Fix png decoding on x86.

Line sizes are only 8-byte aligned, so use unaliged loads
for add_bytes_l2 pointers.
Increasing the alignment requirement to 16 seemed a bit extreme
(png may be used for rather small sizes).
Also fix a mov that had its arguments swapped, leading
add_bytes_l2 being applied on up to 8 bytes too few.

Signed-off-by: Reimar Döffinger <Reimar.Doeffinger@gmx.de>

libavcodec/pngdsp.h

@@ -26,8 +26,8 @@
 
 typedef struct PNGDSPContext {
     void (*add_bytes_l2)(uint8_t *dst  /* align 16 */,
-                         uint8_t *src1 /* align 16 */,
-                         uint8_t *src2 /* align 16 */, int w);
+                         uint8_t *src1,
+                         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

@@ -43,12 +43,12 @@ 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
@@ -56,12 +56,12 @@ cglobal add_bytes_l2, 4, 6, %1, dst, src1, src2, wa, w, i
 
 %if mmsize == 16
     ; vector loop
-    mov                 wq, waq
+    mov                waq, wq
     and                waq, ~7
     jmp .end_l
 .loop_l:
-    movq               mm0, [src1q+iq]
-    paddb              mm0, [src2q+iq]
+    movq               mm0, [src2q+iq]
+    paddb              mm0, [src1q+iq]
     movq  [dstq+iq       ], mm0
     add                 iq, 8
 .end_l: