1 //===-- X86InstrInfo.cpp - X86 Instruction Information --------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file contains the X86 implementation of the TargetInstrInfo class.
10 //
11 //===----------------------------------------------------------------------===//
12
13 #include "X86InstrInfo.h"
14 #include "X86.h"
15 #include "X86InstrBuilder.h"
16 #include "X86InstrFoldTables.h"
17 #include "X86MachineFunctionInfo.h"
18 #include "X86Subtarget.h"
19 #include "X86TargetMachine.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/ADT/Sequence.h"
22 #include "llvm/CodeGen/LivePhysRegs.h"
23 #include "llvm/CodeGen/LiveVariables.h"
24 #include "llvm/CodeGen/MachineConstantPool.h"
25 #include "llvm/CodeGen/MachineDominators.h"
26 #include "llvm/CodeGen/MachineFrameInfo.h"
27 #include "llvm/CodeGen/MachineInstrBuilder.h"
28 #include "llvm/CodeGen/MachineModuleInfo.h"
29 #include "llvm/CodeGen/MachineRegisterInfo.h"
30 #include "llvm/CodeGen/StackMaps.h"
31 #include "llvm/IR/DerivedTypes.h"
32 #include "llvm/IR/Function.h"
33 #include "llvm/IR/DebugInfoMetadata.h"
34 #include "llvm/MC/MCAsmInfo.h"
35 #include "llvm/MC/MCExpr.h"
36 #include "llvm/MC/MCInst.h"
37 #include "llvm/Support/CommandLine.h"
38 #include "llvm/Support/Debug.h"
39 #include "llvm/Support/ErrorHandling.h"
40 #include "llvm/Support/raw_ostream.h"
41 #include "llvm/Target/TargetOptions.h"
42
43 using namespace llvm;
44
45 #define DEBUG_TYPE "x86-instr-info"
46
47 #define GET_INSTRINFO_CTOR_DTOR
48 #include "X86GenInstrInfo.inc"
49
50 static cl::opt<bool>
51 NoFusing("disable-spill-fusing",
52 cl::desc("Disable fusing of spill code into instructions"),
53 cl::Hidden);
54 static cl::opt<bool>
55 PrintFailedFusing("print-failed-fuse-candidates",
56 cl::desc("Print instructions that the allocator wants to"
57 " fuse, but the X86 backend currently can't"),
58 cl::Hidden);
59 static cl::opt<bool>
60 ReMatPICStubLoad("remat-pic-stub-load",
61 cl::desc("Re-materialize load from stub in PIC mode"),
62 cl::init(false), cl::Hidden);
63 static cl::opt<unsigned>
64 PartialRegUpdateClearance("partial-reg-update-clearance",
65 cl::desc("Clearance between two register writes "
66 "for inserting XOR to avoid partial "
67 "register update"),
68 cl::init(64), cl::Hidden);
69 static cl::opt<unsigned>
70 UndefRegClearance("undef-reg-clearance",
71 cl::desc("How many idle instructions we would like before "
72 "certain undef register reads"),
73 cl::init(128), cl::Hidden);
74
75
76 // Pin the vtable to this file.
anchor()77 void X86InstrInfo::anchor() {}
78
X86InstrInfo(X86Subtarget & STI)79 X86InstrInfo::X86InstrInfo(X86Subtarget &STI)
80 : X86GenInstrInfo((STI.isTarget64BitLP64() ? X86::ADJCALLSTACKDOWN64
81 : X86::ADJCALLSTACKDOWN32),
82 (STI.isTarget64BitLP64() ? X86::ADJCALLSTACKUP64
83 : X86::ADJCALLSTACKUP32),
84 X86::CATCHRET,
85 (STI.is64Bit() ? X86::RETQ : X86::RETL)),
86 Subtarget(STI), RI(STI.getTargetTriple()) {
87 }
88
89 bool
isCoalescableExtInstr(const MachineInstr & MI,unsigned & SrcReg,unsigned & DstReg,unsigned & SubIdx) const90 X86InstrInfo::isCoalescableExtInstr(const MachineInstr &MI,
91 unsigned &SrcReg, unsigned &DstReg,
92 unsigned &SubIdx) const {
93 switch (MI.getOpcode()) {
94 default: break;
95 case X86::MOVSX16rr8:
96 case X86::MOVZX16rr8:
97 case X86::MOVSX32rr8:
98 case X86::MOVZX32rr8:
99 case X86::MOVSX64rr8:
100 if (!Subtarget.is64Bit())
101 // It's not always legal to reference the low 8-bit of the larger
102 // register in 32-bit mode.
103 return false;
104 LLVM_FALLTHROUGH;
105 case X86::MOVSX32rr16:
106 case X86::MOVZX32rr16:
107 case X86::MOVSX64rr16:
108 case X86::MOVSX64rr32: {
109 if (MI.getOperand(0).getSubReg() || MI.getOperand(1).getSubReg())
110 // Be conservative.
111 return false;
112 SrcReg = MI.getOperand(1).getReg();
113 DstReg = MI.getOperand(0).getReg();
114 switch (MI.getOpcode()) {
115 default: llvm_unreachable("Unreachable!");
116 case X86::MOVSX16rr8:
117 case X86::MOVZX16rr8:
118 case X86::MOVSX32rr8:
119 case X86::MOVZX32rr8:
120 case X86::MOVSX64rr8:
121 SubIdx = X86::sub_8bit;
122 break;
123 case X86::MOVSX32rr16:
124 case X86::MOVZX32rr16:
125 case X86::MOVSX64rr16:
126 SubIdx = X86::sub_16bit;
127 break;
128 case X86::MOVSX64rr32:
129 SubIdx = X86::sub_32bit;
130 break;
131 }
132 return true;
133 }
134 }
135 return false;
136 }
137
getSPAdjust(const MachineInstr & MI) const138 int X86InstrInfo::getSPAdjust(const MachineInstr &MI) const {
139 const MachineFunction *MF = MI.getParent()->getParent();
140 const TargetFrameLowering *TFI = MF->getSubtarget().getFrameLowering();
141
142 if (isFrameInstr(MI)) {
143 unsigned StackAlign = TFI->getStackAlignment();
144 int SPAdj = alignTo(getFrameSize(MI), StackAlign);
145 SPAdj -= getFrameAdjustment(MI);
146 if (!isFrameSetup(MI))
147 SPAdj = -SPAdj;
148 return SPAdj;
149 }
150
151 // To know whether a call adjusts the stack, we need information
152 // that is bound to the following ADJCALLSTACKUP pseudo.
153 // Look for the next ADJCALLSTACKUP that follows the call.
154 if (MI.isCall()) {
155 const MachineBasicBlock *MBB = MI.getParent();
156 auto I = ++MachineBasicBlock::const_iterator(MI);
157 for (auto E = MBB->end(); I != E; ++I) {
158 if (I->getOpcode() == getCallFrameDestroyOpcode() ||
159 I->isCall())
160 break;
161 }
162
163 // If we could not find a frame destroy opcode, then it has already
164 // been simplified, so we don't care.
165 if (I->getOpcode() != getCallFrameDestroyOpcode())
166 return 0;
167
168 return -(I->getOperand(1).getImm());
169 }
170
171 // Currently handle only PUSHes we can reasonably expect to see
172 // in call sequences
173 switch (MI.getOpcode()) {
174 default:
175 return 0;
176 case X86::PUSH32i8:
177 case X86::PUSH32r:
178 case X86::PUSH32rmm:
179 case X86::PUSH32rmr:
180 case X86::PUSHi32:
181 return 4;
182 case X86::PUSH64i8:
183 case X86::PUSH64r:
184 case X86::PUSH64rmm:
185 case X86::PUSH64rmr:
186 case X86::PUSH64i32:
187 return 8;
188 }
189 }
190
191 /// Return true and the FrameIndex if the specified
192 /// operand and follow operands form a reference to the stack frame.
isFrameOperand(const MachineInstr & MI,unsigned int Op,int & FrameIndex) const193 bool X86InstrInfo::isFrameOperand(const MachineInstr &MI, unsigned int Op,
194 int &FrameIndex) const {
195 if (MI.getOperand(Op + X86::AddrBaseReg).isFI() &&
196 MI.getOperand(Op + X86::AddrScaleAmt).isImm() &&
197 MI.getOperand(Op + X86::AddrIndexReg).isReg() &&
198 MI.getOperand(Op + X86::AddrDisp).isImm() &&
199 MI.getOperand(Op + X86::AddrScaleAmt).getImm() == 1 &&
200 MI.getOperand(Op + X86::AddrIndexReg).getReg() == 0 &&
201 MI.getOperand(Op + X86::AddrDisp).getImm() == 0) {
202 FrameIndex = MI.getOperand(Op + X86::AddrBaseReg).getIndex();
203 return true;
204 }
205 return false;
206 }
207
isFrameLoadOpcode(int Opcode,unsigned & MemBytes)208 static bool isFrameLoadOpcode(int Opcode, unsigned &MemBytes) {
209 switch (Opcode) {
210 default:
211 return false;
212 case X86::MOV8rm:
213 case X86::KMOVBkm:
214 MemBytes = 1;
215 return true;
216 case X86::MOV16rm:
217 case X86::KMOVWkm:
218 MemBytes = 2;
219 return true;
220 case X86::MOV32rm:
221 case X86::MOVSSrm:
222 case X86::MOVSSrm_alt:
223 case X86::VMOVSSrm:
224 case X86::VMOVSSrm_alt:
225 case X86::VMOVSSZrm:
226 case X86::VMOVSSZrm_alt:
227 case X86::KMOVDkm:
228 MemBytes = 4;
229 return true;
230 case X86::MOV64rm:
231 case X86::LD_Fp64m:
232 case X86::MOVSDrm:
233 case X86::MOVSDrm_alt:
234 case X86::VMOVSDrm:
235 case X86::VMOVSDrm_alt:
236 case X86::VMOVSDZrm:
237 case X86::VMOVSDZrm_alt:
238 case X86::MMX_MOVD64rm:
239 case X86::MMX_MOVQ64rm:
240 case X86::KMOVQkm:
241 MemBytes = 8;
242 return true;
243 case X86::MOVAPSrm:
244 case X86::MOVUPSrm:
245 case X86::MOVAPDrm:
246 case X86::MOVUPDrm:
247 case X86::MOVDQArm:
248 case X86::MOVDQUrm:
249 case X86::VMOVAPSrm:
250 case X86::VMOVUPSrm:
251 case X86::VMOVAPDrm:
252 case X86::VMOVUPDrm:
253 case X86::VMOVDQArm:
254 case X86::VMOVDQUrm:
255 case X86::VMOVAPSZ128rm:
256 case X86::VMOVUPSZ128rm:
257 case X86::VMOVAPSZ128rm_NOVLX:
258 case X86::VMOVUPSZ128rm_NOVLX:
259 case X86::VMOVAPDZ128rm:
260 case X86::VMOVUPDZ128rm:
261 case X86::VMOVDQU8Z128rm:
262 case X86::VMOVDQU16Z128rm:
263 case X86::VMOVDQA32Z128rm:
264 case X86::VMOVDQU32Z128rm:
265 case X86::VMOVDQA64Z128rm:
266 case X86::VMOVDQU64Z128rm:
267 MemBytes = 16;
268 return true;
269 case X86::VMOVAPSYrm:
270 case X86::VMOVUPSYrm:
271 case X86::VMOVAPDYrm:
272 case X86::VMOVUPDYrm:
273 case X86::VMOVDQAYrm:
274 case X86::VMOVDQUYrm:
275 case X86::VMOVAPSZ256rm:
276 case X86::VMOVUPSZ256rm:
277 case X86::VMOVAPSZ256rm_NOVLX:
278 case X86::VMOVUPSZ256rm_NOVLX:
279 case X86::VMOVAPDZ256rm:
280 case X86::VMOVUPDZ256rm:
281 case X86::VMOVDQU8Z256rm:
282 case X86::VMOVDQU16Z256rm:
283 case X86::VMOVDQA32Z256rm:
284 case X86::VMOVDQU32Z256rm:
285 case X86::VMOVDQA64Z256rm:
286 case X86::VMOVDQU64Z256rm:
287 MemBytes = 32;
288 return true;
289 case X86::VMOVAPSZrm:
290 case X86::VMOVUPSZrm:
291 case X86::VMOVAPDZrm:
292 case X86::VMOVUPDZrm:
293 case X86::VMOVDQU8Zrm:
294 case X86::VMOVDQU16Zrm:
295 case X86::VMOVDQA32Zrm:
296 case X86::VMOVDQU32Zrm:
297 case X86::VMOVDQA64Zrm:
298 case X86::VMOVDQU64Zrm:
299 MemBytes = 64;
300 return true;
301 }
302 }
303
isFrameStoreOpcode(int Opcode,unsigned & MemBytes)304 static bool isFrameStoreOpcode(int Opcode, unsigned &MemBytes) {
305 switch (Opcode) {
306 default:
307 return false;
308 case X86::MOV8mr:
309 case X86::KMOVBmk:
310 MemBytes = 1;
311 return true;
312 case X86::MOV16mr:
313 case X86::KMOVWmk:
314 MemBytes = 2;
315 return true;
316 case X86::MOV32mr:
317 case X86::MOVSSmr:
318 case X86::VMOVSSmr:
319 case X86::VMOVSSZmr:
320 case X86::KMOVDmk:
321 MemBytes = 4;
322 return true;
323 case X86::MOV64mr:
324 case X86::ST_FpP64m:
325 case X86::MOVSDmr:
326 case X86::VMOVSDmr:
327 case X86::VMOVSDZmr:
328 case X86::MMX_MOVD64mr:
329 case X86::MMX_MOVQ64mr:
330 case X86::MMX_MOVNTQmr:
331 case X86::KMOVQmk:
332 MemBytes = 8;
333 return true;
334 case X86::MOVAPSmr:
335 case X86::MOVUPSmr:
336 case X86::MOVAPDmr:
337 case X86::MOVUPDmr:
338 case X86::MOVDQAmr:
339 case X86::MOVDQUmr:
340 case X86::VMOVAPSmr:
341 case X86::VMOVUPSmr:
342 case X86::VMOVAPDmr:
343 case X86::VMOVUPDmr:
344 case X86::VMOVDQAmr:
345 case X86::VMOVDQUmr:
346 case X86::VMOVUPSZ128mr:
347 case X86::VMOVAPSZ128mr:
348 case X86::VMOVUPSZ128mr_NOVLX:
349 case X86::VMOVAPSZ128mr_NOVLX:
350 case X86::VMOVUPDZ128mr:
351 case X86::VMOVAPDZ128mr:
352 case X86::VMOVDQA32Z128mr:
353 case X86::VMOVDQU32Z128mr:
354 case X86::VMOVDQA64Z128mr:
355 case X86::VMOVDQU64Z128mr:
356 case X86::VMOVDQU8Z128mr:
357 case X86::VMOVDQU16Z128mr:
358 MemBytes = 16;
359 return true;
360 case X86::VMOVUPSYmr:
361 case X86::VMOVAPSYmr:
362 case X86::VMOVUPDYmr:
363 case X86::VMOVAPDYmr:
364 case X86::VMOVDQUYmr:
365 case X86::VMOVDQAYmr:
366 case X86::VMOVUPSZ256mr:
367 case X86::VMOVAPSZ256mr:
368 case X86::VMOVUPSZ256mr_NOVLX:
369 case X86::VMOVAPSZ256mr_NOVLX:
370 case X86::VMOVUPDZ256mr:
371 case X86::VMOVAPDZ256mr:
372 case X86::VMOVDQU8Z256mr:
373 case X86::VMOVDQU16Z256mr:
374 case X86::VMOVDQA32Z256mr:
375 case X86::VMOVDQU32Z256mr:
376 case X86::VMOVDQA64Z256mr:
377 case X86::VMOVDQU64Z256mr:
378 MemBytes = 32;
379 return true;
380 case X86::VMOVUPSZmr:
381 case X86::VMOVAPSZmr:
382 case X86::VMOVUPDZmr:
383 case X86::VMOVAPDZmr:
384 case X86::VMOVDQU8Zmr:
385 case X86::VMOVDQU16Zmr:
386 case X86::VMOVDQA32Zmr:
387 case X86::VMOVDQU32Zmr:
388 case X86::VMOVDQA64Zmr:
389 case X86::VMOVDQU64Zmr:
390 MemBytes = 64;
391 return true;
392 }
393 return false;
394 }
395
isLoadFromStackSlot(const MachineInstr & MI,int & FrameIndex) const396 unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr &MI,
397 int &FrameIndex) const {
398 unsigned Dummy;
399 return X86InstrInfo::isLoadFromStackSlot(MI, FrameIndex, Dummy);
400 }
401
isLoadFromStackSlot(const MachineInstr & MI,int & FrameIndex,unsigned & MemBytes) const402 unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr &MI,
403 int &FrameIndex,
404 unsigned &MemBytes) const {
405 if (isFrameLoadOpcode(MI.getOpcode(), MemBytes))
406 if (MI.getOperand(0).getSubReg() == 0 && isFrameOperand(MI, 1, FrameIndex))
407 return MI.getOperand(0).getReg();
408 return 0;
409 }
410
isLoadFromStackSlotPostFE(const MachineInstr & MI,int & FrameIndex) const411 unsigned X86InstrInfo::isLoadFromStackSlotPostFE(const MachineInstr &MI,
412 int &FrameIndex) const {
413 unsigned Dummy;
414 if (isFrameLoadOpcode(MI.getOpcode(), Dummy)) {
415 unsigned Reg;
416 if ((Reg = isLoadFromStackSlot(MI, FrameIndex)))
417 return Reg;
418 // Check for post-frame index elimination operations
419 SmallVector<const MachineMemOperand *, 1> Accesses;
420 if (hasLoadFromStackSlot(MI, Accesses)) {
421 FrameIndex =
422 cast<FixedStackPseudoSourceValue>(Accesses.front()->getPseudoValue())
423 ->getFrameIndex();
424 return 1;
425 }
426 }
427 return 0;
428 }
429
isStoreToStackSlot(const MachineInstr & MI,int & FrameIndex) const430 unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr &MI,
431 int &FrameIndex) const {
432 unsigned Dummy;
433 return X86InstrInfo::isStoreToStackSlot(MI, FrameIndex, Dummy);
434 }
435
isStoreToStackSlot(const MachineInstr & MI,int & FrameIndex,unsigned & MemBytes) const436 unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr &MI,
437 int &FrameIndex,
438 unsigned &MemBytes) const {
439 if (isFrameStoreOpcode(MI.getOpcode(), MemBytes))
440 if (MI.getOperand(X86::AddrNumOperands).getSubReg() == 0 &&
441 isFrameOperand(MI, 0, FrameIndex))
442 return MI.getOperand(X86::AddrNumOperands).getReg();
443 return 0;
444 }
445
isStoreToStackSlotPostFE(const MachineInstr & MI,int & FrameIndex) const446 unsigned X86InstrInfo::isStoreToStackSlotPostFE(const MachineInstr &MI,
447 int &FrameIndex) const {
448 unsigned Dummy;
449 if (isFrameStoreOpcode(MI.getOpcode(), Dummy)) {
450 unsigned Reg;
451 if ((Reg = isStoreToStackSlot(MI, FrameIndex)))
452 return Reg;
453 // Check for post-frame index elimination operations
454 SmallVector<const MachineMemOperand *, 1> Accesses;
455 if (hasStoreToStackSlot(MI, Accesses)) {
456 FrameIndex =
457 cast<FixedStackPseudoSourceValue>(Accesses.front()->getPseudoValue())
458 ->getFrameIndex();
459 return 1;
460 }
461 }
462 return 0;
463 }
464
465 /// Return true if register is PIC base; i.e.g defined by X86::MOVPC32r.
regIsPICBase(unsigned BaseReg,const MachineRegisterInfo & MRI)466 static bool regIsPICBase(unsigned BaseReg, const MachineRegisterInfo &MRI) {
467 // Don't waste compile time scanning use-def chains of physregs.
468 if (!Register::isVirtualRegister(BaseReg))
469 return false;
470 bool isPICBase = false;
471 for (MachineRegisterInfo::def_instr_iterator I = MRI.def_instr_begin(BaseReg),
472 E = MRI.def_instr_end(); I != E; ++I) {
473 MachineInstr *DefMI = &*I;
474 if (DefMI->getOpcode() != X86::MOVPC32r)
475 return false;
476 assert(!isPICBase && "More than one PIC base?");
477 isPICBase = true;
478 }
479 return isPICBase;
480 }
481
isReallyTriviallyReMaterializable(const MachineInstr & MI,AAResults * AA) const482 bool X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr &MI,
483 AAResults *AA) const {
484 switch (MI.getOpcode()) {
485 default:
486 // This function should only be called for opcodes with the ReMaterializable
487 // flag set.
488 llvm_unreachable("Unknown rematerializable operation!");
489 break;
490
491 case X86::LOAD_STACK_GUARD:
492 case X86::AVX1_SETALLONES:
493 case X86::AVX2_SETALLONES:
494 case X86::AVX512_128_SET0:
495 case X86::AVX512_256_SET0:
496 case X86::AVX512_512_SET0:
497 case X86::AVX512_512_SETALLONES:
498 case X86::AVX512_FsFLD0SD:
499 case X86::AVX512_FsFLD0SS:
500 case X86::AVX512_FsFLD0F128:
501 case X86::AVX_SET0:
502 case X86::FsFLD0SD:
503 case X86::FsFLD0SS:
504 case X86::FsFLD0F128:
505 case X86::KSET0D:
506 case X86::KSET0Q:
507 case X86::KSET0W:
508 case X86::KSET1D:
509 case X86::KSET1Q:
510 case X86::KSET1W:
511 case X86::MMX_SET0:
512 case X86::MOV32ImmSExti8:
513 case X86::MOV32r0:
514 case X86::MOV32r1:
515 case X86::MOV32r_1:
516 case X86::MOV32ri64:
517 case X86::MOV64ImmSExti8:
518 case X86::V_SET0:
519 case X86::V_SETALLONES:
520 case X86::MOV16ri:
521 case X86::MOV32ri:
522 case X86::MOV64ri:
523 case X86::MOV64ri32:
524 case X86::MOV8ri:
525 return true;
526
527 case X86::MOV8rm:
528 case X86::MOV8rm_NOREX:
529 case X86::MOV16rm:
530 case X86::MOV32rm:
531 case X86::MOV64rm:
532 case X86::MOVSSrm:
533 case X86::MOVSSrm_alt:
534 case X86::MOVSDrm:
535 case X86::MOVSDrm_alt:
536 case X86::MOVAPSrm:
537 case X86::MOVUPSrm:
538 case X86::MOVAPDrm:
539 case X86::MOVUPDrm:
540 case X86::MOVDQArm:
541 case X86::MOVDQUrm:
542 case X86::VMOVSSrm:
543 case X86::VMOVSSrm_alt:
544 case X86::VMOVSDrm:
545 case X86::VMOVSDrm_alt:
546 case X86::VMOVAPSrm:
547 case X86::VMOVUPSrm:
548 case X86::VMOVAPDrm:
549 case X86::VMOVUPDrm:
550 case X86::VMOVDQArm:
551 case X86::VMOVDQUrm:
552 case X86::VMOVAPSYrm:
553 case X86::VMOVUPSYrm:
554 case X86::VMOVAPDYrm:
555 case X86::VMOVUPDYrm:
556 case X86::VMOVDQAYrm:
557 case X86::VMOVDQUYrm:
558 case X86::MMX_MOVD64rm:
559 case X86::MMX_MOVQ64rm:
560 // AVX-512
561 case X86::VMOVSSZrm:
562 case X86::VMOVSSZrm_alt:
563 case X86::VMOVSDZrm:
564 case X86::VMOVSDZrm_alt:
565 case X86::VMOVAPDZ128rm:
566 case X86::VMOVAPDZ256rm:
567 case X86::VMOVAPDZrm:
568 case X86::VMOVAPSZ128rm:
569 case X86::VMOVAPSZ256rm:
570 case X86::VMOVAPSZ128rm_NOVLX:
571 case X86::VMOVAPSZ256rm_NOVLX:
572 case X86::VMOVAPSZrm:
573 case X86::VMOVDQA32Z128rm:
574 case X86::VMOVDQA32Z256rm:
575 case X86::VMOVDQA32Zrm:
576 case X86::VMOVDQA64Z128rm:
577 case X86::VMOVDQA64Z256rm:
578 case X86::VMOVDQA64Zrm:
579 case X86::VMOVDQU16Z128rm:
580 case X86::VMOVDQU16Z256rm:
581 case X86::VMOVDQU16Zrm:
582 case X86::VMOVDQU32Z128rm:
583 case X86::VMOVDQU32Z256rm:
584 case X86::VMOVDQU32Zrm:
585 case X86::VMOVDQU64Z128rm:
586 case X86::VMOVDQU64Z256rm:
587 case X86::VMOVDQU64Zrm:
588 case X86::VMOVDQU8Z128rm:
589 case X86::VMOVDQU8Z256rm:
590 case X86::VMOVDQU8Zrm:
591 case X86::VMOVUPDZ128rm:
592 case X86::VMOVUPDZ256rm:
593 case X86::VMOVUPDZrm:
594 case X86::VMOVUPSZ128rm:
595 case X86::VMOVUPSZ256rm:
596 case X86::VMOVUPSZ128rm_NOVLX:
597 case X86::VMOVUPSZ256rm_NOVLX:
598 case X86::VMOVUPSZrm: {
599 // Loads from constant pools are trivially rematerializable.
600 if (MI.getOperand(1 + X86::AddrBaseReg).isReg() &&
601 MI.getOperand(1 + X86::AddrScaleAmt).isImm() &&
602 MI.getOperand(1 + X86::AddrIndexReg).isReg() &&
603 MI.getOperand(1 + X86::AddrIndexReg).getReg() == 0 &&
604 MI.isDereferenceableInvariantLoad(AA)) {
605 Register BaseReg = MI.getOperand(1 + X86::AddrBaseReg).getReg();
606 if (BaseReg == 0 || BaseReg == X86::RIP)
607 return true;
608 // Allow re-materialization of PIC load.
609 if (!ReMatPICStubLoad && MI.getOperand(1 + X86::AddrDisp).isGlobal())
610 return false;
611 const MachineFunction &MF = *MI.getParent()->getParent();
612 const MachineRegisterInfo &MRI = MF.getRegInfo();
613 return regIsPICBase(BaseReg, MRI);
614 }
615 return false;
616 }
617
618 case X86::LEA32r:
619 case X86::LEA64r: {
620 if (MI.getOperand(1 + X86::AddrScaleAmt).isImm() &&
621 MI.getOperand(1 + X86::AddrIndexReg).isReg() &&
622 MI.getOperand(1 + X86::AddrIndexReg).getReg() == 0 &&
623 !MI.getOperand(1 + X86::AddrDisp).isReg()) {
624 // lea fi#, lea GV, etc. are all rematerializable.
625 if (!MI.getOperand(1 + X86::AddrBaseReg).isReg())
626 return true;
627 Register BaseReg = MI.getOperand(1 + X86::AddrBaseReg).getReg();
628 if (BaseReg == 0)
629 return true;
630 // Allow re-materialization of lea PICBase + x.
631 const MachineFunction &MF = *MI.getParent()->getParent();
632 const MachineRegisterInfo &MRI = MF.getRegInfo();
633 return regIsPICBase(BaseReg, MRI);
634 }
635 return false;
636 }
637 }
638 }
639
reMaterialize(MachineBasicBlock & MBB,MachineBasicBlock::iterator I,unsigned DestReg,unsigned SubIdx,const MachineInstr & Orig,const TargetRegisterInfo & TRI) const640 void X86InstrInfo::reMaterialize(MachineBasicBlock &MBB,
641 MachineBasicBlock::iterator I,
642 unsigned DestReg, unsigned SubIdx,
643 const MachineInstr &Orig,
644 const TargetRegisterInfo &TRI) const {
645 bool ClobbersEFLAGS = Orig.modifiesRegister(X86::EFLAGS, &TRI);
646 if (ClobbersEFLAGS && !isSafeToClobberEFLAGS(MBB, I)) {
647 // The instruction clobbers EFLAGS. Re-materialize as MOV32ri to avoid side
648 // effects.
649 int Value;
650 switch (Orig.getOpcode()) {
651 case X86::MOV32r0: Value = 0; break;
652 case X86::MOV32r1: Value = 1; break;
653 case X86::MOV32r_1: Value = -1; break;
654 default:
655 llvm_unreachable("Unexpected instruction!");
656 }
657
658 const DebugLoc &DL = Orig.getDebugLoc();
659 BuildMI(MBB, I, DL, get(X86::MOV32ri))
660 .add(Orig.getOperand(0))
661 .addImm(Value);
662 } else {
663 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(&Orig);
664 MBB.insert(I, MI);
665 }
666
667 MachineInstr &NewMI = *std::prev(I);
668 NewMI.substituteRegister(Orig.getOperand(0).getReg(), DestReg, SubIdx, TRI);
669 }
670
671 /// True if MI has a condition code def, e.g. EFLAGS, that is not marked dead.
hasLiveCondCodeDef(MachineInstr & MI) const672 bool X86InstrInfo::hasLiveCondCodeDef(MachineInstr &MI) const {
673 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
674 MachineOperand &MO = MI.getOperand(i);
675 if (MO.isReg() && MO.isDef() &&
676 MO.getReg() == X86::EFLAGS && !MO.isDead()) {
677 return true;
678 }
679 }
680 return false;
681 }
682
683 /// Check whether the shift count for a machine operand is non-zero.
getTruncatedShiftCount(const MachineInstr & MI,unsigned ShiftAmtOperandIdx)684 inline static unsigned getTruncatedShiftCount(const MachineInstr &MI,
685 unsigned ShiftAmtOperandIdx) {
686 // The shift count is six bits with the REX.W prefix and five bits without.
687 unsigned ShiftCountMask = (MI.getDesc().TSFlags & X86II::REX_W) ? 63 : 31;
688 unsigned Imm = MI.getOperand(ShiftAmtOperandIdx).getImm();
689 return Imm & ShiftCountMask;
690 }
691
692 /// Check whether the given shift count is appropriate
693 /// can be represented by a LEA instruction.
isTruncatedShiftCountForLEA(unsigned ShAmt)694 inline static bool isTruncatedShiftCountForLEA(unsigned ShAmt) {
695 // Left shift instructions can be transformed into load-effective-address
696 // instructions if we can encode them appropriately.
697 // A LEA instruction utilizes a SIB byte to encode its scale factor.
698 // The SIB.scale field is two bits wide which means that we can encode any
699 // shift amount less than 4.
700 return ShAmt < 4 && ShAmt > 0;
701 }
702
classifyLEAReg(MachineInstr & MI,const MachineOperand & Src,unsigned Opc,bool AllowSP,Register & NewSrc,bool & isKill,MachineOperand & ImplicitOp,LiveVariables * LV) const703 bool X86InstrInfo::classifyLEAReg(MachineInstr &MI, const MachineOperand &Src,
704 unsigned Opc, bool AllowSP, Register &NewSrc,
705 bool &isKill, MachineOperand &ImplicitOp,
706 LiveVariables *LV) const {
707 MachineFunction &MF = *MI.getParent()->getParent();
708 const TargetRegisterClass *RC;
709 if (AllowSP) {
710 RC = Opc != X86::LEA32r ? &X86::GR64RegClass : &X86::GR32RegClass;
711 } else {
712 RC = Opc != X86::LEA32r ?
713 &X86::GR64_NOSPRegClass : &X86::GR32_NOSPRegClass;
714 }
715 Register SrcReg = Src.getReg();
716
717 // For both LEA64 and LEA32 the register already has essentially the right
718 // type (32-bit or 64-bit) we may just need to forbid SP.
719 if (Opc != X86::LEA64_32r) {
720 NewSrc = SrcReg;
721 isKill = Src.isKill();
722 assert(!Src.isUndef() && "Undef op doesn't need optimization");
723
724 if (Register::isVirtualRegister(NewSrc) &&
725 !MF.getRegInfo().constrainRegClass(NewSrc, RC))
726 return false;
727
728 return true;
729 }
730
731 // This is for an LEA64_32r and incoming registers are 32-bit. One way or
732 // another we need to add 64-bit registers to the final MI.
733 if (Register::isPhysicalRegister(SrcReg)) {
734 ImplicitOp = Src;
735 ImplicitOp.setImplicit();
736
737 NewSrc = getX86SubSuperRegister(Src.getReg(), 64);
738 isKill = Src.isKill();
739 assert(!Src.isUndef() && "Undef op doesn't need optimization");
740 } else {
741 // Virtual register of the wrong class, we have to create a temporary 64-bit
742 // vreg to feed into the LEA.
743 NewSrc = MF.getRegInfo().createVirtualRegister(RC);
744 MachineInstr *Copy =
745 BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), get(TargetOpcode::COPY))
746 .addReg(NewSrc, RegState::Define | RegState::Undef, X86::sub_32bit)
747 .add(Src);
748
749 // Which is obviously going to be dead after we're done with it.
750 isKill = true;
751
752 if (LV)
753 LV->replaceKillInstruction(SrcReg, MI, *Copy);
754 }
755
756 // We've set all the parameters without issue.
757 return true;
758 }
759
convertToThreeAddressWithLEA(unsigned MIOpc,MachineFunction::iterator & MFI,MachineInstr & MI,LiveVariables * LV,bool Is8BitOp) const760 MachineInstr *X86InstrInfo::convertToThreeAddressWithLEA(
761 unsigned MIOpc, MachineFunction::iterator &MFI, MachineInstr &MI,
762 LiveVariables *LV, bool Is8BitOp) const {
763 // We handle 8-bit adds and various 16-bit opcodes in the switch below.
764 MachineRegisterInfo &RegInfo = MFI->getParent()->getRegInfo();
765 assert((Is8BitOp || RegInfo.getTargetRegisterInfo()->getRegSizeInBits(
766 *RegInfo.getRegClass(MI.getOperand(0).getReg())) == 16) &&
767 "Unexpected type for LEA transform");
768
769 // TODO: For a 32-bit target, we need to adjust the LEA variables with
770 // something like this:
771 // Opcode = X86::LEA32r;
772 // InRegLEA = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
773 // OutRegLEA =
774 // Is8BitOp ? RegInfo.createVirtualRegister(&X86::GR32ABCD_RegClass)
775 // : RegInfo.createVirtualRegister(&X86::GR32RegClass);
776 if (!Subtarget.is64Bit())
777 return nullptr;
778
779 unsigned Opcode = X86::LEA64_32r;
780 Register InRegLEA = RegInfo.createVirtualRegister(&X86::GR64_NOSPRegClass);
781 Register OutRegLEA = RegInfo.createVirtualRegister(&X86::GR32RegClass);
782
783 // Build and insert into an implicit UNDEF value. This is OK because
784 // we will be shifting and then extracting the lower 8/16-bits.
785 // This has the potential to cause partial register stall. e.g.
786 // movw (%rbp,%rcx,2), %dx
787 // leal -65(%rdx), %esi
788 // But testing has shown this *does* help performance in 64-bit mode (at
789 // least on modern x86 machines).
790 MachineBasicBlock::iterator MBBI = MI.getIterator();
791 Register Dest = MI.getOperand(0).getReg();
792 Register Src = MI.getOperand(1).getReg();
793 bool IsDead = MI.getOperand(0).isDead();
794 bool IsKill = MI.getOperand(1).isKill();
795 unsigned SubReg = Is8BitOp ? X86::sub_8bit : X86::sub_16bit;
796 assert(!MI.getOperand(1).isUndef() && "Undef op doesn't need optimization");
797 BuildMI(*MFI, MBBI, MI.getDebugLoc(), get(X86::IMPLICIT_DEF), InRegLEA);
798 MachineInstr *InsMI =
799 BuildMI(*MFI, MBBI, MI.getDebugLoc(), get(TargetOpcode::COPY))
800 .addReg(InRegLEA, RegState::Define, SubReg)
801 .addReg(Src, getKillRegState(IsKill));
802
803 MachineInstrBuilder MIB =
804 BuildMI(*MFI, MBBI, MI.getDebugLoc(), get(Opcode), OutRegLEA);
805 switch (MIOpc) {
806 default: llvm_unreachable("Unreachable!");
807 case X86::SHL8ri:
808 case X86::SHL16ri: {
809 unsigned ShAmt = MI.getOperand(2).getImm();
810 MIB.addReg(0).addImm(1ULL << ShAmt)
811 .addReg(InRegLEA, RegState::Kill).addImm(0).addReg(0);
812 break;
813 }
814 case X86::INC8r:
815 case X86::INC16r:
816 addRegOffset(MIB, InRegLEA, true, 1);
817 break;
818 case X86::DEC8r:
819 case X86::DEC16r:
820 addRegOffset(MIB, InRegLEA, true, -1);
821 break;
822 case X86::ADD8ri:
823 case X86::ADD8ri_DB:
824 case X86::ADD16ri:
825 case X86::ADD16ri8:
826 case X86::ADD16ri_DB:
827 case X86::ADD16ri8_DB:
828 addRegOffset(MIB, InRegLEA, true, MI.getOperand(2).getImm());
829 break;
830 case X86::ADD8rr:
831 case X86::ADD8rr_DB:
832 case X86::ADD16rr:
833 case X86::ADD16rr_DB: {
834 Register Src2 = MI.getOperand(2).getReg();
835 bool IsKill2 = MI.getOperand(2).isKill();
836 assert(!MI.getOperand(2).isUndef() && "Undef op doesn't need optimization");
837 unsigned InRegLEA2 = 0;
838 MachineInstr *InsMI2 = nullptr;
839 if (Src == Src2) {
840 // ADD8rr/ADD16rr killed %reg1028, %reg1028
841 // just a single insert_subreg.
842 addRegReg(MIB, InRegLEA, true, InRegLEA, false);
843 } else {
844 if (Subtarget.is64Bit())
845 InRegLEA2 = RegInfo.createVirtualRegister(&X86::GR64_NOSPRegClass);
846 else
847 InRegLEA2 = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
848 // Build and insert into an implicit UNDEF value. This is OK because
849 // we will be shifting and then extracting the lower 8/16-bits.
850 BuildMI(*MFI, &*MIB, MI.getDebugLoc(), get(X86::IMPLICIT_DEF), InRegLEA2);
851 InsMI2 = BuildMI(*MFI, &*MIB, MI.getDebugLoc(), get(TargetOpcode::COPY))
852 .addReg(InRegLEA2, RegState::Define, SubReg)
853 .addReg(Src2, getKillRegState(IsKill2));
854 addRegReg(MIB, InRegLEA, true, InRegLEA2, true);
855 }
856 if (LV && IsKill2 && InsMI2)
857 LV->replaceKillInstruction(Src2, MI, *InsMI2);
858 break;
859 }
860 }
861
862 MachineInstr *NewMI = MIB;
863 MachineInstr *ExtMI =
864 BuildMI(*MFI, MBBI, MI.getDebugLoc(), get(TargetOpcode::COPY))
865 .addReg(Dest, RegState::Define | getDeadRegState(IsDead))
866 .addReg(OutRegLEA, RegState::Kill, SubReg);
867
868 if (LV) {
869 // Update live variables.
870 LV->getVarInfo(InRegLEA).Kills.push_back(NewMI);
871 LV->getVarInfo(OutRegLEA).Kills.push_back(ExtMI);
872 if (IsKill)
873 LV->replaceKillInstruction(Src, MI, *InsMI);
874 if (IsDead)
875 LV->replaceKillInstruction(Dest, MI, *ExtMI);
876 }
877
878 return ExtMI;
879 }
880
881 /// This method must be implemented by targets that
882 /// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
883 /// may be able to convert a two-address instruction into a true
884 /// three-address instruction on demand. This allows the X86 target (for
885 /// example) to convert ADD and SHL instructions into LEA instructions if they
886 /// would require register copies due to two-addressness.
887 ///
888 /// This method returns a null pointer if the transformation cannot be
889 /// performed, otherwise it returns the new instruction.
890 ///
891 MachineInstr *
convertToThreeAddress(MachineFunction::iterator & MFI,MachineInstr & MI,LiveVariables * LV) const892 X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
893 MachineInstr &MI, LiveVariables *LV) const {
894 // The following opcodes also sets the condition code register(s). Only
895 // convert them to equivalent lea if the condition code register def's
896 // are dead!
897 if (hasLiveCondCodeDef(MI))
898 return nullptr;
899
900 MachineFunction &MF = *MI.getParent()->getParent();
901 // All instructions input are two-addr instructions. Get the known operands.
902 const MachineOperand &Dest = MI.getOperand(0);
903 const MachineOperand &Src = MI.getOperand(1);
904
905 // Ideally, operations with undef should be folded before we get here, but we
906 // can't guarantee it. Bail out because optimizing undefs is a waste of time.
907 // Without this, we have to forward undef state to new register operands to
908 // avoid machine verifier errors.
909 if (Src.isUndef())
910 return nullptr;
911 if (MI.getNumOperands() > 2)
912 if (MI.getOperand(2).isReg() && MI.getOperand(2).isUndef())
913 return nullptr;
914
915 MachineInstr *NewMI = nullptr;
916 bool Is64Bit = Subtarget.is64Bit();
917
918 bool Is8BitOp = false;
919 unsigned MIOpc = MI.getOpcode();
920 switch (MIOpc) {
921 default: llvm_unreachable("Unreachable!");
922 case X86::SHL64ri: {
923 assert(MI.getNumOperands() >= 3 && "Unknown shift instruction!");
924 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
925 if (!isTruncatedShiftCountForLEA(ShAmt)) return nullptr;
926
927 // LEA can't handle RSP.
928 if (Register::isVirtualRegister(Src.getReg()) &&
929 !MF.getRegInfo().constrainRegClass(Src.getReg(),
930 &X86::GR64_NOSPRegClass))
931 return nullptr;
932
933 NewMI = BuildMI(MF, MI.getDebugLoc(), get(X86::LEA64r))
934 .add(Dest)
935 .addReg(0)
936 .addImm(1ULL << ShAmt)
937 .add(Src)
938 .addImm(0)
939 .addReg(0);
940 break;
941 }
942 case X86::SHL32ri: {
943 assert(MI.getNumOperands() >= 3 && "Unknown shift instruction!");
944 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
945 if (!isTruncatedShiftCountForLEA(ShAmt)) return nullptr;
946
947 unsigned Opc = Is64Bit ? X86::LEA64_32r : X86::LEA32r;
948
949 // LEA can't handle ESP.
950 bool isKill;
951 Register SrcReg;
952 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
953 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false,
954 SrcReg, isKill, ImplicitOp, LV))
955 return nullptr;
956
957 MachineInstrBuilder MIB =
958 BuildMI(MF, MI.getDebugLoc(), get(Opc))
959 .add(Dest)
960 .addReg(0)
961 .addImm(1ULL << ShAmt)
962 .addReg(SrcReg, getKillRegState(isKill))
963 .addImm(0)
964 .addReg(0);
965 if (ImplicitOp.getReg() != 0)
966 MIB.add(ImplicitOp);
967 NewMI = MIB;
968
969 break;
970 }
971 case X86::SHL8ri:
972 Is8BitOp = true;
973 LLVM_FALLTHROUGH;
974 case X86::SHL16ri: {
975 assert(MI.getNumOperands() >= 3 && "Unknown shift instruction!");
976 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
977 if (!isTruncatedShiftCountForLEA(ShAmt))
978 return nullptr;
979 return convertToThreeAddressWithLEA(MIOpc, MFI, MI, LV, Is8BitOp);
980 }
981 case X86::INC64r:
982 case X86::INC32r: {
983 assert(MI.getNumOperands() >= 2 && "Unknown inc instruction!");
984 unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r :
985 (Is64Bit ? X86::LEA64_32r : X86::LEA32r);
986 bool isKill;
987 Register SrcReg;
988 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
989 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false, SrcReg, isKill,
990 ImplicitOp, LV))
991 return nullptr;
992
993 MachineInstrBuilder MIB =
994 BuildMI(MF, MI.getDebugLoc(), get(Opc))
995 .add(Dest)
996 .addReg(SrcReg, getKillRegState(isKill));
997 if (ImplicitOp.getReg() != 0)
998 MIB.add(ImplicitOp);
999
1000 NewMI = addOffset(MIB, 1);
1001 break;
1002 }
1003 case X86::DEC64r:
1004 case X86::DEC32r: {
1005 assert(MI.getNumOperands() >= 2 && "Unknown dec instruction!");
1006 unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r
1007 : (Is64Bit ? X86::LEA64_32r : X86::LEA32r);
1008
1009 bool isKill;
1010 Register SrcReg;
1011 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
1012 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false, SrcReg, isKill,
1013 ImplicitOp, LV))
1014 return nullptr;
1015
1016 MachineInstrBuilder MIB = BuildMI(MF, MI.getDebugLoc(), get(Opc))
1017 .add(Dest)
1018 .addReg(SrcReg, getKillRegState(isKill));
1019 if (ImplicitOp.getReg() != 0)
1020 MIB.add(ImplicitOp);
1021
1022 NewMI = addOffset(MIB, -1);
1023
1024 break;
1025 }
1026 case X86::DEC8r:
1027 case X86::INC8r:
1028 Is8BitOp = true;
1029 LLVM_FALLTHROUGH;
1030 case X86::DEC16r:
1031 case X86::INC16r:
1032 return convertToThreeAddressWithLEA(MIOpc, MFI, MI, LV, Is8BitOp);
1033 case X86::ADD64rr:
1034 case X86::ADD64rr_DB:
1035 case X86::ADD32rr:
1036 case X86::ADD32rr_DB: {
1037 assert(MI.getNumOperands() >= 3 && "Unknown add instruction!");
1038 unsigned Opc;
1039 if (MIOpc == X86::ADD64rr || MIOpc == X86::ADD64rr_DB)
1040 Opc = X86::LEA64r;
1041 else
1042 Opc = Is64Bit ? X86::LEA64_32r : X86::LEA32r;
1043
1044 bool isKill;
1045 Register SrcReg;
1046 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
1047 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ true,
1048 SrcReg, isKill, ImplicitOp, LV))
1049 return nullptr;
1050
1051 const MachineOperand &Src2 = MI.getOperand(2);
1052 bool isKill2;
1053 Register SrcReg2;
1054 MachineOperand ImplicitOp2 = MachineOperand::CreateReg(0, false);
1055 if (!classifyLEAReg(MI, Src2, Opc, /*AllowSP=*/ false,
1056 SrcReg2, isKill2, ImplicitOp2, LV))
1057 return nullptr;
1058
1059 MachineInstrBuilder MIB = BuildMI(MF, MI.getDebugLoc(), get(Opc)).add(Dest);
1060 if (ImplicitOp.getReg() != 0)
1061 MIB.add(ImplicitOp);
1062 if (ImplicitOp2.getReg() != 0)
1063 MIB.add(ImplicitOp2);
1064
1065 NewMI = addRegReg(MIB, SrcReg, isKill, SrcReg2, isKill2);
1066 if (LV && Src2.isKill())
1067 LV->replaceKillInstruction(SrcReg2, MI, *NewMI);
1068 break;
1069 }
1070 case X86::ADD8rr:
1071 case X86::ADD8rr_DB:
1072 Is8BitOp = true;
1073 LLVM_FALLTHROUGH;
1074 case X86::ADD16rr:
1075 case X86::ADD16rr_DB:
1076 return convertToThreeAddressWithLEA(MIOpc, MFI, MI, LV, Is8BitOp);
1077 case X86::ADD64ri32:
1078 case X86::ADD64ri8:
1079 case X86::ADD64ri32_DB:
1080 case X86::ADD64ri8_DB:
1081 assert(MI.getNumOperands() >= 3 && "Unknown add instruction!");
1082 NewMI = addOffset(
1083 BuildMI(MF, MI.getDebugLoc(), get(X86::LEA64r)).add(Dest).add(Src),
1084 MI.getOperand(2));
1085 break;
1086 case X86::ADD32ri:
1087 case X86::ADD32ri8:
1088 case X86::ADD32ri_DB:
1089 case X86::ADD32ri8_DB: {
1090 assert(MI.getNumOperands() >= 3 && "Unknown add instruction!");
1091 unsigned Opc = Is64Bit ? X86::LEA64_32r : X86::LEA32r;
1092
1093 bool isKill;
1094 Register SrcReg;
1095 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
1096 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ true,
1097 SrcReg, isKill, ImplicitOp, LV))
1098 return nullptr;
1099
1100 MachineInstrBuilder MIB = BuildMI(MF, MI.getDebugLoc(), get(Opc))
1101 .add(Dest)
1102 .addReg(SrcReg, getKillRegState(isKill));
1103 if (ImplicitOp.getReg() != 0)
1104 MIB.add(ImplicitOp);
1105
1106 NewMI = addOffset(MIB, MI.getOperand(2));
1107 break;
1108 }
1109 case X86::ADD8ri:
1110 case X86::ADD8ri_DB:
1111 Is8BitOp = true;
1112 LLVM_FALLTHROUGH;
1113 case X86::ADD16ri:
1114 case X86::ADD16ri8:
1115 case X86::ADD16ri_DB:
1116 case X86::ADD16ri8_DB:
1117 return convertToThreeAddressWithLEA(MIOpc, MFI, MI, LV, Is8BitOp);
1118 case X86::SUB8ri:
1119 case X86::SUB16ri8:
1120 case X86::SUB16ri:
1121 /// FIXME: Support these similar to ADD8ri/ADD16ri*.
