1 //===-- X86InstrInfo.cpp - X86 Instruction Information --------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file contains the X86 implementation of the TargetInstrInfo class.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "X86InstrInfo.h"
14 #include "X86.h"
15 #include "X86InstrBuilder.h"
16 #include "X86InstrFoldTables.h"
17 #include "X86MachineFunctionInfo.h"
18 #include "X86Subtarget.h"
19 #include "X86TargetMachine.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/ADT/Sequence.h"
22 #include "llvm/CodeGen/LiveIntervals.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/MachineInstr.h"
29 #include "llvm/CodeGen/MachineInstrBuilder.h"
30 #include "llvm/CodeGen/MachineModuleInfo.h"
31 #include "llvm/CodeGen/MachineOperand.h"
32 #include "llvm/CodeGen/MachineRegisterInfo.h"
33 #include "llvm/CodeGen/StackMaps.h"
34 #include "llvm/IR/DebugInfoMetadata.h"
35 #include "llvm/IR/DerivedTypes.h"
36 #include "llvm/IR/Function.h"
37 #include "llvm/IR/InstrTypes.h"
38 #include "llvm/MC/MCAsmInfo.h"
39 #include "llvm/MC/MCExpr.h"
40 #include "llvm/MC/MCInst.h"
41 #include "llvm/Support/CommandLine.h"
42 #include "llvm/Support/Debug.h"
43 #include "llvm/Support/ErrorHandling.h"
44 #include "llvm/Support/raw_ostream.h"
45 #include "llvm/Target/TargetOptions.h"
46 
47 using namespace llvm;
48 
49 #define DEBUG_TYPE "x86-instr-info"
50 
51 #define GET_INSTRINFO_CTOR_DTOR
52 #include "X86GenInstrInfo.inc"
53 
54 static cl::opt<bool>
55     NoFusing("disable-spill-fusing",
56              cl::desc("Disable fusing of spill code into instructions"),
57              cl::Hidden);
58 static cl::opt<bool>
59 PrintFailedFusing("print-failed-fuse-candidates",
60                   cl::desc("Print instructions that the allocator wants to"
61                            " fuse, but the X86 backend currently can't"),
62                   cl::Hidden);
63 static cl::opt<bool>
64 ReMatPICStubLoad("remat-pic-stub-load",
65                  cl::desc("Re-materialize load from stub in PIC mode"),
66                  cl::init(false), cl::Hidden);
67 static cl::opt<unsigned>
68 PartialRegUpdateClearance("partial-reg-update-clearance",
69                           cl::desc("Clearance between two register writes "
70                                    "for inserting XOR to avoid partial "
71                                    "register update"),
72                           cl::init(64), cl::Hidden);
73 static cl::opt<unsigned>
74 UndefRegClearance("undef-reg-clearance",
75                   cl::desc("How many idle instructions we would like before "
76                            "certain undef register reads"),
77                   cl::init(128), cl::Hidden);
78 
79 
80 // Pin the vtable to this file.
81 void X86InstrInfo::anchor() {}
82 
83 X86InstrInfo::X86InstrInfo(X86Subtarget &STI)
84     : X86GenInstrInfo((STI.isTarget64BitLP64() ? X86::ADJCALLSTACKDOWN64
85                                                : X86::ADJCALLSTACKDOWN32),
86                       (STI.isTarget64BitLP64() ? X86::ADJCALLSTACKUP64
87                                                : X86::ADJCALLSTACKUP32),
88                       X86::CATCHRET,
89                       (STI.is64Bit() ? X86::RET64 : X86::RET32)),
90       Subtarget(STI), RI(STI.getTargetTriple()) {
91 }
92 
93 bool
94 X86InstrInfo::isCoalescableExtInstr(const MachineInstr &MI,
95                                     Register &SrcReg, Register &DstReg,
96                                     unsigned &SubIdx) const {
97   switch (MI.getOpcode()) {
98   default: break;
99   case X86::MOVSX16rr8:
100   case X86::MOVZX16rr8:
101   case X86::MOVSX32rr8:
102   case X86::MOVZX32rr8:
103   case X86::MOVSX64rr8:
104     if (!Subtarget.is64Bit())
105       // It's not always legal to reference the low 8-bit of the larger
106       // register in 32-bit mode.
107       return false;
108     LLVM_FALLTHROUGH;
109   case X86::MOVSX32rr16:
110   case X86::MOVZX32rr16:
111   case X86::MOVSX64rr16:
112   case X86::MOVSX64rr32: {
113     if (MI.getOperand(0).getSubReg() || MI.getOperand(1).getSubReg())
114       // Be conservative.
115       return false;
116     SrcReg = MI.getOperand(1).getReg();
117     DstReg = MI.getOperand(0).getReg();
118     switch (MI.getOpcode()) {
119     default: llvm_unreachable("Unreachable!");
120     case X86::MOVSX16rr8:
121     case X86::MOVZX16rr8:
122     case X86::MOVSX32rr8:
123     case X86::MOVZX32rr8:
124     case X86::MOVSX64rr8:
125       SubIdx = X86::sub_8bit;
126       break;
127     case X86::MOVSX32rr16:
128     case X86::MOVZX32rr16:
129     case X86::MOVSX64rr16:
130       SubIdx = X86::sub_16bit;
131       break;
132     case X86::MOVSX64rr32:
133       SubIdx = X86::sub_32bit;
134       break;
135     }
136     return true;
137   }
138   }
139   return false;
140 }
141 
142 bool X86InstrInfo::isDataInvariant(MachineInstr &MI) {
143   if (MI.mayLoad() || MI.mayStore())
144     return false;
145 
146   // Some target-independent operations that trivially lower to data-invariant
147   // instructions.
148   if (MI.isCopyLike() || MI.isInsertSubreg())
149     return true;
150 
151   unsigned Opcode = MI.getOpcode();
152   using namespace X86;
153   // On x86 it is believed that imul is constant time w.r.t. the loaded data.
154   // However, they set flags and are perhaps the most surprisingly constant
155   // time operations so we call them out here separately.
156   if (isIMUL(Opcode))
157     return true;
158   // Bit scanning and counting instructions that are somewhat surprisingly
159   // constant time as they scan across bits and do other fairly complex
160   // operations like popcnt, but are believed to be constant time on x86.
161   // However, these set flags.
162   if (isBSF(Opcode) || isBSR(Opcode) || isLZCNT(Opcode) || isPOPCNT(Opcode) ||
163       isTZCNT(Opcode))
164     return true;
165   // Bit manipulation instructions are effectively combinations of basic
166   // arithmetic ops, and should still execute in constant time. These also
167   // set flags.
168   if (isBLCFILL(Opcode) || isBLCI(Opcode) || isBLCIC(Opcode) ||
169       isBLCMSK(Opcode) || isBLCS(Opcode) || isBLSFILL(Opcode) ||
170       isBLSI(Opcode) || isBLSIC(Opcode) || isBLSMSK(Opcode) || isBLSR(Opcode) ||
171       isTZMSK(Opcode))
172     return true;
173   // Bit extracting and clearing instructions should execute in constant time,
174   // and set flags.
175   if (isBEXTR(Opcode) || isBZHI(Opcode))
176     return true;
177   // Shift and rotate.
178   if (isROL(Opcode) || isROR(Opcode) || isSAR(Opcode) || isSHL(Opcode) ||
179       isSHR(Opcode) || isSHLD(Opcode) || isSHRD(Opcode))
180     return true;
181   // Basic arithmetic is constant time on the input but does set flags.
182   if (isADC(Opcode) || isADD(Opcode) || isAND(Opcode) || isOR(Opcode) ||
183       isSBB(Opcode) || isSUB(Opcode) || isXOR(Opcode))
184     return true;
185   // Arithmetic with just 32-bit and 64-bit variants and no immediates.
186   if (isADCX(Opcode) || isADOX(Opcode) || isANDN(Opcode))
187     return true;
188   // Unary arithmetic operations.
189   if (isDEC(Opcode) || isINC(Opcode) || isNEG(Opcode))
190     return true;
191   // Unlike other arithmetic, NOT doesn't set EFLAGS.
192   if (isNOT(Opcode))
193     return true;
194   // Various move instructions used to zero or sign extend things. Note that we
195   // intentionally don't support the _NOREX variants as we can't handle that
196   // register constraint anyways.
197   if (isMOVSX(Opcode) || isMOVZX(Opcode) || isMOVSXD(Opcode) || isMOV(Opcode))
198     return true;
199   // Arithmetic instructions that are both constant time and don't set flags.
200   if (isRORX(Opcode) || isSARX(Opcode) || isSHLX(Opcode) || isSHRX(Opcode))
201     return true;
202   // LEA doesn't actually access memory, and its arithmetic is constant time.
203   if (isLEA(Opcode))
204     return true;
205   // By default, assume that the instruction is not data invariant.
206   return false;
207 }
208 
209 bool X86InstrInfo::isDataInvariantLoad(MachineInstr &MI) {
210   switch (MI.getOpcode()) {
211   default:
212     // By default, assume that the load will immediately leak.
213     return false;
214 
215   // On x86 it is believed that imul is constant time w.r.t. the loaded data.
216   // However, they set flags and are perhaps the most surprisingly constant
217   // time operations so we call them out here separately.
218   case X86::IMUL16rm:
219   case X86::IMUL16rmi8:
220   case X86::IMUL16rmi:
221   case X86::IMUL32rm:
222   case X86::IMUL32rmi8:
223   case X86::IMUL32rmi:
224   case X86::IMUL64rm:
225   case X86::IMUL64rmi32:
226   case X86::IMUL64rmi8:
227 
228   // Bit scanning and counting instructions that are somewhat surprisingly
229   // constant time as they scan across bits and do other fairly complex
230   // operations like popcnt, but are believed to be constant time on x86.
231   // However, these set flags.
232   case X86::BSF16rm:
233   case X86::BSF32rm:
234   case X86::BSF64rm:
235   case X86::BSR16rm:
236   case X86::BSR32rm:
237   case X86::BSR64rm:
238   case X86::LZCNT16rm:
239   case X86::LZCNT32rm:
240   case X86::LZCNT64rm:
241   case X86::POPCNT16rm:
242   case X86::POPCNT32rm:
243   case X86::POPCNT64rm:
244   case X86::TZCNT16rm:
245   case X86::TZCNT32rm:
246   case X86::TZCNT64rm:
247 
248   // Bit manipulation instructions are effectively combinations of basic
249   // arithmetic ops, and should still execute in constant time. These also
250   // set flags.
251   case X86::BLCFILL32rm:
252   case X86::BLCFILL64rm:
253   case X86::BLCI32rm:
254   case X86::BLCI64rm:
255   case X86::BLCIC32rm:
256   case X86::BLCIC64rm:
257   case X86::BLCMSK32rm:
258   case X86::BLCMSK64rm:
259   case X86::BLCS32rm:
260   case X86::BLCS64rm:
261   case X86::BLSFILL32rm:
262   case X86::BLSFILL64rm:
263   case X86::BLSI32rm:
264   case X86::BLSI64rm:
265   case X86::BLSIC32rm:
266   case X86::BLSIC64rm:
267   case X86::BLSMSK32rm:
268   case X86::BLSMSK64rm:
269   case X86::BLSR32rm:
270   case X86::BLSR64rm:
271   case X86::TZMSK32rm:
272   case X86::TZMSK64rm:
273 
274   // Bit extracting and clearing instructions should execute in constant time,
275   // and set flags.
276   case X86::BEXTR32rm:
277   case X86::BEXTR64rm:
278   case X86::BEXTRI32mi:
279   case X86::BEXTRI64mi:
280   case X86::BZHI32rm:
281   case X86::BZHI64rm:
282 
283   // Basic arithmetic is constant time on the input but does set flags.
284   case X86::ADC8rm:
285   case X86::ADC16rm:
286   case X86::ADC32rm:
287   case X86::ADC64rm:
288   case X86::ADCX32rm:
289   case X86::ADCX64rm:
290   case X86::ADD8rm:
291   case X86::ADD16rm:
292   case X86::ADD32rm:
293   case X86::ADD64rm:
294   case X86::ADOX32rm:
295   case X86::ADOX64rm:
296   case X86::AND8rm:
297   case X86::AND16rm:
298   case X86::AND32rm:
299   case X86::AND64rm:
300   case X86::ANDN32rm:
301   case X86::ANDN64rm:
302   case X86::OR8rm:
303   case X86::OR16rm:
304   case X86::OR32rm:
305   case X86::OR64rm:
306   case X86::SBB8rm:
307   case X86::SBB16rm:
308   case X86::SBB32rm:
309   case X86::SBB64rm:
310   case X86::SUB8rm:
311   case X86::SUB16rm:
312   case X86::SUB32rm:
313   case X86::SUB64rm:
314   case X86::XOR8rm:
315   case X86::XOR16rm:
316   case X86::XOR32rm:
317   case X86::XOR64rm:
318 
319   // Integer multiply w/o affecting flags is still believed to be constant
320   // time on x86. Called out separately as this is among the most surprising
321   // instructions to exhibit that behavior.
322   case X86::MULX32rm:
323   case X86::MULX64rm:
324 
325   // Arithmetic instructions that are both constant time and don't set flags.
326   case X86::RORX32mi:
327   case X86::RORX64mi:
328   case X86::SARX32rm:
329   case X86::SARX64rm:
330   case X86::SHLX32rm:
331   case X86::SHLX64rm:
332   case X86::SHRX32rm:
333   case X86::SHRX64rm:
334 
335   // Conversions are believed to be constant time and don't set flags.
336   case X86::CVTTSD2SI64rm:
337   case X86::VCVTTSD2SI64rm:
338   case X86::VCVTTSD2SI64Zrm:
339   case X86::CVTTSD2SIrm:
340   case X86::VCVTTSD2SIrm:
341   case X86::VCVTTSD2SIZrm:
342   case X86::CVTTSS2SI64rm:
343   case X86::VCVTTSS2SI64rm:
344   case X86::VCVTTSS2SI64Zrm:
345   case X86::CVTTSS2SIrm:
346   case X86::VCVTTSS2SIrm:
347   case X86::VCVTTSS2SIZrm:
348   case X86::CVTSI2SDrm:
349   case X86::VCVTSI2SDrm:
350   case X86::VCVTSI2SDZrm:
351   case X86::CVTSI2SSrm:
352   case X86::VCVTSI2SSrm:
353   case X86::VCVTSI2SSZrm:
354   case X86::CVTSI642SDrm:
355   case X86::VCVTSI642SDrm:
356   case X86::VCVTSI642SDZrm:
357   case X86::CVTSI642SSrm:
358   case X86::VCVTSI642SSrm:
359   case X86::VCVTSI642SSZrm:
360   case X86::CVTSS2SDrm:
361   case X86::VCVTSS2SDrm:
362   case X86::VCVTSS2SDZrm:
363   case X86::CVTSD2SSrm:
364   case X86::VCVTSD2SSrm:
365   case X86::VCVTSD2SSZrm:
366   // AVX512 added unsigned integer conversions.
367   case X86::VCVTTSD2USI64Zrm:
368   case X86::VCVTTSD2USIZrm:
369   case X86::VCVTTSS2USI64Zrm:
370   case X86::VCVTTSS2USIZrm:
371   case X86::VCVTUSI2SDZrm:
372   case X86::VCVTUSI642SDZrm:
373   case X86::VCVTUSI2SSZrm:
374   case X86::VCVTUSI642SSZrm:
375 
376   // Loads to register don't set flags.
377   case X86::MOV8rm:
378   case X86::MOV8rm_NOREX:
379   case X86::MOV16rm:
380   case X86::MOV32rm:
381   case X86::MOV64rm:
382   case X86::MOVSX16rm8:
383   case X86::MOVSX32rm16:
384   case X86::MOVSX32rm8:
385   case X86::MOVSX32rm8_NOREX:
386   case X86::MOVSX64rm16:
387   case X86::MOVSX64rm32:
388   case X86::MOVSX64rm8:
389   case X86::MOVZX16rm8:
390   case X86::MOVZX32rm16:
391   case X86::MOVZX32rm8:
392   case X86::MOVZX32rm8_NOREX:
393   case X86::MOVZX64rm16:
394   case X86::MOVZX64rm8:
395     return true;
396   }
397 }
398 
399 int X86InstrInfo::getSPAdjust(const MachineInstr &MI) const {
400   const MachineFunction *MF = MI.getParent()->getParent();
401   const TargetFrameLowering *TFI = MF->getSubtarget().getFrameLowering();
402 
403   if (isFrameInstr(MI)) {
404     int SPAdj = alignTo(getFrameSize(MI), TFI->getStackAlign());
405     SPAdj -= getFrameAdjustment(MI);
406     if (!isFrameSetup(MI))
407       SPAdj = -SPAdj;
408     return SPAdj;
409   }
410 
411   // To know whether a call adjusts the stack, we need information
412   // that is bound to the following ADJCALLSTACKUP pseudo.
413   // Look for the next ADJCALLSTACKUP that follows the call.
414   if (MI.isCall()) {
415     const MachineBasicBlock *MBB = MI.getParent();
416     auto I = ++MachineBasicBlock::const_iterator(MI);
417     for (auto E = MBB->end(); I != E; ++I) {
418       if (I->getOpcode() == getCallFrameDestroyOpcode() ||
419           I->isCall())
420         break;
421     }
422 
423     // If we could not find a frame destroy opcode, then it has already
424     // been simplified, so we don't care.
425     if (I->getOpcode() != getCallFrameDestroyOpcode())
426       return 0;
427 
428     return -(I->getOperand(1).getImm());
429   }
430 
431   // Currently handle only PUSHes we can reasonably expect to see
432   // in call sequences
433   switch (MI.getOpcode()) {
434   default:
435     return 0;
436   case X86::PUSH32i8:
437   case X86::PUSH32r:
438   case X86::PUSH32rmm:
439   case X86::PUSH32rmr:
440   case X86::PUSHi32:
441     return 4;
442   case X86::PUSH64i8:
443   case X86::PUSH64r:
444   case X86::PUSH64rmm:
445   case X86::PUSH64rmr:
446   case X86::PUSH64i32:
447     return 8;
448   }
449 }
450 
451 /// Return true and the FrameIndex if the specified
452 /// operand and follow operands form a reference to the stack frame.
453 bool X86InstrInfo::isFrameOperand(const MachineInstr &MI, unsigned int Op,
454                                   int &FrameIndex) const {
455   if (MI.getOperand(Op + X86::AddrBaseReg).isFI() &&
456       MI.getOperand(Op + X86::AddrScaleAmt).isImm() &&
457       MI.getOperand(Op + X86::AddrIndexReg).isReg() &&
458       MI.getOperand(Op + X86::AddrDisp).isImm() &&
459       MI.getOperand(Op + X86::AddrScaleAmt).getImm() == 1 &&
460       MI.getOperand(Op + X86::AddrIndexReg).getReg() == 0 &&
461       MI.getOperand(Op + X86::AddrDisp).getImm() == 0) {
462     FrameIndex = MI.getOperand(Op + X86::AddrBaseReg).getIndex();
463     return true;
464   }
465   return false;
466 }
467 
468 static bool isFrameLoadOpcode(int Opcode, unsigned &MemBytes) {
469   switch (Opcode) {
470   default:
471     return false;
472   case X86::MOV8rm:
473   case X86::KMOVBkm:
474     MemBytes = 1;
475     return true;
476   case X86::MOV16rm:
477   case X86::KMOVWkm:
478   case X86::VMOVSHZrm:
479   case X86::VMOVSHZrm_alt:
480     MemBytes = 2;
481     return true;
482   case X86::MOV32rm:
483   case X86::MOVSSrm:
484   case X86::MOVSSrm_alt:
485   case X86::VMOVSSrm:
486   case X86::VMOVSSrm_alt:
487   case X86::VMOVSSZrm:
488   case X86::VMOVSSZrm_alt:
489   case X86::KMOVDkm:
490     MemBytes = 4;
491     return true;
492   case X86::MOV64rm:
493   case X86::LD_Fp64m:
494   case X86::MOVSDrm:
495   case X86::MOVSDrm_alt:
496   case X86::VMOVSDrm:
497   case X86::VMOVSDrm_alt:
498   case X86::VMOVSDZrm:
499   case X86::VMOVSDZrm_alt:
500   case X86::MMX_MOVD64rm:
501   case X86::MMX_MOVQ64rm:
502   case X86::KMOVQkm:
503     MemBytes = 8;
504     return true;
505   case X86::MOVAPSrm:
506   case X86::MOVUPSrm:
507   case X86::MOVAPDrm:
508   case X86::MOVUPDrm:
509   case X86::MOVDQArm:
510   case X86::MOVDQUrm:
511   case X86::VMOVAPSrm:
512   case X86::VMOVUPSrm:
513   case X86::VMOVAPDrm:
514   case X86::VMOVUPDrm:
515   case X86::VMOVDQArm:
516   case X86::VMOVDQUrm:
517   case X86::VMOVAPSZ128rm:
518   case X86::VMOVUPSZ128rm:
519   case X86::VMOVAPSZ128rm_NOVLX:
520   case X86::VMOVUPSZ128rm_NOVLX:
521   case X86::VMOVAPDZ128rm:
522   case X86::VMOVUPDZ128rm:
523   case X86::VMOVDQU8Z128rm:
524   case X86::VMOVDQU16Z128rm:
525   case X86::VMOVDQA32Z128rm:
526   case X86::VMOVDQU32Z128rm:
527   case X86::VMOVDQA64Z128rm:
528   case X86::VMOVDQU64Z128rm:
529     MemBytes = 16;
530     return true;
531   case X86::VMOVAPSYrm:
532   case X86::VMOVUPSYrm:
533   case X86::VMOVAPDYrm:
534   case X86::VMOVUPDYrm:
535   case X86::VMOVDQAYrm:
536   case X86::VMOVDQUYrm:
537   case X86::VMOVAPSZ256rm:
538   case X86::VMOVUPSZ256rm:
539   case X86::VMOVAPSZ256rm_NOVLX:
540   case X86::VMOVUPSZ256rm_NOVLX:
541   case X86::VMOVAPDZ256rm:
542   case X86::VMOVUPDZ256rm:
543   case X86::VMOVDQU8Z256rm:
544   case X86::VMOVDQU16Z256rm:
545   case X86::VMOVDQA32Z256rm:
546   case X86::VMOVDQU32Z256rm:
547   case X86::VMOVDQA64Z256rm:
548   case X86::VMOVDQU64Z256rm:
549     MemBytes = 32;
550     return true;
551   case X86::VMOVAPSZrm:
552   case X86::VMOVUPSZrm:
553   case X86::VMOVAPDZrm:
554   case X86::VMOVUPDZrm:
555   case X86::VMOVDQU8Zrm:
556   case X86::VMOVDQU16Zrm:
557   case X86::VMOVDQA32Zrm:
558   case X86::VMOVDQU32Zrm:
559   case X86::VMOVDQA64Zrm:
560   case X86::VMOVDQU64Zrm:
561     MemBytes = 64;
562     return true;
563   }
564 }
565 
566 static bool isFrameStoreOpcode(int Opcode, unsigned &MemBytes) {
567   switch (Opcode) {
568   default:
569     return false;
570   case X86::MOV8mr:
571   case X86::KMOVBmk:
572     MemBytes = 1;
573     return true;
574   case X86::MOV16mr:
575   case X86::KMOVWmk:
576   case X86::VMOVSHZmr:
577     MemBytes = 2;
578     return true;
579   case X86::MOV32mr:
580   case X86::MOVSSmr:
581   case X86::VMOVSSmr:
582   case X86::VMOVSSZmr:
583   case X86::KMOVDmk:
584     MemBytes = 4;
585     return true;
586   case X86::MOV64mr:
587   case X86::ST_FpP64m:
588   case X86::MOVSDmr:
589   case X86::VMOVSDmr:
590   case X86::VMOVSDZmr:
591   case X86::MMX_MOVD64mr:
592   case X86::MMX_MOVQ64mr:
593   case X86::MMX_MOVNTQmr:
594   case X86::KMOVQmk:
595     MemBytes = 8;
596     return true;
597   case X86::MOVAPSmr:
598   case X86::MOVUPSmr:
599   case X86::MOVAPDmr:
600   case X86::MOVUPDmr:
601   case X86::MOVDQAmr:
602   case X86::MOVDQUmr:
603   case X86::VMOVAPSmr:
604   case X86::VMOVUPSmr:
605   case X86::VMOVAPDmr:
606   case X86::VMOVUPDmr:
607   case X86::VMOVDQAmr:
608   case X86::VMOVDQUmr:
609   case X86::VMOVUPSZ128mr:
610   case X86::VMOVAPSZ128mr:
611   case X86::VMOVUPSZ128mr_NOVLX:
612   case X86::VMOVAPSZ128mr_NOVLX:
613   case X86::VMOVUPDZ128mr:
614   case X86::VMOVAPDZ128mr:
615   case X86::VMOVDQA32Z128mr:
616   case X86::VMOVDQU32Z128mr:
617   case X86::VMOVDQA64Z128mr:
618   case X86::VMOVDQU64Z128mr:
619   case X86::VMOVDQU8Z128mr:
620   case X86::VMOVDQU16Z128mr:
621     MemBytes = 16;
622     return true;
623   case X86::VMOVUPSYmr:
624   case X86::VMOVAPSYmr:
625   case X86::VMOVUPDYmr:
626   case X86::VMOVAPDYmr:
627   case X86::VMOVDQUYmr:
628   case X86::VMOVDQAYmr:
629   case X86::VMOVUPSZ256mr:
630   case X86::VMOVAPSZ256mr:
631   case X86::VMOVUPSZ256mr_NOVLX:
632   case X86::VMOVAPSZ256mr_NOVLX:
633   case X86::VMOVUPDZ256mr:
634   case X86::VMOVAPDZ256mr:
635   case X86::VMOVDQU8Z256mr:
636   case X86::VMOVDQU16Z256mr:
637   case X86::VMOVDQA32Z256mr:
638   case X86::VMOVDQU32Z256mr:
639   case X86::VMOVDQA64Z256mr:
640   case X86::VMOVDQU64Z256mr:
641     MemBytes = 32;
642     return true;
643   case X86::VMOVUPSZmr:
644   case X86::VMOVAPSZmr:
645   case X86::VMOVUPDZmr:
646   case X86::VMOVAPDZmr:
647   case X86::VMOVDQU8Zmr:
648   case X86::VMOVDQU16Zmr:
649   case X86::VMOVDQA32Zmr:
650   case X86::VMOVDQU32Zmr:
651   case X86::VMOVDQA64Zmr:
652   case X86::VMOVDQU64Zmr:
653     MemBytes = 64;
654     return true;
655   }
656   return false;
657 }
658 
659 unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr &MI,
660                                            int &FrameIndex) const {
661   unsigned Dummy;
662   return X86InstrInfo::isLoadFromStackSlot(MI, FrameIndex, Dummy);
663 }
664 
665 unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr &MI,
666                                            int &FrameIndex,
667                                            unsigned &MemBytes) const {
668   if (isFrameLoadOpcode(MI.getOpcode(), MemBytes))
669     if (MI.getOperand(0).getSubReg() == 0 && isFrameOperand(MI, 1, FrameIndex))
670       return MI.getOperand(0).getReg();
671   return 0;
672 }
673 
674 unsigned X86InstrInfo::isLoadFromStackSlotPostFE(const MachineInstr &MI,
675                                                  int &FrameIndex) const {
676   unsigned Dummy;
677   if (isFrameLoadOpcode(MI.getOpcode(), Dummy)) {
678     unsigned Reg;
679     if ((Reg = isLoadFromStackSlot(MI, FrameIndex)))
680       return Reg;
681     // Check for post-frame index elimination operations
682     SmallVector<const MachineMemOperand *, 1> Accesses;
683     if (hasLoadFromStackSlot(MI, Accesses)) {
684       FrameIndex =
685           cast<FixedStackPseudoSourceValue>(Accesses.front()->getPseudoValue())
686               ->getFrameIndex();
687       return MI.getOperand(0).getReg();
688     }
689   }
690   return 0;
691 }
692 
693 unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr &MI,
694                                           int &FrameIndex) const {
695   unsigned Dummy;
696   return X86InstrInfo::isStoreToStackSlot(MI, FrameIndex, Dummy);
697 }
698 
699 unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr &MI,
700                                           int &FrameIndex,
701                                           unsigned &MemBytes) const {
702   if (isFrameStoreOpcode(MI.getOpcode(), MemBytes))
703     if (MI.getOperand(X86::AddrNumOperands).getSubReg() == 0 &&
704         isFrameOperand(MI, 0, FrameIndex))
705       return MI.getOperand(X86::AddrNumOperands).getReg();
706   return 0;
707 }
708 
709 unsigned X86InstrInfo::isStoreToStackSlotPostFE(const MachineInstr &MI,
710                                                 int &FrameIndex) const {
711   unsigned Dummy;
712   if (isFrameStoreOpcode(MI.getOpcode(), Dummy)) {
713     unsigned Reg;
714     if ((Reg = isStoreToStackSlot(MI, FrameIndex)))
715       return Reg;
716     // Check for post-frame index elimination operations
717     SmallVector<const MachineMemOperand *, 1> Accesses;
718     if (hasStoreToStackSlot(MI, Accesses)) {
719       FrameIndex =
720           cast<FixedStackPseudoSourceValue>(Accesses.front()->getPseudoValue())
721               ->getFrameIndex();
722       return MI.getOperand(X86::AddrNumOperands).getReg();
723     }
724   }
725   return 0;
726 }
727 
728 /// Return true if register is PIC base; i.e.g defined by X86::MOVPC32r.
729 static bool regIsPICBase(Register BaseReg, const MachineRegisterInfo &MRI) {
730   // Don't waste compile time scanning use-def chains of physregs.
731   if (!BaseReg.isVirtual())
732     return false;
733   bool isPICBase = false;
734   for (MachineRegisterInfo::def_instr_iterator I = MRI.def_instr_begin(BaseReg),
735          E = MRI.def_instr_end(); I != E; ++I) {
736     MachineInstr *DefMI = &*I;
737     if (DefMI->getOpcode() != X86::MOVPC32r)
738       return false;
739     assert(!isPICBase && "More than one PIC base?");
740     isPICBase = true;
741   }
742   return isPICBase;
743 }
744 
745 bool X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr &MI,
746                                                      AAResults *AA) const {
747   switch (MI.getOpcode()) {
748   default:
749     // This function should only be called for opcodes with the ReMaterializable
750     // flag set.
751     llvm_unreachable("Unknown rematerializable operation!");
752     break;
753 
754   case X86::LOAD_STACK_GUARD:
755   case X86::AVX1_SETALLONES:
756   case X86::AVX2_SETALLONES:
757   case X86::AVX512_128_SET0:
758   case X86::AVX512_256_SET0:
759   case X86::AVX512_512_SET0:
760   case X86::AVX512_512_SETALLONES:
761   case X86::AVX512_FsFLD0SD:
762   case X86::AVX512_FsFLD0SH:
763   case X86::AVX512_FsFLD0SS:
764   case X86::AVX512_FsFLD0F128:
765   case X86::AVX_SET0:
766   case X86::FsFLD0SD:
767   case X86::FsFLD0SS:
768   case X86::FsFLD0F128:
769   case X86::KSET0D:
770   case X86::KSET0Q:
771   case X86::KSET0W:
772   case X86::KSET1D:
773   case X86::KSET1Q:
774   case X86::KSET1W:
775   case X86::MMX_SET0:
776   case X86::MOV32ImmSExti8:
777   case X86::MOV32r0:
778   case X86::MOV32r1:
779   case X86::MOV32r_1:
780   case X86::MOV32ri64:
781   case X86::MOV64ImmSExti8:
782   case X86::V_SET0:
783   case X86::V_SETALLONES:
784   case X86::MOV16ri:
785   case X86::MOV32ri:
786   case X86::MOV64ri:
787   case X86::MOV64ri32:
788   case X86::MOV8ri:
789   case X86::PTILEZEROV:
790     return true;
791 
792   case X86::MOV8rm:
793   case X86::MOV8rm_NOREX:
794   case X86::MOV16rm:
795   case X86::MOV32rm:
796   case X86::MOV64rm:
797   case X86::MOVSSrm:
798   case X86::MOVSSrm_alt:
799   case X86::MOVSDrm:
800   case X86::MOVSDrm_alt:
801   case X86::MOVAPSrm:
802   case X86::MOVUPSrm:
803   case X86::MOVAPDrm:
804   case X86::MOVUPDrm:
805   case X86::MOVDQArm:
806   case X86::MOVDQUrm:
807   case X86::VMOVSSrm:
808   case X86::VMOVSSrm_alt:
809   case X86::VMOVSDrm:
810   case X86::VMOVSDrm_alt:
811   case X86::VMOVAPSrm:
812   case X86::VMOVUPSrm:
813   case X86::VMOVAPDrm:
814   case X86::VMOVUPDrm:
815   case X86::VMOVDQArm:
816   case X86::VMOVDQUrm:
817   case X86::VMOVAPSYrm:
818   case X86::VMOVUPSYrm:
819   case X86::VMOVAPDYrm:
820   case X86::VMOVUPDYrm:
821   case X86::VMOVDQAYrm:
822   case X86::VMOVDQUYrm:
823   case X86::MMX_MOVD64rm:
824   case X86::MMX_MOVQ64rm:
825   // AVX-512
826   case X86::VMOVSSZrm:
827   case X86::VMOVSSZrm_alt:
828   case X86::VMOVSDZrm:
829   case X86::VMOVSDZrm_alt:
830   case X86::VMOVSHZrm:
831   case X86::VMOVSHZrm_alt:
832   case X86::VMOVAPDZ128rm:
833   case X86::VMOVAPDZ256rm:
834   case X86::VMOVAPDZrm:
835   case X86::VMOVAPSZ128rm:
836   case X86::VMOVAPSZ256rm:
837   case X86::VMOVAPSZ128rm_NOVLX:
838   case X86::VMOVAPSZ256rm_NOVLX:
839   case X86::VMOVAPSZrm:
840   case X86::VMOVDQA32Z128rm:
841   case X86::VMOVDQA32Z256rm:
842   case X86::VMOVDQA32Zrm:
843   case X86::VMOVDQA64Z128rm:
844   case X86::VMOVDQA64Z256rm:
845   case X86::VMOVDQA64Zrm:
846   case X86::VMOVDQU16Z128rm:
847   case X86::VMOVDQU16Z256rm:
848   case X86::VMOVDQU16Zrm:
849   case X86::VMOVDQU32Z128rm:
850   case X86::VMOVDQU32Z256rm:
851   case X86::VMOVDQU32Zrm:
852   case X86::VMOVDQU64Z128rm:
853   case X86::VMOVDQU64Z256rm:
854   case X86::VMOVDQU64Zrm:
855   case X86::VMOVDQU8Z128rm:
856   case X86::VMOVDQU8Z256rm:
857   case X86::VMOVDQU8Zrm:
858   case X86::VMOVUPDZ128rm:
859   case X86::VMOVUPDZ256rm:
860   case X86::VMOVUPDZrm:
861   case X86::VMOVUPSZ128rm:
862   case X86::VMOVUPSZ256rm:
863   case X86::VMOVUPSZ128rm_NOVLX:
864   case X86::VMOVUPSZ256rm_NOVLX:
865   case X86::VMOVUPSZrm: {
866     // Loads from constant pools are trivially rematerializable.
867     if (MI.getOperand(1 + X86::AddrBaseReg).isReg() &&
868         MI.getOperand(1 + X86::AddrScaleAmt).isImm() &&
869         MI.getOperand(1 + X86::AddrIndexReg).isReg() &&
870         MI.getOperand(1 + X86::AddrIndexReg).getReg() == 0 &&
871         MI.isDereferenceableInvariantLoad(AA)) {
872       Register BaseReg = MI.getOperand(1 + X86::AddrBaseReg).getReg();
873       if (BaseReg == 0 || BaseReg == X86::RIP)
874         return true;
875       // Allow re-materialization of PIC load.
876       if (!ReMatPICStubLoad && MI.getOperand(1 + X86::AddrDisp).isGlobal())
877         return false;
878       const MachineFunction &MF = *MI.getParent()->getParent();
879       const MachineRegisterInfo &MRI = MF.getRegInfo();
880       return regIsPICBase(BaseReg, MRI);
881     }
882     return false;
883   }
884 
885   case X86::LEA32r:
886   case X86::LEA64r: {
887     if (MI.getOperand(1 + X86::AddrScaleAmt).isImm() &&
888         MI.getOperand(1 + X86::AddrIndexReg).isReg() &&
889         MI.getOperand(1 + X86::AddrIndexReg).getReg() == 0 &&
890         !MI.getOperand(1 + X86::AddrDisp).isReg()) {
891       // lea fi#, lea GV, etc. are all rematerializable.
892       if (!MI.getOperand(1 + X86::AddrBaseReg).isReg())
893         return true;
894       Register BaseReg = MI.getOperand(1 + X86::AddrBaseReg).getReg();
895       if (BaseReg == 0)
896         return true;
897       // Allow re-materialization of lea PICBase + x.
898       const MachineFunction &MF = *MI.getParent()->getParent();
899       const MachineRegisterInfo &MRI = MF.getRegInfo();
900       return regIsPICBase(BaseReg, MRI);
901     }
902     return false;
903   }
904   }
905 }
906 
907 void X86InstrInfo::reMaterialize(MachineBasicBlock &MBB,
908                                  MachineBasicBlock::iterator I,
909                                  Register DestReg, unsigned SubIdx,
910                                  const MachineInstr &Orig,
911                                  const TargetRegisterInfo &TRI) const {
912   bool ClobbersEFLAGS = Orig.modifiesRegister(X86::EFLAGS, &TRI);
913   if (ClobbersEFLAGS && MBB.computeRegisterLiveness(&TRI, X86::EFLAGS, I) !=
914                             MachineBasicBlock::LQR_Dead) {
915     // The instruction clobbers EFLAGS. Re-materialize as MOV32ri to avoid side
916     // effects.
917     int Value;
918     switch (Orig.getOpcode()) {
919     case X86::MOV32r0:  Value = 0; break;
920     case X86::MOV32r1:  Value = 1; break;
921     case X86::MOV32r_1: Value = -1; break;
922     default:
923       llvm_unreachable("Unexpected instruction!");
924     }
925 
926     const DebugLoc &DL = Orig.getDebugLoc();
927     BuildMI(MBB, I, DL, get(X86::MOV32ri))
928         .add(Orig.getOperand(0))
929         .addImm(Value);
930   } else {
931     MachineInstr *MI = MBB.getParent()->CloneMachineInstr(&Orig);
932     MBB.insert(I, MI);
933   }
934 
935   MachineInstr &NewMI = *std::prev(I);
936   NewMI.substituteRegister(Orig.getOperand(0).getReg(), DestReg, SubIdx, TRI);
937 }
938 
939 /// True if MI has a condition code def, e.g. EFLAGS, that is not marked dead.
940 bool X86InstrInfo::hasLiveCondCodeDef(MachineInstr &MI) const {
941   for (const MachineOperand &MO : MI.operands()) {
942     if (MO.isReg() && MO.isDef() &&
943         MO.getReg() == X86::EFLAGS && !MO.isDead()) {
944       return true;
945     }
946   }
947   return false;
948 }
949 
950 /// Check whether the shift count for a machine operand is non-zero.
951 inline static unsigned getTruncatedShiftCount(const MachineInstr &MI,
952                                               unsigned ShiftAmtOperandIdx) {
953   // The shift count is six bits with the REX.W prefix and five bits without.
954   unsigned ShiftCountMask = (MI.getDesc().TSFlags & X86II::REX_W) ? 63 : 31;
955   unsigned Imm = MI.getOperand(ShiftAmtOperandIdx).getImm();
956   return Imm & ShiftCountMask;
957 }
958 
959 /// Check whether the given shift count is appropriate
960 /// can be represented by a LEA instruction.
961 inline static bool isTruncatedShiftCountForLEA(unsigned ShAmt) {
962   // Left shift instructions can be transformed into load-effective-address
963   // instructions if we can encode them appropriately.
964   // A LEA instruction utilizes a SIB byte to encode its scale factor.
965   // The SIB.scale field is two bits wide which means that we can encode any
966   // shift amount less than 4.
967   return ShAmt < 4 && ShAmt > 0;
968 }
969 
970 static bool findRedundantFlagInstr(MachineInstr &CmpInstr,
971                                    MachineInstr &CmpValDefInstr,
972                                    const MachineRegisterInfo *MRI,
973                                    MachineInstr **AndInstr,
974                                    const TargetRegisterInfo *TRI,
975                                    bool &NoSignFlag, bool &ClearsOverflowFlag) {
976   if (CmpValDefInstr.getOpcode() != X86::SUBREG_TO_REG)
977     return false;
978 
979   if (CmpInstr.getOpcode() != X86::TEST64rr)
980     return false;
981 
982   // CmpInstr is a TEST64rr instruction, and `X86InstrInfo::analyzeCompare`
983   // guarantees that it's analyzable only if two registers are identical.
984   assert(
985       (CmpInstr.getOperand(0).getReg() == CmpInstr.getOperand(1).getReg()) &&
986       "CmpInstr is an analyzable TEST64rr, and `X86InstrInfo::analyzeCompare` "
987       "requires two reg operands are the same.");
988 
989   // Caller (`X86InstrInfo::optimizeCompareInstr`) guarantees that
990   // `CmpValDefInstr` defines the value that's used by `CmpInstr`; in this case
991   // if `CmpValDefInstr` sets the EFLAGS, it is likely that `CmpInstr` is
992   // redundant.
993   assert(
994       (MRI->getVRegDef(CmpInstr.getOperand(0).getReg()) == &CmpValDefInstr) &&
995       "Caller guarantees that TEST64rr is a user of SUBREG_TO_REG.");
996 
997   // As seen in X86 td files, CmpValDefInstr.getOperand(1).getImm() is typically
998   // 0.
999   if (CmpValDefInstr.getOperand(1).getImm() != 0)
1000     return false;
1001 
1002   // As seen in X86 td files, CmpValDefInstr.getOperand(3) is typically
1003   // sub_32bit or sub_xmm.
1004   if (CmpValDefInstr.getOperand(3).getImm() != X86::sub_32bit)
1005     return false;
1006 
1007   MachineInstr *VregDefInstr =
1008       MRI->getVRegDef(CmpValDefInstr.getOperand(2).getReg());
1009 
1010   assert(VregDefInstr && "Must have a definition (SSA)");
1011 
1012   // Requires `CmpValDefInstr` and `VregDefInstr` are from the same MBB
1013   // to simplify the subsequent analysis.
1014   //
1015   // FIXME: If `VregDefInstr->getParent()` is the only predecessor of
1016   // `CmpValDefInstr.getParent()`, this could be handled.
1017   if (VregDefInstr->getParent() != CmpValDefInstr.getParent())
1018     return false;
1019 
1020   if (X86::isAND(VregDefInstr->getOpcode())) {
1021     // Get a sequence of instructions like
1022     //   %reg = and* ...                    // Set EFLAGS
1023     //   ...                                // EFLAGS not changed
1024     //   %extended_reg = subreg_to_reg 0, %reg, %subreg.sub_32bit
1025     //   test64rr %extended_reg, %extended_reg, implicit-def $eflags
1026     //
1027     // If subsequent readers use a subset of bits that don't change
1028     // after `and*` instructions, it's likely that the test64rr could
1029     // be optimized away.
1030     for (const MachineInstr &Instr :
1031          make_range(std::next(MachineBasicBlock::iterator(VregDefInstr)),
1032                     MachineBasicBlock::iterator(CmpValDefInstr))) {
1033       // There are instructions between 'VregDefInstr' and
1034       // 'CmpValDefInstr' that modifies EFLAGS.
1035       if (Instr.modifiesRegister(X86::EFLAGS, TRI))
1036         return false;
1037     }
1038 
1039     *AndInstr = VregDefInstr;
1040 
1041     // AND instruction will essentially update SF and clear OF, so
1042     // NoSignFlag should be false in the sense that SF is modified by `AND`.
1043     //
1044     // However, the implementation artifically sets `NoSignFlag` to true
1045     // to poison the SF bit; that is to say, if SF is looked at later, the
1046     // optimization (to erase TEST64rr) will be disabled.
1047     //
1048     // The reason to poison SF bit is that SF bit value could be different
1049     // in the `AND` and `TEST` operation; signed bit is not known for `AND`,
1050     // and is known to be 0 as a result of `TEST64rr`.
1051     //
1052     // FIXME: As opposed to poisoning the SF bit directly, consider peeking into
1053     // the AND instruction and using the static information to guide peephole
1054     // optimization if possible. For example, it's possible to fold a
1055     // conditional move into a copy if the relevant EFLAG bits could be deduced
1056     // from an immediate operand of and operation.
1057     //
1058     NoSignFlag = true;
1059     // ClearsOverflowFlag is true for AND operation (no surprise).
1060     ClearsOverflowFlag = true;
1061     return true;
1062   }
1063   return false;
1064 }
1065 
1066 bool X86InstrInfo::classifyLEAReg(MachineInstr &MI, const MachineOperand &Src,
1067                                   unsigned Opc, bool AllowSP, Register &NewSrc,
1068                                   bool &isKill, MachineOperand &ImplicitOp,
1069                                   LiveVariables *LV, LiveIntervals *LIS) const {
1070   MachineFunction &MF = *MI.getParent()->getParent();
1071   const TargetRegisterClass *RC;
1072   if (AllowSP) {
1073     RC = Opc != X86::LEA32r ? &X86::GR64RegClass : &X86::GR32RegClass;
1074   } else {
1075     RC = Opc != X86::LEA32r ?
1076       &X86::GR64_NOSPRegClass : &X86::GR32_NOSPRegClass;
1077   }
1078   Register SrcReg = Src.getReg();
1079   isKill = MI.killsRegister(SrcReg);
1080 
1081   // For both LEA64 and LEA32 the register already has essentially the right
1082   // type (32-bit or 64-bit) we may just need to forbid SP.
1083   if (Opc != X86::LEA64_32r) {
1084     NewSrc = SrcReg;
1085     assert(!Src.isUndef() && "Undef op doesn't need optimization");
1086 
1087     if (NewSrc.isVirtual() && !MF.getRegInfo().constrainRegClass(NewSrc, RC))
1088       return false;
1089 
1090     return true;
1091   }
1092 
1093   // This is for an LEA64_32r and incoming registers are 32-bit. One way or
1094   // another we need to add 64-bit registers to the final MI.
1095   if (SrcReg.isPhysical()) {
1096     ImplicitOp = Src;
1097     ImplicitOp.setImplicit();
1098 
1099     NewSrc = getX86SubSuperRegister(SrcReg, 64);
1100     assert(!Src.isUndef() && "Undef op doesn't need optimization");
1101   } else {
1102     // Virtual register of the wrong class, we have to create a temporary 64-bit
1103     // vreg to feed into the LEA.
1104     NewSrc = MF.getRegInfo().createVirtualRegister(RC);
1105     MachineInstr *Copy =
1106         BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), get(TargetOpcode::COPY))
1107             .addReg(NewSrc, RegState::Define | RegState::Undef, X86::sub_32bit)
1108             .addReg(SrcReg, getKillRegState(isKill));
1109 
1110     // Which is obviously going to be dead after we're done with it.
1111     isKill = true;
1112 
1113     if (LV)
1114       LV->replaceKillInstruction(SrcReg, MI, *Copy);
1115 
1116     if (LIS) {
1117       SlotIndex CopyIdx = LIS->InsertMachineInstrInMaps(*Copy);
1118       SlotIndex Idx = LIS->getInstructionIndex(MI);
1119       LiveInterval &LI = LIS->getInterval(SrcReg);
1120       LiveRange::Segment *S = LI.getSegmentContaining(Idx);
1121       if (S->end.getBaseIndex() == Idx)
1122         S->end = CopyIdx.getRegSlot();
1123     }
1124   }
1125 
1126   // We've set all the parameters without issue.
1127   return true;
1128 }
1129 
1130 MachineInstr *X86InstrInfo::convertToThreeAddressWithLEA(unsigned MIOpc,
1131                                                          MachineInstr &MI,
1132                                                          LiveVariables *LV,
1133                                                          LiveIntervals *LIS,
1134                                                          bool Is8BitOp) const {
1135   // We handle 8-bit adds and various 16-bit opcodes in the switch below.
1136   MachineBasicBlock &MBB = *MI.getParent();
1137   MachineRegisterInfo &RegInfo = MBB.getParent()->getRegInfo();
1138   assert((Is8BitOp || RegInfo.getTargetRegisterInfo()->getRegSizeInBits(
1139               *RegInfo.getRegClass(MI.getOperand(0).getReg())) == 16) &&
1140          "Unexpected type for LEA transform");
1141 
1142   // TODO: For a 32-bit target, we need to adjust the LEA variables with
1143   // something like this:
1144   //   Opcode = X86::LEA32r;
1145   //   InRegLEA = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
1146   //   OutRegLEA =
1147   //       Is8BitOp ? RegInfo.createVirtualRegister(&X86::GR32ABCD_RegClass)
1148   //                : RegInfo.createVirtualRegister(&X86::GR32RegClass);
1149   if (!Subtarget.is64Bit())
1150     return nullptr;
1151 
1152   unsigned Opcode = X86::LEA64_32r;
1153   Register InRegLEA = RegInfo.createVirtualRegister(&X86::GR64_NOSPRegClass);
1154   Register OutRegLEA = RegInfo.createVirtualRegister(&X86::GR32RegClass);
1155   Register InRegLEA2;
1156 
1157   // Build and insert into an implicit UNDEF value. This is OK because
1158   // we will be shifting and then extracting the lower 8/16-bits.
1159   // This has the potential to cause partial register stall. e.g.
1160   //   movw    (%rbp,%rcx,2), %dx
1161   //   leal    -65(%rdx), %esi
1162   // But testing has shown this *does* help performance in 64-bit mode (at
1163   // least on modern x86 machines).
1164   MachineBasicBlock::iterator MBBI = MI.getIterator();
1165   Register Dest = MI.getOperand(0).getReg();
1166   Register Src = MI.getOperand(1).getReg();
1167   Register Src2;
1168   bool IsDead = MI.getOperand(0).isDead();
1169   bool IsKill = MI.getOperand(1).isKill();
1170   unsigned SubReg = Is8BitOp ? X86::sub_8bit : X86::sub_16bit;
1171   assert(!MI.getOperand(1).isUndef() && "Undef op doesn't need optimization");
1172   MachineInstr *ImpDef =
1173       BuildMI(MBB, MBBI, MI.getDebugLoc(), get(X86::IMPLICIT_DEF), InRegLEA);
1174   MachineInstr *InsMI =
1175       BuildMI(MBB, MBBI, MI.getDebugLoc(), get(TargetOpcode::COPY))
1176           .addReg(InRegLEA, RegState::Define, SubReg)
1177           .addReg(Src, getKillRegState(IsKill));
1178   MachineInstr *ImpDef2 = nullptr;
1179   MachineInstr *InsMI2 = nullptr;
1180 
1181   MachineInstrBuilder MIB =
1182       BuildMI(MBB, MBBI, MI.getDebugLoc(), get(Opcode), OutRegLEA);
1183   switch (MIOpc) {
1184   default: llvm_unreachable("Unreachable!");
1185   case X86::SHL8ri:
1186   case X86::SHL16ri: {
1187     unsigned ShAmt = MI.getOperand(2).getImm();
1188     MIB.addReg(0).addImm(1ULL << ShAmt)
1189        .addReg(InRegLEA, RegState::Kill).addImm(0).addReg(0);
1190     break;
1191   }
1192   case X86::INC8r:
1193   case X86::INC16r:
1194     addRegOffset(MIB, InRegLEA, true, 1);
1195     break;
1196   case X86::DEC8r:
1197   case X86::DEC16r:
1198     addRegOffset(MIB, InRegLEA, true, -1);
1199     break;
1200   case X86::ADD8ri:
1201   case X86::ADD8ri_DB:
1202   case X86::ADD16ri:
1203   case X86::ADD16ri8:
1204   case X86::ADD16ri_DB:
1205   case X86::ADD16ri8_DB:
1206     addRegOffset(MIB, InRegLEA, true, MI.getOperand(2).getImm());
1207     break;
1208   case X86::ADD8rr:
1209   case X86::ADD8rr_DB:
1210   case X86::ADD16rr:
1211   case X86::ADD16rr_DB: {
1212     Src2 = MI.getOperand(2).getReg();
1213     bool IsKill2 = MI.getOperand(2).isKill();
1214     assert(!MI.getOperand(2).isUndef() && "Undef op doesn't need optimization");
1215     if (Src == Src2) {
1216       // ADD8rr/ADD16rr killed %reg1028, %reg1028
1217       // just a single insert_subreg.
1218       addRegReg(MIB, InRegLEA, true, InRegLEA, false);
1219     } else {
1220       if (Subtarget.is64Bit())
1221         InRegLEA2 = RegInfo.createVirtualRegister(&X86::GR64_NOSPRegClass);
1222       else
1223         InRegLEA2 = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
1224       // Build and insert into an implicit UNDEF value. This is OK because
1225       // we will be shifting and then extracting the lower 8/16-bits.
1226       ImpDef2 = BuildMI(MBB, &*MIB, MI.getDebugLoc(), get(X86::IMPLICIT_DEF),
1227                         InRegLEA2);
1228       InsMI2 = BuildMI(MBB, &*MIB, MI.getDebugLoc(), get(TargetOpcode::COPY))
1229                    .addReg(InRegLEA2, RegState::Define, SubReg)
1230                    .addReg(Src2, getKillRegState(IsKill2));
1231       addRegReg(MIB, InRegLEA, true, InRegLEA2, true);
1232     }
1233     if (LV && IsKill2 && InsMI2)
1234       LV->replaceKillInstruction(Src2, MI, *InsMI2);
1235     break;
1236   }
1237   }
1238 
1239   MachineInstr *NewMI = MIB;
1240   MachineInstr *ExtMI =
1241       BuildMI(MBB, MBBI, MI.getDebugLoc(), get(TargetOpcode::COPY))
1242           .addReg(Dest, RegState::Define | getDeadRegState(IsDead))
1243           .addReg(OutRegLEA, RegState::Kill, SubReg);
1244 
1245   if (LV) {
1246     // Update live variables.
1247     LV->getVarInfo(InRegLEA).Kills.push_back(NewMI);
1248     LV->getVarInfo(OutRegLEA).Kills.push_back(ExtMI);
1249     if (IsKill)
1250       LV->replaceKillInstruction(Src, MI, *InsMI);
1251     if (IsDead)
1252       LV->replaceKillInstruction(Dest, MI, *ExtMI);
1253   }
1254 
1255   if (LIS) {
1256     LIS->InsertMachineInstrInMaps(*ImpDef);
1257     SlotIndex InsIdx = LIS->InsertMachineInstrInMaps(*InsMI);
1258     if (ImpDef2)
1259       LIS->InsertMachineInstrInMaps(*ImpDef2);
1260     SlotIndex Ins2Idx;
1261     if (InsMI2)
1262       Ins2Idx = LIS->InsertMachineInstrInMaps(*InsMI2);
1263     SlotIndex NewIdx = LIS->ReplaceMachineInstrInMaps(MI, *NewMI);
1264     SlotIndex ExtIdx = LIS->InsertMachineInstrInMaps(*ExtMI);
1265     LIS->getInterval(InRegLEA);
1266     LIS->getInterval(OutRegLEA);
1267     if (InRegLEA2)
1268       LIS->getInterval(InRegLEA2);
1269 
1270     // Move the use of Src up to InsMI.
1271     LiveInterval &SrcLI = LIS->getInterval(Src);
1272     LiveRange::Segment *SrcSeg = SrcLI.getSegmentContaining(NewIdx);
1273     if (SrcSeg->end == NewIdx.getRegSlot())
1274       SrcSeg->end = InsIdx.getRegSlot();
1275 
1276     if (InsMI2) {
1277       // Move the use of Src2 up to InsMI2.
1278       LiveInterval &Src2LI = LIS->getInterval(Src2);
1279       LiveRange::Segment *Src2Seg = Src2LI.getSegmentContaining(NewIdx);
1280       if (Src2Seg->end == NewIdx.getRegSlot())
1281         Src2Seg->end = Ins2Idx.getRegSlot();
1282     }
1283 
1284     // Move the definition of Dest down to ExtMI.
1285     LiveInterval &DestLI = LIS->getInterval(Dest);
1286     LiveRange::Segment *DestSeg =
1287         DestLI.getSegmentContaining(NewIdx.getRegSlot());
1288     assert(DestSeg->start == NewIdx.getRegSlot() &&
1289            DestSeg->valno->def == NewIdx.getRegSlot());
1290     DestSeg->start = ExtIdx.getRegSlot();
1291     DestSeg->valno->def = ExtIdx.getRegSlot();
1292   }
1293 
1294   return ExtMI;
1295 }
1296 
1297 /// This method must be implemented by targets that
1298 /// set the M_CONVERTIBLE_TO_3_ADDR flag.  When this flag is set, the target
1299 /// may be able to convert a two-address instruction into a true
1300 /// three-address instruction on demand.  This allows the X86 target (for
1301 /// example) to convert ADD and SHL instructions into LEA instructions if they
1302 /// would require register copies due to two-addressness.
1303 ///
1304 /// This method returns a null pointer if the transformation cannot be
1305 /// performed, otherwise it returns the new instruction.
1306 ///
1307 MachineInstr *X86InstrInfo::convertToThreeAddress(MachineInstr &MI,
1308                                                   LiveVariables *LV,
1309                                                   LiveIntervals *LIS) const {
1310   // The following opcodes also sets the condition code register(s). Only
1311   // convert them to equivalent lea if the condition code register def's
1312   // are dead!
1313   if (hasLiveCondCodeDef(MI))
1314     return nullptr;
1315 
1316   MachineFunction &MF = *MI.getParent()->getParent();
1317   // All instructions input are two-addr instructions.  Get the known operands.
1318   const MachineOperand &Dest = MI.getOperand(0);
1319   const MachineOperand &Src = MI.getOperand(1);
1320 
1321   // Ideally, operations with undef should be folded before we get here, but we
1322   // can't guarantee it. Bail out because optimizing undefs is a waste of time.
1323   // Without this, we have to forward undef state to new register operands to
1324   // avoid machine verifier errors.
1325   if (Src.isUndef())
1326     return nullptr;
1327   if (MI.getNumOperands() > 2)
1328     if (MI.getOperand(2).isReg() && MI.getOperand(2).isUndef())
1329       return nullptr;
1330 
1331   MachineInstr *NewMI = nullptr;
1332   Register SrcReg, SrcReg2;
1333   bool Is64Bit = Subtarget.is64Bit();
1334 
1335   bool Is8BitOp = false;
1336   unsigned MIOpc = MI.getOpcode();
1337   switch (MIOpc) {
1338   default: llvm_unreachable("Unreachable!");
1339   case X86::SHL64ri: {
1340     assert(MI.getNumOperands() >= 3 && "Unknown shift instruction!");
1341     unsigned ShAmt = getTruncatedShiftCount(MI, 2);
1342     if (!isTruncatedShiftCountForLEA(ShAmt)) return nullptr;
1343 
1344     // LEA can't handle RSP.
1345     if (Src.getReg().isVirtual() && !MF.getRegInfo().constrainRegClass(
1346                                         Src.getReg(), &X86::GR64_NOSPRegClass))
1347       return nullptr;
1348 
1349     NewMI = BuildMI(MF, MI.getDebugLoc(), get(X86::LEA64r))
1350                 .add(Dest)
1351                 .addReg(0)
1352                 .addImm(1ULL << ShAmt)
1353                 .add(Src)
1354                 .addImm(0)
1355                 .addReg(0);
1356     break;
1357   }
1358   case X86::SHL32ri: {
1359     assert(MI.getNumOperands() >= 3 && "Unknown shift instruction!");
1360     unsigned ShAmt = getTruncatedShiftCount(MI, 2);
1361     if (!isTruncatedShiftCountForLEA(ShAmt)) return nullptr;
1362 
1363     unsigned Opc = Is64Bit ? X86::LEA64_32r : X86::LEA32r;
1364 
1365     // LEA can't handle ESP.
1366     bool isKill;
1367     MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
1368     if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/false, SrcReg, isKill,
1369                         ImplicitOp, LV, LIS))
1370       return nullptr;
1371 
1372     MachineInstrBuilder MIB =
1373         BuildMI(MF, MI.getDebugLoc(), get(Opc))
1374             .add(Dest)
1375             .addReg(0)
1376             .addImm(1ULL << ShAmt)
1377             .addReg(SrcReg, getKillRegState(isKill))
1378             .addImm(0)
1379             .addReg(0);
1380     if (ImplicitOp.getReg() != 0)
1381       MIB.add(ImplicitOp);
1382     NewMI = MIB;
1383 
1384     break;
1385   }
1386   case X86::SHL8ri:
1387     Is8BitOp = true;
1388     LLVM_FALLTHROUGH;
1389   case X86::SHL16ri: {
1390     assert(MI.getNumOperands() >= 3 && "Unknown shift instruction!");
1391     unsigned ShAmt = getTruncatedShiftCount(MI, 2);
1392     if (!isTruncatedShiftCountForLEA(ShAmt))
1393       return nullptr;
1394     return convertToThreeAddressWithLEA(MIOpc, MI, LV, LIS, Is8BitOp);
1395   }
1396   case X86::INC64r:
1397   case X86::INC32r: {
1398     assert(MI.getNumOperands() >= 2 && "Unknown inc instruction!");
1399     unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r :
1400         (Is64Bit ? X86::LEA64_32r : X86::LEA32r);
1401     bool isKill;
1402     MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
1403     if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/false, SrcReg, isKill,
1404                         ImplicitOp, LV, LIS))
1405       return nullptr;
1406 
1407     MachineInstrBuilder MIB =
1408         BuildMI(MF, MI.getDebugLoc(), get(Opc))
1409             .add(Dest)
1410             .addReg(SrcReg, getKillRegState(isKill));
1411     if (ImplicitOp.getReg() != 0)
1412       MIB.add(ImplicitOp);
1413 
1414     NewMI = addOffset(MIB, 1);
1415     break;
1416   }
1417   case X86::DEC64r:
1418   case X86::DEC32r: {
1419     assert(MI.getNumOperands() >= 2 && "Unknown dec instruction!");
1420     unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r
1421         : (Is64Bit ? X86::LEA64_32r : X86::LEA32r);
1422 
1423     bool isKill;
1424     MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
1425     if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/false, SrcReg, isKill,
1426                         ImplicitOp, LV, LIS))
1427       return nullptr;
1428 
1429     MachineInstrBuilder MIB = BuildMI(MF, MI.getDebugLoc(), get(Opc))
1430                                   .add(Dest)
1431                                   .addReg(SrcReg, getKillRegState(isKill));
1432     if (ImplicitOp.getReg() != 0)
1433       MIB.add(ImplicitOp);
1434 
1435     NewMI = addOffset(MIB, -1);
1436 
1437     break;
1438   }
1439   case X86::DEC8r:
1440   case X86::INC8r:
1441     Is8BitOp = true;
1442     LLVM_FALLTHROUGH;
1443   case X86::DEC16r:
1444   case X86::INC16r:
1445     return convertToThreeAddressWithLEA(MIOpc, MI, LV, LIS, Is8BitOp);
1446   case X86::ADD64rr:
1447   case X86::ADD64rr_DB:
1448   case X86::ADD32rr:
1449   case X86::ADD32rr_DB: {
1450     assert(MI.getNumOperands() >= 3 && "Unknown add instruction!");
1451     unsigned Opc;
1452     if (MIOpc == X86::ADD64rr || MIOpc == X86::ADD64rr_DB)
1453       Opc = X86::LEA64r;
1454     else
1455       Opc = Is64Bit ? X86::LEA64_32r : X86::LEA32r;
1456 
1457     const MachineOperand &Src2 = MI.getOperand(2);
1458     bool isKill2;
1459     MachineOperand ImplicitOp2 = MachineOperand::CreateReg(0, false);
1460     if (!classifyLEAReg(MI, Src2, Opc, /*AllowSP=*/false, SrcReg2, isKill2,
1461                         ImplicitOp2, LV, LIS))
1462       return nullptr;
1463 
1464     bool isKill;
1465     MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
1466     if (Src.getReg() == Src2.getReg()) {
1467       // Don't call classify LEAReg a second time on the same register, in case
1468       // the first call inserted a COPY from Src2 and marked it as killed.
