xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AMDGPU/SIOptimizeExecMasking.cpp (revision 5ca8e32633c4ffbbcd6762e5888b6a4ba0708c6c)
1 //===-- SIOptimizeExecMasking.cpp -----------------------------------------===//
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 #include "AMDGPU.h"
10 #include "GCNSubtarget.h"
11 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
12 #include "SIRegisterInfo.h"
13 #include "llvm/CodeGen/LivePhysRegs.h"
14 #include "llvm/CodeGen/MachineFunctionPass.h"
15 #include "llvm/CodeGen/MachineOperand.h"
16 #include "llvm/CodeGen/TargetRegisterInfo.h"
17 #include "llvm/InitializePasses.h"
18 
19 using namespace llvm;
20 
21 #define DEBUG_TYPE "si-optimize-exec-masking"
22 
23 namespace {
24 
25 class SIOptimizeExecMasking : public MachineFunctionPass {
26   MachineFunction *MF = nullptr;
27   const GCNSubtarget *ST = nullptr;
28   const SIRegisterInfo *TRI = nullptr;
29   const SIInstrInfo *TII = nullptr;
30   const MachineRegisterInfo *MRI = nullptr;
31   MCRegister Exec;
32 
33   DenseMap<MachineInstr *, MachineInstr *> SaveExecVCmpMapping;
34   SmallVector<std::pair<MachineInstr *, MachineInstr *>, 1> OrXors;
35 
36   Register isCopyFromExec(const MachineInstr &MI) const;
37   Register isCopyToExec(const MachineInstr &MI) const;
38   bool removeTerminatorBit(MachineInstr &MI) const;
39   MachineBasicBlock::reverse_iterator
40   fixTerminators(MachineBasicBlock &MBB) const;
41   MachineBasicBlock::reverse_iterator
42   findExecCopy(MachineBasicBlock &MBB,
43                MachineBasicBlock::reverse_iterator I) const;
44 
45   bool isRegisterInUseBetween(MachineInstr &Stop, MachineInstr &Start,
46                               MCRegister Reg, bool UseLiveOuts = false,
47                               bool IgnoreStart = false) const;
48   bool isRegisterInUseAfter(MachineInstr &Stop, MCRegister Reg) const;
49   MachineInstr *findInstrBackwards(MachineInstr &Origin,
50                                    std::function<bool(MachineInstr *)> Pred,
51                                    ArrayRef<MCRegister> NonModifiableRegs,
52                                    unsigned MaxInstructions = 20) const;
53   bool optimizeExecSequence();
54   void tryRecordVCmpxAndSaveexecSequence(MachineInstr &MI);
55   bool optimizeVCMPSaveExecSequence(MachineInstr &SaveExecInstr,
56                                     MachineInstr &VCmp, MCRegister Exec) const;
57 
58   void tryRecordOrSaveexecXorSequence(MachineInstr &MI);
59   bool optimizeOrSaveexecXorSequences();
60 
61 public:
62   static char ID;
63 
64   SIOptimizeExecMasking() : MachineFunctionPass(ID) {
65     initializeSIOptimizeExecMaskingPass(*PassRegistry::getPassRegistry());
66   }
67 
68   bool runOnMachineFunction(MachineFunction &MF) override;
69 
70   StringRef getPassName() const override {
71     return "SI optimize exec mask operations";
72   }
73 
74   void getAnalysisUsage(AnalysisUsage &AU) const override {
75     AU.setPreservesCFG();
76     MachineFunctionPass::getAnalysisUsage(AU);
77   }
78 };
79 
80 } // End anonymous namespace.
81 
82 INITIALIZE_PASS_BEGIN(SIOptimizeExecMasking, DEBUG_TYPE,
83                       "SI optimize exec mask operations", false, false)
84 INITIALIZE_PASS_DEPENDENCY(LiveIntervals)
85 INITIALIZE_PASS_END(SIOptimizeExecMasking, DEBUG_TYPE,
86                     "SI optimize exec mask operations", false, false)
87 
88 char SIOptimizeExecMasking::ID = 0;
89 
90 char &llvm::SIOptimizeExecMaskingID = SIOptimizeExecMasking::ID;
91 
92 /// If \p MI is a copy from exec, return the register copied to.
