Relanding #143479 after fixes. Removed `NumberOfFeatures` from the `FeatureIndex` enum as the number of features used depends on whether IR2Vec embeddings are used.
This commit is contained in:
committed by
GitHub
parent
026aae7047
commit
dcf8ec9218
@@ -15,6 +15,7 @@
|
||||
#define LLVM_ANALYSIS_FUNCTIONPROPERTIESANALYSIS_H
|
||||
|
||||
#include "llvm/ADT/DenseSet.h"
|
||||
#include "llvm/Analysis/IR2Vec.h"
|
||||
#include "llvm/IR/Dominators.h"
|
||||
#include "llvm/IR/PassManager.h"
|
||||
#include "llvm/Support/Compiler.h"
|
||||
@@ -32,17 +33,19 @@ class FunctionPropertiesInfo {
|
||||
void updateAggregateStats(const Function &F, const LoopInfo &LI);
|
||||
void reIncludeBB(const BasicBlock &BB);
|
||||
|
||||
ir2vec::Embedding FunctionEmbedding = ir2vec::Embedding(0.0);
|
||||
std::optional<ir2vec::Vocab> IR2VecVocab;
|
||||
|
||||
public:
|
||||
LLVM_ABI static FunctionPropertiesInfo
|
||||
getFunctionPropertiesInfo(const Function &F, const DominatorTree &DT,
|
||||
const LoopInfo &LI);
|
||||
const LoopInfo &LI,
|
||||
const IR2VecVocabResult *VocabResult);
|
||||
|
||||
LLVM_ABI static FunctionPropertiesInfo
|
||||
getFunctionPropertiesInfo(Function &F, FunctionAnalysisManager &FAM);
|
||||
|
||||
bool operator==(const FunctionPropertiesInfo &FPI) const {
|
||||
return std::memcmp(this, &FPI, sizeof(FunctionPropertiesInfo)) == 0;
|
||||
}
|
||||
bool operator==(const FunctionPropertiesInfo &FPI) const;
|
||||
|
||||
bool operator!=(const FunctionPropertiesInfo &FPI) const {
|
||||
return !(*this == FPI);
|
||||
@@ -137,6 +140,19 @@ public:
|
||||
int64_t CallReturnsVectorPointerCount = 0;
|
||||
int64_t CallWithManyArgumentsCount = 0;
|
||||
int64_t CallWithPointerArgumentCount = 0;
|
||||
|
||||
const ir2vec::Embedding &getFunctionEmbedding() const {
|
||||
return FunctionEmbedding;
|
||||
}
|
||||
|
||||
const std::optional<ir2vec::Vocab> &getIR2VecVocab() const {
|
||||
return IR2VecVocab;
|
||||
}
|
||||
|
||||
// Helper intended to be useful for unittests
|
||||
void setFunctionEmbeddingForTest(const ir2vec::Embedding &Embedding) {
|
||||
FunctionEmbedding = Embedding;
|
||||
}
|
||||
};
|
||||
|
||||
// Analysis pass
|
||||
@@ -193,6 +209,7 @@ private:
|
||||
DominatorTree &getUpdatedDominatorTree(FunctionAnalysisManager &FAM) const;
|
||||
|
||||
DenseSet<const BasicBlock *> Successors;
|
||||
DenseSet<const BasicBlock *> CallUsers;
|
||||
|
||||
// Edges we might potentially need to remove from the dominator tree.
|
||||
SmallVector<DominatorTree::UpdateType, 2> DomTreeUpdates;
|
||||
|
||||
@@ -331,6 +331,10 @@ public:
|
||||
};
|
||||
|
||||
Result run(Module &M, ModuleAnalysisManager &MAM) { return Result(M, MAM); }
|
||||
|
||||
private:
|
||||
static bool initializeIR2VecVocabIfRequested(Module &M,
|
||||
ModuleAnalysisManager &MAM);
|
||||
};
|
||||
|
||||
/// Printer pass for the InlineAdvisorAnalysis results.
|
||||
|
||||
@@ -132,6 +132,11 @@ constexpr bool isHeuristicInlineCostFeature(InlineCostFeatureIndex Feature) {
|
||||
"not " \
|
||||
"fully inlined by ElimAvailExtern)")
|
||||
|
||||
// Not all features listed in FeatureIndex are used by the ML model.
|
||||
// Specifically, callee_embedding and caller_embedding are used only when the
|
||||
// usage of IR2Vec embeddings is explicitly enabled. Meaning, the size/number of
|
||||
// features is not static. So, we cannot determine number of features based on
|
||||
// the number of elements in this enum.
|
||||
// clang-format off
|
||||
enum class FeatureIndex : size_t {
|
||||
#define POPULATE_INDICES(DTYPE, SHAPE, NAME, COMMENT) NAME,
|
||||
@@ -142,7 +147,11 @@ enum class FeatureIndex : size_t {
|
||||
INLINE_FEATURE_ITERATOR(POPULATE_INDICES)
|
||||
#undef POPULATE_INDICES
|
||||
|
||||
NumberOfFeatures
|
||||
// IR2Vec embeddings
|
||||
// Dimensions of embeddings are not known in the compile time (until vocab is
|
||||
// read). Hence macros cannot be used here.
