| 1 | //===---- MachineCombiner.cpp - Instcombining on SSA form machine code ----===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // The machine combiner pass uses machine trace metrics to ensure the combined |
| 10 | // instructions do not lengthen the critical path or the resource depth. |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #include "llvm/CodeGen/MachineCombiner.h" |
| 14 | #include "llvm/ADT/DenseMap.h" |
| 15 | #include "llvm/ADT/Statistic.h" |
| 16 | #include "llvm/Analysis/ProfileSummaryInfo.h" |
| 17 | #include "llvm/CodeGen/LazyMachineBlockFrequencyInfo.h" |
| 18 | #include "llvm/CodeGen/MachineBlockFrequencyInfo.h" |
| 19 | #include "llvm/CodeGen/MachineCombinerPattern.h" |
| 20 | #include "llvm/CodeGen/MachineDominators.h" |
| 21 | #include "llvm/CodeGen/MachineFunction.h" |
| 22 | #include "llvm/CodeGen/MachineFunctionAnalysis.h" |
| 23 | #include "llvm/CodeGen/MachineFunctionPass.h" |
| 24 | #include "llvm/CodeGen/MachineLoopInfo.h" |
| 25 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
| 26 | #include "llvm/CodeGen/MachineSizeOpts.h" |
| 27 | #include "llvm/CodeGen/MachineTraceMetrics.h" |
| 28 | #include "llvm/CodeGen/RegisterClassInfo.h" |
| 29 | #include "llvm/CodeGen/TargetInstrInfo.h" |
| 30 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
| 31 | #include "llvm/CodeGen/TargetSchedule.h" |
| 32 | #include "llvm/CodeGen/TargetSubtargetInfo.h" |
| 33 | #include "llvm/InitializePasses.h" |
| 34 | #include "llvm/Support/CommandLine.h" |
| 35 | #include "llvm/Support/Debug.h" |
| 36 | #include "llvm/Support/raw_ostream.h" |
| 37 | |
| 38 | using namespace llvm; |
| 39 | |
| 40 | #define DEBUG_TYPE "machine-combiner" |
| 41 | |
| 42 | STATISTIC(NumInstCombined, "Number of machineinst combined" ); |
| 43 | |
| 44 | static cl::opt<unsigned> |
| 45 | inc_threshold("machine-combiner-inc-threshold" , cl::Hidden, |
| 46 | cl::desc("Incremental depth computation will be used for basic " |
| 47 | "blocks with more instructions." ), cl::init(Val: 500)); |
| 48 | |
| 49 | static cl::opt<bool> dump_intrs("machine-combiner-dump-subst-intrs" , cl::Hidden, |
| 50 | cl::desc("Dump all substituted intrs" ), |
| 51 | cl::init(Val: false)); |
| 52 | |
| 53 | #ifdef EXPENSIVE_CHECKS |
| 54 | static cl::opt<bool> VerifyPatternOrder( |
| 55 | "machine-combiner-verify-pattern-order" , cl::Hidden, |
| 56 | cl::desc( |
| 57 | "Verify that the generated patterns are ordered by increasing latency" ), |
| 58 | cl::init(true)); |
| 59 | #else |
| 60 | static cl::opt<bool> VerifyPatternOrder( |
| 61 | "machine-combiner-verify-pattern-order" , cl::Hidden, |
| 62 | cl::desc( |
| 63 | "Verify that the generated patterns are ordered by increasing latency" ), |
| 64 | cl::init(Val: false)); |
| 65 | #endif |
| 66 | |
| 67 | namespace { |
| 68 | class MachineCombinerImpl { |
| 69 | const TargetSubtargetInfo *STI = nullptr; |
| 70 | const TargetInstrInfo *TII = nullptr; |
| 71 | const TargetRegisterInfo *TRI = nullptr; |
| 72 | MCSchedModel SchedModel; |
| 73 | MachineRegisterInfo *MRI = nullptr; |
| 74 | MachineLoopInfo *MLI = nullptr; // Current MachineLoopInfo |
| 75 | MachineTraceMetrics *Traces = nullptr; |
| 76 | MachineTraceMetrics::Ensemble *TraceEnsemble = nullptr; |
| 77 | MachineBlockFrequencyInfo *MBFI = nullptr; |
| 78 | ProfileSummaryInfo *PSI = nullptr; |
| 79 | RegisterClassInfo *RegClassInfo = nullptr; |
| 80 | |
| 81 | TargetSchedModel TSchedModel; |
| 82 | |
| 83 | public: |
| 84 | MachineCombinerImpl() = default; |
| 85 | bool run(MachineFunction &MF, MachineLoopInfo *MLI, |
| 86 | MachineTraceMetrics *Traces, ProfileSummaryInfo *PSI, |
| 87 | MachineBlockFrequencyInfo *MBFI, RegisterClassInfo *RegClassInfo); |
| 88 | |
| 89 | private: |
| 90 | bool combineInstructions(MachineBasicBlock *); |
| 91 | MachineInstr *getOperandDef(const MachineOperand &MO); |
| 92 | bool isTransientMI(const MachineInstr *MI); |
| 93 | unsigned getDepth(SmallVectorImpl<MachineInstr *> &InsInstrs, |
| 94 | DenseMap<Register, unsigned> &InstrIdxForVirtReg, |
| 95 | MachineTraceMetrics::Trace BlockTrace, |
| 96 | const MachineBasicBlock &MBB); |
| 97 | unsigned getLatency(MachineInstr *Root, MachineInstr *NewRoot, |
| 98 | MachineTraceMetrics::Trace BlockTrace); |
| 99 | bool improvesCriticalPathLen(MachineBasicBlock *MBB, MachineInstr *Root, |
| 100 | MachineTraceMetrics::Trace BlockTrace, |
| 101 | SmallVectorImpl<MachineInstr *> &InsInstrs, |
| 102 | SmallVectorImpl<MachineInstr *> &DelInstrs, |
| 103 | DenseMap<Register, unsigned> &InstrIdxForVirtReg, |
| 104 | unsigned Pattern, bool SlackIsAccurate); |
| 105 | bool reduceRegisterPressure(MachineInstr &Root, MachineBasicBlock *MBB, |
