MachineSink.cpp revision 218893
1//===-- MachineSink.cpp - Sinking for machine instructions ----------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This pass moves instructions into successor blocks when possible, so that 11// they aren't executed on paths where their results aren't needed. 12// 13// This pass is not intended to be a replacement or a complete alternative 14// for an LLVM-IR-level sinking pass. It is only designed to sink simple 15// constructs that are not exposed before lowering and instruction selection. 16// 17//===----------------------------------------------------------------------===// 18 19#define DEBUG_TYPE "machine-sink" 20#include "llvm/CodeGen/Passes.h" 21#include "llvm/CodeGen/MachineRegisterInfo.h" 22#include "llvm/CodeGen/MachineDominators.h" 23#include "llvm/CodeGen/MachineLoopInfo.h" 24#include "llvm/Analysis/AliasAnalysis.h" 25#include "llvm/Target/TargetRegisterInfo.h" 26#include "llvm/Target/TargetInstrInfo.h" 27#include "llvm/Target/TargetMachine.h" 28#include "llvm/ADT/SmallSet.h" 29#include "llvm/ADT/Statistic.h" 30#include "llvm/Support/CommandLine.h" 31#include "llvm/Support/Debug.h" 32#include "llvm/Support/raw_ostream.h" 33using namespace llvm; 34 35static cl::opt<bool> 36SplitEdges("machine-sink-split", 37 cl::desc("Split critical edges during machine sinking"), 38 cl::init(true), cl::Hidden); 39 40STATISTIC(NumSunk, "Number of machine instructions sunk"); 41STATISTIC(NumSplit, "Number of critical edges split"); 42STATISTIC(NumCoalesces, "Number of copies coalesced"); 43 44namespace { 45 class MachineSinking : public MachineFunctionPass { 46 const TargetInstrInfo *TII; 47 const TargetRegisterInfo *TRI; 48 MachineRegisterInfo *MRI; // Machine register information 49 MachineDominatorTree *DT; // Machine dominator tree 50 MachineLoopInfo *LI; 51 AliasAnalysis *AA; 52 BitVector AllocatableSet; // Which physregs are allocatable? 53 54 // Remember which edges have been considered for breaking. 55 SmallSet<std::pair<MachineBasicBlock*,MachineBasicBlock*>, 8> 56 CEBCandidates; 57 58 public: 59 static char ID; // Pass identification 60 MachineSinking() : MachineFunctionPass(ID) { 61 initializeMachineSinkingPass(*PassRegistry::getPassRegistry()); 62 } 63 64 virtual bool runOnMachineFunction(MachineFunction &MF); 65 66 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 67 AU.setPreservesCFG(); 68 MachineFunctionPass::getAnalysisUsage(AU); 69 AU.addRequired<AliasAnalysis>(); 70 AU.addRequired<MachineDominatorTree>(); 71 AU.addRequired<MachineLoopInfo>(); 72 AU.addPreserved<MachineDominatorTree>(); 73 AU.addPreserved<MachineLoopInfo>(); 74 } 75 76 virtual void releaseMemory() { 77 CEBCandidates.clear(); 78 } 79 80 private: 81 bool ProcessBlock(MachineBasicBlock &MBB); 82 bool isWorthBreakingCriticalEdge(MachineInstr *MI, 83 MachineBasicBlock *From, 84 MachineBasicBlock *To); 85 MachineBasicBlock *SplitCriticalEdge(MachineInstr *MI, 86 MachineBasicBlock *From, 87 MachineBasicBlock *To, 88 bool BreakPHIEdge); 89 bool SinkInstruction(MachineInstr *MI, bool &SawStore); 90 bool AllUsesDominatedByBlock(unsigned Reg, MachineBasicBlock *MBB, 91 MachineBasicBlock *DefMBB, 92 bool &BreakPHIEdge, bool &LocalUse) const; 93 bool PerformTrivialForwardCoalescing(MachineInstr *MI, 94 MachineBasicBlock *MBB); 95 }; 96} // end anonymous namespace 97 98char MachineSinking::ID = 0; 99INITIALIZE_PASS_BEGIN(MachineSinking, "machine-sink", 100 "Machine code sinking", false, false) 101INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree) 102INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo) 103INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 104INITIALIZE_PASS_END(MachineSinking, "machine-sink", 105 "Machine code sinking", false, false) 106 107FunctionPass *llvm::createMachineSinkingPass() { return new MachineSinking(); } 108 109bool MachineSinking::PerformTrivialForwardCoalescing(MachineInstr *MI, 110 MachineBasicBlock *MBB) { 111 if (!MI->isCopy()) 112 return false; 113 114 unsigned SrcReg = MI->getOperand(1).getReg(); 115 unsigned DstReg = MI->getOperand(0).getReg(); 116 if (!TargetRegisterInfo::isVirtualRegister(SrcReg) || 117 !TargetRegisterInfo::isVirtualRegister(DstReg) || 118 !MRI->hasOneNonDBGUse(SrcReg)) 119 return false; 120 121 const TargetRegisterClass *SRC = MRI->getRegClass(SrcReg); 122 const TargetRegisterClass *DRC = MRI->getRegClass(DstReg); 123 if (SRC != DRC) 124 return false; 125 126 MachineInstr *DefMI = MRI->getVRegDef(SrcReg); 127 if (DefMI->isCopyLike()) 128 return false; 129 DEBUG(dbgs() << "Coalescing: " << *DefMI); 130 DEBUG(dbgs() << "*** to: " << *MI); 131 MRI->replaceRegWith(DstReg, SrcReg); 132 MI->eraseFromParent(); 133 ++NumCoalesces; 134 return true; 135} 136 137/// AllUsesDominatedByBlock - Return true if all uses of the specified register 138/// occur in blocks dominated by the specified block. If any use is in the 139/// definition block, then return false since it is never legal to move def 140/// after uses. 141bool 142MachineSinking::AllUsesDominatedByBlock(unsigned Reg, 143 MachineBasicBlock *MBB, 144 MachineBasicBlock *DefMBB, 145 bool &BreakPHIEdge, 146 bool &LocalUse) const { 147 assert(TargetRegisterInfo::isVirtualRegister(Reg) && 148 "Only makes sense for vregs"); 149 150 if (MRI->use_nodbg_empty(Reg)) 151 return true; 152 153 // Ignoring debug uses is necessary so debug info doesn't affect the code. 154 // This may leave a referencing dbg_value in the original block, before 155 // the definition of the vreg. Dwarf generator handles this although the 156 // user might not get the right info at runtime. 157 158 // BreakPHIEdge is true if all the uses are in the successor MBB being sunken 159 // into and they are all PHI nodes. In this case, machine-sink must break 160 // the critical edge first. e.g. 161 // 162 // BB#1: derived from LLVM BB %bb4.preheader 163 // Predecessors according to CFG: BB#0 164 // ... 165 // %reg16385<def> = DEC64_32r %reg16437, %EFLAGS<imp-def,dead> 166 // ... 167 // JE_4 <BB#37>, %EFLAGS<imp-use> 168 // Successors according to CFG: BB#37 BB#2 169 // 170 // BB#2: derived from LLVM BB %bb.nph 171 // Predecessors according to CFG: BB#0 BB#1 172 // %reg16386<def> = PHI %reg16434, <BB#0>, %reg16385, <BB#1> 173 BreakPHIEdge = true; 174 for (MachineRegisterInfo::use_nodbg_iterator 175 I = MRI->use_nodbg_begin(Reg), E = MRI->use_nodbg_end(); 176 I != E; ++I) { 177 MachineInstr *UseInst = &*I; 178 MachineBasicBlock *UseBlock = UseInst->getParent(); 179 if (!(UseBlock == MBB && UseInst->isPHI() && 180 UseInst->getOperand(I.getOperandNo()+1).getMBB() == DefMBB)) { 181 BreakPHIEdge = false; 182 break; 183 } 184 } 185 if (BreakPHIEdge) 186 return true; 187 188 for (MachineRegisterInfo::use_nodbg_iterator 189 I = MRI->use_nodbg_begin(Reg), E = MRI->use_nodbg_end(); 190 I != E; ++I) { 191 // Determine the block of the use. 192 MachineInstr *UseInst = &*I; 193 MachineBasicBlock *UseBlock = UseInst->getParent(); 194 if (UseInst->isPHI()) { 195 // PHI nodes use the operand in the predecessor block, not the block with 196 // the PHI. 197 UseBlock = UseInst->getOperand(I.getOperandNo()+1).getMBB(); 198 } else if (UseBlock == DefMBB) { 199 LocalUse = true; 200 return false; 201 } 202 203 // Check that it dominates. 