summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorManman Ren <mren@apple.com>2012-08-28 22:21:25 +0000
committerManman Ren <mren@apple.com>2012-08-28 22:21:25 +0000
commitd26200423ee818e54d4088bd0c499caf840d866d (patch)
treeebcd7b539baa265749b4e8bd01de29b8b3c2ecbe
parent714313b4828cec98b086b54b356407540aa775c4 (diff)
downloadllvm-d26200423ee818e54d4088bd0c499caf840d866d.tar.gz
llvm-d26200423ee818e54d4088bd0c499caf840d866d.tar.bz2
llvm-d26200423ee818e54d4088bd0c499caf840d866d.tar.xz
Profile: set branch weight metadata with data generated from profiling.
This patch implements ProfileDataLoader which loads profile data generated by -insert-edge-profiling and updates branch weight metadata accordingly. Patch by Alastair Murray. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@162799 91177308-0d34-0410-b5e6-96231b3b80d8
-rw-r--r--include/llvm/Analysis/Passes.h8
-rw-r--r--include/llvm/Analysis/ProfileDataLoader.h148
-rw-r--r--include/llvm/InitializePasses.h1
-rw-r--r--include/llvm/LinkAllPasses.h1
-rw-r--r--lib/Analysis/Analysis.cpp1
-rw-r--r--lib/Analysis/CMakeLists.txt2
-rw-r--r--lib/Analysis/ProfileDataLoader.cpp186
-rw-r--r--lib/Analysis/ProfileDataLoaderPass.cpp188
-rw-r--r--lib/Analysis/ProfileInfo.cpp26
-rw-r--r--test/Analysis/Profiling/load-branch-weights-ifs.ll119
-rw-r--r--test/Analysis/Profiling/load-branch-weights-loops.ll185
-rw-r--r--test/Analysis/Profiling/load-branch-weights-switches.ll162
12 files changed, 1001 insertions, 26 deletions
diff --git a/include/llvm/Analysis/Passes.h b/include/llvm/Analysis/Passes.h
index a22bd12dec..c52f846b5c 100644
--- a/include/llvm/Analysis/Passes.h
+++ b/include/llvm/Analysis/Passes.h
@@ -103,6 +103,14 @@ namespace llvm {
//===--------------------------------------------------------------------===//
//
+ // createProfileMetadataLoaderPass - This pass loads information from a
+ // profile dump file and sets branch weight metadata.
+ //
+ ModulePass *createProfileMetadataLoaderPass();
+ extern char &ProfileMetadataLoaderPassID;
+
+ //===--------------------------------------------------------------------===//
+ //
// createNoProfileInfoPass - This pass implements the default "no profile".
//
ImmutablePass *createNoProfileInfoPass();
diff --git a/include/llvm/Analysis/ProfileDataLoader.h b/include/llvm/Analysis/ProfileDataLoader.h
new file mode 100644
index 0000000000..3d15bda364
--- /dev/null
+++ b/include/llvm/Analysis/ProfileDataLoader.h
@@ -0,0 +1,148 @@
+//===- ProfileDataLoader.h - Load & convert profile info ----*- C++ -*-===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// The ProfileDataLoader class is used to load profiling data from a dump file.
+// The ProfileDataT<FType, BType> class is used to store the mapping of this
+// data to control flow edges.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef LLVM_ANALYSIS_PROFILEDATALOADER_H
+#define LLVM_ANALYSIS_PROFILEDATALOADER_H
+
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/ErrorHandling.h"
+#include <vector>
+#include <string>
+#include <map>
+
+namespace llvm {
+
+class ModulePass;
+class Function;
+class BasicBlock;
+
+// Helpers for dumping edges to dbgs().
+raw_ostream& operator<<(raw_ostream &O, std::pair<const BasicBlock *,
+ const BasicBlock *> E);
+raw_ostream& operator<<(raw_ostream &O, const BasicBlock *BB);
+raw_ostream& operator<<(raw_ostream &O, const Function *F);
+
+/// \brief The ProfileDataT<FType, BType> class is used to store the mapping of
+/// profiling data to control flow edges.
+///
+/// An edge is defined by its source and sink basic blocks.
+template<class FType, class BType>
+class ProfileDataT {
+ public:
+ // The profiling information defines an Edge by its source and sink basic
+ // blocks.
+ typedef std::pair<const BType*, const BType*> Edge;
+
+ private:
+ typedef std::map<Edge, unsigned> EdgeWeights;
+
+ /// \brief Count the number of times a transition between two blocks is
+ /// executed.
+ ///
+ /// As a special case, we also hold an edge from the null BasicBlock to the
+ /// entry block to indicate how many times the function was entered.
+ std::map<const FType*, EdgeWeights> EdgeInformation;
+
+ public:
+ static char ID; // Class identification, replacement for typeinfo
+ ProfileDataT() {};
+ ~ProfileDataT() {};
+
+ /// getFunction() - Returns the Function for an Edge.
+ static const FType* getFunction(Edge e) {
+ // e.first may be NULL
+ assert( ((!e.first) || (e.first->getParent() == e.second->getParent()))
+ && "A ProfileData::Edge can not be between two functions");
+ assert(e.second && "A ProfileData::Edge must have a real sink");
+ return e.second->getParent();
+ }
+
+ /// getEdge() - Creates an Edge between two BasicBlocks.
+ static Edge getEdge(const BType *Src, const BType *Dest) {
+ return std::make_pair(Src, Dest);
+ }
+
+ /// getEdgeWeight - Return the number of times that a given edge was
+ /// executed.
