On 04/05/16 11:26, Eric Botcazou wrote: >> Given how many latent bugs it has shown up I think that alone would make >> it valuable to have enabled at -O2. > > It might be worthwhile to test it on embedded architectures because modern > x86 > and PowerPC processors are probably not very sensitive to this kind of tweaks. >
On ARM / AArch32 I haven't seen any performance data yet - the one place we are concerned about the impact is on Thumb2 code size as regrename may end up inadvertently putting more things in high registers. We'll check on that separately, having been away recently I haven't done any recent measurements with CSiBe - I'll try and get that done this week. Our bots have been a bit too flaky recently for me to say this with any certainty at the minute. On AArch64 (thanks to Wilco for some benchmarking), we've generally seen a small upside in our benchmarks (a couple of proprietary suites that I cannot name and SPEC2k(6)) aside from one major regression which is arguably an issue in the backend pattern for that particular intrinsic (aes) and would have been visible with -frename-regs anyway ! That needs to be fixed irrespective of turning this on by default. Wilco tells me that on SPEC2k6 we see a code size improvement by removing quite a lot of redundant moves though the performance difference appears to be in the noise on SPEC2k6. It does appear to look positive for aarch64 at first glance. regards Ramana
