optimize add_8888_8888

-dmtc1: 89
+dmtc1: 85

-lwc1: 91
+lwc1: 95

-lwu: 6
+lwu: 2

Conflicts:

	pixman/pixman-mmx.c

1 files changed, 15 insertions, 4 deletions

diff --git a/pixman/pixman-mmx.c b/pixman/pixman-mmx.c
index d1e9c3e..1d18529 100644
--- a/pixman/pixman-mmx.c
+++ b/pixman/pixman-mmx.c
@@ -358,6 +358,17 @@ static __inline__ uint32_t ldl_u(uint32_t *p)
 }
 
 static force_inline __m64
+load (const uint32_t *v)
+{
+    __m64 ret;
+    asm("lwc1 %0, %1\n\t"
+	: "=f" (ret)
+	: "m" (*v)
+    );
+    return ret;
+}
+
+static force_inline __m64
 load8888 (const uint32_t *v)
 {
 #ifdef USE_LOONGSON_SIMD
@@ -2882,8 +2893,8 @@ mmx_composite_add_8888_8888 (pixman_implementation_t *imp,
 
 	while (w && (unsigned long)dst & 7)
 	{
-	    store(dst, _mm_adds_pu8 (_mm_cvtsi32_si64 (*src),
-	                             _mm_cvtsi32_si64 (*dst)));
+	    store (dst, _mm_adds_pu8 (load ((const uint32_t *)src),
+	                              load ((const uint32_t *)dst)));
 	    dst++;
 	    src++;
 	    w--;
@@ -2900,8 +2911,8 @@ mmx_composite_add_8888_8888 (pixman_implementation_t *imp,
 
 	if (w)
 	{
-	    store(dst, _mm_adds_pu8 (_mm_cvtsi32_si64 (*src),
-	                             _mm_cvtsi32_si64 (*dst)));
+	    store (dst, _mm_adds_pu8 (load ((const uint32_t *)src),
+	                              load ((const uint32_t *)dst)));
 
 	}
     }