1 //===- ModuloSchedule.cpp - Software pipeline schedule expansion ----------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 9 #include "llvm/CodeGen/ModuloSchedule.h" 10 #include "llvm/ADT/StringExtras.h" 11 #include "llvm/Analysis/MemoryLocation.h" 12 #include "llvm/CodeGen/LiveIntervals.h" 13 #include "llvm/CodeGen/MachineInstrBuilder.h" 14 #include "llvm/CodeGen/MachineLoopInfo.h" 15 #include "llvm/CodeGen/MachineRegisterInfo.h" 16 #include "llvm/InitializePasses.h" 17 #include "llvm/MC/MCContext.h" 18 #include "llvm/Support/Debug.h" 19 #include "llvm/Support/ErrorHandling.h" 20 #include "llvm/Support/raw_ostream.h" 21 22 #define DEBUG_TYPE "pipeliner" 23 using namespace llvm; 24 25 void ModuloSchedule::print(raw_ostream &OS) { 26 for (MachineInstr *MI : ScheduledInstrs) 27 OS << "[stage " << getStage(MI) << " @" << getCycle(MI) << "c] " << *MI; 28 } 29 30 //===----------------------------------------------------------------------===// 31 // ModuloScheduleExpander implementation 32 //===----------------------------------------------------------------------===// 33 34 /// Return the register values for the operands of a Phi instruction. 35 /// This function assume the instruction is a Phi. 36 static void getPhiRegs(MachineInstr &Phi, MachineBasicBlock *Loop, 37 unsigned &InitVal, unsigned &LoopVal) { 38 assert(Phi.isPHI() && "Expecting a Phi."); 39 40 InitVal = 0; 41 LoopVal = 0; 42 for (unsigned i = 1, e = Phi.getNumOperands(); i != e; i += 2) 43 if (Phi.getOperand(i + 1).getMBB() != Loop) 44 InitVal = Phi.getOperand(i).getReg(); 45 else 46 LoopVal = Phi.getOperand(i).getReg(); 47 48 assert(InitVal != 0 && LoopVal != 0 && "Unexpected Phi structure."); 49 } 50 51 /// Return the Phi register value that comes from the incoming block. 52 static unsigned getInitPhiReg(MachineInstr &Phi, MachineBasicBlock *LoopBB) { 53 for (unsigned i = 1, e = Phi.getNumOperands(); i != e; i += 2) 54 if (Phi.getOperand(i + 1).getMBB() != LoopBB) 55 return Phi.getOperand(i).getReg(); 56 return 0; 57 } 58 59 /// Return the Phi register value that comes the loop block. 60 static unsigned getLoopPhiReg(MachineInstr &Phi, MachineBasicBlock *LoopBB) { 61 for (unsigned i = 1, e = Phi.getNumOperands(); i != e; i += 2) 62 if (Phi.getOperand(i + 1).getMBB() == LoopBB) 63 return Phi.getOperand(i).getReg(); 64 return 0; 65 } 66 67 void ModuloScheduleExpander::expand() { 68 BB = Schedule.getLoop()->getTopBlock(); 69 Preheader = *BB->pred_begin(); 70 if (Preheader == BB) 71 Preheader = *std::next(BB->pred_begin()); 72 73 // Iterate over the definitions in each instruction, and compute the 74 // stage difference for each use. Keep the maximum value. 75 for (MachineInstr *MI : Schedule.getInstructions()) { 76 int DefStage = Schedule.getStage(MI); 77 for (const MachineOperand &Op : MI->all_defs()) { 78 Register Reg = Op.getReg(); 79 unsigned MaxDiff = 0; 80 bool PhiIsSwapped = false; 81 for (MachineOperand &UseOp : MRI.use_operands(Reg)) { 82 MachineInstr *UseMI = UseOp.getParent(); 83 int UseStage = Schedule.getStage(UseMI); 84 unsigned Diff = 0; 85 if (UseStage != -1 && UseStage >= DefStage) 86 Diff = UseStage - DefStage; 87 if (MI->isPHI()) { 88 if (isLoopCarried(*MI)) 89 ++Diff; 90 else 91 PhiIsSwapped = true; 92 } 93 MaxDiff = std::max(Diff, MaxDiff); 94 } 95 RegToStageDiff[Reg] = std::make_pair(MaxDiff, PhiIsSwapped); 96 } 97 } 98 99 generatePipelinedLoop(); 100 } 101 102 void ModuloScheduleExpander::generatePipelinedLoop() { 103 LoopInfo = TII->analyzeLoopForPipelining(BB); 104 assert(LoopInfo && "Must be able to analyze loop!"); 105 106 // Create a new basic block for the kernel and add it to the CFG. 107 MachineBasicBlock *KernelBB = MF.CreateMachineBasicBlock(BB->getBasicBlock()); 108 109 unsigned MaxStageCount = Schedule.getNumStages() - 1; 110 111 // Remember the registers that are used in different stages. The index is 112 // the iteration, or stage, that the instruction is scheduled in. This is 113 // a map between register names in the original block and the names created 114 // in each stage of the pipelined loop. 115 ValueMapTy *VRMap = new ValueMapTy[(MaxStageCount + 1) * 2]; 116 117 // The renaming destination by Phis for the registers across stages. 118 // This map is updated during Phis generation to point to the most recent 119 // renaming destination. 120 ValueMapTy *VRMapPhi = new ValueMapTy[(MaxStageCount + 1) * 2]; 121 122 InstrMapTy InstrMap; 123 124 SmallVector<MachineBasicBlock *, 4> PrologBBs; 125 126 // Generate the prolog instructions that set up the pipeline. 127 generateProlog(MaxStageCount, KernelBB, VRMap, PrologBBs); 128 MF.insert(BB->getIterator(), KernelBB); 129 130 // Rearrange the instructions to generate the new, pipelined loop, 131 // and update register names as needed. 132 for (MachineInstr *CI : Schedule.getInstructions()) { 133 if (CI->isPHI()) 134 continue; 135 unsigned StageNum = Schedule.getStage(CI); 136 MachineInstr *NewMI = cloneInstr(CI, MaxStageCount, StageNum); 137 updateInstruction(NewMI, false, MaxStageCount, StageNum, VRMap); 138 KernelBB->push_back(NewMI); 139 InstrMap[NewMI] = CI; 140 } 141 142 // Copy any terminator instructions to the new kernel, and update 143 // names as needed. 144 for (MachineInstr &MI : BB->terminators()) { 145 MachineInstr *NewMI = MF.CloneMachineInstr(&MI); 146 updateInstruction(NewMI, false, MaxStageCount, 0, VRMap); 147 KernelBB->push_back(NewMI); 148 InstrMap[NewMI] = &MI; 149 } 150 151 NewKernel = KernelBB; 152 KernelBB->transferSuccessors(BB); 153 KernelBB->replaceSuccessor(BB, KernelBB); 154 155 generateExistingPhis(KernelBB, PrologBBs.back(), KernelBB, KernelBB, VRMap, 156 InstrMap, MaxStageCount, MaxStageCount, false); 157 generatePhis(KernelBB, PrologBBs.back(), KernelBB, KernelBB, VRMap, VRMapPhi, 158 InstrMap, MaxStageCount, MaxStageCount, false); 159 160 LLVM_DEBUG(dbgs() << "New block\n"; KernelBB->dump();); 161 162 SmallVector<MachineBasicBlock *, 4> EpilogBBs; 163 // Generate the epilog instructions to complete the pipeline. 164 generateEpilog(MaxStageCount, KernelBB, BB, VRMap, VRMapPhi, EpilogBBs, 165 PrologBBs); 166 167 // We need this step because the register allocation doesn't handle some 168 // situations well, so we insert copies to help out. 169 splitLifetimes(KernelBB, EpilogBBs); 170 171 // Remove dead instructions due to loop induction variables. 172 removeDeadInstructions(KernelBB, EpilogBBs); 173 174 // Add branches between prolog and epilog blocks. 175 addBranches(*Preheader, PrologBBs, KernelBB, EpilogBBs, VRMap); 176 177 delete[] VRMap; 178 delete[] VRMapPhi; 179 } 180 181 void ModuloScheduleExpander::cleanup() { 182 // Remove the original loop since it's no longer referenced. 183 for (auto &I : *BB) 184 LIS.RemoveMachineInstrFromMaps(I); 185 BB->clear(); 186 BB->eraseFromParent(); 187 } 188 189 /// Generate the pipeline prolog code. 190 void ModuloScheduleExpander::generateProlog(unsigned LastStage, 191 MachineBasicBlock *KernelBB, 192 ValueMapTy *VRMap, 193 MBBVectorTy &PrologBBs) { 194 MachineBasicBlock *PredBB = Preheader; 195 InstrMapTy InstrMap; 196 197 // Generate a basic block for each stage, not including the last stage, 198 // which will be generated in the kernel. Each basic block may contain 199 // instructions from multiple stages/iterations. 200 for (unsigned i = 0; i < LastStage; ++i) { 201 // Create and insert the prolog basic block prior to the original loop 202 // basic block. The original loop is removed later. 203 MachineBasicBlock *NewBB = MF.CreateMachineBasicBlock(BB->getBasicBlock()); 204 PrologBBs.push_back(NewBB); 205 MF.insert(BB->getIterator(), NewBB); 206 NewBB->transferSuccessors(PredBB); 207 PredBB->addSuccessor(NewBB); 208 PredBB = NewBB; 209 210 // Generate instructions for each appropriate stage. Process instructions 211 // in original program order. 212 for (int StageNum = i; StageNum >= 0; --StageNum) { 213 for (MachineBasicBlock::iterator BBI = BB->instr_begin(), 214 BBE = BB->getFirstTerminator(); 215 BBI != BBE; ++BBI) { 216 if (Schedule.getStage(&*BBI) == StageNum) { 217 if (BBI->isPHI()) 218 continue; 219 MachineInstr *NewMI = 220 cloneAndChangeInstr(&*BBI, i, (unsigned)StageNum); 221 updateInstruction(NewMI, false, i, (unsigned)StageNum, VRMap); 222 NewBB->push_back(NewMI); 223 InstrMap[NewMI] = &*BBI; 224 } 225 } 226 } 227 rewritePhiValues(NewBB, i, VRMap, InstrMap); 228 LLVM_DEBUG({ 229 dbgs() << "prolog:\n"; 230 NewBB->dump(); 231 }); 232 } 233 234 PredBB->replaceSuccessor(BB, KernelBB); 235 236 // Check if we need to remove the branch from the preheader to the original 237 // loop, and replace it with a branch to the new loop. 238 unsigned numBranches = TII->removeBranch(*Preheader); 239 if (numBranches) { 240 SmallVector<MachineOperand, 0> Cond; 241 TII->insertBranch(*Preheader, PrologBBs[0], nullptr, Cond, DebugLoc()); 242 } 243 } 244 245 /// Generate the pipeline epilog code. The epilog code finishes the iterations 246 /// that were started in either the prolog or the kernel. We create a basic 247 /// block for each stage that needs to complete. 248 void ModuloScheduleExpander::generateEpilog( 249 unsigned LastStage, MachineBasicBlock *KernelBB, MachineBasicBlock *OrigBB, 250 ValueMapTy *VRMap, ValueMapTy *VRMapPhi, MBBVectorTy &EpilogBBs, 251 MBBVectorTy &PrologBBs) { 252 // We need to change the branch from the kernel to the first epilog block, so 253 // this call to analyze branch uses the kernel rather than the original BB. 254 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; 255 SmallVector<MachineOperand, 4> Cond; 256 bool checkBranch = TII->analyzeBranch(*KernelBB, TBB, FBB, Cond); 257 assert(!checkBranch && "generateEpilog must be able to analyze the branch"); 258 if (checkBranch) 259 return; 260 261 MachineBasicBlock::succ_iterator LoopExitI = KernelBB->succ_begin(); 262 if (*LoopExitI == KernelBB) 263 ++LoopExitI; 264 assert(LoopExitI != KernelBB->succ_end() && "Expecting a successor"); 265 MachineBasicBlock *LoopExitBB = *LoopExitI; 266 267 MachineBasicBlock *PredBB = KernelBB; 268 MachineBasicBlock *EpilogStart = LoopExitBB; 269 InstrMapTy InstrMap; 270 271 // Generate a basic block for each stage, not including the last stage, 272 // which was generated for the kernel. Each basic block may contain 273 // instructions from multiple stages/iterations. 274 int EpilogStage = LastStage + 1; 275 for (unsigned i = LastStage; i >= 1; --i, ++EpilogStage) { 276 MachineBasicBlock *NewBB = MF.CreateMachineBasicBlock(); 277 EpilogBBs.push_back(NewBB); 278 MF.insert(BB->getIterator(), NewBB); 279 280 PredBB->replaceSuccessor(LoopExitBB, NewBB); 281 NewBB->addSuccessor(LoopExitBB); 282 283 if (EpilogStart == LoopExitBB) 284 EpilogStart = NewBB; 285 286 // Add instructions to the epilog depending on the current block. 