1122 return nullptr;
1123 case X86::SUB32ri8:
1124 case X86::SUB32ri: {
1125 if (!MI.getOperand(2).isImm())
1126 return nullptr;
1127 int64_t Imm = MI.getOperand(2).getImm();
1128 if (!isInt<32>(-Imm))
1129 return nullptr;
1130
1131 assert(MI.getNumOperands() >= 3 && "Unknown add instruction!");
1132 unsigned Opc = Is64Bit ? X86::LEA64_32r : X86::LEA32r;
1133
1134 bool isKill;
1135 Register SrcReg;
1136 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
1137 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ true,
1138 SrcReg, isKill, ImplicitOp, LV))
1139 return nullptr;
1140
1141 MachineInstrBuilder MIB = BuildMI(MF, MI.getDebugLoc(), get(Opc))
1142 .add(Dest)
1143 .addReg(SrcReg, getKillRegState(isKill));
1144 if (ImplicitOp.getReg() != 0)
1145 MIB.add(ImplicitOp);
1146
1147 NewMI = addOffset(MIB, -Imm);
1148 break;
1149 }
1150
1151 case X86::SUB64ri8:
1152 case X86::SUB64ri32: {
1153 if (!MI.getOperand(2).isImm())
1154 return nullptr;
1155 int64_t Imm = MI.getOperand(2).getImm();
1156 if (!isInt<32>(-Imm))
1157 return nullptr;
1158
1159 assert(MI.getNumOperands() >= 3 && "Unknown sub instruction!");
1160
1161 MachineInstrBuilder MIB = BuildMI(MF, MI.getDebugLoc(),
1162 get(X86::LEA64r)).add(Dest).add(Src);
1163 NewMI = addOffset(MIB, -Imm);
1164 break;
1165 }
1166
1167 case X86::VMOVDQU8Z128rmk:
1168 case X86::VMOVDQU8Z256rmk:
1169 case X86::VMOVDQU8Zrmk:
1170 case X86::VMOVDQU16Z128rmk:
1171 case X86::VMOVDQU16Z256rmk:
1172 case X86::VMOVDQU16Zrmk:
1173 case X86::VMOVDQU32Z128rmk: case X86::VMOVDQA32Z128rmk:
1174 case X86::VMOVDQU32Z256rmk: case X86::VMOVDQA32Z256rmk:
1175 case X86::VMOVDQU32Zrmk: case X86::VMOVDQA32Zrmk:
1176 case X86::VMOVDQU64Z128rmk: case X86::VMOVDQA64Z128rmk:
1177 case X86::VMOVDQU64Z256rmk: case X86::VMOVDQA64Z256rmk:
1178 case X86::VMOVDQU64Zrmk: case X86::VMOVDQA64Zrmk:
1179 case X86::VMOVUPDZ128rmk: case X86::VMOVAPDZ128rmk:
1180 case X86::VMOVUPDZ256rmk: case X86::VMOVAPDZ256rmk:
1181 case X86::VMOVUPDZrmk: case X86::VMOVAPDZrmk:
1182 case X86::VMOVUPSZ128rmk: case X86::VMOVAPSZ128rmk:
1183 case X86::VMOVUPSZ256rmk: case X86::VMOVAPSZ256rmk:
1184 case X86::VMOVUPSZrmk: case X86::VMOVAPSZrmk:
1185 case X86::VBROADCASTSDZ256mk:
1186 case X86::VBROADCASTSDZmk:
1187 case X86::VBROADCASTSSZ128mk:
1188 case X86::VBROADCASTSSZ256mk:
1189 case X86::VBROADCASTSSZmk:
1190 case X86::VPBROADCASTDZ128mk:
1191 case X86::VPBROADCASTDZ256mk:
1192 case X86::VPBROADCASTDZmk:
1193 case X86::VPBROADCASTQZ128mk:
1194 case X86::VPBROADCASTQZ256mk:
1195 case X86::VPBROADCASTQZmk: {
1196 unsigned Opc;
1197 switch (MIOpc) {
1198 default: llvm_unreachable("Unreachable!");
1199 case X86::VMOVDQU8Z128rmk: Opc = X86::VPBLENDMBZ128rmk; break;
1200 case X86::VMOVDQU8Z256rmk: Opc = X86::VPBLENDMBZ256rmk; break;
1201 case X86::VMOVDQU8Zrmk: Opc = X86::VPBLENDMBZrmk; break;
1202 case X86::VMOVDQU16Z128rmk: Opc = X86::VPBLENDMWZ128rmk; break;
1203 case X86::VMOVDQU16Z256rmk: Opc = X86::VPBLENDMWZ256rmk; break;
1204 case X86::VMOVDQU16Zrmk: Opc = X86::VPBLENDMWZrmk; break;
1205 case X86::VMOVDQU32Z128rmk: Opc = X86::VPBLENDMDZ128rmk; break;
1206 case X86::VMOVDQU32Z256rmk: Opc = X86::VPBLENDMDZ256rmk; break;
1207 case X86::VMOVDQU32Zrmk: Opc = X86::VPBLENDMDZrmk; break;
1208 case X86::VMOVDQU64Z128rmk: Opc = X86::VPBLENDMQZ128rmk; break;
1209 case X86::VMOVDQU64Z256rmk: Opc = X86::VPBLENDMQZ256rmk; break;
1210 case X86::VMOVDQU64Zrmk: Opc = X86::VPBLENDMQZrmk; break;
1211 case X86::VMOVUPDZ128rmk: Opc = X86::VBLENDMPDZ128rmk; break;
1212 case X86::VMOVUPDZ256rmk: Opc = X86::VBLENDMPDZ256rmk; break;
1213 case X86::VMOVUPDZrmk: Opc = X86::VBLENDMPDZrmk; break;
1214 case X86::VMOVUPSZ128rmk: Opc = X86::VBLENDMPSZ128rmk; break;
1215 case X86::VMOVUPSZ256rmk: Opc = X86::VBLENDMPSZ256rmk; break;
1216 case X86::VMOVUPSZrmk: Opc = X86::VBLENDMPSZrmk; break;
1217 case X86::VMOVDQA32Z128rmk: Opc = X86::VPBLENDMDZ128rmk; break;
1218 case X86::VMOVDQA32Z256rmk: Opc = X86::VPBLENDMDZ256rmk; break;
1219 case X86::VMOVDQA32Zrmk: Opc = X86::VPBLENDMDZrmk; break;
1220 case X86::VMOVDQA64Z128rmk: Opc = X86::VPBLENDMQZ128rmk; break;
1221 case X86::VMOVDQA64Z256rmk: Opc = X86::VPBLENDMQZ256rmk; break;
1222 case X86::VMOVDQA64Zrmk: Opc = X86::VPBLENDMQZrmk; break;
1223 case X86::VMOVAPDZ128rmk: Opc = X86::VBLENDMPDZ128rmk; break;
1224 case X86::VMOVAPDZ256rmk: Opc = X86::VBLENDMPDZ256rmk; break;
1225 case X86::VMOVAPDZrmk: Opc = X86::VBLENDMPDZrmk; break;
1226 case X86::VMOVAPSZ128rmk: Opc = X86::VBLENDMPSZ128rmk; break;
1227 case X86::VMOVAPSZ256rmk: Opc = X86::VBLENDMPSZ256rmk; break;
1228 case X86::VMOVAPSZrmk: Opc = X86::VBLENDMPSZrmk; break;
1229 case X86::VBROADCASTSDZ256mk: Opc = X86::VBLENDMPDZ256rmbk; break;
1230 case X86::VBROADCASTSDZmk: Opc = X86::VBLENDMPDZrmbk; break;
1231 case X86::VBROADCASTSSZ128mk: Opc = X86::VBLENDMPSZ128rmbk; break;
1232 case X86::VBROADCASTSSZ256mk: Opc = X86::VBLENDMPSZ256rmbk; break;
1233 case X86::VBROADCASTSSZmk: Opc = X86::VBLENDMPSZrmbk; break;
1234 case X86::VPBROADCASTDZ128mk: Opc = X86::VPBLENDMDZ128rmbk; break;
1235 case X86::VPBROADCASTDZ256mk: Opc = X86::VPBLENDMDZ256rmbk; break;
1236 case X86::VPBROADCASTDZmk: Opc = X86::VPBLENDMDZrmbk; break;
1237 case X86::VPBROADCASTQZ128mk: Opc = X86::VPBLENDMQZ128rmbk; break;
1238 case X86::VPBROADCASTQZ256mk: Opc = X86::VPBLENDMQZ256rmbk; break;
1239 case X86::VPBROADCASTQZmk: Opc = X86::VPBLENDMQZrmbk; break;
1240 }
1241
1242 NewMI = BuildMI(MF, MI.getDebugLoc(), get(Opc))
1243 .add(Dest)
1244 .add(MI.getOperand(2))
1245 .add(Src)
1246 .add(MI.getOperand(3))
1247 .add(MI.getOperand(4))
1248 .add(MI.getOperand(5))
1249 .add(MI.getOperand(6))
1250 .add(MI.getOperand(7));
1251 break;
1252 }
1253
1254 case X86::VMOVDQU8Z128rrk:
1255 case X86::VMOVDQU8Z256rrk:
1256 case X86::VMOVDQU8Zrrk:
1257 case X86::VMOVDQU16Z128rrk:
1258 case X86::VMOVDQU16Z256rrk:
1259 case X86::VMOVDQU16Zrrk:
1260 case X86::VMOVDQU32Z128rrk: case X86::VMOVDQA32Z128rrk:
1261 case X86::VMOVDQU32Z256rrk: case X86::VMOVDQA32Z256rrk:
1262 case X86::VMOVDQU32Zrrk: case X86::VMOVDQA32Zrrk:
1263 case X86::VMOVDQU64Z128rrk: case X86::VMOVDQA64Z128rrk:
1264 case X86::VMOVDQU64Z256rrk: case X86::VMOVDQA64Z256rrk:
1265 case X86::VMOVDQU64Zrrk: case X86::VMOVDQA64Zrrk:
1266 case X86::VMOVUPDZ128rrk: case X86::VMOVAPDZ128rrk:
1267 case X86::VMOVUPDZ256rrk: case X86::VMOVAPDZ256rrk:
1268 case X86::VMOVUPDZrrk: case X86::VMOVAPDZrrk:
1269 case X86::VMOVUPSZ128rrk: case X86::VMOVAPSZ128rrk:
1270 case X86::VMOVUPSZ256rrk: case X86::VMOVAPSZ256rrk:
1271 case X86::VMOVUPSZrrk: case X86::VMOVAPSZrrk: {
1272 unsigned Opc;
1273 switch (MIOpc) {
1274 default: llvm_unreachable("Unreachable!");
1275 case X86::VMOVDQU8Z128rrk: Opc = X86::VPBLENDMBZ128rrk; break;
1276 case X86::VMOVDQU8Z256rrk: Opc = X86::VPBLENDMBZ256rrk; break;
1277 case X86::VMOVDQU8Zrrk: Opc = X86::VPBLENDMBZrrk; break;
1278 case X86::VMOVDQU16Z128rrk: Opc = X86::VPBLENDMWZ128rrk; break;
1279 case X86::VMOVDQU16Z256rrk: Opc = X86::VPBLENDMWZ256rrk; break;
1280 case X86::VMOVDQU16Zrrk: Opc = X86::VPBLENDMWZrrk; break;
1281 case X86::VMOVDQU32Z128rrk: Opc = X86::VPBLENDMDZ128rrk; break;
1282 case X86::VMOVDQU32Z256rrk: Opc = X86::VPBLENDMDZ256rrk; break;
1283 case X86::VMOVDQU32Zrrk: Opc = X86::VPBLENDMDZrrk; break;
1284 case X86::VMOVDQU64Z128rrk: Opc = X86::VPBLENDMQZ128rrk; break;
1285 case X86::VMOVDQU64Z256rrk: Opc = X86::VPBLENDMQZ256rrk; break;
1286 case X86::VMOVDQU64Zrrk: Opc = X86::VPBLENDMQZrrk; break;
1287 case X86::VMOVUPDZ128rrk: Opc = X86::VBLENDMPDZ128rrk; break;
1288 case X86::VMOVUPDZ256rrk: Opc = X86::VBLENDMPDZ256rrk; break;
1289 case X86::VMOVUPDZrrk: Opc = X86::VBLENDMPDZrrk; break;
1290 case X86::VMOVUPSZ128rrk: Opc = X86::VBLENDMPSZ128rrk; break;
1291 case X86::VMOVUPSZ256rrk: Opc = X86::VBLENDMPSZ256rrk; break;
1292 case X86::VMOVUPSZrrk: Opc = X86::VBLENDMPSZrrk; break;
1293 case X86::VMOVDQA32Z128rrk: Opc = X86::VPBLENDMDZ128rrk; break;
1294 case X86::VMOVDQA32Z256rrk: Opc = X86::VPBLENDMDZ256rrk; break;
1295 case X86::VMOVDQA32Zrrk: Opc = X86::VPBLENDMDZrrk; break;
1296 case X86::VMOVDQA64Z128rrk: Opc = X86::VPBLENDMQZ128rrk; break;
1297 case X86::VMOVDQA64Z256rrk: Opc = X86::VPBLENDMQZ256rrk; break;
1298 case X86::VMOVDQA64Zrrk: Opc = X86::VPBLENDMQZrrk; break;
1299 case X86::VMOVAPDZ128rrk: Opc = X86::VBLENDMPDZ128rrk; break;
1300 case X86::VMOVAPDZ256rrk: Opc = X86::VBLENDMPDZ256rrk; break;
1301 case X86::VMOVAPDZrrk: Opc = X86::VBLENDMPDZrrk; break;
1302 case X86::VMOVAPSZ128rrk: Opc = X86::VBLENDMPSZ128rrk; break;
1303 case X86::VMOVAPSZ256rrk: Opc = X86::VBLENDMPSZ256rrk; break;
1304 case X86::VMOVAPSZrrk: Opc = X86::VBLENDMPSZrrk; break;
1305 }
1306
1307 NewMI = BuildMI(MF, MI.getDebugLoc(), get(Opc))
1308 .add(Dest)
1309 .add(MI.getOperand(2))
1310 .add(Src)
1311 .add(MI.getOperand(3));
1312 break;
1313 }
1314 }
1315
1316 if (!NewMI) return nullptr;
1317
1318 if (LV) { // Update live variables
1319 if (Src.isKill())
1320 LV->replaceKillInstruction(Src.getReg(), MI, *NewMI);
1321 if (Dest.isDead())
1322 LV->replaceKillInstruction(Dest.getReg(), MI, *NewMI);
1323 }
1324
1325 MFI->insert(MI.getIterator(), NewMI); // Insert the new inst
1326 return NewMI;
1327 }
1328
1329 /// This determines which of three possible cases of a three source commute
1330 /// the source indexes correspond to taking into account any mask operands.
1331 /// All prevents commuting a passthru operand. Returns -1 if the commute isn't
1332 /// possible.
1333 /// Case 0 - Possible to commute the first and second operands.
1334 /// Case 1 - Possible to commute the first and third operands.
1335 /// Case 2 - Possible to commute the second and third operands.
getThreeSrcCommuteCase(uint64_t TSFlags,unsigned SrcOpIdx1,unsigned SrcOpIdx2)1336 static unsigned getThreeSrcCommuteCase(uint64_t TSFlags, unsigned SrcOpIdx1,
1337 unsigned SrcOpIdx2) {
1338 // Put the lowest index to SrcOpIdx1 to simplify the checks below.
1339 if (SrcOpIdx1 > SrcOpIdx2)
1340 std::swap(SrcOpIdx1, SrcOpIdx2);
1341
1342 unsigned Op1 = 1, Op2 = 2, Op3 = 3;
1343 if (X86II::isKMasked(TSFlags)) {
1344 Op2++;
1345 Op3++;
1346 }
1347
1348 if (SrcOpIdx1 == Op1 && SrcOpIdx2 == Op2)
1349 return 0;
1350 if (SrcOpIdx1 == Op1 && SrcOpIdx2 == Op3)
1351 return 1;
1352 if (SrcOpIdx1 == Op2 && SrcOpIdx2 == Op3)
1353 return 2;
1354 llvm_unreachable("Unknown three src commute case.");
1355 }
1356
getFMA3OpcodeToCommuteOperands(const MachineInstr & MI,unsigned SrcOpIdx1,unsigned SrcOpIdx2,const X86InstrFMA3Group & FMA3Group) const1357 unsigned X86InstrInfo::getFMA3OpcodeToCommuteOperands(
1358 const MachineInstr &MI, unsigned SrcOpIdx1, unsigned SrcOpIdx2,
1359 const X86InstrFMA3Group &FMA3Group) const {
1360
1361 unsigned Opc = MI.getOpcode();
1362
1363 // TODO: Commuting the 1st operand of FMA*_Int requires some additional
1364 // analysis. The commute optimization is legal only if all users of FMA*_Int
1365 // use only the lowest element of the FMA*_Int instruction. Such analysis are
1366 // not implemented yet. So, just return 0 in that case.
1367 // When such analysis are available this place will be the right place for
1368 // calling it.
1369 assert(!(FMA3Group.isIntrinsic() && (SrcOpIdx1 == 1 || SrcOpIdx2 == 1)) &&
1370 "Intrinsic instructions can't commute operand 1");
1371
1372 // Determine which case this commute is or if it can't be done.
1373 unsigned Case = getThreeSrcCommuteCase(MI.getDesc().TSFlags, SrcOpIdx1,
1374 SrcOpIdx2);
1375 assert(Case < 3 && "Unexpected case number!");
1376
1377 // Define the FMA forms mapping array that helps to map input FMA form
1378 // to output FMA form to preserve the operation semantics after
1379 // commuting the operands.
1380 const unsigned Form132Index = 0;
1381 const unsigned Form213Index = 1;
1382 const unsigned Form231Index = 2;
1383 static const unsigned FormMapping[][3] = {
1384 // 0: SrcOpIdx1 == 1 && SrcOpIdx2 == 2;
1385 // FMA132 A, C, b; ==> FMA231 C, A, b;
1386 // FMA213 B, A, c; ==> FMA213 A, B, c;
1387 // FMA231 C, A, b; ==> FMA132 A, C, b;
1388 { Form231Index, Form213Index, Form132Index },
1389 // 1: SrcOpIdx1 == 1 && SrcOpIdx2 == 3;
1390 // FMA132 A, c, B; ==> FMA132 B, c, A;
1391 // FMA213 B, a, C; ==> FMA231 C, a, B;
1392 // FMA231 C, a, B; ==> FMA213 B, a, C;
1393 { Form132Index, Form231Index, Form213Index },
1394 // 2: SrcOpIdx1 == 2 && SrcOpIdx2 == 3;
1395 // FMA132 a, C, B; ==> FMA213 a, B, C;
1396 // FMA213 b, A, C; ==> FMA132 b, C, A;
1397 // FMA231 c, A, B; ==> FMA231 c, B, A;
1398 { Form213Index, Form132Index, Form231Index }
1399 };
1400
1401 unsigned FMAForms[3];
1402 FMAForms[0] = FMA3Group.get132Opcode();
1403 FMAForms[1] = FMA3Group.get213Opcode();
1404 FMAForms[2] = FMA3Group.get231Opcode();
1405 unsigned FormIndex;
1406 for (FormIndex = 0; FormIndex < 3; FormIndex++)
1407 if (Opc == FMAForms[FormIndex])
1408 break;
1409
1410 // Everything is ready, just adjust the FMA opcode and return it.
1411 FormIndex = FormMapping[Case][FormIndex];
1412 return FMAForms[FormIndex];
1413 }
1414
commuteVPTERNLOG(MachineInstr & MI,unsigned SrcOpIdx1,unsigned SrcOpIdx2)1415 static void commuteVPTERNLOG(MachineInstr &MI, unsigned SrcOpIdx1,
1416 unsigned SrcOpIdx2) {
1417 // Determine which case this commute is or if it can't be done.
1418 unsigned Case = getThreeSrcCommuteCase(MI.getDesc().TSFlags, SrcOpIdx1,
1419 SrcOpIdx2);
1420 assert(Case < 3 && "Unexpected case value!");
1421
1422 // For each case we need to swap two pairs of bits in the final immediate.
1423 static const uint8_t SwapMasks[3][4] = {
1424 { 0x04, 0x10, 0x08, 0x20 }, // Swap bits 2/4 and 3/5.
1425 { 0x02, 0x10, 0x08, 0x40 }, // Swap bits 1/4 and 3/6.
1426 { 0x02, 0x04, 0x20, 0x40 }, // Swap bits 1/2 and 5/6.
1427 };
1428
1429 uint8_t Imm = MI.getOperand(MI.getNumOperands()-1).getImm();
1430 // Clear out the bits we are swapping.
1431 uint8_t NewImm = Imm & ~(SwapMasks[Case][0] | SwapMasks[Case][1] |
1432 SwapMasks[Case][2] | SwapMasks[Case][3]);
1433 // If the immediate had a bit of the pair set, then set the opposite bit.
1434 if (Imm & SwapMasks[Case][0]) NewImm |= SwapMasks[Case][1];
1435 if (Imm & SwapMasks[Case][1]) NewImm |= SwapMasks[Case][0];
1436 if (Imm & SwapMasks[Case][2]) NewImm |= SwapMasks[Case][3];
1437 if (Imm & SwapMasks[Case][3]) NewImm |= SwapMasks[Case][2];
1438 MI.getOperand(MI.getNumOperands()-1).setImm(NewImm);
1439 }
1440
1441 // Returns true if this is a VPERMI2 or VPERMT2 instruction that can be
1442 // commuted.
isCommutableVPERMV3Instruction(unsigned Opcode)1443 static bool isCommutableVPERMV3Instruction(unsigned Opcode) {
1444 #define VPERM_CASES(Suffix) \
1445 case X86::VPERMI2##Suffix##128rr: case X86::VPERMT2##Suffix##128rr: \
1446 case X86::VPERMI2##Suffix##256rr: case X86::VPERMT2##Suffix##256rr: \
1447 case X86::VPERMI2##Suffix##rr: case X86::VPERMT2##Suffix##rr: \
1448 case X86::VPERMI2##Suffix##128rm: case X86::VPERMT2##Suffix##128rm: \
1449 case X86::VPERMI2##Suffix##256rm: case X86::VPERMT2##Suffix##256rm: \
1450 case X86::VPERMI2##Suffix##rm: case X86::VPERMT2##Suffix##rm: \
1451 case X86::VPERMI2##Suffix##128rrkz: case X86::VPERMT2##Suffix##128rrkz: \
1452 case X86::VPERMI2##Suffix##256rrkz: case X86::VPERMT2##Suffix##256rrkz: \
1453 case X86::VPERMI2##Suffix##rrkz: case X86::VPERMT2##Suffix##rrkz: \
1454 case X86::VPERMI2##Suffix##128rmkz: case X86::VPERMT2##Suffix##128rmkz: \
1455 case X86::VPERMI2##Suffix##256rmkz: case X86::VPERMT2##Suffix##256rmkz: \
1456 case X86::VPERMI2##Suffix##rmkz: case X86::VPERMT2##Suffix##rmkz:
1457
1458 #define VPERM_CASES_BROADCAST(Suffix) \
1459 VPERM_CASES(Suffix) \
1460 case X86::VPERMI2##Suffix##128rmb: case X86::VPERMT2##Suffix##128rmb: \
1461 case X86::VPERMI2##Suffix##256rmb: case X86::VPERMT2##Suffix##256rmb: \
1462 case X86::VPERMI2##Suffix##rmb: case X86::VPERMT2##Suffix##rmb: \
1463 case X86::VPERMI2##Suffix##128rmbkz: case X86::VPERMT2##Suffix##128rmbkz: \
1464 case X86::VPERMI2##Suffix##256rmbkz: case X86::VPERMT2##Suffix##256rmbkz: \
1465 case X86::VPERMI2##Suffix##rmbkz: case X86::VPERMT2##Suffix##rmbkz:
1466
1467 switch (Opcode) {
1468 default: return false;
1469 VPERM_CASES(B)
1470 VPERM_CASES_BROADCAST(D)
1471 VPERM_CASES_BROADCAST(PD)
1472 VPERM_CASES_BROADCAST(PS)
1473 VPERM_CASES_BROADCAST(Q)
1474 VPERM_CASES(W)
1475 return true;
1476 }
1477 #undef VPERM_CASES_BROADCAST
1478 #undef VPERM_CASES
1479 }
1480
1481 // Returns commuted opcode for VPERMI2 and VPERMT2 instructions by switching
1482 // from the I opcode to the T opcode and vice versa.
getCommutedVPERMV3Opcode(unsigned Opcode)1483 static unsigned getCommutedVPERMV3Opcode(unsigned Opcode) {
1484 #define VPERM_CASES(Orig, New) \
1485 case X86::Orig##128rr: return X86::New##128rr; \
1486 case X86::Orig##128rrkz: return X86::New##128rrkz; \
1487 case X86::Orig##128rm: return X86::New##128rm; \
1488 case X86::Orig##128rmkz: return X86::New##128rmkz; \
1489 case X86::Orig##256rr: return X86::New##256rr; \
1490 case X86::Orig##256rrkz: return X86::New##256rrkz; \
1491 case X86::Orig##256rm: return X86::New##256rm; \
1492 case X86::Orig##256rmkz: return X86::New##256rmkz; \
1493 case X86::Orig##rr: return X86::New##rr; \
1494 case X86::Orig##rrkz: return X86::New##rrkz; \
1495 case X86::Orig##rm: return X86::New##rm; \
1496 case X86::Orig##rmkz: return X86::New##rmkz;
1497
1498 #define VPERM_CASES_BROADCAST(Orig, New) \
1499 VPERM_CASES(Orig, New) \
1500 case X86::Orig##128rmb: return X86::New##128rmb; \
1501 case X86::Orig##128rmbkz: return X86::New##128rmbkz; \
1502 case X86::Orig##256rmb: return X86::New##256rmb; \
1503 case X86::Orig##256rmbkz: return X86::New##256rmbkz; \
1504 case X86::Orig##rmb: return X86::New##rmb; \
1505 case X86::Orig##rmbkz: return X86::New##rmbkz;
1506
1507 switch (Opcode) {
1508 VPERM_CASES(VPERMI2B, VPERMT2B)
1509 VPERM_CASES_BROADCAST(VPERMI2D, VPERMT2D)
1510 VPERM_CASES_BROADCAST(VPERMI2PD, VPERMT2PD)
1511 VPERM_CASES_BROADCAST(VPERMI2PS, VPERMT2PS)
1512 VPERM_CASES_BROADCAST(VPERMI2Q, VPERMT2Q)
1513 VPERM_CASES(VPERMI2W, VPERMT2W)
1514 VPERM_CASES(VPERMT2B, VPERMI2B)
1515 VPERM_CASES_BROADCAST(VPERMT2D, VPERMI2D)
1516 VPERM_CASES_BROADCAST(VPERMT2PD, VPERMI2PD)
1517 VPERM_CASES_BROADCAST(VPERMT2PS, VPERMI2PS)
1518 VPERM_CASES_BROADCAST(VPERMT2Q, VPERMI2Q)
1519 VPERM_CASES(VPERMT2W, VPERMI2W)
1520 }
1521
1522 llvm_unreachable("Unreachable!");
1523 #undef VPERM_CASES_BROADCAST
1524 #undef VPERM_CASES
1525 }
1526
commuteInstructionImpl(MachineInstr & MI,bool NewMI,unsigned OpIdx1,unsigned OpIdx2) const1527 MachineInstr *X86InstrInfo::commuteInstructionImpl(MachineInstr &MI, bool NewMI,
1528 unsigned OpIdx1,
1529 unsigned OpIdx2) const {
1530 auto cloneIfNew = [NewMI](MachineInstr &MI) -> MachineInstr & {
1531 if (NewMI)
1532 return *MI.getParent()->getParent()->CloneMachineInstr(&MI);
1533 return MI;
1534 };
1535
1536 switch (MI.getOpcode()) {
1537 case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I)
1538 case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I)
1539 case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I)
1540 case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I)
1541 case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I)
1542 case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I)
1543 unsigned Opc;
1544 unsigned Size;
1545 switch (MI.getOpcode()) {
1546 default: llvm_unreachable("Unreachable!");
1547 case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break;
1548 case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break;
1549 case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break;
1550 case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break;
1551 case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break;
1552 case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break;
1553 }
1554 unsigned Amt = MI.getOperand(3).getImm();
1555 auto &WorkingMI = cloneIfNew(MI);
1556 WorkingMI.setDesc(get(Opc));
1557 WorkingMI.getOperand(3).setImm(Size - Amt);
1558 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1559 OpIdx1, OpIdx2);
1560 }
1561 case X86::PFSUBrr:
1562 case X86::PFSUBRrr: {
1563 // PFSUB x, y: x = x - y
1564 // PFSUBR x, y: x = y - x
1565 unsigned Opc =
1566 (X86::PFSUBRrr == MI.getOpcode() ? X86::PFSUBrr : X86::PFSUBRrr);
1567 auto &WorkingMI = cloneIfNew(MI);
1568 WorkingMI.setDesc(get(Opc));
1569 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1570 OpIdx1, OpIdx2);
1571 }
1572 case X86::BLENDPDrri:
1573 case X86::BLENDPSrri:
1574 case X86::VBLENDPDrri:
1575 case X86::VBLENDPSrri:
1576 // If we're optimizing for size, try to use MOVSD/MOVSS.
1577 if (MI.getParent()->getParent()->getFunction().hasOptSize()) {
1578 unsigned Mask, Opc;
1579 switch (MI.getOpcode()) {
1580 default: llvm_unreachable("Unreachable!");
1581 case X86::BLENDPDrri: Opc = X86::MOVSDrr; Mask = 0x03; break;
1582 case X86::BLENDPSrri: Opc = X86::MOVSSrr; Mask = 0x0F; break;
1583 case X86::VBLENDPDrri: Opc = X86::VMOVSDrr; Mask = 0x03; break;
1584 case X86::VBLENDPSrri: Opc = X86::VMOVSSrr; Mask = 0x0F; break;
1585 }
1586 if ((MI.getOperand(3).getImm() ^ Mask) == 1) {
1587 auto &WorkingMI = cloneIfNew(MI);
1588 WorkingMI.setDesc(get(Opc));
1589 WorkingMI.RemoveOperand(3);
1590 return TargetInstrInfo::commuteInstructionImpl(WorkingMI,
1591 /*NewMI=*/false,
1592 OpIdx1, OpIdx2);
1593 }
1594 }
1595 LLVM_FALLTHROUGH;
1596 case X86::PBLENDWrri:
1597 case X86::VBLENDPDYrri:
1598 case X86::VBLENDPSYrri:
1599 case X86::VPBLENDDrri:
1600 case X86::VPBLENDWrri:
1601 case X86::VPBLENDDYrri:
1602 case X86::VPBLENDWYrri:{
1603 int8_t Mask;
1604 switch (MI.getOpcode()) {
1605 default: llvm_unreachable("Unreachable!");
1606 case X86::BLENDPDrri: Mask = (int8_t)0x03; break;
1607 case X86::BLENDPSrri: Mask = (int8_t)0x0F; break;
1608 case X86::PBLENDWrri: Mask = (int8_t)0xFF; break;
1609 case X86::VBLENDPDrri: Mask = (int8_t)0x03; break;
1610 case X86::VBLENDPSrri: Mask = (int8_t)0x0F; break;
1611 case X86::VBLENDPDYrri: Mask = (int8_t)0x0F; break;
1612 case X86::VBLENDPSYrri: Mask = (int8_t)0xFF; break;
1613 case X86::VPBLENDDrri: Mask = (int8_t)0x0F; break;
1614 case X86::VPBLENDWrri: Mask = (int8_t)0xFF; break;
1615 case X86::VPBLENDDYrri: Mask = (int8_t)0xFF; break;
1616 case X86::VPBLENDWYrri: Mask = (int8_t)0xFF; break;
1617 }
1618 // Only the least significant bits of Imm are used.
1619 // Using int8_t to ensure it will be sign extended to the int64_t that
1620 // setImm takes in order to match isel behavior.
1621 int8_t Imm = MI.getOperand(3).getImm() & Mask;
1622 auto &WorkingMI = cloneIfNew(MI);
1623 WorkingMI.getOperand(3).setImm(Mask ^ Imm);
1624 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1625 OpIdx1, OpIdx2);
1626 }
1627 case X86::INSERTPSrr:
1628 case X86::VINSERTPSrr:
1629 case X86::VINSERTPSZrr: {
1630 unsigned Imm = MI.getOperand(MI.getNumOperands() - 1).getImm();
1631 unsigned ZMask = Imm & 15;
1632 unsigned DstIdx = (Imm >> 4) & 3;
1633 unsigned SrcIdx = (Imm >> 6) & 3;
1634
1635 // We can commute insertps if we zero 2 of the elements, the insertion is
1636 // "inline" and we don't override the insertion with a zero.
1637 if (DstIdx == SrcIdx && (ZMask & (1 << DstIdx)) == 0 &&
1638 countPopulation(ZMask) == 2) {
1639 unsigned AltIdx = findFirstSet((ZMask | (1 << DstIdx)) ^ 15);
1640 assert(AltIdx < 4 && "Illegal insertion index");
1641 unsigned AltImm = (AltIdx << 6) | (AltIdx << 4) | ZMask;
1642 auto &WorkingMI = cloneIfNew(MI);
1643 WorkingMI.getOperand(MI.getNumOperands() - 1).setImm(AltImm);
1644 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1645 OpIdx1, OpIdx2);
1646 }
1647 return nullptr;
1648 }
1649 case X86::MOVSDrr:
1650 case X86::MOVSSrr:
1651 case X86::VMOVSDrr:
1652 case X86::VMOVSSrr:{
1653 // On SSE41 or later we can commute a MOVSS/MOVSD to a BLENDPS/BLENDPD.
1654 if (Subtarget.hasSSE41()) {
1655 unsigned Mask, Opc;
1656 switch (MI.getOpcode()) {
1657 default: llvm_unreachable("Unreachable!");
1658 case X86::MOVSDrr: Opc = X86::BLENDPDrri; Mask = 0x02; break;
1659 case X86::MOVSSrr: Opc = X86::BLENDPSrri; Mask = 0x0E; break;
1660 case X86::VMOVSDrr: Opc = X86::VBLENDPDrri; Mask = 0x02; break;
1661 case X86::VMOVSSrr: Opc = X86::VBLENDPSrri; Mask = 0x0E; break;
1662 }
1663
1664 auto &WorkingMI = cloneIfNew(MI);
1665 WorkingMI.setDesc(get(Opc));
1666 WorkingMI.addOperand(MachineOperand::CreateImm(Mask));
1667 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1668 OpIdx1, OpIdx2);
1669 }
1670
1671 // Convert to SHUFPD.
1672 assert(MI.getOpcode() == X86::MOVSDrr &&
1673 "Can only commute MOVSDrr without SSE4.1");
1674
1675 auto &WorkingMI = cloneIfNew(MI);
1676 WorkingMI.setDesc(get(X86::SHUFPDrri));
1677 WorkingMI.addOperand(MachineOperand::CreateImm(0x02));
1678 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1679 OpIdx1, OpIdx2);
1680 }
1681 case X86::SHUFPDrri: {
1682 // Commute to MOVSD.
1683 assert(MI.getOperand(3).getImm() == 0x02 && "Unexpected immediate!");
1684 auto &WorkingMI = cloneIfNew(MI);
1685 WorkingMI.setDesc(get(X86::MOVSDrr));
1686 WorkingMI.RemoveOperand(3);
1687 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1688 OpIdx1, OpIdx2);
1689 }
1690 case X86::PCLMULQDQrr:
1691 case X86::VPCLMULQDQrr:
1692 case X86::VPCLMULQDQYrr:
1693 case X86::VPCLMULQDQZrr:
1694 case X86::VPCLMULQDQZ128rr:
1695 case X86::VPCLMULQDQZ256rr: {
1696 // SRC1 64bits = Imm[0] ? SRC1[127:64] : SRC1[63:0]
1697 // SRC2 64bits = Imm[4] ? SRC2[127:64] : SRC2[63:0]
1698 unsigned Imm = MI.getOperand(3).getImm();
1699 unsigned Src1Hi = Imm & 0x01;
1700 unsigned Src2Hi = Imm & 0x10;
1701 auto &WorkingMI = cloneIfNew(MI);
1702 WorkingMI.getOperand(3).setImm((Src1Hi << 4) | (Src2Hi >> 4));
1703 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1704 OpIdx1, OpIdx2);
1705 }
1706 case X86::VPCMPBZ128rri: case X86::VPCMPUBZ128rri:
1707 case X86::VPCMPBZ256rri: case X86::VPCMPUBZ256rri:
1708 case X86::VPCMPBZrri: case X86::VPCMPUBZrri:
1709 case X86::VPCMPDZ128rri: case X86::VPCMPUDZ128rri:
1710 case X86::VPCMPDZ256rri: case X86::VPCMPUDZ256rri:
1711 case X86::VPCMPDZrri: case X86::VPCMPUDZrri:
1712 case X86::VPCMPQZ128rri: case X86::VPCMPUQZ128rri:
1713 case X86::VPCMPQZ256rri: case X86::VPCMPUQZ256rri:
1714 case X86::VPCMPQZrri: case X86::VPCMPUQZrri:
1715 case X86::VPCMPWZ128rri: case X86::VPCMPUWZ128rri:
1716 case X86::VPCMPWZ256rri: case X86::VPCMPUWZ256rri:
1717 case X86::VPCMPWZrri: case X86::VPCMPUWZrri:
1718 case X86::VPCMPBZ128rrik: case X86::VPCMPUBZ128rrik:
1719 case X86::VPCMPBZ256rrik: case X86::VPCMPUBZ256rrik:
1720 case X86::VPCMPBZrrik: case X86::VPCMPUBZrrik:
1721 case X86::VPCMPDZ128rrik: case X86::VPCMPUDZ128rrik:
1722 case X86::VPCMPDZ256rrik: case X86::VPCMPUDZ256rrik:
1723 case X86::VPCMPDZrrik: case X86::VPCMPUDZrrik:
1724 case X86::VPCMPQZ128rrik: case X86::VPCMPUQZ128rrik:
1725 case X86::VPCMPQZ256rrik: case X86::VPCMPUQZ256rrik:
1726 case X86::VPCMPQZrrik: case X86::VPCMPUQZrrik:
1727 case X86::VPCMPWZ128rrik: case X86::VPCMPUWZ128rrik:
1728 case X86::VPCMPWZ256rrik: case X86::VPCMPUWZ256rrik:
1729 case X86::VPCMPWZrrik: case X86::VPCMPUWZrrik: {
1730 // Flip comparison mode immediate (if necessary).
1731 unsigned Imm = MI.getOperand(MI.getNumOperands() - 1).getImm() & 0x7;
1732 Imm = X86::getSwappedVPCMPImm(Imm);
1733 auto &WorkingMI = cloneIfNew(MI);
1734 WorkingMI.getOperand(MI.getNumOperands() - 1).setImm(Imm);
1735 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1736 OpIdx1, OpIdx2);
1737 }
1738 case X86::VPCOMBri: case X86::VPCOMUBri:
1739 case X86::VPCOMDri: case X86::VPCOMUDri:
1740 case X86::VPCOMQri: case X86::VPCOMUQri:
1741 case X86::VPCOMWri: case X86::VPCOMUWri: {
1742 // Flip comparison mode immediate (if necessary).
1743 unsigned Imm = MI.getOperand(3).getImm() & 0x7;
1744 Imm = X86::getSwappedVPCOMImm(Imm);
1745 auto &WorkingMI = cloneIfNew(MI);
1746 WorkingMI.getOperand(3).setImm(Imm);
1747 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1748 OpIdx1, OpIdx2);
1749 }
1750 case X86::VCMPSDZrr:
1751 case X86::VCMPSSZrr:
1752 case X86::VCMPPDZrri:
1753 case X86::VCMPPSZrri:
1754 case X86::VCMPPDZ128rri:
1755 case X86::VCMPPSZ128rri:
1756 case X86::VCMPPDZ256rri:
1757 case X86::VCMPPSZ256rri:
1758 case X86::VCMPPDZrrik:
1759 case X86::VCMPPSZrrik:
1760 case X86::VCMPPDZ128rrik:
1761 case X86::VCMPPSZ128rrik:
1762 case X86::VCMPPDZ256rrik:
1763 case X86::VCMPPSZ256rrik: {
1764 unsigned Imm =
1765 MI.getOperand(MI.getNumExplicitOperands() - 1).getImm() & 0x1f;
1766 Imm = X86::getSwappedVCMPImm(Imm);
1767 auto &WorkingMI = cloneIfNew(MI);
1768 WorkingMI.getOperand(MI.getNumExplicitOperands() - 1).setImm(Imm);
1769 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1770 OpIdx1, OpIdx2);
1771 }
1772 case X86::VPERM2F128rr:
1773 case X86::VPERM2I128rr: {
1774 // Flip permute source immediate.
1775 // Imm & 0x02: lo = if set, select Op1.lo/hi else Op0.lo/hi.
1776 // Imm & 0x20: hi = if set, select Op1.lo/hi else Op0.lo/hi.
1777 int8_t Imm = MI.getOperand(3).getImm() & 0xFF;
1778 auto &WorkingMI = cloneIfNew(MI);
1779 WorkingMI.getOperand(3).setImm(Imm ^ 0x22);
1780 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1781 OpIdx1, OpIdx2);
1782 }
1783 case X86::MOVHLPSrr:
1784 case X86::UNPCKHPDrr:
1785 case X86::VMOVHLPSrr:
1786 case X86::VUNPCKHPDrr:
1787 case X86::VMOVHLPSZrr:
1788 case X86::VUNPCKHPDZ128rr: {
1789 assert(Subtarget.hasSSE2() && "Commuting MOVHLP/UNPCKHPD requires SSE2!");
1790
1791 unsigned Opc = MI.getOpcode();
1792 switch (Opc) {
1793 default: llvm_unreachable("Unreachable!");
1794 case X86::MOVHLPSrr: Opc = X86::UNPCKHPDrr; break;
1795 case X86::UNPCKHPDrr: Opc = X86::MOVHLPSrr; break;
1796 case X86::VMOVHLPSrr: Opc = X86::VUNPCKHPDrr; break;
1797 case X86::VUNPCKHPDrr: Opc = X86::VMOVHLPSrr; break;
1798 case X86::VMOVHLPSZrr: Opc = X86::VUNPCKHPDZ128rr; break;
1799 case X86::VUNPCKHPDZ128rr: Opc = X86::VMOVHLPSZrr; break;
1800 }
1801 auto &WorkingMI = cloneIfNew(MI);
1802 WorkingMI.setDesc(get(Opc));
1803 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1804 OpIdx1, OpIdx2);
1805 }
1806 case X86::CMOV16rr: case X86::CMOV32rr: case X86::CMOV64rr: {
1807 auto &WorkingMI = cloneIfNew(MI);
1808 unsigned OpNo = MI.getDesc().getNumOperands() - 1;
1809 X86::CondCode CC = static_cast<X86::CondCode>(MI.getOperand(OpNo).getImm());
1810 WorkingMI.getOperand(OpNo).setImm(X86::GetOppositeBranchCondition(CC));
1811 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1812 OpIdx1, OpIdx2);
1813 }
1814 case X86::VPTERNLOGDZrri: case X86::VPTERNLOGDZrmi:
1815 case X86::VPTERNLOGDZ128rri: case X86::VPTERNLOGDZ128rmi:
1816 case X86::VPTERNLOGDZ256rri: case X86::VPTERNLOGDZ256rmi:
1817 case X86::VPTERNLOGQZrri: case X86::VPTERNLOGQZrmi:
1818 case X86::VPTERNLOGQZ128rri: case X86::VPTERNLOGQZ128rmi:
1819 case X86::VPTERNLOGQZ256rri: case X86::VPTERNLOGQZ256rmi:
1820 case X86::VPTERNLOGDZrrik:
1821 case X86::VPTERNLOGDZ128rrik:
1822 case X86::VPTERNLOGDZ256rrik:
1823 case X86::VPTERNLOGQZrrik:
1824 case X86::VPTERNLOGQZ128rrik:
1825 case X86::VPTERNLOGQZ256rrik:
1826 case X86::VPTERNLOGDZrrikz: case X86::VPTERNLOGDZrmikz:
1827 case X86::VPTERNLOGDZ128rrikz: case X86::VPTERNLOGDZ128rmikz:
1828 case X86::VPTERNLOGDZ256rrikz: case X86::VPTERNLOGDZ256rmikz:
1829 case X86::VPTERNLOGQZrrikz: case X86::VPTERNLOGQZrmikz:
1830 case X86::VPTERNLOGQZ128rrikz: case X86::VPTERNLOGQZ128rmikz:
1831 case X86::VPTERNLOGQZ256rrikz: case X86::VPTERNLOGQZ256rmikz:
1832 case X86::VPTERNLOGDZ128rmbi:
1833 case X86::VPTERNLOGDZ256rmbi:
1834 case X86::VPTERNLOGDZrmbi:
1835 case X86::VPTERNLOGQZ128rmbi:
1836 case X86::VPTERNLOGQZ256rmbi:
1837 case X86::VPTERNLOGQZrmbi:
1838 case X86::VPTERNLOGDZ128rmbikz:
1839 case X86::VPTERNLOGDZ256rmbikz:
1840 case X86::VPTERNLOGDZrmbikz:
1841 case X86::VPTERNLOGQZ128rmbikz:
1842 case X86::VPTERNLOGQZ256rmbikz:
1843 case X86::VPTERNLOGQZrmbikz: {
1844 auto &WorkingMI = cloneIfNew(MI);
1845 commuteVPTERNLOG(WorkingMI, OpIdx1, OpIdx2);
1846 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1847 OpIdx1, OpIdx2);
1848 }
1849 default: {
1850 if (isCommutableVPERMV3Instruction(MI.getOpcode())) {
1851 unsigned Opc = getCommutedVPERMV3Opcode(MI.getOpcode());
1852 auto &WorkingMI = cloneIfNew(MI);
1853 WorkingMI.setDesc(get(Opc));
1854 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1855 OpIdx1, OpIdx2);
1856 }
1857
1858 const X86InstrFMA3Group *FMA3Group = getFMA3Group(MI.getOpcode(),
1859 MI.getDesc().TSFlags);
1860 if (FMA3Group) {
1861 unsigned Opc =
1862 getFMA3OpcodeToCommuteOperands(MI, OpIdx1, OpIdx2, *FMA3Group);
1863 auto &WorkingMI = cloneIfNew(MI);
1864 WorkingMI.setDesc(get(Opc));
1865 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1866 OpIdx1, OpIdx2);
1867 }
1868
1869 return TargetInstrInfo::commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2);
1870 }
1871 }
1872 }
1873
1874 bool
findThreeSrcCommutedOpIndices(const MachineInstr & MI,unsigned & SrcOpIdx1,unsigned & SrcOpIdx2,bool IsIntrinsic) const1875 X86InstrInfo::findThreeSrcCommutedOpIndices(const MachineInstr &MI,
1876 unsigned &SrcOpIdx1,
1877 unsigned &SrcOpIdx2,
1878 bool IsIntrinsic) const {
1879 uint64_t TSFlags = MI.getDesc().TSFlags;
1880
1881 unsigned FirstCommutableVecOp = 1;
1882 unsigned LastCommutableVecOp = 3;
1883 unsigned KMaskOp = -1U;
1884 if (X86II::isKMasked(TSFlags)) {
1885 // For k-zero-masked operations it is Ok to commute the first vector
1886 // operand.
1887 // For regular k-masked operations a conservative choice is done as the
1888 // elements of the first vector operand, for which the corresponding bit
1889 // in the k-mask operand is set to 0, are copied to the result of the
1890 // instruction.
1891 // TODO/FIXME: The commute still may be legal if it is known that the
1892 // k-mask operand is set to either all ones or all zeroes.
1893 // It is also Ok to commute the 1st operand if all users of MI use only
1894 // the elements enabled by the k-mask operand. For example,
1895 // v4 = VFMADD213PSZrk v1, k, v2, v3; // v1[i] = k[i] ? v2[i]*v1[i]+v3[i]
1896 // : v1[i];
1897 // VMOVAPSZmrk <mem_addr>, k, v4; // this is the ONLY user of v4 ->
1898 // // Ok, to commute v1 in FMADD213PSZrk.
1899
1900 // The k-mask operand has index = 2 for masked and zero-masked operations.
1901 KMaskOp = 2;
1902
1903 // The operand with index = 1 is used as a source for those elements for
1904 // which the corresponding bit in the k-mask is set to 0.
1905 if (X86II::isKMergeMasked(TSFlags))
1906 FirstCommutableVecOp = 3;
1907
1908 LastCommutableVecOp++;
1909 } else if (IsIntrinsic) {
1910 // Commuting the first operand of an intrinsic instruction isn't possible
1911 // unless we can prove that only the lowest element of the result is used.
1912 FirstCommutableVecOp = 2;
1913 }
1914
1915 if (isMem(MI, LastCommutableVecOp))
1916 LastCommutableVecOp--;
1917
1918 // Only the first RegOpsNum operands are commutable.
1919 // Also, the value 'CommuteAnyOperandIndex' is valid here as it means
1920 // that the operand is not specified/fixed.
1921 if (SrcOpIdx1 != CommuteAnyOperandIndex &&
1922 (SrcOpIdx1 < FirstCommutableVecOp || SrcOpIdx1 > LastCommutableVecOp ||
1923 SrcOpIdx1 == KMaskOp))
1924 return false;
1925 if (SrcOpIdx2 != CommuteAnyOperandIndex &&
1926 (SrcOpIdx2 < FirstCommutableVecOp || SrcOpIdx2 > LastCommutableVecOp ||
1927 SrcOpIdx2 == KMaskOp))
1928 return false;
1929
1930 // Look for two different register operands assumed to be commutable
1931 // regardless of the FMA opcode. The FMA opcode is adjusted later.
1932 if (SrcOpIdx1 == CommuteAnyOperandIndex ||
1933 SrcOpIdx2 == CommuteAnyOperandIndex) {
1934 unsigned CommutableOpIdx2 = SrcOpIdx2;
1935
1936 // At least one of operands to be commuted is not specified and
1937 // this method is free to choose appropriate commutable operands.
1938 if (SrcOpIdx1 == SrcOpIdx2)
1939 // Both of operands are not fixed. By default set one of commutable
1940 // operands to the last register operand of the instruction.
1941 CommutableOpIdx2 = LastCommutableVecOp;
1942 else if (SrcOpIdx2 == CommuteAnyOperandIndex)
1943 // Only one of operands is not fixed.
1944 CommutableOpIdx2 = SrcOpIdx1;
1945
1946 // CommutableOpIdx2 is well defined now. Let's choose another commutable
1947 // operand and assign its index to CommutableOpIdx1.
1948 Register Op2Reg = MI.getOperand(CommutableOpIdx2).getReg();
1949
1950 unsigned CommutableOpIdx1;
1951 for (CommutableOpIdx1 = LastCommutableVecOp;
1952 CommutableOpIdx1 >= FirstCommutableVecOp; CommutableOpIdx1--) {
1953 // Just ignore and skip the k-mask operand.
1954 if (CommutableOpIdx1 == KMaskOp)
1955 continue;
1956
1957 // The commuted operands must have different registers.
1958 // Otherwise, the commute transformation does not change anything and
1959 // is useless then.
1960 if (Op2Reg != MI.getOperand(CommutableOpIdx1).getReg())
1961 break;
1962 }
1963
1964 // No appropriate commutable operands were found.
1965 if (CommutableOpIdx1 < FirstCommutableVecOp)
1966 return false;
1967
1968 // Assign the found pair of commutable indices to SrcOpIdx1 and SrcOpidx2
1969 // to return those values.
1970 if (!fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2,
1971 CommutableOpIdx1, CommutableOpIdx2))
1972 return false;
1973 }
1974
1975 return true;
1976 }
1977
findCommutedOpIndices(const MachineInstr & MI,unsigned & SrcOpIdx1,unsigned & SrcOpIdx2) const1978 bool X86InstrInfo::findCommutedOpIndices(const MachineInstr &MI,
1979 unsigned &SrcOpIdx1,
1980 unsigned &SrcOpIdx2) const {
1981 const MCInstrDesc &Desc = MI.getDesc();
1982 if (!Desc.isCommutable())
1983 return false;
1984
1985 switch (MI.getOpcode()) {
1986 case X86::CMPSDrr:
1987 case X86::CMPSSrr:
1988 case X86::CMPPDrri:
1989 case X86::CMPPSrri:
1990 case X86::VCMPSDrr:
1991 case X86::VCMPSSrr:
1992 case X86::VCMPPDrri:
1993 case X86::VCMPPSrri:
1994 case X86::VCMPPDYrri:
1995 case X86::VCMPPSYrri:
1996 case X86::VCMPSDZrr:
1997 case X86::VCMPSSZrr:
1998 case X86::VCMPPDZrri:
1999 case X86::VCMPPSZrri:
2000 case X86::VCMPPDZ128rri:
2001 case X86::VCMPPSZ128rri:
2002 case X86::VCMPPDZ256rri:
2003 case X86::VCMPPSZ256rri:
2004 case X86::VCMPPDZrrik:
2005 case X86::VCMPPSZrrik:
2006 case X86::VCMPPDZ128rrik:
2007 case X86::VCMPPSZ128rrik:
2008 case X86::VCMPPDZ256rrik:
2009 case X86::VCMPPSZ256rrik: {
2010 unsigned OpOffset = X86II::isKMasked(Desc.TSFlags) ? 1 : 0;
2011
2012 // Float comparison can be safely commuted for
2013 // Ordered/Unordered/Equal/NotEqual tests
2014 unsigned Imm = MI.getOperand(3 + OpOffset).getImm() & 0x7;
2015 switch (Imm) {
2016 default:
2017 // EVEX versions can be commuted.
2018 if ((Desc.TSFlags & X86II::EncodingMask) == X86II::EVEX)
2019 break;
2020 return false;
2021 case 0x00: // EQUAL
2022 case 0x03: // UNORDERED
2023 case 0x04: // NOT EQUAL
2024 case 0x07: // ORDERED
2025 break;
2026 }
2027
2028 // The indices of the commutable operands are 1 and 2 (or 2 and 3
2029 // when masked).
2030 // Assign them to the returned operand indices here.
2031 return fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2, 1 + OpOffset,
2032 2 + OpOffset);
2033 }
2034 case X86::MOVSSrr:
2035 // X86::MOVSDrr is always commutable. MOVSS is only commutable if we can
2036 // form sse4.1 blend. We assume VMOVSSrr/VMOVSDrr is always commutable since
2037 // AVX implies sse4.1.
2038 if (Subtarget.hasSSE41())
2039 return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2);
2040 return false;
2041 case X86::SHUFPDrri:
2042 // We can commute this to MOVSD.
2043 if (MI.getOperand(3).getImm() == 0x02)
2044 return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2);
2045 return false;
2046 case X86::MOVHLPSrr:
2047 case X86::UNPCKHPDrr:
2048 case X86::VMOVHLPSrr:
2049 case X86::VUNPCKHPDrr:
2050 case X86::VMOVHLPSZrr:
2051 case X86::VUNPCKHPDZ128rr:
2052 if (Subtarget.hasSSE2())
2053 return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2);
2054 return false;
2055 case X86::VPTERNLOGDZrri: case X86::VPTERNLOGDZrmi:
2056 case X86::VPTERNLOGDZ128rri: case X86::VPTERNLOGDZ128rmi:
2057 case X86::VPTERNLOGDZ256rri: case X86::VPTERNLOGDZ256rmi:
2058 case X86::VPTERNLOGQZrri: case X86::VPTERNLOGQZrmi:
2059 case X86::VPTERNLOGQZ128rri: case X86::VPTERNLOGQZ128rmi:
2060 case X86::VPTERNLOGQZ256rri: case X86::VPTERNLOGQZ256rmi:
2061 case X86::VPTERNLOGDZrrik:
2062 case X86::VPTERNLOGDZ128rrik:
2063 case X86::VPTERNLOGDZ256rrik:
2064 case X86::VPTERNLOGQZrrik:
2065 case X86::VPTERNLOGQZ128rrik:
2066 case X86::VPTERNLOGQZ256rrik:
2067 case X86::VPTERNLOGDZrrikz: case X86::VPTERNLOGDZrmikz:
2068 case X86::VPTERNLOGDZ128rrikz: case X86::VPTERNLOGDZ128rmikz:
2069 case X86::VPTERNLOGDZ256rrikz: case X86::VPTERNLOGDZ256rmikz:
2070 case X86::VPTERNLOGQZrrikz: case X86::VPTERNLOGQZrmikz:
2071 case X86::VPTERNLOGQZ128rrikz: case X86::VPTERNLOGQZ128rmikz:
2072 case X86::VPTERNLOGQZ256rrikz: case X86::VPTERNLOGQZ256rmikz:
2073 case X86::VPTERNLOGDZ128rmbi:
2074 case X86::VPTERNLOGDZ256rmbi:
2075 case X86::VPTERNLOGDZrmbi:
2076 case X86::VPTERNLOGQZ128rmbi:
2077 case X86::VPTERNLOGQZ256rmbi:
2078 case X86::VPTERNLOGQZrmbi:
2079 case X86::VPTERNLOGDZ128rmbikz:
2080 case X86::VPTERNLOGDZ256rmbikz:
2081 case X86::VPTERNLOGDZrmbikz:
2082 case X86::VPTERNLOGQZ128rmbikz:
2083 case X86::VPTERNLOGQZ256rmbikz:
2084 case X86::VPTERNLOGQZrmbikz:
2085 return findThreeSrcCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2);
2086 case X86::VPDPWSSDZ128r:
2087 case X86::VPDPWSSDZ128rk:
2088 case X86::VPDPWSSDZ128rkz:
2089 case X86::VPDPWSSDZ256r:
2090 case X86::VPDPWSSDZ256rk:
2091 case X86::VPDPWSSDZ256rkz:
2092 case X86::VPDPWSSDZr:
2093 case X86::VPDPWSSDZrk:
2094 case X86::VPDPWSSDZrkz:
2095 case X86::VPDPWSSDSZ128r:
2096 case X86::VPDPWSSDSZ128rk:
2097 case X86::VPDPWSSDSZ128rkz:
2098 case X86::VPDPWSSDSZ256r:
2099 case X86::VPDPWSSDSZ256rk:
2100 case X86::VPDPWSSDSZ256rkz:
2101 case X86::VPDPWSSDSZr:
2102 case X86::VPDPWSSDSZrk:
2103 case X86::VPDPWSSDSZrkz:
2104 case X86::VPMADD52HUQZ128r:
2105 case X86::VPMADD52HUQZ128rk:
2106 case X86::VPMADD52HUQZ128rkz:
2107 case X86::VPMADD52HUQZ256r:
2108 case X86::VPMADD52HUQZ256rk:
2109 case X86::VPMADD52HUQZ256rkz:
2110 case X86::VPMADD52HUQZr:
2111 case X86::VPMADD52HUQZrk:
2112 case X86::VPMADD52HUQZrkz:
2113 case X86::VPMADD52LUQZ128r:
2114 case X86::VPMADD52LUQZ128rk:
2115 case X86::VPMADD52LUQZ128rkz:
2116 case X86::VPMADD52LUQZ256r:
2117 case X86::VPMADD52LUQZ256rk:
2118 case X86::VPMADD52LUQZ256rkz:
2119 case X86::VPMADD52LUQZr:
2120 case X86::VPMADD52LUQZrk:
2121 case X86::VPMADD52LUQZrkz: {
2122 unsigned CommutableOpIdx1 = 2;
2123 unsigned CommutableOpIdx2 = 3;
2124 if (X86II::isKMasked(Desc.TSFlags)) {
2125 // Skip the mask register.
2126 ++CommutableOpIdx1;
2127 ++CommutableOpIdx2;
2128 }
2129 if (!fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2,
2130 CommutableOpIdx1, CommutableOpIdx2))
2131 return false;
2132 if (!MI.getOperand(SrcOpIdx1).isReg() ||
2133 !MI.getOperand(SrcOpIdx2).isReg())
2134 // No idea.
2135 return false;
2136 return true;
2137 }
2138
2139 default:
2140 const X86InstrFMA3Group *FMA3Group = getFMA3Group(MI.getOpcode(),
2141 MI.getDesc().TSFlags);
2142 if (FMA3Group)
2143 return findThreeSrcCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2,
2144 FMA3Group->isIntrinsic());
2145
2146 // Handled masked instructions since we need to skip over the mask input
2147 // and the preserved input.
2148 if (X86II::isKMasked(Desc.TSFlags)) {
2149 // First assume that the first input is the mask operand and skip past it.
2150 unsigned CommutableOpIdx1 = Desc.getNumDefs() + 1;
2151 unsigned CommutableOpIdx2 = Desc.getNumDefs() + 2;
2152 // Check if the first input is tied. If there isn't one then we only
2153 // need to skip the mask operand which we did above.
