Speedups: remove dependency on c++ (#2796)

* Speedups: remove dependency on c++

* Speedups: intset: handle malloc failing

* Speedups: intset: fix corner case for int64 on 32bit systems

original idea was to only use bucket->val if int<pointer,
but we always have a union now anyway

* Speedups: add size comment to player_set bucket configuration

* test: more tests for LocationStore.find_item

* test: require _speedups in CI

This kind of tests that the build succeeds.

* test: even more tests for LocationStore.find_item

* Speedups: intset uniform comment style

* Speedups: intset: avoid memory leak when realloc fails

* Speedups: intset: make `gcc -pedantic -std=c99 -fanalyzer` without warnings

Unnamed unions are not in C99, this got fixed.
The overhead of setting count=0 is minimal or optimized-out and silences -fanalizer (see comment).

* Speedups: don't leak memory in case of exception

* Speedups: intset: validate alloc and free

This won't happen in our cython, but it's still a good addition.

* CI: add test framework for C/C++ code

* CI: ctest: fix cwd

* Speedups: intset: ignore msvc warning

* Tests: intset: revert attempt at no-asan

We solve this with env vars in ctest now, and this fails for msvc.

* Test: cpp: docs: fix typo

* Test: cpp: docs: fix another typo

* Test: intset: proper bucket count for Negative test

INTxx_MIN % 1 would not produce a negative number, so the test was flawed.

.github/workflows/ctest.yml

@@ -0,0 +1,54 @@
+# Run CMake / CTest C++ unit tests
+
+name: ctest
+
+on:
+  push:
+    paths:
+      - '**.cc?'
+      - '**.cpp'
+      - '**.cxx'
+      - '**.hh?'
+      - '**.hpp'
+      - '**.hxx'
+      - '**.CMakeLists'
+      - '.github/workflows/ctest.yml'
+  pull_request:
+    paths:
+      - '**.cc?'
+      - '**.cpp'
+      - '**.cxx'
+      - '**.hh?'
+      - '**.hpp'
+      - '**.hxx'
+      - '**.CMakeLists'
+      - '.github/workflows/ctest.yml'
+
+jobs:
+  ctest:
+    runs-on: ${{ matrix.os }}
+    name: Test C++ ${{ matrix.os }}
+
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest, windows-latest]
+
+    steps:
+    - uses: actions/checkout@v4
+    - uses: ilammy/msvc-dev-cmd@v1
+      if: startsWith(matrix.os,'windows')
+    - uses: Bacondish2023/setup-googletest@v1
+      with:
+        build-type: 'Release'
+    - name: Build tests
+      run: |
+        cd test/cpp
+        mkdir build
+        cmake -S . -B build/ -DCMAKE_BUILD_TYPE=Release
+        cmake --build build/ --config Release
+        ls
+    - name: Run tests
+      run: |
+        cd test/cpp
+        ctest --test-dir build/ -C Release --output-on-failure

.gitignore

@@ -178,6 +178,7 @@ dmypy.json
 cython_debug/
 
 # Cython intermediates
+_speedups.c
 _speedups.cpp
 _speedups.html
 

_speedups.pyx

@@ -1,5 +1,6 @@
 #cython: language_level=3
-#distutils: language = c++
+#distutils: language = c
+#distutils: depends = intset.h
 
 """
 Provides faster implementation of some core parts.
@@ -13,7 +14,6 @@ from cpython cimport PyObject
 from typing import Any, Dict, Iterable, Iterator, Generator, Sequence, Tuple, TypeVar, Union, Set, List, TYPE_CHECKING
 from cymem.cymem cimport Pool
 from libc.stdint cimport int64_t, uint32_t
-from libcpp.set cimport set as std_set
 from collections import defaultdict
 
 cdef extern from *:
@@ -31,6 +31,27 @@ ctypedef int64_t ap_id_t
 cdef ap_player_t MAX_PLAYER_ID = 1000000  # limit the size of indexing array
 cdef size_t INVALID_SIZE = <size_t>(-1)  # this is all 0xff... adding 1 results in 0, but it's not negative
 
