Update chunkedvector.c

Update cmake_build.py

Update CMakeLists.txt

Update CMakeLists.txt

.github/workflows/main.yml

@@ -162,7 +162,7 @@ jobs:
           uses: jirutka/setup-alpine@v1
           with:
             arch: ${{ matrix.arch }}
-            packages: gcc g++ make cmake libc-dev linux-headers python3
+            packages: gcc g++ make cmake libc-dev linux-headers python3 git
         - name: Build and Test
           run: |
             echo "Building for architecture: ${{ matrix.arch }}"

.gitignore

@@ -40,3 +40,4 @@ docs/C-API/functions.md
 
 
 cmake-build-*
+tmp/

CMakeLists.txt

@@ -5,6 +5,7 @@ project(pocketpy)
 set(CMAKE_C_STANDARD 11)
 set(CMAKE_C_STANDARD_REQUIRED ON)
 
+include(FetchContent)
 include(CMakeOptions.txt)
 
 if(WIN32)
@@ -13,8 +14,7 @@ endif()
 
 if(MSVC)
     set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /utf-8 /jumptablerdata")
-    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /experimental:c11atomics")
-    add_compile_options(/wd4267 /wd4244 /wd4146)
+    add_compile_options(/wd4267 /wd4244 /wd4146 /experimental:c11atomics)
 
     if(NOT CMAKE_BUILD_TYPE STREQUAL "Debug")
         set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /Ox")
@@ -76,6 +76,28 @@ else()
     add_definitions(-DPK_ENABLE_CUSTOM_SNAME=0)
 endif()
 
+if(PK_ENABLE_MIMALLOC)
+    message(">> Fetching mimalloc")
+    FetchContent_Declare(
+        mimalloc
+        GIT_REPOSITORY https://github.com/microsoft/mimalloc.git
+        GIT_TAG        v3.1.5
+    )
+
+    set(MI_OVERRIDE OFF CACHE BOOL "" FORCE)
+    set(MI_NO_USE_CXX ON CACHE BOOL "" FORCE)
+    set(MI_BUILD_SHARED OFF CACHE BOOL "" FORCE)
+    set(MI_BUILD_OBJECT OFF CACHE BOOL "" FORCE)
+    set(MI_BUILD_STATIC ON CACHE BOOL "" FORCE)
+    set(MI_BUILD_TESTS OFF CACHE BOOL "" FORCE)
+    FetchContent_MakeAvailable(mimalloc)
+    include_directories(${mimalloc_SOURCE_DIR}/include)
+
+    add_definitions(-DPK_ENABLE_MIMALLOC=1)
+else()
+    add_definitions(-DPK_ENABLE_MIMALLOC=0)
+endif()
+
 
 if(PK_BUILD_MODULE_LZ4)
     add_subdirectory(3rd/lz4)
@@ -87,6 +109,7 @@ if(PK_BUILD_MODULE_LIBHV)
     add_definitions(-DPK_BUILD_MODULE_LIBHV)
 endif()
 
+
 if(PK_BUILD_SHARED_LIB)
     message(">> Building shared library")
     add_library(${PROJECT_NAME} SHARED ${POCKETPY_SRC})
@@ -150,3 +173,7 @@ endif()
 if(PK_BUILD_MODULE_LIBHV)
     target_link_libraries(${PROJECT_NAME} libhv_bindings)
 endif()
+
+if(PK_ENABLE_MIMALLOC)
+    target_link_libraries(${PROJECT_NAME} mimalloc-static)
+endif()

CMakeOptions.txt

@@ -11,6 +11,7 @@ option(PK_ENABLE_THREADS "" ON)
 option(PK_ENABLE_DETERMINISM "" OFF)
 option(PK_ENABLE_WATCHDOG "" OFF)
 option(PK_ENABLE_CUSTOM_SNAME "" OFF)
+option(PK_ENABLE_MIMALLOC "" OFF)
 
 # modules
 option(PK_BUILD_MODULE_LZ4 "" OFF)

cmake_build.py

@@ -20,7 +20,7 @@ assert config in ['Debug', 'Release', 'RelWithDebInfo']
 
 os.chdir("build")
 
