Shih-wei Liao | e264f62 | 2010-02-10 11:10:31 -0800 | [diff] [blame] | 1 | //===-- LiveIntervalAnalysis.cpp - Live Interval Analysis -----------------===// |
| 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 LiveInterval analysis pass which is used |
| 11 | // by the Linear Scan Register allocator. This pass linearizes the |
| 12 | // basic blocks of the function in DFS order and uses the |
| 13 | // LiveVariables pass to conservatively compute live intervals for |
| 14 | // each virtual and physical register. |
| 15 | // |
| 16 | //===----------------------------------------------------------------------===// |
| 17 | |
| 18 | #define DEBUG_TYPE "liveintervals" |
| 19 | #include "llvm/CodeGen/LiveIntervalAnalysis.h" |
| 20 | #include "VirtRegMap.h" |
| 21 | #include "llvm/Value.h" |
| 22 | #include "llvm/Analysis/AliasAnalysis.h" |
| 23 | #include "llvm/CodeGen/LiveVariables.h" |
| 24 | #include "llvm/CodeGen/MachineFrameInfo.h" |
| 25 | #include "llvm/CodeGen/MachineInstr.h" |
| 26 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
| 27 | #include "llvm/CodeGen/MachineLoopInfo.h" |
| 28 | #include "llvm/CodeGen/MachineMemOperand.h" |
| 29 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
| 30 | #include "llvm/CodeGen/Passes.h" |
| 31 | #include "llvm/CodeGen/ProcessImplicitDefs.h" |
| 32 | #include "llvm/Target/TargetRegisterInfo.h" |
| 33 | #include "llvm/Target/TargetInstrInfo.h" |
| 34 | #include "llvm/Target/TargetMachine.h" |
| 35 | #include "llvm/Target/TargetOptions.h" |
| 36 | #include "llvm/Support/CommandLine.h" |
| 37 | #include "llvm/Support/Debug.h" |
| 38 | #include "llvm/Support/ErrorHandling.h" |
| 39 | #include "llvm/Support/raw_ostream.h" |
| 40 | #include "llvm/ADT/DepthFirstIterator.h" |
| 41 | #include "llvm/ADT/SmallSet.h" |
| 42 | #include "llvm/ADT/Statistic.h" |
| 43 | #include "llvm/ADT/STLExtras.h" |
| 44 | #include <algorithm> |
| 45 | #include <limits> |
| 46 | #include <cmath> |
| 47 | using namespace llvm; |
| 48 | |
| 49 | // Hidden options for help debugging. |
| 50 | static cl::opt<bool> DisableReMat("disable-rematerialization", |
| 51 | cl::init(false), cl::Hidden); |
| 52 | |
| 53 | static cl::opt<bool> EnableFastSpilling("fast-spill", |
| 54 | cl::init(false), cl::Hidden); |
| 55 | |
| 56 | STATISTIC(numIntervals , "Number of original intervals"); |
| 57 | STATISTIC(numFolds , "Number of loads/stores folded into instructions"); |
| 58 | STATISTIC(numSplits , "Number of intervals split"); |
| 59 | |
| 60 | char LiveIntervals::ID = 0; |
| 61 | static RegisterPass<LiveIntervals> X("liveintervals", "Live Interval Analysis"); |
| 62 | |
| 63 | void LiveIntervals::getAnalysisUsage(AnalysisUsage &AU) const { |
| 64 | AU.setPreservesCFG(); |
| 65 | AU.addRequired<AliasAnalysis>(); |
| 66 | AU.addPreserved<AliasAnalysis>(); |
| 67 | AU.addPreserved<LiveVariables>(); |
| 68 | AU.addRequired<LiveVariables>(); |
| 69 | AU.addPreservedID(MachineLoopInfoID); |
| 70 | AU.addPreservedID(MachineDominatorsID); |
| 71 | |
| 72 | if (!StrongPHIElim) { |
| 73 | AU.addPreservedID(PHIEliminationID); |
| 74 | AU.addRequiredID(PHIEliminationID); |
| 75 | } |
| 76 | |
| 77 | AU.addRequiredID(TwoAddressInstructionPassID); |
| 78 | AU.addPreserved<ProcessImplicitDefs>(); |
| 79 | AU.addRequired<ProcessImplicitDefs>(); |
| 80 | AU.addPreserved<SlotIndexes>(); |
| 81 | AU.addRequiredTransitive<SlotIndexes>(); |
| 82 | MachineFunctionPass::getAnalysisUsage(AU); |
| 83 | } |
| 84 | |
| 85 | void LiveIntervals::releaseMemory() { |
| 86 | // Free the live intervals themselves. |
| 87 | for (DenseMap<unsigned, LiveInterval*>::iterator I = r2iMap_.begin(), |
| 88 | E = r2iMap_.end(); I != E; ++I) |
| 89 | delete I->second; |
| 90 | |
| 91 | r2iMap_.clear(); |
| 92 | |
| 93 | // Release VNInfo memroy regions after all VNInfo objects are dtor'd. |
| 94 | VNInfoAllocator.Reset(); |
| 95 | while (!CloneMIs.empty()) { |
| 96 | MachineInstr *MI = CloneMIs.back(); |
| 97 | CloneMIs.pop_back(); |
| 98 | mf_->DeleteMachineInstr(MI); |
| 99 | } |
| 100 | } |
| 101 | |
| 102 | /// runOnMachineFunction - Register allocate the whole function |
| 103 | /// |
| 104 | bool LiveIntervals::runOnMachineFunction(MachineFunction &fn) { |
| 105 | mf_ = &fn; |
| 106 | mri_ = &mf_->getRegInfo(); |
| 107 | tm_ = &fn.getTarget(); |
| 108 | tri_ = tm_->getRegisterInfo(); |
| 109 | tii_ = tm_->getInstrInfo(); |
| 110 | aa_ = &getAnalysis<AliasAnalysis>(); |
| 111 | lv_ = &getAnalysis<LiveVariables>(); |
| 112 | indexes_ = &getAnalysis<SlotIndexes>(); |
| 113 | allocatableRegs_ = tri_->getAllocatableSet(fn); |
| 114 | |
| 115 | computeIntervals(); |
| 116 | |
| 117 | numIntervals += getNumIntervals(); |
| 118 | |
| 119 | DEBUG(dump()); |
| 120 | return true; |
| 121 | } |
| 122 | |
| 123 | /// print - Implement the dump method. |
| 124 | void LiveIntervals::print(raw_ostream &OS, const Module* ) const { |
| 125 | OS << "********** INTERVALS **********\n"; |
| 126 | for (const_iterator I = begin(), E = end(); I != E; ++I) { |
| 127 | I->second->print(OS, tri_); |
| 128 | OS << "\n"; |
| 129 | } |
| 130 | |
| 131 | printInstrs(OS); |
| 132 | } |
| 133 | |
| 134 | void LiveIntervals::printInstrs(raw_ostream &OS) const { |
| 135 | OS << "********** MACHINEINSTRS **********\n"; |
| 136 | |
| 137 | for (MachineFunction::iterator mbbi = mf_->begin(), mbbe = mf_->end(); |
| 138 | mbbi != mbbe; ++mbbi) { |
| 139 | OS << "BB#" << mbbi->getNumber() |
| 140 | << ":\t\t# derived from " << mbbi->getName() << "\n"; |
| 141 | for (MachineBasicBlock::iterator mii = mbbi->begin(), |
| 142 | mie = mbbi->end(); mii != mie; ++mii) { |
| 143 | if (mii->isDebugValue()) |
| 144 | OS << SlotIndex::getEmptyKey() << '\t' << *mii; |
| 145 | else |
| 146 | OS << getInstructionIndex(mii) << '\t' << *mii; |
| 147 | } |
| 148 | } |
| 149 | } |
| 150 | |
| 151 | void LiveIntervals::dumpInstrs() const { |
| 152 | printInstrs(dbgs()); |
| 153 | } |
| 154 | |
| 155 | bool LiveIntervals::conflictsWithPhysReg(const LiveInterval &li, |
| 156 | VirtRegMap &vrm, unsigned reg) { |
| 157 | // We don't handle fancy stuff crossing basic block boundaries |
| 158 | if (li.ranges.size() != 1) |
| 159 | return true; |
| 160 | const LiveRange &range = li.ranges.front(); |
| 161 | SlotIndex idx = range.start.getBaseIndex(); |
| 162 | SlotIndex end = range.end.getPrevSlot().getBaseIndex().getNextIndex(); |
| 163 | |
| 164 | // Skip deleted instructions |
| 165 | MachineInstr *firstMI = getInstructionFromIndex(idx); |
| 166 | while (!firstMI && idx != end) { |
| 167 | idx = idx.getNextIndex(); |
| 168 | firstMI = getInstructionFromIndex(idx); |
| 169 | } |
| 170 | if (!firstMI) |
| 171 | return false; |
| 172 | |
| 173 | // Find last instruction in range |
| 174 | SlotIndex lastIdx = end.getPrevIndex(); |
| 175 | MachineInstr *lastMI = getInstructionFromIndex(lastIdx); |
| 176 | while (!lastMI && lastIdx != idx) { |
| 177 | lastIdx = lastIdx.getPrevIndex(); |
| 178 | lastMI = getInstructionFromIndex(lastIdx); |
| 179 | } |
| 180 | if (!lastMI) |
| 181 | return false; |
| 182 | |
| 183 | // Range cannot cross basic block boundaries or terminators |
| 184 | MachineBasicBlock *MBB = firstMI->getParent(); |
| 185 | if (MBB != lastMI->getParent() || lastMI->getDesc().isTerminator()) |
| 186 | return true; |
| 187 | |
| 188 | MachineBasicBlock::const_iterator E = lastMI; |
| 189 | ++E; |
| 190 | for (MachineBasicBlock::const_iterator I = firstMI; I != E; ++I) { |
| 191 | const MachineInstr &MI = *I; |
| 192 | |
| 193 | // Allow copies to and from li.reg |
| 194 | unsigned SrcReg, DstReg, SrcSubReg, DstSubReg; |
| 195 | if (tii_->isMoveInstr(MI, SrcReg, DstReg, SrcSubReg, DstSubReg)) |
| 196 | if (SrcReg == li.reg || DstReg == li.reg) |
| 197 | continue; |
| 198 | |
| 199 | // Check for operands using reg |
| 200 | for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) { |
| 201 | const MachineOperand& mop = MI.getOperand(i); |
| 202 | if (!mop.isReg()) |
| 203 | continue; |
| 204 | unsigned PhysReg = mop.getReg(); |
| 205 | if (PhysReg == 0 || PhysReg == li.reg) |
| 206 | continue; |
| 207 | if (TargetRegisterInfo::isVirtualRegister(PhysReg)) { |
| 208 | if (!vrm.hasPhys(PhysReg)) |
| 209 | continue; |
| 210 | PhysReg = vrm.getPhys(PhysReg); |
| 211 | } |
| 212 | if (PhysReg && tri_->regsOverlap(PhysReg, reg)) |
| 213 | return true; |
| 214 | } |
| 215 | } |
| 216 | |
| 217 | // No conflicts found. |
| 218 | return false; |
| 219 | } |
| 220 | |
| 221 | /// conflictsWithPhysRegRef - Similar to conflictsWithPhysRegRef except |
| 222 | /// it can check use as well. |
| 223 | bool LiveIntervals::conflictsWithPhysRegRef(LiveInterval &li, |
| 224 | unsigned Reg, bool CheckUse, |
| 225 | SmallPtrSet<MachineInstr*,32> &JoinedCopies) { |
| 226 | for (LiveInterval::Ranges::const_iterator |
| 227 | I = li.ranges.begin(), E = li.ranges.end(); I != E; ++I) { |
| 228 | for (SlotIndex index = I->start.getBaseIndex(), |
| 229 | end = I->end.getPrevSlot().getBaseIndex().getNextIndex(); |
| 230 | index != end; |
| 231 | index = index.getNextIndex()) { |
| 232 | MachineInstr *MI = getInstructionFromIndex(index); |
| 233 | if (!MI) |
| 234 | continue; // skip deleted instructions |
| 235 | |
| 236 | if (JoinedCopies.count(MI)) |
| 237 | continue; |
| 238 | for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { |
| 239 | MachineOperand& MO = MI->getOperand(i); |
| 240 | if (!MO.isReg()) |
| 241 | continue; |
| 242 | if (MO.isUse() && !CheckUse) |
| 243 | continue; |
| 244 | unsigned PhysReg = MO.getReg(); |
| 245 | if (PhysReg == 0 || TargetRegisterInfo::isVirtualRegister(PhysReg)) |
| 246 | continue; |
| 247 | if (tri_->isSubRegister(Reg, PhysReg)) |
| 248 | return true; |
| 249 | } |
| 250 | } |
| 251 | } |
| 252 | |
| 253 | return false; |
| 254 | } |
| 255 | |
| 256 | #ifndef NDEBUG |
| 257 | static void printRegName(unsigned reg, const TargetRegisterInfo* tri_) { |
| 258 | if (TargetRegisterInfo::isPhysicalRegister(reg)) |
| 259 | dbgs() << tri_->getName(reg); |
| 260 | else |
| 261 | dbgs() << "%reg" << reg; |
| 262 | } |
| 263 | #endif |
| 264 | |
| 265 | void LiveIntervals::handleVirtualRegisterDef(MachineBasicBlock *mbb, |
| 266 | MachineBasicBlock::iterator mi, |
| 267 | SlotIndex MIIdx, |
| 268 | MachineOperand& MO, |
| 269 | unsigned MOIdx, |
| 270 | LiveInterval &interval) { |
| 271 | DEBUG({ |
| 272 | dbgs() << "\t\tregister: "; |
| 273 | printRegName(interval.reg, tri_); |
| 274 | }); |
| 275 | |
| 276 | // Virtual registers may be defined multiple times (due to phi |
| 277 | // elimination and 2-addr elimination). Much of what we do only has to be |
| 278 | // done once for the vreg. We use an empty interval to detect the first |
| 279 | // time we see a vreg. |
| 280 | LiveVariables::VarInfo& vi = lv_->getVarInfo(interval.reg); |
| 281 | if (interval.empty()) { |
| 282 | // Get the Idx of the defining instructions. |
| 283 | SlotIndex defIndex = MIIdx.getDefIndex(); |
| 284 | // Earlyclobbers move back one, so that they overlap the live range |
| 285 | // of inputs. |
| 286 | if (MO.isEarlyClobber()) |
| 287 | defIndex = MIIdx.getUseIndex(); |
| 288 | VNInfo *ValNo; |
| 289 | MachineInstr *CopyMI = NULL; |
| 290 | unsigned SrcReg, DstReg, SrcSubReg, DstSubReg; |