| 106 | SmallVectorImpl<MachineInstr *> &InsInstrs, |
| 107 | SmallVectorImpl<MachineInstr *> &DelInstrs, |
| 108 | unsigned Pattern); |
| 109 | bool preservesResourceLen(MachineBasicBlock *MBB, |
| 110 | MachineTraceMetrics::Trace BlockTrace, |
| 111 | SmallVectorImpl<MachineInstr *> &InsInstrs, |
| 112 | SmallVectorImpl<MachineInstr *> &DelInstrs); |
| 113 | void instr2instrSC(SmallVectorImpl<MachineInstr *> &Instrs, |
| 114 | SmallVectorImpl<const MCSchedClassDesc *> &); |
| 115 | std::pair<unsigned, unsigned> |
| 116 | getLatenciesForInstrSequences(MachineInstr &MI, |
| 117 | SmallVectorImpl<MachineInstr *> &InsInstrs, |
| 118 | SmallVectorImpl<MachineInstr *> &DelInstrs, |
| 119 | MachineTraceMetrics::Trace BlockTrace); |
| 120 | |
| 121 | CombinerObjective getCombinerObjective(unsigned Pattern); |
| 122 | }; |
| 123 | |
| 124 | class MachineCombinerLegacy : public MachineFunctionPass { |
| 125 | public: |
| 126 | static char ID; |
| 127 | MachineCombinerLegacy() : MachineFunctionPass(ID) {} |
| 128 | void getAnalysisUsage(AnalysisUsage &AU) const override; |
| 129 | bool runOnMachineFunction(MachineFunction &MF) override; |
| 130 | StringRef getPassName() const override { return "Machine InstCombiner" ; } |
| 131 | }; |
| 132 | } // namespace |
| 133 | |
| 134 | char MachineCombinerLegacy::ID = 0; |
| 135 | char &llvm::MachineCombinerID = MachineCombinerLegacy::ID; |
| 136 | |
| 137 | INITIALIZE_PASS_BEGIN(MachineCombinerLegacy, DEBUG_TYPE, "Machine InstCombiner" , |
| 138 | false, false) |
| 139 | INITIALIZE_PASS_DEPENDENCY(MachineLoopInfoWrapperPass) |
| 140 | INITIALIZE_PASS_DEPENDENCY(MachineRegisterClassInfoWrapperPass) |
| 141 | INITIALIZE_PASS_DEPENDENCY(MachineTraceMetricsWrapperPass) |
| 142 | INITIALIZE_PASS_END(MachineCombinerLegacy, DEBUG_TYPE, "Machine InstCombiner" , |
| 143 | false, false) |
| 144 | |
| 145 | void MachineCombinerLegacy::getAnalysisUsage(AnalysisUsage &AU) const { |
| 146 | AU.setPreservesCFG(); |
| 147 | AU.addRequired<MachineLoopInfoWrapperPass>(); |
| 148 | AU.addRequired<MachineRegisterClassInfoWrapperPass>(); |
| 149 | AU.addRequired<MachineTraceMetricsWrapperPass>(); |
| 150 | AU.addPreserved<MachineTraceMetricsWrapperPass>(); |
| 151 | AU.addRequired<LazyMachineBlockFrequencyInfoPass>(); |
| 152 | AU.addRequired<ProfileSummaryInfoWrapperPass>(); |
| 153 | MachineFunctionPass::getAnalysisUsage(AU); |
| 154 | } |
| 155 | |
| 156 | MachineInstr *MachineCombinerImpl::getOperandDef(const MachineOperand &MO) { |
| 157 | MachineInstr *DefInstr = nullptr; |
| 158 | // We need a virtual register definition. |
| 159 | if (MO.isReg() && MO.getReg().isVirtual()) |
| 160 | DefInstr = MRI->getUniqueVRegDef(Reg: MO.getReg()); |
| 161 | return DefInstr; |
| 162 | } |
| 163 | |
| 164 | /// Return true if MI is unlikely to generate an actual target instruction. |
| 165 | bool MachineCombinerImpl::isTransientMI(const MachineInstr *MI) { |
| 166 | if (!MI->isCopy()) |
| 167 | return MI->isTransient(); |
| 168 | |
| 169 | // If MI is a COPY, check if its src and dst registers can be coalesced. |
| 170 | Register Dst = MI->getOperand(i: 0).getReg(); |
| 171 | Register Src = MI->getOperand(i: 1).getReg(); |
| 172 | |
| 173 | if (!MI->isFullCopy()) { |
| 174 | // If src RC contains super registers of dst RC, it can also be coalesced. |
| 175 | if (MI->getOperand(i: 0).getSubReg() || Src.isPhysical() || Dst.isPhysical()) |
| 176 | return false; |
| 177 | |
| 178 | auto SrcSub = MI->getOperand(i: 1).getSubReg(); |
| 179 | auto SrcRC = MRI->getRegClass(Reg: Src); |
| 180 | auto DstRC = MRI->getRegClass(Reg: Dst); |
| 181 | return TRI->getMatchingSuperRegClass(A: SrcRC, B: DstRC, Idx: SrcSub) != nullptr; |
| 182 | } |
| 183 | |
| 184 | if (Src.isPhysical() && Dst.isPhysical()) |
| 185 | return Src == Dst; |
| 186 | |
| 187 | if (Src.isVirtual() && Dst.isVirtual()) { |
| 188 | auto SrcRC = MRI->getRegClass(Reg: Src); |
| 189 | auto DstRC = MRI->getRegClass(Reg: Dst); |
| 190 | return SrcRC->hasSuperClassEq(RC: DstRC) || SrcRC->hasSubClassEq(RC: DstRC); |
| 191 | } |
| 192 | |
| 193 | if (Src.isVirtual()) |
| 194 | std::swap(a&: Src, b&: Dst); |
| 195 | |
| 196 | // Now Src is physical register, Dst is virtual register. |
| 197 | auto DstRC = MRI->getRegClass(Reg: Dst); |
| 198 | return DstRC->contains(Reg: Src); |
| 199 | } |
| 200 | |
| 201 | /// Computes depth of instructions in vector \InsInstr. |
| 202 | /// |
| 203 | /// \param InsInstrs is a vector of machine instructions |
| 204 | /// \param InstrIdxForVirtReg is a dense map of virtual register to index |
| 205 | /// of defining machine instruction in \p InsInstrs |
