1 //===-- X86InstrInfo.cpp - X86 Instruction Information --------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains the X86 implementation of the TargetInstrInfo class.
12 //===----------------------------------------------------------------------===//
14 #include "X86InstrInfo.h"
16 #include "X86InstrBuilder.h"
17 #include "X86InstrFoldTables.h"
18 #include "X86MachineFunctionInfo.h"
19 #include "X86Subtarget.h"
20 #include "X86TargetMachine.h"
21 #include "llvm/ADT/STLExtras.h"
22 #include "llvm/ADT/Sequence.h"
23 #include "llvm/CodeGen/LivePhysRegs.h"
24 #include "llvm/CodeGen/LiveVariables.h"
25 #include "llvm/CodeGen/MachineConstantPool.h"
26 #include "llvm/CodeGen/MachineDominators.h"
27 #include "llvm/CodeGen/MachineFrameInfo.h"
28 #include "llvm/CodeGen/MachineInstrBuilder.h"
29 #include "llvm/CodeGen/MachineModuleInfo.h"
30 #include "llvm/CodeGen/MachineRegisterInfo.h"
31 #include "llvm/CodeGen/StackMaps.h"
32 #include "llvm/IR/DerivedTypes.h"
33 #include "llvm/IR/Function.h"
34 #include "llvm/IR/LLVMContext.h"
35 #include "llvm/MC/MCAsmInfo.h"
36 #include "llvm/MC/MCExpr.h"
37 #include "llvm/MC/MCInst.h"
38 #include "llvm/Support/CommandLine.h"
39 #include "llvm/Support/Debug.h"
40 #include "llvm/Support/ErrorHandling.h"
41 #include "llvm/Support/raw_ostream.h"
42 #include "llvm/Target/TargetOptions.h"
46 #define DEBUG_TYPE "x86-instr-info"
48 #define GET_INSTRINFO_CTOR_DTOR
49 #include "X86GenInstrInfo.inc"
52 NoFusing("disable-spill-fusing",
53 cl::desc("Disable fusing of spill code into instructions"),
56 PrintFailedFusing("print-failed-fuse-candidates",
57 cl::desc("Print instructions that the allocator wants to"
58 " fuse, but the X86 backend currently can't"),
61 ReMatPICStubLoad("remat-pic-stub-load",
62 cl::desc("Re-materialize load from stub in PIC mode"),
63 cl::init(false), cl::Hidden);
64 static cl::opt<unsigned>
65 PartialRegUpdateClearance("partial-reg-update-clearance",
66 cl::desc("Clearance between two register writes "
67 "for inserting XOR to avoid partial "
69 cl::init(64), cl::Hidden);
70 static cl::opt<unsigned>
71 UndefRegClearance("undef-reg-clearance",
72 cl::desc("How many idle instructions we would like before "
73 "certain undef register reads"),
74 cl::init(128), cl::Hidden);
77 // Pin the vtable to this file.
78 void X86InstrInfo::anchor() {}
80 X86InstrInfo::X86InstrInfo(X86Subtarget &STI)
81 : X86GenInstrInfo((STI.isTarget64BitLP64() ? X86::ADJCALLSTACKDOWN64
82 : X86::ADJCALLSTACKDOWN32),
83 (STI.isTarget64BitLP64() ? X86::ADJCALLSTACKUP64
84 : X86::ADJCALLSTACKUP32),
86 (STI.is64Bit() ? X86::RETQ : X86::RETL)),
87 Subtarget(STI), RI(STI.getTargetTriple()) {
91 X86InstrInfo::isCoalescableExtInstr(const MachineInstr &MI,
92 unsigned &SrcReg, unsigned &DstReg,
93 unsigned &SubIdx) const {
94 switch (MI.getOpcode()) {
100 case X86::MOVSX64rr8:
101 if (!Subtarget.is64Bit())
102 // It's not always legal to reference the low 8-bit of the larger
103 // register in 32-bit mode.
106 case X86::MOVSX32rr16:
107 case X86::MOVZX32rr16:
108 case X86::MOVSX64rr16:
109 case X86::MOVSX64rr32: {
110 if (MI.getOperand(0).getSubReg() || MI.getOperand(1).getSubReg())
113 SrcReg = MI.getOperand(1).getReg();
114 DstReg = MI.getOperand(0).getReg();
115 switch (MI.getOpcode()) {
116 default: llvm_unreachable("Unreachable!");
117 case X86::MOVSX16rr8:
118 case X86::MOVZX16rr8:
119 case X86::MOVSX32rr8:
120 case X86::MOVZX32rr8:
121 case X86::MOVSX64rr8:
122 SubIdx = X86::sub_8bit;
124 case X86::MOVSX32rr16:
125 case X86::MOVZX32rr16:
126 case X86::MOVSX64rr16:
127 SubIdx = X86::sub_16bit;
129 case X86::MOVSX64rr32:
130 SubIdx = X86::sub_32bit;
139 int X86InstrInfo::getSPAdjust(const MachineInstr &MI) const {
140 const MachineFunction *MF = MI.getParent()->getParent();
141 const TargetFrameLowering *TFI = MF->getSubtarget().getFrameLowering();
143 if (isFrameInstr(MI)) {
144 unsigned StackAlign = TFI->getStackAlignment();
145 int SPAdj = alignTo(getFrameSize(MI), StackAlign);
146 SPAdj -= getFrameAdjustment(MI);
147 if (!isFrameSetup(MI))
152 // To know whether a call adjusts the stack, we need information
153 // that is bound to the following ADJCALLSTACKUP pseudo.
154 // Look for the next ADJCALLSTACKUP that follows the call.
156 const MachineBasicBlock *MBB = MI.getParent();
157 auto I = ++MachineBasicBlock::const_iterator(MI);
158 for (auto E = MBB->end(); I != E; ++I) {
159 if (I->getOpcode() == getCallFrameDestroyOpcode() ||
164 // If we could not find a frame destroy opcode, then it has already
165 // been simplified, so we don't care.
166 if (I->getOpcode() != getCallFrameDestroyOpcode())
169 return -(I->getOperand(1).getImm());
172 // Currently handle only PUSHes we can reasonably expect to see
174 switch (MI.getOpcode()) {
192 /// Return true and the FrameIndex if the specified
193 /// operand and follow operands form a reference to the stack frame.
194 bool X86InstrInfo::isFrameOperand(const MachineInstr &MI, unsigned int Op,
195 int &FrameIndex) const {
196 if (MI.getOperand(Op + X86::AddrBaseReg).isFI() &&
197 MI.getOperand(Op + X86::AddrScaleAmt).isImm() &&
198 MI.getOperand(Op + X86::AddrIndexReg).isReg() &&
199 MI.getOperand(Op + X86::AddrDisp).isImm() &&
200 MI.getOperand(Op + X86::AddrScaleAmt).getImm() == 1 &&
201 MI.getOperand(Op + X86::AddrIndexReg).getReg() == 0 &&
202 MI.getOperand(Op + X86::AddrDisp).getImm() == 0) {
203 FrameIndex = MI.getOperand(Op + X86::AddrBaseReg).getIndex();
209 static bool isFrameLoadOpcode(int Opcode, unsigned &MemBytes) {
233 case X86::MMX_MOVD64rm:
234 case X86::MMX_MOVQ64rm:
250 case X86::VMOVAPSZ128rm:
251 case X86::VMOVUPSZ128rm:
252 case X86::VMOVAPSZ128rm_NOVLX:
253 case X86::VMOVUPSZ128rm_NOVLX:
254 case X86::VMOVAPDZ128rm:
255 case X86::VMOVUPDZ128rm:
256 case X86::VMOVDQU8Z128rm:
257 case X86::VMOVDQU16Z128rm:
258 case X86::VMOVDQA32Z128rm:
259 case X86::VMOVDQU32Z128rm:
260 case X86::VMOVDQA64Z128rm:
261 case X86::VMOVDQU64Z128rm:
264 case X86::VMOVAPSYrm:
265 case X86::VMOVUPSYrm:
266 case X86::VMOVAPDYrm:
267 case X86::VMOVUPDYrm:
268 case X86::VMOVDQAYrm:
269 case X86::VMOVDQUYrm:
270 case X86::VMOVAPSZ256rm:
271 case X86::VMOVUPSZ256rm:
272 case X86::VMOVAPSZ256rm_NOVLX:
273 case X86::VMOVUPSZ256rm_NOVLX:
274 case X86::VMOVAPDZ256rm:
275 case X86::VMOVUPDZ256rm:
276 case X86::VMOVDQU8Z256rm:
277 case X86::VMOVDQU16Z256rm:
278 case X86::VMOVDQA32Z256rm:
279 case X86::VMOVDQU32Z256rm:
280 case X86::VMOVDQA64Z256rm:
281 case X86::VMOVDQU64Z256rm:
284 case X86::VMOVAPSZrm:
285 case X86::VMOVUPSZrm:
286 case X86::VMOVAPDZrm:
287 case X86::VMOVUPDZrm:
288 case X86::VMOVDQU8Zrm:
289 case X86::VMOVDQU16Zrm:
290 case X86::VMOVDQA32Zrm:
291 case X86::VMOVDQU32Zrm:
292 case X86::VMOVDQA64Zrm:
293 case X86::VMOVDQU64Zrm:
299 static bool isFrameStoreOpcode(int Opcode, unsigned &MemBytes) {
323 case X86::MMX_MOVD64mr:
324 case X86::MMX_MOVQ64mr:
325 case X86::MMX_MOVNTQmr:
341 case X86::VMOVUPSZ128mr:
342 case X86::VMOVAPSZ128mr:
343 case X86::VMOVUPSZ128mr_NOVLX:
344 case X86::VMOVAPSZ128mr_NOVLX:
345 case X86::VMOVUPDZ128mr:
346 case X86::VMOVAPDZ128mr:
347 case X86::VMOVDQA32Z128mr:
348 case X86::VMOVDQU32Z128mr:
349 case X86::VMOVDQA64Z128mr:
350 case X86::VMOVDQU64Z128mr:
351 case X86::VMOVDQU8Z128mr:
352 case X86::VMOVDQU16Z128mr:
355 case X86::VMOVUPSYmr:
356 case X86::VMOVAPSYmr:
357 case X86::VMOVUPDYmr:
358 case X86::VMOVAPDYmr:
359 case X86::VMOVDQUYmr:
360 case X86::VMOVDQAYmr:
361 case X86::VMOVUPSZ256mr:
362 case X86::VMOVAPSZ256mr:
363 case X86::VMOVUPSZ256mr_NOVLX:
364 case X86::VMOVAPSZ256mr_NOVLX:
365 case X86::VMOVUPDZ256mr:
366 case X86::VMOVAPDZ256mr:
367 case X86::VMOVDQU8Z256mr:
368 case X86::VMOVDQU16Z256mr:
369 case X86::VMOVDQA32Z256mr:
370 case X86::VMOVDQU32Z256mr:
371 case X86::VMOVDQA64Z256mr:
372 case X86::VMOVDQU64Z256mr:
375 case X86::VMOVUPSZmr:
376 case X86::VMOVAPSZmr:
377 case X86::VMOVUPDZmr:
378 case X86::VMOVAPDZmr:
379 case X86::VMOVDQU8Zmr:
380 case X86::VMOVDQU16Zmr:
381 case X86::VMOVDQA32Zmr:
382 case X86::VMOVDQU32Zmr:
383 case X86::VMOVDQA64Zmr:
384 case X86::VMOVDQU64Zmr:
391 unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr &MI,
392 int &FrameIndex) const {
394 return X86InstrInfo::isLoadFromStackSlot(MI, FrameIndex, Dummy);
397 unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr &MI,
399 unsigned &MemBytes) const {
400 if (isFrameLoadOpcode(MI.getOpcode(), MemBytes))
401 if (MI.getOperand(0).getSubReg() == 0 && isFrameOperand(MI, 1, FrameIndex))
402 return MI.getOperand(0).getReg();
406 unsigned X86InstrInfo::isLoadFromStackSlotPostFE(const MachineInstr &MI,
407 int &FrameIndex) const {
409 if (isFrameLoadOpcode(MI.getOpcode(), Dummy)) {
411 if ((Reg = isLoadFromStackSlot(MI, FrameIndex)))
413 // Check for post-frame index elimination operations
414 SmallVector<const MachineMemOperand *, 1> Accesses;
415 if (hasLoadFromStackSlot(MI, Accesses)) {
417 cast<FixedStackPseudoSourceValue>(Accesses.front()->getPseudoValue())
425 unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr &MI,
426 int &FrameIndex) const {
428 return X86InstrInfo::isStoreToStackSlot(MI, FrameIndex, Dummy);
431 unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr &MI,
433 unsigned &MemBytes) const {
434 if (isFrameStoreOpcode(MI.getOpcode(), MemBytes))
435 if (MI.getOperand(X86::AddrNumOperands).getSubReg() == 0 &&
436 isFrameOperand(MI, 0, FrameIndex))
437 return MI.getOperand(X86::AddrNumOperands).getReg();
441 unsigned X86InstrInfo::isStoreToStackSlotPostFE(const MachineInstr &MI,
442 int &FrameIndex) const {
444 if (isFrameStoreOpcode(MI.getOpcode(), Dummy)) {
446 if ((Reg = isStoreToStackSlot(MI, FrameIndex)))
448 // Check for post-frame index elimination operations
449 SmallVector<const MachineMemOperand *, 1> Accesses;
450 if (hasStoreToStackSlot(MI, Accesses)) {
452 cast<FixedStackPseudoSourceValue>(Accesses.front()->getPseudoValue())
460 /// Return true if register is PIC base; i.e.g defined by X86::MOVPC32r.
461 static bool regIsPICBase(unsigned BaseReg, const MachineRegisterInfo &MRI) {
462 // Don't waste compile time scanning use-def chains of physregs.
463 if (!TargetRegisterInfo::isVirtualRegister(BaseReg))
465 bool isPICBase = false;
466 for (MachineRegisterInfo::def_instr_iterator I = MRI.def_instr_begin(BaseReg),
467 E = MRI.def_instr_end(); I != E; ++I) {
468 MachineInstr *DefMI = &*I;
469 if (DefMI->getOpcode() != X86::MOVPC32r)
471 assert(!isPICBase && "More than one PIC base?");
477 bool X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr &MI,
478 AliasAnalysis *AA) const {
479 switch (MI.getOpcode()) {
482 case X86::MOV8rm_NOREX:
503 case X86::VMOVAPSYrm:
504 case X86::VMOVUPSYrm:
505 case X86::VMOVAPDYrm:
506 case X86::VMOVUPDYrm:
507 case X86::VMOVDQAYrm:
508 case X86::VMOVDQUYrm:
509 case X86::MMX_MOVD64rm:
510 case X86::MMX_MOVQ64rm:
514 case X86::VMOVAPDZ128rm:
515 case X86::VMOVAPDZ256rm:
516 case X86::VMOVAPDZrm:
517 case X86::VMOVAPSZ128rm:
518 case X86::VMOVAPSZ256rm:
519 case X86::VMOVAPSZ128rm_NOVLX:
520 case X86::VMOVAPSZ256rm_NOVLX:
521 case X86::VMOVAPSZrm:
522 case X86::VMOVDQA32Z128rm:
523 case X86::VMOVDQA32Z256rm:
524 case X86::VMOVDQA32Zrm:
525 case X86::VMOVDQA64Z128rm:
526 case X86::VMOVDQA64Z256rm:
527 case X86::VMOVDQA64Zrm:
528 case X86::VMOVDQU16Z128rm:
529 case X86::VMOVDQU16Z256rm:
530 case X86::VMOVDQU16Zrm:
531 case X86::VMOVDQU32Z128rm:
532 case X86::VMOVDQU32Z256rm:
533 case X86::VMOVDQU32Zrm:
534 case X86::VMOVDQU64Z128rm:
535 case X86::VMOVDQU64Z256rm:
536 case X86::VMOVDQU64Zrm:
537 case X86::VMOVDQU8Z128rm:
538 case X86::VMOVDQU8Z256rm:
539 case X86::VMOVDQU8Zrm:
540 case X86::VMOVUPDZ128rm:
541 case X86::VMOVUPDZ256rm:
542 case X86::VMOVUPDZrm:
543 case X86::VMOVUPSZ128rm:
544 case X86::VMOVUPSZ256rm:
545 case X86::VMOVUPSZ128rm_NOVLX:
546 case X86::VMOVUPSZ256rm_NOVLX:
547 case X86::VMOVUPSZrm: {
548 // Loads from constant pools are trivially rematerializable.
549 if (MI.getOperand(1 + X86::AddrBaseReg).isReg() &&
550 MI.getOperand(1 + X86::AddrScaleAmt).isImm() &&
551 MI.getOperand(1 + X86::AddrIndexReg).isReg() &&
552 MI.getOperand(1 + X86::AddrIndexReg).getReg() == 0 &&
553 MI.isDereferenceableInvariantLoad(AA)) {
554 unsigned BaseReg = MI.getOperand(1 + X86::AddrBaseReg).getReg();
555 if (BaseReg == 0 || BaseReg == X86::RIP)
557 // Allow re-materialization of PIC load.
558 if (!ReMatPICStubLoad && MI.getOperand(1 + X86::AddrDisp).isGlobal())
560 const MachineFunction &MF = *MI.getParent()->getParent();
561 const MachineRegisterInfo &MRI = MF.getRegInfo();
562 return regIsPICBase(BaseReg, MRI);
569 if (MI.getOperand(1 + X86::AddrScaleAmt).isImm() &&
570 MI.getOperand(1 + X86::AddrIndexReg).isReg() &&
571 MI.getOperand(1 + X86::AddrIndexReg).getReg() == 0 &&
572 !MI.getOperand(1 + X86::AddrDisp).isReg()) {
573 // lea fi#, lea GV, etc. are all rematerializable.
574 if (!MI.getOperand(1 + X86::AddrBaseReg).isReg())
576 unsigned BaseReg = MI.getOperand(1 + X86::AddrBaseReg).getReg();
579 // Allow re-materialization of lea PICBase + x.
580 const MachineFunction &MF = *MI.getParent()->getParent();
581 const MachineRegisterInfo &MRI = MF.getRegInfo();
582 return regIsPICBase(BaseReg, MRI);
588 // All other instructions marked M_REMATERIALIZABLE are always trivially
593 bool X86InstrInfo::isSafeToClobberEFLAGS(MachineBasicBlock &MBB,
594 MachineBasicBlock::iterator I) const {
595 MachineBasicBlock::iterator E = MBB.end();
597 // For compile time consideration, if we are not able to determine the
598 // safety after visiting 4 instructions in each direction, we will assume
600 MachineBasicBlock::iterator Iter = I;
601 for (unsigned i = 0; Iter != E && i < 4; ++i) {
602 bool SeenDef = false;
603 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
604 MachineOperand &MO = Iter->getOperand(j);
605 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
609 if (MO.getReg() == X86::EFLAGS) {
617 // This instruction defines EFLAGS, no need to look any further.
620 // Skip over debug instructions.
621 while (Iter != E && Iter->isDebugInstr())
625 // It is safe to clobber EFLAGS at the end of a block of no successor has it
628 for (MachineBasicBlock *S : MBB.successors())
629 if (S->isLiveIn(X86::EFLAGS))
634 MachineBasicBlock::iterator B = MBB.begin();
636 for (unsigned i = 0; i < 4; ++i) {
637 // If we make it to the beginning of the block, it's safe to clobber
638 // EFLAGS iff EFLAGS is not live-in.
640 return !MBB.isLiveIn(X86::EFLAGS);
643 // Skip over debug instructions.
644 while (Iter != B && Iter->isDebugInstr())
647 bool SawKill = false;
648 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
649 MachineOperand &MO = Iter->getOperand(j);
650 // A register mask may clobber EFLAGS, but we should still look for a
652 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
654 if (MO.isReg() && MO.getReg() == X86::EFLAGS) {
655 if (MO.isDef()) return MO.isDead();
656 if (MO.isKill()) SawKill = true;
661 // This instruction kills EFLAGS and doesn't redefine it, so
662 // there's no need to look further.
666 // Conservative answer.
670 void X86InstrInfo::reMaterialize(MachineBasicBlock &MBB,
671 MachineBasicBlock::iterator I,
672 unsigned DestReg, unsigned SubIdx,
673 const MachineInstr &Orig,
674 const TargetRegisterInfo &TRI) const {
675 bool ClobbersEFLAGS = false;
676 for (const MachineOperand &MO : Orig.operands()) {
677 if (MO.isReg() && MO.isDef() && MO.getReg() == X86::EFLAGS) {
678 ClobbersEFLAGS = true;
683 if (ClobbersEFLAGS && !isSafeToClobberEFLAGS(MBB, I)) {
684 // The instruction clobbers EFLAGS. Re-materialize as MOV32ri to avoid side
687 switch (Orig.getOpcode()) {
688 case X86::MOV32r0: Value = 0; break;
689 case X86::MOV32r1: Value = 1; break;
690 case X86::MOV32r_1: Value = -1; break;
692 llvm_unreachable("Unexpected instruction!");
695 const DebugLoc &DL = Orig.getDebugLoc();
696 BuildMI(MBB, I, DL, get(X86::MOV32ri))
697 .add(Orig.getOperand(0))
700 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(&Orig);
704 MachineInstr &NewMI = *std::prev(I);
705 NewMI.substituteRegister(Orig.getOperand(0).getReg(), DestReg, SubIdx, TRI);
708 /// True if MI has a condition code def, e.g. EFLAGS, that is not marked dead.
709 bool X86InstrInfo::hasLiveCondCodeDef(MachineInstr &MI) const {
710 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
711 MachineOperand &MO = MI.getOperand(i);
712 if (MO.isReg() && MO.isDef() &&
713 MO.getReg() == X86::EFLAGS && !MO.isDead()) {
720 /// Check whether the shift count for a machine operand is non-zero.
721 inline static unsigned getTruncatedShiftCount(MachineInstr &MI,
722 unsigned ShiftAmtOperandIdx) {
723 // The shift count is six bits with the REX.W prefix and five bits without.
724 unsigned ShiftCountMask = (MI.getDesc().TSFlags & X86II::REX_W) ? 63 : 31;
725 unsigned Imm = MI.getOperand(ShiftAmtOperandIdx).getImm();
726 return Imm & ShiftCountMask;
729 /// Check whether the given shift count is appropriate
730 /// can be represented by a LEA instruction.
731 inline static bool isTruncatedShiftCountForLEA(unsigned ShAmt) {
732 // Left shift instructions can be transformed into load-effective-address
733 // instructions if we can encode them appropriately.
734 // A LEA instruction utilizes a SIB byte to encode its scale factor.
735 // The SIB.scale field is two bits wide which means that we can encode any
736 // shift amount less than 4.
737 return ShAmt < 4 && ShAmt > 0;
740 bool X86InstrInfo::classifyLEAReg(MachineInstr &MI, const MachineOperand &Src,
741 unsigned Opc, bool AllowSP, unsigned &NewSrc,
742 bool &isKill, bool &isUndef,
743 MachineOperand &ImplicitOp,
744 LiveVariables *LV) const {
745 MachineFunction &MF = *MI.getParent()->getParent();
746 const TargetRegisterClass *RC;
748 RC = Opc != X86::LEA32r ? &X86::GR64RegClass : &X86::GR32RegClass;
750 RC = Opc != X86::LEA32r ?
751 &X86::GR64_NOSPRegClass : &X86::GR32_NOSPRegClass;
753 unsigned SrcReg = Src.getReg();
755 // For both LEA64 and LEA32 the register already has essentially the right
756 // type (32-bit or 64-bit) we may just need to forbid SP.
757 if (Opc != X86::LEA64_32r) {
759 isKill = Src.isKill();
760 isUndef = Src.isUndef();
762 if (TargetRegisterInfo::isVirtualRegister(NewSrc) &&
763 !MF.getRegInfo().constrainRegClass(NewSrc, RC))
769 // This is for an LEA64_32r and incoming registers are 32-bit. One way or
770 // another we need to add 64-bit registers to the final MI.
771 if (TargetRegisterInfo::isPhysicalRegister(SrcReg)) {
773 ImplicitOp.setImplicit();
775 NewSrc = getX86SubSuperRegister(Src.getReg(), 64);
776 isKill = Src.isKill();
777 isUndef = Src.isUndef();
779 // Virtual register of the wrong class, we have to create a temporary 64-bit
780 // vreg to feed into the LEA.
781 NewSrc = MF.getRegInfo().createVirtualRegister(RC);
783 BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), get(TargetOpcode::COPY))
784 .addReg(NewSrc, RegState::Define | RegState::Undef, X86::sub_32bit)
787 // Which is obviously going to be dead after we're done with it.
792 LV->replaceKillInstruction(SrcReg, MI, *Copy);
795 // We've set all the parameters without issue.
799 /// Helper for convertToThreeAddress when 16-bit LEA is disabled, use 32-bit
800 /// LEA to form 3-address code by promoting to a 32-bit superregister and then
801 /// truncating back down to a 16-bit subregister.
802 MachineInstr *X86InstrInfo::convertToThreeAddressWithLEA(
803 unsigned MIOpc, MachineFunction::iterator &MFI, MachineInstr &MI,
804 LiveVariables *LV) const {
805 MachineBasicBlock::iterator MBBI = MI.getIterator();
806 unsigned Dest = MI.getOperand(0).getReg();
807 unsigned Src = MI.getOperand(1).getReg();
808 bool isDead = MI.getOperand(0).isDead();
809 bool isKill = MI.getOperand(1).isKill();
811 MachineRegisterInfo &RegInfo = MFI->getParent()->getRegInfo();
812 unsigned leaOutReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
813 unsigned Opc, leaInReg;
814 if (Subtarget.is64Bit()) {
815 Opc = X86::LEA64_32r;
816 leaInReg = RegInfo.createVirtualRegister(&X86::GR64_NOSPRegClass);
819 leaInReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
822 // Build and insert into an implicit UNDEF value. This is OK because
823 // well be shifting and then extracting the lower 16-bits.
824 // This has the potential to cause partial register stall. e.g.
825 // movw (%rbp,%rcx,2), %dx
826 // leal -65(%rdx), %esi
827 // But testing has shown this *does* help performance in 64-bit mode (at
828 // least on modern x86 machines).
829 BuildMI(*MFI, MBBI, MI.getDebugLoc(), get(X86::IMPLICIT_DEF), leaInReg);
830 MachineInstr *InsMI =
831 BuildMI(*MFI, MBBI, MI.getDebugLoc(), get(TargetOpcode::COPY))
832 .addReg(leaInReg, RegState::Define, X86::sub_16bit)
833 .addReg(Src, getKillRegState(isKill));
835 MachineInstrBuilder MIB =
836 BuildMI(*MFI, MBBI, MI.getDebugLoc(), get(Opc), leaOutReg);
838 default: llvm_unreachable("Unreachable!");
840 unsigned ShAmt = MI.getOperand(2).getImm();
841 MIB.addReg(0).addImm(1ULL << ShAmt)
842 .addReg(leaInReg, RegState::Kill).addImm(0).addReg(0);
846 addRegOffset(MIB, leaInReg, true, 1);
849 addRegOffset(MIB, leaInReg, true, -1);
853 case X86::ADD16ri_DB:
854 case X86::ADD16ri8_DB:
855 addRegOffset(MIB, leaInReg, true, MI.getOperand(2).getImm());
858 case X86::ADD16rr_DB: {
859 unsigned Src2 = MI.getOperand(2).getReg();
860 bool isKill2 = MI.getOperand(2).isKill();
861 unsigned leaInReg2 = 0;
862 MachineInstr *InsMI2 = nullptr;
864 // ADD16rr killed %reg1028, %reg1028
865 // just a single insert_subreg.
866 addRegReg(MIB, leaInReg, true, leaInReg, false);
868 if (Subtarget.is64Bit())
869 leaInReg2 = RegInfo.createVirtualRegister(&X86::GR64_NOSPRegClass);
871 leaInReg2 = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
872 // Build and insert into an implicit UNDEF value. This is OK because
873 // well be shifting and then extracting the lower 16-bits.
874 BuildMI(*MFI, &*MIB, MI.getDebugLoc(), get(X86::IMPLICIT_DEF), leaInReg2);
875 InsMI2 = BuildMI(*MFI, &*MIB, MI.getDebugLoc(), get(TargetOpcode::COPY))
876 .addReg(leaInReg2, RegState::Define, X86::sub_16bit)
877 .addReg(Src2, getKillRegState(isKill2));
878 addRegReg(MIB, leaInReg, true, leaInReg2, true);
880 if (LV && isKill2 && InsMI2)
881 LV->replaceKillInstruction(Src2, MI, *InsMI2);
886 MachineInstr *NewMI = MIB;
887 MachineInstr *ExtMI =
888 BuildMI(*MFI, MBBI, MI.getDebugLoc(), get(TargetOpcode::COPY))
889 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
890 .addReg(leaOutReg, RegState::Kill, X86::sub_16bit);
893 // Update live variables
894 LV->getVarInfo(leaInReg).Kills.push_back(NewMI);
895 LV->getVarInfo(leaOutReg).Kills.push_back(ExtMI);
897 LV->replaceKillInstruction(Src, MI, *InsMI);
899 LV->replaceKillInstruction(Dest, MI, *ExtMI);
905 /// This method must be implemented by targets that
906 /// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
907 /// may be able to convert a two-address instruction into a true
908 /// three-address instruction on demand. This allows the X86 target (for
909 /// example) to convert ADD and SHL instructions into LEA instructions if they
910 /// would require register copies due to two-addressness.
912 /// This method returns a null pointer if the transformation cannot be
913 /// performed, otherwise it returns the new instruction.
916 X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
917 MachineInstr &MI, LiveVariables *LV) const {
918 // The following opcodes also sets the condition code register(s). Only
919 // convert them to equivalent lea if the condition code register def's
921 if (hasLiveCondCodeDef(MI))
924 MachineFunction &MF = *MI.getParent()->getParent();
925 // All instructions input are two-addr instructions. Get the known operands.
926 const MachineOperand &Dest = MI.getOperand(0);
927 const MachineOperand &Src = MI.getOperand(1);
929 MachineInstr *NewMI = nullptr;
930 // FIXME: 16-bit LEA's are really slow on Athlons, but not bad on P4's. When
931 // we have better subtarget support, enable the 16-bit LEA generation here.
932 // 16-bit LEA is also slow on Core2.
933 bool DisableLEA16 = true;
934 bool is64Bit = Subtarget.is64Bit();
936 unsigned MIOpc = MI.getOpcode();
938 default: return nullptr;
940 assert(MI.getNumOperands() >= 3 && "Unknown shift instruction!");
941 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
942 if (!isTruncatedShiftCountForLEA(ShAmt)) return nullptr;
944 // LEA can't handle RSP.
945 if (TargetRegisterInfo::isVirtualRegister(Src.getReg()) &&
946 !MF.getRegInfo().constrainRegClass(Src.getReg(),
947 &X86::GR64_NOSPRegClass))
950 NewMI = BuildMI(MF, MI.getDebugLoc(), get(X86::LEA64r))
953 .addImm(1ULL << ShAmt)
960 assert(MI.getNumOperands() >= 3 && "Unknown shift instruction!");
961 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
962 if (!isTruncatedShiftCountForLEA(ShAmt)) return nullptr;
964 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
966 // LEA can't handle ESP.
967 bool isKill, isUndef;
969 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
970 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false,
971 SrcReg, isKill, isUndef, ImplicitOp, LV))
974 MachineInstrBuilder MIB =
975 BuildMI(MF, MI.getDebugLoc(), get(Opc))
978 .addImm(1ULL << ShAmt)
979 .addReg(SrcReg, getKillRegState(isKill) | getUndefRegState(isUndef))
982 if (ImplicitOp.getReg() != 0)
989 assert(MI.getNumOperands() >= 3 && "Unknown shift instruction!");
990 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
991 if (!isTruncatedShiftCountForLEA(ShAmt)) return nullptr;
994 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MI, LV)
996 NewMI = BuildMI(MF, MI.getDebugLoc(), get(X86::LEA16r))
999 .addImm(1ULL << ShAmt)
1007 assert(MI.getNumOperands() >= 2 && "Unknown inc instruction!");
1008 unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r
1009 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
1010 bool isKill, isUndef;
1012 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
1013 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false,
1014 SrcReg, isKill, isUndef, ImplicitOp, LV))
1017 MachineInstrBuilder MIB =
1018 BuildMI(MF, MI.getDebugLoc(), get(Opc))
1021 getKillRegState(isKill) | getUndefRegState(isUndef));
1022 if (ImplicitOp.getReg() != 0)
1023 MIB.add(ImplicitOp);
1025 NewMI = addOffset(MIB, 1);
1030 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MI, LV)
1032 assert(MI.getNumOperands() >= 2 && "Unknown inc instruction!");
1034 BuildMI(MF, MI.getDebugLoc(), get(X86::LEA16r)).add(Dest).add(Src), 1);
1038 assert(MI.getNumOperands() >= 2 && "Unknown dec instruction!");
1039 unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r
1040 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
1042 bool isKill, isUndef;
1044 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
1045 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false,
1046 SrcReg, isKill, isUndef, ImplicitOp, LV))
1049 MachineInstrBuilder MIB = BuildMI(MF, MI.getDebugLoc(), get(Opc))
1051 .addReg(SrcReg, getUndefRegState(isUndef) |
1052 getKillRegState(isKill));
1053 if (ImplicitOp.getReg() != 0)
1054 MIB.add(ImplicitOp);
1056 NewMI = addOffset(MIB, -1);
1062 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MI, LV)
1064 assert(MI.getNumOperands() >= 2 && "Unknown dec instruction!");
1066 BuildMI(MF, MI.getDebugLoc(), get(X86::LEA16r)).add(Dest).add(Src), -1);
1069 case X86::ADD64rr_DB:
1071 case X86::ADD32rr_DB: {
1072 assert(MI.getNumOperands() >= 3 && "Unknown add instruction!");
1074 if (MIOpc == X86::ADD64rr || MIOpc == X86::ADD64rr_DB)
1077 Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
1079 bool isKill, isUndef;
1081 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
1082 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ true,
1083 SrcReg, isKill, isUndef, ImplicitOp, LV))
1086 const MachineOperand &Src2 = MI.getOperand(2);
1087 bool isKill2, isUndef2;
1089 MachineOperand ImplicitOp2 = MachineOperand::CreateReg(0, false);
1090 if (!classifyLEAReg(MI, Src2, Opc, /*AllowSP=*/ false,
1091 SrcReg2, isKill2, isUndef2, ImplicitOp2, LV))
1094 MachineInstrBuilder MIB = BuildMI(MF, MI.getDebugLoc(), get(Opc)).add(Dest);
1095 if (ImplicitOp.getReg() != 0)
1096 MIB.add(ImplicitOp);
1097 if (ImplicitOp2.getReg() != 0)
1098 MIB.add(ImplicitOp2);
1100 NewMI = addRegReg(MIB, SrcReg, isKill, SrcReg2, isKill2);
1102 // Preserve undefness of the operands.
1103 NewMI->getOperand(1).setIsUndef(isUndef);
1104 NewMI->getOperand(3).setIsUndef(isUndef2);
1106 if (LV && Src2.isKill())
1107 LV->replaceKillInstruction(SrcReg2, MI, *NewMI);
1111 case X86::ADD16rr_DB: {
1113 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MI, LV)
1115 assert(MI.getNumOperands() >= 3 && "Unknown add instruction!");
1116 unsigned Src2 = MI.getOperand(2).getReg();
1117 bool isKill2 = MI.getOperand(2).isKill();
1118 NewMI = addRegReg(BuildMI(MF, MI.getDebugLoc(), get(X86::LEA16r)).add(Dest),
1119 Src.getReg(), Src.isKill(), Src2, isKill2);
1121 // Preserve undefness of the operands.
1122 bool isUndef = MI.getOperand(1).isUndef();
1123 bool isUndef2 = MI.getOperand(2).isUndef();
1124 NewMI->getOperand(1).setIsUndef(isUndef);
1125 NewMI->getOperand(3).setIsUndef(isUndef2);
1128 LV->replaceKillInstruction(Src2, MI, *NewMI);
1131 case X86::ADD64ri32:
1133 case X86::ADD64ri32_DB:
1134 case X86::ADD64ri8_DB:
1135 assert(MI.getNumOperands() >= 3 && "Unknown add instruction!");
1137 BuildMI(MF, MI.getDebugLoc(), get(X86::LEA64r)).add(Dest).add(Src),
1142 case X86::ADD32ri_DB:
1143 case X86::ADD32ri8_DB: {
1144 assert(MI.getNumOperands() >= 3 && "Unknown add instruction!");
1145 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
1147 bool isKill, isUndef;
1149 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
1150 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ true,
1151 SrcReg, isKill, isUndef, ImplicitOp, LV))
1154 MachineInstrBuilder MIB = BuildMI(MF, MI.getDebugLoc(), get(Opc))
1156 .addReg(SrcReg, getUndefRegState(isUndef) |
1157 getKillRegState(isKill));
1158 if (ImplicitOp.getReg() != 0)
1159 MIB.add(ImplicitOp);
1161 NewMI = addOffset(MIB, MI.getOperand(2));
1166 case X86::ADD16ri_DB:
1167 case X86::ADD16ri8_DB:
1169 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MI, LV)
1171 assert(MI.getNumOperands() >= 3 && "Unknown add instruction!");
1173 BuildMI(MF, MI.getDebugLoc(), get(X86::LEA16r)).add(Dest).add(Src),
1177 case X86::VMOVDQU8Z128rmk:
1178 case X86::VMOVDQU8Z256rmk:
1179 case X86::VMOVDQU8Zrmk:
1180 case X86::VMOVDQU16Z128rmk:
1181 case X86::VMOVDQU16Z256rmk:
1182 case X86::VMOVDQU16Zrmk:
1183 case X86::VMOVDQU32Z128rmk: case X86::VMOVDQA32Z128rmk:
1184 case X86::VMOVDQU32Z256rmk: case X86::VMOVDQA32Z256rmk:
1185 case X86::VMOVDQU32Zrmk: case X86::VMOVDQA32Zrmk:
1186 case X86::VMOVDQU64Z128rmk: case X86::VMOVDQA64Z128rmk:
1187 case X86::VMOVDQU64Z256rmk: case X86::VMOVDQA64Z256rmk:
1188 case X86::VMOVDQU64Zrmk: case X86::VMOVDQA64Zrmk:
1189 case X86::VMOVUPDZ128rmk: case X86::VMOVAPDZ128rmk:
1190 case X86::VMOVUPDZ256rmk: case X86::VMOVAPDZ256rmk:
1191 case X86::VMOVUPDZrmk: case X86::VMOVAPDZrmk:
1192 case X86::VMOVUPSZ128rmk: case X86::VMOVAPSZ128rmk:
1193 case X86::VMOVUPSZ256rmk: case X86::VMOVAPSZ256rmk:
1194 case X86::VMOVUPSZrmk: case X86::VMOVAPSZrmk: {
1197 default: llvm_unreachable("Unreachable!");
1198 case X86::VMOVDQU8Z128rmk: Opc = X86::VPBLENDMBZ128rmk; break;
1199 case X86::VMOVDQU8Z256rmk: Opc = X86::VPBLENDMBZ256rmk; break;
1200 case X86::VMOVDQU8Zrmk: Opc = X86::VPBLENDMBZrmk; break;
1201 case X86::VMOVDQU16Z128rmk: Opc = X86::VPBLENDMWZ128rmk; break;
1202 case X86::VMOVDQU16Z256rmk: Opc = X86::VPBLENDMWZ256rmk; break;
1203 case X86::VMOVDQU16Zrmk: Opc = X86::VPBLENDMWZrmk; break;
1204 case X86::VMOVDQU32Z128rmk: Opc = X86::VPBLENDMDZ128rmk; break;
1205 case X86::VMOVDQU32Z256rmk: Opc = X86::VPBLENDMDZ256rmk; break;
1206 case X86::VMOVDQU32Zrmk: Opc = X86::VPBLENDMDZrmk; break;
1207 case X86::VMOVDQU64Z128rmk: Opc = X86::VPBLENDMQZ128rmk; break;
1208 case X86::VMOVDQU64Z256rmk: Opc = X86::VPBLENDMQZ256rmk; break;
1209 case X86::VMOVDQU64Zrmk: Opc = X86::VPBLENDMQZrmk; break;
1210 case X86::VMOVUPDZ128rmk: Opc = X86::VBLENDMPDZ128rmk; break;
1211 case X86::VMOVUPDZ256rmk: Opc = X86::VBLENDMPDZ256rmk; break;
1212 case X86::VMOVUPDZrmk: Opc = X86::VBLENDMPDZrmk; break;
1213 case X86::VMOVUPSZ128rmk: Opc = X86::VBLENDMPSZ128rmk; break;
1214 case X86::VMOVUPSZ256rmk: Opc = X86::VBLENDMPSZ256rmk; break;
1215 case X86::VMOVUPSZrmk: Opc = X86::VBLENDMPSZrmk; break;
1216 case X86::VMOVDQA32Z128rmk: Opc = X86::VPBLENDMDZ128rmk; break;
1217 case X86::VMOVDQA32Z256rmk: Opc = X86::VPBLENDMDZ256rmk; break;
1218 case X86::VMOVDQA32Zrmk: Opc = X86::VPBLENDMDZrmk; break;
1219 case X86::VMOVDQA64Z128rmk: Opc = X86::VPBLENDMQZ128rmk; break;
1220 case X86::VMOVDQA64Z256rmk: Opc = X86::VPBLENDMQZ256rmk; break;
1221 case X86::VMOVDQA64Zrmk: Opc = X86::VPBLENDMQZrmk; break;
1222 case X86::VMOVAPDZ128rmk: Opc = X86::VBLENDMPDZ128rmk; break;
1223 case X86::VMOVAPDZ256rmk: Opc = X86::VBLENDMPDZ256rmk; break;
1224 case X86::VMOVAPDZrmk: Opc = X86::VBLENDMPDZrmk; break;
1225 case X86::VMOVAPSZ128rmk: Opc = X86::VBLENDMPSZ128rmk; break;
1226 case X86::VMOVAPSZ256rmk: Opc = X86::VBLENDMPSZ256rmk; break;
1227 case X86::VMOVAPSZrmk: Opc = X86::VBLENDMPSZrmk; break;
1230 NewMI = BuildMI(MF, MI.getDebugLoc(), get(Opc))
1232 .add(MI.getOperand(2))
1234 .add(MI.getOperand(3))
1235 .add(MI.getOperand(4))
1236 .add(MI.getOperand(5))
1237 .add(MI.getOperand(6))
1238 .add(MI.getOperand(7));
1241 case X86::VMOVDQU8Z128rrk:
1242 case X86::VMOVDQU8Z256rrk:
1243 case X86::VMOVDQU8Zrrk:
1244 case X86::VMOVDQU16Z128rrk:
1245 case X86::VMOVDQU16Z256rrk:
1246 case X86::VMOVDQU16Zrrk:
1247 case X86::VMOVDQU32Z128rrk: case X86::VMOVDQA32Z128rrk:
1248 case X86::VMOVDQU32Z256rrk: case X86::VMOVDQA32Z256rrk:
1249 case X86::VMOVDQU32Zrrk: case X86::VMOVDQA32Zrrk:
1250 case X86::VMOVDQU64Z128rrk: case X86::VMOVDQA64Z128rrk:
1251 case X86::VMOVDQU64Z256rrk: case X86::VMOVDQA64Z256rrk:
1252 case X86::VMOVDQU64Zrrk: case X86::VMOVDQA64Zrrk:
1253 case X86::VMOVUPDZ128rrk: case X86::VMOVAPDZ128rrk:
1254 case X86::VMOVUPDZ256rrk: case X86::VMOVAPDZ256rrk:
1255 case X86::VMOVUPDZrrk: case X86::VMOVAPDZrrk:
1256 case X86::VMOVUPSZ128rrk: case X86::VMOVAPSZ128rrk:
1257 case X86::VMOVUPSZ256rrk: case X86::VMOVAPSZ256rrk:
1258 case X86::VMOVUPSZrrk: case X86::VMOVAPSZrrk: {
1261 default: llvm_unreachable("Unreachable!");
1262 case X86::VMOVDQU8Z128rrk: Opc = X86::VPBLENDMBZ128rrk; break;
1263 case X86::VMOVDQU8Z256rrk: Opc = X86::VPBLENDMBZ256rrk; break;
1264 case X86::VMOVDQU8Zrrk: Opc = X86::VPBLENDMBZrrk; break;
1265 case X86::VMOVDQU16Z128rrk: Opc = X86::VPBLENDMWZ128rrk; break;
1266 case X86::VMOVDQU16Z256rrk: Opc = X86::VPBLENDMWZ256rrk; break;
1267 case X86::VMOVDQU16Zrrk: Opc = X86::VPBLENDMWZrrk; break;
1268 case X86::VMOVDQU32Z128rrk: Opc = X86::VPBLENDMDZ128rrk; break;
1269 case X86::VMOVDQU32Z256rrk: Opc = X86::VPBLENDMDZ256rrk; break;
1270 case X86::VMOVDQU32Zrrk: Opc = X86::VPBLENDMDZrrk; break;
1271 case X86::VMOVDQU64Z128rrk: Opc = X86::VPBLENDMQZ128rrk; break;
1272 case X86::VMOVDQU64Z256rrk: Opc = X86::VPBLENDMQZ256rrk; break;
1273 case X86::VMOVDQU64Zrrk: Opc = X86::VPBLENDMQZrrk; break;
1274 case X86::VMOVUPDZ128rrk: Opc = X86::VBLENDMPDZ128rrk; break;
1275 case X86::VMOVUPDZ256rrk: Opc = X86::VBLENDMPDZ256rrk; break;
1276 case X86::VMOVUPDZrrk: Opc = X86::VBLENDMPDZrrk; break;
1277 case X86::VMOVUPSZ128rrk: Opc = X86::VBLENDMPSZ128rrk; break;
1278 case X86::VMOVUPSZ256rrk: Opc = X86::VBLENDMPSZ256rrk; break;
1279 case X86::VMOVUPSZrrk: Opc = X86::VBLENDMPSZrrk; break;
1280 case X86::VMOVDQA32Z128rrk: Opc = X86::VPBLENDMDZ128rrk; break;
1281 case X86::VMOVDQA32Z256rrk: Opc = X86::VPBLENDMDZ256rrk; break;
1282 case X86::VMOVDQA32Zrrk: Opc = X86::VPBLENDMDZrrk; break;
1283 case X86::VMOVDQA64Z128rrk: Opc = X86::VPBLENDMQZ128rrk; break;
1284 case X86::VMOVDQA64Z256rrk: Opc = X86::VPBLENDMQZ256rrk; break;
1285 case X86::VMOVDQA64Zrrk: Opc = X86::VPBLENDMQZrrk; break;
1286 case X86::VMOVAPDZ128rrk: Opc = X86::VBLENDMPDZ128rrk; break;
1287 case X86::VMOVAPDZ256rrk: Opc = X86::VBLENDMPDZ256rrk; break;
1288 case X86::VMOVAPDZrrk: Opc = X86::VBLENDMPDZrrk; break;
1289 case X86::VMOVAPSZ128rrk: Opc = X86::VBLENDMPSZ128rrk; break;
1290 case X86::VMOVAPSZ256rrk: Opc = X86::VBLENDMPSZ256rrk; break;
1291 case X86::VMOVAPSZrrk: Opc = X86::VBLENDMPSZrrk; break;
1294 NewMI = BuildMI(MF, MI.getDebugLoc(), get(Opc))
1296 .add(MI.getOperand(2))
1298 .add(MI.getOperand(3));
1303 if (!NewMI) return nullptr;
1305 if (LV) { // Update live variables
1307 LV->replaceKillInstruction(Src.getReg(), MI, *NewMI);
1309 LV->replaceKillInstruction(Dest.getReg(), MI, *NewMI);
1312 MFI->insert(MI.getIterator(), NewMI); // Insert the new inst
1316 /// This determines which of three possible cases of a three source commute
1317 /// the source indexes correspond to taking into account any mask operands.
