1 //===-- X86InstrInfo.cpp - X86 Instruction Information --------------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file contains the X86 implementation of the TargetInstrInfo class.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "X86InstrInfo.h"
15 #include "X86.h"
16 #include "X86InstrBuilder.h"
17 #include "X86InstrFoldTables.h"
18 #include "X86MachineFunctionInfo.h"
19 #include "X86Subtarget.h"
20 #include "X86TargetMachine.h"
21 #include "llvm/ADT/STLExtras.h"
22 #include "llvm/ADT/Sequence.h"
23 #include "llvm/CodeGen/LivePhysRegs.h"
24 #include "llvm/CodeGen/LiveVariables.h"
25 #include "llvm/CodeGen/MachineConstantPool.h"
26 #include "llvm/CodeGen/MachineDominators.h"
27 #include "llvm/CodeGen/MachineFrameInfo.h"
28 #include "llvm/CodeGen/MachineInstrBuilder.h"
29 #include "llvm/CodeGen/MachineModuleInfo.h"
30 #include "llvm/CodeGen/MachineRegisterInfo.h"
31 #include "llvm/CodeGen/StackMaps.h"
32 #include "llvm/IR/DerivedTypes.h"
33 #include "llvm/IR/Function.h"
34 #include "llvm/IR/LLVMContext.h"
35 #include "llvm/MC/MCAsmInfo.h"
36 #include "llvm/MC/MCExpr.h"
37 #include "llvm/MC/MCInst.h"
38 #include "llvm/Support/CommandLine.h"
39 #include "llvm/Support/Debug.h"
40 #include "llvm/Support/ErrorHandling.h"
41 #include "llvm/Support/raw_ostream.h"
42 #include "llvm/Target/TargetOptions.h"
43 
44 using namespace llvm;
45 
46 #define DEBUG_TYPE "x86-instr-info"
47 
48 #define GET_INSTRINFO_CTOR_DTOR
49 #include "X86GenInstrInfo.inc"
50 
51 static cl::opt<bool>
52     NoFusing("disable-spill-fusing",
53              cl::desc("Disable fusing of spill code into instructions"),
54              cl::Hidden);
55 static cl::opt<bool>
56 PrintFailedFusing("print-failed-fuse-candidates",
57                   cl::desc("Print instructions that the allocator wants to"
58                            " fuse, but the X86 backend currently can't"),
59                   cl::Hidden);
60 static cl::opt<bool>
61 ReMatPICStubLoad("remat-pic-stub-load",
62                  cl::desc("Re-materialize load from stub in PIC mode"),
63                  cl::init(false), cl::Hidden);
64 static cl::opt<unsigned>
65 PartialRegUpdateClearance("partial-reg-update-clearance",
66                           cl::desc("Clearance between two register writes "
67                                    "for inserting XOR to avoid partial "
68                                    "register update"),
69                           cl::init(64), cl::Hidden);
70 static cl::opt<unsigned>
71 UndefRegClearance("undef-reg-clearance",
72                   cl::desc("How many idle instructions we would like before "
73                            "certain undef register reads"),
74                   cl::init(128), cl::Hidden);
75 
76 
77 // Pin the vtable to this file.
anchor()78 void X86InstrInfo::anchor() {}
79 
X86InstrInfo(X86Subtarget & STI)80 X86InstrInfo::X86InstrInfo(X86Subtarget &STI)
81     : X86GenInstrInfo((STI.isTarget64BitLP64() ? X86::ADJCALLSTACKDOWN64
82                                                : X86::ADJCALLSTACKDOWN32),
83                       (STI.isTarget64BitLP64() ? X86::ADJCALLSTACKUP64
84                                                : X86::ADJCALLSTACKUP32),
85                       X86::CATCHRET,
86                       (STI.is64Bit() ? X86::RETQ : X86::RETL)),
87       Subtarget(STI), RI(STI.getTargetTriple()) {
88 }
89 
90 bool
isCoalescableExtInstr(const MachineInstr & MI,unsigned & SrcReg,unsigned & DstReg,unsigned & SubIdx) const91 X86InstrInfo::isCoalescableExtInstr(const MachineInstr &MI,
92                                     unsigned &SrcReg, unsigned &DstReg,
93                                     unsigned &SubIdx) const {
94   switch (MI.getOpcode()) {
95   default: break;
96   case X86::MOVSX16rr8:
97   case X86::MOVZX16rr8:
98   case X86::MOVSX32rr8:
99   case X86::MOVZX32rr8:
100   case X86::MOVSX64rr8:
101     if (!Subtarget.is64Bit())
102       // It's not always legal to reference the low 8-bit of the larger
103       // register in 32-bit mode.
104       return false;
105     LLVM_FALLTHROUGH;
106   case X86::MOVSX32rr16:
107   case X86::MOVZX32rr16:
108   case X86::MOVSX64rr16:
109   case X86::MOVSX64rr32: {
110     if (MI.getOperand(0).getSubReg() || MI.getOperand(1).getSubReg())
111       // Be conservative.
112       return false;
113     SrcReg = MI.getOperand(1).getReg();
114     DstReg = MI.getOperand(0).getReg();
115     switch (MI.getOpcode()) {
116     default: llvm_unreachable("Unreachable!");
117     case X86::MOVSX16rr8:
118     case X86::MOVZX16rr8:
119     case X86::MOVSX32rr8:
120     case X86::MOVZX32rr8:
121     case X86::MOVSX64rr8:
122       SubIdx = X86::sub_8bit;
123       break;
124     case X86::MOVSX32rr16:
125     case X86::MOVZX32rr16:
126     case X86::MOVSX64rr16:
127       SubIdx = X86::sub_16bit;
128       break;
129     case X86::MOVSX64rr32:
130       SubIdx = X86::sub_32bit;
131       break;
132     }
133     return true;
134   }
135   }
136   return false;
137 }
138 
getSPAdjust(const MachineInstr & MI) const139 int X86InstrInfo::getSPAdjust(const MachineInstr &MI) const {
140   const MachineFunction *MF = MI.getParent()->getParent();
141   const TargetFrameLowering *TFI = MF->getSubtarget().getFrameLowering();
142 
143   if (isFrameInstr(MI)) {
144     unsigned StackAlign = TFI->getStackAlignment();
145     int SPAdj = alignTo(getFrameSize(MI), StackAlign);
146     SPAdj -= getFrameAdjustment(MI);
147     if (!isFrameSetup(MI))
148       SPAdj = -SPAdj;
149     return SPAdj;
150   }
151 
152   // To know whether a call adjusts the stack, we need information
153   // that is bound to the following ADJCALLSTACKUP pseudo.
154   // Look for the next ADJCALLSTACKUP that follows the call.
155   if (MI.isCall()) {
156     const MachineBasicBlock *MBB = MI.getParent();
157     auto I = ++MachineBasicBlock::const_iterator(MI);
158     for (auto E = MBB->end(); I != E; ++I) {
159       if (I->getOpcode() == getCallFrameDestroyOpcode() ||
160           I->isCall())
161         break;
162     }
163 
164     // If we could not find a frame destroy opcode, then it has already
165     // been simplified, so we don't care.
166     if (I->getOpcode() != getCallFrameDestroyOpcode())
167       return 0;
168 
169     return -(I->getOperand(1).getImm());
170   }
171 
172   // Currently handle only PUSHes we can reasonably expect to see
173   // in call sequences
174   switch (MI.getOpcode()) {
175   default:
176     return 0;
177   case X86::PUSH32i8:
178   case X86::PUSH32r:
179   case X86::PUSH32rmm:
180   case X86::PUSH32rmr:
181   case X86::PUSHi32:
182     return 4;
183   case X86::PUSH64i8:
184   case X86::PUSH64r:
185   case X86::PUSH64rmm:
186   case X86::PUSH64rmr:
187   case X86::PUSH64i32:
188     return 8;
189   }
190 }
191 
192 /// Return true and the FrameIndex if the specified
193 /// operand and follow operands form a reference to the stack frame.
isFrameOperand(const MachineInstr & MI,unsigned int Op,int & FrameIndex) const194 bool X86InstrInfo::isFrameOperand(const MachineInstr &MI, unsigned int Op,
195                                   int &FrameIndex) const {
196   if (MI.getOperand(Op + X86::AddrBaseReg).isFI() &&
197       MI.getOperand(Op + X86::AddrScaleAmt).isImm() &&
198       MI.getOperand(Op + X86::AddrIndexReg).isReg() &&
199       MI.getOperand(Op + X86::AddrDisp).isImm() &&
200       MI.getOperand(Op + X86::AddrScaleAmt).getImm() == 1 &&
201       MI.getOperand(Op + X86::AddrIndexReg).getReg() == 0 &&
202       MI.getOperand(Op + X86::AddrDisp).getImm() == 0) {
203     FrameIndex = MI.getOperand(Op + X86::AddrBaseReg).getIndex();
204     return true;
205   }
206   return false;
207 }
208 
isFrameLoadOpcode(int Opcode,unsigned & MemBytes)209 static bool isFrameLoadOpcode(int Opcode, unsigned &MemBytes) {
210   switch (Opcode) {
211   default:
212     return false;
213   case X86::MOV8rm:
214   case X86::KMOVBkm:
215     MemBytes = 1;
216     return true;
217   case X86::MOV16rm:
218   case X86::KMOVWkm:
219     MemBytes = 2;
220     return true;
221   case X86::MOV32rm:
222   case X86::MOVSSrm:
223   case X86::VMOVSSZrm:
224   case X86::VMOVSSrm:
225   case X86::KMOVDkm:
226     MemBytes = 4;
227     return true;
228   case X86::MOV64rm:
229   case X86::LD_Fp64m:
230   case X86::MOVSDrm:
231   case X86::VMOVSDrm:
232   case X86::VMOVSDZrm:
233   case X86::MMX_MOVD64rm:
234   case X86::MMX_MOVQ64rm:
235   case X86::KMOVQkm:
236     MemBytes = 8;
237     return true;
238   case X86::MOVAPSrm:
239   case X86::MOVUPSrm:
240   case X86::MOVAPDrm:
241   case X86::MOVUPDrm:
242   case X86::MOVDQArm:
243   case X86::MOVDQUrm:
244   case X86::VMOVAPSrm:
245   case X86::VMOVUPSrm:
246   case X86::VMOVAPDrm:
247   case X86::VMOVUPDrm:
248   case X86::VMOVDQArm:
249   case X86::VMOVDQUrm:
250   case X86::VMOVAPSZ128rm:
251   case X86::VMOVUPSZ128rm:
252   case X86::VMOVAPSZ128rm_NOVLX:
253   case X86::VMOVUPSZ128rm_NOVLX:
254   case X86::VMOVAPDZ128rm:
255   case X86::VMOVUPDZ128rm:
256   case X86::VMOVDQU8Z128rm:
257   case X86::VMOVDQU16Z128rm:
258   case X86::VMOVDQA32Z128rm:
259   case X86::VMOVDQU32Z128rm:
260   case X86::VMOVDQA64Z128rm:
261   case X86::VMOVDQU64Z128rm:
262     MemBytes = 16;
263     return true;
264   case X86::VMOVAPSYrm:
265   case X86::VMOVUPSYrm:
266   case X86::VMOVAPDYrm:
267   case X86::VMOVUPDYrm:
268   case X86::VMOVDQAYrm:
269   case X86::VMOVDQUYrm:
270   case X86::VMOVAPSZ256rm:
271   case X86::VMOVUPSZ256rm:
272   case X86::VMOVAPSZ256rm_NOVLX:
273   case X86::VMOVUPSZ256rm_NOVLX:
274   case X86::VMOVAPDZ256rm:
275   case X86::VMOVUPDZ256rm:
276   case X86::VMOVDQU8Z256rm:
277   case X86::VMOVDQU16Z256rm:
278   case X86::VMOVDQA32Z256rm:
279   case X86::VMOVDQU32Z256rm:
280   case X86::VMOVDQA64Z256rm:
281   case X86::VMOVDQU64Z256rm:
282     MemBytes = 32;
283     return true;
284   case X86::VMOVAPSZrm:
285   case X86::VMOVUPSZrm:
286   case X86::VMOVAPDZrm:
287   case X86::VMOVUPDZrm:
288   case X86::VMOVDQU8Zrm:
289   case X86::VMOVDQU16Zrm:
290   case X86::VMOVDQA32Zrm:
291   case X86::VMOVDQU32Zrm:
292   case X86::VMOVDQA64Zrm:
293   case X86::VMOVDQU64Zrm:
294     MemBytes = 64;
295     return true;
296   }
297 }
298 
isFrameStoreOpcode(int Opcode,unsigned & MemBytes)299 static bool isFrameStoreOpcode(int Opcode, unsigned &MemBytes) {
300   switch (Opcode) {
301   default:
302     return false;
303   case X86::MOV8mr:
304   case X86::KMOVBmk:
305     MemBytes = 1;
306     return true;
307   case X86::MOV16mr:
308   case X86::KMOVWmk:
309     MemBytes = 2;
310     return true;
311   case X86::MOV32mr:
312   case X86::MOVSSmr:
313   case X86::VMOVSSmr:
314   case X86::VMOVSSZmr:
315   case X86::KMOVDmk:
316     MemBytes = 4;
317     return true;
318   case X86::MOV64mr:
319   case X86::ST_FpP64m:
320   case X86::MOVSDmr:
321   case X86::VMOVSDmr:
322   case X86::VMOVSDZmr:
323   case X86::MMX_MOVD64mr:
324   case X86::MMX_MOVQ64mr:
325   case X86::MMX_MOVNTQmr:
326   case X86::KMOVQmk:
327     MemBytes = 8;
328     return true;
329   case X86::MOVAPSmr:
330   case X86::MOVUPSmr:
331   case X86::MOVAPDmr:
332   case X86::MOVUPDmr:
333   case X86::MOVDQAmr:
334   case X86::MOVDQUmr:
335   case X86::VMOVAPSmr:
336   case X86::VMOVUPSmr:
337   case X86::VMOVAPDmr:
338   case X86::VMOVUPDmr:
339   case X86::VMOVDQAmr:
340   case X86::VMOVDQUmr:
341   case X86::VMOVUPSZ128mr:
342   case X86::VMOVAPSZ128mr:
343   case X86::VMOVUPSZ128mr_NOVLX:
344   case X86::VMOVAPSZ128mr_NOVLX:
345   case X86::VMOVUPDZ128mr:
346   case X86::VMOVAPDZ128mr:
347   case X86::VMOVDQA32Z128mr:
348   case X86::VMOVDQU32Z128mr:
349   case X86::VMOVDQA64Z128mr:
350   case X86::VMOVDQU64Z128mr:
351   case X86::VMOVDQU8Z128mr:
352   case X86::VMOVDQU16Z128mr:
353     MemBytes = 16;
354     return true;
355   case X86::VMOVUPSYmr:
356   case X86::VMOVAPSYmr:
357   case X86::VMOVUPDYmr:
358   case X86::VMOVAPDYmr:
359   case X86::VMOVDQUYmr:
360   case X86::VMOVDQAYmr:
361   case X86::VMOVUPSZ256mr:
362   case X86::VMOVAPSZ256mr:
363   case X86::VMOVUPSZ256mr_NOVLX:
364   case X86::VMOVAPSZ256mr_NOVLX:
365   case X86::VMOVUPDZ256mr:
366   case X86::VMOVAPDZ256mr:
367   case X86::VMOVDQU8Z256mr:
368   case X86::VMOVDQU16Z256mr:
369   case X86::VMOVDQA32Z256mr:
370   case X86::VMOVDQU32Z256mr:
371   case X86::VMOVDQA64Z256mr:
372   case X86::VMOVDQU64Z256mr:
373     MemBytes = 32;
374     return true;
375   case X86::VMOVUPSZmr:
376   case X86::VMOVAPSZmr:
377   case X86::VMOVUPDZmr:
378   case X86::VMOVAPDZmr:
379   case X86::VMOVDQU8Zmr:
380   case X86::VMOVDQU16Zmr:
381   case X86::VMOVDQA32Zmr:
382   case X86::VMOVDQU32Zmr:
383   case X86::VMOVDQA64Zmr:
384   case X86::VMOVDQU64Zmr:
385     MemBytes = 64;
386     return true;
387   }
388   return false;
389 }
390 
isLoadFromStackSlot(const MachineInstr & MI,int & FrameIndex) const391 unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr &MI,
392                                            int &FrameIndex) const {
393   unsigned Dummy;
394   return X86InstrInfo::isLoadFromStackSlot(MI, FrameIndex, Dummy);
395 }
396 
isLoadFromStackSlot(const MachineInstr & MI,int & FrameIndex,unsigned & MemBytes) const397 unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr &MI,
398                                            int &FrameIndex,
399                                            unsigned &MemBytes) const {
400   if (isFrameLoadOpcode(MI.getOpcode(), MemBytes))
401     if (MI.getOperand(0).getSubReg() == 0 && isFrameOperand(MI, 1, FrameIndex))
402       return MI.getOperand(0).getReg();
403   return 0;
404 }
405 
isLoadFromStackSlotPostFE(const MachineInstr & MI,int & FrameIndex) const406 unsigned X86InstrInfo::isLoadFromStackSlotPostFE(const MachineInstr &MI,
407                                                  int &FrameIndex) const {
408   unsigned Dummy;
409   if (isFrameLoadOpcode(MI.getOpcode(), Dummy)) {
410     unsigned Reg;
411     if ((Reg = isLoadFromStackSlot(MI, FrameIndex)))
412       return Reg;
413     // Check for post-frame index elimination operations
414     SmallVector<const MachineMemOperand *, 1> Accesses;
415     if (hasLoadFromStackSlot(MI, Accesses)) {
416       FrameIndex =
417           cast<FixedStackPseudoSourceValue>(Accesses.front()->getPseudoValue())
418               ->getFrameIndex();
419       return 1;
420     }
421   }
422   return 0;
423 }
424 
isStoreToStackSlot(const MachineInstr & MI,int & FrameIndex) const425 unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr &MI,
426                                           int &FrameIndex) const {
427   unsigned Dummy;
428   return X86InstrInfo::isStoreToStackSlot(MI, FrameIndex, Dummy);
429 }
430 
isStoreToStackSlot(const MachineInstr & MI,int & FrameIndex,unsigned & MemBytes) const431 unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr &MI,
432                                           int &FrameIndex,
433                                           unsigned &MemBytes) const {
434   if (isFrameStoreOpcode(MI.getOpcode(), MemBytes))
435     if (MI.getOperand(X86::AddrNumOperands).getSubReg() == 0 &&
436         isFrameOperand(MI, 0, FrameIndex))
437       return MI.getOperand(X86::AddrNumOperands).getReg();
438   return 0;
439 }
440 
isStoreToStackSlotPostFE(const MachineInstr & MI,int & FrameIndex) const441 unsigned X86InstrInfo::isStoreToStackSlotPostFE(const MachineInstr &MI,
442                                                 int &FrameIndex) const {
443   unsigned Dummy;
444   if (isFrameStoreOpcode(MI.getOpcode(), Dummy)) {
445     unsigned Reg;
446     if ((Reg = isStoreToStackSlot(MI, FrameIndex)))
447       return Reg;
448     // Check for post-frame index elimination operations
449     SmallVector<const MachineMemOperand *, 1> Accesses;
450     if (hasStoreToStackSlot(MI, Accesses)) {
451       FrameIndex =
452           cast<FixedStackPseudoSourceValue>(Accesses.front()->getPseudoValue())
453               ->getFrameIndex();
454       return 1;
455     }
456   }
457   return 0;
458 }
459 
460 /// Return true if register is PIC base; i.e.g defined by X86::MOVPC32r.
regIsPICBase(unsigned BaseReg,const MachineRegisterInfo & MRI)461 static bool regIsPICBase(unsigned BaseReg, const MachineRegisterInfo &MRI) {
462   // Don't waste compile time scanning use-def chains of physregs.
463   if (!TargetRegisterInfo::isVirtualRegister(BaseReg))
464     return false;
465   bool isPICBase = false;
466   for (MachineRegisterInfo::def_instr_iterator I = MRI.def_instr_begin(BaseReg),
467          E = MRI.def_instr_end(); I != E; ++I) {
468     MachineInstr *DefMI = &*I;
469     if (DefMI->getOpcode() != X86::MOVPC32r)
470       return false;
471     assert(!isPICBase && "More than one PIC base?");
472     isPICBase = true;
473   }
474   return isPICBase;
475 }
476 
isReallyTriviallyReMaterializable(const MachineInstr & MI,AliasAnalysis * AA) const477 bool X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr &MI,
478                                                      AliasAnalysis *AA) const {
479   switch (MI.getOpcode()) {
480   default: break;
481   case X86::MOV8rm:
482   case X86::MOV8rm_NOREX:
483   case X86::MOV16rm:
484   case X86::MOV32rm:
485   case X86::MOV64rm:
486   case X86::LD_Fp64m:
487   case X86::MOVSSrm:
488   case X86::MOVSDrm:
489   case X86::MOVAPSrm:
490   case X86::MOVUPSrm:
491   case X86::MOVAPDrm:
492   case X86::MOVUPDrm:
493   case X86::MOVDQArm:
494   case X86::MOVDQUrm:
495   case X86::VMOVSSrm:
496   case X86::VMOVSDrm:
497   case X86::VMOVAPSrm:
498   case X86::VMOVUPSrm:
499   case X86::VMOVAPDrm:
500   case X86::VMOVUPDrm:
501   case X86::VMOVDQArm:
502   case X86::VMOVDQUrm:
503   case X86::VMOVAPSYrm:
504   case X86::VMOVUPSYrm:
505   case X86::VMOVAPDYrm:
506   case X86::VMOVUPDYrm:
507   case X86::VMOVDQAYrm:
508   case X86::VMOVDQUYrm:
509   case X86::MMX_MOVD64rm:
510   case X86::MMX_MOVQ64rm:
511   // AVX-512
512   case X86::VMOVSSZrm:
513   case X86::VMOVSDZrm:
514   case X86::VMOVAPDZ128rm:
515   case X86::VMOVAPDZ256rm:
516   case X86::VMOVAPDZrm:
517   case X86::VMOVAPSZ128rm:
518   case X86::VMOVAPSZ256rm:
519   case X86::VMOVAPSZ128rm_NOVLX:
520   case X86::VMOVAPSZ256rm_NOVLX:
521   case X86::VMOVAPSZrm:
522   case X86::VMOVDQA32Z128rm:
523   case X86::VMOVDQA32Z256rm:
524   case X86::VMOVDQA32Zrm:
525   case X86::VMOVDQA64Z128rm:
526   case X86::VMOVDQA64Z256rm:
527   case X86::VMOVDQA64Zrm:
528   case X86::VMOVDQU16Z128rm:
529   case X86::VMOVDQU16Z256rm:
530   case X86::VMOVDQU16Zrm:
531   case X86::VMOVDQU32Z128rm:
532   case X86::VMOVDQU32Z256rm:
533   case X86::VMOVDQU32Zrm:
534   case X86::VMOVDQU64Z128rm:
535   case X86::VMOVDQU64Z256rm:
536   case X86::VMOVDQU64Zrm:
537   case X86::VMOVDQU8Z128rm:
538   case X86::VMOVDQU8Z256rm:
539   case X86::VMOVDQU8Zrm:
540   case X86::VMOVUPDZ128rm:
541   case X86::VMOVUPDZ256rm:
542   case X86::VMOVUPDZrm:
543   case X86::VMOVUPSZ128rm:
544   case X86::VMOVUPSZ256rm:
545   case X86::VMOVUPSZ128rm_NOVLX:
546   case X86::VMOVUPSZ256rm_NOVLX:
547   case X86::VMOVUPSZrm: {
548     // Loads from constant pools are trivially rematerializable.
549     if (MI.getOperand(1 + X86::AddrBaseReg).isReg() &&
550         MI.getOperand(1 + X86::AddrScaleAmt).isImm() &&
551         MI.getOperand(1 + X86::AddrIndexReg).isReg() &&
552         MI.getOperand(1 + X86::AddrIndexReg).getReg() == 0 &&
553         MI.isDereferenceableInvariantLoad(AA)) {
554       unsigned BaseReg = MI.getOperand(1 + X86::AddrBaseReg).getReg();
555       if (BaseReg == 0 || BaseReg == X86::RIP)
556         return true;
557       // Allow re-materialization of PIC load.
558       if (!ReMatPICStubLoad && MI.getOperand(1 + X86::AddrDisp).isGlobal())
559         return false;
560       const MachineFunction &MF = *MI.getParent()->getParent();
561       const MachineRegisterInfo &MRI = MF.getRegInfo();
562       return regIsPICBase(BaseReg, MRI);
563     }
564     return false;
565   }
566 
567   case X86::LEA32r:
568   case X86::LEA64r: {
569     if (MI.getOperand(1 + X86::AddrScaleAmt).isImm() &&
570         MI.getOperand(1 + X86::AddrIndexReg).isReg() &&
571         MI.getOperand(1 + X86::AddrIndexReg).getReg() == 0 &&
572         !MI.getOperand(1 + X86::AddrDisp).isReg()) {
573       // lea fi#, lea GV, etc. are all rematerializable.
574       if (!MI.getOperand(1 + X86::AddrBaseReg).isReg())
575         return true;
576       unsigned BaseReg = MI.getOperand(1 + X86::AddrBaseReg).getReg();
577       if (BaseReg == 0)
578         return true;
579       // Allow re-materialization of lea PICBase + x.
580       const MachineFunction &MF = *MI.getParent()->getParent();
581       const MachineRegisterInfo &MRI = MF.getRegInfo();
582       return regIsPICBase(BaseReg, MRI);
583     }
584     return false;
585   }
586   }
587 
588   // All other instructions marked M_REMATERIALIZABLE are always trivially
589   // rematerializable.
590   return true;
591 }
592 
isSafeToClobberEFLAGS(MachineBasicBlock & MBB,MachineBasicBlock::iterator I) const593 bool X86InstrInfo::isSafeToClobberEFLAGS(MachineBasicBlock &MBB,
594                                          MachineBasicBlock::iterator I) const {
595   MachineBasicBlock::iterator E = MBB.end();
596 
597   // For compile time consideration, if we are not able to determine the
598   // safety after visiting 4 instructions in each direction, we will assume
599   // it's not safe.
600   MachineBasicBlock::iterator Iter = I;
601   for (unsigned i = 0; Iter != E && i < 4; ++i) {
602     bool SeenDef = false;
603     for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
604       MachineOperand &MO = Iter->getOperand(j);
605       if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
606         SeenDef = true;
607       if (!MO.isReg())
608         continue;
609       if (MO.getReg() == X86::EFLAGS) {
610         if (MO.isUse())
611           return false;
612         SeenDef = true;
613       }
614     }
615 
616     if (SeenDef)
617       // This instruction defines EFLAGS, no need to look any further.
618       return true;
619     ++Iter;
620     // Skip over debug instructions.
621     while (Iter != E && Iter->isDebugInstr())
622       ++Iter;
623   }
624 
625   // It is safe to clobber EFLAGS at the end of a block of no successor has it
626   // live in.
627   if (Iter == E) {
628     for (MachineBasicBlock *S : MBB.successors())
629       if (S->isLiveIn(X86::EFLAGS))
630         return false;
631     return true;
632   }
633 
634   MachineBasicBlock::iterator B = MBB.begin();
635   Iter = I;
636   for (unsigned i = 0; i < 4; ++i) {
637     // If we make it to the beginning of the block, it's safe to clobber
638     // EFLAGS iff EFLAGS is not live-in.
639     if (Iter == B)
640       return !MBB.isLiveIn(X86::EFLAGS);
641 
642     --Iter;
643     // Skip over debug instructions.
644     while (Iter != B && Iter->isDebugInstr())
645       --Iter;
646 
647     bool SawKill = false;
648     for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
649       MachineOperand &MO = Iter->getOperand(j);
650       // A register mask may clobber EFLAGS, but we should still look for a
651       // live EFLAGS def.
652       if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
653         SawKill = true;
654       if (MO.isReg() && MO.getReg() == X86::EFLAGS) {
655         if (MO.isDef()) return MO.isDead();
656         if (MO.isKill()) SawKill = true;
657       }
658     }
659 
660     if (SawKill)
661       // This instruction kills EFLAGS and doesn't redefine it, so
662       // there's no need to look further.
663       return true;
664   }
665 
666   // Conservative answer.
667   return false;
668 }
669 
reMaterialize(MachineBasicBlock & MBB,MachineBasicBlock::iterator I,unsigned DestReg,unsigned SubIdx,const MachineInstr & Orig,const TargetRegisterInfo & TRI) const670 void X86InstrInfo::reMaterialize(MachineBasicBlock &MBB,
671                                  MachineBasicBlock::iterator I,
672                                  unsigned DestReg, unsigned SubIdx,
673                                  const MachineInstr &Orig,
674                                  const TargetRegisterInfo &TRI) const {
675   bool ClobbersEFLAGS = false;
676   for (const MachineOperand &MO : Orig.operands()) {
677     if (MO.isReg() && MO.isDef() && MO.getReg() == X86::EFLAGS) {
678       ClobbersEFLAGS = true;
679       break;
680     }
681   }
682 
683   if (ClobbersEFLAGS && !isSafeToClobberEFLAGS(MBB, I)) {
684     // The instruction clobbers EFLAGS. Re-materialize as MOV32ri to avoid side
685     // effects.
686     int Value;
687     switch (Orig.getOpcode()) {
688     case X86::MOV32r0:  Value = 0; break;
689     case X86::MOV32r1:  Value = 1; break;
690     case X86::MOV32r_1: Value = -1; break;
691     default:
692       llvm_unreachable("Unexpected instruction!");
693     }
694 
695     const DebugLoc &DL = Orig.getDebugLoc();
696     BuildMI(MBB, I, DL, get(X86::MOV32ri))
697         .add(Orig.getOperand(0))
698         .addImm(Value);
699   } else {
700     MachineInstr *MI = MBB.getParent()->CloneMachineInstr(&Orig);
701     MBB.insert(I, MI);
702   }
703 
704   MachineInstr &NewMI = *std::prev(I);
705   NewMI.substituteRegister(Orig.getOperand(0).getReg(), DestReg, SubIdx, TRI);
706 }
707 
708 /// True if MI has a condition code def, e.g. EFLAGS, that is not marked dead.
hasLiveCondCodeDef(MachineInstr & MI) const709 bool X86InstrInfo::hasLiveCondCodeDef(MachineInstr &MI) const {
710   for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
711     MachineOperand &MO = MI.getOperand(i);
712     if (MO.isReg() && MO.isDef() &&
713         MO.getReg() == X86::EFLAGS && !MO.isDead()) {
714       return true;
715     }
716   }
717   return false;
718 }
719 
720 /// Check whether the shift count for a machine operand is non-zero.
getTruncatedShiftCount(const MachineInstr & MI,unsigned ShiftAmtOperandIdx)721 inline static unsigned getTruncatedShiftCount(const MachineInstr &MI,
722                                               unsigned ShiftAmtOperandIdx) {
723   // The shift count is six bits with the REX.W prefix and five bits without.
724   unsigned ShiftCountMask = (MI.getDesc().TSFlags & X86II::REX_W) ? 63 : 31;
725   unsigned Imm = MI.getOperand(ShiftAmtOperandIdx).getImm();
726   return Imm & ShiftCountMask;
727 }
728 
729 /// Check whether the given shift count is appropriate
730 /// can be represented by a LEA instruction.
isTruncatedShiftCountForLEA(unsigned ShAmt)731 inline static bool isTruncatedShiftCountForLEA(unsigned ShAmt) {
732   // Left shift instructions can be transformed into load-effective-address
733   // instructions if we can encode them appropriately.
734   // A LEA instruction utilizes a SIB byte to encode its scale factor.
735   // The SIB.scale field is two bits wide which means that we can encode any
736   // shift amount less than 4.
737   return ShAmt < 4 && ShAmt > 0;
738 }
739 
classifyLEAReg(MachineInstr & MI,const MachineOperand & Src,unsigned Opc,bool AllowSP,unsigned & NewSrc,bool & isKill,MachineOperand & ImplicitOp,LiveVariables * LV) const740 bool X86InstrInfo::classifyLEAReg(MachineInstr &MI, const MachineOperand &Src,
741                                   unsigned Opc, bool AllowSP, unsigned &NewSrc,
742                                   bool &isKill, MachineOperand &ImplicitOp,
743                                   LiveVariables *LV) const {
744   MachineFunction &MF = *MI.getParent()->getParent();
745   const TargetRegisterClass *RC;
746   if (AllowSP) {
747     RC = Opc != X86::LEA32r ? &X86::GR64RegClass : &X86::GR32RegClass;
748   } else {
749     RC = Opc != X86::LEA32r ?
750       &X86::GR64_NOSPRegClass : &X86::GR32_NOSPRegClass;
751   }
752   unsigned SrcReg = Src.getReg();
753 
754   // For both LEA64 and LEA32 the register already has essentially the right
755   // type (32-bit or 64-bit) we may just need to forbid SP.
756   if (Opc != X86::LEA64_32r) {
757     NewSrc = SrcReg;
758     isKill = Src.isKill();
759     assert(!Src.isUndef() && "Undef op doesn't need optimization");
760 
761     if (TargetRegisterInfo::isVirtualRegister(NewSrc) &&
762         !MF.getRegInfo().constrainRegClass(NewSrc, RC))
763       return false;
764 
765     return true;
766   }
767 
768   // This is for an LEA64_32r and incoming registers are 32-bit. One way or
769   // another we need to add 64-bit registers to the final MI.
770   if (TargetRegisterInfo::isPhysicalRegister(SrcReg)) {
771     ImplicitOp = Src;
772     ImplicitOp.setImplicit();
773 
774     NewSrc = getX86SubSuperRegister(Src.getReg(), 64);
775     isKill = Src.isKill();
776     assert(!Src.isUndef() && "Undef op doesn't need optimization");
777   } else {
778     // Virtual register of the wrong class, we have to create a temporary 64-bit
779     // vreg to feed into the LEA.
780     NewSrc = MF.getRegInfo().createVirtualRegister(RC);
781     MachineInstr *Copy =
782         BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), get(TargetOpcode::COPY))
783             .addReg(NewSrc, RegState::Define | RegState::Undef, X86::sub_32bit)
784             .add(Src);
785 
786     // Which is obviously going to be dead after we're done with it.
787     isKill = true;
788 
789     if (LV)
790       LV->replaceKillInstruction(SrcReg, MI, *Copy);
791   }
792 
793   // We've set all the parameters without issue.
794   return true;
795 }
796 
convertToThreeAddressWithLEA(unsigned MIOpc,MachineFunction::iterator & MFI,MachineInstr & MI,LiveVariables * LV) const797 MachineInstr *X86InstrInfo::convertToThreeAddressWithLEA(
798     unsigned MIOpc, MachineFunction::iterator &MFI, MachineInstr &MI,
799     LiveVariables *LV) const {
800   // We handle 8-bit adds and various 16-bit opcodes in the switch below.
801   bool Is16BitOp = !(MIOpc == X86::ADD8rr || MIOpc == X86::ADD8ri);
802   MachineRegisterInfo &RegInfo = MFI->getParent()->getRegInfo();
803   assert((!Is16BitOp || RegInfo.getTargetRegisterInfo()->getRegSizeInBits(
804               *RegInfo.getRegClass(MI.getOperand(0).getReg())) == 16) &&
805          "Unexpected type for LEA transform");
806 
807   // TODO: For a 32-bit target, we need to adjust the LEA variables with
808   // something like this:
809   //   Opcode = X86::LEA32r;
810   //   InRegLEA = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
811   //   OutRegLEA =
812   //       Is8BitOp ? RegInfo.createVirtualRegister(&X86::GR32ABCD_RegClass)
813   //                : RegInfo.createVirtualRegister(&X86::GR32RegClass);
814   if (!Subtarget.is64Bit())
815     return nullptr;
816 
817   unsigned Opcode = X86::LEA64_32r;
818   unsigned InRegLEA = RegInfo.createVirtualRegister(&X86::GR64_NOSPRegClass);
819   unsigned OutRegLEA = RegInfo.createVirtualRegister(&X86::GR32RegClass);
820 
821   // Build and insert into an implicit UNDEF value. This is OK because
822   // we will be shifting and then extracting the lower 8/16-bits.
823   // This has the potential to cause partial register stall. e.g.
824   //   movw    (%rbp,%rcx,2), %dx
825   //   leal    -65(%rdx), %esi
826   // But testing has shown this *does* help performance in 64-bit mode (at
827   // least on modern x86 machines).
828   MachineBasicBlock::iterator MBBI = MI.getIterator();
829   unsigned Dest = MI.getOperand(0).getReg();
830   unsigned Src = MI.getOperand(1).getReg();
831   bool IsDead = MI.getOperand(0).isDead();
832   bool IsKill = MI.getOperand(1).isKill();
833   unsigned SubReg = Is16BitOp ? X86::sub_16bit : X86::sub_8bit;
834   assert(!MI.getOperand(1).isUndef() && "Undef op doesn't need optimization");
835   BuildMI(*MFI, MBBI, MI.getDebugLoc(), get(X86::IMPLICIT_DEF), InRegLEA);
836   MachineInstr *InsMI =
837       BuildMI(*MFI, MBBI, MI.getDebugLoc(), get(TargetOpcode::COPY))
838           .addReg(InRegLEA, RegState::Define, SubReg)
839           .addReg(Src, getKillRegState(IsKill));
840 
841   MachineInstrBuilder MIB =
842       BuildMI(*MFI, MBBI, MI.getDebugLoc(), get(Opcode), OutRegLEA);
843   switch (MIOpc) {
844   default: llvm_unreachable("Unreachable!");
845   case X86::SHL16ri: {
846     unsigned ShAmt = MI.getOperand(2).getImm();
847     MIB.addReg(0).addImm(1ULL << ShAmt)
848        .addReg(InRegLEA, RegState::Kill).addImm(0).addReg(0);
849     break;
850   }
851   case X86::INC16r:
852     addRegOffset(MIB, InRegLEA, true, 1);
853     break;
854   case X86::DEC16r:
855     addRegOffset(MIB, InRegLEA, true, -1);
856     break;
857   case X86::ADD8ri:
858   case X86::ADD16ri:
859   case X86::ADD16ri8:
860   case X86::ADD16ri_DB:
861   case X86::ADD16ri8_DB:
862     addRegOffset(MIB, InRegLEA, true, MI.getOperand(2).getImm());
863     break;
864   case X86::ADD8rr:
865   case X86::ADD16rr:
866   case X86::ADD16rr_DB: {
867     unsigned Src2 = MI.getOperand(2).getReg();
868     bool IsKill2 = MI.getOperand(2).isKill();
869     assert(!MI.getOperand(2).isUndef() && "Undef op doesn't need optimization");
870     unsigned InRegLEA2 = 0;
871     MachineInstr *InsMI2 = nullptr;
872     if (Src == Src2) {
873       // ADD8rr/ADD16rr killed %reg1028, %reg1028
874       // just a single insert_subreg.
875       addRegReg(MIB, InRegLEA, true, InRegLEA, false);
876     } else {
877       if (Subtarget.is64Bit())
878         InRegLEA2 = RegInfo.createVirtualRegister(&X86::GR64_NOSPRegClass);
879       else
880         InRegLEA2 = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
881       // Build and insert into an implicit UNDEF value. This is OK because
882       // we will be shifting and then extracting the lower 8/16-bits.
883       BuildMI(*MFI, &*MIB, MI.getDebugLoc(), get(X86::IMPLICIT_DEF), InRegLEA2);
884       InsMI2 = BuildMI(*MFI, &*MIB, MI.getDebugLoc(), get(TargetOpcode::COPY))
885                    .addReg(InRegLEA2, RegState::Define, SubReg)
886                    .addReg(Src2, getKillRegState(IsKill2));
887       addRegReg(MIB, InRegLEA, true, InRegLEA2, true);
888     }
889     if (LV && IsKill2 && InsMI2)
890       LV->replaceKillInstruction(Src2, MI, *InsMI2);
891     break;
892   }
893   }
894 
895   MachineInstr *NewMI = MIB;
896   MachineInstr *ExtMI =
897       BuildMI(*MFI, MBBI, MI.getDebugLoc(), get(TargetOpcode::COPY))
898           .addReg(Dest, RegState::Define | getDeadRegState(IsDead))
899           .addReg(OutRegLEA, RegState::Kill, SubReg);
900 
901   if (LV) {
902     // Update live variables.
903     LV->getVarInfo(InRegLEA).Kills.push_back(NewMI);
904     LV->getVarInfo(OutRegLEA).Kills.push_back(ExtMI);
905     if (IsKill)
906       LV->replaceKillInstruction(Src, MI, *InsMI);
907     if (IsDead)
908       LV->replaceKillInstruction(Dest, MI, *ExtMI);
909   }
910 
911   return ExtMI;
912 }
913 
914 /// This method must be implemented by targets that
915 /// set the M_CONVERTIBLE_TO_3_ADDR flag.  When this flag is set, the target
916 /// may be able to convert a two-address instruction into a true
917 /// three-address instruction on demand.  This allows the X86 target (for
918 /// example) to convert ADD and SHL instructions into LEA instructions if they
919 /// would require register copies due to two-addressness.
920 ///
921 /// This method returns a null pointer if the transformation cannot be
922 /// performed, otherwise it returns the new instruction.
923 ///
924 MachineInstr *
convertToThreeAddress(MachineFunction::iterator & MFI,MachineInstr & MI,LiveVariables * LV) const925 X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
926                                     MachineInstr &MI, LiveVariables *LV) const {
927   // The following opcodes also sets the condition code register(s). Only
928   // convert them to equivalent lea if the condition code register def's
929   // are dead!
930   if (hasLiveCondCodeDef(MI))
931     return nullptr;
932 
933   MachineFunction &MF = *MI.getParent()->getParent();
934   // All instructions input are two-addr instructions.  Get the known operands.
935   const MachineOperand &Dest = MI.getOperand(0);
936   const MachineOperand &Src = MI.getOperand(1);
937 
938   // Ideally, operations with undef should be folded before we get here, but we
939   // can't guarantee it. Bail out because optimizing undefs is a waste of time.
940   // Without this, we have to forward undef state to new register operands to
941   // avoid machine verifier errors.
942   if (Src.isUndef())
943     return nullptr;
944   if (MI.getNumOperands() > 2)
945     if (MI.getOperand(2).isReg() && MI.getOperand(2).isUndef())
946       return nullptr;
947 
948   MachineInstr *NewMI = nullptr;
949   bool Is64Bit = Subtarget.is64Bit();
950 
951   unsigned MIOpc = MI.getOpcode();
952   switch (MIOpc) {
953   default: return nullptr;
954   case X86::SHL64ri: {
955     assert(MI.getNumOperands() >= 3 && "Unknown shift instruction!");
956     unsigned ShAmt = getTruncatedShiftCount(MI, 2);
957     if (!isTruncatedShiftCountForLEA(ShAmt)) return nullptr;
958 
959     // LEA can't handle RSP.
960     if (TargetRegisterInfo::isVirtualRegister(Src.getReg()) &&
961         !MF.getRegInfo().constrainRegClass(Src.getReg(),
962                                            &X86::GR64_NOSPRegClass))
963       return nullptr;
964 
965     NewMI = BuildMI(MF, MI.getDebugLoc(), get(X86::LEA64r))
966                 .add(Dest)
967                 .addReg(0)
968                 .addImm(1ULL << ShAmt)
969                 .add(Src)
970                 .addImm(0)
971                 .addReg(0);
972     break;
973   }
974   case X86::SHL32ri: {
975     assert(MI.getNumOperands() >= 3 && "Unknown shift instruction!");
976     unsigned ShAmt = getTruncatedShiftCount(MI, 2);
977     if (!isTruncatedShiftCountForLEA(ShAmt)) return nullptr;
978 
979     unsigned Opc = Is64Bit ? X86::LEA64_32r : X86::LEA32r;
980 
981     // LEA can't handle ESP.
982     bool isKill;
983     unsigned SrcReg;
984     MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
985     if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false,
986                         SrcReg, isKill, ImplicitOp, LV))
987       return nullptr;
988 
989     MachineInstrBuilder MIB =
990         BuildMI(MF, MI.getDebugLoc(), get(Opc))
991             .add(Dest)
992             .addReg(0)
993             .addImm(1ULL << ShAmt)
994             .addReg(SrcReg, getKillRegState(isKill))
995             .addImm(0)
996             .addReg(0);
997     if (ImplicitOp.getReg() != 0)
998       MIB.add(ImplicitOp);
999     NewMI = MIB;
1000 
1001     break;
1002   }
1003   case X86::SHL16ri: {
1004     assert(MI.getNumOperands() >= 3 && "Unknown shift instruction!");
1005     unsigned ShAmt = getTruncatedShiftCount(MI, 2);
1006     if (!isTruncatedShiftCountForLEA(ShAmt))
1007       return nullptr;
1008     return convertToThreeAddressWithLEA(MIOpc, MFI, MI, LV);
1009   }
1010   case X86::INC64r:
1011   case X86::INC32r: {
1012     assert(MI.getNumOperands() >= 2 && "Unknown inc instruction!");
1013     unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r :
1014         (Is64Bit ? X86::LEA64_32r : X86::LEA32r);
1015     bool isKill;
1016     unsigned SrcReg;
1017     MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
1018     if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false, SrcReg, isKill,
1019                         ImplicitOp, LV))
1020       return nullptr;
1021 
1022     MachineInstrBuilder MIB =
1023         BuildMI(MF, MI.getDebugLoc(), get(Opc))
1024             .add(Dest)
1025             .addReg(SrcReg, getKillRegState(isKill));
1026     if (ImplicitOp.getReg() != 0)
1027       MIB.add(ImplicitOp);
1028 
1029     NewMI = addOffset(MIB, 1);
1030     break;
1031   }
1032   case X86::INC16r:
1033     return convertToThreeAddressWithLEA(MIOpc, MFI, MI, LV);
1034   case X86::DEC64r:
1035   case X86::DEC32r: {
1036     assert(MI.getNumOperands() >= 2 && "Unknown dec instruction!");
1037     unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r
1038         : (Is64Bit ? X86::LEA64_32r : X86::LEA32r);
1039 
1040     bool isKill;
1041     unsigned SrcReg;
1042     MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
1043     if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false, SrcReg, isKill,
1044                         ImplicitOp, LV))
1045       return nullptr;
1046 
1047     MachineInstrBuilder MIB = BuildMI(MF, MI.getDebugLoc(), get(Opc))
1048                                   .add(Dest)
1049                                   .addReg(SrcReg, getKillRegState(isKill));
1050     if (ImplicitOp.getReg() != 0)
1051       MIB.add(ImplicitOp);
1052 
1053     NewMI = addOffset(MIB, -1);
1054 
1055     break;
1056   }
1057   case X86::DEC16r:
1058     return convertToThreeAddressWithLEA(MIOpc, MFI, MI, LV);
1059   case X86::ADD64rr:
1060   case X86::ADD64rr_DB:
1061   case X86::ADD32rr:
1062   case X86::ADD32rr_DB: {
1063     assert(MI.getNumOperands() >= 3 && "Unknown add instruction!");
1064     unsigned Opc;
1065     if (MIOpc == X86::ADD64rr || MIOpc == X86::ADD64rr_DB)
1066       Opc = X86::LEA64r;
1067     else
1068       Opc = Is64Bit ? X86::LEA64_32r : X86::LEA32r;
1069 
1070     bool isKill;
1071     unsigned SrcReg;
1072     MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
1073     if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ true,
1074                         SrcReg, isKill, ImplicitOp, LV))
1075       return nullptr;
1076 
1077     const MachineOperand &Src2 = MI.getOperand(2);
1078     bool isKill2;
1079     unsigned SrcReg2;
1080     MachineOperand ImplicitOp2 = MachineOperand::CreateReg(0, false);
1081     if (!classifyLEAReg(MI, Src2, Opc, /*AllowSP=*/ false,
1082                         SrcReg2, isKill2, ImplicitOp2, LV))
1083       return nullptr;
1084 
1085     MachineInstrBuilder MIB = BuildMI(MF, MI.getDebugLoc(), get(Opc)).add(Dest);
1086     if (ImplicitOp.getReg() != 0)
1087       MIB.add(ImplicitOp);
1088     if (ImplicitOp2.getReg() != 0)
1089       MIB.add(ImplicitOp2);
1090 
1091     NewMI = addRegReg(MIB, SrcReg, isKill, SrcReg2, isKill2);
1092     if (LV && Src2.isKill())
1093       LV->replaceKillInstruction(SrcReg2, MI, *NewMI);
1094     break;
1095   }
1096   case X86::ADD8rr:
1097   case X86::ADD16rr:
1098   case X86::ADD16rr_DB:
1099     return convertToThreeAddressWithLEA(MIOpc, MFI, MI, LV);
1100   case X86::ADD64ri32:
1101   case X86::ADD64ri8:
1102   case X86::ADD64ri32_DB:
1103   case X86::ADD64ri8_DB:
1104     assert(MI.getNumOperands() >= 3 && "Unknown add instruction!");
1105     NewMI = addOffset(
1106         BuildMI(MF, MI.getDebugLoc(), get(X86::LEA64r)).add(Dest).add(Src),
1107         MI.getOperand(2));
1108     break;
1109   case X86::ADD32ri:
1110   case X86::ADD32ri8:
1111   case X86::ADD32ri_DB:
1112   case X86::ADD32ri8_DB: {
1113     assert(MI.getNumOperands() >= 3 && "Unknown add instruction!");
1114     unsigned Opc = Is64Bit ? X86::LEA64_32r : X86::LEA32r;
1115 
1116     bool isKill;
1117     unsigned SrcReg;
1118     MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
1119     if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ true,
1120                         SrcReg, isKill, ImplicitOp, LV))
1121       return nullptr;
1122 
1123     MachineInstrBuilder MIB = BuildMI(MF, MI.getDebugLoc(), get(Opc))
1124                                   .add(Dest)
1125                                   .addReg(SrcReg, getKillRegState(isKill));
1126     if (ImplicitOp.getReg() != 0)
1127       MIB.add(ImplicitOp);
1128 
1129     NewMI = addOffset(MIB, MI.getOperand(2));
1130     break;
1131   }
1132   case X86::ADD8ri:
1133   case X86::ADD16ri:
1134   case X86::ADD16ri8:
1135   case X86::ADD16ri_DB:
1136   case X86::ADD16ri8_DB:
1137     return convertToThreeAddressWithLEA(MIOpc, MFI, MI, LV);
1138   case X86::VMOVDQU8Z128rmk:
1139   case X86::VMOVDQU8Z256rmk:
1140   case X86::VMOVDQU8Zrmk:
1141   case X86::VMOVDQU16Z128rmk:
1142   case X86::VMOVDQU16Z256rmk:
1143   case X86::VMOVDQU16Zrmk:
1144   case X86::VMOVDQU32Z128rmk: case X86::VMOVDQA32Z128rmk:
1145   case X86::VMOVDQU32Z256rmk: case X86::VMOVDQA32Z256rmk:
1146   case X86::VMOVDQU32Zrmk:    case X86::VMOVDQA32Zrmk:
1147   case X86::VMOVDQU64Z128rmk: case X86::VMOVDQA64Z128rmk:
1148   case X86::VMOVDQU64Z256rmk: case X86::VMOVDQA64Z256rmk:
1149   case X86::VMOVDQU64Zrmk:    case X86::VMOVDQA64Zrmk:
1150   case X86::VMOVUPDZ128rmk:   case X86::VMOVAPDZ128rmk:
1151   case X86::VMOVUPDZ256rmk:   case X86::VMOVAPDZ256rmk:
1152   case X86::VMOVUPDZrmk:      case X86::VMOVAPDZrmk:
1153   case X86::VMOVUPSZ128rmk:   case X86::VMOVAPSZ128rmk:
1154   case X86::VMOVUPSZ256rmk:   case X86::VMOVAPSZ256rmk:
1155   case X86::VMOVUPSZrmk:      case X86::VMOVAPSZrmk: {
1156     unsigned Opc;
1157     switch (MIOpc) {
1158     default: llvm_unreachable("Unreachable!");
1159     case X86::VMOVDQU8Z128rmk:  Opc = X86::VPBLENDMBZ128rmk; break;
1160     case X86::VMOVDQU8Z256rmk:  Opc = X86::VPBLENDMBZ256rmk; break;
1161     case X86::VMOVDQU8Zrmk:     Opc = X86::VPBLENDMBZrmk;    break;
1162     case X86::VMOVDQU16Z128rmk: Opc = X86::VPBLENDMWZ128rmk; break;
1163     case X86::VMOVDQU16Z256rmk: Opc = X86::VPBLENDMWZ256rmk; break;
1164     case X86::VMOVDQU16Zrmk:    Opc = X86::VPBLENDMWZrmk;    break;
1165     case X86::VMOVDQU32Z128rmk: Opc = X86::VPBLENDMDZ128rmk; break;
1166     case X86::VMOVDQU32Z256rmk: Opc = X86::VPBLENDMDZ256rmk; break;
1167     case X86::VMOVDQU32Zrmk:    Opc = X86::VPBLENDMDZrmk;    break;
1168     case X86::VMOVDQU64Z128rmk: Opc = X86::VPBLENDMQZ128rmk; break;
1169     case X86::VMOVDQU64Z256rmk: Opc = X86::VPBLENDMQZ256rmk; break;
1170     case X86::VMOVDQU64Zrmk:    Opc = X86::VPBLENDMQZrmk;    break;
1171     case X86::VMOVUPDZ128rmk:   Opc = X86::VBLENDMPDZ128rmk; break;
1172     case X86::VMOVUPDZ256rmk:   Opc = X86::VBLENDMPDZ256rmk; break;
1173     case X86::VMOVUPDZrmk:      Opc = X86::VBLENDMPDZrmk;    break;
1174     case X86::VMOVUPSZ128rmk:   Opc = X86::VBLENDMPSZ128rmk; break;
1175     case X86::VMOVUPSZ256rmk:   Opc = X86::VBLENDMPSZ256rmk; break;
1176     case X86::VMOVUPSZrmk:      Opc = X86::VBLENDMPSZrmk;    break;
1177     case X86::VMOVDQA32Z128rmk: Opc = X86::VPBLENDMDZ128rmk; break;
1178     case X86::VMOVDQA32Z256rmk: Opc = X86::VPBLENDMDZ256rmk; break;
1179     case X86::VMOVDQA32Zrmk:    Opc = X86::VPBLENDMDZrmk;    break;
1180     case X86::VMOVDQA64Z128rmk: Opc = X86::VPBLENDMQZ128rmk; break;
1181     case X86::VMOVDQA64Z256rmk: Opc = X86::VPBLENDMQZ256rmk; break;
1182     case X86::VMOVDQA64Zrmk:    Opc = X86::VPBLENDMQZrmk;    break;
1183     case X86::VMOVAPDZ128rmk:   Opc = X86::VBLENDMPDZ128rmk; break;
1184     case X86::VMOVAPDZ256rmk:   Opc = X86::VBLENDMPDZ256rmk; break;
1185     case X86::VMOVAPDZrmk:      Opc = X86::VBLENDMPDZrmk;    break;
1186     case X86::VMOVAPSZ128rmk:   Opc = X86::VBLENDMPSZ128rmk; break;
1187     case X86::VMOVAPSZ256rmk:   Opc = X86::VBLENDMPSZ256rmk; break;
1188     case X86::VMOVAPSZrmk:      Opc = X86::VBLENDMPSZrmk;    break;
1189     }
1190 
1191     NewMI = BuildMI(MF, MI.getDebugLoc(), get(Opc))
1192               .add(Dest)
1193               .add(MI.getOperand(2))
1194               .add(Src)
1195               .add(MI.getOperand(3))
1196               .add(MI.getOperand(4))
1197               .add(MI.getOperand(5))
1198               .add(MI.getOperand(6))
1199               .add(MI.getOperand(7));
1200     break;
1201   }
1202   case X86::VMOVDQU8Z128rrk:
1203   case X86::VMOVDQU8Z256rrk:
1204   case X86::VMOVDQU8Zrrk:
1205   case X86::VMOVDQU16Z128rrk:
1206   case X86::VMOVDQU16Z256rrk:
1207   case X86::VMOVDQU16Zrrk:
1208   case X86::VMOVDQU32Z128rrk: case X86::VMOVDQA32Z128rrk:
1209   case X86::VMOVDQU32Z256rrk: case X86::VMOVDQA32Z256rrk:
1210   case X86::VMOVDQU32Zrrk:    case X86::VMOVDQA32Zrrk:
1211   case X86::VMOVDQU64Z128rrk: case X86::VMOVDQA64Z128rrk:
1212   case X86::VMOVDQU64Z256rrk: case X86::VMOVDQA64Z256rrk:
1213   case X86::VMOVDQU64Zrrk:    case X86::VMOVDQA64Zrrk:
1214   case X86::VMOVUPDZ128rrk:   case X86::VMOVAPDZ128rrk:
1215   case X86::VMOVUPDZ256rrk:   case X86::VMOVAPDZ256rrk:
1216   case X86::VMOVUPDZrrk:      case X86::VMOVAPDZrrk:
1217   case X86::VMOVUPSZ128rrk:   case X86::VMOVAPSZ128rrk:
1218   case X86::VMOVUPSZ256rrk:   case X86::VMOVAPSZ256rrk:
1219   case X86::VMOVUPSZrrk:      case X86::VMOVAPSZrrk: {
1220     unsigned Opc;
1221     switch (MIOpc) {
1222     default: llvm_unreachable("Unreachable!");
1223     case X86::VMOVDQU8Z128rrk:  Opc = X86::VPBLENDMBZ128rrk; break;
1224     case X86::VMOVDQU8Z256rrk:  Opc = X86::VPBLENDMBZ256rrk; break;
1225     case X86::VMOVDQU8Zrrk:     Opc = X86::VPBLENDMBZrrk;    break;
1226     case X86::VMOVDQU16Z128rrk: Opc = X86::VPBLENDMWZ128rrk; break;
1227     case X86::VMOVDQU16Z256rrk: Opc = X86::VPBLENDMWZ256rrk; break;
1228     case X86::VMOVDQU16Zrrk:    Opc = X86::VPBLENDMWZrrk;    break;
1229     case X86::VMOVDQU32Z128rrk: Opc = X86::VPBLENDMDZ128rrk; break;
1230     case X86::VMOVDQU32Z256rrk: Opc = X86::VPBLENDMDZ256rrk; break;
1231     case X86::VMOVDQU32Zrrk:    Opc = X86::VPBLENDMDZrrk;    break;
1232     case X86::VMOVDQU64Z128rrk: Opc = X86::VPBLENDMQZ128rrk; break;
1233     case X86::VMOVDQU64Z256rrk: Opc = X86::VPBLENDMQZ256rrk; break;
1234     case X86::VMOVDQU64Zrrk:    Opc = X86::VPBLENDMQZrrk;    break;
1235     case X86::VMOVUPDZ128rrk:   Opc = X86::VBLENDMPDZ128rrk; break;
1236     case X86::VMOVUPDZ256rrk:   Opc = X86::VBLENDMPDZ256rrk; break;
1237     case X86::VMOVUPDZrrk:      Opc = X86::VBLENDMPDZrrk;    break;
1238     case X86::VMOVUPSZ128rrk:   Opc = X86::VBLENDMPSZ128rrk; break;
1239     case X86::VMOVUPSZ256rrk:   Opc = X86::VBLENDMPSZ256rrk; break;
1240     case X86::VMOVUPSZrrk:      Opc = X86::VBLENDMPSZrrk;    break;
1241     case X86::VMOVDQA32Z128rrk: Opc = X86::VPBLENDMDZ128rrk; break;
1242     case X86::VMOVDQA32Z256rrk: Opc = X86::VPBLENDMDZ256rrk; break;
1243     case X86::VMOVDQA32Zrrk:    Opc = X86::VPBLENDMDZrrk;    break;
1244     case X86::VMOVDQA64Z128rrk: Opc = X86::VPBLENDMQZ128rrk; break;
1245     case X86::VMOVDQA64Z256rrk: Opc = X86::VPBLENDMQZ256rrk; break;
1246     case X86::VMOVDQA64Zrrk:    Opc = X86::VPBLENDMQZrrk;    break;
1247     case X86::VMOVAPDZ128rrk:   Opc = X86::VBLENDMPDZ128rrk; break;
1248     case X86::VMOVAPDZ256rrk:   Opc = X86::VBLENDMPDZ256rrk; break;
1249     case X86::VMOVAPDZrrk:      Opc = X86::VBLENDMPDZrrk;    break;
1250     case X86::VMOVAPSZ128rrk:   Opc = X86::VBLENDMPSZ128rrk; break;
1251     case X86::VMOVAPSZ256rrk:   Opc = X86::VBLENDMPSZ256rrk; break;
1252     case X86::VMOVAPSZrrk:      Opc = X86::VBLENDMPSZrrk;    break;
1253     }
1254 
1255     NewMI = BuildMI(MF, MI.getDebugLoc(), get(Opc))
1256               .add(Dest)
1257               .add(MI.getOperand(2))
1258               .add(Src)
1259               .add(MI.getOperand(3));
1260     break;
1261   }
1262   }
1263 
1264   if (!NewMI) return nullptr;
1265 
1266   if (LV) {  // Update live variables
1267     if (Src.isKill())
1268       LV->replaceKillInstruction(Src.getReg(), MI, *NewMI);
1269     if (Dest.isDead())
1270       LV->replaceKillInstruction(Dest.getReg(), MI, *NewMI);
1271   }
1272 
1273   MFI->insert(MI.getIterator(), NewMI); // Insert the new inst
1274   return NewMI;
1275 }
1276 
1277 /// This determines which of three possible cases of a three source commute
1278 /// the source indexes correspond to taking into account any mask operands.