-code = os.system(f"cmake .. -DPK_ENABLE_DETERMINISM=ON -DCMAKE_BUILD_TYPE={config} {extra_flags}")
+code = os.system(f"cmake .. -DPK_ENABLE_MIMALLOC=ON -DPK_ENABLE_DETERMINISM=ON -DCMAKE_BUILD_TYPE={config} {extra_flags}")
 assert code == 0
 code = os.system(f"cmake --build . --config {config} -j 4")
 assert code == 0

include/pocketpy/common/chunkedvector.h

@@ -1,6 +1,7 @@
 #pragma once
 
 #include "pocketpy/common/vector.h"
+
 typedef struct c11_chunkedvector_chunk {
     int length;
     int capacity;

include/pocketpy/common/utils.h

@@ -1,6 +1,8 @@
 #pragma once
 
 #include <stdio.h>
+#include <stdlib.h>
+#include <assert.h>
 
 #define PK_REGION(name) 1
 

include/pocketpy/common/vector.h

@@ -1,7 +1,5 @@
 #pragma once
 
-#include <stdlib.h>
-#include <assert.h>
 #include <string.h>
 #include <stdbool.h>
 

include/pocketpy/config.h

@@ -27,24 +27,15 @@
 #define PK_ENABLE_CUSTOM_SNAME      0                
 #endif
 
+#ifndef PK_ENABLE_MIMALLOC          // can be overridden by cmake
+#define PK_ENABLE_MIMALLOC          0                
+#endif
+
 // GC min threshold
 #ifndef PK_GC_MIN_THRESHOLD         // can be overridden by cmake
     #define PK_GC_MIN_THRESHOLD     32768
 #endif
 
-// Memory allocation functions
-#ifndef PK_MALLOC
-    #ifndef __cplusplus
-        #define PK_MALLOC(size)             malloc(size)
-        #define PK_REALLOC(ptr, size)       realloc(ptr, size)
-        #define PK_FREE(ptr)                free(ptr)
-    #else
-        #define PK_MALLOC(size)             std::malloc(size)
-        #define PK_REALLOC(ptr, size)       std::realloc(ptr, size)
-        #define PK_FREE(ptr)                std::free(ptr)
-    #endif
-#endif
-
 // This is the maximum size of the value stack in py_TValue units
 // The actual size in bytes equals `sizeof(py_TValue) * PK_VM_STACK_SIZE`
 #ifndef PK_VM_STACK_SIZE            // can be overridden by cmake
@@ -90,3 +81,25 @@
 #else
     #define PK_THREAD_LOCAL
 #endif
+
+// Memory allocation functions
+#ifndef PK_MALLOC
+    #if PK_ENABLE_MIMALLOC
+        #include "mimalloc.h"
+        #define PK_MALLOC(size)                 mi_malloc(size)
+        #define PK_REALLOC(ptr, size)           mi_realloc(ptr, size)
+        #define PK_FREE(ptr)                    mi_free(ptr)
+    #else
+        #ifndef __cplusplus
+            #include <stdlib.h>
+            #define PK_MALLOC(size)             malloc(size)
+            #define PK_REALLOC(ptr, size)       realloc(ptr, size)
+            #define PK_FREE(ptr)                free(ptr)
+        #else
+            #include <cstdlib>
+            #define PK_MALLOC(size)             std::malloc(size)
+            #define PK_REALLOC(ptr, size)       std::realloc(ptr, size)
+            #define PK_FREE(ptr)                std::free(ptr)
+        #endif
+    #endif
+#endif

src/common/algorithm.c

@@ -1,7 +1,6 @@
 #include "pocketpy/common/algorithm.h"
 #include "pocketpy/config.h"
 #include <string.h>
-#include <stdlib.h>
 
 static bool _stable_sort_merge(char* a,
                   char* a_end,

src/common/chunkedvector.c

@@ -1,15 +1,12 @@
 #include "pocketpy/common/chunkedvector.h"
-
-#include <stdlib.h>
-#include <string.h>
-#include "pocketpy/common/utils.h"
-#include "pocketpy/config.h"
+#include "pocketpy/pocketpy.h"
+#include <assert.h>
 
 #if defined(_MSC_VER)
 #include <intrin.h>
 #endif
 
-inline static int c11_bit_length(unsigned long x) {
+PK_INLINE static int c11__bit_length(unsigned long x) {
 #if(defined(__clang__) || defined(__GNUC__))
     return x == 0 ? 0 : (int)sizeof(unsigned long) * 8 - __builtin_clzl(x);
 #elif defined(_MSC_VER)
@@ -39,7 +36,6 @@ void c11_chunkedvector__ctor(c11_chunkedvector* self, int elem_size, int initial
     self->initial_chunks = initial_chunks;
     void* chunks_data = PK_MALLOC(elem_size * ((1U << (unsigned int)initial_chunks) - 1));
     for(int i = 0; i < initial_chunks; i++) {
-        // TODO: optimize?
         c11_chunkedvector_chunk chunk = {.length = 0,
                                          .capacity = 1U << i,
                                          .data = (char*)chunks_data + elem_size * ((1U << i) - 1U)};
@@ -59,10 +55,11 @@ void c11_chunkedvector__dtor(c11_chunkedvector* self) {
 