1318 /// All prevents commuting a passthru operand. Returns -1 if the commute isn't
1320 /// Case 0 - Possible to commute the first and second operands.
1321 /// Case 1 - Possible to commute the first and third operands.
1322 /// Case 2 - Possible to commute the second and third operands.
1323 static unsigned getThreeSrcCommuteCase(uint64_t TSFlags, unsigned SrcOpIdx1,
1324 unsigned SrcOpIdx2) {
1325 // Put the lowest index to SrcOpIdx1 to simplify the checks below.
1326 if (SrcOpIdx1 > SrcOpIdx2)
1327 std::swap(SrcOpIdx1, SrcOpIdx2);
1329 unsigned Op1 = 1, Op2 = 2, Op3 = 3;
1330 if (X86II::isKMasked(TSFlags)) {
1335 if (SrcOpIdx1 == Op1 && SrcOpIdx2 == Op2)
1337 if (SrcOpIdx1 == Op1 && SrcOpIdx2 == Op3)
1339 if (SrcOpIdx1 == Op2 && SrcOpIdx2 == Op3)
1341 llvm_unreachable("Unknown three src commute case.");
1344 unsigned X86InstrInfo::getFMA3OpcodeToCommuteOperands(
1345 const MachineInstr &MI, unsigned SrcOpIdx1, unsigned SrcOpIdx2,
1346 const X86InstrFMA3Group &FMA3Group) const {
1348 unsigned Opc = MI.getOpcode();
1350 // TODO: Commuting the 1st operand of FMA*_Int requires some additional
1351 // analysis. The commute optimization is legal only if all users of FMA*_Int
1352 // use only the lowest element of the FMA*_Int instruction. Such analysis are
1353 // not implemented yet. So, just return 0 in that case.
1354 // When such analysis are available this place will be the right place for
1356 assert(!(FMA3Group.isIntrinsic() && (SrcOpIdx1 == 1 || SrcOpIdx2 == 1)) &&
1357 "Intrinsic instructions can't commute operand 1");
1359 // Determine which case this commute is or if it can't be done.
1360 unsigned Case = getThreeSrcCommuteCase(MI.getDesc().TSFlags, SrcOpIdx1,
1362 assert(Case < 3 && "Unexpected case number!");
1364 // Define the FMA forms mapping array that helps to map input FMA form
1365 // to output FMA form to preserve the operation semantics after
1366 // commuting the operands.
1367 const unsigned Form132Index = 0;
1368 const unsigned Form213Index = 1;
1369 const unsigned Form231Index = 2;
1370 static const unsigned FormMapping[][3] = {
1371 // 0: SrcOpIdx1 == 1 && SrcOpIdx2 == 2;
1372 // FMA132 A, C, b; ==> FMA231 C, A, b;
1373 // FMA213 B, A, c; ==> FMA213 A, B, c;
1374 // FMA231 C, A, b; ==> FMA132 A, C, b;
1375 { Form231Index, Form213Index, Form132Index },
1376 // 1: SrcOpIdx1 == 1 && SrcOpIdx2 == 3;
1377 // FMA132 A, c, B; ==> FMA132 B, c, A;
1378 // FMA213 B, a, C; ==> FMA231 C, a, B;
1379 // FMA231 C, a, B; ==> FMA213 B, a, C;
1380 { Form132Index, Form231Index, Form213Index },
1381 // 2: SrcOpIdx1 == 2 && SrcOpIdx2 == 3;
1382 // FMA132 a, C, B; ==> FMA213 a, B, C;
1383 // FMA213 b, A, C; ==> FMA132 b, C, A;
1384 // FMA231 c, A, B; ==> FMA231 c, B, A;
1385 { Form213Index, Form132Index, Form231Index }
1388 unsigned FMAForms[3];
1389 FMAForms[0] = FMA3Group.get132Opcode();
1390 FMAForms[1] = FMA3Group.get213Opcode();
1391 FMAForms[2] = FMA3Group.get231Opcode();
1393 for (FormIndex = 0; FormIndex < 3; FormIndex++)
1394 if (Opc == FMAForms[FormIndex])
1397 // Everything is ready, just adjust the FMA opcode and return it.
1398 FormIndex = FormMapping[Case][FormIndex];
1399 return FMAForms[FormIndex];
1402 static void commuteVPTERNLOG(MachineInstr &MI, unsigned SrcOpIdx1,
1403 unsigned SrcOpIdx2) {
1404 // Determine which case this commute is or if it can't be done.
1405 unsigned Case = getThreeSrcCommuteCase(MI.getDesc().TSFlags, SrcOpIdx1,
1407 assert(Case < 3 && "Unexpected case value!");
1409 // For each case we need to swap two pairs of bits in the final immediate.
1410 static const uint8_t SwapMasks[3][4] = {
1411 { 0x04, 0x10, 0x08, 0x20 }, // Swap bits 2/4 and 3/5.
1412 { 0x02, 0x10, 0x08, 0x40 }, // Swap bits 1/4 and 3/6.
1413 { 0x02, 0x04, 0x20, 0x40 }, // Swap bits 1/2 and 5/6.
1416 uint8_t Imm = MI.getOperand(MI.getNumOperands()-1).getImm();
1417 // Clear out the bits we are swapping.
1418 uint8_t NewImm = Imm & ~(SwapMasks[Case][0] | SwapMasks[Case][1] |
1419 SwapMasks[Case][2] | SwapMasks[Case][3]);
1420 // If the immediate had a bit of the pair set, then set the opposite bit.
1421 if (Imm & SwapMasks[Case][0]) NewImm |= SwapMasks[Case][1];
1422 if (Imm & SwapMasks[Case][1]) NewImm |= SwapMasks[Case][0];
1423 if (Imm & SwapMasks[Case][2]) NewImm |= SwapMasks[Case][3];
1424 if (Imm & SwapMasks[Case][3]) NewImm |= SwapMasks[Case][2];
1425 MI.getOperand(MI.getNumOperands()-1).setImm(NewImm);
1428 // Returns true if this is a VPERMI2 or VPERMT2 instruction that can be
1430 static bool isCommutableVPERMV3Instruction(unsigned Opcode) {
1431 #define VPERM_CASES(Suffix) \
1432 case X86::VPERMI2##Suffix##128rr: case X86::VPERMT2##Suffix##128rr: \
1433 case X86::VPERMI2##Suffix##256rr: case X86::VPERMT2##Suffix##256rr: \
1434 case X86::VPERMI2##Suffix##rr: case X86::VPERMT2##Suffix##rr: \
1435 case X86::VPERMI2##Suffix##128rm: case X86::VPERMT2##Suffix##128rm: \
1436 case X86::VPERMI2##Suffix##256rm: case X86::VPERMT2##Suffix##256rm: \
1437 case X86::VPERMI2##Suffix##rm: case X86::VPERMT2##Suffix##rm: \
1438 case X86::VPERMI2##Suffix##128rrkz: case X86::VPERMT2##Suffix##128rrkz: \
1439 case X86::VPERMI2##Suffix##256rrkz: case X86::VPERMT2##Suffix##256rrkz: \
1440 case X86::VPERMI2##Suffix##rrkz: case X86::VPERMT2##Suffix##rrkz: \
1441 case X86::VPERMI2##Suffix##128rmkz: case X86::VPERMT2##Suffix##128rmkz: \
1442 case X86::VPERMI2##Suffix##256rmkz: case X86::VPERMT2##Suffix##256rmkz: \
1443 case X86::VPERMI2##Suffix##rmkz: case X86::VPERMT2##Suffix##rmkz:
1445 #define VPERM_CASES_BROADCAST(Suffix) \
1446 VPERM_CASES(Suffix) \
1447 case X86::VPERMI2##Suffix##128rmb: case X86::VPERMT2##Suffix##128rmb: \
1448 case X86::VPERMI2##Suffix##256rmb: case X86::VPERMT2##Suffix##256rmb: \
1449 case X86::VPERMI2##Suffix##rmb: case X86::VPERMT2##Suffix##rmb: \
1450 case X86::VPERMI2##Suffix##128rmbkz: case X86::VPERMT2##Suffix##128rmbkz: \
1451 case X86::VPERMI2##Suffix##256rmbkz: case X86::VPERMT2##Suffix##256rmbkz: \
1452 case X86::VPERMI2##Suffix##rmbkz: case X86::VPERMT2##Suffix##rmbkz:
1455 default: return false;
1457 VPERM_CASES_BROADCAST(D)
1458 VPERM_CASES_BROADCAST(PD)
1459 VPERM_CASES_BROADCAST(PS)
1460 VPERM_CASES_BROADCAST(Q)
1464 #undef VPERM_CASES_BROADCAST
1468 // Returns commuted opcode for VPERMI2 and VPERMT2 instructions by switching
1469 // from the I opcode to the T opcode and vice versa.
1470 static unsigned getCommutedVPERMV3Opcode(unsigned Opcode) {
1471 #define VPERM_CASES(Orig, New) \
1472 case X86::Orig##128rr: return X86::New##128rr; \
1473 case X86::Orig##128rrkz: return X86::New##128rrkz; \
1474 case X86::Orig##128rm: return X86::New##128rm; \
1475 case X86::Orig##128rmkz: return X86::New##128rmkz; \
1476 case X86::Orig##256rr: return X86::New##256rr; \
1477 case X86::Orig##256rrkz: return X86::New##256rrkz; \
1478 case X86::Orig##256rm: return X86::New##256rm; \
1479 case X86::Orig##256rmkz: return X86::New##256rmkz; \
1480 case X86::Orig##rr: return X86::New##rr; \
1481 case X86::Orig##rrkz: return X86::New##rrkz; \
1482 case X86::Orig##rm: return X86::New##rm; \
1483 case X86::Orig##rmkz: return X86::New##rmkz;
1485 #define VPERM_CASES_BROADCAST(Orig, New) \
1486 VPERM_CASES(Orig, New) \
1487 case X86::Orig##128rmb: return X86::New##128rmb; \
1488 case X86::Orig##128rmbkz: return X86::New##128rmbkz; \
1489 case X86::Orig##256rmb: return X86::New##256rmb; \
1490 case X86::Orig##256rmbkz: return X86::New##256rmbkz; \
1491 case X86::Orig##rmb: return X86::New##rmb; \
1492 case X86::Orig##rmbkz: return X86::New##rmbkz;
1495 VPERM_CASES(VPERMI2B, VPERMT2B)
1496 VPERM_CASES_BROADCAST(VPERMI2D, VPERMT2D)
1497 VPERM_CASES_BROADCAST(VPERMI2PD, VPERMT2PD)
1498 VPERM_CASES_BROADCAST(VPERMI2PS, VPERMT2PS)
1499 VPERM_CASES_BROADCAST(VPERMI2Q, VPERMT2Q)
1500 VPERM_CASES(VPERMI2W, VPERMT2W)
1501 VPERM_CASES(VPERMT2B, VPERMI2B)
1502 VPERM_CASES_BROADCAST(VPERMT2D, VPERMI2D)
1503 VPERM_CASES_BROADCAST(VPERMT2PD, VPERMI2PD)
1504 VPERM_CASES_BROADCAST(VPERMT2PS, VPERMI2PS)
1505 VPERM_CASES_BROADCAST(VPERMT2Q, VPERMI2Q)
1506 VPERM_CASES(VPERMT2W, VPERMI2W)
1509 llvm_unreachable("Unreachable!");
1510 #undef VPERM_CASES_BROADCAST
1514 MachineInstr *X86InstrInfo::commuteInstructionImpl(MachineInstr &MI, bool NewMI,
1516 unsigned OpIdx2) const {
1517 auto cloneIfNew = [NewMI](MachineInstr &MI) -> MachineInstr & {
1519 return *MI.getParent()->getParent()->CloneMachineInstr(&MI);
1523 switch (MI.getOpcode()) {
1524 case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I)
1525 case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I)
1526 case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I)
1527 case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I)
1528 case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I)
1529 case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I)
1532 switch (MI.getOpcode()) {
1533 default: llvm_unreachable("Unreachable!");
1534 case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break;
1535 case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break;
1536 case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break;
1537 case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break;
1538 case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break;
1539 case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break;
1541 unsigned Amt = MI.getOperand(3).getImm();
1542 auto &WorkingMI = cloneIfNew(MI);
1543 WorkingMI.setDesc(get(Opc));
1544 WorkingMI.getOperand(3).setImm(Size - Amt);
1545 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1549 case X86::PFSUBRrr: {
1550 // PFSUB x, y: x = x - y
1551 // PFSUBR x, y: x = y - x
1553 (X86::PFSUBRrr == MI.getOpcode() ? X86::PFSUBrr : X86::PFSUBRrr);
1554 auto &WorkingMI = cloneIfNew(MI);
1555 WorkingMI.setDesc(get(Opc));
1556 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1559 case X86::BLENDPDrri:
1560 case X86::BLENDPSrri:
1561 case X86::VBLENDPDrri:
1562 case X86::VBLENDPSrri:
1563 // If we're optimizing for size, try to use MOVSD/MOVSS.
1564 if (MI.getParent()->getParent()->getFunction().optForSize()) {
1566 switch (MI.getOpcode()) {
1567 default: llvm_unreachable("Unreachable!");
1568 case X86::BLENDPDrri: Opc = X86::MOVSDrr; Mask = 0x03; break;
1569 case X86::BLENDPSrri: Opc = X86::MOVSSrr; Mask = 0x0F; break;
1570 case X86::VBLENDPDrri: Opc = X86::VMOVSDrr; Mask = 0x03; break;
1571 case X86::VBLENDPSrri: Opc = X86::VMOVSSrr; Mask = 0x0F; break;
1573 if ((MI.getOperand(3).getImm() ^ Mask) == 1) {
1574 auto &WorkingMI = cloneIfNew(MI);
1575 WorkingMI.setDesc(get(Opc));
1576 WorkingMI.RemoveOperand(3);
1577 return TargetInstrInfo::commuteInstructionImpl(WorkingMI,
1583 case X86::PBLENDWrri:
1584 case X86::VBLENDPDYrri:
1585 case X86::VBLENDPSYrri:
1586 case X86::VPBLENDDrri:
1587 case X86::VPBLENDWrri:
1588 case X86::VPBLENDDYrri:
1589 case X86::VPBLENDWYrri:{
1591 switch (MI.getOpcode()) {
1592 default: llvm_unreachable("Unreachable!");
1593 case X86::BLENDPDrri: Mask = 0x03; break;
1594 case X86::BLENDPSrri: Mask = 0x0F; break;
1595 case X86::PBLENDWrri: Mask = 0xFF; break;
1596 case X86::VBLENDPDrri: Mask = 0x03; break;
1597 case X86::VBLENDPSrri: Mask = 0x0F; break;
1598 case X86::VBLENDPDYrri: Mask = 0x0F; break;
1599 case X86::VBLENDPSYrri: Mask = 0xFF; break;
1600 case X86::VPBLENDDrri: Mask = 0x0F; break;
1601 case X86::VPBLENDWrri: Mask = 0xFF; break;
1602 case X86::VPBLENDDYrri: Mask = 0xFF; break;
1603 case X86::VPBLENDWYrri: Mask = 0xFF; break;
1605 // Only the least significant bits of Imm are used.
1606 unsigned Imm = MI.getOperand(3).getImm() & Mask;
1607 auto &WorkingMI = cloneIfNew(MI);
1608 WorkingMI.getOperand(3).setImm(Mask ^ Imm);
1609 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1615 case X86::VMOVSSrr:{
1616 // On SSE41 or later we can commute a MOVSS/MOVSD to a BLENDPS/BLENDPD.
1617 assert(Subtarget.hasSSE41() && "Commuting MOVSD/MOVSS requires SSE41!");
1620 switch (MI.getOpcode()) {
1621 default: llvm_unreachable("Unreachable!");
1622 case X86::MOVSDrr: Opc = X86::BLENDPDrri; Mask = 0x02; break;
1623 case X86::MOVSSrr: Opc = X86::BLENDPSrri; Mask = 0x0E; break;
1624 case X86::VMOVSDrr: Opc = X86::VBLENDPDrri; Mask = 0x02; break;
1625 case X86::VMOVSSrr: Opc = X86::VBLENDPSrri; Mask = 0x0E; break;
1628 auto &WorkingMI = cloneIfNew(MI);
1629 WorkingMI.setDesc(get(Opc));
1630 WorkingMI.addOperand(MachineOperand::CreateImm(Mask));
1631 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1634 case X86::PCLMULQDQrr:
1635 case X86::VPCLMULQDQrr:
1636 case X86::VPCLMULQDQYrr:
1637 case X86::VPCLMULQDQZrr:
1638 case X86::VPCLMULQDQZ128rr:
1639 case X86::VPCLMULQDQZ256rr: {
1640 // SRC1 64bits = Imm[0] ? SRC1[127:64] : SRC1[63:0]
1641 // SRC2 64bits = Imm[4] ? SRC2[127:64] : SRC2[63:0]
1642 unsigned Imm = MI.getOperand(3).getImm();
1643 unsigned Src1Hi = Imm & 0x01;
1644 unsigned Src2Hi = Imm & 0x10;
1645 auto &WorkingMI = cloneIfNew(MI);
1646 WorkingMI.getOperand(3).setImm((Src1Hi << 4) | (Src2Hi >> 4));
1647 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1650 case X86::VPCMPBZ128rri: case X86::VPCMPUBZ128rri:
1651 case X86::VPCMPBZ256rri: case X86::VPCMPUBZ256rri:
1652 case X86::VPCMPBZrri: case X86::VPCMPUBZrri:
1653 case X86::VPCMPDZ128rri: case X86::VPCMPUDZ128rri:
1654 case X86::VPCMPDZ256rri: case X86::VPCMPUDZ256rri:
1655 case X86::VPCMPDZrri: case X86::VPCMPUDZrri:
1656 case X86::VPCMPQZ128rri: case X86::VPCMPUQZ128rri:
1657 case X86::VPCMPQZ256rri: case X86::VPCMPUQZ256rri:
1658 case X86::VPCMPQZrri: case X86::VPCMPUQZrri:
1659 case X86::VPCMPWZ128rri: case X86::VPCMPUWZ128rri:
1660 case X86::VPCMPWZ256rri: case X86::VPCMPUWZ256rri:
1661 case X86::VPCMPWZrri: case X86::VPCMPUWZrri:
1662 case X86::VPCMPBZ128rrik: case X86::VPCMPUBZ128rrik:
1663 case X86::VPCMPBZ256rrik: case X86::VPCMPUBZ256rrik:
1664 case X86::VPCMPBZrrik: case X86::VPCMPUBZrrik:
1665 case X86::VPCMPDZ128rrik: case X86::VPCMPUDZ128rrik:
1666 case X86::VPCMPDZ256rrik: case X86::VPCMPUDZ256rrik:
1667 case X86::VPCMPDZrrik: case X86::VPCMPUDZrrik:
1668 case X86::VPCMPQZ128rrik: case X86::VPCMPUQZ128rrik:
1669 case X86::VPCMPQZ256rrik: case X86::VPCMPUQZ256rrik:
1670 case X86::VPCMPQZrrik: case X86::VPCMPUQZrrik:
1671 case X86::VPCMPWZ128rrik: case X86::VPCMPUWZ128rrik:
1672 case X86::VPCMPWZ256rrik: case X86::VPCMPUWZ256rrik:
1673 case X86::VPCMPWZrrik: case X86::VPCMPUWZrrik: {
1674 // Flip comparison mode immediate (if necessary).
1675 unsigned Imm = MI.getOperand(MI.getNumOperands() - 1).getImm() & 0x7;
1676 Imm = X86::getSwappedVPCMPImm(Imm);
1677 auto &WorkingMI = cloneIfNew(MI);
1678 WorkingMI.getOperand(MI.getNumOperands() - 1).setImm(Imm);
1679 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1682 case X86::VPCOMBri: case X86::VPCOMUBri:
1683 case X86::VPCOMDri: case X86::VPCOMUDri:
1684 case X86::VPCOMQri: case X86::VPCOMUQri:
1685 case X86::VPCOMWri: case X86::VPCOMUWri: {
1686 // Flip comparison mode immediate (if necessary).
1687 unsigned Imm = MI.getOperand(3).getImm() & 0x7;
1688 Imm = X86::getSwappedVPCOMImm(Imm);
1689 auto &WorkingMI = cloneIfNew(MI);
1690 WorkingMI.getOperand(3).setImm(Imm);
1691 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1694 case X86::VPERM2F128rr:
1695 case X86::VPERM2I128rr: {
1696 // Flip permute source immediate.
1697 // Imm & 0x02: lo = if set, select Op1.lo/hi else Op0.lo/hi.
1698 // Imm & 0x20: hi = if set, select Op1.lo/hi else Op0.lo/hi.
1699 unsigned Imm = MI.getOperand(3).getImm() & 0xFF;
1700 auto &WorkingMI = cloneIfNew(MI);
1701 WorkingMI.getOperand(3).setImm(Imm ^ 0x22);
1702 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1705 case X86::MOVHLPSrr:
1706 case X86::UNPCKHPDrr:
1707 case X86::VMOVHLPSrr:
1708 case X86::VUNPCKHPDrr:
1709 case X86::VMOVHLPSZrr:
1710 case X86::VUNPCKHPDZ128rr: {
1711 assert(Subtarget.hasSSE2() && "Commuting MOVHLP/UNPCKHPD requires SSE2!");
1713 unsigned Opc = MI.getOpcode();
1715 default: llvm_unreachable("Unreachable!");
1716 case X86::MOVHLPSrr: Opc = X86::UNPCKHPDrr; break;
1717 case X86::UNPCKHPDrr: Opc = X86::MOVHLPSrr; break;
1718 case X86::VMOVHLPSrr: Opc = X86::VUNPCKHPDrr; break;
1719 case X86::VUNPCKHPDrr: Opc = X86::VMOVHLPSrr; break;
1720 case X86::VMOVHLPSZrr: Opc = X86::VUNPCKHPDZ128rr; break;
1721 case X86::VUNPCKHPDZ128rr: Opc = X86::VMOVHLPSZrr; break;
1723 auto &WorkingMI = cloneIfNew(MI);
1724 WorkingMI.setDesc(get(Opc));
1725 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1728 case X86::CMOVB16rr: case X86::CMOVB32rr: case X86::CMOVB64rr:
1729 case X86::CMOVAE16rr: case X86::CMOVAE32rr: case X86::CMOVAE64rr:
1730 case X86::CMOVE16rr: case X86::CMOVE32rr: case X86::CMOVE64rr:
1731 case X86::CMOVNE16rr: case X86::CMOVNE32rr: case X86::CMOVNE64rr:
1732 case X86::CMOVBE16rr: case X86::CMOVBE32rr: case X86::CMOVBE64rr:
1733 case X86::CMOVA16rr: case X86::CMOVA32rr: case X86::CMOVA64rr:
1734 case X86::CMOVL16rr: case X86::CMOVL32rr: case X86::CMOVL64rr:
1735 case X86::CMOVGE16rr: case X86::CMOVGE32rr: case X86::CMOVGE64rr:
1736 case X86::CMOVLE16rr: case X86::CMOVLE32rr: case X86::CMOVLE64rr:
1737 case X86::CMOVG16rr: case X86::CMOVG32rr: case X86::CMOVG64rr:
1738 case X86::CMOVS16rr: case X86::CMOVS32rr: case X86::CMOVS64rr:
1739 case X86::CMOVNS16rr: case X86::CMOVNS32rr: case X86::CMOVNS64rr:
1740 case X86::CMOVP16rr: case X86::CMOVP32rr: case X86::CMOVP64rr:
1741 case X86::CMOVNP16rr: case X86::CMOVNP32rr: case X86::CMOVNP64rr:
1742 case X86::CMOVO16rr: case X86::CMOVO32rr: case X86::CMOVO64rr:
1743 case X86::CMOVNO16rr: case X86::CMOVNO32rr: case X86::CMOVNO64rr: {
1745 switch (MI.getOpcode()) {
1746 default: llvm_unreachable("Unreachable!");
1747 case X86::CMOVB16rr: Opc = X86::CMOVAE16rr; break;
1748 case X86::CMOVB32rr: Opc = X86::CMOVAE32rr; break;
1749 case X86::CMOVB64rr: Opc = X86::CMOVAE64rr; break;
1750 case X86::CMOVAE16rr: Opc = X86::CMOVB16rr; break;
1751 case X86::CMOVAE32rr: Opc = X86::CMOVB32rr; break;
1752 case X86::CMOVAE64rr: Opc = X86::CMOVB64rr; break;
1753 case X86::CMOVE16rr: Opc = X86::CMOVNE16rr; break;
1754 case X86::CMOVE32rr: Opc = X86::CMOVNE32rr; break;
1755 case X86::CMOVE64rr: Opc = X86::CMOVNE64rr; break;
1756 case X86::CMOVNE16rr: Opc = X86::CMOVE16rr; break;
1757 case X86::CMOVNE32rr: Opc = X86::CMOVE32rr; break;
1758 case X86::CMOVNE64rr: Opc = X86::CMOVE64rr; break;
1759 case X86::CMOVBE16rr: Opc = X86::CMOVA16rr; break;
1760 case X86::CMOVBE32rr: Opc = X86::CMOVA32rr; break;
1761 case X86::CMOVBE64rr: Opc = X86::CMOVA64rr; break;
1762 case X86::CMOVA16rr: Opc = X86::CMOVBE16rr; break;
1763 case X86::CMOVA32rr: Opc = X86::CMOVBE32rr; break;
1764 case X86::CMOVA64rr: Opc = X86::CMOVBE64rr; break;
1765 case X86::CMOVL16rr: Opc = X86::CMOVGE16rr; break;
1766 case X86::CMOVL32rr: Opc = X86::CMOVGE32rr; break;
1767 case X86::CMOVL64rr: Opc = X86::CMOVGE64rr; break;
1768 case X86::CMOVGE16rr: Opc = X86::CMOVL16rr; break;
1769 case X86::CMOVGE32rr: Opc = X86::CMOVL32rr; break;
1770 case X86::CMOVGE64rr: Opc = X86::CMOVL64rr; break;
1771 case X86::CMOVLE16rr: Opc = X86::CMOVG16rr; break;
1772 case X86::CMOVLE32rr: Opc = X86::CMOVG32rr; break;
1773 case X86::CMOVLE64rr: Opc = X86::CMOVG64rr; break;
1774 case X86::CMOVG16rr: Opc = X86::CMOVLE16rr; break;
1775 case X86::CMOVG32rr: Opc = X86::CMOVLE32rr; break;
1776 case X86::CMOVG64rr: Opc = X86::CMOVLE64rr; break;
1777 case X86::CMOVS16rr: Opc = X86::CMOVNS16rr; break;
1778 case X86::CMOVS32rr: Opc = X86::CMOVNS32rr; break;
1779 case X86::CMOVS64rr: Opc = X86::CMOVNS64rr; break;
1780 case X86::CMOVNS16rr: Opc = X86::CMOVS16rr; break;
1781 case X86::CMOVNS32rr: Opc = X86::CMOVS32rr; break;
1782 case X86::CMOVNS64rr: Opc = X86::CMOVS64rr; break;
1783 case X86::CMOVP16rr: Opc = X86::CMOVNP16rr; break;
1784 case X86::CMOVP32rr: Opc = X86::CMOVNP32rr; break;
1785 case X86::CMOVP64rr: Opc = X86::CMOVNP64rr; break;
1786 case X86::CMOVNP16rr: Opc = X86::CMOVP16rr; break;
1787 case X86::CMOVNP32rr: Opc = X86::CMOVP32rr; break;
1788 case X86::CMOVNP64rr: Opc = X86::CMOVP64rr; break;
1789 case X86::CMOVO16rr: Opc = X86::CMOVNO16rr; break;
1790 case X86::CMOVO32rr: Opc = X86::CMOVNO32rr; break;
1791 case X86::CMOVO64rr: Opc = X86::CMOVNO64rr; break;
1792 case X86::CMOVNO16rr: Opc = X86::CMOVO16rr; break;
1793 case X86::CMOVNO32rr: Opc = X86::CMOVO32rr; break;
1794 case X86::CMOVNO64rr: Opc = X86::CMOVO64rr; break;
1796 auto &WorkingMI = cloneIfNew(MI);
1797 WorkingMI.setDesc(get(Opc));
1798 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1801 case X86::VPTERNLOGDZrri: case X86::VPTERNLOGDZrmi:
1802 case X86::VPTERNLOGDZ128rri: case X86::VPTERNLOGDZ128rmi:
1803 case X86::VPTERNLOGDZ256rri: case X86::VPTERNLOGDZ256rmi:
1804 case X86::VPTERNLOGQZrri: case X86::VPTERNLOGQZrmi:
1805 case X86::VPTERNLOGQZ128rri: case X86::VPTERNLOGQZ128rmi:
1806 case X86::VPTERNLOGQZ256rri: case X86::VPTERNLOGQZ256rmi:
1807 case X86::VPTERNLOGDZrrik:
1808 case X86::VPTERNLOGDZ128rrik:
1809 case X86::VPTERNLOGDZ256rrik:
1810 case X86::VPTERNLOGQZrrik:
1811 case X86::VPTERNLOGQZ128rrik:
1812 case X86::VPTERNLOGQZ256rrik:
1813 case X86::VPTERNLOGDZrrikz: case X86::VPTERNLOGDZrmikz:
1814 case X86::VPTERNLOGDZ128rrikz: case X86::VPTERNLOGDZ128rmikz:
1815 case X86::VPTERNLOGDZ256rrikz: case X86::VPTERNLOGDZ256rmikz:
1816 case X86::VPTERNLOGQZrrikz: case X86::VPTERNLOGQZrmikz:
1817 case X86::VPTERNLOGQZ128rrikz: case X86::VPTERNLOGQZ128rmikz:
1818 case X86::VPTERNLOGQZ256rrikz: case X86::VPTERNLOGQZ256rmikz:
1819 case X86::VPTERNLOGDZ128rmbi:
1820 case X86::VPTERNLOGDZ256rmbi:
1821 case X86::VPTERNLOGDZrmbi:
1822 case X86::VPTERNLOGQZ128rmbi:
1823 case X86::VPTERNLOGQZ256rmbi:
1824 case X86::VPTERNLOGQZrmbi:
1825 case X86::VPTERNLOGDZ128rmbikz:
1826 case X86::VPTERNLOGDZ256rmbikz:
1827 case X86::VPTERNLOGDZrmbikz:
1828 case X86::VPTERNLOGQZ128rmbikz:
1829 case X86::VPTERNLOGQZ256rmbikz:
1830 case X86::VPTERNLOGQZrmbikz: {
1831 auto &WorkingMI = cloneIfNew(MI);
1832 commuteVPTERNLOG(WorkingMI, OpIdx1, OpIdx2);
1833 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1837 if (isCommutableVPERMV3Instruction(MI.getOpcode())) {
1838 unsigned Opc = getCommutedVPERMV3Opcode(MI.getOpcode());
1839 auto &WorkingMI = cloneIfNew(MI);
1840 WorkingMI.setDesc(get(Opc));
1841 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1845 const X86InstrFMA3Group *FMA3Group = getFMA3Group(MI.getOpcode(),
1846 MI.getDesc().TSFlags);
1849 getFMA3OpcodeToCommuteOperands(MI, OpIdx1, OpIdx2, *FMA3Group);
1850 auto &WorkingMI = cloneIfNew(MI);
1851 WorkingMI.setDesc(get(Opc));
1852 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1856 return TargetInstrInfo::commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2);
1862 X86InstrInfo::findThreeSrcCommutedOpIndices(const MachineInstr &MI,
1863 unsigned &SrcOpIdx1,
1864 unsigned &SrcOpIdx2,
1865 bool IsIntrinsic) const {
1866 uint64_t TSFlags = MI.getDesc().TSFlags;
1868 unsigned FirstCommutableVecOp = 1;
1869 unsigned LastCommutableVecOp = 3;
1870 unsigned KMaskOp = -1U;
1871 if (X86II::isKMasked(TSFlags)) {
1872 // For k-zero-masked operations it is Ok to commute the first vector
1874 // For regular k-masked operations a conservative choice is done as the
1875 // elements of the first vector operand, for which the corresponding bit
1876 // in the k-mask operand is set to 0, are copied to the result of the
1878 // TODO/FIXME: The commute still may be legal if it is known that the
1879 // k-mask operand is set to either all ones or all zeroes.
1880 // It is also Ok to commute the 1st operand if all users of MI use only
1881 // the elements enabled by the k-mask operand. For example,
1882 // v4 = VFMADD213PSZrk v1, k, v2, v3; // v1[i] = k[i] ? v2[i]*v1[i]+v3[i]
1884 // VMOVAPSZmrk <mem_addr>, k, v4; // this is the ONLY user of v4 ->
1885 // // Ok, to commute v1 in FMADD213PSZrk.
1887 // The k-mask operand has index = 2 for masked and zero-masked operations.
1890 // The operand with index = 1 is used as a source for those elements for
1891 // which the corresponding bit in the k-mask is set to 0.
1892 if (X86II::isKMergeMasked(TSFlags))
1893 FirstCommutableVecOp = 3;
1895 LastCommutableVecOp++;
1896 } else if (IsIntrinsic) {
1897 // Commuting the first operand of an intrinsic instruction isn't possible
1898 // unless we can prove that only the lowest element of the result is used.
1899 FirstCommutableVecOp = 2;
1902 if (isMem(MI, LastCommutableVecOp))
1903 LastCommutableVecOp--;
1905 // Only the first RegOpsNum operands are commutable.
1906 // Also, the value 'CommuteAnyOperandIndex' is valid here as it means
1907 // that the operand is not specified/fixed.
1908 if (SrcOpIdx1 != CommuteAnyOperandIndex &&
1909 (SrcOpIdx1 < FirstCommutableVecOp || SrcOpIdx1 > LastCommutableVecOp ||
1910 SrcOpIdx1 == KMaskOp))
1912 if (SrcOpIdx2 != CommuteAnyOperandIndex &&
1913 (SrcOpIdx2 < FirstCommutableVecOp || SrcOpIdx2 > LastCommutableVecOp ||
1914 SrcOpIdx2 == KMaskOp))
1917 // Look for two different register operands assumed to be commutable
1918 // regardless of the FMA opcode. The FMA opcode is adjusted later.
1919 if (SrcOpIdx1 == CommuteAnyOperandIndex ||
1920 SrcOpIdx2 == CommuteAnyOperandIndex) {
1921 unsigned CommutableOpIdx1 = SrcOpIdx1;
1922 unsigned CommutableOpIdx2 = SrcOpIdx2;
1924 // At least one of operands to be commuted is not specified and
1925 // this method is free to choose appropriate commutable operands.
1926 if (SrcOpIdx1 == SrcOpIdx2)
1927 // Both of operands are not fixed. By default set one of commutable
1928 // operands to the last register operand of the instruction.
1929 CommutableOpIdx2 = LastCommutableVecOp;
1930 else if (SrcOpIdx2 == CommuteAnyOperandIndex)
1931 // Only one of operands is not fixed.
1932 CommutableOpIdx2 = SrcOpIdx1;
1934 // CommutableOpIdx2 is well defined now. Let's choose another commutable
1935 // operand and assign its index to CommutableOpIdx1.
1936 unsigned Op2Reg = MI.getOperand(CommutableOpIdx2).getReg();
1937 for (CommutableOpIdx1 = LastCommutableVecOp;
1938 CommutableOpIdx1 >= FirstCommutableVecOp; CommutableOpIdx1--) {
1939 // Just ignore and skip the k-mask operand.
1940 if (CommutableOpIdx1 == KMaskOp)
1943 // The commuted operands must have different registers.
1944 // Otherwise, the commute transformation does not change anything and
1946 if (Op2Reg != MI.getOperand(CommutableOpIdx1).getReg())
1950 // No appropriate commutable operands were found.
1951 if (CommutableOpIdx1 < FirstCommutableVecOp)
1954 // Assign the found pair of commutable indices to SrcOpIdx1 and SrcOpidx2
1955 // to return those values.
1956 if (!fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2,
1957 CommutableOpIdx1, CommutableOpIdx2))
1964 bool X86InstrInfo::findCommutedOpIndices(MachineInstr &MI, unsigned &SrcOpIdx1,
1965 unsigned &SrcOpIdx2) const {
1966 const MCInstrDesc &Desc = MI.getDesc();
1967 if (!Desc.isCommutable())
1970 switch (MI.getOpcode()) {
1977 case X86::VCMPPDrri:
1978 case X86::VCMPPSrri:
1979 case X86::VCMPPDYrri:
1980 case X86::VCMPPSYrri:
1981 case X86::VCMPSDZrr:
1982 case X86::VCMPSSZrr:
1983 case X86::VCMPPDZrri:
1984 case X86::VCMPPSZrri:
1985 case X86::VCMPPDZ128rri:
1986 case X86::VCMPPSZ128rri:
1987 case X86::VCMPPDZ256rri:
1988 case X86::VCMPPSZ256rri: {
1989 // Float comparison can be safely commuted for
1990 // Ordered/Unordered/Equal/NotEqual tests
1991 unsigned Imm = MI.getOperand(3).getImm() & 0x7;
1994 case 0x03: // UNORDERED
1995 case 0x04: // NOT EQUAL
1996 case 0x07: // ORDERED
1997 // The indices of the commutable operands are 1 and 2.
1998 // Assign them to the returned operand indices here.
1999 return fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2, 1, 2);
2007 if (Subtarget.hasSSE41())
2008 return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2);
2010 case X86::MOVHLPSrr:
2011 case X86::UNPCKHPDrr:
2012 case X86::VMOVHLPSrr:
2013 case X86::VUNPCKHPDrr:
2014 case X86::VMOVHLPSZrr:
2015 case X86::VUNPCKHPDZ128rr:
2016 if (Subtarget.hasSSE2())
2017 return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2);
2019 case X86::VPTERNLOGDZrri: case X86::VPTERNLOGDZrmi:
2020 case X86::VPTERNLOGDZ128rri: case X86::VPTERNLOGDZ128rmi:
2021 case X86::VPTERNLOGDZ256rri: case X86::VPTERNLOGDZ256rmi:
2022 case X86::VPTERNLOGQZrri: case X86::VPTERNLOGQZrmi:
2023 case X86::VPTERNLOGQZ128rri: case X86::VPTERNLOGQZ128rmi:
2024 case X86::VPTERNLOGQZ256rri: case X86::VPTERNLOGQZ256rmi:
2025 case X86::VPTERNLOGDZrrik:
2026 case X86::VPTERNLOGDZ128rrik:
2027 case X86::VPTERNLOGDZ256rrik:
2028 case X86::VPTERNLOGQZrrik:
2029 case X86::VPTERNLOGQZ128rrik:
2030 case X86::VPTERNLOGQZ256rrik:
2031 case X86::VPTERNLOGDZrrikz: case X86::VPTERNLOGDZrmikz:
2032 case X86::VPTERNLOGDZ128rrikz: case X86::VPTERNLOGDZ128rmikz:
2033 case X86::VPTERNLOGDZ256rrikz: case X86::VPTERNLOGDZ256rmikz:
2034 case X86::VPTERNLOGQZrrikz: case X86::VPTERNLOGQZrmikz:
2035 case X86::VPTERNLOGQZ128rrikz: case X86::VPTERNLOGQZ128rmikz:
2036 case X86::VPTERNLOGQZ256rrikz: case X86::VPTERNLOGQZ256rmikz:
2037 case X86::VPTERNLOGDZ128rmbi:
2038 case X86::VPTERNLOGDZ256rmbi:
2039 case X86::VPTERNLOGDZrmbi:
2040 case X86::VPTERNLOGQZ128rmbi:
2041 case X86::VPTERNLOGQZ256rmbi:
2042 case X86::VPTERNLOGQZrmbi:
2043 case X86::VPTERNLOGDZ128rmbikz:
2044 case X86::VPTERNLOGDZ256rmbikz:
2045 case X86::VPTERNLOGDZrmbikz:
2046 case X86::VPTERNLOGQZ128rmbikz:
2047 case X86::VPTERNLOGQZ256rmbikz:
2048 case X86::VPTERNLOGQZrmbikz:
2049 return findThreeSrcCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2);
2050 case X86::VPMADD52HUQZ128r:
2051 case X86::VPMADD52HUQZ128rk:
2052 case X86::VPMADD52HUQZ128rkz:
2053 case X86::VPMADD52HUQZ256r:
2054 case X86::VPMADD52HUQZ256rk:
2055 case X86::VPMADD52HUQZ256rkz:
2056 case X86::VPMADD52HUQZr:
2057 case X86::VPMADD52HUQZrk:
2058 case X86::VPMADD52HUQZrkz:
2059 case X86::VPMADD52LUQZ128r:
2060 case X86::VPMADD52LUQZ128rk:
2061 case X86::VPMADD52LUQZ128rkz:
2062 case X86::VPMADD52LUQZ256r:
2063 case X86::VPMADD52LUQZ256rk:
2064 case X86::VPMADD52LUQZ256rkz:
2065 case X86::VPMADD52LUQZr:
2066 case X86::VPMADD52LUQZrk:
2067 case X86::VPMADD52LUQZrkz: {
2068 unsigned CommutableOpIdx1 = 2;
2069 unsigned CommutableOpIdx2 = 3;
2070 if (X86II::isKMasked(Desc.TSFlags)) {
2071 // Skip the mask register.
2075 if (!fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2,
2076 CommutableOpIdx1, CommutableOpIdx2))
2078 if (!MI.getOperand(SrcOpIdx1).isReg() ||
2079 !MI.getOperand(SrcOpIdx2).isReg())
2086 const X86InstrFMA3Group *FMA3Group = getFMA3Group(MI.getOpcode(),
2087 MI.getDesc().TSFlags);
2089 return findThreeSrcCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2,
2090 FMA3Group->isIntrinsic());
2092 // Handled masked instructions since we need to skip over the mask input
2093 // and the preserved input.
2094 if (X86II::isKMasked(Desc.TSFlags)) {
2095 // First assume that the first input is the mask operand and skip past it.