| 206 | /// \param BlockTrace is a trace of machine instructions |
| 207 | /// |
| 208 | /// \returns Depth of last instruction in \InsInstrs ("NewRoot") |
| 209 | unsigned |
| 210 | MachineCombinerImpl::getDepth(SmallVectorImpl<MachineInstr *> &InsInstrs, |
| 211 | DenseMap<Register, unsigned> &InstrIdxForVirtReg, |
| 212 | MachineTraceMetrics::Trace BlockTrace, |
| 213 | const MachineBasicBlock &MBB) { |
| 214 | SmallVector<unsigned, 16> InstrDepth; |
| 215 | // For each instruction in the new sequence compute the depth based on the |
| 216 | // operands. Use the trace information when possible. For new operands which |
| 217 | // are tracked in the InstrIdxForVirtReg map depth is looked up in InstrDepth |
| 218 | for (auto *InstrPtr : InsInstrs) { // for each Use |
| 219 | unsigned IDepth = 0; |
| 220 | for (const MachineOperand &MO : InstrPtr->all_uses()) { |
| 221 | // Check for virtual register operand. |
| 222 | if (!MO.getReg().isVirtual()) |
| 223 | continue; |
| 224 | unsigned DepthOp = 0; |
| 225 | unsigned LatencyOp = 0; |
| 226 | auto II = InstrIdxForVirtReg.find(Val: MO.getReg()); |
| 227 | if (II != InstrIdxForVirtReg.end()) { |
| 228 | // Operand is new virtual register not in trace |
| 229 | assert(II->second < InstrDepth.size() && "Bad Index" ); |
| 230 | MachineInstr *DefInstr = InsInstrs[II->second]; |
| 231 | assert(DefInstr && |
| 232 | "There must be a definition for a new virtual register" ); |
| 233 | DepthOp = InstrDepth[II->second]; |
| 234 | int DefIdx = |
| 235 | DefInstr->findRegisterDefOperandIdx(Reg: MO.getReg(), /*TRI=*/nullptr); |
| 236 | int UseIdx = |
| 237 | InstrPtr->findRegisterUseOperandIdx(Reg: MO.getReg(), /*TRI=*/nullptr); |
| 238 | LatencyOp = TSchedModel.computeOperandLatency(DefMI: DefInstr, DefOperIdx: DefIdx, |
| 239 | UseMI: InstrPtr, UseOperIdx: UseIdx); |
| 240 | } else { |
| 241 | MachineInstr *DefInstr = getOperandDef(MO); |
| 242 | if (DefInstr && (TII->getMachineCombinerTraceStrategy() != |
| 243 | MachineTraceStrategy::TS_Local || |
| 244 | DefInstr->getParent() == &MBB)) { |
| 245 | DepthOp = BlockTrace.getInstrCycles(MI: *DefInstr).Depth; |
| 246 | if (!isTransientMI(MI: DefInstr)) |
| 247 | LatencyOp = TSchedModel.computeOperandLatency( |
| 248 | DefMI: DefInstr, |
| 249 | DefOperIdx: DefInstr->findRegisterDefOperandIdx(Reg: MO.getReg(), |
| 250 | /*TRI=*/nullptr), |
| 251 | UseMI: InstrPtr, |
| 252 | UseOperIdx: InstrPtr->findRegisterUseOperandIdx(Reg: MO.getReg(), |
| 253 | /*TRI=*/nullptr)); |
| 254 | } |
| 255 | } |
| 256 | IDepth = std::max(a: IDepth, b: DepthOp + LatencyOp); |
| 257 | } |
| 258 | InstrDepth.push_back(Elt: IDepth); |
| 259 | } |
| 260 | unsigned NewRootIdx = InsInstrs.size() - 1; |
| 261 | return InstrDepth[NewRootIdx]; |
| 262 | } |
| 263 | |
| 264 | /// Computes instruction latency as max of latency of defined operands. |
| 265 | /// |
| 266 | /// \param Root is a machine instruction that could be replaced by NewRoot. |
| 267 | /// It is used to compute a more accurate latency information for NewRoot in |
| 268 | /// case there is a dependent instruction in the same trace (\p BlockTrace) |
| 269 | /// \param NewRoot is the instruction for which the latency is computed |
| 270 | /// \param BlockTrace is a trace of machine instructions |
| 271 | /// |
| 272 | /// \returns Latency of \p NewRoot |
| 273 | unsigned |
| 274 | MachineCombinerImpl::getLatency(MachineInstr *Root, MachineInstr *NewRoot, |
| 275 | MachineTraceMetrics::Trace BlockTrace) { |
| 276 | // Check each definition in NewRoot and compute the latency |
| 277 | unsigned NewRootLatency = 0; |
| 278 | |
| 279 | for (const MachineOperand &MO : NewRoot->all_defs()) { |
| 280 | // Check for virtual register operand. |
| 281 | if (!MO.getReg().isVirtual()) |
| 282 | continue; |
| 283 | // Get the first instruction that uses MO |
| 284 | MachineRegisterInfo::reg_iterator RI = MRI->reg_begin(RegNo: MO.getReg()); |
| 285 | RI++; |
| 286 | if (RI == MRI->reg_end()) |
| 287 | continue; |
| 288 | MachineInstr *UseMO = RI->getParent(); |
| 289 | unsigned LatencyOp = 0; |
| 290 | if (UseMO && BlockTrace.isDepInTrace(DefMI: *Root, UseMI: *UseMO)) { |
| 291 | LatencyOp = TSchedModel.computeOperandLatency( |
| 292 | DefMI: NewRoot, |
| 293 | DefOperIdx: NewRoot->findRegisterDefOperandIdx(Reg: MO.getReg(), /*TRI=*/nullptr), |
| 294 | UseMI: UseMO, |
| 295 | UseOperIdx: UseMO->findRegisterUseOperandIdx(Reg: MO.getReg(), /*TRI=*/nullptr)); |
| 296 | } else { |
| 297 | LatencyOp = TSchedModel.computeInstrLatency(MI: NewRoot); |
| 298 | } |
| 299 | NewRootLatency = std::max(a: NewRootLatency, b: LatencyOp); |
| 300 | } |
| 301 | return NewRootLatency; |