+# configure INTSET for player
+cdef extern from *:
+    """
+    #define INTSET_NAME ap_player_set
+    #define INTSET_TYPE uint32_t  // has to match ap_player_t
+    """
+
+# create INTSET for player
+cdef extern from "intset.h":
+    """
+    #undef INTSET_NAME
+    #undef INTSET_TYPE
+    """
+    ctypedef struct ap_player_set:
+        pass
+
+    ap_player_set* ap_player_set_new(size_t bucket_count) nogil
+    void ap_player_set_free(ap_player_set* set) nogil
+    bint ap_player_set_add(ap_player_set* set, ap_player_t val) nogil
+    bint ap_player_set_contains(ap_player_set* set, ap_player_t val) nogil
+
 
 cdef struct LocationEntry:
     # layout is so that
@@ -185,7 +206,7 @@ cdef class LocationStore:
     def find_item(self, slots: Set[int], seeked_item_id: int) -> Generator[Tuple[int, int, int, int, int], None, None]:
         cdef ap_id_t item = seeked_item_id
         cdef ap_player_t receiver
-        cdef std_set[ap_player_t] receivers
+        cdef ap_player_set* receivers
         cdef size_t slot_count = len(slots)
         if slot_count == 1:
             # specialized implementation for single slot
@@ -197,13 +218,20 @@ cdef class LocationStore:
                             yield entry.sender, entry.location, entry.item, entry.receiver, entry.flags
         elif slot_count:
             # generic implementation with lookup in set
-            for receiver in slots:
+            receivers = ap_player_set_new(min(1023, slot_count))  # limit top level struct to 16KB
-                receivers.insert(receiver)
+            if not receivers:
-            with nogil:
+                raise MemoryError()
-                for entry in self.entries[:self.entry_count]:
+            try:
-                    if entry.item == item and receivers.count(entry.receiver):
+                for receiver in slots:
-                        with gil:
+                    if not ap_player_set_add(receivers, receiver):
-                            yield entry.sender, entry.location, entry.item, entry.receiver, entry.flags
+                        raise MemoryError()
+                with nogil:
+                    for entry in self.entries[:self.entry_count]:
+                        if entry.item == item and ap_player_set_contains(receivers, entry.receiver):
+                            with gil:
+                                yield entry.sender, entry.location, entry.item, entry.receiver, entry.flags
+            finally:
+                ap_player_set_free(receivers)
 
     def get_for_player(self, slot: int) -> Dict[int, Set[int]]:
         cdef ap_player_t receiver = slot

_speedups.pyxbld

@@ -1,8 +1,10 @@
-# This file is required to get pyximport to work with C++.
+# This file is used when doing pyximport
-# Switching from std::set to a pure C implementation is still on the table to simplify everything.
+import os
 
 def make_ext(modname, pyxfilename):
     from distutils.extension import Extension
     return Extension(name=modname,
                      sources=[pyxfilename],
-                     language='c++')
+                     depends=["intset.h"],
+                     include_dirs=[os.getcwd()],
+                     language="c")