2154 if ((MI.getDesc().getOperandConstraint(Desc.getNumDefs(),
2155 MCOI::TIED_TO) != -1)) {
2156 // If this is zero masking instruction with a tied operand, we need to
2157 // move the first index back to the first input since this must
2158 // be a 3 input instruction and we want the first two non-mask inputs.
2159 // Otherwise this is a 2 input instruction with a preserved input and
2160 // mask, so we need to move the indices to skip one more input.
2161 if (X86II::isKMergeMasked(Desc.TSFlags)) {
2162 ++CommutableOpIdx1;
2163 ++CommutableOpIdx2;
2164 } else {
2165 --CommutableOpIdx1;
2166 }
2167 }
2168
2169 if (!fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2,
2170 CommutableOpIdx1, CommutableOpIdx2))
2171 return false;
2172
2173 if (!MI.getOperand(SrcOpIdx1).isReg() ||
2174 !MI.getOperand(SrcOpIdx2).isReg())
2175 // No idea.
2176 return false;
2177 return true;
2178 }
2179
2180 return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2);
2181 }
2182 return false;
2183 }
2184
getCondFromBranch(const MachineInstr & MI)2185 X86::CondCode X86::getCondFromBranch(const MachineInstr &MI) {
2186 switch (MI.getOpcode()) {
2187 default: return X86::COND_INVALID;
2188 case X86::JCC_1:
2189 return static_cast<X86::CondCode>(
2190 MI.getOperand(MI.getDesc().getNumOperands() - 1).getImm());
2191 }
2192 }
2193
2194 /// Return condition code of a SETCC opcode.
getCondFromSETCC(const MachineInstr & MI)2195 X86::CondCode X86::getCondFromSETCC(const MachineInstr &MI) {
2196 switch (MI.getOpcode()) {
2197 default: return X86::COND_INVALID;
2198 case X86::SETCCr: case X86::SETCCm:
2199 return static_cast<X86::CondCode>(
2200 MI.getOperand(MI.getDesc().getNumOperands() - 1).getImm());
2201 }
2202 }
2203
2204 /// Return condition code of a CMov opcode.
getCondFromCMov(const MachineInstr & MI)2205 X86::CondCode X86::getCondFromCMov(const MachineInstr &MI) {
2206 switch (MI.getOpcode()) {
2207 default: return X86::COND_INVALID;
2208 case X86::CMOV16rr: case X86::CMOV32rr: case X86::CMOV64rr:
2209 case X86::CMOV16rm: case X86::CMOV32rm: case X86::CMOV64rm:
2210 return static_cast<X86::CondCode>(
2211 MI.getOperand(MI.getDesc().getNumOperands() - 1).getImm());
2212 }
2213 }
2214
2215 /// Return the inverse of the specified condition,
2216 /// e.g. turning COND_E to COND_NE.
GetOppositeBranchCondition(X86::CondCode CC)2217 X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) {
2218 switch (CC) {
2219 default: llvm_unreachable("Illegal condition code!");
2220 case X86::COND_E: return X86::COND_NE;
2221 case X86::COND_NE: return X86::COND_E;
2222 case X86::COND_L: return X86::COND_GE;
2223 case X86::COND_LE: return X86::COND_G;
2224 case X86::COND_G: return X86::COND_LE;
2225 case X86::COND_GE: return X86::COND_L;
2226 case X86::COND_B: return X86::COND_AE;
2227 case X86::COND_BE: return X86::COND_A;
2228 case X86::COND_A: return X86::COND_BE;
2229 case X86::COND_AE: return X86::COND_B;
2230 case X86::COND_S: return X86::COND_NS;
2231 case X86::COND_NS: return X86::COND_S;
2232 case X86::COND_P: return X86::COND_NP;
2233 case X86::COND_NP: return X86::COND_P;
2234 case X86::COND_O: return X86::COND_NO;
2235 case X86::COND_NO: return X86::COND_O;
2236 case X86::COND_NE_OR_P: return X86::COND_E_AND_NP;
2237 case X86::COND_E_AND_NP: return X86::COND_NE_OR_P;
2238 }
2239 }
2240
2241 /// Assuming the flags are set by MI(a,b), return the condition code if we
2242 /// modify the instructions such that flags are set by MI(b,a).
getSwappedCondition(X86::CondCode CC)2243 static X86::CondCode getSwappedCondition(X86::CondCode CC) {
2244 switch (CC) {
2245 default: return X86::COND_INVALID;
2246 case X86::COND_E: return X86::COND_E;
2247 case X86::COND_NE: return X86::COND_NE;
2248 case X86::COND_L: return X86::COND_G;
2249 case X86::COND_LE: return X86::COND_GE;
2250 case X86::COND_G: return X86::COND_L;
2251 case X86::COND_GE: return X86::COND_LE;
2252 case X86::COND_B: return X86::COND_A;
2253 case X86::COND_BE: return X86::COND_AE;
2254 case X86::COND_A: return X86::COND_B;
2255 case X86::COND_AE: return X86::COND_BE;
2256 }
2257 }
2258
2259 std::pair<X86::CondCode, bool>
getX86ConditionCode(CmpInst::Predicate Predicate)2260 X86::getX86ConditionCode(CmpInst::Predicate Predicate) {
2261 X86::CondCode CC = X86::COND_INVALID;
2262 bool NeedSwap = false;
2263 switch (Predicate) {
2264 default: break;
2265 // Floating-point Predicates
2266 case CmpInst::FCMP_UEQ: CC = X86::COND_E; break;
2267 case CmpInst::FCMP_OLT: NeedSwap = true; LLVM_FALLTHROUGH;
2268 case CmpInst::FCMP_OGT: CC = X86::COND_A; break;
2269 case CmpInst::FCMP_OLE: NeedSwap = true; LLVM_FALLTHROUGH;
2270 case CmpInst::FCMP_OGE: CC = X86::COND_AE; break;
2271 case CmpInst::FCMP_UGT: NeedSwap = true; LLVM_FALLTHROUGH;
2272 case CmpInst::FCMP_ULT: CC = X86::COND_B; break;
2273 case CmpInst::FCMP_UGE: NeedSwap = true; LLVM_FALLTHROUGH;
2274 case CmpInst::FCMP_ULE: CC = X86::COND_BE; break;
2275 case CmpInst::FCMP_ONE: CC = X86::COND_NE; break;
2276 case CmpInst::FCMP_UNO: CC = X86::COND_P; break;
2277 case CmpInst::FCMP_ORD: CC = X86::COND_NP; break;
2278 case CmpInst::FCMP_OEQ: LLVM_FALLTHROUGH;
2279 case CmpInst::FCMP_UNE: CC = X86::COND_INVALID; break;
2280
2281 // Integer Predicates
2282 case CmpInst::ICMP_EQ: CC = X86::COND_E; break;
2283 case CmpInst::ICMP_NE: CC = X86::COND_NE; break;
2284 case CmpInst::ICMP_UGT: CC = X86::COND_A; break;
2285 case CmpInst::ICMP_UGE: CC = X86::COND_AE; break;
2286 case CmpInst::ICMP_ULT: CC = X86::COND_B; break;
2287 case CmpInst::ICMP_ULE: CC = X86::COND_BE; break;
2288 case CmpInst::ICMP_SGT: CC = X86::COND_G; break;
2289 case CmpInst::ICMP_SGE: CC = X86::COND_GE; break;
2290 case CmpInst::ICMP_SLT: CC = X86::COND_L; break;
2291 case CmpInst::ICMP_SLE: CC = X86::COND_LE; break;
2292 }
2293
2294 return std::make_pair(CC, NeedSwap);
2295 }
2296
2297 /// Return a setcc opcode based on whether it has memory operand.
getSETOpc(bool HasMemoryOperand)2298 unsigned X86::getSETOpc(bool HasMemoryOperand) {
2299 return HasMemoryOperand ? X86::SETCCr : X86::SETCCm;
2300 }
2301
2302 /// Return a cmov opcode for the given register size in bytes, and operand type.
getCMovOpcode(unsigned RegBytes,bool HasMemoryOperand)2303 unsigned X86::getCMovOpcode(unsigned RegBytes, bool HasMemoryOperand) {
2304 switch(RegBytes) {
2305 default: llvm_unreachable("Illegal register size!");
2306 case 2: return HasMemoryOperand ? X86::CMOV16rm : X86::CMOV16rr;
2307 case 4: return HasMemoryOperand ? X86::CMOV32rm : X86::CMOV32rr;
2308 case 8: return HasMemoryOperand ? X86::CMOV64rm : X86::CMOV64rr;
2309 }
2310 }
2311
2312 /// Get the VPCMP immediate for the given condition.
getVPCMPImmForCond(ISD::CondCode CC)2313 unsigned X86::getVPCMPImmForCond(ISD::CondCode CC) {
2314 switch (CC) {
2315 default: llvm_unreachable("Unexpected SETCC condition");
2316 case ISD::SETNE: return 4;
2317 case ISD::SETEQ: return 0;
2318 case ISD::SETULT:
2319 case ISD::SETLT: return 1;
2320 case ISD::SETUGT:
2321 case ISD::SETGT: return 6;
2322 case ISD::SETUGE:
2323 case ISD::SETGE: return 5;
2324 case ISD::SETULE:
2325 case ISD::SETLE: return 2;
2326 }
2327 }
2328
2329 /// Get the VPCMP immediate if the operands are swapped.
getSwappedVPCMPImm(unsigned Imm)2330 unsigned X86::getSwappedVPCMPImm(unsigned Imm) {
2331 switch (Imm) {
2332 default: llvm_unreachable("Unreachable!");
2333 case 0x01: Imm = 0x06; break; // LT -> NLE
2334 case 0x02: Imm = 0x05; break; // LE -> NLT
2335 case 0x05: Imm = 0x02; break; // NLT -> LE
2336 case 0x06: Imm = 0x01; break; // NLE -> LT
2337 case 0x00: // EQ
2338 case 0x03: // FALSE
2339 case 0x04: // NE
2340 case 0x07: // TRUE
2341 break;
2342 }
2343
2344 return Imm;
2345 }
2346
2347 /// Get the VPCOM immediate if the operands are swapped.
getSwappedVPCOMImm(unsigned Imm)2348 unsigned X86::getSwappedVPCOMImm(unsigned Imm) {
2349 switch (Imm) {
2350 default: llvm_unreachable("Unreachable!");
2351 case 0x00: Imm = 0x02; break; // LT -> GT
2352 case 0x01: Imm = 0x03; break; // LE -> GE
2353 case 0x02: Imm = 0x00; break; // GT -> LT
2354 case 0x03: Imm = 0x01; break; // GE -> LE
2355 case 0x04: // EQ
2356 case 0x05: // NE
2357 case 0x06: // FALSE
2358 case 0x07: // TRUE
2359 break;
2360 }
2361
2362 return Imm;
2363 }
2364
2365 /// Get the VCMP immediate if the operands are swapped.
getSwappedVCMPImm(unsigned Imm)2366 unsigned X86::getSwappedVCMPImm(unsigned Imm) {
2367 // Only need the lower 2 bits to distinquish.
2368 switch (Imm & 0x3) {
2369 default: llvm_unreachable("Unreachable!");
2370 case 0x00: case 0x03:
2371 // EQ/NE/TRUE/FALSE/ORD/UNORD don't change immediate when commuted.
2372 break;
2373 case 0x01: case 0x02:
2374 // Need to toggle bits 3:0. Bit 4 stays the same.
2375 Imm ^= 0xf;
2376 break;
2377 }
2378
2379 return Imm;
2380 }
2381
isUnpredicatedTerminator(const MachineInstr & MI) const2382 bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr &MI) const {
2383 if (!MI.isTerminator()) return false;
2384
2385 // Conditional branch is a special case.
2386 if (MI.isBranch() && !MI.isBarrier())
2387 return true;
2388 if (!MI.isPredicable())
2389 return true;
2390 return !isPredicated(MI);
2391 }
2392
isUnconditionalTailCall(const MachineInstr & MI) const2393 bool X86InstrInfo::isUnconditionalTailCall(const MachineInstr &MI) const {
2394 switch (MI.getOpcode()) {
2395 case X86::TCRETURNdi:
2396 case X86::TCRETURNri:
2397 case X86::TCRETURNmi:
2398 case X86::TCRETURNdi64:
2399 case X86::TCRETURNri64:
2400 case X86::TCRETURNmi64:
2401 return true;
2402 default:
2403 return false;
2404 }
2405 }
2406
canMakeTailCallConditional(SmallVectorImpl<MachineOperand> & BranchCond,const MachineInstr & TailCall) const2407 bool X86InstrInfo::canMakeTailCallConditional(
2408 SmallVectorImpl<MachineOperand> &BranchCond,
2409 const MachineInstr &TailCall) const {
2410 if (TailCall.getOpcode() != X86::TCRETURNdi &&
2411 TailCall.getOpcode() != X86::TCRETURNdi64) {
2412 // Only direct calls can be done with a conditional branch.
2413 return false;
2414 }
2415
2416 const MachineFunction *MF = TailCall.getParent()->getParent();
2417 if (Subtarget.isTargetWin64() && MF->hasWinCFI()) {
2418 // Conditional tail calls confuse the Win64 unwinder.
2419 return false;
2420 }
2421
2422 assert(BranchCond.size() == 1);
2423 if (BranchCond[0].getImm() > X86::LAST_VALID_COND) {
2424 // Can't make a conditional tail call with this condition.
2425 return false;
2426 }
2427
2428 const X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
2429 if (X86FI->getTCReturnAddrDelta() != 0 ||
2430 TailCall.getOperand(1).getImm() != 0) {
2431 // A conditional tail call cannot do any stack adjustment.
2432 return false;
2433 }
2434
2435 return true;
2436 }
2437
replaceBranchWithTailCall(MachineBasicBlock & MBB,SmallVectorImpl<MachineOperand> & BranchCond,const MachineInstr & TailCall) const2438 void X86InstrInfo::replaceBranchWithTailCall(
2439 MachineBasicBlock &MBB, SmallVectorImpl<MachineOperand> &BranchCond,
2440 const MachineInstr &TailCall) const {
2441 assert(canMakeTailCallConditional(BranchCond, TailCall));
2442
2443 MachineBasicBlock::iterator I = MBB.end();
2444 while (I != MBB.begin()) {
2445 --I;
2446 if (I->isDebugInstr())
2447 continue;
2448 if (!I->isBranch())
2449 assert(0 && "Can't find the branch to replace!");
2450
2451 X86::CondCode CC = X86::getCondFromBranch(*I);
2452 assert(BranchCond.size() == 1);
2453 if (CC != BranchCond[0].getImm())
2454 continue;
2455
2456 break;
2457 }
2458
2459 unsigned Opc = TailCall.getOpcode() == X86::TCRETURNdi ? X86::TCRETURNdicc
2460 : X86::TCRETURNdi64cc;
2461
2462 auto MIB = BuildMI(MBB, I, MBB.findDebugLoc(I), get(Opc));
2463 MIB->addOperand(TailCall.getOperand(0)); // Destination.
2464 MIB.addImm(0); // Stack offset (not used).
2465 MIB->addOperand(BranchCond[0]); // Condition.
2466 MIB.copyImplicitOps(TailCall); // Regmask and (imp-used) parameters.
2467
2468 // Add implicit uses and defs of all live regs potentially clobbered by the
2469 // call. This way they still appear live across the call.
2470 LivePhysRegs LiveRegs(getRegisterInfo());
2471 LiveRegs.addLiveOuts(MBB);
2472 SmallVector<std::pair<MCPhysReg, const MachineOperand *>, 8> Clobbers;
2473 LiveRegs.stepForward(*MIB, Clobbers);
2474 for (const auto &C : Clobbers) {
2475 MIB.addReg(C.first, RegState::Implicit);
2476 MIB.addReg(C.first, RegState::Implicit | RegState::Define);
2477 }
2478
2479 I->eraseFromParent();
2480 }
2481
2482 // Given a MBB and its TBB, find the FBB which was a fallthrough MBB (it may
2483 // not be a fallthrough MBB now due to layout changes). Return nullptr if the
2484 // fallthrough MBB cannot be identified.
getFallThroughMBB(MachineBasicBlock * MBB,MachineBasicBlock * TBB)2485 static MachineBasicBlock *getFallThroughMBB(MachineBasicBlock *MBB,
2486 MachineBasicBlock *TBB) {
2487 // Look for non-EHPad successors other than TBB. If we find exactly one, it
2488 // is the fallthrough MBB. If we find zero, then TBB is both the target MBB
2489 // and fallthrough MBB. If we find more than one, we cannot identify the
2490 // fallthrough MBB and should return nullptr.
2491 MachineBasicBlock *FallthroughBB = nullptr;
2492 for (auto SI = MBB->succ_begin(), SE = MBB->succ_end(); SI != SE; ++SI) {
2493 if ((*SI)->isEHPad() || (*SI == TBB && FallthroughBB))
2494 continue;
2495 // Return a nullptr if we found more than one fallthrough successor.
2496 if (FallthroughBB && FallthroughBB != TBB)
2497 return nullptr;
2498 FallthroughBB = *SI;
2499 }
2500 return FallthroughBB;
2501 }
2502
AnalyzeBranchImpl(MachineBasicBlock & MBB,MachineBasicBlock * & TBB,MachineBasicBlock * & FBB,SmallVectorImpl<MachineOperand> & Cond,SmallVectorImpl<MachineInstr * > & CondBranches,bool AllowModify) const2503 bool X86InstrInfo::AnalyzeBranchImpl(
2504 MachineBasicBlock &MBB, MachineBasicBlock *&TBB, MachineBasicBlock *&FBB,
2505 SmallVectorImpl<MachineOperand> &Cond,
2506 SmallVectorImpl<MachineInstr *> &CondBranches, bool AllowModify) const {
2507
2508 // Start from the bottom of the block and work up, examining the
2509 // terminator instructions.
2510 MachineBasicBlock::iterator I = MBB.end();
2511 MachineBasicBlock::iterator UnCondBrIter = MBB.end();
2512 while (I != MBB.begin()) {
2513 --I;
2514 if (I->isDebugInstr())
2515 continue;
2516
2517 // Working from the bottom, when we see a non-terminator instruction, we're
2518 // done.
2519 if (!isUnpredicatedTerminator(*I))
2520 break;
2521
2522 // A terminator that isn't a branch can't easily be handled by this
2523 // analysis.
2524 if (!I->isBranch())
2525 return true;
2526
2527 // Handle unconditional branches.
2528 if (I->getOpcode() == X86::JMP_1) {
2529 UnCondBrIter = I;
2530
2531 if (!AllowModify) {
2532 TBB = I->getOperand(0).getMBB();
2533 continue;
2534 }
2535
2536 // If the block has any instructions after a JMP, delete them.
2537 while (std::next(I) != MBB.end())
2538 std::next(I)->eraseFromParent();
2539
2540 Cond.clear();
2541 FBB = nullptr;
2542
2543 // Delete the JMP if it's equivalent to a fall-through.
2544 if (MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) {
2545 TBB = nullptr;
2546 I->eraseFromParent();
2547 I = MBB.end();
2548 UnCondBrIter = MBB.end();
2549 continue;
2550 }
2551
2552 // TBB is used to indicate the unconditional destination.
2553 TBB = I->getOperand(0).getMBB();
2554 continue;
2555 }
2556
2557 // Handle conditional branches.
2558 X86::CondCode BranchCode = X86::getCondFromBranch(*I);
2559 if (BranchCode == X86::COND_INVALID)
2560 return true; // Can't handle indirect branch.
2561
2562 // In practice we should never have an undef eflags operand, if we do
2563 // abort here as we are not prepared to preserve the flag.
2564 if (I->findRegisterUseOperand(X86::EFLAGS)->isUndef())
2565 return true;
2566
2567 // Working from the bottom, handle the first conditional branch.
2568 if (Cond.empty()) {
2569 MachineBasicBlock *TargetBB = I->getOperand(0).getMBB();
2570 if (AllowModify && UnCondBrIter != MBB.end() &&
2571 MBB.isLayoutSuccessor(TargetBB)) {
2572 // If we can modify the code and it ends in something like:
2573 //
2574 // jCC L1
2575 // jmp L2
2576 // L1:
2577 // ...
2578 // L2:
2579 //
2580 // Then we can change this to:
2581 //
2582 // jnCC L2
2583 // L1:
2584 // ...
2585 // L2:
2586 //
2587 // Which is a bit more efficient.
2588 // We conditionally jump to the fall-through block.
2589 BranchCode = GetOppositeBranchCondition(BranchCode);
2590 MachineBasicBlock::iterator OldInst = I;
2591
2592 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(X86::JCC_1))
2593 .addMBB(UnCondBrIter->getOperand(0).getMBB())
2594 .addImm(BranchCode);
2595 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(X86::JMP_1))
2596 .addMBB(TargetBB);
2597
2598 OldInst->eraseFromParent();
2599 UnCondBrIter->eraseFromParent();
2600
2601 // Restart the analysis.
2602 UnCondBrIter = MBB.end();
2603 I = MBB.end();
2604 continue;
2605 }
2606
2607 FBB = TBB;
2608 TBB = I->getOperand(0).getMBB();
2609 Cond.push_back(MachineOperand::CreateImm(BranchCode));
2610 CondBranches.push_back(&*I);
2611 continue;
2612 }
2613
2614 // Handle subsequent conditional branches. Only handle the case where all
2615 // conditional branches branch to the same destination and their condition
2616 // opcodes fit one of the special multi-branch idioms.
2617 assert(Cond.size() == 1);
2618 assert(TBB);
2619
2620 // If the conditions are the same, we can leave them alone.
2621 X86::CondCode OldBranchCode = (X86::CondCode)Cond[0].getImm();
2622 auto NewTBB = I->getOperand(0).getMBB();
2623 if (OldBranchCode == BranchCode && TBB == NewTBB)
2624 continue;
2625
2626 // If they differ, see if they fit one of the known patterns. Theoretically,
2627 // we could handle more patterns here, but we shouldn't expect to see them
2628 // if instruction selection has done a reasonable job.
2629 if (TBB == NewTBB &&
2630 ((OldBranchCode == X86::COND_P && BranchCode == X86::COND_NE) ||
2631 (OldBranchCode == X86::COND_NE && BranchCode == X86::COND_P))) {
2632 BranchCode = X86::COND_NE_OR_P;
2633 } else if ((OldBranchCode == X86::COND_NP && BranchCode == X86::COND_NE) ||
2634 (OldBranchCode == X86::COND_E && BranchCode == X86::COND_P)) {
2635 if (NewTBB != (FBB ? FBB : getFallThroughMBB(&MBB, TBB)))
2636 return true;
2637
2638 // X86::COND_E_AND_NP usually has two different branch destinations.
2639 //
2640 // JP B1
2641 // JE B2
2642 // JMP B1
2643 // B1:
2644 // B2:
2645 //
2646 // Here this condition branches to B2 only if NP && E. It has another
2647 // equivalent form:
2648 //
2649 // JNE B1
2650 // JNP B2
2651 // JMP B1
2652 // B1:
2653 // B2:
2654 //
2655 // Similarly it branches to B2 only if E && NP. That is why this condition
2656 // is named with COND_E_AND_NP.
2657 BranchCode = X86::COND_E_AND_NP;
2658 } else
2659 return true;
2660
2661 // Update the MachineOperand.
2662 Cond[0].setImm(BranchCode);
2663 CondBranches.push_back(&*I);
2664 }
2665
2666 return false;
2667 }
2668
analyzeBranch(MachineBasicBlock & MBB,MachineBasicBlock * & TBB,MachineBasicBlock * & FBB,SmallVectorImpl<MachineOperand> & Cond,bool AllowModify) const2669 bool X86InstrInfo::analyzeBranch(MachineBasicBlock &MBB,
2670 MachineBasicBlock *&TBB,
2671 MachineBasicBlock *&FBB,
2672 SmallVectorImpl<MachineOperand> &Cond,
2673 bool AllowModify) const {
2674 SmallVector<MachineInstr *, 4> CondBranches;
2675 return AnalyzeBranchImpl(MBB, TBB, FBB, Cond, CondBranches, AllowModify);
2676 }
2677
analyzeBranchPredicate(MachineBasicBlock & MBB,MachineBranchPredicate & MBP,bool AllowModify) const2678 bool X86InstrInfo::analyzeBranchPredicate(MachineBasicBlock &MBB,
2679 MachineBranchPredicate &MBP,
2680 bool AllowModify) const {
2681 using namespace std::placeholders;
2682
2683 SmallVector<MachineOperand, 4> Cond;
2684 SmallVector<MachineInstr *, 4> CondBranches;
2685 if (AnalyzeBranchImpl(MBB, MBP.TrueDest, MBP.FalseDest, Cond, CondBranches,
2686 AllowModify))
2687 return true;
2688
2689 if (Cond.size() != 1)
2690 return true;
2691
2692 assert(MBP.TrueDest && "expected!");
2693
2694 if (!MBP.FalseDest)
2695 MBP.FalseDest = MBB.getNextNode();
2696
2697 const TargetRegisterInfo *TRI = &getRegisterInfo();
2698
2699 MachineInstr *ConditionDef = nullptr;
2700 bool SingleUseCondition = true;
2701
2702 for (auto I = std::next(MBB.rbegin()), E = MBB.rend(); I != E; ++I) {
2703 if (I->modifiesRegister(X86::EFLAGS, TRI)) {
2704 ConditionDef = &*I;
2705 break;
2706 }
2707
2708 if (I->readsRegister(X86::EFLAGS, TRI))
2709 SingleUseCondition = false;
2710 }
2711
2712 if (!ConditionDef)
2713 return true;
2714
2715 if (SingleUseCondition) {
2716 for (auto *Succ : MBB.successors())
2717 if (Succ->isLiveIn(X86::EFLAGS))
2718 SingleUseCondition = false;
2719 }
2720
2721 MBP.ConditionDef = ConditionDef;
2722 MBP.SingleUseCondition = SingleUseCondition;
2723
2724 // Currently we only recognize the simple pattern:
2725 //
2726 // test %reg, %reg
2727 // je %label
2728 //
2729 const unsigned TestOpcode =
2730 Subtarget.is64Bit() ? X86::TEST64rr : X86::TEST32rr;
2731
2732 if (ConditionDef->getOpcode() == TestOpcode &&
2733 ConditionDef->getNumOperands() == 3 &&
2734 ConditionDef->getOperand(0).isIdenticalTo(ConditionDef->getOperand(1)) &&
2735 (Cond[0].getImm() == X86::COND_NE || Cond[0].getImm() == X86::COND_E)) {
2736 MBP.LHS = ConditionDef->getOperand(0);
2737 MBP.RHS = MachineOperand::CreateImm(0);
2738 MBP.Predicate = Cond[0].getImm() == X86::COND_NE
2739 ? MachineBranchPredicate::PRED_NE
2740 : MachineBranchPredicate::PRED_EQ;
2741 return false;
2742 }
2743
2744 return true;
2745 }
2746
removeBranch(MachineBasicBlock & MBB,int * BytesRemoved) const2747 unsigned X86InstrInfo::removeBranch(MachineBasicBlock &MBB,
2748 int *BytesRemoved) const {
2749 assert(!BytesRemoved && "code size not handled");
2750
2751 MachineBasicBlock::iterator I = MBB.end();
2752 unsigned Count = 0;
2753
2754 while (I != MBB.begin()) {
2755 --I;
2756 if (I->isDebugInstr())
2757 continue;
2758 if (I->getOpcode() != X86::JMP_1 &&
2759 X86::getCondFromBranch(*I) == X86::COND_INVALID)
2760 break;
2761 // Remove the branch.
2762 I->eraseFromParent();
2763 I = MBB.end();
2764 ++Count;
2765 }
2766
2767 return Count;
2768 }
2769
insertBranch(MachineBasicBlock & MBB,MachineBasicBlock * TBB,MachineBasicBlock * FBB,ArrayRef<MachineOperand> Cond,const DebugLoc & DL,int * BytesAdded) const2770 unsigned X86InstrInfo::insertBranch(MachineBasicBlock &MBB,
2771 MachineBasicBlock *TBB,
2772 MachineBasicBlock *FBB,
2773 ArrayRef<MachineOperand> Cond,
2774 const DebugLoc &DL,
2775 int *BytesAdded) const {
2776 // Shouldn't be a fall through.
2777 assert(TBB && "insertBranch must not be told to insert a fallthrough");
2778 assert((Cond.size() == 1 || Cond.size() == 0) &&
2779 "X86 branch conditions have one component!");
2780 assert(!BytesAdded && "code size not handled");
2781
2782 if (Cond.empty()) {
2783 // Unconditional branch?
2784 assert(!FBB && "Unconditional branch with multiple successors!");
2785 BuildMI(&MBB, DL, get(X86::JMP_1)).addMBB(TBB);
2786 return 1;
2787 }
2788
2789 // If FBB is null, it is implied to be a fall-through block.
2790 bool FallThru = FBB == nullptr;
2791
2792 // Conditional branch.
2793 unsigned Count = 0;
2794 X86::CondCode CC = (X86::CondCode)Cond[0].getImm();
2795 switch (CC) {
2796 case X86::COND_NE_OR_P:
2797 // Synthesize NE_OR_P with two branches.
2798 BuildMI(&MBB, DL, get(X86::JCC_1)).addMBB(TBB).addImm(X86::COND_NE);
2799 ++Count;
2800 BuildMI(&MBB, DL, get(X86::JCC_1)).addMBB(TBB).addImm(X86::COND_P);
2801 ++Count;
2802 break;
2803 case X86::COND_E_AND_NP:
2804 // Use the next block of MBB as FBB if it is null.
2805 if (FBB == nullptr) {
2806 FBB = getFallThroughMBB(&MBB, TBB);
2807 assert(FBB && "MBB cannot be the last block in function when the false "
2808 "body is a fall-through.");
2809 }
2810 // Synthesize COND_E_AND_NP with two branches.
2811 BuildMI(&MBB, DL, get(X86::JCC_1)).addMBB(FBB).addImm(X86::COND_NE);
2812 ++Count;
2813 BuildMI(&MBB, DL, get(X86::JCC_1)).addMBB(TBB).addImm(X86::COND_NP);
2814 ++Count;
2815 break;
2816 default: {
2817 BuildMI(&MBB, DL, get(X86::JCC_1)).addMBB(TBB).addImm(CC);
2818 ++Count;
2819 }
2820 }
2821 if (!FallThru) {
2822 // Two-way Conditional branch. Insert the second branch.
2823 BuildMI(&MBB, DL, get(X86::JMP_1)).addMBB(FBB);
2824 ++Count;
2825 }
2826 return Count;
2827 }
2828
2829 bool X86InstrInfo::
canInsertSelect(const MachineBasicBlock & MBB,ArrayRef<MachineOperand> Cond,unsigned TrueReg,unsigned FalseReg,int & CondCycles,int & TrueCycles,int & FalseCycles) const2830 canInsertSelect(const MachineBasicBlock &MBB,
2831 ArrayRef<MachineOperand> Cond,
2832 unsigned TrueReg, unsigned FalseReg,
2833 int &CondCycles, int &TrueCycles, int &FalseCycles) const {
2834 // Not all subtargets have cmov instructions.
2835 if (!Subtarget.hasCMov())
2836 return false;
2837 if (Cond.size() != 1)
2838 return false;
2839 // We cannot do the composite conditions, at least not in SSA form.
2840 if ((X86::CondCode)Cond[0].getImm() > X86::LAST_VALID_COND)
2841 return false;
2842
2843 // Check register classes.
2844 const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2845 const TargetRegisterClass *RC =
2846 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg));
2847 if (!RC)
2848 return false;
2849
2850 // We have cmov instructions for 16, 32, and 64 bit general purpose registers.
2851 if (X86::GR16RegClass.hasSubClassEq(RC) ||
2852 X86::GR32RegClass.hasSubClassEq(RC) ||
2853 X86::GR64RegClass.hasSubClassEq(RC)) {
2854 // This latency applies to Pentium M, Merom, Wolfdale, Nehalem, and Sandy
2855 // Bridge. Probably Ivy Bridge as well.
2856 CondCycles = 2;
2857 TrueCycles = 2;
2858 FalseCycles = 2;
2859 return true;
2860 }
2861
2862 // Can't do vectors.
2863 return false;
2864 }
2865
insertSelect(MachineBasicBlock & MBB,MachineBasicBlock::iterator I,const DebugLoc & DL,unsigned DstReg,ArrayRef<MachineOperand> Cond,unsigned TrueReg,unsigned FalseReg) const2866 void X86InstrInfo::insertSelect(MachineBasicBlock &MBB,
2867 MachineBasicBlock::iterator I,
2868 const DebugLoc &DL, unsigned DstReg,
2869 ArrayRef<MachineOperand> Cond, unsigned TrueReg,
2870 unsigned FalseReg) const {
2871 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2872 const TargetRegisterInfo &TRI = *MRI.getTargetRegisterInfo();
2873 const TargetRegisterClass &RC = *MRI.getRegClass(DstReg);
2874 assert(Cond.size() == 1 && "Invalid Cond array");
2875 unsigned Opc = X86::getCMovOpcode(TRI.getRegSizeInBits(RC) / 8,
2876 false /*HasMemoryOperand*/);
2877 BuildMI(MBB, I, DL, get(Opc), DstReg)
2878 .addReg(FalseReg)
2879 .addReg(TrueReg)
2880 .addImm(Cond[0].getImm());
2881 }
2882
2883 /// Test if the given register is a physical h register.
isHReg(unsigned Reg)2884 static bool isHReg(unsigned Reg) {
2885 return X86::GR8_ABCD_HRegClass.contains(Reg);
2886 }
2887
2888 // Try and copy between VR128/VR64 and GR64 registers.
CopyToFromAsymmetricReg(unsigned DestReg,unsigned SrcReg,const X86Subtarget & Subtarget)2889 static unsigned CopyToFromAsymmetricReg(unsigned DestReg, unsigned SrcReg,
2890 const X86Subtarget &Subtarget) {
2891 bool HasAVX = Subtarget.hasAVX();
2892 bool HasAVX512 = Subtarget.hasAVX512();
2893
2894 // SrcReg(MaskReg) -> DestReg(GR64)
2895 // SrcReg(MaskReg) -> DestReg(GR32)
2896
2897 // All KMASK RegClasses hold the same k registers, can be tested against anyone.
2898 if (X86::VK16RegClass.contains(SrcReg)) {
2899 if (X86::GR64RegClass.contains(DestReg)) {
2900 assert(Subtarget.hasBWI());
2901 return X86::KMOVQrk;
2902 }
2903 if (X86::GR32RegClass.contains(DestReg))
2904 return Subtarget.hasBWI() ? X86::KMOVDrk : X86::KMOVWrk;
2905 }
2906
2907 // SrcReg(GR64) -> DestReg(MaskReg)
2908 // SrcReg(GR32) -> DestReg(MaskReg)
2909
2910 // All KMASK RegClasses hold the same k registers, can be tested against anyone.
2911 if (X86::VK16RegClass.contains(DestReg)) {
2912 if (X86::GR64RegClass.contains(SrcReg)) {
2913 assert(Subtarget.hasBWI());
2914 return X86::KMOVQkr;
2915 }
2916 if (X86::GR32RegClass.contains(SrcReg))
2917 return Subtarget.hasBWI() ? X86::KMOVDkr : X86::KMOVWkr;
2918 }
2919
2920
2921 // SrcReg(VR128) -> DestReg(GR64)
2922 // SrcReg(VR64) -> DestReg(GR64)
2923 // SrcReg(GR64) -> DestReg(VR128)
2924 // SrcReg(GR64) -> DestReg(VR64)
2925
2926 if (X86::GR64RegClass.contains(DestReg)) {
2927 if (X86::VR128XRegClass.contains(SrcReg))
2928 // Copy from a VR128 register to a GR64 register.
2929 return HasAVX512 ? X86::VMOVPQIto64Zrr :
2930 HasAVX ? X86::VMOVPQIto64rr :
2931 X86::MOVPQIto64rr;
2932 if (X86::VR64RegClass.contains(SrcReg))
2933 // Copy from a VR64 register to a GR64 register.
2934 return X86::MMX_MOVD64from64rr;
2935 } else if (X86::GR64RegClass.contains(SrcReg)) {
2936 // Copy from a GR64 register to a VR128 register.
2937 if (X86::VR128XRegClass.contains(DestReg))
2938 return HasAVX512 ? X86::VMOV64toPQIZrr :
2939 HasAVX ? X86::VMOV64toPQIrr :
2940 X86::MOV64toPQIrr;
2941 // Copy from a GR64 register to a VR64 register.
2942 if (X86::VR64RegClass.contains(DestReg))
2943 return X86::MMX_MOVD64to64rr;
2944 }
2945
2946 // SrcReg(VR128) -> DestReg(GR32)
2947 // SrcReg(GR32) -> DestReg(VR128)
2948
2949 if (X86::GR32RegClass.contains(DestReg) &&
2950 X86::VR128XRegClass.contains(SrcReg))
2951 // Copy from a VR128 register to a GR32 register.
2952 return HasAVX512 ? X86::VMOVPDI2DIZrr :
2953 HasAVX ? X86::VMOVPDI2DIrr :
2954 X86::MOVPDI2DIrr;
2955
2956 if (X86::VR128XRegClass.contains(DestReg) &&
2957 X86::GR32RegClass.contains(SrcReg))
2958 // Copy from a VR128 register to a VR128 register.
2959 return HasAVX512 ? X86::VMOVDI2PDIZrr :
2960 HasAVX ? X86::VMOVDI2PDIrr :
2961 X86::MOVDI2PDIrr;
2962 return 0;
2963 }
2964
copyPhysReg(MachineBasicBlock & MBB,MachineBasicBlock::iterator MI,const DebugLoc & DL,MCRegister DestReg,MCRegister SrcReg,bool KillSrc) const2965 void X86InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
2966 MachineBasicBlock::iterator MI,
2967 const DebugLoc &DL, MCRegister DestReg,
2968 MCRegister SrcReg, bool KillSrc) const {
2969 // First deal with the normal symmetric copies.
2970 bool HasAVX = Subtarget.hasAVX();
2971 bool HasVLX = Subtarget.hasVLX();
2972 unsigned Opc = 0;
2973 if (X86::GR64RegClass.contains(DestReg, SrcReg))
2974 Opc = X86::MOV64rr;
2975 else if (X86::GR32RegClass.contains(DestReg, SrcReg))
2976 Opc = X86::MOV32rr;
2977 else if (X86::GR16RegClass.contains(DestReg, SrcReg))
2978 Opc = X86::MOV16rr;
2979 else if (X86::GR8RegClass.contains(DestReg, SrcReg)) {
2980 // Copying to or from a physical H register on x86-64 requires a NOREX
2981 // move. Otherwise use a normal move.
2982 if ((isHReg(DestReg) || isHReg(SrcReg)) &&
2983 Subtarget.is64Bit()) {
2984 Opc = X86::MOV8rr_NOREX;
2985 // Both operands must be encodable without an REX prefix.
2986 assert(X86::GR8_NOREXRegClass.contains(SrcReg, DestReg) &&
2987 "8-bit H register can not be copied outside GR8_NOREX");
2988 } else
2989 Opc = X86::MOV8rr;
2990 }
2991 else if (X86::VR64RegClass.contains(DestReg, SrcReg))
2992 Opc = X86::MMX_MOVQ64rr;
2993 else if (X86::VR128XRegClass.contains(DestReg, SrcReg)) {
2994 if (HasVLX)
2995 Opc = X86::VMOVAPSZ128rr;
2996 else if (X86::VR128RegClass.contains(DestReg, SrcReg))
2997 Opc = HasAVX ? X86::VMOVAPSrr : X86::MOVAPSrr;
2998 else {
2999 // If this an extended register and we don't have VLX we need to use a
3000 // 512-bit move.
3001 Opc = X86::VMOVAPSZrr;
3002 const TargetRegisterInfo *TRI = &getRegisterInfo();
3003 DestReg = TRI->getMatchingSuperReg(DestReg, X86::sub_xmm,
3004 &X86::VR512RegClass);
3005 SrcReg = TRI->getMatchingSuperReg(SrcReg, X86::sub_xmm,
3006 &X86::VR512RegClass);
3007 }
3008 } else if (X86::VR256XRegClass.contains(DestReg, SrcReg)) {
3009 if (HasVLX)
3010 Opc = X86::VMOVAPSZ256rr;
3011 else if (X86::VR256RegClass.contains(DestReg, SrcReg))
3012 Opc = X86::VMOVAPSYrr;
3013 else {
3014 // If this an extended register and we don't have VLX we need to use a
3015 // 512-bit move.
3016 Opc = X86::VMOVAPSZrr;
3017 const TargetRegisterInfo *TRI = &getRegisterInfo();
3018 DestReg = TRI->getMatchingSuperReg(DestReg, X86::sub_ymm,
3019 &X86::VR512RegClass);
3020 SrcReg = TRI->getMatchingSuperReg(SrcReg, X86::sub_ymm,
3021 &X86::VR512RegClass);
3022 }
3023 } else if (X86::VR512RegClass.contains(DestReg, SrcReg))
3024 Opc = X86::VMOVAPSZrr;
3025 // All KMASK RegClasses hold the same k registers, can be tested against anyone.
3026 else if (X86::VK16RegClass.contains(DestReg, SrcReg))
3027 Opc = Subtarget.hasBWI() ? X86::KMOVQkk : X86::KMOVWkk;
3028 if (!Opc)
3029 Opc = CopyToFromAsymmetricReg(DestReg, SrcReg, Subtarget);
3030
3031 if (Opc) {
3032 BuildMI(MBB, MI, DL, get(Opc), DestReg)
3033 .addReg(SrcReg, getKillRegState(KillSrc));
3034 return;
3035 }
3036
3037 if (SrcReg == X86::EFLAGS || DestReg == X86::EFLAGS) {
3038 // FIXME: We use a fatal error here because historically LLVM has tried
3039 // lower some of these physreg copies and we want to ensure we get
3040 // reasonable bug reports if someone encounters a case no other testing
3041 // found. This path should be removed after the LLVM 7 release.
3042 report_fatal_error("Unable to copy EFLAGS physical register!");
3043 }
3044
3045 LLVM_DEBUG(dbgs() << "Cannot copy " << RI.getName(SrcReg) << " to "
3046 << RI.getName(DestReg) << '\n');
3047 report_fatal_error("Cannot emit physreg copy instruction");
3048 }
3049
3050 Optional<DestSourcePair>
isCopyInstrImpl(const MachineInstr & MI) const3051 X86InstrInfo::isCopyInstrImpl(const MachineInstr &MI) const {
3052 if (MI.isMoveReg())
3053 return DestSourcePair{MI.getOperand(0), MI.getOperand(1)};
3054 return None;
3055 }
3056
getLoadStoreRegOpcode(unsigned Reg,const TargetRegisterClass * RC,bool isStackAligned,const X86Subtarget & STI,bool load)3057 static unsigned getLoadStoreRegOpcode(unsigned Reg,
3058 const TargetRegisterClass *RC,
3059 bool isStackAligned,
3060 const X86Subtarget &STI,
3061 bool load) {
3062 bool HasAVX = STI.hasAVX();
3063 bool HasAVX512 = STI.hasAVX512();
3064 bool HasVLX = STI.hasVLX();
3065
3066 switch (STI.getRegisterInfo()->getSpillSize(*RC)) {
3067 default:
3068 llvm_unreachable("Unknown spill size");
3069 case 1:
3070 assert(X86::GR8RegClass.hasSubClassEq(RC) && "Unknown 1-byte regclass");
3071 if (STI.is64Bit())
3072 // Copying to or from a physical H register on x86-64 requires a NOREX
3073 // move. Otherwise use a normal move.
3074 if (isHReg(Reg) || X86::GR8_ABCD_HRegClass.hasSubClassEq(RC))
3075 return load ? X86::MOV8rm_NOREX : X86::MOV8mr_NOREX;
3076 return load ? X86::MOV8rm : X86::MOV8mr;
3077 case 2:
3078 if (X86::VK16RegClass.hasSubClassEq(RC))
3079 return load ? X86::KMOVWkm : X86::KMOVWmk;
3080 assert(X86::GR16RegClass.hasSubClassEq(RC) && "Unknown 2-byte regclass");
3081 return load ? X86::MOV16rm : X86::MOV16mr;
3082 case 4:
3083 if (X86::GR32RegClass.hasSubClassEq(RC))
3084 return load ? X86::MOV32rm : X86::MOV32mr;
3085 if (X86::FR32XRegClass.hasSubClassEq(RC))
3086 return load ?
3087 (HasAVX512 ? X86::VMOVSSZrm_alt :
3088 HasAVX ? X86::VMOVSSrm_alt :
3089 X86::MOVSSrm_alt) :
3090 (HasAVX512 ? X86::VMOVSSZmr :
3091 HasAVX ? X86::VMOVSSmr :
3092 X86::MOVSSmr);
3093 if (X86::RFP32RegClass.hasSubClassEq(RC))
3094 return load ? X86::LD_Fp32m : X86::ST_Fp32m;
3095 if (X86::VK32RegClass.hasSubClassEq(RC)) {
3096 assert(STI.hasBWI() && "KMOVD requires BWI");
3097 return load ? X86::KMOVDkm : X86::KMOVDmk;
3098 }
3099 // All of these mask pair classes have the same spill size, the same kind
3100 // of kmov instructions can be used with all of them.
3101 if (X86::VK1PAIRRegClass.hasSubClassEq(RC) ||
3102 X86::VK2PAIRRegClass.hasSubClassEq(RC) ||
3103 X86::VK4PAIRRegClass.hasSubClassEq(RC) ||
3104 X86::VK8PAIRRegClass.hasSubClassEq(RC) ||
3105 X86::VK16PAIRRegClass.hasSubClassEq(RC))
3106 return load ? X86::MASKPAIR16LOAD : X86::MASKPAIR16STORE;
3107 llvm_unreachable("Unknown 4-byte regclass");
3108 case 8:
3109 if (X86::GR64RegClass.hasSubClassEq(RC))
3110 return load ? X86::MOV64rm : X86::MOV64mr;
3111 if (X86::FR64XRegClass.hasSubClassEq(RC))
3112 return load ?
3113 (HasAVX512 ? X86::VMOVSDZrm_alt :
3114 HasAVX ? X86::VMOVSDrm_alt :
3115 X86::MOVSDrm_alt) :
3116 (HasAVX512 ? X86::VMOVSDZmr :
3117 HasAVX ? X86::VMOVSDmr :
3118 X86::MOVSDmr);
3119 if (X86::VR64RegClass.hasSubClassEq(RC))
3120 return load ? X86::MMX_MOVQ64rm : X86::MMX_MOVQ64mr;
3121 if (X86::RFP64RegClass.hasSubClassEq(RC))
3122 return load ? X86::LD_Fp64m : X86::ST_Fp64m;
3123 if (X86::VK64RegClass.hasSubClassEq(RC)) {
3124 assert(STI.hasBWI() && "KMOVQ requires BWI");
3125 return load ? X86::KMOVQkm : X86::KMOVQmk;
3126 }
3127 llvm_unreachable("Unknown 8-byte regclass");
3128 case 10:
3129 assert(X86::RFP80RegClass.hasSubClassEq(RC) && "Unknown 10-byte regclass");
3130 return load ? X86::LD_Fp80m : X86::ST_FpP80m;
3131 case 16: {
3132 if (X86::VR128XRegClass.hasSubClassEq(RC)) {
3133 // If stack is realigned we can use aligned stores.
3134 if (isStackAligned)
3135 return load ?
3136 (HasVLX ? X86::VMOVAPSZ128rm :
3137 HasAVX512 ? X86::VMOVAPSZ128rm_NOVLX :
3138 HasAVX ? X86::VMOVAPSrm :
3139 X86::MOVAPSrm):
3140 (HasVLX ? X86::VMOVAPSZ128mr :
3141 HasAVX512 ? X86::VMOVAPSZ128mr_NOVLX :
3142 HasAVX ? X86::VMOVAPSmr :
3143 X86::MOVAPSmr);
3144 else
3145 return load ?
3146 (HasVLX ? X86::VMOVUPSZ128rm :
3147 HasAVX512 ? X86::VMOVUPSZ128rm_NOVLX :
3148 HasAVX ? X86::VMOVUPSrm :
3149 X86::MOVUPSrm):
3150 (HasVLX ? X86::VMOVUPSZ128mr :
3151 HasAVX512 ? X86::VMOVUPSZ128mr_NOVLX :
3152 HasAVX ? X86::VMOVUPSmr :
3153 X86::MOVUPSmr);
3154 }
3155 if (X86::BNDRRegClass.hasSubClassEq(RC)) {
3156 if (STI.is64Bit())
3157 return load ? X86::BNDMOV64rm : X86::BNDMOV64mr;
3158 else
3159 return load ? X86::BNDMOV32rm : X86::BNDMOV32mr;
3160 }
3161 llvm_unreachable("Unknown 16-byte regclass");
3162 }
3163 case 32:
3164 assert(X86::VR256XRegClass.hasSubClassEq(RC) && "Unknown 32-byte regclass");
3165 // If stack is realigned we can use aligned stores.
3166 if (isStackAligned)
3167 return load ?
3168 (HasVLX ? X86::VMOVAPSZ256rm :
3169 HasAVX512 ? X86::VMOVAPSZ256rm_NOVLX :
3170 X86::VMOVAPSYrm) :
3171 (HasVLX ? X86::VMOVAPSZ256mr :
3172 HasAVX512 ? X86::VMOVAPSZ256mr_NOVLX :
3173 X86::VMOVAPSYmr);
3174 else
3175 return load ?
3176 (HasVLX ? X86::VMOVUPSZ256rm :
3177 HasAVX512 ? X86::VMOVUPSZ256rm_NOVLX :
3178 X86::VMOVUPSYrm) :
3179 (HasVLX ? X86::VMOVUPSZ256mr :
3180 HasAVX512 ? X86::VMOVUPSZ256mr_NOVLX :
3181 X86::VMOVUPSYmr);
3182 case 64:
3183 assert(X86::VR512RegClass.hasSubClassEq(RC) && "Unknown 64-byte regclass");
3184 assert(STI.hasAVX512() && "Using 512-bit register requires AVX512");
3185 if (isStackAligned)
3186 return load ? X86::VMOVAPSZrm : X86::VMOVAPSZmr;
3187 else
3188 return load ? X86::VMOVUPSZrm : X86::VMOVUPSZmr;
3189 }
3190 }
3191
getMemOperandWithOffset(const MachineInstr & MemOp,const MachineOperand * & BaseOp,int64_t & Offset,const TargetRegisterInfo * TRI) const3192 bool X86InstrInfo::getMemOperandWithOffset(
3193 const MachineInstr &MemOp, const MachineOperand *&BaseOp, int64_t &Offset,
3194 const TargetRegisterInfo *TRI) const {
3195 const MCInstrDesc &Desc = MemOp.getDesc();
3196 int MemRefBegin = X86II::getMemoryOperandNo(Desc.TSFlags);
3197 if (MemRefBegin < 0)
3198 return false;
3199
3200 MemRefBegin += X86II::getOperandBias(Desc);
3201
3202 BaseOp = &MemOp.getOperand(MemRefBegin + X86::AddrBaseReg);
3203 if (!BaseOp->isReg()) // Can be an MO_FrameIndex
3204 return false;
3205
3206 if (MemOp.getOperand(MemRefBegin + X86::AddrScaleAmt).getImm() != 1)
3207 return false;
3208
3209 if (MemOp.getOperand(MemRefBegin + X86::AddrIndexReg).getReg() !=
3210 X86::NoRegister)
3211 return false;
3212
3213 const MachineOperand &DispMO = MemOp.getOperand(MemRefBegin + X86::AddrDisp);
3214
3215 // Displacement can be symbolic
3216 if (!DispMO.isImm())
3217 return false;
3218
3219 Offset = DispMO.getImm();
3220
3221 if (!BaseOp->isReg())
3222 return false;
3223
3224 return true;
3225 }
3226
getStoreRegOpcode(unsigned SrcReg,const TargetRegisterClass * RC,bool isStackAligned,const X86Subtarget & STI)3227 static unsigned getStoreRegOpcode(unsigned SrcReg,
3228 const TargetRegisterClass *RC,
3229 bool isStackAligned,
3230 const X86Subtarget &STI) {
3231 return getLoadStoreRegOpcode(SrcReg, RC, isStackAligned, STI, false);
3232 }
3233
3234
getLoadRegOpcode(unsigned DestReg,const TargetRegisterClass * RC,bool isStackAligned,const X86Subtarget & STI)3235 static unsigned getLoadRegOpcode(unsigned DestReg,
3236 const TargetRegisterClass *RC,
3237 bool isStackAligned,
3238 const X86Subtarget &STI) {
3239 return getLoadStoreRegOpcode(DestReg, RC, isStackAligned, STI, true);
3240 }
3241
storeRegToStackSlot(MachineBasicBlock & MBB,MachineBasicBlock::iterator MI,unsigned SrcReg,bool isKill,int FrameIdx,const TargetRegisterClass * RC,const TargetRegisterInfo * TRI) const3242 void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
3243 MachineBasicBlock::iterator MI,
3244 unsigned SrcReg, bool isKill, int FrameIdx,
3245 const TargetRegisterClass *RC,
3246 const TargetRegisterInfo *TRI) const {
3247 const MachineFunction &MF = *MBB.getParent();
3248 assert(MF.getFrameInfo().getObjectSize(FrameIdx) >= TRI->getSpillSize(*RC) &&
3249 "Stack slot too small for store");
3250 unsigned Alignment = std::max<uint32_t>(TRI->getSpillSize(*RC), 16);
3251 bool isAligned =
3252 (Subtarget.getFrameLowering()->getStackAlignment() >= Alignment) ||
3253 RI.canRealignStack(MF);
3254 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, Subtarget);
3255 addFrameReference(BuildMI(MBB, MI, DebugLoc(), get(Opc)), FrameIdx)
3256 .addReg(SrcReg, getKillRegState(isKill));
3257 }
3258
loadRegFromStackSlot(MachineBasicBlock & MBB,MachineBasicBlock::iterator MI,unsigned DestReg,int FrameIdx,const TargetRegisterClass * RC,const TargetRegisterInfo * TRI) const3259 void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
3260 MachineBasicBlock::iterator MI,
3261 unsigned DestReg, int FrameIdx,
3262 const TargetRegisterClass *RC,
3263 const TargetRegisterInfo *TRI) const {
3264 const MachineFunction &MF = *MBB.getParent();
3265 unsigned Alignment = std::max<uint32_t>(TRI->getSpillSize(*RC), 16);
3266 bool isAligned =
3267 (Subtarget.getFrameLowering()->getStackAlignment() >= Alignment) ||
3268 RI.canRealignStack(MF);
3269 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, Subtarget);
3270 addFrameReference(BuildMI(MBB, MI, DebugLoc(), get(Opc), DestReg), FrameIdx);
3271 }
3272
analyzeCompare(const MachineInstr & MI,unsigned & SrcReg,unsigned & SrcReg2,int & CmpMask,int & CmpValue) const3273 bool X86InstrInfo::analyzeCompare(const MachineInstr &MI, unsigned &SrcReg,
3274 unsigned &SrcReg2, int &CmpMask,
3275 int &CmpValue) const {
3276 switch (MI.getOpcode()) {
3277 default: break;
3278 case X86::CMP64ri32:
3279 case X86::CMP64ri8:
3280 case X86::CMP32ri:
3281 case X86::CMP32ri8:
3282 case X86::CMP16ri:
3283 case X86::CMP16ri8:
3284 case X86::CMP8ri:
3285 SrcReg = MI.getOperand(0).getReg();
3286 SrcReg2 = 0;
3287 if (MI.getOperand(1).isImm()) {
3288 CmpMask = ~0;
3289 CmpValue = MI.getOperand(1).getImm();
3290 } else {
3291 CmpMask = CmpValue = 0;
3292 }
3293 return true;
3294 // A SUB can be used to perform comparison.