|
||||
callee_embedding,
|
||||
caller_embedding
|
||||
};
|
||||
// clang-format on
|
||||
|
||||
@@ -151,10 +160,7 @@ inlineCostFeatureToMlFeature(InlineCostFeatureIndex Feature) {
|
||||
return static_cast<FeatureIndex>(static_cast<size_t>(Feature));
|
||||
}
|
||||
|
||||
constexpr size_t NumberOfFeatures =
|
||||
static_cast<size_t>(FeatureIndex::NumberOfFeatures);
|
||||
|
||||
LLVM_ABI extern const std::vector<TensorSpec> FeatureMap;
|
||||
LLVM_ABI extern std::vector<TensorSpec> FeatureMap;
|
||||
|
||||
LLVM_ABI extern const char *const DecisionName;
|
||||
LLVM_ABI extern const TensorSpec InlineDecisionSpec;
|
||||
|
||||
@@ -82,6 +82,7 @@ private:
|
||||
int64_t NodeCount = 0;
|
||||
int64_t EdgeCount = 0;
|
||||
int64_t EdgesOfLastSeenNodes = 0;
|
||||
const bool UseIR2Vec;
|
||||
|
||||
std::map<const LazyCallGraph::Node *, unsigned> FunctionLevels;
|
||||
const int32_t InitialIRSize = 0;
|
||||
|
||||
@@ -199,6 +199,29 @@ void FunctionPropertiesInfo::updateForBB(const BasicBlock &BB,
|
||||
#undef CHECK_OPERAND
|
||||
}
|
||||
}
|
||||
|
||||
if (IR2VecVocab) {
|
||||
// We instantiate the IR2Vec embedder each time, as having an unique
|
||||
// pointer to the embedder as member of the class would make it
|
||||
// non-copyable. Instantiating the embedder in itself is not costly.
|
||||
auto EmbOrErr = ir2vec::Embedder::create(IR2VecKind::Symbolic,
|
||||
*BB.getParent(), *IR2VecVocab);
|
||||
if (Error Err = EmbOrErr.takeError()) {
|
||||
handleAllErrors(std::move(Err), [&](const ErrorInfoBase &EI) {
|
||||
BB.getContext().emitError("Error creating IR2Vec embeddings: " +
|
||||
EI.message());
|
||||
});
|
||||
return;
|
||||
}
|
||||
auto Embedder = std::move(*EmbOrErr);
|
||||
const auto &BBEmbedding = Embedder->getBBVector(BB);
|
||||
// Subtract BBEmbedding from Function embedding if the direction is -1,
|
||||
// and add it if the direction is +1.
|
||||
if (Direction == -1)
|
||||
FunctionEmbedding -= BBEmbedding;
|
||||
else
|
||||
FunctionEmbedding += BBEmbedding;
|
||||
}
|
||||
}
|
||||
|
||||
void FunctionPropertiesInfo::updateAggregateStats(const Function &F,
|
||||
@@ -220,14 +243,24 @@ void FunctionPropertiesInfo::updateAggregateStats(const Function &F,
|
||||
|
||||
FunctionPropertiesInfo FunctionPropertiesInfo::getFunctionPropertiesInfo(
|
||||
Function &F, FunctionAnalysisManager &FAM) {
|
||||
// We use the cached result of the IR2VecVocabAnalysis run by
|
||||
// InlineAdvisorAnalysis. If the IR2VecVocabAnalysis is not run, we don't
|
||||
// use IR2Vec embeddings.
|
||||
auto VocabResult = FAM.getResult<ModuleAnalysisManagerFunctionProxy>(F)
|
||||
.getCachedResult<IR2VecVocabAnalysis>(*F.getParent());
|
||||
return getFunctionPropertiesInfo(F, FAM.getResult<DominatorTreeAnalysis>(F),
|
||||
FAM.getResult<LoopAnalysis>(F));
|
||||
FAM.getResult<LoopAnalysis>(F), VocabResult);
|
||||
}
|
||||
|
||||
FunctionPropertiesInfo FunctionPropertiesInfo::getFunctionPropertiesInfo(
|
||||
const Function &F, const DominatorTree &DT, const LoopInfo &LI) {
|
||||
const Function &F, const DominatorTree &DT, const LoopInfo &LI,
|
||||
const IR2VecVocabResult *VocabResult) {
|
||||
|
||||
FunctionPropertiesInfo FPI;
|
||||
if (VocabResult && VocabResult->isValid()) {
|
||||
FPI.IR2VecVocab = VocabResult->getVocabulary();
|
||||
FPI.FunctionEmbedding = ir2vec::Embedding(VocabResult->getDimension(), 0.0);
|
||||
}
|
||||
for (const auto &BB : F)
|
||||
if (DT.isReachableFromEntry(&BB))
|
||||
FPI.reIncludeBB(BB);
|
||||
@@ -235,6 +268,66 @@ FunctionPropertiesInfo FunctionPropertiesInfo::getFunctionPropertiesInfo(
|
||||
return FPI;
|
||||
}
|
||||
|
||||
bool FunctionPropertiesInfo::operator==(
|
||||
const FunctionPropertiesInfo &FPI) const {
|
||||
if (BasicBlockCount != FPI.BasicBlockCount ||
|
||||
BlocksReachedFromConditionalInstruction !=
|
||||
FPI.BlocksReachedFromConditionalInstruction ||
|
||||
Uses != FPI.Uses ||
|
||||
DirectCallsToDefinedFunctions != FPI.DirectCallsToDefinedFunctions ||
|
||||
LoadInstCount != FPI.LoadInstCount ||
|
||||
StoreInstCount != FPI.StoreInstCount ||
|
||||
MaxLoopDepth != FPI.MaxLoopDepth ||
|
||||
TopLevelLoopCount != FPI.TopLevelLoopCount ||
|
||||
TotalInstructionCount != FPI.TotalInstructionCount ||
|
||||
BasicBlocksWithSingleSuccessor != FPI.BasicBlocksWithSingleSuccessor ||
|
||||
BasicBlocksWithTwoSuccessors != FPI.BasicBlocksWithTwoSuccessors ||
|
||||
BasicBlocksWithMoreThanTwoSuccessors !=
|
||||
FPI.BasicBlocksWithMoreThanTwoSuccessors ||
|
||||
BasicBlocksWithSinglePredecessor !=
|
||||
FPI.BasicBlocksWithSinglePredecessor ||
|
||||
BasicBlocksWithTwoPredecessors != FPI.BasicBlocksWithTwoPredecessors ||
|
||||
BasicBlocksWithMoreThanTwoPredecessors !=
|
||||
FPI.BasicBlocksWithMoreThanTwoPredecessors ||
|
||||
BigBasicBlocks != FPI.BigBasicBlocks ||
|
||||
MediumBasicBlocks != FPI.MediumBasicBlocks ||
