RegisterCoalescer.cpp revision 34af6f597b09c13fba7d3a1960c0810cfc30beff
1//===- RegisterCoalescer.cpp - Generic Register Coalescing Interface -------==// 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 file implements the generic RegisterCoalescer interface which 11// is used as the common interface used by all clients and 12// implementations of register coalescing. 13// 14//===----------------------------------------------------------------------===// 15 16#define DEBUG_TYPE "regalloc" 17#include "RegisterCoalescer.h" 18#include "LiveDebugVariables.h" 19#include "VirtRegMap.h" 20 21#include "llvm/Pass.h" 22#include "llvm/Value.h" 23#include "llvm/ADT/OwningPtr.h" 24#include "llvm/ADT/STLExtras.h" 25#include "llvm/ADT/SmallSet.h" 26#include "llvm/ADT/Statistic.h" 27#include "llvm/Analysis/AliasAnalysis.h" 28#include "llvm/CodeGen/LiveIntervalAnalysis.h" 29#include "llvm/CodeGen/LiveIntervalAnalysis.h" 30#include "llvm/CodeGen/LiveRangeEdit.h" 31#include "llvm/CodeGen/MachineFrameInfo.h" 32#include "llvm/CodeGen/MachineInstr.h" 33#include "llvm/CodeGen/MachineInstr.h" 34#include "llvm/CodeGen/MachineLoopInfo.h" 35#include "llvm/CodeGen/MachineRegisterInfo.h" 36#include "llvm/CodeGen/MachineRegisterInfo.h" 37#include "llvm/CodeGen/Passes.h" 38#include "llvm/CodeGen/RegisterClassInfo.h" 39#include "llvm/Support/CommandLine.h" 40#include "llvm/Support/Debug.h" 41#include "llvm/Support/ErrorHandling.h" 42#include "llvm/Support/raw_ostream.h" 43#include "llvm/Target/TargetInstrInfo.h" 44#include "llvm/Target/TargetInstrInfo.h" 45#include "llvm/Target/TargetMachine.h" 46#include "llvm/Target/TargetOptions.h" 47#include "llvm/Target/TargetRegisterInfo.h" 48#include <algorithm> 49#include <cmath> 50using namespace llvm; 51 52STATISTIC(numJoins , "Number of interval joins performed"); 53STATISTIC(numCrossRCs , "Number of cross class joins performed"); 54STATISTIC(numCommutes , "Number of instruction commuting performed"); 55STATISTIC(numExtends , "Number of copies extended"); 56STATISTIC(NumReMats , "Number of instructions re-materialized"); 57STATISTIC(NumInflated , "Number of register classes inflated"); 58 59static cl::opt<bool> 60EnableJoining("join-liveintervals", 61 cl::desc("Coalesce copies (default=true)"), 62 cl::init(true)); 63 64static cl::opt<bool> 65VerifyCoalescing("verify-coalescing", 66 cl::desc("Verify machine instrs before and after register coalescing"), 67 cl::Hidden); 68 69namespace { 70 class RegisterCoalescer : public MachineFunctionPass, 71 private LiveRangeEdit::Delegate { 72 MachineFunction* MF; 73 MachineRegisterInfo* MRI; 74 const TargetMachine* TM; 75 const TargetRegisterInfo* TRI; 76 const TargetInstrInfo* TII; 77 LiveIntervals *LIS; 78 LiveDebugVariables *LDV; 79 const MachineLoopInfo* Loops; 80 AliasAnalysis *AA; 81 RegisterClassInfo RegClassInfo; 82 83 /// WorkList - Copy instructions yet to be coalesced. 84 SmallVector<MachineInstr*, 8> WorkList; 85 86 /// ErasedInstrs - Set of instruction pointers that have been erased, and 87 /// that may be present in WorkList. 88 SmallPtrSet<MachineInstr*, 8> ErasedInstrs; 89 90 /// Dead instructions that are about to be deleted. 91 SmallVector<MachineInstr*, 8> DeadDefs; 92 93 /// Virtual registers to be considered for register class inflation. 94 SmallVector<unsigned, 8> InflateRegs; 95 96 /// Recursively eliminate dead defs in DeadDefs. 97 void eliminateDeadDefs(); 98 99 /// LiveRangeEdit callback. 100 void LRE_WillEraseInstruction(MachineInstr *MI); 101 102 /// joinAllIntervals - join compatible live intervals 103 void joinAllIntervals(); 104 105 /// copyCoalesceInMBB - Coalesce copies in the specified MBB, putting 106 /// copies that cannot yet be coalesced into WorkList. 107 void copyCoalesceInMBB(MachineBasicBlock *MBB); 108 109 /// copyCoalesceWorkList - Try to coalesce all copies in WorkList after 110 /// position From. Return true if any progress was made. 111 bool copyCoalesceWorkList(unsigned From = 0); 112 113 /// joinCopy - Attempt to join intervals corresponding to SrcReg/DstReg, 114 /// which are the src/dst of the copy instruction CopyMI. This returns 115 /// true if the copy was successfully coalesced away. If it is not 116 /// currently possible to coalesce this interval, but it may be possible if 117 /// other things get coalesced, then it returns true by reference in 118 /// 'Again'. 119 bool joinCopy(MachineInstr *TheCopy, bool &Again); 120 121 /// joinIntervals - Attempt to join these two intervals. On failure, this 122 /// returns false. The output "SrcInt" will not have been modified, so we 123 /// can use this information below to update aliases. 124 bool joinIntervals(CoalescerPair &CP); 125 126 /// Attempt joining with a reserved physreg. 127 bool joinReservedPhysReg(CoalescerPair &CP); 128 129 /// adjustCopiesBackFrom - We found a non-trivially-coalescable copy. If 130 /// the source value number is defined by a copy from the destination reg 131 /// see if we can merge these two destination reg valno# into a single 132 /// value number, eliminating a copy. 133 bool adjustCopiesBackFrom(const CoalescerPair &CP, MachineInstr *CopyMI); 134 135 /// hasOtherReachingDefs - Return true if there are definitions of IntB 136 /// other than BValNo val# that can reach uses of AValno val# of IntA. 137 bool hasOtherReachingDefs(LiveInterval &IntA, LiveInterval &IntB, 138 VNInfo *AValNo, VNInfo *BValNo); 139 140 /// removeCopyByCommutingDef - We found a non-trivially-coalescable copy. 141 /// If the source value number is defined by a commutable instruction and 142 /// its other operand is coalesced to the copy dest register, see if we 143 /// can transform the copy into a noop by commuting the definition. 144 bool removeCopyByCommutingDef(const CoalescerPair &CP,MachineInstr *CopyMI); 145 146 /// reMaterializeTrivialDef - If the source of a copy is defined by a 147 /// trivial computation, replace the copy by rematerialize the definition. 148 bool reMaterializeTrivialDef(LiveInterval &SrcInt, unsigned DstReg, 149 MachineInstr *CopyMI); 150 151 /// canJoinPhys - Return true if a physreg copy should be joined. 152 bool canJoinPhys(CoalescerPair &CP); 153 154 /// updateRegDefsUses - Replace all defs and uses of SrcReg to DstReg and 155 /// update the subregister number if it is not zero. If DstReg is a 156 /// physical register and the existing subregister number of the def / use 157 /// being updated is not zero, make sure to set it to the correct physical 158 /// subregister. 159 void updateRegDefsUses(unsigned SrcReg, unsigned DstReg, unsigned SubIdx); 160 161 /// eliminateUndefCopy - Handle copies of undef values. 162 bool eliminateUndefCopy(MachineInstr *CopyMI, const CoalescerPair &CP); 163 164 public: 165 static char ID; // Class identification, replacement for typeinfo 166 RegisterCoalescer() : MachineFunctionPass(ID) { 167 initializeRegisterCoalescerPass(*PassRegistry::getPassRegistry()); 168 } 169 170 virtual void getAnalysisUsage(AnalysisUsage &AU) const; 171 172 virtual void releaseMemory(); 173 174 /// runOnMachineFunction - pass entry point 175 virtual bool runOnMachineFunction(MachineFunction&); 176 177 /// print - Implement the dump method. 