diff --git a/llvm/include/llvm/InitializePasses.h b/llvm/include/llvm/InitializePasses.h --- a/llvm/include/llvm/InitializePasses.h +++ b/llvm/include/llvm/InitializePasses.h @@ -369,6 +369,7 @@ void initializeReassociateLegacyPassPass(PassRegistry&); void initializeRedundantDbgInstEliminationPass(PassRegistry&); void initializeRegAllocEvictionAdvisorAnalysisPass(PassRegistry &); +void initializeRegAllocPriorityAdvisorAnalysisPass(PassRegistry &); void initializeRegAllocFastPass(PassRegistry&); void initializeRegAllocScoringPass(PassRegistry &); void initializeRegBankSelectPass(PassRegistry&); diff --git a/llvm/lib/Analysis/models/gen-regalloc-priority-test-model.py b/llvm/lib/Analysis/models/gen-regalloc-priority-test-model.py new file mode 100644 --- /dev/null +++ b/llvm/lib/Analysis/models/gen-regalloc-priority-test-model.py @@ -0,0 +1,95 @@ +"""Generate a mock model for LLVM tests for Register Allocation. +The generated model is not a neural net - it is just a tf.function with the +correct input and output parameters. +""" +## By construction, the mock model will always output the first liverange that can be evicted. + +import os +import sys +import tensorflow as tf +POLICY_DECISION_LABEL = 'priority' +POLICY_OUTPUT_SPEC = """ +[ + { + "logging_name": "priority", + "tensor_spec": { + "name": "StatefulPartitionedCall", + "port": 0, + "type": "float", + "shape": [ + 1 + ] + } + } +] +""" +PER_LIVEINTERVAL_INT64_FEATURE_LIST = [ + 'li_size', 'stage' +] +PER_LIVEINTERVAL_FLOAT32_FEATURE_LIST = ['weight' +] +PER_LIVEINTERVAL_FEATURE_LIST = PER_LIVEINTERVAL_FLOAT32_FEATURE_LIST + \ + PER_LIVEINTERVAL_INT64_FEATURE_LIST +CONTEXT_FEATURE_LIST = ('discount', 'reward', 'step_type') + + +def get_input_signature(): + """Returns (time_step_spec, action_spec) for LLVM register allocation.""" + inputs = dict( + (key, tf.TensorSpec(dtype=tf.int64, shape=(), name=key)) + for key in PER_LIVEINTERVAL_INT64_FEATURE_LIST) + inputs.update( + dict((key, + tf.TensorSpec(dtype=tf.float32, shape=(), name=key)) + for key in PER_LIVEINTERVAL_FLOAT32_FEATURE_LIST)) + inputs.update( + dict((key, tf.TensorSpec(dtype=tf.float32, shape=(), name=key)) + for key in ['discount', 'reward'])) + inputs.update( + dict((key, tf.TensorSpec(dtype=tf.int32, shape=(), name=key)) + for key in ['step_type'])) + return inputs + + +def get_output_spec_path(path): + return os.path.join(path, 'output_spec.json') + + +def build_mock_model(path): + """Build and save the mock model with the given signature.""" + module = tf.Module() + # We have to set this useless variable in order for the TF C API to correctly + # intake it + module.var = tf.Variable(0, dtype=tf.float32) + + def action(*inputs): + s1 = tf.reduce_sum([ + tf.cast(inputs[0][key], tf.float32) for key in PER_LIVEINTERVAL_FEATURE_LIST + ], + axis=0) + s2 = tf.reduce_sum( + [tf.cast(inputs[0][key], tf.float32) for key in CONTEXT_FEATURE_LIST]) + # Add a large number so s won't be 0. + s = s1 + s2 + result = s + module.var + return {POLICY_DECISION_LABEL: result} + module.action = tf.function()(action) + action = { + 'action': module.action.get_concrete_function(get_input_signature()) + } + + tf.saved_model.save(module, path, signatures=action) + output_spec_path = get_output_spec_path(path) + with open(output_spec_path, 'w') as f: + print(f'Writing output spec to {output_spec_path}.') + f.write(POLICY_OUTPUT_SPEC) + + +def main(argv): + assert len(argv) == 2 + model_path = argv[1] + build_mock_model(model_path) + + +if __name__ == '__main__': + main(sys.argv) diff --git a/llvm/lib/CodeGen/CMakeLists.txt b/llvm/lib/CodeGen/CMakeLists.txt --- a/llvm/lib/CodeGen/CMakeLists.txt +++ b/llvm/lib/CodeGen/CMakeLists.txt @@ -144,6 +144,7 @@ MIRSampleProfile.cpp MIRYamlMapping.cpp MLRegallocEvictAdvisor.cpp + MLRegallocPriorityAdvisor.cpp ModuloSchedule.cpp MultiHazardRecognizer.cpp PatchableFunction.cpp @@ -171,6 +172,7 @@ RegAllocBase.cpp RegAllocBasic.cpp RegAllocEvictionAdvisor.cpp + RegAllocPriorityAdvisor.cpp RegAllocFast.cpp RegAllocGreedy.cpp RegAllocPBQP.cpp diff --git a/llvm/lib/CodeGen/MLRegallocPriorityAdvisor.cpp b/llvm/lib/CodeGen/MLRegallocPriorityAdvisor.cpp new file mode 100644 --- /dev/null +++ b/llvm/lib/CodeGen/MLRegallocPriorityAdvisor.cpp @@ -0,0 +1,290 @@ +#include "AllocationOrder.h" +#include "RegAllocGreedy.h" +#include "llvm/Analysis/AliasAnalysis.h" +#include "llvm/Analysis/MLModelRunner.h" +#include "llvm/Analysis/ReleaseModeModelRunner.h" +#include "llvm/Analysis/TensorSpec.h" +#include "llvm/CodeGen/CalcSpillWeights.h" +#include "llvm/CodeGen/LiveRegMatrix.h" +#include "llvm/CodeGen/MachineBlockFrequencyInfo.h" +#include "llvm/CodeGen/MachineFunction.h" +#include "llvm/CodeGen/MachineLoopInfo.h" +#include "llvm/CodeGen/MachineRegisterInfo.h" +#include "llvm/CodeGen/Passes.h" +#include "llvm/CodeGen/RegisterClassInfo.h" +#include "llvm/CodeGen/VirtRegMap.h" +#include "llvm/InitializePasses.h" +#include "llvm/Pass.h" +#include "llvm/PassRegistry.h" +#include "llvm/Support/CommandLine.h" + +#include "llvm/Analysis/ModelUnderTrainingRunner.h" +#include "llvm/Analysis/NoInferenceModelRunner.h" + +#include "RegAllocScore.h" +#include "llvm/Analysis/Utils/TFUtils.h" + +using namespace llvm; + +static cl::opt TrainingLog( + "regalloc-prio-training-log", cl::Hidden, + cl::desc("Training log for the register allocator priority model")); + +static cl::opt ModelUnderTraining( + "regalloc-prio-model", cl::Hidden, + cl::desc("The model being trained for register allocation priority")); + +namespace llvm { + +static const std::vector PerLiveRangeShape{1}; + +#define RA_PRIORITY_FEATURES_LIST(M) \ + M(int64_t, li_size, PerLiveRangeShape, "size") \ + M(int64_t, stage, PerLiveRangeShape, "stage") \ + M(float, weight, PerLiveRangeShape, "weight") + +#define DecisionName "priority" + +// Named features index. +enum FeatureIDs { +#define _FEATURE_IDX(_, name, __, ___) name, + RA_PRIORITY_FEATURES_LIST(_FEATURE_IDX) +#undef _FEATURE_IDX + FeatureCount +}; + +class MLPriorityAdvisor : public RegAllocPriorityAdvisor { +public: + MLPriorityAdvisor(const MachineFunction &MF, const RAGreedy &RA, + MLModelRunner *Runner); + void logReward(float reward) override {} + +protected: + const RegAllocPriorityAdvisor &getDefaultAdvisor() const { + return static_cast(DefaultAdvisor); + } + + // The assumption is that if the Runner could not be constructed, we emit-ed + // error, and we shouldn't be asking for it here. + const MLModelRunner &getRunner() const { return *Runner; } + + float tryFindPriority(unsigned Prio, unsigned Size, LiveRangeStage Stage, + float Weight) const override; + +private: + const DefaultPriorityAdvisor DefaultAdvisor; + MLModelRunner *const Runner; +}; + +#define _DECL_FEATURES(type, name, shape, _) \ + TensorSpec::createSpec(#name, shape), + +static const std::vector InputFeatures{ + {RA_PRIORITY_FEATURES_LIST(_DECL_FEATURES)}, +}; +#undef _DECL_FEATURES + +// =================================== +// Release (AOT) - specifics +// =================================== +class ReleaseModePriorityAdvisorAnalysis final + : public RegAllocPriorityAdvisorAnalysis { +public: + ReleaseModePriorityAdvisorAnalysis() + : RegAllocPriorityAdvisorAnalysis(AdvisorMode::Release) {} + // support for isa<> and dyn_cast. + static bool classof(const RegAllocPriorityAdvisorAnalysis *R) { + return R->getAdvisorMode() == AdvisorMode::Release; + } + +private: + void getAnalysisUsage(AnalysisUsage &AU) const override { + RegAllocPriorityAdvisorAnalysis::getAnalysisUsage(AU); + } + + std::unique_ptr + getAdvisor(const MachineFunction &MF, const RAGreedy &RA) override { + if (!Runner) + Runner = std::make_unique>( + MF.getFunction().getContext(), InputFeatures, DecisionName); + return std::make_unique(MF, RA, Runner.get()); + } + std::unique_ptr> Runner; +}; + +static const TensorSpec Output = + TensorSpec::createSpec(DecisionName, {1}); +static const TensorSpec Reward = TensorSpec::createSpec("reward", {1}); + +#define _DECL_TRAIN_FEATURES(type, name, shape, _) \ + TensorSpec::createSpec(std::string("action_") + #name, shape), + +static const std::vector TrainingInputFeatures{ + {RA_PRIORITY_FEATURES_LIST(_DECL_TRAIN_FEATURES) + TensorSpec::createSpec("action_discount", {1}), + TensorSpec::createSpec("action_step_type", {1}), + TensorSpec::createSpec("action_reward", {1})}}; +#undef _DECL_TRAIN_FEATURES + +class DevelopmentModePriorityAdvisor : public MLPriorityAdvisor { +public: + DevelopmentModePriorityAdvisor(const MachineFunction &MF, const RAGreedy &RA, + MLModelRunner *Runner, Logger *Log) + : MLPriorityAdvisor(MF, RA, Runner), Log(Log) {} + + void logReward(float reward) override { Log->logFloatFinalReward(reward); } + +private: + float tryFindPriority(unsigned Prio, unsigned Size, LiveRangeStage Stage, + float Weight) const override; + Logger *const Log; +}; + +class DevelopmentModePriorityAdvisorAnalysis final + : public RegAllocPriorityAdvisorAnalysis { +public: + DevelopmentModePriorityAdvisorAnalysis() + : RegAllocPriorityAdvisorAnalysis(AdvisorMode::Development) {} + // support for isa<> and dyn_cast. + static bool classof(const RegAllocPriorityAdvisorAnalysis *R) { + return R->getAdvisorMode() == AdvisorMode::Development; + } + + /// get the logger for the given function, or nullptr if we didn't collect + /// one. This is used to inject the score by the RegAllocScoring pass. + Logger *getLogger(const MachineFunction &MF) const { + auto I = LogMap.find(MF.getName()); + if (I == LogMap.end()) + return nullptr; + return I->second.get(); + } + +private: + void getAnalysisUsage(AnalysisUsage &AU) const override { + RegAllocPriorityAdvisorAnalysis::getAnalysisUsage(AU); + } + + // Save all the logs (when requested). + bool doFinalization(Module &M) override { + if (TrainingLog.empty()) + return false; + std::error_code EC; + auto OS = std::make_unique(TrainingLog, EC); + if (EC) { + M.getContext().emitError(EC.message() + ":" + TrainingLog); + return false; + } + Logger::flushLogs(*OS, LogMap); + return false; + } + + std::unique_ptr + getAdvisor(const MachineFunction &MF, const RAGreedy &RA) override { + + LLVMContext &Ctx = MF.getFunction().getContext(); + if (ModelUnderTraining.empty() && TrainingLog.empty()) { + Ctx.emitError("Regalloc development mode should be requested with at " + "least logging enabled and/or a training model"); + return nullptr; + } + if (!Runner) { + if (ModelUnderTraining.empty()) + Runner = std::make_unique(Ctx, InputFeatures); + else + Runner = ModelUnderTrainingRunner::createAndEnsureValid( + Ctx, ModelUnderTraining, DecisionName, TrainingInputFeatures); + if (!Runner) { + Ctx.emitError("Regalloc: could not set up the model runner"); + return nullptr; + } + } + + Logger *Log = nullptr; + if (!TrainingLog.empty()) { + std::vector LFS; + for (const auto &FS : InputFeatures) + LFS.push_back({FS, None}); + if (auto *MUTR = dyn_cast(Runner.get())) + if (MUTR->outputLoggedFeatureSpecs().size() > 1) + append_range(LFS, drop_begin(MUTR->outputLoggedFeatureSpecs())); + // We always log the output; in particular, if we're not evaluating, we + // don't have an output spec json file. That's why we handle the + // 'normal' output separately. + LFS.push_back({Output, None}); + auto I = LogMap.insert(std::make_pair( + MF.getFunction().getName(), + std::make_unique(LFS, Reward, /*IncludeReward*/ true))); + assert(I.second); + Log = I.first->second.get(); + } + + return std::make_unique(MF, RA, + Runner.get(), Log); + } + + std::unique_ptr Runner; + StringMap> LogMap; +}; + +} // namespace llvm + +RegAllocPriorityAdvisorAnalysis *llvm::createDevelopmentModePriorityAdvisor() { + return new DevelopmentModePriorityAdvisorAnalysis(); +} + +RegAllocPriorityAdvisorAnalysis *llvm::createReleaseModePriorityAdvisor() { + return new ReleaseModePriorityAdvisorAnalysis(); +} + +MLPriorityAdvisor::MLPriorityAdvisor(const MachineFunction &MF, + const RAGreedy &RA, MLModelRunner *Runner) + : RegAllocPriorityAdvisor(MF, RA), DefaultAdvisor(MF, RA), + Runner(std::move(Runner)) { + assert(this->Runner); +} + +float MLPriorityAdvisor::tryFindPriority(unsigned Prio, unsigned Size, + LiveRangeStage Stage, + float Weight) const { + *Runner->getTensor(0) = static_cast(Size); + *Runner->getTensor(1) = static_cast(Stage); + *Runner->getTensor(2) = static_cast(Weight); + + float Ret = Runner->evaluate(); + return Ret; +} + +float DevelopmentModePriorityAdvisor::tryFindPriority(unsigned Prio, + unsigned Size, + LiveRangeStage Stage, + float Weight) const { + float Ret = 0; + + if (isa(getRunner())) { + Ret = MLPriorityAdvisor::tryFindPriority(Prio, Size, Stage, Weight); + } else { + Ret = static_cast(Prio); + } + + if (TrainingLog.empty()) + return Ret; + + size_t CurrentFeature = 0; + for (; CurrentFeature < InputFeatures.size(); ++CurrentFeature) { + Log->logSpecifiedTensorValue( + CurrentFeature, reinterpret_cast( + getRunner().getTensorUntyped(CurrentFeature))); + } + + if (auto *MUTR = dyn_cast(&getRunner())) { + for (size_t I = 1; I < MUTR->outputLoggedFeatureSpecs().size(); + ++I, ++CurrentFeature) + Log->logSpecifiedTensorValue( + CurrentFeature, + reinterpret_cast( + MUTR->lastEvaluationResult()->getUntypedTensorValue(I))); + } + Log->logFloatValue(CurrentFeature, &Ret); + + return Ret; +} diff --git a/llvm/lib/CodeGen/RegAllocGreedy.h b/llvm/lib/CodeGen/RegAllocGreedy.h deleted file mode 100644 --- a/llvm/lib/CodeGen/RegAllocGreedy.h +++ /dev/null @@ -1,498 +0,0 @@ -//==- RegAllocGreedy.h ------- greedy register allocator ----------*-C++-*-==// -// -// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. -// See https://llvm.org/LICENSE.txt for license information. -// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception -// -//===----------------------------------------------------------------------===// -// This file defines the RAGreedy function pass for register allocation in -// optimized builds. -//===----------------------------------------------------------------------===// - -#ifndef LLVM_CODEGEN_REGALLOCGREEDY_H_ -#define LLVM_CODEGEN_REGALLOCGREEDY_H_ - -#include "InterferenceCache.h" -#include "RegAllocBase.h" -#include "RegAllocEvictionAdvisor.h" -#include "SpillPlacement.h" -#include "SplitKit.h" -#include "llvm/ADT/ArrayRef.h" -#include "llvm/ADT/BitVector.h" -#include "llvm/ADT/DenseMap.h" -#include "llvm/ADT/IndexedMap.h" -#include "llvm/ADT/SetVector.h" -#include "llvm/ADT/SmallPtrSet.h" -#include "llvm/ADT/SmallVector.h" -#include "llvm/ADT/StringRef.h" -#include "llvm/Analysis/AliasAnalysis.h" -#include "llvm/CodeGen/CalcSpillWeights.h" -#include "llvm/CodeGen/LiveInterval.h" -#include "llvm/CodeGen/LiveRangeEdit.h" -#include "llvm/CodeGen/MachineFunction.h" -#include "llvm/CodeGen/MachineFunctionPass.h" -#include "llvm/CodeGen/RegisterClassInfo.h" -#include "llvm/CodeGen/Spiller.h" -#include "llvm/CodeGen/TargetRegisterInfo.h" -#include -#include -#include -#include -#include - -namespace llvm { -class AllocationOrder; -class AnalysisUsage; -class EdgeBundles; -class LiveDebugVariables; -class LiveIntervals; -class LiveRegMatrix; -class MachineBasicBlock; -class MachineBlockFrequencyInfo; -class MachineDominatorTree; -class MachineLoop; -class MachineLoopInfo; -class MachineOptimizationRemarkEmitter; -class MachineOptimizationRemarkMissed; -class SlotIndex; -class SlotIndexes; -class TargetInstrInfo; -class VirtRegMap; - -class LLVM_LIBRARY_VISIBILITY RAGreedy : public MachineFunctionPass, - public RegAllocBase, - private LiveRangeEdit::Delegate { - // Interface to eviction advisers -public: - /// Track allocation stage and eviction loop prevention during allocation. - class ExtraRegInfo final { - // RegInfo - Keep additional information about each live range. - struct RegInfo { - LiveRangeStage Stage = RS_New; - - // Cascade - Eviction loop prevention. See - // canEvictInterferenceBasedOnCost(). - unsigned Cascade = 0; - - RegInfo() = default; - }; - - IndexedMap Info; - unsigned NextCascade = 1; - - public: - ExtraRegInfo() = default; - ExtraRegInfo(const ExtraRegInfo &) = delete; - - LiveRangeStage getStage(Register Reg) const { return Info[Reg].Stage; } - - LiveRangeStage getStage(const LiveInterval &VirtReg) const { - return getStage(VirtReg.reg()); - } - - void setStage(Register Reg, LiveRangeStage Stage) { - Info.grow(Reg.id()); - Info[Reg].Stage = Stage; - } - - void setStage(const LiveInterval &VirtReg, LiveRangeStage Stage) { - setStage(VirtReg.reg(), Stage); - } - - /// Return the current stage of the register, if present, otherwise - /// initialize it and return that. - LiveRangeStage getOrInitStage(Register Reg) { - Info.grow(Reg.id()); - return getStage(Reg); - } - - unsigned getCascade(Register Reg) const { return Info[Reg].Cascade; } - - void setCascade(Register Reg, unsigned Cascade) { - Info.grow(Reg.id()); - Info[Reg].Cascade = Cascade; - } - - unsigned getOrAssignNewCascade(Register Reg) { - unsigned Cascade = getCascade(Reg); - if (!Cascade) { - Cascade = NextCascade++; - setCascade(Reg, Cascade); - } - return Cascade; - } - - unsigned getCascadeOrCurrentNext(Register Reg) const { - unsigned Cascade = getCascade(Reg); - if (!Cascade) - Cascade = NextCascade; - return Cascade; - } - - template - void setStage(Iterator Begin, Iterator End, LiveRangeStage NewStage) { - for (; Begin != End; ++Begin) { - Register Reg = *Begin; - Info.grow(Reg.id()); - if (Info[Reg].Stage == RS_New) - Info[Reg].Stage = NewStage; - } - } - void LRE_DidCloneVirtReg(Register New, Register Old); - }; - - LiveRegMatrix *getInterferenceMatrix() const { return Matrix; } - LiveIntervals *getLiveIntervals() const { return LIS; } - VirtRegMap *getVirtRegMap() const { return VRM; } - const RegisterClassInfo &getRegClassInfo() const { return RegClassInfo; } - const ExtraRegInfo &getExtraInfo() const { return *ExtraInfo; } - size_t getQueueSize() const { return Queue.size(); } - // end (interface to eviction advisers) - -private: - // Convenient shortcuts. - using PQueue = std::priority_queue>; - using SmallLISet = SmallPtrSet; - - // We need to track all tentative recolorings so we can roll back any - // successful and unsuccessful recoloring attempts. - using RecoloringStack = - SmallVector, 8>; - - // context - MachineFunction *MF; - - // Shortcuts to some useful interface. - const TargetInstrInfo *TII; - - // analyses - SlotIndexes *Indexes; - MachineBlockFrequencyInfo *MBFI; - MachineDominatorTree *DomTree; - MachineLoopInfo *Loops; - MachineOptimizationRemarkEmitter *ORE; - EdgeBundles *Bundles; - SpillPlacement *SpillPlacer; - LiveDebugVariables *DebugVars; - AliasAnalysis *AA; - - // state - std::unique_ptr SpillerInstance; - PQueue Queue; - std::unique_ptr VRAI; - Optional ExtraInfo; - std::unique_ptr EvictAdvisor; - - // Enum CutOffStage to keep a track whether the register allocation failed - // because of the cutoffs encountered in last chance recoloring. - // Note: This is used as bitmask. New value should be next power of 2. - enum CutOffStage { - // No cutoffs encountered - CO_None = 0, - - // lcr-max-depth cutoff encountered - CO_Depth = 1, - - // lcr-max-interf cutoff encountered - CO_Interf = 2 - }; - - uint8_t CutOffInfo; - -#ifndef NDEBUG - static const char *const StageName[]; -#endif - - /// EvictionTrack - Keeps track of past evictions in order to optimize region - /// split decision. - class EvictionTrack { - - public: - using EvictorInfo = - std::pair; - using EvicteeInfo = llvm::DenseMap; - - private: - /// Each Vreg that has been evicted in the last stage of selectOrSplit will - /// be mapped to the evictor Vreg and the PhysReg it was evicted from. - EvicteeInfo Evictees; - - public: - /// Clear all eviction information. - void clear() { Evictees.clear(); } - - /// Clear eviction information for the given evictee Vreg. - /// E.g. when Vreg get's a new allocation, the old eviction info is no - /// longer relevant. - /// \param Evictee The evictee Vreg for whom we want to clear collected - /// eviction info. - void clearEvicteeInfo(Register Evictee) { Evictees.erase(Evictee); } - - /// Track new eviction. - /// The Evictor vreg has evicted the Evictee vreg from Physreg. - /// \param PhysReg The physical register Evictee was evicted from. - /// \param Evictor The evictor Vreg that evicted Evictee. - /// \param Evictee The evictee Vreg. - void addEviction(MCRegister PhysReg, Register Evictor, Register Evictee) { - Evictees[Evictee].first = Evictor; - Evictees[Evictee].second = PhysReg; - } - - /// Return the Evictor Vreg which evicted Evictee Vreg from PhysReg. - /// \param Evictee The evictee vreg. - /// \return The Evictor vreg which evicted Evictee vreg from PhysReg. 0 if - /// nobody has evicted Evictee from PhysReg. - EvictorInfo getEvictor(Register Evictee) { - if (Evictees.count(Evictee)) { - return Evictees[Evictee]; - } - - return EvictorInfo(0, 0); - } - }; - - // Keeps track of past evictions in order to optimize region split decision. - EvictionTrack LastEvicted; - - // splitting state. - std::unique_ptr SA; - std::unique_ptr SE; - - /// Cached per-block interference maps - InterferenceCache IntfCache; - - /// All basic blocks where the current register has uses. - SmallVector SplitConstraints; - - /// Global live range splitting candidate info. - struct GlobalSplitCandidate { - // Register intended for assignment, or 0. - MCRegister PhysReg; - - // SplitKit interval index for this candidate. - unsigned IntvIdx; - - // Interference for PhysReg. - InterferenceCache::Cursor Intf; - - // Bundles where this candidate should be live. - BitVector LiveBundles; - SmallVector ActiveBlocks; - - void reset(InterferenceCache &Cache, MCRegister Reg) { - PhysReg = Reg; - IntvIdx = 0; - Intf.setPhysReg(Cache, Reg); - LiveBundles.clear(); - ActiveBlocks.clear(); - } - - // Set B[I] = C for every live bundle where B[I] was NoCand. - unsigned getBundles(SmallVectorImpl &B, unsigned C) { - unsigned Count = 0; - for (unsigned I : LiveBundles.set_bits()) - if (B[I] == NoCand) { - B[I] = C; - Count++; - } - return Count; - } - }; - - /// Candidate info for each PhysReg in AllocationOrder. - /// This vector never shrinks, but grows to the size of the largest register - /// class. - SmallVector GlobalCand; - - enum : unsigned { NoCand = ~0u }; - - /// Candidate map. Each edge bundle is assigned to a GlobalCand entry, or to - /// NoCand which indicates the stack interval. - SmallVector BundleCand; - - /// Callee-save register cost, calculated once per machine function. - BlockFrequency CSRCost; - - /// Set of broken hints that may be reconciled later because of eviction. - SmallSetVector SetOfBrokenHints; - - /// The register cost values. This list will be recreated for each Machine - /// Function - ArrayRef RegCosts; - - /// Flags for the live range priority calculation, determined once per - /// machine function. - bool RegClassPriorityTrumpsGlobalness; - -public: - RAGreedy(const RegClassFilterFunc F = allocateAllRegClasses); - - /// Return the pass name. - StringRef getPassName() const override { return "Greedy Register Allocator"; } - - /// RAGreedy analysis usage. - void getAnalysisUsage(AnalysisUsage &AU) const override; - void releaseMemory() override; - Spiller &spiller() override { return *SpillerInstance; } - void enqueueImpl(const LiveInterval *LI) override; - const LiveInterval *dequeue() override; - MCRegister selectOrSplit(const LiveInterval &, - SmallVectorImpl &) override; - void aboutToRemoveInterval(const LiveInterval &) override; - - /// Perform register allocation. - bool runOnMachineFunction(MachineFunction &mf) override; - - MachineFunctionProperties getRequiredProperties() const override { - return MachineFunctionProperties().set( - MachineFunctionProperties::Property::NoPHIs); - } - - MachineFunctionProperties getClearedProperties() const override { - return MachineFunctionProperties().set( - MachineFunctionProperties::Property::IsSSA); - } - - static char ID; - -private: - MCRegister selectOrSplitImpl(const LiveInterval &, - SmallVectorImpl &, SmallVirtRegSet &, - RecoloringStack &, unsigned = 0); - - bool LRE_CanEraseVirtReg(Register) override; - void LRE_WillShrinkVirtReg(Register) override; - void LRE_DidCloneVirtReg(Register, Register) override; - void enqueue(PQueue &CurQueue, const LiveInterval *LI); - const LiveInterval *dequeue(PQueue &CurQueue); - - BlockFrequency calcSpillCost(); - bool addSplitConstraints(InterferenceCache::Cursor, BlockFrequency &); - bool addThroughConstraints(InterferenceCache::Cursor, ArrayRef); - bool growRegion(GlobalSplitCandidate &Cand); - bool splitCanCauseEvictionChain(Register Evictee, GlobalSplitCandidate &Cand, - unsigned BBNumber, - const AllocationOrder &Order); - BlockFrequency calcGlobalSplitCost(GlobalSplitCandidate &, - const AllocationOrder &Order); - bool calcCompactRegion(GlobalSplitCandidate &); - void splitAroundRegion(LiveRangeEdit &, ArrayRef); - void calcGapWeights(MCRegister, SmallVectorImpl &); - bool canEvictInterferenceInRange(const LiveInterval &VirtReg, - MCRegister PhysReg, SlotIndex Start, - SlotIndex End, EvictionCost &MaxCost) const; - MCRegister getCheapestEvicteeWeight(const AllocationOrder &Order, - const LiveInterval &VirtReg, - SlotIndex Start, SlotIndex End, - float *BestEvictWeight) const; - void evictInterference(const LiveInterval &, MCRegister, - SmallVectorImpl &); - bool mayRecolorAllInterferences(MCRegister PhysReg, - const LiveInterval &VirtReg, - SmallLISet &RecoloringCandidates, - const SmallVirtRegSet &FixedRegisters); - - MCRegister tryAssign(const LiveInterval &, AllocationOrder &, - SmallVectorImpl &, const SmallVirtRegSet &); - MCRegister tryEvict(const LiveInterval &, AllocationOrder &, - SmallVectorImpl &, uint8_t, - const SmallVirtRegSet &); - MCRegister tryRegionSplit(const LiveInterval &, AllocationOrder &, - SmallVectorImpl &); - /// Calculate cost of region splitting. - unsigned calculateRegionSplitCost(const LiveInterval &VirtReg, - AllocationOrder &Order, - BlockFrequency &BestCost, - unsigned &NumCands, bool IgnoreCSR); - /// Perform region splitting. - unsigned doRegionSplit(const LiveInterval &VirtReg, unsigned BestCand, - bool HasCompact, SmallVectorImpl &NewVRegs); - /// Check other options before using a callee-saved register for the first - /// time. - MCRegister tryAssignCSRFirstTime(const LiveInterval &VirtReg, - AllocationOrder &Order, MCRegister PhysReg, - uint8_t &CostPerUseLimit, - SmallVectorImpl &NewVRegs); - void initializeCSRCost(); - unsigned tryBlockSplit(const LiveInterval &, AllocationOrder &, - SmallVectorImpl &); - unsigned tryInstructionSplit(const LiveInterval &, AllocationOrder &, - SmallVectorImpl &); - unsigned tryLocalSplit(const LiveInterval &, AllocationOrder &, - SmallVectorImpl &); - unsigned trySplit(const LiveInterval &, AllocationOrder &, - SmallVectorImpl &, const SmallVirtRegSet &); - unsigned tryLastChanceRecoloring(const LiveInterval &, AllocationOrder &, - SmallVectorImpl &, - SmallVirtRegSet &, RecoloringStack &, - unsigned); - bool tryRecoloringCandidates(PQueue &, SmallVectorImpl &, - SmallVirtRegSet &, RecoloringStack &, unsigned); - void tryHintRecoloring(const LiveInterval &); - void tryHintsRecoloring(); - - /// Model the information carried by one end of a copy. - struct HintInfo { - /// The frequency of the copy. - BlockFrequency Freq; - /// The virtual register or physical register. - Register Reg; - /// Its currently assigned register. - /// In case of a physical register Reg == PhysReg. - MCRegister PhysReg; - - HintInfo(BlockFrequency Freq, Register Reg, MCRegister PhysReg) - : Freq(Freq), Reg(Reg), PhysReg(PhysReg) {} - }; - using HintsInfo = SmallVector; - - BlockFrequency getBrokenHintFreq(const HintsInfo &, MCRegister); - void collectHintInfo(Register, HintsInfo &); - - /// Greedy RA statistic to remark. - struct RAGreedyStats { - unsigned Reloads = 0; - unsigned FoldedReloads = 0; - unsigned ZeroCostFoldedReloads = 0; - unsigned Spills = 0; - unsigned FoldedSpills = 0; - unsigned Copies = 0; - float ReloadsCost = 0.0f; - float FoldedReloadsCost = 0.0f; - float SpillsCost = 0.0f; - float FoldedSpillsCost = 0.0f; - float CopiesCost = 0.0f; - - bool isEmpty() { - return !(Reloads || FoldedReloads || Spills || FoldedSpills || - ZeroCostFoldedReloads || Copies); - } - - void add(RAGreedyStats other) { - Reloads += other.Reloads; - FoldedReloads += other.FoldedReloads; - ZeroCostFoldedReloads += other.ZeroCostFoldedReloads; - Spills += other.Spills; - FoldedSpills += other.FoldedSpills; - Copies += other.Copies; - ReloadsCost += other.ReloadsCost; - FoldedReloadsCost += other.FoldedReloadsCost; - SpillsCost += other.SpillsCost; - FoldedSpillsCost += other.FoldedSpillsCost; - CopiesCost += other.CopiesCost; - } - - void report(MachineOptimizationRemarkMissed &R); - }; - - /// Compute statistic for a basic block. - RAGreedyStats computeStats(MachineBasicBlock &MBB); - - /// Compute and report statistic through a remark. - RAGreedyStats reportStats(MachineLoop *L); - - /// Report the statistic for each loop. - void reportStats(); -}; -} // namespace llvm -#endif // #ifndef LLVM_CODEGEN_REGALLOCGREEDY_H_ diff --git a/llvm/lib/CodeGen/RegAllocGreedy.cpp b/llvm/lib/CodeGen/RegAllocGreedy.cpp deleted file mode 100644 --- a/llvm/lib/CodeGen/RegAllocGreedy.cpp +++ /dev/null @@ -1,2741 +0,0 @@ -//===- RegAllocGreedy.cpp - greedy register allocator ---------------------===// -// -// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. -// See https://llvm.org/LICENSE.txt for license information. -// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception -// -//===----------------------------------------------------------------------===// -// -// This file defines the RAGreedy function pass for register allocation in -// optimized builds. -// -//===----------------------------------------------------------------------===// - -#include "RegAllocGreedy.h" -#include "AllocationOrder.h" -#include "InterferenceCache.h" -#include "LiveDebugVariables.h" -#include "RegAllocBase.h" -#include "RegAllocEvictionAdvisor.h" -#include "SpillPlacement.h" -#include "SplitKit.h" -#include "llvm/ADT/ArrayRef.h" -#include "llvm/ADT/BitVector.h" -#include "llvm/ADT/IndexedMap.h" -#include "llvm/ADT/SetVector.h" -#include "llvm/ADT/SmallPtrSet.h" -#include "llvm/ADT/SmallSet.h" -#include "llvm/ADT/SmallVector.h" -#include "llvm/ADT/Statistic.h" -#include "llvm/ADT/StringRef.h" -#include "llvm/Analysis/AliasAnalysis.h" -#include "llvm/Analysis/OptimizationRemarkEmitter.h" -#include "llvm/CodeGen/CalcSpillWeights.h" -#include "llvm/CodeGen/EdgeBundles.h" -#include "llvm/CodeGen/LiveInterval.h" -#include "llvm/CodeGen/LiveIntervalUnion.h" -#include "llvm/CodeGen/LiveIntervals.h" -#include "llvm/CodeGen/LiveRangeEdit.h" -#include "llvm/CodeGen/LiveRegMatrix.h" -#include "llvm/CodeGen/LiveStacks.h" -#include "llvm/CodeGen/MachineBasicBlock.h" -#include "llvm/CodeGen/MachineBlockFrequencyInfo.h" -#include "llvm/CodeGen/MachineDominators.h" -#include "llvm/CodeGen/MachineFrameInfo.h" -#include "llvm/CodeGen/MachineFunction.h" -#include "llvm/CodeGen/MachineFunctionPass.h" -#include "llvm/CodeGen/MachineInstr.h" -#include "llvm/CodeGen/MachineLoopInfo.h" -#include "llvm/CodeGen/MachineOperand.h" -#include "llvm/CodeGen/MachineOptimizationRemarkEmitter.h" -#include "llvm/CodeGen/MachineRegisterInfo.h" -#include "llvm/CodeGen/RegAllocRegistry.h" -#include "llvm/CodeGen/RegisterClassInfo.h" -#include "llvm/CodeGen/SlotIndexes.h" -#include "llvm/CodeGen/Spiller.h" -#include "llvm/CodeGen/TargetInstrInfo.h" -#include "llvm/CodeGen/TargetRegisterInfo.h" -#include "llvm/CodeGen/TargetSubtargetInfo.h" -#include "llvm/CodeGen/VirtRegMap.h" -#include "llvm/IR/DebugInfoMetadata.h" -#include "llvm/IR/Function.h" -#include "llvm/IR/LLVMContext.h" -#include "llvm/InitializePasses.h" -#include "llvm/MC/MCRegisterInfo.h" -#include "llvm/Pass.h" -#include "llvm/Support/BlockFrequency.h" -#include "llvm/Support/BranchProbability.h" -#include "llvm/Support/CommandLine.h" -#include "llvm/Support/Debug.h" -#include "llvm/Support/MathExtras.h" -#include "llvm/Support/Timer.h" -#include "llvm/Support/raw_ostream.h" -#include -#include -#include -#include - -using namespace llvm; - -#define DEBUG_TYPE "regalloc" - -STATISTIC(NumGlobalSplits, "Number of split global live ranges"); -STATISTIC(NumLocalSplits, "Number of split local live ranges"); -STATISTIC(NumEvicted, "Number of interferences evicted"); - -static cl::opt SplitSpillMode( - "split-spill-mode", cl::Hidden, - cl::desc("Spill mode for splitting live ranges"), - cl::values(clEnumValN(SplitEditor::SM_Partition, "default", "Default"), - clEnumValN(SplitEditor::SM_Size, "size", "Optimize for size"), - clEnumValN(SplitEditor::SM_Speed, "speed", "Optimize for speed")), - cl::init(SplitEditor::SM_Speed)); - -static cl::opt -LastChanceRecoloringMaxDepth("lcr-max-depth", cl::Hidden, - cl::desc("Last chance recoloring max depth"), - cl::init(5)); - -static cl::opt LastChanceRecoloringMaxInterference( - "lcr-max-interf", cl::Hidden, - cl::desc("Last chance recoloring maximum number of considered" - " interference at a time"), - cl::init(8)); - -static cl::opt ExhaustiveSearch( - "exhaustive-register-search", cl::NotHidden, - cl::desc("Exhaustive Search for registers bypassing the depth " - "and interference cutoffs of last chance recoloring"), - cl::Hidden); - -static cl::opt EnableDeferredSpilling( - "enable-deferred-spilling", cl::Hidden, - cl::desc("Instead of spilling a variable right away, defer the actual " - "code insertion to the end of the allocation. That way the " - "allocator might still find a suitable coloring for this " - "variable because of other evicted variables."), - cl::init(false)); - -// FIXME: Find a good default for this flag and remove the flag. -static cl::opt -CSRFirstTimeCost("regalloc-csr-first-time-cost", - cl::desc("Cost for first time use of callee-saved register."), - cl::init(0), cl::Hidden); - -static cl::opt GrowRegionComplexityBudget( - "grow-region-complexity-budget", - cl::desc("growRegion() does not scale with the number of BB edges, so " - "limit its budget and bail out once we reach the limit."), - cl::init(10000), cl::Hidden); - -static cl::opt GreedyRegClassPriorityTrumpsGlobalness( - "greedy-regclass-priority-trumps-globalness", - cl::desc("Change