|
||||
SmallBasicBlocks != FPI.SmallBasicBlocks ||
|
||||
CastInstructionCount != FPI.CastInstructionCount ||
|
||||
FloatingPointInstructionCount != FPI.FloatingPointInstructionCount ||
|
||||
IntegerInstructionCount != FPI.IntegerInstructionCount ||
|
||||
ConstantIntOperandCount != FPI.ConstantIntOperandCount ||
|
||||
ConstantFPOperandCount != FPI.ConstantFPOperandCount ||
|
||||
ConstantOperandCount != FPI.ConstantOperandCount ||
|
||||
InstructionOperandCount != FPI.InstructionOperandCount ||
|
||||
BasicBlockOperandCount != FPI.BasicBlockOperandCount ||
|
||||
GlobalValueOperandCount != FPI.GlobalValueOperandCount ||
|
||||
InlineAsmOperandCount != FPI.InlineAsmOperandCount ||
|
||||
ArgumentOperandCount != FPI.ArgumentOperandCount ||
|
||||
UnknownOperandCount != FPI.UnknownOperandCount ||
|
||||
CriticalEdgeCount != FPI.CriticalEdgeCount ||
|
||||
ControlFlowEdgeCount != FPI.ControlFlowEdgeCount ||
|
||||
UnconditionalBranchCount != FPI.UnconditionalBranchCount ||
|
||||
IntrinsicCount != FPI.IntrinsicCount ||
|
||||
DirectCallCount != FPI.DirectCallCount ||
|
||||
IndirectCallCount != FPI.IndirectCallCount ||
|
||||
CallReturnsIntegerCount != FPI.CallReturnsIntegerCount ||
|
||||
CallReturnsFloatCount != FPI.CallReturnsFloatCount ||
|
||||
CallReturnsPointerCount != FPI.CallReturnsPointerCount ||
|
||||
CallReturnsVectorIntCount != FPI.CallReturnsVectorIntCount ||
|
||||
CallReturnsVectorFloatCount != FPI.CallReturnsVectorFloatCount ||
|
||||
CallReturnsVectorPointerCount != FPI.CallReturnsVectorPointerCount ||
|
||||
CallWithManyArgumentsCount != FPI.CallWithManyArgumentsCount ||
|
||||
CallWithPointerArgumentCount != FPI.CallWithPointerArgumentCount) {
|
||||
return false;
|
||||
}
|
||||
// Check the equality of the function embeddings. We don't check the equality
|
||||
// of Vocabulary as it remains the same.
|
||||
if (!FunctionEmbedding.approximatelyEquals(FPI.FunctionEmbedding))
|
||||
return false;
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
void FunctionPropertiesInfo::print(raw_ostream &OS) const {
|
||||
#define PRINT_PROPERTY(PROP_NAME) OS << #PROP_NAME ": " << PROP_NAME << "\n";
|
||||
|
||||
@@ -322,6 +415,16 @@ FunctionPropertiesUpdater::FunctionPropertiesUpdater(
|
||||
// The caller's entry BB may change due to new alloca instructions.
|
||||
LikelyToChangeBBs.insert(&*Caller.begin());
|
||||
|
||||
// The users of the value returned by call instruction can change
|
||||
// leading to the change in embeddings being computed, when used.
|
||||
// We conservatively add the BBs with such uses to LikelyToChangeBBs.
|
||||
for (const auto *User : CB.users())
|
||||
CallUsers.insert(dyn_cast<Instruction>(User)->getParent());
|
||||
// CallSiteBB can be removed from CallUsers if present, it's taken care
|
||||
// separately.
|
||||
CallUsers.erase(&CallSiteBB);
|
||||
LikelyToChangeBBs.insert_range(CallUsers);
|
||||
|
||||
// The successors may become unreachable in the case of `invoke` inlining.
|
||||
// We track successors separately, too, because they form a boundary, together
|
||||
// with the CB BB ('Entry') between which the inlined callee will be pasted.
|
||||
@@ -435,6 +538,9 @@ void FunctionPropertiesUpdater::finish(FunctionAnalysisManager &FAM) const {
|
||||
if (&CallSiteBB != &*Caller.begin())
|
||||
Reinclude.insert(&*Caller.begin());
|
||||
|
||||
// Reinclude the BBs which use the values returned by call instruction
|
||||
Reinclude.insert_range(CallUsers);
|
||||
|
||||
// Distribute the successors to the 2 buckets.
|
||||
for (const auto *Succ : Successors)
|
||||
if (DT.isReachableFromEntry(Succ))
|
||||
@@ -486,6 +592,9 @@ bool FunctionPropertiesUpdater::isUpdateValid(Function &F,
|
||||
return false;
|
||||
DominatorTree DT(F);
|
||||
LoopInfo LI(DT);
|
||||
auto Fresh = FunctionPropertiesInfo::getFunctionPropertiesInfo(F, DT, LI);
|
||||
auto VocabResult = FAM.getResult<ModuleAnalysisManagerFunctionProxy>(F)
|
||||
.getCachedResult<IR2VecVocabAnalysis>(*F.getParent());
|
||||
auto Fresh =
|
||||
FunctionPropertiesInfo::getFunctionPropertiesInfo(F, DT, LI, VocabResult);
|
||||
return FPI == Fresh;
|
||||
}
|
||||
|
||||
@@ -16,6 +16,7 @@
|
||||
#include "llvm/ADT/StringExtras.h"
|
||||
#include "llvm/Analysis/AssumptionCache.h"
|
||||
#include "llvm/Analysis/EphemeralValuesCache.h"
|
||||
#include "llvm/Analysis/IR2Vec.h"
|
||||
#include "llvm/Analysis/InlineCost.h"
|
||||
#include "llvm/Analysis/OptimizationRemarkEmitter.h"
|
||||
#include "llvm/Analysis/ProfileSummaryInfo.h"
|
||||
@@ -64,6 +65,13 @@ static cl::opt<bool>
|
||||
cl::desc("If true, annotate inline advisor remarks "
|
||||
"with LTO and pass information."));
|
||||
|
||||
// This flag is used to enable IR2Vec embeddings in the ML inliner; Only valid
|
||||
// with ML inliner. The vocab file is used to initialize the embeddings.
