Canonical structures have major performance impact
As I mentioned on Mattermost, I had a fun experiment collapsing the sbi
and bi
canonical structures.
- 37.40% overall on lambdarust-weak, with improvements for individual files up to 72.22%.
- 23.28% overall on Iron, with improvements for individual files up to 38.13%.
- 3.8% overall on Iris, with improvements for individual files up to 13.16%.
- 5.2% overall on lamdarust master, with improvements for individual files up to 10.18%.
These differences are major, especially for the projects that use BI formers (monPred in lambdarust-weak and fracPred in Iron)! So it looks like there is really something we should investigate.