[Reland] [CGData] Lazy loading support for stable function map #154491
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This is an attempt to reland llvm#151660 by including a missing STL header found by a buildbot failure. The stable function map could be huge for a large application. Fully loading it is slow and consumes a significant amount of memory, which is unnecessary and drastically slows down compilation especially for non-LTO and distributed-ThinLTO setups. This patch introduces an opt-in lazy loading support for the stable function map. The detailed changes are: - `StableFunctionMap` - The map now stores entries in an `EntryStorage` struct, which includes offsets for serialized entries and a `std::once_flag` for thread-safe lazy loading. - The underlying map type is changed from `DenseMap` to `std::unordered_map` for compatibility with `std::once_flag`. - `contains()`, `size()` and `at()` are implemented to only load requested entries on demand. - Lazy Loading Mechanism - When reading indexed codegen data, if the newly-introduced `-indexed-codegen-data-lazy-loading` flag is set, the stable function map is not fully deserialized up front. The binary format for the stable function map now includes offsets and sizes to support lazy loading. - The safety of lazy loading is guarded by the once flag per function hash. This guarantees that even in a multi-threaded environment, the deserialization for a given function hash will happen exactly once. The first thread to request it performs the load, and subsequent threads will wait for it to complete before using the data. For single-threaded builds, the overhead is negligible (a single check on the once flag). For multi-threaded scenarios, users can omit the flag to retain the previous eager-loading behavior.
llvmbot
added
llvm:codegen
LTO
|
@llvm/pr-subscribers-lto Author: Zhaoxuan Jiang (nocchijiang) ChangesThis is an attempt to reland #151660 by including a missing STL header found by a buildbot failure. The stable function map could be huge for a large application. Fully loading it is slow and consumes a significant amount of memory, which is unnecessary and drastically slows down compilation especially for non-LTO and distributed-ThinLTO setups. This patch introduces an opt-in lazy loading support for the stable function map. The detailed changes are:
Patch is 38.85 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/154491.diff 20 Files Affected:
diff --git a/llvm/include/llvm/CGData/CodeGenData.h b/llvm/include/llvm/CGData/CodeGenData.h
index 38b96b72ccac6..e44497a408245 100644
--- a/llvm/include/llvm/CGData/CodeGenData.h
+++ b/llvm/include/llvm/CGData/CodeGenData.h
@@ -285,6 +285,9 @@ enum CGDataVersion {
// Version 3 adds the total size of the Names in the stable function map so
// we can skip reading them into the memory for non-assertion builds.
Version3 = 3,
+ // Version 4 adjusts the structure of stable function merging map for
+ // efficient lazy loading support.
+ Version4 = 4,
CurrentVersion = CG_DATA_INDEX_VERSION
};
const uint64_t Version = CGDataVersion::CurrentVersion;
diff --git a/llvm/include/llvm/CGData/CodeGenData.inc b/llvm/include/llvm/CGData/CodeGenData.inc
index 94de4c0b017a2..d5fbe2fb97718 100644
--- a/llvm/include/llvm/CGData/CodeGenData.inc
+++ b/llvm/include/llvm/CGData/CodeGenData.inc
@@ -49,4 +49,4 @@ CG_DATA_SECT_ENTRY(CG_merge, CG_DATA_QUOTE(CG_DATA_MERGE_COMMON),
#endif
/* Indexed codegen data format version (start from 1). */
-#define CG_DATA_INDEX_VERSION 3
+#define CG_DATA_INDEX_VERSION 4
diff --git a/llvm/include/llvm/CGData/StableFunctionMap.h b/llvm/include/llvm/CGData/StableFunctionMap.h
index bcb72e8216973..8c7e83f4da1b1 100644
--- a/llvm/include/llvm/CGData/StableFunctionMap.h
+++ b/llvm/include/llvm/CGData/StableFunctionMap.h
@@ -20,6 +20,9 @@
#include "llvm/ADT/StringMap.h"
#include "llvm/IR/StructuralHash.h"
#include "llvm/Support/Compiler.h"
+#include "llvm/Support/MemoryBuffer.h"
+#include <mutex>
+#include <unordered_map>
namespace llvm {
@@ -72,11 +75,37 @@ struct StableFunctionMap {
IndexOperandHashMap(std::move(IndexOperandHashMap)) {}
};
- using HashFuncsMapType =
- DenseMap<stable_hash, SmallVector<std::unique_ptr<StableFunctionEntry>>>;
+ using StableFunctionEntries =
+ SmallVector<std::unique_ptr<StableFunctionEntry>>;
+
+ /// In addition to the deserialized StableFunctionEntry, the struct stores
+ /// the offsets of corresponding serialized stable function entries, and a
+ /// once flag for safe lazy loading in a multithreaded environment.
