From 87414bc8be82751a584dc8fa92f04708abf25be0 Mon Sep 17 00:00:00 2001
From: Philip Rinn <rinni@gmx.net>
Date: Fri, 29 Mar 2013 15:19:22 +0100
Subject: [PATCH] Solve bug on Windows when using OpenMP+SSE in a more elegant
 way (see issue 1806)

---
 CMakeLists.txt          | 13 ++-----------
 rtengine/CMakeLists.txt |  2 +-
 rtengine/gauss.h        | 25 ++++++++++++++++++-------
 rtengine/ipsharpen.cc   |  5 ++++-
 4 files changed, 25 insertions(+), 20 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 5b5eb37e4..d79f358ce 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -30,16 +30,7 @@ set (CACHE_NAME_SUFFIX "" CACHE STRING "RawTherapee's cache folder suffix (leave
 set (PROC_TARGET_NUMBER 0 CACHE STRING "Selected target processor from the list above (taken from ProcessorTargets.cmake)")
 
 # The following line set special compilation flags for RTEngine, and will be added to CMAKE_CXX_FLAGS
-# Due to a bug in GCC when using OpenMP+SSE, -mstackrealign will be added to your flags
-set (RTENGINE_CXX_FLAGS "" CACHE STRING "Special compilation flags for RTEngine; -mstackrealign will be added to your flags")
-
-# mandatory flags for rtengine for all platforms, depending on the bit depth
-set (RTENGINE_CXX_FLAGS_MANDATORY "-mstackrealign")
-if (CMAKE_SIZEOF_VOID_P EQUAL 4)
-	set (RTENGINE_CXX_FLAGS_MANDATORY "${RTENGINE_CXX_FLAGS_MANDATORY} -mpreferred-stack-boundary=4")
-#else (CMAKE_SIZEOF_VOID_P EQUAL 8)
-#	set (RTENGINE_CXX_FLAGS_MANDATORY "${RTENGINE_CXX_FLAGS_MANDATORY} ")
-endif (CMAKE_SIZEOF_VOID_P EQUAL 4)
+set (RTENGINE_CXX_FLAGS "" CACHE STRING "Special compilation flags for RTEngine")
 
 #loading the processor targets list
 include (ProcessorTargets.cmake)
@@ -281,7 +272,7 @@ else ()
 endif ()
 
 # Get c++ and linker flags for rtengine (the gui's c++ flags may have less flags)
-set(CXX_FLAGS "${CMAKE_CXX_FLAGS} ${CMAKE_CXX_FLAGS_${UPPER_CMAKE_BUILD_TYPE}} ${RTENGINE_CXX_FLAGS} ${RTENGINE_CXX_FLAGS_MANDATORY}")
+set(CXX_FLAGS "${CMAKE_CXX_FLAGS} ${CMAKE_CXX_FLAGS_${UPPER_CMAKE_BUILD_TYPE}} ${RTENGINE_CXX_FLAGS}")
 set(LFLAGS "${CMAKE_EXE_LINKER_FLAGS} ${CMAKE_EXE_LINKER_FLAGS_${UPPER_CMAKE_BUILD_TYPE}}")
 
 set(ABOUT_COMMAND_WITH_ARGS ${CMAKE_COMMAND}
diff --git a/rtengine/CMakeLists.txt b/rtengine/CMakeLists.txt
index 2e22f2144..dfac79d92 100644
--- a/rtengine/CMakeLists.txt
+++ b/rtengine/CMakeLists.txt
@@ -29,7 +29,7 @@ IF (BUILD_SHARED_LIBS)
     install (TARGETS rtengine DESTINATION ${LIBDIR})
 ENDIF (BUILD_SHARED_LIBS)
 