3295 case X86::SUB64rm:
3296 case X86::SUB32rm:
3297 case X86::SUB16rm:
3298 case X86::SUB8rm:
3299 SrcReg = MI.getOperand(1).getReg();
3300 SrcReg2 = 0;
3301 CmpMask = 0;
3302 CmpValue = 0;
3303 return true;
3304 case X86::SUB64rr:
3305 case X86::SUB32rr:
3306 case X86::SUB16rr:
3307 case X86::SUB8rr:
3308 SrcReg = MI.getOperand(1).getReg();
3309 SrcReg2 = MI.getOperand(2).getReg();
3310 CmpMask = 0;
3311 CmpValue = 0;
3312 return true;
3313 case X86::SUB64ri32:
3314 case X86::SUB64ri8:
3315 case X86::SUB32ri:
3316 case X86::SUB32ri8:
3317 case X86::SUB16ri:
3318 case X86::SUB16ri8:
3319 case X86::SUB8ri:
3320 SrcReg = MI.getOperand(1).getReg();
3321 SrcReg2 = 0;
3322 if (MI.getOperand(2).isImm()) {
3323 CmpMask = ~0;
3324 CmpValue = MI.getOperand(2).getImm();
3325 } else {
3326 CmpMask = CmpValue = 0;
3327 }
3328 return true;
3329 case X86::CMP64rr:
3330 case X86::CMP32rr:
3331 case X86::CMP16rr:
3332 case X86::CMP8rr:
3333 SrcReg = MI.getOperand(0).getReg();
3334 SrcReg2 = MI.getOperand(1).getReg();
3335 CmpMask = 0;
3336 CmpValue = 0;
3337 return true;
3338 case X86::TEST8rr:
3339 case X86::TEST16rr:
3340 case X86::TEST32rr:
3341 case X86::TEST64rr:
3342 SrcReg = MI.getOperand(0).getReg();
3343 if (MI.getOperand(1).getReg() != SrcReg)
3344 return false;
3345 // Compare against zero.
3346 SrcReg2 = 0;
3347 CmpMask = ~0;
3348 CmpValue = 0;
3349 return true;
3350 }
3351 return false;
3352 }
3353
3354 /// Check whether the first instruction, whose only
3355 /// purpose is to update flags, can be made redundant.
3356 /// CMPrr can be made redundant by SUBrr if the operands are the same.
3357 /// This function can be extended later on.
3358 /// SrcReg, SrcRegs: register operands for FlagI.
3359 /// ImmValue: immediate for FlagI if it takes an immediate.
isRedundantFlagInstr(const MachineInstr & FlagI,unsigned SrcReg,unsigned SrcReg2,int ImmMask,int ImmValue,const MachineInstr & OI)3360 inline static bool isRedundantFlagInstr(const MachineInstr &FlagI,
3361 unsigned SrcReg, unsigned SrcReg2,
3362 int ImmMask, int ImmValue,
3363 const MachineInstr &OI) {
3364 if (((FlagI.getOpcode() == X86::CMP64rr && OI.getOpcode() == X86::SUB64rr) ||
3365 (FlagI.getOpcode() == X86::CMP32rr && OI.getOpcode() == X86::SUB32rr) ||
3366 (FlagI.getOpcode() == X86::CMP16rr && OI.getOpcode() == X86::SUB16rr) ||
3367 (FlagI.getOpcode() == X86::CMP8rr && OI.getOpcode() == X86::SUB8rr)) &&
3368 ((OI.getOperand(1).getReg() == SrcReg &&
3369 OI.getOperand(2).getReg() == SrcReg2) ||
3370 (OI.getOperand(1).getReg() == SrcReg2 &&
3371 OI.getOperand(2).getReg() == SrcReg)))
3372 return true;
3373
3374 if (ImmMask != 0 &&
3375 ((FlagI.getOpcode() == X86::CMP64ri32 &&
3376 OI.getOpcode() == X86::SUB64ri32) ||
3377 (FlagI.getOpcode() == X86::CMP64ri8 &&
3378 OI.getOpcode() == X86::SUB64ri8) ||
3379 (FlagI.getOpcode() == X86::CMP32ri && OI.getOpcode() == X86::SUB32ri) ||
3380 (FlagI.getOpcode() == X86::CMP32ri8 &&
3381 OI.getOpcode() == X86::SUB32ri8) ||
3382 (FlagI.getOpcode() == X86::CMP16ri && OI.getOpcode() == X86::SUB16ri) ||
3383 (FlagI.getOpcode() == X86::CMP16ri8 &&
3384 OI.getOpcode() == X86::SUB16ri8) ||
3385 (FlagI.getOpcode() == X86::CMP8ri && OI.getOpcode() == X86::SUB8ri)) &&
3386 OI.getOperand(1).getReg() == SrcReg &&
3387 OI.getOperand(2).getImm() == ImmValue)
3388 return true;
3389 return false;
3390 }
3391
3392 /// Check whether the definition can be converted
3393 /// to remove a comparison against zero.
isDefConvertible(const MachineInstr & MI,bool & NoSignFlag)3394 inline static bool isDefConvertible(const MachineInstr &MI, bool &NoSignFlag) {
3395 NoSignFlag = false;
3396
3397 switch (MI.getOpcode()) {
3398 default: return false;
3399
3400 // The shift instructions only modify ZF if their shift count is non-zero.
3401 // N.B.: The processor truncates the shift count depending on the encoding.
3402 case X86::SAR8ri: case X86::SAR16ri: case X86::SAR32ri:case X86::SAR64ri:
3403 case X86::SHR8ri: case X86::SHR16ri: case X86::SHR32ri:case X86::SHR64ri:
3404 return getTruncatedShiftCount(MI, 2) != 0;
3405
3406 // Some left shift instructions can be turned into LEA instructions but only
3407 // if their flags aren't used. Avoid transforming such instructions.
3408 case X86::SHL8ri: case X86::SHL16ri: case X86::SHL32ri:case X86::SHL64ri:{
3409 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
3410 if (isTruncatedShiftCountForLEA(ShAmt)) return false;
3411 return ShAmt != 0;
3412 }
3413
3414 case X86::SHRD16rri8:case X86::SHRD32rri8:case X86::SHRD64rri8:
3415 case X86::SHLD16rri8:case X86::SHLD32rri8:case X86::SHLD64rri8:
3416 return getTruncatedShiftCount(MI, 3) != 0;
3417
3418 case X86::SUB64ri32: case X86::SUB64ri8: case X86::SUB32ri:
3419 case X86::SUB32ri8: case X86::SUB16ri: case X86::SUB16ri8:
3420 case X86::SUB8ri: case X86::SUB64rr: case X86::SUB32rr:
3421 case X86::SUB16rr: case X86::SUB8rr: case X86::SUB64rm:
3422 case X86::SUB32rm: case X86::SUB16rm: case X86::SUB8rm:
3423 case X86::DEC64r: case X86::DEC32r: case X86::DEC16r: case X86::DEC8r:
3424 case X86::ADD64ri32: case X86::ADD64ri8: case X86::ADD32ri:
3425 case X86::ADD32ri8: case X86::ADD16ri: case X86::ADD16ri8:
3426 case X86::ADD8ri: case X86::ADD64rr: case X86::ADD32rr:
3427 case X86::ADD16rr: case X86::ADD8rr: case X86::ADD64rm:
3428 case X86::ADD32rm: case X86::ADD16rm: case X86::ADD8rm:
3429 case X86::INC64r: case X86::INC32r: case X86::INC16r: case X86::INC8r:
3430 case X86::AND64ri32: case X86::AND64ri8: case X86::AND32ri:
3431 case X86::AND32ri8: case X86::AND16ri: case X86::AND16ri8:
3432 case X86::AND8ri: case X86::AND64rr: case X86::AND32rr:
3433 case X86::AND16rr: case X86::AND8rr: case X86::AND64rm:
3434 case X86::AND32rm: case X86::AND16rm: case X86::AND8rm:
3435 case X86::XOR64ri32: case X86::XOR64ri8: case X86::XOR32ri:
3436 case X86::XOR32ri8: case X86::XOR16ri: case X86::XOR16ri8:
3437 case X86::XOR8ri: case X86::XOR64rr: case X86::XOR32rr:
3438 case X86::XOR16rr: case X86::XOR8rr: case X86::XOR64rm:
3439 case X86::XOR32rm: case X86::XOR16rm: case X86::XOR8rm:
3440 case X86::OR64ri32: case X86::OR64ri8: case X86::OR32ri:
3441 case X86::OR32ri8: case X86::OR16ri: case X86::OR16ri8:
3442 case X86::OR8ri: case X86::OR64rr: case X86::OR32rr:
3443 case X86::OR16rr: case X86::OR8rr: case X86::OR64rm:
3444 case X86::OR32rm: case X86::OR16rm: case X86::OR8rm:
3445 case X86::ADC64ri32: case X86::ADC64ri8: case X86::ADC32ri:
3446 case X86::ADC32ri8: case X86::ADC16ri: case X86::ADC16ri8:
3447 case X86::ADC8ri: case X86::ADC64rr: case X86::ADC32rr:
3448 case X86::ADC16rr: case X86::ADC8rr: case X86::ADC64rm:
3449 case X86::ADC32rm: case X86::ADC16rm: case X86::ADC8rm:
3450 case X86::SBB64ri32: case X86::SBB64ri8: case X86::SBB32ri:
3451 case X86::SBB32ri8: case X86::SBB16ri: case X86::SBB16ri8:
3452 case X86::SBB8ri: case X86::SBB64rr: case X86::SBB32rr:
3453 case X86::SBB16rr: case X86::SBB8rr: case X86::SBB64rm:
3454 case X86::SBB32rm: case X86::SBB16rm: case X86::SBB8rm:
3455 case X86::NEG8r: case X86::NEG16r: case X86::NEG32r: case X86::NEG64r:
3456 case X86::SAR8r1: case X86::SAR16r1: case X86::SAR32r1:case X86::SAR64r1:
3457 case X86::SHR8r1: case X86::SHR16r1: case X86::SHR32r1:case X86::SHR64r1:
3458 case X86::SHL8r1: case X86::SHL16r1: case X86::SHL32r1:case X86::SHL64r1:
3459 case X86::ANDN32rr: case X86::ANDN32rm:
3460 case X86::ANDN64rr: case X86::ANDN64rm:
3461 case X86::BLSI32rr: case X86::BLSI32rm:
3462 case X86::BLSI64rr: case X86::BLSI64rm:
3463 case X86::BLSMSK32rr:case X86::BLSMSK32rm:
3464 case X86::BLSMSK64rr:case X86::BLSMSK64rm:
3465 case X86::BLSR32rr: case X86::BLSR32rm:
3466 case X86::BLSR64rr: case X86::BLSR64rm:
3467 case X86::BZHI32rr: case X86::BZHI32rm:
3468 case X86::BZHI64rr: case X86::BZHI64rm:
3469 case X86::LZCNT16rr: case X86::LZCNT16rm:
3470 case X86::LZCNT32rr: case X86::LZCNT32rm:
3471 case X86::LZCNT64rr: case X86::LZCNT64rm:
3472 case X86::POPCNT16rr:case X86::POPCNT16rm:
3473 case X86::POPCNT32rr:case X86::POPCNT32rm:
3474 case X86::POPCNT64rr:case X86::POPCNT64rm:
3475 case X86::TZCNT16rr: case X86::TZCNT16rm:
3476 case X86::TZCNT32rr: case X86::TZCNT32rm:
3477 case X86::TZCNT64rr: case X86::TZCNT64rm:
3478 case X86::BLCFILL32rr: case X86::BLCFILL32rm:
3479 case X86::BLCFILL64rr: case X86::BLCFILL64rm:
3480 case X86::BLCI32rr: case X86::BLCI32rm:
3481 case X86::BLCI64rr: case X86::BLCI64rm:
3482 case X86::BLCIC32rr: case X86::BLCIC32rm:
3483 case X86::BLCIC64rr: case X86::BLCIC64rm:
3484 case X86::BLCMSK32rr: case X86::BLCMSK32rm:
3485 case X86::BLCMSK64rr: case X86::BLCMSK64rm:
3486 case X86::BLCS32rr: case X86::BLCS32rm:
3487 case X86::BLCS64rr: case X86::BLCS64rm:
3488 case X86::BLSFILL32rr: case X86::BLSFILL32rm:
3489 case X86::BLSFILL64rr: case X86::BLSFILL64rm:
3490 case X86::BLSIC32rr: case X86::BLSIC32rm:
3491 case X86::BLSIC64rr: case X86::BLSIC64rm:
3492 case X86::T1MSKC32rr: case X86::T1MSKC32rm:
3493 case X86::T1MSKC64rr: case X86::T1MSKC64rm:
3494 case X86::TZMSK32rr: case X86::TZMSK32rm:
3495 case X86::TZMSK64rr: case X86::TZMSK64rm:
3496 return true;
3497 case X86::BEXTR32rr: case X86::BEXTR64rr:
3498 case X86::BEXTR32rm: case X86::BEXTR64rm:
3499 case X86::BEXTRI32ri: case X86::BEXTRI32mi:
3500 case X86::BEXTRI64ri: case X86::BEXTRI64mi:
3501 // BEXTR doesn't update the sign flag so we can't use it.
3502 NoSignFlag = true;
3503 return true;
3504 }
3505 }
3506
3507 /// Check whether the use can be converted to remove a comparison against zero.
isUseDefConvertible(const MachineInstr & MI)3508 static X86::CondCode isUseDefConvertible(const MachineInstr &MI) {
3509 switch (MI.getOpcode()) {
3510 default: return X86::COND_INVALID;
3511 case X86::NEG8r:
3512 case X86::NEG16r:
3513 case X86::NEG32r:
3514 case X86::NEG64r:
3515 return X86::COND_AE;
3516 case X86::LZCNT16rr:
3517 case X86::LZCNT32rr:
3518 case X86::LZCNT64rr:
3519 return X86::COND_B;
3520 case X86::POPCNT16rr:
3521 case X86::POPCNT32rr:
3522 case X86::POPCNT64rr:
3523 return X86::COND_E;
3524 case X86::TZCNT16rr:
3525 case X86::TZCNT32rr:
3526 case X86::TZCNT64rr:
3527 return X86::COND_B;
3528 case X86::BSF16rr:
3529 case X86::BSF32rr:
3530 case X86::BSF64rr:
3531 case X86::BSR16rr:
3532 case X86::BSR32rr:
3533 case X86::BSR64rr:
3534 return X86::COND_E;
3535 case X86::BLSI32rr:
3536 case X86::BLSI64rr:
3537 return X86::COND_AE;
3538 case X86::BLSR32rr:
3539 case X86::BLSR64rr:
3540 case X86::BLSMSK32rr:
3541 case X86::BLSMSK64rr:
3542 return X86::COND_B;
3543 // TODO: TBM instructions.
3544 }
3545 }
3546
3547 /// Check if there exists an earlier instruction that
3548 /// operates on the same source operands and sets flags in the same way as
3549 /// Compare; remove Compare if possible.
optimizeCompareInstr(MachineInstr & CmpInstr,unsigned SrcReg,unsigned SrcReg2,int CmpMask,int CmpValue,const MachineRegisterInfo * MRI) const3550 bool X86InstrInfo::optimizeCompareInstr(MachineInstr &CmpInstr, unsigned SrcReg,
3551 unsigned SrcReg2, int CmpMask,
3552 int CmpValue,
3553 const MachineRegisterInfo *MRI) const {
3554 // Check whether we can replace SUB with CMP.
3555 switch (CmpInstr.getOpcode()) {
3556 default: break;
3557 case X86::SUB64ri32:
3558 case X86::SUB64ri8:
3559 case X86::SUB32ri:
3560 case X86::SUB32ri8:
3561 case X86::SUB16ri:
3562 case X86::SUB16ri8:
3563 case X86::SUB8ri:
3564 case X86::SUB64rm:
3565 case X86::SUB32rm:
3566 case X86::SUB16rm:
3567 case X86::SUB8rm:
3568 case X86::SUB64rr:
3569 case X86::SUB32rr:
3570 case X86::SUB16rr:
3571 case X86::SUB8rr: {
3572 if (!MRI->use_nodbg_empty(CmpInstr.getOperand(0).getReg()))
3573 return false;
3574 // There is no use of the destination register, we can replace SUB with CMP.
3575 unsigned NewOpcode = 0;
3576 switch (CmpInstr.getOpcode()) {
3577 default: llvm_unreachable("Unreachable!");
3578 case X86::SUB64rm: NewOpcode = X86::CMP64rm; break;
3579 case X86::SUB32rm: NewOpcode = X86::CMP32rm; break;
3580 case X86::SUB16rm: NewOpcode = X86::CMP16rm; break;
3581 case X86::SUB8rm: NewOpcode = X86::CMP8rm; break;
3582 case X86::SUB64rr: NewOpcode = X86::CMP64rr; break;
3583 case X86::SUB32rr: NewOpcode = X86::CMP32rr; break;
3584 case X86::SUB16rr: NewOpcode = X86::CMP16rr; break;
3585 case X86::SUB8rr: NewOpcode = X86::CMP8rr; break;
3586 case X86::SUB64ri32: NewOpcode = X86::CMP64ri32; break;
3587 case X86::SUB64ri8: NewOpcode = X86::CMP64ri8; break;
3588 case X86::SUB32ri: NewOpcode = X86::CMP32ri; break;
3589 case X86::SUB32ri8: NewOpcode = X86::CMP32ri8; break;
3590 case X86::SUB16ri: NewOpcode = X86::CMP16ri; break;
3591 case X86::SUB16ri8: NewOpcode = X86::CMP16ri8; break;
3592 case X86::SUB8ri: NewOpcode = X86::CMP8ri; break;
3593 }
3594 CmpInstr.setDesc(get(NewOpcode));
3595 CmpInstr.RemoveOperand(0);
3596 // Fall through to optimize Cmp if Cmp is CMPrr or CMPri.
3597 if (NewOpcode == X86::CMP64rm || NewOpcode == X86::CMP32rm ||
3598 NewOpcode == X86::CMP16rm || NewOpcode == X86::CMP8rm)
3599 return false;
3600 }
3601 }
3602
3603 // Get the unique definition of SrcReg.
3604 MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg);
3605 if (!MI) return false;
3606
3607 // CmpInstr is the first instruction of the BB.
3608 MachineBasicBlock::iterator I = CmpInstr, Def = MI;
3609
3610 // If we are comparing against zero, check whether we can use MI to update
3611 // EFLAGS. If MI is not in the same BB as CmpInstr, do not optimize.
3612 bool IsCmpZero = (CmpMask != 0 && CmpValue == 0);
3613 if (IsCmpZero && MI->getParent() != CmpInstr.getParent())
3614 return false;
3615
3616 // If we have a use of the source register between the def and our compare
3617 // instruction we can eliminate the compare iff the use sets EFLAGS in the
3618 // right way.
3619 bool ShouldUpdateCC = false;
3620 bool NoSignFlag = false;
3621 X86::CondCode NewCC = X86::COND_INVALID;
3622 if (IsCmpZero && !isDefConvertible(*MI, NoSignFlag)) {
3623 // Scan forward from the use until we hit the use we're looking for or the
3624 // compare instruction.
3625 for (MachineBasicBlock::iterator J = MI;; ++J) {
3626 // Do we have a convertible instruction?
3627 NewCC = isUseDefConvertible(*J);
3628 if (NewCC != X86::COND_INVALID && J->getOperand(1).isReg() &&
3629 J->getOperand(1).getReg() == SrcReg) {
3630 assert(J->definesRegister(X86::EFLAGS) && "Must be an EFLAGS def!");
3631 ShouldUpdateCC = true; // Update CC later on.
3632 // This is not a def of SrcReg, but still a def of EFLAGS. Keep going
3633 // with the new def.
3634 Def = J;
3635 MI = &*Def;
3636 break;
3637 }
3638
3639 if (J == I)
3640 return false;
3641 }
3642 }
3643
3644 // We are searching for an earlier instruction that can make CmpInstr
3645 // redundant and that instruction will be saved in Sub.
3646 MachineInstr *Sub = nullptr;
3647 const TargetRegisterInfo *TRI = &getRegisterInfo();
3648
3649 // We iterate backward, starting from the instruction before CmpInstr and
3650 // stop when reaching the definition of a source register or done with the BB.
3651 // RI points to the instruction before CmpInstr.
3652 // If the definition is in this basic block, RE points to the definition;
3653 // otherwise, RE is the rend of the basic block.
3654 MachineBasicBlock::reverse_iterator
3655 RI = ++I.getReverse(),
3656 RE = CmpInstr.getParent() == MI->getParent()
3657 ? Def.getReverse() /* points to MI */
3658 : CmpInstr.getParent()->rend();
3659 MachineInstr *Movr0Inst = nullptr;
3660 for (; RI != RE; ++RI) {
3661 MachineInstr &Instr = *RI;
3662 // Check whether CmpInstr can be made redundant by the current instruction.
3663 if (!IsCmpZero && isRedundantFlagInstr(CmpInstr, SrcReg, SrcReg2, CmpMask,
3664 CmpValue, Instr)) {
3665 Sub = &Instr;
3666 break;
3667 }
3668
3669 if (Instr.modifiesRegister(X86::EFLAGS, TRI) ||
3670 Instr.readsRegister(X86::EFLAGS, TRI)) {
3671 // This instruction modifies or uses EFLAGS.
3672
3673 // MOV32r0 etc. are implemented with xor which clobbers condition code.
3674 // They are safe to move up, if the definition to EFLAGS is dead and
3675 // earlier instructions do not read or write EFLAGS.
3676 if (!Movr0Inst && Instr.getOpcode() == X86::MOV32r0 &&
3677 Instr.registerDefIsDead(X86::EFLAGS, TRI)) {
3678 Movr0Inst = &Instr;
3679 continue;
3680 }
3681
3682 // We can't remove CmpInstr.
3683 return false;
3684 }
3685 }
3686
3687 // Return false if no candidates exist.
3688 if (!IsCmpZero && !Sub)
3689 return false;
3690
3691 bool IsSwapped =
3692 (SrcReg2 != 0 && Sub && Sub->getOperand(1).getReg() == SrcReg2 &&
3693 Sub->getOperand(2).getReg() == SrcReg);
3694
3695 // Scan forward from the instruction after CmpInstr for uses of EFLAGS.
3696 // It is safe to remove CmpInstr if EFLAGS is redefined or killed.
3697 // If we are done with the basic block, we need to check whether EFLAGS is
3698 // live-out.
3699 bool IsSafe = false;
3700 SmallVector<std::pair<MachineInstr*, X86::CondCode>, 4> OpsToUpdate;
3701 MachineBasicBlock::iterator E = CmpInstr.getParent()->end();
3702 for (++I; I != E; ++I) {
3703 const MachineInstr &Instr = *I;
3704 bool ModifyEFLAGS = Instr.modifiesRegister(X86::EFLAGS, TRI);
3705 bool UseEFLAGS = Instr.readsRegister(X86::EFLAGS, TRI);
3706 // We should check the usage if this instruction uses and updates EFLAGS.
3707 if (!UseEFLAGS && ModifyEFLAGS) {
3708 // It is safe to remove CmpInstr if EFLAGS is updated again.
3709 IsSafe = true;
3710 break;
3711 }
3712 if (!UseEFLAGS && !ModifyEFLAGS)
3713 continue;
3714
3715 // EFLAGS is used by this instruction.
3716 X86::CondCode OldCC = X86::COND_INVALID;
3717 if (IsCmpZero || IsSwapped) {
3718 // We decode the condition code from opcode.
3719 if (Instr.isBranch())
3720 OldCC = X86::getCondFromBranch(Instr);
3721 else {
3722 OldCC = X86::getCondFromSETCC(Instr);
3723 if (OldCC == X86::COND_INVALID)
3724 OldCC = X86::getCondFromCMov(Instr);
3725 }
3726 if (OldCC == X86::COND_INVALID) return false;
3727 }
3728 X86::CondCode ReplacementCC = X86::COND_INVALID;
3729 if (IsCmpZero) {
3730 switch (OldCC) {
3731 default: break;
3732 case X86::COND_A: case X86::COND_AE:
3733 case X86::COND_B: case X86::COND_BE:
3734 case X86::COND_G: case X86::COND_GE:
3735 case X86::COND_L: case X86::COND_LE:
3736 case X86::COND_O: case X86::COND_NO:
3737 // CF and OF are used, we can't perform this optimization.
3738 return false;
3739 case X86::COND_S: case X86::COND_NS:
3740 // If SF is used, but the instruction doesn't update the SF, then we
3741 // can't do the optimization.
3742 if (NoSignFlag)
3743 return false;
3744 break;
3745 }
3746
3747 // If we're updating the condition code check if we have to reverse the
3748 // condition.
3749 if (ShouldUpdateCC)
3750 switch (OldCC) {
3751 default:
3752 return false;
3753 case X86::COND_E:
3754 ReplacementCC = NewCC;
3755 break;
3756 case X86::COND_NE:
3757 ReplacementCC = GetOppositeBranchCondition(NewCC);
3758 break;
3759 }
3760 } else if (IsSwapped) {
3761 // If we have SUB(r1, r2) and CMP(r2, r1), the condition code needs
3762 // to be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc.
3763 // We swap the condition code and synthesize the new opcode.
3764 ReplacementCC = getSwappedCondition(OldCC);
3765 if (ReplacementCC == X86::COND_INVALID) return false;
3766 }
3767
3768 if ((ShouldUpdateCC || IsSwapped) && ReplacementCC != OldCC) {
3769 // Push the MachineInstr to OpsToUpdate.
3770 // If it is safe to remove CmpInstr, the condition code of these
3771 // instructions will be modified.
3772 OpsToUpdate.push_back(std::make_pair(&*I, ReplacementCC));
3773 }
3774 if (ModifyEFLAGS || Instr.killsRegister(X86::EFLAGS, TRI)) {
3775 // It is safe to remove CmpInstr if EFLAGS is updated again or killed.
3776 IsSafe = true;
3777 break;
3778 }
3779 }
3780
3781 // If EFLAGS is not killed nor re-defined, we should check whether it is
3782 // live-out. If it is live-out, do not optimize.
3783 if ((IsCmpZero || IsSwapped) && !IsSafe) {
3784 MachineBasicBlock *MBB = CmpInstr.getParent();
3785 for (MachineBasicBlock *Successor : MBB->successors())
3786 if (Successor->isLiveIn(X86::EFLAGS))
3787 return false;
3788 }
3789
3790 // The instruction to be updated is either Sub or MI.
3791 Sub = IsCmpZero ? MI : Sub;
3792 // Move Movr0Inst to the appropriate place before Sub.
3793 if (Movr0Inst) {
3794 // Look backwards until we find a def that doesn't use the current EFLAGS.
3795 Def = Sub;
3796 MachineBasicBlock::reverse_iterator InsertI = Def.getReverse(),
3797 InsertE = Sub->getParent()->rend();
3798 for (; InsertI != InsertE; ++InsertI) {
3799 MachineInstr *Instr = &*InsertI;
3800 if (!Instr->readsRegister(X86::EFLAGS, TRI) &&
3801 Instr->modifiesRegister(X86::EFLAGS, TRI)) {
3802 Sub->getParent()->remove(Movr0Inst);
3803 Instr->getParent()->insert(MachineBasicBlock::iterator(Instr),
3804 Movr0Inst);
3805 break;
3806 }
3807 }
3808 if (InsertI == InsertE)
3809 return false;
3810 }
3811
3812 // Make sure Sub instruction defines EFLAGS and mark the def live.
3813 MachineOperand *FlagDef = Sub->findRegisterDefOperand(X86::EFLAGS);
3814 assert(FlagDef && "Unable to locate a def EFLAGS operand");
3815 FlagDef->setIsDead(false);
3816
3817 CmpInstr.eraseFromParent();
3818
3819 // Modify the condition code of instructions in OpsToUpdate.
3820 for (auto &Op : OpsToUpdate) {
3821 Op.first->getOperand(Op.first->getDesc().getNumOperands() - 1)
3822 .setImm(Op.second);
3823 }
3824 return true;
3825 }
3826
3827 /// Try to remove the load by folding it to a register
3828 /// operand at the use. We fold the load instructions if load defines a virtual
3829 /// register, the virtual register is used once in the same BB, and the
3830 /// instructions in-between do not load or store, and have no side effects.
optimizeLoadInstr(MachineInstr & MI,const MachineRegisterInfo * MRI,unsigned & FoldAsLoadDefReg,MachineInstr * & DefMI) const3831 MachineInstr *X86InstrInfo::optimizeLoadInstr(MachineInstr &MI,
3832 const MachineRegisterInfo *MRI,
3833 unsigned &FoldAsLoadDefReg,
3834 MachineInstr *&DefMI) const {
3835 // Check whether we can move DefMI here.
3836 DefMI = MRI->getVRegDef(FoldAsLoadDefReg);
3837 assert(DefMI);
3838 bool SawStore = false;
3839 if (!DefMI->isSafeToMove(nullptr, SawStore))
3840 return nullptr;
3841
3842 // Collect information about virtual register operands of MI.
3843 SmallVector<unsigned, 1> SrcOperandIds;
3844 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
3845 MachineOperand &MO = MI.getOperand(i);
3846 if (!MO.isReg())
3847 continue;
3848 Register Reg = MO.getReg();
3849 if (Reg != FoldAsLoadDefReg)
3850 continue;
3851 // Do not fold if we have a subreg use or a def.
3852 if (MO.getSubReg() || MO.isDef())
3853 return nullptr;
3854 SrcOperandIds.push_back(i);
3855 }
3856 if (SrcOperandIds.empty())
3857 return nullptr;
3858
3859 // Check whether we can fold the def into SrcOperandId.
3860 if (MachineInstr *FoldMI = foldMemoryOperand(MI, SrcOperandIds, *DefMI)) {
3861 FoldAsLoadDefReg = 0;
3862 return FoldMI;
3863 }
3864
3865 return nullptr;
3866 }
3867
3868 /// Expand a single-def pseudo instruction to a two-addr
3869 /// instruction with two undef reads of the register being defined.
3870 /// This is used for mapping:
3871 /// %xmm4 = V_SET0
3872 /// to:
3873 /// %xmm4 = PXORrr undef %xmm4, undef %xmm4
3874 ///
Expand2AddrUndef(MachineInstrBuilder & MIB,const MCInstrDesc & Desc)3875 static bool Expand2AddrUndef(MachineInstrBuilder &MIB,
3876 const MCInstrDesc &Desc) {
3877 assert(Desc.getNumOperands() == 3 && "Expected two-addr instruction.");
3878 Register Reg = MIB->getOperand(0).getReg();
3879 MIB->setDesc(Desc);
3880
3881 // MachineInstr::addOperand() will insert explicit operands before any
3882 // implicit operands.
3883 MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
3884 // But we don't trust that.
3885 assert(MIB->getOperand(1).getReg() == Reg &&
3886 MIB->getOperand(2).getReg() == Reg && "Misplaced operand");
3887 return true;
3888 }
3889
3890 /// Expand a single-def pseudo instruction to a two-addr
3891 /// instruction with two %k0 reads.
3892 /// This is used for mapping:
3893 /// %k4 = K_SET1
3894 /// to:
3895 /// %k4 = KXNORrr %k0, %k0
Expand2AddrKreg(MachineInstrBuilder & MIB,const MCInstrDesc & Desc,unsigned Reg)3896 static bool Expand2AddrKreg(MachineInstrBuilder &MIB,
3897 const MCInstrDesc &Desc, unsigned Reg) {
3898 assert(Desc.getNumOperands() == 3 && "Expected two-addr instruction.");
3899 MIB->setDesc(Desc);
3900 MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
3901 return true;
3902 }
3903
expandMOV32r1(MachineInstrBuilder & MIB,const TargetInstrInfo & TII,bool MinusOne)3904 static bool expandMOV32r1(MachineInstrBuilder &MIB, const TargetInstrInfo &TII,
3905 bool MinusOne) {
3906 MachineBasicBlock &MBB = *MIB->getParent();
3907 DebugLoc DL = MIB->getDebugLoc();
3908 Register Reg = MIB->getOperand(0).getReg();
3909
3910 // Insert the XOR.
3911 BuildMI(MBB, MIB.getInstr(), DL, TII.get(X86::XOR32rr), Reg)
3912 .addReg(Reg, RegState::Undef)
3913 .addReg(Reg, RegState::Undef);
3914
3915 // Turn the pseudo into an INC or DEC.
3916 MIB->setDesc(TII.get(MinusOne ? X86::DEC32r : X86::INC32r));
3917 MIB.addReg(Reg);
3918
3919 return true;
3920 }
3921
ExpandMOVImmSExti8(MachineInstrBuilder & MIB,const TargetInstrInfo & TII,const X86Subtarget & Subtarget)3922 static bool ExpandMOVImmSExti8(MachineInstrBuilder &MIB,
3923 const TargetInstrInfo &TII,
3924 const X86Subtarget &Subtarget) {
3925 MachineBasicBlock &MBB = *MIB->getParent();
3926 DebugLoc DL = MIB->getDebugLoc();
3927 int64_t Imm = MIB->getOperand(1).getImm();
3928 assert(Imm != 0 && "Using push/pop for 0 is not efficient.");
3929 MachineBasicBlock::iterator I = MIB.getInstr();
3930
3931 int StackAdjustment;
3932
3933 if (Subtarget.is64Bit()) {
3934 assert(MIB->getOpcode() == X86::MOV64ImmSExti8 ||
3935 MIB->getOpcode() == X86::MOV32ImmSExti8);
3936
3937 // Can't use push/pop lowering if the function might write to the red zone.
3938 X86MachineFunctionInfo *X86FI =
3939 MBB.getParent()->getInfo<X86MachineFunctionInfo>();
3940 if (X86FI->getUsesRedZone()) {
3941 MIB->setDesc(TII.get(MIB->getOpcode() ==
3942 X86::MOV32ImmSExti8 ? X86::MOV32ri : X86::MOV64ri));
3943 return true;
3944 }
3945
3946 // 64-bit mode doesn't have 32-bit push/pop, so use 64-bit operations and
3947 // widen the register if necessary.
3948 StackAdjustment = 8;
3949 BuildMI(MBB, I, DL, TII.get(X86::PUSH64i8)).addImm(Imm);
3950 MIB->setDesc(TII.get(X86::POP64r));
3951 MIB->getOperand(0)
3952 .setReg(getX86SubSuperRegister(MIB->getOperand(0).getReg(), 64));
3953 } else {
3954 assert(MIB->getOpcode() == X86::MOV32ImmSExti8);
3955 StackAdjustment = 4;
3956 BuildMI(MBB, I, DL, TII.get(X86::PUSH32i8)).addImm(Imm);
3957 MIB->setDesc(TII.get(X86::POP32r));
3958 }
3959
3960 // Build CFI if necessary.
3961 MachineFunction &MF = *MBB.getParent();
3962 const X86FrameLowering *TFL = Subtarget.getFrameLowering();
3963 bool IsWin64Prologue = MF.getTarget().getMCAsmInfo()->usesWindowsCFI();
3964 bool NeedsDwarfCFI = !IsWin64Prologue && MF.needsFrameMoves();
3965 bool EmitCFI = !TFL->hasFP(MF) && NeedsDwarfCFI;
3966 if (EmitCFI) {
3967 TFL->BuildCFI(MBB, I, DL,
3968 MCCFIInstruction::createAdjustCfaOffset(nullptr, StackAdjustment));
3969 TFL->BuildCFI(MBB, std::next(I), DL,
3970 MCCFIInstruction::createAdjustCfaOffset(nullptr, -StackAdjustment));
3971 }
3972
3973 return true;
3974 }
3975
3976 // LoadStackGuard has so far only been implemented for 64-bit MachO. Different
3977 // code sequence is needed for other targets.
expandLoadStackGuard(MachineInstrBuilder & MIB,const TargetInstrInfo & TII)3978 static void expandLoadStackGuard(MachineInstrBuilder &MIB,
3979 const TargetInstrInfo &TII) {
3980 MachineBasicBlock &MBB = *MIB->getParent();
3981 DebugLoc DL = MIB->getDebugLoc();
3982 Register Reg = MIB->getOperand(0).getReg();
3983 const GlobalValue *GV =
3984 cast<GlobalValue>((*MIB->memoperands_begin())->getValue());
3985 auto Flags = MachineMemOperand::MOLoad |
3986 MachineMemOperand::MODereferenceable |
3987 MachineMemOperand::MOInvariant;
3988 MachineMemOperand *MMO = MBB.getParent()->getMachineMemOperand(
3989 MachinePointerInfo::getGOT(*MBB.getParent()), Flags, 8, 8);
3990 MachineBasicBlock::iterator I = MIB.getInstr();
3991
3992 BuildMI(MBB, I, DL, TII.get(X86::MOV64rm), Reg).addReg(X86::RIP).addImm(1)
3993 .addReg(0).addGlobalAddress(GV, 0, X86II::MO_GOTPCREL).addReg(0)
3994 .addMemOperand(MMO);
3995 MIB->setDebugLoc(DL);
3996 MIB->setDesc(TII.get(X86::MOV64rm));
3997 MIB.addReg(Reg, RegState::Kill).addImm(1).addReg(0).addImm(0).addReg(0);
3998 }
3999
expandXorFP(MachineInstrBuilder & MIB,const TargetInstrInfo & TII)4000 static bool expandXorFP(MachineInstrBuilder &MIB, const TargetInstrInfo &TII) {
4001 MachineBasicBlock &MBB = *MIB->getParent();
4002 MachineFunction &MF = *MBB.getParent();
4003 const X86Subtarget &Subtarget = MF.getSubtarget<X86Subtarget>();
4004 const X86RegisterInfo *TRI = Subtarget.getRegisterInfo();
4005 unsigned XorOp =
4006 MIB->getOpcode() == X86::XOR64_FP ? X86::XOR64rr : X86::XOR32rr;
4007 MIB->setDesc(TII.get(XorOp));
4008 MIB.addReg(TRI->getFrameRegister(MF), RegState::Undef);
4009 return true;
4010 }
4011
4012 // This is used to handle spills for 128/256-bit registers when we have AVX512,
4013 // but not VLX. If it uses an extended register we need to use an instruction
4014 // 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)4015 static bool expandNOVLXLoad(MachineInstrBuilder &MIB,
4016 const TargetRegisterInfo *TRI,
4017 const MCInstrDesc &LoadDesc,
4018 const MCInstrDesc &BroadcastDesc,
4019 unsigned SubIdx) {
4020 Register DestReg = MIB->getOperand(0).getReg();
4021 // Check if DestReg is XMM16-31 or YMM16-31.
4022 if (TRI->getEncodingValue(DestReg) < 16) {
4023 // We can use a normal VEX encoded load.
4024 MIB->setDesc(LoadDesc);
4025 } else {
4026 // Use a 128/256-bit VBROADCAST instruction.
4027 MIB->setDesc(BroadcastDesc);
4028 // Change the destination to a 512-bit register.
4029 DestReg = TRI->getMatchingSuperReg(DestReg, SubIdx, &X86::VR512RegClass);
4030 MIB->getOperand(0).setReg(DestReg);
4031 }
4032 return true;
4033 }
4034
4035 // This is used to handle spills for 128/256-bit registers when we have AVX512,
4036 // but not VLX. If it uses an extended register we need to use an instruction
4037 // 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)4038 static bool expandNOVLXStore(MachineInstrBuilder &MIB,
4039 const TargetRegisterInfo *TRI,
4040 const MCInstrDesc &StoreDesc,
4041 const MCInstrDesc &ExtractDesc,
4042 unsigned SubIdx) {
4043 Register SrcReg = MIB->getOperand(X86::AddrNumOperands).getReg();
4044 // Check if DestReg is XMM16-31 or YMM16-31.
4045 if (TRI->getEncodingValue(SrcReg) < 16) {
4046 // We can use a normal VEX encoded store.
4047 MIB->setDesc(StoreDesc);
4048 } else {
4049 // Use a VEXTRACTF instruction.
4050 MIB->setDesc(ExtractDesc);
4051 // Change the destination to a 512-bit register.
4052 SrcReg = TRI->getMatchingSuperReg(SrcReg, SubIdx, &X86::VR512RegClass);
4053 MIB->getOperand(X86::AddrNumOperands).setReg(SrcReg);
4054 MIB.addImm(0x0); // Append immediate to extract from the lower bits.
4055 }
4056
4057 return true;
4058 }
4059
expandSHXDROT(MachineInstrBuilder & MIB,const MCInstrDesc & Desc)4060 static bool expandSHXDROT(MachineInstrBuilder &MIB, const MCInstrDesc &Desc) {
4061 MIB->setDesc(Desc);
4062 int64_t ShiftAmt = MIB->getOperand(2).getImm();
4063 // Temporarily remove the immediate so we can add another source register.
4064 MIB->RemoveOperand(2);
4065 // Add the register. Don't copy the kill flag if there is one.
4066 MIB.addReg(MIB->getOperand(1).getReg(),
4067 getUndefRegState(MIB->getOperand(1).isUndef()));
4068 // Add back the immediate.
4069 MIB.addImm(ShiftAmt);
4070 return true;
4071 }
4072
expandPostRAPseudo(MachineInstr & MI) const4073 bool X86InstrInfo::expandPostRAPseudo(MachineInstr &MI) const {
4074 bool HasAVX = Subtarget.hasAVX();
4075 MachineInstrBuilder MIB(*MI.getParent()->getParent(), MI);
4076 switch (MI.getOpcode()) {
4077 case X86::MOV32r0:
4078 return Expand2AddrUndef(MIB, get(X86::XOR32rr));
4079 case X86::MOV32r1:
4080 return expandMOV32r1(MIB, *this, /*MinusOne=*/ false);
4081 case X86::MOV32r_1:
4082 return expandMOV32r1(MIB, *this, /*MinusOne=*/ true);
4083 case X86::MOV32ImmSExti8:
4084 case X86::MOV64ImmSExti8:
4085 return ExpandMOVImmSExti8(MIB, *this, Subtarget);
4086 case X86::SETB_C8r:
4087 return Expand2AddrUndef(MIB, get(X86::SBB8rr));
4088 case X86::SETB_C16r:
4089 return Expand2AddrUndef(MIB, get(X86::SBB16rr));
4090 case X86::SETB_C32r:
4091 return Expand2AddrUndef(MIB, get(X86::SBB32rr));
4092 case X86::SETB_C64r:
4093 return Expand2AddrUndef(MIB, get(X86::SBB64rr));
4094 case X86::MMX_SET0:
4095 return Expand2AddrUndef(MIB, get(X86::MMX_PXORirr));
4096 case X86::V_SET0:
4097 case X86::FsFLD0SS:
4098 case X86::FsFLD0SD:
4099 case X86::FsFLD0F128:
4100 return Expand2AddrUndef(MIB, get(HasAVX ? X86::VXORPSrr : X86::XORPSrr));
4101 case X86::AVX_SET0: {
4102 assert(HasAVX && "AVX not supported");
4103 const TargetRegisterInfo *TRI = &getRegisterInfo();
4104 Register SrcReg = MIB->getOperand(0).getReg();
4105 Register XReg = TRI->getSubReg(SrcReg, X86::sub_xmm);
4106 MIB->getOperand(0).setReg(XReg);
4107 Expand2AddrUndef(MIB, get(X86::VXORPSrr));
4108 MIB.addReg(SrcReg, RegState::ImplicitDefine);
4109 return true;
4110 }
4111 case X86::AVX512_128_SET0:
4112 case X86::AVX512_FsFLD0SS:
4113 case X86::AVX512_FsFLD0SD:
4114 case X86::AVX512_FsFLD0F128: {
4115 bool HasVLX = Subtarget.hasVLX();
4116 Register SrcReg = MIB->getOperand(0).getReg();
4117 const TargetRegisterInfo *TRI = &getRegisterInfo();
4118 if (HasVLX || TRI->getEncodingValue(SrcReg) < 16)
4119 return Expand2AddrUndef(MIB,
4120 get(HasVLX ? X86::VPXORDZ128rr : X86::VXORPSrr));
4121 // Extended register without VLX. Use a larger XOR.
4122 SrcReg =
4123 TRI->getMatchingSuperReg(SrcReg, X86::sub_xmm, &X86::VR512RegClass);
4124 MIB->getOperand(0).setReg(SrcReg);
4125 return Expand2AddrUndef(MIB, get(X86::VPXORDZrr));
4126 }
4127 case X86::AVX512_256_SET0:
4128 case X86::AVX512_512_SET0: {
4129 bool HasVLX = Subtarget.hasVLX();
4130 Register SrcReg = MIB->getOperand(0).getReg();
4131 const TargetRegisterInfo *TRI = &getRegisterInfo();
4132 if (HasVLX || TRI->getEncodingValue(SrcReg) < 16) {
4133 Register XReg = TRI->getSubReg(SrcReg, X86::sub_xmm);
4134 MIB->getOperand(0).setReg(XReg);
4135 Expand2AddrUndef(MIB,
4136 get(HasVLX ? X86::VPXORDZ128rr : X86::VXORPSrr));
4137 MIB.addReg(SrcReg, RegState::ImplicitDefine);
4138 return true;
4139 }
4140 if (MI.getOpcode() == X86::AVX512_256_SET0) {
4141 // No VLX so we must reference a zmm.
4142 unsigned ZReg =
4143 TRI->getMatchingSuperReg(SrcReg, X86::sub_ymm, &X86::VR512RegClass);
4144 MIB->getOperand(0).setReg(ZReg);
4145 }
4146 return Expand2AddrUndef(MIB, get(X86::VPXORDZrr));
4147 }
4148 case X86::V_SETALLONES:
4149 return Expand2AddrUndef(MIB, get(HasAVX ? X86::VPCMPEQDrr : X86::PCMPEQDrr));
4150 case X86::AVX2_SETALLONES:
4151 return Expand2AddrUndef(MIB, get(X86::VPCMPEQDYrr));
4152 case X86::AVX1_SETALLONES: {
4153 Register Reg = MIB->getOperand(0).getReg();
4154 // VCMPPSYrri with an immediate 0xf should produce VCMPTRUEPS.
4155 MIB->setDesc(get(X86::VCMPPSYrri));
4156 MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef).addImm(0xf);
4157 return true;
4158 }
4159 case X86::AVX512_512_SETALLONES: {
4160 Register Reg = MIB->getOperand(0).getReg();
4161 MIB->setDesc(get(X86::VPTERNLOGDZrri));
4162 // VPTERNLOGD needs 3 register inputs and an immediate.
4163 // 0xff will return 1s for any input.
4164 MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef)
4165 .addReg(Reg, RegState::Undef).addImm(0xff);
4166 return true;
4167 }
4168 case X86::AVX512_512_SEXT_MASK_32:
4169 case X86::AVX512_512_SEXT_MASK_64: {
4170 Register Reg = MIB->getOperand(0).getReg();
4171 Register MaskReg = MIB->getOperand(1).getReg();
4172 unsigned MaskState = getRegState(MIB->getOperand(1));
4173 unsigned Opc = (MI.getOpcode() == X86::AVX512_512_SEXT_MASK_64) ?
4174 X86::VPTERNLOGQZrrikz : X86::VPTERNLOGDZrrikz;
4175 MI.RemoveOperand(1);
4176 MIB->setDesc(get(Opc));
4177 // VPTERNLOG needs 3 register inputs and an immediate.
4178 // 0xff will return 1s for any input.
4179 MIB.addReg(Reg, RegState::Undef).addReg(MaskReg, MaskState)
4180 .addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef).addImm(0xff);
4181 return true;
4182 }
4183 case X86::VMOVAPSZ128rm_NOVLX:
4184 return expandNOVLXLoad(MIB, &getRegisterInfo(), get(X86::VMOVAPSrm),
4185 get(X86::VBROADCASTF32X4rm), X86::sub_xmm);
4186 case X86::VMOVUPSZ128rm_NOVLX:
4187 return expandNOVLXLoad(MIB, &getRegisterInfo(), get(X86::VMOVUPSrm),
4188 get(X86::VBROADCASTF32X4rm), X86::sub_xmm);
4189 case X86::VMOVAPSZ256rm_NOVLX:
4190 return expandNOVLXLoad(MIB, &getRegisterInfo(), get(X86::VMOVAPSYrm),
4191 get(X86::VBROADCASTF64X4rm), X86::sub_ymm);
4192 case X86::VMOVUPSZ256rm_NOVLX:
4193 return expandNOVLXLoad(MIB, &getRegisterInfo(), get(X86::VMOVUPSYrm),
4194 get(X86::VBROADCASTF64X4rm), X86::sub_ymm);
4195 case X86::VMOVAPSZ128mr_NOVLX:
4196 return expandNOVLXStore(MIB, &getRegisterInfo(), get(X86::VMOVAPSmr),
4197 get(X86::VEXTRACTF32x4Zmr), X86::sub_xmm);
4198 case X86::VMOVUPSZ128mr_NOVLX:
4199 return expandNOVLXStore(MIB, &getRegisterInfo(), get(X86::VMOVUPSmr),
4200 get(X86::VEXTRACTF32x4Zmr), X86::sub_xmm);
4201 case X86::VMOVAPSZ256mr_NOVLX:
4202 return expandNOVLXStore(MIB, &getRegisterInfo(), get(X86::VMOVAPSYmr),
4203 get(X86::VEXTRACTF64x4Zmr), X86::sub_ymm);
4204 case X86::VMOVUPSZ256mr_NOVLX:
4205 return expandNOVLXStore(MIB, &getRegisterInfo(), get(X86::VMOVUPSYmr),
4206 get(X86::VEXTRACTF64x4Zmr), X86::sub_ymm);
4207 case X86::MOV32ri64: {
4208 Register Reg = MIB->getOperand(0).getReg();
4209 Register Reg32 = RI.getSubReg(Reg, X86::sub_32bit);
4210 MI.setDesc(get(X86::MOV32ri));
4211 MIB->getOperand(0).setReg(Reg32);
4212 MIB.addReg(Reg, RegState::ImplicitDefine);
4213 return true;
4214 }
4215
4216 // KNL does not recognize dependency-breaking idioms for mask registers,
4217 // so kxnor %k1, %k1, %k2 has a RAW dependence on %k1.
4218 // Using %k0 as the undef input register is a performance heuristic based
4219 // on the assumption that %k0 is used less frequently than the other mask
4220 // registers, since it is not usable as a write mask.
4221 // FIXME: A more advanced approach would be to choose the best input mask
4222 // register based on context.
4223 case X86::KSET0W: return Expand2AddrKreg(MIB, get(X86::KXORWrr), X86::K0);
4224 case X86::KSET0D: return Expand2AddrKreg(MIB, get(X86::KXORDrr), X86::K0);
4225 case X86::KSET0Q: return Expand2AddrKreg(MIB, get(X86::KXORQrr), X86::K0);
4226 case X86::KSET1W: return Expand2AddrKreg(MIB, get(X86::KXNORWrr), X86::K0);
4227 case X86::KSET1D: return Expand2AddrKreg(MIB, get(X86::KXNORDrr), X86::K0);
4228 case X86::KSET1Q: return Expand2AddrKreg(MIB, get(X86::KXNORQrr), X86::K0);
4229 case TargetOpcode::LOAD_STACK_GUARD:
4230 expandLoadStackGuard(MIB, *this);
4231 return true;
4232 case X86::XOR64_FP:
4233 case X86::XOR32_FP:
4234 return expandXorFP(MIB, *this);
4235 case X86::SHLDROT32ri: return expandSHXDROT(MIB, get(X86::SHLD32rri8));
4236 case X86::SHLDROT64ri: return expandSHXDROT(MIB, get(X86::SHLD64rri8));
4237 case X86::SHRDROT32ri: return expandSHXDROT(MIB, get(X86::SHRD32rri8));
4238 case X86::SHRDROT64ri: return expandSHXDROT(MIB, get(X86::SHRD64rri8));
4239 case X86::ADD8rr_DB: MIB->setDesc(get(X86::OR8rr)); break;
4240 case X86::ADD16rr_DB: MIB->setDesc(get(X86::OR16rr)); break;
4241 case X86::ADD32rr_DB: MIB->setDesc(get(X86::OR32rr)); break;
4242 case X86::ADD64rr_DB: MIB->setDesc(get(X86::OR64rr)); break;
4243 case X86::ADD8ri_DB: MIB->setDesc(get(X86::OR8ri)); break;
4244 case X86::ADD16ri_DB: MIB->setDesc(get(X86::OR16ri)); break;
4245 case X86::ADD32ri_DB: MIB->setDesc(get(X86::OR32ri)); break;
4246 case X86::ADD64ri32_DB: MIB->setDesc(get(X86::OR64ri32)); break;
4247 case X86::ADD16ri8_DB: MIB->setDesc(get(X86::OR16ri8)); break;
4248 case X86::ADD32ri8_DB: MIB->setDesc(get(X86::OR32ri8)); break;
4249 case X86::ADD64ri8_DB: MIB->setDesc(get(X86::OR64ri8)); break;
4250 }
4251 return false;
4252 }
4253
4254 /// Return true for all instructions that only update
4255 /// the first 32 or 64-bits of the destination register and leave the rest
4256 /// unmodified. This can be used to avoid folding loads if the instructions
4257 /// only update part of the destination register, and the non-updated part is
4258 /// not needed. e.g. cvtss2sd, sqrtss. Unfolding the load from these
4259 /// instructions breaks the partial register dependency and it can improve
4260 /// performance. e.g.:
4261 ///
4262 /// movss (%rdi), %xmm0
4263 /// cvtss2sd %xmm0, %xmm0
4264 ///
4265 /// Instead of
4266 /// cvtss2sd (%rdi), %xmm0
4267 ///
4268 /// FIXME: This should be turned into a TSFlags.
4269 ///
hasPartialRegUpdate(unsigned Opcode,const X86Subtarget & Subtarget,bool ForLoadFold=false)4270 static bool hasPartialRegUpdate(unsigned Opcode,
4271 const X86Subtarget &Subtarget,
4272 bool ForLoadFold = false) {
4273 switch (Opcode) {
4274 case X86::CVTSI2SSrr:
4275 case X86::CVTSI2SSrm:
4276 case X86::CVTSI642SSrr:
4277 case X86::CVTSI642SSrm:
4278 case X86::CVTSI2SDrr:
4279 case X86::CVTSI2SDrm:
4280 case X86::CVTSI642SDrr:
4281 case X86::CVTSI642SDrm:
4282 // Load folding won't effect the undef register update since the input is
4283 // a GPR.