178 virtual void print(raw_ostream &O, const Module* = 0) const; 179 }; 180} /// end anonymous namespace 181 182char &llvm::RegisterCoalescerID = RegisterCoalescer::ID; 183 184INITIALIZE_PASS_BEGIN(RegisterCoalescer, "simple-register-coalescing", 185 "Simple Register Coalescing", false, false) 186INITIALIZE_PASS_DEPENDENCY(LiveIntervals) 187INITIALIZE_PASS_DEPENDENCY(LiveDebugVariables) 188INITIALIZE_PASS_DEPENDENCY(SlotIndexes) 189INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo) 190INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 191INITIALIZE_PASS_END(RegisterCoalescer, "simple-register-coalescing", 192 "Simple Register Coalescing", false, false) 193 194char RegisterCoalescer::ID = 0; 195 196static unsigned compose(const TargetRegisterInfo &tri, unsigned a, unsigned b) { 197 if (!a) return b; 198 if (!b) return a; 199 return tri.composeSubRegIndices(a, b); 200} 201 202static bool isMoveInstr(const TargetRegisterInfo &tri, const MachineInstr *MI, 203 unsigned &Src, unsigned &Dst, 204 unsigned &SrcSub, unsigned &DstSub) { 205 if (MI->isCopy()) { 206 Dst = MI->getOperand(0).getReg(); 207 DstSub = MI->getOperand(0).getSubReg(); 208 Src = MI->getOperand(1).getReg(); 209 SrcSub = MI->getOperand(1).getSubReg(); 210 } else if (MI->isSubregToReg()) { 211 Dst = MI->getOperand(0).getReg(); 212 DstSub = compose(tri, MI->getOperand(0).getSubReg(), 213 MI->getOperand(3).getImm()); 214 Src = MI->getOperand(2).getReg(); 215 SrcSub = MI->getOperand(2).getSubReg(); 216 } else 217 return false; 218 return true; 219} 220 221bool CoalescerPair::setRegisters(const MachineInstr *MI) { 222 SrcReg = DstReg = 0; 223 SrcIdx = DstIdx = 0; 224 NewRC = 0; 225 Flipped = CrossClass = false; 226 227 unsigned Src, Dst, SrcSub, DstSub; 228 if (!isMoveInstr(TRI, MI, Src, Dst, SrcSub, DstSub)) 229 return false; 230 Partial = SrcSub || DstSub; 231 232 // If one register is a physreg, it must be Dst. 233 if (TargetRegisterInfo::isPhysicalRegister(Src)) { 234 if (TargetRegisterInfo::isPhysicalRegister(Dst)) 235 return false; 236 std::swap(Src, Dst); 237 std::swap(SrcSub, DstSub); 238 Flipped = true; 239 } 240 241 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo(); 242 243 if (TargetRegisterInfo::isPhysicalRegister(Dst)) { 244 // Eliminate DstSub on a physreg. 245 if (DstSub) { 246 Dst = TRI.getSubReg(Dst, DstSub); 247 if (!Dst) return false; 248 DstSub = 0; 249 } 250 251 // Eliminate SrcSub by picking a corresponding Dst superregister. 252 if (SrcSub) { 253 Dst = TRI.getMatchingSuperReg(Dst, SrcSub, MRI.getRegClass(Src)); 254 if (!Dst) return false; 255 SrcSub = 0; 256 } else if (!MRI.getRegClass(Src)->contains(Dst)) { 257 return false; 258 } 259 } else { 260 // Both registers are virtual. 261 const TargetRegisterClass *SrcRC = MRI.getRegClass(Src); 262 const TargetRegisterClass *DstRC = MRI.getRegClass(Dst); 263 264 // Both registers have subreg indices. 265 if (SrcSub && DstSub) { 266 // Copies between different sub-registers are never coalescable. 267 if (Src == Dst && SrcSub != DstSub) 268 return false; 269 270 NewRC = TRI.getCommonSuperRegClass(SrcRC, SrcSub, DstRC, DstSub, 271 SrcIdx, DstIdx); 272 if (!NewRC) 273 return false; 274 } else if (DstSub) { 275 // SrcReg will be merged with a sub-register of DstReg. 276 SrcIdx = DstSub; 277 NewRC = TRI.getMatchingSuperRegClass(DstRC, SrcRC, DstSub); 278 } else if (SrcSub) { 279 // DstReg will be merged with a sub-register of SrcReg. 280 DstIdx = SrcSub; 281 NewRC = TRI.getMatchingSuperRegClass(SrcRC, DstRC, SrcSub); 282 } else { 283 // This is a straight copy without sub-registers. 284 NewRC = TRI.getCommonSubClass(DstRC, SrcRC); 285 } 286 287 // The combined constraint may be impossible to satisfy. 288 if (!NewRC) 289 return false; 290 291 // Prefer SrcReg to be a sub-register of DstReg. 292 // FIXME: Coalescer should support subregs symmetrically. 293 if (DstIdx && !SrcIdx) { 294 std::swap(Src, Dst); 295 std::swap(SrcIdx, DstIdx); 296 Flipped = !Flipped; 297 } 298 299 CrossClass = NewRC != DstRC || NewRC != SrcRC; 300 } 301 // Check our invariants 302 assert(TargetRegisterInfo::isVirtualRegister(Src) && "Src must be virtual"); 303 assert(!(TargetRegisterInfo::isPhysicalRegister(Dst) && DstSub) && 304 "Cannot have a physical SubIdx"); 305 SrcReg = Src; 306 DstReg = Dst; 307 return true; 308} 309 310bool CoalescerPair::flip() { 311 if (TargetRegisterInfo::isPhysicalRegister(DstReg)) 312 return false; 313 std::swap(SrcReg, DstReg); 314 std::swap(SrcIdx, DstIdx); 315 Flipped = !Flipped; 316 return true; 317} 318 319bool CoalescerPair::isCoalescable(const MachineInstr *MI) const { 320 if (!MI) 321 return false; 322 unsigned Src, Dst, SrcSub, DstSub; 323 if (!isMoveInstr(TRI, MI, Src, Dst, SrcSub, DstSub)) 324 return false; 325 326 // Find the virtual register that is SrcReg. 327 if (Dst == SrcReg) { 328 std::swap(Src, Dst); 329 std::swap(SrcSub, DstSub); 330 } else if (Src != SrcReg) { 331 return false; 332 } 333 334 // Now check that Dst matches DstReg. 335 if (TargetRegisterInfo::isPhysicalRegister(DstReg)) { 336 if (!TargetRegisterInfo::isPhysicalRegister(Dst)) 337 return false; 338 assert(!DstIdx && !SrcIdx && "Inconsistent CoalescerPair state."); 339 // DstSub could be set for a physreg from INSERT_SUBREG. 340 if (DstSub) 341 Dst = TRI.getSubReg(Dst, DstSub); 342 // Full copy of Src. 343 if (!SrcSub) 344 return DstReg == Dst; 345 // This is a partial register copy. Check that the parts match. 346 return TRI.getSubReg(DstReg, SrcSub) == Dst; 347 } else { 348 // DstReg is virtual. 349 if (DstReg != Dst) 350 return false; 351 // Registers match, do the subregisters line up? 352 return compose(TRI, SrcIdx, SrcSub) == compose(TRI, DstIdx, DstSub); 353 } 354} 355 356void RegisterCoalescer::getAnalysisUsage(AnalysisUsage &AU) const { 357 AU.setPreservesCFG(); 358 AU.addRequired<AliasAnalysis>(); 359 AU.addRequired<LiveIntervals>(); 360 AU.addPreserved<LiveIntervals>(); 361 AU.addRequired<LiveDebugVariables>(); 362 AU.addPreserved<LiveDebugVariables>(); 363 AU.addPreserved<SlotIndexes>(); 364 AU.addRequired<MachineLoopInfo>(); 365 AU.addPreserved<MachineLoopInfo>(); 366 AU.addPreservedID(MachineDominatorsID); 367 MachineFunctionPass::getAnalysisUsage(AU); 368} 369 370void RegisterCoalescer::eliminateDeadDefs() { 371 SmallVector<LiveInterval*, 8> NewRegs; 372 LiveRangeEdit(0, NewRegs, *MF, *LIS, 0, this).eliminateDeadDefs(DeadDefs); 373} 374 375// Callback from eliminateDeadDefs(). 376void RegisterCoalescer::LRE_WillEraseInstruction(MachineInstr *MI) { 377 // MI may be in WorkList. Make sure we don't visit it. 378 ErasedInstrs.insert(MI); 379} 380 381/// adjustCopiesBackFrom - We found a non-trivially-coalescable copy with IntA 382/// being the source and IntB being the dest, thus this defines a value number 383/// in IntB. If the source value number (in IntA) is defined by a copy from B, 384/// see if we can merge these two pieces of B into a single value number, 385/// eliminating a copy. For example: 386/// 387/// A3 = B0 388/// ... 389/// B1 = A3 <- this copy 390/// 391/// In this case, B0 can be extended to where the B1 copy lives, allowing the B1 392/// value number to be replaced with B0 (which simplifies the B liveinterval). 393/// 394/// This returns true if an interval was modified. 395/// 396bool RegisterCoalescer::adjustCopiesBackFrom(const CoalescerPair &CP, 397 MachineInstr *CopyMI) { 398 assert(!CP.isPartial() && "This doesn't work for partial copies."); 399 assert(!CP.isPhys() && "This doesn't work for physreg copies."); 400 401 LiveInterval &IntA = 402 LIS->getInterval(CP.isFlipped() ? CP.getDstReg() : CP.getSrcReg()); 403 LiveInterval &IntB = 404 LIS->getInterval(CP.isFlipped() ? CP.getSrcReg() : CP.getDstReg()); 405 SlotIndex CopyIdx = LIS->getInstructionIndex(CopyMI).getRegSlot(); 406 407 // BValNo is a value number in B that is defined by a copy from A. 'B3' in 408 // the example above. 