+ struct EntryStorage {
+ /// The actual storage of deserialized stable function entries. If the map
+ /// is lazily loaded, this will be empty until the first access by the
+ /// corresponding function hash.
+ StableFunctionEntries Entries;
+
+ private:
+ /// This is used to deserialize the entry lazily. Each element is the
+ /// corresponding serialized stable function entry's offset in the memory
+ /// buffer (StableFunctionMap::Buffer).
+ /// The offsets are only populated when loading the map lazily, otherwise
+ /// it is empty.
+ SmallVector<uint64_t> Offsets;
+ std::once_flag LazyLoadFlag;
+ friend struct StableFunctionMap;
+ friend struct StableFunctionMapRecord;
+ };
+
+ // Note: DenseMap requires value type to be copyable even if only using
+ // in-place insertion. Use STL instead. This also affects the
+ // deletion-while-iteration in finalize().
+ using HashFuncsMapType = std::unordered_map<stable_hash, EntryStorage>;
/// Get the HashToFuncs map for serialization.
- const HashFuncsMapType &getFunctionMap() const { return HashToFuncs; }
+ const HashFuncsMapType &getFunctionMap() const;
/// Get the NameToId vector for serialization.
ArrayRef<std::string> getNames() const { return IdToName; }
@@ -99,6 +128,19 @@ struct StableFunctionMap {
/// \returns true if there is no stable function entry.
bool empty() const { return size() == 0; }
+ /// \returns true if there is an entry for the given function hash.
+ /// This does not trigger lazy loading.
+ bool contains(HashFuncsMapType::key_type FunctionHash) const {
+ return HashToFuncs.count(FunctionHash) > 0;
+ }
+
+ /// \returns the stable function entries for the given function hash. If the
+ /// map is lazily loaded, it will deserialize the entries if it is not already
+ /// done, other requests to the same hash at the same time will be blocked
+ /// until the entries are deserialized.
+ const StableFunctionEntries &
+ at(HashFuncsMapType::key_type FunctionHash) const;
+
enum SizeType {
UniqueHashCount, // The number of unique hashes in HashToFuncs.
TotalFunctionCount, // The number of total functions in HashToFuncs.
@@ -119,17 +161,31 @@ struct StableFunctionMap {
/// `StableFunctionEntry` is ready for insertion.
void insert(std::unique_ptr<StableFunctionEntry> FuncEntry) {
assert(!Finalized && "Cannot insert after finalization");
- HashToFuncs[FuncEntry->Hash].emplace_back(std::move(FuncEntry));
+ HashToFuncs[FuncEntry->Hash].Entries.emplace_back(std::move(FuncEntry));
}
+ void deserializeLazyLoadingEntry(HashFuncsMapType::iterator It) const;
+
+ /// Eagerly deserialize all the unloaded entries in the lazy loading map.
+ void deserializeLazyLoadingEntries() const;
+
+ bool isLazilyLoaded() const { return (bool)Buffer; }
+
/// A map from a stable_hash to a vector of functions with that hash.