1469       isKill = isKill2;
1470       SrcReg = SrcReg2;
1471     } else {
1472       if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/true, SrcReg, isKill,
1473                           ImplicitOp, LV, LIS))
1474         return nullptr;
1475     }
1476 
1477     MachineInstrBuilder MIB = BuildMI(MF, MI.getDebugLoc(), get(Opc)).add(Dest);
1478     if (ImplicitOp.getReg() != 0)
1479       MIB.add(ImplicitOp);
1480     if (ImplicitOp2.getReg() != 0)
1481       MIB.add(ImplicitOp2);
1482 
1483     NewMI = addRegReg(MIB, SrcReg, isKill, SrcReg2, isKill2);
1484     if (LV && Src2.isKill())
1485       LV->replaceKillInstruction(SrcReg2, MI, *NewMI);
1486     break;
1487   }
1488   case X86::ADD8rr:
1489   case X86::ADD8rr_DB:
1490     Is8BitOp = true;
1491     LLVM_FALLTHROUGH;
1492   case X86::ADD16rr:
1493   case X86::ADD16rr_DB:
1494     return convertToThreeAddressWithLEA(MIOpc, MI, LV, LIS, Is8BitOp);
1495   case X86::ADD64ri32:
1496   case X86::ADD64ri8:
1497   case X86::ADD64ri32_DB:
1498   case X86::ADD64ri8_DB:
1499     assert(MI.getNumOperands() >= 3 && "Unknown add instruction!");
1500     NewMI = addOffset(
1501         BuildMI(MF, MI.getDebugLoc(), get(X86::LEA64r)).add(Dest).add(Src),
1502         MI.getOperand(2));
1503     break;
1504   case X86::ADD32ri:
1505   case X86::ADD32ri8:
1506   case X86::ADD32ri_DB:
1507   case X86::ADD32ri8_DB: {
1508     assert(MI.getNumOperands() >= 3 && "Unknown add instruction!");
1509     unsigned Opc = Is64Bit ? X86::LEA64_32r : X86::LEA32r;
1510 
1511     bool isKill;
1512     MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
1513     if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/true, SrcReg, isKill,
1514                         ImplicitOp, LV, LIS))
1515       return nullptr;
1516 
1517     MachineInstrBuilder MIB = BuildMI(MF, MI.getDebugLoc(), get(Opc))
1518                                   .add(Dest)
1519                                   .addReg(SrcReg, getKillRegState(isKill));
1520     if (ImplicitOp.getReg() != 0)
1521       MIB.add(ImplicitOp);
1522 
1523     NewMI = addOffset(MIB, MI.getOperand(2));
1524     break;
1525   }
1526   case X86::ADD8ri:
1527   case X86::ADD8ri_DB:
1528     Is8BitOp = true;
1529     LLVM_FALLTHROUGH;
1530   case X86::ADD16ri:
1531   case X86::ADD16ri8:
1532   case X86::ADD16ri_DB:
1533   case X86::ADD16ri8_DB:
1534     return convertToThreeAddressWithLEA(MIOpc, MI, LV, LIS, Is8BitOp);
1535   case X86::SUB8ri:
1536   case X86::SUB16ri8:
1537   case X86::SUB16ri:
1538     /// FIXME: Support these similar to ADD8ri/ADD16ri*.
1539     return nullptr;
1540   case X86::SUB32ri8:
1541   case X86::SUB32ri: {
1542     if (!MI.getOperand(2).isImm())
1543       return nullptr;
1544     int64_t Imm = MI.getOperand(2).getImm();
1545     if (!isInt<32>(-Imm))
1546       return nullptr;
1547 
1548     assert(MI.getNumOperands() >= 3 && "Unknown add instruction!");
1549     unsigned Opc = Is64Bit ? X86::LEA64_32r : X86::LEA32r;
1550 
1551     bool isKill;
1552     MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
1553     if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/true, SrcReg, isKill,
1554                         ImplicitOp, LV, LIS))
1555       return nullptr;
1556 
1557     MachineInstrBuilder MIB = BuildMI(MF, MI.getDebugLoc(), get(Opc))
1558                                   .add(Dest)
1559                                   .addReg(SrcReg, getKillRegState(isKill));
1560     if (ImplicitOp.getReg() != 0)
1561       MIB.add(ImplicitOp);
1562 
1563     NewMI = addOffset(MIB, -Imm);
1564     break;
1565   }
1566 
1567   case X86::SUB64ri8:
1568   case X86::SUB64ri32: {
1569     if (!MI.getOperand(2).isImm())
1570       return nullptr;
1571     int64_t Imm = MI.getOperand(2).getImm();
1572     if (!isInt<32>(-Imm))
1573       return nullptr;
1574 
1575     assert(MI.getNumOperands() >= 3 && "Unknown sub instruction!");
1576 
1577     MachineInstrBuilder MIB = BuildMI(MF, MI.getDebugLoc(),
1578                                       get(X86::LEA64r)).add(Dest).add(Src);
1579     NewMI = addOffset(MIB, -Imm);
1580     break;
1581   }
1582 
1583   case X86::VMOVDQU8Z128rmk:
1584   case X86::VMOVDQU8Z256rmk:
1585   case X86::VMOVDQU8Zrmk:
1586   case X86::VMOVDQU16Z128rmk:
1587   case X86::VMOVDQU16Z256rmk:
1588   case X86::VMOVDQU16Zrmk:
1589   case X86::VMOVDQU32Z128rmk: case X86::VMOVDQA32Z128rmk:
1590   case X86::VMOVDQU32Z256rmk: case X86::VMOVDQA32Z256rmk:
1591   case X86::VMOVDQU32Zrmk:    case X86::VMOVDQA32Zrmk:
1592   case X86::VMOVDQU64Z128rmk: case X86::VMOVDQA64Z128rmk:
1593   case X86::VMOVDQU64Z256rmk: case X86::VMOVDQA64Z256rmk:
1594   case X86::VMOVDQU64Zrmk:    case X86::VMOVDQA64Zrmk:
1595   case X86::VMOVUPDZ128rmk:   case X86::VMOVAPDZ128rmk:
1596   case X86::VMOVUPDZ256rmk:   case X86::VMOVAPDZ256rmk:
1597   case X86::VMOVUPDZrmk:      case X86::VMOVAPDZrmk:
1598   case X86::VMOVUPSZ128rmk:   case X86::VMOVAPSZ128rmk:
1599   case X86::VMOVUPSZ256rmk:   case X86::VMOVAPSZ256rmk:
1600   case X86::VMOVUPSZrmk:      case X86::VMOVAPSZrmk:
1601   case X86::VBROADCASTSDZ256rmk:
1602   case X86::VBROADCASTSDZrmk:
1603   case X86::VBROADCASTSSZ128rmk:
1604   case X86::VBROADCASTSSZ256rmk:
1605   case X86::VBROADCASTSSZrmk:
1606   case X86::VPBROADCASTDZ128rmk:
1607   case X86::VPBROADCASTDZ256rmk:
1608   case X86::VPBROADCASTDZrmk:
1609   case X86::VPBROADCASTQZ128rmk:
1610   case X86::VPBROADCASTQZ256rmk:
1611   case X86::VPBROADCASTQZrmk: {
1612     unsigned Opc;
1613     switch (MIOpc) {
1614     default: llvm_unreachable("Unreachable!");
1615     case X86::VMOVDQU8Z128rmk:     Opc = X86::VPBLENDMBZ128rmk; break;
1616     case X86::VMOVDQU8Z256rmk:     Opc = X86::VPBLENDMBZ256rmk; break;
1617     case X86::VMOVDQU8Zrmk:        Opc = X86::VPBLENDMBZrmk;    break;
1618     case X86::VMOVDQU16Z128rmk:    Opc = X86::VPBLENDMWZ128rmk; break;
1619     case X86::VMOVDQU16Z256rmk:    Opc = X86::VPBLENDMWZ256rmk; break;
1620     case X86::VMOVDQU16Zrmk:       Opc = X86::VPBLENDMWZrmk;    break;
1621     case X86::VMOVDQU32Z128rmk:    Opc = X86::VPBLENDMDZ128rmk; break;
1622     case X86::VMOVDQU32Z256rmk:    Opc = X86::VPBLENDMDZ256rmk; break;
1623     case X86::VMOVDQU32Zrmk:       Opc = X86::VPBLENDMDZrmk;    break;
1624     case X86::VMOVDQU64Z128rmk:    Opc = X86::VPBLENDMQZ128rmk; break;
1625     case X86::VMOVDQU64Z256rmk:    Opc = X86::VPBLENDMQZ256rmk; break;
1626     case X86::VMOVDQU64Zrmk:       Opc = X86::VPBLENDMQZrmk;    break;
1627     case X86::VMOVUPDZ128rmk:      Opc = X86::VBLENDMPDZ128rmk; break;
1628     case X86::VMOVUPDZ256rmk:      Opc = X86::VBLENDMPDZ256rmk; break;
1629     case X86::VMOVUPDZrmk:         Opc = X86::VBLENDMPDZrmk;    break;
1630     case X86::VMOVUPSZ128rmk:      Opc = X86::VBLENDMPSZ128rmk; break;
1631     case X86::VMOVUPSZ256rmk:      Opc = X86::VBLENDMPSZ256rmk; break;
1632     case X86::VMOVUPSZrmk:         Opc = X86::VBLENDMPSZrmk;    break;
1633     case X86::VMOVDQA32Z128rmk:    Opc = X86::VPBLENDMDZ128rmk; break;
1634     case X86::VMOVDQA32Z256rmk:    Opc = X86::VPBLENDMDZ256rmk; break;
1635     case X86::VMOVDQA32Zrmk:       Opc = X86::VPBLENDMDZrmk;    break;
1636     case X86::VMOVDQA64Z128rmk:    Opc = X86::VPBLENDMQZ128rmk; break;
1637     case X86::VMOVDQA64Z256rmk:    Opc = X86::VPBLENDMQZ256rmk; break;
1638     case X86::VMOVDQA64Zrmk:       Opc = X86::VPBLENDMQZrmk;    break;
1639     case X86::VMOVAPDZ128rmk:      Opc = X86::VBLENDMPDZ128rmk; break;
1640     case X86::VMOVAPDZ256rmk:      Opc = X86::VBLENDMPDZ256rmk; break;
1641     case X86::VMOVAPDZrmk:         Opc = X86::VBLENDMPDZrmk;    break;
1642     case X86::VMOVAPSZ128rmk:      Opc = X86::VBLENDMPSZ128rmk; break;
1643     case X86::VMOVAPSZ256rmk:      Opc = X86::VBLENDMPSZ256rmk; break;
1644     case X86::VMOVAPSZrmk:         Opc = X86::VBLENDMPSZrmk;    break;
1645     case X86::VBROADCASTSDZ256rmk: Opc = X86::VBLENDMPDZ256rmbk; break;
1646     case X86::VBROADCASTSDZrmk:    Opc = X86::VBLENDMPDZrmbk;    break;
1647     case X86::VBROADCASTSSZ128rmk: Opc = X86::VBLENDMPSZ128rmbk; break;
1648     case X86::VBROADCASTSSZ256rmk: Opc = X86::VBLENDMPSZ256rmbk; break;
1649     case X86::VBROADCASTSSZrmk:    Opc = X86::VBLENDMPSZrmbk;    break;
1650     case X86::VPBROADCASTDZ128rmk: Opc = X86::VPBLENDMDZ128rmbk; break;
1651     case X86::VPBROADCASTDZ256rmk: Opc = X86::VPBLENDMDZ256rmbk; break;
1652     case X86::VPBROADCASTDZrmk:    Opc = X86::VPBLENDMDZrmbk;    break;
1653     case X86::VPBROADCASTQZ128rmk: Opc = X86::VPBLENDMQZ128rmbk; break;
1654     case X86::VPBROADCASTQZ256rmk: Opc = X86::VPBLENDMQZ256rmbk; break;
1655     case X86::VPBROADCASTQZrmk:    Opc = X86::VPBLENDMQZrmbk;    break;
1656     }
1657 
1658     NewMI = BuildMI(MF, MI.getDebugLoc(), get(Opc))
1659               .add(Dest)
1660               .add(MI.getOperand(2))
1661               .add(Src)
1662               .add(MI.getOperand(3))
1663               .add(MI.getOperand(4))
1664               .add(MI.getOperand(5))
1665               .add(MI.getOperand(6))
1666               .add(MI.getOperand(7));
1667     break;
1668   }
1669 
1670   case X86::VMOVDQU8Z128rrk:
1671   case X86::VMOVDQU8Z256rrk:
1672   case X86::VMOVDQU8Zrrk:
1673   case X86::VMOVDQU16Z128rrk:
1674   case X86::VMOVDQU16Z256rrk:
1675   case X86::VMOVDQU16Zrrk:
1676   case X86::VMOVDQU32Z128rrk: case X86::VMOVDQA32Z128rrk:
1677   case X86::VMOVDQU32Z256rrk: case X86::VMOVDQA32Z256rrk:
1678   case X86::VMOVDQU32Zrrk:    case X86::VMOVDQA32Zrrk:
1679   case X86::VMOVDQU64Z128rrk: case X86::VMOVDQA64Z128rrk:
1680   case X86::VMOVDQU64Z256rrk: case X86::VMOVDQA64Z256rrk:
1681   case X86::VMOVDQU64Zrrk:    case X86::VMOVDQA64Zrrk:
1682   case X86::VMOVUPDZ128rrk:   case X86::VMOVAPDZ128rrk:
1683   case X86::VMOVUPDZ256rrk:   case X86::VMOVAPDZ256rrk:
1684   case X86::VMOVUPDZrrk:      case X86::VMOVAPDZrrk:
1685   case X86::VMOVUPSZ128rrk:   case X86::VMOVAPSZ128rrk:
1686   case X86::VMOVUPSZ256rrk:   case X86::VMOVAPSZ256rrk:
1687   case X86::VMOVUPSZrrk:      case X86::VMOVAPSZrrk: {
1688     unsigned Opc;
1689     switch (MIOpc) {
1690     default: llvm_unreachable("Unreachable!");
1691     case X86::VMOVDQU8Z128rrk:  Opc = X86::VPBLENDMBZ128rrk; break;
1692     case X86::VMOVDQU8Z256rrk:  Opc = X86::VPBLENDMBZ256rrk; break;
1693     case X86::VMOVDQU8Zrrk:     Opc = X86::VPBLENDMBZrrk;    break;
1694     case X86::VMOVDQU16Z128rrk: Opc = X86::VPBLENDMWZ128rrk; break;
1695     case X86::VMOVDQU16Z256rrk: Opc = X86::VPBLENDMWZ256rrk; break;
1696     case X86::VMOVDQU16Zrrk:    Opc = X86::VPBLENDMWZrrk;    break;
1697     case X86::VMOVDQU32Z128rrk: Opc = X86::VPBLENDMDZ128rrk; break;
1698     case X86::VMOVDQU32Z256rrk: Opc = X86::VPBLENDMDZ256rrk; break;
1699     case X86::VMOVDQU32Zrrk:    Opc = X86::VPBLENDMDZrrk;    break;
1700     case X86::VMOVDQU64Z128rrk: Opc = X86::VPBLENDMQZ128rrk; break;
1701     case X86::VMOVDQU64Z256rrk: Opc = X86::VPBLENDMQZ256rrk; break;
1702     case X86::VMOVDQU64Zrrk:    Opc = X86::VPBLENDMQZrrk;    break;
1703     case X86::VMOVUPDZ128rrk:   Opc = X86::VBLENDMPDZ128rrk; break;
1704     case X86::VMOVUPDZ256rrk:   Opc = X86::VBLENDMPDZ256rrk; break;
1705     case X86::VMOVUPDZrrk:      Opc = X86::VBLENDMPDZrrk;    break;
1706     case X86::VMOVUPSZ128rrk:   Opc = X86::VBLENDMPSZ128rrk; break;
1707     case X86::VMOVUPSZ256rrk:   Opc = X86::VBLENDMPSZ256rrk; break;
1708     case X86::VMOVUPSZrrk:      Opc = X86::VBLENDMPSZrrk;    break;
1709     case X86::VMOVDQA32Z128rrk: Opc = X86::VPBLENDMDZ128rrk; break;
1710     case X86::VMOVDQA32Z256rrk: Opc = X86::VPBLENDMDZ256rrk; break;
1711     case X86::VMOVDQA32Zrrk:    Opc = X86::VPBLENDMDZrrk;    break;
1712     case X86::VMOVDQA64Z128rrk: Opc = X86::VPBLENDMQZ128rrk; break;
1713     case X86::VMOVDQA64Z256rrk: Opc = X86::VPBLENDMQZ256rrk; break;
1714     case X86::VMOVDQA64Zrrk:    Opc = X86::VPBLENDMQZrrk;    break;
1715     case X86::VMOVAPDZ128rrk:   Opc = X86::VBLENDMPDZ128rrk; break;
1716     case X86::VMOVAPDZ256rrk:   Opc = X86::VBLENDMPDZ256rrk; break;
1717     case X86::VMOVAPDZrrk:      Opc = X86::VBLENDMPDZrrk;    break;
1718     case X86::VMOVAPSZ128rrk:   Opc = X86::VBLENDMPSZ128rrk; break;
1719     case X86::VMOVAPSZ256rrk:   Opc = X86::VBLENDMPSZ256rrk; break;
1720     case X86::VMOVAPSZrrk:      Opc = X86::VBLENDMPSZrrk;    break;
1721     }
1722 
1723     NewMI = BuildMI(MF, MI.getDebugLoc(), get(Opc))
1724               .add(Dest)
1725               .add(MI.getOperand(2))
1726               .add(Src)
1727               .add(MI.getOperand(3));
1728     break;
1729   }
1730   }
1731 
1732   if (!NewMI) return nullptr;
1733 
1734   if (LV) {  // Update live variables
1735     if (Src.isKill())
1736       LV->replaceKillInstruction(Src.getReg(), MI, *NewMI);
1737     if (Dest.isDead())
1738       LV->replaceKillInstruction(Dest.getReg(), MI, *NewMI);
1739   }
1740 
1741   MachineBasicBlock &MBB = *MI.getParent();
1742   MBB.insert(MI.getIterator(), NewMI); // Insert the new inst
1743 
1744   if (LIS) {
1745     LIS->ReplaceMachineInstrInMaps(MI, *NewMI);
1746     if (SrcReg)
1747       LIS->getInterval(SrcReg);
1748     if (SrcReg2)
1749       LIS->getInterval(SrcReg2);
1750   }
1751 
1752   return NewMI;
1753 }
1754 
1755 /// This determines which of three possible cases of a three source commute
1756 /// the source indexes correspond to taking into account any mask operands.
1757 /// All prevents commuting a passthru operand. Returns -1 if the commute isn't
1758 /// possible.
1759 /// Case 0 - Possible to commute the first and second operands.
1760 /// Case 1 - Possible to commute the first and third operands.
1761 /// Case 2 - Possible to commute the second and third operands.
1762 static unsigned getThreeSrcCommuteCase(uint64_t TSFlags, unsigned SrcOpIdx1,
1763                                        unsigned SrcOpIdx2) {
1764   // Put the lowest index to SrcOpIdx1 to simplify the checks below.
1765   if (SrcOpIdx1 > SrcOpIdx2)
1766     std::swap(SrcOpIdx1, SrcOpIdx2);
1767 
1768   unsigned Op1 = 1, Op2 = 2, Op3 = 3;
1769   if (X86II::isKMasked(TSFlags)) {
1770     Op2++;
1771     Op3++;
1772   }
1773 
1774   if (SrcOpIdx1 == Op1 && SrcOpIdx2 == Op2)
1775     return 0;
1776   if (SrcOpIdx1 == Op1 && SrcOpIdx2 == Op3)
1777     return 1;
1778   if (SrcOpIdx1 == Op2 && SrcOpIdx2 == Op3)
1779     return 2;
1780   llvm_unreachable("Unknown three src commute case.");
1781 }
1782 
1783 unsigned X86InstrInfo::getFMA3OpcodeToCommuteOperands(
1784     const MachineInstr &MI, unsigned SrcOpIdx1, unsigned SrcOpIdx2,
1785     const X86InstrFMA3Group &FMA3Group) const {
1786 
1787   unsigned Opc = MI.getOpcode();
1788 
1789   // TODO: Commuting the 1st operand of FMA*_Int requires some additional
1790   // analysis. The commute optimization is legal only if all users of FMA*_Int
1791   // use only the lowest element of the FMA*_Int instruction. Such analysis are
1792   // not implemented yet. So, just return 0 in that case.
1793   // When such analysis are available this place will be the right place for
1794   // calling it.
1795   assert(!(FMA3Group.isIntrinsic() && (SrcOpIdx1 == 1 || SrcOpIdx2 == 1)) &&
1796          "Intrinsic instructions can't commute operand 1");
1797 
1798   // Determine which case this commute is or if it can't be done.
1799   unsigned Case = getThreeSrcCommuteCase(MI.getDesc().TSFlags, SrcOpIdx1,
1800                                          SrcOpIdx2);
1801   assert(Case < 3 && "Unexpected case number!");
1802 
1803   // Define the FMA forms mapping array that helps to map input FMA form
1804   // to output FMA form to preserve the operation semantics after
1805   // commuting the operands.
1806   const unsigned Form132Index = 0;
1807   const unsigned Form213Index = 1;
1808   const unsigned Form231Index = 2;
1809   static const unsigned FormMapping[][3] = {
1810     // 0: SrcOpIdx1 == 1 && SrcOpIdx2 == 2;
1811     // FMA132 A, C, b; ==> FMA231 C, A, b;
1812     // FMA213 B, A, c; ==> FMA213 A, B, c;
1813     // FMA231 C, A, b; ==> FMA132 A, C, b;
1814     { Form231Index, Form213Index, Form132Index },
1815     // 1: SrcOpIdx1 == 1 && SrcOpIdx2 == 3;
1816     // FMA132 A, c, B; ==> FMA132 B, c, A;
1817     // FMA213 B, a, C; ==> FMA231 C, a, B;
1818     // FMA231 C, a, B; ==> FMA213 B, a, C;
1819     { Form132Index, Form231Index, Form213Index },
1820     // 2: SrcOpIdx1 == 2 && SrcOpIdx2 == 3;
1821     // FMA132 a, C, B; ==> FMA213 a, B, C;
1822     // FMA213 b, A, C; ==> FMA132 b, C, A;
1823     // FMA231 c, A, B; ==> FMA231 c, B, A;
1824     { Form213Index, Form132Index, Form231Index }
1825   };
1826 
1827   unsigned FMAForms[3];
1828   FMAForms[0] = FMA3Group.get132Opcode();
1829   FMAForms[1] = FMA3Group.get213Opcode();
1830   FMAForms[2] = FMA3Group.get231Opcode();
1831 
1832   // Everything is ready, just adjust the FMA opcode and return it.
1833   for (unsigned FormIndex = 0; FormIndex < 3; FormIndex++)
1834     if (Opc == FMAForms[FormIndex])
1835       return FMAForms[FormMapping[Case][FormIndex]];
1836 
1837   llvm_unreachable("Illegal FMA3 format");
1838 }
1839 
1840 static void commuteVPTERNLOG(MachineInstr &MI, unsigned SrcOpIdx1,
1841                              unsigned SrcOpIdx2) {
1842   // Determine which case this commute is or if it can't be done.
1843   unsigned Case = getThreeSrcCommuteCase(MI.getDesc().TSFlags, SrcOpIdx1,
1844                                          SrcOpIdx2);
1845   assert(Case < 3 && "Unexpected case value!");
1846 
1847   // For each case we need to swap two pairs of bits in the final immediate.
1848   static const uint8_t SwapMasks[3][4] = {
1849     { 0x04, 0x10, 0x08, 0x20 }, // Swap bits 2/4 and 3/5.
1850     { 0x02, 0x10, 0x08, 0x40 }, // Swap bits 1/4 and 3/6.
1851     { 0x02, 0x04, 0x20, 0x40 }, // Swap bits 1/2 and 5/6.
1852   };
1853 
1854   uint8_t Imm = MI.getOperand(MI.getNumOperands()-1).getImm();
1855   // Clear out the bits we are swapping.
1856   uint8_t NewImm = Imm & ~(SwapMasks[Case][0] | SwapMasks[Case][1] |
1857                            SwapMasks[Case][2] | SwapMasks[Case][3]);
1858   // If the immediate had a bit of the pair set, then set the opposite bit.
1859   if (Imm & SwapMasks[Case][0]) NewImm |= SwapMasks[Case][1];
1860   if (Imm & SwapMasks[Case][1]) NewImm |= SwapMasks[Case][0];
1861   if (Imm & SwapMasks[Case][2]) NewImm |= SwapMasks[Case][3];
1862   if (Imm & SwapMasks[Case][3]) NewImm |= SwapMasks[Case][2];
1863   MI.getOperand(MI.getNumOperands()-1).setImm(NewImm);
1864 }
1865 
1866 // Returns true if this is a VPERMI2 or VPERMT2 instruction that can be
1867 // commuted.
1868 static bool isCommutableVPERMV3Instruction(unsigned Opcode) {
1869 #define VPERM_CASES(Suffix) \
1870   case X86::VPERMI2##Suffix##128rr:    case X86::VPERMT2##Suffix##128rr:    \
1871   case X86::VPERMI2##Suffix##256rr:    case X86::VPERMT2##Suffix##256rr:    \
1872   case X86::VPERMI2##Suffix##rr:       case X86::VPERMT2##Suffix##rr:       \
1873   case X86::VPERMI2##Suffix##128rm:    case X86::VPERMT2##Suffix##128rm:    \
1874   case X86::VPERMI2##Suffix##256rm:    case X86::VPERMT2##Suffix##256rm:    \
1875   case X86::VPERMI2##Suffix##rm:       case X86::VPERMT2##Suffix##rm:       \
1876   case X86::VPERMI2##Suffix##128rrkz:  case X86::VPERMT2##Suffix##128rrkz:  \
1877   case X86::VPERMI2##Suffix##256rrkz:  case X86::VPERMT2##Suffix##256rrkz:  \
1878   case X86::VPERMI2##Suffix##rrkz:     case X86::VPERMT2##Suffix##rrkz:     \
1879   case X86::VPERMI2##Suffix##128rmkz:  case X86::VPERMT2##Suffix##128rmkz:  \
1880   case X86::VPERMI2##Suffix##256rmkz:  case X86::VPERMT2##Suffix##256rmkz:  \
1881   case X86::VPERMI2##Suffix##rmkz:     case X86::VPERMT2##Suffix##rmkz:
1882 
1883 #define VPERM_CASES_BROADCAST(Suffix) \
1884   VPERM_CASES(Suffix) \
1885   case X86::VPERMI2##Suffix##128rmb:   case X86::VPERMT2##Suffix##128rmb:   \
1886   case X86::VPERMI2##Suffix##256rmb:   case X86::VPERMT2##Suffix##256rmb:   \
1887   case X86::VPERMI2##Suffix##rmb:      case X86::VPERMT2##Suffix##rmb:      \
1888   case X86::VPERMI2##Suffix##128rmbkz: case X86::VPERMT2##Suffix##128rmbkz: \
1889   case X86::VPERMI2##Suffix##256rmbkz: case X86::VPERMT2##Suffix##256rmbkz: \
1890   case X86::VPERMI2##Suffix##rmbkz:    case X86::VPERMT2##Suffix##rmbkz:
1891 
1892   switch (Opcode) {
1893   default: return false;
1894   VPERM_CASES(B)
1895   VPERM_CASES_BROADCAST(D)
1896   VPERM_CASES_BROADCAST(PD)
1897   VPERM_CASES_BROADCAST(PS)
1898   VPERM_CASES_BROADCAST(Q)
1899   VPERM_CASES(W)
1900     return true;
1901   }
1902 #undef VPERM_CASES_BROADCAST
1903 #undef VPERM_CASES
1904 }
1905 
1906 // Returns commuted opcode for VPERMI2 and VPERMT2 instructions by switching
1907 // from the I opcode to the T opcode and vice versa.
1908 static unsigned getCommutedVPERMV3Opcode(unsigned Opcode) {
1909 #define VPERM_CASES(Orig, New) \
1910   case X86::Orig##128rr:    return X86::New##128rr;   \
1911   case X86::Orig##128rrkz:  return X86::New##128rrkz; \
1912   case X86::Orig##128rm:    return X86::New##128rm;   \
1913   case X86::Orig##128rmkz:  return X86::New##128rmkz; \
1914   case X86::Orig##256rr:    return X86::New##256rr;   \
1915   case X86::Orig##256rrkz:  return X86::New##256rrkz; \
1916   case X86::Orig##256rm:    return X86::New##256rm;   \
1917   case X86::Orig##256rmkz:  return X86::New##256rmkz; \
1918   case X86::Orig##rr:       return X86::New##rr;      \
1919   case X86::Orig##rrkz:     return X86::New##rrkz;    \
1920   case X86::Orig##rm:       return X86::New##rm;      \
1921   case X86::Orig##rmkz:     return X86::New##rmkz;
1922 
1923 #define VPERM_CASES_BROADCAST(Orig, New) \
1924   VPERM_CASES(Orig, New) \
1925   case X86::Orig##128rmb:   return X86::New##128rmb;   \
1926   case X86::Orig##128rmbkz: return X86::New##128rmbkz; \
1927   case X86::Orig##256rmb:   return X86::New##256rmb;   \
1928   case X86::Orig##256rmbkz: return X86::New##256rmbkz; \
1929   case X86::Orig##rmb:      return X86::New##rmb;      \
1930   case X86::Orig##rmbkz:    return X86::New##rmbkz;
1931 
1932   switch (Opcode) {
1933   VPERM_CASES(VPERMI2B, VPERMT2B)
1934   VPERM_CASES_BROADCAST(VPERMI2D,  VPERMT2D)
1935   VPERM_CASES_BROADCAST(VPERMI2PD, VPERMT2PD)
1936   VPERM_CASES_BROADCAST(VPERMI2PS, VPERMT2PS)
1937   VPERM_CASES_BROADCAST(VPERMI2Q,  VPERMT2Q)
1938   VPERM_CASES(VPERMI2W, VPERMT2W)
1939   VPERM_CASES(VPERMT2B, VPERMI2B)
1940   VPERM_CASES_BROADCAST(VPERMT2D,  VPERMI2D)
1941   VPERM_CASES_BROADCAST(VPERMT2PD, VPERMI2PD)
1942   VPERM_CASES_BROADCAST(VPERMT2PS, VPERMI2PS)
1943   VPERM_CASES_BROADCAST(VPERMT2Q,  VPERMI2Q)
1944   VPERM_CASES(VPERMT2W, VPERMI2W)
1945   }
1946 
1947   llvm_unreachable("Unreachable!");
1948 #undef VPERM_CASES_BROADCAST
1949 #undef VPERM_CASES
1950 }
1951 
1952 MachineInstr *X86InstrInfo::commuteInstructionImpl(MachineInstr &MI, bool NewMI,
1953                                                    unsigned OpIdx1,
1954                                                    unsigned OpIdx2) const {
1955   auto cloneIfNew = [NewMI](MachineInstr &MI) -> MachineInstr & {
1956     if (NewMI)
1957       return *MI.getParent()->getParent()->CloneMachineInstr(&MI);
1958     return MI;
1959   };
1960 
1961   switch (MI.getOpcode()) {
1962   case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I)
1963   case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I)
1964   case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I)
1965   case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I)
1966   case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I)
1967   case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I)
1968     unsigned Opc;
1969     unsigned Size;
1970     switch (MI.getOpcode()) {
1971     default: llvm_unreachable("Unreachable!");
1972     case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break;
1973     case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break;
1974     case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break;
1975     case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break;
1976     case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break;
1977     case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break;
1978     }
1979     unsigned Amt = MI.getOperand(3).getImm();
1980     auto &WorkingMI = cloneIfNew(MI);
1981     WorkingMI.setDesc(get(Opc));
1982     WorkingMI.getOperand(3).setImm(Size - Amt);
1983     return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1984                                                    OpIdx1, OpIdx2);
1985   }
1986   case X86::PFSUBrr:
1987   case X86::PFSUBRrr: {
1988     // PFSUB  x, y: x = x - y
1989     // PFSUBR x, y: x = y - x
1990     unsigned Opc =
1991         (X86::PFSUBRrr == MI.getOpcode() ? X86::PFSUBrr : X86::PFSUBRrr);
1992     auto &WorkingMI = cloneIfNew(MI);
1993     WorkingMI.setDesc(get(Opc));
1994     return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1995                                                    OpIdx1, OpIdx2);
1996   }
1997   case X86::BLENDPDrri:
1998   case X86::BLENDPSrri:
1999   case X86::VBLENDPDrri:
2000   case X86::VBLENDPSrri:
2001     // If we're optimizing for size, try to use MOVSD/MOVSS.
2002     if (MI.getParent()->getParent()->getFunction().hasOptSize()) {
2003       unsigned Mask, Opc;
2004       switch (MI.getOpcode()) {
2005       default: llvm_unreachable("Unreachable!");
2006       case X86::BLENDPDrri:  Opc = X86::MOVSDrr;  Mask = 0x03; break;
2007       case X86::BLENDPSrri:  Opc = X86::MOVSSrr;  Mask = 0x0F; break;
2008       case X86::VBLENDPDrri: Opc = X86::VMOVSDrr; Mask = 0x03; break;
2009       case X86::VBLENDPSrri: Opc = X86::VMOVSSrr; Mask = 0x0F; break;
2010       }
2011       if ((MI.getOperand(3).getImm() ^ Mask) == 1) {
2012         auto &WorkingMI = cloneIfNew(MI);
2013         WorkingMI.setDesc(get(Opc));
2014         WorkingMI.removeOperand(3);
2015         return TargetInstrInfo::commuteInstructionImpl(WorkingMI,
2016                                                        /*NewMI=*/false,
2017                                                        OpIdx1, OpIdx2);
2018       }
2019     }
2020     LLVM_FALLTHROUGH;
2021   case X86::PBLENDWrri:
2022   case X86::VBLENDPDYrri:
2023   case X86::VBLENDPSYrri:
2024   case X86::VPBLENDDrri:
2025   case X86::VPBLENDWrri:
2026   case X86::VPBLENDDYrri:
2027   case X86::VPBLENDWYrri:{
2028     int8_t Mask;
2029     switch (MI.getOpcode()) {
2030     default: llvm_unreachable("Unreachable!");
2031     case X86::BLENDPDrri:    Mask = (int8_t)0x03; break;
2032     case X86::BLENDPSrri:    Mask = (int8_t)0x0F; break;
2033     case X86::PBLENDWrri:    Mask = (int8_t)0xFF; break;
2034     case X86::VBLENDPDrri:   Mask = (int8_t)0x03; break;
2035     case X86::VBLENDPSrri:   Mask = (int8_t)0x0F; break;
2036     case X86::VBLENDPDYrri:  Mask = (int8_t)0x0F; break;
2037     case X86::VBLENDPSYrri:  Mask = (int8_t)0xFF; break;
2038     case X86::VPBLENDDrri:   Mask = (int8_t)0x0F; break;
2039     case X86::VPBLENDWrri:   Mask = (int8_t)0xFF; break;
2040     case X86::VPBLENDDYrri:  Mask = (int8_t)0xFF; break;
2041     case X86::VPBLENDWYrri:  Mask = (int8_t)0xFF; break;
2042     }
2043     // Only the least significant bits of Imm are used.
2044     // Using int8_t to ensure it will be sign extended to the int64_t that
2045     // setImm takes in order to match isel behavior.
2046     int8_t Imm = MI.getOperand(3).getImm() & Mask;
2047     auto &WorkingMI = cloneIfNew(MI);
2048     WorkingMI.getOperand(3).setImm(Mask ^ Imm);
2049     return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
2050                                                    OpIdx1, OpIdx2);
2051   }
2052   case X86::INSERTPSrr:
2053   case X86::VINSERTPSrr:
2054   case X86::VINSERTPSZrr: {
2055     unsigned Imm = MI.getOperand(MI.getNumOperands() - 1).getImm();
2056     unsigned ZMask = Imm & 15;
2057     unsigned DstIdx = (Imm >> 4) & 3;
2058     unsigned SrcIdx = (Imm >> 6) & 3;
2059 
2060     // We can commute insertps if we zero 2 of the elements, the insertion is
2061     // "inline" and we don't override the insertion with a zero.
2062     if (DstIdx == SrcIdx && (ZMask & (1 << DstIdx)) == 0 &&
2063         countPopulation(ZMask) == 2) {
2064       unsigned AltIdx = findFirstSet((ZMask | (1 << DstIdx)) ^ 15);
2065       assert(AltIdx < 4 && "Illegal insertion index");
2066       unsigned AltImm = (AltIdx << 6) | (AltIdx << 4) | ZMask;
2067       auto &WorkingMI = cloneIfNew(MI);
2068       WorkingMI.getOperand(MI.getNumOperands() - 1).setImm(AltImm);
2069       return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
2070                                                      OpIdx1, OpIdx2);
2071     }
2072     return nullptr;
2073   }
2074   case X86::MOVSDrr:
2075   case X86::MOVSSrr:
2076   case X86::VMOVSDrr:
2077   case X86::VMOVSSrr:{
2078     // On SSE41 or later we can commute a MOVSS/MOVSD to a BLENDPS/BLENDPD.
2079     if (Subtarget.hasSSE41()) {
2080       unsigned Mask, Opc;
2081       switch (MI.getOpcode()) {
2082       default: llvm_unreachable("Unreachable!");
2083       case X86::MOVSDrr:  Opc = X86::BLENDPDrri;  Mask = 0x02; break;
2084       case X86::MOVSSrr:  Opc = X86::BLENDPSrri;  Mask = 0x0E; break;
2085       case X86::VMOVSDrr: Opc = X86::VBLENDPDrri; Mask = 0x02; break;
2086       case X86::VMOVSSrr: Opc = X86::VBLENDPSrri; Mask = 0x0E; break;
2087       }
2088 
2089       auto &WorkingMI = cloneIfNew(MI);
2090       WorkingMI.setDesc(get(Opc));
2091       WorkingMI.addOperand(MachineOperand::CreateImm(Mask));
2092       return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
2093                                                      OpIdx1, OpIdx2);
2094     }
2095 
2096     // Convert to SHUFPD.
2097     assert(MI.getOpcode() == X86::MOVSDrr &&
2098            "Can only commute MOVSDrr without SSE4.1");
2099 
2100     auto &WorkingMI = cloneIfNew(MI);
2101     WorkingMI.setDesc(get(X86::SHUFPDrri));
2102     WorkingMI.addOperand(MachineOperand::CreateImm(0x02));
2103     return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
2104                                                    OpIdx1, OpIdx2);
2105   }
2106   case X86::SHUFPDrri: {
2107     // Commute to MOVSD.
2108     assert(MI.getOperand(3).getImm() == 0x02 && "Unexpected immediate!");
2109     auto &WorkingMI = cloneIfNew(MI);
2110     WorkingMI.setDesc(get(X86::MOVSDrr));
2111     WorkingMI.removeOperand(3);
2112     return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
2113                                                    OpIdx1, OpIdx2);
2114   }
2115   case X86::PCLMULQDQrr:
2116   case X86::VPCLMULQDQrr:
2117   case X86::VPCLMULQDQYrr:
2118   case X86::VPCLMULQDQZrr:
2119   case X86::VPCLMULQDQZ128rr:
2120   case X86::VPCLMULQDQZ256rr: {
2121     // SRC1 64bits = Imm[0] ? SRC1[127:64] : SRC1[63:0]
2122     // SRC2 64bits = Imm[4] ? SRC2[127:64] : SRC2[63:0]
2123     unsigned Imm = MI.getOperand(3).getImm();
2124     unsigned Src1Hi = Imm & 0x01;
2125     unsigned Src2Hi = Imm & 0x10;
2126     auto &WorkingMI = cloneIfNew(MI);
2127     WorkingMI.getOperand(3).setImm((Src1Hi << 4) | (Src2Hi >> 4));
2128     return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
2129                                                    OpIdx1, OpIdx2);
2130   }
2131   case X86::VPCMPBZ128rri:  case X86::VPCMPUBZ128rri:
2132   case X86::VPCMPBZ256rri:  case X86::VPCMPUBZ256rri:
2133   case X86::VPCMPBZrri:     case X86::VPCMPUBZrri:
2134   case X86::VPCMPDZ128rri:  case X86::VPCMPUDZ128rri:
2135   case X86::VPCMPDZ256rri:  case X86::VPCMPUDZ256rri:
2136   case X86::VPCMPDZrri:     case X86::VPCMPUDZrri:
2137   case X86::VPCMPQZ128rri:  case X86::VPCMPUQZ128rri:
2138   case X86::VPCMPQZ256rri:  case X86::VPCMPUQZ256rri:
2139   case X86::VPCMPQZrri:     case X86::VPCMPUQZrri:
2140   case X86::VPCMPWZ128rri:  case X86::VPCMPUWZ128rri:
2141   case X86::VPCMPWZ256rri:  case X86::VPCMPUWZ256rri:
2142   case X86::VPCMPWZrri:     case X86::VPCMPUWZrri:
2143   case X86::VPCMPBZ128rrik: case X86::VPCMPUBZ128rrik:
2144   case X86::VPCMPBZ256rrik: case X86::VPCMPUBZ256rrik:
2145   case X86::VPCMPBZrrik:    case X86::VPCMPUBZrrik:
2146   case X86::VPCMPDZ128rrik: case X86::VPCMPUDZ128rrik:
2147   case X86::VPCMPDZ256rrik: case X86::VPCMPUDZ256rrik:
2148   case X86::VPCMPDZrrik:    case X86::VPCMPUDZrrik:
2149   case X86::VPCMPQZ128rrik: case X86::VPCMPUQZ128rrik:
2150   case X86::VPCMPQZ256rrik: case X86::VPCMPUQZ256rrik:
2151   case X86::VPCMPQZrrik:    case X86::VPCMPUQZrrik:
2152   case X86::VPCMPWZ128rrik: case X86::VPCMPUWZ128rrik:
2153   case X86::VPCMPWZ256rrik: case X86::VPCMPUWZ256rrik:
2154   case X86::VPCMPWZrrik:    case X86::VPCMPUWZrrik: {
2155     // Flip comparison mode immediate (if necessary).
2156     unsigned Imm = MI.getOperand(MI.getNumOperands() - 1).getImm() & 0x7;
2157     Imm = X86::getSwappedVPCMPImm(Imm);
2158     auto &WorkingMI = cloneIfNew(MI);
2159     WorkingMI.getOperand(MI.getNumOperands() - 1).setImm(Imm);
2160     return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
2161                                                    OpIdx1, OpIdx2);
2162   }
2163   case X86::VPCOMBri: case X86::VPCOMUBri:
2164   case X86::VPCOMDri: case X86::VPCOMUDri:
2165   case X86::VPCOMQri: case X86::VPCOMUQri:
2166   case X86::VPCOMWri: case X86::VPCOMUWri: {
2167     // Flip comparison mode immediate (if necessary).
2168     unsigned Imm = MI.getOperand(3).getImm() & 0x7;
2169     Imm = X86::getSwappedVPCOMImm(Imm);
2170     auto &WorkingMI = cloneIfNew(MI);
2171     WorkingMI.getOperand(3).setImm(Imm);
2172     return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
2173                                                    OpIdx1, OpIdx2);
2174   }
2175   case X86::VCMPSDZrr:
2176   case X86::VCMPSSZrr:
2177   case X86::VCMPPDZrri:
2178   case X86::VCMPPSZrri:
2179   case X86::VCMPSHZrr:
2180   case X86::VCMPPHZrri:
2181   case X86::VCMPPHZ128rri:
2182   case X86::VCMPPHZ256rri:
2183   case X86::VCMPPDZ128rri:
2184   case X86::VCMPPSZ128rri:
2185   case X86::VCMPPDZ256rri:
2186   case X86::VCMPPSZ256rri:
2187   case X86::VCMPPDZrrik:
2188   case X86::VCMPPSZrrik:
2189   case X86::VCMPPDZ128rrik:
2190   case X86::VCMPPSZ128rrik:
2191   case X86::VCMPPDZ256rrik:
2192   case X86::VCMPPSZ256rrik: {
2193     unsigned Imm =
2194                 MI.getOperand(MI.getNumExplicitOperands() - 1).getImm() & 0x1f;
2195     Imm = X86::getSwappedVCMPImm(Imm);
2196     auto &WorkingMI = cloneIfNew(MI);
2197     WorkingMI.getOperand(MI.getNumExplicitOperands() - 1).setImm(Imm);
2198     return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
2199                                                    OpIdx1, OpIdx2);
2200   }
2201   case X86::VPERM2F128rr:
2202   case X86::VPERM2I128rr: {
2203     // Flip permute source immediate.
2204     // Imm & 0x02: lo = if set, select Op1.lo/hi else Op0.lo/hi.
2205     // Imm & 0x20: hi = if set, select Op1.lo/hi else Op0.lo/hi.
2206     int8_t Imm = MI.getOperand(3).getImm() & 0xFF;
2207     auto &WorkingMI = cloneIfNew(MI);
2208     WorkingMI.getOperand(3).setImm(Imm ^ 0x22);
2209     return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
2210                                                    OpIdx1, OpIdx2);
2211   }
2212   case X86::MOVHLPSrr:
2213   case X86::UNPCKHPDrr:
2214   case X86::VMOVHLPSrr:
2215   case X86::VUNPCKHPDrr:
2216   case X86::VMOVHLPSZrr:
2217   case X86::VUNPCKHPDZ128rr: {
2218     assert(Subtarget.hasSSE2() && "Commuting MOVHLP/UNPCKHPD requires SSE2!");
2219 
2220     unsigned Opc = MI.getOpcode();
2221     switch (Opc) {
2222     default: llvm_unreachable("Unreachable!");
2223     case X86::MOVHLPSrr:       Opc = X86::UNPCKHPDrr;      break;
2224     case X86::UNPCKHPDrr:      Opc = X86::MOVHLPSrr;       break;
2225     case X86::VMOVHLPSrr:      Opc = X86::VUNPCKHPDrr;     break;
2226     case X86::VUNPCKHPDrr:     Opc = X86::VMOVHLPSrr;      break;
2227     case X86::VMOVHLPSZrr:     Opc = X86::VUNPCKHPDZ128rr; break;
2228     case X86::VUNPCKHPDZ128rr: Opc = X86::VMOVHLPSZrr;     break;
2229     }
2230     auto &WorkingMI = cloneIfNew(MI);
2231     WorkingMI.setDesc(get(Opc));
2232     return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
2233                                                    OpIdx1, OpIdx2);
2234   }
2235   case X86::CMOV16rr:  case X86::CMOV32rr:  case X86::CMOV64rr: {
2236     auto &WorkingMI = cloneIfNew(MI);
2237     unsigned OpNo = MI.getDesc().getNumOperands() - 1;
2238     X86::CondCode CC = static_cast<X86::CondCode>(MI.getOperand(OpNo).getImm());
2239     WorkingMI.getOperand(OpNo).setImm(X86::GetOppositeBranchCondition(CC));
2240     return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
2241                                                    OpIdx1, OpIdx2);
2242   }
2243   case X86::VPTERNLOGDZrri:      case X86::VPTERNLOGDZrmi:
2244   case X86::VPTERNLOGDZ128rri:   case X86::VPTERNLOGDZ128rmi:
2245   case X86::VPTERNLOGDZ256rri:   case X86::VPTERNLOGDZ256rmi:
2246   case X86::VPTERNLOGQZrri:      case X86::VPTERNLOGQZrmi:
2247   case X86::VPTERNLOGQZ128rri:   case X86::VPTERNLOGQZ128rmi:
2248   case X86::VPTERNLOGQZ256rri:   case X86::VPTERNLOGQZ256rmi:
2249   case X86::VPTERNLOGDZrrik:
2250   case X86::VPTERNLOGDZ128rrik:
2251   case X86::VPTERNLOGDZ256rrik:
2252   case X86::VPTERNLOGQZrrik:
2253   case X86::VPTERNLOGQZ128rrik:
2254   case X86::VPTERNLOGQZ256rrik:
2255   case X86::VPTERNLOGDZrrikz:    case X86::VPTERNLOGDZrmikz:
2256   case X86::VPTERNLOGDZ128rrikz: case X86::VPTERNLOGDZ128rmikz:
2257   case X86::VPTERNLOGDZ256rrikz: case X86::VPTERNLOGDZ256rmikz:
2258   case X86::VPTERNLOGQZrrikz:    case X86::VPTERNLOGQZrmikz:
2259   case X86::VPTERNLOGQZ128rrikz: case X86::VPTERNLOGQZ128rmikz:
2260   case X86::VPTERNLOGQZ256rrikz: case X86::VPTERNLOGQZ256rmikz:
2261   case X86::VPTERNLOGDZ128rmbi:
2262   case X86::VPTERNLOGDZ256rmbi:
2263   case X86::VPTERNLOGDZrmbi:
2264   case X86::VPTERNLOGQZ128rmbi:
2265   case X86::VPTERNLOGQZ256rmbi:
2266   case X86::VPTERNLOGQZrmbi:
2267   case X86::VPTERNLOGDZ128rmbikz:
2268   case X86::VPTERNLOGDZ256rmbikz:
2269   case X86::VPTERNLOGDZrmbikz:
2270   case X86::VPTERNLOGQZ128rmbikz:
2271   case X86::VPTERNLOGQZ256rmbikz:
2272   case X86::VPTERNLOGQZrmbikz: {
2273     auto &WorkingMI = cloneIfNew(MI);
2274     commuteVPTERNLOG(WorkingMI, OpIdx1, OpIdx2);
2275     return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
2276                                                    OpIdx1, OpIdx2);
2277   }
2278   default: {
2279     if (isCommutableVPERMV3Instruction(MI.getOpcode())) {
2280       unsigned Opc = getCommutedVPERMV3Opcode(MI.getOpcode());
2281       auto &WorkingMI = cloneIfNew(MI);
2282       WorkingMI.setDesc(get(Opc));
2283       return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
2284                                                      OpIdx1, OpIdx2);
2285     }
2286 
2287     const X86InstrFMA3Group *FMA3Group = getFMA3Group(MI.getOpcode(),
2288                                                       MI.getDesc().TSFlags);
2289     if (FMA3Group) {
2290       unsigned Opc =
2291         getFMA3OpcodeToCommuteOperands(MI, OpIdx1, OpIdx2, *FMA3Group);
2292       auto &WorkingMI = cloneIfNew(MI);
2293       WorkingMI.setDesc(get(Opc));
2294       return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
2295                                                      OpIdx1, OpIdx2);
2296     }
2297 
2298     return TargetInstrInfo::commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2);
2299   }
2300   }
2301 }
2302 
2303 bool
2304 X86InstrInfo::findThreeSrcCommutedOpIndices(const MachineInstr &MI,
2305                                             unsigned &SrcOpIdx1,
2306                                             unsigned &SrcOpIdx2,
2307                                             bool IsIntrinsic) const {
2308   uint64_t TSFlags = MI.getDesc().TSFlags;
2309 
2310   unsigned FirstCommutableVecOp = 1;
2311   unsigned LastCommutableVecOp = 3;
2312   unsigned KMaskOp = -1U;
2313   if (X86II::isKMasked(TSFlags)) {
2314     // For k-zero-masked operations it is Ok to commute the first vector
2315     // operand. Unless this is an intrinsic instruction.
2316     // For regular k-masked operations a conservative choice is done as the
2317     // elements of the first vector operand, for which the corresponding bit
2318     // in the k-mask operand is set to 0, are copied to the result of the
2319     // instruction.
2320     // TODO/FIXME: The commute still may be legal if it is known that the
2321     // k-mask operand is set to either all ones or all zeroes.
2322     // It is also Ok to commute the 1st operand if all users of MI use only
2323     // the elements enabled by the k-mask operand. For example,
2324     //   v4 = VFMADD213PSZrk v1, k, v2, v3; // v1[i] = k[i] ? v2[i]*v1[i]+v3[i]
2325     //                                                     : v1[i];
2326     //   VMOVAPSZmrk <mem_addr>, k, v4; // this is the ONLY user of v4 ->
2327     //                                  // Ok, to commute v1 in FMADD213PSZrk.
2328 
2329     // The k-mask operand has index = 2 for masked and zero-masked operations.
2330     KMaskOp = 2;
2331 
2332     // The operand with index = 1 is used as a source for those elements for
2333     // which the corresponding bit in the k-mask is set to 0.
2334     if (X86II::isKMergeMasked(TSFlags) || IsIntrinsic)
2335       FirstCommutableVecOp = 3;
2336 
2337     LastCommutableVecOp++;
2338   } else if (IsIntrinsic) {
2339     // Commuting the first operand of an intrinsic instruction isn't possible
2340     // unless we can prove that only the lowest element of the result is used.
2341     FirstCommutableVecOp = 2;
2342   }
2343 
2344   if (isMem(MI, LastCommutableVecOp))
2345     LastCommutableVecOp--;
2346 
2347   // Only the first RegOpsNum operands are commutable.
2348   // Also, the value 'CommuteAnyOperandIndex' is valid here as it means
2349   // that the operand is not specified/fixed.
2350   if (SrcOpIdx1 != CommuteAnyOperandIndex &&
2351       (SrcOpIdx1 < FirstCommutableVecOp || SrcOpIdx1 > LastCommutableVecOp ||
2352        SrcOpIdx1 == KMaskOp))
2353     return false;
2354   if (SrcOpIdx2 != CommuteAnyOperandIndex &&
2355       (SrcOpIdx2 < FirstCommutableVecOp || SrcOpIdx2 > LastCommutableVecOp ||
2356        SrcOpIdx2 == KMaskOp))
2357     return false;
2358 
2359   // Look for two different register operands assumed to be commutable
2360   // regardless of the FMA opcode. The FMA opcode is adjusted later.
2361   if (SrcOpIdx1 == CommuteAnyOperandIndex ||
2362       SrcOpIdx2 == CommuteAnyOperandIndex) {
2363     unsigned CommutableOpIdx2 = SrcOpIdx2;
2364 
2365     // At least one of operands to be commuted is not specified and
2366     // this method is free to choose appropriate commutable operands.
2367     if (SrcOpIdx1 == SrcOpIdx2)
2368       // Both of operands are not fixed. By default set one of commutable
2369       // operands to the last register operand of the instruction.
2370       CommutableOpIdx2 = LastCommutableVecOp;
2371     else if (SrcOpIdx2 == CommuteAnyOperandIndex)
2372       // Only one of operands is not fixed.
2373       CommutableOpIdx2 = SrcOpIdx1;
2374 
2375     // CommutableOpIdx2 is well defined now. Let's choose another commutable
2376     // operand and assign its index to CommutableOpIdx1.
2377     Register Op2Reg = MI.getOperand(CommutableOpIdx2).getReg();
2378 
2379     unsigned CommutableOpIdx1;
2380     for (CommutableOpIdx1 = LastCommutableVecOp;
2381          CommutableOpIdx1 >= FirstCommutableVecOp; CommutableOpIdx1--) {
2382       // Just ignore and skip the k-mask operand.
2383       if (CommutableOpIdx1 == KMaskOp)
2384         continue;
2385 
2386       // The commuted operands must have different registers.
2387       // Otherwise, the commute transformation does not change anything and
2388       // is useless then.
2389       if (Op2Reg != MI.getOperand(CommutableOpIdx1).getReg())
2390         break;
2391     }
2392 
2393     // No appropriate commutable operands were found.
2394     if (CommutableOpIdx1 < FirstCommutableVecOp)
2395       return false;
2396 
2397     // Assign the found pair of commutable indices to SrcOpIdx1 and SrcOpidx2
2398     // to return those values.
2399     if (!fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2,
2400                               CommutableOpIdx1, CommutableOpIdx2))
2401       return false;
2402   }
2403 
2404   return true;
2405 }
2406 
2407 bool X86InstrInfo::findCommutedOpIndices(const MachineInstr &MI,
2408                                          unsigned &SrcOpIdx1,
2409                                          unsigned &SrcOpIdx2) const {
2410   const MCInstrDesc &Desc = MI.getDesc();
2411   if (!Desc.isCommutable())
2412     return false;
2413 
2414   switch (MI.getOpcode()) {
2415   case X86::CMPSDrr:
2416   case X86::CMPSSrr:
2417   case X86::CMPPDrri:
2418   case X86::CMPPSrri:
2419   case X86::VCMPSDrr:
2420   case X86::VCMPSSrr:
2421   case X86::VCMPPDrri:
2422   case X86::VCMPPSrri:
2423   case X86::VCMPPDYrri:
2424   case X86::VCMPPSYrri:
2425   case X86::VCMPSDZrr:
2426   case X86::VCMPSSZrr:
2427   case X86::VCMPPDZrri:
2428   case X86::VCMPPSZrri:
2429   case X86::VCMPSHZrr:
2430   case X86::VCMPPHZrri:
2431   case X86::VCMPPHZ128rri:
2432   case X86::VCMPPHZ256rri:
2433   case X86::VCMPPDZ128rri:
2434   case X86::VCMPPSZ128rri:
2435   case X86::VCMPPDZ256rri:
2436   case X86::VCMPPSZ256rri:
2437   case X86::VCMPPDZrrik:
2438   case X86::VCMPPSZrrik:
2439   case X86::VCMPPDZ128rrik:
2440   case X86::VCMPPSZ128rrik:
2441   case X86::VCMPPDZ256rrik:
2442   case X86::VCMPPSZ256rrik: {
2443     unsigned OpOffset = X86II::isKMasked(Desc.TSFlags) ? 1 : 0;
2444 
2445     // Float comparison can be safely commuted for
2446     // Ordered/Unordered/Equal/NotEqual tests
2447     unsigned Imm = MI.getOperand(3 + OpOffset).getImm() & 0x7;
2448     switch (Imm) {
2449     default:
2450       // EVEX versions can be commuted.
2451       if ((Desc.TSFlags & X86II::EncodingMask) == X86II::EVEX)
2452         break;
2453       return false;
2454     case 0x00: // EQUAL
2455     case 0x03: // UNORDERED
2456     case 0x04: // NOT EQUAL
2457     case 0x07: // ORDERED
2458       break;
2459     }
2460 
2461     // The indices of the commutable operands are 1 and 2 (or 2 and 3
2462     // when masked).
2463     // Assign them to the returned operand indices here.
2464     return fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2, 1 + OpOffset,
2465                                 2 + OpOffset);
2466   }
2467   case X86::MOVSSrr:
2468     // X86::MOVSDrr is always commutable. MOVSS is only commutable if we can
2469     // form sse4.1 blend. We assume VMOVSSrr/VMOVSDrr is always commutable since
2470     // AVX implies sse4.1.
2471     if (Subtarget.hasSSE41())
2472       return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2);
2473     return false;
2474   case X86::SHUFPDrri:
2475     // We can commute this to MOVSD.
2476     if (MI.getOperand(3).getImm() == 0x02)
2477       return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2);
2478     return false;
2479   case X86::MOVHLPSrr:
2480   case X86::UNPCKHPDrr:
2481   case X86::VMOVHLPSrr:
2482   case X86::VUNPCKHPDrr:
2483   case X86::VMOVHLPSZrr:
2484   case X86::VUNPCKHPDZ128rr:
2485     if (Subtarget.hasSSE2())
2486       return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2);
2487     return false;
2488   case X86::VPTERNLOGDZrri:      case X86::VPTERNLOGDZrmi:
2489   case X86::VPTERNLOGDZ128rri:   case X86::VPTERNLOGDZ128rmi:
2490   case X86::VPTERNLOGDZ256rri:   case X86::VPTERNLOGDZ256rmi:
2491   case X86::VPTERNLOGQZrri:      case X86::VPTERNLOGQZrmi:
2492   case X86::VPTERNLOGQZ128rri:   case X86::VPTERNLOGQZ128rmi:
2493   case X86::VPTERNLOGQZ256rri:   case X86::VPTERNLOGQZ256rmi:
2494   case X86::VPTERNLOGDZrrik:
2495   case X86::VPTERNLOGDZ128rrik:
2496   case X86::VPTERNLOGDZ256rrik:
2497   case X86::VPTERNLOGQZrrik:
2498   case X86::VPTERNLOGQZ128rrik:
2499   case X86::VPTERNLOGQZ256rrik:
2500   case X86::VPTERNLOGDZrrikz:    case X86::VPTERNLOGDZrmikz:
2501   case X86::VPTERNLOGDZ128rrikz: case X86::VPTERNLOGDZ128rmikz:
2502   case X86::VPTERNLOGDZ256rrikz: case X86::VPTERNLOGDZ256rmikz:
2503   case X86::VPTERNLOGQZrrikz:    case X86::VPTERNLOGQZrmikz:
2504   case X86::VPTERNLOGQZ128rrikz: case X86::VPTERNLOGQZ128rmikz:
2505   case X86::VPTERNLOGQZ256rrikz: case X86::VPTERNLOGQZ256rmikz:
2506   case X86::VPTERNLOGDZ128rmbi:
2507   case X86::VPTERNLOGDZ256rmbi:
2508   case X86::VPTERNLOGDZrmbi:
2509   case X86::VPTERNLOGQZ128rmbi:
2510   case X86::VPTERNLOGQZ256rmbi:
2511   case X86::VPTERNLOGQZrmbi:
2512   case X86::VPTERNLOGDZ128rmbikz:
2513   case X86::VPTERNLOGDZ256rmbikz:
2514   case X86::VPTERNLOGDZrmbikz:
2515   case X86::VPTERNLOGQZ128rmbikz:
2516   case X86::VPTERNLOGQZ256rmbikz:
2517   case X86::VPTERNLOGQZrmbikz:
2518     return findThreeSrcCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2);
2519   case X86::VPDPWSSDYrr:
2520   case X86::VPDPWSSDrr:
2521   case X86::VPDPWSSDSYrr:
2522   case X86::VPDPWSSDSrr:
2523   case X86::VPDPWSSDZ128r:
2524   case X86::VPDPWSSDZ128rk:
2525   case X86::VPDPWSSDZ128rkz:
2526   case X86::VPDPWSSDZ256r:
2527   case X86::VPDPWSSDZ256rk:
2528   case X86::VPDPWSSDZ256rkz:
2529   case X86::VPDPWSSDZr:
2530   case X86::VPDPWSSDZrk:
2531   case X86::VPDPWSSDZrkz:
2532   case X86::VPDPWSSDSZ128r:
2533   case X86::VPDPWSSDSZ128rk:
2534   case X86::VPDPWSSDSZ128rkz:
2535   case X86::VPDPWSSDSZ256r:
2536   case X86::VPDPWSSDSZ256rk:
2537   case X86::VPDPWSSDSZ256rkz:
2538   case X86::VPDPWSSDSZr:
2539   case X86::VPDPWSSDSZrk:
2540   case X86::VPDPWSSDSZrkz:
2541   case X86::VPMADD52HUQZ128r:
2542   case X86::VPMADD52HUQZ128rk:
2543   case X86::VPMADD52HUQZ128rkz:
2544   case X86::VPMADD52HUQZ256r:
2545   case X86::VPMADD52HUQZ256rk:
2546   case X86::VPMADD52HUQZ256rkz:
2547   case X86::VPMADD52HUQZr:
2548   case X86::VPMADD52HUQZrk:
2549   case X86::VPMADD52HUQZrkz:
2550   case X86::VPMADD52LUQZ128r:
2551   case X86::VPMADD52LUQZ128rk:
2552   case X86::VPMADD52LUQZ128rkz:
2553   case X86::VPMADD52LUQZ256r:
2554   case X86::VPMADD52LUQZ256rk:
2555   case X86::VPMADD52LUQZ256rkz:
2556   case X86::VPMADD52LUQZr:
2557   case X86::VPMADD52LUQZrk:
2558   case X86::VPMADD52LUQZrkz:
2559   case X86::VFMADDCPHZr:
2560   case X86::VFMADDCPHZrk:
2561   case X86::VFMADDCPHZrkz:
2562   case X86::VFMADDCPHZ128r:
2563   case X86::VFMADDCPHZ128rk:
2564   case X86::VFMADDCPHZ128rkz:
2565   case X86::VFMADDCPHZ256r:
2566   case X86::VFMADDCPHZ256rk:
2567   case X86::VFMADDCPHZ256rkz:
2568   case X86::VFMADDCSHZr:
2569   case X86::VFMADDCSHZrk:
2570   case X86::VFMADDCSHZrkz: {
2571     unsigned CommutableOpIdx1 = 2;
2572     unsigned CommutableOpIdx2 = 3;
2573     if (X86II::isKMasked(Desc.TSFlags)) {
2574       // Skip the mask register.
