Reduce TypedArena creations in check_match.

[nnethercote]

check_match creates a new TypedArena for every call to
create_and_enter. DHAT tells me that each TypedArena typically is
barely used, with typically a single allocation per arena.

This commit moves the TypedArena creation outwards a bit, into
check_match, and then passes it into create_and_enter. This reduces
the number of arenas created by about 4-5x, for a very small perf win.
(Moving the arena creation further outwards is hard because
check_match is a query.)

r? @oli-obk

[nnethercote]

@bors try @rust-timer queue

[rust-timer]

Awaiting bors try build completion

[bors]

⌛ Trying commit cbc577f with merge a303191d82d1d452f5dfddc9efd98c3341e64b66...

[bors]

☀️ Try build successful - checks-actions, checks-azure
Build commit: a303191d82d1d452f5dfddc9efd98c3341e64b66 (a303191d82d1d452f5dfddc9efd98c3341e64b66)

[rust-timer]

Queued a303191d82d1d452f5dfddc9efd98c3341e64b66 with parent 97f3eee, future comparison URL.

[oli-obk]

Why are we even using an arena here? We use Box<PatKind<'tcx>> for cycles (and PatKind just uses Pat by value, not interned). Is this a leftover from previous refactorings or was the plan to use interned pats in PatKind and stop boxin PatKind?

cc @varkor do you remember anything here?

[varkor]

I don't remember why TypedArena is used here. cc @Nadrieril who probably has this code more in-cache.

[Nadrieril]

It's used a lot in _match, where the matrix stores only arena-allocated Pats. Since the matrix gets cloned a lot, I imagine this is good for performance, as opposed to owned Pats ?

[nnethercote]

It's used a lot in _match, where the matrix stores only arena-allocated Pats. Since the matrix gets cloned a lot, I imagine this is good for performance, as opposed to owned Pats ?

I tried removing the arena and although I never got it close to compiling (so many lifetime changes needed) I concluded that this was probably the reason.

[oli-obk]

@bors r+

[bors]

📌 Commit d26d187 has been approved by oli-obk

[rust-timer]

Finished benchmarking try commit a303191d82d1d452f5dfddc9efd98c3341e64b66, comparison URL.