1279 /// All prevents commuting a passthru operand. Returns -1 if the commute isn't
1280 /// possible.
1281 /// Case 0 - Possible to commute the first and second operands.
1282 /// Case 1 - Possible to commute the first and third operands.
1283 /// Case 2 - Possible to commute the second and third operands.
getThreeSrcCommuteCase(uint64_t TSFlags,unsigned SrcOpIdx1,unsigned SrcOpIdx2)1284 static unsigned getThreeSrcCommuteCase(uint64_t TSFlags, unsigned SrcOpIdx1,
1285                                        unsigned SrcOpIdx2) {
1286   // Put the lowest index to SrcOpIdx1 to simplify the checks below.
1287   if (SrcOpIdx1 > SrcOpIdx2)
1288     std::swap(SrcOpIdx1, SrcOpIdx2);
1289 
1290   unsigned Op1 = 1, Op2 = 2, Op3 = 3;
1291   if (X86II::isKMasked(TSFlags)) {
1292     Op2++;
1293     Op3++;
1294   }
1295 
1296   if (SrcOpIdx1 == Op1 && SrcOpIdx2 == Op2)
1297     return 0;
1298   if (SrcOpIdx1 == Op1 && SrcOpIdx2 == Op3)
1299     return 1;
1300   if (SrcOpIdx1 == Op2 && SrcOpIdx2 == Op3)
1301     return 2;
1302   llvm_unreachable("Unknown three src commute case.");
1303 }
1304 
getFMA3OpcodeToCommuteOperands(const MachineInstr & MI,unsigned SrcOpIdx1,unsigned SrcOpIdx2,const X86InstrFMA3Group & FMA3Group) const1305 unsigned X86InstrInfo::getFMA3OpcodeToCommuteOperands(
1306     const MachineInstr &MI, unsigned SrcOpIdx1, unsigned SrcOpIdx2,
1307     const X86InstrFMA3Group &FMA3Group) const {
1308 
1309   unsigned Opc = MI.getOpcode();
1310 
1311   // TODO: Commuting the 1st operand of FMA*_Int requires some additional
1312   // analysis. The commute optimization is legal only if all users of FMA*_Int
1313   // use only the lowest element of the FMA*_Int instruction. Such analysis are
1314   // not implemented yet. So, just return 0 in that case.
1315   // When such analysis are available this place will be the right place for
1316   // calling it.
1317   assert(!(FMA3Group.isIntrinsic() && (SrcOpIdx1 == 1 || SrcOpIdx2 == 1)) &&
1318          "Intrinsic instructions can't commute operand 1");
1319 
1320   // Determine which case this commute is or if it can't be done.
1321   unsigned Case = getThreeSrcCommuteCase(MI.getDesc().TSFlags, SrcOpIdx1,
1322                                          SrcOpIdx2);
1323   assert(Case < 3 && "Unexpected case number!");
1324 
1325   // Define the FMA forms mapping array that helps to map input FMA form
1326   // to output FMA form to preserve the operation semantics after
1327   // commuting the operands.
1328   const unsigned Form132Index = 0;
1329   const unsigned Form213Index = 1;
1330   const unsigned Form231Index = 2;
1331   static const unsigned FormMapping[][3] = {
1332     // 0: SrcOpIdx1 == 1 && SrcOpIdx2 == 2;
1333     // FMA132 A, C, b; ==> FMA231 C, A, b;
1334     // FMA213 B, A, c; ==> FMA213 A, B, c;
1335     // FMA231 C, A, b; ==> FMA132 A, C, b;
1336     { Form231Index, Form213Index, Form132Index },
1337     // 1: SrcOpIdx1 == 1 && SrcOpIdx2 == 3;
1338     // FMA132 A, c, B; ==> FMA132 B, c, A;
1339     // FMA213 B, a, C; ==> FMA231 C, a, B;
1340     // FMA231 C, a, B; ==> FMA213 B, a, C;
1341     { Form132Index, Form231Index, Form213Index },
1342     // 2: SrcOpIdx1 == 2 && SrcOpIdx2 == 3;
1343     // FMA132 a, C, B; ==> FMA213 a, B, C;
1344     // FMA213 b, A, C; ==> FMA132 b, C, A;
1345     // FMA231 c, A, B; ==> FMA231 c, B, A;
1346     { Form213Index, Form132Index, Form231Index }
1347   };
1348 
1349   unsigned FMAForms[3];
1350   FMAForms[0] = FMA3Group.get132Opcode();
1351   FMAForms[1] = FMA3Group.get213Opcode();
1352   FMAForms[2] = FMA3Group.get231Opcode();
1353   unsigned FormIndex;
1354   for (FormIndex = 0; FormIndex < 3; FormIndex++)
1355     if (Opc == FMAForms[FormIndex])
1356       break;
1357 
1358   // Everything is ready, just adjust the FMA opcode and return it.
1359   FormIndex = FormMapping[Case][FormIndex];
1360   return FMAForms[FormIndex];
1361 }
1362 
commuteVPTERNLOG(MachineInstr & MI,unsigned SrcOpIdx1,unsigned SrcOpIdx2)1363 static void commuteVPTERNLOG(MachineInstr &MI, unsigned SrcOpIdx1,
1364                              unsigned SrcOpIdx2) {
1365   // Determine which case this commute is or if it can't be done.
1366   unsigned Case = getThreeSrcCommuteCase(MI.getDesc().TSFlags, SrcOpIdx1,
1367                                          SrcOpIdx2);
1368   assert(Case < 3 && "Unexpected case value!");
1369 
1370   // For each case we need to swap two pairs of bits in the final immediate.
1371   static const uint8_t SwapMasks[3][4] = {
1372     { 0x04, 0x10, 0x08, 0x20 }, // Swap bits 2/4 and 3/5.
1373     { 0x02, 0x10, 0x08, 0x40 }, // Swap bits 1/4 and 3/6.
1374     { 0x02, 0x04, 0x20, 0x40 }, // Swap bits 1/2 and 5/6.
1375   };
1376 
1377   uint8_t Imm = MI.getOperand(MI.getNumOperands()-1).getImm();
1378   // Clear out the bits we are swapping.
1379   uint8_t NewImm = Imm & ~(SwapMasks[Case][0] | SwapMasks[Case][1] |
1380                            SwapMasks[Case][2] | SwapMasks[Case][3]);
1381   // If the immediate had a bit of the pair set, then set the opposite bit.
1382   if (Imm & SwapMasks[Case][0]) NewImm |= SwapMasks[Case][1];
1383   if (Imm & SwapMasks[Case][1]) NewImm |= SwapMasks[Case][0];
1384   if (Imm & SwapMasks[Case][2]) NewImm |= SwapMasks[Case][3];
1385   if (Imm & SwapMasks[Case][3]) NewImm |= SwapMasks[Case][2];
1386   MI.getOperand(MI.getNumOperands()-1).setImm(NewImm);
1387 }
1388 
1389 // Returns true if this is a VPERMI2 or VPERMT2 instruction that can be
1390 // commuted.
isCommutableVPERMV3Instruction(unsigned Opcode)1391 static bool isCommutableVPERMV3Instruction(unsigned Opcode) {
1392 #define VPERM_CASES(Suffix) \
1393   case X86::VPERMI2##Suffix##128rr:    case X86::VPERMT2##Suffix##128rr:    \
1394   case X86::VPERMI2##Suffix##256rr:    case X86::VPERMT2##Suffix##256rr:    \
1395   case X86::VPERMI2##Suffix##rr:       case X86::VPERMT2##Suffix##rr:       \
1396   case X86::VPERMI2##Suffix##128rm:    case X86::VPERMT2##Suffix##128rm:    \
1397   case X86::VPERMI2##Suffix##256rm:    case X86::VPERMT2##Suffix##256rm:    \
1398   case X86::VPERMI2##Suffix##rm:       case X86::VPERMT2##Suffix##rm:       \
1399   case X86::VPERMI2##Suffix##128rrkz:  case X86::VPERMT2##Suffix##128rrkz:  \
1400   case X86::VPERMI2##Suffix##256rrkz:  case X86::VPERMT2##Suffix##256rrkz:  \
1401   case X86::VPERMI2##Suffix##rrkz:     case X86::VPERMT2##Suffix##rrkz:     \
1402   case X86::VPERMI2##Suffix##128rmkz:  case X86::VPERMT2##Suffix##128rmkz:  \
1403   case X86::VPERMI2##Suffix##256rmkz:  case X86::VPERMT2##Suffix##256rmkz:  \
1404   case X86::VPERMI2##Suffix##rmkz:     case X86::VPERMT2##Suffix##rmkz:
1405 
1406 #define VPERM_CASES_BROADCAST(Suffix) \
1407   VPERM_CASES(Suffix) \
1408   case X86::VPERMI2##Suffix##128rmb:   case X86::VPERMT2##Suffix##128rmb:   \
1409   case X86::VPERMI2##Suffix##256rmb:   case X86::VPERMT2##Suffix##256rmb:   \
1410   case X86::VPERMI2##Suffix##rmb:      case X86::VPERMT2##Suffix##rmb:      \
1411   case X86::VPERMI2##Suffix##128rmbkz: case X86::VPERMT2##Suffix##128rmbkz: \
1412   case X86::VPERMI2##Suffix##256rmbkz: case X86::VPERMT2##Suffix##256rmbkz: \
1413   case X86::VPERMI2##Suffix##rmbkz:    case X86::VPERMT2##Suffix##rmbkz:
1414 
1415   switch (Opcode) {
1416   default: return false;
1417   VPERM_CASES(B)
1418   VPERM_CASES_BROADCAST(D)
1419   VPERM_CASES_BROADCAST(PD)
1420   VPERM_CASES_BROADCAST(PS)
1421   VPERM_CASES_BROADCAST(Q)
1422   VPERM_CASES(W)
1423     return true;
1424   }
1425 #undef VPERM_CASES_BROADCAST
1426 #undef VPERM_CASES
1427 }
1428 
1429 // Returns commuted opcode for VPERMI2 and VPERMT2 instructions by switching
1430 // from the I opcode to the T opcode and vice versa.
getCommutedVPERMV3Opcode(unsigned Opcode)1431 static unsigned getCommutedVPERMV3Opcode(unsigned Opcode) {
1432 #define VPERM_CASES(Orig, New) \
1433   case X86::Orig##128rr:    return X86::New##128rr;   \
1434   case X86::Orig##128rrkz:  return X86::New##128rrkz; \
1435   case X86::Orig##128rm:    return X86::New##128rm;   \
1436   case X86::Orig##128rmkz:  return X86::New##128rmkz; \
1437   case X86::Orig##256rr:    return X86::New##256rr;   \
1438   case X86::Orig##256rrkz:  return X86::New##256rrkz; \
1439   case X86::Orig##256rm:    return X86::New##256rm;   \
1440   case X86::Orig##256rmkz:  return X86::New##256rmkz; \
1441   case X86::Orig##rr:       return X86::New##rr;      \
1442   case X86::Orig##rrkz:     return X86::New##rrkz;    \
1443   case X86::Orig##rm:       return X86::New##rm;      \
1444   case X86::Orig##rmkz:     return X86::New##rmkz;
1445 
1446 #define VPERM_CASES_BROADCAST(Orig, New) \
1447   VPERM_CASES(Orig, New) \
1448   case X86::Orig##128rmb:   return X86::New##128rmb;   \
1449   case X86::Orig##128rmbkz: return X86::New##128rmbkz; \
1450   case X86::Orig##256rmb:   return X86::New##256rmb;   \
1451   case X86::Orig##256rmbkz: return X86::New##256rmbkz; \
1452   case X86::Orig##rmb:      return X86::New##rmb;      \
1453   case X86::Orig##rmbkz:    return X86::New##rmbkz;
1454 
1455   switch (Opcode) {
1456   VPERM_CASES(VPERMI2B, VPERMT2B)
1457   VPERM_CASES_BROADCAST(VPERMI2D,  VPERMT2D)
1458   VPERM_CASES_BROADCAST(VPERMI2PD, VPERMT2PD)
1459   VPERM_CASES_BROADCAST(VPERMI2PS, VPERMT2PS)
1460   VPERM_CASES_BROADCAST(VPERMI2Q,  VPERMT2Q)
1461   VPERM_CASES(VPERMI2W, VPERMT2W)
1462   VPERM_CASES(VPERMT2B, VPERMI2B)
1463   VPERM_CASES_BROADCAST(VPERMT2D,  VPERMI2D)
1464   VPERM_CASES_BROADCAST(VPERMT2PD, VPERMI2PD)
1465   VPERM_CASES_BROADCAST(VPERMT2PS, VPERMI2PS)
1466   VPERM_CASES_BROADCAST(VPERMT2Q,  VPERMI2Q)
1467   VPERM_CASES(VPERMT2W, VPERMI2W)
1468   }
1469 
1470   llvm_unreachable("Unreachable!");
1471 #undef VPERM_CASES_BROADCAST
1472 #undef VPERM_CASES
1473 }
1474 
commuteInstructionImpl(MachineInstr & MI,bool NewMI,unsigned OpIdx1,unsigned OpIdx2) const1475 MachineInstr *X86InstrInfo::commuteInstructionImpl(MachineInstr &MI, bool NewMI,
1476                                                    unsigned OpIdx1,
1477                                                    unsigned OpIdx2) const {
1478   auto cloneIfNew = [NewMI](MachineInstr &MI) -> MachineInstr & {
1479     if (NewMI)
1480       return *MI.getParent()->getParent()->CloneMachineInstr(&MI);
1481     return MI;
1482   };
1483 
1484   switch (MI.getOpcode()) {
1485   case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I)
1486   case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I)
1487   case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I)
1488   case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I)
1489   case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I)
1490   case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I)
1491     unsigned Opc;
1492     unsigned Size;
1493     switch (MI.getOpcode()) {
1494     default: llvm_unreachable("Unreachable!");
1495     case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break;
1496     case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break;
1497     case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break;
1498     case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break;
1499     case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break;
1500     case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break;
1501     }
1502     unsigned Amt = MI.getOperand(3).getImm();
1503     auto &WorkingMI = cloneIfNew(MI);
1504     WorkingMI.setDesc(get(Opc));
1505     WorkingMI.getOperand(3).setImm(Size - Amt);
1506     return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1507                                                    OpIdx1, OpIdx2);
1508   }
1509   case X86::PFSUBrr:
1510   case X86::PFSUBRrr: {
1511     // PFSUB  x, y: x = x - y
1512     // PFSUBR x, y: x = y - x
1513     unsigned Opc =
1514         (X86::PFSUBRrr == MI.getOpcode() ? X86::PFSUBrr : X86::PFSUBRrr);
1515     auto &WorkingMI = cloneIfNew(MI);
1516     WorkingMI.setDesc(get(Opc));
1517     return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1518                                                    OpIdx1, OpIdx2);
1519   }
1520   case X86::BLENDPDrri:
1521   case X86::BLENDPSrri:
1522   case X86::VBLENDPDrri:
1523   case X86::VBLENDPSrri:
1524     // If we're optimizing for size, try to use MOVSD/MOVSS.
1525     if (MI.getParent()->getParent()->getFunction().optForSize()) {
1526       unsigned Mask, Opc;
1527       switch (MI.getOpcode()) {
1528       default: llvm_unreachable("Unreachable!");
1529       case X86::BLENDPDrri:  Opc = X86::MOVSDrr;  Mask = 0x03; break;
1530       case X86::BLENDPSrri:  Opc = X86::MOVSSrr;  Mask = 0x0F; break;
1531       case X86::VBLENDPDrri: Opc = X86::VMOVSDrr; Mask = 0x03; break;
1532       case X86::VBLENDPSrri: Opc = X86::VMOVSSrr; Mask = 0x0F; break;
1533       }
1534       if ((MI.getOperand(3).getImm() ^ Mask) == 1) {
1535         auto &WorkingMI = cloneIfNew(MI);
1536         WorkingMI.setDesc(get(Opc));
1537         WorkingMI.RemoveOperand(3);
1538         return TargetInstrInfo::commuteInstructionImpl(WorkingMI,
1539                                                        /*NewMI=*/false,
1540                                                        OpIdx1, OpIdx2);
1541       }
1542     }
1543     LLVM_FALLTHROUGH;
1544   case X86::PBLENDWrri:
1545   case X86::VBLENDPDYrri:
1546   case X86::VBLENDPSYrri:
1547   case X86::VPBLENDDrri:
1548   case X86::VPBLENDWrri:
1549   case X86::VPBLENDDYrri:
1550   case X86::VPBLENDWYrri:{
1551     unsigned Mask;
1552     switch (MI.getOpcode()) {
1553     default: llvm_unreachable("Unreachable!");
1554     case X86::BLENDPDrri:    Mask = 0x03; break;
1555     case X86::BLENDPSrri:    Mask = 0x0F; break;
1556     case X86::PBLENDWrri:    Mask = 0xFF; break;
1557     case X86::VBLENDPDrri:   Mask = 0x03; break;
1558     case X86::VBLENDPSrri:   Mask = 0x0F; break;
1559     case X86::VBLENDPDYrri:  Mask = 0x0F; break;
1560     case X86::VBLENDPSYrri:  Mask = 0xFF; break;
1561     case X86::VPBLENDDrri:   Mask = 0x0F; break;
1562     case X86::VPBLENDWrri:   Mask = 0xFF; break;
1563     case X86::VPBLENDDYrri:  Mask = 0xFF; break;
1564     case X86::VPBLENDWYrri:  Mask = 0xFF; break;
1565     }
1566     // Only the least significant bits of Imm are used.
1567     unsigned Imm = MI.getOperand(3).getImm() & Mask;
1568     auto &WorkingMI = cloneIfNew(MI);
1569     WorkingMI.getOperand(3).setImm(Mask ^ Imm);
1570     return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1571                                                    OpIdx1, OpIdx2);
1572   }
1573   case X86::MOVSDrr:
1574   case X86::MOVSSrr:
1575   case X86::VMOVSDrr:
1576   case X86::VMOVSSrr:{
1577     // On SSE41 or later we can commute a MOVSS/MOVSD to a BLENDPS/BLENDPD.
1578     assert(Subtarget.hasSSE41() && "Commuting MOVSD/MOVSS requires SSE41!");
1579 
1580     unsigned Mask, Opc;
1581     switch (MI.getOpcode()) {
1582     default: llvm_unreachable("Unreachable!");
1583     case X86::MOVSDrr:  Opc = X86::BLENDPDrri;  Mask = 0x02; break;
1584     case X86::MOVSSrr:  Opc = X86::BLENDPSrri;  Mask = 0x0E; break;
1585     case X86::VMOVSDrr: Opc = X86::VBLENDPDrri; Mask = 0x02; break;
1586     case X86::VMOVSSrr: Opc = X86::VBLENDPSrri; Mask = 0x0E; break;
1587     }
1588 
1589     auto &WorkingMI = cloneIfNew(MI);
1590     WorkingMI.setDesc(get(Opc));
1591     WorkingMI.addOperand(MachineOperand::CreateImm(Mask));
1592     return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1593                                                    OpIdx1, OpIdx2);
1594   }
1595   case X86::PCLMULQDQrr:
1596   case X86::VPCLMULQDQrr:
1597   case X86::VPCLMULQDQYrr:
1598   case X86::VPCLMULQDQZrr:
1599   case X86::VPCLMULQDQZ128rr:
1600   case X86::VPCLMULQDQZ256rr: {
1601     // SRC1 64bits = Imm[0] ? SRC1[127:64] : SRC1[63:0]
1602     // SRC2 64bits = Imm[4] ? SRC2[127:64] : SRC2[63:0]
1603     unsigned Imm = MI.getOperand(3).getImm();
1604     unsigned Src1Hi = Imm & 0x01;
1605     unsigned Src2Hi = Imm & 0x10;
1606     auto &WorkingMI = cloneIfNew(MI);
1607     WorkingMI.getOperand(3).setImm((Src1Hi << 4) | (Src2Hi >> 4));
1608     return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1609                                                    OpIdx1, OpIdx2);
1610   }
1611   case X86::VPCMPBZ128rri:  case X86::VPCMPUBZ128rri:
1612   case X86::VPCMPBZ256rri:  case X86::VPCMPUBZ256rri:
1613   case X86::VPCMPBZrri:     case X86::VPCMPUBZrri:
1614   case X86::VPCMPDZ128rri:  case X86::VPCMPUDZ128rri:
1615   case X86::VPCMPDZ256rri:  case X86::VPCMPUDZ256rri:
1616   case X86::VPCMPDZrri:     case X86::VPCMPUDZrri:
1617   case X86::VPCMPQZ128rri:  case X86::VPCMPUQZ128rri:
1618   case X86::VPCMPQZ256rri:  case X86::VPCMPUQZ256rri:
1619   case X86::VPCMPQZrri:     case X86::VPCMPUQZrri:
1620   case X86::VPCMPWZ128rri:  case X86::VPCMPUWZ128rri:
1621   case X86::VPCMPWZ256rri:  case X86::VPCMPUWZ256rri:
1622   case X86::VPCMPWZrri:     case X86::VPCMPUWZrri:
1623   case X86::VPCMPBZ128rrik: case X86::VPCMPUBZ128rrik:
1624   case X86::VPCMPBZ256rrik: case X86::VPCMPUBZ256rrik:
1625   case X86::VPCMPBZrrik:    case X86::VPCMPUBZrrik:
1626   case X86::VPCMPDZ128rrik: case X86::VPCMPUDZ128rrik:
1627   case X86::VPCMPDZ256rrik: case X86::VPCMPUDZ256rrik:
1628   case X86::VPCMPDZrrik:    case X86::VPCMPUDZrrik:
1629   case X86::VPCMPQZ128rrik: case X86::VPCMPUQZ128rrik:
1630   case X86::VPCMPQZ256rrik: case X86::VPCMPUQZ256rrik:
1631   case X86::VPCMPQZrrik:    case X86::VPCMPUQZrrik:
1632   case X86::VPCMPWZ128rrik: case X86::VPCMPUWZ128rrik:
1633   case X86::VPCMPWZ256rrik: case X86::VPCMPUWZ256rrik:
1634   case X86::VPCMPWZrrik:    case X86::VPCMPUWZrrik: {
1635     // Flip comparison mode immediate (if necessary).
1636     unsigned Imm = MI.getOperand(MI.getNumOperands() - 1).getImm() & 0x7;
1637     Imm = X86::getSwappedVPCMPImm(Imm);
1638     auto &WorkingMI = cloneIfNew(MI);
1639     WorkingMI.getOperand(MI.getNumOperands() - 1).setImm(Imm);
1640     return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1641                                                    OpIdx1, OpIdx2);
1642   }
1643   case X86::VPCOMBri: case X86::VPCOMUBri:
1644   case X86::VPCOMDri: case X86::VPCOMUDri:
1645   case X86::VPCOMQri: case X86::VPCOMUQri:
1646   case X86::VPCOMWri: case X86::VPCOMUWri: {
1647     // Flip comparison mode immediate (if necessary).
1648     unsigned Imm = MI.getOperand(3).getImm() & 0x7;
1649     Imm = X86::getSwappedVPCOMImm(Imm);
1650     auto &WorkingMI = cloneIfNew(MI);
1651     WorkingMI.getOperand(3).setImm(Imm);
1652     return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1653                                                    OpIdx1, OpIdx2);
1654   }
1655   case X86::VPERM2F128rr:
1656   case X86::VPERM2I128rr: {
1657     // Flip permute source immediate.
1658     // Imm & 0x02: lo = if set, select Op1.lo/hi else Op0.lo/hi.
1659     // Imm & 0x20: hi = if set, select Op1.lo/hi else Op0.lo/hi.
1660     unsigned Imm = MI.getOperand(3).getImm() & 0xFF;
1661     auto &WorkingMI = cloneIfNew(MI);
1662     WorkingMI.getOperand(3).setImm(Imm ^ 0x22);
1663     return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1664                                                    OpIdx1, OpIdx2);
1665   }
1666   case X86::MOVHLPSrr:
1667   case X86::UNPCKHPDrr:
1668   case X86::VMOVHLPSrr:
1669   case X86::VUNPCKHPDrr:
1670   case X86::VMOVHLPSZrr:
1671   case X86::VUNPCKHPDZ128rr: {
1672     assert(Subtarget.hasSSE2() && "Commuting MOVHLP/UNPCKHPD requires SSE2!");
1673 
1674     unsigned Opc = MI.getOpcode();
1675     switch (Opc) {
1676     default: llvm_unreachable("Unreachable!");
1677     case X86::MOVHLPSrr:       Opc = X86::UNPCKHPDrr;      break;
1678     case X86::UNPCKHPDrr:      Opc = X86::MOVHLPSrr;       break;
1679     case X86::VMOVHLPSrr:      Opc = X86::VUNPCKHPDrr;     break;
1680     case X86::VUNPCKHPDrr:     Opc = X86::VMOVHLPSrr;      break;
1681     case X86::VMOVHLPSZrr:     Opc = X86::VUNPCKHPDZ128rr; break;
1682     case X86::VUNPCKHPDZ128rr: Opc = X86::VMOVHLPSZrr;     break;
1683     }
1684     auto &WorkingMI = cloneIfNew(MI);
1685     WorkingMI.setDesc(get(Opc));
1686     return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1687                                                    OpIdx1, OpIdx2);
1688   }
1689   case X86::CMOVB16rr:  case X86::CMOVB32rr:  case X86::CMOVB64rr:
1690   case X86::CMOVAE16rr: case X86::CMOVAE32rr: case X86::CMOVAE64rr:
1691   case X86::CMOVE16rr:  case X86::CMOVE32rr:  case X86::CMOVE64rr:
1692   case X86::CMOVNE16rr: case X86::CMOVNE32rr: case X86::CMOVNE64rr:
1693   case X86::CMOVBE16rr: case X86::CMOVBE32rr: case X86::CMOVBE64rr:
1694   case X86::CMOVA16rr:  case X86::CMOVA32rr:  case X86::CMOVA64rr:
1695   case X86::CMOVL16rr:  case X86::CMOVL32rr:  case X86::CMOVL64rr:
1696   case X86::CMOVGE16rr: case X86::CMOVGE32rr: case X86::CMOVGE64rr:
1697   case X86::CMOVLE16rr: case X86::CMOVLE32rr: case X86::CMOVLE64rr:
1698   case X86::CMOVG16rr:  case X86::CMOVG32rr:  case X86::CMOVG64rr:
1699   case X86::CMOVS16rr:  case X86::CMOVS32rr:  case X86::CMOVS64rr:
1700   case X86::CMOVNS16rr: case X86::CMOVNS32rr: case X86::CMOVNS64rr:
1701   case X86::CMOVP16rr:  case X86::CMOVP32rr:  case X86::CMOVP64rr:
1702   case X86::CMOVNP16rr: case X86::CMOVNP32rr: case X86::CMOVNP64rr:
1703   case X86::CMOVO16rr:  case X86::CMOVO32rr:  case X86::CMOVO64rr:
1704   case X86::CMOVNO16rr: case X86::CMOVNO32rr: case X86::CMOVNO64rr: {
1705     unsigned Opc;
1706     switch (MI.getOpcode()) {
1707     default: llvm_unreachable("Unreachable!");
1708     case X86::CMOVB16rr:  Opc = X86::CMOVAE16rr; break;
1709     case X86::CMOVB32rr:  Opc = X86::CMOVAE32rr; break;
1710     case X86::CMOVB64rr:  Opc = X86::CMOVAE64rr; break;
1711     case X86::CMOVAE16rr: Opc = X86::CMOVB16rr; break;
1712     case X86::CMOVAE32rr: Opc = X86::CMOVB32rr; break;
1713     case X86::CMOVAE64rr: Opc = X86::CMOVB64rr; break;
1714     case X86::CMOVE16rr:  Opc = X86::CMOVNE16rr; break;
1715     case X86::CMOVE32rr:  Opc = X86::CMOVNE32rr; break;
1716     case X86::CMOVE64rr:  Opc = X86::CMOVNE64rr; break;
1717     case X86::CMOVNE16rr: Opc = X86::CMOVE16rr; break;
1718     case X86::CMOVNE32rr: Opc = X86::CMOVE32rr; break;
1719     case X86::CMOVNE64rr: Opc = X86::CMOVE64rr; break;
1720     case X86::CMOVBE16rr: Opc = X86::CMOVA16rr; break;
1721     case X86::CMOVBE32rr: Opc = X86::CMOVA32rr; break;
1722     case X86::CMOVBE64rr: Opc = X86::CMOVA64rr; break;
1723     case X86::CMOVA16rr:  Opc = X86::CMOVBE16rr; break;
1724     case X86::CMOVA32rr:  Opc = X86::CMOVBE32rr; break;
1725     case X86::CMOVA64rr:  Opc = X86::CMOVBE64rr; break;
1726     case X86::CMOVL16rr:  Opc = X86::CMOVGE16rr; break;
1727     case X86::CMOVL32rr:  Opc = X86::CMOVGE32rr; break;
1728     case X86::CMOVL64rr:  Opc = X86::CMOVGE64rr; break;
1729     case X86::CMOVGE16rr: Opc = X86::CMOVL16rr; break;
1730     case X86::CMOVGE32rr: Opc = X86::CMOVL32rr; break;
1731     case X86::CMOVGE64rr: Opc = X86::CMOVL64rr; break;
1732     case X86::CMOVLE16rr: Opc = X86::CMOVG16rr; break;
1733     case X86::CMOVLE32rr: Opc = X86::CMOVG32rr; break;
1734     case X86::CMOVLE64rr: Opc = X86::CMOVG64rr; break;
1735     case X86::CMOVG16rr:  Opc = X86::CMOVLE16rr; break;
1736     case X86::CMOVG32rr:  Opc = X86::CMOVLE32rr; break;
1737     case X86::CMOVG64rr:  Opc = X86::CMOVLE64rr; break;
1738     case X86::CMOVS16rr:  Opc = X86::CMOVNS16rr; break;
1739     case X86::CMOVS32rr:  Opc = X86::CMOVNS32rr; break;
1740     case X86::CMOVS64rr:  Opc = X86::CMOVNS64rr; break;
1741     case X86::CMOVNS16rr: Opc = X86::CMOVS16rr; break;
1742     case X86::CMOVNS32rr: Opc = X86::CMOVS32rr; break;
1743     case X86::CMOVNS64rr: Opc = X86::CMOVS64rr; break;
1744     case X86::CMOVP16rr:  Opc = X86::CMOVNP16rr; break;
1745     case X86::CMOVP32rr:  Opc = X86::CMOVNP32rr; break;
1746     case X86::CMOVP64rr:  Opc = X86::CMOVNP64rr; break;
1747     case X86::CMOVNP16rr: Opc = X86::CMOVP16rr; break;
1748     case X86::CMOVNP32rr: Opc = X86::CMOVP32rr; break;
1749     case X86::CMOVNP64rr: Opc = X86::CMOVP64rr; break;
1750     case X86::CMOVO16rr:  Opc = X86::CMOVNO16rr; break;
1751     case X86::CMOVO32rr:  Opc = X86::CMOVNO32rr; break;
1752     case X86::CMOVO64rr:  Opc = X86::CMOVNO64rr; break;
1753     case X86::CMOVNO16rr: Opc = X86::CMOVO16rr; break;
1754     case X86::CMOVNO32rr: Opc = X86::CMOVO32rr; break;
1755     case X86::CMOVNO64rr: Opc = X86::CMOVO64rr; break;
1756     }
1757     auto &WorkingMI = cloneIfNew(MI);
1758     WorkingMI.setDesc(get(Opc));
1759     return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1760                                                    OpIdx1, OpIdx2);
1761   }
1762   case X86::VPTERNLOGDZrri:      case X86::VPTERNLOGDZrmi:
1763   case X86::VPTERNLOGDZ128rri:   case X86::VPTERNLOGDZ128rmi:
1764   case X86::VPTERNLOGDZ256rri:   case X86::VPTERNLOGDZ256rmi:
1765   case X86::VPTERNLOGQZrri:      case X86::VPTERNLOGQZrmi:
1766   case X86::VPTERNLOGQZ128rri:   case X86::VPTERNLOGQZ128rmi:
1767   case X86::VPTERNLOGQZ256rri:   case X86::VPTERNLOGQZ256rmi:
1768   case X86::VPTERNLOGDZrrik:
1769   case X86::VPTERNLOGDZ128rrik:
1770   case X86::VPTERNLOGDZ256rrik:
1771   case X86::VPTERNLOGQZrrik:
1772   case X86::VPTERNLOGQZ128rrik:
1773   case X86::VPTERNLOGQZ256rrik:
1774   case X86::VPTERNLOGDZrrikz:    case X86::VPTERNLOGDZrmikz:
1775   case X86::VPTERNLOGDZ128rrikz: case X86::VPTERNLOGDZ128rmikz:
1776   case X86::VPTERNLOGDZ256rrikz: case X86::VPTERNLOGDZ256rmikz:
1777   case X86::VPTERNLOGQZrrikz:    case X86::VPTERNLOGQZrmikz:
1778   case X86::VPTERNLOGQZ128rrikz: case X86::VPTERNLOGQZ128rmikz:
1779   case X86::VPTERNLOGQZ256rrikz: case X86::VPTERNLOGQZ256rmikz:
1780   case X86::VPTERNLOGDZ128rmbi:
1781   case X86::VPTERNLOGDZ256rmbi:
1782   case X86::VPTERNLOGDZrmbi:
1783   case X86::VPTERNLOGQZ128rmbi:
1784   case X86::VPTERNLOGQZ256rmbi:
1785   case X86::VPTERNLOGQZrmbi:
1786   case X86::VPTERNLOGDZ128rmbikz:
1787   case X86::VPTERNLOGDZ256rmbikz:
1788   case X86::VPTERNLOGDZrmbikz:
1789   case X86::VPTERNLOGQZ128rmbikz:
1790   case X86::VPTERNLOGQZ256rmbikz:
1791   case X86::VPTERNLOGQZrmbikz: {
1792     auto &WorkingMI = cloneIfNew(MI);
1793     commuteVPTERNLOG(WorkingMI, OpIdx1, OpIdx2);
1794     return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1795                                                    OpIdx1, OpIdx2);
1796   }
1797   default: {
1798     if (isCommutableVPERMV3Instruction(MI.getOpcode())) {
1799       unsigned Opc = getCommutedVPERMV3Opcode(MI.getOpcode());
1800       auto &WorkingMI = cloneIfNew(MI);
1801       WorkingMI.setDesc(get(Opc));
1802       return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1803                                                      OpIdx1, OpIdx2);
1804     }
1805 
1806     const X86InstrFMA3Group *FMA3Group = getFMA3Group(MI.getOpcode(),
1807                                                       MI.getDesc().TSFlags);
1808     if (FMA3Group) {
1809       unsigned Opc =
1810         getFMA3OpcodeToCommuteOperands(MI, OpIdx1, OpIdx2, *FMA3Group);
1811       auto &WorkingMI = cloneIfNew(MI);
1812       WorkingMI.setDesc(get(Opc));
1813       return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1814                                                      OpIdx1, OpIdx2);
1815     }
1816 
1817     return TargetInstrInfo::commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2);
1818   }
1819   }
1820 }
1821 
1822 bool
findThreeSrcCommutedOpIndices(const MachineInstr & MI,unsigned & SrcOpIdx1,unsigned & SrcOpIdx2,bool IsIntrinsic) const1823 X86InstrInfo::findThreeSrcCommutedOpIndices(const MachineInstr &MI,
1824                                             unsigned &SrcOpIdx1,
1825                                             unsigned &SrcOpIdx2,
1826                                             bool IsIntrinsic) const {
1827   uint64_t TSFlags = MI.getDesc().TSFlags;
1828 
1829   unsigned FirstCommutableVecOp = 1;
1830   unsigned LastCommutableVecOp = 3;
1831   unsigned KMaskOp = -1U;
1832   if (X86II::isKMasked(TSFlags)) {
1833     // For k-zero-masked operations it is Ok to commute the first vector
1834     // operand.
1835     // For regular k-masked operations a conservative choice is done as the
1836     // elements of the first vector operand, for which the corresponding bit
1837     // in the k-mask operand is set to 0, are copied to the result of the
1838     // instruction.
1839     // TODO/FIXME: The commute still may be legal if it is known that the
1840     // k-mask operand is set to either all ones or all zeroes.
1841     // It is also Ok to commute the 1st operand if all users of MI use only
1842     // the elements enabled by the k-mask operand. For example,
1843     //   v4 = VFMADD213PSZrk v1, k, v2, v3; // v1[i] = k[i] ? v2[i]*v1[i]+v3[i]
1844     //                                                     : v1[i];
1845     //   VMOVAPSZmrk <mem_addr>, k, v4; // this is the ONLY user of v4 ->
1846     //                                  // Ok, to commute v1 in FMADD213PSZrk.
1847 
1848     // The k-mask operand has index = 2 for masked and zero-masked operations.
1849     KMaskOp = 2;
1850 
1851     // The operand with index = 1 is used as a source for those elements for
1852     // which the corresponding bit in the k-mask is set to 0.
1853     if (X86II::isKMergeMasked(TSFlags))
1854       FirstCommutableVecOp = 3;
1855 
1856     LastCommutableVecOp++;
1857   } else if (IsIntrinsic) {
1858     // Commuting the first operand of an intrinsic instruction isn't possible
1859     // unless we can prove that only the lowest element of the result is used.
1860     FirstCommutableVecOp = 2;
1861   }
1862 
1863   if (isMem(MI, LastCommutableVecOp))
1864     LastCommutableVecOp--;
1865 
1866   // Only the first RegOpsNum operands are commutable.
1867   // Also, the value 'CommuteAnyOperandIndex' is valid here as it means
1868   // that the operand is not specified/fixed.
1869   if (SrcOpIdx1 != CommuteAnyOperandIndex &&
1870       (SrcOpIdx1 < FirstCommutableVecOp || SrcOpIdx1 > LastCommutableVecOp ||
1871        SrcOpIdx1 == KMaskOp))
1872     return false;
1873   if (SrcOpIdx2 != CommuteAnyOperandIndex &&
1874       (SrcOpIdx2 < FirstCommutableVecOp || SrcOpIdx2 > LastCommutableVecOp ||
1875        SrcOpIdx2 == KMaskOp))
1876     return false;
1877 
1878   // Look for two different register operands assumed to be commutable
1879   // regardless of the FMA opcode. The FMA opcode is adjusted later.
1880   if (SrcOpIdx1 == CommuteAnyOperandIndex ||
1881       SrcOpIdx2 == CommuteAnyOperandIndex) {
1882     unsigned CommutableOpIdx1 = SrcOpIdx1;
1883     unsigned CommutableOpIdx2 = SrcOpIdx2;
1884 
1885     // At least one of operands to be commuted is not specified and
1886     // this method is free to choose appropriate commutable operands.
1887     if (SrcOpIdx1 == SrcOpIdx2)
1888       // Both of operands are not fixed. By default set one of commutable
1889       // operands to the last register operand of the instruction.
1890       CommutableOpIdx2 = LastCommutableVecOp;
1891     else if (SrcOpIdx2 == CommuteAnyOperandIndex)
1892       // Only one of operands is not fixed.
1893       CommutableOpIdx2 = SrcOpIdx1;
1894 
1895     // CommutableOpIdx2 is well defined now. Let's choose another commutable
1896     // operand and assign its index to CommutableOpIdx1.
1897     unsigned Op2Reg = MI.getOperand(CommutableOpIdx2).getReg();
1898     for (CommutableOpIdx1 = LastCommutableVecOp;
1899          CommutableOpIdx1 >= FirstCommutableVecOp; CommutableOpIdx1--) {
1900       // Just ignore and skip the k-mask operand.
1901       if (CommutableOpIdx1 == KMaskOp)
1902         continue;
1903 
1904       // The commuted operands must have different registers.
1905       // Otherwise, the commute transformation does not change anything and
1906       // is useless then.
1907       if (Op2Reg != MI.getOperand(CommutableOpIdx1).getReg())
1908         break;
1909     }
1910 
1911     // No appropriate commutable operands were found.
1912     if (CommutableOpIdx1 < FirstCommutableVecOp)
1913       return false;
1914 
1915     // Assign the found pair of commutable indices to SrcOpIdx1 and SrcOpidx2
1916     // to return those values.
1917     if (!fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2,
1918                               CommutableOpIdx1, CommutableOpIdx2))
1919       return false;
1920   }
1921 
1922   return true;
1923 }
1924 
findCommutedOpIndices(MachineInstr & MI,unsigned & SrcOpIdx1,unsigned & SrcOpIdx2) const1925 bool X86InstrInfo::findCommutedOpIndices(MachineInstr &MI, unsigned &SrcOpIdx1,
1926                                          unsigned &SrcOpIdx2) const {
1927   const MCInstrDesc &Desc = MI.getDesc();
1928   if (!Desc.isCommutable())
1929     return false;
1930 
1931   switch (MI.getOpcode()) {
1932   case X86::CMPSDrr:
1933   case X86::CMPSSrr:
1934   case X86::CMPPDrri:
1935   case X86::CMPPSrri:
1936   case X86::VCMPSDrr:
1937   case X86::VCMPSSrr:
1938   case X86::VCMPPDrri:
1939   case X86::VCMPPSrri:
1940   case X86::VCMPPDYrri:
1941   case X86::VCMPPSYrri:
1942   case X86::VCMPSDZrr:
1943   case X86::VCMPSSZrr:
1944   case X86::VCMPPDZrri:
1945   case X86::VCMPPSZrri:
1946   case X86::VCMPPDZ128rri:
1947   case X86::VCMPPSZ128rri:
1948   case X86::VCMPPDZ256rri:
1949   case X86::VCMPPSZ256rri: {
1950     // Float comparison can be safely commuted for
1951     // Ordered/Unordered/Equal/NotEqual tests
1952     unsigned Imm = MI.getOperand(3).getImm() & 0x7;
1953     switch (Imm) {
1954     case 0x00: // EQUAL
1955     case 0x03: // UNORDERED
1956     case 0x04: // NOT EQUAL
1957     case 0x07: // ORDERED
1958       // The indices of the commutable operands are 1 and 2.
1959       // Assign them to the returned operand indices here.
1960       return fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2, 1, 2);
1961     }
1962     return false;
1963   }
1964   case X86::MOVSDrr:
1965   case X86::MOVSSrr:
1966   case X86::VMOVSDrr:
1967   case X86::VMOVSSrr:
1968     if (Subtarget.hasSSE41())
1969       return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2);
1970     return false;
1971   case X86::MOVHLPSrr:
1972   case X86::UNPCKHPDrr:
1973   case X86::VMOVHLPSrr:
1974   case X86::VUNPCKHPDrr:
1975   case X86::VMOVHLPSZrr:
1976   case X86::VUNPCKHPDZ128rr:
1977     if (Subtarget.hasSSE2())
1978       return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2);
1979     return false;
1980   case X86::VPTERNLOGDZrri:      case X86::VPTERNLOGDZrmi:
1981   case X86::VPTERNLOGDZ128rri:   case X86::VPTERNLOGDZ128rmi:
1982   case X86::VPTERNLOGDZ256rri:   case X86::VPTERNLOGDZ256rmi:
1983   case X86::VPTERNLOGQZrri:      case X86::VPTERNLOGQZrmi:
1984   case X86::VPTERNLOGQZ128rri:   case X86::VPTERNLOGQZ128rmi:
1985   case X86::VPTERNLOGQZ256rri:   case X86::VPTERNLOGQZ256rmi:
1986   case X86::VPTERNLOGDZrrik:
1987   case X86::VPTERNLOGDZ128rrik:
1988   case X86::VPTERNLOGDZ256rrik:
1989   case X86::VPTERNLOGQZrrik:
1990   case X86::VPTERNLOGQZ128rrik:
1991   case X86::VPTERNLOGQZ256rrik:
1992   case X86::VPTERNLOGDZrrikz:    case X86::VPTERNLOGDZrmikz:
1993   case X86::VPTERNLOGDZ128rrikz: case X86::VPTERNLOGDZ128rmikz:
1994   case X86::VPTERNLOGDZ256rrikz: case X86::VPTERNLOGDZ256rmikz:
1995   case X86::VPTERNLOGQZrrikz:    case X86::VPTERNLOGQZrmikz:
1996   case X86::VPTERNLOGQZ128rrikz: case X86::VPTERNLOGQZ128rmikz:
1997   case X86::VPTERNLOGQZ256rrikz: case X86::VPTERNLOGQZ256rmikz:
1998   case X86::VPTERNLOGDZ128rmbi:
1999   case X86::VPTERNLOGDZ256rmbi:
2000   case X86::VPTERNLOGDZrmbi:
2001   case X86::VPTERNLOGQZ128rmbi:
2002   case X86::VPTERNLOGQZ256rmbi:
2003   case X86::VPTERNLOGQZrmbi:
2004   case X86::VPTERNLOGDZ128rmbikz:
2005   case X86::VPTERNLOGDZ256rmbikz:
2006   case X86::VPTERNLOGDZrmbikz:
2007   case X86::VPTERNLOGQZ128rmbikz:
2008   case X86::VPTERNLOGQZ256rmbikz:
2009   case X86::VPTERNLOGQZrmbikz:
2010     return findThreeSrcCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2);
2011   case X86::VPMADD52HUQZ128r:
2012   case X86::VPMADD52HUQZ128rk:
2013   case X86::VPMADD52HUQZ128rkz:
2014   case X86::VPMADD52HUQZ256r:
2015   case X86::VPMADD52HUQZ256rk:
2016   case X86::VPMADD52HUQZ256rkz:
2017   case X86::VPMADD52HUQZr:
2018   case X86::VPMADD52HUQZrk:
2019   case X86::VPMADD52HUQZrkz:
2020   case X86::VPMADD52LUQZ128r:
2021   case X86::VPMADD52LUQZ128rk:
2022   case X86::VPMADD52LUQZ128rkz:
2023   case X86::VPMADD52LUQZ256r:
2024   case X86::VPMADD52LUQZ256rk:
2025   case X86::VPMADD52LUQZ256rkz:
2026   case X86::VPMADD52LUQZr:
2027   case X86::VPMADD52LUQZrk:
2028   case X86::VPMADD52LUQZrkz: {
2029     unsigned CommutableOpIdx1 = 2;
2030     unsigned CommutableOpIdx2 = 3;
2031     if (X86II::isKMasked(Desc.TSFlags)) {
2032       // Skip the mask register.
2033       ++CommutableOpIdx1;
2034       ++CommutableOpIdx2;
2035     }
2036     if (!fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2,
2037                               CommutableOpIdx1, CommutableOpIdx2))
2038       return false;
2039     if (!MI.getOperand(SrcOpIdx1).isReg() ||
2040         !MI.getOperand(SrcOpIdx2).isReg())
2041       // No idea.