- HashFuncsMapType HashToFuncs;
+ mutable HashFuncsMapType HashToFuncs;
/// A vector of strings to hold names.
SmallVector<std::string> IdToName;
/// A map from StringRef (name) to an ID.
StringMap<unsigned> NameToId;
/// True if the function map is finalized with minimal content.
bool Finalized = false;
+ /// The memory buffer that contains the serialized stable function map for
+ /// lazy loading.
+ /// Non-empty only if this StableFunctionMap is created from a MemoryBuffer
+ /// (i.e. by IndexedCodeGenDataReader::read()) and lazily deserialized.
+ std::shared_ptr<MemoryBuffer> Buffer;
+ /// Whether to read stable function names from the buffer.
+ bool ReadStableFunctionMapNames = true;
friend struct StableFunctionMapRecord;
};
diff --git a/llvm/include/llvm/CGData/StableFunctionMapRecord.h b/llvm/include/llvm/CGData/StableFunctionMapRecord.h
index a75cb12a70ba6..2d8b573a3cb46 100644
--- a/llvm/include/llvm/CGData/StableFunctionMapRecord.h
+++ b/llvm/include/llvm/CGData/StableFunctionMapRecord.h
@@ -24,6 +24,26 @@
namespace llvm {
+/// The structure of the serialized stable function map is as follows:
+/// - Number of unique function/module names
+/// - Total size of unique function/module names for opt-in skipping
+/// - Unique function/module names
+/// - Padding to align to 4 bytes
+/// - Number of StableFunctionEntries
+/// - Hashes of each StableFunctionEntry
+/// - Fixed-size fields for each StableFunctionEntry (the order is consistent
+/// with the hashes above):
+/// - FunctionNameId
+/// - ModuleNameId
+/// - InstCount
+/// - Relative offset to the beginning of IndexOperandHashes for this entry
+/// - Total size of variable-sized IndexOperandHashes for lazy-loading support
+/// - Variable-sized IndexOperandHashes for each StableFunctionEntry:
+/// - Number of IndexOperandHashes
+/// - Contents of each IndexOperandHashes
+/// - InstIndex
+/// - OpndIndex
+/// - OpndHash
struct StableFunctionMapRecord {
std::unique_ptr<StableFunctionMap> FunctionMap;
@@ -40,13 +60,25 @@ struct StableFunctionMapRecord {
const StableFunctionMap *FunctionMap,
std::vector<CGDataPatchItem> &PatchItems);
+ /// A static helper function to deserialize the stable function map entry.
+ /// Ptr should be pointing to the start of the fixed-sized fields of the
+ /// entry when passed in.
+ LLVM_ABI static void deserializeEntry(const unsigned char *Ptr,
+ stable_hash Hash,
+ StableFunctionMap *FunctionMap);
+
/// Serialize the stable function map to a raw_ostream.
LLVM_ABI void serialize(raw_ostream &OS,
std::vector<CGDataPatchItem> &PatchItems) const;
/// Deserialize the stable function map from a raw_ostream.
- LLVM_ABI void deserialize(const unsigned char *&Ptr,
- bool ReadStableFunctionMapNames = true);
+ LLVM_ABI void deserialize(const unsigned char *&Ptr);
+
+ /// Lazily deserialize the stable function map from `Buffer` starting at
+ /// `Offset`. The individual stable function entry would be read lazily from
+ /// `Buffer` when the function map is accessed.
+ LLVM_ABI void lazyDeserialize(std::shared_ptr<MemoryBuffer> Buffer,
+ uint64_t Offset);
/// Serialize the stable function map to a YAML stream.
LLVM_ABI void serializeYAML(yaml::Output &YOS) const;
@@ -70,6 +102,18 @@ struct StableFunctionMapRecord {
yaml::Output YOS(OS);
serializeYAML(YOS);
}
+
+ /// Set whether to read stable function names from the buffer.
+ /// Has no effect if the function map is read from a YAML stream.