the greedy register allocator's live range priority " - "calculation to make the AllocationPriority of the register class " - "more important then whether the range is global"), - cl::Hidden); - -static RegisterRegAlloc greedyRegAlloc("greedy", "greedy register allocator", - createGreedyRegisterAllocator); - -char RAGreedy::ID = 0; -char &llvm::RAGreedyID = RAGreedy::ID; - -INITIALIZE_PASS_BEGIN(RAGreedy, "greedy", - "Greedy Register Allocator", false, false) -INITIALIZE_PASS_DEPENDENCY(LiveDebugVariables) -INITIALIZE_PASS_DEPENDENCY(SlotIndexes) -INITIALIZE_PASS_DEPENDENCY(LiveIntervals) -INITIALIZE_PASS_DEPENDENCY(RegisterCoalescer) -INITIALIZE_PASS_DEPENDENCY(MachineScheduler) -INITIALIZE_PASS_DEPENDENCY(LiveStacks) -INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree) -INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo) -INITIALIZE_PASS_DEPENDENCY(VirtRegMap) -INITIALIZE_PASS_DEPENDENCY(LiveRegMatrix) -INITIALIZE_PASS_DEPENDENCY(EdgeBundles) -INITIALIZE_PASS_DEPENDENCY(SpillPlacement) -INITIALIZE_PASS_DEPENDENCY(MachineOptimizationRemarkEmitterPass) -INITIALIZE_PASS_DEPENDENCY(RegAllocEvictionAdvisorAnalysis) -INITIALIZE_PASS_END(RAGreedy, "greedy", - "Greedy Register Allocator", false, false) - -#ifndef NDEBUG -const char *const RAGreedy::StageName[] = { - "RS_New", - "RS_Assign", - "RS_Split", - "RS_Split2", - "RS_Spill", - "RS_Memory", - "RS_Done" -}; -#endif - -// Hysteresis to use when comparing floats. -// This helps stabilize decisions based on float comparisons. -const float Hysteresis = (2007 / 2048.0f); // 0.97998046875 - -FunctionPass* llvm::createGreedyRegisterAllocator() { - return new RAGreedy(); -} - -namespace llvm { -FunctionPass* createGreedyRegisterAllocator( - std::function Ftor); - -} - -FunctionPass* llvm::createGreedyRegisterAllocator( - std::function Ftor) { - return new RAGreedy(Ftor); -} - -RAGreedy::RAGreedy(RegClassFilterFunc F): - MachineFunctionPass(ID), - RegAllocBase(F) { -} - -void RAGreedy::getAnalysisUsage(AnalysisUsage &AU) const { - AU.setPreservesCFG(); - AU.addRequired(); - AU.addPreserved(); - AU.addRequired(); - AU.addPreserved(); - AU.addRequired(); - AU.addPreserved(); - AU.addRequired(); - AU.addPreserved(); - AU.addRequired(); - AU.addPreserved(); - AU.addRequired(); - AU.addPreserved(); - AU.addRequired(); - AU.addPreserved(); - AU.addRequired(); - AU.addPreserved(); - AU.addRequired(); - AU.addPreserved(); - AU.addRequired(); - AU.addPreserved(); - AU.addRequired(); - AU.addRequired(); - AU.addRequired(); - AU.addRequired(); - MachineFunctionPass::getAnalysisUsage(AU); -} - -//===----------------------------------------------------------------------===// -// LiveRangeEdit delegate methods -//===----------------------------------------------------------------------===// - -bool RAGreedy::LRE_CanEraseVirtReg(Register VirtReg) { - LiveInterval &LI = LIS->getInterval(VirtReg); - if (VRM->hasPhys(VirtReg)) { - Matrix->unassign(LI); - aboutToRemoveInterval(LI); - return true; - } - // Unassigned virtreg is probably in the priority queue. - // RegAllocBase will erase it after dequeueing. - // Nonetheless, clear the live-range so that the debug - // dump will show the right state for that VirtReg. - LI.clear(); - return false; -} - -void RAGreedy::LRE_WillShrinkVirtReg(Register VirtReg) { - if (!VRM->hasPhys(VirtReg)) - return; - - // Register is assigned, put it back on the queue for reassignment. - LiveInterval &LI = LIS->getInterval(VirtReg); - Matrix->unassign(LI); - RegAllocBase::enqueue(&LI); -} - -void RAGreedy::LRE_DidCloneVirtReg(Register New, Register Old) { - ExtraInfo->LRE_DidCloneVirtReg(New, Old); -} - -void RAGreedy::ExtraRegInfo::LRE_DidCloneVirtReg(Register New, Register Old) { - // Cloning a register we haven't even heard about yet? Just ignore it. - if (!Info.inBounds(Old)) - return; - - // LRE may clone a virtual register because dead code elimination causes it to - // be split into connected components. The new components are much smaller - // than the original, so they should get a new chance at being assigned. - // same stage as the parent. - Info[Old].Stage = RS_Assign; - Info.grow(New.id()); - Info[New] = Info[Old]; -} - -void RAGreedy::releaseMemory() { - SpillerInstance.reset(); - GlobalCand.clear(); -} - -void RAGreedy::enqueueImpl(const LiveInterval *LI) { enqueue(Queue, LI); } - -void RAGreedy::enqueue(PQueue &CurQueue, const LiveInterval *LI) { - // Prioritize live ranges by size, assigning larger ranges first. - // The queue holds (size, reg) pairs. - const unsigned Size = LI->getSize(); - const Register Reg = LI->reg(); - assert(Reg.isVirtual() && "Can only enqueue virtual registers"); - unsigned Prio; - - auto Stage = ExtraInfo->getOrInitStage(Reg); - if (Stage == RS_New) { - Stage = RS_Assign; - ExtraInfo->setStage(Reg, Stage); - } - if (Stage == RS_Split) { - // Unsplit ranges that couldn't be allocated immediately are deferred until - // everything else has been allocated. - Prio = Size; - } else if (Stage == RS_Memory) { - // Memory operand should be considered last. - // Change the priority such that Memory operand are assigned in - // the reverse order that they came in. - // TODO: Make this a member variable and probably do something about hints. - static unsigned MemOp = 0; - Prio = MemOp++; - } else { - // Giant live ranges fall back to the global assignment heuristic, which - // prevents excessive spilling in pathological cases. - bool ReverseLocal = TRI->reverseLocalAssignment(); - const TargetRegisterClass &RC = *MRI->getRegClass(Reg); - bool ForceGlobal = - !ReverseLocal && (Size / SlotIndex::InstrDist) > - (2 * RegClassInfo.getNumAllocatableRegs(&RC)); - unsigned GlobalBit = 0; - - if (Stage == RS_Assign && !ForceGlobal && !LI->empty() && - LIS->intervalIsInOneMBB(*LI)) { - // Allocate original local ranges in linear instruction order. Since they - // are singly defined, this produces optimal coloring in the absence of - // global interference and other constraints. - if (!ReverseLocal) - Prio = LI->beginIndex().getInstrDistance(Indexes->getLastIndex()); - else { - // Allocating bottom up may allow many short LRGs to be assigned first - // to one of the cheap registers. This could be much faster for very - // large blocks on targets with many physical registers. - Prio = Indexes->getZeroIndex().getInstrDistance(LI->endIndex()); - } - } else { - // Allocate global and split ranges in long->short order. Long ranges that - // don't fit should be spilled (or split) ASAP so they don't create - // interference. Mark a bit to prioritize global above local ranges. - Prio = Size; - GlobalBit = 1; - } - if (RegClassPriorityTrumpsGlobalness) - Prio |= RC.AllocationPriority << 25 | GlobalBit << 24; - else - Prio |= GlobalBit << 29 | RC.AllocationPriority << 24; - - // Mark a higher bit to prioritize global and local above RS_Split. - Prio |= (1u << 31); - - // Boost ranges that have a physical register hint. - if (VRM->hasKnownPreference(Reg)) - Prio |= (1u << 30); - } - // The virtual register number is a tie breaker for same-sized ranges. - // Give lower vreg numbers higher priority to assign them first. - CurQueue.push(std::make_pair(Prio, ~Reg)); -} - -const LiveInterval *RAGreedy::dequeue() { return dequeue(Queue); } - -const LiveInterval *RAGreedy::dequeue(PQueue &CurQueue) { - if (CurQueue.empty()) - return nullptr; - LiveInterval *LI = &LIS->getInterval(~CurQueue.top().second); - CurQueue.pop(); - return LI; -} - -//===----------------------------------------------------------------------===// -// Direct Assignment -//===----------------------------------------------------------------------===// - -/// tryAssign - Try to assign VirtReg to an available register. -MCRegister RAGreedy::tryAssign(const LiveInterval &VirtReg, - AllocationOrder &Order, - SmallVectorImpl &NewVRegs, - const SmallVirtRegSet &FixedRegisters) { - MCRegister PhysReg; - for (auto I = Order.begin(), E = Order.end(); I != E && !PhysReg; ++I) { - assert(*I); - if (!Matrix->checkInterference(VirtReg, *I)) { - if (I.isHint()) - return *I; - else - PhysReg = *I; - } - } - if (!PhysReg.isValid()) - return PhysReg; - - // PhysReg is available, but there may be a better choice. - - // If we missed a simple hint, try to cheaply evict interference from the - // preferred register. - if (Register Hint = MRI->getSimpleHint(VirtReg.reg())) - if (Order.isHint(Hint)) { - MCRegister PhysHint = Hint.asMCReg(); - LLVM_DEBUG(dbgs() << "missed hint " << printReg(PhysHint, TRI) << '\n'); - - if (EvictAdvisor->canEvictHintInterference(VirtReg, PhysHint, - FixedRegisters)) { - evictInterference(VirtReg, PhysHint, NewVRegs); - return PhysHint; - } - // Record the missed hint, we may be able to recover - // at the end if the surrounding allocation changed. - SetOfBrokenHints.insert(&VirtReg); - } - - // Try to evict interference from a cheaper alternative. - uint8_t Cost = RegCosts[PhysReg]; - - // Most registers have 0 additional cost. - if (!Cost) - return PhysReg; - - LLVM_DEBUG(dbgs() << printReg(PhysReg, TRI) << " is available at cost " - << (unsigned)Cost << '\n'); - MCRegister CheapReg = tryEvict(VirtReg, Order, NewVRegs, Cost, FixedRegisters); - return CheapReg ? CheapReg : PhysReg; -} - -//===----------------------------------------------------------------------===// -// Interference eviction -//===----------------------------------------------------------------------===// - -Register RegAllocEvictionAdvisor::canReassign(const LiveInterval &VirtReg, - Register PrevReg) const { - auto Order = - AllocationOrder::create(VirtReg.reg(), *VRM, RegClassInfo, Matrix); - MCRegister PhysReg; - for (auto I = Order.begin(), E = Order.end(); I != E && !PhysReg; ++I) { - if ((*I).id() == PrevReg.id()) - continue; - - MCRegUnitIterator Units(*I, TRI); - for (; Units.isValid(); ++Units) { - // Instantiate a "subquery", not to be confused with the Queries array. - LiveIntervalUnion::Query subQ(VirtReg, Matrix->getLiveUnions()[*Units]); - if (subQ.checkInterference()) - break; - } - // If no units have interference, break out with the current PhysReg. - if (!Units.isValid()) - PhysReg = *I; - } - if (PhysReg) - LLVM_DEBUG(dbgs() << "can reassign: " << VirtReg << " from " - << printReg(PrevReg, TRI) << " to " - << printReg(PhysReg, TRI) << '\n'); - return PhysReg; -} - -/// Return true if all interferences between VirtReg and PhysReg between -/// Start and End can be evicted. -/// -/// \param VirtReg Live range that is about to be assigned. -/// \param PhysReg Desired register for assignment. -/// \param Start Start of range to look for interferences. -/// \param End End of range to look for interferences. -/// \param MaxCost Only look for cheaper candidates and update with new cost -/// when returning true. -/// \return True when interference can be evicted cheaper than MaxCost. -bool RAGreedy::canEvictInterferenceInRange(const LiveInterval &VirtReg, - MCRegister PhysReg, SlotIndex Start, - SlotIndex End, - EvictionCost &MaxCost) const { - EvictionCost Cost; - - for (MCRegUnitIterator Units(PhysReg, TRI); Units.isValid(); ++Units) { - LiveIntervalUnion::Query &Q = Matrix->query(VirtReg, *Units); - - // Check if any interfering live range is heavier than MaxWeight. - for (const LiveInterval *Intf : reverse(Q.interferingVRegs())) { - // Check if interference overlast the segment in interest. - if (!Intf->overlaps(Start, End)) - continue; - - // Never evict spill products. They cannot split or spill. - if (ExtraInfo->getStage(*Intf) == RS_Done) - return false; - - // Would this break a satisfied hint? - bool BreaksHint = VRM->hasPreferredPhys(Intf->reg()); - // Update eviction cost. - Cost.BrokenHints += BreaksHint; - Cost.MaxWeight = std::max(Cost.MaxWeight, Intf->weight()); - // Abort if this would be too expensive. - if (!(Cost < MaxCost)) - return false; - } - } - - if (Cost.MaxWeight == 0) - return false; - - MaxCost = Cost; - return true; -} - -/// Return the physical register that will be best -/// candidate for eviction by a local split interval that will be created -/// between Start and End. -/// -/// \param Order The allocation order -/// \param VirtReg Live range that is about to be assigned. -/// \param Start Start of range to look for interferences -/// \param End End of range to look for interferences -/// \param BestEvictweight The eviction cost of that eviction -/// \return The PhysReg which is the best candidate for eviction and the -/// eviction cost in BestEvictweight -MCRegister RAGreedy::getCheapestEvicteeWeight(const AllocationOrder &Order, - const LiveInterval &VirtReg, - SlotIndex Start, SlotIndex End, - float *BestEvictweight) const { - EvictionCost BestEvictCost; - BestEvictCost.setMax(); - BestEvictCost.MaxWeight = VirtReg.weight(); - MCRegister BestEvicteePhys; - - // Go over all physical registers and find the best candidate for eviction - for (MCRegister PhysReg : Order.getOrder()) { - - if (!canEvictInterferenceInRange(VirtReg, PhysReg, Start, End, - BestEvictCost)) - continue; - - // Best so far. - BestEvicteePhys = PhysReg; - } - *BestEvictweight = BestEvictCost.MaxWeight; - return BestEvicteePhys; -} - -/// evictInterference - Evict any interferring registers that prevent VirtReg -/// from being assigned to Physreg. This assumes that canEvictInterference -/// returned true. -void RAGreedy::evictInterference(const LiveInterval &VirtReg, - MCRegister PhysReg, - SmallVectorImpl &NewVRegs) { - // Make sure that VirtReg has a cascade number, and assign that cascade - // number to every evicted register. These live ranges than then only be - // evicted by a newer cascade, preventing infinite loops. - unsigned Cascade = ExtraInfo->getOrAssignNewCascade(VirtReg.reg()); - - LLVM_DEBUG(dbgs() << "evicting " << printReg(PhysReg, TRI) - << " interference: Cascade " << Cascade << '\n'); - - // Collect all interfering virtregs first. - SmallVector Intfs; - for (MCRegUnitIterator Units(PhysReg, TRI); Units.isValid(); ++Units) { - LiveIntervalUnion::Query &Q = Matrix->query(VirtReg, *Units); - // We usually have the interfering VRegs cached so collectInterferingVRegs() - // should be fast, we may need to recalculate if when different physregs - // overlap the same register unit so we had different SubRanges queried - // against it. - ArrayRef IVR = Q.interferingVRegs(); - Intfs.append(IVR.begin(), IVR.end()); - } - - // Evict them second. This will invalidate the queries. - for (const LiveInterval *Intf : Intfs) { - // The same VirtReg may be present in multiple RegUnits. Skip duplicates. - if (!VRM->hasPhys(Intf->reg())) - continue; - - LastEvicted.addEviction(PhysReg, VirtReg.reg(), Intf->reg()); - - Matrix->unassign(*Intf); - assert((ExtraInfo->getCascade(Intf->reg()) < Cascade || - VirtReg.isSpillable() < Intf->isSpillable()) && - "Cannot decrease cascade number, illegal eviction"); - ExtraInfo->setCascade(Intf->reg(), Cascade); - ++NumEvicted; - NewVRegs.push_back(Intf->reg()); - } -} - -/// Returns true if the given \p PhysReg is a callee saved register and has not -/// been used for allocation yet. -bool RegAllocEvictionAdvisor::isUnusedCalleeSavedReg(MCRegister PhysReg) const { - MCRegister CSR = RegClassInfo.getLastCalleeSavedAlias(PhysReg); - if (!CSR) - return false; - - return !Matrix->isPhysRegUsed(PhysReg); -} - -Optional -RegAllocEvictionAdvisor::getOrderLimit(const LiveInterval &VirtReg, - const AllocationOrder &Order, - unsigned CostPerUseLimit) const { - unsigned OrderLimit = Order.getOrder().size(); - - if (CostPerUseLimit < uint8_t(~0u)) { - // Check of any registers in RC are below CostPerUseLimit. - const TargetRegisterClass *RC = MRI->getRegClass(VirtReg.reg()); - uint8_t MinCost = RegClassInfo.getMinCost(RC); - if (MinCost >= CostPerUseLimit) { - LLVM_DEBUG(dbgs() << TRI->getRegClassName(RC) << " minimum cost = " - << MinCost << ", no cheaper registers to be found.\n"); - return None; - } - - // It is normal for register classes to have a long tail of registers with - // the same cost. We don't need to look at them if they're too expensive. - if (RegCosts[Order.getOrder().back()] >= CostPerUseLimit) { - OrderLimit = RegClassInfo.getLastCostChange(RC); - LLVM_DEBUG(dbgs() << "Only trying the first " << OrderLimit - << " regs.