| 1 | //===- RegAllocGreedy.cpp - greedy register allocator ---------------------===// |
| 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 | // This file defines the RAGreedy function pass for register allocation in |
| 10 | // optimized builds. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #include "RegAllocGreedy.h" |
| 15 | #include "AllocationOrder.h" |
| 16 | #include "InlineSpiller.h" |
| 17 | #include "InterferenceCache.h" |
| 18 | #include "RegAllocBase.h" |
| 19 | #include "SplitKit.h" |
| 20 | #include "llvm/ADT/ArrayRef.h" |
| 21 | #include "llvm/ADT/BitVector.h" |
| 22 | #include "llvm/ADT/IndexedMap.h" |
| 23 | #include "llvm/ADT/SmallSet.h" |
| 24 | #include "llvm/ADT/SmallVector.h" |
| 25 | #include "llvm/ADT/Statistic.h" |
| 26 | #include "llvm/ADT/StringRef.h" |
| 27 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" |
| 28 | #include "llvm/CodeGen/CalcSpillWeights.h" |
| 29 | #include "llvm/CodeGen/EdgeBundles.h" |
| 30 | #include "llvm/CodeGen/LiveDebugVariables.h" |
| 31 | #include "llvm/CodeGen/LiveInterval.h" |
| 32 | #include "llvm/CodeGen/LiveIntervalUnion.h" |
| 33 | #include "llvm/CodeGen/LiveIntervals.h" |
| 34 | #include "llvm/CodeGen/LiveRangeEdit.h" |
| 35 | #include "llvm/CodeGen/LiveRegMatrix.h" |
| 36 | #include "llvm/CodeGen/LiveStacks.h" |
| 37 | #include "llvm/CodeGen/MachineBasicBlock.h" |
| 38 | #include "llvm/CodeGen/MachineBlockFrequencyInfo.h" |
| 39 | #include "llvm/CodeGen/MachineDominators.h" |
| 40 | #include "llvm/CodeGen/MachineFrameInfo.h" |
| 41 | #include "llvm/CodeGen/MachineFunction.h" |
| 42 | #include "llvm/CodeGen/MachineFunctionPass.h" |
| 43 | #include "llvm/CodeGen/MachineInstr.h" |
| 44 | #include "llvm/CodeGen/MachineLoopInfo.h" |
| 45 | #include "llvm/CodeGen/MachineOperand.h" |
| 46 | #include "llvm/CodeGen/MachineOptimizationRemarkEmitter.h" |
| 47 | #include "llvm/CodeGen/MachinePassManager.h" |
| 48 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
| 49 | #include "llvm/CodeGen/RegAllocEvictionAdvisor.h" |
| 50 | #include "llvm/CodeGen/RegAllocGreedyPass.h" |
| 51 | #include "llvm/CodeGen/RegAllocPriorityAdvisor.h" |
| 52 | #include "llvm/CodeGen/RegAllocRegistry.h" |
| 53 | #include "llvm/CodeGen/RegisterClassInfo.h" |
| 54 | #include "llvm/CodeGen/SlotIndexes.h" |
| 55 | #include "llvm/CodeGen/SpillPlacement.h" |
| 56 | #include "llvm/CodeGen/Spiller.h" |
| 57 | #include "llvm/CodeGen/TargetInstrInfo.h" |
| 58 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
| 59 | #include "llvm/CodeGen/TargetSubtargetInfo.h" |
| 60 | #include "llvm/CodeGen/VirtRegMap.h" |
| 61 | #include "llvm/IR/Analysis.h" |
| 62 | #include "llvm/IR/DebugInfoMetadata.h" |
| 63 | #include "llvm/IR/Function.h" |
| 64 | #include "llvm/IR/LLVMContext.h" |
| 65 | #include "llvm/IR/PassTimingInfo.h" |
| 66 | #include "llvm/Pass.h" |
| 67 | #include "llvm/Support/BlockFrequency.h" |
| 68 | #include "llvm/Support/BranchProbability.h" |
| 69 | #include "llvm/Support/CommandLine.h" |
| 70 | #include "llvm/Support/Debug.h" |
| 71 | #include "llvm/Support/MathExtras.h" |
| 72 | #include "llvm/Support/Timer.h" |
| 73 | #include "llvm/Support/raw_ostream.h" |
| 74 | #include <algorithm> |
| 75 | #include <cassert> |
| 76 | #include <cstdint> |
| 77 | #include <utility> |
| 78 | |
| 79 | using namespace llvm; |
| 80 | |
| 81 | #define DEBUG_TYPE "regalloc" |
| 82 | |
| 83 | STATISTIC(NumGlobalSplits, "Number of split global live ranges" ); |
| 84 | STATISTIC(NumLocalSplits, "Number of split local live ranges" ); |
| 85 | STATISTIC(NumEvicted, "Number of interferences evicted" ); |
| 86 | |
| 87 | static cl::opt<SplitEditor::ComplementSpillMode> SplitSpillMode( |
| 88 | "split-spill-mode" , cl::Hidden, |
| 89 | cl::desc("Spill mode for splitting live ranges" ), |
| 90 | cl::values(clEnumValN(SplitEditor::SM_Partition, "default" , "Default" ), |
| 91 | clEnumValN(SplitEditor::SM_Size, "size" , "Optimize for size" ), |
| 92 | clEnumValN(SplitEditor::SM_Speed, "speed" , "Optimize for speed" )), |
| 93 | cl::init(Val: SplitEditor::SM_Speed)); |
| 94 | |
| 95 | static cl::opt<unsigned> |
| 96 | LastChanceRecoloringMaxDepth("lcr-max-depth" , cl::Hidden, |
| 97 | cl::desc("Last chance recoloring max depth" ), |
| 98 | cl::init(Val: 5)); |
| 99 | |
| 100 | static cl::opt<unsigned> LastChanceRecoloringMaxInterference( |
| 101 | "lcr-max-interf" , cl::Hidden, |
| 102 | cl::desc("Last chance recoloring maximum number of considered" |
| 103 | " interference at a time" ), |
| 104 | cl::init(Val: 8)); |
| 105 | |
| 106 | static cl::opt<bool> ExhaustiveSearch( |
| 107 | "exhaustive-register-search" , cl::NotHidden, |
| 108 | cl::desc("Exhaustive Search for registers bypassing the depth " |
| 109 | "and interference cutoffs of last chance recoloring" ), |
| 110 | cl::Hidden); |
| 111 | |
| 112 | static cl::opt<bool> CompactSlotIndexes( |
| 113 | "greedy-compact-slot-indexes" , cl::Hidden, cl::init(Val: false), |
| 114 | cl::desc("Drop SlotIndexes entries for erased instructions before " |
| 115 | "allocating, so live range sizes reflect real code distance" )); |
| 116 | |
| 117 | // This option should be deprecated! |
| 118 | // FIXME: Find a good default for this flag and remove the flag. |
| 119 | static cl::opt<unsigned> |
| 120 | CSRFirstTimeCost("regalloc-csr-first-time-cost" , |
| 121 | cl::desc("Cost for first time use of callee-saved register." ), |
| 122 | cl::init(Val: 0), cl::Hidden); |
| 123 | |
| 124 | static cl::opt<unsigned> CSRCostScale( |
| 125 | "regalloc-csr-cost-scale" , |
| 126 | cl::desc("Scale for the callee-saved register cost, in percentage." ), |
| 127 | cl::init(Val: 80), cl::Hidden); |
| 128 | |
| 129 | static cl::opt<unsigned long> GrowRegionComplexityBudget( |
| 130 | "grow-region-complexity-budget" , |
| 131 | cl::desc("growRegion() does not scale with the number of BB edges, so " |
| 132 | "limit its budget and bail out once we reach the limit." ), |
| 133 | cl::init(Val: 10000), cl::Hidden); |
| 134 | |
| 135 | static cl::opt<bool> GreedyRegClassPriorityTrumpsGlobalness( |
| 136 | "greedy-regclass-priority-trumps-globalness" , |
| 137 | cl::desc("Change the greedy register allocator's live range priority " |
| 138 | "calculation to make the AllocationPriority of the register class " |
| 139 | "more important then whether the range is global" ), |
| 140 | cl::Hidden); |
| 141 | |
| 142 | static cl::opt<bool> GreedyReverseLocalAssignment( |
| 143 | "greedy-reverse-local-assignment" , |
| 144 | cl::desc("Reverse allocation order of local live ranges, such that " |
| 145 | "shorter local live ranges will tend to be allocated first" ), |
| 146 | cl::Hidden); |
| 147 | |
| 148 | static cl::opt<unsigned> SplitThresholdForRegWithHint( |
| 149 | "split-threshold-for-reg-with-hint" , |
| 150 | cl::desc("The threshold for splitting a virtual register with a hint, in " |
| 151 | "percentage" ), |
| 152 | cl::init(Val: 75), cl::Hidden); |
| 153 | |
| 154 | static RegisterRegAlloc greedyRegAlloc("greedy" , "greedy register allocator" , |
| 155 | createGreedyRegisterAllocator); |
| 156 | |
| 157 | namespace { |
| 158 | class RAGreedyLegacy : public MachineFunctionPass { |
| 159 | RegAllocFilterFunc F; |
| 160 | |
| 161 | public: |
| 162 | RAGreedyLegacy(const RegAllocFilterFunc F = nullptr); |
| 163 | |
| 164 | static char ID; |
| 165 | /// Return the pass name. |
| 166 | StringRef getPassName() const override { return "Greedy Register Allocator" ; } |
| 167 | |
| 168 | /// RAGreedy analysis usage. |
| 169 | void getAnalysisUsage(AnalysisUsage &AU) const override; |
| 170 | /// Perform register allocation. |
| 171 | bool runOnMachineFunction(MachineFunction &mf) override; |
| 172 | |
| 173 | MachineFunctionProperties getRequiredProperties() const override { |
| 174 | return MachineFunctionProperties().setNoPHIs(); |
| 175 | } |
| 176 | |
| 177 | MachineFunctionProperties getClearedProperties() const override { |
| 178 | return MachineFunctionProperties().setIsSSA(); |
| 179 | } |
| 180 | }; |
| 181 | |
| 182 | } // end anonymous namespace |
| 183 | |
| 184 | RAGreedyLegacy::RAGreedyLegacy(const RegAllocFilterFunc F) |
| 185 | : MachineFunctionPass(ID), F(std::move(F)) {} |
| 186 | |
| 187 | struct RAGreedy::RequiredAnalyses { |
| 188 | VirtRegMap *VRM = nullptr; |
| 189 | LiveIntervals *LIS = nullptr; |
| 190 | LiveRegMatrix *LRM = nullptr; |
| 191 | SlotIndexes *Indexes = nullptr; |
| 192 | MachineBlockFrequencyInfo *MBFI = nullptr; |
| 193 | MachineDominatorTree *DomTree = nullptr; |
| 194 | MachineLoopInfo *Loops = nullptr; |
| 195 | MachineOptimizationRemarkEmitter *ORE = nullptr; |
| 196 | EdgeBundles *Bundles = nullptr; |
| 197 | SpillPlacement *SpillPlacer = nullptr; |
| 198 | LiveDebugVariables *DebugVars = nullptr; |
| 199 | RegisterClassInfo *RCI = nullptr; |
| 200 | |
| 201 | // Used by InlineSpiller |
| 202 | LiveStacks *LSS; |
| 203 | // Proxies for eviction and priority advisors |
| 204 | RegAllocEvictionAdvisorProvider *EvictProvider; |
| 205 | RegAllocPriorityAdvisorProvider *PriorityProvider; |
| 206 | |
| 207 | RequiredAnalyses() = delete; |
| 208 | RequiredAnalyses(Pass &P); |
| 209 | RequiredAnalyses(MachineFunction &MF, MachineFunctionAnalysisManager &MFAM); |
| 210 | }; |
| 211 | |
| 212 | RAGreedy::RAGreedy(RequiredAnalyses &Analyses, const RegAllocFilterFunc F) |
| 213 | : RegAllocBase(F) { |
| 214 | VRM = Analyses.VRM; |
| 215 | LIS = Analyses.LIS; |
| 216 | Matrix = Analyses.LRM; |
| 217 | Indexes = Analyses.Indexes; |
| 218 | MBFI = Analyses.MBFI; |
| 219 | DomTree = Analyses.DomTree; |
| 220 | Loops = Analyses.Loops; |
| 221 | ORE = Analyses.ORE; |
| 222 | Bundles = Analyses.Bundles; |
| 223 | SpillPlacer = Analyses.SpillPlacer; |
| 224 | DebugVars = Analyses.DebugVars; |
| 225 | LSS = Analyses.LSS; |
| 226 | EvictProvider = Analyses.EvictProvider; |
| 227 | PriorityProvider = Analyses.PriorityProvider; |
| 228 | RegClassInfo = Analyses.RCI; |
| 229 | } |
| 230 | |
| 231 | void RAGreedyPass::printPipeline( |
| 232 | raw_ostream &OS, |
| 233 | function_ref<StringRef(StringRef)> MapClassName2PassName) const { |
| 234 | StringRef FilterName = Opts.FilterName.empty() ? "all" : Opts.FilterName; |
| 235 | OS << "greedy<" << FilterName << '>'; |
| 236 | } |
| 237 | |
| 238 | RAGreedy::RequiredAnalyses::RequiredAnalyses( |
| 239 | MachineFunction &MF, MachineFunctionAnalysisManager &MFAM) { |
| 240 | LIS = &MFAM.getResult<LiveIntervalsAnalysis>(IR&: MF); |
| 241 | LRM = &MFAM.getResult<LiveRegMatrixAnalysis>(IR&: MF); |
| 242 | LSS = &MFAM.getResult<LiveStacksAnalysis>(IR&: MF); |
| 243 | Indexes = &MFAM.getResult<SlotIndexesAnalysis>(IR&: MF); |
| 244 | MBFI = &MFAM.getResult<MachineBlockFrequencyAnalysis>(IR&: MF); |
| 245 | DomTree = &MFAM.getResult<MachineDominatorTreeAnalysis>(IR&: MF); |
| 246 | ORE = &MFAM.getResult<MachineOptimizationRemarkEmitterAnalysis>(IR&: MF); |
| 247 | Loops = &MFAM.getResult<MachineLoopAnalysis>(IR&: MF); |
| 248 | Bundles = &MFAM.getResult<EdgeBundlesAnalysis>(IR&: MF); |
| 249 | SpillPlacer = &MFAM.getResult<SpillPlacementAnalysis>(IR&: MF); |
| 250 | DebugVars = &MFAM.getResult<LiveDebugVariablesAnalysis>(IR&: MF); |
| 251 | EvictProvider = MFAM.getResult<RegAllocEvictionAdvisorAnalysis>(IR&: MF).Provider; |
| 252 | PriorityProvider = |
| 253 | MFAM.getResult<RegAllocPriorityAdvisorAnalysis>(IR&: MF).Provider; |
| 254 | VRM = &MFAM.getResult<VirtRegMapAnalysis>(IR&: MF); |
| 255 | RCI = &MFAM.getResult<MachineRegisterClassAnalysis>(IR&: MF); |
| 256 | } |
| 257 | |
| 258 | PreservedAnalyses RAGreedyPass::run(MachineFunction &MF, |
| 259 | MachineFunctionAnalysisManager &MFAM) { |
| 260 | MFPropsModifier _(*this, MF); |
| 261 | |
| 262 | RAGreedy::RequiredAnalyses Analyses(MF, MFAM); |
| 263 | RAGreedy Impl(Analyses, Opts.Filter); |
| 264 | |
| 265 | bool Changed = Impl.run(mf&: MF); |
| 266 | if (!Changed) |
| 267 | return PreservedAnalyses::all(); |
| 268 | auto PA = getMachineFunctionPassPreservedAnalyses(); |
| 269 | PA.preserveSet<CFGAnalyses>(); |
| 270 | PA.preserve<LiveIntervalsAnalysis>(); |
| 271 | PA.preserve<SlotIndexesAnalysis>(); |
| 272 | PA.preserve<LiveDebugVariablesAnalysis>(); |
| 273 | PA.preserve<LiveStacksAnalysis>(); |
| 274 | PA.preserve<VirtRegMapAnalysis>(); |
| 275 | PA.preserve<LiveRegMatrixAnalysis>(); |
| 276 | return PA; |
| 277 | } |
| 278 | |
| 279 | RAGreedy::RequiredAnalyses::RequiredAnalyses(Pass &P) { |
| 280 | VRM = &P.getAnalysis<VirtRegMapWrapperLegacy>().getVRM(); |
| 281 | LIS = &P.getAnalysis<LiveIntervalsWrapperPass>().getLIS(); |
| 282 | LSS = &P.getAnalysis<LiveStacksWrapperLegacy>().getLS(); |
| 283 | LRM = &P.getAnalysis<LiveRegMatrixWrapperLegacy>().getLRM(); |
| 284 | Indexes = &P.getAnalysis<SlotIndexesWrapperPass>().getSI(); |
| 285 | MBFI = &P.getAnalysis<MachineBlockFrequencyInfoWrapperPass>().getMBFI(); |
| 286 | DomTree = &P.getAnalysis<MachineDominatorTreeWrapperPass>().getDomTree(); |
| 287 | ORE = &P.getAnalysis<MachineOptimizationRemarkEmitterPass>().getORE(); |
| 288 | Loops = &P.getAnalysis<MachineLoopInfoWrapperPass>().getLI(); |
| 289 | Bundles = &P.getAnalysis<EdgeBundlesWrapperLegacy>().getEdgeBundles(); |
| 290 | SpillPlacer = &P.getAnalysis<SpillPlacementWrapperLegacy>().getResult(); |
| 291 | DebugVars = &P.getAnalysis<LiveDebugVariablesWrapperLegacy>().getLDV(); |
| 292 | EvictProvider = |
| 293 | &P.getAnalysis<RegAllocEvictionAdvisorAnalysisLegacy>().getProvider(); |
| 294 | PriorityProvider = |
| 295 | &P.getAnalysis<RegAllocPriorityAdvisorAnalysisLegacy>().getProvider(); |
| 296 | RCI = &P.getAnalysis<MachineRegisterClassInfoWrapperPass>().getRCI(); |
| 297 | } |
| 298 | |
| 299 | bool RAGreedyLegacy::runOnMachineFunction(MachineFunction &MF) { |
| 300 | RAGreedy::RequiredAnalyses Analyses(*this); |
| 301 | RAGreedy Impl(Analyses, F); |
| 302 | return Impl.run(mf&: MF); |
| 303 | } |
| 304 | |
| 305 | char RAGreedyLegacy::ID = 0; |
| 306 | char &llvm::RAGreedyLegacyID = RAGreedyLegacy::ID; |
| 307 | |
| 308 | INITIALIZE_PASS_BEGIN(RAGreedyLegacy, "greedy" , "Greedy Register Allocator" , |
| 309 | false, false) |
| 310 | INITIALIZE_PASS_DEPENDENCY(LiveDebugVariablesWrapperLegacy) |
| 311 | INITIALIZE_PASS_DEPENDENCY(SlotIndexesWrapperPass) |
| 312 | INITIALIZE_PASS_DEPENDENCY(LiveIntervalsWrapperPass) |
| 313 | INITIALIZE_PASS_DEPENDENCY(RegisterCoalescerLegacy) |
| 314 | INITIALIZE_PASS_DEPENDENCY(MachineSchedulerLegacy) |
| 315 | INITIALIZE_PASS_DEPENDENCY(LiveStacksWrapperLegacy) |
| 316 | INITIALIZE_PASS_DEPENDENCY(MachineDominatorTreeWrapperPass) |
| 317 | INITIALIZE_PASS_DEPENDENCY(MachineLoopInfoWrapperPass) |
| 318 | INITIALIZE_PASS_DEPENDENCY(VirtRegMapWrapperLegacy) |
| 319 | INITIALIZE_PASS_DEPENDENCY(LiveRegMatrixWrapperLegacy) |
| 320 | INITIALIZE_PASS_DEPENDENCY(EdgeBundlesWrapperLegacy) |
| 321 | INITIALIZE_PASS_DEPENDENCY(SpillPlacementWrapperLegacy) |
| 322 | INITIALIZE_PASS_DEPENDENCY(MachineOptimizationRemarkEmitterPass) |
| 323 | INITIALIZE_PASS_DEPENDENCY(MachineRegisterClassInfoWrapperPass) |
| 324 | INITIALIZE_PASS_DEPENDENCY(RegAllocEvictionAdvisorAnalysisLegacy) |
| 325 | INITIALIZE_PASS_DEPENDENCY(RegAllocPriorityAdvisorAnalysisLegacy) |
| 326 | INITIALIZE_PASS_END(RAGreedyLegacy, "greedy" , "Greedy Register Allocator" , |
| 327 | false, false) |
| 328 | |
| 329 | #ifndef NDEBUG |
| 330 | const char *const RAGreedy::StageName[] = { |
| 331 | "RS_New" , |
| 332 | "RS_Assign" , |
| 333 | "RS_Split" , |
| 334 | "RS_Split2" , |
| 335 | "RS_Spill" , |
| 336 | "RS_Done" |
| 337 | }; |
| 338 | #endif |
| 339 | |
| 340 | // Hysteresis to use when comparing floats. |
| 341 | // This helps stabilize decisions based on float comparisons. |
| 342 | const float Hysteresis = (2007 / 2048.0f); // 0.97998046875 |
| 343 | |
| 344 | FunctionPass* llvm::createGreedyRegisterAllocator() { |
| 345 | return new RAGreedyLegacy(); |
| 346 | } |
| 347 | |
| 348 | FunctionPass *llvm::createGreedyRegisterAllocator(RegAllocFilterFunc Ftor) { |
| 349 | return new RAGreedyLegacy(Ftor); |
| 350 | } |
| 351 | |
| 352 | void RAGreedyLegacy::getAnalysisUsage(AnalysisUsage &AU) const { |
| 353 | AU.setPreservesCFG(); |
| 354 | AU.addRequired<MachineBlockFrequencyInfoWrapperPass>(); |
| 355 | AU.addRequired<LiveIntervalsWrapperPass>(); |
| 356 | AU.addPreserved<LiveIntervalsWrapperPass>(); |
| 357 | AU.addRequired<SlotIndexesWrapperPass>(); |
| 358 | AU.addPreserved<SlotIndexesWrapperPass>(); |
| 359 | AU.addRequired<LiveDebugVariablesWrapperLegacy>(); |
| 360 | AU.addPreserved<LiveDebugVariablesWrapperLegacy>(); |
| 361 | AU.addRequired<LiveStacksWrapperLegacy>(); |
| 362 | AU.addPreserved<LiveStacksWrapperLegacy>(); |
| 363 | AU.addRequired<MachineDominatorTreeWrapperPass>(); |
| 364 | AU.addRequired<MachineLoopInfoWrapperPass>(); |
| 365 | AU.addRequired<VirtRegMapWrapperLegacy>(); |
| 366 | AU.addPreserved<VirtRegMapWrapperLegacy>(); |
| 367 | AU.addRequired<LiveRegMatrixWrapperLegacy>(); |
| 368 | AU.addPreserved<LiveRegMatrixWrapperLegacy>(); |
| 369 | AU.addRequired<EdgeBundlesWrapperLegacy>(); |
| 370 | AU.addRequired<SpillPlacementWrapperLegacy>(); |
| 371 | AU.addRequired<MachineOptimizationRemarkEmitterPass>(); |
| 372 | AU.addRequired<MachineRegisterClassInfoWrapperPass>(); |
| 373 | AU.addRequired<RegAllocEvictionAdvisorAnalysisLegacy>(); |
| 374 | AU.addRequired<RegAllocPriorityAdvisorAnalysisLegacy>(); |
| 375 | MachineFunctionPass::getAnalysisUsage(AU); |
| 376 | } |
| 377 | |
| 378 | //===----------------------------------------------------------------------===// |
| 379 | // LiveRangeEdit delegate methods |
| 380 | //===----------------------------------------------------------------------===// |
| 381 | |
| 382 | bool RAGreedy::LRE_CanEraseVirtReg(Register VirtReg) { |
| 383 | LiveInterval &LI = LIS->getInterval(Reg: VirtReg); |
