15% regression in compiler memory usage for inflate crate

[alexcrichton]

According to perf.rust-lang.org the inflate-0.1.0 benchmark regressed 15% in memory usage over the past few weeks.

There was an initial big regression due to #39409 and #43576 recovered most of it but unfortunately there's still a 15% overall regression from before :(

[alexcrichton]

A similar, but not as large regression can be seen for the deep vector benchmark

[arielb1]

Looks expected, but someone should run a profiler to confirm it's what it looks like.

[nikomatsakis]

Believed to be fallout from EndRegion. In that case, we can probably just close as "expected", if unfortunate. Those are however likely to go away with the migration to NLL, so hopefully we will recover eventually. It'd be nice to do a quick profile to double check.

[nikomatsakis]

triage: P-medium

[alexcrichton]

1.21 is now in beta

@nikomatsakis should this not be on the milestone for 1.21?

[nikomatsakis]

@alexcrichton we were hoping to verify the results and make sure we at least know the cause. @pnkfelix has done some of the legwork here but I think we're still not clear that EndRegion is to blame, yet.

[pnkfelix]

Just some quick data to supply here: I switched to extern crate alloc_system; inrustc in order to get massif working.

When using valgrind --tool=massif, I was getting peak memory numbers for inflate like this:

Skip EndRegion	Emit EndRegion	ratio
198.1 MB	207.2 MB	1.046

This led me to think that there might be some other issue causing the regression being reported.

But after @nikomatsakis pushed back a little, I switched to using /bin/time, which on Linux includes peak memory usage in its report. The numbers then bounced around some, but on average they looked like this:

Skip EndRegion	Emit EndRegion	ratio
259.4 MB	329.4 MB	1.27

That was a lot scarier, and thus led me to file PR #44249.

(I don't know why the massif report is so different from the one I get from /bin/time. I have found myself repeating the phrase "'tis an ill workman that quarrels with his own tools" a lot over the past few days...)

[alexcrichton]

Hm that is indeed very odd! I would have thought that massif wouldn't tamper too much with the output... I wonder if thats something like really bad heap fragmentation where massif sums up literal allocation sizes whereas /bin/time sums up the actual pages being used?

[nikomatsakis]

Looking at perf, I see that the regression has been fixed by #44249. We'll have to wrangle this when it comes to switch to MIR borrowck, but we'll worry about then.