llvm-project/llvm/lib/ProfileData/SampleProf.cpp
William Huang 7624de5bea [llvm-profdata] Refactoring Sample Profile Reader to increase FDO build speed using MD5 as key to Sample Profile map
This is phase 1 of multiple planned improvements on the sample profile loader.   The major change is to use MD5 hash code ((instead of the function itself) as the key to look up the function offset table and the profiles, which significantly reduce the time it takes to construct the map.

The optimization is based on the fact that many practical sample profiles are using MD5 values for function names to reduce profile size, so we shouldn't need to convert the MD5 to a string and then to a SampleContext and use it as the map's key, because it's extremely slow.

Several changes to note:

(1) For non-CS SampleContext, if it is already MD5 string, the hash value will be its integral value, instead of hashing the MD5 again. In phase 2 this is going to be optimized further using a union to represent MD5 function (without converting it to string) and regular function names.

(2) The SampleProfileMap is a wrapper to *map<uint64_t, FunctionSamples>, while providing interface allowing using SampleContext as key, so that existing code still work. It will check for MD5 collision (unlikely but not too unlikely, since we only takes the lower 64 bits) and handle it to at least guarantee compilation correctness (conflicting old profile is dropped, instead of returning an old profile with inconsistent context). Other code should not try to use MD5 as key to access the map directly, because it will not be able to handle MD5 collision at all. (see exception at (5) )

(3) Any SampleProfileMap::emplace() followed by SampleContext assignment if newly inserted, should be replaced with SampleProfileMap::Create(), which does the same thing.

(4) Previously we ensure an invariant that in SampleProfileMap, the key is equal to the Context of the value, for profile map that is eventually being used for output (as in llvm-profdata/llvm-profgen). Since the key became MD5 hash, only the value keeps the context now, in several places where an intermediate SampleProfileMap is created, each new FunctionSample's context is set immediately after insertion, which is necessary to "remember" the context otherwise irretrievable.

(5) When reading a profile, we cache the MD5 values of all functions, because they are used at least twice (one to index into FuncOffsetTable, the other into SampleProfileMap, more if there are additional sections), in this case the SampleProfileMap is directly accessed with MD5 value so that we don't recalculate it each time (expensive)

Performance impact:
When reading a ~1GB extbinary profile (fixed length MD5, not compressed) with 10 million function names and 2.5 million top level functions (non CS functions, each function has varying nesting level from 0 to 20), this patch improves the function offset table loading time by 20%, and improves full profile read by 5%.

Reviewed By: davidxl, snehasish

Differential Revision: https://reviews.llvm.org/D147740
2023-08-17 20:10:45 +00:00