409 LiveInterval::iterator BLR = IntB.FindLiveRangeContaining(CopyIdx); 410 if (BLR == IntB.end()) return false; 411 VNInfo *BValNo = BLR->valno; 412 413 // Get the location that B is defined at. Two options: either this value has 414 // an unknown definition point or it is defined at CopyIdx. If unknown, we 415 // can't process it. 416 if (BValNo->def != CopyIdx) return false; 417 418 // AValNo is the value number in A that defines the copy, A3 in the example. 419 SlotIndex CopyUseIdx = CopyIdx.getRegSlot(true); 420 LiveInterval::iterator ALR = IntA.FindLiveRangeContaining(CopyUseIdx); 421 // The live range might not exist after fun with physreg coalescing. 422 if (ALR == IntA.end()) return false; 423 VNInfo *AValNo = ALR->valno; 424 425 // If AValNo is defined as a copy from IntB, we can potentially process this. 426 // Get the instruction that defines this value number. 427 MachineInstr *ACopyMI = LIS->getInstructionFromIndex(AValNo->def); 428 if (!CP.isCoalescable(ACopyMI)) 429 return false; 430 431 // Get the LiveRange in IntB that this value number starts with. 432 LiveInterval::iterator ValLR = 433 IntB.FindLiveRangeContaining(AValNo->def.getPrevSlot()); 434 if (ValLR == IntB.end()) 435 return false; 436 437 // Make sure that the end of the live range is inside the same block as 438 // CopyMI. 439 MachineInstr *ValLREndInst = 440 LIS->getInstructionFromIndex(ValLR->end.getPrevSlot()); 441 if (!ValLREndInst || ValLREndInst->getParent() != CopyMI->getParent()) 442 return false; 443 444 // Okay, we now know that ValLR ends in the same block that the CopyMI 445 // live-range starts. If there are no intervening live ranges between them in 446 // IntB, we can merge them. 447 if (ValLR+1 != BLR) return false; 448 449 DEBUG(dbgs() << "Extending: " << PrintReg(IntB.reg, TRI)); 450 451 SlotIndex FillerStart = ValLR->end, FillerEnd = BLR->start; 452 // We are about to delete CopyMI, so need to remove it as the 'instruction 453 // that defines this value #'. Update the valnum with the new defining 454 // instruction #. 455 BValNo->def = FillerStart; 456 457 // Okay, we can merge them. We need to insert a new liverange: 458 // [ValLR.end, BLR.begin) of either value number, then we merge the 459 // two value numbers. 460 IntB.addRange(LiveRange(FillerStart, FillerEnd, BValNo)); 461 462 // Okay, merge "B1" into the same value number as "B0". 463 if (BValNo != ValLR->valno) { 464 // If B1 is killed by a PHI, then the merged live range must also be killed 465 // by the same PHI, as B0 and B1 can not overlap. 466 bool HasPHIKill = BValNo->hasPHIKill(); 467 IntB.MergeValueNumberInto(BValNo, ValLR->valno); 468 if (HasPHIKill) 469 ValLR->valno->setHasPHIKill(true); 470 } 471 DEBUG(dbgs() << " result = " << IntB << '\n'); 472 473 // If the source instruction was killing the source register before the 474 // merge, unset the isKill marker given the live range has been extended. 475 int UIdx = ValLREndInst->findRegisterUseOperandIdx(IntB.reg, true); 476 if (UIdx != -1) { 477 ValLREndInst->getOperand(UIdx).setIsKill(false); 478 } 479 480 // Rewrite the copy. If the copy instruction was killing the destination 481 // register before the merge, find the last use and trim the live range. That 482 // will also add the isKill marker. 483 CopyMI->substituteRegister(IntA.reg, IntB.reg, 0, *TRI); 484 if (ALR->end == CopyIdx) 485 LIS->shrinkToUses(&IntA); 486 487 ++numExtends; 488 return true; 489} 490 491/// hasOtherReachingDefs - Return true if there are definitions of IntB 492/// other than BValNo val# that can reach uses of AValno val# of IntA. 493bool RegisterCoalescer::hasOtherReachingDefs(LiveInterval &IntA, 494 LiveInterval &IntB, 495 VNInfo *AValNo, 496 VNInfo *BValNo) { 497 for (LiveInterval::iterator AI = IntA.begin(), AE = IntA.end(); 498 AI != AE; ++AI) { 499 if (AI->valno != AValNo) continue; 500 LiveInterval::Ranges::iterator BI = 501 std::upper_bound(IntB.ranges.begin(), IntB.ranges.end(), AI->start); 502 if (BI != IntB.ranges.begin()) 503 --BI; 504 for (; BI != IntB.ranges.end() && AI->end >= BI->start; ++BI) { 505 if (BI->valno == BValNo) 506 continue; 507 if (BI->start <= AI->start && BI->end > AI->start) 508 return true; 509 if (BI->start > AI->start && BI->start < AI->end) 510 return true; 511 } 512 } 513 return false; 514} 515 516/// removeCopyByCommutingDef - We found a non-trivially-coalescable copy with 517/// IntA being the source and IntB being the dest, thus this defines a value 518/// number in IntB. If the source value number (in IntA) is defined by a 519/// commutable instruction and its other operand is coalesced to the copy dest 520/// register, see if we can transform the copy into a noop by commuting the 521/// definition. For example, 522/// 523/// A3 = op A2 B0<kill> 524/// ... 525/// B1 = A3 <- this copy 526/// ... 527/// = op A3 <- more uses 528/// 529/// ==> 530/// 531/// B2 = op B0 A2<kill> 532/// ... 533/// B1 = B2 <- now an identify copy 534/// ... 535/// = op B2 <- more uses 536/// 537/// This returns true if an interval was modified. 538/// 539bool RegisterCoalescer::removeCopyByCommutingDef(const CoalescerPair &CP, 540 MachineInstr *CopyMI) { 541 assert (!CP.isPhys()); 542 543 SlotIndex CopyIdx = LIS->getInstructionIndex(CopyMI).getRegSlot(); 544 545 LiveInterval &IntA = 546 LIS->getInterval(CP.isFlipped() ? CP.getDstReg() : CP.getSrcReg()); 547 LiveInterval &IntB = 548 LIS->getInterval(CP.isFlipped() ? CP.getSrcReg() : CP.getDstReg()); 549 550 // BValNo is a value number in B that is defined by a copy from A. 'B3' in 551 // the example above. 552 VNInfo *BValNo = IntB.getVNInfoAt(CopyIdx); 553 if (!BValNo || BValNo->def != CopyIdx) 554 return false; 555 556 assert(BValNo->def == CopyIdx && "Copy doesn't define the value?"); 557 558 // AValNo is the value number in A that defines the copy, A3 in the example. 559 VNInfo *AValNo = IntA.getVNInfoAt(CopyIdx.getRegSlot(true)); 560 assert(AValNo && "COPY source not live"); 561 562 // If other defs can reach uses of this def, then it's not safe to perform 563 // the optimization. 564 if (AValNo->isPHIDef() || AValNo->isUnused() || AValNo->hasPHIKill()) 565 return false; 566 MachineInstr *DefMI = LIS->getInstructionFromIndex(AValNo->def); 567 if (!DefMI) 568 return false; 569 if (!DefMI->isCommutable()) 570 return false; 571 // If DefMI is a two-address instruction then commuting it will change the 572 // destination register. 573 int DefIdx = DefMI->findRegisterDefOperandIdx(IntA.reg); 574 assert(DefIdx != -1); 575 unsigned UseOpIdx; 576 if (!DefMI->isRegTiedToUseOperand(DefIdx, &UseOpIdx)) 577 return false; 578 unsigned Op1, Op2, NewDstIdx; 579 if (!TII->findCommutedOpIndices(DefMI, Op1, Op2)) 580 return false; 581 if (Op1 == UseOpIdx) 582 NewDstIdx = Op2; 583 else if (Op2 == UseOpIdx) 584 NewDstIdx = Op1; 585 else 586 return false; 587 588 MachineOperand &NewDstMO = DefMI->getOperand(NewDstIdx); 589 unsigned NewReg = NewDstMO.getReg(); 590 if (NewReg != IntB.reg || !NewDstMO.isKill()) 591 return false; 592 593 // Make sure there are no other definitions of IntB that would reach the 594 // uses which the new definition can reach. 595 if (hasOtherReachingDefs(IntA, IntB, AValNo, BValNo)) 596 return false; 597 598 // If some of the uses of IntA.reg is already coalesced away, return false. 599 // It's not possible to determine whether it's safe to perform the coalescing. 600 for (MachineRegisterInfo::use_nodbg_iterator UI = 601 MRI->use_nodbg_begin(IntA.reg), 602 UE = MRI->use_nodbg_end(); UI != UE; ++UI) { 603 MachineInstr *UseMI = &*UI; 604 SlotIndex UseIdx = LIS->getInstructionIndex(UseMI); 605 LiveInterval::iterator ULR = IntA.FindLiveRangeContaining(UseIdx); 606 if (ULR == IntA.end() || ULR->valno != AValNo) 607 continue; 608 // If this use is tied to a def, we can't rewrite the register. 