| 384 | if (VRM->hasPhys(virtReg: VirtReg)) { |
| 385 | Matrix->unassign(VirtReg: LI); |
| 386 | aboutToRemoveInterval(LI); |
| 387 | return true; |
| 388 | } |
| 389 | // Unassigned virtreg is probably in the priority queue. |
| 390 | // RegAllocBase will erase it after dequeueing. |
| 391 | // Nonetheless, clear the live-range so that the debug |
| 392 | // dump will show the right state for that VirtReg. |
| 393 | LI.clear(); |
| 394 | return false; |
| 395 | } |
| 396 | |
| 397 | void RAGreedy::LRE_WillShrinkVirtReg(Register VirtReg) { |
| 398 | if (!VRM->hasPhys(virtReg: VirtReg)) |
| 399 | return; |
| 400 | |
| 401 | // Register is assigned, put it back on the queue for reassignment. |
| 402 | LiveInterval &LI = LIS->getInterval(Reg: VirtReg); |
| 403 | Matrix->unassign(VirtReg: LI); |
| 404 | RegAllocBase::enqueue(LI: &LI); |
| 405 | } |
| 406 | |
| 407 | void RAGreedy::LRE_DidCloneVirtReg(Register New, Register Old) { |
| 408 | ExtraInfo->LRE_DidCloneVirtReg(New, Old); |
| 409 | } |
| 410 | |
| 411 | void RAGreedy::ExtraRegInfo::(Register New, Register Old) { |
| 412 | // Cloning a register we haven't even heard about yet? Just ignore it. |
| 413 | if (!Info.inBounds(N: Old)) |
| 414 | return; |
| 415 | |
| 416 | // LRE may clone a virtual register because dead code elimination causes it to |
| 417 | // be split into connected components. The new components are much smaller |
| 418 | // than the original, so they should get a new chance at being assigned. |
| 419 | // same stage as the parent. |
| 420 | Info[Old].Stage = RS_Assign; |
| 421 | Info.grow(N: New.id()); |
| 422 | Info[New] = Info[Old]; |
| 423 | } |
| 424 | |
| 425 | void RAGreedy::releaseMemory() { |
| 426 | SpillerInstance.reset(); |
| 427 | GlobalCand.clear(); |
| 428 | } |
| 429 | |
| 430 | void RAGreedy::enqueueImpl(const LiveInterval *LI) { enqueue(CurQueue&: Queue, LI); } |
| 431 | |
| 432 | void RAGreedy::enqueue(PQueue &CurQueue, const LiveInterval *LI) { |
| 433 | // Prioritize live ranges by size, assigning larger ranges first. |
| 434 | // The queue holds (size, reg) pairs. |
| 435 | const Register Reg = LI->reg(); |
| 436 | assert(Reg.isVirtual() && "Can only enqueue virtual registers" ); |
| 437 | |
| 438 | auto Stage = ExtraInfo->getOrInitStage(Reg); |
| 439 | if (Stage == RS_New) { |
| 440 | Stage = RS_Assign; |
| 441 | ExtraInfo->setStage(Reg, Stage); |
| 442 | } |
| 443 | |
| 444 | unsigned Ret = PriorityAdvisor->getPriority(LI: *LI); |
| 445 | |
| 446 | // The virtual register number is a tie breaker for same-sized ranges. |
| 447 | // Give lower vreg numbers higher priority to assign them first. |
| 448 | CurQueue.push(x: std::make_pair(x&: Ret, y: ~Reg.id())); |
| 449 | } |
| 450 | |
| 451 | unsigned DefaultPriorityAdvisor::getPriority(const LiveInterval &LI) const { |
| 452 | const unsigned Size = LI.getSize(); |
| 453 | const Register Reg = LI.reg(); |
| 454 | unsigned Prio; |
| 455 | LiveRangeStage Stage = RA.getExtraInfo().getStage(VirtReg: LI); |
| 456 | |
| 457 | if (Stage == RS_Split) { |
| 458 | // Unsplit ranges that couldn't be allocated immediately are deferred until |
| 459 | // everything else has been allocated. |
| 460 | Prio = Size; |
| 461 | } else { |
| 462 | // Giant live ranges fall back to the global assignment heuristic, which |
| 463 | // prevents excessive spilling in pathological cases. |
| 464 | const TargetRegisterClass &RC = *MRI->getRegClass(Reg); |
| 465 | bool ForceGlobal = RC.GlobalPriority || |
| 466 | (!ReverseLocalAssignment && |
| 467 | (Size / SlotIndex::InstrDist) > |
| 468 | (2 * RegClassInfo.getNumAllocatableRegs(RC: &RC))); |
| 469 | unsigned GlobalBit = 0; |
| 470 | |
| 471 | if (Stage == RS_Assign && !ForceGlobal && !LI.empty() && |
| 472 | LIS->intervalIsInOneMBB(LI)) { |
| 473 | // Allocate original local ranges in linear instruction order. Since they |
| 474 | // are singly defined, this produces optimal coloring in the absence of |
| 475 | // global interference and other constraints. |
| 476 | if (!ReverseLocalAssignment) |
| 477 | Prio = LI.beginIndex().getApproxInstrDistance(other: Indexes->getLastIndex()); |
| 478 | else { |
| 479 | // Allocating bottom up may allow many short LRGs to be assigned first |
| 480 | // to one of the cheap registers. This could be much faster for very |
| 481 | // large blocks on targets with many physical registers. |
| 482 | Prio = Indexes->getZeroIndex().getApproxInstrDistance(other: LI.endIndex()); |
| 483 | } |
| 484 | } else { |
| 485 | // Allocate global and split ranges in long->short order. Long ranges that |
| 486 | // don't fit should be spilled (or split) ASAP so they don't create |
| 487 | // interference. Mark a bit to prioritize global above local ranges. |
| 488 | Prio = Size; |
| 489 | GlobalBit = 1; |
| 490 | } |
| 491 | |
| 492 | // Priority bit layout: |
| 493 | // 31 RS_Assign priority |
| 494 | // 30 Preference priority |
| 495 | // if (RegClassPriorityTrumpsGlobalness) |
| 496 | // 29-25 AllocPriority |
| 497 | // 24 GlobalBit |
| 498 | // else |
| 499 | // 29 Global bit |
| 500 | // 28-24 AllocPriority |
| 501 | // 0-23 Size/Instr distance |
| 502 | |
| 503 | // Clamp the size to fit with the priority masking scheme |
| 504 | Prio = std::min(a: Prio, b: (unsigned)maxUIntN(N: 24)); |
| 505 | assert(isUInt<5>(RC.AllocationPriority) && "allocation priority overflow" ); |
| 506 | |
| 507 | if (RegClassPriorityTrumpsGlobalness) |
| 508 | Prio |= RC.AllocationPriority << 25 | GlobalBit << 24; |
| 509 | else |
| 510 | Prio |= GlobalBit << 29 | RC.AllocationPriority << 24; |
| 511 | |
| 512 | // Mark a higher bit to prioritize global and local above RS_Split. |
| 513 | Prio |= (1u << 31); |
| 514 | |
| 515 | // Boost ranges that have a physical register hint. |
| 516 | if (VRM->hasKnownPreference(VirtReg: Reg)) |
| 517 | Prio |= (1u << 30); |
| 518 | } |
| 519 | |
| 520 | return Prio; |
| 521 | } |
| 522 | |
| 523 | unsigned DummyPriorityAdvisor::getPriority(const LiveInterval &LI) const { |
| 524 | // Prioritize by virtual register number, lowest first. |
| 525 | Register Reg = LI.reg(); |
| 526 | return ~Reg.virtRegIndex(); |
| 527 | } |
| 528 | |
| 529 | const LiveInterval *RAGreedy::dequeue() { return dequeue(CurQueue&: Queue); } |
| 530 | |
| 531 | const LiveInterval *RAGreedy::dequeue(PQueue &CurQueue) { |
| 532 | if (CurQueue.empty()) |
| 533 | return nullptr; |
| 534 | LiveInterval *LI = &LIS->getInterval(Reg: ~CurQueue.top().second); |
| 535 | CurQueue.pop(); |
| 536 | return LI; |
| 537 | } |
| 538 | |
| 539 | //===----------------------------------------------------------------------===// |
| 540 | // Direct Assignment |
| 541 | //===----------------------------------------------------------------------===// |
| 542 | |
| 543 | /// tryAssign - Try to assign VirtReg to an available register. |
| 544 | MCRegister RAGreedy::tryAssign(const LiveInterval &VirtReg, |
| 545 | AllocationOrder &Order, |
| 546 | SmallVectorImpl<Register> &NewVRegs, |
| 547 | const SmallVirtRegSet &FixedRegisters) { |
| 548 | MCRegister PhysReg; |
| 549 | for (auto I = Order.begin(), E = Order.end(); I != E && !PhysReg; ++I) { |
| 550 | assert(*I); |
| 551 | if (!Matrix->checkInterference(VirtReg, PhysReg: *I)) { |
| 552 | if (I.isHint()) |
| 553 | return *I; |
| 554 | else |
| 555 | PhysReg = *I; |
| 556 | } |
| 557 | } |
| 558 | if (!PhysReg.isValid()) |
| 559 | return PhysReg; |
| 560 | |
| 561 | // PhysReg is available, but there may be a better choice. |
| 562 | |
| 563 | // If we missed a simple hint, try to cheaply evict interference from the |
| 564 | // preferred register. |
| 565 | if (Register Hint = MRI->getSimpleHint(VReg: VirtReg.reg())) |
| 566 | if (Order.isHint(Reg: Hint)) { |
| 567 | MCRegister PhysHint = Hint.asMCReg(); |
| 568 | LLVM_DEBUG(dbgs() << "missed hint " << printReg(PhysHint, TRI) << '\n'); |
| 569 | |
| 570 | if (EvictAdvisor->canEvictHintInterference(VirtReg, PhysReg: PhysHint, |
| 571 | FixedRegisters)) { |
| 572 | evictInterference(VirtReg, PhysHint, NewVRegs); |
| 573 | return PhysHint; |
| 574 | } |
| 575 | |
| 576 | // We can also split the virtual register in cold blocks. |
| 577 | if (trySplitAroundHintReg(Hint: PhysHint, VirtReg, NewVRegs, Order)) |
| 578 | return MCRegister(); |
| 579 | |
| 580 | // Record the missed hint, we may be able to recover |
| 581 | // at the end if the surrounding allocation changed. |
| 582 | SetOfBrokenHints.insert(X: &VirtReg); |
| 583 | } |
| 584 | |
| 585 | // Try to evict interference from a cheaper alternative. |
| 586 | uint8_t Cost = RegCosts[PhysReg.id()]; |
| 587 | |
| 588 | // Most registers have 0 additional cost. |
| 589 | if (!Cost) |
| 590 | return PhysReg; |
| 591 | |
| 592 | LLVM_DEBUG(dbgs() << printReg(PhysReg, TRI) << " is available at cost " |
| 593 | << (unsigned)Cost << '\n'); |
| 594 | MCRegister CheapReg = tryEvict(VirtReg, Order, NewVRegs, Cost, FixedRegisters); |
| 595 | return CheapReg ? CheapReg : PhysReg; |
| 596 | } |
| 597 | |
| 598 | //===----------------------------------------------------------------------===// |
| 599 | // Interference eviction |
| 600 | //===----------------------------------------------------------------------===// |
| 601 | |
| 602 | bool RegAllocEvictionAdvisor::canReassign(const LiveInterval &VirtReg, |
| 603 | MCRegister FromReg) const { |
| 604 | auto HasRegUnitInterference = [&](MCRegUnit Unit) { |
| 605 | // Instantiate a "subquery", not to be confused with the Queries array. |
| 606 | LiveIntervalUnion::Query SubQ( |
| 607 | VirtReg, Matrix->getLiveUnions()[static_cast<unsigned>(Unit)]); |
| 608 | return SubQ.checkInterference(); |
| 609 | }; |
| 610 | |
| 611 | for (MCRegister Reg : |
| 612 | AllocationOrder::create(VirtReg: VirtReg.reg(), VRM: *VRM, RegClassInfo, Matrix)) { |
| 613 | if (Reg == FromReg) |
| 614 | continue; |
| 615 | // If no units have interference, reassignment is possible. |
| 616 | if (none_of(Range: TRI->regunits(Reg), P: HasRegUnitInterference)) { |
| 617 | LLVM_DEBUG(dbgs() << "can reassign: " << VirtReg << " from " |
| 618 | << printReg(FromReg, TRI) << " to " |
| 619 | << printReg(Reg, TRI) << '\n'); |
| 620 | return true; |
| 621 | } |
| 622 | } |
| 623 | return false; |
| 624 | } |
| 625 | |
| 626 | /// evictInterference - Evict any interferring registers that prevent VirtReg |
| 627 | /// from being assigned to Physreg. This assumes that canEvictInterference |
| 628 | /// returned true. |
| 629 | void RAGreedy::evictInterference(const LiveInterval &VirtReg, |
| 630 | MCRegister PhysReg, |
| 631 | SmallVectorImpl<Register> &NewVRegs) { |
| 632 | // Make sure that VirtReg has a cascade number, and assign that cascade |
| 633 | // number to every evicted register. These live ranges than then only be |
| 634 | // evicted by a newer cascade, preventing infinite loops. |
| 635 | unsigned Cascade = ExtraInfo->getOrAssignNewCascade(Reg: VirtReg.reg()); |
| 636 | |
| 637 | LLVM_DEBUG(dbgs() << "evicting " << printReg(PhysReg, TRI) |
| 638 | << " interference: Cascade " << Cascade << '\n'); |
| 639 | |
| 640 | // Collect all interfering virtregs first. |
| 641 | SmallVector<const LiveInterval *, 8> Intfs; |
| 642 | for (MCRegUnit Unit : TRI->regunits(Reg: PhysReg)) { |
| 643 | LiveIntervalUnion::Query &Q = Matrix->query(LR: VirtReg, RegUnit: Unit); |
| 644 | // We usually have the interfering VRegs cached so collectInterferingVRegs() |
| 645 | // should be fast, we may need to recalculate if when different physregs |
| 646 | // overlap the same register unit so we had different SubRanges queried |
| 647 | // against it. |
| 648 | ArrayRef<const LiveInterval *> IVR = Q.interferingVRegs(); |
| 649 | Intfs.append(in_start: IVR.begin(), in_end: IVR.end()); |
| 650 | } |
| 651 | |
| 652 | // Evict them second. This will invalidate the queries. |
| 653 | for (const LiveInterval *Intf : Intfs) { |
| 654 | // The same VirtReg may be present in multiple RegUnits. Skip duplicates. |
| 655 | if (!VRM->hasPhys(virtReg: Intf->reg())) |
| 656 | continue; |
| 657 | |
| 658 | Matrix->unassign(VirtReg: *Intf); |
| 659 | assert((ExtraInfo->getCascade(Intf->reg()) < Cascade || |
| 660 | (Cascade < ExtraInfo->getCascade(Intf->reg()) && |
| 661 | EvictAdvisor->isUrgentEviction(VirtReg, *Intf)) || |
| 662 | VirtReg.isSpillable() < Intf->isSpillable()) && |
| 663 | "Cannot decrease cascade number, illegal eviction" ); |
| 664 | ExtraInfo->setCascade(Reg: Intf->reg(), Cascade); |
| 665 | ++NumEvicted; |
| 666 | NewVRegs.push_back(Elt: Intf->reg()); |
| 667 | } |
| 668 | } |
| 669 | |
| 670 | MCRegister |
| 671 | RegAllocEvictionAdvisor::getUnusedCalleeSavedReg(const LiveInterval &VirtReg, |
| 672 | MCRegister PhysReg) const { |
| 673 | auto IsUnusedCSR = [&](MCRegUnit Unit) { |
| 674 | MCRegister CSR = RegClassInfo.getCalleeSavedAlias(Unit); |
| 675 | if (CSR && !Matrix->isPhysRegUsed(PhysReg: CSR)) |
| 676 | return CSR; |
| 677 | return MCRegister(); |
| 678 | }; |
| 679 | |
| 680 | for (MCRegUnitMaskIterator Units(PhysReg, TRI); Units.isValid(); ++Units) { |
| 681 | MCRegUnit Unit = (*Units).first; |
| 682 | LaneBitmask Mask = (*Units).second; |
| 683 | if (VirtReg.hasSubRanges() && |
| 684 | llvm::none_of(Range: VirtReg.subranges(), |
| 685 | P: [&](const LiveInterval::SubRange &S) { |
| 686 | return (S.LaneMask & Mask).any(); |
| 687 | })) |
| 688 | continue; |
| 689 | if (MCRegister CSR = IsUnusedCSR(Unit)) |
| 690 | return CSR; |
| 691 | } |
| 692 | return MCRegister(); |
| 693 | } |
| 694 | |
| 695 | std::optional<unsigned> |
| 696 | RegAllocEvictionAdvisor::getOrderLimit(const LiveInterval &VirtReg, |
| 697 | const AllocationOrder &Order, |
| 698 | unsigned CostPerUseLimit) const { |
| 699 | unsigned OrderLimit = Order.getOrder().size(); |
| 700 | |
| 701 | if (CostPerUseLimit < uint8_t(~0u)) { |
| 702 | // Check of any registers in RC are below CostPerUseLimit. |
| 703 | const TargetRegisterClass *RC = MRI->getRegClass(Reg: VirtReg.reg()); |
| 704 | uint8_t MinCost = RegClassInfo.getMinCost(RC); |
| 705 | if (MinCost >= CostPerUseLimit) { |
| 706 | LLVM_DEBUG(dbgs() << TRI->getRegClassName(RC) << " minimum cost = " |
| 707 | << MinCost << ", no cheaper registers to be found.\n" ); |
| 708 | return std::nullopt; |
| 709 | } |
| 710 | |
| 711 | // It is normal for register classes to have a long tail of registers with |
| 712 | // the same cost. We don't need to look at them if they're too expensive. |
| 713 | // LastCostChange is an index into the original RegisterClassInfo order, so |
| 714 | // it cannot be used to shorten a custom order. |
| 715 | if (!Order.hasCustomOrder() && |
| 716 | RegCosts[Order.getOrder().back()] >= CostPerUseLimit) { |
| 717 | OrderLimit = RegClassInfo.getLastCostChange(RC); |
| 718 | LLVM_DEBUG(dbgs() << "Only trying the first " << OrderLimit |
| 719 | << " regs.\n" ); |
| 720 | } |
| 721 | } |
| 722 | return OrderLimit; |
| 723 | } |
| 724 | |
| 725 | bool RegAllocEvictionAdvisor::canAllocatePhysReg(const LiveInterval &VirtReg, |
| 726 | unsigned CostPerUseLimit, |
| 727 | MCRegister PhysReg) const { |
| 728 | if (RegCosts[PhysReg.id()] >= CostPerUseLimit) |
| 729 | return false; |
| 730 | // The first use of a callee-saved register in a function has cost 1. |
| 731 | // Don't start using a CSR when the CostPerUseLimit is low. |
| 732 | if (CostPerUseLimit == 1) { |
| 733 | if (MCRegister CSR = getUnusedCalleeSavedReg(VirtReg, PhysReg)) { |
| 734 | LLVM_DEBUG(dbgs() << printReg(PhysReg, TRI) << " would clobber CSR " |
| 735 | << printReg(CSR, TRI) << '\n'); |
| 736 | return false; |
| 737 | } |
| 738 | } |
| 739 | return true; |
| 740 | } |
| 741 | |
| 742 | /// tryEvict - Try to evict all interferences for a physreg. |
| 743 | /// @param VirtReg Currently unassigned virtual register. |
| 744 | /// @param Order Physregs to try. |
| 745 | /// @return Physreg to assign VirtReg, or 0. |
| 746 | MCRegister RAGreedy::tryEvict(const LiveInterval &VirtReg, |
| 747 | AllocationOrder &Order, |
| 748 | SmallVectorImpl<Register> &NewVRegs, |
| 749 | uint8_t CostPerUseLimit, |
| 750 | const SmallVirtRegSet &FixedRegisters) { |
| 751 | NamedRegionTimer T("evict" , "Evict" , TimerGroupName, TimerGroupDescription, |
| 752 | TimePassesIsEnabled); |
| 753 | |
| 754 | MCRegister BestPhys = EvictAdvisor->tryFindEvictionCandidate( |
| 755 | VirtReg, Order, CostPerUseLimit, FixedRegisters); |
| 756 | if (BestPhys.isValid()) |
| 757 | evictInterference(VirtReg, PhysReg: BestPhys, NewVRegs); |
| 758 | return BestPhys; |
| 759 | } |
| 760 | |
| 761 | //===----------------------------------------------------------------------===// |
| 762 | // Region Splitting |
| 763 | //===----------------------------------------------------------------------===// |
| 764 | |
| 765 | /// addSplitConstraints - Fill out the SplitConstraints vector based on the |
| 766 | /// interference pattern in Physreg and its aliases. Add the constraints to |
| 767 | /// SpillPlacement and return the static cost of this split in Cost, assuming |
| 768 | /// that all preferences in SplitConstraints are met. |
| 769 | /// Return false if there are no bundles with positive bias, or if the static |
| 770 | /// cost reaches MaxCost. |
| 771 | bool RAGreedy::addSplitConstraints(InterferenceCache::Cursor Intf, |
| 772 | BlockFrequency &Cost, |
| 773 | BlockFrequency MaxCost) { |
| 774 | ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks(); |
| 775 | |
| 776 | // Reset interference dependent info. |
| 777 | SplitConstraints.resize(N: UseBlocks.size()); |
| 778 | BlockFrequency StaticCost = BlockFrequency(0); |
| 779 | for (unsigned I = 0; I != UseBlocks.size(); ++I) { |
| 780 | const SplitAnalysis::BlockInfo &BI = UseBlocks[I]; |
| 781 | SpillPlacement::BlockConstraint &BC = SplitConstraints[I]; |
| 782 | |
| 783 | BC.Number = BI.MBB->getNumber(); |
| 784 | Intf.moveToBlock(MBBNum: BC.Number); |
