On 1/29/20 5:35 PM, Aleksandar Markovic wrote: > 00:51 Čet, 30.01.2020. Richard Henderson <richard.hender...@linaro.org > <mailto:richard.hender...@linaro.org>> је написао/ла: >> >> The first two address the performance regression noticed >> by Howard Spoelstra. The last two are just something I >> noticed at the same time. >> > > But, performance regression, according to Howard bisect analysis, happened > because of the change in target-independant code, and the fix presented here > is > in target-specific code. This defies basic logic and deserves clear and > detailed explanation. > > My additional concern, of course, is: Are other targets exposed to performance > degradation, and why?
Potentially, yes. However: It requires lots of loads in a loop, on a hot path. I would not have guessed that the ppc32 Load Multiple Word (et al) was on a hot path at all, since the instructions are deprecated. But that's what an ancient os gets you, I suppose. r~