609 if (UseMI->isRegTiedToDefOperand(UI.getOperandNo())) 610 return false; 611 } 612 613 DEBUG(dbgs() << "\tremoveCopyByCommutingDef: " << AValNo->def << '\t' 614 << *DefMI); 615 616 // At this point we have decided that it is legal to do this 617 // transformation. Start by commuting the instruction. 618 MachineBasicBlock *MBB = DefMI->getParent(); 619 MachineInstr *NewMI = TII->commuteInstruction(DefMI); 620 if (!NewMI) 621 return false; 622 if (TargetRegisterInfo::isVirtualRegister(IntA.reg) && 623 TargetRegisterInfo::isVirtualRegister(IntB.reg) && 624 !MRI->constrainRegClass(IntB.reg, MRI->getRegClass(IntA.reg))) 625 return false; 626 if (NewMI != DefMI) { 627 LIS->ReplaceMachineInstrInMaps(DefMI, NewMI); 628 MachineBasicBlock::iterator Pos = DefMI; 629 MBB->insert(Pos, NewMI); 630 MBB->erase(DefMI); 631 } 632 unsigned OpIdx = NewMI->findRegisterUseOperandIdx(IntA.reg, false); 633 NewMI->getOperand(OpIdx).setIsKill(); 634 635 // If ALR and BLR overlaps and end of BLR extends beyond end of ALR, e.g. 636 // A = or A, B 637 // ... 638 // B = A 639 // ... 640 // C = A<kill> 641 // ... 642 // = B 643 644 // Update uses of IntA of the specific Val# with IntB. 645 for (MachineRegisterInfo::use_iterator UI = MRI->use_begin(IntA.reg), 646 UE = MRI->use_end(); UI != UE;) { 647 MachineOperand &UseMO = UI.getOperand(); 648 MachineInstr *UseMI = &*UI; 649 ++UI; 650 if (UseMI->isDebugValue()) { 651 // FIXME These don't have an instruction index. Not clear we have enough 652 // info to decide whether to do this replacement or not. For now do it. 653 UseMO.setReg(NewReg); 654 continue; 655 } 656 SlotIndex UseIdx = LIS->getInstructionIndex(UseMI).getRegSlot(true); 657 LiveInterval::iterator ULR = IntA.FindLiveRangeContaining(UseIdx); 658 if (ULR == IntA.end() || ULR->valno != AValNo) 659 continue; 660 // Kill flags are no longer accurate. They are recomputed after RA. 661 UseMO.setIsKill(false); 662 if (TargetRegisterInfo::isPhysicalRegister(NewReg)) 663 UseMO.substPhysReg(NewReg, *TRI); 664 else 665 UseMO.setReg(NewReg); 666 if (UseMI == CopyMI) 667 continue; 668 if (!UseMI->isCopy()) 669 continue; 670 if (UseMI->getOperand(0).getReg() != IntB.reg || 671 UseMI->getOperand(0).getSubReg()) 672 continue; 673 674 // This copy will become a noop. If it's defining a new val#, merge it into 675 // BValNo. 676 SlotIndex DefIdx = UseIdx.getRegSlot(); 677 VNInfo *DVNI = IntB.getVNInfoAt(DefIdx); 678 if (!DVNI) 679 continue; 680 DEBUG(dbgs() << "\t\tnoop: " << DefIdx << '\t' << *UseMI); 681 assert(DVNI->def == DefIdx); 682 BValNo = IntB.MergeValueNumberInto(BValNo, DVNI); 683 ErasedInstrs.insert(UseMI); 684 LIS->RemoveMachineInstrFromMaps(UseMI); 685 UseMI->eraseFromParent(); 686 } 687 688 // Extend BValNo by merging in IntA live ranges of AValNo. Val# definition 689 // is updated. 690 VNInfo *ValNo = BValNo; 691 ValNo->def = AValNo->def; 692 for (LiveInterval::iterator AI = IntA.begin(), AE = IntA.end(); 693 AI != AE; ++AI) { 694 if (AI->valno != AValNo) continue; 695 IntB.addRange(LiveRange(AI->start, AI->end, ValNo)); 696 } 697 DEBUG(dbgs() << "\t\textended: " << IntB << '\n'); 698 699 IntA.removeValNo(AValNo); 700 DEBUG(dbgs() << "\t\ttrimmed: " << IntA << '\n'); 701 ++numCommutes; 702 return true; 703} 704 705/// reMaterializeTrivialDef - If the source of a copy is defined by a trivial 706/// computation, replace the copy by rematerialize the definition. 707bool RegisterCoalescer::reMaterializeTrivialDef(LiveInterval &SrcInt, 708 unsigned DstReg, 709 MachineInstr *CopyMI) { 710 SlotIndex CopyIdx = LIS->getInstructionIndex(CopyMI).getRegSlot(true); 711 LiveInterval::iterator SrcLR = SrcInt.FindLiveRangeContaining(CopyIdx); 712 assert(SrcLR != SrcInt.end() && "Live range not found!"); 713 VNInfo *ValNo = SrcLR->valno; 714 if (ValNo->isPHIDef() || ValNo->isUnused()) 715 return false; 716 MachineInstr *DefMI = LIS->getInstructionFromIndex(ValNo->def); 717 if (!DefMI) 718 return false; 719 assert(DefMI && "Defining instruction disappeared"); 720 if (!DefMI->isAsCheapAsAMove()) 721 return false; 722 if (!TII->isTriviallyReMaterializable(DefMI, AA)) 723 return false; 724 bool SawStore = false; 725 if (!DefMI->isSafeToMove(TII, AA, SawStore)) 726 return false; 727 const MCInstrDesc &MCID = DefMI->getDesc(); 728 if (MCID.getNumDefs() != 1) 729 return false; 730 if (!DefMI->isImplicitDef()) { 731 // Make sure the copy destination register class fits the instruction 732 // definition register class. The mismatch can happen as a result of earlier 733 // extract_subreg, insert_subreg, subreg_to_reg coalescing. 734 const TargetRegisterClass *RC = TII->getRegClass(MCID, 0, TRI, *MF); 735 if (TargetRegisterInfo::isVirtualRegister(DstReg)) { 736 if (MRI->getRegClass(DstReg) != RC) 737 return false; 738 } else if (!RC->contains(DstReg)) 739 return false; 740 } 741 742 MachineBasicBlock *MBB = CopyMI->getParent(); 743 MachineBasicBlock::iterator MII = 744 llvm::next(MachineBasicBlock::iterator(CopyMI)); 745 TII->reMaterialize(*MBB, MII, DstReg, 0, DefMI, *TRI); 746 MachineInstr *NewMI = prior(MII); 747 748 // NewMI may have dead implicit defs (E.g. EFLAGS for MOV<bits>r0 on X86). 749 // We need to remember these so we can add intervals once we insert 750 // NewMI into SlotIndexes. 751 SmallVector<unsigned, 4> NewMIImplDefs; 752 for (unsigned i = NewMI->getDesc().getNumOperands(), 753 e = NewMI->getNumOperands(); i != e; ++i) { 754 MachineOperand &MO = NewMI->getOperand(i); 755 if (MO.isReg()) { 756 assert(MO.isDef() && MO.isImplicit() && MO.isDead() && 757 TargetRegisterInfo::isPhysicalRegister(MO.getReg())); 758 NewMIImplDefs.push_back(MO.getReg()); 759 } 760 } 761 762 // CopyMI may have implicit operands, transfer them over to the newly 763 // rematerialized instruction. And update implicit def interval valnos. 764 for (unsigned i = CopyMI->getDesc().getNumOperands(), 765 e = CopyMI->getNumOperands(); i != e; ++i) { 766 MachineOperand &MO = CopyMI->getOperand(i); 767 if (MO.isReg()) { 768 assert(MO.isImplicit() && "No explicit operands after implict operands."); 769 // Discard VReg implicit defs. 770 if (TargetRegisterInfo::isPhysicalRegister(MO.getReg())) { 771 NewMI->addOperand(MO); 772 } 773 } 774 } 775 776 LIS->ReplaceMachineInstrInMaps(CopyMI, NewMI); 777 778 SlotIndex NewMIIdx = LIS->getInstructionIndex(NewMI); 779 for (unsigned i = 0, e = NewMIImplDefs.size(); i != e; ++i) { 780 unsigned Reg = NewMIImplDefs[i]; 781 for (MCRegUnitIterator Units(Reg, TRI); Units.isValid(); ++Units) 782 if (LiveInterval *LI = LIS->getCachedRegUnit(*Units)) 783 LI->createDeadDef(NewMIIdx.getRegSlot(), LIS->getVNInfoAllocator()); 784 } 785 786 CopyMI->eraseFromParent(); 787 ErasedInstrs.insert(CopyMI); 788 DEBUG(dbgs() << "Remat: " << *NewMI); 789 ++NumReMats; 790 791 // The source interval can become smaller because we removed a use. 792 LIS->shrinkToUses(&SrcInt, &DeadDefs); 793 if (!DeadDefs.empty()) 794 eliminateDeadDefs(); 795 796 return true; 797} 798 799/// eliminateUndefCopy - ProcessImpicitDefs may leave some copies of <undef> 800/// values, it only removes local variables. When we have a copy like: 801/// 802/// %vreg1 = COPY %vreg2<undef> 803/// 804/// We delete the copy and remove the corresponding value number from %vreg1. 805/// Any uses of that value number are marked as <undef>. 806bool RegisterCoalescer::eliminateUndefCopy(MachineInstr *CopyMI, 807 const CoalescerPair &CP) { 808 SlotIndex Idx = LIS->getInstructionIndex(CopyMI); 809 LiveInterval *SrcInt = &LIS->getInterval(CP.getSrcReg()); 810 if (SrcInt->liveAt(Idx)) 811 return false; 812 LiveInterval *DstInt = &LIS->getInterval(CP.getDstReg()); 813 if (DstInt->liveAt(Idx)) 814 return false; 815 816 // No intervals are live-in to CopyMI - it is undef. 