| 785 | BC.Entry = BI.LiveIn ? SpillPlacement::PrefReg : SpillPlacement::DontCare; |
| 786 | BC.Exit = (BI.LiveOut && |
| 787 | !LIS->getInstructionFromIndex(index: BI.LastInstr)->isImplicitDef()) |
| 788 | ? SpillPlacement::PrefReg |
| 789 | : SpillPlacement::DontCare; |
| 790 | BC.ChangesValue = BI.FirstDef.isValid(); |
| 791 | |
| 792 | if (!Intf.hasInterference()) |
| 793 | continue; |
| 794 | |
| 795 | // Number of spill code instructions to insert. |
| 796 | unsigned Ins = 0; |
| 797 | |
| 798 | // Interference for the live-in value. |
| 799 | if (BI.LiveIn) { |
| 800 | if (Intf.first() <= Indexes->getMBBStartIdx(mbb: BI.MBB)) { |
| 801 | BC.Entry = SpillPlacement::MustSpill; |
| 802 | ++Ins; |
| 803 | } else if (Intf.first() < BI.FirstInstr) { |
| 804 | BC.Entry = SpillPlacement::PrefSpill; |
| 805 | ++Ins; |
| 806 | } else if (Intf.first() < BI.LastInstr) { |
| 807 | ++Ins; |
| 808 | } |
| 809 | |
| 810 | // Abort if the spill cannot be inserted at the MBB' start |
| 811 | if (((BC.Entry == SpillPlacement::MustSpill) || |
| 812 | (BC.Entry == SpillPlacement::PrefSpill)) && |
| 813 | SlotIndex::isEarlierInstr(A: BI.FirstInstr, |
| 814 | B: SA->getFirstSplitPoint(Num: BC.Number))) |
| 815 | return false; |
| 816 | } |
| 817 | |
| 818 | // Interference for the live-out value. |
| 819 | if (BI.LiveOut) { |
| 820 | if (Intf.last() >= SA->getLastSplitPoint(Num: BC.Number)) { |
| 821 | BC.Exit = SpillPlacement::MustSpill; |
| 822 | ++Ins; |
| 823 | } else if (Intf.last() > BI.LastInstr) { |
| 824 | BC.Exit = SpillPlacement::PrefSpill; |
| 825 | ++Ins; |
| 826 | } else if (Intf.last() > BI.FirstInstr) { |
| 827 | ++Ins; |
| 828 | } |
| 829 | } |
| 830 | |
| 831 | // Accumulate the total frequency of inserted spill code. |
| 832 | while (Ins--) |
| 833 | StaticCost += SpillPlacer->getBlockFrequency(Number: BC.Number); |
| 834 | |
| 835 | if (StaticCost >= MaxCost) |
| 836 | return false; |
| 837 | } |
| 838 | Cost = StaticCost; |
| 839 | |
| 840 | // Add constraints for use-blocks. Note that these are the only constraints |
| 841 | // that may add a positive bias, it is downhill from here. |
| 842 | SpillPlacer->addConstraints(LiveBlocks: SplitConstraints); |
| 843 | return SpillPlacer->scanActiveBundles(); |
| 844 | } |
| 845 | |
| 846 | /// addThroughConstraints - Add constraints and links to SpillPlacer from the |
| 847 | /// live-through blocks in Blocks. |
| 848 | bool RAGreedy::addThroughConstraints(InterferenceCache::Cursor Intf, |
| 849 | ArrayRef<unsigned> Blocks) { |
| 850 | const unsigned GroupSize = 8; |
| 851 | SpillPlacement::BlockConstraint BCS[GroupSize]; |
| 852 | unsigned TBS[GroupSize]; |
| 853 | unsigned B = 0, T = 0; |
| 854 | |
| 855 | for (unsigned Number : Blocks) { |
| 856 | Intf.moveToBlock(MBBNum: Number); |
| 857 | |
| 858 | if (!Intf.hasInterference()) { |
| 859 | assert(T < GroupSize && "Array overflow" ); |
| 860 | TBS[T] = Number; |
| 861 | if (++T == GroupSize) { |
| 862 | SpillPlacer->addLinks(Links: ArrayRef(TBS, T)); |
| 863 | T = 0; |
| 864 | } |
| 865 | continue; |
| 866 | } |
| 867 | |
| 868 | assert(B < GroupSize && "Array overflow" ); |
| 869 | BCS[B].Number = Number; |
| 870 | |
| 871 | // Abort if the spill cannot be inserted at the MBB' start |
| 872 | MachineBasicBlock *MBB = MF->getBlockNumbered(N: Number); |
| 873 | auto FirstNonDebugInstr = MBB->getFirstNonDebugInstr(); |
| 874 | if (FirstNonDebugInstr != MBB->end() && |
| 875 | SlotIndex::isEarlierInstr(A: LIS->getInstructionIndex(Instr: *FirstNonDebugInstr), |
| 876 | B: SA->getFirstSplitPoint(Num: Number))) |
| 877 | return false; |
| 878 | |
| 879 | // Interference for the live-in value. |
| 880 | Register Reg = SA->getParent().reg(); |
| 881 | auto InsertPt = MBB->SkipPHIsLabelsAndDebug(I: MBB->begin(), Reg); |
| 882 | SlotIndex InsertIdx = InsertPt == MBB->end() |
| 883 | ? Indexes->getMBBEndIdx(mbb: MBB) |
| 884 | : LIS->getInstructionIndex(Instr: *InsertPt); |
| 885 | if (Intf.first() <= Indexes->getMBBStartIdx(mbb: MBB) || |
| 886 | SlotIndex::isEarlierInstr(A: Intf.first(), B: InsertIdx)) |
| 887 | BCS[B].Entry = SpillPlacement::MustSpill; |
| 888 | else |
| 889 | BCS[B].Entry = SpillPlacement::PrefSpill; |
| 890 | |
| 891 | // Interference for the live-out value. |
| 892 | if (Intf.last() >= SA->getLastSplitPoint(Num: Number)) |
| 893 | BCS[B].Exit = SpillPlacement::MustSpill; |
| 894 | else |
| 895 | BCS[B].Exit = SpillPlacement::PrefSpill; |
| 896 | |
| 897 | if (++B == GroupSize) { |
| 898 | SpillPlacer->addConstraints(LiveBlocks: ArrayRef(BCS, B)); |
| 899 | B = 0; |
| 900 | } |
| 901 | } |
| 902 | |
| 903 | SpillPlacer->addConstraints(LiveBlocks: ArrayRef(BCS, B)); |
| 904 | SpillPlacer->addLinks(Links: ArrayRef(TBS, T)); |
| 905 | return true; |
| 906 | } |
| 907 | |
| 908 | bool RAGreedy::growRegion(GlobalSplitCandidate &Cand) { |
| 909 | // Keep track of through blocks that have not been added to SpillPlacer. |
| 910 | BitVector Todo = SA->getThroughBlocks(); |
| 911 | SmallVectorImpl<unsigned> &ActiveBlocks = Cand.ActiveBlocks; |
| 912 | unsigned AddedTo = 0; |
| 913 | #ifndef NDEBUG |
| 914 | unsigned Visited = 0; |
| 915 | #endif |
| 916 | |
| 917 | unsigned long Budget = GrowRegionComplexityBudget; |
| 918 | while (true) { |
| 919 | ArrayRef<unsigned> NewBundles = SpillPlacer->getRecentPositive(); |
| 920 | // Find new through blocks in the periphery of PrefRegBundles. |
| 921 | for (unsigned Bundle : NewBundles) { |
| 922 | // Look at all blocks connected to Bundle in the full graph. |
| 923 | ArrayRef<unsigned> Blocks = Bundles->getBlocks(Bundle); |
| 924 | // Limit compilation time by bailing out after we use all our budget. |
| 925 | if (Blocks.size() >= Budget) |
| 926 | return false; |
| 927 | Budget -= Blocks.size(); |
| 928 | for (unsigned Block : Blocks) { |
| 929 | if (!Todo.test(Idx: Block)) |
| 930 | continue; |
| 931 | Todo.reset(Idx: Block); |
| 932 | // This is a new through block. Add it to SpillPlacer later. |
| 933 | ActiveBlocks.push_back(Elt: Block); |
| 934 | #ifndef NDEBUG |
| 935 | ++Visited; |
| 936 | #endif |
| 937 | } |
| 938 | } |
| 939 | // Any new blocks to add? |
| 940 | if (ActiveBlocks.size() == AddedTo) |
| 941 | break; |
| 942 | |
| 943 | // Compute through constraints from the interference, or assume that all |
| 944 | // through blocks prefer spilling when forming compact regions. |
| 945 | auto NewBlocks = ArrayRef(ActiveBlocks).slice(N: AddedTo); |
| 946 | if (Cand.PhysReg) { |
| 947 | if (!addThroughConstraints(Intf: Cand.Intf, Blocks: NewBlocks)) |
| 948 | return false; |
| 949 | } else { |
| 950 | // Providing that the variable being spilled does not look like a loop |
| 951 | // induction variable, which is expensive to spill around and better |
| 952 | // pushed into a condition inside the loop if possible, provide a strong |
| 953 | // negative bias on through blocks to prevent unwanted liveness on loop |
| 954 | // backedges. |
| 955 | bool PrefSpill = true; |
| 956 | if (SA->looksLikeLoopIV() && NewBlocks.size() >= 2) { |
| 957 | // Check that the current bundle is adding a Header + start+end of |
| 958 | // loop-internal blocks. If the block is indeed a header, don't make |
| 959 | // the NewBlocks as PrefSpill to allow the variable to be live in |
| 960 | // Header<->Latch. |
| 961 | MachineLoop *L = Loops->getLoopFor(BB: MF->getBlockNumbered(N: NewBlocks[0])); |
| 962 | if (L && L->getHeader()->getNumber() == (int)NewBlocks[0] && |
| 963 | all_of(Range: NewBlocks.drop_front(), P: [&](unsigned Block) { |
| 964 | return L == Loops->getLoopFor(BB: MF->getBlockNumbered(N: Block)); |
| 965 | })) |
| 966 | PrefSpill = false; |
| 967 | } |
| 968 | if (PrefSpill) |
| 969 | SpillPlacer->addPrefSpill(Blocks: NewBlocks, /* Strong= */ true); |
| 970 | } |
| 971 | AddedTo = ActiveBlocks.size(); |
| 972 | |
| 973 | // Perhaps iterating can enable more bundles? |
| 974 | SpillPlacer->iterate(); |
| 975 | } |
| 976 | LLVM_DEBUG(dbgs() << ", v=" << Visited); |
| 977 | return true; |
| 978 | } |
| 979 | |
| 980 | /// calcCompactRegion - Compute the set of edge bundles that should be live |
| 981 | /// when splitting the current live range into compact regions. Compact |
| 982 | /// regions can be computed without looking at interference. They are the |
| 983 | /// regions formed by removing all the live-through blocks from the live range. |
| 984 | /// |
| 985 | /// Returns false if the current live range is already compact, or if the |
| 986 | /// compact regions would form single block regions anyway. |
| 987 | bool RAGreedy::calcCompactRegion(GlobalSplitCandidate &Cand) { |
| 988 | // Without any through blocks, the live range is already compact. |
| 989 | if (!SA->getNumThroughBlocks()) |
| 990 | return false; |
| 991 | |
| 992 | // Compact regions don't correspond to any physreg. |
| 993 | Cand.reset(Cache&: IntfCache, Reg: MCRegister::NoRegister); |
| 994 | |
| 995 | LLVM_DEBUG(dbgs() << "Compact region bundles" ); |
| 996 | |
| 997 | // Use the spill placer to determine the live bundles. GrowRegion pretends |
| 998 | // that all the through blocks have interference when PhysReg is unset. |
| 999 | SpillPlacer->prepare(RegBundles&: Cand.LiveBundles); |
| 1000 | |
| 1001 | // The static split cost will be zero since Cand.Intf reports no interference. |
| 1002 | BlockFrequency Cost; |
| 1003 | if (!addSplitConstraints(Intf: Cand.Intf, Cost, MaxCost: BlockFrequency::max())) { |
| 1004 | LLVM_DEBUG(dbgs() << ", none.\n" ); |
| 1005 | return false; |
| 1006 | } |
| 1007 | |
| 1008 | if (!growRegion(Cand)) { |
| 1009 | LLVM_DEBUG(dbgs() << ", cannot spill all interferences.\n" ); |
| 1010 | return false; |
| 1011 | } |
| 1012 | |
| 1013 | SpillPlacer->finish(); |
| 1014 | |
| 1015 | if (!Cand.LiveBundles.any()) { |
| 1016 | LLVM_DEBUG(dbgs() << ", none.\n" ); |
| 1017 | return false; |
| 1018 | } |
| 1019 | |
| 1020 | LLVM_DEBUG({ |
| 1021 | for (int I : Cand.LiveBundles.set_bits()) |
| 1022 | dbgs() << " EB#" << I; |
| 1023 | dbgs() << ".\n" ; |
| 1024 | }); |
| 1025 | return true; |
| 1026 | } |
| 1027 | |
| 1028 | /// calcBlockSplitCost - Compute how expensive it would be to split the live |
| 1029 | /// range in SA around all use blocks instead of forming bundle regions. |
| 1030 | BlockFrequency RAGreedy::calcBlockSplitCost() { |
| 1031 | BlockFrequency Cost = BlockFrequency(0); |
| 1032 | ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks(); |
| 1033 | for (const SplitAnalysis::BlockInfo &BI : UseBlocks) { |
| 1034 | unsigned Number = BI.MBB->getNumber(); |
| 1035 | // We normally only need one spill instruction - a load or a store. |
| 1036 | Cost += SpillPlacer->getBlockFrequency(Number); |
| 1037 | |
| 1038 | // Unless the value is redefined in the block. |
| 1039 | if (BI.LiveIn && BI.LiveOut && BI.FirstDef) |
| 1040 | Cost += SpillPlacer->getBlockFrequency(Number); |
| 1041 | } |
| 1042 | return Cost; |
| 1043 | } |
| 1044 | |
| 1045 | /// calcGlobalSplitCost - Return the global split cost of following the split |
| 1046 | /// pattern in LiveBundles. This cost should be added to the local cost of the |
| 1047 | /// interference pattern in SplitConstraints. |
| 1048 | /// |
| 1049 | BlockFrequency RAGreedy::calcGlobalSplitCost(GlobalSplitCandidate &Cand, |
| 1050 | const AllocationOrder &Order) { |
| 1051 | BlockFrequency GlobalCost = BlockFrequency(0); |
| 1052 | const BitVector &LiveBundles = Cand.LiveBundles; |
| 1053 | ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks(); |
| 1054 | for (unsigned I = 0; I != UseBlocks.size(); ++I) { |
| 1055 | const SplitAnalysis::BlockInfo &BI = UseBlocks[I]; |
| 1056 | SpillPlacement::BlockConstraint &BC = SplitConstraints[I]; |
| 1057 | bool RegIn = LiveBundles[Bundles->getBundle(N: BC.Number, Out: false)]; |
| 1058 | bool RegOut = LiveBundles[Bundles->getBundle(N: BC.Number, Out: true)]; |
| 1059 | unsigned Ins = 0; |
| 1060 | |
| 1061 | Cand.Intf.moveToBlock(MBBNum: BC.Number); |
| 1062 | |
| 1063 | if (BI.LiveIn) |
| 1064 | Ins += RegIn != (BC.Entry == SpillPlacement::PrefReg); |
| 1065 | if (BI.LiveOut) |
| 1066 | Ins += RegOut != (BC.Exit == SpillPlacement::PrefReg); |
| 1067 | while (Ins--) |
| 1068 | GlobalCost += SpillPlacer->getBlockFrequency(Number: BC.Number); |
| 1069 | } |
| 1070 | |
| 1071 | for (unsigned Number : Cand.ActiveBlocks) { |
| 1072 | bool RegIn = LiveBundles[Bundles->getBundle(N: Number, Out: false)]; |
| 1073 | bool RegOut = LiveBundles[Bundles->getBundle(N: Number, Out: true)]; |
| 1074 | if (!RegIn && !RegOut) |
| 1075 | continue; |
| 1076 | if (RegIn && RegOut) { |
| 1077 | // We need double spill code if this block has interference. |
| 1078 | Cand.Intf.moveToBlock(MBBNum: Number); |
| 1079 | if (Cand.Intf.hasInterference()) { |
| 1080 | GlobalCost += SpillPlacer->getBlockFrequency(Number); |
| 1081 | GlobalCost += SpillPlacer->getBlockFrequency(Number); |
| 1082 | } |
| 1083 | continue; |
| 1084 | } |
| 1085 | // live-in / stack-out or stack-in live-out. |
| 1086 | GlobalCost += SpillPlacer->getBlockFrequency(Number); |
| 1087 | } |
| 1088 | return GlobalCost; |
| 1089 | } |
| 1090 | |
| 1091 | /// splitAroundRegion - Split the current live range around the regions |
| 1092 | /// determined by BundleCand and GlobalCand. |
| 1093 | /// |
| 1094 | /// Before calling this function, GlobalCand and BundleCand must be initialized |
| 1095 | /// so each bundle is assigned to a valid candidate, or NoCand for the |
| 1096 | /// stack-bound bundles. The shared SA/SE SplitAnalysis and SplitEditor |
| 1097 | /// objects must be initialized for the current live range, and intervals |
| 1098 | /// created for the used candidates. |
| 1099 | /// |
| 1100 | /// @param LREdit The LiveRangeEdit object handling the current split. |
| 1101 | /// @param UsedCands List of used GlobalCand entries. Every BundleCand value |
| 1102 | /// must appear in this list. |
| 1103 | void RAGreedy::splitAroundRegion(LiveRangeEdit &LREdit, |
| 1104 | ArrayRef<unsigned> UsedCands) { |
| 1105 | // These are the intervals created for new global ranges. We may create more |
| 1106 | // intervals for local ranges. |
| 1107 | const unsigned NumGlobalIntvs = LREdit.size(); |
| 1108 | LLVM_DEBUG(dbgs() << "splitAroundRegion with " << NumGlobalIntvs |
| 1109 | << " globals.\n" ); |
| 1110 | assert(NumGlobalIntvs && "No global intervals configured" ); |
| 1111 | |
| 1112 | // Isolate even single instructions when dealing with a proper sub-class. |
| 1113 | // That guarantees register class inflation for the stack interval because it |
| 1114 | // is all copies. |
| 1115 | Register Reg = SA->getParent().reg(); |
| 1116 | bool SingleInstrs = RegClassInfo->isProperSubClass(RC: MRI->getRegClass(Reg)); |
| 1117 | |
| 1118 | // First handle all the blocks with uses. |
| 1119 | ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks(); |
| 1120 | for (const SplitAnalysis::BlockInfo &BI : UseBlocks) { |
| 1121 | unsigned Number = BI.MBB->getNumber(); |
| 1122 | unsigned IntvIn = 0, IntvOut = 0; |
| 1123 | SlotIndex IntfIn, IntfOut; |
| 1124 | if (BI.LiveIn) { |
| 1125 | unsigned CandIn = BundleCand[Bundles->getBundle(N: Number, Out: false)]; |
| 1126 | if (CandIn != NoCand) { |
| 1127 | GlobalSplitCandidate &Cand = GlobalCand[CandIn]; |
| 1128 | IntvIn = Cand.IntvIdx; |
| 1129 | Cand.Intf.moveToBlock(MBBNum: Number); |
| 1130 | IntfIn = Cand.Intf.first(); |
| 1131 | } |
| 1132 | } |
| 1133 | if (BI.LiveOut) { |
| 1134 | unsigned CandOut = BundleCand[Bundles->getBundle(N: Number, Out: true)]; |
| 1135 | if (CandOut != NoCand) { |
| 1136 | GlobalSplitCandidate &Cand = GlobalCand[CandOut]; |
| 1137 | IntvOut = Cand.IntvIdx; |
| 1138 | Cand.Intf.moveToBlock(MBBNum: Number); |
| 1139 | IntfOut = Cand.Intf.last(); |
| 1140 | } |
| 1141 | } |
| 1142 | |
| 1143 | // Create separate intervals for isolated blocks with multiple uses. |
| 1144 | if (!IntvIn && !IntvOut) { |
| 1145 | LLVM_DEBUG(dbgs() << printMBBReference(*BI.MBB) << " isolated.\n" ); |
| 1146 | if (SA->shouldSplitSingleBlock(BI, SingleInstrs)) |
| 1147 | SE->splitSingleBlock(BI); |
| 1148 | continue; |
| 1149 | } |
| 1150 | |
| 1151 | if (IntvIn && IntvOut) |
| 1152 | SE->splitLiveThroughBlock(MBBNum: Number, IntvIn, LeaveBefore: IntfIn, IntvOut, EnterAfter: IntfOut); |
| 1153 | else if (IntvIn) |
| 1154 | SE->splitRegInBlock(BI, IntvIn, LeaveBefore: IntfIn); |
| 1155 | else |
| 1156 | SE->splitRegOutBlock(BI, IntvOut, EnterAfter: IntfOut); |
| 1157 | } |
| 1158 | |
| 1159 | // Handle live-through blocks. The relevant live-through blocks are stored in |
| 1160 | // the ActiveBlocks list with each candidate. We need to filter out |
| 1161 | // duplicates. |
| 1162 | BitVector Todo = SA->getThroughBlocks(); |
| 1163 | for (unsigned UsedCand : UsedCands) { |
| 1164 | ArrayRef<unsigned> Blocks = GlobalCand[UsedCand].ActiveBlocks; |
| 1165 | for (unsigned Number : Blocks) { |
| 1166 | if (!Todo.test(Idx: Number)) |
| 1167 | continue; |
| 1168 | Todo.reset(Idx: Number); |
| 1169 | |
| 1170 | unsigned IntvIn = 0, IntvOut = 0; |
| 1171 | SlotIndex IntfIn, IntfOut; |
| 1172 | |
| 1173 | unsigned CandIn = BundleCand[Bundles->getBundle(N: Number, Out: false)]; |
| 1174 | if (CandIn != NoCand) { |
| 1175 | GlobalSplitCandidate &Cand = GlobalCand[CandIn]; |
| 1176 | IntvIn = Cand.IntvIdx; |
| 1177 | Cand.Intf.moveToBlock(MBBNum: Number); |
| 1178 | IntfIn = Cand.Intf.first(); |
| 1179 | } |
| 1180 | |
| 1181 | unsigned CandOut = BundleCand[Bundles->getBundle(N: Number, Out: true)]; |
| 1182 | if (CandOut != NoCand) { |
| 1183 | GlobalSplitCandidate &Cand = GlobalCand[CandOut]; |
| 1184 | IntvOut = Cand.IntvIdx; |
| 1185 | Cand.Intf.moveToBlock(MBBNum: Number); |
| 1186 | IntfOut = Cand.Intf.last(); |
| 1187 | } |
| 1188 | if (!IntvIn && !IntvOut) |
| 1189 | continue; |