+ void setReadStableFunctionMapNames(bool Read) {
+ assert(
+ FunctionMap->empty() &&
+ "Cannot change ReadStableFunctionMapNames after the map is populated");
+ FunctionMap->ReadStableFunctionMapNames = Read;
+ }
+
+private:
+ void deserialize(const unsigned char *&Ptr, bool Lazy);
};
} // namespace llvm
diff --git a/llvm/lib/CGData/CodeGenData.cpp b/llvm/lib/CGData/CodeGenData.cpp
index cd012342e1958..b4f08c3d13b0d 100644
--- a/llvm/lib/CGData/CodeGenData.cpp
+++ b/llvm/lib/CGData/CodeGenData.cpp
@@ -186,7 +186,7 @@ Expected<Header> Header::readFromBuffer(const unsigned char *Curr) {
return make_error<CGDataError>(cgdata_error::unsupported_version);
H.DataKind = endian::readNext<uint32_t, endianness::little, unaligned>(Curr);
- static_assert(IndexedCGData::CGDataVersion::CurrentVersion == Version3,
+ static_assert(IndexedCGData::CGDataVersion::CurrentVersion == Version4,
"Please update the offset computation below if a new field has "
"been added to the header.");
H.OutlinedHashTreeOffset =
diff --git a/llvm/lib/CGData/CodeGenDataReader.cpp b/llvm/lib/CGData/CodeGenDataReader.cpp
index 0ab35499c8986..fc59be8df525a 100644
--- a/llvm/lib/CGData/CodeGenDataReader.cpp
+++ b/llvm/lib/CGData/CodeGenDataReader.cpp
@@ -26,6 +26,12 @@ static cl::opt<bool> IndexedCodeGenDataReadFunctionMapNames(
"disabled to save memory and time for final consumption of the "
"indexed CodeGenData in production."));
+cl::opt<bool> IndexedCodeGenDataLazyLoading(
+ "indexed-codegen-data-lazy-loading", cl::init(false), cl::Hidden,
+ cl::desc(
+ "Lazily load indexed CodeGenData. Enable to save memory and time "
+ "for final consumption of the indexed CodeGenData in production."));
+
namespace llvm {
static Expected<std::unique_ptr<MemoryBuffer>>
@@ -109,11 +115,20 @@ Error IndexedCodeGenDataReader::read() {
return error(cgdata_error::eof);
HashTreeRecord.deserialize(Ptr);
}
+
+ // TODO: lazy loading support for outlined hash tree.
+ std::shared_ptr<MemoryBuffer> SharedDataBuffer = std::move(DataBuffer);
if (hasStableFunctionMap()) {
const unsigned char *Ptr = Start + Header.StableFunctionMapOffset;
if (Ptr >= End)
return error(cgdata_error::eof);
- FunctionMapRecord.deserialize(Ptr, IndexedCodeGenDataReadFunctionMapNames);
+ FunctionMapRecord.setReadStableFunctionMapNames(
+ IndexedCodeGenDataReadFunctionMapNames);
+ if (IndexedCodeGenDataLazyLoading)
+ FunctionMapRecord.lazyDeserialize(SharedDataBuffer,
+ Header.StableFunctionMapOffset);
+ else
+ FunctionMapRecord.deserialize(Ptr);
}
return success();
diff --git a/llvm/lib/CGData/StableFunctionMap.cpp b/llvm/lib/CGData/StableFunctionMap.cpp
index 87f1e76afb60b..46e04bd850104 100644
--- a/llvm/lib/CGData/StableFunctionMap.cpp
+++ b/llvm/lib/CGData/StableFunctionMap.cpp
@@ -15,6 +15,7 @@
#include "llvm/CGData/StableFunctionMap.h"
#include "llvm/ADT/SmallSet.h"
+#include "llvm/CGData/StableFunctionMapRecord.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
@@ -93,9 +94,10 @@ void StableFunctionMap::insert(const StableFunction &Func) {
void StableFunctionMap::merge(const StableFunctionMap &OtherMap) {
assert(!Finalized && "Cannot merge after finalization");
+ deserializeLazyLoadingEntries();
for (auto &[Hash, Funcs] : OtherMap.HashToFuncs) {