2042       return false;
2043     return true;
2044   }
2045 
2046   default:
2047     const X86InstrFMA3Group *FMA3Group = getFMA3Group(MI.getOpcode(),
2048                                                       MI.getDesc().TSFlags);
2049     if (FMA3Group)
2050       return findThreeSrcCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2,
2051                                            FMA3Group->isIntrinsic());
2052 
2053     // Handled masked instructions since we need to skip over the mask input
2054     // and the preserved input.
2055     if (X86II::isKMasked(Desc.TSFlags)) {
2056       // First assume that the first input is the mask operand and skip past it.
2057       unsigned CommutableOpIdx1 = Desc.getNumDefs() + 1;
2058       unsigned CommutableOpIdx2 = Desc.getNumDefs() + 2;
2059       // Check if the first input is tied. If there isn't one then we only
2060       // need to skip the mask operand which we did above.
2061       if ((MI.getDesc().getOperandConstraint(Desc.getNumDefs(),
2062                                              MCOI::TIED_TO) != -1)) {
2063         // If this is zero masking instruction with a tied operand, we need to
2064         // move the first index back to the first input since this must
2065         // be a 3 input instruction and we want the first two non-mask inputs.
2066         // Otherwise this is a 2 input instruction with a preserved input and
2067         // mask, so we need to move the indices to skip one more input.
2068         if (X86II::isKMergeMasked(Desc.TSFlags)) {
2069           ++CommutableOpIdx1;
2070           ++CommutableOpIdx2;
2071         } else {
2072           --CommutableOpIdx1;
2073         }
2074       }
2075 
2076       if (!fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2,
2077                                 CommutableOpIdx1, CommutableOpIdx2))
2078         return false;
2079 
2080       if (!MI.getOperand(SrcOpIdx1).isReg() ||
2081           !MI.getOperand(SrcOpIdx2).isReg())
2082         // No idea.
2083         return false;
2084       return true;
2085     }
2086 
2087     return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2);
2088   }
2089   return false;
2090 }
2091 
getCondFromBranchOpc(unsigned BrOpc)2092 X86::CondCode X86::getCondFromBranchOpc(unsigned BrOpc) {
2093   switch (BrOpc) {
2094   default: return X86::COND_INVALID;
2095   case X86::JE_1:  return X86::COND_E;
2096   case X86::JNE_1: return X86::COND_NE;
2097   case X86::JL_1:  return X86::COND_L;
2098   case X86::JLE_1: return X86::COND_LE;
2099   case X86::JG_1:  return X86::COND_G;
2100   case X86::JGE_1: return X86::COND_GE;
2101   case X86::JB_1:  return X86::COND_B;
2102   case X86::JBE_1: return X86::COND_BE;
2103   case X86::JA_1:  return X86::COND_A;
2104   case X86::JAE_1: return X86::COND_AE;
2105   case X86::JS_1:  return X86::COND_S;
2106   case X86::JNS_1: return X86::COND_NS;
2107   case X86::JP_1:  return X86::COND_P;
2108   case X86::JNP_1: return X86::COND_NP;
2109   case X86::JO_1:  return X86::COND_O;
2110   case X86::JNO_1: return X86::COND_NO;
2111   }
2112 }
2113 
2114 /// Return condition code of a SET opcode.
getCondFromSETOpc(unsigned Opc)2115 X86::CondCode X86::getCondFromSETOpc(unsigned Opc) {
2116   switch (Opc) {
2117   default: return X86::COND_INVALID;
2118   case X86::SETAr:  case X86::SETAm:  return X86::COND_A;
2119   case X86::SETAEr: case X86::SETAEm: return X86::COND_AE;
2120   case X86::SETBr:  case X86::SETBm:  return X86::COND_B;
2121   case X86::SETBEr: case X86::SETBEm: return X86::COND_BE;
2122   case X86::SETEr:  case X86::SETEm:  return X86::COND_E;
2123   case X86::SETGr:  case X86::SETGm:  return X86::COND_G;
2124   case X86::SETGEr: case X86::SETGEm: return X86::COND_GE;
2125   case X86::SETLr:  case X86::SETLm:  return X86::COND_L;
2126   case X86::SETLEr: case X86::SETLEm: return X86::COND_LE;
2127   case X86::SETNEr: case X86::SETNEm: return X86::COND_NE;
2128   case X86::SETNOr: case X86::SETNOm: return X86::COND_NO;
2129   case X86::SETNPr: case X86::SETNPm: return X86::COND_NP;
2130   case X86::SETNSr: case X86::SETNSm: return X86::COND_NS;
2131   case X86::SETOr:  case X86::SETOm:  return X86::COND_O;
2132   case X86::SETPr:  case X86::SETPm:  return X86::COND_P;
2133   case X86::SETSr:  case X86::SETSm:  return X86::COND_S;
2134   }
2135 }
2136 
2137 /// Return condition code of a CMov opcode.
getCondFromCMovOpc(unsigned Opc)2138 X86::CondCode X86::getCondFromCMovOpc(unsigned Opc) {
2139   switch (Opc) {
2140   default: return X86::COND_INVALID;
2141   case X86::CMOVA16rm:  case X86::CMOVA16rr:  case X86::CMOVA32rm:
2142   case X86::CMOVA32rr:  case X86::CMOVA64rm:  case X86::CMOVA64rr:
2143     return X86::COND_A;
2144   case X86::CMOVAE16rm: case X86::CMOVAE16rr: case X86::CMOVAE32rm:
2145   case X86::CMOVAE32rr: case X86::CMOVAE64rm: case X86::CMOVAE64rr:
2146     return X86::COND_AE;
2147   case X86::CMOVB16rm:  case X86::CMOVB16rr:  case X86::CMOVB32rm:
2148   case X86::CMOVB32rr:  case X86::CMOVB64rm:  case X86::CMOVB64rr:
2149     return X86::COND_B;
2150   case X86::CMOVBE16rm: case X86::CMOVBE16rr: case X86::CMOVBE32rm:
2151   case X86::CMOVBE32rr: case X86::CMOVBE64rm: case X86::CMOVBE64rr:
2152     return X86::COND_BE;
2153   case X86::CMOVE16rm:  case X86::CMOVE16rr:  case X86::CMOVE32rm:
2154   case X86::CMOVE32rr:  case X86::CMOVE64rm:  case X86::CMOVE64rr:
2155     return X86::COND_E;
2156   case X86::CMOVG16rm:  case X86::CMOVG16rr:  case X86::CMOVG32rm:
2157   case X86::CMOVG32rr:  case X86::CMOVG64rm:  case X86::CMOVG64rr:
2158     return X86::COND_G;
2159   case X86::CMOVGE16rm: case X86::CMOVGE16rr: case X86::CMOVGE32rm:
2160   case X86::CMOVGE32rr: case X86::CMOVGE64rm: case X86::CMOVGE64rr:
2161     return X86::COND_GE;
2162   case X86::CMOVL16rm:  case X86::CMOVL16rr:  case X86::CMOVL32rm:
2163   case X86::CMOVL32rr:  case X86::CMOVL64rm:  case X86::CMOVL64rr:
2164     return X86::COND_L;
2165   case X86::CMOVLE16rm: case X86::CMOVLE16rr: case X86::CMOVLE32rm:
2166   case X86::CMOVLE32rr: case X86::CMOVLE64rm: case X86::CMOVLE64rr:
2167     return X86::COND_LE;
2168   case X86::CMOVNE16rm: case X86::CMOVNE16rr: case X86::CMOVNE32rm:
2169   case X86::CMOVNE32rr: case X86::CMOVNE64rm: case X86::CMOVNE64rr:
2170     return X86::COND_NE;
2171   case X86::CMOVNO16rm: case X86::CMOVNO16rr: case X86::CMOVNO32rm:
2172   case X86::CMOVNO32rr: case X86::CMOVNO64rm: case X86::CMOVNO64rr:
2173     return X86::COND_NO;
2174   case X86::CMOVNP16rm: case X86::CMOVNP16rr: case X86::CMOVNP32rm:
2175   case X86::CMOVNP32rr: case X86::CMOVNP64rm: case X86::CMOVNP64rr:
2176     return X86::COND_NP;
2177   case X86::CMOVNS16rm: case X86::CMOVNS16rr: case X86::CMOVNS32rm:
2178   case X86::CMOVNS32rr: case X86::CMOVNS64rm: case X86::CMOVNS64rr:
2179     return X86::COND_NS;
2180   case X86::CMOVO16rm:  case X86::CMOVO16rr:  case X86::CMOVO32rm:
2181   case X86::CMOVO32rr:  case X86::CMOVO64rm:  case X86::CMOVO64rr:
2182     return X86::COND_O;
2183   case X86::CMOVP16rm:  case X86::CMOVP16rr:  case X86::CMOVP32rm:
2184   case X86::CMOVP32rr:  case X86::CMOVP64rm:  case X86::CMOVP64rr:
2185     return X86::COND_P;
2186   case X86::CMOVS16rm:  case X86::CMOVS16rr:  case X86::CMOVS32rm:
2187   case X86::CMOVS32rr:  case X86::CMOVS64rm:  case X86::CMOVS64rr:
2188     return X86::COND_S;
2189   }
2190 }
2191 
GetCondBranchFromCond(X86::CondCode CC)2192 unsigned X86::GetCondBranchFromCond(X86::CondCode CC) {
2193   switch (CC) {
2194   default: llvm_unreachable("Illegal condition code!");
2195   case X86::COND_E:  return X86::JE_1;
2196   case X86::COND_NE: return X86::JNE_1;
2197   case X86::COND_L:  return X86::JL_1;
2198   case X86::COND_LE: return X86::JLE_1;
2199   case X86::COND_G:  return X86::JG_1;
2200   case X86::COND_GE: return X86::JGE_1;
2201   case X86::COND_B:  return X86::JB_1;
2202   case X86::COND_BE: return X86::JBE_1;
2203   case X86::COND_A:  return X86::JA_1;
2204   case X86::COND_AE: return X86::JAE_1;
2205   case X86::COND_S:  return X86::JS_1;
2206   case X86::COND_NS: return X86::JNS_1;
2207   case X86::COND_P:  return X86::JP_1;
2208   case X86::COND_NP: return X86::JNP_1;
2209   case X86::COND_O:  return X86::JO_1;
2210   case X86::COND_NO: return X86::JNO_1;
2211   }
2212 }
2213 
2214 /// Return the inverse of the specified condition,
2215 /// e.g. turning COND_E to COND_NE.
GetOppositeBranchCondition(X86::CondCode CC)2216 X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) {
2217   switch (CC) {
2218   default: llvm_unreachable("Illegal condition code!");
2219   case X86::COND_E:  return X86::COND_NE;
2220   case X86::COND_NE: return X86::COND_E;
2221   case X86::COND_L:  return X86::COND_GE;
2222   case X86::COND_LE: return X86::COND_G;
2223   case X86::COND_G:  return X86::COND_LE;
2224   case X86::COND_GE: return X86::COND_L;
2225   case X86::COND_B:  return X86::COND_AE;
2226   case X86::COND_BE: return X86::COND_A;
2227   case X86::COND_A:  return X86::COND_BE;
2228   case X86::COND_AE: return X86::COND_B;
2229   case X86::COND_S:  return X86::COND_NS;
2230   case X86::COND_NS: return X86::COND_S;
2231   case X86::COND_P:  return X86::COND_NP;
2232   case X86::COND_NP: return X86::COND_P;
2233   case X86::COND_O:  return X86::COND_NO;
2234   case X86::COND_NO: return X86::COND_O;
2235   case X86::COND_NE_OR_P:  return X86::COND_E_AND_NP;
2236   case X86::COND_E_AND_NP: return X86::COND_NE_OR_P;
2237   }
2238 }
2239 
2240 /// Assuming the flags are set by MI(a,b), return the condition code if we
2241 /// modify the instructions such that flags are set by MI(b,a).
getSwappedCondition(X86::CondCode CC)2242 static X86::CondCode getSwappedCondition(X86::CondCode CC) {
2243   switch (CC) {
2244   default: return X86::COND_INVALID;
2245   case X86::COND_E:  return X86::COND_E;
2246   case X86::COND_NE: return X86::COND_NE;
2247   case X86::COND_L:  return X86::COND_G;
2248   case X86::COND_LE: return X86::COND_GE;
2249   case X86::COND_G:  return X86::COND_L;
2250   case X86::COND_GE: return X86::COND_LE;
2251   case X86::COND_B:  return X86::COND_A;
2252   case X86::COND_BE: return X86::COND_AE;
2253   case X86::COND_A:  return X86::COND_B;
2254   case X86::COND_AE: return X86::COND_BE;
2255   }
2256 }
2257 
2258 std::pair<X86::CondCode, bool>
getX86ConditionCode(CmpInst::Predicate Predicate)2259 X86::getX86ConditionCode(CmpInst::Predicate Predicate) {
2260   X86::CondCode CC = X86::COND_INVALID;
2261   bool NeedSwap = false;
2262   switch (Predicate) {
2263   default: break;
2264   // Floating-point Predicates
2265   case CmpInst::FCMP_UEQ: CC = X86::COND_E;       break;
2266   case CmpInst::FCMP_OLT: NeedSwap = true;        LLVM_FALLTHROUGH;
2267   case CmpInst::FCMP_OGT: CC = X86::COND_A;       break;
2268   case CmpInst::FCMP_OLE: NeedSwap = true;        LLVM_FALLTHROUGH;
2269   case CmpInst::FCMP_OGE: CC = X86::COND_AE;      break;
2270   case CmpInst::FCMP_UGT: NeedSwap = true;        LLVM_FALLTHROUGH;
2271   case CmpInst::FCMP_ULT: CC = X86::COND_B;       break;
2272   case CmpInst::FCMP_UGE: NeedSwap = true;        LLVM_FALLTHROUGH;
2273   case CmpInst::FCMP_ULE: CC = X86::COND_BE;      break;
2274   case CmpInst::FCMP_ONE: CC = X86::COND_NE;      break;
2275   case CmpInst::FCMP_UNO: CC = X86::COND_P;       break;
2276   case CmpInst::FCMP_ORD: CC = X86::COND_NP;      break;
2277   case CmpInst::FCMP_OEQ:                         LLVM_FALLTHROUGH;
2278   case CmpInst::FCMP_UNE: CC = X86::COND_INVALID; break;
2279 
2280   // Integer Predicates
2281   case CmpInst::ICMP_EQ:  CC = X86::COND_E;       break;
2282   case CmpInst::ICMP_NE:  CC = X86::COND_NE;      break;
2283   case CmpInst::ICMP_UGT: CC = X86::COND_A;       break;
2284   case CmpInst::ICMP_UGE: CC = X86::COND_AE;      break;
2285   case CmpInst::ICMP_ULT: CC = X86::COND_B;       break;
2286   case CmpInst::ICMP_ULE: CC = X86::COND_BE;      break;
2287   case CmpInst::ICMP_SGT: CC = X86::COND_G;       break;
2288   case CmpInst::ICMP_SGE: CC = X86::COND_GE;      break;
2289   case CmpInst::ICMP_SLT: CC = X86::COND_L;       break;
2290   case CmpInst::ICMP_SLE: CC = X86::COND_LE;      break;
2291   }
2292 
2293   return std::make_pair(CC, NeedSwap);
2294 }
2295 
2296 /// Return a set opcode for the given condition and
2297 /// whether it has memory operand.
getSETFromCond(CondCode CC,bool HasMemoryOperand)2298 unsigned X86::getSETFromCond(CondCode CC, bool HasMemoryOperand) {
2299   static const uint16_t Opc[16][2] = {
2300     { X86::SETAr,  X86::SETAm  },
2301     { X86::SETAEr, X86::SETAEm },
2302     { X86::SETBr,  X86::SETBm  },
2303     { X86::SETBEr, X86::SETBEm },
2304     { X86::SETEr,  X86::SETEm  },
2305     { X86::SETGr,  X86::SETGm  },
2306     { X86::SETGEr, X86::SETGEm },
2307     { X86::SETLr,  X86::SETLm  },
2308     { X86::SETLEr, X86::SETLEm },
2309     { X86::SETNEr, X86::SETNEm },
2310     { X86::SETNOr, X86::SETNOm },
2311     { X86::SETNPr, X86::SETNPm },
2312     { X86::SETNSr, X86::SETNSm },
2313     { X86::SETOr,  X86::SETOm  },
2314     { X86::SETPr,  X86::SETPm  },
2315     { X86::SETSr,  X86::SETSm  }
2316   };
2317 
2318   assert(CC <= LAST_VALID_COND && "Can only handle standard cond codes");
2319   return Opc[CC][HasMemoryOperand ? 1 : 0];
2320 }
2321 
2322 /// Return a cmov opcode for the given condition,
2323 /// register size in bytes, and operand type.
getCMovFromCond(CondCode CC,unsigned RegBytes,bool HasMemoryOperand)2324 unsigned X86::getCMovFromCond(CondCode CC, unsigned RegBytes,
2325                               bool HasMemoryOperand) {
2326   static const uint16_t Opc[32][3] = {
2327     { X86::CMOVA16rr,  X86::CMOVA32rr,  X86::CMOVA64rr  },
2328     { X86::CMOVAE16rr, X86::CMOVAE32rr, X86::CMOVAE64rr },
2329     { X86::CMOVB16rr,  X86::CMOVB32rr,  X86::CMOVB64rr  },
2330     { X86::CMOVBE16rr, X86::CMOVBE32rr, X86::CMOVBE64rr },
2331     { X86::CMOVE16rr,  X86::CMOVE32rr,  X86::CMOVE64rr  },
2332     { X86::CMOVG16rr,  X86::CMOVG32rr,  X86::CMOVG64rr  },
2333     { X86::CMOVGE16rr, X86::CMOVGE32rr, X86::CMOVGE64rr },
2334     { X86::CMOVL16rr,  X86::CMOVL32rr,  X86::CMOVL64rr  },
2335     { X86::CMOVLE16rr, X86::CMOVLE32rr, X86::CMOVLE64rr },
2336     { X86::CMOVNE16rr, X86::CMOVNE32rr, X86::CMOVNE64rr },
2337     { X86::CMOVNO16rr, X86::CMOVNO32rr, X86::CMOVNO64rr },
2338     { X86::CMOVNP16rr, X86::CMOVNP32rr, X86::CMOVNP64rr },
2339     { X86::CMOVNS16rr, X86::CMOVNS32rr, X86::CMOVNS64rr },
2340     { X86::CMOVO16rr,  X86::CMOVO32rr,  X86::CMOVO64rr  },
2341     { X86::CMOVP16rr,  X86::CMOVP32rr,  X86::CMOVP64rr  },
2342     { X86::CMOVS16rr,  X86::CMOVS32rr,  X86::CMOVS64rr  },
2343     { X86::CMOVA16rm,  X86::CMOVA32rm,  X86::CMOVA64rm  },
2344     { X86::CMOVAE16rm, X86::CMOVAE32rm, X86::CMOVAE64rm },
2345     { X86::CMOVB16rm,  X86::CMOVB32rm,  X86::CMOVB64rm  },
2346     { X86::CMOVBE16rm, X86::CMOVBE32rm, X86::CMOVBE64rm },
2347     { X86::CMOVE16rm,  X86::CMOVE32rm,  X86::CMOVE64rm  },
2348     { X86::CMOVG16rm,  X86::CMOVG32rm,  X86::CMOVG64rm  },
2349     { X86::CMOVGE16rm, X86::CMOVGE32rm, X86::CMOVGE64rm },
2350     { X86::CMOVL16rm,  X86::CMOVL32rm,  X86::CMOVL64rm  },
2351     { X86::CMOVLE16rm, X86::CMOVLE32rm, X86::CMOVLE64rm },
2352     { X86::CMOVNE16rm, X86::CMOVNE32rm, X86::CMOVNE64rm },
2353     { X86::CMOVNO16rm, X86::CMOVNO32rm, X86::CMOVNO64rm },
2354     { X86::CMOVNP16rm, X86::CMOVNP32rm, X86::CMOVNP64rm },
2355     { X86::CMOVNS16rm, X86::CMOVNS32rm, X86::CMOVNS64rm },
2356     { X86::CMOVO16rm,  X86::CMOVO32rm,  X86::CMOVO64rm  },
2357     { X86::CMOVP16rm,  X86::CMOVP32rm,  X86::CMOVP64rm  },
2358     { X86::CMOVS16rm,  X86::CMOVS32rm,  X86::CMOVS64rm  }
2359   };
2360 
2361   assert(CC < 16 && "Can only handle standard cond codes");
2362   unsigned Idx = HasMemoryOperand ? 16+CC : CC;
2363   switch(RegBytes) {
2364   default: llvm_unreachable("Illegal register size!");
2365   case 2: return Opc[Idx][0];
2366   case 4: return Opc[Idx][1];
2367   case 8: return Opc[Idx][2];
2368   }
2369 }
2370 
2371 /// Get the VPCMP immediate for the given condition.
getVPCMPImmForCond(ISD::CondCode CC)2372 unsigned X86::getVPCMPImmForCond(ISD::CondCode CC) {
2373   switch (CC) {
2374   default: llvm_unreachable("Unexpected SETCC condition");
2375   case ISD::SETNE:  return 4;
2376   case ISD::SETEQ:  return 0;
2377   case ISD::SETULT:
2378   case ISD::SETLT: return 1;
2379   case ISD::SETUGT:
2380   case ISD::SETGT: return 6;
2381   case ISD::SETUGE:
2382   case ISD::SETGE: return 5;
2383   case ISD::SETULE:
2384   case ISD::SETLE: return 2;
2385   }
2386 }
2387 
2388 /// Get the VPCMP immediate if the opcodes are swapped.
getSwappedVPCMPImm(unsigned Imm)2389 unsigned X86::getSwappedVPCMPImm(unsigned Imm) {
2390   switch (Imm) {
2391   default: llvm_unreachable("Unreachable!");
2392   case 0x01: Imm = 0x06; break; // LT  -> NLE
2393   case 0x02: Imm = 0x05; break; // LE  -> NLT
2394   case 0x05: Imm = 0x02; break; // NLT -> LE
2395   case 0x06: Imm = 0x01; break; // NLE -> LT
2396   case 0x00: // EQ
2397   case 0x03: // FALSE
2398   case 0x04: // NE
2399   case 0x07: // TRUE
2400     break;
2401   }
2402 
2403   return Imm;
2404 }
2405 
2406 /// Get the VPCOM immediate if the opcodes are swapped.
getSwappedVPCOMImm(unsigned Imm)2407 unsigned X86::getSwappedVPCOMImm(unsigned Imm) {
2408   switch (Imm) {
2409   default: llvm_unreachable("Unreachable!");
2410   case 0x00: Imm = 0x02; break; // LT -> GT
2411   case 0x01: Imm = 0x03; break; // LE -> GE
2412   case 0x02: Imm = 0x00; break; // GT -> LT
2413   case 0x03: Imm = 0x01; break; // GE -> LE
2414   case 0x04: // EQ
2415   case 0x05: // NE
2416   case 0x06: // FALSE
2417   case 0x07: // TRUE
2418     break;
2419   }
2420 
2421   return Imm;
2422 }
2423 
isUnpredicatedTerminator(const MachineInstr & MI) const2424 bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr &MI) const {
2425   if (!MI.isTerminator()) return false;
2426 
2427   // Conditional branch is a special case.
2428   if (MI.isBranch() && !MI.isBarrier())
2429     return true;
2430   if (!MI.isPredicable())
2431     return true;
2432   return !isPredicated(MI);
2433 }
2434 
isUnconditionalTailCall(const MachineInstr & MI) const2435 bool X86InstrInfo::isUnconditionalTailCall(const MachineInstr &MI) const {
2436   switch (MI.getOpcode()) {
2437   case X86::TCRETURNdi:
2438   case X86::TCRETURNri:
2439   case X86::TCRETURNmi:
2440   case X86::TCRETURNdi64:
2441   case X86::TCRETURNri64:
2442   case X86::TCRETURNmi64:
2443     return true;
2444   default:
2445     return false;
2446   }
2447 }
2448 
canMakeTailCallConditional(SmallVectorImpl<MachineOperand> & BranchCond,const MachineInstr & TailCall) const2449 bool X86InstrInfo::canMakeTailCallConditional(
2450     SmallVectorImpl<MachineOperand> &BranchCond,
2451     const MachineInstr &TailCall) const {
2452   if (TailCall.getOpcode() != X86::TCRETURNdi &&
2453       TailCall.getOpcode() != X86::TCRETURNdi64) {
2454     // Only direct calls can be done with a conditional branch.
2455     return false;
2456   }
2457 
2458   const MachineFunction *MF = TailCall.getParent()->getParent();
2459   if (Subtarget.isTargetWin64() && MF->hasWinCFI()) {
2460     // Conditional tail calls confuse the Win64 unwinder.
2461     return false;
2462   }
2463 
2464   assert(BranchCond.size() == 1);
2465   if (BranchCond[0].getImm() > X86::LAST_VALID_COND) {
2466     // Can't make a conditional tail call with this condition.
2467     return false;
2468   }
2469 
2470   const X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
2471   if (X86FI->getTCReturnAddrDelta() != 0 ||
2472       TailCall.getOperand(1).getImm() != 0) {
2473     // A conditional tail call cannot do any stack adjustment.
2474     return false;
2475   }
2476 
2477   return true;
2478 }
2479 
replaceBranchWithTailCall(MachineBasicBlock & MBB,SmallVectorImpl<MachineOperand> & BranchCond,const MachineInstr & TailCall) const2480 void X86InstrInfo::replaceBranchWithTailCall(
2481     MachineBasicBlock &MBB, SmallVectorImpl<MachineOperand> &BranchCond,
2482     const MachineInstr &TailCall) const {
2483   assert(canMakeTailCallConditional(BranchCond, TailCall));
2484 
2485   MachineBasicBlock::iterator I = MBB.end();
2486   while (I != MBB.begin()) {
2487     --I;
2488     if (I->isDebugInstr())
2489       continue;
2490     if (!I->isBranch())
2491       assert(0 && "Can't find the branch to replace!");
2492 
2493     X86::CondCode CC = X86::getCondFromBranchOpc(I->getOpcode());
2494     assert(BranchCond.size() == 1);
2495     if (CC != BranchCond[0].getImm())
2496       continue;
2497 
2498     break;
2499   }
2500 
2501   unsigned Opc = TailCall.getOpcode() == X86::TCRETURNdi ? X86::TCRETURNdicc
2502                                                          : X86::TCRETURNdi64cc;
2503 
2504   auto MIB = BuildMI(MBB, I, MBB.findDebugLoc(I), get(Opc));
2505   MIB->addOperand(TailCall.getOperand(0)); // Destination.
2506   MIB.addImm(0); // Stack offset (not used).
2507   MIB->addOperand(BranchCond[0]); // Condition.
2508   MIB.copyImplicitOps(TailCall); // Regmask and (imp-used) parameters.
2509 
2510   // Add implicit uses and defs of all live regs potentially clobbered by the
2511   // call. This way they still appear live across the call.
2512   LivePhysRegs LiveRegs(getRegisterInfo());
2513   LiveRegs.addLiveOuts(MBB);
2514   SmallVector<std::pair<MCPhysReg, const MachineOperand *>, 8> Clobbers;
2515   LiveRegs.stepForward(*MIB, Clobbers);
2516   for (const auto &C : Clobbers) {
2517     MIB.addReg(C.first, RegState::Implicit);
2518     MIB.addReg(C.first, RegState::Implicit | RegState::Define);
2519   }
2520 
2521   I->eraseFromParent();
2522 }
2523 
2524 // Given a MBB and its TBB, find the FBB which was a fallthrough MBB (it may
2525 // not be a fallthrough MBB now due to layout changes). Return nullptr if the
2526 // fallthrough MBB cannot be identified.
getFallThroughMBB(MachineBasicBlock * MBB,MachineBasicBlock * TBB)2527 static MachineBasicBlock *getFallThroughMBB(MachineBasicBlock *MBB,
2528                                             MachineBasicBlock *TBB) {
2529   // Look for non-EHPad successors other than TBB. If we find exactly one, it
2530   // is the fallthrough MBB. If we find zero, then TBB is both the target MBB
2531   // and fallthrough MBB. If we find more than one, we cannot identify the
2532   // fallthrough MBB and should return nullptr.
2533   MachineBasicBlock *FallthroughBB = nullptr;
2534   for (auto SI = MBB->succ_begin(), SE = MBB->succ_end(); SI != SE; ++SI) {
2535     if ((*SI)->isEHPad() || (*SI == TBB && FallthroughBB))
2536       continue;
2537     // Return a nullptr if we found more than one fallthrough successor.
2538     if (FallthroughBB && FallthroughBB != TBB)
2539       return nullptr;
2540     FallthroughBB = *SI;
2541   }
2542   return FallthroughBB;
2543 }
2544 
AnalyzeBranchImpl(MachineBasicBlock & MBB,MachineBasicBlock * & TBB,MachineBasicBlock * & FBB,SmallVectorImpl<MachineOperand> & Cond,SmallVectorImpl<MachineInstr * > & CondBranches,bool AllowModify) const2545 bool X86InstrInfo::AnalyzeBranchImpl(
2546     MachineBasicBlock &MBB, MachineBasicBlock *&TBB, MachineBasicBlock *&FBB,
2547     SmallVectorImpl<MachineOperand> &Cond,
2548     SmallVectorImpl<MachineInstr *> &CondBranches, bool AllowModify) const {
2549 
2550   // Start from the bottom of the block and work up, examining the
2551   // terminator instructions.
2552   MachineBasicBlock::iterator I = MBB.end();
2553   MachineBasicBlock::iterator UnCondBrIter = MBB.end();
2554   while (I != MBB.begin()) {
2555     --I;
2556     if (I->isDebugInstr())
2557       continue;
2558 
2559     // Working from the bottom, when we see a non-terminator instruction, we're
2560     // done.
2561     if (!isUnpredicatedTerminator(*I))
2562       break;
2563 
2564     // A terminator that isn't a branch can't easily be handled by this
2565     // analysis.
2566     if (!I->isBranch())
2567       return true;
2568 
2569     // Handle unconditional branches.
2570     if (I->getOpcode() == X86::JMP_1) {
2571       UnCondBrIter = I;
2572 
2573       if (!AllowModify) {
2574         TBB = I->getOperand(0).getMBB();
2575         continue;
2576       }
2577 
2578       // If the block has any instructions after a JMP, delete them.
2579       while (std::next(I) != MBB.end())
2580         std::next(I)->eraseFromParent();
2581 
2582       Cond.clear();
2583       FBB = nullptr;
2584 
2585       // Delete the JMP if it's equivalent to a fall-through.
2586       if (MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) {
2587         TBB = nullptr;
2588         I->eraseFromParent();
2589         I = MBB.end();
2590         UnCondBrIter = MBB.end();
2591         continue;
2592       }
2593 
2594       // TBB is used to indicate the unconditional destination.
2595       TBB = I->getOperand(0).getMBB();
2596       continue;
2597     }
2598 
2599     // Handle conditional branches.
2600     X86::CondCode BranchCode = X86::getCondFromBranchOpc(I->getOpcode());
2601     if (BranchCode == X86::COND_INVALID)
2602       return true;  // Can't handle indirect branch.
2603 
2604     // In practice we should never have an undef eflags operand, if we do
2605     // abort here as we are not prepared to preserve the flag.
2606     if (I->getOperand(1).isUndef())
2607       return true;
2608 
2609     // Working from the bottom, handle the first conditional branch.
2610     if (Cond.empty()) {
2611       MachineBasicBlock *TargetBB = I->getOperand(0).getMBB();
2612       if (AllowModify && UnCondBrIter != MBB.end() &&
2613           MBB.isLayoutSuccessor(TargetBB)) {
2614         // If we can modify the code and it ends in something like:
2615         //
2616         //     jCC L1
2617         //     jmp L2
2618         //   L1:
2619         //     ...
2620         //   L2:
2621         //
2622         // Then we can change this to:
2623         //
2624         //     jnCC L2
2625         //   L1:
2626         //     ...
2627         //   L2:
2628         //
2629         // Which is a bit more efficient.
2630         // We conditionally jump to the fall-through block.
2631         BranchCode = GetOppositeBranchCondition(BranchCode);
2632         unsigned JNCC = GetCondBranchFromCond(BranchCode);
2633         MachineBasicBlock::iterator OldInst = I;
2634 
2635         BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(JNCC))
2636           .addMBB(UnCondBrIter->getOperand(0).getMBB());
2637         BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(X86::JMP_1))
2638           .addMBB(TargetBB);
2639 
2640         OldInst->eraseFromParent();
2641         UnCondBrIter->eraseFromParent();
2642 
2643         // Restart the analysis.
2644         UnCondBrIter = MBB.end();
2645         I = MBB.end();
2646         continue;
2647       }
2648 
2649       FBB = TBB;
2650       TBB = I->getOperand(0).getMBB();
2651       Cond.push_back(MachineOperand::CreateImm(BranchCode));
2652       CondBranches.push_back(&*I);
2653       continue;
2654     }
2655 
2656     // Handle subsequent conditional branches. Only handle the case where all
2657     // conditional branches branch to the same destination and their condition
2658     // opcodes fit one of the special multi-branch idioms.
2659     assert(Cond.size() == 1);
2660     assert(TBB);
2661 
2662     // If the conditions are the same, we can leave them alone.
2663     X86::CondCode OldBranchCode = (X86::CondCode)Cond[0].getImm();
2664     auto NewTBB = I->getOperand(0).getMBB();
2665     if (OldBranchCode == BranchCode && TBB == NewTBB)
2666       continue;
2667 
2668     // If they differ, see if they fit one of the known patterns. Theoretically,
2669     // we could handle more patterns here, but we shouldn't expect to see them
2670     // if instruction selection has done a reasonable job.
2671     if (TBB == NewTBB &&
2672                ((OldBranchCode == X86::COND_P && BranchCode == X86::COND_NE) ||
2673                 (OldBranchCode == X86::COND_NE && BranchCode == X86::COND_P))) {
2674       BranchCode = X86::COND_NE_OR_P;
2675     } else if ((OldBranchCode == X86::COND_NP && BranchCode == X86::COND_NE) ||
2676                (OldBranchCode == X86::COND_E && BranchCode == X86::COND_P)) {
2677       if (NewTBB != (FBB ? FBB : getFallThroughMBB(&MBB, TBB)))
2678         return true;
2679 
2680       // X86::COND_E_AND_NP usually has two different branch destinations.
2681       //
2682       // JP B1
2683       // JE B2
2684       // JMP B1
2685       // B1:
2686       // B2:
2687       //
2688       // Here this condition branches to B2 only if NP && E. It has another
2689       // equivalent form:
2690       //
2691       // JNE B1
2692       // JNP B2
2693       // JMP B1
2694       // B1:
2695       // B2:
2696       //
2697       // Similarly it branches to B2 only if E && NP. That is why this condition
2698       // is named with COND_E_AND_NP.
2699       BranchCode = X86::COND_E_AND_NP;
2700     } else
2701       return true;
2702 
2703     // Update the MachineOperand.
2704     Cond[0].setImm(BranchCode);
2705     CondBranches.push_back(&*I);
2706   }
2707 
2708   return false;
2709 }
2710 
analyzeBranch(MachineBasicBlock & MBB,MachineBasicBlock * & TBB,MachineBasicBlock * & FBB,SmallVectorImpl<MachineOperand> & Cond,bool AllowModify) const2711 bool X86InstrInfo::analyzeBranch(MachineBasicBlock &MBB,
2712                                  MachineBasicBlock *&TBB,
2713                                  MachineBasicBlock *&FBB,
2714                                  SmallVectorImpl<MachineOperand> &Cond,
2715                                  bool AllowModify) const {
2716   SmallVector<MachineInstr *, 4> CondBranches;
2717   return AnalyzeBranchImpl(MBB, TBB, FBB, Cond, CondBranches, AllowModify);
2718 }
2719 
analyzeBranchPredicate(MachineBasicBlock & MBB,MachineBranchPredicate & MBP,bool AllowModify) const2720 bool X86InstrInfo::analyzeBranchPredicate(MachineBasicBlock &MBB,
2721                                           MachineBranchPredicate &MBP,
2722                                           bool AllowModify) const {
2723   using namespace std::placeholders;
2724 
2725   SmallVector<MachineOperand, 4> Cond;
2726   SmallVector<MachineInstr *, 4> CondBranches;
2727   if (AnalyzeBranchImpl(MBB, MBP.TrueDest, MBP.FalseDest, Cond, CondBranches,
2728                         AllowModify))
2729     return true;
2730 
2731   if (Cond.size() != 1)
2732     return true;
2733 
2734   assert(MBP.TrueDest && "expected!");
2735 
2736   if (!MBP.FalseDest)
2737     MBP.FalseDest = MBB.getNextNode();
2738 
2739   const TargetRegisterInfo *TRI = &getRegisterInfo();
2740 
2741   MachineInstr *ConditionDef = nullptr;
2742   bool SingleUseCondition = true;
2743 
2744   for (auto I = std::next(MBB.rbegin()), E = MBB.rend(); I != E; ++I) {
2745     if (I->modifiesRegister(X86::EFLAGS, TRI)) {
2746       ConditionDef = &*I;
2747       break;
2748     }
2749 
2750     if (I->readsRegister(X86::EFLAGS, TRI))
2751       SingleUseCondition = false;
2752   }
2753 
2754   if (!ConditionDef)
2755     return true;
2756 
2757   if (SingleUseCondition) {
2758     for (auto *Succ : MBB.successors())
2759       if (Succ->isLiveIn(X86::EFLAGS))
2760         SingleUseCondition = false;
2761   }
2762 
2763   MBP.ConditionDef = ConditionDef;
2764   MBP.SingleUseCondition = SingleUseCondition;
2765 
2766   // Currently we only recognize the simple pattern:
2767   //
2768   //   test %reg, %reg
2769   //   je %label
2770   //
2771   const unsigned TestOpcode =
2772       Subtarget.is64Bit() ? X86::TEST64rr : X86::TEST32rr;
2773 
2774   if (ConditionDef->getOpcode() == TestOpcode &&
2775       ConditionDef->getNumOperands() == 3 &&
2776       ConditionDef->getOperand(0).isIdenticalTo(ConditionDef->getOperand(1)) &&
2777       (Cond[0].getImm() == X86::COND_NE || Cond[0].getImm() == X86::COND_E)) {
2778     MBP.LHS = ConditionDef->getOperand(0);
2779     MBP.RHS = MachineOperand::CreateImm(0);
2780     MBP.Predicate = Cond[0].getImm() == X86::COND_NE
2781                         ? MachineBranchPredicate::PRED_NE
2782                         : MachineBranchPredicate::PRED_EQ;
2783     return false;
2784   }
2785 
2786   return true;
2787 }
2788 
removeBranch(MachineBasicBlock & MBB,int * BytesRemoved) const2789 unsigned X86InstrInfo::removeBranch(MachineBasicBlock &MBB,
2790                                     int *BytesRemoved) const {
2791   assert(!BytesRemoved && "code size not handled");
2792 
2793   MachineBasicBlock::iterator I = MBB.end();
2794   unsigned Count = 0;
2795 
2796   while (I != MBB.begin()) {
2797     --I;
2798     if (I->isDebugInstr())
2799       continue;
2800     if (I->getOpcode() != X86::JMP_1 &&
2801         X86::getCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
2802       break;
2803     // Remove the branch.
2804     I->eraseFromParent();
2805     I = MBB.end();
2806     ++Count;
2807   }
2808 
2809   return Count;
2810 }
2811 
insertBranch(MachineBasicBlock & MBB,MachineBasicBlock * TBB,MachineBasicBlock * FBB,ArrayRef<MachineOperand> Cond,const DebugLoc & DL,int * BytesAdded) const2812 unsigned X86InstrInfo::insertBranch(MachineBasicBlock &MBB,
2813                                     MachineBasicBlock *TBB,
2814                                     MachineBasicBlock *FBB,
2815                                     ArrayRef<MachineOperand> Cond,
2816                                     const DebugLoc &DL,
2817                                     int *BytesAdded) const {
2818   // Shouldn't be a fall through.
2819   assert(TBB && "insertBranch must not be told to insert a fallthrough");
2820   assert((Cond.size() == 1 || Cond.size() == 0) &&
2821          "X86 branch conditions have one component!");
2822   assert(!BytesAdded && "code size not handled");
2823 
2824   if (Cond.empty()) {
2825     // Unconditional branch?
2826     assert(!FBB && "Unconditional branch with multiple successors!");
2827     BuildMI(&MBB, DL, get(X86::JMP_1)).addMBB(TBB);
2828     return 1;
2829   }
2830 
2831   // If FBB is null, it is implied to be a fall-through block.
2832   bool FallThru = FBB == nullptr;
2833 
2834   // Conditional branch.
2835   unsigned Count = 0;
2836   X86::CondCode CC = (X86::CondCode)Cond[0].getImm();
2837   switch (CC) {
2838   case X86::COND_NE_OR_P:
2839     // Synthesize NE_OR_P with two branches.
2840     BuildMI(&MBB, DL, get(X86::JNE_1)).addMBB(TBB);
2841     ++Count;
2842     BuildMI(&MBB, DL, get(X86::JP_1)).addMBB(TBB);
2843     ++Count;
2844     break;
2845   case X86::COND_E_AND_NP:
2846     // Use the next block of MBB as FBB if it is null.
2847     if (FBB == nullptr) {
2848       FBB = getFallThroughMBB(&MBB, TBB);
2849       assert(FBB && "MBB cannot be the last block in function when the false "
2850                     "body is a fall-through.");
2851     }
2852     // Synthesize COND_E_AND_NP with two branches.
2853     BuildMI(&MBB, DL, get(X86::JNE_1)).addMBB(FBB);
2854     ++Count;
2855     BuildMI(&MBB, DL, get(X86::JNP_1)).addMBB(TBB);
2856     ++Count;
2857     break;
2858   default: {
2859     unsigned Opc = GetCondBranchFromCond(CC);
2860     BuildMI(&MBB, DL, get(Opc)).addMBB(TBB);
2861     ++Count;
2862   }
2863   }
2864   if (!FallThru) {
2865     // Two-way Conditional branch. Insert the second branch.
2866     BuildMI(&MBB, DL, get(X86::JMP_1)).addMBB(FBB);
2867     ++Count;
2868   }
2869   return Count;
2870 }
2871 
2872 bool X86InstrInfo::
canInsertSelect(const MachineBasicBlock & MBB,ArrayRef<MachineOperand> Cond,unsigned TrueReg,unsigned FalseReg,int & CondCycles,int & TrueCycles,int & FalseCycles) const2873 canInsertSelect(const MachineBasicBlock &MBB,
2874                 ArrayRef<MachineOperand> Cond,
2875                 unsigned TrueReg, unsigned FalseReg,
2876                 int &CondCycles, int &TrueCycles, int &FalseCycles) const {
2877   // Not all subtargets have cmov instructions.
2878   if (!Subtarget.hasCMov())
2879     return false;
2880   if (Cond.size() != 1)
2881     return false;
2882   // We cannot do the composite conditions, at least not in SSA form.
2883   if ((X86::CondCode)Cond[0].getImm() > X86::COND_S)
2884     return false;
2885 
2886   // Check register classes.
2887   const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2888   const TargetRegisterClass *RC =
2889     RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg));
2890   if (!RC)
2891     return false;
2892 
2893   // We have cmov instructions for 16, 32, and 64 bit general purpose registers.
2894   if (X86::GR16RegClass.hasSubClassEq(RC) ||
2895       X86::GR32RegClass.hasSubClassEq(RC) ||
2896       X86::GR64RegClass.hasSubClassEq(RC)) {
2897     // This latency applies to Pentium M, Merom, Wolfdale, Nehalem, and Sandy
2898     // Bridge. Probably Ivy Bridge as well.
2899     CondCycles = 2;
2900     TrueCycles = 2;
2901     FalseCycles = 2;
2902     return true;
2903   }
2904 
2905   // Can't do vectors.
2906   return false;
2907 }
2908 
insertSelect(MachineBasicBlock & MBB,MachineBasicBlock::iterator I,const DebugLoc & DL,unsigned DstReg,ArrayRef<MachineOperand> Cond,unsigned TrueReg,unsigned FalseReg) const2909 void X86InstrInfo::insertSelect(MachineBasicBlock &MBB,
2910                                 MachineBasicBlock::iterator I,
2911                                 const DebugLoc &DL, unsigned DstReg,
2912                                 ArrayRef<MachineOperand> Cond, unsigned TrueReg,
2913                                 unsigned FalseReg) const {
2914   MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2915   const TargetRegisterInfo &TRI = *MRI.getTargetRegisterInfo();
2916   const TargetRegisterClass &RC = *MRI.getRegClass(DstReg);
2917   assert(Cond.size() == 1 && "Invalid Cond array");
2918   unsigned Opc = getCMovFromCond((X86::CondCode)Cond[0].getImm(),
2919                                  TRI.getRegSizeInBits(RC) / 8,
2920                                  false /*HasMemoryOperand*/);
2921   BuildMI(MBB, I, DL, get(Opc), DstReg).addReg(FalseReg).addReg(TrueReg);
2922 }
2923 
2924 /// Test if the given register is a physical h register.
isHReg(unsigned Reg)2925 static bool isHReg(unsigned Reg) {
2926   return X86::GR8_ABCD_HRegClass.contains(Reg);
2927 }
2928 
2929 // Try and copy between VR128/VR64 and GR64 registers.
CopyToFromAsymmetricReg(unsigned DestReg,unsigned SrcReg,const X86Subtarget & Subtarget)2930 static unsigned CopyToFromAsymmetricReg(unsigned DestReg, unsigned SrcReg,
2931                                         const X86Subtarget &Subtarget) {
2932   bool HasAVX = Subtarget.hasAVX();
2933   bool HasAVX512 = Subtarget.hasAVX512();
2934 
2935   // SrcReg(MaskReg) -> DestReg(GR64)
2936   // SrcReg(MaskReg) -> DestReg(GR32)
2937 
2938   // All KMASK RegClasses hold the same k registers, can be tested against anyone.
2939   if (X86::VK16RegClass.contains(SrcReg)) {
2940     if (X86::GR64RegClass.contains(DestReg)) {
2941       assert(Subtarget.hasBWI());
2942       return X86::KMOVQrk;
2943     }
2944     if (X86::GR32RegClass.contains(DestReg))
2945       return Subtarget.hasBWI() ? X86::KMOVDrk : X86::KMOVWrk;
2946   }
2947 
2948   // SrcReg(GR64) -> DestReg(MaskReg)
2949   // SrcReg(GR32) -> DestReg(MaskReg)
2950 
2951   // All KMASK RegClasses hold the same k registers, can be tested against anyone.
2952   if (X86::VK16RegClass.contains(DestReg)) {
2953     if (X86::GR64RegClass.contains(SrcReg)) {
2954       assert(Subtarget.hasBWI());
2955       return X86::KMOVQkr;
2956     }
2957     if (X86::GR32RegClass.contains(SrcReg))
2958       return Subtarget.hasBWI() ? X86::KMOVDkr : X86::KMOVWkr;
2959   }
2960 
2961 
2962   // SrcReg(VR128) -> DestReg(GR64)
2963   // SrcReg(VR64)  -> DestReg(GR64)
2964   // SrcReg(GR64)  -> DestReg(VR128)
2965   // SrcReg(GR64)  -> DestReg(VR64)
2966 
2967   if (X86::GR64RegClass.contains(DestReg)) {
2968     if (X86::VR128XRegClass.contains(SrcReg))
2969       // Copy from a VR128 register to a GR64 register.
2970       return HasAVX512 ? X86::VMOVPQIto64Zrr :
2971              HasAVX    ? X86::VMOVPQIto64rr  :
2972                          X86::MOVPQIto64rr;
2973     if (X86::VR64RegClass.contains(SrcReg))
2974       // Copy from a VR64 register to a GR64 register.
2975       return X86::MMX_MOVD64from64rr;
2976   } else if (X86::GR64RegClass.contains(SrcReg)) {
2977     // Copy from a GR64 register to a VR128 register.
2978     if (X86::VR128XRegClass.contains(DestReg))
2979       return HasAVX512 ? X86::VMOV64toPQIZrr :
2980              HasAVX    ? X86::VMOV64toPQIrr  :
2981                          X86::MOV64toPQIrr;
2982     // Copy from a GR64 register to a VR64 register.
2983     if (X86::VR64RegClass.contains(DestReg))
2984       return X86::MMX_MOVD64to64rr;
2985   }
2986 
2987   // SrcReg(FR32) -> DestReg(GR32)
2988   // SrcReg(GR32) -> DestReg(FR32)
2989 
2990   if (X86::GR32RegClass.contains(DestReg) &&
2991       X86::FR32XRegClass.contains(SrcReg))
2992     // Copy from a FR32 register to a GR32 register.
2993     return HasAVX512 ? X86::VMOVSS2DIZrr :
2994            HasAVX    ? X86::VMOVSS2DIrr  :
2995                        X86::MOVSS2DIrr;
2996 
2997   if (X86::FR32XRegClass.contains(DestReg) &&
2998       X86::GR32RegClass.contains(SrcReg))
2999     // Copy from a GR32 register to a FR32 register.
3000     return HasAVX512 ? X86::VMOVDI2SSZrr :
3001            HasAVX    ? X86::VMOVDI2SSrr  :
3002                        X86::MOVDI2SSrr;
3003   return 0;
3004 }
3005 
copyPhysReg(MachineBasicBlock & MBB,MachineBasicBlock::iterator MI,const DebugLoc & DL,unsigned DestReg,unsigned SrcReg,bool KillSrc) const3006 void X86InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
3007                                MachineBasicBlock::iterator MI,
3008                                const DebugLoc &DL, unsigned DestReg,
3009                                unsigned SrcReg, bool KillSrc) const {
3010   // First deal with the normal symmetric copies.
3011   bool HasAVX = Subtarget.hasAVX();
3012   bool HasVLX = Subtarget.hasVLX();
3013   unsigned Opc = 0;
3014   if (X86::GR64RegClass.contains(DestReg, SrcReg))
3015     Opc = X86::MOV64rr;
3016   else if (X86::GR32RegClass.contains(DestReg, SrcReg))
3017     Opc = X86::MOV32rr;
3018   else if (X86::GR16RegClass.contains(DestReg, SrcReg))
3019     Opc = X86::MOV16rr;
3020   else if (X86::GR8RegClass.contains(DestReg, SrcReg)) {
3021     // Copying to or from a physical H register on x86-64 requires a NOREX
3022     // move.  Otherwise use a normal move.
3023     if ((isHReg(DestReg) || isHReg(SrcReg)) &&
3024         Subtarget.is64Bit()) {
3025       Opc = X86::MOV8rr_NOREX;
3026       // Both operands must be encodable without an REX prefix.
3027       assert(X86::GR8_NOREXRegClass.contains(SrcReg, DestReg) &&
3028              "8-bit H register can not be copied outside GR8_NOREX");
3029     } else
3030       Opc = X86::MOV8rr;
3031   }
3032   else if (X86::VR64RegClass.contains(DestReg, SrcReg))
3033     Opc = X86::MMX_MOVQ64rr;
3034   else if (X86::VR128XRegClass.contains(DestReg, SrcReg)) {
3035     if (HasVLX)
3036       Opc = X86::VMOVAPSZ128rr;
3037     else if (X86::VR128RegClass.contains(DestReg, SrcReg))
3038       Opc = HasAVX ? X86::VMOVAPSrr : X86::MOVAPSrr;
3039     else {
3040       // If this an extended register and we don't have VLX we need to use a
3041       // 512-bit move.
3042       Opc = X86::VMOVAPSZrr;
3043       const TargetRegisterInfo *TRI = &getRegisterInfo();
3044       DestReg = TRI->getMatchingSuperReg(DestReg, X86::sub_xmm,
3045                                          &X86::VR512RegClass);
3046       SrcReg = TRI->getMatchingSuperReg(SrcReg, X86::sub_xmm,
3047                                         &X86::VR512RegClass);
3048     }
3049   } else if (X86::VR256XRegClass.contains(DestReg, SrcReg)) {
3050     if (HasVLX)
3051       Opc = X86::VMOVAPSZ256rr;
3052     else if (X86::VR256RegClass.contains(DestReg, SrcReg))
3053       Opc = X86::VMOVAPSYrr;
3054     else {
3055       // If this an extended register and we don't have VLX we need to use a
3056       // 512-bit move.
3057       Opc = X86::VMOVAPSZrr;
3058       const TargetRegisterInfo *TRI = &getRegisterInfo();
3059       DestReg = TRI->getMatchingSuperReg(DestReg, X86::sub_ymm,
3060                                          &X86::VR512RegClass);
3061       SrcReg = TRI->getMatchingSuperReg(SrcReg, X86::sub_ymm,
3062                                         &X86::VR512RegClass);
3063     }
3064   } else if (X86::VR512RegClass.contains(DestReg, SrcReg))
3065     Opc = X86::VMOVAPSZrr;
3066   // All KMASK RegClasses hold the same k registers, can be tested against anyone.
3067   else if (X86::VK16RegClass.contains(DestReg, SrcReg))
3068     Opc = Subtarget.hasBWI() ? X86::KMOVQkk : X86::KMOVWkk;
3069   if (!Opc)
3070     Opc = CopyToFromAsymmetricReg(DestReg, SrcReg, Subtarget);
3071 
3072   if (Opc) {
3073     BuildMI(MBB, MI, DL, get(Opc), DestReg)
3074       .addReg(SrcReg, getKillRegState(KillSrc));
3075     return;
3076   }
3077 
3078   if (SrcReg == X86::EFLAGS || DestReg == X86::EFLAGS) {
3079     // FIXME: We use a fatal error here because historically LLVM has tried
3080     // lower some of these physreg copies and we want to ensure we get
3081     // reasonable bug reports if someone encounters a case no other testing
3082     // found. This path should be removed after the LLVM 7 release.
