30-60% perf regression in ctfe-stress

[jonas-schievink]

This happened on 2020-01-11: https://perf.rust-lang.org/compare.html?start=bfd04876b93ad5c013d90bc46937e28b6ee1a3f4&end=1389494ac145a84dba025ff65969f7ab150c3f02&stat=instructions:u

Caused by #67000

cc @oli-obk @spastorino

(I could've sworn we already tracked this regression, but apparently not)

[oli-obk]

We knew about this regression in the PR, maybe you remember that discussion. In #67000 (comment) I finally came to the conclusion that the reason we regressed perf here is that before that PR we didn't actually evaluate the promoteds, we just created pointers to them. #67000 causes all promoteds to get evaluated immediately when their use site is evaluated. Basically any real program would already have done this work as collect would have evaluated the constant. Only if you have loads and loads of dead promoteds can you see the difference.

[jonas-schievink]

Okay, that sounds much less alarming then. Thanks for the clarification!

[nikomatsakis]

Visited during compiler-team triage. Based on @oli-obk's comment, going to close as "won't fix". Thanks @jonas-schievink!