xref: /freebsd/contrib/llvm-project/llvm/lib/CodeGen/ModuloSchedule.cpp (revision 3f0efe05432b1633991114ca4ca330102a561959)
1 //===- ModuloSchedule.cpp - Software pipeline schedule expansion ----------===//
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 "llvm/CodeGen/ModuloSchedule.h"
10 #include "llvm/ADT/StringExtras.h"
11 #include "llvm/Analysis/MemoryLocation.h"
12 #include "llvm/CodeGen/LiveIntervals.h"
13 #include "llvm/CodeGen/MachineInstrBuilder.h"
14 #include "llvm/CodeGen/MachineLoopInfo.h"
15 #include "llvm/CodeGen/MachineRegisterInfo.h"
16 #include "llvm/InitializePasses.h"
17 #include "llvm/MC/MCContext.h"
18 #include "llvm/Support/Debug.h"
19 #include "llvm/Support/ErrorHandling.h"
20 #include "llvm/Support/raw_ostream.h"
21 
22 #define DEBUG_TYPE "pipeliner"
23 using namespace llvm;
24 
25 void ModuloSchedule::print(raw_ostream &OS) {
26   for (MachineInstr *MI : ScheduledInstrs)
27     OS << "[stage " << getStage(MI) << " @" << getCycle(MI) << "c] " << *MI;
28 }
29 
30 //===----------------------------------------------------------------------===//
31 // ModuloScheduleExpander implementation
32 //===----------------------------------------------------------------------===//
33 
34 /// Return the register values for  the operands of a Phi instruction.
35 /// This function assume the instruction is a Phi.
36 static void getPhiRegs(MachineInstr &Phi, MachineBasicBlock *Loop,
37                        unsigned &InitVal, unsigned &LoopVal) {
38   assert(Phi.isPHI() && "Expecting a Phi.");
39 
40   InitVal = 0;
41   LoopVal = 0;
42   for (unsigned i = 1, e = Phi.getNumOperands(); i != e; i += 2)
43     if (Phi.getOperand(i + 1).getMBB() != Loop)
44       InitVal = Phi.getOperand(i).getReg();
45     else
46       LoopVal = Phi.getOperand(i).getReg();
47 
48   assert(InitVal != 0 && LoopVal != 0 && "Unexpected Phi structure.");
49 }
50 
51 /// Return the Phi register value that comes from the incoming block.
52 static unsigned getInitPhiReg(MachineInstr &Phi, MachineBasicBlock *LoopBB) {
53   for (unsigned i = 1, e = Phi.getNumOperands(); i != e; i += 2)
54     if (Phi.getOperand(i + 1).getMBB() != LoopBB)
55       return Phi.getOperand(i).getReg();
56   return 0;
57 }
58 
59 /// Return the Phi register value that comes the loop block.
60 static unsigned getLoopPhiReg(MachineInstr &Phi, MachineBasicBlock *LoopBB) {
61   for (unsigned i = 1, e = Phi.getNumOperands(); i != e; i += 2)
62     if (Phi.getOperand(i + 1).getMBB() == LoopBB)
63       return Phi.getOperand(i).getReg();
64   return 0;
65 }
66 
67 void ModuloScheduleExpander::expand() {
68   BB = Schedule.getLoop()->getTopBlock();
69   Preheader = *BB->pred_begin();
70   if (Preheader == BB)
71     Preheader = *std::next(BB->pred_begin());
72 
73   // Iterate over the definitions in each instruction, and compute the
74   // stage difference for each use.  Keep the maximum value.
75   for (MachineInstr *MI : Schedule.getInstructions()) {
76     int DefStage = Schedule.getStage(MI);
77     for (const MachineOperand &Op : MI->all_defs()) {
78       Register Reg = Op.getReg();
79       unsigned MaxDiff = 0;
80       bool PhiIsSwapped = false;
81       for (MachineOperand &UseOp : MRI.use_operands(Reg)) {
82         MachineInstr *UseMI = UseOp.getParent();
83         int UseStage = Schedule.getStage(UseMI);
84         unsigned Diff = 0;
85         if (UseStage != -1 && UseStage >= DefStage)
86           Diff = UseStage - DefStage;
87         if (MI->isPHI()) {
88           if (isLoopCarried(*MI))
89             ++Diff;
90           else
91             PhiIsSwapped = true;
92         }
93         MaxDiff = std::max(Diff, MaxDiff);
94       }
95       RegToStageDiff[Reg] = std::make_pair(MaxDiff, PhiIsSwapped);
96     }
97   }
98 
99   generatePipelinedLoop();
100 }
101 
102 void ModuloScheduleExpander::generatePipelinedLoop() {
103   LoopInfo = TII->analyzeLoopForPipelining(BB);
104   assert(LoopInfo && "Must be able to analyze loop!");
105 
106   // Create a new basic block for the kernel and add it to the CFG.
107   MachineBasicBlock *KernelBB = MF.CreateMachineBasicBlock(BB->getBasicBlock());
108 
109   unsigned MaxStageCount = Schedule.getNumStages() - 1;
110 
111   // Remember the registers that are used in different stages. The index is
112   // the iteration, or stage, that the instruction is scheduled in.  This is
113   // a map between register names in the original block and the names created
114   // in each stage of the pipelined loop.
115   ValueMapTy *VRMap = new ValueMapTy[(MaxStageCount + 1) * 2];
116 
117   // The renaming destination by Phis for the registers across stages.
118   // This map is updated during Phis generation to point to the most recent
119   // renaming destination.
120   ValueMapTy *VRMapPhi = new ValueMapTy[(MaxStageCount + 1) * 2];
121 
122   InstrMapTy InstrMap;
123 
124   SmallVector<MachineBasicBlock *, 4> PrologBBs;
125 
126   // Generate the prolog instructions that set up the pipeline.
127   generateProlog(MaxStageCount, KernelBB, VRMap, PrologBBs);
128   MF.insert(BB->getIterator(), KernelBB);
129 
130   // Rearrange the instructions to generate the new, pipelined loop,
131   // and update register names as needed.
132   for (MachineInstr *CI : Schedule.getInstructions()) {
133     if (CI->isPHI())
134       continue;
135     unsigned StageNum = Schedule.getStage(CI);
136     MachineInstr *NewMI = cloneInstr(CI, MaxStageCount, StageNum);
137     updateInstruction(NewMI, false, MaxStageCount, StageNum, VRMap);
138     KernelBB->push_back(NewMI);
139     InstrMap[NewMI] = CI;
140   }
141 
142   // Copy any terminator instructions to the new kernel, and update
143   // names as needed.
144   for (MachineInstr &MI : BB->terminators()) {
145     MachineInstr *NewMI = MF.CloneMachineInstr(&MI);
146     updateInstruction(NewMI, false, MaxStageCount, 0, VRMap);
147     KernelBB->push_back(NewMI);
148     InstrMap[NewMI] = &MI;
149   }
150 
151   NewKernel = KernelBB;
152   KernelBB->transferSuccessors(BB);
153   KernelBB->replaceSuccessor(BB, KernelBB);
154 
155   generateExistingPhis(KernelBB, PrologBBs.back(), KernelBB, KernelBB, VRMap,
156                        InstrMap, MaxStageCount, MaxStageCount, false);
157   generatePhis(KernelBB, PrologBBs.back(), KernelBB, KernelBB, VRMap, VRMapPhi,
158                InstrMap, MaxStageCount, MaxStageCount, false);
159 
160   LLVM_DEBUG(dbgs() << "New block\n"; KernelBB->dump(););
161 
162   SmallVector<MachineBasicBlock *, 4> EpilogBBs;
163   // Generate the epilog instructions to complete the pipeline.
164   generateEpilog(MaxStageCount, KernelBB, BB, VRMap, VRMapPhi, EpilogBBs,
165                  PrologBBs);
166 
167   // We need this step because the register allocation doesn't handle some
168   // situations well, so we insert copies to help out.
169   splitLifetimes(KernelBB, EpilogBBs);
170 
171   // Remove dead instructions due to loop induction variables.
172   removeDeadInstructions(KernelBB, EpilogBBs);
173 
174   // Add branches between prolog and epilog blocks.
175   addBranches(*Preheader, PrologBBs, KernelBB, EpilogBBs, VRMap);
176 
177   delete[] VRMap;
178   delete[] VRMapPhi;
179 }
180 
181 void ModuloScheduleExpander::cleanup() {
182   // Remove the original loop since it's no longer referenced.
183   for (auto &I : *BB)
184     LIS.RemoveMachineInstrFromMaps(I);
185   BB->clear();
186   BB->eraseFromParent();
187 }
188 
189 /// Generate the pipeline prolog code.
190 void ModuloScheduleExpander::generateProlog(unsigned LastStage,
191                                             MachineBasicBlock *KernelBB,
192                                             ValueMapTy *VRMap,
193                                             MBBVectorTy &PrologBBs) {
194   MachineBasicBlock *PredBB = Preheader;
195   InstrMapTy InstrMap;
196 
197   // Generate a basic block for each stage, not including the last stage,
198   // which will be generated in the kernel. Each basic block may contain
199   // instructions from multiple stages/iterations.
200   for (unsigned i = 0; i < LastStage; ++i) {
201     // Create and insert the prolog basic block prior to the original loop
202     // basic block.  The original loop is removed later.
203     MachineBasicBlock *NewBB = MF.CreateMachineBasicBlock(BB->getBasicBlock());
204     PrologBBs.push_back(NewBB);
205     MF.insert(BB->getIterator(), NewBB);
206     NewBB->transferSuccessors(PredBB);
207     PredBB->addSuccessor(NewBB);
208     PredBB = NewBB;
209 
210     // Generate instructions for each appropriate stage. Process instructions
211     // in original program order.
212     for (int StageNum = i; StageNum >= 0; --StageNum) {
213       for (MachineBasicBlock::iterator BBI = BB->instr_begin(),
214                                        BBE = BB->getFirstTerminator();
215            BBI != BBE; ++BBI) {
216         if (Schedule.getStage(&*BBI) == StageNum) {
217           if (BBI->isPHI())
218             continue;
219           MachineInstr *NewMI =
220               cloneAndChangeInstr(&*BBI, i, (unsigned)StageNum);
221           updateInstruction(NewMI, false, i, (unsigned)StageNum, VRMap);
222           NewBB->push_back(NewMI);
223           InstrMap[NewMI] = &*BBI;
224         }
225       }
226     }
227     rewritePhiValues(NewBB, i, VRMap, InstrMap);
228     LLVM_DEBUG({
229       dbgs() << "prolog:\n";
230       NewBB->dump();
231     });
232   }
233 
234   PredBB->replaceSuccessor(BB, KernelBB);
235 
236   // Check if we need to remove the branch from the preheader to the original
237   // loop, and replace it with a branch to the new loop.
238   unsigned numBranches = TII->removeBranch(*Preheader);
239   if (numBranches) {
240     SmallVector<MachineOperand, 0> Cond;
241     TII->insertBranch(*Preheader, PrologBBs[0], nullptr, Cond, DebugLoc());
242   }
243 }
244 
245 /// Generate the pipeline epilog code. The epilog code finishes the iterations
246 /// that were started in either the prolog or the kernel.  We create a basic
247 /// block for each stage that needs to complete.
248 void ModuloScheduleExpander::generateEpilog(
249     unsigned LastStage, MachineBasicBlock *KernelBB, MachineBasicBlock *OrigBB,
250     ValueMapTy *VRMap, ValueMapTy *VRMapPhi, MBBVectorTy &EpilogBBs,
251     MBBVectorTy &PrologBBs) {
252   // We need to change the branch from the kernel to the first epilog block, so
253   // this call to analyze branch uses the kernel rather than the original BB.
254   MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
255   SmallVector<MachineOperand, 4> Cond;
256   bool checkBranch = TII->analyzeBranch(*KernelBB, TBB, FBB, Cond);
257   assert(!checkBranch && "generateEpilog must be able to analyze the branch");
258   if (checkBranch)
259     return;
260 
261   MachineBasicBlock::succ_iterator LoopExitI = KernelBB->succ_begin();
262   if (*LoopExitI == KernelBB)
263     ++LoopExitI;
264   assert(LoopExitI != KernelBB->succ_end() && "Expecting a successor");
265   MachineBasicBlock *LoopExitBB = *LoopExitI;
266 
267   MachineBasicBlock *PredBB = KernelBB;
268   MachineBasicBlock *EpilogStart = LoopExitBB;
269   InstrMapTy InstrMap;
270 
271   // Generate a basic block for each stage, not including the last stage,
272   // which was generated for the kernel.  Each basic block may contain
273   // instructions from multiple stages/iterations.
274   int EpilogStage = LastStage + 1;
275   for (unsigned i = LastStage; i >= 1; --i, ++EpilogStage) {
276     MachineBasicBlock *NewBB = MF.CreateMachineBasicBlock();
277     EpilogBBs.push_back(NewBB);
278     MF.insert(BB->getIterator(), NewBB);
279 
280     PredBB->replaceSuccessor(LoopExitBB, NewBB);
281     NewBB->addSuccessor(LoopExitBB);
282 
283     if (EpilogStart == LoopExitBB)
284       EpilogStart = NewBB;
285 
286     // Add instructions to the epilog depending on the current block.
287     // Process instructions in original program order.
