[clang] Introduce paged vector (PR #66430)
Giulio Eulisse via cfe-commits
cfe-commits at lists.llvm.org
Fri Sep 15 00:33:56 PDT 2023
https://github.com/ktf updated https://github.com/llvm/llvm-project/pull/66430
>From b952f0577dfe8112f394bd5392212f843c0cc86e Mon Sep 17 00:00:00 2001
From: Giulio Eulisse <10544+ktf at users.noreply.github.com>
Date: Thu, 14 Sep 2023 21:58:21 +0200
Subject: [PATCH 1/4] Introduce PagedVector class
The goal of the class is to be an (almost) drop in replacement for
SmallVector and std::vector when those are presized and filled later,
as it happens in SourceManager and ASTReader.
By splitting the actual vector in pages of the same size and allocating
the pages only when they are needed, using this containers reduces the
memory usage by a factor 4 for the cases relevant to the ALICE
experiment ROOT / cling usage.
---
llvm/include/llvm/ADT/PagedVector.h | 96 +++++++++++++++++++++++++++++
1 file changed, 96 insertions(+)
create mode 100644 llvm/include/llvm/ADT/PagedVector.h
diff --git a/llvm/include/llvm/ADT/PagedVector.h b/llvm/include/llvm/ADT/PagedVector.h
new file mode 100644
index 000000000000000..dab0d249aa706e4
--- /dev/null
+++ b/llvm/include/llvm/ADT/PagedVector.h
@@ -0,0 +1,96 @@
+//===- llvm/ADT/PagedVector.h - 'Lazyly allocated' vectors --------*- C++
+//-*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This file defines the PagedVector class.
+//
+//===----------------------------------------------------------------------===//
+#ifndef LLVM_ADT_PAGEDVECTOR_H
+#define LLVM_ADT_PAGEDVECTOR_H
+
+#include <vector>
+
+// Notice that this does not have iterators, because if you
+// have iterators it probably means you are going to touch
+// all the memory in any case, so better use a std::vector in
+// the first place.
+template <typename T, int PAGE_SIZE = 1024 / sizeof(T)> class PagedVector {
+ size_t Size = 0;
+ // Index of where to find a given page in the data
+ mutable std::vector<int> Lookup;
+ // Actual data
+ mutable std::vector<T> Data;
+
+public:
+ // Add a range to the vector.
+ // When vector is accessed, it will call the callback to fill the range
+ // with data.
+
+ // Lookup an element at position i.
+ // If the given range is not filled, it will be filled.
+ // If the given range is filled, return the element.
+ T &operator[](int Index) const { return at(Index); }
+
+ T &at(int Index) const {
+ auto &PageId = Lookup[Index / PAGE_SIZE];
+ // If the range is not filled, fill it
+ if (PageId == -1) {
+ int OldSize = Data.size();
+ PageId = OldSize / PAGE_SIZE;
+ // Allocate the memory
+ Data.resize(OldSize + PAGE_SIZE);
+ // Fill the whole capacity with empty elements
+ for (int I = 0; I < PAGE_SIZE; ++I) {
+ Data[I + OldSize] = T();
+ }
+ }
+ // Return the element
+ return Data[Index % PAGE_SIZE + PAGE_SIZE * PageId];
+ }
+
+ // Return the size of the vector
+ size_t capacity() const { return Lookup.size() * PAGE_SIZE; }
+
+ size_t size() const { return Size; }
+
+ // Expands the vector to the given size.
+ // If the vector is already bigger, does nothing.