\n"); - } - } - return OrderLimit; -} - -bool RegAllocEvictionAdvisor::canAllocatePhysReg(unsigned CostPerUseLimit, - MCRegister PhysReg) const { - if (RegCosts[PhysReg] >= CostPerUseLimit) - return false; - // The first use of a callee-saved register in a function has cost 1. - // Don't start using a CSR when the CostPerUseLimit is low. - if (CostPerUseLimit == 1 && isUnusedCalleeSavedReg(PhysReg)) { - LLVM_DEBUG( - dbgs() << printReg(PhysReg, TRI) << " would clobber CSR " - << printReg(RegClassInfo.getLastCalleeSavedAlias(PhysReg), TRI) - << '\n'); - return false; - } - return true; -} - -/// tryEvict - Try to evict all interferences for a physreg. -/// @param VirtReg Currently unassigned virtual register. -/// @param Order Physregs to try. -/// @return Physreg to assign VirtReg, or 0. -MCRegister RAGreedy::tryEvict(const LiveInterval &VirtReg, - AllocationOrder &Order, - SmallVectorImpl &NewVRegs, - uint8_t CostPerUseLimit, - const SmallVirtRegSet &FixedRegisters) { - NamedRegionTimer T("evict", "Evict", TimerGroupName, TimerGroupDescription, - TimePassesIsEnabled); - - MCRegister BestPhys = EvictAdvisor->tryFindEvictionCandidate( - VirtReg, Order, CostPerUseLimit, FixedRegisters); - if (BestPhys.isValid()) - evictInterference(VirtReg, BestPhys, NewVRegs); - return BestPhys; -} - -//===----------------------------------------------------------------------===// -// Region Splitting -//===----------------------------------------------------------------------===// - -/// addSplitConstraints - Fill out the SplitConstraints vector based on the -/// interference pattern in Physreg and its aliases. Add the constraints to -/// SpillPlacement and return the static cost of this split in Cost, assuming -/// that all preferences in SplitConstraints are met. -/// Return false if there are no bundles with positive bias. -bool RAGreedy::addSplitConstraints(InterferenceCache::Cursor Intf, - BlockFrequency &Cost) { - ArrayRef UseBlocks = SA->getUseBlocks(); - - // Reset interference dependent info. - SplitConstraints.resize(UseBlocks.size()); - BlockFrequency StaticCost = 0; - for (unsigned I = 0; I != UseBlocks.size(); ++I) { - const SplitAnalysis::BlockInfo &BI = UseBlocks[I]; - SpillPlacement::BlockConstraint &BC = SplitConstraints[I]; - - BC.Number = BI.MBB->getNumber(); - Intf.moveToBlock(BC.Number); - BC.Entry = BI.LiveIn ? SpillPlacement::PrefReg : SpillPlacement::DontCare; - BC.Exit = (BI.LiveOut && - !LIS->getInstructionFromIndex(BI.LastInstr)->isImplicitDef()) - ? SpillPlacement::PrefReg - : SpillPlacement::DontCare; - BC.ChangesValue = BI.FirstDef.isValid(); - - if (!Intf.hasInterference()) - continue; - - // Number of spill code instructions to insert. - unsigned Ins = 0; - - // Interference for the live-in value. - if (BI.LiveIn) { - if (Intf.first() <= Indexes->getMBBStartIdx(BC.Number)) { - BC.Entry = SpillPlacement::MustSpill; - ++Ins; - } else if (Intf.first() < BI.FirstInstr) { - BC.Entry = SpillPlacement::PrefSpill; - ++Ins; - } else if (Intf.first() < BI.LastInstr) { - ++Ins; - } - - // Abort if the spill cannot be inserted at the MBB' start - if (((BC.Entry == SpillPlacement::MustSpill) || - (BC.Entry == SpillPlacement::PrefSpill)) && - SlotIndex::isEarlierInstr(BI.FirstInstr, - SA->getFirstSplitPoint(BC.Number))) - return false; - } - - // Interference for the live-out value. - if (BI.LiveOut) { - if (Intf.last() >= SA->getLastSplitPoint(BC.Number)) { - BC.Exit = SpillPlacement::MustSpill; - ++Ins; - } else if (Intf.last() > BI.LastInstr) { - BC.Exit = SpillPlacement::PrefSpill; - ++Ins; - } else if (Intf.last() > BI.FirstInstr) { - ++Ins; - } - } - - // Accumulate the total frequency of inserted spill code. - while (Ins--) - StaticCost += SpillPlacer->getBlockFrequency(BC.Number); - } - Cost = StaticCost; - - // Add constraints for use-blocks. Note that these are the only constraints - // that may add a positive bias, it is downhill from here. - SpillPlacer->addConstraints(SplitConstraints); - return SpillPlacer->scanActiveBundles(); -} - -/// addThroughConstraints - Add constraints and links to SpillPlacer from the -/// live-through blocks in Blocks. -bool RAGreedy::addThroughConstraints(InterferenceCache::Cursor Intf, - ArrayRef Blocks) { - const unsigned GroupSize = 8; - SpillPlacement::BlockConstraint BCS[GroupSize]; - unsigned TBS[GroupSize]; - unsigned B = 0, T = 0; - - for (unsigned Number : Blocks) { - Intf.moveToBlock(Number); - - if (!Intf.hasInterference()) { - assert(T < GroupSize && "Array overflow"); - TBS[T] = Number; - if (++T == GroupSize) { - SpillPlacer->addLinks(makeArrayRef(TBS, T)); - T = 0; - } - continue; - } - - assert(B < GroupSize && "Array overflow"); - BCS[B].Number = Number; - - // Abort if the spill cannot be inserted at the MBB' start - MachineBasicBlock *MBB = MF->getBlockNumbered(Number); - auto FirstNonDebugInstr = MBB->getFirstNonDebugInstr(); - if (FirstNonDebugInstr != MBB->end() && - SlotIndex::isEarlierInstr(LIS->getInstructionIndex(*FirstNonDebugInstr), - SA->getFirstSplitPoint(Number))) - return false; - // Interference for the live-in value. - if (Intf.first() <= Indexes->getMBBStartIdx(Number)) - BCS[B].Entry = SpillPlacement::MustSpill; - else - BCS[B].Entry = SpillPlacement::PrefSpill; - - // Interference for the live-out value. - if (Intf.last() >= SA->getLastSplitPoint(Number)) - BCS[B].Exit = SpillPlacement::MustSpill; - else - BCS[B].Exit = SpillPlacement::PrefSpill; - - if (++B == GroupSize) { - SpillPlacer->addConstraints(makeArrayRef(BCS, B)); - B = 0; - } - } - - SpillPlacer->addConstraints(makeArrayRef(BCS, B)); - SpillPlacer->addLinks(makeArrayRef(TBS, T)); - return true; -} - -bool RAGreedy::growRegion(GlobalSplitCandidate &Cand) { - // Keep track of through blocks that have not been added to SpillPlacer. - BitVector Todo = SA->getThroughBlocks(); - SmallVectorImpl &ActiveBlocks = Cand.ActiveBlocks; - unsigned AddedTo = 0; -#ifndef NDEBUG - unsigned Visited = 0; -#endif - - unsigned long Budget = GrowRegionComplexityBudget; - while (true) { - ArrayRef NewBundles = SpillPlacer->getRecentPositive(); - // Find new through blocks in the periphery of PrefRegBundles. - for (unsigned Bundle : NewBundles) { - // Look at all blocks connected to Bundle in the full graph. - ArrayRef Blocks = Bundles->getBlocks(Bundle); - // Limit compilation time by bailing out after we use all our budget. - if (Blocks.size() >= Budget) - return false; - Budget -= Blocks.size(); - for (unsigned Block : Blocks) { - if (!Todo.test(Block)) - continue; - Todo.reset(Block); - // This is a new through block. Add it to SpillPlacer later. - ActiveBlocks.push_back(Block); -#ifndef NDEBUG - ++Visited; -#endif - } - } - // Any new blocks to add? - if (ActiveBlocks.size() == AddedTo) - break; - - // Compute through constraints from the interference, or assume that all - // through blocks prefer spilling when forming compact regions. - auto NewBlocks = makeArrayRef(ActiveBlocks).slice(AddedTo); - if (Cand.PhysReg) { - if (!addThroughConstraints(Cand.Intf, NewBlocks)) - return false; - } else - // Provide a strong negative bias on through blocks to prevent unwanted - // liveness on loop backedges. - SpillPlacer->addPrefSpill(NewBlocks, /* Strong= */ true); - AddedTo = ActiveBlocks.size(); - - // Perhaps iterating can enable more bundles? - SpillPlacer->iterate(); - } - LLVM_DEBUG(dbgs() << ", v=" << Visited); - return true; -} - -/// calcCompactRegion - Compute the set of edge bundles that should be live -/// when splitting the current live range into compact regions. Compact -/// regions can be computed without looking at interference. They are the -/// regions formed by removing all the live-through blocks from the live range. -/// -/// Returns false if the current live range is already compact, or if the -/// compact regions would form single block regions anyway. -bool RAGreedy::calcCompactRegion(GlobalSplitCandidate &Cand) { - // Without any through blocks, the live range is already compact. - if (!SA->getNumThroughBlocks()) - return false; - - // Compact regions don't correspond to any physreg. - Cand.reset(IntfCache, MCRegister::NoRegister); - - LLVM_DEBUG(dbgs() << "Compact region bundles"); - - // Use the spill placer to determine the live bundles. GrowRegion pretends - // that all the through blocks have interference when PhysReg is unset. - SpillPlacer->prepare(Cand.LiveBundles); - - // The static split cost will be zero since Cand.Intf reports no interference. - BlockFrequency Cost; - if (!addSplitConstraints(Cand.Intf, Cost)) { - LLVM_DEBUG(dbgs() << ", none.\n"); - return false; - } - - if (!growRegion(Cand)) { - LLVM_DEBUG(dbgs() << ", cannot spill all interferences.\n"); - return false; - } - - SpillPlacer->finish(); - - if (!Cand.LiveBundles.any()) { - LLVM_DEBUG(dbgs() << ", none.\n"); - return false; - } - - LLVM_DEBUG({ - for (int I : Cand.LiveBundles.set_bits()) - dbgs() << " EB#" << I; - dbgs() << ".\n"; - }); - return true; -} - -/// calcSpillCost - Compute how expensive it would be to split the live range in -/// SA around all use blocks instead of forming bundle regions. -BlockFrequency RAGreedy::calcSpillCost() { - BlockFrequency Cost = 0; - ArrayRef UseBlocks = SA->getUseBlocks(); - for (const SplitAnalysis::BlockInfo &BI : UseBlocks) { - unsigned Number = BI.MBB->getNumber(); - // We normally only need one spill instruction - a load or a store. - Cost += SpillPlacer->getBlockFrequency(Number); - - // Unless the value is redefined in the block. - if (BI.LiveIn && BI.LiveOut && BI.FirstDef) - Cost += SpillPlacer->getBlockFrequency(Number); - } - return Cost; -} - -/// Check if splitting Evictee will create a local split interval in -/// basic block number BBNumber that may cause a bad eviction chain. This is -/// intended to prevent bad eviction sequences like: -/// movl %ebp, 8(%esp) # 4-byte Spill -/// movl %ecx, %ebp -/// movl %ebx, %ecx -/// movl %edi, %ebx -/// movl %edx, %edi -/// cltd -/// idivl %esi -/// movl %edi, %edx -/// movl %ebx, %edi -/// movl %ecx, %ebx -/// movl %ebp, %ecx -/// movl 16(%esp), %ebp # 4 - byte Reload -/// -/// Such sequences are created in 2 scenarios: -/// -/// Scenario #1: -/// %0 is evicted from physreg0 by %1. -/// Evictee %0 is intended for region splitting with split candidate -/// physreg0 (the reg %0 was evicted from). -/// Region splitting creates a local interval because of interference with the -/// evictor %1 (normally region splitting creates 2 interval, the "by reg" -/// and "by stack" intervals and local interval created when interference -/// occurs). -/// One of the split intervals ends up evicting %2 from physreg1. -/// Evictee %2 is intended for region splitting with split candidate -/// physreg1. -/// One of the split intervals ends up evicting %3 from physreg2, etc. -/// -/// Scenario #2 -/// %0 is evicted from physreg0 by %1. -/// %2 is evicted from physreg2 by %3 etc. -/// Evictee %0 is intended for region splitting with split candidate -/// physreg1. -/// Region splitting creates a local interval because of interference with the -/// evictor %1. -/// One of the split intervals ends up evicting back original evictor %1 -/// from physreg0 (the reg %0 was evicted from). -/// Another evictee %2 is intended for region splitting with split candidate -/// physreg1. -/// One of the split intervals ends up evicting %3 from physreg2, etc. -/// -/// \param Evictee The register considered to be split. -/// \param Cand The split candidate that determines the physical register -/// we are splitting for and the interferences. -/// \param BBNumber The number of a BB for which the region split process will -/// create a local split interval. -/// \param Order The physical registers that may get evicted by a split -/// artifact of Evictee. -/// \return True if splitting Evictee may cause a bad eviction chain, false -/// otherwise. -bool RAGreedy::splitCanCauseEvictionChain(Register Evictee, - GlobalSplitCandidate &Cand, - unsigned BBNumber, - const AllocationOrder &Order) { - EvictionTrack::EvictorInfo VregEvictorInfo = LastEvicted.getEvictor(Evictee); - unsigned Evictor = VregEvictorInfo.first; - MCRegister PhysReg = VregEvictorInfo.second; - - // No actual evictor. - if (!Evictor || !PhysReg) - return false; - - float MaxWeight = 0; - MCRegister FutureEvictedPhysReg = - getCheapestEvicteeWeight(Order, LIS->getInterval(Evictee), - Cand.Intf.first(), Cand.Intf.last(), &MaxWeight); - - // The bad eviction chain occurs when either the split candidate is the - // evicting reg or one of the split artifact will evict the evicting reg. - if ((PhysReg != Cand.PhysReg) && (PhysReg != FutureEvictedPhysReg)) - return false; - - Cand.Intf.moveToBlock(BBNumber); - - // Check to see if the Evictor contains interference (with Evictee) in the - // given BB. If so, this interference caused the eviction of Evictee from - // PhysReg. This suggest that we will create a local interval during the - // region split to avoid this interference This local interval may cause a bad - // eviction chain. - if (!LIS->hasInterval(Evictor)) - return false; - LiveInterval &EvictorLI = LIS->getInterval(Evictor); - if (EvictorLI.FindSegmentContaining(Cand.Intf.first()) == EvictorLI.end()) - return false; - - // Now, check to see if the local interval we will create is going to be - // expensive enough to evict somebody If so, this may cause a bad eviction - // chain. - float splitArtifactWeight = - VRAI->futureWeight(LIS->getInterval(Evictee), - Cand.Intf.first().getPrevIndex(), Cand.Intf.last()); - if (splitArtifactWeight >= 0 && splitArtifactWeight < MaxWeight) - return false; - - return true; -} - -/// calcGlobalSplitCost - Return the global split cost of following the split -/// pattern in LiveBundles. This cost should be added to the local cost of the -/// interference pattern in SplitConstraints. -/// -BlockFrequency RAGreedy::calcGlobalSplitCost(GlobalSplitCandidate &Cand, - const AllocationOrder &Order) { - BlockFrequency GlobalCost = 0; - const BitVector &LiveBundles = Cand.LiveBundles; - ArrayRef UseBlocks = SA->getUseBlocks(); - for (unsigned I = 0; I != UseBlocks.size(); ++I) { - const SplitAnalysis::BlockInfo &BI = UseBlocks[I]; - SpillPlacement::BlockConstraint &BC = SplitConstraints[I]; - bool RegIn = LiveBundles[Bundles->getBundle(BC.Number, false)]; - bool RegOut = LiveBundles[Bundles->getBundle(BC.Number, true)]; - unsigned Ins = 0; - - Cand.Intf.moveToBlock(BC.Number); - - if (BI.LiveIn) - Ins += RegIn != (BC.Entry == SpillPlacement::PrefReg); - if (BI.LiveOut) - Ins += RegOut != (BC.Exit == SpillPlacement::PrefReg); - while (Ins--) - GlobalCost += SpillPlacer->getBlockFrequency(BC.Number); - } - - for (unsigned Number : Cand.ActiveBlocks) { - bool RegIn = LiveBundles[Bundles->getBundle(Number, false)]; - bool RegOut = LiveBundles[Bundles->getBundle(Number, true)]; - if (!RegIn && !RegOut) - continue; - if (RegIn && RegOut) { - // We need double spill code if this block has interference. - Cand.Intf.moveToBlock(Number); - if (Cand.Intf.hasInterference()) { - GlobalCost += SpillPlacer->getBlockFrequency(Number); - GlobalCost += SpillPlacer->getBlockFrequency(Number); - } - continue; - } - // live-in / stack-out or stack-in live-out. - GlobalCost += SpillPlacer->getBlockFrequency(Number); - } - return GlobalCost; -} - -/// splitAroundRegion - Split the current live range around the regions -/// determined by BundleCand and GlobalCand. -/// -/// Before calling this function, GlobalCand and BundleCand must be initialized -/// so each bundle is assigned to a valid candidate, or NoCand for the -/// stack-bound bundles. The shared SA/SE SplitAnalysis and SplitEditor -/// objects must be initialized for the current live range, and intervals -/// created for the used candidates. -/// -/// @param LREdit The LiveRangeEdit object handling the current split. -/// @param UsedCands List of used GlobalCand entries. Every BundleCand value -/// must appear in this list. -void RAGreedy::splitAroundRegion(LiveRangeEdit &LREdit, - ArrayRef UsedCands) { - // These are the intervals created for new global ranges. We may create more - // intervals for local ranges. - const unsigned NumGlobalIntvs = LREdit.size(); - LLVM_DEBUG(dbgs() << "splitAroundRegion with " << NumGlobalIntvs - << " globals.\n"); - assert(NumGlobalIntvs && "No global intervals configured"); - - // Isolate even single instructions when dealing with a proper sub-class. - // That guarantees register class inflation for the stack interval because it - // is all copies. - Register Reg = SA->getParent().reg(); - bool SingleInstrs = RegClassInfo.isProperSubClass(MRI->getRegClass(Reg)); - - // First handle all the blocks with uses. - ArrayRef UseBlocks = SA->getUseBlocks(); - for (const SplitAnalysis::BlockInfo &BI : UseBlocks) { - unsigned Number = BI.MBB->getNumber(); - unsigned IntvIn = 0, IntvOut = 0; - SlotIndex IntfIn, IntfOut; - if (BI.LiveIn) { - unsigned CandIn = BundleCand[Bundles->getBundle(Number, false)]; - if (CandIn != NoCand) { - GlobalSplitCandidate &Cand = GlobalCand[CandIn]; - IntvIn = Cand.IntvIdx; - Cand.Intf.moveToBlock(Number); - IntfIn = Cand.Intf.first(); - } - } - if (BI.LiveOut) { - unsigned CandOut = BundleCand[Bundles->getBundle(Number, true)]; - if (CandOut != NoCand) { - GlobalSplitCandidate &Cand = GlobalCand[CandOut]; - IntvOut = Cand.IntvIdx; - Cand.Intf.moveToBlock(Number); - IntfOut = Cand.Intf.last(); - } - } - - // Create separate intervals for isolated blocks with multiple uses. - if (!IntvIn && !IntvOut) { - LLVM_DEBUG(dbgs() << printMBBReference(*BI.MBB) << " isolated.