Trans refactor was major perf perturbance

[graydon]

I can't say exactly whether it's good or bad, but the perf numbers went sideways a fair bit with the refactoring of trans in 5e36a99. Some numbers dropped, some went up a lot (factor of 2 on fibo and sudoku), and we lost about 5s on self-compile time.

All of this is potentially tolerable, I'm not suggesting we back it out (so nice!) but it's probably worth looking at a couple of the smaller benchmarks at least, maybe comparing a count-llvm-insns graph and/or the LLVM bitcode output, to see if something changed obviously (by mistake) in the code we're generating.

[nikomatsakis]

not unexpected, will take a look

[nikomatsakis]

The main cause seems to be larger stack size due to, I believe, deferred loads. This causes morestack to be invoked more often, leading to slower performance. I am looking into whether it's worth fixing this.

[bblum]

i am surprised that morestack would alone be responsible for up to 2x hits on some test cases. maybe believable, though, if these rare ones are stack-heavy(?) tests.

[catamorphism]

Commit ee4ba44 appears to cause a segfault in run-pass/cci_capture_clause.

[nikomatsakis]

@bblum fibonacci is basically all about calls. Basically all it does is morestack.

[nikomatsakis]

@catamorphism where are you observing this segfault? I don't see it locally and the bots seem fine.

[nikomatsakis]

fibo perf looks fine now. rustc performance is still lower, though that can be a result of many factors. When I test locally, however, rustc performs precisely the same both before and after the patch.

[catamorphism]

@nikomatsakis It segfaults on my machine. The executable x86_64-apple-darwin/test/run-pass/cci_capture_clause.stage2-x86_64-apple-darwin segfaults when run.

[catamorphism]

It also causes one of the libstd tests to segfault on my other machine. I haven't figured out which one yet. Reverting that patch fixes it.

[nikomatsakis]

Segfault fixed in 14303ba

[nikomatsakis]

I still haven't tracked down the msgsend-pipes.rs slowdown. I don't see it locally, so it's probably yet another stack usage thing.

[nikomatsakis]

Deferring to 0.5---I think this is basically "done" anyhow.