Continue disabling harness features, achieving best-case times that match Peter's, on the three tests with new CSVs here.
The fully-disabled harness probably still fools the optimizer because it uses the Buhr-Dice "pass function" trick in place of the disabled observation and iterators array. That these times are now often beating Peter's requires further investigation.
Make STL run with (cpp-sltref) and without (cpp-stlrefAlloGlib) llheap.
Include peter-*.csv from runs on Swift. (Thesis still builds with original numbers from labpc.)
Add data post-processing for sorting and aggregation, leveraging thesis's plots' framework.
Add makefile-saved setups for matching Peter.
Add ability to disable "observation" (use of volatiles) and iterators array (support for remelem).
Add integrated clean-build-run targets for varying the "disable" switches.
Strengthen disable-interleaving mode to avoid the extra loops on remove.
Still to entertain some of this stripping being unnecessary for achieving near Peter-match.
Still to compare costs of the relevant stripped features, and recommend harness variation(s) to for use in thesis.