93 Register SIOptimizeExecMasking::isCopyFromExec(const MachineInstr &MI) const {
94   switch (MI.getOpcode()) {
95   case AMDGPU::COPY:
96   case AMDGPU::S_MOV_B64:
97   case AMDGPU::S_MOV_B64_term:
98   case AMDGPU::S_MOV_B32:
99   case AMDGPU::S_MOV_B32_term: {
100     const MachineOperand &Src = MI.getOperand(1);
101     if (Src.isReg() && Src.getReg() == Exec)
102       return MI.getOperand(0).getReg();
103   }
104   }
105 
106   return AMDGPU::NoRegister;
107 }
108 
109 /// If \p MI is a copy to exec, return the register copied from.
110 Register SIOptimizeExecMasking::isCopyToExec(const MachineInstr &MI) const {
111   switch (MI.getOpcode()) {
112   case AMDGPU::COPY:
113   case AMDGPU::S_MOV_B64:
114   case AMDGPU::S_MOV_B32: {
115     const MachineOperand &Dst = MI.getOperand(0);
116     if (Dst.isReg() && Dst.getReg() == Exec && MI.getOperand(1).isReg())
117       return MI.getOperand(1).getReg();
118     break;
119   }
120   case AMDGPU::S_MOV_B64_term:
121   case AMDGPU::S_MOV_B32_term:
122     llvm_unreachable("should have been replaced");
123   }
124 
125   return Register();
126 }
127 
128 /// If \p MI is a logical operation on an exec value,
129 /// return the register copied to.
130 static Register isLogicalOpOnExec(const MachineInstr &MI) {
131   switch (MI.getOpcode()) {
132   case AMDGPU::S_AND_B64:
133   case AMDGPU::S_OR_B64:
134   case AMDGPU::S_XOR_B64:
135   case AMDGPU::S_ANDN2_B64:
136   case AMDGPU::S_ORN2_B64:
137   case AMDGPU::S_NAND_B64:
138   case AMDGPU::S_NOR_B64:
139   case AMDGPU::S_XNOR_B64: {
140     const MachineOperand &Src1 = MI.getOperand(1);
141     if (Src1.isReg() && Src1.getReg() == AMDGPU::EXEC)
142       return MI.getOperand(0).getReg();
143     const MachineOperand &Src2 = MI.getOperand(2);
144     if (Src2.isReg() && Src2.getReg() == AMDGPU::EXEC)
145       return MI.getOperand(0).getReg();
146     break;
147   }
148   case AMDGPU::S_AND_B32:
149   case AMDGPU::S_OR_B32:
150   case AMDGPU::S_XOR_B32:
151   case AMDGPU::S_ANDN2_B32:
152   case AMDGPU::S_ORN2_B32:
153   case AMDGPU::S_NAND_B32:
154   case AMDGPU::S_NOR_B32:
155   case AMDGPU::S_XNOR_B32: {
156     const MachineOperand &Src1 = MI.getOperand(1);
157     if (Src1.isReg() && Src1.getReg() == AMDGPU::EXEC_LO)
158       return MI.getOperand(0).getReg();
159     const MachineOperand &Src2 = MI.getOperand(2);
160     if (Src2.isReg() && Src2.getReg() == AMDGPU::EXEC_LO)
161       return MI.getOperand(0).getReg();
162     break;
163   }
164   }
165 
166   return AMDGPU::NoRegister;
167 }
168 
169 static unsigned getSaveExecOp(unsigned Opc) {
170   switch (Opc) {
171   case AMDGPU::S_AND_B64:
172     return AMDGPU::S_AND_SAVEEXEC_B64;
173   case AMDGPU::S_OR_B64:
174     return AMDGPU::S_OR_SAVEEXEC_B64;
175   case AMDGPU::S_XOR_B64:
176     return AMDGPU::S_XOR_SAVEEXEC_B64;
177   case AMDGPU::S_ANDN2_B64:
178     return AMDGPU::S_ANDN2_SAVEEXEC_B64;
179   case AMDGPU::S_ORN2_B64:
180     return AMDGPU::S_ORN2_SAVEEXEC_B64;
181   case AMDGPU::S_NAND_B64:
182     return AMDGPU::S_NAND_SAVEEXEC_B64;
183   case AMDGPU::S_NOR_B64:
184     return AMDGPU::S_NOR_SAVEEXEC_B64;
185   case AMDGPU::S_XNOR_B64:
186     return AMDGPU::S_XNOR_SAVEEXEC_B64;
187   case AMDGPU::S_AND_B32:
188     return AMDGPU::S_AND_SAVEEXEC_B32;
189   case AMDGPU::S_OR_B32:
190     return AMDGPU::S_OR_SAVEEXEC_B32;
191   case AMDGPU::S_XOR_B32:
192     return AMDGPU::S_XOR_SAVEEXEC_B32;
193   case AMDGPU::S_ANDN2_B32:
194     return AMDGPU::S_ANDN2_SAVEEXEC_B32;
195   case AMDGPU::S_ORN2_B32:
196     return AMDGPU::S_ORN2_SAVEEXEC_B32;
197   case AMDGPU::S_NAND_B32:
198     return AMDGPU::S_NAND_SAVEEXEC_B32;
199   case AMDGPU::S_NOR_B32:
200     return AMDGPU::S_NOR_SAVEEXEC_B32;
201   case AMDGPU::S_XNOR_B32:
202     return AMDGPU::S_XNOR_SAVEEXEC_B32;
203   default:
204     return AMDGPU::INSTRUCTION_LIST_END;
205   }
206 }
207 
208 // These are only terminators to get correct spill code placement during
209 // register allocation, so turn them back into normal instructions.