288     for (unsigned StageNum = i; StageNum <= LastStage; ++StageNum) {
289       for (auto &BBI : *BB) {
290         if (BBI.isPHI())
291           continue;
292         MachineInstr *In = &BBI;
293         if ((unsigned)Schedule.getStage(In) == StageNum) {
294           // Instructions with memoperands in the epilog are updated with
295           // conservative values.
296           MachineInstr *NewMI = cloneInstr(In, UINT_MAX, 0);
297           updateInstruction(NewMI, i == 1, EpilogStage, 0, VRMap);
298           NewBB->push_back(NewMI);
299           InstrMap[NewMI] = In;
300         }
301       }
302     }
303     generateExistingPhis(NewBB, PrologBBs[i - 1], PredBB, KernelBB, VRMap,
304                          InstrMap, LastStage, EpilogStage, i == 1);
305     generatePhis(NewBB, PrologBBs[i - 1], PredBB, KernelBB, VRMap, VRMapPhi,
306                  InstrMap, LastStage, EpilogStage, i == 1);
307     PredBB = NewBB;
308 
309     LLVM_DEBUG({
310       dbgs() << "epilog:\n";
311       NewBB->dump();
312     });
313   }
314 
315   // Fix any Phi nodes in the loop exit block.
316   LoopExitBB->replacePhiUsesWith(BB, PredBB);
317 
318   // Create a branch to the new epilog from the kernel.
319   // Remove the original branch and add a new branch to the epilog.
320   TII->removeBranch(*KernelBB);
321   assert((OrigBB == TBB || OrigBB == FBB) &&
322          "Unable to determine looping branch direction");
323   if (OrigBB != TBB)
324     TII->insertBranch(*KernelBB, EpilogStart, KernelBB, Cond, DebugLoc());
325   else
326     TII->insertBranch(*KernelBB, KernelBB, EpilogStart, Cond, DebugLoc());
327   // Add a branch to the loop exit.
328   if (EpilogBBs.size() > 0) {
329     MachineBasicBlock *LastEpilogBB = EpilogBBs.back();
330     SmallVector<MachineOperand, 4> Cond1;
331     TII->insertBranch(*LastEpilogBB, LoopExitBB, nullptr, Cond1, DebugLoc());
332   }
333 }
334 
335 /// Replace all uses of FromReg that appear outside the specified
336 /// basic block with ToReg.
337 static void replaceRegUsesAfterLoop(unsigned FromReg, unsigned ToReg,
338                                     MachineBasicBlock *MBB,
339                                     MachineRegisterInfo &MRI,
340                                     LiveIntervals &LIS) {
341   for (MachineOperand &O :
342        llvm::make_early_inc_range(MRI.use_operands(FromReg)))
343     if (O.getParent()->getParent() != MBB)
344       O.setReg(ToReg);
345   if (!LIS.hasInterval(ToReg))
346     LIS.createEmptyInterval(ToReg);
347 }
348 
349 /// Return true if the register has a use that occurs outside the
350 /// specified loop.
351 static bool hasUseAfterLoop(unsigned Reg, MachineBasicBlock *BB,
352                             MachineRegisterInfo &MRI) {
353   for (const MachineOperand &MO : MRI.use_operands(Reg))
354     if (MO.getParent()->getParent() != BB)
355       return true;
356   return false;
357 }
358 
359 /// Generate Phis for the specific block in the generated pipelined code.
360 /// This function looks at the Phis from the original code to guide the
361 /// creation of new Phis.
362 void ModuloScheduleExpander::generateExistingPhis(
363     MachineBasicBlock *NewBB, MachineBasicBlock *BB1, MachineBasicBlock *BB2,
364     MachineBasicBlock *KernelBB, ValueMapTy *VRMap, InstrMapTy &InstrMap,
365     unsigned LastStageNum, unsigned CurStageNum, bool IsLast) {
366   // Compute the stage number for the initial value of the Phi, which
367   // comes from the prolog. The prolog to use depends on to which kernel/
368   // epilog that we're adding the Phi.
369   unsigned PrologStage = 0;
370   unsigned PrevStage = 0;
371   bool InKernel = (LastStageNum == CurStageNum);
372   if (InKernel) {
373     PrologStage = LastStageNum - 1;
374     PrevStage = CurStageNum;
375   } else {
376     PrologStage = LastStageNum - (CurStageNum - LastStageNum);
377     PrevStage = LastStageNum + (CurStageNum - LastStageNum) - 1;
378   }
379 
380   for (MachineBasicBlock::iterator BBI = BB->instr_begin(),
381                                    BBE = BB->getFirstNonPHI();
382        BBI != BBE; ++BBI) {
383     Register Def = BBI->getOperand(0).getReg();
384 
385     unsigned InitVal = 0;
386     unsigned LoopVal = 0;
387     getPhiRegs(*BBI, BB, InitVal, LoopVal);
388 
389     unsigned PhiOp1 = 0;
390     // The Phi value from the loop body typically is defined in the loop, but
391     // not always. So, we need to check if the value is defined in the loop.
392     unsigned PhiOp2 = LoopVal;
393     if (VRMap[LastStageNum].count(LoopVal))
394       PhiOp2 = VRMap[LastStageNum][LoopVal];
395 
396     int StageScheduled = Schedule.getStage(&*BBI);
397     int LoopValStage = Schedule.getStage(MRI.getVRegDef(LoopVal));
398     unsigned NumStages = getStagesForReg(Def, CurStageNum);
399     if (NumStages == 0) {
400       // We don't need to generate a Phi anymore, but we need to rename any uses
401       // of the Phi value.
402       unsigned NewReg = VRMap[PrevStage][LoopVal];
403       rewriteScheduledInstr(NewBB, InstrMap, CurStageNum, 0, &*BBI, Def,
404                             InitVal, NewReg);
405       if (VRMap[CurStageNum].count(LoopVal))
406         VRMap[CurStageNum][Def] = VRMap[CurStageNum][LoopVal];
407     }
408     // Adjust the number of Phis needed depending on the number of prologs left,
409     // and the distance from where the Phi is first scheduled. The number of
410     // Phis cannot exceed the number of prolog stages. Each stage can
411     // potentially define two values.
412     unsigned MaxPhis = PrologStage + 2;
413     if (!InKernel && (int)PrologStage <= LoopValStage)
414       MaxPhis = std::max((int)MaxPhis - (int)LoopValStage, 1);
415     unsigned NumPhis = std::min(NumStages, MaxPhis);
416 
417     unsigned NewReg = 0;
418     unsigned AccessStage = (LoopValStage != -1) ? LoopValStage : StageScheduled;
419     // In the epilog, we may need to look back one stage to get the correct
420     // Phi name, because the epilog and prolog blocks execute the same stage.
421     // The correct name is from the previous block only when the Phi has
422     // been completely scheduled prior to the epilog, and Phi value is not
423     // needed in multiple stages.
424     int StageDiff = 0;
425     if (!InKernel && StageScheduled >= LoopValStage && AccessStage == 0 &&
426         NumPhis == 1)
427       StageDiff = 1;
428     // Adjust the computations below when the phi and the loop definition
429     // are scheduled in different stages.
430     if (InKernel && LoopValStage != -1 && StageScheduled > LoopValStage)
431       StageDiff = StageScheduled - LoopValStage;
432     for (unsigned np = 0; np < NumPhis; ++np) {
433       // If the Phi hasn't been scheduled, then use the initial Phi operand
434       // value. Otherwise, use the scheduled version of the instruction. This
435       // is a little complicated when a Phi references another Phi.
436       if (np > PrologStage || StageScheduled >= (int)LastStageNum)
437         PhiOp1 = InitVal;
438       // Check if the Phi has already been scheduled in a prolog stage.
439       else if (PrologStage >= AccessStage + StageDiff + np &&
440                VRMap[PrologStage - StageDiff - np].count(LoopVal) != 0)
441         PhiOp1 = VRMap[PrologStage - StageDiff - np][LoopVal];
442       // Check if the Phi has already been scheduled, but the loop instruction
443       // is either another Phi, or doesn't occur in the loop.
444       else if (PrologStage >= AccessStage + StageDiff + np) {
445         // If the Phi references another Phi, we need to examine the other
446         // Phi to get the correct value.
447         PhiOp1 = LoopVal;
448         MachineInstr *InstOp1 = MRI.getVRegDef(PhiOp1);
449         int Indirects = 1;
450         while (InstOp1 && InstOp1->isPHI() && InstOp1->getParent() == BB) {
451           int PhiStage = Schedule.getStage(InstOp1);
452           if ((int)(PrologStage - StageDiff - np) < PhiStage + Indirects)
453             PhiOp1 = getInitPhiReg(*InstOp1, BB);
454           else
455             PhiOp1 = getLoopPhiReg(*InstOp1, BB);
456           InstOp1 = MRI.getVRegDef(PhiOp1);
457           int PhiOpStage = Schedule.getStage(InstOp1);
458           int StageAdj = (PhiOpStage != -1 ? PhiStage - PhiOpStage : 0);
459           if (PhiOpStage != -1 && PrologStage - StageAdj >= Indirects + np &&
460               VRMap[PrologStage - StageAdj - Indirects - np].count(PhiOp1)) {
461             PhiOp1 = VRMap[PrologStage - StageAdj - Indirects - np][PhiOp1];
462             break;
463           }
464           ++Indirects;
465         }
466       } else
467         PhiOp1 = InitVal;
468       // If this references a generated Phi in the kernel, get the Phi operand
469       // from the incoming block.
470       if (MachineInstr *InstOp1 = MRI.getVRegDef(PhiOp1))
471         if (InstOp1->isPHI() && InstOp1->getParent() == KernelBB)
472           PhiOp1 = getInitPhiReg(*InstOp1, KernelBB);
473 
474       MachineInstr *PhiInst = MRI.getVRegDef(LoopVal);
475       bool LoopDefIsPhi = PhiInst && PhiInst->isPHI();
476       // In the epilog, a map lookup is needed to get the value from the kernel,
477       // or previous epilog block. How is does this depends on if the
478       // instruction is scheduled in the previous block.
479       if (!InKernel) {
480         int StageDiffAdj = 0;
481         if (LoopValStage != -1 && StageScheduled > LoopValStage)
482           StageDiffAdj = StageScheduled - LoopValStage;
483         // Use the loop value defined in the kernel, unless the kernel
484         // contains the last definition of the Phi.
485         if (np == 0 && PrevStage == LastStageNum &&
486             (StageScheduled != 0 || LoopValStage != 0) &&
487             VRMap[PrevStage - StageDiffAdj].count(LoopVal))
488           PhiOp2 = VRMap[PrevStage - StageDiffAdj][LoopVal];
489         // Use the value defined by the Phi. We add one because we switch
490         // from looking at the loop value to the Phi definition.
491         else if (np > 0 && PrevStage == LastStageNum &&
492                  VRMap[PrevStage - np + 1].count(Def))
493           PhiOp2 = VRMap[PrevStage - np + 1][Def];
494         // Use the loop value defined in the kernel.
495         else if (static_cast<unsigned>(LoopValStage) > PrologStage + 1 &&
496                  VRMap[PrevStage - StageDiffAdj - np].count(LoopVal))
497           PhiOp2 = VRMap[PrevStage - StageDiffAdj - np][LoopVal];
498         // Use the value defined by the Phi, unless we're generating the first
499         // epilog and the Phi refers to a Phi in a different stage.
500         else if (VRMap[PrevStage - np].count(Def) &&
501                  (!LoopDefIsPhi || (PrevStage != LastStageNum) ||
502                   (LoopValStage == StageScheduled)))
503           PhiOp2 = VRMap[PrevStage - np][Def];
504       }
505 
506       // Check if we can reuse an existing Phi. This occurs when a Phi
507       // references another Phi, and the other Phi is scheduled in an
508       // earlier stage. We can try to reuse an existing Phi up until the last
509       // stage of the current Phi.
510       if (LoopDefIsPhi) {
511         if (static_cast<int>(PrologStage - np) >= StageScheduled) {
512           int LVNumStages = getStagesForPhi(LoopVal);
513           int StageDiff = (StageScheduled - LoopValStage);
514           LVNumStages -= StageDiff;
515           // Make sure the loop value Phi has been processed already.
516           if (LVNumStages > (int)np && VRMap[CurStageNum].count(LoopVal)) {
517             NewReg = PhiOp2;
518             unsigned ReuseStage = CurStageNum;
519             if (isLoopCarried(*PhiInst))
520               ReuseStage -= LVNumStages;
521             // Check if the Phi to reuse has been generated yet. If not, then
522             // there is nothing to reuse.
523             if (VRMap[ReuseStage - np].count(LoopVal)) {
524               NewReg = VRMap[ReuseStage - np][LoopVal];
525 
526               rewriteScheduledInstr(NewBB, InstrMap, CurStageNum, np, &*BBI,
527                                     Def, NewReg);
528               // Update the map with the new Phi name.
529               VRMap[CurStageNum - np][Def] = NewReg;
530               PhiOp2 = NewReg;
531               if (VRMap[LastStageNum - np - 1].count(LoopVal))
532                 PhiOp2 = VRMap[LastStageNum - np - 1][LoopVal];
533 
534               if (IsLast && np == NumPhis - 1)
535                 replaceRegUsesAfterLoop(Def, NewReg, BB, MRI, LIS);
536               continue;
537             }
538           }
539         }
540         if (InKernel && StageDiff > 0 &&
541             VRMap[CurStageNum - StageDiff - np].count(LoopVal))
542           PhiOp2 = VRMap[CurStageNum - StageDiff - np][LoopVal];
543       }
544 
545       const TargetRegisterClass *RC = MRI.getRegClass(Def);
546       NewReg = MRI.createVirtualRegister(RC);
547 
548       MachineInstrBuilder NewPhi =
549           BuildMI(*NewBB, NewBB->getFirstNonPHI(), DebugLoc(),
550                   TII->get(TargetOpcode::PHI), NewReg);
551       NewPhi.addReg(PhiOp1).addMBB(BB1);
552       NewPhi.addReg(PhiOp2).addMBB(BB2);
553       if (np == 0)
554         InstrMap[NewPhi] = &*BBI;
555 
556       // We define the Phis after creating the new pipelined code, so
557       // we need to rename the Phi values in scheduled instructions.
