048 MCSymbol *MILabel = Ctx.createTempSymbol();
1049 OutStreamer->emitLabel(Symbol: MILabel);
1050 SM.recordPatchPoint(L: *MILabel, MI);
1051
1052 PatchPointOpers opers(&MI);
1053 unsigned ScratchIdx = opers.getNextScratchIdx();
1054 unsigned EncodedBytes = 0;
1055 const MachineOperand &CalleeMO = opers.getCallTarget();
1056
1057 // Check for null target. If target is non-null (i.e. is non-zero or is
1058 // symbolic) then emit a call.
1059 if (!(CalleeMO.isImm() && !CalleeMO.getImm())) {
1060 MCOperand CalleeMCOp;
1061 switch (CalleeMO.getType()) {
1062 default:
1063 /// FIXME: Add a verifier check for bad callee types.
1064 llvm_unreachable("Unrecognized callee operand type.");
1065 case MachineOperand::MO_Immediate:
1066 if (CalleeMO.getImm())
1067 CalleeMCOp = MCOperand::createImm(Val: CalleeMO.getImm());
1068 break;
1069 case MachineOperand::MO_ExternalSymbol:
1070 case MachineOperand::MO_GlobalAddress:
1071 CalleeMCOp = MCIL.LowerSymbolOperand(MO: CalleeMO,
1072 Sym: MCIL.GetSymbolFromOperand(MO: CalleeMO));
1073 break;
1074 }
1075
1076 // Emit MOV to materialize the target address and the CALL to target.
1077 // This is encoded with 12-13 bytes, depending on which register is used.
1078 Register ScratchReg = MI.getOperand(i: ScratchIdx).getReg();
1079 if (X86II::isX86_64ExtendedReg(RegNo: ScratchReg))
1080 EncodedBytes = 13;
1081 else
1082 EncodedBytes = 12;
1083
1084 EmitAndCountInstruction(
1085 Inst&: MCInstBuilder(X86::MOV64ri).addReg(Reg: ScratchReg).addOperand(Op: CalleeMCOp));
1086 // FIXME: Add retpoline support and remove this.
1087 if (Subtarget->useIndirectThunkCalls())
1088 report_fatal_error(
1089 reason: "Lowering patchpoint with thunks not yet implemented.");
1090 EmitAndCountInstruction(Inst&: MCInstBuilder(X86::CALL64r).addReg(Reg: ScratchReg));
1091 }
1092
1093 // Emit padding.
1094 unsigned NumBytes = opers.getNumPatchBytes();
1095 assert(NumBytes >= EncodedBytes &&
1096 "Patchpoint can't request size less than the length of a call.");
1097
1098 emitX86Nops(OS&: *OutStreamer, NumBytes: NumBytes - EncodedBytes, Subtarget);
1099}
1100
1101void X86AsmPrinter::LowerPATCHABLE_EVENT_CALL(const MachineInstr &MI,
1102 X86MCInstLower &MCIL) {
1103 assert(Subtarget->is64Bit() && "XRay custom events only supports X86-64");
1104
1105 NoAutoPaddingScope NoPadScope(*OutStreamer);
1106
1107 // We want to emit the following pattern, which follows the x86 calling
1108 // convention to prepare for the trampoline call to be patched in.
1109 //
1110 // .p2align 1, ...
1111 // .Lxray_event_sled_N:
1112 // jmp +N // jump across the instrumentation sled
1113 // ... // set up arguments in register
1114 // callq __xray_CustomEvent@plt // force dependency to symbol
1115 // ...
1116 // <jump here>
1117 //
1118 // After patching, it would look something like:
1119 //
1120 // nopw (2-byte nop)
1121 // ...
1122 // callq __xrayCustomEvent // already lowered
1123 // ...
1124 //
1125 // ---
1126 // First we emit the label and the jump.
1127 auto CurSled = OutContext.createTempSymbol(Name: "xray_event_sled_", AlwaysAddSuffix: true);
1128 OutStreamer->AddComment(T: "# XRay Custom Event Log");
1129 OutStreamer->emitCodeAlignment(Alignment: Align(2), STI: &getSubtargetInfo());
1130 OutStreamer->emitLabel(Symbol: CurSled);
1131
1132 // Use a two-byte `jmp`. This version of JMP takes an 8-bit relative offset as
1133 // an operand (computed as an offset from the jmp instruction).
1134 // FIXME: Find another less hacky way do force the relative jump.
1135 OutStreamer->emitBinaryData(Data: "\xeb\x0f");
1136
1137 // The default C calling convention will place two arguments into %rcx and
1138 // %rdx -- so we only work with those.
1139 const Register DestRegs[] = {X86::RDI, X86::RSI};
1140 bool UsedMask[] = {false, false};
1141 // Filled out in loop.
1142 Register SrcRegs[] = {0, 0};
1143
1144 // Then we put the operands in the %rdi and %rsi registers. We spill the
1145 // values in the register before we clobber them, and mark them as used in
1146 // UsedMask. In case the arguments are already in the correct register, we use
1147 // emit nops appropriately sized to keep the sled the same size in every
1148 // situation.
1149 for (unsigned I = 0; I < MI.getNumOperands(); ++I)
1150 if (auto Op = MCIL.LowerMachineOperand(MI: &MI, MO: MI.getOperand(i: I));
1151 Op.isValid()) {
1152 assert(Op.isReg() && "Only support arguments in registers");
1153 SrcRegs[I] = getX86SubSuperRegister(Reg: Op.getReg(), Size: 64);
1154 assert(SrcRegs[I].isValid() && "Invalid operand");
1155 if (SrcRegs[I] != DestRegs[I]) {
1156 UsedMask[I] = true;
1157 EmitAndCountInstruction(
1158 Inst&: MCInstBuilder(X86::PUSH64r).addReg(Reg: DestRegs[I]));
1159 } else {
1160 emitX86Nops(OS&: *OutStreamer, NumBytes: 4, Subtarget);
1161 }
1162 }
1163
1164 // Now that the register values are stashed, mov arguments into place.
1165 // FIXME: This doesn't work if one of the later SrcRegs is equal to an
1166 // earlier DestReg. We will have already overwritten over the register before
1167 // we can copy from it.
1168 for (unsigned I = 0; I < MI.getNumOperands(); ++I)
1169 if (SrcRegs[I] != DestRegs[I])
1170 EmitAndCountInstruction(
1171 Inst&: MCInstBuilder(X86::MOV64rr).addReg(Reg: DestRegs[I]).addReg(Reg: SrcRegs[I]));
1172
1173 // We emit a hard dependency on the __xray_CustomEvent symbol, which is the
1174 // name of the trampoline to be implemented by the XRay runtime.
1175 auto TSym = OutContext.getOrCreateSymbol(Name: "__xray_CustomEvent");
1176 MachineOperand TOp = MachineOperand::CreateMCSymbol(Sym: TSym);
1177 if (isPositionIndependent())
1178 TOp.setTargetFlags(X86II::MO_PLT);
1179
1180 // Emit the call instruction.
1181 EmitAndCountInstruction(Inst&: MCInstBuilder(X86::CALL64pcrel32)
1182 .addOperand(Op: MCIL.LowerSymbolOperand(MO: TOp, Sym: TSym)));
1183
1184 // Restore caller-saved and used registers.
1185 for (unsigned I = sizeof UsedMask; I-- > 0;)
1186 if (UsedMask[I])
1187 EmitAndCountInstruction(Inst&: MCInstBuilder(X86::POP64r).addReg(Reg: DestRegs[I]));
1188 else
1189 emitX86Nops(OS&: *OutStreamer, NumBytes: 1, Subtarget);
1190
1191 OutStreamer->AddComment(T: "xray custom event end.");
1192
1193 // Record the sled version. Version 0 of this sled was spelled differently, so
1194 // we let the runtime handle the different offsets we're using. Version 2
1195 // changed the absolute address to a PC-relative address.
1196 recordSled(Sled: CurSled, MI, Kind: SledKind::CUSTOM_EVENT, Version: 2);
1197}
1198
1199void X86AsmPrinter::LowerPATCHABLE_TYPED_EVENT_CALL(const MachineInstr &MI,
1200 X86MCInstLower &MCIL) {
1201 assert(Subtarget->is64Bit() && "XRay typed events only supports X86-64");
1202
1203 NoAutoPaddingScope NoPadScope(*OutStreamer);
1204
1205 // We want to emit the following pattern, which follows the x86 calling
1206 // convention to prepare for the trampoline call to be patched in.
1207 //
1208 // .p2align 1, ...
1209 // .Lxray_event_sled_N:
1210 // jmp +N // jump across the instrumentation sled
1211 // ... // set up arguments in register
1212 // callq __xray_TypedEvent@plt // force dependency to symbol
1213 // ...
1214 // <jump here>
1215 //
1216 // After patching, it would look something like:
1217 //
1218 // nopw (2-byte nop)
1219 // ...
1220 // callq __xrayTypedEvent // already lowered
1221 // ...
1222 //
1223 // ---
1224 // First we emit the label and the jump.
1225 auto CurSled = OutContext.createTempSymbol(Name: "xray_typed_event_sled_", AlwaysAddSuffix: true);
1226 OutStreamer->AddComment(T: "# XRay Typed Event Log");
1227 OutStreamer->emitCodeAlignment(Alignment: Align(2), STI: &getSubtargetInfo());
1228 OutStreamer->emitLabel(Symbol: CurSled);
1229
1230 // Use a two-byte `jmp`. This version of JMP takes an 8-bit relative offset as
1231 // an operand (computed as an offset from the jmp instruction).
