MachineSink.cpp revision e5e7946018844978d0ac09fdb35998a53b43ad34
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/Statistic.h" 29#include "llvm/Support/CommandLine.h" 30#include "llvm/Support/Debug.h" 31#include "llvm/Support/raw_ostream.h" 32using namespace llvm; 33 34static cl::opt<bool> 35SplitEdges("machine-sink-split", 36 cl::desc("Split critical edges during machine sinking"), 37 cl::init(false), cl::Hidden); 38static cl::opt<unsigned> 39SplitLimit("split-limit", 40 cl::init(~0u), cl::Hidden); 41 42STATISTIC(NumSunk, "Number of machine instructions sunk"); 43STATISTIC(NumSplit, "Number of critical edges split"); 44 45namespace { 46 class MachineSinking : public MachineFunctionPass { 47 const TargetInstrInfo *TII; 48 const TargetRegisterInfo *TRI; 49 MachineRegisterInfo *RegInfo; // Machine register information 50 MachineDominatorTree *DT; // Machine dominator tree 51 MachineLoopInfo *LI; 52 AliasAnalysis *AA; 53 BitVector AllocatableSet; // Which physregs are allocatable? 54 55 public: 56 static char ID; // Pass identification 57 MachineSinking() : MachineFunctionPass(ID) {} 58 59 virtual bool runOnMachineFunction(MachineFunction &MF); 60 61 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 62 AU.setPreservesCFG(); 63 MachineFunctionPass::getAnalysisUsage(AU); 64 AU.addRequired<AliasAnalysis>(); 65 AU.addRequired<MachineDominatorTree>(); 66 AU.addRequired<MachineLoopInfo>(); 67 AU.addPreserved<MachineDominatorTree>(); 68 AU.addPreserved<MachineLoopInfo>(); 69 } 70 private: 71 bool ProcessBlock(MachineBasicBlock &MBB); 72 MachineBasicBlock *SplitCriticalEdge(MachineBasicBlock *From, 73 MachineBasicBlock *To); 74 bool SinkInstruction(MachineInstr *MI, bool &SawStore); 75 bool AllUsesDominatedByBlock(unsigned Reg, MachineBasicBlock *MBB, 76 MachineBasicBlock *DefMBB, bool &LocalUse) const; 77 }; 78} // end anonymous namespace 79 80char MachineSinking::ID = 0; 81INITIALIZE_PASS(MachineSinking, "machine-sink", 82 "Machine code sinking", false, false); 83 84FunctionPass *llvm::createMachineSinkingPass() { return new MachineSinking(); } 85 86/// AllUsesDominatedByBlock - Return true if all uses of the specified register 87/// occur in blocks dominated by the specified block. If any use is in the 88/// definition block, then return false since it is never legal to move def 89/// after uses. 90bool MachineSinking::AllUsesDominatedByBlock(unsigned Reg, 91 MachineBasicBlock *MBB, 92 MachineBasicBlock *DefMBB, 93 bool &LocalUse) const { 94 assert(TargetRegisterInfo::isVirtualRegister(Reg) && 95 "Only makes sense for vregs"); 96 // Ignoring debug uses is necessary so debug info doesn't affect the code. 97 // This may leave a referencing dbg_value in the original block, before 98 // the definition of the vreg. Dwarf generator handles this although the 99 // user might not get the right info at runtime. 100 for (MachineRegisterInfo::use_nodbg_iterator 101 I = RegInfo->use_nodbg_begin(Reg), E = RegInfo->use_nodbg_end(); 102 I != E; ++I) { 103 // Determine the block of the use. 104 MachineInstr *UseInst = &*I; 105 MachineBasicBlock *UseBlock = UseInst->getParent(); 106 107 if (UseInst->isPHI()) { 108 // PHI nodes use the operand in the predecessor block, not the block with 109 // the PHI. 110 UseBlock = UseInst->getOperand(I.getOperandNo()+1).getMBB(); 111 } else if (UseBlock == DefMBB) { 112 LocalUse = true; 113 return false; 114 } 115 116 // Check that it dominates. 