2575       ++CommutableOpIdx1;
2576       ++CommutableOpIdx2;
2577     }
2578     if (!fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2,
2579                               CommutableOpIdx1, CommutableOpIdx2))
2580       return false;
2581     if (!MI.getOperand(SrcOpIdx1).isReg() ||
2582         !MI.getOperand(SrcOpIdx2).isReg())
2583       // No idea.
2584       return false;
2585     return true;
2586   }
2587 
2588   default:
2589     const X86InstrFMA3Group *FMA3Group = getFMA3Group(MI.getOpcode(),
2590                                                       MI.getDesc().TSFlags);
2591     if (FMA3Group)
2592       return findThreeSrcCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2,
2593                                            FMA3Group->isIntrinsic());
2594 
2595     // Handled masked instructions since we need to skip over the mask input
2596     // and the preserved input.
2597     if (X86II::isKMasked(Desc.TSFlags)) {
2598       // First assume that the first input is the mask operand and skip past it.
2599       unsigned CommutableOpIdx1 = Desc.getNumDefs() + 1;
2600       unsigned CommutableOpIdx2 = Desc.getNumDefs() + 2;
2601       // Check if the first input is tied. If there isn't one then we only
2602       // need to skip the mask operand which we did above.
2603       if ((MI.getDesc().getOperandConstraint(Desc.getNumDefs(),
2604                                              MCOI::TIED_TO) != -1)) {
2605         // If this is zero masking instruction with a tied operand, we need to
2606         // move the first index back to the first input since this must
2607         // be a 3 input instruction and we want the first two non-mask inputs.
2608         // Otherwise this is a 2 input instruction with a preserved input and
2609         // mask, so we need to move the indices to skip one more input.
2610         if (X86II::isKMergeMasked(Desc.TSFlags)) {
2611           ++CommutableOpIdx1;
2612           ++CommutableOpIdx2;
2613         } else {
2614           --CommutableOpIdx1;
2615         }
2616       }
2617 
2618       if (!fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2,
2619                                 CommutableOpIdx1, CommutableOpIdx2))
2620         return false;
2621 
2622       if (!MI.getOperand(SrcOpIdx1).isReg() ||
2623           !MI.getOperand(SrcOpIdx2).isReg())
2624         // No idea.
2625         return false;
2626       return true;
2627     }
2628 
2629     return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2);
2630   }
2631   return false;
2632 }
2633 
2634 static bool isConvertibleLEA(MachineInstr *MI) {
2635   unsigned Opcode = MI->getOpcode();
2636   if (Opcode != X86::LEA32r && Opcode != X86::LEA64r &&
2637       Opcode != X86::LEA64_32r)
2638     return false;
2639 
2640   const MachineOperand &Scale = MI->getOperand(1 + X86::AddrScaleAmt);
2641   const MachineOperand &Disp = MI->getOperand(1 + X86::AddrDisp);
2642   const MachineOperand &Segment = MI->getOperand(1 + X86::AddrSegmentReg);
2643 
2644   if (Segment.getReg() != 0 || !Disp.isImm() || Disp.getImm() != 0 ||
2645       Scale.getImm() > 1)
2646     return false;
2647 
2648   return true;
2649 }
2650 
2651 bool X86InstrInfo::hasCommutePreference(MachineInstr &MI, bool &Commute) const {
2652   // Currently we're interested in following sequence only.
2653   //   r3 = lea r1, r2
2654   //   r5 = add r3, r4
2655   // Both r3 and r4 are killed in add, we hope the add instruction has the
2656   // operand order
2657   //   r5 = add r4, r3
2658   // So later in X86FixupLEAs the lea instruction can be rewritten as add.
2659   unsigned Opcode = MI.getOpcode();
2660   if (Opcode != X86::ADD32rr && Opcode != X86::ADD64rr)
2661     return false;
2662 
2663   const MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
2664   Register Reg1 = MI.getOperand(1).getReg();
2665   Register Reg2 = MI.getOperand(2).getReg();
2666 
2667   // Check if Reg1 comes from LEA in the same MBB.
2668   if (MachineInstr *Inst = MRI.getUniqueVRegDef(Reg1)) {
2669     if (isConvertibleLEA(Inst) && Inst->getParent() == MI.getParent()) {
2670       Commute = true;
2671       return true;
2672     }
2673   }
2674 
2675   // Check if Reg2 comes from LEA in the same MBB.
2676   if (MachineInstr *Inst = MRI.getUniqueVRegDef(Reg2)) {
2677     if (isConvertibleLEA(Inst) && Inst->getParent() == MI.getParent()) {
2678       Commute = false;
2679       return true;
2680     }
2681   }
2682 
2683   return false;
2684 }
2685 
2686 int X86::getCondSrcNoFromDesc(const MCInstrDesc &MCID) {
2687   unsigned Opcode = MCID.getOpcode();
2688   if (!(X86::isJCC(Opcode) || X86::isSETCC(Opcode) || X86::isCMOVCC(Opcode)))
2689     return -1;
2690   // Assume that condition code is always the last use operand.
2691   unsigned NumUses = MCID.getNumOperands() - MCID.getNumDefs();
2692   return NumUses - 1;
2693 }
2694 
2695 X86::CondCode X86::getCondFromMI(const MachineInstr &MI) {
2696   const MCInstrDesc &MCID = MI.getDesc();
2697   int CondNo = getCondSrcNoFromDesc(MCID);
2698   if (CondNo < 0)
2699     return X86::COND_INVALID;
2700   CondNo += MCID.getNumDefs();
2701   return static_cast<X86::CondCode>(MI.getOperand(CondNo).getImm());
2702 }
2703 
2704 X86::CondCode X86::getCondFromBranch(const MachineInstr &MI) {
2705   return X86::isJCC(MI.getOpcode()) ? X86::getCondFromMI(MI)
2706                                     : X86::COND_INVALID;
2707 }
2708 
2709 X86::CondCode X86::getCondFromSETCC(const MachineInstr &MI) {
2710   return X86::isSETCC(MI.getOpcode()) ? X86::getCondFromMI(MI)
2711                                       : X86::COND_INVALID;
2712 }
2713 
2714 X86::CondCode X86::getCondFromCMov(const MachineInstr &MI) {
2715   return X86::isCMOVCC(MI.getOpcode()) ? X86::getCondFromMI(MI)
2716                                        : X86::COND_INVALID;
2717 }
2718 
2719 /// Return the inverse of the specified condition,
2720 /// e.g. turning COND_E to COND_NE.
2721 X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) {
2722   switch (CC) {
2723   default: llvm_unreachable("Illegal condition code!");
2724   case X86::COND_E:  return X86::COND_NE;
2725   case X86::COND_NE: return X86::COND_E;
2726   case X86::COND_L:  return X86::COND_GE;
2727   case X86::COND_LE: return X86::COND_G;
2728   case X86::COND_G:  return X86::COND_LE;
2729   case X86::COND_GE: return X86::COND_L;
2730   case X86::COND_B:  return X86::COND_AE;
2731   case X86::COND_BE: return X86::COND_A;
2732   case X86::COND_A:  return X86::COND_BE;
2733   case X86::COND_AE: return X86::COND_B;
2734   case X86::COND_S:  return X86::COND_NS;
2735   case X86::COND_NS: return X86::COND_S;
2736   case X86::COND_P:  return X86::COND_NP;
2737   case X86::COND_NP: return X86::COND_P;
2738   case X86::COND_O:  return X86::COND_NO;
2739   case X86::COND_NO: return X86::COND_O;
2740   case X86::COND_NE_OR_P:  return X86::COND_E_AND_NP;
2741   case X86::COND_E_AND_NP: return X86::COND_NE_OR_P;
2742   }
2743 }
2744 
2745 /// Assuming the flags are set by MI(a,b), return the condition code if we
2746 /// modify the instructions such that flags are set by MI(b,a).
2747 static X86::CondCode getSwappedCondition(X86::CondCode CC) {
2748   switch (CC) {
2749   default: return X86::COND_INVALID;
2750   case X86::COND_E:  return X86::COND_E;
2751   case X86::COND_NE: return X86::COND_NE;
2752   case X86::COND_L:  return X86::COND_G;
2753   case X86::COND_LE: return X86::COND_GE;
2754   case X86::COND_G:  return X86::COND_L;
2755   case X86::COND_GE: return X86::COND_LE;
2756   case X86::COND_B:  return X86::COND_A;
2757   case X86::COND_BE: return X86::COND_AE;
2758   case X86::COND_A:  return X86::COND_B;
2759   case X86::COND_AE: return X86::COND_BE;
2760   }
2761 }
2762 
2763 std::pair<X86::CondCode, bool>
2764 X86::getX86ConditionCode(CmpInst::Predicate Predicate) {
2765   X86::CondCode CC = X86::COND_INVALID;
2766   bool NeedSwap = false;
2767   switch (Predicate) {
2768   default: break;
2769   // Floating-point Predicates
2770   case CmpInst::FCMP_UEQ: CC = X86::COND_E;       break;
2771   case CmpInst::FCMP_OLT: NeedSwap = true;        LLVM_FALLTHROUGH;
2772   case CmpInst::FCMP_OGT: CC = X86::COND_A;       break;
2773   case CmpInst::FCMP_OLE: NeedSwap = true;        LLVM_FALLTHROUGH;
2774   case CmpInst::FCMP_OGE: CC = X86::COND_AE;      break;
2775   case CmpInst::FCMP_UGT: NeedSwap = true;        LLVM_FALLTHROUGH;
2776   case CmpInst::FCMP_ULT: CC = X86::COND_B;       break;
2777   case CmpInst::FCMP_UGE: NeedSwap = true;        LLVM_FALLTHROUGH;
2778   case CmpInst::FCMP_ULE: CC = X86::COND_BE;      break;
2779   case CmpInst::FCMP_ONE: CC = X86::COND_NE;      break;
2780   case CmpInst::FCMP_UNO: CC = X86::COND_P;       break;
2781   case CmpInst::FCMP_ORD: CC = X86::COND_NP;      break;
2782   case CmpInst::FCMP_OEQ:                         LLVM_FALLTHROUGH;
2783   case CmpInst::FCMP_UNE: CC = X86::COND_INVALID; break;
2784 
2785   // Integer Predicates
2786   case CmpInst::ICMP_EQ:  CC = X86::COND_E;       break;
2787   case CmpInst::ICMP_NE:  CC = X86::COND_NE;      break;
2788   case CmpInst::ICMP_UGT: CC = X86::COND_A;       break;
2789   case CmpInst::ICMP_UGE: CC = X86::COND_AE;      break;
2790   case CmpInst::ICMP_ULT: CC = X86::COND_B;       break;
2791   case CmpInst::ICMP_ULE: CC = X86::COND_BE;      break;
2792   case CmpInst::ICMP_SGT: CC = X86::COND_G;       break;
2793   case CmpInst::ICMP_SGE: CC = X86::COND_GE;      break;
2794   case CmpInst::ICMP_SLT: CC = X86::COND_L;       break;
2795   case CmpInst::ICMP_SLE: CC = X86::COND_LE;      break;
2796   }
2797 
2798   return std::make_pair(CC, NeedSwap);
2799 }
2800 
2801 /// Return a cmov opcode for the given register size in bytes, and operand type.
2802 unsigned X86::getCMovOpcode(unsigned RegBytes, bool HasMemoryOperand) {
2803   switch(RegBytes) {
2804   default: llvm_unreachable("Illegal register size!");
2805   case 2: return HasMemoryOperand ? X86::CMOV16rm : X86::CMOV16rr;
2806   case 4: return HasMemoryOperand ? X86::CMOV32rm : X86::CMOV32rr;
2807   case 8: return HasMemoryOperand ? X86::CMOV64rm : X86::CMOV64rr;
2808   }
2809 }
2810 
2811 /// Get the VPCMP immediate for the given condition.
2812 unsigned X86::getVPCMPImmForCond(ISD::CondCode CC) {
2813   switch (CC) {
2814   default: llvm_unreachable("Unexpected SETCC condition");
2815   case ISD::SETNE:  return 4;
2816   case ISD::SETEQ:  return 0;
2817   case ISD::SETULT:
2818   case ISD::SETLT: return 1;
2819   case ISD::SETUGT:
2820   case ISD::SETGT: return 6;
2821   case ISD::SETUGE:
2822   case ISD::SETGE: return 5;
2823   case ISD::SETULE:
2824   case ISD::SETLE: return 2;
2825   }
2826 }
2827 
2828 /// Get the VPCMP immediate if the operands are swapped.
2829 unsigned X86::getSwappedVPCMPImm(unsigned Imm) {
2830   switch (Imm) {
2831   default: llvm_unreachable("Unreachable!");
2832   case 0x01: Imm = 0x06; break; // LT  -> NLE
2833   case 0x02: Imm = 0x05; break; // LE  -> NLT
2834   case 0x05: Imm = 0x02; break; // NLT -> LE
2835   case 0x06: Imm = 0x01; break; // NLE -> LT
2836   case 0x00: // EQ
2837   case 0x03: // FALSE
2838   case 0x04: // NE
2839   case 0x07: // TRUE
2840     break;
2841   }
2842 
2843   return Imm;
2844 }
2845 
2846 /// Get the VPCOM immediate if the operands are swapped.
2847 unsigned X86::getSwappedVPCOMImm(unsigned Imm) {
2848   switch (Imm) {
2849   default: llvm_unreachable("Unreachable!");
2850   case 0x00: Imm = 0x02; break; // LT -> GT
2851   case 0x01: Imm = 0x03; break; // LE -> GE
2852   case 0x02: Imm = 0x00; break; // GT -> LT
2853   case 0x03: Imm = 0x01; break; // GE -> LE
2854   case 0x04: // EQ
2855   case 0x05: // NE
2856   case 0x06: // FALSE
2857   case 0x07: // TRUE
2858     break;
2859   }
2860 
2861   return Imm;
2862 }
2863 
2864 /// Get the VCMP immediate if the operands are swapped.
2865 unsigned X86::getSwappedVCMPImm(unsigned Imm) {
2866   // Only need the lower 2 bits to distinquish.
2867   switch (Imm & 0x3) {
2868   default: llvm_unreachable("Unreachable!");
2869   case 0x00: case 0x03:
2870     // EQ/NE/TRUE/FALSE/ORD/UNORD don't change immediate when commuted.
2871     break;
2872   case 0x01: case 0x02:
2873     // Need to toggle bits 3:0. Bit 4 stays the same.
2874     Imm ^= 0xf;
2875     break;
2876   }
2877 
2878   return Imm;
2879 }
2880 
2881 /// Return true if the Reg is X87 register.
2882 static bool isX87Reg(unsigned Reg) {
2883   return (Reg == X86::FPCW || Reg == X86::FPSW ||
2884           (Reg >= X86::ST0 && Reg <= X86::ST7));
2885 }
2886 
2887 /// check if the instruction is X87 instruction
2888 bool X86::isX87Instruction(MachineInstr &MI) {
2889   for (const MachineOperand &MO : MI.operands()) {
2890     if (!MO.isReg())
2891       continue;
2892     if (isX87Reg(MO.getReg()))
2893       return true;
2894   }
2895   return false;
2896 }
2897 
2898 bool X86InstrInfo::isUnconditionalTailCall(const MachineInstr &MI) const {
2899   switch (MI.getOpcode()) {
2900   case X86::TCRETURNdi:
2901   case X86::TCRETURNri:
2902   case X86::TCRETURNmi:
2903   case X86::TCRETURNdi64:
2904   case X86::TCRETURNri64:
2905   case X86::TCRETURNmi64:
2906     return true;
2907   default:
2908     return false;
2909   }
2910 }
2911 
2912 bool X86InstrInfo::canMakeTailCallConditional(
2913     SmallVectorImpl<MachineOperand> &BranchCond,
2914     const MachineInstr &TailCall) const {
2915   if (TailCall.getOpcode() != X86::TCRETURNdi &&
2916       TailCall.getOpcode() != X86::TCRETURNdi64) {
2917     // Only direct calls can be done with a conditional branch.
2918     return false;
2919   }
2920 
2921   const MachineFunction *MF = TailCall.getParent()->getParent();
2922   if (Subtarget.isTargetWin64() && MF->hasWinCFI()) {
2923     // Conditional tail calls confuse the Win64 unwinder.
2924     return false;
2925   }
2926 
2927   assert(BranchCond.size() == 1);
2928   if (BranchCond[0].getImm() > X86::LAST_VALID_COND) {
2929     // Can't make a conditional tail call with this condition.
2930     return false;
2931   }
2932 
2933   const X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
2934   if (X86FI->getTCReturnAddrDelta() != 0 ||
2935       TailCall.getOperand(1).getImm() != 0) {
2936     // A conditional tail call cannot do any stack adjustment.
2937     return false;
2938   }
2939 
2940   return true;
2941 }
2942 
2943 void X86InstrInfo::replaceBranchWithTailCall(
2944     MachineBasicBlock &MBB, SmallVectorImpl<MachineOperand> &BranchCond,
2945     const MachineInstr &TailCall) const {
2946   assert(canMakeTailCallConditional(BranchCond, TailCall));
2947 
2948   MachineBasicBlock::iterator I = MBB.end();
2949   while (I != MBB.begin()) {
2950     --I;
2951     if (I->isDebugInstr())
2952       continue;
2953     if (!I->isBranch())
2954       assert(0 && "Can't find the branch to replace!");
2955 
2956     X86::CondCode CC = X86::getCondFromBranch(*I);
2957     assert(BranchCond.size() == 1);
2958     if (CC != BranchCond[0].getImm())
2959       continue;
2960 
2961     break;
2962   }
2963 
2964   unsigned Opc = TailCall.getOpcode() == X86::TCRETURNdi ? X86::TCRETURNdicc
2965                                                          : X86::TCRETURNdi64cc;
2966 
2967   auto MIB = BuildMI(MBB, I, MBB.findDebugLoc(I), get(Opc));
2968   MIB->addOperand(TailCall.getOperand(0)); // Destination.
2969   MIB.addImm(0); // Stack offset (not used).
2970   MIB->addOperand(BranchCond[0]); // Condition.
2971   MIB.copyImplicitOps(TailCall); // Regmask and (imp-used) parameters.
2972 
2973   // Add implicit uses and defs of all live regs potentially clobbered by the
2974   // call. This way they still appear live across the call.
2975   LivePhysRegs LiveRegs(getRegisterInfo());
2976   LiveRegs.addLiveOuts(MBB);
2977   SmallVector<std::pair<MCPhysReg, const MachineOperand *>, 8> Clobbers;
2978   LiveRegs.stepForward(*MIB, Clobbers);
2979   for (const auto &C : Clobbers) {
2980     MIB.addReg(C.first, RegState::Implicit);
2981     MIB.addReg(C.first, RegState::Implicit | RegState::Define);
2982   }
2983 
2984   I->eraseFromParent();
2985 }
2986 
2987 // Given a MBB and its TBB, find the FBB which was a fallthrough MBB (it may
2988 // not be a fallthrough MBB now due to layout changes). Return nullptr if the
2989 // fallthrough MBB cannot be identified.
2990 static MachineBasicBlock *getFallThroughMBB(MachineBasicBlock *MBB,
2991                                             MachineBasicBlock *TBB) {
2992   // Look for non-EHPad successors other than TBB. If we find exactly one, it
2993   // is the fallthrough MBB. If we find zero, then TBB is both the target MBB
2994   // and fallthrough MBB. If we find more than one, we cannot identify the
2995   // fallthrough MBB and should return nullptr.
2996   MachineBasicBlock *FallthroughBB = nullptr;
2997   for (MachineBasicBlock *Succ : MBB->successors()) {
2998     if (Succ->isEHPad() || (Succ == TBB && FallthroughBB))
2999       continue;
3000     // Return a nullptr if we found more than one fallthrough successor.
3001     if (FallthroughBB && FallthroughBB != TBB)
3002       return nullptr;
3003     FallthroughBB = Succ;
3004   }
3005   return FallthroughBB;
3006 }
3007 
3008 bool X86InstrInfo::AnalyzeBranchImpl(
3009     MachineBasicBlock &MBB, MachineBasicBlock *&TBB, MachineBasicBlock *&FBB,
3010     SmallVectorImpl<MachineOperand> &Cond,
3011     SmallVectorImpl<MachineInstr *> &CondBranches, bool AllowModify) const {
3012 
3013   // Start from the bottom of the block and work up, examining the
3014   // terminator instructions.
3015   MachineBasicBlock::iterator I = MBB.end();
3016   MachineBasicBlock::iterator UnCondBrIter = MBB.end();
3017   while (I != MBB.begin()) {
3018     --I;
3019     if (I->isDebugInstr())
3020       continue;
3021 
3022     // Working from the bottom, when we see a non-terminator instruction, we're
3023     // done.
3024     if (!isUnpredicatedTerminator(*I))
3025       break;
3026 
3027     // A terminator that isn't a branch can't easily be handled by this
3028     // analysis.
3029     if (!I->isBranch())
3030       return true;
3031 
3032     // Handle unconditional branches.
3033     if (I->getOpcode() == X86::JMP_1) {
3034       UnCondBrIter = I;
3035 
3036       if (!AllowModify) {
3037         TBB = I->getOperand(0).getMBB();
3038         continue;
3039       }
3040 
3041       // If the block has any instructions after a JMP, delete them.
3042       MBB.erase(std::next(I), MBB.end());
3043 
3044       Cond.clear();
3045       FBB = nullptr;
3046 
3047       // Delete the JMP if it's equivalent to a fall-through.
3048       if (MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) {
3049         TBB = nullptr;
3050         I->eraseFromParent();
3051         I = MBB.end();
3052         UnCondBrIter = MBB.end();
3053         continue;
3054       }
3055 
3056       // TBB is used to indicate the unconditional destination.
3057       TBB = I->getOperand(0).getMBB();
3058       continue;
3059     }
3060 
3061     // Handle conditional branches.
3062     X86::CondCode BranchCode = X86::getCondFromBranch(*I);
3063     if (BranchCode == X86::COND_INVALID)
3064       return true;  // Can't handle indirect branch.
3065 
3066     // In practice we should never have an undef eflags operand, if we do
3067     // abort here as we are not prepared to preserve the flag.
3068     if (I->findRegisterUseOperand(X86::EFLAGS)->isUndef())
3069       return true;
3070 
3071     // Working from the bottom, handle the first conditional branch.
3072     if (Cond.empty()) {
3073       MachineBasicBlock *TargetBB = I->getOperand(0).getMBB();
3074       if (AllowModify && UnCondBrIter != MBB.end() &&
3075           MBB.isLayoutSuccessor(TargetBB)) {
3076         // If we can modify the code and it ends in something like:
3077         //
3078         //     jCC L1
3079         //     jmp L2
3080         //   L1:
3081         //     ...
3082         //   L2:
3083         //
3084         // Then we can change this to:
3085         //
3086         //     jnCC L2
3087         //   L1:
3088         //     ...
3089         //   L2:
3090         //
3091         // Which is a bit more efficient.
3092         // We conditionally jump to the fall-through block.
3093         BranchCode = GetOppositeBranchCondition(BranchCode);
3094         MachineBasicBlock::iterator OldInst = I;
3095 
3096         BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(X86::JCC_1))
3097           .addMBB(UnCondBrIter->getOperand(0).getMBB())
3098           .addImm(BranchCode);
3099         BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(X86::JMP_1))
3100           .addMBB(TargetBB);
3101 
3102         OldInst->eraseFromParent();
3103         UnCondBrIter->eraseFromParent();
3104 
3105         // Restart the analysis.
3106         UnCondBrIter = MBB.end();
3107         I = MBB.end();
3108         continue;
3109       }
3110 
3111       FBB = TBB;
3112       TBB = I->getOperand(0).getMBB();
3113       Cond.push_back(MachineOperand::CreateImm(BranchCode));
3114       CondBranches.push_back(&*I);
3115       continue;
3116     }
3117 
3118     // Handle subsequent conditional branches. Only handle the case where all
3119     // conditional branches branch to the same destination and their condition
3120     // opcodes fit one of the special multi-branch idioms.
3121     assert(Cond.size() == 1);
3122     assert(TBB);
3123 
3124     // If the conditions are the same, we can leave them alone.
3125     X86::CondCode OldBranchCode = (X86::CondCode)Cond[0].getImm();
3126     auto NewTBB = I->getOperand(0).getMBB();
3127     if (OldBranchCode == BranchCode && TBB == NewTBB)
3128       continue;
3129 
3130     // If they differ, see if they fit one of the known patterns. Theoretically,
3131     // we could handle more patterns here, but we shouldn't expect to see them
3132     // if instruction selection has done a reasonable job.
3133     if (TBB == NewTBB &&
3134                ((OldBranchCode == X86::COND_P && BranchCode == X86::COND_NE) ||
3135                 (OldBranchCode == X86::COND_NE && BranchCode == X86::COND_P))) {
3136       BranchCode = X86::COND_NE_OR_P;
3137     } else if ((OldBranchCode == X86::COND_NP && BranchCode == X86::COND_NE) ||
3138                (OldBranchCode == X86::COND_E && BranchCode == X86::COND_P)) {
3139       if (NewTBB != (FBB ? FBB : getFallThroughMBB(&MBB, TBB)))
3140         return true;
3141 
3142       // X86::COND_E_AND_NP usually has two different branch destinations.
3143       //
3144       // JP B1
3145       // JE B2
3146       // JMP B1
3147       // B1:
3148       // B2:
3149       //
3150       // Here this condition branches to B2 only if NP && E. It has another
3151       // equivalent form:
3152       //
3153       // JNE B1
3154       // JNP B2
3155       // JMP B1
3156       // B1:
3157       // B2:
3158       //
3159       // Similarly it branches to B2 only if E && NP. That is why this condition
3160       // is named with COND_E_AND_NP.
3161       BranchCode = X86::COND_E_AND_NP;
3162     } else
3163       return true;
3164 
3165     // Update the MachineOperand.
3166     Cond[0].setImm(BranchCode);
3167     CondBranches.push_back(&*I);
3168   }
3169 
3170   return false;
3171 }
3172 
3173 bool X86InstrInfo::analyzeBranch(MachineBasicBlock &MBB,
3174                                  MachineBasicBlock *&TBB,
3175                                  MachineBasicBlock *&FBB,
3176                                  SmallVectorImpl<MachineOperand> &Cond,
3177                                  bool AllowModify) const {
3178   SmallVector<MachineInstr *, 4> CondBranches;
3179   return AnalyzeBranchImpl(MBB, TBB, FBB, Cond, CondBranches, AllowModify);
3180 }
3181 
3182 bool X86InstrInfo::analyzeBranchPredicate(MachineBasicBlock &MBB,
3183                                           MachineBranchPredicate &MBP,
3184                                           bool AllowModify) const {
3185   using namespace std::placeholders;
3186 
3187   SmallVector<MachineOperand, 4> Cond;
3188   SmallVector<MachineInstr *, 4> CondBranches;
3189   if (AnalyzeBranchImpl(MBB, MBP.TrueDest, MBP.FalseDest, Cond, CondBranches,
3190                         AllowModify))
3191     return true;
3192 
3193   if (Cond.size() != 1)
3194     return true;
3195 
3196   assert(MBP.TrueDest && "expected!");
3197 
3198   if (!MBP.FalseDest)
3199     MBP.FalseDest = MBB.getNextNode();
3200 
3201   const TargetRegisterInfo *TRI = &getRegisterInfo();
3202 
3203   MachineInstr *ConditionDef = nullptr;
3204   bool SingleUseCondition = true;
3205 
3206   for (MachineInstr &MI : llvm::drop_begin(llvm::reverse(MBB))) {
3207     if (MI.modifiesRegister(X86::EFLAGS, TRI)) {
3208       ConditionDef = &MI;
3209       break;
3210     }
3211 
3212     if (MI.readsRegister(X86::EFLAGS, TRI))
3213       SingleUseCondition = false;
3214   }
3215 
3216   if (!ConditionDef)
3217     return true;
3218 
3219   if (SingleUseCondition) {
3220     for (auto *Succ : MBB.successors())
3221       if (Succ->isLiveIn(X86::EFLAGS))
3222         SingleUseCondition = false;
3223   }
3224 
3225   MBP.ConditionDef = ConditionDef;
3226   MBP.SingleUseCondition = SingleUseCondition;
3227 
3228   // Currently we only recognize the simple pattern:
3229   //
3230   //   test %reg, %reg
3231   //   je %label
3232   //
3233   const unsigned TestOpcode =
3234       Subtarget.is64Bit() ? X86::TEST64rr : X86::TEST32rr;
3235 
3236   if (ConditionDef->getOpcode() == TestOpcode &&
3237       ConditionDef->getNumOperands() == 3 &&
3238       ConditionDef->getOperand(0).isIdenticalTo(ConditionDef->getOperand(1)) &&
3239       (Cond[0].getImm() == X86::COND_NE || Cond[0].getImm() == X86::COND_E)) {
3240     MBP.LHS = ConditionDef->getOperand(0);
3241     MBP.RHS = MachineOperand::CreateImm(0);
3242     MBP.Predicate = Cond[0].getImm() == X86::COND_NE
3243                         ? MachineBranchPredicate::PRED_NE
3244                         : MachineBranchPredicate::PRED_EQ;
3245     return false;
3246   }
3247 
3248   return true;
3249 }
3250 
3251 unsigned X86InstrInfo::removeBranch(MachineBasicBlock &MBB,
3252                                     int *BytesRemoved) const {
3253   assert(!BytesRemoved && "code size not handled");
3254 
3255   MachineBasicBlock::iterator I = MBB.end();
3256   unsigned Count = 0;
3257 
3258   while (I != MBB.begin()) {
3259     --I;
3260     if (I->isDebugInstr())
3261       continue;
3262     if (I->getOpcode() != X86::JMP_1 &&
3263         X86::getCondFromBranch(*I) == X86::COND_INVALID)
3264       break;
3265     // Remove the branch.
3266     I->eraseFromParent();
3267     I = MBB.end();
3268     ++Count;
3269   }
3270 
3271   return Count;
3272 }
3273 
3274 unsigned X86InstrInfo::insertBranch(MachineBasicBlock &MBB,
3275                                     MachineBasicBlock *TBB,
3276                                     MachineBasicBlock *FBB,
3277                                     ArrayRef<MachineOperand> Cond,
3278                                     const DebugLoc &DL,
3279                                     int *BytesAdded) const {
3280   // Shouldn't be a fall through.
3281   assert(TBB && "insertBranch must not be told to insert a fallthrough");
3282   assert((Cond.size() == 1 || Cond.size() == 0) &&
3283          "X86 branch conditions have one component!");
3284   assert(!BytesAdded && "code size not handled");
3285 
3286   if (Cond.empty()) {
3287     // Unconditional branch?
3288     assert(!FBB && "Unconditional branch with multiple successors!");
3289     BuildMI(&MBB, DL, get(X86::JMP_1)).addMBB(TBB);
3290     return 1;
3291   }
3292 
3293   // If FBB is null, it is implied to be a fall-through block.
3294   bool FallThru = FBB == nullptr;
3295 
3296   // Conditional branch.
3297   unsigned Count = 0;
3298   X86::CondCode CC = (X86::CondCode)Cond[0].getImm();
3299   switch (CC) {
3300   case X86::COND_NE_OR_P:
3301     // Synthesize NE_OR_P with two branches.
3302     BuildMI(&MBB, DL, get(X86::JCC_1)).addMBB(TBB).addImm(X86::COND_NE);
3303     ++Count;
3304     BuildMI(&MBB, DL, get(X86::JCC_1)).addMBB(TBB).addImm(X86::COND_P);
3305     ++Count;
3306     break;
3307   case X86::COND_E_AND_NP:
3308     // Use the next block of MBB as FBB if it is null.
3309     if (FBB == nullptr) {
3310       FBB = getFallThroughMBB(&MBB, TBB);
3311       assert(FBB && "MBB cannot be the last block in function when the false "
3312                     "body is a fall-through.");
3313     }
3314     // Synthesize COND_E_AND_NP with two branches.
3315     BuildMI(&MBB, DL, get(X86::JCC_1)).addMBB(FBB).addImm(X86::COND_NE);
3316     ++Count;
3317     BuildMI(&MBB, DL, get(X86::JCC_1)).addMBB(TBB).addImm(X86::COND_NP);
3318     ++Count;
3319     break;
3320   default: {
3321     BuildMI(&MBB, DL, get(X86::JCC_1)).addMBB(TBB).addImm(CC);
3322     ++Count;
3323   }
3324   }
3325   if (!FallThru) {
3326     // Two-way Conditional branch. Insert the second branch.
3327     BuildMI(&MBB, DL, get(X86::JMP_1)).addMBB(FBB);
3328     ++Count;
3329   }
3330   return Count;
3331 }
3332 
3333 bool X86InstrInfo::canInsertSelect(const MachineBasicBlock &MBB,
3334                                    ArrayRef<MachineOperand> Cond,
3335                                    Register DstReg, Register TrueReg,
3336                                    Register FalseReg, int &CondCycles,
3337                                    int &TrueCycles, int &FalseCycles) const {
3338   // Not all subtargets have cmov instructions.
3339   if (!Subtarget.canUseCMOV())
3340     return false;
3341   if (Cond.size() != 1)
3342     return false;
3343   // We cannot do the composite conditions, at least not in SSA form.
3344   if ((X86::CondCode)Cond[0].getImm() > X86::LAST_VALID_COND)
3345     return false;
3346 
3347   // Check register classes.
3348   const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
3349   const TargetRegisterClass *RC =
3350     RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg));
3351   if (!RC)
3352     return false;
3353 
3354   // We have cmov instructions for 16, 32, and 64 bit general purpose registers.
3355   if (X86::GR16RegClass.hasSubClassEq(RC) ||
3356       X86::GR32RegClass.hasSubClassEq(RC) ||
3357       X86::GR64RegClass.hasSubClassEq(RC)) {
3358     // This latency applies to Pentium M, Merom, Wolfdale, Nehalem, and Sandy
3359     // Bridge. Probably Ivy Bridge as well.
3360     CondCycles = 2;
3361     TrueCycles = 2;
3362     FalseCycles = 2;
3363     return true;
3364   }
3365 
3366   // Can't do vectors.
3367   return false;
3368 }
3369 
3370 void X86InstrInfo::insertSelect(MachineBasicBlock &MBB,
3371                                 MachineBasicBlock::iterator I,
3372                                 const DebugLoc &DL, Register DstReg,
3373                                 ArrayRef<MachineOperand> Cond, Register TrueReg,
3374                                 Register FalseReg) const {
3375   MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
3376   const TargetRegisterInfo &TRI = *MRI.getTargetRegisterInfo();
3377   const TargetRegisterClass &RC = *MRI.getRegClass(DstReg);
3378   assert(Cond.size() == 1 && "Invalid Cond array");
3379   unsigned Opc = X86::getCMovOpcode(TRI.getRegSizeInBits(RC) / 8,
3380                                     false /*HasMemoryOperand*/);
3381   BuildMI(MBB, I, DL, get(Opc), DstReg)
3382       .addReg(FalseReg)
3383       .addReg(TrueReg)
3384       .addImm(Cond[0].getImm());
3385 }
3386 
3387 /// Test if the given register is a physical h register.
3388 static bool isHReg(unsigned Reg) {
3389   return X86::GR8_ABCD_HRegClass.contains(Reg);
3390 }
3391 
3392 // Try and copy between VR128/VR64 and GR64 registers.
3393 static unsigned CopyToFromAsymmetricReg(unsigned DestReg, unsigned SrcReg,
3394                                         const X86Subtarget &Subtarget) {
3395   bool HasAVX = Subtarget.hasAVX();
3396   bool HasAVX512 = Subtarget.hasAVX512();
3397 
3398   // SrcReg(MaskReg) -> DestReg(GR64)
3399   // SrcReg(MaskReg) -> DestReg(GR32)
3400 
3401   // All KMASK RegClasses hold the same k registers, can be tested against anyone.
3402   if (X86::VK16RegClass.contains(SrcReg)) {
3403     if (X86::GR64RegClass.contains(DestReg)) {
3404       assert(Subtarget.hasBWI());
3405       return X86::KMOVQrk;
3406     }
3407     if (X86::GR32RegClass.contains(DestReg))
3408       return Subtarget.hasBWI() ? X86::KMOVDrk : X86::KMOVWrk;
3409   }
3410 
3411   // SrcReg(GR64) -> DestReg(MaskReg)
3412   // SrcReg(GR32) -> DestReg(MaskReg)
3413 
3414   // All KMASK RegClasses hold the same k registers, can be tested against anyone.
3415   if (X86::VK16RegClass.contains(DestReg)) {
3416     if (X86::GR64RegClass.contains(SrcReg)) {
3417       assert(Subtarget.hasBWI());
3418       return X86::KMOVQkr;
3419     }
3420     if (X86::GR32RegClass.contains(SrcReg))
3421       return Subtarget.hasBWI() ? X86::KMOVDkr : X86::KMOVWkr;
3422   }
3423 
3424 
3425   // SrcReg(VR128) -> DestReg(GR64)
3426   // SrcReg(VR64)  -> DestReg(GR64)
3427   // SrcReg(GR64)  -> DestReg(VR128)
3428   // SrcReg(GR64)  -> DestReg(VR64)
3429 
3430   if (X86::GR64RegClass.contains(DestReg)) {
3431     if (X86::VR128XRegClass.contains(SrcReg))
3432       // Copy from a VR128 register to a GR64 register.
3433       return HasAVX512 ? X86::VMOVPQIto64Zrr :
3434              HasAVX    ? X86::VMOVPQIto64rr  :
3435                          X86::MOVPQIto64rr;
3436     if (X86::VR64RegClass.contains(SrcReg))
3437       // Copy from a VR64 register to a GR64 register.
3438       return X86::MMX_MOVD64from64rr;
3439   } else if (X86::GR64RegClass.contains(SrcReg)) {
3440     // Copy from a GR64 register to a VR128 register.
3441     if (X86::VR128XRegClass.contains(DestReg))
3442       return HasAVX512 ? X86::VMOV64toPQIZrr :
3443              HasAVX    ? X86::VMOV64toPQIrr  :
3444                          X86::MOV64toPQIrr;
3445     // Copy from a GR64 register to a VR64 register.
3446     if (X86::VR64RegClass.contains(DestReg))
3447       return X86::MMX_MOVD64to64rr;
3448   }
3449 
3450   // SrcReg(VR128) -> DestReg(GR32)
3451   // SrcReg(GR32)  -> DestReg(VR128)
3452 
3453   if (X86::GR32RegClass.contains(DestReg) &&
3454       X86::VR128XRegClass.contains(SrcReg))
3455     // Copy from a VR128 register to a GR32 register.
3456     return HasAVX512 ? X86::VMOVPDI2DIZrr :
3457            HasAVX    ? X86::VMOVPDI2DIrr  :
3458                        X86::MOVPDI2DIrr;
3459 
3460   if (X86::VR128XRegClass.contains(DestReg) &&
3461       X86::GR32RegClass.contains(SrcReg))
3462     // Copy from a VR128 register to a VR128 register.
3463     return HasAVX512 ? X86::VMOVDI2PDIZrr :
3464            HasAVX    ? X86::VMOVDI2PDIrr  :
3465                        X86::MOVDI2PDIrr;
3466   return 0;
3467 }
3468 
3469 void X86InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
3470                                MachineBasicBlock::iterator MI,
3471                                const DebugLoc &DL, MCRegister DestReg,
3472                                MCRegister SrcReg, bool KillSrc) const {
3473   // First deal with the normal symmetric copies.
3474   bool HasAVX = Subtarget.hasAVX();
3475   bool HasVLX = Subtarget.hasVLX();
3476   unsigned Opc = 0;
3477   if (X86::GR64RegClass.contains(DestReg, SrcReg))
3478     Opc = X86::MOV64rr;
3479   else if (X86::GR32RegClass.contains(DestReg, SrcReg))
3480     Opc = X86::MOV32rr;
3481   else if (X86::GR16RegClass.contains(DestReg, SrcReg))
3482     Opc = X86::MOV16rr;
3483   else if (X86::GR8RegClass.contains(DestReg, SrcReg)) {
3484     // Copying to or from a physical H register on x86-64 requires a NOREX
3485     // move.  Otherwise use a normal move.
3486     if ((isHReg(DestReg) || isHReg(SrcReg)) &&
3487         Subtarget.is64Bit()) {
3488       Opc = X86::MOV8rr_NOREX;
3489       // Both operands must be encodable without an REX prefix.
3490       assert(X86::GR8_NOREXRegClass.contains(SrcReg, DestReg) &&
3491              "8-bit H register can not be copied outside GR8_NOREX");
3492     } else
3493       Opc = X86::MOV8rr;
3494   }
3495   else if (X86::VR64RegClass.contains(DestReg, SrcReg))
3496     Opc = X86::MMX_MOVQ64rr;
3497   else if (X86::VR128XRegClass.contains(DestReg, SrcReg)) {
3498     if (HasVLX)
3499       Opc = X86::VMOVAPSZ128rr;
3500     else if (X86::VR128RegClass.contains(DestReg, SrcReg))
3501       Opc = HasAVX ? X86::VMOVAPSrr : X86::MOVAPSrr;
3502     else {
3503       // If this an extended register and we don't have VLX we need to use a
3504       // 512-bit move.
3505       Opc = X86::VMOVAPSZrr;
3506       const TargetRegisterInfo *TRI = &getRegisterInfo();
3507       DestReg = TRI->getMatchingSuperReg(DestReg, X86::sub_xmm,
3508                                          &X86::VR512RegClass);
3509       SrcReg = TRI->getMatchingSuperReg(SrcReg, X86::sub_xmm,
3510                                         &X86::VR512RegClass);
3511     }
3512   } else if (X86::VR256XRegClass.contains(DestReg, SrcReg)) {
3513     if (HasVLX)
3514       Opc = X86::VMOVAPSZ256rr;
3515     else if (X86::VR256RegClass.contains(DestReg, SrcReg))
3516       Opc = X86::VMOVAPSYrr;
3517     else {
3518       // If this an extended register and we don't have VLX we need to use a
3519       // 512-bit move.
3520       Opc = X86::VMOVAPSZrr;
3521       const TargetRegisterInfo *TRI = &getRegisterInfo();
3522       DestReg = TRI->getMatchingSuperReg(DestReg, X86::sub_ymm,
3523                                          &X86::VR512RegClass);
3524       SrcReg = TRI->getMatchingSuperReg(SrcReg, X86::sub_ymm,
3525                                         &X86::VR512RegClass);
3526     }
3527   } else if (X86::VR512RegClass.contains(DestReg, SrcReg))
3528     Opc = X86::VMOVAPSZrr;
3529   // All KMASK RegClasses hold the same k registers, can be tested against anyone.
3530   else if (X86::VK16RegClass.contains(DestReg, SrcReg))
3531     Opc = Subtarget.hasBWI() ? X86::KMOVQkk : X86::KMOVWkk;
3532   if (!Opc)
3533     Opc = CopyToFromAsymmetricReg(DestReg, SrcReg, Subtarget);
3534 
3535   if (Opc) {
3536     BuildMI(MBB, MI, DL, get(Opc), DestReg)
3537       .addReg(SrcReg, getKillRegState(KillSrc));
3538     return;
3539   }
3540 
3541   if (SrcReg == X86::EFLAGS || DestReg == X86::EFLAGS) {
3542     // FIXME: We use a fatal error here because historically LLVM has tried
3543     // lower some of these physreg copies and we want to ensure we get
3544     // reasonable bug reports if someone encounters a case no other testing
3545     // found. This path should be removed after the LLVM 7 release.
3546     report_fatal_error("Unable to copy EFLAGS physical register!");
3547   }
3548 
3549   LLVM_DEBUG(dbgs() << "Cannot copy " << RI.getName(SrcReg) << " to "
3550                     << RI.getName(DestReg) << '\n');
3551   report_fatal_error("Cannot emit physreg copy instruction");
3552 }
3553 
3554 Optional<DestSourcePair>
3555 X86InstrInfo::isCopyInstrImpl(const MachineInstr &MI) const {
3556   if (MI.isMoveReg())
3557     return DestSourcePair{MI.getOperand(0), MI.getOperand(1)};
3558   return None;
3559 }
3560 
3561 static unsigned getLoadStoreRegOpcode(Register Reg,
3562                                       const TargetRegisterClass *RC,
3563                                       bool IsStackAligned,
3564                                       const X86Subtarget &STI, bool load) {
3565   bool HasAVX = STI.hasAVX();
3566   bool HasAVX512 = STI.hasAVX512();
3567   bool HasVLX = STI.hasVLX();
3568 
3569   switch (STI.getRegisterInfo()->getSpillSize(*RC)) {
3570   default:
3571     llvm_unreachable("Unknown spill size");
3572   case 1:
3573     assert(X86::GR8RegClass.hasSubClassEq(RC) && "Unknown 1-byte regclass");
3574     if (STI.is64Bit())
3575       // Copying to or from a physical H register on x86-64 requires a NOREX
3576       // move.  Otherwise use a normal move.
3577       if (isHReg(Reg) || X86::GR8_ABCD_HRegClass.hasSubClassEq(RC))
3578         return load ? X86::MOV8rm_NOREX : X86::MOV8mr_NOREX;
3579     return load ? X86::MOV8rm : X86::MOV8mr;
3580   case 2:
3581     if (X86::VK16RegClass.hasSubClassEq(RC))
3582       return load ? X86::KMOVWkm : X86::KMOVWmk;
3583     if (X86::FR16XRegClass.hasSubClassEq(RC)) {
3584       assert(STI.hasFP16());
3585       return load ? X86::VMOVSHZrm_alt : X86::VMOVSHZmr;
3586     }
3587     assert(X86::GR16RegClass.hasSubClassEq(RC) && "Unknown 2-byte regclass");
3588     return load ? X86::MOV16rm : X86::MOV16mr;
3589   case 4:
3590     if (X86::GR32RegClass.hasSubClassEq(RC))
3591       return load ? X86::MOV32rm : X86::MOV32mr;
3592     if (X86::FR32XRegClass.hasSubClassEq(RC))
3593       return load ?
3594         (HasAVX512 ? X86::VMOVSSZrm_alt :
3595          HasAVX    ? X86::VMOVSSrm_alt :
3596                      X86::MOVSSrm_alt) :
3597         (HasAVX512 ? X86::VMOVSSZmr :
3598          HasAVX    ? X86::VMOVSSmr :
3599                      X86::MOVSSmr);
3600     if (X86::RFP32RegClass.hasSubClassEq(RC))
3601       return load ? X86::LD_Fp32m : X86::ST_Fp32m;
3602     if (X86::VK32RegClass.hasSubClassEq(RC)) {
3603       assert(STI.hasBWI() && "KMOVD requires BWI");
3604       return load ? X86::KMOVDkm : X86::KMOVDmk;
3605     }
3606     // All of these mask pair classes have the same spill size, the same kind
3607     // of kmov instructions can be used with all of them.
3608     if (X86::VK1PAIRRegClass.hasSubClassEq(RC) ||
3609         X86::VK2PAIRRegClass.hasSubClassEq(RC) ||
3610         X86::VK4PAIRRegClass.hasSubClassEq(RC) ||
3611         X86::VK8PAIRRegClass.hasSubClassEq(RC) ||
3612         X86::VK16PAIRRegClass.hasSubClassEq(RC))
3613       return load ? X86::MASKPAIR16LOAD : X86::MASKPAIR16STORE;
3614     llvm_unreachable("Unknown 4-byte regclass");
3615   case 8:
3616     if (X86::GR64RegClass.hasSubClassEq(RC))
3617       return load ? X86::MOV64rm : X86::MOV64mr;
3618     if (X86::FR64XRegClass.hasSubClassEq(RC))
3619       return load ?
3620         (HasAVX512 ? X86::VMOVSDZrm_alt :
3621          HasAVX    ? X86::VMOVSDrm_alt :
3622                      X86::MOVSDrm_alt) :
3623         (HasAVX512 ? X86::VMOVSDZmr :
3624          HasAVX    ? X86::VMOVSDmr :
3625                      X86::MOVSDmr);
3626     if (X86::VR64RegClass.hasSubClassEq(RC))
3627       return load ? X86::MMX_MOVQ64rm : X86::MMX_MOVQ64mr;
3628     if (X86::RFP64RegClass.hasSubClassEq(RC))
3629       return load ? X86::LD_Fp64m : X86::ST_Fp64m;
3630     if (X86::VK64RegClass.hasSubClassEq(RC)) {
3631       assert(STI.hasBWI() && "KMOVQ requires BWI");
3632       return load ? X86::KMOVQkm : X86::KMOVQmk;
3633     }
3634     llvm_unreachable("Unknown 8-byte regclass");
3635   case 10:
3636     assert(X86::RFP80RegClass.hasSubClassEq(RC) && "Unknown 10-byte regclass");
3637     return load ? X86::LD_Fp80m : X86::ST_FpP80m;
3638   case 16: {
3639     if (X86::VR128XRegClass.hasSubClassEq(RC)) {
3640       // If stack is realigned we can use aligned stores.
3641       if (IsStackAligned)
3642         return load ?
3643           (HasVLX    ? X86::VMOVAPSZ128rm :
3644            HasAVX512 ? X86::VMOVAPSZ128rm_NOVLX :
3645            HasAVX    ? X86::VMOVAPSrm :
3646                        X86::MOVAPSrm):
3647           (HasVLX    ? X86::VMOVAPSZ128mr :
3648            HasAVX512 ? X86::VMOVAPSZ128mr_NOVLX :
3649            HasAVX    ? X86::VMOVAPSmr :
3650                        X86::MOVAPSmr);
3651       else
3652         return load ?
3653           (HasVLX    ? X86::VMOVUPSZ128rm :
3654            HasAVX512 ? X86::VMOVUPSZ128rm_NOVLX :
3655            HasAVX    ? X86::VMOVUPSrm :
3656                        X86::MOVUPSrm):
3657           (HasVLX    ? X86::VMOVUPSZ128mr :
3658            HasAVX512 ? X86::VMOVUPSZ128mr_NOVLX :
3659            HasAVX    ? X86::VMOVUPSmr :
3660                        X86::MOVUPSmr);
3661     }
3662     llvm_unreachable("Unknown 16-byte regclass");
3663   }
3664   case 32:
3665     assert(X86::VR256XRegClass.hasSubClassEq(RC) && "Unknown 32-byte regclass");
3666     // If stack is realigned we can use aligned stores.
3667     if (IsStackAligned)
3668       return load ?
3669         (HasVLX    ? X86::VMOVAPSZ256rm :
3670          HasAVX512 ? X86::VMOVAPSZ256rm_NOVLX :
3671                      X86::VMOVAPSYrm) :
3672         (HasVLX    ? X86::VMOVAPSZ256mr :
3673          HasAVX512 ? X86::VMOVAPSZ256mr_NOVLX :
3674                      X86::VMOVAPSYmr);
3675     else
3676       return load ?
3677         (HasVLX    ? X86::VMOVUPSZ256rm :
3678          HasAVX512 ? X86::VMOVUPSZ256rm_NOVLX :
3679                      X86::VMOVUPSYrm) :
3680         (HasVLX    ? X86::VMOVUPSZ256mr :
3681          HasAVX512 ? X86::VMOVUPSZ256mr_NOVLX :
3682                      X86::VMOVUPSYmr);
3683   case 64:
3684     assert(X86::VR512RegClass.hasSubClassEq(RC) && "Unknown 64-byte regclass");
3685     assert(STI.hasAVX512() && "Using 512-bit register requires AVX512");
3686     if (IsStackAligned)
3687       return load ? X86::VMOVAPSZrm : X86::VMOVAPSZmr;
3688     else
3689       return load ? X86::VMOVUPSZrm : X86::VMOVUPSZmr;
3690   }
3691 }
3692 
3693 Optional<ExtAddrMode>
3694 X86InstrInfo::getAddrModeFromMemoryOp(const MachineInstr &MemI,
3695                                       const TargetRegisterInfo *TRI) const {
3696   const MCInstrDesc &Desc = MemI.getDesc();
3697   int MemRefBegin = X86II::getMemoryOperandNo(Desc.TSFlags);
3698   if (MemRefBegin < 0)
3699     return None;
3700 
3701   MemRefBegin += X86II::getOperandBias(Desc);
3702 
3703   auto &BaseOp = MemI.getOperand(MemRefBegin + X86::AddrBaseReg);
3704   if (!BaseOp.isReg()) // Can be an MO_FrameIndex
3705     return None;
3706 
3707   const MachineOperand &DispMO = MemI.getOperand(MemRefBegin + X86::AddrDisp);
3708   // Displacement can be symbolic
3709   if (!DispMO.isImm())
3710     return None;
3711 
3712   ExtAddrMode AM;
3713   AM.BaseReg = BaseOp.getReg();
3714   AM.ScaledReg = MemI.getOperand(MemRefBegin + X86::AddrIndexReg).getReg();
3715   AM.Scale = MemI.getOperand(MemRefBegin + X86::AddrScaleAmt).getImm();
3716   AM.Displacement = DispMO.getImm();
3717   return AM;
3718 }
3719 
3720 bool X86InstrInfo::verifyInstruction(const MachineInstr &MI,
3721                                      StringRef &ErrInfo) const {
3722   Optional<ExtAddrMode> AMOrNone = getAddrModeFromMemoryOp(MI, nullptr);
3723   if (!AMOrNone)
3724     return true;
3725 
3726   ExtAddrMode AM = *AMOrNone;
3727 
3728   if (AM.ScaledReg != X86::NoRegister) {
3729     switch (AM.Scale) {
3730     case 1:
3731     case 2:
3732     case 4:
3733     case 8:
3734       break;
3735     default:
3736       ErrInfo = "Scale factor in address must be 1, 2, 4 or 8";
3737       return false;
3738     }
3739   }
3740   if (!isInt<32>(AM.Displacement)) {
3741     ErrInfo = "Displacement in address must fit into 32-bit signed "
3742               "integer";
3743     return false;
3744   }
3745 
3746   return true;
3747 }
3748 
3749 bool X86InstrInfo::getConstValDefinedInReg(const MachineInstr &MI,
3750                                            const Register Reg,
3751                                            int64_t &ImmVal) const {
3752   if (MI.getOpcode() != X86::MOV32ri && MI.getOpcode() != X86::MOV64ri)
3753     return false;
3754   // Mov Src can be a global address.
3755   if (!MI.getOperand(1).isImm() || MI.getOperand(0).getReg() != Reg)
3756     return false;
3757   ImmVal = MI.getOperand(1).getImm();
3758   return true;
3759 }
3760 
3761 bool X86InstrInfo::preservesZeroValueInReg(
3762     const MachineInstr *MI, const Register NullValueReg,
3763     const TargetRegisterInfo *TRI) const {
3764   if (!MI->modifiesRegister(NullValueReg, TRI))
3765     return true;
3766   switch (MI->getOpcode()) {
3767   // Shift right/left of a null unto itself is still a null, i.e. rax = shl rax
3768   // X.
3769   case X86::SHR64ri:
3770   case X86::SHR32ri:
3771   case X86::SHL64ri:
3772   case X86::SHL32ri:
3773     assert(MI->getOperand(0).isDef() && MI->getOperand(1).isUse() &&
3774            "expected for shift opcode!");
3775     return MI->getOperand(0).getReg() == NullValueReg &&
3776            MI->getOperand(1).getReg() == NullValueReg;
3777   // Zero extend of a sub-reg of NullValueReg into itself does not change the
3778   // null value.
3779   case X86::MOV32rr:
3780     return llvm::all_of(MI->operands(), [&](const MachineOperand &MO) {
3781       return TRI->isSubRegisterEq(NullValueReg, MO.getReg());
3782     });
3783   default:
3784     return false;
3785   }
3786   llvm_unreachable("Should be handled above!");
3787 }
3788 
3789 bool X86InstrInfo::getMemOperandsWithOffsetWidth(
3790     const MachineInstr &MemOp, SmallVectorImpl<const MachineOperand *> &BaseOps,
3791     int64_t &Offset, bool &OffsetIsScalable, unsigned &Width,
3792     const TargetRegisterInfo *TRI) const {
3793   const MCInstrDesc &Desc = MemOp.getDesc();
3794   int MemRefBegin = X86II::getMemoryOperandNo(Desc.TSFlags);
3795   if (MemRefBegin < 0)
3796     return false;
3797 
3798   MemRefBegin += X86II::getOperandBias(Desc);
3799 
3800   const MachineOperand *BaseOp =
3801       &MemOp.getOperand(MemRefBegin + X86::AddrBaseReg);
3802   if (!BaseOp->isReg()) // Can be an MO_FrameIndex
3803     return false;
3804 
3805   if (MemOp.getOperand(MemRefBegin + X86::AddrScaleAmt).getImm() != 1)
3806     return false;
3807 
3808   if (MemOp.getOperand(MemRefBegin + X86::AddrIndexReg).getReg() !=
3809       X86::NoRegister)
3810     return false;
3811 
3812   const MachineOperand &DispMO = MemOp.getOperand(MemRefBegin + X86::AddrDisp);
3813 
3814   // Displacement can be symbolic
3815   if (!DispMO.isImm())
3816     return false;
3817 
3818   Offset = DispMO.getImm();
3819 
3820   if (!BaseOp->isReg())
3821     return false;
3822 
3823   OffsetIsScalable = false;
3824   // FIXME: Relying on memoperands() may not be right thing to do here. Check
3825   // with X86 maintainers, and fix it accordingly. For now, it is ok, since
3826   // there is no use of `Width` for X86 back-end at the moment.
3827   Width =
3828       !MemOp.memoperands_empty() ? MemOp.memoperands().front()->getSize() : 0;
3829   BaseOps.push_back(BaseOp);
3830   return true;
3831 }
3832 
3833 static unsigned getStoreRegOpcode(Register SrcReg,
3834                                   const TargetRegisterClass *RC,
3835                                   bool IsStackAligned,
3836                                   const X86Subtarget &STI) {
3837   return getLoadStoreRegOpcode(SrcReg, RC, IsStackAligned, STI, false);
3838 }
3839 
3840 static unsigned getLoadRegOpcode(Register DestReg,
3841                                  const TargetRegisterClass *RC,
3842                                  bool IsStackAligned, const X86Subtarget &STI) {
3843   return getLoadStoreRegOpcode(DestReg, RC, IsStackAligned, STI, true);
3844 }
3845 
3846 void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
3847                                        MachineBasicBlock::iterator MI,
3848                                        Register SrcReg, bool isKill, int FrameIdx,
3849                                        const TargetRegisterClass *RC,
3850                                        const TargetRegisterInfo *TRI) const {
3851   const MachineFunction &MF = *MBB.getParent();
3852   const MachineFrameInfo &MFI = MF.getFrameInfo();
3853   assert(MFI.getObjectSize(FrameIdx) >= TRI->getSpillSize(*RC) &&
3854          "Stack slot too small for store");
3855   if (RC->getID() == X86::TILERegClassID) {
3856     unsigned Opc = X86::TILESTORED;
3857     // tilestored %tmm, (%sp, %idx)
3858     MachineRegisterInfo &RegInfo = MBB.getParent()->getRegInfo();
3859     Register VirtReg = RegInfo.createVirtualRegister(&X86::GR64_NOSPRegClass);
3860     BuildMI(MBB, MI, DebugLoc(), get(X86::MOV64ri), VirtReg).addImm(64);
3861     MachineInstr *NewMI =
3862         addFrameReference(BuildMI(MBB, MI, DebugLoc(), get(Opc)), FrameIdx)
3863             .addReg(SrcReg, getKillRegState(isKill));
3864     MachineOperand &MO = NewMI->getOperand(2);
3865     MO.setReg(VirtReg);
3866     MO.setIsKill(true);
3867   } else {
3868     unsigned Alignment = std::max<uint32_t>(TRI->getSpillSize(*RC), 16);
3869     bool isAligned =
3870         (Subtarget.getFrameLowering()->getStackAlign() >= Alignment) ||
3871         (RI.canRealignStack(MF) && !MFI.isFixedObjectIndex(FrameIdx));
3872     unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, Subtarget);
3873     addFrameReference(BuildMI(MBB, MI, DebugLoc(), get(Opc)), FrameIdx)
3874         .addReg(SrcReg, getKillRegState(isKill));
3875   }
3876 }
3877 
3878 void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
3879                                         MachineBasicBlock::iterator MI,
3880                                         Register DestReg, int FrameIdx,
3881                                         const TargetRegisterClass *RC,
3882                                         const TargetRegisterInfo *TRI) const {
3883   if (RC->getID() == X86::TILERegClassID) {
3884     unsigned Opc = X86::TILELOADD;
3885     // tileloadd (%sp, %idx), %tmm
3886     MachineRegisterInfo &RegInfo = MBB.getParent()->getRegInfo();
3887     Register VirtReg = RegInfo.createVirtualRegister(&X86::GR64_NOSPRegClass);
3888     MachineInstr *NewMI =
3889         BuildMI(MBB, MI, DebugLoc(), get(X86::MOV64ri), VirtReg).addImm(64);
3890     NewMI = addFrameReference(BuildMI(MBB, MI, DebugLoc(), get(Opc), DestReg),
3891                               FrameIdx);
3892     MachineOperand &MO = NewMI->getOperand(3);
3893     MO.setReg(VirtReg);
3894     MO.setIsKill(true);
3895   } else {
3896     const MachineFunction &MF = *MBB.getParent();
3897     const MachineFrameInfo &MFI = MF.getFrameInfo();
3898     unsigned Alignment = std::max<uint32_t>(TRI->getSpillSize(*RC), 16);
3899     bool isAligned =
3900         (Subtarget.getFrameLowering()->getStackAlign() >= Alignment) ||
3901         (RI.canRealignStack(MF) && !MFI.isFixedObjectIndex(FrameIdx));
3902     unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, Subtarget);
3903     addFrameReference(BuildMI(MBB, MI, DebugLoc(), get(Opc), DestReg),
3904                       FrameIdx);
3905   }
3906 }
3907 
3908 bool X86InstrInfo::analyzeCompare(const MachineInstr &MI, Register &SrcReg,
3909                                   Register &SrcReg2, int64_t &CmpMask,
3910                                   int64_t &CmpValue) const {
3911   switch (MI.getOpcode()) {
3912   default: break;
3913   case X86::CMP64ri32:
3914   case X86::CMP64ri8:
3915   case X86::CMP32ri:
3916   case X86::CMP32ri8:
3917   case X86::CMP16ri:
3918   case X86::CMP16ri8:
3919   case X86::CMP8ri:
3920     SrcReg = MI.getOperand(0).getReg();
3921     SrcReg2 = 0;
3922     if (MI.getOperand(1).isImm()) {
3923       CmpMask = ~0;
3924       CmpValue = MI.getOperand(1).getImm();
3925     } else {
3926       CmpMask = CmpValue = 0;
3927     }
3928     return true;
3929   // A SUB can be used to perform comparison.