204 if (!DT->dominates(MBB, UseBlock)) 205 return false; 206 } 207 208 return true; 209} 210 211bool MachineSinking::runOnMachineFunction(MachineFunction &MF) { 212 DEBUG(dbgs() << "******** Machine Sinking ********\n"); 213 214 const TargetMachine &TM = MF.getTarget(); 215 TII = TM.getInstrInfo(); 216 TRI = TM.getRegisterInfo(); 217 MRI = &MF.getRegInfo(); 218 DT = &getAnalysis<MachineDominatorTree>(); 219 LI = &getAnalysis<MachineLoopInfo>(); 220 AA = &getAnalysis<AliasAnalysis>(); 221 AllocatableSet = TRI->getAllocatableSet(MF); 222 223 bool EverMadeChange = false; 224 225 while (1) { 226 bool MadeChange = false; 227 228 // Process all basic blocks. 229 CEBCandidates.clear(); 230 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); 231 I != E; ++I) 232 MadeChange |= ProcessBlock(*I); 233 234 // If this iteration over the code changed anything, keep iterating. 235 if (!MadeChange) break; 236 EverMadeChange = true; 237 } 238 return EverMadeChange; 239} 240 241bool MachineSinking::ProcessBlock(MachineBasicBlock &MBB) { 242 // Can't sink anything out of a block that has less than two successors. 243 if (MBB.succ_size() <= 1 || MBB.empty()) return false; 244 245 // Don't bother sinking code out of unreachable blocks. In addition to being 246 // unprofitable, it can also lead to infinite looping, because in an 247 // unreachable loop there may be nowhere to stop. 248 if (!DT->isReachableFromEntry(&MBB)) return false; 249 250 bool MadeChange = false; 251 252 // Walk the basic block bottom-up. Remember if we saw a store. 253 MachineBasicBlock::iterator I = MBB.end(); 254 --I; 255 bool ProcessedBegin, SawStore = false; 256 do { 257 MachineInstr *MI = I; // The instruction to sink. 258 259 // Predecrement I (if it's not begin) so that it isn't invalidated by 260 // sinking. 261 ProcessedBegin = I == MBB.begin(); 262 if (!ProcessedBegin) 263 --I; 264 265 if (MI->isDebugValue()) 266 continue; 267 268 if (PerformTrivialForwardCoalescing(MI, &MBB)) 269 continue; 270 271 if (SinkInstruction(MI, SawStore)) 272 ++NumSunk, MadeChange = true; 273 274 // If we just processed the first instruction in the block, we're done. 275 } while (!ProcessedBegin); 276 277 return MadeChange; 278} 279 280bool MachineSinking::isWorthBreakingCriticalEdge(MachineInstr *MI, 281 MachineBasicBlock *From, 282 MachineBasicBlock *To) { 283 // FIXME: Need much better heuristics. 284 285 // If the pass has already considered breaking this edge (during this pass 286 // through the function), then let's go ahead and break it. This means 287 // sinking multiple "cheap" instructions into the same block. 288 if (!CEBCandidates.insert(std::make_pair(From, To))) 289 return true; 290 291 if (!MI->isCopy() && !MI->getDesc().isAsCheapAsAMove()) 292 return true; 293 294 // MI is cheap, we probably don't want to break the critical edge for it. 295 // However, if this would allow some definitions of its source operands 296 // to be sunk then it's probably worth it. 297 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 298 const MachineOperand &MO = MI->getOperand(i); 299 if (!MO.isReg()) continue; 300 unsigned Reg = MO.getReg(); 301 if (Reg == 0 || !TargetRegisterInfo::isPhysicalRegister(Reg)) 302 continue; 303 if (MRI->hasOneNonDBGUse(Reg)) 304 return true; 305 } 306 307 return false; 308} 309 310MachineBasicBlock *MachineSinking::SplitCriticalEdge(MachineInstr *MI, 311 MachineBasicBlock *FromBB, 312 MachineBasicBlock *ToBB, 313 bool BreakPHIEdge) { 314 if (!isWorthBreakingCriticalEdge(MI, FromBB, ToBB)) 315 return 0; 316 317 // Avoid breaking back edge. From == To means backedge for single BB loop. 318 if (!SplitEdges || FromBB == ToBB) 319 return 0; 320 321 // Check for backedges of more "complex" loops. 