2096 unsigned CommutableOpIdx1 = Desc.getNumDefs() + 1;
2097 unsigned CommutableOpIdx2 = Desc.getNumDefs() + 2;
2098 // Check if the first input is tied. If there isn't one then we only
2099 // need to skip the mask operand which we did above.
2100 if ((MI.getDesc().getOperandConstraint(Desc.getNumDefs(),
2101 MCOI::TIED_TO) != -1)) {
2102 // If this is zero masking instruction with a tied operand, we need to
2103 // move the first index back to the first input since this must
2104 // be a 3 input instruction and we want the first two non-mask inputs.
2105 // Otherwise this is a 2 input instruction with a preserved input and
2106 // mask, so we need to move the indices to skip one more input.
2107 if (X86II::isKMergeMasked(Desc.TSFlags)) {
2115 if (!fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2,
2116 CommutableOpIdx1, CommutableOpIdx2))
2119 if (!MI.getOperand(SrcOpIdx1).isReg() ||
2120 !MI.getOperand(SrcOpIdx2).isReg())
2126 return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2);
2131 X86::CondCode X86::getCondFromBranchOpc(unsigned BrOpc) {
2133 default: return X86::COND_INVALID;
2134 case X86::JE_1: return X86::COND_E;
2135 case X86::JNE_1: return X86::COND_NE;
2136 case X86::JL_1: return X86::COND_L;
2137 case X86::JLE_1: return X86::COND_LE;
2138 case X86::JG_1: return X86::COND_G;
2139 case X86::JGE_1: return X86::COND_GE;
2140 case X86::JB_1: return X86::COND_B;
2141 case X86::JBE_1: return X86::COND_BE;
2142 case X86::JA_1: return X86::COND_A;
2143 case X86::JAE_1: return X86::COND_AE;
2144 case X86::JS_1: return X86::COND_S;
2145 case X86::JNS_1: return X86::COND_NS;
2146 case X86::JP_1: return X86::COND_P;
2147 case X86::JNP_1: return X86::COND_NP;
2148 case X86::JO_1: return X86::COND_O;
2149 case X86::JNO_1: return X86::COND_NO;
2153 /// Return condition code of a SET opcode.
2154 X86::CondCode X86::getCondFromSETOpc(unsigned Opc) {
2156 default: return X86::COND_INVALID;
2157 case X86::SETAr: case X86::SETAm: return X86::COND_A;
2158 case X86::SETAEr: case X86::SETAEm: return X86::COND_AE;
2159 case X86::SETBr: case X86::SETBm: return X86::COND_B;
2160 case X86::SETBEr: case X86::SETBEm: return X86::COND_BE;
2161 case X86::SETEr: case X86::SETEm: return X86::COND_E;
2162 case X86::SETGr: case X86::SETGm: return X86::COND_G;
2163 case X86::SETGEr: case X86::SETGEm: return X86::COND_GE;
2164 case X86::SETLr: case X86::SETLm: return X86::COND_L;
2165 case X86::SETLEr: case X86::SETLEm: return X86::COND_LE;
2166 case X86::SETNEr: case X86::SETNEm: return X86::COND_NE;
2167 case X86::SETNOr: case X86::SETNOm: return X86::COND_NO;
2168 case X86::SETNPr: case X86::SETNPm: return X86::COND_NP;
2169 case X86::SETNSr: case X86::SETNSm: return X86::COND_NS;
2170 case X86::SETOr: case X86::SETOm: return X86::COND_O;
2171 case X86::SETPr: case X86::SETPm: return X86::COND_P;
2172 case X86::SETSr: case X86::SETSm: return X86::COND_S;
2176 /// Return condition code of a CMov opcode.
2177 X86::CondCode X86::getCondFromCMovOpc(unsigned Opc) {
2179 default: return X86::COND_INVALID;
2180 case X86::CMOVA16rm: case X86::CMOVA16rr: case X86::CMOVA32rm:
2181 case X86::CMOVA32rr: case X86::CMOVA64rm: case X86::CMOVA64rr:
2183 case X86::CMOVAE16rm: case X86::CMOVAE16rr: case X86::CMOVAE32rm:
2184 case X86::CMOVAE32rr: case X86::CMOVAE64rm: case X86::CMOVAE64rr:
2185 return X86::COND_AE;
2186 case X86::CMOVB16rm: case X86::CMOVB16rr: case X86::CMOVB32rm:
2187 case X86::CMOVB32rr: case X86::CMOVB64rm: case X86::CMOVB64rr:
2189 case X86::CMOVBE16rm: case X86::CMOVBE16rr: case X86::CMOVBE32rm:
2190 case X86::CMOVBE32rr: case X86::CMOVBE64rm: case X86::CMOVBE64rr:
2191 return X86::COND_BE;
2192 case X86::CMOVE16rm: case X86::CMOVE16rr: case X86::CMOVE32rm:
2193 case X86::CMOVE32rr: case X86::CMOVE64rm: case X86::CMOVE64rr:
2195 case X86::CMOVG16rm: case X86::CMOVG16rr: case X86::CMOVG32rm:
2196 case X86::CMOVG32rr: case X86::CMOVG64rm: case X86::CMOVG64rr:
2198 case X86::CMOVGE16rm: case X86::CMOVGE16rr: case X86::CMOVGE32rm:
2199 case X86::CMOVGE32rr: case X86::CMOVGE64rm: case X86::CMOVGE64rr:
2200 return X86::COND_GE;
2201 case X86::CMOVL16rm: case X86::CMOVL16rr: case X86::CMOVL32rm:
2202 case X86::CMOVL32rr: case X86::CMOVL64rm: case X86::CMOVL64rr:
2204 case X86::CMOVLE16rm: case X86::CMOVLE16rr: case X86::CMOVLE32rm:
2205 case X86::CMOVLE32rr: case X86::CMOVLE64rm: case X86::CMOVLE64rr:
2206 return X86::COND_LE;
2207 case X86::CMOVNE16rm: case X86::CMOVNE16rr: case X86::CMOVNE32rm:
2208 case X86::CMOVNE32rr: case X86::CMOVNE64rm: case X86::CMOVNE64rr:
2209 return X86::COND_NE;
2210 case X86::CMOVNO16rm: case X86::CMOVNO16rr: case X86::CMOVNO32rm:
2211 case X86::CMOVNO32rr: case X86::CMOVNO64rm: case X86::CMOVNO64rr:
2212 return X86::COND_NO;
2213 case X86::CMOVNP16rm: case X86::CMOVNP16rr: case X86::CMOVNP32rm:
2214 case X86::CMOVNP32rr: case X86::CMOVNP64rm: case X86::CMOVNP64rr:
2215 return X86::COND_NP;
2216 case X86::CMOVNS16rm: case X86::CMOVNS16rr: case X86::CMOVNS32rm:
2217 case X86::CMOVNS32rr: case X86::CMOVNS64rm: case X86::CMOVNS64rr:
2218 return X86::COND_NS;
2219 case X86::CMOVO16rm: case X86::CMOVO16rr: case X86::CMOVO32rm:
2220 case X86::CMOVO32rr: case X86::CMOVO64rm: case X86::CMOVO64rr:
2222 case X86::CMOVP16rm: case X86::CMOVP16rr: case X86::CMOVP32rm:
2223 case X86::CMOVP32rr: case X86::CMOVP64rm: case X86::CMOVP64rr:
2225 case X86::CMOVS16rm: case X86::CMOVS16rr: case X86::CMOVS32rm:
2226 case X86::CMOVS32rr: case X86::CMOVS64rm: case X86::CMOVS64rr:
2231 unsigned X86::GetCondBranchFromCond(X86::CondCode CC) {
2233 default: llvm_unreachable("Illegal condition code!");
2234 case X86::COND_E: return X86::JE_1;
2235 case X86::COND_NE: return X86::JNE_1;
2236 case X86::COND_L: return X86::JL_1;
2237 case X86::COND_LE: return X86::JLE_1;
2238 case X86::COND_G: return X86::JG_1;
2239 case X86::COND_GE: return X86::JGE_1;
2240 case X86::COND_B: return X86::JB_1;
2241 case X86::COND_BE: return X86::JBE_1;
2242 case X86::COND_A: return X86::JA_1;
2243 case X86::COND_AE: return X86::JAE_1;
2244 case X86::COND_S: return X86::JS_1;
2245 case X86::COND_NS: return X86::JNS_1;
2246 case X86::COND_P: return X86::JP_1;
2247 case X86::COND_NP: return X86::JNP_1;
2248 case X86::COND_O: return X86::JO_1;
2249 case X86::COND_NO: return X86::JNO_1;
2253 /// Return the inverse of the specified condition,
2254 /// e.g. turning COND_E to COND_NE.
2255 X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) {
2257 default: llvm_unreachable("Illegal condition code!");
2258 case X86::COND_E: return X86::COND_NE;
2259 case X86::COND_NE: return X86::COND_E;
2260 case X86::COND_L: return X86::COND_GE;
2261 case X86::COND_LE: return X86::COND_G;
2262 case X86::COND_G: return X86::COND_LE;
2263 case X86::COND_GE: return X86::COND_L;
2264 case X86::COND_B: return X86::COND_AE;
2265 case X86::COND_BE: return X86::COND_A;
2266 case X86::COND_A: return X86::COND_BE;
2267 case X86::COND_AE: return X86::COND_B;
2268 case X86::COND_S: return X86::COND_NS;
2269 case X86::COND_NS: return X86::COND_S;
2270 case X86::COND_P: return X86::COND_NP;
2271 case X86::COND_NP: return X86::COND_P;
2272 case X86::COND_O: return X86::COND_NO;
2273 case X86::COND_NO: return X86::COND_O;
2274 case X86::COND_NE_OR_P: return X86::COND_E_AND_NP;
2275 case X86::COND_E_AND_NP: return X86::COND_NE_OR_P;
2279 /// Assuming the flags are set by MI(a,b), return the condition code if we
2280 /// modify the instructions such that flags are set by MI(b,a).
2281 static X86::CondCode getSwappedCondition(X86::CondCode CC) {
2283 default: return X86::COND_INVALID;
2284 case X86::COND_E: return X86::COND_E;
2285 case X86::COND_NE: return X86::COND_NE;
2286 case X86::COND_L: return X86::COND_G;
2287 case X86::COND_LE: return X86::COND_GE;
2288 case X86::COND_G: return X86::COND_L;
2289 case X86::COND_GE: return X86::COND_LE;
2290 case X86::COND_B: return X86::COND_A;
2291 case X86::COND_BE: return X86::COND_AE;
2292 case X86::COND_A: return X86::COND_B;
2293 case X86::COND_AE: return X86::COND_BE;
2297 std::pair<X86::CondCode, bool>
2298 X86::getX86ConditionCode(CmpInst::Predicate Predicate) {
2299 X86::CondCode CC = X86::COND_INVALID;
2300 bool NeedSwap = false;
2301 switch (Predicate) {
2303 // Floating-point Predicates
2304 case CmpInst::FCMP_UEQ: CC = X86::COND_E; break;
2305 case CmpInst::FCMP_OLT: NeedSwap = true; LLVM_FALLTHROUGH;
2306 case CmpInst::FCMP_OGT: CC = X86::COND_A; break;
2307 case CmpInst::FCMP_OLE: NeedSwap = true; LLVM_FALLTHROUGH;
2308 case CmpInst::FCMP_OGE: CC = X86::COND_AE; break;
2309 case CmpInst::FCMP_UGT: NeedSwap = true; LLVM_FALLTHROUGH;
2310 case CmpInst::FCMP_ULT: CC = X86::COND_B; break;
2311 case CmpInst::FCMP_UGE: NeedSwap = true; LLVM_FALLTHROUGH;
2312 case CmpInst::FCMP_ULE: CC = X86::COND_BE; break;
2313 case CmpInst::FCMP_ONE: CC = X86::COND_NE; break;
2314 case CmpInst::FCMP_UNO: CC = X86::COND_P; break;
2315 case CmpInst::FCMP_ORD: CC = X86::COND_NP; break;
2316 case CmpInst::FCMP_OEQ: LLVM_FALLTHROUGH;
2317 case CmpInst::FCMP_UNE: CC = X86::COND_INVALID; break;
2319 // Integer Predicates
2320 case CmpInst::ICMP_EQ: CC = X86::COND_E; break;
2321 case CmpInst::ICMP_NE: CC = X86::COND_NE; break;
2322 case CmpInst::ICMP_UGT: CC = X86::COND_A; break;
2323 case CmpInst::ICMP_UGE: CC = X86::COND_AE; break;
2324 case CmpInst::ICMP_ULT: CC = X86::COND_B; break;
2325 case CmpInst::ICMP_ULE: CC = X86::COND_BE; break;
2326 case CmpInst::ICMP_SGT: CC = X86::COND_G; break;
2327 case CmpInst::ICMP_SGE: CC = X86::COND_GE; break;
2328 case CmpInst::ICMP_SLT: CC = X86::COND_L; break;
2329 case CmpInst::ICMP_SLE: CC = X86::COND_LE; break;
2332 return std::make_pair(CC, NeedSwap);
2335 /// Return a set opcode for the given condition and
2336 /// whether it has memory operand.
2337 unsigned X86::getSETFromCond(CondCode CC, bool HasMemoryOperand) {
2338 static const uint16_t Opc[16][2] = {
2339 { X86::SETAr, X86::SETAm },
2340 { X86::SETAEr, X86::SETAEm },
2341 { X86::SETBr, X86::SETBm },
2342 { X86::SETBEr, X86::SETBEm },
2343 { X86::SETEr, X86::SETEm },
2344 { X86::SETGr, X86::SETGm },
2345 { X86::SETGEr, X86::SETGEm },
2346 { X86::SETLr, X86::SETLm },
2347 { X86::SETLEr, X86::SETLEm },
2348 { X86::SETNEr, X86::SETNEm },
2349 { X86::SETNOr, X86::SETNOm },
2350 { X86::SETNPr, X86::SETNPm },
2351 { X86::SETNSr, X86::SETNSm },
2352 { X86::SETOr, X86::SETOm },
2353 { X86::SETPr, X86::SETPm },
2354 { X86::SETSr, X86::SETSm }
2357 assert(CC <= LAST_VALID_COND && "Can only handle standard cond codes");
2358 return Opc[CC][HasMemoryOperand ? 1 : 0];
2361 /// Return a cmov opcode for the given condition,
2362 /// register size in bytes, and operand type.
2363 unsigned X86::getCMovFromCond(CondCode CC, unsigned RegBytes,
2364 bool HasMemoryOperand) {
2365 static const uint16_t Opc[32][3] = {
2366 { X86::CMOVA16rr, X86::CMOVA32rr, X86::CMOVA64rr },
2367 { X86::CMOVAE16rr, X86::CMOVAE32rr, X86::CMOVAE64rr },
2368 { X86::CMOVB16rr, X86::CMOVB32rr, X86::CMOVB64rr },
2369 { X86::CMOVBE16rr, X86::CMOVBE32rr, X86::CMOVBE64rr },
2370 { X86::CMOVE16rr, X86::CMOVE32rr, X86::CMOVE64rr },
2371 { X86::CMOVG16rr, X86::CMOVG32rr, X86::CMOVG64rr },
2372 { X86::CMOVGE16rr, X86::CMOVGE32rr, X86::CMOVGE64rr },
2373 { X86::CMOVL16rr, X86::CMOVL32rr, X86::CMOVL64rr },
2374 { X86::CMOVLE16rr, X86::CMOVLE32rr, X86::CMOVLE64rr },
2375 { X86::CMOVNE16rr, X86::CMOVNE32rr, X86::CMOVNE64rr },
2376 { X86::CMOVNO16rr, X86::CMOVNO32rr, X86::CMOVNO64rr },
2377 { X86::CMOVNP16rr, X86::CMOVNP32rr, X86::CMOVNP64rr },
2378 { X86::CMOVNS16rr, X86::CMOVNS32rr, X86::CMOVNS64rr },
2379 { X86::CMOVO16rr, X86::CMOVO32rr, X86::CMOVO64rr },
2380 { X86::CMOVP16rr, X86::CMOVP32rr, X86::CMOVP64rr },
2381 { X86::CMOVS16rr, X86::CMOVS32rr, X86::CMOVS64rr },
2382 { X86::CMOVA16rm, X86::CMOVA32rm, X86::CMOVA64rm },
2383 { X86::CMOVAE16rm, X86::CMOVAE32rm, X86::CMOVAE64rm },
2384 { X86::CMOVB16rm, X86::CMOVB32rm, X86::CMOVB64rm },
2385 { X86::CMOVBE16rm, X86::CMOVBE32rm, X86::CMOVBE64rm },
2386 { X86::CMOVE16rm, X86::CMOVE32rm, X86::CMOVE64rm },
2387 { X86::CMOVG16rm, X86::CMOVG32rm, X86::CMOVG64rm },
2388 { X86::CMOVGE16rm, X86::CMOVGE32rm, X86::CMOVGE64rm },
2389 { X86::CMOVL16rm, X86::CMOVL32rm, X86::CMOVL64rm },
2390 { X86::CMOVLE16rm, X86::CMOVLE32rm, X86::CMOVLE64rm },
2391 { X86::CMOVNE16rm, X86::CMOVNE32rm, X86::CMOVNE64rm },
2392 { X86::CMOVNO16rm, X86::CMOVNO32rm, X86::CMOVNO64rm },
2393 { X86::CMOVNP16rm, X86::CMOVNP32rm, X86::CMOVNP64rm },
2394 { X86::CMOVNS16rm, X86::CMOVNS32rm, X86::CMOVNS64rm },
2395 { X86::CMOVO16rm, X86::CMOVO32rm, X86::CMOVO64rm },
2396 { X86::CMOVP16rm, X86::CMOVP32rm, X86::CMOVP64rm },
2397 { X86::CMOVS16rm, X86::CMOVS32rm, X86::CMOVS64rm }
2400 assert(CC < 16 && "Can only handle standard cond codes");
2401 unsigned Idx = HasMemoryOperand ? 16+CC : CC;
2403 default: llvm_unreachable("Illegal register size!");
2404 case 2: return Opc[Idx][0];
2405 case 4: return Opc[Idx][1];
2406 case 8: return Opc[Idx][2];
2410 /// Get the VPCMP immediate for the given condition.
2411 unsigned X86::getVPCMPImmForCond(ISD::CondCode CC) {
2413 default: llvm_unreachable("Unexpected SETCC condition");
2414 case ISD::SETNE: return 4;
2415 case ISD::SETEQ: return 0;
2417 case ISD::SETLT: return 1;
2419 case ISD::SETGT: return 6;
2421 case ISD::SETGE: return 5;
2423 case ISD::SETLE: return 2;
2427 /// Get the VPCMP immediate if the opcodes are swapped.
2428 unsigned X86::getSwappedVPCMPImm(unsigned Imm) {
2430 default: llvm_unreachable("Unreachable!");
2431 case 0x01: Imm = 0x06; break; // LT -> NLE
2432 case 0x02: Imm = 0x05; break; // LE -> NLT
2433 case 0x05: Imm = 0x02; break; // NLT -> LE
2434 case 0x06: Imm = 0x01; break; // NLE -> LT
2445 /// Get the VPCOM immediate if the opcodes are swapped.
2446 unsigned X86::getSwappedVPCOMImm(unsigned Imm) {
2448 default: llvm_unreachable("Unreachable!");
2449 case 0x00: Imm = 0x02; break; // LT -> GT
2450 case 0x01: Imm = 0x03; break; // LE -> GE
2451 case 0x02: Imm = 0x00; break; // GT -> LT
2452 case 0x03: Imm = 0x01; break; // GE -> LE
2463 bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr &MI) const {
2464 if (!MI.isTerminator()) return false;
2466 // Conditional branch is a special case.
2467 if (MI.isBranch() && !MI.isBarrier())
2469 if (!MI.isPredicable())
2471 return !isPredicated(MI);
2474 bool X86InstrInfo::isUnconditionalTailCall(const MachineInstr &MI) const {
2475 switch (MI.getOpcode()) {
2476 case X86::TCRETURNdi:
2477 case X86::TCRETURNri:
2478 case X86::TCRETURNmi:
2479 case X86::TCRETURNdi64:
2480 case X86::TCRETURNri64:
2481 case X86::TCRETURNmi64:
2488 bool X86InstrInfo::canMakeTailCallConditional(
2489 SmallVectorImpl<MachineOperand> &BranchCond,
2490 const MachineInstr &TailCall) const {
2491 if (TailCall.getOpcode() != X86::TCRETURNdi &&
2492 TailCall.getOpcode() != X86::TCRETURNdi64) {
2493 // Only direct calls can be done with a conditional branch.
2497 const MachineFunction *MF = TailCall.getParent()->getParent();
2498 if (Subtarget.isTargetWin64() && MF->hasWinCFI()) {
2499 // Conditional tail calls confuse the Win64 unwinder.
2503 assert(BranchCond.size() == 1);
2504 if (BranchCond[0].getImm() > X86::LAST_VALID_COND) {
2505 // Can't make a conditional tail call with this condition.
2509 const X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
2510 if (X86FI->getTCReturnAddrDelta() != 0 ||
2511 TailCall.getOperand(1).getImm() != 0) {
2512 // A conditional tail call cannot do any stack adjustment.
2519 void X86InstrInfo::replaceBranchWithTailCall(
2520 MachineBasicBlock &MBB, SmallVectorImpl<MachineOperand> &BranchCond,
2521 const MachineInstr &TailCall) const {
2522 assert(canMakeTailCallConditional(BranchCond, TailCall));
2524 MachineBasicBlock::iterator I = MBB.end();
2525 while (I != MBB.begin()) {
2527 if (I->isDebugInstr())
2530 assert(0 && "Can't find the branch to replace!");
2532 X86::CondCode CC = X86::getCondFromBranchOpc(I->getOpcode());
2533 assert(BranchCond.size() == 1);
2534 if (CC != BranchCond[0].getImm())
2540 unsigned Opc = TailCall.getOpcode() == X86::TCRETURNdi ? X86::TCRETURNdicc
2541 : X86::TCRETURNdi64cc;
2543 auto MIB = BuildMI(MBB, I, MBB.findDebugLoc(I), get(Opc));
2544 MIB->addOperand(TailCall.getOperand(0)); // Destination.
2545 MIB.addImm(0); // Stack offset (not used).
2546 MIB->addOperand(BranchCond[0]); // Condition.
2547 MIB.copyImplicitOps(TailCall); // Regmask and (imp-used) parameters.
2549 // Add implicit uses and defs of all live regs potentially clobbered by the
2550 // call. This way they still appear live across the call.
2551 LivePhysRegs LiveRegs(getRegisterInfo());
2552 LiveRegs.addLiveOuts(MBB);
2553 SmallVector<std::pair<unsigned, const MachineOperand *>, 8> Clobbers;
2554 LiveRegs.stepForward(*MIB, Clobbers);
2555 for (const auto &C : Clobbers) {
2556 MIB.addReg(C.first, RegState::Implicit);
2557 MIB.addReg(C.first, RegState::Implicit | RegState::Define);
2560 I->eraseFromParent();
2563 // Given a MBB and its TBB, find the FBB which was a fallthrough MBB (it may
2564 // not be a fallthrough MBB now due to layout changes). Return nullptr if the
2565 // fallthrough MBB cannot be identified.
2566 static MachineBasicBlock *getFallThroughMBB(MachineBasicBlock *MBB,
2567 MachineBasicBlock *TBB) {
2568 // Look for non-EHPad successors other than TBB. If we find exactly one, it
2569 // is the fallthrough MBB. If we find zero, then TBB is both the target MBB
2570 // and fallthrough MBB. If we find more than one, we cannot identify the
2571 // fallthrough MBB and should return nullptr.
2572 MachineBasicBlock *FallthroughBB = nullptr;
2573 for (auto SI = MBB->succ_begin(), SE = MBB->succ_end(); SI != SE; ++SI) {
2574 if ((*SI)->isEHPad() || (*SI == TBB && FallthroughBB))
2576 // Return a nullptr if we found more than one fallthrough successor.
2577 if (FallthroughBB && FallthroughBB != TBB)
2579 FallthroughBB = *SI;
2581 return FallthroughBB;
2584 bool X86InstrInfo::AnalyzeBranchImpl(
2585 MachineBasicBlock &MBB, MachineBasicBlock *&TBB, MachineBasicBlock *&FBB,
2586 SmallVectorImpl<MachineOperand> &Cond,
2587 SmallVectorImpl<MachineInstr *> &CondBranches, bool AllowModify) const {
2589 // Start from the bottom of the block and work up, examining the
2590 // terminator instructions.
2591 MachineBasicBlock::iterator I = MBB.end();
2592 MachineBasicBlock::iterator UnCondBrIter = MBB.end();
2593 while (I != MBB.begin()) {
2595 if (I->isDebugInstr())
2598 // Working from the bottom, when we see a non-terminator instruction, we're
2600 if (!isUnpredicatedTerminator(*I))
2603 // A terminator that isn't a branch can't easily be handled by this
2608 // Handle unconditional branches.
2609 if (I->getOpcode() == X86::JMP_1) {
2613 TBB = I->getOperand(0).getMBB();
2617 // If the block has any instructions after a JMP, delete them.
2618 while (std::next(I) != MBB.end())
2619 std::next(I)->eraseFromParent();
2624 // Delete the JMP if it's equivalent to a fall-through.
2625 if (MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) {
2627 I->eraseFromParent();
2629 UnCondBrIter = MBB.end();
2633 // TBB is used to indicate the unconditional destination.
2634 TBB = I->getOperand(0).getMBB();
2638 // Handle conditional branches.
2639 X86::CondCode BranchCode = X86::getCondFromBranchOpc(I->getOpcode());
2640 if (BranchCode == X86::COND_INVALID)
2641 return true; // Can't handle indirect branch.
2643 // Working from the bottom, handle the first conditional branch.
2645 MachineBasicBlock *TargetBB = I->getOperand(0).getMBB();
2646 if (AllowModify && UnCondBrIter != MBB.end() &&
2647 MBB.isLayoutSuccessor(TargetBB)) {
2648 // If we can modify the code and it ends in something like:
2656 // Then we can change this to:
2663 // Which is a bit more efficient.
2664 // We conditionally jump to the fall-through block.
2665 BranchCode = GetOppositeBranchCondition(BranchCode);
2666 unsigned JNCC = GetCondBranchFromCond(BranchCode);
2667 MachineBasicBlock::iterator OldInst = I;
2669 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(JNCC))
2670 .addMBB(UnCondBrIter->getOperand(0).getMBB());
2671 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(X86::JMP_1))
2674 OldInst->eraseFromParent();
2675 UnCondBrIter->eraseFromParent();
2677 // Restart the analysis.
2678 UnCondBrIter = MBB.end();
2684 TBB = I->getOperand(0).getMBB();
2685 Cond.push_back(MachineOperand::CreateImm(BranchCode));
2686 CondBranches.push_back(&*I);
2690 // Handle subsequent conditional branches. Only handle the case where all
2691 // conditional branches branch to the same destination and their condition
2692 // opcodes fit one of the special multi-branch idioms.
2693 assert(Cond.size() == 1);
2696 // If the conditions are the same, we can leave them alone.
2697 X86::CondCode OldBranchCode = (X86::CondCode)Cond[0].getImm();
2698 auto NewTBB = I->getOperand(0).getMBB();
2699 if (OldBranchCode == BranchCode && TBB == NewTBB)
2702 // If they differ, see if they fit one of the known patterns. Theoretically,
2703 // we could handle more patterns here, but we shouldn't expect to see them
2704 // if instruction selection has done a reasonable job.
2705 if (TBB == NewTBB &&
2706 ((OldBranchCode == X86::COND_P && BranchCode == X86::COND_NE) ||
2707 (OldBranchCode == X86::COND_NE && BranchCode == X86::COND_P))) {
2708 BranchCode = X86::COND_NE_OR_P;
2709 } else if ((OldBranchCode == X86::COND_NP && BranchCode == X86::COND_NE) ||
2710 (OldBranchCode == X86::COND_E && BranchCode == X86::COND_P)) {
2711 if (NewTBB != (FBB ? FBB : getFallThroughMBB(&MBB, TBB)))
2714 // X86::COND_E_AND_NP usually has two different branch destinations.
2722 // Here this condition branches to B2 only if NP && E. It has another
2731 // Similarly it branches to B2 only if E && NP. That is why this condition
2732 // is named with COND_E_AND_NP.
2733 BranchCode = X86::COND_E_AND_NP;
2737 // Update the MachineOperand.
2738 Cond[0].setImm(BranchCode);
2739 CondBranches.push_back(&*I);
2745 bool X86InstrInfo::analyzeBranch(MachineBasicBlock &MBB,
2746 MachineBasicBlock *&TBB,
2747 MachineBasicBlock *&FBB,
2748 SmallVectorImpl<MachineOperand> &Cond,
2749 bool AllowModify) const {
2750 SmallVector<MachineInstr *, 4> CondBranches;
2751 return AnalyzeBranchImpl(MBB, TBB, FBB, Cond, CondBranches, AllowModify);
2754 bool X86InstrInfo::analyzeBranchPredicate(MachineBasicBlock &MBB,
2755 MachineBranchPredicate &MBP,
2756 bool AllowModify) const {
2757 using namespace std::placeholders;
2759 SmallVector<MachineOperand, 4> Cond;
2760 SmallVector<MachineInstr *, 4> CondBranches;
2761 if (AnalyzeBranchImpl(MBB, MBP.TrueDest, MBP.FalseDest, Cond, CondBranches,
2765 if (Cond.size() != 1)
2768 assert(MBP.TrueDest && "expected!");
2771 MBP.FalseDest = MBB.getNextNode();
2773 const TargetRegisterInfo *TRI = &getRegisterInfo();
2775 MachineInstr *ConditionDef = nullptr;
2776 bool SingleUseCondition = true;
2778 for (auto I = std::next(MBB.rbegin()), E = MBB.rend(); I != E; ++I) {
2779 if (I->modifiesRegister(X86::EFLAGS, TRI)) {
2784 if (I->readsRegister(X86::EFLAGS, TRI))
2785 SingleUseCondition = false;
2791 if (SingleUseCondition) {
2792 for (auto *Succ : MBB.successors())
2793 if (Succ->isLiveIn(X86::EFLAGS))
2794 SingleUseCondition = false;
2797 MBP.ConditionDef = ConditionDef;
2798 MBP.SingleUseCondition = SingleUseCondition;
2800 // Currently we only recognize the simple pattern:
2805 const unsigned TestOpcode =
2806 Subtarget.is64Bit() ? X86::TEST64rr : X86::TEST32rr;
2808 if (ConditionDef->getOpcode() == TestOpcode &&
2809 ConditionDef->getNumOperands() == 3 &&
2810 ConditionDef->getOperand(0).isIdenticalTo(ConditionDef->getOperand(1)) &&
2811 (Cond[0].getImm() == X86::COND_NE || Cond[0].getImm() == X86::COND_E)) {
2812 MBP.LHS = ConditionDef->getOperand(0);
2813 MBP.RHS = MachineOperand::CreateImm(0);
2814 MBP.Predicate = Cond[0].getImm() == X86::COND_NE
2815 ? MachineBranchPredicate::PRED_NE
2816 : MachineBranchPredicate::PRED_EQ;
2823 unsigned X86InstrInfo::removeBranch(MachineBasicBlock &MBB,
2824 int *BytesRemoved) const {
2825 assert(!BytesRemoved && "code size not handled");
2827 MachineBasicBlock::iterator I = MBB.end();
2830 while (I != MBB.begin()) {
2832 if (I->isDebugInstr())
2834 if (I->getOpcode() != X86::JMP_1 &&
2835 X86::getCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
2837 // Remove the branch.
2838 I->eraseFromParent();
2846 unsigned X86InstrInfo::insertBranch(MachineBasicBlock &MBB,
2847 MachineBasicBlock *TBB,
2848 MachineBasicBlock *FBB,
2849 ArrayRef<MachineOperand> Cond,
2851 int *BytesAdded) const {
2852 // Shouldn't be a fall through.
2853 assert(TBB && "insertBranch must not be told to insert a fallthrough");
2854 assert((Cond.size() == 1 || Cond.size() == 0) &&
2855 "X86 branch conditions have one component!");
2856 assert(!BytesAdded && "code size not handled");
2859 // Unconditional branch?
2860 assert(!FBB && "Unconditional branch with multiple successors!");
2861 BuildMI(&MBB, DL, get(X86::JMP_1)).addMBB(TBB);
2865 // If FBB is null, it is implied to be a fall-through block.
2866 bool FallThru = FBB == nullptr;
2868 // Conditional branch.
2870 X86::CondCode CC = (X86::CondCode)Cond[0].getImm();
2872 case X86::COND_NE_OR_P:
2873 // Synthesize NE_OR_P with two branches.
2874 BuildMI(&MBB, DL, get(X86::JNE_1)).addMBB(TBB);
2876 BuildMI(&MBB, DL, get(X86::JP_1)).addMBB(TBB);
2879 case X86::COND_E_AND_NP:
2880 // Use the next block of MBB as FBB if it is null.
2881 if (FBB == nullptr) {
2882 FBB = getFallThroughMBB(&MBB, TBB);
2883 assert(FBB && "MBB cannot be the last block in function when the false "
2884 "body is a fall-through.");
2886 // Synthesize COND_E_AND_NP with two branches.
2887 BuildMI(&MBB, DL, get(X86::JNE_1)).addMBB(FBB);
2889 BuildMI(&MBB, DL, get(X86::JNP_1)).addMBB(TBB);
2893 unsigned Opc = GetCondBranchFromCond(CC);
2894 BuildMI(&MBB, DL, get(Opc)).addMBB(TBB);
2899 // Two-way Conditional branch. Insert the second branch.
2900 BuildMI(&MBB, DL, get(X86::JMP_1)).addMBB(FBB);
2907 canInsertSelect(const MachineBasicBlock &MBB,
2908 ArrayRef<MachineOperand> Cond,
2909 unsigned TrueReg, unsigned FalseReg,
2910 int &CondCycles, int &TrueCycles, int &FalseCycles) const {
2911 // Not all subtargets have cmov instructions.
2912 if (!Subtarget.hasCMov())
2914 if (Cond.size() != 1)
2916 // We cannot do the composite conditions, at least not in SSA form.
2917 if ((X86::CondCode)Cond[0].getImm() > X86::COND_S)
2920 // Check register classes.
2921 const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2922 const TargetRegisterClass *RC =
2923 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg));
2927 // We have cmov instructions for 16, 32, and 64 bit general purpose registers.
2928 if (X86::GR16RegClass.hasSubClassEq(RC) ||
2929 X86::GR32RegClass.hasSubClassEq(RC) ||
2930 X86::GR64RegClass.hasSubClassEq(RC)) {
2931 // This latency applies to Pentium M, Merom, Wolfdale, Nehalem, and Sandy
2932 // Bridge. Probably Ivy Bridge as well.
2939 // Can't do vectors.
2943 void X86InstrInfo::insertSelect(MachineBasicBlock &MBB,
2944 MachineBasicBlock::iterator I,
2945 const DebugLoc &DL, unsigned DstReg,
2946 ArrayRef<MachineOperand> Cond, unsigned TrueReg,
2947 unsigned FalseReg) const {
2948 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2949 const TargetRegisterInfo &TRI = *MRI.getTargetRegisterInfo();
2950 const TargetRegisterClass &RC = *MRI.getRegClass(DstReg);
2951 assert(Cond.size() == 1 && "Invalid Cond array");
2952 unsigned Opc = getCMovFromCond((X86::CondCode)Cond[0].getImm(),
2953 TRI.getRegSizeInBits(RC) / 8,
2954 false /*HasMemoryOperand*/);
2955 BuildMI(MBB, I, DL, get(Opc), DstReg).addReg(FalseReg).addReg(TrueReg);
2958 /// Test if the given register is a physical h register.
2959 static bool isHReg(unsigned Reg) {
2960 return X86::GR8_ABCD_HRegClass.contains(Reg);
2963 // Try and copy between VR128/VR64 and GR64 registers.
2964 static unsigned CopyToFromAsymmetricReg(unsigned DestReg, unsigned SrcReg,
2965 const X86Subtarget &Subtarget) {
2966 bool HasAVX = Subtarget.hasAVX();
2967 bool HasAVX512 = Subtarget.hasAVX512();
2969 // SrcReg(MaskReg) -> DestReg(GR64)
2970 // SrcReg(MaskReg) -> DestReg(GR32)
2972 // All KMASK RegClasses hold the same k registers, can be tested against anyone.
2973 if (X86::VK16RegClass.contains(SrcReg)) {
2974 if (X86::GR64RegClass.contains(DestReg)) {
2975 assert(Subtarget.hasBWI());
2976 return X86::KMOVQrk;
2978 if (X86::GR32RegClass.contains(DestReg))
2979 return Subtarget.hasBWI() ? X86::KMOVDrk : X86::KMOVWrk;
2982 // SrcReg(GR64) -> DestReg(MaskReg)
2983 // SrcReg(GR32) -> DestReg(MaskReg)
2985 // All KMASK RegClasses hold the same k registers, can be tested against anyone.
2986 if (X86::VK16RegClass.contains(DestReg)) {
2987 if (X86::GR64RegClass.contains(SrcReg)) {
2988 assert(Subtarget.hasBWI());
2989 return X86::KMOVQkr;
2991 if (X86::GR32RegClass.contains(SrcReg))
2992 return Subtarget.hasBWI() ? X86::KMOVDkr : X86::KMOVWkr;
2996 // SrcReg(VR128) -> DestReg(GR64)
2997 // SrcReg(VR64) -> DestReg(GR64)
2998 // SrcReg(GR64) -> DestReg(VR128)
2999 // SrcReg(GR64) -> DestReg(VR64)
3001 if (X86::GR64RegClass.contains(DestReg)) {
3002 if (X86::VR128XRegClass.contains(SrcReg))
3003 // Copy from a VR128 register to a GR64 register.
3004 return HasAVX512 ? X86::VMOVPQIto64Zrr :
3005 HasAVX ? X86::VMOVPQIto64rr :
3007 if (X86::VR64RegClass.contains(SrcReg))
3008 // Copy from a VR64 register to a GR64 register.
3009 return X86::MMX_MOVD64from64rr;
3010 } else if (X86::GR64RegClass.contains(SrcReg)) {
3011 // Copy from a GR64 register to a VR128 register.
3012 if (X86::VR128XRegClass.contains(DestReg))
3013 return HasAVX512 ? X86::VMOV64toPQIZrr :
3014 HasAVX ? X86::VMOV64toPQIrr :
3016 // Copy from a GR64 register to a VR64 register.
3017 if (X86::VR64RegClass.contains(DestReg))
3018 return X86::MMX_MOVD64to64rr;
3021 // SrcReg(FR32) -> DestReg(GR32)
3022 // SrcReg(GR32) -> DestReg(FR32)
3024 if (X86::GR32RegClass.contains(DestReg) &&
3025 X86::FR32XRegClass.contains(SrcReg))
3026 // Copy from a FR32 register to a GR32 register.
3027 return HasAVX512 ? X86::VMOVSS2DIZrr :
3028 HasAVX ? X86::VMOVSS2DIrr :
3031 if (X86::FR32XRegClass.contains(DestReg) &&
3032 X86::GR32RegClass.contains(SrcReg))
3033 // Copy from a GR32 register to a FR32 register.
3034 return HasAVX512 ? X86::VMOVDI2SSZrr :
3035 HasAVX ? X86::VMOVDI2SSrr :
3040 void X86InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
3041 MachineBasicBlock::iterator MI,
3042 const DebugLoc &DL, unsigned DestReg,
3043 unsigned SrcReg, bool KillSrc) const {
3044 // First deal with the normal symmetric copies.
3045 bool HasAVX = Subtarget.hasAVX();
3046 bool HasVLX = Subtarget.hasVLX();
3048 if (X86::GR64RegClass.contains(DestReg, SrcReg))
3050 else if (X86::GR32RegClass.contains(DestReg, SrcReg))
3052 else if (X86::GR16RegClass.contains(DestReg, SrcReg))
3054 else if (X86::GR8RegClass.contains(DestReg, SrcReg)) {
3055 // Copying to or from a physical H register on x86-64 requires a NOREX
3056 // move. Otherwise use a normal move.
3057 if ((isHReg(DestReg) || isHReg(SrcReg)) &&
3058 Subtarget.is64Bit()) {
3059 Opc = X86::MOV8rr_NOREX;
3060 // Both operands must be encodable without an REX prefix.
3061 assert(X86::GR8_NOREXRegClass.contains(SrcReg, DestReg) &&
3062 "8-bit H register can not be copied outside GR8_NOREX");
3066 else if (X86::VR64RegClass.contains(DestReg, SrcReg))
3067 Opc = X86::MMX_MOVQ64rr;
3068 else if (X86::VR128XRegClass.contains(DestReg, SrcReg)) {
3070 Opc = X86::VMOVAPSZ128rr;
3071 else if (X86::VR128RegClass.contains(DestReg, SrcReg))
3072 Opc = HasAVX ? X86::VMOVAPSrr : X86::MOVAPSrr;
3074 // If this an extended register and we don't have VLX we need to use a
3076 Opc = X86::VMOVAPSZrr;
3077 const TargetRegisterInfo *TRI = &getRegisterInfo();
3078 DestReg = TRI->getMatchingSuperReg(DestReg, X86::sub_xmm,
3079 &X86::VR512RegClass);
3080 SrcReg = TRI->getMatchingSuperReg(SrcReg, X86::sub_xmm,
3081 &X86::VR512RegClass);
3083 } else if (X86::VR256XRegClass.contains(DestReg, SrcReg)) {
3085 Opc = X86::VMOVAPSZ256rr;
3086 else if (X86::VR256RegClass.contains(DestReg, SrcReg))
3087 Opc = X86::VMOVAPSYrr;
3089 // If this an extended register and we don't have VLX we need to use a
3091 Opc = X86::VMOVAPSZrr;
3092 const TargetRegisterInfo *TRI = &getRegisterInfo();
3093 DestReg = TRI->getMatchingSuperReg(DestReg, X86::sub_ymm,
3094 &X86::VR512RegClass);
3095 SrcReg = TRI->getMatchingSuperReg(SrcReg, X86::sub_ymm,
3096 &X86::VR512RegClass);
3098 } else if (X86::VR512RegClass.contains(DestReg, SrcReg))
3099 Opc = X86::VMOVAPSZrr;
3100 // All KMASK RegClasses hold the same k registers, can be tested against anyone.
3101 else if (X86::VK16RegClass.contains(DestReg, SrcReg))
3102 Opc = Subtarget.hasBWI() ? X86::KMOVQkk : X86::KMOVWkk;
3104 Opc = CopyToFromAsymmetricReg(DestReg, SrcReg, Subtarget);
3107 BuildMI(MBB, MI, DL, get(Opc), DestReg)
3108 .addReg(SrcReg, getKillRegState(KillSrc));
3112 if (SrcReg == X86::EFLAGS || DestReg == X86::EFLAGS) {
3113 // FIXME: We use a fatal error here because historically LLVM has tried
3114 // lower some of these physreg copies and we want to ensure we get
3115 // reasonable bug reports if someone encounters a case no other testing
3116 // found. This path should be removed after the LLVM 7 release.
3117 report_fatal_error("Unable to copy EFLAGS physical register!");
3120 LLVM_DEBUG(dbgs() << "Cannot copy " << RI.getName(SrcReg) << " to "
3121 << RI.getName(DestReg) << '\n');
3122 report_fatal_error("Cannot emit physreg copy instruction");
3125 bool X86InstrInfo::isCopyInstrImpl(const MachineInstr &MI,
3126 const MachineOperand *&Src,
3127 const MachineOperand *&Dest) const {
3128 if (MI.isMoveReg()) {
3129 Dest = &MI.getOperand(0);
3130 Src = &MI.getOperand(1);
3136 static unsigned getLoadStoreRegOpcode(unsigned Reg,
3137 const TargetRegisterClass *RC,
3138 bool isStackAligned,
3139 const X86Subtarget &STI,
3141 bool HasAVX = STI.hasAVX();
3142 bool HasAVX512 = STI.hasAVX512();
3143 bool HasVLX = STI.hasVLX();
3145 switch (STI.getRegisterInfo()->getSpillSize(*RC)) {
3147 llvm_unreachable("Unknown spill size");
3149 assert(X86::GR8RegClass.hasSubClassEq(RC) && "Unknown 1-byte regclass");
3151 // Copying to or from a physical H register on x86-64 requires a NOREX
3152 // move. Otherwise use a normal move.