3930   case X86::SUB64rm:
3931   case X86::SUB32rm:
3932   case X86::SUB16rm:
3933   case X86::SUB8rm:
3934     SrcReg = MI.getOperand(1).getReg();
3935     SrcReg2 = 0;
3936     CmpMask = 0;
3937     CmpValue = 0;
3938     return true;
3939   case X86::SUB64rr:
3940   case X86::SUB32rr:
3941   case X86::SUB16rr:
3942   case X86::SUB8rr:
3943     SrcReg = MI.getOperand(1).getReg();
3944     SrcReg2 = MI.getOperand(2).getReg();
3945     CmpMask = 0;
3946     CmpValue = 0;
3947     return true;
3948   case X86::SUB64ri32:
3949   case X86::SUB64ri8:
3950   case X86::SUB32ri:
3951   case X86::SUB32ri8:
3952   case X86::SUB16ri:
3953   case X86::SUB16ri8:
3954   case X86::SUB8ri:
3955     SrcReg = MI.getOperand(1).getReg();
3956     SrcReg2 = 0;
3957     if (MI.getOperand(2).isImm()) {
3958       CmpMask = ~0;
3959       CmpValue = MI.getOperand(2).getImm();
3960     } else {
3961       CmpMask = CmpValue = 0;
3962     }
3963     return true;
3964   case X86::CMP64rr:
3965   case X86::CMP32rr:
3966   case X86::CMP16rr:
3967   case X86::CMP8rr:
3968     SrcReg = MI.getOperand(0).getReg();
3969     SrcReg2 = MI.getOperand(1).getReg();
3970     CmpMask = 0;
3971     CmpValue = 0;
3972     return true;
3973   case X86::TEST8rr:
3974   case X86::TEST16rr:
3975   case X86::TEST32rr:
3976   case X86::TEST64rr:
3977     SrcReg = MI.getOperand(0).getReg();
3978     if (MI.getOperand(1).getReg() != SrcReg)
3979       return false;
3980     // Compare against zero.
3981     SrcReg2 = 0;
3982     CmpMask = ~0;
3983     CmpValue = 0;
3984     return true;
3985   }
3986   return false;
3987 }
3988 
3989 bool X86InstrInfo::isRedundantFlagInstr(const MachineInstr &FlagI,
3990                                         Register SrcReg, Register SrcReg2,
3991                                         int64_t ImmMask, int64_t ImmValue,
3992                                         const MachineInstr &OI, bool *IsSwapped,
3993                                         int64_t *ImmDelta) const {
3994   switch (OI.getOpcode()) {
3995   case X86::CMP64rr:
3996   case X86::CMP32rr:
3997   case X86::CMP16rr:
3998   case X86::CMP8rr:
3999   case X86::SUB64rr:
4000   case X86::SUB32rr:
4001   case X86::SUB16rr:
4002   case X86::SUB8rr: {
4003     Register OISrcReg;
4004     Register OISrcReg2;
4005     int64_t OIMask;
4006     int64_t OIValue;
4007     if (!analyzeCompare(OI, OISrcReg, OISrcReg2, OIMask, OIValue) ||
4008         OIMask != ImmMask || OIValue != ImmValue)
4009       return false;
4010     if (SrcReg == OISrcReg && SrcReg2 == OISrcReg2) {
4011       *IsSwapped = false;
4012       return true;
4013     }
4014     if (SrcReg == OISrcReg2 && SrcReg2 == OISrcReg) {
4015       *IsSwapped = true;
4016       return true;
4017     }
4018     return false;
4019   }
4020   case X86::CMP64ri32:
4021   case X86::CMP64ri8:
4022   case X86::CMP32ri:
4023   case X86::CMP32ri8:
4024   case X86::CMP16ri:
4025   case X86::CMP16ri8:
4026   case X86::CMP8ri:
4027   case X86::SUB64ri32:
4028   case X86::SUB64ri8:
4029   case X86::SUB32ri:
4030   case X86::SUB32ri8:
4031   case X86::SUB16ri:
4032   case X86::SUB16ri8:
4033   case X86::SUB8ri:
4034   case X86::TEST64rr:
4035   case X86::TEST32rr:
4036   case X86::TEST16rr:
4037   case X86::TEST8rr: {
4038     if (ImmMask != 0) {
4039       Register OISrcReg;
4040       Register OISrcReg2;
4041       int64_t OIMask;
4042       int64_t OIValue;
4043       if (analyzeCompare(OI, OISrcReg, OISrcReg2, OIMask, OIValue) &&
4044           SrcReg == OISrcReg && ImmMask == OIMask) {
4045         if (OIValue == ImmValue) {
4046           *ImmDelta = 0;
4047           return true;
4048         } else if (static_cast<uint64_t>(ImmValue) ==
4049                    static_cast<uint64_t>(OIValue) - 1) {
4050           *ImmDelta = -1;
4051           return true;
4052         } else if (static_cast<uint64_t>(ImmValue) ==
4053                    static_cast<uint64_t>(OIValue) + 1) {
4054           *ImmDelta = 1;
4055           return true;
4056         } else {
4057           return false;
4058         }
4059       }
4060     }
4061     return FlagI.isIdenticalTo(OI);
4062   }
4063   default:
4064     return false;
4065   }
4066 }
4067 
4068 /// Check whether the definition can be converted
4069 /// to remove a comparison against zero.
4070 inline static bool isDefConvertible(const MachineInstr &MI, bool &NoSignFlag,
4071                                     bool &ClearsOverflowFlag) {
4072   NoSignFlag = false;
4073   ClearsOverflowFlag = false;
4074 
4075   switch (MI.getOpcode()) {
4076   default: return false;
4077 
4078   // The shift instructions only modify ZF if their shift count is non-zero.
4079   // N.B.: The processor truncates the shift count depending on the encoding.
4080   case X86::SAR8ri:    case X86::SAR16ri:  case X86::SAR32ri:case X86::SAR64ri:
4081   case X86::SHR8ri:    case X86::SHR16ri:  case X86::SHR32ri:case X86::SHR64ri:
4082      return getTruncatedShiftCount(MI, 2) != 0;
4083 
4084   // Some left shift instructions can be turned into LEA instructions but only
4085   // if their flags aren't used. Avoid transforming such instructions.
4086   case X86::SHL8ri:    case X86::SHL16ri:  case X86::SHL32ri:case X86::SHL64ri:{
4087     unsigned ShAmt = getTruncatedShiftCount(MI, 2);
4088     if (isTruncatedShiftCountForLEA(ShAmt)) return false;
4089     return ShAmt != 0;
4090   }
4091 
4092   case X86::SHRD16rri8:case X86::SHRD32rri8:case X86::SHRD64rri8:
4093   case X86::SHLD16rri8:case X86::SHLD32rri8:case X86::SHLD64rri8:
4094      return getTruncatedShiftCount(MI, 3) != 0;
4095 
4096   case X86::SUB64ri32: case X86::SUB64ri8: case X86::SUB32ri:
4097   case X86::SUB32ri8:  case X86::SUB16ri:  case X86::SUB16ri8:
4098   case X86::SUB8ri:    case X86::SUB64rr:  case X86::SUB32rr:
4099   case X86::SUB16rr:   case X86::SUB8rr:   case X86::SUB64rm:
4100   case X86::SUB32rm:   case X86::SUB16rm:  case X86::SUB8rm:
4101   case X86::DEC64r:    case X86::DEC32r:   case X86::DEC16r: case X86::DEC8r:
4102   case X86::ADD64ri32: case X86::ADD64ri8: case X86::ADD32ri:
4103   case X86::ADD32ri8:  case X86::ADD16ri:  case X86::ADD16ri8:
4104   case X86::ADD8ri:    case X86::ADD64rr:  case X86::ADD32rr:
4105   case X86::ADD16rr:   case X86::ADD8rr:   case X86::ADD64rm:
4106   case X86::ADD32rm:   case X86::ADD16rm:  case X86::ADD8rm:
4107   case X86::INC64r:    case X86::INC32r:   case X86::INC16r: case X86::INC8r:
4108   case X86::ADC64ri32: case X86::ADC64ri8: case X86::ADC32ri:
4109   case X86::ADC32ri8:  case X86::ADC16ri:  case X86::ADC16ri8:
4110   case X86::ADC8ri:    case X86::ADC64rr:  case X86::ADC32rr:
4111   case X86::ADC16rr:   case X86::ADC8rr:   case X86::ADC64rm:
4112   case X86::ADC32rm:   case X86::ADC16rm:  case X86::ADC8rm:
4113   case X86::SBB64ri32: case X86::SBB64ri8: case X86::SBB32ri:
4114   case X86::SBB32ri8:  case X86::SBB16ri:  case X86::SBB16ri8:
4115   case X86::SBB8ri:    case X86::SBB64rr:  case X86::SBB32rr:
4116   case X86::SBB16rr:   case X86::SBB8rr:   case X86::SBB64rm:
4117   case X86::SBB32rm:   case X86::SBB16rm:  case X86::SBB8rm:
4118   case X86::NEG8r:     case X86::NEG16r:   case X86::NEG32r: case X86::NEG64r:
4119   case X86::SAR8r1:    case X86::SAR16r1:  case X86::SAR32r1:case X86::SAR64r1:
4120   case X86::SHR8r1:    case X86::SHR16r1:  case X86::SHR32r1:case X86::SHR64r1:
4121   case X86::SHL8r1:    case X86::SHL16r1:  case X86::SHL32r1:case X86::SHL64r1:
4122   case X86::LZCNT16rr: case X86::LZCNT16rm:
4123   case X86::LZCNT32rr: case X86::LZCNT32rm:
4124   case X86::LZCNT64rr: case X86::LZCNT64rm:
4125   case X86::POPCNT16rr:case X86::POPCNT16rm:
4126   case X86::POPCNT32rr:case X86::POPCNT32rm:
4127   case X86::POPCNT64rr:case X86::POPCNT64rm:
4128   case X86::TZCNT16rr: case X86::TZCNT16rm:
4129   case X86::TZCNT32rr: case X86::TZCNT32rm:
4130   case X86::TZCNT64rr: case X86::TZCNT64rm:
4131     return true;
4132   case X86::AND64ri32:   case X86::AND64ri8:  case X86::AND32ri:
4133   case X86::AND32ri8:    case X86::AND16ri:   case X86::AND16ri8:
4134   case X86::AND8ri:      case X86::AND64rr:   case X86::AND32rr:
4135   case X86::AND16rr:     case X86::AND8rr:    case X86::AND64rm:
4136   case X86::AND32rm:     case X86::AND16rm:   case X86::AND8rm:
4137   case X86::XOR64ri32:   case X86::XOR64ri8:  case X86::XOR32ri:
4138   case X86::XOR32ri8:    case X86::XOR16ri:   case X86::XOR16ri8:
4139   case X86::XOR8ri:      case X86::XOR64rr:   case X86::XOR32rr:
4140   case X86::XOR16rr:     case X86::XOR8rr:    case X86::XOR64rm:
4141   case X86::XOR32rm:     case X86::XOR16rm:   case X86::XOR8rm:
4142   case X86::OR64ri32:    case X86::OR64ri8:   case X86::OR32ri:
4143   case X86::OR32ri8:     case X86::OR16ri:    case X86::OR16ri8:
4144   case X86::OR8ri:       case X86::OR64rr:    case X86::OR32rr:
4145   case X86::OR16rr:      case X86::OR8rr:     case X86::OR64rm:
4146   case X86::OR32rm:      case X86::OR16rm:    case X86::OR8rm:
4147   case X86::ANDN32rr:    case X86::ANDN32rm:
4148   case X86::ANDN64rr:    case X86::ANDN64rm:
4149   case X86::BLSI32rr:    case X86::BLSI32rm:
4150   case X86::BLSI64rr:    case X86::BLSI64rm:
4151   case X86::BLSMSK32rr:  case X86::BLSMSK32rm:
4152   case X86::BLSMSK64rr:  case X86::BLSMSK64rm:
4153   case X86::BLSR32rr:    case X86::BLSR32rm:
4154   case X86::BLSR64rr:    case X86::BLSR64rm:
4155   case X86::BLCFILL32rr: case X86::BLCFILL32rm:
4156   case X86::BLCFILL64rr: case X86::BLCFILL64rm:
4157   case X86::BLCI32rr:    case X86::BLCI32rm:
4158   case X86::BLCI64rr:    case X86::BLCI64rm:
4159   case X86::BLCIC32rr:   case X86::BLCIC32rm:
4160   case X86::BLCIC64rr:   case X86::BLCIC64rm:
4161   case X86::BLCMSK32rr:  case X86::BLCMSK32rm:
4162   case X86::BLCMSK64rr:  case X86::BLCMSK64rm:
4163   case X86::BLCS32rr:    case X86::BLCS32rm:
4164   case X86::BLCS64rr:    case X86::BLCS64rm:
4165   case X86::BLSFILL32rr: case X86::BLSFILL32rm:
4166   case X86::BLSFILL64rr: case X86::BLSFILL64rm:
4167   case X86::BLSIC32rr:   case X86::BLSIC32rm:
4168   case X86::BLSIC64rr:   case X86::BLSIC64rm:
4169   case X86::BZHI32rr:    case X86::BZHI32rm:
4170   case X86::BZHI64rr:    case X86::BZHI64rm:
4171   case X86::T1MSKC32rr:  case X86::T1MSKC32rm:
4172   case X86::T1MSKC64rr:  case X86::T1MSKC64rm:
4173   case X86::TZMSK32rr:   case X86::TZMSK32rm:
4174   case X86::TZMSK64rr:   case X86::TZMSK64rm:
4175     // These instructions clear the overflow flag just like TEST.
4176     // FIXME: These are not the only instructions in this switch that clear the
4177     // overflow flag.
4178     ClearsOverflowFlag = true;
4179     return true;
4180   case X86::BEXTR32rr:   case X86::BEXTR64rr:
4181   case X86::BEXTR32rm:   case X86::BEXTR64rm:
4182   case X86::BEXTRI32ri:  case X86::BEXTRI32mi:
4183   case X86::BEXTRI64ri:  case X86::BEXTRI64mi:
4184     // BEXTR doesn't update the sign flag so we can't use it. It does clear
4185     // the overflow flag, but that's not useful without the sign flag.
4186     NoSignFlag = true;
4187     return true;
4188   }
4189 }
4190 
4191 /// Check whether the use can be converted to remove a comparison against zero.
4192 static X86::CondCode isUseDefConvertible(const MachineInstr &MI) {
4193   switch (MI.getOpcode()) {
4194   default: return X86::COND_INVALID;
4195   case X86::NEG8r:
4196   case X86::NEG16r:
4197   case X86::NEG32r:
4198   case X86::NEG64r:
4199     return X86::COND_AE;
4200   case X86::LZCNT16rr:
4201   case X86::LZCNT32rr:
4202   case X86::LZCNT64rr:
4203     return X86::COND_B;
4204   case X86::POPCNT16rr:
4205   case X86::POPCNT32rr:
4206   case X86::POPCNT64rr:
4207     return X86::COND_E;
4208   case X86::TZCNT16rr:
4209   case X86::TZCNT32rr:
4210   case X86::TZCNT64rr:
4211     return X86::COND_B;
4212   case X86::BSF16rr:
4213   case X86::BSF32rr:
4214   case X86::BSF64rr:
4215   case X86::BSR16rr:
4216   case X86::BSR32rr:
4217   case X86::BSR64rr:
4218     return X86::COND_E;
4219   case X86::BLSI32rr:
4220   case X86::BLSI64rr:
4221     return X86::COND_AE;
4222   case X86::BLSR32rr:
4223   case X86::BLSR64rr:
4224   case X86::BLSMSK32rr:
4225   case X86::BLSMSK64rr:
4226     return X86::COND_B;
4227   // TODO: TBM instructions.
4228   }
4229 }
4230 
4231 /// Check if there exists an earlier instruction that
4232 /// operates on the same source operands and sets flags in the same way as
4233 /// Compare; remove Compare if possible.
4234 bool X86InstrInfo::optimizeCompareInstr(MachineInstr &CmpInstr, Register SrcReg,
4235                                         Register SrcReg2, int64_t CmpMask,
4236                                         int64_t CmpValue,
4237                                         const MachineRegisterInfo *MRI) const {
4238   // Check whether we can replace SUB with CMP.
4239   switch (CmpInstr.getOpcode()) {
4240   default: break;
4241   case X86::SUB64ri32:
4242   case X86::SUB64ri8:
4243   case X86::SUB32ri:
4244   case X86::SUB32ri8:
4245   case X86::SUB16ri:
4246   case X86::SUB16ri8:
4247   case X86::SUB8ri:
4248   case X86::SUB64rm:
4249   case X86::SUB32rm:
4250   case X86::SUB16rm:
4251   case X86::SUB8rm:
4252   case X86::SUB64rr:
4253   case X86::SUB32rr:
4254   case X86::SUB16rr:
4255   case X86::SUB8rr: {
4256     if (!MRI->use_nodbg_empty(CmpInstr.getOperand(0).getReg()))
4257       return false;
4258     // There is no use of the destination register, we can replace SUB with CMP.
4259     unsigned NewOpcode = 0;
4260     switch (CmpInstr.getOpcode()) {
4261     default: llvm_unreachable("Unreachable!");
4262     case X86::SUB64rm:   NewOpcode = X86::CMP64rm;   break;
4263     case X86::SUB32rm:   NewOpcode = X86::CMP32rm;   break;
4264     case X86::SUB16rm:   NewOpcode = X86::CMP16rm;   break;
4265     case X86::SUB8rm:    NewOpcode = X86::CMP8rm;    break;
4266     case X86::SUB64rr:   NewOpcode = X86::CMP64rr;   break;
4267     case X86::SUB32rr:   NewOpcode = X86::CMP32rr;   break;
4268     case X86::SUB16rr:   NewOpcode = X86::CMP16rr;   break;
4269     case X86::SUB8rr:    NewOpcode = X86::CMP8rr;    break;
4270     case X86::SUB64ri32: NewOpcode = X86::CMP64ri32; break;
4271     case X86::SUB64ri8:  NewOpcode = X86::CMP64ri8;  break;
4272     case X86::SUB32ri:   NewOpcode = X86::CMP32ri;   break;
4273     case X86::SUB32ri8:  NewOpcode = X86::CMP32ri8;  break;
4274     case X86::SUB16ri:   NewOpcode = X86::CMP16ri;   break;
4275     case X86::SUB16ri8:  NewOpcode = X86::CMP16ri8;  break;
4276     case X86::SUB8ri:    NewOpcode = X86::CMP8ri;    break;
4277     }
4278     CmpInstr.setDesc(get(NewOpcode));
4279     CmpInstr.removeOperand(0);
4280     // Mutating this instruction invalidates any debug data associated with it.
4281     CmpInstr.dropDebugNumber();
4282     // Fall through to optimize Cmp if Cmp is CMPrr or CMPri.
4283     if (NewOpcode == X86::CMP64rm || NewOpcode == X86::CMP32rm ||
4284         NewOpcode == X86::CMP16rm || NewOpcode == X86::CMP8rm)
4285       return false;
4286   }
4287   }
4288 
4289   // The following code tries to remove the comparison by re-using EFLAGS
4290   // from earlier instructions.
4291 
4292   bool IsCmpZero = (CmpMask != 0 && CmpValue == 0);
4293 
4294   // Transformation currently requires SSA values.
4295   if (SrcReg2.isPhysical())
4296     return false;
4297   MachineInstr *SrcRegDef = MRI->getVRegDef(SrcReg);
4298   assert(SrcRegDef && "Must have a definition (SSA)");
4299 
4300   MachineInstr *MI = nullptr;
4301   MachineInstr *Sub = nullptr;
4302   MachineInstr *Movr0Inst = nullptr;
4303   bool NoSignFlag = false;
4304   bool ClearsOverflowFlag = false;
4305   bool ShouldUpdateCC = false;
4306   bool IsSwapped = false;
4307   X86::CondCode NewCC = X86::COND_INVALID;
4308   int64_t ImmDelta = 0;
4309 
4310   // Search backward from CmpInstr for the next instruction defining EFLAGS.
4311   const TargetRegisterInfo *TRI = &getRegisterInfo();
4312   MachineBasicBlock &CmpMBB = *CmpInstr.getParent();
4313   MachineBasicBlock::reverse_iterator From =
4314       std::next(MachineBasicBlock::reverse_iterator(CmpInstr));
4315   for (MachineBasicBlock *MBB = &CmpMBB;;) {
4316     for (MachineInstr &Inst : make_range(From, MBB->rend())) {
4317       // Try to use EFLAGS from the instruction defining %SrcReg. Example:
4318       //     %eax = addl ...
4319       //     ...                // EFLAGS not changed
4320       //     testl %eax, %eax   // <-- can be removed
4321       if (&Inst == SrcRegDef) {
4322         if (IsCmpZero &&
4323             isDefConvertible(Inst, NoSignFlag, ClearsOverflowFlag)) {
4324           MI = &Inst;
4325           break;
4326         }
4327 
4328         // Look back for the following pattern, in which case the test64rr
4329         // instruction could be erased.
4330         //
4331         // Example:
4332         //  %reg = and32ri %in_reg, 5
4333         //  ...                         // EFLAGS not changed.
4334         //  %src_reg = subreg_to_reg 0, %reg, %subreg.sub_index
4335         //  test64rr %src_reg, %src_reg, implicit-def $eflags
4336         MachineInstr *AndInstr = nullptr;
4337         if (IsCmpZero &&
4338             findRedundantFlagInstr(CmpInstr, Inst, MRI, &AndInstr, TRI,
4339                                    NoSignFlag, ClearsOverflowFlag)) {
4340           assert(AndInstr != nullptr && X86::isAND(AndInstr->getOpcode()));
4341           MI = AndInstr;
4342           break;
4343         }
4344         // Cannot find other candidates before definition of SrcReg.
4345         return false;
4346       }
4347 
4348       if (Inst.modifiesRegister(X86::EFLAGS, TRI)) {
4349         // Try to use EFLAGS produced by an instruction reading %SrcReg.
4350         // Example:
4351         //      %eax = ...
4352         //      ...
4353         //      popcntl %eax
4354         //      ...                 // EFLAGS not changed
4355         //      testl %eax, %eax    // <-- can be removed
4356         if (IsCmpZero) {
4357           NewCC = isUseDefConvertible(Inst);
4358           if (NewCC != X86::COND_INVALID && Inst.getOperand(1).isReg() &&
4359               Inst.getOperand(1).getReg() == SrcReg) {
4360             ShouldUpdateCC = true;
4361             MI = &Inst;
4362             break;
4363           }
4364         }
4365 
4366         // Try to use EFLAGS from an instruction with similar flag results.
4367         // Example:
4368         //     sub x, y  or  cmp x, y
4369         //     ...           // EFLAGS not changed
4370         //     cmp x, y      // <-- can be removed
4371         if (isRedundantFlagInstr(CmpInstr, SrcReg, SrcReg2, CmpMask, CmpValue,
4372                                  Inst, &IsSwapped, &ImmDelta)) {
4373           Sub = &Inst;
4374           break;
4375         }
4376 
4377         // MOV32r0 is implemented with xor which clobbers condition code. It is
4378         // safe to move up, if the definition to EFLAGS is dead and earlier
4379         // instructions do not read or write EFLAGS.
4380         if (!Movr0Inst && Inst.getOpcode() == X86::MOV32r0 &&
4381             Inst.registerDefIsDead(X86::EFLAGS, TRI)) {
4382           Movr0Inst = &Inst;
4383           continue;
4384         }
4385 
4386         // Cannot do anything for any other EFLAG changes.
4387         return false;
4388       }
4389     }
4390 
4391     if (MI || Sub)
4392       break;
4393 
4394     // Reached begin of basic block. Continue in predecessor if there is
4395     // exactly one.
4396     if (MBB->pred_size() != 1)
4397       return false;
4398     MBB = *MBB->pred_begin();
4399     From = MBB->rbegin();
4400   }
4401 
4402   // Scan forward from the instruction after CmpInstr for uses of EFLAGS.
4403   // It is safe to remove CmpInstr if EFLAGS is redefined or killed.
4404   // If we are done with the basic block, we need to check whether EFLAGS is
4405   // live-out.
4406   bool FlagsMayLiveOut = true;
4407   SmallVector<std::pair<MachineInstr*, X86::CondCode>, 4> OpsToUpdate;
4408   MachineBasicBlock::iterator AfterCmpInstr =
4409       std::next(MachineBasicBlock::iterator(CmpInstr));
4410   for (MachineInstr &Instr : make_range(AfterCmpInstr, CmpMBB.end())) {
4411     bool ModifyEFLAGS = Instr.modifiesRegister(X86::EFLAGS, TRI);
4412     bool UseEFLAGS = Instr.readsRegister(X86::EFLAGS, TRI);
4413     // We should check the usage if this instruction uses and updates EFLAGS.
4414     if (!UseEFLAGS && ModifyEFLAGS) {
4415       // It is safe to remove CmpInstr if EFLAGS is updated again.
4416       FlagsMayLiveOut = false;
4417       break;
4418     }
4419     if (!UseEFLAGS && !ModifyEFLAGS)
4420       continue;
4421 
4422     // EFLAGS is used by this instruction.
4423     X86::CondCode OldCC = X86::COND_INVALID;
4424     if (MI || IsSwapped || ImmDelta != 0) {
4425       // We decode the condition code from opcode.
4426       if (Instr.isBranch())
4427         OldCC = X86::getCondFromBranch(Instr);
4428       else {
4429         OldCC = X86::getCondFromSETCC(Instr);
4430         if (OldCC == X86::COND_INVALID)
4431           OldCC = X86::getCondFromCMov(Instr);
4432       }
4433       if (OldCC == X86::COND_INVALID) return false;
4434     }
4435     X86::CondCode ReplacementCC = X86::COND_INVALID;
4436     if (MI) {
4437       switch (OldCC) {
4438       default: break;
4439       case X86::COND_A: case X86::COND_AE:
4440       case X86::COND_B: case X86::COND_BE:
4441         // CF is used, we can't perform this optimization.
4442         return false;
4443       case X86::COND_G: case X86::COND_GE:
4444       case X86::COND_L: case X86::COND_LE:
4445       case X86::COND_O: case X86::COND_NO:
4446         // If OF is used, the instruction needs to clear it like CmpZero does.
4447         if (!ClearsOverflowFlag)
4448           return false;
4449         break;
4450       case X86::COND_S: case X86::COND_NS:
4451         // If SF is used, but the instruction doesn't update the SF, then we
4452         // can't do the optimization.
4453         if (NoSignFlag)
4454           return false;
4455         break;
4456       }
4457 
4458       // If we're updating the condition code check if we have to reverse the
4459       // condition.
4460       if (ShouldUpdateCC)
4461         switch (OldCC) {
4462         default:
4463           return false;
4464         case X86::COND_E:
4465           ReplacementCC = NewCC;
4466           break;
4467         case X86::COND_NE:
4468           ReplacementCC = GetOppositeBranchCondition(NewCC);
4469           break;
4470         }
4471     } else if (IsSwapped) {
4472       // If we have SUB(r1, r2) and CMP(r2, r1), the condition code needs
4473       // to be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc.
4474       // We swap the condition code and synthesize the new opcode.
4475       ReplacementCC = getSwappedCondition(OldCC);
4476       if (ReplacementCC == X86::COND_INVALID)
4477         return false;
4478       ShouldUpdateCC = true;
4479     } else if (ImmDelta != 0) {
4480       unsigned BitWidth = TRI->getRegSizeInBits(*MRI->getRegClass(SrcReg));
4481       // Shift amount for min/max constants to adjust for 8/16/32 instruction
4482       // sizes.
4483       switch (OldCC) {
4484       case X86::COND_L: // x <s (C + 1)  -->  x <=s C
4485         if (ImmDelta != 1 || APInt::getSignedMinValue(BitWidth) == CmpValue)
4486           return false;
4487         ReplacementCC = X86::COND_LE;
4488         break;
4489       case X86::COND_B: // x <u (C + 1)  -->  x <=u C
4490         if (ImmDelta != 1 || CmpValue == 0)
4491           return false;
4492         ReplacementCC = X86::COND_BE;
4493         break;
4494       case X86::COND_GE: // x >=s (C + 1)  -->  x >s C
4495         if (ImmDelta != 1 || APInt::getSignedMinValue(BitWidth) == CmpValue)
4496           return false;
4497         ReplacementCC = X86::COND_G;
4498         break;
4499       case X86::COND_AE: // x >=u (C + 1)  -->  x >u C
4500         if (ImmDelta != 1 || CmpValue == 0)
4501           return false;
4502         ReplacementCC = X86::COND_A;
4503         break;
4504       case X86::COND_G: // x >s (C - 1)  -->  x >=s C
4505         if (ImmDelta != -1 || APInt::getSignedMaxValue(BitWidth) == CmpValue)
4506           return false;
4507         ReplacementCC = X86::COND_GE;
4508         break;
4509       case X86::COND_A: // x >u (C - 1)  -->  x >=u C
4510         if (ImmDelta != -1 || APInt::getMaxValue(BitWidth) == CmpValue)
4511           return false;
4512         ReplacementCC = X86::COND_AE;
4513         break;
4514       case X86::COND_LE: // x <=s (C - 1)  -->  x <s C
4515         if (ImmDelta != -1 || APInt::getSignedMaxValue(BitWidth) == CmpValue)
4516           return false;
4517         ReplacementCC = X86::COND_L;
4518         break;
4519       case X86::COND_BE: // x <=u (C - 1)  -->  x <u C
4520         if (ImmDelta != -1 || APInt::getMaxValue(BitWidth) == CmpValue)
4521           return false;
4522         ReplacementCC = X86::COND_B;
4523         break;
4524       default:
4525         return false;
4526       }
4527       ShouldUpdateCC = true;
4528     }
4529 
4530     if (ShouldUpdateCC && ReplacementCC != OldCC) {
4531       // Push the MachineInstr to OpsToUpdate.
4532       // If it is safe to remove CmpInstr, the condition code of these
4533       // instructions will be modified.
4534       OpsToUpdate.push_back(std::make_pair(&Instr, ReplacementCC));
4535     }
4536     if (ModifyEFLAGS || Instr.killsRegister(X86::EFLAGS, TRI)) {
4537       // It is safe to remove CmpInstr if EFLAGS is updated again or killed.
4538       FlagsMayLiveOut = false;
4539       break;
4540     }
4541   }
4542 
4543   // If we have to update users but EFLAGS is live-out abort, since we cannot
4544   // easily find all of the users.
4545   if ((MI != nullptr || ShouldUpdateCC) && FlagsMayLiveOut) {
4546     for (MachineBasicBlock *Successor : CmpMBB.successors())
4547       if (Successor->isLiveIn(X86::EFLAGS))
4548         return false;
4549   }
4550 
4551   // The instruction to be updated is either Sub or MI.
4552   assert((MI == nullptr || Sub == nullptr) && "Should not have Sub and MI set");
4553   Sub = MI != nullptr ? MI : Sub;
4554   MachineBasicBlock *SubBB = Sub->getParent();
4555   // Move Movr0Inst to the appropriate place before Sub.
4556   if (Movr0Inst) {
4557     // Only move within the same block so we don't accidentally move to a
4558     // block with higher execution frequency.
4559     if (&CmpMBB != SubBB)
4560       return false;
4561     // Look backwards until we find a def that doesn't use the current EFLAGS.
4562     MachineBasicBlock::reverse_iterator InsertI = Sub,
4563                                         InsertE = Sub->getParent()->rend();
4564     for (; InsertI != InsertE; ++InsertI) {
4565       MachineInstr *Instr = &*InsertI;
4566       if (!Instr->readsRegister(X86::EFLAGS, TRI) &&
4567           Instr->modifiesRegister(X86::EFLAGS, TRI)) {
4568         Movr0Inst->getParent()->remove(Movr0Inst);
4569         Instr->getParent()->insert(MachineBasicBlock::iterator(Instr),
4570                                    Movr0Inst);
4571         break;
4572       }
4573     }
4574     if (InsertI == InsertE)
4575       return false;
4576   }
4577 
4578   // Make sure Sub instruction defines EFLAGS and mark the def live.
4579   MachineOperand *FlagDef = Sub->findRegisterDefOperand(X86::EFLAGS);
4580   assert(FlagDef && "Unable to locate a def EFLAGS operand");
4581   FlagDef->setIsDead(false);
4582 
4583   CmpInstr.eraseFromParent();
4584 
4585   // Modify the condition code of instructions in OpsToUpdate.
4586   for (auto &Op : OpsToUpdate) {
4587     Op.first->getOperand(Op.first->getDesc().getNumOperands() - 1)
4588         .setImm(Op.second);
4589   }
4590   // Add EFLAGS to block live-ins between CmpBB and block of flags producer.
4591   for (MachineBasicBlock *MBB = &CmpMBB; MBB != SubBB;
4592        MBB = *MBB->pred_begin()) {
4593     assert(MBB->pred_size() == 1 && "Expected exactly one predecessor");
4594     if (!MBB->isLiveIn(X86::EFLAGS))
4595       MBB->addLiveIn(X86::EFLAGS);
4596   }
4597   return true;
4598 }
4599 
4600 /// Try to remove the load by folding it to a register
4601 /// operand at the use. We fold the load instructions if load defines a virtual
4602 /// register, the virtual register is used once in the same BB, and the
4603 /// instructions in-between do not load or store, and have no side effects.
4604 MachineInstr *X86InstrInfo::optimizeLoadInstr(MachineInstr &MI,
4605                                               const MachineRegisterInfo *MRI,
4606                                               Register &FoldAsLoadDefReg,
4607                                               MachineInstr *&DefMI) const {
4608   // Check whether we can move DefMI here.
4609   DefMI = MRI->getVRegDef(FoldAsLoadDefReg);
4610   assert(DefMI);
4611   bool SawStore = false;
4612   if (!DefMI->isSafeToMove(nullptr, SawStore))
4613     return nullptr;
4614 
4615   // Collect information about virtual register operands of MI.
4616   SmallVector<unsigned, 1> SrcOperandIds;
4617   for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
4618     MachineOperand &MO = MI.getOperand(i);
4619     if (!MO.isReg())
4620       continue;
4621     Register Reg = MO.getReg();
4622     if (Reg != FoldAsLoadDefReg)
4623       continue;
4624     // Do not fold if we have a subreg use or a def.
4625     if (MO.getSubReg() || MO.isDef())
4626       return nullptr;
4627     SrcOperandIds.push_back(i);
4628   }
4629   if (SrcOperandIds.empty())
4630     return nullptr;
4631 
4632   // Check whether we can fold the def into SrcOperandId.
4633   if (MachineInstr *FoldMI = foldMemoryOperand(MI, SrcOperandIds, *DefMI)) {
4634     FoldAsLoadDefReg = 0;
4635     return FoldMI;
4636   }
4637 
4638   return nullptr;
4639 }
4640 
4641 /// Expand a single-def pseudo instruction to a two-addr
4642 /// instruction with two undef reads of the register being defined.
4643 /// This is used for mapping:
4644 ///   %xmm4 = V_SET0
4645 /// to:
4646 ///   %xmm4 = PXORrr undef %xmm4, undef %xmm4
4647 ///
4648 static bool Expand2AddrUndef(MachineInstrBuilder &MIB,
4649                              const MCInstrDesc &Desc) {
4650   assert(Desc.getNumOperands() == 3 && "Expected two-addr instruction.");
4651   Register Reg = MIB.getReg(0);
4652   MIB->setDesc(Desc);
4653 
4654   // MachineInstr::addOperand() will insert explicit operands before any
4655   // implicit operands.
4656   MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
4657   // But we don't trust that.
4658   assert(MIB.getReg(1) == Reg &&
4659          MIB.getReg(2) == Reg && "Misplaced operand");
4660   return true;
4661 }
4662 
4663 /// Expand a single-def pseudo instruction to a two-addr
4664 /// instruction with two %k0 reads.
4665 /// This is used for mapping:
4666 ///   %k4 = K_SET1
4667 /// to:
4668 ///   %k4 = KXNORrr %k0, %k0
4669 static bool Expand2AddrKreg(MachineInstrBuilder &MIB, const MCInstrDesc &Desc,
4670                             Register Reg) {
4671   assert(Desc.getNumOperands() == 3 && "Expected two-addr instruction.");
4672   MIB->setDesc(Desc);
4673   MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
4674   return true;
4675 }
4676 
4677 static bool expandMOV32r1(MachineInstrBuilder &MIB, const TargetInstrInfo &TII,
4678                           bool MinusOne) {
4679   MachineBasicBlock &MBB = *MIB->getParent();
4680   const DebugLoc &DL = MIB->getDebugLoc();
4681   Register Reg = MIB.getReg(0);
4682 
4683   // Insert the XOR.
4684   BuildMI(MBB, MIB.getInstr(), DL, TII.get(X86::XOR32rr), Reg)
4685       .addReg(Reg, RegState::Undef)
4686       .addReg(Reg, RegState::Undef);
4687 
4688   // Turn the pseudo into an INC or DEC.
4689   MIB->setDesc(TII.get(MinusOne ? X86::DEC32r : X86::INC32r));
4690   MIB.addReg(Reg);
4691 
4692   return true;
4693 }
4694 
4695 static bool ExpandMOVImmSExti8(MachineInstrBuilder &MIB,
4696                                const TargetInstrInfo &TII,
4697                                const X86Subtarget &Subtarget) {
4698   MachineBasicBlock &MBB = *MIB->getParent();
4699   const DebugLoc &DL = MIB->getDebugLoc();
4700   int64_t Imm = MIB->getOperand(1).getImm();
4701   assert(Imm != 0 && "Using push/pop for 0 is not efficient.");
4702   MachineBasicBlock::iterator I = MIB.getInstr();
4703 
4704   int StackAdjustment;
4705 
4706   if (Subtarget.is64Bit()) {
4707     assert(MIB->getOpcode() == X86::MOV64ImmSExti8 ||
4708            MIB->getOpcode() == X86::MOV32ImmSExti8);
4709 
4710     // Can't use push/pop lowering if the function might write to the red zone.
4711     X86MachineFunctionInfo *X86FI =
4712         MBB.getParent()->getInfo<X86MachineFunctionInfo>();
4713     if (X86FI->getUsesRedZone()) {
4714       MIB->setDesc(TII.get(MIB->getOpcode() ==
4715                            X86::MOV32ImmSExti8 ? X86::MOV32ri : X86::MOV64ri));
4716       return true;
4717     }
4718 
4719     // 64-bit mode doesn't have 32-bit push/pop, so use 64-bit operations and
4720     // widen the register if necessary.
4721     StackAdjustment = 8;
4722     BuildMI(MBB, I, DL, TII.get(X86::PUSH64i8)).addImm(Imm);
4723     MIB->setDesc(TII.get(X86::POP64r));
4724     MIB->getOperand(0)
4725         .setReg(getX86SubSuperRegister(MIB.getReg(0), 64));
4726   } else {
4727     assert(MIB->getOpcode() == X86::MOV32ImmSExti8);
4728     StackAdjustment = 4;
4729     BuildMI(MBB, I, DL, TII.get(X86::PUSH32i8)).addImm(Imm);
4730     MIB->setDesc(TII.get(X86::POP32r));
4731   }
4732   MIB->removeOperand(1);
4733   MIB->addImplicitDefUseOperands(*MBB.getParent());
4734 
4735   // Build CFI if necessary.
4736   MachineFunction &MF = *MBB.getParent();
4737   const X86FrameLowering *TFL = Subtarget.getFrameLowering();
4738   bool IsWin64Prologue = MF.getTarget().getMCAsmInfo()->usesWindowsCFI();
4739   bool NeedsDwarfCFI = !IsWin64Prologue && MF.needsFrameMoves();
4740   bool EmitCFI = !TFL->hasFP(MF) && NeedsDwarfCFI;
4741   if (EmitCFI) {
4742     TFL->BuildCFI(MBB, I, DL,
4743         MCCFIInstruction::createAdjustCfaOffset(nullptr, StackAdjustment));
4744     TFL->BuildCFI(MBB, std::next(I), DL,
4745         MCCFIInstruction::createAdjustCfaOffset(nullptr, -StackAdjustment));
4746   }
4747 
4748   return true;
4749 }
4750 
4751 // LoadStackGuard has so far only been implemented for 64-bit MachO. Different
4752 // code sequence is needed for other targets.
4753 static void expandLoadStackGuard(MachineInstrBuilder &MIB,
4754                                  const TargetInstrInfo &TII) {
4755   MachineBasicBlock &MBB = *MIB->getParent();
4756   const DebugLoc &DL = MIB->getDebugLoc();
4757   Register Reg = MIB.getReg(0);
4758   const GlobalValue *GV =
4759       cast<GlobalValue>((*MIB->memoperands_begin())->getValue());
4760   auto Flags = MachineMemOperand::MOLoad |
4761                MachineMemOperand::MODereferenceable |
4762                MachineMemOperand::MOInvariant;
4763   MachineMemOperand *MMO = MBB.getParent()->getMachineMemOperand(
4764       MachinePointerInfo::getGOT(*MBB.getParent()), Flags, 8, Align(8));
4765   MachineBasicBlock::iterator I = MIB.getInstr();
4766 
4767   BuildMI(MBB, I, DL, TII.get(X86::MOV64rm), Reg).addReg(X86::RIP).addImm(1)
4768       .addReg(0).addGlobalAddress(GV, 0, X86II::MO_GOTPCREL).addReg(0)
4769       .addMemOperand(MMO);
4770   MIB->setDebugLoc(DL);
4771   MIB->setDesc(TII.get(X86::MOV64rm));
4772   MIB.addReg(Reg, RegState::Kill).addImm(1).addReg(0).addImm(0).addReg(0);
4773 }
4774 
4775 static bool expandXorFP(MachineInstrBuilder &MIB, const TargetInstrInfo &TII) {
4776   MachineBasicBlock &MBB = *MIB->getParent();
4777   MachineFunction &MF = *MBB.getParent();
4778   const X86Subtarget &Subtarget = MF.getSubtarget<X86Subtarget>();
4779   const X86RegisterInfo *TRI = Subtarget.getRegisterInfo();
4780   unsigned XorOp =
4781       MIB->getOpcode() == X86::XOR64_FP ? X86::XOR64rr : X86::XOR32rr;
4782   MIB->setDesc(TII.get(XorOp));
4783   MIB.addReg(TRI->getFrameRegister(MF), RegState::Undef);
4784   return true;
4785 }
4786 
4787 // This is used to handle spills for 128/256-bit registers when we have AVX512,
4788 // but not VLX. If it uses an extended register we need to use an instruction
4789 // that loads the lower 128/256-bit, but is available with only AVX512F.
4790 static bool expandNOVLXLoad(MachineInstrBuilder &MIB,
4791                             const TargetRegisterInfo *TRI,
4792                             const MCInstrDesc &LoadDesc,
4793                             const MCInstrDesc &BroadcastDesc,
4794                             unsigned SubIdx) {
4795   Register DestReg = MIB.getReg(0);
4796   // Check if DestReg is XMM16-31 or YMM16-31.
4797   if (TRI->getEncodingValue(DestReg) < 16) {
4798     // We can use a normal VEX encoded load.
4799     MIB->setDesc(LoadDesc);
4800   } else {
4801     // Use a 128/256-bit VBROADCAST instruction.
4802     MIB->setDesc(BroadcastDesc);
4803     // Change the destination to a 512-bit register.
4804     DestReg = TRI->getMatchingSuperReg(DestReg, SubIdx, &X86::VR512RegClass);
4805     MIB->getOperand(0).setReg(DestReg);
4806   }
4807   return true;
4808 }
4809 
4810 // This is used to handle spills for 128/256-bit registers when we have AVX512,
4811 // but not VLX. If it uses an extended register we need to use an instruction
4812 // that stores the lower 128/256-bit, but is available with only AVX512F.
4813 static bool expandNOVLXStore(MachineInstrBuilder &MIB,
4814                              const TargetRegisterInfo *TRI,
4815                              const MCInstrDesc &StoreDesc,
4816                              const MCInstrDesc &ExtractDesc,
4817                              unsigned SubIdx) {
4818   Register SrcReg = MIB.getReg(X86::AddrNumOperands);
4819   // Check if DestReg is XMM16-31 or YMM16-31.
4820   if (TRI->getEncodingValue(SrcReg) < 16) {
4821     // We can use a normal VEX encoded store.
4822     MIB->setDesc(StoreDesc);
4823   } else {
4824     // Use a VEXTRACTF instruction.
4825     MIB->setDesc(ExtractDesc);
4826     // Change the destination to a 512-bit register.
4827     SrcReg = TRI->getMatchingSuperReg(SrcReg, SubIdx, &X86::VR512RegClass);
4828     MIB->getOperand(X86::AddrNumOperands).setReg(SrcReg);
4829     MIB.addImm(0x0); // Append immediate to extract from the lower bits.
4830   }
4831 
4832   return true;
4833 }
4834 
4835 static bool expandSHXDROT(MachineInstrBuilder &MIB, const MCInstrDesc &Desc) {
4836   MIB->setDesc(Desc);
4837   int64_t ShiftAmt = MIB->getOperand(2).getImm();
4838   // Temporarily remove the immediate so we can add another source register.
4839   MIB->removeOperand(2);
4840   // Add the register. Don't copy the kill flag if there is one.
4841   MIB.addReg(MIB.getReg(1),
4842              getUndefRegState(MIB->getOperand(1).isUndef()));
4843   // Add back the immediate.
4844   MIB.addImm(ShiftAmt);
4845   return true;
4846 }
4847 
4848 bool X86InstrInfo::expandPostRAPseudo(MachineInstr &MI) const {
4849   bool HasAVX = Subtarget.hasAVX();
4850   MachineInstrBuilder MIB(*MI.getParent()->getParent(), MI);
4851   switch (MI.getOpcode()) {
4852   case X86::MOV32r0:
4853     return Expand2AddrUndef(MIB, get(X86::XOR32rr));
4854   case X86::MOV32r1:
4855     return expandMOV32r1(MIB, *this, /*MinusOne=*/ false);
4856   case X86::MOV32r_1:
4857     return expandMOV32r1(MIB, *this, /*MinusOne=*/ true);
4858   case X86::MOV32ImmSExti8:
4859   case X86::MOV64ImmSExti8:
4860     return ExpandMOVImmSExti8(MIB, *this, Subtarget);
4861   case X86::SETB_C32r:
4862     return Expand2AddrUndef(MIB, get(X86::SBB32rr));
4863   case X86::SETB_C64r:
4864     return Expand2AddrUndef(MIB, get(X86::SBB64rr));
4865   case X86::MMX_SET0:
4866     return Expand2AddrUndef(MIB, get(X86::MMX_PXORrr));
4867   case X86::V_SET0:
4868   case X86::FsFLD0SS:
4869   case X86::FsFLD0SD:
4870   case X86::FsFLD0F128:
4871     return Expand2AddrUndef(MIB, get(HasAVX ? X86::VXORPSrr : X86::XORPSrr));
4872   case X86::AVX_SET0: {
4873     assert(HasAVX && "AVX not supported");
4874     const TargetRegisterInfo *TRI = &getRegisterInfo();
4875     Register SrcReg = MIB.getReg(0);
4876     Register XReg = TRI->getSubReg(SrcReg, X86::sub_xmm);
4877     MIB->getOperand(0).setReg(XReg);
4878     Expand2AddrUndef(MIB, get(X86::VXORPSrr));
4879     MIB.addReg(SrcReg, RegState::ImplicitDefine);
4880     return true;
4881   }
4882   case X86::AVX512_128_SET0:
4883   case X86::AVX512_FsFLD0SH:
4884   case X86::AVX512_FsFLD0SS:
4885   case X86::AVX512_FsFLD0SD:
4886   case X86::AVX512_FsFLD0F128: {
4887     bool HasVLX = Subtarget.hasVLX();
4888     Register SrcReg = MIB.getReg(0);
4889     const TargetRegisterInfo *TRI = &getRegisterInfo();
4890     if (HasVLX || TRI->getEncodingValue(SrcReg) < 16)
4891       return Expand2AddrUndef(MIB,
4892                               get(HasVLX ? X86::VPXORDZ128rr : X86::VXORPSrr));
4893     // Extended register without VLX. Use a larger XOR.
4894     SrcReg =
4895         TRI->getMatchingSuperReg(SrcReg, X86::sub_xmm, &X86::VR512RegClass);
4896     MIB->getOperand(0).setReg(SrcReg);
4897     return Expand2AddrUndef(MIB, get(X86::VPXORDZrr));
4898   }
4899   case X86::AVX512_256_SET0:
4900   case X86::AVX512_512_SET0: {
4901     bool HasVLX = Subtarget.hasVLX();
4902     Register SrcReg = MIB.getReg(0);
4903     const TargetRegisterInfo *TRI = &getRegisterInfo();
4904     if (HasVLX || TRI->getEncodingValue(SrcReg) < 16) {
4905       Register XReg = TRI->getSubReg(SrcReg, X86::sub_xmm);
4906       MIB->getOperand(0).setReg(XReg);
4907       Expand2AddrUndef(MIB,
4908                        get(HasVLX ? X86::VPXORDZ128rr : X86::VXORPSrr));
4909       MIB.addReg(SrcReg, RegState::ImplicitDefine);
4910       return true;
4911     }
4912     if (MI.getOpcode() == X86::AVX512_256_SET0) {
4913       // No VLX so we must reference a zmm.
4914       unsigned ZReg =
4915         TRI->getMatchingSuperReg(SrcReg, X86::sub_ymm, &X86::VR512RegClass);
4916       MIB->getOperand(0).setReg(ZReg);
4917     }
4918     return Expand2AddrUndef(MIB, get(X86::VPXORDZrr));
4919   }
4920   case X86::V_SETALLONES:
4921     return Expand2AddrUndef(MIB, get(HasAVX ? X86::VPCMPEQDrr : X86::PCMPEQDrr));
4922   case X86::AVX2_SETALLONES:
4923     return Expand2AddrUndef(MIB, get(X86::VPCMPEQDYrr));
4924   case X86::AVX1_SETALLONES: {
4925     Register Reg = MIB.getReg(0);
4926     // VCMPPSYrri with an immediate 0xf should produce VCMPTRUEPS.
4927     MIB->setDesc(get(X86::VCMPPSYrri));
4928     MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef).addImm(0xf);
4929     return true;
4930   }
4931   case X86::AVX512_512_SETALLONES: {
4932     Register Reg = MIB.getReg(0);
4933     MIB->setDesc(get(X86::VPTERNLOGDZrri));
4934     // VPTERNLOGD needs 3 register inputs and an immediate.
4935     // 0xff will return 1s for any input.
4936     MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef)
4937        .addReg(Reg, RegState::Undef).addImm(0xff);
4938     return true;
4939   }
4940   case X86::AVX512_512_SEXT_MASK_32:
4941   case X86::AVX512_512_SEXT_MASK_64: {
4942     Register Reg = MIB.getReg(0);
4943     Register MaskReg = MIB.getReg(1);
4944     unsigned MaskState = getRegState(MIB->getOperand(1));
4945     unsigned Opc = (MI.getOpcode() == X86::AVX512_512_SEXT_MASK_64) ?
4946                    X86::VPTERNLOGQZrrikz : X86::VPTERNLOGDZrrikz;
4947     MI.removeOperand(1);
4948     MIB->setDesc(get(Opc));
4949     // VPTERNLOG needs 3 register inputs and an immediate.
4950     // 0xff will return 1s for any input.
4951     MIB.addReg(Reg, RegState::Undef).addReg(MaskReg, MaskState)
4952        .addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef).addImm(0xff);
4953     return true;
4954   }
4955   case X86::VMOVAPSZ128rm_NOVLX:
4956     return expandNOVLXLoad(MIB, &getRegisterInfo(), get(X86::VMOVAPSrm),
4957                            get(X86::VBROADCASTF32X4rm), X86::sub_xmm);
4958   case X86::VMOVUPSZ128rm_NOVLX:
4959     return expandNOVLXLoad(MIB, &getRegisterInfo(), get(X86::VMOVUPSrm),
4960                            get(X86::VBROADCASTF32X4rm), X86::sub_xmm);
4961   case X86::VMOVAPSZ256rm_NOVLX:
4962     return expandNOVLXLoad(MIB, &getRegisterInfo(), get(X86::VMOVAPSYrm),
4963                            get(X86::VBROADCASTF64X4rm), X86::sub_ymm);
4964   case X86::VMOVUPSZ256rm_NOVLX:
4965     return expandNOVLXLoad(MIB, &getRegisterInfo(), get(X86::VMOVUPSYrm),
4966                            get(X86::VBROADCASTF64X4rm), X86::sub_ymm);
4967   case X86::VMOVAPSZ128mr_NOVLX:
4968     return expandNOVLXStore(MIB, &getRegisterInfo(), get(X86::VMOVAPSmr),
4969                             get(X86::VEXTRACTF32x4Zmr), X86::sub_xmm);
4970   case X86::VMOVUPSZ128mr_NOVLX:
4971     return expandNOVLXStore(MIB, &getRegisterInfo(), get(X86::VMOVUPSmr),
4972                             get(X86::VEXTRACTF32x4Zmr), X86::sub_xmm);
4973   case X86::VMOVAPSZ256mr_NOVLX:
4974     return expandNOVLXStore(MIB, &getRegisterInfo(), get(X86::VMOVAPSYmr),
4975                             get(X86::VEXTRACTF64x4Zmr), X86::sub_ymm);
4976   case X86::VMOVUPSZ256mr_NOVLX:
4977     return expandNOVLXStore(MIB, &getRegisterInfo(), get(X86::VMOVUPSYmr),
4978                             get(X86::VEXTRACTF64x4Zmr), X86::sub_ymm);
4979   case X86::MOV32ri64: {
4980     Register Reg = MIB.getReg(0);
4981     Register Reg32 = RI.getSubReg(Reg, X86::sub_32bit);
4982     MI.setDesc(get(X86::MOV32ri));
4983     MIB->getOperand(0).setReg(Reg32);
4984     MIB.addReg(Reg, RegState::ImplicitDefine);
4985     return true;
4986   }
4987 
4988   // KNL does not recognize dependency-breaking idioms for mask registers,
4989   // so kxnor %k1, %k1, %k2 has a RAW dependence on %k1.
4990   // Using %k0 as the undef input register is a performance heuristic based
4991   // on the assumption that %k0 is used less frequently than the other mask
4992   // registers, since it is not usable as a write mask.
4993   // FIXME: A more advanced approach would be to choose the best input mask
4994   // register based on context.
4995   case X86::KSET0W: return Expand2AddrKreg(MIB, get(X86::KXORWrr), X86::K0);
4996   case X86::KSET0D: return Expand2AddrKreg(MIB, get(X86::KXORDrr), X86::K0);
4997   case X86::KSET0Q: return Expand2AddrKreg(MIB, get(X86::KXORQrr), X86::K0);
4998   case X86::KSET1W: return Expand2AddrKreg(MIB, get(X86::KXNORWrr), X86::K0);
4999   case X86::KSET1D: return Expand2AddrKreg(MIB, get(X86::KXNORDrr), X86::K0);
5000   case X86::KSET1Q: return Expand2AddrKreg(MIB, get(X86::KXNORQrr), X86::K0);
5001   case TargetOpcode::LOAD_STACK_GUARD:
5002     expandLoadStackGuard(MIB, *this);
5003     return true;
5004   case X86::XOR64_FP:
5005   case X86::XOR32_FP:
5006     return expandXorFP(MIB, *this);
5007   case X86::SHLDROT32ri: return expandSHXDROT(MIB, get(X86::SHLD32rri8));
5008   case X86::SHLDROT64ri: return expandSHXDROT(MIB, get(X86::SHLD64rri8));
5009   case X86::SHRDROT32ri: return expandSHXDROT(MIB, get(X86::SHRD32rri8));
5010   case X86::SHRDROT64ri: return expandSHXDROT(MIB, get(X86::SHRD64rri8));
5011   case X86::ADD8rr_DB:    MIB->setDesc(get(X86::OR8rr));    break;
5012   case X86::ADD16rr_DB:   MIB->setDesc(get(X86::OR16rr));   break;
5013   case X86::ADD32rr_DB:   MIB->setDesc(get(X86::OR32rr));   break;
5014   case X86::ADD64rr_DB:   MIB->setDesc(get(X86::OR64rr));   break;
5015   case X86::ADD8ri_DB:    MIB->setDesc(get(X86::OR8ri));    break;
5016   case X86::ADD16ri_DB:   MIB->setDesc(get(X86::OR16ri));   break;
5017   case X86::ADD32ri_DB:   MIB->setDesc(get(X86::OR32ri));   break;
5018   case X86::ADD64ri32_DB: MIB->setDesc(get(X86::OR64ri32)); break;
5019   case X86::ADD16ri8_DB:  MIB->setDesc(get(X86::OR16ri8));  break;
5020   case X86::ADD32ri8_DB:  MIB->setDesc(get(X86::OR32ri8));  break;
5021   case X86::ADD64ri8_DB:  MIB->setDesc(get(X86::OR64ri8));  break;
5022   }
5023   return false;
5024 }
5025 
5026 /// Return true for all instructions that only update
5027 /// the first 32 or 64-bits of the destination register and leave the rest
5028 /// unmodified. This can be used to avoid folding loads if the instructions
5029 /// only update part of the destination register, and the non-updated part is
5030 /// not needed. e.g. cvtss2sd, sqrtss. Unfolding the load from these
5031 /// instructions breaks the partial register dependency and it can improve
5032 /// performance. e.g.:
5033 ///
5034 ///   movss (%rdi), %xmm0
5035 ///   cvtss2sd %xmm0, %xmm0
5036 ///
5037 /// Instead of
5038 ///   cvtss2sd (%rdi), %xmm0
5039 ///
5040 /// FIXME: This should be turned into a TSFlags.
5041 ///
5042 static bool hasPartialRegUpdate(unsigned Opcode,
5043                                 const X86Subtarget &Subtarget,
5044                                 bool ForLoadFold = false) {
5045   switch (Opcode) {
5046   case X86::CVTSI2SSrr:
5047   case X86::CVTSI2SSrm:
5048   case X86::CVTSI642SSrr:
5049   case X86::CVTSI642SSrm:
5050   case X86::CVTSI2SDrr:
5051   case X86::CVTSI2SDrm:
5052   case X86::CVTSI642SDrr:
5053   case X86::CVTSI642SDrm:
5054     // Load folding won't effect the undef register update since the input is
5055     // a GPR.