1232 // FIXME: Find another less hacky way do force the relative jump.
1233 OutStreamer->emitBinaryData(Data: "\xeb\x14");
1234
1235 // An x86-64 convention may place three arguments into %rcx, %rdx, and R8,
1236 // so we'll work with those. Or we may be called via SystemV, in which case
1237 // we don't have to do any translation.
1238 const Register DestRegs[] = {X86::RDI, X86::RSI, X86::RDX};
1239 bool UsedMask[] = {false, false, false};
1240
1241 // Will fill out src regs in the loop.
1242 Register SrcRegs[] = {0, 0, 0};
1243
1244 // Then we put the operands in the SystemV registers. We spill the values in
1245 // the registers before we clobber them, and mark them as used in UsedMask.
1246 // In case the arguments are already in the correct register, we emit nops
1247 // appropriately sized to keep the sled the same size in every situation.
1248 for (unsigned I = 0; I < MI.getNumOperands(); ++I)
1249 if (auto Op = MCIL.LowerMachineOperand(MI: &MI, MO: MI.getOperand(i: I));
1250 Op.isValid()) {
1251 // TODO: Is register only support adequate?
1252 assert(Op.isReg() && "Only supports arguments in registers");
1253 SrcRegs[I] = getX86SubSuperRegister(Reg: Op.getReg(), Size: 64);
1254 assert(SrcRegs[I].isValid() && "Invalid operand");
1255 if (SrcRegs[I] != DestRegs[I]) {
1256 UsedMask[I] = true;
1257 EmitAndCountInstruction(
1258 Inst&: MCInstBuilder(X86::PUSH64r).addReg(Reg: DestRegs[I]));
1259 } else {
1260 emitX86Nops(OS&: *OutStreamer, NumBytes: 4, Subtarget);
1261 }
1262 }
1263
1264 // In the above loop we only stash all of the destination registers or emit
1265 // nops if the arguments are already in the right place. Doing the actually
1266 // moving is postponed until after all the registers are stashed so nothing
1267 // is clobbers. We've already added nops to account for the size of mov and
1268 // push if the register is in the right place, so we only have to worry about
1269 // emitting movs.
1270 // FIXME: This doesn't work if one of the later SrcRegs is equal to an
1271 // earlier DestReg. We will have already overwritten over the register before
1272 // we can copy from it.
1273 for (unsigned I = 0; I < MI.getNumOperands(); ++I)
1274 if (UsedMask[I])
1275 EmitAndCountInstruction(
1276 Inst&: MCInstBuilder(X86::MOV64rr).addReg(Reg: DestRegs[I]).addReg(Reg: SrcRegs[I]));
1277
1278 // We emit a hard dependency on the __xray_TypedEvent symbol, which is the
1279 // name of the trampoline to be implemented by the XRay runtime.
1280 auto TSym = OutContext.getOrCreateSymbol(Name: "__xray_TypedEvent");
1281 MachineOperand TOp = MachineOperand::CreateMCSymbol(Sym: TSym);
1282 if (isPositionIndependent())
1283 TOp.setTargetFlags(X86II::MO_PLT);
1284
1285 // Emit the call instruction.
1286 EmitAndCountInstruction(Inst&: MCInstBuilder(X86::CALL64pcrel32)
1287 .addOperand(Op: MCIL.LowerSymbolOperand(MO: TOp, Sym: TSym)));
1288
1289 // Restore caller-saved and used registers.
1290 for (unsigned I = sizeof UsedMask; I-- > 0;)
1291 if (UsedMask[I])
1292 EmitAndCountInstruction(Inst&: MCInstBuilder(X86::POP64r).addReg(Reg: DestRegs[I]));
1293 else
1294 emitX86Nops(OS&: *OutStreamer, NumBytes: 1, Subtarget);
1295
1296 OutStreamer->AddComment(T: "xray typed event end.");
1297
1298 // Record the sled version.
1299 recordSled(Sled: CurSled, MI, Kind: SledKind::TYPED_EVENT, Version: 2);
1300}
1301
1302void X86AsmPrinter::LowerPATCHABLE_FUNCTION_ENTER(const MachineInstr &MI,
1303 X86MCInstLower &MCIL) {
1304
1305 NoAutoPaddingScope NoPadScope(*OutStreamer);
1306
1307 const Function &F = MF->getFunction();
1308 if (F.hasFnAttribute(Kind: "patchable-function-entry")) {
1309 unsigned Num;
1310 if (F.getFnAttribute(Kind: "patchable-function-entry")
1311 .getValueAsString()
1312 .getAsInteger(Radix: 10, Result&: Num))
1313 return;
1314 emitX86Nops(OS&: *OutStreamer, NumBytes: Num, Subtarget);
1315 return;
1316 }
1317 // We want to emit the following pattern:
1318 //
1319 // .p2align 1, ...
1320 // .Lxray_sled_N:
1321 // jmp .tmpN
1322 // # 9 bytes worth of noops
1323 //
1324 // We need the 9 bytes because at runtime, we'd be patching over the full 11
1325 // bytes with the following pattern:
1326 //
1327 // mov %r10, <function id, 32-bit> // 6 bytes
1328 // call <relative offset, 32-bits> // 5 bytes
1329 //
1330 auto CurSled = OutContext.createTempSymbol(Name: "xray_sled_", AlwaysAddSuffix: true);
1331 OutStreamer->emitCodeAlignment(Alignment: Align(2), STI: &getSubtargetInfo());
1332 OutStreamer->emitLabel(Symbol: CurSled);
1333
1334 // Use a two-byte `jmp`. This version of JMP takes an 8-bit relative offset as
1335 // an operand (computed as an offset from the jmp instruction).
1336 // FIXME: Find another less hacky way do force the relative jump.
1337 OutStreamer->emitBytes(Data: "\xeb\x09");
1338 emitX86Nops(OS&: *OutStreamer, NumBytes: 9, Subtarget);
1339 recordSled(Sled: CurSled, MI, Kind: SledKind::FUNCTION_ENTER, Version: 2);
1340}
1341
1342void X86AsmPrinter::LowerPATCHABLE_RET(const MachineInstr &MI,
1343 X86MCInstLower &MCIL) {
1344 NoAutoPaddingScope NoPadScope(*OutStreamer);
1345
1346 // Since PATCHABLE_RET takes the opcode of the return statement as an
1347 // argument, we use that to emit the correct form of the RET that we want.
1348 // i.e. when we see this:
1349 //
1350 // PATCHABLE_RET X86::RET ...
1351 //
1352 // We should emit the RET followed by sleds.
1353 //
1354 // .p2align 1, ...
1355 // .Lxray_sled_N:
1356 // ret # or equivalent instruction
1357 // # 10 bytes worth of noops
1358 //
1359 // This just makes sure that the alignment for the next instruction is 2.
1360 auto CurSled = OutContext.createTempSymbol(Name: "xray_sled_", AlwaysAddSuffix: true);
1361 OutStreamer->emitCodeAlignment(Alignment: Align(2), STI: &getSubtargetInfo());
1362 OutStreamer->emitLabel(Symbol: CurSled);
1363 unsigned OpCode = MI.getOperand(i: 0).getImm();
1364 MCInst Ret;
1365 Ret.setOpcode(OpCode);
1366 for (auto &MO : drop_begin(RangeOrContainer: MI.operands()))
1367 if (auto Op = MCIL.LowerMachineOperand(MI: &MI, MO); Op.isValid())
1368 Ret.addOperand(Op);
1369 OutStreamer->emitInstruction(Inst: Ret, STI: getSubtargetInfo());
1370 emitX86Nops(OS&: *OutStreamer, NumBytes: 10, Subtarget);
1371 recordSled(Sled: CurSled, MI, Kind: SledKind::FUNCTION_EXIT, Version: 2);
1372}
1373
1374void X86AsmPrinter::LowerPATCHABLE_TAIL_CALL(const MachineInstr &MI,
1375 X86MCInstLower &MCIL) {
1376 MCInst TC;
1377 TC.setOpcode(convertTailJumpOpcode(Opcode: MI.getOperand(i: 0).getImm()));
1378 // Drop the tail jump opcode.
1379 auto TCOperands = drop_begin(RangeOrContainer: MI.operands());
1380 bool IsConditional = TC.getOpcode() == X86::JCC_1;
1381 MCSymbol *FallthroughLabel;
1382 if (IsConditional) {
1383 // Rewrite:
1384 // je target
1385 //
1386 // To:
1387 // jne .fallthrough
1388 // .p2align 1, ...
1389 // .Lxray_sled_N:
1390 // SLED_CODE
1391 // jmp target
1392 // .fallthrough:
1393 FallthroughLabel = OutContext.createTempSymbol();
1394 EmitToStreamer(
1395 S&: *OutStreamer,
1396 Inst: MCInstBuilder(X86::JCC_1)
1397 .addExpr(Val: MCSymbolRefExpr::create(Symbol: FallthroughLabel, Ctx&: OutContext))
1398 .addImm(Val: X86::GetOppositeBranchCondition(
1399 CC: static_cast<X86::CondCode>(MI.getOperand(i: 2).getImm()))));
1400 TC.setOpcode(X86::JMP_1);
1401 // Drop the condition code.
1402 TCOperands = drop_end(RangeOrContainer&: TCOperands);
1403 }
1404
1405 NoAutoPaddingScope NoPadScope(*OutStreamer);
1406
1407 // Like PATCHABLE_RET, we have the actual instruction in the operands to this
1408 // instruction so we lower that particular instruction and its operands.