558 
559       unsigned PrevReg = 0;
560       if (InKernel && VRMap[PrevStage - np].count(LoopVal))
561         PrevReg = VRMap[PrevStage - np][LoopVal];
562       rewriteScheduledInstr(NewBB, InstrMap, CurStageNum, np, &*BBI, Def,
563                             NewReg, PrevReg);
564       // If the Phi has been scheduled, use the new name for rewriting.
565       if (VRMap[CurStageNum - np].count(Def)) {
566         unsigned R = VRMap[CurStageNum - np][Def];
567         rewriteScheduledInstr(NewBB, InstrMap, CurStageNum, np, &*BBI, R,
568                               NewReg);
569       }
570 
571       // Check if we need to rename any uses that occurs after the loop. The
572       // register to replace depends on whether the Phi is scheduled in the
573       // epilog.
574       if (IsLast && np == NumPhis - 1)
575         replaceRegUsesAfterLoop(Def, NewReg, BB, MRI, LIS);
576 
577       // In the kernel, a dependent Phi uses the value from this Phi.
578       if (InKernel)
579         PhiOp2 = NewReg;
580 
581       // Update the map with the new Phi name.
582       VRMap[CurStageNum - np][Def] = NewReg;
583     }
584 
585     while (NumPhis++ < NumStages) {
586       rewriteScheduledInstr(NewBB, InstrMap, CurStageNum, NumPhis, &*BBI, Def,
587                             NewReg, 0);
588     }
589 
590     // Check if we need to rename a Phi that has been eliminated due to
591     // scheduling.
592     if (NumStages == 0 && IsLast && VRMap[CurStageNum].count(LoopVal))
593       replaceRegUsesAfterLoop(Def, VRMap[CurStageNum][LoopVal], BB, MRI, LIS);
594   }
595 }
596 
597 /// Generate Phis for the specified block in the generated pipelined code.
598 /// These are new Phis needed because the definition is scheduled after the
599 /// use in the pipelined sequence.
600 void ModuloScheduleExpander::generatePhis(
601     MachineBasicBlock *NewBB, MachineBasicBlock *BB1, MachineBasicBlock *BB2,
602     MachineBasicBlock *KernelBB, ValueMapTy *VRMap, ValueMapTy *VRMapPhi,
603     InstrMapTy &InstrMap, unsigned LastStageNum, unsigned CurStageNum,
604     bool IsLast) {
605   // Compute the stage number that contains the initial Phi value, and
606   // the Phi from the previous stage.
607   unsigned PrologStage = 0;
608   unsigned PrevStage = 0;
609   unsigned StageDiff = CurStageNum - LastStageNum;
610   bool InKernel = (StageDiff == 0);
611   if (InKernel) {
612     PrologStage = LastStageNum - 1;
613     PrevStage = CurStageNum;
614   } else {
615     PrologStage = LastStageNum - StageDiff;
616     PrevStage = LastStageNum + StageDiff - 1;
617   }
618 
619   for (MachineBasicBlock::iterator BBI = BB->getFirstNonPHI(),
620                                    BBE = BB->instr_end();
621        BBI != BBE; ++BBI) {
622     for (unsigned i = 0, e = BBI->getNumOperands(); i != e; ++i) {
623       MachineOperand &MO = BBI->getOperand(i);
624       if (!MO.isReg() || !MO.isDef() || !MO.getReg().isVirtual())
625         continue;
626 
627       int StageScheduled = Schedule.getStage(&*BBI);
628       assert(StageScheduled != -1 && "Expecting scheduled instruction.");
629       Register Def = MO.getReg();
630       unsigned NumPhis = getStagesForReg(Def, CurStageNum);
631       // An instruction scheduled in stage 0 and is used after the loop
632       // requires a phi in the epilog for the last definition from either
633       // the kernel or prolog.
634       if (!InKernel && NumPhis == 0 && StageScheduled == 0 &&
635           hasUseAfterLoop(Def, BB, MRI))
636         NumPhis = 1;
637       if (!InKernel && (unsigned)StageScheduled > PrologStage)
638         continue;
639 
640       unsigned PhiOp2;
641       if (InKernel) {
642         PhiOp2 = VRMap[PrevStage][Def];
643         if (MachineInstr *InstOp2 = MRI.getVRegDef(PhiOp2))
644           if (InstOp2->isPHI() && InstOp2->getParent() == NewBB)
645             PhiOp2 = getLoopPhiReg(*InstOp2, BB2);
646       }
647       // The number of Phis can't exceed the number of prolog stages. The
648       // prolog stage number is zero based.
649       if (NumPhis > PrologStage + 1 - StageScheduled)
650         NumPhis = PrologStage + 1 - StageScheduled;
651       for (unsigned np = 0; np < NumPhis; ++np) {
652         // Example for
653         // Org:
654         //   %Org = ... (Scheduled at Stage#0, NumPhi = 2)
655         //
656         // Prolog0 (Stage0):
657         //   %Clone0 = ...
658         // Prolog1 (Stage1):
659         //   %Clone1 = ...
660         // Kernel (Stage2):
661         //   %Phi0 = Phi %Clone1, Prolog1, %Clone2, Kernel
662         //   %Phi1 = Phi %Clone0, Prolog1, %Phi0, Kernel
663         //   %Clone2 = ...
664         // Epilog0 (Stage3):
665         //   %Phi2 = Phi %Clone1, Prolog1, %Clone2, Kernel
666         //   %Phi3 = Phi %Clone0, Prolog1, %Phi0, Kernel
667         // Epilog1 (Stage4):
668         //   %Phi4 = Phi %Clone0, Prolog0, %Phi2, Epilog0
669         //
670         // VRMap = {0: %Clone0, 1: %Clone1, 2: %Clone2}
671         // VRMapPhi (after Kernel) = {0: %Phi1, 1: %Phi0}
672         // VRMapPhi (after Epilog0) = {0: %Phi3, 1: %Phi2}
673 
674         unsigned PhiOp1 = VRMap[PrologStage][Def];
675         if (np <= PrologStage)
676           PhiOp1 = VRMap[PrologStage - np][Def];
677         if (!InKernel) {
678           if (PrevStage == LastStageNum && np == 0)
679             PhiOp2 = VRMap[LastStageNum][Def];
680           else
681             PhiOp2 = VRMapPhi[PrevStage - np][Def];
682         }
683 
684         const TargetRegisterClass *RC = MRI.getRegClass(Def);
685         Register NewReg = MRI.createVirtualRegister(RC);
686 
687         MachineInstrBuilder NewPhi =
688             BuildMI(*NewBB, NewBB->getFirstNonPHI(), DebugLoc(),
689                     TII->get(TargetOpcode::PHI), NewReg);
690         NewPhi.addReg(PhiOp1).addMBB(BB1);
691         NewPhi.addReg(PhiOp2).addMBB(BB2);
692         if (np == 0)
693           InstrMap[NewPhi] = &*BBI;
694 
695         // Rewrite uses and update the map. The actions depend upon whether
696         // we generating code for the kernel or epilog blocks.
697         if (InKernel) {
698           rewriteScheduledInstr(NewBB, InstrMap, CurStageNum, np, &*BBI, PhiOp1,
699                                 NewReg);
700           rewriteScheduledInstr(NewBB, InstrMap, CurStageNum, np, &*BBI, PhiOp2,
701                                 NewReg);
702 
703           PhiOp2 = NewReg;
704           VRMapPhi[PrevStage - np - 1][Def] = NewReg;
705         } else {
706           VRMapPhi[CurStageNum - np][Def] = NewReg;
707           if (np == NumPhis - 1)
708             rewriteScheduledInstr(NewBB, InstrMap, CurStageNum, np, &*BBI, Def,
709                                   NewReg);
710         }
711         if (IsLast && np == NumPhis - 1)
712           replaceRegUsesAfterLoop(Def, NewReg, BB, MRI, LIS);
713       }
714     }
715   }
716 }
717 
718 /// Remove instructions that generate values with no uses.
719 /// Typically, these are induction variable operations that generate values
720 /// used in the loop itself.  A dead instruction has a definition with
721 /// no uses, or uses that occur in the original loop only.
722 void ModuloScheduleExpander::removeDeadInstructions(MachineBasicBlock *KernelBB,
723                                                     MBBVectorTy &EpilogBBs) {
724   // For each epilog block, check that the value defined by each instruction
725   // is used.  If not, delete it.
726   for (MachineBasicBlock *MBB : llvm::reverse(EpilogBBs))
727     for (MachineBasicBlock::reverse_instr_iterator MI = MBB->instr_rbegin(),
728                                                    ME = MBB->instr_rend();
729          MI != ME;) {
730       // From DeadMachineInstructionElem. Don't delete inline assembly.
731       if (MI->isInlineAsm()) {
732         ++MI;
733         continue;
734       }
735       bool SawStore = false;
736       // Check if it's safe to remove the instruction due to side effects.
737       // We can, and want to, remove Phis here.
738       if (!MI->isSafeToMove(nullptr, SawStore) && !MI->isPHI()) {
739         ++MI;
740         continue;
741       }
742       bool used = true;
743       for (const MachineOperand &MO : MI->all_defs()) {
744         Register reg = MO.getReg();
745         // Assume physical registers are used, unless they are marked dead.
746         if (reg.isPhysical()) {
747           used = !MO.isDead();
748           if (used)
749             break;
750           continue;
751         }
752         unsigned realUses = 0;
753         for (const MachineOperand &U : MRI.use_operands(reg)) {
754           // Check if there are any uses that occur only in the original
755           // loop.  If so, that's not a real use.
756           if (U.getParent()->getParent() != BB) {
757             realUses++;
758             used = true;
759             break;
760           }
761         }
762         if (realUses > 0)
763           break;
764         used = false;
765       }
766       if (!used) {
767         LIS.RemoveMachineInstrFromMaps(*MI);
768         MI++->eraseFromParent();
769         continue;
770       }
771       ++MI;
772     }
773   // In the kernel block, check if we can remove a Phi that generates a value
774   // used in an instruction removed in the epilog block.
775   for (MachineInstr &MI : llvm::make_early_inc_range(KernelBB->phis())) {
776     Register reg = MI.getOperand(0).getReg();
777     if (MRI.use_begin(reg) == MRI.use_end()) {
778       LIS.RemoveMachineInstrFromMaps(MI);
779       MI.eraseFromParent();
780     }
781   }
782 }
783 
784 /// For loop carried definitions, we split the lifetime of a virtual register
785 /// that has uses past the definition in the next iteration. A copy with a new
786 /// virtual register is inserted before the definition, which helps with
787 /// generating a better register assignment.
788 ///
789 ///   v1 = phi(a, v2)     v1 = phi(a, v2)
790 ///   v2 = phi(b, v3)     v2 = phi(b, v3)
791 ///   v3 = ..             v4 = copy v1
792 ///   .. = V1             v3 = ..
793 ///                       .. = v4
794 void ModuloScheduleExpander::splitLifetimes(MachineBasicBlock *KernelBB,
795                                             MBBVectorTy &EpilogBBs) {
796   const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
797   for (auto &PHI : KernelBB->phis()) {
798     Register Def = PHI.getOperand(0).getReg();
799     // Check for any Phi definition that used as an operand of another Phi
800     // in the same block.
801     for (MachineRegisterInfo::use_instr_iterator I = MRI.use_instr_begin(Def),
802                                                  E = MRI.use_instr_end();
803          I != E; ++I) {
804       if (I->isPHI() && I->getParent() == KernelBB) {
805         // Get the loop carried definition.
806         unsigned LCDef = getLoopPhiReg(PHI, KernelBB);
807         if (!LCDef)
808           continue;
809         MachineInstr *MI = MRI.getVRegDef(LCDef);
810         if (!MI || MI->getParent() != KernelBB || MI->isPHI())
811           continue;
812         // Search through the rest of the block looking for uses of the Phi
813         // definition. If one occurs, then split the lifetime.
814         unsigned SplitReg = 0;
815         for (auto &BBJ : make_range(MachineBasicBlock::instr_iterator(MI),
816                                     KernelBB->instr_end()))
817           if (BBJ.readsRegister(Def)) {
818             // We split the lifetime when we find the first use.
819             if (SplitReg == 0) {
820               SplitReg = MRI.createVirtualRegister(MRI.getRegClass(Def));
821               BuildMI(*KernelBB, MI, MI->getDebugLoc(),
822                       TII->get(TargetOpcode::COPY), SplitReg)
823                   .addReg(Def);
824             }
825             BBJ.substituteRegister(Def, SplitReg, 0, *TRI);
826           }
827         if (!SplitReg)
828           continue;
829         // Search through each of the epilog blocks for any uses to be renamed.