322 if (LI->getLoopFor(FromBB) == LI->getLoopFor(ToBB) && 323 LI->isLoopHeader(ToBB)) 324 return 0; 325 326 // It's not always legal to break critical edges and sink the computation 327 // to the edge. 328 // 329 // BB#1: 330 // v1024 331 // Beq BB#3 332 // <fallthrough> 333 // BB#2: 334 // ... no uses of v1024 335 // <fallthrough> 336 // BB#3: 337 // ... 338 // = v1024 339 // 340 // If BB#1 -> BB#3 edge is broken and computation of v1024 is inserted: 341 // 342 // BB#1: 343 // ... 344 // Bne BB#2 345 // BB#4: 346 // v1024 = 347 // B BB#3 348 // BB#2: 349 // ... no uses of v1024 350 // <fallthrough> 351 // BB#3: 352 // ... 353 // = v1024 354 // 355 // This is incorrect since v1024 is not computed along the BB#1->BB#2->BB#3 356 // flow. We need to ensure the new basic block where the computation is 357 // sunk to dominates all the uses. 358 // It's only legal to break critical edge and sink the computation to the 359 // new block if all the predecessors of "To", except for "From", are 360 // not dominated by "From". Given SSA property, this means these 361 // predecessors are dominated by "To". 362 // 363 // There is no need to do this check if all the uses are PHI nodes. PHI 364 // sources are only defined on the specific predecessor edges. 365 if (!BreakPHIEdge) { 366 for (MachineBasicBlock::pred_iterator PI = ToBB->pred_begin(), 367 E = ToBB->pred_end(); PI != E; ++PI) { 368 if (*PI == FromBB) 369 continue; 370 if (!DT->dominates(ToBB, *PI)) 371 return 0; 372 } 373 } 374 375 return FromBB->SplitCriticalEdge(ToBB, this); 376} 377 378static bool AvoidsSinking(MachineInstr *MI, MachineRegisterInfo *MRI) { 379 return MI->isInsertSubreg() || MI->isSubregToReg() || MI->isRegSequence(); 380} 381 382/// SinkInstruction - Determine whether it is safe to sink the specified machine 383/// instruction out of its current block into a successor. 384bool MachineSinking::SinkInstruction(MachineInstr *MI, bool &SawStore) { 385 // Don't sink insert_subreg, subreg_to_reg, reg_sequence. These are meant to 386 // be close to the source to make it easier to coalesce. 387 if (AvoidsSinking(MI, MRI)) 388 return false; 389 390 // Check if it's safe to move the instruction. 391 if (!MI->isSafeToMove(TII, AA, SawStore)) 392 return false; 393 394 // FIXME: This should include support for sinking instructions within the 395 // block they are currently in to shorten the live ranges. We often get 396 // instructions sunk into the top of a large block, but it would be better to 397 // also sink them down before their first use in the block. This xform has to 398 // be careful not to *increase* register pressure though, e.g. sinking 399 // "x = y + z" down if it kills y and z would increase the live ranges of y 400 // and z and only shrink the live range of x. 401 402 // Loop over all the operands of the specified instruction. If there is 403 // anything we can't handle, bail out. 404 MachineBasicBlock *ParentBlock = MI->getParent(); 405 406 // SuccToSinkTo - This is the successor to sink this instruction to, once we 407 // decide. 