\n"); - if (SA->shouldSplitSingleBlock(BI, SingleInstrs)) - SE->splitSingleBlock(BI); - continue; - } - - if (IntvIn && IntvOut) - SE->splitLiveThroughBlock(Number, IntvIn, IntfIn, IntvOut, IntfOut); - else if (IntvIn) - SE->splitRegInBlock(BI, IntvIn, IntfIn); - else - SE->splitRegOutBlock(BI, IntvOut, IntfOut); - } - - // Handle live-through blocks. The relevant live-through blocks are stored in - // the ActiveBlocks list with each candidate. We need to filter out - // duplicates. - BitVector Todo = SA->getThroughBlocks(); - for (unsigned UsedCand : UsedCands) { - ArrayRef Blocks = GlobalCand[UsedCand].ActiveBlocks; - for (unsigned Number : Blocks) { - if (!Todo.test(Number)) - continue; - Todo.reset(Number); - - unsigned IntvIn = 0, IntvOut = 0; - SlotIndex IntfIn, IntfOut; - - unsigned CandIn = BundleCand[Bundles->getBundle(Number, false)]; - if (CandIn != NoCand) { - GlobalSplitCandidate &Cand = GlobalCand[CandIn]; - IntvIn = Cand.IntvIdx; - Cand.Intf.moveToBlock(Number); - IntfIn = Cand.Intf.first(); - } - - unsigned CandOut = BundleCand[Bundles->getBundle(Number, true)]; - if (CandOut != NoCand) { - GlobalSplitCandidate &Cand = GlobalCand[CandOut]; - IntvOut = Cand.IntvIdx; - Cand.Intf.moveToBlock(Number); - IntfOut = Cand.Intf.last(); - } - if (!IntvIn && !IntvOut) - continue; - SE->splitLiveThroughBlock(Number, IntvIn, IntfIn, IntvOut, IntfOut); - } - } - - ++NumGlobalSplits; - - SmallVector IntvMap; - SE->finish(&IntvMap); - DebugVars->splitRegister(Reg, LREdit.regs(), *LIS); - - unsigned OrigBlocks = SA->getNumLiveBlocks(); - - // Sort out the new intervals created by splitting. We get four kinds: - // - Remainder intervals should not be split again. - // - Candidate intervals can be assigned to Cand.PhysReg. - // - Block-local splits are candidates for local splitting. - // - DCE leftovers should go back on the queue. - for (unsigned I = 0, E = LREdit.size(); I != E; ++I) { - const LiveInterval &Reg = LIS->getInterval(LREdit.get(I)); - - // Ignore old intervals from DCE. - if (ExtraInfo->getOrInitStage(Reg.reg()) != RS_New) - continue; - - // Remainder interval. Don't try splitting again, spill if it doesn't - // allocate. - if (IntvMap[I] == 0) { - ExtraInfo->setStage(Reg, RS_Spill); - continue; - } - - // Global intervals. Allow repeated splitting as long as the number of live - // blocks is strictly decreasing. - if (IntvMap[I] < NumGlobalIntvs) { - if (SA->countLiveBlocks(&Reg) >= OrigBlocks) { - LLVM_DEBUG(dbgs() << "Main interval covers the same " << OrigBlocks - << " blocks as original.\n"); - // Don't allow repeated splitting as a safe guard against looping. - ExtraInfo->setStage(Reg, RS_Split2); - } - continue; - } - - // Other intervals are treated as new. This includes local intervals created - // for blocks with multiple uses, and anything created by DCE. - } - - if (VerifyEnabled) - MF->verify(this, "After splitting live range around region"); -} - -MCRegister RAGreedy::tryRegionSplit(const LiveInterval &VirtReg, - AllocationOrder &Order, - SmallVectorImpl &NewVRegs) { - if (!TRI->shouldRegionSplitForVirtReg(*MF, VirtReg)) - return MCRegister::NoRegister; - unsigned NumCands = 0; - BlockFrequency SpillCost = calcSpillCost(); - BlockFrequency BestCost; - - // Check if we can split this live range around a compact region. - bool HasCompact = calcCompactRegion(GlobalCand.front()); - if (HasCompact) { - // Yes, keep GlobalCand[0] as the compact region candidate. - NumCands = 1; - BestCost = BlockFrequency::getMaxFrequency(); - } else { - // No benefit from the compact region, our fallback will be per-block - // splitting. Make sure we find a solution that is cheaper than spilling. - BestCost = SpillCost; - LLVM_DEBUG(dbgs() << "Cost of isolating all blocks = "; - MBFI->printBlockFreq(dbgs(), BestCost) << '\n'); - } - - unsigned BestCand = calculateRegionSplitCost(VirtReg, Order, BestCost, - NumCands, false /*IgnoreCSR*/); - - // No solutions found, fall back to single block splitting. - if (!HasCompact && BestCand == NoCand) - return MCRegister::NoRegister; - - return doRegionSplit(VirtReg, BestCand, HasCompact, NewVRegs); -} - -unsigned RAGreedy::calculateRegionSplitCost(const LiveInterval &VirtReg, - AllocationOrder &Order, - BlockFrequency &BestCost, - unsigned &NumCands, - bool IgnoreCSR) { - unsigned BestCand = NoCand; - for (MCPhysReg PhysReg : Order) { - assert(PhysReg); - if (IgnoreCSR && EvictAdvisor->isUnusedCalleeSavedReg(PhysReg)) - continue; - - // Discard bad candidates before we run out of interference cache cursors. - // This will only affect register classes with a lot of registers (>32). - if (NumCands == IntfCache.getMaxCursors()) { - unsigned WorstCount = ~0u; - unsigned Worst = 0; - for (unsigned CandIndex = 0; CandIndex != NumCands; ++CandIndex) { - if (CandIndex == BestCand || !GlobalCand[CandIndex].PhysReg) - continue; - unsigned Count = GlobalCand[CandIndex].LiveBundles.count(); - if (Count < WorstCount) { - Worst = CandIndex; - WorstCount = Count; - } - } - --NumCands; - GlobalCand[Worst] = GlobalCand[NumCands]; - if (BestCand == NumCands) - BestCand = Worst; - } - - if (GlobalCand.size() <= NumCands) - GlobalCand.resize(NumCands+1); - GlobalSplitCandidate &Cand = GlobalCand[NumCands]; - Cand.reset(IntfCache, PhysReg); - - SpillPlacer->prepare(Cand.LiveBundles); - BlockFrequency Cost; - if (!addSplitConstraints(Cand.Intf, Cost)) { - LLVM_DEBUG(dbgs() << printReg(PhysReg, TRI) << "\tno positive bundles\n"); - continue; - } - LLVM_DEBUG(dbgs() << printReg(PhysReg, TRI) << "\tstatic = "; - MBFI->printBlockFreq(dbgs(), Cost)); - if (Cost >= BestCost) { - LLVM_DEBUG({ - if (BestCand == NoCand) - dbgs() << " worse than no bundles\n"; - else - dbgs() << " worse than " - << printReg(GlobalCand[BestCand].PhysReg, TRI) << '\n'; - }); - continue; - } - if (!growRegion(Cand)) { - LLVM_DEBUG(dbgs() << ", cannot spill all interferences.\n"); - continue; - } - - SpillPlacer->finish(); - - // No live bundles, defer to splitSingleBlocks(). - if (!Cand.LiveBundles.any()) { - LLVM_DEBUG(dbgs() << " no bundles.\n"); - continue; - } - - Cost += calcGlobalSplitCost(Cand, Order); - LLVM_DEBUG({ - dbgs() << ", total = "; - MBFI->printBlockFreq(dbgs(), Cost) << " with bundles"; - for (int I : Cand.LiveBundles.set_bits()) - dbgs() << " EB#" << I; - dbgs() << ".\n"; - }); - if (Cost < BestCost) { - BestCand = NumCands; - BestCost = Cost; - } - ++NumCands; - } - - return BestCand; -} - -unsigned RAGreedy::doRegionSplit(const LiveInterval &VirtReg, unsigned BestCand, - bool HasCompact, - SmallVectorImpl &NewVRegs) { - SmallVector UsedCands; - // Prepare split editor. - LiveRangeEdit LREdit(&VirtReg, NewVRegs, *MF, *LIS, VRM, this, &DeadRemats); - SE->reset(LREdit, SplitSpillMode); - - // Assign all edge bundles to the preferred candidate, or NoCand. - BundleCand.assign(Bundles->getNumBundles(), NoCand); - - // Assign bundles for the best candidate region. - if (BestCand != NoCand) { - GlobalSplitCandidate &Cand = GlobalCand[BestCand]; - if (unsigned B = Cand.getBundles(BundleCand, BestCand)) { - UsedCands.push_back(BestCand); - Cand.IntvIdx = SE->openIntv(); - LLVM_DEBUG(dbgs() << "Split for " << printReg(Cand.PhysReg, TRI) << " in " - << B << " bundles, intv " << Cand.IntvIdx << ".\n"); - (void)B; - } - } - - // Assign bundles for the compact region. - if (HasCompact) { - GlobalSplitCandidate &Cand = GlobalCand.front(); - assert(!Cand.PhysReg && "Compact region has no physreg"); - if (unsigned B = Cand.getBundles(BundleCand, 0)) { - UsedCands.push_back(0); - Cand.IntvIdx = SE->openIntv(); - LLVM_DEBUG(dbgs() << "Split for compact region in " << B - << " bundles, intv " << Cand.IntvIdx << ".\n"); - (void)B; - } - } - - splitAroundRegion(LREdit, UsedCands); - return 0; -} - -//===----------------------------------------------------------------------===// -// Per-Block Splitting -//===----------------------------------------------------------------------===// - -/// tryBlockSplit - Split a global live range around every block with uses. This -/// creates a lot of local live ranges, that will be split by tryLocalSplit if -/// they don't allocate. -unsigned RAGreedy::tryBlockSplit(const LiveInterval &VirtReg, - AllocationOrder &Order, - SmallVectorImpl &NewVRegs) { - assert(&SA->getParent() == &VirtReg && "Live range wasn't analyzed"); - Register Reg = VirtReg.reg(); - bool SingleInstrs = RegClassInfo.isProperSubClass(MRI->getRegClass(Reg)); - LiveRangeEdit LREdit(&VirtReg, NewVRegs, *MF, *LIS, VRM, this, &DeadRemats); - SE->reset(LREdit, SplitSpillMode); - ArrayRef UseBlocks = SA->getUseBlocks(); - for (const SplitAnalysis::BlockInfo &BI : UseBlocks) { - if (SA->shouldSplitSingleBlock(BI, SingleInstrs)) - SE->splitSingleBlock(BI); - } - // No blocks were split. - if (LREdit.empty()) - return 0; - - // We did split for some blocks. - SmallVector IntvMap; - SE->finish(&IntvMap); - - // Tell LiveDebugVariables about the new ranges. - DebugVars->splitRegister(Reg, LREdit.regs(), *LIS); - - // Sort out the new intervals created by splitting. The remainder interval - // goes straight to spilling, the new local ranges get to stay RS_New. - for (unsigned I = 0, E = LREdit.size(); I != E; ++I) { - const LiveInterval &LI = LIS->getInterval(LREdit.get(I)); - if (ExtraInfo->getOrInitStage(LI.reg()) == RS_New && IntvMap[I] == 0) - ExtraInfo->setStage(LI, RS_Spill); - } - - if (VerifyEnabled) - MF->verify(this, "After splitting live range around basic blocks"); - return 0; -} - -//===----------------------------------------------------------------------===// -// Per-Instruction Splitting -//===----------------------------------------------------------------------===// - -/// Get the number of allocatable registers that match the constraints of \p Reg -/// on \p MI and that are also in \p SuperRC. -static unsigned getNumAllocatableRegsForConstraints( - const MachineInstr *MI, Register Reg, const TargetRegisterClass *SuperRC, - const TargetInstrInfo *TII, const TargetRegisterInfo *TRI, - const RegisterClassInfo &RCI) { - assert(SuperRC && "Invalid register class"); - - const TargetRegisterClass *ConstrainedRC = - MI->getRegClassConstraintEffectForVReg(Reg, SuperRC, TII, TRI, - /* ExploreBundle */ true); - if (!ConstrainedRC) - return 0; - return RCI.getNumAllocatableRegs(ConstrainedRC); -} - -/// tryInstructionSplit - Split a live range around individual instructions. -/// This is normally not worthwhile since the spiller is doing essentially the -/// same thing. However, when the live range is in a constrained register -/// class, it may help to insert copies such that parts of the live range can -/// be moved to a larger register class. -/// -/// This is similar to spilling to a larger register class. -unsigned RAGreedy::tryInstructionSplit(const LiveInterval &VirtReg, - AllocationOrder &Order, - SmallVectorImpl &NewVRegs) { - const TargetRegisterClass *CurRC = MRI->getRegClass(VirtReg.reg()); - // There is no point to this if there are no larger sub-classes. - if (!RegClassInfo.isProperSubClass(CurRC)) - return 0; - - // Always enable split spill mode, since we're effectively spilling to a - // register. - LiveRangeEdit LREdit(&VirtReg, NewVRegs, *MF, *LIS, VRM, this, &DeadRemats); - SE->reset(LREdit, SplitEditor::SM_Size); - - ArrayRef Uses = SA->getUseSlots(); - if (Uses.size() <= 1) - return 0; - - LLVM_DEBUG(dbgs() << "Split around " << Uses.size() - << " individual instrs.\n"); - - const TargetRegisterClass *SuperRC = - TRI->getLargestLegalSuperClass(CurRC, *MF); - unsigned SuperRCNumAllocatableRegs = - RegClassInfo.getNumAllocatableRegs(SuperRC); - // Split around every non-copy instruction if this split will relax - // the constraints on the virtual register. - // Otherwise, splitting just inserts uncoalescable copies that do not help - // the allocation. - for (const SlotIndex Use : Uses) { - if (const MachineInstr *MI = Indexes->getInstructionFromIndex(Use)) - if (MI->isFullCopy() || - SuperRCNumAllocatableRegs == - getNumAllocatableRegsForConstraints(MI, VirtReg.reg(), SuperRC, - TII, TRI, RegClassInfo)) { - LLVM_DEBUG(dbgs() << " skip:\t" << Use << '\t' << *MI); - continue; - } - SE->openIntv(); - SlotIndex SegStart = SE->enterIntvBefore(Use); - SlotIndex SegStop = SE->leaveIntvAfter(Use); - SE->useIntv(SegStart, SegStop); - } - - if (LREdit.empty()) { - LLVM_DEBUG(dbgs() << "All uses were copies.\n"); - return 0; - } - - SmallVector IntvMap; - SE->finish(&IntvMap); - DebugVars->splitRegister(VirtReg.reg(), LREdit.regs(), *LIS); - // Assign all new registers to RS_Spill. This was the last chance. - ExtraInfo->setStage(LREdit.begin(), LREdit.end(), RS_Spill); - return 0; -} - -//===----------------------------------------------------------------------===// -// Local Splitting -//===----------------------------------------------------------------------===// - -/// calcGapWeights - Compute the maximum spill weight that needs to be evicted -/// in order to use PhysReg between two entries in SA->UseSlots. -/// -/// GapWeight[I] represents the gap between UseSlots[I] and UseSlots[I + 1]. -/// -void RAGreedy::calcGapWeights(MCRegister PhysReg, - SmallVectorImpl &GapWeight) { - assert(SA->getUseBlocks().size() == 1 && "Not a local interval"); - const SplitAnalysis::BlockInfo &BI = SA->getUseBlocks().front(); - ArrayRef Uses = SA->getUseSlots(); - const unsigned NumGaps = Uses.size()-1; - - // Start and end points for the interference check. - SlotIndex StartIdx = - BI.LiveIn ? BI.FirstInstr.getBaseIndex() : BI.FirstInstr; - SlotIndex StopIdx = - BI.LiveOut ? BI.LastInstr.getBoundaryIndex() : BI.LastInstr; - - GapWeight.assign(NumGaps, 0.0f); - - // Add interference from each overlapping register. - for (MCRegUnitIterator Units(PhysReg, TRI); Units.isValid(); ++Units) { - if (!Matrix->query(const_cast(SA->getParent()), *Units) - .checkInterference()) - continue; - - // We know that VirtReg is a continuous interval from FirstInstr to - // LastInstr, so we don't need InterferenceQuery. - // - // Interference that overlaps an instruction is counted in both gaps - // surrounding the instruction. The exception is interference before - // StartIdx and after StopIdx. - // - LiveIntervalUnion::SegmentIter IntI = - Matrix->getLiveUnions()[*Units] .find(StartIdx); - for (unsigned Gap = 0; IntI.valid() && IntI.start() < StopIdx; ++IntI) { - // Skip the gaps before IntI. - while (Uses[Gap+1].getBoundaryIndex() < IntI.start()) - if (++Gap == NumGaps) - break; - if (Gap == NumGaps) - break; - - // Update the gaps covered by IntI. - const float weight = IntI.value()->weight(); - for (; Gap != NumGaps; ++Gap) { - GapWeight[Gap] = std::max(GapWeight[Gap], weight); - if (Uses[Gap+1].getBaseIndex() >= IntI.stop()) - break; - } - if (Gap == NumGaps) - break; - } - } - - // Add fixed interference. - for (MCRegUnitIterator Units(PhysReg, TRI); Units.isValid(); ++Units) { - const LiveRange &LR = LIS->getRegUnit(*Units); - LiveRange::const_iterator I = LR.find(StartIdx); - LiveRange::const_iterator E = LR.end(); - - // Same loop as above. Mark any overlapped gaps as HUGE_VALF. - for (unsigned Gap = 0; I != E && I->start < StopIdx; ++I) { - while (Uses[Gap+1].getBoundaryIndex() < I->start) - if (++Gap == NumGaps) - break; - if (Gap == NumGaps) - break; - - for (; Gap != NumGaps; ++Gap) { - GapWeight[Gap] = huge_valf; - if (Uses[Gap+1].getBaseIndex() >= I->end) - break; - } - if (Gap == NumGaps) - break; - } - } -} - -/// tryLocalSplit - Try to split VirtReg into smaller intervals inside its only -/// basic block. -/// -unsigned RAGreedy::tryLocalSplit(const LiveInterval &VirtReg, - AllocationOrder &Order, - SmallVectorImpl &NewVRegs) { - // TODO: the function currently only handles a single UseBlock; it should be - // possible to generalize. - if (SA->getUseBlocks().size() != 1) - return 0; - - const SplitAnalysis::BlockInfo &BI = SA->getUseBlocks().front(); - - // Note that it is possible to have an interval that is live-in or live-out - // while only covering a single block - A phi-def can use undef values from - // predecessors, and the block could be a single-block loop. - // We don't bother doing anything clever about such a case, we simply assume - // that the interval is continuous from FirstInstr to LastInstr. We should - // make sure that we don't do anything illegal to such an interval, though. - - ArrayRef Uses = SA->getUseSlots(); - if (Uses.size() <= 2) - return 0; - const unsigned NumGaps = Uses.size()-1; - - LLVM_DEBUG({ - dbgs() << "tryLocalSplit: "; - for (const auto &Use : Uses) - dbgs() << ' ' << Use; - dbgs() << '\n'; - }); - - // If VirtReg is live across any register mask operands, compute a list of - // gaps with register masks. - SmallVector RegMaskGaps; - if (Matrix->checkRegMaskInterference(VirtReg)) { - // Get regmask slots for the whole block. - ArrayRef RMS = LIS->getRegMaskSlotsInBlock(BI.MBB->getNumber()); - LLVM_DEBUG(dbgs() << RMS.size() << " regmasks in block:"); - // Constrain to VirtReg's live range. - unsigned RI = - llvm::lower_bound(RMS, Uses.front().getRegSlot()) - RMS.begin(); - unsigned RE = RMS.size(); - for (unsigned I = 0; I != NumGaps && RI != RE; ++I) { - // Look for Uses[I] <= RMS <= Uses[I + 1]. - assert(!SlotIndex::isEarlierInstr(RMS[RI], Uses[I])); - if (SlotIndex::isEarlierInstr(Uses[I + 1], RMS[RI])) - continue; - // Skip a regmask on the same instruction as the last use. It doesn't - // overlap the live range. - if (SlotIndex::isSameInstr(Uses[I + 1], RMS[RI]) && I + 1 == NumGaps) - break; - LLVM_DEBUG(dbgs() << ' ' << RMS[RI] << ':' << Uses[I] << '-' - << Uses[I + 1]); - RegMaskGaps.push_back(I); - // Advance