Am Montag, dem 11.05.2026 um 19:30 +0300 schrieb Alexander Monakov: > Hello, > > On Mon, 11 May 2026, Martin Uecker wrote: > > So, annoyingly, I got this wrong. > > > > Bootstrapped and regression tested on x86_64. > > In the testcase, wrong code is introduced by a fairly late DSE pass after > ivopts rewrites the memory access in the loop. I'm curious, do you know > if the types in the IR are inconsistent from the start (but the inconsistency > is benign), or does iv-opts introduce an incorrectly-typed access somehow?
I haven't understood where exactly this goes wrong, but we get aliasing for pointers to compatible but non-identical function types wrong in general (PR114959), so it was a mistake to rely on this in the patch. Martin
