On Tue, Mar 05, 2024 at 09:29:38AM +0100, Richard Biener wrote:
> I wonder if we need to care about extra temporaries on RTL before
> RA, thus whether always using a temporary would be OK?
I'd still need to do the resolve_reg_p check, otherwise if it is e.g. a
memory, the copying to temporary would be significantly harder because
src isn't a CONCATN in that case.
I think the reg_overlap_mentioned_p is worth it, because the normal
case is that there is no overlap, usually expansion forces distinct
pseudos. If this was more common problem, it wouldn't take so long
to discover it and wouldn't need -fno-tree-coalesce-vars.
And the extra pseudo could result in different RA decisions.
Jakub
