[PATCH] sra->srl reduction and PowerPC test change

[Bill Schmidt]

Hi Richard,

I can't sign off on the codegen part (though it looks fine to me), but
the patch for the PowerPC test LGTM.  Thanks for the improvement!

Bill

On Wed, 2013-10-16 at 15:13 +0100, Richard Sandiford wrote:
> SimplifyDemandedBits has code to turn sra into srl if the shifted-in
> sign bits don't matter.  This makes it easier to combine the shift with
> others (such as a shift left).
> 
> The attached patch extends this to cases where the only bit that matters
> is a sign bit.  E.g. (and (sra (i32 x) 31) 2) -> (and (srl (i32 x) 30) 2).
> 
> I think the codegen part is obvious enough to just apply (but please let
> me know if I'm wrong).  However, the optimisation changes the output of
> test/CodeGen/PowerPC/rlwimi-and.ll.  The old code was:
> 
>         li 3, 0
>         lbz 12, 0(3)
>         cmplwi 0, 3, 0
>         slwi 4, 12, 31
>         rlwinm 3, 12, 0, 31, 31
>         srawi 4, 4, 31
>         rlwinm 4, 4, 0, 23, 23
>         rlwimi 4, 3, 0, 24, 22
>         sth 4, 0(3)
> 
> The new code is:
> 
>         li 3, 0
>         lbz 12, 0(3)
>         cmplwi 0, 3, 0
>         slwi 4, 12, 31
>         rlwinm 3, 12, 0, 31, 31
>         rlwimi 3, 4, 9, 23, 23
>         sth 3, 0(3)
> 
> Is the test change OK?  As mentioned in the FIXME, I think we could also
> fold the SLWI into the RLWIMI, but that'd probably be a back-end change.
> 
> Thanks,
> Richard
>