ri to the next gap. A regmask on one of the uses counts in - // both gaps. - while (RI != RE && SlotIndex::isEarlierInstr(RMS[RI], Uses[I + 1])) - ++RI; - } - LLVM_DEBUG(dbgs() << '\n'); - } - - // Since we allow local split results to be split again, there is a risk of - // creating infinite loops. It is tempting to require that the new live - // ranges have less instructions than the original. That would guarantee - // convergence, but it is too strict. A live range with 3 instructions can be - // split 2+3 (including the COPY), and we want to allow that. - // - // Instead we use these rules: - // - // 1. Allow any split for ranges with getStage() < RS_Split2. (Except for the - // noop split, of course). - // 2. Require progress be made for ranges with getStage() == RS_Split2. All - // the new ranges must have fewer instructions than before the split. - // 3. New ranges with the same number of instructions are marked RS_Split2, - // smaller ranges are marked RS_New. - // - // These rules allow a 3 -> 2+3 split once, which we need. They also prevent - // excessive splitting and infinite loops. - // - bool ProgressRequired = ExtraInfo->getStage(VirtReg) >= RS_Split2; - - // Best split candidate. - unsigned BestBefore = NumGaps; - unsigned BestAfter = 0; - float BestDiff = 0; - - const float blockFreq = - SpillPlacer->getBlockFrequency(BI.MBB->getNumber()).getFrequency() * - (1.0f / MBFI->getEntryFreq()); - SmallVector GapWeight; - - for (MCPhysReg PhysReg : Order) { - assert(PhysReg); - // Keep track of the largest spill weight that would need to be evicted in - // order to make use of PhysReg between UseSlots[I] and UseSlots[I + 1]. - calcGapWeights(PhysReg, GapWeight); - - // Remove any gaps with regmask clobbers. - if (Matrix->checkRegMaskInterference(VirtReg, PhysReg)) - for (unsigned I = 0, E = RegMaskGaps.size(); I != E; ++I) - GapWeight[RegMaskGaps[I]] = huge_valf; - - // Try to find the best sequence of gaps to close. - // The new spill weight must be larger than any gap interference. - - // We will split before Uses[SplitBefore] and after Uses[SplitAfter]. - unsigned SplitBefore = 0, SplitAfter = 1; - - // MaxGap should always be max(GapWeight[SplitBefore..SplitAfter-1]). - // It is the spill weight that needs to be evicted. - float MaxGap = GapWeight[0]; - - while (true) { - // Live before/after split? - const bool LiveBefore = SplitBefore != 0 || BI.LiveIn; - const bool LiveAfter = SplitAfter != NumGaps || BI.LiveOut; - - LLVM_DEBUG(dbgs() << printReg(PhysReg, TRI) << ' ' << Uses[SplitBefore] - << '-' << Uses[SplitAfter] << " I=" << MaxGap); - - // Stop before the interval gets so big we wouldn't be making progress. - if (!LiveBefore && !LiveAfter) { - LLVM_DEBUG(dbgs() << " all\n"); - break; - } - // Should the interval be extended or shrunk? - bool Shrink = true; - - // How many gaps would the new range have? - unsigned NewGaps = LiveBefore + SplitAfter - SplitBefore + LiveAfter; - - // Legally, without causing looping? - bool Legal = !ProgressRequired || NewGaps < NumGaps; - - if (Legal && MaxGap < huge_valf) { - // Estimate the new spill weight. Each instruction reads or writes the - // register. Conservatively assume there are no read-modify-write - // instructions. - // - // Try to guess the size of the new interval. - const float EstWeight = normalizeSpillWeight( - blockFreq * (NewGaps + 1), - Uses[SplitBefore].distance(Uses[SplitAfter]) + - (LiveBefore + LiveAfter) * SlotIndex::InstrDist, - 1); - // Would this split be possible to allocate? - // Never allocate all gaps, we wouldn't be making progress. - LLVM_DEBUG(dbgs() << " w=" << EstWeight); - if (EstWeight * Hysteresis >= MaxGap) { - Shrink = false; - float Diff = EstWeight - MaxGap; - if (Diff > BestDiff) { - LLVM_DEBUG(dbgs() << " (best)"); - BestDiff = Hysteresis * Diff; - BestBefore = SplitBefore; - BestAfter = SplitAfter; - } - } - } - - // Try to shrink. - if (Shrink) { - if (++SplitBefore < SplitAfter) { - LLVM_DEBUG(dbgs() << " shrink\n"); - // Recompute the max when necessary. - if (GapWeight[SplitBefore - 1] >= MaxGap) { - MaxGap = GapWeight[SplitBefore]; - for (unsigned I = SplitBefore + 1; I != SplitAfter; ++I) - MaxGap = std::max(MaxGap, GapWeight[I]); - } - continue; - } - MaxGap = 0; - } - - // Try to extend the interval. - if (SplitAfter >= NumGaps) { - LLVM_DEBUG(dbgs() << " end\n"); - break; - } - - LLVM_DEBUG(dbgs() << " extend\n"); - MaxGap = std::max(MaxGap, GapWeight[SplitAfter++]); - } - } - - // Didn't find any candidates? - if (BestBefore == NumGaps) - return 0; - - LLVM_DEBUG(dbgs() << "Best local split range: " << Uses[BestBefore] << '-' - << Uses[BestAfter] << ", " << BestDiff << ", " - << (BestAfter - BestBefore + 1) << " instrs\n"); - - LiveRangeEdit LREdit(&VirtReg, NewVRegs, *MF, *LIS, VRM, this, &DeadRemats); - SE->reset(LREdit); - - SE->openIntv(); - SlotIndex SegStart = SE->enterIntvBefore(Uses[BestBefore]); - SlotIndex SegStop = SE->leaveIntvAfter(Uses[BestAfter]); - SE->useIntv(SegStart, SegStop); - SmallVector IntvMap; - SE->finish(&IntvMap); - DebugVars->splitRegister(VirtReg.reg(), LREdit.regs(), *LIS); - // If the new range has the same number of instructions as before, mark it as - // RS_Split2 so the next split will be forced to make progress. Otherwise, - // leave the new intervals as RS_New so they can compete. - bool LiveBefore = BestBefore != 0 || BI.LiveIn; - bool LiveAfter = BestAfter != NumGaps || BI.LiveOut; - unsigned NewGaps = LiveBefore + BestAfter - BestBefore + LiveAfter; - if (NewGaps >= NumGaps) { - LLVM_DEBUG(dbgs() << "Tagging non-progress ranges:"); - assert(!ProgressRequired && "Didn't make progress when it was required."); - for (unsigned I = 0, E = IntvMap.size(); I != E; ++I) - if (IntvMap[I] == 1) { - ExtraInfo->setStage(LIS->getInterval(LREdit.get(I)), RS_Split2); - LLVM_DEBUG(dbgs() << ' ' << printReg(LREdit.get(I))); - } - LLVM_DEBUG(dbgs() << '\n'); - } - ++NumLocalSplits; - - return 0; -} - -//===----------------------------------------------------------------------===// -// Live Range Splitting -//===----------------------------------------------------------------------===// - -/// trySplit - Try to split VirtReg or one of its interferences, making it -/// assignable. -/// @return Physreg when VirtReg may be assigned and/or new NewVRegs. -unsigned RAGreedy::trySplit(const LiveInterval &VirtReg, AllocationOrder &Order, - SmallVectorImpl &NewVRegs, - const SmallVirtRegSet &FixedRegisters) { - // Ranges must be Split2 or less. - if (ExtraInfo->getStage(VirtReg) >= RS_Spill) - return 0; - - // Local intervals are handled separately. - if (LIS->intervalIsInOneMBB(VirtReg)) { - NamedRegionTimer T("local_split", "Local Splitting", TimerGroupName, - TimerGroupDescription, TimePassesIsEnabled); - SA->analyze(&VirtReg); - Register PhysReg = tryLocalSplit(VirtReg, Order, NewVRegs); - if (PhysReg || !NewVRegs.empty()) - return PhysReg; - return tryInstructionSplit(VirtReg, Order, NewVRegs); - } - - NamedRegionTimer T("global_split", "Global Splitting", TimerGroupName, - TimerGroupDescription, TimePassesIsEnabled); - - SA->analyze(&VirtReg); - - // First try to split around a region spanning multiple blocks. RS_Split2 - // ranges already made dubious progress with region splitting, so they go - // straight to single block splitting. - if (ExtraInfo->getStage(VirtReg) < RS_Split2) { - MCRegister PhysReg = tryRegionSplit(VirtReg, Order, NewVRegs); - if (PhysReg || !NewVRegs.empty()) - return PhysReg; - } - - // Then isolate blocks. - return tryBlockSplit(VirtReg, Order, NewVRegs); -} - -//===----------------------------------------------------------------------===// -// Last Chance Recoloring -//===----------------------------------------------------------------------===// - -/// Return true if \p reg has any tied def operand. -static bool hasTiedDef(MachineRegisterInfo *MRI, unsigned reg) { - for (const MachineOperand &MO : MRI->def_operands(reg)) - if (MO.isTied()) - return true; - - return false; -} - -/// Return true if the existing assignment of \p Intf overlaps, but is not the -/// same, as \p PhysReg. -static bool assignedRegPartiallyOverlaps(const TargetRegisterInfo &TRI, - const VirtRegMap &VRM, - MCRegister PhysReg, - const LiveInterval &Intf) { - MCRegister AssignedReg = VRM.getPhys(Intf.reg()); - if (PhysReg == AssignedReg) - return false; - return TRI.regsOverlap(PhysReg, AssignedReg); -} - -/// mayRecolorAllInterferences - Check if the virtual registers that -/// interfere with \p VirtReg on \p PhysReg (or one of its aliases) may be -/// recolored to free \p PhysReg. -/// When true is returned, \p RecoloringCandidates has been augmented with all -/// the live intervals that need to be recolored in order to free \p PhysReg -/// for \p VirtReg. -/// \p FixedRegisters contains all the virtual registers that cannot be -/// recolored. -bool RAGreedy::mayRecolorAllInterferences( - MCRegister PhysReg, const LiveInterval &VirtReg, - SmallLISet &RecoloringCandidates, const SmallVirtRegSet &FixedRegisters) { - const TargetRegisterClass *CurRC = MRI->getRegClass(VirtReg.reg()); - - for (MCRegUnitIterator Units(PhysReg, TRI); Units.isValid(); ++Units) { - LiveIntervalUnion::Query &Q = Matrix->query(VirtReg, *Units); - // If there is LastChanceRecoloringMaxInterference or more interferences, - // chances are one would not be recolorable. - if (Q.interferingVRegs(LastChanceRecoloringMaxInterference).size() >= - LastChanceRecoloringMaxInterference && - !ExhaustiveSearch) { - LLVM_DEBUG(dbgs() << "Early abort: too many interferences.\n"); - CutOffInfo |= CO_Interf; - return false; - } - for (const LiveInterval *Intf : reverse(Q.interferingVRegs())) { - // If Intf is done and sits on the same register class as VirtReg, it - // would not be recolorable as it is in the same state as - // VirtReg. However there are at least two exceptions. - // - // If VirtReg has tied defs and Intf doesn't, then - // there is still a point in examining if it can be recolorable. - // - // Additionally, if the register class has overlapping tuple members, it - // may still be recolorable using a different tuple. This is more likely - // if the existing assignment aliases with the candidate. - // - if (((ExtraInfo->getStage(*Intf) == RS_Done && - MRI->getRegClass(Intf->reg()) == CurRC && - !assignedRegPartiallyOverlaps(*TRI, *VRM, PhysReg, *Intf)) && - !(hasTiedDef(MRI, VirtReg.reg()) && - !hasTiedDef(MRI, Intf->reg()))) || - FixedRegisters.count(Intf->reg())) { - LLVM_DEBUG( - dbgs() << "Early abort: the interference is not recolorable.\n"); - return false; - } - RecoloringCandidates.insert(Intf); - } - } - return true; -} - -/// tryLastChanceRecoloring - Try to assign a color to \p VirtReg by recoloring -/// its interferences. -/// Last chance recoloring chooses a color for \p VirtReg and recolors every -/// virtual register that was using it. The recoloring process may recursively -/// use the last chance recoloring. Therefore, when a virtual register has been -/// assigned a color by this mechanism, it is marked as Fixed, i.e., it cannot -/// be last-chance-recolored again during this recoloring "session". -/// E.g., -/// Let -/// vA can use {R1, R2 } -/// vB can use { R2, R3} -/// vC can use {R1 } -/// Where vA, vB, and vC cannot be split anymore (they are reloads for -/// instance) and they all interfere. -/// -/// vA is assigned R1 -/// vB is assigned R2 -/// vC tries to evict vA but vA is already done. -/// Regular register allocation fails. -/// -/// Last chance recoloring kicks in: -/// vC does as if vA was evicted => vC uses R1. -/// vC is marked as fixed. -/// vA needs to find a color. -/// None are available. -/// vA cannot evict vC: vC is a fixed virtual register now. -/// vA does as if vB was evicted => vA uses R2. -/// vB needs to find a color. -/// R3 is available. -/// Recoloring => vC = R1, vA = R2, vB = R3 -/// -/// \p Order defines the preferred allocation order for \p VirtReg. -/// \p NewRegs will contain any new virtual register that have been created -/// (split, spill) during the process and that must be assigned. -/// \p FixedRegisters contains all the virtual registers that cannot be -/// recolored. -/// -/// \p RecolorStack tracks the original assignments of successfully recolored -/// registers. -/// -/// \p Depth gives the current depth of the last chance recoloring. -/// \return a physical register that can be used for VirtReg or ~0u if none -/// exists. -unsigned RAGreedy::tryLastChanceRecoloring(const LiveInterval &VirtReg, - AllocationOrder &Order, - SmallVectorImpl &NewVRegs, - SmallVirtRegSet &FixedRegisters, - RecoloringStack &RecolorStack, - unsigned Depth) { - if (!TRI->shouldUseLastChanceRecoloringForVirtReg(*MF, VirtReg)) - return ~0u; - - LLVM_DEBUG(dbgs() << "Try last chance recoloring for " << VirtReg << '\n'); - - const ssize_t EntryStackSize = RecolorStack.size(); - - // Ranges must be Done. - assert((ExtraInfo->getStage(VirtReg) >= RS_Done || !VirtReg.isSpillable()) && - "Last chance recoloring should really be last chance"); - // Set the max depth to LastChanceRecoloringMaxDepth. - // We may want to reconsider that if we end up with a too large search space - // for target with hundreds of registers. - // Indeed, in that case we may want to cut the search space earlier. - if (Depth >= LastChanceRecoloringMaxDepth && !ExhaustiveSearch) { - LLVM_DEBUG(dbgs() << "Abort because max depth has been reached.\n"); - CutOffInfo |= CO_Depth; - return ~0u; - } - - // Set of Live intervals that will need to be recolored. - SmallLISet RecoloringCandidates; - - // Mark VirtReg as fixed, i.e., it will not be recolored pass this point in - // this recoloring "session". - assert(!FixedRegisters.count(VirtReg.reg())); - FixedRegisters.insert(VirtReg.reg()); - SmallVector CurrentNewVRegs; - - for (MCRegister PhysReg : Order) { - assert(PhysReg.isValid()); - LLVM_DEBUG(dbgs() << "Try to assign: " << VirtReg << " to " - << printReg(PhysReg, TRI) << '\n'); - RecoloringCandidates.clear(); - CurrentNewVRegs.clear(); - - // It is only possible to recolor virtual register interference. - if (Matrix->checkInterference(VirtReg, PhysReg) > - LiveRegMatrix::IK_VirtReg) { - LLVM_DEBUG( - dbgs() << "Some interferences are not with virtual registers.\n"); - - continue; - } - - // Early give up on this PhysReg if it is obvious we cannot recolor all - // the interferences. - if (!mayRecolorAllInterferences(PhysReg, VirtReg, RecoloringCandidates, - FixedRegisters)) { - LLVM_DEBUG(dbgs() << "Some interferences cannot be recolored.\n"); - continue; - } - - // RecoloringCandidates contains all the virtual registers that interfere - // with VirtReg on PhysReg (or one of its aliases). Enqueue them for - // recoloring and perform the actual recoloring. - PQueue RecoloringQueue; - for (const LiveInterval *RC : RecoloringCandidates) { - Register ItVirtReg = RC->reg(); - enqueue(RecoloringQueue, RC); - assert(VRM->hasPhys(ItVirtReg) && - "Interferences are supposed to be with allocated variables"); - - // Record the current allocation. - RecolorStack.push_back(std::make_pair(RC, VRM->getPhys(ItVirtReg))); - - // unset the related struct. - Matrix->unassign(*RC); - } - - // Do as if VirtReg was assigned to PhysReg so that the underlying - // recoloring has the right information about the interferes and - // available colors. - Matrix->assign(VirtReg, PhysReg); - - // Save the current recoloring state. - // If we cannot recolor all the interferences, we will have to start again - // at this point for the next physical register. - SmallVirtRegSet SaveFixedRegisters(FixedRegisters); - if (tryRecoloringCandidates(RecoloringQueue, CurrentNewVRegs, - FixedRegisters, RecolorStack, Depth)) { - // Push the queued vregs into the main queue. - for (Register NewVReg : CurrentNewVRegs) - NewVRegs.push_back(NewVReg); - // Do not mess up with the global assignment process. - // I.e., VirtReg must be unassigned. - Matrix->unassign(VirtReg); - return PhysReg; - } - - LLVM_DEBUG(dbgs() << "Fail to assign: " << VirtReg << " to " - << printReg(PhysReg, TRI) << '\n'); - - // The recoloring attempt failed, undo the changes. - FixedRegisters = SaveFixedRegisters; - Matrix->unassign(VirtReg); - - // For a newly created vreg which is also in RecoloringCandidates, - // don't add it to NewVRegs because its physical register will be restored - // below. Other vregs in CurrentNewVRegs are created by calling - // selectOrSplit and should be added into NewVRegs. - for (Register &R : CurrentNewVRegs) { - if (RecoloringCandidates.count(&LIS->getInterval(R))) - continue; - NewVRegs.push_back(R); - } - - // Roll back our unsuccessful recoloring. Also roll back any successful - // recolorings in any recursive recoloring attempts, since it's possible - // they would have introduced conflicts with assignments we will be - // restoring further up the stack. Perform all unassignments prior to - // reassigning, since sub-recolorings may have conflicted with the registers - // we are going to restore to their original assignments. - for (ssize_t I = RecolorStack.size() - 1; I >= EntryStackSize; --I) { - const LiveInterval *LI; - MCRegister PhysReg; - std::tie(LI, PhysReg) = RecolorStack[I]; - - if (VRM->hasPhys(LI->reg())) - Matrix->unassign(*LI); - } - - for (size_t I = EntryStackSize; I != RecolorStack.size(); ++I) { - const LiveInterval *LI; - MCRegister PhysReg; - std::tie(LI, PhysReg) = RecolorStack[I]; - Matrix->assign(*LI, PhysReg); - } - - // Pop the stack of recoloring attempts. - RecolorStack.resize(EntryStackSize); - } - - // Last chance recoloring did not worked either, give up. - return ~0u; -} - -/// tryRecoloringCandidates - Try to assign a new color to every register -/// in \RecoloringQueue. -/// \p NewRegs will contain any new virtual register created during the -/// recoloring process. -/// \p FixedRegisters[in/out] contains all the registers that have been -/// recolored. -/// \return true if all virtual registers in RecoloringQueue were successfully -/// recolored, false otherwise. -bool RAGreedy::tryRecoloringCandidates(PQueue &RecoloringQueue, - SmallVectorImpl &NewVRegs, - SmallVirtRegSet &FixedRegisters, - RecoloringStack &RecolorStack, - unsigned Depth) { - while (!RecoloringQueue.empty()) { - const LiveInterval *LI = dequeue(RecoloringQueue); - LLVM_DEBUG(dbgs() << "Try to recolor: " << *LI << '\n'); - MCRegister PhysReg = selectOrSplitImpl(*LI, NewVRegs, FixedRegisters, - RecolorStack, Depth + 1); - // When splitting happens, the live-range may actually be empty. - // In that case, this is okay to continue the recoloring even - // if we did not find an alternative color for it. Indeed, - // there will not be anything to color for LI in the end. - if (PhysReg == ~0u || (!PhysReg && !LI->empty())) - return false; - - if (!PhysReg) { - assert(LI->empty() && "Only empty live-range do not require a register"); - LLVM_DEBUG(dbgs() << "Recoloring of " << *LI - << " succeeded. Empty LI.