817 if (CP.isFlipped()) 818 DstInt = SrcInt; 819 SrcInt = 0; 820 821 VNInfo *DeadVNI = DstInt->getVNInfoAt(Idx.getRegSlot()); 822 assert(DeadVNI && "No value defined in DstInt"); 823 DstInt->removeValNo(DeadVNI); 824 825 // Find new undef uses. 826 for (MachineRegisterInfo::reg_nodbg_iterator 827 I = MRI->reg_nodbg_begin(DstInt->reg), E = MRI->reg_nodbg_end(); 828 I != E; ++I) { 829 MachineOperand &MO = I.getOperand(); 830 if (MO.isDef() || MO.isUndef()) 831 continue; 832 MachineInstr *MI = MO.getParent(); 833 SlotIndex Idx = LIS->getInstructionIndex(MI); 834 if (DstInt->liveAt(Idx)) 835 continue; 836 MO.setIsUndef(true); 837 DEBUG(dbgs() << "\tnew undef: " << Idx << '\t' << *MI); 838 } 839 return true; 840} 841 842/// updateRegDefsUses - Replace all defs and uses of SrcReg to DstReg and 843/// update the subregister number if it is not zero. If DstReg is a 844/// physical register and the existing subregister number of the def / use 845/// being updated is not zero, make sure to set it to the correct physical 846/// subregister. 847void RegisterCoalescer::updateRegDefsUses(unsigned SrcReg, 848 unsigned DstReg, 849 unsigned SubIdx) { 850 bool DstIsPhys = TargetRegisterInfo::isPhysicalRegister(DstReg); 851 LiveInterval *DstInt = DstIsPhys ? 0 : &LIS->getInterval(DstReg); 852 853 // Update LiveDebugVariables. 854 LDV->renameRegister(SrcReg, DstReg, SubIdx); 855 856 for (MachineRegisterInfo::reg_iterator I = MRI->reg_begin(SrcReg); 857 MachineInstr *UseMI = I.skipInstruction();) { 858 SmallVector<unsigned,8> Ops; 859 bool Reads, Writes; 860 tie(Reads, Writes) = UseMI->readsWritesVirtualRegister(SrcReg, &Ops); 861 862 // If SrcReg wasn't read, it may still be the case that DstReg is live-in 863 // because SrcReg is a sub-register. 864 if (DstInt && !Reads && SubIdx) 865 Reads = DstInt->liveAt(LIS->getInstructionIndex(UseMI)); 866 867 // Replace SrcReg with DstReg in all UseMI operands. 868 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 869 MachineOperand &MO = UseMI->getOperand(Ops[i]); 870 871 // Adjust <undef> flags in case of sub-register joins. We don't want to 872 // turn a full def into a read-modify-write sub-register def and vice 873 // versa. 874 if (SubIdx && MO.isDef()) 875 MO.setIsUndef(!Reads); 876 877 if (DstIsPhys) 878 MO.substPhysReg(DstReg, *TRI); 879 else 880 MO.substVirtReg(DstReg, SubIdx, *TRI); 881 } 882 883 DEBUG({ 884 dbgs() << "\t\tupdated: "; 885 if (!UseMI->isDebugValue()) 886 dbgs() << LIS->getInstructionIndex(UseMI) << "\t"; 887 dbgs() << *UseMI; 888 }); 889 } 890} 891 892/// canJoinPhys - Return true if a copy involving a physreg should be joined. 893bool RegisterCoalescer::canJoinPhys(CoalescerPair &CP) { 894 /// Always join simple intervals that are defined by a single copy from a 895 /// reserved register. This doesn't increase register pressure, so it is 896 /// always beneficial. 897 if (!RegClassInfo.isReserved(CP.getDstReg())) { 898 DEBUG(dbgs() << "\tCan only merge into reserved registers.\n"); 899 return false; 900 } 901 902 LiveInterval &JoinVInt = LIS->getInterval(CP.getSrcReg()); 903 if (CP.isFlipped() && JoinVInt.containsOneValue()) 904 return true; 905 906 DEBUG(dbgs() << "\tCannot join defs into reserved register.\n"); 907 return false; 908} 909 910/// joinCopy - Attempt to join intervals corresponding to SrcReg/DstReg, 911/// which are the src/dst of the copy instruction CopyMI. This returns true 912/// if the copy was successfully coalesced away. If it is not currently 913/// possible to coalesce this interval, but it may be possible if other 914/// things get coalesced, then it returns true by reference in 'Again'. 915bool RegisterCoalescer::joinCopy(MachineInstr *CopyMI, bool &Again) { 916 917 Again = false; 918 DEBUG(dbgs() << LIS->getInstructionIndex(CopyMI) << '\t' << *CopyMI); 919 920 CoalescerPair CP(*TRI); 921 if (!CP.setRegisters(CopyMI)) { 922 DEBUG(dbgs() << "\tNot coalescable.\n"); 923 return false; 924 } 925 926 // Dead code elimination. This really should be handled by MachineDCE, but 927 // sometimes dead copies slip through, and we can't generate invalid live 928 // ranges. 929 if (!CP.isPhys() && CopyMI->allDefsAreDead()) { 930 DEBUG(dbgs() << "\tCopy is dead.\n"); 931 DeadDefs.push_back(CopyMI); 932 eliminateDeadDefs(); 933 return true; 934 } 935 936 // Eliminate undefs. 937 if (!CP.isPhys() && eliminateUndefCopy(CopyMI, CP)) { 938 DEBUG(dbgs() << "\tEliminated copy of <undef> value.\n"); 939 LIS->RemoveMachineInstrFromMaps(CopyMI); 940 CopyMI->eraseFromParent(); 941 return false; // Not coalescable. 942 } 943 944 // Coalesced copies are normally removed immediately, but transformations 945 // like removeCopyByCommutingDef() can inadvertently create identity copies. 946 // When that happens, just join the values and remove the copy. 947 if (CP.getSrcReg() == CP.getDstReg()) { 948 LiveInterval &LI = LIS->getInterval(CP.getSrcReg()); 949 DEBUG(dbgs() << "\tCopy already coalesced: " << LI << '\n'); 950 LiveRangeQuery LRQ(LI, LIS->getInstructionIndex(CopyMI)); 951 if (VNInfo *DefVNI = LRQ.valueDefined()) { 952 VNInfo *ReadVNI = LRQ.valueIn(); 953 assert(ReadVNI && "No value before copy and no <undef> flag."); 954 assert(ReadVNI != DefVNI && "Cannot read and define the same value."); 955 LI.MergeValueNumberInto(DefVNI, ReadVNI); 956 DEBUG(dbgs() << "\tMerged values: " << LI << '\n'); 957 } 958 LIS->RemoveMachineInstrFromMaps(CopyMI); 959 CopyMI->eraseFromParent(); 960 return true; 961 } 962 963 // Enforce policies. 964 if (CP.isPhys()) { 965 DEBUG(dbgs() << "\tConsidering merging " << PrintReg(CP.getSrcReg(), TRI) 966 << " with " << PrintReg(CP.getDstReg(), TRI, CP.getSrcIdx()) 967 << '\n'); 968 if (!canJoinPhys(CP)) { 969 // Before giving up coalescing, if definition of source is defined by 970 // trivial computation, try rematerializing it. 971 if (!CP.isFlipped() && 972 reMaterializeTrivialDef(LIS->getInterval(CP.getSrcReg()), 973 CP.getDstReg(), CopyMI)) 974 return true; 975 return false; 976 } 977 } else { 978 DEBUG({ 979 dbgs() << "\tConsidering merging to " << CP.getNewRC()->getName() 980 << " with "; 981 if (CP.getDstIdx() && CP.getSrcIdx()) 982 dbgs() << PrintReg(CP.getDstReg()) << " in " 983 << TRI->getSubRegIndexName(CP.getDstIdx()) << " and " 984 << PrintReg(CP.getSrcReg()) << " in " 985 << TRI->getSubRegIndexName(CP.getSrcIdx()) << '\n'; 986 else 987 dbgs() << PrintReg(CP.getSrcReg(), TRI) << " in " 988 << PrintReg(CP.getDstReg(), TRI, CP.getSrcIdx()) << '\n'; 989 }); 990 991 // When possible, let DstReg be the larger interval. 992 if (!CP.isPartial() && LIS->getInterval(CP.getSrcReg()).ranges.size() > 993 LIS->getInterval(CP.getDstReg()).ranges.size()) 994 CP.flip(); 995 } 996 997 // Okay, attempt to join these two intervals. On failure, this returns false. 998 // Otherwise, if one of the intervals being joined is a physreg, this method 999 // always canonicalizes DstInt to be it. The output "SrcInt" will not have 1000 // been modified, so we can use this information below to update aliases. 1001 if (!joinIntervals(CP)) { 1002 // Coalescing failed. 1003 1004 // If definition of source is defined by trivial computation, try 1005 // rematerializing it. 1006 if (!CP.isFlipped() && 1007 reMaterializeTrivialDef(LIS->getInterval(CP.getSrcReg()), 1008 CP.getDstReg(), CopyMI)) 1009 return true; 1010 1011 // If we can eliminate the copy without merging the live ranges, do so now. 1012 if (!CP.isPartial() && !CP.isPhys()) { 1013 if (adjustCopiesBackFrom(CP, CopyMI) || 1014 removeCopyByCommutingDef(CP, CopyMI)) { 1015 LIS->RemoveMachineInstrFromMaps(CopyMI); 1016 CopyMI->eraseFromParent(); 1017 DEBUG(dbgs() << "\tTrivial!