wip

include/customizer/cell_customizer.hpp

@@ -116,7 +116,6 @@ class CellCustomizer
                    const std::vector<bool> &allowed_nodes,
                    CellMetric &metric) const
     {
-        // std::cerr << "Customizing cells\n";
         const auto number_of_nodes = graph.GetNumberOfNodes();
         HeapPtr heaps([number_of_nodes]{
             return Heap{number_of_nodes};

include/util/pool_allocator.hpp

@@ -4,6 +4,7 @@
 #include <array>
 #include <bit>
 #include <boost/assert.hpp>
+#include <cstddef>
 #include <cstdlib>
 #include <iostream>
 #include <mutex>
@@ -14,40 +15,42 @@
 namespace osrm::util
 {
 
-class MemoryManager
+class MemoryPool
 {
 private:
-    constexpr static size_t MIN_ITEMS_IN_BLOCK = 1024;
+    constexpr static size_t MIN_CHUNK_SIZE_BYTES = 4096;
 
 public:
-    static std::shared_ptr<MemoryManager> instance()
+    static std::shared_ptr<MemoryPool> instance()
     {
-        static thread_local std::shared_ptr<MemoryManager> instance;
+        static thread_local std::shared_ptr<MemoryPool> instance;
         if (!instance)
         {
-            instance = std::shared_ptr<MemoryManager>(new MemoryManager());
+            instance = std::shared_ptr<MemoryPool>(new MemoryPool());
         }
         return instance;
     }
 
-    // TODO: alignment!!!
     template <typename T>
-    T *allocate(std::size_t n)
+    T *allocate(std::size_t items_count)
     {
-        size_t free_list_index = get_next_power_of_two_exponent(n * sizeof(T));
+        static_assert(alignof(T) <= alignof(std::max_align_t), "Type is over-aligned for this allocator.");
+
+        size_t free_list_index = get_next_power_of_two_exponent(items_count * sizeof(T));
         auto &free_list = free_lists_[free_list_index];
-        const auto items_in_block = 1u << free_list_index;
         if (free_list.empty())
         {
-            // Check if there is space in current block
+            size_t block_size_in_bytes = 1u << free_list_index;
-            if (current_block_left_items_ < items_in_block)
+            block_size_in_bytes = align_up(block_size_in_bytes, alignof(std::max_align_t));
+            // Check if there is space in current memory chunk
+            if (current_chunk_left_bytes_ < block_size_in_bytes)
             {
-                allocate_block<T>(items_in_block);
+                allocate_chunk(block_size_in_bytes);
             }
 
-            free_list.push_back(current_block_ptr_);
+            free_list.push_back(current_chunk_ptr_);
-            current_block_left_items_ -= items_in_block;
+            current_chunk_left_bytes_ -= block_size_in_bytes;
-            current_block_ptr_ += items_in_block * sizeof(T);
+            current_chunk_ptr_ += block_size_in_bytes;
         }
         auto ptr = static_cast<T*>(free_list.back());
         free_list.pop_back();
@@ -61,49 +64,54 @@ public:
         free_lists_[free_list_index].push_back(p);
     }
 
-    ~MemoryManager()
+    ~MemoryPool()
     {
-        std::cerr << "~MemoryManager()" << std::endl;
+        for (auto chunk : chunks_)
-        for (auto block : blocks_)
         {
-            std::free(block);
+            // NOLINTNEXTLINE(cppcoreguidelines-no-malloc)
+            std::free(chunk);
         }
     }
 
 private:
-    MemoryManager() = default;
+    MemoryPool() = default;
-    MemoryManager(const MemoryManager &) = delete;
+    MemoryPool(const MemoryPool &) = delete;
-    MemoryManager &operator=(const MemoryManager &) = delete;
+    MemoryPool &operator=(const MemoryPool &) = delete;
 
-    size_t get_next_power_of_two_exponent(size_t n) const
+    inline size_t get_next_power_of_two_exponent(size_t n) const
     {
         BOOST_ASSERT(n > 0);
         return (sizeof(size_t) * 8) - std::countl_zero(n - 1);
     }
 
-    template <typename T>
+    inline size_t align_up(size_t n, size_t alignment)
-    void allocate_block(size_t items_in_block)
     {
-        items_in_block = std::max(items_in_block, MIN_ITEMS_IN_BLOCK);
+        return (n + alignment - 1) & ~(alignment - 1);
+    }
 