1409 // Unlike PATCHABLE_RET though, we put the sled before the JMP, much like how
1410 // we do it for PATCHABLE_FUNCTION_ENTER. The sled should be very similar to
1411 // the PATCHABLE_FUNCTION_ENTER case, followed by the lowering of the actual
1412 // tail call much like how we have it in PATCHABLE_RET.
1413 auto CurSled = OutContext.createTempSymbol(Name: "xray_sled_", AlwaysAddSuffix: true);
1414 OutStreamer->emitCodeAlignment(Alignment: Align(2), STI: &getSubtargetInfo());
1415 OutStreamer->emitLabel(Symbol: CurSled);
1416 auto Target = OutContext.createTempSymbol();
1417
1418 // Use a two-byte `jmp`. This version of JMP takes an 8-bit relative offset as
1419 // an operand (computed as an offset from the jmp instruction).
1420 // FIXME: Find another less hacky way do force the relative jump.
1421 OutStreamer->emitBytes(Data: "\xeb\x09");
1422 emitX86Nops(OS&: *OutStreamer, NumBytes: 9, Subtarget);
1423 OutStreamer->emitLabel(Symbol: Target);
1424 recordSled(Sled: CurSled, MI, Kind: SledKind::TAIL_CALL, Version: 2);
1425
1426 // Before emitting the instruction, add a comment to indicate that this is
1427 // indeed a tail call.
1428 OutStreamer->AddComment(T: "TAILCALL");
1429 for (auto &MO : TCOperands)
1430 if (auto Op = MCIL.LowerMachineOperand(MI: &MI, MO); Op.isValid())
1431 TC.addOperand(Op);
1432 OutStreamer->emitInstruction(Inst: TC, STI: getSubtargetInfo());
1433
1434 if (IsConditional)
1435 OutStreamer->emitLabel(Symbol: FallthroughLabel);
1436}
1437
1438// Returns instruction preceding MBBI in MachineFunction.
1439// If MBBI is the first instruction of the first basic block, returns null.
1440static MachineBasicBlock::const_iterator
1441PrevCrossBBInst(MachineBasicBlock::const_iterator MBBI) {
1442 const MachineBasicBlock *MBB = MBBI->getParent();
1443 while (MBBI == MBB->begin()) {
1444 if (MBB == &MBB->getParent()->front())
1445 return MachineBasicBlock::const_iterator();
1446 MBB = MBB->getPrevNode();
1447 MBBI = MBB->end();
1448 }
1449 --MBBI;
1450 return MBBI;
1451}
1452
1453static unsigned getSrcIdx(const MachineInstr* MI, unsigned SrcIdx) {
1454 if (X86II::isKMasked(TSFlags: MI->getDesc().TSFlags)) {
1455 // Skip mask operand.
1456 ++SrcIdx;
1457 if (X86II::isKMergeMasked(TSFlags: MI->getDesc().TSFlags)) {
1458 // Skip passthru operand.
1459 ++SrcIdx;
1460 }
1461 }
1462 return SrcIdx;
1463}
1464
1465static void printDstRegisterName(raw_ostream &CS, const MachineInstr *MI,
1466 unsigned SrcOpIdx) {
1467 const MachineOperand &DstOp = MI->getOperand(i: 0);
1468 CS << X86ATTInstPrinter::getRegisterName(Reg: DstOp.getReg());
1469
1470 // Handle AVX512 MASK/MASXZ write mask comments.
1471 // MASK: zmmX {%kY}
1472 // MASKZ: zmmX {%kY} {z}
1473 if (X86II::isKMasked(TSFlags: MI->getDesc().TSFlags)) {
1474 const MachineOperand &WriteMaskOp = MI->getOperand(i: SrcOpIdx - 1);
1475 StringRef Mask = X86ATTInstPrinter::getRegisterName(Reg: WriteMaskOp.getReg());
1476 CS << " {%" << Mask << "}";
1477 if (!X86II::isKMergeMasked(TSFlags: MI->getDesc().TSFlags)) {
1478 CS << " {z}";
1479 }
1480 }
1481}
1482
1483static void printShuffleMask(raw_ostream &CS, StringRef Src1Name,
1484 StringRef Src2Name, ArrayRef<int> Mask) {
1485 // One source operand, fix the mask to print all elements in one span.
1486 SmallVector<int, 8> ShuffleMask(Mask);
1487 if (Src1Name == Src2Name)
1488 for (int i = 0, e = ShuffleMask.size(); i != e; ++i)
1489 if (ShuffleMask[i] >= e)
1490 ShuffleMask[i] -= e;
1491
1492 for (int i = 0, e = ShuffleMask.size(); i != e; ++i) {
1493 if (i != 0)
1494 CS << ",";
1495 if (ShuffleMask[i] == SM_SentinelZero) {
1496 CS << "zero";
1497 continue;
1498 }
1499
1500 // Otherwise, it must come from src1 or src2. Print the span of elements
1501 // that comes from this src.
1502 bool isSrc1 = ShuffleMask[i] < (int)e;
1503 CS << (isSrc1 ? Src1Name : Src2Name) << '[';
1504
1505 bool IsFirst = true;
1506 while (i != e && ShuffleMask[i] != SM_SentinelZero &&
1507 (ShuffleMask[i] < (int)e) == isSrc1) {
1508 if (!IsFirst)
1509 CS << ',';
1510 else
1511 IsFirst = false;
1512 if (ShuffleMask[i] == SM_SentinelUndef)
1513 CS << "u";
1514 else
1515 CS << ShuffleMask[i] % (int)e;
1516 ++i;
1517 }
1518 CS << ']';
1519 --i; // For loop increments element #.
1520 }
1521}
1522
1523static std::string getShuffleComment(const MachineInstr *MI, unsigned SrcOp1Idx,
1524 unsigned SrcOp2Idx, ArrayRef<int> Mask) {
1525 std::string Comment;
1526
1527 const MachineOperand &SrcOp1 = MI->getOperand(i: SrcOp1Idx);
1528 const MachineOperand &SrcOp2 = MI->getOperand(i: SrcOp2Idx);
1529 StringRef Src1Name = SrcOp1.isReg()
1530 ? X86ATTInstPrinter::getRegisterName(Reg: SrcOp1.getReg())
1531 : "mem";
1532 StringRef Src2Name = SrcOp2.isReg()
1533 ? X86ATTInstPrinter::getRegisterName(Reg: SrcOp2.getReg())
1534 : "mem";
1535
1536 raw_string_ostream CS(Comment);
1537 printDstRegisterName(CS, MI, SrcOpIdx: SrcOp1Idx);
1538 CS << " = ";
1539 printShuffleMask(CS, Src1Name, Src2Name, Mask);
1540 CS.flush();
1541
1542 return Comment;
1543}
1544
1545static void printConstant(const APInt &Val, raw_ostream &CS,
1546 bool PrintZero = false) {
1547 if (Val.getBitWidth() <= 64) {
1548 CS << (PrintZero ? 0ULL : Val.getZExtValue());
1549 } else {
1550 // print multi-word constant as (w0,w1)
1551 CS << "(";
1552 for (int i = 0, N = Val.getNumWords(); i < N; ++i) {
1553 if (i > 0)
1554 CS << ",";
1555 CS << (PrintZero ? 0ULL : Val.getRawData()[i]);
1556 }
1557 CS << ")";
1558 }
1559}
1560
1561static void printConstant(const APFloat &Flt, raw_ostream &CS,
1562 bool PrintZero = false) {
1563 SmallString<32> Str;
1564 // Force scientific notation to distinguish from integers.
1565 if (PrintZero)
1566 APFloat::getZero(Sem: Flt.getSemantics()).toString(Str, FormatPrecision: 0, FormatMaxPadding: 0);
1567 else
1568 Flt.toString(Str, FormatPrecision: 0, FormatMaxPadding: 0);
1569 CS << Str;
1570}
1571
1572static void printConstant(const Constant *COp, unsigned BitWidth,
1573 raw_ostream &CS, bool PrintZero = false) {
1574 if (isa<UndefValue>(Val: COp)) {
1575 CS << "u";
1576 } else if (auto *CI = dyn_cast<ConstantInt>(Val: COp)) {
1577 printConstant(Val: CI->getValue(), CS, PrintZero);
1578 } else if (auto *CF = dyn_cast<ConstantFP>(Val: COp)) {
1579 printConstant(Flt: CF->getValueAPF(), CS, PrintZero);
1580 } else if (auto *CDS = dyn_cast<ConstantDataSequential>(Val: COp)) {
1581 Type *EltTy = CDS->getElementType();
1582 bool IsInteger = EltTy->isIntegerTy();
1583 bool IsFP = EltTy->isHalfTy() || EltTy->isFloatTy() || EltTy->isDoubleTy();
1584 unsigned EltBits = EltTy->getPrimitiveSizeInBits();
1585 unsigned E = std::min(a: BitWidth / EltBits, b: CDS->getNumElements());
1586 assert((BitWidth % EltBits) == 0 && "Element size mismatch");
1587 for (unsigned I = 0; I != E; ++I) {
1588 if (I != 0)
1589 CS << ",";
1590 if (IsInteger)
1591 printConstant(Val: CDS->getElementAsAPInt(i: I), CS, PrintZero);
1592 else if (IsFP)
1593 printConstant(Flt: CDS->getElementAsAPFloat(i: I), CS, PrintZero);
1594 else
1595 CS << "?";
1596 }
1597 } else if (auto *CV = dyn_cast<ConstantVector>(Val: COp)) {
1598 unsigned EltBits = CV->getType()->getScalarSizeInBits();
1599 unsigned E = std::min(a: BitWidth / EltBits, b: CV->getNumOperands());
1600 assert((BitWidth % EltBits) == 0 && "Element size mismatch");
1601 for (unsigned I = 0; I != E; ++I) {
1602 if (I != 0)
1603 CS << ",";
1604 printConstant(COp: CV->getOperand(i_nocapture: I), BitWidth: EltBits, CS, PrintZero);
1605 }
1606 } else {
1607 CS << "?";
1608 }
1609}
1610
1611static void printZeroUpperMove(const MachineInstr *MI, MCStreamer &OutStreamer,
1612 int SclWidth, int VecWidth,
1613 const char *ShuffleComment) {
1614 unsigned SrcIdx = getSrcIdx(MI, SrcIdx: 1);
1615
1616 std::string Comment;
1617 raw_string_ostream CS(Comment);
1618 printDstRegisterName(CS, MI, SrcOpIdx: SrcIdx);
1619 CS << " = ";
1620
1621 if (auto *C = X86::getConstantFromPool(MI: *MI, OpNo: SrcIdx)) {
1622 CS << "[";
1623 printConstant(COp: C, BitWidth: SclWidth, CS);
1624 for (int I = 1, E = VecWidth / SclWidth; I < E; ++I) {
1625 CS << ",";
1626 printConstant(COp: C, BitWidth: SclWidth, CS, PrintZero: true);
1627 }
1628 CS << "]";
1629 OutStreamer.AddComment(T: CS.str());
1630 return; // early-out
1631 }
1632
1633 // We didn't find a constant load, fallback to a shuffle mask decode.