+ unsigned getEdgeWeight(Edge e) const {
+ const FType *f = getFunction(e);
+ assert( (EdgeInformation.find(f) != EdgeInformation.end())
+ && "No profiling information for function");
+ EdgeWeights weights = EdgeInformation.find(f)->second;
+
+ assert( (weights.find(e) != weights.end())
+ && "No profiling information for edge");
+ return weights.find(e)->second;
+ }
+
+ /// addEdgeWeight - Add 'weight' to the already stored execution count for
+ /// this edge.
+ void addEdgeWeight(Edge e, unsigned weight) {
+ EdgeInformation[getFunction(e)][e] += weight;
+ }
+};
+
+typedef ProfileDataT<Function, BasicBlock> ProfileData;
+//typedef ProfileDataT<MachineFunction, MachineBasicBlock> MachineProfileData;
+
+/// The ProfileDataLoader class is used to load raw profiling data from the
+/// dump file.
+class ProfileDataLoader {
+private:
+ /// The name of the file where the raw profiling data is stored.
+ const std::string &Filename;
+
+ /// A vector of the command line arguments used when the target program was
+ /// run to generate profiling data. One entry per program run.
+ std::vector<std::string> CommandLines;
+
+ /// The raw values for how many times each edge was traversed, values from
+ /// multiple program runs are accumulated.
+ std::vector<unsigned> EdgeCounts;
+
+public:
+ /// ProfileDataLoader ctor - Read the specified profiling data file, exiting
+ /// the program if the file is invalid or broken.
+ ProfileDataLoader(const char *ToolName, const std::string &Filename);
+
+ /// A special value used to represent the weight of an edge which has not
+ /// been counted yet.
+ static const unsigned Uncounted;
+
+ /// The maximum value that can be stored in a profiling counter.
+ static const unsigned MaxCount;
+
+ /// getNumExecutions - Return the number of times the target program was run
+ /// to generate this profiling data.
+ unsigned getNumExecutions() const { return CommandLines.size(); }
+
+ /// getExecution - Return the command line parameters used to generate the
+ /// i'th set of profiling data.
+ const std::string& getExecution(unsigned i) const { return CommandLines[i]; }
+
+ const std::string& getFileName() const { return Filename; }
+
+ /// getRawEdgeCounts - Return the raw profiling data, this is just a list of
+ /// numbers with no mappings to edges.
+ const std::vector<unsigned>& getRawEdgeCounts() const { return EdgeCounts; }
+};
+
+/// createProfileMetadataLoaderPass - This function returns a Pass that loads
+/// the profiling information for the module from the specified filename.
+ModulePass *createProfileMetadataLoaderPass(const std::string &Filename);
+
+} // End llvm namespace
+
+#endif
diff --git a/include/llvm/InitializePasses.h b/include/llvm/InitializePasses.h
index de97957a84..994311334c 100644
--- a/include/llvm/InitializePasses.h
+++ b/include/llvm/InitializePasses.h
@@ -141,6 +141,7 @@ void initializeLiveRegMatrixPass(PassRegistry&);
void initializeLiveStacksPass(PassRegistry&);
void initializeLiveVariablesPass(PassRegistry&);
void initializeLoaderPassPass(PassRegistry&);
+void initializeProfileMetadataLoaderPassPass(PassRegistry&);
void initializePathProfileLoaderPassPass(PassRegistry&);
void initializeLocalStackSlotPassPass(PassRegistry&);
void initializeLoopDeletionPass(PassRegistry&);
diff --git a/include/llvm/LinkAllPasses.h b/include/llvm/LinkAllPasses.h
index 697c94c094..fe4c92a295 100644
--- a/include/llvm/LinkAllPasses.h
+++ b/include/llvm/LinkAllPasses.h
@@ -107,6 +107,7 @@ namespace {
(void) llvm::createProfileVerifierPass();
(void) llvm::createPathProfileVerifierPass();
(void) llvm::createProfileLoaderPass();
+ (void) llvm::createProfileMetadataLoaderPass();
(void) llvm::createPathProfileLoaderPass();
(void) llvm::createPromoteMemoryToRegisterPass();
(void) llvm::createDemoteRegisterToMemoryPass();
diff --git a/lib/Analysis/Analysis.cpp b/lib/Analysis/Analysis.cpp
index 0ba6af93b5..87a75fd3b1 100644
--- a/lib/Analysis/Analysis.cpp
+++ b/lib/Analysis/Analysis.cpp
@@ -61,6 +61,7 @@ void llvm::initializeAnalysis(PassRegistry &Registry) {
initializePathProfileLoaderPassPass(Registry);
initializeProfileVerifierPassPass(Registry);
initializePathProfileVerifierPass(Registry);
+ initializeProfileMetadataLoaderPassPass(Registry);
initializeRegionInfoPass(Registry);
initializeRegionViewerPass(Registry);
initializeRegionPrinterPass(Registry);
diff --git a/lib/Analysis/CMakeLists.txt b/lib/Analysis/CMakeLists.txt
index 96e68b4199..e461848e86 100644
--- a/lib/Analysis/CMakeLists.txt
+++ b/lib/Analysis/CMakeLists.txt
@@ -44,6 +44,8 @@ add_llvm_library(LLVMAnalysis
ProfileInfoLoader.cpp
ProfileInfoLoaderPass.cpp
ProfileVerifierPass.cpp
+ ProfileDataLoader.cpp
+ ProfileDataLoaderPass.cpp
RegionInfo.cpp
RegionPass.cpp
RegionPrinter.cpp
diff --git a/lib/Analysis/ProfileDataLoader.cpp b/lib/Analysis/ProfileDataLoader.cpp
new file mode 100644
index 0000000000..000649228a
--- /dev/null
+++ b/lib/Analysis/ProfileDataLoader.cpp
@@ -0,0 +1,186 @@
+//===- ProfileDataLoader.cpp - Load profile information from disk ---------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// The ProfileDataLoader class is used to load raw profiling data from the dump
+// file.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Module.h"
+#include "llvm/InstrTypes.h"
+#include "llvm/Analysis/ProfileDataLoader.h"
+#include "llvm/Analysis/ProfileDataTypes.h"
+#include "llvm/Support/raw_ostream.h"
+#include <cstdio>
+#include <cstdlib>
+using namespace llvm;
+
+namespace llvm {
+
+template<>
+char ProfileDataT<Function,BasicBlock>::ID = 0;
+
+raw_ostream& operator<<(raw_ostream &O, const Function *F) {
+ return O << F->getName();
+}
+
+raw_ostream& operator<<(raw_ostream &O, const BasicBlock *BB) {
+ return O << BB->getName();
+}
+
+raw_ostream& operator<<(raw_ostream &O, std::pair<const BasicBlock *, const BasicBlock *> E) {
+ O << "(";
+
+ if (E.first)
+ O << E.first;
+ else
+ O << "0";
+
+ O << ",";
+
+ if (E.second)
+ O << E.second;
+ else
+ O << "0";
+
+ return O << ")";
+}
+
+} // namespace llvm
+
+/// ByteSwap - Byteswap 'Var' if 'Really' is true. Required when the compiler
+/// host and target have different endianness.