3083     report_fatal_error("Unable to copy EFLAGS physical register!");
3084   }
3085 
3086   LLVM_DEBUG(dbgs() << "Cannot copy " << RI.getName(SrcReg) << " to "
3087                     << RI.getName(DestReg) << '\n');
3088   report_fatal_error("Cannot emit physreg copy instruction");
3089 }
3090 
isCopyInstrImpl(const MachineInstr & MI,const MachineOperand * & Src,const MachineOperand * & Dest) const3091 bool X86InstrInfo::isCopyInstrImpl(const MachineInstr &MI,
3092                                    const MachineOperand *&Src,
3093                                    const MachineOperand *&Dest) const {
3094   if (MI.isMoveReg()) {
3095     Dest = &MI.getOperand(0);
3096     Src = &MI.getOperand(1);
3097     return true;
3098   }
3099   return false;
3100 }
3101 
getLoadStoreRegOpcode(unsigned Reg,const TargetRegisterClass * RC,bool isStackAligned,const X86Subtarget & STI,bool load)3102 static unsigned getLoadStoreRegOpcode(unsigned Reg,
3103                                       const TargetRegisterClass *RC,
3104                                       bool isStackAligned,
3105                                       const X86Subtarget &STI,
3106                                       bool load) {
3107   bool HasAVX = STI.hasAVX();
3108   bool HasAVX512 = STI.hasAVX512();
3109   bool HasVLX = STI.hasVLX();
3110 
3111   switch (STI.getRegisterInfo()->getSpillSize(*RC)) {
3112   default:
3113     llvm_unreachable("Unknown spill size");
3114   case 1:
3115     assert(X86::GR8RegClass.hasSubClassEq(RC) && "Unknown 1-byte regclass");
3116     if (STI.is64Bit())
3117       // Copying to or from a physical H register on x86-64 requires a NOREX
3118       // move.  Otherwise use a normal move.
3119       if (isHReg(Reg) || X86::GR8_ABCD_HRegClass.hasSubClassEq(RC))
3120         return load ? X86::MOV8rm_NOREX : X86::MOV8mr_NOREX;
3121     return load ? X86::MOV8rm : X86::MOV8mr;
3122   case 2:
3123     if (X86::VK16RegClass.hasSubClassEq(RC))
3124       return load ? X86::KMOVWkm : X86::KMOVWmk;
3125     assert(X86::GR16RegClass.hasSubClassEq(RC) && "Unknown 2-byte regclass");
3126     return load ? X86::MOV16rm : X86::MOV16mr;
3127   case 4:
3128     if (X86::GR32RegClass.hasSubClassEq(RC))
3129       return load ? X86::MOV32rm : X86::MOV32mr;
3130     if (X86::FR32XRegClass.hasSubClassEq(RC))
3131       return load ?
3132         (HasAVX512 ? X86::VMOVSSZrm : HasAVX ? X86::VMOVSSrm : X86::MOVSSrm) :
3133         (HasAVX512 ? X86::VMOVSSZmr : HasAVX ? X86::VMOVSSmr : X86::MOVSSmr);
3134     if (X86::RFP32RegClass.hasSubClassEq(RC))
3135       return load ? X86::LD_Fp32m : X86::ST_Fp32m;
3136     if (X86::VK32RegClass.hasSubClassEq(RC)) {
3137       assert(STI.hasBWI() && "KMOVD requires BWI");
3138       return load ? X86::KMOVDkm : X86::KMOVDmk;
3139     }
3140     llvm_unreachable("Unknown 4-byte regclass");
3141   case 8:
3142     if (X86::GR64RegClass.hasSubClassEq(RC))
3143       return load ? X86::MOV64rm : X86::MOV64mr;
3144     if (X86::FR64XRegClass.hasSubClassEq(RC))
3145       return load ?
3146         (HasAVX512 ? X86::VMOVSDZrm : HasAVX ? X86::VMOVSDrm : X86::MOVSDrm) :
3147         (HasAVX512 ? X86::VMOVSDZmr : HasAVX ? X86::VMOVSDmr : X86::MOVSDmr);
3148     if (X86::VR64RegClass.hasSubClassEq(RC))
3149       return load ? X86::MMX_MOVQ64rm : X86::MMX_MOVQ64mr;
3150     if (X86::RFP64RegClass.hasSubClassEq(RC))
3151       return load ? X86::LD_Fp64m : X86::ST_Fp64m;
3152     if (X86::VK64RegClass.hasSubClassEq(RC)) {
3153       assert(STI.hasBWI() && "KMOVQ requires BWI");
3154       return load ? X86::KMOVQkm : X86::KMOVQmk;
3155     }
3156     llvm_unreachable("Unknown 8-byte regclass");
3157   case 10:
3158     assert(X86::RFP80RegClass.hasSubClassEq(RC) && "Unknown 10-byte regclass");
3159     return load ? X86::LD_Fp80m : X86::ST_FpP80m;
3160   case 16: {
3161     if (X86::VR128XRegClass.hasSubClassEq(RC)) {
3162       // If stack is realigned we can use aligned stores.
3163       if (isStackAligned)
3164         return load ?
3165           (HasVLX    ? X86::VMOVAPSZ128rm :
3166            HasAVX512 ? X86::VMOVAPSZ128rm_NOVLX :
3167            HasAVX    ? X86::VMOVAPSrm :
3168                        X86::MOVAPSrm):
3169           (HasVLX    ? X86::VMOVAPSZ128mr :
3170            HasAVX512 ? X86::VMOVAPSZ128mr_NOVLX :
3171            HasAVX    ? X86::VMOVAPSmr :
3172                        X86::MOVAPSmr);
3173       else
3174         return load ?
3175           (HasVLX    ? X86::VMOVUPSZ128rm :
3176            HasAVX512 ? X86::VMOVUPSZ128rm_NOVLX :
3177            HasAVX    ? X86::VMOVUPSrm :
3178                        X86::MOVUPSrm):
3179           (HasVLX    ? X86::VMOVUPSZ128mr :
3180            HasAVX512 ? X86::VMOVUPSZ128mr_NOVLX :
3181            HasAVX    ? X86::VMOVUPSmr :
3182                        X86::MOVUPSmr);
3183     }
3184     if (X86::BNDRRegClass.hasSubClassEq(RC)) {
3185       if (STI.is64Bit())
3186         return load ? X86::BNDMOV64rm : X86::BNDMOV64mr;
3187       else
3188         return load ? X86::BNDMOV32rm : X86::BNDMOV32mr;
3189     }
3190     llvm_unreachable("Unknown 16-byte regclass");
3191   }
3192   case 32:
3193     assert(X86::VR256XRegClass.hasSubClassEq(RC) && "Unknown 32-byte regclass");
3194     // If stack is realigned we can use aligned stores.
3195     if (isStackAligned)
3196       return load ?
3197         (HasVLX    ? X86::VMOVAPSZ256rm :
3198          HasAVX512 ? X86::VMOVAPSZ256rm_NOVLX :
3199                      X86::VMOVAPSYrm) :
3200         (HasVLX    ? X86::VMOVAPSZ256mr :
3201          HasAVX512 ? X86::VMOVAPSZ256mr_NOVLX :
3202                      X86::VMOVAPSYmr);
3203     else
3204       return load ?
3205         (HasVLX    ? X86::VMOVUPSZ256rm :
3206          HasAVX512 ? X86::VMOVUPSZ256rm_NOVLX :
3207                      X86::VMOVUPSYrm) :
3208         (HasVLX    ? X86::VMOVUPSZ256mr :
3209          HasAVX512 ? X86::VMOVUPSZ256mr_NOVLX :
3210                      X86::VMOVUPSYmr);
3211   case 64:
3212     assert(X86::VR512RegClass.hasSubClassEq(RC) && "Unknown 64-byte regclass");
3213     assert(STI.hasAVX512() && "Using 512-bit register requires AVX512");
3214     if (isStackAligned)
3215       return load ? X86::VMOVAPSZrm : X86::VMOVAPSZmr;
3216     else
3217       return load ? X86::VMOVUPSZrm : X86::VMOVUPSZmr;
3218   }
3219 }
3220 
getMemOperandWithOffset(MachineInstr & MemOp,MachineOperand * & BaseOp,int64_t & Offset,const TargetRegisterInfo * TRI) const3221 bool X86InstrInfo::getMemOperandWithOffset(
3222     MachineInstr &MemOp, MachineOperand *&BaseOp, int64_t &Offset,
3223     const TargetRegisterInfo *TRI) const {
3224   const MCInstrDesc &Desc = MemOp.getDesc();
3225   int MemRefBegin = X86II::getMemoryOperandNo(Desc.TSFlags);
3226   if (MemRefBegin < 0)
3227     return false;
3228 
3229   MemRefBegin += X86II::getOperandBias(Desc);
3230 
3231   BaseOp = &MemOp.getOperand(MemRefBegin + X86::AddrBaseReg);
3232   if (!BaseOp->isReg()) // Can be an MO_FrameIndex
3233     return false;
3234 
3235   if (MemOp.getOperand(MemRefBegin + X86::AddrScaleAmt).getImm() != 1)
3236     return false;
3237 
3238   if (MemOp.getOperand(MemRefBegin + X86::AddrIndexReg).getReg() !=
3239       X86::NoRegister)
3240     return false;
3241 
3242   const MachineOperand &DispMO = MemOp.getOperand(MemRefBegin + X86::AddrDisp);
3243 
3244   // Displacement can be symbolic
3245   if (!DispMO.isImm())
3246     return false;
3247 
3248   Offset = DispMO.getImm();
3249 
3250   assert(BaseOp->isReg() && "getMemOperandWithOffset only supports base "
3251                             "operands of type register.");
3252   return true;
3253 }
3254 
getStoreRegOpcode(unsigned SrcReg,const TargetRegisterClass * RC,bool isStackAligned,const X86Subtarget & STI)3255 static unsigned getStoreRegOpcode(unsigned SrcReg,
3256                                   const TargetRegisterClass *RC,
3257                                   bool isStackAligned,
3258                                   const X86Subtarget &STI) {
3259   return getLoadStoreRegOpcode(SrcReg, RC, isStackAligned, STI, false);
3260 }
3261 
3262 
getLoadRegOpcode(unsigned DestReg,const TargetRegisterClass * RC,bool isStackAligned,const X86Subtarget & STI)3263 static unsigned getLoadRegOpcode(unsigned DestReg,
3264                                  const TargetRegisterClass *RC,
3265                                  bool isStackAligned,
3266                                  const X86Subtarget &STI) {
3267   return getLoadStoreRegOpcode(DestReg, RC, isStackAligned, STI, true);
3268 }
3269 
storeRegToStackSlot(MachineBasicBlock & MBB,MachineBasicBlock::iterator MI,unsigned SrcReg,bool isKill,int FrameIdx,const TargetRegisterClass * RC,const TargetRegisterInfo * TRI) const3270 void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
3271                                        MachineBasicBlock::iterator MI,
3272                                        unsigned SrcReg, bool isKill, int FrameIdx,
3273                                        const TargetRegisterClass *RC,
3274                                        const TargetRegisterInfo *TRI) const {
3275   const MachineFunction &MF = *MBB.getParent();
3276   assert(MF.getFrameInfo().getObjectSize(FrameIdx) >= TRI->getSpillSize(*RC) &&
3277          "Stack slot too small for store");
3278   unsigned Alignment = std::max<uint32_t>(TRI->getSpillSize(*RC), 16);
3279   bool isAligned =
3280       (Subtarget.getFrameLowering()->getStackAlignment() >= Alignment) ||
3281       RI.canRealignStack(MF);
3282   unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, Subtarget);
3283   addFrameReference(BuildMI(MBB, MI, DebugLoc(), get(Opc)), FrameIdx)
3284     .addReg(SrcReg, getKillRegState(isKill));
3285 }
3286 
storeRegToAddr(MachineFunction & MF,unsigned SrcReg,bool isKill,SmallVectorImpl<MachineOperand> & Addr,const TargetRegisterClass * RC,ArrayRef<MachineMemOperand * > MMOs,SmallVectorImpl<MachineInstr * > & NewMIs) const3287 void X86InstrInfo::storeRegToAddr(
3288     MachineFunction &MF, unsigned SrcReg, bool isKill,
3289     SmallVectorImpl<MachineOperand> &Addr, const TargetRegisterClass *RC,
3290     ArrayRef<MachineMemOperand *> MMOs,
3291     SmallVectorImpl<MachineInstr *> &NewMIs) const {
3292   const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
3293   unsigned Alignment = std::max<uint32_t>(TRI.getSpillSize(*RC), 16);
3294   bool isAligned = !MMOs.empty() && MMOs.front()->getAlignment() >= Alignment;
3295   unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, Subtarget);
3296   DebugLoc DL;
3297   MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc));
3298   for (unsigned i = 0, e = Addr.size(); i != e; ++i)
3299     MIB.add(Addr[i]);
3300   MIB.addReg(SrcReg, getKillRegState(isKill));
3301   MIB.setMemRefs(MMOs);
3302   NewMIs.push_back(MIB);
3303 }
3304 
3305 
loadRegFromStackSlot(MachineBasicBlock & MBB,MachineBasicBlock::iterator MI,unsigned DestReg,int FrameIdx,const TargetRegisterClass * RC,const TargetRegisterInfo * TRI) const3306 void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
3307                                         MachineBasicBlock::iterator MI,
3308                                         unsigned DestReg, int FrameIdx,
3309                                         const TargetRegisterClass *RC,
3310                                         const TargetRegisterInfo *TRI) const {
3311   const MachineFunction &MF = *MBB.getParent();
3312   unsigned Alignment = std::max<uint32_t>(TRI->getSpillSize(*RC), 16);
3313   bool isAligned =
3314       (Subtarget.getFrameLowering()->getStackAlignment() >= Alignment) ||
3315       RI.canRealignStack(MF);
3316   unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, Subtarget);
3317   addFrameReference(BuildMI(MBB, MI, DebugLoc(), get(Opc), DestReg), FrameIdx);
3318 }
3319 
loadRegFromAddr(MachineFunction & MF,unsigned DestReg,SmallVectorImpl<MachineOperand> & Addr,const TargetRegisterClass * RC,ArrayRef<MachineMemOperand * > MMOs,SmallVectorImpl<MachineInstr * > & NewMIs) const3320 void X86InstrInfo::loadRegFromAddr(
3321     MachineFunction &MF, unsigned DestReg,
3322     SmallVectorImpl<MachineOperand> &Addr, const TargetRegisterClass *RC,
3323     ArrayRef<MachineMemOperand *> MMOs,
3324     SmallVectorImpl<MachineInstr *> &NewMIs) const {
3325   const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
3326   unsigned Alignment = std::max<uint32_t>(TRI.getSpillSize(*RC), 16);
3327   bool isAligned = !MMOs.empty() && MMOs.front()->getAlignment() >= Alignment;
3328   unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, Subtarget);
3329   DebugLoc DL;
3330   MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc), DestReg);
3331   for (unsigned i = 0, e = Addr.size(); i != e; ++i)
3332     MIB.add(Addr[i]);
3333   MIB.setMemRefs(MMOs);
3334   NewMIs.push_back(MIB);
3335 }
3336 
analyzeCompare(const MachineInstr & MI,unsigned & SrcReg,unsigned & SrcReg2,int & CmpMask,int & CmpValue) const3337 bool X86InstrInfo::analyzeCompare(const MachineInstr &MI, unsigned &SrcReg,
3338                                   unsigned &SrcReg2, int &CmpMask,
3339                                   int &CmpValue) const {
3340   switch (MI.getOpcode()) {
3341   default: break;
3342   case X86::CMP64ri32:
3343   case X86::CMP64ri8:
3344   case X86::CMP32ri:
3345   case X86::CMP32ri8:
3346   case X86::CMP16ri:
3347   case X86::CMP16ri8:
3348   case X86::CMP8ri:
3349     SrcReg = MI.getOperand(0).getReg();
3350     SrcReg2 = 0;
3351     if (MI.getOperand(1).isImm()) {
3352       CmpMask = ~0;
3353       CmpValue = MI.getOperand(1).getImm();
3354     } else {
3355       CmpMask = CmpValue = 0;
3356     }
3357     return true;
3358   // A SUB can be used to perform comparison.
3359   case X86::SUB64rm:
3360   case X86::SUB32rm:
3361   case X86::SUB16rm:
3362   case X86::SUB8rm:
3363     SrcReg = MI.getOperand(1).getReg();
3364     SrcReg2 = 0;
3365     CmpMask = 0;
3366     CmpValue = 0;
3367     return true;
3368   case X86::SUB64rr:
3369   case X86::SUB32rr:
3370   case X86::SUB16rr:
3371   case X86::SUB8rr:
3372     SrcReg = MI.getOperand(1).getReg();
3373     SrcReg2 = MI.getOperand(2).getReg();
3374     CmpMask = 0;
3375     CmpValue = 0;
3376     return true;
3377   case X86::SUB64ri32:
3378   case X86::SUB64ri8:
3379   case X86::SUB32ri:
3380   case X86::SUB32ri8:
3381   case X86::SUB16ri:
3382   case X86::SUB16ri8:
3383   case X86::SUB8ri:
3384     SrcReg = MI.getOperand(1).getReg();
3385     SrcReg2 = 0;
3386     if (MI.getOperand(2).isImm()) {
3387       CmpMask = ~0;
3388       CmpValue = MI.getOperand(2).getImm();
3389     } else {
3390       CmpMask = CmpValue = 0;
3391     }
3392     return true;
3393   case X86::CMP64rr:
3394   case X86::CMP32rr:
3395   case X86::CMP16rr:
3396   case X86::CMP8rr:
3397     SrcReg = MI.getOperand(0).getReg();
3398     SrcReg2 = MI.getOperand(1).getReg();
3399     CmpMask = 0;
3400     CmpValue = 0;
3401     return true;
3402   case X86::TEST8rr:
3403   case X86::TEST16rr:
3404   case X86::TEST32rr:
3405   case X86::TEST64rr:
3406     SrcReg = MI.getOperand(0).getReg();
3407     if (MI.getOperand(1).getReg() != SrcReg)
3408       return false;
3409     // Compare against zero.
3410     SrcReg2 = 0;
3411     CmpMask = ~0;
3412     CmpValue = 0;
3413     return true;
3414   }
3415   return false;
3416 }
3417 
3418 /// Check whether the first instruction, whose only
3419 /// purpose is to update flags, can be made redundant.
3420 /// CMPrr can be made redundant by SUBrr if the operands are the same.
3421 /// This function can be extended later on.
3422 /// SrcReg, SrcRegs: register operands for FlagI.
3423 /// ImmValue: immediate for FlagI if it takes an immediate.
isRedundantFlagInstr(const MachineInstr & FlagI,unsigned SrcReg,unsigned SrcReg2,int ImmMask,int ImmValue,const MachineInstr & OI)3424 inline static bool isRedundantFlagInstr(const MachineInstr &FlagI,
3425                                         unsigned SrcReg, unsigned SrcReg2,
3426                                         int ImmMask, int ImmValue,
3427                                         const MachineInstr &OI) {
3428   if (((FlagI.getOpcode() == X86::CMP64rr && OI.getOpcode() == X86::SUB64rr) ||
3429        (FlagI.getOpcode() == X86::CMP32rr && OI.getOpcode() == X86::SUB32rr) ||
3430        (FlagI.getOpcode() == X86::CMP16rr && OI.getOpcode() == X86::SUB16rr) ||
3431        (FlagI.getOpcode() == X86::CMP8rr && OI.getOpcode() == X86::SUB8rr)) &&
3432       ((OI.getOperand(1).getReg() == SrcReg &&
3433         OI.getOperand(2).getReg() == SrcReg2) ||
3434        (OI.getOperand(1).getReg() == SrcReg2 &&
3435         OI.getOperand(2).getReg() == SrcReg)))
3436     return true;
3437 
3438   if (ImmMask != 0 &&
3439       ((FlagI.getOpcode() == X86::CMP64ri32 &&
3440         OI.getOpcode() == X86::SUB64ri32) ||
3441        (FlagI.getOpcode() == X86::CMP64ri8 &&
3442         OI.getOpcode() == X86::SUB64ri8) ||
3443        (FlagI.getOpcode() == X86::CMP32ri && OI.getOpcode() == X86::SUB32ri) ||
3444        (FlagI.getOpcode() == X86::CMP32ri8 &&
3445         OI.getOpcode() == X86::SUB32ri8) ||
3446        (FlagI.getOpcode() == X86::CMP16ri && OI.getOpcode() == X86::SUB16ri) ||
3447        (FlagI.getOpcode() == X86::CMP16ri8 &&
3448         OI.getOpcode() == X86::SUB16ri8) ||
3449        (FlagI.getOpcode() == X86::CMP8ri && OI.getOpcode() == X86::SUB8ri)) &&
3450       OI.getOperand(1).getReg() == SrcReg &&
3451       OI.getOperand(2).getImm() == ImmValue)
3452     return true;
3453   return false;
3454 }
3455 
3456 /// Check whether the definition can be converted
3457 /// to remove a comparison against zero.
isDefConvertible(const MachineInstr & MI,bool & NoSignFlag)3458 inline static bool isDefConvertible(const MachineInstr &MI, bool &NoSignFlag) {
3459   NoSignFlag = false;
3460 
3461   switch (MI.getOpcode()) {
3462   default: return false;
3463 
3464   // The shift instructions only modify ZF if their shift count is non-zero.
3465   // N.B.: The processor truncates the shift count depending on the encoding.
3466   case X86::SAR8ri:    case X86::SAR16ri:  case X86::SAR32ri:case X86::SAR64ri:
3467   case X86::SHR8ri:    case X86::SHR16ri:  case X86::SHR32ri:case X86::SHR64ri:
3468      return getTruncatedShiftCount(MI, 2) != 0;
3469 
3470   // Some left shift instructions can be turned into LEA instructions but only
3471   // if their flags aren't used. Avoid transforming such instructions.
3472   case X86::SHL8ri:    case X86::SHL16ri:  case X86::SHL32ri:case X86::SHL64ri:{
3473     unsigned ShAmt = getTruncatedShiftCount(MI, 2);
3474     if (isTruncatedShiftCountForLEA(ShAmt)) return false;
3475     return ShAmt != 0;
3476   }
3477 
3478   case X86::SHRD16rri8:case X86::SHRD32rri8:case X86::SHRD64rri8:
3479   case X86::SHLD16rri8:case X86::SHLD32rri8:case X86::SHLD64rri8:
3480      return getTruncatedShiftCount(MI, 3) != 0;
3481 
3482   case X86::SUB64ri32: case X86::SUB64ri8: case X86::SUB32ri:
3483   case X86::SUB32ri8:  case X86::SUB16ri:  case X86::SUB16ri8:
3484   case X86::SUB8ri:    case X86::SUB64rr:  case X86::SUB32rr:
3485   case X86::SUB16rr:   case X86::SUB8rr:   case X86::SUB64rm:
3486   case X86::SUB32rm:   case X86::SUB16rm:  case X86::SUB8rm:
3487   case X86::DEC64r:    case X86::DEC32r:   case X86::DEC16r: case X86::DEC8r:
3488   case X86::ADD64ri32: case X86::ADD64ri8: case X86::ADD32ri:
3489   case X86::ADD32ri8:  case X86::ADD16ri:  case X86::ADD16ri8:
3490   case X86::ADD8ri:    case X86::ADD64rr:  case X86::ADD32rr:
3491   case X86::ADD16rr:   case X86::ADD8rr:   case X86::ADD64rm:
3492   case X86::ADD32rm:   case X86::ADD16rm:  case X86::ADD8rm:
3493   case X86::INC64r:    case X86::INC32r:   case X86::INC16r: case X86::INC8r:
3494   case X86::AND64ri32: case X86::AND64ri8: case X86::AND32ri:
3495   case X86::AND32ri8:  case X86::AND16ri:  case X86::AND16ri8:
3496   case X86::AND8ri:    case X86::AND64rr:  case X86::AND32rr:
3497   case X86::AND16rr:   case X86::AND8rr:   case X86::AND64rm:
3498   case X86::AND32rm:   case X86::AND16rm:  case X86::AND8rm:
3499   case X86::XOR64ri32: case X86::XOR64ri8: case X86::XOR32ri:
3500   case X86::XOR32ri8:  case X86::XOR16ri:  case X86::XOR16ri8:
3501   case X86::XOR8ri:    case X86::XOR64rr:  case X86::XOR32rr:
3502   case X86::XOR16rr:   case X86::XOR8rr:   case X86::XOR64rm:
3503   case X86::XOR32rm:   case X86::XOR16rm:  case X86::XOR8rm:
3504   case X86::OR64ri32:  case X86::OR64ri8:  case X86::OR32ri:
3505   case X86::OR32ri8:   case X86::OR16ri:   case X86::OR16ri8:
3506   case X86::OR8ri:     case X86::OR64rr:   case X86::OR32rr:
3507   case X86::OR16rr:    case X86::OR8rr:    case X86::OR64rm:
3508   case X86::OR32rm:    case X86::OR16rm:   case X86::OR8rm:
3509   case X86::ADC64ri32: case X86::ADC64ri8: case X86::ADC32ri:
3510   case X86::ADC32ri8:  case X86::ADC16ri:  case X86::ADC16ri8:
3511   case X86::ADC8ri:    case X86::ADC64rr:  case X86::ADC32rr:
3512   case X86::ADC16rr:   case X86::ADC8rr:   case X86::ADC64rm:
3513   case X86::ADC32rm:   case X86::ADC16rm:  case X86::ADC8rm:
3514   case X86::SBB64ri32: case X86::SBB64ri8: case X86::SBB32ri:
3515   case X86::SBB32ri8:  case X86::SBB16ri:  case X86::SBB16ri8:
3516   case X86::SBB8ri:    case X86::SBB64rr:  case X86::SBB32rr:
3517   case X86::SBB16rr:   case X86::SBB8rr:   case X86::SBB64rm:
3518   case X86::SBB32rm:   case X86::SBB16rm:  case X86::SBB8rm:
3519   case X86::NEG8r:     case X86::NEG16r:   case X86::NEG32r: case X86::NEG64r:
3520   case X86::SAR8r1:    case X86::SAR16r1:  case X86::SAR32r1:case X86::SAR64r1:
3521   case X86::SHR8r1:    case X86::SHR16r1:  case X86::SHR32r1:case X86::SHR64r1:
3522   case X86::SHL8r1:    case X86::SHL16r1:  case X86::SHL32r1:case X86::SHL64r1:
3523   case X86::ANDN32rr:  case X86::ANDN32rm:
3524   case X86::ANDN64rr:  case X86::ANDN64rm:
3525   case X86::BLSI32rr:  case X86::BLSI32rm:
3526   case X86::BLSI64rr:  case X86::BLSI64rm:
3527   case X86::BLSMSK32rr:case X86::BLSMSK32rm:
3528   case X86::BLSMSK64rr:case X86::BLSMSK64rm:
3529   case X86::BLSR32rr:  case X86::BLSR32rm:
3530   case X86::BLSR64rr:  case X86::BLSR64rm:
3531   case X86::BZHI32rr:  case X86::BZHI32rm:
3532   case X86::BZHI64rr:  case X86::BZHI64rm:
3533   case X86::LZCNT16rr: case X86::LZCNT16rm:
3534   case X86::LZCNT32rr: case X86::LZCNT32rm:
3535   case X86::LZCNT64rr: case X86::LZCNT64rm:
3536   case X86::POPCNT16rr:case X86::POPCNT16rm:
3537   case X86::POPCNT32rr:case X86::POPCNT32rm:
3538   case X86::POPCNT64rr:case X86::POPCNT64rm:
3539   case X86::TZCNT16rr: case X86::TZCNT16rm:
3540   case X86::TZCNT32rr: case X86::TZCNT32rm:
3541   case X86::TZCNT64rr: case X86::TZCNT64rm:
3542   case X86::BLCFILL32rr: case X86::BLCFILL32rm:
3543   case X86::BLCFILL64rr: case X86::BLCFILL64rm:
3544   case X86::BLCI32rr:    case X86::BLCI32rm:
3545   case X86::BLCI64rr:    case X86::BLCI64rm:
3546   case X86::BLCIC32rr:   case X86::BLCIC32rm:
3547   case X86::BLCIC64rr:   case X86::BLCIC64rm:
3548   case X86::BLCMSK32rr:  case X86::BLCMSK32rm:
3549   case X86::BLCMSK64rr:  case X86::BLCMSK64rm:
3550   case X86::BLCS32rr:    case X86::BLCS32rm:
3551   case X86::BLCS64rr:    case X86::BLCS64rm:
3552   case X86::BLSFILL32rr: case X86::BLSFILL32rm:
3553   case X86::BLSFILL64rr: case X86::BLSFILL64rm:
3554   case X86::BLSIC32rr:   case X86::BLSIC32rm:
3555   case X86::BLSIC64rr:   case X86::BLSIC64rm:
3556   case X86::T1MSKC32rr:  case X86::T1MSKC32rm:
3557   case X86::T1MSKC64rr:  case X86::T1MSKC64rm:
3558   case X86::TZMSK32rr:   case X86::TZMSK32rm:
3559   case X86::TZMSK64rr:   case X86::TZMSK64rm:
3560     return true;
3561   case X86::BEXTR32rr:   case X86::BEXTR64rr:
3562   case X86::BEXTR32rm:   case X86::BEXTR64rm:
3563   case X86::BEXTRI32ri:  case X86::BEXTRI32mi:
3564   case X86::BEXTRI64ri:  case X86::BEXTRI64mi:
3565     // BEXTR doesn't update the sign flag so we can't use it.
3566     NoSignFlag = true;
3567     return true;
3568   }
3569 }
3570 
3571 /// Check whether the use can be converted to remove a comparison against zero.
isUseDefConvertible(const MachineInstr & MI)3572 static X86::CondCode isUseDefConvertible(const MachineInstr &MI) {
3573   switch (MI.getOpcode()) {
3574   default: return X86::COND_INVALID;
3575   case X86::LZCNT16rr: case X86::LZCNT16rm:
3576   case X86::LZCNT32rr: case X86::LZCNT32rm:
3577   case X86::LZCNT64rr: case X86::LZCNT64rm:
3578     return X86::COND_B;
3579   case X86::POPCNT16rr:case X86::POPCNT16rm:
3580   case X86::POPCNT32rr:case X86::POPCNT32rm:
3581   case X86::POPCNT64rr:case X86::POPCNT64rm:
3582     return X86::COND_E;
3583   case X86::TZCNT16rr: case X86::TZCNT16rm:
3584   case X86::TZCNT32rr: case X86::TZCNT32rm:
3585   case X86::TZCNT64rr: case X86::TZCNT64rm:
3586     return X86::COND_B;
3587   case X86::BSF16rr: case X86::BSF16rm:
3588   case X86::BSF32rr: case X86::BSF32rm:
3589   case X86::BSF64rr: case X86::BSF64rm:
3590   case X86::BSR16rr: case X86::BSR16rm:
3591   case X86::BSR32rr: case X86::BSR32rm:
3592   case X86::BSR64rr: case X86::BSR64rm:
3593     return X86::COND_E;
3594   }
3595 }
3596 
3597 /// Check if there exists an earlier instruction that
3598 /// operates on the same source operands and sets flags in the same way as
3599 /// Compare; remove Compare if possible.
optimizeCompareInstr(MachineInstr & CmpInstr,unsigned SrcReg,unsigned SrcReg2,int CmpMask,int CmpValue,const MachineRegisterInfo * MRI) const3600 bool X86InstrInfo::optimizeCompareInstr(MachineInstr &CmpInstr, unsigned SrcReg,
3601                                         unsigned SrcReg2, int CmpMask,
3602                                         int CmpValue,
3603                                         const MachineRegisterInfo *MRI) const {
3604   // Check whether we can replace SUB with CMP.
3605   unsigned NewOpcode = 0;
3606   switch (CmpInstr.getOpcode()) {
3607   default: break;
3608   case X86::SUB64ri32:
3609   case X86::SUB64ri8:
3610   case X86::SUB32ri:
3611   case X86::SUB32ri8:
3612   case X86::SUB16ri:
3613   case X86::SUB16ri8:
3614   case X86::SUB8ri:
3615   case X86::SUB64rm:
3616   case X86::SUB32rm:
3617   case X86::SUB16rm:
3618   case X86::SUB8rm:
3619   case X86::SUB64rr:
3620   case X86::SUB32rr:
3621   case X86::SUB16rr:
3622   case X86::SUB8rr: {
3623     if (!MRI->use_nodbg_empty(CmpInstr.getOperand(0).getReg()))
3624       return false;
3625     // There is no use of the destination register, we can replace SUB with CMP.
3626     switch (CmpInstr.getOpcode()) {
3627     default: llvm_unreachable("Unreachable!");
3628     case X86::SUB64rm:   NewOpcode = X86::CMP64rm;   break;
3629     case X86::SUB32rm:   NewOpcode = X86::CMP32rm;   break;
3630     case X86::SUB16rm:   NewOpcode = X86::CMP16rm;   break;
3631     case X86::SUB8rm:    NewOpcode = X86::CMP8rm;    break;
3632     case X86::SUB64rr:   NewOpcode = X86::CMP64rr;   break;
3633     case X86::SUB32rr:   NewOpcode = X86::CMP32rr;   break;
3634     case X86::SUB16rr:   NewOpcode = X86::CMP16rr;   break;
3635     case X86::SUB8rr:    NewOpcode = X86::CMP8rr;    break;
3636     case X86::SUB64ri32: NewOpcode = X86::CMP64ri32; break;
3637     case X86::SUB64ri8:  NewOpcode = X86::CMP64ri8;  break;
3638     case X86::SUB32ri:   NewOpcode = X86::CMP32ri;   break;
3639     case X86::SUB32ri8:  NewOpcode = X86::CMP32ri8;  break;
3640     case X86::SUB16ri:   NewOpcode = X86::CMP16ri;   break;
3641     case X86::SUB16ri8:  NewOpcode = X86::CMP16ri8;  break;
3642     case X86::SUB8ri:    NewOpcode = X86::CMP8ri;    break;
3643     }
3644     CmpInstr.setDesc(get(NewOpcode));
3645     CmpInstr.RemoveOperand(0);
3646     // Fall through to optimize Cmp if Cmp is CMPrr or CMPri.
3647     if (NewOpcode == X86::CMP64rm || NewOpcode == X86::CMP32rm ||
3648         NewOpcode == X86::CMP16rm || NewOpcode == X86::CMP8rm)
3649       return false;
3650   }
3651   }
3652 
3653   // Get the unique definition of SrcReg.
3654   MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg);
3655   if (!MI) return false;
3656 
3657   // CmpInstr is the first instruction of the BB.
3658   MachineBasicBlock::iterator I = CmpInstr, Def = MI;
3659 
3660   // If we are comparing against zero, check whether we can use MI to update
3661   // EFLAGS. If MI is not in the same BB as CmpInstr, do not optimize.
3662   bool IsCmpZero = (CmpMask != 0 && CmpValue == 0);
3663   if (IsCmpZero && MI->getParent() != CmpInstr.getParent())
3664     return false;
3665 
3666   // If we have a use of the source register between the def and our compare
3667   // instruction we can eliminate the compare iff the use sets EFLAGS in the
3668   // right way.
3669   bool ShouldUpdateCC = false;
3670   bool NoSignFlag = false;
3671   X86::CondCode NewCC = X86::COND_INVALID;
3672   if (IsCmpZero && !isDefConvertible(*MI, NoSignFlag)) {
3673     // Scan forward from the use until we hit the use we're looking for or the
3674     // compare instruction.
3675     for (MachineBasicBlock::iterator J = MI;; ++J) {
3676       // Do we have a convertible instruction?
3677       NewCC = isUseDefConvertible(*J);
3678       if (NewCC != X86::COND_INVALID && J->getOperand(1).isReg() &&
3679           J->getOperand(1).getReg() == SrcReg) {
3680         assert(J->definesRegister(X86::EFLAGS) && "Must be an EFLAGS def!");
3681         ShouldUpdateCC = true; // Update CC later on.
3682         // This is not a def of SrcReg, but still a def of EFLAGS. Keep going
3683         // with the new def.
3684         Def = J;
3685         MI = &*Def;
3686         break;
3687       }
3688 
3689       if (J == I)
3690         return false;
3691     }
3692   }
3693 
3694   // We are searching for an earlier instruction that can make CmpInstr
3695   // redundant and that instruction will be saved in Sub.
3696   MachineInstr *Sub = nullptr;
3697   const TargetRegisterInfo *TRI = &getRegisterInfo();
3698 
3699   // We iterate backward, starting from the instruction before CmpInstr and
3700   // stop when reaching the definition of a source register or done with the BB.
3701   // RI points to the instruction before CmpInstr.
3702   // If the definition is in this basic block, RE points to the definition;
3703   // otherwise, RE is the rend of the basic block.
3704   MachineBasicBlock::reverse_iterator
3705       RI = ++I.getReverse(),
3706       RE = CmpInstr.getParent() == MI->getParent()
3707                ? Def.getReverse() /* points to MI */
3708                : CmpInstr.getParent()->rend();
3709   MachineInstr *Movr0Inst = nullptr;
3710   for (; RI != RE; ++RI) {
3711     MachineInstr &Instr = *RI;
3712     // Check whether CmpInstr can be made redundant by the current instruction.
3713     if (!IsCmpZero && isRedundantFlagInstr(CmpInstr, SrcReg, SrcReg2, CmpMask,
3714                                            CmpValue, Instr)) {
3715       Sub = &Instr;
3716       break;
3717     }
3718 
3719     if (Instr.modifiesRegister(X86::EFLAGS, TRI) ||
3720         Instr.readsRegister(X86::EFLAGS, TRI)) {
3721       // This instruction modifies or uses EFLAGS.
3722 
3723       // MOV32r0 etc. are implemented with xor which clobbers condition code.
3724       // They are safe to move up, if the definition to EFLAGS is dead and
3725       // earlier instructions do not read or write EFLAGS.
3726       if (!Movr0Inst && Instr.getOpcode() == X86::MOV32r0 &&
3727           Instr.registerDefIsDead(X86::EFLAGS, TRI)) {
3728         Movr0Inst = &Instr;
3729         continue;
3730       }
3731 
3732       // We can't remove CmpInstr.
3733       return false;
3734     }
3735   }
3736 
3737   // Return false if no candidates exist.
3738   if (!IsCmpZero && !Sub)
3739     return false;
3740 
3741   bool IsSwapped = (SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 &&
3742                     Sub->getOperand(2).getReg() == SrcReg);
3743 
3744   // Scan forward from the instruction after CmpInstr for uses of EFLAGS.
3745   // It is safe to remove CmpInstr if EFLAGS is redefined or killed.
3746   // If we are done with the basic block, we need to check whether EFLAGS is
3747   // live-out.
3748   bool IsSafe = false;
3749   SmallVector<std::pair<MachineInstr*, unsigned /*NewOpc*/>, 4> OpsToUpdate;
3750   MachineBasicBlock::iterator E = CmpInstr.getParent()->end();
3751   for (++I; I != E; ++I) {
3752     const MachineInstr &Instr = *I;
3753     bool ModifyEFLAGS = Instr.modifiesRegister(X86::EFLAGS, TRI);
3754     bool UseEFLAGS = Instr.readsRegister(X86::EFLAGS, TRI);
3755     // We should check the usage if this instruction uses and updates EFLAGS.
3756     if (!UseEFLAGS && ModifyEFLAGS) {
3757       // It is safe to remove CmpInstr if EFLAGS is updated again.
3758       IsSafe = true;
3759       break;
3760     }
3761     if (!UseEFLAGS && !ModifyEFLAGS)
3762       continue;
3763 
3764     // EFLAGS is used by this instruction.
3765     X86::CondCode OldCC = X86::COND_INVALID;
3766     bool OpcIsSET = false;
3767     if (IsCmpZero || IsSwapped) {
3768       // We decode the condition code from opcode.
3769       if (Instr.isBranch())
3770         OldCC = X86::getCondFromBranchOpc(Instr.getOpcode());
3771       else {
3772         OldCC = X86::getCondFromSETOpc(Instr.getOpcode());
3773         if (OldCC != X86::COND_INVALID)
3774           OpcIsSET = true;
3775         else
3776           OldCC = X86::getCondFromCMovOpc(Instr.getOpcode());
3777       }
3778       if (OldCC == X86::COND_INVALID) return false;
3779     }
3780     X86::CondCode ReplacementCC = X86::COND_INVALID;
3781     if (IsCmpZero) {
3782       switch (OldCC) {
3783       default: break;
3784       case X86::COND_A: case X86::COND_AE:
3785       case X86::COND_B: case X86::COND_BE:
3786       case X86::COND_G: case X86::COND_GE:
3787       case X86::COND_L: case X86::COND_LE:
3788       case X86::COND_O: case X86::COND_NO:
3789         // CF and OF are used, we can't perform this optimization.
3790         return false;
3791       case X86::COND_S: case X86::COND_NS:
3792         // If SF is used, but the instruction doesn't update the SF, then we
3793         // can't do the optimization.
3794         if (NoSignFlag)
3795           return false;
3796         break;
3797       }
3798 
3799       // If we're updating the condition code check if we have to reverse the
3800       // condition.
3801       if (ShouldUpdateCC)
3802         switch (OldCC) {
3803         default:
3804           return false;
3805         case X86::COND_E:
3806           ReplacementCC = NewCC;
3807           break;
3808         case X86::COND_NE:
3809           ReplacementCC = GetOppositeBranchCondition(NewCC);
3810           break;
3811         }
3812     } else if (IsSwapped) {
3813       // If we have SUB(r1, r2) and CMP(r2, r1), the condition code needs
3814       // to be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc.
3815       // We swap the condition code and synthesize the new opcode.
3816       ReplacementCC = getSwappedCondition(OldCC);
3817       if (ReplacementCC == X86::COND_INVALID) return false;
3818     }
3819 
3820     if ((ShouldUpdateCC || IsSwapped) && ReplacementCC != OldCC) {
3821       // Synthesize the new opcode.
3822       bool HasMemoryOperand = Instr.hasOneMemOperand();
3823       unsigned NewOpc;
3824       if (Instr.isBranch())
3825         NewOpc = GetCondBranchFromCond(ReplacementCC);
3826       else if(OpcIsSET)
3827         NewOpc = getSETFromCond(ReplacementCC, HasMemoryOperand);
3828       else {
3829         unsigned DstReg = Instr.getOperand(0).getReg();
3830         const TargetRegisterClass *DstRC = MRI->getRegClass(DstReg);
3831         NewOpc = getCMovFromCond(ReplacementCC, TRI->getRegSizeInBits(*DstRC)/8,
3832                                  HasMemoryOperand);
3833       }
3834 
3835       // Push the MachineInstr to OpsToUpdate.
3836       // If it is safe to remove CmpInstr, the condition code of these
3837       // instructions will be modified.
3838       OpsToUpdate.push_back(std::make_pair(&*I, NewOpc));
3839     }
3840     if (ModifyEFLAGS || Instr.killsRegister(X86::EFLAGS, TRI)) {
3841       // It is safe to remove CmpInstr if EFLAGS is updated again or killed.
3842       IsSafe = true;
3843       break;
3844     }
3845   }
3846 
3847   // If EFLAGS is not killed nor re-defined, we should check whether it is
3848   // live-out. If it is live-out, do not optimize.
3849   if ((IsCmpZero || IsSwapped) && !IsSafe) {
3850     MachineBasicBlock *MBB = CmpInstr.getParent();
3851     for (MachineBasicBlock *Successor : MBB->successors())
3852       if (Successor->isLiveIn(X86::EFLAGS))
3853         return false;
3854   }
3855 
3856   // The instruction to be updated is either Sub or MI.
3857   Sub = IsCmpZero ? MI : Sub;
3858   // Move Movr0Inst to the appropriate place before Sub.
3859   if (Movr0Inst) {
3860     // Look backwards until we find a def that doesn't use the current EFLAGS.
3861     Def = Sub;
3862     MachineBasicBlock::reverse_iterator InsertI = Def.getReverse(),
3863                                         InsertE = Sub->getParent()->rend();
3864     for (; InsertI != InsertE; ++InsertI) {
3865       MachineInstr *Instr = &*InsertI;
3866       if (!Instr->readsRegister(X86::EFLAGS, TRI) &&
3867           Instr->modifiesRegister(X86::EFLAGS, TRI)) {
3868         Sub->getParent()->remove(Movr0Inst);
3869         Instr->getParent()->insert(MachineBasicBlock::iterator(Instr),
3870                                    Movr0Inst);
3871         break;
3872       }
3873     }
3874     if (InsertI == InsertE)
3875       return false;
3876   }
3877 
3878   // Make sure Sub instruction defines EFLAGS and mark the def live.
3879   unsigned i = 0, e = Sub->getNumOperands();
3880   for (; i != e; ++i) {
3881     MachineOperand &MO = Sub->getOperand(i);
3882     if (MO.isReg() && MO.isDef() && MO.getReg() == X86::EFLAGS) {
3883       MO.setIsDead(false);
3884       break;
3885     }
3886   }
3887   assert(i != e && "Unable to locate a def EFLAGS operand");
3888 
3889   CmpInstr.eraseFromParent();
3890 
3891   // Modify the condition code of instructions in OpsToUpdate.
3892   for (auto &Op : OpsToUpdate)
3893     Op.first->setDesc(get(Op.second));
3894   return true;
3895 }
3896 
3897 /// Try to remove the load by folding it to a register
3898 /// operand at the use. We fold the load instructions if load defines a virtual
3899 /// register, the virtual register is used once in the same BB, and the
3900 /// instructions in-between do not load or store, and have no side effects.
optimizeLoadInstr(MachineInstr & MI,const MachineRegisterInfo * MRI,unsigned & FoldAsLoadDefReg,MachineInstr * & DefMI) const3901 MachineInstr *X86InstrInfo::optimizeLoadInstr(MachineInstr &MI,
3902                                               const MachineRegisterInfo *MRI,
3903                                               unsigned &FoldAsLoadDefReg,
3904                                               MachineInstr *&DefMI) const {
3905   // Check whether we can move DefMI here.
3906   DefMI = MRI->getVRegDef(FoldAsLoadDefReg);
3907   assert(DefMI);
3908   bool SawStore = false;
3909   if (!DefMI->isSafeToMove(nullptr, SawStore))
3910     return nullptr;
3911 
3912   // Collect information about virtual register operands of MI.
3913   SmallVector<unsigned, 1> SrcOperandIds;
3914   for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
3915     MachineOperand &MO = MI.getOperand(i);
3916     if (!MO.isReg())
3917       continue;
3918     unsigned Reg = MO.getReg();
3919     if (Reg != FoldAsLoadDefReg)
3920       continue;
3921     // Do not fold if we have a subreg use or a def.
3922     if (MO.getSubReg() || MO.isDef())
3923       return nullptr;
3924     SrcOperandIds.push_back(i);
3925   }
3926   if (SrcOperandIds.empty())
3927     return nullptr;
3928 
3929   // Check whether we can fold the def into SrcOperandId.
3930   if (MachineInstr *FoldMI = foldMemoryOperand(MI, SrcOperandIds, *DefMI)) {
3931     FoldAsLoadDefReg = 0;
3932     return FoldMI;
3933   }
3934 
3935   return nullptr;
3936 }
3937 
3938 /// Expand a single-def pseudo instruction to a two-addr
3939 /// instruction with two undef reads of the register being defined.
3940 /// This is used for mapping:
3941 ///   %xmm4 = V_SET0
3942 /// to:
3943 ///   %xmm4 = PXORrr undef %xmm4, undef %xmm4
3944 ///
Expand2AddrUndef(MachineInstrBuilder & MIB,const MCInstrDesc & Desc)3945 static bool Expand2AddrUndef(MachineInstrBuilder &MIB,
3946                              const MCInstrDesc &Desc) {
3947   assert(Desc.getNumOperands() == 3 && "Expected two-addr instruction.");
3948   unsigned Reg = MIB->getOperand(0).getReg();
3949   MIB->setDesc(Desc);
3950 
3951   // MachineInstr::addOperand() will insert explicit operands before any
3952   // implicit operands.
3953   MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
3954   // But we don't trust that.
3955   assert(MIB->getOperand(1).getReg() == Reg &&
3956          MIB->getOperand(2).getReg() == Reg && "Misplaced operand");
3957   return true;
3958 }
3959 
3960 /// Expand a single-def pseudo instruction to a two-addr
3961 /// instruction with two %k0 reads.
3962 /// This is used for mapping:
3963 ///   %k4 = K_SET1
3964 /// to:
3965 ///   %k4 = KXNORrr %k0, %k0
Expand2AddrKreg(MachineInstrBuilder & MIB,const MCInstrDesc & Desc,unsigned Reg)3966 static bool Expand2AddrKreg(MachineInstrBuilder &MIB,
3967                             const MCInstrDesc &Desc, unsigned Reg) {
3968   assert(Desc.getNumOperands() == 3 && "Expected two-addr instruction.");
3969   MIB->setDesc(Desc);
3970   MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
3971   return true;
3972 }
3973 
expandMOV32r1(MachineInstrBuilder & MIB,const TargetInstrInfo & TII,bool MinusOne)3974 static bool expandMOV32r1(MachineInstrBuilder &MIB, const TargetInstrInfo &TII,
3975                           bool MinusOne) {
3976   MachineBasicBlock &MBB = *MIB->getParent();
3977   DebugLoc DL = MIB->getDebugLoc();
3978   unsigned Reg = MIB->getOperand(0).getReg();
3979 
3980   // Insert the XOR.
3981   BuildMI(MBB, MIB.getInstr(), DL, TII.get(X86::XOR32rr), Reg)
3982       .addReg(Reg, RegState::Undef)
3983       .addReg(Reg, RegState::Undef);
3984 
3985   // Turn the pseudo into an INC or DEC.
3986   MIB->setDesc(TII.get(MinusOne ? X86::DEC32r : X86::INC32r));
3987   MIB.addReg(Reg);
3988 
3989   return true;
3990 }
3991 
ExpandMOVImmSExti8(MachineInstrBuilder & MIB,const TargetInstrInfo & TII,const X86Subtarget & Subtarget)3992 static bool ExpandMOVImmSExti8(MachineInstrBuilder &MIB,
3993                                const TargetInstrInfo &TII,
3994                                const X86Subtarget &Subtarget) {
3995   MachineBasicBlock &MBB = *MIB->getParent();
3996   DebugLoc DL = MIB->getDebugLoc();
3997   int64_t Imm = MIB->getOperand(1).getImm();
3998   assert(Imm != 0 && "Using push/pop for 0 is not efficient.");
3999   MachineBasicBlock::iterator I = MIB.getInstr();
4000 
4001   int StackAdjustment;
4002 
4003   if (Subtarget.is64Bit()) {
4004     assert(MIB->getOpcode() == X86::MOV64ImmSExti8 ||
4005            MIB->getOpcode() == X86::MOV32ImmSExti8);
4006 
4007     // Can't use push/pop lowering if the function might write to the red zone.
4008     X86MachineFunctionInfo *X86FI =
4009         MBB.getParent()->getInfo<X86MachineFunctionInfo>();
4010     if (X86FI->getUsesRedZone()) {
4011       MIB->setDesc(TII.get(MIB->getOpcode() ==
4012                            X86::MOV32ImmSExti8 ? X86::MOV32ri : X86::MOV64ri));
4013       return true;
4014     }
4015 
4016     // 64-bit mode doesn't have 32-bit push/pop, so use 64-bit operations and
4017     // widen the register if necessary.
4018     StackAdjustment = 8;
4019     BuildMI(MBB, I, DL, TII.get(X86::PUSH64i8)).addImm(Imm);
4020     MIB->setDesc(TII.get(X86::POP64r));
4021     MIB->getOperand(0)
4022         .setReg(getX86SubSuperRegister(MIB->getOperand(0).getReg(), 64));
4023   } else {
4024     assert(MIB->getOpcode() == X86::MOV32ImmSExti8);
4025     StackAdjustment = 4;
4026     BuildMI(MBB, I, DL, TII.get(X86::PUSH32i8)).addImm(Imm);
4027     MIB->setDesc(TII.get(X86::POP32r));
4028   }
4029 
4030   // Build CFI if necessary.
4031   MachineFunction &MF = *MBB.getParent();
4032   const X86FrameLowering *TFL = Subtarget.getFrameLowering();
4033   bool IsWin64Prologue = MF.getTarget().getMCAsmInfo()->usesWindowsCFI();
4034   bool NeedsDwarfCFI =
4035       !IsWin64Prologue &&
4036       (MF.getMMI().hasDebugInfo() || MF.getFunction().needsUnwindTableEntry());
4037   bool EmitCFI = !TFL->hasFP(MF) && NeedsDwarfCFI;
4038   if (EmitCFI) {
4039     TFL->BuildCFI(MBB, I, DL,
4040         MCCFIInstruction::createAdjustCfaOffset(nullptr, StackAdjustment));
4041     TFL->BuildCFI(MBB, std::next(I), DL,
4042         MCCFIInstruction::createAdjustCfaOffset(nullptr, -StackAdjustment));
4043   }
4044 
4045   return true;
4046 }
4047 
4048 // LoadStackGuard has so far only been implemented for 64-bit MachO. Different
4049 // code sequence is needed for other targets.
expandLoadStackGuard(MachineInstrBuilder & MIB,const TargetInstrInfo & TII)4050 static void expandLoadStackGuard(MachineInstrBuilder &MIB,
4051                                  const TargetInstrInfo &TII) {
4052   MachineBasicBlock &MBB = *MIB->getParent();
4053   DebugLoc DL = MIB->getDebugLoc();
4054   unsigned Reg = MIB->getOperand(0).getReg();
4055   const GlobalValue *GV =
4056       cast<GlobalValue>((*MIB->memoperands_begin())->getValue());
4057   auto Flags = MachineMemOperand::MOLoad |
4058                MachineMemOperand::MODereferenceable |
4059                MachineMemOperand::MOInvariant;
4060   MachineMemOperand *MMO = MBB.getParent()->getMachineMemOperand(
4061       MachinePointerInfo::getGOT(*MBB.getParent()), Flags, 8, 8);
4062   MachineBasicBlock::iterator I = MIB.getInstr();
4063 
4064   BuildMI(MBB, I, DL, TII.get(X86::MOV64rm), Reg).addReg(X86::RIP).addImm(1)
4065       .addReg(0).addGlobalAddress(GV, 0, X86II::MO_GOTPCREL).addReg(0)
4066       .addMemOperand(MMO);
4067   MIB->setDebugLoc(DL);
4068   MIB->setDesc(TII.get(X86::MOV64rm));
4069   MIB.addReg(Reg, RegState::Kill).addImm(1).addReg(0).addImm(0).addReg(0);
4070 }
4071 
expandXorFP(MachineInstrBuilder & MIB,const TargetInstrInfo & TII)4072 static bool expandXorFP(MachineInstrBuilder &MIB, const TargetInstrInfo &TII) {
4073   MachineBasicBlock &MBB = *MIB->getParent();
4074   MachineFunction &MF = *MBB.getParent();
4075   const X86Subtarget &Subtarget = MF.getSubtarget<X86Subtarget>();
4076   const X86RegisterInfo *TRI = Subtarget.getRegisterInfo();
4077   unsigned XorOp =
4078       MIB->getOpcode() == X86::XOR64_FP ? X86::XOR64rr : X86::XOR32rr;
4079   MIB->setDesc(TII.get(XorOp));
4080   MIB.addReg(TRI->getFrameRegister(MF), RegState::Undef);
4081   return true;
4082 }
4083 
4084 // This is used to handle spills for 128/256-bit registers when we have AVX512,
4085 // but not VLX. If it uses an extended register we need to use an instruction
4086 // that loads the lower 128/256-bit, but is available with only AVX512F.
expandNOVLXLoad(MachineInstrBuilder & MIB,const TargetRegisterInfo * TRI,const MCInstrDesc & LoadDesc,const MCInstrDesc & BroadcastDesc,unsigned SubIdx)4087 static bool expandNOVLXLoad(MachineInstrBuilder &MIB,
4088                             const TargetRegisterInfo *TRI,
4089                             const MCInstrDesc &LoadDesc,
4090                             const MCInstrDesc &BroadcastDesc,
4091                             unsigned SubIdx) {
4092   unsigned DestReg = MIB->getOperand(0).getReg();
4093   // Check if DestReg is XMM16-31 or YMM16-31.