287 // Process instructions in original program order. 288 for (unsigned StageNum = i; StageNum <= LastStage; ++StageNum) { 289 for (auto &BBI : *BB) { 290 if (BBI.isPHI()) 291 continue; 292 MachineInstr *In = &BBI; 293 if ((unsigned)Schedule.getStage(In) == StageNum) { 294 // Instructions with memoperands in the epilog are updated with 295 // conservative values. 296 MachineInstr *NewMI = cloneInstr(In, UINT_MAX, 0); 297 updateInstruction(NewMI, i == 1, EpilogStage, 0, VRMap); 298 NewBB->push_back(NewMI); 299 InstrMap[NewMI] = In; 300 } 301 } 302 } 303 generateExistingPhis(NewBB, PrologBBs[i - 1], PredBB, KernelBB, VRMap, 304 InstrMap, LastStage, EpilogStage, i == 1); 305 generatePhis(NewBB, PrologBBs[i - 1], PredBB, KernelBB, VRMap, VRMapPhi, 306 InstrMap, LastStage, EpilogStage, i == 1); 307 PredBB = NewBB; 308 309 LLVM_DEBUG({ 310 dbgs() << "epilog:\n"; 311 NewBB->dump(); 312 }); 313 } 314 315 // Fix any Phi nodes in the loop exit block. 316 LoopExitBB->replacePhiUsesWith(BB, PredBB); 317 318 // Create a branch to the new epilog from the kernel. 319 // Remove the original branch and add a new branch to the epilog. 320 TII->removeBranch(*KernelBB); 321 assert((OrigBB == TBB || OrigBB == FBB) && 322 "Unable to determine looping branch direction"); 323 if (OrigBB != TBB) 324 TII->insertBranch(*KernelBB, EpilogStart, KernelBB, Cond, DebugLoc()); 325 else 326 TII->insertBranch(*KernelBB, KernelBB, EpilogStart, Cond, DebugLoc()); 327 // Add a branch to the loop exit. 328 if (EpilogBBs.size() > 0) { 329 MachineBasicBlock *LastEpilogBB = EpilogBBs.back(); 330 SmallVector<MachineOperand, 4> Cond1; 331 TII->insertBranch(*LastEpilogBB, LoopExitBB, nullptr, Cond1, DebugLoc()); 332 } 333 } 334 335 /// Replace all uses of FromReg that appear outside the specified 336 /// basic block with ToReg. 337 static void replaceRegUsesAfterLoop(unsigned FromReg, unsigned ToReg, 338 MachineBasicBlock *MBB, 339 MachineRegisterInfo &MRI, 340 LiveIntervals &LIS) { 341 for (MachineOperand &O : 342 llvm::make_early_inc_range(MRI.use_operands(FromReg))) 343 if (O.getParent()->getParent() != MBB) 344 O.setReg(ToReg); 345 if (!LIS.hasInterval(ToReg)) 346 LIS.createEmptyInterval(ToReg); 347 } 348 349 /// Return true if the register has a use that occurs outside the 350 /// specified loop. 351 static bool hasUseAfterLoop(unsigned Reg, MachineBasicBlock *BB, 352 MachineRegisterInfo &MRI) { 353 for (const MachineOperand &MO : MRI.use_operands(Reg)) 354 if (MO.getParent()->getParent() != BB) 355 return true; 356 return false; 357 } 358 359 /// Generate Phis for the specific block in the generated pipelined code. 360 /// This function looks at the Phis from the original code to guide the 361 /// creation of new Phis. 362 void ModuloScheduleExpander::generateExistingPhis( 363 MachineBasicBlock *NewBB, MachineBasicBlock *BB1, MachineBasicBlock *BB2, 364 MachineBasicBlock *KernelBB, ValueMapTy *VRMap, InstrMapTy &InstrMap, 365 unsigned LastStageNum, unsigned CurStageNum, bool IsLast) { 366 // Compute the stage number for the initial value of the Phi, which 367 // comes from the prolog. The prolog to use depends on to which kernel/ 368 // epilog that we're adding the Phi. 369 unsigned PrologStage = 0; 370 unsigned PrevStage = 0; 371 bool InKernel = (LastStageNum == CurStageNum); 372 if (InKernel) { 373 PrologStage = LastStageNum - 1; 374 PrevStage = CurStageNum; 375 } else { 376 PrologStage = LastStageNum - (CurStageNum - LastStageNum); 377 PrevStage = LastStageNum + (CurStageNum - LastStageNum) - 1; 378 } 379 380 for (MachineBasicBlock::iterator BBI = BB->instr_begin(), 381 BBE = BB->getFirstNonPHI(); 382 BBI != BBE; ++BBI) { 383 Register Def = BBI->getOperand(0).getReg(); 384 385 unsigned InitVal = 0; 386 unsigned LoopVal = 0; 387 getPhiRegs(*BBI, BB, InitVal, LoopVal); 388 389 unsigned PhiOp1 = 0; 390 // The Phi value from the loop body typically is defined in the loop, but 391 // not always. So, we need to check if the value is defined in the loop. 392 unsigned PhiOp2 = LoopVal; 393 if (VRMap[LastStageNum].count(LoopVal)) 394 PhiOp2 = VRMap[LastStageNum][LoopVal]; 395 396 int StageScheduled = Schedule.getStage(&*BBI); 397 int LoopValStage = Schedule.getStage(MRI.getVRegDef(LoopVal)); 398 unsigned NumStages = getStagesForReg(Def, CurStageNum); 399 if (NumStages == 0) { 400 // We don't need to generate a Phi anymore, but we need to rename any uses 401 // of the Phi value. 402 unsigned NewReg = VRMap[PrevStage][LoopVal]; 403 rewriteScheduledInstr(NewBB, InstrMap, CurStageNum, 0, &*BBI, Def, 404 InitVal, NewReg); 405 if (VRMap[CurStageNum].count(LoopVal)) 406 VRMap[CurStageNum][Def] = VRMap[CurStageNum][LoopVal]; 407 } 408 // Adjust the number of Phis needed depending on the number of prologs left, 409 // and the distance from where the Phi is first scheduled. The number of 410 // Phis cannot exceed the number of prolog stages. Each stage can 411 // potentially define two values. 412 unsigned MaxPhis = PrologStage + 2; 413 if (!InKernel && (int)PrologStage <= LoopValStage) 414 MaxPhis = std::max((int)MaxPhis - (int)LoopValStage, 1); 415 unsigned NumPhis = std::min(NumStages, MaxPhis); 416 417 unsigned NewReg = 0; 418 unsigned AccessStage = (LoopValStage != -1) ? LoopValStage : StageScheduled; 419 // In the epilog, we may need to look back one stage to get the correct 420 // Phi name, because the epilog and prolog blocks execute the same stage. 421 // The correct name is from the previous block only when the Phi has 422 // been completely scheduled prior to the epilog, and Phi value is not 423 // needed in multiple stages. 424 int StageDiff = 0; 425 if (!InKernel && StageScheduled >= LoopValStage && AccessStage == 0 && 426 NumPhis == 1) 427 StageDiff = 1; 428 // Adjust the computations below when the phi and the loop definition 429 // are scheduled in different stages. 430 if (InKernel && LoopValStage != -1 && StageScheduled > LoopValStage) 431 StageDiff = StageScheduled - LoopValStage; 432 for (unsigned np = 0; np < NumPhis; ++np) { 433 // If the Phi hasn't been scheduled, then use the initial Phi operand 434 // value. Otherwise, use the scheduled version of the instruction. This 435 // is a little complicated when a Phi references another Phi. 436 if (np > PrologStage || StageScheduled >= (int)LastStageNum) 437 PhiOp1 = InitVal; 438 // Check if the Phi has already been scheduled in a prolog stage. 439 else if (PrologStage >= AccessStage + StageDiff + np && 440 VRMap[PrologStage - StageDiff - np].count(LoopVal) != 0) 441 PhiOp1 = VRMap[PrologStage - StageDiff - np][LoopVal]; 442 // Check if the Phi has already been scheduled, but the loop instruction 443 // is either another Phi, or doesn't occur in the loop. 444 else if (PrologStage >= AccessStage + StageDiff + np) { 445 // If the Phi references another Phi, we need to examine the other 446 // Phi to get the correct value. 447 PhiOp1 = LoopVal; 448 MachineInstr *InstOp1 = MRI.getVRegDef(PhiOp1); 449 int Indirects = 1; 450 while (InstOp1 && InstOp1->isPHI() && InstOp1->getParent() == BB) { 451 int PhiStage = Schedule.getStage(InstOp1); 452 if ((int)(PrologStage - StageDiff - np) < PhiStage + Indirects) 453 PhiOp1 = getInitPhiReg(*InstOp1, BB); 454 else 455 PhiOp1 = getLoopPhiReg(*InstOp1, BB); 456 InstOp1 = MRI.getVRegDef(PhiOp1); 457 int PhiOpStage = Schedule.getStage(InstOp1); 458 int StageAdj = (PhiOpStage != -1 ? PhiStage - PhiOpStage : 0); 459 if (PhiOpStage != -1 && PrologStage - StageAdj >= Indirects + np && 460 VRMap[PrologStage - StageAdj - Indirects - np].count(PhiOp1)) { 461 PhiOp1 = VRMap[PrologStage - StageAdj - Indirects - np][PhiOp1]; 462 break; 463 } 464 ++Indirects; 465 } 466 } else 467 PhiOp1 = InitVal; 468 // If this references a generated Phi in the kernel, get the Phi operand 469 // from the incoming block. 470 if (MachineInstr *InstOp1 = MRI.getVRegDef(PhiOp1)) 471 if (InstOp1->isPHI() && InstOp1->getParent() == KernelBB) 472 PhiOp1 = getInitPhiReg(*InstOp1, KernelBB); 473 474 MachineInstr *PhiInst = MRI.getVRegDef(LoopVal); 475 bool LoopDefIsPhi = PhiInst && PhiInst->isPHI(); 476 // In the epilog, a map lookup is needed to get the value from the kernel, 477 // or previous epilog block. How is does this depends on if the 478 // instruction is scheduled in the previous block. 479 if (!InKernel) { 480 int StageDiffAdj = 0; 481 if (LoopValStage != -1 && StageScheduled > LoopValStage) 482 StageDiffAdj = StageScheduled - LoopValStage; 483 // Use the loop value defined in the kernel, unless the kernel 484 // contains the last definition of the Phi. 485 if (np == 0 && PrevStage == LastStageNum && 486 (StageScheduled != 0 || LoopValStage != 0) && 487 VRMap[PrevStage - StageDiffAdj].count(LoopVal)) 488 PhiOp2 = VRMap[PrevStage - StageDiffAdj][LoopVal]; 489 // Use the value defined by the Phi. We add one because we switch 490 // from looking at the loop value to the Phi definition. 491 else if (np > 0 && PrevStage == LastStageNum && 492 VRMap[PrevStage - np + 1].count(Def)) 493 PhiOp2 = VRMap[PrevStage - np + 1][Def]; 494 // Use the loop value defined in the kernel. 495 else if (static_cast<unsigned>(LoopValStage) > PrologStage + 1 && 496 VRMap[PrevStage - StageDiffAdj - np].count(LoopVal)) 497 PhiOp2 = VRMap[PrevStage - StageDiffAdj - np][LoopVal]; 498 // Use the value defined by the Phi, unless we're generating the first 499 // epilog and the Phi refers to a Phi in a different stage. 500 else if (VRMap[PrevStage - np].count(Def) && 501 (!LoopDefIsPhi || (PrevStage != LastStageNum) || 502 (LoopValStage == StageScheduled))) 503 PhiOp2 = VRMap[PrevStage - np][Def]; 504 } 505 506 // Check if we can reuse an existing Phi. This occurs when a Phi 507 // references another Phi, and the other Phi is scheduled in an 508 // earlier stage. We can try to reuse an existing Phi up until the last 509 // stage of the current Phi. 510 if (LoopDefIsPhi) { 511 if (static_cast<int>(PrologStage - np) >= StageScheduled) { 512 int LVNumStages = getStagesForPhi(LoopVal); 513 int StageDiff = (StageScheduled - LoopValStage); 514 LVNumStages -= StageDiff; 515 // Make sure the loop value Phi has been processed already. 516 if (LVNumStages > (int)np && VRMap[CurStageNum].count(LoopVal)) { 517 NewReg = PhiOp2; 518 unsigned ReuseStage = CurStageNum; 519 if (isLoopCarried(*PhiInst)) 520 ReuseStage -= LVNumStages; 521 // Check if the Phi to reuse has been generated yet. If not, then 522 // there is nothing to reuse. 