514 lines
19 KiB
C++

//=-- SampleProf.cpp - Sample profiling format support --------------------===//
//
// 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 contains common definitions used in the reading and writing of
// sample profile data.
//
//===----------------------------------------------------------------------===//
#include "llvm/ProfileData/SampleProf.h"
#include "llvm/Config/llvm-config.h"
#include "llvm/IR/DebugInfoMetadata.h"
#include "llvm/IR/PseudoProbe.h"
#include "llvm/ProfileData/SampleProfReader.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/raw_ostream.h"
#include <string>
#include <system_error>
using namespace llvm;
using namespace sampleprof;
static cl::opt<uint64_t> ProfileSymbolListCutOff(
"profile-symbol-list-cutoff", cl::Hidden, cl::init(-1),
cl::desc("Cutoff value about how many symbols in profile symbol list "
"will be used. This is very useful for performance debugging"));
static cl::opt<bool> GenerateMergedBaseProfiles(
"generate-merged-base-profiles",
cl::desc("When generating nested context-sensitive profiles, always "
"generate extra base profile for function with all its context "
"profiles merged into it."));
namespace llvm {
namespace sampleprof {
bool FunctionSamples::ProfileIsProbeBased = false;
bool FunctionSamples::ProfileIsCS = false;
bool FunctionSamples::ProfileIsPreInlined = false;
bool FunctionSamples::UseMD5 = false;
bool FunctionSamples::HasUniqSuffix = true;
bool FunctionSamples::ProfileIsFS = false;
} // namespace sampleprof
} // namespace llvm
namespace {
// FIXME: This class is only here to support the transition to llvm::Error. It
// will be removed once this transition is complete. Clients should prefer to
// deal with the Error value directly, rather than converting to error_code.
class SampleProfErrorCategoryType : public std::error_category {
const char *name() const noexcept override { return "llvm.sampleprof"; }
std::string message(int IE) const override {
sampleprof_error E = static_cast<sampleprof_error>(IE);
switch (E) {
case sampleprof_error::success:
return "Success";
case sampleprof_error::bad_magic:
return "Invalid sample profile data (bad magic)";
case sampleprof_error::unsupported_version:
return "Unsupported sample profile format version";
case sampleprof_error::too_large:
return "Too much profile data";
case sampleprof_error::truncated:
return "Truncated profile data";
case sampleprof_error::malformed:
return "Malformed sample profile data";
case sampleprof_error::unrecognized_format:
return "Unrecognized sample profile encoding format";
case sampleprof_error::unsupported_writing_format:
return "Profile encoding format unsupported for writing operations";
case sampleprof_error::truncated_name_table:
return "Truncated function name table";
case sampleprof_error::not_implemented:
return "Unimplemented feature";
case sampleprof_error::counter_overflow:
return "Counter overflow";
case sampleprof_error::ostream_seek_unsupported:
return "Ostream does not support seek";
case sampleprof_error::uncompress_failed:
return "Uncompress failure";
case sampleprof_error::zlib_unavailable:
return "Zlib is unavailable";
case sampleprof_error::hash_mismatch:
return "Function hash mismatch";
}
llvm_unreachable("A value of sampleprof_error has no message.");
}
};
} // end anonymous namespace
const std::error_category &llvm::sampleprof_category() {
static SampleProfErrorCategoryType ErrorCategory;
return ErrorCategory;
}
void LineLocation::print(raw_ostream &OS) const {
OS << LineOffset;
if (Discriminator > 0)
OS << "." << Discriminator;
}
raw_ostream &llvm::sampleprof::operator<<(raw_ostream &OS,
const LineLocation &Loc) {
Loc.print(OS);
return OS;
}
/// Merge the samples in \p Other into this record.
/// Optionally scale sample counts by \p Weight.
sampleprof_error SampleRecord::merge(const SampleRecord &Other,
uint64_t Weight) {
sampleprof_error Result;
Result = addSamples(Other.getSamples(), Weight);
for (const auto &I : Other.getCallTargets()) {
MergeResult(Result, addCalledTarget(I.first(), I.second, Weight));
}
return Result;
}
#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
LLVM_DUMP_METHOD void LineLocation::dump() const { print(dbgs()); }
#endif
/// Print the sample record to the stream \p OS indented by \p Indent.
void SampleRecord::print(raw_ostream &OS, unsigned Indent) const {