-set_target_properties (rtengine PROPERTIES COMPILE_FLAGS "${RTENGINE_CXX_FLAGS} ${RTENGINE_CXX_FLAGS_MANDATORY}")
+set_target_properties (rtengine PROPERTIES COMPILE_FLAGS "${RTENGINE_CXX_FLAGS}")
 
 target_link_libraries (rtengine rtexif ${EXTRA_LIB} ${GOBJECT_LIBRARIES} ${GTHREAD_LIBRARIES}
     ${GLIB2_LIBRARIES} ${GLIBMM_LIBRARIES} ${LCMS_LIBRARIES} ${EXPAT_LIBRARIES} ${FFTW3F_LIBRARIES} ${IPTCDATA_LIBRARIES}
diff --git a/rtengine/gauss.h b/rtengine/gauss.h
index 322fb1774..7605b8789 100644
--- a/rtengine/gauss.h
+++ b/rtengine/gauss.h
@@ -77,9 +77,12 @@ template<class T> void gaussVertical3 (T** src, T** dst, AlignedBufferMP<double>
     }
 }
 
-#ifdef __SSE__
+#ifdef __SSE__
+#ifdef WIN32
+template<class T> __attribute__((force_align_arg_pointer)) void gaussVertical3Sse (T** src, T** dst, int W, int H, const float c0, const float c1) {
+#else
 template<class T> void gaussVertical3Sse (T** src, T** dst, int W, int H, const float c0, const float c1) {
-
+#endif
     __m128 Tv,Tm1v,Tp1v;
     __m128 c0v,c1v;
     c0v = _mm_set1_ps(c0);
@@ -115,9 +118,11 @@ template<class T> void gaussVertical3Sse (T** src, T** dst, int W, int H, const
 }
 
 
-
+#ifdef WIN32
+template<class T> __attribute__((force_align_arg_pointer)) void gaussHorizontal3Sse (T** src, T** dst, int W, int H, const float c0, const float c1) {
+#else
 template<class T> void gaussHorizontal3Sse (T** src, T** dst, int W, int H, const float c0, const float c1) {
-
+#endif
     float tmp[W][4] __attribute__ ((aligned (16)));
 
     __m128 Tv,Tm1v,Tp1v;
@@ -170,8 +175,11 @@ template<class T> void gaussHorizontal3Sse (T** src, T** dst, int W, int H, cons
 
 
 // fast gaussian approximation if the support window is large
+#ifdef WIN32
+template<class T> __attribute__((force_align_arg_pointer)) void gaussHorizontalSse (T** src, T** dst, int W, int H, float sigma) {
+#else
 template<class T> void gaussHorizontalSse (T** src, T** dst, int W, int H, float sigma) {
-
+#endif
     if (sigma<0.25) {
         // dont perform filtering
         if (src!=dst)
@@ -406,9 +414,12 @@ template<class T> void gaussHorizontal (T** src, T** dst, AlignedBufferMP<double
 #endif
 }
 
-#ifdef __SSE__
+#ifdef __SSE__
+#ifdef WIN32
+template<class T> __attribute__((force_align_arg_pointer)) void gaussVerticalSse (T** src, T** dst, int W, int H, float sigma) {
+#else
 template<class T> void gaussVerticalSse (T** src, T** dst, int W, int H, float sigma) {
-
+#endif
     if (sigma<0.25) {
         // dont perform filtering
         if (src!=dst)
diff --git a/rtengine/ipsharpen.cc b/rtengine/ipsharpen.cc
index 60c066967..0421347a9 100644
--- a/rtengine/ipsharpen.cc
+++ b/rtengine/ipsharpen.cc
@@ -40,8 +40,11 @@ namespace rtengine {
 
 
 extern const Settings* settings;
-
+#if defined( __SSE__ ) && defined( WIN32 )
+__attribute__((force_align_arg_pointer)) void ImProcFunctions::dcdamping (float** aI, float** aO, float damping, int W, int H) {
+#else
 void ImProcFunctions::dcdamping (float** aI, float** aO, float damping, int W, int H) {
+#endif
 
 	const float dampingFac=-2.0/(damping*damping);