210 bool SIOptimizeExecMasking::removeTerminatorBit(MachineInstr &MI) const {
211   switch (MI.getOpcode()) {
212   case AMDGPU::S_MOV_B32_term: {
213     bool RegSrc = MI.getOperand(1).isReg();
214     MI.setDesc(TII->get(RegSrc ? AMDGPU::COPY : AMDGPU::S_MOV_B32));
215     return true;
216   }
217   case AMDGPU::S_MOV_B64_term: {
218     bool RegSrc = MI.getOperand(1).isReg();
219     MI.setDesc(TII->get(RegSrc ? AMDGPU::COPY : AMDGPU::S_MOV_B64));
220     return true;
221   }
222   case AMDGPU::S_XOR_B64_term: {
223     // This is only a terminator to get the correct spill code placement during
224     // register allocation.
225     MI.setDesc(TII->get(AMDGPU::S_XOR_B64));
226     return true;
227   }
228   case AMDGPU::S_XOR_B32_term: {
229     // This is only a terminator to get the correct spill code placement during
230     // register allocation.
231     MI.setDesc(TII->get(AMDGPU::S_XOR_B32));
232     return true;
233   }
234   case AMDGPU::S_OR_B64_term: {
235     // This is only a terminator to get the correct spill code placement during
236     // register allocation.
237     MI.setDesc(TII->get(AMDGPU::S_OR_B64));
238     return true;
239   }
240   case AMDGPU::S_OR_B32_term: {
241     // This is only a terminator to get the correct spill code placement during
242     // register allocation.
243     MI.setDesc(TII->get(AMDGPU::S_OR_B32));
244     return true;
245   }
246   case AMDGPU::S_ANDN2_B64_term: {
247     // This is only a terminator to get the correct spill code placement during
248     // register allocation.
249     MI.setDesc(TII->get(AMDGPU::S_ANDN2_B64));
250     return true;
251   }
252   case AMDGPU::S_ANDN2_B32_term: {
253     // This is only a terminator to get the correct spill code placement during
254     // register allocation.
255     MI.setDesc(TII->get(AMDGPU::S_ANDN2_B32));
256     return true;
257   }
258   case AMDGPU::S_AND_B64_term: {
259     // This is only a terminator to get the correct spill code placement during
260     // register allocation.
261     MI.setDesc(TII->get(AMDGPU::S_AND_B64));
262     return true;
263   }
264   case AMDGPU::S_AND_B32_term: {
265     // This is only a terminator to get the correct spill code placement during
266     // register allocation.
267     MI.setDesc(TII->get(AMDGPU::S_AND_B32));
268     return true;
269   }
270   default:
271     return false;
272   }
273 }
274 
275 // Turn all pseudoterminators in the block into their equivalent non-terminator
276 // instructions. Returns the reverse iterator to the first non-terminator
277 // instruction in the block.