OS << NumSamples;
if (hasCalls()) {
OS << ", calls:";
for (const auto &I : getSortedCallTargets())
OS << " " << I.first << ":" << I.second;
}
OS << "\n";
}
#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
LLVM_DUMP_METHOD void SampleRecord::dump() const { print(dbgs(), 0); }
#endif
raw_ostream &llvm::sampleprof::operator<<(raw_ostream &OS,
const SampleRecord &Sample) {
Sample.print(OS, 0);
return OS;
}
/// Print the samples collected for a function on stream \p OS.
void FunctionSamples::print(raw_ostream &OS, unsigned Indent) const {
if (getFunctionHash())
OS << "CFG checksum " << getFunctionHash() << "\n";
OS << TotalSamples << ", " << TotalHeadSamples << ", " << BodySamples.size()
<< " sampled lines\n";
OS.indent(Indent);
if (!BodySamples.empty()) {
OS << "Samples collected in the function's body {\n";
SampleSorter<LineLocation, SampleRecord> SortedBodySamples(BodySamples);
for (const auto &SI : SortedBodySamples.get()) {
OS.indent(Indent + 2);
OS << SI->first << ": " << SI->second;
}
OS.indent(Indent);
OS << "}\n";
} else {
OS << "No samples collected in the function's body\n";
}
OS.indent(Indent);
if (!CallsiteSamples.empty()) {
OS << "Samples collected in inlined callsites {\n";
SampleSorter<LineLocation, FunctionSamplesMap> SortedCallsiteSamples(
CallsiteSamples);
for (const auto &CS : SortedCallsiteSamples.get()) {
for (const auto &FS : CS->second) {
OS.indent(Indent + 2);
OS << CS->first << ": inlined callee: " << FS.second.getName() << ": ";
FS.second.print(OS, Indent + 4);
}
}
OS.indent(Indent);
OS << "}\n";
} else {
OS << "No inlined callsites in this function\n";
}
}
raw_ostream &llvm::sampleprof::operator<<(raw_ostream &OS,
const FunctionSamples &FS) {
FS.print(OS);
return OS;
}
void sampleprof::sortFuncProfiles(
const SampleProfileMap &ProfileMap,
std::vector<NameFunctionSamples> &SortedProfiles) {
for (const auto &I : ProfileMap) {
SortedProfiles.push_back(std::make_pair(I.first, &I.second));
}
llvm::stable_sort(SortedProfiles, [](const NameFunctionSamples &A,
const NameFunctionSamples &B) {
if (A.second->getTotalSamples() == B.second->getTotalSamples())
return A.second->getContext() < B.second->getContext();
return A.second->getTotalSamples() > B.second->getTotalSamples();
});
}
unsigned FunctionSamples::getOffset(const DILocation *DIL) {
return (DIL->getLine() - DIL->getScope()->getSubprogram()->getLine()) &
0xffff;
}
LineLocation FunctionSamples::getCallSiteIdentifier(const DILocation *DIL,
bool ProfileIsFS) {
if (FunctionSamples::ProfileIsProbeBased) {
// In a pseudo-probe based profile, a callsite is simply represented by the
// ID of the probe associated with the call instruction. The probe ID is
// encoded in the Discriminator field of the call instruction's debug
// metadata.
return LineLocation(PseudoProbeDwarfDiscriminator::extractProbeIndex(
DIL->getDiscriminator()),
0);
} else {
unsigned Discriminator =
ProfileIsFS ? DIL->getDiscriminator() : DIL->getBaseDiscriminator();
return LineLocation(FunctionSamples::getOffset(DIL), Discriminator);
}
}
uint64_t FunctionSamples::getCallSiteHash(StringRef CalleeName,
const LineLocation &Callsite) {
uint64_t NameHash = std::hash<std::string>{}(CalleeName.str());
uint64_t LocId =
(((uint64_t)Callsite.LineOffset) << 32) | Callsite.Discriminator;
return NameHash + (LocId << 5) + LocId;
}
const FunctionSamples *FunctionSamples::findFunctionSamples(
const DILocation *DIL, SampleProfileReaderItaniumRemapper *Remapper) const {
assert(DIL);
SmallVector<std::pair<LineLocation, StringRef>, 10> S;
const DILocation *PrevDIL = DIL;
for (DIL = DIL->getInlinedAt(); DIL; DIL = DIL->getInlinedAt()) {
// Use C++ linkage name if possible.
StringRef Name = PrevDIL->getScope()->getSubprogram()->getLinkageName();
if (Name.empty())
Name = PrevDIL->getScope()->getSubprogram()->getName();
S.emplace_back(FunctionSamples::getCallSiteIdentifier(
DIL, FunctionSamples::ProfileIsFS),
Name);
PrevDIL = DIL;
}
if (S.size() == 0)
return this;
const FunctionSamples *FS = this;
for (int i = S.size() - 1; i >= 0 && FS != nullptr; i--) {
FS = FS->findFunctionSamplesAt(S[i].first, S[i].second, Remapper);
}
return FS;
}
void FunctionSamples::findAllNames(DenseSet<StringRef> &NameSet) const {
NameSet.insert(getName());
for (const auto &BS : BodySamples)