3153 if (isHReg(Reg) || X86::GR8_ABCD_HRegClass.hasSubClassEq(RC))
3154 return load ? X86::MOV8rm_NOREX : X86::MOV8mr_NOREX;
3155 return load ? X86::MOV8rm : X86::MOV8mr;
3157 if (X86::VK16RegClass.hasSubClassEq(RC))
3158 return load ? X86::KMOVWkm : X86::KMOVWmk;
3159 assert(X86::GR16RegClass.hasSubClassEq(RC) && "Unknown 2-byte regclass");
3160 return load ? X86::MOV16rm : X86::MOV16mr;
3162 if (X86::GR32RegClass.hasSubClassEq(RC))
3163 return load ? X86::MOV32rm : X86::MOV32mr;
3164 if (X86::FR32XRegClass.hasSubClassEq(RC))
3166 (HasAVX512 ? X86::VMOVSSZrm : HasAVX ? X86::VMOVSSrm : X86::MOVSSrm) :
3167 (HasAVX512 ? X86::VMOVSSZmr : HasAVX ? X86::VMOVSSmr : X86::MOVSSmr);
3168 if (X86::RFP32RegClass.hasSubClassEq(RC))
3169 return load ? X86::LD_Fp32m : X86::ST_Fp32m;
3170 if (X86::VK32RegClass.hasSubClassEq(RC)) {
3171 assert(STI.hasBWI() && "KMOVD requires BWI");
3172 return load ? X86::KMOVDkm : X86::KMOVDmk;
3174 llvm_unreachable("Unknown 4-byte regclass");
3176 if (X86::GR64RegClass.hasSubClassEq(RC))
3177 return load ? X86::MOV64rm : X86::MOV64mr;
3178 if (X86::FR64XRegClass.hasSubClassEq(RC))
3180 (HasAVX512 ? X86::VMOVSDZrm : HasAVX ? X86::VMOVSDrm : X86::MOVSDrm) :
3181 (HasAVX512 ? X86::VMOVSDZmr : HasAVX ? X86::VMOVSDmr : X86::MOVSDmr);
3182 if (X86::VR64RegClass.hasSubClassEq(RC))
3183 return load ? X86::MMX_MOVQ64rm : X86::MMX_MOVQ64mr;
3184 if (X86::RFP64RegClass.hasSubClassEq(RC))
3185 return load ? X86::LD_Fp64m : X86::ST_Fp64m;
3186 if (X86::VK64RegClass.hasSubClassEq(RC)) {
3187 assert(STI.hasBWI() && "KMOVQ requires BWI");
3188 return load ? X86::KMOVQkm : X86::KMOVQmk;
3190 llvm_unreachable("Unknown 8-byte regclass");
3192 assert(X86::RFP80RegClass.hasSubClassEq(RC) && "Unknown 10-byte regclass");
3193 return load ? X86::LD_Fp80m : X86::ST_FpP80m;
3195 if (X86::VR128XRegClass.hasSubClassEq(RC)) {
3196 // If stack is realigned we can use aligned stores.
3199 (HasVLX ? X86::VMOVAPSZ128rm :
3200 HasAVX512 ? X86::VMOVAPSZ128rm_NOVLX :
3201 HasAVX ? X86::VMOVAPSrm :
3203 (HasVLX ? X86::VMOVAPSZ128mr :
3204 HasAVX512 ? X86::VMOVAPSZ128mr_NOVLX :
3205 HasAVX ? X86::VMOVAPSmr :
3209 (HasVLX ? X86::VMOVUPSZ128rm :
3210 HasAVX512 ? X86::VMOVUPSZ128rm_NOVLX :
3211 HasAVX ? X86::VMOVUPSrm :
3213 (HasVLX ? X86::VMOVUPSZ128mr :
3214 HasAVX512 ? X86::VMOVUPSZ128mr_NOVLX :
3215 HasAVX ? X86::VMOVUPSmr :
3218 if (X86::BNDRRegClass.hasSubClassEq(RC)) {
3220 return load ? X86::BNDMOV64rm : X86::BNDMOV64mr;
3222 return load ? X86::BNDMOV32rm : X86::BNDMOV32mr;
3224 llvm_unreachable("Unknown 16-byte regclass");
3227 assert(X86::VR256XRegClass.hasSubClassEq(RC) && "Unknown 32-byte regclass");
3228 // If stack is realigned we can use aligned stores.
3231 (HasVLX ? X86::VMOVAPSZ256rm :
3232 HasAVX512 ? X86::VMOVAPSZ256rm_NOVLX :
3234 (HasVLX ? X86::VMOVAPSZ256mr :
3235 HasAVX512 ? X86::VMOVAPSZ256mr_NOVLX :
3239 (HasVLX ? X86::VMOVUPSZ256rm :
3240 HasAVX512 ? X86::VMOVUPSZ256rm_NOVLX :
3242 (HasVLX ? X86::VMOVUPSZ256mr :
3243 HasAVX512 ? X86::VMOVUPSZ256mr_NOVLX :
3246 assert(X86::VR512RegClass.hasSubClassEq(RC) && "Unknown 64-byte regclass");
3247 assert(STI.hasAVX512() && "Using 512-bit register requires AVX512");
3249 return load ? X86::VMOVAPSZrm : X86::VMOVAPSZmr;
3251 return load ? X86::VMOVUPSZrm : X86::VMOVUPSZmr;
3255 bool X86InstrInfo::getMemOpBaseRegImmOfs(MachineInstr &MemOp, unsigned &BaseReg,
3257 const TargetRegisterInfo *TRI) const {
3258 const MCInstrDesc &Desc = MemOp.getDesc();
3259 int MemRefBegin = X86II::getMemoryOperandNo(Desc.TSFlags);
3260 if (MemRefBegin < 0)
3263 MemRefBegin += X86II::getOperandBias(Desc);
3265 MachineOperand &BaseMO = MemOp.getOperand(MemRefBegin + X86::AddrBaseReg);
3266 if (!BaseMO.isReg()) // Can be an MO_FrameIndex
3269 BaseReg = BaseMO.getReg();
3270 if (MemOp.getOperand(MemRefBegin + X86::AddrScaleAmt).getImm() != 1)
3273 if (MemOp.getOperand(MemRefBegin + X86::AddrIndexReg).getReg() !=
3277 const MachineOperand &DispMO = MemOp.getOperand(MemRefBegin + X86::AddrDisp);
3279 // Displacement can be symbolic
3280 if (!DispMO.isImm())
3283 Offset = DispMO.getImm();
3288 static unsigned getStoreRegOpcode(unsigned SrcReg,
3289 const TargetRegisterClass *RC,
3290 bool isStackAligned,
3291 const X86Subtarget &STI) {
3292 return getLoadStoreRegOpcode(SrcReg, RC, isStackAligned, STI, false);
3296 static unsigned getLoadRegOpcode(unsigned DestReg,
3297 const TargetRegisterClass *RC,
3298 bool isStackAligned,
3299 const X86Subtarget &STI) {
3300 return getLoadStoreRegOpcode(DestReg, RC, isStackAligned, STI, true);
3303 void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
3304 MachineBasicBlock::iterator MI,
3305 unsigned SrcReg, bool isKill, int FrameIdx,
3306 const TargetRegisterClass *RC,
3307 const TargetRegisterInfo *TRI) const {
3308 const MachineFunction &MF = *MBB.getParent();
3309 assert(MF.getFrameInfo().getObjectSize(FrameIdx) >= TRI->getSpillSize(*RC) &&
3310 "Stack slot too small for store");
3311 unsigned Alignment = std::max<uint32_t>(TRI->getSpillSize(*RC), 16);
3313 (Subtarget.getFrameLowering()->getStackAlignment() >= Alignment) ||
3314 RI.canRealignStack(MF);
3315 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, Subtarget);
3316 addFrameReference(BuildMI(MBB, MI, DebugLoc(), get(Opc)), FrameIdx)
3317 .addReg(SrcReg, getKillRegState(isKill));
3320 void X86InstrInfo::storeRegToAddr(
3321 MachineFunction &MF, unsigned SrcReg, bool isKill,
3322 SmallVectorImpl<MachineOperand> &Addr, const TargetRegisterClass *RC,
3323 ArrayRef<MachineMemOperand *> MMOs,
3324 SmallVectorImpl<MachineInstr *> &NewMIs) const {
3325 const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
3326 unsigned Alignment = std::max<uint32_t>(TRI.getSpillSize(*RC), 16);
3327 bool isAligned = !MMOs.empty() && MMOs.front()->getAlignment() >= Alignment;
3328 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, Subtarget);
3330 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc));
3331 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
3333 MIB.addReg(SrcReg, getKillRegState(isKill));
3334 MIB.setMemRefs(MMOs);
3335 NewMIs.push_back(MIB);
3339 void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
3340 MachineBasicBlock::iterator MI,
3341 unsigned DestReg, int FrameIdx,
3342 const TargetRegisterClass *RC,
3343 const TargetRegisterInfo *TRI) const {
3344 const MachineFunction &MF = *MBB.getParent();
3345 unsigned Alignment = std::max<uint32_t>(TRI->getSpillSize(*RC), 16);
3347 (Subtarget.getFrameLowering()->getStackAlignment() >= Alignment) ||
3348 RI.canRealignStack(MF);
3349 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, Subtarget);
3350 addFrameReference(BuildMI(MBB, MI, DebugLoc(), get(Opc), DestReg), FrameIdx);
3353 void X86InstrInfo::loadRegFromAddr(
3354 MachineFunction &MF, unsigned DestReg,
3355 SmallVectorImpl<MachineOperand> &Addr, const TargetRegisterClass *RC,
3356 ArrayRef<MachineMemOperand *> MMOs,
3357 SmallVectorImpl<MachineInstr *> &NewMIs) const {
3358 const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
3359 unsigned Alignment = std::max<uint32_t>(TRI.getSpillSize(*RC), 16);
3360 bool isAligned = !MMOs.empty() && MMOs.front()->getAlignment() >= Alignment;
3361 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, Subtarget);
3363 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc), DestReg);
3364 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
3366 MIB.setMemRefs(MMOs);
3367 NewMIs.push_back(MIB);
3370 bool X86InstrInfo::analyzeCompare(const MachineInstr &MI, unsigned &SrcReg,
3371 unsigned &SrcReg2, int &CmpMask,
3372 int &CmpValue) const {
3373 switch (MI.getOpcode()) {
3375 case X86::CMP64ri32:
3382 SrcReg = MI.getOperand(0).getReg();
3384 if (MI.getOperand(1).isImm()) {
3386 CmpValue = MI.getOperand(1).getImm();
3388 CmpMask = CmpValue = 0;
3391 // A SUB can be used to perform comparison.
3396 SrcReg = MI.getOperand(1).getReg();
3405 SrcReg = MI.getOperand(1).getReg();
3406 SrcReg2 = MI.getOperand(2).getReg();
3410 case X86::SUB64ri32:
3417 SrcReg = MI.getOperand(1).getReg();
3419 if (MI.getOperand(2).isImm()) {
3421 CmpValue = MI.getOperand(2).getImm();
3423 CmpMask = CmpValue = 0;
3430 SrcReg = MI.getOperand(0).getReg();
3431 SrcReg2 = MI.getOperand(1).getReg();
3439 SrcReg = MI.getOperand(0).getReg();
3440 if (MI.getOperand(1).getReg() != SrcReg)
3442 // Compare against zero.
3451 /// Check whether the first instruction, whose only
3452 /// purpose is to update flags, can be made redundant.
3453 /// CMPrr can be made redundant by SUBrr if the operands are the same.
3454 /// This function can be extended later on.
3455 /// SrcReg, SrcRegs: register operands for FlagI.
3456 /// ImmValue: immediate for FlagI if it takes an immediate.
3457 inline static bool isRedundantFlagInstr(MachineInstr &FlagI, unsigned SrcReg,
3458 unsigned SrcReg2, int ImmMask,
3459 int ImmValue, MachineInstr &OI) {
3460 if (((FlagI.getOpcode() == X86::CMP64rr && OI.getOpcode() == X86::SUB64rr) ||
3461 (FlagI.getOpcode() == X86::CMP32rr && OI.getOpcode() == X86::SUB32rr) ||
3462 (FlagI.getOpcode() == X86::CMP16rr && OI.getOpcode() == X86::SUB16rr) ||
3463 (FlagI.getOpcode() == X86::CMP8rr && OI.getOpcode() == X86::SUB8rr)) &&
3464 ((OI.getOperand(1).getReg() == SrcReg &&
3465 OI.getOperand(2).getReg() == SrcReg2) ||
3466 (OI.getOperand(1).getReg() == SrcReg2 &&
3467 OI.getOperand(2).getReg() == SrcReg)))
3471 ((FlagI.getOpcode() == X86::CMP64ri32 &&
3472 OI.getOpcode() == X86::SUB64ri32) ||
3473 (FlagI.getOpcode() == X86::CMP64ri8 &&
3474 OI.getOpcode() == X86::SUB64ri8) ||
3475 (FlagI.getOpcode() == X86::CMP32ri && OI.getOpcode() == X86::SUB32ri) ||
3476 (FlagI.getOpcode() == X86::CMP32ri8 &&
3477 OI.getOpcode() == X86::SUB32ri8) ||
3478 (FlagI.getOpcode() == X86::CMP16ri && OI.getOpcode() == X86::SUB16ri) ||
3479 (FlagI.getOpcode() == X86::CMP16ri8 &&
3480 OI.getOpcode() == X86::SUB16ri8) ||
3481 (FlagI.getOpcode() == X86::CMP8ri && OI.getOpcode() == X86::SUB8ri)) &&
3482 OI.getOperand(1).getReg() == SrcReg &&
3483 OI.getOperand(2).getImm() == ImmValue)
3488 /// Check whether the definition can be converted
3489 /// to remove a comparison against zero.
3490 inline static bool isDefConvertible(MachineInstr &MI) {
3491 switch (MI.getOpcode()) {
3492 default: return false;
3494 // The shift instructions only modify ZF if their shift count is non-zero.
3495 // N.B.: The processor truncates the shift count depending on the encoding.
3496 case X86::SAR8ri: case X86::SAR16ri: case X86::SAR32ri:case X86::SAR64ri:
3497 case X86::SHR8ri: case X86::SHR16ri: case X86::SHR32ri:case X86::SHR64ri:
3498 return getTruncatedShiftCount(MI, 2) != 0;
3500 // Some left shift instructions can be turned into LEA instructions but only
3501 // if their flags aren't used. Avoid transforming such instructions.
3502 case X86::SHL8ri: case X86::SHL16ri: case X86::SHL32ri:case X86::SHL64ri:{
3503 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
3504 if (isTruncatedShiftCountForLEA(ShAmt)) return false;
3508 case X86::SHRD16rri8:case X86::SHRD32rri8:case X86::SHRD64rri8:
3509 case X86::SHLD16rri8:case X86::SHLD32rri8:case X86::SHLD64rri8:
3510 return getTruncatedShiftCount(MI, 3) != 0;
3512 case X86::SUB64ri32: case X86::SUB64ri8: case X86::SUB32ri:
3513 case X86::SUB32ri8: case X86::SUB16ri: case X86::SUB16ri8:
3514 case X86::SUB8ri: case X86::SUB64rr: case X86::SUB32rr:
3515 case X86::SUB16rr: case X86::SUB8rr: case X86::SUB64rm:
3516 case X86::SUB32rm: case X86::SUB16rm: case X86::SUB8rm:
3517 case X86::DEC64r: case X86::DEC32r: case X86::DEC16r: case X86::DEC8r:
3518 case X86::ADD64ri32: case X86::ADD64ri8: case X86::ADD32ri:
3519 case X86::ADD32ri8: case X86::ADD16ri: case X86::ADD16ri8:
3520 case X86::ADD8ri: case X86::ADD64rr: case X86::ADD32rr:
3521 case X86::ADD16rr: case X86::ADD8rr: case X86::ADD64rm:
3522 case X86::ADD32rm: case X86::ADD16rm: case X86::ADD8rm:
3523 case X86::INC64r: case X86::INC32r: case X86::INC16r: case X86::INC8r:
3524 case X86::AND64ri32: case X86::AND64ri8: case X86::AND32ri:
3525 case X86::AND32ri8: case X86::AND16ri: case X86::AND16ri8:
3526 case X86::AND8ri: case X86::AND64rr: case X86::AND32rr:
3527 case X86::AND16rr: case X86::AND8rr: case X86::AND64rm:
3528 case X86::AND32rm: case X86::AND16rm: case X86::AND8rm:
3529 case X86::XOR64ri32: case X86::XOR64ri8: case X86::XOR32ri:
3530 case X86::XOR32ri8: case X86::XOR16ri: case X86::XOR16ri8:
3531 case X86::XOR8ri: case X86::XOR64rr: case X86::XOR32rr:
3532 case X86::XOR16rr: case X86::XOR8rr: case X86::XOR64rm:
3533 case X86::XOR32rm: case X86::XOR16rm: case X86::XOR8rm:
3534 case X86::OR64ri32: case X86::OR64ri8: case X86::OR32ri:
3535 case X86::OR32ri8: case X86::OR16ri: case X86::OR16ri8:
3536 case X86::OR8ri: case X86::OR64rr: case X86::OR32rr:
3537 case X86::OR16rr: case X86::OR8rr: case X86::OR64rm:
3538 case X86::OR32rm: case X86::OR16rm: case X86::OR8rm:
3539 case X86::ADC64ri32: case X86::ADC64ri8: case X86::ADC32ri:
3540 case X86::ADC32ri8: case X86::ADC16ri: case X86::ADC16ri8:
3541 case X86::ADC8ri: case X86::ADC64rr: case X86::ADC32rr:
3542 case X86::ADC16rr: case X86::ADC8rr: case X86::ADC64rm:
3543 case X86::ADC32rm: case X86::ADC16rm: case X86::ADC8rm:
3544 case X86::SBB64ri32: case X86::SBB64ri8: case X86::SBB32ri:
3545 case X86::SBB32ri8: case X86::SBB16ri: case X86::SBB16ri8:
3546 case X86::SBB8ri: case X86::SBB64rr: case X86::SBB32rr:
3547 case X86::SBB16rr: case X86::SBB8rr: case X86::SBB64rm:
3548 case X86::SBB32rm: case X86::SBB16rm: case X86::SBB8rm:
3549 case X86::NEG8r: case X86::NEG16r: case X86::NEG32r: case X86::NEG64r:
3550 case X86::SAR8r1: case X86::SAR16r1: case X86::SAR32r1:case X86::SAR64r1:
3551 case X86::SHR8r1: case X86::SHR16r1: case X86::SHR32r1:case X86::SHR64r1:
3552 case X86::SHL8r1: case X86::SHL16r1: case X86::SHL32r1:case X86::SHL64r1:
3553 case X86::ANDN32rr: case X86::ANDN32rm:
3554 case X86::ANDN64rr: case X86::ANDN64rm:
3555 case X86::BEXTR32rr: case X86::BEXTR64rr:
3556 case X86::BEXTR32rm: case X86::BEXTR64rm:
3557 case X86::BLSI32rr: case X86::BLSI32rm:
3558 case X86::BLSI64rr: case X86::BLSI64rm:
3559 case X86::BLSMSK32rr:case X86::BLSMSK32rm:
3560 case X86::BLSMSK64rr:case X86::BLSMSK64rm:
3561 case X86::BLSR32rr: case X86::BLSR32rm:
3562 case X86::BLSR64rr: case X86::BLSR64rm:
3563 case X86::BZHI32rr: case X86::BZHI32rm:
3564 case X86::BZHI64rr: case X86::BZHI64rm:
3565 case X86::LZCNT16rr: case X86::LZCNT16rm:
3566 case X86::LZCNT32rr: case X86::LZCNT32rm:
3567 case X86::LZCNT64rr: case X86::LZCNT64rm:
3568 case X86::POPCNT16rr:case X86::POPCNT16rm:
3569 case X86::POPCNT32rr:case X86::POPCNT32rm:
3570 case X86::POPCNT64rr:case X86::POPCNT64rm:
3571 case X86::TZCNT16rr: case X86::TZCNT16rm:
3572 case X86::TZCNT32rr: case X86::TZCNT32rm:
3573 case X86::TZCNT64rr: case X86::TZCNT64rm:
3574 case X86::BEXTRI32ri: case X86::BEXTRI32mi:
3575 case X86::BEXTRI64ri: case X86::BEXTRI64mi:
3576 case X86::BLCFILL32rr: case X86::BLCFILL32rm:
3577 case X86::BLCFILL64rr: case X86::BLCFILL64rm:
3578 case X86::BLCI32rr: case X86::BLCI32rm:
3579 case X86::BLCI64rr: case X86::BLCI64rm:
3580 case X86::BLCIC32rr: case X86::BLCIC32rm:
3581 case X86::BLCIC64rr: case X86::BLCIC64rm:
3582 case X86::BLCMSK32rr: case X86::BLCMSK32rm:
3583 case X86::BLCMSK64rr: case X86::BLCMSK64rm:
3584 case X86::BLCS32rr: case X86::BLCS32rm:
3585 case X86::BLCS64rr: case X86::BLCS64rm:
3586 case X86::BLSFILL32rr: case X86::BLSFILL32rm:
3587 case X86::BLSFILL64rr: case X86::BLSFILL64rm:
3588 case X86::BLSIC32rr: case X86::BLSIC32rm:
3589 case X86::BLSIC64rr: case X86::BLSIC64rm:
3594 /// Check whether the use can be converted to remove a comparison against zero.
3595 static X86::CondCode isUseDefConvertible(MachineInstr &MI) {
3596 switch (MI.getOpcode()) {
3597 default: return X86::COND_INVALID;
3598 case X86::LZCNT16rr: case X86::LZCNT16rm:
3599 case X86::LZCNT32rr: case X86::LZCNT32rm:
3600 case X86::LZCNT64rr: case X86::LZCNT64rm:
3602 case X86::POPCNT16rr:case X86::POPCNT16rm:
3603 case X86::POPCNT32rr:case X86::POPCNT32rm:
3604 case X86::POPCNT64rr:case X86::POPCNT64rm:
3606 case X86::TZCNT16rr: case X86::TZCNT16rm:
3607 case X86::TZCNT32rr: case X86::TZCNT32rm:
3608 case X86::TZCNT64rr: case X86::TZCNT64rm:
3620 /// Check if there exists an earlier instruction that
3621 /// operates on the same source operands and sets flags in the same way as
3622 /// Compare; remove Compare if possible.
3623 bool X86InstrInfo::optimizeCompareInstr(MachineInstr &CmpInstr, unsigned SrcReg,
3624 unsigned SrcReg2, int CmpMask,
3626 const MachineRegisterInfo *MRI) const {
3627 // Check whether we can replace SUB with CMP.
3628 unsigned NewOpcode = 0;
3629 switch (CmpInstr.getOpcode()) {
3631 case X86::SUB64ri32:
3646 if (!MRI->use_nodbg_empty(CmpInstr.getOperand(0).getReg()))
3648 // There is no use of the destination register, we can replace SUB with CMP.
3649 switch (CmpInstr.getOpcode()) {
3650 default: llvm_unreachable("Unreachable!");
3651 case X86::SUB64rm: NewOpcode = X86::CMP64rm; break;
3652 case X86::SUB32rm: NewOpcode = X86::CMP32rm; break;
3653 case X86::SUB16rm: NewOpcode = X86::CMP16rm; break;
3654 case X86::SUB8rm: NewOpcode = X86::CMP8rm; break;
3655 case X86::SUB64rr: NewOpcode = X86::CMP64rr; break;
3656 case X86::SUB32rr: NewOpcode = X86::CMP32rr; break;
3657 case X86::SUB16rr: NewOpcode = X86::CMP16rr; break;
3658 case X86::SUB8rr: NewOpcode = X86::CMP8rr; break;
3659 case X86::SUB64ri32: NewOpcode = X86::CMP64ri32; break;
3660 case X86::SUB64ri8: NewOpcode = X86::CMP64ri8; break;
3661 case X86::SUB32ri: NewOpcode = X86::CMP32ri; break;
3662 case X86::SUB32ri8: NewOpcode = X86::CMP32ri8; break;
3663 case X86::SUB16ri: NewOpcode = X86::CMP16ri; break;
3664 case X86::SUB16ri8: NewOpcode = X86::CMP16ri8; break;
3665 case X86::SUB8ri: NewOpcode = X86::CMP8ri; break;
3667 CmpInstr.setDesc(get(NewOpcode));
3668 CmpInstr.RemoveOperand(0);
3669 // Fall through to optimize Cmp if Cmp is CMPrr or CMPri.
3670 if (NewOpcode == X86::CMP64rm || NewOpcode == X86::CMP32rm ||
3671 NewOpcode == X86::CMP16rm || NewOpcode == X86::CMP8rm)
3676 // Get the unique definition of SrcReg.
3677 MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg);
3678 if (!MI) return false;
3680 // CmpInstr is the first instruction of the BB.
3681 MachineBasicBlock::iterator I = CmpInstr, Def = MI;
3683 // If we are comparing against zero, check whether we can use MI to update
3684 // EFLAGS. If MI is not in the same BB as CmpInstr, do not optimize.
3685 bool IsCmpZero = (CmpMask != 0 && CmpValue == 0);
3686 if (IsCmpZero && MI->getParent() != CmpInstr.getParent())
3689 // If we have a use of the source register between the def and our compare
3690 // instruction we can eliminate the compare iff the use sets EFLAGS in the
3692 bool ShouldUpdateCC = false;
3693 X86::CondCode NewCC = X86::COND_INVALID;
3694 if (IsCmpZero && !isDefConvertible(*MI)) {
3695 // Scan forward from the use until we hit the use we're looking for or the
3696 // compare instruction.
3697 for (MachineBasicBlock::iterator J = MI;; ++J) {
3698 // Do we have a convertible instruction?
3699 NewCC = isUseDefConvertible(*J);
3700 if (NewCC != X86::COND_INVALID && J->getOperand(1).isReg() &&
3701 J->getOperand(1).getReg() == SrcReg) {
3702 assert(J->definesRegister(X86::EFLAGS) && "Must be an EFLAGS def!");
3703 ShouldUpdateCC = true; // Update CC later on.
3704 // This is not a def of SrcReg, but still a def of EFLAGS. Keep going
3705 // with the new def.
3716 // We are searching for an earlier instruction that can make CmpInstr
3717 // redundant and that instruction will be saved in Sub.
3718 MachineInstr *Sub = nullptr;
3719 const TargetRegisterInfo *TRI = &getRegisterInfo();
3721 // We iterate backward, starting from the instruction before CmpInstr and
3722 // stop when reaching the definition of a source register or done with the BB.
3723 // RI points to the instruction before CmpInstr.
3724 // If the definition is in this basic block, RE points to the definition;
3725 // otherwise, RE is the rend of the basic block.
3726 MachineBasicBlock::reverse_iterator
3727 RI = ++I.getReverse(),
3728 RE = CmpInstr.getParent() == MI->getParent()
3729 ? Def.getReverse() /* points to MI */
3730 : CmpInstr.getParent()->rend();
3731 MachineInstr *Movr0Inst = nullptr;
3732 for (; RI != RE; ++RI) {
3733 MachineInstr &Instr = *RI;
3734 // Check whether CmpInstr can be made redundant by the current instruction.
3735 if (!IsCmpZero && isRedundantFlagInstr(CmpInstr, SrcReg, SrcReg2, CmpMask,
3741 if (Instr.modifiesRegister(X86::EFLAGS, TRI) ||
3742 Instr.readsRegister(X86::EFLAGS, TRI)) {
3743 // This instruction modifies or uses EFLAGS.
3745 // MOV32r0 etc. are implemented with xor which clobbers condition code.
3746 // They are safe to move up, if the definition to EFLAGS is dead and
3747 // earlier instructions do not read or write EFLAGS.
3748 if (!Movr0Inst && Instr.getOpcode() == X86::MOV32r0 &&
3749 Instr.registerDefIsDead(X86::EFLAGS, TRI)) {
3754 // We can't remove CmpInstr.
3759 // Return false if no candidates exist.
3760 if (!IsCmpZero && !Sub)
3763 bool IsSwapped = (SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 &&
3764 Sub->getOperand(2).getReg() == SrcReg);
3766 // Scan forward from the instruction after CmpInstr for uses of EFLAGS.
3767 // It is safe to remove CmpInstr if EFLAGS is redefined or killed.
3768 // If we are done with the basic block, we need to check whether EFLAGS is
3770 bool IsSafe = false;
3771 SmallVector<std::pair<MachineInstr*, unsigned /*NewOpc*/>, 4> OpsToUpdate;
3772 MachineBasicBlock::iterator E = CmpInstr.getParent()->end();
3773 for (++I; I != E; ++I) {
3774 const MachineInstr &Instr = *I;
3775 bool ModifyEFLAGS = Instr.modifiesRegister(X86::EFLAGS, TRI);
3776 bool UseEFLAGS = Instr.readsRegister(X86::EFLAGS, TRI);
3777 // We should check the usage if this instruction uses and updates EFLAGS.
3778 if (!UseEFLAGS && ModifyEFLAGS) {
3779 // It is safe to remove CmpInstr if EFLAGS is updated again.
3783 if (!UseEFLAGS && !ModifyEFLAGS)
3786 // EFLAGS is used by this instruction.
3787 X86::CondCode OldCC = X86::COND_INVALID;
3788 bool OpcIsSET = false;
3789 if (IsCmpZero || IsSwapped) {
3790 // We decode the condition code from opcode.
3791 if (Instr.isBranch())
3792 OldCC = X86::getCondFromBranchOpc(Instr.getOpcode());
3794 OldCC = X86::getCondFromSETOpc(Instr.getOpcode());
3795 if (OldCC != X86::COND_INVALID)
3798 OldCC = X86::getCondFromCMovOpc(Instr.getOpcode());
3800 if (OldCC == X86::COND_INVALID) return false;
3802 X86::CondCode ReplacementCC = X86::COND_INVALID;
3806 case X86::COND_A: case X86::COND_AE:
3807 case X86::COND_B: case X86::COND_BE:
3808 case X86::COND_G: case X86::COND_GE:
3809 case X86::COND_L: case X86::COND_LE:
3810 case X86::COND_O: case X86::COND_NO:
3811 // CF and OF are used, we can't perform this optimization.
3815 // If we're updating the condition code check if we have to reverse the
3822 ReplacementCC = NewCC;
3825 ReplacementCC = GetOppositeBranchCondition(NewCC);
3828 } else if (IsSwapped) {
3829 // If we have SUB(r1, r2) and CMP(r2, r1), the condition code needs
3830 // to be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc.
3831 // We swap the condition code and synthesize the new opcode.
3832 ReplacementCC = getSwappedCondition(OldCC);
3833 if (ReplacementCC == X86::COND_INVALID) return false;
3836 if ((ShouldUpdateCC || IsSwapped) && ReplacementCC != OldCC) {
3837 // Synthesize the new opcode.
3838 bool HasMemoryOperand = Instr.hasOneMemOperand();
3840 if (Instr.isBranch())
3841 NewOpc = GetCondBranchFromCond(ReplacementCC);
3843 NewOpc = getSETFromCond(ReplacementCC, HasMemoryOperand);
3845 unsigned DstReg = Instr.getOperand(0).getReg();
3846 const TargetRegisterClass *DstRC = MRI->getRegClass(DstReg);
3847 NewOpc = getCMovFromCond(ReplacementCC, TRI->getRegSizeInBits(*DstRC)/8,
3851 // Push the MachineInstr to OpsToUpdate.
3852 // If it is safe to remove CmpInstr, the condition code of these
3853 // instructions will be modified.
3854 OpsToUpdate.push_back(std::make_pair(&*I, NewOpc));
3856 if (ModifyEFLAGS || Instr.killsRegister(X86::EFLAGS, TRI)) {
3857 // It is safe to remove CmpInstr if EFLAGS is updated again or killed.
3863 // If EFLAGS is not killed nor re-defined, we should check whether it is
3864 // live-out. If it is live-out, do not optimize.
3865 if ((IsCmpZero || IsSwapped) && !IsSafe) {
3866 MachineBasicBlock *MBB = CmpInstr.getParent();
3867 for (MachineBasicBlock *Successor : MBB->successors())
3868 if (Successor->isLiveIn(X86::EFLAGS))
3872 // The instruction to be updated is either Sub or MI.
3873 Sub = IsCmpZero ? MI : Sub;
3874 // Move Movr0Inst to the appropriate place before Sub.
3876 // Look backwards until we find a def that doesn't use the current EFLAGS.
3878 MachineBasicBlock::reverse_iterator InsertI = Def.getReverse(),
3879 InsertE = Sub->getParent()->rend();
3880 for (; InsertI != InsertE; ++InsertI) {
3881 MachineInstr *Instr = &*InsertI;
3882 if (!Instr->readsRegister(X86::EFLAGS, TRI) &&
3883 Instr->modifiesRegister(X86::EFLAGS, TRI)) {
3884 Sub->getParent()->remove(Movr0Inst);
3885 Instr->getParent()->insert(MachineBasicBlock::iterator(Instr),
3890 if (InsertI == InsertE)
3894 // Make sure Sub instruction defines EFLAGS and mark the def live.
3895 unsigned i = 0, e = Sub->getNumOperands();
3896 for (; i != e; ++i) {
3897 MachineOperand &MO = Sub->getOperand(i);
3898 if (MO.isReg() && MO.isDef() && MO.getReg() == X86::EFLAGS) {
3899 MO.setIsDead(false);
3903 assert(i != e && "Unable to locate a def EFLAGS operand");
3905 CmpInstr.eraseFromParent();
3907 // Modify the condition code of instructions in OpsToUpdate.
3908 for (auto &Op : OpsToUpdate)
3909 Op.first->setDesc(get(Op.second));
3913 /// Try to remove the load by folding it to a register
3914 /// operand at the use. We fold the load instructions if load defines a virtual
3915 /// register, the virtual register is used once in the same BB, and the
3916 /// instructions in-between do not load or store, and have no side effects.
3917 MachineInstr *X86InstrInfo::optimizeLoadInstr(MachineInstr &MI,
3918 const MachineRegisterInfo *MRI,
3919 unsigned &FoldAsLoadDefReg,
3920 MachineInstr *&DefMI) const {
3921 // Check whether we can move DefMI here.
3922 DefMI = MRI->getVRegDef(FoldAsLoadDefReg);
3924 bool SawStore = false;
3925 if (!DefMI->isSafeToMove(nullptr, SawStore))
3928 // Collect information about virtual register operands of MI.
3929 SmallVector<unsigned, 1> SrcOperandIds;
3930 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
3931 MachineOperand &MO = MI.getOperand(i);
3934 unsigned Reg = MO.getReg();
3935 if (Reg != FoldAsLoadDefReg)
3937 // Do not fold if we have a subreg use or a def.
3938 if (MO.getSubReg() || MO.isDef())
3940 SrcOperandIds.push_back(i);
3942 if (SrcOperandIds.empty())
3945 // Check whether we can fold the def into SrcOperandId.
3946 if (MachineInstr *FoldMI = foldMemoryOperand(MI, SrcOperandIds, *DefMI)) {
3947 FoldAsLoadDefReg = 0;
3954 /// Expand a single-def pseudo instruction to a two-addr
3955 /// instruction with two undef reads of the register being defined.
3956 /// This is used for mapping:
3959 /// %xmm4 = PXORrr undef %xmm4, undef %xmm4
3961 static bool Expand2AddrUndef(MachineInstrBuilder &MIB,
3962 const MCInstrDesc &Desc) {
3963 assert(Desc.getNumOperands() == 3 && "Expected two-addr instruction.");
3964 unsigned Reg = MIB->getOperand(0).getReg();
3967 // MachineInstr::addOperand() will insert explicit operands before any
3968 // implicit operands.
3969 MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
3970 // But we don't trust that.
3971 assert(MIB->getOperand(1).getReg() == Reg &&
3972 MIB->getOperand(2).getReg() == Reg && "Misplaced operand");
3976 /// Expand a single-def pseudo instruction to a two-addr
3977 /// instruction with two %k0 reads.
3978 /// This is used for mapping:
3981 /// %k4 = KXNORrr %k0, %k0
3982 static bool Expand2AddrKreg(MachineInstrBuilder &MIB,
3983 const MCInstrDesc &Desc, unsigned Reg) {
3984 assert(Desc.getNumOperands() == 3 && "Expected two-addr instruction.");
3986 MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
3990 static bool expandMOV32r1(MachineInstrBuilder &MIB, const TargetInstrInfo &TII,
3992 MachineBasicBlock &MBB = *MIB->getParent();
3993 DebugLoc DL = MIB->getDebugLoc();
3994 unsigned Reg = MIB->getOperand(0).getReg();
3997 BuildMI(MBB, MIB.getInstr(), DL, TII.get(X86::XOR32rr), Reg)
3998 .addReg(Reg, RegState::Undef)
3999 .addReg(Reg, RegState::Undef);
4001 // Turn the pseudo into an INC or DEC.
4002 MIB->setDesc(TII.get(MinusOne ? X86::DEC32r : X86::INC32r));
4008 static bool ExpandMOVImmSExti8(MachineInstrBuilder &MIB,
4009 const TargetInstrInfo &TII,
4010 const X86Subtarget &Subtarget) {
4011 MachineBasicBlock &MBB = *MIB->getParent();
4012 DebugLoc DL = MIB->getDebugLoc();
4013 int64_t Imm = MIB->getOperand(1).getImm();
4014 assert(Imm != 0 && "Using push/pop for 0 is not efficient.");
4015 MachineBasicBlock::iterator I = MIB.getInstr();
4017 int StackAdjustment;
4019 if (Subtarget.is64Bit()) {
4020 assert(MIB->getOpcode() == X86::MOV64ImmSExti8 ||
4021 MIB->getOpcode() == X86::MOV32ImmSExti8);
4023 // Can't use push/pop lowering if the function might write to the red zone.
4024 X86MachineFunctionInfo *X86FI =
4025 MBB.getParent()->getInfo<X86MachineFunctionInfo>();
4026 if (X86FI->getUsesRedZone()) {
4027 MIB->setDesc(TII.get(MIB->getOpcode() ==
4028 X86::MOV32ImmSExti8 ? X86::MOV32ri : X86::MOV64ri));
4032 // 64-bit mode doesn't have 32-bit push/pop, so use 64-bit operations and
4033 // widen the register if necessary.
4034 StackAdjustment = 8;
4035 BuildMI(MBB, I, DL, TII.get(X86::PUSH64i8)).addImm(Imm);
4036 MIB->setDesc(TII.get(X86::POP64r));
4038 .setReg(getX86SubSuperRegister(MIB->getOperand(0).getReg(), 64));
4040 assert(MIB->getOpcode() == X86::MOV32ImmSExti8);
4041 StackAdjustment = 4;
4042 BuildMI(MBB, I, DL, TII.get(X86::PUSH32i8)).addImm(Imm);
4043 MIB->setDesc(TII.get(X86::POP32r));
4046 // Build CFI if necessary.
4047 MachineFunction &MF = *MBB.getParent();
4048 const X86FrameLowering *TFL = Subtarget.getFrameLowering();
4049 bool IsWin64Prologue = MF.getTarget().getMCAsmInfo()->usesWindowsCFI();
4050 bool NeedsDwarfCFI =
4052 (MF.getMMI().hasDebugInfo() || MF.getFunction().needsUnwindTableEntry());
4053 bool EmitCFI = !TFL->hasFP(MF) && NeedsDwarfCFI;
4055 TFL->BuildCFI(MBB, I, DL,
4056 MCCFIInstruction::createAdjustCfaOffset(nullptr, StackAdjustment));
4057 TFL->BuildCFI(MBB, std::next(I), DL,
4058 MCCFIInstruction::createAdjustCfaOffset(nullptr, -StackAdjustment));
4064 // LoadStackGuard has so far only been implemented for 64-bit MachO. Different
4065 // code sequence is needed for other targets.
4066 static void expandLoadStackGuard(MachineInstrBuilder &MIB,
4067 const TargetInstrInfo &TII) {
4068 MachineBasicBlock &MBB = *MIB->getParent();
4069 DebugLoc DL = MIB->getDebugLoc();
4070 unsigned Reg = MIB->getOperand(0).getReg();
4071 const GlobalValue *GV =
4072 cast<GlobalValue>((*MIB->memoperands_begin())->getValue());
4073 auto Flags = MachineMemOperand::MOLoad |
4074 MachineMemOperand::MODereferenceable |
4075 MachineMemOperand::MOInvariant;
4076 MachineMemOperand *MMO = MBB.getParent()->getMachineMemOperand(
4077 MachinePointerInfo::getGOT(*MBB.getParent()), Flags, 8, 8);
4078 MachineBasicBlock::iterator I = MIB.getInstr();
4080 BuildMI(MBB, I, DL, TII.get(X86::MOV64rm), Reg).addReg(X86::RIP).addImm(1)
4081 .addReg(0).addGlobalAddress(GV, 0, X86II::MO_GOTPCREL).addReg(0)
4082 .addMemOperand(MMO);
4083 MIB->setDebugLoc(DL);
4084 MIB->setDesc(TII.get(X86::MOV64rm));
4085 MIB.addReg(Reg, RegState::Kill).addImm(1).addReg(0).addImm(0).addReg(0);
4088 static bool expandXorFP(MachineInstrBuilder &MIB, const TargetInstrInfo &TII) {
4089 MachineBasicBlock &MBB = *MIB->getParent();
4090 MachineFunction &MF = *MBB.getParent();
4091 const X86Subtarget &Subtarget = MF.getSubtarget<X86Subtarget>();
4092 const X86RegisterInfo *TRI = Subtarget.getRegisterInfo();
4094 MIB->getOpcode() == X86::XOR64_FP ? X86::XOR64rr : X86::XOR32rr;
4095 MIB->setDesc(TII.get(XorOp));
4096 MIB.addReg(TRI->getFrameRegister(MF), RegState::Undef);
4100 // This is used to handle spills for 128/256-bit registers when we have AVX512,
4101 // but not VLX. If it uses an extended register we need to use an instruction
4102 // that loads the lower 128/256-bit, but is available with only AVX512F.
4103 static bool expandNOVLXLoad(MachineInstrBuilder &MIB,
4104 const TargetRegisterInfo *TRI,
4105 const MCInstrDesc &LoadDesc,
4106 const MCInstrDesc &BroadcastDesc,
4108 unsigned DestReg = MIB->getOperand(0).getReg();
4109 // Check if DestReg is XMM16-31 or YMM16-31.
4110 if (TRI->getEncodingValue(DestReg) < 16) {
4111 // We can use a normal VEX encoded load.
4112 MIB->setDesc(LoadDesc);
4114 // Use a 128/256-bit VBROADCAST instruction.
4115 MIB->setDesc(BroadcastDesc);
4116 // Change the destination to a 512-bit register.
4117 DestReg = TRI->getMatchingSuperReg(DestReg, SubIdx, &X86::VR512RegClass);
4118 MIB->getOperand(0).setReg(DestReg);
4123 // This is used to handle spills for 128/256-bit registers when we have AVX512,
4124 // but not VLX. If it uses an extended register we need to use an instruction
4125 // that stores the lower 128/256-bit, but is available with only AVX512F.
4126 static bool expandNOVLXStore(MachineInstrBuilder &MIB,
4127 const TargetRegisterInfo *TRI,
4128 const MCInstrDesc &StoreDesc,
4129 const MCInstrDesc &ExtractDesc,
4131 unsigned SrcReg = MIB->getOperand(X86::AddrNumOperands).getReg();
4132 // Check if DestReg is XMM16-31 or YMM16-31.
4133 if (TRI->getEncodingValue(SrcReg) < 16) {
4134 // We can use a normal VEX encoded store.
4135 MIB->setDesc(StoreDesc);
4137 // Use a VEXTRACTF instruction.
4138 MIB->setDesc(ExtractDesc);
4139 // Change the destination to a 512-bit register.
4140 SrcReg = TRI->getMatchingSuperReg(SrcReg, SubIdx, &X86::VR512RegClass);
4141 MIB->getOperand(X86::AddrNumOperands).setReg(SrcReg);
4142 MIB.addImm(0x0); // Append immediate to extract from the lower bits.
4147 bool X86InstrInfo::expandPostRAPseudo(MachineInstr &MI) const {
4148 bool HasAVX = Subtarget.hasAVX();
4149 MachineInstrBuilder MIB(*MI.getParent()->getParent(), MI);
4150 switch (MI.getOpcode()) {
4152 return Expand2AddrUndef(MIB, get(X86::XOR32rr));
4154 return expandMOV32r1(MIB, *this, /*MinusOne=*/ false);
4156 return expandMOV32r1(MIB, *this, /*MinusOne=*/ true);
4157 case X86::MOV32ImmSExti8:
4158 case X86::MOV64ImmSExti8:
4159 return ExpandMOVImmSExti8(MIB, *this, Subtarget);
4161 return Expand2AddrUndef(MIB, get(X86::SBB8rr));
4162 case X86::SETB_C16r:
4163 return Expand2AddrUndef(MIB, get(X86::SBB16rr));
4164 case X86::SETB_C32r:
4165 return Expand2AddrUndef(MIB, get(X86::SBB32rr));
4166 case X86::SETB_C64r:
4167 return Expand2AddrUndef(MIB, get(X86::SBB64rr));
4169 return Expand2AddrUndef(MIB, get(X86::MMX_PXORirr));
4173 return Expand2AddrUndef(MIB, get(HasAVX ? X86::VXORPSrr : X86::XORPSrr));
4174 case X86::AVX_SET0: {
4175 assert(HasAVX && "AVX not supported");
4176 const TargetRegisterInfo *TRI = &getRegisterInfo();
4177 unsigned SrcReg = MIB->getOperand(0).getReg();
4178 unsigned XReg = TRI->getSubReg(SrcReg, X86::sub_xmm);
4179 MIB->getOperand(0).setReg(XReg);
4180 Expand2AddrUndef(MIB, get(X86::VXORPSrr));
4181 MIB.addReg(SrcReg, RegState::ImplicitDefine);
4184 case X86::AVX512_128_SET0:
4185 case X86::AVX512_FsFLD0SS:
4186 case X86::AVX512_FsFLD0SD: {
4187 bool HasVLX = Subtarget.hasVLX();
4188 unsigned SrcReg = MIB->getOperand(0).getReg();
4189 const TargetRegisterInfo *TRI = &getRegisterInfo();
4190 if (HasVLX || TRI->getEncodingValue(SrcReg) < 16)
4191 return Expand2AddrUndef(MIB,
4192 get(HasVLX ? X86::VPXORDZ128rr : X86::VXORPSrr));
4193 // Extended register without VLX. Use a larger XOR.