117 if (!DT->dominates(MBB, UseBlock)) 118 return false; 119 } 120 121 return true; 122} 123 124bool MachineSinking::runOnMachineFunction(MachineFunction &MF) { 125 DEBUG(dbgs() << "******** Machine Sinking ********\n"); 126 127 const TargetMachine &TM = MF.getTarget(); 128 TII = TM.getInstrInfo(); 129 TRI = TM.getRegisterInfo(); 130 RegInfo = &MF.getRegInfo(); 131 DT = &getAnalysis<MachineDominatorTree>(); 132 LI = &getAnalysis<MachineLoopInfo>(); 133 AA = &getAnalysis<AliasAnalysis>(); 134 AllocatableSet = TRI->getAllocatableSet(MF); 135 136 bool EverMadeChange = false; 137 138 while (1) { 139 bool MadeChange = false; 140 141 // Process all basic blocks. 142 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); 143 I != E; ++I) 144 MadeChange |= ProcessBlock(*I); 145 146 // If this iteration over the code changed anything, keep iterating. 147 if (!MadeChange) break; 148 EverMadeChange = true; 149 } 150 return EverMadeChange; 151} 152 153bool MachineSinking::ProcessBlock(MachineBasicBlock &MBB) { 154 // Can't sink anything out of a block that has less than two successors. 155 if (MBB.succ_size() <= 1 || MBB.empty()) return false; 156 157 // Don't bother sinking code out of unreachable blocks. In addition to being 158 // unprofitable, it can also lead to infinite looping, because in an 159 // unreachable loop there may be nowhere to stop. 160 if (!DT->isReachableFromEntry(&MBB)) return false; 161 162 bool MadeChange = false; 163 164 // Walk the basic block bottom-up. Remember if we saw a store. 165 MachineBasicBlock::iterator I = MBB.end(); 166 --I; 167 bool ProcessedBegin, SawStore = false; 168 do { 169 MachineInstr *MI = I; // The instruction to sink. 170 171 // Predecrement I (if it's not begin) so that it isn't invalidated by 172 // sinking. 173 ProcessedBegin = I == MBB.begin(); 174 if (!ProcessedBegin) 175 --I; 176 177 if (MI->isDebugValue()) 178 continue; 179 180 if (SinkInstruction(MI, SawStore)) 181 ++NumSunk, MadeChange = true; 182 183 // If we just processed the first instruction in the block, we're done. 184 } while (!ProcessedBegin); 185 186 return MadeChange; 187} 188 189MachineBasicBlock *MachineSinking::SplitCriticalEdge(MachineBasicBlock *FromBB, 190 MachineBasicBlock *ToBB) { 191 // Avoid breaking back edge. From == To means backedge for single BB loop. 192 if (!SplitEdges || NumSplit == SplitLimit || FromBB == ToBB) 193 return 0; 194 195 // Check for more "complex" loops. 196 if (LI->getLoopFor(FromBB) != LI->getLoopFor(ToBB) || 197 !LI->isLoopHeader(ToBB)) { 198 // It's not always legal to break critical edges and sink the computation 199 // to the edge. 200 // 201 // BB#1: 202 // v1024 203 // Beq BB#3 204 // <fallthrough> 205 // BB#2: 206 // ... no uses of v1024 207 // <fallthrough> 208 // BB#3: 209 // ... 210 // = v1024 211 // 212 // If BB#1 -> BB#3 edge is broken and computation of v1024 is inserted: 213 // 214 // BB#1: 215 // ... 216 // Bne BB#2 217 // BB#4: 218 // v1024 = 219 // B BB#3 220 // BB#2: 221 // ... no uses of v1024 222 // <fallthrough> 223 // BB#3: 224 // ... 225 // = v1024 226 // 227 // This is incorrect since v1024 is not computed along the BB#1->BB#2->BB#3 228 // flow. We need to ensure the new basic block where the computation is 229 // sunk to dominates all the uses. 230 // It's only legal to break critical edge and sink the computation to the 231 // new block if all the predecessors of "To", except for "From", are 232 // not dominated by "From". Given SSA property, this means these 233 // predecessors are dominated by "To". 234 for (MachineBasicBlock::pred_iterator PI = ToBB->pred_begin(), 235 E = ToBB->pred_end(); PI != E; ++PI) { 236 if (*PI == FromBB) 237 continue; 238 if (!DT->dominates(ToBB, *PI)) 239 return 0; 240 } 241 242 // FIXME: Determine if it's cost effective to break this edge. 243 return FromBB->SplitCriticalEdge(ToBB, this); 244 } 245 246 return 0; 247} 248 249/// SinkInstruction - Determine whether it is safe to sink the specified machine 250/// instruction out of its current block into a successor. 