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 | static cl::opt<bool> SwapBranchTargetsMVE( |
26 | "pipeliner-swap-branch-targets-mve" , cl::Hidden, cl::init(Val: false), |
27 | cl::desc("Swap target blocks of a conditional branch for MVE expander" )); |
28 | |
29 | void ModuloSchedule::print(raw_ostream &OS) { |
30 | for (MachineInstr *MI : ScheduledInstrs) |
31 | OS << "[stage " << getStage(MI) << " @" << getCycle(MI) << "c] " << *MI; |
32 | } |
33 | |
34 | //===----------------------------------------------------------------------===// |
35 | // ModuloScheduleExpander implementation |
36 | //===----------------------------------------------------------------------===// |
37 | |
38 | /// Return the register values for the operands of a Phi instruction. |
39 | /// This function assume the instruction is a Phi. |
40 | static void getPhiRegs(MachineInstr &Phi, MachineBasicBlock *Loop, |
41 | unsigned &InitVal, unsigned &LoopVal) { |
42 | assert(Phi.isPHI() && "Expecting a Phi." ); |
43 | |
44 | InitVal = 0; |
45 | LoopVal = 0; |
46 | for (unsigned i = 1, e = Phi.getNumOperands(); i != e; i += 2) |
47 | if (Phi.getOperand(i: i + 1).getMBB() != Loop) |
48 | InitVal = Phi.getOperand(i).getReg(); |
49 | else |
50 | LoopVal = Phi.getOperand(i).getReg(); |
51 | |
52 | assert(InitVal != 0 && LoopVal != 0 && "Unexpected Phi structure." ); |
53 | } |
54 | |
55 | /// Return the Phi register value that comes from the incoming block. |
56 | static unsigned getInitPhiReg(MachineInstr &Phi, MachineBasicBlock *LoopBB) { |
57 | for (unsigned i = 1, e = Phi.getNumOperands(); i != e; i += 2) |
58 | if (Phi.getOperand(i: i + 1).getMBB() != LoopBB) |
59 | return Phi.getOperand(i).getReg(); |
60 | return 0; |
61 | } |
62 | |
63 | /// Return the Phi register value that comes the loop block. |
64 | static unsigned getLoopPhiReg(MachineInstr &Phi, MachineBasicBlock *LoopBB) { |
65 | for (unsigned i = 1, e = Phi.getNumOperands(); i != e; i += 2) |
66 | if (Phi.getOperand(i: i + 1).getMBB() == LoopBB) |
67 | return Phi.getOperand(i).getReg(); |
68 | return 0; |
69 | } |
70 | |
71 | void ModuloScheduleExpander::expand() { |
72 | BB = Schedule.getLoop()->getTopBlock(); |
73 | Preheader = *BB->pred_begin(); |
74 | if (Preheader == BB) |
75 | Preheader = *std::next(x: BB->pred_begin()); |
76 | |
77 | // Iterate over the definitions in each instruction, and compute the |
78 | // stage difference for each use. Keep the maximum value. |
79 | for (MachineInstr *MI : Schedule.getInstructions()) { |
80 | int DefStage = Schedule.getStage(MI); |
81 | for (const MachineOperand &Op : MI->all_defs()) { |
82 | Register Reg = Op.getReg(); |
83 | unsigned MaxDiff = 0; |
84 | bool PhiIsSwapped = false; |
85 | for (MachineOperand &UseOp : MRI.use_operands(Reg)) { |
86 | MachineInstr *UseMI = UseOp.getParent(); |
87 | int UseStage = Schedule.getStage(MI: UseMI); |
88 | unsigned Diff = 0; |
89 | if (UseStage != -1 && UseStage >= DefStage) |
90 | Diff = UseStage - DefStage; |
91 | if (MI->isPHI()) { |
92 | if (isLoopCarried(Phi&: *MI)) |
93 | ++Diff; |
94 | else |
95 | PhiIsSwapped = true; |
96 | } |
97 | MaxDiff = std::max(a: Diff, b: MaxDiff); |
98 | } |
99 | RegToStageDiff[Reg] = std::make_pair(x&: MaxDiff, y&: PhiIsSwapped); |
100 | } |
101 | } |
102 | |
103 | generatePipelinedLoop(); |
104 | } |
105 | |
106 | void ModuloScheduleExpander::generatePipelinedLoop() { |
107 | LoopInfo = TII->analyzeLoopForPipelining(LoopBB: BB); |
108 | assert(LoopInfo && "Must be able to analyze loop!" ); |
109 | |
110 | // Create a new basic block for the kernel and add it to the CFG. |
111 | MachineBasicBlock *KernelBB = MF.CreateMachineBasicBlock(BB: BB->getBasicBlock()); |
112 | |
113 | unsigned MaxStageCount = Schedule.getNumStages() - 1; |
114 | |
115 | // Remember the registers that are used in different stages. The index is |
116 | // the iteration, or stage, that the instruction is scheduled in. This is |
117 | // a map between register names in the original block and the names created |
118 | // in each stage of the pipelined loop. |
119 | ValueMapTy *VRMap = new ValueMapTy[(MaxStageCount + 1) * 2]; |
120 | |
121 | // The renaming destination by Phis for the registers across stages. |
122 | // This map is updated during Phis generation to point to the most recent |
123 | // renaming destination. |
124 | ValueMapTy *VRMapPhi = new ValueMapTy[(MaxStageCount + 1) * 2]; |
125 | |
126 | InstrMapTy InstrMap; |
127 | |
128 | SmallVector<MachineBasicBlock *, 4> PrologBBs; |
129 | |
130 | // Generate the prolog instructions that set up the pipeline. |
131 | generateProlog(LastStage: MaxStageCount, KernelBB, VRMap, PrologBBs); |
132 | MF.insert(MBBI: BB->getIterator(), MBB: KernelBB); |
133 | LIS.insertMBBInMaps(MBB: KernelBB); |
134 | |
135 | // Rearrange the instructions to generate the new, pipelined loop, |
136 | // and update register names as needed. |
137 | for (MachineInstr *CI : Schedule.getInstructions()) { |
138 | if (CI->isPHI()) |
139 | continue; |
140 | unsigned StageNum = Schedule.getStage(MI: CI); |
141 | MachineInstr *NewMI = cloneInstr(OldMI: CI, CurStageNum: MaxStageCount, InstStageNum: StageNum); |
142 | updateInstruction(NewMI, LastDef: false, CurStageNum: MaxStageCount, InstrStageNum: StageNum, VRMap); |
143 | KernelBB->push_back(MI: NewMI); |
144 | InstrMap[NewMI] = CI; |
145 | } |
146 | |
147 | // Copy any terminator instructions to the new kernel, and update |
148 | // names as needed. |
149 | for (MachineInstr &MI : BB->terminators()) { |
150 | MachineInstr *NewMI = MF.CloneMachineInstr(Orig: &MI); |
151 | updateInstruction(NewMI, LastDef: false, CurStageNum: MaxStageCount, InstrStageNum: 0, VRMap); |
152 | KernelBB->push_back(MI: NewMI); |
153 | InstrMap[NewMI] = &MI; |
154 | } |
155 | |
156 | NewKernel = KernelBB; |
157 | KernelBB->transferSuccessors(FromMBB: BB); |
158 | KernelBB->replaceSuccessor(Old: BB, New: KernelBB); |
159 | |
160 | generateExistingPhis(NewBB: KernelBB, BB1: PrologBBs.back(), BB2: KernelBB, KernelBB, VRMap, |
161 | InstrMap, LastStageNum: MaxStageCount, CurStageNum: MaxStageCount, IsLast: false); |
162 | generatePhis(NewBB: KernelBB, BB1: PrologBBs.back(), BB2: KernelBB, KernelBB, VRMap, VRMapPhi, |
163 | InstrMap, LastStageNum: MaxStageCount, CurStageNum: MaxStageCount, IsLast: false); |
164 | |
165 | LLVM_DEBUG(dbgs() << "New block\n" ; KernelBB->dump();); |
166 | |
167 | SmallVector<MachineBasicBlock *, 4> EpilogBBs; |
168 | // Generate the epilog instructions to complete the pipeline. |
169 | generateEpilog(LastStage: MaxStageCount, KernelBB, OrigBB: BB, VRMap, VRMapPhi, EpilogBBs, |
170 | PrologBBs); |
171 | |
172 | // We need this step because the register allocation doesn't handle some |
173 | // situations well, so we insert copies to help out. |
174 | splitLifetimes(KernelBB, EpilogBBs); |
175 | |
176 | // Remove dead instructions due to loop induction variables. |
177 | removeDeadInstructions(KernelBB, EpilogBBs); |
178 | |
179 | // Add branches between prolog and epilog blocks. |
180 | addBranches(PreheaderBB&: *Preheader, PrologBBs, KernelBB, EpilogBBs, VRMap); |
181 | |
182 | delete[] VRMap; |
183 | delete[] VRMapPhi; |
184 | } |
185 | |
186 | void ModuloScheduleExpander::cleanup() { |
187 | // Remove the original loop since it's no longer referenced. |
188 | for (auto &I : *BB) |
189 | LIS.RemoveMachineInstrFromMaps(MI&: I); |
190 | BB->clear(); |
191 | BB->eraseFromParent(); |
192 | } |
193 | |
194 | /// Generate the pipeline prolog code. |
195 | void ModuloScheduleExpander::generateProlog(unsigned LastStage, |
196 | MachineBasicBlock *KernelBB, |
197 | ValueMapTy *VRMap, |
198 | MBBVectorTy &PrologBBs) { |
199 | MachineBasicBlock *PredBB = Preheader; |
200 | InstrMapTy InstrMap; |
201 | |
202 | // Generate a basic block for each stage, not including the last stage, |
203 | // which will be generated in the kernel. Each basic block may contain |
204 | // instructions from multiple stages/iterations. |
205 | for (unsigned i = 0; i < LastStage; ++i) { |
206 | // Create and insert the prolog basic block prior to the original loop |
207 | // basic block. The original loop is removed later. |
208 | MachineBasicBlock *NewBB = MF.CreateMachineBasicBlock(BB: BB->getBasicBlock()); |
209 | PrologBBs.push_back(Elt: NewBB); |
210 | MF.insert(MBBI: BB->getIterator(), MBB: NewBB); |
211 | NewBB->transferSuccessors(FromMBB: PredBB); |
212 | PredBB->addSuccessor(Succ: NewBB); |
213 | PredBB = NewBB; |
214 | LIS.insertMBBInMaps(MBB: NewBB); |
215 | |
216 | // Generate instructions for each appropriate stage. Process instructions |
217 | // in original program order. |
218 | for (int StageNum = i; StageNum >= 0; --StageNum) { |
219 | for (MachineBasicBlock::iterator BBI = BB->instr_begin(), |
220 | BBE = BB->getFirstTerminator(); |
221 | BBI != BBE; ++BBI) { |
222 | if (Schedule.getStage(MI: &*BBI) == StageNum) { |
223 | if (BBI->isPHI()) |
224 | continue; |
225 | MachineInstr *NewMI = |
226 | cloneAndChangeInstr(OldMI: &*BBI, CurStageNum: i, InstStageNum: (unsigned)StageNum); |
227 | updateInstruction(NewMI, LastDef: false, CurStageNum: i, InstrStageNum: (unsigned)StageNum, VRMap); |
228 | NewBB->push_back(MI: NewMI); |
229 | InstrMap[NewMI] = &*BBI; |
230 | } |
231 | } |
232 | } |
233 | rewritePhiValues(NewBB, StageNum: i, VRMap, InstrMap); |
234 | LLVM_DEBUG({ |
235 | dbgs() << "prolog:\n" ; |
236 | NewBB->dump(); |
237 | }); |
238 | } |
239 | |
240 | PredBB->replaceSuccessor(Old: BB, New: KernelBB); |
241 | |
242 | // Check if we need to remove the branch from the preheader to the original |
243 | // loop, and replace it with a branch to the new loop. |
244 | unsigned numBranches = TII->removeBranch(MBB&: *Preheader); |
245 | if (numBranches) { |
246 | SmallVector<MachineOperand, 0> Cond; |
247 | TII->insertBranch(MBB&: *Preheader, TBB: PrologBBs[0], FBB: nullptr, Cond, DL: DebugLoc()); |
248 | } |
249 | } |
250 | |
251 | /// Generate the pipeline epilog code. The epilog code finishes the iterations |
252 | /// that were started in either the prolog or the kernel. We create a basic |
253 | /// block for each stage that needs to complete. |
254 | void ModuloScheduleExpander::generateEpilog( |
255 | unsigned LastStage, MachineBasicBlock *KernelBB, MachineBasicBlock *OrigBB, |
256 | ValueMapTy *VRMap, ValueMapTy *VRMapPhi, MBBVectorTy &EpilogBBs, |
257 | MBBVectorTy &PrologBBs) { |
258 | // We need to change the branch from the kernel to the first epilog block, so |
259 | // this call to analyze branch uses the kernel rather than the original BB. |
260 | MachineBasicBlock *TBB = nullptr, *FBB = nullptr; |
261 | SmallVector<MachineOperand, 4> Cond; |
262 | bool checkBranch = TII->analyzeBranch(MBB&: *KernelBB, TBB, FBB, Cond); |
263 | assert(!checkBranch && "generateEpilog must be able to analyze the branch" ); |
264 | if (checkBranch) |
265 | return; |
266 | |
267 | MachineBasicBlock::succ_iterator LoopExitI = KernelBB->succ_begin(); |
268 | if (*LoopExitI == KernelBB) |
269 | ++LoopExitI; |
270 | assert(LoopExitI != KernelBB->succ_end() && "Expecting a successor" ); |
271 | MachineBasicBlock *LoopExitBB = *LoopExitI; |
272 | |
273 | MachineBasicBlock *PredBB = KernelBB; |
274 | MachineBasicBlock *EpilogStart = LoopExitBB; |
275 | InstrMapTy InstrMap; |
276 | |
277 | // Generate a basic block for each stage, not including the last stage, |
278 | // which was generated for the kernel. Each basic block may contain |
279 | // instructions from multiple stages/iterations. |
280 | int EpilogStage = LastStage + 1; |
281 | for (unsigned i = LastStage; i >= 1; --i, ++EpilogStage) { |
282 | MachineBasicBlock *NewBB = MF.CreateMachineBasicBlock(); |
283 | EpilogBBs.push_back(Elt: NewBB); |
284 | MF.insert(MBBI: BB->getIterator(), MBB: NewBB); |
285 | |
286 | PredBB->replaceSuccessor(Old: LoopExitBB, New: NewBB); |
287 | NewBB->addSuccessor(Succ: LoopExitBB); |
288 | LIS.insertMBBInMaps(MBB: NewBB); |
289 | |
290 | if (EpilogStart == LoopExitBB) |
291 | EpilogStart = NewBB; |
292 | |
293 | // Add instructions to the epilog depending on the current block. |
294 | // Process instructions in original program order. |
295 | for (unsigned StageNum = i; StageNum <= LastStage; ++StageNum) { |
296 | for (auto &BBI : *BB) { |
297 | if (BBI.isPHI()) |
298 | continue; |
299 | MachineInstr *In = &BBI; |
300 | if ((unsigned)Schedule.getStage(MI: In) == StageNum) { |
301 | // Instructions with memoperands in the epilog are updated with |
302 | // conservative values. |
303 | MachineInstr *NewMI = cloneInstr(OldMI: In, UINT_MAX, InstStageNum: 0); |
304 | updateInstruction(NewMI, LastDef: i == 1, CurStageNum: EpilogStage, InstrStageNum: 0, VRMap); |
305 | NewBB->push_back(MI: NewMI); |
306 | InstrMap[NewMI] = In; |
307 | } |
308 | } |
309 | } |
310 | generateExistingPhis(NewBB, BB1: PrologBBs[i - 1], BB2: PredBB, KernelBB, VRMap, |
311 | InstrMap, LastStageNum: LastStage, CurStageNum: EpilogStage, IsLast: i == 1); |
312 | generatePhis(NewBB, BB1: PrologBBs[i - 1], BB2: PredBB, KernelBB, VRMap, VRMapPhi, |
313 | InstrMap, LastStageNum: LastStage, CurStageNum: EpilogStage, IsLast: i == 1); |
314 | PredBB = NewBB; |
315 | |
316 | LLVM_DEBUG({ |
317 | dbgs() << "epilog:\n" ; |
318 | NewBB->dump(); |
319 | }); |
320 | } |
321 | |
322 | // Fix any Phi nodes in the loop exit block. |
323 | LoopExitBB->replacePhiUsesWith(Old: BB, New: PredBB); |
324 | |
325 | // Create a branch to the new epilog from the kernel. |
326 | // Remove the original branch and add a new branch to the epilog. |
327 | TII->removeBranch(MBB&: *KernelBB); |
328 | assert((OrigBB == TBB || OrigBB == FBB) && |
329 | "Unable to determine looping branch direction" ); |
330 | if (OrigBB != TBB) |
331 | TII->insertBranch(MBB&: *KernelBB, TBB: EpilogStart, FBB: KernelBB, Cond, DL: DebugLoc()); |
332 | else |
333 | TII->insertBranch(MBB&: *KernelBB, TBB: KernelBB, FBB: EpilogStart, Cond, DL: DebugLoc()); |
334 | // Add a branch to the loop exit. |
335 | if (EpilogBBs.size() > 0) { |
336 | MachineBasicBlock *LastEpilogBB = EpilogBBs.back(); |
337 | SmallVector<MachineOperand, 4> Cond1; |
338 | TII->insertBranch(MBB&: *LastEpilogBB, TBB: LoopExitBB, FBB: nullptr, Cond: Cond1, DL: DebugLoc()); |
339 | } |
340 | } |
341 | |
342 | /// Replace all uses of FromReg that appear outside the specified |
343 | /// basic block with ToReg. |
344 | static void replaceRegUsesAfterLoop(unsigned FromReg, unsigned ToReg, |
345 | MachineBasicBlock *MBB, |
346 | MachineRegisterInfo &MRI, |
347 | LiveIntervals &LIS) { |
348 | for (MachineOperand &O : |
349 | llvm::make_early_inc_range(Range: MRI.use_operands(Reg: FromReg))) |
350 | if (O.getParent()->getParent() != MBB) |
351 | O.setReg(ToReg); |
352 | if (!LIS.hasInterval(Reg: ToReg)) |
353 | LIS.createEmptyInterval(Reg: ToReg); |
354 | } |
355 | |
356 | /// Return true if the register has a use that occurs outside the |
357 | /// specified loop. |
358 | static bool hasUseAfterLoop(unsigned Reg, MachineBasicBlock *BB, |
359 | MachineRegisterInfo &MRI) { |
360 | for (const MachineOperand &MO : MRI.use_operands(Reg)) |
361 | if (MO.getParent()->getParent() != BB) |
362 | return true; |
363 | return false; |
364 | } |
365 | |
366 | /// Generate Phis for the specific block in the generated pipelined code. |
367 | /// This function looks at the Phis from the original code to guide the |
368 | /// creation of new Phis. |
369 | void ModuloScheduleExpander::generateExistingPhis( |
370 | MachineBasicBlock *NewBB, MachineBasicBlock *BB1, MachineBasicBlock *BB2, |
371 | MachineBasicBlock *KernelBB, ValueMapTy *VRMap, InstrMapTy &InstrMap, |
372 | unsigned LastStageNum, unsigned CurStageNum, bool IsLast) { |
