refactor: use simd bit packing for index

Author: variar

3rdparty/CMakeLists.txt

@@ -386,6 +386,29 @@ if(KLOGG_USE_SENTRY)
   endif()
 endif(KLOGG_USE_SENTRY)
 
+cpmaddpackage(
+  NAME
+  simdcomp
+  GITHUB_REPOSITORY
+  lemire/simdcomp
+  GIT_TAG
+  009c67807670d16f8984c0534aef0e630e5465a4
+  DOWNLOAD_ONLY
+  YES
+)
+if(simdcomp_ADDED)
+  add_library(simdcomp STATIC 
+    ${simdcomp_SOURCE_DIR}/src/avxbitpacking.c
+    ${simdcomp_SOURCE_DIR}/src/simdfor.c
+    ${simdcomp_SOURCE_DIR}/src/simdcomputil.c
+    ${simdcomp_SOURCE_DIR}/src/simdbitpacking.c
+    ${simdcomp_SOURCE_DIR}/src/simdintegratedbitpacking.c
+    ${simdcomp_SOURCE_DIR}/src/simdpackedsearch.c
+    ${simdcomp_SOURCE_DIR}/src/simdpackedselect.c
+  )
+  target_include_directories(simdcomp PUBLIC ${simdcomp_SOURCE_DIR}/include)
+endif()
+
 set(klogg_cpm_targets
     xxhash
     Catch2
@@ -407,6 +430,7 @@ set(klogg_cpm_targets
     crashpad_compat
     crashpad_util
     mini_chromium
+    simdcomp
 )
 foreach(target ${klogg_cpm_targets})
   if(TARGET ${target})

src/logdata/CMakeLists.txt

@@ -1,7 +1,6 @@
 add_library(
   klogg_logdata STATIC
   ${CMAKE_CURRENT_SOURCE_DIR}/include/abstractlogdata.h
-  ${CMAKE_CURRENT_SOURCE_DIR}/include/blockpool.h
   ${CMAKE_CURRENT_SOURCE_DIR}/include/compressedlinestorage.h
   ${CMAKE_CURRENT_SOURCE_DIR}/include/encodingdetector.h
   ${CMAKE_CURRENT_SOURCE_DIR}/include/linepositionarray.h
@@ -16,7 +15,6 @@ add_library(
   ${CMAKE_CURRENT_SOURCE_DIR}/include/filedigest.h
   ${CMAKE_CURRENT_SOURCE_DIR}/include/readablesize.h
   ${CMAKE_CURRENT_SOURCE_DIR}/src/abstractlogdata.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/src/blockpool.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/src/compressedlinestorage.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/src/encodingdetector.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/src/logdata.cpp
@@ -54,6 +52,7 @@ target_link_libraries(
          kdtoolbox
          robin_hood
          simdutf
+         simdcomp
          klogg_mimalloc_wrapper
 )
 

src/logdata/include/blockpool.h

@@ -1,22 +1,3 @@
-/*
- * Copyright (C) 2015 Nicolas Bonnefon and other contributors
- *
- * This file is part of glogg.
- *
- * glogg is free software: you can redistribute it and/or modify
- * it under the terms of the GNU General Public License as published by
- * the Free Software Foundation, either version 3 of the License, or
- * (at your option) any later version.
- *
- * glogg is distributed in the hope that it will be useful,
- * but WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
- * GNU General Public License for more details.
- *
- * You should have received a copy of the GNU General Public License
- * along with glogg.  If not, see <http://www.gnu.org/licenses/>.
- */
-
 /*
  * Copyright (C) 2016 -- 2019 Anton Filimonov and other contributors
  *
@@ -36,88 +17,36 @@
  * along with klogg.  If not, see <http://www.gnu.org/licenses/>.
  */
 
+#include <array>
 #include <cstddef>
 #include <cstdint>
 #include <vector>
 
-#include "blockpool.h"
 #include "linetypes.h"
 #include <type_safe/strong_typedef.hpp>
 
-
 // This class is a compressed storage backend for LinePositionArray
 // It emulates the interface of a vector, but take advantage of the nature
 // of the stored data (increasing end of line addresses) to apply some
 // compression in memory, while still providing fast, constant-time look-up.
 
-/* The current algorithm takes advantage of the fact most lines are reasonably
- * short, it codes each line on:
- * - Line < 127 bytes    : 1 byte
- * - 127 < line < 16383  : 2 bytes
- * - line > 16383        : 6 bytes (or 10 bytes)
- * Uncompressed backend stores line on 4 bytes or 8 bytes.
- *
- * The algorithm is quite simple, the file is first divided in two parts:
- * - The lines whose end are located before UINT32_MAX
- * - The lines whose end are located after UINT32_MAX
- * Those end of lines are stored separately in the table32 and the table64
- * respectively.
- *
- * The EOL list is then divided in blocks of IndexBlockSize (256) lines.
- * A block index vector (per table) contains pointers to each block.
- *
- * Each block is then defined as such:
- * Block32 (sizes in byte)
- * 00 - Absolute EOF address (4 bytes)
- * 04 - ( 0xxx xxxx  if second line is < 127 ) (1 byte, relative)
- *    - ( 10xx xxxx
- *        xxxx xxxx  if second line is < 16383 ) (2 bytes, relative)
- *    - ( 1111 1111
- *        xxxx xxxx
- *        xxxx xxxx  if second line is > 16383 ) (6 bytes, absolute)
- * ...
- * (126 more lines)
- *
- * Block64 (sizes in byte)
- * 00 - Absolute EOF address (8 bytes)
- * 08 - ( 0xxx xxxx  if second line is < 127 ) (1 byte, relative)
- *    - ( 10xx xxxx
- *        xxxx xxxx  if second line is < 16383 ) (2 bytes, relative)
- *    - ( 1111 1111
- *        xxxx xxxx
- *        xxxx xxxx
- *        xxxx xxxx
- *        xxxx xxxx  if second line is > 16383 ) (10 bytes, absolute)
- * ...
- * (126 more lines)
- *
- * Absolute addressing has been adopted for line > 16383 to bound memory usage in case
- * of pathologically long (MBs or GBs) lines, even if it is a bit less efficient for
- * long-ish (30 KB) lines.
- *
- * The table32 always starts at 0, the table64 starts at first_long_line_
- */
-
-#ifndef COMPRESSEDLINESTORAGE_H
-#define COMPRESSEDLINESTORAGE_H
+#ifndef SIMDCOMPRESSEDLINESTORAGE_H
+#define SIMDCOMPRESSEDLINESTORAGE_H
 
