[Bug rtl-optimization/82628] [8 Regression] wrong code at -Os on x86_64-linux-gnu in the 32-bit mode

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=82628

--- Comment #4 from Uroš Bizjak <ubizjak at gmail dot com> ---
(In reply to Uroš Bizjak from comment #3)
> Looks like combine failure to me:
>
> I don't think that combining:
> 
> (set (reg:CC 17 flags)
>     (compare:CC (subreg:SI (reg/v:DI 89 [ d ]) 0)
>         (const_int 1 [0x1])))
> 
> with
> 
> (insn 20 19 21 4 (parallel [
>             (set (reg:CCGZ 17 flags)
>                 (compare:CCGZ (reg:SI 94)
>                     (plus:SI (ltu:SI (reg:CC 17 flags)
>                             (const_int 0 [0]))
>                         (subreg:SI (reg/v:DI 89 [ d ]) 4))))
>             (clobber (scratch:SI))
> 
> should result in unconditional jump:
> 
> (set (pc)
>     (label_ref 16))

Please note two different subregs of the same register. Maybe somewhere combine
figures out that two different subregs of the same register are the same (which
are not - see the offsets!) and removes insn due to the invalid
"simplification".

Segher?