Re: [PATCH] Fix PR70497, missed "subreg" CSE on GIMPLE

["H.J. Lu" <hjl dot tools at gmail dot com>]

On Mon, May 9, 2016 at 5:59 AM, Richard Biener <rguenther@suse.de> wrote:
>
> The following patch implements CSEing of "subreg" reads from memory
> like (from the testcase in the PR)
>
> union U { int i[16]; char c; };
>
> char foo(int i)
> {
>   union U u;
>   u.i[0] = i;
>   return u.c;
> }
>
> CSEing u.c as (char)i and thus removing u during GIMPLE optimizations.
>
> The patch always goes via generating BIT_FIELD_REFs and letting them
> be simplified via the match-and-simplify machinery.  This means it
> replaces handling of complex component and vector extracts we've been
> able to do before.
>
> I didn't restrict the kind of BIT_FIELD_REFs much apart from requiring
> byte-size accesses.  I did inspect code generated on powerpc (big-endian)
> for the testcase though (also to verify any endianess issues) and didn't
> spot anything wrong (even for non-lowpart "subregs").
>
> Bootstrapped on x86_64-unknown-linux-gnu, testing in progress.
>
> Richard.
>
> 2016-05-09  Richard Biener  <rguenther@suse.de>
>
>         PR tree-optimization/70497
>         * tree-ssa-sccvn.c (vn_nary_build_or_lookup): New function
>         split out from ...
>         (visit_reference_op_load): ... here.
>         (vn_reference_lookup_3): Use it to handle subreg-like accesses
>         with simplified BIT_FIELD_REFs.
>         * tree-ssa-pre.c (eliminate_insert): Handle inserting BIT_FIELD_REFs.
>         * tree-complex.c (extract_component): Handle BIT_FIELD_REFs
>         correctly.
>
>         * gcc.dg/torture/20160404-1.c: New testcase.
>         * gcc.dg/tree-ssa/ssa-fre-54.c: Likewise.
>

This caused:

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71055

H.J.