373 | // Compute the stage number for the initial value of the Phi, which |
374 | // comes from the prolog. The prolog to use depends on to which kernel/ |
375 | // epilog that we're adding the Phi. |
376 | unsigned PrologStage = 0; |
377 | unsigned PrevStage = 0; |
378 | bool InKernel = (LastStageNum == CurStageNum); |
379 | if (InKernel) { |
380 | PrologStage = LastStageNum - 1; |
381 | PrevStage = CurStageNum; |
382 | } else { |
383 | PrologStage = LastStageNum - (CurStageNum - LastStageNum); |
384 | PrevStage = LastStageNum + (CurStageNum - LastStageNum) - 1; |
385 | } |
386 | |
387 | for (MachineBasicBlock::iterator BBI = BB->instr_begin(), |
388 | BBE = BB->getFirstNonPHI(); |
389 | BBI != BBE; ++BBI) { |
390 | Register Def = BBI->getOperand(i: 0).getReg(); |
391 | |
392 | unsigned InitVal = 0; |
393 | unsigned LoopVal = 0; |
394 | getPhiRegs(Phi&: *BBI, Loop: BB, InitVal, LoopVal); |
395 | |
396 | unsigned PhiOp1 = 0; |
397 | // The Phi value from the loop body typically is defined in the loop, but |
398 | // not always. So, we need to check if the value is defined in the loop. |
399 | unsigned PhiOp2 = LoopVal; |
400 | if (VRMap[LastStageNum].count(Val: LoopVal)) |
401 | PhiOp2 = VRMap[LastStageNum][LoopVal]; |
402 | |
403 | int StageScheduled = Schedule.getStage(MI: &*BBI); |
404 | int LoopValStage = Schedule.getStage(MI: MRI.getVRegDef(Reg: LoopVal)); |
405 | unsigned NumStages = getStagesForReg(Reg: Def, CurStage: CurStageNum); |
406 | if (NumStages == 0) { |
407 | // We don't need to generate a Phi anymore, but we need to rename any uses |
408 | // of the Phi value. |
409 | unsigned NewReg = VRMap[PrevStage][LoopVal]; |
410 | rewriteScheduledInstr(BB: NewBB, InstrMap, CurStageNum, PhiNum: 0, Phi: &*BBI, OldReg: Def, |
411 | NewReg: InitVal, PrevReg: NewReg); |
412 | if (VRMap[CurStageNum].count(Val: LoopVal)) |
413 | VRMap[CurStageNum][Def] = VRMap[CurStageNum][LoopVal]; |
414 | } |
415 | // Adjust the number of Phis needed depending on the number of prologs left, |
416 | // and the distance from where the Phi is first scheduled. The number of |
417 | // Phis cannot exceed the number of prolog stages. Each stage can |
418 | // potentially define two values. |
419 | unsigned MaxPhis = PrologStage + 2; |
420 | if (!InKernel && (int)PrologStage <= LoopValStage) |
421 | MaxPhis = std::max(a: (int)MaxPhis - (int)LoopValStage, b: 1); |
422 | unsigned NumPhis = std::min(a: NumStages, b: MaxPhis); |
423 | |
424 | unsigned NewReg = 0; |
425 | unsigned AccessStage = (LoopValStage != -1) ? LoopValStage : StageScheduled; |
426 | // In the epilog, we may need to look back one stage to get the correct |
427 | // Phi name, because the epilog and prolog blocks execute the same stage. |
428 | // The correct name is from the previous block only when the Phi has |
429 | // been completely scheduled prior to the epilog, and Phi value is not |
430 | // needed in multiple stages. |
431 | int StageDiff = 0; |
432 | if (!InKernel && StageScheduled >= LoopValStage && AccessStage == 0 && |
433 | NumPhis == 1) |
434 | StageDiff = 1; |
435 | // Adjust the computations below when the phi and the loop definition |
436 | // are scheduled in different stages. |
437 | if (InKernel && LoopValStage != -1 && StageScheduled > LoopValStage) |
438 | StageDiff = StageScheduled - LoopValStage; |
439 | for (unsigned np = 0; np < NumPhis; ++np) { |
440 | // If the Phi hasn't been scheduled, then use the initial Phi operand |
441 | // value. Otherwise, use the scheduled version of the instruction. This |
442 | // is a little complicated when a Phi references another Phi. |
443 | if (np > PrologStage || StageScheduled >= (int)LastStageNum) |
444 | PhiOp1 = InitVal; |
445 | // Check if the Phi has already been scheduled in a prolog stage. |
446 | else if (PrologStage >= AccessStage + StageDiff + np && |
447 | VRMap[PrologStage - StageDiff - np].count(Val: LoopVal) != 0) |
448 | PhiOp1 = VRMap[PrologStage - StageDiff - np][LoopVal]; |
449 | // Check if the Phi has already been scheduled, but the loop instruction |
450 | // is either another Phi, or doesn't occur in the loop. |
451 | else if (PrologStage >= AccessStage + StageDiff + np) { |
452 | // If the Phi references another Phi, we need to examine the other |
453 | // Phi to get the correct value. |
454 | PhiOp1 = LoopVal; |
455 | MachineInstr *InstOp1 = MRI.getVRegDef(Reg: PhiOp1); |
456 | int Indirects = 1; |
457 | while (InstOp1 && InstOp1->isPHI() && InstOp1->getParent() == BB) { |
458 | int PhiStage = Schedule.getStage(MI: InstOp1); |
459 | if ((int)(PrologStage - StageDiff - np) < PhiStage + Indirects) |
460 | PhiOp1 = getInitPhiReg(Phi&: *InstOp1, LoopBB: BB); |
461 | else |
462 | PhiOp1 = getLoopPhiReg(Phi&: *InstOp1, LoopBB: BB); |
463 | InstOp1 = MRI.getVRegDef(Reg: PhiOp1); |
464 | int PhiOpStage = Schedule.getStage(MI: InstOp1); |
465 | int StageAdj = (PhiOpStage != -1 ? PhiStage - PhiOpStage : 0); |
466 | if (PhiOpStage != -1 && PrologStage - StageAdj >= Indirects + np && |
467 | VRMap[PrologStage - StageAdj - Indirects - np].count(Val: PhiOp1)) { |
468 | PhiOp1 = VRMap[PrologStage - StageAdj - Indirects - np][PhiOp1]; |
469 | break; |
470 | } |
471 | ++Indirects; |
472 | } |
473 | } else |
474 | PhiOp1 = InitVal; |
475 | // If this references a generated Phi in the kernel, get the Phi operand |
476 | // from the incoming block. |
477 | if (MachineInstr *InstOp1 = MRI.getVRegDef(Reg: PhiOp1)) |
478 | if (InstOp1->isPHI() && InstOp1->getParent() == KernelBB) |
479 | PhiOp1 = getInitPhiReg(Phi&: *InstOp1, LoopBB: KernelBB); |
480 | |
481 | MachineInstr *PhiInst = MRI.getVRegDef(Reg: LoopVal); |
482 | bool LoopDefIsPhi = PhiInst && PhiInst->isPHI(); |
483 | // In the epilog, a map lookup is needed to get the value from the kernel, |
484 | // or previous epilog block. How is does this depends on if the |
485 | // instruction is scheduled in the previous block. |
486 | if (!InKernel) { |
487 | int StageDiffAdj = 0; |
488 | if (LoopValStage != -1 && StageScheduled > LoopValStage) |
489 | StageDiffAdj = StageScheduled - LoopValStage; |
490 | // Use the loop value defined in the kernel, unless the kernel |
491 | // contains the last definition of the Phi. |
492 | if (np == 0 && PrevStage == LastStageNum && |
493 | (StageScheduled != 0 || LoopValStage != 0) && |
494 | VRMap[PrevStage - StageDiffAdj].count(Val: LoopVal)) |
495 | PhiOp2 = VRMap[PrevStage - StageDiffAdj][LoopVal]; |
496 | // Use the value defined by the Phi. We add one because we switch |
497 | // from looking at the loop value to the Phi definition. |
498 | else if (np > 0 && PrevStage == LastStageNum && |
499 | VRMap[PrevStage - np + 1].count(Val: Def)) |
500 | PhiOp2 = VRMap[PrevStage - np + 1][Def]; |
501 | // Use the loop value defined in the kernel. |
502 | else if (static_cast<unsigned>(LoopValStage) > PrologStage + 1 && |
503 | VRMap[PrevStage - StageDiffAdj - np].count(Val: LoopVal)) |
504 | PhiOp2 = VRMap[PrevStage - StageDiffAdj - np][LoopVal]; |
505 | // Use the value defined by the Phi, unless we're generating the first |
506 | // epilog and the Phi refers to a Phi in a different stage. |
507 | else if (VRMap[PrevStage - np].count(Val: Def) && |
508 | (!LoopDefIsPhi || (PrevStage != LastStageNum) || |
509 | (LoopValStage == StageScheduled))) |
510 | PhiOp2 = VRMap[PrevStage - np][Def]; |
511 | } |
512 | |
513 | // Check if we can reuse an existing Phi. This occurs when a Phi |
514 | // references another Phi, and the other Phi is scheduled in an |
515 | // earlier stage. We can try to reuse an existing Phi up until the last |
516 | // stage of the current Phi. |
517 | if (LoopDefIsPhi) { |
518 | if (static_cast<int>(PrologStage - np) >= StageScheduled) { |
519 | int LVNumStages = getStagesForPhi(Reg: LoopVal); |
520 | int StageDiff = (StageScheduled - LoopValStage); |
521 | LVNumStages -= StageDiff; |
522 | // Make sure the loop value Phi has been processed already. |
523 | if (LVNumStages > (int)np && VRMap[CurStageNum].count(Val: LoopVal)) { |
524 | NewReg = PhiOp2; |
525 | unsigned ReuseStage = CurStageNum; |
526 | if (isLoopCarried(Phi&: *PhiInst)) |
527 | ReuseStage -= LVNumStages; |
528 | // Check if the Phi to reuse has been generated yet. If not, then |
529 | // there is nothing to reuse. |
530 | if (VRMap[ReuseStage - np].count(Val: LoopVal)) { |
531 | NewReg = VRMap[ReuseStage - np][LoopVal]; |
532 | |
533 | rewriteScheduledInstr(BB: NewBB, InstrMap, CurStageNum, PhiNum: np, Phi: &*BBI, |
534 | OldReg: Def, NewReg); |
535 | // Update the map with the new Phi name. |
536 | VRMap[CurStageNum - np][Def] = NewReg; |
537 | PhiOp2 = NewReg; |
538 | if (VRMap[LastStageNum - np - 1].count(Val: LoopVal)) |
539 | PhiOp2 = VRMap[LastStageNum - np - 1][LoopVal]; |
540 | |
541 | if (IsLast && np == NumPhis - 1) |
542 | replaceRegUsesAfterLoop(FromReg: Def, ToReg: NewReg, MBB: BB, MRI, LIS); |
543 | continue; |
544 | } |
545 | } |
546 | } |
547 | if (InKernel && StageDiff > 0 && |
548 | VRMap[CurStageNum - StageDiff - np].count(Val: LoopVal)) |
549 | PhiOp2 = VRMap[CurStageNum - StageDiff - np][LoopVal]; |
550 | } |
551 | |
552 | const TargetRegisterClass *RC = MRI.getRegClass(Reg: Def); |
553 | NewReg = MRI.createVirtualRegister(RegClass: RC); |
554 | |
555 | MachineInstrBuilder NewPhi = |
556 | BuildMI(BB&: *NewBB, I: NewBB->getFirstNonPHI(), MIMD: DebugLoc(), |
557 | MCID: TII->get(Opcode: TargetOpcode::PHI), DestReg: NewReg); |
558 | NewPhi.addReg(RegNo: PhiOp1).addMBB(MBB: BB1); |
559 | NewPhi.addReg(RegNo: PhiOp2).addMBB(MBB: BB2); |
560 | if (np == 0) |
561 | InstrMap[NewPhi] = &*BBI; |
562 | |
563 | // We define the Phis after creating the new pipelined code, so |
564 | // we need to rename the Phi values in scheduled instructions. |
565 | |
566 | unsigned PrevReg = 0; |
567 | if (InKernel && VRMap[PrevStage - np].count(Val: LoopVal)) |
568 | PrevReg = VRMap[PrevStage - np][LoopVal]; |
569 | rewriteScheduledInstr(BB: NewBB, InstrMap, CurStageNum, PhiNum: np, Phi: &*BBI, OldReg: Def, |
570 | NewReg, PrevReg); |
571 | // If the Phi has been scheduled, use the new name for rewriting. |
572 | if (VRMap[CurStageNum - np].count(Val: Def)) { |
573 | unsigned R = VRMap[CurStageNum - np][Def]; |
574 | rewriteScheduledInstr(BB: NewBB, InstrMap, CurStageNum, PhiNum: np, Phi: &*BBI, OldReg: R, |
575 | NewReg); |
576 | } |
577 | |
578 | // Check if we need to rename any uses that occurs after the loop. The |
579 | // register to replace depends on whether the Phi is scheduled in the |
580 | // epilog. |
581 | if (IsLast && np == NumPhis - 1) |
582 | replaceRegUsesAfterLoop(FromReg: Def, ToReg: NewReg, MBB: BB, MRI, LIS); |
583 | |
584 | // In the kernel, a dependent Phi uses the value from this Phi. |
585 | if (InKernel) |
586 | PhiOp2 = NewReg; |
587 | |
588 | // Update the map with the new Phi name. |
589 | VRMap[CurStageNum - np][Def] = NewReg; |
590 | } |
591 | |
592 | while (NumPhis++ < NumStages) { |
593 | rewriteScheduledInstr(BB: NewBB, InstrMap, CurStageNum, PhiNum: NumPhis, Phi: &*BBI, OldReg: Def, |
594 | NewReg, PrevReg: 0); |
595 | } |
596 | |
597 | // Check if we need to rename a Phi that has been eliminated due to |
598 | // scheduling. |
599 | if (NumStages == 0 && IsLast && VRMap[CurStageNum].count(Val: LoopVal)) |
600 | replaceRegUsesAfterLoop(FromReg: Def, ToReg: VRMap[CurStageNum][LoopVal], MBB: BB, MRI, LIS); |
601 | } |
602 | } |
603 | |
604 | /// Generate Phis for the specified block in the generated pipelined code. |
605 | /// These are new Phis needed because the definition is scheduled after the |
606 | /// use in the pipelined sequence. |
607 | void ModuloScheduleExpander::generatePhis( |
608 | MachineBasicBlock *NewBB, MachineBasicBlock *BB1, MachineBasicBlock *BB2, |
609 | MachineBasicBlock *KernelBB, ValueMapTy *VRMap, ValueMapTy *VRMapPhi, |
610 | InstrMapTy &InstrMap, unsigned LastStageNum, unsigned CurStageNum, |
611 | bool IsLast) { |
612 | // Compute the stage number that contains the initial Phi value, and |
613 | // the Phi from the previous stage. |
614 | unsigned PrologStage = 0; |
615 | unsigned PrevStage = 0; |
616 | unsigned StageDiff = CurStageNum - LastStageNum; |
617 | bool InKernel = (StageDiff == 0); |
618 | if (InKernel) { |
619 | PrologStage = LastStageNum - 1; |
620 | PrevStage = CurStageNum; |
621 | } else { |
622 | PrologStage = LastStageNum - StageDiff; |
623 | PrevStage = LastStageNum + StageDiff - 1; |
624 | } |
625 | |
626 | for (MachineBasicBlock::iterator BBI = BB->getFirstNonPHI(), |
627 | BBE = BB->instr_end(); |
628 | BBI != BBE; ++BBI) { |
629 | for (unsigned i = 0, e = BBI->getNumOperands(); i != e; ++i) { |
630 | MachineOperand &MO = BBI->getOperand(i); |
631 | if (!MO.isReg() || !MO.isDef() || !MO.getReg().isVirtual()) |
632 | continue; |
633 | |
634 | int StageScheduled = Schedule.getStage(MI: &*BBI); |
635 | assert(StageScheduled != -1 && "Expecting scheduled instruction." ); |
636 | Register Def = MO.getReg(); |
637 | unsigned NumPhis = getStagesForReg(Reg: Def, CurStage: CurStageNum); |
638 | // An instruction scheduled in stage 0 and is used after the loop |
639 | // requires a phi in the epilog for the last definition from either |
640 | // the kernel or prolog. |
641 | if (!InKernel && NumPhis == 0 && StageScheduled == 0 && |
642 | hasUseAfterLoop(Reg: Def, BB, MRI)) |
643 | NumPhis = 1; |
644 | if (!InKernel && (unsigned)StageScheduled > PrologStage) |
645 | continue; |
646 | |
647 | unsigned PhiOp2; |
648 | if (InKernel) { |
649 | PhiOp2 = VRMap[PrevStage][Def]; |
650 | if (MachineInstr *InstOp2 = MRI.getVRegDef(Reg: PhiOp2)) |
651 | if (InstOp2->isPHI() && InstOp2->getParent() == NewBB) |
652 | PhiOp2 = getLoopPhiReg(Phi&: *InstOp2, LoopBB: BB2); |
653 | } |
654 | // The number of Phis can't exceed the number of prolog stages. The |
655 | // prolog stage number is zero based. |
656 | if (NumPhis > PrologStage + 1 - StageScheduled) |
657 | NumPhis = PrologStage + 1 - StageScheduled; |
658 | for (unsigned np = 0; np < NumPhis; ++np) { |
659 | // Example for |
660 | // Org: |
661 | // %Org = ... (Scheduled at Stage#0, NumPhi = 2) |
662 | // |
663 | // Prolog0 (Stage0): |
664 | // %Clone0 = ... |
665 | // Prolog1 (Stage1): |
666 | // %Clone1 = ... |
667 | // Kernel (Stage2): |
668 | // %Phi0 = Phi %Clone1, Prolog1, %Clone2, Kernel |
669 | // %Phi1 = Phi %Clone0, Prolog1, %Phi0, Kernel |
670 | // %Clone2 = ... |
671 | // Epilog0 (Stage3): |
672 | // %Phi2 = Phi %Clone1, Prolog1, %Clone2, Kernel |
673 | // %Phi3 = Phi %Clone0, Prolog1, %Phi0, Kernel |
674 | // Epilog1 (Stage4): |
675 | // %Phi4 = Phi %Clone0, Prolog0, %Phi2, Epilog0 |
676 | // |
677 | // VRMap = {0: %Clone0, 1: %Clone1, 2: %Clone2} |
678 | // VRMapPhi (after Kernel) = {0: %Phi1, 1: %Phi0} |
679 | // VRMapPhi (after Epilog0) = {0: %Phi3, 1: %Phi2} |
680 | |
681 | unsigned PhiOp1 = VRMap[PrologStage][Def]; |
682 | if (np <= PrologStage) |
683 | PhiOp1 = VRMap[PrologStage - np][Def]; |
684 | if (!InKernel) { |
685 | if (PrevStage == LastStageNum && np == 0) |
686 | PhiOp2 = VRMap[LastStageNum][Def]; |
687 | else |
688 | PhiOp2 = VRMapPhi[PrevStage - np][Def]; |
689 | } |
690 | |
691 | const TargetRegisterClass *RC = MRI.getRegClass(Reg: Def); |
692 | Register NewReg = MRI.createVirtualRegister(RegClass: RC); |
693 | |
694 | MachineInstrBuilder NewPhi = |
695 | BuildMI(BB&: *NewBB, I: NewBB->getFirstNonPHI(), MIMD: DebugLoc(), |
696 | MCID: TII->get(Opcode: TargetOpcode::PHI), DestReg: NewReg); |
697 | NewPhi.addReg(RegNo: PhiOp1).addMBB(MBB: BB1); |
698 | NewPhi.addReg(RegNo: PhiOp2).addMBB(MBB: BB2); |
699 | if (np == 0) |
700 | InstrMap[NewPhi] = &*BBI; |
701 | |
702 | // Rewrite uses and update the map. The actions depend upon whether |
703 | // we generating code for the kernel or epilog blocks. |