1634 CS << ShuffleComment;
1635 OutStreamer.AddComment(T: CS.str());
1636}
1637
1638static void printBroadcast(const MachineInstr *MI, MCStreamer &OutStreamer,
1639 int Repeats, int BitWidth) {
1640 unsigned SrcIdx = getSrcIdx(MI, SrcIdx: 1);
1641 if (auto *C = X86::getConstantFromPool(MI: *MI, OpNo: SrcIdx)) {
1642 std::string Comment;
1643 raw_string_ostream CS(Comment);
1644 printDstRegisterName(CS, MI, SrcOpIdx: SrcIdx);
1645 CS << " = [";
1646 for (int l = 0; l != Repeats; ++l) {
1647 if (l != 0)
1648 CS << ",";
1649 printConstant(COp: C, BitWidth, CS);
1650 }
1651 CS << "]";
1652 OutStreamer.AddComment(T: CS.str());
1653 }
1654}
1655
1656static bool printExtend(const MachineInstr *MI, MCStreamer &OutStreamer,
1657 int SrcEltBits, int DstEltBits, bool IsSext) {
1658 unsigned SrcIdx = getSrcIdx(MI, SrcIdx: 1);
1659 auto *C = X86::getConstantFromPool(MI: *MI, OpNo: SrcIdx);
1660 if (C && C->getType()->getScalarSizeInBits() == unsigned(SrcEltBits)) {
1661 if (auto *CDS = dyn_cast<ConstantDataSequential>(Val: C)) {
1662 int NumElts = CDS->getNumElements();
1663 std::string Comment;
1664 raw_string_ostream CS(Comment);
1665 printDstRegisterName(CS, MI, SrcOpIdx: SrcIdx);
1666 CS << " = [";
1667 for (int i = 0; i != NumElts; ++i) {
1668 if (i != 0)
1669 CS << ",";
1670 if (CDS->getElementType()->isIntegerTy()) {
1671 APInt Elt = CDS->getElementAsAPInt(i);
1672 Elt = IsSext ? Elt.sext(width: DstEltBits) : Elt.zext(width: DstEltBits);
1673 printConstant(Val: Elt, CS);
1674 } else
1675 CS << "?";
1676 }
1677 CS << "]";
1678 OutStreamer.AddComment(T: CS.str());
1679 return true;
1680 }
1681 }
1682
1683 return false;
1684}
1685static void printSignExtend(const MachineInstr *MI, MCStreamer &OutStreamer,
1686 int SrcEltBits, int DstEltBits) {
1687 printExtend(MI, OutStreamer, SrcEltBits, DstEltBits, IsSext: true);
1688}
1689static void printZeroExtend(const MachineInstr *MI, MCStreamer &OutStreamer,
1690 int SrcEltBits, int DstEltBits) {
1691 if (printExtend(MI, OutStreamer, SrcEltBits, DstEltBits, IsSext: false))
1692 return;
1693
1694 // We didn't find a constant load, fallback to a shuffle mask decode.
1695 std::string Comment;
1696 raw_string_ostream CS(Comment);
1697 printDstRegisterName(CS, MI, SrcOpIdx: getSrcIdx(MI, SrcIdx: 1));
1698 CS << " = ";
1699
1700 SmallVector<int> Mask;
1701 unsigned Width = X86::getVectorRegisterWidth(Info: MI->getDesc().operands()[0]);
1702 assert((Width % DstEltBits) == 0 && (DstEltBits % SrcEltBits) == 0 &&
1703 "Illegal extension ratio");
1704 DecodeZeroExtendMask(SrcScalarBits: SrcEltBits, DstScalarBits: DstEltBits, NumDstElts: Width / DstEltBits, IsAnyExtend: false, ShuffleMask&: Mask);
1705 printShuffleMask(CS, Src1Name: "mem", Src2Name: "", Mask);
1706
1707 OutStreamer.AddComment(T: CS.str());
1708}
1709
1710void X86AsmPrinter::EmitSEHInstruction(const MachineInstr *MI) {
1711 assert(MF->hasWinCFI() && "SEH_ instruction in function without WinCFI?");
1712 assert((getSubtarget().isOSWindows() || TM.getTargetTriple().isUEFI()) &&
1713 "SEH_ instruction Windows and UEFI only");
1714
1715 // Use the .cv_fpo directives if we're emitting CodeView on 32-bit x86.
1716 if (EmitFPOData) {
1717 X86TargetStreamer *XTS =
1718 static_cast<X86TargetStreamer *>(OutStreamer->getTargetStreamer());
1719 switch (MI->getOpcode()) {
1720 case X86::SEH_PushReg:
1721 XTS->emitFPOPushReg(Reg: MI->getOperand(i: 0).getImm());
1722 break;
1723 case X86::SEH_StackAlloc:
1724 XTS->emitFPOStackAlloc(StackAlloc: MI->getOperand(i: 0).getImm());
1725 break;
1726 case X86::SEH_StackAlign:
1727 XTS->emitFPOStackAlign(Align: MI->getOperand(i: 0).getImm());
1728 break;
1729 case X86::SEH_SetFrame:
1730 assert(MI->getOperand(1).getImm() == 0 &&
1731 ".cv_fpo_setframe takes no offset");
1732 XTS->emitFPOSetFrame(Reg: MI->getOperand(i: 0).getImm());
1733 break;
1734 case X86::SEH_EndPrologue:
1735 XTS->emitFPOEndPrologue();
1736 break;
1737 case X86::SEH_SaveReg:
1738 case X86::SEH_SaveXMM:
1739 case X86::SEH_PushFrame:
1740 llvm_unreachable("SEH_ directive incompatible with FPO");
1741 break;
1742 default:
1743 llvm_unreachable("expected SEH_ instruction");
1744 }
1745 return;
1746 }
1747
1748 // Otherwise, use the .seh_ directives for all other Windows platforms.
1749 switch (MI->getOpcode()) {
1750 case X86::SEH_PushReg:
1751 OutStreamer->emitWinCFIPushReg(Register: MI->getOperand(i: 0).getImm());
1752 break;
1753
1754 case X86::SEH_SaveReg:
1755 OutStreamer->emitWinCFISaveReg(Register: MI->getOperand(i: 0).getImm(),
1756 Offset: MI->getOperand(i: 1).getImm());
1757 break;
1758
1759 case X86::SEH_SaveXMM:
1760 OutStreamer->emitWinCFISaveXMM(Register: MI->getOperand(i: 0).getImm(),
1761 Offset: MI->getOperand(i: 1).getImm());
1762 break;
1763
1764 case X86::SEH_StackAlloc:
1765 OutStreamer->emitWinCFIAllocStack(Size: MI->getOperand(i: 0).getImm());
1766 break;
1767
1768 case X86::SEH_SetFrame:
1769 OutStreamer->emitWinCFISetFrame(Register: MI->getOperand(i: 0).getImm(),
1770 Offset: MI->getOperand(i: 1).getImm());
1771 break;
1772
1773 case X86::SEH_PushFrame:
1774 OutStreamer->emitWinCFIPushFrame(Code: MI->getOperand(i: 0).getImm());
1775 break;
1776
1777 case X86::SEH_EndPrologue:
1778 OutStreamer->emitWinCFIEndProlog();
1779 break;
1780
1781 default:
1782 llvm_unreachable("expected SEH_ instruction");
1783 }
1784}
1785
1786static void addConstantComments(const MachineInstr *MI,
1787 MCStreamer &OutStreamer) {
1788 switch (MI->getOpcode()) {
1789 // Lower PSHUFB and VPERMILP normally but add a comment if we can find
1790 // a constant shuffle mask. We won't be able to do this at the MC layer
1791 // because the mask isn't an immediate.