523 if (VRMap[ReuseStage - np].count(LoopVal)) { 524 NewReg = VRMap[ReuseStage - np][LoopVal]; 525 526 rewriteScheduledInstr(NewBB, InstrMap, CurStageNum, np, &*BBI, 527 Def, NewReg); 528 // Update the map with the new Phi name. 529 VRMap[CurStageNum - np][Def] = NewReg; 530 PhiOp2 = NewReg; 531 if (VRMap[LastStageNum - np - 1].count(LoopVal)) 532 PhiOp2 = VRMap[LastStageNum - np - 1][LoopVal]; 533 534 if (IsLast && np == NumPhis - 1) 535 replaceRegUsesAfterLoop(Def, NewReg, BB, MRI, LIS); 536 continue; 537 } 538 } 539 } 540 if (InKernel && StageDiff > 0 && 541 VRMap[CurStageNum - StageDiff - np].count(LoopVal)) 542 PhiOp2 = VRMap[CurStageNum - StageDiff - np][LoopVal]; 543 } 544 545 const TargetRegisterClass *RC = MRI.getRegClass(Def); 546 NewReg = MRI.createVirtualRegister(RC); 547 548 MachineInstrBuilder NewPhi = 549 BuildMI(*NewBB, NewBB->getFirstNonPHI(), DebugLoc(), 550 TII->get(TargetOpcode::PHI), NewReg); 551 NewPhi.addReg(PhiOp1).addMBB(BB1); 552 NewPhi.addReg(PhiOp2).addMBB(BB2); 553 if (np == 0) 554 InstrMap[NewPhi] = &*BBI; 555 556 // We define the Phis after creating the new pipelined code, so 557 // we need to rename the Phi values in scheduled instructions. 558 559 unsigned PrevReg = 0; 560 if (InKernel && VRMap[PrevStage - np].count(LoopVal)) 561 PrevReg = VRMap[PrevStage - np][LoopVal]; 562 rewriteScheduledInstr(NewBB, InstrMap, CurStageNum, np, &*BBI, Def, 563 NewReg, PrevReg); 564 // If the Phi has been scheduled, use the new name for rewriting. 565 if (VRMap[CurStageNum - np].count(Def)) { 566 unsigned R = VRMap[CurStageNum - np][Def]; 567 rewriteScheduledInstr(NewBB, InstrMap, CurStageNum, np, &*BBI, R, 568 NewReg); 569 } 570 571 // Check if we need to rename any uses that occurs after the loop. The 572 // register to replace depends on whether the Phi is scheduled in the 573 // epilog. 574 if (IsLast && np == NumPhis - 1) 575 replaceRegUsesAfterLoop(Def, NewReg, BB, MRI, LIS); 576 577 // In the kernel, a dependent Phi uses the value from this Phi. 578 if (InKernel) 579 PhiOp2 = NewReg; 580 581 // Update the map with the new Phi name. 582 VRMap[CurStageNum - np][Def] = NewReg; 583 } 584 585 while (NumPhis++ < NumStages) { 586 rewriteScheduledInstr(NewBB, InstrMap, CurStageNum, NumPhis, &*BBI, Def, 587 NewReg, 0); 588 } 589 590 // Check if we need to rename a Phi that has been eliminated due to 591 // scheduling. 592 if (NumStages == 0 && IsLast && VRMap[CurStageNum].count(LoopVal)) 593 replaceRegUsesAfterLoop(Def, VRMap[CurStageNum][LoopVal], BB, MRI, LIS); 594 } 595 } 596 597 /// Generate Phis for the specified block in the generated pipelined code. 598 /// These are new Phis needed because the definition is scheduled after the 599 /// use in the pipelined sequence. 600 void ModuloScheduleExpander::generatePhis( 601 MachineBasicBlock *NewBB, MachineBasicBlock *BB1, MachineBasicBlock *BB2, 602 MachineBasicBlock *KernelBB, ValueMapTy *VRMap, ValueMapTy *VRMapPhi, 603 InstrMapTy &InstrMap, unsigned LastStageNum, unsigned CurStageNum, 604 bool IsLast) { 605 // Compute the stage number that contains the initial Phi value, and 606 // the Phi from the previous stage. 607 unsigned PrologStage = 0; 608 unsigned PrevStage = 0; 609 unsigned StageDiff = CurStageNum - LastStageNum; 610 bool InKernel = (StageDiff == 0); 611 if (InKernel) { 612 PrologStage = LastStageNum - 1; 613 PrevStage = CurStageNum; 614 } else { 615 PrologStage = LastStageNum - StageDiff; 616 PrevStage = LastStageNum + StageDiff - 1; 617 } 618 619 for (MachineBasicBlock::iterator BBI = BB->getFirstNonPHI(), 620 BBE = BB->instr_end(); 621 BBI != BBE; ++BBI) { 622 for (unsigned i = 0, e = BBI->getNumOperands(); i != e; ++i) { 623 MachineOperand &MO = BBI->getOperand(i); 624 if (!MO.isReg() || !MO.isDef() || !MO.getReg().isVirtual()) 625 continue; 626 627 int StageScheduled = Schedule.getStage(&*BBI); 628 assert(StageScheduled != -1 && "Expecting scheduled instruction."); 629 Register Def = MO.getReg(); 630 unsigned NumPhis = getStagesForReg(Def, CurStageNum); 631 // An instruction scheduled in stage 0 and is used after the loop 632 // requires a phi in the epilog for the last definition from either 633 // the kernel or prolog. 634 if (!InKernel && NumPhis == 0 && StageScheduled == 0 && 635 hasUseAfterLoop(Def, BB, MRI)) 636 NumPhis = 1; 637 if (!InKernel && (unsigned)StageScheduled > PrologStage) 638 continue; 639 640 unsigned PhiOp2; 641 if (InKernel) { 642 PhiOp2 = VRMap[PrevStage][Def]; 643 if (MachineInstr *InstOp2 = MRI.getVRegDef(PhiOp2)) 644 if (InstOp2->isPHI() && InstOp2->getParent() == NewBB) 645 PhiOp2 = getLoopPhiReg(*InstOp2, BB2); 646 } 647 // The number of Phis can't exceed the number of prolog stages. The 648 // prolog stage number is zero based. 649 if (NumPhis > PrologStage + 1 - StageScheduled) 650 NumPhis = PrologStage + 1 - StageScheduled; 651 for (unsigned np = 0; np < NumPhis; ++np) { 652 // Example for 653 // Org: 654 // %Org = ... (Scheduled at Stage#0, NumPhi = 2) 655 // 656 // Prolog0 (Stage0): 657 // %Clone0 = ... 658 // Prolog1 (Stage1): 659 // %Clone1 = ... 660 // Kernel (Stage2): 661 // %Phi0 = Phi %Clone1, Prolog1, %Clone2, Kernel 662 // %Phi1 = Phi %Clone0, Prolog1, %Phi0, Kernel 663 // %Clone2 = ... 664 // Epilog0 (Stage3): 665 // %Phi2 = Phi %Clone1, Prolog1, %Clone2, Kernel 666 // %Phi3 = Phi %Clone0, Prolog1, %Phi0, Kernel 667 // Epilog1 (Stage4): 668 // %Phi4 = Phi %Clone0, Prolog0, %Phi2, Epilog0 669 // 670 // VRMap = {0: %Clone0, 1: %Clone1, 2: %Clone2} 671 // VRMapPhi (after Kernel) = {0: %Phi1, 1: %Phi0} 672 // VRMapPhi (after Epilog0) = {0: %Phi3, 1: %Phi2} 673 674 unsigned PhiOp1 = VRMap[PrologStage][Def]; 675 if (np <= PrologStage) 676 PhiOp1 = VRMap[PrologStage - np][Def]; 677 if (!InKernel) { 678 if (PrevStage == LastStageNum && np == 0) 679 PhiOp2 = VRMap[LastStageNum][Def]; 680 else 681 PhiOp2 = VRMapPhi[PrevStage - np][Def]; 682 } 683 684 const TargetRegisterClass *RC = MRI.getRegClass(Def); 685 Register NewReg = MRI.createVirtualRegister(RC); 686 687 MachineInstrBuilder NewPhi = 688 BuildMI(*NewBB, NewBB->getFirstNonPHI(), DebugLoc(), 689 TII->get(TargetOpcode::PHI), NewReg); 690 NewPhi.addReg(PhiOp1).addMBB(BB1); 691 NewPhi.addReg(PhiOp2).addMBB(BB2); 692 if (np == 0) 693 InstrMap[NewPhi] = &*BBI; 694 695 // Rewrite uses and update the map. The actions depend upon whether 696 // we generating code for the kernel or epilog blocks. 697 if (InKernel) { 698 rewriteScheduledInstr(NewBB, InstrMap, CurStageNum, np, &*BBI, PhiOp1, 699 NewReg); 700 rewriteScheduledInstr(NewBB, InstrMap, CurStageNum, np, &*BBI, PhiOp2, 701 NewReg); 702 703 PhiOp2 = NewReg; 704 VRMapPhi[PrevStage - np - 1][Def] = NewReg; 705 } else { 706 VRMapPhi[CurStageNum - np][Def] = NewReg; 707 if (np == NumPhis - 1) 708 rewriteScheduledInstr(NewBB, InstrMap, CurStageNum, np, &*BBI, Def, 709 NewReg); 710 } 711 if (IsLast && np == NumPhis - 1) 712 replaceRegUsesAfterLoop(Def, NewReg, BB, MRI, LIS); 713 } 714 } 715 } 716 } 717 718 /// Remove instructions that generate values with no uses. 719 /// Typically, these are induction variable operations that generate values 720 /// used in the loop itself. A dead instruction has a definition with 721 /// no uses, or uses that occur in the original loop only. 722 void ModuloScheduleExpander::removeDeadInstructions(MachineBasicBlock *KernelBB, 723 MBBVectorTy &EpilogBBs) { 724 // For each epilog block, check that the value defined by each instruction 725 // is used. If not, delete it. 726 for (MachineBasicBlock *MBB : llvm::reverse(EpilogBBs)) 727 for (MachineBasicBlock::reverse_instr_iterator MI = MBB->instr_rbegin(), 728 ME = MBB->instr_rend(); 729 MI != ME;) { 730 // From DeadMachineInstructionElem. Don't delete inline assembly. 731 if (MI->isInlineAsm()) { 732 ++MI; 733 continue; 734 } 735 bool SawStore = false; 736 // Check if it's safe to remove the instruction due to side effects. 737 // We can, and want to, remove Phis here. 738 if (!MI->isSafeToMove(nullptr, SawStore) && !MI->isPHI()) { 739 ++MI; 740 continue; 741 } 742 bool used = true; 743 for (const MachineOperand &MO : MI->all_defs()) { 744 Register reg = MO.getReg(); 745 // Assume physical registers are used, unless they are marked dead. 746 if (reg.isPhysical()) { 747 used = !MO.isDead(); 748 if (used) 749 break; 750 continue; 751 } 752 unsigned realUses = 0; 753 for (const MachineOperand &U : MRI.use_operands(reg)) { 754 // Check if there are any uses that occur only in the original 755 // loop. If so, that's not a real use. 756 if (U.getParent()->getParent() != BB) { 757 realUses++; 758 used = true; 759 break; 760 } 761 } 762 if (realUses > 0) 763 break; 764 used = false; 765 } 766 if (!used) { 767 LIS.RemoveMachineInstrFromMaps(*MI); 768 MI++->eraseFromParent(); 769 continue; 770 } 771 ++MI; 772 } 773 // In the kernel block, check if we can remove a Phi that generates a value 774 // used in an instruction removed in the epilog block. 775 for (MachineInstr &MI : llvm::make_early_inc_range(KernelBB->phis())) { 776 Register reg = MI.getOperand(0).getReg(); 777 if (MRI.use_begin(reg) == MRI.use_end()) { 778 LIS.RemoveMachineInstrFromMaps(MI); 779 MI.eraseFromParent(); 780 } 781 } 782 } 783 784 /// For loop carried definitions, we split the lifetime of a virtual register 785 /// that has uses past the definition in the next iteration. A copy with a new 786 /// virtual register is inserted before the definition, which helps with 787 /// generating a better register assignment. 788 /// 789 /// v1 = phi(a, v2) v1 = phi(a, v2) 790 /// v2 = phi(b, v3) v2 = phi(b, v3) 791 /// v3 = .. v4 = copy v1 792 /// .. = V1 v3 = .. 793 /// .. = v4 794 void ModuloScheduleExpander::splitLifetimes(MachineBasicBlock *KernelBB, 795 MBBVectorTy &EpilogBBs) { 796 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 797 for (auto &PHI : KernelBB->phis()) { 798 Register Def = PHI.getOperand(0).getReg(); 799 // Check for any Phi definition that used as an operand of another Phi 800 // in the same block. 801 for (MachineRegisterInfo::use_instr_iterator I = MRI.use_instr_begin(Def), 802 E = MRI.use_instr_end(); 803 I != E; ++I) { 804 if (I->isPHI() && I->getParent() == KernelBB) { 805 // Get the loop carried definition. 806 unsigned LCDef = getLoopPhiReg(PHI, KernelBB); 807 if (!LCDef) 808 continue; 809 MachineInstr *MI = MRI.getVRegDef(LCDef); 810 if (!MI || MI->getParent() != KernelBB || MI->isPHI()) 811 continue; 812 // Search through the rest of the block looking for uses of the Phi 813 // definition. If one occurs, then split the lifetime. 814 unsigned SplitReg = 0; 815 for (auto &BBJ : make_range(MachineBasicBlock::instr_iterator(MI), 816 KernelBB->instr_end())) 817 if (BBJ.readsRegister(Def)) { 818 // We split the lifetime when we find the first use. 819 if (SplitReg == 0) { 820 SplitReg = MRI.createVirtualRegister(MRI.getRegClass(Def)); 821 BuildMI(*KernelBB, MI, MI->getDebugLoc(), 822 TII->get(TargetOpcode::COPY), SplitReg) 823 .addReg(Def); 824 } 825 BBJ.substituteRegister(Def, SplitReg, 0, *TRI); 826 } 827 if (!SplitReg) 828 continue; 829 // Search through each of the epilog blocks for any uses to be renamed. 