| 291 | if (mi->isExtractSubreg() || mi->isInsertSubreg() || mi->isSubregToReg() || |
| 292 | tii_->isMoveInstr(*mi, SrcReg, DstReg, SrcSubReg, DstSubReg)) |
| 293 | CopyMI = mi; |
| 294 | // Earlyclobbers move back one. |
| 295 | ValNo = interval.getNextValue(defIndex, CopyMI, true, VNInfoAllocator); |
| 296 | |
| 297 | assert(ValNo->id == 0 && "First value in interval is not 0?"); |
| 298 | |
| 299 | // Loop over all of the blocks that the vreg is defined in. There are |
| 300 | // two cases we have to handle here. The most common case is a vreg |
| 301 | // whose lifetime is contained within a basic block. In this case there |
| 302 | // will be a single kill, in MBB, which comes after the definition. |
| 303 | if (vi.Kills.size() == 1 && vi.Kills[0]->getParent() == mbb) { |
| 304 | // FIXME: what about dead vars? |
| 305 | SlotIndex killIdx; |
| 306 | if (vi.Kills[0] != mi) |
| 307 | killIdx = getInstructionIndex(vi.Kills[0]).getDefIndex(); |
| 308 | else |
| 309 | killIdx = defIndex.getStoreIndex(); |
| 310 | |
| 311 | // If the kill happens after the definition, we have an intra-block |
| 312 | // live range. |
| 313 | if (killIdx > defIndex) { |
| 314 | assert(vi.AliveBlocks.empty() && |
| 315 | "Shouldn't be alive across any blocks!"); |
| 316 | LiveRange LR(defIndex, killIdx, ValNo); |
| 317 | interval.addRange(LR); |
| 318 | DEBUG(dbgs() << " +" << LR << "\n"); |
| 319 | ValNo->addKill(killIdx); |
| 320 | return; |
| 321 | } |
| 322 | } |
| 323 | |
| 324 | // The other case we handle is when a virtual register lives to the end |
| 325 | // of the defining block, potentially live across some blocks, then is |
| 326 | // live into some number of blocks, but gets killed. Start by adding a |
| 327 | // range that goes from this definition to the end of the defining block. |
| 328 | LiveRange NewLR(defIndex, getMBBEndIdx(mbb), ValNo); |
| 329 | DEBUG(dbgs() << " +" << NewLR); |
| 330 | interval.addRange(NewLR); |
| 331 | |
Shih-wei Liao | e445432 | 2010-04-07 12:21:42 -0700 | [diff] [blame] | 332 | bool PHIJoin = lv_->isPHIJoin(interval.reg); |
| 333 | |
| 334 | if (PHIJoin) { |
| 335 | // A phi join register is killed at the end of the MBB and revived as a new |
| 336 | // valno in the killing blocks. |
| 337 | assert(vi.AliveBlocks.empty() && "Phi join can't pass through blocks"); |
| 338 | DEBUG(dbgs() << " phi-join"); |
| 339 | ValNo->addKill(indexes_->getTerminatorGap(mbb)); |
| 340 | ValNo->setHasPHIKill(true); |
| 341 | } else { |
| 342 | // Iterate over all of the blocks that the variable is completely |
| 343 | // live in, adding [insrtIndex(begin), instrIndex(end)+4) to the |
| 344 | // live interval. |
| 345 | for (SparseBitVector<>::iterator I = vi.AliveBlocks.begin(), |
| 346 | E = vi.AliveBlocks.end(); I != E; ++I) { |
| 347 | MachineBasicBlock *aliveBlock = mf_->getBlockNumbered(*I); |
| 348 | LiveRange LR(getMBBStartIdx(aliveBlock), getMBBEndIdx(aliveBlock), ValNo); |
| 349 | interval.addRange(LR); |
| 350 | DEBUG(dbgs() << " +" << LR); |
| 351 | } |
Shih-wei Liao | e264f62 | 2010-02-10 11:10:31 -0800 | [diff] [blame] | 352 | } |
| 353 | |
| 354 | // Finally, this virtual register is live from the start of any killing |
| 355 | // block to the 'use' slot of the killing instruction. |
| 356 | for (unsigned i = 0, e = vi.Kills.size(); i != e; ++i) { |
| 357 | MachineInstr *Kill = vi.Kills[i]; |
Shih-wei Liao | e445432 | 2010-04-07 12:21:42 -0700 | [diff] [blame] | 358 | SlotIndex Start = getMBBStartIdx(Kill->getParent()); |
| 359 | SlotIndex killIdx = getInstructionIndex(Kill).getDefIndex(); |
| 360 | |
| 361 | // Create interval with one of a NEW value number. Note that this value |
| 362 | // number isn't actually defined by an instruction, weird huh? :) |
| 363 | if (PHIJoin) { |
| 364 | ValNo = interval.getNextValue(SlotIndex(Start, true), 0, false, |
| 365 | VNInfoAllocator); |
| 366 | ValNo->setIsPHIDef(true); |
| 367 | } |
| 368 | LiveRange LR(Start, killIdx, ValNo); |
Shih-wei Liao | e264f62 | 2010-02-10 11:10:31 -0800 | [diff] [blame] | 369 | interval.addRange(LR); |
| 370 | ValNo->addKill(killIdx); |
| 371 | DEBUG(dbgs() << " +" << LR); |
| 372 | } |
| 373 | |
| 374 | } else { |
| 375 | // If this is the second time we see a virtual register definition, it |
| 376 | // must be due to phi elimination or two addr elimination. If this is |
| 377 | // the result of two address elimination, then the vreg is one of the |
| 378 | // def-and-use register operand. |
| 379 | if (mi->isRegTiedToUseOperand(MOIdx)) { |
| 380 | // If this is a two-address definition, then we have already processed |
| 381 | // the live range. The only problem is that we didn't realize there |
| 382 | // are actually two values in the live interval. Because of this we |
| 383 | // need to take the LiveRegion that defines this register and split it |
| 384 | // into two values. |
| 385 | assert(interval.containsOneValue()); |
| 386 | SlotIndex DefIndex = interval.getValNumInfo(0)->def.getDefIndex(); |
| 387 | SlotIndex RedefIndex = MIIdx.getDefIndex(); |
| 388 | if (MO.isEarlyClobber()) |
| 389 | RedefIndex = MIIdx.getUseIndex(); |
| 390 | |
| 391 | const LiveRange *OldLR = |
| 392 | interval.getLiveRangeContaining(RedefIndex.getUseIndex()); |
| 393 | VNInfo *OldValNo = OldLR->valno; |
| 394 | |
| 395 | // Delete the initial value, which should be short and continuous, |
| 396 | // because the 2-addr copy must be in the same MBB as the redef. |
| 397 | interval.removeRange(DefIndex, RedefIndex); |
| 398 | |
| 399 | // Two-address vregs should always only be redefined once. This means |
| 400 | // that at this point, there should be exactly one value number in it. |
| 401 | assert(interval.containsOneValue() && "Unexpected 2-addr liveint!"); |
| 402 | |
| 403 | // The new value number (#1) is defined by the instruction we claimed |
| 404 | // defined value #0. |
| 405 | VNInfo *ValNo = interval.getNextValue(OldValNo->def, OldValNo->getCopy(), |
| 406 | false, // update at * |
| 407 | VNInfoAllocator); |
| 408 | ValNo->setFlags(OldValNo->getFlags()); // * <- updating here |
| 409 | |
| 410 | // Value#0 is now defined by the 2-addr instruction. |
| 411 | OldValNo->def = RedefIndex; |
| 412 | OldValNo->setCopy(0); |
| 413 | |
| 414 | // Add the new live interval which replaces the range for the input copy. |
| 415 | LiveRange LR(DefIndex, RedefIndex, ValNo); |
| 416 | DEBUG(dbgs() << " replace range with " << LR); |
| 417 | interval.addRange(LR); |
| 418 | ValNo->addKill(RedefIndex); |
| 419 | |
| 420 | // If this redefinition is dead, we need to add a dummy unit live |
| 421 | // range covering the def slot. |
| 422 | if (MO.isDead()) |
| 423 | interval.addRange(LiveRange(RedefIndex, RedefIndex.getStoreIndex(), |
| 424 | OldValNo)); |
| 425 | |
| 426 | DEBUG({ |
| 427 | dbgs() << " RESULT: "; |
| 428 | interval.print(dbgs(), tri_); |
| 429 | }); |
| 430 | } else { |
Shih-wei Liao | e445432 | 2010-04-07 12:21:42 -0700 | [diff] [blame] | 431 | assert(lv_->isPHIJoin(interval.reg) && "Multiply defined register"); |
Shih-wei Liao | e264f62 | 2010-02-10 11:10:31 -0800 | [diff] [blame] | 432 | // In the case of PHI elimination, each variable definition is only |
| 433 | // live until the end of the block. We've already taken care of the |
| 434 | // rest of the live range. |
Shih-wei Liao | e445432 | 2010-04-07 12:21:42 -0700 | [diff] [blame] | 435 | |
Shih-wei Liao | e264f62 | 2010-02-10 11:10:31 -0800 | [diff] [blame] | 436 | SlotIndex defIndex = MIIdx.getDefIndex(); |
| 437 | if (MO.isEarlyClobber()) |
| 438 | defIndex = MIIdx.getUseIndex(); |
| 439 | |
| 440 | VNInfo *ValNo; |
| 441 | MachineInstr *CopyMI = NULL; |
| 442 | unsigned SrcReg, DstReg, SrcSubReg, DstSubReg; |
| 443 | if (mi->isExtractSubreg() || mi->isInsertSubreg() || mi->isSubregToReg()|| |
| 444 | tii_->isMoveInstr(*mi, SrcReg, DstReg, SrcSubReg, DstSubReg)) |
| 445 | CopyMI = mi; |
| 446 | ValNo = interval.getNextValue(defIndex, CopyMI, true, VNInfoAllocator); |
| 447 | |
| 448 | SlotIndex killIndex = getMBBEndIdx(mbb); |
| 449 | LiveRange LR(defIndex, killIndex, ValNo); |
| 450 | interval.addRange(LR); |
| 451 | ValNo->addKill(indexes_->getTerminatorGap(mbb)); |
| 452 | ValNo->setHasPHIKill(true); |
Shih-wei Liao | e445432 | 2010-04-07 12:21:42 -0700 | [diff] [blame] | 453 | DEBUG(dbgs() << " phi-join +" << LR); |
Shih-wei Liao | e264f62 | 2010-02-10 11:10:31 -0800 | [diff] [blame] | 454 | } |
| 455 | } |
| 456 | |
| 457 | DEBUG(dbgs() << '\n'); |
| 458 | } |
| 459 | |
| 460 | void LiveIntervals::handlePhysicalRegisterDef(MachineBasicBlock *MBB, |
| 461 | MachineBasicBlock::iterator mi, |
| 462 | SlotIndex MIIdx, |
| 463 | MachineOperand& MO, |
| 464 | LiveInterval &interval, |
| 465 | MachineInstr *CopyMI) { |
| 466 | // A physical register cannot be live across basic block, so its |
| 467 | // lifetime must end somewhere in its defining basic block. |
| 468 | DEBUG({ |
| 469 | dbgs() << "\t\tregister: "; |
| 470 | printRegName(interval.reg, tri_); |
| 471 | }); |
| 472 | |
| 473 | SlotIndex baseIndex = MIIdx; |
| 474 | SlotIndex start = baseIndex.getDefIndex(); |
| 475 | // Earlyclobbers move back one. |
| 476 | if (MO.isEarlyClobber()) |
| 477 | start = MIIdx.getUseIndex(); |
| 478 | SlotIndex end = start; |
| 479 | |
| 480 | // If it is not used after definition, it is considered dead at |
| 481 | // the instruction defining it. Hence its interval is: |
| 482 | // [defSlot(def), defSlot(def)+1) |
| 483 | // For earlyclobbers, the defSlot was pushed back one; the extra |
| 484 | // advance below compensates. |
| 485 | if (MO.isDead()) { |
| 486 | DEBUG(dbgs() << " dead"); |
| 487 | end = start.getStoreIndex(); |
| 488 | goto exit; |
| 489 | } |
| 490 | |
| 491 | // If it is not dead on definition, it must be killed by a |
| 492 | // subsequent instruction. Hence its interval is: |
| 493 | // [defSlot(def), useSlot(kill)+1) |
| 494 | baseIndex = baseIndex.getNextIndex(); |
| 495 | while (++mi != MBB->end()) { |
| 496 | |
| 497 | if (mi->isDebugValue()) |
| 498 | continue; |
| 499 | if (getInstructionFromIndex(baseIndex) == 0) |
| 500 | baseIndex = indexes_->getNextNonNullIndex(baseIndex); |
| 501 | |
| 502 | if (mi->killsRegister(interval.reg, tri_)) { |
| 503 | DEBUG(dbgs() << " killed"); |
| 504 | end = baseIndex.getDefIndex(); |
| 505 | goto exit; |
| 506 | } else { |
| 507 | int DefIdx = mi->findRegisterDefOperandIdx(interval.reg, false, tri_); |
| 508 | if (DefIdx != -1) { |
| 509 | if (mi->isRegTiedToUseOperand(DefIdx)) { |
| 510 | // Two-address instruction. |
| 511 | end = baseIndex.getDefIndex(); |
| 512 | } else { |
| 513 | // Another instruction redefines the register before it is ever read. |
| 514 | // Then the register is essentially dead at the instruction that |
| 515 | // defines it. Hence its interval is: |
| 516 | // [defSlot(def), defSlot(def)+1) |
| 517 | DEBUG(dbgs() << " dead"); |
| 518 | end = start.getStoreIndex(); |
| 519 | } |
| 520 | goto exit; |
| 521 | } |
| 522 | } |
| 523 | |
| 524 | baseIndex = baseIndex.getNextIndex(); |
| 525 | } |
| 526 | |
| 527 | // The only case we should have a dead physreg here without a killing or |
| 528 | // instruction where we know it's dead is if it is live-in to the function |