830         for (auto &Epilog : EpilogBBs)
831           for (auto &I : *Epilog)
832             if (I.readsRegister(Def))
833               I.substituteRegister(Def, SplitReg, 0, *TRI);
834         break;
835       }
836     }
837   }
838 }
839 
840 /// Remove the incoming block from the Phis in a basic block.
841 static void removePhis(MachineBasicBlock *BB, MachineBasicBlock *Incoming) {
842   for (MachineInstr &MI : *BB) {
843     if (!MI.isPHI())
844       break;
845     for (unsigned i = 1, e = MI.getNumOperands(); i != e; i += 2)
846       if (MI.getOperand(i + 1).getMBB() == Incoming) {
847         MI.removeOperand(i + 1);
848         MI.removeOperand(i);
849         break;
850       }
851   }
852 }
853 
854 /// Create branches from each prolog basic block to the appropriate epilog
855 /// block.  These edges are needed if the loop ends before reaching the
856 /// kernel.
857 void ModuloScheduleExpander::addBranches(MachineBasicBlock &PreheaderBB,
858                                          MBBVectorTy &PrologBBs,
859                                          MachineBasicBlock *KernelBB,
860                                          MBBVectorTy &EpilogBBs,
861                                          ValueMapTy *VRMap) {
862   assert(PrologBBs.size() == EpilogBBs.size() && "Prolog/Epilog mismatch");
863   MachineBasicBlock *LastPro = KernelBB;
864   MachineBasicBlock *LastEpi = KernelBB;
865 
866   // Start from the blocks connected to the kernel and work "out"
867   // to the first prolog and the last epilog blocks.
868   SmallVector<MachineInstr *, 4> PrevInsts;
869   unsigned MaxIter = PrologBBs.size() - 1;
870   for (unsigned i = 0, j = MaxIter; i <= MaxIter; ++i, --j) {
871     // Add branches to the prolog that go to the corresponding
872     // epilog, and the fall-thru prolog/kernel block.
873     MachineBasicBlock *Prolog = PrologBBs[j];
874     MachineBasicBlock *Epilog = EpilogBBs[i];
875 
876     SmallVector<MachineOperand, 4> Cond;
877     std::optional<bool> StaticallyGreater =
878         LoopInfo->createTripCountGreaterCondition(j + 1, *Prolog, Cond);
879     unsigned numAdded = 0;
880     if (!StaticallyGreater) {
881       Prolog->addSuccessor(Epilog);
882       numAdded = TII->insertBranch(*Prolog, Epilog, LastPro, Cond, DebugLoc());
883     } else if (*StaticallyGreater == false) {
884       Prolog->addSuccessor(Epilog);
885       Prolog->removeSuccessor(LastPro);
886       LastEpi->removeSuccessor(Epilog);
887       numAdded = TII->insertBranch(*Prolog, Epilog, nullptr, Cond, DebugLoc());
888       removePhis(Epilog, LastEpi);
889       // Remove the blocks that are no longer referenced.
890       if (LastPro != LastEpi) {
891         LastEpi->clear();
892         LastEpi->eraseFromParent();
893       }
894       if (LastPro == KernelBB) {
895         LoopInfo->disposed();
896         NewKernel = nullptr;
897       }
898       LastPro->clear();
899       LastPro->eraseFromParent();
900     } else {
901       numAdded = TII->insertBranch(*Prolog, LastPro, nullptr, Cond, DebugLoc());
902       removePhis(Epilog, Prolog);
903     }
904     LastPro = Prolog;
905     LastEpi = Epilog;
906     for (MachineBasicBlock::reverse_instr_iterator I = Prolog->instr_rbegin(),
907                                                    E = Prolog->instr_rend();
908          I != E && numAdded > 0; ++I, --numAdded)
909       updateInstruction(&*I, false, j, 0, VRMap);
910   }
911 
912   if (NewKernel) {
913     LoopInfo->setPreheader(PrologBBs[MaxIter]);
914     LoopInfo->adjustTripCount(-(MaxIter + 1));
915   }
916 }
917 
918 /// Return true if we can compute the amount the instruction changes
919 /// during each iteration. Set Delta to the amount of the change.
920 bool ModuloScheduleExpander::computeDelta(MachineInstr &MI, unsigned &Delta) {
921   const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
922   const MachineOperand *BaseOp;
923   int64_t Offset;
924   bool OffsetIsScalable;
925   if (!TII->getMemOperandWithOffset(MI, BaseOp, Offset, OffsetIsScalable, TRI))
926     return false;
927 
928   // FIXME: This algorithm assumes instructions have fixed-size offsets.
929   if (OffsetIsScalable)
930     return false;
931 
932   if (!BaseOp->isReg())
933     return false;
934 
935   Register BaseReg = BaseOp->getReg();
936 
937   MachineRegisterInfo &MRI = MF.getRegInfo();
938   // Check if there is a Phi. If so, get the definition in the loop.
939   MachineInstr *BaseDef = MRI.getVRegDef(BaseReg);
940   if (BaseDef && BaseDef->isPHI()) {
941     BaseReg = getLoopPhiReg(*BaseDef, MI.getParent());
942     BaseDef = MRI.getVRegDef(BaseReg);
943   }
944   if (!BaseDef)
945     return false;
946 
947   int D = 0;
948   if (!TII->getIncrementValue(*BaseDef, D) && D >= 0)
949     return false;
950 
951   Delta = D;
952   return true;
953 }
954 
955 /// Update the memory operand with a new offset when the pipeliner
956 /// generates a new copy of the instruction that refers to a
957 /// different memory location.
958 void ModuloScheduleExpander::updateMemOperands(MachineInstr &NewMI,
959                                                MachineInstr &OldMI,
960                                                unsigned Num) {
961   if (Num == 0)
962     return;
963   // If the instruction has memory operands, then adjust the offset
964   // when the instruction appears in different stages.
965   if (NewMI.memoperands_empty())
966     return;
967   SmallVector<MachineMemOperand *, 2> NewMMOs;
968   for (MachineMemOperand *MMO : NewMI.memoperands()) {
969     // TODO: Figure out whether isAtomic is really necessary (see D57601).
970     if (MMO->isVolatile() || MMO->isAtomic() ||
971         (MMO->isInvariant() && MMO->isDereferenceable()) ||
972         (!MMO->getValue())) {
973       NewMMOs.push_back(MMO);
974       continue;
975     }
976     unsigned Delta;
977     if (Num != UINT_MAX && computeDelta(OldMI, Delta)) {
978       int64_t AdjOffset = Delta * Num;
979       NewMMOs.push_back(
980           MF.getMachineMemOperand(MMO, AdjOffset, MMO->getSize()));
981     } else {
982       NewMMOs.push_back(
983           MF.getMachineMemOperand(MMO, 0, MemoryLocation::UnknownSize));
984     }
985   }
986   NewMI.setMemRefs(MF, NewMMOs);
987 }
988 
989 /// Clone the instruction for the new pipelined loop and update the
990 /// memory operands, if needed.
991 MachineInstr *ModuloScheduleExpander::cloneInstr(MachineInstr *OldMI,
992                                                  unsigned CurStageNum,
993                                                  unsigned InstStageNum) {
994   MachineInstr *NewMI = MF.CloneMachineInstr(OldMI);
995   updateMemOperands(*NewMI, *OldMI, CurStageNum - InstStageNum);
996   return NewMI;
997 }
998 
999 /// Clone the instruction for the new pipelined loop. If needed, this
1000 /// function updates the instruction using the values saved in the
1001 /// InstrChanges structure.
1002 MachineInstr *ModuloScheduleExpander::cloneAndChangeInstr(
1003     MachineInstr *OldMI, unsigned CurStageNum, unsigned InstStageNum) {
1004   MachineInstr *NewMI = MF.CloneMachineInstr(OldMI);
1005   auto It = InstrChanges.find(OldMI);
1006   if (It != InstrChanges.end()) {
1007     std::pair<unsigned, int64_t> RegAndOffset = It->second;
1008     unsigned BasePos, OffsetPos;
1009     if (!TII->getBaseAndOffsetPosition(*OldMI, BasePos, OffsetPos))
1010       return nullptr;
1011     int64_t NewOffset = OldMI->getOperand(OffsetPos).getImm();
1012     MachineInstr *LoopDef = findDefInLoop(RegAndOffset.first);
1013     if (Schedule.getStage(LoopDef) > (signed)InstStageNum)
1014       NewOffset += RegAndOffset.second * (CurStageNum - InstStageNum);
1015     NewMI->getOperand(OffsetPos).setImm(NewOffset);
1016   }
1017   updateMemOperands(*NewMI, *OldMI, CurStageNum - InstStageNum);
1018   return NewMI;
1019 }
1020 
1021 /// Update the machine instruction with new virtual registers.  This
1022 /// function may change the definitions and/or uses.
1023 void ModuloScheduleExpander::updateInstruction(MachineInstr *NewMI,
1024                                                bool LastDef,
1025                                                unsigned CurStageNum,
1026                                                unsigned InstrStageNum,
1027                                                ValueMapTy *VRMap) {
1028   for (MachineOperand &MO : NewMI->operands()) {
1029     if (!MO.isReg() || !MO.getReg().isVirtual())
1030       continue;
1031     Register reg = MO.getReg();
1032     if (MO.isDef()) {
1033       // Create a new virtual register for the definition.
1034       const TargetRegisterClass *RC = MRI.getRegClass(reg);
1035       Register NewReg = MRI.createVirtualRegister(RC);
1036       MO.setReg(NewReg);
1037       VRMap[CurStageNum][reg] = NewReg;
1038       if (LastDef)
1039         replaceRegUsesAfterLoop(reg, NewReg, BB, MRI, LIS);
1040     } else if (MO.isUse()) {
1041       MachineInstr *Def = MRI.getVRegDef(reg);
1042       // Compute the stage that contains the last definition for instruction.
1043       int DefStageNum = Schedule.getStage(Def);
1044       unsigned StageNum = CurStageNum;
1045       if (DefStageNum != -1 && (int)InstrStageNum > DefStageNum) {
1046         // Compute the difference in stages between the defintion and the use.
1047         unsigned StageDiff = (InstrStageNum - DefStageNum);
1048         // Make an adjustment to get the last definition.
1049         StageNum -= StageDiff;
1050       }
1051       if (VRMap[StageNum].count(reg))
1052         MO.setReg(VRMap[StageNum][reg]);
1053     }
1054   }
1055 }
1056 
1057 /// Return the instruction in the loop that defines the register.
1058 /// If the definition is a Phi, then follow the Phi operand to
1059 /// the instruction in the loop.
1060 MachineInstr *ModuloScheduleExpander::findDefInLoop(unsigned Reg) {
1061   SmallPtrSet<MachineInstr *, 8> Visited;
1062   MachineInstr *Def = MRI.getVRegDef(Reg);
1063   while (Def->isPHI()) {
1064     if (!Visited.insert(Def).second)
1065       break;
1066     for (unsigned i = 1, e = Def->getNumOperands(); i < e; i += 2)
1067       if (Def->getOperand(i + 1).getMBB() == BB) {
1068         Def = MRI.getVRegDef(Def->getOperand(i).getReg());
1069         break;
1070       }
1071   }
1072   return Def;
1073 }
1074 
1075 /// Return the new name for the value from the previous stage.
1076 unsigned ModuloScheduleExpander::getPrevMapVal(
1077     unsigned StageNum, unsigned PhiStage, unsigned LoopVal, unsigned LoopStage,
1078     ValueMapTy *VRMap, MachineBasicBlock *BB) {
1079   unsigned PrevVal = 0;
1080   if (StageNum > PhiStage) {
1081     MachineInstr *LoopInst = MRI.getVRegDef(LoopVal);
1082     if (PhiStage == LoopStage && VRMap[StageNum - 1].count(LoopVal))
1083       // The name is defined in the previous stage.
1084       PrevVal = VRMap[StageNum - 1][LoopVal];
1085     else if (VRMap[StageNum].count(LoopVal))
1086       // The previous name is defined in the current stage when the instruction
1087       // order is swapped.
1088       PrevVal = VRMap[StageNum][LoopVal];
1089     else if (!LoopInst->isPHI() || LoopInst->getParent() != BB)
1090       // The loop value hasn't yet been scheduled.
1091       PrevVal = LoopVal;
1092     else if (StageNum == PhiStage + 1)
1093       // The loop value is another phi, which has not been scheduled.
1094       PrevVal = getInitPhiReg(*LoopInst, BB);
1095     else if (StageNum > PhiStage + 1 && LoopInst->getParent() == BB)
1096       // The loop value is another phi, which has been scheduled.
1097       PrevVal =
1098           getPrevMapVal(StageNum - 1, PhiStage, getLoopPhiReg(*LoopInst, BB),
1099                         LoopStage, VRMap, BB);
1100   }
1101   return PrevVal;
1102 }
1103 
1104 /// Rewrite the Phi values in the specified block to use the mappings
1105 /// from the initial operand. Once the Phi is scheduled, we switch
1106 /// to using the loop value instead of the Phi value, so those names
1107 /// do not need to be rewritten.