4195 TRI->getMatchingSuperReg(SrcReg, X86::sub_xmm, &X86::VR512RegClass);
4196 MIB->getOperand(0).setReg(SrcReg);
4197 return Expand2AddrUndef(MIB, get(X86::VPXORDZrr));
4199 case X86::AVX512_256_SET0:
4200 case X86::AVX512_512_SET0: {
4201 bool HasVLX = Subtarget.hasVLX();
4202 unsigned SrcReg = MIB->getOperand(0).getReg();
4203 const TargetRegisterInfo *TRI = &getRegisterInfo();
4204 if (HasVLX || TRI->getEncodingValue(SrcReg) < 16) {
4205 unsigned XReg = TRI->getSubReg(SrcReg, X86::sub_xmm);
4206 MIB->getOperand(0).setReg(XReg);
4207 Expand2AddrUndef(MIB,
4208 get(HasVLX ? X86::VPXORDZ128rr : X86::VXORPSrr));
4209 MIB.addReg(SrcReg, RegState::ImplicitDefine);
4212 return Expand2AddrUndef(MIB, get(X86::VPXORDZrr));
4214 case X86::V_SETALLONES:
4215 return Expand2AddrUndef(MIB, get(HasAVX ? X86::VPCMPEQDrr : X86::PCMPEQDrr));
4216 case X86::AVX2_SETALLONES:
4217 return Expand2AddrUndef(MIB, get(X86::VPCMPEQDYrr));
4218 case X86::AVX1_SETALLONES: {
4219 unsigned Reg = MIB->getOperand(0).getReg();
4220 // VCMPPSYrri with an immediate 0xf should produce VCMPTRUEPS.
4221 MIB->setDesc(get(X86::VCMPPSYrri));
4222 MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef).addImm(0xf);
4225 case X86::AVX512_512_SETALLONES: {
4226 unsigned Reg = MIB->getOperand(0).getReg();
4227 MIB->setDesc(get(X86::VPTERNLOGDZrri));
4228 // VPTERNLOGD needs 3 register inputs and an immediate.
4229 // 0xff will return 1s for any input.
4230 MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef)
4231 .addReg(Reg, RegState::Undef).addImm(0xff);
4234 case X86::AVX512_512_SEXT_MASK_32:
4235 case X86::AVX512_512_SEXT_MASK_64: {
4236 unsigned Reg = MIB->getOperand(0).getReg();
4237 unsigned MaskReg = MIB->getOperand(1).getReg();
4238 unsigned MaskState = getRegState(MIB->getOperand(1));
4239 unsigned Opc = (MI.getOpcode() == X86::AVX512_512_SEXT_MASK_64) ?
4240 X86::VPTERNLOGQZrrikz : X86::VPTERNLOGDZrrikz;
4241 MI.RemoveOperand(1);
4242 MIB->setDesc(get(Opc));
4243 // VPTERNLOG needs 3 register inputs and an immediate.
4244 // 0xff will return 1s for any input.
4245 MIB.addReg(Reg, RegState::Undef).addReg(MaskReg, MaskState)
4246 .addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef).addImm(0xff);
4249 case X86::VMOVAPSZ128rm_NOVLX:
4250 return expandNOVLXLoad(MIB, &getRegisterInfo(), get(X86::VMOVAPSrm),
4251 get(X86::VBROADCASTF32X4rm), X86::sub_xmm);
4252 case X86::VMOVUPSZ128rm_NOVLX:
4253 return expandNOVLXLoad(MIB, &getRegisterInfo(), get(X86::VMOVUPSrm),
4254 get(X86::VBROADCASTF32X4rm), X86::sub_xmm);
4255 case X86::VMOVAPSZ256rm_NOVLX:
4256 return expandNOVLXLoad(MIB, &getRegisterInfo(), get(X86::VMOVAPSYrm),
4257 get(X86::VBROADCASTF64X4rm), X86::sub_ymm);
4258 case X86::VMOVUPSZ256rm_NOVLX:
4259 return expandNOVLXLoad(MIB, &getRegisterInfo(), get(X86::VMOVUPSYrm),
4260 get(X86::VBROADCASTF64X4rm), X86::sub_ymm);
4261 case X86::VMOVAPSZ128mr_NOVLX:
4262 return expandNOVLXStore(MIB, &getRegisterInfo(), get(X86::VMOVAPSmr),
4263 get(X86::VEXTRACTF32x4Zmr), X86::sub_xmm);
4264 case X86::VMOVUPSZ128mr_NOVLX:
4265 return expandNOVLXStore(MIB, &getRegisterInfo(), get(X86::VMOVUPSmr),
4266 get(X86::VEXTRACTF32x4Zmr), X86::sub_xmm);
4267 case X86::VMOVAPSZ256mr_NOVLX:
4268 return expandNOVLXStore(MIB, &getRegisterInfo(), get(X86::VMOVAPSYmr),
4269 get(X86::VEXTRACTF64x4Zmr), X86::sub_ymm);
4270 case X86::VMOVUPSZ256mr_NOVLX:
4271 return expandNOVLXStore(MIB, &getRegisterInfo(), get(X86::VMOVUPSYmr),
4272 get(X86::VEXTRACTF64x4Zmr), X86::sub_ymm);
4273 case X86::MOV32ri64: {
4274 unsigned Reg = MIB->getOperand(0).getReg();
4275 unsigned Reg32 = RI.getSubReg(Reg, X86::sub_32bit);
4276 MI.setDesc(get(X86::MOV32ri));
4277 MIB->getOperand(0).setReg(Reg32);
4278 MIB.addReg(Reg, RegState::ImplicitDefine);
4282 // KNL does not recognize dependency-breaking idioms for mask registers,
4283 // so kxnor %k1, %k1, %k2 has a RAW dependence on %k1.
4284 // Using %k0 as the undef input register is a performance heuristic based
4285 // on the assumption that %k0 is used less frequently than the other mask
4286 // registers, since it is not usable as a write mask.
4287 // FIXME: A more advanced approach would be to choose the best input mask
4288 // register based on context.
4289 case X86::KSET0W: return Expand2AddrKreg(MIB, get(X86::KXORWrr), X86::K0);
4290 case X86::KSET0D: return Expand2AddrKreg(MIB, get(X86::KXORDrr), X86::K0);
4291 case X86::KSET0Q: return Expand2AddrKreg(MIB, get(X86::KXORQrr), X86::K0);
4292 case X86::KSET1W: return Expand2AddrKreg(MIB, get(X86::KXNORWrr), X86::K0);
4293 case X86::KSET1D: return Expand2AddrKreg(MIB, get(X86::KXNORDrr), X86::K0);
4294 case X86::KSET1Q: return Expand2AddrKreg(MIB, get(X86::KXNORQrr), X86::K0);
4295 case TargetOpcode::LOAD_STACK_GUARD:
4296 expandLoadStackGuard(MIB, *this);
4300 return expandXorFP(MIB, *this);
4305 /// Return true for all instructions that only update
4306 /// the first 32 or 64-bits of the destination register and leave the rest
4307 /// unmodified. This can be used to avoid folding loads if the instructions
4308 /// only update part of the destination register, and the non-updated part is
4309 /// not needed. e.g. cvtss2sd, sqrtss. Unfolding the load from these
4310 /// instructions breaks the partial register dependency and it can improve
4311 /// performance. e.g.:
4313 /// movss (%rdi), %xmm0
4314 /// cvtss2sd %xmm0, %xmm0
4317 /// cvtss2sd (%rdi), %xmm0
4319 /// FIXME: This should be turned into a TSFlags.
4321 static bool hasPartialRegUpdate(unsigned Opcode,
4322 const X86Subtarget &Subtarget) {
4324 case X86::CVTSI2SSrr:
4325 case X86::CVTSI2SSrm:
4326 case X86::CVTSI642SSrr:
4327 case X86::CVTSI642SSrm:
4328 case X86::CVTSI2SDrr:
4329 case X86::CVTSI2SDrm:
4330 case X86::CVTSI642SDrr:
4331 case X86::CVTSI642SDrm:
4332 case X86::CVTSD2SSrr:
4333 case X86::CVTSD2SSrm:
4334 case X86::CVTSS2SDrr:
4335 case X86::CVTSS2SDrm:
4342 case X86::RCPSSr_Int:
4343 case X86::RCPSSm_Int:
4350 case X86::RSQRTSSr_Int:
4351 case X86::RSQRTSSm_Int:
4354 case X86::SQRTSSr_Int:
4355 case X86::SQRTSSm_Int:
4358 case X86::SQRTSDr_Int:
4359 case X86::SQRTSDm_Int:
4362 case X86::POPCNT32rm:
4363 case X86::POPCNT32rr:
4364 case X86::POPCNT64rm:
4365 case X86::POPCNT64rr:
4366 return Subtarget.hasPOPCNTFalseDeps();
4367 case X86::LZCNT32rm:
4368 case X86::LZCNT32rr:
4369 case X86::LZCNT64rm:
4370 case X86::LZCNT64rr:
4371 case X86::TZCNT32rm:
4372 case X86::TZCNT32rr:
4373 case X86::TZCNT64rm:
4374 case X86::TZCNT64rr:
4375 return Subtarget.hasLZCNTFalseDeps();
4381 /// Inform the BreakFalseDeps pass how many idle
4382 /// instructions we would like before a partial register update.
4383 unsigned X86InstrInfo::getPartialRegUpdateClearance(
4384 const MachineInstr &MI, unsigned OpNum,
4385 const TargetRegisterInfo *TRI) const {
4386 if (OpNum != 0 || !hasPartialRegUpdate(MI.getOpcode(), Subtarget))
4389 // If MI is marked as reading Reg, the partial register update is wanted.
4390 const MachineOperand &MO = MI.getOperand(0);
4391 unsigned Reg = MO.getReg();
4392 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
4393 if (MO.readsReg() || MI.readsVirtualRegister(Reg))
4396 if (MI.readsRegister(Reg, TRI))
4400 // If any instructions in the clearance range are reading Reg, insert a
4401 // dependency breaking instruction, which is inexpensive and is likely to
4402 // be hidden in other instruction's cycles.
4403 return PartialRegUpdateClearance;
4406 // Return true for any instruction the copies the high bits of the first source
4407 // operand into the unused high bits of the destination operand.
4408 static bool hasUndefRegUpdate(unsigned Opcode) {
4410 case X86::VCVTSI2SSrr:
4411 case X86::VCVTSI2SSrm:
4412 case X86::VCVTSI2SSrr_Int:
4413 case X86::VCVTSI2SSrm_Int:
4414 case X86::VCVTSI642SSrr:
4415 case X86::VCVTSI642SSrm:
4416 case X86::VCVTSI642SSrr_Int:
4417 case X86::VCVTSI642SSrm_Int:
4418 case X86::VCVTSI2SDrr:
4419 case X86::VCVTSI2SDrm:
4420 case X86::VCVTSI2SDrr_Int:
4421 case X86::VCVTSI2SDrm_Int:
4422 case X86::VCVTSI642SDrr:
4423 case X86::VCVTSI642SDrm:
4424 case X86::VCVTSI642SDrr_Int:
4425 case X86::VCVTSI642SDrm_Int:
4426 case X86::VCVTSD2SSrr:
4427 case X86::VCVTSD2SSrm:
4428 case X86::VCVTSD2SSrr_Int:
4429 case X86::VCVTSD2SSrm_Int:
4430 case X86::VCVTSS2SDrr:
4431 case X86::VCVTSS2SDrm:
4432 case X86::VCVTSS2SDrr_Int:
4433 case X86::VCVTSS2SDrm_Int:
4435 case X86::VRCPSSr_Int:
4437 case X86::VRCPSSm_Int:
4438 case X86::VROUNDSDr:
4439 case X86::VROUNDSDm:
4440 case X86::VROUNDSDr_Int:
4441 case X86::VROUNDSDm_Int:
4442 case X86::VROUNDSSr:
4443 case X86::VROUNDSSm:
4444 case X86::VROUNDSSr_Int:
4445 case X86::VROUNDSSm_Int:
4446 case X86::VRSQRTSSr:
4447 case X86::VRSQRTSSr_Int:
4448 case X86::VRSQRTSSm:
4449 case X86::VRSQRTSSm_Int:
4451 case X86::VSQRTSSr_Int:
4453 case X86::VSQRTSSm_Int:
4455 case X86::VSQRTSDr_Int:
4457 case X86::VSQRTSDm_Int:
4459 case X86::VCVTSI2SSZrr:
4460 case X86::VCVTSI2SSZrm:
4461 case X86::VCVTSI2SSZrr_Int:
4462 case X86::VCVTSI2SSZrrb_Int:
4463 case X86::VCVTSI2SSZrm_Int:
4464 case X86::VCVTSI642SSZrr:
4465 case X86::VCVTSI642SSZrm:
4466 case X86::VCVTSI642SSZrr_Int:
4467 case X86::VCVTSI642SSZrrb_Int:
4468 case X86::VCVTSI642SSZrm_Int:
4469 case X86::VCVTSI2SDZrr:
4470 case X86::VCVTSI2SDZrm:
4471 case X86::VCVTSI2SDZrr_Int:
4472 case X86::VCVTSI2SDZrrb_Int:
4473 case X86::VCVTSI2SDZrm_Int:
4474 case X86::VCVTSI642SDZrr:
4475 case X86::VCVTSI642SDZrm:
4476 case X86::VCVTSI642SDZrr_Int:
4477 case X86::VCVTSI642SDZrrb_Int:
4478 case X86::VCVTSI642SDZrm_Int:
4479 case X86::VCVTUSI2SSZrr:
4480 case X86::VCVTUSI2SSZrm:
4481 case X86::VCVTUSI2SSZrr_Int:
4482 case X86::VCVTUSI2SSZrrb_Int:
4483 case X86::VCVTUSI2SSZrm_Int:
4484 case X86::VCVTUSI642SSZrr:
4485 case X86::VCVTUSI642SSZrm:
4486 case X86::VCVTUSI642SSZrr_Int:
4487 case X86::VCVTUSI642SSZrrb_Int:
4488 case X86::VCVTUSI642SSZrm_Int:
4489 case X86::VCVTUSI2SDZrr:
4490 case X86::VCVTUSI2SDZrm:
4491 case X86::VCVTUSI2SDZrr_Int:
4492 case X86::VCVTUSI2SDZrm_Int:
4493 case X86::VCVTUSI642SDZrr:
4494 case X86::VCVTUSI642SDZrm:
4495 case X86::VCVTUSI642SDZrr_Int:
4496 case X86::VCVTUSI642SDZrrb_Int:
4497 case X86::VCVTUSI642SDZrm_Int:
4498 case X86::VCVTSD2SSZrr:
4499 case X86::VCVTSD2SSZrr_Int:
4500 case X86::VCVTSD2SSZrrb_Int:
4501 case X86::VCVTSD2SSZrm:
4502 case X86::VCVTSD2SSZrm_Int:
4503 case X86::VCVTSS2SDZrr:
4504 case X86::VCVTSS2SDZrr_Int:
4505 case X86::VCVTSS2SDZrrb_Int:
4506 case X86::VCVTSS2SDZrm:
4507 case X86::VCVTSS2SDZrm_Int:
4508 case X86::VGETEXPSDZr:
4509 case X86::VGETEXPSDZrb:
4510 case X86::VGETEXPSDZm:
4511 case X86::VGETEXPSSZr:
4512 case X86::VGETEXPSSZrb:
4513 case X86::VGETEXPSSZm:
4514 case X86::VGETMANTSDZrri:
4515 case X86::VGETMANTSDZrrib:
4516 case X86::VGETMANTSDZrmi:
4517 case X86::VGETMANTSSZrri:
4518 case X86::VGETMANTSSZrrib:
4519 case X86::VGETMANTSSZrmi:
4520 case X86::VRNDSCALESDZr:
4521 case X86::VRNDSCALESDZr_Int:
4522 case X86::VRNDSCALESDZrb_Int:
4523 case X86::VRNDSCALESDZm:
4524 case X86::VRNDSCALESDZm_Int:
4525 case X86::VRNDSCALESSZr:
4526 case X86::VRNDSCALESSZr_Int:
4527 case X86::VRNDSCALESSZrb_Int:
4528 case X86::VRNDSCALESSZm:
4529 case X86::VRNDSCALESSZm_Int:
4530 case X86::VRCP14SDZrr:
4531 case X86::VRCP14SDZrm:
4532 case X86::VRCP14SSZrr:
4533 case X86::VRCP14SSZrm:
4534 case X86::VRCP28SDZr:
4535 case X86::VRCP28SDZrb:
4536 case X86::VRCP28SDZm:
4537 case X86::VRCP28SSZr:
4538 case X86::VRCP28SSZrb:
4539 case X86::VRCP28SSZm:
4540 case X86::VREDUCESSZrmi:
4541 case X86::VREDUCESSZrri:
4542 case X86::VREDUCESSZrrib:
4543 case X86::VRSQRT14SDZrr:
4544 case X86::VRSQRT14SDZrm:
4545 case X86::VRSQRT14SSZrr:
4546 case X86::VRSQRT14SSZrm:
4547 case X86::VRSQRT28SDZr:
4548 case X86::VRSQRT28SDZrb:
4549 case X86::VRSQRT28SDZm:
4550 case X86::VRSQRT28SSZr:
4551 case X86::VRSQRT28SSZrb:
4552 case X86::VRSQRT28SSZm:
4553 case X86::VSQRTSSZr:
4554 case X86::VSQRTSSZr_Int:
4555 case X86::VSQRTSSZrb_Int:
4556 case X86::VSQRTSSZm:
4557 case X86::VSQRTSSZm_Int:
4558 case X86::VSQRTSDZr:
4559 case X86::VSQRTSDZr_Int:
4560 case X86::VSQRTSDZrb_Int:
4561 case X86::VSQRTSDZm:
4562 case X86::VSQRTSDZm_Int:
4569 /// Inform the BreakFalseDeps pass how many idle instructions we would like
4570 /// before certain undef register reads.
4572 /// This catches the VCVTSI2SD family of instructions:
4574 /// vcvtsi2sdq %rax, undef %xmm0, %xmm14
4576 /// We should to be careful *not* to catch VXOR idioms which are presumably
4577 /// handled specially in the pipeline:
4579 /// vxorps undef %xmm1, undef %xmm1, %xmm1
4581 /// Like getPartialRegUpdateClearance, this makes a strong assumption that the
4582 /// high bits that are passed-through are not live.
4584 X86InstrInfo::getUndefRegClearance(const MachineInstr &MI, unsigned &OpNum,
4585 const TargetRegisterInfo *TRI) const {
4586 if (!hasUndefRegUpdate(MI.getOpcode()))
4589 // Set the OpNum parameter to the first source operand.
4592 const MachineOperand &MO = MI.getOperand(OpNum);
4593 if (MO.isUndef() && TargetRegisterInfo::isPhysicalRegister(MO.getReg())) {
4594 return UndefRegClearance;
4599 void X86InstrInfo::breakPartialRegDependency(
4600 MachineInstr &MI, unsigned OpNum, const TargetRegisterInfo *TRI) const {
4601 unsigned Reg = MI.getOperand(OpNum).getReg();
4602 // If MI kills this register, the false dependence is already broken.
4603 if (MI.killsRegister(Reg, TRI))
4606 if (X86::VR128RegClass.contains(Reg)) {
4607 // These instructions are all floating point domain, so xorps is the best
4609 unsigned Opc = Subtarget.hasAVX() ? X86::VXORPSrr : X86::XORPSrr;
4610 BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), get(Opc), Reg)
4611 .addReg(Reg, RegState::Undef)
4612 .addReg(Reg, RegState::Undef);
4613 MI.addRegisterKilled(Reg, TRI, true);
4614 } else if (X86::VR256RegClass.contains(Reg)) {
4615 // Use vxorps to clear the full ymm register.
4616 // It wants to read and write the xmm sub-register.
4617 unsigned XReg = TRI->getSubReg(Reg, X86::sub_xmm);
4618 BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), get(X86::VXORPSrr), XReg)
4619 .addReg(XReg, RegState::Undef)
4620 .addReg(XReg, RegState::Undef)
4621 .addReg(Reg, RegState::ImplicitDefine);
4622 MI.addRegisterKilled(Reg, TRI, true);
4623 } else if (X86::GR64RegClass.contains(Reg)) {
4624 // Using XOR32rr because it has shorter encoding and zeros up the upper bits
4626 unsigned XReg = TRI->getSubReg(Reg, X86::sub_32bit);
4627 BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), get(X86::XOR32rr), XReg)
4628 .addReg(XReg, RegState::Undef)
4629 .addReg(XReg, RegState::Undef)
4630 .addReg(Reg, RegState::ImplicitDefine);
4631 MI.addRegisterKilled(Reg, TRI, true);
4632 } else if (X86::GR32RegClass.contains(Reg)) {
4633 BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), get(X86::XOR32rr), Reg)
4634 .addReg(Reg, RegState::Undef)
4635 .addReg(Reg, RegState::Undef);
4636 MI.addRegisterKilled(Reg, TRI, true);
4640 static void addOperands(MachineInstrBuilder &MIB, ArrayRef<MachineOperand> MOs,
4641 int PtrOffset = 0) {
4642 unsigned NumAddrOps = MOs.size();
4644 if (NumAddrOps < 4) {
4645 // FrameIndex only - add an immediate offset (whether its zero or not).
4646 for (unsigned i = 0; i != NumAddrOps; ++i)
4648 addOffset(MIB, PtrOffset);
4650 // General Memory Addressing - we need to add any offset to an existing
4652 assert(MOs.size() == 5 && "Unexpected memory operand list length");
4653 for (unsigned i = 0; i != NumAddrOps; ++i) {
4654 const MachineOperand &MO = MOs[i];
4655 if (i == 3 && PtrOffset != 0) {
4656 MIB.addDisp(MO, PtrOffset);
4664 static void updateOperandRegConstraints(MachineFunction &MF,
4665 MachineInstr &NewMI,
4666 const TargetInstrInfo &TII) {
4667 MachineRegisterInfo &MRI = MF.getRegInfo();
4668 const TargetRegisterInfo &TRI = *MRI.getTargetRegisterInfo();
4670 for (int Idx : llvm::seq<int>(0, NewMI.getNumOperands())) {
4671 MachineOperand &MO = NewMI.getOperand(Idx);
4672 // We only need to update constraints on virtual register operands.
4675 unsigned Reg = MO.getReg();
4676 if (!TRI.isVirtualRegister(Reg))
4679 auto *NewRC = MRI.constrainRegClass(
4680 Reg, TII.getRegClass(NewMI.getDesc(), Idx, &TRI, MF));
4683 dbgs() << "WARNING: Unable to update register constraint for operand "
4684 << Idx << " of instruction:\n";
4685 NewMI.dump(); dbgs() << "\n");
4690 static MachineInstr *FuseTwoAddrInst(MachineFunction &MF, unsigned Opcode,
4691 ArrayRef<MachineOperand> MOs,
4692 MachineBasicBlock::iterator InsertPt,
4694 const TargetInstrInfo &TII) {
4695 // Create the base instruction with the memory operand as the first part.
4696 // Omit the implicit operands, something BuildMI can't do.
4697 MachineInstr *NewMI =
4698 MF.CreateMachineInstr(TII.get(Opcode), MI.getDebugLoc(), true);
4699 MachineInstrBuilder MIB(MF, NewMI);
4700 addOperands(MIB, MOs);
4702 // Loop over the rest of the ri operands, converting them over.
4703 unsigned NumOps = MI.getDesc().getNumOperands() - 2;
4704 for (unsigned i = 0; i != NumOps; ++i) {
4705 MachineOperand &MO = MI.getOperand(i + 2);
4708 for (unsigned i = NumOps + 2, e = MI.getNumOperands(); i != e; ++i) {
4709 MachineOperand &MO = MI.getOperand(i);
4713 updateOperandRegConstraints(MF, *NewMI, TII);
4715 MachineBasicBlock *MBB = InsertPt->getParent();
4716 MBB->insert(InsertPt, NewMI);
4721 static MachineInstr *FuseInst(MachineFunction &MF, unsigned Opcode,
4722 unsigned OpNo, ArrayRef<MachineOperand> MOs,
4723 MachineBasicBlock::iterator InsertPt,
4724 MachineInstr &MI, const TargetInstrInfo &TII,
4725 int PtrOffset = 0) {
4726 // Omit the implicit operands, something BuildMI can't do.
4727 MachineInstr *NewMI =
4728 MF.CreateMachineInstr(TII.get(Opcode), MI.getDebugLoc(), true);
4729 MachineInstrBuilder MIB(MF, NewMI);
4731 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
4732 MachineOperand &MO = MI.getOperand(i);
4734 assert(MO.isReg() && "Expected to fold into reg operand!");
4735 addOperands(MIB, MOs, PtrOffset);
4741 updateOperandRegConstraints(MF, *NewMI, TII);
4743 MachineBasicBlock *MBB = InsertPt->getParent();
4744 MBB->insert(InsertPt, NewMI);
4749 static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode,
4750 ArrayRef<MachineOperand> MOs,
4751 MachineBasicBlock::iterator InsertPt,
4753 MachineInstrBuilder MIB = BuildMI(*InsertPt->getParent(), InsertPt,
4754 MI.getDebugLoc(), TII.get(Opcode));
4755 addOperands(MIB, MOs);
4756 return MIB.addImm(0);
4759 MachineInstr *X86InstrInfo::foldMemoryOperandCustom(
4760 MachineFunction &MF, MachineInstr &MI, unsigned OpNum,
4761 ArrayRef<MachineOperand> MOs, MachineBasicBlock::iterator InsertPt,
4762 unsigned Size, unsigned Align) const {
4763 switch (MI.getOpcode()) {
4764 case X86::INSERTPSrr:
4765 case X86::VINSERTPSrr:
4766 case X86::VINSERTPSZrr:
4767 // Attempt to convert the load of inserted vector into a fold load
4768 // of a single float.
4770 unsigned Imm = MI.getOperand(MI.getNumOperands() - 1).getImm();
4771 unsigned ZMask = Imm & 15;
4772 unsigned DstIdx = (Imm >> 4) & 3;
4773 unsigned SrcIdx = (Imm >> 6) & 3;
4775 const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
4776 const TargetRegisterClass *RC = getRegClass(MI.getDesc(), OpNum, &RI, MF);
4777 unsigned RCSize = TRI.getRegSizeInBits(*RC) / 8;
4778 if (Size <= RCSize && 4 <= Align) {
4779 int PtrOffset = SrcIdx * 4;
4780 unsigned NewImm = (DstIdx << 4) | ZMask;
4781 unsigned NewOpCode =
4782 (MI.getOpcode() == X86::VINSERTPSZrr) ? X86::VINSERTPSZrm :
4783 (MI.getOpcode() == X86::VINSERTPSrr) ? X86::VINSERTPSrm :
4785 MachineInstr *NewMI =
4786 FuseInst(MF, NewOpCode, OpNum, MOs, InsertPt, MI, *this, PtrOffset);
4787 NewMI->getOperand(NewMI->getNumOperands() - 1).setImm(NewImm);
4792 case X86::MOVHLPSrr:
4793 case X86::VMOVHLPSrr:
4794 case X86::VMOVHLPSZrr:
4795 // Move the upper 64-bits of the second operand to the lower 64-bits.
4796 // To fold the load, adjust the pointer to the upper and use (V)MOVLPS.
4797 // TODO: In most cases AVX doesn't have a 8-byte alignment requirement.
4799 const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
4800 const TargetRegisterClass *RC = getRegClass(MI.getDesc(), OpNum, &RI, MF);
4801 unsigned RCSize = TRI.getRegSizeInBits(*RC) / 8;
4802 if (Size <= RCSize && 8 <= Align) {
4803 unsigned NewOpCode =
4804 (MI.getOpcode() == X86::VMOVHLPSZrr) ? X86::VMOVLPSZ128rm :
4805 (MI.getOpcode() == X86::VMOVHLPSrr) ? X86::VMOVLPSrm :
4807 MachineInstr *NewMI =
4808 FuseInst(MF, NewOpCode, OpNum, MOs, InsertPt, MI, *this, 8);
4818 static bool shouldPreventUndefRegUpdateMemFold(MachineFunction &MF, MachineInstr &MI) {
4819 if (MF.getFunction().optForSize() || !hasUndefRegUpdate(MI.getOpcode()) ||
4820 !MI.getOperand(1).isReg())
4823 // The are two cases we need to handle depending on where in the pipeline
4824 // the folding attempt is being made.
4825 // -Register has the undef flag set.
4826 // -Register is produced by the IMPLICIT_DEF instruction.
4828 if (MI.getOperand(1).isUndef())
4831 MachineRegisterInfo &RegInfo = MF.getRegInfo();
4832 MachineInstr *VRegDef = RegInfo.getUniqueVRegDef(MI.getOperand(1).getReg());
4833 return VRegDef && VRegDef->isImplicitDef();
4837 MachineInstr *X86InstrInfo::foldMemoryOperandImpl(
4838 MachineFunction &MF, MachineInstr &MI, unsigned OpNum,
4839 ArrayRef<MachineOperand> MOs, MachineBasicBlock::iterator InsertPt,
4840 unsigned Size, unsigned Align, bool AllowCommute) const {
4841 bool isSlowTwoMemOps = Subtarget.slowTwoMemOps();
4842 bool isTwoAddrFold = false;
4844 // For CPUs that favor the register form of a call or push,
4845 // do not fold loads into calls or pushes, unless optimizing for size
4847 if (isSlowTwoMemOps && !MF.getFunction().optForMinSize() &&
4848 (MI.getOpcode() == X86::CALL32r || MI.getOpcode() == X86::CALL64r ||
4849 MI.getOpcode() == X86::PUSH16r || MI.getOpcode() == X86::PUSH32r ||
4850 MI.getOpcode() == X86::PUSH64r))
4853 // Avoid partial and undef register update stalls unless optimizing for size.
4854 if (!MF.getFunction().optForSize() &&
4855 (hasPartialRegUpdate(MI.getOpcode(), Subtarget) ||
4856 shouldPreventUndefRegUpdateMemFold(MF, MI)))
4859 unsigned NumOps = MI.getDesc().getNumOperands();
4861 NumOps > 1 && MI.getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
4863 // FIXME: AsmPrinter doesn't know how to handle
4864 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
4865 if (MI.getOpcode() == X86::ADD32ri &&
4866 MI.getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
4869 // GOTTPOFF relocation loads can only be folded into add instructions.
4870 // FIXME: Need to exclude other relocations that only support specific
4872 if (MOs.size() == X86::AddrNumOperands &&
4873 MOs[X86::AddrDisp].getTargetFlags() == X86II::MO_GOTTPOFF &&
4874 MI.getOpcode() != X86::ADD64rr)
4877 MachineInstr *NewMI = nullptr;
4879 // Attempt to fold any custom cases we have.
4880 if (MachineInstr *CustomMI =
4881 foldMemoryOperandCustom(MF, MI, OpNum, MOs, InsertPt, Size, Align))
4884 const X86MemoryFoldTableEntry *I = nullptr;
4886 // Folding a memory location into the two-address part of a two-address
4887 // instruction is different than folding it other places. It requires
4888 // replacing the *two* registers with the memory location.
4889 if (isTwoAddr && NumOps >= 2 && OpNum < 2 && MI.getOperand(0).isReg() &&
4890 MI.getOperand(1).isReg() &&
4891 MI.getOperand(0).getReg() == MI.getOperand(1).getReg()) {
4892 I = lookupTwoAddrFoldTable(MI.getOpcode());
4893 isTwoAddrFold = true;
4896 if (MI.getOpcode() == X86::MOV32r0) {
4897 NewMI = MakeM0Inst(*this, X86::MOV32mi, MOs, InsertPt, MI);
4903 I = lookupFoldTable(MI.getOpcode(), OpNum);
4907 unsigned Opcode = I->DstOp;
4908 unsigned MinAlign = (I->Flags & TB_ALIGN_MASK) >> TB_ALIGN_SHIFT;
4909 if (Align < MinAlign)
4911 bool NarrowToMOV32rm = false;
4913 const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
4914 const TargetRegisterClass *RC = getRegClass(MI.getDesc(), OpNum,
4916 unsigned RCSize = TRI.getRegSizeInBits(*RC) / 8;
4917 if (Size < RCSize) {
4918 // Check if it's safe to fold the load. If the size of the object is
4919 // narrower than the load width, then it's not.
4920 if (Opcode != X86::MOV64rm || RCSize != 8 || Size != 4)
4922 // If this is a 64-bit load, but the spill slot is 32, then we can do
4923 // a 32-bit load which is implicitly zero-extended. This likely is
4924 // due to live interval analysis remat'ing a load from stack slot.
4925 if (MI.getOperand(0).getSubReg() || MI.getOperand(1).getSubReg())
4927 Opcode = X86::MOV32rm;
4928 NarrowToMOV32rm = true;
4933 NewMI = FuseTwoAddrInst(MF, Opcode, MOs, InsertPt, MI, *this);
4935 NewMI = FuseInst(MF, Opcode, OpNum, MOs, InsertPt, MI, *this);
4937 if (NarrowToMOV32rm) {
4938 // If this is the special case where we use a MOV32rm to load a 32-bit
4939 // value and zero-extend the top bits. Change the destination register
4941 unsigned DstReg = NewMI->getOperand(0).getReg();
4942 if (TargetRegisterInfo::isPhysicalRegister(DstReg))
4943 NewMI->getOperand(0).setReg(RI.getSubReg(DstReg, X86::sub_32bit));
4945 NewMI->getOperand(0).setSubReg(X86::sub_32bit);
4950 // If the instruction and target operand are commutable, commute the
4951 // instruction and try again.
4953 unsigned CommuteOpIdx1 = OpNum, CommuteOpIdx2 = CommuteAnyOperandIndex;
4954 if (findCommutedOpIndices(MI, CommuteOpIdx1, CommuteOpIdx2)) {
4955 bool HasDef = MI.getDesc().getNumDefs();
4956 unsigned Reg0 = HasDef ? MI.getOperand(0).getReg() : 0;
4957 unsigned Reg1 = MI.getOperand(CommuteOpIdx1).getReg();
4958 unsigned Reg2 = MI.getOperand(CommuteOpIdx2).getReg();
4960 0 == MI.getDesc().getOperandConstraint(CommuteOpIdx1, MCOI::TIED_TO);
4962 0 == MI.getDesc().getOperandConstraint(CommuteOpIdx2, MCOI::TIED_TO);
4964 // If either of the commutable operands are tied to the destination
4965 // then we can not commute + fold.
4966 if ((HasDef && Reg0 == Reg1 && Tied1) ||
4967 (HasDef && Reg0 == Reg2 && Tied2))
4970 MachineInstr *CommutedMI =
4971 commuteInstruction(MI, false, CommuteOpIdx1, CommuteOpIdx2);
4973 // Unable to commute.
4976 if (CommutedMI != &MI) {
4977 // New instruction. We can't fold from this.
4978 CommutedMI->eraseFromParent();
4982 // Attempt to fold with the commuted version of the instruction.
4983 NewMI = foldMemoryOperandImpl(MF, MI, CommuteOpIdx2, MOs, InsertPt,
4984 Size, Align, /*AllowCommute=*/false);
4988 // Folding failed again - undo the commute before returning.
4989 MachineInstr *UncommutedMI =
4990 commuteInstruction(MI, false, CommuteOpIdx1, CommuteOpIdx2);
4991 if (!UncommutedMI) {
4992 // Unable to commute.
4995 if (UncommutedMI != &MI) {
4996 // New instruction. It doesn't need to be kept.
4997 UncommutedMI->eraseFromParent();
5001 // Return here to prevent duplicate fuse failure report.
5007 if (PrintFailedFusing && !MI.isCopy())
5008 dbgs() << "We failed to fuse operand " << OpNum << " in " << MI;
5013 X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF, MachineInstr &MI,
5014 ArrayRef<unsigned> Ops,
5015 MachineBasicBlock::iterator InsertPt,
5016 int FrameIndex, LiveIntervals *LIS) const {
5017 // Check switch flag
5021 // Avoid partial and undef register update stalls unless optimizing for size.
5022 if (!MF.getFunction().optForSize() &&
5023 (hasPartialRegUpdate(MI.getOpcode(), Subtarget) ||
5024 shouldPreventUndefRegUpdateMemFold(MF, MI)))
5027 // Don't fold subreg spills, or reloads that use a high subreg.
5028 for (auto Op : Ops) {
5029 MachineOperand &MO = MI.getOperand(Op);
5030 auto SubReg = MO.getSubReg();
5031 if (SubReg && (MO.isDef() || SubReg == X86::sub_8bit_hi))
5035 const MachineFrameInfo &MFI = MF.getFrameInfo();
5036 unsigned Size = MFI.getObjectSize(FrameIndex);
5037 unsigned Alignment = MFI.getObjectAlignment(FrameIndex);
5038 // If the function stack isn't realigned we don't want to fold instructions
5039 // that need increased alignment.
5040 if (!RI.needsStackRealignment(MF))
5042 std::min(Alignment, Subtarget.getFrameLowering()->getStackAlignment());
5043 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
5044 unsigned NewOpc = 0;
5045 unsigned RCSize = 0;
5046 switch (MI.getOpcode()) {
5047 default: return nullptr;
5048 case X86::TEST8rr: NewOpc = X86::CMP8ri; RCSize = 1; break;
5049 case X86::TEST16rr: NewOpc = X86::CMP16ri8; RCSize = 2; break;
5050 case X86::TEST32rr: NewOpc = X86::CMP32ri8; RCSize = 4; break;
5051 case X86::TEST64rr: NewOpc = X86::CMP64ri8; RCSize = 8; break;
5053 // Check if it's safe to fold the load. If the size of the object is
5054 // narrower than the load width, then it's not.
5057 // Change to CMPXXri r, 0 first.
5058 MI.setDesc(get(NewOpc));
5059 MI.getOperand(1).ChangeToImmediate(0);
5060 } else if (Ops.size() != 1)
5063 return foldMemoryOperandImpl(MF, MI, Ops[0],
5064 MachineOperand::CreateFI(FrameIndex), InsertPt,
5065 Size, Alignment, /*AllowCommute=*/true);
5068 /// Check if \p LoadMI is a partial register load that we can't fold into \p MI
5069 /// because the latter uses contents that wouldn't be defined in the folded
5070 /// version. For instance, this transformation isn't legal:
5071 /// movss (%rdi), %xmm0
5072 /// addps %xmm0, %xmm0
5074 /// addps (%rdi), %xmm0
5076 /// But this one is:
5077 /// movss (%rdi), %xmm0
5078 /// addss %xmm0, %xmm0
5080 /// addss (%rdi), %xmm0
5082 static bool isNonFoldablePartialRegisterLoad(const MachineInstr &LoadMI,
5083 const MachineInstr &UserMI,
5084 const MachineFunction &MF) {
5085 unsigned Opc = LoadMI.getOpcode();
5086 unsigned UserOpc = UserMI.getOpcode();
5087 const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
5088 const TargetRegisterClass *RC =
5089 MF.getRegInfo().getRegClass(LoadMI.getOperand(0).getReg());
5090 unsigned RegSize = TRI.getRegSizeInBits(*RC);
5092 if ((Opc == X86::MOVSSrm || Opc == X86::VMOVSSrm || Opc == X86::VMOVSSZrm) &&
5094 // These instructions only load 32 bits, we can't fold them if the
5095 // destination register is wider than 32 bits (4 bytes), and its user
5096 // instruction isn't scalar (SS).
5098 case X86::ADDSSrr_Int: case X86::VADDSSrr_Int: case X86::VADDSSZrr_Int:
5099 case X86::CMPSSrr_Int: case X86::VCMPSSrr_Int: case X86::VCMPSSZrr_Int:
5100 case X86::DIVSSrr_Int: case X86::VDIVSSrr_Int: case X86::VDIVSSZrr_Int:
5101 case X86::MAXSSrr_Int: case X86::VMAXSSrr_Int: case X86::VMAXSSZrr_Int:
5102 case X86::MINSSrr_Int: case X86::VMINSSrr_Int: case X86::VMINSSZrr_Int:
5103 case X86::MULSSrr_Int: case X86::VMULSSrr_Int: case X86::VMULSSZrr_Int:
5104 case X86::SUBSSrr_Int: case X86::VSUBSSrr_Int: case X86::VSUBSSZrr_Int:
5105 case X86::VADDSSZrr_Intk: case X86::VADDSSZrr_Intkz:
5106 case X86::VDIVSSZrr_Intk: case X86::VDIVSSZrr_Intkz:
5107 case X86::VMAXSSZrr_Intk: case X86::VMAXSSZrr_Intkz:
5108 case X86::VMINSSZrr_Intk: case X86::VMINSSZrr_Intkz:
5109 case X86::VMULSSZrr_Intk: case X86::VMULSSZrr_Intkz:
5110 case X86::VSUBSSZrr_Intk: case X86::VSUBSSZrr_Intkz:
5111 case X86::VFMADDSS4rr_Int: case X86::VFNMADDSS4rr_Int:
5112 case X86::VFMSUBSS4rr_Int: case X86::VFNMSUBSS4rr_Int:
5113 case X86::VFMADD132SSr_Int: case X86::VFNMADD132SSr_Int:
5114 case X86::VFMADD213SSr_Int: case X86::VFNMADD213SSr_Int:
5115 case X86::VFMADD231SSr_Int: case X86::VFNMADD231SSr_Int:
5116 case X86::VFMSUB132SSr_Int: case X86::VFNMSUB132SSr_Int:
5117 case X86::VFMSUB213SSr_Int: case X86::VFNMSUB213SSr_Int:
5118 case X86::VFMSUB231SSr_Int: case X86::VFNMSUB231SSr_Int:
5119 case X86::VFMADD132SSZr_Int: case X86::VFNMADD132SSZr_Int:
5120 case X86::VFMADD213SSZr_Int: case X86::VFNMADD213SSZr_Int:
5121 case X86::VFMADD231SSZr_Int: case X86::VFNMADD231SSZr_Int:
5122 case X86::VFMSUB132SSZr_Int: case X86::VFNMSUB132SSZr_Int:
5123 case X86::VFMSUB213SSZr_Int: case X86::VFNMSUB213SSZr_Int:
5124 case X86::VFMSUB231SSZr_Int: case X86::VFNMSUB231SSZr_Int:
5125 case X86::VFMADD132SSZr_Intk: case X86::VFNMADD132SSZr_Intk:
5126 case X86::VFMADD213SSZr_Intk: case X86::VFNMADD213SSZr_Intk:
5127 case X86::VFMADD231SSZr_Intk: case X86::VFNMADD231SSZr_Intk:
5128 case X86::VFMSUB132SSZr_Intk: case X86::VFNMSUB132SSZr_Intk:
5129 case X86::VFMSUB213SSZr_Intk: case X86::VFNMSUB213SSZr_Intk:
5130 case X86::VFMSUB231SSZr_Intk: case X86::VFNMSUB231SSZr_Intk:
5131 case X86::VFMADD132SSZr_Intkz: case X86::VFNMADD132SSZr_Intkz:
5132 case X86::VFMADD213SSZr_Intkz: case X86::VFNMADD213SSZr_Intkz:
5133 case X86::VFMADD231SSZr_Intkz: case X86::VFNMADD231SSZr_Intkz:
5134 case X86::VFMSUB132SSZr_Intkz: case X86::VFNMSUB132SSZr_Intkz:
5135 case X86::VFMSUB213SSZr_Intkz: case X86::VFNMSUB213SSZr_Intkz:
5136 case X86::VFMSUB231SSZr_Intkz: case X86::VFNMSUB231SSZr_Intkz:
5143 if ((Opc == X86::MOVSDrm || Opc == X86::VMOVSDrm || Opc == X86::VMOVSDZrm) &&
5145 // These instructions only load 64 bits, we can't fold them if the
5146 // destination register is wider than 64 bits (8 bytes), and its user
5147 // instruction isn't scalar (SD).
