python3: enable optimizations.

[triallax]

This --enable-optimizations flag enables profile-guided optimizations (PGO) and link-time optimization (LTO, but apparently only sometimes, depending on the platform). My rudimentary benchmarking indicates that compared to python3 from the repos, python3 built from this PR yielded performance improvements of around 10%-12% (I tried building python3 from master locally, but that one had horrendous performance for some reason). I'm not a benchmarking expert though, so I wanted others to try out their own benchmarks and post their results here.

On my potato laptop, this change tripled build time from 5 minutes to 15 minutes. I don't think this is an issue however because changes to the python3 template are not done often (averaging once per month in the last 12 months), and the builders are much beefier than my laptop, so their build times will be even shorter anyway.

TODO:

• Is this also needed for python3-tkinter? (I don't think so, but better check)
• Are there similar performance improvements with cross builds? (PGO runs and profiles a Python that is compiled for the host architecture, which is why I want to confirm this)

Testing the changes

• I tested the changes in this PR: YES (only python3, not python3-tkinter)

@ahesford

[ahesford]

Sorry for the delay. I don't really have a problem with enabling optimizations; if they improve runtime performance, we can pay the price in compile time. The question is how much PGO makes improves things for the average user if the optimization is tailored to the build system. I'd like to test a few things, including building on different CPUs to see how well the optimizations translate.

[tornaria]

Sorry for the delay. I don't really have a problem with enabling optimizations; if they improve runtime performance, we can pay the price in compile time. The question is how much PGO makes improves things for the average user if the optimization is tailored to the build system. I'd like to test a few things, including building on different CPUs to see how well the optimizations translate.

As a data point, I built this on my box (i7-9700). Running the sagemath testsuite with this (in the same box) takes ~5% less time to complete. As in 12:00 vs 12:30 for the normal test, and 23:17 vs 24:40 for the long test using 8 threads.

[ahesford]

I built this PR on a Zen 3 CPU and ran a test that built a sparse finite-element matrix and then used SuperLU to performan LU factorization. Although the SuperLU performance won't change, the setup does some heavy looping in Python and the overall time for the routine dropped from 2.78 sec to 2.3 sec, each trial averaged over several runs. I installed the Zen-built package on another system with a 9th-generation Intel CPU and re-ran the same test, with the times dropping from 3.7 sec to 2.75 sec.

It seems these optimizations offer tangible benefits, and they translate to different systems.

[tornaria]

Just for the record, the sagemath (10.1.rc0) testsuite with python3-3.11.4_1:

Total time for all tests: 843.5 seconds
    cpu time: 5039.4 seconds
    cumulative wall time: 5700.8 seconds

After updating to python3-3.11.4_2 (no other change, no recompilation):

Total time for all tests: 790.8 seconds
    cpu time: 4601.8 seconds
    cumulative wall time: 5175.3 seconds

That's a neat 5-10 % speedup! Thanks!