\n"); - continue; - } - LLVM_DEBUG(dbgs() << "Recoloring of " << *LI - << " succeeded with: " << printReg(PhysReg, TRI) << '\n'); - - Matrix->assign(*LI, PhysReg); - FixedRegisters.insert(LI->reg()); - } - return true; -} - -//===----------------------------------------------------------------------===// -// Main Entry Point -//===----------------------------------------------------------------------===// - -MCRegister RAGreedy::selectOrSplit(const LiveInterval &VirtReg, - SmallVectorImpl &NewVRegs) { - CutOffInfo = CO_None; - LLVMContext &Ctx = MF->getFunction().getContext(); - SmallVirtRegSet FixedRegisters; - RecoloringStack RecolorStack; - MCRegister Reg = - selectOrSplitImpl(VirtReg, NewVRegs, FixedRegisters, RecolorStack); - if (Reg == ~0U && (CutOffInfo != CO_None)) { - uint8_t CutOffEncountered = CutOffInfo & (CO_Depth | CO_Interf); - if (CutOffEncountered == CO_Depth) - Ctx.emitError("register allocation failed: maximum depth for recoloring " - "reached. Use -fexhaustive-register-search to skip " - "cutoffs"); - else if (CutOffEncountered == CO_Interf) - Ctx.emitError("register allocation failed: maximum interference for " - "recoloring reached. Use -fexhaustive-register-search " - "to skip cutoffs"); - else if (CutOffEncountered == (CO_Depth | CO_Interf)) - Ctx.emitError("register allocation failed: maximum interference and " - "depth for recoloring reached. Use " - "-fexhaustive-register-search to skip cutoffs"); - } - return Reg; -} - -/// Using a CSR for the first time has a cost because it causes push|pop -/// to be added to prologue|epilogue. Splitting a cold section of the live -/// range can have lower cost than using the CSR for the first time; -/// Spilling a live range in the cold path can have lower cost than using -/// the CSR for the first time. Returns the physical register if we decide -/// to use the CSR; otherwise return 0. -MCRegister RAGreedy::tryAssignCSRFirstTime( - const LiveInterval &VirtReg, AllocationOrder &Order, MCRegister PhysReg, - uint8_t &CostPerUseLimit, SmallVectorImpl &NewVRegs) { - if (ExtraInfo->getStage(VirtReg) == RS_Spill && VirtReg.isSpillable()) { - // We choose spill over using the CSR for the first time if the spill cost - // is lower than CSRCost. - SA->analyze(&VirtReg); - if (calcSpillCost() >= CSRCost) - return PhysReg; - - // We are going to spill, set CostPerUseLimit to 1 to make sure that - // we will not use a callee-saved register in tryEvict. - CostPerUseLimit = 1; - return 0; - } - if (ExtraInfo->getStage(VirtReg) < RS_Split) { - // We choose pre-splitting over using the CSR for the first time if - // the cost of splitting is lower than CSRCost. - SA->analyze(&VirtReg); - unsigned NumCands = 0; - BlockFrequency BestCost = CSRCost; // Don't modify CSRCost. - unsigned BestCand = calculateRegionSplitCost(VirtReg, Order, BestCost, - NumCands, true /*IgnoreCSR*/); - if (BestCand == NoCand) - // Use the CSR if we can't find a region split below CSRCost. - return PhysReg; - - // Perform the actual pre-splitting. - doRegionSplit(VirtReg, BestCand, false/*HasCompact*/, NewVRegs); - return 0; - } - return PhysReg; -} - -void RAGreedy::aboutToRemoveInterval(const LiveInterval &LI) { - // Do not keep invalid information around. - SetOfBrokenHints.remove(&LI); -} - -void RAGreedy::initializeCSRCost() { - // We use the larger one out of the command-line option and the value report - // by TRI. - CSRCost = BlockFrequency( - std::max((unsigned)CSRFirstTimeCost, TRI->getCSRFirstUseCost())); - if (!CSRCost.getFrequency()) - return; - - // Raw cost is relative to Entry == 2^14; scale it appropriately. - uint64_t ActualEntry = MBFI->getEntryFreq(); - if (!ActualEntry) { - CSRCost = 0; - return; - } - uint64_t FixedEntry = 1 << 14; - if (ActualEntry < FixedEntry) - CSRCost *= BranchProbability(ActualEntry, FixedEntry); - else if (ActualEntry <= UINT32_MAX) - // Invert the fraction and divide. - CSRCost /= BranchProbability(FixedEntry, ActualEntry); - else - // Can't use BranchProbability in general, since it takes 32-bit numbers. - CSRCost = CSRCost.getFrequency() * (ActualEntry / FixedEntry); -} - -/// Collect the hint info for \p Reg. -/// The results are stored into \p Out. -/// \p Out is not cleared before being populated. -void RAGreedy::collectHintInfo(Register Reg, HintsInfo &Out) { - for (const MachineInstr &Instr : MRI->reg_nodbg_instructions(Reg)) { - if (!Instr.isFullCopy()) - continue; - // Look for the other end of the copy. - Register OtherReg = Instr.getOperand(0).getReg(); - if (OtherReg == Reg) { - OtherReg = Instr.getOperand(1).getReg(); - if (OtherReg == Reg) - continue; - } - // Get the current assignment. - MCRegister OtherPhysReg = - OtherReg.isPhysical() ? OtherReg.asMCReg() : VRM->getPhys(OtherReg); - // Push the collected information. - Out.push_back(HintInfo(MBFI->getBlockFreq(Instr.getParent()), OtherReg, - OtherPhysReg)); - } -} - -/// Using the given \p List, compute the cost of the broken hints if -/// \p PhysReg was used. -/// \return The cost of \p List for \p PhysReg. -BlockFrequency RAGreedy::getBrokenHintFreq(const HintsInfo &List, - MCRegister PhysReg) { - BlockFrequency Cost = 0; - for (const HintInfo &Info : List) { - if (Info.PhysReg != PhysReg) - Cost += Info.Freq; - } - return Cost; -} - -/// Using the register assigned to \p VirtReg, try to recolor -/// all the live ranges that are copy-related with \p VirtReg. -/// The recoloring is then propagated to all the live-ranges that have -/// been recolored and so on, until no more copies can be coalesced or -/// it is not profitable. -/// For a given live range, profitability is determined by the sum of the -/// frequencies of the non-identity copies it would introduce with the old -/// and new register. -void RAGreedy::tryHintRecoloring(const LiveInterval &VirtReg) { - // We have a broken hint, check if it is possible to fix it by - // reusing PhysReg for the copy-related live-ranges. Indeed, we evicted - // some register and PhysReg may be available for the other live-ranges. - SmallSet Visited; - SmallVector RecoloringCandidates; - HintsInfo Info; - Register Reg = VirtReg.reg(); - MCRegister PhysReg = VRM->getPhys(Reg); - // Start the recoloring algorithm from the input live-interval, then - // it will propagate to the ones that are copy-related with it. - Visited.insert(Reg); - RecoloringCandidates.push_back(Reg); - - LLVM_DEBUG(dbgs() << "Trying to reconcile hints for: " << printReg(Reg, TRI) - << '(' << printReg(PhysReg, TRI) << ")\n"); - - do { - Reg = RecoloringCandidates.pop_back_val(); - - // We cannot recolor physical register. - if (Register::isPhysicalRegister(Reg)) - continue; - - // This may be a skipped class - if (!VRM->hasPhys(Reg)) { - assert(!ShouldAllocateClass(*TRI, *MRI->getRegClass(Reg)) && - "We have an unallocated variable which should have been handled"); - continue; - } - - // Get the live interval mapped with this virtual register to be able - // to check for the interference with the new color. - LiveInterval &LI = LIS->getInterval(Reg); - MCRegister CurrPhys = VRM->getPhys(Reg); - // Check that the new color matches the register class constraints and - // that it is free for this live range. - if (CurrPhys != PhysReg && (!MRI->getRegClass(Reg)->contains(PhysReg) || - Matrix->checkInterference(LI, PhysReg))) - continue; - - LLVM_DEBUG(dbgs() << printReg(Reg, TRI) << '(' << printReg(CurrPhys, TRI) - << ") is recolorable.\n"); - - // Gather the hint info. - Info.clear(); - collectHintInfo(Reg, Info); - // Check if recoloring the live-range will increase the cost of the - // non-identity copies. - if (CurrPhys != PhysReg) { - LLVM_DEBUG(dbgs() << "Checking profitability:\n"); - BlockFrequency OldCopiesCost = getBrokenHintFreq(Info, CurrPhys); - BlockFrequency NewCopiesCost = getBrokenHintFreq(Info, PhysReg); - LLVM_DEBUG(dbgs() << "Old Cost: " << OldCopiesCost.getFrequency() - << "\nNew Cost: " << NewCopiesCost.getFrequency() - << '\n'); - if (OldCopiesCost < NewCopiesCost) { - LLVM_DEBUG(dbgs() << "=> Not profitable.\n"); - continue; - } - // At this point, the cost is either cheaper or equal. If it is - // equal, we consider this is profitable because it may expose - // more recoloring opportunities. - LLVM_DEBUG(dbgs() << "=> Profitable.\n"); - // Recolor the live-range. - Matrix->unassign(LI); - Matrix->assign(LI, PhysReg); - } - // Push all copy-related live-ranges to keep reconciling the broken - // hints. - for (const HintInfo &HI : Info) { - if (Visited.insert(HI.Reg).second) - RecoloringCandidates.push_back(HI.Reg); - } - } while (!RecoloringCandidates.empty()); -} - -/// Try to recolor broken hints. -/// Broken hints may be repaired by recoloring when an evicted variable -/// freed up a register for a larger live-range. -/// Consider the following example: -/// BB1: -/// a = -/// b = -/// BB2: -/// ... -/// = b -/// = a -/// Let us assume b gets split: -/// BB1: -/// a = -/// b = -/// BB2: -/// c = b -/// ... -/// d = c -/// = d -/// = a -/// Because of how the allocation work, b, c, and d may be assigned different -/// colors. Now, if a gets evicted later: -/// BB1: -/// a = -/// st a, SpillSlot -/// b = -/// BB2: -/// c = b -/// ... -/// d = c -/// = d -/// e = ld SpillSlot -/// = e -/// This is likely that we can assign the same register for b, c, and d, -/// getting rid of 2 copies. -void RAGreedy::tryHintsRecoloring() { - for (const LiveInterval *LI : SetOfBrokenHints) { - assert(Register::isVirtualRegister(LI->reg()) && - "Recoloring is possible only for virtual registers"); - // Some dead defs may be around (e.g., because of debug uses). - // Ignore those. - if (!VRM->hasPhys(LI->reg())) - continue; - tryHintRecoloring(*LI); - } -} - -MCRegister RAGreedy::selectOrSplitImpl(const LiveInterval &VirtReg, - SmallVectorImpl &NewVRegs, - SmallVirtRegSet &FixedRegisters, - RecoloringStack &RecolorStack, - unsigned Depth) { - uint8_t CostPerUseLimit = uint8_t(~0u); - // First try assigning a free register. - auto Order = - AllocationOrder::create(VirtReg.reg(), *VRM, RegClassInfo, Matrix); - if (MCRegister PhysReg = - tryAssign(VirtReg, Order, NewVRegs, FixedRegisters)) { - // If VirtReg got an assignment, the eviction info is no longer relevant. - LastEvicted.clearEvicteeInfo(VirtReg.reg()); - // When NewVRegs is not empty, we may have made decisions such as evicting - // a virtual register, go with the earlier decisions and use the physical - // register. - if (CSRCost.getFrequency() && - EvictAdvisor->isUnusedCalleeSavedReg(PhysReg) && NewVRegs.empty()) { - MCRegister CSRReg = tryAssignCSRFirstTime(VirtReg, Order, PhysReg, - CostPerUseLimit, NewVRegs); - if (CSRReg || !NewVRegs.empty()) - // Return now if we decide to use a CSR or create new vregs due to - // pre-splitting. - return CSRReg; - } else - return PhysReg; - } - - LiveRangeStage Stage = ExtraInfo->getStage(VirtReg); - LLVM_DEBUG(dbgs() << StageName[Stage] << " Cascade " - << ExtraInfo->getCascade(VirtReg.reg()) << '\n'); - - // Try to evict a less worthy live range, but only for ranges from the primary - // queue. The RS_Split ranges already failed to do this, and they should not - // get a second chance until they have been split. - if (Stage != RS_Split) - if (Register PhysReg = - tryEvict(VirtReg, Order, NewVRegs, CostPerUseLimit, - FixedRegisters)) { - Register Hint = MRI->getSimpleHint(VirtReg.reg()); - // If VirtReg has a hint and that hint is broken record this - // virtual register as a recoloring candidate for broken hint. - // Indeed, since we evicted a variable in its neighborhood it is - // likely we can at least partially recolor some of the - // copy-related live-ranges. - if (Hint && Hint != PhysReg) - SetOfBrokenHints.insert(&VirtReg); - // If VirtReg eviction someone, the eviction info for it as an evictee is - // no longer relevant. - LastEvicted.clearEvicteeInfo(VirtReg.reg()); - return PhysReg; - } - - assert((NewVRegs.empty() || Depth) && "Cannot append to existing NewVRegs"); - - // The first time we see a live range, don't try to split or spill. - // Wait until the second time, when all smaller ranges have been allocated. - // This gives a better picture of the interference to split around. - if (Stage < RS_Split) { - ExtraInfo->setStage(VirtReg, RS_Split); - LLVM_DEBUG(dbgs() << "wait for second round\n"); - NewVRegs.push_back(VirtReg.reg()); - return 0; - } - - if (Stage < RS_Spill) { - // Try splitting VirtReg or interferences. - unsigned NewVRegSizeBefore = NewVRegs.size(); - Register PhysReg = trySplit(VirtReg, Order, NewVRegs, FixedRegisters); - if (PhysReg || (NewVRegs.size() - NewVRegSizeBefore)) { - // If VirtReg got split, the eviction info is no longer relevant. - LastEvicted.clearEvicteeInfo(VirtReg.reg()); - return PhysReg; - } - } - - // If we couldn't allocate a register from spilling, there is probably some - // invalid inline assembly. The base class will report it. - if (Stage >= RS_Done || !VirtReg.isSpillable()) { - return tryLastChanceRecoloring(VirtReg, Order, NewVRegs, FixedRegisters, - RecolorStack, Depth); - } - - // Finally spill VirtReg itself. - if ((EnableDeferredSpilling || - TRI->shouldUseDeferredSpillingForVirtReg(*MF, VirtReg)) && - ExtraInfo->getStage(VirtReg) < RS_Memory) { - // TODO: This is experimental and in particular, we do not model - // the live range splitting done by spilling correctly. - // We would need a deep integration with the spiller to do the - // right thing here. Anyway, that is still good for early testing. - ExtraInfo->setStage(VirtReg, RS_Memory); - LLVM_DEBUG(dbgs() << "Do as if this register is in memory\n"); - NewVRegs.push_back(VirtReg.reg()); - } else { - NamedRegionTimer T("spill", "Spiller", TimerGroupName, - TimerGroupDescription, TimePassesIsEnabled); - LiveRangeEdit LRE(&VirtReg, NewVRegs, *MF, *LIS, VRM, this, &DeadRemats); - spiller().spill(LRE); - ExtraInfo->setStage(NewVRegs.begin(), NewVRegs.end(), RS_Done); - - // Tell LiveDebugVariables about the new ranges. Ranges not being covered by - // the new regs are kept in LDV (still mapping to the old register), until - // we rewrite spilled locations in LDV at a later stage. - DebugVars->splitRegister(VirtReg.reg(), LRE.regs(), *LIS); - - if (VerifyEnabled) - MF->verify(this, "After spilling"); - } - - // The live virtual register requesting allocation was spilled, so tell - // the caller not to allocate anything during this round. - return 0; -} - -void RAGreedy::RAGreedyStats::report(MachineOptimizationRemarkMissed &R) { - using namespace ore; - if (Spills) { - R << NV("NumSpills", Spills) << " spills "; - R << NV("TotalSpillsCost", SpillsCost) << " total spills cost "; - } - if (FoldedSpills) { - R << NV("NumFoldedSpills", FoldedSpills) << " folded spills "; - R << NV("TotalFoldedSpillsCost", FoldedSpillsCost) - << " total folded spills cost "; - } - if (Reloads) { - R << NV("NumReloads", Reloads) << " reloads "; - R << NV("TotalReloadsCost", ReloadsCost) << " total reloads cost "; - } - if (FoldedReloads) { - R << NV("NumFoldedReloads", FoldedReloads) << " folded reloads "; - R << NV("TotalFoldedReloadsCost", FoldedReloadsCost) - << " total folded reloads cost "; - } - if (ZeroCostFoldedReloads) - R << NV("NumZeroCostFoldedReloads", ZeroCostFoldedReloads) - << " zero cost folded reloads "; - if (Copies) { - R << NV("NumVRCopies", Copies) << " virtual registers copies "; - R << NV("TotalCopiesCost", CopiesCost) << " total copies cost "; - } -} - -RAGreedy::RAGreedyStats RAGreedy::computeStats(MachineBasicBlock &MBB) { - RAGreedyStats Stats; - const MachineFrameInfo &MFI = MF->getFrameInfo(); - int FI; - - auto isSpillSlotAccess = [&MFI](const MachineMemOperand *A) { - return MFI.isSpillSlotObjectIndex(cast( - A->getPseudoValue())->getFrameIndex()); - }; - auto isPatchpointInstr = [](const MachineInstr &MI) { - return MI.getOpcode() == TargetOpcode::PATCHPOINT || - MI.getOpcode() == TargetOpcode::STACKMAP || - MI.getOpcode() == TargetOpcode::STATEPOINT; - }; - for (MachineInstr &MI : MBB) { - if (MI.isCopy()) { - MachineOperand &Dest = MI.getOperand(0); - MachineOperand &Src = MI.getOperand(1); - if (Dest.isReg() && Src.isReg() && Dest.getReg().isVirtual() && - Src.getReg().isVirtual()) - ++Stats.Copies; - continue; - } - - SmallVector Accesses; - if (TII->isLoadFromStackSlot(MI, FI) && MFI.isSpillSlotObjectIndex(FI)) { - ++Stats.Reloads; - continue; - } - if (TII->isStoreToStackSlot(MI, FI) && MFI.isSpillSlotObjectIndex(FI)) { - ++Stats.Spills; - continue; - } - if (TII->hasLoadFromStackSlot(MI, Accesses) && - llvm::any_of(Accesses, isSpillSlotAccess)) { - if (!isPatchpointInstr(MI)) { - Stats.FoldedReloads += Accesses.size(); - continue; - } - // For statepoint there may be folded and zero cost folded stack reloads. - std::pair NonZeroCostRange = - TII->getPatchpointUnfoldableRange(MI); - SmallSet FoldedReloads; - SmallSet ZeroCostFoldedReloads; - for (unsigned Idx = 0, E = MI.getNumOperands(); Idx < E; ++Idx) { - MachineOperand &MO = MI.getOperand(Idx); - if (!MO.isFI() || !MFI.isSpillSlotObjectIndex(MO.getIndex())) - continue; - if (Idx >= NonZeroCostRange.first && Idx < NonZeroCostRange.second) - FoldedReloads.insert(MO.getIndex()); - else - ZeroCostFoldedReloads.insert(MO.getIndex()); - } - // If stack slot is used in folded reload it is not zero cost then. - for (unsigned Slot : FoldedReloads) - ZeroCostFoldedReloads.erase(Slot); - Stats.FoldedReloads += FoldedReloads.size(); - Stats.ZeroCostFoldedReloads += ZeroCostFoldedReloads.size(); - continue; - } - Accesses.clear(); - if (TII->hasStoreToStackSlot(MI, Accesses) && - llvm::any_of(Accesses, isSpillSlotAccess)) { - Stats.FoldedSpills += Accesses.size(); - } - } - // Set cost of collected statistic by multiplication to relative frequency of - // this basic block. - float RelFreq = MBFI->getBlockFreqRelativeToEntryBlock(&MBB); - Stats.ReloadsCost = RelFreq * Stats.Reloads; - Stats.FoldedReloadsCost = RelFreq * Stats.FoldedReloads; - Stats.SpillsCost = RelFreq * Stats.Spills; - Stats.FoldedSpillsCost = RelFreq * Stats.FoldedSpills; - Stats.CopiesCost = RelFreq * Stats.Copies; - return Stats; -} - -RAGreedy::RAGreedyStats RAGreedy::reportStats(MachineLoop *L) { - RAGreedyStats Stats; - - // Sum up the spill and reloads in subloops. - for (MachineLoop *SubLoop : *L) - Stats.add(reportStats(SubLoop)); - - for (MachineBasicBlock *MBB : L->getBlocks()) - // Handle blocks that were not included in subloops. - if (Loops->getLoopFor(MBB) == L) - Stats.add(computeStats(*MBB)); - - if (!Stats.isEmpty()) { - using namespace ore; - - ORE->emit([&]() { - MachineOptimizationRemarkMissed R(DEBUG_TYPE, "LoopSpillReloadCopies", - L->getStartLoc(), L->getHeader()); - Stats.report(R); - R << "generated in loop"; - return R; - }); - } - return Stats; -} - -void RAGreedy::reportStats() { - if (!ORE->allowExtraAnalysis(DEBUG_TYPE)) - return; - RAGreedyStats Stats; - for (MachineLoop *L : *Loops) - Stats.add(reportStats(L)); - // Process non-loop blocks. - for (MachineBasicBlock &MBB : *MF) - if (!Loops->getLoopFor(&MBB)) - Stats.add(computeStats(MBB)); - if (!Stats.isEmpty()) { - using namespace ore; - - ORE->emit([&]() { - DebugLoc Loc; - if (auto *SP = MF->getFunction().getSubprogram()) - Loc = DILocation::get(SP->getContext(), SP->getLine(), 1, SP); - MachineOptimizationRemarkMissed R(DEBUG_TYPE, "SpillReloadCopies", Loc, - &MF->front()); - Stats.report(R); - R << "generated in function"; - return R; - }); - } -} - -bool RAGreedy::runOnMachineFunction(MachineFunction &mf) { - LLVM_DEBUG(dbgs() << "********** GREEDY REGISTER ALLOCATION **********\n" - << "********** Function: " << mf.getName() << '\n'); - - MF = &mf; - TII = MF->getSubtarget().getInstrInfo(); - - if (VerifyEnabled) - MF->verify(this, "Before greedy register allocator"); - - RegAllocBase::init(getAnalysis(), - getAnalysis(), - getAnalysis()); - Indexes = &getAnalysis(); - MBFI = &getAnalysis(); - DomTree = &getAnalysis(); - ORE = &getAnalysis().getORE(); - Loops = &getAnalysis(); - Bundles = &getAnalysis(); - SpillPlacer = &getAnalysis(); - DebugVars = &getAnalysis(); - AA = &getAnalysis().getAAResults(); - - initializeCSRCost(); - - RegCosts = TRI->getRegisterCosts(*MF); - RegClassPriorityTrumpsGlobalness = - GreedyRegClassPriorityTrumpsGlobalness.getNumOccurrences() - ? GreedyRegClassPriorityTrumpsGlobalness - : TRI->regClassPriorityTrumpsGlobalness(*MF); - - ExtraInfo.emplace(); - EvictAdvisor = - getAnalysis().getAdvisor(*MF, *this); - - VRAI = std::make_unique(*MF, *LIS, *VRM, *Loops, *MBFI); - SpillerInstance.reset(createInlineSpiller(*this, *MF, *VRM, *VRAI)); - - VRAI->calculateSpillWeightsAndHints(); - - LLVM_DEBUG(LIS->dump()); - - SA.reset(new SplitAnalysis(*VRM, *LIS, *Loops)); - SE.reset(new SplitEditor(*SA, *AA, *LIS, *VRM, *DomTree, *MBFI, *VRAI)); - - IntfCache.init(MF, Matrix->getLiveUnions(), Indexes, LIS, TRI); - GlobalCand.resize(32); // This will grow as needed. - SetOfBrokenHints.clear(); - LastEvicted.clear(); - - allocatePhysRegs(); - tryHintsRecoloring(); - - if (VerifyEnabled) - MF->verify(this, "Before post optimization"); - postOptimization(); - reportStats(); - - releaseMemory(); - return true; -} diff --git a/llvm/lib/CodeGen/RegAllocPriorityAdvisor.h b/llvm/lib/CodeGen/RegAllocPriorityAdvisor.h new file mode 100644 --- /dev/null +++ b/llvm/lib/CodeGen/RegAllocPriorityAdvisor.h @@ -0,0 +1,67 @@ +#include "llvm/Pass.h" + +namespace llvm { + +class MachineFunction; +class RAGreedy; + +class RegAllocPriorityAdvisor { +public: + RegAllocPriorityAdvisor(const RegAllocPriorityAdvisor &) = delete; + RegAllocPriorityAdvisor(RegAllocPriorityAdvisor &&) = delete; + virtual ~RegAllocPriorityAdvisor() = default; + + virtual float tryFindPriority(unsigned Prio, unsigned Size, + LiveRangeStage Stage, float Weight) const = 0; + virtual void logReward(float reward) = 0; + +protected: + RegAllocPriorityAdvisor(const MachineFunction &MF, const RAGreedy &RA); + + const MachineFunction &MF; + const RAGreedy &RA; +}; + +class RegAllocPriorityAdvisorAnalysis : public ImmutablePass { +public: + enum class AdvisorMode : int { Default, Release, Development }; + + RegAllocPriorityAdvisorAnalysis(AdvisorMode Mode) + : ImmutablePass(ID), Mode(Mode){}; + static char ID; + + /// Get an advisor for the given context (i.e. machine function, etc) + virtual std::unique_ptr + getAdvisor(const MachineFunction &MF, const RAGreedy &RA) = 0; + AdvisorMode getAdvisorMode() const { return Mode; } + +protected: + // This analysis preserves everything, and subclasses may have additional + // requirements. + void getAnalysisUsage(AnalysisUsage &AU) const override { + AU.setPreservesAll(); + } + +private: + StringRef getPassName() const override; + const AdvisorMode Mode; +}; + +template <> Pass *callDefaultCtor(); + +RegAllocPriorityAdvisorAnalysis *createReleaseModePriorityAdvisor(); + +RegAllocPriorityAdvisorAnalysis *createDevelopmentModePriorityAdvisor(); + +class DefaultPriorityAdvisor : public RegAllocPriorityAdvisor { +public: + DefaultPriorityAdvisor(const MachineFunction &MF, const RAGreedy &RA) + : RegAllocPriorityAdvisor(MF, RA) {} + + void logReward(float reward) override{}; + +private: + float tryFindPriority(unsigned Prio, unsigned Size, LiveRangeStage Stage, + float Weight) const override; +}; +} // namespace llvm \ No newline at end of file diff --git a/llvm/lib/CodeGen/RegAllocPriorityAdvisor.cpp b/llvm/lib/CodeGen/RegAllocPriorityAdvisor.cpp new file mode 100644 --- /dev/null +++ b/llvm/lib/CodeGen/RegAllocPriorityAdvisor.cpp @@ -0,0 +1,94 @@ +#include "RegAllocGreedy.h" +#include "llvm/CodeGen/MachineFunction.h" +#include "llvm/InitializePasses.h" +#include "llvm/Pass.h" + +using namespace llvm; + +static cl::opt Mode( + "regalloc-prio-enable-advisor", cl::Hidden, + cl::init(RegAllocPriorityAdvisorAnalysis::AdvisorMode::Default), + cl::desc("Enable regalloc advisor mode"), + cl::values( + clEnumValN(RegAllocPriorityAdvisorAnalysis::AdvisorMode::Default, + "default", "Default"), + clEnumValN(RegAllocPriorityAdvisorAnalysis::AdvisorMode::Release, + "release", "precompiled"), + clEnumValN(RegAllocPriorityAdvisorAnalysis::AdvisorMode::Development, + "development", "for training"))); + +char RegAllocPriorityAdvisorAnalysis::ID = 0; +INITIALIZE_PASS(RegAllocPriorityAdvisorAnalysis, "regalloc-prio", + "RegAllocPriorityAdvisorAnalysis", false, true) + +namespace { +class DefaultPriorityAdvisorAnalysis final + : public RegAllocPriorityAdvisorAnalysis { +public: + DefaultPriorityAdvisorAnalysis(bool NotAsRequested) + : RegAllocPriorityAdvisorAnalysis(AdvisorMode::Default), + NotAsRequested(NotAsRequested) {} + + // support for isa<> and dyn_cast. + static bool classof(const RegAllocPriorityAdvisorAnalysis *R) { + return R->getAdvisorMode() == AdvisorMode::Default; + } + +private: + std::unique_ptr + getAdvisor(const MachineFunction &MF, const RAGreedy &RA) override { + return std::make_unique(MF, RA); + } + bool doInitialization(Module &M) override { + if (NotAsRequested) + M.getContext().emitError("Requested regalloc priority advisor analysis " + "could be created. Using default"); + return RegAllocPriorityAdvisorAnalysis::doInitialization(M); + } + const bool NotAsRequested; +}; +} // namespace + +template <> Pass *llvm::callDefaultCtor() { + Pass *Ret = nullptr; + switch (Mode) { + case RegAllocPriorityAdvisorAnalysis::AdvisorMode::Default: + Ret = new DefaultPriorityAdvisorAnalysis(/*NotAsRequested*/ false); + break; + case RegAllocPriorityAdvisorAnalysis::AdvisorMode::Development: +#if defined(LLVM_HAVE_TF_API) + Ret = createDevelopmentModePriorityAdvisor(); +#endif + break; + case RegAllocPriorityAdvisorAnalysis::AdvisorMode::Release: +#if defined(LLVM_HAVE_TF_AOT) + Ret = createReleaseModePriorityAdvisor(); +#endif + break; + } + if (Ret) + return Ret; + return new DefaultPriorityAdvisorAnalysis(/*NotAsRequested*/ true); +} + +StringRef RegAllocPriorityAdvisorAnalysis::getPassName() const { + switch (getAdvisorMode()) { + case AdvisorMode::Default: + return "Default Regalloc Priority Advisor"; + case AdvisorMode::Release: + return "Release mode Regalloc Priority Advisor"; + case AdvisorMode::Development: + return "Development mode Regalloc Priority Advisor"; + } + llvm_unreachable("Unknown advisor kind"); +} + +RegAllocPriorityAdvisor::RegAllocPriorityAdvisor(const MachineFunction &MF, + const RAGreedy &RA) + : MF(MF), RA(RA) {} + +float DefaultPriorityAdvisor::tryFindPriority(unsigned Prio, unsigned Size, + LiveRangeStage Stage, + float Weight) const { + return static_cast(Prio); +} \ No newline at end of file diff --git a/llvm/test/CodeGen/MLRegalloc/Inputs/reference-prio-log-noml.txt b/llvm/test/CodeGen/MLRegalloc/Inputs/reference-prio-log-noml.txt new file mode 100644 --- /dev/null +++ b/llvm/test/CodeGen/MLRegalloc/Inputs/reference-prio-log-noml.txt @@ -0,0 +1,361 @@ +fields { + key: "SyFgets" + value { + string_value: "feature_lists { + feature_list { + key: \"li_size\" value { + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } } } + feature_list { + key: \"priority\" value { + feature { float_list { value: 2.68435814e+09 } } + feature { float_list { value: 2.68435814e+09 } } + feature { float_list { value: 2.68435866e+09 } } + feature { float_list { value: 2.68435789e+09 } } + feature { float_list { value: 3.75810074e+09 } } + feature { float_list { value: 3.7580969e+09 } } + feature { float_list { value: 2.14748518e+09 } } + feature { float_list { value: 2.14748493e+09 } } + feature { float_list { value: 2.68435814e+09 } } + feature { float_list { value: 2.14748493e+09 } } + feature { float_list { value: 2.14748493e+09 } } + feature { float_list { value: 2.68435456e+09 } } + feature { float_list { value: 2.68435456e+09 } } + feature { float_list { value: 2.68435763e+09 } } + feature { float_list { value: 2.68435456e+09 } } + feature { float_list { value: 2.68435456e+09 } } + feature { float_list { value: 2.14748467e+09 } } + feature { float_list { value: 2.68435456e+09 } } + feature { float_list { value: 2.68435456e+09 } } + feature { float_list { value: 2.68435456e+09 } } + feature { float_list { value: 2.68435763e+09 } } + feature { float_list { value: 2.68435763e+09 } } + feature { float_list { value: 2.1474839e+09 } } + feature { float_list { value: 2.1474839e+09 } } + feature { float_list { value: 2.68435456e+09 } } + feature { float_list { value: 2.1474839e+09 } } + feature { float_list { value: 3.22122547e+09 } } + feature { float_list { value: 2.68435456e+09 } } + feature { float_list { value: 2.14748365e+09 } } + feature { float_list { value: 2.14748493e+09 } } + feature { float_list { value: 2.14748493e+09 } } + feature { float_list { value: 2.68435814e+09 } } + feature { float_list { value: 2.68435456e+09 } } + feature { float_list { value: 2.68435584e+09 } } + feature { float_list { value: 2.68435482e+09 } } + feature { float_list { value: 2.68435482e+09 } } + feature { float_list { value: 2.68435763e+09 } } + feature { float_list { value: 2.68435584e+09 } } + feature { float_list { value: 2.14748365e+09 } } + feature { float_list { value: 3.75810074e+09 } } + feature { float_list { value: 2.68435814e+09 } } + feature { float_list { value: 3584 } } + feature { float_list { value: 2.68435814e+09 } } + feature { float_list { value: 3550 } } + feature { float_list { value: 2958 } } + feature { float_list { value: 2974 } } + feature { float_list { value: 2.68435814e+09 } } + feature { float_list { value: 3534 } } + feature { float_list { value: 2.68435866e+09 } } + feature { float_list { value: 3998 } } + feature { float_list { value: 3.75809946e+09 } } + feature { float_list { value: 3.75810074e+09 } } + feature { float_list { value: 4336 } } + feature { float_list { value: 2.68435814e+09 } } + feature { float_list { value: 3.7580969e+09 } } + feature { float_list { value: 2.68435482e+09 } } + feature { float_list { value: 2.68435482e+09 } } + feature { float_list { value: 2.68435814e+09 } } + feature { float_list { value: 2.68435482e+09 } } + feature { float_list { value: 2.68435456e+09 } } + feature { float_list { value: 2.68435456e+09 } } + feature { float_list { value: 2.68435814e+09 } } + feature { float_list { value: 2.14748493e+09 } } + feature { float_list { value: 2.68435456e+09 } } + feature { float_list { value: 2.68435814e+09 } } + feature { float_list { value: 2.14748493e+09 } } + feature { float_list { value: 2.68435456e+09 } } + feature { float_list { value: 2.68435456e+09 } } + feature { float_list { value: 2.68435456e+09 } } } } + feature_list { + key: \"reward\" value { + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 36.9006081 } } } } + feature_list { + key: \"stage\" value { + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } + feature { int64_list { value: 0 } } } } + feature_list { + key: \"weight\" value { + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } + feature { float_list { value: 0 } } } } } " + } +} diff --git a/llvm/test/CodeGen/MLRegalloc/dev-mode-prio-logging.ll b/llvm/test/CodeGen/MLRegalloc/dev-mode-prio-logging.ll new file mode 100644 --- /dev/null +++ b/llvm/test/CodeGen/MLRegalloc/dev-mode-prio-logging.ll @@ -0,0 +1,30 @@ +; REQUIRES: have_tf_api +; REQUIRES: x86_64-linux +; +; Check that we log correctly, both with a learned policy, and the default policy +; +; RUN: llc -mtriple=x86_64-linux-unknown -regalloc=greedy -regalloc-prio-enable-advisor=development \ +; RUN: -regalloc-prio-training-log=%t1 -tfutils-text-log < %S/Inputs/input.ll +; RUN: sed -i 's/ \+/ /g' %t1 +; RUN: sed -i 's/\\n key:/\n key:/g' %t1 +; RUN: sed -i 's/\\n feature/\n feature/g' %t1 +; RUN: sed -i 's/\\n/ /g' %t1 +; RUN: FileCheck --input-file %t1 %s --check-prefixes=CHECK,NOML +; RUN: diff %t1 %S/Inputs/reference-prio-log-noml.txt + +; RUN: rm -rf %t && mkdir %t +; RUN: %python %S/../../../lib/Analysis/models/gen-regalloc-priority-test-model.py %t +; RUN: llc -mtriple=x86_64-linux-unknown -regalloc=greedy -regalloc-prio-enable-advisor=development \ +; RUN: -regalloc-prio-training-log=%t2 -tfutils-text-log -regalloc-prio-model=%t < %S/Inputs/input.ll +; RUN: sed -i 's/ \+/ /g' %t2 +; RUN: sed -i 's/\\n key:/\n key:/g' %t2 +; RUN: sed -i 's/\\n feature/\n feature/g' %t2 +; RUN: sed -i 's/\\n/ /g' %t2 +; RUN: FileCheck --input-file %t2 %s --check-prefixes=CHECK,ML + +; CHECK-NOT: nan +; CHECK-LABEL: key: \"priority\" +; NOML-NEXT: feature { float_list { value: 2.68435814e+09 } } +; ML-NEXT: feature { float_list { value: 3551 } } +; CHECK-LABEL: key: \"reward\" +