alternate matrix transpose code-path for pre-SSE4.1

CMakeLists.txt

@@ -6,11 +6,7 @@ if (NOT DEFINED ATHENA_INCLUDE_DIR)
 endif()
 include_directories(include ${ATHENA_INCLUDE_DIR})
 
-if(NOT MSVC)
+set(SOURCES
-set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -msse4.1 -msse4.2 -std=c++14")
-endif()
-
-add_library(zeus
     src/CVector3f.cpp
     src/Math.cpp
     src/CQuaternion.cpp
@@ -23,7 +19,30 @@ add_library(zeus
     src/CRectangle.cpp
     src/CVector4f.cpp
     src/CMatrix4f.cpp
-    src/CAABox.cpp
+    src/CAABox.cpp)
+
+# SSELegacy.cpp compiled separately to escape the effects of link-time optimization
+if(NOT MSVC)
+set_source_files_properties(${SOURCES} PROPERTIES COMPILE_FLAGS "-msse4.1 -msse4.2 -std=c++14")
+string(REPLACE "-flto" "" CUSTOM_FLAGS ${CMAKE_CXX_FLAGS})
+set(CUSTOM_FLAGS "${CUSTOM_FLAGS} -O3 -msse3 -std=c++14")
+else()
+string(REPLACE "/GL" "" CUSTOM_FLAGS ${CMAKE_CXX_FLAGS})
+endif()
+separate_arguments(CUSTOM_FLAGS UNIX_COMMAND ${CUSTOM_FLAGS})
+add_custom_command(
+    OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/src/SSELegacy.o
+    COMMAND ${CMAKE_CXX_COMPILER}
+    ARGS ${CUSTOM_FLAGS} -c ${CMAKE_CURRENT_SOURCE_DIR}/src/SSELegacy.cpp
+                         -o ${CMAKE_CURRENT_BINARY_DIR}/src/SSELegacy.o
+                         -I ${CMAKE_CURRENT_SOURCE_DIR}/include
+                         -I ${ATHENA_INCLUDE_DIR}
+
+    MAIN_DEPENDENCY src/SSELegacy.cpp)
+
+add_library(zeus
+    ${SOURCES}
+    ${CMAKE_CURRENT_BINARY_DIR}/src/SSELegacy.o
 
     include/zeus/Math.hpp
     include/zeus/CQuaternion.hpp

include/zeus/CMatrix3f.hpp

@@ -137,7 +137,9 @@ public:
     static const CMatrix3f skIdentityMatrix3f;
     
     void transpose();
+    void transposeSSE3();
     CMatrix3f transposed() const;
+    CMatrix3f transposedSSE3() const;
 
     inline void invert() {*this = inverted();}
     CMatrix3f inverted() const;

include/zeus/CMatrix4f.hpp

@@ -96,41 +96,8 @@ public:
         return vec[i];
     }
 
-    inline CMatrix4f transposed()
+    CMatrix4f transposed() const;
-    {
+    CMatrix4f transposedSSE3() const;
-        CMatrix4f ret;
-#if __SSE__
-        __m128 T0 = _mm_unpacklo_ps(vec[0].mVec128, vec[1].mVec128);
-        __m128 T2 = _mm_unpacklo_ps(vec[2].mVec128, vec[3].mVec128);
-        __m128 T1 = _mm_unpackhi_ps(vec[0].mVec128, vec[1].mVec128);
-        __m128 T3 = _mm_unpackhi_ps(vec[2].mVec128, vec[3].mVec128);
-        ret.vec[0].mVec128 = _mm_movelh_ps(T0, T2);
-        ret.vec[1].mVec128 = _mm_movehl_ps(T2, T0);
-        ret.vec[2].mVec128 = _mm_movelh_ps(T1, T3);
-        ret.vec[3].mVec128 = _mm_movehl_ps(T3, T1);
-#else
-        ret.m[0][0] = m[0][0];
-        ret.m[1][0] = m[0][1];
-        ret.m[2][0] = m[0][2];
-        ret.m[3][0] = m[0][3];
-
-        ret.m[0][1] = m[1][0];
-        ret.m[1][1] = m[1][1];
-        ret.m[2][1] = m[1][2];
-        ret.m[3][1] = m[1][3];
-
-        ret.m[0][2] = m[2][0];
-        ret.m[1][2] = m[2][1];
-        ret.m[2][2] = m[2][2];
-        ret.m[3][2] = m[2][3];
-
-        ret.m[0][3] = m[3][0];
-        ret.m[1][3] = m[3][1];
-        ret.m[2][3] = m[3][2];
-        ret.m[3][3] = m[3][3];
-#endif
-        return ret;
-    }
 
