wf: handle "livelock" checking before reaching WfPredicates::compute.

[eddyb]

For wf::obligations's "livelock" handling, this PR shouldn't cause any behavioral changes, as the check moved to it should be equivalent to the old one in WfPredicates::compute.

However, it fixes #70168 by making other users of WfPredicates::compute (that is, wf::predicate_obligations and compute's own upvar handling) correct for ty::Infer, in that they now get a WellFormed(ty::Infer(_)) obligation instead of silently ignoring the type.

r? @nikomatsakis

[eddyb]

@bors try @rust-timer queue (there may be new noop shallow_resolve calls introduced)

[rust-timer]

Awaiting bors try build completion

[bors]

⌛ Trying commit 1b15eb9a35ef930ba373fc945ee84dcabededdff with merge 05f512090ead730e044e52571dba2da5ff8f9dcf...

[eddyb]

(I forgot to push some comment tweaks, hopefully it doesn't kill the try build)

[eddyb]

@rust-timer build 05f512090ead730e044e52571dba2da5ff8f9dcf

[rust-timer]

Queued 05f512090ead730e044e52571dba2da5ff8f9dcf with parent f4c675c, future comparison URL.

[rust-timer]

Finished benchmarking try commit 05f512090ead730e044e52571dba2da5ff8f9dcf, comparison URL.

[eddyb]

The worst regression, clap-rs-check clean has 143 less specializes calls? What's up with that?
But typeck_tables_of getting slower makes sense, and clap-rs-debug clean also has that.

I guess I'll try to come up with a synthetic stress test that hits the worst case of this PR (wf::obligations with ty::Infer that resolves to a different ty::Infer), and see what I can do to alleviate it.

But still, 7.5% regression of typeck_tables_of, from this change? That means a lot of clap-rs' type-checking time is spent on solving WF obligations.

[bors]

☔ The latest upstream changes (presumably #70205) made this pull request unmergeable. Please resolve the merge conflicts.

[eddyb]

Wait I forgot about this, this is technically a bugfix (before, if compute returned false here, those inference variables would be forgotten about).

So it might make sense that it's slower. Let me, uh, make a PR that only changes this, and check that for perf regressions.

EDIT: Actually, there's the entirety of #70168, predicate_obligations being the other user of .compute w/o checking the return value. I might be able to pinpoint where the extra obligations are coming from.

[eddyb]

This is it.

typeck_tables_of on clap-rs gets slightly faster (instead of being a significant regression), if I ignore upvar types that are ty::Infer(ty::TyVar(_)) and don't resolve.

But that's a WF hole, right? Do we just accept the regression?

[nikomatsakis]

Sorry, I lost track of this PR. I agree that this is a bug-fix.

[eddyb]

I added some fast paths, let's see if they improve anything (but also see #70170 (comment)):

@bors try @rust-timer queue

[rust-timer]

Awaiting bors try build completion

[bors]

⌛ Trying commit ef83aa34b11d92b10f53b514d04a5c2b475707d2 with merge bb0cbcbd0a0e408606210fa280cf86c90448764c...

[bors]

☀️ Try build successful - checks-azure
Build commit: bb0cbcbd0a0e408606210fa280cf86c90448764c (bb0cbcbd0a0e408606210fa280cf86c90448764c)

[rust-timer]

Queued bb0cbcbd0a0e408606210fa280cf86c90448764c with parent 424c793, future comparison URL.

[rust-timer]

Finished benchmarking try commit bb0cbcbd0a0e408606210fa280cf86c90448764c, comparison URL.

[eddyb]

Hehe, I think my second commit, wf: {Int,Float}Var can only infer to always-WF ints/floats., made the old macro-expanded version of inflate that perf.rust-lang.org tests, significantly faster to compile, it seems, so I guess it was spending a lot of time on WF-ing ty::IntVars?

I'm not surprised, there were a lot of macro-replicated integer literals in that code.

[nikomatsakis]

I'm inclined to merge this PR @eddyb -- r=me on the code itself -- the clap-rs-debug perf regression seems small and I agree it's a bug fix, and there are some significant (albeit somewhat surprising?) improvements, too. I guess your explanation makes sense, though.

[eddyb]

Let's get some new numbers, post-rust-lang/rustc-perf#645:

@bors try @rust-timer queue

[rust-timer]

Awaiting bors try build completion

[bors]

⌛ Trying commit 05a872d with merge ad57f0a2528d29aca128300e3fa1b174a9eb497e...

[bors]

☀️ Try build successful - checks-azure
Build commit: ad57f0a2528d29aca128300e3fa1b174a9eb497e (ad57f0a2528d29aca128300e3fa1b174a9eb497e)

[rust-timer]

Queued ad57f0a2528d29aca128300e3fa1b174a9eb497e with parent e94eaa6, future comparison URL.

[rust-timer]

Finished benchmarking try commit ad57f0a2528d29aca128300e3fa1b174a9eb497e, comparison URL.

[eddyb]

@bors r=nikomatsakis rollup=never

[bors]

📌 Commit 05a872d has been approved by nikomatsakis

[bors]

⌛ Testing commit 05a872d with merge 08dfbfb...

[bors]

☀️ Test successful - checks-azure
Approved by: nikomatsakis
Pushing 08dfbfb to master...