251bool MachineSinking::SinkInstruction(MachineInstr *MI, bool &SawStore) { 252 // Check if it's safe to move the instruction. 253 if (!MI->isSafeToMove(TII, AA, SawStore)) 254 return false; 255 256 // FIXME: This should include support for sinking instructions within the 257 // block they are currently in to shorten the live ranges. We often get 258 // instructions sunk into the top of a large block, but it would be better to 259 // also sink them down before their first use in the block. This xform has to 260 // be careful not to *increase* register pressure though, e.g. sinking 261 // "x = y + z" down if it kills y and z would increase the live ranges of y 262 // and z and only shrink the live range of x. 263 264 // Loop over all the operands of the specified instruction. If there is 265 // anything we can't handle, bail out. 266 MachineBasicBlock *ParentBlock = MI->getParent(); 267 268 // SuccToSinkTo - This is the successor to sink this instruction to, once we 269 // decide. 270 MachineBasicBlock *SuccToSinkTo = 0; 271 272 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 273 const MachineOperand &MO = MI->getOperand(i); 274 if (!MO.isReg()) continue; // Ignore non-register operands. 275 276 unsigned Reg = MO.getReg(); 277 if (Reg == 0) continue; 278 279 if (TargetRegisterInfo::isPhysicalRegister(Reg)) { 280 if (MO.isUse()) { 281 // If the physreg has no defs anywhere, it's just an ambient register 282 // and we can freely move its uses. Alternatively, if it's allocatable, 283 // it could get allocated to something with a def during allocation. 284 if (!RegInfo->def_empty(Reg)) 285 return false; 286 287 if (AllocatableSet.test(Reg)) 288 return false; 289 290 // Check for a def among the register's aliases too. 291 for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) { 292 unsigned AliasReg = *Alias; 293 if (!RegInfo->def_empty(AliasReg)) 294 return false; 295 296 if (AllocatableSet.test(AliasReg)) 297 return false; 298 } 299 } else if (!MO.isDead()) { 300 // A def that isn't dead. We can't move it. 301 return false; 302 } 303 } else { 304 // Virtual register uses are always safe to sink. 305 if (MO.isUse()) continue; 306 307 // If it's not safe to move defs of the register class, then abort. 308 if (!TII->isSafeToMoveRegClassDefs(RegInfo->getRegClass(Reg))) 309 return false; 310 311 // FIXME: This picks a successor to sink into based on having one 312 // successor that dominates all the uses. However, there are cases where 313 // sinking can happen but where the sink point isn't a successor. For 314 // example: 315 // 316 // x = computation 317 // if () {} else {} 318 // use x 319 // 320 // the instruction could be sunk over the whole diamond for the 321 // if/then/else (or loop, etc), allowing it to be sunk into other blocks 322 // after that. 323 324 // Virtual register defs can only be sunk if all their uses are in blocks 325 // dominated by one of the successors. 326 if (SuccToSinkTo) { 327 // If a previous operand picked a block to sink to, then this operand 328 // must be sinkable to the same block. 329 bool LocalUse = false; 330 if (!AllUsesDominatedByBlock(Reg, SuccToSinkTo, ParentBlock, LocalUse)) 331 return false; 332 333 continue; 334 } 335 336 // Otherwise, we should look at all the successors and decide which one 337 // we should sink to. 338 for (MachineBasicBlock::succ_iterator SI = ParentBlock->succ_begin(), 339 E = ParentBlock->succ_end(); SI != E; ++SI) { 340 bool LocalUse = false; 341 if (AllUsesDominatedByBlock(Reg, *SI, ParentBlock, LocalUse)) { 342 SuccToSinkTo = *SI; 343 break; 344 } 345 if (LocalUse) 346 // Def is used locally, it's never safe to move this def. 347 return false; 348 } 349 350 // If we couldn't find a block to sink to, ignore this instruction. 