1792 case X86::PSHUFBrm:
1793 case X86::VPSHUFBrm:
1794 case X86::VPSHUFBYrm:
1795 case X86::VPSHUFBZ128rm:
1796 case X86::VPSHUFBZ128rmk:
1797 case X86::VPSHUFBZ128rmkz:
1798 case X86::VPSHUFBZ256rm:
1799 case X86::VPSHUFBZ256rmk:
1800 case X86::VPSHUFBZ256rmkz:
1801 case X86::VPSHUFBZrm:
1802 case X86::VPSHUFBZrmk:
1803 case X86::VPSHUFBZrmkz: {
1804 unsigned SrcIdx = getSrcIdx(MI, SrcIdx: 1);
1805 if (auto *C = X86::getConstantFromPool(MI: *MI, OpNo: SrcIdx + 1)) {
1806 unsigned Width = X86::getVectorRegisterWidth(Info: MI->getDesc().operands()[0]);
1807 SmallVector<int, 64> Mask;
1808 DecodePSHUFBMask(C, Width, ShuffleMask&: Mask);
1809 if (!Mask.empty())
1810 OutStreamer.AddComment(T: getShuffleComment(MI, SrcOp1Idx: SrcIdx, SrcOp2Idx: SrcIdx, Mask));
1811 }
1812 break;
1813 }
1814
1815 case X86::VPERMILPSrm:
1816 case X86::VPERMILPSYrm:
1817 case X86::VPERMILPSZ128rm:
1818 case X86::VPERMILPSZ128rmk:
1819 case X86::VPERMILPSZ128rmkz:
1820 case X86::VPERMILPSZ256rm:
1821 case X86::VPERMILPSZ256rmk:
1822 case X86::VPERMILPSZ256rmkz:
1823 case X86::VPERMILPSZrm:
1824 case X86::VPERMILPSZrmk:
1825 case X86::VPERMILPSZrmkz: {
1826 unsigned SrcIdx = getSrcIdx(MI, SrcIdx: 1);
1827 if (auto *C = X86::getConstantFromPool(MI: *MI, OpNo: SrcIdx + 1)) {
1828 unsigned Width = X86::getVectorRegisterWidth(Info: MI->getDesc().operands()[0]);
1829 SmallVector<int, 16> Mask;
1830 DecodeVPERMILPMask(C, ElSize: 32, Width, ShuffleMask&: Mask);
1831 if (!Mask.empty())
1832 OutStreamer.AddComment(T: getShuffleComment(MI, SrcOp1Idx: SrcIdx, SrcOp2Idx: SrcIdx, Mask));
1833 }
1834 break;
1835 }
1836 case X86::VPERMILPDrm:
1837 case X86::VPERMILPDYrm:
1838 case X86::VPERMILPDZ128rm:
1839 case X86::VPERMILPDZ128rmk:
1840 case X86::VPERMILPDZ128rmkz:
1841 case X86::VPERMILPDZ256rm:
1842 case X86::VPERMILPDZ256rmk:
1843 case X86::VPERMILPDZ256rmkz:
1844 case X86::VPERMILPDZrm:
1845 case X86::VPERMILPDZrmk:
1846 case X86::VPERMILPDZrmkz: {
1847 unsigned SrcIdx = getSrcIdx(MI, SrcIdx: 1);
1848 if (auto *C = X86::getConstantFromPool(MI: *MI, OpNo: SrcIdx + 1)) {
1849 unsigned Width = X86::getVectorRegisterWidth(Info: MI->getDesc().operands()[0]);
1850 SmallVector<int, 16> Mask;
1851 DecodeVPERMILPMask(C, ElSize: 64, Width, ShuffleMask&: Mask);
1852 if (!Mask.empty())
1853 OutStreamer.AddComment(T: getShuffleComment(MI, SrcOp1Idx: SrcIdx, SrcOp2Idx: SrcIdx, Mask));
1854 }
1855 break;
1856 }
1857
1858 case X86::VPERMIL2PDrm:
1859 case X86::VPERMIL2PSrm:
1860 case X86::VPERMIL2PDYrm:
1861 case X86::VPERMIL2PSYrm: {
1862 assert(MI->getNumOperands() >= (3 + X86::AddrNumOperands + 1) &&
1863 "Unexpected number of operands!");
1864
1865 const MachineOperand &CtrlOp = MI->getOperand(i: MI->getNumOperands() - 1);
1866 if (!CtrlOp.isImm())
1867 break;
1868
1869 unsigned ElSize;
1870 switch (MI->getOpcode()) {
1871 default: llvm_unreachable("Invalid opcode");
1872 case X86::VPERMIL2PSrm: case X86::VPERMIL2PSYrm: ElSize = 32; break;
1873 case X86::VPERMIL2PDrm: case X86::VPERMIL2PDYrm: ElSize = 64; break;
1874 }
1875
1876 if (auto *C = X86::getConstantFromPool(MI: *MI, OpNo: 3)) {
1877 unsigned Width = X86::getVectorRegisterWidth(Info: MI->getDesc().operands()[0]);
1878 SmallVector<int, 16> Mask;
1879 DecodeVPERMIL2PMask(C, M2Z: (unsigned)CtrlOp.getImm(), ElSize, Width, ShuffleMask&: Mask);
1880 if (!Mask.empty())
1881 OutStreamer.AddComment(T: getShuffleComment(MI, SrcOp1Idx: 1, SrcOp2Idx: 2, Mask));
1882 }
1883 break;
1884 }
1885
1886 case X86::VPPERMrrm: {
1887 if (auto *C = X86::getConstantFromPool(MI: *MI, OpNo: 3)) {
1888 unsigned Width = X86::getVectorRegisterWidth(Info: MI->getDesc().operands()[0]);
1889 SmallVector<int, 16> Mask;
1890 DecodeVPPERMMask(C, Width, ShuffleMask&: Mask);
1891 if (!Mask.empty())
1892 OutStreamer.AddComment(T: getShuffleComment(MI, SrcOp1Idx: 1, SrcOp2Idx: 2, Mask));
1893 }
1894 break;
1895 }
1896
1897 case X86::MMX_MOVQ64rm: {
1898 if (auto *C = X86::getConstantFromPool(MI: *MI, OpNo: 1)) {
1899 std::string Comment;
1900 raw_string_ostream CS(Comment);
1901 const MachineOperand &DstOp = MI->getOperand(i: 0);
1902 CS << X86ATTInstPrinter::getRegisterName(Reg: DstOp.getReg()) << " = ";
1903 if (auto *CF = dyn_cast<ConstantFP>(Val: C)) {
1904 CS << "0x" << toString(I: CF->getValueAPF().bitcastToAPInt(), Radix: 16, Signed: false);
1905 OutStreamer.AddComment(T: CS.str());
1906 }
1907 }
1908 break;
1909 }
1910
1911#define INSTR_CASE(Prefix, Instr, Suffix, Postfix) \
1912 case X86::Prefix##Instr##Suffix##rm##Postfix:
1913
1914#define CASE_ARITH_RM(Instr) \
1915 INSTR_CASE(, Instr, , ) /* SSE */ \
1916 INSTR_CASE(V, Instr, , ) /* AVX-128 */ \
1917 INSTR_CASE(V, Instr, Y, ) /* AVX-256 */ \
1918 INSTR_CASE(V, Instr, Z128, ) \
1919 INSTR_CASE(V, Instr, Z128, k) \
1920 INSTR_CASE(V, Instr, Z128, kz) \
1921 INSTR_CASE(V, Instr, Z256, ) \
1922 INSTR_CASE(V, Instr, Z256, k) \
1923 INSTR_CASE(V, Instr, Z256, kz) \
1924 INSTR_CASE(V, Instr, Z, ) \
1925 INSTR_CASE(V, Instr, Z, k) \
1926 INSTR_CASE(V, Instr, Z, kz)
1927
1928 // TODO: Add additional instructions when useful.
