[llvm-dev] Question about canonicalizing cmp+select

[Roman Lebedev via llvm-dev]

On Tue, Jul 3, 2018 at 9:06 PM, Yuan Lin via llvm-dev
<llvm-dev at lists.llvm.org> wrote:
> Hi, Sanjay/all,
>
>   I noticed in rL331486 that some compare-select optimizations are disabled
> in favor of providing canonicalized cmp+select to the backend.
>
>   I am currently working on a private backend target, and the target has a
> small code size limit.  With this change, some of the apps went over the
> codesize limit.  As an example,
>
> C code:
>   b = (a > -1) ? 4 : 5;
>
> ll code:
> Before rL331486:
>   %0 = lshr i16 %a.0.a.0., 15
>   %1 = or i16 %0, 4
>
> After rL331486:
>   %cmp = icmp sgt i16 %a.0.a.0., -1
>   %cond = select i1 %cmp, i16 4, i16 5
>
>   With the various encoding restrictions of my particular target, the
> cmp/select generated slightly larger code size.  However, because the apps
> were very close to the code size limit, this slight change pushed them over
> the limit.
>
>   If I still prefer to have this optimization performed, then is my best
> strategy moving forward to implement this optimization as a peephole opt in
> my backend?
I personally think it should probably be a DAGCombine transform,
driven by a Target Transform Info hook (e.g. like hasAndNot()).

> Thanks,
> --Yuan
Roman.

> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>