\n"); 1018 return true; 1019 } 1020 } 1021 1022 // Otherwise, we are unable to join the intervals. 1023 DEBUG(dbgs() << "\tInterference!\n"); 1024 Again = true; // May be possible to coalesce later. 1025 return false; 1026 } 1027 1028 // Coalescing to a virtual register that is of a sub-register class of the 1029 // other. Make sure the resulting register is set to the right register class. 1030 if (CP.isCrossClass()) { 1031 ++numCrossRCs; 1032 MRI->setRegClass(CP.getDstReg(), CP.getNewRC()); 1033 } 1034 1035 // Removing sub-register copies can ease the register class constraints. 1036 // Make sure we attempt to inflate the register class of DstReg. 1037 if (!CP.isPhys() && RegClassInfo.isProperSubClass(CP.getNewRC())) 1038 InflateRegs.push_back(CP.getDstReg()); 1039 1040 // CopyMI has been erased by joinIntervals at this point. Remove it from 1041 // ErasedInstrs since copyCoalesceWorkList() won't add a successful join back 1042 // to the work list. This keeps ErasedInstrs from growing needlessly. 1043 ErasedInstrs.erase(CopyMI); 1044 1045 // Rewrite all SrcReg operands to DstReg. 1046 // Also update DstReg operands to include DstIdx if it is set. 1047 if (CP.getDstIdx()) 1048 updateRegDefsUses(CP.getDstReg(), CP.getDstReg(), CP.getDstIdx()); 1049 updateRegDefsUses(CP.getSrcReg(), CP.getDstReg(), CP.getSrcIdx()); 1050 1051 // SrcReg is guaranteed to be the register whose live interval that is 1052 // being merged. 1053 LIS->removeInterval(CP.getSrcReg()); 1054 1055 // Update regalloc hint. 1056 TRI->UpdateRegAllocHint(CP.getSrcReg(), CP.getDstReg(), *MF); 1057 1058 DEBUG({ 1059 dbgs() << "\tJoined. Result = " << PrintReg(CP.getDstReg(), TRI); 1060 if (!CP.isPhys()) 1061 dbgs() << LIS->getInterval(CP.getDstReg()); 1062 dbgs() << '\n'; 1063 }); 1064 1065 ++numJoins; 1066 return true; 1067} 1068 1069/// Attempt joining with a reserved physreg. 1070bool RegisterCoalescer::joinReservedPhysReg(CoalescerPair &CP) { 1071 assert(CP.isPhys() && "Must be a physreg copy"); 1072 assert(RegClassInfo.isReserved(CP.getDstReg()) && "Not a reserved register"); 1073 LiveInterval &RHS = LIS->getInterval(CP.getSrcReg()); 1074 DEBUG(dbgs() << "\t\tRHS = " << PrintReg(CP.getSrcReg()) << ' ' << RHS 1075 << '\n'); 1076 1077 assert(CP.isFlipped() && RHS.containsOneValue() && 1078 "Invalid join with reserved register"); 1079 1080 // Optimization for reserved registers like ESP. We can only merge with a 1081 // reserved physreg if RHS has a single value that is a copy of CP.DstReg(). 1082 // The live range of the reserved register will look like a set of dead defs 1083 // - we don't properly track the live range of reserved registers. 1084 1085 // Deny any overlapping intervals. This depends on all the reserved 1086 // register live ranges to look like dead defs. 1087 for (MCRegUnitIterator UI(CP.getDstReg(), TRI); UI.isValid(); ++UI) 1088 if (RHS.overlaps(LIS->getRegUnit(*UI))) { 1089 DEBUG(dbgs() << "\t\tInterference: " << PrintRegUnit(*UI, TRI) << '\n'); 1090 return false; 1091 } 1092 1093 // Skip any value computations, we are not adding new values to the 1094 // reserved register. Also skip merging the live ranges, the reserved 1095 // register live range doesn't need to be accurate as long as all the 1096 // defs are there. 1097 1098 // We don't track kills for reserved registers. 1099 MRI->clearKillFlags(CP.getSrcReg()); 1100 1101 return true; 1102} 1103 1104/// ComputeUltimateVN - Assuming we are going to join two live intervals, 1105/// compute what the resultant value numbers for each value in the input two 1106/// ranges will be. This is complicated by copies between the two which can 1107/// and will commonly cause multiple value numbers to be merged into one. 1108/// 1109/// VN is the value number that we're trying to resolve. InstDefiningValue 1110/// keeps track of the new InstDefiningValue assignment for the result 1111/// LiveInterval. ThisFromOther/OtherFromThis are sets that keep track of 1112/// whether a value in this or other is a copy from the opposite set. 1113/// ThisValNoAssignments/OtherValNoAssignments keep track of value #'s that have 1114/// already been assigned. 1115/// 1116/// ThisFromOther[x] - If x is defined as a copy from the other interval, this 1117/// contains the value number the copy is from. 1118/// 1119static unsigned ComputeUltimateVN(VNInfo *VNI, 1120 SmallVector<VNInfo*, 16> &NewVNInfo, 1121 DenseMap<VNInfo*, VNInfo*> &ThisFromOther, 1122 DenseMap<VNInfo*, VNInfo*> &OtherFromThis, 1123 SmallVector<int, 16> &ThisValNoAssignments, 1124 SmallVector<int, 16> &OtherValNoAssignments) { 1125 unsigned VN = VNI->id; 1126 1127 // If the VN has already been computed, just return it. 1128 if (ThisValNoAssignments[VN] >= 0) 1129 return ThisValNoAssignments[VN]; 1130 assert(ThisValNoAssignments[VN] != -2 && "Cyclic value numbers"); 1131 1132 // If this val is not a copy from the other val, then it must be a new value 1133 // number in the destination. 1134 DenseMap<VNInfo*, VNInfo*>::iterator I = ThisFromOther.find(VNI); 1135 if (I == ThisFromOther.end()) { 1136 NewVNInfo.push_back(VNI); 1137 return ThisValNoAssignments[VN] = NewVNInfo.size()-1; 1138 } 1139 VNInfo *OtherValNo = I->second; 1140 1141 // Otherwise, this *is* a copy from the RHS. If the other side has already 1142 // been computed, return it. 1143 if (OtherValNoAssignments[OtherValNo->id] >= 0) 1144 return ThisValNoAssignments[VN] = OtherValNoAssignments[OtherValNo->id]; 1145 1146 // Mark this value number as currently being computed, then ask what the 1147 // ultimate value # of the other value is. 1148 ThisValNoAssignments[VN] = -2; 1149 unsigned UltimateVN = 1150 ComputeUltimateVN(OtherValNo, NewVNInfo, OtherFromThis, ThisFromOther, 1151 OtherValNoAssignments, ThisValNoAssignments); 1152 return ThisValNoAssignments[VN] = UltimateVN; 1153} 1154 1155 1156// Find out if we have something like 1157// A = X 1158// B = X 1159// if so, we can pretend this is actually 1160// A = X 1161// B = A 1162// which allows us to coalesce A and B. 1163// VNI is the definition of B. LR is the life range of A that includes 1164// the slot just before B. If we return true, we add "B = X" to DupCopies. 1165// This implies that A dominates B. 1166static bool RegistersDefinedFromSameValue(LiveIntervals &li, 1167 const TargetRegisterInfo &tri, 1168 CoalescerPair &CP, 1169 VNInfo *VNI, 1170 VNInfo *OtherVNI, 1171 SmallVector<MachineInstr*, 8> &DupCopies) { 1172 // FIXME: This is very conservative. For example, we don't handle 1173 // physical registers. 1174 1175 MachineInstr *MI = li.getInstructionFromIndex(VNI->def); 1176 1177 if (!MI || CP.isPartial() || CP.isPhys()) 1178 return false; 1179 1180 unsigned A = CP.getDstReg(); 1181 if (!TargetRegisterInfo::isVirtualRegister(A)) 1182 return false; 1183 1184 unsigned B = CP.getSrcReg(); 1185 if (!TargetRegisterInfo::isVirtualRegister(B)) 1186 return false; 1187 1188 MachineInstr *OtherMI = li.getInstructionFromIndex(OtherVNI->def); 1189 if (!OtherMI) 1190 return false; 1191 1192 if (MI->isImplicitDef()) { 1193 DupCopies.push_back(MI); 1194 return true; 1195 } else { 1196 if (!MI->isFullCopy()) 1197 return false; 1198 unsigned Src = MI->getOperand(1).getReg(); 1199 if (!TargetRegisterInfo::isVirtualRegister(Src)) 1200 return false; 1201 if (!OtherMI->isFullCopy()) 1202 return false; 1203 unsigned OtherSrc = OtherMI->getOperand(1).getReg(); 1204 if (!TargetRegisterInfo::isVirtualRegister(OtherSrc)) 1205 return false; 1206 1207 if (Src != OtherSrc) 1208 return false; 1209 1210 // If the copies use two different value numbers of X, we cannot merge 1211 // A and B. 1212 LiveInterval &SrcInt = li.getInterval(Src); 1213 // getVNInfoBefore returns NULL for undef copies. In this case, the 1214 // optimization is still safe. 1215 if (SrcInt.getVNInfoBefore(OtherVNI->def) != 1216 SrcInt.getVNInfoBefore(VNI->def)) 1217 return false; 1218 1219 DupCopies.push_back(MI); 1220 return true; 1221 } 1222} 1223 1224/// joinIntervals - Attempt to join these two intervals. On failure, this 1225/// returns false. 