704 | if (InKernel) { |
705 | rewriteScheduledInstr(BB: NewBB, InstrMap, CurStageNum, PhiNum: np, Phi: &*BBI, OldReg: PhiOp1, |
706 | NewReg); |
707 | rewriteScheduledInstr(BB: NewBB, InstrMap, CurStageNum, PhiNum: np, Phi: &*BBI, OldReg: PhiOp2, |
708 | NewReg); |
709 | |
710 | PhiOp2 = NewReg; |
711 | VRMapPhi[PrevStage - np - 1][Def] = NewReg; |
712 | } else { |
713 | VRMapPhi[CurStageNum - np][Def] = NewReg; |
714 | if (np == NumPhis - 1) |
715 | rewriteScheduledInstr(BB: NewBB, InstrMap, CurStageNum, PhiNum: np, Phi: &*BBI, OldReg: Def, |
716 | NewReg); |
717 | } |
718 | if (IsLast && np == NumPhis - 1) |
719 | replaceRegUsesAfterLoop(FromReg: Def, ToReg: NewReg, MBB: BB, MRI, LIS); |
720 | } |
721 | } |
722 | } |
723 | } |
724 | |
725 | /// Remove instructions that generate values with no uses. |
726 | /// Typically, these are induction variable operations that generate values |
727 | /// used in the loop itself. A dead instruction has a definition with |
728 | /// no uses, or uses that occur in the original loop only. |
729 | void ModuloScheduleExpander::removeDeadInstructions(MachineBasicBlock *KernelBB, |
730 | MBBVectorTy &EpilogBBs) { |
731 | // For each epilog block, check that the value defined by each instruction |
732 | // is used. If not, delete it. |
733 | for (MachineBasicBlock *MBB : llvm::reverse(C&: EpilogBBs)) |
734 | for (MachineBasicBlock::reverse_instr_iterator MI = MBB->instr_rbegin(), |
735 | ME = MBB->instr_rend(); |
736 | MI != ME;) { |
737 | // From DeadMachineInstructionElem. Don't delete inline assembly. |
738 | if (MI->isInlineAsm()) { |
739 | ++MI; |
740 | continue; |
741 | } |
742 | bool SawStore = false; |
743 | // Check if it's safe to remove the instruction due to side effects. |
744 | // We can, and want to, remove Phis here. |
745 | if (!MI->isSafeToMove(AA: nullptr, SawStore) && !MI->isPHI()) { |
746 | ++MI; |
747 | continue; |
748 | } |
749 | bool used = true; |
750 | for (const MachineOperand &MO : MI->all_defs()) { |
751 | Register reg = MO.getReg(); |
752 | // Assume physical registers are used, unless they are marked dead. |
753 | if (reg.isPhysical()) { |
754 | used = !MO.isDead(); |
755 | if (used) |
756 | break; |
757 | continue; |
758 | } |
759 | unsigned realUses = 0; |
760 | for (const MachineOperand &U : MRI.use_operands(Reg: reg)) { |
761 | // Check if there are any uses that occur only in the original |
762 | // loop. If so, that's not a real use. |
763 | if (U.getParent()->getParent() != BB) { |
764 | realUses++; |
765 | used = true; |
766 | break; |
767 | } |
768 | } |
769 | if (realUses > 0) |
770 | break; |
771 | used = false; |
772 | } |
773 | if (!used) { |
774 | LIS.RemoveMachineInstrFromMaps(MI&: *MI); |
775 | MI++->eraseFromParent(); |
776 | continue; |
777 | } |
778 | ++MI; |
779 | } |
780 | // In the kernel block, check if we can remove a Phi that generates a value |
781 | // used in an instruction removed in the epilog block. |
782 | for (MachineInstr &MI : llvm::make_early_inc_range(Range: KernelBB->phis())) { |
783 | Register reg = MI.getOperand(i: 0).getReg(); |
784 | if (MRI.use_begin(RegNo: reg) == MRI.use_end()) { |
785 | LIS.RemoveMachineInstrFromMaps(MI); |
786 | MI.eraseFromParent(); |
787 | } |
788 | } |
789 | } |
790 | |
791 | /// For loop carried definitions, we split the lifetime of a virtual register |
792 | /// that has uses past the definition in the next iteration. A copy with a new |
793 | /// virtual register is inserted before the definition, which helps with |
794 | /// generating a better register assignment. |
795 | /// |
796 | /// v1 = phi(a, v2) v1 = phi(a, v2) |
797 | /// v2 = phi(b, v3) v2 = phi(b, v3) |
798 | /// v3 = .. v4 = copy v1 |
799 | /// .. = V1 v3 = .. |
800 | /// .. = v4 |
801 | void ModuloScheduleExpander::splitLifetimes(MachineBasicBlock *KernelBB, |
802 | MBBVectorTy &EpilogBBs) { |
803 | const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); |
804 | for (auto &PHI : KernelBB->phis()) { |
805 | Register Def = PHI.getOperand(i: 0).getReg(); |
806 | // Check for any Phi definition that used as an operand of another Phi |
807 | // in the same block. |
808 | for (MachineRegisterInfo::use_instr_iterator I = MRI.use_instr_begin(RegNo: Def), |
809 | E = MRI.use_instr_end(); |
810 | I != E; ++I) { |
811 | if (I->isPHI() && I->getParent() == KernelBB) { |
812 | // Get the loop carried definition. |
813 | unsigned LCDef = getLoopPhiReg(Phi&: PHI, LoopBB: KernelBB); |
814 | if (!LCDef) |
815 | continue; |
816 | MachineInstr *MI = MRI.getVRegDef(Reg: LCDef); |
817 | if (!MI || MI->getParent() != KernelBB || MI->isPHI()) |
818 | continue; |
819 | // Search through the rest of the block looking for uses of the Phi |
820 | // definition. If one occurs, then split the lifetime. |
821 | unsigned SplitReg = 0; |
822 | for (auto &BBJ : make_range(x: MachineBasicBlock::instr_iterator(MI), |
823 | y: KernelBB->instr_end())) |
824 | if (BBJ.readsRegister(Reg: Def, /*TRI=*/nullptr)) { |
825 | // We split the lifetime when we find the first use. |
826 | if (SplitReg == 0) { |
827 | SplitReg = MRI.createVirtualRegister(RegClass: MRI.getRegClass(Reg: Def)); |
828 | BuildMI(BB&: *KernelBB, I: MI, MIMD: MI->getDebugLoc(), |
829 | MCID: TII->get(Opcode: TargetOpcode::COPY), DestReg: SplitReg) |
830 | .addReg(RegNo: Def); |
831 | } |
832 | BBJ.substituteRegister(FromReg: Def, ToReg: SplitReg, SubIdx: 0, RegInfo: *TRI); |
833 | } |
834 | if (!SplitReg) |
835 | continue; |
836 | // Search through each of the epilog blocks for any uses to be renamed. |
837 | for (auto &Epilog : EpilogBBs) |
838 | for (auto &I : *Epilog) |
839 | if (I.readsRegister(Reg: Def, /*TRI=*/nullptr)) |
840 | I.substituteRegister(FromReg: Def, ToReg: SplitReg, SubIdx: 0, RegInfo: *TRI); |
841 | break; |
842 | } |
843 | } |
844 | } |
845 | } |
846 | |
847 | /// Remove the incoming block from the Phis in a basic block. |
848 | static void removePhis(MachineBasicBlock *BB, MachineBasicBlock *Incoming) { |
849 | for (MachineInstr &MI : *BB) { |
850 | if (!MI.isPHI()) |
851 | break; |
852 | for (unsigned i = 1, e = MI.getNumOperands(); i != e; i += 2) |
853 | if (MI.getOperand(i: i + 1).getMBB() == Incoming) { |
854 | MI.removeOperand(OpNo: i + 1); |
855 | MI.removeOperand(OpNo: i); |
856 | break; |
857 | } |
858 | } |
859 | } |
860 | |
861 | /// Create branches from each prolog basic block to the appropriate epilog |
862 | /// block. These edges are needed if the loop ends before reaching the |
863 | /// kernel. |
864 | void ModuloScheduleExpander::addBranches(MachineBasicBlock &, |
865 | MBBVectorTy &PrologBBs, |
866 | MachineBasicBlock *KernelBB, |
867 | MBBVectorTy &EpilogBBs, |
868 | ValueMapTy *VRMap) { |
869 | assert(PrologBBs.size() == EpilogBBs.size() && "Prolog/Epilog mismatch" ); |
870 | MachineBasicBlock *LastPro = KernelBB; |
871 | MachineBasicBlock *LastEpi = KernelBB; |
872 | |
873 | // Start from the blocks connected to the kernel and work "out" |
874 | // to the first prolog and the last epilog blocks. |
875 | SmallVector<MachineInstr *, 4> PrevInsts; |
876 | unsigned MaxIter = PrologBBs.size() - 1; |
877 | for (unsigned i = 0, j = MaxIter; i <= MaxIter; ++i, --j) { |
878 | // Add branches to the prolog that go to the corresponding |
879 | // epilog, and the fall-thru prolog/kernel block. |
880 | MachineBasicBlock *Prolog = PrologBBs[j]; |
881 | MachineBasicBlock *Epilog = EpilogBBs[i]; |
882 | |
883 | SmallVector<MachineOperand, 4> Cond; |
884 | std::optional<bool> StaticallyGreater = |
885 | LoopInfo->createTripCountGreaterCondition(TC: j + 1, MBB&: *Prolog, Cond); |
886 | unsigned numAdded = 0; |
887 | if (!StaticallyGreater) { |
888 | Prolog->addSuccessor(Succ: Epilog); |
889 | numAdded = TII->insertBranch(MBB&: *Prolog, TBB: Epilog, FBB: LastPro, Cond, DL: DebugLoc()); |
890 | } else if (*StaticallyGreater == false) { |
891 | Prolog->addSuccessor(Succ: Epilog); |
892 | Prolog->removeSuccessor(Succ: LastPro); |
893 | LastEpi->removeSuccessor(Succ: Epilog); |
894 | numAdded = TII->insertBranch(MBB&: *Prolog, TBB: Epilog, FBB: nullptr, Cond, DL: DebugLoc()); |
895 | removePhis(BB: Epilog, Incoming: LastEpi); |
896 | // Remove the blocks that are no longer referenced. |
897 | if (LastPro != LastEpi) { |
898 | LastEpi->clear(); |
899 | LastEpi->eraseFromParent(); |
900 | } |
901 | if (LastPro == KernelBB) { |
902 | LoopInfo->disposed(); |
903 | NewKernel = nullptr; |
904 | } |
905 | LastPro->clear(); |
906 | LastPro->eraseFromParent(); |
907 | } else { |
908 | numAdded = TII->insertBranch(MBB&: *Prolog, TBB: LastPro, FBB: nullptr, Cond, DL: DebugLoc()); |
909 | removePhis(BB: Epilog, Incoming: Prolog); |
910 | } |
911 | LastPro = Prolog; |
912 | LastEpi = Epilog; |
913 | for (MachineBasicBlock::reverse_instr_iterator I = Prolog->instr_rbegin(), |
914 | E = Prolog->instr_rend(); |
915 | I != E && numAdded > 0; ++I, --numAdded) |
916 | updateInstruction(NewMI: &*I, LastDef: false, CurStageNum: j, InstrStageNum: 0, VRMap); |
917 | } |
918 | |
919 | if (NewKernel) { |
920 | LoopInfo->setPreheader(PrologBBs[MaxIter]); |
921 | LoopInfo->adjustTripCount(TripCountAdjust: -(MaxIter + 1)); |
922 | } |
923 | } |
924 | |
925 | /// Return true if we can compute the amount the instruction changes |
926 | /// during each iteration. Set Delta to the amount of the change. |
927 | bool ModuloScheduleExpander::computeDelta(MachineInstr &MI, unsigned &Delta) { |
928 | const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); |
929 | const MachineOperand *BaseOp; |
930 | int64_t Offset; |
931 | bool OffsetIsScalable; |
932 | if (!TII->getMemOperandWithOffset(MI, BaseOp, Offset, OffsetIsScalable, TRI)) |
933 | return false; |
934 | |
935 | // FIXME: This algorithm assumes instructions have fixed-size offsets. |
936 | if (OffsetIsScalable) |
937 | return false; |
938 | |
939 | if (!BaseOp->isReg()) |
940 | return false; |
941 | |
942 | Register BaseReg = BaseOp->getReg(); |
943 | |
944 | MachineRegisterInfo &MRI = MF.getRegInfo(); |
945 | // Check if there is a Phi. If so, get the definition in the loop. |
946 | MachineInstr *BaseDef = MRI.getVRegDef(Reg: BaseReg); |
947 | if (BaseDef && BaseDef->isPHI()) { |
948 | BaseReg = getLoopPhiReg(Phi&: *BaseDef, LoopBB: MI.getParent()); |
949 | BaseDef = MRI.getVRegDef(Reg: BaseReg); |
950 | } |
951 | if (!BaseDef) |
952 | return false; |
953 | |
954 | int D = 0; |
955 | if (!TII->getIncrementValue(MI: *BaseDef, Value&: D) && D >= 0) |
956 | return false; |
957 | |
958 | Delta = D; |
959 | return true; |
960 | } |
961 | |
962 | /// Update the memory operand with a new offset when the pipeliner |
963 | /// generates a new copy of the instruction that refers to a |
964 | /// different memory location. |
965 | void ModuloScheduleExpander::updateMemOperands(MachineInstr &NewMI, |
966 | MachineInstr &OldMI, |
967 | unsigned Num) { |
968 | if (Num == 0) |
969 | return; |
970 | // If the instruction has memory operands, then adjust the offset |
971 | // when the instruction appears in different stages. |
972 | if (NewMI.memoperands_empty()) |
973 | return; |
974 | SmallVector<MachineMemOperand *, 2> NewMMOs; |
975 | for (MachineMemOperand *MMO : NewMI.memoperands()) { |
976 | // TODO: Figure out whether isAtomic is really necessary (see D57601). |
977 | if (MMO->isVolatile() || MMO->isAtomic() || |
978 | (MMO->isInvariant() && MMO->isDereferenceable()) || |
979 | (!MMO->getValue())) { |
980 | NewMMOs.push_back(Elt: MMO); |
981 | continue; |
982 | } |
983 | unsigned Delta; |
984 | if (Num != UINT_MAX && computeDelta(MI&: OldMI, Delta)) { |
985 | int64_t AdjOffset = Delta * Num; |
986 | NewMMOs.push_back( |
987 | Elt: MF.getMachineMemOperand(MMO, Offset: AdjOffset, Size: MMO->getSize())); |
988 | } else { |
989 | NewMMOs.push_back(Elt: MF.getMachineMemOperand( |
990 | MMO, Offset: 0, Size: LocationSize::beforeOrAfterPointer())); |
991 | } |
992 | } |
993 | NewMI.setMemRefs(MF, MemRefs: NewMMOs); |
994 | } |
995 | |
996 | /// Clone the instruction for the new pipelined loop and update the |
997 | /// memory operands, if needed. |
998 | MachineInstr *ModuloScheduleExpander::cloneInstr(MachineInstr *OldMI, |
999 | unsigned CurStageNum, |
1000 | unsigned InstStageNum) { |
1001 | MachineInstr *NewMI = MF.CloneMachineInstr(Orig: OldMI); |
1002 | updateMemOperands(NewMI&: *NewMI, OldMI&: *OldMI, Num: CurStageNum - InstStageNum); |
1003 | return NewMI; |
1004 | } |
1005 | |
1006 | /// Clone the instruction for the new pipelined loop. If needed, this |
1007 | /// function updates the instruction using the values saved in the |
1008 | /// InstrChanges structure. |
1009 | MachineInstr *ModuloScheduleExpander::cloneAndChangeInstr( |
1010 | MachineInstr *OldMI, unsigned CurStageNum, unsigned InstStageNum) { |
1011 | MachineInstr *NewMI = MF.CloneMachineInstr(Orig: OldMI); |
1012 | auto It = InstrChanges.find(Val: OldMI); |
1013 | if (It != InstrChanges.end()) { |
1014 | std::pair<unsigned, int64_t> RegAndOffset = It->second; |
1015 | unsigned BasePos, OffsetPos; |
1016 | if (!TII->getBaseAndOffsetPosition(MI: *OldMI, BasePos, OffsetPos)) |
1017 | return nullptr; |
1018 | int64_t NewOffset = OldMI->getOperand(i: OffsetPos).getImm(); |
1019 | MachineInstr *LoopDef = findDefInLoop(Reg: RegAndOffset.first); |
1020 | if (Schedule.getStage(MI: LoopDef) > (signed)InstStageNum) |
1021 | NewOffset += RegAndOffset.second * (CurStageNum - InstStageNum); |
1022 | NewMI->getOperand(i: OffsetPos).setImm(NewOffset); |
1023 | } |
1024 | updateMemOperands(NewMI&: *NewMI, OldMI&: *OldMI, Num: CurStageNum - InstStageNum); |
1025 | return NewMI; |
1026 | } |
1027 | |
1028 | /// Update the machine instruction with new virtual registers. This |
1029 | /// function may change the definitions and/or uses. |
1030 | void ModuloScheduleExpander::updateInstruction(MachineInstr *NewMI, |
1031 | bool LastDef, |
1032 | unsigned CurStageNum, |
1033 | unsigned InstrStageNum, |
1034 | ValueMapTy *VRMap) { |
1035 | for (MachineOperand &MO : NewMI->operands()) { |
1036 | if (!MO.isReg() || !MO.getReg().isVirtual()) |
1037 | continue; |
1038 | Register reg = MO.getReg(); |
1039 | if (MO.isDef()) { |
1040 | // Create a new virtual register for the definition. |
1041 | const TargetRegisterClass *RC = MRI.getRegClass(Reg: reg); |
1042 | Register NewReg = MRI.createVirtualRegister(RegClass: RC); |
1043 | MO.setReg(NewReg); |
1044 | VRMap[CurStageNum][reg] = NewReg; |
1045 | if (LastDef) |
1046 | replaceRegUsesAfterLoop(FromReg: reg, ToReg: NewReg, MBB: BB, MRI, LIS); |
1047 | } else if (MO.isUse()) { |
1048 | MachineInstr *Def = MRI.getVRegDef(Reg: reg); |
1049 | // Compute the stage that contains the last definition for instruction. |
1050 | int DefStageNum = Schedule.getStage(MI: Def); |
1051 | unsigned StageNum = CurStageNum; |
1052 | if (DefStageNum != -1 && (int)InstrStageNum > DefStageNum) { |
1053 | // Compute the difference in stages between the defintion and the use. |
1054 | unsigned StageDiff = (InstrStageNum - DefStageNum); |
1055 | // Make an adjustment to get the last definition. |
1056 | StageNum -= StageDiff; |
1057 | } |
1058 | if (VRMap[StageNum].count(Val: reg)) |
1059 | MO.setReg(VRMap[StageNum][reg]); |
1060 | } |
1061 | } |
1062 | } |
1063 | |
1064 | /// Return the instruction in the loop that defines the register. |
1065 | /// If the definition is a Phi, then follow the Phi operand to |
1066 | /// the instruction in the loop. |
1067 | MachineInstr *ModuloScheduleExpander::findDefInLoop(unsigned Reg) { |
1068 | SmallPtrSet<MachineInstr *, 8> Visited; |
1069 | MachineInstr *Def = MRI.getVRegDef(Reg); |
1070 | while (Def->isPHI()) { |
1071 | if (!Visited.insert(Ptr: Def).second) |
1072 | break; |