| 529 | // and never used. Another possible case is the implicit use of the |
| 530 | // physical register has been deleted by two-address pass. |
| 531 | end = start.getStoreIndex(); |
| 532 | |
| 533 | exit: |
| 534 | assert(start < end && "did not find end of interval?"); |
| 535 | |
| 536 | // Already exists? Extend old live interval. |
| 537 | LiveInterval::iterator OldLR = interval.FindLiveRangeContaining(start); |
| 538 | bool Extend = OldLR != interval.end(); |
| 539 | VNInfo *ValNo = Extend |
| 540 | ? OldLR->valno : interval.getNextValue(start, CopyMI, true, VNInfoAllocator); |
| 541 | if (MO.isEarlyClobber() && Extend) |
| 542 | ValNo->setHasRedefByEC(true); |
| 543 | LiveRange LR(start, end, ValNo); |
| 544 | interval.addRange(LR); |
| 545 | LR.valno->addKill(end); |
| 546 | DEBUG(dbgs() << " +" << LR << '\n'); |
| 547 | } |
| 548 | |
| 549 | void LiveIntervals::handleRegisterDef(MachineBasicBlock *MBB, |
| 550 | MachineBasicBlock::iterator MI, |
| 551 | SlotIndex MIIdx, |
| 552 | MachineOperand& MO, |
| 553 | unsigned MOIdx) { |
| 554 | if (TargetRegisterInfo::isVirtualRegister(MO.getReg())) |
| 555 | handleVirtualRegisterDef(MBB, MI, MIIdx, MO, MOIdx, |
| 556 | getOrCreateInterval(MO.getReg())); |
| 557 | else if (allocatableRegs_[MO.getReg()]) { |
| 558 | MachineInstr *CopyMI = NULL; |
| 559 | unsigned SrcReg, DstReg, SrcSubReg, DstSubReg; |
| 560 | if (MI->isExtractSubreg() || MI->isInsertSubreg() || MI->isSubregToReg() || |
| 561 | tii_->isMoveInstr(*MI, SrcReg, DstReg, SrcSubReg, DstSubReg)) |
| 562 | CopyMI = MI; |
| 563 | handlePhysicalRegisterDef(MBB, MI, MIIdx, MO, |
| 564 | getOrCreateInterval(MO.getReg()), CopyMI); |
| 565 | // Def of a register also defines its sub-registers. |
| 566 | for (const unsigned* AS = tri_->getSubRegisters(MO.getReg()); *AS; ++AS) |
| 567 | // If MI also modifies the sub-register explicitly, avoid processing it |
| 568 | // more than once. Do not pass in TRI here so it checks for exact match. |
| 569 | if (!MI->modifiesRegister(*AS)) |
| 570 | handlePhysicalRegisterDef(MBB, MI, MIIdx, MO, |
| 571 | getOrCreateInterval(*AS), 0); |
| 572 | } |
| 573 | } |
| 574 | |
| 575 | void LiveIntervals::handleLiveInRegister(MachineBasicBlock *MBB, |
| 576 | SlotIndex MIIdx, |
| 577 | LiveInterval &interval, bool isAlias) { |
| 578 | DEBUG({ |
| 579 | dbgs() << "\t\tlivein register: "; |
| 580 | printRegName(interval.reg, tri_); |
| 581 | }); |
| 582 | |
| 583 | // Look for kills, if it reaches a def before it's killed, then it shouldn't |
| 584 | // be considered a livein. |
| 585 | MachineBasicBlock::iterator mi = MBB->begin(); |
| 586 | SlotIndex baseIndex = MIIdx; |
| 587 | SlotIndex start = baseIndex; |
| 588 | if (getInstructionFromIndex(baseIndex) == 0) |
| 589 | baseIndex = indexes_->getNextNonNullIndex(baseIndex); |
| 590 | |
| 591 | SlotIndex end = baseIndex; |
| 592 | bool SeenDefUse = false; |
| 593 | |
| 594 | MachineBasicBlock::iterator E = MBB->end(); |
| 595 | while (mi != E) { |
| 596 | if (mi->isDebugValue()) { |
| 597 | ++mi; |
Shih-wei Liao | e445432 | 2010-04-07 12:21:42 -0700 | [diff] [blame] | 598 | if (mi != E && !mi->isDebugValue()) { |
| 599 | baseIndex = indexes_->getNextNonNullIndex(baseIndex); |
| 600 | } |
Shih-wei Liao | e264f62 | 2010-02-10 11:10:31 -0800 | [diff] [blame] | 601 | continue; |
| 602 | } |
| 603 | if (mi->killsRegister(interval.reg, tri_)) { |
| 604 | DEBUG(dbgs() << " killed"); |
| 605 | end = baseIndex.getDefIndex(); |
| 606 | SeenDefUse = true; |
| 607 | break; |
| 608 | } else if (mi->modifiesRegister(interval.reg, tri_)) { |
| 609 | // Another instruction redefines the register before it is ever read. |
| 610 | // Then the register is essentially dead at the instruction that defines |
| 611 | // it. Hence its interval is: |
| 612 | // [defSlot(def), defSlot(def)+1) |
| 613 | DEBUG(dbgs() << " dead"); |
| 614 | end = start.getStoreIndex(); |
| 615 | SeenDefUse = true; |
| 616 | break; |
| 617 | } |
| 618 | |
| 619 | ++mi; |
| 620 | if (mi != E && !mi->isDebugValue()) { |
| 621 | baseIndex = indexes_->getNextNonNullIndex(baseIndex); |
| 622 | } |
| 623 | } |
| 624 | |
| 625 | // Live-in register might not be used at all. |
| 626 | if (!SeenDefUse) { |
| 627 | if (isAlias) { |
| 628 | DEBUG(dbgs() << " dead"); |
| 629 | end = MIIdx.getStoreIndex(); |
| 630 | } else { |
| 631 | DEBUG(dbgs() << " live through"); |
| 632 | end = baseIndex; |
| 633 | } |
| 634 | } |
| 635 | |
| 636 | VNInfo *vni = |
| 637 | interval.getNextValue(SlotIndex(getMBBStartIdx(MBB), true), |
| 638 | 0, false, VNInfoAllocator); |
| 639 | vni->setIsPHIDef(true); |
| 640 | LiveRange LR(start, end, vni); |
| 641 | |
| 642 | interval.addRange(LR); |
| 643 | LR.valno->addKill(end); |
| 644 | DEBUG(dbgs() << " +" << LR << '\n'); |
| 645 | } |
| 646 | |
| 647 | /// computeIntervals - computes the live intervals for virtual |
| 648 | /// registers. for some ordering of the machine instructions [1,N] a |
| 649 | /// live interval is an interval [i, j) where 1 <= i <= j < N for |
| 650 | /// which a variable is live |
| 651 | void LiveIntervals::computeIntervals() { |
| 652 | DEBUG(dbgs() << "********** COMPUTING LIVE INTERVALS **********\n" |
| 653 | << "********** Function: " |
| 654 | << ((Value*)mf_->getFunction())->getName() << '\n'); |
| 655 | |
| 656 | SmallVector<unsigned, 8> UndefUses; |
| 657 | for (MachineFunction::iterator MBBI = mf_->begin(), E = mf_->end(); |
| 658 | MBBI != E; ++MBBI) { |
| 659 | MachineBasicBlock *MBB = MBBI; |
| 660 | if (MBB->empty()) |
| 661 | continue; |
| 662 | |
| 663 | // Track the index of the current machine instr. |
| 664 | SlotIndex MIIndex = getMBBStartIdx(MBB); |
| 665 | DEBUG(dbgs() << MBB->getName() << ":\n"); |
| 666 | |
| 667 | // Create intervals for live-ins to this BB first. |
| 668 | for (MachineBasicBlock::const_livein_iterator LI = MBB->livein_begin(), |
| 669 | LE = MBB->livein_end(); LI != LE; ++LI) { |
| 670 | handleLiveInRegister(MBB, MIIndex, getOrCreateInterval(*LI)); |
| 671 | // Multiple live-ins can alias the same register. |
| 672 | for (const unsigned* AS = tri_->getSubRegisters(*LI); *AS; ++AS) |
| 673 | if (!hasInterval(*AS)) |
| 674 | handleLiveInRegister(MBB, MIIndex, getOrCreateInterval(*AS), |
| 675 | true); |
| 676 | } |
| 677 | |
| 678 | // Skip over empty initial indices. |
| 679 | if (getInstructionFromIndex(MIIndex) == 0) |
| 680 | MIIndex = indexes_->getNextNonNullIndex(MIIndex); |
| 681 | |
| 682 | for (MachineBasicBlock::iterator MI = MBB->begin(), miEnd = MBB->end(); |
| 683 | MI != miEnd; ++MI) { |
| 684 | DEBUG(dbgs() << MIIndex << "\t" << *MI); |
| 685 | if (MI->isDebugValue()) |
| 686 | continue; |
| 687 | |
| 688 | // Handle defs. |
| 689 | for (int i = MI->getNumOperands() - 1; i >= 0; --i) { |
| 690 | MachineOperand &MO = MI->getOperand(i); |
| 691 | if (!MO.isReg() || !MO.getReg()) |
| 692 | continue; |
| 693 | |
| 694 | // handle register defs - build intervals |
| 695 | if (MO.isDef()) |
| 696 | handleRegisterDef(MBB, MI, MIIndex, MO, i); |
| 697 | else if (MO.isUndef()) |
| 698 | UndefUses.push_back(MO.getReg()); |
| 699 | } |
| 700 | |
| 701 | // Move to the next instr slot. |
| 702 | MIIndex = indexes_->getNextNonNullIndex(MIIndex); |
| 703 | } |
| 704 | } |
| 705 | |
| 706 | // Create empty intervals for registers defined by implicit_def's (except |
| 707 | // for those implicit_def that define values which are liveout of their |
| 708 | // blocks. |
| 709 | for (unsigned i = 0, e = UndefUses.size(); i != e; ++i) { |
| 710 | unsigned UndefReg = UndefUses[i]; |
| 711 | (void)getOrCreateInterval(UndefReg); |
| 712 | } |
| 713 | } |
| 714 | |
| 715 | LiveInterval* LiveIntervals::createInterval(unsigned reg) { |
| 716 | float Weight = TargetRegisterInfo::isPhysicalRegister(reg) ? HUGE_VALF : 0.0F; |
| 717 | return new LiveInterval(reg, Weight); |
| 718 | } |
| 719 | |
| 720 | /// dupInterval - Duplicate a live interval. The caller is responsible for |
| 721 | /// managing the allocated memory. |
| 722 | LiveInterval* LiveIntervals::dupInterval(LiveInterval *li) { |
| 723 | LiveInterval *NewLI = createInterval(li->reg); |
| 724 | NewLI->Copy(*li, mri_, getVNInfoAllocator()); |
| 725 | return NewLI; |
| 726 | } |
| 727 | |
| 728 | /// getVNInfoSourceReg - Helper function that parses the specified VNInfo |
| 729 | /// copy field and returns the source register that defines it. |
| 730 | unsigned LiveIntervals::getVNInfoSourceReg(const VNInfo *VNI) const { |
| 731 | if (!VNI->getCopy()) |
| 732 | return 0; |
| 733 | |
| 734 | if (VNI->getCopy()->isExtractSubreg()) { |
| 735 | // If it's extracting out of a physical register, return the sub-register. |
| 736 | unsigned Reg = VNI->getCopy()->getOperand(1).getReg(); |
| 737 | if (TargetRegisterInfo::isPhysicalRegister(Reg)) { |
| 738 | unsigned SrcSubReg = VNI->getCopy()->getOperand(2).getImm(); |
| 739 | unsigned DstSubReg = VNI->getCopy()->getOperand(0).getSubReg(); |
| 740 | if (SrcSubReg == DstSubReg) |
| 741 | // %reg1034:3<def> = EXTRACT_SUBREG %EDX, 3 |
| 742 | // reg1034 can still be coalesced to EDX. |
| 743 | return Reg; |
| 744 | assert(DstSubReg == 0); |
| 745 | Reg = tri_->getSubReg(Reg, VNI->getCopy()->getOperand(2).getImm()); |
| 746 | } |
| 747 | return Reg; |
| 748 | } else if (VNI->getCopy()->isInsertSubreg() || |
| 749 | VNI->getCopy()->isSubregToReg()) |
| 750 | return VNI->getCopy()->getOperand(2).getReg(); |
| 751 | |
| 752 | unsigned SrcReg, DstReg, SrcSubReg, DstSubReg; |
| 753 | if (tii_->isMoveInstr(*VNI->getCopy(), SrcReg, DstReg, SrcSubReg, DstSubReg)) |
| 754 | return SrcReg; |
| 755 | llvm_unreachable("Unrecognized copy instruction!"); |
| 756 | return 0; |
| 757 | } |
| 758 | |
| 759 | //===----------------------------------------------------------------------===// |
| 760 | // Register allocator hooks. |
| 761 | // |
| 762 | |
| 763 | /// getReMatImplicitUse - If the remat definition MI has one (for now, we only |
| 764 | /// allow one) virtual register operand, then its uses are implicitly using |
| 765 | /// the register. Returns the virtual register. |
| 766 | unsigned LiveIntervals::getReMatImplicitUse(const LiveInterval &li, |
| 767 | MachineInstr *MI) const { |
| 768 | unsigned RegOp = 0; |
| 769 | for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { |
| 770 | MachineOperand &MO = MI->getOperand(i); |
| 771 | if (!MO.isReg() || !MO.isUse()) |
| 772 | continue; |
| 773 | unsigned Reg = MO.getReg(); |
| 774 | if (Reg == 0 || Reg == li.reg) |
| 775 | continue; |
| 776 | |
| 777 | if (TargetRegisterInfo::isPhysicalRegister(Reg) && |
| 778 | !allocatableRegs_[Reg]) |
| 779 | continue; |
| 780 | // FIXME: For now, only remat MI with at most one register operand. |
| 781 | assert(!RegOp && |
| 782 | "Can't rematerialize instruction with multiple register operand!"); |
| 783 | RegOp = MO.getReg(); |
| 784 | #ifndef NDEBUG |
| 785 | break; |
| 786 | #endif |
| 787 | } |
| 788 | return RegOp; |
| 789 | } |
| 790 | |
| 791 | /// isValNoAvailableAt - Return true if the val# of the specified interval |
| 792 | /// which reaches the given instruction also reaches the specified use index. |
| 793 | bool LiveIntervals::isValNoAvailableAt(const LiveInterval &li, MachineInstr *MI, |
| 794 | SlotIndex UseIdx) const { |
