[Bug libgcj/51500] [4.7 regression] 106 unexpected libjava testsuite failures with mingw32

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51500

--- Comment #14 from Uros Bizjak <ubizjak at gmail dot com> 2012-01-30 19:58:43 
UTC ---
(In reply to comment #13)
> Right the following sequence seems to be better IMHO:
> 
> pop %eax
> push %ecx
> push %eax
> mov __ctx, %eax
> call fun
> pop %ecx
> mov %ecx, %(esp)
> ret

I'd avoid push and pop insns and go directly with moves and %esp adjustments.
Please see i.e. libitm/config/x86/sjlj.S for some ideas.