351 if (SuccToSinkTo == 0) 352 return false; 353 } 354 } 355 356 // If there are no outputs, it must have side-effects. 357 if (SuccToSinkTo == 0) 358 return false; 359 360 // It's not safe to sink instructions to EH landing pad. Control flow into 361 // landing pad is implicitly defined. 362 if (SuccToSinkTo->isLandingPad()) 363 return false; 364 365 // It is not possible to sink an instruction into its own block. This can 366 // happen with loops. 367 if (MI->getParent() == SuccToSinkTo) 368 return false; 369 370 // If the instruction to move defines a dead physical register which is live 371 // when leaving the basic block, don't move it because it could turn into a 372 // "zombie" define of that preg. E.g., EFLAGS. (<rdar://problem/8030636>) 373 for (unsigned I = 0, E = MI->getNumOperands(); I != E; ++I) { 374 const MachineOperand &MO = MI->getOperand(I); 375 if (!MO.isReg()) continue; 376 unsigned Reg = MO.getReg(); 377 if (Reg == 0 || !TargetRegisterInfo::isPhysicalRegister(Reg)) continue; 378 if (SuccToSinkTo->isLiveIn(Reg)) 379 return false; 380 } 381 382 DEBUG(dbgs() << "Sink instr " << *MI << "\tinto block " << *SuccToSinkTo); 383 384 // If the block has multiple predecessors, this would introduce computation on 385 // a path that it doesn't already exist. We could split the critical edge, 386 // but for now we just punt. 387 // FIXME: Split critical edges if not backedges. 388 if (SuccToSinkTo->pred_size() > 1) { 389 // We cannot sink a load across a critical edge - there may be stores in 390 // other code paths. 391 bool TryBreak = false; 392 bool store = true; 393 if (!MI->isSafeToMove(TII, AA, store)) { 394 DEBUG(dbgs() << " *** PUNTING: Won't sink load along critical edge.\n"); 395 TryBreak = true; 396 } 397 398 // We don't want to sink across a critical edge if we don't dominate the 399 // successor. We could be introducing calculations to new code paths. 400 if (!TryBreak && !DT->dominates(ParentBlock, SuccToSinkTo)) { 401 DEBUG(dbgs() << " *** PUNTING: Critical edge found\n"); 402 TryBreak = true; 403 } 404 405 // Don't sink instructions into a loop. 406 if (!TryBreak && LI->isLoopHeader(SuccToSinkTo)) { 407 DEBUG(dbgs() << " *** PUNTING: Loop header found\n"); 408 TryBreak = true; 409 } 410 411 // Otherwise we are OK with sinking along a critical edge. 412 if (!TryBreak) 413 DEBUG(dbgs() << "Sinking along critical edge.\n"); 414 else { 415 MachineBasicBlock *NewSucc = SplitCriticalEdge(ParentBlock, SuccToSinkTo); 416 if (!NewSucc) { 417 DEBUG(dbgs() << 418 " *** PUNTING: Not legal or profitable to break critical edge\n"); 419 return false; 420 } else { 421 DEBUG(dbgs() << "*** Splitting critical edge:" 422 " BB#" << ParentBlock->getNumber() 423 << " -- BB#" << NewSucc->getNumber() 424 << " -- BB#" << SuccToSinkTo->getNumber() << '\n'); 425 SuccToSinkTo = NewSucc; 426 ++NumSplit; 427 } 428 } 429 } 430 431 // Determine where to insert into. Skip phi nodes. 432 MachineBasicBlock::iterator InsertPos = SuccToSinkTo->begin(); 433 while (InsertPos != SuccToSinkTo->end() && InsertPos->isPHI()) 434 ++InsertPos; 435 436 // Move the instruction. 437 SuccToSinkTo->splice(InsertPos, ParentBlock, MI, 438 ++MachineBasicBlock::iterator(MI)); 439 440 // Conservatively, clear any kill flags, since it's possible that they are no 441 // longer correct. 442 MI->clearKillInfo(); 443 444 return true; 445} 446