| 795 | SlotIndex Index = getInstructionIndex(MI); |
| 796 | VNInfo *ValNo = li.FindLiveRangeContaining(Index)->valno; |
| 797 | LiveInterval::const_iterator UI = li.FindLiveRangeContaining(UseIdx); |
| 798 | return UI != li.end() && UI->valno == ValNo; |
| 799 | } |
| 800 | |
| 801 | /// isReMaterializable - Returns true if the definition MI of the specified |
| 802 | /// val# of the specified interval is re-materializable. |
| 803 | bool LiveIntervals::isReMaterializable(const LiveInterval &li, |
| 804 | const VNInfo *ValNo, MachineInstr *MI, |
| 805 | SmallVectorImpl<LiveInterval*> &SpillIs, |
| 806 | bool &isLoad) { |
| 807 | if (DisableReMat) |
| 808 | return false; |
| 809 | |
| 810 | if (!tii_->isTriviallyReMaterializable(MI, aa_)) |
| 811 | return false; |
| 812 | |
| 813 | // Target-specific code can mark an instruction as being rematerializable |
| 814 | // if it has one virtual reg use, though it had better be something like |
| 815 | // a PIC base register which is likely to be live everywhere. |
| 816 | unsigned ImpUse = getReMatImplicitUse(li, MI); |
| 817 | if (ImpUse) { |
| 818 | const LiveInterval &ImpLi = getInterval(ImpUse); |
| 819 | for (MachineRegisterInfo::use_iterator ri = mri_->use_begin(li.reg), |
| 820 | re = mri_->use_end(); ri != re; ++ri) { |
| 821 | MachineInstr *UseMI = &*ri; |
| 822 | SlotIndex UseIdx = getInstructionIndex(UseMI); |
| 823 | if (li.FindLiveRangeContaining(UseIdx)->valno != ValNo) |
| 824 | continue; |
| 825 | if (!isValNoAvailableAt(ImpLi, MI, UseIdx)) |
| 826 | return false; |
| 827 | } |
| 828 | |
| 829 | // If a register operand of the re-materialized instruction is going to |
| 830 | // be spilled next, then it's not legal to re-materialize this instruction. |
| 831 | for (unsigned i = 0, e = SpillIs.size(); i != e; ++i) |
| 832 | if (ImpUse == SpillIs[i]->reg) |
| 833 | return false; |
| 834 | } |
| 835 | return true; |
| 836 | } |
| 837 | |
| 838 | /// isReMaterializable - Returns true if the definition MI of the specified |
| 839 | /// val# of the specified interval is re-materializable. |
| 840 | bool LiveIntervals::isReMaterializable(const LiveInterval &li, |
| 841 | const VNInfo *ValNo, MachineInstr *MI) { |
| 842 | SmallVector<LiveInterval*, 4> Dummy1; |
| 843 | bool Dummy2; |
| 844 | return isReMaterializable(li, ValNo, MI, Dummy1, Dummy2); |
| 845 | } |
| 846 | |
| 847 | /// isReMaterializable - Returns true if every definition of MI of every |
| 848 | /// val# of the specified interval is re-materializable. |
| 849 | bool LiveIntervals::isReMaterializable(const LiveInterval &li, |
| 850 | SmallVectorImpl<LiveInterval*> &SpillIs, |
| 851 | bool &isLoad) { |
| 852 | isLoad = false; |
| 853 | for (LiveInterval::const_vni_iterator i = li.vni_begin(), e = li.vni_end(); |
| 854 | i != e; ++i) { |
| 855 | const VNInfo *VNI = *i; |
| 856 | if (VNI->isUnused()) |
| 857 | continue; // Dead val#. |
| 858 | // Is the def for the val# rematerializable? |
| 859 | if (!VNI->isDefAccurate()) |
| 860 | return false; |
| 861 | MachineInstr *ReMatDefMI = getInstructionFromIndex(VNI->def); |
| 862 | bool DefIsLoad = false; |
| 863 | if (!ReMatDefMI || |
| 864 | !isReMaterializable(li, VNI, ReMatDefMI, SpillIs, DefIsLoad)) |
| 865 | return false; |
| 866 | isLoad |= DefIsLoad; |
| 867 | } |
| 868 | return true; |
| 869 | } |
| 870 | |
| 871 | /// FilterFoldedOps - Filter out two-address use operands. Return |
| 872 | /// true if it finds any issue with the operands that ought to prevent |
| 873 | /// folding. |
| 874 | static bool FilterFoldedOps(MachineInstr *MI, |
| 875 | SmallVector<unsigned, 2> &Ops, |
| 876 | unsigned &MRInfo, |
| 877 | SmallVector<unsigned, 2> &FoldOps) { |
| 878 | MRInfo = 0; |
| 879 | for (unsigned i = 0, e = Ops.size(); i != e; ++i) { |
| 880 | unsigned OpIdx = Ops[i]; |
| 881 | MachineOperand &MO = MI->getOperand(OpIdx); |
| 882 | // FIXME: fold subreg use. |
| 883 | if (MO.getSubReg()) |
| 884 | return true; |
| 885 | if (MO.isDef()) |
| 886 | MRInfo |= (unsigned)VirtRegMap::isMod; |
| 887 | else { |
| 888 | // Filter out two-address use operand(s). |
| 889 | if (MI->isRegTiedToDefOperand(OpIdx)) { |
| 890 | MRInfo = VirtRegMap::isModRef; |
| 891 | continue; |
| 892 | } |
| 893 | MRInfo |= (unsigned)VirtRegMap::isRef; |
| 894 | } |
| 895 | FoldOps.push_back(OpIdx); |
| 896 | } |
| 897 | return false; |
| 898 | } |
| 899 | |
| 900 | |
| 901 | /// tryFoldMemoryOperand - Attempts to fold either a spill / restore from |
| 902 | /// slot / to reg or any rematerialized load into ith operand of specified |
| 903 | /// MI. If it is successul, MI is updated with the newly created MI and |
| 904 | /// returns true. |
| 905 | bool LiveIntervals::tryFoldMemoryOperand(MachineInstr* &MI, |
| 906 | VirtRegMap &vrm, MachineInstr *DefMI, |
| 907 | SlotIndex InstrIdx, |
| 908 | SmallVector<unsigned, 2> &Ops, |
| 909 | bool isSS, int Slot, unsigned Reg) { |
| 910 | // If it is an implicit def instruction, just delete it. |
| 911 | if (MI->isImplicitDef()) { |
| 912 | RemoveMachineInstrFromMaps(MI); |
| 913 | vrm.RemoveMachineInstrFromMaps(MI); |
| 914 | MI->eraseFromParent(); |
| 915 | ++numFolds; |
| 916 | return true; |
| 917 | } |
| 918 | |
| 919 | // Filter the list of operand indexes that are to be folded. Abort if |
| 920 | // any operand will prevent folding. |
| 921 | unsigned MRInfo = 0; |
| 922 | SmallVector<unsigned, 2> FoldOps; |
| 923 | if (FilterFoldedOps(MI, Ops, MRInfo, FoldOps)) |
| 924 | return false; |
| 925 | |
| 926 | // The only time it's safe to fold into a two address instruction is when |
| 927 | // it's folding reload and spill from / into a spill stack slot. |
| 928 | if (DefMI && (MRInfo & VirtRegMap::isMod)) |
| 929 | return false; |
| 930 | |
| 931 | MachineInstr *fmi = isSS ? tii_->foldMemoryOperand(*mf_, MI, FoldOps, Slot) |
| 932 | : tii_->foldMemoryOperand(*mf_, MI, FoldOps, DefMI); |
| 933 | if (fmi) { |
| 934 | // Remember this instruction uses the spill slot. |
| 935 | if (isSS) vrm.addSpillSlotUse(Slot, fmi); |
| 936 | |
| 937 | // Attempt to fold the memory reference into the instruction. If |
| 938 | // we can do this, we don't need to insert spill code. |
| 939 | MachineBasicBlock &MBB = *MI->getParent(); |
| 940 | if (isSS && !mf_->getFrameInfo()->isImmutableObjectIndex(Slot)) |
| 941 | vrm.virtFolded(Reg, MI, fmi, (VirtRegMap::ModRef)MRInfo); |
| 942 | vrm.transferSpillPts(MI, fmi); |
| 943 | vrm.transferRestorePts(MI, fmi); |
| 944 | vrm.transferEmergencySpills(MI, fmi); |
| 945 | ReplaceMachineInstrInMaps(MI, fmi); |
| 946 | MI = MBB.insert(MBB.erase(MI), fmi); |
| 947 | ++numFolds; |
| 948 | return true; |
| 949 | } |
| 950 | return false; |
| 951 | } |
| 952 | |
| 953 | /// canFoldMemoryOperand - Returns true if the specified load / store |
| 954 | /// folding is possible. |
| 955 | bool LiveIntervals::canFoldMemoryOperand(MachineInstr *MI, |
| 956 | SmallVector<unsigned, 2> &Ops, |
| 957 | bool ReMat) const { |
| 958 | // Filter the list of operand indexes that are to be folded. Abort if |
| 959 | // any operand will prevent folding. |
| 960 | unsigned MRInfo = 0; |
| 961 | SmallVector<unsigned, 2> FoldOps; |
| 962 | if (FilterFoldedOps(MI, Ops, MRInfo, FoldOps)) |
| 963 | return false; |
| 964 | |
| 965 | // It's only legal to remat for a use, not a def. |
| 966 | if (ReMat && (MRInfo & VirtRegMap::isMod)) |
| 967 | return false; |
| 968 | |
| 969 | return tii_->canFoldMemoryOperand(MI, FoldOps); |
| 970 | } |
| 971 | |
| 972 | bool LiveIntervals::intervalIsInOneMBB(const LiveInterval &li) const { |
| 973 | LiveInterval::Ranges::const_iterator itr = li.ranges.begin(); |
| 974 | |
| 975 | MachineBasicBlock *mbb = indexes_->getMBBCoveringRange(itr->start, itr->end); |
| 976 | |
| 977 | if (mbb == 0) |
| 978 | return false; |
| 979 | |
| 980 | for (++itr; itr != li.ranges.end(); ++itr) { |
| 981 | MachineBasicBlock *mbb2 = |
| 982 | indexes_->getMBBCoveringRange(itr->start, itr->end); |
| 983 | |
| 984 | if (mbb2 != mbb) |
| 985 | return false; |
| 986 | } |
| 987 | |
| 988 | return true; |
| 989 | } |
| 990 | |
| 991 | /// rewriteImplicitOps - Rewrite implicit use operands of MI (i.e. uses of |
| 992 | /// interval on to-be re-materialized operands of MI) with new register. |
| 993 | void LiveIntervals::rewriteImplicitOps(const LiveInterval &li, |
| 994 | MachineInstr *MI, unsigned NewVReg, |
| 995 | VirtRegMap &vrm) { |
| 996 | // There is an implicit use. That means one of the other operand is |
| 997 | // being remat'ed and the remat'ed instruction has li.reg as an |
| 998 | // use operand. Make sure we rewrite that as well. |
| 999 | for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { |
| 1000 | MachineOperand &MO = MI->getOperand(i); |
| 1001 | if (!MO.isReg()) |
| 1002 | continue; |
| 1003 | unsigned Reg = MO.getReg(); |
| 1004 | if (Reg == 0 || TargetRegisterInfo::isPhysicalRegister(Reg)) |
| 1005 | continue; |
| 1006 | if (!vrm.isReMaterialized(Reg)) |
| 1007 | continue; |
| 1008 | MachineInstr *ReMatMI = vrm.getReMaterializedMI(Reg); |
| 1009 | MachineOperand *UseMO = ReMatMI->findRegisterUseOperand(li.reg); |
| 1010 | if (UseMO) |
| 1011 | UseMO->setReg(NewVReg); |
| 1012 | } |
| 1013 | } |
| 1014 | |
| 1015 | /// rewriteInstructionForSpills, rewriteInstructionsForSpills - Helper functions |
| 1016 | /// for addIntervalsForSpills to rewrite uses / defs for the given live range. |
| 1017 | bool LiveIntervals:: |
| 1018 | rewriteInstructionForSpills(const LiveInterval &li, const VNInfo *VNI, |
| 1019 | bool TrySplit, SlotIndex index, SlotIndex end, |
| 1020 | MachineInstr *MI, |
| 1021 | MachineInstr *ReMatOrigDefMI, MachineInstr *ReMatDefMI, |
| 1022 | unsigned Slot, int LdSlot, |
| 1023 | bool isLoad, bool isLoadSS, bool DefIsReMat, bool CanDelete, |
| 1024 | VirtRegMap &vrm, |
| 1025 | const TargetRegisterClass* rc, |
| 1026 | SmallVector<int, 4> &ReMatIds, |
| 1027 | const MachineLoopInfo *loopInfo, |
| 1028 | unsigned &NewVReg, unsigned ImpUse, bool &HasDef, bool &HasUse, |
| 1029 | DenseMap<unsigned,unsigned> &MBBVRegsMap, |
| 1030 | std::vector<LiveInterval*> &NewLIs) { |
| 1031 | bool CanFold = false; |
| 1032 | RestartInstruction: |
| 1033 | for (unsigned i = 0; i != MI->getNumOperands(); ++i) { |
| 1034 | MachineOperand& mop = MI->getOperand(i); |
| 1035 | if (!mop.isReg()) |
| 1036 | continue; |
| 1037 | unsigned Reg = mop.getReg(); |
| 1038 | unsigned RegI = Reg; |
| 1039 | if (Reg == 0 || TargetRegisterInfo::isPhysicalRegister(Reg)) |
| 1040 | continue; |
| 1041 | if (Reg != li.reg) |
| 1042 | continue; |
| 1043 | |
| 1044 | bool TryFold = !DefIsReMat; |
| 1045 | bool FoldSS = true; // Default behavior unless it's a remat. |
| 1046 | int FoldSlot = Slot; |
| 1047 | if (DefIsReMat) { |
| 1048 | // If this is the rematerializable definition MI itself and |
| 1049 | // all of its uses are rematerialized, simply delete it. |
| 1050 | if (MI == ReMatOrigDefMI && CanDelete) { |
| 1051 | DEBUG(dbgs() << "\t\t\t\tErasing re-materializable def: " |
| 1052 | << MI << '\n'); |
| 1053 | RemoveMachineInstrFromMaps(MI); |
| 1054 | vrm.RemoveMachineInstrFromMaps(MI); |
| 1055 | MI->eraseFromParent(); |
| 1056 | break; |
| 1057 | } |
| 1058 | |
| 1059 | // If def for this use can't be rematerialized, then try folding. |