5149 case X86::ADDSDrr_Int: case X86::VADDSDrr_Int: case X86::VADDSDZrr_Int:
5150 case X86::CMPSDrr_Int: case X86::VCMPSDrr_Int: case X86::VCMPSDZrr_Int:
5151 case X86::DIVSDrr_Int: case X86::VDIVSDrr_Int: case X86::VDIVSDZrr_Int:
5152 case X86::MAXSDrr_Int: case X86::VMAXSDrr_Int: case X86::VMAXSDZrr_Int:
5153 case X86::MINSDrr_Int: case X86::VMINSDrr_Int: case X86::VMINSDZrr_Int:
5154 case X86::MULSDrr_Int: case X86::VMULSDrr_Int: case X86::VMULSDZrr_Int:
5155 case X86::SUBSDrr_Int: case X86::VSUBSDrr_Int: case X86::VSUBSDZrr_Int:
5156 case X86::VADDSDZrr_Intk: case X86::VADDSDZrr_Intkz:
5157 case X86::VDIVSDZrr_Intk: case X86::VDIVSDZrr_Intkz:
5158 case X86::VMAXSDZrr_Intk: case X86::VMAXSDZrr_Intkz:
5159 case X86::VMINSDZrr_Intk: case X86::VMINSDZrr_Intkz:
5160 case X86::VMULSDZrr_Intk: case X86::VMULSDZrr_Intkz:
5161 case X86::VSUBSDZrr_Intk: case X86::VSUBSDZrr_Intkz:
5162 case X86::VFMADDSD4rr_Int: case X86::VFNMADDSD4rr_Int:
5163 case X86::VFMSUBSD4rr_Int: case X86::VFNMSUBSD4rr_Int:
5164 case X86::VFMADD132SDr_Int: case X86::VFNMADD132SDr_Int:
5165 case X86::VFMADD213SDr_Int: case X86::VFNMADD213SDr_Int:
5166 case X86::VFMADD231SDr_Int: case X86::VFNMADD231SDr_Int:
5167 case X86::VFMSUB132SDr_Int: case X86::VFNMSUB132SDr_Int:
5168 case X86::VFMSUB213SDr_Int: case X86::VFNMSUB213SDr_Int:
5169 case X86::VFMSUB231SDr_Int: case X86::VFNMSUB231SDr_Int:
5170 case X86::VFMADD132SDZr_Int: case X86::VFNMADD132SDZr_Int:
5171 case X86::VFMADD213SDZr_Int: case X86::VFNMADD213SDZr_Int:
5172 case X86::VFMADD231SDZr_Int: case X86::VFNMADD231SDZr_Int:
5173 case X86::VFMSUB132SDZr_Int: case X86::VFNMSUB132SDZr_Int:
5174 case X86::VFMSUB213SDZr_Int: case X86::VFNMSUB213SDZr_Int:
5175 case X86::VFMSUB231SDZr_Int: case X86::VFNMSUB231SDZr_Int:
5176 case X86::VFMADD132SDZr_Intk: case X86::VFNMADD132SDZr_Intk:
5177 case X86::VFMADD213SDZr_Intk: case X86::VFNMADD213SDZr_Intk:
5178 case X86::VFMADD231SDZr_Intk: case X86::VFNMADD231SDZr_Intk:
5179 case X86::VFMSUB132SDZr_Intk: case X86::VFNMSUB132SDZr_Intk:
5180 case X86::VFMSUB213SDZr_Intk: case X86::VFNMSUB213SDZr_Intk:
5181 case X86::VFMSUB231SDZr_Intk: case X86::VFNMSUB231SDZr_Intk:
5182 case X86::VFMADD132SDZr_Intkz: case X86::VFNMADD132SDZr_Intkz:
5183 case X86::VFMADD213SDZr_Intkz: case X86::VFNMADD213SDZr_Intkz:
5184 case X86::VFMADD231SDZr_Intkz: case X86::VFNMADD231SDZr_Intkz:
5185 case X86::VFMSUB132SDZr_Intkz: case X86::VFNMSUB132SDZr_Intkz:
5186 case X86::VFMSUB213SDZr_Intkz: case X86::VFNMSUB213SDZr_Intkz:
5187 case X86::VFMSUB231SDZr_Intkz: case X86::VFNMSUB231SDZr_Intkz:
5197 MachineInstr *X86InstrInfo::foldMemoryOperandImpl(
5198 MachineFunction &MF, MachineInstr &MI, ArrayRef<unsigned> Ops,
5199 MachineBasicBlock::iterator InsertPt, MachineInstr &LoadMI,
5200 LiveIntervals *LIS) const {
5202 // TODO: Support the case where LoadMI loads a wide register, but MI
5203 // only uses a subreg.
5204 for (auto Op : Ops) {
5205 if (MI.getOperand(Op).getSubReg())
5209 // If loading from a FrameIndex, fold directly from the FrameIndex.
5210 unsigned NumOps = LoadMI.getDesc().getNumOperands();
5212 if (isLoadFromStackSlot(LoadMI, FrameIndex)) {
5213 if (isNonFoldablePartialRegisterLoad(LoadMI, MI, MF))
5215 return foldMemoryOperandImpl(MF, MI, Ops, InsertPt, FrameIndex, LIS);
5218 // Check switch flag
5219 if (NoFusing) return nullptr;
5221 // Avoid partial and undef register update stalls unless optimizing for size.
5222 if (!MF.getFunction().optForSize() &&
5223 (hasPartialRegUpdate(MI.getOpcode(), Subtarget) ||
5224 shouldPreventUndefRegUpdateMemFold(MF, MI)))
5227 // Determine the alignment of the load.
5228 unsigned Alignment = 0;
5229 if (LoadMI.hasOneMemOperand())
5230 Alignment = (*LoadMI.memoperands_begin())->getAlignment();
5232 switch (LoadMI.getOpcode()) {
5233 case X86::AVX512_512_SET0:
5234 case X86::AVX512_512_SETALLONES:
5237 case X86::AVX2_SETALLONES:
5238 case X86::AVX1_SETALLONES:
5240 case X86::AVX512_256_SET0:
5244 case X86::V_SETALLONES:
5245 case X86::AVX512_128_SET0:
5250 case X86::AVX512_FsFLD0SD:
5254 case X86::AVX512_FsFLD0SS:
5260 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
5261 unsigned NewOpc = 0;
5262 switch (MI.getOpcode()) {
5263 default: return nullptr;
5264 case X86::TEST8rr: NewOpc = X86::CMP8ri; break;
5265 case X86::TEST16rr: NewOpc = X86::CMP16ri8; break;
5266 case X86::TEST32rr: NewOpc = X86::CMP32ri8; break;
5267 case X86::TEST64rr: NewOpc = X86::CMP64ri8; break;
5269 // Change to CMPXXri r, 0 first.
5270 MI.setDesc(get(NewOpc));
5271 MI.getOperand(1).ChangeToImmediate(0);
5272 } else if (Ops.size() != 1)
5275 // Make sure the subregisters match.
5276 // Otherwise we risk changing the size of the load.
5277 if (LoadMI.getOperand(0).getSubReg() != MI.getOperand(Ops[0]).getSubReg())
5280 SmallVector<MachineOperand,X86::AddrNumOperands> MOs;
5281 switch (LoadMI.getOpcode()) {
5284 case X86::V_SETALLONES:
5285 case X86::AVX2_SETALLONES:
5286 case X86::AVX1_SETALLONES:
5288 case X86::AVX512_128_SET0:
5289 case X86::AVX512_256_SET0:
5290 case X86::AVX512_512_SET0:
5291 case X86::AVX512_512_SETALLONES:
5293 case X86::AVX512_FsFLD0SD:
5295 case X86::AVX512_FsFLD0SS: {
5296 // Folding a V_SET0 or V_SETALLONES as a load, to ease register pressure.
5297 // Create a constant-pool entry and operands to load from it.
5299 // Medium and large mode can't fold loads this way.
5300 if (MF.getTarget().getCodeModel() != CodeModel::Small &&
5301 MF.getTarget().getCodeModel() != CodeModel::Kernel)
5304 // x86-32 PIC requires a PIC base register for constant pools.
5305 unsigned PICBase = 0;
5306 if (MF.getTarget().isPositionIndependent()) {
5307 if (Subtarget.is64Bit())
5310 // FIXME: PICBase = getGlobalBaseReg(&MF);
5311 // This doesn't work for several reasons.
5312 // 1. GlobalBaseReg may have been spilled.
5313 // 2. It may not be live at MI.
5317 // Create a constant-pool entry.
5318 MachineConstantPool &MCP = *MF.getConstantPool();
5320 unsigned Opc = LoadMI.getOpcode();
5321 if (Opc == X86::FsFLD0SS || Opc == X86::AVX512_FsFLD0SS)
5322 Ty = Type::getFloatTy(MF.getFunction().getContext());
5323 else if (Opc == X86::FsFLD0SD || Opc == X86::AVX512_FsFLD0SD)
5324 Ty = Type::getDoubleTy(MF.getFunction().getContext());
5325 else if (Opc == X86::AVX512_512_SET0 || Opc == X86::AVX512_512_SETALLONES)
5326 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction().getContext()),16);
5327 else if (Opc == X86::AVX2_SETALLONES || Opc == X86::AVX_SET0 ||
5328 Opc == X86::AVX512_256_SET0 || Opc == X86::AVX1_SETALLONES)
5329 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction().getContext()), 8);
5330 else if (Opc == X86::MMX_SET0)
5331 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction().getContext()), 2);
5333 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction().getContext()), 4);
5335 bool IsAllOnes = (Opc == X86::V_SETALLONES || Opc == X86::AVX2_SETALLONES ||
5336 Opc == X86::AVX512_512_SETALLONES ||
5337 Opc == X86::AVX1_SETALLONES);
5338 const Constant *C = IsAllOnes ? Constant::getAllOnesValue(Ty) :
5339 Constant::getNullValue(Ty);
5340 unsigned CPI = MCP.getConstantPoolIndex(C, Alignment);
5342 // Create operands to load from the constant pool entry.
5343 MOs.push_back(MachineOperand::CreateReg(PICBase, false));
5344 MOs.push_back(MachineOperand::CreateImm(1));
5345 MOs.push_back(MachineOperand::CreateReg(0, false));
5346 MOs.push_back(MachineOperand::CreateCPI(CPI, 0));
5347 MOs.push_back(MachineOperand::CreateReg(0, false));
5351 if (isNonFoldablePartialRegisterLoad(LoadMI, MI, MF))
5354 // Folding a normal load. Just copy the load's address operands.
5355 MOs.append(LoadMI.operands_begin() + NumOps - X86::AddrNumOperands,
5356 LoadMI.operands_begin() + NumOps);
5360 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, InsertPt,
5361 /*Size=*/0, Alignment, /*AllowCommute=*/true);
5364 static SmallVector<MachineMemOperand *, 2>
5365 extractLoadMMOs(ArrayRef<MachineMemOperand *> MMOs, MachineFunction &MF) {
5366 SmallVector<MachineMemOperand *, 2> LoadMMOs;
5368 for (MachineMemOperand *MMO : MMOs) {
5372 if (!MMO->isStore()) {
5374 LoadMMOs.push_back(MMO);
5376 // Clone the MMO and unset the store flag.
5377 LoadMMOs.push_back(MF.getMachineMemOperand(
5378 MMO->getPointerInfo(), MMO->getFlags() & ~MachineMemOperand::MOStore,
5379 MMO->getSize(), MMO->getBaseAlignment(), MMO->getAAInfo(), nullptr,
5380 MMO->getSyncScopeID(), MMO->getOrdering(),
5381 MMO->getFailureOrdering()));
5388 static SmallVector<MachineMemOperand *, 2>
5389 extractStoreMMOs(ArrayRef<MachineMemOperand *> MMOs, MachineFunction &MF) {
5390 SmallVector<MachineMemOperand *, 2> StoreMMOs;
5392 for (MachineMemOperand *MMO : MMOs) {
5393 if (!MMO->isStore())
5396 if (!MMO->isLoad()) {
5398 StoreMMOs.push_back(MMO);
5400 // Clone the MMO and unset the load flag.
5401 StoreMMOs.push_back(MF.getMachineMemOperand(
5402 MMO->getPointerInfo(), MMO->getFlags() & ~MachineMemOperand::MOLoad,
5403 MMO->getSize(), MMO->getBaseAlignment(), MMO->getAAInfo(), nullptr,
5404 MMO->getSyncScopeID(), MMO->getOrdering(),
5405 MMO->getFailureOrdering()));
5412 bool X86InstrInfo::unfoldMemoryOperand(
5413 MachineFunction &MF, MachineInstr &MI, unsigned Reg, bool UnfoldLoad,
5414 bool UnfoldStore, SmallVectorImpl<MachineInstr *> &NewMIs) const {
5415 const X86MemoryFoldTableEntry *I = lookupUnfoldTable(MI.getOpcode());
5418 unsigned Opc = I->DstOp;
5419 unsigned Index = I->Flags & TB_INDEX_MASK;
5420 bool FoldedLoad = I->Flags & TB_FOLDED_LOAD;
5421 bool FoldedStore = I->Flags & TB_FOLDED_STORE;
5422 if (UnfoldLoad && !FoldedLoad)
5424 UnfoldLoad &= FoldedLoad;
5425 if (UnfoldStore && !FoldedStore)
5427 UnfoldStore &= FoldedStore;
5429 const MCInstrDesc &MCID = get(Opc);
5430 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
5431 // TODO: Check if 32-byte or greater accesses are slow too?
5432 if (!MI.hasOneMemOperand() && RC == &X86::VR128RegClass &&
5433 Subtarget.isUnalignedMem16Slow())
5434 // Without memoperands, loadRegFromAddr and storeRegToStackSlot will
5435 // conservatively assume the address is unaligned. That's bad for
5438 SmallVector<MachineOperand, X86::AddrNumOperands> AddrOps;
5439 SmallVector<MachineOperand,2> BeforeOps;
5440 SmallVector<MachineOperand,2> AfterOps;
5441 SmallVector<MachineOperand,4> ImpOps;
5442 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
5443 MachineOperand &Op = MI.getOperand(i);
5444 if (i >= Index && i < Index + X86::AddrNumOperands)
5445 AddrOps.push_back(Op);
5446 else if (Op.isReg() && Op.isImplicit())
5447 ImpOps.push_back(Op);
5449 BeforeOps.push_back(Op);
5451 AfterOps.push_back(Op);
5454 // Emit the load instruction.
5456 auto MMOs = extractLoadMMOs(MI.memoperands(), MF);
5457 loadRegFromAddr(MF, Reg, AddrOps, RC, MMOs, NewMIs);
5459 // Address operands cannot be marked isKill.
5460 for (unsigned i = 1; i != 1 + X86::AddrNumOperands; ++i) {
5461 MachineOperand &MO = NewMIs[0]->getOperand(i);
5463 MO.setIsKill(false);
5468 // Emit the data processing instruction.
5469 MachineInstr *DataMI = MF.CreateMachineInstr(MCID, MI.getDebugLoc(), true);
5470 MachineInstrBuilder MIB(MF, DataMI);
5473 MIB.addReg(Reg, RegState::Define);
5474 for (MachineOperand &BeforeOp : BeforeOps)
5478 for (MachineOperand &AfterOp : AfterOps)
5480 for (MachineOperand &ImpOp : ImpOps) {
5481 MIB.addReg(ImpOp.getReg(),
5482 getDefRegState(ImpOp.isDef()) |
5483 RegState::Implicit |
5484 getKillRegState(ImpOp.isKill()) |
5485 getDeadRegState(ImpOp.isDead()) |
5486 getUndefRegState(ImpOp.isUndef()));
5488 // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
5489 switch (DataMI->getOpcode()) {
5491 case X86::CMP64ri32:
5498 MachineOperand &MO0 = DataMI->getOperand(0);
5499 MachineOperand &MO1 = DataMI->getOperand(1);
5500 if (MO1.getImm() == 0) {
5502 switch (DataMI->getOpcode()) {
5503 default: llvm_unreachable("Unreachable!");
5505 case X86::CMP64ri32: NewOpc = X86::TEST64rr; break;
5507 case X86::CMP32ri: NewOpc = X86::TEST32rr; break;
5509 case X86::CMP16ri: NewOpc = X86::TEST16rr; break;
5510 case X86::CMP8ri: NewOpc = X86::TEST8rr; break;
5512 DataMI->setDesc(get(NewOpc));
5513 MO1.ChangeToRegister(MO0.getReg(), false);
5517 NewMIs.push_back(DataMI);
5519 // Emit the store instruction.
5521 const TargetRegisterClass *DstRC = getRegClass(MCID, 0, &RI, MF);
5522 auto MMOs = extractStoreMMOs(MI.memoperands(), MF);
5523 storeRegToAddr(MF, Reg, true, AddrOps, DstRC, MMOs, NewMIs);
5530 X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
5531 SmallVectorImpl<SDNode*> &NewNodes) const {
5532 if (!N->isMachineOpcode())
5535 const X86MemoryFoldTableEntry *I = lookupUnfoldTable(N->getMachineOpcode());
5538 unsigned Opc = I->DstOp;
5539 unsigned Index = I->Flags & TB_INDEX_MASK;
5540 bool FoldedLoad = I->Flags & TB_FOLDED_LOAD;
5541 bool FoldedStore = I->Flags & TB_FOLDED_STORE;
5542 const MCInstrDesc &MCID = get(Opc);
5543 MachineFunction &MF = DAG.getMachineFunction();
5544 const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
5545 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
5546 unsigned NumDefs = MCID.NumDefs;
5547 std::vector<SDValue> AddrOps;
5548 std::vector<SDValue> BeforeOps;
5549 std::vector<SDValue> AfterOps;
5551 unsigned NumOps = N->getNumOperands();
5552 for (unsigned i = 0; i != NumOps-1; ++i) {
5553 SDValue Op = N->getOperand(i);
5554 if (i >= Index-NumDefs && i < Index-NumDefs + X86::AddrNumOperands)
5555 AddrOps.push_back(Op);
5556 else if (i < Index-NumDefs)
5557 BeforeOps.push_back(Op);
5558 else if (i > Index-NumDefs)
5559 AfterOps.push_back(Op);
5561 SDValue Chain = N->getOperand(NumOps-1);
5562 AddrOps.push_back(Chain);
5564 // Emit the load instruction.
5565 SDNode *Load = nullptr;
5567 EVT VT = *TRI.legalclasstypes_begin(*RC);
5568 auto MMOs = extractLoadMMOs(cast<MachineSDNode>(N)->memoperands(), MF);
5569 if (MMOs.empty() && RC == &X86::VR128RegClass &&
5570 Subtarget.isUnalignedMem16Slow())
5571 // Do not introduce a slow unaligned load.
5573 // FIXME: If a VR128 can have size 32, we should be checking if a 32-byte
5574 // memory access is slow above.
5575 unsigned Alignment = std::max<uint32_t>(TRI.getSpillSize(*RC), 16);
5576 bool isAligned = !MMOs.empty() && MMOs.front()->getAlignment() >= Alignment;
5577 Load = DAG.getMachineNode(getLoadRegOpcode(0, RC, isAligned, Subtarget), dl,
5578 VT, MVT::Other, AddrOps);
5579 NewNodes.push_back(Load);
5581 // Preserve memory reference information.
5582 DAG.setNodeMemRefs(cast<MachineSDNode>(Load), MMOs);
5585 // Emit the data processing instruction.
5586 std::vector<EVT> VTs;
5587 const TargetRegisterClass *DstRC = nullptr;
5588 if (MCID.getNumDefs() > 0) {
5589 DstRC = getRegClass(MCID, 0, &RI, MF);
5590 VTs.push_back(*TRI.legalclasstypes_begin(*DstRC));
5592 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
5593 EVT VT = N->getValueType(i);
5594 if (VT != MVT::Other && i >= (unsigned)MCID.getNumDefs())
5598 BeforeOps.push_back(SDValue(Load, 0));
5599 BeforeOps.insert(BeforeOps.end(), AfterOps.begin(), AfterOps.end());
5600 // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
5603 case X86::CMP64ri32:
5610 if (isNullConstant(BeforeOps[1])) {
5612 default: llvm_unreachable("Unreachable!");
5614 case X86::CMP64ri32: Opc = X86::TEST64rr; break;
5616 case X86::CMP32ri: Opc = X86::TEST32rr; break;
5618 case X86::CMP16ri: Opc = X86::TEST16rr; break;
5619 case X86::CMP8ri: Opc = X86::TEST8rr; break;
5621 BeforeOps[1] = BeforeOps[0];
5624 SDNode *NewNode= DAG.getMachineNode(Opc, dl, VTs, BeforeOps);
5625 NewNodes.push_back(NewNode);
5627 // Emit the store instruction.
5630 AddrOps.push_back(SDValue(NewNode, 0));
5631 AddrOps.push_back(Chain);
5632 auto MMOs = extractStoreMMOs(cast<MachineSDNode>(N)->memoperands(), MF);
5633 if (MMOs.empty() && RC == &X86::VR128RegClass &&
5634 Subtarget.isUnalignedMem16Slow())
5635 // Do not introduce a slow unaligned store.
5637 // FIXME: If a VR128 can have size 32, we should be checking if a 32-byte
5638 // memory access is slow above.
5639 unsigned Alignment = std::max<uint32_t>(TRI.getSpillSize(*RC), 16);
5640 bool isAligned = !MMOs.empty() && MMOs.front()->getAlignment() >= Alignment;
5642 DAG.getMachineNode(getStoreRegOpcode(0, DstRC, isAligned, Subtarget),
5643 dl, MVT::Other, AddrOps);
5644 NewNodes.push_back(Store);
5646 // Preserve memory reference information.
5647 DAG.setNodeMemRefs(cast<MachineSDNode>(Store), MMOs);
5653 unsigned X86InstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc,
5654 bool UnfoldLoad, bool UnfoldStore,
5655 unsigned *LoadRegIndex) const {
5656 const X86MemoryFoldTableEntry *I = lookupUnfoldTable(Opc);
5659 bool FoldedLoad = I->Flags & TB_FOLDED_LOAD;
5660 bool FoldedStore = I->Flags & TB_FOLDED_STORE;
5661 if (UnfoldLoad && !FoldedLoad)
5663 if (UnfoldStore && !FoldedStore)
5666 *LoadRegIndex = I->Flags & TB_INDEX_MASK;
5671 X86InstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
5672 int64_t &Offset1, int64_t &Offset2) const {
5673 if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode())
5675 unsigned Opc1 = Load1->getMachineOpcode();
5676 unsigned Opc2 = Load2->getMachineOpcode();
5678 default: return false;
5688 case X86::MMX_MOVD64rm:
5689 case X86::MMX_MOVQ64rm:
5696 // AVX load instructions
5699 case X86::VMOVAPSrm:
5700 case X86::VMOVUPSrm:
5701 case X86::VMOVAPDrm:
5702 case X86::VMOVUPDrm:
5703 case X86::VMOVDQArm:
5704 case X86::VMOVDQUrm:
5705 case X86::VMOVAPSYrm:
5706 case X86::VMOVUPSYrm:
5707 case X86::VMOVAPDYrm:
5708 case X86::VMOVUPDYrm:
5709 case X86::VMOVDQAYrm:
5710 case X86::VMOVDQUYrm:
5711 // AVX512 load instructions
5712 case X86::VMOVSSZrm:
5713 case X86::VMOVSDZrm:
5714 case X86::VMOVAPSZ128rm:
5715 case X86::VMOVUPSZ128rm:
5716 case X86::VMOVAPSZ128rm_NOVLX:
5717 case X86::VMOVUPSZ128rm_NOVLX:
5718 case X86::VMOVAPDZ128rm:
5719 case X86::VMOVUPDZ128rm:
5720 case X86::VMOVDQU8Z128rm:
5721 case X86::VMOVDQU16Z128rm:
5722 case X86::VMOVDQA32Z128rm:
5723 case X86::VMOVDQU32Z128rm:
5724 case X86::VMOVDQA64Z128rm:
5725 case X86::VMOVDQU64Z128rm:
5726 case X86::VMOVAPSZ256rm:
5727 case X86::VMOVUPSZ256rm:
5728 case X86::VMOVAPSZ256rm_NOVLX:
5729 case X86::VMOVUPSZ256rm_NOVLX:
5730 case X86::VMOVAPDZ256rm:
5731 case X86::VMOVUPDZ256rm:
5732 case X86::VMOVDQU8Z256rm:
5733 case X86::VMOVDQU16Z256rm:
5734 case X86::VMOVDQA32Z256rm:
5735 case X86::VMOVDQU32Z256rm:
5736 case X86::VMOVDQA64Z256rm:
5737 case X86::VMOVDQU64Z256rm:
5738 case X86::VMOVAPSZrm:
5739 case X86::VMOVUPSZrm:
5740 case X86::VMOVAPDZrm:
5741 case X86::VMOVUPDZrm:
5742 case X86::VMOVDQU8Zrm:
5743 case X86::VMOVDQU16Zrm:
5744 case X86::VMOVDQA32Zrm:
5745 case X86::VMOVDQU32Zrm:
5746 case X86::VMOVDQA64Zrm:
5747 case X86::VMOVDQU64Zrm:
5755 default: return false;
5765 case X86::MMX_MOVD64rm:
5766 case X86::MMX_MOVQ64rm:
5773 // AVX load instructions
5776 case X86::VMOVAPSrm:
5777 case X86::VMOVUPSrm:
5778 case X86::VMOVAPDrm:
5779 case X86::VMOVUPDrm:
5780 case X86::VMOVDQArm:
5781 case X86::VMOVDQUrm:
5782 case X86::VMOVAPSYrm:
5783 case X86::VMOVUPSYrm:
5784 case X86::VMOVAPDYrm:
5785 case X86::VMOVUPDYrm:
5786 case X86::VMOVDQAYrm:
5787 case X86::VMOVDQUYrm:
5788 // AVX512 load instructions
5789 case X86::VMOVSSZrm:
5790 case X86::VMOVSDZrm:
5791 case X86::VMOVAPSZ128rm:
5792 case X86::VMOVUPSZ128rm:
5793 case X86::VMOVAPSZ128rm_NOVLX:
5794 case X86::VMOVUPSZ128rm_NOVLX:
5795 case X86::VMOVAPDZ128rm:
5796 case X86::VMOVUPDZ128rm:
5797 case X86::VMOVDQU8Z128rm:
5798 case X86::VMOVDQU16Z128rm:
5799 case X86::VMOVDQA32Z128rm:
5800 case X86::VMOVDQU32Z128rm:
5801 case X86::VMOVDQA64Z128rm:
5802 case X86::VMOVDQU64Z128rm:
5803 case X86::VMOVAPSZ256rm:
5804 case X86::VMOVUPSZ256rm:
5805 case X86::VMOVAPSZ256rm_NOVLX:
5806 case X86::VMOVUPSZ256rm_NOVLX:
5807 case X86::VMOVAPDZ256rm:
5808 case X86::VMOVUPDZ256rm:
5809 case X86::VMOVDQU8Z256rm:
5810 case X86::VMOVDQU16Z256rm:
5811 case X86::VMOVDQA32Z256rm:
5812 case X86::VMOVDQU32Z256rm:
5813 case X86::VMOVDQA64Z256rm:
5814 case X86::VMOVDQU64Z256rm:
5815 case X86::VMOVAPSZrm:
5816 case X86::VMOVUPSZrm:
5817 case X86::VMOVAPDZrm:
5818 case X86::VMOVUPDZrm:
5819 case X86::VMOVDQU8Zrm:
5820 case X86::VMOVDQU16Zrm:
5821 case X86::VMOVDQA32Zrm:
5822 case X86::VMOVDQU32Zrm:
5823 case X86::VMOVDQA64Zrm:
5824 case X86::VMOVDQU64Zrm:
5832 // Lambda to check if both the loads have the same value for an operand index.
5833 auto HasSameOp = [&](int I) {
5834 return Load1->getOperand(I) == Load2->getOperand(I);
5837 // All operands except the displacement should match.
5838 if (!HasSameOp(X86::AddrBaseReg) || !HasSameOp(X86::AddrScaleAmt) ||
5839 !HasSameOp(X86::AddrIndexReg) || !HasSameOp(X86::AddrSegmentReg))
5842 // Chain Operand must be the same.
5846 // Now let's examine if the displacements are constants.
5847 auto Disp1 = dyn_cast<ConstantSDNode>(Load1->getOperand(X86::AddrDisp));
5848 auto Disp2 = dyn_cast<ConstantSDNode>(Load2->getOperand(X86::AddrDisp));
5849 if (!Disp1 || !Disp2)
5852 Offset1 = Disp1->getSExtValue();
5853 Offset2 = Disp2->getSExtValue();
5857 bool X86InstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
5858 int64_t Offset1, int64_t Offset2,
5859 unsigned NumLoads) const {
5860 assert(Offset2 > Offset1);
5861 if ((Offset2 - Offset1) / 8 > 64)
5864 unsigned Opc1 = Load1->getMachineOpcode();
5865 unsigned Opc2 = Load2->getMachineOpcode();
5867 return false; // FIXME: overly conservative?
5874 case X86::MMX_MOVD64rm:
5875 case X86::MMX_MOVQ64rm:
5879 EVT VT = Load1->getValueType(0);
5880 switch (VT.getSimpleVT().SimpleTy) {
5882 // XMM registers. In 64-bit mode we can be a bit more aggressive since we
5883 // have 16 of them to play with.
5884 if (Subtarget.is64Bit()) {
5887 } else if (NumLoads) {
5906 reverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
5907 assert(Cond.size() == 1 && "Invalid X86 branch condition!");
5908 X86::CondCode CC = static_cast<X86::CondCode>(Cond[0].getImm());
5909 Cond[0].setImm(GetOppositeBranchCondition(CC));
5914 isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
5915 // FIXME: Return false for x87 stack register classes for now. We can't
5916 // allow any loads of these registers before FpGet_ST0_80.
5917 return !(RC == &X86::CCRRegClass || RC == &X86::DFCCRRegClass ||
5918 RC == &X86::RFP32RegClass || RC == &X86::RFP64RegClass ||
5919 RC == &X86::RFP80RegClass);
5922 /// Return a virtual register initialized with the
5923 /// the global base register value. Output instructions required to
5924 /// initialize the register in the function entry block, if necessary.
5926 /// TODO: Eliminate this and move the code to X86MachineFunctionInfo.
5928 unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const {
5929 assert((!Subtarget.is64Bit() ||
5930 MF->getTarget().getCodeModel() == CodeModel::Medium ||
5931 MF->getTarget().getCodeModel() == CodeModel::Large) &&
5932 "X86-64 PIC uses RIP relative addressing");
5934 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
5935 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
5936 if (GlobalBaseReg != 0)
5937 return GlobalBaseReg;
5939 // Create the register. The code to initialize it is inserted
5940 // later, by the CGBR pass (below).
5941 MachineRegisterInfo &RegInfo = MF->getRegInfo();
5942 GlobalBaseReg = RegInfo.createVirtualRegister(
5943 Subtarget.is64Bit() ? &X86::GR64_NOSPRegClass : &X86::GR32_NOSPRegClass);
5944 X86FI->setGlobalBaseReg(GlobalBaseReg);
5945 return GlobalBaseReg;
5948 // These are the replaceable SSE instructions. Some of these have Int variants
5949 // that we don't include here. We don't want to replace instructions selected
5951 static const uint16_t ReplaceableInstrs[][3] = {
5952 //PackedSingle PackedDouble PackedInt
5953 { X86::MOVAPSmr, X86::MOVAPDmr, X86::MOVDQAmr },
5954 { X86::MOVAPSrm, X86::MOVAPDrm, X86::MOVDQArm },
5955 { X86::MOVAPSrr, X86::MOVAPDrr, X86::MOVDQArr },
5956 { X86::MOVUPSmr, X86::MOVUPDmr, X86::MOVDQUmr },
5957 { X86::MOVUPSrm, X86::MOVUPDrm, X86::MOVDQUrm },
5958 { X86::MOVLPSmr, X86::MOVLPDmr, X86::MOVPQI2QImr },
5959 { X86::MOVSDmr, X86::MOVSDmr, X86::MOVPQI2QImr },
5960 { X86::MOVSSmr, X86::MOVSSmr, X86::MOVPDI2DImr },
5961 { X86::MOVSDrm, X86::MOVSDrm, X86::MOVQI2PQIrm },
5962 { X86::MOVSSrm, X86::MOVSSrm, X86::MOVDI2PDIrm },
5963 { X86::MOVNTPSmr, X86::MOVNTPDmr, X86::MOVNTDQmr },
5964 { X86::ANDNPSrm, X86::ANDNPDrm, X86::PANDNrm },
5965 { X86::ANDNPSrr, X86::ANDNPDrr, X86::PANDNrr },
5966 { X86::ANDPSrm, X86::ANDPDrm, X86::PANDrm },
5967 { X86::ANDPSrr, X86::ANDPDrr, X86::PANDrr },
5968 { X86::ORPSrm, X86::ORPDrm, X86::PORrm },
5969 { X86::ORPSrr, X86::ORPDrr, X86::PORrr },
5970 { X86::XORPSrm, X86::XORPDrm, X86::PXORrm },
5971 { X86::XORPSrr, X86::XORPDrr, X86::PXORrr },
5972 { X86::UNPCKLPDrm, X86::UNPCKLPDrm, X86::PUNPCKLQDQrm },
5973 { X86::MOVLHPSrr, X86::UNPCKLPDrr, X86::PUNPCKLQDQrr },
5974 { X86::UNPCKHPDrm, X86::UNPCKHPDrm, X86::PUNPCKHQDQrm },
5975 { X86::UNPCKHPDrr, X86::UNPCKHPDrr, X86::PUNPCKHQDQrr },
5976 { X86::UNPCKLPSrm, X86::UNPCKLPSrm, X86::PUNPCKLDQrm },
5977 { X86::UNPCKLPSrr, X86::UNPCKLPSrr, X86::PUNPCKLDQrr },
5978 { X86::UNPCKHPSrm, X86::UNPCKHPSrm, X86::PUNPCKHDQrm },
5979 { X86::UNPCKHPSrr, X86::UNPCKHPSrr, X86::PUNPCKHDQrr },
5980 { X86::EXTRACTPSmr, X86::EXTRACTPSmr, X86::PEXTRDmr },
5981 { X86::EXTRACTPSrr, X86::EXTRACTPSrr, X86::PEXTRDrr },
5982 // AVX 128-bit support
5983 { X86::VMOVAPSmr, X86::VMOVAPDmr, X86::VMOVDQAmr },
5984 { X86::VMOVAPSrm, X86::VMOVAPDrm, X86::VMOVDQArm },
5985 { X86::VMOVAPSrr, X86::VMOVAPDrr, X86::VMOVDQArr },
5986 { X86::VMOVUPSmr, X86::VMOVUPDmr, X86::VMOVDQUmr },
5987 { X86::VMOVUPSrm, X86::VMOVUPDrm, X86::VMOVDQUrm },
5988 { X86::VMOVLPSmr, X86::VMOVLPDmr, X86::VMOVPQI2QImr },
5989 { X86::VMOVSDmr, X86::VMOVSDmr, X86::VMOVPQI2QImr },
5990 { X86::VMOVSSmr, X86::VMOVSSmr, X86::VMOVPDI2DImr },
5991 { X86::VMOVSDrm, X86::VMOVSDrm, X86::VMOVQI2PQIrm },
5992 { X86::VMOVSSrm, X86::VMOVSSrm, X86::VMOVDI2PDIrm },
5993 { X86::VMOVNTPSmr, X86::VMOVNTPDmr, X86::VMOVNTDQmr },
5994 { X86::VANDNPSrm, X86::VANDNPDrm, X86::VPANDNrm },
5995 { X86::VANDNPSrr, X86::VANDNPDrr, X86::VPANDNrr },
5996 { X86::VANDPSrm, X86::VANDPDrm, X86::VPANDrm },
5997 { X86::VANDPSrr, X86::VANDPDrr, X86::VPANDrr },
5998 { X86::VORPSrm, X86::VORPDrm, X86::VPORrm },
5999 { X86::VORPSrr, X86::VORPDrr, X86::VPORrr },
6000 { X86::VXORPSrm, X86::VXORPDrm, X86::VPXORrm },
6001 { X86::VXORPSrr, X86::VXORPDrr, X86::VPXORrr },
6002 { X86::VUNPCKLPDrm, X86::VUNPCKLPDrm, X86::VPUNPCKLQDQrm },
6003 { X86::VMOVLHPSrr, X86::VUNPCKLPDrr, X86::VPUNPCKLQDQrr },
6004 { X86::VUNPCKHPDrm, X86::VUNPCKHPDrm, X86::VPUNPCKHQDQrm },
6005 { X86::VUNPCKHPDrr, X86::VUNPCKHPDrr, X86::VPUNPCKHQDQrr },
6006 { X86::VUNPCKLPSrm, X86::VUNPCKLPSrm, X86::VPUNPCKLDQrm },
6007 { X86::VUNPCKLPSrr, X86::VUNPCKLPSrr, X86::VPUNPCKLDQrr },
6008 { X86::VUNPCKHPSrm, X86::VUNPCKHPSrm, X86::VPUNPCKHDQrm },
6009 { X86::VUNPCKHPSrr, X86::VUNPCKHPSrr, X86::VPUNPCKHDQrr },
6010 { X86::VEXTRACTPSmr, X86::VEXTRACTPSmr, X86::VPEXTRDmr },
6011 { X86::VEXTRACTPSrr, X86::VEXTRACTPSrr, X86::VPEXTRDrr },
6012 // AVX 256-bit support
6013 { X86::VMOVAPSYmr, X86::VMOVAPDYmr, X86::VMOVDQAYmr },
6014 { X86::VMOVAPSYrm, X86::VMOVAPDYrm, X86::VMOVDQAYrm },
6015 { X86::VMOVAPSYrr, X86::VMOVAPDYrr, X86::VMOVDQAYrr },
6016 { X86::VMOVUPSYmr, X86::VMOVUPDYmr, X86::VMOVDQUYmr },
6017 { X86::VMOVUPSYrm, X86::VMOVUPDYrm, X86::VMOVDQUYrm },
6018 { X86::VMOVNTPSYmr, X86::VMOVNTPDYmr, X86::VMOVNTDQYmr },
6019 { X86::VPERMPSYrm, X86::VPERMPSYrm, X86::VPERMDYrm },
6020 { X86::VPERMPSYrr, X86::VPERMPSYrr, X86::VPERMDYrr },
6021 { X86::VPERMPDYmi, X86::VPERMPDYmi, X86::VPERMQYmi },
6022 { X86::VPERMPDYri, X86::VPERMPDYri, X86::VPERMQYri },
6024 { X86::VMOVLPSZ128mr, X86::VMOVLPDZ128mr, X86::VMOVPQI2QIZmr },
6025 { X86::VMOVNTPSZ128mr, X86::VMOVNTPDZ128mr, X86::VMOVNTDQZ128mr },
6026 { X86::VMOVNTPSZ256mr, X86::VMOVNTPDZ256mr, X86::VMOVNTDQZ256mr },
6027 { X86::VMOVNTPSZmr, X86::VMOVNTPDZmr, X86::VMOVNTDQZmr },
6028 { X86::VMOVSDZmr, X86::VMOVSDZmr, X86::VMOVPQI2QIZmr },
6029 { X86::VMOVSSZmr, X86::VMOVSSZmr, X86::VMOVPDI2DIZmr },
6030 { X86::VMOVSDZrm, X86::VMOVSDZrm, X86::VMOVQI2PQIZrm },
6031 { X86::VMOVSSZrm, X86::VMOVSSZrm, X86::VMOVDI2PDIZrm },
6032 { X86::VBROADCASTSSZ128r, X86::VBROADCASTSSZ128r, X86::VPBROADCASTDZ128r },
6033 { X86::VBROADCASTSSZ128m, X86::VBROADCASTSSZ128m, X86::VPBROADCASTDZ128m },
6034 { X86::VBROADCASTSSZ256r, X86::VBROADCASTSSZ256r, X86::VPBROADCASTDZ256r },
6035 { X86::VBROADCASTSSZ256m, X86::VBROADCASTSSZ256m, X86::VPBROADCASTDZ256m },
6036 { X86::VBROADCASTSSZr, X86::VBROADCASTSSZr, X86::VPBROADCASTDZr },
6037 { X86::VBROADCASTSSZm, X86::VBROADCASTSSZm, X86::VPBROADCASTDZm },
6038 { X86::VBROADCASTSDZ256r, X86::VBROADCASTSDZ256r, X86::VPBROADCASTQZ256r },
6039 { X86::VBROADCASTSDZ256m, X86::VBROADCASTSDZ256m, X86::VPBROADCASTQZ256m },
6040 { X86::VBROADCASTSDZr, X86::VBROADCASTSDZr, X86::VPBROADCASTQZr },
6041 { X86::VBROADCASTSDZm, X86::VBROADCASTSDZm, X86::VPBROADCASTQZm },
6042 { X86::VINSERTF32x4Zrr, X86::VINSERTF32x4Zrr, X86::VINSERTI32x4Zrr },
6043 { X86::VINSERTF32x4Zrm, X86::VINSERTF32x4Zrm, X86::VINSERTI32x4Zrm },
6044 { X86::VINSERTF32x8Zrr, X86::VINSERTF32x8Zrr, X86::VINSERTI32x8Zrr },
6045 { X86::VINSERTF32x8Zrm, X86::VINSERTF32x8Zrm, X86::VINSERTI32x8Zrm },
6046 { X86::VINSERTF64x2Zrr, X86::VINSERTF64x2Zrr, X86::VINSERTI64x2Zrr },
6047 { X86::VINSERTF64x2Zrm, X86::VINSERTF64x2Zrm, X86::VINSERTI64x2Zrm },
6048 { X86::VINSERTF64x4Zrr, X86::VINSERTF64x4Zrr, X86::VINSERTI64x4Zrr },
6049 { X86::VINSERTF64x4Zrm, X86::VINSERTF64x4Zrm, X86::VINSERTI64x4Zrm },
6050 { X86::VINSERTF32x4Z256rr,X86::VINSERTF32x4Z256rr,X86::VINSERTI32x4Z256rr },
6051 { X86::VINSERTF32x4Z256rm,X86::VINSERTF32x4Z256rm,X86::VINSERTI32x4Z256rm },
6052 { X86::VINSERTF64x2Z256rr,X86::VINSERTF64x2Z256rr,X86::VINSERTI64x2Z256rr },
6053 { X86::VINSERTF64x2Z256rm,X86::VINSERTF64x2Z256rm,X86::VINSERTI64x2Z256rm },
6054 { X86::VEXTRACTF32x4Zrr, X86::VEXTRACTF32x4Zrr, X86::VEXTRACTI32x4Zrr },
6055 { X86::VEXTRACTF32x4Zmr, X86::VEXTRACTF32x4Zmr, X86::VEXTRACTI32x4Zmr },
6056 { X86::VEXTRACTF32x8Zrr, X86::VEXTRACTF32x8Zrr, X86::VEXTRACTI32x8Zrr },
6057 { X86::VEXTRACTF32x8Zmr, X86::VEXTRACTF32x8Zmr, X86::VEXTRACTI32x8Zmr },
6058 { X86::VEXTRACTF64x2Zrr, X86::VEXTRACTF64x2Zrr, X86::VEXTRACTI64x2Zrr },
6059 { X86::VEXTRACTF64x2Zmr, X86::VEXTRACTF64x2Zmr, X86::VEXTRACTI64x2Zmr },
6060 { X86::VEXTRACTF64x4Zrr, X86::VEXTRACTF64x4Zrr, X86::VEXTRACTI64x4Zrr },
6061 { X86::VEXTRACTF64x4Zmr, X86::VEXTRACTF64x4Zmr, X86::VEXTRACTI64x4Zmr },
6062 { X86::VEXTRACTF32x4Z256rr,X86::VEXTRACTF32x4Z256rr,X86::VEXTRACTI32x4Z256rr },
6063 { X86::VEXTRACTF32x4Z256mr,X86::VEXTRACTF32x4Z256mr,X86::VEXTRACTI32x4Z256mr },
6064 { X86::VEXTRACTF64x2Z256rr,X86::VEXTRACTF64x2Z256rr,X86::VEXTRACTI64x2Z256rr },
6065 { X86::VEXTRACTF64x2Z256mr,X86::VEXTRACTF64x2Z256mr,X86::VEXTRACTI64x2Z256mr },
6066 { X86::VPERMILPSmi, X86::VPERMILPSmi, X86::VPSHUFDmi },
6067 { X86::VPERMILPSri, X86::VPERMILPSri, X86::VPSHUFDri },
6068 { X86::VPERMILPSZ128mi, X86::VPERMILPSZ128mi, X86::VPSHUFDZ128mi },
6069 { X86::VPERMILPSZ128ri, X86::VPERMILPSZ128ri, X86::VPSHUFDZ128ri },
6070 { X86::VPERMILPSZ256mi, X86::VPERMILPSZ256mi, X86::VPSHUFDZ256mi },
6071 { X86::VPERMILPSZ256ri, X86::VPERMILPSZ256ri, X86::VPSHUFDZ256ri },
6072 { X86::VPERMILPSZmi, X86::VPERMILPSZmi, X86::VPSHUFDZmi },
6073 { X86::VPERMILPSZri, X86::VPERMILPSZri, X86::VPSHUFDZri },
6074 { X86::VPERMPSZ256rm, X86::VPERMPSZ256rm, X86::VPERMDZ256rm },
6075 { X86::VPERMPSZ256rr, X86::VPERMPSZ256rr, X86::VPERMDZ256rr },
6076 { X86::VPERMPDZ256mi, X86::VPERMPDZ256mi, X86::VPERMQZ256mi },
6077 { X86::VPERMPDZ256ri, X86::VPERMPDZ256ri, X86::VPERMQZ256ri },
6078 { X86::VPERMPDZ256rm, X86::VPERMPDZ256rm, X86::VPERMQZ256rm },
6079 { X86::VPERMPDZ256rr, X86::VPERMPDZ256rr, X86::VPERMQZ256rr },
6080 { X86::VPERMPSZrm, X86::VPERMPSZrm, X86::VPERMDZrm },
6081 { X86::VPERMPSZrr, X86::VPERMPSZrr, X86::VPERMDZrr },
6082 { X86::VPERMPDZmi, X86::VPERMPDZmi, X86::VPERMQZmi },
6083 { X86::VPERMPDZri, X86::VPERMPDZri, X86::VPERMQZri },
6084 { X86::VPERMPDZrm, X86::VPERMPDZrm, X86::VPERMQZrm },
6085 { X86::VPERMPDZrr, X86::VPERMPDZrr, X86::VPERMQZrr },
6086 { X86::VUNPCKLPDZ256rm, X86::VUNPCKLPDZ256rm, X86::VPUNPCKLQDQZ256rm },
6087 { X86::VUNPCKLPDZ256rr, X86::VUNPCKLPDZ256rr, X86::VPUNPCKLQDQZ256rr },
6088 { X86::VUNPCKHPDZ256rm, X86::VUNPCKHPDZ256rm, X86::VPUNPCKHQDQZ256rm },
6089 { X86::VUNPCKHPDZ256rr, X86::VUNPCKHPDZ256rr, X86::VPUNPCKHQDQZ256rr },
6090 { X86::VUNPCKLPSZ256rm, X86::VUNPCKLPSZ256rm, X86::VPUNPCKLDQZ256rm },
6091 { X86::VUNPCKLPSZ256rr, X86::VUNPCKLPSZ256rr, X86::VPUNPCKLDQZ256rr },
6092 { X86::VUNPCKHPSZ256rm, X86::VUNPCKHPSZ256rm, X86::VPUNPCKHDQZ256rm },
6093 { X86::VUNPCKHPSZ256rr, X86::VUNPCKHPSZ256rr, X86::VPUNPCKHDQZ256rr },
6094 { X86::VUNPCKLPDZ128rm, X86::VUNPCKLPDZ128rm, X86::VPUNPCKLQDQZ128rm },
6095 { X86::VMOVLHPSZrr, X86::VUNPCKLPDZ128rr, X86::VPUNPCKLQDQZ128rr },
6096 { X86::VUNPCKHPDZ128rm, X86::VUNPCKHPDZ128rm, X86::VPUNPCKHQDQZ128rm },
6097 { X86::VUNPCKHPDZ128rr, X86::VUNPCKHPDZ128rr, X86::VPUNPCKHQDQZ128rr },
6098 { X86::VUNPCKLPSZ128rm, X86::VUNPCKLPSZ128rm, X86::VPUNPCKLDQZ128rm },
6099 { X86::VUNPCKLPSZ128rr, X86::VUNPCKLPSZ128rr, X86::VPUNPCKLDQZ128rr },
6100 { X86::VUNPCKHPSZ128rm, X86::VUNPCKHPSZ128rm, X86::VPUNPCKHDQZ128rm },
6101 { X86::VUNPCKHPSZ128rr, X86::VUNPCKHPSZ128rr, X86::VPUNPCKHDQZ128rr },
6102 { X86::VUNPCKLPDZrm, X86::VUNPCKLPDZrm, X86::VPUNPCKLQDQZrm },
6103 { X86::VUNPCKLPDZrr, X86::VUNPCKLPDZrr, X86::VPUNPCKLQDQZrr },
6104 { X86::VUNPCKHPDZrm, X86::VUNPCKHPDZrm, X86::VPUNPCKHQDQZrm },
6105 { X86::VUNPCKHPDZrr, X86::VUNPCKHPDZrr, X86::VPUNPCKHQDQZrr },
6106 { X86::VUNPCKLPSZrm, X86::VUNPCKLPSZrm, X86::VPUNPCKLDQZrm },
6107 { X86::VUNPCKLPSZrr, X86::VUNPCKLPSZrr, X86::VPUNPCKLDQZrr },
6108 { X86::VUNPCKHPSZrm, X86::VUNPCKHPSZrm, X86::VPUNPCKHDQZrm },
6109 { X86::VUNPCKHPSZrr, X86::VUNPCKHPSZrr, X86::VPUNPCKHDQZrr },
6110 { X86::VEXTRACTPSZmr, X86::VEXTRACTPSZmr, X86::VPEXTRDZmr },
6111 { X86::VEXTRACTPSZrr, X86::VEXTRACTPSZrr, X86::VPEXTRDZrr },
6114 static const uint16_t ReplaceableInstrsAVX2[][3] = {
6115 //PackedSingle PackedDouble PackedInt
6116 { X86::VANDNPSYrm, X86::VANDNPDYrm, X86::VPANDNYrm },
6117 { X86::VANDNPSYrr, X86::VANDNPDYrr, X86::VPANDNYrr },
6118 { X86::VANDPSYrm, X86::VANDPDYrm, X86::VPANDYrm },
6119 { X86::VANDPSYrr, X86::VANDPDYrr, X86::VPANDYrr },
6120 { X86::VORPSYrm, X86::VORPDYrm, X86::VPORYrm },
6121 { X86::VORPSYrr, X86::VORPDYrr, X86::VPORYrr },
6122 { X86::VXORPSYrm, X86::VXORPDYrm, X86::VPXORYrm },
6123 { X86::VXORPSYrr, X86::VXORPDYrr, X86::VPXORYrr },
6124 { X86::VPERM2F128rm, X86::VPERM2F128rm, X86::VPERM2I128rm },
6125 { X86::VPERM2F128rr, X86::VPERM2F128rr, X86::VPERM2I128rr },
6126 { X86::VBROADCASTSSrm, X86::VBROADCASTSSrm, X86::VPBROADCASTDrm},
6127 { X86::VBROADCASTSSrr, X86::VBROADCASTSSrr, X86::VPBROADCASTDrr},
6128 { X86::VBROADCASTSSYrr, X86::VBROADCASTSSYrr, X86::VPBROADCASTDYrr},
6129 { X86::VBROADCASTSSYrm, X86::VBROADCASTSSYrm, X86::VPBROADCASTDYrm},
6130 { X86::VBROADCASTSDYrr, X86::VBROADCASTSDYrr, X86::VPBROADCASTQYrr},
6131 { X86::VBROADCASTSDYrm, X86::VBROADCASTSDYrm, X86::VPBROADCASTQYrm},
6132 { X86::VBROADCASTF128, X86::VBROADCASTF128, X86::VBROADCASTI128 },
6133 { X86::VBLENDPSYrri, X86::VBLENDPSYrri, X86::VPBLENDDYrri },
6134 { X86::VBLENDPSYrmi, X86::VBLENDPSYrmi, X86::VPBLENDDYrmi },
6135 { X86::VPERMILPSYmi, X86::VPERMILPSYmi, X86::VPSHUFDYmi },
6136 { X86::VPERMILPSYri, X86::VPERMILPSYri, X86::VPSHUFDYri },
6137 { X86::VUNPCKLPDYrm, X86::VUNPCKLPDYrm, X86::VPUNPCKLQDQYrm },
6138 { X86::VUNPCKLPDYrr, X86::VUNPCKLPDYrr, X86::VPUNPCKLQDQYrr },
6139 { X86::VUNPCKHPDYrm, X86::VUNPCKHPDYrm, X86::VPUNPCKHQDQYrm },
6140 { X86::VUNPCKHPDYrr, X86::VUNPCKHPDYrr, X86::VPUNPCKHQDQYrr },
6141 { X86::VUNPCKLPSYrm, X86::VUNPCKLPSYrm, X86::VPUNPCKLDQYrm },
6142 { X86::VUNPCKLPSYrr, X86::VUNPCKLPSYrr, X86::VPUNPCKLDQYrr },
6143 { X86::VUNPCKHPSYrm, X86::VUNPCKHPSYrm, X86::VPUNPCKHDQYrm },
6144 { X86::VUNPCKHPSYrr, X86::VUNPCKHPSYrr, X86::VPUNPCKHDQYrr },
6147 static const uint16_t ReplaceableInstrsAVX2InsertExtract[][3] = {
6148 //PackedSingle PackedDouble PackedInt
6149 { X86::VEXTRACTF128mr, X86::VEXTRACTF128mr, X86::VEXTRACTI128mr },
6150 { X86::VEXTRACTF128rr, X86::VEXTRACTF128rr, X86::VEXTRACTI128rr },
6151 { X86::VINSERTF128rm, X86::VINSERTF128rm, X86::VINSERTI128rm },
6152 { X86::VINSERTF128rr, X86::VINSERTF128rr, X86::VINSERTI128rr },
6155 static const uint16_t ReplaceableInstrsAVX512[][4] = {
6156 // Two integer columns for 64-bit and 32-bit elements.