intset.h

@@ -0,0 +1,135 @@
+/* A specialized unordered_set implementation for literals, where bucket_count
+ * is defined at initialization rather than increased automatically.
+ */
+#include <stddef.h>
+#include <stdbool.h>
+#include <stdlib.h>
+#include <string.h>
+
+#ifndef INTSET_NAME
+#error "Please #define INTSET_NAME ... before including intset.h"
+#endif
+
+#ifndef INTSET_TYPE
+#error "Please #define INTSET_TYPE ... before including intset.h"
+#endif
+
+/* macros to generate unique names from INTSET_NAME */
+#ifndef INTSET_CONCAT
+#define INTSET_CONCAT_(a, b) a ## b
+#define INTSET_CONCAT(a, b) INTSET_CONCAT_(a, b)
+#define INTSET_FUNC_(a, b) INTSET_CONCAT(a, _ ## b)
+#endif
+
+#define INTSET_FUNC(name) INTSET_FUNC_(INTSET_NAME, name)
+#define INTSET_BUCKET INTSET_CONCAT(INTSET_NAME, Bucket)
+#define INTSET_UNION INTSET_CONCAT(INTSET_NAME, Union)
+
+#if defined(_MSC_VER)
+#pragma warning(push)
+#pragma warning(disable : 4200)
+#endif
+
+
+typedef struct {
+    size_t count;
+    union INTSET_UNION {
+        INTSET_TYPE val;
+        INTSET_TYPE *data;
+    } v;
+} INTSET_BUCKET;
+
+typedef struct {
+    size_t bucket_count;
+    INTSET_BUCKET buckets[];
+} INTSET_NAME;
+
+static INTSET_NAME *INTSET_FUNC(new)(size_t buckets)
+{
+    size_t i, size;
+    INTSET_NAME *set;
+
+    if (buckets < 1)
+        buckets = 1;
+    if ((SIZE_MAX - sizeof(INTSET_NAME)) / sizeof(INTSET_BUCKET) < buckets)
+        return NULL;
+    size = sizeof(INTSET_NAME) + buckets * sizeof(INTSET_BUCKET);
+    set = (INTSET_NAME*)malloc(size);
+    if (!set)
+        return NULL;
+    memset(set, 0, size); /* gcc -fanalyzer does not understand this sets all buckets' count to 0 */
+    for (i = 0; i < buckets; i++) {
+        set->buckets[i].count = 0;
+    }
+    set->bucket_count = buckets;
+    return set;
+}
+
+static void INTSET_FUNC(free)(INTSET_NAME *set)
+{
+    size_t i;
+    if (!set)
+        return;
+    for (i = 0; i < set->bucket_count; i++) {
+        if (set->buckets[i].count > 1)
+            free(set->buckets[i].v.data);
+    }
+    free(set);
+}
+
+static bool INTSET_FUNC(contains)(INTSET_NAME *set, INTSET_TYPE val)
+{
+    size_t i;
+    INTSET_BUCKET* bucket = &set->buckets[(size_t)val % set->bucket_count];
+    if (bucket->count == 1)
+        return bucket->v.val == val;
+    for (i = 0; i < bucket->count; ++i) {
+        if (bucket->v.data[i] == val)
+            return true;
+    }
+    return false;
+}
+
+static bool INTSET_FUNC(add)(INTSET_NAME *set, INTSET_TYPE val)
+{
+    INTSET_BUCKET* bucket;
+
+    if (INTSET_FUNC(contains)(set, val))
+        return true; /* ok */
+
+    bucket = &set->buckets[(size_t)val % set->bucket_count];
+    if (bucket->count == 0) {
+        bucket->v.val = val;
+        bucket->count = 1;
+    } else if (bucket->count == 1) {
+        INTSET_TYPE old = bucket->v.val;
+        bucket->v.data = (INTSET_TYPE*)malloc(2 * sizeof(INTSET_TYPE));
+        if (!bucket->v.data) {
+            bucket->v.val = old;
+            return false; /* error */
+        }
+        bucket->v.data[0] = old;
+        bucket->v.data[1] = val;
+        bucket->count = 2;
+    } else {
+        size_t new_bucket_size;
+        INTSET_TYPE* new_bucket_data;
+
+        new_bucket_size = (bucket->count + 1) * sizeof(INTSET_TYPE);
+        new_bucket_data = (INTSET_TYPE*)realloc(bucket->v.data, new_bucket_size);
+        if (!new_bucket_data)
+            return false; /* error */
+        bucket->v.data = new_bucket_data;
+        bucket->v.data[bucket->count++] = val;
+    }
+    return true; /* success */
+}
+
+
+#if defined(_MSC_VER)
+#pragma warning(pop)
+#endif
+
+#undef INTSET_FUNC
+#undef INTSET_BUCKET
+#undef INTSET_UNION