830 for (auto &Epilog : EpilogBBs) 831 for (auto &I : *Epilog) 832 if (I.readsRegister(Def)) 833 I.substituteRegister(Def, SplitReg, 0, *TRI); 834 break; 835 } 836 } 837 } 838 } 839 840 /// Remove the incoming block from the Phis in a basic block. 841 static void removePhis(MachineBasicBlock *BB, MachineBasicBlock *Incoming) { 842 for (MachineInstr &MI : *BB) { 843 if (!MI.isPHI()) 844 break; 845 for (unsigned i = 1, e = MI.getNumOperands(); i != e; i += 2) 846 if (MI.getOperand(i + 1).getMBB() == Incoming) { 847 MI.removeOperand(i + 1); 848 MI.removeOperand(i); 849 break; 850 } 851 } 852 } 853 854 /// Create branches from each prolog basic block to the appropriate epilog 855 /// block. These edges are needed if the loop ends before reaching the 856 /// kernel. 857 void ModuloScheduleExpander::addBranches(MachineBasicBlock &PreheaderBB, 858 MBBVectorTy &PrologBBs, 859 MachineBasicBlock *KernelBB, 860 MBBVectorTy &EpilogBBs, 861 ValueMapTy *VRMap) { 862 assert(PrologBBs.size() == EpilogBBs.size() && "Prolog/Epilog mismatch"); 863 MachineBasicBlock *LastPro = KernelBB; 864 MachineBasicBlock *LastEpi = KernelBB; 865 866 // Start from the blocks connected to the kernel and work "out" 867 // to the first prolog and the last epilog blocks. 868 SmallVector<MachineInstr *, 4> PrevInsts; 869 unsigned MaxIter = PrologBBs.size() - 1; 870 for (unsigned i = 0, j = MaxIter; i <= MaxIter; ++i, --j) { 871 // Add branches to the prolog that go to the corresponding 872 // epilog, and the fall-thru prolog/kernel block. 873 MachineBasicBlock *Prolog = PrologBBs[j]; 874 MachineBasicBlock *Epilog = EpilogBBs[i]; 875 876 SmallVector<MachineOperand, 4> Cond; 877 std::optional<bool> StaticallyGreater = 878 LoopInfo->createTripCountGreaterCondition(j + 1, *Prolog, Cond); 879 unsigned numAdded = 0; 880 if (!StaticallyGreater) { 881 Prolog->addSuccessor(Epilog); 882 numAdded = TII->insertBranch(*Prolog, Epilog, LastPro, Cond, DebugLoc()); 883 } else if (*StaticallyGreater == false) { 884 Prolog->addSuccessor(Epilog); 885 Prolog->removeSuccessor(LastPro); 886 LastEpi->removeSuccessor(Epilog); 887 numAdded = TII->insertBranch(*Prolog, Epilog, nullptr, Cond, DebugLoc()); 888 removePhis(Epilog, LastEpi); 889 // Remove the blocks that are no longer referenced. 890 if (LastPro != LastEpi) { 891 LastEpi->clear(); 892 LastEpi->eraseFromParent(); 893 } 894 if (LastPro == KernelBB) { 895 LoopInfo->disposed(); 896 NewKernel = nullptr; 897 } 898 LastPro->clear(); 899 LastPro->eraseFromParent(); 900 } else { 901 numAdded = TII->insertBranch(*Prolog, LastPro, nullptr, Cond, DebugLoc()); 902 removePhis(Epilog, Prolog); 903 } 904 LastPro = Prolog; 905 LastEpi = Epilog; 906 for (MachineBasicBlock::reverse_instr_iterator I = Prolog->instr_rbegin(), 907 E = Prolog->instr_rend(); 908 I != E && numAdded > 0; ++I, --numAdded) 909 updateInstruction(&*I, false, j, 0, VRMap); 910 } 911 912 if (NewKernel) { 913 LoopInfo->setPreheader(PrologBBs[MaxIter]); 914 LoopInfo->adjustTripCount(-(MaxIter + 1)); 915 } 916 } 917 918 /// Return true if we can compute the amount the instruction changes 919 /// during each iteration. Set Delta to the amount of the change. 920 bool ModuloScheduleExpander::computeDelta(MachineInstr &MI, unsigned &Delta) { 921 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 922 const MachineOperand *BaseOp; 923 int64_t Offset; 924 bool OffsetIsScalable; 925 if (!TII->getMemOperandWithOffset(MI, BaseOp, Offset, OffsetIsScalable, TRI)) 926 return false; 927 928 // FIXME: This algorithm assumes instructions have fixed-size offsets. 929 if (OffsetIsScalable) 930 return false; 931 932 if (!BaseOp->isReg()) 933 return false; 934 935 Register BaseReg = BaseOp->getReg(); 936 937 MachineRegisterInfo &MRI = MF.getRegInfo(); 938 // Check if there is a Phi. If so, get the definition in the loop. 939 MachineInstr *BaseDef = MRI.getVRegDef(BaseReg); 940 if (BaseDef && BaseDef->isPHI()) { 941 BaseReg = getLoopPhiReg(*BaseDef, MI.getParent()); 942 BaseDef = MRI.getVRegDef(BaseReg); 943 } 944 if (!BaseDef) 945 return false; 946 947 int D = 0; 948 if (!TII->getIncrementValue(*BaseDef, D) && D >= 0) 949 return false; 950 951 Delta = D; 952 return true; 953 } 954 955 /// Update the memory operand with a new offset when the pipeliner 956 /// generates a new copy of the instruction that refers to a 957 /// different memory location. 958 void ModuloScheduleExpander::updateMemOperands(MachineInstr &NewMI, 959 MachineInstr &OldMI, 960 unsigned Num) { 961 if (Num == 0) 962 return; 963 // If the instruction has memory operands, then adjust the offset 964 // when the instruction appears in different stages. 965 if (NewMI.memoperands_empty()) 966 return; 967 SmallVector<MachineMemOperand *, 2> NewMMOs; 968 for (MachineMemOperand *MMO : NewMI.memoperands()) { 969 // TODO: Figure out whether isAtomic is really necessary (see D57601). 970 if (MMO->isVolatile() || MMO->isAtomic() || 971 (MMO->isInvariant() && MMO->isDereferenceable()) || 972 (!MMO->getValue())) { 973 NewMMOs.push_back(MMO); 974 continue; 975 } 976 unsigned Delta; 977 if (Num != UINT_MAX && computeDelta(OldMI, Delta)) { 978 int64_t AdjOffset = Delta * Num; 979 NewMMOs.push_back( 980 MF.getMachineMemOperand(MMO, AdjOffset, MMO->getSize())); 981 } else { 982 NewMMOs.push_back( 983 MF.getMachineMemOperand(MMO, 0, MemoryLocation::UnknownSize)); 984 } 985 } 986 NewMI.setMemRefs(MF, NewMMOs); 987 } 988 989 /// Clone the instruction for the new pipelined loop and update the 990 /// memory operands, if needed. 991 MachineInstr *ModuloScheduleExpander::cloneInstr(MachineInstr *OldMI, 992 unsigned CurStageNum, 993 unsigned InstStageNum) { 994 MachineInstr *NewMI = MF.CloneMachineInstr(OldMI); 995 updateMemOperands(*NewMI, *OldMI, CurStageNum - InstStageNum); 996 return NewMI; 997 } 998 999 /// Clone the instruction for the new pipelined loop. If needed, this 1000 /// function updates the instruction using the values saved in the 1001 /// InstrChanges structure. 1002 MachineInstr *ModuloScheduleExpander::cloneAndChangeInstr( 1003 MachineInstr *OldMI, unsigned CurStageNum, unsigned InstStageNum) { 1004 MachineInstr *NewMI = MF.CloneMachineInstr(OldMI); 1005 auto It = InstrChanges.find(OldMI); 1006 if (It != InstrChanges.end()) { 1007 std::pair<unsigned, int64_t> RegAndOffset = It->second; 1008 unsigned BasePos, OffsetPos; 1009 if (!TII->getBaseAndOffsetPosition(*OldMI, BasePos, OffsetPos)) 1010 return nullptr; 1011 int64_t NewOffset = OldMI->getOperand(OffsetPos).getImm(); 1012 MachineInstr *LoopDef = findDefInLoop(RegAndOffset.first); 1013 if (Schedule.getStage(LoopDef) > (signed)InstStageNum) 1014 NewOffset += RegAndOffset.second * (CurStageNum - InstStageNum); 1015 NewMI->getOperand(OffsetPos).setImm(NewOffset); 1016 } 1017 updateMemOperands(*NewMI, *OldMI, CurStageNum - InstStageNum); 1018 return NewMI; 1019 } 1020 1021 /// Update the machine instruction with new virtual registers. This 1022 /// function may change the definitions and/or uses. 1023 void ModuloScheduleExpander::updateInstruction(MachineInstr *NewMI, 1024 bool LastDef, 1025 unsigned CurStageNum, 1026 unsigned InstrStageNum, 1027 ValueMapTy *VRMap) { 1028 for (MachineOperand &MO : NewMI->operands()) { 1029 if (!MO.isReg() || !MO.getReg().isVirtual()) 1030 continue; 1031 Register reg = MO.getReg(); 1032 if (MO.isDef()) { 1033 // Create a new virtual register for the definition. 1034 const TargetRegisterClass *RC = MRI.getRegClass(reg); 1035 Register NewReg = MRI.createVirtualRegister(RC); 1036 MO.setReg(NewReg); 1037 VRMap[CurStageNum][reg] = NewReg; 1038 if (LastDef) 1039 replaceRegUsesAfterLoop(reg, NewReg, BB, MRI, LIS); 1040 } else if (MO.isUse()) { 1041 MachineInstr *Def = MRI.getVRegDef(reg); 1042 // Compute the stage that contains the last definition for instruction. 1043 int DefStageNum = Schedule.getStage(Def); 1044 unsigned StageNum = CurStageNum; 1045 if (DefStageNum != -1 && (int)InstrStageNum > DefStageNum) { 1046 // Compute the difference in stages between the defintion and the use. 1047 unsigned StageDiff = (InstrStageNum - DefStageNum); 1048 // Make an adjustment to get the last definition. 1049 StageNum -= StageDiff; 1050 } 1051 if (VRMap[StageNum].count(reg)) 1052 MO.setReg(VRMap[StageNum][reg]); 1053 } 1054 } 1055 } 1056 1057 /// Return the instruction in the loop that defines the register. 1058 /// If the definition is a Phi, then follow the Phi operand to 1059 /// the instruction in the loop. 1060 MachineInstr *ModuloScheduleExpander::findDefInLoop(unsigned Reg) { 1061 SmallPtrSet<MachineInstr *, 8> Visited; 1062 MachineInstr *Def = MRI.getVRegDef(Reg); 1063 while (Def->isPHI()) { 1064 if (!Visited.insert(Def).second) 1065 break; 1066 for (unsigned i = 1, e = Def->getNumOperands(); i < e; i += 2) 1067 if (Def->getOperand(i + 1).getMBB() == BB) { 1068 Def = MRI.getVRegDef(Def->getOperand(i).getReg()); 1069 break; 1070 } 1071 } 1072 return Def; 1073 } 1074 1075 /// Return the new name for the value from the previous stage. 1076 unsigned ModuloScheduleExpander::getPrevMapVal( 1077 unsigned StageNum, unsigned PhiStage, unsigned LoopVal, unsigned LoopStage, 1078 ValueMapTy *VRMap, MachineBasicBlock *BB) { 1079 unsigned PrevVal = 0; 1080 if (StageNum > PhiStage) { 1081 MachineInstr *LoopInst = MRI.getVRegDef(LoopVal); 1082 if (PhiStage == LoopStage && VRMap[StageNum - 1].count(LoopVal)) 1083 // The name is defined in the previous stage. 1084 PrevVal = VRMap[StageNum - 1][LoopVal]; 1085 else if (VRMap[StageNum].count(LoopVal)) 1086 // The previous name is defined in the current stage when the instruction 1087 // order is swapped. 1088 PrevVal = VRMap[StageNum][LoopVal]; 1089 else if (!LoopInst->isPHI() || LoopInst->getParent() != BB) 1090 // The loop value hasn't yet been scheduled. 1091 PrevVal = LoopVal; 1092 else if (StageNum == PhiStage + 1) 1093 // The loop value is another phi, which has not been scheduled. 1094 PrevVal = getInitPhiReg(*LoopInst, BB); 1095 else if (StageNum > PhiStage + 1 && LoopInst->getParent() == BB) 1096 // The loop value is another phi, which has been scheduled. 1097 PrevVal = 1098 getPrevMapVal(StageNum - 1, PhiStage, getLoopPhiReg(*LoopInst, BB), 1099 LoopStage, VRMap, BB); 1100 } 1101 return PrevVal; 1102 } 1103 1104 /// Rewrite the Phi values in the specified block to use the mappings 1105 /// from the initial operand. Once the Phi is scheduled, we switch 1106 /// to using the loop value instead of the Phi value, so those names 1107 /// do not need to be rewritten. 