1108 void ModuloScheduleExpander::rewritePhiValues(MachineBasicBlock *NewBB,
1109                                               unsigned StageNum,
1110                                               ValueMapTy *VRMap,
1111                                               InstrMapTy &InstrMap) {
1112   for (auto &PHI : BB->phis()) {
1113     unsigned InitVal = 0;
1114     unsigned LoopVal = 0;
1115     getPhiRegs(PHI, BB, InitVal, LoopVal);
1116     Register PhiDef = PHI.getOperand(0).getReg();
1117 
1118     unsigned PhiStage = (unsigned)Schedule.getStage(MRI.getVRegDef(PhiDef));
1119     unsigned LoopStage = (unsigned)Schedule.getStage(MRI.getVRegDef(LoopVal));
1120     unsigned NumPhis = getStagesForPhi(PhiDef);
1121     if (NumPhis > StageNum)
1122       NumPhis = StageNum;
1123     for (unsigned np = 0; np <= NumPhis; ++np) {
1124       unsigned NewVal =
1125           getPrevMapVal(StageNum - np, PhiStage, LoopVal, LoopStage, VRMap, BB);
1126       if (!NewVal)
1127         NewVal = InitVal;
1128       rewriteScheduledInstr(NewBB, InstrMap, StageNum - np, np, &PHI, PhiDef,
1129                             NewVal);
1130     }
1131   }
1132 }
1133 
1134 /// Rewrite a previously scheduled instruction to use the register value
1135 /// from the new instruction. Make sure the instruction occurs in the
1136 /// basic block, and we don't change the uses in the new instruction.
1137 void ModuloScheduleExpander::rewriteScheduledInstr(
1138     MachineBasicBlock *BB, InstrMapTy &InstrMap, unsigned CurStageNum,
1139     unsigned PhiNum, MachineInstr *Phi, unsigned OldReg, unsigned NewReg,
1140     unsigned PrevReg) {
1141   bool InProlog = (CurStageNum < (unsigned)Schedule.getNumStages() - 1);
1142   int StagePhi = Schedule.getStage(Phi) + PhiNum;
1143   // Rewrite uses that have been scheduled already to use the new
1144   // Phi register.
1145   for (MachineOperand &UseOp :
1146        llvm::make_early_inc_range(MRI.use_operands(OldReg))) {
1147     MachineInstr *UseMI = UseOp.getParent();
1148     if (UseMI->getParent() != BB)
1149       continue;
1150     if (UseMI->isPHI()) {
1151       if (!Phi->isPHI() && UseMI->getOperand(0).getReg() == NewReg)
1152         continue;
1153       if (getLoopPhiReg(*UseMI, BB) != OldReg)
1154         continue;
1155     }
1156     InstrMapTy::iterator OrigInstr = InstrMap.find(UseMI);
1157     assert(OrigInstr != InstrMap.end() && "Instruction not scheduled.");
1158     MachineInstr *OrigMI = OrigInstr->second;
1159     int StageSched = Schedule.getStage(OrigMI);
1160     int CycleSched = Schedule.getCycle(OrigMI);
1161     unsigned ReplaceReg = 0;
1162     // This is the stage for the scheduled instruction.
1163     if (StagePhi == StageSched && Phi->isPHI()) {
1164       int CyclePhi = Schedule.getCycle(Phi);
1165       if (PrevReg && InProlog)
1166         ReplaceReg = PrevReg;
1167       else if (PrevReg && !isLoopCarried(*Phi) &&
1168                (CyclePhi <= CycleSched || OrigMI->isPHI()))
1169         ReplaceReg = PrevReg;
1170       else
1171         ReplaceReg = NewReg;
1172     }
1173     // The scheduled instruction occurs before the scheduled Phi, and the
1174     // Phi is not loop carried.
1175     if (!InProlog && StagePhi + 1 == StageSched && !isLoopCarried(*Phi))
1176       ReplaceReg = NewReg;
1177     if (StagePhi > StageSched && Phi->isPHI())
1178       ReplaceReg = NewReg;
1179     if (!InProlog && !Phi->isPHI() && StagePhi < StageSched)
1180       ReplaceReg = NewReg;
1181     if (ReplaceReg) {
1182       const TargetRegisterClass *NRC =
1183           MRI.constrainRegClass(ReplaceReg, MRI.getRegClass(OldReg));
1184       if (NRC)
1185         UseOp.setReg(ReplaceReg);
1186       else {
1187         Register SplitReg = MRI.createVirtualRegister(MRI.getRegClass(OldReg));
1188         BuildMI(*BB, UseMI, UseMI->getDebugLoc(), TII->get(TargetOpcode::COPY),
1189                 SplitReg)
1190             .addReg(ReplaceReg);
1191         UseOp.setReg(SplitReg);
1192       }
1193     }
1194   }
1195 }
1196 
1197 bool ModuloScheduleExpander::isLoopCarried(MachineInstr &Phi) {
1198   if (!Phi.isPHI())
1199     return false;
1200   int DefCycle = Schedule.getCycle(&Phi);
1201   int DefStage = Schedule.getStage(&Phi);
1202 
1203   unsigned InitVal = 0;
1204   unsigned LoopVal = 0;
1205   getPhiRegs(Phi, Phi.getParent(), InitVal, LoopVal);
1206   MachineInstr *Use = MRI.getVRegDef(LoopVal);
1207   if (!Use || Use->isPHI())
1208     return true;
1209   int LoopCycle = Schedule.getCycle(Use);
1210   int LoopStage = Schedule.getStage(Use);
1211   return (LoopCycle > DefCycle) || (LoopStage <= DefStage);
1212 }
1213 
1214 //===----------------------------------------------------------------------===//
1215 // PeelingModuloScheduleExpander implementation
1216 //===----------------------------------------------------------------------===//
1217 // This is a reimplementation of ModuloScheduleExpander that works by creating
1218 // a fully correct steady-state kernel and peeling off the prolog and epilogs.
1219 //===----------------------------------------------------------------------===//
1220 
1221 namespace {
1222 // Remove any dead phis in MBB. Dead phis either have only one block as input
1223 // (in which case they are the identity) or have no uses.
1224 void EliminateDeadPhis(MachineBasicBlock *MBB, MachineRegisterInfo &MRI,
1225                        LiveIntervals *LIS, bool KeepSingleSrcPhi = false) {
1226   bool Changed = true;
1227   while (Changed) {
1228     Changed = false;
1229     for (MachineInstr &MI : llvm::make_early_inc_range(MBB->phis())) {
1230       assert(MI.isPHI());
1231       if (MRI.use_empty(MI.getOperand(0).getReg())) {
1232         if (LIS)
1233           LIS->RemoveMachineInstrFromMaps(MI);
1234         MI.eraseFromParent();
1235         Changed = true;
1236       } else if (!KeepSingleSrcPhi && MI.getNumExplicitOperands() == 3) {
1237         const TargetRegisterClass *ConstrainRegClass =
1238             MRI.constrainRegClass(MI.getOperand(1).getReg(),
1239                                   MRI.getRegClass(MI.getOperand(0).getReg()));
1240         assert(ConstrainRegClass &&
1241                "Expected a valid constrained register class!");
1242         (void)ConstrainRegClass;
1243         MRI.replaceRegWith(MI.getOperand(0).getReg(),
1244                            MI.getOperand(1).getReg());
1245         if (LIS)
1246           LIS->RemoveMachineInstrFromMaps(MI);
1247         MI.eraseFromParent();
1248         Changed = true;
1249       }
1250     }
1251   }
1252 }
1253 
1254 /// Rewrites the kernel block in-place to adhere to the given schedule.
1255 /// KernelRewriter holds all of the state required to perform the rewriting.
1256 class KernelRewriter {
1257   ModuloSchedule &S;
1258   MachineBasicBlock *BB;
1259   MachineBasicBlock *PreheaderBB, *ExitBB;
1260   MachineRegisterInfo &MRI;
1261   const TargetInstrInfo *TII;
1262   LiveIntervals *LIS;
1263 
1264   // Map from register class to canonical undef register for that class.
1265   DenseMap<const TargetRegisterClass *, Register> Undefs;
1266   // Map from <LoopReg, InitReg> to phi register for all created phis. Note that
1267   // this map is only used when InitReg is non-undef.
1268   DenseMap<std::pair<unsigned, unsigned>, Register> Phis;
1269   // Map from LoopReg to phi register where the InitReg is undef.
1270   DenseMap<Register, Register> UndefPhis;
1271 
1272   // Reg is used by MI. Return the new register MI should use to adhere to the
1273   // schedule. Insert phis as necessary.
1274   Register remapUse(Register Reg, MachineInstr &MI);
1275   // Insert a phi that carries LoopReg from the loop body and InitReg otherwise.
1276   // If InitReg is not given it is chosen arbitrarily. It will either be undef
1277   // or will be chosen so as to share another phi.
1278   Register phi(Register LoopReg, std::optional<Register> InitReg = {},
1279                const TargetRegisterClass *RC = nullptr);
1280   // Create an undef register of the given register class.
1281   Register undef(const TargetRegisterClass *RC);
1282 
1283 public:
1284   KernelRewriter(MachineLoop &L, ModuloSchedule &S, MachineBasicBlock *LoopBB,
1285                  LiveIntervals *LIS = nullptr);
1286   void rewrite();
1287 };
1288 } // namespace
1289 
1290 KernelRewriter::KernelRewriter(MachineLoop &L, ModuloSchedule &S,
1291                                MachineBasicBlock *LoopBB, LiveIntervals *LIS)
1292     : S(S), BB(LoopBB), PreheaderBB(L.getLoopPreheader()),
1293       ExitBB(L.getExitBlock()), MRI(BB->getParent()->getRegInfo()),
1294       TII(BB->getParent()->getSubtarget().getInstrInfo()), LIS(LIS) {
1295   PreheaderBB = *BB->pred_begin();
1296   if (PreheaderBB == BB)
1297     PreheaderBB = *std::next(BB->pred_begin());
1298 }
1299 
1300 void KernelRewriter::rewrite() {
1301   // Rearrange the loop to be in schedule order. Note that the schedule may
1302   // contain instructions that are not owned by the loop block (InstrChanges and
1303   // friends), so we gracefully handle unowned instructions and delete any
1304   // instructions that weren't in the schedule.
1305   auto InsertPt = BB->getFirstTerminator();
1306   MachineInstr *FirstMI = nullptr;
1307   for (MachineInstr *MI : S.getInstructions()) {
1308     if (MI->isPHI())
1309       continue;
1310     if (MI->getParent())
1311       MI->removeFromParent();
1312     BB->insert(InsertPt, MI);
1313     if (!FirstMI)
1314       FirstMI = MI;
1315   }
1316   assert(FirstMI && "Failed to find first MI in schedule");
1317 
1318   // At this point all of the scheduled instructions are between FirstMI
1319   // and the end of the block. Kill from the first non-phi to FirstMI.
1320   for (auto I = BB->getFirstNonPHI(); I != FirstMI->getIterator();) {
1321     if (LIS)
1322       LIS->RemoveMachineInstrFromMaps(*I);
1323     (I++)->eraseFromParent();
1324   }
1325 
1326   // Now remap every instruction in the loop.
1327   for (MachineInstr &MI : *BB) {
1328     if (MI.isPHI() || MI.isTerminator())
1329       continue;
1330     for (MachineOperand &MO : MI.uses()) {
1331       if (!MO.isReg() || MO.getReg().isPhysical() || MO.isImplicit())
1332         continue;
1333       Register Reg = remapUse(MO.getReg(), MI);
1334       MO.setReg(Reg);
1335     }
1336   }
1337   EliminateDeadPhis(BB, MRI, LIS);
1338 
1339   // Ensure a phi exists for all instructions that are either referenced by
1340   // an illegal phi or by an instruction outside the loop. This allows us to
1341   // treat remaps of these values the same as "normal" values that come from
1342   // loop-carried phis.
1343   for (auto MI = BB->getFirstNonPHI(); MI != BB->end(); ++MI) {
1344     if (MI->isPHI()) {
1345       Register R = MI->getOperand(0).getReg();
1346       phi(R);
1347       continue;
1348     }
1349 
1350     for (MachineOperand &Def : MI->defs()) {
1351       for (MachineInstr &MI : MRI.use_instructions(Def.getReg())) {
1352         if (MI.getParent() != BB) {
1353           phi(Def.getReg());
1354           break;
1355         }
1356       }
1357     }
1358   }
1359 }
1360 
1361 Register KernelRewriter::remapUse(Register Reg, MachineInstr &MI) {
1362   MachineInstr *Producer = MRI.getUniqueVRegDef(Reg);
1363   if (!Producer)
1364     return Reg;
1365 
1366   int ConsumerStage = S.getStage(&MI);
1367   if (!Producer->isPHI()) {
1368     // Non-phi producers are simple to remap. Insert as many phis as the
1369     // difference between the consumer and producer stages.
1370     if (Producer->getParent() != BB)
1371       // Producer was not inside the loop. Use the register as-is.
1372       return Reg;
1373     int ProducerStage = S.getStage(Producer);
1374     assert(ConsumerStage != -1 &&
1375            "In-loop consumer should always be scheduled!");
1376     assert(ConsumerStage >= ProducerStage);
1377     unsigned StageDiff = ConsumerStage - ProducerStage;
1378 
1379     for (unsigned I = 0; I < StageDiff; ++I)
1380       Reg = phi(Reg);
1381     return Reg;
1382   }
1383 
1384   // First, dive through the phi chain to find the defaults for the generated
1385   // phis.