+ void expand(size_t NewSize) {
+ // You cannot shrink the vector, otherwise
+ // you would have to invalidate
+ assert(NewSize >= Size);
+ if (NewSize <= Size) {
+ return;
+ }
+ if (NewSize <= capacity()) {
+ Size = NewSize;
+ return;
+ }
+ auto Pages = NewSize / PAGE_SIZE;
+ auto Remainder = NewSize % PAGE_SIZE;
+ if (Remainder) {
+ Pages += 1;
+ }
+ assert(Pages > Lookup.size());
+ Lookup.resize(Pages, -1);
+ Size = NewSize;
+ }
+
+ // Return true if the vector is empty
+ bool empty() const { return Size == 0; }
+ /// Clear the vector
+ void clear() {
+ Size = 0;
+ Lookup.clear();
+ Data.clear();
+ }
+
+ std::vector<T> const &materialised() const { return Data; }
+};
+
+#endif // LLVM_ADT_PAGEDVECTOR_H
>From f0c75c8af0fea27b5b758e2e5775787f33595de3 Mon Sep 17 00:00:00 2001
From: Giulio Eulisse <10544+ktf at users.noreply.github.com>
Date: Thu, 14 Sep 2023 22:20:29 +0200
Subject: [PATCH 2/4] Use PagedVector to reduce memory usage of sparsely
populated vectors
---
clang/include/clang/Basic/SourceManager.h | 3 ++-
clang/include/clang/Serialization/ASTReader.h | 5 +++--
clang/lib/Basic/SourceManager.cpp | 12 ++++++------
clang/lib/Serialization/ASTReader.cpp | 9 +++++----
4 files changed, 16 insertions(+), 13 deletions(-)
diff --git a/clang/include/clang/Basic/SourceManager.h b/clang/include/clang/Basic/SourceManager.h
index 2f846502d6f3327..b1942a3d86afc37 100644
--- a/clang/include/clang/Basic/SourceManager.h
+++ b/clang/include/clang/Basic/SourceManager.h
@@ -43,6 +43,7 @@
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/DenseSet.h"
#include "llvm/ADT/IntrusiveRefCntPtr.h"
+#include "llvm/ADT/PagedVector.h"
#include "llvm/ADT/PointerIntPair.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/StringRef.h"
@@ -699,7 +700,7 @@ class SourceManager : public RefCountedBase<SourceManager> {
///
/// Negative FileIDs are indexes into this table. To get from ID to an index,
/// use (-ID - 2).
- SmallVector<SrcMgr::SLocEntry, 0> LoadedSLocEntryTable;
+ PagedVector<SrcMgr::SLocEntry> LoadedSLocEntryTable;
/// The starting offset of the next local SLocEntry.
///
diff --git a/clang/include/clang/Serialization/ASTReader.h b/clang/include/clang/Serialization/ASTReader.h
index dc1eb21c27801fe..567aecc8246e761 100644
--- a/clang/include/clang/Serialization/ASTReader.h
+++ b/clang/include/clang/Serialization/ASTReader.h
@@ -38,6 +38,7 @@
#include "llvm/ADT/DenseSet.h"
#include "llvm/ADT/IntrusiveRefCntPtr.h"
#include "llvm/ADT/MapVector.h"
+#include "llvm/ADT/PagedVector.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/SetVector.h"
#include "llvm/ADT/SmallPtrSet.h"
@@ -487,7 +488,7 @@ class ASTReader
///
/// When the pointer at index I is non-NULL, the type with
/// ID = (I + 1) << FastQual::Width has already been loaded
- std::vector<QualType> TypesLoaded;
+ PagedVector<QualType> TypesLoaded;
using GlobalTypeMapType =
ContinuousRangeMap<serialization::TypeID, ModuleFile *, 4>;
@@ -501,7 +502,7 @@ class ASTReader
///
/// When the pointer at index I is non-NULL, the declaration with ID
/// = I + 1 has already been loaded.