1108 void ModuloScheduleExpander::rewritePhiValues(MachineBasicBlock *NewBB, 1109 unsigned StageNum, 1110 ValueMapTy *VRMap, 1111 InstrMapTy &InstrMap) { 1112 for (auto &PHI : BB->phis()) { 1113 unsigned InitVal = 0; 1114 unsigned LoopVal = 0; 1115 getPhiRegs(PHI, BB, InitVal, LoopVal); 1116 Register PhiDef = PHI.getOperand(0).getReg(); 1117 1118 unsigned PhiStage = (unsigned)Schedule.getStage(MRI.getVRegDef(PhiDef)); 1119 unsigned LoopStage = (unsigned)Schedule.getStage(MRI.getVRegDef(LoopVal)); 1120 unsigned NumPhis = getStagesForPhi(PhiDef); 1121 if (NumPhis > StageNum) 1122 NumPhis = StageNum; 1123 for (unsigned np = 0; np <= NumPhis; ++np) { 1124 unsigned NewVal = 1125 getPrevMapVal(StageNum - np, PhiStage, LoopVal, LoopStage, VRMap, BB); 1126 if (!NewVal) 1127 NewVal = InitVal; 1128 rewriteScheduledInstr(NewBB, InstrMap, StageNum - np, np, &PHI, PhiDef, 1129 NewVal); 1130 } 1131 } 1132 } 1133 1134 /// Rewrite a previously scheduled instruction to use the register value 1135 /// from the new instruction. Make sure the instruction occurs in the 1136 /// basic block, and we don't change the uses in the new instruction. 1137 void ModuloScheduleExpander::rewriteScheduledInstr( 1138 MachineBasicBlock *BB, InstrMapTy &InstrMap, unsigned CurStageNum, 1139 unsigned PhiNum, MachineInstr *Phi, unsigned OldReg, unsigned NewReg, 1140 unsigned PrevReg) { 1141 bool InProlog = (CurStageNum < (unsigned)Schedule.getNumStages() - 1); 1142 int StagePhi = Schedule.getStage(Phi) + PhiNum; 1143 // Rewrite uses that have been scheduled already to use the new 1144 // Phi register. 1145 for (MachineOperand &UseOp : 1146 llvm::make_early_inc_range(MRI.use_operands(OldReg))) { 1147 MachineInstr *UseMI = UseOp.getParent(); 1148 if (UseMI->getParent() != BB) 1149 continue; 1150 if (UseMI->isPHI()) { 1151 if (!Phi->isPHI() && UseMI->getOperand(0).getReg() == NewReg) 1152 continue; 1153 if (getLoopPhiReg(*UseMI, BB) != OldReg) 1154 continue; 1155 } 1156 InstrMapTy::iterator OrigInstr = InstrMap.find(UseMI); 1157 assert(OrigInstr != InstrMap.end() && "Instruction not scheduled."); 1158 MachineInstr *OrigMI = OrigInstr->second; 1159 int StageSched = Schedule.getStage(OrigMI); 1160 int CycleSched = Schedule.getCycle(OrigMI); 1161 unsigned ReplaceReg = 0; 1162 // This is the stage for the scheduled instruction. 1163 if (StagePhi == StageSched && Phi->isPHI()) { 1164 int CyclePhi = Schedule.getCycle(Phi); 1165 if (PrevReg && InProlog) 1166 ReplaceReg = PrevReg; 1167 else if (PrevReg && !isLoopCarried(*Phi) && 1168 (CyclePhi <= CycleSched || OrigMI->isPHI())) 1169 ReplaceReg = PrevReg; 1170 else 1171 ReplaceReg = NewReg; 1172 } 1173 // The scheduled instruction occurs before the scheduled Phi, and the 1174 // Phi is not loop carried. 1175 if (!InProlog && StagePhi + 1 == StageSched && !isLoopCarried(*Phi)) 1176 ReplaceReg = NewReg; 1177 if (StagePhi > StageSched && Phi->isPHI()) 1178 ReplaceReg = NewReg; 1179 if (!InProlog && !Phi->isPHI() && StagePhi < StageSched) 1180 ReplaceReg = NewReg; 1181 if (ReplaceReg) { 1182 const TargetRegisterClass *NRC = 1183 MRI.constrainRegClass(ReplaceReg, MRI.getRegClass(OldReg)); 1184 if (NRC) 1185 UseOp.setReg(ReplaceReg); 1186 else { 1187 Register SplitReg = MRI.createVirtualRegister(MRI.getRegClass(OldReg)); 1188 BuildMI(*BB, UseMI, UseMI->getDebugLoc(), TII->get(TargetOpcode::COPY), 1189 SplitReg) 1190 .addReg(ReplaceReg); 1191 UseOp.setReg(SplitReg); 1192 } 1193 } 1194 } 1195 } 1196 1197 bool ModuloScheduleExpander::isLoopCarried(MachineInstr &Phi) { 1198 if (!Phi.isPHI()) 1199 return false; 1200 int DefCycle = Schedule.getCycle(&Phi); 1201 int DefStage = Schedule.getStage(&Phi); 1202 1203 unsigned InitVal = 0; 1204 unsigned LoopVal = 0; 1205 getPhiRegs(Phi, Phi.getParent(), InitVal, LoopVal); 1206 MachineInstr *Use = MRI.getVRegDef(LoopVal); 1207 if (!Use || Use->isPHI()) 1208 return true; 1209 int LoopCycle = Schedule.getCycle(Use); 1210 int LoopStage = Schedule.getStage(Use); 1211 return (LoopCycle > DefCycle) || (LoopStage <= DefStage); 1212 } 1213 1214 //===----------------------------------------------------------------------===// 1215 // PeelingModuloScheduleExpander implementation 1216 //===----------------------------------------------------------------------===// 1217 // This is a reimplementation of ModuloScheduleExpander that works by creating 1218 // a fully correct steady-state kernel and peeling off the prolog and epilogs. 1219 //===----------------------------------------------------------------------===// 1220 1221 namespace { 1222 // Remove any dead phis in MBB. Dead phis either have only one block as input 1223 // (in which case they are the identity) or have no uses. 1224 void EliminateDeadPhis(MachineBasicBlock *MBB, MachineRegisterInfo &MRI, 1225 LiveIntervals *LIS, bool KeepSingleSrcPhi = false) { 1226 bool Changed = true; 1227 while (Changed) { 1228 Changed = false; 1229 for (MachineInstr &MI : llvm::make_early_inc_range(MBB->phis())) { 1230 assert(MI.isPHI()); 1231 if (MRI.use_empty(MI.getOperand(0).getReg())) { 1232 if (LIS) 1233 LIS->RemoveMachineInstrFromMaps(MI); 1234 MI.eraseFromParent(); 1235 Changed = true; 1236 } else if (!KeepSingleSrcPhi && MI.getNumExplicitOperands() == 3) { 1237 const TargetRegisterClass *ConstrainRegClass = 1238 MRI.constrainRegClass(MI.getOperand(1).getReg(), 1239 MRI.getRegClass(MI.getOperand(0).getReg())); 1240 assert(ConstrainRegClass && 1241 "Expected a valid constrained register class!"); 1242 (void)ConstrainRegClass; 1243 MRI.replaceRegWith(MI.getOperand(0).getReg(), 1244 MI.getOperand(1).getReg()); 1245 if (LIS) 1246 LIS->RemoveMachineInstrFromMaps(MI); 1247 MI.eraseFromParent(); 1248 Changed = true; 1249 } 1250 } 1251 } 1252 } 1253 1254 /// Rewrites the kernel block in-place to adhere to the given schedule. 1255 /// KernelRewriter holds all of the state required to perform the rewriting. 1256 class KernelRewriter { 1257 ModuloSchedule &S; 1258 MachineBasicBlock *BB; 1259 MachineBasicBlock *PreheaderBB, *ExitBB; 1260 MachineRegisterInfo &MRI; 1261 const TargetInstrInfo *TII; 1262 LiveIntervals *LIS; 1263 1264 // Map from register class to canonical undef register for that class. 1265 DenseMap<const TargetRegisterClass *, Register> Undefs; 1266 // Map from <LoopReg, InitReg> to phi register for all created phis. Note that 1267 // this map is only used when InitReg is non-undef. 1268 DenseMap<std::pair<unsigned, unsigned>, Register> Phis; 1269 // Map from LoopReg to phi register where the InitReg is undef. 1270 DenseMap<Register, Register> UndefPhis; 1271 1272 // Reg is used by MI. Return the new register MI should use to adhere to the 1273 // schedule. Insert phis as necessary. 1274 Register remapUse(Register Reg, MachineInstr &MI); 1275 // Insert a phi that carries LoopReg from the loop body and InitReg otherwise. 1276 // If InitReg is not given it is chosen arbitrarily. It will either be undef 1277 // or will be chosen so as to share another phi. 1278 Register phi(Register LoopReg, std::optional<Register> InitReg = {}, 1279 const TargetRegisterClass *RC = nullptr); 1280 // Create an undef register of the given register class. 1281 Register undef(const TargetRegisterClass *RC); 1282 1283 public: 1284 KernelRewriter(MachineLoop &L, ModuloSchedule &S, MachineBasicBlock *LoopBB, 1285 LiveIntervals *LIS = nullptr); 1286 void rewrite(); 1287 }; 1288 } // namespace 1289 1290 KernelRewriter::KernelRewriter(MachineLoop &L, ModuloSchedule &S, 1291 MachineBasicBlock *LoopBB, LiveIntervals *LIS) 1292 : S(S), BB(LoopBB), PreheaderBB(L.getLoopPreheader()), 1293 ExitBB(L.getExitBlock()), MRI(BB->getParent()->getRegInfo()), 1294 TII(BB->getParent()->getSubtarget().getInstrInfo()), LIS(LIS) { 1295 PreheaderBB = *BB->pred_begin(); 1296 if (PreheaderBB == BB) 1297 PreheaderBB = *std::next(BB->pred_begin()); 1298 } 1299 1300 void KernelRewriter::rewrite() { 1301 // Rearrange the loop to be in schedule order. Note that the schedule may 1302 // contain instructions that are not owned by the loop block (InstrChanges and 1303 // friends), so we gracefully handle unowned instructions and delete any 1304 // instructions that weren't in the schedule. 1305 auto InsertPt = BB->getFirstTerminator(); 1306 MachineInstr *FirstMI = nullptr; 1307 for (MachineInstr *MI : S.getInstructions()) { 1308 if (MI->isPHI()) 1309 continue; 1310 if (MI->getParent()) 1311 MI->removeFromParent(); 1312 BB->insert(InsertPt, MI); 1313 if (!FirstMI) 1314 FirstMI = MI; 1315 } 1316 assert(FirstMI && "Failed to find first MI in schedule"); 1317 1318 // At this point all of the scheduled instructions are between FirstMI 1319 // and the end of the block. Kill from the first non-phi to FirstMI. 1320 for (auto I = BB->getFirstNonPHI(); I != FirstMI->getIterator();) { 1321 if (LIS) 1322 LIS->RemoveMachineInstrFromMaps(*I); 1323 (I++)->eraseFromParent(); 1324 } 1325 1326 // Now remap every instruction in the loop. 1327 for (MachineInstr &MI : *BB) { 1328 if (MI.isPHI() || MI.isTerminator()) 1329 continue; 1330 for (MachineOperand &MO : MI.uses()) { 1331 if (!MO.isReg() || MO.getReg().isPhysical() || MO.isImplicit()) 1332 continue; 1333 Register Reg = remapUse(MO.getReg(), MI); 1334 MO.setReg(Reg); 1335 } 1336 } 1337 EliminateDeadPhis(BB, MRI, LIS); 1338 1339 // Ensure a phi exists for all instructions that are either referenced by 1340 // an illegal phi or by an instruction outside the loop. This allows us to 1341 // treat remaps of these values the same as "normal" values that come from 1342 // loop-carried phis. 1343 for (auto MI = BB->getFirstNonPHI(); MI != BB->end(); ++MI) { 1344 if (MI->isPHI()) { 1345 Register R = MI->getOperand(0).getReg(); 1346 phi(R); 1347 continue; 1348 } 1349 1350 for (MachineOperand &Def : MI->defs()) { 1351 for (MachineInstr &MI : MRI.use_instructions(Def.getReg())) { 1352 if (MI.getParent() != BB) { 1353 phi(Def.getReg()); 1354 break; 1355 } 1356 } 1357 } 1358 } 1359 } 1360 1361 Register KernelRewriter::remapUse(Register Reg, MachineInstr &MI) { 1362 MachineInstr *Producer = MRI.getUniqueVRegDef(Reg); 1363 if (!Producer) 1364 return Reg; 1365 1366 int ConsumerStage = S.getStage(&MI); 1367 if (!Producer->isPHI()) { 1368 // Non-phi producers are simple to remap. Insert as many phis as the 1369 // difference between the consumer and producer stages. 