[FFmpeg-devel] [PATCH/RFC] mathops/x86: work around inline asm miscompilation with GCC 4.8.1

[Carl Eugen Hoyos]

Michael Niedermayer <michaelni <at> gmx.at> writes:

> > I pin-pointed the specific uses of the function in 
> > question that cause it, but reading the generated 
> > asm didn't show anything obvious yet, but it could 
> > be the smallest thing.
> > Its the 4 mid_pred calls in vc1_pred_b_mv, if 
> > anyone else wants to have a look.

(Only the two first ones.)

> can you post the generated asm? i mean the good one 
> without volatile and the bad with ?

gcc 4.8.0, 32bit
The relevant parts of the source where the change 
makes a difference for fate are lines 2156 and 2235 
of libavcodec/vc1dec.c:
px = mid_pred(A[0], B[0], C[0]);

Following is the diff between the original 
("notworking.txt") and the patched assembly 
("working.txt"). I only changed the two 
relevant function calls, not any other part 
of the code:
--- notworking.txt      2013-06-21 09:31:13.065457215 +0200
+++ working.txt 2013-06-21 09:29:33.042439057 +0200
@@ -149,26 +149,26 @@
    0x08557c88 <+552>:   cmpl   $0x1,0x28(%esp)
    0x08557c8d <+557>:   lea    (%esi,%ebx,4),%esi
    0x08557c90 <+560>:   mov    %esi,0x2c(%esp)
    0x08557c94 <+564>:   je     0x85581f8 <vc1_pred_b_mv+1944>
    0x08557c9a <+570>:   add    %ebx,%ecx
    0x08557c9c <+572>:   mov    0x1c(%esp),%ebx
    0x08557ca0 <+576>:   lea    (%ebx,%ecx,4),%ebx
    0x08557ca3 <+579>:   movswl (%ebx),%esi
    0x08557ca6 <+582>:   mov    %ebx,0x30(%esp)
    0x08557caa <+586>:   mov    0x24(%esp),%ebx
-   0x08557cae <+590>:   mov    %esi,%edx
-   0x08557cb0 <+592>:   mov    %esi,%ecx
-   0x08557cb2 <+594>:   mov    0x10(%esp),%esi
-   0x08557cb6 <+598>:   movswl (%ebx),%ebx
-   0x08557cb9 <+601>:   mov    %ebx,0x10(%esp)
-   0x08557cbd <+605>:   mov    0x2c(%esp),%ebx
+   0x08557cae <+590>:   mov    %esi,%ecx
+   0x08557cb0 <+592>:   mov    %esi,%edx
+   0x08557cb2 <+594>:   movswl (%ebx),%ebx
+   0x08557cb5 <+597>:   mov    %ebx,0x10(%esp)
+   0x08557cb9 <+601>:   mov    0x2c(%esp),%ebx
+   0x08557cbd <+605>:   mov    0x10(%esp),%esi
    0x08557cc1 <+609>:   movswl (%ebx),%ebx
    0x08557cc4 <+612>:   cmp    %ecx,%ebx
    0x08557cc6 <+614>:   cmovg  %ebx,%edx
    0x08557cc9 <+617>:   cmovg  %ecx,%ebx
    0x08557ccc <+620>:   cmp    %esi,%ebx
    0x08557cce <+622>:   cmovl  %esi,%ebx
    0x08557cd1 <+625>:   cmp    %ebx,%edx
    0x08557cd3 <+627>:   cmovg  %ebx,%edx
    0x08557cd6 <+630>:   mov    0x24(%esp),%esi
    0x08557cda <+634>:   mov    0x30(%esp),%ebx
@@ -302,26 +302,26 @@
    0x08557e88 <+1064>:  cmpl   $0x1,0x1c(%esp)
    0x08557e8d <+1069>:  lea    (%esi,%ebx,4),%esi
    0x08557e90 <+1072>:  mov    %esi,0x24(%esp)
    0x08557e94 <+1076>:  je     0x8558222 <vc1_pred_b_mv+1986>
    0x08557e9a <+1082>:  add    %ebx,%ecx
    0x08557e9c <+1084>:  mov    0x20(%esp),%ebx
    0x08557ea0 <+1088>:  lea    (%ebx,%ecx,4),%ebx
    0x08557ea3 <+1091>:  movswl (%ebx),%esi
    0x08557ea6 <+1094>:  mov    %ebx,0x28(%esp)
    0x08557eaa <+1098>:  mov    0x18(%esp),%ebx
-   0x08557eae <+1102>:  mov    %esi,%edx
-   0x08557eb0 <+1104>:  mov    %esi,%ecx
-   0x08557eb2 <+1106>:  mov    0x14(%esp),%esi
-   0x08557eb6 <+1110>:  movswl (%ebx),%ebx
-   0x08557eb9 <+1113>:  mov    %ebx,0x14(%esp)
-   0x08557ebd <+1117>:  mov    0x24(%esp),%ebx
+   0x08557eae <+1102>:  mov    %esi,%ecx
+   0x08557eb0 <+1104>:  mov    %esi,%edx
+   0x08557eb2 <+1106>:  movswl (%ebx),%ebx
+   0x08557eb5 <+1109>:  mov    %ebx,0x14(%esp)
+   0x08557eb9 <+1113>:  mov    0x24(%esp),%ebx
+   0x08557ebd <+1117>:  mov    0x14(%esp),%esi
    0x08557ec1 <+1121>:  movswl (%ebx),%ebx
    0x08557ec4 <+1124>:  cmp    %ecx,%ebx
    0x08557ec6 <+1126>:  cmovg  %ebx,%edx
    0x08557ec9 <+1129>:  cmovg  %ecx,%ebx
    0x08557ecc <+1132>:  cmp    %esi,%ebx
    0x08557ece <+1134>:  cmovl  %esi,%ebx
    0x08557ed1 <+1137>:  cmp    %ebx,%edx
    0x08557ed3 <+1139>:  cmovg  %ebx,%edx
    0x08557ed6 <+1142>:  mov    0x18(%esp),%esi
    0x08557eda <+1146>:  mov    0x28(%esp),%ebx


Hope that helps, Carl Eugen