+static inline unsigned ByteSwap(unsigned Var, bool Really) {
+ if (!Really) return Var;
+ return ((Var & (255U<< 0U)) << 24U) |
+ ((Var & (255U<< 8U)) << 8U) |
+ ((Var & (255U<<16U)) >> 8U) |
+ ((Var & (255U<<24U)) >> 24U);
+}
+
+/// AddCounts - Add 'A' and 'B', accounting for the fact that the value of one
+/// (or both) may not be defined.
+static unsigned AddCounts(unsigned A, unsigned B) {
+ // If either value is undefined, use the other.
+ // Undefined + undefined = undefined.
+ if (A == ProfileDataLoader::Uncounted) return B;
+ if (B == ProfileDataLoader::Uncounted) return A;
+
+ // Saturate to the maximum storable value. This could change taken/nottaken
+ // ratios, but is presumably better than wrapping and thus potentially
+ // inverting ratios.
+ unsigned long long tmp = (unsigned long long)A + (unsigned long long)B;
+ if (tmp > (unsigned long long)ProfileDataLoader::MaxCount)
+ tmp = ProfileDataLoader::MaxCount;
+ return (unsigned)tmp;
+}
+
+/// ReadProfilingData - Load 'NumEntries' items of type 'T' from file 'F'
+template <typename T>
+static void ReadProfilingData(const char *ToolName, FILE *F,
+ std::vector<T> &Data, size_t NumEntries) {
+ // Read in the block of data...
+ if (fread(&Data[0], sizeof(T), NumEntries, F) != NumEntries) {
+ errs() << ToolName << ": profiling data truncated!\n";
+ perror(0);
+ exit(1);
+ }
+}
+
+/// ReadProfilingNumEntries - Read how many entries are in this profiling data
+/// packet.
+static unsigned ReadProfilingNumEntries(const char *ToolName, FILE *F,
+ bool ShouldByteSwap) {
+ std::vector<unsigned> NumEntries(1);
+ ReadProfilingData<unsigned>(ToolName, F, NumEntries, 1);
+ return ByteSwap(NumEntries[0], ShouldByteSwap);
+}
+
+/// ReadProfilingBlock - Read the number of entries in the next profiling data
+/// packet and then accumulate the entries into 'Data'.
+static void ReadProfilingBlock(const char *ToolName, FILE *F,
+ bool ShouldByteSwap,
+ std::vector<unsigned> &Data) {
+ // Read the number of entries...
+ unsigned NumEntries = ReadProfilingNumEntries(ToolName, F, ShouldByteSwap);
+
+ // Read in the data.
+ std::vector<unsigned> TempSpace(NumEntries);
+ ReadProfilingData<unsigned>(ToolName, F, TempSpace, (size_t)NumEntries);
+
+ // Make sure we have enough space ...
+ if (Data.size() < NumEntries)
+ Data.resize(NumEntries, ProfileDataLoader::Uncounted);
+
+ // Accumulate the data we just read into the existing data.
+ for (unsigned i = 0; i < NumEntries; ++i) {
+ Data[i] = AddCounts(ByteSwap(TempSpace[i], ShouldByteSwap), Data[i]);
+ }
+}
+
+/// ReadProfilingArgBlock - Read the command line arguments that the progam was
+/// run with when the current profiling data packet(s) were generated.
+static void ReadProfilingArgBlock(const char *ToolName, FILE *F,
+ bool ShouldByteSwap,
+ std::vector<std::string> &CommandLines) {
+ // Read the number of bytes ...
+ unsigned ArgLength = ReadProfilingNumEntries(ToolName, F, ShouldByteSwap);
+
+ // Read in the arguments (if there are any to read). Round up the length to
+ // the nearest 4-byte multiple.
+ std::vector<char> Args(ArgLength+4);
+ if (ArgLength)
+ ReadProfilingData<char>(ToolName, F, Args, (ArgLength+3) & ~3);
+
+ // Store the arguments.
+ CommandLines.push_back(std::string(&Args[0], &Args[ArgLength]));
+}
+
+const unsigned ProfileDataLoader::Uncounted = ~0U;
+const unsigned ProfileDataLoader::MaxCount = ~0U - 1U;
+
+/// ProfileDataLoader ctor - Read the specified profiling data file, exiting
+/// the program if the file is invalid or broken.
+ProfileDataLoader::ProfileDataLoader(const char *ToolName,
+ const std::string &Filename)
+ : Filename(Filename) {
+ FILE *F = fopen(Filename.c_str(), "rb");
+ if (F == 0) {
+ errs() << ToolName << ": Error opening '" << Filename << "': ";
+ perror(0);
+ exit(1);
+ }
+
+ // Keep reading packets until we run out of them.
