move SSE2sum code to separate cxx file...

and compile it with -arch:SSE2 if Windows.

This code however gets called only if cpuid::hasSSE2()
is true, so this does not cause problems with machines
without SSE2 support.

Change-Id: Ice23ac71d4c577b8811b08c74a3ca500a94fdc09
Reviewed-on: https://gerrit.libreoffice.org/80847
Tested-by: Jenkins
Reviewed-by: Luboš Luňák <l.lunak@collabora.com>

3 files changed, 77 insertions, 50 deletions

diff --git a/sc/Library_sc.mk b/sc/Library_sc.mk
index d594ca110f6d..9c057eaf6a6d 100644
--- a/sc/Library_sc.mk
+++ b/sc/Library_sc.mk
@@ -98,6 +98,17 @@ $(eval $(call gb_Library_use_libraries,sc,\
     xo \
 ))
 
+ifeq ($(OS),WNT)
+$(eval $(call gb_Library_add_exception_objects,sc,\
+    sc/source/core/tool/arraysumSSE2, -arch:SSE2 \
+))
+
+else
+$(eval $(call gb_Library_add_exception_objects,sc,\
+    sc/source/core/tool/arraysumSSE2 \
+))
+endif
+
 $(eval $(call gb_Library_add_exception_objects,sc,\
     sc/source/core/data/attarray \
     sc/source/core/data/attrib \
diff --git a/sc/source/core/inc/arraysumfunctor.hxx b/sc/source/core/inc/arraysumfunctor.hxx
index 7ef8a7face05..f1182874b1a1 100644
--- a/sc/source/core/inc/arraysumfunctor.hxx
+++ b/sc/source/core/inc/arraysumfunctor.hxx
@@ -14,7 +14,6 @@
 #include <cstdint>
 #include <rtl/math.hxx>
 
-#include <tools/simdsupport.hxx>
 #include <tools/simd.hxx>
 #include <tools/cpuid.hxx>
 
@@ -87,56 +86,8 @@ public:
     }
 
 private:
-    double executeSSE2(size_t& i, const double* pCurrent) const
-    {
-#if defined(LO_SSE2_AVAILABLE)
-        double fSum = 0.0;
-        size_t nRealSize = mnSize - i;
-        size_t nUnrolledSize = nRealSize - (nRealSize % 8);
-
-        if (nUnrolledSize > 0)
-        {
-            __m128d sum1 = _mm_setzero_pd();
-            __m128d sum2 = _mm_setzero_pd();
-            __m128d sum3 = _mm_setzero_pd();
-            __m128d sum4 = _mm_setzero_pd();
-
-            for (; i < nUnrolledSize; i += 8)
-            {
-                __m128d load1 = _mm_load_pd(pCurrent);
-                sum1 = _mm_add_pd(sum1, load1);
-                pCurrent += 2;
-
-                __m128d load2 = _mm_load_pd(pCurrent);
-                sum2 = _mm_add_pd(sum2, load2);
-                pCurrent += 2;
-
-                __m128d load3 = _mm_load_pd(pCurrent);
-                sum3 = _mm_add_pd(sum3, load3);
-                pCurrent += 2;
-
-                __m128d load4 = _mm_load_pd(pCurrent);
-                sum4 = _mm_add_pd(sum4, load4);
-                pCurrent += 2;
-            }
-            sum1 = _mm_add_pd(_mm_add_pd(sum1, sum2), _mm_add_pd(sum3, sum4));
-
-            double temp;
-
-            _mm_storel_pd(&temp, sum1);
-            fSum += temp;
-
-            _mm_storeh_pd(&temp, sum1);
-            fSum += temp;
-        }
-        return fSum;
-#else
-        (void) i;
-        (void) pCurrent;
-        return 0.0;
-#endif
-    }
 
+    double executeSSE2(size_t& i, const double* pCurrent) const;
     double executeUnrolled(size_t& i, const double* pCurrent) const
     {
         size_t nRealSize = mnSize - i;
diff --git a/sc/source/core/tool/arraysumSSE2.cxx b/sc/source/core/tool/arraysumSSE2.cxx
new file mode 100644
index 000000000000..894675335834
--- /dev/null
+++ b/sc/source/core/tool/arraysumSSE2.cxx
@@ -0,0 +1,65 @@
+/* -*- Mode: C++; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4 -*- */
+/*
+ * This file is part of the LibreOffice project.
+ *
+ * This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/.
+ *
+ */
+
+#include <arraysumfunctor.hxx>
+#include <tools/simdsupport.hxx>
+
+namespace sc
+{
+double ArraySumFunctor::executeSSE2(size_t& i, const double* pCurrent) const
+{
+#if defined(LO_SSE2_AVAILABLE)
+    double fSum = 0.0;
+    size_t nRealSize = mnSize - i;
+    size_t nUnrolledSize = nRealSize - (nRealSize % 8);
+
+    if (nUnrolledSize > 0)
+    {
+        __m128d sum1 = _mm_setzero_pd();
+        __m128d sum2 = _mm_setzero_pd();
+        __m128d sum3 = _mm_setzero_pd();
+        __m128d sum4 = _mm_setzero_pd();
+
+        for (; i < nUnrolledSize; i += 8)
+        {
+            __m128d load1 = _mm_load_pd(pCurrent);
+            sum1 = _mm_add_pd(sum1, load1);
+            pCurrent += 2;
+
+            __m128d load2 = _mm_load_pd(pCurrent);
+            sum2 = _mm_add_pd(sum2, load2);
+            pCurrent += 2;
+
+            __m128d load3 = _mm_load_pd(pCurrent);
+            sum3 = _mm_add_pd(sum3, load3);
+            pCurrent += 2;
+
+            __m128d load4 = _mm_load_pd(pCurrent);
+            sum4 = _mm_add_pd(sum4, load4);
+            pCurrent += 2;
+        }
+        sum1 = _mm_add_pd(_mm_add_pd(sum1, sum2), _mm_add_pd(sum3, sum4));
+
+        double temp;
+
+        _mm_storel_pd(&temp, sum1);
+        fSum += temp;
+
+        _mm_storeh_pd(&temp, sum1);
+        fSum += temp;
+    }
+    return fSum;
+#else
+    (void)i;
+    (void)pCurrent;
+    return 0.0;
+#endif
+}
+}
\ No newline at end of file