for (const auto &TS : BS.second.getCallTargets())
NameSet.insert(TS.getKey());
for (const auto &CS : CallsiteSamples) {
for (const auto &NameFS : CS.second) {
NameSet.insert(NameFS.first);
NameFS.second.findAllNames(NameSet);
}
}
}
const FunctionSamples *FunctionSamples::findFunctionSamplesAt(
const LineLocation &Loc, StringRef CalleeName,
SampleProfileReaderItaniumRemapper *Remapper) const {
CalleeName = getCanonicalFnName(CalleeName);
std::string CalleeGUID;
CalleeName = getRepInFormat(CalleeName, UseMD5, CalleeGUID);
auto iter = CallsiteSamples.find(mapIRLocToProfileLoc(Loc));
if (iter == CallsiteSamples.end())
return nullptr;
auto FS = iter->second.find(CalleeName);
if (FS != iter->second.end())
return &FS->second;
if (Remapper) {
if (auto NameInProfile = Remapper->lookUpNameInProfile(CalleeName)) {
auto FS = iter->second.find(*NameInProfile);
if (FS != iter->second.end())
return &FS->second;
}
}
// If we cannot find exact match of the callee name, return the FS with
// the max total count. Only do this when CalleeName is not provided,
// i.e., only for indirect calls.
if (!CalleeName.empty())
return nullptr;
uint64_t MaxTotalSamples = 0;
const FunctionSamples *R = nullptr;
for (const auto &NameFS : iter->second)
if (NameFS.second.getTotalSamples() >= MaxTotalSamples) {
MaxTotalSamples = NameFS.second.getTotalSamples();
R = &NameFS.second;
}
return R;
}
#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
LLVM_DUMP_METHOD void FunctionSamples::dump() const { print(dbgs(), 0); }
#endif
std::error_code ProfileSymbolList::read(const uint8_t *Data,
uint64_t ListSize) {
const char *ListStart = reinterpret_cast<const char *>(Data);
uint64_t Size = 0;
uint64_t StrNum = 0;
while (Size < ListSize && StrNum < ProfileSymbolListCutOff) {
StringRef Str(ListStart + Size);
add(Str);
Size += Str.size() + 1;
StrNum++;
}
if (Size != ListSize && StrNum != ProfileSymbolListCutOff)
return sampleprof_error::malformed;
return sampleprof_error::success;
}
void SampleContextTrimmer::trimAndMergeColdContextProfiles(
uint64_t ColdCountThreshold, bool TrimColdContext, bool MergeColdContext,
uint32_t ColdContextFrameLength, bool TrimBaseProfileOnly) {
if (!TrimColdContext && !MergeColdContext)
return;
// Nothing to merge if sample threshold is zero
if (ColdCountThreshold == 0)
return;
// Trimming base profiles only is mainly to honor the preinliner decsion. When
// MergeColdContext is true preinliner decsion is not honored anyway so turn
// off TrimBaseProfileOnly.
if (MergeColdContext)
TrimBaseProfileOnly = false;
// Filter the cold profiles from ProfileMap and move them into a tmp
// container
std::vector<std::pair<hash_code, const FunctionSamples *>> ColdProfiles;
for (const auto &I : ProfileMap) {
const SampleContext &Context = I.second.getContext();
const FunctionSamples &FunctionProfile = I.second;
if (FunctionProfile.getTotalSamples() < ColdCountThreshold &&
(!TrimBaseProfileOnly || Context.isBaseContext()))
ColdProfiles.emplace_back(I.first, &I.second);
}
// Remove the cold profile from ProfileMap and merge them into
// MergedProfileMap by the last K frames of context
SampleProfileMap MergedProfileMap;
for (const auto &I : ColdProfiles) {
if (MergeColdContext) {
auto MergedContext = I.second->getContext().getContextFrames();
if (ColdContextFrameLength < MergedContext.size())
MergedContext = MergedContext.take_back(ColdContextFrameLength);
// Need to set MergedProfile's context here otherwise it will be lost.
FunctionSamples &MergedProfile = MergedProfileMap.Create(MergedContext);
MergedProfile.merge(*I.second);
}
ProfileMap.erase(I.first);
}
// Move the merged profiles into ProfileMap;
for (const auto &I : MergedProfileMap) {
// Filter the cold merged profile
if (TrimColdContext && I.second.getTotalSamples() < ColdCountThreshold &&
ProfileMap.find(I.second.getContext()) == ProfileMap.end())
continue;
// Merge the profile if the original profile exists, otherwise just insert
// as a new profile. If inserted as a new profile from MergedProfileMap, it
// already has the right context.
auto Ret = ProfileMap.emplace(I.second.getContext(), FunctionSamples());
FunctionSamples &OrigProfile = Ret.first->second;
OrigProfile.merge(I.second);
}
}
std::error_code ProfileSymbolList::write(raw_ostream &OS) {
// Sort the symbols before output. If doing compression.