|
||||
static cl::opt<std::string> IR2VecVocabFile(
|
||||
"ml-inliner-ir2vec-vocab-file", cl::Hidden,
|
||||
cl::desc("Vocab file for IR2Vec; Setting this enables "
|
||||
"configuring the model to use IR2Vec embeddings."));
|
||||
|
||||
namespace llvm {
|
||||
extern cl::opt<InlinerFunctionImportStatsOpts> InlinerFunctionImportStats;
|
||||
} // namespace llvm
|
||||
@@ -206,6 +214,20 @@ void InlineAdvice::recordInliningWithCalleeDeleted() {
|
||||
AnalysisKey InlineAdvisorAnalysis::Key;
|
||||
AnalysisKey PluginInlineAdvisorAnalysis::Key;
|
||||
|
||||
bool InlineAdvisorAnalysis::initializeIR2VecVocabIfRequested(
|
||||
Module &M, ModuleAnalysisManager &MAM) {
|
||||
if (!IR2VecVocabFile.empty()) {
|
||||
auto IR2VecVocabResult = MAM.getResult<IR2VecVocabAnalysis>(M);
|
||||
if (!IR2VecVocabResult.isValid()) {
|
||||
M.getContext().emitError("Failed to load IR2Vec vocabulary");
|
||||
return false;
|
||||
}
|
||||
}
|
||||
// No vocab file specified is OK; We just don't use IR2Vec
|
||||
// embeddings.
|
||||
return true;
|
||||
}
|
||||
|
||||
bool InlineAdvisorAnalysis::Result::tryCreate(
|
||||
InlineParams Params, InliningAdvisorMode Mode,
|
||||
const ReplayInlinerSettings &ReplaySettings, InlineContext IC) {
|
||||
@@ -231,14 +253,21 @@ bool InlineAdvisorAnalysis::Result::tryCreate(
|
||||
/* EmitRemarks =*/true, IC);
|
||||
}
|
||||
break;
|
||||
// Run IR2VecVocabAnalysis once per module to get the vocabulary.
|
||||
// We run it here because it is immutable and we want to avoid running it
|
||||
// multiple times.
|
||||
case InliningAdvisorMode::Development:
|
||||
#ifdef LLVM_HAVE_TFLITE
|
||||
LLVM_DEBUG(dbgs() << "Using development-mode inliner policy.\n");
|
||||
if (!InlineAdvisorAnalysis::initializeIR2VecVocabIfRequested(M, MAM))
|
||||
return false;
|
||||
Advisor = llvm::getDevelopmentModeAdvisor(M, MAM, GetDefaultAdvice);
|
||||
#endif
|
||||
break;
|
||||
case InliningAdvisorMode::Release:
|
||||
LLVM_DEBUG(dbgs() << "Using release-mode inliner policy.\n");
|
||||
if (!InlineAdvisorAnalysis::initializeIR2VecVocabIfRequested(M, MAM))
|
||||
return false;
|
||||
Advisor = llvm::getReleaseModeAdvisor(M, MAM, GetDefaultAdvice);
|
||||
break;
|
||||
}
|
||||
|
||||
@@ -107,7 +107,7 @@ static cl::opt<bool> KeepFPICache(
|
||||
cl::init(false));
|
||||
|
||||
// clang-format off
|
||||
const std::vector<TensorSpec> llvm::FeatureMap{
|
||||
std::vector<TensorSpec> llvm::FeatureMap{
|
||||
#define POPULATE_NAMES(DTYPE, SHAPE, NAME, __) TensorSpec::createSpec<DTYPE>(#NAME, SHAPE),
|
||||
// InlineCost features - these must come first
|
||||
INLINE_COST_FEATURE_ITERATOR(POPULATE_NAMES)
|
||||
@@ -144,6 +144,7 @@ MLInlineAdvisor::MLInlineAdvisor(
|
||||
M, MAM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager()),
|
||||
ModelRunner(std::move(Runner)), GetDefaultAdvice(GetDefaultAdvice),
|
||||
CG(MAM.getResult<LazyCallGraphAnalysis>(M)),
|
||||
UseIR2Vec(MAM.getCachedResult<IR2VecVocabAnalysis>(M) != nullptr),
|
||||
InitialIRSize(getModuleIRSize()), CurrentIRSize(InitialIRSize),
|
||||
PSI(MAM.getResult<ProfileSummaryAnalysis>(M)) {
|
||||
assert(ModelRunner);
|
||||
@@ -186,6 +187,19 @@ MLInlineAdvisor::MLInlineAdvisor(
|
||||
EdgeCount += getLocalCalls(KVP.first->getFunction());
|
||||
}
|
||||
NodeCount = AllNodes.size();
|
||||
|
||||
if (auto IR2VecVocabResult = MAM.getCachedResult<IR2VecVocabAnalysis>(M)) {
|
||||
if (!IR2VecVocabResult->isValid()) {
|
||||
M.getContext().emitError("IR2VecVocabAnalysis is not valid");
|
||||
return;
|
||||
}
|
||||
// Add the IR2Vec features to the feature map
|
||||
auto IR2VecDim = IR2VecVocabResult->getDimension();
|
||||
FeatureMap.push_back(
|
||||
TensorSpec::createSpec<float>("callee_embedding", {IR2VecDim}));
|
||||
FeatureMap.push_back(
|
||||
TensorSpec::createSpec<float>("caller_embedding", {IR2VecDim}));
|
||||
}
|
||||
}
|
||||
|
||||
unsigned MLInlineAdvisor::getInitialFunctionLevel(const Function &F) const {
|
||||
@@ -433,6 +447,22 @@ std::unique_ptr<InlineAdvice> MLInlineAdvisor::getAdviceImpl(CallBase &CB) {
|
||||
*ModelRunner->getTensor<int64_t>(FeatureIndex::is_caller_avail_external) =
|
||||
Caller.hasAvailableExternallyLinkage();
|
||||
|
||||
if (UseIR2Vec) {
|
||||
// Python side expects float embeddings. The IR2Vec embeddings are doubles
|
||||
// as of now due to the restriction of fromJSON method used by the
|
||||
// readVocabulary method in ir2vec::Embeddings.
