> Date: Thu, 11 May 2023 12:15:20 -0600 > From: Jeff Law <jeffreya...@gmail.com>
> On 5/11/23 10:55, Paul Koning wrote: > > > > > >> On May 11, 2023, at 11:05 AM, Hans-Peter Nilsson via Gcc-patches > >> <gcc-patches@gcc.gnu.org> wrote: > >> > >> ... > >> Yes, very interesting. Thank you for sharing this. I've > >> seen regressions with LRA for CRIS too, for > >> "double-register-sized" types, which for CRIS, a 32-bit > >> target, translates to 64-bit types (DFmode and DImode), and > >> where LRA does a much worse job than reload; spills a lot > >> more often to stack, even after trying every > >> register-allocation-related hook I found (and also an LRA > >> patch which helped only by a fraction, but regressed results > >> on x86_64-linux, so let's quickly forget it again). > > > > That observation makes me a bit worried. While CRIS may not be a priority > > platform, that description makes it sound like a case that would be > > significant in any 32 bit platform, which would include priority ones like > > i386 and ARM. > If I understood things correctly, it seems to impact more when the > target exposes double-word patterns but doesn't actually have > instructions for those operations. That's an implementation pattern > we've largely been moving away from over the last decade or so. That description doesn't really match CRIS though. The "ax" prefix used in DImode patterns links the next instruction to include the carry. Thus better than an "open-coded" version. In comparison, CRIS doesn't define separable patterns (anddi3, iordi3 etc.) But, there's a movdi expander and splitter - with a long reload-related comment. Perhaps I can do away with that even though having some arithmetic and compares in DImode. Thanks for the hint. brgds, H-P
