Wes McKinney comments

Results 203 comments of


                                            Wes McKinney

[C++][DONOTMERGE] Use -O2 instead of -O3 for RELEASE builds

@ursabot please benchmark

[C++][DONOTMERGE] Use -O2 instead of -O3 for RELEASE builds

cc @pitrou

[C++][DONOTMERGE] Use -O2 instead of -O3 for RELEASE builds

Here's a dump of symbols that shrink the most in -O2: https://gist.github.com/wesm/4a2815077ed37b671d6160b8abec5e7c I'd be interested to see if e.g. unsafe numeric casts are significantly affected by this

[C++][DONOTMERGE] Use -O2 instead of -O3 for RELEASE builds

I would guess there are places that benefit from loop unswitching also.

[C++][DONOTMERGE] Use -O2 instead of -O3 for RELEASE builds

I would say that we should just keep O3 and keep an eye on symbol sizes in case we need to intervene occasionally. On the whole I think the symbol...

[C++][DONOTMERGE] Use -O2 instead of -O3 for RELEASE builds

@ursabot please benchmark lang=C++

[C++][DONOTMERGE] Use -O2 instead of -O3 for RELEASE builds

I added `-ftree-vectorize` for gcc

[C++][DONOTMERGE] Use -O2 instead of -O3 for RELEASE builds

> You undid the changes I had already pushed :-( @pitrou, ugh sorry about that, I think I hadn't reloaded the browser. I will restore

[C++][DONOTMERGE] Use -O2 instead of -O3 for RELEASE builds

@save-buffer thanks for your comments > Change the code to use fewer templates and more raw for loops. If we're feeling really adventurous, we could write a Python or Jinja...

[C++][DONOTMERGE] Use -O2 instead of -O3 for RELEASE builds

In https://conbench.ursa.dev/compare/runs/e938638743e84794ad829524fae04fbd...20727b1b390e4b30be10f49db7f06f3f/ it seems that there are several hundred microbenchmarks with > 10% performance regressions but also over 100 microbenchmarks with > 10% performance improvement. I'd say it's a coin...