| 1060 | // If def is rematerializable and it's a load, also try folding. |
| 1061 | TryFold = !ReMatDefMI || (ReMatDefMI && (MI == ReMatOrigDefMI || isLoad)); |
| 1062 | if (isLoad) { |
| 1063 | // Try fold loads (from stack slot, constant pool, etc.) into uses. |
| 1064 | FoldSS = isLoadSS; |
| 1065 | FoldSlot = LdSlot; |
| 1066 | } |
| 1067 | } |
| 1068 | |
| 1069 | // Scan all of the operands of this instruction rewriting operands |
| 1070 | // to use NewVReg instead of li.reg as appropriate. We do this for |
| 1071 | // two reasons: |
| 1072 | // |
| 1073 | // 1. If the instr reads the same spilled vreg multiple times, we |
| 1074 | // want to reuse the NewVReg. |
| 1075 | // 2. If the instr is a two-addr instruction, we are required to |
| 1076 | // keep the src/dst regs pinned. |
| 1077 | // |
| 1078 | // Keep track of whether we replace a use and/or def so that we can |
| 1079 | // create the spill interval with the appropriate range. |
| 1080 | |
| 1081 | HasUse = mop.isUse(); |
| 1082 | HasDef = mop.isDef(); |
| 1083 | SmallVector<unsigned, 2> Ops; |
| 1084 | Ops.push_back(i); |
| 1085 | for (unsigned j = i+1, e = MI->getNumOperands(); j != e; ++j) { |
| 1086 | const MachineOperand &MOj = MI->getOperand(j); |
| 1087 | if (!MOj.isReg()) |
| 1088 | continue; |
| 1089 | unsigned RegJ = MOj.getReg(); |
| 1090 | if (RegJ == 0 || TargetRegisterInfo::isPhysicalRegister(RegJ)) |
| 1091 | continue; |
| 1092 | if (RegJ == RegI) { |
| 1093 | Ops.push_back(j); |
| 1094 | if (!MOj.isUndef()) { |
| 1095 | HasUse |= MOj.isUse(); |
| 1096 | HasDef |= MOj.isDef(); |
| 1097 | } |
| 1098 | } |
| 1099 | } |
| 1100 | |
| 1101 | // Create a new virtual register for the spill interval. |
| 1102 | // Create the new register now so we can map the fold instruction |
| 1103 | // to the new register so when it is unfolded we get the correct |
| 1104 | // answer. |
| 1105 | bool CreatedNewVReg = false; |
| 1106 | if (NewVReg == 0) { |
| 1107 | NewVReg = mri_->createVirtualRegister(rc); |
| 1108 | vrm.grow(); |
| 1109 | CreatedNewVReg = true; |
| 1110 | |
| 1111 | // The new virtual register should get the same allocation hints as the |
| 1112 | // old one. |
| 1113 | std::pair<unsigned, unsigned> Hint = mri_->getRegAllocationHint(Reg); |
| 1114 | if (Hint.first || Hint.second) |
| 1115 | mri_->setRegAllocationHint(NewVReg, Hint.first, Hint.second); |
| 1116 | } |
| 1117 | |
| 1118 | if (!TryFold) |
| 1119 | CanFold = false; |
| 1120 | else { |
| 1121 | // Do not fold load / store here if we are splitting. We'll find an |
| 1122 | // optimal point to insert a load / store later. |
| 1123 | if (!TrySplit) { |
| 1124 | if (tryFoldMemoryOperand(MI, vrm, ReMatDefMI, index, |
| 1125 | Ops, FoldSS, FoldSlot, NewVReg)) { |
| 1126 | // Folding the load/store can completely change the instruction in |
| 1127 | // unpredictable ways, rescan it from the beginning. |
| 1128 | |
| 1129 | if (FoldSS) { |
| 1130 | // We need to give the new vreg the same stack slot as the |
| 1131 | // spilled interval. |
| 1132 | vrm.assignVirt2StackSlot(NewVReg, FoldSlot); |
| 1133 | } |
| 1134 | |
| 1135 | HasUse = false; |
| 1136 | HasDef = false; |
| 1137 | CanFold = false; |
| 1138 | if (isNotInMIMap(MI)) |
| 1139 | break; |
| 1140 | goto RestartInstruction; |
| 1141 | } |
| 1142 | } else { |
| 1143 | // We'll try to fold it later if it's profitable. |
| 1144 | CanFold = canFoldMemoryOperand(MI, Ops, DefIsReMat); |
| 1145 | } |
| 1146 | } |
| 1147 | |
| 1148 | mop.setReg(NewVReg); |
| 1149 | if (mop.isImplicit()) |
| 1150 | rewriteImplicitOps(li, MI, NewVReg, vrm); |
| 1151 | |
| 1152 | // Reuse NewVReg for other reads. |
| 1153 | for (unsigned j = 0, e = Ops.size(); j != e; ++j) { |
| 1154 | MachineOperand &mopj = MI->getOperand(Ops[j]); |
| 1155 | mopj.setReg(NewVReg); |
| 1156 | if (mopj.isImplicit()) |
| 1157 | rewriteImplicitOps(li, MI, NewVReg, vrm); |
| 1158 | } |
| 1159 | |
| 1160 | if (CreatedNewVReg) { |
| 1161 | if (DefIsReMat) { |
| 1162 | vrm.setVirtIsReMaterialized(NewVReg, ReMatDefMI); |
| 1163 | if (ReMatIds[VNI->id] == VirtRegMap::MAX_STACK_SLOT) { |
| 1164 | // Each valnum may have its own remat id. |
| 1165 | ReMatIds[VNI->id] = vrm.assignVirtReMatId(NewVReg); |
| 1166 | } else { |
| 1167 | vrm.assignVirtReMatId(NewVReg, ReMatIds[VNI->id]); |
| 1168 | } |
| 1169 | if (!CanDelete || (HasUse && HasDef)) { |
| 1170 | // If this is a two-addr instruction then its use operands are |
| 1171 | // rematerializable but its def is not. It should be assigned a |
| 1172 | // stack slot. |
| 1173 | vrm.assignVirt2StackSlot(NewVReg, Slot); |
| 1174 | } |
| 1175 | } else { |
| 1176 | vrm.assignVirt2StackSlot(NewVReg, Slot); |
| 1177 | } |
| 1178 | } else if (HasUse && HasDef && |
| 1179 | vrm.getStackSlot(NewVReg) == VirtRegMap::NO_STACK_SLOT) { |
| 1180 | // If this interval hasn't been assigned a stack slot (because earlier |
| 1181 | // def is a deleted remat def), do it now. |
| 1182 | assert(Slot != VirtRegMap::NO_STACK_SLOT); |
| 1183 | vrm.assignVirt2StackSlot(NewVReg, Slot); |
| 1184 | } |
| 1185 | |
| 1186 | // Re-matting an instruction with virtual register use. Add the |
| 1187 | // register as an implicit use on the use MI. |
| 1188 | if (DefIsReMat && ImpUse) |
| 1189 | MI->addOperand(MachineOperand::CreateReg(ImpUse, false, true)); |
| 1190 | |
| 1191 | // Create a new register interval for this spill / remat. |
| 1192 | LiveInterval &nI = getOrCreateInterval(NewVReg); |
| 1193 | if (CreatedNewVReg) { |
| 1194 | NewLIs.push_back(&nI); |
| 1195 | MBBVRegsMap.insert(std::make_pair(MI->getParent()->getNumber(), NewVReg)); |
| 1196 | if (TrySplit) |
| 1197 | vrm.setIsSplitFromReg(NewVReg, li.reg); |
| 1198 | } |
| 1199 | |
| 1200 | if (HasUse) { |
| 1201 | if (CreatedNewVReg) { |
| 1202 | LiveRange LR(index.getLoadIndex(), index.getDefIndex(), |
| 1203 | nI.getNextValue(SlotIndex(), 0, false, VNInfoAllocator)); |
| 1204 | DEBUG(dbgs() << " +" << LR); |
| 1205 | nI.addRange(LR); |
| 1206 | } else { |
| 1207 | // Extend the split live interval to this def / use. |
| 1208 | SlotIndex End = index.getDefIndex(); |
| 1209 | LiveRange LR(nI.ranges[nI.ranges.size()-1].end, End, |
| 1210 | nI.getValNumInfo(nI.getNumValNums()-1)); |
| 1211 | DEBUG(dbgs() << " +" << LR); |
| 1212 | nI.addRange(LR); |
| 1213 | } |
| 1214 | } |
| 1215 | if (HasDef) { |
| 1216 | LiveRange LR(index.getDefIndex(), index.getStoreIndex(), |
| 1217 | nI.getNextValue(SlotIndex(), 0, false, VNInfoAllocator)); |
| 1218 | DEBUG(dbgs() << " +" << LR); |
| 1219 | nI.addRange(LR); |
| 1220 | } |
| 1221 | |
| 1222 | DEBUG({ |
| 1223 | dbgs() << "\t\t\t\tAdded new interval: "; |
| 1224 | nI.print(dbgs(), tri_); |
| 1225 | dbgs() << '\n'; |
| 1226 | }); |
| 1227 | } |
| 1228 | return CanFold; |
| 1229 | } |
| 1230 | bool LiveIntervals::anyKillInMBBAfterIdx(const LiveInterval &li, |
| 1231 | const VNInfo *VNI, |
| 1232 | MachineBasicBlock *MBB, |
| 1233 | SlotIndex Idx) const { |
| 1234 | SlotIndex End = getMBBEndIdx(MBB); |
| 1235 | for (unsigned j = 0, ee = VNI->kills.size(); j != ee; ++j) { |
| 1236 | if (VNI->kills[j].isPHI()) |
| 1237 | continue; |
| 1238 | |
| 1239 | SlotIndex KillIdx = VNI->kills[j]; |
| 1240 | if (KillIdx > Idx && KillIdx <= End) |
| 1241 | return true; |
| 1242 | } |
| 1243 | return false; |
| 1244 | } |
| 1245 | |
| 1246 | /// RewriteInfo - Keep track of machine instrs that will be rewritten |
| 1247 | /// during spilling. |
| 1248 | namespace { |
| 1249 | struct RewriteInfo { |
| 1250 | SlotIndex Index; |
| 1251 | MachineInstr *MI; |
| 1252 | bool HasUse; |
| 1253 | bool HasDef; |
| 1254 | RewriteInfo(SlotIndex i, MachineInstr *mi, bool u, bool d) |
| 1255 | : Index(i), MI(mi), HasUse(u), HasDef(d) {} |
| 1256 | }; |
| 1257 | |
| 1258 | struct RewriteInfoCompare { |
| 1259 | bool operator()(const RewriteInfo &LHS, const RewriteInfo &RHS) const { |
| 1260 | return LHS.Index < RHS.Index; |
| 1261 | } |
| 1262 | }; |
| 1263 | } |
| 1264 | |
| 1265 | void LiveIntervals:: |
| 1266 | rewriteInstructionsForSpills(const LiveInterval &li, bool TrySplit, |
| 1267 | LiveInterval::Ranges::const_iterator &I, |
| 1268 | MachineInstr *ReMatOrigDefMI, MachineInstr *ReMatDefMI, |
| 1269 | unsigned Slot, int LdSlot, |
| 1270 | bool isLoad, bool isLoadSS, bool DefIsReMat, bool CanDelete, |
| 1271 | VirtRegMap &vrm, |
| 1272 | const TargetRegisterClass* rc, |
| 1273 | SmallVector<int, 4> &ReMatIds, |
| 1274 | const MachineLoopInfo *loopInfo, |
| 1275 | BitVector &SpillMBBs, |
| 1276 | DenseMap<unsigned, std::vector<SRInfo> > &SpillIdxes, |
| 1277 | BitVector &RestoreMBBs, |
| 1278 | DenseMap<unsigned, std::vector<SRInfo> > &RestoreIdxes, |
| 1279 | DenseMap<unsigned,unsigned> &MBBVRegsMap, |
| 1280 | std::vector<LiveInterval*> &NewLIs) { |
| 1281 | bool AllCanFold = true; |
| 1282 | unsigned NewVReg = 0; |
| 1283 | SlotIndex start = I->start.getBaseIndex(); |
| 1284 | SlotIndex end = I->end.getPrevSlot().getBaseIndex().getNextIndex(); |
| 1285 | |
| 1286 | // First collect all the def / use in this live range that will be rewritten. |
| 1287 | // Make sure they are sorted according to instruction index. |
| 1288 | std::vector<RewriteInfo> RewriteMIs; |
| 1289 | for (MachineRegisterInfo::reg_iterator ri = mri_->reg_begin(li.reg), |
| 1290 | re = mri_->reg_end(); ri != re; ) { |
| 1291 | MachineInstr *MI = &*ri; |
| 1292 | MachineOperand &O = ri.getOperand(); |
| 1293 | ++ri; |
| 1294 | if (MI->isDebugValue()) { |
| 1295 | // Remove debug info for now. |
| 1296 | O.setReg(0U); |
| 1297 | DEBUG(dbgs() << "Removing debug info due to spill:" << "\t" << *MI); |
| 1298 | continue; |
| 1299 | } |
| 1300 | assert(!O.isImplicit() && "Spilling register that's used as implicit use?"); |
| 1301 | SlotIndex index = getInstructionIndex(MI); |
| 1302 | if (index < start || index >= end) |
| 1303 | continue; |
| 1304 | |
| 1305 | if (O.isUndef()) |
| 1306 | // Must be defined by an implicit def. It should not be spilled. Note, |
| 1307 | // this is for correctness reason. e.g. |
| 1308 | // 8 %reg1024<def> = IMPLICIT_DEF |
| 1309 | // 12 %reg1024<def> = INSERT_SUBREG %reg1024<kill>, %reg1025, 2 |
| 1310 | // The live range [12, 14) are not part of the r1024 live interval since |
| 1311 | // it's defined by an implicit def. It will not conflicts with live |
| 1312 | // interval of r1025. Now suppose both registers are spilled, you can |
| 1313 | // easily see a situation where both registers are reloaded before |
| 1314 | // the INSERT_SUBREG and both target registers that would overlap. |
| 1315 | continue; |
| 1316 | RewriteMIs.push_back(RewriteInfo(index, MI, O.isUse(), O.isDef())); |
| 1317 | } |
| 1318 | std::sort(RewriteMIs.begin(), RewriteMIs.end(), RewriteInfoCompare()); |
| 1319 | |
| 1320 | unsigned ImpUse = DefIsReMat ? getReMatImplicitUse(li, ReMatDefMI) : 0; |
| 1321 | // Now rewrite the defs and uses. |
| 1322 | for (unsigned i = 0, e = RewriteMIs.size(); i != e; ) { |
| 1323 | RewriteInfo &rwi = RewriteMIs[i]; |
| 1324 | ++i; |
| 1325 | SlotIndex index = rwi.Index; |
| 1326 | bool MIHasUse = rwi.HasUse; |
| 1327 | bool MIHasDef = rwi.HasDef; |
| 1328 | MachineInstr *MI = rwi.MI; |
| 1329 | // If MI def and/or use the same register multiple times, then there |
| 1330 | // are multiple entries. |
| 1331 | unsigned NumUses = MIHasUse; |
| 1332 | while (i != e && RewriteMIs[i].MI == MI) { |
| 1333 | assert(RewriteMIs[i].Index == index); |
| 1334 | bool isUse = RewriteMIs[i].HasUse; |
| 1335 | if (isUse) ++NumUses; |
| 1336 | MIHasUse |= isUse; |
| 1337 | MIHasDef |= RewriteMIs[i].HasDef; |
| 1338 | ++i; |
| 1339 | } |
| 1340 | MachineBasicBlock *MBB = MI->getParent(); |
| 1341 | |
| 1342 | if (ImpUse && MI != ReMatDefMI) { |
Shih-wei Liao | e445432 | 2010-04-07 12:21:42 -0700 | [diff] [blame] | 1343 | // Re-matting an instruction with virtual register use. Prevent interval |
| 1344 | // from being spilled. |
| 1345 | getInterval(ImpUse).markNotSpillable(); |
Shih-wei Liao | e264f62 | 2010-02-10 11:10:31 -0800 | [diff] [blame] | 1346 | } |
| 1347 | |
| 1348 | unsigned MBBId = MBB->getNumber(); |
| 1349 | unsigned ThisVReg = 0; |
| 1350 | if (TrySplit) { |
| 1351 | DenseMap<unsigned,unsigned>::iterator NVI = MBBVRegsMap.find(MBBId); |
| 1352 | if (NVI != MBBVRegsMap.end()) { |
| 1353 | ThisVReg = NVI->second; |
| 1354 | // One common case: |
| 1355 | // x = use |
| 1356 | // ... |
| 1357 | // ... |
| 1358 | // def = ... |
| 1359 | // = use |
| 1360 | // It's better to start a new interval to avoid artifically |
| 1361 | // extend the new interval. |
| 1362 | if (MIHasDef && !MIHasUse) { |
| 1363 | MBBVRegsMap.erase(MBB->getNumber()); |
| 1364 | ThisVReg = 0; |
| 1365 | } |
| 1366 | } |
| 1367 | } |
| 1368 | |
| 1369 | bool IsNew = ThisVReg == 0; |
| 1370 | if (IsNew) { |
| 1371 | // This ends the previous live interval. If all of its def / use |
| 1372 | // can be folded, give it a low spill weight. |
| 1373 | if (NewVReg && TrySplit && AllCanFold) { |
| 1374 | LiveInterval &nI = getOrCreateInterval(NewVReg); |
| 1375 | nI.weight /= 10.0F; |
| 1376 | } |
| 1377 | AllCanFold = true; |
| 1378 | } |
| 1379 | NewVReg = ThisVReg; |
| 1380 | |
| 1381 | bool HasDef = false; |
| 1382 | bool HasUse = false; |
| 1383 | bool CanFold = rewriteInstructionForSpills(li, I->valno, TrySplit, |
| 1384 | index, end, MI, ReMatOrigDefMI, ReMatDefMI, |
| 1385 | Slot, LdSlot, isLoad, isLoadSS, DefIsReMat, |
| 1386 | CanDelete, vrm, rc, ReMatIds, loopInfo, NewVReg, |
| 1387 | ImpUse, HasDef, HasUse, MBBVRegsMap, NewLIs); |
| 1388 | if (!HasDef && !HasUse) |
| 1389 | continue; |
| 1390 | |
| 1391 | AllCanFold &= CanFold; |
| 1392 | |
| 1393 | // Update weight of spill interval. |
| 1394 | LiveInterval &nI = getOrCreateInterval(NewVReg); |
| 1395 | if (!TrySplit) { |
| 1396 | // The spill weight is now infinity as it cannot be spilled again. |
Shih-wei Liao | e445432 | 2010-04-07 12:21:42 -0700 | [diff] [blame] | 1397 | nI.markNotSpillable(); |
Shih-wei Liao | e264f62 | 2010-02-10 11:10:31 -0800 | [diff] [blame] | 1398 | continue; |
| 1399 | } |
| 1400 | |
| 1401 | // Keep track of the last def and first use in each MBB. |
| 1402 | if (HasDef) { |
| 1403 | if (MI != ReMatOrigDefMI || !CanDelete) { |
| 1404 | bool HasKill = false; |
| 1405 | if (!HasUse) |
| 1406 | HasKill = anyKillInMBBAfterIdx(li, I->valno, MBB, index.getDefIndex()); |
| 1407 | else { |
| 1408 | // If this is a two-address code, then this index starts a new VNInfo. |
| 1409 | const VNInfo *VNI = li.findDefinedVNInfoForRegInt(index.getDefIndex()); |
| 1410 | if (VNI) |
| 1411 | HasKill = anyKillInMBBAfterIdx(li, VNI, MBB, index.getDefIndex()); |
| 1412 | } |
| 1413 | DenseMap<unsigned, std::vector<SRInfo> >::iterator SII = |
| 1414 | SpillIdxes.find(MBBId); |
| 1415 | if (!HasKill) { |
| 1416 | if (SII == SpillIdxes.end()) { |
| 1417 | std::vector<SRInfo> S; |
| 1418 | S.push_back(SRInfo(index, NewVReg, true)); |
| 1419 | SpillIdxes.insert(std::make_pair(MBBId, S)); |
| 1420 | } else if (SII->second.back().vreg != NewVReg) { |
| 1421 | SII->second.push_back(SRInfo(index, NewVReg, true)); |
| 1422 | } else if (index > SII->second.back().index) { |
| 1423 | // If there is an earlier def and this is a two-address |
| 1424 | // instruction, then it's not possible to fold the store (which |
| 1425 | // would also fold the load). |
| 1426 | SRInfo &Info = SII->second.back(); |
| 1427 | Info.index = index; |
| 1428 | Info.canFold = !HasUse; |
| 1429 | } |
| 1430 | SpillMBBs.set(MBBId); |
| 1431 | } else if (SII != SpillIdxes.end() && |
| 1432 | SII->second.back().vreg == NewVReg && |
| 1433 | index > SII->second.back().index) { |
| 1434 | // There is an earlier def that's not killed (must be two-address). |
| 1435 | // The spill is no longer needed. |
| 1436 | SII->second.pop_back(); |
| 1437 | if (SII->second.empty()) { |
| 1438 | SpillIdxes.erase(MBBId); |
| 1439 | SpillMBBs.reset(MBBId); |
| 1440 | } |
| 1441 | } |
| 1442 | } |
| 1443 | } |
| 1444 | |
| 1445 | if (HasUse) { |
| 1446 | DenseMap<unsigned, std::vector<SRInfo> >::iterator SII = |
| 1447 | SpillIdxes.find(MBBId); |
| 1448 | if (SII != SpillIdxes.end() && |
| 1449 | SII->second.back().vreg == NewVReg && |
| 1450 | index > SII->second.back().index) |
| 1451 | // Use(s) following the last def, it's not safe to fold the spill. |
| 1452 | SII->second.back().canFold = false; |
| 1453 | DenseMap<unsigned, std::vector<SRInfo> >::iterator RII = |
| 1454 | RestoreIdxes.find(MBBId); |
| 1455 | if (RII != RestoreIdxes.end() && RII->second.back().vreg == NewVReg) |
| 1456 | // If we are splitting live intervals, only fold if it's the first |
| 1457 | // use and there isn't another use later in the MBB. |
| 1458 | RII->second.back().canFold = false; |
| 1459 | else if (IsNew) { |
| 1460 | // Only need a reload if there isn't an earlier def / use. |
| 1461 | if (RII == RestoreIdxes.end()) { |
| 1462 | std::vector<SRInfo> Infos; |
| 1463 | Infos.push_back(SRInfo(index, NewVReg, true)); |
| 1464 | RestoreIdxes.insert(std::make_pair(MBBId, Infos)); |
| 1465 | } else { |
| 1466 | RII->second.push_back(SRInfo(index, NewVReg, true)); |
| 1467 | } |
| 1468 | RestoreMBBs.set(MBBId); |
| 1469 | } |
| 1470 | } |
| 1471 | |
| 1472 | // Update spill weight. |
| 1473 | unsigned loopDepth = loopInfo->getLoopDepth(MBB); |
| 1474 | nI.weight += getSpillWeight(HasDef, HasUse, loopDepth); |
| 1475 | } |
| 1476 | |
| 1477 | if (NewVReg && TrySplit && AllCanFold) { |
| 1478 | // If all of its def / use can be folded, give it a low spill weight. |
| 1479 | LiveInterval &nI = getOrCreateInterval(NewVReg); |
| 1480 | nI.weight /= 10.0F; |
| 1481 | } |
| 1482 | } |
| 1483 | |
| 1484 | bool LiveIntervals::alsoFoldARestore(int Id, SlotIndex index, |
| 1485 | unsigned vr, BitVector &RestoreMBBs, |
| 1486 | DenseMap<unsigned,std::vector<SRInfo> > &RestoreIdxes) { |
| 1487 | if (!RestoreMBBs[Id]) |
| 1488 | return false; |
| 1489 | std::vector<SRInfo> &Restores = RestoreIdxes[Id]; |
| 1490 | for (unsigned i = 0, e = Restores.size(); i != e; ++i) |
| 1491 | if (Restores[i].index == index && |
| 1492 | Restores[i].vreg == vr && |
| 1493 | Restores[i].canFold) |
| 1494 | return true; |
| 1495 | return false; |
| 1496 | } |
| 1497 | |
| 1498 | void LiveIntervals::eraseRestoreInfo(int Id, SlotIndex index, |
| 1499 | unsigned vr, BitVector &RestoreMBBs, |
| 1500 | DenseMap<unsigned,std::vector<SRInfo> > &RestoreIdxes) { |
| 1501 | if (!RestoreMBBs[Id]) |
| 1502 | return; |
| 1503 | std::vector<SRInfo> &Restores = RestoreIdxes[Id]; |
| 1504 | for (unsigned i = 0, e = Restores.size(); i != e; ++i) |
| 1505 | if (Restores[i].index == index && Restores[i].vreg) |
| 1506 | Restores[i].index = SlotIndex(); |
| 1507 | } |
| 1508 | |
| 1509 | /// handleSpilledImpDefs - Remove IMPLICIT_DEF instructions which are being |
| 1510 | /// spilled and create empty intervals for their uses. |
| 1511 | void |
| 1512 | LiveIntervals::handleSpilledImpDefs(const LiveInterval &li, VirtRegMap &vrm, |
| 1513 | const TargetRegisterClass* rc, |
| 1514 | std::vector<LiveInterval*> &NewLIs) { |
| 1515 | for (MachineRegisterInfo::reg_iterator ri = mri_->reg_begin(li.reg), |
| 1516 | re = mri_->reg_end(); ri != re; ) { |
| 1517 | MachineOperand &O = ri.getOperand(); |
| 1518 | MachineInstr *MI = &*ri; |
| 1519 | ++ri; |
| 1520 | if (O.isDef()) { |
| 1521 | assert(MI->isImplicitDef() && |
| 1522 | "Register def was not rewritten?"); |
| 1523 | RemoveMachineInstrFromMaps(MI); |
| 1524 | vrm.RemoveMachineInstrFromMaps(MI); |
| 1525 | MI->eraseFromParent(); |
| 1526 | } else { |
| 1527 | // This must be an use of an implicit_def so it's not part of the live |
| 1528 | // interval. Create a new empty live interval for it. |
| 1529 | // FIXME: Can we simply erase some of the instructions? e.g. Stores? |
| 1530 | unsigned NewVReg = mri_->createVirtualRegister(rc); |
| 1531 | vrm.grow(); |
| 1532 | vrm.setIsImplicitlyDefined(NewVReg); |
| 1533 | NewLIs.push_back(&getOrCreateInterval(NewVReg)); |
| 1534 | for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { |
| 1535 | MachineOperand &MO = MI->getOperand(i); |
| 1536 | if (MO.isReg() && MO.getReg() == li.reg) { |
| 1537 | MO.setReg(NewVReg); |
| 1538 | MO.setIsUndef(); |
| 1539 | } |
| 1540 | } |
| 1541 | } |
| 1542 | } |
| 1543 | } |
| 1544 | |
Shih-wei Liao | e445432 | 2010-04-07 12:21:42 -0700 | [diff] [blame] | 1545 | float |
| 1546 | LiveIntervals::getSpillWeight(bool isDef, bool isUse, unsigned loopDepth) { |
| 1547 | // Limit the loop depth ridiculousness. |
| 1548 | if (loopDepth > 200) |
| 1549 | loopDepth = 200; |
| 1550 | |
| 1551 | // The loop depth is used to roughly estimate the number of times the |
| 1552 | // instruction is executed. Something like 10^d is simple, but will quickly |
| 1553 | // overflow a float. This expression behaves like 10^d for small d, but is |
| 1554 | // more tempered for large d. At d=200 we get 6.7e33 which leaves a bit of |
| 1555 | // headroom before overflow. |
| 1556 | float lc = powf(1 + (100.0f / (loopDepth+10)), (float)loopDepth); |
| 1557 | |
| 1558 | return (isDef + isUse) * lc; |
| 1559 | } |
| 1560 | |
| 1561 | void |
| 1562 | LiveIntervals::normalizeSpillWeights(std::vector<LiveInterval*> &NewLIs) { |
| 1563 | for (unsigned i = 0, e = NewLIs.size(); i != e; ++i) |
| 1564 | normalizeSpillWeight(*NewLIs[i]); |
| 1565 | } |
| 1566 | |
Shih-wei Liao | e264f62 | 2010-02-10 11:10:31 -0800 | [diff] [blame] | 1567 | std::vector<LiveInterval*> LiveIntervals:: |
| 1568 | addIntervalsForSpillsFast(const LiveInterval &li, |
| 1569 | const MachineLoopInfo *loopInfo, |
| 1570 | VirtRegMap &vrm) { |
| 1571 | unsigned slot = vrm.assignVirt2StackSlot(li.reg); |
| 1572 | |
| 1573 | std::vector<LiveInterval*> added; |
| 1574 | |
Shih-wei Liao | e445432 | 2010-04-07 12:21:42 -0700 | [diff] [blame] | 1575 | assert(li.isSpillable() && "attempt to spill already spilled interval!"); |
Shih-wei Liao | e264f62 | 2010-02-10 11:10:31 -0800 | [diff] [blame] | 1576 | |
| 1577 | DEBUG({ |
| 1578 | dbgs() << "\t\t\t\tadding intervals for spills for interval: "; |
| 1579 | li.dump(); |
| 1580 | dbgs() << '\n'; |
| 1581 | }); |
| 1582 | |
| 1583 | const TargetRegisterClass* rc = mri_->getRegClass(li.reg); |
| 1584 | |
| 1585 | MachineRegisterInfo::reg_iterator RI = mri_->reg_begin(li.reg); |
| 1586 | while (RI != mri_->reg_end()) { |
| 1587 | MachineInstr* MI = &*RI; |
| 1588 | |
| 1589 | SmallVector<unsigned, 2> Indices; |
| 1590 | bool HasUse = false; |
| 1591 | bool HasDef = false; |
| 1592 | |