4094   if (TRI->getEncodingValue(DestReg) < 16) {
4095     // We can use a normal VEX encoded load.
4096     MIB->setDesc(LoadDesc);
4097   } else {
4098     // Use a 128/256-bit VBROADCAST instruction.
4099     MIB->setDesc(BroadcastDesc);
4100     // Change the destination to a 512-bit register.
4101     DestReg = TRI->getMatchingSuperReg(DestReg, SubIdx, &X86::VR512RegClass);
4102     MIB->getOperand(0).setReg(DestReg);
4103   }
4104   return true;
4105 }
4106 
4107 // This is used to handle spills for 128/256-bit registers when we have AVX512,
4108 // but not VLX. If it uses an extended register we need to use an instruction
4109 // that stores the lower 128/256-bit, but is available with only AVX512F.
expandNOVLXStore(MachineInstrBuilder & MIB,const TargetRegisterInfo * TRI,const MCInstrDesc & StoreDesc,const MCInstrDesc & ExtractDesc,unsigned SubIdx)4110 static bool expandNOVLXStore(MachineInstrBuilder &MIB,
4111                              const TargetRegisterInfo *TRI,
4112                              const MCInstrDesc &StoreDesc,
4113                              const MCInstrDesc &ExtractDesc,
4114                              unsigned SubIdx) {
4115   unsigned SrcReg = MIB->getOperand(X86::AddrNumOperands).getReg();
4116   // Check if DestReg is XMM16-31 or YMM16-31.
4117   if (TRI->getEncodingValue(SrcReg) < 16) {
4118     // We can use a normal VEX encoded store.
4119     MIB->setDesc(StoreDesc);
4120   } else {
4121     // Use a VEXTRACTF instruction.
4122     MIB->setDesc(ExtractDesc);
4123     // Change the destination to a 512-bit register.
4124     SrcReg = TRI->getMatchingSuperReg(SrcReg, SubIdx, &X86::VR512RegClass);
4125     MIB->getOperand(X86::AddrNumOperands).setReg(SrcReg);
4126     MIB.addImm(0x0); // Append immediate to extract from the lower bits.
4127   }
4128 
4129   return true;
4130 }
expandPostRAPseudo(MachineInstr & MI) const4131 bool X86InstrInfo::expandPostRAPseudo(MachineInstr &MI) const {
4132   bool HasAVX = Subtarget.hasAVX();
4133   MachineInstrBuilder MIB(*MI.getParent()->getParent(), MI);
4134   switch (MI.getOpcode()) {
4135   case X86::MOV32r0:
4136     return Expand2AddrUndef(MIB, get(X86::XOR32rr));
4137   case X86::MOV32r1:
4138     return expandMOV32r1(MIB, *this, /*MinusOne=*/ false);
4139   case X86::MOV32r_1:
4140     return expandMOV32r1(MIB, *this, /*MinusOne=*/ true);
4141   case X86::MOV32ImmSExti8:
4142   case X86::MOV64ImmSExti8:
4143     return ExpandMOVImmSExti8(MIB, *this, Subtarget);
4144   case X86::SETB_C8r:
4145     return Expand2AddrUndef(MIB, get(X86::SBB8rr));
4146   case X86::SETB_C16r:
4147     return Expand2AddrUndef(MIB, get(X86::SBB16rr));
4148   case X86::SETB_C32r:
4149     return Expand2AddrUndef(MIB, get(X86::SBB32rr));
4150   case X86::SETB_C64r:
4151     return Expand2AddrUndef(MIB, get(X86::SBB64rr));
4152   case X86::MMX_SET0:
4153     return Expand2AddrUndef(MIB, get(X86::MMX_PXORirr));
4154   case X86::V_SET0:
4155   case X86::FsFLD0SS:
4156   case X86::FsFLD0SD:
4157     return Expand2AddrUndef(MIB, get(HasAVX ? X86::VXORPSrr : X86::XORPSrr));
4158   case X86::AVX_SET0: {
4159     assert(HasAVX && "AVX not supported");
4160     const TargetRegisterInfo *TRI = &getRegisterInfo();
4161     unsigned SrcReg = MIB->getOperand(0).getReg();
4162     unsigned XReg = TRI->getSubReg(SrcReg, X86::sub_xmm);
4163     MIB->getOperand(0).setReg(XReg);
4164     Expand2AddrUndef(MIB, get(X86::VXORPSrr));
4165     MIB.addReg(SrcReg, RegState::ImplicitDefine);
4166     return true;
4167   }
4168   case X86::AVX512_128_SET0:
4169   case X86::AVX512_FsFLD0SS:
4170   case X86::AVX512_FsFLD0SD: {
4171     bool HasVLX = Subtarget.hasVLX();
4172     unsigned SrcReg = MIB->getOperand(0).getReg();
4173     const TargetRegisterInfo *TRI = &getRegisterInfo();
4174     if (HasVLX || TRI->getEncodingValue(SrcReg) < 16)
4175       return Expand2AddrUndef(MIB,
4176                               get(HasVLX ? X86::VPXORDZ128rr : X86::VXORPSrr));
4177     // Extended register without VLX. Use a larger XOR.
4178     SrcReg =
4179         TRI->getMatchingSuperReg(SrcReg, X86::sub_xmm, &X86::VR512RegClass);
4180     MIB->getOperand(0).setReg(SrcReg);
4181     return Expand2AddrUndef(MIB, get(X86::VPXORDZrr));
4182   }
4183   case X86::AVX512_256_SET0:
4184   case X86::AVX512_512_SET0: {
4185     bool HasVLX = Subtarget.hasVLX();
4186     unsigned SrcReg = MIB->getOperand(0).getReg();
4187     const TargetRegisterInfo *TRI = &getRegisterInfo();
4188     if (HasVLX || TRI->getEncodingValue(SrcReg) < 16) {
4189       unsigned XReg = TRI->getSubReg(SrcReg, X86::sub_xmm);
4190       MIB->getOperand(0).setReg(XReg);
4191       Expand2AddrUndef(MIB,
4192                        get(HasVLX ? X86::VPXORDZ128rr : X86::VXORPSrr));
4193       MIB.addReg(SrcReg, RegState::ImplicitDefine);
4194       return true;
4195     }
4196     return Expand2AddrUndef(MIB, get(X86::VPXORDZrr));
4197   }
4198   case X86::V_SETALLONES:
4199     return Expand2AddrUndef(MIB, get(HasAVX ? X86::VPCMPEQDrr : X86::PCMPEQDrr));
4200   case X86::AVX2_SETALLONES:
4201     return Expand2AddrUndef(MIB, get(X86::VPCMPEQDYrr));
4202   case X86::AVX1_SETALLONES: {
4203     unsigned Reg = MIB->getOperand(0).getReg();
4204     // VCMPPSYrri with an immediate 0xf should produce VCMPTRUEPS.
4205     MIB->setDesc(get(X86::VCMPPSYrri));
4206     MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef).addImm(0xf);
4207     return true;
4208   }
4209   case X86::AVX512_512_SETALLONES: {
4210     unsigned Reg = MIB->getOperand(0).getReg();
4211     MIB->setDesc(get(X86::VPTERNLOGDZrri));
4212     // VPTERNLOGD needs 3 register inputs and an immediate.
4213     // 0xff will return 1s for any input.
4214     MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef)
4215        .addReg(Reg, RegState::Undef).addImm(0xff);
4216     return true;
4217   }
4218   case X86::AVX512_512_SEXT_MASK_32:
4219   case X86::AVX512_512_SEXT_MASK_64: {
4220     unsigned Reg = MIB->getOperand(0).getReg();
4221     unsigned MaskReg = MIB->getOperand(1).getReg();
4222     unsigned MaskState = getRegState(MIB->getOperand(1));
4223     unsigned Opc = (MI.getOpcode() == X86::AVX512_512_SEXT_MASK_64) ?
4224                    X86::VPTERNLOGQZrrikz : X86::VPTERNLOGDZrrikz;
4225     MI.RemoveOperand(1);
4226     MIB->setDesc(get(Opc));
4227     // VPTERNLOG needs 3 register inputs and an immediate.
4228     // 0xff will return 1s for any input.
4229     MIB.addReg(Reg, RegState::Undef).addReg(MaskReg, MaskState)
4230        .addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef).addImm(0xff);
4231     return true;
4232   }
4233   case X86::VMOVAPSZ128rm_NOVLX:
4234     return expandNOVLXLoad(MIB, &getRegisterInfo(), get(X86::VMOVAPSrm),
4235                            get(X86::VBROADCASTF32X4rm), X86::sub_xmm);
4236   case X86::VMOVUPSZ128rm_NOVLX:
4237     return expandNOVLXLoad(MIB, &getRegisterInfo(), get(X86::VMOVUPSrm),
4238                            get(X86::VBROADCASTF32X4rm), X86::sub_xmm);
4239   case X86::VMOVAPSZ256rm_NOVLX:
4240     return expandNOVLXLoad(MIB, &getRegisterInfo(), get(X86::VMOVAPSYrm),
4241                            get(X86::VBROADCASTF64X4rm), X86::sub_ymm);
4242   case X86::VMOVUPSZ256rm_NOVLX:
4243     return expandNOVLXLoad(MIB, &getRegisterInfo(), get(X86::VMOVUPSYrm),
4244                            get(X86::VBROADCASTF64X4rm), X86::sub_ymm);
4245   case X86::VMOVAPSZ128mr_NOVLX:
4246     return expandNOVLXStore(MIB, &getRegisterInfo(), get(X86::VMOVAPSmr),
4247                             get(X86::VEXTRACTF32x4Zmr), X86::sub_xmm);
4248   case X86::VMOVUPSZ128mr_NOVLX:
4249     return expandNOVLXStore(MIB, &getRegisterInfo(), get(X86::VMOVUPSmr),
4250                             get(X86::VEXTRACTF32x4Zmr), X86::sub_xmm);
4251   case X86::VMOVAPSZ256mr_NOVLX:
4252     return expandNOVLXStore(MIB, &getRegisterInfo(), get(X86::VMOVAPSYmr),
4253                             get(X86::VEXTRACTF64x4Zmr), X86::sub_ymm);
4254   case X86::VMOVUPSZ256mr_NOVLX:
4255     return expandNOVLXStore(MIB, &getRegisterInfo(), get(X86::VMOVUPSYmr),
4256                             get(X86::VEXTRACTF64x4Zmr), X86::sub_ymm);
4257   case X86::MOV32ri64: {
4258     unsigned Reg = MIB->getOperand(0).getReg();
4259     unsigned Reg32 = RI.getSubReg(Reg, X86::sub_32bit);
4260     MI.setDesc(get(X86::MOV32ri));
4261     MIB->getOperand(0).setReg(Reg32);
4262     MIB.addReg(Reg, RegState::ImplicitDefine);
4263     return true;
4264   }
4265 
4266   // KNL does not recognize dependency-breaking idioms for mask registers,
4267   // so kxnor %k1, %k1, %k2 has a RAW dependence on %k1.
4268   // Using %k0 as the undef input register is a performance heuristic based
4269   // on the assumption that %k0 is used less frequently than the other mask
4270   // registers, since it is not usable as a write mask.
4271   // FIXME: A more advanced approach would be to choose the best input mask
4272   // register based on context.
4273   case X86::KSET0W: return Expand2AddrKreg(MIB, get(X86::KXORWrr), X86::K0);
4274   case X86::KSET0D: return Expand2AddrKreg(MIB, get(X86::KXORDrr), X86::K0);
4275   case X86::KSET0Q: return Expand2AddrKreg(MIB, get(X86::KXORQrr), X86::K0);
4276   case X86::KSET1W: return Expand2AddrKreg(MIB, get(X86::KXNORWrr), X86::K0);
4277   case X86::KSET1D: return Expand2AddrKreg(MIB, get(X86::KXNORDrr), X86::K0);
4278   case X86::KSET1Q: return Expand2AddrKreg(MIB, get(X86::KXNORQrr), X86::K0);
4279   case TargetOpcode::LOAD_STACK_GUARD:
4280     expandLoadStackGuard(MIB, *this);
4281     return true;
4282   case X86::XOR64_FP:
4283   case X86::XOR32_FP:
4284     return expandXorFP(MIB, *this);
4285   }
4286   return false;
4287 }
4288 
4289 /// Return true for all instructions that only update
4290 /// the first 32 or 64-bits of the destination register and leave the rest
4291 /// unmodified. This can be used to avoid folding loads if the instructions
4292 /// only update part of the destination register, and the non-updated part is
4293 /// not needed. e.g. cvtss2sd, sqrtss. Unfolding the load from these
4294 /// instructions breaks the partial register dependency and it can improve
4295 /// performance. e.g.:
4296 ///
4297 ///   movss (%rdi), %xmm0
4298 ///   cvtss2sd %xmm0, %xmm0
4299 ///
4300 /// Instead of
4301 ///   cvtss2sd (%rdi), %xmm0
4302 ///
4303 /// FIXME: This should be turned into a TSFlags.
4304 ///
hasPartialRegUpdate(unsigned Opcode,const X86Subtarget & Subtarget)4305 static bool hasPartialRegUpdate(unsigned Opcode,
4306                                 const X86Subtarget &Subtarget) {
4307   switch (Opcode) {
4308   case X86::CVTSI2SSrr:
4309   case X86::CVTSI2SSrm:
4310   case X86::CVTSI642SSrr:
4311   case X86::CVTSI642SSrm:
4312   case X86::CVTSI2SDrr:
4313   case X86::CVTSI2SDrm:
4314   case X86::CVTSI642SDrr:
4315   case X86::CVTSI642SDrm:
4316   case X86::CVTSD2SSrr:
4317   case X86::CVTSD2SSrm:
4318   case X86::CVTSS2SDrr:
4319   case X86::CVTSS2SDrm:
4320   case X86::MOVHPDrm:
4321   case X86::MOVHPSrm:
4322   case X86::MOVLPDrm:
4323   case X86::MOVLPSrm:
4324   case X86::RCPSSr:
4325   case X86::RCPSSm:
4326   case X86::RCPSSr_Int:
4327   case X86::RCPSSm_Int:
4328   case X86::ROUNDSDr:
4329   case X86::ROUNDSDm:
4330   case X86::ROUNDSSr:
4331   case X86::ROUNDSSm:
4332   case X86::RSQRTSSr:
4333   case X86::RSQRTSSm:
4334   case X86::RSQRTSSr_Int:
4335   case X86::RSQRTSSm_Int:
4336   case X86::SQRTSSr:
4337   case X86::SQRTSSm:
4338   case X86::SQRTSSr_Int:
4339   case X86::SQRTSSm_Int:
4340   case X86::SQRTSDr:
4341   case X86::SQRTSDm:
4342   case X86::SQRTSDr_Int:
4343   case X86::SQRTSDm_Int:
4344     return true;
4345   // GPR
4346   case X86::POPCNT32rm:
4347   case X86::POPCNT32rr:
4348   case X86::POPCNT64rm:
4349   case X86::POPCNT64rr:
4350     return Subtarget.hasPOPCNTFalseDeps();
4351   case X86::LZCNT32rm:
4352   case X86::LZCNT32rr:
4353   case X86::LZCNT64rm:
4354   case X86::LZCNT64rr:
4355   case X86::TZCNT32rm:
4356   case X86::TZCNT32rr:
4357   case X86::TZCNT64rm:
4358   case X86::TZCNT64rr:
4359     return Subtarget.hasLZCNTFalseDeps();
4360   }
4361 
4362   return false;
4363 }
4364 
4365 /// Inform the BreakFalseDeps pass how many idle
4366 /// instructions we would like before a partial register update.
getPartialRegUpdateClearance(const MachineInstr & MI,unsigned OpNum,const TargetRegisterInfo * TRI) const4367 unsigned X86InstrInfo::getPartialRegUpdateClearance(
4368     const MachineInstr &MI, unsigned OpNum,
4369     const TargetRegisterInfo *TRI) const {
4370   if (OpNum != 0 || !hasPartialRegUpdate(MI.getOpcode(), Subtarget))
4371     return 0;
4372 
4373   // If MI is marked as reading Reg, the partial register update is wanted.
4374   const MachineOperand &MO = MI.getOperand(0);
4375   unsigned Reg = MO.getReg();
4376   if (TargetRegisterInfo::isVirtualRegister(Reg)) {
4377     if (MO.readsReg() || MI.readsVirtualRegister(Reg))
4378       return 0;
4379   } else {
4380     if (MI.readsRegister(Reg, TRI))
4381       return 0;
4382   }
4383 
4384   // If any instructions in the clearance range are reading Reg, insert a
4385   // dependency breaking instruction, which is inexpensive and is likely to
4386   // be hidden in other instruction's cycles.
4387   return PartialRegUpdateClearance;
4388 }
4389 
4390 // Return true for any instruction the copies the high bits of the first source
4391 // operand into the unused high bits of the destination operand.
hasUndefRegUpdate(unsigned Opcode)4392 static bool hasUndefRegUpdate(unsigned Opcode) {
4393   switch (Opcode) {
4394   case X86::VCVTSI2SSrr:
4395   case X86::VCVTSI2SSrm:
4396   case X86::VCVTSI2SSrr_Int:
4397   case X86::VCVTSI2SSrm_Int:
4398   case X86::VCVTSI642SSrr:
4399   case X86::VCVTSI642SSrm:
4400   case X86::VCVTSI642SSrr_Int:
4401   case X86::VCVTSI642SSrm_Int:
4402   case X86::VCVTSI2SDrr:
4403   case X86::VCVTSI2SDrm:
4404   case X86::VCVTSI2SDrr_Int:
4405   case X86::VCVTSI2SDrm_Int:
4406   case X86::VCVTSI642SDrr:
4407   case X86::VCVTSI642SDrm:
4408   case X86::VCVTSI642SDrr_Int:
4409   case X86::VCVTSI642SDrm_Int:
4410   case X86::VCVTSD2SSrr:
4411   case X86::VCVTSD2SSrm:
4412   case X86::VCVTSD2SSrr_Int:
4413   case X86::VCVTSD2SSrm_Int:
4414   case X86::VCVTSS2SDrr:
4415   case X86::VCVTSS2SDrm:
4416   case X86::VCVTSS2SDrr_Int:
4417   case X86::VCVTSS2SDrm_Int:
4418   case X86::VRCPSSr:
4419   case X86::VRCPSSr_Int:
4420   case X86::VRCPSSm:
4421   case X86::VRCPSSm_Int:
4422   case X86::VROUNDSDr:
4423   case X86::VROUNDSDm:
4424   case X86::VROUNDSDr_Int:
4425   case X86::VROUNDSDm_Int:
4426   case X86::VROUNDSSr:
4427   case X86::VROUNDSSm:
4428   case X86::VROUNDSSr_Int:
4429   case X86::VROUNDSSm_Int:
4430   case X86::VRSQRTSSr:
4431   case X86::VRSQRTSSr_Int:
4432   case X86::VRSQRTSSm:
4433   case X86::VRSQRTSSm_Int:
4434   case X86::VSQRTSSr:
4435   case X86::VSQRTSSr_Int:
4436   case X86::VSQRTSSm:
4437   case X86::VSQRTSSm_Int:
4438   case X86::VSQRTSDr:
4439   case X86::VSQRTSDr_Int:
4440   case X86::VSQRTSDm:
4441   case X86::VSQRTSDm_Int:
4442   // AVX-512
4443   case X86::VCVTSI2SSZrr:
4444   case X86::VCVTSI2SSZrm:
4445   case X86::VCVTSI2SSZrr_Int:
4446   case X86::VCVTSI2SSZrrb_Int:
4447   case X86::VCVTSI2SSZrm_Int:
4448   case X86::VCVTSI642SSZrr:
4449   case X86::VCVTSI642SSZrm:
4450   case X86::VCVTSI642SSZrr_Int:
4451   case X86::VCVTSI642SSZrrb_Int:
4452   case X86::VCVTSI642SSZrm_Int:
4453   case X86::VCVTSI2SDZrr:
4454   case X86::VCVTSI2SDZrm:
4455   case X86::VCVTSI2SDZrr_Int:
4456   case X86::VCVTSI2SDZrrb_Int:
4457   case X86::VCVTSI2SDZrm_Int:
4458   case X86::VCVTSI642SDZrr:
4459   case X86::VCVTSI642SDZrm:
4460   case X86::VCVTSI642SDZrr_Int:
4461   case X86::VCVTSI642SDZrrb_Int:
4462   case X86::VCVTSI642SDZrm_Int:
4463   case X86::VCVTUSI2SSZrr:
4464   case X86::VCVTUSI2SSZrm:
4465   case X86::VCVTUSI2SSZrr_Int:
4466   case X86::VCVTUSI2SSZrrb_Int:
4467   case X86::VCVTUSI2SSZrm_Int:
4468   case X86::VCVTUSI642SSZrr:
4469   case X86::VCVTUSI642SSZrm:
4470   case X86::VCVTUSI642SSZrr_Int:
4471   case X86::VCVTUSI642SSZrrb_Int:
4472   case X86::VCVTUSI642SSZrm_Int:
4473   case X86::VCVTUSI2SDZrr:
4474   case X86::VCVTUSI2SDZrm:
4475   case X86::VCVTUSI2SDZrr_Int:
4476   case X86::VCVTUSI2SDZrm_Int:
4477   case X86::VCVTUSI642SDZrr:
4478   case X86::VCVTUSI642SDZrm:
4479   case X86::VCVTUSI642SDZrr_Int:
4480   case X86::VCVTUSI642SDZrrb_Int:
4481   case X86::VCVTUSI642SDZrm_Int:
4482   case X86::VCVTSD2SSZrr:
4483   case X86::VCVTSD2SSZrr_Int:
4484   case X86::VCVTSD2SSZrrb_Int:
4485   case X86::VCVTSD2SSZrm:
4486   case X86::VCVTSD2SSZrm_Int:
4487   case X86::VCVTSS2SDZrr:
4488   case X86::VCVTSS2SDZrr_Int:
4489   case X86::VCVTSS2SDZrrb_Int:
4490   case X86::VCVTSS2SDZrm:
4491   case X86::VCVTSS2SDZrm_Int:
4492   case X86::VGETEXPSDZr:
4493   case X86::VGETEXPSDZrb:
4494   case X86::VGETEXPSDZm:
4495   case X86::VGETEXPSSZr:
4496   case X86::VGETEXPSSZrb:
4497   case X86::VGETEXPSSZm:
4498   case X86::VGETMANTSDZrri:
4499   case X86::VGETMANTSDZrrib:
4500   case X86::VGETMANTSDZrmi:
4501   case X86::VGETMANTSSZrri:
4502   case X86::VGETMANTSSZrrib:
4503   case X86::VGETMANTSSZrmi:
4504   case X86::VRNDSCALESDZr:
4505   case X86::VRNDSCALESDZr_Int:
4506   case X86::VRNDSCALESDZrb_Int:
4507   case X86::VRNDSCALESDZm:
4508   case X86::VRNDSCALESDZm_Int:
4509   case X86::VRNDSCALESSZr:
4510   case X86::VRNDSCALESSZr_Int:
4511   case X86::VRNDSCALESSZrb_Int:
4512   case X86::VRNDSCALESSZm:
4513   case X86::VRNDSCALESSZm_Int:
4514   case X86::VRCP14SDZrr:
4515   case X86::VRCP14SDZrm:
4516   case X86::VRCP14SSZrr:
4517   case X86::VRCP14SSZrm:
4518   case X86::VRCP28SDZr:
4519   case X86::VRCP28SDZrb:
4520   case X86::VRCP28SDZm:
4521   case X86::VRCP28SSZr:
4522   case X86::VRCP28SSZrb:
4523   case X86::VRCP28SSZm:
4524   case X86::VREDUCESSZrmi:
4525   case X86::VREDUCESSZrri:
4526   case X86::VREDUCESSZrrib:
4527   case X86::VRSQRT14SDZrr:
4528   case X86::VRSQRT14SDZrm:
4529   case X86::VRSQRT14SSZrr:
4530   case X86::VRSQRT14SSZrm:
4531   case X86::VRSQRT28SDZr:
4532   case X86::VRSQRT28SDZrb:
4533   case X86::VRSQRT28SDZm:
4534   case X86::VRSQRT28SSZr:
4535   case X86::VRSQRT28SSZrb:
4536   case X86::VRSQRT28SSZm:
4537   case X86::VSQRTSSZr:
4538   case X86::VSQRTSSZr_Int:
4539   case X86::VSQRTSSZrb_Int:
4540   case X86::VSQRTSSZm:
4541   case X86::VSQRTSSZm_Int:
4542   case X86::VSQRTSDZr:
4543   case X86::VSQRTSDZr_Int:
4544   case X86::VSQRTSDZrb_Int:
4545   case X86::VSQRTSDZm:
4546   case X86::VSQRTSDZm_Int:
4547     return true;
4548   }
4549 
4550   return false;
4551 }
4552 
4553 /// Inform the BreakFalseDeps pass how many idle instructions we would like
4554 /// before certain undef register reads.
4555 ///
4556 /// This catches the VCVTSI2SD family of instructions:
4557 ///
4558 /// vcvtsi2sdq %rax, undef %xmm0, %xmm14
4559 ///
4560 /// We should to be careful *not* to catch VXOR idioms which are presumably
4561 /// handled specially in the pipeline:
4562 ///
4563 /// vxorps undef %xmm1, undef %xmm1, %xmm1
4564 ///
4565 /// Like getPartialRegUpdateClearance, this makes a strong assumption that the
4566 /// high bits that are passed-through are not live.
4567 unsigned
getUndefRegClearance(const MachineInstr & MI,unsigned & OpNum,const TargetRegisterInfo * TRI) const4568 X86InstrInfo::getUndefRegClearance(const MachineInstr &MI, unsigned &OpNum,
4569                                    const TargetRegisterInfo *TRI) const {
4570   if (!hasUndefRegUpdate(MI.getOpcode()))
4571     return 0;
4572 
4573   // Set the OpNum parameter to the first source operand.
4574   OpNum = 1;
4575 
4576   const MachineOperand &MO = MI.getOperand(OpNum);
4577   if (MO.isUndef() && TargetRegisterInfo::isPhysicalRegister(MO.getReg())) {
4578     return UndefRegClearance;
4579   }
4580   return 0;
4581 }
4582 
breakPartialRegDependency(MachineInstr & MI,unsigned OpNum,const TargetRegisterInfo * TRI) const4583 void X86InstrInfo::breakPartialRegDependency(
4584     MachineInstr &MI, unsigned OpNum, const TargetRegisterInfo *TRI) const {
4585   unsigned Reg = MI.getOperand(OpNum).getReg();
4586   // If MI kills this register, the false dependence is already broken.
4587   if (MI.killsRegister(Reg, TRI))
4588     return;
4589 
4590   if (X86::VR128RegClass.contains(Reg)) {
4591     // These instructions are all floating point domain, so xorps is the best
4592     // choice.
4593     unsigned Opc = Subtarget.hasAVX() ? X86::VXORPSrr : X86::XORPSrr;
4594     BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), get(Opc), Reg)
4595         .addReg(Reg, RegState::Undef)
4596         .addReg(Reg, RegState::Undef);
4597     MI.addRegisterKilled(Reg, TRI, true);
4598   } else if (X86::VR256RegClass.contains(Reg)) {
4599     // Use vxorps to clear the full ymm register.
4600     // It wants to read and write the xmm sub-register.
4601     unsigned XReg = TRI->getSubReg(Reg, X86::sub_xmm);
4602     BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), get(X86::VXORPSrr), XReg)
4603         .addReg(XReg, RegState::Undef)
4604         .addReg(XReg, RegState::Undef)
4605         .addReg(Reg, RegState::ImplicitDefine);
4606     MI.addRegisterKilled(Reg, TRI, true);
4607   } else if (X86::GR64RegClass.contains(Reg)) {
4608     // Using XOR32rr because it has shorter encoding and zeros up the upper bits
4609     // as well.
4610     unsigned XReg = TRI->getSubReg(Reg, X86::sub_32bit);
4611     BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), get(X86::XOR32rr), XReg)
4612         .addReg(XReg, RegState::Undef)
4613         .addReg(XReg, RegState::Undef)
4614         .addReg(Reg, RegState::ImplicitDefine);
4615     MI.addRegisterKilled(Reg, TRI, true);
4616   } else if (X86::GR32RegClass.contains(Reg)) {
4617     BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), get(X86::XOR32rr), Reg)
4618         .addReg(Reg, RegState::Undef)
4619         .addReg(Reg, RegState::Undef);
4620     MI.addRegisterKilled(Reg, TRI, true);
4621   }
4622 }
4623 
addOperands(MachineInstrBuilder & MIB,ArrayRef<MachineOperand> MOs,int PtrOffset=0)4624 static void addOperands(MachineInstrBuilder &MIB, ArrayRef<MachineOperand> MOs,
4625                         int PtrOffset = 0) {
4626   unsigned NumAddrOps = MOs.size();
4627 
4628   if (NumAddrOps < 4) {
4629     // FrameIndex only - add an immediate offset (whether its zero or not).
4630     for (unsigned i = 0; i != NumAddrOps; ++i)
4631       MIB.add(MOs[i]);
4632     addOffset(MIB, PtrOffset);
4633   } else {
4634     // General Memory Addressing - we need to add any offset to an existing
4635     // offset.
4636     assert(MOs.size() == 5 && "Unexpected memory operand list length");
4637     for (unsigned i = 0; i != NumAddrOps; ++i) {
4638       const MachineOperand &MO = MOs[i];
4639       if (i == 3 && PtrOffset != 0) {
4640         MIB.addDisp(MO, PtrOffset);
4641       } else {
4642         MIB.add(MO);
4643       }
4644     }
4645   }
4646 }
4647 
updateOperandRegConstraints(MachineFunction & MF,MachineInstr & NewMI,const TargetInstrInfo & TII)4648 static void updateOperandRegConstraints(MachineFunction &MF,
4649                                         MachineInstr &NewMI,
4650                                         const TargetInstrInfo &TII) {
4651   MachineRegisterInfo &MRI = MF.getRegInfo();
4652   const TargetRegisterInfo &TRI = *MRI.getTargetRegisterInfo();
4653 
4654   for (int Idx : llvm::seq<int>(0, NewMI.getNumOperands())) {
4655     MachineOperand &MO = NewMI.getOperand(Idx);
4656     // We only need to update constraints on virtual register operands.
4657     if (!MO.isReg())
4658       continue;
4659     unsigned Reg = MO.getReg();
4660     if (!TRI.isVirtualRegister(Reg))
4661       continue;
4662 
4663     auto *NewRC = MRI.constrainRegClass(
4664         Reg, TII.getRegClass(NewMI.getDesc(), Idx, &TRI, MF));
4665     if (!NewRC) {
4666       LLVM_DEBUG(
4667           dbgs() << "WARNING: Unable to update register constraint for operand "
4668                  << Idx << " of instruction:\n";
4669           NewMI.dump(); dbgs() << "\n");
4670     }
4671   }
4672 }
4673 
FuseTwoAddrInst(MachineFunction & MF,unsigned Opcode,ArrayRef<MachineOperand> MOs,MachineBasicBlock::iterator InsertPt,MachineInstr & MI,const TargetInstrInfo & TII)4674 static MachineInstr *FuseTwoAddrInst(MachineFunction &MF, unsigned Opcode,
4675                                      ArrayRef<MachineOperand> MOs,
4676                                      MachineBasicBlock::iterator InsertPt,
4677                                      MachineInstr &MI,
4678                                      const TargetInstrInfo &TII) {
4679   // Create the base instruction with the memory operand as the first part.
4680   // Omit the implicit operands, something BuildMI can't do.
4681   MachineInstr *NewMI =
4682       MF.CreateMachineInstr(TII.get(Opcode), MI.getDebugLoc(), true);
4683   MachineInstrBuilder MIB(MF, NewMI);
4684   addOperands(MIB, MOs);
4685 
4686   // Loop over the rest of the ri operands, converting them over.
4687   unsigned NumOps = MI.getDesc().getNumOperands() - 2;
4688   for (unsigned i = 0; i != NumOps; ++i) {
4689     MachineOperand &MO = MI.getOperand(i + 2);
4690     MIB.add(MO);
4691   }
4692   for (unsigned i = NumOps + 2, e = MI.getNumOperands(); i != e; ++i) {
4693     MachineOperand &MO = MI.getOperand(i);
4694     MIB.add(MO);
4695   }
4696 
4697   updateOperandRegConstraints(MF, *NewMI, TII);
4698 
4699   MachineBasicBlock *MBB = InsertPt->getParent();
4700   MBB->insert(InsertPt, NewMI);
4701 
4702   return MIB;
4703 }
4704 
FuseInst(MachineFunction & MF,unsigned Opcode,unsigned OpNo,ArrayRef<MachineOperand> MOs,MachineBasicBlock::iterator InsertPt,MachineInstr & MI,const TargetInstrInfo & TII,int PtrOffset=0)4705 static MachineInstr *FuseInst(MachineFunction &MF, unsigned Opcode,
4706                               unsigned OpNo, ArrayRef<MachineOperand> MOs,
4707                               MachineBasicBlock::iterator InsertPt,
4708                               MachineInstr &MI, const TargetInstrInfo &TII,
4709                               int PtrOffset = 0) {
4710   // Omit the implicit operands, something BuildMI can't do.
4711   MachineInstr *NewMI =
4712       MF.CreateMachineInstr(TII.get(Opcode), MI.getDebugLoc(), true);
4713   MachineInstrBuilder MIB(MF, NewMI);
4714 
4715   for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
4716     MachineOperand &MO = MI.getOperand(i);
4717     if (i == OpNo) {
4718       assert(MO.isReg() && "Expected to fold into reg operand!");
4719       addOperands(MIB, MOs, PtrOffset);
4720     } else {
4721       MIB.add(MO);
4722     }
4723   }
4724 
4725   updateOperandRegConstraints(MF, *NewMI, TII);
4726 
4727   MachineBasicBlock *MBB = InsertPt->getParent();
4728   MBB->insert(InsertPt, NewMI);
4729 
4730   return MIB;
4731 }
4732 
MakeM0Inst(const TargetInstrInfo & TII,unsigned Opcode,ArrayRef<MachineOperand> MOs,MachineBasicBlock::iterator InsertPt,MachineInstr & MI)4733 static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode,
4734                                 ArrayRef<MachineOperand> MOs,
4735                                 MachineBasicBlock::iterator InsertPt,
4736                                 MachineInstr &MI) {
4737   MachineInstrBuilder MIB = BuildMI(*InsertPt->getParent(), InsertPt,
4738                                     MI.getDebugLoc(), TII.get(Opcode));
4739   addOperands(MIB, MOs);
4740   return MIB.addImm(0);
4741 }
4742 
foldMemoryOperandCustom(MachineFunction & MF,MachineInstr & MI,unsigned OpNum,ArrayRef<MachineOperand> MOs,MachineBasicBlock::iterator InsertPt,unsigned Size,unsigned Align) const4743 MachineInstr *X86InstrInfo::foldMemoryOperandCustom(
4744     MachineFunction &MF, MachineInstr &MI, unsigned OpNum,
4745     ArrayRef<MachineOperand> MOs, MachineBasicBlock::iterator InsertPt,
4746     unsigned Size, unsigned Align) const {
4747   switch (MI.getOpcode()) {
4748   case X86::INSERTPSrr:
4749   case X86::VINSERTPSrr:
4750   case X86::VINSERTPSZrr:
4751     // Attempt to convert the load of inserted vector into a fold load
4752     // of a single float.
4753     if (OpNum == 2) {
4754       unsigned Imm = MI.getOperand(MI.getNumOperands() - 1).getImm();
4755       unsigned ZMask = Imm & 15;
4756       unsigned DstIdx = (Imm >> 4) & 3;
4757       unsigned SrcIdx = (Imm >> 6) & 3;
4758 
4759       const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
4760       const TargetRegisterClass *RC = getRegClass(MI.getDesc(), OpNum, &RI, MF);
4761       unsigned RCSize = TRI.getRegSizeInBits(*RC) / 8;
4762       if (Size <= RCSize && 4 <= Align) {
4763         int PtrOffset = SrcIdx * 4;
4764         unsigned NewImm = (DstIdx << 4) | ZMask;
4765         unsigned NewOpCode =
4766             (MI.getOpcode() == X86::VINSERTPSZrr) ? X86::VINSERTPSZrm :
4767             (MI.getOpcode() == X86::VINSERTPSrr)  ? X86::VINSERTPSrm  :
4768                                                     X86::INSERTPSrm;
4769         MachineInstr *NewMI =
4770             FuseInst(MF, NewOpCode, OpNum, MOs, InsertPt, MI, *this, PtrOffset);
4771         NewMI->getOperand(NewMI->getNumOperands() - 1).setImm(NewImm);
4772         return NewMI;
4773       }
4774     }
4775     break;
4776   case X86::MOVHLPSrr:
4777   case X86::VMOVHLPSrr:
4778   case X86::VMOVHLPSZrr:
4779     // Move the upper 64-bits of the second operand to the lower 64-bits.
4780     // To fold the load, adjust the pointer to the upper and use (V)MOVLPS.
4781     // TODO: In most cases AVX doesn't have a 8-byte alignment requirement.
4782     if (OpNum == 2) {
4783       const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
4784       const TargetRegisterClass *RC = getRegClass(MI.getDesc(), OpNum, &RI, MF);
4785       unsigned RCSize = TRI.getRegSizeInBits(*RC) / 8;
4786       if (Size <= RCSize && 8 <= Align) {
4787         unsigned NewOpCode =
4788             (MI.getOpcode() == X86::VMOVHLPSZrr) ? X86::VMOVLPSZ128rm :
4789             (MI.getOpcode() == X86::VMOVHLPSrr)  ? X86::VMOVLPSrm     :
4790                                                    X86::MOVLPSrm;
4791         MachineInstr *NewMI =
4792             FuseInst(MF, NewOpCode, OpNum, MOs, InsertPt, MI, *this, 8);
4793         return NewMI;
4794       }
4795     }
4796     break;
4797   };
4798 
4799   return nullptr;
4800 }
4801 
shouldPreventUndefRegUpdateMemFold(MachineFunction & MF,MachineInstr & MI)4802 static bool shouldPreventUndefRegUpdateMemFold(MachineFunction &MF, MachineInstr &MI) {
4803   if (MF.getFunction().optForSize() || !hasUndefRegUpdate(MI.getOpcode()) ||
4804       !MI.getOperand(1).isReg())
4805     return false;
4806 
4807   // The are two cases we need to handle depending on where in the pipeline
4808   // the folding attempt is being made.
4809   // -Register has the undef flag set.
4810   // -Register is produced by the IMPLICIT_DEF instruction.
4811 
4812   if (MI.getOperand(1).isUndef())
4813     return true;
4814 
4815   MachineRegisterInfo &RegInfo = MF.getRegInfo();
4816   MachineInstr *VRegDef = RegInfo.getUniqueVRegDef(MI.getOperand(1).getReg());
4817   return VRegDef && VRegDef->isImplicitDef();
4818 }
4819 
4820 
foldMemoryOperandImpl(MachineFunction & MF,MachineInstr & MI,unsigned OpNum,ArrayRef<MachineOperand> MOs,MachineBasicBlock::iterator InsertPt,unsigned Size,unsigned Align,bool AllowCommute) const4821 MachineInstr *X86InstrInfo::foldMemoryOperandImpl(
4822     MachineFunction &MF, MachineInstr &MI, unsigned OpNum,
4823     ArrayRef<MachineOperand> MOs, MachineBasicBlock::iterator InsertPt,
4824     unsigned Size, unsigned Align, bool AllowCommute) const {
4825   bool isSlowTwoMemOps = Subtarget.slowTwoMemOps();
4826   bool isTwoAddrFold = false;
4827 
4828   // For CPUs that favor the register form of a call or push,
4829   // do not fold loads into calls or pushes, unless optimizing for size
4830   // aggressively.
4831   if (isSlowTwoMemOps && !MF.getFunction().optForMinSize() &&
4832       (MI.getOpcode() == X86::CALL32r || MI.getOpcode() == X86::CALL64r ||
4833        MI.getOpcode() == X86::PUSH16r || MI.getOpcode() == X86::PUSH32r ||
4834        MI.getOpcode() == X86::PUSH64r))
4835     return nullptr;
4836 
4837   // Avoid partial and undef register update stalls unless optimizing for size.
4838   if (!MF.getFunction().optForSize() &&
4839       (hasPartialRegUpdate(MI.getOpcode(), Subtarget) ||
4840        shouldPreventUndefRegUpdateMemFold(MF, MI)))
4841     return nullptr;
4842 
4843   unsigned NumOps = MI.getDesc().getNumOperands();
4844   bool isTwoAddr =
4845       NumOps > 1 && MI.getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
4846 
4847   // FIXME: AsmPrinter doesn't know how to handle
4848   // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
4849   if (MI.getOpcode() == X86::ADD32ri &&
4850       MI.getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
4851     return nullptr;
4852 
4853   // GOTTPOFF relocation loads can only be folded into add instructions.
4854   // FIXME: Need to exclude other relocations that only support specific
4855   // instructions.
4856   if (MOs.size() == X86::AddrNumOperands &&
4857       MOs[X86::AddrDisp].getTargetFlags() == X86II::MO_GOTTPOFF &&
4858       MI.getOpcode() != X86::ADD64rr)
4859     return nullptr;
4860 
4861   MachineInstr *NewMI = nullptr;
4862 
4863   // Attempt to fold any custom cases we have.
4864   if (MachineInstr *CustomMI =
4865           foldMemoryOperandCustom(MF, MI, OpNum, MOs, InsertPt, Size, Align))
4866     return CustomMI;
4867 
4868   const X86MemoryFoldTableEntry *I = nullptr;
4869 
4870   // Folding a memory location into the two-address part of a two-address
4871   // instruction is different than folding it other places.  It requires
4872   // replacing the *two* registers with the memory location.
4873   if (isTwoAddr && NumOps >= 2 && OpNum < 2 && MI.getOperand(0).isReg() &&
4874       MI.getOperand(1).isReg() &&
4875       MI.getOperand(0).getReg() == MI.getOperand(1).getReg()) {
4876     I = lookupTwoAddrFoldTable(MI.getOpcode());
4877     isTwoAddrFold = true;
4878   } else {
4879     if (OpNum == 0) {
4880       if (MI.getOpcode() == X86::MOV32r0) {
4881         NewMI = MakeM0Inst(*this, X86::MOV32mi, MOs, InsertPt, MI);
4882         if (NewMI)
4883           return NewMI;
4884       }
4885     }
4886 
4887     I = lookupFoldTable(MI.getOpcode(), OpNum);
4888   }
4889 
4890   if (I != nullptr) {
4891     unsigned Opcode = I->DstOp;
4892     unsigned MinAlign = (I->Flags & TB_ALIGN_MASK) >> TB_ALIGN_SHIFT;
4893     if (Align < MinAlign)
4894       return nullptr;
4895     bool NarrowToMOV32rm = false;
4896     if (Size) {
4897       const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
4898       const TargetRegisterClass *RC = getRegClass(MI.getDesc(), OpNum,
4899                                                   &RI, MF);
4900       unsigned RCSize = TRI.getRegSizeInBits(*RC) / 8;
4901       if (Size < RCSize) {
4902         // Check if it's safe to fold the load. If the size of the object is
4903         // narrower than the load width, then it's not.
4904         if (Opcode != X86::MOV64rm || RCSize != 8 || Size != 4)
4905           return nullptr;
4906         // If this is a 64-bit load, but the spill slot is 32, then we can do
4907         // a 32-bit load which is implicitly zero-extended. This likely is
4908         // due to live interval analysis remat'ing a load from stack slot.
4909         if (MI.getOperand(0).getSubReg() || MI.getOperand(1).getSubReg())
4910           return nullptr;
4911         Opcode = X86::MOV32rm;
4912         NarrowToMOV32rm = true;
4913       }
4914     }
4915 
4916     if (isTwoAddrFold)
4917       NewMI = FuseTwoAddrInst(MF, Opcode, MOs, InsertPt, MI, *this);
4918     else
4919       NewMI = FuseInst(MF, Opcode, OpNum, MOs, InsertPt, MI, *this);
4920 
4921     if (NarrowToMOV32rm) {
4922       // If this is the special case where we use a MOV32rm to load a 32-bit
4923       // value and zero-extend the top bits. Change the destination register
4924       // to a 32-bit one.
4925       unsigned DstReg = NewMI->getOperand(0).getReg();
4926       if (TargetRegisterInfo::isPhysicalRegister(DstReg))
4927         NewMI->getOperand(0).setReg(RI.getSubReg(DstReg, X86::sub_32bit));
4928       else
4929         NewMI->getOperand(0).setSubReg(X86::sub_32bit);
4930     }
4931     return NewMI;
4932   }
4933 
4934   // If the instruction and target operand are commutable, commute the
4935   // instruction and try again.
4936   if (AllowCommute) {
4937     unsigned CommuteOpIdx1 = OpNum, CommuteOpIdx2 = CommuteAnyOperandIndex;
4938     if (findCommutedOpIndices(MI, CommuteOpIdx1, CommuteOpIdx2)) {
4939       bool HasDef = MI.getDesc().getNumDefs();
4940       unsigned Reg0 = HasDef ? MI.getOperand(0).getReg() : 0;
4941       unsigned Reg1 = MI.getOperand(CommuteOpIdx1).getReg();
4942       unsigned Reg2 = MI.getOperand(CommuteOpIdx2).getReg();
4943       bool Tied1 =
4944           0 == MI.getDesc().getOperandConstraint(CommuteOpIdx1, MCOI::TIED_TO);
4945       bool Tied2 =
4946           0 == MI.getDesc().getOperandConstraint(CommuteOpIdx2, MCOI::TIED_TO);
4947 
4948       // If either of the commutable operands are tied to the destination
4949       // then we can not commute + fold.
4950       if ((HasDef && Reg0 == Reg1 && Tied1) ||
4951           (HasDef && Reg0 == Reg2 && Tied2))
4952         return nullptr;
4953 
4954       MachineInstr *CommutedMI =
4955           commuteInstruction(MI, false, CommuteOpIdx1, CommuteOpIdx2);
4956       if (!CommutedMI) {
4957         // Unable to commute.
4958         return nullptr;
4959       }
4960       if (CommutedMI != &MI) {
4961         // New instruction. We can't fold from this.
4962         CommutedMI->eraseFromParent();
4963         return nullptr;
4964       }
4965 
4966       // Attempt to fold with the commuted version of the instruction.
4967       NewMI = foldMemoryOperandImpl(MF, MI, CommuteOpIdx2, MOs, InsertPt,
4968                                     Size, Align, /*AllowCommute=*/false);
4969       if (NewMI)
4970         return NewMI;
4971 
4972       // Folding failed again - undo the commute before returning.
4973       MachineInstr *UncommutedMI =
4974           commuteInstruction(MI, false, CommuteOpIdx1, CommuteOpIdx2);
4975       if (!UncommutedMI) {
4976         // Unable to commute.
4977         return nullptr;
4978       }
4979       if (UncommutedMI != &MI) {
4980         // New instruction. It doesn't need to be kept.
4981         UncommutedMI->eraseFromParent();
4982         return nullptr;
4983       }
4984 
4985       // Return here to prevent duplicate fuse failure report.
4986       return nullptr;
4987     }
4988   }
4989 
4990   // No fusion
4991   if (PrintFailedFusing && !MI.isCopy())
4992     dbgs() << "We failed to fuse operand " << OpNum << " in " << MI;
4993   return nullptr;
4994 }
4995 
4996 MachineInstr *
foldMemoryOperandImpl(MachineFunction & MF,MachineInstr & MI,ArrayRef<unsigned> Ops,MachineBasicBlock::iterator InsertPt,int FrameIndex,LiveIntervals * LIS) const4997 X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF, MachineInstr &MI,
4998                                     ArrayRef<unsigned> Ops,
4999                                     MachineBasicBlock::iterator InsertPt,
5000                                     int FrameIndex, LiveIntervals *LIS) const {
5001   // Check switch flag
5002   if (NoFusing)
5003     return nullptr;
5004 
5005   // Avoid partial and undef register update stalls unless optimizing for size.
5006   if (!MF.getFunction().optForSize() &&
5007       (hasPartialRegUpdate(MI.getOpcode(), Subtarget) ||
5008        shouldPreventUndefRegUpdateMemFold(MF, MI)))
5009     return nullptr;
5010 
5011   // Don't fold subreg spills, or reloads that use a high subreg.
5012   for (auto Op : Ops) {
5013     MachineOperand &MO = MI.getOperand(Op);
5014     auto SubReg = MO.getSubReg();
5015     if (SubReg && (MO.isDef() || SubReg == X86::sub_8bit_hi))
5016       return nullptr;
5017   }
5018 
5019   const MachineFrameInfo &MFI = MF.getFrameInfo();
5020   unsigned Size = MFI.getObjectSize(FrameIndex);
5021   unsigned Alignment = MFI.getObjectAlignment(FrameIndex);
5022   // If the function stack isn't realigned we don't want to fold instructions
5023   // that need increased alignment.
5024   if (!RI.needsStackRealignment(MF))
5025     Alignment =
5026         std::min(Alignment, Subtarget.getFrameLowering()->getStackAlignment());
5027   if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
5028     unsigned NewOpc = 0;
5029     unsigned RCSize = 0;
5030     switch (MI.getOpcode()) {
5031     default: return nullptr;
5032     case X86::TEST8rr:  NewOpc = X86::CMP8ri; RCSize = 1; break;
5033     case X86::TEST16rr: NewOpc = X86::CMP16ri8; RCSize = 2; break;
5034     case X86::TEST32rr: NewOpc = X86::CMP32ri8; RCSize = 4; break;
5035     case X86::TEST64rr: NewOpc = X86::CMP64ri8; RCSize = 8; break;
5036     }
5037     // Check if it's safe to fold the load. If the size of the object is
5038     // narrower than the load width, then it's not.
5039     if (Size < RCSize)
5040       return nullptr;
5041     // Change to CMPXXri r, 0 first.
5042     MI.setDesc(get(NewOpc));
5043     MI.getOperand(1).ChangeToImmediate(0);
5044   } else if (Ops.size() != 1)
5045     return nullptr;
5046 
5047   return foldMemoryOperandImpl(MF, MI, Ops[0],
5048                                MachineOperand::CreateFI(FrameIndex), InsertPt,
5049                                Size, Alignment, /*AllowCommute=*/true);
5050 }
5051 
5052 /// Check if \p LoadMI is a partial register load that we can't fold into \p MI
5053 /// because the latter uses contents that wouldn't be defined in the folded
5054 /// version.  For instance, this transformation isn't legal:
5055 ///   movss (%rdi), %xmm0
5056 ///   addps %xmm0, %xmm0
5057 /// ->
5058 ///   addps (%rdi), %xmm0
5059 ///
5060 /// But this one is:
5061 ///   movss (%rdi), %xmm0
5062 ///   addss %xmm0, %xmm0
5063 /// ->
5064 ///   addss (%rdi), %xmm0
5065 ///
isNonFoldablePartialRegisterLoad(const MachineInstr & LoadMI,const MachineInstr & UserMI,const MachineFunction & MF)5066 static bool isNonFoldablePartialRegisterLoad(const MachineInstr &LoadMI,
5067                                              const MachineInstr &UserMI,
5068                                              const MachineFunction &MF) {
5069   unsigned Opc = LoadMI.getOpcode();
5070   unsigned UserOpc = UserMI.getOpcode();
5071   const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
5072   const TargetRegisterClass *RC =
5073       MF.getRegInfo().getRegClass(LoadMI.getOperand(0).getReg());
5074   unsigned RegSize = TRI.getRegSizeInBits(*RC);
5075 
5076   if ((Opc == X86::MOVSSrm || Opc == X86::VMOVSSrm || Opc == X86::VMOVSSZrm) &&
5077       RegSize > 32) {
5078     // These instructions only load 32 bits, we can't fold them if the
5079     // destination register is wider than 32 bits (4 bytes), and its user
5080     // instruction isn't scalar (SS).