1929 CASE_ARITH_RM(PMADDUBSW) {
1930 unsigned SrcIdx = getSrcIdx(MI, SrcIdx: 1);
1931 if (auto *C = X86::getConstantFromPool(MI: *MI, OpNo: SrcIdx + 1)) {
1932 if (C->getType()->getScalarSizeInBits() == 8) {
1933 std::string Comment;
1934 raw_string_ostream CS(Comment);
1935 unsigned VectorWidth =
1936 X86::getVectorRegisterWidth(Info: MI->getDesc().operands()[0]);
1937 CS << "[";
1938 printConstant(COp: C, BitWidth: VectorWidth, CS);
1939 CS << "]";
1940 OutStreamer.AddComment(T: CS.str());
1941 }
1942 }
1943 break;
1944 }
1945
1946 CASE_ARITH_RM(PMADDWD)
1947 CASE_ARITH_RM(PMULLW)
1948 CASE_ARITH_RM(PMULHW)
1949 CASE_ARITH_RM(PMULHUW)
1950 CASE_ARITH_RM(PMULHRSW) {
1951 unsigned SrcIdx = getSrcIdx(MI, SrcIdx: 1);
1952 if (auto *C = X86::getConstantFromPool(MI: *MI, OpNo: SrcIdx + 1)) {
1953 if (C->getType()->getScalarSizeInBits() == 16) {
1954 std::string Comment;
1955 raw_string_ostream CS(Comment);
1956 unsigned VectorWidth =
1957 X86::getVectorRegisterWidth(Info: MI->getDesc().operands()[0]);
1958 CS << "[";
1959 printConstant(COp: C, BitWidth: VectorWidth, CS);
1960 CS << "]";
1961 OutStreamer.AddComment(T: CS.str());
1962 }
1963 }
1964 break;
1965 }
1966
1967#define MASK_AVX512_CASE(Instr) \
1968 case Instr: \
1969 case Instr##k: \
1970 case Instr##kz:
1971
1972 case X86::MOVSDrm:
1973 case X86::VMOVSDrm:
1974 MASK_AVX512_CASE(X86::VMOVSDZrm)
1975 case X86::MOVSDrm_alt:
1976 case X86::VMOVSDrm_alt:
1977 case X86::VMOVSDZrm_alt:
1978 case X86::MOVQI2PQIrm:
1979 case X86::VMOVQI2PQIrm:
1980 case X86::VMOVQI2PQIZrm:
1981 printZeroUpperMove(MI, OutStreamer, SclWidth: 64, VecWidth: 128, ShuffleComment: "mem[0],zero");
1982 break;
1983
1984 MASK_AVX512_CASE(X86::VMOVSHZrm)
1985 case X86::VMOVSHZrm_alt:
1986 printZeroUpperMove(MI, OutStreamer, SclWidth: 16, VecWidth: 128,
1987 ShuffleComment: "mem[0],zero,zero,zero,zero,zero,zero,zero");
1988 break;
1989
1990 case X86::MOVSSrm:
1991 case X86::VMOVSSrm:
1992 MASK_AVX512_CASE(X86::VMOVSSZrm)
1993 case X86::MOVSSrm_alt:
1994 case X86::VMOVSSrm_alt:
1995 case X86::VMOVSSZrm_alt:
1996 case X86::MOVDI2PDIrm:
1997 case X86::VMOVDI2PDIrm:
1998 case X86::VMOVDI2PDIZrm:
1999 printZeroUpperMove(MI, OutStreamer, SclWidth: 32, VecWidth: 128, ShuffleComment: "mem[0],zero,zero,zero");
2000 break;
2001
2002#define MOV_CASE(Prefix, Suffix) \
2003 case X86::Prefix##MOVAPD##Suffix##rm: \
2004 case X86::Prefix##MOVAPS##Suffix##rm: \
2005 case X86::Prefix##MOVUPD##Suffix##rm: \
2006 case X86::Prefix##MOVUPS##Suffix##rm: \
2007 case X86::Prefix##MOVDQA##Suffix##rm: \
2008 case X86::Prefix##MOVDQU##Suffix##rm:
2009
2010#define MOV_AVX512_CASE(Suffix, Postfix) \
2011 case X86::VMOVDQA64##Suffix##rm##Postfix: \
2012 case X86::VMOVDQA32##Suffix##rm##Postfix: \
2013 case X86::VMOVDQU64##Suffix##rm##Postfix: \
2014 case X86::VMOVDQU32##Suffix##rm##Postfix: \
2015 case X86::VMOVDQU16##Suffix##rm##Postfix: \
2016 case X86::VMOVDQU8##Suffix##rm##Postfix: \
2017 case X86::VMOVAPS##Suffix##rm##Postfix: \
2018 case X86::VMOVAPD##Suffix##rm##Postfix: \
2019 case X86::VMOVUPS##Suffix##rm##Postfix: \
2020 case X86::VMOVUPD##Suffix##rm##Postfix:
2021
2022#define CASE_128_MOV_RM() \
2023 MOV_CASE(, ) /* SSE */ \
2024 MOV_CASE(V, ) /* AVX-128 */ \
2025 MOV_AVX512_CASE(Z128, ) \
2026 MOV_AVX512_CASE(Z128, k) \
2027 MOV_AVX512_CASE(Z128, kz)
2028
2029#define CASE_256_MOV_RM() \
2030 MOV_CASE(V, Y) /* AVX-256 */ \
2031 MOV_AVX512_CASE(Z256, ) \
2032 MOV_AVX512_CASE(Z256, k) \
2033 MOV_AVX512_CASE(Z256, kz) \
2034
2035#define CASE_512_MOV_RM() \
2036 MOV_AVX512_CASE(Z, ) \
2037 MOV_AVX512_CASE(Z, k) \
2038 MOV_AVX512_CASE(Z, kz) \
2039
2040 // For loads from a constant pool to a vector register, print the constant
2041 // loaded.
2042 CASE_128_MOV_RM()
2043 printBroadcast(MI, OutStreamer, Repeats: 1, BitWidth: 128);
2044 break;
2045 CASE_256_MOV_RM()
2046 printBroadcast(MI, OutStreamer, Repeats: 1, BitWidth: 256);
2047 break;
2048 CASE_512_MOV_RM()
2049 printBroadcast(MI, OutStreamer, Repeats: 1, BitWidth: 512);
2050 break;
2051 case X86::VBROADCASTF128rm:
2052 case X86::VBROADCASTI128rm:
2053 MASK_AVX512_CASE(X86::VBROADCASTF32X4Z256rm)
2054 MASK_AVX512_CASE(X86::VBROADCASTF64X2Z128rm)
2055 MASK_AVX512_CASE(X86::VBROADCASTI32X4Z256rm)
2056 MASK_AVX512_CASE(X86::VBROADCASTI64X2Z128rm)
2057 printBroadcast(MI, OutStreamer, Repeats: 2, BitWidth: 128);
2058 break;
2059 MASK_AVX512_CASE(X86::VBROADCASTF32X4rm)
2060 MASK_AVX512_CASE(X86::VBROADCASTF64X2rm)
2061 MASK_AVX512_CASE(X86::VBROADCASTI32X4rm)
2062 MASK_AVX512_CASE(X86::VBROADCASTI64X2rm)
2063 printBroadcast(MI, OutStreamer, Repeats: 4, BitWidth: 128);
2064 break;
2065 MASK_AVX512_CASE(X86::VBROADCASTF32X8rm)
2066 MASK_AVX512_CASE(X86::VBROADCASTF64X4rm)
2067 MASK_AVX512_CASE(X86::VBROADCASTI32X8rm)
2068 MASK_AVX512_CASE(X86::VBROADCASTI64X4rm)
2069 printBroadcast(MI, OutStreamer, Repeats: 2, BitWidth: 256);
2070 break;
2071
2072 // For broadcast loads from a constant pool to a vector register, repeatedly
2073 // print the constant loaded.
2074 case X86::MOVDDUPrm:
2075 case X86::VMOVDDUPrm:
2076 MASK_AVX512_CASE(X86::VMOVDDUPZ128rm)
2077 case X86::VPBROADCASTQrm:
2078 MASK_AVX512_CASE(X86::VPBROADCASTQZ128rm)
2079 printBroadcast(MI, OutStreamer, Repeats: 2, BitWidth: 64);
2080 break;
2081 case X86::VBROADCASTSDYrm:
2082 MASK_AVX512_CASE(X86::VBROADCASTSDZ256rm)
2083 case X86::VPBROADCASTQYrm:
2084 MASK_AVX512_CASE(X86::VPBROADCASTQZ256rm)
2085 printBroadcast(MI, OutStreamer, Repeats: 4, BitWidth: 64);
2086 break;
2087 MASK_AVX512_CASE(X86::VBROADCASTSDZrm)
2088 MASK_AVX512_CASE(X86::VPBROADCASTQZrm)
2089 printBroadcast(MI, OutStreamer, Repeats: 8, BitWidth: 64);
2090 break;
2091 case X86::VBROADCASTSSrm:
2092 MASK_AVX512_CASE(X86::VBROADCASTSSZ128rm)
2093 case X86::VPBROADCASTDrm:
2094 MASK_AVX512_CASE(X86::VPBROADCASTDZ128rm)
2095 printBroadcast(MI, OutStreamer, Repeats: 4, BitWidth: 32);
2096 break;
2097 case X86::VBROADCASTSSYrm:
2098 MASK_AVX512_CASE(X86::VBROADCASTSSZ256rm)
2099 case X86::VPBROADCASTDYrm:
2100 MASK_AVX512_CASE(X86::VPBROADCASTDZ256rm)
2101 printBroadcast(MI, OutStreamer, Repeats: 8, BitWidth: 32);
2102 break;
2103 MASK_AVX512_CASE(X86::VBROADCASTSSZrm)
2104 MASK_AVX512_CASE(X86::VPBROADCASTDZrm)
2105 printBroadcast(MI, OutStreamer, Repeats: 16, BitWidth: 32);
2106 break;
2107 case X86::VPBROADCASTWrm:
2108 MASK_AVX512_CASE(X86::VPBROADCASTWZ128rm)
2109 printBroadcast(MI, OutStreamer, Repeats: 8, BitWidth: 16);
2110 break;
2111 case X86::VPBROADCASTWYrm:
2112 MASK_AVX512_CASE(X86::VPBROADCASTWZ256rm)
2113 printBroadcast(MI, OutStreamer, Repeats: 16, BitWidth: 16);
2114 break;
2115 MASK_AVX512_CASE(X86::VPBROADCASTWZrm)
2116 printBroadcast(MI, OutStreamer, Repeats: 32, BitWidth: 16);
2117 break;
2118 case X86::VPBROADCASTBrm:
2119 MASK_AVX512_CASE(X86::VPBROADCASTBZ128rm)
2120 printBroadcast(MI, OutStreamer, Repeats: 16, BitWidth: 8);
2121 break;
2122 case X86::VPBROADCASTBYrm:
2123 MASK_AVX512_CASE(X86::VPBROADCASTBZ256rm)