| 302 | } |
| 303 | |
| 304 | CombinerObjective MachineCombinerImpl::getCombinerObjective(unsigned Pattern) { |
| 305 | // TODO: If C++ ever gets a real enum class, make this part of the |
| 306 | // MachineCombinerPattern class. |
| 307 | switch (Pattern) { |
| 308 | case MachineCombinerPattern::REASSOC_AX_BY: |
| 309 | case MachineCombinerPattern::REASSOC_AX_YB: |
| 310 | case MachineCombinerPattern::REASSOC_XA_BY: |
| 311 | case MachineCombinerPattern::REASSOC_XA_YB: |
| 312 | return CombinerObjective::MustReduceDepth; |
| 313 | default: |
| 314 | return TII->getCombinerObjective(Pattern); |
| 315 | } |
| 316 | } |
| 317 | |
| 318 | /// Estimate the latency of the new and original instruction sequence by summing |
| 319 | /// up the latencies of the inserted and deleted instructions. This assumes |
| 320 | /// that the inserted and deleted instructions are dependent instruction chains, |
| 321 | /// which might not hold in all cases. |
| 322 | std::pair<unsigned, unsigned> |
| 323 | MachineCombinerImpl::getLatenciesForInstrSequences( |
| 324 | MachineInstr &MI, SmallVectorImpl<MachineInstr *> &InsInstrs, |
| 325 | SmallVectorImpl<MachineInstr *> &DelInstrs, |
| 326 | MachineTraceMetrics::Trace BlockTrace) { |
| 327 | assert(!InsInstrs.empty() && "Only support sequences that insert instrs." ); |
| 328 | unsigned NewRootLatency = 0; |
| 329 | // NewRoot is the last instruction in the \p InsInstrs vector. |
| 330 | MachineInstr *NewRoot = InsInstrs.back(); |
| 331 | for (unsigned i = 0; i < InsInstrs.size() - 1; i++) |
| 332 | NewRootLatency += TSchedModel.computeInstrLatency(MI: InsInstrs[i]); |
| 333 | NewRootLatency += getLatency(Root: &MI, NewRoot, BlockTrace); |
| 334 | |
| 335 | unsigned RootLatency = 0; |
| 336 | for (auto *I : DelInstrs) |
| 337 | RootLatency += TSchedModel.computeInstrLatency(MI: I); |
| 338 | |
| 339 | return {NewRootLatency, RootLatency}; |
| 340 | } |
| 341 | |
| 342 | bool MachineCombinerImpl::reduceRegisterPressure( |
| 343 | MachineInstr &Root, MachineBasicBlock *MBB, |
| 344 | SmallVectorImpl<MachineInstr *> &InsInstrs, |
| 345 | SmallVectorImpl<MachineInstr *> &DelInstrs, unsigned Pattern) { |
| 346 | // FIXME: for now, we don't do any check for the register pressure patterns. |
| 347 | // We treat them as always profitable. But we can do better if we make |
| 348 | // RegPressureTracker class be aware of TIE attribute. Then we can get an |
| 349 | // accurate compare of register pressure with DelInstrs or InsInstrs. |
| 350 | return true; |
| 351 | } |
| 352 | |
| 353 | /// The DAGCombine code sequence ends in MI (Machine Instruction) Root. |
| 354 | /// The new code sequence ends in MI NewRoot. A necessary condition for the new |
| 355 | /// sequence to replace the old sequence is that it cannot lengthen the critical |
| 356 | /// path. The definition of "improve" may be restricted by specifying that the |
| 357 | /// new path improves the data dependency chain (MustReduceDepth). |
| 358 | bool MachineCombinerImpl::improvesCriticalPathLen( |
| 359 | MachineBasicBlock *MBB, MachineInstr *Root, |
| 360 | MachineTraceMetrics::Trace BlockTrace, |
| 361 | SmallVectorImpl<MachineInstr *> &InsInstrs, |
| 362 | SmallVectorImpl<MachineInstr *> &DelInstrs, |
| 363 | DenseMap<Register, unsigned> &InstrIdxForVirtReg, unsigned Pattern, |
| 364 | bool SlackIsAccurate) { |
| 365 | // Get depth and latency of NewRoot and Root. |
| 366 | unsigned NewRootDepth = |
| 367 | getDepth(InsInstrs, InstrIdxForVirtReg, BlockTrace, MBB: *MBB); |
| 368 | unsigned RootDepth = BlockTrace.getInstrCycles(MI: *Root).Depth; |
| 369 | |
| 370 | LLVM_DEBUG(dbgs() << " Dependence data for " << *Root << "\tNewRootDepth: " |
| 371 | << NewRootDepth << "\tRootDepth: " << RootDepth); |
| 372 | |
| 373 | // For a transform such as reassociation, the cost equation is |
| 374 | // conservatively calculated so that we must improve the depth (data |
| 375 | // dependency cycles) in the critical path to proceed with the transform. |
| 376 | // Being conservative also protects against inaccuracies in the underlying |
| 377 | // machine trace metrics and CPU models. |
| 378 | if (getCombinerObjective(Pattern) == CombinerObjective::MustReduceDepth) { |
| 379 | LLVM_DEBUG(dbgs() << "\tIt MustReduceDepth " ); |
| 380 | LLVM_DEBUG(NewRootDepth < RootDepth |
| 381 | ? dbgs() << "\t and it does it\n" |
| 382 | : dbgs() << "\t but it does NOT do it\n" ); |
| 383 | return NewRootDepth < RootDepth; |
| 384 | } |
| 385 | |
| 386 | // A more flexible cost calculation for the critical path includes the slack |
| 387 | // of the original code sequence. This may allow the transform to proceed |
