Antoine Pitrou
Antoine Pitrou
@lwhite1 Do you feel acquainted enough with the C data interface to give this a look?
Perhaps @lidavidm can give another look? :-)
Hmm, perhaps the bit-util micro-benchmarks are a bit pathologic, but other regressions seem real and quite significant...
There's `-ftree-vectorize` which apparently enables all vectorization (both "tree" and "SLP").
Also, apparently with gcc 12 the following flags are enabled with `-O2`: `-ftree-loop-vectorize -ftree-slp-vectorize -fvect-cost-model=very-cheap`
@ursabot please benchmark lang=C++
Ok, so I run locally with gcc 9.4.0 (Ubuntu 20.04) on AMD Zen 2. ### Build times (with ccache disabled) * `-O3`: ``` real 1m25,888s user 21m58,527s sys 0m48,888s ```...
Hmm, I notice that `-DNDEBUG` was not passed anymore, so perhaps that affected some benchmarks :-( Edit: re-ran benchmarks and updated the gists above.
@ursabot please benchmark lang=C++
cc @save-buffer @westonpace for further opinions