|
||||
auto setEmbedding = [&](const ir2vec::Embedding &Embedding,
|
||||
FeatureIndex Index) {
|
||||
llvm::transform(Embedding, ModelRunner->getTensor<float>(Index),
|
||||
[](double Val) { return static_cast<float>(Val); });
|
||||
};
|
||||
|
||||
setEmbedding(CalleeBefore.getFunctionEmbedding(),
|
||||
FeatureIndex::callee_embedding);
|
||||
setEmbedding(CallerBefore.getFunctionEmbedding(),
|
||||
FeatureIndex::caller_embedding);
|
||||
}
|
||||
|
||||
// Add the cost features
|
||||
for (size_t I = 0;
|
||||
I < static_cast<size_t>(InlineCostFeatureIndex::NumberOfFeatures); ++I) {
|
||||
@@ -441,8 +471,7 @@ std::unique_ptr<InlineAdvice> MLInlineAdvisor::getAdviceImpl(CallBase &CB) {
|
||||
}
|
||||
// This one would have been set up to be right at the end.
|
||||
if (!InteractiveChannelBaseName.empty() && InteractiveIncludeDefault)
|
||||
*ModelRunner->getTensor<int64_t>(FeatureIndex::NumberOfFeatures) =
|
||||
GetDefaultAdvice(CB);
|
||||
*ModelRunner->getTensor<int64_t>(FeatureMap.size()) = GetDefaultAdvice(CB);
|
||||
return getAdviceFromModel(CB, ORE);
|
||||
}
|
||||
|
||||
@@ -520,7 +549,7 @@ void MLInlineAdvice::reportContextForRemark(
|
||||
DiagnosticInfoOptimizationBase &OR) {
|
||||
using namespace ore;
|
||||
OR << NV("Callee", Callee->getName());
|
||||
for (size_t I = 0; I < NumberOfFeatures; ++I)
|
||||
for (size_t I = 0; I < FeatureMap.size(); ++I)
|
||||
OR << NV(FeatureMap[I].name(),
|
||||
*getAdvisor()->getModelRunner().getTensor<int64_t>(I));
|
||||
OR << NV("ShouldInline", isInliningRecommended());
|
||||
|
||||
@@ -8,6 +8,7 @@
|
||||
|
||||
#include "llvm/Analysis/FunctionPropertiesAnalysis.h"
|
||||
#include "llvm/Analysis/AliasAnalysis.h"
|
||||
#include "llvm/Analysis/IR2Vec.h"
|
||||
#include "llvm/Analysis/LoopInfo.h"
|
||||
#include "llvm/AsmParser/Parser.h"
|
||||
#include "llvm/IR/Dominators.h"
|
||||
@@ -20,15 +21,20 @@
|
||||
#include "llvm/Support/Compiler.h"
|
||||
#include "llvm/Support/SourceMgr.h"
|
||||
#include "llvm/Transforms/Utils/Cloning.h"
|
||||
#include "gmock/gmock.h"
|
||||
#include "gtest/gtest.h"
|
||||
#include <cstring>
|
||||
|
||||
using namespace llvm;
|
||||
using namespace testing;
|
||||
|
||||
namespace llvm {
|
||||
LLVM_ABI extern cl::opt<bool> EnableDetailedFunctionProperties;
|
||||
LLVM_ABI extern cl::opt<bool> BigBasicBlockInstructionThreshold;
|
||||
LLVM_ABI extern cl::opt<bool> MediumBasicBlockInstrutionThreshold;
|
||||
LLVM_ABI extern cl::opt<float> ir2vec::OpcWeight;
|
||||
LLVM_ABI extern cl::opt<float> ir2vec::TypeWeight;
|
||||
LLVM_ABI extern cl::opt<float> ir2vec::ArgWeight;
|
||||
} // namespace llvm
|
||||
|
||||
namespace {
|
||||
@@ -36,17 +42,65 @@ namespace {
|
||||
class FunctionPropertiesAnalysisTest : public testing::Test {
|
||||
public:
|
||||
FunctionPropertiesAnalysisTest() {
|
||||
createTestVocabulary(1);
|
||||
MAM.registerPass([&] { return IR2VecVocabAnalysis(Vocabulary); });
|
||||
MAM.registerPass([&] { return PassInstrumentationAnalysis(); });
|
||||
FAM.registerPass([&] { return ModuleAnalysisManagerFunctionProxy(MAM); });
|
||||
FAM.registerPass([&] { return DominatorTreeAnalysis(); });
|
||||
FAM.registerPass([&] { return LoopAnalysis(); });
|
||||
FAM.registerPass([&] { return PassInstrumentationAnalysis(); });
|
||||
|
||||
ir2vec::OpcWeight = 1.0;
|
||||
ir2vec::TypeWeight = 1.0;
|
||||
ir2vec::ArgWeight = 1.0;
|
||||
}
|
||||
|
||||
private:
|
||||
float OriginalOpcWeight = ir2vec::OpcWeight;
|
||||
float OriginalTypeWeight = ir2vec::TypeWeight;
|
||||
float OriginalArgWeight = ir2vec::ArgWeight;
|
||||
|
||||
void createTestVocabulary(unsigned Dim) {
|
||||
llvm::SmallVector<std::string, 32> SampleEntities = {
|
||||
"add", "sub", "mul", "icmp", "br",
|
||||
"ret", "store", "load", "alloca", "phi",
|
||||
"call", "voidTy", "floatTy", "integerTy", "functionTy",
|
||||
"structTy", "arrayTy", "pointerTy", "vectorTy", "emptyTy",
|
||||
"labelTy", "tokenTy", "metadataTy", "unknownTy", "function",
|
||||
"pointer", "constant", "variable", "getelementptr", "invoke",
|
||||
"landingpad", "resume", "catch", "cleanup"};
|
||||
float EmbVal = 0.1;
|
||||
|
||||
// Helper lambda to add entries to the vocabulary
|
||||
auto addEntry = [&](std::string key) {
|
||||
Vocabulary[key] = ir2vec::Embedding(Dim, EmbVal);
|
||||
EmbVal += 0.1;
|
||||
};
|
||||
|
||||
for (auto &Name : SampleEntities)
|
||||
addEntry(Name);
|
||||
return;
|
||||
}
|
||||
|
||||
protected:
|
||||
std::unique_ptr<DominatorTree> DT;
|
||||
std::unique_ptr<LoopInfo> LI;
|
||||
FunctionAnalysisManager FAM;
|
||||
ModuleAnalysisManager MAM;
|
||||
ir2vec::Vocab Vocabulary;
|
||||
|
||||
void TearDown() override {
|
||||
// Restore original IR2Vec weights
|
||||
ir2vec::OpcWeight = OriginalOpcWeight;
|
||||
ir2vec::TypeWeight = OriginalTypeWeight;
|
||||
ir2vec::ArgWeight = OriginalArgWeight;
|
||||
}
|
||||
|
||||
FunctionPropertiesInfo buildFPI(Function &F) {
|
||||
// FunctionPropertiesInfo assumes IR2VecVocabAnalysis has been run to
|
||||
// use IR2Vec.