| 1593 | for (unsigned i = 0; i != MI->getNumOperands(); ++i) { |
| 1594 | MachineOperand& mop = MI->getOperand(i); |
| 1595 | if (!mop.isReg() || mop.getReg() != li.reg) continue; |
| 1596 | |
| 1597 | HasUse |= MI->getOperand(i).isUse(); |
| 1598 | HasDef |= MI->getOperand(i).isDef(); |
| 1599 | |
| 1600 | Indices.push_back(i); |
| 1601 | } |
| 1602 | |
| 1603 | if (!tryFoldMemoryOperand(MI, vrm, NULL, getInstructionIndex(MI), |
| 1604 | Indices, true, slot, li.reg)) { |
| 1605 | unsigned NewVReg = mri_->createVirtualRegister(rc); |
| 1606 | vrm.grow(); |
| 1607 | vrm.assignVirt2StackSlot(NewVReg, slot); |
| 1608 | |
| 1609 | // create a new register for this spill |
| 1610 | LiveInterval &nI = getOrCreateInterval(NewVReg); |
Shih-wei Liao | e445432 | 2010-04-07 12:21:42 -0700 | [diff] [blame] | 1611 | nI.markNotSpillable(); |
Shih-wei Liao | e264f62 | 2010-02-10 11:10:31 -0800 | [diff] [blame] | 1612 | |
| 1613 | // Rewrite register operands to use the new vreg. |
| 1614 | for (SmallVectorImpl<unsigned>::iterator I = Indices.begin(), |
| 1615 | E = Indices.end(); I != E; ++I) { |
| 1616 | MI->getOperand(*I).setReg(NewVReg); |
| 1617 | |
| 1618 | if (MI->getOperand(*I).isUse()) |
| 1619 | MI->getOperand(*I).setIsKill(true); |
| 1620 | } |
| 1621 | |
| 1622 | // Fill in the new live interval. |
| 1623 | SlotIndex index = getInstructionIndex(MI); |
| 1624 | if (HasUse) { |
| 1625 | LiveRange LR(index.getLoadIndex(), index.getUseIndex(), |
| 1626 | nI.getNextValue(SlotIndex(), 0, false, |
| 1627 | getVNInfoAllocator())); |
| 1628 | DEBUG(dbgs() << " +" << LR); |
| 1629 | nI.addRange(LR); |
| 1630 | vrm.addRestorePoint(NewVReg, MI); |
| 1631 | } |
| 1632 | if (HasDef) { |
| 1633 | LiveRange LR(index.getDefIndex(), index.getStoreIndex(), |
| 1634 | nI.getNextValue(SlotIndex(), 0, false, |
| 1635 | getVNInfoAllocator())); |
| 1636 | DEBUG(dbgs() << " +" << LR); |
| 1637 | nI.addRange(LR); |
| 1638 | vrm.addSpillPoint(NewVReg, true, MI); |
| 1639 | } |
| 1640 | |
| 1641 | added.push_back(&nI); |
| 1642 | |
| 1643 | DEBUG({ |
| 1644 | dbgs() << "\t\t\t\tadded new interval: "; |
| 1645 | nI.dump(); |
| 1646 | dbgs() << '\n'; |
| 1647 | }); |
| 1648 | } |
| 1649 | |
| 1650 | |
| 1651 | RI = mri_->reg_begin(li.reg); |
| 1652 | } |
| 1653 | |
| 1654 | return added; |
| 1655 | } |
| 1656 | |
| 1657 | std::vector<LiveInterval*> LiveIntervals:: |
| 1658 | addIntervalsForSpills(const LiveInterval &li, |
| 1659 | SmallVectorImpl<LiveInterval*> &SpillIs, |
| 1660 | const MachineLoopInfo *loopInfo, VirtRegMap &vrm) { |
| 1661 | |
| 1662 | if (EnableFastSpilling) |
| 1663 | return addIntervalsForSpillsFast(li, loopInfo, vrm); |
| 1664 | |
Shih-wei Liao | e445432 | 2010-04-07 12:21:42 -0700 | [diff] [blame] | 1665 | assert(li.isSpillable() && "attempt to spill already spilled interval!"); |
Shih-wei Liao | e264f62 | 2010-02-10 11:10:31 -0800 | [diff] [blame] | 1666 | |
| 1667 | DEBUG({ |
| 1668 | dbgs() << "\t\t\t\tadding intervals for spills for interval: "; |
| 1669 | li.print(dbgs(), tri_); |
| 1670 | dbgs() << '\n'; |
| 1671 | }); |
| 1672 | |
| 1673 | // Each bit specify whether a spill is required in the MBB. |
| 1674 | BitVector SpillMBBs(mf_->getNumBlockIDs()); |
| 1675 | DenseMap<unsigned, std::vector<SRInfo> > SpillIdxes; |
| 1676 | BitVector RestoreMBBs(mf_->getNumBlockIDs()); |
| 1677 | DenseMap<unsigned, std::vector<SRInfo> > RestoreIdxes; |
| 1678 | DenseMap<unsigned,unsigned> MBBVRegsMap; |
| 1679 | std::vector<LiveInterval*> NewLIs; |
| 1680 | const TargetRegisterClass* rc = mri_->getRegClass(li.reg); |
| 1681 | |
| 1682 | unsigned NumValNums = li.getNumValNums(); |
| 1683 | SmallVector<MachineInstr*, 4> ReMatDefs; |
| 1684 | ReMatDefs.resize(NumValNums, NULL); |
| 1685 | SmallVector<MachineInstr*, 4> ReMatOrigDefs; |
| 1686 | ReMatOrigDefs.resize(NumValNums, NULL); |
| 1687 | SmallVector<int, 4> ReMatIds; |
| 1688 | ReMatIds.resize(NumValNums, VirtRegMap::MAX_STACK_SLOT); |
| 1689 | BitVector ReMatDelete(NumValNums); |
| 1690 | unsigned Slot = VirtRegMap::MAX_STACK_SLOT; |
| 1691 | |
| 1692 | // Spilling a split live interval. It cannot be split any further. Also, |
| 1693 | // it's also guaranteed to be a single val# / range interval. |
| 1694 | if (vrm.getPreSplitReg(li.reg)) { |
| 1695 | vrm.setIsSplitFromReg(li.reg, 0); |
| 1696 | // Unset the split kill marker on the last use. |
| 1697 | SlotIndex KillIdx = vrm.getKillPoint(li.reg); |
| 1698 | if (KillIdx != SlotIndex()) { |
| 1699 | MachineInstr *KillMI = getInstructionFromIndex(KillIdx); |
| 1700 | assert(KillMI && "Last use disappeared?"); |
| 1701 | int KillOp = KillMI->findRegisterUseOperandIdx(li.reg, true); |
| 1702 | assert(KillOp != -1 && "Last use disappeared?"); |
| 1703 | KillMI->getOperand(KillOp).setIsKill(false); |
| 1704 | } |
| 1705 | vrm.removeKillPoint(li.reg); |
| 1706 | bool DefIsReMat = vrm.isReMaterialized(li.reg); |
| 1707 | Slot = vrm.getStackSlot(li.reg); |
| 1708 | assert(Slot != VirtRegMap::MAX_STACK_SLOT); |
| 1709 | MachineInstr *ReMatDefMI = DefIsReMat ? |
| 1710 | vrm.getReMaterializedMI(li.reg) : NULL; |
| 1711 | int LdSlot = 0; |
| 1712 | bool isLoadSS = DefIsReMat && tii_->isLoadFromStackSlot(ReMatDefMI, LdSlot); |
| 1713 | bool isLoad = isLoadSS || |
| 1714 | (DefIsReMat && (ReMatDefMI->getDesc().canFoldAsLoad())); |
| 1715 | bool IsFirstRange = true; |
| 1716 | for (LiveInterval::Ranges::const_iterator |
| 1717 | I = li.ranges.begin(), E = li.ranges.end(); I != E; ++I) { |
| 1718 | // If this is a split live interval with multiple ranges, it means there |
| 1719 | // are two-address instructions that re-defined the value. Only the |
| 1720 | // first def can be rematerialized! |
| 1721 | if (IsFirstRange) { |
| 1722 | // Note ReMatOrigDefMI has already been deleted. |
| 1723 | rewriteInstructionsForSpills(li, false, I, NULL, ReMatDefMI, |
| 1724 | Slot, LdSlot, isLoad, isLoadSS, DefIsReMat, |
| 1725 | false, vrm, rc, ReMatIds, loopInfo, |
| 1726 | SpillMBBs, SpillIdxes, RestoreMBBs, RestoreIdxes, |
| 1727 | MBBVRegsMap, NewLIs); |
| 1728 | } else { |
| 1729 | rewriteInstructionsForSpills(li, false, I, NULL, 0, |
| 1730 | Slot, 0, false, false, false, |
| 1731 | false, vrm, rc, ReMatIds, loopInfo, |
| 1732 | SpillMBBs, SpillIdxes, RestoreMBBs, RestoreIdxes, |
| 1733 | MBBVRegsMap, NewLIs); |
| 1734 | } |
| 1735 | IsFirstRange = false; |
| 1736 | } |
| 1737 | |
| 1738 | handleSpilledImpDefs(li, vrm, rc, NewLIs); |
Shih-wei Liao | e445432 | 2010-04-07 12:21:42 -0700 | [diff] [blame] | 1739 | normalizeSpillWeights(NewLIs); |
Shih-wei Liao | e264f62 | 2010-02-10 11:10:31 -0800 | [diff] [blame] | 1740 | return NewLIs; |
| 1741 | } |
| 1742 | |
| 1743 | bool TrySplit = !intervalIsInOneMBB(li); |
| 1744 | if (TrySplit) |
| 1745 | ++numSplits; |
| 1746 | bool NeedStackSlot = false; |
| 1747 | for (LiveInterval::const_vni_iterator i = li.vni_begin(), e = li.vni_end(); |
| 1748 | i != e; ++i) { |
| 1749 | const VNInfo *VNI = *i; |
| 1750 | unsigned VN = VNI->id; |
| 1751 | if (VNI->isUnused()) |
| 1752 | continue; // Dead val#. |
| 1753 | // Is the def for the val# rematerializable? |
| 1754 | MachineInstr *ReMatDefMI = VNI->isDefAccurate() |
| 1755 | ? getInstructionFromIndex(VNI->def) : 0; |
| 1756 | bool dummy; |
| 1757 | if (ReMatDefMI && isReMaterializable(li, VNI, ReMatDefMI, SpillIs, dummy)) { |
| 1758 | // Remember how to remat the def of this val#. |
| 1759 | ReMatOrigDefs[VN] = ReMatDefMI; |
| 1760 | // Original def may be modified so we have to make a copy here. |
| 1761 | MachineInstr *Clone = mf_->CloneMachineInstr(ReMatDefMI); |
| 1762 | CloneMIs.push_back(Clone); |
| 1763 | ReMatDefs[VN] = Clone; |
| 1764 | |
| 1765 | bool CanDelete = true; |
| 1766 | if (VNI->hasPHIKill()) { |
| 1767 | // A kill is a phi node, not all of its uses can be rematerialized. |
| 1768 | // It must not be deleted. |
| 1769 | CanDelete = false; |
| 1770 | // Need a stack slot if there is any live range where uses cannot be |
| 1771 | // rematerialized. |
| 1772 | NeedStackSlot = true; |
| 1773 | } |
| 1774 | if (CanDelete) |
| 1775 | ReMatDelete.set(VN); |
| 1776 | } else { |
| 1777 | // Need a stack slot if there is any live range where uses cannot be |
| 1778 | // rematerialized. |
| 1779 | NeedStackSlot = true; |
| 1780 | } |
| 1781 | } |
| 1782 | |
| 1783 | // One stack slot per live interval. |
| 1784 | if (NeedStackSlot && vrm.getPreSplitReg(li.reg) == 0) { |
| 1785 | if (vrm.getStackSlot(li.reg) == VirtRegMap::NO_STACK_SLOT) |
| 1786 | Slot = vrm.assignVirt2StackSlot(li.reg); |
| 1787 | |
| 1788 | // This case only occurs when the prealloc splitter has already assigned |
| 1789 | // a stack slot to this vreg. |
| 1790 | else |
| 1791 | Slot = vrm.getStackSlot(li.reg); |
| 1792 | } |
| 1793 | |
| 1794 | // Create new intervals and rewrite defs and uses. |
| 1795 | for (LiveInterval::Ranges::const_iterator |
| 1796 | I = li.ranges.begin(), E = li.ranges.end(); I != E; ++I) { |
| 1797 | MachineInstr *ReMatDefMI = ReMatDefs[I->valno->id]; |
| 1798 | MachineInstr *ReMatOrigDefMI = ReMatOrigDefs[I->valno->id]; |
| 1799 | bool DefIsReMat = ReMatDefMI != NULL; |
| 1800 | bool CanDelete = ReMatDelete[I->valno->id]; |
| 1801 | int LdSlot = 0; |
| 1802 | bool isLoadSS = DefIsReMat && tii_->isLoadFromStackSlot(ReMatDefMI, LdSlot); |
| 1803 | bool isLoad = isLoadSS || |
| 1804 | (DefIsReMat && ReMatDefMI->getDesc().canFoldAsLoad()); |
| 1805 | rewriteInstructionsForSpills(li, TrySplit, I, ReMatOrigDefMI, ReMatDefMI, |
| 1806 | Slot, LdSlot, isLoad, isLoadSS, DefIsReMat, |
| 1807 | CanDelete, vrm, rc, ReMatIds, loopInfo, |
| 1808 | SpillMBBs, SpillIdxes, RestoreMBBs, RestoreIdxes, |
| 1809 | MBBVRegsMap, NewLIs); |
| 1810 | } |
| 1811 | |
| 1812 | // Insert spills / restores if we are splitting. |
| 1813 | if (!TrySplit) { |
| 1814 | handleSpilledImpDefs(li, vrm, rc, NewLIs); |
Shih-wei Liao | e445432 | 2010-04-07 12:21:42 -0700 | [diff] [blame] | 1815 | normalizeSpillWeights(NewLIs); |
Shih-wei Liao | e264f62 | 2010-02-10 11:10:31 -0800 | [diff] [blame] | 1816 | return NewLIs; |
| 1817 | } |
| 1818 | |
| 1819 | SmallPtrSet<LiveInterval*, 4> AddedKill; |
| 1820 | SmallVector<unsigned, 2> Ops; |
| 1821 | if (NeedStackSlot) { |
| 1822 | int Id = SpillMBBs.find_first(); |
| 1823 | while (Id != -1) { |
| 1824 | std::vector<SRInfo> &spills = SpillIdxes[Id]; |
| 1825 | for (unsigned i = 0, e = spills.size(); i != e; ++i) { |
| 1826 | SlotIndex index = spills[i].index; |
| 1827 | unsigned VReg = spills[i].vreg; |
| 1828 | LiveInterval &nI = getOrCreateInterval(VReg); |
| 1829 | bool isReMat = vrm.isReMaterialized(VReg); |
| 1830 | MachineInstr *MI = getInstructionFromIndex(index); |
| 1831 | bool CanFold = false; |
| 1832 | bool FoundUse = false; |
| 1833 | Ops.clear(); |
| 1834 | if (spills[i].canFold) { |
| 1835 | CanFold = true; |
| 1836 | for (unsigned j = 0, ee = MI->getNumOperands(); j != ee; ++j) { |
| 1837 | MachineOperand &MO = MI->getOperand(j); |
| 1838 | if (!MO.isReg() || MO.getReg() != VReg) |
| 1839 | continue; |
| 1840 | |
| 1841 | Ops.push_back(j); |
| 1842 | if (MO.isDef()) |
| 1843 | continue; |
| 1844 | if (isReMat || |
| 1845 | (!FoundUse && !alsoFoldARestore(Id, index, VReg, |
| 1846 | RestoreMBBs, RestoreIdxes))) { |
| 1847 | // MI has two-address uses of the same register. If the use |