5056     return !ForLoadFold;
5057   case X86::CVTSD2SSrr:
5058   case X86::CVTSD2SSrm:
5059   case X86::CVTSS2SDrr:
5060   case X86::CVTSS2SDrm:
5061   case X86::MOVHPDrm:
5062   case X86::MOVHPSrm:
5063   case X86::MOVLPDrm:
5064   case X86::MOVLPSrm:
5065   case X86::RCPSSr:
5066   case X86::RCPSSm:
5067   case X86::RCPSSr_Int:
5068   case X86::RCPSSm_Int:
5069   case X86::ROUNDSDr:
5070   case X86::ROUNDSDm:
5071   case X86::ROUNDSSr:
5072   case X86::ROUNDSSm:
5073   case X86::RSQRTSSr:
5074   case X86::RSQRTSSm:
5075   case X86::RSQRTSSr_Int:
5076   case X86::RSQRTSSm_Int:
5077   case X86::SQRTSSr:
5078   case X86::SQRTSSm:
5079   case X86::SQRTSSr_Int:
5080   case X86::SQRTSSm_Int:
5081   case X86::SQRTSDr:
5082   case X86::SQRTSDm:
5083   case X86::SQRTSDr_Int:
5084   case X86::SQRTSDm_Int:
5085     return true;
5086   case X86::VFCMULCPHZ128rm:
5087   case X86::VFCMULCPHZ128rmb:
5088   case X86::VFCMULCPHZ128rmbkz:
5089   case X86::VFCMULCPHZ128rmkz:
5090   case X86::VFCMULCPHZ128rr:
5091   case X86::VFCMULCPHZ128rrkz:
5092   case X86::VFCMULCPHZ256rm:
5093   case X86::VFCMULCPHZ256rmb:
5094   case X86::VFCMULCPHZ256rmbkz:
5095   case X86::VFCMULCPHZ256rmkz:
5096   case X86::VFCMULCPHZ256rr:
5097   case X86::VFCMULCPHZ256rrkz:
5098   case X86::VFCMULCPHZrm:
5099   case X86::VFCMULCPHZrmb:
5100   case X86::VFCMULCPHZrmbkz:
5101   case X86::VFCMULCPHZrmkz:
5102   case X86::VFCMULCPHZrr:
5103   case X86::VFCMULCPHZrrb:
5104   case X86::VFCMULCPHZrrbkz:
5105   case X86::VFCMULCPHZrrkz:
5106   case X86::VFMULCPHZ128rm:
5107   case X86::VFMULCPHZ128rmb:
5108   case X86::VFMULCPHZ128rmbkz:
5109   case X86::VFMULCPHZ128rmkz:
5110   case X86::VFMULCPHZ128rr:
5111   case X86::VFMULCPHZ128rrkz:
5112   case X86::VFMULCPHZ256rm:
5113   case X86::VFMULCPHZ256rmb:
5114   case X86::VFMULCPHZ256rmbkz:
5115   case X86::VFMULCPHZ256rmkz:
5116   case X86::VFMULCPHZ256rr:
5117   case X86::VFMULCPHZ256rrkz:
5118   case X86::VFMULCPHZrm:
5119   case X86::VFMULCPHZrmb:
5120   case X86::VFMULCPHZrmbkz:
5121   case X86::VFMULCPHZrmkz:
5122   case X86::VFMULCPHZrr:
5123   case X86::VFMULCPHZrrb:
5124   case X86::VFMULCPHZrrbkz:
5125   case X86::VFMULCPHZrrkz:
5126   case X86::VFCMULCSHZrm:
5127   case X86::VFCMULCSHZrmkz:
5128   case X86::VFCMULCSHZrr:
5129   case X86::VFCMULCSHZrrb:
5130   case X86::VFCMULCSHZrrbkz:
5131   case X86::VFCMULCSHZrrkz:
5132   case X86::VFMULCSHZrm:
5133   case X86::VFMULCSHZrmkz:
5134   case X86::VFMULCSHZrr:
5135   case X86::VFMULCSHZrrb:
5136   case X86::VFMULCSHZrrbkz:
5137   case X86::VFMULCSHZrrkz:
5138     return Subtarget.hasMULCFalseDeps();
5139   case X86::VPERMDYrm:
5140   case X86::VPERMDYrr:
5141   case X86::VPERMQYmi:
5142   case X86::VPERMQYri:
5143   case X86::VPERMPSYrm:
5144   case X86::VPERMPSYrr:
5145   case X86::VPERMPDYmi:
5146   case X86::VPERMPDYri:
5147   case X86::VPERMDZ256rm:
5148   case X86::VPERMDZ256rmb:
5149   case X86::VPERMDZ256rmbkz:
5150   case X86::VPERMDZ256rmkz:
5151   case X86::VPERMDZ256rr:
5152   case X86::VPERMDZ256rrkz:
5153   case X86::VPERMDZrm:
5154   case X86::VPERMDZrmb:
5155   case X86::VPERMDZrmbkz:
5156   case X86::VPERMDZrmkz:
5157   case X86::VPERMDZrr:
5158   case X86::VPERMDZrrkz:
5159   case X86::VPERMQZ256mbi:
5160   case X86::VPERMQZ256mbikz:
5161   case X86::VPERMQZ256mi:
5162   case X86::VPERMQZ256mikz:
5163   case X86::VPERMQZ256ri:
5164   case X86::VPERMQZ256rikz:
5165   case X86::VPERMQZ256rm:
5166   case X86::VPERMQZ256rmb:
5167   case X86::VPERMQZ256rmbkz:
5168   case X86::VPERMQZ256rmkz:
5169   case X86::VPERMQZ256rr:
5170   case X86::VPERMQZ256rrkz:
5171   case X86::VPERMQZmbi:
5172   case X86::VPERMQZmbikz:
5173   case X86::VPERMQZmi:
5174   case X86::VPERMQZmikz:
5175   case X86::VPERMQZri:
5176   case X86::VPERMQZrikz:
5177   case X86::VPERMQZrm:
5178   case X86::VPERMQZrmb:
5179   case X86::VPERMQZrmbkz:
5180   case X86::VPERMQZrmkz:
5181   case X86::VPERMQZrr:
5182   case X86::VPERMQZrrkz:
5183   case X86::VPERMPSZ256rm:
5184   case X86::VPERMPSZ256rmb:
5185   case X86::VPERMPSZ256rmbkz:
5186   case X86::VPERMPSZ256rmkz:
5187   case X86::VPERMPSZ256rr:
5188   case X86::VPERMPSZ256rrkz:
5189   case X86::VPERMPSZrm:
5190   case X86::VPERMPSZrmb:
5191   case X86::VPERMPSZrmbkz:
5192   case X86::VPERMPSZrmkz:
5193   case X86::VPERMPSZrr:
5194   case X86::VPERMPSZrrkz:
5195   case X86::VPERMPDZ256mbi:
5196   case X86::VPERMPDZ256mbikz:
5197   case X86::VPERMPDZ256mi:
5198   case X86::VPERMPDZ256mikz:
5199   case X86::VPERMPDZ256ri:
5200   case X86::VPERMPDZ256rikz:
5201   case X86::VPERMPDZ256rm:
5202   case X86::VPERMPDZ256rmb:
5203   case X86::VPERMPDZ256rmbkz:
5204   case X86::VPERMPDZ256rmkz:
5205   case X86::VPERMPDZ256rr:
5206   case X86::VPERMPDZ256rrkz:
5207   case X86::VPERMPDZmbi:
5208   case X86::VPERMPDZmbikz:
5209   case X86::VPERMPDZmi:
5210   case X86::VPERMPDZmikz:
5211   case X86::VPERMPDZri:
5212   case X86::VPERMPDZrikz:
5213   case X86::VPERMPDZrm:
5214   case X86::VPERMPDZrmb:
5215   case X86::VPERMPDZrmbkz:
5216   case X86::VPERMPDZrmkz:
5217   case X86::VPERMPDZrr:
5218   case X86::VPERMPDZrrkz:
5219     return Subtarget.hasPERMFalseDeps();
5220   case X86::VRANGEPDZ128rmbi:
5221   case X86::VRANGEPDZ128rmbikz:
5222   case X86::VRANGEPDZ128rmi:
5223   case X86::VRANGEPDZ128rmikz:
5224   case X86::VRANGEPDZ128rri:
5225   case X86::VRANGEPDZ128rrikz:
5226   case X86::VRANGEPDZ256rmbi:
5227   case X86::VRANGEPDZ256rmbikz:
5228   case X86::VRANGEPDZ256rmi:
5229   case X86::VRANGEPDZ256rmikz:
5230   case X86::VRANGEPDZ256rri:
5231   case X86::VRANGEPDZ256rrikz:
5232   case X86::VRANGEPDZrmbi:
5233   case X86::VRANGEPDZrmbikz:
5234   case X86::VRANGEPDZrmi:
5235   case X86::VRANGEPDZrmikz:
5236   case X86::VRANGEPDZrri:
5237   case X86::VRANGEPDZrrib:
5238   case X86::VRANGEPDZrribkz:
5239   case X86::VRANGEPDZrrikz:
5240   case X86::VRANGEPSZ128rmbi:
5241   case X86::VRANGEPSZ128rmbikz:
5242   case X86::VRANGEPSZ128rmi:
5243   case X86::VRANGEPSZ128rmikz:
5244   case X86::VRANGEPSZ128rri:
5245   case X86::VRANGEPSZ128rrikz:
5246   case X86::VRANGEPSZ256rmbi:
5247   case X86::VRANGEPSZ256rmbikz:
5248   case X86::VRANGEPSZ256rmi:
5249   case X86::VRANGEPSZ256rmikz:
5250   case X86::VRANGEPSZ256rri:
5251   case X86::VRANGEPSZ256rrikz:
5252   case X86::VRANGEPSZrmbi:
5253   case X86::VRANGEPSZrmbikz:
5254   case X86::VRANGEPSZrmi:
5255   case X86::VRANGEPSZrmikz:
5256   case X86::VRANGEPSZrri:
5257   case X86::VRANGEPSZrrib:
5258   case X86::VRANGEPSZrribkz:
5259   case X86::VRANGEPSZrrikz:
5260   case X86::VRANGESDZrmi:
5261   case X86::VRANGESDZrmikz:
5262   case X86::VRANGESDZrri:
5263   case X86::VRANGESDZrrib:
5264   case X86::VRANGESDZrribkz:
5265   case X86::VRANGESDZrrikz:
5266   case X86::VRANGESSZrmi:
5267   case X86::VRANGESSZrmikz:
5268   case X86::VRANGESSZrri:
5269   case X86::VRANGESSZrrib:
5270   case X86::VRANGESSZrribkz:
5271   case X86::VRANGESSZrrikz:
5272     return Subtarget.hasRANGEFalseDeps();
5273   case X86::VGETMANTSSZrmi:
5274   case X86::VGETMANTSSZrmikz:
5275   case X86::VGETMANTSSZrri:
5276   case X86::VGETMANTSSZrrib:
5277   case X86::VGETMANTSSZrribkz:
5278   case X86::VGETMANTSSZrrikz:
5279   case X86::VGETMANTSDZrmi:
5280   case X86::VGETMANTSDZrmikz:
5281   case X86::VGETMANTSDZrri:
5282   case X86::VGETMANTSDZrrib:
5283   case X86::VGETMANTSDZrribkz:
5284   case X86::VGETMANTSDZrrikz:
5285   case X86::VGETMANTSHZrmi:
5286   case X86::VGETMANTSHZrmikz:
5287   case X86::VGETMANTSHZrri:
5288   case X86::VGETMANTSHZrrib:
5289   case X86::VGETMANTSHZrribkz:
5290   case X86::VGETMANTSHZrrikz:
5291   case X86::VGETMANTPSZ128rmbi:
5292   case X86::VGETMANTPSZ128rmbikz:
5293   case X86::VGETMANTPSZ128rmi:
5294   case X86::VGETMANTPSZ128rmikz:
5295   case X86::VGETMANTPSZ256rmbi:
5296   case X86::VGETMANTPSZ256rmbikz:
5297   case X86::VGETMANTPSZ256rmi:
5298   case X86::VGETMANTPSZ256rmikz:
5299   case X86::VGETMANTPSZrmbi:
5300   case X86::VGETMANTPSZrmbikz:
5301   case X86::VGETMANTPSZrmi:
5302   case X86::VGETMANTPSZrmikz:
5303   case X86::VGETMANTPDZ128rmbi:
5304   case X86::VGETMANTPDZ128rmbikz:
5305   case X86::VGETMANTPDZ128rmi:
5306   case X86::VGETMANTPDZ128rmikz:
5307   case X86::VGETMANTPDZ256rmbi:
5308   case X86::VGETMANTPDZ256rmbikz:
5309   case X86::VGETMANTPDZ256rmi:
5310   case X86::VGETMANTPDZ256rmikz:
5311   case X86::VGETMANTPDZrmbi:
5312   case X86::VGETMANTPDZrmbikz:
5313   case X86::VGETMANTPDZrmi:
5314   case X86::VGETMANTPDZrmikz:
5315     return Subtarget.hasGETMANTFalseDeps();
5316   case X86::VPMULLQZ128rm:
5317   case X86::VPMULLQZ128rmb:
5318   case X86::VPMULLQZ128rmbkz:
5319   case X86::VPMULLQZ128rmkz:
5320   case X86::VPMULLQZ128rr:
5321   case X86::VPMULLQZ128rrkz:
5322   case X86::VPMULLQZ256rm:
5323   case X86::VPMULLQZ256rmb:
5324   case X86::VPMULLQZ256rmbkz:
5325   case X86::VPMULLQZ256rmkz:
5326   case X86::VPMULLQZ256rr:
5327   case X86::VPMULLQZ256rrkz:
5328   case X86::VPMULLQZrm:
5329   case X86::VPMULLQZrmb:
5330   case X86::VPMULLQZrmbkz:
5331   case X86::VPMULLQZrmkz:
5332   case X86::VPMULLQZrr:
5333   case X86::VPMULLQZrrkz:
5334     return Subtarget.hasMULLQFalseDeps();
5335   // GPR
5336   case X86::POPCNT32rm:
5337   case X86::POPCNT32rr:
5338   case X86::POPCNT64rm:
5339   case X86::POPCNT64rr:
5340     return Subtarget.hasPOPCNTFalseDeps();
5341   case X86::LZCNT32rm:
5342   case X86::LZCNT32rr:
5343   case X86::LZCNT64rm:
5344   case X86::LZCNT64rr:
5345   case X86::TZCNT32rm:
5346   case X86::TZCNT32rr:
5347   case X86::TZCNT64rm:
5348   case X86::TZCNT64rr:
5349     return Subtarget.hasLZCNTFalseDeps();
5350   }
5351 
5352   return false;
5353 }
5354 
5355 /// Inform the BreakFalseDeps pass how many idle
5356 /// instructions we would like before a partial register update.
5357 unsigned X86InstrInfo::getPartialRegUpdateClearance(
5358     const MachineInstr &MI, unsigned OpNum,
5359     const TargetRegisterInfo *TRI) const {
5360   if (OpNum != 0 || !hasPartialRegUpdate(MI.getOpcode(), Subtarget))
5361     return 0;
5362 
5363   // If MI is marked as reading Reg, the partial register update is wanted.
5364   const MachineOperand &MO = MI.getOperand(0);
5365   Register Reg = MO.getReg();
5366   if (Reg.isVirtual()) {
5367     if (MO.readsReg() || MI.readsVirtualRegister(Reg))
5368       return 0;
5369   } else {
5370     if (MI.readsRegister(Reg, TRI))
5371       return 0;
5372   }
5373 
5374   // If any instructions in the clearance range are reading Reg, insert a
5375   // dependency breaking instruction, which is inexpensive and is likely to
5376   // be hidden in other instruction's cycles.
5377   return PartialRegUpdateClearance;
5378 }
5379 
5380 // Return true for any instruction the copies the high bits of the first source
5381 // operand into the unused high bits of the destination operand.
5382 // Also returns true for instructions that have two inputs where one may
5383 // be undef and we want it to use the same register as the other input.
5384 static bool hasUndefRegUpdate(unsigned Opcode, unsigned OpNum,
5385                               bool ForLoadFold = false) {
5386   // Set the OpNum parameter to the first source operand.
5387   switch (Opcode) {
5388   case X86::MMX_PUNPCKHBWrr:
5389   case X86::MMX_PUNPCKHWDrr:
5390   case X86::MMX_PUNPCKHDQrr:
5391   case X86::MMX_PUNPCKLBWrr:
5392   case X86::MMX_PUNPCKLWDrr:
5393   case X86::MMX_PUNPCKLDQrr:
5394   case X86::MOVHLPSrr:
5395   case X86::PACKSSWBrr:
5396   case X86::PACKUSWBrr:
5397   case X86::PACKSSDWrr:
5398   case X86::PACKUSDWrr:
5399   case X86::PUNPCKHBWrr:
5400   case X86::PUNPCKLBWrr:
5401   case X86::PUNPCKHWDrr:
5402   case X86::PUNPCKLWDrr:
5403   case X86::PUNPCKHDQrr:
5404   case X86::PUNPCKLDQrr:
5405   case X86::PUNPCKHQDQrr:
5406   case X86::PUNPCKLQDQrr:
5407   case X86::SHUFPDrri:
5408   case X86::SHUFPSrri:
5409     // These instructions are sometimes used with an undef first or second
5410     // source. Return true here so BreakFalseDeps will assign this source to the
5411     // same register as the first source to avoid a false dependency.
5412     // Operand 1 of these instructions is tied so they're separate from their
5413     // VEX counterparts.
5414     return OpNum == 2 && !ForLoadFold;
5415 
5416   case X86::VMOVLHPSrr:
5417   case X86::VMOVLHPSZrr:
5418   case X86::VPACKSSWBrr:
5419   case X86::VPACKUSWBrr:
5420   case X86::VPACKSSDWrr:
5421   case X86::VPACKUSDWrr:
5422   case X86::VPACKSSWBZ128rr:
5423   case X86::VPACKUSWBZ128rr:
5424   case X86::VPACKSSDWZ128rr:
5425   case X86::VPACKUSDWZ128rr:
5426   case X86::VPERM2F128rr:
5427   case X86::VPERM2I128rr:
5428   case X86::VSHUFF32X4Z256rri:
5429   case X86::VSHUFF32X4Zrri:
5430   case X86::VSHUFF64X2Z256rri:
5431   case X86::VSHUFF64X2Zrri:
5432   case X86::VSHUFI32X4Z256rri:
5433   case X86::VSHUFI32X4Zrri:
5434   case X86::VSHUFI64X2Z256rri:
5435   case X86::VSHUFI64X2Zrri:
5436   case X86::VPUNPCKHBWrr:
5437   case X86::VPUNPCKLBWrr:
5438   case X86::VPUNPCKHBWYrr:
5439   case X86::VPUNPCKLBWYrr:
5440   case X86::VPUNPCKHBWZ128rr:
5441   case X86::VPUNPCKLBWZ128rr:
5442   case X86::VPUNPCKHBWZ256rr:
5443   case X86::VPUNPCKLBWZ256rr:
5444   case X86::VPUNPCKHBWZrr:
5445   case X86::VPUNPCKLBWZrr:
5446   case X86::VPUNPCKHWDrr:
5447   case X86::VPUNPCKLWDrr:
5448   case X86::VPUNPCKHWDYrr:
5449   case X86::VPUNPCKLWDYrr:
5450   case X86::VPUNPCKHWDZ128rr:
5451   case X86::VPUNPCKLWDZ128rr:
5452   case X86::VPUNPCKHWDZ256rr:
5453   case X86::VPUNPCKLWDZ256rr:
5454   case X86::VPUNPCKHWDZrr:
5455   case X86::VPUNPCKLWDZrr:
5456   case X86::VPUNPCKHDQrr:
5457   case X86::VPUNPCKLDQrr:
5458   case X86::VPUNPCKHDQYrr:
5459   case X86::VPUNPCKLDQYrr:
5460   case X86::VPUNPCKHDQZ128rr:
5461   case X86::VPUNPCKLDQZ128rr:
5462   case X86::VPUNPCKHDQZ256rr:
5463   case X86::VPUNPCKLDQZ256rr:
5464   case X86::VPUNPCKHDQZrr:
5465   case X86::VPUNPCKLDQZrr:
5466   case X86::VPUNPCKHQDQrr:
5467   case X86::VPUNPCKLQDQrr:
5468   case X86::VPUNPCKHQDQYrr:
5469   case X86::VPUNPCKLQDQYrr:
5470   case X86::VPUNPCKHQDQZ128rr:
5471   case X86::VPUNPCKLQDQZ128rr:
5472   case X86::VPUNPCKHQDQZ256rr:
5473   case X86::VPUNPCKLQDQZ256rr:
5474   case X86::VPUNPCKHQDQZrr:
5475   case X86::VPUNPCKLQDQZrr:
5476     // These instructions are sometimes used with an undef first or second
5477     // source. Return true here so BreakFalseDeps will assign this source to the
5478     // same register as the first source to avoid a false dependency.
5479     return (OpNum == 1 || OpNum == 2) && !ForLoadFold;
5480 
5481   case X86::VCVTSI2SSrr:
5482   case X86::VCVTSI2SSrm:
5483   case X86::VCVTSI2SSrr_Int:
5484   case X86::VCVTSI2SSrm_Int:
5485   case X86::VCVTSI642SSrr:
5486   case X86::VCVTSI642SSrm:
5487   case X86::VCVTSI642SSrr_Int:
5488   case X86::VCVTSI642SSrm_Int:
5489   case X86::VCVTSI2SDrr:
5490   case X86::VCVTSI2SDrm:
5491   case X86::VCVTSI2SDrr_Int:
5492   case X86::VCVTSI2SDrm_Int:
5493   case X86::VCVTSI642SDrr:
5494   case X86::VCVTSI642SDrm:
5495   case X86::VCVTSI642SDrr_Int:
5496   case X86::VCVTSI642SDrm_Int:
5497   // AVX-512
5498   case X86::VCVTSI2SSZrr:
5499   case X86::VCVTSI2SSZrm:
5500   case X86::VCVTSI2SSZrr_Int:
5501   case X86::VCVTSI2SSZrrb_Int:
5502   case X86::VCVTSI2SSZrm_Int:
5503   case X86::VCVTSI642SSZrr:
5504   case X86::VCVTSI642SSZrm:
5505   case X86::VCVTSI642SSZrr_Int:
5506   case X86::VCVTSI642SSZrrb_Int:
5507   case X86::VCVTSI642SSZrm_Int:
5508   case X86::VCVTSI2SDZrr:
5509   case X86::VCVTSI2SDZrm:
5510   case X86::VCVTSI2SDZrr_Int:
5511   case X86::VCVTSI2SDZrm_Int:
5512   case X86::VCVTSI642SDZrr:
5513   case X86::VCVTSI642SDZrm:
5514   case X86::VCVTSI642SDZrr_Int:
5515   case X86::VCVTSI642SDZrrb_Int:
5516   case X86::VCVTSI642SDZrm_Int:
5517   case X86::VCVTUSI2SSZrr:
5518   case X86::VCVTUSI2SSZrm:
5519   case X86::VCVTUSI2SSZrr_Int:
5520   case X86::VCVTUSI2SSZrrb_Int:
5521   case X86::VCVTUSI2SSZrm_Int:
5522   case X86::VCVTUSI642SSZrr:
5523   case X86::VCVTUSI642SSZrm:
5524   case X86::VCVTUSI642SSZrr_Int:
5525   case X86::VCVTUSI642SSZrrb_Int:
5526   case X86::VCVTUSI642SSZrm_Int:
5527   case X86::VCVTUSI2SDZrr:
5528   case X86::VCVTUSI2SDZrm:
5529   case X86::VCVTUSI2SDZrr_Int:
5530   case X86::VCVTUSI2SDZrm_Int:
5531   case X86::VCVTUSI642SDZrr:
5532   case X86::VCVTUSI642SDZrm:
5533   case X86::VCVTUSI642SDZrr_Int:
5534   case X86::VCVTUSI642SDZrrb_Int:
5535   case X86::VCVTUSI642SDZrm_Int:
5536   case X86::VCVTSI2SHZrr:
5537   case X86::VCVTSI2SHZrm:
5538   case X86::VCVTSI2SHZrr_Int:
5539   case X86::VCVTSI2SHZrrb_Int:
5540   case X86::VCVTSI2SHZrm_Int:
5541   case X86::VCVTSI642SHZrr:
5542   case X86::VCVTSI642SHZrm:
5543   case X86::VCVTSI642SHZrr_Int:
5544   case X86::VCVTSI642SHZrrb_Int:
5545   case X86::VCVTSI642SHZrm_Int:
5546   case X86::VCVTUSI2SHZrr:
5547   case X86::VCVTUSI2SHZrm:
5548   case X86::VCVTUSI2SHZrr_Int:
5549   case X86::VCVTUSI2SHZrrb_Int:
5550   case X86::VCVTUSI2SHZrm_Int:
5551   case X86::VCVTUSI642SHZrr:
5552   case X86::VCVTUSI642SHZrm:
5553   case X86::VCVTUSI642SHZrr_Int:
5554   case X86::VCVTUSI642SHZrrb_Int:
5555   case X86::VCVTUSI642SHZrm_Int:
5556     // Load folding won't effect the undef register update since the input is
5557     // a GPR.
5558     return OpNum == 1 && !ForLoadFold;
5559   case X86::VCVTSD2SSrr:
5560   case X86::VCVTSD2SSrm:
5561   case X86::VCVTSD2SSrr_Int:
5562   case X86::VCVTSD2SSrm_Int:
5563   case X86::VCVTSS2SDrr:
5564   case X86::VCVTSS2SDrm:
5565   case X86::VCVTSS2SDrr_Int:
5566   case X86::VCVTSS2SDrm_Int:
5567   case X86::VRCPSSr:
5568   case X86::VRCPSSr_Int:
5569   case X86::VRCPSSm:
5570   case X86::VRCPSSm_Int:
5571   case X86::VROUNDSDr:
5572   case X86::VROUNDSDm:
5573   case X86::VROUNDSDr_Int:
5574   case X86::VROUNDSDm_Int:
5575   case X86::VROUNDSSr:
5576   case X86::VROUNDSSm:
5577   case X86::VROUNDSSr_Int:
5578   case X86::VROUNDSSm_Int:
5579   case X86::VRSQRTSSr:
5580   case X86::VRSQRTSSr_Int:
5581   case X86::VRSQRTSSm:
5582   case X86::VRSQRTSSm_Int:
5583   case X86::VSQRTSSr:
5584   case X86::VSQRTSSr_Int:
5585   case X86::VSQRTSSm:
5586   case X86::VSQRTSSm_Int:
5587   case X86::VSQRTSDr:
5588   case X86::VSQRTSDr_Int:
5589   case X86::VSQRTSDm:
5590   case X86::VSQRTSDm_Int:
5591   // AVX-512
5592   case X86::VCVTSD2SSZrr:
5593   case X86::VCVTSD2SSZrr_Int:
5594   case X86::VCVTSD2SSZrrb_Int:
5595   case X86::VCVTSD2SSZrm:
5596   case X86::VCVTSD2SSZrm_Int:
5597   case X86::VCVTSS2SDZrr:
5598   case X86::VCVTSS2SDZrr_Int:
5599   case X86::VCVTSS2SDZrrb_Int:
5600   case X86::VCVTSS2SDZrm:
5601   case X86::VCVTSS2SDZrm_Int:
5602   case X86::VGETEXPSDZr:
5603   case X86::VGETEXPSDZrb:
5604   case X86::VGETEXPSDZm:
5605   case X86::VGETEXPSSZr:
5606   case X86::VGETEXPSSZrb:
5607   case X86::VGETEXPSSZm:
5608   case X86::VGETMANTSDZrri:
5609   case X86::VGETMANTSDZrrib:
5610   case X86::VGETMANTSDZrmi:
5611   case X86::VGETMANTSSZrri:
5612   case X86::VGETMANTSSZrrib:
5613   case X86::VGETMANTSSZrmi:
5614   case X86::VRNDSCALESDZr:
5615   case X86::VRNDSCALESDZr_Int:
5616   case X86::VRNDSCALESDZrb_Int:
5617   case X86::VRNDSCALESDZm:
5618   case X86::VRNDSCALESDZm_Int:
5619   case X86::VRNDSCALESSZr:
5620   case X86::VRNDSCALESSZr_Int:
5621   case X86::VRNDSCALESSZrb_Int:
5622   case X86::VRNDSCALESSZm:
5623   case X86::VRNDSCALESSZm_Int:
5624   case X86::VRCP14SDZrr:
5625   case X86::VRCP14SDZrm:
5626   case X86::VRCP14SSZrr:
5627   case X86::VRCP14SSZrm:
5628   case X86::VRCPSHZrr:
5629   case X86::VRCPSHZrm:
5630   case X86::VRSQRTSHZrr:
5631   case X86::VRSQRTSHZrm:
5632   case X86::VREDUCESHZrmi:
5633   case X86::VREDUCESHZrri:
5634   case X86::VREDUCESHZrrib:
5635   case X86::VGETEXPSHZr:
5636   case X86::VGETEXPSHZrb:
5637   case X86::VGETEXPSHZm:
5638   case X86::VGETMANTSHZrri:
5639   case X86::VGETMANTSHZrrib:
5640   case X86::VGETMANTSHZrmi:
5641   case X86::VRNDSCALESHZr:
5642   case X86::VRNDSCALESHZr_Int:
5643   case X86::VRNDSCALESHZrb_Int:
5644   case X86::VRNDSCALESHZm:
5645   case X86::VRNDSCALESHZm_Int:
5646   case X86::VSQRTSHZr:
5647   case X86::VSQRTSHZr_Int:
5648   case X86::VSQRTSHZrb_Int:
5649   case X86::VSQRTSHZm:
5650   case X86::VSQRTSHZm_Int:
5651   case X86::VRCP28SDZr:
5652   case X86::VRCP28SDZrb:
5653   case X86::VRCP28SDZm:
5654   case X86::VRCP28SSZr:
5655   case X86::VRCP28SSZrb:
5656   case X86::VRCP28SSZm:
5657   case X86::VREDUCESSZrmi:
5658   case X86::VREDUCESSZrri:
5659   case X86::VREDUCESSZrrib:
5660   case X86::VRSQRT14SDZrr:
5661   case X86::VRSQRT14SDZrm:
5662   case X86::VRSQRT14SSZrr:
5663   case X86::VRSQRT14SSZrm:
5664   case X86::VRSQRT28SDZr:
5665   case X86::VRSQRT28SDZrb:
5666   case X86::VRSQRT28SDZm:
5667   case X86::VRSQRT28SSZr:
5668   case X86::VRSQRT28SSZrb:
5669   case X86::VRSQRT28SSZm:
5670   case X86::VSQRTSSZr:
5671   case X86::VSQRTSSZr_Int:
5672   case X86::VSQRTSSZrb_Int:
5673   case X86::VSQRTSSZm:
5674   case X86::VSQRTSSZm_Int:
5675   case X86::VSQRTSDZr:
5676   case X86::VSQRTSDZr_Int:
5677   case X86::VSQRTSDZrb_Int:
5678   case X86::VSQRTSDZm:
5679   case X86::VSQRTSDZm_Int:
5680   case X86::VCVTSD2SHZrr:
5681   case X86::VCVTSD2SHZrr_Int:
5682   case X86::VCVTSD2SHZrrb_Int:
5683   case X86::VCVTSD2SHZrm:
5684   case X86::VCVTSD2SHZrm_Int:
5685   case X86::VCVTSS2SHZrr:
5686   case X86::VCVTSS2SHZrr_Int:
5687   case X86::VCVTSS2SHZrrb_Int:
5688   case X86::VCVTSS2SHZrm:
5689   case X86::VCVTSS2SHZrm_Int:
5690   case X86::VCVTSH2SDZrr:
5691   case X86::VCVTSH2SDZrr_Int:
5692   case X86::VCVTSH2SDZrrb_Int:
5693   case X86::VCVTSH2SDZrm:
5694   case X86::VCVTSH2SDZrm_Int:
5695   case X86::VCVTSH2SSZrr:
5696   case X86::VCVTSH2SSZrr_Int:
5697   case X86::VCVTSH2SSZrrb_Int:
5698   case X86::VCVTSH2SSZrm:
5699   case X86::VCVTSH2SSZrm_Int:
5700     return OpNum == 1;
5701   case X86::VMOVSSZrrk:
5702   case X86::VMOVSDZrrk:
5703     return OpNum == 3 && !ForLoadFold;
5704   case X86::VMOVSSZrrkz:
5705   case X86::VMOVSDZrrkz:
5706     return OpNum == 2 && !ForLoadFold;
5707   }
5708 
5709   return false;
5710 }
5711 
5712 /// Inform the BreakFalseDeps pass how many idle instructions we would like
5713 /// before certain undef register reads.
5714 ///
5715 /// This catches the VCVTSI2SD family of instructions:
5716 ///
5717 /// vcvtsi2sdq %rax, undef %xmm0, %xmm14
5718 ///
5719 /// We should to be careful *not* to catch VXOR idioms which are presumably
5720 /// handled specially in the pipeline:
5721 ///
5722 /// vxorps undef %xmm1, undef %xmm1, %xmm1
5723 ///
5724 /// Like getPartialRegUpdateClearance, this makes a strong assumption that the
5725 /// high bits that are passed-through are not live.
5726 unsigned
5727 X86InstrInfo::getUndefRegClearance(const MachineInstr &MI, unsigned OpNum,
5728                                    const TargetRegisterInfo *TRI) const {
5729   const MachineOperand &MO = MI.getOperand(OpNum);
5730   if (Register::isPhysicalRegister(MO.getReg()) &&
5731       hasUndefRegUpdate(MI.getOpcode(), OpNum))
5732     return UndefRegClearance;
5733 
5734   return 0;
5735 }
5736 
5737 void X86InstrInfo::breakPartialRegDependency(
5738     MachineInstr &MI, unsigned OpNum, const TargetRegisterInfo *TRI) const {
5739   Register Reg = MI.getOperand(OpNum).getReg();
5740   // If MI kills this register, the false dependence is already broken.
5741   if (MI.killsRegister(Reg, TRI))
5742     return;
5743 
5744   if (X86::VR128RegClass.contains(Reg)) {
5745     // These instructions are all floating point domain, so xorps is the best
5746     // choice.
5747     unsigned Opc = Subtarget.hasAVX() ? X86::VXORPSrr : X86::XORPSrr;
5748     BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), get(Opc), Reg)
5749         .addReg(Reg, RegState::Undef)
5750         .addReg(Reg, RegState::Undef);
5751     MI.addRegisterKilled(Reg, TRI, true);
5752   } else if (X86::VR256RegClass.contains(Reg)) {
5753     // Use vxorps to clear the full ymm register.
5754     // It wants to read and write the xmm sub-register.
5755     Register XReg = TRI->getSubReg(Reg, X86::sub_xmm);
5756     BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), get(X86::VXORPSrr), XReg)
5757         .addReg(XReg, RegState::Undef)
5758         .addReg(XReg, RegState::Undef)
5759         .addReg(Reg, RegState::ImplicitDefine);
5760     MI.addRegisterKilled(Reg, TRI, true);
5761   } else if (X86::VR128XRegClass.contains(Reg)) {
5762     // Only handle VLX targets.
5763     if (!Subtarget.hasVLX())
5764       return;
5765     // Since vxorps requires AVX512DQ, vpxord should be the best choice.
5766     BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), get(X86::VPXORDZ128rr), Reg)
5767         .addReg(Reg, RegState::Undef)
5768         .addReg(Reg, RegState::Undef);
5769     MI.addRegisterKilled(Reg, TRI, true);
5770   } else if (X86::VR256XRegClass.contains(Reg) ||
5771              X86::VR512RegClass.contains(Reg)) {
5772     // Only handle VLX targets.
5773     if (!Subtarget.hasVLX())
5774       return;
5775     // Use vpxord to clear the full ymm/zmm register.
5776     // It wants to read and write the xmm sub-register.
5777     Register XReg = TRI->getSubReg(Reg, X86::sub_xmm);
5778     BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), get(X86::VPXORDZ128rr), XReg)
5779         .addReg(XReg, RegState::Undef)
5780         .addReg(XReg, RegState::Undef)
5781         .addReg(Reg, RegState::ImplicitDefine);
5782     MI.addRegisterKilled(Reg, TRI, true);
5783   } else if (X86::GR64RegClass.contains(Reg)) {
5784     // Using XOR32rr because it has shorter encoding and zeros up the upper bits
5785     // as well.
5786     Register XReg = TRI->getSubReg(Reg, X86::sub_32bit);
5787     BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), get(X86::XOR32rr), XReg)
5788         .addReg(XReg, RegState::Undef)
5789         .addReg(XReg, RegState::Undef)
5790         .addReg(Reg, RegState::ImplicitDefine);
5791     MI.addRegisterKilled(Reg, TRI, true);
5792   } else if (X86::GR32RegClass.contains(Reg)) {
5793     BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), get(X86::XOR32rr), Reg)
5794         .addReg(Reg, RegState::Undef)
5795         .addReg(Reg, RegState::Undef);
5796     MI.addRegisterKilled(Reg, TRI, true);
5797   }
5798 }
5799 
5800 static void addOperands(MachineInstrBuilder &MIB, ArrayRef<MachineOperand> MOs,
5801                         int PtrOffset = 0) {
5802   unsigned NumAddrOps = MOs.size();
5803 
5804   if (NumAddrOps < 4) {
5805     // FrameIndex only - add an immediate offset (whether its zero or not).
5806     for (unsigned i = 0; i != NumAddrOps; ++i)
5807       MIB.add(MOs[i]);
5808     addOffset(MIB, PtrOffset);
5809   } else {
5810     // General Memory Addressing - we need to add any offset to an existing
5811     // offset.
5812     assert(MOs.size() == 5 && "Unexpected memory operand list length");
5813     for (unsigned i = 0; i != NumAddrOps; ++i) {
5814       const MachineOperand &MO = MOs[i];
5815       if (i == 3 && PtrOffset != 0) {
5816         MIB.addDisp(MO, PtrOffset);
5817       } else {
5818         MIB.add(MO);
5819       }
5820     }
5821   }
5822 }
5823 
5824 static void updateOperandRegConstraints(MachineFunction &MF,
5825                                         MachineInstr &NewMI,
5826                                         const TargetInstrInfo &TII) {
5827   MachineRegisterInfo &MRI = MF.getRegInfo();
5828   const TargetRegisterInfo &TRI = *MRI.getTargetRegisterInfo();
5829 
5830   for (int Idx : llvm::seq<int>(0, NewMI.getNumOperands())) {
5831     MachineOperand &MO = NewMI.getOperand(Idx);
5832     // We only need to update constraints on virtual register operands.
5833     if (!MO.isReg())
5834       continue;
5835     Register Reg = MO.getReg();
5836     if (!Reg.isVirtual())
5837       continue;
5838 
5839     auto *NewRC = MRI.constrainRegClass(
5840         Reg, TII.getRegClass(NewMI.getDesc(), Idx, &TRI, MF));
5841     if (!NewRC) {
5842       LLVM_DEBUG(
5843           dbgs() << "WARNING: Unable to update register constraint for operand "
5844                  << Idx << " of instruction:\n";
5845           NewMI.dump(); dbgs() << "\n");
5846     }
5847   }
5848 }
5849 
5850 static MachineInstr *FuseTwoAddrInst(MachineFunction &MF, unsigned Opcode,
5851                                      ArrayRef<MachineOperand> MOs,
5852                                      MachineBasicBlock::iterator InsertPt,
5853                                      MachineInstr &MI,
5854                                      const TargetInstrInfo &TII) {
5855   // Create the base instruction with the memory operand as the first part.
5856   // Omit the implicit operands, something BuildMI can't do.
5857   MachineInstr *NewMI =
5858       MF.CreateMachineInstr(TII.get(Opcode), MI.getDebugLoc(), true);
5859   MachineInstrBuilder MIB(MF, NewMI);
5860   addOperands(MIB, MOs);
5861 
5862   // Loop over the rest of the ri operands, converting them over.
5863   unsigned NumOps = MI.getDesc().getNumOperands() - 2;
5864   for (unsigned i = 0; i != NumOps; ++i) {
5865     MachineOperand &MO = MI.getOperand(i + 2);
5866     MIB.add(MO);
5867   }
5868   for (const MachineOperand &MO : llvm::drop_begin(MI.operands(), NumOps + 2))
5869     MIB.add(MO);
5870 
5871   updateOperandRegConstraints(MF, *NewMI, TII);
5872 
5873   MachineBasicBlock *MBB = InsertPt->getParent();
5874   MBB->insert(InsertPt, NewMI);
5875 
5876   return MIB;
5877 }
5878 
5879 static MachineInstr *FuseInst(MachineFunction &MF, unsigned Opcode,
5880                               unsigned OpNo, ArrayRef<MachineOperand> MOs,
5881                               MachineBasicBlock::iterator InsertPt,
5882                               MachineInstr &MI, const TargetInstrInfo &TII,
5883                               int PtrOffset = 0) {
5884   // Omit the implicit operands, something BuildMI can't do.
5885   MachineInstr *NewMI =
5886       MF.CreateMachineInstr(TII.get(Opcode), MI.getDebugLoc(), true);
5887   MachineInstrBuilder MIB(MF, NewMI);
5888 
5889   for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
5890     MachineOperand &MO = MI.getOperand(i);
5891     if (i == OpNo) {
5892       assert(MO.isReg() && "Expected to fold into reg operand!");
5893       addOperands(MIB, MOs, PtrOffset);
5894     } else {
5895       MIB.add(MO);
5896     }
5897   }
5898 
5899   updateOperandRegConstraints(MF, *NewMI, TII);
5900 
5901   // Copy the NoFPExcept flag from the instruction we're fusing.
5902   if (MI.getFlag(MachineInstr::MIFlag::NoFPExcept))
5903     NewMI->setFlag(MachineInstr::MIFlag::NoFPExcept);
5904 
5905   MachineBasicBlock *MBB = InsertPt->getParent();
5906   MBB->insert(InsertPt, NewMI);
5907 
5908   return MIB;
5909 }
5910 
5911 static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode,
5912                                 ArrayRef<MachineOperand> MOs,
5913                                 MachineBasicBlock::iterator InsertPt,
5914                                 MachineInstr &MI) {
5915   MachineInstrBuilder MIB = BuildMI(*InsertPt->getParent(), InsertPt,
5916                                     MI.getDebugLoc(), TII.get(Opcode));
5917   addOperands(MIB, MOs);
5918   return MIB.addImm(0);
5919 }
5920 
5921 MachineInstr *X86InstrInfo::foldMemoryOperandCustom(
5922     MachineFunction &MF, MachineInstr &MI, unsigned OpNum,
5923     ArrayRef<MachineOperand> MOs, MachineBasicBlock::iterator InsertPt,
5924     unsigned Size, Align Alignment) const {
5925   switch (MI.getOpcode()) {
5926   case X86::INSERTPSrr:
5927   case X86::VINSERTPSrr:
5928   case X86::VINSERTPSZrr:
5929     // Attempt to convert the load of inserted vector into a fold load
5930     // of a single float.
5931     if (OpNum == 2) {
5932       unsigned Imm = MI.getOperand(MI.getNumOperands() - 1).getImm();
5933       unsigned ZMask = Imm & 15;
5934       unsigned DstIdx = (Imm >> 4) & 3;
5935       unsigned SrcIdx = (Imm >> 6) & 3;
5936 
5937       const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
5938       const TargetRegisterClass *RC = getRegClass(MI.getDesc(), OpNum, &RI, MF);
5939       unsigned RCSize = TRI.getRegSizeInBits(*RC) / 8;
5940       if ((Size == 0 || Size >= 16) && RCSize >= 16 && Alignment >= Align(4)) {
5941         int PtrOffset = SrcIdx * 4;
5942         unsigned NewImm = (DstIdx << 4) | ZMask;
5943         unsigned NewOpCode =
5944             (MI.getOpcode() == X86::VINSERTPSZrr) ? X86::VINSERTPSZrm :
5945             (MI.getOpcode() == X86::VINSERTPSrr)  ? X86::VINSERTPSrm  :
5946                                                     X86::INSERTPSrm;
5947         MachineInstr *NewMI =
5948             FuseInst(MF, NewOpCode, OpNum, MOs, InsertPt, MI, *this, PtrOffset);
5949         NewMI->getOperand(NewMI->getNumOperands() - 1).setImm(NewImm);
5950         return NewMI;
5951       }
5952     }
5953     break;
5954   case X86::MOVHLPSrr:
5955   case X86::VMOVHLPSrr:
5956   case X86::VMOVHLPSZrr:
5957     // Move the upper 64-bits of the second operand to the lower 64-bits.
5958     // To fold the load, adjust the pointer to the upper and use (V)MOVLPS.
5959     // TODO: In most cases AVX doesn't have a 8-byte alignment requirement.
5960     if (OpNum == 2) {
5961       const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
5962       const TargetRegisterClass *RC = getRegClass(MI.getDesc(), OpNum, &RI, MF);
5963       unsigned RCSize = TRI.getRegSizeInBits(*RC) / 8;
5964       if ((Size == 0 || Size >= 16) && RCSize >= 16 && Alignment >= Align(8)) {
5965         unsigned NewOpCode =
5966             (MI.getOpcode() == X86::VMOVHLPSZrr) ? X86::VMOVLPSZ128rm :
5967             (MI.getOpcode() == X86::VMOVHLPSrr)  ? X86::VMOVLPSrm     :
5968                                                    X86::MOVLPSrm;
5969         MachineInstr *NewMI =
5970             FuseInst(MF, NewOpCode, OpNum, MOs, InsertPt, MI, *this, 8);
5971         return NewMI;
5972       }
5973     }
5974     break;
5975   case X86::UNPCKLPDrr:
5976     // If we won't be able to fold this to the memory form of UNPCKL, use
5977     // MOVHPD instead. Done as custom because we can't have this in the load
5978     // table twice.
5979     if (OpNum == 2) {
5980       const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
5981       const TargetRegisterClass *RC = getRegClass(MI.getDesc(), OpNum, &RI, MF);
5982       unsigned RCSize = TRI.getRegSizeInBits(*RC) / 8;
5983       if ((Size == 0 || Size >= 16) && RCSize >= 16 && Alignment < Align(16)) {
5984         MachineInstr *NewMI =
5985             FuseInst(MF, X86::MOVHPDrm, OpNum, MOs, InsertPt, MI, *this);
5986         return NewMI;
5987       }
5988     }
5989     break;
5990   }
5991 
5992   return nullptr;
5993 }
5994 
5995 static bool shouldPreventUndefRegUpdateMemFold(MachineFunction &MF,
5996                                                MachineInstr &MI) {
5997   if (!hasUndefRegUpdate(MI.getOpcode(), 1, /*ForLoadFold*/true) ||
5998       !MI.getOperand(1).isReg())
5999     return false;
6000 
6001   // The are two cases we need to handle depending on where in the pipeline
6002   // the folding attempt is being made.
6003   // -Register has the undef flag set.
6004   // -Register is produced by the IMPLICIT_DEF instruction.
6005 
6006   if (MI.getOperand(1).isUndef())
6007     return true;
6008 
6009   MachineRegisterInfo &RegInfo = MF.getRegInfo();
6010   MachineInstr *VRegDef = RegInfo.getUniqueVRegDef(MI.getOperand(1).getReg());
6011   return VRegDef && VRegDef->isImplicitDef();
6012 }
6013 
6014 MachineInstr *X86InstrInfo::foldMemoryOperandImpl(
6015     MachineFunction &MF, MachineInstr &MI, unsigned OpNum,
6016     ArrayRef<MachineOperand> MOs, MachineBasicBlock::iterator InsertPt,
6017     unsigned Size, Align Alignment, bool AllowCommute) const {
6018   bool isSlowTwoMemOps = Subtarget.slowTwoMemOps();
6019   bool isTwoAddrFold = false;
6020 
6021   // For CPUs that favor the register form of a call or push,
6022   // do not fold loads into calls or pushes, unless optimizing for size
6023   // aggressively.
6024   if (isSlowTwoMemOps && !MF.getFunction().hasMinSize() &&
6025       (MI.getOpcode() == X86::CALL32r || MI.getOpcode() == X86::CALL64r ||
6026        MI.getOpcode() == X86::PUSH16r || MI.getOpcode() == X86::PUSH32r ||
6027        MI.getOpcode() == X86::PUSH64r))
6028     return nullptr;
6029 
6030   // Avoid partial and undef register update stalls unless optimizing for size.
6031   if (!MF.getFunction().hasOptSize() &&
6032       (hasPartialRegUpdate(MI.getOpcode(), Subtarget, /*ForLoadFold*/true) ||
6033        shouldPreventUndefRegUpdateMemFold(MF, MI)))
6034     return nullptr;
6035 
6036   unsigned NumOps = MI.getDesc().getNumOperands();
6037   bool isTwoAddr =
6038       NumOps > 1 && MI.getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
6039 
6040   // FIXME: AsmPrinter doesn't know how to handle
6041   // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
6042   if (MI.getOpcode() == X86::ADD32ri &&
6043       MI.getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
6044     return nullptr;
6045 
6046   // GOTTPOFF relocation loads can only be folded into add instructions.
6047   // FIXME: Need to exclude other relocations that only support specific
6048   // instructions.
6049   if (MOs.size() == X86::AddrNumOperands &&
6050       MOs[X86::AddrDisp].getTargetFlags() == X86II::MO_GOTTPOFF &&
6051       MI.getOpcode() != X86::ADD64rr)
6052     return nullptr;
6053 
6054   MachineInstr *NewMI = nullptr;
6055 
6056   // Attempt to fold any custom cases we have.
6057   if (MachineInstr *CustomMI = foldMemoryOperandCustom(
6058           MF, MI, OpNum, MOs, InsertPt, Size, Alignment))
6059     return CustomMI;
6060 
6061   const X86MemoryFoldTableEntry *I = nullptr;
6062 
6063   // Folding a memory location into the two-address part of a two-address
6064   // instruction is different than folding it other places.  It requires
6065   // replacing the *two* registers with the memory location.
6066   if (isTwoAddr && NumOps >= 2 && OpNum < 2 && MI.getOperand(0).isReg() &&
6067       MI.getOperand(1).isReg() &&
6068       MI.getOperand(0).getReg() == MI.getOperand(1).getReg()) {
6069     I = lookupTwoAddrFoldTable(MI.getOpcode());
6070     isTwoAddrFold = true;
6071   } else {
6072     if (OpNum == 0) {
6073       if (MI.getOpcode() == X86::MOV32r0) {
6074         NewMI = MakeM0Inst(*this, X86::MOV32mi, MOs, InsertPt, MI);
6075         if (NewMI)
6076           return NewMI;
6077       }
6078     }
6079 
6080     I = lookupFoldTable(MI.getOpcode(), OpNum);
6081   }
6082 
6083   if (I != nullptr) {
6084     unsigned Opcode = I->DstOp;
6085     bool FoldedLoad =
6086         isTwoAddrFold || (OpNum == 0 && I->Flags & TB_FOLDED_LOAD) || OpNum > 0;
6087     bool FoldedStore =
6088         isTwoAddrFold || (OpNum == 0 && I->Flags & TB_FOLDED_STORE);
6089     MaybeAlign MinAlign =
6090         decodeMaybeAlign((I->Flags & TB_ALIGN_MASK) >> TB_ALIGN_SHIFT);
6091     if (MinAlign && Alignment < *MinAlign)
6092       return nullptr;
6093     bool NarrowToMOV32rm = false;
6094     if (Size) {
6095       const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
6096       const TargetRegisterClass *RC = getRegClass(MI.getDesc(), OpNum,
6097                                                   &RI, MF);
6098       unsigned RCSize = TRI.getRegSizeInBits(*RC) / 8;
6099       // Check if it's safe to fold the load. If the size of the object is
6100       // narrower than the load width, then it's not.
6101       // FIXME: Allow scalar intrinsic instructions like ADDSSrm_Int.
6102       if (FoldedLoad && Size < RCSize) {
6103         // If this is a 64-bit load, but the spill slot is 32, then we can do
6104         // a 32-bit load which is implicitly zero-extended. This likely is
6105         // due to live interval analysis remat'ing a load from stack slot.
6106         if (Opcode != X86::MOV64rm || RCSize != 8 || Size != 4)
6107           return nullptr;
6108         if (MI.getOperand(0).getSubReg() || MI.getOperand(1).getSubReg())
6109           return nullptr;
6110         Opcode = X86::MOV32rm;
6111         NarrowToMOV32rm = true;
6112       }
6113       // For stores, make sure the size of the object is equal to the size of
6114       // the store. If the object is larger, the extra bits would be garbage. If
6115       // the object is smaller we might overwrite another object or fault.
6116       if (FoldedStore && Size != RCSize)
6117         return nullptr;
6118     }
6119 
6120     if (isTwoAddrFold)
6121       NewMI = FuseTwoAddrInst(MF, Opcode, MOs, InsertPt, MI, *this);
6122     else
6123       NewMI = FuseInst(MF, Opcode, OpNum, MOs, InsertPt, MI, *this);
6124 
6125     if (NarrowToMOV32rm) {
6126       // If this is the special case where we use a MOV32rm to load a 32-bit
6127       // value and zero-extend the top bits. Change the destination register
6128       // to a 32-bit one.
6129       Register DstReg = NewMI->getOperand(0).getReg();
6130       if (DstReg.isPhysical())
6131         NewMI->getOperand(0).setReg(RI.getSubReg(DstReg, X86::sub_32bit));
6132       else
6133         NewMI->getOperand(0).setSubReg(X86::sub_32bit);
6134     }
6135     return NewMI;
6136   }
6137 
6138   // If the instruction and target operand are commutable, commute the
6139   // instruction and try again.
6140   if (AllowCommute) {
6141     unsigned CommuteOpIdx1 = OpNum, CommuteOpIdx2 = CommuteAnyOperandIndex;
6142     if (findCommutedOpIndices(MI, CommuteOpIdx1, CommuteOpIdx2)) {
6143       bool HasDef = MI.getDesc().getNumDefs();
6144       Register Reg0 = HasDef ? MI.getOperand(0).getReg() : Register();
6145       Register Reg1 = MI.getOperand(CommuteOpIdx1).getReg();
6146       Register Reg2 = MI.getOperand(CommuteOpIdx2).getReg();
6147       bool Tied1 =
6148           0 == MI.getDesc().getOperandConstraint(CommuteOpIdx1, MCOI::TIED_TO);
6149       bool Tied2 =
6150           0 == MI.getDesc().getOperandConstraint(CommuteOpIdx2, MCOI::TIED_TO);
6151 
6152       // If either of the commutable operands are tied to the destination
6153       // then we can not commute + fold.
6154       if ((HasDef && Reg0 == Reg1 && Tied1) ||
6155           (HasDef && Reg0 == Reg2 && Tied2))
6156         return nullptr;
6157 
6158       MachineInstr *CommutedMI =
6159           commuteInstruction(MI, false, CommuteOpIdx1, CommuteOpIdx2);
6160       if (!CommutedMI) {
6161         // Unable to commute.
6162         return nullptr;
6163       }
6164       if (CommutedMI != &MI) {
6165         // New instruction. We can't fold from this.
6166         CommutedMI->eraseFromParent();
6167         return nullptr;
6168       }
6169 
6170       // Attempt to fold with the commuted version of the instruction.
6171       NewMI = foldMemoryOperandImpl(MF, MI, CommuteOpIdx2, MOs, InsertPt, Size,
6172                                     Alignment, /*AllowCommute=*/false);
6173       if (NewMI)
6174         return NewMI;
6175 
6176       // Folding failed again - undo the commute before returning.
6177       MachineInstr *UncommutedMI =
6178           commuteInstruction(MI, false, CommuteOpIdx1, CommuteOpIdx2);
6179       if (!UncommutedMI) {
6180         // Unable to commute.
6181         return nullptr;
6182       }
6183       if (UncommutedMI != &MI) {
6184         // New instruction. It doesn't need to be kept.
6185         UncommutedMI->eraseFromParent();
6186         return nullptr;
6187       }
6188 
6189       // Return here to prevent duplicate fuse failure report.
6190       return nullptr;
6191     }
6192   }
6193 
6194   // No fusion
6195   if (PrintFailedFusing && !MI.isCopy())
6196     dbgs() << "We failed to fuse operand " << OpNum << " in " << MI;
6197   return nullptr;
6198 }
6199 
6200 MachineInstr *
6201 X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF, MachineInstr &MI,
6202                                     ArrayRef<unsigned> Ops,
6203                                     MachineBasicBlock::iterator InsertPt,
6204                                     int FrameIndex, LiveIntervals *LIS,
6205                                     VirtRegMap *VRM) const {
6206   // Check switch flag
6207   if (NoFusing)
6208     return nullptr;
6209 
6210   // Avoid partial and undef register update stalls unless optimizing for size.
6211   if (!MF.getFunction().hasOptSize() &&
6212       (hasPartialRegUpdate(MI.getOpcode(), Subtarget, /*ForLoadFold*/true) ||
6213        shouldPreventUndefRegUpdateMemFold(MF, MI)))
6214     return nullptr;
6215 
6216   // Don't fold subreg spills, or reloads that use a high subreg.
6217   for (auto Op : Ops) {
6218     MachineOperand &MO = MI.getOperand(Op);
6219     auto SubReg = MO.getSubReg();
6220     if (SubReg && (MO.isDef() || SubReg == X86::sub_8bit_hi))
6221       return nullptr;
6222   }
6223 
6224   const MachineFrameInfo &MFI = MF.getFrameInfo();
6225   unsigned Size = MFI.getObjectSize(FrameIndex);
6226   Align Alignment = MFI.getObjectAlign(FrameIndex);
6227   // If the function stack isn't realigned we don't want to fold instructions
6228   // that need increased alignment.
6229   if (!RI.hasStackRealignment(MF))
6230     Alignment =
6231         std::min(Alignment, Subtarget.getFrameLowering()->getStackAlign());
6232   if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
6233     unsigned NewOpc = 0;
6234     unsigned RCSize = 0;
6235     switch (MI.getOpcode()) {
6236     default: return nullptr;
6237     case X86::TEST8rr:  NewOpc = X86::CMP8ri; RCSize = 1; break;
6238     case X86::TEST16rr: NewOpc = X86::CMP16ri8; RCSize = 2; break;
6239     case X86::TEST32rr: NewOpc = X86::CMP32ri8; RCSize = 4; break;
6240     case X86::TEST64rr: NewOpc = X86::CMP64ri8; RCSize = 8; break;
6241     }
6242     // Check if it's safe to fold the load. If the size of the object is
6243     // narrower than the load width, then it's not.
6244     if (Size < RCSize)
6245       return nullptr;
6246     // Change to CMPXXri r, 0 first.
6247     MI.setDesc(get(NewOpc));
6248     MI.getOperand(1).ChangeToImmediate(0);
6249   } else if (Ops.size() != 1)
6250     return nullptr;
6251 
6252   return foldMemoryOperandImpl(MF, MI, Ops[0],
6253                                MachineOperand::CreateFI(FrameIndex), InsertPt,
6254                                Size, Alignment, /*AllowCommute=*/true);
6255 }
6256 
6257 /// Check if \p LoadMI is a partial register load that we can't fold into \p MI
6258 /// because the latter uses contents that wouldn't be defined in the folded
6259 /// version.  For instance, this transformation isn't legal:
6260 ///   movss (%rdi), %xmm0
6261 ///   addps %xmm0, %xmm0
6262 /// ->
6263 ///   addps (%rdi), %xmm0
6264 ///
6265 /// But this one is:
6266 ///   movss (%rdi), %xmm0
6267 ///   addss %xmm0, %xmm0
6268 /// ->
6269 ///   addss (%rdi), %xmm0
6270 ///
6271 static bool isNonFoldablePartialRegisterLoad(const MachineInstr &LoadMI,
6272                                              const MachineInstr &UserMI,
6273                                              const MachineFunction &MF) {
6274   unsigned Opc = LoadMI.getOpcode();
6275   unsigned UserOpc = UserMI.getOpcode();
6276   const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
6277   const TargetRegisterClass *RC =
6278       MF.getRegInfo().getRegClass(LoadMI.getOperand(0).getReg());
6279   unsigned RegSize = TRI.getRegSizeInBits(*RC);
6280 
6281   if ((Opc == X86::MOVSSrm || Opc == X86::VMOVSSrm || Opc == X86::VMOVSSZrm ||
6282        Opc == X86::MOVSSrm_alt || Opc == X86::VMOVSSrm_alt ||
6283        Opc == X86::VMOVSSZrm_alt) &&
6284       RegSize > 32) {
6285     // These instructions only load 32 bits, we can't fold them if the
6286     // destination register is wider than 32 bits (4 bytes), and its user
6287     // instruction isn't scalar (SS).
6288     switch (UserOpc) {
6289     case X86::CVTSS2SDrr_Int:
6290     case X86::VCVTSS2SDrr_Int:
6291     case X86::VCVTSS2SDZrr_Int:
6292     case X86::VCVTSS2SDZrr_Intk:
6293     case X86::VCVTSS2SDZrr_Intkz:
6294     case X86::CVTSS2SIrr_Int:     case X86::CVTSS2SI64rr_Int:
6295     case X86::VCVTSS2SIrr_Int:    case X86::VCVTSS2SI64rr_Int:
6296     case X86::VCVTSS2SIZrr_Int:   case X86::VCVTSS2SI64Zrr_Int:
6297     case X86::CVTTSS2SIrr_Int:    case X86::CVTTSS2SI64rr_Int:
6298     case X86::VCVTTSS2SIrr_Int:   case X86::VCVTTSS2SI64rr_Int:
6299     case X86::VCVTTSS2SIZrr_Int:  case X86::VCVTTSS2SI64Zrr_Int:
6300     case X86::VCVTSS2USIZrr_Int:  case X86::VCVTSS2USI64Zrr_Int:
6301     case X86::VCVTTSS2USIZrr_Int: case X86::VCVTTSS2USI64Zrr_Int:
6302     case X86::RCPSSr_Int:   case X86::VRCPSSr_Int:
6303     case X86::RSQRTSSr_Int: case X86::VRSQRTSSr_Int:
6304     case X86::ROUNDSSr_Int: case X86::VROUNDSSr_Int:
6305     case X86::COMISSrr_Int: case X86::VCOMISSrr_Int: case X86::VCOMISSZrr_Int:
6306     case X86::UCOMISSrr_Int:case X86::VUCOMISSrr_Int:case X86::VUCOMISSZrr_Int:
6307     case X86::ADDSSrr_Int: case X86::VADDSSrr_Int: case X86::VADDSSZrr_Int:
6308     case X86::CMPSSrr_Int: case X86::VCMPSSrr_Int: case X86::VCMPSSZrr_Int:
6309     case X86::DIVSSrr_Int: case X86::VDIVSSrr_Int: case X86::VDIVSSZrr_Int:
6310     case X86::MAXSSrr_Int: case X86::VMAXSSrr_Int: case X86::VMAXSSZrr_Int:
6311     case X86::MINSSrr_Int: case X86::VMINSSrr_Int: case X86::VMINSSZrr_Int:
6312     case X86::MULSSrr_Int: case X86::VMULSSrr_Int: case X86::VMULSSZrr_Int:
6313     case X86::SQRTSSr_Int: case X86::VSQRTSSr_Int: case X86::VSQRTSSZr_Int:
6314     case X86::SUBSSrr_Int: case X86::VSUBSSrr_Int: case X86::VSUBSSZrr_Int:
6315     case X86::VADDSSZrr_Intk: case X86::VADDSSZrr_Intkz:
6316     case X86::VCMPSSZrr_Intk:
6317     case X86::VDIVSSZrr_Intk: case X86::VDIVSSZrr_Intkz:
6318     case X86::VMAXSSZrr_Intk: case X86::VMAXSSZrr_Intkz:
6319     case X86::VMINSSZrr_Intk: case X86::VMINSSZrr_Intkz:
6320     case X86::VMULSSZrr_Intk: case X86::VMULSSZrr_Intkz:
6321     case X86::VSQRTSSZr_Intk: case X86::VSQRTSSZr_Intkz:
6322     case X86::VSUBSSZrr_Intk: case X86::VSUBSSZrr_Intkz:
6323     case X86::VFMADDSS4rr_Int:   case X86::VFNMADDSS4rr_Int:
6324     case X86::VFMSUBSS4rr_Int:   case X86::VFNMSUBSS4rr_Int:
6325     case X86::VFMADD132SSr_Int:  case X86::VFNMADD132SSr_Int:
6326     case X86::VFMADD213SSr_Int:  case X86::VFNMADD213SSr_Int:
6327     case X86::VFMADD231SSr_Int:  case X86::VFNMADD231SSr_Int:
6328     case X86::VFMSUB132SSr_Int:  case X86::VFNMSUB132SSr_Int:
6329     case X86::VFMSUB213SSr_Int:  case X86::VFNMSUB213SSr_Int:
6330     case X86::VFMSUB231SSr_Int:  case X86::VFNMSUB231SSr_Int:
6331     case X86::VFMADD132SSZr_Int: case X86::VFNMADD132SSZr_Int:
6332     case X86::VFMADD213SSZr_Int: case X86::VFNMADD213SSZr_Int:
6333     case X86::VFMADD231SSZr_Int: case X86::VFNMADD231SSZr_Int:
6334     case X86::VFMSUB132SSZr_Int: case X86::VFNMSUB132SSZr_Int:
6335     case X86::VFMSUB213SSZr_Int: case X86::VFNMSUB213SSZr_Int:
6336     case X86::VFMSUB231SSZr_Int: case X86::VFNMSUB231SSZr_Int:
6337     case X86::VFMADD132SSZr_Intk: case X86::VFNMADD132SSZr_Intk:
6338     case X86::VFMADD213SSZr_Intk: case X86::VFNMADD213SSZr_Intk:
6339     case X86::VFMADD231SSZr_Intk: case X86::VFNMADD231SSZr_Intk:
6340     case X86::VFMSUB132SSZr_Intk: case X86::VFNMSUB132SSZr_Intk:
6341     case X86::VFMSUB213SSZr_Intk: case X86::VFNMSUB213SSZr_Intk:
6342     case X86::VFMSUB231SSZr_Intk: case X86::VFNMSUB231SSZr_Intk:
6343     case X86::VFMADD132SSZr_Intkz: case X86::VFNMADD132SSZr_Intkz:
6344     case X86::VFMADD213SSZr_Intkz: case X86::VFNMADD213SSZr_Intkz:
6345     case X86::VFMADD231SSZr_Intkz: case X86::VFNMADD231SSZr_Intkz:
6346     case X86::VFMSUB132SSZr_Intkz: case X86::VFNMSUB132SSZr_Intkz:
6347     case X86::VFMSUB213SSZr_Intkz: case X86::VFNMSUB213SSZr_Intkz:
6348     case X86::VFMSUB231SSZr_Intkz: case X86::VFNMSUB231SSZr_Intkz:
6349     case X86::VFIXUPIMMSSZrri:
6350     case X86::VFIXUPIMMSSZrrik:
6351     case X86::VFIXUPIMMSSZrrikz:
6352     case X86::VFPCLASSSSZrr:
6353     case X86::VFPCLASSSSZrrk:
6354     case X86::VGETEXPSSZr:
6355     case X86::VGETEXPSSZrk:
6356     case X86::VGETEXPSSZrkz:
6357     case X86::VGETMANTSSZrri:
6358     case X86::VGETMANTSSZrrik:
6359     case X86::VGETMANTSSZrrikz:
6360     case X86::VRANGESSZrri:
6361     case X86::VRANGESSZrrik:
6362     case X86::VRANGESSZrrikz:
6363     case X86::VRCP14SSZrr:
6364     case X86::VRCP14SSZrrk:
6365     case X86::VRCP14SSZrrkz:
6366     case X86::VRCP28SSZr:
6367     case X86::VRCP28SSZrk:
6368     case X86::VRCP28SSZrkz:
6369     case X86::VREDUCESSZrri:
6370     case X86::VREDUCESSZrrik:
6371     case X86::VREDUCESSZrrikz:
6372     case X86::VRNDSCALESSZr_Int:
6373     case X86::VRNDSCALESSZr_Intk:
6374     case X86::VRNDSCALESSZr_Intkz:
6375     case X86::VRSQRT14SSZrr:
6376     case X86::VRSQRT14SSZrrk:
6377     case X86::VRSQRT14SSZrrkz:
6378     case X86::VRSQRT28SSZr:
6379     case X86::VRSQRT28SSZrk:
6380     case X86::VRSQRT28SSZrkz:
6381     case X86::VSCALEFSSZrr:
6382     case X86::VSCALEFSSZrrk:
6383     case X86::VSCALEFSSZrrkz:
6384       return false;
6385     default:
6386       return true;
6387     }
6388   }
6389 
6390   if ((Opc == X86::MOVSDrm || Opc == X86::VMOVSDrm || Opc == X86::VMOVSDZrm ||
6391        Opc == X86::MOVSDrm_alt || Opc == X86::VMOVSDrm_alt ||
6392        Opc == X86::VMOVSDZrm_alt) &&
6393       RegSize > 64) {
6394     // These instructions only load 64 bits, we can't fold them if the
6395     // destination register is wider than 64 bits (8 bytes), and its user
6396     // instruction isn't scalar (SD).