1370 if (Producer->getParent() != BB) 1371 // Producer was not inside the loop. Use the register as-is. 1372 return Reg; 1373 int ProducerStage = S.getStage(Producer); 1374 assert(ConsumerStage != -1 && 1375 "In-loop consumer should always be scheduled!"); 1376 assert(ConsumerStage >= ProducerStage); 1377 unsigned StageDiff = ConsumerStage - ProducerStage; 1378 1379 for (unsigned I = 0; I < StageDiff; ++I) 1380 Reg = phi(Reg); 1381 return Reg; 1382 } 1383 1384 // First, dive through the phi chain to find the defaults for the generated 1385 // phis. 1386 SmallVector<std::optional<Register>, 4> Defaults; 1387 Register LoopReg = Reg; 1388 auto LoopProducer = Producer; 1389 while (LoopProducer->isPHI() && LoopProducer->getParent() == BB) { 1390 LoopReg = getLoopPhiReg(*LoopProducer, BB); 1391 Defaults.emplace_back(getInitPhiReg(*LoopProducer, BB)); 1392 LoopProducer = MRI.getUniqueVRegDef(LoopReg); 1393 assert(LoopProducer); 1394 } 1395 int LoopProducerStage = S.getStage(LoopProducer); 1396 1397 std::optional<Register> IllegalPhiDefault; 1398 1399 if (LoopProducerStage == -1) { 1400 // Do nothing. 1401 } else if (LoopProducerStage > ConsumerStage) { 1402 // This schedule is only representable if ProducerStage == ConsumerStage+1. 1403 // In addition, Consumer's cycle must be scheduled after Producer in the 1404 // rescheduled loop. This is enforced by the pipeliner's ASAP and ALAP 1405 // functions. 1406 #ifndef NDEBUG // Silence unused variables in non-asserts mode. 1407 int LoopProducerCycle = S.getCycle(LoopProducer); 1408 int ConsumerCycle = S.getCycle(&MI); 1409 #endif 1410 assert(LoopProducerCycle <= ConsumerCycle); 1411 assert(LoopProducerStage == ConsumerStage + 1); 1412 // Peel off the first phi from Defaults and insert a phi between producer 1413 // and consumer. This phi will not be at the front of the block so we 1414 // consider it illegal. It will only exist during the rewrite process; it 1415 // needs to exist while we peel off prologs because these could take the 1416 // default value. After that we can replace all uses with the loop producer 1417 // value. 1418 IllegalPhiDefault = Defaults.front(); 1419 Defaults.erase(Defaults.begin()); 1420 } else { 1421 assert(ConsumerStage >= LoopProducerStage); 1422 int StageDiff = ConsumerStage - LoopProducerStage; 1423 if (StageDiff > 0) { 1424 LLVM_DEBUG(dbgs() << " -- padding defaults array from " << Defaults.size() 1425 << " to " << (Defaults.size() + StageDiff) << "\n"); 1426 // If we need more phis than we have defaults for, pad out with undefs for 1427 // the earliest phis, which are at the end of the defaults chain (the 1428 // chain is in reverse order). 1429 Defaults.resize(Defaults.size() + StageDiff, 1430 Defaults.empty() ? std::optional<Register>() 1431 : Defaults.back()); 1432 } 1433 } 1434 1435 // Now we know the number of stages to jump back, insert the phi chain. 1436 auto DefaultI = Defaults.rbegin(); 1437 while (DefaultI != Defaults.rend()) 1438 LoopReg = phi(LoopReg, *DefaultI++, MRI.getRegClass(Reg)); 1439 1440 if (IllegalPhiDefault) { 1441 // The consumer optionally consumes LoopProducer in the same iteration 1442 // (because the producer is scheduled at an earlier cycle than the consumer) 1443 // or the initial value. To facilitate this we create an illegal block here 1444 // by embedding a phi in the middle of the block. We will fix this up 1445 // immediately prior to pruning. 1446 auto RC = MRI.getRegClass(Reg); 1447 Register R = MRI.createVirtualRegister(RC); 1448 MachineInstr *IllegalPhi = 1449 BuildMI(*BB, MI, DebugLoc(), TII->get(TargetOpcode::PHI), R) 1450 .addReg(*IllegalPhiDefault) 1451 .addMBB(PreheaderBB) // Block choice is arbitrary and has no effect. 1452 .addReg(LoopReg) 1453 .addMBB(BB); // Block choice is arbitrary and has no effect. 1454 // Illegal phi should belong to the producer stage so that it can be 1455 // filtered correctly during peeling. 1456 S.setStage(IllegalPhi, LoopProducerStage); 1457 return R; 1458 } 1459 1460 return LoopReg; 1461 } 1462 1463 Register KernelRewriter::phi(Register LoopReg, std::optional<Register> InitReg, 1464 const TargetRegisterClass *RC) { 1465 // If the init register is not undef, try and find an existing phi. 1466 if (InitReg) { 1467 auto I = Phis.find({LoopReg, *InitReg}); 1468 if (I != Phis.end()) 1469 return I->second; 1470 } else { 1471 for (auto &KV : Phis) { 1472 if (KV.first.first == LoopReg) 1473 return KV.second; 1474 } 1475 } 1476 1477 // InitReg is either undef or no existing phi takes InitReg as input. Try and 1478 // find a phi that takes undef as input. 1479 auto I = UndefPhis.find(LoopReg); 1480 if (I != UndefPhis.end()) { 1481 Register R = I->second; 1482 if (!InitReg) 1483 // Found a phi taking undef as input, and this input is undef so return 1484 // without any more changes. 1485 return R; 1486 // Found a phi taking undef as input, so rewrite it to take InitReg. 1487 MachineInstr *MI = MRI.getVRegDef(R); 1488 MI->getOperand(1).setReg(*InitReg); 1489 Phis.insert({{LoopReg, *InitReg}, R}); 1490 const TargetRegisterClass *ConstrainRegClass = 1491 MRI.constrainRegClass(R, MRI.getRegClass(*InitReg)); 1492 assert(ConstrainRegClass && "Expected a valid constrained register class!"); 1493 (void)ConstrainRegClass; 1494 UndefPhis.erase(I); 1495 return R; 1496 } 1497 1498 // Failed to find any existing phi to reuse, so create a new one. 1499 if (!RC) 1500 RC = MRI.getRegClass(LoopReg); 1501 Register R = MRI.createVirtualRegister(RC); 1502 if (InitReg) { 1503 const TargetRegisterClass *ConstrainRegClass = 1504 MRI.constrainRegClass(R, MRI.getRegClass(*InitReg)); 1505 assert(ConstrainRegClass && "Expected a valid constrained register class!"); 1506 (void)ConstrainRegClass; 1507 } 1508 BuildMI(*BB, BB->getFirstNonPHI(), DebugLoc(), TII->get(TargetOpcode::PHI), R) 1509 .addReg(InitReg ? *InitReg : undef(RC)) 1510 .addMBB(PreheaderBB) 1511 .addReg(LoopReg) 1512 .addMBB(BB); 1513 if (!InitReg) 1514 UndefPhis[LoopReg] = R; 1515 else 1516 Phis[{LoopReg, *InitReg}] = R; 1517 return R; 1518 } 1519 1520 Register KernelRewriter::undef(const TargetRegisterClass *RC) { 1521 Register &R = Undefs[RC]; 1522 if (R == 0) { 1523 // Create an IMPLICIT_DEF that defines this register if we need it. 1524 // All uses of this should be removed by the time we have finished unrolling 1525 // prologs and epilogs. 1526 R = MRI.createVirtualRegister(RC); 1527 auto *InsertBB = &PreheaderBB->getParent()->front(); 1528 BuildMI(*InsertBB, InsertBB->getFirstTerminator(), DebugLoc(), 1529 TII->get(TargetOpcode::IMPLICIT_DEF), R); 1530 } 1531 return R; 1532 } 1533 1534 namespace { 1535 /// Describes an operand in the kernel of a pipelined loop. Characteristics of 1536 /// the operand are discovered, such as how many in-loop PHIs it has to jump 1537 /// through and defaults for these phis. 1538 class KernelOperandInfo { 1539 MachineBasicBlock *BB; 1540 MachineRegisterInfo &MRI; 1541 SmallVector<Register, 4> PhiDefaults; 1542 MachineOperand *Source; 1543 MachineOperand *Target; 1544 1545 public: 1546 KernelOperandInfo(MachineOperand *MO, MachineRegisterInfo &MRI, 1547 const SmallPtrSetImpl<MachineInstr *> &IllegalPhis) 1548 : MRI(MRI) { 1549 Source = MO; 1550 BB = MO->getParent()->getParent(); 1551 while (isRegInLoop(MO)) { 1552 MachineInstr *MI = MRI.getVRegDef(MO->getReg()); 1553 if (MI->isFullCopy()) { 1554 MO = &MI->getOperand(1); 1555 continue; 1556 } 1557 if (!MI->isPHI()) 1558 break; 1559 // If this is an illegal phi, don't count it in distance. 1560 if (IllegalPhis.count(MI)) { 1561 MO = &MI->getOperand(3); 1562 continue; 1563 } 1564 1565 Register Default = getInitPhiReg(*MI, BB); 1566 MO = MI->getOperand(2).getMBB() == BB ? &MI->getOperand(1) 1567 : &MI->getOperand(3); 1568 PhiDefaults.push_back(Default); 1569 } 1570 Target = MO; 1571 } 1572 1573 bool operator==(const KernelOperandInfo &Other) const { 1574 return PhiDefaults.size() == Other.PhiDefaults.size(); 1575 } 1576 1577 void print(raw_ostream &OS) const { 1578 OS << "use of " << *Source << ": distance(" << PhiDefaults.size() << ") in " 1579 << *Source->getParent(); 1580 } 1581 1582 private: 1583 bool isRegInLoop(MachineOperand *MO) { 1584 return MO->isReg() && MO->getReg().isVirtual() && 1585 MRI.getVRegDef(MO->getReg())->getParent() == BB; 1586 } 1587 }; 1588 } // namespace 1589 1590 MachineBasicBlock * 1591 PeelingModuloScheduleExpander::peelKernel(LoopPeelDirection LPD) { 1592 MachineBasicBlock *NewBB = PeelSingleBlockLoop(LPD, BB, MRI, TII); 1593 if (LPD == LPD_Front) 1594 PeeledFront.push_back(NewBB); 1595 else 1596 PeeledBack.push_front(NewBB); 1597 for (auto I = BB->begin(), NI = NewBB->begin(); !I->isTerminator(); 1598 ++I, ++NI) { 1599 CanonicalMIs[&*I] = &*I; 1600 CanonicalMIs[&*NI] = &*I; 1601 BlockMIs[{NewBB, &*I}] = &*NI; 1602 BlockMIs[{BB, &*I}] = &*I; 1603 } 1604 return NewBB; 1605 } 1606 1607 void PeelingModuloScheduleExpander::filterInstructions(MachineBasicBlock *MB, 1608 int MinStage) { 1609 for (auto I = MB->getFirstInstrTerminator()->getReverseIterator(); 1610 I != std::next(MB->getFirstNonPHI()->getReverseIterator());) { 1611 MachineInstr *MI = &*I++; 1612 int Stage = getStage(MI); 1613 if (Stage == -1 || Stage >= MinStage) 1614 continue; 1615 1616 for (MachineOperand &DefMO : MI->defs()) { 1617 SmallVector<std::pair<MachineInstr *, Register>, 4> Subs; 1618 for (MachineInstr &UseMI : MRI.use_instructions(DefMO.getReg())) { 1619 // Only PHIs can use values from this block by construction. 1620 // Match with the equivalent PHI in B. 1621 assert(UseMI.isPHI()); 1622 Register Reg = getEquivalentRegisterIn(UseMI.getOperand(0).getReg(), 1623 MI->getParent()); 1624 Subs.emplace_back(&UseMI, Reg); 1625 } 1626 for (auto &Sub : Subs) 1627 Sub.first->substituteRegister(DefMO.getReg(), Sub.second, /*SubIdx=*/0, 1628 *MRI.getTargetRegisterInfo()); 1629 } 1630 if (LIS) 1631 LIS->RemoveMachineInstrFromMaps(*MI); 1632 MI->eraseFromParent(); 1633 } 1634 } 1635 1636 void PeelingModuloScheduleExpander::moveStageBetweenBlocks( 1637 MachineBasicBlock *DestBB, MachineBasicBlock *SourceBB, unsigned Stage) { 1638 auto InsertPt = DestBB->getFirstNonPHI(); 1639 DenseMap<Register, Register> Remaps; 1640 for (MachineInstr &MI : llvm::make_early_inc_range( 1641 llvm::make_range(SourceBB->getFirstNonPHI(), SourceBB->end()))) { 1642 if (MI.isPHI()) { 1643 // This is an illegal PHI. If we move any instructions using an illegal 1644 // PHI, we need to create a legal Phi. 1645 if (getStage(&MI) != Stage) { 1646 // The legal Phi is not necessary if the illegal phi's stage 1647 // is being moved. 1648 Register PhiR = MI.getOperand(0).getReg(); 1649 auto RC = MRI.getRegClass(PhiR); 1650 Register NR = MRI.createVirtualRegister(RC); 1651 MachineInstr *NI = BuildMI(*DestBB, DestBB->getFirstNonPHI(), 1652 DebugLoc(), TII->get(TargetOpcode::PHI), NR) 1653 .addReg(PhiR) 1654 .addMBB(SourceBB); 1655 BlockMIs[{DestBB, CanonicalMIs[&MI]}] = NI; 1656 CanonicalMIs[NI] = CanonicalMIs[&MI]; 1657 Remaps[PhiR] = NR; 1658 } 1659 } 1660 if (getStage(&MI) != Stage) 1661 continue; 1662 MI.removeFromParent(); 1663 DestBB->insert(InsertPt, &MI); 1664 auto *KernelMI = CanonicalMIs[&MI]; 1665 BlockMIs[{DestBB, KernelMI}] = &MI; 1666 BlockMIs.erase({SourceBB, KernelMI}); 1667 } 1668 SmallVector<MachineInstr *, 4> PhiToDelete; 1669 for (MachineInstr &MI : DestBB->phis()) { 1670 assert(MI.getNumOperands() == 3); 1671 MachineInstr *Def = MRI.getVRegDef(MI.getOperand(1).getReg()); 1672 // If the instruction referenced by the phi is moved inside the block 1673 // we don't need the phi anymore. 