| 388 | // even if the instruction depths (data dependency cycles) become worse. |
| 389 | |
| 390 | // Account for the latency of the inserted and deleted instructions by |
| 391 | unsigned NewRootLatency, RootLatency; |
| 392 | if (TII->accumulateInstrSeqToRootLatency(Root&: *Root)) { |
| 393 | std::tie(args&: NewRootLatency, args&: RootLatency) = |
| 394 | getLatenciesForInstrSequences(MI&: *Root, InsInstrs, DelInstrs, BlockTrace); |
| 395 | } else { |
| 396 | NewRootLatency = TSchedModel.computeInstrLatency(MI: InsInstrs.back()); |
| 397 | RootLatency = TSchedModel.computeInstrLatency(MI: Root); |
| 398 | } |
| 399 | |
| 400 | unsigned RootSlack = BlockTrace.getInstrSlack(MI: *Root); |
| 401 | unsigned NewCycleCount = NewRootDepth + NewRootLatency; |
| 402 | unsigned OldCycleCount = |
| 403 | RootDepth + RootLatency + (SlackIsAccurate ? RootSlack : 0); |
| 404 | LLVM_DEBUG(dbgs() << "\n\tNewRootLatency: " << NewRootLatency |
| 405 | << "\tRootLatency: " << RootLatency << "\n\tRootSlack: " |
| 406 | << RootSlack << " SlackIsAccurate=" << SlackIsAccurate |
| 407 | << "\n\tNewRootDepth + NewRootLatency = " << NewCycleCount |
| 408 | << "\n\tRootDepth + RootLatency + RootSlack = " |
| 409 | << OldCycleCount); |
| 410 | LLVM_DEBUG(NewCycleCount <= OldCycleCount |
| 411 | ? dbgs() << "\n\t It IMPROVES PathLen because" |
| 412 | : dbgs() << "\n\t It DOES NOT improve PathLen because" ); |
| 413 | LLVM_DEBUG(dbgs() << "\n\t\tNewCycleCount = " << NewCycleCount |
| 414 | << ", OldCycleCount = " << OldCycleCount << "\n" ); |
| 415 | |
| 416 | return NewCycleCount <= OldCycleCount; |
| 417 | } |
| 418 | |
| 419 | /// helper routine to convert instructions into SC |
| 420 | void MachineCombinerImpl::instr2instrSC( |
| 421 | SmallVectorImpl<MachineInstr *> &Instrs, |
| 422 | SmallVectorImpl<const MCSchedClassDesc *> &) { |
| 423 | for (auto *InstrPtr : Instrs) { |
| 424 | unsigned Opc = InstrPtr->getOpcode(); |
| 425 | unsigned Idx = TII->get(Opcode: Opc).getSchedClass(); |
| 426 | const MCSchedClassDesc *SC = SchedModel.getSchedClassDesc(SchedClassIdx: Idx); |
| 427 | InstrsSC.push_back(Elt: SC); |
| 428 | } |
| 429 | } |
| 430 | |
| 431 | /// True when the new instructions do not increase resource length |
| 432 | bool MachineCombinerImpl::preservesResourceLen( |
| 433 | MachineBasicBlock *MBB, MachineTraceMetrics::Trace BlockTrace, |
| 434 | SmallVectorImpl<MachineInstr *> &InsInstrs, |
| 435 | SmallVectorImpl<MachineInstr *> &DelInstrs) { |
| 436 | if (!TSchedModel.hasInstrSchedModel()) |
| 437 | return true; |
| 438 | |
| 439 | // Compute current resource length |
| 440 | |
| 441 | // ArrayRef<const MachineBasicBlock *> MBBarr(MBB); |
| 442 | SmallVector<const MachineBasicBlock *, 1> MBBarr; |
| 443 | MBBarr.push_back(Elt: MBB); |
| 444 | unsigned ResLenBeforeCombine = BlockTrace.getResourceLength(Extrablocks: MBBarr); |
| 445 | |
| 446 | // Deal with SC rather than Instructions. |
| 447 | SmallVector<const MCSchedClassDesc *, 16> ; |
| 448 | SmallVector<const MCSchedClassDesc *, 16> ; |
| 449 | |
| 450 | instr2instrSC(Instrs&: InsInstrs, InstrsSC&: InsInstrsSC); |
| 451 | instr2instrSC(Instrs&: DelInstrs, InstrsSC&: DelInstrsSC); |
| 452 | |
| 453 | ArrayRef<const MCSchedClassDesc *> MSCInsArr{InsInstrsSC}; |
| 454 | ArrayRef<const MCSchedClassDesc *> MSCDelArr{DelInstrsSC}; |
| 455 | |
| 456 | // Compute new resource length. |
| 457 | unsigned ResLenAfterCombine = |
| 458 | BlockTrace.getResourceLength(Extrablocks: MBBarr, ExtraInstrs: MSCInsArr, RemoveInstrs: MSCDelArr); |
| 459 | |
| 460 | LLVM_DEBUG(dbgs() << "\t\tResource length before replacement: " |
| 461 | << ResLenBeforeCombine |
| 462 | << " and after: " << ResLenAfterCombine << "\n" ); |
| 463 | LLVM_DEBUG( |
| 464 | ResLenAfterCombine <= |
| 465 | ResLenBeforeCombine + TII->getExtendResourceLenLimit() |
| 466 | ? dbgs() << "\t\t As result it IMPROVES/PRESERVES Resource Length\n" |
| 467 | : dbgs() << "\t\t As result it DOES NOT improve/preserve Resource " |
| 468 | "Length\n" ); |
| 469 | |
| 470 | return ResLenAfterCombine <= |
| 471 | ResLenBeforeCombine + TII->getExtendResourceLenLimit(); |
| 472 | } |
| 473 | |
| 474 | /// Inserts InsInstrs and deletes DelInstrs. Incrementally updates instruction |
| 475 | /// depths if requested. |
| 476 | /// |
| 477 | /// \param MBB basic block to insert instructions in |
| 478 | /// \param MI current machine instruction |
| 479 | /// \param InsInstrs new instructions to insert in \p MBB |
| 480 | /// \param DelInstrs instruction to delete from \p MBB |
| 481 | /// \param TraceEnsemble is a pointer to the machine trace information |