1386   SmallVector<std::optional<Register>, 4> Defaults;
1387   Register LoopReg = Reg;
1388   auto LoopProducer = Producer;
1389   while (LoopProducer->isPHI() && LoopProducer->getParent() == BB) {
1390     LoopReg = getLoopPhiReg(*LoopProducer, BB);
1391     Defaults.emplace_back(getInitPhiReg(*LoopProducer, BB));
1392     LoopProducer = MRI.getUniqueVRegDef(LoopReg);
1393     assert(LoopProducer);
1394   }
1395   int LoopProducerStage = S.getStage(LoopProducer);
1396 
1397   std::optional<Register> IllegalPhiDefault;
1398 
1399   if (LoopProducerStage == -1) {
1400     // Do nothing.
1401   } else if (LoopProducerStage > ConsumerStage) {
1402     // This schedule is only representable if ProducerStage == ConsumerStage+1.
1403     // In addition, Consumer's cycle must be scheduled after Producer in the
1404     // rescheduled loop. This is enforced by the pipeliner's ASAP and ALAP
1405     // functions.
1406 #ifndef NDEBUG // Silence unused variables in non-asserts mode.
1407     int LoopProducerCycle = S.getCycle(LoopProducer);
1408     int ConsumerCycle = S.getCycle(&MI);
1409 #endif
1410     assert(LoopProducerCycle <= ConsumerCycle);
1411     assert(LoopProducerStage == ConsumerStage + 1);
1412     // Peel off the first phi from Defaults and insert a phi between producer
1413     // and consumer. This phi will not be at the front of the block so we
1414     // consider it illegal. It will only exist during the rewrite process; it
1415     // needs to exist while we peel off prologs because these could take the
1416     // default value. After that we can replace all uses with the loop producer
1417     // value.
1418     IllegalPhiDefault = Defaults.front();
1419     Defaults.erase(Defaults.begin());
1420   } else {
1421     assert(ConsumerStage >= LoopProducerStage);
1422     int StageDiff = ConsumerStage - LoopProducerStage;
1423     if (StageDiff > 0) {
1424       LLVM_DEBUG(dbgs() << " -- padding defaults array from " << Defaults.size()
1425                         << " to " << (Defaults.size() + StageDiff) << "\n");
1426       // If we need more phis than we have defaults for, pad out with undefs for
1427       // the earliest phis, which are at the end of the defaults chain (the
1428       // chain is in reverse order).
1429       Defaults.resize(Defaults.size() + StageDiff,
1430                       Defaults.empty() ? std::optional<Register>()
1431                                        : Defaults.back());
1432     }
1433   }
1434 
1435   // Now we know the number of stages to jump back, insert the phi chain.
1436   auto DefaultI = Defaults.rbegin();
1437   while (DefaultI != Defaults.rend())
1438     LoopReg = phi(LoopReg, *DefaultI++, MRI.getRegClass(Reg));
1439 
1440   if (IllegalPhiDefault) {
1441     // The consumer optionally consumes LoopProducer in the same iteration
1442     // (because the producer is scheduled at an earlier cycle than the consumer)
1443     // or the initial value. To facilitate this we create an illegal block here
1444     // by embedding a phi in the middle of the block. We will fix this up
1445     // immediately prior to pruning.
1446     auto RC = MRI.getRegClass(Reg);
1447     Register R = MRI.createVirtualRegister(RC);
1448     MachineInstr *IllegalPhi =
1449         BuildMI(*BB, MI, DebugLoc(), TII->get(TargetOpcode::PHI), R)
1450             .addReg(*IllegalPhiDefault)
1451             .addMBB(PreheaderBB) // Block choice is arbitrary and has no effect.
1452             .addReg(LoopReg)
1453             .addMBB(BB); // Block choice is arbitrary and has no effect.
1454     // Illegal phi should belong to the producer stage so that it can be
1455     // filtered correctly during peeling.
1456     S.setStage(IllegalPhi, LoopProducerStage);
1457     return R;
1458   }
1459 
1460   return LoopReg;
1461 }
1462 
1463 Register KernelRewriter::phi(Register LoopReg, std::optional<Register> InitReg,
1464                              const TargetRegisterClass *RC) {
1465   // If the init register is not undef, try and find an existing phi.
1466   if (InitReg) {
1467     auto I = Phis.find({LoopReg, *InitReg});
1468     if (I != Phis.end())
1469       return I->second;
1470   } else {
1471     for (auto &KV : Phis) {
1472       if (KV.first.first == LoopReg)
1473         return KV.second;
1474     }
1475   }
1476 
1477   // InitReg is either undef or no existing phi takes InitReg as input. Try and
1478   // find a phi that takes undef as input.
1479   auto I = UndefPhis.find(LoopReg);
1480   if (I != UndefPhis.end()) {
1481     Register R = I->second;
1482     if (!InitReg)
1483       // Found a phi taking undef as input, and this input is undef so return
1484       // without any more changes.
1485       return R;
1486     // Found a phi taking undef as input, so rewrite it to take InitReg.
1487     MachineInstr *MI = MRI.getVRegDef(R);
1488     MI->getOperand(1).setReg(*InitReg);
1489     Phis.insert({{LoopReg, *InitReg}, R});
1490     const TargetRegisterClass *ConstrainRegClass =
1491         MRI.constrainRegClass(R, MRI.getRegClass(*InitReg));
1492     assert(ConstrainRegClass && "Expected a valid constrained register class!");
1493     (void)ConstrainRegClass;
1494     UndefPhis.erase(I);
1495     return R;
1496   }
1497 
1498   // Failed to find any existing phi to reuse, so create a new one.
1499   if (!RC)
1500     RC = MRI.getRegClass(LoopReg);
1501   Register R = MRI.createVirtualRegister(RC);
1502   if (InitReg) {
1503     const TargetRegisterClass *ConstrainRegClass =
1504         MRI.constrainRegClass(R, MRI.getRegClass(*InitReg));
1505     assert(ConstrainRegClass && "Expected a valid constrained register class!");
1506     (void)ConstrainRegClass;
1507   }
1508   BuildMI(*BB, BB->getFirstNonPHI(), DebugLoc(), TII->get(TargetOpcode::PHI), R)
1509       .addReg(InitReg ? *InitReg : undef(RC))
1510       .addMBB(PreheaderBB)
1511       .addReg(LoopReg)
1512       .addMBB(BB);
1513   if (!InitReg)
1514     UndefPhis[LoopReg] = R;
1515   else
1516     Phis[{LoopReg, *InitReg}] = R;
1517   return R;
1518 }
1519 
1520 Register KernelRewriter::undef(const TargetRegisterClass *RC) {
1521   Register &R = Undefs[RC];
1522   if (R == 0) {
1523     // Create an IMPLICIT_DEF that defines this register if we need it.
1524     // All uses of this should be removed by the time we have finished unrolling
1525     // prologs and epilogs.
1526     R = MRI.createVirtualRegister(RC);
1527     auto *InsertBB = &PreheaderBB->getParent()->front();
1528     BuildMI(*InsertBB, InsertBB->getFirstTerminator(), DebugLoc(),
1529             TII->get(TargetOpcode::IMPLICIT_DEF), R);
1530   }
1531   return R;
1532 }
1533 
1534 namespace {
1535 /// Describes an operand in the kernel of a pipelined loop. Characteristics of
1536 /// the operand are discovered, such as how many in-loop PHIs it has to jump
1537 /// through and defaults for these phis.
1538 class KernelOperandInfo {
1539   MachineBasicBlock *BB;
1540   MachineRegisterInfo &MRI;
1541   SmallVector<Register, 4> PhiDefaults;
1542   MachineOperand *Source;
1543   MachineOperand *Target;
1544 
1545 public:
1546   KernelOperandInfo(MachineOperand *MO, MachineRegisterInfo &MRI,
1547                     const SmallPtrSetImpl<MachineInstr *> &IllegalPhis)
1548       : MRI(MRI) {
1549     Source = MO;
1550     BB = MO->getParent()->getParent();
1551     while (isRegInLoop(MO)) {
1552       MachineInstr *MI = MRI.getVRegDef(MO->getReg());
1553       if (MI->isFullCopy()) {
1554         MO = &MI->getOperand(1);
1555         continue;
1556       }
1557       if (!MI->isPHI())
1558         break;
1559       // If this is an illegal phi, don't count it in distance.
1560       if (IllegalPhis.count(MI)) {
1561         MO = &MI->getOperand(3);
1562         continue;
1563       }
1564 
1565       Register Default = getInitPhiReg(*MI, BB);
1566       MO = MI->getOperand(2).getMBB() == BB ? &MI->getOperand(1)
1567                                             : &MI->getOperand(3);
1568       PhiDefaults.push_back(Default);
1569     }
1570     Target = MO;
1571   }
1572 
1573   bool operator==(const KernelOperandInfo &Other) const {
1574     return PhiDefaults.size() == Other.PhiDefaults.size();
1575   }
1576 
1577   void print(raw_ostream &OS) const {
1578     OS << "use of " << *Source << ": distance(" << PhiDefaults.size() << ") in "
1579        << *Source->getParent();
1580   }
1581 
1582 private:
1583   bool isRegInLoop(MachineOperand *MO) {
1584     return MO->isReg() && MO->getReg().isVirtual() &&
1585            MRI.getVRegDef(MO->getReg())->getParent() == BB;
1586   }
1587 };
1588 } // namespace
1589 
1590 MachineBasicBlock *
1591 PeelingModuloScheduleExpander::peelKernel(LoopPeelDirection LPD) {
1592   MachineBasicBlock *NewBB = PeelSingleBlockLoop(LPD, BB, MRI, TII);
1593   if (LPD == LPD_Front)
1594     PeeledFront.push_back(NewBB);
1595   else
1596     PeeledBack.push_front(NewBB);
1597   for (auto I = BB->begin(), NI = NewBB->begin(); !I->isTerminator();
1598        ++I, ++NI) {
1599     CanonicalMIs[&*I] = &*I;
1600     CanonicalMIs[&*NI] = &*I;
1601     BlockMIs[{NewBB, &*I}] = &*NI;
1602     BlockMIs[{BB, &*I}] = &*I;
1603   }
1604   return NewBB;
1605 }
1606 
1607 void PeelingModuloScheduleExpander::filterInstructions(MachineBasicBlock *MB,
1608                                                        int MinStage) {
1609   for (auto I = MB->getFirstInstrTerminator()->getReverseIterator();
1610        I != std::next(MB->getFirstNonPHI()->getReverseIterator());) {
1611     MachineInstr *MI = &*I++;
1612     int Stage = getStage(MI);
1613     if (Stage == -1 || Stage >= MinStage)
1614       continue;
1615 
1616     for (MachineOperand &DefMO : MI->defs()) {
1617       SmallVector<std::pair<MachineInstr *, Register>, 4> Subs;
1618       for (MachineInstr &UseMI : MRI.use_instructions(DefMO.getReg())) {
1619         // Only PHIs can use values from this block by construction.
1620         // Match with the equivalent PHI in B.
1621         assert(UseMI.isPHI());
1622         Register Reg = getEquivalentRegisterIn(UseMI.getOperand(0).getReg(),
1623                                                MI->getParent());
1624         Subs.emplace_back(&UseMI, Reg);
1625       }
1626       for (auto &Sub : Subs)
1627         Sub.first->substituteRegister(DefMO.getReg(), Sub.second, /*SubIdx=*/0,
1628                                       *MRI.getTargetRegisterInfo());
1629     }
1630     if (LIS)
1631       LIS->RemoveMachineInstrFromMaps(*MI);
1632     MI->eraseFromParent();
1633   }
1634 }
1635 
1636 void PeelingModuloScheduleExpander::moveStageBetweenBlocks(
1637     MachineBasicBlock *DestBB, MachineBasicBlock *SourceBB, unsigned Stage) {
1638   auto InsertPt = DestBB->getFirstNonPHI();
1639   DenseMap<Register, Register> Remaps;
1640   for (MachineInstr &MI : llvm::make_early_inc_range(
1641            llvm::make_range(SourceBB->getFirstNonPHI(), SourceBB->end()))) {
1642     if (MI.isPHI()) {
1643       // This is an illegal PHI. If we move any instructions using an illegal
1644       // PHI, we need to create a legal Phi.
1645       if (getStage(&MI) != Stage) {
1646         // The legal Phi is not necessary if the illegal phi's stage
1647         // is being moved.
1648         Register PhiR = MI.getOperand(0).getReg();
1649         auto RC = MRI.getRegClass(PhiR);
1650         Register NR = MRI.createVirtualRegister(RC);
1651         MachineInstr *NI = BuildMI(*DestBB, DestBB->getFirstNonPHI(),
1652                                    DebugLoc(), TII->get(TargetOpcode::PHI), NR)
1653                                .addReg(PhiR)
1654                                .addMBB(SourceBB);
1655         BlockMIs[{DestBB, CanonicalMIs[&MI]}] = NI;
1656         CanonicalMIs[NI] = CanonicalMIs[&MI];
1657         Remaps[PhiR] = NR;
1658       }
1659     }
1660     if (getStage(&MI) != Stage)
1661       continue;
1662     MI.removeFromParent();
1663     DestBB->insert(InsertPt, &MI);
1664     auto *KernelMI = CanonicalMIs[&MI];
1665     BlockMIs[{DestBB, KernelMI}] = &MI;
1666     BlockMIs.erase({SourceBB, KernelMI});
1667   }
1668   SmallVector<MachineInstr *, 4> PhiToDelete;
1669   for (MachineInstr &MI : DestBB->phis()) {
1670     assert(MI.getNumOperands() == 3);
1671     MachineInstr *Def = MRI.getVRegDef(MI.getOperand(1).getReg());
1672     // If the instruction referenced by the phi is moved inside the block
1673     // we don't need the phi anymore.