1073 | for (unsigned i = 1, e = Def->getNumOperands(); i < e; i += 2) |
1074 | if (Def->getOperand(i: i + 1).getMBB() == BB) { |
1075 | Def = MRI.getVRegDef(Reg: Def->getOperand(i).getReg()); |
1076 | break; |
1077 | } |
1078 | } |
1079 | return Def; |
1080 | } |
1081 | |
1082 | /// Return the new name for the value from the previous stage. |
1083 | unsigned ModuloScheduleExpander::getPrevMapVal( |
1084 | unsigned StageNum, unsigned PhiStage, unsigned LoopVal, unsigned LoopStage, |
1085 | ValueMapTy *VRMap, MachineBasicBlock *BB) { |
1086 | unsigned PrevVal = 0; |
1087 | if (StageNum > PhiStage) { |
1088 | MachineInstr *LoopInst = MRI.getVRegDef(Reg: LoopVal); |
1089 | if (PhiStage == LoopStage && VRMap[StageNum - 1].count(Val: LoopVal)) |
1090 | // The name is defined in the previous stage. |
1091 | PrevVal = VRMap[StageNum - 1][LoopVal]; |
1092 | else if (VRMap[StageNum].count(Val: LoopVal)) |
1093 | // The previous name is defined in the current stage when the instruction |
1094 | // order is swapped. |
1095 | PrevVal = VRMap[StageNum][LoopVal]; |
1096 | else if (!LoopInst->isPHI() || LoopInst->getParent() != BB) |
1097 | // The loop value hasn't yet been scheduled. |
1098 | PrevVal = LoopVal; |
1099 | else if (StageNum == PhiStage + 1) |
1100 | // The loop value is another phi, which has not been scheduled. |
1101 | PrevVal = getInitPhiReg(Phi&: *LoopInst, LoopBB: BB); |
1102 | else if (StageNum > PhiStage + 1 && LoopInst->getParent() == BB) |
1103 | // The loop value is another phi, which has been scheduled. |
1104 | PrevVal = |
1105 | getPrevMapVal(StageNum: StageNum - 1, PhiStage, LoopVal: getLoopPhiReg(Phi&: *LoopInst, LoopBB: BB), |
1106 | LoopStage, VRMap, BB); |
1107 | } |
1108 | return PrevVal; |
1109 | } |
1110 | |
1111 | /// Rewrite the Phi values in the specified block to use the mappings |
1112 | /// from the initial operand. Once the Phi is scheduled, we switch |
1113 | /// to using the loop value instead of the Phi value, so those names |
1114 | /// do not need to be rewritten. |
1115 | void ModuloScheduleExpander::rewritePhiValues(MachineBasicBlock *NewBB, |
1116 | unsigned StageNum, |
1117 | ValueMapTy *VRMap, |
1118 | InstrMapTy &InstrMap) { |
1119 | for (auto &PHI : BB->phis()) { |
1120 | unsigned InitVal = 0; |
1121 | unsigned LoopVal = 0; |
1122 | getPhiRegs(Phi&: PHI, Loop: BB, InitVal, LoopVal); |
1123 | Register PhiDef = PHI.getOperand(i: 0).getReg(); |
1124 | |
1125 | unsigned PhiStage = (unsigned)Schedule.getStage(MI: MRI.getVRegDef(Reg: PhiDef)); |
1126 | unsigned LoopStage = (unsigned)Schedule.getStage(MI: MRI.getVRegDef(Reg: LoopVal)); |
1127 | unsigned NumPhis = getStagesForPhi(Reg: PhiDef); |
1128 | if (NumPhis > StageNum) |
1129 | NumPhis = StageNum; |
1130 | for (unsigned np = 0; np <= NumPhis; ++np) { |
1131 | unsigned NewVal = |
1132 | getPrevMapVal(StageNum: StageNum - np, PhiStage, LoopVal, LoopStage, VRMap, BB); |
1133 | if (!NewVal) |
1134 | NewVal = InitVal; |
1135 | rewriteScheduledInstr(BB: NewBB, InstrMap, CurStageNum: StageNum - np, PhiNum: np, Phi: &PHI, OldReg: PhiDef, |
1136 | NewReg: NewVal); |
1137 | } |
1138 | } |
1139 | } |
1140 | |
1141 | /// Rewrite a previously scheduled instruction to use the register value |
1142 | /// from the new instruction. Make sure the instruction occurs in the |
1143 | /// basic block, and we don't change the uses in the new instruction. |
1144 | void ModuloScheduleExpander::rewriteScheduledInstr( |
1145 | MachineBasicBlock *BB, InstrMapTy &InstrMap, unsigned CurStageNum, |
1146 | unsigned PhiNum, MachineInstr *Phi, unsigned OldReg, unsigned NewReg, |
1147 | unsigned PrevReg) { |
1148 | bool InProlog = (CurStageNum < (unsigned)Schedule.getNumStages() - 1); |
1149 | int StagePhi = Schedule.getStage(MI: Phi) + PhiNum; |
1150 | // Rewrite uses that have been scheduled already to use the new |
1151 | // Phi register. |
1152 | for (MachineOperand &UseOp : |
1153 | llvm::make_early_inc_range(Range: MRI.use_operands(Reg: OldReg))) { |
1154 | MachineInstr *UseMI = UseOp.getParent(); |
1155 | if (UseMI->getParent() != BB) |
1156 | continue; |
1157 | if (UseMI->isPHI()) { |
1158 | if (!Phi->isPHI() && UseMI->getOperand(i: 0).getReg() == NewReg) |
1159 | continue; |
1160 | if (getLoopPhiReg(Phi&: *UseMI, LoopBB: BB) != OldReg) |
1161 | continue; |
1162 | } |
1163 | InstrMapTy::iterator OrigInstr = InstrMap.find(Val: UseMI); |
1164 | assert(OrigInstr != InstrMap.end() && "Instruction not scheduled." ); |
1165 | MachineInstr *OrigMI = OrigInstr->second; |
1166 | int StageSched = Schedule.getStage(MI: OrigMI); |
1167 | int CycleSched = Schedule.getCycle(MI: OrigMI); |
1168 | unsigned ReplaceReg = 0; |
1169 | // This is the stage for the scheduled instruction. |
1170 | if (StagePhi == StageSched && Phi->isPHI()) { |
1171 | int CyclePhi = Schedule.getCycle(MI: Phi); |
1172 | if (PrevReg && InProlog) |
1173 | ReplaceReg = PrevReg; |
1174 | else if (PrevReg && !isLoopCarried(Phi&: *Phi) && |
1175 | (CyclePhi <= CycleSched || OrigMI->isPHI())) |
1176 | ReplaceReg = PrevReg; |
1177 | else |
1178 | ReplaceReg = NewReg; |
1179 | } |
1180 | // The scheduled instruction occurs before the scheduled Phi, and the |
1181 | // Phi is not loop carried. |
1182 | if (!InProlog && StagePhi + 1 == StageSched && !isLoopCarried(Phi&: *Phi)) |
1183 | ReplaceReg = NewReg; |
1184 | if (StagePhi > StageSched && Phi->isPHI()) |
1185 | ReplaceReg = NewReg; |
1186 | if (!InProlog && !Phi->isPHI() && StagePhi < StageSched) |
1187 | ReplaceReg = NewReg; |
1188 | if (ReplaceReg) { |
1189 | const TargetRegisterClass *NRC = |
1190 | MRI.constrainRegClass(Reg: ReplaceReg, RC: MRI.getRegClass(Reg: OldReg)); |
1191 | if (NRC) |
1192 | UseOp.setReg(ReplaceReg); |
1193 | else { |
1194 | Register SplitReg = MRI.createVirtualRegister(RegClass: MRI.getRegClass(Reg: OldReg)); |
1195 | BuildMI(BB&: *BB, I: UseMI, MIMD: UseMI->getDebugLoc(), MCID: TII->get(Opcode: TargetOpcode::COPY), |
1196 | DestReg: SplitReg) |
1197 | .addReg(RegNo: ReplaceReg); |
1198 | UseOp.setReg(SplitReg); |
1199 | } |
1200 | } |
1201 | } |
1202 | } |
1203 | |
1204 | bool ModuloScheduleExpander::isLoopCarried(MachineInstr &Phi) { |
1205 | if (!Phi.isPHI()) |
1206 | return false; |
1207 | int DefCycle = Schedule.getCycle(MI: &Phi); |
1208 | int DefStage = Schedule.getStage(MI: &Phi); |
1209 | |
1210 | unsigned InitVal = 0; |
1211 | unsigned LoopVal = 0; |
1212 | getPhiRegs(Phi, Loop: Phi.getParent(), InitVal, LoopVal); |
1213 | MachineInstr *Use = MRI.getVRegDef(Reg: LoopVal); |
1214 | if (!Use || Use->isPHI()) |
1215 | return true; |
1216 | int LoopCycle = Schedule.getCycle(MI: Use); |
1217 | int LoopStage = Schedule.getStage(MI: Use); |
1218 | return (LoopCycle > DefCycle) || (LoopStage <= DefStage); |
1219 | } |
1220 | |
1221 | //===----------------------------------------------------------------------===// |
1222 | // PeelingModuloScheduleExpander implementation |
1223 | //===----------------------------------------------------------------------===// |
1224 | // This is a reimplementation of ModuloScheduleExpander that works by creating |
1225 | // a fully correct steady-state kernel and peeling off the prolog and epilogs. |
1226 | //===----------------------------------------------------------------------===// |
1227 | |
1228 | namespace { |
1229 | // Remove any dead phis in MBB. Dead phis either have only one block as input |
1230 | // (in which case they are the identity) or have no uses. |
1231 | void EliminateDeadPhis(MachineBasicBlock *MBB, MachineRegisterInfo &MRI, |
1232 | LiveIntervals *LIS, bool KeepSingleSrcPhi = false) { |
1233 | bool Changed = true; |
1234 | while (Changed) { |
1235 | Changed = false; |
1236 | for (MachineInstr &MI : llvm::make_early_inc_range(Range: MBB->phis())) { |
1237 | assert(MI.isPHI()); |
1238 | if (MRI.use_empty(RegNo: MI.getOperand(i: 0).getReg())) { |
1239 | if (LIS) |
1240 | LIS->RemoveMachineInstrFromMaps(MI); |
1241 | MI.eraseFromParent(); |
1242 | Changed = true; |
1243 | } else if (!KeepSingleSrcPhi && MI.getNumExplicitOperands() == 3) { |
1244 | const TargetRegisterClass *ConstrainRegClass = |
1245 | MRI.constrainRegClass(Reg: MI.getOperand(i: 1).getReg(), |
1246 | RC: MRI.getRegClass(Reg: MI.getOperand(i: 0).getReg())); |
1247 | assert(ConstrainRegClass && |
1248 | "Expected a valid constrained register class!" ); |
1249 | (void)ConstrainRegClass; |
1250 | MRI.replaceRegWith(FromReg: MI.getOperand(i: 0).getReg(), |
1251 | ToReg: MI.getOperand(i: 1).getReg()); |
1252 | if (LIS) |
1253 | LIS->RemoveMachineInstrFromMaps(MI); |
1254 | MI.eraseFromParent(); |
1255 | Changed = true; |
1256 | } |
1257 | } |
1258 | } |
1259 | } |
1260 | |
1261 | /// Rewrites the kernel block in-place to adhere to the given schedule. |
1262 | /// KernelRewriter holds all of the state required to perform the rewriting. |
1263 | class KernelRewriter { |
1264 | ModuloSchedule &S; |
1265 | MachineBasicBlock *BB; |
1266 | MachineBasicBlock *, *ExitBB; |
1267 | MachineRegisterInfo &MRI; |
1268 | const TargetInstrInfo *TII; |
1269 | LiveIntervals *LIS; |
1270 | |
1271 | // Map from register class to canonical undef register for that class. |
1272 | DenseMap<const TargetRegisterClass *, Register> Undefs; |
1273 | // Map from <LoopReg, InitReg> to phi register for all created phis. Note that |
1274 | // this map is only used when InitReg is non-undef. |
1275 | DenseMap<std::pair<unsigned, unsigned>, Register> Phis; |
1276 | // Map from LoopReg to phi register where the InitReg is undef. |
1277 | DenseMap<Register, Register> UndefPhis; |
1278 | |
1279 | // Reg is used by MI. Return the new register MI should use to adhere to the |
1280 | // schedule. Insert phis as necessary. |
1281 | Register remapUse(Register Reg, MachineInstr &MI); |
1282 | // Insert a phi that carries LoopReg from the loop body and InitReg otherwise. |
1283 | // If InitReg is not given it is chosen arbitrarily. It will either be undef |
1284 | // or will be chosen so as to share another phi. |
1285 | Register phi(Register LoopReg, std::optional<Register> InitReg = {}, |
1286 | const TargetRegisterClass *RC = nullptr); |
1287 | // Create an undef register of the given register class. |
1288 | Register undef(const TargetRegisterClass *RC); |
1289 | |
1290 | public: |
1291 | KernelRewriter(MachineLoop &L, ModuloSchedule &S, MachineBasicBlock *LoopBB, |
1292 | LiveIntervals *LIS = nullptr); |
1293 | void rewrite(); |
1294 | }; |
1295 | } // namespace |
1296 | |
1297 | KernelRewriter::KernelRewriter(MachineLoop &L, ModuloSchedule &S, |
1298 | MachineBasicBlock *LoopBB, LiveIntervals *LIS) |
1299 | : S(S), BB(LoopBB), PreheaderBB(L.getLoopPreheader()), |
1300 | ExitBB(L.getExitBlock()), MRI(BB->getParent()->getRegInfo()), |
1301 | TII(BB->getParent()->getSubtarget().getInstrInfo()), LIS(LIS) { |
1302 | PreheaderBB = *BB->pred_begin(); |
1303 | if (PreheaderBB == BB) |
1304 | PreheaderBB = *std::next(x: BB->pred_begin()); |
1305 | } |
1306 | |
1307 | void KernelRewriter::rewrite() { |
1308 | // Rearrange the loop to be in schedule order. Note that the schedule may |
1309 | // contain instructions that are not owned by the loop block (InstrChanges and |
1310 | // friends), so we gracefully handle unowned instructions and delete any |
1311 | // instructions that weren't in the schedule. |
1312 | auto InsertPt = BB->getFirstTerminator(); |
1313 | MachineInstr *FirstMI = nullptr; |
1314 | for (MachineInstr *MI : S.getInstructions()) { |
1315 | if (MI->isPHI()) |
1316 | continue; |
1317 | if (MI->getParent()) |
1318 | MI->removeFromParent(); |
1319 | BB->insert(I: InsertPt, MI); |
1320 | if (!FirstMI) |
1321 | FirstMI = MI; |
1322 | } |
1323 | assert(FirstMI && "Failed to find first MI in schedule" ); |
1324 | |
1325 | // At this point all of the scheduled instructions are between FirstMI |
1326 | // and the end of the block. Kill from the first non-phi to FirstMI. |
1327 | for (auto I = BB->getFirstNonPHI(); I != FirstMI->getIterator();) { |
1328 | if (LIS) |
1329 | LIS->RemoveMachineInstrFromMaps(MI&: *I); |
1330 | (I++)->eraseFromParent(); |
1331 | } |
1332 | |
1333 | // Now remap every instruction in the loop. |
1334 | for (MachineInstr &MI : *BB) { |
1335 | if (MI.isPHI() || MI.isTerminator()) |
1336 | continue; |
1337 | for (MachineOperand &MO : MI.uses()) { |
1338 | if (!MO.isReg() || MO.getReg().isPhysical() || MO.isImplicit()) |
1339 | continue; |
1340 | Register Reg = remapUse(Reg: MO.getReg(), MI); |
1341 | MO.setReg(Reg); |
1342 | } |
1343 | } |
1344 | EliminateDeadPhis(MBB: BB, MRI, LIS); |
1345 | |
1346 | // Ensure a phi exists for all instructions that are either referenced by |
1347 | // an illegal phi or by an instruction outside the loop. This allows us to |
1348 | // treat remaps of these values the same as "normal" values that come from |
1349 | // loop-carried phis. |
1350 | for (auto MI = BB->getFirstNonPHI(); MI != BB->end(); ++MI) { |
1351 | if (MI->isPHI()) { |
1352 | Register R = MI->getOperand(i: 0).getReg(); |
1353 | phi(LoopReg: R); |
1354 | continue; |
1355 | } |
1356 | |
1357 | for (MachineOperand &Def : MI->defs()) { |
1358 | for (MachineInstr &MI : MRI.use_instructions(Reg: Def.getReg())) { |
1359 | if (MI.getParent() != BB) { |
1360 | phi(LoopReg: Def.getReg()); |
1361 | break; |
1362 | } |
1363 | } |
1364 | } |
1365 | } |
1366 | } |
1367 | |
1368 | Register KernelRewriter::remapUse(Register Reg, MachineInstr &MI) { |
1369 | MachineInstr *Producer = MRI.getUniqueVRegDef(Reg); |
1370 | if (!Producer) |
1371 | return Reg; |
1372 | |
1373 | int ConsumerStage = S.getStage(MI: &MI); |
1374 | if (!Producer->isPHI()) { |
1375 | // Non-phi producers are simple to remap. Insert as many phis as the |
1376 | // difference between the consumer and producer stages. |
1377 | if (Producer->getParent() != BB) |
1378 | // Producer was not inside the loop. Use the register as-is. |
1379 | return Reg; |
1380 | int ProducerStage = S.getStage(MI: Producer); |
1381 | assert(ConsumerStage != -1 && |
1382 | "In-loop consumer should always be scheduled!" ); |
1383 | assert(ConsumerStage >= ProducerStage); |
1384 | unsigned StageDiff = ConsumerStage - ProducerStage; |
1385 | |
1386 | for (unsigned I = 0; I < StageDiff; ++I) |
1387 | Reg = phi(LoopReg: Reg); |
1388 | return Reg; |
1389 | } |
1390 | |
1391 | // First, dive through the phi chain to find the defaults for the generated |
1392 | // phis. |
1393 | SmallVector<std::optional<Register>, 4> Defaults; |
1394 | Register LoopReg = Reg; |
1395 | auto LoopProducer = Producer; |
1396 | while (LoopProducer->isPHI() && LoopProducer->getParent() == BB) { |
1397 | LoopReg = getLoopPhiReg(Phi&: *LoopProducer, LoopBB: BB); |
1398 | Defaults.emplace_back(Args: getInitPhiReg(Phi&: *LoopProducer, LoopBB: BB)); |
1399 | LoopProducer = MRI.getUniqueVRegDef(Reg: LoopReg); |
1400 | assert(LoopProducer); |
1401 | } |
1402 | int LoopProducerStage = S.getStage(MI: LoopProducer); |
1403 | |
1404 | std::optional<Register> IllegalPhiDefault; |
1405 | |
1406 | if (LoopProducerStage == -1) { |
1407 | // Do nothing. |
1408 | } else if (LoopProducerStage > ConsumerStage) { |
1409 | // This schedule is only representable if ProducerStage == ConsumerStage+1. |
1410 | // In addition, Consumer's cycle must be scheduled after Producer in the |
1411 | // rescheduled loop. This is enforced by the pipeliner's ASAP and ALAP |
1412 | // functions. |
1413 | #ifndef NDEBUG // Silence unused variables in non-asserts mode. |
1414 | int LoopProducerCycle = S.getCycle(LoopProducer); |
1415 | int ConsumerCycle = S.getCycle(&MI); |
1416 | #endif |
1417 | assert(LoopProducerCycle <= ConsumerCycle); |
1418 | assert(LoopProducerStage == ConsumerStage + 1); |
1419 | // Peel off the first phi from Defaults and insert a phi between producer |
