Improve SSE intrinsic support.

.github/workflows/linux.yml

@@ -12,10 +12,10 @@ jobs:
     strategy:
       matrix:
         platform: [clang, gcc]
-        os: [ubuntu-20.04, ubuntu-22.04]
+        os: [ubuntu-22.04, ubuntu-22.04-arm, ubuntu-24.04, ubuntu-24.04-arm, ubuntu-26.04, ubuntu-26.04-arm]
 
     steps:
-    - uses: actions/checkout@v3.3.0
+    - uses: actions/checkout@v7.0.1
       with:
         submodules: 'true'
         fetch-depth: 0

.github/workflows/macos.yml

@@ -8,10 +8,13 @@ on:
 
 jobs:
   macos:
-    runs-on: macos-10.15
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [macos-15, macos-15-intel, macos-26, macos-26-intel]
 
     steps:
-    - uses: actions/checkout@v3.3.0
+    - uses: actions/checkout@v7.0.1
       with:
         submodules: 'true'
         fetch-depth: 0

.github/workflows/windows.yml

@@ -8,18 +8,20 @@ on:
 
 jobs:
   windows_msvc:
-    runs-on: windows-2019
+    runs-on: ${{ matrix.os }}
     strategy:
       matrix:
         platform: [x86, amd64]
+        os: [windows-2022, windows-2025]
+
     steps:
-    - uses: actions/checkout@v3.3.0
+    - uses: actions/checkout@v7.0.1
       with:
         submodules: 'true'
         fetch-depth: 0
         ref: ${{ github.event.pull_request.head.sha }}
 
-    - uses: ilammy/msvc-dev-cmd@v1.12.1
+    - uses: ilammy/msvc-dev-cmd@v1.13.0
       with:
         arch: ${{ matrix.platform }}
 

.gitignore

@@ -34,3 +34,9 @@ CPackSourceConfig.cmake
 *.tmp
 *.bak
 thumbs.db
+
+/CMakeUserPresets.json
+/.cache
+.clang-format
+/.vscode
+/.vs

CMakeLists.txt

@@ -1,10 +1,5 @@
 cmake_minimum_required(VERSION 3.14)
 
-if (NOT CMAKE_BUILD_TYPE AND NOT CMAKE_CONFIGURATION_TYPES)
-    message(STATUS "No build type selected, default to Release")
-    set(CMAKE_BUILD_TYPE "Release")
-endif()
-
 project(gamemath VERSION 1.1.1 LANGUAGES C)
 
 # Decide if this project is the top level project, if so enable development targets.
@@ -214,9 +209,9 @@ if(BUILD_SHARED_LIBS)
 endif()
 
 if(CMAKE_C_COMPILER_ID MATCHES "Clang")
-    target_compile_options(gamemath PRIVATE -fno-builtin -fno-strict-aliasing -Wno-implicit-function-declaration)
+    target_compile_options(gamemath PRIVATE -fno-builtin -fno-strict-aliasing -ffp-contract=off -Wno-implicit-function-declaration)
 elseif(CMAKE_C_COMPILER_ID MATCHES "GNU")
-    target_compile_options(gamemath PRIVATE -fno-gnu89-inline -fno-builtin -fno-strict-aliasing -Wno-implicit-function-declaration)
+    target_compile_options(gamemath PRIVATE -fno-gnu89-inline -fno-builtin -fno-strict-aliasing -ffp-contract=off -Wno-implicit-function-declaration)
 elseif(MSVC)
     target_compile_options(gamemath PRIVATE /wd4146 /wd4244 /wd4554 /wd4723 /wd4756)
 endif()

double/e_sqrt.c

@@ -89,6 +89,11 @@ static	const double	one	= 1.0, tiny=1.0e-300;
 double
 __ieee754_sqrt(double x)
 {
+#ifdef USE_SSE
+	double ret;
+	_mm_store_sd(&ret, _mm_sqrt_pd(_mm_load_sd(&x)));
+	return ret;
+#else
 	double z;
 	int32_t sign = (int)0x80000000;
 	int32_t ix0,s0,q,m,t,i;
@@ -181,6 +186,7 @@ __ieee754_sqrt(double x)
 	ix0 += (m <<20);
 	INSERT_WORDS(z,ix0,ix1);
 	return z;
+#endif
 }
 