408 MachineBasicBlock *SuccToSinkTo = 0; 409 410 bool BreakPHIEdge = false; 411 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 412 const MachineOperand &MO = MI->getOperand(i); 413 if (!MO.isReg()) continue; // Ignore non-register operands. 414 415 unsigned Reg = MO.getReg(); 416 if (Reg == 0) continue; 417 418 if (TargetRegisterInfo::isPhysicalRegister(Reg)) { 419 if (MO.isUse()) { 420 // If the physreg has no defs anywhere, it's just an ambient register 421 // and we can freely move its uses. Alternatively, if it's allocatable, 422 // it could get allocated to something with a def during allocation. 423 if (!MRI->def_empty(Reg)) 424 return false; 425 426 if (AllocatableSet.test(Reg)) 427 return false; 428 429 // Check for a def among the register's aliases too. 430 for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) { 431 unsigned AliasReg = *Alias; 432 if (!MRI->def_empty(AliasReg)) 433 return false; 434 435 if (AllocatableSet.test(AliasReg)) 436 return false; 437 } 438 } else if (!MO.isDead()) { 439 // A def that isn't dead. We can't move it. 440 return false; 441 } 442 } else { 443 // Virtual register uses are always safe to sink. 444 if (MO.isUse()) continue; 445 446 // If it's not safe to move defs of the register class, then abort. 447 if (!TII->isSafeToMoveRegClassDefs(MRI->getRegClass(Reg))) 448 return false; 449 450 // FIXME: This picks a successor to sink into based on having one 451 // successor that dominates all the uses. However, there are cases where 452 // sinking can happen but where the sink point isn't a successor. For 453 // example: 454 // 455 // x = computation 456 // if () {} else {} 457 // use x 458 // 459 // the instruction could be sunk over the whole diamond for the 460 // if/then/else (or loop, etc), allowing it to be sunk into other blocks 461 // after that. 462 463 // Virtual register defs can only be sunk if all their uses are in blocks 464 // dominated by one of the successors. 465 if (SuccToSinkTo) { 466 // If a previous operand picked a block to sink to, then this operand 467 // must be sinkable to the same block. 468 bool LocalUse = false; 469 if (!AllUsesDominatedByBlock(Reg, SuccToSinkTo, ParentBlock, 470 BreakPHIEdge, LocalUse)) 471 return false; 472 473 continue; 474 } 475 476 // Otherwise, we should look at all the successors and decide which one 477 // we should sink to. 478 for (MachineBasicBlock::succ_iterator SI = ParentBlock->succ_begin(), 479 E = ParentBlock->succ_end(); SI != E; ++SI) { 480 bool LocalUse = false; 481 if (AllUsesDominatedByBlock(Reg, *SI, ParentBlock, 482 BreakPHIEdge, LocalUse)) { 483 SuccToSinkTo = *SI; 484 break; 485 } 486 if (LocalUse) 487 // Def is used locally, it's never safe to move this def. 488 return false; 489 } 490 491 // If we couldn't find a block to sink to, ignore this instruction. 492 if (SuccToSinkTo == 0) 493 return false; 494 } 495 } 496 497 // If there are no outputs, it must have side-effects. 498 if (SuccToSinkTo == 0) 499 return false; 500 501 // It's not safe to sink instructions to EH landing pad. Control flow into 502 // landing pad is implicitly defined. 503 if (SuccToSinkTo->isLandingPad()) 504 return false; 505 506 // It is not possible to sink an instruction into its own block. This can 507 // happen with loops. 