On Mon, 5 Oct 2026, Alex Coplan wrote:
> Hi,
>
> r17-71-g3f6d37193c7987ee9362c073dd5da6f7419c87b3 adds the following
> assert in vect_create_new_slp_node:
>
> gcc_assert (code == ERROR_MARK || code == VEC_PERM_EXPR);
>
> which seems to suggest that SLP nodes should only ever have one of these
> two codes.
Yes, this is in preparation (TODO) for dropping 'code' in favor of
some (class?) sub-setting of slp_tree, separating permute from
regular nodes.
> Recently I have been studying how the recognition of complex add with
> rotate works. There is code in vect_build_slp_tree_2 which, in the
> two_operators case, builds synthetic {PLUS,MINUS}_EXPR nodes which are
> then packed together with a permutation. This is later recognized by
> the complex add slp pattern.
>
> I noticed that this code sets the code fields of the synthetic nodes to
> the relevant tree codes, e.g. {PLUS,MINUS}_EXPR, but this isn't actually
> needed, since the pattern code just looks at the code of the
> representative stmt for non-permutation nodes. It's also at odds with
> the recently-added assert above.
>
> This patch therefore drops setting the codes on those nodes, and instead
> leaves them set to ERROR_MARK (meaning that consumers should look at the
> representative, as the patterns currently do). It is intended to be a
> non-functional cleanup.
>
> Bootstrapped/regtested on aarch64-linux-gnu, arm-linux-gnueabihf, and
> x86_64-linux-gnu. OK for trunk?
OK.
Richard.
> Thanks,
> Alex
>
> gcc/ChangeLog:
>
> * tree-vect-slp.cc (vect_build_slp_tree_2): In the two_operators
> case, avoid setting the code field on the plus/minus nodes.
>
--
Richard Biener <[email protected]>
SUSE Software Solutions Germany GmbH,
Frankenstrasse 146, 90461 Nuernberg, Germany;
GF: Stefan Gaiser, Jochen Jaser, Abhinav Puri; (HRB 36809, AG Nuernberg)