|
||||
auto VocabResult = MAM.getResult<IR2VecVocabAnalysis>(*F.getParent());
|
||||
(void)VocabResult;
|
||||
return FunctionPropertiesInfo::getFunctionPropertiesInfo(F, FAM);
|
||||
}
|
||||
|
||||
@@ -71,6 +125,13 @@ protected:
|
||||
return CB;
|
||||
return nullptr;
|
||||
}
|
||||
|
||||
std::unique_ptr<ir2vec::Embedder> createEmbedder(const Function &F) {
|
||||
auto EmbResult =
|
||||
ir2vec::Embedder::create(IR2VecKind::Symbolic, F, Vocabulary);
|
||||
EXPECT_TRUE(static_cast<bool>(EmbResult));
|
||||
return std::move(*EmbResult);
|
||||
}
|
||||
};
|
||||
|
||||
TEST_F(FunctionPropertiesAnalysisTest, BasicTest) {
|
||||
@@ -113,6 +174,8 @@ define internal i32 @top() {
|
||||
EXPECT_EQ(BranchesFeatures.StoreInstCount, 0);
|
||||
EXPECT_EQ(BranchesFeatures.MaxLoopDepth, 0);
|
||||
EXPECT_EQ(BranchesFeatures.TopLevelLoopCount, 0);
|
||||
EXPECT_TRUE(BranchesFeatures.getFunctionEmbedding().approximatelyEquals(
|
||||
createEmbedder(*BranchesFunction)->getFunctionVector()));
|
||||
|
||||
Function *TopFunction = M->getFunction("top");
|
||||
FunctionPropertiesInfo TopFeatures = buildFPI(*TopFunction);
|
||||
@@ -120,6 +183,8 @@ define internal i32 @top() {
|
||||
EXPECT_EQ(TopFeatures.BlocksReachedFromConditionalInstruction, 0);
|
||||
EXPECT_EQ(TopFeatures.Uses, 0);
|
||||
EXPECT_EQ(TopFeatures.DirectCallsToDefinedFunctions, 1);
|
||||
EXPECT_TRUE(TopFeatures.getFunctionEmbedding().approximatelyEquals(
|
||||
createEmbedder(*TopFunction)->getFunctionVector()));
|
||||
EXPECT_EQ(BranchesFeatures.LoadInstCount, 0);
|
||||
EXPECT_EQ(BranchesFeatures.StoreInstCount, 0);
|
||||
EXPECT_EQ(BranchesFeatures.MaxLoopDepth, 0);
|
||||
@@ -159,6 +224,9 @@ define internal i32 @top() {
|
||||
EXPECT_EQ(DetailedBranchesFeatures.CallReturnsPointerCount, 0);
|
||||
EXPECT_EQ(DetailedBranchesFeatures.CallWithManyArgumentsCount, 0);
|
||||
EXPECT_EQ(DetailedBranchesFeatures.CallWithPointerArgumentCount, 0);
|
||||
EXPECT_TRUE(
|
||||
DetailedBranchesFeatures.getFunctionEmbedding().approximatelyEquals(
|
||||
createEmbedder(*BranchesFunction)->getFunctionVector()));
|
||||
EnableDetailedFunctionProperties.setValue(false);
|
||||
}
|
||||
|
||||
@@ -210,6 +278,8 @@ finally:
|
||||
EXPECT_EQ(DetailedF1Properties.CallReturnsPointerCount, 0);
|
||||
EXPECT_EQ(DetailedF1Properties.CallWithManyArgumentsCount, 0);
|
||||
EXPECT_EQ(DetailedF1Properties.CallWithPointerArgumentCount, 0);
|
||||
EXPECT_TRUE(DetailedF1Properties.getFunctionEmbedding().approximatelyEquals(
|
||||
createEmbedder(*F1)->getFunctionVector()));
|
||||
EnableDetailedFunctionProperties.setValue(false);
|
||||
}
|
||||
|
||||
@@ -240,6 +310,8 @@ define i32 @f2(i32 %a) {
|
||||
ExpectedInitial.TotalInstructionCount = 3;
|
||||
ExpectedInitial.Uses = 1;
|
||||
ExpectedInitial.DirectCallsToDefinedFunctions = 1;
|
||||
ExpectedInitial.setFunctionEmbeddingForTest(
|
||||
createEmbedder(*F1)->getFunctionVector());
|
||||
|
||||
FunctionPropertiesInfo ExpectedFinal = ExpectedInitial;
|
||||
ExpectedFinal.DirectCallsToDefinedFunctions = 0;
|
||||
@@ -252,6 +324,9 @@ define i32 @f2(i32 %a) {
|
||||
auto IR = llvm::InlineFunction(*CB, IFI);
|
||||
EXPECT_TRUE(IR.isSuccess());
|
||||
invalidate(*F1);
|
||||
ExpectedFinal.setFunctionEmbeddingForTest(
|
||||
createEmbedder(*F1)->getFunctionVector());
|
||||
|
||||
EXPECT_TRUE(FPU.finishAndTest(FAM));
|
||||
EXPECT_EQ(FPI, ExpectedFinal);
|
||||
}
|
||||
@@ -294,6 +369,8 @@ define i32 @f2(i32 %a) {
|
||||
ExpectedInitial.TotalInstructionCount = 9;
|
||||
ExpectedInitial.Uses = 1;
|
||||
ExpectedInitial.DirectCallsToDefinedFunctions = 1;
|
||||
ExpectedInitial.setFunctionEmbeddingForTest(
|
||||
createEmbedder(*F1)->getFunctionVector());
|
||||
|
||||
FunctionPropertiesInfo ExpectedFinal = ExpectedInitial;
|
||||
ExpectedFinal.DirectCallsToDefinedFunctions = 0;