6157 //PackedSingle PackedDouble PackedInt PackedInt
6158 { X86::VMOVAPSZ128mr, X86::VMOVAPDZ128mr, X86::VMOVDQA64Z128mr, X86::VMOVDQA32Z128mr },
6159 { X86::VMOVAPSZ128rm, X86::VMOVAPDZ128rm, X86::VMOVDQA64Z128rm, X86::VMOVDQA32Z128rm },
6160 { X86::VMOVAPSZ128rr, X86::VMOVAPDZ128rr, X86::VMOVDQA64Z128rr, X86::VMOVDQA32Z128rr },
6161 { X86::VMOVUPSZ128mr, X86::VMOVUPDZ128mr, X86::VMOVDQU64Z128mr, X86::VMOVDQU32Z128mr },
6162 { X86::VMOVUPSZ128rm, X86::VMOVUPDZ128rm, X86::VMOVDQU64Z128rm, X86::VMOVDQU32Z128rm },
6163 { X86::VMOVAPSZ256mr, X86::VMOVAPDZ256mr, X86::VMOVDQA64Z256mr, X86::VMOVDQA32Z256mr },
6164 { X86::VMOVAPSZ256rm, X86::VMOVAPDZ256rm, X86::VMOVDQA64Z256rm, X86::VMOVDQA32Z256rm },
6165 { X86::VMOVAPSZ256rr, X86::VMOVAPDZ256rr, X86::VMOVDQA64Z256rr, X86::VMOVDQA32Z256rr },
6166 { X86::VMOVUPSZ256mr, X86::VMOVUPDZ256mr, X86::VMOVDQU64Z256mr, X86::VMOVDQU32Z256mr },
6167 { X86::VMOVUPSZ256rm, X86::VMOVUPDZ256rm, X86::VMOVDQU64Z256rm, X86::VMOVDQU32Z256rm },
6168 { X86::VMOVAPSZmr, X86::VMOVAPDZmr, X86::VMOVDQA64Zmr, X86::VMOVDQA32Zmr },
6169 { X86::VMOVAPSZrm, X86::VMOVAPDZrm, X86::VMOVDQA64Zrm, X86::VMOVDQA32Zrm },
6170 { X86::VMOVAPSZrr, X86::VMOVAPDZrr, X86::VMOVDQA64Zrr, X86::VMOVDQA32Zrr },
6171 { X86::VMOVUPSZmr, X86::VMOVUPDZmr, X86::VMOVDQU64Zmr, X86::VMOVDQU32Zmr },
6172 { X86::VMOVUPSZrm, X86::VMOVUPDZrm, X86::VMOVDQU64Zrm, X86::VMOVDQU32Zrm },
6175 static const uint16_t ReplaceableInstrsAVX512DQ[][4] = {
6176 // Two integer columns for 64-bit and 32-bit elements.
6177 //PackedSingle PackedDouble PackedInt PackedInt
6178 { X86::VANDNPSZ128rm, X86::VANDNPDZ128rm, X86::VPANDNQZ128rm, X86::VPANDNDZ128rm },
6179 { X86::VANDNPSZ128rr, X86::VANDNPDZ128rr, X86::VPANDNQZ128rr, X86::VPANDNDZ128rr },
6180 { X86::VANDPSZ128rm, X86::VANDPDZ128rm, X86::VPANDQZ128rm, X86::VPANDDZ128rm },
6181 { X86::VANDPSZ128rr, X86::VANDPDZ128rr, X86::VPANDQZ128rr, X86::VPANDDZ128rr },
6182 { X86::VORPSZ128rm, X86::VORPDZ128rm, X86::VPORQZ128rm, X86::VPORDZ128rm },
6183 { X86::VORPSZ128rr, X86::VORPDZ128rr, X86::VPORQZ128rr, X86::VPORDZ128rr },
6184 { X86::VXORPSZ128rm, X86::VXORPDZ128rm, X86::VPXORQZ128rm, X86::VPXORDZ128rm },
6185 { X86::VXORPSZ128rr, X86::VXORPDZ128rr, X86::VPXORQZ128rr, X86::VPXORDZ128rr },
6186 { X86::VANDNPSZ256rm, X86::VANDNPDZ256rm, X86::VPANDNQZ256rm, X86::VPANDNDZ256rm },
6187 { X86::VANDNPSZ256rr, X86::VANDNPDZ256rr, X86::VPANDNQZ256rr, X86::VPANDNDZ256rr },
6188 { X86::VANDPSZ256rm, X86::VANDPDZ256rm, X86::VPANDQZ256rm, X86::VPANDDZ256rm },
6189 { X86::VANDPSZ256rr, X86::VANDPDZ256rr, X86::VPANDQZ256rr, X86::VPANDDZ256rr },
6190 { X86::VORPSZ256rm, X86::VORPDZ256rm, X86::VPORQZ256rm, X86::VPORDZ256rm },
6191 { X86::VORPSZ256rr, X86::VORPDZ256rr, X86::VPORQZ256rr, X86::VPORDZ256rr },
6192 { X86::VXORPSZ256rm, X86::VXORPDZ256rm, X86::VPXORQZ256rm, X86::VPXORDZ256rm },
6193 { X86::VXORPSZ256rr, X86::VXORPDZ256rr, X86::VPXORQZ256rr, X86::VPXORDZ256rr },
6194 { X86::VANDNPSZrm, X86::VANDNPDZrm, X86::VPANDNQZrm, X86::VPANDNDZrm },
6195 { X86::VANDNPSZrr, X86::VANDNPDZrr, X86::VPANDNQZrr, X86::VPANDNDZrr },
6196 { X86::VANDPSZrm, X86::VANDPDZrm, X86::VPANDQZrm, X86::VPANDDZrm },
6197 { X86::VANDPSZrr, X86::VANDPDZrr, X86::VPANDQZrr, X86::VPANDDZrr },
6198 { X86::VORPSZrm, X86::VORPDZrm, X86::VPORQZrm, X86::VPORDZrm },
6199 { X86::VORPSZrr, X86::VORPDZrr, X86::VPORQZrr, X86::VPORDZrr },
6200 { X86::VXORPSZrm, X86::VXORPDZrm, X86::VPXORQZrm, X86::VPXORDZrm },
6201 { X86::VXORPSZrr, X86::VXORPDZrr, X86::VPXORQZrr, X86::VPXORDZrr },
6204 static const uint16_t ReplaceableInstrsAVX512DQMasked[][4] = {
6205 // Two integer columns for 64-bit and 32-bit elements.
6206 //PackedSingle PackedDouble
6207 //PackedInt PackedInt
6208 { X86::VANDNPSZ128rmk, X86::VANDNPDZ128rmk,
6209 X86::VPANDNQZ128rmk, X86::VPANDNDZ128rmk },
6210 { X86::VANDNPSZ128rmkz, X86::VANDNPDZ128rmkz,
6211 X86::VPANDNQZ128rmkz, X86::VPANDNDZ128rmkz },
6212 { X86::VANDNPSZ128rrk, X86::VANDNPDZ128rrk,
6213 X86::VPANDNQZ128rrk, X86::VPANDNDZ128rrk },
6214 { X86::VANDNPSZ128rrkz, X86::VANDNPDZ128rrkz,
6215 X86::VPANDNQZ128rrkz, X86::VPANDNDZ128rrkz },
6216 { X86::VANDPSZ128rmk, X86::VANDPDZ128rmk,
6217 X86::VPANDQZ128rmk, X86::VPANDDZ128rmk },
6218 { X86::VANDPSZ128rmkz, X86::VANDPDZ128rmkz,
6219 X86::VPANDQZ128rmkz, X86::VPANDDZ128rmkz },
6220 { X86::VANDPSZ128rrk, X86::VANDPDZ128rrk,
6221 X86::VPANDQZ128rrk, X86::VPANDDZ128rrk },
6222 { X86::VANDPSZ128rrkz, X86::VANDPDZ128rrkz,
6223 X86::VPANDQZ128rrkz, X86::VPANDDZ128rrkz },
6224 { X86::VORPSZ128rmk, X86::VORPDZ128rmk,
6225 X86::VPORQZ128rmk, X86::VPORDZ128rmk },
6226 { X86::VORPSZ128rmkz, X86::VORPDZ128rmkz,
6227 X86::VPORQZ128rmkz, X86::VPORDZ128rmkz },
6228 { X86::VORPSZ128rrk, X86::VORPDZ128rrk,
6229 X86::VPORQZ128rrk, X86::VPORDZ128rrk },
6230 { X86::VORPSZ128rrkz, X86::VORPDZ128rrkz,
6231 X86::VPORQZ128rrkz, X86::VPORDZ128rrkz },
6232 { X86::VXORPSZ128rmk, X86::VXORPDZ128rmk,
6233 X86::VPXORQZ128rmk, X86::VPXORDZ128rmk },
6234 { X86::VXORPSZ128rmkz, X86::VXORPDZ128rmkz,
6235 X86::VPXORQZ128rmkz, X86::VPXORDZ128rmkz },
6236 { X86::VXORPSZ128rrk, X86::VXORPDZ128rrk,
6237 X86::VPXORQZ128rrk, X86::VPXORDZ128rrk },
6238 { X86::VXORPSZ128rrkz, X86::VXORPDZ128rrkz,
6239 X86::VPXORQZ128rrkz, X86::VPXORDZ128rrkz },
6240 { X86::VANDNPSZ256rmk, X86::VANDNPDZ256rmk,
6241 X86::VPANDNQZ256rmk, X86::VPANDNDZ256rmk },
6242 { X86::VANDNPSZ256rmkz, X86::VANDNPDZ256rmkz,
6243 X86::VPANDNQZ256rmkz, X86::VPANDNDZ256rmkz },
6244 { X86::VANDNPSZ256rrk, X86::VANDNPDZ256rrk,
6245 X86::VPANDNQZ256rrk, X86::VPANDNDZ256rrk },
6246 { X86::VANDNPSZ256rrkz, X86::VANDNPDZ256rrkz,
6247 X86::VPANDNQZ256rrkz, X86::VPANDNDZ256rrkz },
6248 { X86::VANDPSZ256rmk, X86::VANDPDZ256rmk,
6249 X86::VPANDQZ256rmk, X86::VPANDDZ256rmk },
6250 { X86::VANDPSZ256rmkz, X86::VANDPDZ256rmkz,
6251 X86::VPANDQZ256rmkz, X86::VPANDDZ256rmkz },
6252 { X86::VANDPSZ256rrk, X86::VANDPDZ256rrk,
6253 X86::VPANDQZ256rrk, X86::VPANDDZ256rrk },
6254 { X86::VANDPSZ256rrkz, X86::VANDPDZ256rrkz,
6255 X86::VPANDQZ256rrkz, X86::VPANDDZ256rrkz },
6256 { X86::VORPSZ256rmk, X86::VORPDZ256rmk,
6257 X86::VPORQZ256rmk, X86::VPORDZ256rmk },
6258 { X86::VORPSZ256rmkz, X86::VORPDZ256rmkz,
6259 X86::VPORQZ256rmkz, X86::VPORDZ256rmkz },
6260 { X86::VORPSZ256rrk, X86::VORPDZ256rrk,
6261 X86::VPORQZ256rrk, X86::VPORDZ256rrk },
6262 { X86::VORPSZ256rrkz, X86::VORPDZ256rrkz,
6263 X86::VPORQZ256rrkz, X86::VPORDZ256rrkz },
6264 { X86::VXORPSZ256rmk, X86::VXORPDZ256rmk,
6265 X86::VPXORQZ256rmk, X86::VPXORDZ256rmk },
6266 { X86::VXORPSZ256rmkz, X86::VXORPDZ256rmkz,
6267 X86::VPXORQZ256rmkz, X86::VPXORDZ256rmkz },
6268 { X86::VXORPSZ256rrk, X86::VXORPDZ256rrk,
6269 X86::VPXORQZ256rrk, X86::VPXORDZ256rrk },
6270 { X86::VXORPSZ256rrkz, X86::VXORPDZ256rrkz,
6271 X86::VPXORQZ256rrkz, X86::VPXORDZ256rrkz },
6272 { X86::VANDNPSZrmk, X86::VANDNPDZrmk,
6273 X86::VPANDNQZrmk, X86::VPANDNDZrmk },
6274 { X86::VANDNPSZrmkz, X86::VANDNPDZrmkz,
6275 X86::VPANDNQZrmkz, X86::VPANDNDZrmkz },
6276 { X86::VANDNPSZrrk, X86::VANDNPDZrrk,
6277 X86::VPANDNQZrrk, X86::VPANDNDZrrk },
6278 { X86::VANDNPSZrrkz, X86::VANDNPDZrrkz,
6279 X86::VPANDNQZrrkz, X86::VPANDNDZrrkz },
6280 { X86::VANDPSZrmk, X86::VANDPDZrmk,
6281 X86::VPANDQZrmk, X86::VPANDDZrmk },
6282 { X86::VANDPSZrmkz, X86::VANDPDZrmkz,
6283 X86::VPANDQZrmkz, X86::VPANDDZrmkz },
6284 { X86::VANDPSZrrk, X86::VANDPDZrrk,
6285 X86::VPANDQZrrk, X86::VPANDDZrrk },
6286 { X86::VANDPSZrrkz, X86::VANDPDZrrkz,
6287 X86::VPANDQZrrkz, X86::VPANDDZrrkz },
6288 { X86::VORPSZrmk, X86::VORPDZrmk,
6289 X86::VPORQZrmk, X86::VPORDZrmk },
6290 { X86::VORPSZrmkz, X86::VORPDZrmkz,
6291 X86::VPORQZrmkz, X86::VPORDZrmkz },
6292 { X86::VORPSZrrk, X86::VORPDZrrk,
6293 X86::VPORQZrrk, X86::VPORDZrrk },
6294 { X86::VORPSZrrkz, X86::VORPDZrrkz,
6295 X86::VPORQZrrkz, X86::VPORDZrrkz },
6296 { X86::VXORPSZrmk, X86::VXORPDZrmk,
6297 X86::VPXORQZrmk, X86::VPXORDZrmk },
6298 { X86::VXORPSZrmkz, X86::VXORPDZrmkz,
6299 X86::VPXORQZrmkz, X86::VPXORDZrmkz },
6300 { X86::VXORPSZrrk, X86::VXORPDZrrk,
6301 X86::VPXORQZrrk, X86::VPXORDZrrk },
6302 { X86::VXORPSZrrkz, X86::VXORPDZrrkz,
6303 X86::VPXORQZrrkz, X86::VPXORDZrrkz },
6304 // Broadcast loads can be handled the same as masked operations to avoid
6305 // changing element size.
6306 { X86::VANDNPSZ128rmb, X86::VANDNPDZ128rmb,
6307 X86::VPANDNQZ128rmb, X86::VPANDNDZ128rmb },
6308 { X86::VANDPSZ128rmb, X86::VANDPDZ128rmb,
6309 X86::VPANDQZ128rmb, X86::VPANDDZ128rmb },
6310 { X86::VORPSZ128rmb, X86::VORPDZ128rmb,
6311 X86::VPORQZ128rmb, X86::VPORDZ128rmb },
6312 { X86::VXORPSZ128rmb, X86::VXORPDZ128rmb,
6313 X86::VPXORQZ128rmb, X86::VPXORDZ128rmb },
6314 { X86::VANDNPSZ256rmb, X86::VANDNPDZ256rmb,
6315 X86::VPANDNQZ256rmb, X86::VPANDNDZ256rmb },
6316 { X86::VANDPSZ256rmb, X86::VANDPDZ256rmb,
6317 X86::VPANDQZ256rmb, X86::VPANDDZ256rmb },
6318 { X86::VORPSZ256rmb, X86::VORPDZ256rmb,
6319 X86::VPORQZ256rmb, X86::VPORDZ256rmb },
6320 { X86::VXORPSZ256rmb, X86::VXORPDZ256rmb,
6321 X86::VPXORQZ256rmb, X86::VPXORDZ256rmb },
6322 { X86::VANDNPSZrmb, X86::VANDNPDZrmb,
6323 X86::VPANDNQZrmb, X86::VPANDNDZrmb },
6324 { X86::VANDPSZrmb, X86::VANDPDZrmb,
6325 X86::VPANDQZrmb, X86::VPANDDZrmb },
6326 { X86::VANDPSZrmb, X86::VANDPDZrmb,
6327 X86::VPANDQZrmb, X86::VPANDDZrmb },
6328 { X86::VORPSZrmb, X86::VORPDZrmb,
6329 X86::VPORQZrmb, X86::VPORDZrmb },
6330 { X86::VXORPSZrmb, X86::VXORPDZrmb,
6331 X86::VPXORQZrmb, X86::VPXORDZrmb },
6332 { X86::VANDNPSZ128rmbk, X86::VANDNPDZ128rmbk,
6333 X86::VPANDNQZ128rmbk, X86::VPANDNDZ128rmbk },
6334 { X86::VANDPSZ128rmbk, X86::VANDPDZ128rmbk,
6335 X86::VPANDQZ128rmbk, X86::VPANDDZ128rmbk },
6336 { X86::VORPSZ128rmbk, X86::VORPDZ128rmbk,
6337 X86::VPORQZ128rmbk, X86::VPORDZ128rmbk },
6338 { X86::VXORPSZ128rmbk, X86::VXORPDZ128rmbk,
6339 X86::VPXORQZ128rmbk, X86::VPXORDZ128rmbk },
6340 { X86::VANDNPSZ256rmbk, X86::VANDNPDZ256rmbk,
6341 X86::VPANDNQZ256rmbk, X86::VPANDNDZ256rmbk },
6342 { X86::VANDPSZ256rmbk, X86::VANDPDZ256rmbk,
6343 X86::VPANDQZ256rmbk, X86::VPANDDZ256rmbk },
6344 { X86::VORPSZ256rmbk, X86::VORPDZ256rmbk,
6345 X86::VPORQZ256rmbk, X86::VPORDZ256rmbk },
6346 { X86::VXORPSZ256rmbk, X86::VXORPDZ256rmbk,
6347 X86::VPXORQZ256rmbk, X86::VPXORDZ256rmbk },
6348 { X86::VANDNPSZrmbk, X86::VANDNPDZrmbk,
6349 X86::VPANDNQZrmbk, X86::VPANDNDZrmbk },
6350 { X86::VANDPSZrmbk, X86::VANDPDZrmbk,
6351 X86::VPANDQZrmbk, X86::VPANDDZrmbk },
6352 { X86::VANDPSZrmbk, X86::VANDPDZrmbk,
6353 X86::VPANDQZrmbk, X86::VPANDDZrmbk },
6354 { X86::VORPSZrmbk, X86::VORPDZrmbk,
6355 X86::VPORQZrmbk, X86::VPORDZrmbk },
6356 { X86::VXORPSZrmbk, X86::VXORPDZrmbk,
6357 X86::VPXORQZrmbk, X86::VPXORDZrmbk },
6358 { X86::VANDNPSZ128rmbkz,X86::VANDNPDZ128rmbkz,
6359 X86::VPANDNQZ128rmbkz,X86::VPANDNDZ128rmbkz},
6360 { X86::VANDPSZ128rmbkz, X86::VANDPDZ128rmbkz,
6361 X86::VPANDQZ128rmbkz, X86::VPANDDZ128rmbkz },
6362 { X86::VORPSZ128rmbkz, X86::VORPDZ128rmbkz,
6363 X86::VPORQZ128rmbkz, X86::VPORDZ128rmbkz },
6364 { X86::VXORPSZ128rmbkz, X86::VXORPDZ128rmbkz,
6365 X86::VPXORQZ128rmbkz, X86::VPXORDZ128rmbkz },
6366 { X86::VANDNPSZ256rmbkz,X86::VANDNPDZ256rmbkz,
6367 X86::VPANDNQZ256rmbkz,X86::VPANDNDZ256rmbkz},
6368 { X86::VANDPSZ256rmbkz, X86::VANDPDZ256rmbkz,
6369 X86::VPANDQZ256rmbkz, X86::VPANDDZ256rmbkz },
6370 { X86::VORPSZ256rmbkz, X86::VORPDZ256rmbkz,
6371 X86::VPORQZ256rmbkz, X86::VPORDZ256rmbkz },
6372 { X86::VXORPSZ256rmbkz, X86::VXORPDZ256rmbkz,
6373 X86::VPXORQZ256rmbkz, X86::VPXORDZ256rmbkz },
6374 { X86::VANDNPSZrmbkz, X86::VANDNPDZrmbkz,
6375 X86::VPANDNQZrmbkz, X86::VPANDNDZrmbkz },
6376 { X86::VANDPSZrmbkz, X86::VANDPDZrmbkz,
6377 X86::VPANDQZrmbkz, X86::VPANDDZrmbkz },
6378 { X86::VANDPSZrmbkz, X86::VANDPDZrmbkz,
6379 X86::VPANDQZrmbkz, X86::VPANDDZrmbkz },
6380 { X86::VORPSZrmbkz, X86::VORPDZrmbkz,
6381 X86::VPORQZrmbkz, X86::VPORDZrmbkz },
6382 { X86::VXORPSZrmbkz, X86::VXORPDZrmbkz,
6383 X86::VPXORQZrmbkz, X86::VPXORDZrmbkz },
6386 // NOTE: These should only be used by the custom domain methods.
6387 static const uint16_t ReplaceableCustomInstrs[][3] = {
6388 //PackedSingle PackedDouble PackedInt
6389 { X86::BLENDPSrmi, X86::BLENDPDrmi, X86::PBLENDWrmi },
6390 { X86::BLENDPSrri, X86::BLENDPDrri, X86::PBLENDWrri },
6391 { X86::VBLENDPSrmi, X86::VBLENDPDrmi, X86::VPBLENDWrmi },
6392 { X86::VBLENDPSrri, X86::VBLENDPDrri, X86::VPBLENDWrri },
6393 { X86::VBLENDPSYrmi, X86::VBLENDPDYrmi, X86::VPBLENDWYrmi },
6394 { X86::VBLENDPSYrri, X86::VBLENDPDYrri, X86::VPBLENDWYrri },
6396 static const uint16_t ReplaceableCustomAVX2Instrs[][3] = {
6397 //PackedSingle PackedDouble PackedInt
6398 { X86::VBLENDPSrmi, X86::VBLENDPDrmi, X86::VPBLENDDrmi },
6399 { X86::VBLENDPSrri, X86::VBLENDPDrri, X86::VPBLENDDrri },
6400 { X86::VBLENDPSYrmi, X86::VBLENDPDYrmi, X86::VPBLENDDYrmi },
6401 { X86::VBLENDPSYrri, X86::VBLENDPDYrri, X86::VPBLENDDYrri },
6404 // Special table for changing EVEX logic instructions to VEX.
6405 // TODO: Should we run EVEX->VEX earlier?
6406 static const uint16_t ReplaceableCustomAVX512LogicInstrs[][4] = {
6407 // Two integer columns for 64-bit and 32-bit elements.
6408 //PackedSingle PackedDouble PackedInt PackedInt
6409 { X86::VANDNPSrm, X86::VANDNPDrm, X86::VPANDNQZ128rm, X86::VPANDNDZ128rm },
6410 { X86::VANDNPSrr, X86::VANDNPDrr, X86::VPANDNQZ128rr, X86::VPANDNDZ128rr },
6411 { X86::VANDPSrm, X86::VANDPDrm, X86::VPANDQZ128rm, X86::VPANDDZ128rm },
6412 { X86::VANDPSrr, X86::VANDPDrr, X86::VPANDQZ128rr, X86::VPANDDZ128rr },
6413 { X86::VORPSrm, X86::VORPDrm, X86::VPORQZ128rm, X86::VPORDZ128rm },
6414 { X86::VORPSrr, X86::VORPDrr, X86::VPORQZ128rr, X86::VPORDZ128rr },
6415 { X86::VXORPSrm, X86::VXORPDrm, X86::VPXORQZ128rm, X86::VPXORDZ128rm },
6416 { X86::VXORPSrr, X86::VXORPDrr, X86::VPXORQZ128rr, X86::VPXORDZ128rr },
6417 { X86::VANDNPSYrm, X86::VANDNPDYrm, X86::VPANDNQZ256rm, X86::VPANDNDZ256rm },
6418 { X86::VANDNPSYrr, X86::VANDNPDYrr, X86::VPANDNQZ256rr, X86::VPANDNDZ256rr },
6419 { X86::VANDPSYrm, X86::VANDPDYrm, X86::VPANDQZ256rm, X86::VPANDDZ256rm },
6420 { X86::VANDPSYrr, X86::VANDPDYrr, X86::VPANDQZ256rr, X86::VPANDDZ256rr },
6421 { X86::VORPSYrm, X86::VORPDYrm, X86::VPORQZ256rm, X86::VPORDZ256rm },
6422 { X86::VORPSYrr, X86::VORPDYrr, X86::VPORQZ256rr, X86::VPORDZ256rr },
6423 { X86::VXORPSYrm, X86::VXORPDYrm, X86::VPXORQZ256rm, X86::VPXORDZ256rm },
6424 { X86::VXORPSYrr, X86::VXORPDYrr, X86::VPXORQZ256rr, X86::VPXORDZ256rr },
6427 // FIXME: Some shuffle and unpack instructions have equivalents in different
6428 // domains, but they require a bit more work than just switching opcodes.
6430 static const uint16_t *lookup(unsigned opcode, unsigned domain,
6431 ArrayRef<uint16_t[3]> Table) {
6432 for (const uint16_t (&Row)[3] : Table)
6433 if (Row[domain-1] == opcode)
6438 static const uint16_t *lookupAVX512(unsigned opcode, unsigned domain,
6439 ArrayRef<uint16_t[4]> Table) {
6440 // If this is the integer domain make sure to check both integer columns.
6441 for (const uint16_t (&Row)[4] : Table)
6442 if (Row[domain-1] == opcode || (domain == 3 && Row[3] == opcode))
6447 // Helper to attempt to widen/narrow blend masks.
6448 static bool AdjustBlendMask(unsigned OldMask, unsigned OldWidth,
6449 unsigned NewWidth, unsigned *pNewMask = nullptr) {
6450 assert(((OldWidth % NewWidth) == 0 || (NewWidth % OldWidth) == 0) &&
6451 "Illegal blend mask scale");
6452 unsigned NewMask = 0;
6454 if ((OldWidth % NewWidth) == 0) {
6455 unsigned Scale = OldWidth / NewWidth;
6456 unsigned SubMask = (1u << Scale) - 1;
6457 for (unsigned i = 0; i != NewWidth; ++i) {
6458 unsigned Sub = (OldMask >> (i * Scale)) & SubMask;
6460 NewMask |= (1u << i);
6461 else if (Sub != 0x0)
6465 unsigned Scale = NewWidth / OldWidth;
6466 unsigned SubMask = (1u << Scale) - 1;
6467 for (unsigned i = 0; i != OldWidth; ++i) {
6468 if (OldMask & (1 << i)) {
6469 NewMask |= (SubMask << (i * Scale));
6475 *pNewMask = NewMask;
6479 uint16_t X86InstrInfo::getExecutionDomainCustom(const MachineInstr &MI) const {
6480 unsigned Opcode = MI.getOpcode();
6481 unsigned NumOperands = MI.getDesc().getNumOperands();
6483 auto GetBlendDomains = [&](unsigned ImmWidth, bool Is256) {
6484 uint16_t validDomains = 0;
6485 if (MI.getOperand(NumOperands - 1).isImm()) {
6486 unsigned Imm = MI.getOperand(NumOperands - 1).getImm();
6487 if (AdjustBlendMask(Imm, ImmWidth, Is256 ? 8 : 4))
6488 validDomains |= 0x2; // PackedSingle
6489 if (AdjustBlendMask(Imm, ImmWidth, Is256 ? 4 : 2))
6490 validDomains |= 0x4; // PackedDouble
6491 if (!Is256 || Subtarget.hasAVX2())
6492 validDomains |= 0x8; // PackedInt
6494 return validDomains;
6498 case X86::BLENDPDrmi:
6499 case X86::BLENDPDrri:
6500 case X86::VBLENDPDrmi:
6501 case X86::VBLENDPDrri:
6502 return GetBlendDomains(2, false);
6503 case X86::VBLENDPDYrmi:
6504 case X86::VBLENDPDYrri:
6505 return GetBlendDomains(4, true);
6506 case X86::BLENDPSrmi:
6507 case X86::BLENDPSrri:
6508 case X86::VBLENDPSrmi:
6509 case X86::VBLENDPSrri:
6510 case X86::VPBLENDDrmi:
6511 case X86::VPBLENDDrri:
6512 return GetBlendDomains(4, false);
6513 case X86::VBLENDPSYrmi:
6514 case X86::VBLENDPSYrri:
6515 case X86::VPBLENDDYrmi:
6516 case X86::VPBLENDDYrri:
6517 return GetBlendDomains(8, true);
6518 case X86::PBLENDWrmi:
6519 case X86::PBLENDWrri:
6520 case X86::VPBLENDWrmi:
6521 case X86::VPBLENDWrri:
6522 // Treat VPBLENDWY as a 128-bit vector as it repeats the lo/hi masks.
6523 case X86::VPBLENDWYrmi:
6524 case X86::VPBLENDWYrri:
6525 return GetBlendDomains(8, false);
6526 case X86::VPANDDZ128rr: case X86::VPANDDZ128rm:
6527 case X86::VPANDDZ256rr: case X86::VPANDDZ256rm:
6528 case X86::VPANDQZ128rr: case X86::VPANDQZ128rm:
6529 case X86::VPANDQZ256rr: case X86::VPANDQZ256rm:
6530 case X86::VPANDNDZ128rr: case X86::VPANDNDZ128rm:
6531 case X86::VPANDNDZ256rr: case X86::VPANDNDZ256rm:
6532 case X86::VPANDNQZ128rr: case X86::VPANDNQZ128rm:
6533 case X86::VPANDNQZ256rr: case X86::VPANDNQZ256rm:
6534 case X86::VPORDZ128rr: case X86::VPORDZ128rm:
6535 case X86::VPORDZ256rr: case X86::VPORDZ256rm:
6536 case X86::VPORQZ128rr: case X86::VPORQZ128rm:
6537 case X86::VPORQZ256rr: case X86::VPORQZ256rm:
6538 case X86::VPXORDZ128rr: case X86::VPXORDZ128rm:
6539 case X86::VPXORDZ256rr: case X86::VPXORDZ256rm:
6540 case X86::VPXORQZ128rr: case X86::VPXORQZ128rm:
6541 case X86::VPXORQZ256rr: case X86::VPXORQZ256rm:
6542 // If we don't have DQI see if we can still switch from an EVEX integer
6543 // instruction to a VEX floating point instruction.
6544 if (Subtarget.hasDQI())
6547 if (RI.getEncodingValue(MI.getOperand(0).getReg()) >= 16)
6549 if (RI.getEncodingValue(MI.getOperand(1).getReg()) >= 16)
6551 // Register forms will have 3 operands. Memory form will have more.
6552 if (NumOperands == 3 &&
6553 RI.getEncodingValue(MI.getOperand(2).getReg()) >= 16)
6556 // All domains are valid.
6558 case X86::MOVHLPSrr:
6559 // We can swap domains when both inputs are the same register.
6560 // FIXME: This doesn't catch all the cases we would like. If the input
6561 // register isn't KILLed by the instruction, the two address instruction
6562 // pass puts a COPY on one input. The other input uses the original
6563 // register. This prevents the same physical register from being used by
6565 if (MI.getOperand(1).getReg() == MI.getOperand(2).getReg() &&
6566 MI.getOperand(0).getSubReg() == 0 &&
6567 MI.getOperand(1).getSubReg() == 0 &&
6568 MI.getOperand(2).getSubReg() == 0)
6575 bool X86InstrInfo::setExecutionDomainCustom(MachineInstr &MI,
6576 unsigned Domain) const {
6577 assert(Domain > 0 && Domain < 4 && "Invalid execution domain");
6578 uint16_t dom = (MI.getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
6579 assert(dom && "Not an SSE instruction");
6581 unsigned Opcode = MI.getOpcode();
6582 unsigned NumOperands = MI.getDesc().getNumOperands();
6584 auto SetBlendDomain = [&](unsigned ImmWidth, bool Is256) {
6585 if (MI.getOperand(NumOperands - 1).isImm()) {
6586 unsigned Imm = MI.getOperand(NumOperands - 1).getImm() & 255;
6587 Imm = (ImmWidth == 16 ? ((Imm << 8) | Imm) : Imm);
6588 unsigned NewImm = Imm;
6590 const uint16_t *table = lookup(Opcode, dom, ReplaceableCustomInstrs);
6592 table = lookup(Opcode, dom, ReplaceableCustomAVX2Instrs);
6594 if (Domain == 1) { // PackedSingle
6595 AdjustBlendMask(Imm, ImmWidth, Is256 ? 8 : 4, &NewImm);
6596 } else if (Domain == 2) { // PackedDouble
6597 AdjustBlendMask(Imm, ImmWidth, Is256 ? 4 : 2, &NewImm);
6598 } else if (Domain == 3) { // PackedInt
6599 if (Subtarget.hasAVX2()) {
6600 // If we are already VPBLENDW use that, else use VPBLENDD.
6601 if ((ImmWidth / (Is256 ? 2 : 1)) != 8) {
6602 table = lookup(Opcode, dom, ReplaceableCustomAVX2Instrs);
6603 AdjustBlendMask(Imm, ImmWidth, Is256 ? 8 : 4, &NewImm);
6606 assert(!Is256 && "128-bit vector expected");
6607 AdjustBlendMask(Imm, ImmWidth, 8, &NewImm);
6611 assert(table && table[Domain - 1] && "Unknown domain op");
6612 MI.setDesc(get(table[Domain - 1]));
6613 MI.getOperand(NumOperands - 1).setImm(NewImm & 255);
6619 case X86::BLENDPDrmi:
6620 case X86::BLENDPDrri:
6621 case X86::VBLENDPDrmi:
6622 case X86::VBLENDPDrri:
6623 return SetBlendDomain(2, false);
6624 case X86::VBLENDPDYrmi:
6625 case X86::VBLENDPDYrri:
6626 return SetBlendDomain(4, true);
6627 case X86::BLENDPSrmi:
6628 case X86::BLENDPSrri:
6629 case X86::VBLENDPSrmi:
6630 case X86::VBLENDPSrri:
6631 case X86::VPBLENDDrmi:
6632 case X86::VPBLENDDrri:
6633 return SetBlendDomain(4, false);
6634 case X86::VBLENDPSYrmi:
6635 case X86::VBLENDPSYrri:
6636 case X86::VPBLENDDYrmi:
6637 case X86::VPBLENDDYrri:
6638 return SetBlendDomain(8, true);
6639 case X86::PBLENDWrmi:
6640 case X86::PBLENDWrri:
6641 case X86::VPBLENDWrmi:
6642 case X86::VPBLENDWrri:
6643 return SetBlendDomain(8, false);
6644 case X86::VPBLENDWYrmi:
6645 case X86::VPBLENDWYrri:
6646 return SetBlendDomain(16, true);
6647 case X86::VPANDDZ128rr: case X86::VPANDDZ128rm:
6648 case X86::VPANDDZ256rr: case X86::VPANDDZ256rm:
6649 case X86::VPANDQZ128rr: case X86::VPANDQZ128rm:
6650 case X86::VPANDQZ256rr: case X86::VPANDQZ256rm:
6651 case X86::VPANDNDZ128rr: case X86::VPANDNDZ128rm:
6652 case X86::VPANDNDZ256rr: case X86::VPANDNDZ256rm:
6653 case X86::VPANDNQZ128rr: case X86::VPANDNQZ128rm:
6654 case X86::VPANDNQZ256rr: case X86::VPANDNQZ256rm:
6655 case X86::VPORDZ128rr: case X86::VPORDZ128rm:
6656 case X86::VPORDZ256rr: case X86::VPORDZ256rm:
6657 case X86::VPORQZ128rr: case X86::VPORQZ128rm:
6658 case X86::VPORQZ256rr: case X86::VPORQZ256rm:
6659 case X86::VPXORDZ128rr: case X86::VPXORDZ128rm:
6660 case X86::VPXORDZ256rr: case X86::VPXORDZ256rm:
6661 case X86::VPXORQZ128rr: case X86::VPXORQZ128rm:
6662 case X86::VPXORQZ256rr: case X86::VPXORQZ256rm: {
6663 // Without DQI, convert EVEX instructions to VEX instructions.