5081     switch (UserOpc) {
5082     case X86::ADDSSrr_Int: case X86::VADDSSrr_Int: case X86::VADDSSZrr_Int:
5083     case X86::CMPSSrr_Int: case X86::VCMPSSrr_Int: case X86::VCMPSSZrr_Int:
5084     case X86::DIVSSrr_Int: case X86::VDIVSSrr_Int: case X86::VDIVSSZrr_Int:
5085     case X86::MAXSSrr_Int: case X86::VMAXSSrr_Int: case X86::VMAXSSZrr_Int:
5086     case X86::MINSSrr_Int: case X86::VMINSSrr_Int: case X86::VMINSSZrr_Int:
5087     case X86::MULSSrr_Int: case X86::VMULSSrr_Int: case X86::VMULSSZrr_Int:
5088     case X86::SUBSSrr_Int: case X86::VSUBSSrr_Int: case X86::VSUBSSZrr_Int:
5089     case X86::VADDSSZrr_Intk: case X86::VADDSSZrr_Intkz:
5090     case X86::VDIVSSZrr_Intk: case X86::VDIVSSZrr_Intkz:
5091     case X86::VMAXSSZrr_Intk: case X86::VMAXSSZrr_Intkz:
5092     case X86::VMINSSZrr_Intk: case X86::VMINSSZrr_Intkz:
5093     case X86::VMULSSZrr_Intk: case X86::VMULSSZrr_Intkz:
5094     case X86::VSUBSSZrr_Intk: case X86::VSUBSSZrr_Intkz:
5095     case X86::VFMADDSS4rr_Int:   case X86::VFNMADDSS4rr_Int:
5096     case X86::VFMSUBSS4rr_Int:   case X86::VFNMSUBSS4rr_Int:
5097     case X86::VFMADD132SSr_Int:  case X86::VFNMADD132SSr_Int:
5098     case X86::VFMADD213SSr_Int:  case X86::VFNMADD213SSr_Int:
5099     case X86::VFMADD231SSr_Int:  case X86::VFNMADD231SSr_Int:
5100     case X86::VFMSUB132SSr_Int:  case X86::VFNMSUB132SSr_Int:
5101     case X86::VFMSUB213SSr_Int:  case X86::VFNMSUB213SSr_Int:
5102     case X86::VFMSUB231SSr_Int:  case X86::VFNMSUB231SSr_Int:
5103     case X86::VFMADD132SSZr_Int: case X86::VFNMADD132SSZr_Int:
5104     case X86::VFMADD213SSZr_Int: case X86::VFNMADD213SSZr_Int:
5105     case X86::VFMADD231SSZr_Int: case X86::VFNMADD231SSZr_Int:
5106     case X86::VFMSUB132SSZr_Int: case X86::VFNMSUB132SSZr_Int:
5107     case X86::VFMSUB213SSZr_Int: case X86::VFNMSUB213SSZr_Int:
5108     case X86::VFMSUB231SSZr_Int: case X86::VFNMSUB231SSZr_Int:
5109     case X86::VFMADD132SSZr_Intk: case X86::VFNMADD132SSZr_Intk:
5110     case X86::VFMADD213SSZr_Intk: case X86::VFNMADD213SSZr_Intk:
5111     case X86::VFMADD231SSZr_Intk: case X86::VFNMADD231SSZr_Intk:
5112     case X86::VFMSUB132SSZr_Intk: case X86::VFNMSUB132SSZr_Intk:
5113     case X86::VFMSUB213SSZr_Intk: case X86::VFNMSUB213SSZr_Intk:
5114     case X86::VFMSUB231SSZr_Intk: case X86::VFNMSUB231SSZr_Intk:
5115     case X86::VFMADD132SSZr_Intkz: case X86::VFNMADD132SSZr_Intkz:
5116     case X86::VFMADD213SSZr_Intkz: case X86::VFNMADD213SSZr_Intkz:
5117     case X86::VFMADD231SSZr_Intkz: case X86::VFNMADD231SSZr_Intkz:
5118     case X86::VFMSUB132SSZr_Intkz: case X86::VFNMSUB132SSZr_Intkz:
5119     case X86::VFMSUB213SSZr_Intkz: case X86::VFNMSUB213SSZr_Intkz:
5120     case X86::VFMSUB231SSZr_Intkz: case X86::VFNMSUB231SSZr_Intkz:
5121       return false;
5122     default:
5123       return true;
5124     }
5125   }
5126 
5127   if ((Opc == X86::MOVSDrm || Opc == X86::VMOVSDrm || Opc == X86::VMOVSDZrm) &&
5128       RegSize > 64) {
5129     // These instructions only load 64 bits, we can't fold them if the
5130     // destination register is wider than 64 bits (8 bytes), and its user
5131     // instruction isn't scalar (SD).
5132     switch (UserOpc) {
5133     case X86::ADDSDrr_Int: case X86::VADDSDrr_Int: case X86::VADDSDZrr_Int:
5134     case X86::CMPSDrr_Int: case X86::VCMPSDrr_Int: case X86::VCMPSDZrr_Int:
5135     case X86::DIVSDrr_Int: case X86::VDIVSDrr_Int: case X86::VDIVSDZrr_Int:
5136     case X86::MAXSDrr_Int: case X86::VMAXSDrr_Int: case X86::VMAXSDZrr_Int:
5137     case X86::MINSDrr_Int: case X86::VMINSDrr_Int: case X86::VMINSDZrr_Int:
5138     case X86::MULSDrr_Int: case X86::VMULSDrr_Int: case X86::VMULSDZrr_Int:
5139     case X86::SUBSDrr_Int: case X86::VSUBSDrr_Int: case X86::VSUBSDZrr_Int:
5140     case X86::VADDSDZrr_Intk: case X86::VADDSDZrr_Intkz:
5141     case X86::VDIVSDZrr_Intk: case X86::VDIVSDZrr_Intkz:
5142     case X86::VMAXSDZrr_Intk: case X86::VMAXSDZrr_Intkz:
5143     case X86::VMINSDZrr_Intk: case X86::VMINSDZrr_Intkz:
5144     case X86::VMULSDZrr_Intk: case X86::VMULSDZrr_Intkz:
5145     case X86::VSUBSDZrr_Intk: case X86::VSUBSDZrr_Intkz:
5146     case X86::VFMADDSD4rr_Int:   case X86::VFNMADDSD4rr_Int:
5147     case X86::VFMSUBSD4rr_Int:   case X86::VFNMSUBSD4rr_Int:
5148     case X86::VFMADD132SDr_Int:  case X86::VFNMADD132SDr_Int:
5149     case X86::VFMADD213SDr_Int:  case X86::VFNMADD213SDr_Int:
5150     case X86::VFMADD231SDr_Int:  case X86::VFNMADD231SDr_Int:
5151     case X86::VFMSUB132SDr_Int:  case X86::VFNMSUB132SDr_Int:
5152     case X86::VFMSUB213SDr_Int:  case X86::VFNMSUB213SDr_Int:
5153     case X86::VFMSUB231SDr_Int:  case X86::VFNMSUB231SDr_Int:
5154     case X86::VFMADD132SDZr_Int: case X86::VFNMADD132SDZr_Int:
5155     case X86::VFMADD213SDZr_Int: case X86::VFNMADD213SDZr_Int:
5156     case X86::VFMADD231SDZr_Int: case X86::VFNMADD231SDZr_Int:
5157     case X86::VFMSUB132SDZr_Int: case X86::VFNMSUB132SDZr_Int:
5158     case X86::VFMSUB213SDZr_Int: case X86::VFNMSUB213SDZr_Int:
5159     case X86::VFMSUB231SDZr_Int: case X86::VFNMSUB231SDZr_Int:
5160     case X86::VFMADD132SDZr_Intk: case X86::VFNMADD132SDZr_Intk:
5161     case X86::VFMADD213SDZr_Intk: case X86::VFNMADD213SDZr_Intk:
5162     case X86::VFMADD231SDZr_Intk: case X86::VFNMADD231SDZr_Intk:
5163     case X86::VFMSUB132SDZr_Intk: case X86::VFNMSUB132SDZr_Intk:
5164     case X86::VFMSUB213SDZr_Intk: case X86::VFNMSUB213SDZr_Intk:
5165     case X86::VFMSUB231SDZr_Intk: case X86::VFNMSUB231SDZr_Intk:
5166     case X86::VFMADD132SDZr_Intkz: case X86::VFNMADD132SDZr_Intkz:
5167     case X86::VFMADD213SDZr_Intkz: case X86::VFNMADD213SDZr_Intkz:
5168     case X86::VFMADD231SDZr_Intkz: case X86::VFNMADD231SDZr_Intkz:
5169     case X86::VFMSUB132SDZr_Intkz: case X86::VFNMSUB132SDZr_Intkz:
5170     case X86::VFMSUB213SDZr_Intkz: case X86::VFNMSUB213SDZr_Intkz:
5171     case X86::VFMSUB231SDZr_Intkz: case X86::VFNMSUB231SDZr_Intkz:
5172       return false;
5173     default:
5174       return true;
5175     }
5176   }
5177 
5178   return false;
5179 }
5180 
foldMemoryOperandImpl(MachineFunction & MF,MachineInstr & MI,ArrayRef<unsigned> Ops,MachineBasicBlock::iterator InsertPt,MachineInstr & LoadMI,LiveIntervals * LIS) const5181 MachineInstr *X86InstrInfo::foldMemoryOperandImpl(
5182     MachineFunction &MF, MachineInstr &MI, ArrayRef<unsigned> Ops,
5183     MachineBasicBlock::iterator InsertPt, MachineInstr &LoadMI,
5184     LiveIntervals *LIS) const {
5185 
5186   // TODO: Support the case where LoadMI loads a wide register, but MI
5187   // only uses a subreg.
5188   for (auto Op : Ops) {
5189     if (MI.getOperand(Op).getSubReg())
5190       return nullptr;
5191   }
5192 
5193   // If loading from a FrameIndex, fold directly from the FrameIndex.
5194   unsigned NumOps = LoadMI.getDesc().getNumOperands();
5195   int FrameIndex;
5196   if (isLoadFromStackSlot(LoadMI, FrameIndex)) {
5197     if (isNonFoldablePartialRegisterLoad(LoadMI, MI, MF))
5198       return nullptr;
5199     return foldMemoryOperandImpl(MF, MI, Ops, InsertPt, FrameIndex, LIS);
5200   }
5201 
5202   // Check switch flag
5203   if (NoFusing) return nullptr;
5204 
5205   // Avoid partial and undef register update stalls unless optimizing for size.
5206   if (!MF.getFunction().optForSize() &&
5207       (hasPartialRegUpdate(MI.getOpcode(), Subtarget) ||
5208        shouldPreventUndefRegUpdateMemFold(MF, MI)))
5209     return nullptr;
5210 
5211   // Determine the alignment of the load.
5212   unsigned Alignment = 0;
5213   if (LoadMI.hasOneMemOperand())
5214     Alignment = (*LoadMI.memoperands_begin())->getAlignment();
5215   else
5216     switch (LoadMI.getOpcode()) {
5217     case X86::AVX512_512_SET0:
5218     case X86::AVX512_512_SETALLONES:
5219       Alignment = 64;
5220       break;
5221     case X86::AVX2_SETALLONES:
5222     case X86::AVX1_SETALLONES:
5223     case X86::AVX_SET0:
5224     case X86::AVX512_256_SET0:
5225       Alignment = 32;
5226       break;
5227     case X86::V_SET0:
5228     case X86::V_SETALLONES:
5229     case X86::AVX512_128_SET0:
5230       Alignment = 16;
5231       break;
5232     case X86::MMX_SET0:
5233     case X86::FsFLD0SD:
5234     case X86::AVX512_FsFLD0SD:
5235       Alignment = 8;
5236       break;
5237     case X86::FsFLD0SS:
5238     case X86::AVX512_FsFLD0SS:
5239       Alignment = 4;
5240       break;
5241     default:
5242       return nullptr;
5243     }
5244   if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
5245     unsigned NewOpc = 0;
5246     switch (MI.getOpcode()) {
5247     default: return nullptr;
5248     case X86::TEST8rr:  NewOpc = X86::CMP8ri; break;
5249     case X86::TEST16rr: NewOpc = X86::CMP16ri8; break;
5250     case X86::TEST32rr: NewOpc = X86::CMP32ri8; break;
5251     case X86::TEST64rr: NewOpc = X86::CMP64ri8; break;
5252     }
5253     // Change to CMPXXri r, 0 first.
5254     MI.setDesc(get(NewOpc));
5255     MI.getOperand(1).ChangeToImmediate(0);
5256   } else if (Ops.size() != 1)
5257     return nullptr;
5258 
5259   // Make sure the subregisters match.
5260   // Otherwise we risk changing the size of the load.
5261   if (LoadMI.getOperand(0).getSubReg() != MI.getOperand(Ops[0]).getSubReg())
5262     return nullptr;
5263 
5264   SmallVector<MachineOperand,X86::AddrNumOperands> MOs;
5265   switch (LoadMI.getOpcode()) {
5266   case X86::MMX_SET0:
5267   case X86::V_SET0:
5268   case X86::V_SETALLONES:
5269   case X86::AVX2_SETALLONES:
5270   case X86::AVX1_SETALLONES:
5271   case X86::AVX_SET0:
5272   case X86::AVX512_128_SET0:
5273   case X86::AVX512_256_SET0:
5274   case X86::AVX512_512_SET0:
5275   case X86::AVX512_512_SETALLONES:
5276   case X86::FsFLD0SD:
5277   case X86::AVX512_FsFLD0SD:
5278   case X86::FsFLD0SS:
5279   case X86::AVX512_FsFLD0SS: {
5280     // Folding a V_SET0 or V_SETALLONES as a load, to ease register pressure.
5281     // Create a constant-pool entry and operands to load from it.
5282 
5283     // Medium and large mode can't fold loads this way.
5284     if (MF.getTarget().getCodeModel() != CodeModel::Small &&
5285         MF.getTarget().getCodeModel() != CodeModel::Kernel)
5286       return nullptr;
5287 
5288     // x86-32 PIC requires a PIC base register for constant pools.
5289     unsigned PICBase = 0;
5290     if (MF.getTarget().isPositionIndependent()) {
5291       if (Subtarget.is64Bit())
5292         PICBase = X86::RIP;
5293       else
5294         // FIXME: PICBase = getGlobalBaseReg(&MF);
5295         // This doesn't work for several reasons.
5296         // 1. GlobalBaseReg may have been spilled.
5297         // 2. It may not be live at MI.
5298         return nullptr;
5299     }
5300 
5301     // Create a constant-pool entry.
5302     MachineConstantPool &MCP = *MF.getConstantPool();
5303     Type *Ty;
5304     unsigned Opc = LoadMI.getOpcode();
5305     if (Opc == X86::FsFLD0SS || Opc == X86::AVX512_FsFLD0SS)
5306       Ty = Type::getFloatTy(MF.getFunction().getContext());
5307     else if (Opc == X86::FsFLD0SD || Opc == X86::AVX512_FsFLD0SD)
5308       Ty = Type::getDoubleTy(MF.getFunction().getContext());
5309     else if (Opc == X86::AVX512_512_SET0 || Opc == X86::AVX512_512_SETALLONES)
5310       Ty = VectorType::get(Type::getInt32Ty(MF.getFunction().getContext()),16);
5311     else if (Opc == X86::AVX2_SETALLONES || Opc == X86::AVX_SET0 ||
5312              Opc == X86::AVX512_256_SET0 || Opc == X86::AVX1_SETALLONES)
5313       Ty = VectorType::get(Type::getInt32Ty(MF.getFunction().getContext()), 8);
5314     else if (Opc == X86::MMX_SET0)
5315       Ty = VectorType::get(Type::getInt32Ty(MF.getFunction().getContext()), 2);
5316     else
5317       Ty = VectorType::get(Type::getInt32Ty(MF.getFunction().getContext()), 4);
5318 
5319     bool IsAllOnes = (Opc == X86::V_SETALLONES || Opc == X86::AVX2_SETALLONES ||
5320                       Opc == X86::AVX512_512_SETALLONES ||
5321                       Opc == X86::AVX1_SETALLONES);
5322     const Constant *C = IsAllOnes ? Constant::getAllOnesValue(Ty) :
5323                                     Constant::getNullValue(Ty);
5324     unsigned CPI = MCP.getConstantPoolIndex(C, Alignment);
5325 
5326     // Create operands to load from the constant pool entry.
5327     MOs.push_back(MachineOperand::CreateReg(PICBase, false));
5328     MOs.push_back(MachineOperand::CreateImm(1));
5329     MOs.push_back(MachineOperand::CreateReg(0, false));
5330     MOs.push_back(MachineOperand::CreateCPI(CPI, 0));
5331     MOs.push_back(MachineOperand::CreateReg(0, false));
5332     break;
5333   }
5334   default: {
5335     if (isNonFoldablePartialRegisterLoad(LoadMI, MI, MF))
5336       return nullptr;
5337 
5338     // Folding a normal load. Just copy the load's address operands.
5339     MOs.append(LoadMI.operands_begin() + NumOps - X86::AddrNumOperands,
5340                LoadMI.operands_begin() + NumOps);
5341     break;
5342   }
5343   }
5344   return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, InsertPt,
5345                                /*Size=*/0, Alignment, /*AllowCommute=*/true);
5346 }
5347 
5348 static SmallVector<MachineMemOperand *, 2>
extractLoadMMOs(ArrayRef<MachineMemOperand * > MMOs,MachineFunction & MF)5349 extractLoadMMOs(ArrayRef<MachineMemOperand *> MMOs, MachineFunction &MF) {
5350   SmallVector<MachineMemOperand *, 2> LoadMMOs;
5351 
5352   for (MachineMemOperand *MMO : MMOs) {
5353     if (!MMO->isLoad())
5354       continue;
5355 
5356     if (!MMO->isStore()) {
5357       // Reuse the MMO.
5358       LoadMMOs.push_back(MMO);
5359     } else {
5360       // Clone the MMO and unset the store flag.
5361       LoadMMOs.push_back(MF.getMachineMemOperand(
5362           MMO->getPointerInfo(), MMO->getFlags() & ~MachineMemOperand::MOStore,
5363           MMO->getSize(), MMO->getBaseAlignment(), MMO->getAAInfo(), nullptr,
5364           MMO->getSyncScopeID(), MMO->getOrdering(),
5365           MMO->getFailureOrdering()));
5366     }
5367   }
5368 
5369   return LoadMMOs;
5370 }
5371 
5372 static SmallVector<MachineMemOperand *, 2>
extractStoreMMOs(ArrayRef<MachineMemOperand * > MMOs,MachineFunction & MF)5373 extractStoreMMOs(ArrayRef<MachineMemOperand *> MMOs, MachineFunction &MF) {
5374   SmallVector<MachineMemOperand *, 2> StoreMMOs;
5375 
5376   for (MachineMemOperand *MMO : MMOs) {
5377     if (!MMO->isStore())
5378       continue;
5379 
5380     if (!MMO->isLoad()) {
5381       // Reuse the MMO.
5382       StoreMMOs.push_back(MMO);
5383     } else {
5384       // Clone the MMO and unset the load flag.
5385       StoreMMOs.push_back(MF.getMachineMemOperand(
5386           MMO->getPointerInfo(), MMO->getFlags() & ~MachineMemOperand::MOLoad,
5387           MMO->getSize(), MMO->getBaseAlignment(), MMO->getAAInfo(), nullptr,
5388           MMO->getSyncScopeID(), MMO->getOrdering(),
5389           MMO->getFailureOrdering()));
5390     }
5391   }
5392 
5393   return StoreMMOs;
5394 }
5395 
unfoldMemoryOperand(MachineFunction & MF,MachineInstr & MI,unsigned Reg,bool UnfoldLoad,bool UnfoldStore,SmallVectorImpl<MachineInstr * > & NewMIs) const5396 bool X86InstrInfo::unfoldMemoryOperand(
5397     MachineFunction &MF, MachineInstr &MI, unsigned Reg, bool UnfoldLoad,
5398     bool UnfoldStore, SmallVectorImpl<MachineInstr *> &NewMIs) const {
5399   const X86MemoryFoldTableEntry *I = lookupUnfoldTable(MI.getOpcode());
5400   if (I == nullptr)
5401     return false;
5402   unsigned Opc = I->DstOp;
5403   unsigned Index = I->Flags & TB_INDEX_MASK;
5404   bool FoldedLoad = I->Flags & TB_FOLDED_LOAD;
5405   bool FoldedStore = I->Flags & TB_FOLDED_STORE;
5406   if (UnfoldLoad && !FoldedLoad)
5407     return false;
5408   UnfoldLoad &= FoldedLoad;
5409   if (UnfoldStore && !FoldedStore)
5410     return false;
5411   UnfoldStore &= FoldedStore;
5412 
5413   const MCInstrDesc &MCID = get(Opc);
5414   const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
5415   // TODO: Check if 32-byte or greater accesses are slow too?
5416   if (!MI.hasOneMemOperand() && RC == &X86::VR128RegClass &&
5417       Subtarget.isUnalignedMem16Slow())
5418     // Without memoperands, loadRegFromAddr and storeRegToStackSlot will
5419     // conservatively assume the address is unaligned. That's bad for
5420     // performance.
5421     return false;
5422   SmallVector<MachineOperand, X86::AddrNumOperands> AddrOps;
5423   SmallVector<MachineOperand,2> BeforeOps;
5424   SmallVector<MachineOperand,2> AfterOps;
5425   SmallVector<MachineOperand,4> ImpOps;
5426   for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
5427     MachineOperand &Op = MI.getOperand(i);
5428     if (i >= Index && i < Index + X86::AddrNumOperands)
5429       AddrOps.push_back(Op);
5430     else if (Op.isReg() && Op.isImplicit())
5431       ImpOps.push_back(Op);
5432     else if (i < Index)
5433       BeforeOps.push_back(Op);
5434     else if (i > Index)
5435       AfterOps.push_back(Op);
5436   }
5437 
5438   // Emit the load instruction.
5439   if (UnfoldLoad) {
5440     auto MMOs = extractLoadMMOs(MI.memoperands(), MF);
5441     loadRegFromAddr(MF, Reg, AddrOps, RC, MMOs, NewMIs);
5442     if (UnfoldStore) {
5443       // Address operands cannot be marked isKill.
5444       for (unsigned i = 1; i != 1 + X86::AddrNumOperands; ++i) {
5445         MachineOperand &MO = NewMIs[0]->getOperand(i);
5446         if (MO.isReg())
5447           MO.setIsKill(false);
5448       }
5449     }
5450   }
5451 
5452   // Emit the data processing instruction.
5453   MachineInstr *DataMI = MF.CreateMachineInstr(MCID, MI.getDebugLoc(), true);
5454   MachineInstrBuilder MIB(MF, DataMI);
5455 
5456   if (FoldedStore)
5457     MIB.addReg(Reg, RegState::Define);
5458   for (MachineOperand &BeforeOp : BeforeOps)
5459     MIB.add(BeforeOp);
5460   if (FoldedLoad)
5461     MIB.addReg(Reg);
5462   for (MachineOperand &AfterOp : AfterOps)
5463     MIB.add(AfterOp);
5464   for (MachineOperand &ImpOp : ImpOps) {
5465     MIB.addReg(ImpOp.getReg(),
5466                getDefRegState(ImpOp.isDef()) |
5467                RegState::Implicit |
5468                getKillRegState(ImpOp.isKill()) |
5469                getDeadRegState(ImpOp.isDead()) |
5470                getUndefRegState(ImpOp.isUndef()));
5471   }
5472   // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
5473   switch (DataMI->getOpcode()) {
5474   default: break;
5475   case X86::CMP64ri32:
5476   case X86::CMP64ri8:
5477   case X86::CMP32ri:
5478   case X86::CMP32ri8:
5479   case X86::CMP16ri:
5480   case X86::CMP16ri8:
5481   case X86::CMP8ri: {
5482     MachineOperand &MO0 = DataMI->getOperand(0);
5483     MachineOperand &MO1 = DataMI->getOperand(1);
5484     if (MO1.getImm() == 0) {
5485       unsigned NewOpc;
5486       switch (DataMI->getOpcode()) {
5487       default: llvm_unreachable("Unreachable!");
5488       case X86::CMP64ri8:
5489       case X86::CMP64ri32: NewOpc = X86::TEST64rr; break;
5490       case X86::CMP32ri8:
5491       case X86::CMP32ri:   NewOpc = X86::TEST32rr; break;
5492       case X86::CMP16ri8:
5493       case X86::CMP16ri:   NewOpc = X86::TEST16rr; break;
5494       case X86::CMP8ri:    NewOpc = X86::TEST8rr; break;
5495       }
5496       DataMI->setDesc(get(NewOpc));
5497       MO1.ChangeToRegister(MO0.getReg(), false);
5498     }
5499   }
5500   }
5501   NewMIs.push_back(DataMI);
5502 
5503   // Emit the store instruction.
5504   if (UnfoldStore) {
5505     const TargetRegisterClass *DstRC = getRegClass(MCID, 0, &RI, MF);
5506     auto MMOs = extractStoreMMOs(MI.memoperands(), MF);
5507     storeRegToAddr(MF, Reg, true, AddrOps, DstRC, MMOs, NewMIs);
5508   }
5509 
5510   return true;
5511 }
5512 
5513 bool
unfoldMemoryOperand(SelectionDAG & DAG,SDNode * N,SmallVectorImpl<SDNode * > & NewNodes) const5514 X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
5515                                   SmallVectorImpl<SDNode*> &NewNodes) const {
5516   if (!N->isMachineOpcode())
5517     return false;
5518 
5519   const X86MemoryFoldTableEntry *I = lookupUnfoldTable(N->getMachineOpcode());
5520   if (I == nullptr)
5521     return false;
5522   unsigned Opc = I->DstOp;
5523   unsigned Index = I->Flags & TB_INDEX_MASK;
5524   bool FoldedLoad = I->Flags & TB_FOLDED_LOAD;
5525   bool FoldedStore = I->Flags & TB_FOLDED_STORE;
5526   const MCInstrDesc &MCID = get(Opc);
5527   MachineFunction &MF = DAG.getMachineFunction();
5528   const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
5529   const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
5530   unsigned NumDefs = MCID.NumDefs;
5531   std::vector<SDValue> AddrOps;
5532   std::vector<SDValue> BeforeOps;
5533   std::vector<SDValue> AfterOps;
5534   SDLoc dl(N);
5535   unsigned NumOps = N->getNumOperands();
5536   for (unsigned i = 0; i != NumOps-1; ++i) {
5537     SDValue Op = N->getOperand(i);
5538     if (i >= Index-NumDefs && i < Index-NumDefs + X86::AddrNumOperands)
5539       AddrOps.push_back(Op);
5540     else if (i < Index-NumDefs)
5541       BeforeOps.push_back(Op);
5542     else if (i > Index-NumDefs)
5543       AfterOps.push_back(Op);
5544   }
5545   SDValue Chain = N->getOperand(NumOps-1);
5546   AddrOps.push_back(Chain);
5547 
5548   // Emit the load instruction.
5549   SDNode *Load = nullptr;
5550   if (FoldedLoad) {
5551     EVT VT = *TRI.legalclasstypes_begin(*RC);
5552     auto MMOs = extractLoadMMOs(cast<MachineSDNode>(N)->memoperands(), MF);
5553     if (MMOs.empty() && RC == &X86::VR128RegClass &&
5554         Subtarget.isUnalignedMem16Slow())
5555       // Do not introduce a slow unaligned load.
5556       return false;
5557     // FIXME: If a VR128 can have size 32, we should be checking if a 32-byte
5558     // memory access is slow above.
5559     unsigned Alignment = std::max<uint32_t>(TRI.getSpillSize(*RC), 16);
5560     bool isAligned = !MMOs.empty() && MMOs.front()->getAlignment() >= Alignment;
5561     Load = DAG.getMachineNode(getLoadRegOpcode(0, RC, isAligned, Subtarget), dl,
5562                               VT, MVT::Other, AddrOps);
5563     NewNodes.push_back(Load);
5564 
5565     // Preserve memory reference information.
5566     DAG.setNodeMemRefs(cast<MachineSDNode>(Load), MMOs);
5567   }
5568 
5569   // Emit the data processing instruction.
5570   std::vector<EVT> VTs;
5571   const TargetRegisterClass *DstRC = nullptr;
5572   if (MCID.getNumDefs() > 0) {
5573     DstRC = getRegClass(MCID, 0, &RI, MF);
5574     VTs.push_back(*TRI.legalclasstypes_begin(*DstRC));
5575   }
5576   for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
5577     EVT VT = N->getValueType(i);
5578     if (VT != MVT::Other && i >= (unsigned)MCID.getNumDefs())
5579       VTs.push_back(VT);
5580   }
5581   if (Load)
5582     BeforeOps.push_back(SDValue(Load, 0));
5583   BeforeOps.insert(BeforeOps.end(), AfterOps.begin(), AfterOps.end());
5584   // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
5585   switch (Opc) {
5586     default: break;
5587     case X86::CMP64ri32:
5588     case X86::CMP64ri8:
5589     case X86::CMP32ri:
5590     case X86::CMP32ri8:
5591     case X86::CMP16ri:
5592     case X86::CMP16ri8:
5593     case X86::CMP8ri:
5594       if (isNullConstant(BeforeOps[1])) {
5595         switch (Opc) {
5596           default: llvm_unreachable("Unreachable!");
5597           case X86::CMP64ri8:
5598           case X86::CMP64ri32: Opc = X86::TEST64rr; break;
5599           case X86::CMP32ri8:
5600           case X86::CMP32ri:   Opc = X86::TEST32rr; break;
5601           case X86::CMP16ri8:
5602           case X86::CMP16ri:   Opc = X86::TEST16rr; break;
5603           case X86::CMP8ri:    Opc = X86::TEST8rr; break;
5604         }
5605         BeforeOps[1] = BeforeOps[0];
5606       }
5607   }
5608   SDNode *NewNode= DAG.getMachineNode(Opc, dl, VTs, BeforeOps);
5609   NewNodes.push_back(NewNode);
5610 
5611   // Emit the store instruction.
5612   if (FoldedStore) {
5613     AddrOps.pop_back();
5614     AddrOps.push_back(SDValue(NewNode, 0));
5615     AddrOps.push_back(Chain);
5616     auto MMOs = extractStoreMMOs(cast<MachineSDNode>(N)->memoperands(), MF);
5617     if (MMOs.empty() && RC == &X86::VR128RegClass &&
5618         Subtarget.isUnalignedMem16Slow())
5619       // Do not introduce a slow unaligned store.
5620       return false;
5621     // FIXME: If a VR128 can have size 32, we should be checking if a 32-byte
5622     // memory access is slow above.
5623     unsigned Alignment = std::max<uint32_t>(TRI.getSpillSize(*RC), 16);
5624     bool isAligned = !MMOs.empty() && MMOs.front()->getAlignment() >= Alignment;
5625     SDNode *Store =
5626         DAG.getMachineNode(getStoreRegOpcode(0, DstRC, isAligned, Subtarget),
5627                            dl, MVT::Other, AddrOps);
5628     NewNodes.push_back(Store);
5629 
5630     // Preserve memory reference information.
5631     DAG.setNodeMemRefs(cast<MachineSDNode>(Store), MMOs);
5632   }
5633 
5634   return true;
5635 }
5636 
getOpcodeAfterMemoryUnfold(unsigned Opc,bool UnfoldLoad,bool UnfoldStore,unsigned * LoadRegIndex) const5637 unsigned X86InstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc,
5638                                       bool UnfoldLoad, bool UnfoldStore,
5639                                       unsigned *LoadRegIndex) const {
5640   const X86MemoryFoldTableEntry *I = lookupUnfoldTable(Opc);
5641   if (I == nullptr)
5642     return 0;
5643   bool FoldedLoad = I->Flags & TB_FOLDED_LOAD;
5644   bool FoldedStore = I->Flags & TB_FOLDED_STORE;
5645   if (UnfoldLoad && !FoldedLoad)
5646     return 0;
5647   if (UnfoldStore && !FoldedStore)
5648     return 0;
5649   if (LoadRegIndex)
5650     *LoadRegIndex = I->Flags & TB_INDEX_MASK;
5651   return I->DstOp;
5652 }
5653 
5654 bool
areLoadsFromSameBasePtr(SDNode * Load1,SDNode * Load2,int64_t & Offset1,int64_t & Offset2) const5655 X86InstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
5656                                      int64_t &Offset1, int64_t &Offset2) const {
5657   if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode())
5658     return false;
5659   unsigned Opc1 = Load1->getMachineOpcode();
5660   unsigned Opc2 = Load2->getMachineOpcode();
5661   switch (Opc1) {
5662   default: return false;
5663   case X86::MOV8rm:
5664   case X86::MOV16rm:
5665   case X86::MOV32rm:
5666   case X86::MOV64rm:
5667   case X86::LD_Fp32m:
5668   case X86::LD_Fp64m:
5669   case X86::LD_Fp80m:
5670   case X86::MOVSSrm:
5671   case X86::MOVSDrm:
5672   case X86::MMX_MOVD64rm:
5673   case X86::MMX_MOVQ64rm:
5674   case X86::MOVAPSrm:
5675   case X86::MOVUPSrm:
5676   case X86::MOVAPDrm:
5677   case X86::MOVUPDrm:
5678   case X86::MOVDQArm:
5679   case X86::MOVDQUrm:
5680   // AVX load instructions
5681   case X86::VMOVSSrm:
5682   case X86::VMOVSDrm:
5683   case X86::VMOVAPSrm:
5684   case X86::VMOVUPSrm:
5685   case X86::VMOVAPDrm:
5686   case X86::VMOVUPDrm:
5687   case X86::VMOVDQArm:
5688   case X86::VMOVDQUrm:
5689   case X86::VMOVAPSYrm:
5690   case X86::VMOVUPSYrm:
5691   case X86::VMOVAPDYrm:
5692   case X86::VMOVUPDYrm:
5693   case X86::VMOVDQAYrm:
5694   case X86::VMOVDQUYrm:
5695   // AVX512 load instructions
5696   case X86::VMOVSSZrm:
5697   case X86::VMOVSDZrm:
5698   case X86::VMOVAPSZ128rm:
5699   case X86::VMOVUPSZ128rm:
5700   case X86::VMOVAPSZ128rm_NOVLX:
5701   case X86::VMOVUPSZ128rm_NOVLX:
5702   case X86::VMOVAPDZ128rm:
5703   case X86::VMOVUPDZ128rm:
5704   case X86::VMOVDQU8Z128rm:
5705   case X86::VMOVDQU16Z128rm:
5706   case X86::VMOVDQA32Z128rm:
5707   case X86::VMOVDQU32Z128rm:
5708   case X86::VMOVDQA64Z128rm:
5709   case X86::VMOVDQU64Z128rm:
5710   case X86::VMOVAPSZ256rm:
5711   case X86::VMOVUPSZ256rm:
5712   case X86::VMOVAPSZ256rm_NOVLX:
5713   case X86::VMOVUPSZ256rm_NOVLX:
5714   case X86::VMOVAPDZ256rm:
5715   case X86::VMOVUPDZ256rm:
5716   case X86::VMOVDQU8Z256rm:
5717   case X86::VMOVDQU16Z256rm:
5718   case X86::VMOVDQA32Z256rm:
5719   case X86::VMOVDQU32Z256rm:
5720   case X86::VMOVDQA64Z256rm:
5721   case X86::VMOVDQU64Z256rm:
5722   case X86::VMOVAPSZrm:
5723   case X86::VMOVUPSZrm:
5724   case X86::VMOVAPDZrm:
5725   case X86::VMOVUPDZrm:
5726   case X86::VMOVDQU8Zrm:
5727   case X86::VMOVDQU16Zrm:
5728   case X86::VMOVDQA32Zrm:
5729   case X86::VMOVDQU32Zrm:
5730   case X86::VMOVDQA64Zrm:
5731   case X86::VMOVDQU64Zrm:
5732   case X86::KMOVBkm:
5733   case X86::KMOVWkm:
5734   case X86::KMOVDkm:
5735   case X86::KMOVQkm:
5736     break;
5737   }
5738   switch (Opc2) {
5739   default: return false;
5740   case X86::MOV8rm:
5741   case X86::MOV16rm:
5742   case X86::MOV32rm:
5743   case X86::MOV64rm:
5744   case X86::LD_Fp32m:
5745   case X86::LD_Fp64m:
5746   case X86::LD_Fp80m:
5747   case X86::MOVSSrm:
5748   case X86::MOVSDrm:
5749   case X86::MMX_MOVD64rm:
5750   case X86::MMX_MOVQ64rm:
5751   case X86::MOVAPSrm:
5752   case X86::MOVUPSrm:
5753   case X86::MOVAPDrm:
5754   case X86::MOVUPDrm:
5755   case X86::MOVDQArm:
5756   case X86::MOVDQUrm:
5757   // AVX load instructions
5758   case X86::VMOVSSrm:
5759   case X86::VMOVSDrm:
5760   case X86::VMOVAPSrm:
5761   case X86::VMOVUPSrm:
5762   case X86::VMOVAPDrm:
5763   case X86::VMOVUPDrm:
5764   case X86::VMOVDQArm:
5765   case X86::VMOVDQUrm:
5766   case X86::VMOVAPSYrm:
5767   case X86::VMOVUPSYrm:
5768   case X86::VMOVAPDYrm:
5769   case X86::VMOVUPDYrm:
5770   case X86::VMOVDQAYrm:
5771   case X86::VMOVDQUYrm:
5772   // AVX512 load instructions
5773   case X86::VMOVSSZrm:
5774   case X86::VMOVSDZrm:
5775   case X86::VMOVAPSZ128rm:
5776   case X86::VMOVUPSZ128rm:
5777   case X86::VMOVAPSZ128rm_NOVLX:
5778   case X86::VMOVUPSZ128rm_NOVLX:
5779   case X86::VMOVAPDZ128rm:
5780   case X86::VMOVUPDZ128rm:
5781   case X86::VMOVDQU8Z128rm:
5782   case X86::VMOVDQU16Z128rm:
5783   case X86::VMOVDQA32Z128rm:
5784   case X86::VMOVDQU32Z128rm:
5785   case X86::VMOVDQA64Z128rm:
5786   case X86::VMOVDQU64Z128rm:
5787   case X86::VMOVAPSZ256rm:
5788   case X86::VMOVUPSZ256rm:
5789   case X86::VMOVAPSZ256rm_NOVLX:
5790   case X86::VMOVUPSZ256rm_NOVLX:
5791   case X86::VMOVAPDZ256rm:
5792   case X86::VMOVUPDZ256rm:
5793   case X86::VMOVDQU8Z256rm:
5794   case X86::VMOVDQU16Z256rm:
5795   case X86::VMOVDQA32Z256rm:
5796   case X86::VMOVDQU32Z256rm:
5797   case X86::VMOVDQA64Z256rm:
5798   case X86::VMOVDQU64Z256rm:
5799   case X86::VMOVAPSZrm:
5800   case X86::VMOVUPSZrm:
5801   case X86::VMOVAPDZrm:
5802   case X86::VMOVUPDZrm:
5803   case X86::VMOVDQU8Zrm:
5804   case X86::VMOVDQU16Zrm:
5805   case X86::VMOVDQA32Zrm:
5806   case X86::VMOVDQU32Zrm:
5807   case X86::VMOVDQA64Zrm:
5808   case X86::VMOVDQU64Zrm:
5809   case X86::KMOVBkm:
5810   case X86::KMOVWkm:
5811   case X86::KMOVDkm:
5812   case X86::KMOVQkm:
5813     break;
5814   }
5815 
5816   // Lambda to check if both the loads have the same value for an operand index.
5817   auto HasSameOp = [&](int I) {
5818     return Load1->getOperand(I) == Load2->getOperand(I);
5819   };
5820 
5821   // All operands except the displacement should match.
5822   if (!HasSameOp(X86::AddrBaseReg) || !HasSameOp(X86::AddrScaleAmt) ||
5823       !HasSameOp(X86::AddrIndexReg) || !HasSameOp(X86::AddrSegmentReg))
5824     return false;
5825 
5826   // Chain Operand must be the same.
5827   if (!HasSameOp(5))
5828     return false;
5829 
5830   // Now let's examine if the displacements are constants.
5831   auto Disp1 = dyn_cast<ConstantSDNode>(Load1->getOperand(X86::AddrDisp));
5832   auto Disp2 = dyn_cast<ConstantSDNode>(Load2->getOperand(X86::AddrDisp));
5833   if (!Disp1 || !Disp2)
5834     return false;
5835 
5836   Offset1 = Disp1->getSExtValue();
5837   Offset2 = Disp2->getSExtValue();
5838   return true;
5839 }
5840 
shouldScheduleLoadsNear(SDNode * Load1,SDNode * Load2,int64_t Offset1,int64_t Offset2,unsigned NumLoads) const5841 bool X86InstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
5842                                            int64_t Offset1, int64_t Offset2,
5843                                            unsigned NumLoads) const {
5844   assert(Offset2 > Offset1);
5845   if ((Offset2 - Offset1) / 8 > 64)
5846     return false;
5847 
5848   unsigned Opc1 = Load1->getMachineOpcode();
5849   unsigned Opc2 = Load2->getMachineOpcode();
5850   if (Opc1 != Opc2)
5851     return false;  // FIXME: overly conservative?
5852 
5853   switch (Opc1) {
5854   default: break;
5855   case X86::LD_Fp32m:
5856   case X86::LD_Fp64m:
5857   case X86::LD_Fp80m:
5858   case X86::MMX_MOVD64rm:
5859   case X86::MMX_MOVQ64rm:
5860     return false;
5861   }
5862 
5863   EVT VT = Load1->getValueType(0);
5864   switch (VT.getSimpleVT().SimpleTy) {
5865   default:
5866     // XMM registers. In 64-bit mode we can be a bit more aggressive since we
5867     // have 16 of them to play with.
5868     if (Subtarget.is64Bit()) {
5869       if (NumLoads >= 3)
5870         return false;
5871     } else if (NumLoads) {
5872       return false;
5873     }
5874     break;
5875   case MVT::i8:
5876   case MVT::i16:
5877   case MVT::i32:
5878   case MVT::i64:
5879   case MVT::f32:
5880   case MVT::f64:
5881     if (NumLoads)
5882       return false;
5883     break;
5884   }
5885 
5886   return true;
5887 }
5888 
5889 bool X86InstrInfo::
reverseBranchCondition(SmallVectorImpl<MachineOperand> & Cond) const5890 reverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
5891   assert(Cond.size() == 1 && "Invalid X86 branch condition!");
5892   X86::CondCode CC = static_cast<X86::CondCode>(Cond[0].getImm());
5893   Cond[0].setImm(GetOppositeBranchCondition(CC));
5894   return false;
5895 }
5896 
5897 bool X86InstrInfo::
isSafeToMoveRegClassDefs(const TargetRegisterClass * RC) const5898 isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
5899   // FIXME: Return false for x87 stack register classes for now. We can't
5900   // allow any loads of these registers before FpGet_ST0_80.
5901   return !(RC == &X86::CCRRegClass || RC == &X86::DFCCRRegClass ||
5902            RC == &X86::RFP32RegClass || RC == &X86::RFP64RegClass ||
5903            RC == &X86::RFP80RegClass);
5904 }
5905 
5906 /// Return a virtual register initialized with the
5907 /// the global base register value. Output instructions required to
5908 /// initialize the register in the function entry block, if necessary.
5909 ///
5910 /// TODO: Eliminate this and move the code to X86MachineFunctionInfo.
5911 ///
getGlobalBaseReg(MachineFunction * MF) const5912 unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const {
5913   assert((!Subtarget.is64Bit() ||
5914           MF->getTarget().getCodeModel() == CodeModel::Medium ||
5915           MF->getTarget().getCodeModel() == CodeModel::Large) &&
5916          "X86-64 PIC uses RIP relative addressing");
5917 
5918   X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
5919   unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
5920   if (GlobalBaseReg != 0)
5921     return GlobalBaseReg;
5922 
5923   // Create the register. The code to initialize it is inserted
5924   // later, by the CGBR pass (below).
5925   MachineRegisterInfo &RegInfo = MF->getRegInfo();
5926   GlobalBaseReg = RegInfo.createVirtualRegister(
5927       Subtarget.is64Bit() ? &X86::GR64_NOSPRegClass : &X86::GR32_NOSPRegClass);
5928   X86FI->setGlobalBaseReg(GlobalBaseReg);
5929   return GlobalBaseReg;
5930 }
5931 
5932 // These are the replaceable SSE instructions. Some of these have Int variants
5933 // that we don't include here. We don't want to replace instructions selected
5934 // by intrinsics.