|
||||
@@ -307,6 +384,9 @@ define i32 @f2(i32 %a) {
|
||||
EXPECT_TRUE(IR.isSuccess());
|
||||
invalidate(*F1);
|
||||
EXPECT_TRUE(FPU.finishAndTest(FAM));
|
||||
|
||||
ExpectedFinal.setFunctionEmbeddingForTest(
|
||||
createEmbedder(*F1)->getFunctionVector());
|
||||
EXPECT_EQ(FPI, ExpectedFinal);
|
||||
}
|
||||
|
||||
@@ -356,6 +436,8 @@ exit:
|
||||
ExpectedInitial.TotalInstructionCount = 9;
|
||||
ExpectedInitial.Uses = 1;
|
||||
ExpectedInitial.DirectCallsToDefinedFunctions = 1;
|
||||
ExpectedInitial.setFunctionEmbeddingForTest(
|
||||
createEmbedder(*F1)->getFunctionVector());
|
||||
|
||||
FunctionPropertiesInfo ExpectedFinal;
|
||||
ExpectedFinal.BasicBlockCount = 6;
|
||||
@@ -374,6 +456,9 @@ exit:
|
||||
EXPECT_TRUE(IR.isSuccess());
|
||||
invalidate(*F1);
|
||||
EXPECT_TRUE(FPU.finishAndTest(FAM));
|
||||
|
||||
ExpectedFinal.setFunctionEmbeddingForTest(
|
||||
createEmbedder(*F1)->getFunctionVector());
|
||||
EXPECT_EQ(FPI, ExpectedFinal);
|
||||
}
|
||||
|
||||
@@ -422,6 +507,8 @@ declare i32 @__gxx_personality_v0(...)
|
||||
EXPECT_EQ(static_cast<size_t>(FPI.BasicBlockCount), F1->size());
|
||||
EXPECT_EQ(static_cast<size_t>(FPI.TotalInstructionCount),
|
||||
F1->getInstructionCount());
|
||||
EXPECT_TRUE(FPI.getFunctionEmbedding().approximatelyEquals(
|
||||
createEmbedder(*F1)->getFunctionVector()));
|
||||
}
|
||||
|
||||
TEST_F(FunctionPropertiesAnalysisTest, InvokeUnreachableHandler) {
|
||||
@@ -475,6 +562,8 @@ declare i32 @__gxx_personality_v0(...)
|
||||
EXPECT_EQ(static_cast<size_t>(FPI.BasicBlockCount), F1->size() - 1);
|
||||
EXPECT_EQ(static_cast<size_t>(FPI.TotalInstructionCount),
|
||||
F1->getInstructionCount() - 2);
|
||||
EXPECT_TRUE(FPI.getFunctionEmbedding().approximatelyEquals(
|
||||
createEmbedder(*F1)->getFunctionVector()));
|
||||
EXPECT_EQ(FPI, FunctionPropertiesInfo::getFunctionPropertiesInfo(*F1, FAM));
|
||||
}
|
||||
|
||||
@@ -581,6 +670,8 @@ lpad:
|
||||
EXPECT_EQ(static_cast<size_t>(FPI.BasicBlockCount), F1->size() - 1);
|
||||
EXPECT_EQ(static_cast<size_t>(FPI.TotalInstructionCount),
|
||||
F1->getInstructionCount() - 2);
|
||||
EXPECT_TRUE(FPI.getFunctionEmbedding().approximatelyEquals(
|
||||
createEmbedder(*F1)->getFunctionVector()));
|
||||
EXPECT_EQ(FPI, FunctionPropertiesInfo::getFunctionPropertiesInfo(*F1, FAM));
|
||||
}
|
||||
|
||||
@@ -637,6 +728,8 @@ lpad:
|
||||
EXPECT_EQ(static_cast<size_t>(FPI.BasicBlockCount), F1->size() - 1);
|
||||
EXPECT_EQ(static_cast<size_t>(FPI.TotalInstructionCount),
|
||||
F1->getInstructionCount() - 2);
|
||||
EXPECT_TRUE(FPI.getFunctionEmbedding().approximatelyEquals(
|
||||
createEmbedder(*F1)->getFunctionVector()));
|
||||
EXPECT_EQ(FPI, FunctionPropertiesInfo::getFunctionPropertiesInfo(*F1, FAM));
|
||||
}
|
||||
|
||||
@@ -689,6 +782,8 @@ end:
|
||||
ExpectedInitial.DirectCallsToDefinedFunctions = 1;
|
||||
ExpectedInitial.MaxLoopDepth = 1;
|
||||
ExpectedInitial.TopLevelLoopCount = 1;
|
||||
ExpectedInitial.setFunctionEmbeddingForTest(
|
||||
createEmbedder(*F1)->getFunctionVector());
|
||||
|
||||
FunctionPropertiesInfo ExpectedFinal = ExpectedInitial;
|
||||
ExpectedFinal.BasicBlockCount = 6;
|
||||
@@ -705,6 +800,9 @@ end:
|
||||
EXPECT_TRUE(IR.isSuccess());
|
||||
invalidate(*F1);
|
||||
EXPECT_TRUE(FPU.finishAndTest(FAM));
|
||||
|
||||
ExpectedFinal.setFunctionEmbeddingForTest(
|
||||
createEmbedder(*F1)->getFunctionVector());
|
||||
EXPECT_EQ(FPI, ExpectedFinal);
|
||||
}
|
||||
|
||||
@@ -733,7 +831,7 @@ extra:
|
||||
extra2:
|
||||
br label %cond.end
|
||||
|
||||
cond.end: ; preds = %cond.false, %cond.true
|
||||
cond.end: ; preds = %extra2, %cond.true
|
||||
%cond = phi i64 [ %conv2, %cond.true ], [ %call3, %extra ]
|
||||
ret i64 %cond
|
||||
}
|
||||
@@ -757,6 +855,8 @@ declare void @llvm.trap()