2124 printBroadcast(MI, OutStreamer, Repeats: 32, BitWidth: 8);
2125 break;
2126 MASK_AVX512_CASE(X86::VPBROADCASTBZrm)
2127 printBroadcast(MI, OutStreamer, Repeats: 64, BitWidth: 8);
2128 break;
2129
2130#define MOVX_CASE(Prefix, Ext, Type, Suffix, Postfix) \
2131 case X86::Prefix##PMOV##Ext##Type##Suffix##rm##Postfix:
2132
2133#define CASE_MOVX_RM(Ext, Type) \
2134 MOVX_CASE(, Ext, Type, , ) \
2135 MOVX_CASE(V, Ext, Type, , ) \
2136 MOVX_CASE(V, Ext, Type, Y, ) \
2137 MOVX_CASE(V, Ext, Type, Z128, ) \
2138 MOVX_CASE(V, Ext, Type, Z128, k ) \
2139 MOVX_CASE(V, Ext, Type, Z128, kz ) \
2140 MOVX_CASE(V, Ext, Type, Z256, ) \
2141 MOVX_CASE(V, Ext, Type, Z256, k ) \
2142 MOVX_CASE(V, Ext, Type, Z256, kz ) \
2143 MOVX_CASE(V, Ext, Type, Z, ) \
2144 MOVX_CASE(V, Ext, Type, Z, k ) \
2145 MOVX_CASE(V, Ext, Type, Z, kz )
2146
2147 CASE_MOVX_RM(SX, BD)
2148 printSignExtend(MI, OutStreamer, SrcEltBits: 8, DstEltBits: 32);
2149 break;
2150 CASE_MOVX_RM(SX, BQ)
2151 printSignExtend(MI, OutStreamer, SrcEltBits: 8, DstEltBits: 64);
2152 break;
2153 CASE_MOVX_RM(SX, BW)
2154 printSignExtend(MI, OutStreamer, SrcEltBits: 8, DstEltBits: 16);
2155 break;
2156 CASE_MOVX_RM(SX, DQ)
2157 printSignExtend(MI, OutStreamer, SrcEltBits: 32, DstEltBits: 64);
2158 break;
2159 CASE_MOVX_RM(SX, WD)
2160 printSignExtend(MI, OutStreamer, SrcEltBits: 16, DstEltBits: 32);
2161 break;
2162 CASE_MOVX_RM(SX, WQ)
2163 printSignExtend(MI, OutStreamer, SrcEltBits: 16, DstEltBits: 64);
2164 break;
2165
2166 CASE_MOVX_RM(ZX, BD)
2167 printZeroExtend(MI, OutStreamer, SrcEltBits: 8, DstEltBits: 32);
2168 break;
2169 CASE_MOVX_RM(ZX, BQ)
2170 printZeroExtend(MI, OutStreamer, SrcEltBits: 8, DstEltBits: 64);
2171 break;
2172 CASE_MOVX_RM(ZX, BW)
2173 printZeroExtend(MI, OutStreamer, SrcEltBits: 8, DstEltBits: 16);
2174 break;
2175 CASE_MOVX_RM(ZX, DQ)
2176 printZeroExtend(MI, OutStreamer, SrcEltBits: 32, DstEltBits: 64);
2177 break;
2178 CASE_MOVX_RM(ZX, WD)
2179 printZeroExtend(MI, OutStreamer, SrcEltBits: 16, DstEltBits: 32);
2180 break;
2181 CASE_MOVX_RM(ZX, WQ)
2182 printZeroExtend(MI, OutStreamer, SrcEltBits: 16, DstEltBits: 64);
2183 break;
2184 }
2185}
2186
2187void X86AsmPrinter::emitInstruction(const MachineInstr *MI) {
2188 // FIXME: Enable feature predicate checks once all the test pass.
2189 // X86_MC::verifyInstructionPredicates(MI->getOpcode(),
2190 // Subtarget->getFeatureBits());
2191
2192 X86MCInstLower MCInstLowering(*MF, *this);
2193 const X86RegisterInfo *RI =
2194 MF->getSubtarget<X86Subtarget>().getRegisterInfo();
2195
2196 if (MI->getOpcode() == X86::OR64rm) {
2197 for (auto &Opd : MI->operands()) {
2198 if (Opd.isSymbol() && StringRef(Opd.getSymbolName()) ==
2199 "swift_async_extendedFramePointerFlags") {
2200 ShouldEmitWeakSwiftAsyncExtendedFramePointerFlags = true;
2201 }
2202 }
2203 }
2204
2205 // Add comments for values loaded from constant pool.
2206 if (OutStreamer->isVerboseAsm())
2207 addConstantComments(MI, OutStreamer&: *OutStreamer);
2208
2209 // Add a comment about EVEX compression
2210 if (TM.Options.MCOptions.ShowMCEncoding) {
2211 if (MI->getAsmPrinterFlags() & X86::AC_EVEX_2_LEGACY)
2212 OutStreamer->AddComment(T: "EVEX TO LEGACY Compression ", EOL: false);
2213 else if (MI->getAsmPrinterFlags() & X86::AC_EVEX_2_VEX)
2214 OutStreamer->AddComment(T: "EVEX TO VEX Compression ", EOL: false);
2215 else if (MI->getAsmPrinterFlags() & X86::AC_EVEX_2_EVEX)
2216 OutStreamer->AddComment(T: "EVEX TO EVEX Compression ", EOL: false);
2217 }
2218
2219 switch (MI->getOpcode()) {
2220 case TargetOpcode::DBG_VALUE:
2221 llvm_unreachable("Should be handled target independently");
2222
2223 case X86::EH_RETURN:
2224 case X86::EH_RETURN64: {
2225 // Lower these as normal, but add some comments.
2226 Register Reg = MI->getOperand(i: 0).getReg();
2227 OutStreamer->AddComment(T: StringRef("eh_return, addr: %") +
2228 X86ATTInstPrinter::getRegisterName(Reg));
2229 break;
2230 }
2231 case X86::CLEANUPRET: {
2232 // Lower these as normal, but add some comments.
2233 OutStreamer->AddComment(T: "CLEANUPRET");
2234 break;
2235 }
2236
2237 case X86::CATCHRET: {
2238 // Lower these as normal, but add some comments.
2239 OutStreamer->AddComment(T: "CATCHRET");
2240 break;
2241 }
2242
2243 case X86::ENDBR32:
2244 case X86::ENDBR64: {
2245 // CurrentPatchableFunctionEntrySym can be CurrentFnBegin only for
2246 // -fpatchable-function-entry=N,0. The entry MBB is guaranteed to be
2247 // non-empty. If MI is the initial ENDBR, place the
2248 // __patchable_function_entries label after ENDBR.
2249 if (CurrentPatchableFunctionEntrySym &&
2250 CurrentPatchableFunctionEntrySym == CurrentFnBegin &&
2251 MI == &MF->front().front()) {
2252 MCInst Inst;
2253 MCInstLowering.Lower(MI, OutMI&: Inst);
2254 EmitAndCountInstruction(Inst);
2255 CurrentPatchableFunctionEntrySym = createTempSymbol(Name: "patch");
2256 OutStreamer->emitLabel(Symbol: CurrentPatchableFunctionEntrySym);
2257 return;
2258 }
2259 break;
2260 }
2261
2262 case X86::TAILJMPd64:
2263 if (IndCSPrefix && MI->hasRegisterImplicitUseOperand(Reg: X86::R11))
2264 EmitAndCountInstruction(Inst&: MCInstBuilder(X86::CS_PREFIX));
2265 [[fallthrough]];
2266 case X86::TAILJMPr:
2267 case X86::TAILJMPm:
2268 case X86::TAILJMPd:
2269 case X86::TAILJMPd_CC:
2270 case X86::TAILJMPr64:
2271 case X86::TAILJMPm64:
2272 case X86::TAILJMPd64_CC:
2273 case X86::TAILJMPr64_REX:
2274 case X86::TAILJMPm64_REX:
2275 // Lower these as normal, but add some comments.
2276 OutStreamer->AddComment(T: "TAILCALL");
2277 break;
2278
2279 case X86::TLS_addr32:
2280 case X86::TLS_addr64:
2281 case X86::TLS_addrX32:
2282 case X86::TLS_base_addr32:
2283 case X86::TLS_base_addr64:
2284 case X86::TLS_base_addrX32:
2285 case X86::TLS_desc32:
2286 case X86::TLS_desc64:
2287 return LowerTlsAddr(MCInstLowering, MI: *MI);
2288
2289 case X86::MOVPC32r: {
2290 // This is a pseudo op for a two instruction sequence with a label, which
2291 // looks like:
2292 // call "L1$pb"
2293 // "L1$pb":
2294 // popl %esi
2295
2296 // Emit the call.
2297 MCSymbol *PICBase = MF->getPICBaseSymbol();
2298 // FIXME: We would like an efficient form for this, so we don't have to do a
2299 // lot of extra uniquing.
2300 EmitAndCountInstruction(
2301 Inst&: MCInstBuilder(X86::CALLpcrel32)
2302 .addExpr(Val: MCSymbolRefExpr::create(Symbol: PICBase, Ctx&: OutContext)));
2303
2304 const X86FrameLowering *FrameLowering =
2305 MF->getSubtarget<X86Subtarget>().getFrameLowering();
2306 bool hasFP = FrameLowering->hasFP(MF: *MF);
2307
2308 // TODO: This is needed only if we require precise CFA.
2309 bool HasActiveDwarfFrame = OutStreamer->getNumFrameInfos() &&
2310 !OutStreamer->getDwarfFrameInfos().back().End;
2311
2312 int stackGrowth = -RI->getSlotSize();
2313
2314 if (HasActiveDwarfFrame && !hasFP) {
2315 OutStreamer->emitCFIAdjustCfaOffset(Adjustment: -stackGrowth);
2316 MF->getInfo<X86MachineFunctionInfo>()->setHasCFIAdjustCfa(true);
2317 }
2318
2319 // Emit the label.