6397     switch (UserOpc) {
6398     case X86::CVTSD2SSrr_Int:
6399     case X86::VCVTSD2SSrr_Int:
6400     case X86::VCVTSD2SSZrr_Int:
6401     case X86::VCVTSD2SSZrr_Intk:
6402     case X86::VCVTSD2SSZrr_Intkz:
6403     case X86::CVTSD2SIrr_Int:     case X86::CVTSD2SI64rr_Int:
6404     case X86::VCVTSD2SIrr_Int:    case X86::VCVTSD2SI64rr_Int:
6405     case X86::VCVTSD2SIZrr_Int:   case X86::VCVTSD2SI64Zrr_Int:
6406     case X86::CVTTSD2SIrr_Int:    case X86::CVTTSD2SI64rr_Int:
6407     case X86::VCVTTSD2SIrr_Int:   case X86::VCVTTSD2SI64rr_Int:
6408     case X86::VCVTTSD2SIZrr_Int:  case X86::VCVTTSD2SI64Zrr_Int:
6409     case X86::VCVTSD2USIZrr_Int:  case X86::VCVTSD2USI64Zrr_Int:
6410     case X86::VCVTTSD2USIZrr_Int: case X86::VCVTTSD2USI64Zrr_Int:
6411     case X86::ROUNDSDr_Int: case X86::VROUNDSDr_Int:
6412     case X86::COMISDrr_Int: case X86::VCOMISDrr_Int: case X86::VCOMISDZrr_Int:
6413     case X86::UCOMISDrr_Int:case X86::VUCOMISDrr_Int:case X86::VUCOMISDZrr_Int:
6414     case X86::ADDSDrr_Int: case X86::VADDSDrr_Int: case X86::VADDSDZrr_Int:
6415     case X86::CMPSDrr_Int: case X86::VCMPSDrr_Int: case X86::VCMPSDZrr_Int:
6416     case X86::DIVSDrr_Int: case X86::VDIVSDrr_Int: case X86::VDIVSDZrr_Int:
6417     case X86::MAXSDrr_Int: case X86::VMAXSDrr_Int: case X86::VMAXSDZrr_Int:
6418     case X86::MINSDrr_Int: case X86::VMINSDrr_Int: case X86::VMINSDZrr_Int:
6419     case X86::MULSDrr_Int: case X86::VMULSDrr_Int: case X86::VMULSDZrr_Int:
6420     case X86::SQRTSDr_Int: case X86::VSQRTSDr_Int: case X86::VSQRTSDZr_Int:
6421     case X86::SUBSDrr_Int: case X86::VSUBSDrr_Int: case X86::VSUBSDZrr_Int:
6422     case X86::VADDSDZrr_Intk: case X86::VADDSDZrr_Intkz:
6423     case X86::VCMPSDZrr_Intk:
6424     case X86::VDIVSDZrr_Intk: case X86::VDIVSDZrr_Intkz:
6425     case X86::VMAXSDZrr_Intk: case X86::VMAXSDZrr_Intkz:
6426     case X86::VMINSDZrr_Intk: case X86::VMINSDZrr_Intkz:
6427     case X86::VMULSDZrr_Intk: case X86::VMULSDZrr_Intkz:
6428     case X86::VSQRTSDZr_Intk: case X86::VSQRTSDZr_Intkz:
6429     case X86::VSUBSDZrr_Intk: case X86::VSUBSDZrr_Intkz:
6430     case X86::VFMADDSD4rr_Int:   case X86::VFNMADDSD4rr_Int:
6431     case X86::VFMSUBSD4rr_Int:   case X86::VFNMSUBSD4rr_Int:
6432     case X86::VFMADD132SDr_Int:  case X86::VFNMADD132SDr_Int:
6433     case X86::VFMADD213SDr_Int:  case X86::VFNMADD213SDr_Int:
6434     case X86::VFMADD231SDr_Int:  case X86::VFNMADD231SDr_Int:
6435     case X86::VFMSUB132SDr_Int:  case X86::VFNMSUB132SDr_Int:
6436     case X86::VFMSUB213SDr_Int:  case X86::VFNMSUB213SDr_Int:
6437     case X86::VFMSUB231SDr_Int:  case X86::VFNMSUB231SDr_Int:
6438     case X86::VFMADD132SDZr_Int: case X86::VFNMADD132SDZr_Int:
6439     case X86::VFMADD213SDZr_Int: case X86::VFNMADD213SDZr_Int:
6440     case X86::VFMADD231SDZr_Int: case X86::VFNMADD231SDZr_Int:
6441     case X86::VFMSUB132SDZr_Int: case X86::VFNMSUB132SDZr_Int:
6442     case X86::VFMSUB213SDZr_Int: case X86::VFNMSUB213SDZr_Int:
6443     case X86::VFMSUB231SDZr_Int: case X86::VFNMSUB231SDZr_Int:
6444     case X86::VFMADD132SDZr_Intk: case X86::VFNMADD132SDZr_Intk:
6445     case X86::VFMADD213SDZr_Intk: case X86::VFNMADD213SDZr_Intk:
6446     case X86::VFMADD231SDZr_Intk: case X86::VFNMADD231SDZr_Intk:
6447     case X86::VFMSUB132SDZr_Intk: case X86::VFNMSUB132SDZr_Intk:
6448     case X86::VFMSUB213SDZr_Intk: case X86::VFNMSUB213SDZr_Intk:
6449     case X86::VFMSUB231SDZr_Intk: case X86::VFNMSUB231SDZr_Intk:
6450     case X86::VFMADD132SDZr_Intkz: case X86::VFNMADD132SDZr_Intkz:
6451     case X86::VFMADD213SDZr_Intkz: case X86::VFNMADD213SDZr_Intkz:
6452     case X86::VFMADD231SDZr_Intkz: case X86::VFNMADD231SDZr_Intkz:
6453     case X86::VFMSUB132SDZr_Intkz: case X86::VFNMSUB132SDZr_Intkz:
6454     case X86::VFMSUB213SDZr_Intkz: case X86::VFNMSUB213SDZr_Intkz:
6455     case X86::VFMSUB231SDZr_Intkz: case X86::VFNMSUB231SDZr_Intkz:
6456     case X86::VFIXUPIMMSDZrri:
6457     case X86::VFIXUPIMMSDZrrik:
6458     case X86::VFIXUPIMMSDZrrikz:
6459     case X86::VFPCLASSSDZrr:
6460     case X86::VFPCLASSSDZrrk:
6461     case X86::VGETEXPSDZr:
6462     case X86::VGETEXPSDZrk:
6463     case X86::VGETEXPSDZrkz:
6464     case X86::VGETMANTSDZrri:
6465     case X86::VGETMANTSDZrrik:
6466     case X86::VGETMANTSDZrrikz:
6467     case X86::VRANGESDZrri:
6468     case X86::VRANGESDZrrik:
6469     case X86::VRANGESDZrrikz:
6470     case X86::VRCP14SDZrr:
6471     case X86::VRCP14SDZrrk:
6472     case X86::VRCP14SDZrrkz:
6473     case X86::VRCP28SDZr:
6474     case X86::VRCP28SDZrk:
6475     case X86::VRCP28SDZrkz:
6476     case X86::VREDUCESDZrri:
6477     case X86::VREDUCESDZrrik:
6478     case X86::VREDUCESDZrrikz:
6479     case X86::VRNDSCALESDZr_Int:
6480     case X86::VRNDSCALESDZr_Intk:
6481     case X86::VRNDSCALESDZr_Intkz:
6482     case X86::VRSQRT14SDZrr:
6483     case X86::VRSQRT14SDZrrk:
6484     case X86::VRSQRT14SDZrrkz:
6485     case X86::VRSQRT28SDZr:
6486     case X86::VRSQRT28SDZrk:
6487     case X86::VRSQRT28SDZrkz:
6488     case X86::VSCALEFSDZrr:
6489     case X86::VSCALEFSDZrrk:
6490     case X86::VSCALEFSDZrrkz:
6491       return false;
6492     default:
6493       return true;
6494     }
6495   }
6496 
6497   if ((Opc == X86::VMOVSHZrm || Opc == X86::VMOVSHZrm_alt) && RegSize > 16) {
6498     // These instructions only load 16 bits, we can't fold them if the
6499     // destination register is wider than 16 bits (2 bytes), and its user
6500     // instruction isn't scalar (SH).
6501     switch (UserOpc) {
6502     case X86::VADDSHZrr_Int:
6503     case X86::VCMPSHZrr_Int:
6504     case X86::VDIVSHZrr_Int:
6505     case X86::VMAXSHZrr_Int:
6506     case X86::VMINSHZrr_Int:
6507     case X86::VMULSHZrr_Int:
6508     case X86::VSUBSHZrr_Int:
6509     case X86::VADDSHZrr_Intk: case X86::VADDSHZrr_Intkz:
6510     case X86::VCMPSHZrr_Intk:
6511     case X86::VDIVSHZrr_Intk: case X86::VDIVSHZrr_Intkz:
6512     case X86::VMAXSHZrr_Intk: case X86::VMAXSHZrr_Intkz:
6513     case X86::VMINSHZrr_Intk: case X86::VMINSHZrr_Intkz:
6514     case X86::VMULSHZrr_Intk: case X86::VMULSHZrr_Intkz:
6515     case X86::VSUBSHZrr_Intk: case X86::VSUBSHZrr_Intkz:
6516     case X86::VFMADD132SHZr_Int: case X86::VFNMADD132SHZr_Int:
6517     case X86::VFMADD213SHZr_Int: case X86::VFNMADD213SHZr_Int:
6518     case X86::VFMADD231SHZr_Int: case X86::VFNMADD231SHZr_Int:
6519     case X86::VFMSUB132SHZr_Int: case X86::VFNMSUB132SHZr_Int:
6520     case X86::VFMSUB213SHZr_Int: case X86::VFNMSUB213SHZr_Int:
6521     case X86::VFMSUB231SHZr_Int: case X86::VFNMSUB231SHZr_Int:
6522     case X86::VFMADD132SHZr_Intk: case X86::VFNMADD132SHZr_Intk:
6523     case X86::VFMADD213SHZr_Intk: case X86::VFNMADD213SHZr_Intk:
6524     case X86::VFMADD231SHZr_Intk: case X86::VFNMADD231SHZr_Intk:
6525     case X86::VFMSUB132SHZr_Intk: case X86::VFNMSUB132SHZr_Intk:
6526     case X86::VFMSUB213SHZr_Intk: case X86::VFNMSUB213SHZr_Intk:
6527     case X86::VFMSUB231SHZr_Intk: case X86::VFNMSUB231SHZr_Intk:
6528     case X86::VFMADD132SHZr_Intkz: case X86::VFNMADD132SHZr_Intkz:
6529     case X86::VFMADD213SHZr_Intkz: case X86::VFNMADD213SHZr_Intkz:
6530     case X86::VFMADD231SHZr_Intkz: case X86::VFNMADD231SHZr_Intkz:
6531     case X86::VFMSUB132SHZr_Intkz: case X86::VFNMSUB132SHZr_Intkz:
6532     case X86::VFMSUB213SHZr_Intkz: case X86::VFNMSUB213SHZr_Intkz:
6533     case X86::VFMSUB231SHZr_Intkz: case X86::VFNMSUB231SHZr_Intkz:
6534       return false;
6535     default:
6536       return true;
6537     }
6538   }
6539 
6540   return false;
6541 }
6542 
6543 MachineInstr *X86InstrInfo::foldMemoryOperandImpl(
6544     MachineFunction &MF, MachineInstr &MI, ArrayRef<unsigned> Ops,
6545     MachineBasicBlock::iterator InsertPt, MachineInstr &LoadMI,
6546     LiveIntervals *LIS) const {
6547 
6548   // TODO: Support the case where LoadMI loads a wide register, but MI
6549   // only uses a subreg.
6550   for (auto Op : Ops) {
6551     if (MI.getOperand(Op).getSubReg())
6552       return nullptr;
6553   }
6554 
6555   // If loading from a FrameIndex, fold directly from the FrameIndex.
6556   unsigned NumOps = LoadMI.getDesc().getNumOperands();
6557   int FrameIndex;
6558   if (isLoadFromStackSlot(LoadMI, FrameIndex)) {
6559     if (isNonFoldablePartialRegisterLoad(LoadMI, MI, MF))
6560       return nullptr;
6561     return foldMemoryOperandImpl(MF, MI, Ops, InsertPt, FrameIndex, LIS);
6562   }
6563 
6564   // Check switch flag
6565   if (NoFusing) return nullptr;
6566 
6567   // Avoid partial and undef register update stalls unless optimizing for size.
6568   if (!MF.getFunction().hasOptSize() &&
6569       (hasPartialRegUpdate(MI.getOpcode(), Subtarget, /*ForLoadFold*/true) ||
6570        shouldPreventUndefRegUpdateMemFold(MF, MI)))
6571     return nullptr;
6572 
6573   // Determine the alignment of the load.
6574   Align Alignment;
6575   if (LoadMI.hasOneMemOperand())
6576     Alignment = (*LoadMI.memoperands_begin())->getAlign();
6577   else
6578     switch (LoadMI.getOpcode()) {
6579     case X86::AVX512_512_SET0:
6580     case X86::AVX512_512_SETALLONES:
6581       Alignment = Align(64);
6582       break;
6583     case X86::AVX2_SETALLONES:
6584     case X86::AVX1_SETALLONES:
6585     case X86::AVX_SET0:
6586     case X86::AVX512_256_SET0:
6587       Alignment = Align(32);
6588       break;
6589     case X86::V_SET0:
6590     case X86::V_SETALLONES:
6591     case X86::AVX512_128_SET0:
6592     case X86::FsFLD0F128:
6593     case X86::AVX512_FsFLD0F128:
6594       Alignment = Align(16);
6595       break;
6596     case X86::MMX_SET0:
6597     case X86::FsFLD0SD:
6598     case X86::AVX512_FsFLD0SD:
6599       Alignment = Align(8);
6600       break;
6601     case X86::FsFLD0SS:
6602     case X86::AVX512_FsFLD0SS:
6603       Alignment = Align(4);
6604       break;
6605     case X86::AVX512_FsFLD0SH:
6606       Alignment = Align(2);
6607       break;
6608     default:
6609       return nullptr;
6610     }
6611   if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
6612     unsigned NewOpc = 0;
6613     switch (MI.getOpcode()) {
6614     default: return nullptr;
6615     case X86::TEST8rr:  NewOpc = X86::CMP8ri; break;
6616     case X86::TEST16rr: NewOpc = X86::CMP16ri8; break;
6617     case X86::TEST32rr: NewOpc = X86::CMP32ri8; break;
6618     case X86::TEST64rr: NewOpc = X86::CMP64ri8; break;
6619     }
6620     // Change to CMPXXri r, 0 first.
6621     MI.setDesc(get(NewOpc));
6622     MI.getOperand(1).ChangeToImmediate(0);
6623   } else if (Ops.size() != 1)
6624     return nullptr;
6625 
6626   // Make sure the subregisters match.
6627   // Otherwise we risk changing the size of the load.
6628   if (LoadMI.getOperand(0).getSubReg() != MI.getOperand(Ops[0]).getSubReg())
6629     return nullptr;
6630 
6631   SmallVector<MachineOperand,X86::AddrNumOperands> MOs;
6632   switch (LoadMI.getOpcode()) {
6633   case X86::MMX_SET0:
6634   case X86::V_SET0:
6635   case X86::V_SETALLONES:
6636   case X86::AVX2_SETALLONES:
6637   case X86::AVX1_SETALLONES:
6638   case X86::AVX_SET0:
6639   case X86::AVX512_128_SET0:
6640   case X86::AVX512_256_SET0:
6641   case X86::AVX512_512_SET0:
6642   case X86::AVX512_512_SETALLONES:
6643   case X86::AVX512_FsFLD0SH:
6644   case X86::FsFLD0SD:
6645   case X86::AVX512_FsFLD0SD:
6646   case X86::FsFLD0SS:
6647   case X86::AVX512_FsFLD0SS:
6648   case X86::FsFLD0F128:
6649   case X86::AVX512_FsFLD0F128: {
6650     // Folding a V_SET0 or V_SETALLONES as a load, to ease register pressure.
6651     // Create a constant-pool entry and operands to load from it.
6652 
6653     // Medium and large mode can't fold loads this way.
6654     if (MF.getTarget().getCodeModel() != CodeModel::Small &&
6655         MF.getTarget().getCodeModel() != CodeModel::Kernel)
6656       return nullptr;
6657 
6658     // x86-32 PIC requires a PIC base register for constant pools.
6659     unsigned PICBase = 0;
6660     // Since we're using Small or Kernel code model, we can always use
6661     // RIP-relative addressing for a smaller encoding.
6662     if (Subtarget.is64Bit()) {
6663       PICBase = X86::RIP;
6664     } else if (MF.getTarget().isPositionIndependent()) {
6665       // FIXME: PICBase = getGlobalBaseReg(&MF);
6666       // This doesn't work for several reasons.
6667       // 1. GlobalBaseReg may have been spilled.
6668       // 2. It may not be live at MI.
6669       return nullptr;
6670     }
6671 
6672     // Create a constant-pool entry.
6673     MachineConstantPool &MCP = *MF.getConstantPool();
6674     Type *Ty;
6675     unsigned Opc = LoadMI.getOpcode();
6676     if (Opc == X86::FsFLD0SS || Opc == X86::AVX512_FsFLD0SS)
6677       Ty = Type::getFloatTy(MF.getFunction().getContext());
6678     else if (Opc == X86::FsFLD0SD || Opc == X86::AVX512_FsFLD0SD)
6679       Ty = Type::getDoubleTy(MF.getFunction().getContext());
6680     else if (Opc == X86::FsFLD0F128 || Opc == X86::AVX512_FsFLD0F128)
6681       Ty = Type::getFP128Ty(MF.getFunction().getContext());
6682     else if (Opc == X86::AVX512_FsFLD0SH)
6683       Ty = Type::getHalfTy(MF.getFunction().getContext());
6684     else if (Opc == X86::AVX512_512_SET0 || Opc == X86::AVX512_512_SETALLONES)
6685       Ty = FixedVectorType::get(Type::getInt32Ty(MF.getFunction().getContext()),
6686                                 16);
6687     else if (Opc == X86::AVX2_SETALLONES || Opc == X86::AVX_SET0 ||
6688              Opc == X86::AVX512_256_SET0 || Opc == X86::AVX1_SETALLONES)
6689       Ty = FixedVectorType::get(Type::getInt32Ty(MF.getFunction().getContext()),
6690                                 8);
6691     else if (Opc == X86::MMX_SET0)
6692       Ty = FixedVectorType::get(Type::getInt32Ty(MF.getFunction().getContext()),
6693                                 2);
6694     else
6695       Ty = FixedVectorType::get(Type::getInt32Ty(MF.getFunction().getContext()),
6696                                 4);
6697 
6698     bool IsAllOnes = (Opc == X86::V_SETALLONES || Opc == X86::AVX2_SETALLONES ||
6699                       Opc == X86::AVX512_512_SETALLONES ||
6700                       Opc == X86::AVX1_SETALLONES);
6701     const Constant *C = IsAllOnes ? Constant::getAllOnesValue(Ty) :
6702                                     Constant::getNullValue(Ty);
6703     unsigned CPI = MCP.getConstantPoolIndex(C, Alignment);
6704 
6705     // Create operands to load from the constant pool entry.
6706     MOs.push_back(MachineOperand::CreateReg(PICBase, false));
6707     MOs.push_back(MachineOperand::CreateImm(1));
6708     MOs.push_back(MachineOperand::CreateReg(0, false));
6709     MOs.push_back(MachineOperand::CreateCPI(CPI, 0));
6710     MOs.push_back(MachineOperand::CreateReg(0, false));
6711     break;
6712   }
6713   default: {
6714     if (isNonFoldablePartialRegisterLoad(LoadMI, MI, MF))
6715       return nullptr;
6716 
6717     // Folding a normal load. Just copy the load's address operands.
6718     MOs.append(LoadMI.operands_begin() + NumOps - X86::AddrNumOperands,
6719                LoadMI.operands_begin() + NumOps);
6720     break;
6721   }
6722   }
6723   return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, InsertPt,
6724                                /*Size=*/0, Alignment, /*AllowCommute=*/true);
6725 }
6726 
6727 static SmallVector<MachineMemOperand *, 2>
6728 extractLoadMMOs(ArrayRef<MachineMemOperand *> MMOs, MachineFunction &MF) {
6729   SmallVector<MachineMemOperand *, 2> LoadMMOs;
6730 
6731   for (MachineMemOperand *MMO : MMOs) {
6732     if (!MMO->isLoad())
6733       continue;
6734 
6735     if (!MMO->isStore()) {
6736       // Reuse the MMO.
6737       LoadMMOs.push_back(MMO);
6738     } else {
6739       // Clone the MMO and unset the store flag.
6740       LoadMMOs.push_back(MF.getMachineMemOperand(
6741           MMO, MMO->getFlags() & ~MachineMemOperand::MOStore));
6742     }
6743   }
6744 
6745   return LoadMMOs;
6746 }
6747 
6748 static SmallVector<MachineMemOperand *, 2>
6749 extractStoreMMOs(ArrayRef<MachineMemOperand *> MMOs, MachineFunction &MF) {
6750   SmallVector<MachineMemOperand *, 2> StoreMMOs;
6751 
6752   for (MachineMemOperand *MMO : MMOs) {
6753     if (!MMO->isStore())
6754       continue;
6755 
6756     if (!MMO->isLoad()) {
6757       // Reuse the MMO.
6758       StoreMMOs.push_back(MMO);
6759     } else {
6760       // Clone the MMO and unset the load flag.
6761       StoreMMOs.push_back(MF.getMachineMemOperand(
6762           MMO, MMO->getFlags() & ~MachineMemOperand::MOLoad));
6763     }
6764   }
6765 
6766   return StoreMMOs;
6767 }
6768 
6769 static unsigned getBroadcastOpcode(const X86MemoryFoldTableEntry *I,
6770                                    const TargetRegisterClass *RC,
6771                                    const X86Subtarget &STI) {
6772   assert(STI.hasAVX512() && "Expected at least AVX512!");
6773   unsigned SpillSize = STI.getRegisterInfo()->getSpillSize(*RC);
6774   assert((SpillSize == 64 || STI.hasVLX()) &&
6775          "Can't broadcast less than 64 bytes without AVX512VL!");
6776 
6777   switch (I->Flags & TB_BCAST_MASK) {
6778   default: llvm_unreachable("Unexpected broadcast type!");
6779   case TB_BCAST_D:
6780     switch (SpillSize) {
6781     default: llvm_unreachable("Unknown spill size");
6782     case 16: return X86::VPBROADCASTDZ128rm;
6783     case 32: return X86::VPBROADCASTDZ256rm;
6784     case 64: return X86::VPBROADCASTDZrm;
6785     }
6786     break;
6787   case TB_BCAST_Q:
6788     switch (SpillSize) {
6789     default: llvm_unreachable("Unknown spill size");
6790     case 16: return X86::VPBROADCASTQZ128rm;
6791     case 32: return X86::VPBROADCASTQZ256rm;
6792     case 64: return X86::VPBROADCASTQZrm;
6793     }
6794     break;
6795   case TB_BCAST_SS:
6796     switch (SpillSize) {
6797     default: llvm_unreachable("Unknown spill size");
6798     case 16: return X86::VBROADCASTSSZ128rm;
6799     case 32: return X86::VBROADCASTSSZ256rm;
6800     case 64: return X86::VBROADCASTSSZrm;
6801     }
6802     break;
6803   case TB_BCAST_SD:
6804     switch (SpillSize) {
6805     default: llvm_unreachable("Unknown spill size");
6806     case 16: return X86::VMOVDDUPZ128rm;
6807     case 32: return X86::VBROADCASTSDZ256rm;
6808     case 64: return X86::VBROADCASTSDZrm;
6809     }
6810     break;
6811   }
6812 }
6813 
6814 bool X86InstrInfo::unfoldMemoryOperand(
6815     MachineFunction &MF, MachineInstr &MI, unsigned Reg, bool UnfoldLoad,
6816     bool UnfoldStore, SmallVectorImpl<MachineInstr *> &NewMIs) const {
6817   const X86MemoryFoldTableEntry *I = lookupUnfoldTable(MI.getOpcode());
6818   if (I == nullptr)
6819     return false;
6820   unsigned Opc = I->DstOp;
6821   unsigned Index = I->Flags & TB_INDEX_MASK;
6822   bool FoldedLoad = I->Flags & TB_FOLDED_LOAD;
6823   bool FoldedStore = I->Flags & TB_FOLDED_STORE;
6824   bool FoldedBCast = I->Flags & TB_FOLDED_BCAST;
6825   if (UnfoldLoad && !FoldedLoad)
6826     return false;
6827   UnfoldLoad &= FoldedLoad;
6828   if (UnfoldStore && !FoldedStore)
6829     return false;
6830   UnfoldStore &= FoldedStore;
6831 
6832   const MCInstrDesc &MCID = get(Opc);
6833 
6834   const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
6835   const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
6836   // TODO: Check if 32-byte or greater accesses are slow too?
6837   if (!MI.hasOneMemOperand() && RC == &X86::VR128RegClass &&
6838       Subtarget.isUnalignedMem16Slow())
6839     // Without memoperands, loadRegFromAddr and storeRegToStackSlot will
6840     // conservatively assume the address is unaligned. That's bad for
6841     // performance.
6842     return false;
6843   SmallVector<MachineOperand, X86::AddrNumOperands> AddrOps;
6844   SmallVector<MachineOperand,2> BeforeOps;
6845   SmallVector<MachineOperand,2> AfterOps;
6846   SmallVector<MachineOperand,4> ImpOps;
6847   for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
6848     MachineOperand &Op = MI.getOperand(i);
6849     if (i >= Index && i < Index + X86::AddrNumOperands)
6850       AddrOps.push_back(Op);
6851     else if (Op.isReg() && Op.isImplicit())
6852       ImpOps.push_back(Op);
6853     else if (i < Index)
6854       BeforeOps.push_back(Op);
6855     else if (i > Index)
6856       AfterOps.push_back(Op);
6857   }
6858 
6859   // Emit the load or broadcast instruction.
6860   if (UnfoldLoad) {
6861     auto MMOs = extractLoadMMOs(MI.memoperands(), MF);
6862 
6863     unsigned Opc;
6864     if (FoldedBCast) {
6865       Opc = getBroadcastOpcode(I, RC, Subtarget);
6866     } else {
6867       unsigned Alignment = std::max<uint32_t>(TRI.getSpillSize(*RC), 16);
6868       bool isAligned = !MMOs.empty() && MMOs.front()->getAlign() >= Alignment;
6869       Opc = getLoadRegOpcode(Reg, RC, isAligned, Subtarget);
6870     }
6871 
6872     DebugLoc DL;
6873     MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc), Reg);
6874     for (unsigned i = 0, e = AddrOps.size(); i != e; ++i)
6875       MIB.add(AddrOps[i]);
6876     MIB.setMemRefs(MMOs);
6877     NewMIs.push_back(MIB);
6878 
6879     if (UnfoldStore) {
6880       // Address operands cannot be marked isKill.
6881       for (unsigned i = 1; i != 1 + X86::AddrNumOperands; ++i) {
6882         MachineOperand &MO = NewMIs[0]->getOperand(i);
6883         if (MO.isReg())
6884           MO.setIsKill(false);
6885       }
6886     }
6887   }
6888 
6889   // Emit the data processing instruction.
6890   MachineInstr *DataMI = MF.CreateMachineInstr(MCID, MI.getDebugLoc(), true);
6891   MachineInstrBuilder MIB(MF, DataMI);
6892 
6893   if (FoldedStore)
6894     MIB.addReg(Reg, RegState::Define);
6895   for (MachineOperand &BeforeOp : BeforeOps)
6896     MIB.add(BeforeOp);
6897   if (FoldedLoad)
6898     MIB.addReg(Reg);
6899   for (MachineOperand &AfterOp : AfterOps)
6900     MIB.add(AfterOp);
6901   for (MachineOperand &ImpOp : ImpOps) {
6902     MIB.addReg(ImpOp.getReg(),
6903                getDefRegState(ImpOp.isDef()) |
6904                RegState::Implicit |
6905                getKillRegState(ImpOp.isKill()) |
6906                getDeadRegState(ImpOp.isDead()) |
6907                getUndefRegState(ImpOp.isUndef()));
6908   }
6909   // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
6910   switch (DataMI->getOpcode()) {
6911   default: break;
6912   case X86::CMP64ri32:
6913   case X86::CMP64ri8:
6914   case X86::CMP32ri:
6915   case X86::CMP32ri8:
6916   case X86::CMP16ri:
6917   case X86::CMP16ri8:
6918   case X86::CMP8ri: {
6919     MachineOperand &MO0 = DataMI->getOperand(0);
6920     MachineOperand &MO1 = DataMI->getOperand(1);
6921     if (MO1.isImm() && MO1.getImm() == 0) {
6922       unsigned NewOpc;
6923       switch (DataMI->getOpcode()) {
6924       default: llvm_unreachable("Unreachable!");
6925       case X86::CMP64ri8:
6926       case X86::CMP64ri32: NewOpc = X86::TEST64rr; break;
6927       case X86::CMP32ri8:
6928       case X86::CMP32ri:   NewOpc = X86::TEST32rr; break;
6929       case X86::CMP16ri8:
6930       case X86::CMP16ri:   NewOpc = X86::TEST16rr; break;
6931       case X86::CMP8ri:    NewOpc = X86::TEST8rr; break;
6932       }
6933       DataMI->setDesc(get(NewOpc));
6934       MO1.ChangeToRegister(MO0.getReg(), false);
6935     }
6936   }
6937   }
6938   NewMIs.push_back(DataMI);
6939 
6940   // Emit the store instruction.
6941   if (UnfoldStore) {
6942     const TargetRegisterClass *DstRC = getRegClass(MCID, 0, &RI, MF);
6943     auto MMOs = extractStoreMMOs(MI.memoperands(), MF);
6944     unsigned Alignment = std::max<uint32_t>(TRI.getSpillSize(*DstRC), 16);
6945     bool isAligned = !MMOs.empty() && MMOs.front()->getAlign() >= Alignment;
6946     unsigned Opc = getStoreRegOpcode(Reg, DstRC, isAligned, Subtarget);
6947     DebugLoc DL;
6948     MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc));
6949     for (unsigned i = 0, e = AddrOps.size(); i != e; ++i)
6950       MIB.add(AddrOps[i]);
6951     MIB.addReg(Reg, RegState::Kill);
6952     MIB.setMemRefs(MMOs);
6953     NewMIs.push_back(MIB);
6954   }
6955 
6956   return true;
6957 }
6958 
6959 bool
6960 X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
6961                                   SmallVectorImpl<SDNode*> &NewNodes) const {
6962   if (!N->isMachineOpcode())
6963     return false;
6964 
6965   const X86MemoryFoldTableEntry *I = lookupUnfoldTable(N->getMachineOpcode());
6966   if (I == nullptr)
6967     return false;
6968   unsigned Opc = I->DstOp;
6969   unsigned Index = I->Flags & TB_INDEX_MASK;
6970   bool FoldedLoad = I->Flags & TB_FOLDED_LOAD;
6971   bool FoldedStore = I->Flags & TB_FOLDED_STORE;
6972   bool FoldedBCast = I->Flags & TB_FOLDED_BCAST;
6973   const MCInstrDesc &MCID = get(Opc);
6974   MachineFunction &MF = DAG.getMachineFunction();
6975   const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
6976   const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
6977   unsigned NumDefs = MCID.NumDefs;
6978   std::vector<SDValue> AddrOps;
6979   std::vector<SDValue> BeforeOps;
6980   std::vector<SDValue> AfterOps;
6981   SDLoc dl(N);
6982   unsigned NumOps = N->getNumOperands();
6983   for (unsigned i = 0; i != NumOps-1; ++i) {
6984     SDValue Op = N->getOperand(i);
6985     if (i >= Index-NumDefs && i < Index-NumDefs + X86::AddrNumOperands)
6986       AddrOps.push_back(Op);
6987     else if (i < Index-NumDefs)
6988       BeforeOps.push_back(Op);
6989     else if (i > Index-NumDefs)
6990       AfterOps.push_back(Op);
6991   }
6992   SDValue Chain = N->getOperand(NumOps-1);
6993   AddrOps.push_back(Chain);
6994 
6995   // Emit the load instruction.
6996   SDNode *Load = nullptr;
6997   if (FoldedLoad) {
6998     EVT VT = *TRI.legalclasstypes_begin(*RC);
6999     auto MMOs = extractLoadMMOs(cast<MachineSDNode>(N)->memoperands(), MF);
7000     if (MMOs.empty() && RC == &X86::VR128RegClass &&
7001         Subtarget.isUnalignedMem16Slow())
7002       // Do not introduce a slow unaligned load.
7003       return false;
7004     // FIXME: If a VR128 can have size 32, we should be checking if a 32-byte
7005     // memory access is slow above.
7006 
7007     unsigned Opc;
7008     if (FoldedBCast) {
7009       Opc = getBroadcastOpcode(I, RC, Subtarget);
7010     } else {
7011       unsigned Alignment = std::max<uint32_t>(TRI.getSpillSize(*RC), 16);
7012       bool isAligned = !MMOs.empty() && MMOs.front()->getAlign() >= Alignment;
7013       Opc = getLoadRegOpcode(0, RC, isAligned, Subtarget);
7014     }
7015 
7016     Load = DAG.getMachineNode(Opc, dl, VT, MVT::Other, AddrOps);
7017     NewNodes.push_back(Load);
7018 
7019     // Preserve memory reference information.
7020     DAG.setNodeMemRefs(cast<MachineSDNode>(Load), MMOs);
7021   }
7022 
7023   // Emit the data processing instruction.
7024   std::vector<EVT> VTs;
7025   const TargetRegisterClass *DstRC = nullptr;
7026   if (MCID.getNumDefs() > 0) {
7027     DstRC = getRegClass(MCID, 0, &RI, MF);
7028     VTs.push_back(*TRI.legalclasstypes_begin(*DstRC));
7029   }
7030   for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
7031     EVT VT = N->getValueType(i);
7032     if (VT != MVT::Other && i >= (unsigned)MCID.getNumDefs())
7033       VTs.push_back(VT);
7034   }
7035   if (Load)
7036     BeforeOps.push_back(SDValue(Load, 0));
7037   llvm::append_range(BeforeOps, AfterOps);
7038   // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
7039   switch (Opc) {
7040     default: break;
7041     case X86::CMP64ri32:
7042     case X86::CMP64ri8:
7043     case X86::CMP32ri:
7044     case X86::CMP32ri8:
7045     case X86::CMP16ri:
7046     case X86::CMP16ri8:
7047     case X86::CMP8ri:
7048       if (isNullConstant(BeforeOps[1])) {
7049         switch (Opc) {
7050           default: llvm_unreachable("Unreachable!");
7051           case X86::CMP64ri8:
7052           case X86::CMP64ri32: Opc = X86::TEST64rr; break;
7053           case X86::CMP32ri8:
7054           case X86::CMP32ri:   Opc = X86::TEST32rr; break;
7055           case X86::CMP16ri8:
7056           case X86::CMP16ri:   Opc = X86::TEST16rr; break;
7057           case X86::CMP8ri:    Opc = X86::TEST8rr; break;
7058         }
7059         BeforeOps[1] = BeforeOps[0];
7060       }
7061   }
7062   SDNode *NewNode= DAG.getMachineNode(Opc, dl, VTs, BeforeOps);
7063   NewNodes.push_back(NewNode);
7064 
7065   // Emit the store instruction.
7066   if (FoldedStore) {
7067     AddrOps.pop_back();
7068     AddrOps.push_back(SDValue(NewNode, 0));
7069     AddrOps.push_back(Chain);
7070     auto MMOs = extractStoreMMOs(cast<MachineSDNode>(N)->memoperands(), MF);
7071     if (MMOs.empty() && RC == &X86::VR128RegClass &&
7072         Subtarget.isUnalignedMem16Slow())
7073       // Do not introduce a slow unaligned store.
7074       return false;
7075     // FIXME: If a VR128 can have size 32, we should be checking if a 32-byte
7076     // memory access is slow above.
7077     unsigned Alignment = std::max<uint32_t>(TRI.getSpillSize(*RC), 16);
7078     bool isAligned = !MMOs.empty() && MMOs.front()->getAlign() >= Alignment;
7079     SDNode *Store =
7080         DAG.getMachineNode(getStoreRegOpcode(0, DstRC, isAligned, Subtarget),
7081                            dl, MVT::Other, AddrOps);
7082     NewNodes.push_back(Store);
7083 
7084     // Preserve memory reference information.
7085     DAG.setNodeMemRefs(cast<MachineSDNode>(Store), MMOs);
7086   }
7087 
7088   return true;
7089 }
7090 
7091 unsigned X86InstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc,
7092                                       bool UnfoldLoad, bool UnfoldStore,
7093                                       unsigned *LoadRegIndex) const {
7094   const X86MemoryFoldTableEntry *I = lookupUnfoldTable(Opc);
7095   if (I == nullptr)
7096     return 0;
7097   bool FoldedLoad = I->Flags & TB_FOLDED_LOAD;
7098   bool FoldedStore = I->Flags & TB_FOLDED_STORE;
7099   if (UnfoldLoad && !FoldedLoad)
7100     return 0;
7101   if (UnfoldStore && !FoldedStore)
7102     return 0;
7103   if (LoadRegIndex)
7104     *LoadRegIndex = I->Flags & TB_INDEX_MASK;
7105   return I->DstOp;
7106 }
7107 
7108 bool
7109 X86InstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
7110                                      int64_t &Offset1, int64_t &Offset2) const {
7111   if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode())
7112     return false;
7113   unsigned Opc1 = Load1->getMachineOpcode();
7114   unsigned Opc2 = Load2->getMachineOpcode();
7115   switch (Opc1) {
7116   default: return false;
7117   case X86::MOV8rm:
7118   case X86::MOV16rm:
7119   case X86::MOV32rm:
7120   case X86::MOV64rm:
7121   case X86::LD_Fp32m:
7122   case X86::LD_Fp64m:
7123   case X86::LD_Fp80m:
7124   case X86::MOVSSrm:
7125   case X86::MOVSSrm_alt:
7126   case X86::MOVSDrm:
7127   case X86::MOVSDrm_alt:
7128   case X86::MMX_MOVD64rm:
7129   case X86::MMX_MOVQ64rm:
7130   case X86::MOVAPSrm:
7131   case X86::MOVUPSrm:
7132   case X86::MOVAPDrm:
7133   case X86::MOVUPDrm:
7134   case X86::MOVDQArm:
7135   case X86::MOVDQUrm:
7136   // AVX load instructions
7137   case X86::VMOVSSrm:
7138   case X86::VMOVSSrm_alt:
7139   case X86::VMOVSDrm:
7140   case X86::VMOVSDrm_alt:
7141   case X86::VMOVAPSrm:
7142   case X86::VMOVUPSrm:
7143   case X86::VMOVAPDrm:
7144   case X86::VMOVUPDrm:
7145   case X86::VMOVDQArm:
7146   case X86::VMOVDQUrm:
7147   case X86::VMOVAPSYrm:
7148   case X86::VMOVUPSYrm:
7149   case X86::VMOVAPDYrm:
7150   case X86::VMOVUPDYrm:
7151   case X86::VMOVDQAYrm:
7152   case X86::VMOVDQUYrm:
7153   // AVX512 load instructions
7154   case X86::VMOVSSZrm:
7155   case X86::VMOVSSZrm_alt:
7156   case X86::VMOVSDZrm:
7157   case X86::VMOVSDZrm_alt:
7158   case X86::VMOVAPSZ128rm:
7159   case X86::VMOVUPSZ128rm:
7160   case X86::VMOVAPSZ128rm_NOVLX:
7161   case X86::VMOVUPSZ128rm_NOVLX:
7162   case X86::VMOVAPDZ128rm:
7163   case X86::VMOVUPDZ128rm:
7164   case X86::VMOVDQU8Z128rm:
7165   case X86::VMOVDQU16Z128rm:
7166   case X86::VMOVDQA32Z128rm:
7167   case X86::VMOVDQU32Z128rm:
7168   case X86::VMOVDQA64Z128rm:
7169   case X86::VMOVDQU64Z128rm:
7170   case X86::VMOVAPSZ256rm:
7171   case X86::VMOVUPSZ256rm:
7172   case X86::VMOVAPSZ256rm_NOVLX:
7173   case X86::VMOVUPSZ256rm_NOVLX:
7174   case X86::VMOVAPDZ256rm:
7175   case X86::VMOVUPDZ256rm:
7176   case X86::VMOVDQU8Z256rm:
7177   case X86::VMOVDQU16Z256rm:
7178   case X86::VMOVDQA32Z256rm:
7179   case X86::VMOVDQU32Z256rm:
7180   case X86::VMOVDQA64Z256rm:
7181   case X86::VMOVDQU64Z256rm:
7182   case X86::VMOVAPSZrm:
7183   case X86::VMOVUPSZrm:
7184   case X86::VMOVAPDZrm:
7185   case X86::VMOVUPDZrm:
7186   case X86::VMOVDQU8Zrm:
7187   case X86::VMOVDQU16Zrm:
7188   case X86::VMOVDQA32Zrm:
7189   case X86::VMOVDQU32Zrm:
7190   case X86::VMOVDQA64Zrm:
7191   case X86::VMOVDQU64Zrm:
7192   case X86::KMOVBkm:
7193   case X86::KMOVWkm:
7194   case X86::KMOVDkm:
7195   case X86::KMOVQkm:
7196     break;
7197   }
7198   switch (Opc2) {
7199   default: return false;
7200   case X86::MOV8rm:
7201   case X86::MOV16rm:
7202   case X86::MOV32rm:
7203   case X86::MOV64rm:
7204   case X86::LD_Fp32m:
7205   case X86::LD_Fp64m:
7206   case X86::LD_Fp80m:
7207   case X86::MOVSSrm:
7208   case X86::MOVSSrm_alt:
7209   case X86::MOVSDrm:
7210   case X86::MOVSDrm_alt:
7211   case X86::MMX_MOVD64rm:
7212   case X86::MMX_MOVQ64rm:
7213   case X86::MOVAPSrm:
7214   case X86::MOVUPSrm:
7215   case X86::MOVAPDrm:
7216   case X86::MOVUPDrm:
7217   case X86::MOVDQArm:
7218   case X86::MOVDQUrm:
7219   // AVX load instructions
7220   case X86::VMOVSSrm:
7221   case X86::VMOVSSrm_alt:
7222   case X86::VMOVSDrm:
7223   case X86::VMOVSDrm_alt:
7224   case X86::VMOVAPSrm:
7225   case X86::VMOVUPSrm:
7226   case X86::VMOVAPDrm:
7227   case X86::VMOVUPDrm:
7228   case X86::VMOVDQArm:
7229   case X86::VMOVDQUrm:
7230   case X86::VMOVAPSYrm:
7231   case X86::VMOVUPSYrm:
7232   case X86::VMOVAPDYrm:
7233   case X86::VMOVUPDYrm:
7234   case X86::VMOVDQAYrm:
7235   case X86::VMOVDQUYrm:
7236   // AVX512 load instructions
7237   case X86::VMOVSSZrm:
7238   case X86::VMOVSSZrm_alt:
7239   case X86::VMOVSDZrm:
7240   case X86::VMOVSDZrm_alt:
7241   case X86::VMOVAPSZ128rm:
7242   case X86::VMOVUPSZ128rm:
7243   case X86::VMOVAPSZ128rm_NOVLX:
7244   case X86::VMOVUPSZ128rm_NOVLX:
7245   case X86::VMOVAPDZ128rm:
7246   case X86::VMOVUPDZ128rm:
7247   case X86::VMOVDQU8Z128rm:
7248   case X86::VMOVDQU16Z128rm:
7249   case X86::VMOVDQA32Z128rm:
7250   case X86::VMOVDQU32Z128rm:
7251   case X86::VMOVDQA64Z128rm:
7252   case X86::VMOVDQU64Z128rm:
7253   case X86::VMOVAPSZ256rm:
7254   case X86::VMOVUPSZ256rm:
7255   case X86::VMOVAPSZ256rm_NOVLX:
7256   case X86::VMOVUPSZ256rm_NOVLX:
7257   case X86::VMOVAPDZ256rm:
7258   case X86::VMOVUPDZ256rm:
7259   case X86::VMOVDQU8Z256rm:
7260   case X86::VMOVDQU16Z256rm:
7261   case X86::VMOVDQA32Z256rm:
7262   case X86::VMOVDQU32Z256rm:
7263   case X86::VMOVDQA64Z256rm:
7264   case X86::VMOVDQU64Z256rm:
7265   case X86::VMOVAPSZrm:
7266   case X86::VMOVUPSZrm:
7267   case X86::VMOVAPDZrm:
7268   case X86::VMOVUPDZrm:
7269   case X86::VMOVDQU8Zrm:
7270   case X86::VMOVDQU16Zrm:
7271   case X86::VMOVDQA32Zrm:
7272   case X86::VMOVDQU32Zrm:
7273   case X86::VMOVDQA64Zrm:
7274   case X86::VMOVDQU64Zrm:
7275   case X86::KMOVBkm:
7276   case X86::KMOVWkm:
7277   case X86::KMOVDkm:
7278   case X86::KMOVQkm:
7279     break;
7280   }
7281 
7282   // Lambda to check if both the loads have the same value for an operand index.
7283   auto HasSameOp = [&](int I) {
7284     return Load1->getOperand(I) == Load2->getOperand(I);
7285   };
7286 
7287   // All operands except the displacement should match.
7288   if (!HasSameOp(X86::AddrBaseReg) || !HasSameOp(X86::AddrScaleAmt) ||
7289       !HasSameOp(X86::AddrIndexReg) || !HasSameOp(X86::AddrSegmentReg))
7290     return false;
7291 
7292   // Chain Operand must be the same.
7293   if (!HasSameOp(5))
7294     return false;
7295 
7296   // Now let's examine if the displacements are constants.
7297   auto Disp1 = dyn_cast<ConstantSDNode>(Load1->getOperand(X86::AddrDisp));
7298   auto Disp2 = dyn_cast<ConstantSDNode>(Load2->getOperand(X86::AddrDisp));
7299   if (!Disp1 || !Disp2)
7300     return false;
7301 
7302   Offset1 = Disp1->getSExtValue();
7303   Offset2 = Disp2->getSExtValue();
7304   return true;
7305 }
7306 
7307 bool X86InstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
7308                                            int64_t Offset1, int64_t Offset2,
7309                                            unsigned NumLoads) const {
7310   assert(Offset2 > Offset1);
7311   if ((Offset2 - Offset1) / 8 > 64)
7312     return false;
7313 
7314   unsigned Opc1 = Load1->getMachineOpcode();
7315   unsigned Opc2 = Load2->getMachineOpcode();
7316   if (Opc1 != Opc2)
7317     return false;  // FIXME: overly conservative?
7318 
7319   switch (Opc1) {
7320   default: break;
7321   case X86::LD_Fp32m:
7322   case X86::LD_Fp64m:
7323   case X86::LD_Fp80m:
7324   case X86::MMX_MOVD64rm:
7325   case X86::MMX_MOVQ64rm:
7326     return false;
7327   }
7328 
7329   EVT VT = Load1->getValueType(0);
7330   switch (VT.getSimpleVT().SimpleTy) {
7331   default:
7332     // XMM registers. In 64-bit mode we can be a bit more aggressive since we
7333     // have 16 of them to play with.
7334     if (Subtarget.is64Bit()) {
7335       if (NumLoads >= 3)
7336         return false;
7337     } else if (NumLoads) {
7338       return false;
7339     }
7340     break;
7341   case MVT::i8:
7342   case MVT::i16:
7343   case MVT::i32:
7344   case MVT::i64:
7345   case MVT::f32:
7346   case MVT::f64:
7347     if (NumLoads)
7348       return false;
7349     break;
7350   }
7351 
7352   return true;
7353 }
7354 
7355 bool X86InstrInfo::isSchedulingBoundary(const MachineInstr &MI,
7356                                         const MachineBasicBlock *MBB,
7357                                         const MachineFunction &MF) const {
7358 
7359   // ENDBR instructions should not be scheduled around.
7360   unsigned Opcode = MI.getOpcode();
7361   if (Opcode == X86::ENDBR64 || Opcode == X86::ENDBR32 ||
7362       Opcode == X86::PLDTILECFGV)
7363     return true;
7364 
7365   return TargetInstrInfo::isSchedulingBoundary(MI, MBB, MF);
7366 }
7367 
7368 bool X86InstrInfo::
7369 reverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
7370   assert(Cond.size() == 1 && "Invalid X86 branch condition!");
7371   X86::CondCode CC = static_cast<X86::CondCode>(Cond[0].getImm());
7372   Cond[0].setImm(GetOppositeBranchCondition(CC));
7373   return false;
7374 }
7375 
7376 bool X86InstrInfo::
7377 isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
7378   // FIXME: Return false for x87 stack register classes for now. We can't
7379   // allow any loads of these registers before FpGet_ST0_80.
7380   return !(RC == &X86::CCRRegClass || RC == &X86::DFCCRRegClass ||
7381            RC == &X86::RFP32RegClass || RC == &X86::RFP64RegClass ||
7382            RC == &X86::RFP80RegClass);
7383 }
7384 
7385 /// Return a virtual register initialized with the
7386 /// the global base register value. Output instructions required to
7387 /// initialize the register in the function entry block, if necessary.
7388 ///
7389 /// TODO: Eliminate this and move the code to X86MachineFunctionInfo.
7390 ///
7391 unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const {
7392   assert((!Subtarget.is64Bit() ||
7393           MF->getTarget().getCodeModel() == CodeModel::Medium ||
7394           MF->getTarget().getCodeModel() == CodeModel::Large) &&
7395          "X86-64 PIC uses RIP relative addressing");
7396 
7397   X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
7398   Register GlobalBaseReg = X86FI->getGlobalBaseReg();
7399   if (GlobalBaseReg != 0)
7400     return GlobalBaseReg;
7401 
7402   // Create the register. The code to initialize it is inserted
7403   // later, by the CGBR pass (below).
7404   MachineRegisterInfo &RegInfo = MF->getRegInfo();
7405   GlobalBaseReg = RegInfo.createVirtualRegister(
7406       Subtarget.is64Bit() ? &X86::GR64_NOSPRegClass : &X86::GR32_NOSPRegClass);
7407   X86FI->setGlobalBaseReg(GlobalBaseReg);
7408   return GlobalBaseReg;
7409 }
7410 
7411 // These are the replaceable SSE instructions. Some of these have Int variants
7412 // that we don't include here. We don't want to replace instructions selected
7413 // by intrinsics.