// It will make the compression much more effective.
std::vector<StringRef> SortedList(Syms.begin(), Syms.end());
llvm::sort(SortedList);
std::string OutputString;
for (auto &Sym : SortedList) {
OutputString.append(Sym.str());
OutputString.append(1, '\0');
}
OS << OutputString;
return sampleprof_error::success;
}
void ProfileSymbolList::dump(raw_ostream &OS) const {
OS << "======== Dump profile symbol list ========\n";
std::vector<StringRef> SortedList(Syms.begin(), Syms.end());
llvm::sort(SortedList);
for (auto &Sym : SortedList)
OS << Sym << "\n";
}
ProfileConverter::FrameNode *
ProfileConverter::FrameNode::getOrCreateChildFrame(const LineLocation &CallSite,
StringRef CalleeName) {
uint64_t Hash = FunctionSamples::getCallSiteHash(CalleeName, CallSite);
auto It = AllChildFrames.find(Hash);
if (It != AllChildFrames.end()) {
assert(It->second.FuncName == CalleeName &&
"Hash collision for child context node");
return &It->second;
}
AllChildFrames[Hash] = FrameNode(CalleeName, nullptr, CallSite);
return &AllChildFrames[Hash];
}
ProfileConverter::ProfileConverter(SampleProfileMap &Profiles)
: ProfileMap(Profiles) {
for (auto &FuncSample : Profiles) {
FunctionSamples *FSamples = &FuncSample.second;
auto *NewNode = getOrCreateContextPath(FSamples->getContext());
assert(!NewNode->FuncSamples && "New node cannot have sample profile");
NewNode->FuncSamples = FSamples;
}
}
ProfileConverter::FrameNode *
ProfileConverter::getOrCreateContextPath(const SampleContext &Context) {
auto Node = &RootFrame;
LineLocation CallSiteLoc(0, 0);
for (auto &Callsite : Context.getContextFrames()) {
Node = Node->getOrCreateChildFrame(CallSiteLoc, Callsite.FuncName);
CallSiteLoc = Callsite.Location;
}
return Node;
}
void ProfileConverter::convertCSProfiles(ProfileConverter::FrameNode &Node) {
// Process each child profile. Add each child profile to callsite profile map
// of the current node `Node` if `Node` comes with a profile. Otherwise
// promote the child profile to a standalone profile.
auto *NodeProfile = Node.FuncSamples;
for (auto &It : Node.AllChildFrames) {
auto &ChildNode = It.second;
convertCSProfiles(ChildNode);
auto *ChildProfile = ChildNode.FuncSamples;
if (!ChildProfile)
continue;
SampleContext OrigChildContext = ChildProfile->getContext();
hash_code OrigChildContextHash = OrigChildContext.getHashCode();
// Reset the child context to be contextless.
ChildProfile->getContext().setName(OrigChildContext.getName());
if (NodeProfile) {
// Add child profile to the callsite profile map.
auto &SamplesMap = NodeProfile->functionSamplesAt(ChildNode.CallSiteLoc);
SamplesMap.emplace(OrigChildContext.getName().str(), *ChildProfile);
NodeProfile->addTotalSamples(ChildProfile->getTotalSamples());
// Remove the corresponding body sample for the callsite and update the
// total weight.
auto Count = NodeProfile->removeCalledTargetAndBodySample(
ChildNode.CallSiteLoc.LineOffset, ChildNode.CallSiteLoc.Discriminator,
OrigChildContext.getName());
NodeProfile->removeTotalSamples(Count);
}
hash_code NewChildProfileHash(0);
// Separate child profile to be a standalone profile, if the current parent
// profile doesn't exist. This is a duplicating operation when the child
// profile is already incorporated into the parent which is still useful and
// thus done optionally. It is seen that duplicating context profiles into
// base profiles improves the code quality for thinlto build by allowing a
// profile in the prelink phase for to-be-fully-inlined functions.
if (!NodeProfile) {
ProfileMap[ChildProfile->getContext()].merge(*ChildProfile);
NewChildProfileHash = ChildProfile->getContext().getHashCode();
} else if (GenerateMergedBaseProfiles) {
ProfileMap[ChildProfile->getContext()].merge(*ChildProfile);
NewChildProfileHash = ChildProfile->getContext().getHashCode();
auto &SamplesMap = NodeProfile->functionSamplesAt(ChildNode.CallSiteLoc);
SamplesMap[ChildProfile->getName().str()].getContext().setAttribute(
ContextDuplicatedIntoBase);
}
// Remove the original child profile. Check if MD5 of new child profile
// collides with old profile, in this case the [] operator already
// overwritten it without the need of erase.
if (NewChildProfileHash != OrigChildContextHash)
ProfileMap.erase(OrigChildContextHash);
}
}
void ProfileConverter::convertCSProfiles() { convertCSProfiles(RootFrame); }