 void* c11_chunkedvector__emplace(c11_chunkedvector* self) {
     if(self->length == self->capacity) {
-        #ifndef NDEBUG
-            c11_chunkedvector_chunk last_chunk = c11_vector__back(c11_chunkedvector_chunk, &self->chunks);
-            assert(last_chunk.capacity == last_chunk.length);
-        #endif
+#ifndef NDEBUG
+        c11_chunkedvector_chunk last_chunk =
+            c11_vector__back(c11_chunkedvector_chunk, &self->chunks);
+        assert(last_chunk.capacity == last_chunk.length);
+#endif
         c11_chunkedvector_chunk chunk = {
             .length = 0,
             .capacity = 1U << (unsigned int)self->chunks.length,
@@ -70,9 +67,14 @@ void* c11_chunkedvector__emplace(c11_chunkedvector* self) {
         self->capacity += chunk.capacity;
         c11_vector__push(c11_chunkedvector_chunk, &self->chunks, chunk);
     }
-    int last_chunk_index = c11_bit_length(self->length + 1) - 1;
+#if 1
+    int last_chunk_index = c11__bit_length(self->length + 1) - 1;
     c11_chunkedvector_chunk* last_chunk =
         c11__at(c11_chunkedvector_chunk, &self->chunks, last_chunk_index);
+#else
+    // This is not correct, because there is some pre-allocated chunks
+    c11_chunkedvector_chunk* last_chunk = &c11_vector__back(c11_chunkedvector_chunk, &self->chunks);
+#endif
     void* p = (char*)last_chunk->data + self->elem_size * last_chunk->length;
     last_chunk->length++;
     self->length++;
@@ -80,7 +82,7 @@ void* c11_chunkedvector__emplace(c11_chunkedvector* self) {
 }
 
 void* c11_chunkedvector__at(c11_chunkedvector* self, int index) {
-    int chunk_index = c11_bit_length(index + 1) - 1;
+    int chunk_index = c11__bit_length(index + 1) - 1;
     c11_chunkedvector_chunk* chunk = c11__at(c11_chunkedvector_chunk, &self->chunks, chunk_index);
     return (char*)chunk->data + (index + 1 - (1U << (unsigned int)chunk_index)) * self->elem_size;
 }

src/common/memorypool.c

@@ -1,7 +1,6 @@
 #include "pocketpy/common/memorypool.h"
-#include "pocketpy/pocketpy.h"
+#include "pocketpy/config.h"
 
-#include <stdlib.h>
 #include <stdbool.h>
 
 void FixedMemoryPool__ctor(FixedMemoryPool* self, int BlockSize, int BlockCount) {

src/common/name.c

@@ -4,6 +4,7 @@
 #include "pocketpy/common/str.h"
 #include "pocketpy/pocketpy.h"
 #include "pocketpy/common/threads.h"
+#include <assert.h>
 
 typedef struct NameBucket NameBucket;
 

src/common/sourcedata.c

@@ -1,7 +1,6 @@
 #include "pocketpy/objects/sourcedata.h"
 #include "pocketpy/common/sstream.h"
 #include <ctype.h>
-#include <stdlib.h>
 #include <string.h>
 
 static void SourceData__ctor(struct SourceData* self,

src/common/sstream.c

@@ -4,8 +4,8 @@
 #include "pocketpy/pocketpy.h"
 
 #include <stdarg.h>
-#include <stdio.h>
 #include <assert.h>
+#include <stdio.h>
 #include <ctype.h>
 #include <math.h>
 

src/common/str.c

@@ -2,11 +2,9 @@
 #include "pocketpy/common/sstream.h"
 #include "pocketpy/common/utils.h"
 
-#include <assert.h>
 #include <string.h>
-#include <stdlib.h>
-#include <ctype.h>
 #include <stdio.h>
+#include <assert.h>
 
 c11_string* c11_string__new(const char* data) { return c11_string__new2(data, strlen(data)); }
 

src/common/vector.c

@@ -1,6 +1,5 @@
 #include "pocketpy/common/vector.h"
 
-#include <stdlib.h>
 #include <string.h>
 #include "pocketpy/common/utils.h"
 #include "pocketpy/config.h"

src/interpreter/ceval.c

@@ -7,6 +7,7 @@
 #include "pocketpy/pocketpy.h"
 #include "pocketpy/objects/error.h"
 #include <stdbool.h>
+#include <assert.h>
 #include <time.h>
 
 static bool stack_format_object(VM* self, c11_sv spec);

src/interpreter/frame.c

@@ -5,6 +5,7 @@
 #include "pocketpy/objects/codeobject.h"
 #include "pocketpy/pocketpy.h"
 #include <stdbool.h>
+#include <assert.h>
 
 void ValueStack__ctor(ValueStack* self) {
     self->sp = self->begin;

src/interpreter/generator.c

@@ -4,6 +4,7 @@
 #include "pocketpy/objects/base.h"
 #include "pocketpy/pocketpy.h"
 #include <stdbool.h>
+#include <assert.h>
 