2320 OutStreamer->emitLabel(Symbol: PICBase);
2321
2322 // popl $reg
2323 EmitAndCountInstruction(
2324 Inst&: MCInstBuilder(X86::POP32r).addReg(Reg: MI->getOperand(i: 0).getReg()));
2325
2326 if (HasActiveDwarfFrame && !hasFP) {
2327 OutStreamer->emitCFIAdjustCfaOffset(Adjustment: stackGrowth);
2328 }
2329 return;
2330 }
2331
2332 case X86::ADD32ri: {
2333 // Lower the MO_GOT_ABSOLUTE_ADDRESS form of ADD32ri.
2334 if (MI->getOperand(i: 2).getTargetFlags() != X86II::MO_GOT_ABSOLUTE_ADDRESS)
2335 break;
2336
2337 // Okay, we have something like:
2338 // EAX = ADD32ri EAX, MO_GOT_ABSOLUTE_ADDRESS(@MYGLOBAL)
2339
2340 // For this, we want to print something like:
2341 // MYGLOBAL + (. - PICBASE)
2342 // However, we can't generate a ".", so just emit a new label here and refer
2343 // to it.
2344 MCSymbol *DotSym = OutContext.createTempSymbol();
2345 OutStreamer->emitLabel(Symbol: DotSym);
2346
2347 // Now that we have emitted the label, lower the complex operand expression.
2348 MCSymbol *OpSym = MCInstLowering.GetSymbolFromOperand(MO: MI->getOperand(i: 2));
2349
2350 const MCExpr *DotExpr = MCSymbolRefExpr::create(Symbol: DotSym, Ctx&: OutContext);
2351 const MCExpr *PICBase =
2352 MCSymbolRefExpr::create(Symbol: MF->getPICBaseSymbol(), Ctx&: OutContext);
2353 DotExpr = MCBinaryExpr::createSub(LHS: DotExpr, RHS: PICBase, Ctx&: OutContext);
2354
2355 DotExpr = MCBinaryExpr::createAdd(
2356 LHS: MCSymbolRefExpr::create(Symbol: OpSym, Ctx&: OutContext), RHS: DotExpr, Ctx&: OutContext);
2357
2358 EmitAndCountInstruction(Inst&: MCInstBuilder(X86::ADD32ri)
2359 .addReg(Reg: MI->getOperand(i: 0).getReg())
2360 .addReg(Reg: MI->getOperand(i: 1).getReg())
2361 .addExpr(Val: DotExpr));
2362 return;
2363 }
2364 case TargetOpcode::STATEPOINT:
2365 return LowerSTATEPOINT(MI: *MI, MCIL&: MCInstLowering);
2366
2367 case TargetOpcode::FAULTING_OP:
2368 return LowerFAULTING_OP(FaultingMI: *MI, MCIL&: MCInstLowering);
2369
2370 case TargetOpcode::FENTRY_CALL:
2371 return LowerFENTRY_CALL(MI: *MI, MCIL&: MCInstLowering);
2372
2373 case TargetOpcode::PATCHABLE_OP:
2374 return LowerPATCHABLE_OP(MI: *MI, MCIL&: MCInstLowering);
2375
2376 case TargetOpcode::STACKMAP:
2377 return LowerSTACKMAP(MI: *MI);
2378
2379 case TargetOpcode::PATCHPOINT:
2380 return LowerPATCHPOINT(MI: *MI, MCIL&: MCInstLowering);
2381
2382 case TargetOpcode::PATCHABLE_FUNCTION_ENTER:
2383 return LowerPATCHABLE_FUNCTION_ENTER(MI: *MI, MCIL&: MCInstLowering);
2384
2385 case TargetOpcode::PATCHABLE_RET:
2386 return LowerPATCHABLE_RET(MI: *MI, MCIL&: MCInstLowering);
2387
2388 case TargetOpcode::PATCHABLE_TAIL_CALL:
2389 return LowerPATCHABLE_TAIL_CALL(MI: *MI, MCIL&: MCInstLowering);
2390
2391 case TargetOpcode::PATCHABLE_EVENT_CALL:
2392 return LowerPATCHABLE_EVENT_CALL(MI: *MI, MCIL&: MCInstLowering);
2393
2394 case TargetOpcode::PATCHABLE_TYPED_EVENT_CALL:
2395 return LowerPATCHABLE_TYPED_EVENT_CALL(MI: *MI, MCIL&: MCInstLowering);
2396
2397 case X86::MORESTACK_RET:
2398 EmitAndCountInstruction(Inst&: MCInstBuilder(getRetOpcode(Subtarget: *Subtarget)));
2399 return;
2400
2401 case X86::KCFI_CHECK:
2402 return LowerKCFI_CHECK(MI: *MI);
2403
2404 case X86::ASAN_CHECK_MEMACCESS:
2405 return LowerASAN_CHECK_MEMACCESS(MI: *MI);
2406
2407 case X86::MORESTACK_RET_RESTORE_R10:
2408 // Return, then restore R10.
2409 EmitAndCountInstruction(Inst&: MCInstBuilder(getRetOpcode(Subtarget: *Subtarget)));
2410 EmitAndCountInstruction(
2411 Inst&: MCInstBuilder(X86::MOV64rr).addReg(Reg: X86::R10).addReg(Reg: X86::RAX));
2412 return;
2413
2414 case X86::SEH_PushReg:
2415 case X86::SEH_SaveReg:
2416 case X86::SEH_SaveXMM:
2417 case X86::SEH_StackAlloc:
2418 case X86::SEH_StackAlign:
2419 case X86::SEH_SetFrame:
2420 case X86::SEH_PushFrame:
2421 case X86::SEH_EndPrologue:
2422 EmitSEHInstruction(MI);
2423 return;
2424
2425 case X86::SEH_Epilogue: {
2426 assert(MF->hasWinCFI() && "SEH_ instruction in function without WinCFI?");
2427 MachineBasicBlock::const_iterator MBBI(MI);
2428 // Check if preceded by a call and emit nop if so.
2429 for (MBBI = PrevCrossBBInst(MBBI);
2430 MBBI != MachineBasicBlock::const_iterator();
2431 MBBI = PrevCrossBBInst(MBBI)) {
2432 // Pseudo instructions that aren't a call are assumed to not emit any
2433 // code. If they do, we worst case generate unnecessary noops after a
2434 // call.
2435 if (MBBI->isCall() || !MBBI->isPseudo()) {
2436 if (MBBI->isCall())
2437 EmitAndCountInstruction(Inst&: MCInstBuilder(X86::NOOP));
2438 break;
2439 }
2440 }
2441 return;
2442 }
2443 case X86::UBSAN_UD1:
2444 EmitAndCountInstruction(Inst&: MCInstBuilder(X86::UD1Lm)
2445 .addReg(Reg: X86::EAX)
2446 .addReg(Reg: X86::EAX)
2447 .addImm(Val: 1)
2448 .addReg(Reg: X86::NoRegister)
2449 .addImm(Val: MI->getOperand(i: 0).getImm())
2450 .addReg(Reg: X86::NoRegister));
2451 return;
2452 case X86::CALL64pcrel32:
2453 if (IndCSPrefix && MI->hasRegisterImplicitUseOperand(Reg: X86::R11))
2454 EmitAndCountInstruction(Inst&: MCInstBuilder(X86::CS_PREFIX));
2455 break;
2456 case X86::JCC_1:
2457 // Two instruction prefixes (2EH for branch not-taken and 3EH for branch
2458 // taken) are used as branch hints. Here we add branch taken prefix for
2459 // jump instruction with higher probability than threshold.
2460 if (getSubtarget().hasBranchHint() && EnableBranchHint) {
2461 const MachineBranchProbabilityInfo *MBPI =
2462 &getAnalysis<MachineBranchProbabilityInfoWrapperPass>().getMBPI();
2463 MachineBasicBlock *DestBB = MI->getOperand(i: 0).getMBB();
2464 BranchProbability EdgeProb =
2465 MBPI->getEdgeProbability(Src: MI->getParent(), Dst: DestBB);
2466 BranchProbability Threshold(BranchHintProbabilityThreshold, 100);
2467 if (EdgeProb > Threshold)
2468 EmitAndCountInstruction(Inst&: MCInstBuilder(X86::DS_PREFIX));
2469 }
2470 break;
2471 }
2472
2473 MCInst TmpInst;
2474 MCInstLowering.Lower(MI, OutMI&: TmpInst);
2475
2476 // Stackmap shadows cannot include branch targets, so we can count the bytes
2477 // in a call towards the shadow, but must ensure that the no thread returns
2478 // in to the stackmap shadow. The only way to achieve this is if the call
2479 // is at the end of the shadow.
2480 if (MI->isCall()) {
2481 // Count then size of the call towards the shadow
2482 SMShadowTracker.count(Inst&: TmpInst, STI: getSubtargetInfo(), CodeEmitter: CodeEmitter.get());
2483 // Then flush the shadow so that we fill with nops before the call, not
2484 // after it.
2485 SMShadowTracker.emitShadowPadding(OutStreamer&: *OutStreamer, STI: getSubtargetInfo());
2486 // Then emit the call
2487 OutStreamer->emitInstruction(Inst: TmpInst, STI: getSubtargetInfo());
2488 return;
2489 }
2490
2491 EmitAndCountInstruction(Inst&: TmpInst);
2492}
2493