[Bug rtl-optimization/85925] [7/8/9 regression] Mis-compilation at -02, masking with 257 goes wrong in combine

[ebotcazou at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85925

--- Comment #11 from Eric Botcazou <ebotcazou at gcc dot gnu.org> ---
> I get this:
> 
>        alpha  100.905%  
>          arm  100.072%  
>          c6x  100.000%  
>         csky  100.063%  
>        h8300  100.000%  
>         i386  100.000%  
>   microblaze  100.001%  
>         mips  100.052%  
>       mips64  100.043%  
>        nios2  100.000%  
>       parisc  100.049%  
>     parisc64  100.031%  
>    powerpc64   99.952%   
>      riscv64  100.049%  
>         s390  100.000%  
>           sh  100.006%  
>      shnommu  100.002%  
>        sparc  100.003%  
>      sparc64  100.007%  
>       x86_64  100.000%

Thanks.  Note that the patch only changes something for
WORD_REGISTER_OPERATIONS
targets though, so you can exclude h8300, i386, powerpc64, s390 and x86_64.

Can you easily re-do the test with the reversion of the nonzero_bits1 change,
i.e. the patch that you had/have in your tree?

> The alpha one looks really bad.  Only the powerpc64 is a win.

See above, this must be something else.