5935 static const uint16_t ReplaceableInstrs[][3] = {
5936   //PackedSingle     PackedDouble    PackedInt
5937   { X86::MOVAPSmr,   X86::MOVAPDmr,  X86::MOVDQAmr  },
5938   { X86::MOVAPSrm,   X86::MOVAPDrm,  X86::MOVDQArm  },
5939   { X86::MOVAPSrr,   X86::MOVAPDrr,  X86::MOVDQArr  },
5940   { X86::MOVUPSmr,   X86::MOVUPDmr,  X86::MOVDQUmr  },
5941   { X86::MOVUPSrm,   X86::MOVUPDrm,  X86::MOVDQUrm  },
5942   { X86::MOVLPSmr,   X86::MOVLPDmr,  X86::MOVPQI2QImr },
5943   { X86::MOVSDmr,    X86::MOVSDmr,   X86::MOVPQI2QImr },
5944   { X86::MOVSSmr,    X86::MOVSSmr,   X86::MOVPDI2DImr },
5945   { X86::MOVSDrm,    X86::MOVSDrm,   X86::MOVQI2PQIrm },
5946   { X86::MOVSSrm,    X86::MOVSSrm,   X86::MOVDI2PDIrm },
5947   { X86::MOVNTPSmr,  X86::MOVNTPDmr, X86::MOVNTDQmr },
5948   { X86::ANDNPSrm,   X86::ANDNPDrm,  X86::PANDNrm   },
5949   { X86::ANDNPSrr,   X86::ANDNPDrr,  X86::PANDNrr   },
5950   { X86::ANDPSrm,    X86::ANDPDrm,   X86::PANDrm    },
5951   { X86::ANDPSrr,    X86::ANDPDrr,   X86::PANDrr    },
5952   { X86::ORPSrm,     X86::ORPDrm,    X86::PORrm     },
5953   { X86::ORPSrr,     X86::ORPDrr,    X86::PORrr     },
5954   { X86::XORPSrm,    X86::XORPDrm,   X86::PXORrm    },
5955   { X86::XORPSrr,    X86::XORPDrr,   X86::PXORrr    },
5956   { X86::UNPCKLPDrm, X86::UNPCKLPDrm, X86::PUNPCKLQDQrm },
5957   { X86::MOVLHPSrr,  X86::UNPCKLPDrr, X86::PUNPCKLQDQrr },
5958   { X86::UNPCKHPDrm, X86::UNPCKHPDrm, X86::PUNPCKHQDQrm },
5959   { X86::UNPCKHPDrr, X86::UNPCKHPDrr, X86::PUNPCKHQDQrr },
5960   { X86::UNPCKLPSrm, X86::UNPCKLPSrm, X86::PUNPCKLDQrm },
5961   { X86::UNPCKLPSrr, X86::UNPCKLPSrr, X86::PUNPCKLDQrr },
5962   { X86::UNPCKHPSrm, X86::UNPCKHPSrm, X86::PUNPCKHDQrm },
5963   { X86::UNPCKHPSrr, X86::UNPCKHPSrr, X86::PUNPCKHDQrr },
5964   { X86::EXTRACTPSmr, X86::EXTRACTPSmr, X86::PEXTRDmr },
5965   { X86::EXTRACTPSrr, X86::EXTRACTPSrr, X86::PEXTRDrr },
5966   // AVX 128-bit support
5967   { X86::VMOVAPSmr,  X86::VMOVAPDmr,  X86::VMOVDQAmr  },
5968   { X86::VMOVAPSrm,  X86::VMOVAPDrm,  X86::VMOVDQArm  },
5969   { X86::VMOVAPSrr,  X86::VMOVAPDrr,  X86::VMOVDQArr  },
5970   { X86::VMOVUPSmr,  X86::VMOVUPDmr,  X86::VMOVDQUmr  },
5971   { X86::VMOVUPSrm,  X86::VMOVUPDrm,  X86::VMOVDQUrm  },
5972   { X86::VMOVLPSmr,  X86::VMOVLPDmr,  X86::VMOVPQI2QImr },
5973   { X86::VMOVSDmr,   X86::VMOVSDmr,   X86::VMOVPQI2QImr },
5974   { X86::VMOVSSmr,   X86::VMOVSSmr,   X86::VMOVPDI2DImr },
5975   { X86::VMOVSDrm,   X86::VMOVSDrm,   X86::VMOVQI2PQIrm },
5976   { X86::VMOVSSrm,   X86::VMOVSSrm,   X86::VMOVDI2PDIrm },
5977   { X86::VMOVNTPSmr, X86::VMOVNTPDmr, X86::VMOVNTDQmr },
5978   { X86::VANDNPSrm,  X86::VANDNPDrm,  X86::VPANDNrm   },
5979   { X86::VANDNPSrr,  X86::VANDNPDrr,  X86::VPANDNrr   },
5980   { X86::VANDPSrm,   X86::VANDPDrm,   X86::VPANDrm    },
5981   { X86::VANDPSrr,   X86::VANDPDrr,   X86::VPANDrr    },
5982   { X86::VORPSrm,    X86::VORPDrm,    X86::VPORrm     },
5983   { X86::VORPSrr,    X86::VORPDrr,    X86::VPORrr     },
5984   { X86::VXORPSrm,   X86::VXORPDrm,   X86::VPXORrm    },
5985   { X86::VXORPSrr,   X86::VXORPDrr,   X86::VPXORrr    },
5986   { X86::VUNPCKLPDrm, X86::VUNPCKLPDrm, X86::VPUNPCKLQDQrm },
5987   { X86::VMOVLHPSrr,  X86::VUNPCKLPDrr, X86::VPUNPCKLQDQrr },
5988   { X86::VUNPCKHPDrm, X86::VUNPCKHPDrm, X86::VPUNPCKHQDQrm },
5989   { X86::VUNPCKHPDrr, X86::VUNPCKHPDrr, X86::VPUNPCKHQDQrr },
5990   { X86::VUNPCKLPSrm, X86::VUNPCKLPSrm, X86::VPUNPCKLDQrm },
5991   { X86::VUNPCKLPSrr, X86::VUNPCKLPSrr, X86::VPUNPCKLDQrr },
5992   { X86::VUNPCKHPSrm, X86::VUNPCKHPSrm, X86::VPUNPCKHDQrm },
5993   { X86::VUNPCKHPSrr, X86::VUNPCKHPSrr, X86::VPUNPCKHDQrr },
5994   { X86::VEXTRACTPSmr, X86::VEXTRACTPSmr, X86::VPEXTRDmr },
5995   { X86::VEXTRACTPSrr, X86::VEXTRACTPSrr, X86::VPEXTRDrr },
5996   // AVX 256-bit support
5997   { X86::VMOVAPSYmr,   X86::VMOVAPDYmr,   X86::VMOVDQAYmr  },
5998   { X86::VMOVAPSYrm,   X86::VMOVAPDYrm,   X86::VMOVDQAYrm  },
5999   { X86::VMOVAPSYrr,   X86::VMOVAPDYrr,   X86::VMOVDQAYrr  },
6000   { X86::VMOVUPSYmr,   X86::VMOVUPDYmr,   X86::VMOVDQUYmr  },
6001   { X86::VMOVUPSYrm,   X86::VMOVUPDYrm,   X86::VMOVDQUYrm  },
6002   { X86::VMOVNTPSYmr,  X86::VMOVNTPDYmr,  X86::VMOVNTDQYmr },
6003   { X86::VPERMPSYrm,   X86::VPERMPSYrm,   X86::VPERMDYrm },
6004   { X86::VPERMPSYrr,   X86::VPERMPSYrr,   X86::VPERMDYrr },
6005   { X86::VPERMPDYmi,   X86::VPERMPDYmi,   X86::VPERMQYmi },
6006   { X86::VPERMPDYri,   X86::VPERMPDYri,   X86::VPERMQYri },
6007   // AVX512 support
6008   { X86::VMOVLPSZ128mr,  X86::VMOVLPDZ128mr,  X86::VMOVPQI2QIZmr  },
6009   { X86::VMOVNTPSZ128mr, X86::VMOVNTPDZ128mr, X86::VMOVNTDQZ128mr },
6010   { X86::VMOVNTPSZ256mr, X86::VMOVNTPDZ256mr, X86::VMOVNTDQZ256mr },
6011   { X86::VMOVNTPSZmr,    X86::VMOVNTPDZmr,    X86::VMOVNTDQZmr    },
6012   { X86::VMOVSDZmr,      X86::VMOVSDZmr,      X86::VMOVPQI2QIZmr  },
6013   { X86::VMOVSSZmr,      X86::VMOVSSZmr,      X86::VMOVPDI2DIZmr  },
6014   { X86::VMOVSDZrm,      X86::VMOVSDZrm,      X86::VMOVQI2PQIZrm  },
6015   { X86::VMOVSSZrm,      X86::VMOVSSZrm,      X86::VMOVDI2PDIZrm  },
6016   { X86::VBROADCASTSSZ128r, X86::VBROADCASTSSZ128r, X86::VPBROADCASTDZ128r },
6017   { X86::VBROADCASTSSZ128m, X86::VBROADCASTSSZ128m, X86::VPBROADCASTDZ128m },
6018   { X86::VBROADCASTSSZ256r, X86::VBROADCASTSSZ256r, X86::VPBROADCASTDZ256r },
6019   { X86::VBROADCASTSSZ256m, X86::VBROADCASTSSZ256m, X86::VPBROADCASTDZ256m },
6020   { X86::VBROADCASTSSZr,    X86::VBROADCASTSSZr,    X86::VPBROADCASTDZr },
6021   { X86::VBROADCASTSSZm,    X86::VBROADCASTSSZm,    X86::VPBROADCASTDZm },
6022   { X86::VBROADCASTSDZ256r, X86::VBROADCASTSDZ256r, X86::VPBROADCASTQZ256r },
6023   { X86::VBROADCASTSDZ256m, X86::VBROADCASTSDZ256m, X86::VPBROADCASTQZ256m },
6024   { X86::VBROADCASTSDZr,    X86::VBROADCASTSDZr,    X86::VPBROADCASTQZr },
6025   { X86::VBROADCASTSDZm,    X86::VBROADCASTSDZm,    X86::VPBROADCASTQZm },
6026   { X86::VINSERTF32x4Zrr,   X86::VINSERTF32x4Zrr,   X86::VINSERTI32x4Zrr },
6027   { X86::VINSERTF32x4Zrm,   X86::VINSERTF32x4Zrm,   X86::VINSERTI32x4Zrm },
6028   { X86::VINSERTF32x8Zrr,   X86::VINSERTF32x8Zrr,   X86::VINSERTI32x8Zrr },
6029   { X86::VINSERTF32x8Zrm,   X86::VINSERTF32x8Zrm,   X86::VINSERTI32x8Zrm },
6030   { X86::VINSERTF64x2Zrr,   X86::VINSERTF64x2Zrr,   X86::VINSERTI64x2Zrr },
6031   { X86::VINSERTF64x2Zrm,   X86::VINSERTF64x2Zrm,   X86::VINSERTI64x2Zrm },
6032   { X86::VINSERTF64x4Zrr,   X86::VINSERTF64x4Zrr,   X86::VINSERTI64x4Zrr },
6033   { X86::VINSERTF64x4Zrm,   X86::VINSERTF64x4Zrm,   X86::VINSERTI64x4Zrm },
6034   { X86::VINSERTF32x4Z256rr,X86::VINSERTF32x4Z256rr,X86::VINSERTI32x4Z256rr },
6035   { X86::VINSERTF32x4Z256rm,X86::VINSERTF32x4Z256rm,X86::VINSERTI32x4Z256rm },
6036   { X86::VINSERTF64x2Z256rr,X86::VINSERTF64x2Z256rr,X86::VINSERTI64x2Z256rr },
6037   { X86::VINSERTF64x2Z256rm,X86::VINSERTF64x2Z256rm,X86::VINSERTI64x2Z256rm },
6038   { X86::VEXTRACTF32x4Zrr,   X86::VEXTRACTF32x4Zrr,   X86::VEXTRACTI32x4Zrr },
6039   { X86::VEXTRACTF32x4Zmr,   X86::VEXTRACTF32x4Zmr,   X86::VEXTRACTI32x4Zmr },
6040   { X86::VEXTRACTF32x8Zrr,   X86::VEXTRACTF32x8Zrr,   X86::VEXTRACTI32x8Zrr },
6041   { X86::VEXTRACTF32x8Zmr,   X86::VEXTRACTF32x8Zmr,   X86::VEXTRACTI32x8Zmr },
6042   { X86::VEXTRACTF64x2Zrr,   X86::VEXTRACTF64x2Zrr,   X86::VEXTRACTI64x2Zrr },
6043   { X86::VEXTRACTF64x2Zmr,   X86::VEXTRACTF64x2Zmr,   X86::VEXTRACTI64x2Zmr },
6044   { X86::VEXTRACTF64x4Zrr,   X86::VEXTRACTF64x4Zrr,   X86::VEXTRACTI64x4Zrr },
6045   { X86::VEXTRACTF64x4Zmr,   X86::VEXTRACTF64x4Zmr,   X86::VEXTRACTI64x4Zmr },
6046   { X86::VEXTRACTF32x4Z256rr,X86::VEXTRACTF32x4Z256rr,X86::VEXTRACTI32x4Z256rr },
6047   { X86::VEXTRACTF32x4Z256mr,X86::VEXTRACTF32x4Z256mr,X86::VEXTRACTI32x4Z256mr },
6048   { X86::VEXTRACTF64x2Z256rr,X86::VEXTRACTF64x2Z256rr,X86::VEXTRACTI64x2Z256rr },
6049   { X86::VEXTRACTF64x2Z256mr,X86::VEXTRACTF64x2Z256mr,X86::VEXTRACTI64x2Z256mr },
6050   { X86::VPERMILPSmi,        X86::VPERMILPSmi,        X86::VPSHUFDmi },
6051   { X86::VPERMILPSri,        X86::VPERMILPSri,        X86::VPSHUFDri },
6052   { X86::VPERMILPSZ128mi,    X86::VPERMILPSZ128mi,    X86::VPSHUFDZ128mi },
6053   { X86::VPERMILPSZ128ri,    X86::VPERMILPSZ128ri,    X86::VPSHUFDZ128ri },
6054   { X86::VPERMILPSZ256mi,    X86::VPERMILPSZ256mi,    X86::VPSHUFDZ256mi },
6055   { X86::VPERMILPSZ256ri,    X86::VPERMILPSZ256ri,    X86::VPSHUFDZ256ri },
6056   { X86::VPERMILPSZmi,       X86::VPERMILPSZmi,       X86::VPSHUFDZmi },
6057   { X86::VPERMILPSZri,       X86::VPERMILPSZri,       X86::VPSHUFDZri },
6058   { X86::VPERMPSZ256rm,      X86::VPERMPSZ256rm,      X86::VPERMDZ256rm },
6059   { X86::VPERMPSZ256rr,      X86::VPERMPSZ256rr,      X86::VPERMDZ256rr },
6060   { X86::VPERMPDZ256mi,      X86::VPERMPDZ256mi,      X86::VPERMQZ256mi },
6061   { X86::VPERMPDZ256ri,      X86::VPERMPDZ256ri,      X86::VPERMQZ256ri },
6062   { X86::VPERMPDZ256rm,      X86::VPERMPDZ256rm,      X86::VPERMQZ256rm },
6063   { X86::VPERMPDZ256rr,      X86::VPERMPDZ256rr,      X86::VPERMQZ256rr },
6064   { X86::VPERMPSZrm,         X86::VPERMPSZrm,         X86::VPERMDZrm },
6065   { X86::VPERMPSZrr,         X86::VPERMPSZrr,         X86::VPERMDZrr },
6066   { X86::VPERMPDZmi,         X86::VPERMPDZmi,         X86::VPERMQZmi },
6067   { X86::VPERMPDZri,         X86::VPERMPDZri,         X86::VPERMQZri },
6068   { X86::VPERMPDZrm,         X86::VPERMPDZrm,         X86::VPERMQZrm },
6069   { X86::VPERMPDZrr,         X86::VPERMPDZrr,         X86::VPERMQZrr },
6070   { X86::VUNPCKLPDZ256rm,    X86::VUNPCKLPDZ256rm,    X86::VPUNPCKLQDQZ256rm },
6071   { X86::VUNPCKLPDZ256rr,    X86::VUNPCKLPDZ256rr,    X86::VPUNPCKLQDQZ256rr },
6072   { X86::VUNPCKHPDZ256rm,    X86::VUNPCKHPDZ256rm,    X86::VPUNPCKHQDQZ256rm },
6073   { X86::VUNPCKHPDZ256rr,    X86::VUNPCKHPDZ256rr,    X86::VPUNPCKHQDQZ256rr },
6074   { X86::VUNPCKLPSZ256rm,    X86::VUNPCKLPSZ256rm,    X86::VPUNPCKLDQZ256rm },
6075   { X86::VUNPCKLPSZ256rr,    X86::VUNPCKLPSZ256rr,    X86::VPUNPCKLDQZ256rr },
6076   { X86::VUNPCKHPSZ256rm,    X86::VUNPCKHPSZ256rm,    X86::VPUNPCKHDQZ256rm },
6077   { X86::VUNPCKHPSZ256rr,    X86::VUNPCKHPSZ256rr,    X86::VPUNPCKHDQZ256rr },
6078   { X86::VUNPCKLPDZ128rm,    X86::VUNPCKLPDZ128rm,    X86::VPUNPCKLQDQZ128rm },
6079   { X86::VMOVLHPSZrr,        X86::VUNPCKLPDZ128rr,    X86::VPUNPCKLQDQZ128rr },
6080   { X86::VUNPCKHPDZ128rm,    X86::VUNPCKHPDZ128rm,    X86::VPUNPCKHQDQZ128rm },
6081   { X86::VUNPCKHPDZ128rr,    X86::VUNPCKHPDZ128rr,    X86::VPUNPCKHQDQZ128rr },
6082   { X86::VUNPCKLPSZ128rm,    X86::VUNPCKLPSZ128rm,    X86::VPUNPCKLDQZ128rm },
6083   { X86::VUNPCKLPSZ128rr,    X86::VUNPCKLPSZ128rr,    X86::VPUNPCKLDQZ128rr },
6084   { X86::VUNPCKHPSZ128rm,    X86::VUNPCKHPSZ128rm,    X86::VPUNPCKHDQZ128rm },
6085   { X86::VUNPCKHPSZ128rr,    X86::VUNPCKHPSZ128rr,    X86::VPUNPCKHDQZ128rr },
6086   { X86::VUNPCKLPDZrm,       X86::VUNPCKLPDZrm,       X86::VPUNPCKLQDQZrm },
6087   { X86::VUNPCKLPDZrr,       X86::VUNPCKLPDZrr,       X86::VPUNPCKLQDQZrr },
6088   { X86::VUNPCKHPDZrm,       X86::VUNPCKHPDZrm,       X86::VPUNPCKHQDQZrm },
6089   { X86::VUNPCKHPDZrr,       X86::VUNPCKHPDZrr,       X86::VPUNPCKHQDQZrr },
6090   { X86::VUNPCKLPSZrm,       X86::VUNPCKLPSZrm,       X86::VPUNPCKLDQZrm },
6091   { X86::VUNPCKLPSZrr,       X86::VUNPCKLPSZrr,       X86::VPUNPCKLDQZrr },
6092   { X86::VUNPCKHPSZrm,       X86::VUNPCKHPSZrm,       X86::VPUNPCKHDQZrm },
6093   { X86::VUNPCKHPSZrr,       X86::VUNPCKHPSZrr,       X86::VPUNPCKHDQZrr },
6094   { X86::VEXTRACTPSZmr,      X86::VEXTRACTPSZmr,      X86::VPEXTRDZmr },
6095   { X86::VEXTRACTPSZrr,      X86::VEXTRACTPSZrr,      X86::VPEXTRDZrr },
6096 };
6097 
6098 static const uint16_t ReplaceableInstrsAVX2[][3] = {
6099   //PackedSingle       PackedDouble       PackedInt
6100   { X86::VANDNPSYrm,   X86::VANDNPDYrm,   X86::VPANDNYrm   },
6101   { X86::VANDNPSYrr,   X86::VANDNPDYrr,   X86::VPANDNYrr   },
6102   { X86::VANDPSYrm,    X86::VANDPDYrm,    X86::VPANDYrm    },
6103   { X86::VANDPSYrr,    X86::VANDPDYrr,    X86::VPANDYrr    },
6104   { X86::VORPSYrm,     X86::VORPDYrm,     X86::VPORYrm     },
6105   { X86::VORPSYrr,     X86::VORPDYrr,     X86::VPORYrr     },
6106   { X86::VXORPSYrm,    X86::VXORPDYrm,    X86::VPXORYrm    },
6107   { X86::VXORPSYrr,    X86::VXORPDYrr,    X86::VPXORYrr    },
6108   { X86::VPERM2F128rm,   X86::VPERM2F128rm,   X86::VPERM2I128rm },
6109   { X86::VPERM2F128rr,   X86::VPERM2F128rr,   X86::VPERM2I128rr },
6110   { X86::VBROADCASTSSrm, X86::VBROADCASTSSrm, X86::VPBROADCASTDrm},
6111   { X86::VBROADCASTSSrr, X86::VBROADCASTSSrr, X86::VPBROADCASTDrr},
6112   { X86::VBROADCASTSSYrr, X86::VBROADCASTSSYrr, X86::VPBROADCASTDYrr},
6113   { X86::VBROADCASTSSYrm, X86::VBROADCASTSSYrm, X86::VPBROADCASTDYrm},
6114   { X86::VBROADCASTSDYrr, X86::VBROADCASTSDYrr, X86::VPBROADCASTQYrr},
6115   { X86::VBROADCASTSDYrm, X86::VBROADCASTSDYrm, X86::VPBROADCASTQYrm},
6116   { X86::VBROADCASTF128,  X86::VBROADCASTF128,  X86::VBROADCASTI128 },
6117   { X86::VBLENDPSYrri,    X86::VBLENDPSYrri,    X86::VPBLENDDYrri },
6118   { X86::VBLENDPSYrmi,    X86::VBLENDPSYrmi,    X86::VPBLENDDYrmi },
6119   { X86::VPERMILPSYmi,    X86::VPERMILPSYmi,    X86::VPSHUFDYmi },
6120   { X86::VPERMILPSYri,    X86::VPERMILPSYri,    X86::VPSHUFDYri },
6121   { X86::VUNPCKLPDYrm,    X86::VUNPCKLPDYrm,    X86::VPUNPCKLQDQYrm },
6122   { X86::VUNPCKLPDYrr,    X86::VUNPCKLPDYrr,    X86::VPUNPCKLQDQYrr },
6123   { X86::VUNPCKHPDYrm,    X86::VUNPCKHPDYrm,    X86::VPUNPCKHQDQYrm },
6124   { X86::VUNPCKHPDYrr,    X86::VUNPCKHPDYrr,    X86::VPUNPCKHQDQYrr },
6125   { X86::VUNPCKLPSYrm,    X86::VUNPCKLPSYrm,    X86::VPUNPCKLDQYrm },
6126   { X86::VUNPCKLPSYrr,    X86::VUNPCKLPSYrr,    X86::VPUNPCKLDQYrr },
6127   { X86::VUNPCKHPSYrm,    X86::VUNPCKHPSYrm,    X86::VPUNPCKHDQYrm },
6128   { X86::VUNPCKHPSYrr,    X86::VUNPCKHPSYrr,    X86::VPUNPCKHDQYrr },
6129 };
6130 
6131 static const uint16_t ReplaceableInstrsAVX2InsertExtract[][3] = {
6132   //PackedSingle       PackedDouble       PackedInt
6133   { X86::VEXTRACTF128mr, X86::VEXTRACTF128mr, X86::VEXTRACTI128mr },
6134   { X86::VEXTRACTF128rr, X86::VEXTRACTF128rr, X86::VEXTRACTI128rr },
6135   { X86::VINSERTF128rm,  X86::VINSERTF128rm,  X86::VINSERTI128rm },
6136   { X86::VINSERTF128rr,  X86::VINSERTF128rr,  X86::VINSERTI128rr },
6137 };
6138 
6139 static const uint16_t ReplaceableInstrsAVX512[][4] = {
6140   // Two integer columns for 64-bit and 32-bit elements.
6141   //PackedSingle        PackedDouble        PackedInt             PackedInt
6142   { X86::VMOVAPSZ128mr, X86::VMOVAPDZ128mr, X86::VMOVDQA64Z128mr, X86::VMOVDQA32Z128mr  },
6143   { X86::VMOVAPSZ128rm, X86::VMOVAPDZ128rm, X86::VMOVDQA64Z128rm, X86::VMOVDQA32Z128rm  },
6144   { X86::VMOVAPSZ128rr, X86::VMOVAPDZ128rr, X86::VMOVDQA64Z128rr, X86::VMOVDQA32Z128rr  },
6145   { X86::VMOVUPSZ128mr, X86::VMOVUPDZ128mr, X86::VMOVDQU64Z128mr, X86::VMOVDQU32Z128mr  },
6146   { X86::VMOVUPSZ128rm, X86::VMOVUPDZ128rm, X86::VMOVDQU64Z128rm, X86::VMOVDQU32Z128rm  },
6147   { X86::VMOVAPSZ256mr, X86::VMOVAPDZ256mr, X86::VMOVDQA64Z256mr, X86::VMOVDQA32Z256mr  },
6148   { X86::VMOVAPSZ256rm, X86::VMOVAPDZ256rm, X86::VMOVDQA64Z256rm, X86::VMOVDQA32Z256rm  },
6149   { X86::VMOVAPSZ256rr, X86::VMOVAPDZ256rr, X86::VMOVDQA64Z256rr, X86::VMOVDQA32Z256rr  },
6150   { X86::VMOVUPSZ256mr, X86::VMOVUPDZ256mr, X86::VMOVDQU64Z256mr, X86::VMOVDQU32Z256mr  },
6151   { X86::VMOVUPSZ256rm, X86::VMOVUPDZ256rm, X86::VMOVDQU64Z256rm, X86::VMOVDQU32Z256rm  },
6152   { X86::VMOVAPSZmr,    X86::VMOVAPDZmr,    X86::VMOVDQA64Zmr,    X86::VMOVDQA32Zmr     },
6153   { X86::VMOVAPSZrm,    X86::VMOVAPDZrm,    X86::VMOVDQA64Zrm,    X86::VMOVDQA32Zrm     },
6154   { X86::VMOVAPSZrr,    X86::VMOVAPDZrr,    X86::VMOVDQA64Zrr,    X86::VMOVDQA32Zrr     },
6155   { X86::VMOVUPSZmr,    X86::VMOVUPDZmr,    X86::VMOVDQU64Zmr,    X86::VMOVDQU32Zmr     },
6156   { X86::VMOVUPSZrm,    X86::VMOVUPDZrm,    X86::VMOVDQU64Zrm,    X86::VMOVDQU32Zrm     },
6157 };
6158 
6159 static const uint16_t ReplaceableInstrsAVX512DQ[][4] = {
6160   // Two integer columns for 64-bit and 32-bit elements.
6161   //PackedSingle        PackedDouble        PackedInt           PackedInt
6162   { X86::VANDNPSZ128rm, X86::VANDNPDZ128rm, X86::VPANDNQZ128rm, X86::VPANDNDZ128rm },
6163   { X86::VANDNPSZ128rr, X86::VANDNPDZ128rr, X86::VPANDNQZ128rr, X86::VPANDNDZ128rr },
6164   { X86::VANDPSZ128rm,  X86::VANDPDZ128rm,  X86::VPANDQZ128rm,  X86::VPANDDZ128rm  },
6165   { X86::VANDPSZ128rr,  X86::VANDPDZ128rr,  X86::VPANDQZ128rr,  X86::VPANDDZ128rr  },
6166   { X86::VORPSZ128rm,   X86::VORPDZ128rm,   X86::VPORQZ128rm,   X86::VPORDZ128rm   },
6167   { X86::VORPSZ128rr,   X86::VORPDZ128rr,   X86::VPORQZ128rr,   X86::VPORDZ128rr   },
6168   { X86::VXORPSZ128rm,  X86::VXORPDZ128rm,  X86::VPXORQZ128rm,  X86::VPXORDZ128rm  },
6169   { X86::VXORPSZ128rr,  X86::VXORPDZ128rr,  X86::VPXORQZ128rr,  X86::VPXORDZ128rr  },
6170   { X86::VANDNPSZ256rm, X86::VANDNPDZ256rm, X86::VPANDNQZ256rm, X86::VPANDNDZ256rm },
6171   { X86::VANDNPSZ256rr, X86::VANDNPDZ256rr, X86::VPANDNQZ256rr, X86::VPANDNDZ256rr },
6172   { X86::VANDPSZ256rm,  X86::VANDPDZ256rm,  X86::VPANDQZ256rm,  X86::VPANDDZ256rm  },
6173   { X86::VANDPSZ256rr,  X86::VANDPDZ256rr,  X86::VPANDQZ256rr,  X86::VPANDDZ256rr  },
6174   { X86::VORPSZ256rm,   X86::VORPDZ256rm,   X86::VPORQZ256rm,   X86::VPORDZ256rm   },
6175   { X86::VORPSZ256rr,   X86::VORPDZ256rr,   X86::VPORQZ256rr,   X86::VPORDZ256rr   },
6176   { X86::VXORPSZ256rm,  X86::VXORPDZ256rm,  X86::VPXORQZ256rm,  X86::VPXORDZ256rm  },
6177   { X86::VXORPSZ256rr,  X86::VXORPDZ256rr,  X86::VPXORQZ256rr,  X86::VPXORDZ256rr  },
6178   { X86::VANDNPSZrm,    X86::VANDNPDZrm,    X86::VPANDNQZrm,    X86::VPANDNDZrm    },
6179   { X86::VANDNPSZrr,    X86::VANDNPDZrr,    X86::VPANDNQZrr,    X86::VPANDNDZrr    },
6180   { X86::VANDPSZrm,     X86::VANDPDZrm,     X86::VPANDQZrm,     X86::VPANDDZrm     },
6181   { X86::VANDPSZrr,     X86::VANDPDZrr,     X86::VPANDQZrr,     X86::VPANDDZrr     },
6182   { X86::VORPSZrm,      X86::VORPDZrm,      X86::VPORQZrm,      X86::VPORDZrm      },
6183   { X86::VORPSZrr,      X86::VORPDZrr,      X86::VPORQZrr,      X86::VPORDZrr      },
6184   { X86::VXORPSZrm,     X86::VXORPDZrm,     X86::VPXORQZrm,     X86::VPXORDZrm     },
6185   { X86::VXORPSZrr,     X86::VXORPDZrr,     X86::VPXORQZrr,     X86::VPXORDZrr     },
6186 };
6187 
6188 static const uint16_t ReplaceableInstrsAVX512DQMasked[][4] = {
6189   // Two integer columns for 64-bit and 32-bit elements.
6190   //PackedSingle          PackedDouble
6191   //PackedInt             PackedInt
6192   { X86::VANDNPSZ128rmk,  X86::VANDNPDZ128rmk,
6193     X86::VPANDNQZ128rmk,  X86::VPANDNDZ128rmk  },
6194   { X86::VANDNPSZ128rmkz, X86::VANDNPDZ128rmkz,
6195     X86::VPANDNQZ128rmkz, X86::VPANDNDZ128rmkz },
6196   { X86::VANDNPSZ128rrk,  X86::VANDNPDZ128rrk,
6197     X86::VPANDNQZ128rrk,  X86::VPANDNDZ128rrk  },
6198   { X86::VANDNPSZ128rrkz, X86::VANDNPDZ128rrkz,
6199     X86::VPANDNQZ128rrkz, X86::VPANDNDZ128rrkz },
6200   { X86::VANDPSZ128rmk,   X86::VANDPDZ128rmk,
6201     X86::VPANDQZ128rmk,   X86::VPANDDZ128rmk   },
6202   { X86::VANDPSZ128rmkz,  X86::VANDPDZ128rmkz,
6203     X86::VPANDQZ128rmkz,  X86::VPANDDZ128rmkz  },
6204   { X86::VANDPSZ128rrk,   X86::VANDPDZ128rrk,
6205     X86::VPANDQZ128rrk,   X86::VPANDDZ128rrk   },
6206   { X86::VANDPSZ128rrkz,  X86::VANDPDZ128rrkz,
6207     X86::VPANDQZ128rrkz,  X86::VPANDDZ128rrkz  },
6208   { X86::VORPSZ128rmk,    X86::VORPDZ128rmk,
6209     X86::VPORQZ128rmk,    X86::VPORDZ128rmk    },
6210   { X86::VORPSZ128rmkz,   X86::VORPDZ128rmkz,
6211     X86::VPORQZ128rmkz,   X86::VPORDZ128rmkz   },
6212   { X86::VORPSZ128rrk,    X86::VORPDZ128rrk,
6213     X86::VPORQZ128rrk,    X86::VPORDZ128rrk    },
6214   { X86::VORPSZ128rrkz,   X86::VORPDZ128rrkz,
6215     X86::VPORQZ128rrkz,   X86::VPORDZ128rrkz   },
6216   { X86::VXORPSZ128rmk,   X86::VXORPDZ128rmk,
6217     X86::VPXORQZ128rmk,   X86::VPXORDZ128rmk   },
6218   { X86::VXORPSZ128rmkz,  X86::VXORPDZ128rmkz,
6219     X86::VPXORQZ128rmkz,  X86::VPXORDZ128rmkz  },
6220   { X86::VXORPSZ128rrk,   X86::VXORPDZ128rrk,
6221     X86::VPXORQZ128rrk,   X86::VPXORDZ128rrk   },
6222   { X86::VXORPSZ128rrkz,  X86::VXORPDZ128rrkz,
6223     X86::VPXORQZ128rrkz,  X86::VPXORDZ128rrkz  },
6224   { X86::VANDNPSZ256rmk,  X86::VANDNPDZ256rmk,
6225     X86::VPANDNQZ256rmk,  X86::VPANDNDZ256rmk  },
6226   { X86::VANDNPSZ256rmkz, X86::VANDNPDZ256rmkz,
6227     X86::VPANDNQZ256rmkz, X86::VPANDNDZ256rmkz },
6228   { X86::VANDNPSZ256rrk,  X86::VANDNPDZ256rrk,
6229     X86::VPANDNQZ256rrk,  X86::VPANDNDZ256rrk  },
6230   { X86::VANDNPSZ256rrkz, X86::VANDNPDZ256rrkz,
6231     X86::VPANDNQZ256rrkz, X86::VPANDNDZ256rrkz },
6232   { X86::VANDPSZ256rmk,   X86::VANDPDZ256rmk,
6233     X86::VPANDQZ256rmk,   X86::VPANDDZ256rmk   },
6234   { X86::VANDPSZ256rmkz,  X86::VANDPDZ256rmkz,
6235     X86::VPANDQZ256rmkz,  X86::VPANDDZ256rmkz  },
6236   { X86::VANDPSZ256rrk,   X86::VANDPDZ256rrk,
6237     X86::VPANDQZ256rrk,   X86::VPANDDZ256rrk   },
6238   { X86::VANDPSZ256rrkz,  X86::VANDPDZ256rrkz,
6239     X86::VPANDQZ256rrkz,  X86::VPANDDZ256rrkz  },
6240   { X86::VORPSZ256rmk,    X86::VORPDZ256rmk,
6241     X86::VPORQZ256rmk,    X86::VPORDZ256rmk    },
6242   { X86::VORPSZ256rmkz,   X86::VORPDZ256rmkz,
6243     X86::VPORQZ256rmkz,   X86::VPORDZ256rmkz   },
6244   { X86::VORPSZ256rrk,    X86::VORPDZ256rrk,
6245     X86::VPORQZ256rrk,    X86::VPORDZ256rrk    },
6246   { X86::VORPSZ256rrkz,   X86::VORPDZ256rrkz,
6247     X86::VPORQZ256rrkz,   X86::VPORDZ256rrkz   },
6248   { X86::VXORPSZ256rmk,   X86::VXORPDZ256rmk,
6249     X86::VPXORQZ256rmk,   X86::VPXORDZ256rmk   },
6250   { X86::VXORPSZ256rmkz,  X86::VXORPDZ256rmkz,
6251     X86::VPXORQZ256rmkz,  X86::VPXORDZ256rmkz  },
6252   { X86::VXORPSZ256rrk,   X86::VXORPDZ256rrk,
6253     X86::VPXORQZ256rrk,   X86::VPXORDZ256rrk   },
6254   { X86::VXORPSZ256rrkz,  X86::VXORPDZ256rrkz,
6255     X86::VPXORQZ256rrkz,  X86::VPXORDZ256rrkz  },
6256   { X86::VANDNPSZrmk,     X86::VANDNPDZrmk,
6257     X86::VPANDNQZrmk,     X86::VPANDNDZrmk     },
6258   { X86::VANDNPSZrmkz,    X86::VANDNPDZrmkz,
6259     X86::VPANDNQZrmkz,    X86::VPANDNDZrmkz    },
6260   { X86::VANDNPSZrrk,     X86::VANDNPDZrrk,
6261     X86::VPANDNQZrrk,     X86::VPANDNDZrrk     },
6262   { X86::VANDNPSZrrkz,    X86::VANDNPDZrrkz,
6263     X86::VPANDNQZrrkz,    X86::VPANDNDZrrkz    },
6264   { X86::VANDPSZrmk,      X86::VANDPDZrmk,
6265     X86::VPANDQZrmk,      X86::VPANDDZrmk      },
6266   { X86::VANDPSZrmkz,     X86::VANDPDZrmkz,
6267     X86::VPANDQZrmkz,     X86::VPANDDZrmkz     },
6268   { X86::VANDPSZrrk,      X86::VANDPDZrrk,
6269     X86::VPANDQZrrk,      X86::VPANDDZrrk      },
6270   { X86::VANDPSZrrkz,     X86::VANDPDZrrkz,
6271     X86::VPANDQZrrkz,     X86::VPANDDZrrkz     },
6272   { X86::VORPSZrmk,       X86::VORPDZrmk,
6273     X86::VPORQZrmk,       X86::VPORDZrmk       },
6274   { X86::VORPSZrmkz,      X86::VORPDZrmkz,
6275     X86::VPORQZrmkz,      X86::VPORDZrmkz      },
6276   { X86::VORPSZrrk,       X86::VORPDZrrk,
6277     X86::VPORQZrrk,       X86::VPORDZrrk       },
6278   { X86::VORPSZrrkz,      X86::VORPDZrrkz,
6279     X86::VPORQZrrkz,      X86::VPORDZrrkz      },
6280   { X86::VXORPSZrmk,      X86::VXORPDZrmk,
6281     X86::VPXORQZrmk,      X86::VPXORDZrmk      },
6282   { X86::VXORPSZrmkz,     X86::VXORPDZrmkz,
6283     X86::VPXORQZrmkz,     X86::VPXORDZrmkz     },
6284   { X86::VXORPSZrrk,      X86::VXORPDZrrk,
6285     X86::VPXORQZrrk,      X86::VPXORDZrrk      },
6286   { X86::VXORPSZrrkz,     X86::VXORPDZrrkz,
6287     X86::VPXORQZrrkz,     X86::VPXORDZrrkz     },
6288   // Broadcast loads can be handled the same as masked operations to avoid
6289   // changing element size.
6290   { X86::VANDNPSZ128rmb,  X86::VANDNPDZ128rmb,
6291     X86::VPANDNQZ128rmb,  X86::VPANDNDZ128rmb  },
6292   { X86::VANDPSZ128rmb,   X86::VANDPDZ128rmb,
6293     X86::VPANDQZ128rmb,   X86::VPANDDZ128rmb   },
6294   { X86::VORPSZ128rmb,    X86::VORPDZ128rmb,
6295     X86::VPORQZ128rmb,    X86::VPORDZ128rmb    },
6296   { X86::VXORPSZ128rmb,   X86::VXORPDZ128rmb,
6297     X86::VPXORQZ128rmb,   X86::VPXORDZ128rmb   },
6298   { X86::VANDNPSZ256rmb,  X86::VANDNPDZ256rmb,
6299     X86::VPANDNQZ256rmb,  X86::VPANDNDZ256rmb  },
6300   { X86::VANDPSZ256rmb,   X86::VANDPDZ256rmb,
6301     X86::VPANDQZ256rmb,   X86::VPANDDZ256rmb   },
6302   { X86::VORPSZ256rmb,    X86::VORPDZ256rmb,
6303     X86::VPORQZ256rmb,    X86::VPORDZ256rmb    },
6304   { X86::VXORPSZ256rmb,   X86::VXORPDZ256rmb,
6305     X86::VPXORQZ256rmb,   X86::VPXORDZ256rmb   },
6306   { X86::VANDNPSZrmb,     X86::VANDNPDZrmb,
6307     X86::VPANDNQZrmb,     X86::VPANDNDZrmb     },
6308   { X86::VANDPSZrmb,      X86::VANDPDZrmb,
6309     X86::VPANDQZrmb,      X86::VPANDDZrmb      },
6310   { X86::VANDPSZrmb,      X86::VANDPDZrmb,
6311     X86::VPANDQZrmb,      X86::VPANDDZrmb      },
6312   { X86::VORPSZrmb,       X86::VORPDZrmb,
6313     X86::VPORQZrmb,       X86::VPORDZrmb       },
6314   { X86::VXORPSZrmb,      X86::VXORPDZrmb,
6315     X86::VPXORQZrmb,      X86::VPXORDZrmb      },
6316   { X86::VANDNPSZ128rmbk, X86::VANDNPDZ128rmbk,
6317     X86::VPANDNQZ128rmbk, X86::VPANDNDZ128rmbk },
6318   { X86::VANDPSZ128rmbk,  X86::VANDPDZ128rmbk,
6319     X86::VPANDQZ128rmbk,  X86::VPANDDZ128rmbk  },
6320   { X86::VORPSZ128rmbk,   X86::VORPDZ128rmbk,
6321     X86::VPORQZ128rmbk,   X86::VPORDZ128rmbk   },
6322   { X86::VXORPSZ128rmbk,  X86::VXORPDZ128rmbk,
6323     X86::VPXORQZ128rmbk,  X86::VPXORDZ128rmbk  },
6324   { X86::VANDNPSZ256rmbk, X86::VANDNPDZ256rmbk,
6325     X86::VPANDNQZ256rmbk, X86::VPANDNDZ256rmbk },
6326   { X86::VANDPSZ256rmbk,  X86::VANDPDZ256rmbk,
6327     X86::VPANDQZ256rmbk,  X86::VPANDDZ256rmbk  },
6328   { X86::VORPSZ256rmbk,   X86::VORPDZ256rmbk,
6329     X86::VPORQZ256rmbk,   X86::VPORDZ256rmbk   },
6330   { X86::VXORPSZ256rmbk,  X86::VXORPDZ256rmbk,
6331     X86::VPXORQZ256rmbk,  X86::VPXORDZ256rmbk  },
6332   { X86::VANDNPSZrmbk,    X86::VANDNPDZrmbk,
6333     X86::VPANDNQZrmbk,    X86::VPANDNDZrmbk    },
6334   { X86::VANDPSZrmbk,     X86::VANDPDZrmbk,
6335     X86::VPANDQZrmbk,     X86::VPANDDZrmbk     },
6336   { X86::VANDPSZrmbk,     X86::VANDPDZrmbk,
6337     X86::VPANDQZrmbk,     X86::VPANDDZrmbk     },
6338   { X86::VORPSZrmbk,      X86::VORPDZrmbk,
6339     X86::VPORQZrmbk,      X86::VPORDZrmbk      },
6340   { X86::VXORPSZrmbk,     X86::VXORPDZrmbk,
6341     X86::VPXORQZrmbk,     X86::VPXORDZrmbk     },
6342   { X86::VANDNPSZ128rmbkz,X86::VANDNPDZ128rmbkz,
6343     X86::VPANDNQZ128rmbkz,X86::VPANDNDZ128rmbkz},
6344   { X86::VANDPSZ128rmbkz, X86::VANDPDZ128rmbkz,
6345     X86::VPANDQZ128rmbkz, X86::VPANDDZ128rmbkz },
6346   { X86::VORPSZ128rmbkz,  X86::VORPDZ128rmbkz,
6347     X86::VPORQZ128rmbkz,  X86::VPORDZ128rmbkz  },
6348   { X86::VXORPSZ128rmbkz, X86::VXORPDZ128rmbkz,
6349     X86::VPXORQZ128rmbkz, X86::VPXORDZ128rmbkz },
6350   { X86::VANDNPSZ256rmbkz,X86::VANDNPDZ256rmbkz,
6351     X86::VPANDNQZ256rmbkz,X86::VPANDNDZ256rmbkz},
6352   { X86::VANDPSZ256rmbkz, X86::VANDPDZ256rmbkz,
6353     X86::VPANDQZ256rmbkz, X86::VPANDDZ256rmbkz },
6354   { X86::VORPSZ256rmbkz,  X86::VORPDZ256rmbkz,
6355     X86::VPORQZ256rmbkz,  X86::VPORDZ256rmbkz  },
6356   { X86::VXORPSZ256rmbkz, X86::VXORPDZ256rmbkz,
6357     X86::VPXORQZ256rmbkz, X86::VPXORDZ256rmbkz },
6358   { X86::VANDNPSZrmbkz,   X86::VANDNPDZrmbkz,
6359     X86::VPANDNQZrmbkz,   X86::VPANDNDZrmbkz   },
6360   { X86::VANDPSZrmbkz,    X86::VANDPDZrmbkz,
6361     X86::VPANDQZrmbkz,    X86::VPANDDZrmbkz    },
6362   { X86::VANDPSZrmbkz,    X86::VANDPDZrmbkz,
6363     X86::VPANDQZrmbkz,    X86::VPANDDZrmbkz    },
6364   { X86::VORPSZrmbkz,     X86::VORPDZrmbkz,
6365     X86::VPORQZrmbkz,     X86::VPORDZrmbkz     },
6366   { X86::VXORPSZrmbkz,    X86::VXORPDZrmbkz,
6367     X86::VPXORQZrmbkz,    X86::VPXORDZrmbkz    },
6368 };
6369 
6370 // NOTE: These should only be used by the custom domain methods.
6371 static const uint16_t ReplaceableCustomInstrs[][3] = {
6372   //PackedSingle             PackedDouble             PackedInt
6373   { X86::BLENDPSrmi,         X86::BLENDPDrmi,         X86::PBLENDWrmi   },
6374   { X86::BLENDPSrri,         X86::BLENDPDrri,         X86::PBLENDWrri   },
6375   { X86::VBLENDPSrmi,        X86::VBLENDPDrmi,        X86::VPBLENDWrmi  },
6376   { X86::VBLENDPSrri,        X86::VBLENDPDrri,        X86::VPBLENDWrri  },
6377   { X86::VBLENDPSYrmi,       X86::VBLENDPDYrmi,       X86::VPBLENDWYrmi },
6378   { X86::VBLENDPSYrri,       X86::VBLENDPDYrri,       X86::VPBLENDWYrri },
6379 };
6380 static const uint16_t ReplaceableCustomAVX2Instrs[][3] = {
6381   //PackedSingle             PackedDouble             PackedInt
6382   { X86::VBLENDPSrmi,        X86::VBLENDPDrmi,        X86::VPBLENDDrmi  },
6383   { X86::VBLENDPSrri,        X86::VBLENDPDrri,        X86::VPBLENDDrri  },
6384   { X86::VBLENDPSYrmi,       X86::VBLENDPDYrmi,       X86::VPBLENDDYrmi },
6385   { X86::VBLENDPSYrri,       X86::VBLENDPDYrri,       X86::VPBLENDDYrri },
6386 };
6387 
6388 // Special table for changing EVEX logic instructions to VEX.
6389 // TODO: Should we run EVEX->VEX earlier?
6390 static const uint16_t ReplaceableCustomAVX512LogicInstrs[][4] = {
6391   // Two integer columns for 64-bit and 32-bit elements.
6392   //PackedSingle     PackedDouble     PackedInt           PackedInt
6393   { X86::VANDNPSrm,  X86::VANDNPDrm,  X86::VPANDNQZ128rm, X86::VPANDNDZ128rm },
6394   { X86::VANDNPSrr,  X86::VANDNPDrr,  X86::VPANDNQZ128rr, X86::VPANDNDZ128rr },
6395   { X86::VANDPSrm,   X86::VANDPDrm,   X86::VPANDQZ128rm,  X86::VPANDDZ128rm  },
6396   { X86::VANDPSrr,   X86::VANDPDrr,   X86::VPANDQZ128rr,  X86::VPANDDZ128rr  },
6397   { X86::VORPSrm,    X86::VORPDrm,    X86::VPORQZ128rm,   X86::VPORDZ128rm   },
6398   { X86::VORPSrr,    X86::VORPDrr,    X86::VPORQZ128rr,   X86::VPORDZ128rr   },
6399   { X86::VXORPSrm,   X86::VXORPDrm,   X86::VPXORQZ128rm,  X86::VPXORDZ128rm  },
6400   { X86::VXORPSrr,   X86::VXORPDrr,   X86::VPXORQZ128rr,  X86::VPXORDZ128rr  },
6401   { X86::VANDNPSYrm, X86::VANDNPDYrm, X86::VPANDNQZ256rm, X86::VPANDNDZ256rm },
6402   { X86::VANDNPSYrr, X86::VANDNPDYrr, X86::VPANDNQZ256rr, X86::VPANDNDZ256rr },
6403   { X86::VANDPSYrm,  X86::VANDPDYrm,  X86::VPANDQZ256rm,  X86::VPANDDZ256rm  },
6404   { X86::VANDPSYrr,  X86::VANDPDYrr,  X86::VPANDQZ256rr,  X86::VPANDDZ256rr  },
6405   { X86::VORPSYrm,   X86::VORPDYrm,   X86::VPORQZ256rm,   X86::VPORDZ256rm   },
6406   { X86::VORPSYrr,   X86::VORPDYrr,   X86::VPORQZ256rr,   X86::VPORDZ256rr   },
6407   { X86::VXORPSYrm,  X86::VXORPDYrm,  X86::VPXORQZ256rm,  X86::VPXORDZ256rm  },
6408   { X86::VXORPSYrr,  X86::VXORPDYrr,  X86::VPXORQZ256rr,  X86::VPXORDZ256rr  },
6409 };
6410 
6411 // FIXME: Some shuffle and unpack instructions have equivalents in different
6412 // domains, but they require a bit more work than just switching opcodes.
6413 
lookup(unsigned opcode,unsigned domain,ArrayRef<uint16_t[3]> Table)6414 static const uint16_t *lookup(unsigned opcode, unsigned domain,
6415                               ArrayRef<uint16_t[3]> Table) {
6416   for (const uint16_t (&Row)[3] : Table)
6417     if (Row[domain-1] == opcode)
6418       return Row;
6419   return nullptr;
6420 }
6421 
lookupAVX512(unsigned opcode,unsigned domain,ArrayRef<uint16_t[4]> Table)6422 static const uint16_t *lookupAVX512(unsigned opcode, unsigned domain,
6423                                     ArrayRef<uint16_t[4]> Table) {
6424   // If this is the integer domain make sure to check both integer columns.
6425   for (const uint16_t (&Row)[4] : Table)
6426     if (Row[domain-1] == opcode || (domain == 3 && Row[3] == opcode))
6427       return Row;
6428   return nullptr;
6429 }
6430 
6431 // Helper to attempt to widen/narrow blend masks.
AdjustBlendMask(unsigned OldMask,unsigned OldWidth,unsigned NewWidth,unsigned * pNewMask=nullptr)6432 static bool AdjustBlendMask(unsigned OldMask, unsigned OldWidth,
6433                             unsigned NewWidth, unsigned *pNewMask = nullptr) {
6434   assert(((OldWidth % NewWidth) == 0 || (NewWidth % OldWidth) == 0) &&
6435          "Illegal blend mask scale");
6436   unsigned NewMask = 0;
6437 
6438   if ((OldWidth % NewWidth) == 0) {
6439     unsigned Scale = OldWidth / NewWidth;
6440     unsigned SubMask = (1u << Scale) - 1;
6441     for (unsigned i = 0; i != NewWidth; ++i) {
6442       unsigned Sub = (OldMask >> (i * Scale)) & SubMask;
6443       if (Sub == SubMask)
6444         NewMask |= (1u << i);
6445       else if (Sub != 0x0)
6446         return false;
6447     }
6448   } else {
6449     unsigned Scale = NewWidth / OldWidth;
6450     unsigned SubMask = (1u << Scale) - 1;
6451     for (unsigned i = 0; i != OldWidth; ++i) {
6452       if (OldMask & (1 << i)) {
6453         NewMask |= (SubMask << (i * Scale));
6454       }
6455     }
6456   }
6457 
6458   if (pNewMask)
6459     *pNewMask = NewMask;
6460   return true;
6461 }
6462 
getExecutionDomainCustom(const MachineInstr & MI) const6463 uint16_t X86InstrInfo::getExecutionDomainCustom(const MachineInstr &MI) const {
6464   unsigned Opcode = MI.getOpcode();
6465   unsigned NumOperands = MI.getDesc().getNumOperands();
6466 
6467   auto GetBlendDomains = [&](unsigned ImmWidth, bool Is256) {
6468     uint16_t validDomains = 0;
6469     if (MI.getOperand(NumOperands - 1).isImm()) {
6470       unsigned Imm = MI.getOperand(NumOperands - 1).getImm();
6471       if (AdjustBlendMask(Imm, ImmWidth, Is256 ? 8 : 4))
6472         validDomains |= 0x2; // PackedSingle
6473       if (AdjustBlendMask(Imm, ImmWidth, Is256 ? 4 : 2))
6474         validDomains |= 0x4; // PackedDouble
6475       if (!Is256 || Subtarget.hasAVX2())
6476         validDomains |= 0x8; // PackedInt
6477     }
6478     return validDomains;
6479   };
6480 
6481   switch (Opcode) {
6482   case X86::BLENDPDrmi:
6483   case X86::BLENDPDrri:
6484   case X86::VBLENDPDrmi:
6485   case X86::VBLENDPDrri:
6486     return GetBlendDomains(2, false);
6487   case X86::VBLENDPDYrmi:
6488   case X86::VBLENDPDYrri:
6489     return GetBlendDomains(4, true);
6490   case X86::BLENDPSrmi:
6491   case X86::BLENDPSrri:
6492   case X86::VBLENDPSrmi:
6493   case X86::VBLENDPSrri:
6494   case X86::VPBLENDDrmi:
6495   case X86::VPBLENDDrri:
6496     return GetBlendDomains(4, false);
6497   case X86::VBLENDPSYrmi:
6498   case X86::VBLENDPSYrri:
6499   case X86::VPBLENDDYrmi:
6500   case X86::VPBLENDDYrri:
6501     return GetBlendDomains(8, true);
6502   case X86::PBLENDWrmi:
6503   case X86::PBLENDWrri:
6504   case X86::VPBLENDWrmi:
6505   case X86::VPBLENDWrri:
6506   // Treat VPBLENDWY as a 128-bit vector as it repeats the lo/hi masks.
6507   case X86::VPBLENDWYrmi:
6508   case X86::VPBLENDWYrri:
6509     return GetBlendDomains(8, false);
6510   case X86::VPANDDZ128rr:  case X86::VPANDDZ128rm:
6511   case X86::VPANDDZ256rr:  case X86::VPANDDZ256rm:
6512   case X86::VPANDQZ128rr:  case X86::VPANDQZ128rm:
6513   case X86::VPANDQZ256rr:  case X86::VPANDQZ256rm:
6514   case X86::VPANDNDZ128rr: case X86::VPANDNDZ128rm:
6515   case X86::VPANDNDZ256rr: case X86::VPANDNDZ256rm:
6516   case X86::VPANDNQZ128rr: case X86::VPANDNQZ128rm:
6517   case X86::VPANDNQZ256rr: case X86::VPANDNQZ256rm:
6518   case X86::VPORDZ128rr:   case X86::VPORDZ128rm:
6519   case X86::VPORDZ256rr:   case X86::VPORDZ256rm:
6520   case X86::VPORQZ128rr:   case X86::VPORQZ128rm:
6521   case X86::VPORQZ256rr:   case X86::VPORQZ256rm:
6522   case X86::VPXORDZ128rr:  case X86::VPXORDZ128rm:
6523   case X86::VPXORDZ256rr:  case X86::VPXORDZ256rm:
6524   case X86::VPXORQZ128rr:  case X86::VPXORQZ128rm:
6525   case X86::VPXORQZ256rr:  case X86::VPXORQZ256rm:
6526     // If we don't have DQI see if we can still switch from an EVEX integer
6527     // instruction to a VEX floating point instruction.
6528     if (Subtarget.hasDQI())
6529       return 0;
6530 
6531     if (RI.getEncodingValue(MI.getOperand(0).getReg()) >= 16)
6532       return 0;
6533     if (RI.getEncodingValue(MI.getOperand(1).getReg()) >= 16)
6534       return 0;
6535     // Register forms will have 3 operands. Memory form will have more.
6536     if (NumOperands == 3 &&
6537         RI.getEncodingValue(MI.getOperand(2).getReg()) >= 16)
6538       return 0;
6539 
6540     // All domains are valid.
6541     return 0xe;
6542   case X86::MOVHLPSrr:
6543     // We can swap domains when both inputs are the same register.
6544     // FIXME: This doesn't catch all the cases we would like. If the input
6545     // register isn't KILLed by the instruction, the two address instruction
6546     // pass puts a COPY on one input. The other input uses the original
6547     // register. This prevents the same physical register from being used by
6548     // both inputs.
6549     if (MI.getOperand(1).getReg() == MI.getOperand(2).getReg() &&
6550         MI.getOperand(0).getSubReg() == 0 &&
6551         MI.getOperand(1).getSubReg() == 0 &&
6552         MI.getOperand(2).getSubReg() == 0)
6553       return 0x6;
6554     return 0;
6555   }
6556   return 0;
6557 }
6558 
setExecutionDomainCustom(MachineInstr & MI,unsigned Domain) const6559 bool X86InstrInfo::setExecutionDomainCustom(MachineInstr &MI,
6560                                             unsigned Domain) const {
6561   assert(Domain > 0 && Domain < 4 && "Invalid execution domain");
6562   uint16_t dom = (MI.getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
6563   assert(dom && "Not an SSE instruction");
6564 
6565   unsigned Opcode = MI.getOpcode();
6566   unsigned NumOperands = MI.getDesc().getNumOperands();
6567 
6568   auto SetBlendDomain = [&](unsigned ImmWidth, bool Is256) {
6569     if (MI.getOperand(NumOperands - 1).isImm()) {
6570       unsigned Imm = MI.getOperand(NumOperands - 1).getImm() & 255;
6571       Imm = (ImmWidth == 16 ? ((Imm << 8) | Imm) : Imm);
6572       unsigned NewImm = Imm;
6573 
6574       const uint16_t *table = lookup(Opcode, dom, ReplaceableCustomInstrs);
6575       if (!table)
6576         table = lookup(Opcode, dom, ReplaceableCustomAVX2Instrs);
6577 
6578       if (Domain == 1) { // PackedSingle
6579         AdjustBlendMask(Imm, ImmWidth, Is256 ? 8 : 4, &NewImm);
6580       } else if (Domain == 2) { // PackedDouble
6581         AdjustBlendMask(Imm, ImmWidth, Is256 ? 4 : 2, &NewImm);
6582       } else if (Domain == 3) { // PackedInt
6583         if (Subtarget.hasAVX2()) {
6584           // If we are already VPBLENDW use that, else use VPBLENDD.
6585           if ((ImmWidth / (Is256 ? 2 : 1)) != 8) {
6586             table = lookup(Opcode, dom, ReplaceableCustomAVX2Instrs);
6587             AdjustBlendMask(Imm, ImmWidth, Is256 ? 8 : 4, &NewImm);
6588           }
6589         } else {
6590           assert(!Is256 && "128-bit vector expected");
6591           AdjustBlendMask(Imm, ImmWidth, 8, &NewImm);
6592         }
6593       }
6594 
6595       assert(table && table[Domain - 1] && "Unknown domain op");
6596       MI.setDesc(get(table[Domain - 1]));
6597       MI.getOperand(NumOperands - 1).setImm(NewImm & 255);
6598     }
6599     return true;
6600   };
6601 
6602   switch (Opcode) {
6603   case X86::BLENDPDrmi:
6604   case X86::BLENDPDrri:
6605   case X86::VBLENDPDrmi:
6606   case X86::VBLENDPDrri:
6607     return SetBlendDomain(2, false);
6608   case X86::VBLENDPDYrmi:
6609   case X86::VBLENDPDYrri:
6610     return SetBlendDomain(4, true);
6611   case X86::BLENDPSrmi:
6612   case X86::BLENDPSrri:
6613   case X86::VBLENDPSrmi:
6614   case X86::VBLENDPSrri:
6615   case X86::VPBLENDDrmi:
6616   case X86::VPBLENDDrri:
6617     return SetBlendDomain(4, false);
6618   case X86::VBLENDPSYrmi:
6619   case X86::VBLENDPSYrri:
6620   case X86::VPBLENDDYrmi:
6621   case X86::VPBLENDDYrri:
6622     return SetBlendDomain(8, true);
6623   case X86::PBLENDWrmi:
6624   case X86::PBLENDWrri:
6625   case X86::VPBLENDWrmi:
6626   case X86::VPBLENDWrri:
6627     return SetBlendDomain(8, false);
6628   case X86::VPBLENDWYrmi:
6629   case X86::VPBLENDWYrri:
6630     return SetBlendDomain(16, true);
6631   case X86::VPANDDZ128rr:  case X86::VPANDDZ128rm:
6632   case X86::VPANDDZ256rr:  case X86::VPANDDZ256rm:
6633   case X86::VPANDQZ128rr:  case X86::VPANDQZ128rm:
6634   case X86::VPANDQZ256rr:  case X86::VPANDQZ256rm:
6635   case X86::VPANDNDZ128rr: case X86::VPANDNDZ128rm:
6636   case X86::VPANDNDZ256rr: case X86::VPANDNDZ256rm:
6637   case X86::VPANDNQZ128rr: case X86::VPANDNQZ128rm:
6638   case X86::VPANDNQZ256rr: case X86::VPANDNQZ256rm:
6639   case X86::VPORDZ128rr:   case X86::VPORDZ128rm:
6640   case X86::VPORDZ256rr:   case X86::VPORDZ256rm:
6641   case X86::VPORQZ128rr:   case X86::VPORQZ128rm:
6642   case X86::VPORQZ256rr:   case X86::VPORQZ256rm:
6643   case X86::VPXORDZ128rr:  case X86::VPXORDZ128rm:
6644   case X86::VPXORDZ256rr:  case X86::VPXORDZ256rm:
6645   case X86::VPXORQZ128rr:  case X86::VPXORQZ128rm:
6646   case X86::VPXORQZ256rr:  case X86::VPXORQZ256rm: {
6647     // Without DQI, convert EVEX instructions to VEX instructions.
6648     if (Subtarget.hasDQI())
6649       return false;
6650 
6651     const uint16_t *table = lookupAVX512(MI.getOpcode(), dom,
6652                                          ReplaceableCustomAVX512LogicInstrs);
6653     assert(table && "Instruction not found in table?");
6654     // Don't change integer Q instructions to D instructions and
6655     // use D intructions if we started with a PS instruction.
6656     if (Domain == 3 && (dom == 1 || table[3] == MI.getOpcode()))
6657       Domain = 4;
6658     MI.setDesc(get(table[Domain - 1]));
6659     return true;
6660   }
6661   case X86::UNPCKHPDrr:
6662   case X86::MOVHLPSrr:
6663     // We just need to commute the instruction which will switch the domains.
6664     if (Domain != dom && Domain != 3 &&
6665         MI.getOperand(1).getReg() == MI.getOperand(2).getReg() &&
6666         MI.getOperand(0).getSubReg() == 0 &&
6667         MI.getOperand(1).getSubReg() == 0 &&
6668         MI.getOperand(2).getSubReg() == 0) {
6669       commuteInstruction(MI, false);
6670       return true;
6671     }
6672     // We must always return true for MOVHLPSrr.
6673     if (Opcode == X86::MOVHLPSrr)
6674       return true;
6675   }
6676   return false;
6677 }
6678 
6679 std::pair<uint16_t, uint16_t>
getExecutionDomain(const MachineInstr & MI) const6680 X86InstrInfo::getExecutionDomain(const MachineInstr &MI) const {
6681   uint16_t domain = (MI.getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
6682   unsigned opcode = MI.getOpcode();
6683   uint16_t validDomains = 0;
6684   if (domain) {
6685     // Attempt to match for custom instructions.
