[Bug debug/96354] [10/11 Regression] ICE in maybe_canonicalize_mem_ref_addr, at gimple-fold.c:4903 since r10-2271-gd81ab49d0586fca0

["rguenther at suse dot de" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96354

--- Comment #16 from rguenther at suse dot de <rguenther at suse dot de> ---
On Wed, 29 Jul 2020, jakub at gcc dot gnu.org wrote:

> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96354
> 
> --- Comment #15 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
> We have indeed:
>   # DEBUG D#2 => MEM[(double *)&<retval>]
> and on the caller side:
>   D.2566[_9] = foo<3, 3> (D.2559, D.2572); [return slot optimization]
> So, that is why to &<retval> &D.2566[_9] is propagated.
> Now, if I add to foo a call to some function template that takes &t as
> argument, this is handled correctly because id->regimplify is set to true and
> the
> qux<C<3> > (&D.2572[_9]);
> call is fixed up by gimple_regimplify_operands.  That function really isn't
> called on debug stmts though (and not prepared to be called for them).
> If I use
> template <int N, int M> C<N> foo (D<N>, C<M>) { C<N> t; double d = 1.25;
> __builtin_memcpy (&t, &d, sizeof (double)); return t; }
> instead so that before that inlining we get
>   MEM <long unsigned int> [(char * {ref-all})&<retval>] = _4;
> then it is indeed again gimple_regimplify_operands that fixes up the
> MEM <long unsigned int> [(char * {ref-all})&D.2569[_9]] = 4608308318706860032;
> into:
>   _21 = &D.2569[_9];
>   MEM <long unsigned int> [(char * {ref-all})_21] = 4608308318706860032;

So we could avoid (some) regimplification if we'd dealt with this
gimplification step during return value setup.  We then only might
not re-propagate things and so

 <retval>.b;

might be forever

 tem_1 = &D.123[j_3];
 MEM[tem_1].b;

instead of

 D.123[j_3].b;

doing the re-gimplification when we replace things from the
decl map might be possible as well of course (and we then
would have "context" and could do special things when in
ADDR_EXPR context).  Not sure how ugly that is.

Not sure how bad the non-propagation above is either and how
often it triggers.