[Bug target/51821] [4.5/4.6/4.7 Regression] 64bit > 32bit conversion produces incorrect results with optimizations

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51821

--- Comment #3 from Uros Bizjak <ubizjak at gmail dot com> 2012-01-11 12:33:59 UTC ---
(In reply to comment #2)
> I believe this is a target issue.  We produce
> 
> test:
> .LFB0:
>         .cfi_startproc
>         movl    4(%esp), %ecx
>         movl    $-1, %eax
>         xorl    %edx, %edx
>         sall    %cl, %eax
>         testb   $32, %cl
>         cmovne  %eax, %edx   <-- wrong
>         cmovne  %edx, %eax
>         ret
> 
> but as this is a SHIFT_COUNT_TRUNCATED target the sall instruction does
> nothing, so shifting DImode by 32 is not implemented correctly on i?86
> (ix86_split_ashl).

There is one cmovne too much. -O2 -m32 -msse2 is needed to trigger this bug,
without cmove we produce correct code.