- std::vector<Decl *> DeclsLoaded;
+ PagedVector<Decl *> DeclsLoaded;
using GlobalDeclMapType =
ContinuousRangeMap<serialization::DeclID, ModuleFile *, 4>;
diff --git a/clang/lib/Basic/SourceManager.cpp b/clang/lib/Basic/SourceManager.cpp
index 0521ac7b30339ab..c028afe63ac85ad 100644
--- a/clang/lib/Basic/SourceManager.cpp
+++ b/clang/lib/Basic/SourceManager.cpp
@@ -458,7 +458,7 @@ SourceManager::AllocateLoadedSLocEntries(unsigned NumSLocEntries,
CurrentLoadedOffset - TotalSize < NextLocalOffset) {
return std::make_pair(0, 0);
}
- LoadedSLocEntryTable.resize(LoadedSLocEntryTable.size() + NumSLocEntries);
+ LoadedSLocEntryTable.expand(LoadedSLocEntryTable.size() + NumSLocEntries);
SLocEntryLoaded.resize(LoadedSLocEntryTable.size());
CurrentLoadedOffset -= TotalSize;
int ID = LoadedSLocEntryTable.size();
@@ -2344,11 +2344,11 @@ SourceManager::MemoryBufferSizes SourceManager::getMemoryBufferSizes() const {
}
size_t SourceManager::getDataStructureSizes() const {
- size_t size = llvm::capacity_in_bytes(MemBufferInfos)
- + llvm::capacity_in_bytes(LocalSLocEntryTable)
- + llvm::capacity_in_bytes(LoadedSLocEntryTable)
- + llvm::capacity_in_bytes(SLocEntryLoaded)
- + llvm::capacity_in_bytes(FileInfos);
+ size_t size = llvm::capacity_in_bytes(MemBufferInfos) +
+ llvm::capacity_in_bytes(LocalSLocEntryTable) +
+ llvm::capacity_in_bytes(LoadedSLocEntryTable.materialised()) +
+ llvm::capacity_in_bytes(SLocEntryLoaded) +
+ llvm::capacity_in_bytes(FileInfos);
if (OverriddenFilesInfo)
size += llvm::capacity_in_bytes(OverriddenFilesInfo->OverriddenFiles);
diff --git a/clang/lib/Serialization/ASTReader.cpp b/clang/lib/Serialization/ASTReader.cpp
index 0952244d037a77c..18b1a9a2480a73e 100644
--- a/clang/lib/Serialization/ASTReader.cpp
+++ b/clang/lib/Serialization/ASTReader.cpp
@@ -3260,7 +3260,7 @@ llvm::Error ASTReader::ReadASTBlock(ModuleFile &F,
std::make_pair(LocalBaseTypeIndex,
F.BaseTypeIndex - LocalBaseTypeIndex));
- TypesLoaded.resize(TypesLoaded.size() + F.LocalNumTypes);
+ TypesLoaded.expand(TypesLoaded.size() + F.LocalNumTypes);
}
break;
}
@@ -3290,7 +3290,7 @@ llvm::Error ASTReader::ReadASTBlock(ModuleFile &F,
// module.
F.GlobalToLocalDeclIDs[&F] = LocalBaseDeclID;
- DeclsLoaded.resize(DeclsLoaded.size() + F.LocalNumDecls);
+ DeclsLoaded.expand(DeclsLoaded.size() + F.LocalNumDecls);
}
break;
}
@@ -7944,9 +7944,10 @@ void ASTReader::PrintStats() {
std::fprintf(stderr, "*** AST File Statistics:\n");
unsigned NumTypesLoaded =
- TypesLoaded.size() - llvm::count(TypesLoaded, QualType());
+ TypesLoaded.size() - llvm::count(TypesLoaded.materialised(), QualType());
unsigned NumDeclsLoaded =
- DeclsLoaded.size() - llvm::count(DeclsLoaded, (Decl *)nullptr);
+ DeclsLoaded.size() -
+ llvm::count(DeclsLoaded.materialised(), (Decl *)nullptr);
unsigned NumIdentifiersLoaded =
IdentifiersLoaded.size() -
llvm::count(IdentifiersLoaded, (IdentifierInfo *)nullptr);
>From 6d5095809dcb95fbad4231925e4b8c08d094143c Mon Sep 17 00:00:00 2001
From: Giulio Eulisse <10544+ktf at users.noreply.github.com>
Date: Fri, 15 Sep 2023 09:30:15 +0200
Subject: [PATCH 3/4] Update documentation
---
llvm/docs/ProgrammersManual.rst | 29 +++++++++++++++++++++++++++++
1 file changed, 29 insertions(+)
diff --git a/llvm/docs/ProgrammersManual.rst b/llvm/docs/ProgrammersManual.rst
index 43dd985d9779ed2..8c26e01728b6910 100644
--- a/llvm/docs/ProgrammersManual.rst
+++ b/llvm/docs/ProgrammersManual.rst
@@ -1625,6 +1625,35 @@ SmallVector has grown a few other minor advantages over std::vector, causing
and is no longer "private to the implementation". A name like
``SmallVectorHeader`` might be more appropriate.