| 1848 | // isn't the first and only use in the BB, then we can't fold |
| 1849 | // it. FIXME: Move this to rewriteInstructionsForSpills. |
| 1850 | CanFold = false; |
| 1851 | break; |
| 1852 | } |
| 1853 | FoundUse = true; |
| 1854 | } |
| 1855 | } |
| 1856 | // Fold the store into the def if possible. |
| 1857 | bool Folded = false; |
| 1858 | if (CanFold && !Ops.empty()) { |
| 1859 | if (tryFoldMemoryOperand(MI, vrm, NULL, index, Ops, true, Slot,VReg)){ |
| 1860 | Folded = true; |
| 1861 | if (FoundUse) { |
| 1862 | // Also folded uses, do not issue a load. |
| 1863 | eraseRestoreInfo(Id, index, VReg, RestoreMBBs, RestoreIdxes); |
| 1864 | nI.removeRange(index.getLoadIndex(), index.getDefIndex()); |
| 1865 | } |
| 1866 | nI.removeRange(index.getDefIndex(), index.getStoreIndex()); |
| 1867 | } |
| 1868 | } |
| 1869 | |
| 1870 | // Otherwise tell the spiller to issue a spill. |
| 1871 | if (!Folded) { |
| 1872 | LiveRange *LR = &nI.ranges[nI.ranges.size()-1]; |
| 1873 | bool isKill = LR->end == index.getStoreIndex(); |
| 1874 | if (!MI->registerDefIsDead(nI.reg)) |
| 1875 | // No need to spill a dead def. |
| 1876 | vrm.addSpillPoint(VReg, isKill, MI); |
| 1877 | if (isKill) |
| 1878 | AddedKill.insert(&nI); |
| 1879 | } |
| 1880 | } |
| 1881 | Id = SpillMBBs.find_next(Id); |
| 1882 | } |
| 1883 | } |
| 1884 | |
| 1885 | int Id = RestoreMBBs.find_first(); |
| 1886 | while (Id != -1) { |
| 1887 | std::vector<SRInfo> &restores = RestoreIdxes[Id]; |
| 1888 | for (unsigned i = 0, e = restores.size(); i != e; ++i) { |
| 1889 | SlotIndex index = restores[i].index; |
| 1890 | if (index == SlotIndex()) |
| 1891 | continue; |
| 1892 | unsigned VReg = restores[i].vreg; |
| 1893 | LiveInterval &nI = getOrCreateInterval(VReg); |
| 1894 | bool isReMat = vrm.isReMaterialized(VReg); |
| 1895 | MachineInstr *MI = getInstructionFromIndex(index); |
| 1896 | bool CanFold = false; |
| 1897 | Ops.clear(); |
| 1898 | if (restores[i].canFold) { |
| 1899 | CanFold = true; |
| 1900 | for (unsigned j = 0, ee = MI->getNumOperands(); j != ee; ++j) { |
| 1901 | MachineOperand &MO = MI->getOperand(j); |
| 1902 | if (!MO.isReg() || MO.getReg() != VReg) |
| 1903 | continue; |
| 1904 | |
| 1905 | if (MO.isDef()) { |
| 1906 | // If this restore were to be folded, it would have been folded |
| 1907 | // already. |
| 1908 | CanFold = false; |
| 1909 | break; |
| 1910 | } |
| 1911 | Ops.push_back(j); |
| 1912 | } |
| 1913 | } |
| 1914 | |
| 1915 | // Fold the load into the use if possible. |
| 1916 | bool Folded = false; |
| 1917 | if (CanFold && !Ops.empty()) { |
| 1918 | if (!isReMat) |
| 1919 | Folded = tryFoldMemoryOperand(MI, vrm, NULL,index,Ops,true,Slot,VReg); |
| 1920 | else { |
| 1921 | MachineInstr *ReMatDefMI = vrm.getReMaterializedMI(VReg); |
| 1922 | int LdSlot = 0; |
| 1923 | bool isLoadSS = tii_->isLoadFromStackSlot(ReMatDefMI, LdSlot); |
| 1924 | // If the rematerializable def is a load, also try to fold it. |
| 1925 | if (isLoadSS || ReMatDefMI->getDesc().canFoldAsLoad()) |
| 1926 | Folded = tryFoldMemoryOperand(MI, vrm, ReMatDefMI, index, |
| 1927 | Ops, isLoadSS, LdSlot, VReg); |
| 1928 | if (!Folded) { |
| 1929 | unsigned ImpUse = getReMatImplicitUse(li, ReMatDefMI); |
| 1930 | if (ImpUse) { |
| 1931 | // Re-matting an instruction with virtual register use. Add the |
Shih-wei Liao | e445432 | 2010-04-07 12:21:42 -0700 | [diff] [blame] | 1932 | // register as an implicit use on the use MI and mark the register |
| 1933 | // interval as unspillable. |
Shih-wei Liao | e264f62 | 2010-02-10 11:10:31 -0800 | [diff] [blame] | 1934 | LiveInterval &ImpLi = getInterval(ImpUse); |
Shih-wei Liao | e445432 | 2010-04-07 12:21:42 -0700 | [diff] [blame] | 1935 | ImpLi.markNotSpillable(); |
Shih-wei Liao | e264f62 | 2010-02-10 11:10:31 -0800 | [diff] [blame] | 1936 | MI->addOperand(MachineOperand::CreateReg(ImpUse, false, true)); |
| 1937 | } |
| 1938 | } |
| 1939 | } |
| 1940 | } |
| 1941 | // If folding is not possible / failed, then tell the spiller to issue a |
| 1942 | // load / rematerialization for us. |
| 1943 | if (Folded) |
| 1944 | nI.removeRange(index.getLoadIndex(), index.getDefIndex()); |
| 1945 | else |
| 1946 | vrm.addRestorePoint(VReg, MI); |
| 1947 | } |
| 1948 | Id = RestoreMBBs.find_next(Id); |
| 1949 | } |
| 1950 | |
| 1951 | // Finalize intervals: add kills, finalize spill weights, and filter out |
| 1952 | // dead intervals. |
| 1953 | std::vector<LiveInterval*> RetNewLIs; |
| 1954 | for (unsigned i = 0, e = NewLIs.size(); i != e; ++i) { |
| 1955 | LiveInterval *LI = NewLIs[i]; |
| 1956 | if (!LI->empty()) { |
| 1957 | LI->weight /= SlotIndex::NUM * getApproximateInstructionCount(*LI); |
| 1958 | if (!AddedKill.count(LI)) { |
| 1959 | LiveRange *LR = &LI->ranges[LI->ranges.size()-1]; |
| 1960 | SlotIndex LastUseIdx = LR->end.getBaseIndex(); |
| 1961 | MachineInstr *LastUse = getInstructionFromIndex(LastUseIdx); |
| 1962 | int UseIdx = LastUse->findRegisterUseOperandIdx(LI->reg, false); |
| 1963 | assert(UseIdx != -1); |
| 1964 | if (!LastUse->isRegTiedToDefOperand(UseIdx)) { |
| 1965 | LastUse->getOperand(UseIdx).setIsKill(); |
| 1966 | vrm.addKillPoint(LI->reg, LastUseIdx); |
| 1967 | } |
| 1968 | } |
| 1969 | RetNewLIs.push_back(LI); |
| 1970 | } |
| 1971 | } |
| 1972 | |
| 1973 | handleSpilledImpDefs(li, vrm, rc, RetNewLIs); |
Shih-wei Liao | e445432 | 2010-04-07 12:21:42 -0700 | [diff] [blame] | 1974 | normalizeSpillWeights(RetNewLIs); |
Shih-wei Liao | e264f62 | 2010-02-10 11:10:31 -0800 | [diff] [blame] | 1975 | return RetNewLIs; |
| 1976 | } |
| 1977 | |
| 1978 | /// hasAllocatableSuperReg - Return true if the specified physical register has |
| 1979 | /// any super register that's allocatable. |
| 1980 | bool LiveIntervals::hasAllocatableSuperReg(unsigned Reg) const { |
| 1981 | for (const unsigned* AS = tri_->getSuperRegisters(Reg); *AS; ++AS) |
| 1982 | if (allocatableRegs_[*AS] && hasInterval(*AS)) |
| 1983 | return true; |
| 1984 | return false; |
| 1985 | } |
| 1986 | |
| 1987 | /// getRepresentativeReg - Find the largest super register of the specified |
| 1988 | /// physical register. |
| 1989 | unsigned LiveIntervals::getRepresentativeReg(unsigned Reg) const { |
| 1990 | // Find the largest super-register that is allocatable. |
| 1991 | unsigned BestReg = Reg; |
| 1992 | for (const unsigned* AS = tri_->getSuperRegisters(Reg); *AS; ++AS) { |
| 1993 | unsigned SuperReg = *AS; |
| 1994 | if (!hasAllocatableSuperReg(SuperReg) && hasInterval(SuperReg)) { |
| 1995 | BestReg = SuperReg; |
| 1996 | break; |
| 1997 | } |
| 1998 | } |
| 1999 | return BestReg; |
| 2000 | } |
| 2001 | |
| 2002 | /// getNumConflictsWithPhysReg - Return the number of uses and defs of the |
| 2003 | /// specified interval that conflicts with the specified physical register. |
| 2004 | unsigned LiveIntervals::getNumConflictsWithPhysReg(const LiveInterval &li, |
| 2005 | unsigned PhysReg) const { |
| 2006 | unsigned NumConflicts = 0; |
| 2007 | const LiveInterval &pli = getInterval(getRepresentativeReg(PhysReg)); |
| 2008 | for (MachineRegisterInfo::reg_iterator I = mri_->reg_begin(li.reg), |
| 2009 | E = mri_->reg_end(); I != E; ++I) { |
| 2010 | MachineOperand &O = I.getOperand(); |
| 2011 | MachineInstr *MI = O.getParent(); |
| 2012 | SlotIndex Index = getInstructionIndex(MI); |
| 2013 | if (pli.liveAt(Index)) |
| 2014 | ++NumConflicts; |
| 2015 | } |
| 2016 | return NumConflicts; |
| 2017 | } |
| 2018 | |
| 2019 | /// spillPhysRegAroundRegDefsUses - Spill the specified physical register |
| 2020 | /// around all defs and uses of the specified interval. Return true if it |
| 2021 | /// was able to cut its interval. |
| 2022 | bool LiveIntervals::spillPhysRegAroundRegDefsUses(const LiveInterval &li, |
| 2023 | unsigned PhysReg, VirtRegMap &vrm) { |
| 2024 | unsigned SpillReg = getRepresentativeReg(PhysReg); |
| 2025 | |
| 2026 | for (const unsigned *AS = tri_->getAliasSet(PhysReg); *AS; ++AS) |
| 2027 | // If there are registers which alias PhysReg, but which are not a |
| 2028 | // sub-register of the chosen representative super register. Assert |
| 2029 | // since we can't handle it yet. |
| 2030 | assert(*AS == SpillReg || !allocatableRegs_[*AS] || !hasInterval(*AS) || |
| 2031 | tri_->isSuperRegister(*AS, SpillReg)); |
| 2032 | |
| 2033 | bool Cut = false; |
| 2034 | SmallVector<unsigned, 4> PRegs; |
| 2035 | if (hasInterval(SpillReg)) |
| 2036 | PRegs.push_back(SpillReg); |
| 2037 | else { |
| 2038 | SmallSet<unsigned, 4> Added; |
| 2039 | for (const unsigned* AS = tri_->getSubRegisters(SpillReg); *AS; ++AS) |
| 2040 | if (Added.insert(*AS) && hasInterval(*AS)) { |
| 2041 | PRegs.push_back(*AS); |
| 2042 | for (const unsigned* ASS = tri_->getSubRegisters(*AS); *ASS; ++ASS) |
| 2043 | Added.insert(*ASS); |
| 2044 | } |
| 2045 | } |
| 2046 | |
| 2047 | SmallPtrSet<MachineInstr*, 8> SeenMIs; |
| 2048 | for (MachineRegisterInfo::reg_iterator I = mri_->reg_begin(li.reg), |
| 2049 | E = mri_->reg_end(); I != E; ++I) { |
| 2050 | MachineOperand &O = I.getOperand(); |
| 2051 | MachineInstr *MI = O.getParent(); |
| 2052 | if (SeenMIs.count(MI)) |
| 2053 | continue; |
| 2054 | SeenMIs.insert(MI); |
| 2055 | SlotIndex Index = getInstructionIndex(MI); |
| 2056 | for (unsigned i = 0, e = PRegs.size(); i != e; ++i) { |
| 2057 | unsigned PReg = PRegs[i]; |
| 2058 | LiveInterval &pli = getInterval(PReg); |
| 2059 | if (!pli.liveAt(Index)) |
| 2060 | continue; |
| 2061 | vrm.addEmergencySpill(PReg, MI); |
| 2062 | SlotIndex StartIdx = Index.getLoadIndex(); |
| 2063 | SlotIndex EndIdx = Index.getNextIndex().getBaseIndex(); |
| 2064 | if (pli.isInOneLiveRange(StartIdx, EndIdx)) { |
| 2065 | pli.removeRange(StartIdx, EndIdx); |
| 2066 | Cut = true; |
| 2067 | } else { |
| 2068 | std::string msg; |
| 2069 | raw_string_ostream Msg(msg); |
| 2070 | Msg << "Ran out of registers during register allocation!"; |
| 2071 | if (MI->isInlineAsm()) { |
| 2072 | Msg << "\nPlease check your inline asm statement for invalid " |
| 2073 | << "constraints:\n"; |
| 2074 | MI->print(Msg, tm_); |
| 2075 | } |
| 2076 | llvm_report_error(Msg.str()); |
| 2077 | } |
| 2078 | for (const unsigned* AS = tri_->getSubRegisters(PReg); *AS; ++AS) { |
| 2079 | if (!hasInterval(*AS)) |
| 2080 | continue; |
| 2081 | LiveInterval &spli = getInterval(*AS); |
| 2082 | if (spli.liveAt(Index)) |
| 2083 | spli.removeRange(Index.getLoadIndex(), |
| 2084 | Index.getNextIndex().getBaseIndex()); |
| 2085 | } |
| 2086 | } |
| 2087 | } |
| 2088 | return Cut; |
| 2089 | } |
| 2090 | |
| 2091 | LiveRange LiveIntervals::addLiveRangeToEndOfBlock(unsigned reg, |
| 2092 | MachineInstr* startInst) { |
| 2093 | LiveInterval& Interval = getOrCreateInterval(reg); |
| 2094 | VNInfo* VN = Interval.getNextValue( |
| 2095 | SlotIndex(getInstructionIndex(startInst).getDefIndex()), |
| 2096 | startInst, true, getVNInfoAllocator()); |
| 2097 | VN->setHasPHIKill(true); |
| 2098 | VN->kills.push_back(indexes_->getTerminatorGap(startInst->getParent())); |
| 2099 | LiveRange LR( |
| 2100 | SlotIndex(getInstructionIndex(startInst).getDefIndex()), |
| 2101 | getMBBEndIdx(startInst->getParent()), VN); |
| 2102 | Interval.addRange(LR); |
| 2103 | |
| 2104 | return LR; |
| 2105 | } |
| 2106 | |