mmx: optimize over_n_8_0565

Load mask into vector register and improve instruction-level parallelism
by calculating independent intermediate results together.

Loongson:
over_n_8_0565 =  L1:  14.35  L2:  14.26  M: 12.00 ( 16.00%)  HT: 11.30 VT: 10.78  R: 10.34  RT:  7.12 (  75Kops/s)
over_n_8_0565 =  L1:  23.13  L2:  23.35  M: 18.03 ( 24.28%)  HT: 16.06 VT: 15.52  R: 14.08  RT:  8.29 (  83Kops/s)

1 files changed, 23 insertions, 23 deletions

diff --git a/pixman/pixman-mmx.c b/pixman/pixman-mmx.c
index 08eb25d..d526d8a 100644
--- a/pixman/pixman-mmx.c
+++ b/pixman/pixman-mmx.c
@@ -2246,35 +2246,35 @@ mmx_composite_over_n_8_0565 (pixman_implementation_t *imp,
 
 	while (w >= 4)
 	{
-	    uint64_t m0, m1, m2, m3;
-	    m0 = *mask;
-	    m1 = *(mask + 1);
-	    m2 = *(mask + 2);
-	    m3 = *(mask + 3);
+	    __m64 vmask = load8888u ((uint32_t *)mask);
 
-	    if (srca == 0xff && (m0 & m1 & m2 & m3) == 0xff)
+	    if (srca == 0xff && is_equal (vmask, MC (4x00ff)))
 	    {
 		*(__m64 *)dst = srcsrcsrcsrc;
 	    }
-	    else if (m0 | m1 | m2 | m3)
+	    else if (!is_zero (vmask))
 	    {
-		__m64 vdest;
-		__m64 vm0, vm1, vm2, vm3;
-
-		vdest = *(__m64 *)dst;
+		__m64 vdest = *(__m64 *)dst;
 
-		vm0 = to_m64 (m0);
-		vdest = pack_565 (in_over (vsrc, vsrca, expand_alpha_rev (vm0),
-					   expand565 (vdest, 0)), vdest, 0);
-		vm1 = to_m64 (m1);
-		vdest = pack_565 (in_over (vsrc, vsrca, expand_alpha_rev (vm1),
-					   expand565 (vdest, 1)), vdest, 1);
-		vm2 = to_m64 (m2);
-		vdest = pack_565 (in_over (vsrc, vsrca, expand_alpha_rev (vm2),
-					   expand565 (vdest, 2)), vdest, 2);
-		vm3 = to_m64 (m3);
-		vdest = pack_565 (in_over (vsrc, vsrca, expand_alpha_rev (vm3),
-					   expand565 (vdest, 3)), vdest, 3);
+		const __m64 vm0 = _mm_shuffle_pi16 (vmask, _MM_SHUFFLE (0, 0, 0, 0));
+		const __m64 vm1 = _mm_shuffle_pi16 (vmask, _MM_SHUFFLE (1, 1, 1, 1));
+		const __m64 vm2 = _mm_shuffle_pi16 (vmask, _MM_SHUFFLE (2, 2, 2, 2));
+		const __m64 vm3 = _mm_shuffle_pi16 (vmask, _MM_SHUFFLE (3, 3, 3, 3));
+		
+		const __m64 vd0 = expand565 (vdest, 0);
+		const __m64 vd1 = expand565 (vdest, 1);
+		const __m64 vd2 = expand565 (vdest, 2);
+		const __m64 vd3 = expand565 (vdest, 3);
+		
+		const __m64 io0 = in_over (vsrc, vsrca, vm0, vd0);
+		const __m64 io1 = in_over (vsrc, vsrca, vm1, vd1);
+		const __m64 io2 = in_over (vsrc, vsrca, vm2, vd2);
+		const __m64 io3 = in_over (vsrc, vsrca, vm3, vd3);
+
+		vdest = pack_565 (io0, vdest, 0);
+		vdest = pack_565 (io1, vdest, 1);
+		vdest = pack_565 (io2, vdest, 2);
+		vdest = pack_565 (io3, vdest, 3);
 
 		*(__m64 *)dst = vdest;
 	    }