+ unsigned PacketType;
+ while (fread(&PacketType, sizeof(unsigned), 1, F) == 1) {
+ // If the low eight bits of the packet are zero, we must be dealing with an
+ // endianness mismatch. Byteswap all words read from the profiling
+ // information. This can happen when the compiler host and target have
+ // different endianness.
+ bool ShouldByteSwap = (char)PacketType == 0;
+ PacketType = ByteSwap(PacketType, ShouldByteSwap);
+
+ switch (PacketType) {
+ case ArgumentInfo:
+ ReadProfilingArgBlock(ToolName, F, ShouldByteSwap, CommandLines);
+ break;
+
+ case EdgeInfo:
+ ReadProfilingBlock(ToolName, F, ShouldByteSwap, EdgeCounts);
+ break;
+
+ default:
+ errs() << ToolName << ": Unknown packet type #" << PacketType << "!\n";
+ exit(1);
+ }
+ }
+
+ fclose(F);
+}
diff --git a/lib/Analysis/ProfileDataLoaderPass.cpp b/lib/Analysis/ProfileDataLoaderPass.cpp
new file mode 100644
index 0000000000..2a61a0b6ed
--- /dev/null
+++ b/lib/Analysis/ProfileDataLoaderPass.cpp
@@ -0,0 +1,188 @@
+//===- ProfileDataLoaderPass.cpp - Set branch weight metadata from prof ---===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This pass loads profiling data from a dump file and sets branch weight
+// metadata.
+//
+// TODO: Replace all "profile-metadata-loader" strings with "profile-loader"
+// once ProfileInfo etc. has been removed.
+//
+//===----------------------------------------------------------------------===//
+#define DEBUG_TYPE "profile-metadata-loader"
+#include "llvm/BasicBlock.h"
+#include "llvm/InstrTypes.h"
+#include "llvm/Module.h"
+#include "llvm/LLVMContext.h"
+#include "llvm/MDBuilder.h"
+#include "llvm/Metadata.h"
+#include "llvm/Pass.h"
+#include "llvm/Analysis/Passes.h"
+#include "llvm/Analysis/ProfileDataLoader.h"
+#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/CFG.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/raw_ostream.h"
+#include "llvm/Support/Format.h"
+#include "llvm/ADT/Statistic.h"
+#include <vector>
+using namespace llvm;
+
+STATISTIC(NumEdgesRead, "The # of edges read.");
+STATISTIC(NumTermsAnnotated, "The # of terminator instructions annotated.");
+
+static cl::opt<std::string>
+ProfileMetadataFilename("profile-file", cl::init("llvmprof.out"),
+ cl::value_desc("filename"),
+ cl::desc("Profile file loaded by -profile-metadata-loader"));
+
+namespace {
+ /// This pass loads profiling data from a dump file and sets branch weight
+ /// metadata.
+ class ProfileMetadataLoaderPass : public ModulePass {
+ std::string Filename;
+ public:
+ static char ID; // Class identification, replacement for typeinfo
+ explicit ProfileMetadataLoaderPass(const std::string &filename = "")
+ : ModulePass(ID), Filename(filename) {
+ initializeProfileMetadataLoaderPassPass(*PassRegistry::getPassRegistry());
+ if (filename.empty()) Filename = ProfileMetadataFilename;
+ }
+
+ virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ AU.setPreservesAll();
+ }
+
+ virtual const char *getPassName() const {
+ return "Profile loader";
+ }
+
+ virtual void readEdge(unsigned, ProfileData&, ProfileData::Edge,
+ std::vector<unsigned>&);
+ virtual unsigned matchEdges(Module&, ProfileData&, std::vector<unsigned>&);
+ virtual void setBranchWeightMetadata(Module&, ProfileData&);
+
+ virtual bool runOnModule(Module &M);
+ };
+} // End of anonymous namespace
+
+char ProfileMetadataLoaderPass::ID = 0;
+INITIALIZE_PASS_BEGIN(ProfileMetadataLoaderPass, "profile-metadata-loader",
+ "Load profile information from llvmprof.out", false, true)
+INITIALIZE_PASS_END(ProfileMetadataLoaderPass, "profile-metadata-loader",
+ "Load profile information from llvmprof.out", false, true)
+
+char &llvm::ProfileMetadataLoaderPassID = ProfileMetadataLoaderPass::ID;
+
+/// createProfileMetadataLoaderPass - This function returns a Pass that loads
+/// the profiling information for the module from the specified filename,
+/// making it available to the optimizers.
+ModulePass *llvm::createProfileMetadataLoaderPass() {
+ return new ProfileMetadataLoaderPass();
+}
+ModulePass *llvm::createProfileMetadataLoaderPass(const std::string &Filename) {
+ return new ProfileMetadataLoaderPass(Filename);
+}
+
+/// readEdge - Take the value from a profile counter and assign it to an edge.
+void ProfileMetadataLoaderPass::readEdge(unsigned ReadCount,
+ ProfileData &PB, ProfileData::Edge e,
+ std::vector<unsigned> &Counters) {
+ if (ReadCount < Counters.size()) {
+ unsigned weight = Counters[ReadCount];
+ assert(weight != ProfileDataLoader::Uncounted);
+ PB.addEdgeWeight(e, weight);
+
+ DEBUG(dbgs() << "-- Read Edge Counter for " << e
+ << " (# "<< (ReadCount) << "): "
+ << PB.getEdgeWeight(e) << "\n");
+ }
+}
+
+/// matchEdges - Link every profile counter with an edge.