4284 return !ForLoadFold;
4285 case X86::CVTSD2SSrr:
4286 case X86::CVTSD2SSrm:
4287 case X86::CVTSS2SDrr:
4288 case X86::CVTSS2SDrm:
4289 case X86::MOVHPDrm:
4290 case X86::MOVHPSrm:
4291 case X86::MOVLPDrm:
4292 case X86::MOVLPSrm:
4293 case X86::RCPSSr:
4294 case X86::RCPSSm:
4295 case X86::RCPSSr_Int:
4296 case X86::RCPSSm_Int:
4297 case X86::ROUNDSDr:
4298 case X86::ROUNDSDm:
4299 case X86::ROUNDSSr:
4300 case X86::ROUNDSSm:
4301 case X86::RSQRTSSr:
4302 case X86::RSQRTSSm:
4303 case X86::RSQRTSSr_Int:
4304 case X86::RSQRTSSm_Int:
4305 case X86::SQRTSSr:
4306 case X86::SQRTSSm:
4307 case X86::SQRTSSr_Int:
4308 case X86::SQRTSSm_Int:
4309 case X86::SQRTSDr:
4310 case X86::SQRTSDm:
4311 case X86::SQRTSDr_Int:
4312 case X86::SQRTSDm_Int:
4313 return true;
4314 // GPR
4315 case X86::POPCNT32rm:
4316 case X86::POPCNT32rr:
4317 case X86::POPCNT64rm:
4318 case X86::POPCNT64rr:
4319 return Subtarget.hasPOPCNTFalseDeps();
4320 case X86::LZCNT32rm:
4321 case X86::LZCNT32rr:
4322 case X86::LZCNT64rm:
4323 case X86::LZCNT64rr:
4324 case X86::TZCNT32rm:
4325 case X86::TZCNT32rr:
4326 case X86::TZCNT64rm:
4327 case X86::TZCNT64rr:
4328 return Subtarget.hasLZCNTFalseDeps();
4329 }
4330
4331 return false;
4332 }
4333
4334 /// Inform the BreakFalseDeps pass how many idle
4335 /// instructions we would like before a partial register update.
getPartialRegUpdateClearance(const MachineInstr & MI,unsigned OpNum,const TargetRegisterInfo * TRI) const4336 unsigned X86InstrInfo::getPartialRegUpdateClearance(
4337 const MachineInstr &MI, unsigned OpNum,
4338 const TargetRegisterInfo *TRI) const {
4339 if (OpNum != 0 || !hasPartialRegUpdate(MI.getOpcode(), Subtarget))
4340 return 0;
4341
4342 // If MI is marked as reading Reg, the partial register update is wanted.
4343 const MachineOperand &MO = MI.getOperand(0);
4344 Register Reg = MO.getReg();
4345 if (Register::isVirtualRegister(Reg)) {
4346 if (MO.readsReg() || MI.readsVirtualRegister(Reg))
4347 return 0;
4348 } else {
4349 if (MI.readsRegister(Reg, TRI))
4350 return 0;
4351 }
4352
4353 // If any instructions in the clearance range are reading Reg, insert a
4354 // dependency breaking instruction, which is inexpensive and is likely to
4355 // be hidden in other instruction's cycles.
4356 return PartialRegUpdateClearance;
4357 }
4358
4359 // Return true for any instruction the copies the high bits of the first source
4360 // operand into the unused high bits of the destination operand.
hasUndefRegUpdate(unsigned Opcode,unsigned & OpNum,bool ForLoadFold=false)4361 static bool hasUndefRegUpdate(unsigned Opcode, unsigned &OpNum,
4362 bool ForLoadFold = false) {
4363 // Set the OpNum parameter to the first source operand.
4364 OpNum = 1;
4365 switch (Opcode) {
4366 case X86::VCVTSI2SSrr:
4367 case X86::VCVTSI2SSrm:
4368 case X86::VCVTSI2SSrr_Int:
4369 case X86::VCVTSI2SSrm_Int:
4370 case X86::VCVTSI642SSrr:
4371 case X86::VCVTSI642SSrm:
4372 case X86::VCVTSI642SSrr_Int:
4373 case X86::VCVTSI642SSrm_Int:
4374 case X86::VCVTSI2SDrr:
4375 case X86::VCVTSI2SDrm:
4376 case X86::VCVTSI2SDrr_Int:
4377 case X86::VCVTSI2SDrm_Int:
4378 case X86::VCVTSI642SDrr:
4379 case X86::VCVTSI642SDrm:
4380 case X86::VCVTSI642SDrr_Int:
4381 case X86::VCVTSI642SDrm_Int:
4382 // AVX-512
4383 case X86::VCVTSI2SSZrr:
4384 case X86::VCVTSI2SSZrm:
4385 case X86::VCVTSI2SSZrr_Int:
4386 case X86::VCVTSI2SSZrrb_Int:
4387 case X86::VCVTSI2SSZrm_Int:
4388 case X86::VCVTSI642SSZrr:
4389 case X86::VCVTSI642SSZrm:
4390 case X86::VCVTSI642SSZrr_Int:
4391 case X86::VCVTSI642SSZrrb_Int:
4392 case X86::VCVTSI642SSZrm_Int:
4393 case X86::VCVTSI2SDZrr:
4394 case X86::VCVTSI2SDZrm:
4395 case X86::VCVTSI2SDZrr_Int:
4396 case X86::VCVTSI2SDZrm_Int:
4397 case X86::VCVTSI642SDZrr:
4398 case X86::VCVTSI642SDZrm:
4399 case X86::VCVTSI642SDZrr_Int:
4400 case X86::VCVTSI642SDZrrb_Int:
4401 case X86::VCVTSI642SDZrm_Int:
4402 case X86::VCVTUSI2SSZrr:
4403 case X86::VCVTUSI2SSZrm:
4404 case X86::VCVTUSI2SSZrr_Int:
4405 case X86::VCVTUSI2SSZrrb_Int:
4406 case X86::VCVTUSI2SSZrm_Int:
4407 case X86::VCVTUSI642SSZrr:
4408 case X86::VCVTUSI642SSZrm:
4409 case X86::VCVTUSI642SSZrr_Int:
4410 case X86::VCVTUSI642SSZrrb_Int:
4411 case X86::VCVTUSI642SSZrm_Int:
4412 case X86::VCVTUSI2SDZrr:
4413 case X86::VCVTUSI2SDZrm:
4414 case X86::VCVTUSI2SDZrr_Int:
4415 case X86::VCVTUSI2SDZrm_Int:
4416 case X86::VCVTUSI642SDZrr:
4417 case X86::VCVTUSI642SDZrm:
4418 case X86::VCVTUSI642SDZrr_Int:
4419 case X86::VCVTUSI642SDZrrb_Int:
4420 case X86::VCVTUSI642SDZrm_Int:
4421 // Load folding won't effect the undef register update since the input is
4422 // a GPR.
4423 return !ForLoadFold;
4424 case X86::VCVTSD2SSrr:
4425 case X86::VCVTSD2SSrm:
4426 case X86::VCVTSD2SSrr_Int:
4427 case X86::VCVTSD2SSrm_Int:
4428 case X86::VCVTSS2SDrr:
4429 case X86::VCVTSS2SDrm:
4430 case X86::VCVTSS2SDrr_Int:
4431 case X86::VCVTSS2SDrm_Int:
4432 case X86::VRCPSSr:
4433 case X86::VRCPSSr_Int:
4434 case X86::VRCPSSm:
4435 case X86::VRCPSSm_Int:
4436 case X86::VROUNDSDr:
4437 case X86::VROUNDSDm:
4438 case X86::VROUNDSDr_Int:
4439 case X86::VROUNDSDm_Int:
4440 case X86::VROUNDSSr:
4441 case X86::VROUNDSSm:
4442 case X86::VROUNDSSr_Int:
4443 case X86::VROUNDSSm_Int:
4444 case X86::VRSQRTSSr:
4445 case X86::VRSQRTSSr_Int:
4446 case X86::VRSQRTSSm:
4447 case X86::VRSQRTSSm_Int:
4448 case X86::VSQRTSSr:
4449 case X86::VSQRTSSr_Int:
4450 case X86::VSQRTSSm:
4451 case X86::VSQRTSSm_Int:
4452 case X86::VSQRTSDr:
4453 case X86::VSQRTSDr_Int:
4454 case X86::VSQRTSDm:
4455 case X86::VSQRTSDm_Int:
4456 // AVX-512
4457 case X86::VCVTSD2SSZrr:
4458 case X86::VCVTSD2SSZrr_Int:
4459 case X86::VCVTSD2SSZrrb_Int:
4460 case X86::VCVTSD2SSZrm:
4461 case X86::VCVTSD2SSZrm_Int:
4462 case X86::VCVTSS2SDZrr:
4463 case X86::VCVTSS2SDZrr_Int:
4464 case X86::VCVTSS2SDZrrb_Int:
4465 case X86::VCVTSS2SDZrm:
4466 case X86::VCVTSS2SDZrm_Int:
4467 case X86::VGETEXPSDZr:
4468 case X86::VGETEXPSDZrb:
4469 case X86::VGETEXPSDZm:
4470 case X86::VGETEXPSSZr:
4471 case X86::VGETEXPSSZrb:
4472 case X86::VGETEXPSSZm:
4473 case X86::VGETMANTSDZrri:
4474 case X86::VGETMANTSDZrrib:
4475 case X86::VGETMANTSDZrmi:
4476 case X86::VGETMANTSSZrri:
4477 case X86::VGETMANTSSZrrib:
4478 case X86::VGETMANTSSZrmi:
4479 case X86::VRNDSCALESDZr:
4480 case X86::VRNDSCALESDZr_Int:
4481 case X86::VRNDSCALESDZrb_Int:
4482 case X86::VRNDSCALESDZm:
4483 case X86::VRNDSCALESDZm_Int:
4484 case X86::VRNDSCALESSZr:
4485 case X86::VRNDSCALESSZr_Int:
4486 case X86::VRNDSCALESSZrb_Int:
4487 case X86::VRNDSCALESSZm:
4488 case X86::VRNDSCALESSZm_Int:
4489 case X86::VRCP14SDZrr:
4490 case X86::VRCP14SDZrm:
4491 case X86::VRCP14SSZrr:
4492 case X86::VRCP14SSZrm:
4493 case X86::VRCP28SDZr:
4494 case X86::VRCP28SDZrb:
4495 case X86::VRCP28SDZm:
4496 case X86::VRCP28SSZr:
4497 case X86::VRCP28SSZrb:
4498 case X86::VRCP28SSZm:
4499 case X86::VREDUCESSZrmi:
4500 case X86::VREDUCESSZrri:
4501 case X86::VREDUCESSZrrib:
4502 case X86::VRSQRT14SDZrr:
4503 case X86::VRSQRT14SDZrm:
4504 case X86::VRSQRT14SSZrr:
4505 case X86::VRSQRT14SSZrm:
4506 case X86::VRSQRT28SDZr:
4507 case X86::VRSQRT28SDZrb:
4508 case X86::VRSQRT28SDZm:
4509 case X86::VRSQRT28SSZr:
4510 case X86::VRSQRT28SSZrb:
4511 case X86::VRSQRT28SSZm:
4512 case X86::VSQRTSSZr:
4513 case X86::VSQRTSSZr_Int:
4514 case X86::VSQRTSSZrb_Int:
4515 case X86::VSQRTSSZm:
4516 case X86::VSQRTSSZm_Int:
4517 case X86::VSQRTSDZr:
4518 case X86::VSQRTSDZr_Int:
4519 case X86::VSQRTSDZrb_Int:
4520 case X86::VSQRTSDZm:
4521 case X86::VSQRTSDZm_Int:
4522 return true;
4523 case X86::VMOVSSZrrk:
4524 case X86::VMOVSDZrrk:
4525 OpNum = 3;
4526 return true;
4527 case X86::VMOVSSZrrkz:
4528 case X86::VMOVSDZrrkz:
4529 OpNum = 2;
4530 return true;
4531 }
4532
4533 return false;
4534 }
4535
4536 /// Inform the BreakFalseDeps pass how many idle instructions we would like
4537 /// before certain undef register reads.
4538 ///
4539 /// This catches the VCVTSI2SD family of instructions:
4540 ///
4541 /// vcvtsi2sdq %rax, undef %xmm0, %xmm14
4542 ///
4543 /// We should to be careful *not* to catch VXOR idioms which are presumably
4544 /// handled specially in the pipeline:
4545 ///
4546 /// vxorps undef %xmm1, undef %xmm1, %xmm1
4547 ///
4548 /// Like getPartialRegUpdateClearance, this makes a strong assumption that the
4549 /// high bits that are passed-through are not live.
4550 unsigned
getUndefRegClearance(const MachineInstr & MI,unsigned & OpNum,const TargetRegisterInfo * TRI) const4551 X86InstrInfo::getUndefRegClearance(const MachineInstr &MI, unsigned &OpNum,
4552 const TargetRegisterInfo *TRI) const {
4553 if (!hasUndefRegUpdate(MI.getOpcode(), OpNum))
4554 return 0;
4555
4556 const MachineOperand &MO = MI.getOperand(OpNum);
4557 if (MO.isUndef() && Register::isPhysicalRegister(MO.getReg())) {
4558 return UndefRegClearance;
4559 }
4560 return 0;
4561 }
4562
breakPartialRegDependency(MachineInstr & MI,unsigned OpNum,const TargetRegisterInfo * TRI) const4563 void X86InstrInfo::breakPartialRegDependency(
4564 MachineInstr &MI, unsigned OpNum, const TargetRegisterInfo *TRI) const {
4565 Register Reg = MI.getOperand(OpNum).getReg();
4566 // If MI kills this register, the false dependence is already broken.
4567 if (MI.killsRegister(Reg, TRI))
4568 return;
4569
4570 if (X86::VR128RegClass.contains(Reg)) {
4571 // These instructions are all floating point domain, so xorps is the best
4572 // choice.
4573 unsigned Opc = Subtarget.hasAVX() ? X86::VXORPSrr : X86::XORPSrr;
4574 BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), get(Opc), Reg)
4575 .addReg(Reg, RegState::Undef)
4576 .addReg(Reg, RegState::Undef);
4577 MI.addRegisterKilled(Reg, TRI, true);
4578 } else if (X86::VR256RegClass.contains(Reg)) {
4579 // Use vxorps to clear the full ymm register.
4580 // It wants to read and write the xmm sub-register.
4581 Register XReg = TRI->getSubReg(Reg, X86::sub_xmm);
4582 BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), get(X86::VXORPSrr), XReg)
4583 .addReg(XReg, RegState::Undef)
4584 .addReg(XReg, RegState::Undef)
4585 .addReg(Reg, RegState::ImplicitDefine);
4586 MI.addRegisterKilled(Reg, TRI, true);
4587 } else if (X86::GR64RegClass.contains(Reg)) {
4588 // Using XOR32rr because it has shorter encoding and zeros up the upper bits
4589 // as well.
4590 Register XReg = TRI->getSubReg(Reg, X86::sub_32bit);
4591 BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), get(X86::XOR32rr), XReg)
4592 .addReg(XReg, RegState::Undef)
4593 .addReg(XReg, RegState::Undef)
4594 .addReg(Reg, RegState::ImplicitDefine);
4595 MI.addRegisterKilled(Reg, TRI, true);
4596 } else if (X86::GR32RegClass.contains(Reg)) {
4597 BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), get(X86::XOR32rr), Reg)
4598 .addReg(Reg, RegState::Undef)
4599 .addReg(Reg, RegState::Undef);
4600 MI.addRegisterKilled(Reg, TRI, true);
4601 }
4602 }
4603
addOperands(MachineInstrBuilder & MIB,ArrayRef<MachineOperand> MOs,int PtrOffset=0)4604 static void addOperands(MachineInstrBuilder &MIB, ArrayRef<MachineOperand> MOs,
4605 int PtrOffset = 0) {
4606 unsigned NumAddrOps = MOs.size();
4607
4608 if (NumAddrOps < 4) {
4609 // FrameIndex only - add an immediate offset (whether its zero or not).
4610 for (unsigned i = 0; i != NumAddrOps; ++i)
4611 MIB.add(MOs[i]);
4612 addOffset(MIB, PtrOffset);
4613 } else {
4614 // General Memory Addressing - we need to add any offset to an existing
4615 // offset.
4616 assert(MOs.size() == 5 && "Unexpected memory operand list length");
4617 for (unsigned i = 0; i != NumAddrOps; ++i) {
4618 const MachineOperand &MO = MOs[i];
4619 if (i == 3 && PtrOffset != 0) {
4620 MIB.addDisp(MO, PtrOffset);
4621 } else {
4622 MIB.add(MO);
4623 }
4624 }
4625 }
4626 }
4627
updateOperandRegConstraints(MachineFunction & MF,MachineInstr & NewMI,const TargetInstrInfo & TII)4628 static void updateOperandRegConstraints(MachineFunction &MF,
4629 MachineInstr &NewMI,
4630 const TargetInstrInfo &TII) {
4631 MachineRegisterInfo &MRI = MF.getRegInfo();
4632 const TargetRegisterInfo &TRI = *MRI.getTargetRegisterInfo();
4633
4634 for (int Idx : llvm::seq<int>(0, NewMI.getNumOperands())) {
4635 MachineOperand &MO = NewMI.getOperand(Idx);
4636 // We only need to update constraints on virtual register operands.
4637 if (!MO.isReg())
4638 continue;
4639 Register Reg = MO.getReg();
4640 if (!Register::isVirtualRegister(Reg))
4641 continue;
4642
4643 auto *NewRC = MRI.constrainRegClass(
4644 Reg, TII.getRegClass(NewMI.getDesc(), Idx, &TRI, MF));
4645 if (!NewRC) {
4646 LLVM_DEBUG(
4647 dbgs() << "WARNING: Unable to update register constraint for operand "
4648 << Idx << " of instruction:\n";
4649 NewMI.dump(); dbgs() << "\n");
4650 }
4651 }
4652 }
4653
FuseTwoAddrInst(MachineFunction & MF,unsigned Opcode,ArrayRef<MachineOperand> MOs,MachineBasicBlock::iterator InsertPt,MachineInstr & MI,const TargetInstrInfo & TII)4654 static MachineInstr *FuseTwoAddrInst(MachineFunction &MF, unsigned Opcode,
4655 ArrayRef<MachineOperand> MOs,
4656 MachineBasicBlock::iterator InsertPt,
4657 MachineInstr &MI,
4658 const TargetInstrInfo &TII) {
4659 // Create the base instruction with the memory operand as the first part.
4660 // Omit the implicit operands, something BuildMI can't do.
4661 MachineInstr *NewMI =
4662 MF.CreateMachineInstr(TII.get(Opcode), MI.getDebugLoc(), true);
4663 MachineInstrBuilder MIB(MF, NewMI);
4664 addOperands(MIB, MOs);
4665
4666 // Loop over the rest of the ri operands, converting them over.
4667 unsigned NumOps = MI.getDesc().getNumOperands() - 2;
4668 for (unsigned i = 0; i != NumOps; ++i) {
4669 MachineOperand &MO = MI.getOperand(i + 2);
4670 MIB.add(MO);
4671 }
4672 for (unsigned i = NumOps + 2, e = MI.getNumOperands(); i != e; ++i) {
4673 MachineOperand &MO = MI.getOperand(i);
4674 MIB.add(MO);
4675 }
4676
4677 updateOperandRegConstraints(MF, *NewMI, TII);
4678
4679 MachineBasicBlock *MBB = InsertPt->getParent();
4680 MBB->insert(InsertPt, NewMI);
4681
4682 return MIB;
4683 }
4684
FuseInst(MachineFunction & MF,unsigned Opcode,unsigned OpNo,ArrayRef<MachineOperand> MOs,MachineBasicBlock::iterator InsertPt,MachineInstr & MI,const TargetInstrInfo & TII,int PtrOffset=0)4685 static MachineInstr *FuseInst(MachineFunction &MF, unsigned Opcode,
4686 unsigned OpNo, ArrayRef<MachineOperand> MOs,
4687 MachineBasicBlock::iterator InsertPt,
4688 MachineInstr &MI, const TargetInstrInfo &TII,
4689 int PtrOffset = 0) {
4690 // Omit the implicit operands, something BuildMI can't do.
4691 MachineInstr *NewMI =
4692 MF.CreateMachineInstr(TII.get(Opcode), MI.getDebugLoc(), true);
4693 MachineInstrBuilder MIB(MF, NewMI);
4694
4695 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
4696 MachineOperand &MO = MI.getOperand(i);
4697 if (i == OpNo) {
4698 assert(MO.isReg() && "Expected to fold into reg operand!");
4699 addOperands(MIB, MOs, PtrOffset);
4700 } else {
4701 MIB.add(MO);
4702 }
4703 }
4704
4705 updateOperandRegConstraints(MF, *NewMI, TII);
4706
4707 // Copy the NoFPExcept flag from the instruction we're fusing.
4708 if (MI.getFlag(MachineInstr::MIFlag::NoFPExcept))
4709 NewMI->setFlag(MachineInstr::MIFlag::NoFPExcept);
4710
4711 MachineBasicBlock *MBB = InsertPt->getParent();
4712 MBB->insert(InsertPt, NewMI);
4713
4714 return MIB;
4715 }
4716
MakeM0Inst(const TargetInstrInfo & TII,unsigned Opcode,ArrayRef<MachineOperand> MOs,MachineBasicBlock::iterator InsertPt,MachineInstr & MI)4717 static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode,
4718 ArrayRef<MachineOperand> MOs,
4719 MachineBasicBlock::iterator InsertPt,
4720 MachineInstr &MI) {
4721 MachineInstrBuilder MIB = BuildMI(*InsertPt->getParent(), InsertPt,
4722 MI.getDebugLoc(), TII.get(Opcode));
4723 addOperands(MIB, MOs);
4724 return MIB.addImm(0);
4725 }
4726
foldMemoryOperandCustom(MachineFunction & MF,MachineInstr & MI,unsigned OpNum,ArrayRef<MachineOperand> MOs,MachineBasicBlock::iterator InsertPt,unsigned Size,unsigned Align) const4727 MachineInstr *X86InstrInfo::foldMemoryOperandCustom(
4728 MachineFunction &MF, MachineInstr &MI, unsigned OpNum,
4729 ArrayRef<MachineOperand> MOs, MachineBasicBlock::iterator InsertPt,
4730 unsigned Size, unsigned Align) const {
4731 switch (MI.getOpcode()) {
4732 case X86::INSERTPSrr:
4733 case X86::VINSERTPSrr:
4734 case X86::VINSERTPSZrr:
4735 // Attempt to convert the load of inserted vector into a fold load
4736 // of a single float.
4737 if (OpNum == 2) {
4738 unsigned Imm = MI.getOperand(MI.getNumOperands() - 1).getImm();
4739 unsigned ZMask = Imm & 15;
4740 unsigned DstIdx = (Imm >> 4) & 3;
4741 unsigned SrcIdx = (Imm >> 6) & 3;
4742
4743 const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
4744 const TargetRegisterClass *RC = getRegClass(MI.getDesc(), OpNum, &RI, MF);
4745 unsigned RCSize = TRI.getRegSizeInBits(*RC) / 8;
4746 if ((Size == 0 || Size >= 16) && RCSize >= 16 && 4 <= Align) {
4747 int PtrOffset = SrcIdx * 4;
4748 unsigned NewImm = (DstIdx << 4) | ZMask;
4749 unsigned NewOpCode =
4750 (MI.getOpcode() == X86::VINSERTPSZrr) ? X86::VINSERTPSZrm :
4751 (MI.getOpcode() == X86::VINSERTPSrr) ? X86::VINSERTPSrm :
4752 X86::INSERTPSrm;
4753 MachineInstr *NewMI =
4754 FuseInst(MF, NewOpCode, OpNum, MOs, InsertPt, MI, *this, PtrOffset);
4755 NewMI->getOperand(NewMI->getNumOperands() - 1).setImm(NewImm);
4756 return NewMI;
4757 }
4758 }
4759 break;
4760 case X86::MOVHLPSrr:
4761 case X86::VMOVHLPSrr:
4762 case X86::VMOVHLPSZrr:
4763 // Move the upper 64-bits of the second operand to the lower 64-bits.
4764 // To fold the load, adjust the pointer to the upper and use (V)MOVLPS.
4765 // TODO: In most cases AVX doesn't have a 8-byte alignment requirement.
4766 if (OpNum == 2) {
4767 const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
4768 const TargetRegisterClass *RC = getRegClass(MI.getDesc(), OpNum, &RI, MF);
4769 unsigned RCSize = TRI.getRegSizeInBits(*RC) / 8;
4770 if ((Size == 0 || Size >= 16) && RCSize >= 16 && 8 <= Align) {
4771 unsigned NewOpCode =
4772 (MI.getOpcode() == X86::VMOVHLPSZrr) ? X86::VMOVLPSZ128rm :
4773 (MI.getOpcode() == X86::VMOVHLPSrr) ? X86::VMOVLPSrm :
4774 X86::MOVLPSrm;
4775 MachineInstr *NewMI =
4776 FuseInst(MF, NewOpCode, OpNum, MOs, InsertPt, MI, *this, 8);
4777 return NewMI;
4778 }
4779 }
4780 break;
4781 case X86::UNPCKLPDrr:
4782 // If we won't be able to fold this to the memory form of UNPCKL, use
4783 // MOVHPD instead. Done as custom because we can't have this in the load
4784 // table twice.
4785 if (OpNum == 2) {
4786 const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
4787 const TargetRegisterClass *RC = getRegClass(MI.getDesc(), OpNum, &RI, MF);
4788 unsigned RCSize = TRI.getRegSizeInBits(*RC) / 8;
4789 if ((Size == 0 || Size >= 16) && RCSize >= 16 && Align < 16) {
4790 MachineInstr *NewMI =
4791 FuseInst(MF, X86::MOVHPDrm, OpNum, MOs, InsertPt, MI, *this);
4792 return NewMI;
4793 }
4794 }
4795 break;
4796 }
4797
4798 return nullptr;
4799 }
4800
shouldPreventUndefRegUpdateMemFold(MachineFunction & MF,MachineInstr & MI)4801 static bool shouldPreventUndefRegUpdateMemFold(MachineFunction &MF,
4802 MachineInstr &MI) {
4803 unsigned Ignored;
4804 if (!hasUndefRegUpdate(MI.getOpcode(), Ignored, /*ForLoadFold*/true) ||
4805 !MI.getOperand(1).isReg())
4806 return false;
4807
4808 // The are two cases we need to handle depending on where in the pipeline
4809 // the folding attempt is being made.
4810 // -Register has the undef flag set.
4811 // -Register is produced by the IMPLICIT_DEF instruction.
4812
4813 if (MI.getOperand(1).isUndef())
4814 return true;
4815
4816 MachineRegisterInfo &RegInfo = MF.getRegInfo();
4817 MachineInstr *VRegDef = RegInfo.getUniqueVRegDef(MI.getOperand(1).getReg());
4818 return VRegDef && VRegDef->isImplicitDef();
4819 }
4820
4821
foldMemoryOperandImpl(MachineFunction & MF,MachineInstr & MI,unsigned OpNum,ArrayRef<MachineOperand> MOs,MachineBasicBlock::iterator InsertPt,unsigned Size,unsigned Align,bool AllowCommute) const4822 MachineInstr *X86InstrInfo::foldMemoryOperandImpl(
4823 MachineFunction &MF, MachineInstr &MI, unsigned OpNum,
4824 ArrayRef<MachineOperand> MOs, MachineBasicBlock::iterator InsertPt,
4825 unsigned Size, unsigned Align, bool AllowCommute) const {
4826 bool isSlowTwoMemOps = Subtarget.slowTwoMemOps();
4827 bool isTwoAddrFold = false;
4828
4829 // For CPUs that favor the register form of a call or push,
4830 // do not fold loads into calls or pushes, unless optimizing for size
4831 // aggressively.
4832 if (isSlowTwoMemOps && !MF.getFunction().hasMinSize() &&
4833 (MI.getOpcode() == X86::CALL32r || MI.getOpcode() == X86::CALL64r ||
4834 MI.getOpcode() == X86::PUSH16r || MI.getOpcode() == X86::PUSH32r ||
4835 MI.getOpcode() == X86::PUSH64r))
4836 return nullptr;
4837
4838 // Avoid partial and undef register update stalls unless optimizing for size.
4839 if (!MF.getFunction().hasOptSize() &&
4840 (hasPartialRegUpdate(MI.getOpcode(), Subtarget, /*ForLoadFold*/true) ||
4841 shouldPreventUndefRegUpdateMemFold(MF, MI)))
4842 return nullptr;
4843
4844 unsigned NumOps = MI.getDesc().getNumOperands();
4845 bool isTwoAddr =
4846 NumOps > 1 && MI.getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
4847
4848 // FIXME: AsmPrinter doesn't know how to handle
4849 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
4850 if (MI.getOpcode() == X86::ADD32ri &&
4851 MI.getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
4852 return nullptr;
4853
4854 // GOTTPOFF relocation loads can only be folded into add instructions.
4855 // FIXME: Need to exclude other relocations that only support specific
4856 // instructions.
4857 if (MOs.size() == X86::AddrNumOperands &&
4858 MOs[X86::AddrDisp].getTargetFlags() == X86II::MO_GOTTPOFF &&
4859 MI.getOpcode() != X86::ADD64rr)
4860 return nullptr;
4861
4862 MachineInstr *NewMI = nullptr;
4863
4864 // Attempt to fold any custom cases we have.
4865 if (MachineInstr *CustomMI =
4866 foldMemoryOperandCustom(MF, MI, OpNum, MOs, InsertPt, Size, Align))
4867 return CustomMI;
4868
4869 const X86MemoryFoldTableEntry *I = nullptr;
4870
4871 // Folding a memory location into the two-address part of a two-address
4872 // instruction is different than folding it other places. It requires
4873 // replacing the *two* registers with the memory location.
4874 if (isTwoAddr && NumOps >= 2 && OpNum < 2 && MI.getOperand(0).isReg() &&
4875 MI.getOperand(1).isReg() &&
4876 MI.getOperand(0).getReg() == MI.getOperand(1).getReg()) {
4877 I = lookupTwoAddrFoldTable(MI.getOpcode());
4878 isTwoAddrFold = true;
4879 } else {
4880 if (OpNum == 0) {
4881 if (MI.getOpcode() == X86::MOV32r0) {
4882 NewMI = MakeM0Inst(*this, X86::MOV32mi, MOs, InsertPt, MI);
4883 if (NewMI)
4884 return NewMI;
4885 }
4886 }
4887
4888 I = lookupFoldTable(MI.getOpcode(), OpNum);
4889 }
4890
4891 if (I != nullptr) {
4892 unsigned Opcode = I->DstOp;
4893 unsigned MinAlign = (I->Flags & TB_ALIGN_MASK) >> TB_ALIGN_SHIFT;
4894 MinAlign = MinAlign ? 1 << (MinAlign - 1) : 0;
4895 if (Align < MinAlign)
4896 return nullptr;
4897 bool NarrowToMOV32rm = false;
4898 if (Size) {
4899 const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
4900 const TargetRegisterClass *RC = getRegClass(MI.getDesc(), OpNum,
4901 &RI, MF);
4902 unsigned RCSize = TRI.getRegSizeInBits(*RC) / 8;
4903 if (Size < RCSize) {
4904 // FIXME: Allow scalar intrinsic instructions like ADDSSrm_Int.
4905 // Check if it's safe to fold the load. If the size of the object is
4906 // narrower than the load width, then it's not.
4907 if (Opcode != X86::MOV64rm || RCSize != 8 || Size != 4)
4908 return nullptr;
4909 // If this is a 64-bit load, but the spill slot is 32, then we can do
4910 // a 32-bit load which is implicitly zero-extended. This likely is
4911 // due to live interval analysis remat'ing a load from stack slot.
4912 if (MI.getOperand(0).getSubReg() || MI.getOperand(1).getSubReg())
4913 return nullptr;
4914 Opcode = X86::MOV32rm;
4915 NarrowToMOV32rm = true;
4916 }
4917 }
4918
4919 if (isTwoAddrFold)
4920 NewMI = FuseTwoAddrInst(MF, Opcode, MOs, InsertPt, MI, *this);
4921 else
4922 NewMI = FuseInst(MF, Opcode, OpNum, MOs, InsertPt, MI, *this);
4923
4924 if (NarrowToMOV32rm) {
4925 // If this is the special case where we use a MOV32rm to load a 32-bit
4926 // value and zero-extend the top bits. Change the destination register
4927 // to a 32-bit one.
4928 Register DstReg = NewMI->getOperand(0).getReg();
4929 if (Register::isPhysicalRegister(DstReg))
4930 NewMI->getOperand(0).setReg(RI.getSubReg(DstReg, X86::sub_32bit));
4931 else
4932 NewMI->getOperand(0).setSubReg(X86::sub_32bit);
4933 }
4934 return NewMI;
4935 }
4936
4937 // If the instruction and target operand are commutable, commute the
4938 // instruction and try again.
4939 if (AllowCommute) {
4940 unsigned CommuteOpIdx1 = OpNum, CommuteOpIdx2 = CommuteAnyOperandIndex;
4941 if (findCommutedOpIndices(MI, CommuteOpIdx1, CommuteOpIdx2)) {
4942 bool HasDef = MI.getDesc().getNumDefs();
4943 Register Reg0 = HasDef ? MI.getOperand(0).getReg() : Register();
4944 Register Reg1 = MI.getOperand(CommuteOpIdx1).getReg();
4945 Register Reg2 = MI.getOperand(CommuteOpIdx2).getReg();
4946 bool Tied1 =
4947 0 == MI.getDesc().getOperandConstraint(CommuteOpIdx1, MCOI::TIED_TO);
4948 bool Tied2 =
4949 0 == MI.getDesc().getOperandConstraint(CommuteOpIdx2, MCOI::TIED_TO);
4950
4951 // If either of the commutable operands are tied to the destination
4952 // then we can not commute + fold.
4953 if ((HasDef && Reg0 == Reg1 && Tied1) ||
4954 (HasDef && Reg0 == Reg2 && Tied2))
4955 return nullptr;
4956
4957 MachineInstr *CommutedMI =
4958 commuteInstruction(MI, false, CommuteOpIdx1, CommuteOpIdx2);
4959 if (!CommutedMI) {
4960 // Unable to commute.
4961 return nullptr;
4962 }
4963 if (CommutedMI != &MI) {
4964 // New instruction. We can't fold from this.
4965 CommutedMI->eraseFromParent();
4966 return nullptr;
4967 }
4968
4969 // Attempt to fold with the commuted version of the instruction.
4970 NewMI = foldMemoryOperandImpl(MF, MI, CommuteOpIdx2, MOs, InsertPt,
4971 Size, Align, /*AllowCommute=*/false);
4972 if (NewMI)
4973 return NewMI;
4974
4975 // Folding failed again - undo the commute before returning.
4976 MachineInstr *UncommutedMI =
4977 commuteInstruction(MI, false, CommuteOpIdx1, CommuteOpIdx2);
4978 if (!UncommutedMI) {
4979 // Unable to commute.
4980 return nullptr;
4981 }
4982 if (UncommutedMI != &MI) {
4983 // New instruction. It doesn't need to be kept.
4984 UncommutedMI->eraseFromParent();
4985 return nullptr;
4986 }
4987
4988 // Return here to prevent duplicate fuse failure report.
4989 return nullptr;
4990 }
4991 }
4992
4993 // No fusion
4994 if (PrintFailedFusing && !MI.isCopy())
4995 dbgs() << "We failed to fuse operand " << OpNum << " in " << MI;
4996 return nullptr;
4997 }
4998
4999 MachineInstr *
foldMemoryOperandImpl(MachineFunction & MF,MachineInstr & MI,ArrayRef<unsigned> Ops,MachineBasicBlock::iterator InsertPt,int FrameIndex,LiveIntervals * LIS,VirtRegMap * VRM) const5000 X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF, MachineInstr &MI,
5001 ArrayRef<unsigned> Ops,
5002 MachineBasicBlock::iterator InsertPt,
5003 int FrameIndex, LiveIntervals *LIS,
5004 VirtRegMap *VRM) const {
5005 // Check switch flag
5006 if (NoFusing)
5007 return nullptr;
5008
5009 // Avoid partial and undef register update stalls unless optimizing for size.
5010 if (!MF.getFunction().hasOptSize() &&
5011 (hasPartialRegUpdate(MI.getOpcode(), Subtarget, /*ForLoadFold*/true) ||
5012 shouldPreventUndefRegUpdateMemFold(MF, MI)))
5013 return nullptr;
5014
5015 // Don't fold subreg spills, or reloads that use a high subreg.
5016 for (auto Op : Ops) {
5017 MachineOperand &MO = MI.getOperand(Op);
5018 auto SubReg = MO.getSubReg();
5019 if (SubReg && (MO.isDef() || SubReg == X86::sub_8bit_hi))
5020 return nullptr;
5021 }
5022
5023 const MachineFrameInfo &MFI = MF.getFrameInfo();
5024 unsigned Size = MFI.getObjectSize(FrameIndex);
5025 unsigned Alignment = MFI.getObjectAlignment(FrameIndex);
5026 // If the function stack isn't realigned we don't want to fold instructions
5027 // that need increased alignment.
5028 if (!RI.needsStackRealignment(MF))
5029 Alignment =
5030 std::min(Alignment, Subtarget.getFrameLowering()->getStackAlignment());
5031 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
5032 unsigned NewOpc = 0;
5033 unsigned RCSize = 0;
5034 switch (MI.getOpcode()) {
5035 default: return nullptr;
5036 case X86::TEST8rr: NewOpc = X86::CMP8ri; RCSize = 1; break;
5037 case X86::TEST16rr: NewOpc = X86::CMP16ri8; RCSize = 2; break;
5038 case X86::TEST32rr: NewOpc = X86::CMP32ri8; RCSize = 4; break;
5039 case X86::TEST64rr: NewOpc = X86::CMP64ri8; RCSize = 8; break;
5040 }
5041 // Check if it's safe to fold the load. If the size of the object is
5042 // narrower than the load width, then it's not.
5043 if (Size < RCSize)
5044 return nullptr;
5045 // Change to CMPXXri r, 0 first.
5046 MI.setDesc(get(NewOpc));
5047 MI.getOperand(1).ChangeToImmediate(0);
5048 } else if (Ops.size() != 1)
5049 return nullptr;
5050
5051 return foldMemoryOperandImpl(MF, MI, Ops[0],
5052 MachineOperand::CreateFI(FrameIndex), InsertPt,
5053 Size, Alignment, /*AllowCommute=*/true);
5054 }
5055
5056 /// Check if \p LoadMI is a partial register load that we can't fold into \p MI
5057 /// because the latter uses contents that wouldn't be defined in the folded
5058 /// version. For instance, this transformation isn't legal:
5059 /// movss (%rdi), %xmm0
5060 /// addps %xmm0, %xmm0
5061 /// ->
5062 /// addps (%rdi), %xmm0
5063 ///
5064 /// But this one is:
5065 /// movss (%rdi), %xmm0
5066 /// addss %xmm0, %xmm0
5067 /// ->
5068 /// addss (%rdi), %xmm0
5069 ///
isNonFoldablePartialRegisterLoad(const MachineInstr & LoadMI,const MachineInstr & UserMI,const MachineFunction & MF)5070 static bool isNonFoldablePartialRegisterLoad(const MachineInstr &LoadMI,
5071 const MachineInstr &UserMI,
5072 const MachineFunction &MF) {
5073 unsigned Opc = LoadMI.getOpcode();
5074 unsigned UserOpc = UserMI.getOpcode();
5075 const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
5076 const TargetRegisterClass *RC =
5077 MF.getRegInfo().getRegClass(LoadMI.getOperand(0).getReg());
5078 unsigned RegSize = TRI.getRegSizeInBits(*RC);
5079
5080 if ((Opc == X86::MOVSSrm || Opc == X86::VMOVSSrm || Opc == X86::VMOVSSZrm ||
5081 Opc == X86::MOVSSrm_alt || Opc == X86::VMOVSSrm_alt ||
5082 Opc == X86::VMOVSSZrm_alt) &&
5083 RegSize > 32) {
5084 // These instructions only load 32 bits, we can't fold them if the
5085 // destination register is wider than 32 bits (4 bytes), and its user
5086 // instruction isn't scalar (SS).
5087 switch (UserOpc) {
5088 case X86::ADDSSrr_Int: case X86::VADDSSrr_Int: case X86::VADDSSZrr_Int:
5089 case X86::CMPSSrr_Int: case X86::VCMPSSrr_Int: case X86::VCMPSSZrr_Int:
5090 case X86::DIVSSrr_Int: case X86::VDIVSSrr_Int: case X86::VDIVSSZrr_Int:
5091 case X86::MAXSSrr_Int: case X86::VMAXSSrr_Int: case X86::VMAXSSZrr_Int:
5092 case X86::MINSSrr_Int: case X86::VMINSSrr_Int: case X86::VMINSSZrr_Int:
5093 case X86::MULSSrr_Int: case X86::VMULSSrr_Int: case X86::VMULSSZrr_Int:
5094 case X86::SUBSSrr_Int: case X86::VSUBSSrr_Int: case X86::VSUBSSZrr_Int:
5095 case X86::VADDSSZrr_Intk: case X86::VADDSSZrr_Intkz:
5096 case X86::VCMPSSZrr_Intk:
5097 case X86::VDIVSSZrr_Intk: case X86::VDIVSSZrr_Intkz:
5098 case X86::VMAXSSZrr_Intk: case X86::VMAXSSZrr_Intkz:
5099 case X86::VMINSSZrr_Intk: case X86::VMINSSZrr_Intkz:
5100 case X86::VMULSSZrr_Intk: case X86::VMULSSZrr_Intkz:
5101 case X86::VSUBSSZrr_Intk: case X86::VSUBSSZrr_Intkz:
5102 case X86::VFMADDSS4rr_Int: case X86::VFNMADDSS4rr_Int:
5103 case X86::VFMSUBSS4rr_Int: case X86::VFNMSUBSS4rr_Int:
5104 case X86::VFMADD132SSr_Int: case X86::VFNMADD132SSr_Int:
5105 case X86::VFMADD213SSr_Int: case X86::VFNMADD213SSr_Int:
5106 case X86::VFMADD231SSr_Int: case X86::VFNMADD231SSr_Int:
5107 case X86::VFMSUB132SSr_Int: case X86::VFNMSUB132SSr_Int:
5108 case X86::VFMSUB213SSr_Int: case X86::VFNMSUB213SSr_Int:
5109 case X86::VFMSUB231SSr_Int: case X86::VFNMSUB231SSr_Int:
5110 case X86::VFMADD132SSZr_Int: case X86::VFNMADD132SSZr_Int:
5111 case X86::VFMADD213SSZr_Int: case X86::VFNMADD213SSZr_Int:
5112 case X86::VFMADD231SSZr_Int: case X86::VFNMADD231SSZr_Int:
5113 case X86::VFMSUB132SSZr_Int: case X86::VFNMSUB132SSZr_Int:
5114 case X86::VFMSUB213SSZr_Int: case X86::VFNMSUB213SSZr_Int:
5115 case X86::VFMSUB231SSZr_Int: case X86::VFNMSUB231SSZr_Int:
5116 case X86::VFMADD132SSZr_Intk: case X86::VFNMADD132SSZr_Intk:
5117 case X86::VFMADD213SSZr_Intk: case X86::VFNMADD213SSZr_Intk:
5118 case X86::VFMADD231SSZr_Intk: case X86::VFNMADD231SSZr_Intk:
5119 case X86::VFMSUB132SSZr_Intk: case X86::VFNMSUB132SSZr_Intk:
5120 case X86::VFMSUB213SSZr_Intk: case X86::VFNMSUB213SSZr_Intk:
5121 case X86::VFMSUB231SSZr_Intk: case X86::VFNMSUB231SSZr_Intk:
5122 case X86::VFMADD132SSZr_Intkz: case X86::VFNMADD132SSZr_Intkz:
5123 case X86::VFMADD213SSZr_Intkz: case X86::VFNMADD213SSZr_Intkz:
5124 case X86::VFMADD231SSZr_Intkz: case X86::VFNMADD231SSZr_Intkz:
5125 case X86::VFMSUB132SSZr_Intkz: case X86::VFNMSUB132SSZr_Intkz:
5126 case X86::VFMSUB213SSZr_Intkz: case X86::VFNMSUB213SSZr_Intkz:
5127 case X86::VFMSUB231SSZr_Intkz: case X86::VFNMSUB231SSZr_Intkz:
5128 return false;
5129 default:
5130 return true;
5131 }
5132 }
5133
5134 if ((Opc == X86::MOVSDrm || Opc == X86::VMOVSDrm || Opc == X86::VMOVSDZrm ||
5135 Opc == X86::MOVSDrm_alt || Opc == X86::VMOVSDrm_alt ||
5136 Opc == X86::VMOVSDZrm_alt) &&
5137 RegSize > 64) {
5138 // These instructions only load 64 bits, we can't fold them if the
5139 // destination register is wider than 64 bits (8 bytes), and its user
5140 // instruction isn't scalar (SD).
5141 switch (UserOpc) {
5142 case X86::ADDSDrr_Int: case X86::VADDSDrr_Int: case X86::VADDSDZrr_Int:
5143 case X86::CMPSDrr_Int: case X86::VCMPSDrr_Int: case X86::VCMPSDZrr_Int:
5144 case X86::DIVSDrr_Int: case X86::VDIVSDrr_Int: case X86::VDIVSDZrr_Int:
5145 case X86::MAXSDrr_Int: case X86::VMAXSDrr_Int: case X86::VMAXSDZrr_Int:
5146 case X86::MINSDrr_Int: case X86::VMINSDrr_Int: case X86::VMINSDZrr_Int:
5147 case X86::MULSDrr_Int: case X86::VMULSDrr_Int: case X86::VMULSDZrr_Int:
5148 case X86::SUBSDrr_Int: case X86::VSUBSDrr_Int: case X86::VSUBSDZrr_Int:
5149 case X86::VADDSDZrr_Intk: case X86::VADDSDZrr_Intkz:
5150 case X86::VCMPSDZrr_Intk:
5151 case X86::VDIVSDZrr_Intk: case X86::VDIVSDZrr_Intkz:
5152 case X86::VMAXSDZrr_Intk: case X86::VMAXSDZrr_Intkz:
5153 case X86::VMINSDZrr_Intk: case X86::VMINSDZrr_Intkz:
5154 case X86::VMULSDZrr_Intk: case X86::VMULSDZrr_Intkz:
5155 case X86::VSUBSDZrr_Intk: case X86::VSUBSDZrr_Intkz:
5156 case X86::VFMADDSD4rr_Int: case X86::VFNMADDSD4rr_Int:
5157 case X86::VFMSUBSD4rr_Int: case X86::VFNMSUBSD4rr_Int:
5158 case X86::VFMADD132SDr_Int: case X86::VFNMADD132SDr_Int:
5159 case X86::VFMADD213SDr_Int: case X86::VFNMADD213SDr_Int:
5160 case X86::VFMADD231SDr_Int: case X86::VFNMADD231SDr_Int:
5161 case X86::VFMSUB132SDr_Int: case X86::VFNMSUB132SDr_Int:
5162 case X86::VFMSUB213SDr_Int: case X86::VFNMSUB213SDr_Int:
5163 case X86::VFMSUB231SDr_Int: case X86::VFNMSUB231SDr_Int:
5164 case X86::VFMADD132SDZr_Int: case X86::VFNMADD132SDZr_Int:
5165 case X86::VFMADD213SDZr_Int: case X86::VFNMADD213SDZr_Int:
5166 case X86::VFMADD231SDZr_Int: case X86::VFNMADD231SDZr_Int:
5167 case X86::VFMSUB132SDZr_Int: case X86::VFNMSUB132SDZr_Int:
5168 case X86::VFMSUB213SDZr_Int: case X86::VFNMSUB213SDZr_Int:
5169 case X86::VFMSUB231SDZr_Int: case X86::VFNMSUB231SDZr_Int:
5170 case X86::VFMADD132SDZr_Intk: case X86::VFNMADD132SDZr_Intk:
5171 case X86::VFMADD213SDZr_Intk: case X86::VFNMADD213SDZr_Intk:
5172 case X86::VFMADD231SDZr_Intk: case X86::VFNMADD231SDZr_Intk:
5173 case X86::VFMSUB132SDZr_Intk: case X86::VFNMSUB132SDZr_Intk:
5174 case X86::VFMSUB213SDZr_Intk: case X86::VFNMSUB213SDZr_Intk:
5175 case X86::VFMSUB231SDZr_Intk: case X86::VFNMSUB231SDZr_Intk:
5176 case X86::VFMADD132SDZr_Intkz: case X86::VFNMADD132SDZr_Intkz:
5177 case X86::VFMADD213SDZr_Intkz: case X86::VFNMADD213SDZr_Intkz:
5178 case X86::VFMADD231SDZr_Intkz: case X86::VFNMADD231SDZr_Intkz:
5179 case X86::VFMSUB132SDZr_Intkz: case X86::VFNMSUB132SDZr_Intkz:
5180 case X86::VFMSUB213SDZr_Intkz: case X86::VFNMSUB213SDZr_Intkz:
5181 case X86::VFMSUB231SDZr_Intkz: case X86::VFNMSUB231SDZr_Intkz:
5182 return false;
5183 default:
5184 return true;
5185 }
5186 }
5187
5188 return false;
5189 }
5190
foldMemoryOperandImpl(MachineFunction & MF,MachineInstr & MI,ArrayRef<unsigned> Ops,MachineBasicBlock::iterator InsertPt,MachineInstr & LoadMI,LiveIntervals * LIS) const5191 MachineInstr *X86InstrInfo::foldMemoryOperandImpl(
5192 MachineFunction &MF, MachineInstr &MI, ArrayRef<unsigned> Ops,
5193 MachineBasicBlock::iterator InsertPt, MachineInstr &LoadMI,
5194 LiveIntervals *LIS) const {
5195
5196 // TODO: Support the case where LoadMI loads a wide register, but MI
5197 // only uses a subreg.
5198 for (auto Op : Ops) {
5199 if (MI.getOperand(Op).getSubReg())
5200 return nullptr;
5201 }
5202
5203 // If loading from a FrameIndex, fold directly from the FrameIndex.
5204 unsigned NumOps = LoadMI.getDesc().getNumOperands();
5205 int FrameIndex;
5206 if (isLoadFromStackSlot(LoadMI, FrameIndex)) {
5207 if (isNonFoldablePartialRegisterLoad(LoadMI, MI, MF))
5208 return nullptr;
5209 return foldMemoryOperandImpl(MF, MI, Ops, InsertPt, FrameIndex, LIS);
5210 }
5211
5212 // Check switch flag
5213 if (NoFusing) return nullptr;
5214
5215 // Avoid partial and undef register update stalls unless optimizing for size.
5216 if (!MF.getFunction().hasOptSize() &&
5217 (hasPartialRegUpdate(MI.getOpcode(), Subtarget, /*ForLoadFold*/true) ||
5218 shouldPreventUndefRegUpdateMemFold(MF, MI)))
5219 return nullptr;
5220
5221 // Determine the alignment of the load.
5222 unsigned Alignment = 0;
5223 if (LoadMI.hasOneMemOperand())
5224 Alignment = (*LoadMI.memoperands_begin())->getAlignment();
5225 else
5226 switch (LoadMI.getOpcode()) {
5227 case X86::AVX512_512_SET0:
5228 case X86::AVX512_512_SETALLONES:
5229 Alignment = 64;
5230 break;
5231 case X86::AVX2_SETALLONES:
5232 case X86::AVX1_SETALLONES:
5233 case X86::AVX_SET0:
5234 case X86::AVX512_256_SET0:
5235 Alignment = 32;
5236 break;
5237 case X86::V_SET0:
5238 case X86::V_SETALLONES:
5239 case X86::AVX512_128_SET0:
5240 case X86::FsFLD0F128:
5241 case X86::AVX512_FsFLD0F128:
5242 Alignment = 16;
5243 break;
5244 case X86::MMX_SET0:
5245 case X86::FsFLD0SD:
5246 case X86::AVX512_FsFLD0SD:
5247 Alignment = 8;
5248 break;
5249 case X86::FsFLD0SS:
5250 case X86::AVX512_FsFLD0SS:
5251 Alignment = 4;
5252 break;
5253 default:
5254 return nullptr;
5255 }
5256 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
5257 unsigned NewOpc = 0;
5258 switch (MI.getOpcode()) {
5259 default: return nullptr;
5260 case X86::TEST8rr: NewOpc = X86::CMP8ri; break;
5261 case X86::TEST16rr: NewOpc = X86::CMP16ri8; break;
5262 case X86::TEST32rr: NewOpc = X86::CMP32ri8; break;
5263 case X86::TEST64rr: NewOpc = X86::CMP64ri8; break;
5264 }
5265 // Change to CMPXXri r, 0 first.
5266 MI.setDesc(get(NewOpc));
5267 MI.getOperand(1).ChangeToImmediate(0);
5268 } else if (Ops.size() != 1)
5269 return nullptr;
5270
5271 // Make sure the subregisters match.
5272 // Otherwise we risk changing the size of the load.
5273 if (LoadMI.getOperand(0).getSubReg() != MI.getOperand(Ops[0]).getSubReg())
5274 return nullptr;
5275
5276 SmallVector<MachineOperand,X86::AddrNumOperands> MOs;
5277 switch (LoadMI.getOpcode()) {
5278 case X86::MMX_SET0:
5279 case X86::V_SET0:
5280 case X86::V_SETALLONES:
5281 case X86::AVX2_SETALLONES:
5282 case X86::AVX1_SETALLONES:
5283 case X86::AVX_SET0:
5284 case X86::AVX512_128_SET0:
5285 case X86::AVX512_256_SET0:
5286 case X86::AVX512_512_SET0:
5287 case X86::AVX512_512_SETALLONES:
5288 case X86::FsFLD0SD:
5289 case X86::AVX512_FsFLD0SD:
5290 case X86::FsFLD0SS:
5291 case X86::AVX512_FsFLD0SS:
5292 case X86::FsFLD0F128:
5293 case X86::AVX512_FsFLD0F128: {
5294 // Folding a V_SET0 or V_SETALLONES as a load, to ease register pressure.
5295 // Create a constant-pool entry and operands to load from it.
5296
5297 // Medium and large mode can't fold loads this way.
5298 if (MF.getTarget().getCodeModel() != CodeModel::Small &&
5299 MF.getTarget().getCodeModel() != CodeModel::Kernel)
5300 return nullptr;
5301
5302 // x86-32 PIC requires a PIC base register for constant pools.
5303 unsigned PICBase = 0;
5304 if (MF.getTarget().isPositionIndependent()) {
5305 if (Subtarget.is64Bit())
5306 PICBase = X86::RIP;
5307 else
5308 // FIXME: PICBase = getGlobalBaseReg(&MF);
5309 // This doesn't work for several reasons.
5310 // 1. GlobalBaseReg may have been spilled.
5311 // 2. It may not be live at MI.
5312 return nullptr;
5313 }
5314
5315 // Create a constant-pool entry.