test/cpp/CMakeLists.txt

@@ -0,0 +1,49 @@
+cmake_minimum_required(VERSION 3.5)
+project(ap-cpp-tests)
+
+enable_testing()
+
+find_package(GTest REQUIRED)
+
+if (CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
+    add_definitions("/source-charset:utf-8")
+    set(CMAKE_CXX_FLAGS_DEBUG "/MTd")
+    set(CMAKE_CXX_FLAGS_RELEASE "/MT")
+elseif (CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
+    # enable static analysis for gcc
+    add_compile_options(-fanalyzer -Werror)
+    # disable stuff that gets triggered by googletest
+    add_compile_options(-Wno-analyzer-malloc-leak)
+    # enable asan for gcc
+    add_compile_options(-fsanitize=address)
+    add_link_options(-fsanitize=address)
+endif ()
+
+add_executable(test_default)
+
+target_include_directories(test_default
+        PRIVATE
+        ${GTEST_INCLUDE_DIRS}
+)
+
+target_link_libraries(test_default
+        ${GTEST_BOTH_LIBRARIES}
+)
+
+add_test(
+        NAME test_default
+        COMMAND  test_default
+)
+
+set_property(
+        TEST test_default
+        PROPERTY ENVIRONMENT "ASAN_OPTIONS=allocator_may_return_null=1"
+)
+
+file(GLOB ITEMS *)
+foreach(item ${ITEMS})
+    if(IS_DIRECTORY ${item} AND EXISTS ${item}/CMakeLists.txt)
+        message(${item})
+        add_subdirectory(${item})
+    endif()
+endforeach()

test/cpp/README.md

@@ -0,0 +1,32 @@
+# C++ tests
+
+Test framework for C and C++ code in AP.
+
+## Adding a Test
+
+### GoogleTest
+
+Adding GoogleTests is as simple as creating a directory with
+* one or more `test_*.cpp` files that define tests using
+  [GoogleTest API](https://google.github.io/googletest/)
+* a `CMakeLists.txt` that adds the .cpp files to `test_default` target using
+  [target_sources](https://cmake.org/cmake/help/latest/command/target_sources.html)
+
+### CTest
+
+If either GoogleTest is not suitable for the test or the build flags / sources / libraries are incompatible,
+you can add another CTest to the project using add_target and add_test, similar to how it's done for `test_default`.
+
+## Running Tests
+
+* Install [CMake](https://cmake.org/).
+* Build and/or install GoogleTest and make sure
+  [CMake can find it](https://cmake.org/cmake/help/latest/module/FindGTest.html), or
+  [create a parent `CMakeLists.txt` that fetches GoogleTest](https://google.github.io/googletest/quickstart-cmake.html).
+* Enter the directory with the top-most `CMakeLists.txt` and run
+  ```sh
+  mkdir build
+  cmake -S . -B build/ -DCMAKE_BUILD_TYPE=Release
+  cmake --build build/ --config Release && \
+  ctest --test-dir build/ -C Release --output-on-failure
+  ```

test/cpp/intset/CMakeLists.txt

@@ -0,0 +1,4 @@
+target_sources(test_default
+        PRIVATE
+        ${CMAKE_CURRENT_SOURCE_DIR}/test_intset.cpp
+)