- auto &ThisFuncs = HashToFuncs[Hash];
- for (auto &Func : Funcs) {
+ auto &ThisFuncs = HashToFuncs[Hash].Entries;
+ for (auto &Func : Funcs.Entries) {
auto FuncNameId =
getIdOrCreateForName(*OtherMap.getNameForId(Func->FunctionNameId));
auto ModuleNameId =
@@ -114,25 +116,63 @@ size_t StableFunctionMap::size(SizeType Type) const {
case UniqueHashCount:
return HashToFuncs.size();
case TotalFunctionCount: {
+ deserializeLazyLoadingEntries();
size_t Count = 0;
for (auto &Funcs : HashToFuncs)
- Count += Funcs.second.size();
+ Count += Funcs.second.Entries.size();
return Count;
}
case MergeableFunctionCount: {
+ deserializeLazyLoadingEntries();
size_t Count = 0;
for (auto &[Hash, Funcs] : HashToFuncs)
- if (Funcs.size() >= 2)
- Count += Funcs.size();
+ if (Funcs.Entries.size() >= 2)
+ Count += Funcs.Entries.size();
return Count;
}
}
llvm_unreachable("Unhandled size type");
}
+const StableFunctionMap::StableFunctionEntries &
+StableFunctionMap::at(HashFuncsMapType::key_type FunctionHash) const {
+ auto It = HashToFuncs.find(FunctionHash);
+ if (isLazilyLoaded())
+ deserializeLazyLoadingEntry(It);
+ return It->second.Entries;
+}
+
+void StableFunctionMap::deserializeLazyLoadingEntry(
+ HashFuncsMapType::iterator It) const {
+ assert(isLazilyLoaded() && "Cannot deserialize non-lazily-loaded map");
+ auto &[Hash, Storage] = *It;
+ std::call_once(Storage.LazyLoadFlag,
+ [this, HashArg = Hash, &StorageArg = Storage]() {
+ for (auto Offset : StorageArg.Offsets)
+ StableFunctionMapRecord::deserializeEntry(
+ reinterpret_cast<const unsigned char *>(Offset),
+ HashArg, const_cast<StableFunctionMap *>(this));
+ });
+}
+
+void StableFunctionMap::deserializeLazyLoadingEntries() const {
+ if (!isLazilyLoaded())
+ return;
+ for (auto It = HashToFuncs.begin(); It != HashToFuncs.end(); ++It)
+ deserializeLazyLoadingEntry(It);
+}
+
+const StableFunctionMap::HashFuncsMapType &
+StableFunctionMap::getFunctionMap() const {
+ // Ensure all entries are deserialized before returning the raw map.
+ if (isLazilyLoaded())
+ deserializeLazyLoadingEntries();
+ return HashToFuncs;
+}
+
using ParamLocs = SmallVector<IndexPair>;
-static void removeIdenticalIndexPair(
- SmallVector<std::unique_ptr<StableFunctionMap::StableFunctionEntry>> &SFS) {
+static void
+removeIdenticalIndexPair(StableFunctionMap::StableFunctionEntries &SFS) {
auto &RSF = SFS[0];
unsigned StableFunctionCount = SFS.size();
@@ -159,9 +199,7 @@ static void removeIdenticalIndexPair(
SF->IndexOperandHashMap->erase(Pair);
}
-static bool isProfitable(
- const SmallVector<std::unique_ptr<StableFunctionMap::StableFunctionEntry>>
- &SFS) {
+static bool isProfitable(const StableFunctionMap::StableFunctionEntries &SFS) {
unsigned StableFunctionCount = SFS.size();
if (StableFunctionCount < GlobalMergingMinMerges)
return false;
@@ -202,8 +240,11 @@ static bool isProfitable(
}
void StableFunctionMap::finalize(bool SkipTrim) {
+ deserializeLazyLoadingEntries();
+ SmallVector<HashFuncsMapType::iterator> ToDelete;
for (auto It = HashToFuncs.begin(); It != HashToFuncs.end(); ++It) {
- auto &[StableHash, SFS] = *It;
+ auto &[StableHash, Storage] = *It;
+ auto &SFS = Storage.Entries;
// Group stable functions by ModuleIdentifier.