| 482 | /// \param RegUnits set of live registers, needed to compute instruction depths |
| 483 | /// \param TII is target instruction info, used to call target hook |
| 484 | /// \param Pattern is used to call target hook finalizeInsInstrs |
| 485 | /// \param IncrementalUpdate if true, compute instruction depths incrementally, |
| 486 | /// otherwise invalidate the trace |
| 487 | static void |
| 488 | insertDeleteInstructions(MachineBasicBlock *MBB, MachineInstr &MI, |
| 489 | SmallVectorImpl<MachineInstr *> &InsInstrs, |
| 490 | SmallVectorImpl<MachineInstr *> &DelInstrs, |
| 491 | MachineTraceMetrics::Ensemble *TraceEnsemble, |
| 492 | LiveRegUnitSet &RegUnits, const TargetInstrInfo *TII, |
| 493 | unsigned Pattern, bool IncrementalUpdate) { |
| 494 | // If we want to fix up some placeholder for some target, do it now. |
| 495 | // We need this because in genAlternativeCodeSequence, we have not decided the |
| 496 | // better pattern InsInstrs or DelInstrs, so we don't want generate some |
| 497 | // sideeffect to the function. For example we need to delay the constant pool |
| 498 | // entry creation here after InsInstrs is selected as better pattern. |
| 499 | // Otherwise the constant pool entry created for InsInstrs will not be deleted |
| 500 | // even if InsInstrs is not the better pattern. |
| 501 | TII->finalizeInsInstrs(Root&: MI, Pattern, InsInstrs); |
| 502 | |
| 503 | for (auto *InstrPtr : InsInstrs) |
| 504 | MBB->insert(I: (MachineBasicBlock::iterator)&MI, MI: InstrPtr); |
| 505 | |
| 506 | for (auto *InstrPtr : DelInstrs) { |
| 507 | InstrPtr->eraseFromParent(); |
| 508 | // Erase all LiveRegs defined by the removed instruction |
| 509 | for (auto *I = RegUnits.begin(); I != RegUnits.end();) { |
| 510 | if (I->MI == InstrPtr) |
| 511 | I = RegUnits.erase(I); |
| 512 | else |
| 513 | I++; |
| 514 | } |
| 515 | } |
| 516 | |
| 517 | if (IncrementalUpdate) |
| 518 | for (auto *InstrPtr : InsInstrs) |
| 519 | TraceEnsemble->updateDepth(MBB, *InstrPtr, RegUnits); |
| 520 | else |
| 521 | TraceEnsemble->invalidate(MBB); |
| 522 | |
| 523 | NumInstCombined++; |
| 524 | } |
| 525 | |
| 526 | /// Substitute a slow code sequence with a faster one by |
| 527 | /// evaluating instruction combining pattern. |
| 528 | /// The prototype of such a pattern is MUl + ADD -> MADD. Performs instruction |
| 529 | /// combining based on machine trace metrics. Only combine a sequence of |
| 530 | /// instructions when this neither lengthens the critical path nor increases |
| 531 | /// resource pressure. When optimizing for codesize always combine when the new |
| 532 | /// sequence is shorter. |
| 533 | bool MachineCombinerImpl::combineInstructions(MachineBasicBlock *MBB) { |
| 534 | bool Changed = false; |
| 535 | LLVM_DEBUG(dbgs() << "Combining MBB " << MBB->getName() << "\n" ); |
| 536 | |
| 537 | bool IncrementalUpdate = false; |
| 538 | auto BlockIter = MBB->begin(); |
| 539 | decltype(BlockIter) LastUpdate; |
| 540 | // Check if the block is in a loop. |
| 541 | const MachineLoop *ML = MLI->getLoopFor(BB: MBB); |
| 542 | if (!TraceEnsemble) |
| 543 | TraceEnsemble = Traces->getEnsemble(TII->getMachineCombinerTraceStrategy()); |
| 544 | |
| 545 | LiveRegUnitSet RegUnits; |
| 546 | RegUnits.setUniverse(TRI->getNumRegUnits()); |
| 547 | |
| 548 | bool OptForSize = llvm::shouldOptimizeForSize(MBB, PSI, MBFI); |
| 549 | |
| 550 | bool DoRegPressureReduce = |
| 551 | TII->shouldReduceRegisterPressure(MBB, RegClassInfo); |
| 552 | |
| 553 | while (BlockIter != MBB->end()) { |
| 554 | auto &MI = *BlockIter++; |
| 555 | SmallVector<unsigned, 16> Patterns; |
| 556 | // The motivating example is: |
| 557 | // |
| 558 | // MUL Other MUL_op1 MUL_op2 Other |
| 559 | // \ / \ | / |
| 560 | // ADD/SUB => MADD/MSUB |
| 561 | // (=Root) (=NewRoot) |
| 562 | |
| 563 | // The DAGCombine code always replaced MUL + ADD/SUB by MADD. While this is |
| 564 | // usually beneficial for code size it unfortunately can hurt performance |
| 565 | // when the ADD is on the critical path, but the MUL is not. With the |
| 566 | // substitution the MUL becomes part of the critical path (in form of the |
| 567 | // MADD) and can lengthen it on architectures where the MADD latency is |
| 568 | // longer than the ADD latency. |
| 569 | // |
| 570 | // For each instruction we check if it can be the root of a combiner |
| 571 | // pattern. Then for each pattern the new code sequence in form of MI is |
| 572 | // generated and evaluated. When the efficiency criteria (don't lengthen |
| 573 | // critical path, don't use more resources) is met the new sequence gets |
| 574 | // hooked up into the basic block before the old sequence is removed. |
