tnl: Optimize SSE load[23]f_1 since they don't need the identity swizzle.

SSE movss from memory zeroes out everything above the destination dword, so
we get the (a, 0) or (a, 0, 0) result that these functions needed.

Bug #16520.

diff --git a/src/mesa/tnl/t_vertex_sse.c b/src/mesa/tnl/t_vertex_sse.c
index d8021a3d2ca07deec976250d42014c42c5bdf2c0..07adc1ed79e854e87e0798a4bf0fbcdad980fe69 100644 (file)
--- a/src/mesa/tnl/t_vertex_sse.c
+++ b/src/mesa/tnl/t_vertex_sse.c
@@ -146,7 +146,8 @@ static void emit_load3f_1( struct x86_program *p,
                           struct x86_reg dest,
                           struct x86_reg arg0 )
 {
-   emit_load4f_1(p, dest, arg0);
+   /* Loading from memory erases the upper bits. */
+   sse_movss(&p->func, dest, arg0);
 }
 
 static void emit_load2f_2( struct x86_program *p, 
@@ -160,7 +161,8 @@ static void emit_load2f_1( struct x86_program *p,
                           struct x86_reg dest,
                           struct x86_reg arg0 )
 {
-   emit_load4f_1(p, dest, arg0);
+   /* Loading from memory erases the upper bits. */
+   sse_movss(&p->func, dest, arg0);
 }
 
 static void emit_load1f_1( struct x86_program *p,