6664 if (Subtarget.hasDQI())
6667 const uint16_t *table = lookupAVX512(MI.getOpcode(), dom,
6668 ReplaceableCustomAVX512LogicInstrs);
6669 assert(table && "Instruction not found in table?");
6670 // Don't change integer Q instructions to D instructions and
6671 // use D intructions if we started with a PS instruction.
6672 if (Domain == 3 && (dom == 1 || table[3] == MI.getOpcode()))
6674 MI.setDesc(get(table[Domain - 1]));
6677 case X86::UNPCKHPDrr:
6678 case X86::MOVHLPSrr:
6679 // We just need to commute the instruction which will switch the domains.
6680 if (Domain != dom && Domain != 3 &&
6681 MI.getOperand(1).getReg() == MI.getOperand(2).getReg() &&
6682 MI.getOperand(0).getSubReg() == 0 &&
6683 MI.getOperand(1).getSubReg() == 0 &&
6684 MI.getOperand(2).getSubReg() == 0) {
6685 commuteInstruction(MI, false);
6688 // We must always return true for MOVHLPSrr.
6689 if (Opcode == X86::MOVHLPSrr)
6695 std::pair<uint16_t, uint16_t>
6696 X86InstrInfo::getExecutionDomain(const MachineInstr &MI) const {
6697 uint16_t domain = (MI.getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
6698 unsigned opcode = MI.getOpcode();
6699 uint16_t validDomains = 0;
6701 // Attempt to match for custom instructions.
6702 validDomains = getExecutionDomainCustom(MI);
6704 return std::make_pair(domain, validDomains);
6706 if (lookup(opcode, domain, ReplaceableInstrs)) {
6708 } else if (lookup(opcode, domain, ReplaceableInstrsAVX2)) {
6709 validDomains = Subtarget.hasAVX2() ? 0xe : 0x6;
6710 } else if (lookup(opcode, domain, ReplaceableInstrsAVX2InsertExtract)) {
6711 // Insert/extract instructions should only effect domain if AVX2
6713 if (!Subtarget.hasAVX2())
6714 return std::make_pair(0, 0);
6716 } else if (lookupAVX512(opcode, domain, ReplaceableInstrsAVX512)) {
6718 } else if (Subtarget.hasDQI() && lookupAVX512(opcode, domain,
6719 ReplaceableInstrsAVX512DQ)) {
6721 } else if (Subtarget.hasDQI()) {
6722 if (const uint16_t *table = lookupAVX512(opcode, domain,
6723 ReplaceableInstrsAVX512DQMasked)) {
6724 if (domain == 1 || (domain == 3 && table[3] == opcode))
6731 return std::make_pair(domain, validDomains);
6734 void X86InstrInfo::setExecutionDomain(MachineInstr &MI, unsigned Domain) const {
6735 assert(Domain>0 && Domain<4 && "Invalid execution domain");
6736 uint16_t dom = (MI.getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
6737 assert(dom && "Not an SSE instruction");
6739 // Attempt to match for custom instructions.
6740 if (setExecutionDomainCustom(MI, Domain))
6743 const uint16_t *table = lookup(MI.getOpcode(), dom, ReplaceableInstrs);
6744 if (!table) { // try the other table
6745 assert((Subtarget.hasAVX2() || Domain < 3) &&
6746 "256-bit vector operations only available in AVX2");
6747 table = lookup(MI.getOpcode(), dom, ReplaceableInstrsAVX2);
6749 if (!table) { // try the other table
6750 assert(Subtarget.hasAVX2() &&
6751 "256-bit insert/extract only available in AVX2");
6752 table = lookup(MI.getOpcode(), dom, ReplaceableInstrsAVX2InsertExtract);
6754 if (!table) { // try the AVX512 table
6755 assert(Subtarget.hasAVX512() && "Requires AVX-512");
6756 table = lookupAVX512(MI.getOpcode(), dom, ReplaceableInstrsAVX512);
6757 // Don't change integer Q instructions to D instructions.
6758 if (table && Domain == 3 && table[3] == MI.getOpcode())
6761 if (!table) { // try the AVX512DQ table
6762 assert((Subtarget.hasDQI() || Domain >= 3) && "Requires AVX-512DQ");
6763 table = lookupAVX512(MI.getOpcode(), dom, ReplaceableInstrsAVX512DQ);
6764 // Don't change integer Q instructions to D instructions and
6765 // use D intructions if we started with a PS instruction.
6766 if (table && Domain == 3 && (dom == 1 || table[3] == MI.getOpcode()))
6769 if (!table) { // try the AVX512DQMasked table
6770 assert((Subtarget.hasDQI() || Domain >= 3) && "Requires AVX-512DQ");
6771 table = lookupAVX512(MI.getOpcode(), dom, ReplaceableInstrsAVX512DQMasked);
6772 if (table && Domain == 3 && (dom == 1 || table[3] == MI.getOpcode()))
6775 assert(table && "Cannot change domain");
6776 MI.setDesc(get(table[Domain - 1]));
6779 /// Return the noop instruction to use for a noop.
6780 void X86InstrInfo::getNoop(MCInst &NopInst) const {
6781 NopInst.setOpcode(X86::NOOP);
6784 bool X86InstrInfo::isHighLatencyDef(int opc) const {
6786 default: return false;
6792 case X86::DIVSDrm_Int:
6794 case X86::DIVSDrr_Int:
6796 case X86::DIVSSrm_Int:
6798 case X86::DIVSSrr_Int:
6804 case X86::SQRTSDm_Int:
6806 case X86::SQRTSDr_Int:
6808 case X86::SQRTSSm_Int:
6810 case X86::SQRTSSr_Int:
6811 // AVX instructions with high latency
6814 case X86::VDIVPDYrm:
6815 case X86::VDIVPDYrr:
6818 case X86::VDIVPSYrm:
6819 case X86::VDIVPSYrr:
6821 case X86::VDIVSDrm_Int:
6823 case X86::VDIVSDrr_Int:
6825 case X86::VDIVSSrm_Int:
6827 case X86::VDIVSSrr_Int:
6830 case X86::VSQRTPDYm:
6831 case X86::VSQRTPDYr:
6834 case X86::VSQRTPSYm:
6835 case X86::VSQRTPSYr:
6837 case X86::VSQRTSDm_Int:
6839 case X86::VSQRTSDr_Int:
6841 case X86::VSQRTSSm_Int:
6843 case X86::VSQRTSSr_Int:
6844 // AVX512 instructions with high latency
6845 case X86::VDIVPDZ128rm:
6846 case X86::VDIVPDZ128rmb:
6847 case X86::VDIVPDZ128rmbk:
6848 case X86::VDIVPDZ128rmbkz:
6849 case X86::VDIVPDZ128rmk:
6850 case X86::VDIVPDZ128rmkz:
6851 case X86::VDIVPDZ128rr:
6852 case X86::VDIVPDZ128rrk:
6853 case X86::VDIVPDZ128rrkz:
6854 case X86::VDIVPDZ256rm:
6855 case X86::VDIVPDZ256rmb:
6856 case X86::VDIVPDZ256rmbk:
6857 case X86::VDIVPDZ256rmbkz:
6858 case X86::VDIVPDZ256rmk:
6859 case X86::VDIVPDZ256rmkz:
6860 case X86::VDIVPDZ256rr:
6861 case X86::VDIVPDZ256rrk:
6862 case X86::VDIVPDZ256rrkz:
6863 case X86::VDIVPDZrrb:
6864 case X86::VDIVPDZrrbk:
6865 case X86::VDIVPDZrrbkz:
6866 case X86::VDIVPDZrm:
6867 case X86::VDIVPDZrmb:
6868 case X86::VDIVPDZrmbk:
6869 case X86::VDIVPDZrmbkz:
6870 case X86::VDIVPDZrmk:
6871 case X86::VDIVPDZrmkz:
6872 case X86::VDIVPDZrr:
6873 case X86::VDIVPDZrrk:
6874 case X86::VDIVPDZrrkz:
6875 case X86::VDIVPSZ128rm:
6876 case X86::VDIVPSZ128rmb:
6877 case X86::VDIVPSZ128rmbk:
6878 case X86::VDIVPSZ128rmbkz:
6879 case X86::VDIVPSZ128rmk:
6880 case X86::VDIVPSZ128rmkz:
6881 case X86::VDIVPSZ128rr:
6882 case X86::VDIVPSZ128rrk:
6883 case X86::VDIVPSZ128rrkz:
6884 case X86::VDIVPSZ256rm:
6885 case X86::VDIVPSZ256rmb:
6886 case X86::VDIVPSZ256rmbk:
6887 case X86::VDIVPSZ256rmbkz:
6888 case X86::VDIVPSZ256rmk:
6889 case X86::VDIVPSZ256rmkz:
6890 case X86::VDIVPSZ256rr:
6891 case X86::VDIVPSZ256rrk:
6892 case X86::VDIVPSZ256rrkz:
6893 case X86::VDIVPSZrrb:
6894 case X86::VDIVPSZrrbk:
6895 case X86::VDIVPSZrrbkz:
6896 case X86::VDIVPSZrm:
6897 case X86::VDIVPSZrmb:
6898 case X86::VDIVPSZrmbk:
6899 case X86::VDIVPSZrmbkz:
6900 case X86::VDIVPSZrmk:
6901 case X86::VDIVPSZrmkz:
6902 case X86::VDIVPSZrr:
6903 case X86::VDIVPSZrrk:
6904 case X86::VDIVPSZrrkz:
6905 case X86::VDIVSDZrm:
6906 case X86::VDIVSDZrr:
6907 case X86::VDIVSDZrm_Int:
6908 case X86::VDIVSDZrm_Intk:
6909 case X86::VDIVSDZrm_Intkz:
6910 case X86::VDIVSDZrr_Int:
6911 case X86::VDIVSDZrr_Intk:
6912 case X86::VDIVSDZrr_Intkz:
6913 case X86::VDIVSDZrrb_Int:
6914 case X86::VDIVSDZrrb_Intk:
6915 case X86::VDIVSDZrrb_Intkz:
6916 case X86::VDIVSSZrm:
6917 case X86::VDIVSSZrr:
6918 case X86::VDIVSSZrm_Int:
6919 case X86::VDIVSSZrm_Intk:
6920 case X86::VDIVSSZrm_Intkz:
6921 case X86::VDIVSSZrr_Int:
6922 case X86::VDIVSSZrr_Intk:
6923 case X86::VDIVSSZrr_Intkz:
6924 case X86::VDIVSSZrrb_Int:
6925 case X86::VDIVSSZrrb_Intk:
6926 case X86::VDIVSSZrrb_Intkz:
6927 case X86::VSQRTPDZ128m:
6928 case X86::VSQRTPDZ128mb:
6929 case X86::VSQRTPDZ128mbk:
6930 case X86::VSQRTPDZ128mbkz:
6931 case X86::VSQRTPDZ128mk:
6932 case X86::VSQRTPDZ128mkz:
6933 case X86::VSQRTPDZ128r:
6934 case X86::VSQRTPDZ128rk:
6935 case X86::VSQRTPDZ128rkz:
6936 case X86::VSQRTPDZ256m:
6937 case X86::VSQRTPDZ256mb:
6938 case X86::VSQRTPDZ256mbk:
6939 case X86::VSQRTPDZ256mbkz:
6940 case X86::VSQRTPDZ256mk:
6941 case X86::VSQRTPDZ256mkz:
6942 case X86::VSQRTPDZ256r:
6943 case X86::VSQRTPDZ256rk:
6944 case X86::VSQRTPDZ256rkz:
6945 case X86::VSQRTPDZm:
6946 case X86::VSQRTPDZmb:
6947 case X86::VSQRTPDZmbk:
6948 case X86::VSQRTPDZmbkz:
6949 case X86::VSQRTPDZmk:
6950 case X86::VSQRTPDZmkz:
6951 case X86::VSQRTPDZr:
6952 case X86::VSQRTPDZrb:
6953 case X86::VSQRTPDZrbk:
6954 case X86::VSQRTPDZrbkz:
6955 case X86::VSQRTPDZrk:
6956 case X86::VSQRTPDZrkz:
6957 case X86::VSQRTPSZ128m:
6958 case X86::VSQRTPSZ128mb:
6959 case X86::VSQRTPSZ128mbk:
6960 case X86::VSQRTPSZ128mbkz:
6961 case X86::VSQRTPSZ128mk:
6962 case X86::VSQRTPSZ128mkz:
6963 case X86::VSQRTPSZ128r:
6964 case X86::VSQRTPSZ128rk:
6965 case X86::VSQRTPSZ128rkz:
6966 case X86::VSQRTPSZ256m:
6967 case X86::VSQRTPSZ256mb:
6968 case X86::VSQRTPSZ256mbk:
6969 case X86::VSQRTPSZ256mbkz:
6970 case X86::VSQRTPSZ256mk:
6971 case X86::VSQRTPSZ256mkz:
6972 case X86::VSQRTPSZ256r:
6973 case X86::VSQRTPSZ256rk:
6974 case X86::VSQRTPSZ256rkz:
6975 case X86::VSQRTPSZm:
6976 case X86::VSQRTPSZmb:
6977 case X86::VSQRTPSZmbk:
6978 case X86::VSQRTPSZmbkz:
6979 case X86::VSQRTPSZmk:
6980 case X86::VSQRTPSZmkz:
6981 case X86::VSQRTPSZr:
6982 case X86::VSQRTPSZrb:
6983 case X86::VSQRTPSZrbk:
6984 case X86::VSQRTPSZrbkz:
6985 case X86::VSQRTPSZrk:
6986 case X86::VSQRTPSZrkz:
6987 case X86::VSQRTSDZm:
6988 case X86::VSQRTSDZm_Int:
6989 case X86::VSQRTSDZm_Intk:
6990 case X86::VSQRTSDZm_Intkz:
6991 case X86::VSQRTSDZr:
6992 case X86::VSQRTSDZr_Int:
6993 case X86::VSQRTSDZr_Intk:
6994 case X86::VSQRTSDZr_Intkz:
6995 case X86::VSQRTSDZrb_Int:
6996 case X86::VSQRTSDZrb_Intk:
6997 case X86::VSQRTSDZrb_Intkz:
6998 case X86::VSQRTSSZm:
6999 case X86::VSQRTSSZm_Int:
7000 case X86::VSQRTSSZm_Intk:
7001 case X86::VSQRTSSZm_Intkz:
7002 case X86::VSQRTSSZr:
7003 case X86::VSQRTSSZr_Int:
7004 case X86::VSQRTSSZr_Intk:
7005 case X86::VSQRTSSZr_Intkz:
7006 case X86::VSQRTSSZrb_Int:
7007 case X86::VSQRTSSZrb_Intk:
7008 case X86::VSQRTSSZrb_Intkz:
7010 case X86::VGATHERDPDYrm:
7011 case X86::VGATHERDPDZ128rm:
7012 case X86::VGATHERDPDZ256rm:
7013 case X86::VGATHERDPDZrm:
7014 case X86::VGATHERDPDrm:
7015 case X86::VGATHERDPSYrm:
7016 case X86::VGATHERDPSZ128rm:
7017 case X86::VGATHERDPSZ256rm:
7018 case X86::VGATHERDPSZrm:
7019 case X86::VGATHERDPSrm:
7020 case X86::VGATHERPF0DPDm:
7021 case X86::VGATHERPF0DPSm:
7022 case X86::VGATHERPF0QPDm:
7023 case X86::VGATHERPF0QPSm:
7024 case X86::VGATHERPF1DPDm:
7025 case X86::VGATHERPF1DPSm:
7026 case X86::VGATHERPF1QPDm:
7027 case X86::VGATHERPF1QPSm:
7028 case X86::VGATHERQPDYrm:
7029 case X86::VGATHERQPDZ128rm:
7030 case X86::VGATHERQPDZ256rm:
7031 case X86::VGATHERQPDZrm:
7032 case X86::VGATHERQPDrm:
7033 case X86::VGATHERQPSYrm:
7034 case X86::VGATHERQPSZ128rm:
7035 case X86::VGATHERQPSZ256rm:
7036 case X86::VGATHERQPSZrm:
7037 case X86::VGATHERQPSrm:
7038 case X86::VPGATHERDDYrm:
7039 case X86::VPGATHERDDZ128rm:
7040 case X86::VPGATHERDDZ256rm:
7041 case X86::VPGATHERDDZrm:
7042 case X86::VPGATHERDDrm:
7043 case X86::VPGATHERDQYrm:
7044 case X86::VPGATHERDQZ128rm:
7045 case X86::VPGATHERDQZ256rm:
7046 case X86::VPGATHERDQZrm:
7047 case X86::VPGATHERDQrm:
7048 case X86::VPGATHERQDYrm:
7049 case X86::VPGATHERQDZ128rm:
7050 case X86::VPGATHERQDZ256rm:
7051 case X86::VPGATHERQDZrm:
7052 case X86::VPGATHERQDrm:
7053 case X86::VPGATHERQQYrm:
7054 case X86::VPGATHERQQZ128rm:
7055 case X86::VPGATHERQQZ256rm:
7056 case X86::VPGATHERQQZrm:
7057 case X86::VPGATHERQQrm:
7058 case X86::VSCATTERDPDZ128mr:
7059 case X86::VSCATTERDPDZ256mr:
7060 case X86::VSCATTERDPDZmr:
7061 case X86::VSCATTERDPSZ128mr:
7062 case X86::VSCATTERDPSZ256mr:
7063 case X86::VSCATTERDPSZmr:
7064 case X86::VSCATTERPF0DPDm:
7065 case X86::VSCATTERPF0DPSm:
7066 case X86::VSCATTERPF0QPDm:
7067 case X86::VSCATTERPF0QPSm:
7068 case X86::VSCATTERPF1DPDm:
7069 case X86::VSCATTERPF1DPSm:
7070 case X86::VSCATTERPF1QPDm:
7071 case X86::VSCATTERPF1QPSm:
7072 case X86::VSCATTERQPDZ128mr:
7073 case X86::VSCATTERQPDZ256mr:
7074 case X86::VSCATTERQPDZmr:
7075 case X86::VSCATTERQPSZ128mr:
7076 case X86::VSCATTERQPSZ256mr:
7077 case X86::VSCATTERQPSZmr:
7078 case X86::VPSCATTERDDZ128mr:
7079 case X86::VPSCATTERDDZ256mr:
7080 case X86::VPSCATTERDDZmr:
7081 case X86::VPSCATTERDQZ128mr:
7082 case X86::VPSCATTERDQZ256mr:
7083 case X86::VPSCATTERDQZmr:
7084 case X86::VPSCATTERQDZ128mr:
7085 case X86::VPSCATTERQDZ256mr:
7086 case X86::VPSCATTERQDZmr:
7087 case X86::VPSCATTERQQZ128mr:
7088 case X86::VPSCATTERQQZ256mr:
7089 case X86::VPSCATTERQQZmr:
7094 bool X86InstrInfo::hasHighOperandLatency(const TargetSchedModel &SchedModel,
7095 const MachineRegisterInfo *MRI,
7096 const MachineInstr &DefMI,
7098 const MachineInstr &UseMI,
7099 unsigned UseIdx) const {
7100 return isHighLatencyDef(DefMI.getOpcode());
7103 bool X86InstrInfo::hasReassociableOperands(const MachineInstr &Inst,
7104 const MachineBasicBlock *MBB) const {
7105 assert((Inst.getNumOperands() == 3 || Inst.getNumOperands() == 4) &&
7106 "Reassociation needs binary operators");
7108 // Integer binary math/logic instructions have a third source operand:
7109 // the EFLAGS register. That operand must be both defined here and never
7110 // used; ie, it must be dead. If the EFLAGS operand is live, then we can
7111 // not change anything because rearranging the operands could affect other
7112 // instructions that depend on the exact status flags (zero, sign, etc.)
7113 // that are set by using these particular operands with this operation.
7114 if (Inst.getNumOperands() == 4) {
7115 assert(Inst.getOperand(3).isReg() &&
7116 Inst.getOperand(3).getReg() == X86::EFLAGS &&
7117 "Unexpected operand in reassociable instruction");
7118 if (!Inst.getOperand(3).isDead())
7122 return TargetInstrInfo::hasReassociableOperands(Inst, MBB);
7125 // TODO: There are many more machine instruction opcodes to match:
7126 // 1. Other data types (integer, vectors)
7127 // 2. Other math / logic operations (xor, or)
7128 // 3. Other forms of the same operation (intrinsics and other variants)
7129 bool X86InstrInfo::isAssociativeAndCommutative(const MachineInstr &Inst) const {
7130 switch (Inst.getOpcode()) {
7161 case X86::VPANDDZ128rr:
7162 case X86::VPANDDZ256rr:
7163 case X86::VPANDDZrr:
7164 case X86::VPANDQZ128rr:
7165 case X86::VPANDQZ256rr:
7166 case X86::VPANDQZrr:
7169 case X86::VPORDZ128rr:
7170 case X86::VPORDZ256rr:
7172 case X86::VPORQZ128rr:
7173 case X86::VPORQZ256rr:
7177 case X86::VPXORDZ128rr:
7178 case X86::VPXORDZ256rr:
7179 case X86::VPXORDZrr:
7180 case X86::VPXORQZ128rr:
7181 case X86::VPXORQZ256rr:
7182 case X86::VPXORQZrr:
7185 case X86::VANDPDYrr:
7186 case X86::VANDPSYrr:
7187 case X86::VANDPDZ128rr:
7188 case X86::VANDPSZ128rr:
7189 case X86::VANDPDZ256rr:
7190 case X86::VANDPSZ256rr:
7191 case X86::VANDPDZrr:
7192 case X86::VANDPSZrr:
7197 case X86::VORPDZ128rr:
7198 case X86::VORPSZ128rr:
7199 case X86::VORPDZ256rr:
7200 case X86::VORPSZ256rr:
7205 case X86::VXORPDYrr:
7206 case X86::VXORPSYrr:
7207 case X86::VXORPDZ128rr:
7208 case X86::VXORPSZ128rr:
7209 case X86::VXORPDZ256rr:
7210 case X86::VXORPSZ256rr:
7211 case X86::VXORPDZrr:
7212 case X86::VXORPSZrr:
7233 case X86::VPADDBYrr:
7234 case X86::VPADDWYrr:
7235 case X86::VPADDDYrr:
7236 case X86::VPADDQYrr:
7237 case X86::VPADDBZ128rr:
7238 case X86::VPADDWZ128rr:
7239 case X86::VPADDDZ128rr:
7240 case X86::VPADDQZ128rr:
7241 case X86::VPADDBZ256rr:
7242 case X86::VPADDWZ256rr:
7243 case X86::VPADDDZ256rr:
7244 case X86::VPADDQZ256rr:
7245 case X86::VPADDBZrr:
7246 case X86::VPADDWZrr:
7247 case X86::VPADDDZrr:
7248 case X86::VPADDQZrr:
7249 case X86::VPMULLWrr:
7250 case X86::VPMULLWYrr:
7251 case X86::VPMULLWZ128rr:
7252 case X86::VPMULLWZ256rr:
7253 case X86::VPMULLWZrr:
7254 case X86::VPMULLDrr:
7255 case X86::VPMULLDYrr:
7256 case X86::VPMULLDZ128rr:
7257 case X86::VPMULLDZ256rr:
7258 case X86::VPMULLDZrr:
7259 case X86::VPMULLQZ128rr:
7260 case X86::VPMULLQZ256rr:
7261 case X86::VPMULLQZrr:
7262 // Normal min/max instructions are not commutative because of NaN and signed
7263 // zero semantics, but these are. Thus, there's no need to check for global
7264 // relaxed math; the instructions themselves have the properties we need.
7273 case X86::VMAXCPDrr:
7274 case X86::VMAXCPSrr:
7275 case X86::VMAXCPDYrr:
7276 case X86::VMAXCPSYrr:
7277 case X86::VMAXCPDZ128rr:
7278 case X86::VMAXCPSZ128rr:
7279 case X86::VMAXCPDZ256rr:
7280 case X86::VMAXCPSZ256rr:
7281 case X86::VMAXCPDZrr:
7282 case X86::VMAXCPSZrr:
7283 case X86::VMAXCSDrr:
7284 case X86::VMAXCSSrr:
7285 case X86::VMAXCSDZrr:
7286 case X86::VMAXCSSZrr:
7287 case X86::VMINCPDrr:
7288 case X86::VMINCPSrr:
7289 case X86::VMINCPDYrr:
7290 case X86::VMINCPSYrr:
7291 case X86::VMINCPDZ128rr:
7292 case X86::VMINCPSZ128rr:
7293 case X86::VMINCPDZ256rr:
7294 case X86::VMINCPSZ256rr:
7295 case X86::VMINCPDZrr:
7296 case X86::VMINCPSZrr:
7297 case X86::VMINCSDrr:
7298 case X86::VMINCSSrr:
7299 case X86::VMINCSDZrr:
7300 case X86::VMINCSSZrr:
7312 case X86::VADDPDYrr:
7313 case X86::VADDPSYrr:
7314 case X86::VADDPDZ128rr:
7315 case X86::VADDPSZ128rr:
7316 case X86::VADDPDZ256rr:
7317 case X86::VADDPSZ256rr:
7318 case X86::VADDPDZrr:
7319 case X86::VADDPSZrr:
7322 case X86::VADDSDZrr:
7323 case X86::VADDSSZrr:
7326 case X86::VMULPDYrr:
7327 case X86::VMULPSYrr:
7328 case X86::VMULPDZ128rr:
7329 case X86::VMULPSZ128rr:
7330 case X86::VMULPDZ256rr:
7331 case X86::VMULPSZ256rr:
7332 case X86::VMULPDZrr:
7333 case X86::VMULPSZrr:
7336 case X86::VMULSDZrr:
7337 case X86::VMULSSZrr:
7338 return Inst.getParent()->getParent()->getTarget().Options.UnsafeFPMath;
7344 /// This is an architecture-specific helper function of reassociateOps.
7345 /// Set special operand attributes for new instructions after reassociation.
7346 void X86InstrInfo::setSpecialOperandAttr(MachineInstr &OldMI1,
7347 MachineInstr &OldMI2,
7348 MachineInstr &NewMI1,
7349 MachineInstr &NewMI2) const {
7350 // Integer instructions define an implicit EFLAGS source register operand as
7351 // the third source (fourth total) operand.
7352 if (OldMI1.getNumOperands() != 4 || OldMI2.getNumOperands() != 4)
7355 assert(NewMI1.getNumOperands() == 4 && NewMI2.getNumOperands() == 4 &&
7356 "Unexpected instruction type for reassociation");
7358 MachineOperand &OldOp1 = OldMI1.getOperand(3);
7359 MachineOperand &OldOp2 = OldMI2.getOperand(3);
7360 MachineOperand &NewOp1 = NewMI1.getOperand(3);
7361 MachineOperand &NewOp2 = NewMI2.getOperand(3);
7363 assert(OldOp1.isReg() && OldOp1.getReg() == X86::EFLAGS && OldOp1.isDead() &&
7364 "Must have dead EFLAGS operand in reassociable instruction");
7365 assert(OldOp2.isReg() && OldOp2.getReg() == X86::EFLAGS && OldOp2.isDead() &&
7366 "Must have dead EFLAGS operand in reassociable instruction");
7371 assert(NewOp1.isReg() && NewOp1.getReg() == X86::EFLAGS &&
7372 "Unexpected operand in reassociable instruction");
7373 assert(NewOp2.isReg() && NewOp2.getReg() == X86::EFLAGS &&
7374 "Unexpected operand in reassociable instruction");
7376 // Mark the new EFLAGS operands as dead to be helpful to subsequent iterations
7377 // of this pass or other passes. The EFLAGS operands must be dead in these new
7378 // instructions because the EFLAGS operands in the original instructions must
7379 // be dead in order for reassociation to occur.
7384 std::pair<unsigned, unsigned>
7385 X86InstrInfo::decomposeMachineOperandsTargetFlags(unsigned TF) const {
7386 return std::make_pair(TF, 0u);
7389 ArrayRef<std::pair<unsigned, const char *>>
7390 X86InstrInfo::getSerializableDirectMachineOperandTargetFlags() const {
7391 using namespace X86II;
7392 static const std::pair<unsigned, const char *> TargetFlags[] = {
7393 {MO_GOT_ABSOLUTE_ADDRESS, "x86-got-absolute-address"},
7394 {MO_PIC_BASE_OFFSET, "x86-pic-base-offset"},
7395 {MO_GOT, "x86-got"},
7396 {MO_GOTOFF, "x86-gotoff"},
7397 {MO_GOTPCREL, "x86-gotpcrel"},
7398 {MO_PLT, "x86-plt"},
7399 {MO_TLSGD, "x86-tlsgd"},
7400 {MO_TLSLD, "x86-tlsld"},
7401 {MO_TLSLDM, "x86-tlsldm"},
7402 {MO_GOTTPOFF, "x86-gottpoff"},
7403 {MO_INDNTPOFF, "x86-indntpoff"},
7404 {MO_TPOFF, "x86-tpoff"},
7405 {MO_DTPOFF, "x86-dtpoff"},
7406 {MO_NTPOFF, "x86-ntpoff"},
7407 {MO_GOTNTPOFF, "x86-gotntpoff"},
7408 {MO_DLLIMPORT, "x86-dllimport"},
7409 {MO_DARWIN_NONLAZY, "x86-darwin-nonlazy"},
7410 {MO_DARWIN_NONLAZY_PIC_BASE, "x86-darwin-nonlazy-pic-base"},
7411 {MO_TLVP, "x86-tlvp"},
7412 {MO_TLVP_PIC_BASE, "x86-tlvp-pic-base"},
7413 {MO_SECREL, "x86-secrel"},
7414 {MO_COFFSTUB, "x86-coffstub"}};
7415 return makeArrayRef(TargetFlags);
7419 /// Create Global Base Reg pass. This initializes the PIC
7420 /// global base register for x86-32.
7421 struct CGBR : public MachineFunctionPass {
7423 CGBR() : MachineFunctionPass(ID) {}
7425 bool runOnMachineFunction(MachineFunction &MF) override {
7426 const X86TargetMachine *TM =
7427 static_cast<const X86TargetMachine *>(&MF.getTarget());
7428 const X86Subtarget &STI = MF.getSubtarget<X86Subtarget>();
7430 // Don't do anything in the 64-bit small and kernel code models. They use
7431 // RIP-relative addressing for everything.
7432 if (STI.is64Bit() && (TM->getCodeModel() == CodeModel::Small ||
7433 TM->getCodeModel() == CodeModel::Kernel))
7436 // Only emit a global base reg in PIC mode.
7437 if (!TM->isPositionIndependent())
7440 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
7441 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
7443 // If we didn't need a GlobalBaseReg, don't insert code.
7444 if (GlobalBaseReg == 0)
7447 // Insert the set of GlobalBaseReg into the first MBB of the function
7448 MachineBasicBlock &FirstMBB = MF.front();
7449 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
7450 DebugLoc DL = FirstMBB.findDebugLoc(MBBI);
7451 MachineRegisterInfo &RegInfo = MF.getRegInfo();
7452 const X86InstrInfo *TII = STI.getInstrInfo();
7455 if (STI.isPICStyleGOT())
7456 PC = RegInfo.createVirtualRegister(&X86::GR32RegClass);
7460 if (STI.is64Bit()) {
7461 if (TM->getCodeModel() == CodeModel::Medium) {
7462 // In the medium code model, use a RIP-relative LEA to materialize the
7464 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::LEA64r), PC)
7468 .addExternalSymbol("_GLOBAL_OFFSET_TABLE_")
7470 } else if (TM->getCodeModel() == CodeModel::Large) {
7471 // Loading the GOT in the large code model requires math with labels,
7472 // so we use a pseudo instruction and expand it during MC emission.
7473 unsigned Scratch = RegInfo.createVirtualRegister(&X86::GR64RegClass);
7474 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOVGOT64r), PC)
7475 .addReg(Scratch, RegState::Undef | RegState::Define)
7476 .addExternalSymbol("_GLOBAL_OFFSET_TABLE_");
7478 llvm_unreachable("unexpected code model");
7481 // Operand of MovePCtoStack is completely ignored by asm printer. It's
7482 // only used in JIT code emission as displacement to pc.
7483 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOVPC32r), PC).addImm(0);
7485 // If we're using vanilla 'GOT' PIC style, we should use relative
7486 // addressing not to pc, but to _GLOBAL_OFFSET_TABLE_ external.
7487 if (STI.isPICStyleGOT()) {
7488 // Generate addl $__GLOBAL_OFFSET_TABLE_ + [.-piclabel],
7490 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD32ri), GlobalBaseReg)
7492 .addExternalSymbol("_GLOBAL_OFFSET_TABLE_",
7493 X86II::MO_GOT_ABSOLUTE_ADDRESS);
7500 StringRef getPassName() const override {
7501 return "X86 PIC Global Base Reg Initialization";
7504 void getAnalysisUsage(AnalysisUsage &AU) const override {
7505 AU.setPreservesCFG();
7506 MachineFunctionPass::getAnalysisUsage(AU);
7513 llvm::createX86GlobalBaseRegPass() { return new CGBR(); }
7516 struct LDTLSCleanup : public MachineFunctionPass {
7518 LDTLSCleanup() : MachineFunctionPass(ID) {}
7520 bool runOnMachineFunction(MachineFunction &MF) override {
7521 if (skipFunction(MF.getFunction()))
7524 X86MachineFunctionInfo *MFI = MF.getInfo<X86MachineFunctionInfo>();
7525 if (MFI->getNumLocalDynamicTLSAccesses() < 2) {
7526 // No point folding accesses if there isn't at least two.
7530 MachineDominatorTree *DT = &getAnalysis<MachineDominatorTree>();
7531 return VisitNode(DT->getRootNode(), 0);
7534 // Visit the dominator subtree rooted at Node in pre-order.
7535 // If TLSBaseAddrReg is non-null, then use that to replace any
7536 // TLS_base_addr instructions. Otherwise, create the register
7537 // when the first such instruction is seen, and then use it
7538 // as we encounter more instructions.
7539 bool VisitNode(MachineDomTreeNode *Node, unsigned TLSBaseAddrReg) {
7540 MachineBasicBlock *BB = Node->getBlock();
7541 bool Changed = false;
7543 // Traverse the current block.
7544 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;
7546 switch (I->getOpcode()) {
7547 case X86::TLS_base_addr32:
7548 case X86::TLS_base_addr64:
7550 I = ReplaceTLSBaseAddrCall(*I, TLSBaseAddrReg);
7552 I = SetRegister(*I, &TLSBaseAddrReg);
7560 // Visit the children of this block in the dominator tree.
7561 for (MachineDomTreeNode::iterator I = Node->begin(), E = Node->end();
7563 Changed |= VisitNode(*I, TLSBaseAddrReg);
7569 // Replace the TLS_base_addr instruction I with a copy from
7570 // TLSBaseAddrReg, returning the new instruction.
7571 MachineInstr *ReplaceTLSBaseAddrCall(MachineInstr &I,
7572 unsigned TLSBaseAddrReg) {
7573 MachineFunction *MF = I.getParent()->getParent();
7574 const X86Subtarget &STI = MF->getSubtarget<X86Subtarget>();
7575 const bool is64Bit = STI.is64Bit();
7576 const X86InstrInfo *TII = STI.getInstrInfo();
7578 // Insert a Copy from TLSBaseAddrReg to RAX/EAX.
7579 MachineInstr *Copy =
7580 BuildMI(*I.getParent(), I, I.getDebugLoc(),
7581 TII->get(TargetOpcode::COPY), is64Bit ? X86::RAX : X86::EAX)
7582 .addReg(TLSBaseAddrReg);
7584 // Erase the TLS_base_addr instruction.
7585 I.eraseFromParent();
7590 // Create a virtual register in *TLSBaseAddrReg, and populate it by
7591 // inserting a copy instruction after I. Returns the new instruction.
7592 MachineInstr *SetRegister(MachineInstr &I, unsigned *TLSBaseAddrReg) {
7593 MachineFunction *MF = I.getParent()->getParent();
7594 const X86Subtarget &STI = MF->getSubtarget<X86Subtarget>();
7595 const bool is64Bit = STI.is64Bit();
7596 const X86InstrInfo *TII = STI.getInstrInfo();
7598 // Create a virtual register for the TLS base address.
7599 MachineRegisterInfo &RegInfo = MF->getRegInfo();
7600 *TLSBaseAddrReg = RegInfo.createVirtualRegister(is64Bit
7601 ? &X86::GR64RegClass
7602 : &X86::GR32RegClass);
7604 // Insert a copy from RAX/EAX to TLSBaseAddrReg.
7605 MachineInstr *Next = I.getNextNode();
7606 MachineInstr *Copy =
7607 BuildMI(*I.getParent(), Next, I.getDebugLoc(),
7608 TII->get(TargetOpcode::COPY), *TLSBaseAddrReg)
7609 .addReg(is64Bit ? X86::RAX : X86::EAX);
7614 StringRef getPassName() const override {
7615 return "Local Dynamic TLS Access Clean-up";
7618 void getAnalysisUsage(AnalysisUsage &AU) const override {
7619 AU.setPreservesCFG();
7620 AU.addRequired<MachineDominatorTree>();
7621 MachineFunctionPass::getAnalysisUsage(AU);
7626 char LDTLSCleanup::ID = 0;
7628 llvm::createCleanupLocalDynamicTLSPass() { return new LDTLSCleanup(); }
7630 /// Constants defining how certain sequences should be outlined.
7632 /// \p MachineOutlinerDefault implies that the function is called with a call
7633 /// instruction, and a return must be emitted for the outlined function frame.
7637 /// I1 OUTLINED_FUNCTION:
7638 /// I2 --> call OUTLINED_FUNCTION I1
7643 /// * Call construction overhead: 1 (call instruction)
7644 /// * Frame construction overhead: 1 (return instruction)
7646 /// \p MachineOutlinerTailCall implies that the function is being tail called.
7647 /// A jump is emitted instead of a call, and the return is already present in
7648 /// the outlined sequence. That is,
7650 /// I1 OUTLINED_FUNCTION:
7651 /// I2 --> jmp OUTLINED_FUNCTION I1
7655 /// * Call construction overhead: 1 (jump instruction)
7656 /// * Frame construction overhead: 0 (don't need to return)
7658 enum MachineOutlinerClass {
7659 MachineOutlinerDefault,
7660 MachineOutlinerTailCall
7663 outliner::OutlinedFunction X86InstrInfo::getOutliningCandidateInfo(
7664 std::vector<outliner::Candidate> &RepeatedSequenceLocs) const {
7665 unsigned SequenceSize =
7666 std::accumulate(RepeatedSequenceLocs[0].front(),
7667 std::next(RepeatedSequenceLocs[0].back()), 0,
7668 [](unsigned Sum, const MachineInstr &MI) {
7669 // FIXME: x86 doesn't implement getInstSizeInBytes, so
7670 // we can't tell the cost. Just assume each instruction
7672 if (MI.isDebugInstr() || MI.isKill())
7677 // FIXME: Use real size in bytes for call and ret instructions.
7678 if (RepeatedSequenceLocs[0].back()->isTerminator()) {
7679 for (outliner::Candidate &C : RepeatedSequenceLocs)
7680 C.setCallInfo(MachineOutlinerTailCall, 1);
7682 return outliner::OutlinedFunction(RepeatedSequenceLocs, SequenceSize,
7683 0, // Number of bytes to emit frame.
7684 MachineOutlinerTailCall // Type of frame.
7688 for (outliner::Candidate &C : RepeatedSequenceLocs)
7689 C.setCallInfo(MachineOutlinerDefault, 1);
7691 return outliner::OutlinedFunction(RepeatedSequenceLocs, SequenceSize, 1,
7692 MachineOutlinerDefault);
7695 bool X86InstrInfo::isFunctionSafeToOutlineFrom(MachineFunction &MF,
7696 bool OutlineFromLinkOnceODRs) const {
7697 const Function &F = MF.getFunction();
7699 // Does the function use a red zone? If it does, then we can't risk messing
7701 if (!F.hasFnAttribute(Attribute::NoRedZone)) {
7702 // It could have a red zone. If it does, then we don't want to touch it.
7703 const X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
7704 if (!X86FI || X86FI->getUsesRedZone())
7708 // If we *don't* want to outline from things that could potentially be deduped
7709 // then return false.
7710 if (!OutlineFromLinkOnceODRs && F.hasLinkOnceODRLinkage())
7713 // This function is viable for outlining, so return true.
7718 X86InstrInfo::getOutliningType(MachineBasicBlock::iterator &MIT, unsigned Flags) const {
7719 MachineInstr &MI = *MIT;
7720 // Don't allow debug values to impact outlining type.
7721 if (MI.isDebugInstr() || MI.isIndirectDebugValue())
7722 return outliner::InstrType::Invisible;
7724 // At this point, KILL instructions don't really tell us much so we can go
7725 // ahead and skip over them.
7727 return outliner::InstrType::Invisible;
7729 // Is this a tail call? If yes, we can outline as a tail call.
7731 return outliner::InstrType::Legal;
7733 // Is this the terminator of a basic block?
7734 if (MI.isTerminator() || MI.isReturn()) {
7736 // Does its parent have any successors in its MachineFunction?
7737 if (MI.getParent()->succ_empty())
7738 return outliner::InstrType::Legal;
7740 // It does, so we can't tail call it.
7741 return outliner::InstrType::Illegal;
7744 // Don't outline anything that modifies or reads from the stack pointer.
7746 // FIXME: There are instructions which are being manually built without
7747 // explicit uses/defs so we also have to check the MCInstrDesc. We should be
7748 // able to remove the extra checks once those are fixed up. For example,
7749 // sometimes we might get something like %rax = POP64r 1. This won't be
7750 // caught by modifiesRegister or readsRegister even though the instruction
7751 // really ought to be formed so that modifiesRegister/readsRegister would
7753 if (MI.modifiesRegister(X86::RSP, &RI) || MI.readsRegister(X86::RSP, &RI) ||
7754 MI.getDesc().hasImplicitUseOfPhysReg(X86::RSP) ||
7755 MI.getDesc().hasImplicitDefOfPhysReg(X86::RSP))
7756 return outliner::InstrType::Illegal;
7758 // Outlined calls change the instruction pointer, so don't read from it.
7759 if (MI.readsRegister(X86::RIP, &RI) ||
7760 MI.getDesc().hasImplicitUseOfPhysReg(X86::RIP) ||
7761 MI.getDesc().hasImplicitDefOfPhysReg(X86::RIP))
7762 return outliner::InstrType::Illegal;
7764 // Positions can't safely be outlined.
7765 if (MI.isPosition())
7766 return outliner::InstrType::Illegal;
7768 // Make sure none of the operands of this instruction do anything tricky.
7769 for (const MachineOperand &MOP : MI.operands())
7770 if (MOP.isCPI() || MOP.isJTI() || MOP.isCFIIndex() || MOP.isFI() ||
7771 MOP.isTargetIndex())
7772 return outliner::InstrType::Illegal;
7774 return outliner::InstrType::Legal;
7777 void X86InstrInfo::buildOutlinedFrame(MachineBasicBlock &MBB,
7778 MachineFunction &MF,
7779 const outliner::OutlinedFunction &OF)
7781 // If we're a tail call, we already have a return, so don't do anything.
7782 if (OF.FrameConstructionID == MachineOutlinerTailCall)
7785 // We're a normal call, so our sequence doesn't have a return instruction.
7787 MachineInstr *retq = BuildMI(MF, DebugLoc(), get(X86::RETQ));
7788 MBB.insert(MBB.end(), retq);
7791 MachineBasicBlock::iterator
7792 X86InstrInfo::insertOutlinedCall(Module &M, MachineBasicBlock &MBB,
7793 MachineBasicBlock::iterator &It,
7794 MachineFunction &MF,
7795 const outliner::Candidate &C) const {
7796 // Is it a tail call?
7797 if (C.CallConstructionID == MachineOutlinerTailCall) {
7798 // Yes, just insert a JMP.
7800 BuildMI(MF, DebugLoc(), get(X86::TAILJMPd64))
7801 .addGlobalAddress(M.getNamedValue(MF.getName())));
7803 // No, insert a call.
7805 BuildMI(MF, DebugLoc(), get(X86::CALL64pcrel32))
7806 .addGlobalAddress(M.getNamedValue(MF.getName())));