+.. _dss_smallvector:
+
+llvm/ADT/PagedVector.h
+^^^^^^^^^^^^^^^^^^^^^^
+
+``PagedVector<Type, PageSize>`` is a random access container that allocates (PageSize) elements
+of type Type when the first element of a page is accessed via the ``operator[]`` or the ``at()``
+method. This is useful for the case in which the number of elements is known in advance and
+their actual initialization is expensive and sparse so that it's only done lazily when the element is
+accessed. Typical cases for this are the ASTReader::TypesLoaded and ASTReader::DeclsLoaded containers.
+
+The main advantage is that a PagedVector allows to delay the actual allocation of the page until it's needed,
+at the extra cost of one integer per page and one extra indirection when accessing elements with their positional
+index.
+
+In order to maximise the memory footprint of this container, it's important to balance the PageSize so that
+it's not too small (otherwise the overhead of the integer per page might become too high) and not too big
+(otherwise the memory is wasted if the page is not fully used).
+
+The PagedVector can only be expanded at the end, and it's not possible to shrink it, to keep the implementation
+simple and not give the false hope that the resize operation is cheap.
+
+Moreover, while retaining the oder of the elements based on their insertion index, like a vector, iterating over
+the elements via begin() and end() is not provided in the API, due to the fact accessing the elements in order would
+allocate all the iterated pages, defeating memory savings and the purpose of the PagedVector.
+
+Finally a ``materialised()`` method is provided to access the elements associated to the accessed pages, which could
+speedup operations that need to iterate over initialized elements in a non-ordered manner.
+
.. _dss_vector:
<vector>
>From b5dc8ed4ae7ccd7702e5a3065b62fb4675f36084 Mon Sep 17 00:00:00 2001
From: Giulio Eulisse <10544+ktf at users.noreply.github.com>
Date: Fri, 15 Sep 2023 09:32:51 +0200
Subject: [PATCH 4/4] Improve comments
---
llvm/include/llvm/ADT/PagedVector.h | 14 +++++++-------
1 file changed, 7 insertions(+), 7 deletions(-)
diff --git a/llvm/include/llvm/ADT/PagedVector.h b/llvm/include/llvm/ADT/PagedVector.h
index dab0d249aa706e4..635f2fd0829b36b 100644
--- a/llvm/include/llvm/ADT/PagedVector.h
+++ b/llvm/include/llvm/ADT/PagedVector.h
@@ -15,6 +15,9 @@
#include <vector>
+// A vector that allocates memory in pages.
+// Order is kept, but memory is allocated only when one element of the page is
+// accessed.
// Notice that this does not have iterators, because if you
// have iterators it probably means you are going to touch
// all the memory in any case, so better use a std::vector in
@@ -23,17 +26,13 @@ template <typename T, int PAGE_SIZE = 1024 / sizeof(T)> class PagedVector {
size_t Size = 0;
// Index of where to find a given page in the data
mutable std::vector<int> Lookup;
- // Actual data
+ // Actual page data
mutable std::vector<T> Data;
public:
- // Add a range to the vector.
- // When vector is accessed, it will call the callback to fill the range
- // with data.
-
// Lookup an element at position i.
- // If the given range is not filled, it will be filled.
- // If the given range is filled, return the element.
+ // If the associated page is not filled, it will be filled with default
+ // constructed elements. If the associated page is filled, return the element.
T &operator[](int Index) const { return at(Index); }
T &at(int Index) const {
@@ -83,6 +82,7 @@ template <typename T, int PAGE_SIZE = 1024 / sizeof(T)> class PagedVector {
// Return true if the vector is empty
bool empty() const { return Size == 0; }
+
/// Clear the vector
void clear() {
Size = 0;
More information about the cfe-commits
mailing list