 /*

double/s_llrint.c

@@ -1,17 +1,9 @@
 #include "math_private.h"
 
-#ifdef USE_SSE
-#if (defined _M_IX86 || defined __i386__)
-#define USE_32BIT
-#endif
-
+#if defined(USE_SSE) && !(defined _M_IX86 || defined __i386__)
 long long gm_llrint(double x)
 {
-#ifdef USE_32BIT
-    return _mm_cvtsd_si32(_mm_load_sd(&x));
-#else
     return _mm_cvtsd_si64(_mm_load_sd(&x));
-#endif
 }
 #else
 #define type		double

double/s_lrint.c

@@ -7,7 +7,7 @@
 
 long gm_lrint(double x)
 {
-#ifdef USE_32BIT
+#if defined(USE_32BIT) || defined(_WIN32) // long is always 32bit on windows
     return (long)_mm_cvtsd_si32(_mm_load_sd(&x));
 #else
     return (long)_mm_cvtsd_si64(_mm_load_sd(&x));

float/e_sqrtf.c

@@ -21,6 +21,11 @@ static	const float	one	= 1.0f, tiny=1.0e-30f;
 float
 __ieee754_sqrtf(float x)
 {
+#ifdef USE_SSE
+	float ret;
+	_mm_store_ss(&ret, _mm_sqrt_ps(_mm_load_ss(&x)));
+	return ret;
+#else
 	float z;
 	int32_t sign = (int)0x80000000;
 	int32_t ix,s,q,m,t,i;
@@ -82,4 +87,5 @@ __ieee754_sqrtf(float x)
 	ix += (m <<23);
 	SET_FLOAT_WORD(z,ix);
 	return z;
+#endif
 }

float/s_llrintf.c

@@ -1,17 +1,9 @@
 #include "math_private.h"
 
-#ifdef USE_SSE
-#if (defined _M_IX86 || defined __i386__)
-#define USE_32BIT
-#endif
-
+#if defined(USE_SSE) && !(defined _M_IX86 || defined __i386__)
 long long gm_llrintf(float x)
 {
-#ifdef USE_32BIT
-    return _mm_cvtss_si32(_mm_load_ss(&x));
-#else
     return _mm_cvtss_si64(_mm_load_ss(&x));
-#endif
 }
 #else
 #define type		float

tests/CMakeLists.txt

@@ -8,4 +8,14 @@ target_link_libraries(test_lrintf_intrin PRIVATE gamemath::gamemath)
 if(CMAKE_SYSTEM_NAME STREQUAL "Linux")
     # Looks like linux needs libm linking for the fenv.h functions?
     target_link_libraries(test_lrintf_intrin PRIVATE m)
-endif()
+endif()
+
+if(CMAKE_C_COMPILER_ID MATCHES "Clang")
+    set(GM_TEST_OPTIONS -fno-builtin -fno-strict-aliasing -ffp-contract=off -Wno-implicit-function-declaration)
+elseif(CMAKE_C_COMPILER_ID MATCHES "GNU")
+    set(GM_TEST_OPTIONS -fno-gnu89-inline -fno-builtin -fno-strict-aliasing -ffp-contract=off -Wno-implicit-function-declaration)
+elseif(MSVC)
+    set(GM_TEST_OPTIONS /fp:strict /wd4146 /wd4244 /wd4554 /wd4723 /wd4756)
+endif()
+
+target_compile_options(test_lrintf_intrin PRIVATE ${GM_TEST_OPTIONS})