278 MachineBasicBlock::reverse_iterator
279 SIOptimizeExecMasking::fixTerminators(MachineBasicBlock &MBB) const {
280   MachineBasicBlock::reverse_iterator I = MBB.rbegin(), E = MBB.rend();
281 
282   bool Seen = false;
283   MachineBasicBlock::reverse_iterator FirstNonTerm = I;
284   for (; I != E; ++I) {
285     if (!I->isTerminator())
286       return Seen ? FirstNonTerm : I;
287 
288     if (removeTerminatorBit(*I)) {
289       if (!Seen) {
290         FirstNonTerm = I;
291         Seen = true;
292       }
293     }
294   }
295 
296   return FirstNonTerm;
297 }
298 
299 MachineBasicBlock::reverse_iterator SIOptimizeExecMasking::findExecCopy(
300     MachineBasicBlock &MBB, MachineBasicBlock::reverse_iterator I) const {
301   const unsigned InstLimit = 25;
302 
303   auto E = MBB.rend();
304   for (unsigned N = 0; N <= InstLimit && I != E; ++I, ++N) {
305     Register CopyFromExec = isCopyFromExec(*I);
306     if (CopyFromExec.isValid())
307       return I;
308   }
309 
310   return E;
311 }
312 
313 // XXX - Seems LivePhysRegs doesn't work correctly since it will incorrectly
314 // report the register as unavailable because a super-register with a lane mask
315 // is unavailable.
316 static bool isLiveOut(const MachineBasicBlock &MBB, unsigned Reg) {
317   for (MachineBasicBlock *Succ : MBB.successors()) {
318     if (Succ->isLiveIn(Reg))
319       return true;
320   }
321 
322   return false;
323 }
324 
325 // Backwards-iterate from Origin (for n=MaxInstructions iterations) until either
326 // the beginning of the BB is reached or Pred evaluates to true - which can be
327 // an arbitrary condition based on the current MachineInstr, for instance an
328 // target instruction. Breaks prematurely by returning nullptr if  one of the
329 // registers given in NonModifiableRegs is modified by the current instruction.
330 MachineInstr *SIOptimizeExecMasking::findInstrBackwards(
331     MachineInstr &Origin, std::function<bool(MachineInstr *)> Pred,
332     ArrayRef<MCRegister> NonModifiableRegs, unsigned MaxInstructions) const {
333   MachineBasicBlock::reverse_iterator A = Origin.getReverseIterator(),
334                                       E = Origin.getParent()->rend();
335   unsigned CurrentIteration = 0;
336 
337   for (++A; CurrentIteration < MaxInstructions && A != E; ++A) {
338     if (A->isDebugInstr())
339       continue;
340 
341     if (Pred(&*A))
342       return &*A;
343 
344     for (MCRegister Reg : NonModifiableRegs) {
345       if (A->modifiesRegister(Reg, TRI))
346         return nullptr;
347     }
348 
349     ++CurrentIteration;
350   }
351 
352   return nullptr;
353 }
354 
355 // Determine if a register Reg is not re-defined and still in use
356 // in the range (Stop..Start].
357 // It does so by backwards calculating liveness from the end of the BB until
358 // either Stop or the beginning of the BB is reached.
359 // After liveness is calculated, we can determine if Reg is still in use and not
360 // defined inbetween the instructions.
361 bool SIOptimizeExecMasking::isRegisterInUseBetween(MachineInstr &Stop,
362                                                    MachineInstr &Start,
363                                                    MCRegister Reg,
364                                                    bool UseLiveOuts,
365                                                    bool IgnoreStart) const {
366   LivePhysRegs LR(*TRI);
367   if (UseLiveOuts)
368     LR.addLiveOuts(*Stop.getParent());
369 
370   MachineBasicBlock::reverse_iterator A(Start);
371 
372   if (IgnoreStart)
373     ++A;
374 
375   for (; A != Stop.getParent()->rend() && A != Stop; ++A) {
376     LR.stepBackward(*A);
377   }
378 
379   return !LR.available(*MRI, Reg);
380 }
381 
382 // Determine if a register Reg is not re-defined and still in use
383 // in the range (Stop..BB.end].
384 bool SIOptimizeExecMasking::isRegisterInUseAfter(MachineInstr &Stop,
385                                                  MCRegister Reg) const {
386   return isRegisterInUseBetween(Stop, *Stop.getParent()->rbegin(), Reg, true);
387 }
388 
389 // Optimize sequences emitted for control flow lowering. They are originally
390 // emitted as the separate operations because spill code may need to be
391 // inserted for the saved copy of exec.
392 //
393 //     x = copy exec
394 //     z = s_<op>_b64 x, y
395 //     exec = copy z
396 // =>
397 //     x = s_<op>_saveexec_b64 y
398 //
399 bool SIOptimizeExecMasking::optimizeExecSequence() {
400   bool Changed = false;
401   for (MachineBasicBlock &MBB : *MF) {
402     MachineBasicBlock::reverse_iterator I = fixTerminators(MBB);
403     MachineBasicBlock::reverse_iterator E = MBB.rend();
404     if (I == E)
405       continue;
406 
407     // It's possible to see other terminator copies after the exec copy. This
408     // can happen if control flow pseudos had their outputs used by phis.