+unsigned ProfileMetadataLoaderPass::matchEdges(Module &M, ProfileData &PB,
+ std::vector<unsigned> &Counters) {
+ if (Counters.size() == 0) return 0;
+
+ unsigned ReadCount = 0;
+
+ for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
+ if (F->isDeclaration()) continue;
+ DEBUG(dbgs() << "Loading edges in '" << F->getName() << "'\n");
+ readEdge(ReadCount++, PB, PB.getEdge(0, &F->getEntryBlock()), Counters);
+ for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
+ TerminatorInst *TI = BB->getTerminator();
+ for (unsigned s = 0, e = TI->getNumSuccessors(); s != e; ++s) {
+ readEdge(ReadCount++, PB, PB.getEdge(BB,TI->getSuccessor(s)),
+ Counters);
+ }
+ }
+ }
+
+ return ReadCount;
+}
+
+/// setBranchWeightMetadata - Translate the counter values associated with each
+/// edge into branch weights for each conditional branch (a branch with 2 or
+/// more desinations).
+void ProfileMetadataLoaderPass::setBranchWeightMetadata(Module &M,
+ ProfileData &PB) {
+ for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
+ if (F->isDeclaration()) continue;
+ DEBUG(dbgs() << "Setting branch metadata in '" << F->getName() << "'\n");
+
+ for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
+ TerminatorInst *TI = BB->getTerminator();
+ unsigned NumSuccessors = TI->getNumSuccessors();
+
+ // If there is only one successor then we can not set a branch
+ // probability as the target is certain.
+ if (NumSuccessors < 2) continue;
+
+ // Load the weights of all edges leading from this terminator.
+ DEBUG(dbgs() << "-- Terminator with " << NumSuccessors
+ << " successors:\n");
+ std::vector<uint32_t> Weights(NumSuccessors);
+ for (unsigned s = 0 ; s < NumSuccessors ; ++s) {
+ ProfileData::Edge edge = PB.getEdge(BB, TI->getSuccessor(s));
+ Weights[s] = (uint32_t)PB.getEdgeWeight(edge);
+ DEBUG(dbgs() << "---- Edge '" << edge << "' has weight "
+ << Weights[s] << "\n");
+ }
+
+ // Set branch weight metadata. This will set branch probabilities of
+ // 100%/0% if that is true of the dynamic execution.
+ // BranchProbabilityInfo can account for this when it loads this metadata
+ // (it gives the unexectuted branch a weight of 1 for the purposes of
+ // probability calculations).
+ MDBuilder MDB(TI->getContext());
+ MDNode *Node = MDB.createBranchWeights(Weights);
+ TI->setMetadata(LLVMContext::MD_prof, Node);
+ NumTermsAnnotated++;
+ }
+ }
+}
+
+bool ProfileMetadataLoaderPass::runOnModule(Module &M) {
+ ProfileDataLoader PDL("profile-data-loader", Filename);
+ ProfileData PB;
+
+ std::vector<unsigned> Counters = PDL.getRawEdgeCounts();
+
+ unsigned ReadCount = matchEdges(M, PB, Counters);
+
+ if (ReadCount != Counters.size()) {
+ errs() << "WARNING: profile information is inconsistent with "
+ << "the current program!\n";
+ }
+ NumEdgesRead = ReadCount;
+
+ setBranchWeightMetadata(M, PB);
+
+ return ReadCount > 0;
+}
diff --git a/lib/Analysis/ProfileInfo.cpp b/lib/Analysis/ProfileInfo.cpp
index 173de2c027..b5b7ac1e50 100644
--- a/lib/Analysis/ProfileInfo.cpp
+++ b/lib/Analysis/ProfileInfo.cpp
@@ -1016,40 +1016,14 @@ void ProfileInfoT<Function,BasicBlock>::repair(const Function *F) {
}
}
-raw_ostream& operator<<(raw_ostream &O, const Function *F) {
- return O << F->getName();
-}
-
raw_ostream& operator<<(raw_ostream &O, const MachineFunction *MF) {
return O << MF->getFunction()->getName() << "(MF)";
}
-raw_ostream& operator<<(raw_ostream &O, const BasicBlock *BB) {
- return O << BB->getName();
-}
-
raw_ostream& operator<<(raw_ostream &O, const MachineBasicBlock *MBB) {
return O << MBB->getBasicBlock()->getName() << "(MB)";
}
-raw_ostream& operator<<(raw_ostream &O, std::pair<const BasicBlock *, const BasicBlock *> E) {
- O << "(";
-
- if (E.first)
- O << E.first;
- else
- O << "0";
-
- O << ",";
-
- if (E.second)
- O << E.second;
- else
- O << "0";
-
- return O << ")";
-}
-
raw_ostream& operator<<(raw_ostream &O, std::pair<const MachineBasicBlock *, const MachineBasicBlock *> E) {
O << "(";
diff --git a/test/Analysis/Profiling/load-branch-weights-ifs.ll b/test/Analysis/Profiling/load-branch-weights-ifs.ll
new file mode 100644
index 0000000000..b5ee2f3466
--- /dev/null
+++ b/test/Analysis/Profiling/load-branch-weights-ifs.ll
@@ -0,0 +1,119 @@
+; RUN: opt -insert-edge-profiling -o %t1 < %s
+; RUN: rm -f %t1.prof_data
+; RUN: lli -load %llvmshlibdir/libprofile_rt%shlibext %t1 \
+; RUN: -llvmprof-output %t1.prof_data
+; RUN: opt -profile-file %t1.prof_data -profile-metadata-loader -S -o - < %s \
+; RUN: | FileCheck %s
+; RUN: rm -f %t1.prof_data
+
+;; func_mod - Branch taken 6 times in 7.
