This wasn't a right decision in the first place, the feature flag was
broken in the last rustfmt release, and syntax highlighting of imports
is more important anyway
This gives a significant speedup, because chalk will call these
functions several times even withing a single revision. The only
significant one here is `impl_data`, but I figured it might be good to
cache others just for consistency.
The results I get are:
Before:
from scratch: 16.081457952s
no change: 15.846493ms
trivial change: 352.95592ms
comment change: 361.998408ms
const change: 457.629212ms
After:
from scratch: 14.910610278s
no change: 14.934647ms
trivial change: 85.633023ms
comment change: 96.433023ms
const change: 171.543296ms
Seems like a nice win!
Note that we can't just remove CheckCanceled trait altogether:
sometimes it's useful to check for cancellation while the query is
running! We do this, for example, in the name resolution fixed-point
loop.
We use panics for cancellation, so we could trigger panic while
holding the solver. std::sync::Mutex will be poisoned as a result,
which and all further attempts to use solver (from other threads) will
panic as well.
This commit switches to parking_lot::Mutex which just unlocks on panic.
Reducing it to 2 was just a failed attempt to see whether that would help fix
some slow cases; in fact, it can create new slow cases by replacing concrete
types by variables.
For Send/Sync/Sized, we don't handle auto traits correctly yet and because they
have a lot of impls, they can easily lead to slowdowns. In the case of
Fn/FnMut/FnOnce, we don't parse the special Fn notation correctly yet and don't
handle closures yet, so we are very unlikely to find an impl.