 void pk_newgenerator(py_Ref out, py_Frame* frame, py_TValue* begin, py_TValue* end) {
     Generator* ud = py_newobject(out, tp_generator, 1, sizeof(Generator));

src/interpreter/heap.c

@@ -3,6 +3,7 @@
 #include "pocketpy/interpreter/objectpool.h"
 #include "pocketpy/objects/base.h"
 #include "pocketpy/pocketpy.h"
+#include <assert.h>
 
 void ManagedHeap__ctor(ManagedHeap* self) {
     MultiPool__ctor(&self->small_objects);

src/interpreter/line_profier.c

@@ -1,4 +1,5 @@
 #include "pocketpy/interpreter/line_profiler.h"
+#include <assert.h>
 
 void LineProfiler__ctor(LineProfiler* self) {
     c11_smallmap_p2i__ctor(&self->records);

src/interpreter/objectpool.c

@@ -1,12 +1,10 @@
 #include "pocketpy/interpreter/objectpool.h"
 
-#include "pocketpy/config.h"
 #include "pocketpy/objects/object.h"
 #include "pocketpy/common/sstream.h"
 
 #include <assert.h>
 #include <stdbool.h>
-#include <stdlib.h>
 #include <string.h>
 
 static PoolArena* PoolArena__new(int block_size) {

src/interpreter/typeinfo.c

@@ -1,4 +1,5 @@
 #include "pocketpy/interpreter/vm.h"
+#include <assert.h>
 
 py_ItemRef pk_tpfindname(py_TypeInfo* ti, py_Name name) {
     assert(ti != NULL);

src/interpreter/vm.c

@@ -1,6 +1,5 @@
 #include "pocketpy/interpreter/vm.h"
 #include "pocketpy/common/memorypool.h"
-#include "pocketpy/common/sstream.h"
 #include "pocketpy/common/utils.h"
 #include "pocketpy/interpreter/generator.h"
 #include "pocketpy/interpreter/modules.h"
@@ -10,6 +9,7 @@
 #include "pocketpy/common/_generated.h"
 #include "pocketpy/pocketpy.h"
 #include <stdbool.h>
+#include <assert.h>
 
 static char* pk_default_importfile(const char* path) {
 #if PK_ENABLE_OS

src/interpreter/vmx.c

@@ -1,4 +1,5 @@
 #include "pocketpy/interpreter/vm.h"
+#include <assert.h>
 
 void pk_print_stack(VM* self, py_Frame* frame, Bytecode byte) {
     return;

src/modules/conio.c

@@ -1,5 +1,4 @@
 #include "pocketpy/pocketpy.h"
-#include <stdlib.h>
 
 #if PY_SYS_PLATFORM == 0
 

src/modules/lz4.c

@@ -1,6 +1,5 @@
 #ifdef PK_BUILD_MODULE_LZ4
 
-#include <stdlib.h>
 #include <string.h>
 #include <assert.h>
 #include "pocketpy/pocketpy.h"

src/modules/traceback.c

@@ -1,6 +1,4 @@
-#include "pocketpy/objects/base.h"
 #include "pocketpy/pocketpy.h"
-#include <stdlib.h>
 
 static bool traceback_format_exc(int argc, py_Ref argv) {
     PY_CHECK_ARGC(0);

src/objects/codeobject.c

@@ -2,6 +2,7 @@
 #include "pocketpy/common/utils.h"
 #include "pocketpy/pocketpy.h"
 #include <stdint.h>
+#include <assert.h>
 
 void Bytecode__set_signed_arg(Bytecode* self, int arg) {
     self->arg = (int16_t)arg;

src/objects/namedict.c

@@ -4,7 +4,6 @@
 #include <stdint.h>
 #include <string.h>
 #include <assert.h>
-#include <stdlib.h>
 
 #define HASH_PROBE_0(__k, ok, i)                                                                   \
     ok = false;                                                                                    \

src2/main.c

@@ -1,6 +1,5 @@
 #include <stdbool.h>
 #include <stdio.h>
-#include <stdlib.h>
 #include <assert.h>
 #include <string.h>
 
@@ -20,7 +19,7 @@ static char* read_file(const char* path) {
     fseek(file, 0, SEEK_END);
     long size = ftell(file);
     fseek(file, 0, SEEK_SET);
-    char* buffer = malloc(size + 1);
+    char* buffer = PK_MALLOC(size + 1);
     size = fread(buffer, 1, size, file);
     buffer[size] = 0;
     return buffer;
@@ -85,7 +84,7 @@ int main(int argc, char** argv) {
         char* source = read_file(filename);
         if(source) {
             if(!py_exec(source, filename, EXEC_MODE, NULL)) py_printexc();
-            free(source);
+            PK_FREE(source);
         }
     }