1420 | // and consumer. This phi will not be at the front of the block so we |
1421 | // consider it illegal. It will only exist during the rewrite process; it |
1422 | // needs to exist while we peel off prologs because these could take the |
1423 | // default value. After that we can replace all uses with the loop producer |
1424 | // value. |
1425 | IllegalPhiDefault = Defaults.front(); |
1426 | Defaults.erase(CI: Defaults.begin()); |
1427 | } else { |
1428 | assert(ConsumerStage >= LoopProducerStage); |
1429 | int StageDiff = ConsumerStage - LoopProducerStage; |
1430 | if (StageDiff > 0) { |
1431 | LLVM_DEBUG(dbgs() << " -- padding defaults array from " << Defaults.size() |
1432 | << " to " << (Defaults.size() + StageDiff) << "\n" ); |
1433 | // If we need more phis than we have defaults for, pad out with undefs for |
1434 | // the earliest phis, which are at the end of the defaults chain (the |
1435 | // chain is in reverse order). |
1436 | Defaults.resize(N: Defaults.size() + StageDiff, |
1437 | NV: Defaults.empty() ? std::optional<Register>() |
1438 | : Defaults.back()); |
1439 | } |
1440 | } |
1441 | |
1442 | // Now we know the number of stages to jump back, insert the phi chain. |
1443 | auto DefaultI = Defaults.rbegin(); |
1444 | while (DefaultI != Defaults.rend()) |
1445 | LoopReg = phi(LoopReg, InitReg: *DefaultI++, RC: MRI.getRegClass(Reg)); |
1446 | |
1447 | if (IllegalPhiDefault) { |
1448 | // The consumer optionally consumes LoopProducer in the same iteration |
1449 | // (because the producer is scheduled at an earlier cycle than the consumer) |
1450 | // or the initial value. To facilitate this we create an illegal block here |
1451 | // by embedding a phi in the middle of the block. We will fix this up |
1452 | // immediately prior to pruning. |
1453 | auto RC = MRI.getRegClass(Reg); |
1454 | Register R = MRI.createVirtualRegister(RegClass: RC); |
1455 | MachineInstr *IllegalPhi = |
1456 | BuildMI(BB&: *BB, I&: MI, MIMD: DebugLoc(), MCID: TII->get(Opcode: TargetOpcode::PHI), DestReg: R) |
1457 | .addReg(RegNo: *IllegalPhiDefault) |
1458 | .addMBB(MBB: PreheaderBB) // Block choice is arbitrary and has no effect. |
1459 | .addReg(RegNo: LoopReg) |
1460 | .addMBB(MBB: BB); // Block choice is arbitrary and has no effect. |
1461 | // Illegal phi should belong to the producer stage so that it can be |
1462 | // filtered correctly during peeling. |
1463 | S.setStage(MI: IllegalPhi, MIStage: LoopProducerStage); |
1464 | return R; |
1465 | } |
1466 | |
1467 | return LoopReg; |
1468 | } |
1469 | |
1470 | Register KernelRewriter::phi(Register LoopReg, std::optional<Register> InitReg, |
1471 | const TargetRegisterClass *RC) { |
1472 | // If the init register is not undef, try and find an existing phi. |
1473 | if (InitReg) { |
1474 | auto I = Phis.find(Val: {LoopReg, *InitReg}); |
1475 | if (I != Phis.end()) |
1476 | return I->second; |
1477 | } else { |
1478 | for (auto &KV : Phis) { |
1479 | if (KV.first.first == LoopReg) |
1480 | return KV.second; |
1481 | } |
1482 | } |
1483 | |
1484 | // InitReg is either undef or no existing phi takes InitReg as input. Try and |
1485 | // find a phi that takes undef as input. |
1486 | auto I = UndefPhis.find(Val: LoopReg); |
1487 | if (I != UndefPhis.end()) { |
1488 | Register R = I->second; |
1489 | if (!InitReg) |
1490 | // Found a phi taking undef as input, and this input is undef so return |
1491 | // without any more changes. |
1492 | return R; |
1493 | // Found a phi taking undef as input, so rewrite it to take InitReg. |
1494 | MachineInstr *MI = MRI.getVRegDef(Reg: R); |
1495 | MI->getOperand(i: 1).setReg(*InitReg); |
1496 | Phis.insert(KV: {{LoopReg, *InitReg}, R}); |
1497 | const TargetRegisterClass *ConstrainRegClass = |
1498 | MRI.constrainRegClass(Reg: R, RC: MRI.getRegClass(Reg: *InitReg)); |
1499 | assert(ConstrainRegClass && "Expected a valid constrained register class!" ); |
1500 | (void)ConstrainRegClass; |
1501 | UndefPhis.erase(I); |
1502 | return R; |
1503 | } |
1504 | |
1505 | // Failed to find any existing phi to reuse, so create a new one. |
1506 | if (!RC) |
1507 | RC = MRI.getRegClass(Reg: LoopReg); |
1508 | Register R = MRI.createVirtualRegister(RegClass: RC); |
1509 | if (InitReg) { |
1510 | const TargetRegisterClass *ConstrainRegClass = |
1511 | MRI.constrainRegClass(Reg: R, RC: MRI.getRegClass(Reg: *InitReg)); |
1512 | assert(ConstrainRegClass && "Expected a valid constrained register class!" ); |
1513 | (void)ConstrainRegClass; |
1514 | } |
1515 | BuildMI(BB&: *BB, I: BB->getFirstNonPHI(), MIMD: DebugLoc(), MCID: TII->get(Opcode: TargetOpcode::PHI), DestReg: R) |
1516 | .addReg(RegNo: InitReg ? *InitReg : undef(RC)) |
1517 | .addMBB(MBB: PreheaderBB) |
1518 | .addReg(RegNo: LoopReg) |
1519 | .addMBB(MBB: BB); |
1520 | if (!InitReg) |
1521 | UndefPhis[LoopReg] = R; |
1522 | else |
1523 | Phis[{LoopReg, *InitReg}] = R; |
1524 | return R; |
1525 | } |
1526 | |
1527 | Register KernelRewriter::undef(const TargetRegisterClass *RC) { |
1528 | Register &R = Undefs[RC]; |
1529 | if (R == 0) { |
1530 | // Create an IMPLICIT_DEF that defines this register if we need it. |
1531 | // All uses of this should be removed by the time we have finished unrolling |
1532 | // prologs and epilogs. |
1533 | R = MRI.createVirtualRegister(RegClass: RC); |
1534 | auto *InsertBB = &PreheaderBB->getParent()->front(); |
1535 | BuildMI(BB&: *InsertBB, I: InsertBB->getFirstTerminator(), MIMD: DebugLoc(), |
1536 | MCID: TII->get(Opcode: TargetOpcode::IMPLICIT_DEF), DestReg: R); |
1537 | } |
1538 | return R; |
1539 | } |
1540 | |
1541 | namespace { |
1542 | /// Describes an operand in the kernel of a pipelined loop. Characteristics of |
1543 | /// the operand are discovered, such as how many in-loop PHIs it has to jump |
1544 | /// through and defaults for these phis. |
1545 | class KernelOperandInfo { |
1546 | MachineBasicBlock *BB; |
1547 | MachineRegisterInfo &MRI; |
1548 | SmallVector<Register, 4> PhiDefaults; |
1549 | MachineOperand *Source; |
1550 | MachineOperand *Target; |
1551 | |
1552 | public: |
1553 | KernelOperandInfo(MachineOperand *MO, MachineRegisterInfo &MRI, |
1554 | const SmallPtrSetImpl<MachineInstr *> &IllegalPhis) |
1555 | : MRI(MRI) { |
1556 | Source = MO; |
1557 | BB = MO->getParent()->getParent(); |
1558 | while (isRegInLoop(MO)) { |
1559 | MachineInstr *MI = MRI.getVRegDef(Reg: MO->getReg()); |
1560 | if (MI->isFullCopy()) { |
1561 | MO = &MI->getOperand(i: 1); |
1562 | continue; |
1563 | } |
1564 | if (!MI->isPHI()) |
1565 | break; |
1566 | // If this is an illegal phi, don't count it in distance. |
1567 | if (IllegalPhis.count(Ptr: MI)) { |
1568 | MO = &MI->getOperand(i: 3); |
1569 | continue; |
1570 | } |
1571 | |
1572 | Register Default = getInitPhiReg(Phi&: *MI, LoopBB: BB); |
1573 | MO = MI->getOperand(i: 2).getMBB() == BB ? &MI->getOperand(i: 1) |
1574 | : &MI->getOperand(i: 3); |
1575 | PhiDefaults.push_back(Elt: Default); |
1576 | } |
1577 | Target = MO; |
1578 | } |
1579 | |
1580 | bool operator==(const KernelOperandInfo &Other) const { |
1581 | return PhiDefaults.size() == Other.PhiDefaults.size(); |
1582 | } |
1583 | |
1584 | void print(raw_ostream &OS) const { |
1585 | OS << "use of " << *Source << ": distance(" << PhiDefaults.size() << ") in " |
1586 | << *Source->getParent(); |
1587 | } |
1588 | |
1589 | private: |
1590 | bool isRegInLoop(MachineOperand *MO) { |
1591 | return MO->isReg() && MO->getReg().isVirtual() && |
1592 | MRI.getVRegDef(Reg: MO->getReg())->getParent() == BB; |
1593 | } |
1594 | }; |
1595 | } // namespace |
1596 | |
1597 | MachineBasicBlock * |
1598 | PeelingModuloScheduleExpander::peelKernel(LoopPeelDirection LPD) { |
1599 | MachineBasicBlock *NewBB = PeelSingleBlockLoop(Direction: LPD, Loop: BB, MRI, TII); |
1600 | if (LPD == LPD_Front) |
1601 | PeeledFront.push_back(x: NewBB); |
1602 | else |
1603 | PeeledBack.push_front(x: NewBB); |
1604 | for (auto I = BB->begin(), NI = NewBB->begin(); !I->isTerminator(); |
1605 | ++I, ++NI) { |
1606 | CanonicalMIs[&*I] = &*I; |
1607 | CanonicalMIs[&*NI] = &*I; |
1608 | BlockMIs[{NewBB, &*I}] = &*NI; |
1609 | BlockMIs[{BB, &*I}] = &*I; |
1610 | } |
1611 | return NewBB; |
1612 | } |
1613 | |
1614 | void PeelingModuloScheduleExpander::filterInstructions(MachineBasicBlock *MB, |
1615 | int MinStage) { |
1616 | for (auto I = MB->getFirstInstrTerminator()->getReverseIterator(); |
1617 | I != std::next(x: MB->getFirstNonPHI()->getReverseIterator());) { |
1618 | MachineInstr *MI = &*I++; |
1619 | int Stage = getStage(MI); |
1620 | if (Stage == -1 || Stage >= MinStage) |
1621 | continue; |
1622 | |
1623 | for (MachineOperand &DefMO : MI->defs()) { |
1624 | SmallVector<std::pair<MachineInstr *, Register>, 4> Subs; |
1625 | for (MachineInstr &UseMI : MRI.use_instructions(Reg: DefMO.getReg())) { |
1626 | // Only PHIs can use values from this block by construction. |
1627 | // Match with the equivalent PHI in B. |
1628 | assert(UseMI.isPHI()); |
1629 | Register Reg = getEquivalentRegisterIn(Reg: UseMI.getOperand(i: 0).getReg(), |
1630 | BB: MI->getParent()); |
1631 | Subs.emplace_back(Args: &UseMI, Args&: Reg); |
1632 | } |
1633 | for (auto &Sub : Subs) |
1634 | Sub.first->substituteRegister(FromReg: DefMO.getReg(), ToReg: Sub.second, /*SubIdx=*/0, |
1635 | RegInfo: *MRI.getTargetRegisterInfo()); |
1636 | } |
1637 | if (LIS) |
1638 | LIS->RemoveMachineInstrFromMaps(MI&: *MI); |
1639 | MI->eraseFromParent(); |
1640 | } |
1641 | } |
1642 | |
1643 | void PeelingModuloScheduleExpander::moveStageBetweenBlocks( |
1644 | MachineBasicBlock *DestBB, MachineBasicBlock *SourceBB, unsigned Stage) { |
1645 | auto InsertPt = DestBB->getFirstNonPHI(); |
1646 | DenseMap<Register, Register> Remaps; |
1647 | for (MachineInstr &MI : llvm::make_early_inc_range( |
1648 | Range: llvm::make_range(x: SourceBB->getFirstNonPHI(), y: SourceBB->end()))) { |
1649 | if (MI.isPHI()) { |
1650 | // This is an illegal PHI. If we move any instructions using an illegal |
1651 | // PHI, we need to create a legal Phi. |
1652 | if (getStage(MI: &MI) != Stage) { |
1653 | // The legal Phi is not necessary if the illegal phi's stage |
1654 | // is being moved. |
1655 | Register PhiR = MI.getOperand(i: 0).getReg(); |
1656 | auto RC = MRI.getRegClass(Reg: PhiR); |
1657 | Register NR = MRI.createVirtualRegister(RegClass: RC); |
1658 | MachineInstr *NI = BuildMI(BB&: *DestBB, I: DestBB->getFirstNonPHI(), |
1659 | MIMD: DebugLoc(), MCID: TII->get(Opcode: TargetOpcode::PHI), DestReg: NR) |
1660 | .addReg(RegNo: PhiR) |
1661 | .addMBB(MBB: SourceBB); |
1662 | BlockMIs[{DestBB, CanonicalMIs[&MI]}] = NI; |
1663 | CanonicalMIs[NI] = CanonicalMIs[&MI]; |
1664 | Remaps[PhiR] = NR; |
1665 | } |
1666 | } |
1667 | if (getStage(MI: &MI) != Stage) |
1668 | continue; |
1669 | MI.removeFromParent(); |
1670 | DestBB->insert(I: InsertPt, MI: &MI); |
1671 | auto *KernelMI = CanonicalMIs[&MI]; |
1672 | BlockMIs[{DestBB, KernelMI}] = &MI; |
1673 | BlockMIs.erase(Val: {SourceBB, KernelMI}); |
1674 | } |
1675 | SmallVector<MachineInstr *, 4> PhiToDelete; |
1676 | for (MachineInstr &MI : DestBB->phis()) { |
1677 | assert(MI.getNumOperands() == 3); |
1678 | MachineInstr *Def = MRI.getVRegDef(Reg: MI.getOperand(i: 1).getReg()); |
1679 | // If the instruction referenced by the phi is moved inside the block |
1680 | // we don't need the phi anymore. |
1681 | if (getStage(MI: Def) == Stage) { |
1682 | Register PhiReg = MI.getOperand(i: 0).getReg(); |
1683 | assert(Def->findRegisterDefOperandIdx(MI.getOperand(1).getReg(), |
1684 | /*TRI=*/nullptr) != -1); |
1685 | MRI.replaceRegWith(FromReg: MI.getOperand(i: 0).getReg(), ToReg: MI.getOperand(i: 1).getReg()); |
1686 | MI.getOperand(i: 0).setReg(PhiReg); |
1687 | PhiToDelete.push_back(Elt: &MI); |
1688 | } |
1689 | } |
1690 | for (auto *P : PhiToDelete) |
1691 | P->eraseFromParent(); |
1692 | InsertPt = DestBB->getFirstNonPHI(); |
1693 | // Helper to clone Phi instructions into the destination block. We clone Phi |
1694 | // greedily to avoid combinatorial explosion of Phi instructions. |
1695 | auto clonePhi = [&](MachineInstr *Phi) { |
1696 | MachineInstr *NewMI = MF.CloneMachineInstr(Orig: Phi); |
1697 | DestBB->insert(I: InsertPt, MI: NewMI); |
1698 | Register OrigR = Phi->getOperand(i: 0).getReg(); |
1699 | Register R = MRI.createVirtualRegister(RegClass: MRI.getRegClass(Reg: OrigR)); |
1700 | NewMI->getOperand(i: 0).setReg(R); |
1701 | NewMI->getOperand(i: 1).setReg(OrigR); |
1702 | NewMI->getOperand(i: 2).setMBB(*DestBB->pred_begin()); |
1703 | Remaps[OrigR] = R; |
1704 | CanonicalMIs[NewMI] = CanonicalMIs[Phi]; |
1705 | BlockMIs[{DestBB, CanonicalMIs[Phi]}] = NewMI; |
1706 | PhiNodeLoopIteration[NewMI] = PhiNodeLoopIteration[Phi]; |
1707 | return R; |
1708 | }; |
1709 | for (auto I = DestBB->getFirstNonPHI(); I != DestBB->end(); ++I) { |
1710 | for (MachineOperand &MO : I->uses()) { |
1711 | if (!MO.isReg()) |
1712 | continue; |
1713 | if (Remaps.count(Val: MO.getReg())) |
1714 | MO.setReg(Remaps[MO.getReg()]); |
1715 | else { |
1716 | // If we are using a phi from the source block we need to add a new phi |
1717 | // pointing to the old one. |
1718 | MachineInstr *Use = MRI.getUniqueVRegDef(Reg: MO.getReg()); |
1719 | if (Use && Use->isPHI() && Use->getParent() == SourceBB) { |
1720 | Register R = clonePhi(Use); |
1721 | MO.setReg(R); |
1722 | } |
1723 | } |
1724 | } |
1725 | } |
1726 | } |
1727 | |
1728 | Register |
1729 | PeelingModuloScheduleExpander::getPhiCanonicalReg(MachineInstr *CanonicalPhi, |
1730 | MachineInstr *Phi) { |
1731 | unsigned distance = PhiNodeLoopIteration[Phi]; |
1732 | MachineInstr *CanonicalUse = CanonicalPhi; |
1733 | Register CanonicalUseReg = CanonicalUse->getOperand(i: 0).getReg(); |
1734 | for (unsigned I = 0; I < distance; ++I) { |
1735 | assert(CanonicalUse->isPHI()); |
1736 | assert(CanonicalUse->getNumOperands() == 5); |
1737 | unsigned LoopRegIdx = 3, InitRegIdx = 1; |
1738 | if (CanonicalUse->getOperand(i: 2).getMBB() == CanonicalUse->getParent()) |
1739 | std::swap(a&: LoopRegIdx, b&: InitRegIdx); |
1740 | CanonicalUseReg = CanonicalUse->getOperand(i: LoopRegIdx).getReg(); |
1741 | CanonicalUse = MRI.getVRegDef(Reg: CanonicalUseReg); |
1742 | } |
1743 | return CanonicalUseReg; |
1744 | } |
1745 | |
1746 | void PeelingModuloScheduleExpander::peelPrologAndEpilogs() { |
1747 | BitVector LS(Schedule.getNumStages(), true); |
1748 | BitVector AS(Schedule.getNumStages(), true); |
1749 | LiveStages[BB] = LS; |
1750 | AvailableStages[BB] = AS; |
1751 | |
1752 | // Peel out the prologs. |
1753 | LS.reset(); |
1754 | for (int I = 0; I < Schedule.getNumStages() - 1; ++I) { |
1755 | LS[I] = true; |
1756 | Prologs.push_back(Elt: peelKernel(LPD: LPD_Front)); |
1757 | LiveStages[Prologs.back()] = LS; |
1758 | AvailableStages[Prologs.back()] = LS; |
1759 | } |
1760 | |
1761 | // Create a block that will end up as the new loop exiting block (dominated by |
1762 | // all prologs and epilogs). It will only contain PHIs, in the same order as |
1763 | // BB's PHIs. This gives us a poor-man's LCSSA with the inductive property |
1764 | // that the exiting block is a (sub) clone of BB. This in turn gives us the |
1765 | // property that any value deffed in BB but used outside of BB is used by a |
1766 | // PHI in the exiting block. |
1767 | MachineBasicBlock *ExitingBB = CreateLCSSAExitingBlock(); |
1768 | EliminateDeadPhis(MBB: ExitingBB, MRI, LIS, /*KeepSingleSrcPhi=*/true); |
1769 | // Push out the epilogs, again in reverse order. |
1770 | // We can't assume anything about the minumum loop trip count at this point, |
1771 | // so emit a fairly complex epilog. |
1772 | |