     inline CVector3f multiplyOneOverW(const CVector3f& point) const
     {

src/CMatrix3f.cpp

@@ -33,6 +33,11 @@ CMatrix3f::CMatrix3f(const CQuaternion& quat)
 void CMatrix3f::transpose()
 {
 #if __SSE__
+    if (!cpuFeatures().SSE41)
+    {
+        transposeSSE3();
+        return;
+    }
     __m128 zero = _mm_xor_ps(vec[0].mVec128, vec[0].mVec128);
     __m128 T0 = _mm_unpacklo_ps(vec[0].mVec128, vec[1].mVec128);
     __m128 T2 = _mm_unpacklo_ps(vec[2].mVec128, zero);
@@ -61,6 +66,8 @@ void CMatrix3f::transpose()
 CMatrix3f CMatrix3f::transposed() const
 {
 #if __SSE__
+    if (!cpuFeatures().SSE41)
+        return transposedSSE3();
     __m128 zero = _mm_xor_ps(vec[0].mVec128, vec[0].mVec128);
     __m128 T0 = _mm_unpacklo_ps(vec[0].mVec128, vec[1].mVec128);
     __m128 T2 = _mm_unpacklo_ps(vec[2].mVec128, zero);

src/CMatrix4f.cpp

@@ -1,3 +1,46 @@
 #include "zeus/CMatrix4f.hpp"
 
-const zeus::CMatrix4f zeus::CMatrix4f::skIdentityMatrix4f = CMatrix4f();
+namespace zeus
+{
+
+const CMatrix4f CMatrix4f::skIdentityMatrix4f = CMatrix4f();
+
+CMatrix4f CMatrix4f::transposed() const
+{
+    CMatrix4f ret;
+#if __SSE__
+    if (!cpuFeatures().SSE41)
+        return transposedSSE3();
+    __m128 T0 = _mm_unpacklo_ps(vec[0].mVec128, vec[1].mVec128);
+    __m128 T2 = _mm_unpacklo_ps(vec[2].mVec128, vec[3].mVec128);
+    __m128 T1 = _mm_unpackhi_ps(vec[0].mVec128, vec[1].mVec128);
+    __m128 T3 = _mm_unpackhi_ps(vec[2].mVec128, vec[3].mVec128);
+    ret.vec[0].mVec128 = _mm_movelh_ps(T0, T2);
+    ret.vec[1].mVec128 = _mm_movehl_ps(T2, T0);
+    ret.vec[2].mVec128 = _mm_movelh_ps(T1, T3);
+    ret.vec[3].mVec128 = _mm_movehl_ps(T3, T1);
+#else
+    ret.m[0][0] = m[0][0];
+    ret.m[1][0] = m[0][1];
+    ret.m[2][0] = m[0][2];
+    ret.m[3][0] = m[0][3];
+
+    ret.m[0][1] = m[1][0];
+    ret.m[1][1] = m[1][1];
+    ret.m[2][1] = m[1][2];
+    ret.m[3][1] = m[1][3];
+
+    ret.m[0][2] = m[2][0];
+    ret.m[1][2] = m[2][1];
+    ret.m[2][2] = m[2][2];
+    ret.m[3][2] = m[2][3];
+
+    ret.m[0][3] = m[3][0];
+    ret.m[1][3] = m[3][1];
+    ret.m[2][3] = m[3][2];
+    ret.m[3][3] = m[3][3];
+#endif
+    return ret;
+}
+
+}