| 1190 | SE->splitLiveThroughBlock(MBBNum: Number, IntvIn, LeaveBefore: IntfIn, IntvOut, EnterAfter: IntfOut); |
| 1191 | } |
| 1192 | } |
| 1193 | |
| 1194 | ++NumGlobalSplits; |
| 1195 | |
| 1196 | SmallVector<unsigned, 8> IntvMap; |
| 1197 | SE->finish(LRMap: &IntvMap); |
| 1198 | DebugVars->splitRegister(OldReg: Reg, NewRegs: LREdit.regs(), LIS&: *LIS); |
| 1199 | |
| 1200 | unsigned OrigBlocks = SA->getNumLiveBlocks(); |
| 1201 | |
| 1202 | // Sort out the new intervals created by splitting. We get four kinds: |
| 1203 | // - Remainder intervals should not be split again. |
| 1204 | // - Candidate intervals can be assigned to Cand.PhysReg. |
| 1205 | // - Block-local splits are candidates for local splitting. |
| 1206 | // - DCE leftovers should go back on the queue. |
| 1207 | for (unsigned I = 0, E = LREdit.size(); I != E; ++I) { |
| 1208 | const LiveInterval &Reg = LIS->getInterval(Reg: LREdit.get(idx: I)); |
| 1209 | |
| 1210 | // Ignore old intervals from DCE. |
| 1211 | if (ExtraInfo->getOrInitStage(Reg: Reg.reg()) != RS_New) |
| 1212 | continue; |
| 1213 | |
| 1214 | // Remainder interval. Don't try splitting again, spill if it doesn't |
| 1215 | // allocate. |
| 1216 | if (IntvMap[I] == 0) { |
| 1217 | ExtraInfo->setStage(VirtReg: Reg, Stage: RS_Spill); |
| 1218 | continue; |
| 1219 | } |
| 1220 | |
| 1221 | // Global intervals. Allow repeated splitting as long as the number of live |
| 1222 | // blocks is strictly decreasing. |
| 1223 | if (IntvMap[I] < NumGlobalIntvs) { |
| 1224 | if (SA->countLiveBlocks(li: &Reg) >= OrigBlocks) { |
| 1225 | LLVM_DEBUG(dbgs() << "Main interval covers the same " << OrigBlocks |
| 1226 | << " blocks as original.\n" ); |
| 1227 | // Don't allow repeated splitting as a safe guard against looping. |
| 1228 | ExtraInfo->setStage(VirtReg: Reg, Stage: RS_Split2); |
| 1229 | } |
| 1230 | continue; |
| 1231 | } |
| 1232 | |
| 1233 | // Other intervals are treated as new. This includes local intervals created |
| 1234 | // for blocks with multiple uses, and anything created by DCE. |
| 1235 | } |
| 1236 | |
| 1237 | if (VerifyEnabled) |
| 1238 | MF->verify(LiveInts: LIS, Indexes, Banner: "After splitting live range around region" , |
| 1239 | OS: &errs()); |
| 1240 | } |
| 1241 | |
| 1242 | MCRegister RAGreedy::tryRegionSplit(const LiveInterval &VirtReg, |
| 1243 | AllocationOrder &Order, |
| 1244 | SmallVectorImpl<Register> &NewVRegs) { |
| 1245 | if (!TRI->shouldRegionSplitForVirtReg(MF: *MF, VirtReg)) |
| 1246 | return MCRegister::NoRegister; |
| 1247 | unsigned NumCands = 0; |
| 1248 | BlockFrequency SpillCost = calcBlockSplitCost(); |
| 1249 | BlockFrequency BestCost; |
| 1250 | |
| 1251 | // Check if we can split this live range around a compact region. |
| 1252 | bool HasCompact = calcCompactRegion(Cand&: GlobalCand.front()); |
| 1253 | if (HasCompact) { |
| 1254 | // Yes, keep GlobalCand[0] as the compact region candidate. |
| 1255 | NumCands = 1; |
| 1256 | BestCost = BlockFrequency::max(); |
| 1257 | } else { |
| 1258 | // No benefit from the compact region, our fallback will be per-block |
| 1259 | // splitting. Make sure we find a solution that is cheaper than spilling. |
| 1260 | BestCost = SpillCost; |
| 1261 | LLVM_DEBUG(dbgs() << "Cost of isolating all blocks = " |
| 1262 | << printBlockFreq(*MBFI, BestCost) << '\n'); |
| 1263 | } |
| 1264 | |
| 1265 | unsigned BestCand = calculateRegionSplitCost(VirtReg, Order, BestCost, |
| 1266 | NumCands, IgnoreCSR: false /*IgnoreCSR*/); |
| 1267 | |
| 1268 | // No solutions found, fall back to single block splitting. |
| 1269 | if (!HasCompact && BestCand == NoCand) |
| 1270 | return MCRegister::NoRegister; |
| 1271 | |
| 1272 | return doRegionSplit(VirtReg, BestCand, HasCompact, NewVRegs); |
| 1273 | } |
| 1274 | |
| 1275 | unsigned RAGreedy::calculateRegionSplitCostAroundReg(MCRegister PhysReg, |
| 1276 | AllocationOrder &Order, |
| 1277 | BlockFrequency &BestCost, |
| 1278 | unsigned &NumCands, |
| 1279 | unsigned &BestCand) { |
| 1280 | // Discard bad candidates before we run out of interference cache cursors. |
| 1281 | // This will only affect register classes with a lot of registers (>32). |
| 1282 | if (NumCands == IntfCache.getMaxCursors()) { |
| 1283 | unsigned WorstCount = ~0u; |
| 1284 | unsigned Worst = 0; |
| 1285 | for (unsigned CandIndex = 0; CandIndex != NumCands; ++CandIndex) { |
| 1286 | if (CandIndex == BestCand || !GlobalCand[CandIndex].PhysReg) |
| 1287 | continue; |
| 1288 | unsigned Count = GlobalCand[CandIndex].LiveBundles.count(); |
| 1289 | if (Count < WorstCount) { |
| 1290 | Worst = CandIndex; |
| 1291 | WorstCount = Count; |
| 1292 | } |
| 1293 | } |
| 1294 | --NumCands; |
| 1295 | GlobalCand[Worst] = GlobalCand[NumCands]; |
| 1296 | if (BestCand == NumCands) |
| 1297 | BestCand = Worst; |
| 1298 | } |
| 1299 | |
| 1300 | if (GlobalCand.size() <= NumCands) |
| 1301 | GlobalCand.resize(N: NumCands+1); |
| 1302 | GlobalSplitCandidate &Cand = GlobalCand[NumCands]; |
| 1303 | Cand.reset(Cache&: IntfCache, Reg: PhysReg); |
| 1304 | |
| 1305 | SpillPlacer->prepare(RegBundles&: Cand.LiveBundles); |
| 1306 | BlockFrequency Cost; |
| 1307 | if (!addSplitConstraints(Intf: Cand.Intf, Cost, MaxCost: BestCost)) { |
| 1308 | LLVM_DEBUG(dbgs() << printReg(PhysReg, TRI) |
| 1309 | << "\tno positive bundles or too expensive\n" ); |
| 1310 | return BestCand; |
| 1311 | } |
| 1312 | LLVM_DEBUG(dbgs() << printReg(PhysReg, TRI) |
| 1313 | << "\tstatic = " << printBlockFreq(*MBFI, Cost)); |
| 1314 | if (Cost >= BestCost) { |
| 1315 | LLVM_DEBUG({ |
| 1316 | if (BestCand == NoCand) |
| 1317 | dbgs() << " worse than no bundles\n" ; |
| 1318 | else |
| 1319 | dbgs() << " worse than " |
| 1320 | << printReg(GlobalCand[BestCand].PhysReg, TRI) << '\n'; |
| 1321 | }); |
| 1322 | return BestCand; |
| 1323 | } |
| 1324 | if (!growRegion(Cand)) { |
| 1325 | LLVM_DEBUG(dbgs() << ", cannot spill all interferences.\n" ); |
| 1326 | return BestCand; |
| 1327 | } |
| 1328 | |
| 1329 | SpillPlacer->finish(); |
| 1330 | |
| 1331 | // No live bundles, defer to splitSingleBlocks(). |
| 1332 | if (!Cand.LiveBundles.any()) { |
| 1333 | LLVM_DEBUG(dbgs() << " no bundles.\n" ); |
| 1334 | return BestCand; |
| 1335 | } |
| 1336 | |
| 1337 | Cost += calcGlobalSplitCost(Cand, Order); |
| 1338 | LLVM_DEBUG({ |
| 1339 | dbgs() << ", total = " << printBlockFreq(*MBFI, Cost) << " with bundles" ; |
| 1340 | for (int I : Cand.LiveBundles.set_bits()) |
| 1341 | dbgs() << " EB#" << I; |
| 1342 | dbgs() << ".\n" ; |
| 1343 | }); |
| 1344 | if (Cost < BestCost) { |
| 1345 | BestCand = NumCands; |
| 1346 | BestCost = Cost; |
| 1347 | } |
| 1348 | ++NumCands; |
| 1349 | |
| 1350 | return BestCand; |
| 1351 | } |
| 1352 | |
| 1353 | unsigned RAGreedy::calculateRegionSplitCost(const LiveInterval &VirtReg, |
| 1354 | AllocationOrder &Order, |
| 1355 | BlockFrequency &BestCost, |
| 1356 | unsigned &NumCands, |
| 1357 | bool IgnoreCSR) { |
| 1358 | unsigned BestCand = NoCand; |
| 1359 | for (MCRegister PhysReg : Order) { |
| 1360 | assert(PhysReg); |
| 1361 | if (IgnoreCSR && EvictAdvisor->getUnusedCalleeSavedReg(VirtReg, PhysReg)) |
| 1362 | continue; |
| 1363 | |
| 1364 | calculateRegionSplitCostAroundReg(PhysReg, Order, BestCost, NumCands, |
| 1365 | BestCand); |
| 1366 | } |
| 1367 | |
| 1368 | return BestCand; |
| 1369 | } |
| 1370 | |
| 1371 | MCRegister RAGreedy::doRegionSplit(const LiveInterval &VirtReg, |
| 1372 | unsigned BestCand, bool HasCompact, |
| 1373 | SmallVectorImpl<Register> &NewVRegs) { |
| 1374 | SmallVector<unsigned, 8> UsedCands; |
| 1375 | // Prepare split editor. |
| 1376 | LiveRangeEdit LREdit(&VirtReg, NewVRegs, *MF, *LIS, VRM, this, &DeadRemats); |
| 1377 | SE->reset(LREdit, SplitSpillMode); |
| 1378 | |
| 1379 | // Assign all edge bundles to the preferred candidate, or NoCand. |
| 1380 | BundleCand.assign(NumElts: Bundles->getNumBundles(), Elt: NoCand); |
| 1381 | |
| 1382 | // Assign bundles for the best candidate region. |
| 1383 | if (BestCand != NoCand) { |
| 1384 | GlobalSplitCandidate &Cand = GlobalCand[BestCand]; |
| 1385 | if (unsigned B = Cand.getBundles(B&: BundleCand, C: BestCand)) { |
| 1386 | UsedCands.push_back(Elt: BestCand); |
| 1387 | Cand.IntvIdx = SE->openIntv(); |
| 1388 | LLVM_DEBUG(dbgs() << "Split for " << printReg(Cand.PhysReg, TRI) << " in " |
| 1389 | << B << " bundles, intv " << Cand.IntvIdx << ".\n" ); |
| 1390 | (void)B; |
| 1391 | } |
| 1392 | } |
| 1393 | |
| 1394 | // Assign bundles for the compact region. |
| 1395 | if (HasCompact) { |
| 1396 | GlobalSplitCandidate &Cand = GlobalCand.front(); |
| 1397 | assert(!Cand.PhysReg && "Compact region has no physreg" ); |
| 1398 | if (unsigned B = Cand.getBundles(B&: BundleCand, C: 0)) { |
| 1399 | UsedCands.push_back(Elt: 0); |
| 1400 | Cand.IntvIdx = SE->openIntv(); |
| 1401 | LLVM_DEBUG(dbgs() << "Split for compact region in " << B |
| 1402 | << " bundles, intv " << Cand.IntvIdx << ".\n" ); |
| 1403 | (void)B; |
| 1404 | } |
| 1405 | } |
| 1406 | |
| 1407 | splitAroundRegion(LREdit, UsedCands); |
| 1408 | return MCRegister(); |
| 1409 | } |
| 1410 | |
| 1411 | // VirtReg has a physical Hint, this function tries to split VirtReg around |
| 1412 | // Hint if we can place new COPY instructions in cold blocks. |
| 1413 | bool RAGreedy::trySplitAroundHintReg(MCRegister Hint, |
| 1414 | const LiveInterval &VirtReg, |
| 1415 | SmallVectorImpl<Register> &NewVRegs, |
| 1416 | AllocationOrder &Order) { |
| 1417 | // Split the VirtReg may generate COPY instructions in multiple cold basic |
| 1418 | // blocks, and increase code size. So we avoid it when the function is |
| 1419 | // optimized for size. |
| 1420 | if (MF->getFunction().hasOptSize()) |
| 1421 | return false; |
| 1422 | |
| 1423 | // Don't allow repeated splitting as a safe guard against looping. |
| 1424 | if (ExtraInfo->getStage(VirtReg) >= RS_Split2) |
| 1425 | return false; |
| 1426 | |
| 1427 | BlockFrequency Cost = BlockFrequency(0); |
| 1428 | Register Reg = VirtReg.reg(); |
| 1429 | |
| 1430 | // Compute the cost of assigning a non Hint physical register to VirtReg. |
| 1431 | // We define it as the total frequency of broken COPY instructions to/from |
| 1432 | // Hint register, and after split, they can be deleted. |
| 1433 | |
| 1434 | // FIXME: This is miscounting the costs with subregisters. In particular, this |
| 1435 | // should support recognizing SplitKit formed copy bundles instead of direct |
| 1436 | // copy instructions, which will appear in the same block. |
| 1437 | for (const MachineOperand &Opnd : MRI->reg_nodbg_operands(Reg)) { |
| 1438 | const MachineInstr &Instr = *Opnd.getParent(); |
| 1439 | if (!Instr.isCopy() || Opnd.isImplicit()) |
| 1440 | continue; |
| 1441 | |
| 1442 | // Look for the other end of the copy. |
| 1443 | const bool IsDef = Opnd.isDef(); |
| 1444 | const MachineOperand &OtherOpnd = Instr.getOperand(i: IsDef); |
| 1445 | Register OtherReg = OtherOpnd.getReg(); |
| 1446 | assert(Reg == Opnd.getReg()); |
| 1447 | if (OtherReg == Reg) |
| 1448 | continue; |
| 1449 | |
| 1450 | unsigned SubReg = Opnd.getSubReg(); |
| 1451 | unsigned OtherSubReg = OtherOpnd.getSubReg(); |
| 1452 | if (SubReg && OtherSubReg && SubReg != OtherSubReg) |
| 1453 | continue; |
| 1454 | |
| 1455 | // Check if VirtReg interferes with OtherReg after this COPY instruction. |
| 1456 | if (Opnd.readsReg()) { |
| 1457 | SlotIndex Index = LIS->getInstructionIndex(Instr).getRegSlot(); |
| 1458 | |
| 1459 | if (SubReg) { |
| 1460 | LaneBitmask Mask = TRI->getSubRegIndexLaneMask(SubIdx: SubReg); |
| 1461 | if (IsDef) |
| 1462 | Mask = ~Mask; |
| 1463 | |
| 1464 | if (any_of(Range: VirtReg.subranges(), P: [=](const LiveInterval::SubRange &S) { |
| 1465 | return (S.LaneMask & Mask).any() && S.liveAt(index: Index); |
| 1466 | })) { |
| 1467 | continue; |
| 1468 | } |
| 1469 | } else { |
| 1470 | if (VirtReg.liveAt(index: Index)) |
| 1471 | continue; |
| 1472 | } |
| 1473 | } |
| 1474 | |
| 1475 | MCRegister OtherPhysReg = |
| 1476 | OtherReg.isPhysical() ? OtherReg.asMCReg() : VRM->getPhys(virtReg: OtherReg); |
| 1477 | MCRegister ThisHint = SubReg ? TRI->getSubReg(Reg: Hint, Idx: SubReg) : Hint; |
| 1478 | if (OtherPhysReg == ThisHint) |
| 1479 | Cost += MBFI->getBlockFreq(MBB: Instr.getParent()); |
| 1480 | } |
| 1481 | |
| 1482 | // Decrease the cost so it will be split in colder blocks. |
| 1483 | BranchProbability Threshold(SplitThresholdForRegWithHint, 100); |
| 1484 | Cost *= Threshold; |
| 1485 | if (Cost == BlockFrequency(0)) |
| 1486 | return false; |
| 1487 | |
| 1488 | unsigned NumCands = 0; |
| 1489 | unsigned BestCand = NoCand; |
| 1490 | SA->analyze(li: &VirtReg); |
| 1491 | calculateRegionSplitCostAroundReg(PhysReg: Hint, Order, BestCost&: Cost, NumCands, BestCand); |
| 1492 | if (BestCand == NoCand) |
| 1493 | return false; |
| 1494 | |
| 1495 | doRegionSplit(VirtReg, BestCand, HasCompact: false/*HasCompact*/, NewVRegs); |
| 1496 | return true; |
| 1497 | } |
| 1498 | |
| 1499 | //===----------------------------------------------------------------------===// |
| 1500 | // Per-Block Splitting |
| 1501 | //===----------------------------------------------------------------------===// |
| 1502 | |
| 1503 | /// tryBlockSplit - Split a global live range around every block with uses. This |
| 1504 | /// creates a lot of local live ranges, that will be split by tryLocalSplit if |
| 1505 | /// they don't allocate. |
| 1506 | MCRegister RAGreedy::tryBlockSplit(const LiveInterval &VirtReg, |
| 1507 | AllocationOrder &Order, |
| 1508 | SmallVectorImpl<Register> &NewVRegs) { |
| 1509 | assert(&SA->getParent() == &VirtReg && "Live range wasn't analyzed" ); |
| 1510 | Register Reg = VirtReg.reg(); |
| 1511 | bool SingleInstrs = RegClassInfo->isProperSubClass(RC: MRI->getRegClass(Reg)); |
| 1512 | LiveRangeEdit LREdit(&VirtReg, NewVRegs, *MF, *LIS, VRM, this, &DeadRemats); |
| 1513 | SE->reset(LREdit, SplitSpillMode); |
| 1514 | ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks(); |
| 1515 | for (const SplitAnalysis::BlockInfo &BI : UseBlocks) { |
| 1516 | if (SA->shouldSplitSingleBlock(BI, SingleInstrs)) |
| 1517 | SE->splitSingleBlock(BI); |
| 1518 | } |
| 1519 | // No blocks were split. |
| 1520 | if (LREdit.empty()) |
| 1521 | return MCRegister(); |
| 1522 | |
| 1523 | // We did split for some blocks. |
| 1524 | SmallVector<unsigned, 8> IntvMap; |
| 1525 | SE->finish(LRMap: &IntvMap); |
| 1526 | |
| 1527 | // Tell LiveDebugVariables about the new ranges. |
| 1528 | DebugVars->splitRegister(OldReg: Reg, NewRegs: LREdit.regs(), LIS&: *LIS); |
| 1529 | |
| 1530 | // Sort out the new intervals created by splitting. The remainder interval |
| 1531 | // goes straight to spilling, the new local ranges get to stay RS_New. |
| 1532 | for (unsigned I = 0, E = LREdit.size(); I != E; ++I) { |
| 1533 | const LiveInterval &LI = LIS->getInterval(Reg: LREdit.get(idx: I)); |
| 1534 | if (ExtraInfo->getOrInitStage(Reg: LI.reg()) == RS_New && IntvMap[I] == 0) |
| 1535 | ExtraInfo->setStage(VirtReg: LI, Stage: RS_Spill); |
| 1536 | } |
| 1537 | |
| 1538 | if (VerifyEnabled) |
| 1539 | MF->verify(LiveInts: LIS, Indexes, Banner: "After splitting live range around basic blocks" , |
| 1540 | OS: &errs()); |
| 1541 | return MCRegister(); |
| 1542 | } |
| 1543 | |
| 1544 | //===----------------------------------------------------------------------===// |
| 1545 | // Per-Instruction Splitting |
| 1546 | //===----------------------------------------------------------------------===// |
| 1547 | |
| 1548 | /// Get the number of allocatable registers that match the constraints of \p Reg |
| 1549 | /// on \p MI and that are also in \p SuperRC. |
| 1550 | static unsigned getNumAllocatableRegsForConstraints( |
| 1551 | const MachineInstr *MI, Register Reg, const TargetRegisterClass *SuperRC, |
| 1552 | const TargetInstrInfo *TII, const TargetRegisterInfo *TRI, |
| 1553 | const RegisterClassInfo &RCI) { |
| 1554 | assert(SuperRC && "Invalid register class" ); |
| 1555 | |
| 1556 | const TargetRegisterClass *ConstrainedRC = |
| 1557 | MI->getRegClassConstraintEffectForVReg(Reg, CurRC: SuperRC, TII, TRI, |
| 1558 | /* ExploreBundle */ true); |
| 1559 | if (!ConstrainedRC) |
| 1560 | return 0; |
| 1561 | return RCI.getNumAllocatableRegs(RC: ConstrainedRC); |
| 1562 | } |
| 1563 | |
| 1564 | static LaneBitmask getInstReadLaneMask(const MachineRegisterInfo &MRI, |
| 1565 | const TargetRegisterInfo &TRI, |
| 1566 | const MachineInstr &FirstMI, |
| 1567 | Register Reg) { |
| 1568 | LaneBitmask Mask; |
| 1569 | SmallVector<std::pair<MachineInstr *, unsigned>, 8> Ops; |
| 1570 | (void)AnalyzeVirtRegInBundle(MI&: const_cast<MachineInstr &>(FirstMI), Reg, Ops: &Ops); |
| 1571 | |
| 1572 | for (auto [MI, OpIdx] : Ops) { |
| 1573 | const MachineOperand &MO = MI->getOperand(i: OpIdx); |
| 1574 | assert(MO.isReg() && MO.getReg() == Reg); |
| 1575 | unsigned SubReg = MO.getSubReg(); |
| 1576 | if (SubReg == 0 && MO.isUse()) { |
| 1577 | if (MO.isUndef()) |
| 1578 | continue; |
| 1579 | return MRI.getMaxLaneMaskForVReg(Reg); |
| 1580 | } |
| 1581 | |
| 1582 | LaneBitmask SubRegMask = TRI.getSubRegIndexLaneMask(SubIdx: SubReg); |
| 1583 | if (MO.isDef()) { |
| 1584 | if (!MO.isUndef()) |
| 1585 | Mask |= ~SubRegMask; |
| 1586 | } else |
| 1587 | Mask |= SubRegMask; |
| 1588 | } |
| 1589 | |
| 1590 | return Mask; |
| 1591 | } |
| 1592 | |
| 1593 | /// Return true if \p MI at \P Use reads a subset of the lanes live in \p |
| 1594 | /// VirtReg. |
| 1595 | static bool readsLaneSubset(const MachineRegisterInfo &MRI, |
| 1596 | const MachineInstr *MI, const LiveInterval &VirtReg, |
| 1597 | const TargetRegisterInfo *TRI, SlotIndex Use, |
| 1598 | const TargetInstrInfo *TII) { |
| 1599 | // Early check the common case. Beware of the semi-formed bundles SplitKit |