 class CompressedLinePositionStorage {
-  public:
-    // Default constructor
-    CompressedLinePositionStorage()
-        : block_index_{ 0 }
-        , long_block_index_{ 0 }
-    {
-    }
+public:
+    CompressedLinePositionStorage();
 
     // Copy constructor would be slow, delete!
     CompressedLinePositionStorage( const CompressedLinePositionStorage& orig ) = delete;
-    CompressedLinePositionStorage& operator=( const CompressedLinePositionStorage& orig ) = delete;
+    CompressedLinePositionStorage& operator=( const CompressedLinePositionStorage& orig )
+        = delete;
 
-    // Move constructor
     CompressedLinePositionStorage( CompressedLinePositionStorage&& orig ) noexcept;
-    // Move assignement
-    CompressedLinePositionStorage& operator=( CompressedLinePositionStorage&& orig ) noexcept;
+    CompressedLinePositionStorage&
+    operator=( CompressedLinePositionStorage&& orig ) noexcept;
+
+    ~CompressedLinePositionStorage() = default;
 
     // Append the passed end-of-line to the storage
     void append( OffsetInFile pos );
@@ -129,76 +58,49 @@ class CompressedLinePositionStorage {
     // Size of the array
     LinesCount size() const
     {
-        return nb_lines_;
+        return nbLines_;
     }
 
     size_t allocatedSize() const;
 
-    struct BlockOffset 
-        : type_safe::strong_typedef<BlockOffset, size_t>
-        , type_safe::strong_typedef_op::increment<BlockOffset>
-        , type_safe::strong_typedef_op::addition<BlockOffset>
-        , type_safe::strong_typedef_op::relational_comparison<BlockOffset>
-        , type_safe::strong_typedef_op::equality_comparison<BlockOffset>
-        , type_safe::strong_typedef_op::explicit_bool<BlockOffset>
-    {
-        using strong_typedef::strong_typedef;
-    };
-
-    // Cache the last position read
-    // This is to speed up consecutive reads (whole page)
-    struct Cache {
-        LineNumber index {std::numeric_limits<LineNumber::UnderlyingType>::max() - 1U};
-        OffsetInFile position {0};
-        BlockOffset offset {0};
-    };
-
     // Element at index
-    OffsetInFile at( size_t i, Cache* lastPosition = nullptr ) const
+    OffsetInFile at( size_t i ) const
     {
-        return at( LineNumber( i ), lastPosition );
+        return at( LineNumber( i ) );
     }
-    OffsetInFile at( LineNumber i, Cache* lastPosition = nullptr ) const;
+    OffsetInFile at( LineNumber i ) const;
 
     // Add one list to the other
     void append_list( const klogg::vector<OffsetInFile>& positions );
 
     // Pop the last element of the storage
     void pop_back();
 
-  private:
+private:
     // Utility for move ctor/assign
     void move_from( CompressedLinePositionStorage&& orig ) noexcept;
 
-    // The two indexes
-    BlockPool<uint32_t> pool32_;
-    BlockPool<OffsetInFile::UnderlyingType> pool64_;
-
-    // Total number of lines in storage
-    LinesCount nb_lines_;
-
-    // Current position (position of the end of the last line added)
-    OffsetInFile current_pos_;
+    void compress_current_block();
+    void uncompress_last_block();
+    struct BlockMetadata {
+        OffsetInFile firstLineOffset{};
+        uint8_t packetBitWidth{};
+        size_t packetStorageOffset{};
+    };
 
-    uint32_t block_index_;
-    uint32_t long_block_index_;
+    klogg::vector<BlockMetadata> blocks_;
+    klogg::vector<uint8_t> packedLinesStorage_;
 
-    // The index of the first line whose end is stored in a block64
-    // this is the origin point for all calculations in block64
-    OptionalLineNumber first_long_line_;
+    klogg::vector<OffsetInFile> currentLinesBlock_;
+    klogg::vector<uint32_t> currentLinesBlockShifted_;
 
-    // Offset of the next position (not yet written) within the current
-    // block. null means there is no current block (previous block
-    // finished or no data)
-    BlockOffset block_offset_;
+    // Total number of lines in storage
+    LinesCount nbLines_;
 
-    // For pop_back:
+    // Current position (position of the end of the last line added)
+    OffsetInFile lastPos_;
 
-    // Previous offset to block element, it is restored when we
-    // "pop_back" the last element.
-    // A null here means pop_back need to free the block
-    // that has just been created.
-    BlockOffset previous_block_offset_;
+    bool canUseSimdSelect_ {false};
 };
 
 #endif