508 if (MI->getParent() == SuccToSinkTo) 509 return false; 510 511 // If the instruction to move defines a dead physical register which is live 512 // when leaving the basic block, don't move it because it could turn into a 513 // "zombie" define of that preg. E.g., EFLAGS. (<rdar://problem/8030636>) 514 for (unsigned I = 0, E = MI->getNumOperands(); I != E; ++I) { 515 const MachineOperand &MO = MI->getOperand(I); 516 if (!MO.isReg()) continue; 517 unsigned Reg = MO.getReg(); 518 if (Reg == 0 || !TargetRegisterInfo::isPhysicalRegister(Reg)) continue; 519 if (SuccToSinkTo->isLiveIn(Reg)) 520 return false; 521 } 522 523 DEBUG(dbgs() << "Sink instr " << *MI << "\tinto block " << *SuccToSinkTo); 524 525 // If the block has multiple predecessors, this would introduce computation on 526 // a path that it doesn't already exist. We could split the critical edge, 527 // but for now we just punt. 528 if (SuccToSinkTo->pred_size() > 1) { 529 // We cannot sink a load across a critical edge - there may be stores in 530 // other code paths. 531 bool TryBreak = false; 532 bool store = true; 533 if (!MI->isSafeToMove(TII, AA, store)) { 534 DEBUG(dbgs() << " *** NOTE: Won't sink load along critical edge.\n"); 535 TryBreak = true; 536 } 537 538 // We don't want to sink across a critical edge if we don't dominate the 539 // successor. We could be introducing calculations to new code paths. 540 if (!TryBreak && !DT->dominates(ParentBlock, SuccToSinkTo)) { 541 DEBUG(dbgs() << " *** NOTE: Critical edge found\n"); 542 TryBreak = true; 543 } 544 545 // Don't sink instructions into a loop. 546 if (!TryBreak && LI->isLoopHeader(SuccToSinkTo)) { 547 DEBUG(dbgs() << " *** NOTE: Loop header found\n"); 548 TryBreak = true; 549 } 550 551 // Otherwise we are OK with sinking along a critical edge. 552 if (!TryBreak) 553 DEBUG(dbgs() << "Sinking along critical edge.\n"); 554 else { 555 MachineBasicBlock *NewSucc = 556 SplitCriticalEdge(MI, ParentBlock, SuccToSinkTo, BreakPHIEdge); 557 if (!NewSucc) { 558 DEBUG(dbgs() << " *** PUNTING: Not legal or profitable to " 559 "break critical edge\n"); 560 return false; 561 } else { 562 DEBUG(dbgs() << " *** Splitting critical edge:" 563 " BB#" << ParentBlock->getNumber() 564 << " -- BB#" << NewSucc->getNumber() 565 << " -- BB#" << SuccToSinkTo->getNumber() << '\n'); 566 SuccToSinkTo = NewSucc; 567 ++NumSplit; 568 BreakPHIEdge = false; 569 } 570 } 571 } 572 573 if (BreakPHIEdge) { 574 // BreakPHIEdge is true if all the uses are in the successor MBB being 575 // sunken into and they are all PHI nodes. In this case, machine-sink must 576 // break the critical edge first. 577 MachineBasicBlock *NewSucc = SplitCriticalEdge(MI, ParentBlock, 578 SuccToSinkTo, BreakPHIEdge); 579 if (!NewSucc) { 580 DEBUG(dbgs() << " *** PUNTING: Not legal or profitable to " 581 "break critical edge\n"); 582 return false; 583 } 584 585 DEBUG(dbgs() << " *** Splitting critical edge:" 586 " BB#" << ParentBlock->getNumber() 587 << " -- BB#" << NewSucc->getNumber() 588 << " -- BB#" << SuccToSinkTo->getNumber() << '\n'); 589 SuccToSinkTo = NewSucc; 590 ++NumSplit; 591 } 592 593 // Determine where to insert into. Skip phi nodes. 594 MachineBasicBlock::iterator InsertPos = SuccToSinkTo->begin(); 595 while (InsertPos != SuccToSinkTo->end() && InsertPos->isPHI()) 596 ++InsertPos; 597 598 // Move the instruction. 599 SuccToSinkTo->splice(InsertPos, ParentBlock, MI, 600 ++MachineBasicBlock::iterator(MI)); 601 602 // Conservatively, clear any kill flags, since it's possible that they are no 603 // longer correct. 604 MI->clearKillInfo(); 605 606 return true; 607} 608