7414 static const uint16_t ReplaceableInstrs[][3] = {
7415   //PackedSingle     PackedDouble    PackedInt
7416   { X86::MOVAPSmr,   X86::MOVAPDmr,  X86::MOVDQAmr  },
7417   { X86::MOVAPSrm,   X86::MOVAPDrm,  X86::MOVDQArm  },
7418   { X86::MOVAPSrr,   X86::MOVAPDrr,  X86::MOVDQArr  },
7419   { X86::MOVUPSmr,   X86::MOVUPDmr,  X86::MOVDQUmr  },
7420   { X86::MOVUPSrm,   X86::MOVUPDrm,  X86::MOVDQUrm  },
7421   { X86::MOVLPSmr,   X86::MOVLPDmr,  X86::MOVPQI2QImr },
7422   { X86::MOVSDmr,    X86::MOVSDmr,   X86::MOVPQI2QImr },
7423   { X86::MOVSSmr,    X86::MOVSSmr,   X86::MOVPDI2DImr },
7424   { X86::MOVSDrm,    X86::MOVSDrm,   X86::MOVQI2PQIrm },
7425   { X86::MOVSDrm_alt,X86::MOVSDrm_alt,X86::MOVQI2PQIrm },
7426   { X86::MOVSSrm,    X86::MOVSSrm,   X86::MOVDI2PDIrm },
7427   { X86::MOVSSrm_alt,X86::MOVSSrm_alt,X86::MOVDI2PDIrm },
7428   { X86::MOVNTPSmr,  X86::MOVNTPDmr, X86::MOVNTDQmr },
7429   { X86::ANDNPSrm,   X86::ANDNPDrm,  X86::PANDNrm   },
7430   { X86::ANDNPSrr,   X86::ANDNPDrr,  X86::PANDNrr   },
7431   { X86::ANDPSrm,    X86::ANDPDrm,   X86::PANDrm    },
7432   { X86::ANDPSrr,    X86::ANDPDrr,   X86::PANDrr    },
7433   { X86::ORPSrm,     X86::ORPDrm,    X86::PORrm     },
7434   { X86::ORPSrr,     X86::ORPDrr,    X86::PORrr     },
7435   { X86::XORPSrm,    X86::XORPDrm,   X86::PXORrm    },
7436   { X86::XORPSrr,    X86::XORPDrr,   X86::PXORrr    },
7437   { X86::UNPCKLPDrm, X86::UNPCKLPDrm, X86::PUNPCKLQDQrm },
7438   { X86::MOVLHPSrr,  X86::UNPCKLPDrr, X86::PUNPCKLQDQrr },
7439   { X86::UNPCKHPDrm, X86::UNPCKHPDrm, X86::PUNPCKHQDQrm },
7440   { X86::UNPCKHPDrr, X86::UNPCKHPDrr, X86::PUNPCKHQDQrr },
7441   { X86::UNPCKLPSrm, X86::UNPCKLPSrm, X86::PUNPCKLDQrm },
7442   { X86::UNPCKLPSrr, X86::UNPCKLPSrr, X86::PUNPCKLDQrr },
7443   { X86::UNPCKHPSrm, X86::UNPCKHPSrm, X86::PUNPCKHDQrm },
7444   { X86::UNPCKHPSrr, X86::UNPCKHPSrr, X86::PUNPCKHDQrr },
7445   { X86::EXTRACTPSmr, X86::EXTRACTPSmr, X86::PEXTRDmr },
7446   { X86::EXTRACTPSrr, X86::EXTRACTPSrr, X86::PEXTRDrr },
7447   // AVX 128-bit support
7448   { X86::VMOVAPSmr,  X86::VMOVAPDmr,  X86::VMOVDQAmr  },
7449   { X86::VMOVAPSrm,  X86::VMOVAPDrm,  X86::VMOVDQArm  },
7450   { X86::VMOVAPSrr,  X86::VMOVAPDrr,  X86::VMOVDQArr  },
7451   { X86::VMOVUPSmr,  X86::VMOVUPDmr,  X86::VMOVDQUmr  },
7452   { X86::VMOVUPSrm,  X86::VMOVUPDrm,  X86::VMOVDQUrm  },
7453   { X86::VMOVLPSmr,  X86::VMOVLPDmr,  X86::VMOVPQI2QImr },
7454   { X86::VMOVSDmr,   X86::VMOVSDmr,   X86::VMOVPQI2QImr },
7455   { X86::VMOVSSmr,   X86::VMOVSSmr,   X86::VMOVPDI2DImr },
7456   { X86::VMOVSDrm,   X86::VMOVSDrm,   X86::VMOVQI2PQIrm },
7457   { X86::VMOVSDrm_alt,X86::VMOVSDrm_alt,X86::VMOVQI2PQIrm },
7458   { X86::VMOVSSrm,   X86::VMOVSSrm,   X86::VMOVDI2PDIrm },
7459   { X86::VMOVSSrm_alt,X86::VMOVSSrm_alt,X86::VMOVDI2PDIrm },
7460   { X86::VMOVNTPSmr, X86::VMOVNTPDmr, X86::VMOVNTDQmr },
7461   { X86::VANDNPSrm,  X86::VANDNPDrm,  X86::VPANDNrm   },
7462   { X86::VANDNPSrr,  X86::VANDNPDrr,  X86::VPANDNrr   },
7463   { X86::VANDPSrm,   X86::VANDPDrm,   X86::VPANDrm    },
7464   { X86::VANDPSrr,   X86::VANDPDrr,   X86::VPANDrr    },
7465   { X86::VORPSrm,    X86::VORPDrm,    X86::VPORrm     },
7466   { X86::VORPSrr,    X86::VORPDrr,    X86::VPORrr     },
7467   { X86::VXORPSrm,   X86::VXORPDrm,   X86::VPXORrm    },
7468   { X86::VXORPSrr,   X86::VXORPDrr,   X86::VPXORrr    },
7469   { X86::VUNPCKLPDrm, X86::VUNPCKLPDrm, X86::VPUNPCKLQDQrm },
7470   { X86::VMOVLHPSrr,  X86::VUNPCKLPDrr, X86::VPUNPCKLQDQrr },
7471   { X86::VUNPCKHPDrm, X86::VUNPCKHPDrm, X86::VPUNPCKHQDQrm },
7472   { X86::VUNPCKHPDrr, X86::VUNPCKHPDrr, X86::VPUNPCKHQDQrr },
7473   { X86::VUNPCKLPSrm, X86::VUNPCKLPSrm, X86::VPUNPCKLDQrm },
7474   { X86::VUNPCKLPSrr, X86::VUNPCKLPSrr, X86::VPUNPCKLDQrr },
7475   { X86::VUNPCKHPSrm, X86::VUNPCKHPSrm, X86::VPUNPCKHDQrm },
7476   { X86::VUNPCKHPSrr, X86::VUNPCKHPSrr, X86::VPUNPCKHDQrr },
7477   { X86::VEXTRACTPSmr, X86::VEXTRACTPSmr, X86::VPEXTRDmr },
7478   { X86::VEXTRACTPSrr, X86::VEXTRACTPSrr, X86::VPEXTRDrr },
7479   // AVX 256-bit support
7480   { X86::VMOVAPSYmr,   X86::VMOVAPDYmr,   X86::VMOVDQAYmr  },
7481   { X86::VMOVAPSYrm,   X86::VMOVAPDYrm,   X86::VMOVDQAYrm  },
7482   { X86::VMOVAPSYrr,   X86::VMOVAPDYrr,   X86::VMOVDQAYrr  },
7483   { X86::VMOVUPSYmr,   X86::VMOVUPDYmr,   X86::VMOVDQUYmr  },
7484   { X86::VMOVUPSYrm,   X86::VMOVUPDYrm,   X86::VMOVDQUYrm  },
7485   { X86::VMOVNTPSYmr,  X86::VMOVNTPDYmr,  X86::VMOVNTDQYmr },
7486   { X86::VPERMPSYrm,   X86::VPERMPSYrm,   X86::VPERMDYrm },
7487   { X86::VPERMPSYrr,   X86::VPERMPSYrr,   X86::VPERMDYrr },
7488   { X86::VPERMPDYmi,   X86::VPERMPDYmi,   X86::VPERMQYmi },
7489   { X86::VPERMPDYri,   X86::VPERMPDYri,   X86::VPERMQYri },
7490   // AVX512 support
7491   { X86::VMOVLPSZ128mr,  X86::VMOVLPDZ128mr,  X86::VMOVPQI2QIZmr  },
7492   { X86::VMOVNTPSZ128mr, X86::VMOVNTPDZ128mr, X86::VMOVNTDQZ128mr },
7493   { X86::VMOVNTPSZ256mr, X86::VMOVNTPDZ256mr, X86::VMOVNTDQZ256mr },
7494   { X86::VMOVNTPSZmr,    X86::VMOVNTPDZmr,    X86::VMOVNTDQZmr    },
7495   { X86::VMOVSDZmr,      X86::VMOVSDZmr,      X86::VMOVPQI2QIZmr  },
7496   { X86::VMOVSSZmr,      X86::VMOVSSZmr,      X86::VMOVPDI2DIZmr  },
7497   { X86::VMOVSDZrm,      X86::VMOVSDZrm,      X86::VMOVQI2PQIZrm  },
7498   { X86::VMOVSDZrm_alt,  X86::VMOVSDZrm_alt,  X86::VMOVQI2PQIZrm  },
7499   { X86::VMOVSSZrm,      X86::VMOVSSZrm,      X86::VMOVDI2PDIZrm  },
7500   { X86::VMOVSSZrm_alt,  X86::VMOVSSZrm_alt,  X86::VMOVDI2PDIZrm  },
7501   { X86::VBROADCASTSSZ128rr,X86::VBROADCASTSSZ128rr,X86::VPBROADCASTDZ128rr },
7502   { X86::VBROADCASTSSZ128rm,X86::VBROADCASTSSZ128rm,X86::VPBROADCASTDZ128rm },
7503   { X86::VBROADCASTSSZ256rr,X86::VBROADCASTSSZ256rr,X86::VPBROADCASTDZ256rr },
7504   { X86::VBROADCASTSSZ256rm,X86::VBROADCASTSSZ256rm,X86::VPBROADCASTDZ256rm },
7505   { X86::VBROADCASTSSZrr,   X86::VBROADCASTSSZrr,   X86::VPBROADCASTDZrr },
7506   { X86::VBROADCASTSSZrm,   X86::VBROADCASTSSZrm,   X86::VPBROADCASTDZrm },
7507   { X86::VMOVDDUPZ128rr,    X86::VMOVDDUPZ128rr,    X86::VPBROADCASTQZ128rr },
7508   { X86::VMOVDDUPZ128rm,    X86::VMOVDDUPZ128rm,    X86::VPBROADCASTQZ128rm },
7509   { X86::VBROADCASTSDZ256rr,X86::VBROADCASTSDZ256rr,X86::VPBROADCASTQZ256rr },
7510   { X86::VBROADCASTSDZ256rm,X86::VBROADCASTSDZ256rm,X86::VPBROADCASTQZ256rm },
7511   { X86::VBROADCASTSDZrr,   X86::VBROADCASTSDZrr,   X86::VPBROADCASTQZrr },
7512   { X86::VBROADCASTSDZrm,   X86::VBROADCASTSDZrm,   X86::VPBROADCASTQZrm },
7513   { X86::VINSERTF32x4Zrr,   X86::VINSERTF32x4Zrr,   X86::VINSERTI32x4Zrr },
7514   { X86::VINSERTF32x4Zrm,   X86::VINSERTF32x4Zrm,   X86::VINSERTI32x4Zrm },
7515   { X86::VINSERTF32x8Zrr,   X86::VINSERTF32x8Zrr,   X86::VINSERTI32x8Zrr },
7516   { X86::VINSERTF32x8Zrm,   X86::VINSERTF32x8Zrm,   X86::VINSERTI32x8Zrm },
7517   { X86::VINSERTF64x2Zrr,   X86::VINSERTF64x2Zrr,   X86::VINSERTI64x2Zrr },
7518   { X86::VINSERTF64x2Zrm,   X86::VINSERTF64x2Zrm,   X86::VINSERTI64x2Zrm },
7519   { X86::VINSERTF64x4Zrr,   X86::VINSERTF64x4Zrr,   X86::VINSERTI64x4Zrr },
7520   { X86::VINSERTF64x4Zrm,   X86::VINSERTF64x4Zrm,   X86::VINSERTI64x4Zrm },
7521   { X86::VINSERTF32x4Z256rr,X86::VINSERTF32x4Z256rr,X86::VINSERTI32x4Z256rr },
7522   { X86::VINSERTF32x4Z256rm,X86::VINSERTF32x4Z256rm,X86::VINSERTI32x4Z256rm },
7523   { X86::VINSERTF64x2Z256rr,X86::VINSERTF64x2Z256rr,X86::VINSERTI64x2Z256rr },
7524   { X86::VINSERTF64x2Z256rm,X86::VINSERTF64x2Z256rm,X86::VINSERTI64x2Z256rm },
7525   { X86::VEXTRACTF32x4Zrr,   X86::VEXTRACTF32x4Zrr,   X86::VEXTRACTI32x4Zrr },
7526   { X86::VEXTRACTF32x4Zmr,   X86::VEXTRACTF32x4Zmr,   X86::VEXTRACTI32x4Zmr },
7527   { X86::VEXTRACTF32x8Zrr,   X86::VEXTRACTF32x8Zrr,   X86::VEXTRACTI32x8Zrr },
7528   { X86::VEXTRACTF32x8Zmr,   X86::VEXTRACTF32x8Zmr,   X86::VEXTRACTI32x8Zmr },
7529   { X86::VEXTRACTF64x2Zrr,   X86::VEXTRACTF64x2Zrr,   X86::VEXTRACTI64x2Zrr },
7530   { X86::VEXTRACTF64x2Zmr,   X86::VEXTRACTF64x2Zmr,   X86::VEXTRACTI64x2Zmr },
7531   { X86::VEXTRACTF64x4Zrr,   X86::VEXTRACTF64x4Zrr,   X86::VEXTRACTI64x4Zrr },
7532   { X86::VEXTRACTF64x4Zmr,   X86::VEXTRACTF64x4Zmr,   X86::VEXTRACTI64x4Zmr },
7533   { X86::VEXTRACTF32x4Z256rr,X86::VEXTRACTF32x4Z256rr,X86::VEXTRACTI32x4Z256rr },
7534   { X86::VEXTRACTF32x4Z256mr,X86::VEXTRACTF32x4Z256mr,X86::VEXTRACTI32x4Z256mr },
7535   { X86::VEXTRACTF64x2Z256rr,X86::VEXTRACTF64x2Z256rr,X86::VEXTRACTI64x2Z256rr },
7536   { X86::VEXTRACTF64x2Z256mr,X86::VEXTRACTF64x2Z256mr,X86::VEXTRACTI64x2Z256mr },
7537   { X86::VPERMILPSmi,        X86::VPERMILPSmi,        X86::VPSHUFDmi },
7538   { X86::VPERMILPSri,        X86::VPERMILPSri,        X86::VPSHUFDri },
7539   { X86::VPERMILPSZ128mi,    X86::VPERMILPSZ128mi,    X86::VPSHUFDZ128mi },
7540   { X86::VPERMILPSZ128ri,    X86::VPERMILPSZ128ri,    X86::VPSHUFDZ128ri },
7541   { X86::VPERMILPSZ256mi,    X86::VPERMILPSZ256mi,    X86::VPSHUFDZ256mi },
7542   { X86::VPERMILPSZ256ri,    X86::VPERMILPSZ256ri,    X86::VPSHUFDZ256ri },
7543   { X86::VPERMILPSZmi,       X86::VPERMILPSZmi,       X86::VPSHUFDZmi },
7544   { X86::VPERMILPSZri,       X86::VPERMILPSZri,       X86::VPSHUFDZri },
7545   { X86::VPERMPSZ256rm,      X86::VPERMPSZ256rm,      X86::VPERMDZ256rm },
7546   { X86::VPERMPSZ256rr,      X86::VPERMPSZ256rr,      X86::VPERMDZ256rr },
7547   { X86::VPERMPDZ256mi,      X86::VPERMPDZ256mi,      X86::VPERMQZ256mi },
7548   { X86::VPERMPDZ256ri,      X86::VPERMPDZ256ri,      X86::VPERMQZ256ri },
7549   { X86::VPERMPDZ256rm,      X86::VPERMPDZ256rm,      X86::VPERMQZ256rm },
7550   { X86::VPERMPDZ256rr,      X86::VPERMPDZ256rr,      X86::VPERMQZ256rr },
7551   { X86::VPERMPSZrm,         X86::VPERMPSZrm,         X86::VPERMDZrm },
7552   { X86::VPERMPSZrr,         X86::VPERMPSZrr,         X86::VPERMDZrr },
7553   { X86::VPERMPDZmi,         X86::VPERMPDZmi,         X86::VPERMQZmi },
7554   { X86::VPERMPDZri,         X86::VPERMPDZri,         X86::VPERMQZri },
7555   { X86::VPERMPDZrm,         X86::VPERMPDZrm,         X86::VPERMQZrm },
7556   { X86::VPERMPDZrr,         X86::VPERMPDZrr,         X86::VPERMQZrr },
7557   { X86::VUNPCKLPDZ256rm,    X86::VUNPCKLPDZ256rm,    X86::VPUNPCKLQDQZ256rm },
7558   { X86::VUNPCKLPDZ256rr,    X86::VUNPCKLPDZ256rr,    X86::VPUNPCKLQDQZ256rr },
7559   { X86::VUNPCKHPDZ256rm,    X86::VUNPCKHPDZ256rm,    X86::VPUNPCKHQDQZ256rm },
7560   { X86::VUNPCKHPDZ256rr,    X86::VUNPCKHPDZ256rr,    X86::VPUNPCKHQDQZ256rr },
7561   { X86::VUNPCKLPSZ256rm,    X86::VUNPCKLPSZ256rm,    X86::VPUNPCKLDQZ256rm },
7562   { X86::VUNPCKLPSZ256rr,    X86::VUNPCKLPSZ256rr,    X86::VPUNPCKLDQZ256rr },
7563   { X86::VUNPCKHPSZ256rm,    X86::VUNPCKHPSZ256rm,    X86::VPUNPCKHDQZ256rm },
7564   { X86::VUNPCKHPSZ256rr,    X86::VUNPCKHPSZ256rr,    X86::VPUNPCKHDQZ256rr },
7565   { X86::VUNPCKLPDZ128rm,    X86::VUNPCKLPDZ128rm,    X86::VPUNPCKLQDQZ128rm },
7566   { X86::VMOVLHPSZrr,        X86::VUNPCKLPDZ128rr,    X86::VPUNPCKLQDQZ128rr },
7567   { X86::VUNPCKHPDZ128rm,    X86::VUNPCKHPDZ128rm,    X86::VPUNPCKHQDQZ128rm },
7568   { X86::VUNPCKHPDZ128rr,    X86::VUNPCKHPDZ128rr,    X86::VPUNPCKHQDQZ128rr },
7569   { X86::VUNPCKLPSZ128rm,    X86::VUNPCKLPSZ128rm,    X86::VPUNPCKLDQZ128rm },
7570   { X86::VUNPCKLPSZ128rr,    X86::VUNPCKLPSZ128rr,    X86::VPUNPCKLDQZ128rr },
7571   { X86::VUNPCKHPSZ128rm,    X86::VUNPCKHPSZ128rm,    X86::VPUNPCKHDQZ128rm },
7572   { X86::VUNPCKHPSZ128rr,    X86::VUNPCKHPSZ128rr,    X86::VPUNPCKHDQZ128rr },
7573   { X86::VUNPCKLPDZrm,       X86::VUNPCKLPDZrm,       X86::VPUNPCKLQDQZrm },
7574   { X86::VUNPCKLPDZrr,       X86::VUNPCKLPDZrr,       X86::VPUNPCKLQDQZrr },
7575   { X86::VUNPCKHPDZrm,       X86::VUNPCKHPDZrm,       X86::VPUNPCKHQDQZrm },
7576   { X86::VUNPCKHPDZrr,       X86::VUNPCKHPDZrr,       X86::VPUNPCKHQDQZrr },
7577   { X86::VUNPCKLPSZrm,       X86::VUNPCKLPSZrm,       X86::VPUNPCKLDQZrm },
7578   { X86::VUNPCKLPSZrr,       X86::VUNPCKLPSZrr,       X86::VPUNPCKLDQZrr },
7579   { X86::VUNPCKHPSZrm,       X86::VUNPCKHPSZrm,       X86::VPUNPCKHDQZrm },
7580   { X86::VUNPCKHPSZrr,       X86::VUNPCKHPSZrr,       X86::VPUNPCKHDQZrr },
7581   { X86::VEXTRACTPSZmr,      X86::VEXTRACTPSZmr,      X86::VPEXTRDZmr },
7582   { X86::VEXTRACTPSZrr,      X86::VEXTRACTPSZrr,      X86::VPEXTRDZrr },
7583 };
7584 
7585 static const uint16_t ReplaceableInstrsAVX2[][3] = {
7586   //PackedSingle       PackedDouble       PackedInt
7587   { X86::VANDNPSYrm,   X86::VANDNPDYrm,   X86::VPANDNYrm   },
7588   { X86::VANDNPSYrr,   X86::VANDNPDYrr,   X86::VPANDNYrr   },
7589   { X86::VANDPSYrm,    X86::VANDPDYrm,    X86::VPANDYrm    },
7590   { X86::VANDPSYrr,    X86::VANDPDYrr,    X86::VPANDYrr    },
7591   { X86::VORPSYrm,     X86::VORPDYrm,     X86::VPORYrm     },
7592   { X86::VORPSYrr,     X86::VORPDYrr,     X86::VPORYrr     },
7593   { X86::VXORPSYrm,    X86::VXORPDYrm,    X86::VPXORYrm    },
7594   { X86::VXORPSYrr,    X86::VXORPDYrr,    X86::VPXORYrr    },
7595   { X86::VPERM2F128rm,   X86::VPERM2F128rm,   X86::VPERM2I128rm },
7596   { X86::VPERM2F128rr,   X86::VPERM2F128rr,   X86::VPERM2I128rr },
7597   { X86::VBROADCASTSSrm, X86::VBROADCASTSSrm, X86::VPBROADCASTDrm},
7598   { X86::VBROADCASTSSrr, X86::VBROADCASTSSrr, X86::VPBROADCASTDrr},
7599   { X86::VMOVDDUPrm,     X86::VMOVDDUPrm,     X86::VPBROADCASTQrm},
7600   { X86::VMOVDDUPrr,     X86::VMOVDDUPrr,     X86::VPBROADCASTQrr},
7601   { X86::VBROADCASTSSYrr, X86::VBROADCASTSSYrr, X86::VPBROADCASTDYrr},
7602   { X86::VBROADCASTSSYrm, X86::VBROADCASTSSYrm, X86::VPBROADCASTDYrm},
7603   { X86::VBROADCASTSDYrr, X86::VBROADCASTSDYrr, X86::VPBROADCASTQYrr},
7604   { X86::VBROADCASTSDYrm, X86::VBROADCASTSDYrm, X86::VPBROADCASTQYrm},
7605   { X86::VBROADCASTF128,  X86::VBROADCASTF128,  X86::VBROADCASTI128 },
7606   { X86::VBLENDPSYrri,    X86::VBLENDPSYrri,    X86::VPBLENDDYrri },
7607   { X86::VBLENDPSYrmi,    X86::VBLENDPSYrmi,    X86::VPBLENDDYrmi },
7608   { X86::VPERMILPSYmi,    X86::VPERMILPSYmi,    X86::VPSHUFDYmi },
7609   { X86::VPERMILPSYri,    X86::VPERMILPSYri,    X86::VPSHUFDYri },
7610   { X86::VUNPCKLPDYrm,    X86::VUNPCKLPDYrm,    X86::VPUNPCKLQDQYrm },
7611   { X86::VUNPCKLPDYrr,    X86::VUNPCKLPDYrr,    X86::VPUNPCKLQDQYrr },
7612   { X86::VUNPCKHPDYrm,    X86::VUNPCKHPDYrm,    X86::VPUNPCKHQDQYrm },
7613   { X86::VUNPCKHPDYrr,    X86::VUNPCKHPDYrr,    X86::VPUNPCKHQDQYrr },
7614   { X86::VUNPCKLPSYrm,    X86::VUNPCKLPSYrm,    X86::VPUNPCKLDQYrm },
7615   { X86::VUNPCKLPSYrr,    X86::VUNPCKLPSYrr,    X86::VPUNPCKLDQYrr },
7616   { X86::VUNPCKHPSYrm,    X86::VUNPCKHPSYrm,    X86::VPUNPCKHDQYrm },
7617   { X86::VUNPCKHPSYrr,    X86::VUNPCKHPSYrr,    X86::VPUNPCKHDQYrr },
7618 };
7619 
7620 static const uint16_t ReplaceableInstrsFP[][3] = {
7621   //PackedSingle         PackedDouble
7622   { X86::MOVLPSrm,       X86::MOVLPDrm,      X86::INSTRUCTION_LIST_END },
7623   { X86::MOVHPSrm,       X86::MOVHPDrm,      X86::INSTRUCTION_LIST_END },
7624   { X86::MOVHPSmr,       X86::MOVHPDmr,      X86::INSTRUCTION_LIST_END },
7625   { X86::VMOVLPSrm,      X86::VMOVLPDrm,     X86::INSTRUCTION_LIST_END },
7626   { X86::VMOVHPSrm,      X86::VMOVHPDrm,     X86::INSTRUCTION_LIST_END },
7627   { X86::VMOVHPSmr,      X86::VMOVHPDmr,     X86::INSTRUCTION_LIST_END },
7628   { X86::VMOVLPSZ128rm,  X86::VMOVLPDZ128rm, X86::INSTRUCTION_LIST_END },
7629   { X86::VMOVHPSZ128rm,  X86::VMOVHPDZ128rm, X86::INSTRUCTION_LIST_END },
7630   { X86::VMOVHPSZ128mr,  X86::VMOVHPDZ128mr, X86::INSTRUCTION_LIST_END },
7631 };
7632 
7633 static const uint16_t ReplaceableInstrsAVX2InsertExtract[][3] = {
7634   //PackedSingle       PackedDouble       PackedInt
7635   { X86::VEXTRACTF128mr, X86::VEXTRACTF128mr, X86::VEXTRACTI128mr },
7636   { X86::VEXTRACTF128rr, X86::VEXTRACTF128rr, X86::VEXTRACTI128rr },
7637   { X86::VINSERTF128rm,  X86::VINSERTF128rm,  X86::VINSERTI128rm },
7638   { X86::VINSERTF128rr,  X86::VINSERTF128rr,  X86::VINSERTI128rr },
7639 };
7640 
7641 static const uint16_t ReplaceableInstrsAVX512[][4] = {
7642   // Two integer columns for 64-bit and 32-bit elements.
7643   //PackedSingle        PackedDouble        PackedInt             PackedInt
7644   { X86::VMOVAPSZ128mr, X86::VMOVAPDZ128mr, X86::VMOVDQA64Z128mr, X86::VMOVDQA32Z128mr  },
7645   { X86::VMOVAPSZ128rm, X86::VMOVAPDZ128rm, X86::VMOVDQA64Z128rm, X86::VMOVDQA32Z128rm  },
7646   { X86::VMOVAPSZ128rr, X86::VMOVAPDZ128rr, X86::VMOVDQA64Z128rr, X86::VMOVDQA32Z128rr  },
7647   { X86::VMOVUPSZ128mr, X86::VMOVUPDZ128mr, X86::VMOVDQU64Z128mr, X86::VMOVDQU32Z128mr  },
7648   { X86::VMOVUPSZ128rm, X86::VMOVUPDZ128rm, X86::VMOVDQU64Z128rm, X86::VMOVDQU32Z128rm  },
7649   { X86::VMOVAPSZ256mr, X86::VMOVAPDZ256mr, X86::VMOVDQA64Z256mr, X86::VMOVDQA32Z256mr  },
7650   { X86::VMOVAPSZ256rm, X86::VMOVAPDZ256rm, X86::VMOVDQA64Z256rm, X86::VMOVDQA32Z256rm  },
7651   { X86::VMOVAPSZ256rr, X86::VMOVAPDZ256rr, X86::VMOVDQA64Z256rr, X86::VMOVDQA32Z256rr  },
7652   { X86::VMOVUPSZ256mr, X86::VMOVUPDZ256mr, X86::VMOVDQU64Z256mr, X86::VMOVDQU32Z256mr  },
7653   { X86::VMOVUPSZ256rm, X86::VMOVUPDZ256rm, X86::VMOVDQU64Z256rm, X86::VMOVDQU32Z256rm  },
7654   { X86::VMOVAPSZmr,    X86::VMOVAPDZmr,    X86::VMOVDQA64Zmr,    X86::VMOVDQA32Zmr     },
7655   { X86::VMOVAPSZrm,    X86::VMOVAPDZrm,    X86::VMOVDQA64Zrm,    X86::VMOVDQA32Zrm     },
7656   { X86::VMOVAPSZrr,    X86::VMOVAPDZrr,    X86::VMOVDQA64Zrr,    X86::VMOVDQA32Zrr     },
7657   { X86::VMOVUPSZmr,    X86::VMOVUPDZmr,    X86::VMOVDQU64Zmr,    X86::VMOVDQU32Zmr     },
7658   { X86::VMOVUPSZrm,    X86::VMOVUPDZrm,    X86::VMOVDQU64Zrm,    X86::VMOVDQU32Zrm     },
7659 };
7660 
7661 static const uint16_t ReplaceableInstrsAVX512DQ[][4] = {
7662   // Two integer columns for 64-bit and 32-bit elements.
7663   //PackedSingle        PackedDouble        PackedInt           PackedInt
7664   { X86::VANDNPSZ128rm, X86::VANDNPDZ128rm, X86::VPANDNQZ128rm, X86::VPANDNDZ128rm },
7665   { X86::VANDNPSZ128rr, X86::VANDNPDZ128rr, X86::VPANDNQZ128rr, X86::VPANDNDZ128rr },
7666   { X86::VANDPSZ128rm,  X86::VANDPDZ128rm,  X86::VPANDQZ128rm,  X86::VPANDDZ128rm  },
7667   { X86::VANDPSZ128rr,  X86::VANDPDZ128rr,  X86::VPANDQZ128rr,  X86::VPANDDZ128rr  },
7668   { X86::VORPSZ128rm,   X86::VORPDZ128rm,   X86::VPORQZ128rm,   X86::VPORDZ128rm   },
7669   { X86::VORPSZ128rr,   X86::VORPDZ128rr,   X86::VPORQZ128rr,   X86::VPORDZ128rr   },
7670   { X86::VXORPSZ128rm,  X86::VXORPDZ128rm,  X86::VPXORQZ128rm,  X86::VPXORDZ128rm  },
7671   { X86::VXORPSZ128rr,  X86::VXORPDZ128rr,  X86::VPXORQZ128rr,  X86::VPXORDZ128rr  },
7672   { X86::VANDNPSZ256rm, X86::VANDNPDZ256rm, X86::VPANDNQZ256rm, X86::VPANDNDZ256rm },
7673   { X86::VANDNPSZ256rr, X86::VANDNPDZ256rr, X86::VPANDNQZ256rr, X86::VPANDNDZ256rr },
7674   { X86::VANDPSZ256rm,  X86::VANDPDZ256rm,  X86::VPANDQZ256rm,  X86::VPANDDZ256rm  },
7675   { X86::VANDPSZ256rr,  X86::VANDPDZ256rr,  X86::VPANDQZ256rr,  X86::VPANDDZ256rr  },
7676   { X86::VORPSZ256rm,   X86::VORPDZ256rm,   X86::VPORQZ256rm,   X86::VPORDZ256rm   },
7677   { X86::VORPSZ256rr,   X86::VORPDZ256rr,   X86::VPORQZ256rr,   X86::VPORDZ256rr   },
7678   { X86::VXORPSZ256rm,  X86::VXORPDZ256rm,  X86::VPXORQZ256rm,  X86::VPXORDZ256rm  },
7679   { X86::VXORPSZ256rr,  X86::VXORPDZ256rr,  X86::VPXORQZ256rr,  X86::VPXORDZ256rr  },
7680   { X86::VANDNPSZrm,    X86::VANDNPDZrm,    X86::VPANDNQZrm,    X86::VPANDNDZrm    },
7681   { X86::VANDNPSZrr,    X86::VANDNPDZrr,    X86::VPANDNQZrr,    X86::VPANDNDZrr    },
7682   { X86::VANDPSZrm,     X86::VANDPDZrm,     X86::VPANDQZrm,     X86::VPANDDZrm     },
7683   { X86::VANDPSZrr,     X86::VANDPDZrr,     X86::VPANDQZrr,     X86::VPANDDZrr     },
7684   { X86::VORPSZrm,      X86::VORPDZrm,      X86::VPORQZrm,      X86::VPORDZrm      },
7685   { X86::VORPSZrr,      X86::VORPDZrr,      X86::VPORQZrr,      X86::VPORDZrr      },
7686   { X86::VXORPSZrm,     X86::VXORPDZrm,     X86::VPXORQZrm,     X86::VPXORDZrm     },
7687   { X86::VXORPSZrr,     X86::VXORPDZrr,     X86::VPXORQZrr,     X86::VPXORDZrr     },
7688 };
7689 
7690 static const uint16_t ReplaceableInstrsAVX512DQMasked[][4] = {
7691   // Two integer columns for 64-bit and 32-bit elements.
7692   //PackedSingle          PackedDouble
7693   //PackedInt             PackedInt
7694   { X86::VANDNPSZ128rmk,  X86::VANDNPDZ128rmk,
7695     X86::VPANDNQZ128rmk,  X86::VPANDNDZ128rmk  },
7696   { X86::VANDNPSZ128rmkz, X86::VANDNPDZ128rmkz,
7697     X86::VPANDNQZ128rmkz, X86::VPANDNDZ128rmkz },
7698   { X86::VANDNPSZ128rrk,  X86::VANDNPDZ128rrk,
7699     X86::VPANDNQZ128rrk,  X86::VPANDNDZ128rrk  },
7700   { X86::VANDNPSZ128rrkz, X86::VANDNPDZ128rrkz,
7701     X86::VPANDNQZ128rrkz, X86::VPANDNDZ128rrkz },
7702   { X86::VANDPSZ128rmk,   X86::VANDPDZ128rmk,
7703     X86::VPANDQZ128rmk,   X86::VPANDDZ128rmk   },
7704   { X86::VANDPSZ128rmkz,  X86::VANDPDZ128rmkz,
7705     X86::VPANDQZ128rmkz,  X86::VPANDDZ128rmkz  },
7706   { X86::VANDPSZ128rrk,   X86::VANDPDZ128rrk,
7707     X86::VPANDQZ128rrk,   X86::VPANDDZ128rrk   },
7708   { X86::VANDPSZ128rrkz,  X86::VANDPDZ128rrkz,
7709     X86::VPANDQZ128rrkz,  X86::VPANDDZ128rrkz  },
7710   { X86::VORPSZ128rmk,    X86::VORPDZ128rmk,
7711     X86::VPORQZ128rmk,    X86::VPORDZ128rmk    },
7712   { X86::VORPSZ128rmkz,   X86::VORPDZ128rmkz,
7713     X86::VPORQZ128rmkz,   X86::VPORDZ128rmkz   },
7714   { X86::VORPSZ128rrk,    X86::VORPDZ128rrk,
7715     X86::VPORQZ128rrk,    X86::VPORDZ128rrk    },
7716   { X86::VORPSZ128rrkz,   X86::VORPDZ128rrkz,
7717     X86::VPORQZ128rrkz,   X86::VPORDZ128rrkz   },
7718   { X86::VXORPSZ128rmk,   X86::VXORPDZ128rmk,
7719     X86::VPXORQZ128rmk,   X86::VPXORDZ128rmk   },
7720   { X86::VXORPSZ128rmkz,  X86::VXORPDZ128rmkz,
7721     X86::VPXORQZ128rmkz,  X86::VPXORDZ128rmkz  },
7722   { X86::VXORPSZ128rrk,   X86::VXORPDZ128rrk,
7723     X86::VPXORQZ128rrk,   X86::VPXORDZ128rrk   },
7724   { X86::VXORPSZ128rrkz,  X86::VXORPDZ128rrkz,
7725     X86::VPXORQZ128rrkz,  X86::VPXORDZ128rrkz  },
7726   { X86::VANDNPSZ256rmk,  X86::VANDNPDZ256rmk,
7727     X86::VPANDNQZ256rmk,  X86::VPANDNDZ256rmk  },
7728   { X86::VANDNPSZ256rmkz, X86::VANDNPDZ256rmkz,
7729     X86::VPANDNQZ256rmkz, X86::VPANDNDZ256rmkz },
7730   { X86::VANDNPSZ256rrk,  X86::VANDNPDZ256rrk,
7731     X86::VPANDNQZ256rrk,  X86::VPANDNDZ256rrk  },
7732   { X86::VANDNPSZ256rrkz, X86::VANDNPDZ256rrkz,
7733     X86::VPANDNQZ256rrkz, X86::VPANDNDZ256rrkz },
7734   { X86::VANDPSZ256rmk,   X86::VANDPDZ256rmk,
7735     X86::VPANDQZ256rmk,   X86::VPANDDZ256rmk   },
7736   { X86::VANDPSZ256rmkz,  X86::VANDPDZ256rmkz,
7737     X86::VPANDQZ256rmkz,  X86::VPANDDZ256rmkz  },
7738   { X86::VANDPSZ256rrk,   X86::VANDPDZ256rrk,
7739     X86::VPANDQZ256rrk,   X86::VPANDDZ256rrk   },
7740   { X86::VANDPSZ256rrkz,  X86::VANDPDZ256rrkz,
7741     X86::VPANDQZ256rrkz,  X86::VPANDDZ256rrkz  },
7742   { X86::VORPSZ256rmk,    X86::VORPDZ256rmk,
7743     X86::VPORQZ256rmk,    X86::VPORDZ256rmk    },
7744   { X86::VORPSZ256rmkz,   X86::VORPDZ256rmkz,
7745     X86::VPORQZ256rmkz,   X86::VPORDZ256rmkz   },
7746   { X86::VORPSZ256rrk,    X86::VORPDZ256rrk,
7747     X86::VPORQZ256rrk,    X86::VPORDZ256rrk    },
7748   { X86::VORPSZ256rrkz,   X86::VORPDZ256rrkz,
7749     X86::VPORQZ256rrkz,   X86::VPORDZ256rrkz   },
7750   { X86::VXORPSZ256rmk,   X86::VXORPDZ256rmk,
7751     X86::VPXORQZ256rmk,   X86::VPXORDZ256rmk   },
7752   { X86::VXORPSZ256rmkz,  X86::VXORPDZ256rmkz,
7753     X86::VPXORQZ256rmkz,  X86::VPXORDZ256rmkz  },
7754   { X86::VXORPSZ256rrk,   X86::VXORPDZ256rrk,
7755     X86::VPXORQZ256rrk,   X86::VPXORDZ256rrk   },
7756   { X86::VXORPSZ256rrkz,  X86::VXORPDZ256rrkz,
7757     X86::VPXORQZ256rrkz,  X86::VPXORDZ256rrkz  },
7758   { X86::VANDNPSZrmk,     X86::VANDNPDZrmk,
7759     X86::VPANDNQZrmk,     X86::VPANDNDZrmk     },
7760   { X86::VANDNPSZrmkz,    X86::VANDNPDZrmkz,
7761     X86::VPANDNQZrmkz,    X86::VPANDNDZrmkz    },
7762   { X86::VANDNPSZrrk,     X86::VANDNPDZrrk,
7763     X86::VPANDNQZrrk,     X86::VPANDNDZrrk     },
7764   { X86::VANDNPSZrrkz,    X86::VANDNPDZrrkz,
7765     X86::VPANDNQZrrkz,    X86::VPANDNDZrrkz    },
7766   { X86::VANDPSZrmk,      X86::VANDPDZrmk,
7767     X86::VPANDQZrmk,      X86::VPANDDZrmk      },
7768   { X86::VANDPSZrmkz,     X86::VANDPDZrmkz,
7769     X86::VPANDQZrmkz,     X86::VPANDDZrmkz     },
7770   { X86::VANDPSZrrk,      X86::VANDPDZrrk,
7771     X86::VPANDQZrrk,      X86::VPANDDZrrk      },
7772   { X86::VANDPSZrrkz,     X86::VANDPDZrrkz,
7773     X86::VPANDQZrrkz,     X86::VPANDDZrrkz     },
7774   { X86::VORPSZrmk,       X86::VORPDZrmk,
7775     X86::VPORQZrmk,       X86::VPORDZrmk       },
7776   { X86::VORPSZrmkz,      X86::VORPDZrmkz,
7777     X86::VPORQZrmkz,      X86::VPORDZrmkz      },
7778   { X86::VORPSZrrk,       X86::VORPDZrrk,
7779     X86::VPORQZrrk,       X86::VPORDZrrk       },
7780   { X86::VORPSZrrkz,      X86::VORPDZrrkz,
7781     X86::VPORQZrrkz,      X86::VPORDZrrkz      },
7782   { X86::VXORPSZrmk,      X86::VXORPDZrmk,
7783     X86::VPXORQZrmk,      X86::VPXORDZrmk      },
7784   { X86::VXORPSZrmkz,     X86::VXORPDZrmkz,
7785     X86::VPXORQZrmkz,     X86::VPXORDZrmkz     },
7786   { X86::VXORPSZrrk,      X86::VXORPDZrrk,
7787     X86::VPXORQZrrk,      X86::VPXORDZrrk      },
7788   { X86::VXORPSZrrkz,     X86::VXORPDZrrkz,
7789     X86::VPXORQZrrkz,     X86::VPXORDZrrkz     },
7790   // Broadcast loads can be handled the same as masked operations to avoid
7791   // changing element size.
7792   { X86::VANDNPSZ128rmb,  X86::VANDNPDZ128rmb,
7793     X86::VPANDNQZ128rmb,  X86::VPANDNDZ128rmb  },
7794   { X86::VANDPSZ128rmb,   X86::VANDPDZ128rmb,
7795     X86::VPANDQZ128rmb,   X86::VPANDDZ128rmb   },
7796   { X86::VORPSZ128rmb,    X86::VORPDZ128rmb,
7797     X86::VPORQZ128rmb,    X86::VPORDZ128rmb    },
7798   { X86::VXORPSZ128rmb,   X86::VXORPDZ128rmb,
7799     X86::VPXORQZ128rmb,   X86::VPXORDZ128rmb   },
7800   { X86::VANDNPSZ256rmb,  X86::VANDNPDZ256rmb,
7801     X86::VPANDNQZ256rmb,  X86::VPANDNDZ256rmb  },
7802   { X86::VANDPSZ256rmb,   X86::VANDPDZ256rmb,
7803     X86::VPANDQZ256rmb,   X86::VPANDDZ256rmb   },
7804   { X86::VORPSZ256rmb,    X86::VORPDZ256rmb,
7805     X86::VPORQZ256rmb,    X86::VPORDZ256rmb    },
7806   { X86::VXORPSZ256rmb,   X86::VXORPDZ256rmb,
7807     X86::VPXORQZ256rmb,   X86::VPXORDZ256rmb   },
7808   { X86::VANDNPSZrmb,     X86::VANDNPDZrmb,
7809     X86::VPANDNQZrmb,     X86::VPANDNDZrmb     },
7810   { X86::VANDPSZrmb,      X86::VANDPDZrmb,
7811     X86::VPANDQZrmb,      X86::VPANDDZrmb      },
7812   { X86::VANDPSZrmb,      X86::VANDPDZrmb,
7813     X86::VPANDQZrmb,      X86::VPANDDZrmb      },
7814   { X86::VORPSZrmb,       X86::VORPDZrmb,
7815     X86::VPORQZrmb,       X86::VPORDZrmb       },
7816   { X86::VXORPSZrmb,      X86::VXORPDZrmb,
7817     X86::VPXORQZrmb,      X86::VPXORDZrmb      },
7818   { X86::VANDNPSZ128rmbk, X86::VANDNPDZ128rmbk,
7819     X86::VPANDNQZ128rmbk, X86::VPANDNDZ128rmbk },
7820   { X86::VANDPSZ128rmbk,  X86::VANDPDZ128rmbk,
7821     X86::VPANDQZ128rmbk,  X86::VPANDDZ128rmbk  },
7822   { X86::VORPSZ128rmbk,   X86::VORPDZ128rmbk,
7823     X86::VPORQZ128rmbk,   X86::VPORDZ128rmbk   },
7824   { X86::VXORPSZ128rmbk,  X86::VXORPDZ128rmbk,
7825     X86::VPXORQZ128rmbk,  X86::VPXORDZ128rmbk  },
7826   { X86::VANDNPSZ256rmbk, X86::VANDNPDZ256rmbk,
7827     X86::VPANDNQZ256rmbk, X86::VPANDNDZ256rmbk },
7828   { X86::VANDPSZ256rmbk,  X86::VANDPDZ256rmbk,
7829     X86::VPANDQZ256rmbk,  X86::VPANDDZ256rmbk  },
7830   { X86::VORPSZ256rmbk,   X86::VORPDZ256rmbk,
7831     X86::VPORQZ256rmbk,   X86::VPORDZ256rmbk   },
7832   { X86::VXORPSZ256rmbk,  X86::VXORPDZ256rmbk,
7833     X86::VPXORQZ256rmbk,  X86::VPXORDZ256rmbk  },
7834   { X86::VANDNPSZrmbk,    X86::VANDNPDZrmbk,
7835     X86::VPANDNQZrmbk,    X86::VPANDNDZrmbk    },
7836   { X86::VANDPSZrmbk,     X86::VANDPDZrmbk,
7837     X86::VPANDQZrmbk,     X86::VPANDDZrmbk     },
7838   { X86::VANDPSZrmbk,     X86::VANDPDZrmbk,
7839     X86::VPANDQZrmbk,     X86::VPANDDZrmbk     },
7840   { X86::VORPSZrmbk,      X86::VORPDZrmbk,
7841     X86::VPORQZrmbk,      X86::VPORDZrmbk      },
7842   { X86::VXORPSZrmbk,     X86::VXORPDZrmbk,
7843     X86::VPXORQZrmbk,     X86::VPXORDZrmbk     },
7844   { X86::VANDNPSZ128rmbkz,X86::VANDNPDZ128rmbkz,
7845     X86::VPANDNQZ128rmbkz,X86::VPANDNDZ128rmbkz},
7846   { X86::VANDPSZ128rmbkz, X86::VANDPDZ128rmbkz,
7847     X86::VPANDQZ128rmbkz, X86::VPANDDZ128rmbkz },
7848   { X86::VORPSZ128rmbkz,  X86::VORPDZ128rmbkz,
7849     X86::VPORQZ128rmbkz,  X86::VPORDZ128rmbkz  },
7850   { X86::VXORPSZ128rmbkz, X86::VXORPDZ128rmbkz,
7851     X86::VPXORQZ128rmbkz, X86::VPXORDZ128rmbkz },
7852   { X86::VANDNPSZ256rmbkz,X86::VANDNPDZ256rmbkz,
7853     X86::VPANDNQZ256rmbkz,X86::VPANDNDZ256rmbkz},
7854   { X86::VANDPSZ256rmbkz, X86::VANDPDZ256rmbkz,
7855     X86::VPANDQZ256rmbkz, X86::VPANDDZ256rmbkz },
7856   { X86::VORPSZ256rmbkz,  X86::VORPDZ256rmbkz,
7857     X86::VPORQZ256rmbkz,  X86::VPORDZ256rmbkz  },
7858   { X86::VXORPSZ256rmbkz, X86::VXORPDZ256rmbkz,
7859     X86::VPXORQZ256rmbkz, X86::VPXORDZ256rmbkz },
7860   { X86::VANDNPSZrmbkz,   X86::VANDNPDZrmbkz,
7861     X86::VPANDNQZrmbkz,   X86::VPANDNDZrmbkz   },
7862   { X86::VANDPSZrmbkz,    X86::VANDPDZrmbkz,
7863     X86::VPANDQZrmbkz,    X86::VPANDDZrmbkz    },
7864   { X86::VANDPSZrmbkz,    X86::VANDPDZrmbkz,
7865     X86::VPANDQZrmbkz,    X86::VPANDDZrmbkz    },
7866   { X86::VORPSZrmbkz,     X86::VORPDZrmbkz,
7867     X86::VPORQZrmbkz,     X86::VPORDZrmbkz     },
7868   { X86::VXORPSZrmbkz,    X86::VXORPDZrmbkz,
7869     X86::VPXORQZrmbkz,    X86::VPXORDZrmbkz    },
7870 };
7871 
7872 // NOTE: These should only be used by the custom domain methods.
7873 static const uint16_t ReplaceableBlendInstrs[][3] = {
7874   //PackedSingle             PackedDouble             PackedInt
7875   { X86::BLENDPSrmi,         X86::BLENDPDrmi,         X86::PBLENDWrmi   },
7876   { X86::BLENDPSrri,         X86::BLENDPDrri,         X86::PBLENDWrri   },
7877   { X86::VBLENDPSrmi,        X86::VBLENDPDrmi,        X86::VPBLENDWrmi  },
7878   { X86::VBLENDPSrri,        X86::VBLENDPDrri,        X86::VPBLENDWrri  },
7879   { X86::VBLENDPSYrmi,       X86::VBLENDPDYrmi,       X86::VPBLENDWYrmi },
7880   { X86::VBLENDPSYrri,       X86::VBLENDPDYrri,       X86::VPBLENDWYrri },
7881 };
7882 static const uint16_t ReplaceableBlendAVX2Instrs[][3] = {
7883   //PackedSingle             PackedDouble             PackedInt
7884   { X86::VBLENDPSrmi,        X86::VBLENDPDrmi,        X86::VPBLENDDrmi  },
7885   { X86::VBLENDPSrri,        X86::VBLENDPDrri,        X86::VPBLENDDrri  },
7886   { X86::VBLENDPSYrmi,       X86::VBLENDPDYrmi,       X86::VPBLENDDYrmi },
7887   { X86::VBLENDPSYrri,       X86::VBLENDPDYrri,       X86::VPBLENDDYrri },
7888 };
7889 
7890 // Special table for changing EVEX logic instructions to VEX.
7891 // TODO: Should we run EVEX->VEX earlier?
7892 static const uint16_t ReplaceableCustomAVX512LogicInstrs[][4] = {
7893   // Two integer columns for 64-bit and 32-bit elements.
7894   //PackedSingle     PackedDouble     PackedInt           PackedInt
7895   { X86::VANDNPSrm,  X86::VANDNPDrm,  X86::VPANDNQZ128rm, X86::VPANDNDZ128rm },
7896   { X86::VANDNPSrr,  X86::VANDNPDrr,  X86::VPANDNQZ128rr, X86::VPANDNDZ128rr },
7897   { X86::VANDPSrm,   X86::VANDPDrm,   X86::VPANDQZ128rm,  X86::VPANDDZ128rm  },
7898   { X86::VANDPSrr,   X86::VANDPDrr,   X86::VPANDQZ128rr,  X86::VPANDDZ128rr  },
7899   { X86::VORPSrm,    X86::VORPDrm,    X86::VPORQZ128rm,   X86::VPORDZ128rm   },
7900   { X86::VORPSrr,    X86::VORPDrr,    X86::VPORQZ128rr,   X86::VPORDZ128rr   },
7901   { X86::VXORPSrm,   X86::VXORPDrm,   X86::VPXORQZ128rm,  X86::VPXORDZ128rm  },
7902   { X86::VXORPSrr,   X86::VXORPDrr,   X86::VPXORQZ128rr,  X86::VPXORDZ128rr  },
7903   { X86::VANDNPSYrm, X86::VANDNPDYrm, X86::VPANDNQZ256rm, X86::VPANDNDZ256rm },
7904   { X86::VANDNPSYrr, X86::VANDNPDYrr, X86::VPANDNQZ256rr, X86::VPANDNDZ256rr },
7905   { X86::VANDPSYrm,  X86::VANDPDYrm,  X86::VPANDQZ256rm,  X86::VPANDDZ256rm  },
7906   { X86::VANDPSYrr,  X86::VANDPDYrr,  X86::VPANDQZ256rr,  X86::VPANDDZ256rr  },
7907   { X86::VORPSYrm,   X86::VORPDYrm,   X86::VPORQZ256rm,   X86::VPORDZ256rm   },
7908   { X86::VORPSYrr,   X86::VORPDYrr,   X86::VPORQZ256rr,   X86::VPORDZ256rr   },
7909   { X86::VXORPSYrm,  X86::VXORPDYrm,  X86::VPXORQZ256rm,  X86::VPXORDZ256rm  },
7910   { X86::VXORPSYrr,  X86::VXORPDYrr,  X86::VPXORQZ256rr,  X86::VPXORDZ256rr  },
7911 };
7912 
7913 // FIXME: Some shuffle and unpack instructions have equivalents in different
7914 // domains, but they require a bit more work than just switching opcodes.
7915 
7916 static const uint16_t *lookup(unsigned opcode, unsigned domain,
7917                               ArrayRef<uint16_t[3]> Table) {
7918   for (const uint16_t (&Row)[3] : Table)
7919     if (Row[domain-1] == opcode)
7920       return Row;
7921   return nullptr;
7922 }
7923 
7924 static const uint16_t *lookupAVX512(unsigned opcode, unsigned domain,
7925                                     ArrayRef<uint16_t[4]> Table) {
7926   // If this is the integer domain make sure to check both integer columns.
7927   for (const uint16_t (&Row)[4] : Table)
7928     if (Row[domain-1] == opcode || (domain == 3 && Row[3] == opcode))
7929       return Row;
7930   return nullptr;
7931 }
7932 
7933 // Helper to attempt to widen/narrow blend masks.
7934 static bool AdjustBlendMask(unsigned OldMask, unsigned OldWidth,
7935                             unsigned NewWidth, unsigned *pNewMask = nullptr) {
7936   assert(((OldWidth % NewWidth) == 0 || (NewWidth % OldWidth) == 0) &&
7937          "Illegal blend mask scale");
7938   unsigned NewMask = 0;
7939 
7940   if ((OldWidth % NewWidth) == 0) {
7941     unsigned Scale = OldWidth / NewWidth;
7942     unsigned SubMask = (1u << Scale) - 1;
7943     for (unsigned i = 0; i != NewWidth; ++i) {
7944       unsigned Sub = (OldMask >> (i * Scale)) & SubMask;
7945       if (Sub == SubMask)
7946         NewMask |= (1u << i);
7947       else if (Sub != 0x0)
7948         return false;
7949     }
7950   } else {
7951     unsigned Scale = NewWidth / OldWidth;
7952     unsigned SubMask = (1u << Scale) - 1;
7953     for (unsigned i = 0; i != OldWidth; ++i) {
7954       if (OldMask & (1 << i)) {
7955         NewMask |= (SubMask << (i * Scale));
7956       }
7957     }
7958   }
7959 
7960   if (pNewMask)
7961     *pNewMask = NewMask;
7962   return true;
7963 }
7964 
7965 uint16_t X86InstrInfo::getExecutionDomainCustom(const MachineInstr &MI) const {
7966   unsigned Opcode = MI.getOpcode();
7967   unsigned NumOperands = MI.getDesc().getNumOperands();
7968 
7969   auto GetBlendDomains = [&](unsigned ImmWidth, bool Is256) {
7970     uint16_t validDomains = 0;
7971     if (MI.getOperand(NumOperands - 1).isImm()) {
7972       unsigned Imm = MI.getOperand(NumOperands - 1).getImm();
7973       if (AdjustBlendMask(Imm, ImmWidth, Is256 ? 8 : 4))
7974         validDomains |= 0x2; // PackedSingle
7975       if (AdjustBlendMask(Imm, ImmWidth, Is256 ? 4 : 2))
7976         validDomains |= 0x4; // PackedDouble
7977       if (!Is256 || Subtarget.hasAVX2())
7978         validDomains |= 0x8; // PackedInt
7979     }
7980     return validDomains;
7981   };
7982 
7983   switch (Opcode) {
7984   case X86::BLENDPDrmi:
7985   case X86::BLENDPDrri:
7986   case X86::VBLENDPDrmi:
7987   case X86::VBLENDPDrri:
7988     return GetBlendDomains(2, false);
7989   case X86::VBLENDPDYrmi:
7990   case X86::VBLENDPDYrri:
7991     return GetBlendDomains(4, true);
7992   case X86::BLENDPSrmi:
7993   case X86::BLENDPSrri:
7994   case X86::VBLENDPSrmi:
7995   case X86::VBLENDPSrri:
7996   case X86::VPBLENDDrmi:
7997   case X86::VPBLENDDrri:
7998     return GetBlendDomains(4, false);
7999   case X86::VBLENDPSYrmi:
8000   case X86::VBLENDPSYrri:
8001   case X86::VPBLENDDYrmi:
8002   case X86::VPBLENDDYrri:
8003     return GetBlendDomains(8, true);
8004   case X86::PBLENDWrmi:
8005   case X86::PBLENDWrri:
8006   case X86::VPBLENDWrmi:
8007   case X86::VPBLENDWrri:
8008   // Treat VPBLENDWY as a 128-bit vector as it repeats the lo/hi masks.
8009   case X86::VPBLENDWYrmi:
8010   case X86::VPBLENDWYrri:
8011     return GetBlendDomains(8, false);
8012   case X86::VPANDDZ128rr:  case X86::VPANDDZ128rm:
8013   case X86::VPANDDZ256rr:  case X86::VPANDDZ256rm:
8014   case X86::VPANDQZ128rr:  case X86::VPANDQZ128rm:
8015   case X86::VPANDQZ256rr:  case X86::VPANDQZ256rm:
8016   case X86::VPANDNDZ128rr: case X86::VPANDNDZ128rm:
8017   case X86::VPANDNDZ256rr: case X86::VPANDNDZ256rm:
8018   case X86::VPANDNQZ128rr: case X86::VPANDNQZ128rm:
8019   case X86::VPANDNQZ256rr: case X86::VPANDNQZ256rm:
8020   case X86::VPORDZ128rr:   case X86::VPORDZ128rm:
8021   case X86::VPORDZ256rr:   case X86::VPORDZ256rm:
8022   case X86::VPORQZ128rr:   case X86::VPORQZ128rm:
8023   case X86::VPORQZ256rr:   case X86::VPORQZ256rm:
8024   case X86::VPXORDZ128rr:  case X86::VPXORDZ128rm:
8025   case X86::VPXORDZ256rr:  case X86::VPXORDZ256rm:
8026   case X86::VPXORQZ128rr:  case X86::VPXORQZ128rm:
8027   case X86::VPXORQZ256rr:  case X86::VPXORQZ256rm:
8028     // If we don't have DQI see if we can still switch from an EVEX integer
8029     // instruction to a VEX floating point instruction.
8030     if (Subtarget.hasDQI())
8031       return 0;
8032 
8033     if (RI.getEncodingValue(MI.getOperand(0).getReg()) >= 16)
8034       return 0;
8035     if (RI.getEncodingValue(MI.getOperand(1).getReg()) >= 16)
8036       return 0;
8037     // Register forms will have 3 operands. Memory form will have more.
8038     if (NumOperands == 3 &&
8039         RI.getEncodingValue(MI.getOperand(2).getReg()) >= 16)
8040       return 0;
8041 
8042     // All domains are valid.
8043     return 0xe;
8044   case X86::MOVHLPSrr:
8045     // We can swap domains when both inputs are the same register.
8046     // FIXME: This doesn't catch all the cases we would like. If the input
8047     // register isn't KILLed by the instruction, the two address instruction
8048     // pass puts a COPY on one input. The other input uses the original
8049     // register. This prevents the same physical register from being used by
8050     // both inputs.
8051     if (MI.getOperand(1).getReg() == MI.getOperand(2).getReg() &&
8052         MI.getOperand(0).getSubReg() == 0 &&
8053         MI.getOperand(1).getSubReg() == 0 &&
8054         MI.getOperand(2).getSubReg() == 0)
8055       return 0x6;
8056     return 0;
8057   case X86::SHUFPDrri:
8058     return 0x6;
8059   }
8060   return 0;
8061 }
8062 
8063 bool X86InstrInfo::setExecutionDomainCustom(MachineInstr &MI,
8064                                             unsigned Domain) const {
8065   assert(Domain > 0 && Domain < 4 && "Invalid execution domain");
8066   uint16_t dom = (MI.getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
8067   assert(dom && "Not an SSE instruction");
8068 
8069   unsigned Opcode = MI.getOpcode();
8070   unsigned NumOperands = MI.getDesc().getNumOperands();
8071 
8072   auto SetBlendDomain = [&](unsigned ImmWidth, bool Is256) {
8073     if (MI.getOperand(NumOperands - 1).isImm()) {
8074       unsigned Imm = MI.getOperand(NumOperands - 1).getImm() & 255;
8075       Imm = (ImmWidth == 16 ? ((Imm << 8) | Imm) : Imm);
8076       unsigned NewImm = Imm;
8077 
8078       const uint16_t *table = lookup(Opcode, dom, ReplaceableBlendInstrs);
8079       if (!table)
8080         table = lookup(Opcode, dom, ReplaceableBlendAVX2Instrs);
8081 
8082       if (Domain == 1) { // PackedSingle
8083         AdjustBlendMask(Imm, ImmWidth, Is256 ? 8 : 4, &NewImm);
8084       } else if (Domain == 2) { // PackedDouble
8085         AdjustBlendMask(Imm, ImmWidth, Is256 ? 4 : 2, &NewImm);
8086       } else if (Domain == 3) { // PackedInt
8087         if (Subtarget.hasAVX2()) {
8088           // If we are already VPBLENDW use that, else use VPBLENDD.
8089           if ((ImmWidth / (Is256 ? 2 : 1)) != 8) {
8090             table = lookup(Opcode, dom, ReplaceableBlendAVX2Instrs);
8091             AdjustBlendMask(Imm, ImmWidth, Is256 ? 8 : 4, &NewImm);
8092           }
8093         } else {
8094           assert(!Is256 && "128-bit vector expected");
8095           AdjustBlendMask(Imm, ImmWidth, 8, &NewImm);
8096         }
8097       }
8098 
8099       assert(table && table[Domain - 1] && "Unknown domain op");
8100       MI.setDesc(get(table[Domain - 1]));
8101       MI.getOperand(NumOperands - 1).setImm(NewImm & 255);
8102     }
8103     return true;
8104   };
8105 
8106   switch (Opcode) {
8107   case X86::BLENDPDrmi:
8108   case X86::BLENDPDrri:
8109   case X86::VBLENDPDrmi:
8110   case X86::VBLENDPDrri:
8111     return SetBlendDomain(2, false);
8112   case X86::VBLENDPDYrmi:
8113   case X86::VBLENDPDYrri:
8114     return SetBlendDomain(4, true);
8115   case X86::BLENDPSrmi:
8116   case X86::BLENDPSrri:
8117   case X86::VBLENDPSrmi:
8118   case X86::VBLENDPSrri:
8119   case X86::VPBLENDDrmi:
8120   case X86::VPBLENDDrri:
8121     return SetBlendDomain(4, false);
8122   case X86::VBLENDPSYrmi:
8123   case X86::VBLENDPSYrri:
8124   case X86::VPBLENDDYrmi:
8125   case X86::VPBLENDDYrri:
8126     return SetBlendDomain(8, true);
8127   case X86::PBLENDWrmi:
8128   case X86::PBLENDWrri:
8129   case X86::VPBLENDWrmi:
8130   case X86::VPBLENDWrri:
8131     return SetBlendDomain(8, false);
8132   case X86::VPBLENDWYrmi:
8133   case X86::VPBLENDWYrri:
8134     return SetBlendDomain(16, true);
8135   case X86::VPANDDZ128rr:  case X86::VPANDDZ128rm:
8136   case X86::VPANDDZ256rr:  case X86::VPANDDZ256rm:
8137   case X86::VPANDQZ128rr:  case X86::VPANDQZ128rm:
8138   case X86::VPANDQZ256rr:  case X86::VPANDQZ256rm:
8139   case X86::VPANDNDZ128rr: case X86::VPANDNDZ128rm:
8140   case X86::VPANDNDZ256rr: case X86::VPANDNDZ256rm:
8141   case X86::VPANDNQZ128rr: case X86::VPANDNQZ128rm:
8142   case X86::VPANDNQZ256rr: case X86::VPANDNQZ256rm:
8143   case X86::VPORDZ128rr:   case X86::VPORDZ128rm:
8144   case X86::VPORDZ256rr:   case X86::VPORDZ256rm:
8145   case X86::VPORQZ128rr:   case X86::VPORQZ128rm:
8146   case X86::VPORQZ256rr:   case X86::VPORQZ256rm:
8147   case X86::VPXORDZ128rr:  case X86::VPXORDZ128rm:
8148   case X86::VPXORDZ256rr:  case X86::VPXORDZ256rm:
8149   case X86::VPXORQZ128rr:  case X86::VPXORQZ128rm:
8150   case X86::VPXORQZ256rr:  case X86::VPXORQZ256rm: {
8151     // Without DQI, convert EVEX instructions to VEX instructions.
8152     if (Subtarget.hasDQI())
8153       return false;
8154 
8155     const uint16_t *table = lookupAVX512(MI.getOpcode(), dom,
8156                                          ReplaceableCustomAVX512LogicInstrs);
8157     assert(table && "Instruction not found in table?");
8158     // Don't change integer Q instructions to D instructions and
8159     // use D intructions if we started with a PS instruction.
8160     if (Domain == 3 && (dom == 1 || table[3] == MI.getOpcode()))
8161       Domain = 4;
8162     MI.setDesc(get(table[Domain - 1]));
8163     return true;
8164   }
8165   case X86::UNPCKHPDrr:
8166   case X86::MOVHLPSrr:
8167     // We just need to commute the instruction which will switch the domains.
8168     if (Domain != dom && Domain != 3 &&
8169         MI.getOperand(1).getReg() == MI.getOperand(2).getReg() &&
8170         MI.getOperand(0).getSubReg() == 0 &&
8171         MI.getOperand(1).getSubReg() == 0 &&
8172         MI.getOperand(2).getSubReg() == 0) {
8173       commuteInstruction(MI, false);
8174       return true;
8175     }
8176     // We must always return true for MOVHLPSrr.
8177     if (Opcode == X86::MOVHLPSrr)
8178       return true;
8179     break;
8180   case X86::SHUFPDrri: {
8181     if (Domain == 1) {
8182       unsigned Imm = MI.getOperand(3).getImm();
8183       unsigned NewImm = 0x44;
8184       if (Imm & 1) NewImm |= 0x0a;
8185       if (Imm & 2) NewImm |= 0xa0;
8186       MI.getOperand(3).setImm(NewImm);
8187       MI.setDesc(get(X86::SHUFPSrri));
8188     }
8189     return true;
8190   }
8191   }
8192   return false;
8193 }
8194 
8195 std::pair<uint16_t, uint16_t>
8196 X86InstrInfo::getExecutionDomain(const MachineInstr &MI) const {
8197   uint16_t domain = (MI.getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
8198   unsigned opcode = MI.getOpcode();
8199   uint16_t validDomains = 0;
8200   if (domain) {
8201     // Attempt to match for custom instructions.
8202     validDomains = getExecutionDomainCustom(MI);
8203     if (validDomains)
8204       return std::make_pair(domain, validDomains);
8205 
8206     if (lookup(opcode, domain, ReplaceableInstrs)) {
8207       validDomains = 0xe;
8208     } else if (lookup(opcode, domain, ReplaceableInstrsAVX2)) {
8209       validDomains = Subtarget.hasAVX2() ? 0xe : 0x6;
8210     } else if (lookup(opcode, domain, ReplaceableInstrsFP)) {
8211       validDomains = 0x6;
8212     } else if (lookup(opcode, domain, ReplaceableInstrsAVX2InsertExtract)) {
8213       // Insert/extract instructions should only effect domain if AVX2
8214       // is enabled.
8215       if (!Subtarget.hasAVX2())
8216         return std::make_pair(0, 0);
8217       validDomains = 0xe;
8218     } else if (lookupAVX512(opcode, domain, ReplaceableInstrsAVX512)) {
8219       validDomains = 0xe;
8220     } else if (Subtarget.hasDQI() && lookupAVX512(opcode, domain,
8221                                                   ReplaceableInstrsAVX512DQ)) {
8222       validDomains = 0xe;
8223     } else if (Subtarget.hasDQI()) {
8224       if (const uint16_t *table = lookupAVX512(opcode, domain,
8225                                              ReplaceableInstrsAVX512DQMasked)) {
8226         if (domain == 1 || (domain == 3 && table[3] == opcode))
8227           validDomains = 0xa;
8228         else
8229           validDomains = 0xc;
8230       }
8231     }
8232   }
8233   return std::make_pair(domain, validDomains);
8234 }
8235 
8236 void X86InstrInfo::setExecutionDomain(MachineInstr &MI, unsigned Domain) const {
8237   assert(Domain>0 && Domain<4 && "Invalid execution domain");
8238   uint16_t dom = (MI.getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
8239   assert(dom && "Not an SSE instruction");
8240 
8241   // Attempt to match for custom instructions.
8242   if (setExecutionDomainCustom(MI, Domain))
8243     return;
8244 
8245   const uint16_t *table = lookup(MI.getOpcode(), dom, ReplaceableInstrs);
8246   if (!table) { // try the other table
8247     assert((Subtarget.hasAVX2() || Domain < 3) &&
8248            "256-bit vector operations only available in AVX2");
8249     table = lookup(MI.getOpcode(), dom, ReplaceableInstrsAVX2);
8250   }
8251   if (!table) { // try the FP table
8252     table = lookup(MI.getOpcode(), dom, ReplaceableInstrsFP);
8253     assert((!table || Domain < 3) &&
8254            "Can only select PackedSingle or PackedDouble");
8255   }
8256   if (!table) { // try the other table
8257     assert(Subtarget.hasAVX2() &&
8258            "256-bit insert/extract only available in AVX2");
8259     table = lookup(MI.getOpcode(), dom, ReplaceableInstrsAVX2InsertExtract);
8260   }
8261   if (!table) { // try the AVX512 table
8262     assert(Subtarget.hasAVX512() && "Requires AVX-512");
8263     table = lookupAVX512(MI.getOpcode(), dom, ReplaceableInstrsAVX512);
8264     // Don't change integer Q instructions to D instructions.
8265     if (table && Domain == 3 && table[3] == MI.getOpcode())
8266       Domain = 4;
8267   }
8268   if (!table) { // try the AVX512DQ table
8269     assert((Subtarget.hasDQI() || Domain >= 3) && "Requires AVX-512DQ");
8270     table = lookupAVX512(MI.getOpcode(), dom, ReplaceableInstrsAVX512DQ);
8271     // Don't change integer Q instructions to D instructions and
8272     // use D instructions if we started with a PS instruction.
8273     if (table && Domain == 3 && (dom == 1 || table[3] == MI.getOpcode()))
8274       Domain = 4;
8275   }
8276   if (!table) { // try the AVX512DQMasked table
8277     assert((Subtarget.hasDQI() || Domain >= 3) && "Requires AVX-512DQ");
8278     table = lookupAVX512(MI.getOpcode(), dom, ReplaceableInstrsAVX512DQMasked);
8279     if (table && Domain == 3 && (dom == 1 || table[3] == MI.getOpcode()))
8280       Domain = 4;
8281   }
8282   assert(table && "Cannot change domain");
8283   MI.setDesc(get(table[Domain - 1]));
8284 }
8285 
8286 /// Return the noop instruction to use for a noop.