1773 | // We first peel number of stages minus one epilogue. Then we remove dead |
1774 | // stages and reorder instructions based on their stage. If we have 3 stages |
1775 | // we generate first: |
1776 | // E0[3, 2, 1] |
1777 | // E1[3', 2'] |
1778 | // E2[3''] |
1779 | // And then we move instructions based on their stages to have: |
1780 | // E0[3] |
1781 | // E1[2, 3'] |
1782 | // E2[1, 2', 3''] |
1783 | // The transformation is legal because we only move instructions past |
1784 | // instructions of a previous loop iteration. |
1785 | for (int I = 1; I <= Schedule.getNumStages() - 1; ++I) { |
1786 | Epilogs.push_back(Elt: peelKernel(LPD: LPD_Back)); |
1787 | MachineBasicBlock *B = Epilogs.back(); |
1788 | filterInstructions(MB: B, MinStage: Schedule.getNumStages() - I); |
1789 | // Keep track at which iteration each phi belongs to. We need it to know |
1790 | // what version of the variable to use during prologue/epilogue stitching. |
1791 | EliminateDeadPhis(MBB: B, MRI, LIS, /*KeepSingleSrcPhi=*/true); |
1792 | for (MachineInstr &Phi : B->phis()) |
1793 | PhiNodeLoopIteration[&Phi] = Schedule.getNumStages() - I; |
1794 | } |
1795 | for (size_t I = 0; I < Epilogs.size(); I++) { |
1796 | LS.reset(); |
1797 | for (size_t J = I; J < Epilogs.size(); J++) { |
1798 | int Iteration = J; |
1799 | unsigned Stage = Schedule.getNumStages() - 1 + I - J; |
1800 | // Move stage one block at a time so that Phi nodes are updated correctly. |
1801 | for (size_t K = Iteration; K > I; K--) |
1802 | moveStageBetweenBlocks(DestBB: Epilogs[K - 1], SourceBB: Epilogs[K], Stage); |
1803 | LS[Stage] = true; |
1804 | } |
1805 | LiveStages[Epilogs[I]] = LS; |
1806 | AvailableStages[Epilogs[I]] = AS; |
1807 | } |
1808 | |
1809 | // Now we've defined all the prolog and epilog blocks as a fallthrough |
1810 | // sequence, add the edges that will be followed if the loop trip count is |
1811 | // lower than the number of stages (connecting prologs directly with epilogs). |
1812 | auto PI = Prologs.begin(); |
1813 | auto EI = Epilogs.begin(); |
1814 | assert(Prologs.size() == Epilogs.size()); |
1815 | for (; PI != Prologs.end(); ++PI, ++EI) { |
1816 | MachineBasicBlock *Pred = *(*EI)->pred_begin(); |
1817 | (*PI)->addSuccessor(Succ: *EI); |
1818 | for (MachineInstr &MI : (*EI)->phis()) { |
1819 | Register Reg = MI.getOperand(i: 1).getReg(); |
1820 | MachineInstr *Use = MRI.getUniqueVRegDef(Reg); |
1821 | if (Use && Use->getParent() == Pred) { |
1822 | MachineInstr *CanonicalUse = CanonicalMIs[Use]; |
1823 | if (CanonicalUse->isPHI()) { |
1824 | // If the use comes from a phi we need to skip as many phi as the |
1825 | // distance between the epilogue and the kernel. Trace through the phi |
1826 | // chain to find the right value. |
1827 | Reg = getPhiCanonicalReg(CanonicalPhi: CanonicalUse, Phi: Use); |
1828 | } |
1829 | Reg = getEquivalentRegisterIn(Reg, BB: *PI); |
1830 | } |
1831 | MI.addOperand(Op: MachineOperand::CreateReg(Reg, /*isDef=*/false)); |
1832 | MI.addOperand(Op: MachineOperand::CreateMBB(MBB: *PI)); |
1833 | } |
1834 | } |
1835 | |
1836 | // Create a list of all blocks in order. |
1837 | SmallVector<MachineBasicBlock *, 8> Blocks; |
1838 | llvm::copy(Range&: PeeledFront, Out: std::back_inserter(x&: Blocks)); |
1839 | Blocks.push_back(Elt: BB); |
1840 | llvm::copy(Range&: PeeledBack, Out: std::back_inserter(x&: Blocks)); |
1841 | |
1842 | // Iterate in reverse order over all instructions, remapping as we go. |
1843 | for (MachineBasicBlock *B : reverse(C&: Blocks)) { |
1844 | for (auto I = B->instr_rbegin(); |
1845 | I != std::next(x: B->getFirstNonPHI()->getReverseIterator());) { |
1846 | MachineBasicBlock::reverse_instr_iterator MI = I++; |
1847 | rewriteUsesOf(MI: &*MI); |
1848 | } |
1849 | } |
1850 | for (auto *MI : IllegalPhisToDelete) { |
1851 | if (LIS) |
1852 | LIS->RemoveMachineInstrFromMaps(MI&: *MI); |
1853 | MI->eraseFromParent(); |
1854 | } |
1855 | IllegalPhisToDelete.clear(); |
1856 | |
1857 | // Now all remapping has been done, we're free to optimize the generated code. |
1858 | for (MachineBasicBlock *B : reverse(C&: Blocks)) |
1859 | EliminateDeadPhis(MBB: B, MRI, LIS); |
1860 | EliminateDeadPhis(MBB: ExitingBB, MRI, LIS); |
1861 | } |
1862 | |
1863 | MachineBasicBlock *PeelingModuloScheduleExpander::CreateLCSSAExitingBlock() { |
1864 | MachineFunction &MF = *BB->getParent(); |
1865 | MachineBasicBlock *Exit = *BB->succ_begin(); |
1866 | if (Exit == BB) |
1867 | Exit = *std::next(x: BB->succ_begin()); |
1868 | |
1869 | MachineBasicBlock *NewBB = MF.CreateMachineBasicBlock(BB: BB->getBasicBlock()); |
1870 | MF.insert(MBBI: std::next(x: BB->getIterator()), MBB: NewBB); |
1871 | |
1872 | // Clone all phis in BB into NewBB and rewrite. |
1873 | for (MachineInstr &MI : BB->phis()) { |
1874 | auto RC = MRI.getRegClass(Reg: MI.getOperand(i: 0).getReg()); |
1875 | Register OldR = MI.getOperand(i: 3).getReg(); |
1876 | Register R = MRI.createVirtualRegister(RegClass: RC); |
1877 | SmallVector<MachineInstr *, 4> Uses; |
1878 | for (MachineInstr &Use : MRI.use_instructions(Reg: OldR)) |
1879 | if (Use.getParent() != BB) |
1880 | Uses.push_back(Elt: &Use); |
1881 | for (MachineInstr *Use : Uses) |
1882 | Use->substituteRegister(FromReg: OldR, ToReg: R, /*SubIdx=*/0, |
1883 | RegInfo: *MRI.getTargetRegisterInfo()); |
1884 | MachineInstr *NI = BuildMI(BB: NewBB, MIMD: DebugLoc(), MCID: TII->get(Opcode: TargetOpcode::PHI), DestReg: R) |
1885 | .addReg(RegNo: OldR) |
1886 | .addMBB(MBB: BB); |
1887 | BlockMIs[{NewBB, &MI}] = NI; |
1888 | CanonicalMIs[NI] = &MI; |
1889 | } |
1890 | BB->replaceSuccessor(Old: Exit, New: NewBB); |
1891 | Exit->replacePhiUsesWith(Old: BB, New: NewBB); |
1892 | NewBB->addSuccessor(Succ: Exit); |
1893 | |
1894 | MachineBasicBlock *TBB = nullptr, *FBB = nullptr; |
1895 | SmallVector<MachineOperand, 4> Cond; |
1896 | bool CanAnalyzeBr = !TII->analyzeBranch(MBB&: *BB, TBB, FBB, Cond); |
1897 | (void)CanAnalyzeBr; |
1898 | assert(CanAnalyzeBr && "Must be able to analyze the loop branch!" ); |
1899 | TII->removeBranch(MBB&: *BB); |
1900 | TII->insertBranch(MBB&: *BB, TBB: TBB == Exit ? NewBB : TBB, FBB: FBB == Exit ? NewBB : FBB, |
1901 | Cond, DL: DebugLoc()); |
1902 | TII->insertUnconditionalBranch(MBB&: *NewBB, DestBB: Exit, DL: DebugLoc()); |
1903 | return NewBB; |
1904 | } |
1905 | |
1906 | Register |
1907 | PeelingModuloScheduleExpander::getEquivalentRegisterIn(Register Reg, |
1908 | MachineBasicBlock *BB) { |
1909 | MachineInstr *MI = MRI.getUniqueVRegDef(Reg); |
1910 | unsigned OpIdx = MI->findRegisterDefOperandIdx(Reg, /*TRI=*/nullptr); |
1911 | return BlockMIs[{BB, CanonicalMIs[MI]}]->getOperand(i: OpIdx).getReg(); |
1912 | } |
1913 | |
1914 | void PeelingModuloScheduleExpander::rewriteUsesOf(MachineInstr *MI) { |
1915 | if (MI->isPHI()) { |
1916 | // This is an illegal PHI. The loop-carried (desired) value is operand 3, |
1917 | // and it is produced by this block. |
1918 | Register PhiR = MI->getOperand(i: 0).getReg(); |
1919 | Register R = MI->getOperand(i: 3).getReg(); |
1920 | int RMIStage = getStage(MI: MRI.getUniqueVRegDef(Reg: R)); |
1921 | if (RMIStage != -1 && !AvailableStages[MI->getParent()].test(Idx: RMIStage)) |
1922 | R = MI->getOperand(i: 1).getReg(); |
1923 | MRI.setRegClass(Reg: R, RC: MRI.getRegClass(Reg: PhiR)); |
1924 | MRI.replaceRegWith(FromReg: PhiR, ToReg: R); |
1925 | // Postpone deleting the Phi as it may be referenced by BlockMIs and used |
1926 | // later to figure out how to remap registers. |
1927 | MI->getOperand(i: 0).setReg(PhiR); |
1928 | IllegalPhisToDelete.push_back(Elt: MI); |
1929 | return; |
1930 | } |
1931 | |
1932 | int Stage = getStage(MI); |
1933 | if (Stage == -1 || LiveStages.count(Val: MI->getParent()) == 0 || |
1934 | LiveStages[MI->getParent()].test(Idx: Stage)) |
1935 | // Instruction is live, no rewriting to do. |
1936 | return; |
1937 | |
1938 | for (MachineOperand &DefMO : MI->defs()) { |
1939 | SmallVector<std::pair<MachineInstr *, Register>, 4> Subs; |
1940 | for (MachineInstr &UseMI : MRI.use_instructions(Reg: DefMO.getReg())) { |
1941 | // Only PHIs can use values from this block by construction. |
1942 | // Match with the equivalent PHI in B. |
1943 | assert(UseMI.isPHI()); |
1944 | Register Reg = getEquivalentRegisterIn(Reg: UseMI.getOperand(i: 0).getReg(), |
1945 | BB: MI->getParent()); |
1946 | Subs.emplace_back(Args: &UseMI, Args&: Reg); |
1947 | } |
1948 | for (auto &Sub : Subs) |
1949 | Sub.first->substituteRegister(FromReg: DefMO.getReg(), ToReg: Sub.second, /*SubIdx=*/0, |
1950 | RegInfo: *MRI.getTargetRegisterInfo()); |
1951 | } |
1952 | if (LIS) |
1953 | LIS->RemoveMachineInstrFromMaps(MI&: *MI); |
1954 | MI->eraseFromParent(); |
1955 | } |
1956 | |
1957 | void PeelingModuloScheduleExpander::fixupBranches() { |
1958 | // Work outwards from the kernel. |
1959 | bool KernelDisposed = false; |
1960 | int TC = Schedule.getNumStages() - 1; |
1961 | for (auto PI = Prologs.rbegin(), EI = Epilogs.rbegin(); PI != Prologs.rend(); |
1962 | ++PI, ++EI, --TC) { |
1963 | MachineBasicBlock *Prolog = *PI; |
1964 | MachineBasicBlock *Fallthrough = *Prolog->succ_begin(); |
1965 | MachineBasicBlock *Epilog = *EI; |
1966 | SmallVector<MachineOperand, 4> Cond; |
1967 | TII->removeBranch(MBB&: *Prolog); |
1968 | std::optional<bool> StaticallyGreater = |
1969 | LoopInfo->createTripCountGreaterCondition(TC, MBB&: *Prolog, Cond); |
1970 | if (!StaticallyGreater) { |
1971 | LLVM_DEBUG(dbgs() << "Dynamic: TC > " << TC << "\n" ); |
1972 | // Dynamically branch based on Cond. |
1973 | TII->insertBranch(MBB&: *Prolog, TBB: Epilog, FBB: Fallthrough, Cond, DL: DebugLoc()); |
1974 | } else if (*StaticallyGreater == false) { |
1975 | LLVM_DEBUG(dbgs() << "Static-false: TC > " << TC << "\n" ); |
1976 | // Prolog never falls through; branch to epilog and orphan interior |
1977 | // blocks. Leave it to unreachable-block-elim to clean up. |
1978 | Prolog->removeSuccessor(Succ: Fallthrough); |
1979 | for (MachineInstr &P : Fallthrough->phis()) { |
1980 | P.removeOperand(OpNo: 2); |
1981 | P.removeOperand(OpNo: 1); |
1982 | } |
1983 | TII->insertUnconditionalBranch(MBB&: *Prolog, DestBB: Epilog, DL: DebugLoc()); |
1984 | KernelDisposed = true; |
1985 | } else { |
1986 | LLVM_DEBUG(dbgs() << "Static-true: TC > " << TC << "\n" ); |
1987 | // Prolog always falls through; remove incoming values in epilog. |
1988 | Prolog->removeSuccessor(Succ: Epilog); |
1989 | for (MachineInstr &P : Epilog->phis()) { |
1990 | P.removeOperand(OpNo: 4); |
1991 | P.removeOperand(OpNo: 3); |
1992 | } |
1993 | } |
1994 | } |
1995 | |
1996 | if (!KernelDisposed) { |
1997 | LoopInfo->adjustTripCount(TripCountAdjust: -(Schedule.getNumStages() - 1)); |
1998 | LoopInfo->setPreheader(Prologs.back()); |
1999 | } else { |
2000 | LoopInfo->disposed(); |
2001 | } |
2002 | } |
2003 | |
2004 | void PeelingModuloScheduleExpander::rewriteKernel() { |
2005 | KernelRewriter KR(*Schedule.getLoop(), Schedule, BB); |
2006 | KR.rewrite(); |
2007 | } |
2008 | |
2009 | void PeelingModuloScheduleExpander::expand() { |
2010 | BB = Schedule.getLoop()->getTopBlock(); |
2011 | Preheader = Schedule.getLoop()->getLoopPreheader(); |
2012 | LLVM_DEBUG(Schedule.dump()); |
2013 | LoopInfo = TII->analyzeLoopForPipelining(LoopBB: BB); |
2014 | assert(LoopInfo); |
2015 | |
2016 | rewriteKernel(); |
2017 | peelPrologAndEpilogs(); |
2018 | fixupBranches(); |
2019 | } |
2020 | |
2021 | void PeelingModuloScheduleExpander::validateAgainstModuloScheduleExpander() { |
2022 | BB = Schedule.getLoop()->getTopBlock(); |
2023 | Preheader = Schedule.getLoop()->getLoopPreheader(); |
2024 | |
2025 | // Dump the schedule before we invalidate and remap all its instructions. |
2026 | // Stash it in a string so we can print it if we found an error. |
2027 | std::string ScheduleDump; |
2028 | raw_string_ostream OS(ScheduleDump); |
2029 | Schedule.print(OS); |
2030 | OS.flush(); |
2031 | |
2032 | // First, run the normal ModuleScheduleExpander. We don't support any |
2033 | // InstrChanges. |
2034 | assert(LIS && "Requires LiveIntervals!" ); |
2035 | ModuloScheduleExpander MSE(MF, Schedule, *LIS, |
2036 | ModuloScheduleExpander::InstrChangesTy()); |
2037 | MSE.expand(); |
2038 | MachineBasicBlock *ExpandedKernel = MSE.getRewrittenKernel(); |
2039 | if (!ExpandedKernel) { |
2040 | // The expander optimized away the kernel. We can't do any useful checking. |
2041 | MSE.cleanup(); |
2042 | return; |
2043 | } |
2044 | // Before running the KernelRewriter, re-add BB into the CFG. |
2045 | Preheader->addSuccessor(Succ: BB); |
2046 | |
2047 | // Now run the new expansion algorithm. |
2048 | KernelRewriter KR(*Schedule.getLoop(), Schedule, BB); |
2049 | KR.rewrite(); |
2050 | peelPrologAndEpilogs(); |
2051 | |
2052 | // Collect all illegal phis that the new algorithm created. We'll give these |
2053 | // to KernelOperandInfo. |
2054 | SmallPtrSet<MachineInstr *, 4> IllegalPhis; |
2055 | for (auto NI = BB->getFirstNonPHI(); NI != BB->end(); ++NI) { |
2056 | if (NI->isPHI()) |
2057 | IllegalPhis.insert(Ptr: &*NI); |
2058 | } |
2059 | |
2060 | // Co-iterate across both kernels. We expect them to be identical apart from |
2061 | // phis and full COPYs (we look through both). |
2062 | SmallVector<std::pair<KernelOperandInfo, KernelOperandInfo>, 8> KOIs; |
2063 | auto OI = ExpandedKernel->begin(); |
2064 | auto NI = BB->begin(); |
2065 | for (; !OI->isTerminator() && !NI->isTerminator(); ++OI, ++NI) { |
2066 | while (OI->isPHI() || OI->isFullCopy()) |
2067 | ++OI; |
2068 | while (NI->isPHI() || NI->isFullCopy()) |
2069 | ++NI; |
2070 | assert(OI->getOpcode() == NI->getOpcode() && "Opcodes don't match?!" ); |
2071 | // Analyze every operand separately. |
2072 | for (auto OOpI = OI->operands_begin(), NOpI = NI->operands_begin(); |
2073 | OOpI != OI->operands_end(); ++OOpI, ++NOpI) |
2074 | KOIs.emplace_back(Args: KernelOperandInfo(&*OOpI, MRI, IllegalPhis), |
2075 | Args: KernelOperandInfo(&*NOpI, MRI, IllegalPhis)); |
2076 | } |
2077 | |
2078 | bool Failed = false; |
2079 | for (auto &OldAndNew : KOIs) { |
2080 | if (OldAndNew.first == OldAndNew.second) |
2081 | continue; |
2082 | Failed = true; |
2083 | errs() << "Modulo kernel validation error: [\n" ; |
2084 | errs() << " [golden] " ; |
2085 | OldAndNew.first.print(OS&: errs()); |
2086 | errs() << " " ; |
2087 | OldAndNew.second.print(OS&: errs()); |
2088 | errs() << "]\n" ; |
2089 | } |
2090 | |
2091 | if (Failed) { |
2092 | errs() << "Golden reference kernel:\n" ; |
2093 | ExpandedKernel->print(OS&: errs()); |
2094 | errs() << "New kernel:\n" ; |
2095 | BB->print(OS&: errs()); |
2096 | errs() << ScheduleDump; |
2097 | report_fatal_error( |
2098 | reason: "Modulo kernel validation (-pipeliner-experimental-cg) failed" ); |
2099 | } |
2100 | |
2101 | // Cleanup by removing BB from the CFG again as the original |
2102 | // ModuloScheduleExpander intended. |
2103 | Preheader->removeSuccessor(Succ: BB); |
2104 | MSE.cleanup(); |
2105 | } |
2106 | |
2107 | MachineInstr *ModuloScheduleExpanderMVE::cloneInstr(MachineInstr *OldMI) { |
2108 | MachineInstr *NewMI = MF.CloneMachineInstr(Orig: OldMI); |
2109 | |
2110 | // TODO: Offset information needs to be corrected. |
2111 | NewMI->dropMemRefs(MF); |
2112 | |
2113 | return NewMI; |
2114 | } |
2115 | |
2116 | /// Create a dedicated exit for Loop. Exit is the original exit for Loop. |
2117 | /// If it is already dedicated exit, return it. Otherwise, insert a new |
2118 | /// block between them and return the new block. |
2119 | static MachineBasicBlock *createDedicatedExit(MachineBasicBlock *Loop, |
2120 | MachineBasicBlock *Exit) { |
2121 | if (Exit->pred_size() == 1) |
2122 | return Exit; |
2123 | |
2124 | MachineFunction *MF = Loop->getParent(); |