409     Register CopyToExec;
410 
411     unsigned SearchCount = 0;
412     const unsigned SearchLimit = 5;
413     while (I != E && SearchCount++ < SearchLimit) {
414       CopyToExec = isCopyToExec(*I);
415       if (CopyToExec)
416         break;
417       ++I;
418     }
419 
420     if (!CopyToExec)
421       continue;
422 
423     // Scan backwards to find the def.
424     auto *CopyToExecInst = &*I;
425     auto CopyFromExecInst = findExecCopy(MBB, I);
426     if (CopyFromExecInst == E) {
427       auto PrepareExecInst = std::next(I);
428       if (PrepareExecInst == E)
429         continue;
430       // Fold exec = COPY (S_AND_B64 reg, exec) -> exec = S_AND_B64 reg, exec
431       if (CopyToExecInst->getOperand(1).isKill() &&
432           isLogicalOpOnExec(*PrepareExecInst) == CopyToExec) {
433         LLVM_DEBUG(dbgs() << "Fold exec copy: " << *PrepareExecInst);
434 
435         PrepareExecInst->getOperand(0).setReg(Exec);
436 
437         LLVM_DEBUG(dbgs() << "into: " << *PrepareExecInst << '\n');
438 
439         CopyToExecInst->eraseFromParent();
440         Changed = true;
441       }
442 
443       continue;
444     }
445 
446     if (isLiveOut(MBB, CopyToExec)) {
447       // The copied register is live out and has a second use in another block.
448       LLVM_DEBUG(dbgs() << "Exec copy source register is live out\n");
449       continue;
450     }
451 
452     Register CopyFromExec = CopyFromExecInst->getOperand(0).getReg();
453     MachineInstr *SaveExecInst = nullptr;
454     SmallVector<MachineInstr *, 4> OtherUseInsts;
455 
456     for (MachineBasicBlock::iterator
457              J = std::next(CopyFromExecInst->getIterator()),
458              JE = I->getIterator();
459          J != JE; ++J) {
460       if (SaveExecInst && J->readsRegister(Exec, TRI)) {
461         LLVM_DEBUG(dbgs() << "exec read prevents saveexec: " << *J << '\n');
462         // Make sure this is inserted after any VALU ops that may have been
463         // scheduled in between.
464         SaveExecInst = nullptr;
465         break;
466       }
467 
468       bool ReadsCopyFromExec = J->readsRegister(CopyFromExec, TRI);
469 
470       if (J->modifiesRegister(CopyToExec, TRI)) {
471         if (SaveExecInst) {
472           LLVM_DEBUG(dbgs() << "Multiple instructions modify "
473                             << printReg(CopyToExec, TRI) << '\n');
474           SaveExecInst = nullptr;
475           break;
476         }
477 
478         unsigned SaveExecOp = getSaveExecOp(J->getOpcode());
479         if (SaveExecOp == AMDGPU::INSTRUCTION_LIST_END)
480           break;
481 
482         if (ReadsCopyFromExec) {
483           SaveExecInst = &*J;
484           LLVM_DEBUG(dbgs() << "Found save exec op: " << *SaveExecInst << '\n');
485           continue;
486         } else {
487           LLVM_DEBUG(dbgs()
488                      << "Instruction does not read exec copy: " << *J << '\n');
489           break;
490         }
491       } else if (ReadsCopyFromExec && !SaveExecInst) {
492         // Make sure no other instruction is trying to use this copy, before it
493         // will be rewritten by the saveexec, i.e. hasOneUse. There may have
494         // been another use, such as an inserted spill. For example:
495         //
496         // %sgpr0_sgpr1 = COPY %exec
497         // spill %sgpr0_sgpr1
498         // %sgpr2_sgpr3 = S_AND_B64 %sgpr0_sgpr1
499         //
500         LLVM_DEBUG(dbgs() << "Found second use of save inst candidate: " << *J
501                           << '\n');
502         break;
503       }
504 
505       if (SaveExecInst && J->readsRegister(CopyToExec, TRI)) {
506         assert(SaveExecInst != &*J);
507         OtherUseInsts.push_back(&*J);
508       }