+define i32 @func_mod(i32 %N) nounwind uwtable {
+entry:
+ %retval = alloca i32, align 4
+ %N.addr = alloca i32, align 4
+ store i32 %N, i32* %N.addr, align 4
+ %0 = load i32* %N.addr, align 4
+ %rem = srem i32 %0, 7
+ %tobool = icmp ne i32 %rem, 0
+ br i1 %tobool, label %if.then, label %if.else
+; CHECK: br i1 %tobool, label %if.then, label %if.else, !prof !0
+
+if.then:
+ store i32 1, i32* %retval
+ br label %return
+
+if.else:
+ store i32 0, i32* %retval
+ br label %return
+
+return:
+ %1 = load i32* %retval
+ ret i32 %1
+}
+
+;; func_const_true - conditional branch which 100% taken probability.
+define i32 @func_const_true(i32 %N) nounwind uwtable {
+entry:
+ %retval = alloca i32, align 4
+ %N.addr = alloca i32, align 4
+ store i32 %N, i32* %N.addr, align 4
+ %0 = load i32* %N.addr, align 4
+ %cmp = icmp eq i32 %0, 1
+ br i1 %cmp, label %if.then, label %if.end
+; CHECK: br i1 %cmp, label %if.then, label %if.end, !prof !1
+
+if.then:
+ store i32 1, i32* %retval
+ br label %return
+
+if.end:
+ store i32 0, i32* %retval
+ br label %return
+
+return:
+ %1 = load i32* %retval
+ ret i32 %1
+}
+
+;; func_const_true - conditional branch which 100% not-taken probability.
+define i32 @func_const_false(i32 %N) nounwind uwtable {
+entry:
+ %retval = alloca i32, align 4
+ %N.addr = alloca i32, align 4
+ store i32 %N, i32* %N.addr, align 4
+ %0 = load i32* %N.addr, align 4
+ %cmp = icmp eq i32 %0, 1
+ br i1 %cmp, label %if.then, label %if.end
+; CHECK: br i1 %cmp, label %if.then, label %if.end, !prof !2
+
+if.then:
+ store i32 1, i32* %retval
+ br label %return
+
+if.end:
+ store i32 0, i32* %retval
+ br label %return
+
+return:
+ %1 = load i32* %retval
+ ret i32 %1
+}
+
+define i32 @main(i32 %argc, i8** %argv) nounwind uwtable {
+entry:
+ %retval = alloca i32, align 4
+ %argc.addr = alloca i32, align 4
+ %argv.addr = alloca i8**, align 8
+ %loop = alloca i32, align 4
+ store i32 0, i32* %retval
+ store i32 0, i32* %loop, align 4
+ br label %for.cond
+
+for.cond:
+ %0 = load i32* %loop, align 4
+ %cmp = icmp slt i32 %0, 7000
+ br i1 %cmp, label %for.body, label %for.end
+; CHECK: br i1 %cmp, label %for.body, label %for.end, !prof !3
+
+for.body:
+ %1 = load i32* %loop, align 4
+ %call = call i32 @func_mod(i32 %1)
+ br label %for.inc
+
+for.inc:
+ %2 = load i32* %loop, align 4
+ %inc = add nsw i32 %2, 1
+ store i32 %inc, i32* %loop, align 4
+ br label %for.cond
+
+for.end:
+ %call1 = call i32 @func_const_true(i32 1)
+ %call2 = call i32 @func_const_false(i32 0)
+ ret i32 0
+}
+
+; CHECK: !0 = metadata !{metadata !"branch_weights", i32 6000, i32 1000}
+; CHECK: !1 = metadata !{metadata !"branch_weights", i32 1, i32 0}
+; CHECK: !2 = metadata !{metadata !"branch_weights", i32 0, i32 1}
+; CHECK: !3 = metadata !{metadata !"branch_weights", i32 7000, i32 1}
+; CHECK-NOT: !4
diff --git a/test/Analysis/Profiling/load-branch-weights-loops.ll b/test/Analysis/Profiling/load-branch-weights-loops.ll
new file mode 100644
index 0000000000..26c89b7c9b
--- /dev/null
+++ b/test/Analysis/Profiling/load-branch-weights-loops.ll
@@ -0,0 +1,185 @@
+; RUN: opt -insert-edge-profiling -o %t1 < %s
+; RUN: rm -f %t1.prof_data
+; RUN: lli -load %llvmshlibdir/libprofile_rt%shlibext %t1 \
+; RUN: -llvmprof-output %t1.prof_data
+; RUN: opt -profile-file %t1.prof_data -profile-metadata-loader -S -o - < %s \
+; RUN: | FileCheck %s
+; RUN: rm -f %t1.prof_data
+
+;; func_for - Test branch probabilities for a vanilla for loop.
+define i32 @func_for(i32 %N) nounwind uwtable {
+entry:
+ %N.addr = alloca i32, align 4
+ %ret = alloca i32, align 4
+ %loop = alloca i32, align 4
+ store i32 %N, i32* %N.addr, align 4
+ store i32 0, i32* %ret, align 4
+ store i32 0, i32* %loop, align 4
+ br label %for.cond
+
+for.cond:
+ %0 = load i32* %loop, align 4
+ %1 = load i32* %N.addr, align 4
+ %cmp = icmp slt i32 %0, %1
+ br i1 %cmp, label %for.body, label %for.end
+; CHECK: br i1 %cmp, label %for.body, label %for.end, !prof !0
+
+for.body:
+ %2 = load i32* %N.addr, align 4
+ %3 = load i32* %ret, align 4
+ %add = add nsw i32 %3, %2
+ store i32 %add, i32* %ret, align 4
+ br label %for.inc
+
+for.inc:
+ %4 = load i32* %loop, align 4
+ %inc = add nsw i32 %4, 1
+ store i32 %inc, i32* %loop, align 4
+ br label %for.cond
+
+for.end:
+ %5 = load i32* %ret, align 4
+ ret i32 %5
+}
+
+;; func_for_odd - Test branch probabilities for a for loop with a continue and
+;; a break.
