Re-add time-passes to the collected data

[Mark-Simulacrum]

It would be good to get opinions from the compiler team. I expected effectively nonexistent demand, but so far both @arielb1 and @eddyb have requested it for the runs I've done on PRs, so maybe it'd be worth re-adding. I could see us adding it only into the comparison view, where you could switch to a specific crate being shown and then you'd see the passes for it. Timing data has fluctuated too much (and still does) in the past for it to be worth it though I think.

cc @michaelwoerister @alexcrichton @nikomatsakis: What are your thoughts?

[alexcrichton]

I was under the impression that "passes" as we know them today are basically going away in the compiler anyway so we wouldn't want to track this over time regardless. It may be worthwhile, though, investigating the ability to profile queries, as in the "near future" I'd imagine that everything is structured as a query?

[Mark-Simulacrum]

Hm, well, I'd certainly be willing to add queries to the data we collect. However, it looks to me that they are as prone (or even more so) to change names and meanings over time, and as such are unstable for long-term tracking. However, when comparing PR to PR (e.g. @bors try runs) perhaps that's not too large a problem. Perhaps we could implement some form of high-level view on queries: maybe something like start -> HIR construction, HIR built, typeck, mir building, borrowck, llvm passes. These are largely drawn out of the air -- I don't really know what's meaningful to keep track of and permanent enough to not change meaning overly much over time.

[nikomatsakis]

I am personally against adding tooling for time-passes, which I consider a dated concept. I believe that query-profiling data would probably be useful though. To me, though, I'd consider this lower priority versus getting benchmarks of the generated code up and going.

[nikomatsakis]

It's true that query names do change though. But how much does that matter? I imagine typically you'd just want to view the composition of a single run in terms of the queries in which it spent its time -- you can compare "by eye" across runs, you're basically just looking for gross discrepancies there. The human can adjust for the different names (which are unlikely anyway as long as the two runs are close together in time).

[Mark-Simulacrum]

Okay, so I take it that time-passes seems like a bad idea due to being dated. With query-profiling data, I'd be open to adding it, but will need to read up on what that means: I'm not entirely familiar with it as a concept. I think we can definitely do something with it, maybe even just having the HTML files that I believe it generates available for perusal for all benchmarks on perf.rlo.