509     }
510 
511     if (!SaveExecInst)
512       continue;
513 
514     LLVM_DEBUG(dbgs() << "Insert save exec op: " << *SaveExecInst << '\n');
515 
516     MachineOperand &Src0 = SaveExecInst->getOperand(1);
517     MachineOperand &Src1 = SaveExecInst->getOperand(2);
518 
519     MachineOperand *OtherOp = nullptr;
520 
521     if (Src0.isReg() && Src0.getReg() == CopyFromExec) {
522       OtherOp = &Src1;
523     } else if (Src1.isReg() && Src1.getReg() == CopyFromExec) {
524       if (!SaveExecInst->isCommutable())
525         break;
526 
527       OtherOp = &Src0;
528     } else
529       llvm_unreachable("unexpected");
530 
531     CopyFromExecInst->eraseFromParent();
532 
533     auto InsPt = SaveExecInst->getIterator();
534     const DebugLoc &DL = SaveExecInst->getDebugLoc();
535 
536     BuildMI(MBB, InsPt, DL, TII->get(getSaveExecOp(SaveExecInst->getOpcode())),
537             CopyFromExec)
538         .addReg(OtherOp->getReg());
539     SaveExecInst->eraseFromParent();
540 
541     CopyToExecInst->eraseFromParent();
542 
543     for (MachineInstr *OtherInst : OtherUseInsts) {
544       OtherInst->substituteRegister(CopyToExec, Exec, AMDGPU::NoSubRegister,
545                                     *TRI);
546     }
547 
548     Changed = true;
549   }
550 
551   return Changed;
552 }
553 
554 // Inserts the optimized s_mov_b32 / v_cmpx sequence based on the
555 // operands extracted from a v_cmp ..., s_and_saveexec pattern.
556 bool SIOptimizeExecMasking::optimizeVCMPSaveExecSequence(
557     MachineInstr &SaveExecInstr, MachineInstr &VCmp, MCRegister Exec) const {
558   const int NewOpcode = AMDGPU::getVCMPXOpFromVCMP(VCmp.getOpcode());
559 
560   if (NewOpcode == -1)
561     return false;
562 
563   MachineOperand *Src0 = TII->getNamedOperand(VCmp, AMDGPU::OpName::src0);
564   MachineOperand *Src1 = TII->getNamedOperand(VCmp, AMDGPU::OpName::src1);
565 
566   Register MoveDest = SaveExecInstr.getOperand(0).getReg();
567 
568   MachineBasicBlock::instr_iterator InsertPosIt = SaveExecInstr.getIterator();
569   if (!SaveExecInstr.uses().empty()) {
570     bool IsSGPR32 = TRI->getRegSizeInBits(MoveDest, *MRI) == 32;
571     unsigned MovOpcode = IsSGPR32 ? AMDGPU::S_MOV_B32 : AMDGPU::S_MOV_B64;
572     BuildMI(*SaveExecInstr.getParent(), InsertPosIt,
573             SaveExecInstr.getDebugLoc(), TII->get(MovOpcode), MoveDest)
574         .addReg(Exec);
575   }
576 
577   // Omit dst as V_CMPX is implicitly writing to EXEC.
578   // Add dummy src and clamp modifiers, if needed.
579   auto Builder = BuildMI(*VCmp.getParent(), std::next(InsertPosIt),
580                          VCmp.getDebugLoc(), TII->get(NewOpcode));
581 
582   auto TryAddImmediateValueFromNamedOperand =
583       [&](unsigned OperandName) -> void {
584     if (auto *Mod = TII->getNamedOperand(VCmp, OperandName))
585       Builder.addImm(Mod->getImm());
586   };
587 
588   TryAddImmediateValueFromNamedOperand(AMDGPU::OpName::src0_modifiers);
589   Builder.add(*Src0);
590 
591   TryAddImmediateValueFromNamedOperand(AMDGPU::OpName::src1_modifiers);
592   Builder.add(*Src1);
593 
594   TryAddImmediateValueFromNamedOperand(AMDGPU::OpName::clamp);
595 
596   // The kill flags may no longer be correct.
597   if (Src0->isReg())
598     MRI->clearKillFlags(Src0->getReg());
599   if (Src1->isReg())
600     MRI->clearKillFlags(Src1->getReg());
601 
602   SaveExecInstr.eraseFromParent();
603   VCmp.eraseFromParent();
604 
605   return true;
606 }
607 
608 // Record (on GFX10.3 and later) occurences of
609 // v_cmp_* SGPR, IMM, VGPR
610 // s_and_saveexec_b32 EXEC_SGPR_DEST, SGPR
611 // to be replaced with
612 // s_mov_b32 EXEC_SGPR_DEST, exec_lo
613 // v_cmpx_* IMM, VGPR
614 // to reduce pipeline stalls.