test/cpp/intset/test_intset.cpp

@@ -0,0 +1,105 @@
+#include <limits>
+#include <cstdint>
+#include <gtest/gtest.h>
+
+// uint32Set
+#define INTSET_NAME uint32Set
+#define INTSET_TYPE uint32_t
+#include "../../../intset.h"
+#undef INTSET_NAME
+#undef INTSET_TYPE
+
+// int64Set
+#define INTSET_NAME int64Set
+#define INTSET_TYPE int64_t
+#include "../../../intset.h"
+
+
+TEST(IntsetTest, ZeroBuckets)
+{
+    // trying to allocate with zero buckets has to either fail or be functioning
+    uint32Set *set = uint32Set_new(0);
+    if (!set)
+        return; // failed -> OK
+
+    EXPECT_FALSE(uint32Set_contains(set, 1));
+    EXPECT_TRUE(uint32Set_add(set, 1));
+    EXPECT_TRUE(uint32Set_contains(set, 1));
+    uint32Set_free(set);
+}
+
+TEST(IntsetTest, Duplicate)
+{
+    // adding the same number again can't fail
+    uint32Set *set = uint32Set_new(2);
+    ASSERT_TRUE(set);
+    EXPECT_TRUE(uint32Set_add(set, 0));
+    EXPECT_TRUE(uint32Set_add(set, 0));
+    EXPECT_TRUE(uint32Set_contains(set, 0));
+    uint32Set_free(set);
+}
+
+TEST(IntsetTest, SetAllocFailure)
+{
+    // try to allocate 100TB of RAM, should fail and return NULL
+    if (sizeof(size_t) < 8)
+        GTEST_SKIP() << "Alloc error not testable on 32bit";
+    int64Set *set = int64Set_new(6250000000000ULL);
+    EXPECT_FALSE(set);
+    int64Set_free(set);
+}
+
+TEST(IntsetTest, SetAllocOverflow)
+{
+    // try to overflow argument passed to malloc
+    int64Set *set = int64Set_new(std::numeric_limits<size_t>::max());
+    EXPECT_FALSE(set);
+    int64Set_free(set);
+}
+
+TEST(IntsetTest, NullFree)
+{
+    // free(NULL) should not try to free buckets
+    uint32Set_free(NULL);
+    int64Set_free(NULL);
+}
+
+TEST(IntsetTest, BucketRealloc)
+{
+    // add a couple of values to the same bucket to test growing the bucket
+    uint32Set* set = uint32Set_new(1);
+    ASSERT_TRUE(set);
+    EXPECT_FALSE(uint32Set_contains(set, 0));
+    EXPECT_TRUE(uint32Set_add(set, 0));
+    EXPECT_TRUE(uint32Set_contains(set, 0));
+    for (uint32_t i = 1; i < 32; ++i) {
+        EXPECT_TRUE(uint32Set_add(set, i));
+        EXPECT_TRUE(uint32Set_contains(set, i - 1));
+        EXPECT_TRUE(uint32Set_contains(set, i));
+        EXPECT_FALSE(uint32Set_contains(set, i + 1));
+    }
+    uint32Set_free(set);
+}
+
+TEST(IntSet, Max)
+{
+    constexpr auto n = std::numeric_limits<uint32_t>::max();
+    uint32Set *set = uint32Set_new(1);
+    ASSERT_TRUE(set);
+    EXPECT_FALSE(uint32Set_contains(set, n));
+    EXPECT_TRUE(uint32Set_add(set, n));
+    EXPECT_TRUE(uint32Set_contains(set, n));
+    uint32Set_free(set);
+}
+
+TEST(InsetTest, Negative)
+{
+    constexpr auto n = std::numeric_limits<int64_t>::min();
+    static_assert(n < 0, "n not negative");
+    int64Set *set = int64Set_new(3);
+    ASSERT_TRUE(set);
+    EXPECT_FALSE(int64Set_contains(set, n));
+    EXPECT_TRUE(int64Set_add(set, n));
+    EXPECT_TRUE(int64Set_contains(set, n));
+    int64Set_free(set);
+}

test/netutils/test_location_store.py

@@ -1,4 +1,5 @@
 # Tests for _speedups.LocationStore and NetUtils._LocationStore
+import os
 import typing
 import unittest
 import warnings
@@ -7,6 +8,8 @@ from NetUtils import LocationStore, _LocationStore
 State = typing.Dict[typing.Tuple[int, int], typing.Set[int]]
 RawLocations = typing.Dict[int, typing.Dict[int, typing.Tuple[int, int, int]]]
 