1674 if (getStage(Def) == Stage) { 1675 Register PhiReg = MI.getOperand(0).getReg(); 1676 assert(Def->findRegisterDefOperandIdx(MI.getOperand(1).getReg()) != -1); 1677 MRI.replaceRegWith(MI.getOperand(0).getReg(), MI.getOperand(1).getReg()); 1678 MI.getOperand(0).setReg(PhiReg); 1679 PhiToDelete.push_back(&MI); 1680 } 1681 } 1682 for (auto *P : PhiToDelete) 1683 P->eraseFromParent(); 1684 InsertPt = DestBB->getFirstNonPHI(); 1685 // Helper to clone Phi instructions into the destination block. We clone Phi 1686 // greedily to avoid combinatorial explosion of Phi instructions. 1687 auto clonePhi = [&](MachineInstr *Phi) { 1688 MachineInstr *NewMI = MF.CloneMachineInstr(Phi); 1689 DestBB->insert(InsertPt, NewMI); 1690 Register OrigR = Phi->getOperand(0).getReg(); 1691 Register R = MRI.createVirtualRegister(MRI.getRegClass(OrigR)); 1692 NewMI->getOperand(0).setReg(R); 1693 NewMI->getOperand(1).setReg(OrigR); 1694 NewMI->getOperand(2).setMBB(*DestBB->pred_begin()); 1695 Remaps[OrigR] = R; 1696 CanonicalMIs[NewMI] = CanonicalMIs[Phi]; 1697 BlockMIs[{DestBB, CanonicalMIs[Phi]}] = NewMI; 1698 PhiNodeLoopIteration[NewMI] = PhiNodeLoopIteration[Phi]; 1699 return R; 1700 }; 1701 for (auto I = DestBB->getFirstNonPHI(); I != DestBB->end(); ++I) { 1702 for (MachineOperand &MO : I->uses()) { 1703 if (!MO.isReg()) 1704 continue; 1705 if (Remaps.count(MO.getReg())) 1706 MO.setReg(Remaps[MO.getReg()]); 1707 else { 1708 // If we are using a phi from the source block we need to add a new phi 1709 // pointing to the old one. 1710 MachineInstr *Use = MRI.getUniqueVRegDef(MO.getReg()); 1711 if (Use && Use->isPHI() && Use->getParent() == SourceBB) { 1712 Register R = clonePhi(Use); 1713 MO.setReg(R); 1714 } 1715 } 1716 } 1717 } 1718 } 1719 1720 Register 1721 PeelingModuloScheduleExpander::getPhiCanonicalReg(MachineInstr *CanonicalPhi, 1722 MachineInstr *Phi) { 1723 unsigned distance = PhiNodeLoopIteration[Phi]; 1724 MachineInstr *CanonicalUse = CanonicalPhi; 1725 Register CanonicalUseReg = CanonicalUse->getOperand(0).getReg(); 1726 for (unsigned I = 0; I < distance; ++I) { 1727 assert(CanonicalUse->isPHI()); 1728 assert(CanonicalUse->getNumOperands() == 5); 1729 unsigned LoopRegIdx = 3, InitRegIdx = 1; 1730 if (CanonicalUse->getOperand(2).getMBB() == CanonicalUse->getParent()) 1731 std::swap(LoopRegIdx, InitRegIdx); 1732 CanonicalUseReg = CanonicalUse->getOperand(LoopRegIdx).getReg(); 1733 CanonicalUse = MRI.getVRegDef(CanonicalUseReg); 1734 } 1735 return CanonicalUseReg; 1736 } 1737 1738 void PeelingModuloScheduleExpander::peelPrologAndEpilogs() { 1739 BitVector LS(Schedule.getNumStages(), true); 1740 BitVector AS(Schedule.getNumStages(), true); 1741 LiveStages[BB] = LS; 1742 AvailableStages[BB] = AS; 1743 1744 // Peel out the prologs. 1745 LS.reset(); 1746 for (int I = 0; I < Schedule.getNumStages() - 1; ++I) { 1747 LS[I] = true; 1748 Prologs.push_back(peelKernel(LPD_Front)); 1749 LiveStages[Prologs.back()] = LS; 1750 AvailableStages[Prologs.back()] = LS; 1751 } 1752 1753 // Create a block that will end up as the new loop exiting block (dominated by 1754 // all prologs and epilogs). It will only contain PHIs, in the same order as 1755 // BB's PHIs. This gives us a poor-man's LCSSA with the inductive property 1756 // that the exiting block is a (sub) clone of BB. This in turn gives us the 1757 // property that any value deffed in BB but used outside of BB is used by a 1758 // PHI in the exiting block. 1759 MachineBasicBlock *ExitingBB = CreateLCSSAExitingBlock(); 1760 EliminateDeadPhis(ExitingBB, MRI, LIS, /*KeepSingleSrcPhi=*/true); 1761 // Push out the epilogs, again in reverse order. 1762 // We can't assume anything about the minumum loop trip count at this point, 1763 // so emit a fairly complex epilog. 1764 1765 // We first peel number of stages minus one epilogue. Then we remove dead 1766 // stages and reorder instructions based on their stage. If we have 3 stages 1767 // we generate first: 1768 // E0[3, 2, 1] 1769 // E1[3', 2'] 1770 // E2[3''] 1771 // And then we move instructions based on their stages to have: 1772 // E0[3] 1773 // E1[2, 3'] 1774 // E2[1, 2', 3''] 1775 // The transformation is legal because we only move instructions past 1776 // instructions of a previous loop iteration. 1777 for (int I = 1; I <= Schedule.getNumStages() - 1; ++I) { 1778 Epilogs.push_back(peelKernel(LPD_Back)); 1779 MachineBasicBlock *B = Epilogs.back(); 1780 filterInstructions(B, Schedule.getNumStages() - I); 1781 // Keep track at which iteration each phi belongs to. We need it to know 1782 // what version of the variable to use during prologue/epilogue stitching. 1783 EliminateDeadPhis(B, MRI, LIS, /*KeepSingleSrcPhi=*/true); 1784 for (MachineInstr &Phi : B->phis()) 1785 PhiNodeLoopIteration[&Phi] = Schedule.getNumStages() - I; 1786 } 1787 for (size_t I = 0; I < Epilogs.size(); I++) { 1788 LS.reset(); 1789 for (size_t J = I; J < Epilogs.size(); J++) { 1790 int Iteration = J; 1791 unsigned Stage = Schedule.getNumStages() - 1 + I - J; 1792 // Move stage one block at a time so that Phi nodes are updated correctly. 1793 for (size_t K = Iteration; K > I; K--) 1794 moveStageBetweenBlocks(Epilogs[K - 1], Epilogs[K], Stage); 1795 LS[Stage] = true; 1796 } 1797 LiveStages[Epilogs[I]] = LS; 1798 AvailableStages[Epilogs[I]] = AS; 1799 } 1800 1801 // Now we've defined all the prolog and epilog blocks as a fallthrough 1802 // sequence, add the edges that will be followed if the loop trip count is 1803 // lower than the number of stages (connecting prologs directly with epilogs). 1804 auto PI = Prologs.begin(); 1805 auto EI = Epilogs.begin(); 1806 assert(Prologs.size() == Epilogs.size()); 1807 for (; PI != Prologs.end(); ++PI, ++EI) { 1808 MachineBasicBlock *Pred = *(*EI)->pred_begin(); 1809 (*PI)->addSuccessor(*EI); 1810 for (MachineInstr &MI : (*EI)->phis()) { 1811 Register Reg = MI.getOperand(1).getReg(); 1812 MachineInstr *Use = MRI.getUniqueVRegDef(Reg); 1813 if (Use && Use->getParent() == Pred) { 1814 MachineInstr *CanonicalUse = CanonicalMIs[Use]; 1815 if (CanonicalUse->isPHI()) { 1816 // If the use comes from a phi we need to skip as many phi as the 1817 // distance between the epilogue and the kernel. Trace through the phi 1818 // chain to find the right value. 1819 Reg = getPhiCanonicalReg(CanonicalUse, Use); 1820 } 1821 Reg = getEquivalentRegisterIn(Reg, *PI); 1822 } 1823 MI.addOperand(MachineOperand::CreateReg(Reg, /*isDef=*/false)); 1824 MI.addOperand(MachineOperand::CreateMBB(*PI)); 1825 } 1826 } 1827 1828 // Create a list of all blocks in order. 1829 SmallVector<MachineBasicBlock *, 8> Blocks; 1830 llvm::copy(PeeledFront, std::back_inserter(Blocks)); 1831 Blocks.push_back(BB); 1832 llvm::copy(PeeledBack, std::back_inserter(Blocks)); 1833 1834 // Iterate in reverse order over all instructions, remapping as we go. 1835 for (MachineBasicBlock *B : reverse(Blocks)) { 1836 for (auto I = B->instr_rbegin(); 1837 I != std::next(B->getFirstNonPHI()->getReverseIterator());) { 1838 MachineBasicBlock::reverse_instr_iterator MI = I++; 1839 rewriteUsesOf(&*MI); 1840 } 1841 } 1842 for (auto *MI : IllegalPhisToDelete) { 1843 if (LIS) 1844 LIS->RemoveMachineInstrFromMaps(*MI); 1845 MI->eraseFromParent(); 1846 } 1847 IllegalPhisToDelete.clear(); 1848 1849 // Now all remapping has been done, we're free to optimize the generated code. 1850 for (MachineBasicBlock *B : reverse(Blocks)) 1851 EliminateDeadPhis(B, MRI, LIS); 1852 EliminateDeadPhis(ExitingBB, MRI, LIS); 1853 } 1854 1855 MachineBasicBlock *PeelingModuloScheduleExpander::CreateLCSSAExitingBlock() { 1856 MachineFunction &MF = *BB->getParent(); 1857 MachineBasicBlock *Exit = *BB->succ_begin(); 1858 if (Exit == BB) 1859 Exit = *std::next(BB->succ_begin()); 1860 1861 MachineBasicBlock *NewBB = MF.CreateMachineBasicBlock(BB->getBasicBlock()); 1862 MF.insert(std::next(BB->getIterator()), NewBB); 1863 1864 // Clone all phis in BB into NewBB and rewrite. 1865 for (MachineInstr &MI : BB->phis()) { 1866 auto RC = MRI.getRegClass(MI.getOperand(0).getReg()); 1867 Register OldR = MI.getOperand(3).getReg(); 1868 Register R = MRI.createVirtualRegister(RC); 1869 SmallVector<MachineInstr *, 4> Uses; 1870 for (MachineInstr &Use : MRI.use_instructions(OldR)) 1871 if (Use.getParent() != BB) 1872 Uses.push_back(&Use); 1873 for (MachineInstr *Use : Uses) 1874 Use->substituteRegister(OldR, R, /*SubIdx=*/0, 1875 *MRI.getTargetRegisterInfo()); 1876 MachineInstr *NI = BuildMI(NewBB, DebugLoc(), TII->get(TargetOpcode::PHI), R) 1877 .addReg(OldR) 1878 .addMBB(BB); 1879 BlockMIs[{NewBB, &MI}] = NI; 1880 CanonicalMIs[NI] = &MI; 1881 } 1882 BB->replaceSuccessor(Exit, NewBB); 1883 Exit->replacePhiUsesWith(BB, NewBB); 1884 NewBB->addSuccessor(Exit); 1885 1886 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; 1887 SmallVector<MachineOperand, 4> Cond; 1888 bool CanAnalyzeBr = !TII->analyzeBranch(*BB, TBB, FBB, Cond); 1889 (void)CanAnalyzeBr; 1890 assert(CanAnalyzeBr && "Must be able to analyze the loop branch!"); 1891 TII->removeBranch(*BB); 1892 TII->insertBranch(*BB, TBB == Exit ? NewBB : TBB, FBB == Exit ? NewBB : FBB, 1893 Cond, DebugLoc()); 1894 TII->insertUnconditionalBranch(*NewBB, Exit, DebugLoc()); 1895 return NewBB; 1896 } 1897 1898 Register 1899 PeelingModuloScheduleExpander::getEquivalentRegisterIn(Register Reg, 1900 MachineBasicBlock *BB) { 1901 MachineInstr *MI = MRI.getUniqueVRegDef(Reg); 1902 unsigned OpIdx = MI->findRegisterDefOperandIdx(Reg); 1903 return BlockMIs[{BB, CanonicalMIs[MI]}]->getOperand(OpIdx).getReg(); 1904 } 1905 1906 void PeelingModuloScheduleExpander::rewriteUsesOf(MachineInstr *MI) { 1907 if (MI->isPHI()) { 1908 // This is an illegal PHI. The loop-carried (desired) value is operand 3, 1909 // and it is produced by this block. 1910 Register PhiR = MI->getOperand(0).getReg(); 1911 Register R = MI->getOperand(3).getReg(); 1912 int RMIStage = getStage(MRI.getUniqueVRegDef(R)); 1913 if (RMIStage != -1 && !AvailableStages[MI->getParent()].test(RMIStage)) 1914 R = MI->getOperand(1).getReg(); 1915 MRI.setRegClass(R, MRI.getRegClass(PhiR)); 1916 MRI.replaceRegWith(PhiR, R); 1917 // Postpone deleting the Phi as it may be referenced by BlockMIs and used 1918 // later to figure out how to remap registers. 