Jim Wilson wrote: > It looks like a slight lose on qdf24xx on SPEC CPU2006 at -O3. I see > about a 0.37% loss on the integer benchmarks, and no significant > change on the FP benchmarks. The integer loss is mainly due to > 458.sjeng which drops 2%. We had tried various values for > max_case_values earlier, and didn't see any performance improvement > from setting it, so we are using the default value.
That's interesting as sjeng shows ~2% gain on Cortex-A72 due to the hot switches being badly laid out... I wonder whether the loss you see is due to code alignment or some other secondary effect. > We've been tracking changes to the FSF tree, and adjust our tuning > structure as necessary, so I'm not too concerned about this. We will > just set the max_case_values field in the tuning structure to get the > result we want. What I am slightly concerned about is that the > max_case_values field is only used at -O3 and above which limits the > usefulness. If a port has specified a value, it probably should be > used for all non-size optimization, which means we should check for > optimize_size first, then check for a cpu specific value, then use the > default. If you do that, then you don't need to change the default to > get better generic/a53 code, you can change it in the generic and/or > a53 tuning tables. Yes it would be better to ensure max_case_values is used at -O2 as well. But even then you'd want a reasonable default so that you only need to set it if you want something very different. > Though I see that the original patch from Samsung that added the > max_case_values field has the -O3 check, so there was apparently some > reason why they wanted it to work that way. The value that the > exynos-m1 is using, 48, looks pretty large, so maybe they thought that > the code size expansion from that is only OK at -O3 and above. Worst > case, we might need two max_case_value fields, one to use at -O1/-O2, > and one to use at -O3. I hope we can improve switch expansion in GCC7 to avoid the worst of the issues, and then we can revisit these settings. Wilco