| 1600 | // creates by setting the bundle flag on copies without a matching BUNDLE. |
| 1601 | |
| 1602 | auto DestSrc = TII->isCopyInstr(MI: *MI); |
| 1603 | if (DestSrc && !MI->isBundled() && |
| 1604 | DestSrc->Destination->getSubReg() == DestSrc->Source->getSubReg()) |
| 1605 | return false; |
| 1606 | |
| 1607 | // FIXME: We're only considering uses, but should be consider defs too? |
| 1608 | LaneBitmask ReadMask = getInstReadLaneMask(MRI, TRI: *TRI, FirstMI: *MI, Reg: VirtReg.reg()); |
| 1609 | |
| 1610 | LaneBitmask LiveAtMask; |
| 1611 | for (const LiveInterval::SubRange &S : VirtReg.subranges()) { |
| 1612 | if (S.liveAt(index: Use)) |
| 1613 | LiveAtMask |= S.LaneMask; |
| 1614 | } |
| 1615 | |
| 1616 | // If the live lanes aren't different from the lanes used by the instruction, |
| 1617 | // this doesn't help. |
| 1618 | return (ReadMask & ~(LiveAtMask & TRI->getCoveringLanes())).any(); |
| 1619 | } |
| 1620 | |
| 1621 | /// tryInstructionSplit - Split a live range around individual instructions. |
| 1622 | /// This is normally not worthwhile since the spiller is doing essentially the |
| 1623 | /// same thing. However, when the live range is in a constrained register |
| 1624 | /// class, it may help to insert copies such that parts of the live range can |
| 1625 | /// be moved to a larger register class. |
| 1626 | /// |
| 1627 | /// This is similar to spilling to a larger register class. |
| 1628 | MCRegister RAGreedy::tryInstructionSplit(const LiveInterval &VirtReg, |
| 1629 | AllocationOrder &Order, |
| 1630 | SmallVectorImpl<Register> &NewVRegs) { |
| 1631 | const TargetRegisterClass *CurRC = MRI->getRegClass(Reg: VirtReg.reg()); |
| 1632 | // There is no point to this if there are no larger sub-classes. |
| 1633 | |
| 1634 | bool SplitSubClass = true; |
| 1635 | if (!RegClassInfo->isProperSubClass(RC: CurRC)) { |
| 1636 | if (!VirtReg.hasSubRanges()) |
| 1637 | return MCRegister(); |
| 1638 | SplitSubClass = false; |
| 1639 | } |
| 1640 | |
| 1641 | // Always enable split spill mode, since we're effectively spilling to a |
| 1642 | // register. |
| 1643 | LiveRangeEdit LREdit(&VirtReg, NewVRegs, *MF, *LIS, VRM, this, &DeadRemats); |
| 1644 | SE->reset(LREdit, SplitEditor::SM_Size); |
| 1645 | |
| 1646 | ArrayRef<SlotIndex> Uses = SA->getUseSlots(); |
| 1647 | if (Uses.size() <= 1) |
| 1648 | return MCRegister(); |
| 1649 | |
| 1650 | LLVM_DEBUG(dbgs() << "Split around " << Uses.size() |
| 1651 | << " individual instrs.\n" ); |
| 1652 | |
| 1653 | const TargetRegisterClass *SuperRC = |
| 1654 | TRI->getLargestLegalSuperClass(RC: CurRC, *MF); |
| 1655 | unsigned SuperRCNumAllocatableRegs = |
| 1656 | RegClassInfo->getNumAllocatableRegs(RC: SuperRC); |
| 1657 | // Split around every non-copy instruction if this split will relax |
| 1658 | // the constraints on the virtual register. |
| 1659 | // Otherwise, splitting just inserts uncoalescable copies that do not help |
| 1660 | // the allocation. |
| 1661 | for (const SlotIndex Use : Uses) { |
| 1662 | if (const MachineInstr *MI = Indexes->getInstructionFromIndex(index: Use)) { |
| 1663 | if (TII->isFullCopyInstr(MI: *MI) || |
| 1664 | (SplitSubClass && |
| 1665 | SuperRCNumAllocatableRegs == |
| 1666 | getNumAllocatableRegsForConstraints(MI, Reg: VirtReg.reg(), SuperRC, |
| 1667 | TII, TRI, RCI: *RegClassInfo)) || |
| 1668 | // TODO: Handle split for subranges with subclass constraints? |
| 1669 | (!SplitSubClass && VirtReg.hasSubRanges() && |
| 1670 | !readsLaneSubset(MRI: *MRI, MI, VirtReg, TRI, Use, TII))) { |
| 1671 | LLVM_DEBUG(dbgs() << " skip:\t" << Use << '\t' << *MI); |
| 1672 | continue; |
| 1673 | } |
| 1674 | } |
| 1675 | SE->openIntv(); |
| 1676 | SlotIndex SegStart = SE->enterIntvBefore(Idx: Use); |
| 1677 | SlotIndex SegStop = SE->leaveIntvAfter(Idx: Use); |
| 1678 | SE->useIntv(Start: SegStart, End: SegStop); |
| 1679 | } |
| 1680 | |
| 1681 | if (LREdit.empty()) { |
| 1682 | LLVM_DEBUG(dbgs() << "All uses were copies.\n" ); |
| 1683 | return MCRegister(); |
| 1684 | } |
| 1685 | |
| 1686 | SmallVector<unsigned, 8> IntvMap; |
| 1687 | SE->finish(LRMap: &IntvMap); |
| 1688 | DebugVars->splitRegister(OldReg: VirtReg.reg(), NewRegs: LREdit.regs(), LIS&: *LIS); |
| 1689 | // Assign all new registers to RS_Spill. This was the last chance. |
| 1690 | ExtraInfo->setStage(Begin: LREdit.begin(), End: LREdit.end(), NewStage: RS_Spill); |
| 1691 | return MCRegister(); |
| 1692 | } |
| 1693 | |
| 1694 | //===----------------------------------------------------------------------===// |
| 1695 | // Local Splitting |
| 1696 | //===----------------------------------------------------------------------===// |
| 1697 | |
| 1698 | /// calcGapWeights - Compute the maximum spill weight that needs to be evicted |
| 1699 | /// in order to use PhysReg between two entries in SA->UseSlots. |
| 1700 | /// |
| 1701 | /// GapWeight[I] represents the gap between UseSlots[I] and UseSlots[I + 1]. |
| 1702 | /// |
| 1703 | void RAGreedy::calcGapWeights(MCRegister PhysReg, |
| 1704 | SmallVectorImpl<float> &GapWeight) { |
| 1705 | assert(SA->getUseBlocks().size() == 1 && "Not a local interval" ); |
| 1706 | const SplitAnalysis::BlockInfo &BI = SA->getUseBlocks().front(); |
| 1707 | ArrayRef<SlotIndex> Uses = SA->getUseSlots(); |
| 1708 | const unsigned NumGaps = Uses.size()-1; |
| 1709 | |
| 1710 | // Start and end points for the interference check. |
| 1711 | SlotIndex StartIdx = |
| 1712 | BI.LiveIn ? BI.FirstInstr.getBaseIndex() : BI.FirstInstr; |
| 1713 | SlotIndex StopIdx = |
| 1714 | BI.LiveOut ? BI.LastInstr.getBoundaryIndex() : BI.LastInstr; |
| 1715 | |
| 1716 | GapWeight.assign(NumElts: NumGaps, Elt: 0.0f); |
| 1717 | |
| 1718 | // Add interference from each overlapping register. |
| 1719 | for (MCRegUnit Unit : TRI->regunits(Reg: PhysReg)) { |
| 1720 | if (!Matrix->query(LR: const_cast<LiveInterval &>(SA->getParent()), RegUnit: Unit) |
| 1721 | .checkInterference()) |
| 1722 | continue; |
| 1723 | |
| 1724 | // We know that VirtReg is a continuous interval from FirstInstr to |
| 1725 | // LastInstr, so we don't need InterferenceQuery. |
| 1726 | // |
| 1727 | // Interference that overlaps an instruction is counted in both gaps |
| 1728 | // surrounding the instruction. The exception is interference before |
| 1729 | // StartIdx and after StopIdx. |
| 1730 | // |
| 1731 | LiveIntervalUnion::SegmentIter IntI = |
| 1732 | Matrix->getLiveUnions()[static_cast<unsigned>(Unit)].find(x: StartIdx); |
| 1733 | for (unsigned Gap = 0; IntI.valid() && IntI.start() < StopIdx; ++IntI) { |
| 1734 | // Skip the gaps before IntI. |
| 1735 | while (Uses[Gap+1].getBoundaryIndex() < IntI.start()) |
| 1736 | if (++Gap == NumGaps) |
| 1737 | break; |
| 1738 | if (Gap == NumGaps) |
| 1739 | break; |
| 1740 | |
| 1741 | // Update the gaps covered by IntI. |
| 1742 | const float weight = IntI.value()->weight(); |
| 1743 | for (; Gap != NumGaps; ++Gap) { |
| 1744 | GapWeight[Gap] = std::max(a: GapWeight[Gap], b: weight); |
| 1745 | if (Uses[Gap+1].getBaseIndex() >= IntI.stop()) |
| 1746 | break; |
| 1747 | } |
| 1748 | if (Gap == NumGaps) |
| 1749 | break; |
| 1750 | } |
| 1751 | } |
| 1752 | |
| 1753 | // Add fixed interference. |
| 1754 | for (MCRegUnit Unit : TRI->regunits(Reg: PhysReg)) { |
| 1755 | const LiveRange &LR = LIS->getRegUnit(Unit); |
| 1756 | LiveRange::const_iterator I = LR.find(Pos: StartIdx); |
| 1757 | LiveRange::const_iterator E = LR.end(); |
| 1758 | |
| 1759 | // Same loop as above. Mark any overlapped gaps as HUGE_VALF. |
| 1760 | for (unsigned Gap = 0; I != E && I->start < StopIdx; ++I) { |
| 1761 | while (Uses[Gap+1].getBoundaryIndex() < I->start) |
| 1762 | if (++Gap == NumGaps) |
| 1763 | break; |
| 1764 | if (Gap == NumGaps) |
| 1765 | break; |
| 1766 | |
| 1767 | for (; Gap != NumGaps; ++Gap) { |
| 1768 | GapWeight[Gap] = huge_valf; |
| 1769 | if (Uses[Gap+1].getBaseIndex() >= I->end) |
| 1770 | break; |
| 1771 | } |
| 1772 | if (Gap == NumGaps) |
| 1773 | break; |
| 1774 | } |
| 1775 | } |
| 1776 | } |
| 1777 | |
| 1778 | /// tryLocalSplit - Try to split VirtReg into smaller intervals inside its only |
| 1779 | /// basic block. |
| 1780 | /// |
| 1781 | MCRegister RAGreedy::tryLocalSplit(const LiveInterval &VirtReg, |
| 1782 | AllocationOrder &Order, |
| 1783 | SmallVectorImpl<Register> &NewVRegs) { |
| 1784 | // TODO: the function currently only handles a single UseBlock; it should be |
| 1785 | // possible to generalize. |
| 1786 | if (SA->getUseBlocks().size() != 1) |
| 1787 | return MCRegister(); |
| 1788 | |
| 1789 | const SplitAnalysis::BlockInfo &BI = SA->getUseBlocks().front(); |
| 1790 | |
| 1791 | // Note that it is possible to have an interval that is live-in or live-out |
| 1792 | // while only covering a single block - A phi-def can use undef values from |
| 1793 | // predecessors, and the block could be a single-block loop. |
| 1794 | // We don't bother doing anything clever about such a case, we simply assume |
| 1795 | // that the interval is continuous from FirstInstr to LastInstr. We should |
| 1796 | // make sure that we don't do anything illegal to such an interval, though. |
| 1797 | |
| 1798 | ArrayRef<SlotIndex> Uses = SA->getUseSlots(); |
| 1799 | if (Uses.size() <= 2) |
| 1800 | return MCRegister(); |
| 1801 | const unsigned NumGaps = Uses.size()-1; |
| 1802 | |
| 1803 | LLVM_DEBUG({ |
| 1804 | dbgs() << "tryLocalSplit: " ; |
| 1805 | for (const auto &Use : Uses) |
| 1806 | dbgs() << ' ' << Use; |
| 1807 | dbgs() << '\n'; |
| 1808 | }); |
| 1809 | |
| 1810 | // If VirtReg is live across any register mask operands, compute a list of |
| 1811 | // gaps with register masks. |
| 1812 | SmallVector<unsigned, 8> RegMaskGaps; |
| 1813 | if (Matrix->checkRegMaskInterference(VirtReg)) { |
| 1814 | // Get regmask slots for the whole block. |
| 1815 | ArrayRef<SlotIndex> RMS = LIS->getRegMaskSlotsInBlock(MBBNum: BI.MBB->getNumber()); |
| 1816 | LLVM_DEBUG(dbgs() << RMS.size() << " regmasks in block:" ); |
| 1817 | // Constrain to VirtReg's live range. |
| 1818 | unsigned RI = |
| 1819 | llvm::lower_bound(Range&: RMS, Value: Uses.front().getRegSlot()) - RMS.begin(); |
| 1820 | unsigned RE = RMS.size(); |
| 1821 | for (unsigned I = 0; I != NumGaps && RI != RE; ++I) { |
| 1822 | // Look for Uses[I] <= RMS <= Uses[I + 1]. |
| 1823 | assert(!SlotIndex::isEarlierInstr(RMS[RI], Uses[I])); |
| 1824 | if (SlotIndex::isEarlierInstr(A: Uses[I + 1], B: RMS[RI])) |
| 1825 | continue; |
| 1826 | // Skip a regmask on the same instruction as the last use. It doesn't |
| 1827 | // overlap the live range. |
| 1828 | if (SlotIndex::isSameInstr(A: Uses[I + 1], B: RMS[RI]) && I + 1 == NumGaps) |
| 1829 | break; |
| 1830 | LLVM_DEBUG(dbgs() << ' ' << RMS[RI] << ':' << Uses[I] << '-' |
| 1831 | << Uses[I + 1]); |
| 1832 | RegMaskGaps.push_back(Elt: I); |
| 1833 | // Advance ri to the next gap. A regmask on one of the uses counts in |
| 1834 | // both gaps. |
| 1835 | while (RI != RE && SlotIndex::isEarlierInstr(A: RMS[RI], B: Uses[I + 1])) |
| 1836 | ++RI; |
| 1837 | } |
| 1838 | LLVM_DEBUG(dbgs() << '\n'); |
| 1839 | } |
| 1840 | |
| 1841 | // Since we allow local split results to be split again, there is a risk of |
| 1842 | // creating infinite loops. It is tempting to require that the new live |
| 1843 | // ranges have less instructions than the original. That would guarantee |
| 1844 | // convergence, but it is too strict. A live range with 3 instructions can be |
| 1845 | // split 2+3 (including the COPY), and we want to allow that. |
| 1846 | // |
| 1847 | // Instead we use these rules: |
| 1848 | // |
| 1849 | // 1. Allow any split for ranges with getStage() < RS_Split2. (Except for the |
| 1850 | // noop split, of course). |
| 1851 | // 2. Require progress be made for ranges with getStage() == RS_Split2. All |
| 1852 | // the new ranges must have fewer instructions than before the split. |
| 1853 | // 3. New ranges with the same number of instructions are marked RS_Split2, |
| 1854 | // smaller ranges are marked RS_New. |
| 1855 | // |
| 1856 | // These rules allow a 3 -> 2+3 split once, which we need. They also prevent |
| 1857 | // excessive splitting and infinite loops. |
| 1858 | // |
| 1859 | bool ProgressRequired = ExtraInfo->getStage(VirtReg) >= RS_Split2; |
| 1860 | |
| 1861 | // Best split candidate. |
| 1862 | unsigned BestBefore = NumGaps; |
| 1863 | unsigned BestAfter = 0; |
| 1864 | float BestDiff = 0; |
| 1865 | |
| 1866 | const float blockFreq = |
| 1867 | SpillPlacer->getBlockFrequency(Number: BI.MBB->getNumber()).getFrequency() * |
| 1868 | (1.0f / MBFI->getEntryFreq().getFrequency()); |
| 1869 | SmallVector<float, 8> GapWeight; |
| 1870 | |
| 1871 | for (MCRegister PhysReg : Order) { |
| 1872 | assert(PhysReg); |
| 1873 | // Keep track of the largest spill weight that would need to be evicted in |
| 1874 | // order to make use of PhysReg between UseSlots[I] and UseSlots[I + 1]. |
| 1875 | calcGapWeights(PhysReg, GapWeight); |
| 1876 | |
| 1877 | // Remove any gaps with regmask clobbers. |
| 1878 | if (Matrix->checkRegMaskInterference(VirtReg, PhysReg)) |
| 1879 | for (unsigned Gap : RegMaskGaps) |
| 1880 | GapWeight[Gap] = huge_valf; |
| 1881 | |
| 1882 | // Try to find the best sequence of gaps to close. |
| 1883 | // The new spill weight must be larger than any gap interference. |
| 1884 | |
| 1885 | // We will split before Uses[SplitBefore] and after Uses[SplitAfter]. |
| 1886 | unsigned SplitBefore = 0, SplitAfter = 1; |
| 1887 | |
| 1888 | // MaxGap should always be max(GapWeight[SplitBefore..SplitAfter-1]). |
| 1889 | // It is the spill weight that needs to be evicted. |
| 1890 | float MaxGap = GapWeight[0]; |
| 1891 | |
| 1892 | while (true) { |
| 1893 | // Live before/after split? |
| 1894 | const bool LiveBefore = SplitBefore != 0 || BI.LiveIn; |
| 1895 | const bool LiveAfter = SplitAfter != NumGaps || BI.LiveOut; |
| 1896 | |
| 1897 | LLVM_DEBUG(dbgs() << printReg(PhysReg, TRI) << ' ' << Uses[SplitBefore] |
| 1898 | << '-' << Uses[SplitAfter] << " I=" << MaxGap); |
| 1899 | |
| 1900 | // Stop before the interval gets so big we wouldn't be making progress. |
| 1901 | if (!LiveBefore && !LiveAfter) { |
| 1902 | LLVM_DEBUG(dbgs() << " all\n" ); |
| 1903 | break; |
| 1904 | } |
| 1905 | // Should the interval be extended or shrunk? |
| 1906 | bool Shrink = true; |
| 1907 | |
| 1908 | // How many gaps would the new range have? |
| 1909 | unsigned NewGaps = LiveBefore + SplitAfter - SplitBefore + LiveAfter; |
| 1910 | |
| 1911 | // Legally, without causing looping? |
| 1912 | bool Legal = !ProgressRequired || NewGaps < NumGaps; |
| 1913 | |
| 1914 | if (Legal && MaxGap < huge_valf) { |
| 1915 | // Estimate the new spill weight. Each instruction reads or writes the |
| 1916 | // register. Conservatively assume there are no read-modify-write |
| 1917 | // instructions. |
| 1918 | // |
| 1919 | // Try to guess the size of the new interval. |
| 1920 | const float EstWeight = normalizeSpillWeight( |
| 1921 | UseDefFreq: blockFreq * (NewGaps + 1), |
| 1922 | Size: Uses[SplitBefore].distance(other: Uses[SplitAfter]) + |
| 1923 | (LiveBefore + LiveAfter) * SlotIndex::InstrDist, |
| 1924 | NumInstr: 1); |
| 1925 | // Would this split be possible to allocate? |
| 1926 | // Never allocate all gaps, we wouldn't be making progress. |
| 1927 | LLVM_DEBUG(dbgs() << " w=" << EstWeight); |
| 1928 | if (EstWeight * Hysteresis >= MaxGap) { |
| 1929 | Shrink = false; |
| 1930 | float Diff = EstWeight - MaxGap; |
| 1931 | if (Diff > BestDiff) { |
| 1932 | LLVM_DEBUG(dbgs() << " (best)" ); |
| 1933 | BestDiff = Hysteresis * Diff; |
| 1934 | BestBefore = SplitBefore; |
| 1935 | BestAfter = SplitAfter; |
| 1936 | } |
| 1937 | } |
| 1938 | } |
| 1939 | |
| 1940 | // Try to shrink. |
| 1941 | if (Shrink) { |
| 1942 | if (++SplitBefore < SplitAfter) { |
| 1943 | LLVM_DEBUG(dbgs() << " shrink\n" ); |
| 1944 | // Recompute the max when necessary. |
| 1945 | if (GapWeight[SplitBefore - 1] >= MaxGap) { |
| 1946 | MaxGap = GapWeight[SplitBefore]; |
| 1947 | for (unsigned I = SplitBefore + 1; I != SplitAfter; ++I) |
| 1948 | MaxGap = std::max(a: MaxGap, b: GapWeight[I]); |
| 1949 | } |
| 1950 | continue; |
| 1951 | } |
| 1952 | MaxGap = 0; |
| 1953 | } |
| 1954 | |
| 1955 | // Try to extend the interval. |
| 1956 | if (SplitAfter >= NumGaps) { |
| 1957 | LLVM_DEBUG(dbgs() << " end\n" ); |
| 1958 | break; |
| 1959 | } |
| 1960 | |
| 1961 | LLVM_DEBUG(dbgs() << " extend\n" ); |
| 1962 | MaxGap = std::max(a: MaxGap, b: GapWeight[SplitAfter++]); |
| 1963 | } |
| 1964 | } |
| 1965 | |
| 1966 | // Didn't find any candidates? |
| 1967 | if (BestBefore == NumGaps) |
| 1968 | return MCRegister(); |
| 1969 | |
| 1970 | LLVM_DEBUG(dbgs() << "Best local split range: " << Uses[BestBefore] << '-' |
| 1971 | << Uses[BestAfter] << ", " << BestDiff << ", " |
| 1972 | << (BestAfter - BestBefore + 1) << " instrs\n" ); |
| 1973 | |
| 1974 | LiveRangeEdit LREdit(&VirtReg, NewVRegs, *MF, *LIS, VRM, this, &DeadRemats); |
| 1975 | SE->reset(LREdit); |
| 1976 | |
| 1977 | SE->openIntv(); |
| 1978 | SlotIndex SegStart = SE->enterIntvBefore(Idx: Uses[BestBefore]); |
| 1979 | SlotIndex SegStop = SE->leaveIntvAfter(Idx: Uses[BestAfter]); |
| 1980 | SE->useIntv(Start: SegStart, End: SegStop); |
| 1981 | SmallVector<unsigned, 8> IntvMap; |
| 1982 | SE->finish(LRMap: &IntvMap); |
| 1983 | DebugVars->splitRegister(OldReg: VirtReg.reg(), NewRegs: LREdit.regs(), LIS&: *LIS); |
| 1984 | // If the new range has the same number of instructions as before, mark it as |
| 1985 | // RS_Split2 so the next split will be forced to make progress. Otherwise, |
| 1986 | // leave the new intervals as RS_New so they can compete. |
| 1987 | bool LiveBefore = BestBefore != 0 || BI.LiveIn; |
| 1988 | bool LiveAfter = BestAfter != NumGaps || BI.LiveOut; |