+define i32 @func_for_odd(i32 %N) nounwind uwtable {
+entry:
+ %N.addr = alloca i32, align 4
+ %ret = alloca i32, align 4
+ %loop = alloca i32, align 4
+ store i32 %N, i32* %N.addr, align 4
+ store i32 0, i32* %ret, align 4
+ store i32 0, i32* %loop, align 4
+ br label %for.cond
+
+for.cond:
+ %0 = load i32* %loop, align 4
+ %1 = load i32* %N.addr, align 4
+ %cmp = icmp slt i32 %0, %1
+ br i1 %cmp, label %for.body, label %for.end
+; CHECK: br i1 %cmp, label %for.body, label %for.end, !prof !1
+
+for.body:
+ %2 = load i32* %loop, align 4
+ %rem = srem i32 %2, 10
+ %tobool = icmp ne i32 %rem, 0
+ br i1 %tobool, label %if.then, label %if.end
+; CHECK: br i1 %tobool, label %if.then, label %if.end, !prof !2
+
+if.then:
+ br label %for.inc
+
+if.end:
+ %3 = load i32* %loop, align 4
+ %cmp1 = icmp eq i32 %3, 500
+ br i1 %cmp1, label %if.then2, label %if.end3
+; CHECK: br i1 %cmp1, label %if.then2, label %if.end3, !prof !3
+
+if.then2:
+ br label %for.end
+
+if.end3:
+ %4 = load i32* %N.addr, align 4
+ %5 = load i32* %ret, align 4
+ %add = add nsw i32 %5, %4
+ store i32 %add, i32* %ret, align 4
+ br label %for.inc
+
+for.inc:
+ %6 = load i32* %loop, align 4
+ %inc = add nsw i32 %6, 1
+ store i32 %inc, i32* %loop, align 4
+ br label %for.cond
+
+for.end:
+ %7 = load i32* %ret, align 4
+ ret i32 %7
+}
+
+;; func_while - Test branch probability in a vanilla while loop.
+define i32 @func_while(i32 %N) nounwind uwtable {
+entry:
+ %N.addr = alloca i32, align 4
+ %ret = alloca i32, align 4
+ %loop = alloca i32, align 4
+ store i32 %N, i32* %N.addr, align 4
+ store i32 0, i32* %ret, align 4
+ store i32 0, i32* %loop, align 4
+ br label %while.cond
+
+while.cond:
+ %0 = load i32* %loop, align 4
+ %1 = load i32* %N.addr, align 4
+ %cmp = icmp slt i32 %0, %1
+ br i1 %cmp, label %while.body, label %while.end
+; CHECK: br i1 %cmp, label %while.body, label %while.end, !prof !0
+
+while.body:
+ %2 = load i32* %N.addr, align 4
+ %3 = load i32* %ret, align 4
+ %add = add nsw i32 %3, %2
+ store i32 %add, i32* %ret, align 4
+ %4 = load i32* %loop, align 4
+ %inc = add nsw i32 %4, 1
+ store i32 %inc, i32* %loop, align 4
+ br label %while.cond
+
+while.end:
+ %5 = load i32* %ret, align 4
+ ret i32 %5
+}
+
+;; func_while - Test branch probability in a vanilla do-while loop.
+define i32 @func_do_while(i32 %N) nounwind uwtable {
+entry:
+ %N.addr = alloca i32, align 4
+ %ret = alloca i32, align 4
+ %loop = alloca i32, align 4
+ store i32 %N, i32* %N.addr, align 4
+ store i32 0, i32* %ret, align 4
+ store i32 0, i32* %loop, align 4
+ br label %do.body
+
+do.body:
+ %0 = load i32* %N.addr, align 4
+ %1 = load i32* %ret, align 4
+ %add = add nsw i32 %1, %0
+ store i32 %add, i32* %ret, align 4
+ %2 = load i32* %loop, align 4
+ %inc = add nsw i32 %2, 1
+ store i32 %inc, i32* %loop, align 4
+ br label %do.cond
+
+do.cond:
+ %3 = load i32* %loop, align 4
+ %4 = load i32* %N.addr, align 4
+ %cmp = icmp slt i32 %3, %4
+ br i1 %cmp, label %do.body, label %do.end
+; CHECK: br i1 %cmp, label %do.body, label %do.end, !prof !4
+
+do.end:
+ %5 = load i32* %ret, align 4
+ ret i32 %5
+}
+
+define i32 @main(i32 %argc, i8** %argv) nounwind uwtable {
+entry:
+ %retval = alloca i32, align 4
+ %argc.addr = alloca i32, align 4
+ %argv.addr = alloca i8**, align 8
+ store i32 0, i32* %retval
+ %call = call i32 @func_for(i32 1000)
+ %call1 = call i32 @func_for_odd(i32 1000)
+ %call2 = call i32 @func_while(i32 1000)
+ %call3 = call i32 @func_do_while(i32 1000)
+ ret i32 0
+}
+
+!0 = metadata !{metadata !"branch_weights", i32 1000, i32 1}
+!1 = metadata !{metadata !"branch_weights", i32 501, i32 0}
+!2 = metadata !{metadata !"branch_weights", i32 450, i32 51}
+!3 = metadata !{metadata !"branch_weights", i32 1, i32 50}
+!4 = metadata !{metadata !"branch_weights", i32 999, i32 1}
+; CHECK-NOT: !5
diff --git a/test/Analysis/Profiling/load-branch-weights-switches.ll b/test/Analysis/Profiling/load-branch-weights-switches.ll
new file mode 100644
index 0000000000..9efe6d0fe4
--- /dev/null
+++ b/test/Analysis/Profiling/load-branch-weights-switches.ll
@@ -0,0 +1,162 @@
+; RUN: opt -insert-edge-profiling -o %t1 < %s
+; RUN: rm -f %t1.prof_data
+; RUN: lli -load %llvmshlibdir/libprofile_rt%shlibext %t1 \
+; RUN: -llvmprof-output %t1.prof_data
+; RUN: opt -profile-file %t1.prof_data -profile-metadata-loader -S -o - < %s \
+; RUN: | FileCheck %s
+; RUN: rm -f %t1.prof_data
+
+;; func_switch - Test branch probabilities for a switch instruction with an
+;; even chance of taking each case (or no case).