615 void SIOptimizeExecMasking::tryRecordVCmpxAndSaveexecSequence(
616     MachineInstr &MI) {
617   if (!ST->hasGFX10_3Insts())
618     return;
619 
620   const unsigned AndSaveExecOpcode =
621       ST->isWave32() ? AMDGPU::S_AND_SAVEEXEC_B32 : AMDGPU::S_AND_SAVEEXEC_B64;
622 
623   if (MI.getOpcode() != AndSaveExecOpcode)
624     return;
625 
626   Register SaveExecDest = MI.getOperand(0).getReg();
627   if (!TRI->isSGPRReg(*MRI, SaveExecDest))
628     return;
629 
630   MachineOperand *SaveExecSrc0 = TII->getNamedOperand(MI, AMDGPU::OpName::src0);
631   if (!SaveExecSrc0->isReg())
632     return;
633 
634   // Tries to find a possibility to optimize a v_cmp ..., s_and_saveexec
635   // sequence by looking at an instance of an s_and_saveexec instruction.
636   // Returns a pointer to the v_cmp instruction if it is safe to replace the
637   // sequence (see the conditions in the function body). This is after register
638   // allocation, so some checks on operand dependencies need to be considered.
639   MachineInstr *VCmp = nullptr;
640 
641   // Try to find the last v_cmp instruction that defs the saveexec input
642   // operand without any write to Exec or the saveexec input operand inbetween.
643   VCmp = findInstrBackwards(
644       MI,
645       [&](MachineInstr *Check) {
646         return AMDGPU::getVCMPXOpFromVCMP(Check->getOpcode()) != -1 &&
647                Check->modifiesRegister(SaveExecSrc0->getReg(), TRI);
648       },
649       {Exec, SaveExecSrc0->getReg()});
650 
651   if (!VCmp)
652     return;
653 
654   MachineOperand *VCmpDest = TII->getNamedOperand(*VCmp, AMDGPU::OpName::sdst);
655   assert(VCmpDest && "Should have an sdst operand!");
656 
657   // Check if any of the v_cmp source operands is written by the saveexec.
658   MachineOperand *Src0 = TII->getNamedOperand(*VCmp, AMDGPU::OpName::src0);
659   if (Src0->isReg() && TRI->isSGPRReg(*MRI, Src0->getReg()) &&
660       MI.modifiesRegister(Src0->getReg(), TRI))
661     return;
662 
663   MachineOperand *Src1 = TII->getNamedOperand(*VCmp, AMDGPU::OpName::src1);
664   if (Src1->isReg() && TRI->isSGPRReg(*MRI, Src1->getReg()) &&
665       MI.modifiesRegister(Src1->getReg(), TRI))
666     return;
667 
668   // Don't do the transformation if the destination operand is included in
669   // it's MBB Live-outs, meaning it's used in any of its successors, leading
670   // to incorrect code if the v_cmp and therefore the def of
671   // the dest operand is removed.
672   if (isLiveOut(*VCmp->getParent(), VCmpDest->getReg()))
673     return;
674 
675   // If the v_cmp target is in use between v_cmp and s_and_saveexec or after the
676   // s_and_saveexec, skip the optimization.
677   if (isRegisterInUseBetween(*VCmp, MI, VCmpDest->getReg(), false, true) ||
678       isRegisterInUseAfter(MI, VCmpDest->getReg()))
679     return;
680 
681   // Try to determine if there is a write to any of the VCmp
682   // operands between the saveexec and the vcmp.
683   // If yes, additional VGPR spilling might need to be inserted. In this case,
684   // it's not worth replacing the instruction sequence.
685   SmallVector<MCRegister, 2> NonDefRegs;
686   if (Src0->isReg())
687     NonDefRegs.push_back(Src0->getReg());
688 
689   if (Src1->isReg())
690     NonDefRegs.push_back(Src1->getReg());
691 
692   if (!findInstrBackwards(
693           MI, [&](MachineInstr *Check) { return Check == VCmp; }, NonDefRegs))
694     return;
695 
696   if (VCmp)
697     SaveExecVCmpMapping[&MI] = VCmp;
698 }
699 
700 // Record occurences of
701 // s_or_saveexec s_o, s_i
702 // s_xor exec, exec, s_o
703 // to be replaced with
704 // s_andn2_saveexec s_o, s_i.