| 1989 | unsigned NewGaps = LiveBefore + BestAfter - BestBefore + LiveAfter; |
| 1990 | if (NewGaps >= NumGaps) { |
| 1991 | LLVM_DEBUG(dbgs() << "Tagging non-progress ranges:" ); |
| 1992 | assert(!ProgressRequired && "Didn't make progress when it was required." ); |
| 1993 | for (unsigned I = 0, E = IntvMap.size(); I != E; ++I) |
| 1994 | if (IntvMap[I] == 1) { |
| 1995 | ExtraInfo->setStage(VirtReg: LIS->getInterval(Reg: LREdit.get(idx: I)), Stage: RS_Split2); |
| 1996 | LLVM_DEBUG(dbgs() << ' ' << printReg(LREdit.get(I))); |
| 1997 | } |
| 1998 | LLVM_DEBUG(dbgs() << '\n'); |
| 1999 | } |
| 2000 | ++NumLocalSplits; |
| 2001 | |
| 2002 | return MCRegister(); |
| 2003 | } |
| 2004 | |
| 2005 | //===----------------------------------------------------------------------===// |
| 2006 | // Live Range Splitting |
| 2007 | //===----------------------------------------------------------------------===// |
| 2008 | |
| 2009 | /// trySplit - Try to split VirtReg or one of its interferences, making it |
| 2010 | /// assignable. |
| 2011 | /// @return Physreg when VirtReg may be assigned and/or new NewVRegs. |
| 2012 | MCRegister RAGreedy::trySplit(const LiveInterval &VirtReg, |
| 2013 | AllocationOrder &Order, |
| 2014 | SmallVectorImpl<Register> &NewVRegs, |
| 2015 | const SmallVirtRegSet &FixedRegisters) { |
| 2016 | // Ranges must be Split2 or less. |
| 2017 | if (ExtraInfo->getStage(VirtReg) >= RS_Spill) |
| 2018 | return MCRegister(); |
| 2019 | |
| 2020 | // Local intervals are handled separately. |
| 2021 | if (LIS->intervalIsInOneMBB(LI: VirtReg)) { |
| 2022 | NamedRegionTimer T("local_split" , "Local Splitting" , TimerGroupName, |
| 2023 | TimerGroupDescription, TimePassesIsEnabled); |
| 2024 | SA->analyze(li: &VirtReg); |
| 2025 | MCRegister PhysReg = tryLocalSplit(VirtReg, Order, NewVRegs); |
| 2026 | if (PhysReg || !NewVRegs.empty()) |
| 2027 | return PhysReg; |
| 2028 | return tryInstructionSplit(VirtReg, Order, NewVRegs); |
| 2029 | } |
| 2030 | |
| 2031 | NamedRegionTimer T("global_split" , "Global Splitting" , TimerGroupName, |
| 2032 | TimerGroupDescription, TimePassesIsEnabled); |
| 2033 | |
| 2034 | SA->analyze(li: &VirtReg); |
| 2035 | |
| 2036 | // First try to split around a region spanning multiple blocks. RS_Split2 |
| 2037 | // ranges already made dubious progress with region splitting, so they go |
| 2038 | // straight to single block splitting. |
| 2039 | if (ExtraInfo->getStage(VirtReg) < RS_Split2) { |
| 2040 | MCRegister PhysReg = tryRegionSplit(VirtReg, Order, NewVRegs); |
| 2041 | if (PhysReg || !NewVRegs.empty()) |
| 2042 | return PhysReg; |
| 2043 | } |
| 2044 | |
| 2045 | // Then isolate blocks. |
| 2046 | return tryBlockSplit(VirtReg, Order, NewVRegs); |
| 2047 | } |
| 2048 | |
| 2049 | //===----------------------------------------------------------------------===// |
| 2050 | // Last Chance Recoloring |
| 2051 | //===----------------------------------------------------------------------===// |
| 2052 | |
| 2053 | /// Return true if \p reg has any tied def operand. |
| 2054 | static bool hasTiedDef(MachineRegisterInfo *MRI, Register reg) { |
| 2055 | for (const MachineOperand &MO : MRI->def_operands(Reg: reg)) |
| 2056 | if (MO.isTied()) |
| 2057 | return true; |
| 2058 | |
| 2059 | return false; |
| 2060 | } |
| 2061 | |
| 2062 | /// Return true if the existing assignment of \p Intf overlaps, but is not the |
| 2063 | /// same, as \p PhysReg. |
| 2064 | static bool assignedRegPartiallyOverlaps(const TargetRegisterInfo &TRI, |
| 2065 | const VirtRegMap &VRM, |
| 2066 | MCRegister PhysReg, |
| 2067 | const LiveInterval &Intf) { |
| 2068 | MCRegister AssignedReg = VRM.getPhys(virtReg: Intf.reg()); |
| 2069 | if (PhysReg == AssignedReg) |
| 2070 | return false; |
| 2071 | return TRI.regsOverlap(RegA: PhysReg, RegB: AssignedReg); |
| 2072 | } |
| 2073 | |
| 2074 | /// mayRecolorAllInterferences - Check if the virtual registers that |
| 2075 | /// interfere with \p VirtReg on \p PhysReg (or one of its aliases) may be |
| 2076 | /// recolored to free \p PhysReg. |
| 2077 | /// When true is returned, \p RecoloringCandidates has been augmented with all |
| 2078 | /// the live intervals that need to be recolored in order to free \p PhysReg |
| 2079 | /// for \p VirtReg. |
| 2080 | /// \p FixedRegisters contains all the virtual registers that cannot be |
| 2081 | /// recolored. |
| 2082 | bool RAGreedy::mayRecolorAllInterferences( |
| 2083 | MCRegister PhysReg, const LiveInterval &VirtReg, |
| 2084 | SmallLISet &RecoloringCandidates, const SmallVirtRegSet &FixedRegisters) { |
| 2085 | const TargetRegisterClass *CurRC = MRI->getRegClass(Reg: VirtReg.reg()); |
| 2086 | |
| 2087 | for (MCRegUnit Unit : TRI->regunits(Reg: PhysReg)) { |
| 2088 | LiveIntervalUnion::Query &Q = Matrix->query(LR: VirtReg, RegUnit: Unit); |
| 2089 | // If there is LastChanceRecoloringMaxInterference or more interferences, |
| 2090 | // chances are one would not be recolorable. |
| 2091 | if (Q.interferingVRegs(MaxInterferingRegs: LastChanceRecoloringMaxInterference).size() >= |
| 2092 | LastChanceRecoloringMaxInterference && |
| 2093 | !ExhaustiveSearch) { |
| 2094 | LLVM_DEBUG(dbgs() << "Early abort: too many interferences.\n" ); |
| 2095 | CutOffInfo |= CO_Interf; |
| 2096 | return false; |
| 2097 | } |
| 2098 | for (const LiveInterval *Intf : reverse(C: Q.interferingVRegs())) { |
| 2099 | // If Intf is done and sits on the same register class as VirtReg, it |
| 2100 | // would not be recolorable as it is in the same state as |
| 2101 | // VirtReg. However there are at least two exceptions. |
| 2102 | // |
| 2103 | // If VirtReg has tied defs and Intf doesn't, then |
| 2104 | // there is still a point in examining if it can be recolorable. |
| 2105 | // |
| 2106 | // Additionally, if the register class has overlapping tuple members, it |
| 2107 | // may still be recolorable using a different tuple. This is more likely |
| 2108 | // if the existing assignment aliases with the candidate. |
| 2109 | // |
| 2110 | if (((ExtraInfo->getStage(VirtReg: *Intf) == RS_Done && |
| 2111 | MRI->getRegClass(Reg: Intf->reg()) == CurRC && |
| 2112 | !assignedRegPartiallyOverlaps(TRI: *TRI, VRM: *VRM, PhysReg, Intf: *Intf)) && |
| 2113 | !(hasTiedDef(MRI, reg: VirtReg.reg()) && |
| 2114 | !hasTiedDef(MRI, reg: Intf->reg()))) || |
| 2115 | FixedRegisters.count(V: Intf->reg())) { |
| 2116 | LLVM_DEBUG( |
| 2117 | dbgs() << "Early abort: the interference is not recolorable.\n" ); |
| 2118 | return false; |
| 2119 | } |
| 2120 | RecoloringCandidates.insert(X: Intf); |
| 2121 | } |
| 2122 | } |
| 2123 | return true; |
| 2124 | } |
| 2125 | |
| 2126 | /// tryLastChanceRecoloring - Try to assign a color to \p VirtReg by recoloring |
| 2127 | /// its interferences. |
| 2128 | /// Last chance recoloring chooses a color for \p VirtReg and recolors every |
| 2129 | /// virtual register that was using it. The recoloring process may recursively |
| 2130 | /// use the last chance recoloring. Therefore, when a virtual register has been |
| 2131 | /// assigned a color by this mechanism, it is marked as Fixed, i.e., it cannot |
| 2132 | /// be last-chance-recolored again during this recoloring "session". |
| 2133 | /// E.g., |
| 2134 | /// Let |
| 2135 | /// vA can use {R1, R2 } |
| 2136 | /// vB can use { R2, R3} |
| 2137 | /// vC can use {R1 } |
| 2138 | /// Where vA, vB, and vC cannot be split anymore (they are reloads for |
| 2139 | /// instance) and they all interfere. |
| 2140 | /// |
| 2141 | /// vA is assigned R1 |
| 2142 | /// vB is assigned R2 |
| 2143 | /// vC tries to evict vA but vA is already done. |
| 2144 | /// Regular register allocation fails. |
| 2145 | /// |
| 2146 | /// Last chance recoloring kicks in: |
| 2147 | /// vC does as if vA was evicted => vC uses R1. |
| 2148 | /// vC is marked as fixed. |
| 2149 | /// vA needs to find a color. |
| 2150 | /// None are available. |
| 2151 | /// vA cannot evict vC: vC is a fixed virtual register now. |
| 2152 | /// vA does as if vB was evicted => vA uses R2. |
| 2153 | /// vB needs to find a color. |
| 2154 | /// R3 is available. |
| 2155 | /// Recoloring => vC = R1, vA = R2, vB = R3 |
| 2156 | /// |
| 2157 | /// \p Order defines the preferred allocation order for \p VirtReg. |
| 2158 | /// \p NewRegs will contain any new virtual register that have been created |
| 2159 | /// (split, spill) during the process and that must be assigned. |
| 2160 | /// \p FixedRegisters contains all the virtual registers that cannot be |
| 2161 | /// recolored. |
| 2162 | /// |
| 2163 | /// \p RecolorStack tracks the original assignments of successfully recolored |
| 2164 | /// registers. |
| 2165 | /// |
| 2166 | /// \p Depth gives the current depth of the last chance recoloring. |
| 2167 | /// \return a physical register that can be used for VirtReg or ~0u if none |
| 2168 | /// exists. |
| 2169 | MCRegister RAGreedy::tryLastChanceRecoloring( |
| 2170 | const LiveInterval &VirtReg, AllocationOrder &Order, |
| 2171 | SmallVectorImpl<Register> &NewVRegs, SmallVirtRegSet &FixedRegisters, |
| 2172 | RecoloringStack &RecolorStack, unsigned Depth) { |
| 2173 | if (!TRI->shouldUseLastChanceRecoloringForVirtReg(MF: *MF, VirtReg)) |
| 2174 | return ~0u; |
| 2175 | |
| 2176 | LLVM_DEBUG(dbgs() << "Try last chance recoloring for " << VirtReg << '\n'); |
| 2177 | |
| 2178 | const ssize_t EntryStackSize = RecolorStack.size(); |
| 2179 | |
| 2180 | // Ranges must be Done. |
| 2181 | assert((ExtraInfo->getStage(VirtReg) >= RS_Done || !VirtReg.isSpillable()) && |
| 2182 | "Last chance recoloring should really be last chance" ); |
| 2183 | // Set the max depth to LastChanceRecoloringMaxDepth. |
| 2184 | // We may want to reconsider that if we end up with a too large search space |
| 2185 | // for target with hundreds of registers. |
| 2186 | // Indeed, in that case we may want to cut the search space earlier. |
| 2187 | if (Depth >= LastChanceRecoloringMaxDepth && !ExhaustiveSearch) { |
| 2188 | LLVM_DEBUG(dbgs() << "Abort because max depth has been reached.\n" ); |
| 2189 | CutOffInfo |= CO_Depth; |
| 2190 | return ~0u; |
| 2191 | } |
| 2192 | |
| 2193 | // Set of Live intervals that will need to be recolored. |
| 2194 | SmallLISet RecoloringCandidates; |
| 2195 | |
| 2196 | // Mark VirtReg as fixed, i.e., it will not be recolored pass this point in |
| 2197 | // this recoloring "session". |
| 2198 | assert(!FixedRegisters.count(VirtReg.reg())); |
| 2199 | FixedRegisters.insert(V: VirtReg.reg()); |
| 2200 | SmallVector<Register, 4> CurrentNewVRegs; |
| 2201 | |
| 2202 | for (MCRegister PhysReg : Order) { |
| 2203 | assert(PhysReg.isValid()); |
| 2204 | LLVM_DEBUG(dbgs() << "Try to assign: " << VirtReg << " to " |
| 2205 | << printReg(PhysReg, TRI) << '\n'); |
| 2206 | RecoloringCandidates.clear(); |
| 2207 | CurrentNewVRegs.clear(); |
| 2208 | |
| 2209 | // It is only possible to recolor virtual register interference. |
| 2210 | if (Matrix->checkInterference(VirtReg, PhysReg) > |
| 2211 | LiveRegMatrix::IK_VirtReg) { |
| 2212 | LLVM_DEBUG( |
| 2213 | dbgs() << "Some interferences are not with virtual registers.\n" ); |
| 2214 | |
| 2215 | continue; |
| 2216 | } |
| 2217 | |
| 2218 | // Early give up on this PhysReg if it is obvious we cannot recolor all |
| 2219 | // the interferences. |
| 2220 | if (!mayRecolorAllInterferences(PhysReg, VirtReg, RecoloringCandidates, |
| 2221 | FixedRegisters)) { |
| 2222 | LLVM_DEBUG(dbgs() << "Some interferences cannot be recolored.\n" ); |
| 2223 | continue; |
| 2224 | } |
| 2225 | |
| 2226 | // RecoloringCandidates contains all the virtual registers that interfere |
| 2227 | // with VirtReg on PhysReg (or one of its aliases). Enqueue them for |
| 2228 | // recoloring and perform the actual recoloring. |
| 2229 | PQueue RecoloringQueue; |
| 2230 | for (const LiveInterval *RC : RecoloringCandidates) { |
| 2231 | Register ItVirtReg = RC->reg(); |
| 2232 | enqueue(CurQueue&: RecoloringQueue, LI: RC); |
| 2233 | assert(VRM->hasPhys(ItVirtReg) && |
| 2234 | "Interferences are supposed to be with allocated variables" ); |
| 2235 | |
| 2236 | // Record the current allocation. |
| 2237 | RecolorStack.push_back(Elt: std::make_pair(x&: RC, y: VRM->getPhys(virtReg: ItVirtReg))); |
| 2238 | |
| 2239 | // unset the related struct. |
| 2240 | Matrix->unassign(VirtReg: *RC); |
| 2241 | } |
| 2242 | |
| 2243 | // Do as if VirtReg was assigned to PhysReg so that the underlying |
| 2244 | // recoloring has the right information about the interferes and |
| 2245 | // available colors. |
| 2246 | Matrix->assign(VirtReg, PhysReg); |
| 2247 | |
| 2248 | // VirtReg may be deleted during tryRecoloringCandidates, save a copy. |
| 2249 | Register ThisVirtReg = VirtReg.reg(); |
| 2250 | |
| 2251 | // Save the current recoloring state. |
| 2252 | // If we cannot recolor all the interferences, we will have to start again |
| 2253 | // at this point for the next physical register. |
| 2254 | SmallVirtRegSet SaveFixedRegisters(FixedRegisters); |
| 2255 | if (tryRecoloringCandidates(RecoloringQueue, CurrentNewVRegs, |
| 2256 | FixedRegisters, RecolorStack, Depth)) { |
| 2257 | // Push the queued vregs into the main queue. |
| 2258 | llvm::append_range(C&: NewVRegs, R&: CurrentNewVRegs); |
| 2259 | // Do not mess up with the global assignment process. |
| 2260 | // I.e., VirtReg must be unassigned. |
| 2261 | if (VRM->hasPhys(virtReg: ThisVirtReg)) { |
| 2262 | Matrix->unassign(VirtReg); |
| 2263 | return PhysReg; |
| 2264 | } |
| 2265 | |
| 2266 | // It is possible VirtReg will be deleted during tryRecoloringCandidates. |
| 2267 | LLVM_DEBUG(dbgs() << "tryRecoloringCandidates deleted a fixed register " |
| 2268 | << printReg(ThisVirtReg) << '\n'); |
| 2269 | FixedRegisters.erase(V: ThisVirtReg); |
| 2270 | return MCRegister(); |
| 2271 | } |
| 2272 | |
| 2273 | LLVM_DEBUG(dbgs() << "Fail to assign: " << VirtReg << " to " |
| 2274 | << printReg(PhysReg, TRI) << '\n'); |
| 2275 | |
| 2276 | // The recoloring attempt failed, undo the changes. |
| 2277 | FixedRegisters = SaveFixedRegisters; |
| 2278 | Matrix->unassign(VirtReg); |
| 2279 | |
| 2280 | // For a newly created vreg which is also in RecoloringCandidates, |
| 2281 | // don't add it to NewVRegs because its physical register will be restored |
| 2282 | // below. Other vregs in CurrentNewVRegs are created by calling |
| 2283 | // selectOrSplit and should be added into NewVRegs. |
| 2284 | for (Register R : CurrentNewVRegs) { |
| 2285 | if (RecoloringCandidates.count(key: &LIS->getInterval(Reg: R))) |
| 2286 | continue; |
| 2287 | NewVRegs.push_back(Elt: R); |
| 2288 | } |
| 2289 | |
| 2290 | // Roll back our unsuccessful recoloring. Also roll back any successful |
| 2291 | // recolorings in any recursive recoloring attempts, since it's possible |
| 2292 | // they would have introduced conflicts with assignments we will be |
| 2293 | // restoring further up the stack. Perform all unassignments prior to |
| 2294 | // reassigning, since sub-recolorings may have conflicted with the registers |
| 2295 | // we are going to restore to their original assignments. |
| 2296 | for (ssize_t I = RecolorStack.size() - 1; I >= EntryStackSize; --I) { |
| 2297 | const LiveInterval *LI; |
| 2298 | MCRegister PhysReg; |
| 2299 | std::tie(args&: LI, args&: PhysReg) = RecolorStack[I]; |
| 2300 | |
| 2301 | if (VRM->hasPhys(virtReg: LI->reg())) |
| 2302 | Matrix->unassign(VirtReg: *LI); |
| 2303 | } |
| 2304 | |
| 2305 | for (size_t I = EntryStackSize; I != RecolorStack.size(); ++I) { |
| 2306 | const LiveInterval *LI; |
| 2307 | MCRegister PhysReg; |
| 2308 | std::tie(args&: LI, args&: PhysReg) = RecolorStack[I]; |
| 2309 | if (!LI->empty() && !MRI->reg_nodbg_empty(RegNo: LI->reg())) |
| 2310 | Matrix->assign(VirtReg: *LI, PhysReg); |
| 2311 | } |
| 2312 | |
| 2313 | // Pop the stack of recoloring attempts. |
| 2314 | RecolorStack.resize(N: EntryStackSize); |
| 2315 | } |
| 2316 | |
| 2317 | // Last chance recoloring did not worked either, give up. |
| 2318 | return ~0u; |
| 2319 | } |
| 2320 | |
| 2321 | /// tryRecoloringCandidates - Try to assign a new color to every register |
| 2322 | /// in \RecoloringQueue. |
| 2323 | /// \p NewRegs will contain any new virtual register created during the |
| 2324 | /// recoloring process. |
| 2325 | /// \p FixedRegisters[in/out] contains all the registers that have been |
| 2326 | /// recolored. |
| 2327 | /// \return true if all virtual registers in RecoloringQueue were successfully |
| 2328 | /// recolored, false otherwise. |
| 2329 | bool RAGreedy::tryRecoloringCandidates(PQueue &RecoloringQueue, |
| 2330 | SmallVectorImpl<Register> &NewVRegs, |
| 2331 | SmallVirtRegSet &FixedRegisters, |
| 2332 | RecoloringStack &RecolorStack, |
| 2333 | unsigned Depth) { |
| 2334 | while (!RecoloringQueue.empty()) { |
| 2335 | const LiveInterval *LI = dequeue(CurQueue&: RecoloringQueue); |
| 2336 | LLVM_DEBUG(dbgs() << "Try to recolor: " << *LI << '\n'); |
| 2337 | MCRegister PhysReg = selectOrSplitImpl(*LI, NewVRegs, FixedRegisters, |
| 2338 | RecolorStack, Depth + 1); |
| 2339 | // When splitting happens, the live-range may actually be empty. |
| 2340 | // In that case, this is okay to continue the recoloring even |
| 2341 | // if we did not find an alternative color for it. Indeed, |
| 2342 | // there will not be anything to color for LI in the end. |
| 2343 | if (PhysReg == ~0u || (!PhysReg && !LI->empty())) |
| 2344 | return false; |
| 2345 | |
| 2346 | if (!PhysReg) { |
| 2347 | assert(LI->empty() && "Only empty live-range do not require a register" ); |
| 2348 | LLVM_DEBUG(dbgs() << "Recoloring of " << *LI |
| 2349 | << " succeeded. Empty LI.\n" ); |
| 2350 | continue; |
| 2351 | } |
| 2352 | LLVM_DEBUG(dbgs() << "Recoloring of " << *LI |