1226bool RegisterCoalescer::joinIntervals(CoalescerPair &CP) { 1227 // Handle physreg joins separately. 1228 if (CP.isPhys()) 1229 return joinReservedPhysReg(CP); 1230 1231 LiveInterval &RHS = LIS->getInterval(CP.getSrcReg()); 1232 DEBUG(dbgs() << "\t\tRHS = " << PrintReg(CP.getSrcReg()) << ' ' << RHS 1233 << '\n'); 1234 1235 // Compute the final value assignment, assuming that the live ranges can be 1236 // coalesced. 1237 SmallVector<int, 16> LHSValNoAssignments; 1238 SmallVector<int, 16> RHSValNoAssignments; 1239 DenseMap<VNInfo*, VNInfo*> LHSValsDefinedFromRHS; 1240 DenseMap<VNInfo*, VNInfo*> RHSValsDefinedFromLHS; 1241 SmallVector<VNInfo*, 16> NewVNInfo; 1242 1243 SmallVector<MachineInstr*, 8> DupCopies; 1244 SmallVector<MachineInstr*, 8> DeadCopies; 1245 1246 LiveInterval &LHS = LIS->getOrCreateInterval(CP.getDstReg()); 1247 DEBUG(dbgs() << "\t\tLHS = " << PrintReg(CP.getDstReg(), TRI) << ' ' << LHS 1248 << '\n'); 1249 1250 // Loop over the value numbers of the LHS, seeing if any are defined from 1251 // the RHS. 1252 for (LiveInterval::vni_iterator i = LHS.vni_begin(), e = LHS.vni_end(); 1253 i != e; ++i) { 1254 VNInfo *VNI = *i; 1255 if (VNI->isUnused() || VNI->isPHIDef()) 1256 continue; 1257 MachineInstr *MI = LIS->getInstructionFromIndex(VNI->def); 1258 assert(MI && "Missing def"); 1259 if (!MI->isCopyLike() && !MI->isImplicitDef()) // Src not defined by a copy? 1260 continue; 1261 1262 // Figure out the value # from the RHS. 1263 VNInfo *OtherVNI = RHS.getVNInfoBefore(VNI->def); 1264 // The copy could be to an aliased physreg. 1265 if (!OtherVNI) 1266 continue; 1267 1268 // DstReg is known to be a register in the LHS interval. If the src is 1269 // from the RHS interval, we can use its value #. 1270 if (CP.isCoalescable(MI)) 1271 DeadCopies.push_back(MI); 1272 else if (!RegistersDefinedFromSameValue(*LIS, *TRI, CP, VNI, OtherVNI, 1273 DupCopies)) 1274 continue; 1275 1276 LHSValsDefinedFromRHS[VNI] = OtherVNI; 1277 } 1278 1279 // Loop over the value numbers of the RHS, seeing if any are defined from 1280 // the LHS. 1281 for (LiveInterval::vni_iterator i = RHS.vni_begin(), e = RHS.vni_end(); 1282 i != e; ++i) { 1283 VNInfo *VNI = *i; 1284 if (VNI->isUnused() || VNI->isPHIDef()) 1285 continue; 1286 MachineInstr *MI = LIS->getInstructionFromIndex(VNI->def); 1287 assert(MI && "Missing def"); 1288 if (!MI->isCopyLike() && !MI->isImplicitDef()) // Src not defined by a copy? 1289 continue; 1290 1291 // Figure out the value # from the LHS. 1292 VNInfo *OtherVNI = LHS.getVNInfoBefore(VNI->def); 1293 // The copy could be to an aliased physreg. 1294 if (!OtherVNI) 1295 continue; 1296 1297 // DstReg is known to be a register in the RHS interval. If the src is 1298 // from the LHS interval, we can use its value #. 1299 if (CP.isCoalescable(MI)) 1300 DeadCopies.push_back(MI); 1301 else if (!RegistersDefinedFromSameValue(*LIS, *TRI, CP, VNI, OtherVNI, 1302 DupCopies)) 1303 continue; 1304 1305 RHSValsDefinedFromLHS[VNI] = OtherVNI; 1306 } 1307 1308 LHSValNoAssignments.resize(LHS.getNumValNums(), -1); 1309 RHSValNoAssignments.resize(RHS.getNumValNums(), -1); 1310 NewVNInfo.reserve(LHS.getNumValNums() + RHS.getNumValNums()); 1311 1312 for (LiveInterval::vni_iterator i = LHS.vni_begin(), e = LHS.vni_end(); 1313 i != e; ++i) { 1314 VNInfo *VNI = *i; 1315 unsigned VN = VNI->id; 1316 if (LHSValNoAssignments[VN] >= 0 || VNI->isUnused()) 1317 continue; 1318 ComputeUltimateVN(VNI, NewVNInfo, 1319 LHSValsDefinedFromRHS, RHSValsDefinedFromLHS, 1320 LHSValNoAssignments, RHSValNoAssignments); 1321 } 1322 for (LiveInterval::vni_iterator i = RHS.vni_begin(), e = RHS.vni_end(); 1323 i != e; ++i) { 1324 VNInfo *VNI = *i; 1325 unsigned VN = VNI->id; 1326 if (RHSValNoAssignments[VN] >= 0 || VNI->isUnused()) 1327 continue; 1328 // If this value number isn't a copy from the LHS, it's a new number. 1329 if (RHSValsDefinedFromLHS.find(VNI) == RHSValsDefinedFromLHS.end()) { 1330 NewVNInfo.push_back(VNI); 1331 RHSValNoAssignments[VN] = NewVNInfo.size()-1; 1332 continue; 1333 } 1334 1335 ComputeUltimateVN(VNI, NewVNInfo, 1336 RHSValsDefinedFromLHS, LHSValsDefinedFromRHS, 1337 RHSValNoAssignments, LHSValNoAssignments); 1338 } 1339 1340 // Armed with the mappings of LHS/RHS values to ultimate values, walk the 1341 // interval lists to see if these intervals are coalescable. 1342 LiveInterval::const_iterator I = LHS.begin(); 1343 LiveInterval::const_iterator IE = LHS.end(); 1344 LiveInterval::const_iterator J = RHS.begin(); 1345 LiveInterval::const_iterator JE = RHS.end(); 1346 1347 // Collect interval end points that will no longer be kills. 1348 SmallVector<MachineInstr*, 8> LHSOldKills; 1349 SmallVector<MachineInstr*, 8> RHSOldKills; 1350 1351 // Skip ahead until the first place of potential sharing. 1352 if (I != IE && J != JE) { 1353 if (I->start < J->start) { 1354 I = std::upper_bound(I, IE, J->start); 1355 if (I != LHS.begin()) --I; 1356 } else if (J->start < I->start) { 1357 J = std::upper_bound(J, JE, I->start); 1358 if (J != RHS.begin()) --J; 1359 } 1360 } 1361 1362 while (I != IE && J != JE) { 1363 // Determine if these two live ranges overlap. 1364 // If so, check value # info to determine if they are really different. 1365 if (I->end > J->start && J->end > I->start) { 1366 // If the live range overlap will map to the same value number in the 1367 // result liverange, we can still coalesce them. If not, we can't. 1368 if (LHSValNoAssignments[I->valno->id] != 1369 RHSValNoAssignments[J->valno->id]) 1370 return false; 1371 1372 // Extended live ranges should no longer be killed. 1373 if (!I->end.isBlock() && I->end < J->end) 1374 if (MachineInstr *MI = LIS->getInstructionFromIndex(I->end)) 1375 LHSOldKills.push_back(MI); 1376 if (!J->end.isBlock() && J->end < I->end) 1377 if (MachineInstr *MI = LIS->getInstructionFromIndex(J->end)) 1378 RHSOldKills.push_back(MI); 1379 } 1380 1381 if (I->end < J->end) 1382 ++I; 1383 else 1384 ++J; 1385 } 1386 1387 // Update kill info. Some live ranges are extended due to copy coalescing. 1388 for (DenseMap<VNInfo*, VNInfo*>::iterator I = LHSValsDefinedFromRHS.begin(), 1389 E = LHSValsDefinedFromRHS.end(); I != E; ++I) { 1390 VNInfo *VNI = I->first; 1391 unsigned LHSValID = LHSValNoAssignments[VNI->id]; 1392 if (VNI->hasPHIKill()) 1393 NewVNInfo[LHSValID]->setHasPHIKill(true); 1394 } 1395 1396 // Update kill info. Some live ranges are extended due to copy coalescing. 1397 for (DenseMap<VNInfo*, VNInfo*>::iterator I = RHSValsDefinedFromLHS.begin(), 1398 E = RHSValsDefinedFromLHS.end(); I != E; ++I) { 1399 VNInfo *VNI = I->first; 1400 unsigned RHSValID = RHSValNoAssignments[VNI->id]; 1401 if (VNI->hasPHIKill()) 1402 NewVNInfo[RHSValID]->setHasPHIKill(true); 1403 } 1404 1405 // Clear kill flags where live ranges are extended. 1406 while (!LHSOldKills.empty()) 1407 LHSOldKills.pop_back_val()->clearRegisterKills(LHS.reg, TRI); 1408 while (!RHSOldKills.empty()) 1409 RHSOldKills.pop_back_val()->clearRegisterKills(RHS.reg, TRI); 1410 1411 if (LHSValNoAssignments.empty()) 1412 LHSValNoAssignments.push_back(-1); 1413 if (RHSValNoAssignments.empty()) 1414 RHSValNoAssignments.push_back(-1); 1415 1416 // Now erase all the redundant copies. 