+ci = bool(os.environ.get("CI"))  # always set in GitHub actions
+
 sample_data: RawLocations = {
     1: {
         11: (21, 2, 7),
@@ -24,6 +27,9 @@ sample_data: RawLocations = {
     3: {
         9: (99, 4, 0),
     },
+    5: {
+        9: (99, 5, 0),
+    }
 }
 
 empty_state: State = {
@@ -45,14 +51,14 @@ class Base:
         store: typing.Union[LocationStore, _LocationStore]
 
         def test_len(self) -> None:
-            self.assertEqual(len(self.store), 4)
+            self.assertEqual(len(self.store), 5)
             self.assertEqual(len(self.store[1]), 3)
 
         def test_key_error(self) -> None:
             with self.assertRaises(KeyError):
                 _ = self.store[0]
             with self.assertRaises(KeyError):
-                _ = self.store[5]
+                _ = self.store[6]
             locations = self.store[1]  # no Exception
             with self.assertRaises(KeyError):
                 _ = locations[7]
@@ -71,7 +77,7 @@ class Base:
             self.assertEqual(self.store[1].get(10, (None, None, None)), (None, None, None))
 
         def test_iter(self) -> None:
-            self.assertEqual(sorted(self.store), [1, 2, 3, 4])
+            self.assertEqual(sorted(self.store), [1, 2, 3, 4, 5])
             self.assertEqual(len(self.store), len(sample_data))
             self.assertEqual(list(self.store[1]), [11, 12, 13])
             self.assertEqual(len(self.store[1]), len(sample_data[1]))
@@ -85,13 +91,26 @@ class Base:
             self.assertEqual(sorted(self.store[1].items())[0][1], self.store[1][11])
 
         def test_find_item(self) -> None:
+            # empty player set
             self.assertEqual(sorted(self.store.find_item(set(), 99)), [])
+            # no such player, single
+            self.assertEqual(sorted(self.store.find_item({6}, 99)), [])
+            # no such player, set
+            self.assertEqual(sorted(self.store.find_item({7, 8, 9}, 99)), [])
+            # no such item
             self.assertEqual(sorted(self.store.find_item({3}, 1)), [])
-            self.assertEqual(sorted(self.store.find_item({5}, 99)), [])
+            # valid matches
             self.assertEqual(sorted(self.store.find_item({3}, 99)),
                              [(4, 9, 99, 3, 0)])
             self.assertEqual(sorted(self.store.find_item({3, 4}, 99)),
                              [(3, 9, 99, 4, 0), (4, 9, 99, 3, 0)])
+            self.assertEqual(sorted(self.store.find_item({2, 3, 4}, 99)),
+                             [(3, 9, 99, 4, 0), (4, 9, 99, 3, 0)])
+            # test hash collision in set
+            self.assertEqual(sorted(self.store.find_item({3, 5}, 99)),
+                             [(4, 9, 99, 3, 0), (5, 9, 99, 5, 0)])
+            self.assertEqual(sorted(self.store.find_item(set(range(2048)), 13)),
+                             [(1, 13, 13, 1, 0)])
 
         def test_get_for_player(self) -> None:
             self.assertEqual(self.store.get_for_player(3), {4: {9}})
@@ -196,18 +215,20 @@ class TestPurePythonLocationStoreConstructor(Base.TestLocationStoreConstructor):
         super().setUp()
 
 
-@unittest.skipIf(LocationStore is _LocationStore, "_speedups not available")
+@unittest.skipIf(LocationStore is _LocationStore and not ci, "_speedups not available")
 class TestSpeedupsLocationStore(Base.TestLocationStore):
     """Run base method tests for cython implementation."""
     def setUp(self) -> None:
+        self.assertFalse(LocationStore is _LocationStore, "Failed to load _speedups")
         self.store = LocationStore(sample_data)
         super().setUp()
 
 
-@unittest.skipIf(LocationStore is _LocationStore, "_speedups not available")
+@unittest.skipIf(LocationStore is _LocationStore and not ci, "_speedups not available")
 class TestSpeedupsLocationStoreConstructor(Base.TestLocationStoreConstructor):
     """Run base constructor tests and tests the additional constraints for cython implementation."""
     def setUp(self) -> None:
+        self.assertFalse(LocationStore is _LocationStore, "Failed to load _speedups")
         self.type = LocationStore
         super().setUp()