shithub: cstory

--- a/CMakeLists.txt

+++ b/CMakeLists.txt

@@ -16,6 +16,8 @@

 set(RENDERER "SDLTexture" CACHE STRING "Which renderer the game should use: 'OpenGL3' for an OpenGL 3.2 renderer, 'OpenGLES2' for an OpenGL ES 2.0 renderer, 'SDLTexture' for SDL2's hardware-accelerated Texture API, 'SDLSurface' for SDL2's software-rendered Surface API, or 'Software' for a handwritten software renderer")

 option(LTO "Enable link-time optimisation" OFF)

+option(NATIVE_OPTIMIZATIONS "Enable processor-specific optimisations (executable might not work on other architectures) (GCC-compatible compilers only)" OFF)

 option(WARNINGS "Enable common compiler warnings (for GCC-compatible compilers and MSVC only)" OFF)

 option(WARNINGS_ALL "Enable ALL compiler warnings (for Clang and MSVC only)" OFF)

 option(WARNINGS_FATAL "Stop compilation on any compiler warning (for GCC-compatible compilers and MSVC only)" OFF)

@@ -446,6 +448,25 @@

 		check_ipo_supported(RESULT result)

 		if(result)

 			set_target_properties(CSE2 PROPERTIES INTERPROCEDURAL_OPTIMIZATION TRUE)

+		endif()

+	endif()

+endif()

+# Enable -march=native if available

+if(NATIVE_OPTIMIZATIONS)

+	include(CheckCXXCompilerFlag)

+	CHECK_CXX_COMPILER_FLAG("-march=native" COMPILER_SUPPORTS_MARCH_NATIVE)	# GCC flag

+	if(COMPILER_SUPPORTS_MARCH_NATIVE)

+		target_compile_options(CSE2 PRIVATE -march=native)

+	else()

+		CHECK_CXX_COMPILER_FLAG("-xHost" COMPILER_SUPPORTS_XHOST)	# ICC (Linux) flag

+		CHECK_CXX_COMPILER_FLAG("/QxHost" COMPILER_SUPPORTS_QXHOST)	# ICC (Windows) flag

+		if(COMPILER_SUPPORTS_XHOST)

+			target_compile_options(CSE2 PRIVATE -xHost)

+		elseif(COMPILER_SUPPORTS_QXHOST)

+			target_compile_options(CSE2 PRIVATE /QxHost)

+		else()

+			message(WARNING "Couldn't activate native optimizations ! (Unsupported compiler)")

 		endif()

 	endif()

 endif()

--- a/README.md

+++ b/README.md

@@ -61,6 +61,7 @@

 Name | Function

 --------|--------

 `-DLTO=ON` | Enable link-time optimisation

+`-DNATIVE_OPTIMIZATIONS=ON` | Enable processor-specific optimisations (executable might not work on other architectures) (GCC-compatible compilers only)

 `-DJAPANESE=ON` | Enable the Japanese-language build (instead of the unofficial Aeon Genesis English translation)

 `-DFIX_BUGS=ON` | Fix various bugs in the game

 `-DDEBUG_SAVE=ON` | Re-enable the ability to drag-and-drop save files onto the window

--- a/src/Backends/Audio/SDL2.cpp

+++ b/src/Backends/Audio/SDL2.cpp

@@ -12,6 +12,7 @@

 #include "../../Organya.h"

 #include "../../WindowsWrapper.h"

+#include "../../CommonDefines.h"

 #define MIN(a, b) ((a) < (b) ? (a) : (b))

 #define MAX(a, b) ((a) > (b) ? (a) : (b))

@@ -70,7 +71,8 @@

 	sound->volume_r = sound->pan_r * sound->volume;

-static void MixSounds(float *stream, unsigned int frames_total)

+// Most CPU-intensive function in the game (2/3rd CPU time consumption in my experience), so marked with attrHot so the compiler considers it a hot spot (as it is) when optimizing

+attrHot static void MixSounds(float *stream, unsigned int frames_total)

 	for (AudioBackend_Sound *sound = sound_list_head; sound != NULL; sound = sound->next)

@@ -85,7 +87,7 @@

 				const float sample2 = (sound->samples[(size_t)sound->position + 1] - 128.0f) / 128.0f;

 				// Perform linear interpolation

-				const float interpolated_sample = sample1 + ((sample2 - sample1) * (float)fmod(sound->position, 1.0));

+				const float interpolated_sample = sample1 + ((sample2 - sample1) * (float)fmod((float)sound->position, 1.0f));

 				*steam_pointer++ += interpolated_sample * sound->volume_l;

 				*steam_pointer++ += interpolated_sample * sound->volume_r;

--- a/src/CommonDefines.h

+++ b/src/CommonDefines.h

@@ -10,6 +10,10 @@

 #define TILES_TO_UNITS(x) ((int)((x) * (0x200 * 0x10)))

 #define UNITS_TO_TILES(x) ((int)((x) / (0x200 * 0x10)))

+#ifdef __GNUC__

+#define attrHot __attribute__((hot))

+#endif

 enum Collisions

 	COLL_LEFT_WALL = 1,     // Touching a left wall