Re: [PATCH] combine, v2: Don't record for UNDO_MODE pointers into regno_reg_rtx array [PR104985]

[Segher Boessenkool <segher@kernel.crashing.org>]

Hi!

So, the core of this problem is once again that regno_reg_rtx is
reallocated.  It will be another decade until we got rid of all fallout
of breaking that guarantee :-(

On Wed, Apr 06, 2022 at 10:50:40AM +0200, Jakub Jelinek wrote:
> On Tue, Apr 05, 2022 at 04:56:55PM -0500, Segher Boessenkool wrote:
> > > -  union { rtx *r; int *i; struct insn_link **l; } where;
> > > +  union { rtx *r; int *i; rtx m; struct insn_link **l; } where;
> > 
> > NAK.  It is not clear at all what "rtx m" means, esp. since there is an
> > "rtx *r" already.  In the PR you said "machine_mode m", that is clear of
> > course, can you do that instead?
> 
> So in that case something like this (i.e. the regno variant, renamed
> to subst_mode from SUBST_MODE, and naming the union member regno rather
> than m)?

I can't say I like that either, the undo for a mode change should just
store the old mode directly, anything else is too fragile.

But, whatever, we'll fix that later.  The patch is okay for trunk.
Thanks!


Segher