On Mon, Jun 29, 2020 at 08:58:57AM +0200, Richard Biener via Gcc-patches wrote:
> Most of the cases I've seen involve transforms that make _b_c_p constant
> on one path and then introduce a new PHI merging two _b_c_p values
> to be then tested in a not simplified condition. I'm not sure how to fend
> off jump threading (yeah, it's nearly always jump threading doing this...)
> doing this but certainly the easiest way would be to simply disallow
> [jump threading] from duplicating _b_c_p calls.
>
> Or fold _b_c_p even earlier (though I definitely saw early backwards
> jump threading mess up such a case).
Yeah, perhaps disallow duplicating bcp calls in the threading before IPA
(for bcp to work well, we want to preserve it until inlining had a chance to
propagate constants in there) and after IPA allow that, but just fold them
into 0 during that on both paths.
Jakub