2125 | const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo(); |
2126 | |
2127 | MachineBasicBlock *NewExit = |
2128 | MF->CreateMachineBasicBlock(BB: Loop->getBasicBlock()); |
2129 | MF->insert(MBBI: Loop->getIterator(), MBB: NewExit); |
2130 | |
2131 | MachineBasicBlock *TBB = nullptr, *FBB = nullptr; |
2132 | SmallVector<MachineOperand, 4> Cond; |
2133 | TII->analyzeBranch(MBB&: *Loop, TBB, FBB, Cond); |
2134 | if (TBB == Loop) |
2135 | FBB = NewExit; |
2136 | else if (FBB == Loop) |
2137 | TBB = NewExit; |
2138 | else |
2139 | llvm_unreachable("unexpected loop structure" ); |
2140 | TII->removeBranch(MBB&: *Loop); |
2141 | TII->insertBranch(MBB&: *Loop, TBB, FBB, Cond, DL: DebugLoc()); |
2142 | Loop->replaceSuccessor(Old: Exit, New: NewExit); |
2143 | TII->insertUnconditionalBranch(MBB&: *NewExit, DestBB: Exit, DL: DebugLoc()); |
2144 | NewExit->addSuccessor(Succ: Exit); |
2145 | |
2146 | Exit->replacePhiUsesWith(Old: Loop, New: NewExit); |
2147 | |
2148 | return NewExit; |
2149 | } |
2150 | |
2151 | /// Insert branch code into the end of MBB. It branches to GreaterThan if the |
2152 | /// remaining trip count for instructions in LastStage0Insts is greater than |
2153 | /// RequiredTC, and to Otherwise otherwise. |
2154 | void ModuloScheduleExpanderMVE::insertCondBranch(MachineBasicBlock &MBB, |
2155 | int RequiredTC, |
2156 | InstrMapTy &LastStage0Insts, |
2157 | MachineBasicBlock &GreaterThan, |
2158 | MachineBasicBlock &Otherwise) { |
2159 | SmallVector<MachineOperand, 4> Cond; |
2160 | LoopInfo->createRemainingIterationsGreaterCondition(TC: RequiredTC, MBB, Cond, |
2161 | LastStage0Insts); |
2162 | |
2163 | if (SwapBranchTargetsMVE) { |
2164 | // Set SwapBranchTargetsMVE to true if a target prefers to replace TBB and |
2165 | // FBB for optimal performance. |
2166 | if (TII->reverseBranchCondition(Cond)) |
2167 | llvm_unreachable("can not reverse branch condition" ); |
2168 | TII->insertBranch(MBB, TBB: &Otherwise, FBB: &GreaterThan, Cond, DL: DebugLoc()); |
2169 | } else { |
2170 | TII->insertBranch(MBB, TBB: &GreaterThan, FBB: &Otherwise, Cond, DL: DebugLoc()); |
2171 | } |
2172 | } |
2173 | |
2174 | /// Generate a pipelined loop that is unrolled by using MVE algorithm and any |
2175 | /// other necessary blocks. The control flow is modified to execute the |
2176 | /// pipelined loop if the trip count satisfies the condition, otherwise the |
2177 | /// original loop. The original loop is also used to execute the remainder |
2178 | /// iterations which occur due to unrolling. |
2179 | void ModuloScheduleExpanderMVE::generatePipelinedLoop() { |
2180 | // The control flow for pipelining with MVE: |
2181 | // |
2182 | // OrigPreheader: |
2183 | // // The block that is originally the loop preheader |
2184 | // goto Check |
2185 | // |
2186 | // Check: |
2187 | // // Check whether the trip count satisfies the requirements to pipeline. |
2188 | // if (LoopCounter > NumStages + NumUnroll - 2) |
2189 | // // The minimum number of iterations to pipeline = |
2190 | // // iterations executed in prolog/epilog (NumStages-1) + |
2191 | // // iterations executed in one kernel run (NumUnroll) |
2192 | // goto Prolog |
2193 | // // fallback to the original loop |
2194 | // goto NewPreheader |
2195 | // |
2196 | // Prolog: |
2197 | // // All prolog stages. There are no direct branches to the epilogue. |
2198 | // goto NewKernel |
2199 | // |
2200 | // NewKernel: |
2201 | // // NumUnroll copies of the kernel |
2202 | // if (LoopCounter > MVE-1) |
2203 | // goto NewKernel |
2204 | // goto Epilog |
2205 | // |
2206 | // Epilog: |
2207 | // // All epilog stages. |
2208 | // if (LoopCounter > 0) |
2209 | // // The remainder is executed in the original loop |
2210 | // goto NewPreheader |
2211 | // goto NewExit |
2212 | // |
2213 | // NewPreheader: |
2214 | // // Newly created preheader for the original loop. |
2215 | // // The initial values of the phis in the loop are merged from two paths. |
2216 | // NewInitVal = Phi OrigInitVal, Check, PipelineLastVal, Epilog |
2217 | // goto OrigKernel |
2218 | // |
2219 | // OrigKernel: |
2220 | // // The original loop block. |
2221 | // if (LoopCounter != 0) |
2222 | // goto OrigKernel |
2223 | // goto NewExit |
2224 | // |
2225 | // NewExit: |
2226 | // // Newly created dedicated exit for the original loop. |
2227 | // // Merge values which are referenced after the loop |
2228 | // Merged = Phi OrigVal, OrigKernel, PipelineVal, Epilog |
2229 | // goto OrigExit |
2230 | // |
2231 | // OrigExit: |
2232 | // // The block that is originally the loop exit. |
2233 | // // If it is already deicated exit, NewExit is not created. |
2234 | |
2235 | // An example of where each stage is executed: |
2236 | // Assume #Stages 3, #MVE 4, #Iterations 12 |
2237 | // Iter 0 1 2 3 4 5 6 7 8 9 10-11 |
2238 | // ------------------------------------------------- |
2239 | // Stage 0 Prolog#0 |
2240 | // Stage 1 0 Prolog#1 |
2241 | // Stage 2 1 0 Kernel Unroll#0 Iter#0 |
2242 | // Stage 2 1 0 Kernel Unroll#1 Iter#0 |
2243 | // Stage 2 1 0 Kernel Unroll#2 Iter#0 |
2244 | // Stage 2 1 0 Kernel Unroll#3 Iter#0 |
2245 | // Stage 2 1 0 Kernel Unroll#0 Iter#1 |
2246 | // Stage 2 1 0 Kernel Unroll#1 Iter#1 |
2247 | // Stage 2 1 0 Kernel Unroll#2 Iter#1 |
2248 | // Stage 2 1 0 Kernel Unroll#3 Iter#1 |
2249 | // Stage 2 1 Epilog#0 |
2250 | // Stage 2 Epilog#1 |
2251 | // Stage 0-2 OrigKernel |
2252 | |
2253 | LoopInfo = TII->analyzeLoopForPipelining(LoopBB: OrigKernel); |
2254 | assert(LoopInfo && "Must be able to analyze loop!" ); |
2255 | |
2256 | calcNumUnroll(); |
2257 | |
2258 | Check = MF.CreateMachineBasicBlock(BB: OrigKernel->getBasicBlock()); |
2259 | Prolog = MF.CreateMachineBasicBlock(BB: OrigKernel->getBasicBlock()); |
2260 | NewKernel = MF.CreateMachineBasicBlock(BB: OrigKernel->getBasicBlock()); |
2261 | Epilog = MF.CreateMachineBasicBlock(BB: OrigKernel->getBasicBlock()); |
2262 | NewPreheader = MF.CreateMachineBasicBlock(BB: OrigKernel->getBasicBlock()); |
2263 | |
2264 | MF.insert(MBBI: OrigKernel->getIterator(), MBB: Check); |
2265 | MF.insert(MBBI: OrigKernel->getIterator(), MBB: Prolog); |
2266 | MF.insert(MBBI: OrigKernel->getIterator(), MBB: NewKernel); |
2267 | MF.insert(MBBI: OrigKernel->getIterator(), MBB: Epilog); |
2268 | MF.insert(MBBI: OrigKernel->getIterator(), MBB: NewPreheader); |
2269 | |
2270 | NewExit = createDedicatedExit(Loop: OrigKernel, Exit: OrigExit); |
2271 | |
2272 | NewPreheader->transferSuccessorsAndUpdatePHIs(FromMBB: OrigPreheader); |
2273 | TII->insertUnconditionalBranch(MBB&: *NewPreheader, DestBB: OrigKernel, DL: DebugLoc()); |
2274 | |
2275 | OrigPreheader->addSuccessor(Succ: Check); |
2276 | TII->removeBranch(MBB&: *OrigPreheader); |
2277 | TII->insertUnconditionalBranch(MBB&: *OrigPreheader, DestBB: Check, DL: DebugLoc()); |
2278 | |
2279 | Check->addSuccessor(Succ: Prolog); |
2280 | Check->addSuccessor(Succ: NewPreheader); |
2281 | |
2282 | Prolog->addSuccessor(Succ: NewKernel); |
2283 | |
2284 | NewKernel->addSuccessor(Succ: NewKernel); |
2285 | NewKernel->addSuccessor(Succ: Epilog); |
2286 | |
2287 | Epilog->addSuccessor(Succ: NewPreheader); |
2288 | Epilog->addSuccessor(Succ: NewExit); |
2289 | |
2290 | InstrMapTy LastStage0Insts; |
2291 | insertCondBranch(MBB&: *Check, RequiredTC: Schedule.getNumStages() + NumUnroll - 2, |
2292 | LastStage0Insts, GreaterThan&: *Prolog, Otherwise&: *NewPreheader); |
2293 | |
2294 | // VRMaps map (prolog/kernel/epilog phase#, original register#) to new |
2295 | // register# |
2296 | SmallVector<ValueMapTy> PrologVRMap, KernelVRMap, EpilogVRMap; |
2297 | generateProlog(VRMap&: PrologVRMap); |
2298 | generateKernel(PrologVRMap, KernelVRMap, LastStage0Insts); |
2299 | generateEpilog(KernelVRMap, EpilogVRMap, LastStage0Insts); |
2300 | } |
2301 | |
2302 | /// Replace MI's use operands according to the maps. |
2303 | void ModuloScheduleExpanderMVE::updateInstrUse( |
2304 | MachineInstr *MI, int StageNum, int PhaseNum, |
2305 | SmallVectorImpl<ValueMapTy> &CurVRMap, |
2306 | SmallVectorImpl<ValueMapTy> *PrevVRMap) { |
2307 | // If MI is in the prolog/kernel/epilog block, CurVRMap is |
2308 | // PrologVRMap/KernelVRMap/EpilogVRMap respectively. |
2309 | // PrevVRMap is nullptr/PhiVRMap/KernelVRMap respectively. |
2310 | // Refer to the appropriate map according to the stage difference between |
2311 | // MI and the definition of an operand. |
2312 | |
2313 | for (MachineOperand &UseMO : MI->uses()) { |
2314 | if (!UseMO.isReg() || !UseMO.getReg().isVirtual()) |
2315 | continue; |
2316 | int DiffStage = 0; |
2317 | Register OrigReg = UseMO.getReg(); |
2318 | MachineInstr *DefInst = MRI.getVRegDef(Reg: OrigReg); |
2319 | if (!DefInst || DefInst->getParent() != OrigKernel) |
2320 | continue; |
2321 | unsigned InitReg = 0; |
2322 | unsigned DefReg = OrigReg; |
2323 | if (DefInst->isPHI()) { |
2324 | ++DiffStage; |
2325 | unsigned LoopReg; |
2326 | getPhiRegs(Phi&: *DefInst, Loop: OrigKernel, InitVal&: InitReg, LoopVal&: LoopReg); |
2327 | // LoopReg is guaranteed to be defined within the loop by canApply() |
2328 | DefReg = LoopReg; |
2329 | DefInst = MRI.getVRegDef(Reg: LoopReg); |
2330 | } |
2331 | unsigned DefStageNum = Schedule.getStage(MI: DefInst); |
2332 | DiffStage += StageNum - DefStageNum; |
2333 | Register NewReg; |
2334 | if (PhaseNum >= DiffStage && CurVRMap[PhaseNum - DiffStage].count(Val: DefReg)) |
2335 | // NewReg is defined in a previous phase of the same block |
2336 | NewReg = CurVRMap[PhaseNum - DiffStage][DefReg]; |
2337 | else if (!PrevVRMap) |
2338 | // Since this is the first iteration, refer the initial register of the |
2339 | // loop |
2340 | NewReg = InitReg; |
2341 | else |
2342 | // Cases where DiffStage is larger than PhaseNum. |
2343 | // If MI is in the kernel block, the value is defined by the previous |
2344 | // iteration and PhiVRMap is referenced. If MI is in the epilog block, the |
2345 | // value is defined in the kernel block and KernelVRMap is referenced. |
2346 | NewReg = (*PrevVRMap)[PrevVRMap->size() - (DiffStage - PhaseNum)][DefReg]; |
2347 | |
2348 | const TargetRegisterClass *NRC = |
2349 | MRI.constrainRegClass(Reg: NewReg, RC: MRI.getRegClass(Reg: OrigReg)); |
2350 | if (NRC) |
2351 | UseMO.setReg(NewReg); |
2352 | else { |
2353 | Register SplitReg = MRI.createVirtualRegister(RegClass: MRI.getRegClass(Reg: OrigReg)); |
2354 | BuildMI(BB&: *OrigKernel, I: MI, MIMD: MI->getDebugLoc(), MCID: TII->get(Opcode: TargetOpcode::COPY), |
2355 | DestReg: SplitReg) |
2356 | .addReg(RegNo: NewReg); |
2357 | UseMO.setReg(SplitReg); |
2358 | } |
2359 | } |
2360 | } |
2361 | |
2362 | /// Return a phi if Reg is referenced by the phi. |
2363 | /// canApply() guarantees that at most only one such phi exists. |
2364 | static MachineInstr *getLoopPhiUser(Register Reg, MachineBasicBlock *Loop) { |
2365 | for (MachineInstr &Phi : Loop->phis()) { |
2366 | unsigned InitVal, LoopVal; |
2367 | getPhiRegs(Phi, Loop, InitVal, LoopVal); |
2368 | if (LoopVal == Reg) |
2369 | return Φ |
2370 | } |
2371 | return nullptr; |
2372 | } |
2373 | |
2374 | /// Generate phis for registers defined by OrigMI. |
2375 | void ModuloScheduleExpanderMVE::generatePhi( |
2376 | MachineInstr *OrigMI, int UnrollNum, |
2377 | SmallVectorImpl<ValueMapTy> &PrologVRMap, |
2378 | SmallVectorImpl<ValueMapTy> &KernelVRMap, |
2379 | SmallVectorImpl<ValueMapTy> &PhiVRMap) { |
2380 | int StageNum = Schedule.getStage(MI: OrigMI); |
2381 | bool UsePrologReg; |
2382 | if (Schedule.getNumStages() - NumUnroll + UnrollNum - 1 >= StageNum) |
2383 | UsePrologReg = true; |
2384 | else if (Schedule.getNumStages() - NumUnroll + UnrollNum == StageNum) |
2385 | UsePrologReg = false; |
2386 | else |
2387 | return; |
2388 | |
2389 | // Examples that show which stages are merged by phi. |
2390 | // Meaning of the symbol following the stage number: |
2391 | // a/b: Stages with the same letter are merged (UsePrologReg == true) |
2392 | // +: Merged with the initial value (UsePrologReg == false) |
2393 | // *: No phis required |
2394 | // |
2395 | // #Stages 3, #MVE 4 |
2396 | // Iter 0 1 2 3 4 5 6 7 8 |
2397 | // ----------------------------------------- |
2398 | // Stage 0a Prolog#0 |
2399 | // Stage 1a 0b Prolog#1 |
2400 | // Stage 2* 1* 0* Kernel Unroll#0 |
2401 | // Stage 2* 1* 0+ Kernel Unroll#1 |
2402 | // Stage 2* 1+ 0a Kernel Unroll#2 |
2403 | // Stage 2+ 1a 0b Kernel Unroll#3 |
2404 | // |
2405 | // #Stages 3, #MVE 2 |
2406 | // Iter 0 1 2 3 4 5 6 7 8 |
2407 | // ----------------------------------------- |
2408 | // Stage 0a Prolog#0 |
2409 | // Stage 1a 0b Prolog#1 |
2410 | // Stage 2* 1+ 0a Kernel Unroll#0 |
2411 | // Stage 2+ 1a 0b Kernel Unroll#1 |
2412 | // |
2413 | // #Stages 3, #MVE 1 |
2414 | // Iter 0 1 2 3 4 5 6 7 8 |
2415 | // ----------------------------------------- |
2416 | // Stage 0* Prolog#0 |
2417 | // Stage 1a 0b Prolog#1 |
2418 | // Stage 2+ 1a 0b Kernel Unroll#0 |
2419 | |
2420 | for (MachineOperand &DefMO : OrigMI->defs()) { |
2421 | if (!DefMO.isReg() || DefMO.isDead()) |
2422 | continue; |
2423 | Register OrigReg = DefMO.getReg(); |
2424 | auto NewReg = KernelVRMap[UnrollNum].find(Val: OrigReg); |
2425 | if (NewReg == KernelVRMap[UnrollNum].end()) |
2426 | continue; |
2427 | Register CorrespondReg; |
2428 | if (UsePrologReg) { |
2429 | int PrologNum = Schedule.getNumStages() - NumUnroll + UnrollNum - 1; |
2430 | CorrespondReg = PrologVRMap[PrologNum][OrigReg]; |
2431 | } else { |
2432 | MachineInstr *Phi = getLoopPhiUser(Reg: OrigReg, Loop: OrigKernel); |
2433 | if (!Phi) |
2434 | continue; |
2435 | CorrespondReg = getInitPhiReg(Phi&: *Phi, LoopBB: OrigKernel); |
2436 | } |
2437 | |
2438 | assert(CorrespondReg.isValid()); |
2439 | Register PhiReg = MRI.createVirtualRegister(RegClass: MRI.getRegClass(Reg: OrigReg)); |
2440 | BuildMI(BB&: *NewKernel, I: NewKernel->getFirstNonPHI(), MIMD: DebugLoc(), |
2441 | MCID: TII->get(Opcode: TargetOpcode::PHI), DestReg: PhiReg) |
2442 | .addReg(RegNo: NewReg->second) |
2443 | .addMBB(MBB: NewKernel) |
2444 | .addReg(RegNo: CorrespondReg) |
2445 | .addMBB(MBB: Prolog); |
2446 | PhiVRMap[UnrollNum][OrigReg] = PhiReg; |
2447 | } |
2448 | } |
2449 | |
2450 | static void replacePhiSrc(MachineInstr &Phi, Register OrigReg, Register NewReg, |
2451 | MachineBasicBlock *NewMBB) { |
2452 | for (unsigned Idx = 1; Idx < Phi.getNumOperands(); Idx += 2) { |
2453 | if (Phi.getOperand(i: Idx).getReg() == OrigReg) { |
2454 | Phi.getOperand(i: Idx).setReg(NewReg); |
2455 | Phi.getOperand(i: Idx + 1).setMBB(NewMBB); |
2456 | return; |
2457 | } |
2458 | } |
2459 | } |
2460 | |
2461 | /// Generate phis that merge values from multiple routes |
2462 | void ModuloScheduleExpanderMVE::mergeRegUsesAfterPipeline(Register OrigReg, |
2463 | Register NewReg) { |
2464 | SmallVector<MachineOperand *> UsesAfterLoop; |
2465 | SmallVector<MachineInstr *> LoopPhis; |
2466 | for (MachineRegisterInfo::use_iterator I = MRI.use_begin(RegNo: OrigReg), |
2467 | E = MRI.use_end(); |
2468 | I != E; ++I) { |
2469 | MachineOperand &O = *I; |
2470 | if (O.getParent()->getParent() != OrigKernel && |
2471 | O.getParent()->getParent() != Prolog && |
2472 | O.getParent()->getParent() != NewKernel && |
2473 | O.getParent()->getParent() != Epilog) |
2474 | UsesAfterLoop.push_back(Elt: &O); |
2475 | if (O.getParent()->getParent() == OrigKernel && O.getParent()->isPHI()) |
2476 | LoopPhis.push_back(Elt: O.getParent()); |
2477 | } |
2478 | |
2479 | // Merge the route that only execute the pipelined loop (when there are no |
2480 | // remaining iterations) with the route that execute the original loop. |
2481 | if (!UsesAfterLoop.empty()) { |
2482 | Register PhiReg = MRI.createVirtualRegister(RegClass: MRI.getRegClass(Reg: OrigReg)); |