|
||||
ExpectedInitial.BlocksReachedFromConditionalInstruction = 2;
|
||||
ExpectedInitial.Uses = 1;
|
||||
ExpectedInitial.DirectCallsToDefinedFunctions = 1;
|
||||
ExpectedInitial.setFunctionEmbeddingForTest(
|
||||
createEmbedder(*F1)->getFunctionVector());
|
||||
|
||||
FunctionPropertiesInfo ExpectedFinal = ExpectedInitial;
|
||||
ExpectedFinal.BasicBlockCount = 4;
|
||||
@@ -772,6 +872,9 @@ declare void @llvm.trap()
|
||||
EXPECT_TRUE(IR.isSuccess());
|
||||
invalidate(*F1);
|
||||
EXPECT_TRUE(FPU.finishAndTest(FAM));
|
||||
|
||||
ExpectedFinal.setFunctionEmbeddingForTest(
|
||||
createEmbedder(*F1)->getFunctionVector());
|
||||
EXPECT_EQ(FPI, ExpectedFinal);
|
||||
}
|
||||
|
||||
@@ -817,6 +920,8 @@ declare void @f3()
|
||||
ExpectedInitial.BlocksReachedFromConditionalInstruction = 0;
|
||||
ExpectedInitial.Uses = 1;
|
||||
ExpectedInitial.DirectCallsToDefinedFunctions = 1;
|
||||
ExpectedInitial.setFunctionEmbeddingForTest(
|
||||
createEmbedder(*F1)->getFunctionVector());
|
||||
|
||||
FunctionPropertiesInfo ExpectedFinal = ExpectedInitial;
|
||||
ExpectedFinal.BasicBlockCount = 6;
|
||||
@@ -832,6 +937,9 @@ declare void @f3()
|
||||
EXPECT_TRUE(IR.isSuccess());
|
||||
invalidate(*F1);
|
||||
EXPECT_TRUE(FPU.finishAndTest(FAM));
|
||||
|
||||
ExpectedFinal.setFunctionEmbeddingForTest(
|
||||
createEmbedder(*F1)->getFunctionVector());
|
||||
EXPECT_EQ(FPI, ExpectedFinal);
|
||||
}
|
||||
|
||||
@@ -885,6 +993,8 @@ define i64 @f1(i64 %e) {
|
||||
EXPECT_EQ(DetailedF1Properties.CallReturnsPointerCount, 0);
|
||||
EXPECT_EQ(DetailedF1Properties.CallWithManyArgumentsCount, 0);
|
||||
EXPECT_EQ(DetailedF1Properties.CallWithPointerArgumentCount, 0);
|
||||
EXPECT_TRUE(DetailedF1Properties.getFunctionEmbedding().approximatelyEquals(
|
||||
createEmbedder(*F1)->getFunctionVector()));
|
||||
EnableDetailedFunctionProperties.setValue(false);
|
||||
}
|
||||
|
||||
@@ -910,6 +1020,8 @@ declare float @llvm.cos.f32(float)
|
||||
EXPECT_EQ(DetailedF1Properties.CallReturnsPointerCount, 0);
|
||||
EXPECT_EQ(DetailedF1Properties.CallWithManyArgumentsCount, 0);
|
||||
EXPECT_EQ(DetailedF1Properties.CallWithPointerArgumentCount, 0);
|
||||
EXPECT_TRUE(DetailedF1Properties.getFunctionEmbedding().approximatelyEquals(
|
||||
createEmbedder(*F1)->getFunctionVector()));
|
||||
EnableDetailedFunctionProperties.setValue(false);
|
||||
}
|
||||
|
||||
@@ -943,6 +1055,8 @@ declare float @f5()
|
||||
EXPECT_EQ(DetailedF1Properties.CallReturnsPointerCount, 1);
|
||||
EXPECT_EQ(DetailedF1Properties.CallWithManyArgumentsCount, 1);
|
||||
EXPECT_EQ(DetailedF1Properties.CallWithPointerArgumentCount, 1);
|
||||
EXPECT_TRUE(DetailedF1Properties.getFunctionEmbedding().approximatelyEquals(
|
||||
createEmbedder(*F1)->getFunctionVector()));
|
||||
EnableDetailedFunctionProperties.setValue(false);
|
||||
}
|
||||
|
||||
@@ -972,6 +1086,8 @@ BottomBlock2:
|
||||
EnableDetailedFunctionProperties.setValue(true);
|
||||
FunctionPropertiesInfo DetailedF1Properties = buildFPI(*F1);
|
||||
EXPECT_EQ(DetailedF1Properties.CriticalEdgeCount, 1);
|
||||
EXPECT_TRUE(DetailedF1Properties.getFunctionEmbedding().approximatelyEquals(
|
||||
createEmbedder(*F1)->getFunctionVector()));
|
||||
EnableDetailedFunctionProperties.setValue(false);
|
||||
}
|
||||
|
||||
@@ -997,6 +1113,8 @@ declare <4 x ptr> @f4()
|
||||
EXPECT_EQ(DetailedF1Properties.CallReturnsVectorIntCount, 1);
|
||||
EXPECT_EQ(DetailedF1Properties.CallReturnsVectorFloatCount, 1);
|
||||
EXPECT_EQ(DetailedF1Properties.CallReturnsVectorPointerCount, 1);
|
||||
EXPECT_TRUE(DetailedF1Properties.getFunctionEmbedding().approximatelyEquals(
|
||||
createEmbedder(*F1)->getFunctionVector()));
|
||||
EnableDetailedFunctionProperties.setValue(false);
|
||||
}
|
||||
|
||||
|
||||
Reference in New Issue
Block a user