| 575 | // |
| 576 | // The algorithm does not try to evaluate all patterns and pick the best. |
| 577 | // This is only an artificial restriction though. In practice there is |
| 578 | // mostly one pattern, and getMachineCombinerPatterns() can order patterns |
| 579 | // based on an internal cost heuristic. If |
| 580 | // machine-combiner-verify-pattern-order is enabled, all patterns are |
| 581 | // checked to ensure later patterns do not provide better latency savings. |
| 582 | |
| 583 | if (!TII->getMachineCombinerPatterns(Root&: MI, Patterns, DoRegPressureReduce)) |
| 584 | continue; |
| 585 | |
| 586 | // Only used when VerifyPatternOrder is enabled. |
| 587 | [[maybe_unused]] long PrevLatencyDiff = std::numeric_limits<long>::max(); |
| 588 | |
| 589 | for (const auto P : Patterns) { |
| 590 | SmallVector<MachineInstr *, 16> InsInstrs; |
| 591 | SmallVector<MachineInstr *, 16> DelInstrs; |
| 592 | DenseMap<Register, unsigned> InstrIdxForVirtReg; |
| 593 | TII->genAlternativeCodeSequence(Root&: MI, Pattern: P, InsInstrs, DelInstrs, |
| 594 | InstIdxForVirtReg&: InstrIdxForVirtReg); |
| 595 | // Found pattern, but did not generate alternative sequence. |
| 596 | // This can happen e.g. when an immediate could not be materialized |
| 597 | // in a single instruction. |
| 598 | if (InsInstrs.empty()) |
| 599 | continue; |
| 600 | |
| 601 | LLVM_DEBUG(if (dump_intrs) { |
| 602 | dbgs() << "\tFor the Pattern (" << (int)P |
| 603 | << ") these instructions could be removed\n" ; |
| 604 | for (auto const *InstrPtr : DelInstrs) |
| 605 | InstrPtr->print(dbgs(), /*IsStandalone*/false, /*SkipOpers*/false, |
| 606 | /*SkipDebugLoc*/false, /*AddNewLine*/true, TII); |
| 607 | dbgs() << "\tThese instructions could replace the removed ones\n" ; |
| 608 | for (auto const *InstrPtr : InsInstrs) |
| 609 | InstrPtr->print(dbgs(), /*IsStandalone*/false, /*SkipOpers*/false, |
| 610 | /*SkipDebugLoc*/false, /*AddNewLine*/true, TII); |
| 611 | }); |
| 612 | |
| 613 | // Check that the difference between original and new latency is |
| 614 | // decreasing for later patterns. This helps to discover sub-optimal |
| 615 | // pattern orderings. |
| 616 | if (VerifyPatternOrder && TSchedModel.hasInstrSchedModelOrItineraries()) { |
| 617 | auto [NewRootLatency, RootLatency] = getLatenciesForInstrSequences( |
| 618 | MI, InsInstrs, DelInstrs, BlockTrace: TraceEnsemble->getTrace(MBB)); |
| 619 | long CurrentLatencyDiff = ((long)RootLatency) - ((long)NewRootLatency); |
| 620 | assert(CurrentLatencyDiff <= PrevLatencyDiff && |
| 621 | "Current pattern is expected to be better than the previous " |
| 622 | "pattern." ); |
| 623 | PrevLatencyDiff = CurrentLatencyDiff; |
| 624 | } |
| 625 | |
| 626 | if (IncrementalUpdate && LastUpdate != BlockIter) { |
| 627 | // Update depths since the last incremental update. |
| 628 | TraceEnsemble->updateDepths(Start: LastUpdate, End: BlockIter, RegUnits); |
| 629 | LastUpdate = BlockIter; |
| 630 | } |
| 631 | |
| 632 | if (DoRegPressureReduce && |
| 633 | getCombinerObjective(Pattern: P) == |
| 634 | CombinerObjective::MustReduceRegisterPressure) { |
| 635 | if (MBB->size() > inc_threshold) { |
| 636 | // Use incremental depth updates for basic blocks above threshold |
| 637 | IncrementalUpdate = true; |
| 638 | LastUpdate = BlockIter; |
| 639 | } |
| 640 | if (reduceRegisterPressure(Root&: MI, MBB, InsInstrs, DelInstrs, Pattern: P)) { |
| 641 | // Replace DelInstrs with InsInstrs. |
| 642 | insertDeleteInstructions(MBB, MI, InsInstrs, DelInstrs, TraceEnsemble, |
| 643 | RegUnits, TII, Pattern: P, IncrementalUpdate); |
| 644 | Changed |= true; |
| 645 | |
| 646 | // Go back to previous instruction as it may have ILP reassociation |
| 647 | // opportunity. |
| 648 | BlockIter--; |
| 649 | break; |
| 650 | } |
| 651 | } |
| 652 | |
| 653 | if (ML && TII->isThroughputPattern(Pattern: P)) { |
| 654 | LLVM_DEBUG(dbgs() << "\t Replacing due to throughput pattern in loop\n" ); |
| 655 | insertDeleteInstructions(MBB, MI, InsInstrs, DelInstrs, TraceEnsemble, |
| 656 | RegUnits, TII, Pattern: P, IncrementalUpdate); |
| 657 | // Eagerly stop after the first pattern fires. |
| 658 | Changed = true; |
| 659 | break; |
| 660 | } else if (OptForSize && InsInstrs.size() < DelInstrs.size()) { |
| 661 | LLVM_DEBUG(dbgs() << "\t Replacing due to OptForSize (" |
| 662 | << InsInstrs.size() << " < " |
| 663 | << DelInstrs.size() << ")\n" ); |
| 664 | insertDeleteInstructions(MBB, MI, InsInstrs, DelInstrs, TraceEnsemble, |
| 665 | RegUnits, TII, Pattern: P, IncrementalUpdate); |
| 666 | // Eagerly stop after the first pattern fires. |
| 667 | Changed = true; |
| 668 | break; |