-        size_t block_size = items_in_block * sizeof(T);
+    inline void* align_pointer(void* ptr, size_t alignment)
-        void *block = std::malloc(block_size);
+    {
-        if (!block)
+        return reinterpret_cast<void*>(align_up(reinterpret_cast<uintptr_t>(ptr), alignment));
+    }
+
+    void allocate_chunk(size_t bytes)
+    {
+        auto chunk_size = std::max(bytes, MIN_CHUNK_SIZE_BYTES);
+        // NOLINTNEXTLINE(cppcoreguidelines-no-malloc)
+        void *chunk = std::malloc(chunk_size);
+        if (!chunk)
         {
             throw std::bad_alloc();
         }
-        total_allocated_ += block_size;
+        chunks_.push_back(chunk);
-        blocks_.push_back(block);
+        current_chunk_ptr_ = static_cast<uint8_t*>(chunk);
-        current_block_ptr_ = static_cast<uint8_t*>(block);
+        current_chunk_left_bytes_ = chunk_size;
-        current_block_left_items_ = items_in_block;
     }
 
     std::array<std::vector<void *>, 32> free_lists_;
-    std::vector<void *> blocks_;
+    std::vector<void *> chunks_;
-    uint8_t *current_block_ptr_ = nullptr;
+    uint8_t *current_chunk_ptr_ = nullptr;
-    size_t current_block_left_items_ = 0;
+    size_t current_chunk_left_bytes_ = 0;
-
-    size_t total_allocated_ = 0;
 };
 
 template <typename T>
@@ -112,10 +120,10 @@ class PoolAllocator
 public:
     using value_type = T;
 
-    PoolAllocator() noexcept : pool(MemoryManager::instance()) {};
+    PoolAllocator() noexcept : pool(MemoryPool::instance()) {};
 
     template <typename U>
-    PoolAllocator(const PoolAllocator<U> &) noexcept : pool(MemoryManager::instance()) {}
+    PoolAllocator(const PoolAllocator<U> &) noexcept : pool(MemoryPool::instance()) {}
 
     template <typename U>
     struct rebind
@@ -133,17 +141,13 @@ public:
         pool->deallocate<T>(p, n);
     }
 
-    ~PoolAllocator() {
-        std::cerr << "~PoolAllocator()" << std::endl;
-    }
-
     PoolAllocator(const PoolAllocator &) = default;
     PoolAllocator &operator=(const PoolAllocator &) = default;
     PoolAllocator(PoolAllocator &&) noexcept = default;
     PoolAllocator &operator=(PoolAllocator &&) noexcept = default;
 
 private:
-    std::shared_ptr<MemoryManager> pool;
+    std::shared_ptr<MemoryPool> pool;
 };
 template <typename T, typename U>
 bool operator==(const PoolAllocator<T> &, const PoolAllocator<U> &)

include/util/query_heap.hpp

@@ -36,10 +36,6 @@ template <typename NodeID, typename Key> class UnorderedMapStorage
   public:
     explicit UnorderedMapStorage(std::size_t) { nodes.rehash(1000); }
 
-    ~UnorderedMapStorage() {
-        std::cerr << "~UnorderedMapStorage()" << std::endl;
-    }
-
     Key &operator[](const NodeID node) { return nodes[node]; }
 
     Key peek_index(const NodeID node) const

src/util/pool_allocator.cpp

@@ -1,5 +0,0 @@
-#include "util/pool_allocator.hpp"
-
-namespace osrm::util
-{
-} // namespace osrm::util

test/data/Makefile

@@ -65,4 +65,4 @@ benchmark: data $(DATA_NAME).requests
 checksum:
 	$(MD5SUM) $(DATA_NAME).osm.pbf $(DATA_NAME).poly > data.md5sum
 
-.PHONY: clean checksum benchmark data
+.PHONY: clean checksum benchmark data

unit_tests/util/pool_allocator.cpp

@@ -85,4 +85,33 @@ BOOST_AUTO_TEST_CASE(unordered_map)
     BOOST_CHECK_EQUAL(map[2], 43);
 }
 
+BOOST_AUTO_TEST_CASE(alignment)
+{
+    PoolAllocator<char> pool_char; 
+    PoolAllocator<double> pool_double; 
+    
+    auto ptr_char = pool_char.allocate(1);
+    auto ptr_double = pool_double.allocate(1);
+
+    BOOST_CHECK_NE(ptr_double, nullptr);
+    BOOST_CHECK_EQUAL(reinterpret_cast<uintptr_t>(ptr_double) % alignof(double), 0);
+    BOOST_CHECK_NE(ptr_char, nullptr);
+    BOOST_CHECK_EQUAL(reinterpret_cast<uintptr_t>(ptr_char) % alignof(char), 0);
+
+   pool_char.deallocate(ptr_char, 1);
+ pool_double.deallocate(ptr_double, 1);
+    
+
+    ptr_char = pool_char.allocate(2);
+    ptr_double = pool_double.allocate(1);
+
+    BOOST_CHECK_NE(ptr_double, nullptr);
+    BOOST_CHECK_EQUAL(reinterpret_cast<uintptr_t>(ptr_double) % alignof(double), 0);
+    BOOST_CHECK_NE(ptr_char, nullptr);
+    BOOST_CHECK_EQUAL(reinterpret_cast<uintptr_t>(ptr_char) % alignof(char), 0);
+
+    pool_char.deallocate(ptr_char, 2);
+    pool_double.deallocate(ptr_double, 1);
+}
+
 BOOST_AUTO_TEST_SUITE_END()