2483 | BuildMI(BB&: *NewExit, I: NewExit->getFirstNonPHI(), MIMD: DebugLoc(), |
2484 | MCID: TII->get(Opcode: TargetOpcode::PHI), DestReg: PhiReg) |
2485 | .addReg(RegNo: OrigReg) |
2486 | .addMBB(MBB: OrigKernel) |
2487 | .addReg(RegNo: NewReg) |
2488 | .addMBB(MBB: Epilog); |
2489 | |
2490 | for (MachineOperand *MO : UsesAfterLoop) |
2491 | MO->setReg(PhiReg); |
2492 | |
2493 | if (!LIS.hasInterval(Reg: PhiReg)) |
2494 | LIS.createEmptyInterval(Reg: PhiReg); |
2495 | } |
2496 | |
2497 | // Merge routes from the pipelined loop and the bypassed route before the |
2498 | // original loop |
2499 | if (!LoopPhis.empty()) { |
2500 | for (MachineInstr *Phi : LoopPhis) { |
2501 | unsigned InitReg, LoopReg; |
2502 | getPhiRegs(Phi&: *Phi, Loop: OrigKernel, InitVal&: InitReg, LoopVal&: LoopReg); |
2503 | Register NewInit = MRI.createVirtualRegister(RegClass: MRI.getRegClass(Reg: InitReg)); |
2504 | BuildMI(BB&: *NewPreheader, I: NewPreheader->getFirstNonPHI(), MIMD: Phi->getDebugLoc(), |
2505 | MCID: TII->get(Opcode: TargetOpcode::PHI), DestReg: NewInit) |
2506 | .addReg(RegNo: InitReg) |
2507 | .addMBB(MBB: Check) |
2508 | .addReg(RegNo: NewReg) |
2509 | .addMBB(MBB: Epilog); |
2510 | replacePhiSrc(Phi&: *Phi, OrigReg: InitReg, NewReg: NewInit, NewMBB: NewPreheader); |
2511 | } |
2512 | } |
2513 | } |
2514 | |
2515 | void ModuloScheduleExpanderMVE::generateProlog( |
2516 | SmallVectorImpl<ValueMapTy> &PrologVRMap) { |
2517 | PrologVRMap.clear(); |
2518 | PrologVRMap.resize(N: Schedule.getNumStages() - 1); |
2519 | DenseMap<MachineInstr *, std::pair<int, int>> NewMIMap; |
2520 | for (int PrologNum = 0; PrologNum < Schedule.getNumStages() - 1; |
2521 | ++PrologNum) { |
2522 | for (MachineInstr *MI : Schedule.getInstructions()) { |
2523 | if (MI->isPHI()) |
2524 | continue; |
2525 | int StageNum = Schedule.getStage(MI); |
2526 | if (StageNum > PrologNum) |
2527 | continue; |
2528 | MachineInstr *NewMI = cloneInstr(OldMI: MI); |
2529 | updateInstrDef(NewMI, VRMap&: PrologVRMap[PrologNum], LastDef: false); |
2530 | NewMIMap[NewMI] = {PrologNum, StageNum}; |
2531 | Prolog->push_back(MI: NewMI); |
2532 | } |
2533 | } |
2534 | |
2535 | for (auto I : NewMIMap) { |
2536 | MachineInstr *MI = I.first; |
2537 | int PrologNum = I.second.first; |
2538 | int StageNum = I.second.second; |
2539 | updateInstrUse(MI, StageNum, PhaseNum: PrologNum, CurVRMap&: PrologVRMap, PrevVRMap: nullptr); |
2540 | } |
2541 | |
2542 | LLVM_DEBUG({ |
2543 | dbgs() << "prolog:\n" ; |
2544 | Prolog->dump(); |
2545 | }); |
2546 | } |
2547 | |
2548 | void ModuloScheduleExpanderMVE::generateKernel( |
2549 | SmallVectorImpl<ValueMapTy> &PrologVRMap, |
2550 | SmallVectorImpl<ValueMapTy> &KernelVRMap, InstrMapTy &LastStage0Insts) { |
2551 | KernelVRMap.clear(); |
2552 | KernelVRMap.resize(N: NumUnroll); |
2553 | SmallVector<ValueMapTy> PhiVRMap; |
2554 | PhiVRMap.resize(N: NumUnroll); |
2555 | DenseMap<MachineInstr *, std::pair<int, int>> NewMIMap; |
2556 | DenseMap<MachineInstr *, MachineInstr *> MIMapLastStage0; |
2557 | for (int UnrollNum = 0; UnrollNum < NumUnroll; ++UnrollNum) { |
2558 | for (MachineInstr *MI : Schedule.getInstructions()) { |
2559 | if (MI->isPHI()) |
2560 | continue; |
2561 | int StageNum = Schedule.getStage(MI); |
2562 | MachineInstr *NewMI = cloneInstr(OldMI: MI); |
2563 | if (UnrollNum == NumUnroll - 1) |
2564 | LastStage0Insts[MI] = NewMI; |
2565 | updateInstrDef(NewMI, VRMap&: KernelVRMap[UnrollNum], |
2566 | LastDef: (UnrollNum == NumUnroll - 1 && StageNum == 0)); |
2567 | generatePhi(OrigMI: MI, UnrollNum, PrologVRMap, KernelVRMap, PhiVRMap); |
2568 | NewMIMap[NewMI] = {UnrollNum, StageNum}; |
2569 | NewKernel->push_back(MI: NewMI); |
2570 | } |
2571 | } |
2572 | |
2573 | for (auto I : NewMIMap) { |
2574 | MachineInstr *MI = I.first; |
2575 | int UnrollNum = I.second.first; |
2576 | int StageNum = I.second.second; |
2577 | updateInstrUse(MI, StageNum, PhaseNum: UnrollNum, CurVRMap&: KernelVRMap, PrevVRMap: &PhiVRMap); |
2578 | } |
2579 | |
2580 | // If remaining trip count is greater than NumUnroll-1, loop continues |
2581 | insertCondBranch(MBB&: *NewKernel, RequiredTC: NumUnroll - 1, LastStage0Insts, GreaterThan&: *NewKernel, |
2582 | Otherwise&: *Epilog); |
2583 | |
2584 | LLVM_DEBUG({ |
2585 | dbgs() << "kernel:\n" ; |
2586 | NewKernel->dump(); |
2587 | }); |
2588 | } |
2589 | |
2590 | void ModuloScheduleExpanderMVE::generateEpilog( |
2591 | SmallVectorImpl<ValueMapTy> &KernelVRMap, |
2592 | SmallVectorImpl<ValueMapTy> &EpilogVRMap, InstrMapTy &LastStage0Insts) { |
2593 | EpilogVRMap.clear(); |
2594 | EpilogVRMap.resize(N: Schedule.getNumStages() - 1); |
2595 | DenseMap<MachineInstr *, std::pair<int, int>> NewMIMap; |
2596 | for (int EpilogNum = 0; EpilogNum < Schedule.getNumStages() - 1; |
2597 | ++EpilogNum) { |
2598 | for (MachineInstr *MI : Schedule.getInstructions()) { |
2599 | if (MI->isPHI()) |
2600 | continue; |
2601 | int StageNum = Schedule.getStage(MI); |
2602 | if (StageNum <= EpilogNum) |
2603 | continue; |
2604 | MachineInstr *NewMI = cloneInstr(OldMI: MI); |
2605 | updateInstrDef(NewMI, VRMap&: EpilogVRMap[EpilogNum], LastDef: StageNum - 1 == EpilogNum); |
2606 | NewMIMap[NewMI] = {EpilogNum, StageNum}; |
2607 | Epilog->push_back(MI: NewMI); |
2608 | } |
2609 | } |
2610 | |
2611 | for (auto I : NewMIMap) { |
2612 | MachineInstr *MI = I.first; |
2613 | int EpilogNum = I.second.first; |
2614 | int StageNum = I.second.second; |
2615 | updateInstrUse(MI, StageNum, PhaseNum: EpilogNum, CurVRMap&: EpilogVRMap, PrevVRMap: &KernelVRMap); |
2616 | } |
2617 | |
2618 | // If there are remaining iterations, they are executed in the original loop. |
2619 | // Instructions related to loop control, such as loop counter comparison, |
2620 | // are indicated by shouldIgnoreForPipelining() and are assumed to be placed |
2621 | // in stage 0. Thus, the map is for the last one in the kernel. |
2622 | insertCondBranch(MBB&: *Epilog, RequiredTC: 0, LastStage0Insts, GreaterThan&: *NewPreheader, Otherwise&: *NewExit); |
2623 | |
2624 | LLVM_DEBUG({ |
2625 | dbgs() << "epilog:\n" ; |
2626 | Epilog->dump(); |
2627 | }); |
2628 | } |
2629 | |
2630 | /// Calculate the number of unroll required and set it to NumUnroll |
2631 | void ModuloScheduleExpanderMVE::calcNumUnroll() { |
2632 | DenseMap<MachineInstr *, unsigned> Inst2Idx; |
2633 | NumUnroll = 1; |
2634 | for (unsigned I = 0; I < Schedule.getInstructions().size(); ++I) |
2635 | Inst2Idx[Schedule.getInstructions()[I]] = I; |
2636 | |
2637 | for (MachineInstr *MI : Schedule.getInstructions()) { |
2638 | if (MI->isPHI()) |
2639 | continue; |
2640 | int StageNum = Schedule.getStage(MI); |
2641 | for (const MachineOperand &MO : MI->uses()) { |
2642 | if (!MO.isReg() || !MO.getReg().isVirtual()) |
2643 | continue; |
2644 | MachineInstr *DefMI = MRI.getVRegDef(Reg: MO.getReg()); |
2645 | if (DefMI->getParent() != OrigKernel) |
2646 | continue; |
2647 | |
2648 | int NumUnrollLocal = 1; |
2649 | if (DefMI->isPHI()) { |
2650 | ++NumUnrollLocal; |
2651 | // canApply() guarantees that DefMI is not phi and is an instruction in |
2652 | // the loop |
2653 | DefMI = MRI.getVRegDef(Reg: getLoopPhiReg(Phi&: *DefMI, LoopBB: OrigKernel)); |
2654 | } |
2655 | NumUnrollLocal += StageNum - Schedule.getStage(MI: DefMI); |
2656 | if (Inst2Idx[MI] <= Inst2Idx[DefMI]) |
2657 | --NumUnrollLocal; |
2658 | NumUnroll = std::max(a: NumUnroll, b: NumUnrollLocal); |
2659 | } |
2660 | } |
2661 | LLVM_DEBUG(dbgs() << "NumUnroll: " << NumUnroll << "\n" ); |
2662 | } |
2663 | |
2664 | /// Create new virtual registers for definitions of NewMI and update NewMI. |
2665 | /// If the definitions are referenced after the pipelined loop, phis are |
2666 | /// created to merge with other routes. |
2667 | void ModuloScheduleExpanderMVE::updateInstrDef(MachineInstr *NewMI, |
2668 | ValueMapTy &VRMap, |
2669 | bool LastDef) { |
2670 | for (MachineOperand &MO : NewMI->operands()) { |
2671 | if (!MO.isReg() || !MO.getReg().isVirtual() || !MO.isDef()) |
2672 | continue; |
2673 | Register Reg = MO.getReg(); |
2674 | const TargetRegisterClass *RC = MRI.getRegClass(Reg); |
2675 | Register NewReg = MRI.createVirtualRegister(RegClass: RC); |
2676 | MO.setReg(NewReg); |
2677 | VRMap[Reg] = NewReg; |
2678 | if (LastDef) |
2679 | mergeRegUsesAfterPipeline(OrigReg: Reg, NewReg); |
2680 | } |
2681 | } |
2682 | |
2683 | void ModuloScheduleExpanderMVE::expand() { |
2684 | OrigKernel = Schedule.getLoop()->getTopBlock(); |
2685 | OrigPreheader = Schedule.getLoop()->getLoopPreheader(); |
2686 | OrigExit = Schedule.getLoop()->getExitBlock(); |
2687 | |
2688 | LLVM_DEBUG(Schedule.dump()); |
2689 | |
2690 | generatePipelinedLoop(); |
2691 | } |
2692 | |
2693 | /// Check if ModuloScheduleExpanderMVE can be applied to L |
2694 | bool ModuloScheduleExpanderMVE::canApply(MachineLoop &L) { |
2695 | if (!L.getExitBlock()) { |
2696 | LLVM_DEBUG( |
2697 | dbgs() << "Can not apply MVE expander: No single exit block.\n" ;); |
2698 | return false; |
2699 | } |
2700 | |
2701 | MachineBasicBlock *BB = L.getTopBlock(); |
2702 | MachineRegisterInfo &MRI = BB->getParent()->getRegInfo(); |
2703 | |
2704 | // Put some constraints on the operands of the phis to simplify the |
2705 | // transformation |
2706 | DenseSet<unsigned> UsedByPhi; |
2707 | for (MachineInstr &MI : BB->phis()) { |
2708 | // Registers defined by phis must be used only inside the loop and be never |
2709 | // used by phis. |
2710 | for (MachineOperand &MO : MI.defs()) |
2711 | if (MO.isReg()) |
2712 | for (MachineInstr &Ref : MRI.use_instructions(Reg: MO.getReg())) |
2713 | if (Ref.getParent() != BB || Ref.isPHI()) { |
2714 | LLVM_DEBUG(dbgs() |
2715 | << "Can not apply MVE expander: A phi result is " |
2716 | "referenced outside of the loop or by phi.\n" ;); |
2717 | return false; |
2718 | } |
2719 | |
2720 | // A source register from the loop block must be defined inside the loop. |
2721 | // A register defined inside the loop must be referenced by only one phi at |
2722 | // most. |
2723 | unsigned InitVal, LoopVal; |
2724 | getPhiRegs(Phi&: MI, Loop: MI.getParent(), InitVal, LoopVal); |
2725 | if (!Register(LoopVal).isVirtual() || |
2726 | MRI.getVRegDef(Reg: LoopVal)->getParent() != BB) { |
2727 | LLVM_DEBUG( |
2728 | dbgs() << "Can not apply MVE expander: A phi source value coming " |
2729 | "from the loop is not defined in the loop.\n" ;); |
2730 | return false; |
2731 | } |
2732 | if (UsedByPhi.count(V: LoopVal)) { |
2733 | LLVM_DEBUG(dbgs() << "Can not apply MVE expander: A value defined in the " |
2734 | "loop is referenced by two or more phis.\n" ;); |
2735 | return false; |
2736 | } |
2737 | UsedByPhi.insert(V: LoopVal); |
2738 | } |
2739 | |
2740 | return true; |
2741 | } |
2742 | |
2743 | //===----------------------------------------------------------------------===// |
2744 | // ModuloScheduleTestPass implementation |
2745 | //===----------------------------------------------------------------------===// |
2746 | // This pass constructs a ModuloSchedule from its module and runs |
2747 | // ModuloScheduleExpander. |
2748 | // |
2749 | // The module is expected to contain a single-block analyzable loop. |
2750 | // The total order of instructions is taken from the loop as-is. |
2751 | // Instructions are expected to be annotated with a PostInstrSymbol. |
2752 | // This PostInstrSymbol must have the following format: |
2753 | // "Stage=%d Cycle=%d". |
2754 | //===----------------------------------------------------------------------===// |
2755 | |
2756 | namespace { |
2757 | class ModuloScheduleTest : public MachineFunctionPass { |
2758 | public: |
2759 | static char ID; |
2760 | |
2761 | ModuloScheduleTest() : MachineFunctionPass(ID) { |
2762 | initializeModuloScheduleTestPass(*PassRegistry::getPassRegistry()); |
2763 | } |
2764 | |
2765 | bool runOnMachineFunction(MachineFunction &MF) override; |
2766 | void runOnLoop(MachineFunction &MF, MachineLoop &L); |
2767 | |
2768 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
2769 | AU.addRequired<MachineLoopInfoWrapperPass>(); |
2770 | AU.addRequired<LiveIntervalsWrapperPass>(); |
2771 | MachineFunctionPass::getAnalysisUsage(AU); |
2772 | } |
2773 | }; |
2774 | } // namespace |
2775 | |
2776 | char ModuloScheduleTest::ID = 0; |
2777 | |
2778 | INITIALIZE_PASS_BEGIN(ModuloScheduleTest, "modulo-schedule-test" , |
2779 | "Modulo Schedule test pass" , false, false) |
2780 | INITIALIZE_PASS_DEPENDENCY(MachineLoopInfoWrapperPass) |
2781 | INITIALIZE_PASS_DEPENDENCY(LiveIntervalsWrapperPass) |
2782 | INITIALIZE_PASS_END(ModuloScheduleTest, "modulo-schedule-test" , |
2783 | "Modulo Schedule test pass" , false, false) |
2784 | |
2785 | bool ModuloScheduleTest::runOnMachineFunction(MachineFunction &MF) { |
2786 | MachineLoopInfo &MLI = getAnalysis<MachineLoopInfoWrapperPass>().getLI(); |
2787 | for (auto *L : MLI) { |
2788 | if (L->getTopBlock() != L->getBottomBlock()) |
2789 | continue; |
2790 | runOnLoop(MF, L&: *L); |
2791 | return false; |
2792 | } |
2793 | return false; |
2794 | } |
2795 | |
2796 | static void parseSymbolString(StringRef S, int &Cycle, int &Stage) { |
2797 | std::pair<StringRef, StringRef> StageAndCycle = getToken(Source: S, Delimiters: "_" ); |
2798 | std::pair<StringRef, StringRef> StageTokenAndValue = |
2799 | getToken(Source: StageAndCycle.first, Delimiters: "-" ); |
2800 | std::pair<StringRef, StringRef> CycleTokenAndValue = |
2801 | getToken(Source: StageAndCycle.second, Delimiters: "-" ); |
2802 | if (StageTokenAndValue.first != "Stage" || |
2803 | CycleTokenAndValue.first != "_Cycle" ) { |
2804 | llvm_unreachable( |
2805 | "Bad post-instr symbol syntax: see comment in ModuloScheduleTest" ); |
2806 | return; |
2807 | } |
2808 | |
2809 | StageTokenAndValue.second.drop_front().getAsInteger(Radix: 10, Result&: Stage); |
2810 | CycleTokenAndValue.second.drop_front().getAsInteger(Radix: 10, Result&: Cycle); |
2811 | |
2812 | dbgs() << " Stage=" << Stage << ", Cycle=" << Cycle << "\n" ; |
2813 | } |
2814 | |
2815 | void ModuloScheduleTest::runOnLoop(MachineFunction &MF, MachineLoop &L) { |
2816 | LiveIntervals &LIS = getAnalysis<LiveIntervalsWrapperPass>().getLIS(); |
2817 | MachineBasicBlock *BB = L.getTopBlock(); |
2818 | dbgs() << "--- ModuloScheduleTest running on BB#" << BB->getNumber() << "\n" ; |
2819 | |
2820 | DenseMap<MachineInstr *, int> Cycle, Stage; |
2821 | std::vector<MachineInstr *> Instrs; |
2822 | for (MachineInstr &MI : *BB) { |
2823 | if (MI.isTerminator()) |
2824 | continue; |
2825 | Instrs.push_back(x: &MI); |
2826 | if (MCSymbol *Sym = MI.getPostInstrSymbol()) { |
2827 | dbgs() << "Parsing post-instr symbol for " << MI; |
2828 | parseSymbolString(S: Sym->getName(), Cycle&: Cycle[&MI], Stage&: Stage[&MI]); |
2829 | } |
2830 | } |
2831 | |
2832 | ModuloSchedule MS(MF, &L, std::move(Instrs), std::move(Cycle), |
2833 | std::move(Stage)); |
2834 | ModuloScheduleExpander MSE( |
2835 | MF, MS, LIS, /*InstrChanges=*/ModuloScheduleExpander::InstrChangesTy()); |
2836 | MSE.expand(); |
2837 | MSE.cleanup(); |
2838 | } |
2839 | |
2840 | //===----------------------------------------------------------------------===// |
2841 | // ModuloScheduleTestAnnotater implementation |
2842 | //===----------------------------------------------------------------------===// |
2843 | |
2844 | void ModuloScheduleTestAnnotater::annotate() { |
2845 | for (MachineInstr *MI : S.getInstructions()) { |
2846 | SmallVector<char, 16> SV; |
2847 | raw_svector_ostream OS(SV); |
2848 | OS << "Stage-" << S.getStage(MI) << "_Cycle-" << S.getCycle(MI); |
2849 | MCSymbol *Sym = MF.getContext().getOrCreateSymbol(Name: OS.str()); |
2850 | MI->setPostInstrSymbol(MF, Symbol: Sym); |
2851 | } |
2852 | } |
2853 | |