pngdsp x86: use unaligned access

For test images manually generated to contain only up prediction,
timing results:
         8380x3032    255x185
before:   138635       1992
after:    139232       1996

Actually jumping to the proper version depending on the alignment:
8380x3032: 138767

A 0.5% speed improvement for gigantic images is not worth the code
duplication.

Fixes ticket #4148

Signed-off-by: Christophe Gisquet <christophe.gisquet@gmail.com>
Tested-by: Benoit Fouet <benoit.fouet@free.fr>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>

libavcodec/pngdsp.h

@@ -25,9 +25,9 @@
 #include <stdint.h>
 
 typedef struct PNGDSPContext {
-    void (*add_bytes_l2)(uint8_t *dst  /* align 16 */,
+    void (*add_bytes_l2)(uint8_t *dst,
                          uint8_t *src1 /* align 16 */,
-                         uint8_t *src2 /* align 16 */, int w);
+                         uint8_t *src2, int w);
 
     /* this might write to dst[w] */
     void (*add_paeth_prediction)(uint8_t *dst, uint8_t *src,

libavcodec/x86/pngdsp.asm

@@ -42,12 +42,12 @@ cglobal add_bytes_l2, 4, 6, %1, dst, src1, src2, wa, w, i
     and                waq, ~(mmsize*2-1)
     jmp .end_v
 .loop_v:
-    mova                m0, [src1q+iq]
-    mova                m1, [src1q+iq+mmsize]
-    paddb               m0, [src2q+iq]
-    paddb               m1, [src2q+iq+mmsize]
-    mova  [dstq+iq       ], m0
-    mova  [dstq+iq+mmsize], m1
+    movu                m0, [src2q+iq]
+    movu                m1, [src2q+iq+mmsize]
+    paddb               m0, [src1q+iq]
+    paddb               m1, [src1q+iq+mmsize]
+    movu  [dstq+iq       ], m0
+    movu  [dstq+iq+mmsize], m1
     add                 iq, mmsize*2
 .end_v:
     cmp                 iq, waq