1919 MI->getOperand(0).setReg(PhiR); 1920 IllegalPhisToDelete.push_back(MI); 1921 return; 1922 } 1923 1924 int Stage = getStage(MI); 1925 if (Stage == -1 || LiveStages.count(MI->getParent()) == 0 || 1926 LiveStages[MI->getParent()].test(Stage)) 1927 // Instruction is live, no rewriting to do. 1928 return; 1929 1930 for (MachineOperand &DefMO : MI->defs()) { 1931 SmallVector<std::pair<MachineInstr *, Register>, 4> Subs; 1932 for (MachineInstr &UseMI : MRI.use_instructions(DefMO.getReg())) { 1933 // Only PHIs can use values from this block by construction. 1934 // Match with the equivalent PHI in B. 1935 assert(UseMI.isPHI()); 1936 Register Reg = getEquivalentRegisterIn(UseMI.getOperand(0).getReg(), 1937 MI->getParent()); 1938 Subs.emplace_back(&UseMI, Reg); 1939 } 1940 for (auto &Sub : Subs) 1941 Sub.first->substituteRegister(DefMO.getReg(), Sub.second, /*SubIdx=*/0, 1942 *MRI.getTargetRegisterInfo()); 1943 } 1944 if (LIS) 1945 LIS->RemoveMachineInstrFromMaps(*MI); 1946 MI->eraseFromParent(); 1947 } 1948 1949 void PeelingModuloScheduleExpander::fixupBranches() { 1950 // Work outwards from the kernel. 1951 bool KernelDisposed = false; 1952 int TC = Schedule.getNumStages() - 1; 1953 for (auto PI = Prologs.rbegin(), EI = Epilogs.rbegin(); PI != Prologs.rend(); 1954 ++PI, ++EI, --TC) { 1955 MachineBasicBlock *Prolog = *PI; 1956 MachineBasicBlock *Fallthrough = *Prolog->succ_begin(); 1957 MachineBasicBlock *Epilog = *EI; 1958 SmallVector<MachineOperand, 4> Cond; 1959 TII->removeBranch(*Prolog); 1960 std::optional<bool> StaticallyGreater = 1961 LoopInfo->createTripCountGreaterCondition(TC, *Prolog, Cond); 1962 if (!StaticallyGreater) { 1963 LLVM_DEBUG(dbgs() << "Dynamic: TC > " << TC << "\n"); 1964 // Dynamically branch based on Cond. 1965 TII->insertBranch(*Prolog, Epilog, Fallthrough, Cond, DebugLoc()); 1966 } else if (*StaticallyGreater == false) { 1967 LLVM_DEBUG(dbgs() << "Static-false: TC > " << TC << "\n"); 1968 // Prolog never falls through; branch to epilog and orphan interior 1969 // blocks. Leave it to unreachable-block-elim to clean up. 1970 Prolog->removeSuccessor(Fallthrough); 1971 for (MachineInstr &P : Fallthrough->phis()) { 1972 P.removeOperand(2); 1973 P.removeOperand(1); 1974 } 1975 TII->insertUnconditionalBranch(*Prolog, Epilog, DebugLoc()); 1976 KernelDisposed = true; 1977 } else { 1978 LLVM_DEBUG(dbgs() << "Static-true: TC > " << TC << "\n"); 1979 // Prolog always falls through; remove incoming values in epilog. 1980 Prolog->removeSuccessor(Epilog); 1981 for (MachineInstr &P : Epilog->phis()) { 1982 P.removeOperand(4); 1983 P.removeOperand(3); 1984 } 1985 } 1986 } 1987 1988 if (!KernelDisposed) { 1989 LoopInfo->adjustTripCount(-(Schedule.getNumStages() - 1)); 1990 LoopInfo->setPreheader(Prologs.back()); 1991 } else { 1992 LoopInfo->disposed(); 1993 } 1994 } 1995 1996 void PeelingModuloScheduleExpander::rewriteKernel() { 1997 KernelRewriter KR(*Schedule.getLoop(), Schedule, BB); 1998 KR.rewrite(); 1999 } 2000 2001 void PeelingModuloScheduleExpander::expand() { 2002 BB = Schedule.getLoop()->getTopBlock(); 2003 Preheader = Schedule.getLoop()->getLoopPreheader(); 2004 LLVM_DEBUG(Schedule.dump()); 2005 LoopInfo = TII->analyzeLoopForPipelining(BB); 2006 assert(LoopInfo); 2007 2008 rewriteKernel(); 2009 peelPrologAndEpilogs(); 2010 fixupBranches(); 2011 } 2012 2013 void PeelingModuloScheduleExpander::validateAgainstModuloScheduleExpander() { 2014 BB = Schedule.getLoop()->getTopBlock(); 2015 Preheader = Schedule.getLoop()->getLoopPreheader(); 2016 2017 // Dump the schedule before we invalidate and remap all its instructions. 2018 // Stash it in a string so we can print it if we found an error. 2019 std::string ScheduleDump; 2020 raw_string_ostream OS(ScheduleDump); 2021 Schedule.print(OS); 2022 OS.flush(); 2023 2024 // First, run the normal ModuleScheduleExpander. We don't support any 2025 // InstrChanges. 2026 assert(LIS && "Requires LiveIntervals!"); 2027 ModuloScheduleExpander MSE(MF, Schedule, *LIS, 2028 ModuloScheduleExpander::InstrChangesTy()); 2029 MSE.expand(); 2030 MachineBasicBlock *ExpandedKernel = MSE.getRewrittenKernel(); 2031 if (!ExpandedKernel) { 2032 // The expander optimized away the kernel. We can't do any useful checking. 2033 MSE.cleanup(); 2034 return; 2035 } 2036 // Before running the KernelRewriter, re-add BB into the CFG. 2037 Preheader->addSuccessor(BB); 2038 2039 // Now run the new expansion algorithm. 2040 KernelRewriter KR(*Schedule.getLoop(), Schedule, BB); 2041 KR.rewrite(); 2042 peelPrologAndEpilogs(); 2043 2044 // Collect all illegal phis that the new algorithm created. We'll give these 2045 // to KernelOperandInfo. 2046 SmallPtrSet<MachineInstr *, 4> IllegalPhis; 2047 for (auto NI = BB->getFirstNonPHI(); NI != BB->end(); ++NI) { 2048 if (NI->isPHI()) 2049 IllegalPhis.insert(&*NI); 2050 } 2051 2052 // Co-iterate across both kernels. We expect them to be identical apart from 2053 // phis and full COPYs (we look through both). 2054 SmallVector<std::pair<KernelOperandInfo, KernelOperandInfo>, 8> KOIs; 2055 auto OI = ExpandedKernel->begin(); 2056 auto NI = BB->begin(); 2057 for (; !OI->isTerminator() && !NI->isTerminator(); ++OI, ++NI) { 2058 while (OI->isPHI() || OI->isFullCopy()) 2059 ++OI; 2060 while (NI->isPHI() || NI->isFullCopy()) 2061 ++NI; 2062 assert(OI->getOpcode() == NI->getOpcode() && "Opcodes don't match?!"); 2063 // Analyze every operand separately. 2064 for (auto OOpI = OI->operands_begin(), NOpI = NI->operands_begin(); 2065 OOpI != OI->operands_end(); ++OOpI, ++NOpI) 2066 KOIs.emplace_back(KernelOperandInfo(&*OOpI, MRI, IllegalPhis), 2067 KernelOperandInfo(&*NOpI, MRI, IllegalPhis)); 2068 } 2069 2070 bool Failed = false; 2071 for (auto &OldAndNew : KOIs) { 2072 if (OldAndNew.first == OldAndNew.second) 2073 continue; 2074 Failed = true; 2075 errs() << "Modulo kernel validation error: [\n"; 2076 errs() << " [golden] "; 2077 OldAndNew.first.print(errs()); 2078 errs() << " "; 2079 OldAndNew.second.print(errs()); 2080 errs() << "]\n"; 2081 } 2082 2083 if (Failed) { 2084 errs() << "Golden reference kernel:\n"; 2085 ExpandedKernel->print(errs()); 2086 errs() << "New kernel:\n"; 2087 BB->print(errs()); 2088 errs() << ScheduleDump; 2089 report_fatal_error( 2090 "Modulo kernel validation (-pipeliner-experimental-cg) failed"); 2091 } 2092 2093 // Cleanup by removing BB from the CFG again as the original 2094 // ModuloScheduleExpander intended. 2095 Preheader->removeSuccessor(BB); 2096 MSE.cleanup(); 2097 } 2098 2099 //===----------------------------------------------------------------------===// 2100 // ModuloScheduleTestPass implementation 2101 //===----------------------------------------------------------------------===// 2102 // This pass constructs a ModuloSchedule from its module and runs 2103 // ModuloScheduleExpander. 2104 // 2105 // The module is expected to contain a single-block analyzable loop. 2106 // The total order of instructions is taken from the loop as-is. 2107 // Instructions are expected to be annotated with a PostInstrSymbol. 2108 // This PostInstrSymbol must have the following format: 2109 // "Stage=%d Cycle=%d". 2110 //===----------------------------------------------------------------------===// 2111 2112 namespace { 2113 class ModuloScheduleTest : public MachineFunctionPass { 2114 public: 2115 static char ID; 2116 2117 ModuloScheduleTest() : MachineFunctionPass(ID) { 2118 initializeModuloScheduleTestPass(*PassRegistry::getPassRegistry()); 2119 } 2120 2121 bool runOnMachineFunction(MachineFunction &MF) override; 2122 void runOnLoop(MachineFunction &MF, MachineLoop &L); 2123 2124 void getAnalysisUsage(AnalysisUsage &AU) const override { 2125 AU.addRequired<MachineLoopInfo>(); 2126 AU.addRequired<LiveIntervals>(); 2127 MachineFunctionPass::getAnalysisUsage(AU); 2128 } 2129 }; 2130 } // namespace 2131 2132 char ModuloScheduleTest::ID = 0; 2133 2134 INITIALIZE_PASS_BEGIN(ModuloScheduleTest, "modulo-schedule-test", 2135 "Modulo Schedule test pass", false, false) 2136 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo) 2137 INITIALIZE_PASS_DEPENDENCY(LiveIntervals) 2138 INITIALIZE_PASS_END(ModuloScheduleTest, "modulo-schedule-test", 2139 "Modulo Schedule test pass", false, false) 2140 2141 bool ModuloScheduleTest::runOnMachineFunction(MachineFunction &MF) { 2142 MachineLoopInfo &MLI = getAnalysis<MachineLoopInfo>(); 2143 for (auto *L : MLI) { 2144 if (L->getTopBlock() != L->getBottomBlock()) 2145 continue; 2146 runOnLoop(MF, *L); 2147 return false; 2148 } 2149 return false; 2150 } 2151 2152 static void parseSymbolString(StringRef S, int &Cycle, int &Stage) { 2153 std::pair<StringRef, StringRef> StageAndCycle = getToken(S, "_"); 2154 std::pair<StringRef, StringRef> StageTokenAndValue = 2155 getToken(StageAndCycle.first, "-"); 2156 std::pair<StringRef, StringRef> CycleTokenAndValue = 2157 getToken(StageAndCycle.second, "-"); 2158 if (StageTokenAndValue.first != "Stage" || 2159 CycleTokenAndValue.first != "_Cycle") { 2160 llvm_unreachable( 2161 "Bad post-instr symbol syntax: see comment in ModuloScheduleTest"); 2162 return; 2163 } 2164 2165 StageTokenAndValue.second.drop_front().getAsInteger(10, Stage); 2166 CycleTokenAndValue.second.drop_front().getAsInteger(10, Cycle); 2167 2168 dbgs() << " Stage=" << Stage << ", Cycle=" << Cycle << "\n"; 2169 } 2170 2171 void ModuloScheduleTest::runOnLoop(MachineFunction &MF, MachineLoop &L) { 2172 LiveIntervals &LIS = getAnalysis<LiveIntervals>(); 2173 MachineBasicBlock *BB = L.getTopBlock(); 2174 dbgs() << "--- ModuloScheduleTest running on BB#" << BB->getNumber() << "\n"; 2175 2176 DenseMap<MachineInstr *, int> Cycle, Stage; 2177 std::vector<MachineInstr *> Instrs; 2178 for (MachineInstr &MI : *BB) { 2179 if (MI.isTerminator()) 2180 continue; 2181 Instrs.push_back(&MI); 2182 if (MCSymbol *Sym = MI.getPostInstrSymbol()) { 2183 dbgs() << "Parsing post-instr symbol for " << MI; 2184 parseSymbolString(Sym->getName(), Cycle[&MI], Stage[&MI]); 2185 } 2186 } 2187 2188 ModuloSchedule MS(MF, &L, std::move(Instrs), std::move(Cycle), 2189 std::move(Stage)); 2190 ModuloScheduleExpander MSE( 2191 MF, MS, LIS, /*InstrChanges=*/ModuloScheduleExpander::InstrChangesTy()); 2192 MSE.expand(); 2193 MSE.cleanup(); 2194 } 2195 2196 //===----------------------------------------------------------------------===// 2197 // ModuloScheduleTestAnnotater implementation 2198 //===----------------------------------------------------------------------===// 2199 2200 void ModuloScheduleTestAnnotater::annotate() { 2201 for (MachineInstr *MI : S.getInstructions()) { 2202 SmallVector<char, 16> SV; 2203 raw_svector_ostream OS(SV); 2204 OS << "Stage-" << S.getStage(MI) << "_Cycle-" << S.getCycle(MI); 2205 MCSymbol *Sym = MF.getContext().getOrCreateSymbol(OS.str()); 2206 MI->setPostInstrSymbol(MF, Sym); 2207 } 2208 } 2209