src/SSELegacy.cpp

@@ -0,0 +1,97 @@
+#include "zeus/CMatrix3f.hpp"
+#include "zeus/CMatrix4f.hpp"
+
+namespace zeus
+{
+
+void CMatrix3f::transposeSSE3()
+{
+#if __SSE__
+    __m128 zero = _mm_xor_ps(vec[0].mVec128, vec[0].mVec128);
+    __m128 T0 = _mm_unpacklo_ps(vec[0].mVec128, vec[1].mVec128);
+    __m128 T2 = _mm_unpacklo_ps(vec[2].mVec128, zero);
+    __m128 T1 = _mm_unpackhi_ps(vec[0].mVec128, vec[1].mVec128);
+    __m128 T3 = _mm_unpackhi_ps(vec[2].mVec128, zero);
+    vec[0].mVec128 = _mm_movelh_ps(T0, T2);
+    vec[1].mVec128 = _mm_movehl_ps(T2, T0);
+    vec[2].mVec128 = _mm_movelh_ps(T1, T3);
+#else
+    float tmp;
+
+    tmp = m[0][1];
+    m[0][1] = m[1][0];
+    m[1][0] = tmp;
+
+    tmp = m[0][2];
+    m[0][2] = m[2][0];
+    m[2][0] = tmp;
+
+    tmp = m[1][2];
+    m[1][2] = m[2][1];
+    m[2][1] = tmp;
+#endif
+}
+
+CMatrix3f CMatrix3f::transposedSSE3() const
+{
+#if __SSE__
+    __m128 zero = _mm_xor_ps(vec[0].mVec128, vec[0].mVec128);
+    __m128 T0 = _mm_unpacklo_ps(vec[0].mVec128, vec[1].mVec128);
+    __m128 T2 = _mm_unpacklo_ps(vec[2].mVec128, zero);
+    __m128 T1 = _mm_unpackhi_ps(vec[0].mVec128, vec[1].mVec128);
+    __m128 T3 = _mm_unpackhi_ps(vec[2].mVec128, zero);
+    return CMatrix3f(_mm_movelh_ps(T0, T2), _mm_movehl_ps(T2, T0), _mm_movelh_ps(T1, T3));
+#else
+    CMatrix3f ret(*this);
+    float tmp;
+
+    tmp = ret.m[0][1];
+    ret.m[0][1] = ret.m[1][0];
+    ret.m[1][0] = tmp;
+
+    tmp = m[0][2];
+    ret.m[0][2] = ret.m[2][0];
+    ret.m[2][0] = tmp;
+
+    tmp = m[1][2];
+    ret.m[1][2] = ret.m[2][1];
+    ret.m[2][1] = tmp;
+
+    return ret;
+#endif
+}
+
+CMatrix4f CMatrix4f::transposedSSE3() const
+{
+#if __SSE__
+    CMatrix3f ret;
+    __m128 T0 = _mm_unpacklo_ps(vec[0].mVec128, vec[1].mVec128);
+    __m128 T2 = _mm_unpacklo_ps(vec[2].mVec128, vec[3].mVec128);
+    __m128 T1 = _mm_unpackhi_ps(vec[0].mVec128, vec[1].mVec128);
+    __m128 T3 = _mm_unpackhi_ps(vec[2].mVec128, vec[3].mVec128);
+    ret.vec[0].mVec128 = _mm_movelh_ps(T0, T2);
+    ret.vec[1].mVec128 = _mm_movehl_ps(T2, T0);
+    ret.vec[2].mVec128 = _mm_movelh_ps(T1, T3);
+    ret.vec[3].mVec128 = _mm_movehl_ps(T3, T1);
+    return ret;
+#else
+    float tmp;
+    CMatrix3f ret(*this);
+
+    tmp = ret.m[0][1];
+    ret.m[0][1] = ret.m[1][0];
+    ret.m[1][0] = tmp;
+
+    tmp = m[0][2];
+    ret.m[0][2] = ret.m[2][0];
+    ret.m[2][0] = tmp;
+
+    tmp = m[1][2];
+    ret.m[1][2] = ret.m[2][1];
+    ret.m[2][1] = tmp;
+
+    return ret;
+#endif
+}
+
+}