llvm::stable_sort(SFS, [&](const std::unique_ptr<StableFunctionEntry> &L,
@@ -236,7 +277,7 @@ void StableFunctionMap::finalize(bool SkipTrim) {
}
}
if (Invalid) {
- HashToFuncs.erase(It);
+ ToDelete.push_back(It);
continue;
}
@@ -248,8 +289,10 @@ void StableFunctionMap::finalize(bool SkipTrim) {
removeIdenticalIndexPair(SFS);
if (!isProfitable(SFS))
- HashToFuncs.erase(It);
+ ToDelete.push_back(It);
}
+ for (auto It : ToDelete)
+ HashToFuncs.erase(It);
Finalized = true;
}
diff --git a/llvm/lib/CGData/StableFunctionMapRecord.cpp b/llvm/lib/CGData/StableFunctionMapRecord.cpp
index 423e068023088..e585995ba6a31 100644
--- a/llvm/lib/CGData/StableFunctionMapRecord.cpp
+++ b/llvm/lib/CGData/StableFunctionMapRecord.cpp
@@ -53,7 +53,7 @@ static SmallVector<const StableFunctionMap::StableFunctionEntry *>
getStableFunctionEntries(const StableFunctionMap &SFM) {
SmallVector<const StableFunctionMap::StableFunctionEntry *> FuncEntries;
for (const auto &P : SFM.getFunctionMap())
- for (auto &Func : P.second)
+ for (auto &Func : P.second.Entries)
FuncEntries.emplace_back(Func.get());
llvm::stable_sort(
@@ -107,14 +107,25 @@ void StableFunctionMapRecord::serialize(
// Write StableFunctionEntries whose pointers are sorted.
auto FuncEntries = getStableFunctionEntries(*FunctionMap);
Writer.write<uint32_t>(FuncEntries.size());
-
- for (const auto *FuncRef : FuncEntries) {
+ for (const auto *FuncRef : FuncEntries)
Writer.write<stable_hash>(FuncRef->Hash);
+ std::vector<uint64_t> IndexOperandHashesOffsets;
+ IndexOperandHashesOffsets.reserve(FuncEntries.size());
+ for (const auto *FuncRef : FuncEntries) {
Writer.write<uint32_t>(FuncRef->FunctionNameId);
Writer.write<uint32_t>(FuncRef->ModuleNameId);
Writer.write<uint32_t>(FuncRef->InstCount);
-
+ const uint64_t Offset = Writer.OS.tell();
+ IndexOperandHashesOffsets.push_back(Offset);
+ Writer.write<uint64_t>(0);
+ }
+ const uint64_t IndexOperandHashesByteSizeOffset = Writer.OS.tell();
+ Writer.write<uint64_t>(0);
+ for (size_t I = 0; I < FuncEntries.size(); ++I) {
+ const uint64_t Offset = Writer.OS.tell() - IndexOperandHashesOffsets[I];
+ PatchItems.emplace_back(IndexOperandHashesOffsets[I], &Offset, 1);
// Emit IndexOperandHashes sorted from IndexOperandHashMap.