1417 for (unsigned i = 0, e = DeadCopies.size(); i != e; ++i) { 1418 MachineInstr *MI = DeadCopies[i]; 1419 if (!ErasedInstrs.insert(MI)) 1420 continue; 1421 DEBUG(dbgs() << "\t\terased:\t" << LIS->getInstructionIndex(MI) 1422 << '\t' << *MI); 1423 LIS->RemoveMachineInstrFromMaps(MI); 1424 MI->eraseFromParent(); 1425 } 1426 1427 SmallVector<unsigned, 8> SourceRegisters; 1428 for (SmallVector<MachineInstr*, 8>::iterator I = DupCopies.begin(), 1429 E = DupCopies.end(); I != E; ++I) { 1430 MachineInstr *MI = *I; 1431 if (!ErasedInstrs.insert(MI)) 1432 continue; 1433 1434 // If MI is a copy, then we have pretended that the assignment to B in 1435 // A = X 1436 // B = X 1437 // was actually a copy from A. Now that we decided to coalesce A and B, 1438 // transform the code into 1439 // A = X 1440 // In the case of the implicit_def, we just have to remove it. 1441 if (!MI->isImplicitDef()) { 1442 unsigned Src = MI->getOperand(1).getReg(); 1443 SourceRegisters.push_back(Src); 1444 } 1445 LIS->RemoveMachineInstrFromMaps(MI); 1446 MI->eraseFromParent(); 1447 } 1448 1449 // If B = X was the last use of X in a liverange, we have to shrink it now 1450 // that B = X is gone. 1451 for (SmallVector<unsigned, 8>::iterator I = SourceRegisters.begin(), 1452 E = SourceRegisters.end(); I != E; ++I) { 1453 LIS->shrinkToUses(&LIS->getInterval(*I)); 1454 } 1455 1456 // If we get here, we know that we can coalesce the live ranges. Ask the 1457 // intervals to coalesce themselves now. 1458 LHS.join(RHS, &LHSValNoAssignments[0], &RHSValNoAssignments[0], NewVNInfo, 1459 MRI); 1460 return true; 1461} 1462 1463namespace { 1464 // DepthMBBCompare - Comparison predicate that sort first based on the loop 1465 // depth of the basic block (the unsigned), and then on the MBB number. 1466 struct DepthMBBCompare { 1467 typedef std::pair<unsigned, MachineBasicBlock*> DepthMBBPair; 1468 bool operator()(const DepthMBBPair &LHS, const DepthMBBPair &RHS) const { 1469 // Deeper loops first 1470 if (LHS.first != RHS.first) 1471 return LHS.first > RHS.first; 1472 1473 // Prefer blocks that are more connected in the CFG. This takes care of 1474 // the most difficult copies first while intervals are short. 1475 unsigned cl = LHS.second->pred_size() + LHS.second->succ_size(); 1476 unsigned cr = RHS.second->pred_size() + RHS.second->succ_size(); 1477 if (cl != cr) 1478 return cl > cr; 1479 1480 // As a last resort, sort by block number. 1481 return LHS.second->getNumber() < RHS.second->getNumber(); 1482 } 1483 }; 1484} 1485 1486// Try joining WorkList copies starting from index From. 1487// Null out any successful joins. 1488bool RegisterCoalescer::copyCoalesceWorkList(unsigned From) { 1489 assert(From <= WorkList.size() && "Out of range"); 1490 bool Progress = false; 1491 for (unsigned i = From, e = WorkList.size(); i != e; ++i) { 1492 if (!WorkList[i]) 1493 continue; 1494 // Skip instruction pointers that have already been erased, for example by 1495 // dead code elimination. 1496 if (ErasedInstrs.erase(WorkList[i])) { 1497 WorkList[i] = 0; 1498 continue; 1499 } 1500 bool Again = false; 1501 bool Success = joinCopy(WorkList[i], Again); 1502 Progress |= Success; 1503 if (Success || !Again) 1504 WorkList[i] = 0; 1505 } 1506 return Progress; 1507} 1508 1509void 1510RegisterCoalescer::copyCoalesceInMBB(MachineBasicBlock *MBB) { 1511 DEBUG(dbgs() << MBB->getName() << ":\n"); 1512 1513 // Collect all copy-like instructions in MBB. Don't start coalescing anything 1514 // yet, it might invalidate the iterator. 1515 const unsigned PrevSize = WorkList.size(); 1516 for (MachineBasicBlock::iterator MII = MBB->begin(), E = MBB->end(); 1517 MII != E; ++MII) 1518 if (MII->isCopyLike()) 1519 WorkList.push_back(MII); 1520 1521 // Try coalescing the collected copies immediately, and remove the nulls. 1522 // This prevents the WorkList from getting too large since most copies are 1523 // joinable on the first attempt. 1524 if (copyCoalesceWorkList(PrevSize)) 1525 WorkList.erase(std::remove(WorkList.begin() + PrevSize, WorkList.end(), 1526 (MachineInstr*)0), WorkList.end()); 1527} 1528 1529void RegisterCoalescer::joinAllIntervals() { 1530 DEBUG(dbgs() << "********** JOINING INTERVALS ***********\n"); 1531 assert(WorkList.empty() && "Old data still around."); 1532 1533 if (Loops->empty()) { 1534 // If there are no loops in the function, join intervals in function order. 1535 for (MachineFunction::iterator I = MF->begin(), E = MF->end(); 1536 I != E; ++I) 1537 copyCoalesceInMBB(I); 1538 } else { 1539 // Otherwise, join intervals in inner loops before other intervals. 1540 // Unfortunately we can't just iterate over loop hierarchy here because 1541 // there may be more MBB's than BB's. Collect MBB's for sorting. 1542 1543 // Join intervals in the function prolog first. We want to join physical 1544 // registers with virtual registers before the intervals got too long. 1545 std::vector<std::pair<unsigned, MachineBasicBlock*> > MBBs; 1546 for (MachineFunction::iterator I = MF->begin(), E = MF->end();I != E;++I){ 1547 MachineBasicBlock *MBB = I; 1548 MBBs.push_back(std::make_pair(Loops->getLoopDepth(MBB), I)); 1549 } 1550 1551 // Sort by loop depth. 1552 std::sort(MBBs.begin(), MBBs.end(), DepthMBBCompare()); 1553 1554 // Finally, join intervals in loop nest order. 1555 for (unsigned i = 0, e = MBBs.size(); i != e; ++i) 1556 copyCoalesceInMBB(MBBs[i].second); 1557 } 1558 1559 // Joining intervals can allow other intervals to be joined. Iteratively join 1560 // until we make no progress. 1561 while (copyCoalesceWorkList()) 1562 /* empty */ ; 1563} 1564 1565void RegisterCoalescer::releaseMemory() { 1566 ErasedInstrs.clear(); 1567 WorkList.clear(); 1568 DeadDefs.clear(); 1569 InflateRegs.clear(); 1570} 1571 1572bool RegisterCoalescer::runOnMachineFunction(MachineFunction &fn) { 1573 MF = &fn; 1574 MRI = &fn.getRegInfo(); 1575 TM = &fn.getTarget(); 1576 TRI = TM->getRegisterInfo(); 1577 TII = TM->getInstrInfo(); 1578 LIS = &getAnalysis<LiveIntervals>(); 1579 LDV = &getAnalysis<LiveDebugVariables>(); 1580 AA = &getAnalysis<AliasAnalysis>(); 1581 Loops = &getAnalysis<MachineLoopInfo>(); 1582 1583 DEBUG(dbgs() << "********** SIMPLE REGISTER COALESCING **********\n" 1584 << "********** Function: " 1585 << ((Value*)MF->getFunction())->getName() << '\n'); 1586 1587 if (VerifyCoalescing) 1588 MF->verify(this, "Before register coalescing"); 1589 1590 RegClassInfo.runOnMachineFunction(fn); 1591 1592 // Join (coalesce) intervals if requested. 1593 if (EnableJoining) 1594 joinAllIntervals(); 1595 1596 // After deleting a lot of copies, register classes may be less constrained. 1597 // Removing sub-register operands may allow GR32_ABCD -> GR32 and DPR_VFP2 -> 1598 // DPR inflation. 1599 array_pod_sort(InflateRegs.begin(), InflateRegs.end()); 1600 InflateRegs.erase(std::unique(InflateRegs.begin(), InflateRegs.end()), 1601 InflateRegs.end()); 1602 DEBUG(dbgs() << "Trying to inflate " << InflateRegs.size() << " regs.\n"); 1603 for (unsigned i = 0, e = InflateRegs.size(); i != e; ++i) { 1604 unsigned Reg = InflateRegs[i]; 1605 if (MRI->reg_nodbg_empty(Reg)) 1606 continue; 1607 if (MRI->recomputeRegClass(Reg, *TM)) { 1608 DEBUG(dbgs() << PrintReg(Reg) << " inflated to " 1609 << MRI->getRegClass(Reg)->getName() << '\n'); 1610 ++NumInflated; 1611 } 1612 } 1613 1614 DEBUG(dump()); 1615 DEBUG(LDV->dump()); 1616 if (VerifyCoalescing) 1617 MF->verify(this, "After register coalescing"); 1618 return true; 1619} 1620 1621/// print - Implement the dump method. 1622void RegisterCoalescer::print(raw_ostream &O, const Module* m) const { 1623 LIS->print(O, m); 1624} 1625