1674     if (getStage(Def) == Stage) {
1675       Register PhiReg = MI.getOperand(0).getReg();
1676       assert(Def->findRegisterDefOperandIdx(MI.getOperand(1).getReg()) != -1);
1677       MRI.replaceRegWith(MI.getOperand(0).getReg(), MI.getOperand(1).getReg());
1678       MI.getOperand(0).setReg(PhiReg);
1679       PhiToDelete.push_back(&MI);
1680     }
1681   }
1682   for (auto *P : PhiToDelete)
1683     P->eraseFromParent();
1684   InsertPt = DestBB->getFirstNonPHI();
1685   // Helper to clone Phi instructions into the destination block. We clone Phi
1686   // greedily to avoid combinatorial explosion of Phi instructions.
1687   auto clonePhi = [&](MachineInstr *Phi) {
1688     MachineInstr *NewMI = MF.CloneMachineInstr(Phi);
1689     DestBB->insert(InsertPt, NewMI);
1690     Register OrigR = Phi->getOperand(0).getReg();
1691     Register R = MRI.createVirtualRegister(MRI.getRegClass(OrigR));
1692     NewMI->getOperand(0).setReg(R);
1693     NewMI->getOperand(1).setReg(OrigR);
1694     NewMI->getOperand(2).setMBB(*DestBB->pred_begin());
1695     Remaps[OrigR] = R;
1696     CanonicalMIs[NewMI] = CanonicalMIs[Phi];
1697     BlockMIs[{DestBB, CanonicalMIs[Phi]}] = NewMI;
1698     PhiNodeLoopIteration[NewMI] = PhiNodeLoopIteration[Phi];
1699     return R;
1700   };
1701   for (auto I = DestBB->getFirstNonPHI(); I != DestBB->end(); ++I) {
1702     for (MachineOperand &MO : I->uses()) {
1703       if (!MO.isReg())
1704         continue;
1705       if (Remaps.count(MO.getReg()))
1706         MO.setReg(Remaps[MO.getReg()]);
1707       else {
1708         // If we are using a phi from the source block we need to add a new phi
1709         // pointing to the old one.
1710         MachineInstr *Use = MRI.getUniqueVRegDef(MO.getReg());
1711         if (Use && Use->isPHI() && Use->getParent() == SourceBB) {
1712           Register R = clonePhi(Use);
1713           MO.setReg(R);
1714         }
1715       }
1716     }
1717   }
1718 }
1719 
1720 Register
1721 PeelingModuloScheduleExpander::getPhiCanonicalReg(MachineInstr *CanonicalPhi,
1722                                                   MachineInstr *Phi) {
1723   unsigned distance = PhiNodeLoopIteration[Phi];
1724   MachineInstr *CanonicalUse = CanonicalPhi;
1725   Register CanonicalUseReg = CanonicalUse->getOperand(0).getReg();
1726   for (unsigned I = 0; I < distance; ++I) {
1727     assert(CanonicalUse->isPHI());
1728     assert(CanonicalUse->getNumOperands() == 5);
1729     unsigned LoopRegIdx = 3, InitRegIdx = 1;
1730     if (CanonicalUse->getOperand(2).getMBB() == CanonicalUse->getParent())
1731       std::swap(LoopRegIdx, InitRegIdx);
1732     CanonicalUseReg = CanonicalUse->getOperand(LoopRegIdx).getReg();
1733     CanonicalUse = MRI.getVRegDef(CanonicalUseReg);
1734   }
1735   return CanonicalUseReg;
1736 }
1737 
1738 void PeelingModuloScheduleExpander::peelPrologAndEpilogs() {
1739   BitVector LS(Schedule.getNumStages(), true);
1740   BitVector AS(Schedule.getNumStages(), true);
1741   LiveStages[BB] = LS;
1742   AvailableStages[BB] = AS;
1743 
1744   // Peel out the prologs.
1745   LS.reset();
1746   for (int I = 0; I < Schedule.getNumStages() - 1; ++I) {
1747     LS[I] = true;
1748     Prologs.push_back(peelKernel(LPD_Front));
1749     LiveStages[Prologs.back()] = LS;
1750     AvailableStages[Prologs.back()] = LS;
1751   }
1752 
1753   // Create a block that will end up as the new loop exiting block (dominated by
1754   // all prologs and epilogs). It will only contain PHIs, in the same order as
1755   // BB's PHIs. This gives us a poor-man's LCSSA with the inductive property
1756   // that the exiting block is a (sub) clone of BB. This in turn gives us the
1757   // property that any value deffed in BB but used outside of BB is used by a
1758   // PHI in the exiting block.
1759   MachineBasicBlock *ExitingBB = CreateLCSSAExitingBlock();
1760   EliminateDeadPhis(ExitingBB, MRI, LIS, /*KeepSingleSrcPhi=*/true);
1761   // Push out the epilogs, again in reverse order.
1762   // We can't assume anything about the minumum loop trip count at this point,
1763   // so emit a fairly complex epilog.
1764 
1765   // We first peel number of stages minus one epilogue. Then we remove dead
1766   // stages and reorder instructions based on their stage. If we have 3 stages
1767   // we generate first:
1768   // E0[3, 2, 1]
1769   // E1[3', 2']
1770   // E2[3'']
1771   // And then we move instructions based on their stages to have:
1772   // E0[3]
1773   // E1[2, 3']
1774   // E2[1, 2', 3'']
1775   // The transformation is legal because we only move instructions past
1776   // instructions of a previous loop iteration.
1777   for (int I = 1; I <= Schedule.getNumStages() - 1; ++I) {
1778     Epilogs.push_back(peelKernel(LPD_Back));
1779     MachineBasicBlock *B = Epilogs.back();
1780     filterInstructions(B, Schedule.getNumStages() - I);
1781     // Keep track at which iteration each phi belongs to. We need it to know
1782     // what version of the variable to use during prologue/epilogue stitching.
1783     EliminateDeadPhis(B, MRI, LIS, /*KeepSingleSrcPhi=*/true);
1784     for (MachineInstr &Phi : B->phis())
1785       PhiNodeLoopIteration[&Phi] = Schedule.getNumStages() - I;
1786   }
1787   for (size_t I = 0; I < Epilogs.size(); I++) {
1788     LS.reset();
1789     for (size_t J = I; J < Epilogs.size(); J++) {
1790       int Iteration = J;
1791       unsigned Stage = Schedule.getNumStages() - 1 + I - J;
1792       // Move stage one block at a time so that Phi nodes are updated correctly.
1793       for (size_t K = Iteration; K > I; K--)
1794         moveStageBetweenBlocks(Epilogs[K - 1], Epilogs[K], Stage);
1795       LS[Stage] = true;
1796     }
1797     LiveStages[Epilogs[I]] = LS;
1798     AvailableStages[Epilogs[I]] = AS;
1799   }
1800 
1801   // Now we've defined all the prolog and epilog blocks as a fallthrough
1802   // sequence, add the edges that will be followed if the loop trip count is
1803   // lower than the number of stages (connecting prologs directly with epilogs).
1804   auto PI = Prologs.begin();
1805   auto EI = Epilogs.begin();
1806   assert(Prologs.size() == Epilogs.size());
1807   for (; PI != Prologs.end(); ++PI, ++EI) {
1808     MachineBasicBlock *Pred = *(*EI)->pred_begin();
1809     (*PI)->addSuccessor(*EI);
1810     for (MachineInstr &MI : (*EI)->phis()) {
1811       Register Reg = MI.getOperand(1).getReg();
1812       MachineInstr *Use = MRI.getUniqueVRegDef(Reg);
1813       if (Use && Use->getParent() == Pred) {
1814         MachineInstr *CanonicalUse = CanonicalMIs[Use];
1815         if (CanonicalUse->isPHI()) {
1816           // If the use comes from a phi we need to skip as many phi as the
1817           // distance between the epilogue and the kernel. Trace through the phi
1818           // chain to find the right value.
1819           Reg = getPhiCanonicalReg(CanonicalUse, Use);
1820         }
1821         Reg = getEquivalentRegisterIn(Reg, *PI);
1822       }
1823       MI.addOperand(MachineOperand::CreateReg(Reg, /*isDef=*/false));
1824       MI.addOperand(MachineOperand::CreateMBB(*PI));
1825     }
1826   }
1827 
1828   // Create a list of all blocks in order.
1829   SmallVector<MachineBasicBlock *, 8> Blocks;
1830   llvm::copy(PeeledFront, std::back_inserter(Blocks));
1831   Blocks.push_back(BB);
1832   llvm::copy(PeeledBack, std::back_inserter(Blocks));
1833 
1834   // Iterate in reverse order over all instructions, remapping as we go.
1835   for (MachineBasicBlock *B : reverse(Blocks)) {
1836     for (auto I = B->instr_rbegin();
1837          I != std::next(B->getFirstNonPHI()->getReverseIterator());) {
1838       MachineBasicBlock::reverse_instr_iterator MI = I++;
1839       rewriteUsesOf(&*MI);
1840     }
1841   }
1842   for (auto *MI : IllegalPhisToDelete) {
1843     if (LIS)
1844       LIS->RemoveMachineInstrFromMaps(*MI);
1845     MI->eraseFromParent();
1846   }
1847   IllegalPhisToDelete.clear();
1848 
1849   // Now all remapping has been done, we're free to optimize the generated code.
1850   for (MachineBasicBlock *B : reverse(Blocks))
1851     EliminateDeadPhis(B, MRI, LIS);
1852   EliminateDeadPhis(ExitingBB, MRI, LIS);
1853 }
1854 
1855 MachineBasicBlock *PeelingModuloScheduleExpander::CreateLCSSAExitingBlock() {
1856   MachineFunction &MF = *BB->getParent();
1857   MachineBasicBlock *Exit = *BB->succ_begin();
1858   if (Exit == BB)
1859     Exit = *std::next(BB->succ_begin());
1860 
1861   MachineBasicBlock *NewBB = MF.CreateMachineBasicBlock(BB->getBasicBlock());
1862   MF.insert(std::next(BB->getIterator()), NewBB);
1863 
1864   // Clone all phis in BB into NewBB and rewrite.
1865   for (MachineInstr &MI : BB->phis()) {
1866     auto RC = MRI.getRegClass(MI.getOperand(0).getReg());
1867     Register OldR = MI.getOperand(3).getReg();
1868     Register R = MRI.createVirtualRegister(RC);
1869     SmallVector<MachineInstr *, 4> Uses;
1870     for (MachineInstr &Use : MRI.use_instructions(OldR))
1871       if (Use.getParent() != BB)
1872         Uses.push_back(&Use);
1873     for (MachineInstr *Use : Uses)
1874       Use->substituteRegister(OldR, R, /*SubIdx=*/0,
1875                               *MRI.getTargetRegisterInfo());
1876     MachineInstr *NI = BuildMI(NewBB, DebugLoc(), TII->get(TargetOpcode::PHI), R)
1877         .addReg(OldR)
1878         .addMBB(BB);
1879     BlockMIs[{NewBB, &MI}] = NI;
1880     CanonicalMIs[NI] = &MI;
1881   }
1882   BB->replaceSuccessor(Exit, NewBB);
1883   Exit->replacePhiUsesWith(BB, NewBB);
1884   NewBB->addSuccessor(Exit);
1885 
1886   MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
1887   SmallVector<MachineOperand, 4> Cond;
1888   bool CanAnalyzeBr = !TII->analyzeBranch(*BB, TBB, FBB, Cond);
1889   (void)CanAnalyzeBr;
1890   assert(CanAnalyzeBr && "Must be able to analyze the loop branch!");
1891   TII->removeBranch(*BB);
1892   TII->insertBranch(*BB, TBB == Exit ? NewBB : TBB, FBB == Exit ? NewBB : FBB,
1893                     Cond, DebugLoc());
1894   TII->insertUnconditionalBranch(*NewBB, Exit, DebugLoc());
1895   return NewBB;
1896 }
1897 
1898 Register
1899 PeelingModuloScheduleExpander::getEquivalentRegisterIn(Register Reg,
1900                                                        MachineBasicBlock *BB) {
1901   MachineInstr *MI = MRI.getUniqueVRegDef(Reg);
1902   unsigned OpIdx = MI->findRegisterDefOperandIdx(Reg);
1903   return BlockMIs[{BB, CanonicalMIs[MI]}]->getOperand(OpIdx).getReg();
1904 }
1905 
1906 void PeelingModuloScheduleExpander::rewriteUsesOf(MachineInstr *MI) {
1907   if (MI->isPHI()) {
1908     // This is an illegal PHI. The loop-carried (desired) value is operand 3,
1909     // and it is produced by this block.
1910     Register PhiR = MI->getOperand(0).getReg();
1911     Register R = MI->getOperand(3).getReg();
1912     int RMIStage = getStage(MRI.getUniqueVRegDef(R));
1913     if (RMIStage != -1 && !AvailableStages[MI->getParent()].test(RMIStage))
1914       R = MI->getOperand(1).getReg();
1915     MRI.setRegClass(R, MRI.getRegClass(PhiR));
1916     MRI.replaceRegWith(PhiR, R);
1917     // Postpone deleting the Phi as it may be referenced by BlockMIs and used
1918     // later to figure out how to remap registers.
1919     MI->getOperand(0).setReg(PhiR);
1920     IllegalPhisToDelete.push_back(MI);
1921     return;
1922   }
1923 
1924   int Stage = getStage(MI);
1925   if (Stage == -1 || LiveStages.count(MI->getParent()) == 0 ||
1926       LiveStages[MI->getParent()].test(Stage))
1927     // Instruction is live, no rewriting to do.