705 void SIOptimizeExecMasking::tryRecordOrSaveexecXorSequence(MachineInstr &MI) {
706   const unsigned XorOpcode =
707       ST->isWave32() ? AMDGPU::S_XOR_B32 : AMDGPU::S_XOR_B64;
708 
709   if (MI.getOpcode() == XorOpcode && &MI != &MI.getParent()->front()) {
710     const MachineOperand &XorDst = MI.getOperand(0);
711     const MachineOperand &XorSrc0 = MI.getOperand(1);
712     const MachineOperand &XorSrc1 = MI.getOperand(2);
713 
714     if (XorDst.isReg() && XorDst.getReg() == Exec && XorSrc0.isReg() &&
715         XorSrc1.isReg() &&
716         (XorSrc0.getReg() == Exec || XorSrc1.getReg() == Exec)) {
717       const unsigned OrSaveexecOpcode = ST->isWave32()
718                                             ? AMDGPU::S_OR_SAVEEXEC_B32
719                                             : AMDGPU::S_OR_SAVEEXEC_B64;
720 
721       // Peek at the previous instruction and check if this is a relevant
722       // s_or_saveexec instruction.
723       MachineInstr &PossibleOrSaveexec = *MI.getPrevNode();
724       if (PossibleOrSaveexec.getOpcode() != OrSaveexecOpcode)
725         return;
726 
727       const MachineOperand &OrDst = PossibleOrSaveexec.getOperand(0);
728       const MachineOperand &OrSrc0 = PossibleOrSaveexec.getOperand(1);
729       if (OrDst.isReg() && OrSrc0.isReg()) {
730         if ((XorSrc0.getReg() == Exec && XorSrc1.getReg() == OrDst.getReg()) ||
731             (XorSrc0.getReg() == OrDst.getReg() && XorSrc1.getReg() == Exec)) {
732           OrXors.emplace_back(&PossibleOrSaveexec, &MI);
733         }
734       }
735     }
736   }
737 }
738 
739 bool SIOptimizeExecMasking::optimizeOrSaveexecXorSequences() {
740   if (OrXors.empty()) {
741     return false;
742   }
743 
744   bool Changed = false;
745   const unsigned Andn2Opcode = ST->isWave32() ? AMDGPU::S_ANDN2_SAVEEXEC_B32
746                                               : AMDGPU::S_ANDN2_SAVEEXEC_B64;
747 
748   for (const auto &Pair : OrXors) {
749     MachineInstr *Or = nullptr;
750     MachineInstr *Xor = nullptr;
751     std::tie(Or, Xor) = Pair;
752     BuildMI(*Or->getParent(), Or->getIterator(), Or->getDebugLoc(),
753             TII->get(Andn2Opcode), Or->getOperand(0).getReg())
754         .addReg(Or->getOperand(1).getReg());
755 
756     Or->eraseFromParent();
757     Xor->eraseFromParent();
758 
759     Changed = true;
760   }
761 
762   return Changed;
763 }
764 
765 bool SIOptimizeExecMasking::runOnMachineFunction(MachineFunction &MF) {
766   if (skipFunction(MF.getFunction()))
767     return false;
768 
769   this->MF = &MF;
770   ST = &MF.getSubtarget<GCNSubtarget>();
771   TRI = ST->getRegisterInfo();
772   TII = ST->getInstrInfo();
773   MRI = &MF.getRegInfo();
774   Exec = TRI->getExec();
775 
776   bool Changed = optimizeExecSequence();
777 
778   OrXors.clear();
779   SaveExecVCmpMapping.clear();
780   static unsigned SearchWindow = 10;
781   for (MachineBasicBlock &MBB : MF) {
782     unsigned SearchCount = 0;
783 
784     for (auto &MI : llvm::reverse(MBB)) {
785       if (MI.isDebugInstr())
786         continue;
787 
788       if (SearchCount >= SearchWindow) {
789         break;
790       }
791 
792       tryRecordOrSaveexecXorSequence(MI);
793       tryRecordVCmpxAndSaveexecSequence(MI);
794 
795       if (MI.modifiesRegister(Exec, TRI)) {
796         break;
797       }
798 
799       ++SearchCount;
800     }
801   }
802 
803   Changed |= optimizeOrSaveexecXorSequences();
804   for (const auto &Entry : SaveExecVCmpMapping) {
805     MachineInstr *SaveExecInstr = Entry.getFirst();
806     MachineInstr *VCmpInstr = Entry.getSecond();
807 
808     Changed |= optimizeVCMPSaveExecSequence(*SaveExecInstr, *VCmpInstr, Exec);
809   }
810 
811   return Changed;
812 }
813