6686     validDomains = getExecutionDomainCustom(MI);
6687     if (validDomains)
6688       return std::make_pair(domain, validDomains);
6689 
6690     if (lookup(opcode, domain, ReplaceableInstrs)) {
6691       validDomains = 0xe;
6692     } else if (lookup(opcode, domain, ReplaceableInstrsAVX2)) {
6693       validDomains = Subtarget.hasAVX2() ? 0xe : 0x6;
6694     } else if (lookup(opcode, domain, ReplaceableInstrsAVX2InsertExtract)) {
6695       // Insert/extract instructions should only effect domain if AVX2
6696       // is enabled.
6697       if (!Subtarget.hasAVX2())
6698         return std::make_pair(0, 0);
6699       validDomains = 0xe;
6700     } else if (lookupAVX512(opcode, domain, ReplaceableInstrsAVX512)) {
6701       validDomains = 0xe;
6702     } else if (Subtarget.hasDQI() && lookupAVX512(opcode, domain,
6703                                                   ReplaceableInstrsAVX512DQ)) {
6704       validDomains = 0xe;
6705     } else if (Subtarget.hasDQI()) {
6706       if (const uint16_t *table = lookupAVX512(opcode, domain,
6707                                              ReplaceableInstrsAVX512DQMasked)) {
6708         if (domain == 1 || (domain == 3 && table[3] == opcode))
6709           validDomains = 0xa;
6710         else
6711           validDomains = 0xc;
6712       }
6713     }
6714   }
6715   return std::make_pair(domain, validDomains);
6716 }
6717 
setExecutionDomain(MachineInstr & MI,unsigned Domain) const6718 void X86InstrInfo::setExecutionDomain(MachineInstr &MI, unsigned Domain) const {
6719   assert(Domain>0 && Domain<4 && "Invalid execution domain");
6720   uint16_t dom = (MI.getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
6721   assert(dom && "Not an SSE instruction");
6722 
6723   // Attempt to match for custom instructions.
6724   if (setExecutionDomainCustom(MI, Domain))
6725     return;
6726 
6727   const uint16_t *table = lookup(MI.getOpcode(), dom, ReplaceableInstrs);
6728   if (!table) { // try the other table
6729     assert((Subtarget.hasAVX2() || Domain < 3) &&
6730            "256-bit vector operations only available in AVX2");
6731     table = lookup(MI.getOpcode(), dom, ReplaceableInstrsAVX2);
6732   }
6733   if (!table) { // try the other table
6734     assert(Subtarget.hasAVX2() &&
6735            "256-bit insert/extract only available in AVX2");
6736     table = lookup(MI.getOpcode(), dom, ReplaceableInstrsAVX2InsertExtract);
6737   }
6738   if (!table) { // try the AVX512 table
6739     assert(Subtarget.hasAVX512() && "Requires AVX-512");
6740     table = lookupAVX512(MI.getOpcode(), dom, ReplaceableInstrsAVX512);
6741     // Don't change integer Q instructions to D instructions.
6742     if (table && Domain == 3 && table[3] == MI.getOpcode())
6743       Domain = 4;
6744   }
6745   if (!table) { // try the AVX512DQ table
6746     assert((Subtarget.hasDQI() || Domain >= 3) && "Requires AVX-512DQ");
6747     table = lookupAVX512(MI.getOpcode(), dom, ReplaceableInstrsAVX512DQ);
6748     // Don't change integer Q instructions to D instructions and
6749     // use D intructions if we started with a PS instruction.
6750     if (table && Domain == 3 && (dom == 1 || table[3] == MI.getOpcode()))
6751       Domain = 4;
6752   }
6753   if (!table) { // try the AVX512DQMasked table
6754     assert((Subtarget.hasDQI() || Domain >= 3) && "Requires AVX-512DQ");
6755     table = lookupAVX512(MI.getOpcode(), dom, ReplaceableInstrsAVX512DQMasked);
6756     if (table && Domain == 3 && (dom == 1 || table[3] == MI.getOpcode()))
6757       Domain = 4;
6758   }
6759   assert(table && "Cannot change domain");
6760   MI.setDesc(get(table[Domain - 1]));
6761 }
6762 
6763 /// Return the noop instruction to use for a noop.
getNoop(MCInst & NopInst) const6764 void X86InstrInfo::getNoop(MCInst &NopInst) const {
6765   NopInst.setOpcode(X86::NOOP);
6766 }
6767 
isHighLatencyDef(int opc) const6768 bool X86InstrInfo::isHighLatencyDef(int opc) const {
6769   switch (opc) {
6770   default: return false;
6771   case X86::DIVPDrm:
6772   case X86::DIVPDrr:
6773   case X86::DIVPSrm:
6774   case X86::DIVPSrr:
6775   case X86::DIVSDrm:
6776   case X86::DIVSDrm_Int:
6777   case X86::DIVSDrr:
6778   case X86::DIVSDrr_Int:
6779   case X86::DIVSSrm:
6780   case X86::DIVSSrm_Int:
6781   case X86::DIVSSrr:
6782   case X86::DIVSSrr_Int:
6783   case X86::SQRTPDm:
6784   case X86::SQRTPDr:
6785   case X86::SQRTPSm:
6786   case X86::SQRTPSr:
6787   case X86::SQRTSDm:
6788   case X86::SQRTSDm_Int:
6789   case X86::SQRTSDr:
6790   case X86::SQRTSDr_Int:
6791   case X86::SQRTSSm:
6792   case X86::SQRTSSm_Int:
6793   case X86::SQRTSSr:
6794   case X86::SQRTSSr_Int:
6795   // AVX instructions with high latency
6796   case X86::VDIVPDrm:
6797   case X86::VDIVPDrr:
6798   case X86::VDIVPDYrm:
6799   case X86::VDIVPDYrr:
6800   case X86::VDIVPSrm:
6801   case X86::VDIVPSrr:
6802   case X86::VDIVPSYrm:
6803   case X86::VDIVPSYrr:
6804   case X86::VDIVSDrm:
6805   case X86::VDIVSDrm_Int:
6806   case X86::VDIVSDrr:
6807   case X86::VDIVSDrr_Int:
6808   case X86::VDIVSSrm:
6809   case X86::VDIVSSrm_Int:
6810   case X86::VDIVSSrr:
6811   case X86::VDIVSSrr_Int:
6812   case X86::VSQRTPDm:
6813   case X86::VSQRTPDr:
6814   case X86::VSQRTPDYm:
6815   case X86::VSQRTPDYr:
6816   case X86::VSQRTPSm:
6817   case X86::VSQRTPSr:
6818   case X86::VSQRTPSYm:
6819   case X86::VSQRTPSYr:
6820   case X86::VSQRTSDm:
6821   case X86::VSQRTSDm_Int:
6822   case X86::VSQRTSDr:
6823   case X86::VSQRTSDr_Int:
6824   case X86::VSQRTSSm:
6825   case X86::VSQRTSSm_Int:
6826   case X86::VSQRTSSr:
6827   case X86::VSQRTSSr_Int:
6828   // AVX512 instructions with high latency
6829   case X86::VDIVPDZ128rm:
6830   case X86::VDIVPDZ128rmb:
6831   case X86::VDIVPDZ128rmbk:
6832   case X86::VDIVPDZ128rmbkz:
6833   case X86::VDIVPDZ128rmk:
6834   case X86::VDIVPDZ128rmkz:
6835   case X86::VDIVPDZ128rr:
6836   case X86::VDIVPDZ128rrk:
6837   case X86::VDIVPDZ128rrkz:
6838   case X86::VDIVPDZ256rm:
6839   case X86::VDIVPDZ256rmb:
6840   case X86::VDIVPDZ256rmbk:
6841   case X86::VDIVPDZ256rmbkz:
6842   case X86::VDIVPDZ256rmk:
6843   case X86::VDIVPDZ256rmkz:
6844   case X86::VDIVPDZ256rr:
6845   case X86::VDIVPDZ256rrk:
6846   case X86::VDIVPDZ256rrkz:
6847   case X86::VDIVPDZrrb:
6848   case X86::VDIVPDZrrbk:
6849   case X86::VDIVPDZrrbkz:
6850   case X86::VDIVPDZrm:
6851   case X86::VDIVPDZrmb:
6852   case X86::VDIVPDZrmbk:
6853   case X86::VDIVPDZrmbkz:
6854   case X86::VDIVPDZrmk:
6855   case X86::VDIVPDZrmkz:
6856   case X86::VDIVPDZrr:
6857   case X86::VDIVPDZrrk:
6858   case X86::VDIVPDZrrkz:
6859   case X86::VDIVPSZ128rm:
6860   case X86::VDIVPSZ128rmb:
6861   case X86::VDIVPSZ128rmbk:
6862   case X86::VDIVPSZ128rmbkz:
6863   case X86::VDIVPSZ128rmk:
6864   case X86::VDIVPSZ128rmkz:
6865   case X86::VDIVPSZ128rr:
6866   case X86::VDIVPSZ128rrk:
6867   case X86::VDIVPSZ128rrkz:
6868   case X86::VDIVPSZ256rm:
6869   case X86::VDIVPSZ256rmb:
6870   case X86::VDIVPSZ256rmbk:
6871   case X86::VDIVPSZ256rmbkz:
6872   case X86::VDIVPSZ256rmk:
6873   case X86::VDIVPSZ256rmkz:
6874   case X86::VDIVPSZ256rr:
6875   case X86::VDIVPSZ256rrk:
6876   case X86::VDIVPSZ256rrkz:
6877   case X86::VDIVPSZrrb:
6878   case X86::VDIVPSZrrbk:
6879   case X86::VDIVPSZrrbkz:
6880   case X86::VDIVPSZrm:
6881   case X86::VDIVPSZrmb:
6882   case X86::VDIVPSZrmbk:
6883   case X86::VDIVPSZrmbkz:
6884   case X86::VDIVPSZrmk:
6885   case X86::VDIVPSZrmkz:
6886   case X86::VDIVPSZrr:
6887   case X86::VDIVPSZrrk:
6888   case X86::VDIVPSZrrkz:
6889   case X86::VDIVSDZrm:
6890   case X86::VDIVSDZrr:
6891   case X86::VDIVSDZrm_Int:
6892   case X86::VDIVSDZrm_Intk:
6893   case X86::VDIVSDZrm_Intkz:
6894   case X86::VDIVSDZrr_Int:
6895   case X86::VDIVSDZrr_Intk:
6896   case X86::VDIVSDZrr_Intkz:
6897   case X86::VDIVSDZrrb_Int:
6898   case X86::VDIVSDZrrb_Intk:
6899   case X86::VDIVSDZrrb_Intkz:
6900   case X86::VDIVSSZrm:
6901   case X86::VDIVSSZrr:
6902   case X86::VDIVSSZrm_Int:
6903   case X86::VDIVSSZrm_Intk:
6904   case X86::VDIVSSZrm_Intkz:
6905   case X86::VDIVSSZrr_Int:
6906   case X86::VDIVSSZrr_Intk:
6907   case X86::VDIVSSZrr_Intkz:
6908   case X86::VDIVSSZrrb_Int:
6909   case X86::VDIVSSZrrb_Intk:
6910   case X86::VDIVSSZrrb_Intkz:
6911   case X86::VSQRTPDZ128m:
6912   case X86::VSQRTPDZ128mb:
6913   case X86::VSQRTPDZ128mbk:
6914   case X86::VSQRTPDZ128mbkz:
6915   case X86::VSQRTPDZ128mk:
6916   case X86::VSQRTPDZ128mkz:
6917   case X86::VSQRTPDZ128r:
6918   case X86::VSQRTPDZ128rk:
6919   case X86::VSQRTPDZ128rkz:
6920   case X86::VSQRTPDZ256m:
6921   case X86::VSQRTPDZ256mb:
6922   case X86::VSQRTPDZ256mbk:
6923   case X86::VSQRTPDZ256mbkz:
6924   case X86::VSQRTPDZ256mk:
6925   case X86::VSQRTPDZ256mkz:
6926   case X86::VSQRTPDZ256r:
6927   case X86::VSQRTPDZ256rk:
6928   case X86::VSQRTPDZ256rkz:
6929   case X86::VSQRTPDZm:
6930   case X86::VSQRTPDZmb:
6931   case X86::VSQRTPDZmbk:
6932   case X86::VSQRTPDZmbkz:
6933   case X86::VSQRTPDZmk:
6934   case X86::VSQRTPDZmkz:
6935   case X86::VSQRTPDZr:
6936   case X86::VSQRTPDZrb:
6937   case X86::VSQRTPDZrbk:
6938   case X86::VSQRTPDZrbkz:
6939   case X86::VSQRTPDZrk:
6940   case X86::VSQRTPDZrkz:
6941   case X86::VSQRTPSZ128m:
6942   case X86::VSQRTPSZ128mb:
6943   case X86::VSQRTPSZ128mbk:
6944   case X86::VSQRTPSZ128mbkz:
6945   case X86::VSQRTPSZ128mk:
6946   case X86::VSQRTPSZ128mkz:
6947   case X86::VSQRTPSZ128r:
6948   case X86::VSQRTPSZ128rk:
6949   case X86::VSQRTPSZ128rkz:
6950   case X86::VSQRTPSZ256m:
6951   case X86::VSQRTPSZ256mb:
6952   case X86::VSQRTPSZ256mbk:
6953   case X86::VSQRTPSZ256mbkz:
6954   case X86::VSQRTPSZ256mk:
6955   case X86::VSQRTPSZ256mkz:
6956   case X86::VSQRTPSZ256r:
6957   case X86::VSQRTPSZ256rk:
6958   case X86::VSQRTPSZ256rkz:
6959   case X86::VSQRTPSZm:
6960   case X86::VSQRTPSZmb:
6961   case X86::VSQRTPSZmbk:
6962   case X86::VSQRTPSZmbkz:
6963   case X86::VSQRTPSZmk:
6964   case X86::VSQRTPSZmkz:
6965   case X86::VSQRTPSZr:
6966   case X86::VSQRTPSZrb:
6967   case X86::VSQRTPSZrbk:
6968   case X86::VSQRTPSZrbkz:
6969   case X86::VSQRTPSZrk:
6970   case X86::VSQRTPSZrkz:
6971   case X86::VSQRTSDZm:
6972   case X86::VSQRTSDZm_Int:
6973   case X86::VSQRTSDZm_Intk:
6974   case X86::VSQRTSDZm_Intkz:
6975   case X86::VSQRTSDZr:
6976   case X86::VSQRTSDZr_Int:
6977   case X86::VSQRTSDZr_Intk:
6978   case X86::VSQRTSDZr_Intkz:
6979   case X86::VSQRTSDZrb_Int:
6980   case X86::VSQRTSDZrb_Intk:
6981   case X86::VSQRTSDZrb_Intkz:
6982   case X86::VSQRTSSZm:
6983   case X86::VSQRTSSZm_Int:
6984   case X86::VSQRTSSZm_Intk:
6985   case X86::VSQRTSSZm_Intkz:
6986   case X86::VSQRTSSZr:
6987   case X86::VSQRTSSZr_Int:
6988   case X86::VSQRTSSZr_Intk:
6989   case X86::VSQRTSSZr_Intkz:
6990   case X86::VSQRTSSZrb_Int:
6991   case X86::VSQRTSSZrb_Intk:
6992   case X86::VSQRTSSZrb_Intkz:
6993 
6994   case X86::VGATHERDPDYrm:
6995   case X86::VGATHERDPDZ128rm:
6996   case X86::VGATHERDPDZ256rm:
6997   case X86::VGATHERDPDZrm:
6998   case X86::VGATHERDPDrm:
6999   case X86::VGATHERDPSYrm:
7000   case X86::VGATHERDPSZ128rm:
7001   case X86::VGATHERDPSZ256rm:
7002   case X86::VGATHERDPSZrm:
7003   case X86::VGATHERDPSrm:
7004   case X86::VGATHERPF0DPDm:
7005   case X86::VGATHERPF0DPSm:
7006   case X86::VGATHERPF0QPDm:
7007   case X86::VGATHERPF0QPSm:
7008   case X86::VGATHERPF1DPDm:
7009   case X86::VGATHERPF1DPSm:
7010   case X86::VGATHERPF1QPDm:
7011   case X86::VGATHERPF1QPSm:
7012   case X86::VGATHERQPDYrm:
7013   case X86::VGATHERQPDZ128rm:
7014   case X86::VGATHERQPDZ256rm:
7015   case X86::VGATHERQPDZrm:
7016   case X86::VGATHERQPDrm:
7017   case X86::VGATHERQPSYrm:
7018   case X86::VGATHERQPSZ128rm:
7019   case X86::VGATHERQPSZ256rm:
7020   case X86::VGATHERQPSZrm:
7021   case X86::VGATHERQPSrm:
7022   case X86::VPGATHERDDYrm:
7023   case X86::VPGATHERDDZ128rm:
7024   case X86::VPGATHERDDZ256rm:
7025   case X86::VPGATHERDDZrm:
7026   case X86::VPGATHERDDrm:
7027   case X86::VPGATHERDQYrm:
7028   case X86::VPGATHERDQZ128rm:
7029   case X86::VPGATHERDQZ256rm:
7030   case X86::VPGATHERDQZrm:
7031   case X86::VPGATHERDQrm:
7032   case X86::VPGATHERQDYrm:
7033   case X86::VPGATHERQDZ128rm:
7034   case X86::VPGATHERQDZ256rm:
7035   case X86::VPGATHERQDZrm:
7036   case X86::VPGATHERQDrm:
7037   case X86::VPGATHERQQYrm:
7038   case X86::VPGATHERQQZ128rm:
7039   case X86::VPGATHERQQZ256rm:
7040   case X86::VPGATHERQQZrm:
7041   case X86::VPGATHERQQrm:
7042   case X86::VSCATTERDPDZ128mr:
7043   case X86::VSCATTERDPDZ256mr:
7044   case X86::VSCATTERDPDZmr:
7045   case X86::VSCATTERDPSZ128mr:
7046   case X86::VSCATTERDPSZ256mr:
7047   case X86::VSCATTERDPSZmr:
7048   case X86::VSCATTERPF0DPDm:
7049   case X86::VSCATTERPF0DPSm:
7050   case X86::VSCATTERPF0QPDm:
7051   case X86::VSCATTERPF0QPSm:
7052   case X86::VSCATTERPF1DPDm:
7053   case X86::VSCATTERPF1DPSm:
7054   case X86::VSCATTERPF1QPDm:
7055   case X86::VSCATTERPF1QPSm:
7056   case X86::VSCATTERQPDZ128mr:
7057   case X86::VSCATTERQPDZ256mr:
7058   case X86::VSCATTERQPDZmr:
7059   case X86::VSCATTERQPSZ128mr:
7060   case X86::VSCATTERQPSZ256mr:
7061   case X86::VSCATTERQPSZmr:
7062   case X86::VPSCATTERDDZ128mr:
7063   case X86::VPSCATTERDDZ256mr:
7064   case X86::VPSCATTERDDZmr:
7065   case X86::VPSCATTERDQZ128mr:
7066   case X86::VPSCATTERDQZ256mr:
7067   case X86::VPSCATTERDQZmr:
7068   case X86::VPSCATTERQDZ128mr:
7069   case X86::VPSCATTERQDZ256mr:
7070   case X86::VPSCATTERQDZmr:
7071   case X86::VPSCATTERQQZ128mr:
7072   case X86::VPSCATTERQQZ256mr:
7073   case X86::VPSCATTERQQZmr:
7074     return true;
7075   }
7076 }
7077 
hasHighOperandLatency(const TargetSchedModel & SchedModel,const MachineRegisterInfo * MRI,const MachineInstr & DefMI,unsigned DefIdx,const MachineInstr & UseMI,unsigned UseIdx) const7078 bool X86InstrInfo::hasHighOperandLatency(const TargetSchedModel &SchedModel,
7079                                          const MachineRegisterInfo *MRI,
7080                                          const MachineInstr &DefMI,
7081                                          unsigned DefIdx,
7082                                          const MachineInstr &UseMI,
7083                                          unsigned UseIdx) const {
7084   return isHighLatencyDef(DefMI.getOpcode());
7085 }
7086 
hasReassociableOperands(const MachineInstr & Inst,const MachineBasicBlock * MBB) const7087 bool X86InstrInfo::hasReassociableOperands(const MachineInstr &Inst,
7088                                            const MachineBasicBlock *MBB) const {
7089   assert((Inst.getNumOperands() == 3 || Inst.getNumOperands() == 4) &&
7090          "Reassociation needs binary operators");
7091 
7092   // Integer binary math/logic instructions have a third source operand:
7093   // the EFLAGS register. That operand must be both defined here and never
7094   // used; ie, it must be dead. If the EFLAGS operand is live, then we can
7095   // not change anything because rearranging the operands could affect other
7096   // instructions that depend on the exact status flags (zero, sign, etc.)
7097   // that are set by using these particular operands with this operation.
7098   if (Inst.getNumOperands() == 4) {
7099     assert(Inst.getOperand(3).isReg() &&
7100            Inst.getOperand(3).getReg() == X86::EFLAGS &&
7101            "Unexpected operand in reassociable instruction");
7102     if (!Inst.getOperand(3).isDead())
7103       return false;
7104   }
7105 
7106   return TargetInstrInfo::hasReassociableOperands(Inst, MBB);
7107 }
7108 
7109 // TODO: There are many more machine instruction opcodes to match:
7110 //       1. Other data types (integer, vectors)
7111 //       2. Other math / logic operations (xor, or)
7112 //       3. Other forms of the same operation (intrinsics and other variants)
isAssociativeAndCommutative(const MachineInstr & Inst) const7113 bool X86InstrInfo::isAssociativeAndCommutative(const MachineInstr &Inst) const {
7114   switch (Inst.getOpcode()) {
7115   case X86::AND8rr:
7116   case X86::AND16rr:
7117   case X86::AND32rr:
7118   case X86::AND64rr:
7119   case X86::OR8rr:
7120   case X86::OR16rr:
7121   case X86::OR32rr:
7122   case X86::OR64rr:
7123   case X86::XOR8rr:
7124   case X86::XOR16rr:
7125   case X86::XOR32rr:
7126   case X86::XOR64rr:
7127   case X86::IMUL16rr:
7128   case X86::IMUL32rr:
7129   case X86::IMUL64rr:
7130   case X86::PANDrr:
7131   case X86::PORrr:
7132   case X86::PXORrr:
7133   case X86::ANDPDrr:
7134   case X86::ANDPSrr:
7135   case X86::ORPDrr:
7136   case X86::ORPSrr:
7137   case X86::XORPDrr:
7138   case X86::XORPSrr:
7139   case X86::PADDBrr:
7140   case X86::PADDWrr:
7141   case X86::PADDDrr:
7142   case X86::PADDQrr:
7143   case X86::VPANDrr:
7144   case X86::VPANDYrr:
7145   case X86::VPANDDZ128rr:
7146   case X86::VPANDDZ256rr:
7147   case X86::VPANDDZrr:
7148   case X86::VPANDQZ128rr:
7149   case X86::VPANDQZ256rr:
7150   case X86::VPANDQZrr:
7151   case X86::VPORrr:
7152   case X86::VPORYrr:
7153   case X86::VPORDZ128rr:
7154   case X86::VPORDZ256rr:
7155   case X86::VPORDZrr:
7156   case X86::VPORQZ128rr:
7157   case X86::VPORQZ256rr:
7158   case X86::VPORQZrr:
7159   case X86::VPXORrr:
7160   case X86::VPXORYrr:
7161   case X86::VPXORDZ128rr:
7162   case X86::VPXORDZ256rr:
7163   case X86::VPXORDZrr:
7164   case X86::VPXORQZ128rr:
7165   case X86::VPXORQZ256rr:
7166   case X86::VPXORQZrr:
7167   case X86::VANDPDrr:
7168   case X86::VANDPSrr:
7169   case X86::VANDPDYrr:
7170   case X86::VANDPSYrr:
7171   case X86::VANDPDZ128rr:
7172   case X86::VANDPSZ128rr:
7173   case X86::VANDPDZ256rr:
7174   case X86::VANDPSZ256rr:
7175   case X86::VANDPDZrr:
7176   case X86::VANDPSZrr:
7177   case X86::VORPDrr:
7178   case X86::VORPSrr:
7179   case X86::VORPDYrr:
7180   case X86::VORPSYrr:
7181   case X86::VORPDZ128rr:
7182   case X86::VORPSZ128rr:
7183   case X86::VORPDZ256rr:
7184   case X86::VORPSZ256rr:
7185   case X86::VORPDZrr:
7186   case X86::VORPSZrr:
7187   case X86::VXORPDrr:
7188   case X86::VXORPSrr:
7189   case X86::VXORPDYrr:
7190   case X86::VXORPSYrr:
7191   case X86::VXORPDZ128rr:
7192   case X86::VXORPSZ128rr:
7193   case X86::VXORPDZ256rr:
7194   case X86::VXORPSZ256rr:
7195   case X86::VXORPDZrr:
7196   case X86::VXORPSZrr:
7197   case X86::KADDBrr:
7198   case X86::KADDWrr:
7199   case X86::KADDDrr:
7200   case X86::KADDQrr:
7201   case X86::KANDBrr:
7202   case X86::KANDWrr:
7203   case X86::KANDDrr:
7204   case X86::KANDQrr:
7205   case X86::KORBrr:
7206   case X86::KORWrr:
7207   case X86::KORDrr:
7208   case X86::KORQrr:
7209   case X86::KXORBrr:
7210   case X86::KXORWrr:
7211   case X86::KXORDrr:
7212   case X86::KXORQrr:
7213   case X86::VPADDBrr:
7214   case X86::VPADDWrr:
7215   case X86::VPADDDrr:
7216   case X86::VPADDQrr:
7217   case X86::VPADDBYrr:
7218   case X86::VPADDWYrr:
7219   case X86::VPADDDYrr:
7220   case X86::VPADDQYrr:
7221   case X86::VPADDBZ128rr:
7222   case X86::VPADDWZ128rr:
7223   case X86::VPADDDZ128rr:
7224   case X86::VPADDQZ128rr:
7225   case X86::VPADDBZ256rr:
7226   case X86::VPADDWZ256rr:
7227   case X86::VPADDDZ256rr:
7228   case X86::VPADDQZ256rr:
7229   case X86::VPADDBZrr:
7230   case X86::VPADDWZrr:
7231   case X86::VPADDDZrr:
7232   case X86::VPADDQZrr:
7233   case X86::VPMULLWrr:
7234   case X86::VPMULLWYrr:
7235   case X86::VPMULLWZ128rr:
7236   case X86::VPMULLWZ256rr:
7237   case X86::VPMULLWZrr:
7238   case X86::VPMULLDrr:
7239   case X86::VPMULLDYrr:
7240   case X86::VPMULLDZ128rr:
7241   case X86::VPMULLDZ256rr:
7242   case X86::VPMULLDZrr:
7243   case X86::VPMULLQZ128rr:
7244   case X86::VPMULLQZ256rr:
7245   case X86::VPMULLQZrr:
7246   // Normal min/max instructions are not commutative because of NaN and signed
7247   // zero semantics, but these are. Thus, there's no need to check for global
7248   // relaxed math; the instructions themselves have the properties we need.
7249   case X86::MAXCPDrr:
7250   case X86::MAXCPSrr:
7251   case X86::MAXCSDrr:
7252   case X86::MAXCSSrr:
7253   case X86::MINCPDrr:
7254   case X86::MINCPSrr:
7255   case X86::MINCSDrr:
7256   case X86::MINCSSrr:
7257   case X86::VMAXCPDrr:
7258   case X86::VMAXCPSrr:
7259   case X86::VMAXCPDYrr:
7260   case X86::VMAXCPSYrr:
7261   case X86::VMAXCPDZ128rr:
7262   case X86::VMAXCPSZ128rr:
7263   case X86::VMAXCPDZ256rr:
7264   case X86::VMAXCPSZ256rr:
7265   case X86::VMAXCPDZrr:
7266   case X86::VMAXCPSZrr:
7267   case X86::VMAXCSDrr:
7268   case X86::VMAXCSSrr:
7269   case X86::VMAXCSDZrr:
7270   case X86::VMAXCSSZrr:
7271   case X86::VMINCPDrr:
7272   case X86::VMINCPSrr:
7273   case X86::VMINCPDYrr:
7274   case X86::VMINCPSYrr:
7275   case X86::VMINCPDZ128rr:
7276   case X86::VMINCPSZ128rr:
7277   case X86::VMINCPDZ256rr:
7278   case X86::VMINCPSZ256rr:
7279   case X86::VMINCPDZrr:
7280   case X86::VMINCPSZrr:
7281   case X86::VMINCSDrr:
7282   case X86::VMINCSSrr:
7283   case X86::VMINCSDZrr:
7284   case X86::VMINCSSZrr:
7285     return true;
7286   case X86::ADDPDrr:
7287   case X86::ADDPSrr:
7288   case X86::ADDSDrr:
7289   case X86::ADDSSrr:
7290   case X86::MULPDrr:
7291   case X86::MULPSrr:
7292   case X86::MULSDrr:
7293   case X86::MULSSrr:
7294   case X86::VADDPDrr:
7295   case X86::VADDPSrr:
7296   case X86::VADDPDYrr:
7297   case X86::VADDPSYrr:
7298   case X86::VADDPDZ128rr:
7299   case X86::VADDPSZ128rr:
7300   case X86::VADDPDZ256rr:
7301   case X86::VADDPSZ256rr:
7302   case X86::VADDPDZrr:
7303   case X86::VADDPSZrr:
7304   case X86::VADDSDrr:
7305   case X86::VADDSSrr:
7306   case X86::VADDSDZrr:
7307   case X86::VADDSSZrr:
7308   case X86::VMULPDrr:
7309   case X86::VMULPSrr:
7310   case X86::VMULPDYrr:
7311   case X86::VMULPSYrr:
7312   case X86::VMULPDZ128rr:
7313   case X86::VMULPSZ128rr:
7314   case X86::VMULPDZ256rr:
7315   case X86::VMULPSZ256rr:
7316   case X86::VMULPDZrr:
7317   case X86::VMULPSZrr:
7318   case X86::VMULSDrr:
7319   case X86::VMULSSrr:
7320   case X86::VMULSDZrr:
7321   case X86::VMULSSZrr:
7322     return Inst.getParent()->getParent()->getTarget().Options.UnsafeFPMath;
7323   default:
7324     return false;
7325   }
7326 }
7327 
7328 /// This is an architecture-specific helper function of reassociateOps.
7329 /// Set special operand attributes for new instructions after reassociation.
setSpecialOperandAttr(MachineInstr & OldMI1,MachineInstr & OldMI2,MachineInstr & NewMI1,MachineInstr & NewMI2) const7330 void X86InstrInfo::setSpecialOperandAttr(MachineInstr &OldMI1,
7331                                          MachineInstr &OldMI2,
7332                                          MachineInstr &NewMI1,
7333                                          MachineInstr &NewMI2) const {
7334   // Integer instructions define an implicit EFLAGS source register operand as
7335   // the third source (fourth total) operand.
7336   if (OldMI1.getNumOperands() != 4 || OldMI2.getNumOperands() != 4)
7337     return;
7338 
7339   assert(NewMI1.getNumOperands() == 4 && NewMI2.getNumOperands() == 4 &&
7340          "Unexpected instruction type for reassociation");
7341 
7342   MachineOperand &OldOp1 = OldMI1.getOperand(3);
7343   MachineOperand &OldOp2 = OldMI2.getOperand(3);
7344   MachineOperand &NewOp1 = NewMI1.getOperand(3);
7345   MachineOperand &NewOp2 = NewMI2.getOperand(3);
7346 
7347   assert(OldOp1.isReg() && OldOp1.getReg() == X86::EFLAGS && OldOp1.isDead() &&
7348          "Must have dead EFLAGS operand in reassociable instruction");
7349   assert(OldOp2.isReg() && OldOp2.getReg() == X86::EFLAGS && OldOp2.isDead() &&
7350          "Must have dead EFLAGS operand in reassociable instruction");
7351 
7352   (void)OldOp1;
7353   (void)OldOp2;
7354 
7355   assert(NewOp1.isReg() && NewOp1.getReg() == X86::EFLAGS &&
7356          "Unexpected operand in reassociable instruction");
7357   assert(NewOp2.isReg() && NewOp2.getReg() == X86::EFLAGS &&
7358          "Unexpected operand in reassociable instruction");
7359 
7360   // Mark the new EFLAGS operands as dead to be helpful to subsequent iterations
7361   // of this pass or other passes. The EFLAGS operands must be dead in these new
7362   // instructions because the EFLAGS operands in the original instructions must
7363   // be dead in order for reassociation to occur.
7364   NewOp1.setIsDead();
7365   NewOp2.setIsDead();
7366 }
7367 
7368 std::pair<unsigned, unsigned>
decomposeMachineOperandsTargetFlags(unsigned TF) const7369 X86InstrInfo::decomposeMachineOperandsTargetFlags(unsigned TF) const {
7370   return std::make_pair(TF, 0u);
7371 }
7372 
7373 ArrayRef<std::pair<unsigned, const char *>>
getSerializableDirectMachineOperandTargetFlags() const7374 X86InstrInfo::getSerializableDirectMachineOperandTargetFlags() const {
7375   using namespace X86II;
7376   static const std::pair<unsigned, const char *> TargetFlags[] = {
7377       {MO_GOT_ABSOLUTE_ADDRESS, "x86-got-absolute-address"},
7378       {MO_PIC_BASE_OFFSET, "x86-pic-base-offset"},
7379       {MO_GOT, "x86-got"},
7380       {MO_GOTOFF, "x86-gotoff"},
7381       {MO_GOTPCREL, "x86-gotpcrel"},
7382       {MO_PLT, "x86-plt"},
7383       {MO_TLSGD, "x86-tlsgd"},
7384       {MO_TLSLD, "x86-tlsld"},
7385       {MO_TLSLDM, "x86-tlsldm"},
7386       {MO_GOTTPOFF, "x86-gottpoff"},
7387       {MO_INDNTPOFF, "x86-indntpoff"},
7388       {MO_TPOFF, "x86-tpoff"},
7389       {MO_DTPOFF, "x86-dtpoff"},
7390       {MO_NTPOFF, "x86-ntpoff"},
7391       {MO_GOTNTPOFF, "x86-gotntpoff"},
7392       {MO_DLLIMPORT, "x86-dllimport"},
7393       {MO_DARWIN_NONLAZY, "x86-darwin-nonlazy"},
7394       {MO_DARWIN_NONLAZY_PIC_BASE, "x86-darwin-nonlazy-pic-base"},
7395       {MO_TLVP, "x86-tlvp"},
7396       {MO_TLVP_PIC_BASE, "x86-tlvp-pic-base"},
7397       {MO_SECREL, "x86-secrel"},
7398       {MO_COFFSTUB, "x86-coffstub"}};
7399   return makeArrayRef(TargetFlags);
7400 }
7401 
7402 namespace {
7403   /// Create Global Base Reg pass. This initializes the PIC
7404   /// global base register for x86-32.
7405   struct CGBR : public MachineFunctionPass {
7406     static char ID;
CGBR__anondb2c66a20511::CGBR7407     CGBR() : MachineFunctionPass(ID) {}
7408 
runOnMachineFunction__anondb2c66a20511::CGBR7409     bool runOnMachineFunction(MachineFunction &MF) override {
7410       const X86TargetMachine *TM =
7411         static_cast<const X86TargetMachine *>(&MF.getTarget());
7412       const X86Subtarget &STI = MF.getSubtarget<X86Subtarget>();
7413 
7414       // Don't do anything in the 64-bit small and kernel code models. They use
7415       // RIP-relative addressing for everything.
7416       if (STI.is64Bit() && (TM->getCodeModel() == CodeModel::Small ||
7417                             TM->getCodeModel() == CodeModel::Kernel))
7418         return false;
7419 
7420       // Only emit a global base reg in PIC mode.
7421       if (!TM->isPositionIndependent())
7422         return false;
7423 
7424       X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
7425       unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
7426 
7427       // If we didn't need a GlobalBaseReg, don't insert code.
7428       if (GlobalBaseReg == 0)
7429         return false;
7430 
7431       // Insert the set of GlobalBaseReg into the first MBB of the function
7432       MachineBasicBlock &FirstMBB = MF.front();
7433       MachineBasicBlock::iterator MBBI = FirstMBB.begin();
7434       DebugLoc DL = FirstMBB.findDebugLoc(MBBI);
7435       MachineRegisterInfo &RegInfo = MF.getRegInfo();
7436       const X86InstrInfo *TII = STI.getInstrInfo();
7437 
7438       unsigned PC;
7439       if (STI.isPICStyleGOT())
7440         PC = RegInfo.createVirtualRegister(&X86::GR32RegClass);
7441       else
7442         PC = GlobalBaseReg;
7443 
7444       if (STI.is64Bit()) {
7445         if (TM->getCodeModel() == CodeModel::Medium) {
7446           // In the medium code model, use a RIP-relative LEA to materialize the
7447           // GOT.
7448           BuildMI(FirstMBB, MBBI, DL, TII->get(X86::LEA64r), PC)
7449               .addReg(X86::RIP)
7450               .addImm(0)
7451               .addReg(0)
7452               .addExternalSymbol("_GLOBAL_OFFSET_TABLE_")
7453               .addReg(0);
7454         } else if (TM->getCodeModel() == CodeModel::Large) {
7455           // In the large code model, we are aiming for this code, though the
7456           // register allocation may vary:
7457           //   leaq .LN$pb(%rip), %rax
7458           //   movq $_GLOBAL_OFFSET_TABLE_ - .LN$pb, %rcx
7459           //   addq %rcx, %rax
7460           // RAX now holds address of _GLOBAL_OFFSET_TABLE_.
7461           unsigned PBReg = RegInfo.createVirtualRegister(&X86::GR64RegClass);
7462           unsigned GOTReg =
7463               RegInfo.createVirtualRegister(&X86::GR64RegClass);
7464           BuildMI(FirstMBB, MBBI, DL, TII->get(X86::LEA64r), PBReg)
7465               .addReg(X86::RIP)
7466               .addImm(0)
7467               .addReg(0)
7468               .addSym(MF.getPICBaseSymbol())
7469               .addReg(0);
7470           std::prev(MBBI)->setPreInstrSymbol(MF, MF.getPICBaseSymbol());
7471           BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOV64ri), GOTReg)
7472               .addExternalSymbol("_GLOBAL_OFFSET_TABLE_",
7473                                  X86II::MO_PIC_BASE_OFFSET);
7474           BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD64rr), PC)
7475               .addReg(PBReg, RegState::Kill)
7476               .addReg(GOTReg, RegState::Kill);
7477         } else {
7478           llvm_unreachable("unexpected code model");
7479         }
7480       } else {
7481         // Operand of MovePCtoStack is completely ignored by asm printer. It's
7482         // only used in JIT code emission as displacement to pc.
7483         BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOVPC32r), PC).addImm(0);
7484 
7485         // If we're using vanilla 'GOT' PIC style, we should use relative
7486         // addressing not to pc, but to _GLOBAL_OFFSET_TABLE_ external.
7487         if (STI.isPICStyleGOT()) {
7488           // Generate addl $__GLOBAL_OFFSET_TABLE_ + [.-piclabel],
7489           // %some_register
7490           BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD32ri), GlobalBaseReg)
7491               .addReg(PC)
7492               .addExternalSymbol("_GLOBAL_OFFSET_TABLE_",
7493                                  X86II::MO_GOT_ABSOLUTE_ADDRESS);
7494         }
7495       }
7496 
7497       return true;
7498     }
7499 
getPassName__anondb2c66a20511::CGBR7500     StringRef getPassName() const override {
7501       return "X86 PIC Global Base Reg Initialization";
7502     }
7503 
getAnalysisUsage__anondb2c66a20511::CGBR7504     void getAnalysisUsage(AnalysisUsage &AU) const override {
7505       AU.setPreservesCFG();
7506       MachineFunctionPass::getAnalysisUsage(AU);
7507     }
7508   };
7509 }
7510 
7511 char CGBR::ID = 0;
7512 FunctionPass*
createX86GlobalBaseRegPass()7513 llvm::createX86GlobalBaseRegPass() { return new CGBR(); }
7514 
7515 namespace {
7516   struct LDTLSCleanup : public MachineFunctionPass {
7517     static char ID;
LDTLSCleanup__anondb2c66a20611::LDTLSCleanup7518     LDTLSCleanup() : MachineFunctionPass(ID) {}
7519 
runOnMachineFunction__anondb2c66a20611::LDTLSCleanup7520     bool runOnMachineFunction(MachineFunction &MF) override {
7521       if (skipFunction(MF.getFunction()))
7522         return false;
7523 
7524       X86MachineFunctionInfo *MFI = MF.getInfo<X86MachineFunctionInfo>();
7525       if (MFI->getNumLocalDynamicTLSAccesses() < 2) {
7526         // No point folding accesses if there isn't at least two.
7527         return false;
7528       }
7529 
7530       MachineDominatorTree *DT = &getAnalysis<MachineDominatorTree>();
7531       return VisitNode(DT->getRootNode(), 0);
7532     }
7533 
7534     // Visit the dominator subtree rooted at Node in pre-order.
7535     // If TLSBaseAddrReg is non-null, then use that to replace any
7536     // TLS_base_addr instructions. Otherwise, create the register
7537     // when the first such instruction is seen, and then use it
7538     // as we encounter more instructions.
VisitNode__anondb2c66a20611::LDTLSCleanup7539     bool VisitNode(MachineDomTreeNode *Node, unsigned TLSBaseAddrReg) {
7540       MachineBasicBlock *BB = Node->getBlock();
7541       bool Changed = false;
7542 
7543       // Traverse the current block.
7544       for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;
7545            ++I) {
7546         switch (I->getOpcode()) {
7547           case X86::TLS_base_addr32:
7548           case X86::TLS_base_addr64:
7549             if (TLSBaseAddrReg)
7550               I = ReplaceTLSBaseAddrCall(*I, TLSBaseAddrReg);
7551             else
7552               I = SetRegister(*I, &TLSBaseAddrReg);
7553             Changed = true;
7554             break;
7555           default:
7556             break;
7557         }
7558       }
7559 
7560       // Visit the children of this block in the dominator tree.
7561       for (MachineDomTreeNode::iterator I = Node->begin(), E = Node->end();
7562            I != E; ++I) {
7563         Changed |= VisitNode(*I, TLSBaseAddrReg);
7564       }
7565 
7566       return Changed;
7567     }
7568 
7569     // Replace the TLS_base_addr instruction I with a copy from
7570     // TLSBaseAddrReg, returning the new instruction.
ReplaceTLSBaseAddrCall__anondb2c66a20611::LDTLSCleanup7571     MachineInstr *ReplaceTLSBaseAddrCall(MachineInstr &I,
7572                                          unsigned TLSBaseAddrReg) {
7573       MachineFunction *MF = I.getParent()->getParent();
7574       const X86Subtarget &STI = MF->getSubtarget<X86Subtarget>();
7575       const bool is64Bit = STI.is64Bit();
7576       const X86InstrInfo *TII = STI.getInstrInfo();
7577 
7578       // Insert a Copy from TLSBaseAddrReg to RAX/EAX.
7579       MachineInstr *Copy =
7580           BuildMI(*I.getParent(), I, I.getDebugLoc(),
7581                   TII->get(TargetOpcode::COPY), is64Bit ? X86::RAX : X86::EAX)
7582               .addReg(TLSBaseAddrReg);
7583 
7584       // Erase the TLS_base_addr instruction.
7585       I.eraseFromParent();
7586 
7587       return Copy;
7588     }
7589 
7590     // Create a virtual register in *TLSBaseAddrReg, and populate it by
7591     // inserting a copy instruction after I. Returns the new instruction.
SetRegister__anondb2c66a20611::LDTLSCleanup7592     MachineInstr *SetRegister(MachineInstr &I, unsigned *TLSBaseAddrReg) {
7593       MachineFunction *MF = I.getParent()->getParent();
7594       const X86Subtarget &STI = MF->getSubtarget<X86Subtarget>();
7595       const bool is64Bit = STI.is64Bit();
7596       const X86InstrInfo *TII = STI.getInstrInfo();
7597 
7598       // Create a virtual register for the TLS base address.
7599       MachineRegisterInfo &RegInfo = MF->getRegInfo();
7600       *TLSBaseAddrReg = RegInfo.createVirtualRegister(is64Bit
7601                                                       ? &X86::GR64RegClass
7602                                                       : &X86::GR32RegClass);
7603 
7604       // Insert a copy from RAX/EAX to TLSBaseAddrReg.
7605       MachineInstr *Next = I.getNextNode();
7606       MachineInstr *Copy =
7607           BuildMI(*I.getParent(), Next, I.getDebugLoc(),
7608                   TII->get(TargetOpcode::COPY), *TLSBaseAddrReg)
7609               .addReg(is64Bit ? X86::RAX : X86::EAX);
7610 
7611       return Copy;
7612     }
7613 
getPassName__anondb2c66a20611::LDTLSCleanup7614     StringRef getPassName() const override {
7615       return "Local Dynamic TLS Access Clean-up";
7616     }
7617 
getAnalysisUsage__anondb2c66a20611::LDTLSCleanup7618     void getAnalysisUsage(AnalysisUsage &AU) const override {
7619       AU.setPreservesCFG();
7620       AU.addRequired<MachineDominatorTree>();
7621       MachineFunctionPass::getAnalysisUsage(AU);
7622     }
7623   };
7624 }
7625 
7626 char LDTLSCleanup::ID = 0;
7627 FunctionPass*
createCleanupLocalDynamicTLSPass()7628 llvm::createCleanupLocalDynamicTLSPass() { return new LDTLSCleanup(); }
7629 
7630 /// Constants defining how certain sequences should be outlined.
7631 ///
7632 /// \p MachineOutlinerDefault implies that the function is called with a call
7633 /// instruction, and a return must be emitted for the outlined function frame.
7634 ///
7635 /// That is,
7636 ///
7637 /// I1                                 OUTLINED_FUNCTION:
7638 /// I2 --> call OUTLINED_FUNCTION       I1
7639 /// I3                                  I2
7640 ///                                     I3
7641 ///                                     ret
7642 ///
7643 /// * Call construction overhead: 1 (call instruction)
7644 /// * Frame construction overhead: 1 (return instruction)
7645 ///
7646 /// \p MachineOutlinerTailCall implies that the function is being tail called.
7647 /// A jump is emitted instead of a call, and the return is already present in
7648 /// the outlined sequence. That is,
7649 ///
7650 /// I1                                 OUTLINED_FUNCTION:
7651 /// I2 --> jmp OUTLINED_FUNCTION       I1
7652 /// ret                                I2
7653 ///                                    ret
7654 ///
7655 /// * Call construction overhead: 1 (jump instruction)
7656 /// * Frame construction overhead: 0 (don't need to return)
7657 ///
7658 enum MachineOutlinerClass {
7659   MachineOutlinerDefault,
7660   MachineOutlinerTailCall
7661 };
7662 
getOutliningCandidateInfo(std::vector<outliner::Candidate> & RepeatedSequenceLocs) const7663 outliner::OutlinedFunction X86InstrInfo::getOutliningCandidateInfo(
7664     std::vector<outliner::Candidate> &RepeatedSequenceLocs) const {
7665   unsigned SequenceSize =
7666       std::accumulate(RepeatedSequenceLocs[0].front(),
7667                       std::next(RepeatedSequenceLocs[0].back()), 0,
7668                       [](unsigned Sum, const MachineInstr &MI) {
7669                         // FIXME: x86 doesn't implement getInstSizeInBytes, so
7670                         // we can't tell the cost.  Just assume each instruction
7671                         // is one byte.
7672                         if (MI.isDebugInstr() || MI.isKill())
7673                           return Sum;
7674                         return Sum + 1;
7675                       });
7676 
7677   // FIXME: Use real size in bytes for call and ret instructions.
7678   if (RepeatedSequenceLocs[0].back()->isTerminator()) {
7679     for (outliner::Candidate &C : RepeatedSequenceLocs)
7680       C.setCallInfo(MachineOutlinerTailCall, 1);
7681 
7682     return outliner::OutlinedFunction(RepeatedSequenceLocs, SequenceSize,
7683                                       0, // Number of bytes to emit frame.
7684                                       MachineOutlinerTailCall // Type of frame.
7685     );
7686   }
7687 
7688   for (outliner::Candidate &C : RepeatedSequenceLocs)
7689     C.setCallInfo(MachineOutlinerDefault, 1);
7690 
7691   return outliner::OutlinedFunction(RepeatedSequenceLocs, SequenceSize, 1,
7692                                     MachineOutlinerDefault);
7693 }
7694 
isFunctionSafeToOutlineFrom(MachineFunction & MF,bool OutlineFromLinkOnceODRs) const7695 bool X86InstrInfo::isFunctionSafeToOutlineFrom(MachineFunction &MF,
7696                                            bool OutlineFromLinkOnceODRs) const {
7697   const Function &F = MF.getFunction();
7698 
7699   // Does the function use a red zone? If it does, then we can't risk messing
7700   // with the stack.
7701   if (!F.hasFnAttribute(Attribute::NoRedZone)) {
7702     // It could have a red zone. If it does, then we don't want to touch it.
7703     const X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
7704     if (!X86FI || X86FI->getUsesRedZone())
7705       return false;
7706   }
7707 
7708   // If we *don't* want to outline from things that could potentially be deduped
7709   // then return false.
7710   if (!OutlineFromLinkOnceODRs && F.hasLinkOnceODRLinkage())
7711       return false;
7712 
7713   // This function is viable for outlining, so return true.
7714   return true;
7715 }
7716 
7717 outliner::InstrType
getOutliningType(MachineBasicBlock::iterator & MIT,unsigned Flags) const7718 X86InstrInfo::getOutliningType(MachineBasicBlock::iterator &MIT,  unsigned Flags) const {
7719   MachineInstr &MI = *MIT;
7720   // Don't allow debug values to impact outlining type.
7721   if (MI.isDebugInstr() || MI.isIndirectDebugValue())
7722     return outliner::InstrType::Invisible;
7723 
7724   // At this point, KILL instructions don't really tell us much so we can go
7725   // ahead and skip over them.
7726   if (MI.isKill())
7727     return outliner::InstrType::Invisible;
7728 
7729   // Is this a tail call? If yes, we can outline as a tail call.
7730   if (isTailCall(MI))
7731     return outliner::InstrType::Legal;
7732 
7733   // Is this the terminator of a basic block?
7734   if (MI.isTerminator() || MI.isReturn()) {
7735 
7736     // Does its parent have any successors in its MachineFunction?
7737     if (MI.getParent()->succ_empty())
7738       return outliner::InstrType::Legal;
7739 
7740     // It does, so we can't tail call it.
7741     return outliner::InstrType::Illegal;
7742   }
7743 
7744   // Don't outline anything that modifies or reads from the stack pointer.
7745   //
7746   // FIXME: There are instructions which are being manually built without
7747   // explicit uses/defs so we also have to check the MCInstrDesc. We should be
7748   // able to remove the extra checks once those are fixed up. For example,
7749   // sometimes we might get something like %rax = POP64r 1. This won't be
7750   // caught by modifiesRegister or readsRegister even though the instruction
7751   // really ought to be formed so that modifiesRegister/readsRegister would
7752   // catch it.
7753   if (MI.modifiesRegister(X86::RSP, &RI) || MI.readsRegister(X86::RSP, &RI) ||
7754       MI.getDesc().hasImplicitUseOfPhysReg(X86::RSP) ||
7755       MI.getDesc().hasImplicitDefOfPhysReg(X86::RSP))
7756     return outliner::InstrType::Illegal;
7757 
7758   // Outlined calls change the instruction pointer, so don't read from it.
7759   if (MI.readsRegister(X86::RIP, &RI) ||
7760       MI.getDesc().hasImplicitUseOfPhysReg(X86::RIP) ||
7761       MI.getDesc().hasImplicitDefOfPhysReg(X86::RIP))
7762     return outliner::InstrType::Illegal;
7763 
7764   // Positions can't safely be outlined.
7765   if (MI.isPosition())
7766     return outliner::InstrType::Illegal;
7767 
7768   // Make sure none of the operands of this instruction do anything tricky.
7769   for (const MachineOperand &MOP : MI.operands())
7770     if (MOP.isCPI() || MOP.isJTI() || MOP.isCFIIndex() || MOP.isFI() ||
7771         MOP.isTargetIndex())
7772       return outliner::InstrType::Illegal;
7773 
7774   return outliner::InstrType::Legal;
7775 }
7776 
buildOutlinedFrame(MachineBasicBlock & MBB,MachineFunction & MF,const outliner::OutlinedFunction & OF) const7777 void X86InstrInfo::buildOutlinedFrame(MachineBasicBlock &MBB,
7778                                           MachineFunction &MF,
7779                                           const outliner::OutlinedFunction &OF)
7780                                           const {
7781   // If we're a tail call, we already have a return, so don't do anything.
7782   if (OF.FrameConstructionID == MachineOutlinerTailCall)
7783     return;
7784 
7785   // We're a normal call, so our sequence doesn't have a return instruction.
7786   // Add it in.
7787   MachineInstr *retq = BuildMI(MF, DebugLoc(), get(X86::RETQ));
7788   MBB.insert(MBB.end(), retq);
7789 }
7790 
7791 MachineBasicBlock::iterator
insertOutlinedCall(Module & M,MachineBasicBlock & MBB,MachineBasicBlock::iterator & It,MachineFunction & MF,const outliner::Candidate & C) const7792 X86InstrInfo::insertOutlinedCall(Module &M, MachineBasicBlock &MBB,
7793                                  MachineBasicBlock::iterator &It,
7794                                  MachineFunction &MF,
7795                                  const outliner::Candidate &C) const {
7796   // Is it a tail call?
7797   if (C.CallConstructionID == MachineOutlinerTailCall) {
7798     // Yes, just insert a JMP.
7799     It = MBB.insert(It,
7800                   BuildMI(MF, DebugLoc(), get(X86::TAILJMPd64))
7801                       .addGlobalAddress(M.getNamedValue(MF.getName())));
7802   } else {
7803     // No, insert a call.
7804     It = MBB.insert(It,
7805                   BuildMI(MF, DebugLoc(), get(X86::CALL64pcrel32))
7806                       .addGlobalAddress(M.getNamedValue(MF.getName())));
7807   }
7808 
7809   return It;
7810 }
7811 
7812 #define GET_INSTRINFO_HELPERS
7813 #include "X86GenInstrInfo.inc"
7814