8287 MCInst X86InstrInfo::getNop() const {
8288   MCInst Nop;
8289   Nop.setOpcode(X86::NOOP);
8290   return Nop;
8291 }
8292 
8293 bool X86InstrInfo::isHighLatencyDef(int opc) const {
8294   switch (opc) {
8295   default: return false;
8296   case X86::DIVPDrm:
8297   case X86::DIVPDrr:
8298   case X86::DIVPSrm:
8299   case X86::DIVPSrr:
8300   case X86::DIVSDrm:
8301   case X86::DIVSDrm_Int:
8302   case X86::DIVSDrr:
8303   case X86::DIVSDrr_Int:
8304   case X86::DIVSSrm:
8305   case X86::DIVSSrm_Int:
8306   case X86::DIVSSrr:
8307   case X86::DIVSSrr_Int:
8308   case X86::SQRTPDm:
8309   case X86::SQRTPDr:
8310   case X86::SQRTPSm:
8311   case X86::SQRTPSr:
8312   case X86::SQRTSDm:
8313   case X86::SQRTSDm_Int:
8314   case X86::SQRTSDr:
8315   case X86::SQRTSDr_Int:
8316   case X86::SQRTSSm:
8317   case X86::SQRTSSm_Int:
8318   case X86::SQRTSSr:
8319   case X86::SQRTSSr_Int:
8320   // AVX instructions with high latency
8321   case X86::VDIVPDrm:
8322   case X86::VDIVPDrr:
8323   case X86::VDIVPDYrm:
8324   case X86::VDIVPDYrr:
8325   case X86::VDIVPSrm:
8326   case X86::VDIVPSrr:
8327   case X86::VDIVPSYrm:
8328   case X86::VDIVPSYrr:
8329   case X86::VDIVSDrm:
8330   case X86::VDIVSDrm_Int:
8331   case X86::VDIVSDrr:
8332   case X86::VDIVSDrr_Int:
8333   case X86::VDIVSSrm:
8334   case X86::VDIVSSrm_Int:
8335   case X86::VDIVSSrr:
8336   case X86::VDIVSSrr_Int:
8337   case X86::VSQRTPDm:
8338   case X86::VSQRTPDr:
8339   case X86::VSQRTPDYm:
8340   case X86::VSQRTPDYr:
8341   case X86::VSQRTPSm:
8342   case X86::VSQRTPSr:
8343   case X86::VSQRTPSYm:
8344   case X86::VSQRTPSYr:
8345   case X86::VSQRTSDm:
8346   case X86::VSQRTSDm_Int:
8347   case X86::VSQRTSDr:
8348   case X86::VSQRTSDr_Int:
8349   case X86::VSQRTSSm:
8350   case X86::VSQRTSSm_Int:
8351   case X86::VSQRTSSr:
8352   case X86::VSQRTSSr_Int:
8353   // AVX512 instructions with high latency
8354   case X86::VDIVPDZ128rm:
8355   case X86::VDIVPDZ128rmb:
8356   case X86::VDIVPDZ128rmbk:
8357   case X86::VDIVPDZ128rmbkz:
8358   case X86::VDIVPDZ128rmk:
8359   case X86::VDIVPDZ128rmkz:
8360   case X86::VDIVPDZ128rr:
8361   case X86::VDIVPDZ128rrk:
8362   case X86::VDIVPDZ128rrkz:
8363   case X86::VDIVPDZ256rm:
8364   case X86::VDIVPDZ256rmb:
8365   case X86::VDIVPDZ256rmbk:
8366   case X86::VDIVPDZ256rmbkz:
8367   case X86::VDIVPDZ256rmk:
8368   case X86::VDIVPDZ256rmkz:
8369   case X86::VDIVPDZ256rr:
8370   case X86::VDIVPDZ256rrk:
8371   case X86::VDIVPDZ256rrkz:
8372   case X86::VDIVPDZrrb:
8373   case X86::VDIVPDZrrbk:
8374   case X86::VDIVPDZrrbkz:
8375   case X86::VDIVPDZrm:
8376   case X86::VDIVPDZrmb:
8377   case X86::VDIVPDZrmbk:
8378   case X86::VDIVPDZrmbkz:
8379   case X86::VDIVPDZrmk:
8380   case X86::VDIVPDZrmkz:
8381   case X86::VDIVPDZrr:
8382   case X86::VDIVPDZrrk:
8383   case X86::VDIVPDZrrkz:
8384   case X86::VDIVPSZ128rm:
8385   case X86::VDIVPSZ128rmb:
8386   case X86::VDIVPSZ128rmbk:
8387   case X86::VDIVPSZ128rmbkz:
8388   case X86::VDIVPSZ128rmk:
8389   case X86::VDIVPSZ128rmkz:
8390   case X86::VDIVPSZ128rr:
8391   case X86::VDIVPSZ128rrk:
8392   case X86::VDIVPSZ128rrkz:
8393   case X86::VDIVPSZ256rm:
8394   case X86::VDIVPSZ256rmb:
8395   case X86::VDIVPSZ256rmbk:
8396   case X86::VDIVPSZ256rmbkz:
8397   case X86::VDIVPSZ256rmk:
8398   case X86::VDIVPSZ256rmkz:
8399   case X86::VDIVPSZ256rr:
8400   case X86::VDIVPSZ256rrk:
8401   case X86::VDIVPSZ256rrkz:
8402   case X86::VDIVPSZrrb:
8403   case X86::VDIVPSZrrbk:
8404   case X86::VDIVPSZrrbkz:
8405   case X86::VDIVPSZrm:
8406   case X86::VDIVPSZrmb:
8407   case X86::VDIVPSZrmbk:
8408   case X86::VDIVPSZrmbkz:
8409   case X86::VDIVPSZrmk:
8410   case X86::VDIVPSZrmkz:
8411   case X86::VDIVPSZrr:
8412   case X86::VDIVPSZrrk:
8413   case X86::VDIVPSZrrkz:
8414   case X86::VDIVSDZrm:
8415   case X86::VDIVSDZrr:
8416   case X86::VDIVSDZrm_Int:
8417   case X86::VDIVSDZrm_Intk:
8418   case X86::VDIVSDZrm_Intkz:
8419   case X86::VDIVSDZrr_Int:
8420   case X86::VDIVSDZrr_Intk:
8421   case X86::VDIVSDZrr_Intkz:
8422   case X86::VDIVSDZrrb_Int:
8423   case X86::VDIVSDZrrb_Intk:
8424   case X86::VDIVSDZrrb_Intkz:
8425   case X86::VDIVSSZrm:
8426   case X86::VDIVSSZrr:
8427   case X86::VDIVSSZrm_Int:
8428   case X86::VDIVSSZrm_Intk:
8429   case X86::VDIVSSZrm_Intkz:
8430   case X86::VDIVSSZrr_Int:
8431   case X86::VDIVSSZrr_Intk:
8432   case X86::VDIVSSZrr_Intkz:
8433   case X86::VDIVSSZrrb_Int:
8434   case X86::VDIVSSZrrb_Intk:
8435   case X86::VDIVSSZrrb_Intkz:
8436   case X86::VSQRTPDZ128m:
8437   case X86::VSQRTPDZ128mb:
8438   case X86::VSQRTPDZ128mbk:
8439   case X86::VSQRTPDZ128mbkz:
8440   case X86::VSQRTPDZ128mk:
8441   case X86::VSQRTPDZ128mkz:
8442   case X86::VSQRTPDZ128r:
8443   case X86::VSQRTPDZ128rk:
8444   case X86::VSQRTPDZ128rkz:
8445   case X86::VSQRTPDZ256m:
8446   case X86::VSQRTPDZ256mb:
8447   case X86::VSQRTPDZ256mbk:
8448   case X86::VSQRTPDZ256mbkz:
8449   case X86::VSQRTPDZ256mk:
8450   case X86::VSQRTPDZ256mkz:
8451   case X86::VSQRTPDZ256r:
8452   case X86::VSQRTPDZ256rk:
8453   case X86::VSQRTPDZ256rkz:
8454   case X86::VSQRTPDZm:
8455   case X86::VSQRTPDZmb:
8456   case X86::VSQRTPDZmbk:
8457   case X86::VSQRTPDZmbkz:
8458   case X86::VSQRTPDZmk:
8459   case X86::VSQRTPDZmkz:
8460   case X86::VSQRTPDZr:
8461   case X86::VSQRTPDZrb:
8462   case X86::VSQRTPDZrbk:
8463   case X86::VSQRTPDZrbkz:
8464   case X86::VSQRTPDZrk:
8465   case X86::VSQRTPDZrkz:
8466   case X86::VSQRTPSZ128m:
8467   case X86::VSQRTPSZ128mb:
8468   case X86::VSQRTPSZ128mbk:
8469   case X86::VSQRTPSZ128mbkz:
8470   case X86::VSQRTPSZ128mk:
8471   case X86::VSQRTPSZ128mkz:
8472   case X86::VSQRTPSZ128r:
8473   case X86::VSQRTPSZ128rk:
8474   case X86::VSQRTPSZ128rkz:
8475   case X86::VSQRTPSZ256m:
8476   case X86::VSQRTPSZ256mb:
8477   case X86::VSQRTPSZ256mbk:
8478   case X86::VSQRTPSZ256mbkz:
8479   case X86::VSQRTPSZ256mk:
8480   case X86::VSQRTPSZ256mkz:
8481   case X86::VSQRTPSZ256r:
8482   case X86::VSQRTPSZ256rk:
8483   case X86::VSQRTPSZ256rkz:
8484   case X86::VSQRTPSZm:
8485   case X86::VSQRTPSZmb:
8486   case X86::VSQRTPSZmbk:
8487   case X86::VSQRTPSZmbkz:
8488   case X86::VSQRTPSZmk:
8489   case X86::VSQRTPSZmkz:
8490   case X86::VSQRTPSZr:
8491   case X86::VSQRTPSZrb:
8492   case X86::VSQRTPSZrbk:
8493   case X86::VSQRTPSZrbkz:
8494   case X86::VSQRTPSZrk:
8495   case X86::VSQRTPSZrkz:
8496   case X86::VSQRTSDZm:
8497   case X86::VSQRTSDZm_Int:
8498   case X86::VSQRTSDZm_Intk:
8499   case X86::VSQRTSDZm_Intkz:
8500   case X86::VSQRTSDZr:
8501   case X86::VSQRTSDZr_Int:
8502   case X86::VSQRTSDZr_Intk:
8503   case X86::VSQRTSDZr_Intkz:
8504   case X86::VSQRTSDZrb_Int:
8505   case X86::VSQRTSDZrb_Intk:
8506   case X86::VSQRTSDZrb_Intkz:
8507   case X86::VSQRTSSZm:
8508   case X86::VSQRTSSZm_Int:
8509   case X86::VSQRTSSZm_Intk:
8510   case X86::VSQRTSSZm_Intkz:
8511   case X86::VSQRTSSZr:
8512   case X86::VSQRTSSZr_Int:
8513   case X86::VSQRTSSZr_Intk:
8514   case X86::VSQRTSSZr_Intkz:
8515   case X86::VSQRTSSZrb_Int:
8516   case X86::VSQRTSSZrb_Intk:
8517   case X86::VSQRTSSZrb_Intkz:
8518 
8519   case X86::VGATHERDPDYrm:
8520   case X86::VGATHERDPDZ128rm:
8521   case X86::VGATHERDPDZ256rm:
8522   case X86::VGATHERDPDZrm:
8523   case X86::VGATHERDPDrm:
8524   case X86::VGATHERDPSYrm:
8525   case X86::VGATHERDPSZ128rm:
8526   case X86::VGATHERDPSZ256rm:
8527   case X86::VGATHERDPSZrm:
8528   case X86::VGATHERDPSrm:
8529   case X86::VGATHERPF0DPDm:
8530   case X86::VGATHERPF0DPSm:
8531   case X86::VGATHERPF0QPDm:
8532   case X86::VGATHERPF0QPSm:
8533   case X86::VGATHERPF1DPDm:
8534   case X86::VGATHERPF1DPSm:
8535   case X86::VGATHERPF1QPDm:
8536   case X86::VGATHERPF1QPSm:
8537   case X86::VGATHERQPDYrm:
8538   case X86::VGATHERQPDZ128rm:
8539   case X86::VGATHERQPDZ256rm:
8540   case X86::VGATHERQPDZrm:
8541   case X86::VGATHERQPDrm:
8542   case X86::VGATHERQPSYrm:
8543   case X86::VGATHERQPSZ128rm:
8544   case X86::VGATHERQPSZ256rm:
8545   case X86::VGATHERQPSZrm:
8546   case X86::VGATHERQPSrm:
8547   case X86::VPGATHERDDYrm:
8548   case X86::VPGATHERDDZ128rm:
8549   case X86::VPGATHERDDZ256rm:
8550   case X86::VPGATHERDDZrm:
8551   case X86::VPGATHERDDrm:
8552   case X86::VPGATHERDQYrm:
8553   case X86::VPGATHERDQZ128rm:
8554   case X86::VPGATHERDQZ256rm:
8555   case X86::VPGATHERDQZrm:
8556   case X86::VPGATHERDQrm:
8557   case X86::VPGATHERQDYrm:
8558   case X86::VPGATHERQDZ128rm:
8559   case X86::VPGATHERQDZ256rm:
8560   case X86::VPGATHERQDZrm:
8561   case X86::VPGATHERQDrm:
8562   case X86::VPGATHERQQYrm:
8563   case X86::VPGATHERQQZ128rm:
8564   case X86::VPGATHERQQZ256rm:
8565   case X86::VPGATHERQQZrm:
8566   case X86::VPGATHERQQrm:
8567   case X86::VSCATTERDPDZ128mr:
8568   case X86::VSCATTERDPDZ256mr:
8569   case X86::VSCATTERDPDZmr:
8570   case X86::VSCATTERDPSZ128mr:
8571   case X86::VSCATTERDPSZ256mr:
8572   case X86::VSCATTERDPSZmr:
8573   case X86::VSCATTERPF0DPDm:
8574   case X86::VSCATTERPF0DPSm:
8575   case X86::VSCATTERPF0QPDm:
8576   case X86::VSCATTERPF0QPSm:
8577   case X86::VSCATTERPF1DPDm:
8578   case X86::VSCATTERPF1DPSm:
8579   case X86::VSCATTERPF1QPDm:
8580   case X86::VSCATTERPF1QPSm:
8581   case X86::VSCATTERQPDZ128mr:
8582   case X86::VSCATTERQPDZ256mr:
8583   case X86::VSCATTERQPDZmr:
8584   case X86::VSCATTERQPSZ128mr:
8585   case X86::VSCATTERQPSZ256mr:
8586   case X86::VSCATTERQPSZmr:
8587   case X86::VPSCATTERDDZ128mr:
8588   case X86::VPSCATTERDDZ256mr:
8589   case X86::VPSCATTERDDZmr:
8590   case X86::VPSCATTERDQZ128mr:
8591   case X86::VPSCATTERDQZ256mr:
8592   case X86::VPSCATTERDQZmr:
8593   case X86::VPSCATTERQDZ128mr:
8594   case X86::VPSCATTERQDZ256mr:
8595   case X86::VPSCATTERQDZmr:
8596   case X86::VPSCATTERQQZ128mr:
8597   case X86::VPSCATTERQQZ256mr:
8598   case X86::VPSCATTERQQZmr:
8599     return true;
8600   }
8601 }
8602 
8603 bool X86InstrInfo::hasHighOperandLatency(const TargetSchedModel &SchedModel,
8604                                          const MachineRegisterInfo *MRI,
8605                                          const MachineInstr &DefMI,
8606                                          unsigned DefIdx,
8607                                          const MachineInstr &UseMI,
8608                                          unsigned UseIdx) const {
8609   return isHighLatencyDef(DefMI.getOpcode());
8610 }
8611 
8612 bool X86InstrInfo::hasReassociableOperands(const MachineInstr &Inst,
8613                                            const MachineBasicBlock *MBB) const {
8614   assert(Inst.getNumExplicitOperands() == 3 && Inst.getNumExplicitDefs() == 1 &&
8615          Inst.getNumDefs() <= 2 && "Reassociation needs binary operators");
8616 
8617   // Integer binary math/logic instructions have a third source operand:
8618   // the EFLAGS register. That operand must be both defined here and never
8619   // used; ie, it must be dead. If the EFLAGS operand is live, then we can
8620   // not change anything because rearranging the operands could affect other
8621   // instructions that depend on the exact status flags (zero, sign, etc.)
8622   // that are set by using these particular operands with this operation.
8623   const MachineOperand *FlagDef = Inst.findRegisterDefOperand(X86::EFLAGS);
8624   assert((Inst.getNumDefs() == 1 || FlagDef) &&
8625          "Implicit def isn't flags?");
8626   if (FlagDef && !FlagDef->isDead())
8627     return false;
8628 
8629   return TargetInstrInfo::hasReassociableOperands(Inst, MBB);
8630 }
8631 
8632 // TODO: There are many more machine instruction opcodes to match:
8633 //       1. Other data types (integer, vectors)
8634 //       2. Other math / logic operations (xor, or)
8635 //       3. Other forms of the same operation (intrinsics and other variants)
8636 bool X86InstrInfo::isAssociativeAndCommutative(const MachineInstr &Inst) const {
8637   switch (Inst.getOpcode()) {
8638   case X86::AND8rr:
8639   case X86::AND16rr:
8640   case X86::AND32rr:
8641   case X86::AND64rr:
8642   case X86::OR8rr:
8643   case X86::OR16rr:
8644   case X86::OR32rr:
8645   case X86::OR64rr:
8646   case X86::XOR8rr:
8647   case X86::XOR16rr:
8648   case X86::XOR32rr:
8649   case X86::XOR64rr:
8650   case X86::IMUL16rr:
8651   case X86::IMUL32rr:
8652   case X86::IMUL64rr:
8653   case X86::PANDrr:
8654   case X86::PORrr:
8655   case X86::PXORrr:
8656   case X86::ANDPDrr:
8657   case X86::ANDPSrr:
8658   case X86::ORPDrr:
8659   case X86::ORPSrr:
8660   case X86::XORPDrr:
8661   case X86::XORPSrr:
8662   case X86::PADDBrr:
8663   case X86::PADDWrr:
8664   case X86::PADDDrr:
8665   case X86::PADDQrr:
8666   case X86::PMULLWrr:
8667   case X86::PMULLDrr:
8668   case X86::PMAXSBrr:
8669   case X86::PMAXSDrr:
8670   case X86::PMAXSWrr:
8671   case X86::PMAXUBrr:
8672   case X86::PMAXUDrr:
8673   case X86::PMAXUWrr:
8674   case X86::PMINSBrr:
8675   case X86::PMINSDrr:
8676   case X86::PMINSWrr:
8677   case X86::PMINUBrr:
8678   case X86::PMINUDrr:
8679   case X86::PMINUWrr:
8680   case X86::VPANDrr:
8681   case X86::VPANDYrr:
8682   case X86::VPANDDZ128rr:
8683   case X86::VPANDDZ256rr:
8684   case X86::VPANDDZrr:
8685   case X86::VPANDQZ128rr:
8686   case X86::VPANDQZ256rr:
8687   case X86::VPANDQZrr:
8688   case X86::VPORrr:
8689   case X86::VPORYrr:
8690   case X86::VPORDZ128rr:
8691   case X86::VPORDZ256rr:
8692   case X86::VPORDZrr:
8693   case X86::VPORQZ128rr:
8694   case X86::VPORQZ256rr:
8695   case X86::VPORQZrr:
8696   case X86::VPXORrr:
8697   case X86::VPXORYrr:
8698   case X86::VPXORDZ128rr:
8699   case X86::VPXORDZ256rr:
8700   case X86::VPXORDZrr:
8701   case X86::VPXORQZ128rr:
8702   case X86::VPXORQZ256rr:
8703   case X86::VPXORQZrr:
8704   case X86::VANDPDrr:
8705   case X86::VANDPSrr:
8706   case X86::VANDPDYrr:
8707   case X86::VANDPSYrr:
8708   case X86::VANDPDZ128rr:
8709   case X86::VANDPSZ128rr:
8710   case X86::VANDPDZ256rr:
8711   case X86::VANDPSZ256rr:
8712   case X86::VANDPDZrr:
8713   case X86::VANDPSZrr:
8714   case X86::VORPDrr:
8715   case X86::VORPSrr:
8716   case X86::VORPDYrr:
8717   case X86::VORPSYrr:
8718   case X86::VORPDZ128rr:
8719   case X86::VORPSZ128rr:
8720   case X86::VORPDZ256rr:
8721   case X86::VORPSZ256rr:
8722   case X86::VORPDZrr:
8723   case X86::VORPSZrr:
8724   case X86::VXORPDrr:
8725   case X86::VXORPSrr:
8726   case X86::VXORPDYrr:
8727   case X86::VXORPSYrr:
8728   case X86::VXORPDZ128rr:
8729   case X86::VXORPSZ128rr:
8730   case X86::VXORPDZ256rr:
8731   case X86::VXORPSZ256rr:
8732   case X86::VXORPDZrr:
8733   case X86::VXORPSZrr:
8734   case X86::KADDBrr:
8735   case X86::KADDWrr:
8736   case X86::KADDDrr:
8737   case X86::KADDQrr:
8738   case X86::KANDBrr:
8739   case X86::KANDWrr:
8740   case X86::KANDDrr:
8741   case X86::KANDQrr:
8742   case X86::KORBrr:
8743   case X86::KORWrr:
8744   case X86::KORDrr:
8745   case X86::KORQrr:
8746   case X86::KXORBrr:
8747   case X86::KXORWrr:
8748   case X86::KXORDrr:
8749   case X86::KXORQrr:
8750   case X86::VPADDBrr:
8751   case X86::VPADDWrr:
8752   case X86::VPADDDrr:
8753   case X86::VPADDQrr:
8754   case X86::VPADDBYrr:
8755   case X86::VPADDWYrr:
8756   case X86::VPADDDYrr:
8757   case X86::VPADDQYrr:
8758   case X86::VPADDBZ128rr:
8759   case X86::VPADDWZ128rr:
8760   case X86::VPADDDZ128rr:
8761   case X86::VPADDQZ128rr:
8762   case X86::VPADDBZ256rr:
8763   case X86::VPADDWZ256rr:
8764   case X86::VPADDDZ256rr:
8765   case X86::VPADDQZ256rr:
8766   case X86::VPADDBZrr:
8767   case X86::VPADDWZrr:
8768   case X86::VPADDDZrr:
8769   case X86::VPADDQZrr:
8770   case X86::VPMULLWrr:
8771   case X86::VPMULLWYrr:
8772   case X86::VPMULLWZ128rr:
8773   case X86::VPMULLWZ256rr:
8774   case X86::VPMULLWZrr:
8775   case X86::VPMULLDrr:
8776   case X86::VPMULLDYrr:
8777   case X86::VPMULLDZ128rr:
8778   case X86::VPMULLDZ256rr:
8779   case X86::VPMULLDZrr:
8780   case X86::VPMULLQZ128rr:
8781   case X86::VPMULLQZ256rr:
8782   case X86::VPMULLQZrr:
8783   case X86::VPMAXSBrr:
8784   case X86::VPMAXSBYrr:
8785   case X86::VPMAXSBZ128rr:
8786   case X86::VPMAXSBZ256rr:
8787   case X86::VPMAXSBZrr:
8788   case X86::VPMAXSDrr:
8789   case X86::VPMAXSDYrr:
8790   case X86::VPMAXSDZ128rr:
8791   case X86::VPMAXSDZ256rr:
8792   case X86::VPMAXSDZrr:
8793   case X86::VPMAXSQZ128rr:
8794   case X86::VPMAXSQZ256rr:
8795   case X86::VPMAXSQZrr:
8796   case X86::VPMAXSWrr:
8797   case X86::VPMAXSWYrr:
8798   case X86::VPMAXSWZ128rr:
8799   case X86::VPMAXSWZ256rr:
8800   case X86::VPMAXSWZrr:
8801   case X86::VPMAXUBrr:
8802   case X86::VPMAXUBYrr:
8803   case X86::VPMAXUBZ128rr:
8804   case X86::VPMAXUBZ256rr:
8805   case X86::VPMAXUBZrr:
8806   case X86::VPMAXUDrr:
8807   case X86::VPMAXUDYrr:
8808   case X86::VPMAXUDZ128rr:
8809   case X86::VPMAXUDZ256rr:
8810   case X86::VPMAXUDZrr:
8811   case X86::VPMAXUQZ128rr:
8812   case X86::VPMAXUQZ256rr:
8813   case X86::VPMAXUQZrr:
8814   case X86::VPMAXUWrr:
8815   case X86::VPMAXUWYrr:
8816   case X86::VPMAXUWZ128rr:
8817   case X86::VPMAXUWZ256rr:
8818   case X86::VPMAXUWZrr:
8819   case X86::VPMINSBrr:
8820   case X86::VPMINSBYrr:
8821   case X86::VPMINSBZ128rr:
8822   case X86::VPMINSBZ256rr:
8823   case X86::VPMINSBZrr:
8824   case X86::VPMINSDrr:
8825   case X86::VPMINSDYrr:
8826   case X86::VPMINSDZ128rr:
8827   case X86::VPMINSDZ256rr:
8828   case X86::VPMINSDZrr:
8829   case X86::VPMINSQZ128rr:
8830   case X86::VPMINSQZ256rr:
8831   case X86::VPMINSQZrr:
8832   case X86::VPMINSWrr:
8833   case X86::VPMINSWYrr:
8834   case X86::VPMINSWZ128rr:
8835   case X86::VPMINSWZ256rr:
8836   case X86::VPMINSWZrr:
8837   case X86::VPMINUBrr:
8838   case X86::VPMINUBYrr:
8839   case X86::VPMINUBZ128rr:
8840   case X86::VPMINUBZ256rr:
8841   case X86::VPMINUBZrr:
8842   case X86::VPMINUDrr:
8843   case X86::VPMINUDYrr:
8844   case X86::VPMINUDZ128rr:
8845   case X86::VPMINUDZ256rr:
8846   case X86::VPMINUDZrr:
8847   case X86::VPMINUQZ128rr:
8848   case X86::VPMINUQZ256rr:
8849   case X86::VPMINUQZrr:
8850   case X86::VPMINUWrr:
8851   case X86::VPMINUWYrr:
8852   case X86::VPMINUWZ128rr:
8853   case X86::VPMINUWZ256rr:
8854   case X86::VPMINUWZrr:
8855   // Normal min/max instructions are not commutative because of NaN and signed
8856   // zero semantics, but these are. Thus, there's no need to check for global
8857   // relaxed math; the instructions themselves have the properties we need.
8858   case X86::MAXCPDrr:
8859   case X86::MAXCPSrr:
8860   case X86::MAXCSDrr:
8861   case X86::MAXCSSrr:
8862   case X86::MINCPDrr:
8863   case X86::MINCPSrr:
8864   case X86::MINCSDrr:
8865   case X86::MINCSSrr:
8866   case X86::VMAXCPDrr:
8867   case X86::VMAXCPSrr:
8868   case X86::VMAXCPDYrr:
8869   case X86::VMAXCPSYrr:
8870   case X86::VMAXCPDZ128rr:
8871   case X86::VMAXCPSZ128rr:
8872   case X86::VMAXCPDZ256rr:
8873   case X86::VMAXCPSZ256rr:
8874   case X86::VMAXCPDZrr:
8875   case X86::VMAXCPSZrr:
8876   case X86::VMAXCSDrr:
8877   case X86::VMAXCSSrr:
8878   case X86::VMAXCSDZrr:
8879   case X86::VMAXCSSZrr:
8880   case X86::VMINCPDrr:
8881   case X86::VMINCPSrr:
8882   case X86::VMINCPDYrr:
8883   case X86::VMINCPSYrr:
8884   case X86::VMINCPDZ128rr:
8885   case X86::VMINCPSZ128rr:
8886   case X86::VMINCPDZ256rr:
8887   case X86::VMINCPSZ256rr:
8888   case X86::VMINCPDZrr:
8889   case X86::VMINCPSZrr:
8890   case X86::VMINCSDrr:
8891   case X86::VMINCSSrr:
8892   case X86::VMINCSDZrr:
8893   case X86::VMINCSSZrr:
8894   case X86::VMAXCPHZ128rr:
8895   case X86::VMAXCPHZ256rr:
8896   case X86::VMAXCPHZrr:
8897   case X86::VMAXCSHZrr:
8898   case X86::VMINCPHZ128rr:
8899   case X86::VMINCPHZ256rr:
8900   case X86::VMINCPHZrr:
8901   case X86::VMINCSHZrr:
8902     return true;
8903   case X86::ADDPDrr:
8904   case X86::ADDPSrr:
8905   case X86::ADDSDrr:
8906   case X86::ADDSSrr:
8907   case X86::MULPDrr:
8908   case X86::MULPSrr:
8909   case X86::MULSDrr:
8910   case X86::MULSSrr:
8911   case X86::VADDPDrr:
8912   case X86::VADDPSrr:
8913   case X86::VADDPDYrr:
8914   case X86::VADDPSYrr:
8915   case X86::VADDPDZ128rr:
8916   case X86::VADDPSZ128rr:
8917   case X86::VADDPDZ256rr:
8918   case X86::VADDPSZ256rr:
8919   case X86::VADDPDZrr:
8920   case X86::VADDPSZrr:
8921   case X86::VADDSDrr:
8922   case X86::VADDSSrr:
8923   case X86::VADDSDZrr:
8924   case X86::VADDSSZrr:
8925   case X86::VMULPDrr:
8926   case X86::VMULPSrr:
8927   case X86::VMULPDYrr:
8928   case X86::VMULPSYrr:
8929   case X86::VMULPDZ128rr:
8930   case X86::VMULPSZ128rr:
8931   case X86::VMULPDZ256rr:
8932   case X86::VMULPSZ256rr:
8933   case X86::VMULPDZrr:
8934   case X86::VMULPSZrr:
8935   case X86::VMULSDrr:
8936   case X86::VMULSSrr:
8937   case X86::VMULSDZrr:
8938   case X86::VMULSSZrr:
8939   case X86::VADDPHZ128rr:
8940   case X86::VADDPHZ256rr:
8941   case X86::VADDPHZrr:
8942   case X86::VADDSHZrr:
8943   case X86::VMULPHZ128rr:
8944   case X86::VMULPHZ256rr:
8945   case X86::VMULPHZrr:
8946   case X86::VMULSHZrr:
8947     return Inst.getFlag(MachineInstr::MIFlag::FmReassoc) &&
8948            Inst.getFlag(MachineInstr::MIFlag::FmNsz);
8949   default:
8950     return false;
8951   }
8952 }
8953 
8954 /// If \p DescribedReg overlaps with the MOVrr instruction's destination
8955 /// register then, if possible, describe the value in terms of the source
8956 /// register.
8957 static Optional<ParamLoadedValue>
8958 describeMOVrrLoadedValue(const MachineInstr &MI, Register DescribedReg,
8959                          const TargetRegisterInfo *TRI) {
8960   Register DestReg = MI.getOperand(0).getReg();
8961   Register SrcReg = MI.getOperand(1).getReg();
8962 
8963   auto Expr = DIExpression::get(MI.getMF()->getFunction().getContext(), {});
8964 
8965   // If the described register is the destination, just return the source.
8966   if (DestReg == DescribedReg)
8967     return ParamLoadedValue(MachineOperand::CreateReg(SrcReg, false), Expr);
8968 
8969   // If the described register is a sub-register of the destination register,
8970   // then pick out the source register's corresponding sub-register.
8971   if (unsigned SubRegIdx = TRI->getSubRegIndex(DestReg, DescribedReg)) {
8972     Register SrcSubReg = TRI->getSubReg(SrcReg, SubRegIdx);
8973     return ParamLoadedValue(MachineOperand::CreateReg(SrcSubReg, false), Expr);
8974   }
8975 
8976   // The remaining case to consider is when the described register is a
8977   // super-register of the destination register. MOV8rr and MOV16rr does not
8978   // write to any of the other bytes in the register, meaning that we'd have to
8979   // describe the value using a combination of the source register and the
8980   // non-overlapping bits in the described register, which is not currently
8981   // possible.
8982   if (MI.getOpcode() == X86::MOV8rr || MI.getOpcode() == X86::MOV16rr ||
8983       !TRI->isSuperRegister(DestReg, DescribedReg))
8984     return None;
8985 
8986   assert(MI.getOpcode() == X86::MOV32rr && "Unexpected super-register case");
8987   return ParamLoadedValue(MachineOperand::CreateReg(SrcReg, false), Expr);
8988 }
8989 
8990 Optional<ParamLoadedValue>
8991 X86InstrInfo::describeLoadedValue(const MachineInstr &MI, Register Reg) const {
8992   const MachineOperand *Op = nullptr;
8993   DIExpression *Expr = nullptr;
8994 
8995   const TargetRegisterInfo *TRI = &getRegisterInfo();
8996 
8997   switch (MI.getOpcode()) {
8998   case X86::LEA32r:
8999   case X86::LEA64r:
9000   case X86::LEA64_32r: {
9001     // We may need to describe a 64-bit parameter with a 32-bit LEA.
9002     if (!TRI->isSuperRegisterEq(MI.getOperand(0).getReg(), Reg))
9003       return None;
9004 
9005     // Operand 4 could be global address. For now we do not support
9006     // such situation.
9007     if (!MI.getOperand(4).isImm() || !MI.getOperand(2).isImm())
9008       return None;
9009 
9010     const MachineOperand &Op1 = MI.getOperand(1);
9011     const MachineOperand &Op2 = MI.getOperand(3);
9012     assert(Op2.isReg() && (Op2.getReg() == X86::NoRegister ||
9013                            Register::isPhysicalRegister(Op2.getReg())));
9014 
9015     // Omit situations like:
9016     // %rsi = lea %rsi, 4, ...
9017     if ((Op1.isReg() && Op1.getReg() == MI.getOperand(0).getReg()) ||
9018         Op2.getReg() == MI.getOperand(0).getReg())
9019       return None;
9020     else if ((Op1.isReg() && Op1.getReg() != X86::NoRegister &&
9021               TRI->regsOverlap(Op1.getReg(), MI.getOperand(0).getReg())) ||
9022              (Op2.getReg() != X86::NoRegister &&
9023               TRI->regsOverlap(Op2.getReg(), MI.getOperand(0).getReg())))
9024       return None;
9025 
9026     int64_t Coef = MI.getOperand(2).getImm();
9027     int64_t Offset = MI.getOperand(4).getImm();
9028     SmallVector<uint64_t, 8> Ops;
9029 
9030     if ((Op1.isReg() && Op1.getReg() != X86::NoRegister)) {
9031       Op = &Op1;
9032     } else if (Op1.isFI())
9033       Op = &Op1;
9034 
9035     if (Op && Op->isReg() && Op->getReg() == Op2.getReg() && Coef > 0) {
9036       Ops.push_back(dwarf::DW_OP_constu);
9037       Ops.push_back(Coef + 1);
9038       Ops.push_back(dwarf::DW_OP_mul);
9039     } else {
9040       if (Op && Op2.getReg() != X86::NoRegister) {
9041         int dwarfReg = TRI->getDwarfRegNum(Op2.getReg(), false);
9042         if (dwarfReg < 0)
9043           return None;
9044         else if (dwarfReg < 32) {
9045           Ops.push_back(dwarf::DW_OP_breg0 + dwarfReg);
9046           Ops.push_back(0);
9047         } else {
9048           Ops.push_back(dwarf::DW_OP_bregx);
9049           Ops.push_back(dwarfReg);
9050           Ops.push_back(0);
9051         }
9052       } else if (!Op) {
9053         assert(Op2.getReg() != X86::NoRegister);
9054         Op = &Op2;
9055       }
9056 
9057       if (Coef > 1) {
9058         assert(Op2.getReg() != X86::NoRegister);
9059         Ops.push_back(dwarf::DW_OP_constu);
9060         Ops.push_back(Coef);
9061         Ops.push_back(dwarf::DW_OP_mul);
9062       }
9063 
9064       if (((Op1.isReg() && Op1.getReg() != X86::NoRegister) || Op1.isFI()) &&
9065           Op2.getReg() != X86::NoRegister) {
9066         Ops.push_back(dwarf::DW_OP_plus);
9067       }
9068     }
9069 
9070     DIExpression::appendOffset(Ops, Offset);
9071     Expr = DIExpression::get(MI.getMF()->getFunction().getContext(), Ops);
9072 
9073     return ParamLoadedValue(*Op, Expr);;
9074   }
9075   case X86::MOV8ri:
9076   case X86::MOV16ri:
9077     // TODO: Handle MOV8ri and MOV16ri.
9078     return None;
9079   case X86::MOV32ri:
9080   case X86::MOV64ri:
9081   case X86::MOV64ri32:
9082     // MOV32ri may be used for producing zero-extended 32-bit immediates in
9083     // 64-bit parameters, so we need to consider super-registers.
9084     if (!TRI->isSuperRegisterEq(MI.getOperand(0).getReg(), Reg))
9085       return None;
9086     return ParamLoadedValue(MI.getOperand(1), Expr);
9087   case X86::MOV8rr:
9088   case X86::MOV16rr:
9089   case X86::MOV32rr:
9090   case X86::MOV64rr:
9091     return describeMOVrrLoadedValue(MI, Reg, TRI);
9092   case X86::XOR32rr: {
9093     // 64-bit parameters are zero-materialized using XOR32rr, so also consider
9094     // super-registers.
9095     if (!TRI->isSuperRegisterEq(MI.getOperand(0).getReg(), Reg))
9096       return None;
9097     if (MI.getOperand(1).getReg() == MI.getOperand(2).getReg())
9098       return ParamLoadedValue(MachineOperand::CreateImm(0), Expr);
9099     return None;
9100   }
9101   case X86::MOVSX64rr32: {
9102     // We may need to describe the lower 32 bits of the MOVSX; for example, in
9103     // cases like this:
9104     //
9105     //  $ebx = [...]
9106     //  $rdi = MOVSX64rr32 $ebx
9107     //  $esi = MOV32rr $edi
9108     if (!TRI->isSubRegisterEq(MI.getOperand(0).getReg(), Reg))
9109       return None;
9110 
9111     Expr = DIExpression::get(MI.getMF()->getFunction().getContext(), {});
9112 
9113     // If the described register is the destination register we need to
9114     // sign-extend the source register from 32 bits. The other case we handle
9115     // is when the described register is the 32-bit sub-register of the
9116     // destination register, in case we just need to return the source
9117     // register.
9118     if (Reg == MI.getOperand(0).getReg())
9119       Expr = DIExpression::appendExt(Expr, 32, 64, true);
9120     else
9121       assert(X86MCRegisterClasses[X86::GR32RegClassID].contains(Reg) &&
9122              "Unhandled sub-register case for MOVSX64rr32");
9123 
9124     return ParamLoadedValue(MI.getOperand(1), Expr);
9125   }
9126   default:
9127     assert(!MI.isMoveImmediate() && "Unexpected MoveImm instruction");
9128     return TargetInstrInfo::describeLoadedValue(MI, Reg);
9129   }
9130 }
9131 
9132 /// This is an architecture-specific helper function of reassociateOps.
9133 /// Set special operand attributes for new instructions after reassociation.
9134 void X86InstrInfo::setSpecialOperandAttr(MachineInstr &OldMI1,
9135                                          MachineInstr &OldMI2,
9136                                          MachineInstr &NewMI1,
9137                                          MachineInstr &NewMI2) const {
9138   // Propagate FP flags from the original instructions.
9139   // But clear poison-generating flags because those may not be valid now.
9140   // TODO: There should be a helper function for copying only fast-math-flags.
9141   uint16_t IntersectedFlags = OldMI1.getFlags() & OldMI2.getFlags();
9142   NewMI1.setFlags(IntersectedFlags);
9143   NewMI1.clearFlag(MachineInstr::MIFlag::NoSWrap);
9144   NewMI1.clearFlag(MachineInstr::MIFlag::NoUWrap);
9145   NewMI1.clearFlag(MachineInstr::MIFlag::IsExact);
9146 
9147   NewMI2.setFlags(IntersectedFlags);
9148   NewMI2.clearFlag(MachineInstr::MIFlag::NoSWrap);
9149   NewMI2.clearFlag(MachineInstr::MIFlag::NoUWrap);
9150   NewMI2.clearFlag(MachineInstr::MIFlag::IsExact);
9151 
9152   // Integer instructions may define an implicit EFLAGS dest register operand.
9153   MachineOperand *OldFlagDef1 = OldMI1.findRegisterDefOperand(X86::EFLAGS);
9154   MachineOperand *OldFlagDef2 = OldMI2.findRegisterDefOperand(X86::EFLAGS);
9155 
9156   assert(!OldFlagDef1 == !OldFlagDef2 &&
9157          "Unexpected instruction type for reassociation");
9158 
9159   if (!OldFlagDef1 || !OldFlagDef2)
9160     return;
9161 
9162   assert(OldFlagDef1->isDead() && OldFlagDef2->isDead() &&
9163          "Must have dead EFLAGS operand in reassociable instruction");
9164 
9165   MachineOperand *NewFlagDef1 = NewMI1.findRegisterDefOperand(X86::EFLAGS);
9166   MachineOperand *NewFlagDef2 = NewMI2.findRegisterDefOperand(X86::EFLAGS);
9167 
9168   assert(NewFlagDef1 && NewFlagDef2 &&
9169          "Unexpected operand in reassociable instruction");
9170 
9171   // Mark the new EFLAGS operands as dead to be helpful to subsequent iterations
9172   // of this pass or other passes. The EFLAGS operands must be dead in these new
9173   // instructions because the EFLAGS operands in the original instructions must
9174   // be dead in order for reassociation to occur.
9175   NewFlagDef1->setIsDead();
9176   NewFlagDef2->setIsDead();
9177 }
9178 
9179 std::pair<unsigned, unsigned>
9180 X86InstrInfo::decomposeMachineOperandsTargetFlags(unsigned TF) const {
9181   return std::make_pair(TF, 0u);
9182 }
9183 
9184 ArrayRef<std::pair<unsigned, const char *>>
9185 X86InstrInfo::getSerializableDirectMachineOperandTargetFlags() const {
9186   using namespace X86II;
9187   static const std::pair<unsigned, const char *> TargetFlags[] = {
9188       {MO_GOT_ABSOLUTE_ADDRESS, "x86-got-absolute-address"},
9189       {MO_PIC_BASE_OFFSET, "x86-pic-base-offset"},
9190       {MO_GOT, "x86-got"},
9191       {MO_GOTOFF, "x86-gotoff"},
9192       {MO_GOTPCREL, "x86-gotpcrel"},
9193       {MO_GOTPCREL_NORELAX, "x86-gotpcrel-norelax"},
9194       {MO_PLT, "x86-plt"},
9195       {MO_TLSGD, "x86-tlsgd"},
9196       {MO_TLSLD, "x86-tlsld"},
9197       {MO_TLSLDM, "x86-tlsldm"},
9198       {MO_GOTTPOFF, "x86-gottpoff"},
9199       {MO_INDNTPOFF, "x86-indntpoff"},
9200       {MO_TPOFF, "x86-tpoff"},
9201       {MO_DTPOFF, "x86-dtpoff"},
9202       {MO_NTPOFF, "x86-ntpoff"},
9203       {MO_GOTNTPOFF, "x86-gotntpoff"},
9204       {MO_DLLIMPORT, "x86-dllimport"},
9205       {MO_DARWIN_NONLAZY, "x86-darwin-nonlazy"},
9206       {MO_DARWIN_NONLAZY_PIC_BASE, "x86-darwin-nonlazy-pic-base"},
9207       {MO_TLVP, "x86-tlvp"},
9208       {MO_TLVP_PIC_BASE, "x86-tlvp-pic-base"},
9209       {MO_SECREL, "x86-secrel"},
9210       {MO_COFFSTUB, "x86-coffstub"}};
9211   return makeArrayRef(TargetFlags);
9212 }
9213 
9214 namespace {
9215   /// Create Global Base Reg pass. This initializes the PIC
9216   /// global base register for x86-32.
9217   struct CGBR : public MachineFunctionPass {
9218     static char ID;
9219     CGBR() : MachineFunctionPass(ID) {}
9220 
9221     bool runOnMachineFunction(MachineFunction &MF) override {
9222       const X86TargetMachine *TM =
9223         static_cast<const X86TargetMachine *>(&MF.getTarget());
9224       const X86Subtarget &STI = MF.getSubtarget<X86Subtarget>();
9225 
9226       // Don't do anything in the 64-bit small and kernel code models. They use
9227       // RIP-relative addressing for everything.
9228       if (STI.is64Bit() && (TM->getCodeModel() == CodeModel::Small ||
9229                             TM->getCodeModel() == CodeModel::Kernel))
9230         return false;
9231 
9232       // Only emit a global base reg in PIC mode.
9233       if (!TM->isPositionIndependent())
9234         return false;
9235 
9236       X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
9237       Register GlobalBaseReg = X86FI->getGlobalBaseReg();
9238 
9239       // If we didn't need a GlobalBaseReg, don't insert code.
9240       if (GlobalBaseReg == 0)
9241         return false;
9242 
9243       // Insert the set of GlobalBaseReg into the first MBB of the function
9244       MachineBasicBlock &FirstMBB = MF.front();
9245       MachineBasicBlock::iterator MBBI = FirstMBB.begin();
9246       DebugLoc DL = FirstMBB.findDebugLoc(MBBI);
9247       MachineRegisterInfo &RegInfo = MF.getRegInfo();
9248       const X86InstrInfo *TII = STI.getInstrInfo();
9249 
9250       Register PC;
9251       if (STI.isPICStyleGOT())
9252         PC = RegInfo.createVirtualRegister(&X86::GR32RegClass);
9253       else
9254         PC = GlobalBaseReg;
9255 
9256       if (STI.is64Bit()) {
9257         if (TM->getCodeModel() == CodeModel::Medium) {
9258           // In the medium code model, use a RIP-relative LEA to materialize the
9259           // GOT.
9260           BuildMI(FirstMBB, MBBI, DL, TII->get(X86::LEA64r), PC)
9261               .addReg(X86::RIP)
9262               .addImm(0)
9263               .addReg(0)
9264               .addExternalSymbol("_GLOBAL_OFFSET_TABLE_")
9265               .addReg(0);
9266         } else if (TM->getCodeModel() == CodeModel::Large) {
9267           // In the large code model, we are aiming for this code, though the
9268           // register allocation may vary:
9269           //   leaq .LN$pb(%rip), %rax
9270           //   movq $_GLOBAL_OFFSET_TABLE_ - .LN$pb, %rcx
9271           //   addq %rcx, %rax
9272           // RAX now holds address of _GLOBAL_OFFSET_TABLE_.
9273           Register PBReg = RegInfo.createVirtualRegister(&X86::GR64RegClass);
9274           Register GOTReg = RegInfo.createVirtualRegister(&X86::GR64RegClass);
9275           BuildMI(FirstMBB, MBBI, DL, TII->get(X86::LEA64r), PBReg)
9276               .addReg(X86::RIP)
9277               .addImm(0)
9278               .addReg(0)
9279               .addSym(MF.getPICBaseSymbol())
9280               .addReg(0);
9281           std::prev(MBBI)->setPreInstrSymbol(MF, MF.getPICBaseSymbol());
9282           BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOV64ri), GOTReg)
9283               .addExternalSymbol("_GLOBAL_OFFSET_TABLE_",
9284                                  X86II::MO_PIC_BASE_OFFSET);
9285           BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD64rr), PC)
9286               .addReg(PBReg, RegState::Kill)
9287               .addReg(GOTReg, RegState::Kill);
9288         } else {
9289           llvm_unreachable("unexpected code model");
9290         }
9291       } else {
9292         // Operand of MovePCtoStack is completely ignored by asm printer. It's
9293         // only used in JIT code emission as displacement to pc.
9294         BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOVPC32r), PC).addImm(0);
9295 
9296         // If we're using vanilla 'GOT' PIC style, we should use relative
9297         // addressing not to pc, but to _GLOBAL_OFFSET_TABLE_ external.
9298         if (STI.isPICStyleGOT()) {
9299           // Generate addl $__GLOBAL_OFFSET_TABLE_ + [.-piclabel],
9300           // %some_register
9301           BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD32ri), GlobalBaseReg)
9302               .addReg(PC)
9303               .addExternalSymbol("_GLOBAL_OFFSET_TABLE_",
9304                                  X86II::MO_GOT_ABSOLUTE_ADDRESS);
9305         }
9306       }
9307 
9308       return true;
9309     }
9310 
9311     StringRef getPassName() const override {
9312       return "X86 PIC Global Base Reg Initialization";
9313     }
9314 
9315     void getAnalysisUsage(AnalysisUsage &AU) const override {
9316       AU.setPreservesCFG();
9317       MachineFunctionPass::getAnalysisUsage(AU);
9318     }
9319   };
9320 } // namespace
9321 
9322 char CGBR::ID = 0;
9323 FunctionPass*
9324 llvm::createX86GlobalBaseRegPass() { return new CGBR(); }
9325 
9326 namespace {
9327   struct LDTLSCleanup : public MachineFunctionPass {
9328     static char ID;
9329     LDTLSCleanup() : MachineFunctionPass(ID) {}
9330 
9331     bool runOnMachineFunction(MachineFunction &MF) override {
9332       if (skipFunction(MF.getFunction()))
9333         return false;
9334 
9335       X86MachineFunctionInfo *MFI = MF.getInfo<X86MachineFunctionInfo>();
9336       if (MFI->getNumLocalDynamicTLSAccesses() < 2) {
9337         // No point folding accesses if there isn't at least two.
9338         return false;
9339       }
9340 
9341       MachineDominatorTree *DT = &getAnalysis<MachineDominatorTree>();
9342       return VisitNode(DT->getRootNode(), 0);
9343     }
9344 
9345     // Visit the dominator subtree rooted at Node in pre-order.
9346     // If TLSBaseAddrReg is non-null, then use that to replace any
9347     // TLS_base_addr instructions. Otherwise, create the register
9348     // when the first such instruction is seen, and then use it
9349     // as we encounter more instructions.
9350     bool VisitNode(MachineDomTreeNode *Node, unsigned TLSBaseAddrReg) {
9351       MachineBasicBlock *BB = Node->getBlock();
9352       bool Changed = false;
9353 
9354       // Traverse the current block.
9355       for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;
9356            ++I) {
9357         switch (I->getOpcode()) {
9358           case X86::TLS_base_addr32:
9359           case X86::TLS_base_addr64:
9360             if (TLSBaseAddrReg)
9361               I = ReplaceTLSBaseAddrCall(*I, TLSBaseAddrReg);
9362             else
9363               I = SetRegister(*I, &TLSBaseAddrReg);
9364             Changed = true;
9365             break;
9366           default:
9367             break;
9368         }
9369       }
9370 
9371       // Visit the children of this block in the dominator tree.
9372       for (auto I = Node->begin(), E = Node->end(); I != E; ++I) {
9373         Changed |= VisitNode(*I, TLSBaseAddrReg);
9374       }
9375 
9376       return Changed;
9377     }
9378 
9379     // Replace the TLS_base_addr instruction I with a copy from
9380     // TLSBaseAddrReg, returning the new instruction.
9381     MachineInstr *ReplaceTLSBaseAddrCall(MachineInstr &I,
9382                                          unsigned TLSBaseAddrReg) {
9383       MachineFunction *MF = I.getParent()->getParent();
9384       const X86Subtarget &STI = MF->getSubtarget<X86Subtarget>();
9385       const bool is64Bit = STI.is64Bit();
9386       const X86InstrInfo *TII = STI.getInstrInfo();
9387 
9388       // Insert a Copy from TLSBaseAddrReg to RAX/EAX.
9389       MachineInstr *Copy =
9390           BuildMI(*I.getParent(), I, I.getDebugLoc(),
9391                   TII->get(TargetOpcode::COPY), is64Bit ? X86::RAX : X86::EAX)
9392               .addReg(TLSBaseAddrReg);
9393 
9394       // Erase the TLS_base_addr instruction.
9395       I.eraseFromParent();
9396 
9397       return Copy;
9398     }
9399 
9400     // Create a virtual register in *TLSBaseAddrReg, and populate it by
9401     // inserting a copy instruction after I. Returns the new instruction.
9402     MachineInstr *SetRegister(MachineInstr &I, unsigned *TLSBaseAddrReg) {
9403       MachineFunction *MF = I.getParent()->getParent();
9404       const X86Subtarget &STI = MF->getSubtarget<X86Subtarget>();
9405       const bool is64Bit = STI.is64Bit();
9406       const X86InstrInfo *TII = STI.getInstrInfo();
9407 
9408       // Create a virtual register for the TLS base address.
9409       MachineRegisterInfo &RegInfo = MF->getRegInfo();
9410       *TLSBaseAddrReg = RegInfo.createVirtualRegister(is64Bit
9411                                                       ? &X86::GR64RegClass
9412                                                       : &X86::GR32RegClass);
9413 
9414       // Insert a copy from RAX/EAX to TLSBaseAddrReg.
9415       MachineInstr *Next = I.getNextNode();
9416       MachineInstr *Copy =
9417           BuildMI(*I.getParent(), Next, I.getDebugLoc(),
9418                   TII->get(TargetOpcode::COPY), *TLSBaseAddrReg)
9419               .addReg(is64Bit ? X86::RAX : X86::EAX);
9420 
9421       return Copy;
9422     }
9423 
9424     StringRef getPassName() const override {
9425       return "Local Dynamic TLS Access Clean-up";
9426     }
9427 
9428     void getAnalysisUsage(AnalysisUsage &AU) const override {
9429       AU.setPreservesCFG();
9430       AU.addRequired<MachineDominatorTree>();
9431       MachineFunctionPass::getAnalysisUsage(AU);
9432     }
9433   };
9434 }
9435 
9436 char LDTLSCleanup::ID = 0;
9437 FunctionPass*
9438 llvm::createCleanupLocalDynamicTLSPass() { return new LDTLSCleanup(); }
9439 
9440 /// Constants defining how certain sequences should be outlined.
9441 ///
9442 /// \p MachineOutlinerDefault implies that the function is called with a call
9443 /// instruction, and a return must be emitted for the outlined function frame.
9444 ///
9445 /// That is,
9446 ///
9447 /// I1                                 OUTLINED_FUNCTION:
9448 /// I2 --> call OUTLINED_FUNCTION       I1
9449 /// I3                                  I2
9450 ///                                     I3
9451 ///                                     ret
9452 ///
9453 /// * Call construction overhead: 1 (call instruction)
9454 /// * Frame construction overhead: 1 (return instruction)
9455 ///
9456 /// \p MachineOutlinerTailCall implies that the function is being tail called.
9457 /// A jump is emitted instead of a call, and the return is already present in
9458 /// the outlined sequence. That is,
9459 ///
9460 /// I1                                 OUTLINED_FUNCTION:
9461 /// I2 --> jmp OUTLINED_FUNCTION       I1
9462 /// ret                                I2
9463 ///                                    ret
9464 ///
9465 /// * Call construction overhead: 1 (jump instruction)
9466 /// * Frame construction overhead: 0 (don't need to return)
9467 ///
9468 enum MachineOutlinerClass {
9469   MachineOutlinerDefault,
9470   MachineOutlinerTailCall
9471 };
9472 
9473 outliner::OutlinedFunction X86InstrInfo::getOutliningCandidateInfo(
9474     std::vector<outliner::Candidate> &RepeatedSequenceLocs) const {
9475   unsigned SequenceSize =
9476       std::accumulate(RepeatedSequenceLocs[0].front(),
9477                       std::next(RepeatedSequenceLocs[0].back()), 0,
9478                       [](unsigned Sum, const MachineInstr &MI) {
9479                         // FIXME: x86 doesn't implement getInstSizeInBytes, so
9480                         // we can't tell the cost.  Just assume each instruction
9481                         // is one byte.
9482                         if (MI.isDebugInstr() || MI.isKill())
9483                           return Sum;
9484                         return Sum + 1;
9485                       });
9486 
9487   // We check to see if CFI Instructions are present, and if they are
9488   // we find the number of CFI Instructions in the candidates.
9489   unsigned CFICount = 0;
9490   for (auto &I : make_range(RepeatedSequenceLocs[0].front(),
9491                             std::next(RepeatedSequenceLocs[0].back()))) {
9492     if (I.isCFIInstruction())
9493       CFICount++;
9494   }
9495 
9496   // We compare the number of found CFI Instructions to  the number of CFI
9497   // instructions in the parent function for each candidate.  We must check this
9498   // since if we outline one of the CFI instructions in a function, we have to
9499   // outline them all for correctness. If we do not, the address offsets will be
9500   // incorrect between the two sections of the program.
9501   for (outliner::Candidate &C : RepeatedSequenceLocs) {
9502     std::vector<MCCFIInstruction> CFIInstructions =
9503         C.getMF()->getFrameInstructions();
9504 
9505     if (CFICount > 0 && CFICount != CFIInstructions.size())
9506       return outliner::OutlinedFunction();
9507   }
9508 
9509   // FIXME: Use real size in bytes for call and ret instructions.
9510   if (RepeatedSequenceLocs[0].back()->isTerminator()) {
9511     for (outliner::Candidate &C : RepeatedSequenceLocs)
9512       C.setCallInfo(MachineOutlinerTailCall, 1);
9513 
9514     return outliner::OutlinedFunction(RepeatedSequenceLocs, SequenceSize,
9515                                       0, // Number of bytes to emit frame.
9516                                       MachineOutlinerTailCall // Type of frame.
9517     );
9518   }
9519 
9520   if (CFICount > 0)
9521     return outliner::OutlinedFunction();
9522 
9523   for (outliner::Candidate &C : RepeatedSequenceLocs)
9524     C.setCallInfo(MachineOutlinerDefault, 1);
9525 
9526   return outliner::OutlinedFunction(RepeatedSequenceLocs, SequenceSize, 1,
9527                                     MachineOutlinerDefault);
9528 }
9529 
9530 bool X86InstrInfo::isFunctionSafeToOutlineFrom(MachineFunction &MF,
9531                                            bool OutlineFromLinkOnceODRs) const {
9532   const Function &F = MF.getFunction();
9533 
9534   // Does the function use a red zone? If it does, then we can't risk messing
9535   // with the stack.
9536   if (Subtarget.getFrameLowering()->has128ByteRedZone(MF)) {
9537     // It could have a red zone. If it does, then we don't want to touch it.
9538     const X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
9539     if (!X86FI || X86FI->getUsesRedZone())
9540       return false;
9541   }
9542 
9543   // If we *don't* want to outline from things that could potentially be deduped
9544   // then return false.
9545   if (!OutlineFromLinkOnceODRs && F.hasLinkOnceODRLinkage())
9546       return false;
9547 
9548   // This function is viable for outlining, so return true.
9549   return true;
9550 }
9551 
9552 outliner::InstrType
9553 X86InstrInfo::getOutliningType(MachineBasicBlock::iterator &MIT,  unsigned Flags) const {
9554   MachineInstr &MI = *MIT;
9555   // Don't allow debug values to impact outlining type.
9556   if (MI.isDebugInstr() || MI.isIndirectDebugValue())
9557     return outliner::InstrType::Invisible;
9558 
9559   // At this point, KILL instructions don't really tell us much so we can go
9560   // ahead and skip over them.
9561   if (MI.isKill())
9562     return outliner::InstrType::Invisible;
9563 
9564   // Is this a tail call? If yes, we can outline as a tail call.
9565   if (isTailCall(MI))
9566     return outliner::InstrType::Legal;
9567 
9568   // Is this the terminator of a basic block?
9569   if (MI.isTerminator() || MI.isReturn()) {
9570 
9571     // Does its parent have any successors in its MachineFunction?
9572     if (MI.getParent()->succ_empty())
9573       return outliner::InstrType::Legal;
9574 
9575     // It does, so we can't tail call it.
9576     return outliner::InstrType::Illegal;
9577   }
9578 
9579   // Don't outline anything that modifies or reads from the stack pointer.
9580   //
9581   // FIXME: There are instructions which are being manually built without
9582   // explicit uses/defs so we also have to check the MCInstrDesc. We should be
9583   // able to remove the extra checks once those are fixed up. For example,
9584   // sometimes we might get something like %rax = POP64r 1. This won't be
9585   // caught by modifiesRegister or readsRegister even though the instruction
9586   // really ought to be formed so that modifiesRegister/readsRegister would
9587   // catch it.
9588   if (MI.modifiesRegister(X86::RSP, &RI) || MI.readsRegister(X86::RSP, &RI) ||
9589       MI.getDesc().hasImplicitUseOfPhysReg(X86::RSP) ||
9590       MI.getDesc().hasImplicitDefOfPhysReg(X86::RSP))
9591     return outliner::InstrType::Illegal;
9592 
9593   // Outlined calls change the instruction pointer, so don't read from it.
9594   if (MI.readsRegister(X86::RIP, &RI) ||
9595       MI.getDesc().hasImplicitUseOfPhysReg(X86::RIP) ||
9596       MI.getDesc().hasImplicitDefOfPhysReg(X86::RIP))
9597     return outliner::InstrType::Illegal;
9598 
9599   // Positions can't safely be outlined.
9600   if (MI.isPosition())
9601     return outliner::InstrType::Illegal;
9602 
9603   // Make sure none of the operands of this instruction do anything tricky.
9604   for (const MachineOperand &MOP : MI.operands())
9605     if (MOP.isCPI() || MOP.isJTI() || MOP.isCFIIndex() || MOP.isFI() ||
9606         MOP.isTargetIndex())
9607       return outliner::InstrType::Illegal;
9608 
9609   return outliner::InstrType::Legal;
9610 }
9611 
9612 void X86InstrInfo::buildOutlinedFrame(MachineBasicBlock &MBB,
9613                                           MachineFunction &MF,
9614                                           const outliner::OutlinedFunction &OF)
9615                                           const {
9616   // If we're a tail call, we already have a return, so don't do anything.
9617   if (OF.FrameConstructionID == MachineOutlinerTailCall)
9618     return;
9619 
9620   // We're a normal call, so our sequence doesn't have a return instruction.
9621   // Add it in.
9622   MachineInstr *retq = BuildMI(MF, DebugLoc(), get(X86::RET64));
9623   MBB.insert(MBB.end(), retq);
9624 }
9625 
9626 MachineBasicBlock::iterator
9627 X86InstrInfo::insertOutlinedCall(Module &M, MachineBasicBlock &MBB,
9628                                  MachineBasicBlock::iterator &It,
9629                                  MachineFunction &MF,
9630                                  outliner::Candidate &C) const {
9631   // Is it a tail call?
9632   if (C.CallConstructionID == MachineOutlinerTailCall) {
9633     // Yes, just insert a JMP.
9634     It = MBB.insert(It,
9635                   BuildMI(MF, DebugLoc(), get(X86::TAILJMPd64))
9636                       .addGlobalAddress(M.getNamedValue(MF.getName())));
9637   } else {
9638     // No, insert a call.
9639     It = MBB.insert(It,
9640                   BuildMI(MF, DebugLoc(), get(X86::CALL64pcrel32))
9641                       .addGlobalAddress(M.getNamedValue(MF.getName())));
9642   }
9643 
9644   return It;
9645 }
9646 
9647 #define GET_INSTRINFO_HELPERS
9648 #include "X86GenInstrInfo.inc"
9649