1928     return;
1929 
1930   for (MachineOperand &DefMO : MI->defs()) {
1931     SmallVector<std::pair<MachineInstr *, Register>, 4> Subs;
1932     for (MachineInstr &UseMI : MRI.use_instructions(DefMO.getReg())) {
1933       // Only PHIs can use values from this block by construction.
1934       // Match with the equivalent PHI in B.
1935       assert(UseMI.isPHI());
1936       Register Reg = getEquivalentRegisterIn(UseMI.getOperand(0).getReg(),
1937                                              MI->getParent());
1938       Subs.emplace_back(&UseMI, Reg);
1939     }
1940     for (auto &Sub : Subs)
1941       Sub.first->substituteRegister(DefMO.getReg(), Sub.second, /*SubIdx=*/0,
1942                                     *MRI.getTargetRegisterInfo());
1943   }
1944   if (LIS)
1945     LIS->RemoveMachineInstrFromMaps(*MI);
1946   MI->eraseFromParent();
1947 }
1948 
1949 void PeelingModuloScheduleExpander::fixupBranches() {
1950   // Work outwards from the kernel.
1951   bool KernelDisposed = false;
1952   int TC = Schedule.getNumStages() - 1;
1953   for (auto PI = Prologs.rbegin(), EI = Epilogs.rbegin(); PI != Prologs.rend();
1954        ++PI, ++EI, --TC) {
1955     MachineBasicBlock *Prolog = *PI;
1956     MachineBasicBlock *Fallthrough = *Prolog->succ_begin();
1957     MachineBasicBlock *Epilog = *EI;
1958     SmallVector<MachineOperand, 4> Cond;
1959     TII->removeBranch(*Prolog);
1960     std::optional<bool> StaticallyGreater =
1961         LoopInfo->createTripCountGreaterCondition(TC, *Prolog, Cond);
1962     if (!StaticallyGreater) {
1963       LLVM_DEBUG(dbgs() << "Dynamic: TC > " << TC << "\n");
1964       // Dynamically branch based on Cond.
1965       TII->insertBranch(*Prolog, Epilog, Fallthrough, Cond, DebugLoc());
1966     } else if (*StaticallyGreater == false) {
1967       LLVM_DEBUG(dbgs() << "Static-false: TC > " << TC << "\n");
1968       // Prolog never falls through; branch to epilog and orphan interior
1969       // blocks. Leave it to unreachable-block-elim to clean up.
1970       Prolog->removeSuccessor(Fallthrough);
1971       for (MachineInstr &P : Fallthrough->phis()) {
1972         P.removeOperand(2);
1973         P.removeOperand(1);
1974       }
1975       TII->insertUnconditionalBranch(*Prolog, Epilog, DebugLoc());
1976       KernelDisposed = true;
1977     } else {
1978       LLVM_DEBUG(dbgs() << "Static-true: TC > " << TC << "\n");
1979       // Prolog always falls through; remove incoming values in epilog.
1980       Prolog->removeSuccessor(Epilog);
1981       for (MachineInstr &P : Epilog->phis()) {
1982         P.removeOperand(4);
1983         P.removeOperand(3);
1984       }
1985     }
1986   }
1987 
1988   if (!KernelDisposed) {
1989     LoopInfo->adjustTripCount(-(Schedule.getNumStages() - 1));
1990     LoopInfo->setPreheader(Prologs.back());
1991   } else {
1992     LoopInfo->disposed();
1993   }
1994 }
1995 
1996 void PeelingModuloScheduleExpander::rewriteKernel() {
1997   KernelRewriter KR(*Schedule.getLoop(), Schedule, BB);
1998   KR.rewrite();
1999 }
2000 
2001 void PeelingModuloScheduleExpander::expand() {
2002   BB = Schedule.getLoop()->getTopBlock();
2003   Preheader = Schedule.getLoop()->getLoopPreheader();
2004   LLVM_DEBUG(Schedule.dump());
2005   LoopInfo = TII->analyzeLoopForPipelining(BB);
2006   assert(LoopInfo);
2007 
2008   rewriteKernel();
2009   peelPrologAndEpilogs();
2010   fixupBranches();
2011 }
2012 
2013 void PeelingModuloScheduleExpander::validateAgainstModuloScheduleExpander() {
2014   BB = Schedule.getLoop()->getTopBlock();
2015   Preheader = Schedule.getLoop()->getLoopPreheader();
2016 
2017   // Dump the schedule before we invalidate and remap all its instructions.
2018   // Stash it in a string so we can print it if we found an error.
2019   std::string ScheduleDump;
2020   raw_string_ostream OS(ScheduleDump);
2021   Schedule.print(OS);
2022   OS.flush();
2023 
2024   // First, run the normal ModuleScheduleExpander. We don't support any
2025   // InstrChanges.
2026   assert(LIS && "Requires LiveIntervals!");
2027   ModuloScheduleExpander MSE(MF, Schedule, *LIS,
2028                              ModuloScheduleExpander::InstrChangesTy());
2029   MSE.expand();
2030   MachineBasicBlock *ExpandedKernel = MSE.getRewrittenKernel();
2031   if (!ExpandedKernel) {
2032     // The expander optimized away the kernel. We can't do any useful checking.
2033     MSE.cleanup();
2034     return;
2035   }
2036   // Before running the KernelRewriter, re-add BB into the CFG.
2037   Preheader->addSuccessor(BB);
2038 
2039   // Now run the new expansion algorithm.
2040   KernelRewriter KR(*Schedule.getLoop(), Schedule, BB);
2041   KR.rewrite();
2042   peelPrologAndEpilogs();
2043 
2044   // Collect all illegal phis that the new algorithm created. We'll give these
2045   // to KernelOperandInfo.
2046   SmallPtrSet<MachineInstr *, 4> IllegalPhis;
2047   for (auto NI = BB->getFirstNonPHI(); NI != BB->end(); ++NI) {
2048     if (NI->isPHI())
2049       IllegalPhis.insert(&*NI);
2050   }
2051 
2052   // Co-iterate across both kernels. We expect them to be identical apart from
2053   // phis and full COPYs (we look through both).
2054   SmallVector<std::pair<KernelOperandInfo, KernelOperandInfo>, 8> KOIs;
2055   auto OI = ExpandedKernel->begin();
2056   auto NI = BB->begin();
2057   for (; !OI->isTerminator() && !NI->isTerminator(); ++OI, ++NI) {
2058     while (OI->isPHI() || OI->isFullCopy())
2059       ++OI;
2060     while (NI->isPHI() || NI->isFullCopy())
2061       ++NI;
2062     assert(OI->getOpcode() == NI->getOpcode() && "Opcodes don't match?!");
2063     // Analyze every operand separately.
2064     for (auto OOpI = OI->operands_begin(), NOpI = NI->operands_begin();
2065          OOpI != OI->operands_end(); ++OOpI, ++NOpI)
2066       KOIs.emplace_back(KernelOperandInfo(&*OOpI, MRI, IllegalPhis),
2067                         KernelOperandInfo(&*NOpI, MRI, IllegalPhis));
2068   }
2069 
2070   bool Failed = false;
2071   for (auto &OldAndNew : KOIs) {
2072     if (OldAndNew.first == OldAndNew.second)
2073       continue;
2074     Failed = true;
2075     errs() << "Modulo kernel validation error: [\n";
2076     errs() << " [golden] ";
2077     OldAndNew.first.print(errs());
2078     errs() << "          ";
2079     OldAndNew.second.print(errs());
2080     errs() << "]\n";
2081   }
2082 
2083   if (Failed) {
2084     errs() << "Golden reference kernel:\n";
2085     ExpandedKernel->print(errs());
2086     errs() << "New kernel:\n";
2087     BB->print(errs());
2088     errs() << ScheduleDump;
2089     report_fatal_error(
2090         "Modulo kernel validation (-pipeliner-experimental-cg) failed");
2091   }
2092 
2093   // Cleanup by removing BB from the CFG again as the original
2094   // ModuloScheduleExpander intended.
2095   Preheader->removeSuccessor(BB);
2096   MSE.cleanup();
2097 }
2098 
2099 //===----------------------------------------------------------------------===//
2100 // ModuloScheduleTestPass implementation
2101 //===----------------------------------------------------------------------===//
2102 // This pass constructs a ModuloSchedule from its module and runs
2103 // ModuloScheduleExpander.
2104 //
2105 // The module is expected to contain a single-block analyzable loop.
2106 // The total order of instructions is taken from the loop as-is.
2107 // Instructions are expected to be annotated with a PostInstrSymbol.
2108 // This PostInstrSymbol must have the following format:
2109 //  "Stage=%d Cycle=%d".
2110 //===----------------------------------------------------------------------===//
2111 
2112 namespace {
2113 class ModuloScheduleTest : public MachineFunctionPass {
2114 public:
2115   static char ID;
2116 
2117   ModuloScheduleTest() : MachineFunctionPass(ID) {
2118     initializeModuloScheduleTestPass(*PassRegistry::getPassRegistry());
2119   }
2120 
2121   bool runOnMachineFunction(MachineFunction &MF) override;
2122   void runOnLoop(MachineFunction &MF, MachineLoop &L);
2123 
2124   void getAnalysisUsage(AnalysisUsage &AU) const override {
2125     AU.addRequired<MachineLoopInfo>();
2126     AU.addRequired<LiveIntervals>();
2127     MachineFunctionPass::getAnalysisUsage(AU);
2128   }
2129 };
2130 } // namespace
2131 
2132 char ModuloScheduleTest::ID = 0;
2133 
2134 INITIALIZE_PASS_BEGIN(ModuloScheduleTest, "modulo-schedule-test",
2135                       "Modulo Schedule test pass", false, false)
2136 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
2137 INITIALIZE_PASS_DEPENDENCY(LiveIntervals)
2138 INITIALIZE_PASS_END(ModuloScheduleTest, "modulo-schedule-test",
2139                     "Modulo Schedule test pass", false, false)
2140 
2141 bool ModuloScheduleTest::runOnMachineFunction(MachineFunction &MF) {
2142   MachineLoopInfo &MLI = getAnalysis<MachineLoopInfo>();
2143   for (auto *L : MLI) {
2144     if (L->getTopBlock() != L->getBottomBlock())
2145       continue;
2146     runOnLoop(MF, *L);
2147     return false;
2148   }
2149   return false;
2150 }
2151 
2152 static void parseSymbolString(StringRef S, int &Cycle, int &Stage) {
2153   std::pair<StringRef, StringRef> StageAndCycle = getToken(S, "_");
2154   std::pair<StringRef, StringRef> StageTokenAndValue =
2155       getToken(StageAndCycle.first, "-");
2156   std::pair<StringRef, StringRef> CycleTokenAndValue =
2157       getToken(StageAndCycle.second, "-");
2158   if (StageTokenAndValue.first != "Stage" ||
2159       CycleTokenAndValue.first != "_Cycle") {
2160     llvm_unreachable(
2161         "Bad post-instr symbol syntax: see comment in ModuloScheduleTest");
2162     return;
2163   }
2164 
2165   StageTokenAndValue.second.drop_front().getAsInteger(10, Stage);
2166   CycleTokenAndValue.second.drop_front().getAsInteger(10, Cycle);
2167 
2168   dbgs() << "  Stage=" << Stage << ", Cycle=" << Cycle << "\n";
2169 }
2170 
2171 void ModuloScheduleTest::runOnLoop(MachineFunction &MF, MachineLoop &L) {
2172   LiveIntervals &LIS = getAnalysis<LiveIntervals>();
2173   MachineBasicBlock *BB = L.getTopBlock();
2174   dbgs() << "--- ModuloScheduleTest running on BB#" << BB->getNumber() << "\n";
2175 
2176   DenseMap<MachineInstr *, int> Cycle, Stage;
2177   std::vector<MachineInstr *> Instrs;
2178   for (MachineInstr &MI : *BB) {
2179     if (MI.isTerminator())
2180       continue;
2181     Instrs.push_back(&MI);
2182     if (MCSymbol *Sym = MI.getPostInstrSymbol()) {
2183       dbgs() << "Parsing post-instr symbol for " << MI;
2184       parseSymbolString(Sym->getName(), Cycle[&MI], Stage[&MI]);
2185     }
2186   }
2187 
2188   ModuloSchedule MS(MF, &L, std::move(Instrs), std::move(Cycle),
2189                     std::move(Stage));
2190   ModuloScheduleExpander MSE(
2191       MF, MS, LIS, /*InstrChanges=*/ModuloScheduleExpander::InstrChangesTy());
2192   MSE.expand();
2193   MSE.cleanup();
2194 }
2195 
2196 //===----------------------------------------------------------------------===//
2197 // ModuloScheduleTestAnnotater implementation
2198 //===----------------------------------------------------------------------===//
2199 
2200 void ModuloScheduleTestAnnotater::annotate() {
2201   for (MachineInstr *MI : S.getInstructions()) {
2202     SmallVector<char, 16> SV;
2203     raw_svector_ostream OS(SV);
2204     OS << "Stage-" << S.getStage(MI) << "_Cycle-" << S.getCycle(MI);
2205     MCSymbol *Sym = MF.getContext().getOrCreateSymbol(OS.str());
2206     MI->setPostInstrSymbol(MF, Sym);
2207   }
2208 }
2209