5316 MachineConstantPool &MCP = *MF.getConstantPool();
5317 Type *Ty;
5318 unsigned Opc = LoadMI.getOpcode();
5319 if (Opc == X86::FsFLD0SS || Opc == X86::AVX512_FsFLD0SS)
5320 Ty = Type::getFloatTy(MF.getFunction().getContext());
5321 else if (Opc == X86::FsFLD0SD || Opc == X86::AVX512_FsFLD0SD)
5322 Ty = Type::getDoubleTy(MF.getFunction().getContext());
5323 else if (Opc == X86::FsFLD0F128 || Opc == X86::AVX512_FsFLD0F128)
5324 Ty = Type::getFP128Ty(MF.getFunction().getContext());
5325 else if (Opc == X86::AVX512_512_SET0 || Opc == X86::AVX512_512_SETALLONES)
5326 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction().getContext()),16);
5327 else if (Opc == X86::AVX2_SETALLONES || Opc == X86::AVX_SET0 ||
5328 Opc == X86::AVX512_256_SET0 || Opc == X86::AVX1_SETALLONES)
5329 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction().getContext()), 8);
5330 else if (Opc == X86::MMX_SET0)
5331 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction().getContext()), 2);
5332 else
5333 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction().getContext()), 4);
5334
5335 bool IsAllOnes = (Opc == X86::V_SETALLONES || Opc == X86::AVX2_SETALLONES ||
5336 Opc == X86::AVX512_512_SETALLONES ||
5337 Opc == X86::AVX1_SETALLONES);
5338 const Constant *C = IsAllOnes ? Constant::getAllOnesValue(Ty) :
5339 Constant::getNullValue(Ty);
5340 unsigned CPI = MCP.getConstantPoolIndex(C, Alignment);
5341
5342 // Create operands to load from the constant pool entry.
5343 MOs.push_back(MachineOperand::CreateReg(PICBase, false));
5344 MOs.push_back(MachineOperand::CreateImm(1));
5345 MOs.push_back(MachineOperand::CreateReg(0, false));
5346 MOs.push_back(MachineOperand::CreateCPI(CPI, 0));
5347 MOs.push_back(MachineOperand::CreateReg(0, false));
5348 break;
5349 }
5350 default: {
5351 if (isNonFoldablePartialRegisterLoad(LoadMI, MI, MF))
5352 return nullptr;
5353
5354 // Folding a normal load. Just copy the load's address operands.
5355 MOs.append(LoadMI.operands_begin() + NumOps - X86::AddrNumOperands,
5356 LoadMI.operands_begin() + NumOps);
5357 break;
5358 }
5359 }
5360 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, InsertPt,
5361 /*Size=*/0, Alignment, /*AllowCommute=*/true);
5362 }
5363
5364 static SmallVector<MachineMemOperand *, 2>
extractLoadMMOs(ArrayRef<MachineMemOperand * > MMOs,MachineFunction & MF)5365 extractLoadMMOs(ArrayRef<MachineMemOperand *> MMOs, MachineFunction &MF) {
5366 SmallVector<MachineMemOperand *, 2> LoadMMOs;
5367
5368 for (MachineMemOperand *MMO : MMOs) {
5369 if (!MMO->isLoad())
5370 continue;
5371
5372 if (!MMO->isStore()) {
5373 // Reuse the MMO.
5374 LoadMMOs.push_back(MMO);
5375 } else {
5376 // Clone the MMO and unset the store flag.
5377 LoadMMOs.push_back(MF.getMachineMemOperand(
5378 MMO, MMO->getFlags() & ~MachineMemOperand::MOStore));
5379 }
5380 }
5381
5382 return LoadMMOs;
5383 }
5384
5385 static SmallVector<MachineMemOperand *, 2>
extractStoreMMOs(ArrayRef<MachineMemOperand * > MMOs,MachineFunction & MF)5386 extractStoreMMOs(ArrayRef<MachineMemOperand *> MMOs, MachineFunction &MF) {
5387 SmallVector<MachineMemOperand *, 2> StoreMMOs;
5388
5389 for (MachineMemOperand *MMO : MMOs) {
5390 if (!MMO->isStore())
5391 continue;
5392
5393 if (!MMO->isLoad()) {
5394 // Reuse the MMO.
5395 StoreMMOs.push_back(MMO);
5396 } else {
5397 // Clone the MMO and unset the load flag.
5398 StoreMMOs.push_back(MF.getMachineMemOperand(
5399 MMO, MMO->getFlags() & ~MachineMemOperand::MOLoad));
5400 }
5401 }
5402
5403 return StoreMMOs;
5404 }
5405
getBroadcastOpcode(const X86MemoryFoldTableEntry * I,const TargetRegisterClass * RC,const X86Subtarget & STI)5406 static unsigned getBroadcastOpcode(const X86MemoryFoldTableEntry *I,
5407 const TargetRegisterClass *RC,
5408 const X86Subtarget &STI) {
5409 assert(STI.hasAVX512() && "Expected at least AVX512!");
5410 unsigned SpillSize = STI.getRegisterInfo()->getSpillSize(*RC);
5411 assert((SpillSize == 64 || STI.hasVLX()) &&
5412 "Can't broadcast less than 64 bytes without AVX512VL!");
5413
5414 switch (I->Flags & TB_BCAST_MASK) {
5415 default: llvm_unreachable("Unexpected broadcast type!");
5416 case TB_BCAST_D:
5417 switch (SpillSize) {
5418 default: llvm_unreachable("Unknown spill size");
5419 case 16: return X86::VPBROADCASTDZ128m;
5420 case 32: return X86::VPBROADCASTDZ256m;
5421 case 64: return X86::VPBROADCASTDZm;
5422 }
5423 break;
5424 case TB_BCAST_Q:
5425 switch (SpillSize) {
5426 default: llvm_unreachable("Unknown spill size");
5427 case 16: return X86::VPBROADCASTQZ128m;
5428 case 32: return X86::VPBROADCASTQZ256m;
5429 case 64: return X86::VPBROADCASTQZm;
5430 }
5431 break;
5432 case TB_BCAST_SS:
5433 switch (SpillSize) {
5434 default: llvm_unreachable("Unknown spill size");
5435 case 16: return X86::VBROADCASTSSZ128m;
5436 case 32: return X86::VBROADCASTSSZ256m;
5437 case 64: return X86::VBROADCASTSSZm;
5438 }
5439 break;
5440 case TB_BCAST_SD:
5441 switch (SpillSize) {
5442 default: llvm_unreachable("Unknown spill size");
5443 case 16: return X86::VMOVDDUPZ128rm;
5444 case 32: return X86::VBROADCASTSDZ256m;
5445 case 64: return X86::VBROADCASTSDZm;
5446 }
5447 break;
5448 }
5449 }
5450
unfoldMemoryOperand(MachineFunction & MF,MachineInstr & MI,unsigned Reg,bool UnfoldLoad,bool UnfoldStore,SmallVectorImpl<MachineInstr * > & NewMIs) const5451 bool X86InstrInfo::unfoldMemoryOperand(
5452 MachineFunction &MF, MachineInstr &MI, unsigned Reg, bool UnfoldLoad,
5453 bool UnfoldStore, SmallVectorImpl<MachineInstr *> &NewMIs) const {
5454 const X86MemoryFoldTableEntry *I = lookupUnfoldTable(MI.getOpcode());
5455 if (I == nullptr)
5456 return false;
5457 unsigned Opc = I->DstOp;
5458 unsigned Index = I->Flags & TB_INDEX_MASK;
5459 bool FoldedLoad = I->Flags & TB_FOLDED_LOAD;
5460 bool FoldedStore = I->Flags & TB_FOLDED_STORE;
5461 bool FoldedBCast = I->Flags & TB_FOLDED_BCAST;
5462 if (UnfoldLoad && !FoldedLoad)
5463 return false;
5464 UnfoldLoad &= FoldedLoad;
5465 if (UnfoldStore && !FoldedStore)
5466 return false;
5467 UnfoldStore &= FoldedStore;
5468
5469 const MCInstrDesc &MCID = get(Opc);
5470
5471 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
5472 const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
5473 // TODO: Check if 32-byte or greater accesses are slow too?
5474 if (!MI.hasOneMemOperand() && RC == &X86::VR128RegClass &&
5475 Subtarget.isUnalignedMem16Slow())
5476 // Without memoperands, loadRegFromAddr and storeRegToStackSlot will
5477 // conservatively assume the address is unaligned. That's bad for
5478 // performance.
5479 return false;
5480 SmallVector<MachineOperand, X86::AddrNumOperands> AddrOps;
5481 SmallVector<MachineOperand,2> BeforeOps;
5482 SmallVector<MachineOperand,2> AfterOps;
5483 SmallVector<MachineOperand,4> ImpOps;
5484 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
5485 MachineOperand &Op = MI.getOperand(i);
5486 if (i >= Index && i < Index + X86::AddrNumOperands)
5487 AddrOps.push_back(Op);
5488 else if (Op.isReg() && Op.isImplicit())
5489 ImpOps.push_back(Op);
5490 else if (i < Index)
5491 BeforeOps.push_back(Op);
5492 else if (i > Index)
5493 AfterOps.push_back(Op);
5494 }
5495
5496 // Emit the load or broadcast instruction.
5497 if (UnfoldLoad) {
5498 auto MMOs = extractLoadMMOs(MI.memoperands(), MF);
5499
5500 unsigned Opc;
5501 if (FoldedBCast) {
5502 Opc = getBroadcastOpcode(I, RC, Subtarget);
5503 } else {
5504 unsigned Alignment = std::max<uint32_t>(TRI.getSpillSize(*RC), 16);
5505 bool isAligned = !MMOs.empty() && MMOs.front()->getAlignment() >= Alignment;
5506 Opc = getLoadRegOpcode(Reg, RC, isAligned, Subtarget);
5507 }
5508
5509 DebugLoc DL;
5510 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc), Reg);
5511 for (unsigned i = 0, e = AddrOps.size(); i != e; ++i)
5512 MIB.add(AddrOps[i]);
5513 MIB.setMemRefs(MMOs);
5514 NewMIs.push_back(MIB);
5515
5516 if (UnfoldStore) {
5517 // Address operands cannot be marked isKill.
5518 for (unsigned i = 1; i != 1 + X86::AddrNumOperands; ++i) {
5519 MachineOperand &MO = NewMIs[0]->getOperand(i);
5520 if (MO.isReg())
5521 MO.setIsKill(false);
5522 }
5523 }
5524 }
5525
5526 // Emit the data processing instruction.
5527 MachineInstr *DataMI = MF.CreateMachineInstr(MCID, MI.getDebugLoc(), true);
5528 MachineInstrBuilder MIB(MF, DataMI);
5529
5530 if (FoldedStore)
5531 MIB.addReg(Reg, RegState::Define);
5532 for (MachineOperand &BeforeOp : BeforeOps)
5533 MIB.add(BeforeOp);
5534 if (FoldedLoad)
5535 MIB.addReg(Reg);
5536 for (MachineOperand &AfterOp : AfterOps)
5537 MIB.add(AfterOp);
5538 for (MachineOperand &ImpOp : ImpOps) {
5539 MIB.addReg(ImpOp.getReg(),
5540 getDefRegState(ImpOp.isDef()) |
5541 RegState::Implicit |
5542 getKillRegState(ImpOp.isKill()) |
5543 getDeadRegState(ImpOp.isDead()) |
5544 getUndefRegState(ImpOp.isUndef()));
5545 }
5546 // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
5547 switch (DataMI->getOpcode()) {
5548 default: break;
5549 case X86::CMP64ri32:
5550 case X86::CMP64ri8:
5551 case X86::CMP32ri:
5552 case X86::CMP32ri8:
5553 case X86::CMP16ri:
5554 case X86::CMP16ri8:
5555 case X86::CMP8ri: {
5556 MachineOperand &MO0 = DataMI->getOperand(0);
5557 MachineOperand &MO1 = DataMI->getOperand(1);
5558 if (MO1.getImm() == 0) {
5559 unsigned NewOpc;
5560 switch (DataMI->getOpcode()) {
5561 default: llvm_unreachable("Unreachable!");
5562 case X86::CMP64ri8:
5563 case X86::CMP64ri32: NewOpc = X86::TEST64rr; break;
5564 case X86::CMP32ri8:
5565 case X86::CMP32ri: NewOpc = X86::TEST32rr; break;
5566 case X86::CMP16ri8:
5567 case X86::CMP16ri: NewOpc = X86::TEST16rr; break;
5568 case X86::CMP8ri: NewOpc = X86::TEST8rr; break;
5569 }
5570 DataMI->setDesc(get(NewOpc));
5571 MO1.ChangeToRegister(MO0.getReg(), false);
5572 }
5573 }
5574 }
5575 NewMIs.push_back(DataMI);
5576
5577 // Emit the store instruction.
5578 if (UnfoldStore) {
5579 const TargetRegisterClass *DstRC = getRegClass(MCID, 0, &RI, MF);
5580 auto MMOs = extractStoreMMOs(MI.memoperands(), MF);
5581 unsigned Alignment = std::max<uint32_t>(TRI.getSpillSize(*DstRC), 16);
5582 bool isAligned = !MMOs.empty() && MMOs.front()->getAlignment() >= Alignment;
5583 unsigned Opc = getStoreRegOpcode(Reg, DstRC, isAligned, Subtarget);
5584 DebugLoc DL;
5585 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc));
5586 for (unsigned i = 0, e = AddrOps.size(); i != e; ++i)
5587 MIB.add(AddrOps[i]);
5588 MIB.addReg(Reg, RegState::Kill);
5589 MIB.setMemRefs(MMOs);
5590 NewMIs.push_back(MIB);
5591 }
5592
5593 return true;
5594 }
5595
5596 bool
unfoldMemoryOperand(SelectionDAG & DAG,SDNode * N,SmallVectorImpl<SDNode * > & NewNodes) const5597 X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
5598 SmallVectorImpl<SDNode*> &NewNodes) const {
5599 if (!N->isMachineOpcode())
5600 return false;
5601
5602 const X86MemoryFoldTableEntry *I = lookupUnfoldTable(N->getMachineOpcode());
5603 if (I == nullptr)
5604 return false;
5605 unsigned Opc = I->DstOp;
5606 unsigned Index = I->Flags & TB_INDEX_MASK;
5607 bool FoldedLoad = I->Flags & TB_FOLDED_LOAD;
5608 bool FoldedStore = I->Flags & TB_FOLDED_STORE;
5609 bool FoldedBCast = I->Flags & TB_FOLDED_BCAST;
5610 const MCInstrDesc &MCID = get(Opc);
5611 MachineFunction &MF = DAG.getMachineFunction();
5612 const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
5613 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
5614 unsigned NumDefs = MCID.NumDefs;
5615 std::vector<SDValue> AddrOps;
5616 std::vector<SDValue> BeforeOps;
5617 std::vector<SDValue> AfterOps;
5618 SDLoc dl(N);
5619 unsigned NumOps = N->getNumOperands();
5620 for (unsigned i = 0; i != NumOps-1; ++i) {
5621 SDValue Op = N->getOperand(i);
5622 if (i >= Index-NumDefs && i < Index-NumDefs + X86::AddrNumOperands)
5623 AddrOps.push_back(Op);
5624 else if (i < Index-NumDefs)
5625 BeforeOps.push_back(Op);
5626 else if (i > Index-NumDefs)
5627 AfterOps.push_back(Op);
5628 }
5629 SDValue Chain = N->getOperand(NumOps-1);
5630 AddrOps.push_back(Chain);
5631
5632 // Emit the load instruction.
5633 SDNode *Load = nullptr;
5634 if (FoldedLoad) {
5635 EVT VT = *TRI.legalclasstypes_begin(*RC);
5636 auto MMOs = extractLoadMMOs(cast<MachineSDNode>(N)->memoperands(), MF);
5637 if (MMOs.empty() && RC == &X86::VR128RegClass &&
5638 Subtarget.isUnalignedMem16Slow())
5639 // Do not introduce a slow unaligned load.
5640 return false;
5641 // FIXME: If a VR128 can have size 32, we should be checking if a 32-byte
5642 // memory access is slow above.
5643
5644 unsigned Opc;
5645 if (FoldedBCast) {
5646 Opc = getBroadcastOpcode(I, RC, Subtarget);
5647 } else {
5648 unsigned Alignment = std::max<uint32_t>(TRI.getSpillSize(*RC), 16);
5649 bool isAligned = !MMOs.empty() && MMOs.front()->getAlignment() >= Alignment;
5650 Opc = getLoadRegOpcode(0, RC, isAligned, Subtarget);
5651 }
5652
5653 Load = DAG.getMachineNode(Opc, dl, VT, MVT::Other, AddrOps);
5654 NewNodes.push_back(Load);
5655
5656 // Preserve memory reference information.
5657 DAG.setNodeMemRefs(cast<MachineSDNode>(Load), MMOs);
5658 }
5659
5660 // Emit the data processing instruction.
5661 std::vector<EVT> VTs;
5662 const TargetRegisterClass *DstRC = nullptr;
5663 if (MCID.getNumDefs() > 0) {
5664 DstRC = getRegClass(MCID, 0, &RI, MF);
5665 VTs.push_back(*TRI.legalclasstypes_begin(*DstRC));
5666 }
5667 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
5668 EVT VT = N->getValueType(i);
5669 if (VT != MVT::Other && i >= (unsigned)MCID.getNumDefs())
5670 VTs.push_back(VT);
5671 }
5672 if (Load)
5673 BeforeOps.push_back(SDValue(Load, 0));
5674 BeforeOps.insert(BeforeOps.end(), AfterOps.begin(), AfterOps.end());
5675 // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
5676 switch (Opc) {
5677 default: break;
5678 case X86::CMP64ri32:
5679 case X86::CMP64ri8:
5680 case X86::CMP32ri:
5681 case X86::CMP32ri8:
5682 case X86::CMP16ri:
5683 case X86::CMP16ri8:
5684 case X86::CMP8ri:
5685 if (isNullConstant(BeforeOps[1])) {
5686 switch (Opc) {
5687 default: llvm_unreachable("Unreachable!");
5688 case X86::CMP64ri8:
5689 case X86::CMP64ri32: Opc = X86::TEST64rr; break;
5690 case X86::CMP32ri8:
5691 case X86::CMP32ri: Opc = X86::TEST32rr; break;
5692 case X86::CMP16ri8:
5693 case X86::CMP16ri: Opc = X86::TEST16rr; break;
5694 case X86::CMP8ri: Opc = X86::TEST8rr; break;
5695 }
5696 BeforeOps[1] = BeforeOps[0];
5697 }
5698 }
5699 SDNode *NewNode= DAG.getMachineNode(Opc, dl, VTs, BeforeOps);
5700 NewNodes.push_back(NewNode);
5701
5702 // Emit the store instruction.
5703 if (FoldedStore) {
5704 AddrOps.pop_back();
5705 AddrOps.push_back(SDValue(NewNode, 0));
5706 AddrOps.push_back(Chain);
5707 auto MMOs = extractStoreMMOs(cast<MachineSDNode>(N)->memoperands(), MF);
5708 if (MMOs.empty() && RC == &X86::VR128RegClass &&
5709 Subtarget.isUnalignedMem16Slow())
5710 // Do not introduce a slow unaligned store.
5711 return false;
5712 // FIXME: If a VR128 can have size 32, we should be checking if a 32-byte
5713 // memory access is slow above.
5714 unsigned Alignment = std::max<uint32_t>(TRI.getSpillSize(*RC), 16);
5715 bool isAligned = !MMOs.empty() && MMOs.front()->getAlignment() >= Alignment;
5716 SDNode *Store =
5717 DAG.getMachineNode(getStoreRegOpcode(0, DstRC, isAligned, Subtarget),
5718 dl, MVT::Other, AddrOps);
5719 NewNodes.push_back(Store);
5720
5721 // Preserve memory reference information.
5722 DAG.setNodeMemRefs(cast<MachineSDNode>(Store), MMOs);
5723 }
5724
5725 return true;
5726 }
5727
getOpcodeAfterMemoryUnfold(unsigned Opc,bool UnfoldLoad,bool UnfoldStore,unsigned * LoadRegIndex) const5728 unsigned X86InstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc,
5729 bool UnfoldLoad, bool UnfoldStore,
5730 unsigned *LoadRegIndex) const {
5731 const X86MemoryFoldTableEntry *I = lookupUnfoldTable(Opc);
5732 if (I == nullptr)
5733 return 0;
5734 bool FoldedLoad = I->Flags & TB_FOLDED_LOAD;
5735 bool FoldedStore = I->Flags & TB_FOLDED_STORE;
5736 if (UnfoldLoad && !FoldedLoad)
5737 return 0;
5738 if (UnfoldStore && !FoldedStore)
5739 return 0;
5740 if (LoadRegIndex)
5741 *LoadRegIndex = I->Flags & TB_INDEX_MASK;
5742 return I->DstOp;
5743 }
5744
5745 bool
areLoadsFromSameBasePtr(SDNode * Load1,SDNode * Load2,int64_t & Offset1,int64_t & Offset2) const5746 X86InstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
5747 int64_t &Offset1, int64_t &Offset2) const {
5748 if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode())
5749 return false;
5750 unsigned Opc1 = Load1->getMachineOpcode();
5751 unsigned Opc2 = Load2->getMachineOpcode();
5752 switch (Opc1) {
5753 default: return false;
5754 case X86::MOV8rm:
5755 case X86::MOV16rm:
5756 case X86::MOV32rm:
5757 case X86::MOV64rm:
5758 case X86::LD_Fp32m:
5759 case X86::LD_Fp64m:
5760 case X86::LD_Fp80m:
5761 case X86::MOVSSrm:
5762 case X86::MOVSSrm_alt:
5763 case X86::MOVSDrm:
5764 case X86::MOVSDrm_alt:
5765 case X86::MMX_MOVD64rm:
5766 case X86::MMX_MOVQ64rm:
5767 case X86::MOVAPSrm:
5768 case X86::MOVUPSrm:
5769 case X86::MOVAPDrm:
5770 case X86::MOVUPDrm:
5771 case X86::MOVDQArm:
5772 case X86::MOVDQUrm:
5773 // AVX load instructions
5774 case X86::VMOVSSrm:
5775 case X86::VMOVSSrm_alt:
5776 case X86::VMOVSDrm:
5777 case X86::VMOVSDrm_alt:
5778 case X86::VMOVAPSrm:
5779 case X86::VMOVUPSrm:
5780 case X86::VMOVAPDrm:
5781 case X86::VMOVUPDrm:
5782 case X86::VMOVDQArm:
5783 case X86::VMOVDQUrm:
5784 case X86::VMOVAPSYrm:
5785 case X86::VMOVUPSYrm:
5786 case X86::VMOVAPDYrm:
5787 case X86::VMOVUPDYrm:
5788 case X86::VMOVDQAYrm:
5789 case X86::VMOVDQUYrm:
5790 // AVX512 load instructions
5791 case X86::VMOVSSZrm:
5792 case X86::VMOVSSZrm_alt:
5793 case X86::VMOVSDZrm:
5794 case X86::VMOVSDZrm_alt:
5795 case X86::VMOVAPSZ128rm:
5796 case X86::VMOVUPSZ128rm:
5797 case X86::VMOVAPSZ128rm_NOVLX:
5798 case X86::VMOVUPSZ128rm_NOVLX:
5799 case X86::VMOVAPDZ128rm:
5800 case X86::VMOVUPDZ128rm:
5801 case X86::VMOVDQU8Z128rm:
5802 case X86::VMOVDQU16Z128rm:
5803 case X86::VMOVDQA32Z128rm:
5804 case X86::VMOVDQU32Z128rm:
5805 case X86::VMOVDQA64Z128rm:
5806 case X86::VMOVDQU64Z128rm:
5807 case X86::VMOVAPSZ256rm:
5808 case X86::VMOVUPSZ256rm:
5809 case X86::VMOVAPSZ256rm_NOVLX:
5810 case X86::VMOVUPSZ256rm_NOVLX:
5811 case X86::VMOVAPDZ256rm:
5812 case X86::VMOVUPDZ256rm:
5813 case X86::VMOVDQU8Z256rm:
5814 case X86::VMOVDQU16Z256rm:
5815 case X86::VMOVDQA32Z256rm:
5816 case X86::VMOVDQU32Z256rm:
5817 case X86::VMOVDQA64Z256rm:
5818 case X86::VMOVDQU64Z256rm:
5819 case X86::VMOVAPSZrm:
5820 case X86::VMOVUPSZrm:
5821 case X86::VMOVAPDZrm:
5822 case X86::VMOVUPDZrm:
5823 case X86::VMOVDQU8Zrm:
5824 case X86::VMOVDQU16Zrm:
5825 case X86::VMOVDQA32Zrm:
5826 case X86::VMOVDQU32Zrm:
5827 case X86::VMOVDQA64Zrm:
5828 case X86::VMOVDQU64Zrm:
5829 case X86::KMOVBkm:
5830 case X86::KMOVWkm:
5831 case X86::KMOVDkm:
5832 case X86::KMOVQkm:
5833 break;
5834 }
5835 switch (Opc2) {
5836 default: return false;
5837 case X86::MOV8rm:
5838 case X86::MOV16rm:
5839 case X86::MOV32rm:
5840 case X86::MOV64rm:
5841 case X86::LD_Fp32m:
5842 case X86::LD_Fp64m:
5843 case X86::LD_Fp80m:
5844 case X86::MOVSSrm:
5845 case X86::MOVSSrm_alt:
5846 case X86::MOVSDrm:
5847 case X86::MOVSDrm_alt:
5848 case X86::MMX_MOVD64rm:
5849 case X86::MMX_MOVQ64rm:
5850 case X86::MOVAPSrm:
5851 case X86::MOVUPSrm:
5852 case X86::MOVAPDrm:
5853 case X86::MOVUPDrm:
5854 case X86::MOVDQArm:
5855 case X86::MOVDQUrm:
5856 // AVX load instructions
5857 case X86::VMOVSSrm:
5858 case X86::VMOVSSrm_alt:
5859 case X86::VMOVSDrm:
5860 case X86::VMOVSDrm_alt:
5861 case X86::VMOVAPSrm:
5862 case X86::VMOVUPSrm:
5863 case X86::VMOVAPDrm:
5864 case X86::VMOVUPDrm:
5865 case X86::VMOVDQArm:
5866 case X86::VMOVDQUrm:
5867 case X86::VMOVAPSYrm:
5868 case X86::VMOVUPSYrm:
5869 case X86::VMOVAPDYrm:
5870 case X86::VMOVUPDYrm:
5871 case X86::VMOVDQAYrm:
5872 case X86::VMOVDQUYrm:
5873 // AVX512 load instructions
5874 case X86::VMOVSSZrm:
5875 case X86::VMOVSSZrm_alt:
5876 case X86::VMOVSDZrm:
5877 case X86::VMOVSDZrm_alt:
5878 case X86::VMOVAPSZ128rm:
5879 case X86::VMOVUPSZ128rm:
5880 case X86::VMOVAPSZ128rm_NOVLX:
5881 case X86::VMOVUPSZ128rm_NOVLX:
5882 case X86::VMOVAPDZ128rm:
5883 case X86::VMOVUPDZ128rm:
5884 case X86::VMOVDQU8Z128rm:
5885 case X86::VMOVDQU16Z128rm:
5886 case X86::VMOVDQA32Z128rm:
5887 case X86::VMOVDQU32Z128rm:
5888 case X86::VMOVDQA64Z128rm:
5889 case X86::VMOVDQU64Z128rm:
5890 case X86::VMOVAPSZ256rm:
5891 case X86::VMOVUPSZ256rm:
5892 case X86::VMOVAPSZ256rm_NOVLX:
5893 case X86::VMOVUPSZ256rm_NOVLX:
5894 case X86::VMOVAPDZ256rm:
5895 case X86::VMOVUPDZ256rm:
5896 case X86::VMOVDQU8Z256rm:
5897 case X86::VMOVDQU16Z256rm:
5898 case X86::VMOVDQA32Z256rm:
5899 case X86::VMOVDQU32Z256rm:
5900 case X86::VMOVDQA64Z256rm:
5901 case X86::VMOVDQU64Z256rm:
5902 case X86::VMOVAPSZrm:
5903 case X86::VMOVUPSZrm:
5904 case X86::VMOVAPDZrm:
5905 case X86::VMOVUPDZrm:
5906 case X86::VMOVDQU8Zrm:
5907 case X86::VMOVDQU16Zrm:
5908 case X86::VMOVDQA32Zrm:
5909 case X86::VMOVDQU32Zrm:
5910 case X86::VMOVDQA64Zrm:
5911 case X86::VMOVDQU64Zrm:
5912 case X86::KMOVBkm:
5913 case X86::KMOVWkm:
5914 case X86::KMOVDkm:
5915 case X86::KMOVQkm:
5916 break;
5917 }
5918
5919 // Lambda to check if both the loads have the same value for an operand index.
5920 auto HasSameOp = [&](int I) {
5921 return Load1->getOperand(I) == Load2->getOperand(I);
5922 };
5923
5924 // All operands except the displacement should match.
5925 if (!HasSameOp(X86::AddrBaseReg) || !HasSameOp(X86::AddrScaleAmt) ||
5926 !HasSameOp(X86::AddrIndexReg) || !HasSameOp(X86::AddrSegmentReg))
5927 return false;
5928
5929 // Chain Operand must be the same.
5930 if (!HasSameOp(5))
5931 return false;
5932
5933 // Now let's examine if the displacements are constants.
5934 auto Disp1 = dyn_cast<ConstantSDNode>(Load1->getOperand(X86::AddrDisp));
5935 auto Disp2 = dyn_cast<ConstantSDNode>(Load2->getOperand(X86::AddrDisp));
5936 if (!Disp1 || !Disp2)
5937 return false;
5938
5939 Offset1 = Disp1->getSExtValue();
5940 Offset2 = Disp2->getSExtValue();
5941 return true;
5942 }
5943
shouldScheduleLoadsNear(SDNode * Load1,SDNode * Load2,int64_t Offset1,int64_t Offset2,unsigned NumLoads) const5944 bool X86InstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
5945 int64_t Offset1, int64_t Offset2,
5946 unsigned NumLoads) const {
5947 assert(Offset2 > Offset1);
5948 if ((Offset2 - Offset1) / 8 > 64)
5949 return false;
5950
5951 unsigned Opc1 = Load1->getMachineOpcode();
5952 unsigned Opc2 = Load2->getMachineOpcode();
5953 if (Opc1 != Opc2)
5954 return false; // FIXME: overly conservative?
5955
5956 switch (Opc1) {
5957 default: break;
5958 case X86::LD_Fp32m:
5959 case X86::LD_Fp64m:
5960 case X86::LD_Fp80m:
5961 case X86::MMX_MOVD64rm:
5962 case X86::MMX_MOVQ64rm:
5963 return false;
5964 }
5965
5966 EVT VT = Load1->getValueType(0);
5967 switch (VT.getSimpleVT().SimpleTy) {
5968 default:
5969 // XMM registers. In 64-bit mode we can be a bit more aggressive since we
5970 // have 16 of them to play with.
5971 if (Subtarget.is64Bit()) {
5972 if (NumLoads >= 3)
5973 return false;
5974 } else if (NumLoads) {
5975 return false;
5976 }
5977 break;
5978 case MVT::i8:
5979 case MVT::i16:
5980 case MVT::i32:
5981 case MVT::i64:
5982 case MVT::f32:
5983 case MVT::f64:
5984 if (NumLoads)
5985 return false;
5986 break;
5987 }
5988
5989 return true;
5990 }
5991
5992 bool X86InstrInfo::
reverseBranchCondition(SmallVectorImpl<MachineOperand> & Cond) const5993 reverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
5994 assert(Cond.size() == 1 && "Invalid X86 branch condition!");
5995 X86::CondCode CC = static_cast<X86::CondCode>(Cond[0].getImm());
5996 Cond[0].setImm(GetOppositeBranchCondition(CC));
5997 return false;
5998 }
5999
6000 bool X86InstrInfo::
isSafeToMoveRegClassDefs(const TargetRegisterClass * RC) const6001 isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
6002 // FIXME: Return false for x87 stack register classes for now. We can't
6003 // allow any loads of these registers before FpGet_ST0_80.
6004 return !(RC == &X86::CCRRegClass || RC == &X86::DFCCRRegClass ||
6005 RC == &X86::RFP32RegClass || RC == &X86::RFP64RegClass ||
6006 RC == &X86::RFP80RegClass);
6007 }
6008
6009 /// Return a virtual register initialized with the
6010 /// the global base register value. Output instructions required to
6011 /// initialize the register in the function entry block, if necessary.
6012 ///
6013 /// TODO: Eliminate this and move the code to X86MachineFunctionInfo.
6014 ///
getGlobalBaseReg(MachineFunction * MF) const6015 unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const {
6016 assert((!Subtarget.is64Bit() ||
6017 MF->getTarget().getCodeModel() == CodeModel::Medium ||
6018 MF->getTarget().getCodeModel() == CodeModel::Large) &&
6019 "X86-64 PIC uses RIP relative addressing");
6020
6021 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
6022 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
6023 if (GlobalBaseReg != 0)
6024 return GlobalBaseReg;
6025
6026 // Create the register. The code to initialize it is inserted
6027 // later, by the CGBR pass (below).
6028 MachineRegisterInfo &RegInfo = MF->getRegInfo();
6029 GlobalBaseReg = RegInfo.createVirtualRegister(
6030 Subtarget.is64Bit() ? &X86::GR64_NOSPRegClass : &X86::GR32_NOSPRegClass);
6031 X86FI->setGlobalBaseReg(GlobalBaseReg);
6032 return GlobalBaseReg;
6033 }
6034
6035 // These are the replaceable SSE instructions. Some of these have Int variants
6036 // that we don't include here. We don't want to replace instructions selected
6037 // by intrinsics.
6038 static const uint16_t ReplaceableInstrs[][3] = {
6039 //PackedSingle PackedDouble PackedInt
6040 { X86::MOVAPSmr, X86::MOVAPDmr, X86::MOVDQAmr },
6041 { X86::MOVAPSrm, X86::MOVAPDrm, X86::MOVDQArm },
6042 { X86::MOVAPSrr, X86::MOVAPDrr, X86::MOVDQArr },
6043 { X86::MOVUPSmr, X86::MOVUPDmr, X86::MOVDQUmr },
6044 { X86::MOVUPSrm, X86::MOVUPDrm, X86::MOVDQUrm },
6045 { X86::MOVLPSmr, X86::MOVLPDmr, X86::MOVPQI2QImr },
6046 { X86::MOVSDmr, X86::MOVSDmr, X86::MOVPQI2QImr },
6047 { X86::MOVSSmr, X86::MOVSSmr, X86::MOVPDI2DImr },
6048 { X86::MOVSDrm, X86::MOVSDrm, X86::MOVQI2PQIrm },
6049 { X86::MOVSDrm_alt,X86::MOVSDrm_alt,X86::MOVQI2PQIrm },
6050 { X86::MOVSSrm, X86::MOVSSrm, X86::MOVDI2PDIrm },
6051 { X86::MOVSSrm_alt,X86::MOVSSrm_alt,X86::MOVDI2PDIrm },
6052 { X86::MOVNTPSmr, X86::MOVNTPDmr, X86::MOVNTDQmr },
6053 { X86::ANDNPSrm, X86::ANDNPDrm, X86::PANDNrm },
6054 { X86::ANDNPSrr, X86::ANDNPDrr, X86::PANDNrr },
6055 { X86::ANDPSrm, X86::ANDPDrm, X86::PANDrm },
6056 { X86::ANDPSrr, X86::ANDPDrr, X86::PANDrr },
6057 { X86::ORPSrm, X86::ORPDrm, X86::PORrm },
6058 { X86::ORPSrr, X86::ORPDrr, X86::PORrr },
6059 { X86::XORPSrm, X86::XORPDrm, X86::PXORrm },
6060 { X86::XORPSrr, X86::XORPDrr, X86::PXORrr },
6061 { X86::UNPCKLPDrm, X86::UNPCKLPDrm, X86::PUNPCKLQDQrm },
6062 { X86::MOVLHPSrr, X86::UNPCKLPDrr, X86::PUNPCKLQDQrr },
6063 { X86::UNPCKHPDrm, X86::UNPCKHPDrm, X86::PUNPCKHQDQrm },
6064 { X86::UNPCKHPDrr, X86::UNPCKHPDrr, X86::PUNPCKHQDQrr },
6065 { X86::UNPCKLPSrm, X86::UNPCKLPSrm, X86::PUNPCKLDQrm },
6066 { X86::UNPCKLPSrr, X86::UNPCKLPSrr, X86::PUNPCKLDQrr },
6067 { X86::UNPCKHPSrm, X86::UNPCKHPSrm, X86::PUNPCKHDQrm },
6068 { X86::UNPCKHPSrr, X86::UNPCKHPSrr, X86::PUNPCKHDQrr },
6069 { X86::EXTRACTPSmr, X86::EXTRACTPSmr, X86::PEXTRDmr },
6070 { X86::EXTRACTPSrr, X86::EXTRACTPSrr, X86::PEXTRDrr },
6071 // AVX 128-bit support
6072 { X86::VMOVAPSmr, X86::VMOVAPDmr, X86::VMOVDQAmr },
6073 { X86::VMOVAPSrm, X86::VMOVAPDrm, X86::VMOVDQArm },
6074 { X86::VMOVAPSrr, X86::VMOVAPDrr, X86::VMOVDQArr },
6075 { X86::VMOVUPSmr, X86::VMOVUPDmr, X86::VMOVDQUmr },
6076 { X86::VMOVUPSrm, X86::VMOVUPDrm, X86::VMOVDQUrm },
6077 { X86::VMOVLPSmr, X86::VMOVLPDmr, X86::VMOVPQI2QImr },
6078 { X86::VMOVSDmr, X86::VMOVSDmr, X86::VMOVPQI2QImr },
6079 { X86::VMOVSSmr, X86::VMOVSSmr, X86::VMOVPDI2DImr },
6080 { X86::VMOVSDrm, X86::VMOVSDrm, X86::VMOVQI2PQIrm },
6081 { X86::VMOVSDrm_alt,X86::VMOVSDrm_alt,X86::VMOVQI2PQIrm },
6082 { X86::VMOVSSrm, X86::VMOVSSrm, X86::VMOVDI2PDIrm },
6083 { X86::VMOVSSrm_alt,X86::VMOVSSrm_alt,X86::VMOVDI2PDIrm },
6084 { X86::VMOVNTPSmr, X86::VMOVNTPDmr, X86::VMOVNTDQmr },
6085 { X86::VANDNPSrm, X86::VANDNPDrm, X86::VPANDNrm },
6086 { X86::VANDNPSrr, X86::VANDNPDrr, X86::VPANDNrr },
6087 { X86::VANDPSrm, X86::VANDPDrm, X86::VPANDrm },
6088 { X86::VANDPSrr, X86::VANDPDrr, X86::VPANDrr },
6089 { X86::VORPSrm, X86::VORPDrm, X86::VPORrm },
6090 { X86::VORPSrr, X86::VORPDrr, X86::VPORrr },
6091 { X86::VXORPSrm, X86::VXORPDrm, X86::VPXORrm },
6092 { X86::VXORPSrr, X86::VXORPDrr, X86::VPXORrr },
6093 { X86::VUNPCKLPDrm, X86::VUNPCKLPDrm, X86::VPUNPCKLQDQrm },
6094 { X86::VMOVLHPSrr, X86::VUNPCKLPDrr, X86::VPUNPCKLQDQrr },
6095 { X86::VUNPCKHPDrm, X86::VUNPCKHPDrm, X86::VPUNPCKHQDQrm },
6096 { X86::VUNPCKHPDrr, X86::VUNPCKHPDrr, X86::VPUNPCKHQDQrr },
6097 { X86::VUNPCKLPSrm, X86::VUNPCKLPSrm, X86::VPUNPCKLDQrm },
6098 { X86::VUNPCKLPSrr, X86::VUNPCKLPSrr, X86::VPUNPCKLDQrr },
6099 { X86::VUNPCKHPSrm, X86::VUNPCKHPSrm, X86::VPUNPCKHDQrm },
6100 { X86::VUNPCKHPSrr, X86::VUNPCKHPSrr, X86::VPUNPCKHDQrr },
6101 { X86::VEXTRACTPSmr, X86::VEXTRACTPSmr, X86::VPEXTRDmr },
6102 { X86::VEXTRACTPSrr, X86::VEXTRACTPSrr, X86::VPEXTRDrr },
6103 // AVX 256-bit support
6104 { X86::VMOVAPSYmr, X86::VMOVAPDYmr, X86::VMOVDQAYmr },
6105 { X86::VMOVAPSYrm, X86::VMOVAPDYrm, X86::VMOVDQAYrm },
6106 { X86::VMOVAPSYrr, X86::VMOVAPDYrr, X86::VMOVDQAYrr },
6107 { X86::VMOVUPSYmr, X86::VMOVUPDYmr, X86::VMOVDQUYmr },
6108 { X86::VMOVUPSYrm, X86::VMOVUPDYrm, X86::VMOVDQUYrm },
6109 { X86::VMOVNTPSYmr, X86::VMOVNTPDYmr, X86::VMOVNTDQYmr },
6110 { X86::VPERMPSYrm, X86::VPERMPSYrm, X86::VPERMDYrm },
6111 { X86::VPERMPSYrr, X86::VPERMPSYrr, X86::VPERMDYrr },
6112 { X86::VPERMPDYmi, X86::VPERMPDYmi, X86::VPERMQYmi },
6113 { X86::VPERMPDYri, X86::VPERMPDYri, X86::VPERMQYri },
6114 // AVX512 support
6115 { X86::VMOVLPSZ128mr, X86::VMOVLPDZ128mr, X86::VMOVPQI2QIZmr },
6116 { X86::VMOVNTPSZ128mr, X86::VMOVNTPDZ128mr, X86::VMOVNTDQZ128mr },
6117 { X86::VMOVNTPSZ256mr, X86::VMOVNTPDZ256mr, X86::VMOVNTDQZ256mr },
6118 { X86::VMOVNTPSZmr, X86::VMOVNTPDZmr, X86::VMOVNTDQZmr },
6119 { X86::VMOVSDZmr, X86::VMOVSDZmr, X86::VMOVPQI2QIZmr },
6120 { X86::VMOVSSZmr, X86::VMOVSSZmr, X86::VMOVPDI2DIZmr },
6121 { X86::VMOVSDZrm, X86::VMOVSDZrm, X86::VMOVQI2PQIZrm },
6122 { X86::VMOVSDZrm_alt, X86::VMOVSDZrm_alt, X86::VMOVQI2PQIZrm },
6123 { X86::VMOVSSZrm, X86::VMOVSSZrm, X86::VMOVDI2PDIZrm },
6124 { X86::VMOVSSZrm_alt, X86::VMOVSSZrm_alt, X86::VMOVDI2PDIZrm },
6125 { X86::VBROADCASTSSZ128r, X86::VBROADCASTSSZ128r, X86::VPBROADCASTDZ128r },
6126 { X86::VBROADCASTSSZ128m, X86::VBROADCASTSSZ128m, X86::VPBROADCASTDZ128m },
6127 { X86::VBROADCASTSSZ256r, X86::VBROADCASTSSZ256r, X86::VPBROADCASTDZ256r },
6128 { X86::VBROADCASTSSZ256m, X86::VBROADCASTSSZ256m, X86::VPBROADCASTDZ256m },
6129 { X86::VBROADCASTSSZr, X86::VBROADCASTSSZr, X86::VPBROADCASTDZr },
6130 { X86::VBROADCASTSSZm, X86::VBROADCASTSSZm, X86::VPBROADCASTDZm },
6131 { X86::VMOVDDUPZ128rr, X86::VMOVDDUPZ128rr, X86::VPBROADCASTQZ128r },
6132 { X86::VMOVDDUPZ128rm, X86::VMOVDDUPZ128rm, X86::VPBROADCASTQZ128m },
6133 { X86::VBROADCASTSDZ256r, X86::VBROADCASTSDZ256r, X86::VPBROADCASTQZ256r },
6134 { X86::VBROADCASTSDZ256m, X86::VBROADCASTSDZ256m, X86::VPBROADCASTQZ256m },
6135 { X86::VBROADCASTSDZr, X86::VBROADCASTSDZr, X86::VPBROADCASTQZr },
6136 { X86::VBROADCASTSDZm, X86::VBROADCASTSDZm, X86::VPBROADCASTQZm },
6137 { X86::VINSERTF32x4Zrr, X86::VINSERTF32x4Zrr, X86::VINSERTI32x4Zrr },
6138 { X86::VINSERTF32x4Zrm, X86::VINSERTF32x4Zrm, X86::VINSERTI32x4Zrm },
6139 { X86::VINSERTF32x8Zrr, X86::VINSERTF32x8Zrr, X86::VINSERTI32x8Zrr },
6140 { X86::VINSERTF32x8Zrm, X86::VINSERTF32x8Zrm, X86::VINSERTI32x8Zrm },
6141 { X86::VINSERTF64x2Zrr, X86::VINSERTF64x2Zrr, X86::VINSERTI64x2Zrr },
6142 { X86::VINSERTF64x2Zrm, X86::VINSERTF64x2Zrm, X86::VINSERTI64x2Zrm },
6143 { X86::VINSERTF64x4Zrr, X86::VINSERTF64x4Zrr, X86::VINSERTI64x4Zrr },
6144 { X86::VINSERTF64x4Zrm, X86::VINSERTF64x4Zrm, X86::VINSERTI64x4Zrm },
6145 { X86::VINSERTF32x4Z256rr,X86::VINSERTF32x4Z256rr,X86::VINSERTI32x4Z256rr },
6146 { X86::VINSERTF32x4Z256rm,X86::VINSERTF32x4Z256rm,X86::VINSERTI32x4Z256rm },
6147 { X86::VINSERTF64x2Z256rr,X86::VINSERTF64x2Z256rr,X86::VINSERTI64x2Z256rr },
6148 { X86::VINSERTF64x2Z256rm,X86::VINSERTF64x2Z256rm,X86::VINSERTI64x2Z256rm },
6149 { X86::VEXTRACTF32x4Zrr, X86::VEXTRACTF32x4Zrr, X86::VEXTRACTI32x4Zrr },
6150 { X86::VEXTRACTF32x4Zmr, X86::VEXTRACTF32x4Zmr, X86::VEXTRACTI32x4Zmr },
6151 { X86::VEXTRACTF32x8Zrr, X86::VEXTRACTF32x8Zrr, X86::VEXTRACTI32x8Zrr },
6152 { X86::VEXTRACTF32x8Zmr, X86::VEXTRACTF32x8Zmr, X86::VEXTRACTI32x8Zmr },
6153 { X86::VEXTRACTF64x2Zrr, X86::VEXTRACTF64x2Zrr, X86::VEXTRACTI64x2Zrr },
6154 { X86::VEXTRACTF64x2Zmr, X86::VEXTRACTF64x2Zmr, X86::VEXTRACTI64x2Zmr },
6155 { X86::VEXTRACTF64x4Zrr, X86::VEXTRACTF64x4Zrr, X86::VEXTRACTI64x4Zrr },
6156 { X86::VEXTRACTF64x4Zmr, X86::VEXTRACTF64x4Zmr, X86::VEXTRACTI64x4Zmr },
6157 { X86::VEXTRACTF32x4Z256rr,X86::VEXTRACTF32x4Z256rr,X86::VEXTRACTI32x4Z256rr },
6158 { X86::VEXTRACTF32x4Z256mr,X86::VEXTRACTF32x4Z256mr,X86::VEXTRACTI32x4Z256mr },
6159 { X86::VEXTRACTF64x2Z256rr,X86::VEXTRACTF64x2Z256rr,X86::VEXTRACTI64x2Z256rr },
6160 { X86::VEXTRACTF64x2Z256mr,X86::VEXTRACTF64x2Z256mr,X86::VEXTRACTI64x2Z256mr },
6161 { X86::VPERMILPSmi, X86::VPERMILPSmi, X86::VPSHUFDmi },
6162 { X86::VPERMILPSri, X86::VPERMILPSri, X86::VPSHUFDri },
6163 { X86::VPERMILPSZ128mi, X86::VPERMILPSZ128mi, X86::VPSHUFDZ128mi },
6164 { X86::VPERMILPSZ128ri, X86::VPERMILPSZ128ri, X86::VPSHUFDZ128ri },
6165 { X86::VPERMILPSZ256mi, X86::VPERMILPSZ256mi, X86::VPSHUFDZ256mi },
6166 { X86::VPERMILPSZ256ri, X86::VPERMILPSZ256ri, X86::VPSHUFDZ256ri },
6167 { X86::VPERMILPSZmi, X86::VPERMILPSZmi, X86::VPSHUFDZmi },
6168 { X86::VPERMILPSZri, X86::VPERMILPSZri, X86::VPSHUFDZri },
6169 { X86::VPERMPSZ256rm, X86::VPERMPSZ256rm, X86::VPERMDZ256rm },
6170 { X86::VPERMPSZ256rr, X86::VPERMPSZ256rr, X86::VPERMDZ256rr },
6171 { X86::VPERMPDZ256mi, X86::VPERMPDZ256mi, X86::VPERMQZ256mi },
6172 { X86::VPERMPDZ256ri, X86::VPERMPDZ256ri, X86::VPERMQZ256ri },
6173 { X86::VPERMPDZ256rm, X86::VPERMPDZ256rm, X86::VPERMQZ256rm },
6174 { X86::VPERMPDZ256rr, X86::VPERMPDZ256rr, X86::VPERMQZ256rr },
6175 { X86::VPERMPSZrm, X86::VPERMPSZrm, X86::VPERMDZrm },
6176 { X86::VPERMPSZrr, X86::VPERMPSZrr, X86::VPERMDZrr },
6177 { X86::VPERMPDZmi, X86::VPERMPDZmi, X86::VPERMQZmi },
6178 { X86::VPERMPDZri, X86::VPERMPDZri, X86::VPERMQZri },
6179 { X86::VPERMPDZrm, X86::VPERMPDZrm, X86::VPERMQZrm },
6180 { X86::VPERMPDZrr, X86::VPERMPDZrr, X86::VPERMQZrr },
6181 { X86::VUNPCKLPDZ256rm, X86::VUNPCKLPDZ256rm, X86::VPUNPCKLQDQZ256rm },
6182 { X86::VUNPCKLPDZ256rr, X86::VUNPCKLPDZ256rr, X86::VPUNPCKLQDQZ256rr },
6183 { X86::VUNPCKHPDZ256rm, X86::VUNPCKHPDZ256rm, X86::VPUNPCKHQDQZ256rm },
6184 { X86::VUNPCKHPDZ256rr, X86::VUNPCKHPDZ256rr, X86::VPUNPCKHQDQZ256rr },
6185 { X86::VUNPCKLPSZ256rm, X86::VUNPCKLPSZ256rm, X86::VPUNPCKLDQZ256rm },
6186 { X86::VUNPCKLPSZ256rr, X86::VUNPCKLPSZ256rr, X86::VPUNPCKLDQZ256rr },
6187 { X86::VUNPCKHPSZ256rm, X86::VUNPCKHPSZ256rm, X86::VPUNPCKHDQZ256rm },
6188 { X86::VUNPCKHPSZ256rr, X86::VUNPCKHPSZ256rr, X86::VPUNPCKHDQZ256rr },
6189 { X86::VUNPCKLPDZ128rm, X86::VUNPCKLPDZ128rm, X86::VPUNPCKLQDQZ128rm },
6190 { X86::VMOVLHPSZrr, X86::VUNPCKLPDZ128rr, X86::VPUNPCKLQDQZ128rr },
6191 { X86::VUNPCKHPDZ128rm, X86::VUNPCKHPDZ128rm, X86::VPUNPCKHQDQZ128rm },
6192 { X86::VUNPCKHPDZ128rr, X86::VUNPCKHPDZ128rr, X86::VPUNPCKHQDQZ128rr },
6193 { X86::VUNPCKLPSZ128rm, X86::VUNPCKLPSZ128rm, X86::VPUNPCKLDQZ128rm },
6194 { X86::VUNPCKLPSZ128rr, X86::VUNPCKLPSZ128rr, X86::VPUNPCKLDQZ128rr },
6195 { X86::VUNPCKHPSZ128rm, X86::VUNPCKHPSZ128rm, X86::VPUNPCKHDQZ128rm },
6196 { X86::VUNPCKHPSZ128rr, X86::VUNPCKHPSZ128rr, X86::VPUNPCKHDQZ128rr },
6197 { X86::VUNPCKLPDZrm, X86::VUNPCKLPDZrm, X86::VPUNPCKLQDQZrm },
6198 { X86::VUNPCKLPDZrr, X86::VUNPCKLPDZrr, X86::VPUNPCKLQDQZrr },
6199 { X86::VUNPCKHPDZrm, X86::VUNPCKHPDZrm, X86::VPUNPCKHQDQZrm },
6200 { X86::VUNPCKHPDZrr, X86::VUNPCKHPDZrr, X86::VPUNPCKHQDQZrr },
6201 { X86::VUNPCKLPSZrm, X86::VUNPCKLPSZrm, X86::VPUNPCKLDQZrm },
6202 { X86::VUNPCKLPSZrr, X86::VUNPCKLPSZrr, X86::VPUNPCKLDQZrr },
6203 { X86::VUNPCKHPSZrm, X86::VUNPCKHPSZrm, X86::VPUNPCKHDQZrm },
6204 { X86::VUNPCKHPSZrr, X86::VUNPCKHPSZrr, X86::VPUNPCKHDQZrr },
6205 { X86::VEXTRACTPSZmr, X86::VEXTRACTPSZmr, X86::VPEXTRDZmr },
6206 { X86::VEXTRACTPSZrr, X86::VEXTRACTPSZrr, X86::VPEXTRDZrr },
6207 };
6208
6209 static const uint16_t ReplaceableInstrsAVX2[][3] = {
6210 //PackedSingle PackedDouble PackedInt
6211 { X86::VANDNPSYrm, X86::VANDNPDYrm, X86::VPANDNYrm },
6212 { X86::VANDNPSYrr, X86::VANDNPDYrr, X86::VPANDNYrr },
6213 { X86::VANDPSYrm, X86::VANDPDYrm, X86::VPANDYrm },
6214 { X86::VANDPSYrr, X86::VANDPDYrr, X86::VPANDYrr },
6215 { X86::VORPSYrm, X86::VORPDYrm, X86::VPORYrm },
6216 { X86::VORPSYrr, X86::VORPDYrr, X86::VPORYrr },
6217 { X86::VXORPSYrm, X86::VXORPDYrm, X86::VPXORYrm },
6218 { X86::VXORPSYrr, X86::VXORPDYrr, X86::VPXORYrr },
6219 { X86::VPERM2F128rm, X86::VPERM2F128rm, X86::VPERM2I128rm },
6220 { X86::VPERM2F128rr, X86::VPERM2F128rr, X86::VPERM2I128rr },
6221 { X86::VBROADCASTSSrm, X86::VBROADCASTSSrm, X86::VPBROADCASTDrm},
6222 { X86::VBROADCASTSSrr, X86::VBROADCASTSSrr, X86::VPBROADCASTDrr},
6223 { X86::VMOVDDUPrm, X86::VMOVDDUPrm, X86::VPBROADCASTQrm},
6224 { X86::VMOVDDUPrr, X86::VMOVDDUPrr, X86::VPBROADCASTQrr},
6225 { X86::VBROADCASTSSYrr, X86::VBROADCASTSSYrr, X86::VPBROADCASTDYrr},
6226 { X86::VBROADCASTSSYrm, X86::VBROADCASTSSYrm, X86::VPBROADCASTDYrm},
6227 { X86::VBROADCASTSDYrr, X86::VBROADCASTSDYrr, X86::VPBROADCASTQYrr},
6228 { X86::VBROADCASTSDYrm, X86::VBROADCASTSDYrm, X86::VPBROADCASTQYrm},
6229 { X86::VBROADCASTF128, X86::VBROADCASTF128, X86::VBROADCASTI128 },
6230 { X86::VBLENDPSYrri, X86::VBLENDPSYrri, X86::VPBLENDDYrri },
6231 { X86::VBLENDPSYrmi, X86::VBLENDPSYrmi, X86::VPBLENDDYrmi },
6232 { X86::VPERMILPSYmi, X86::VPERMILPSYmi, X86::VPSHUFDYmi },
6233 { X86::VPERMILPSYri, X86::VPERMILPSYri, X86::VPSHUFDYri },
6234 { X86::VUNPCKLPDYrm, X86::VUNPCKLPDYrm, X86::VPUNPCKLQDQYrm },
6235 { X86::VUNPCKLPDYrr, X86::VUNPCKLPDYrr, X86::VPUNPCKLQDQYrr },
6236 { X86::VUNPCKHPDYrm, X86::VUNPCKHPDYrm, X86::VPUNPCKHQDQYrm },
6237 { X86::VUNPCKHPDYrr, X86::VUNPCKHPDYrr, X86::VPUNPCKHQDQYrr },
6238 { X86::VUNPCKLPSYrm, X86::VUNPCKLPSYrm, X86::VPUNPCKLDQYrm },
6239 { X86::VUNPCKLPSYrr, X86::VUNPCKLPSYrr, X86::VPUNPCKLDQYrr },
6240 { X86::VUNPCKHPSYrm, X86::VUNPCKHPSYrm, X86::VPUNPCKHDQYrm },
6241 { X86::VUNPCKHPSYrr, X86::VUNPCKHPSYrr, X86::VPUNPCKHDQYrr },
6242 };
6243
6244 static const uint16_t ReplaceableInstrsFP[][3] = {
6245 //PackedSingle PackedDouble
6246 { X86::MOVLPSrm, X86::MOVLPDrm, X86::INSTRUCTION_LIST_END },
6247 { X86::MOVHPSrm, X86::MOVHPDrm, X86::INSTRUCTION_LIST_END },
6248 { X86::MOVHPSmr, X86::MOVHPDmr, X86::INSTRUCTION_LIST_END },
6249 { X86::VMOVLPSrm, X86::VMOVLPDrm, X86::INSTRUCTION_LIST_END },
6250 { X86::VMOVHPSrm, X86::VMOVHPDrm, X86::INSTRUCTION_LIST_END },
6251 { X86::VMOVHPSmr, X86::VMOVHPDmr, X86::INSTRUCTION_LIST_END },
6252 { X86::VMOVLPSZ128rm, X86::VMOVLPDZ128rm, X86::INSTRUCTION_LIST_END },
6253 { X86::VMOVHPSZ128rm, X86::VMOVHPDZ128rm, X86::INSTRUCTION_LIST_END },
6254 { X86::VMOVHPSZ128mr, X86::VMOVHPDZ128mr, X86::INSTRUCTION_LIST_END },
6255 };
6256
6257 static const uint16_t ReplaceableInstrsAVX2InsertExtract[][3] = {
6258 //PackedSingle PackedDouble PackedInt
6259 { X86::VEXTRACTF128mr, X86::VEXTRACTF128mr, X86::VEXTRACTI128mr },
6260 { X86::VEXTRACTF128rr, X86::VEXTRACTF128rr, X86::VEXTRACTI128rr },
6261 { X86::VINSERTF128rm, X86::VINSERTF128rm, X86::VINSERTI128rm },
6262 { X86::VINSERTF128rr, X86::VINSERTF128rr, X86::VINSERTI128rr },
6263 };
6264
6265 static const uint16_t ReplaceableInstrsAVX512[][4] = {
6266 // Two integer columns for 64-bit and 32-bit elements.
