o-
commented
1 month ago Here are some stats for your PR:
- WARNING: the longest CI job test_release_2 took 2.07h
- are-we-fast-r suite improved by 1.03
- real_thing suite regressed by 0.94
- shootout suite regressed by 0.94
- simple suite regressed by 0.78
- summary suite regressed by 0.9
- convolution_v regressed by 0.69
- binarytrees_naive regressed by 0.19
- nbody_naive regressed by 0.74
- mandelbrot_noout_naive regressed by 0.79
- knucleotide regressed by 0.93
- spectralnorm_alt_2 regressed by 0.94
- pidigits regressed by 0.95
- superWhile regressed by 0.03
- scalarFor regressed by 0.91
- lapplyDots regressed by 0.92
- Overall benchmarks regressed by 0.9
Please find your performance results at https://rir-benchmarks.prl.fit.cvut.cz/diff?job_ids[]=7025932406&job_ids[]=6997286723&selection=all
Remove the type feedback versioning. The rationale is that it likely does not add much value while being costly. We used the information about version update before to give extra chance to recompile, but that might be either unproductive or unprecise. Unproductive in the sense that maybe the new information is not useful for the given context. Unprecise in the sense that it does not track the possible type feedback version change in inlinees.
Removes the time base heuristics for recompilation.
Fixes some issues related to inlining