The latter is a new module.
As well as the code motion, some other changes were required.
- `QueryJobId` methods became free functions so they could move while
`QueryJobId` itself stayed put. This was so `QueryMap` and
`QueryJobInfo` could be moved.
- Some visibilities in `rustc_query_system` required changing.
- `collect_active_jobs_from_all_queries` is no longer required in `trait
QueryContext`.
It existed only to bridge the divide between `rustc_query_system` and
`rustc_query_impl`. But enough pieces have been moved from the former to
the latter that the trait is no longer needed. Less indirection and
abstraction makes the code easier to understand.
Query keys must be stably hashable. Currently this requirement is
expressed as a where-clause on `impl QueryDispatcher for
SemiDynamicQueryDispatcher` and a where-clause on
`create_deferred_query_stack_frame`.
This commit removes those where-clause bounds and adds a single bound to
`QueryCache::Key`, which already has some other bounds. I.e. it
consolidates the bounds. It also gives them a name (`QueryCacheKey`) to
avoid repeating them. There is also a related `Key` trait in
`rustc_middle`; it should probably be merged with `QueryCacheKey` in the
future, but not today.
This cleanup helps with the next two commits, which do bigger
rearrangements, and where the where-clauses caused me some difficulties.
The previous commit moved some code from `rustc_query_system`, which
doesn't have access to `TyCtxt` and `QueryCtxt`, to `rustc_query_impl`,
which does. We can now remove quite a bit of indirection.
- Three methods in `trait QueryContext` are no longer needed
(`next_job_id`, `current_query_job`, `start_query`). As a result,
`QueryCtxt`'s trait impls of these methods are changed to inherent
methods.
- `qcx: Q::Qcx` parameters are simplified to `qcx: QueryCtxt<'tcx>`.
- `*qcx.dep_context()` occurrences are simplified to `qcx.tcx`, and
things like `qcx.dep_context().profiler()` become `qcx.tcx.prof`.
- `DepGraphData<<Q::Qcx as HasDepContext>::Deps>` becomes
`DepGraphData<DepsType>`.
In short, various layers of indirection and abstraction are cut away.
The resulting code is simpler, more concrete, and easier to understand.
It's a good demonstration of the benefits of eliminating
`rustc_query_system`, and there will be more to come.
[Note: this commit conceptually moves 75% of a file to another location,
and leaves 25% of it behind. It's impossible to preserve all the git
history. To preserve git history of the moved 75%, in this commit we
rename the file and remove the 25% from it, leaving the code in an
incomplete (uncompilable) state. In the next commit we add back the
25% in the old location.]
We are in the process of eliminating `rustc_query_system`. Chunks of it
are unused by `rustc_middle`, and so can be moved into
`rustc_query_impl`. This commit does some of that.
Mostly it's just moving code from one file to a new file. There are a
couple of non-trivial changes.
- `QueryState` and `ActiveKeyStatus` must remain in `rustc_query_system`
because they are used by `rustc_middle`. But their inherent methods
are not used by `rustc_middle`. So these methods are moved and
converted to free functions.
- The visibility of some things must increase. This includes `DepGraphData`
and some of its methods, which are now used in `rustc_query_impl`.
This is a bit annoying but seems hard to avoid.
What little is left behind in
`compiler/rustc_query_system/src/query/plumbing.rs` will be able to
moved into `rustc_query_impl` or `rustc_middle` in the future.
`QueryStackFrame<I>` is a generic type, and the `I` parameter leaks out
to many other related types. However, it only has two instantiations in
practice:
- `QueryStackFrame<QueryStackFrameExtra>`
- `QueryStackFrame<QueryStackDeferred<'tcx>>`
And most of the places that currently use `QueryStackFrame<I>` are
actually specific to one of the instantiations. This commit removes the
unneeded genericity.
The following types are only ever used with `QueryStackDeferred<'tcx>`:
- QueryMap
- QueryJobInfo
- QueryJob
- QueryWaiter
- QueryLatchInfo
- QueryLatch
- QueryState
- QueryResult
and their `<I>` parameter is changed to `<'tcx>`.
Also, the `QueryContext::QueryInfo` associated type is removed.
`CycleError<I>` and `QueryInfo<I>` are still generic over type, because
they are used with both instantiations.
This required also adding a `<'tcx>` parameter to the traits
`QueryDispatcher` and `QueryContext`, which is annoying but I can't see
how to avoid it.
I tried removing it in #151203, to replace it with something simpler.
But a couple of fuzzing failures have come up and I don't have a clear
picture on how to fix them. So I'm reverting the main part of #151203.
This commit also adds the two fuzzing tests.
Fixes#151226, #151358.
PR #138672 introduced a complex and invasive split of `QueryStackFrame`
to avoid a query cycle. This commit reverts that change because there is
a much simpler change that fixes the problem, which will be in the next
commit.
Since it's inception a long time ago, the parallel compiler and its cfgs
have been a maintenance burden. This was a necessary evil the allow
iteration while not degrading performance because of synchronization
overhead.
But this time is over. Thanks to the amazing work by the parallel
working group (and the dyn sync crimes), the parallel compiler has now
been fast enough to be shipped by default in nightly for quite a while
now.
Stable and beta have still been on the serial compiler, because they
can't use `-Zthreads` anyways.
But this is quite suboptimal:
- the maintenance burden still sucks
- we're not testing the serial compiler in nightly
Because of these reasons, it's time to end it. The serial compiler has
served us well in the years since it was split from the parallel one,
but it's over now.
Let the knight slay one head of the two-headed dragon!
Foreign maps are used to cache external DefIds, typically backed by
metadata decoding. In the future we might skip caching `V` there (since
loading from metadata usually is already cheap enough), but for now this
cuts down on the impact to memory usage and time to None-init a bunch of
memory. Foreign data is usually much sparser, since we're not usually
loading *all* entries from the foreign crate(s).
This may introduce additional mono _but_ may help const fold things
better and especially may help not constructing a `QueryVTable` anymore
which is cheap but not free.
