Hi Alex, On Tue, Sep 22, 2020 at 08:40:07AM +0100, Alex Coplan wrote: > On 21/09/2020 18:35, Segher Boessenkool wrote: > Thanks for doing this testing. The results look good, then: no code size > changes and no build regressions.
No *code* changes. I cannot test aarch64 likme this. > > So, there is no difference for most targets (I checked some targets and > > there really is no difference). The only exception is aarch64 (which > > the kernel calls "arm64"): the unpatched compiler ICEs! (At least three > > times, even). > > Indeed, this is the intended purpose of the patch, see the PR (96998). You want to fix a ICE in LRA caused by an instruction created by LRA, with a patch to combine?! That doesn't sound right. If what you want to do is a) fix the backend bug, and then b) get some extra performance, then do *that*, and keep the patches separate. > > Can you fix this first? There probably is something target-specific > > wrong related to zero_extract. > > The intent is to fix this in combine here. See the earlier replies in > this thread. But that is logically impossible. The infringing insn does not *exist* yet during combine. Segher