6267 //PackedSingle PackedDouble PackedInt PackedInt
6268 { X86::VMOVAPSZ128mr, X86::VMOVAPDZ128mr, X86::VMOVDQA64Z128mr, X86::VMOVDQA32Z128mr },
6269 { X86::VMOVAPSZ128rm, X86::VMOVAPDZ128rm, X86::VMOVDQA64Z128rm, X86::VMOVDQA32Z128rm },
6270 { X86::VMOVAPSZ128rr, X86::VMOVAPDZ128rr, X86::VMOVDQA64Z128rr, X86::VMOVDQA32Z128rr },
6271 { X86::VMOVUPSZ128mr, X86::VMOVUPDZ128mr, X86::VMOVDQU64Z128mr, X86::VMOVDQU32Z128mr },
6272 { X86::VMOVUPSZ128rm, X86::VMOVUPDZ128rm, X86::VMOVDQU64Z128rm, X86::VMOVDQU32Z128rm },
6273 { X86::VMOVAPSZ256mr, X86::VMOVAPDZ256mr, X86::VMOVDQA64Z256mr, X86::VMOVDQA32Z256mr },
6274 { X86::VMOVAPSZ256rm, X86::VMOVAPDZ256rm, X86::VMOVDQA64Z256rm, X86::VMOVDQA32Z256rm },
6275 { X86::VMOVAPSZ256rr, X86::VMOVAPDZ256rr, X86::VMOVDQA64Z256rr, X86::VMOVDQA32Z256rr },
6276 { X86::VMOVUPSZ256mr, X86::VMOVUPDZ256mr, X86::VMOVDQU64Z256mr, X86::VMOVDQU32Z256mr },
6277 { X86::VMOVUPSZ256rm, X86::VMOVUPDZ256rm, X86::VMOVDQU64Z256rm, X86::VMOVDQU32Z256rm },
6278 { X86::VMOVAPSZmr, X86::VMOVAPDZmr, X86::VMOVDQA64Zmr, X86::VMOVDQA32Zmr },
6279 { X86::VMOVAPSZrm, X86::VMOVAPDZrm, X86::VMOVDQA64Zrm, X86::VMOVDQA32Zrm },
6280 { X86::VMOVAPSZrr, X86::VMOVAPDZrr, X86::VMOVDQA64Zrr, X86::VMOVDQA32Zrr },
6281 { X86::VMOVUPSZmr, X86::VMOVUPDZmr, X86::VMOVDQU64Zmr, X86::VMOVDQU32Zmr },
6282 { X86::VMOVUPSZrm, X86::VMOVUPDZrm, X86::VMOVDQU64Zrm, X86::VMOVDQU32Zrm },
6283 };
6284
6285 static const uint16_t ReplaceableInstrsAVX512DQ[][4] = {
6286 // Two integer columns for 64-bit and 32-bit elements.
6287 //PackedSingle PackedDouble PackedInt PackedInt
6288 { X86::VANDNPSZ128rm, X86::VANDNPDZ128rm, X86::VPANDNQZ128rm, X86::VPANDNDZ128rm },
6289 { X86::VANDNPSZ128rr, X86::VANDNPDZ128rr, X86::VPANDNQZ128rr, X86::VPANDNDZ128rr },
6290 { X86::VANDPSZ128rm, X86::VANDPDZ128rm, X86::VPANDQZ128rm, X86::VPANDDZ128rm },
6291 { X86::VANDPSZ128rr, X86::VANDPDZ128rr, X86::VPANDQZ128rr, X86::VPANDDZ128rr },
6292 { X86::VORPSZ128rm, X86::VORPDZ128rm, X86::VPORQZ128rm, X86::VPORDZ128rm },
6293 { X86::VORPSZ128rr, X86::VORPDZ128rr, X86::VPORQZ128rr, X86::VPORDZ128rr },
6294 { X86::VXORPSZ128rm, X86::VXORPDZ128rm, X86::VPXORQZ128rm, X86::VPXORDZ128rm },
6295 { X86::VXORPSZ128rr, X86::VXORPDZ128rr, X86::VPXORQZ128rr, X86::VPXORDZ128rr },
6296 { X86::VANDNPSZ256rm, X86::VANDNPDZ256rm, X86::VPANDNQZ256rm, X86::VPANDNDZ256rm },
6297 { X86::VANDNPSZ256rr, X86::VANDNPDZ256rr, X86::VPANDNQZ256rr, X86::VPANDNDZ256rr },
6298 { X86::VANDPSZ256rm, X86::VANDPDZ256rm, X86::VPANDQZ256rm, X86::VPANDDZ256rm },
6299 { X86::VANDPSZ256rr, X86::VANDPDZ256rr, X86::VPANDQZ256rr, X86::VPANDDZ256rr },
6300 { X86::VORPSZ256rm, X86::VORPDZ256rm, X86::VPORQZ256rm, X86::VPORDZ256rm },
6301 { X86::VORPSZ256rr, X86::VORPDZ256rr, X86::VPORQZ256rr, X86::VPORDZ256rr },
6302 { X86::VXORPSZ256rm, X86::VXORPDZ256rm, X86::VPXORQZ256rm, X86::VPXORDZ256rm },
6303 { X86::VXORPSZ256rr, X86::VXORPDZ256rr, X86::VPXORQZ256rr, X86::VPXORDZ256rr },
6304 { X86::VANDNPSZrm, X86::VANDNPDZrm, X86::VPANDNQZrm, X86::VPANDNDZrm },
6305 { X86::VANDNPSZrr, X86::VANDNPDZrr, X86::VPANDNQZrr, X86::VPANDNDZrr },
6306 { X86::VANDPSZrm, X86::VANDPDZrm, X86::VPANDQZrm, X86::VPANDDZrm },
6307 { X86::VANDPSZrr, X86::VANDPDZrr, X86::VPANDQZrr, X86::VPANDDZrr },
6308 { X86::VORPSZrm, X86::VORPDZrm, X86::VPORQZrm, X86::VPORDZrm },
6309 { X86::VORPSZrr, X86::VORPDZrr, X86::VPORQZrr, X86::VPORDZrr },
6310 { X86::VXORPSZrm, X86::VXORPDZrm, X86::VPXORQZrm, X86::VPXORDZrm },
6311 { X86::VXORPSZrr, X86::VXORPDZrr, X86::VPXORQZrr, X86::VPXORDZrr },
6312 };
6313
6314 static const uint16_t ReplaceableInstrsAVX512DQMasked[][4] = {
6315 // Two integer columns for 64-bit and 32-bit elements.
6316 //PackedSingle PackedDouble
6317 //PackedInt PackedInt
6318 { X86::VANDNPSZ128rmk, X86::VANDNPDZ128rmk,
6319 X86::VPANDNQZ128rmk, X86::VPANDNDZ128rmk },
6320 { X86::VANDNPSZ128rmkz, X86::VANDNPDZ128rmkz,
6321 X86::VPANDNQZ128rmkz, X86::VPANDNDZ128rmkz },
6322 { X86::VANDNPSZ128rrk, X86::VANDNPDZ128rrk,
6323 X86::VPANDNQZ128rrk, X86::VPANDNDZ128rrk },
6324 { X86::VANDNPSZ128rrkz, X86::VANDNPDZ128rrkz,
6325 X86::VPANDNQZ128rrkz, X86::VPANDNDZ128rrkz },
6326 { X86::VANDPSZ128rmk, X86::VANDPDZ128rmk,
6327 X86::VPANDQZ128rmk, X86::VPANDDZ128rmk },
6328 { X86::VANDPSZ128rmkz, X86::VANDPDZ128rmkz,
6329 X86::VPANDQZ128rmkz, X86::VPANDDZ128rmkz },
6330 { X86::VANDPSZ128rrk, X86::VANDPDZ128rrk,
6331 X86::VPANDQZ128rrk, X86::VPANDDZ128rrk },
6332 { X86::VANDPSZ128rrkz, X86::VANDPDZ128rrkz,
6333 X86::VPANDQZ128rrkz, X86::VPANDDZ128rrkz },
6334 { X86::VORPSZ128rmk, X86::VORPDZ128rmk,
6335 X86::VPORQZ128rmk, X86::VPORDZ128rmk },
6336 { X86::VORPSZ128rmkz, X86::VORPDZ128rmkz,
6337 X86::VPORQZ128rmkz, X86::VPORDZ128rmkz },
6338 { X86::VORPSZ128rrk, X86::VORPDZ128rrk,
6339 X86::VPORQZ128rrk, X86::VPORDZ128rrk },
6340 { X86::VORPSZ128rrkz, X86::VORPDZ128rrkz,
6341 X86::VPORQZ128rrkz, X86::VPORDZ128rrkz },
6342 { X86::VXORPSZ128rmk, X86::VXORPDZ128rmk,
6343 X86::VPXORQZ128rmk, X86::VPXORDZ128rmk },
6344 { X86::VXORPSZ128rmkz, X86::VXORPDZ128rmkz,
6345 X86::VPXORQZ128rmkz, X86::VPXORDZ128rmkz },
6346 { X86::VXORPSZ128rrk, X86::VXORPDZ128rrk,
6347 X86::VPXORQZ128rrk, X86::VPXORDZ128rrk },
6348 { X86::VXORPSZ128rrkz, X86::VXORPDZ128rrkz,
6349 X86::VPXORQZ128rrkz, X86::VPXORDZ128rrkz },
6350 { X86::VANDNPSZ256rmk, X86::VANDNPDZ256rmk,
6351 X86::VPANDNQZ256rmk, X86::VPANDNDZ256rmk },
6352 { X86::VANDNPSZ256rmkz, X86::VANDNPDZ256rmkz,
6353 X86::VPANDNQZ256rmkz, X86::VPANDNDZ256rmkz },
6354 { X86::VANDNPSZ256rrk, X86::VANDNPDZ256rrk,
6355 X86::VPANDNQZ256rrk, X86::VPANDNDZ256rrk },
6356 { X86::VANDNPSZ256rrkz, X86::VANDNPDZ256rrkz,
6357 X86::VPANDNQZ256rrkz, X86::VPANDNDZ256rrkz },
6358 { X86::VANDPSZ256rmk, X86::VANDPDZ256rmk,
6359 X86::VPANDQZ256rmk, X86::VPANDDZ256rmk },
6360 { X86::VANDPSZ256rmkz, X86::VANDPDZ256rmkz,
6361 X86::VPANDQZ256rmkz, X86::VPANDDZ256rmkz },
6362 { X86::VANDPSZ256rrk, X86::VANDPDZ256rrk,
6363 X86::VPANDQZ256rrk, X86::VPANDDZ256rrk },
6364 { X86::VANDPSZ256rrkz, X86::VANDPDZ256rrkz,
6365 X86::VPANDQZ256rrkz, X86::VPANDDZ256rrkz },
6366 { X86::VORPSZ256rmk, X86::VORPDZ256rmk,
6367 X86::VPORQZ256rmk, X86::VPORDZ256rmk },
6368 { X86::VORPSZ256rmkz, X86::VORPDZ256rmkz,
6369 X86::VPORQZ256rmkz, X86::VPORDZ256rmkz },
6370 { X86::VORPSZ256rrk, X86::VORPDZ256rrk,
6371 X86::VPORQZ256rrk, X86::VPORDZ256rrk },
6372 { X86::VORPSZ256rrkz, X86::VORPDZ256rrkz,
6373 X86::VPORQZ256rrkz, X86::VPORDZ256rrkz },
6374 { X86::VXORPSZ256rmk, X86::VXORPDZ256rmk,
6375 X86::VPXORQZ256rmk, X86::VPXORDZ256rmk },
6376 { X86::VXORPSZ256rmkz, X86::VXORPDZ256rmkz,
6377 X86::VPXORQZ256rmkz, X86::VPXORDZ256rmkz },
6378 { X86::VXORPSZ256rrk, X86::VXORPDZ256rrk,
6379 X86::VPXORQZ256rrk, X86::VPXORDZ256rrk },
6380 { X86::VXORPSZ256rrkz, X86::VXORPDZ256rrkz,
6381 X86::VPXORQZ256rrkz, X86::VPXORDZ256rrkz },
6382 { X86::VANDNPSZrmk, X86::VANDNPDZrmk,
6383 X86::VPANDNQZrmk, X86::VPANDNDZrmk },
6384 { X86::VANDNPSZrmkz, X86::VANDNPDZrmkz,
6385 X86::VPANDNQZrmkz, X86::VPANDNDZrmkz },
6386 { X86::VANDNPSZrrk, X86::VANDNPDZrrk,
6387 X86::VPANDNQZrrk, X86::VPANDNDZrrk },
6388 { X86::VANDNPSZrrkz, X86::VANDNPDZrrkz,
6389 X86::VPANDNQZrrkz, X86::VPANDNDZrrkz },
6390 { X86::VANDPSZrmk, X86::VANDPDZrmk,
6391 X86::VPANDQZrmk, X86::VPANDDZrmk },
6392 { X86::VANDPSZrmkz, X86::VANDPDZrmkz,
6393 X86::VPANDQZrmkz, X86::VPANDDZrmkz },
6394 { X86::VANDPSZrrk, X86::VANDPDZrrk,
6395 X86::VPANDQZrrk, X86::VPANDDZrrk },
6396 { X86::VANDPSZrrkz, X86::VANDPDZrrkz,
6397 X86::VPANDQZrrkz, X86::VPANDDZrrkz },
6398 { X86::VORPSZrmk, X86::VORPDZrmk,
6399 X86::VPORQZrmk, X86::VPORDZrmk },
6400 { X86::VORPSZrmkz, X86::VORPDZrmkz,
6401 X86::VPORQZrmkz, X86::VPORDZrmkz },
6402 { X86::VORPSZrrk, X86::VORPDZrrk,
6403 X86::VPORQZrrk, X86::VPORDZrrk },
6404 { X86::VORPSZrrkz, X86::VORPDZrrkz,
6405 X86::VPORQZrrkz, X86::VPORDZrrkz },
6406 { X86::VXORPSZrmk, X86::VXORPDZrmk,
6407 X86::VPXORQZrmk, X86::VPXORDZrmk },
6408 { X86::VXORPSZrmkz, X86::VXORPDZrmkz,
6409 X86::VPXORQZrmkz, X86::VPXORDZrmkz },
6410 { X86::VXORPSZrrk, X86::VXORPDZrrk,
6411 X86::VPXORQZrrk, X86::VPXORDZrrk },
6412 { X86::VXORPSZrrkz, X86::VXORPDZrrkz,
6413 X86::VPXORQZrrkz, X86::VPXORDZrrkz },
6414 // Broadcast loads can be handled the same as masked operations to avoid
6415 // changing element size.
6416 { X86::VANDNPSZ128rmb, X86::VANDNPDZ128rmb,
6417 X86::VPANDNQZ128rmb, X86::VPANDNDZ128rmb },
6418 { X86::VANDPSZ128rmb, X86::VANDPDZ128rmb,
6419 X86::VPANDQZ128rmb, X86::VPANDDZ128rmb },
6420 { X86::VORPSZ128rmb, X86::VORPDZ128rmb,
6421 X86::VPORQZ128rmb, X86::VPORDZ128rmb },
6422 { X86::VXORPSZ128rmb, X86::VXORPDZ128rmb,
6423 X86::VPXORQZ128rmb, X86::VPXORDZ128rmb },
6424 { X86::VANDNPSZ256rmb, X86::VANDNPDZ256rmb,
6425 X86::VPANDNQZ256rmb, X86::VPANDNDZ256rmb },
6426 { X86::VANDPSZ256rmb, X86::VANDPDZ256rmb,
6427 X86::VPANDQZ256rmb, X86::VPANDDZ256rmb },
6428 { X86::VORPSZ256rmb, X86::VORPDZ256rmb,
6429 X86::VPORQZ256rmb, X86::VPORDZ256rmb },
6430 { X86::VXORPSZ256rmb, X86::VXORPDZ256rmb,
6431 X86::VPXORQZ256rmb, X86::VPXORDZ256rmb },
6432 { X86::VANDNPSZrmb, X86::VANDNPDZrmb,
6433 X86::VPANDNQZrmb, X86::VPANDNDZrmb },
6434 { X86::VANDPSZrmb, X86::VANDPDZrmb,
6435 X86::VPANDQZrmb, X86::VPANDDZrmb },
6436 { X86::VANDPSZrmb, X86::VANDPDZrmb,
6437 X86::VPANDQZrmb, X86::VPANDDZrmb },
6438 { X86::VORPSZrmb, X86::VORPDZrmb,
6439 X86::VPORQZrmb, X86::VPORDZrmb },
6440 { X86::VXORPSZrmb, X86::VXORPDZrmb,
6441 X86::VPXORQZrmb, X86::VPXORDZrmb },
6442 { X86::VANDNPSZ128rmbk, X86::VANDNPDZ128rmbk,
6443 X86::VPANDNQZ128rmbk, X86::VPANDNDZ128rmbk },
6444 { X86::VANDPSZ128rmbk, X86::VANDPDZ128rmbk,
6445 X86::VPANDQZ128rmbk, X86::VPANDDZ128rmbk },
6446 { X86::VORPSZ128rmbk, X86::VORPDZ128rmbk,
6447 X86::VPORQZ128rmbk, X86::VPORDZ128rmbk },
6448 { X86::VXORPSZ128rmbk, X86::VXORPDZ128rmbk,
6449 X86::VPXORQZ128rmbk, X86::VPXORDZ128rmbk },
6450 { X86::VANDNPSZ256rmbk, X86::VANDNPDZ256rmbk,
6451 X86::VPANDNQZ256rmbk, X86::VPANDNDZ256rmbk },
6452 { X86::VANDPSZ256rmbk, X86::VANDPDZ256rmbk,
6453 X86::VPANDQZ256rmbk, X86::VPANDDZ256rmbk },
6454 { X86::VORPSZ256rmbk, X86::VORPDZ256rmbk,
6455 X86::VPORQZ256rmbk, X86::VPORDZ256rmbk },
6456 { X86::VXORPSZ256rmbk, X86::VXORPDZ256rmbk,
6457 X86::VPXORQZ256rmbk, X86::VPXORDZ256rmbk },
6458 { X86::VANDNPSZrmbk, X86::VANDNPDZrmbk,
6459 X86::VPANDNQZrmbk, X86::VPANDNDZrmbk },
6460 { X86::VANDPSZrmbk, X86::VANDPDZrmbk,
6461 X86::VPANDQZrmbk, X86::VPANDDZrmbk },
6462 { X86::VANDPSZrmbk, X86::VANDPDZrmbk,
6463 X86::VPANDQZrmbk, X86::VPANDDZrmbk },
6464 { X86::VORPSZrmbk, X86::VORPDZrmbk,
6465 X86::VPORQZrmbk, X86::VPORDZrmbk },
6466 { X86::VXORPSZrmbk, X86::VXORPDZrmbk,
6467 X86::VPXORQZrmbk, X86::VPXORDZrmbk },
6468 { X86::VANDNPSZ128rmbkz,X86::VANDNPDZ128rmbkz,
6469 X86::VPANDNQZ128rmbkz,X86::VPANDNDZ128rmbkz},
6470 { X86::VANDPSZ128rmbkz, X86::VANDPDZ128rmbkz,
6471 X86::VPANDQZ128rmbkz, X86::VPANDDZ128rmbkz },
6472 { X86::VORPSZ128rmbkz, X86::VORPDZ128rmbkz,
6473 X86::VPORQZ128rmbkz, X86::VPORDZ128rmbkz },
6474 { X86::VXORPSZ128rmbkz, X86::VXORPDZ128rmbkz,
6475 X86::VPXORQZ128rmbkz, X86::VPXORDZ128rmbkz },
6476 { X86::VANDNPSZ256rmbkz,X86::VANDNPDZ256rmbkz,
6477 X86::VPANDNQZ256rmbkz,X86::VPANDNDZ256rmbkz},
6478 { X86::VANDPSZ256rmbkz, X86::VANDPDZ256rmbkz,
6479 X86::VPANDQZ256rmbkz, X86::VPANDDZ256rmbkz },
6480 { X86::VORPSZ256rmbkz, X86::VORPDZ256rmbkz,
6481 X86::VPORQZ256rmbkz, X86::VPORDZ256rmbkz },
6482 { X86::VXORPSZ256rmbkz, X86::VXORPDZ256rmbkz,
6483 X86::VPXORQZ256rmbkz, X86::VPXORDZ256rmbkz },
6484 { X86::VANDNPSZrmbkz, X86::VANDNPDZrmbkz,
6485 X86::VPANDNQZrmbkz, X86::VPANDNDZrmbkz },
6486 { X86::VANDPSZrmbkz, X86::VANDPDZrmbkz,
6487 X86::VPANDQZrmbkz, X86::VPANDDZrmbkz },
6488 { X86::VANDPSZrmbkz, X86::VANDPDZrmbkz,
6489 X86::VPANDQZrmbkz, X86::VPANDDZrmbkz },
6490 { X86::VORPSZrmbkz, X86::VORPDZrmbkz,
6491 X86::VPORQZrmbkz, X86::VPORDZrmbkz },
6492 { X86::VXORPSZrmbkz, X86::VXORPDZrmbkz,
6493 X86::VPXORQZrmbkz, X86::VPXORDZrmbkz },
6494 };
6495
6496 // NOTE: These should only be used by the custom domain methods.
6497 static const uint16_t ReplaceableBlendInstrs[][3] = {
6498 //PackedSingle PackedDouble PackedInt
6499 { X86::BLENDPSrmi, X86::BLENDPDrmi, X86::PBLENDWrmi },
6500 { X86::BLENDPSrri, X86::BLENDPDrri, X86::PBLENDWrri },
6501 { X86::VBLENDPSrmi, X86::VBLENDPDrmi, X86::VPBLENDWrmi },
6502 { X86::VBLENDPSrri, X86::VBLENDPDrri, X86::VPBLENDWrri },
6503 { X86::VBLENDPSYrmi, X86::VBLENDPDYrmi, X86::VPBLENDWYrmi },
6504 { X86::VBLENDPSYrri, X86::VBLENDPDYrri, X86::VPBLENDWYrri },
6505 };
6506 static const uint16_t ReplaceableBlendAVX2Instrs[][3] = {
6507 //PackedSingle PackedDouble PackedInt
6508 { X86::VBLENDPSrmi, X86::VBLENDPDrmi, X86::VPBLENDDrmi },
6509 { X86::VBLENDPSrri, X86::VBLENDPDrri, X86::VPBLENDDrri },
6510 { X86::VBLENDPSYrmi, X86::VBLENDPDYrmi, X86::VPBLENDDYrmi },
6511 { X86::VBLENDPSYrri, X86::VBLENDPDYrri, X86::VPBLENDDYrri },
6512 };
6513
6514 // Special table for changing EVEX logic instructions to VEX.
6515 // TODO: Should we run EVEX->VEX earlier?
6516 static const uint16_t ReplaceableCustomAVX512LogicInstrs[][4] = {
6517 // Two integer columns for 64-bit and 32-bit elements.
6518 //PackedSingle PackedDouble PackedInt PackedInt
6519 { X86::VANDNPSrm, X86::VANDNPDrm, X86::VPANDNQZ128rm, X86::VPANDNDZ128rm },
6520 { X86::VANDNPSrr, X86::VANDNPDrr, X86::VPANDNQZ128rr, X86::VPANDNDZ128rr },
6521 { X86::VANDPSrm, X86::VANDPDrm, X86::VPANDQZ128rm, X86::VPANDDZ128rm },
6522 { X86::VANDPSrr, X86::VANDPDrr, X86::VPANDQZ128rr, X86::VPANDDZ128rr },
6523 { X86::VORPSrm, X86::VORPDrm, X86::VPORQZ128rm, X86::VPORDZ128rm },
6524 { X86::VORPSrr, X86::VORPDrr, X86::VPORQZ128rr, X86::VPORDZ128rr },
6525 { X86::VXORPSrm, X86::VXORPDrm, X86::VPXORQZ128rm, X86::VPXORDZ128rm },
6526 { X86::VXORPSrr, X86::VXORPDrr, X86::VPXORQZ128rr, X86::VPXORDZ128rr },
6527 { X86::VANDNPSYrm, X86::VANDNPDYrm, X86::VPANDNQZ256rm, X86::VPANDNDZ256rm },
6528 { X86::VANDNPSYrr, X86::VANDNPDYrr, X86::VPANDNQZ256rr, X86::VPANDNDZ256rr },
6529 { X86::VANDPSYrm, X86::VANDPDYrm, X86::VPANDQZ256rm, X86::VPANDDZ256rm },
6530 { X86::VANDPSYrr, X86::VANDPDYrr, X86::VPANDQZ256rr, X86::VPANDDZ256rr },
6531 { X86::VORPSYrm, X86::VORPDYrm, X86::VPORQZ256rm, X86::VPORDZ256rm },
6532 { X86::VORPSYrr, X86::VORPDYrr, X86::VPORQZ256rr, X86::VPORDZ256rr },
6533 { X86::VXORPSYrm, X86::VXORPDYrm, X86::VPXORQZ256rm, X86::VPXORDZ256rm },
6534 { X86::VXORPSYrr, X86::VXORPDYrr, X86::VPXORQZ256rr, X86::VPXORDZ256rr },
6535 };
6536
6537 // FIXME: Some shuffle and unpack instructions have equivalents in different
6538 // domains, but they require a bit more work than just switching opcodes.
6539
lookup(unsigned opcode,unsigned domain,ArrayRef<uint16_t[3]> Table)6540 static const uint16_t *lookup(unsigned opcode, unsigned domain,
6541 ArrayRef<uint16_t[3]> Table) {
6542 for (const uint16_t (&Row)[3] : Table)
6543 if (Row[domain-1] == opcode)
6544 return Row;
6545 return nullptr;
6546 }
6547
lookupAVX512(unsigned opcode,unsigned domain,ArrayRef<uint16_t[4]> Table)6548 static const uint16_t *lookupAVX512(unsigned opcode, unsigned domain,
6549 ArrayRef<uint16_t[4]> Table) {
6550 // If this is the integer domain make sure to check both integer columns.
6551 for (const uint16_t (&Row)[4] : Table)
6552 if (Row[domain-1] == opcode || (domain == 3 && Row[3] == opcode))
6553 return Row;
6554 return nullptr;
6555 }
6556
6557 // Helper to attempt to widen/narrow blend masks.
AdjustBlendMask(unsigned OldMask,unsigned OldWidth,unsigned NewWidth,unsigned * pNewMask=nullptr)6558 static bool AdjustBlendMask(unsigned OldMask, unsigned OldWidth,
6559 unsigned NewWidth, unsigned *pNewMask = nullptr) {
6560 assert(((OldWidth % NewWidth) == 0 || (NewWidth % OldWidth) == 0) &&
6561 "Illegal blend mask scale");
6562 unsigned NewMask = 0;
6563
6564 if ((OldWidth % NewWidth) == 0) {
6565 unsigned Scale = OldWidth / NewWidth;
6566 unsigned SubMask = (1u << Scale) - 1;
6567 for (unsigned i = 0; i != NewWidth; ++i) {
6568 unsigned Sub = (OldMask >> (i * Scale)) & SubMask;
6569 if (Sub == SubMask)
6570 NewMask |= (1u << i);
6571 else if (Sub != 0x0)
6572 return false;
6573 }
6574 } else {
6575 unsigned Scale = NewWidth / OldWidth;
6576 unsigned SubMask = (1u << Scale) - 1;
6577 for (unsigned i = 0; i != OldWidth; ++i) {
6578 if (OldMask & (1 << i)) {
6579 NewMask |= (SubMask << (i * Scale));
6580 }
6581 }
6582 }
6583
6584 if (pNewMask)
6585 *pNewMask = NewMask;
6586 return true;
6587 }
6588
getExecutionDomainCustom(const MachineInstr & MI) const6589 uint16_t X86InstrInfo::getExecutionDomainCustom(const MachineInstr &MI) const {
6590 unsigned Opcode = MI.getOpcode();
6591 unsigned NumOperands = MI.getDesc().getNumOperands();
6592
6593 auto GetBlendDomains = [&](unsigned ImmWidth, bool Is256) {
6594 uint16_t validDomains = 0;
6595 if (MI.getOperand(NumOperands - 1).isImm()) {
6596 unsigned Imm = MI.getOperand(NumOperands - 1).getImm();
6597 if (AdjustBlendMask(Imm, ImmWidth, Is256 ? 8 : 4))
6598 validDomains |= 0x2; // PackedSingle
6599 if (AdjustBlendMask(Imm, ImmWidth, Is256 ? 4 : 2))
6600 validDomains |= 0x4; // PackedDouble
6601 if (!Is256 || Subtarget.hasAVX2())
6602 validDomains |= 0x8; // PackedInt
6603 }
6604 return validDomains;
6605 };
6606
6607 switch (Opcode) {
6608 case X86::BLENDPDrmi:
6609 case X86::BLENDPDrri:
6610 case X86::VBLENDPDrmi:
6611 case X86::VBLENDPDrri:
6612 return GetBlendDomains(2, false);
6613 case X86::VBLENDPDYrmi:
6614 case X86::VBLENDPDYrri:
6615 return GetBlendDomains(4, true);
6616 case X86::BLENDPSrmi:
6617 case X86::BLENDPSrri:
6618 case X86::VBLENDPSrmi:
6619 case X86::VBLENDPSrri:
6620 case X86::VPBLENDDrmi:
6621 case X86::VPBLENDDrri:
6622 return GetBlendDomains(4, false);
6623 case X86::VBLENDPSYrmi:
6624 case X86::VBLENDPSYrri:
6625 case X86::VPBLENDDYrmi:
6626 case X86::VPBLENDDYrri:
6627 return GetBlendDomains(8, true);
6628 case X86::PBLENDWrmi:
6629 case X86::PBLENDWrri:
6630 case X86::VPBLENDWrmi:
6631 case X86::VPBLENDWrri:
6632 // Treat VPBLENDWY as a 128-bit vector as it repeats the lo/hi masks.
6633 case X86::VPBLENDWYrmi:
6634 case X86::VPBLENDWYrri:
6635 return GetBlendDomains(8, false);
6636 case X86::VPANDDZ128rr: case X86::VPANDDZ128rm:
6637 case X86::VPANDDZ256rr: case X86::VPANDDZ256rm:
6638 case X86::VPANDQZ128rr: case X86::VPANDQZ128rm:
6639 case X86::VPANDQZ256rr: case X86::VPANDQZ256rm:
6640 case X86::VPANDNDZ128rr: case X86::VPANDNDZ128rm:
6641 case X86::VPANDNDZ256rr: case X86::VPANDNDZ256rm:
6642 case X86::VPANDNQZ128rr: case X86::VPANDNQZ128rm:
6643 case X86::VPANDNQZ256rr: case X86::VPANDNQZ256rm:
6644 case X86::VPORDZ128rr: case X86::VPORDZ128rm:
6645 case X86::VPORDZ256rr: case X86::VPORDZ256rm:
6646 case X86::VPORQZ128rr: case X86::VPORQZ128rm:
6647 case X86::VPORQZ256rr: case X86::VPORQZ256rm:
6648 case X86::VPXORDZ128rr: case X86::VPXORDZ128rm:
6649 case X86::VPXORDZ256rr: case X86::VPXORDZ256rm:
6650 case X86::VPXORQZ128rr: case X86::VPXORQZ128rm:
6651 case X86::VPXORQZ256rr: case X86::VPXORQZ256rm:
6652 // If we don't have DQI see if we can still switch from an EVEX integer
6653 // instruction to a VEX floating point instruction.
6654 if (Subtarget.hasDQI())
6655 return 0;
6656
6657 if (RI.getEncodingValue(MI.getOperand(0).getReg()) >= 16)
6658 return 0;
6659 if (RI.getEncodingValue(MI.getOperand(1).getReg()) >= 16)
6660 return 0;
6661 // Register forms will have 3 operands. Memory form will have more.
6662 if (NumOperands == 3 &&
6663 RI.getEncodingValue(MI.getOperand(2).getReg()) >= 16)
6664 return 0;
6665
6666 // All domains are valid.
6667 return 0xe;
6668 case X86::MOVHLPSrr:
6669 // We can swap domains when both inputs are the same register.
6670 // FIXME: This doesn't catch all the cases we would like. If the input
6671 // register isn't KILLed by the instruction, the two address instruction
6672 // pass puts a COPY on one input. The other input uses the original
6673 // register. This prevents the same physical register from being used by
6674 // both inputs.
6675 if (MI.getOperand(1).getReg() == MI.getOperand(2).getReg() &&
6676 MI.getOperand(0).getSubReg() == 0 &&
6677 MI.getOperand(1).getSubReg() == 0 &&
6678 MI.getOperand(2).getSubReg() == 0)
6679 return 0x6;
6680 return 0;
6681 case X86::SHUFPDrri:
6682 return 0x6;
6683 }
6684 return 0;
6685 }
6686
setExecutionDomainCustom(MachineInstr & MI,unsigned Domain) const6687 bool X86InstrInfo::setExecutionDomainCustom(MachineInstr &MI,
6688 unsigned Domain) const {
6689 assert(Domain > 0 && Domain < 4 && "Invalid execution domain");
6690 uint16_t dom = (MI.getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
6691 assert(dom && "Not an SSE instruction");
6692
6693 unsigned Opcode = MI.getOpcode();
6694 unsigned NumOperands = MI.getDesc().getNumOperands();
6695
6696 auto SetBlendDomain = [&](unsigned ImmWidth, bool Is256) {
6697 if (MI.getOperand(NumOperands - 1).isImm()) {
6698 unsigned Imm = MI.getOperand(NumOperands - 1).getImm() & 255;
6699 Imm = (ImmWidth == 16 ? ((Imm << 8) | Imm) : Imm);
6700 unsigned NewImm = Imm;
6701
6702 const uint16_t *table = lookup(Opcode, dom, ReplaceableBlendInstrs);
6703 if (!table)
6704 table = lookup(Opcode, dom, ReplaceableBlendAVX2Instrs);
6705
6706 if (Domain == 1) { // PackedSingle
6707 AdjustBlendMask(Imm, ImmWidth, Is256 ? 8 : 4, &NewImm);
6708 } else if (Domain == 2) { // PackedDouble
6709 AdjustBlendMask(Imm, ImmWidth, Is256 ? 4 : 2, &NewImm);
6710 } else if (Domain == 3) { // PackedInt
6711 if (Subtarget.hasAVX2()) {
6712 // If we are already VPBLENDW use that, else use VPBLENDD.
6713 if ((ImmWidth / (Is256 ? 2 : 1)) != 8) {
6714 table = lookup(Opcode, dom, ReplaceableBlendAVX2Instrs);
6715 AdjustBlendMask(Imm, ImmWidth, Is256 ? 8 : 4, &NewImm);
6716 }
6717 } else {
6718 assert(!Is256 && "128-bit vector expected");
6719 AdjustBlendMask(Imm, ImmWidth, 8, &NewImm);
6720 }
6721 }
6722
6723 assert(table && table[Domain - 1] && "Unknown domain op");
6724 MI.setDesc(get(table[Domain - 1]));
6725 MI.getOperand(NumOperands - 1).setImm(NewImm & 255);
6726 }
6727 return true;
6728 };
6729
6730 switch (Opcode) {
6731 case X86::BLENDPDrmi:
6732 case X86::BLENDPDrri:
6733 case X86::VBLENDPDrmi:
6734 case X86::VBLENDPDrri:
6735 return SetBlendDomain(2, false);
6736 case X86::VBLENDPDYrmi:
6737 case X86::VBLENDPDYrri:
6738 return SetBlendDomain(4, true);
6739 case X86::BLENDPSrmi:
6740 case X86::BLENDPSrri:
6741 case X86::VBLENDPSrmi:
6742 case X86::VBLENDPSrri:
6743 case X86::VPBLENDDrmi:
6744 case X86::VPBLENDDrri:
6745 return SetBlendDomain(4, false);
6746 case X86::VBLENDPSYrmi:
6747 case X86::VBLENDPSYrri:
6748 case X86::VPBLENDDYrmi:
6749 case X86::VPBLENDDYrri:
6750 return SetBlendDomain(8, true);
6751 case X86::PBLENDWrmi:
6752 case X86::PBLENDWrri:
6753 case X86::VPBLENDWrmi:
6754 case X86::VPBLENDWrri:
6755 return SetBlendDomain(8, false);
6756 case X86::VPBLENDWYrmi:
6757 case X86::VPBLENDWYrri:
6758 return SetBlendDomain(16, true);
6759 case X86::VPANDDZ128rr: case X86::VPANDDZ128rm:
6760 case X86::VPANDDZ256rr: case X86::VPANDDZ256rm:
6761 case X86::VPANDQZ128rr: case X86::VPANDQZ128rm:
6762 case X86::VPANDQZ256rr: case X86::VPANDQZ256rm:
6763 case X86::VPANDNDZ128rr: case X86::VPANDNDZ128rm:
6764 case X86::VPANDNDZ256rr: case X86::VPANDNDZ256rm:
6765 case X86::VPANDNQZ128rr: case X86::VPANDNQZ128rm:
6766 case X86::VPANDNQZ256rr: case X86::VPANDNQZ256rm:
6767 case X86::VPORDZ128rr: case X86::VPORDZ128rm:
6768 case X86::VPORDZ256rr: case X86::VPORDZ256rm:
6769 case X86::VPORQZ128rr: case X86::VPORQZ128rm:
6770 case X86::VPORQZ256rr: case X86::VPORQZ256rm:
6771 case X86::VPXORDZ128rr: case X86::VPXORDZ128rm:
6772 case X86::VPXORDZ256rr: case X86::VPXORDZ256rm:
6773 case X86::VPXORQZ128rr: case X86::VPXORQZ128rm:
6774 case X86::VPXORQZ256rr: case X86::VPXORQZ256rm: {
6775 // Without DQI, convert EVEX instructions to VEX instructions.
6776 if (Subtarget.hasDQI())
6777 return false;
6778
6779 const uint16_t *table = lookupAVX512(MI.getOpcode(), dom,
6780 ReplaceableCustomAVX512LogicInstrs);
6781 assert(table && "Instruction not found in table?");
6782 // Don't change integer Q instructions to D instructions and
6783 // use D intructions if we started with a PS instruction.
6784 if (Domain == 3 && (dom == 1 || table[3] == MI.getOpcode()))
6785 Domain = 4;
6786 MI.setDesc(get(table[Domain - 1]));
6787 return true;
6788 }
6789 case X86::UNPCKHPDrr:
6790 case X86::MOVHLPSrr:
6791 // We just need to commute the instruction which will switch the domains.
6792 if (Domain != dom && Domain != 3 &&
6793 MI.getOperand(1).getReg() == MI.getOperand(2).getReg() &&
6794 MI.getOperand(0).getSubReg() == 0 &&
6795 MI.getOperand(1).getSubReg() == 0 &&
6796 MI.getOperand(2).getSubReg() == 0) {
6797 commuteInstruction(MI, false);
6798 return true;
6799 }
6800 // We must always return true for MOVHLPSrr.
6801 if (Opcode == X86::MOVHLPSrr)
6802 return true;
6803 break;
6804 case X86::SHUFPDrri: {
6805 if (Domain == 1) {
6806 unsigned Imm = MI.getOperand(3).getImm();
6807 unsigned NewImm = 0x44;
6808 if (Imm & 1) NewImm |= 0x0a;
6809 if (Imm & 2) NewImm |= 0xa0;
6810 MI.getOperand(3).setImm(NewImm);
6811 MI.setDesc(get(X86::SHUFPSrri));
6812 }
6813 return true;
6814 }
6815 }
6816 return false;
6817 }
6818
6819 std::pair<uint16_t, uint16_t>
getExecutionDomain(const MachineInstr & MI) const6820 X86InstrInfo::getExecutionDomain(const MachineInstr &MI) const {
6821 uint16_t domain = (MI.getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
6822 unsigned opcode = MI.getOpcode();
6823 uint16_t validDomains = 0;
6824 if (domain) {
6825 // Attempt to match for custom instructions.
6826 validDomains = getExecutionDomainCustom(MI);
6827 if (validDomains)
6828 return std::make_pair(domain, validDomains);
6829
6830 if (lookup(opcode, domain, ReplaceableInstrs)) {
6831 validDomains = 0xe;
6832 } else if (lookup(opcode, domain, ReplaceableInstrsAVX2)) {
6833 validDomains = Subtarget.hasAVX2() ? 0xe : 0x6;
6834 } else if (lookup(opcode, domain, ReplaceableInstrsFP)) {
6835 validDomains = 0x6;
6836 } else if (lookup(opcode, domain, ReplaceableInstrsAVX2InsertExtract)) {
6837 // Insert/extract instructions should only effect domain if AVX2
6838 // is enabled.
6839 if (!Subtarget.hasAVX2())
6840 return std::make_pair(0, 0);
6841 validDomains = 0xe;
6842 } else if (lookupAVX512(opcode, domain, ReplaceableInstrsAVX512)) {
6843 validDomains = 0xe;
6844 } else if (Subtarget.hasDQI() && lookupAVX512(opcode, domain,
6845 ReplaceableInstrsAVX512DQ)) {
6846 validDomains = 0xe;
6847 } else if (Subtarget.hasDQI()) {
6848 if (const uint16_t *table = lookupAVX512(opcode, domain,
6849 ReplaceableInstrsAVX512DQMasked)) {
6850 if (domain == 1 || (domain == 3 && table[3] == opcode))
6851 validDomains = 0xa;
6852 else
6853 validDomains = 0xc;
6854 }
6855 }
6856 }
6857 return std::make_pair(domain, validDomains);
6858 }
6859
setExecutionDomain(MachineInstr & MI,unsigned Domain) const6860 void X86InstrInfo::setExecutionDomain(MachineInstr &MI, unsigned Domain) const {
6861 assert(Domain>0 && Domain<4 && "Invalid execution domain");
6862 uint16_t dom = (MI.getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
6863 assert(dom && "Not an SSE instruction");
6864
6865 // Attempt to match for custom instructions.
6866 if (setExecutionDomainCustom(MI, Domain))
6867 return;
6868
6869 const uint16_t *table = lookup(MI.getOpcode(), dom, ReplaceableInstrs);
6870 if (!table) { // try the other table
6871 assert((Subtarget.hasAVX2() || Domain < 3) &&
6872 "256-bit vector operations only available in AVX2");
6873 table = lookup(MI.getOpcode(), dom, ReplaceableInstrsAVX2);
6874 }
6875 if (!table) { // try the FP table
6876 table = lookup(MI.getOpcode(), dom, ReplaceableInstrsFP);
6877 assert((!table || Domain < 3) &&
6878 "Can only select PackedSingle or PackedDouble");
6879 }
6880 if (!table) { // try the other table
6881 assert(Subtarget.hasAVX2() &&
6882 "256-bit insert/extract only available in AVX2");
6883 table = lookup(MI.getOpcode(), dom, ReplaceableInstrsAVX2InsertExtract);
6884 }
6885 if (!table) { // try the AVX512 table
6886 assert(Subtarget.hasAVX512() && "Requires AVX-512");
6887 table = lookupAVX512(MI.getOpcode(), dom, ReplaceableInstrsAVX512);
6888 // Don't change integer Q instructions to D instructions.
6889 if (table && Domain == 3 && table[3] == MI.getOpcode())
6890 Domain = 4;
6891 }
6892 if (!table) { // try the AVX512DQ table
6893 assert((Subtarget.hasDQI() || Domain >= 3) && "Requires AVX-512DQ");
6894 table = lookupAVX512(MI.getOpcode(), dom, ReplaceableInstrsAVX512DQ);
6895 // Don't change integer Q instructions to D instructions and
6896 // use D intructions if we started with a PS instruction.