+ const auto *FuncRef = FuncEntries[I];
IndexOperandHashVecType IndexOperandHashes =
getStableIndexOperandHashes(FuncRef);
Writer.write<uint32_t>(IndexOperandHashes.size());
@@ -124,10 +135,62 @@ void StableFunctionMapRecord::serialize(
Writer.write<stable_hash>(IndexOperandHash.second);
}
}
+ // Write the total size of IndexOperandHashes.
+ const uint64_t IndexOperandHashesByteSize =
+ Writer.OS.tell() - IndexOperandHashesByteSizeOffset - sizeof(uint64_t);
+ PatchItems.emplace_back(IndexOperandHashesByteSizeOffset,
+ &IndexOperandHashesByteSize, 1);
+}
+
+void StableFunctionMapRecord::deserializeEntry(const unsigned char *Ptr,
+ stable_hash Hash,
+ StableFunctionMap *FunctionMap) {
+ auto FunctionNameId =
+ endian::readNext<uint32_t, endianness::little, unaligned>(Ptr);
+ if (FunctionMap->ReadStableFunctionMapNames)
+ assert(FunctionMap->getNameForId(FunctionNameId) &&
+ "FunctionNameId out of range");
+ auto ModuleNameId =
+ endian::readNext<uint32_t, endianness::little, unaligned>(Ptr);
+ if (FunctionMap->ReadStableFunctionMapNames)
+ assert(FunctionMap->getNameForId(ModuleNameId) &&
+ "ModuleNameId out of range");
+ auto InstCount =
+ endian::readNext<uint32_t, endianness::little, unaligned>(Ptr);
+
+ // Read IndexOperandHashes to build IndexOperandHashMap
+ auto CurrentPosition = reinterpret_cast<uintptr_t>(Ptr);
+ auto IndexOperandHashesOffset =
+ endian::readNext<uint64_t, endianness::little, unaligned>(Ptr);
+ auto *IndexOperandHashesPtr = reinterpret_cast<const unsigned char *>(
+ CurrentPosition + IndexOperandHashesOffset);
+ auto NumIndexOperandHashes =
+ endian::readNext<uint32_t, endianness::little, unaligned>(
+ IndexOperandHashesPtr);
+ auto IndexOperandHashMap = std::make_unique<IndexOperandHashMapType>();
+ for (unsigned J = 0; J < NumIndexOperandHashes; ++J) {
+ auto InstIndex = endian::readNext<uint32_t, endianness::little, unaligned>(
+ IndexOperandHashesPtr);
+ auto O...
[truncated]
|
|
@nico I wasn't able to reproduce the failed build locally, so I just add the missing include. @kyulee-com Could you please help land the patch (and potentially revert it again, which I hope it does not happen)? |
kyulee-com
approved these changes
Aug 20, 2025
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This is an attempt to reland #151660 by including a missing STL header found by a buildbot failure.
The stable function map could be huge for a large application. Fully loading it is slow and consumes a significant amount of memory, which is unnecessary and drastically slows down compilation especially for non-LTO and distributed-ThinLTO setups. This patch introduces an opt-in lazy loading support for the stable function map. The detailed changes are:
StableFunctionMapThe map now stores entries in an
EntryStoragestruct, which includes offsets for serialized entries and astd::once_flagfor thread-safe lazy loading.The underlying map type is changed from
DenseMaptostd::unordered_mapfor compatibility withstd::once_flag.contains(),size()andat()are implemented to only load requested entries on demand.Lazy Loading Mechanism
When reading indexed codegen data, if the newly-introduced
-indexed-codegen-data-lazy-loadingflag is set, the stable function map is not fully deserialized up front. The binary format for the stable function map now includes offsets and sizes to support lazy loading.The safety of lazy loading is guarded by the once flag per function hash. This guarantees that even in a multi-threaded environment, the deserialization for a given function hash will happen exactly once. The first thread to request it performs the load, and subsequent threads will wait for it to complete before using the data. For single-threaded builds, the overhead is negligible (a single check on the once flag). For multi-threaded scenarios, users can omit the flag to retain the previous eager-loading behavior.