| 669 | } else { |
| 670 | // For big basic blocks, we only compute the full trace the first time |
| 671 | // we hit this. We do not invalidate the trace, but instead update the |
| 672 | // instruction depths incrementally. |
| 673 | // NOTE: Only the instruction depths up to MI are accurate. All other |
| 674 | // trace information is not updated. |
| 675 | MachineTraceMetrics::Trace BlockTrace = TraceEnsemble->getTrace(MBB); |
| 676 | Traces->verifyAnalysis(); |
| 677 | if (improvesCriticalPathLen(MBB, Root: &MI, BlockTrace, InsInstrs, DelInstrs, |
| 678 | InstrIdxForVirtReg, Pattern: P, |
| 679 | SlackIsAccurate: !IncrementalUpdate) && |
| 680 | preservesResourceLen(MBB, BlockTrace, InsInstrs, DelInstrs)) { |
| 681 | if (MBB->size() > inc_threshold) { |
| 682 | // Use incremental depth updates for basic blocks above treshold |
| 683 | IncrementalUpdate = true; |
| 684 | LastUpdate = BlockIter; |
| 685 | } |
| 686 | |
| 687 | insertDeleteInstructions(MBB, MI, InsInstrs, DelInstrs, TraceEnsemble, |
| 688 | RegUnits, TII, Pattern: P, IncrementalUpdate); |
| 689 | |
| 690 | // Eagerly stop after the first pattern fires. |
| 691 | Changed = true; |
| 692 | break; |
| 693 | } |
| 694 | // Cleanup instructions of the alternative code sequence. There is no |
| 695 | // use for them. |
| 696 | MachineFunction *MF = MBB->getParent(); |
| 697 | for (auto *InstrPtr : InsInstrs) |
| 698 | MF->deleteMachineInstr(MI: InstrPtr); |
| 699 | } |
| 700 | InstrIdxForVirtReg.clear(); |
| 701 | } |
| 702 | } |
| 703 | |
| 704 | if (Changed && IncrementalUpdate) |
| 705 | Traces->invalidate(MBB); |
| 706 | return Changed; |
| 707 | } |
| 708 | |
| 709 | bool MachineCombinerImpl::run(MachineFunction &MF, MachineLoopInfo *MLI, |
| 710 | MachineTraceMetrics *Traces, |
| 711 | ProfileSummaryInfo *PSI, |
| 712 | MachineBlockFrequencyInfo *MBFI, |
| 713 | RegisterClassInfo *RegClassInfo) { |
| 714 | STI = &MF.getSubtarget(); |
| 715 | TII = STI->getInstrInfo(); |
| 716 | TRI = STI->getRegisterInfo(); |
| 717 | SchedModel = STI->getSchedModel(); |
| 718 | TSchedModel.init(TSInfo: STI); |
| 719 | MRI = &MF.getRegInfo(); |
| 720 | this->MLI = MLI; |
| 721 | this->Traces = Traces; |
| 722 | this->PSI = PSI; |
| 723 | this->MBFI = MBFI; |
| 724 | this->RegClassInfo = RegClassInfo; |
| 725 | TraceEnsemble = nullptr; |
| 726 | |
| 727 | LLVM_DEBUG(dbgs() << "Machine InstCombiner: " << MF.getName() << '\n'); |
| 728 | if (!TII->useMachineCombiner()) { |
| 729 | LLVM_DEBUG( |
| 730 | dbgs() |
| 731 | << " Skipping pass: Target does not support machine combiner\n" ); |
| 732 | return false; |
| 733 | } |
| 734 | |
| 735 | bool Changed = false; |
| 736 | |
| 737 | // Try to combine instructions. |
| 738 | for (auto &MBB : MF) |
| 739 | Changed |= combineInstructions(MBB: &MBB); |
| 740 | |
| 741 | return Changed; |
| 742 | } |
| 743 | |
| 744 | bool MachineCombinerLegacy::runOnMachineFunction(MachineFunction &MF) { |
| 745 | auto *MLI = &getAnalysis<MachineLoopInfoWrapperPass>().getLI(); |
| 746 | auto *Traces = &getAnalysis<MachineTraceMetricsWrapperPass>().getMTM(); |
| 747 | auto *PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI(); |
| 748 | auto *MBFI = (PSI && PSI->hasProfileSummary()) |
| 749 | ? &getAnalysis<LazyMachineBlockFrequencyInfoPass>().getBFI() |
| 750 | : nullptr; |
| 751 | auto &RegClassInfo = |
| 752 | getAnalysis<MachineRegisterClassInfoWrapperPass>().getRCI(); |
| 753 | return MachineCombinerImpl().run(MF, MLI, Traces, PSI, MBFI, RegClassInfo: &RegClassInfo); |
| 754 | } |
| 755 | |
| 756 | PreservedAnalyses |
| 757 | MachineCombinerPass::run(MachineFunction &MF, |
| 758 | MachineFunctionAnalysisManager &MFAM) { |
| 759 | MFPropsModifier _(*this, MF); |
| 760 | auto &MLI = MFAM.getResult<MachineLoopAnalysis>(IR&: MF); |
| 761 | auto &Traces = MFAM.getResult<MachineTraceMetricsAnalysis>(IR&: MF); |
| 762 | auto *PSI = MFAM.getResult<ModuleAnalysisManagerMachineFunctionProxy>(IR&: MF) |
| 763 | .getCachedResult<ProfileSummaryAnalysis>( |
| 764 | IR&: *MF.getFunction().getParent()); |
| 765 | auto *MBFI = (PSI && PSI->hasProfileSummary()) |
| 766 | ? &MFAM.getResult<MachineBlockFrequencyAnalysis>(IR&: MF) |
| 767 | : nullptr; |
| 768 | auto &RegClassInfo = MFAM.getResult<MachineRegisterClassAnalysis>(IR&: MF); |
| 769 | if (!MachineCombinerImpl().run(MF, MLI: &MLI, Traces: &Traces, PSI, MBFI, RegClassInfo: &RegClassInfo)) |
| 770 | return PreservedAnalyses::all(); |
| 771 | |
| 772 | auto PA = getMachineFunctionPassPreservedAnalyses(); |
| 773 | PA.preserveSet<CFGAnalyses>(); |
| 774 | PA.preserve<MachineTraceMetricsAnalysis>(); |
| 775 | return PA; |
| 776 | } |
| 777 | |