| 2353 | << " succeeded with: " << printReg(PhysReg, TRI) << '\n'); |
| 2354 | |
| 2355 | Matrix->assign(VirtReg: *LI, PhysReg); |
| 2356 | FixedRegisters.insert(V: LI->reg()); |
| 2357 | } |
| 2358 | return true; |
| 2359 | } |
| 2360 | |
| 2361 | //===----------------------------------------------------------------------===// |
| 2362 | // Main Entry Point |
| 2363 | //===----------------------------------------------------------------------===// |
| 2364 | |
| 2365 | MCRegister RAGreedy::selectOrSplit(const LiveInterval &VirtReg, |
| 2366 | SmallVectorImpl<Register> &NewVRegs) { |
| 2367 | CutOffInfo = CO_None; |
| 2368 | LLVMContext &Ctx = MF->getFunction().getContext(); |
| 2369 | SmallVirtRegSet FixedRegisters; |
| 2370 | RecoloringStack RecolorStack; |
| 2371 | MCRegister Reg = |
| 2372 | selectOrSplitImpl(VirtReg, NewVRegs, FixedRegisters, RecolorStack); |
| 2373 | if (Reg == ~0U && (CutOffInfo != CO_None)) { |
| 2374 | uint8_t CutOffEncountered = CutOffInfo & (CO_Depth | CO_Interf); |
| 2375 | if (CutOffEncountered == CO_Depth) |
| 2376 | Ctx.emitError(ErrorStr: "register allocation failed: maximum depth for recoloring " |
| 2377 | "reached. Use -fexhaustive-register-search to skip " |
| 2378 | "cutoffs" ); |
| 2379 | else if (CutOffEncountered == CO_Interf) |
| 2380 | Ctx.emitError(ErrorStr: "register allocation failed: maximum interference for " |
| 2381 | "recoloring reached. Use -fexhaustive-register-search " |
| 2382 | "to skip cutoffs" ); |
| 2383 | else if (CutOffEncountered == (CO_Depth | CO_Interf)) |
| 2384 | Ctx.emitError(ErrorStr: "register allocation failed: maximum interference and " |
| 2385 | "depth for recoloring reached. Use " |
| 2386 | "-fexhaustive-register-search to skip cutoffs" ); |
| 2387 | } |
| 2388 | return Reg; |
| 2389 | } |
| 2390 | |
| 2391 | /// calcSpillCost - Compute how expensive it would be to spill the live range in |
| 2392 | /// LI into memory. |
| 2393 | BlockFrequency RAGreedy::calcSpillCost(const LiveInterval &LI) { |
| 2394 | uint64_t SpillCost = 0; |
| 2395 | SmallPtrSet<MachineInstr *, 8> Visited; |
| 2396 | |
| 2397 | for (MachineRegisterInfo::reg_instr_nodbg_iterator |
| 2398 | I = MRI->reg_instr_nodbg_begin(RegNo: LI.reg()), |
| 2399 | E = MRI->reg_instr_nodbg_end(); |
| 2400 | I != E;) { |
| 2401 | MachineInstr *MI = &*(I++); |
| 2402 | if (MI->isMetaInstruction()) |
| 2403 | continue; |
| 2404 | if (!Visited.insert(Ptr: MI).second) |
| 2405 | continue; |
| 2406 | |
| 2407 | auto [Reads, Writes] = MI->readsWritesVirtualRegister(Reg: LI.reg()); |
| 2408 | auto MBBFreq = SpillPlacer->getBlockFrequency(Number: MI->getParent()->getNumber()); |
| 2409 | SpillCost += (Reads + Writes) * MBBFreq.getFrequency(); |
| 2410 | } |
| 2411 | |
| 2412 | return BlockFrequency(SpillCost); |
| 2413 | } |
| 2414 | |
| 2415 | BlockFrequency RAGreedy::calcRematCost(const LiveInterval &VirtReg, |
| 2416 | AllocationOrder &Order) const { |
| 2417 | if (!VirtReg.isSpillable()) |
| 2418 | return BlockFrequency::max(); |
| 2419 | |
| 2420 | // This logic is intentionally narrow: handle a single concrete value |
| 2421 | // whose def can be cheaply rematerialized at every use. |
| 2422 | if (!VirtReg.containsOneValue()) |
| 2423 | return BlockFrequency::max(); |
| 2424 | const VNInfo *OnlyVNI = VirtReg.getValNumInfo(ValNo: 0); |
| 2425 | if (OnlyVNI->isUnused() || OnlyVNI->isPHIDef()) |
| 2426 | return BlockFrequency::max(); |
| 2427 | |
| 2428 | MachineInstr *DefMI = LIS->getInstructionFromIndex(index: OnlyVNI->def); |
| 2429 | if (!DefMI || !TII->isReMaterializable(MI: *DefMI) || |
| 2430 | !TII->isAsCheapAsAMove(MI: *DefMI)) |
| 2431 | return BlockFrequency::max(); |
| 2432 | |
| 2433 | // Only consider live ranges with register-mask interference. This limits the |
| 2434 | // heuristic to avoiding a first-use CSR required by a call rather than by |
| 2435 | // register pressure. |
| 2436 | if (!Matrix->checkRegMaskInterference(VirtReg)) |
| 2437 | return BlockFrequency::max(); |
| 2438 | |
| 2439 | BlockFrequency RematCost; |
| 2440 | bool HasUse = false; |
| 2441 | for (MachineInstr &UseMI : MRI->use_nodbg_instructions(Reg: VirtReg.reg())) { |
| 2442 | if (!UseMI.readsVirtualRegister(Reg: VirtReg.reg())) |
| 2443 | continue; |
| 2444 | |
| 2445 | if (!canGuaranteeAssignmentAfterRemat(VReg: VirtReg.reg(), MI: UseMI)) |
| 2446 | return BlockFrequency::max(); |
| 2447 | |
| 2448 | SlotIndex UseIdx = LIS->getInstructionIndex(Instr: UseMI).getRegSlot(EC: true); |
| 2449 | if (!VirtRegAuxInfo::allUsesAvailableAt(MI: DefMI, UseIdx, LIS: *LIS, MRI: *MRI, TII: *TII)) |
| 2450 | return BlockFrequency::max(); |
| 2451 | |
| 2452 | // Check if any register that is not a first-use CSR is available as |
| 2453 | // destination for the rematerialization. |
| 2454 | SlotIndex PrevIdx = UseIdx.getPrevSlot(); |
| 2455 | bool HasNoFirstCSRReg = false; |
| 2456 | for (MCRegister RematPhysReg : Order) { |
| 2457 | if (EvictAdvisor->getUnusedCalleeSavedReg(VirtReg, PhysReg: RematPhysReg)) |
| 2458 | continue; |
| 2459 | if (!Matrix->checkInterference(Start: PrevIdx, End: UseIdx, PhysReg: RematPhysReg)) { |
| 2460 | HasNoFirstCSRReg = true; |
| 2461 | break; |
| 2462 | } |
| 2463 | } |
| 2464 | if (!HasNoFirstCSRReg) |
| 2465 | return BlockFrequency::max(); |
| 2466 | |
| 2467 | RematCost += SpillPlacer->getBlockFrequency(Number: UseMI.getParent()->getNumber()); |
| 2468 | HasUse = true; |
| 2469 | } |
| 2470 | |
| 2471 | return HasUse ? RematCost : BlockFrequency::max(); |
| 2472 | } |
| 2473 | |
| 2474 | /// Using a CSR for the first time has a cost because it causes push|pop |
| 2475 | /// to be added to prologue|epilogue. Splitting a cold section of the live |
| 2476 | /// range can have lower cost than using the CSR for the first time; |
| 2477 | /// Spilling a live range in the cold path can have lower cost than using |
| 2478 | /// the CSR for the first time. Returns the physical register if we decide |
| 2479 | /// to use the CSR; otherwise return MCRegister(). |
| 2480 | MCRegister RAGreedy::tryAssignCSRFirstTime( |
| 2481 | const LiveInterval &VirtReg, AllocationOrder &Order, MCRegister PhysReg, |
| 2482 | uint8_t &CostPerUseLimit, SmallVectorImpl<Register> &NewVRegs) { |
| 2483 | LiveRangeStage Stage = ExtraInfo->getStage(VirtReg); |
| 2484 | |
| 2485 | if (Stage == RS_Spill && VirtReg.isSpillable()) { |
| 2486 | // We choose spill or rematerialize over using the CSR for the first time if |
| 2487 | // the spill cost is lower than CSRCost. |
| 2488 | SA->analyze(li: &VirtReg); |
| 2489 | if (calcSpillCost(LI: VirtReg) >= CSRCost && |
| 2490 | calcRematCost(VirtReg, Order) >= CSRCost) { |
| 2491 | return PhysReg; |
| 2492 | } |
| 2493 | |
| 2494 | // We are going to spill or rematerialize, set CostPerUseLimit to 1 to make |
| 2495 | // sure that we will not use a callee-saved register in tryEvict. |
| 2496 | CostPerUseLimit = 1; |
| 2497 | return MCRegister(); |
| 2498 | } |
| 2499 | if (Stage < RS_Split) { |
| 2500 | // We choose pre-splitting over using the CSR for the first time if |
| 2501 | // the cost of splitting is lower than CSRCost. |
| 2502 | SA->analyze(li: &VirtReg); |
| 2503 | unsigned NumCands = 0; |
| 2504 | BlockFrequency BestCost = CSRCost; // Don't modify CSRCost. |
| 2505 | unsigned BestCand = calculateRegionSplitCost(VirtReg, Order, BestCost, |
| 2506 | NumCands, IgnoreCSR: true /*IgnoreCSR*/); |
| 2507 | if (BestCand == NoCand) { |
| 2508 | if (calcRematCost(VirtReg, Order) < CSRCost) { |
| 2509 | CostPerUseLimit = 1; |
| 2510 | return MCRegister(); |
| 2511 | } |
| 2512 | // Use the CSR if we can't find a region split below CSRCost. |
| 2513 | return PhysReg; |
| 2514 | } |
| 2515 | |
| 2516 | // Perform the actual pre-splitting. |
| 2517 | doRegionSplit(VirtReg, BestCand, HasCompact: false/*HasCompact*/, NewVRegs); |
| 2518 | return MCRegister(); |
| 2519 | } |
| 2520 | if (calcRematCost(VirtReg, Order) < CSRCost) { |
| 2521 | CostPerUseLimit = 1; |
| 2522 | return MCRegister(); |
| 2523 | } |
| 2524 | return PhysReg; |
| 2525 | } |
| 2526 | |
| 2527 | void RAGreedy::aboutToRemoveInterval(const LiveInterval &LI) { |
| 2528 | // Do not keep invalid information around. |
| 2529 | SetOfBrokenHints.remove(X: &LI); |
| 2530 | } |
| 2531 | |
| 2532 | void RAGreedy::initializeCSRCost() { |
| 2533 | if (!CSRCostScale.getNumOccurrences() && |
| 2534 | (CSRFirstTimeCost.getNumOccurrences() || TRI->getCSRCost())) { |
| 2535 | // We should deprecate the usage of CSRFirstTimeCost! |
| 2536 | // We use the command-line option if it is explicitly set, otherwise use the |
| 2537 | // larger one out of the command-line option and the value reported by TRI. |
| 2538 | CSRCost = BlockFrequency( |
| 2539 | CSRFirstTimeCost.getNumOccurrences() |
| 2540 | ? CSRFirstTimeCost |
| 2541 | : std::max(a: (unsigned)CSRFirstTimeCost, b: TRI->getCSRCost())); |
| 2542 | if (!CSRCost.getFrequency()) |
| 2543 | return; |
| 2544 | |
| 2545 | // Raw cost is relative to Entry == 2^14; scale it appropriately. |
| 2546 | uint64_t ActualEntry = MBFI->getEntryFreq().getFrequency(); |
| 2547 | if (!ActualEntry) { |
| 2548 | CSRCost = BlockFrequency(0); |
| 2549 | return; |
| 2550 | } |
| 2551 | uint64_t FixedEntry = 1 << 14; |
| 2552 | if (ActualEntry < FixedEntry) { |
| 2553 | CSRCost *= BranchProbability(ActualEntry, FixedEntry); |
| 2554 | } else if (ActualEntry <= UINT32_MAX) { |
| 2555 | // Invert the fraction and divide. |
| 2556 | CSRCost /= BranchProbability(FixedEntry, ActualEntry); |
| 2557 | } else { |
| 2558 | // Can't use BranchProbability in general, since it takes 32-bit numbers. |
| 2559 | CSRCost = |
| 2560 | BlockFrequency(CSRCost.getFrequency() * (ActualEntry / FixedEntry)); |
| 2561 | } |
| 2562 | } else { |
| 2563 | uint64_t EntryFreq = MBFI->getEntryFreq().getFrequency(); |
| 2564 | CSRCost = BlockFrequency(TRI->getCSRFirstUseCost(MF: *MF) * EntryFreq); |
| 2565 | unsigned Scale = TRI->getCSRCostScale(MF: *MF); |
| 2566 | // Command line specified CSRCostScale can override target's default value. |
| 2567 | if (CSRCostScale.getNumOccurrences()) |
| 2568 | Scale = CSRCostScale; |
| 2569 | |
| 2570 | if (Scale < 100) |
| 2571 | CSRCost *= BranchProbability(Scale, 100); |
| 2572 | else |
| 2573 | CSRCost /= BranchProbability(100, Scale); |
| 2574 | } |
| 2575 | } |
| 2576 | |
| 2577 | /// Collect the hint info for \p Reg. |
| 2578 | /// The results are stored into \p Out. |
| 2579 | /// \p Out is not cleared before being populated. |
| 2580 | void RAGreedy::collectHintInfo(Register Reg, HintsInfo &Out) { |
| 2581 | const TargetRegisterClass *RC = MRI->getRegClass(Reg); |
| 2582 | |
| 2583 | for (const MachineOperand &Opnd : MRI->reg_nodbg_operands(Reg)) { |
| 2584 | const MachineInstr &Instr = *Opnd.getParent(); |
| 2585 | if (!Instr.isCopy() || Opnd.isImplicit()) |
| 2586 | continue; |
| 2587 | |
| 2588 | // Look for the other end of the copy. |
| 2589 | const MachineOperand &OtherOpnd = Instr.getOperand(i: Opnd.isDef()); |
| 2590 | Register OtherReg = OtherOpnd.getReg(); |
| 2591 | if (OtherReg == Reg) |
| 2592 | continue; |
| 2593 | unsigned OtherSubReg = OtherOpnd.getSubReg(); |
| 2594 | unsigned SubReg = Opnd.getSubReg(); |
| 2595 | |
| 2596 | // Get the current assignment. |
| 2597 | MCRegister OtherPhysReg; |
| 2598 | if (OtherReg.isPhysical()) { |
| 2599 | if (OtherSubReg) |
| 2600 | OtherPhysReg = TRI->getMatchingSuperReg(Reg: OtherReg, SubIdx: OtherSubReg, RC); |
| 2601 | else if (SubReg) |
| 2602 | OtherPhysReg = TRI->getMatchingSuperReg(Reg: OtherReg, SubIdx: SubReg, RC); |
| 2603 | else |
| 2604 | OtherPhysReg = OtherReg; |
| 2605 | } else { |
| 2606 | OtherPhysReg = VRM->getPhys(virtReg: OtherReg); |
| 2607 | // TODO: Should find matching superregister, but applying this in the |
| 2608 | // non-hint case currently causes regressions |
| 2609 | |
| 2610 | if (SubReg && OtherSubReg && SubReg != OtherSubReg) |
| 2611 | continue; |
| 2612 | } |
| 2613 | |
| 2614 | // Push the collected information. |
| 2615 | if (OtherPhysReg) { |
| 2616 | Out.push_back(Elt: HintInfo(Instr.getParent(), OtherReg, OtherPhysReg)); |
| 2617 | } |
| 2618 | } |
| 2619 | } |
| 2620 | |
| 2621 | /// Using the given \p List, compute the cost of the broken hints if |
| 2622 | /// \p PhysReg was used. |
| 2623 | /// \return The cost of \p List for \p PhysReg. |
| 2624 | BlockFrequency RAGreedy::getBrokenHintFreq(const HintsInfo &List, |
| 2625 | MCRegister PhysReg) { |
| 2626 | BlockFrequency Cost = BlockFrequency(0); |
| 2627 | for (const HintInfo &Info : List) { |
| 2628 | if (Info.PhysReg != PhysReg) |
| 2629 | Cost += MBFI->getBlockFreq(MBB: Info.MBB); |
| 2630 | } |
| 2631 | return Cost; |
| 2632 | } |
| 2633 | |
| 2634 | /// Using the register assigned to \p VirtReg, try to recolor |
| 2635 | /// all the live ranges that are copy-related with \p VirtReg. |
| 2636 | /// The recoloring is then propagated to all the live-ranges that have |
| 2637 | /// been recolored and so on, until no more copies can be coalesced or |
| 2638 | /// it is not profitable. |
| 2639 | /// For a given live range, profitability is determined by the sum of the |
| 2640 | /// frequencies of the non-identity copies it would introduce with the old |
| 2641 | /// and new register. |
| 2642 | /// \p NoRecolorRegs holds registers whose walk is known to recolor nothing. |
| 2643 | /// It is cleared whenever a recoloring happens. |
| 2644 | void RAGreedy::tryHintRecoloring(const LiveInterval &VirtReg, |
| 2645 | DenseSet<Register> &NoRecolorRegs) { |
| 2646 | // We have a broken hint, check if it is possible to fix it by |
| 2647 | // reusing PhysReg for the copy-related live-ranges. Indeed, we evicted |
| 2648 | // some register and PhysReg may be available for the other live-ranges. |
| 2649 | HintsInfo Info; |
| 2650 | Register Reg = VirtReg.reg(); |
| 2651 | MCRegister PhysReg = VRM->getPhys(virtReg: Reg); |
| 2652 | // Start the recoloring algorithm from the input live-interval, then |
| 2653 | // it will propagate to the ones that are copy-related with it. |
| 2654 | SmallDenseSet<Register, 8> Visited = {Reg}; |
| 2655 | SmallVector<Register, 2> RecoloringCandidates = {Reg}; |
| 2656 | SmallVector<Register, 8> Reached; |
| 2657 | bool Changed = false; |
| 2658 | |
| 2659 | LLVM_DEBUG(dbgs() << "Trying to reconcile hints for: " << printReg(Reg, TRI) |
| 2660 | << '(' << printReg(PhysReg, TRI) << ")\n" ); |
| 2661 | |
| 2662 | do { |
| 2663 | Reg = RecoloringCandidates.pop_back_val(); |
| 2664 | |
| 2665 | MCRegister CurrPhys = VRM->getPhys(virtReg: Reg); |
| 2666 | |
| 2667 | // This may be a skipped register. |
| 2668 | if (!CurrPhys) { |
| 2669 | assert(!shouldAllocateRegister(Reg) && |
| 2670 | "We have an unallocated variable which should have been handled" ); |
| 2671 | continue; |
| 2672 | } |
| 2673 | |
| 2674 | // Get the live interval mapped with this virtual register to be able |
| 2675 | // to check for the interference with the new color. |
| 2676 | LiveInterval &LI = LIS->getInterval(Reg); |
| 2677 | // Check that the new color matches the register class constraints and |
| 2678 | // that it is free for this live range. |
| 2679 | if (CurrPhys != PhysReg && (!MRI->getRegClass(Reg)->contains(Reg: PhysReg) || |
| 2680 | Matrix->checkInterference(VirtReg: LI, PhysReg))) |
| 2681 | continue; |
| 2682 | |
| 2683 | LLVM_DEBUG(dbgs() << printReg(Reg, TRI) << '(' << printReg(CurrPhys, TRI) |
| 2684 | << ") is recolorable.\n" ); |
| 2685 | |
| 2686 | // Gather the hint info. |
| 2687 | Info.clear(); |
| 2688 | collectHintInfo(Reg, Out&: Info); |
| 2689 | // Check if recoloring the live-range will increase the cost of the |
| 2690 | // non-identity copies. |
| 2691 | if (CurrPhys != PhysReg) { |
| 2692 | LLVM_DEBUG(dbgs() << "Checking profitability:\n" ); |
| 2693 | BlockFrequency OldCopiesCost = getBrokenHintFreq(List: Info, PhysReg: CurrPhys); |
| 2694 | BlockFrequency NewCopiesCost = getBrokenHintFreq(List: Info, PhysReg); |
| 2695 | LLVM_DEBUG(dbgs() << "Old Cost: " << printBlockFreq(*MBFI, OldCopiesCost) |
| 2696 | << "\nNew Cost: " |
| 2697 | << printBlockFreq(*MBFI, NewCopiesCost) << '\n'); |
| 2698 | if (OldCopiesCost < NewCopiesCost) { |
| 2699 | LLVM_DEBUG(dbgs() << "=> Not profitable.\n" ); |
| 2700 | continue; |
| 2701 | } |
| 2702 | // At this point, the cost is either cheaper or equal. If it is |
| 2703 | // equal, we consider this is profitable because it may expose |
| 2704 | // more recoloring opportunities. |
| 2705 | LLVM_DEBUG(dbgs() << "=> Profitable.\n" ); |
| 2706 | // Recolor the live-range. |
| 2707 | Matrix->unassign(VirtReg: LI); |
| 2708 | Matrix->assign(VirtReg: LI, PhysReg); |
| 2709 | Changed = true; |
| 2710 | } else { |
| 2711 | Reached.push_back(Elt: Reg); |
| 2712 | } |
| 2713 | // Push all copy-related live-ranges to keep reconciling the broken |
| 2714 | // hints. |
| 2715 | for (const HintInfo &HI : Info) { |
| 2716 | // We cannot recolor physical register. |
| 2717 | if (HI.Reg.isVirtual() && Visited.insert(V: HI.Reg).second) |
| 2718 | RecoloringCandidates.push_back(Elt: HI.Reg); |
| 2719 | } |
| 2720 | } while (!RecoloringCandidates.empty()); |
| 2721 | |
| 2722 | if (Changed) |
| 2723 | NoRecolorRegs.clear(); |
| 2724 | else |
| 2725 | NoRecolorRegs.insert_range(R&: Reached); |
| 2726 | } |
| 2727 | |
| 2728 | /// Try to recolor broken hints. |
| 2729 | /// Broken hints may be repaired by recoloring when an evicted variable |
| 2730 | /// freed up a register for a larger live-range. |
| 2731 | /// Consider the following example: |
| 2732 | /// BB1: |
| 2733 | /// a = |
| 2734 | /// b = |
| 2735 | /// BB2: |
| 2736 | /// ... |
| 2737 | /// = b |
| 2738 | /// = a |
| 2739 | /// Let us assume b gets split: |
| 2740 | /// BB1: |
| 2741 | /// a = |
| 2742 | /// b = |
| 2743 | /// BB2: |
| 2744 | /// c = b |
| 2745 | /// ... |