6897 if (table && Domain == 3 && (dom == 1 || table[3] == MI.getOpcode()))
6898 Domain = 4;
6899 }
6900 if (!table) { // try the AVX512DQMasked table
6901 assert((Subtarget.hasDQI() || Domain >= 3) && "Requires AVX-512DQ");
6902 table = lookupAVX512(MI.getOpcode(), dom, ReplaceableInstrsAVX512DQMasked);
6903 if (table && Domain == 3 && (dom == 1 || table[3] == MI.getOpcode()))
6904 Domain = 4;
6905 }
6906 assert(table && "Cannot change domain");
6907 MI.setDesc(get(table[Domain - 1]));
6908 }
6909
6910 /// Return the noop instruction to use for a noop.
getNoop(MCInst & NopInst) const6911 void X86InstrInfo::getNoop(MCInst &NopInst) const {
6912 NopInst.setOpcode(X86::NOOP);
6913 }
6914
isHighLatencyDef(int opc) const6915 bool X86InstrInfo::isHighLatencyDef(int opc) const {
6916 switch (opc) {
6917 default: return false;
6918 case X86::DIVPDrm:
6919 case X86::DIVPDrr:
6920 case X86::DIVPSrm:
6921 case X86::DIVPSrr:
6922 case X86::DIVSDrm:
6923 case X86::DIVSDrm_Int:
6924 case X86::DIVSDrr:
6925 case X86::DIVSDrr_Int:
6926 case X86::DIVSSrm:
6927 case X86::DIVSSrm_Int:
6928 case X86::DIVSSrr:
6929 case X86::DIVSSrr_Int:
6930 case X86::SQRTPDm:
6931 case X86::SQRTPDr:
6932 case X86::SQRTPSm:
6933 case X86::SQRTPSr:
6934 case X86::SQRTSDm:
6935 case X86::SQRTSDm_Int:
6936 case X86::SQRTSDr:
6937 case X86::SQRTSDr_Int:
6938 case X86::SQRTSSm:
6939 case X86::SQRTSSm_Int:
6940 case X86::SQRTSSr:
6941 case X86::SQRTSSr_Int:
6942 // AVX instructions with high latency
6943 case X86::VDIVPDrm:
6944 case X86::VDIVPDrr:
6945 case X86::VDIVPDYrm:
6946 case X86::VDIVPDYrr:
6947 case X86::VDIVPSrm:
6948 case X86::VDIVPSrr:
6949 case X86::VDIVPSYrm:
6950 case X86::VDIVPSYrr:
6951 case X86::VDIVSDrm:
6952 case X86::VDIVSDrm_Int:
6953 case X86::VDIVSDrr:
6954 case X86::VDIVSDrr_Int:
6955 case X86::VDIVSSrm:
6956 case X86::VDIVSSrm_Int:
6957 case X86::VDIVSSrr:
6958 case X86::VDIVSSrr_Int:
6959 case X86::VSQRTPDm:
6960 case X86::VSQRTPDr:
6961 case X86::VSQRTPDYm:
6962 case X86::VSQRTPDYr:
6963 case X86::VSQRTPSm:
6964 case X86::VSQRTPSr:
6965 case X86::VSQRTPSYm:
6966 case X86::VSQRTPSYr:
6967 case X86::VSQRTSDm:
6968 case X86::VSQRTSDm_Int:
6969 case X86::VSQRTSDr:
6970 case X86::VSQRTSDr_Int:
6971 case X86::VSQRTSSm:
6972 case X86::VSQRTSSm_Int:
6973 case X86::VSQRTSSr:
6974 case X86::VSQRTSSr_Int:
6975 // AVX512 instructions with high latency
6976 case X86::VDIVPDZ128rm:
6977 case X86::VDIVPDZ128rmb:
6978 case X86::VDIVPDZ128rmbk:
6979 case X86::VDIVPDZ128rmbkz:
6980 case X86::VDIVPDZ128rmk:
6981 case X86::VDIVPDZ128rmkz:
6982 case X86::VDIVPDZ128rr:
6983 case X86::VDIVPDZ128rrk:
6984 case X86::VDIVPDZ128rrkz:
6985 case X86::VDIVPDZ256rm:
6986 case X86::VDIVPDZ256rmb:
6987 case X86::VDIVPDZ256rmbk:
6988 case X86::VDIVPDZ256rmbkz:
6989 case X86::VDIVPDZ256rmk:
6990 case X86::VDIVPDZ256rmkz:
6991 case X86::VDIVPDZ256rr:
6992 case X86::VDIVPDZ256rrk:
6993 case X86::VDIVPDZ256rrkz:
6994 case X86::VDIVPDZrrb:
6995 case X86::VDIVPDZrrbk:
6996 case X86::VDIVPDZrrbkz:
6997 case X86::VDIVPDZrm:
6998 case X86::VDIVPDZrmb:
6999 case X86::VDIVPDZrmbk:
7000 case X86::VDIVPDZrmbkz:
7001 case X86::VDIVPDZrmk:
7002 case X86::VDIVPDZrmkz:
7003 case X86::VDIVPDZrr:
7004 case X86::VDIVPDZrrk:
7005 case X86::VDIVPDZrrkz:
7006 case X86::VDIVPSZ128rm:
7007 case X86::VDIVPSZ128rmb:
7008 case X86::VDIVPSZ128rmbk:
7009 case X86::VDIVPSZ128rmbkz:
7010 case X86::VDIVPSZ128rmk:
7011 case X86::VDIVPSZ128rmkz:
7012 case X86::VDIVPSZ128rr:
7013 case X86::VDIVPSZ128rrk:
7014 case X86::VDIVPSZ128rrkz:
7015 case X86::VDIVPSZ256rm:
7016 case X86::VDIVPSZ256rmb:
7017 case X86::VDIVPSZ256rmbk:
7018 case X86::VDIVPSZ256rmbkz:
7019 case X86::VDIVPSZ256rmk:
7020 case X86::VDIVPSZ256rmkz:
7021 case X86::VDIVPSZ256rr:
7022 case X86::VDIVPSZ256rrk:
7023 case X86::VDIVPSZ256rrkz:
7024 case X86::VDIVPSZrrb:
7025 case X86::VDIVPSZrrbk:
7026 case X86::VDIVPSZrrbkz:
7027 case X86::VDIVPSZrm:
7028 case X86::VDIVPSZrmb:
7029 case X86::VDIVPSZrmbk:
7030 case X86::VDIVPSZrmbkz:
7031 case X86::VDIVPSZrmk:
7032 case X86::VDIVPSZrmkz:
7033 case X86::VDIVPSZrr:
7034 case X86::VDIVPSZrrk:
7035 case X86::VDIVPSZrrkz:
7036 case X86::VDIVSDZrm:
7037 case X86::VDIVSDZrr:
7038 case X86::VDIVSDZrm_Int:
7039 case X86::VDIVSDZrm_Intk:
7040 case X86::VDIVSDZrm_Intkz:
7041 case X86::VDIVSDZrr_Int:
7042 case X86::VDIVSDZrr_Intk:
7043 case X86::VDIVSDZrr_Intkz:
7044 case X86::VDIVSDZrrb_Int:
7045 case X86::VDIVSDZrrb_Intk:
7046 case X86::VDIVSDZrrb_Intkz:
7047 case X86::VDIVSSZrm:
7048 case X86::VDIVSSZrr:
7049 case X86::VDIVSSZrm_Int:
7050 case X86::VDIVSSZrm_Intk:
7051 case X86::VDIVSSZrm_Intkz:
7052 case X86::VDIVSSZrr_Int:
7053 case X86::VDIVSSZrr_Intk:
7054 case X86::VDIVSSZrr_Intkz:
7055 case X86::VDIVSSZrrb_Int:
7056 case X86::VDIVSSZrrb_Intk:
7057 case X86::VDIVSSZrrb_Intkz:
7058 case X86::VSQRTPDZ128m:
7059 case X86::VSQRTPDZ128mb:
7060 case X86::VSQRTPDZ128mbk:
7061 case X86::VSQRTPDZ128mbkz:
7062 case X86::VSQRTPDZ128mk:
7063 case X86::VSQRTPDZ128mkz:
7064 case X86::VSQRTPDZ128r:
7065 case X86::VSQRTPDZ128rk:
7066 case X86::VSQRTPDZ128rkz:
7067 case X86::VSQRTPDZ256m:
7068 case X86::VSQRTPDZ256mb:
7069 case X86::VSQRTPDZ256mbk:
7070 case X86::VSQRTPDZ256mbkz:
7071 case X86::VSQRTPDZ256mk:
7072 case X86::VSQRTPDZ256mkz:
7073 case X86::VSQRTPDZ256r:
7074 case X86::VSQRTPDZ256rk:
7075 case X86::VSQRTPDZ256rkz:
7076 case X86::VSQRTPDZm:
7077 case X86::VSQRTPDZmb:
7078 case X86::VSQRTPDZmbk:
7079 case X86::VSQRTPDZmbkz:
7080 case X86::VSQRTPDZmk:
7081 case X86::VSQRTPDZmkz:
7082 case X86::VSQRTPDZr:
7083 case X86::VSQRTPDZrb:
7084 case X86::VSQRTPDZrbk:
7085 case X86::VSQRTPDZrbkz:
7086 case X86::VSQRTPDZrk:
7087 case X86::VSQRTPDZrkz:
7088 case X86::VSQRTPSZ128m:
7089 case X86::VSQRTPSZ128mb:
7090 case X86::VSQRTPSZ128mbk:
7091 case X86::VSQRTPSZ128mbkz:
7092 case X86::VSQRTPSZ128mk:
7093 case X86::VSQRTPSZ128mkz:
7094 case X86::VSQRTPSZ128r:
7095 case X86::VSQRTPSZ128rk:
7096 case X86::VSQRTPSZ128rkz:
7097 case X86::VSQRTPSZ256m:
7098 case X86::VSQRTPSZ256mb:
7099 case X86::VSQRTPSZ256mbk:
7100 case X86::VSQRTPSZ256mbkz:
7101 case X86::VSQRTPSZ256mk:
7102 case X86::VSQRTPSZ256mkz:
7103 case X86::VSQRTPSZ256r:
7104 case X86::VSQRTPSZ256rk:
7105 case X86::VSQRTPSZ256rkz:
7106 case X86::VSQRTPSZm:
7107 case X86::VSQRTPSZmb:
7108 case X86::VSQRTPSZmbk:
7109 case X86::VSQRTPSZmbkz:
7110 case X86::VSQRTPSZmk:
7111 case X86::VSQRTPSZmkz:
7112 case X86::VSQRTPSZr:
7113 case X86::VSQRTPSZrb:
7114 case X86::VSQRTPSZrbk:
7115 case X86::VSQRTPSZrbkz:
7116 case X86::VSQRTPSZrk:
7117 case X86::VSQRTPSZrkz:
7118 case X86::VSQRTSDZm:
7119 case X86::VSQRTSDZm_Int:
7120 case X86::VSQRTSDZm_Intk:
7121 case X86::VSQRTSDZm_Intkz:
7122 case X86::VSQRTSDZr:
7123 case X86::VSQRTSDZr_Int:
7124 case X86::VSQRTSDZr_Intk:
7125 case X86::VSQRTSDZr_Intkz:
7126 case X86::VSQRTSDZrb_Int:
7127 case X86::VSQRTSDZrb_Intk:
7128 case X86::VSQRTSDZrb_Intkz:
7129 case X86::VSQRTSSZm:
7130 case X86::VSQRTSSZm_Int:
7131 case X86::VSQRTSSZm_Intk:
7132 case X86::VSQRTSSZm_Intkz:
7133 case X86::VSQRTSSZr:
7134 case X86::VSQRTSSZr_Int:
7135 case X86::VSQRTSSZr_Intk:
7136 case X86::VSQRTSSZr_Intkz:
7137 case X86::VSQRTSSZrb_Int:
7138 case X86::VSQRTSSZrb_Intk:
7139 case X86::VSQRTSSZrb_Intkz:
7140
7141 case X86::VGATHERDPDYrm:
7142 case X86::VGATHERDPDZ128rm:
7143 case X86::VGATHERDPDZ256rm:
7144 case X86::VGATHERDPDZrm:
7145 case X86::VGATHERDPDrm:
7146 case X86::VGATHERDPSYrm:
7147 case X86::VGATHERDPSZ128rm:
7148 case X86::VGATHERDPSZ256rm:
7149 case X86::VGATHERDPSZrm:
7150 case X86::VGATHERDPSrm:
7151 case X86::VGATHERPF0DPDm:
7152 case X86::VGATHERPF0DPSm:
7153 case X86::VGATHERPF0QPDm:
7154 case X86::VGATHERPF0QPSm:
7155 case X86::VGATHERPF1DPDm:
7156 case X86::VGATHERPF1DPSm:
7157 case X86::VGATHERPF1QPDm:
7158 case X86::VGATHERPF1QPSm:
7159 case X86::VGATHERQPDYrm:
7160 case X86::VGATHERQPDZ128rm:
7161 case X86::VGATHERQPDZ256rm:
7162 case X86::VGATHERQPDZrm:
7163 case X86::VGATHERQPDrm:
7164 case X86::VGATHERQPSYrm:
7165 case X86::VGATHERQPSZ128rm:
7166 case X86::VGATHERQPSZ256rm:
7167 case X86::VGATHERQPSZrm:
7168 case X86::VGATHERQPSrm:
7169 case X86::VPGATHERDDYrm:
7170 case X86::VPGATHERDDZ128rm:
7171 case X86::VPGATHERDDZ256rm:
7172 case X86::VPGATHERDDZrm:
7173 case X86::VPGATHERDDrm:
7174 case X86::VPGATHERDQYrm:
7175 case X86::VPGATHERDQZ128rm:
7176 case X86::VPGATHERDQZ256rm:
7177 case X86::VPGATHERDQZrm:
7178 case X86::VPGATHERDQrm:
7179 case X86::VPGATHERQDYrm:
7180 case X86::VPGATHERQDZ128rm:
7181 case X86::VPGATHERQDZ256rm:
7182 case X86::VPGATHERQDZrm:
7183 case X86::VPGATHERQDrm:
7184 case X86::VPGATHERQQYrm:
7185 case X86::VPGATHERQQZ128rm:
7186 case X86::VPGATHERQQZ256rm:
7187 case X86::VPGATHERQQZrm:
7188 case X86::VPGATHERQQrm:
7189 case X86::VSCATTERDPDZ128mr:
7190 case X86::VSCATTERDPDZ256mr:
7191 case X86::VSCATTERDPDZmr:
7192 case X86::VSCATTERDPSZ128mr:
7193 case X86::VSCATTERDPSZ256mr:
7194 case X86::VSCATTERDPSZmr:
7195 case X86::VSCATTERPF0DPDm:
7196 case X86::VSCATTERPF0DPSm:
7197 case X86::VSCATTERPF0QPDm:
7198 case X86::VSCATTERPF0QPSm:
7199 case X86::VSCATTERPF1DPDm:
7200 case X86::VSCATTERPF1DPSm:
7201 case X86::VSCATTERPF1QPDm:
7202 case X86::VSCATTERPF1QPSm:
7203 case X86::VSCATTERQPDZ128mr:
7204 case X86::VSCATTERQPDZ256mr:
7205 case X86::VSCATTERQPDZmr:
7206 case X86::VSCATTERQPSZ128mr:
7207 case X86::VSCATTERQPSZ256mr:
7208 case X86::VSCATTERQPSZmr:
7209 case X86::VPSCATTERDDZ128mr:
7210 case X86::VPSCATTERDDZ256mr:
7211 case X86::VPSCATTERDDZmr:
7212 case X86::VPSCATTERDQZ128mr:
7213 case X86::VPSCATTERDQZ256mr:
7214 case X86::VPSCATTERDQZmr:
7215 case X86::VPSCATTERQDZ128mr:
7216 case X86::VPSCATTERQDZ256mr:
7217 case X86::VPSCATTERQDZmr:
7218 case X86::VPSCATTERQQZ128mr:
7219 case X86::VPSCATTERQQZ256mr:
7220 case X86::VPSCATTERQQZmr:
7221 return true;
7222 }
7223 }
7224
hasHighOperandLatency(const TargetSchedModel & SchedModel,const MachineRegisterInfo * MRI,const MachineInstr & DefMI,unsigned DefIdx,const MachineInstr & UseMI,unsigned UseIdx) const7225 bool X86InstrInfo::hasHighOperandLatency(const TargetSchedModel &SchedModel,
7226 const MachineRegisterInfo *MRI,
7227 const MachineInstr &DefMI,
7228 unsigned DefIdx,
7229 const MachineInstr &UseMI,
7230 unsigned UseIdx) const {
7231 return isHighLatencyDef(DefMI.getOpcode());
7232 }
7233
hasReassociableOperands(const MachineInstr & Inst,const MachineBasicBlock * MBB) const7234 bool X86InstrInfo::hasReassociableOperands(const MachineInstr &Inst,
7235 const MachineBasicBlock *MBB) const {
7236 assert(Inst.getNumExplicitOperands() == 3 && Inst.getNumExplicitDefs() == 1 &&
7237 Inst.getNumDefs() <= 2 && "Reassociation needs binary operators");
7238
7239 // Integer binary math/logic instructions have a third source operand:
7240 // the EFLAGS register. That operand must be both defined here and never
7241 // used; ie, it must be dead. If the EFLAGS operand is live, then we can
7242 // not change anything because rearranging the operands could affect other
7243 // instructions that depend on the exact status flags (zero, sign, etc.)
7244 // that are set by using these particular operands with this operation.
7245 const MachineOperand *FlagDef = Inst.findRegisterDefOperand(X86::EFLAGS);
7246 assert((Inst.getNumDefs() == 1 || FlagDef) &&
7247 "Implicit def isn't flags?");
7248 if (FlagDef && !FlagDef->isDead())
7249 return false;
7250
7251 return TargetInstrInfo::hasReassociableOperands(Inst, MBB);
7252 }
7253
7254 // TODO: There are many more machine instruction opcodes to match:
7255 // 1. Other data types (integer, vectors)
7256 // 2. Other math / logic operations (xor, or)
7257 // 3. Other forms of the same operation (intrinsics and other variants)
isAssociativeAndCommutative(const MachineInstr & Inst) const7258 bool X86InstrInfo::isAssociativeAndCommutative(const MachineInstr &Inst) const {
7259 switch (Inst.getOpcode()) {
7260 case X86::AND8rr:
7261 case X86::AND16rr:
7262 case X86::AND32rr:
7263 case X86::AND64rr:
7264 case X86::OR8rr:
7265 case X86::OR16rr:
7266 case X86::OR32rr:
7267 case X86::OR64rr:
7268 case X86::XOR8rr:
7269 case X86::XOR16rr:
7270 case X86::XOR32rr:
7271 case X86::XOR64rr:
7272 case X86::IMUL16rr:
7273 case X86::IMUL32rr:
7274 case X86::IMUL64rr:
7275 case X86::PANDrr:
7276 case X86::PORrr:
7277 case X86::PXORrr:
7278 case X86::ANDPDrr:
7279 case X86::ANDPSrr:
7280 case X86::ORPDrr:
7281 case X86::ORPSrr:
7282 case X86::XORPDrr:
7283 case X86::XORPSrr:
7284 case X86::PADDBrr:
7285 case X86::PADDWrr:
7286 case X86::PADDDrr:
7287 case X86::PADDQrr:
7288 case X86::PMULLWrr:
7289 case X86::PMULLDrr:
7290 case X86::PMAXSBrr:
7291 case X86::PMAXSDrr:
7292 case X86::PMAXSWrr:
7293 case X86::PMAXUBrr:
7294 case X86::PMAXUDrr:
7295 case X86::PMAXUWrr:
7296 case X86::PMINSBrr:
7297 case X86::PMINSDrr:
7298 case X86::PMINSWrr:
7299 case X86::PMINUBrr:
7300 case X86::PMINUDrr:
7301 case X86::PMINUWrr:
7302 case X86::VPANDrr:
7303 case X86::VPANDYrr:
7304 case X86::VPANDDZ128rr:
7305 case X86::VPANDDZ256rr:
7306 case X86::VPANDDZrr:
7307 case X86::VPANDQZ128rr:
7308 case X86::VPANDQZ256rr:
7309 case X86::VPANDQZrr:
7310 case X86::VPORrr:
7311 case X86::VPORYrr:
7312 case X86::VPORDZ128rr:
7313 case X86::VPORDZ256rr:
7314 case X86::VPORDZrr:
7315 case X86::VPORQZ128rr:
7316 case X86::VPORQZ256rr:
7317 case X86::VPORQZrr:
7318 case X86::VPXORrr:
7319 case X86::VPXORYrr:
7320 case X86::VPXORDZ128rr:
7321 case X86::VPXORDZ256rr:
7322 case X86::VPXORDZrr:
7323 case X86::VPXORQZ128rr:
7324 case X86::VPXORQZ256rr:
7325 case X86::VPXORQZrr:
7326 case X86::VANDPDrr:
7327 case X86::VANDPSrr:
7328 case X86::VANDPDYrr:
7329 case X86::VANDPSYrr:
7330 case X86::VANDPDZ128rr:
7331 case X86::VANDPSZ128rr:
7332 case X86::VANDPDZ256rr:
7333 case X86::VANDPSZ256rr:
7334 case X86::VANDPDZrr:
7335 case X86::VANDPSZrr:
7336 case X86::VORPDrr:
7337 case X86::VORPSrr:
7338 case X86::VORPDYrr:
7339 case X86::VORPSYrr:
7340 case X86::VORPDZ128rr:
7341 case X86::VORPSZ128rr:
7342 case X86::VORPDZ256rr:
7343 case X86::VORPSZ256rr:
7344 case X86::VORPDZrr:
7345 case X86::VORPSZrr:
7346 case X86::VXORPDrr:
7347 case X86::VXORPSrr:
7348 case X86::VXORPDYrr:
7349 case X86::VXORPSYrr:
7350 case X86::VXORPDZ128rr:
7351 case X86::VXORPSZ128rr:
7352 case X86::VXORPDZ256rr:
7353 case X86::VXORPSZ256rr:
7354 case X86::VXORPDZrr:
7355 case X86::VXORPSZrr:
7356 case X86::KADDBrr:
7357 case X86::KADDWrr:
7358 case X86::KADDDrr:
7359 case X86::KADDQrr:
7360 case X86::KANDBrr:
7361 case X86::KANDWrr:
7362 case X86::KANDDrr:
7363 case X86::KANDQrr:
7364 case X86::KORBrr:
7365 case X86::KORWrr:
7366 case X86::KORDrr:
7367 case X86::KORQrr:
7368 case X86::KXORBrr:
7369 case X86::KXORWrr:
7370 case X86::KXORDrr:
7371 case X86::KXORQrr:
7372 case X86::VPADDBrr:
7373 case X86::VPADDWrr:
7374 case X86::VPADDDrr:
7375 case X86::VPADDQrr:
7376 case X86::VPADDBYrr:
7377 case X86::VPADDWYrr:
7378 case X86::VPADDDYrr:
7379 case X86::VPADDQYrr:
7380 case X86::VPADDBZ128rr:
7381 case X86::VPADDWZ128rr:
7382 case X86::VPADDDZ128rr:
7383 case X86::VPADDQZ128rr:
7384 case X86::VPADDBZ256rr:
7385 case X86::VPADDWZ256rr:
7386 case X86::VPADDDZ256rr:
7387 case X86::VPADDQZ256rr:
7388 case X86::VPADDBZrr:
7389 case X86::VPADDWZrr:
7390 case X86::VPADDDZrr:
7391 case X86::VPADDQZrr:
7392 case X86::VPMULLWrr:
7393 case X86::VPMULLWYrr:
7394 case X86::VPMULLWZ128rr:
7395 case X86::VPMULLWZ256rr:
7396 case X86::VPMULLWZrr:
7397 case X86::VPMULLDrr:
7398 case X86::VPMULLDYrr:
7399 case X86::VPMULLDZ128rr:
7400 case X86::VPMULLDZ256rr:
7401 case X86::VPMULLDZrr:
7402 case X86::VPMULLQZ128rr:
7403 case X86::VPMULLQZ256rr:
7404 case X86::VPMULLQZrr:
7405 case X86::VPMAXSBrr:
7406 case X86::VPMAXSBYrr:
7407 case X86::VPMAXSBZ128rr:
7408 case X86::VPMAXSBZ256rr:
7409 case X86::VPMAXSBZrr:
7410 case X86::VPMAXSDrr:
7411 case X86::VPMAXSDYrr:
7412 case X86::VPMAXSDZ128rr:
7413 case X86::VPMAXSDZ256rr:
7414 case X86::VPMAXSDZrr:
7415 case X86::VPMAXSQZ128rr:
7416 case X86::VPMAXSQZ256rr:
7417 case X86::VPMAXSQZrr:
7418 case X86::VPMAXSWrr:
7419 case X86::VPMAXSWYrr:
7420 case X86::VPMAXSWZ128rr:
7421 case X86::VPMAXSWZ256rr:
7422 case X86::VPMAXSWZrr:
7423 case X86::VPMAXUBrr:
7424 case X86::VPMAXUBYrr:
7425 case X86::VPMAXUBZ128rr:
7426 case X86::VPMAXUBZ256rr:
7427 case X86::VPMAXUBZrr:
7428 case X86::VPMAXUDrr:
7429 case X86::VPMAXUDYrr:
7430 case X86::VPMAXUDZ128rr:
7431 case X86::VPMAXUDZ256rr:
7432 case X86::VPMAXUDZrr:
7433 case X86::VPMAXUQZ128rr:
7434 case X86::VPMAXUQZ256rr:
7435 case X86::VPMAXUQZrr:
7436 case X86::VPMAXUWrr:
7437 case X86::VPMAXUWYrr:
7438 case X86::VPMAXUWZ128rr:
7439 case X86::VPMAXUWZ256rr:
7440 case X86::VPMAXUWZrr:
7441 case X86::VPMINSBrr:
7442 case X86::VPMINSBYrr:
7443 case X86::VPMINSBZ128rr:
7444 case X86::VPMINSBZ256rr:
7445 case X86::VPMINSBZrr:
7446 case X86::VPMINSDrr:
7447 case X86::VPMINSDYrr:
7448 case X86::VPMINSDZ128rr:
7449 case X86::VPMINSDZ256rr:
7450 case X86::VPMINSDZrr:
7451 case X86::VPMINSQZ128rr:
7452 case X86::VPMINSQZ256rr:
7453 case X86::VPMINSQZrr:
7454 case X86::VPMINSWrr:
7455 case X86::VPMINSWYrr:
7456 case X86::VPMINSWZ128rr:
7457 case X86::VPMINSWZ256rr:
7458 case X86::VPMINSWZrr:
7459 case X86::VPMINUBrr:
7460 case X86::VPMINUBYrr:
7461 case X86::VPMINUBZ128rr:
7462 case X86::VPMINUBZ256rr:
7463 case X86::VPMINUBZrr:
7464 case X86::VPMINUDrr:
7465 case X86::VPMINUDYrr:
7466 case X86::VPMINUDZ128rr:
7467 case X86::VPMINUDZ256rr:
7468 case X86::VPMINUDZrr:
7469 case X86::VPMINUQZ128rr:
7470 case X86::VPMINUQZ256rr:
7471 case X86::VPMINUQZrr:
7472 case X86::VPMINUWrr:
7473 case X86::VPMINUWYrr:
7474 case X86::VPMINUWZ128rr:
7475 case X86::VPMINUWZ256rr:
7476 case X86::VPMINUWZrr:
7477 // Normal min/max instructions are not commutative because of NaN and signed
7478 // zero semantics, but these are. Thus, there's no need to check for global
7479 // relaxed math; the instructions themselves have the properties we need.
7480 case X86::MAXCPDrr:
7481 case X86::MAXCPSrr:
7482 case X86::MAXCSDrr:
7483 case X86::MAXCSSrr:
7484 case X86::MINCPDrr:
7485 case X86::MINCPSrr:
7486 case X86::MINCSDrr:
7487 case X86::MINCSSrr:
7488 case X86::VMAXCPDrr:
7489 case X86::VMAXCPSrr:
7490 case X86::VMAXCPDYrr:
7491 case X86::VMAXCPSYrr:
7492 case X86::VMAXCPDZ128rr:
7493 case X86::VMAXCPSZ128rr:
7494 case X86::VMAXCPDZ256rr:
7495 case X86::VMAXCPSZ256rr:
7496 case X86::VMAXCPDZrr:
7497 case X86::VMAXCPSZrr:
7498 case X86::VMAXCSDrr:
7499 case X86::VMAXCSSrr:
7500 case X86::VMAXCSDZrr:
7501 case X86::VMAXCSSZrr:
7502 case X86::VMINCPDrr:
7503 case X86::VMINCPSrr:
7504 case X86::VMINCPDYrr:
7505 case X86::VMINCPSYrr:
7506 case X86::VMINCPDZ128rr:
7507 case X86::VMINCPSZ128rr:
7508 case X86::VMINCPDZ256rr:
7509 case X86::VMINCPSZ256rr:
7510 case X86::VMINCPDZrr:
7511 case X86::VMINCPSZrr:
7512 case X86::VMINCSDrr:
7513 case X86::VMINCSSrr:
7514 case X86::VMINCSDZrr:
7515 case X86::VMINCSSZrr:
7516 return true;
7517 case X86::ADDPDrr:
7518 case X86::ADDPSrr:
7519 case X86::ADDSDrr:
7520 case X86::ADDSSrr:
7521 case X86::MULPDrr:
7522 case X86::MULPSrr:
7523 case X86::MULSDrr:
7524 case X86::MULSSrr:
7525 case X86::VADDPDrr:
7526 case X86::VADDPSrr:
7527 case X86::VADDPDYrr:
7528 case X86::VADDPSYrr:
7529 case X86::VADDPDZ128rr:
7530 case X86::VADDPSZ128rr:
7531 case X86::VADDPDZ256rr:
7532 case X86::VADDPSZ256rr:
7533 case X86::VADDPDZrr:
7534 case X86::VADDPSZrr:
7535 case X86::VADDSDrr:
7536 case X86::VADDSSrr:
7537 case X86::VADDSDZrr:
7538 case X86::VADDSSZrr:
7539 case X86::VMULPDrr:
7540 case X86::VMULPSrr:
7541 case X86::VMULPDYrr:
7542 case X86::VMULPSYrr:
7543 case X86::VMULPDZ128rr:
7544 case X86::VMULPSZ128rr:
7545 case X86::VMULPDZ256rr:
7546 case X86::VMULPSZ256rr:
7547 case X86::VMULPDZrr:
7548 case X86::VMULPSZrr:
7549 case X86::VMULSDrr:
7550 case X86::VMULSSrr:
7551 case X86::VMULSDZrr:
7552 case X86::VMULSSZrr:
7553 return Inst.getParent()->getParent()->getTarget().Options.UnsafeFPMath;
7554 default:
7555 return false;
7556 }
7557 }
7558
7559 /// If \p DescribedReg overlaps with the MOVrr instruction's destination
7560 /// register then, if possible, describe the value in terms of the source
7561 /// register.
7562 static Optional<ParamLoadedValue>
describeMOVrrLoadedValue(const MachineInstr & MI,Register DescribedReg,const TargetRegisterInfo * TRI)7563 describeMOVrrLoadedValue(const MachineInstr &MI, Register DescribedReg,
7564 const TargetRegisterInfo *TRI) {
7565 Register DestReg = MI.getOperand(0).getReg();
7566 Register SrcReg = MI.getOperand(1).getReg();
7567
7568 auto Expr = DIExpression::get(MI.getMF()->getFunction().getContext(), {});
7569
7570 // If the described register is the destination, just return the source.
7571 if (DestReg == DescribedReg)
7572 return ParamLoadedValue(MachineOperand::CreateReg(SrcReg, false), Expr);
7573
7574 // If the described register is a sub-register of the destination register,
7575 // then pick out the source register's corresponding sub-register.
7576 if (unsigned SubRegIdx = TRI->getSubRegIndex(DestReg, DescribedReg)) {
7577 unsigned SrcSubReg = TRI->getSubReg(SrcReg, SubRegIdx);
7578 return ParamLoadedValue(MachineOperand::CreateReg(SrcSubReg, false), Expr);
7579 }
7580
7581 // The remaining case to consider is when the described register is a
7582 // super-register of the destination register. MOV8rr and MOV16rr does not
7583 // write to any of the other bytes in the register, meaning that we'd have to
7584 // describe the value using a combination of the source register and the
7585 // non-overlapping bits in the described register, which is not currently
7586 // possible.
7587 if (MI.getOpcode() == X86::MOV8rr || MI.getOpcode() == X86::MOV16rr ||
7588 !TRI->isSuperRegister(DestReg, DescribedReg))
7589 return None;
7590
7591 assert(MI.getOpcode() == X86::MOV32rr && "Unexpected super-register case");
7592 return ParamLoadedValue(MachineOperand::CreateReg(SrcReg, false), Expr);
7593 }
7594
7595 Optional<ParamLoadedValue>
describeLoadedValue(const MachineInstr & MI,Register Reg) const7596 X86InstrInfo::describeLoadedValue(const MachineInstr &MI, Register Reg) const {
7597 const MachineOperand *Op = nullptr;
7598 DIExpression *Expr = nullptr;
7599
7600 const TargetRegisterInfo *TRI = &getRegisterInfo();
7601
7602 switch (MI.getOpcode()) {
7603 case X86::LEA32r:
7604 case X86::LEA64r:
7605 case X86::LEA64_32r: {
7606 // We may need to describe a 64-bit parameter with a 32-bit LEA.
7607 if (!TRI->isSuperRegisterEq(MI.getOperand(0).getReg(), Reg))
7608 return None;
7609
7610 // Operand 4 could be global address. For now we do not support
7611 // such situation.
7612 if (!MI.getOperand(4).isImm() || !MI.getOperand(2).isImm())
7613 return None;
7614
7615 const MachineOperand &Op1 = MI.getOperand(1);
7616 const MachineOperand &Op2 = MI.getOperand(3);
7617 assert(Op2.isReg() && (Op2.getReg() == X86::NoRegister ||
7618 Register::isPhysicalRegister(Op2.getReg())));
7619
7620 // Omit situations like:
7621 // %rsi = lea %rsi, 4, ...
7622 if ((Op1.isReg() && Op1.getReg() == MI.getOperand(0).getReg()) ||
7623 Op2.getReg() == MI.getOperand(0).getReg())
7624 return None;
7625 else if ((Op1.isReg() && Op1.getReg() != X86::NoRegister &&
7626 TRI->regsOverlap(Op1.getReg(), MI.getOperand(0).getReg())) ||
7627 (Op2.getReg() != X86::NoRegister &&
7628 TRI->regsOverlap(Op2.getReg(), MI.getOperand(0).getReg())))
7629 return None;
7630
7631 int64_t Coef = MI.getOperand(2).getImm();
7632 int64_t Offset = MI.getOperand(4).getImm();
7633 SmallVector<uint64_t, 8> Ops;
7634
7635 if ((Op1.isReg() && Op1.getReg() != X86::NoRegister)) {
7636 Op = &Op1;
7637 } else if (Op1.isFI())
7638 Op = &Op1;
7639
7640 if (Op && Op->isReg() && Op->getReg() == Op2.getReg() && Coef > 0) {
7641 Ops.push_back(dwarf::DW_OP_constu);
7642 Ops.push_back(Coef + 1);
7643 Ops.push_back(dwarf::DW_OP_mul);
7644 } else {
7645 if (Op && Op2.getReg() != X86::NoRegister) {
7646 int dwarfReg = TRI->getDwarfRegNum(Op2.getReg(), false);
7647 if (dwarfReg < 0)
7648 return None;
7649 else if (dwarfReg < 32) {
7650 Ops.push_back(dwarf::DW_OP_breg0 + dwarfReg);
7651 Ops.push_back(0);
7652 } else {
7653 Ops.push_back(dwarf::DW_OP_bregx);
7654 Ops.push_back(dwarfReg);
7655 Ops.push_back(0);
7656 }
7657 } else if (!Op) {
7658 assert(Op2.getReg() != X86::NoRegister);
7659 Op = &Op2;
7660 }
7661
7662 if (Coef > 1) {
7663 assert(Op2.getReg() != X86::NoRegister);
7664 Ops.push_back(dwarf::DW_OP_constu);
7665 Ops.push_back(Coef);
7666 Ops.push_back(dwarf::DW_OP_mul);
7667 }
7668
7669 if (((Op1.isReg() && Op1.getReg() != X86::NoRegister) || Op1.isFI()) &&
7670 Op2.getReg() != X86::NoRegister) {
7671 Ops.push_back(dwarf::DW_OP_plus);
7672 }
7673 }
7674
7675 DIExpression::appendOffset(Ops, Offset);
7676 Expr = DIExpression::get(MI.getMF()->getFunction().getContext(), Ops);
7677
7678 return ParamLoadedValue(*Op, Expr);;
7679 }
7680 case X86::MOV32ri:
7681 case X86::MOV64ri:
7682 case X86::MOV64ri32:
7683 // MOV32ri may be used for producing zero-extended 32-bit immediates in
7684 // 64-bit parameters, so we need to consider super-registers.
7685 if (!TRI->isSuperRegisterEq(MI.getOperand(0).getReg(), Reg))
7686 return None;
7687 return ParamLoadedValue(MI.getOperand(1), Expr);
7688 case X86::MOV8rr:
7689 case X86::MOV16rr:
7690 case X86::MOV32rr:
7691 case X86::MOV64rr:
7692 return describeMOVrrLoadedValue(MI, Reg, TRI);
7693 case X86::XOR32rr: {
7694 // 64-bit parameters are zero-materialized using XOR32rr, so also consider
7695 // super-registers.
7696 if (!TRI->isSuperRegisterEq(MI.getOperand(0).getReg(), Reg))
7697 return None;
7698 if (MI.getOperand(1).getReg() == MI.getOperand(2).getReg())
7699 return ParamLoadedValue(MachineOperand::CreateImm(0), Expr);
7700 return None;
7701 }
7702 case X86::MOVSX64rr32: {
7703 // We may need to describe the lower 32 bits of the MOVSX; for example, in
7704 // cases like this:
7705 //
7706 // $ebx = [...]
7707 // $rdi = MOVSX64rr32 $ebx
7708 // $esi = MOV32rr $edi
7709 if (!TRI->isSubRegisterEq(MI.getOperand(0).getReg(), Reg))
7710 return None;
7711
7712 Expr = DIExpression::get(MI.getMF()->getFunction().getContext(), {});
7713
7714 // If the described register is the destination register we need to
7715 // sign-extend the source register from 32 bits. The other case we handle
7716 // is when the described register is the 32-bit sub-register of the
7717 // destination register, in case we just need to return the source
7718 // register.
7719 if (Reg == MI.getOperand(0).getReg())
7720 Expr = DIExpression::appendExt(Expr, 32, 64, true);
7721 else
7722 assert(X86MCRegisterClasses[X86::GR32RegClassID].contains(Reg) &&
7723 "Unhandled sub-register case for MOVSX64rr32");
7724
7725 return ParamLoadedValue(MI.getOperand(1), Expr);
7726 }
7727 default:
7728 assert(!MI.isMoveImmediate() && "Unexpected MoveImm instruction");
7729 return TargetInstrInfo::describeLoadedValue(MI, Reg);
7730 }
7731 }
7732
7733 /// This is an architecture-specific helper function of reassociateOps.
7734 /// Set special operand attributes for new instructions after reassociation.
setSpecialOperandAttr(MachineInstr & OldMI1,MachineInstr & OldMI2,MachineInstr & NewMI1,MachineInstr & NewMI2) const7735 void X86InstrInfo::setSpecialOperandAttr(MachineInstr &OldMI1,
7736 MachineInstr &OldMI2,
7737 MachineInstr &NewMI1,
7738 MachineInstr &NewMI2) const {
7739 // Integer instructions may define an implicit EFLAGS dest register operand.
7740 MachineOperand *OldFlagDef1 = OldMI1.findRegisterDefOperand(X86::EFLAGS);
7741 MachineOperand *OldFlagDef2 = OldMI2.findRegisterDefOperand(X86::EFLAGS);
7742
7743 assert(!OldFlagDef1 == !OldFlagDef2 &&
7744 "Unexpected instruction type for reassociation");
7745
7746 if (!OldFlagDef1 || !OldFlagDef2)
7747 return;
7748
7749 assert(OldFlagDef1->isDead() && OldFlagDef2->isDead() &&
7750 "Must have dead EFLAGS operand in reassociable instruction");
7751
7752 MachineOperand *NewFlagDef1 = NewMI1.findRegisterDefOperand(X86::EFLAGS);
7753 MachineOperand *NewFlagDef2 = NewMI2.findRegisterDefOperand(X86::EFLAGS);
7754
7755 assert(NewFlagDef1 && NewFlagDef2 &&
7756 "Unexpected operand in reassociable instruction");
7757
7758 // Mark the new EFLAGS operands as dead to be helpful to subsequent iterations
7759 // of this pass or other passes. The EFLAGS operands must be dead in these new
7760 // instructions because the EFLAGS operands in the original instructions must
7761 // be dead in order for reassociation to occur.
7762 NewFlagDef1->setIsDead();
7763 NewFlagDef2->setIsDead();
7764 }
7765
7766 std::pair<unsigned, unsigned>
decomposeMachineOperandsTargetFlags(unsigned TF) const7767 X86InstrInfo::decomposeMachineOperandsTargetFlags(unsigned TF) const {
7768 return std::make_pair(TF, 0u);
7769 }
7770
7771 ArrayRef<std::pair<unsigned, const char *>>
getSerializableDirectMachineOperandTargetFlags() const7772 X86InstrInfo::getSerializableDirectMachineOperandTargetFlags() const {
7773 using namespace X86II;
7774 static const std::pair<unsigned, const char *> TargetFlags[] = {
7775 {MO_GOT_ABSOLUTE_ADDRESS, "x86-got-absolute-address"},
7776 {MO_PIC_BASE_OFFSET, "x86-pic-base-offset"},
7777 {MO_GOT, "x86-got"},
7778 {MO_GOTOFF, "x86-gotoff"},
7779 {MO_GOTPCREL, "x86-gotpcrel"},
7780 {MO_PLT, "x86-plt"},
7781 {MO_TLSGD, "x86-tlsgd"},
7782 {MO_TLSLD, "x86-tlsld"},
7783 {MO_TLSLDM, "x86-tlsldm"},
7784 {MO_GOTTPOFF, "x86-gottpoff"},
7785 {MO_INDNTPOFF, "x86-indntpoff"},
7786 {MO_TPOFF, "x86-tpoff"},
7787 {MO_DTPOFF, "x86-dtpoff"},
7788 {MO_NTPOFF, "x86-ntpoff"},
7789 {MO_GOTNTPOFF, "x86-gotntpoff"},
7790 {MO_DLLIMPORT, "x86-dllimport"},
7791 {MO_DARWIN_NONLAZY, "x86-darwin-nonlazy"},
7792 {MO_DARWIN_NONLAZY_PIC_BASE, "x86-darwin-nonlazy-pic-base"},
7793 {MO_TLVP, "x86-tlvp"},
7794 {MO_TLVP_PIC_BASE, "x86-tlvp-pic-base"},
7795 {MO_SECREL, "x86-secrel"},
7796 {MO_COFFSTUB, "x86-coffstub"}};
7797 return makeArrayRef(TargetFlags);
7798 }
7799
7800 namespace {
7801 /// Create Global Base Reg pass. This initializes the PIC
7802 /// global base register for x86-32.
7803 struct CGBR : public MachineFunctionPass {
7804 static char ID;
CGBR__anone5c35a6e0511::CGBR7805 CGBR() : MachineFunctionPass(ID) {}
7806
runOnMachineFunction__anone5c35a6e0511::CGBR7807 bool runOnMachineFunction(MachineFunction &MF) override {
7808 const X86TargetMachine *TM =
7809 static_cast<const X86TargetMachine *>(&MF.getTarget());
7810 const X86Subtarget &STI = MF.getSubtarget<X86Subtarget>();
7811
7812 // Don't do anything in the 64-bit small and kernel code models. They use
7813 // RIP-relative addressing for everything.
7814 if (STI.is64Bit() && (TM->getCodeModel() == CodeModel::Small ||
7815 TM->getCodeModel() == CodeModel::Kernel))
7816 return false;
7817
7818 // Only emit a global base reg in PIC mode.
7819 if (!TM->isPositionIndependent())
7820 return false;
7821
7822 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
7823 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
7824
7825 // If we didn't need a GlobalBaseReg, don't insert code.
7826 if (GlobalBaseReg == 0)
7827 return false;
7828
7829 // Insert the set of GlobalBaseReg into the first MBB of the function
7830 MachineBasicBlock &FirstMBB = MF.front();
7831 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
7832 DebugLoc DL = FirstMBB.findDebugLoc(MBBI);
7833 MachineRegisterInfo &RegInfo = MF.getRegInfo();
7834 const X86InstrInfo *TII = STI.getInstrInfo();
7835
7836 unsigned PC;
7837 if (STI.isPICStyleGOT())
7838 PC = RegInfo.createVirtualRegister(&X86::GR32RegClass);
7839 else
7840 PC = GlobalBaseReg;
7841
7842 if (STI.is64Bit()) {
7843 if (TM->getCodeModel() == CodeModel::Medium) {
7844 // In the medium code model, use a RIP-relative LEA to materialize the
7845 // GOT.
7846 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::LEA64r), PC)
7847 .addReg(X86::RIP)
7848 .addImm(0)
7849 .addReg(0)
7850 .addExternalSymbol("_GLOBAL_OFFSET_TABLE_")
7851 .addReg(0);
7852 } else if (TM->getCodeModel() == CodeModel::Large) {
7853 // In the large code model, we are aiming for this code, though the
7854 // register allocation may vary:
7855 // leaq .LN$pb(%rip), %rax
7856 // movq $_GLOBAL_OFFSET_TABLE_ - .LN$pb, %rcx
7857 // addq %rcx, %rax
7858 // RAX now holds address of _GLOBAL_OFFSET_TABLE_.
7859 Register PBReg = RegInfo.createVirtualRegister(&X86::GR64RegClass);
7860 Register GOTReg = RegInfo.createVirtualRegister(&X86::GR64RegClass);
7861 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::LEA64r), PBReg)
7862 .addReg(X86::RIP)
7863 .addImm(0)
7864 .addReg(0)
7865 .addSym(MF.getPICBaseSymbol())
7866 .addReg(0);
7867 std::prev(MBBI)->setPreInstrSymbol(MF, MF.getPICBaseSymbol());
7868 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOV64ri), GOTReg)
7869 .addExternalSymbol("_GLOBAL_OFFSET_TABLE_",
7870 X86II::MO_PIC_BASE_OFFSET);
7871 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD64rr), PC)
7872 .addReg(PBReg, RegState::Kill)
7873 .addReg(GOTReg, RegState::Kill);
7874 } else {
7875 llvm_unreachable("unexpected code model");
7876 }
7877 } else {
7878 // Operand of MovePCtoStack is completely ignored by asm printer. It's
7879 // only used in JIT code emission as displacement to pc.
7880 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOVPC32r), PC).addImm(0);
7881
7882 // If we're using vanilla 'GOT' PIC style, we should use relative
7883 // addressing not to pc, but to _GLOBAL_OFFSET_TABLE_ external.
7884 if (STI.isPICStyleGOT()) {
7885 // Generate addl $__GLOBAL_OFFSET_TABLE_ + [.-piclabel],
7886 // %some_register
7887 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD32ri), GlobalBaseReg)
7888 .addReg(PC)
7889 .addExternalSymbol("_GLOBAL_OFFSET_TABLE_",
7890 X86II::MO_GOT_ABSOLUTE_ADDRESS);
7891 }
7892 }
7893
7894 return true;
7895 }
7896
getPassName__anone5c35a6e0511::CGBR7897 StringRef getPassName() const override {
7898 return "X86 PIC Global Base Reg Initialization";
7899 }
7900
getAnalysisUsage__anone5c35a6e0511::CGBR7901 void getAnalysisUsage(AnalysisUsage &AU) const override {
7902 AU.setPreservesCFG();
7903 MachineFunctionPass::getAnalysisUsage(AU);
7904 }
7905 };
7906 }
7907
7908 char CGBR::ID = 0;
7909 FunctionPass*
createX86GlobalBaseRegPass()7910 llvm::createX86GlobalBaseRegPass() { return new CGBR(); }
7911
7912 namespace {
7913 struct LDTLSCleanup : public MachineFunctionPass {
7914 static char ID;
LDTLSCleanup__anone5c35a6e0611::LDTLSCleanup7915 LDTLSCleanup() : MachineFunctionPass(ID) {}
7916
runOnMachineFunction__anone5c35a6e0611::LDTLSCleanup7917 bool runOnMachineFunction(MachineFunction &MF) override {
7918 if (skipFunction(MF.getFunction()))
7919 return false;
7920
7921 X86MachineFunctionInfo *MFI = MF.getInfo<X86MachineFunctionInfo>();
7922 if (MFI->getNumLocalDynamicTLSAccesses() < 2) {
7923 // No point folding accesses if there isn't at least two.
7924 return false;
7925 }
7926
7927 MachineDominatorTree *DT = &getAnalysis<MachineDominatorTree>();
7928 return VisitNode(DT->getRootNode(), 0);
7929 }
7930
7931 // Visit the dominator subtree rooted at Node in pre-order.
7932 // If TLSBaseAddrReg is non-null, then use that to replace any
7933 // TLS_base_addr instructions. Otherwise, create the register
7934 // when the first such instruction is seen, and then use it
7935 // as we encounter more instructions.
VisitNode__anone5c35a6e0611::LDTLSCleanup7936 bool VisitNode(MachineDomTreeNode *Node, unsigned TLSBaseAddrReg) {
7937 MachineBasicBlock *BB = Node->getBlock();
7938 bool Changed = false;
7939
7940 // Traverse the current block.
7941 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;
7942 ++I) {
7943 switch (I->getOpcode()) {
7944 case X86::TLS_base_addr32:
7945 case X86::TLS_base_addr64:
7946 if (TLSBaseAddrReg)
7947 I = ReplaceTLSBaseAddrCall(*I, TLSBaseAddrReg);
7948 else
7949 I = SetRegister(*I, &TLSBaseAddrReg);
7950 Changed = true;
7951 break;
7952 default:
7953 break;
7954 }
7955 }
7956
7957 // Visit the children of this block in the dominator tree.
7958 for (MachineDomTreeNode::iterator I = Node->begin(), E = Node->end();
7959 I != E; ++I) {
7960 Changed |= VisitNode(*I, TLSBaseAddrReg);
7961 }
7962
7963 return Changed;
7964 }
7965
7966 // Replace the TLS_base_addr instruction I with a copy from
7967 // TLSBaseAddrReg, returning the new instruction.
ReplaceTLSBaseAddrCall__anone5c35a6e0611::LDTLSCleanup7968 MachineInstr *ReplaceTLSBaseAddrCall(MachineInstr &I,
7969 unsigned TLSBaseAddrReg) {
7970 MachineFunction *MF = I.getParent()->getParent();
7971 const X86Subtarget &STI = MF->getSubtarget<X86Subtarget>();
7972 const bool is64Bit = STI.is64Bit();
7973 const X86InstrInfo *TII = STI.getInstrInfo();
7974
7975 // Insert a Copy from TLSBaseAddrReg to RAX/EAX.
7976 MachineInstr *Copy =
7977 BuildMI(*I.getParent(), I, I.getDebugLoc(),
7978 TII->get(TargetOpcode::COPY), is64Bit ? X86::RAX : X86::EAX)
7979 .addReg(TLSBaseAddrReg);
7980
7981 // Erase the TLS_base_addr instruction.
7982 I.eraseFromParent();
7983
7984 return Copy;
7985 }
7986
7987 // Create a virtual register in *TLSBaseAddrReg, and populate it by
7988 // inserting a copy instruction after I. Returns the new instruction.
SetRegister__anone5c35a6e0611::LDTLSCleanup7989 MachineInstr *SetRegister(MachineInstr &I, unsigned *TLSBaseAddrReg) {
7990 MachineFunction *MF = I.getParent()->getParent();
7991 const X86Subtarget &STI = MF->getSubtarget<X86Subtarget>();
7992 const bool is64Bit = STI.is64Bit();
7993 const X86InstrInfo *TII = STI.getInstrInfo();
7994
7995 // Create a virtual register for the TLS base address.
7996 MachineRegisterInfo &RegInfo = MF->getRegInfo();
7997 *TLSBaseAddrReg = RegInfo.createVirtualRegister(is64Bit
7998 ? &X86::GR64RegClass
7999 : &X86::GR32RegClass);
8000
8001 // Insert a copy from RAX/EAX to TLSBaseAddrReg.
8002 MachineInstr *Next = I.getNextNode();
8003 MachineInstr *Copy =
8004 BuildMI(*I.getParent(), Next, I.getDebugLoc(),
8005 TII->get(TargetOpcode::COPY), *TLSBaseAddrReg)
8006 .addReg(is64Bit ? X86::RAX : X86::EAX);
8007
8008 return Copy;
8009 }
8010
getPassName__anone5c35a6e0611::LDTLSCleanup8011 StringRef getPassName() const override {
8012 return "Local Dynamic TLS Access Clean-up";
8013 }
8014
getAnalysisUsage__anone5c35a6e0611::LDTLSCleanup8015 void getAnalysisUsage(AnalysisUsage &AU) const override {
8016 AU.setPreservesCFG();
8017 AU.addRequired<MachineDominatorTree>();
8018 MachineFunctionPass::getAnalysisUsage(AU);
8019 }
8020 };
8021 }
8022
8023 char LDTLSCleanup::ID = 0;
8024 FunctionPass*
createCleanupLocalDynamicTLSPass()8025 llvm::createCleanupLocalDynamicTLSPass() { return new LDTLSCleanup(); }
8026
8027 /// Constants defining how certain sequences should be outlined.
8028 ///
8029 /// \p MachineOutlinerDefault implies that the function is called with a call
8030 /// instruction, and a return must be emitted for the outlined function frame.
8031 ///
8032 /// That is,
8033 ///
8034 /// I1 OUTLINED_FUNCTION:
8035 /// I2 --> call OUTLINED_FUNCTION I1
8036 /// I3 I2
8037 /// I3
8038 /// ret
8039 ///
8040 /// * Call construction overhead: 1 (call instruction)
8041 /// * Frame construction overhead: 1 (return instruction)
8042 ///
8043 /// \p MachineOutlinerTailCall implies that the function is being tail called.
8044 /// A jump is emitted instead of a call, and the return is already present in
8045 /// the outlined sequence. That is,
8046 ///
8047 /// I1 OUTLINED_FUNCTION:
8048 /// I2 --> jmp OUTLINED_FUNCTION I1
8049 /// ret I2
8050 /// ret
8051 ///
8052 /// * Call construction overhead: 1 (jump instruction)
8053 /// * Frame construction overhead: 0 (don't need to return)
8054 ///
8055 enum MachineOutlinerClass {
8056 MachineOutlinerDefault,
8057 MachineOutlinerTailCall
8058 };
8059
getOutliningCandidateInfo(std::vector<outliner::Candidate> & RepeatedSequenceLocs) const8060 outliner::OutlinedFunction X86InstrInfo::getOutliningCandidateInfo(
8061 std::vector<outliner::Candidate> &RepeatedSequenceLocs) const {
8062 unsigned SequenceSize =
8063 std::accumulate(RepeatedSequenceLocs[0].front(),
8064 std::next(RepeatedSequenceLocs[0].back()), 0,
8065 [](unsigned Sum, const MachineInstr &MI) {
8066 // FIXME: x86 doesn't implement getInstSizeInBytes, so
8067 // we can't tell the cost. Just assume each instruction
8068 // is one byte.
8069 if (MI.isDebugInstr() || MI.isKill())
8070 return Sum;
8071 return Sum + 1;
8072 });
8073
8074 // FIXME: Use real size in bytes for call and ret instructions.
8075 if (RepeatedSequenceLocs[0].back()->isTerminator()) {
8076 for (outliner::Candidate &C : RepeatedSequenceLocs)
8077 C.setCallInfo(MachineOutlinerTailCall, 1);
8078
8079 return outliner::OutlinedFunction(RepeatedSequenceLocs, SequenceSize,
8080 0, // Number of bytes to emit frame.
8081 MachineOutlinerTailCall // Type of frame.
8082 );
8083 }
8084
8085 for (outliner::Candidate &C : RepeatedSequenceLocs)
8086 C.setCallInfo(MachineOutlinerDefault, 1);
8087
8088 return outliner::OutlinedFunction(RepeatedSequenceLocs, SequenceSize, 1,
8089 MachineOutlinerDefault);
8090 }
8091
isFunctionSafeToOutlineFrom(MachineFunction & MF,bool OutlineFromLinkOnceODRs) const8092 bool X86InstrInfo::isFunctionSafeToOutlineFrom(MachineFunction &MF,
8093 bool OutlineFromLinkOnceODRs) const {
8094 const Function &F = MF.getFunction();
8095
8096 // Does the function use a red zone? If it does, then we can't risk messing
8097 // with the stack.
8098 if (Subtarget.getFrameLowering()->has128ByteRedZone(MF)) {
8099 // It could have a red zone. If it does, then we don't want to touch it.
8100 const X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
8101 if (!X86FI || X86FI->getUsesRedZone())
8102 return false;
8103 }
8104
8105 // If we *don't* want to outline from things that could potentially be deduped
8106 // then return false.
8107 if (!OutlineFromLinkOnceODRs && F.hasLinkOnceODRLinkage())
8108 return false;
8109
8110 // This function is viable for outlining, so return true.
8111 return true;
8112 }
8113
8114 outliner::InstrType
getOutliningType(MachineBasicBlock::iterator & MIT,unsigned Flags) const8115 X86InstrInfo::getOutliningType(MachineBasicBlock::iterator &MIT, unsigned Flags) const {
8116 MachineInstr &MI = *MIT;
8117 // Don't allow debug values to impact outlining type.
8118 if (MI.isDebugInstr() || MI.isIndirectDebugValue())
8119 return outliner::InstrType::Invisible;
8120
8121 // At this point, KILL instructions don't really tell us much so we can go
8122 // ahead and skip over them.
8123 if (MI.isKill())
8124 return outliner::InstrType::Invisible;
8125
8126 // Is this a tail call? If yes, we can outline as a tail call.
8127 if (isTailCall(MI))
8128 return outliner::InstrType::Legal;
8129
8130 // Is this the terminator of a basic block?
8131 if (MI.isTerminator() || MI.isReturn()) {
8132
8133 // Does its parent have any successors in its MachineFunction?
8134 if (MI.getParent()->succ_empty())
8135 return outliner::InstrType::Legal;
8136
8137 // It does, so we can't tail call it.
8138 return outliner::InstrType::Illegal;
8139 }
8140
8141 // Don't outline anything that modifies or reads from the stack pointer.
8142 //
8143 // FIXME: There are instructions which are being manually built without
8144 // explicit uses/defs so we also have to check the MCInstrDesc. We should be
8145 // able to remove the extra checks once those are fixed up. For example,
8146 // sometimes we might get something like %rax = POP64r 1. This won't be
8147 // caught by modifiesRegister or readsRegister even though the instruction
8148 // really ought to be formed so that modifiesRegister/readsRegister would
8149 // catch it.
8150 if (MI.modifiesRegister(X86::RSP, &RI) || MI.readsRegister(X86::RSP, &RI) ||
8151 MI.getDesc().hasImplicitUseOfPhysReg(X86::RSP) ||
8152 MI.getDesc().hasImplicitDefOfPhysReg(X86::RSP))
8153 return outliner::InstrType::Illegal;
8154
8155 // Outlined calls change the instruction pointer, so don't read from it.
8156 if (MI.readsRegister(X86::RIP, &RI) ||
8157 MI.getDesc().hasImplicitUseOfPhysReg(X86::RIP) ||
8158 MI.getDesc().hasImplicitDefOfPhysReg(X86::RIP))
8159 return outliner::InstrType::Illegal;
8160
8161 // Positions can't safely be outlined.
8162 if (MI.isPosition())
8163 return outliner::InstrType::Illegal;
8164
8165 // Make sure none of the operands of this instruction do anything tricky.
8166 for (const MachineOperand &MOP : MI.operands())
8167 if (MOP.isCPI() || MOP.isJTI() || MOP.isCFIIndex() || MOP.isFI() ||
8168 MOP.isTargetIndex())
8169 return outliner::InstrType::Illegal;
8170
8171 return outliner::InstrType::Legal;
8172 }
8173
buildOutlinedFrame(MachineBasicBlock & MBB,MachineFunction & MF,const outliner::OutlinedFunction & OF) const8174 void X86InstrInfo::buildOutlinedFrame(MachineBasicBlock &MBB,
8175 MachineFunction &MF,
8176 const outliner::OutlinedFunction &OF)
8177 const {
8178 // If we're a tail call, we already have a return, so don't do anything.
8179 if (OF.FrameConstructionID == MachineOutlinerTailCall)
8180 return;
8181
8182 // We're a normal call, so our sequence doesn't have a return instruction.
8183 // Add it in.
8184 MachineInstr *retq = BuildMI(MF, DebugLoc(), get(X86::RETQ));
8185 MBB.insert(MBB.end(), retq);
8186 }
8187
8188 MachineBasicBlock::iterator
insertOutlinedCall(Module & M,MachineBasicBlock & MBB,MachineBasicBlock::iterator & It,MachineFunction & MF,const outliner::Candidate & C) const8189 X86InstrInfo::insertOutlinedCall(Module &M, MachineBasicBlock &MBB,
8190 MachineBasicBlock::iterator &It,
8191 MachineFunction &MF,
8192 const outliner::Candidate &C) const {
8193 // Is it a tail call?
8194 if (C.CallConstructionID == MachineOutlinerTailCall) {
8195 // Yes, just insert a JMP.
8196 It = MBB.insert(It,
8197 BuildMI(MF, DebugLoc(), get(X86::TAILJMPd64))
8198 .addGlobalAddress(M.getNamedValue(MF.getName())));
8199 } else {
8200 // No, insert a call.
8201 It = MBB.insert(It,
8202 BuildMI(MF, DebugLoc(), get(X86::CALL64pcrel32))
8203 .addGlobalAddress(M.getNamedValue(MF.getName())));
8204 }
8205
8206 return It;
8207 }
8208
8209 #define GET_INSTRINFO_HELPERS
8210 #include "X86GenInstrInfo.inc"
8211