+define i32 @func_switch(i32 %N) nounwind uwtable {
+entry:
+ %retval = alloca i32, align 4
+ %N.addr = alloca i32, align 4
+ store i32 %N, i32* %N.addr, align 4
+ %0 = load i32* %N.addr, align 4
+ %rem = srem i32 %0, 4
+ switch i32 %rem, label %sw.epilog [
+ i32 0, label %sw.bb
+ i32 1, label %sw.bb1
+ i32 2, label %sw.bb2
+ ]
+; CHECK: ], !prof !0
+
+sw.bb:
+ store i32 5, i32* %retval
+ br label %return
+
+sw.bb1:
+ store i32 6, i32* %retval
+ br label %return
+
+sw.bb2:
+ store i32 7, i32* %retval
+ br label %return
+
+sw.epilog:
+ store i32 8, i32* %retval
+ br label %return
+
+return:
+ %1 = load i32* %retval
+ ret i32 %1
+}
+
+;; func_switch_switch - Test branch probabilities in a switch-instruction that
+;; leads to further switch instructions. The first-tier switch occludes some
+;; possibilities in the second-tier switches, leading to some branches having a
+;; 0 probability.
+define i32 @func_switch_switch(i32 %N) nounwind uwtable {
+entry:
+ %retval = alloca i32, align 4
+ %N.addr = alloca i32, align 4
+ store i32 %N, i32* %N.addr, align 4
+ %0 = load i32* %N.addr, align 4
+ %rem = srem i32 %0, 2
+ switch i32 %rem, label %sw.default11 [
+ i32 0, label %sw.bb
+ i32 1, label %sw.bb5
+ ]
+; CHECK: ], !prof !1
+
+sw.bb:
+ %1 = load i32* %N.addr, align 4
+ %rem1 = srem i32 %1, 4
+ switch i32 %rem1, label %sw.default [
+ i32 0, label %sw.bb2
+ i32 1, label %sw.bb3
+ i32 2, label %sw.bb4
+ ]
+; CHECK: ], !prof !2
+
+sw.bb2:
+ store i32 5, i32* %retval
+ br label %return
+
+sw.bb3:
+ store i32 6, i32* %retval
+ br label %return
+
+sw.bb4:
+ store i32 7, i32* %retval
+ br label %return
+
+sw.default:
+ store i32 8, i32* %retval
+ br label %return
+
+sw.bb5:
+ %2 = load i32* %N.addr, align 4
+ %rem6 = srem i32 %2, 4
+ switch i32 %rem6, label %sw.default10 [
+ i32 0, label %sw.bb7
+ i32 1, label %sw.bb8
+ i32 2, label %sw.bb9
+ ]
+; CHECK: ], !prof !3
+
+sw.bb7:
+ store i32 9, i32* %retval
+ br label %return
+
+sw.bb8:
+ store i32 10, i32* %retval
+ br label %return
+
+sw.bb9:
+ store i32 11, i32* %retval
+ br label %return
+
+sw.default10:
+ store i32 12, i32* %retval
+ br label %return
+
+sw.default11:
+ store i32 13, i32* %retval
+ br label %return
+
+return:
+ %3 = load i32* %retval
+ ret i32 %3
+}
+
+define i32 @main(i32 %argc, i8** %argv) nounwind uwtable {
+entry:
+ %retval = alloca i32, align 4
+ %argc.addr = alloca i32, align 4
+ %argv.addr = alloca i8**, align 8
+ %loop = alloca i32, align 4
+ store i32 0, i32* %retval
+ store i32 0, i32* %loop, align 4
+ br label %for.cond
+
+for.cond:
+ %0 = load i32* %loop, align 4
+ %cmp = icmp slt i32 %0, 4000
+ br i1 %cmp, label %for.body, label %for.end
+; CHECK: br i1 %cmp, label %for.body, label %for.end, !prof !4
+
+for.body:
+ %1 = load i32* %loop, align 4
+ %call = call i32 @func_switch(i32 %1)
+ %2 = load i32* %loop, align 4
+ %call1 = call i32 @func_switch_switch(i32 %2)
+ br label %for.inc
+
+for.inc:
+ %3 = load i32* %loop, align 4
+ %inc = add nsw i32 %3, 1
+ store i32 %inc, i32* %loop, align 4
+ br label %for.cond
+
+for.end:
+ ret i32 0
+}
+
+; CHECK: !0 = metadata !{metadata !"branch_weights", i32 1000, i32 1000, i32 1000, i32 1000}
+; CHECK: !1 = metadata !{metadata !"branch_weights", i32 0, i32 2000, i32 2000}
+; CHECK: !2 = metadata !{metadata !"branch_weights", i32 0, i32 1000, i32 0, i32 1000}
+; CHECK: !3 = metadata !{metadata !"branch_weights", i32 1000, i32 0, i32 1000, i32 0}
+; CHECK: !4 = metadata !{metadata !"branch_weights", i32 4000, i32 1}
+; CHECK-NOT: !5