| 2746 | /// d = c |
| 2747 | /// = d |
| 2748 | /// = a |
| 2749 | /// Because of how the allocation work, b, c, and d may be assigned different |
| 2750 | /// colors. Now, if a gets evicted later: |
| 2751 | /// BB1: |
| 2752 | /// a = |
| 2753 | /// st a, SpillSlot |
| 2754 | /// b = |
| 2755 | /// BB2: |
| 2756 | /// c = b |
| 2757 | /// ... |
| 2758 | /// d = c |
| 2759 | /// = d |
| 2760 | /// e = ld SpillSlot |
| 2761 | /// = e |
| 2762 | /// This is likely that we can assign the same register for b, c, and d, |
| 2763 | /// getting rid of 2 copies. |
| 2764 | void RAGreedy::tryHintsRecoloring() { |
| 2765 | DenseSet<Register> NoRecolorRegs; |
| 2766 | for (const LiveInterval *LI : SetOfBrokenHints) { |
| 2767 | assert(LI->reg().isVirtual() && |
| 2768 | "Recoloring is possible only for virtual registers" ); |
| 2769 | // Some dead defs may be around (e.g., because of debug uses). |
| 2770 | // Ignore those. |
| 2771 | if (!VRM->hasPhys(virtReg: LI->reg()) || NoRecolorRegs.contains(V: LI->reg())) |
| 2772 | continue; |
| 2773 | tryHintRecoloring(VirtReg: *LI, NoRecolorRegs); |
| 2774 | } |
| 2775 | } |
| 2776 | |
| 2777 | MCRegister RAGreedy::selectOrSplitImpl(const LiveInterval &VirtReg, |
| 2778 | SmallVectorImpl<Register> &NewVRegs, |
| 2779 | SmallVirtRegSet &FixedRegisters, |
| 2780 | RecoloringStack &RecolorStack, |
| 2781 | unsigned Depth) { |
| 2782 | uint8_t CostPerUseLimit = uint8_t(~0u); |
| 2783 | // First try assigning a free register. |
| 2784 | auto Order = |
| 2785 | AllocationOrder::create(VirtReg: VirtReg.reg(), VRM: *VRM, RegClassInfo: *RegClassInfo, Matrix); |
| 2786 | if (MCRegister PhysReg = |
| 2787 | tryAssign(VirtReg, Order, NewVRegs, FixedRegisters)) { |
| 2788 | // When NewVRegs is not empty, we may have made decisions such as evicting |
| 2789 | // a virtual register, go with the earlier decisions and use the physical |
| 2790 | // register. |
| 2791 | if (CSRCost.getFrequency() && |
| 2792 | EvictAdvisor->getUnusedCalleeSavedReg(VirtReg, PhysReg) && |
| 2793 | NewVRegs.empty()) { |
| 2794 | MCRegister CSRReg = tryAssignCSRFirstTime(VirtReg, Order, PhysReg, |
| 2795 | CostPerUseLimit, NewVRegs); |
| 2796 | if (CSRReg || !NewVRegs.empty()) |
| 2797 | // Return now if we decide to use a CSR or create new vregs due to |
| 2798 | // pre-splitting. |
| 2799 | return CSRReg; |
| 2800 | } else |
| 2801 | return PhysReg; |
| 2802 | } |
| 2803 | // Non empty NewVRegs means VirtReg has been split. |
| 2804 | if (!NewVRegs.empty()) |
| 2805 | return MCRegister(); |
| 2806 | |
| 2807 | LiveRangeStage Stage = ExtraInfo->getStage(VirtReg); |
| 2808 | LLVM_DEBUG(dbgs() << StageName[Stage] << " Cascade " |
| 2809 | << ExtraInfo->getCascade(VirtReg.reg()) << '\n'); |
| 2810 | |
| 2811 | // Try to evict a less worthy live range, but only for ranges from the primary |
| 2812 | // queue. The RS_Split ranges already failed to do this, and they should not |
| 2813 | // get a second chance until they have been split. |
| 2814 | if (Stage != RS_Split) { |
| 2815 | if (MCRegister PhysReg = |
| 2816 | tryEvict(VirtReg, Order, NewVRegs, CostPerUseLimit, |
| 2817 | FixedRegisters)) { |
| 2818 | Register Hint = MRI->getSimpleHint(VReg: VirtReg.reg()); |
| 2819 | // If VirtReg has a hint and that hint is broken record this |
| 2820 | // virtual register as a recoloring candidate for broken hint. |
| 2821 | // Indeed, since we evicted a variable in its neighborhood it is |
| 2822 | // likely we can at least partially recolor some of the |
| 2823 | // copy-related live-ranges. |
| 2824 | if (Hint && Hint != PhysReg) |
| 2825 | SetOfBrokenHints.insert(X: &VirtReg); |
| 2826 | return PhysReg; |
| 2827 | } |
| 2828 | } |
| 2829 | |
| 2830 | assert((NewVRegs.empty() || Depth) && "Cannot append to existing NewVRegs" ); |
| 2831 | |
| 2832 | // The first time we see a live range, don't try to split or spill. |
| 2833 | // Wait until the second time, when all smaller ranges have been allocated. |
| 2834 | // This gives a better picture of the interference to split around. |
| 2835 | if (Stage < RS_Split) { |
| 2836 | ExtraInfo->setStage(VirtReg, Stage: RS_Split); |
| 2837 | LLVM_DEBUG(dbgs() << "wait for second round\n" ); |
| 2838 | NewVRegs.push_back(Elt: VirtReg.reg()); |
| 2839 | return MCRegister(); |
| 2840 | } |
| 2841 | |
| 2842 | if (Stage < RS_Spill && !VirtReg.empty()) { |
| 2843 | // Try splitting VirtReg or interferences. |
| 2844 | unsigned NewVRegSizeBefore = NewVRegs.size(); |
| 2845 | MCRegister PhysReg = trySplit(VirtReg, Order, NewVRegs, FixedRegisters); |
| 2846 | if (PhysReg || (NewVRegs.size() - NewVRegSizeBefore)) |
| 2847 | return PhysReg; |
| 2848 | } |
| 2849 | |
| 2850 | // If we couldn't allocate a register from spilling, there is probably some |
| 2851 | // invalid inline assembly. The base class will report it. |
| 2852 | if (Stage >= RS_Done || !VirtReg.isSpillable()) { |
| 2853 | return tryLastChanceRecoloring(VirtReg, Order, NewVRegs, FixedRegisters, |
| 2854 | RecolorStack, Depth); |
| 2855 | } |
| 2856 | |
| 2857 | // Finally spill VirtReg itself. |
| 2858 | NamedRegionTimer T("spill" , "Spiller" , TimerGroupName, |
| 2859 | TimerGroupDescription, TimePassesIsEnabled); |
| 2860 | LiveRangeEdit LRE(&VirtReg, NewVRegs, *MF, *LIS, VRM, this, &DeadRemats); |
| 2861 | spiller().spill(LRE, Order: &Order); |
| 2862 | ExtraInfo->setStage(Begin: NewVRegs.begin(), End: NewVRegs.end(), NewStage: RS_Done); |
| 2863 | |
| 2864 | // Tell LiveDebugVariables about the new ranges. Ranges not being covered by |
| 2865 | // the new regs are kept in LDV (still mapping to the old register), until |
| 2866 | // we rewrite spilled locations in LDV at a later stage. |
| 2867 | for (Register r : spiller().getSpilledRegs()) |
| 2868 | DebugVars->splitRegister(OldReg: r, NewRegs: LRE.regs(), LIS&: *LIS); |
| 2869 | for (Register r : spiller().getReplacedRegs()) |
| 2870 | DebugVars->splitRegister(OldReg: r, NewRegs: LRE.regs(), LIS&: *LIS); |
| 2871 | |
| 2872 | if (VerifyEnabled) |
| 2873 | MF->verify(LiveInts: LIS, Indexes, Banner: "After spilling" , OS: &errs()); |
| 2874 | |
| 2875 | // The live virtual register requesting allocation was spilled, so tell |
| 2876 | // the caller not to allocate anything during this round. |
| 2877 | return MCRegister(); |
| 2878 | } |
| 2879 | |
| 2880 | void RAGreedy::RAGreedyStats::(MachineOptimizationRemarkMissed &R) { |
| 2881 | using namespace ore; |
| 2882 | if (Spills) { |
| 2883 | R << NV("NumSpills" , Spills) << " spills " ; |
| 2884 | R << NV("TotalSpillsCost" , SpillsCost) << " total spills cost " ; |
| 2885 | } |
| 2886 | if (FoldedSpills) { |
| 2887 | R << NV("NumFoldedSpills" , FoldedSpills) << " folded spills " ; |
| 2888 | R << NV("TotalFoldedSpillsCost" , FoldedSpillsCost) |
| 2889 | << " total folded spills cost " ; |
| 2890 | } |
| 2891 | if (Reloads) { |
| 2892 | R << NV("NumReloads" , Reloads) << " reloads " ; |
| 2893 | R << NV("TotalReloadsCost" , ReloadsCost) << " total reloads cost " ; |
| 2894 | } |
| 2895 | if (FoldedReloads) { |
| 2896 | R << NV("NumFoldedReloads" , FoldedReloads) << " folded reloads " ; |
| 2897 | R << NV("TotalFoldedReloadsCost" , FoldedReloadsCost) |
| 2898 | << " total folded reloads cost " ; |
| 2899 | } |
| 2900 | if (ZeroCostFoldedReloads) |
| 2901 | R << NV("NumZeroCostFoldedReloads" , ZeroCostFoldedReloads) |
| 2902 | << " zero cost folded reloads " ; |
| 2903 | if (Copies) { |
| 2904 | R << NV("NumVRCopies" , Copies) << " virtual registers copies " ; |
| 2905 | R << NV("TotalCopiesCost" , CopiesCost) << " total copies cost " ; |
| 2906 | } |
| 2907 | } |
| 2908 | |
| 2909 | RAGreedy::RAGreedyStats RAGreedy::computeStats(MachineBasicBlock &MBB) { |
| 2910 | RAGreedyStats Stats; |
| 2911 | const MachineFrameInfo &MFI = MF->getFrameInfo(); |
| 2912 | int FI; |
| 2913 | |
| 2914 | auto isSpillSlotAccess = [&MFI](const MachineMemOperand *A) { |
| 2915 | return MFI.isSpillSlotObjectIndex(ObjectIdx: cast<FixedStackPseudoSourceValue>( |
| 2916 | Val: A->getPseudoValue())->getFrameIndex()); |
| 2917 | }; |
| 2918 | auto isPatchpointInstr = [](const MachineInstr &MI) { |
| 2919 | return MI.getOpcode() == TargetOpcode::PATCHPOINT || |
| 2920 | MI.getOpcode() == TargetOpcode::STACKMAP || |
| 2921 | MI.getOpcode() == TargetOpcode::STATEPOINT; |
| 2922 | }; |
| 2923 | for (MachineInstr &MI : MBB) { |
| 2924 | auto DestSrc = TII->isCopyInstr(MI); |
| 2925 | if (DestSrc) { |
| 2926 | const MachineOperand &Dest = *DestSrc->Destination; |
| 2927 | const MachineOperand &Src = *DestSrc->Source; |
| 2928 | Register SrcReg = Src.getReg(); |
| 2929 | Register DestReg = Dest.getReg(); |
| 2930 | // Only count `COPY`s with a virtual register as source or destination. |
| 2931 | if (SrcReg.isVirtual() || DestReg.isVirtual()) { |
| 2932 | if (SrcReg.isVirtual()) { |
| 2933 | SrcReg = VRM->getPhys(virtReg: SrcReg); |
| 2934 | if (SrcReg && Src.getSubReg()) |
| 2935 | SrcReg = TRI->getSubReg(Reg: SrcReg, Idx: Src.getSubReg()); |
| 2936 | } |
| 2937 | if (DestReg.isVirtual()) { |
| 2938 | DestReg = VRM->getPhys(virtReg: DestReg); |
| 2939 | if (DestReg && Dest.getSubReg()) |
| 2940 | DestReg = TRI->getSubReg(Reg: DestReg, Idx: Dest.getSubReg()); |
| 2941 | } |
| 2942 | if (SrcReg != DestReg) |
| 2943 | ++Stats.Copies; |
| 2944 | } |
| 2945 | continue; |
| 2946 | } |
| 2947 | |
| 2948 | SmallVector<const MachineMemOperand *, 2> Accesses; |
| 2949 | if (TII->isLoadFromStackSlot(MI, FrameIndex&: FI) && MFI.isSpillSlotObjectIndex(ObjectIdx: FI)) { |
| 2950 | ++Stats.Reloads; |
| 2951 | continue; |
| 2952 | } |
| 2953 | if (TII->isStoreToStackSlot(MI, FrameIndex&: FI) && MFI.isSpillSlotObjectIndex(ObjectIdx: FI)) { |
| 2954 | ++Stats.Spills; |
| 2955 | continue; |
| 2956 | } |
| 2957 | if (TII->hasLoadFromStackSlot(MI, Accesses) && |
| 2958 | llvm::any_of(Range&: Accesses, P: isSpillSlotAccess)) { |
| 2959 | if (!isPatchpointInstr(MI)) { |
| 2960 | Stats.FoldedReloads += Accesses.size(); |
| 2961 | continue; |
| 2962 | } |
| 2963 | // For statepoint there may be folded and zero cost folded stack reloads. |
| 2964 | std::pair<unsigned, unsigned> NonZeroCostRange = |
| 2965 | TII->getPatchpointUnfoldableRange(MI); |
| 2966 | SmallSet<unsigned, 16> FoldedReloads; |
| 2967 | SmallSet<unsigned, 16> ZeroCostFoldedReloads; |
| 2968 | for (unsigned Idx = 0, E = MI.getNumOperands(); Idx < E; ++Idx) { |
| 2969 | MachineOperand &MO = MI.getOperand(i: Idx); |
| 2970 | if (!MO.isFI() || !MFI.isSpillSlotObjectIndex(ObjectIdx: MO.getIndex())) |
| 2971 | continue; |
| 2972 | if (Idx >= NonZeroCostRange.first && Idx < NonZeroCostRange.second) |
| 2973 | FoldedReloads.insert(V: MO.getIndex()); |
| 2974 | else |
| 2975 | ZeroCostFoldedReloads.insert(V: MO.getIndex()); |
| 2976 | } |
| 2977 | // If stack slot is used in folded reload it is not zero cost then. |
| 2978 | for (unsigned Slot : FoldedReloads) |
| 2979 | ZeroCostFoldedReloads.erase(V: Slot); |
| 2980 | Stats.FoldedReloads += FoldedReloads.size(); |
| 2981 | Stats.ZeroCostFoldedReloads += ZeroCostFoldedReloads.size(); |
| 2982 | continue; |
| 2983 | } |
| 2984 | Accesses.clear(); |
| 2985 | if (TII->hasStoreToStackSlot(MI, Accesses) && |
| 2986 | llvm::any_of(Range&: Accesses, P: isSpillSlotAccess)) { |
| 2987 | Stats.FoldedSpills += Accesses.size(); |
| 2988 | } |
| 2989 | } |
| 2990 | // Set cost of collected statistic by multiplication to relative frequency of |
| 2991 | // this basic block. |
| 2992 | float RelFreq = MBFI->getBlockFreqRelativeToEntryBlock(MBB: &MBB); |
| 2993 | Stats.ReloadsCost = RelFreq * Stats.Reloads; |
| 2994 | Stats.FoldedReloadsCost = RelFreq * Stats.FoldedReloads; |
| 2995 | Stats.SpillsCost = RelFreq * Stats.Spills; |
| 2996 | Stats.FoldedSpillsCost = RelFreq * Stats.FoldedSpills; |
| 2997 | Stats.CopiesCost = RelFreq * Stats.Copies; |
| 2998 | return Stats; |
| 2999 | } |
| 3000 | |
| 3001 | RAGreedy::RAGreedyStats RAGreedy::reportStats(MachineLoop *L) { |
| 3002 | RAGreedyStats Stats; |
| 3003 | |
| 3004 | // Sum up the spill and reloads in subloops. |
| 3005 | for (MachineLoop *SubLoop : *L) |
| 3006 | Stats.add(other: reportStats(L: SubLoop)); |
| 3007 | |
| 3008 | for (MachineBasicBlock *MBB : L->getBlocks()) |
| 3009 | // Handle blocks that were not included in subloops. |
| 3010 | if (Loops->getLoopFor(BB: MBB) == L) |
| 3011 | Stats.add(other: computeStats(MBB&: *MBB)); |
| 3012 | |
| 3013 | if (!Stats.isEmpty()) { |
| 3014 | using namespace ore; |
| 3015 | |
| 3016 | ORE->emit(RemarkBuilder: [&]() { |
| 3017 | MachineOptimizationRemarkMissed R(DEBUG_TYPE, "LoopSpillReloadCopies" , |
| 3018 | L->getStartLoc(), L->getHeader()); |
| 3019 | Stats.report(R); |
| 3020 | R << "generated in loop" ; |
| 3021 | return R; |
| 3022 | }); |
| 3023 | } |
| 3024 | return Stats; |
| 3025 | } |
| 3026 | |
| 3027 | void RAGreedy::reportStats() { |
| 3028 | if (!ORE->allowExtraAnalysis(DEBUG_TYPE)) |
| 3029 | return; |
| 3030 | RAGreedyStats Stats; |
| 3031 | for (MachineLoop *L : *Loops) |
| 3032 | Stats.add(other: reportStats(L)); |
| 3033 | // Process non-loop blocks. |
| 3034 | for (MachineBasicBlock &MBB : *MF) |
| 3035 | if (!Loops->getLoopFor(BB: &MBB)) |
| 3036 | Stats.add(other: computeStats(MBB)); |
| 3037 | if (!Stats.isEmpty()) { |
| 3038 | using namespace ore; |
| 3039 | |
| 3040 | ORE->emit(RemarkBuilder: [&]() { |
| 3041 | DebugLoc Loc; |
| 3042 | if (auto *SP = MF->getFunction().getSubprogram()) |
| 3043 | Loc = DILocation::get(Context&: SP->getContext(), Line: SP->getLine(), Column: 1, Scope: SP); |
| 3044 | MachineOptimizationRemarkMissed R(DEBUG_TYPE, "SpillReloadCopies" , Loc, |
| 3045 | &MF->front()); |
| 3046 | Stats.report(R); |
| 3047 | R << "generated in function" ; |
| 3048 | return R; |
| 3049 | }); |
| 3050 | } |
| 3051 | } |
| 3052 | |
| 3053 | bool RAGreedy::hasVirtRegAlloc() { |
| 3054 | for (unsigned I = 0, E = MRI->getNumVirtRegs(); I != E; ++I) { |
| 3055 | Register Reg = Register::index2VirtReg(Index: I); |
| 3056 | if (MRI->reg_nodbg_empty(RegNo: Reg)) |
| 3057 | continue; |
| 3058 | if (shouldAllocateRegister(Reg)) |
| 3059 | return true; |
| 3060 | } |
| 3061 | |
| 3062 | return false; |
| 3063 | } |
| 3064 | |
| 3065 | bool RAGreedy::run(MachineFunction &mf) { |
| 3066 | LLVM_DEBUG(dbgs() << "********** GREEDY REGISTER ALLOCATION **********\n" |
| 3067 | << "********** Function: " << mf.getName() << '\n'); |
| 3068 | |
| 3069 | MF = &mf; |
| 3070 | TII = MF->getSubtarget().getInstrInfo(); |
| 3071 | |
| 3072 | if (VerifyEnabled) |
| 3073 | MF->verify(LiveInts: LIS, Indexes, Banner: "Before greedy register allocator" , OS: &errs()); |
| 3074 | |
| 3075 | RegAllocBase::init(vrm&: *this->VRM, lis&: *this->LIS, mat&: *this->Matrix, |
| 3076 | rci&: *this->RegClassInfo); |
| 3077 | |
| 3078 | // Early return if there is no virtual register to be allocated to a |
| 3079 | // physical register. |
| 3080 | if (!hasVirtRegAlloc()) |
| 3081 | return false; |
| 3082 | |
| 3083 | // Canonicalize LiveDebugVariables' indexes here and again in |
| 3084 | // emitDebugValues(); those are the only two points where it is clean, as |
| 3085 | // splitting stales indexes again mid-run. Reclaiming erased entries needs |
| 3086 | // its own entry point called from both, not a hook in packIndexes(), which |
| 3087 | // renumberIndexes() also reaches. |
| 3088 | DebugVars->canonicalizeIndexes(SI: *Indexes); |
| 3089 | |
| 3090 | // Renumber to get accurate and consistent results from |
| 3091 | // SlotIndexes::getApproxInstrDistance. |
| 3092 | bool Compact = CompactSlotIndexes.getNumOccurrences() |
| 3093 | ? CompactSlotIndexes |
| 3094 | : TRI->shouldCompactSlotIndexes(MF: *MF); |
| 3095 | if (Compact) { |
| 3096 | // Compaction unlinks what it erases; this is what makes that safe. |
| 3097 | LIS->removeUnreferencedIntervals(); |
| 3098 | Indexes->compactIndexes(); |
| 3099 | } else { |
| 3100 | Indexes->packIndexes(); |
| 3101 | } |
| 3102 | |
| 3103 | initializeCSRCost(); |
| 3104 | |
| 3105 | RegCosts = TRI->getRegisterCosts(MF: *MF); |
| 3106 | RegClassPriorityTrumpsGlobalness = |
| 3107 | GreedyRegClassPriorityTrumpsGlobalness.getNumOccurrences() |
| 3108 | ? GreedyRegClassPriorityTrumpsGlobalness |
| 3109 | : TRI->regClassPriorityTrumpsGlobalness(MF: *MF); |
| 3110 | |
| 3111 | ReverseLocalAssignment = GreedyReverseLocalAssignment.getNumOccurrences() |
| 3112 | ? GreedyReverseLocalAssignment |
| 3113 | : TRI->reverseLocalAssignment(); |
| 3114 | |
| 3115 | ExtraInfo.emplace(); |
| 3116 | |
| 3117 | EvictAdvisor = EvictProvider->getAdvisor(MF: *MF, RA: *this, MBFI, Loops); |
| 3118 | PriorityAdvisor = PriorityProvider->getAdvisor(MF: *MF, RA: *this, SI&: *Indexes); |
| 3119 | |
| 3120 | VRAI = std::make_unique<VirtRegAuxInfo>(args&: *MF, args&: *LIS, args&: *VRM, args&: *Loops, args&: *MBFI); |
| 3121 | SpillerInstance.reset(p: createInlineSpiller(Analyses: {.LIS: *LIS, .LSS: *LSS, .MDT: *DomTree, .MBFI: *MBFI}, MF&: *MF, |
| 3122 | VRM&: *VRM, VRAI&: *VRAI, Matrix)); |
| 3123 | |
| 3124 | VRAI->calculateSpillWeightsAndHints(); |
| 3125 | |
| 3126 | LLVM_DEBUG(LIS->dump()); |
| 3127 | |
| 3128 | SA.reset(p: new SplitAnalysis(*VRM, *LIS, *Loops)); |
| 3129 | SE.reset(p: new SplitEditor(*SA, *LIS, *VRM, *DomTree, *MBFI, *VRAI)); |
| 3130 | |
| 3131 | IntfCache.init(mf: MF, liuarray: Matrix->getLiveUnions(), indexes: Indexes, lis: LIS, tri: TRI); |
| 3132 | GlobalCand.resize(N: 32); // This will grow as needed. |
| 3133 | SetOfBrokenHints.clear(); |
| 3134 | |
| 3135 | allocatePhysRegs(); |
| 3136 | tryHintsRecoloring(); |
| 3137 | |
| 3138 | if (VerifyEnabled) |
| 3139 | MF->verify(LiveInts: LIS, Indexes, Banner: "Before post optimization" , OS: &errs()); |
| 3140 | postOptimization(); |
| 3141 | reportStats(); |
| 3142 | |
| 3143 | releaseMemory(); |
| 3144 | return true; |
| 3145 | } |
| 3146 | |