| Age | Commit message (Collapse) | Author | Lines |
|---|
|
|
|
|
|
Some `let chains` clean-up
Not sure if this kind of clean-up is welcoming because of size, but I decided to try out one
r? compiler
|
|
|
|
|
|
|
|
|
|
Bringing `rustc_rayon_core` in tree as `rustc_thread_pool`
This PR moves [`rustc_rayon_core`](https://github.com/rust-lang/rustc-rayon/tree/5fadf44/rayon-core) from commit `5fadf44` as suggested in [this zulip thread](https://rust-lang.zulipchat.com/#narrow/channel/187679-t-compiler.2Fparallel-rustc/topic/Bringing.20.60rustc_rayon_core.60.20in.20tree). I tried to split the work into separate commits so it is easy to review. The first commit is a simple copy and paste from the fork, and subsequent changes were made to use the new crate and to ensure the new crate complies with different format and lint expectations.
**Call-out:** I was also wondering if I need to make any further changes to accommodate licensing requirements.
r? oli-obk
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Use thread local dep graph encoding
This adds thread local encoding of dep graph nodes. Each thread has a `MemEncoder` that gets flushed to the global `FileEncoder` when it exceeds 64 kB. Each thread also has a local cache of dep indices. This means there can now be empty gaps in `SerializedDepGraph`.
Indices are marked green and also allocated by the new atomic operation `DepNodeColorMap::try_mark_green` as the encoder lock is removed.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Allow out of order dep graph node encoding
This allows out of order dep graph node encoding by also encoding the index instead of using the file node order as the index.
`MemEncoder` is also brought back to life and used for encoding.
Both of these are done to enable thread-local encoding of dep graph nodes.
This is based on https://github.com/rust-lang/rust/pull/139636.
|
|
|
|
|
|
|
|
|
|
|
|
Use a session counter to make anon dep nodes unique
This changes the unique session hash used to ensure unique anon dep nodes per session from a timestamp to a counter.
This is nicer for debugging as it makes the dep graph deterministic.
|
|
Encode dep node edge count as u32 instead of usize
This encodes the dep node edge count as u32 instead of usize. It doesn't need to be usize as the count is limited the the unique number of dep nodes which we use a 32-bit index for. It probably saves some branches for encoding / decoding, but it isn't too hot as the count gets packed in with other fields when it's low.
<table><tr><td rowspan="2">Benchmark</td><td colspan="1"><b>Before</b></th><td colspan="2"><b>After</b></th><td colspan="1"><b>Before</b></th><td colspan="2"><b>After</b></th><td colspan="1"><b>Before</b></th><td colspan="2"><b>After</b></th></tr><tr><td align="right">Time</td><td align="right">Time</td><td align="right">%</th><td align="right">Physical Memory</td><td align="right">Physical Memory</td><td align="right">%</th><td align="right">Committed Memory</td><td align="right">Committed Memory</td><td align="right">%</th></tr><tr><td>🟣 <b>clap</b>:check:unchanged</td><td align="right">0.3343s</td><td align="right">0.3343s</td><td align="right"> -0.00%</td><td align="right">97.02 MiB</td><td align="right">96.89 MiB</td><td align="right"> -0.13%</td><td align="right">167.93 MiB</td><td align="right">167.80 MiB</td><td align="right"> -0.08%</td></tr><tr><td>🟣 <b>hyper</b>:check:unchanged</td><td align="right">0.1353s</td><td align="right">0.1350s</td><td align="right"> -0.19%</td><td align="right">61.91 MiB</td><td align="right">61.83 MiB</td><td align="right"> -0.13%</td><td align="right">124.64 MiB</td><td align="right">124.64 MiB</td><td align="right"> 0.00%</td></tr><tr><td>🟣 <b>regex</b>:check:unchanged</td><td align="right">0.2479s</td><td align="right">0.2481s</td><td align="right"> 0.10%</td><td align="right">78.34 MiB</td><td align="right">78.35 MiB</td><td align="right"> 0.02%</td><td align="right">145.20 MiB</td><td align="right">145.31 MiB</td><td align="right"> 0.08%</td></tr><tr><td>🟣 <b>syn</b>:check:unchanged</td><td align="right">0.5347s</td><td align="right">0.5333s</td><td align="right"> -0.26%</td><td align="right">118.63 MiB</td><td align="right">118.66 MiB</td><td align="right"> 0.02%</td><td align="right">193.05 MiB</td><td align="right">193.11 MiB</td><td align="right"> 0.03%</td></tr><tr><td>Total</td><td align="right">1.2521s</td><td align="right">1.2507s</td><td align="right"> -0.11%</td><td align="right">355.90 MiB</td><td align="right">355.74 MiB</td><td align="right"> -0.05%</td><td align="right">630.83 MiB</td><td align="right">630.87 MiB</td><td align="right"> 0.01%</td></tr><tr><td>Summary</td><td align="right">1.0000s</td><td align="right">0.9991s</td><td align="right"> -0.09%</td><td align="right">1 byte</td><td align="right">1.00 bytes</td><td align="right"> -0.06%</td><td align="right">1 byte</td><td align="right">1.00 bytes</td><td align="right"> 0.01%</td></tr></table>
|
|
|
|
|
|
|
|
|
|
|
|
expand: Leave traces when expanding `cfg` attributes
This is the same as https://github.com/rust-lang/rust/pull/138515, but for `cfg(true)` instead of `cfg_attr`.
The difference is that `cfg(true)`s already left "traces" after themselves - the `cfg` attributes themselves, with `expanded_inert_attrs` set to true, with full tokens, available to proc macros.
This is not a reasonably expected behavior, but it could not be removed without a replacement, because a [major rustdoc feature](https://github.com/rust-lang/rfcs/pull/3631) and a number of clippy lints rely on it. This PR implements a replacement.
This needs a crater run, because it changes observable behavior (in an intended way) - proc macros can no longer see expanded `cfg(true)` attributes.
(Some minor unnecessary special casing for `sym::cfg_attr` is also removed in this PR.)
r? `@nnethercote`
|
|
Avoiding calling queries when collecting active queries
This PR changes active query collection to no longer call queries. Instead the fields needing queries have their computation delayed to when an cycle error is emitted or when printing the query backtrace in a panic.
This is done by splitting the fields in `QueryStackFrame` needing queries into a new `QueryStackFrameExtra` type. When collecting queries `QueryStackFrame` will contain a closure that can create `QueryStackFrameExtra`, which does make use of queries. Calling `lift` on a `QueryStackFrame` or `CycleError` will convert it to a variant containing `QueryStackFrameExtra` using those closures.
This also only calls queries needed to collect information on a cycle errors, instead of information on all active queries.
Calling queries when collecting active queries is a bit odd. Calling queries should not be done in the deadlock handler at all.
This avoids the out of memory scenario in https://github.com/rust-lang/rust/issues/124901.
|
|
Remove `prev_index_to_index` field from `CurrentDepGraph`
The dep graph currently has 2 ways to map a previous index into a current index. The `prev_index_to_index` map stores the current index equivalent of a previous index. For indices which are marked green, we also store the same information in the `DepNodeColorMap`. We actually only need to known the mapping for green nodes however, so this PR removes `prev_index_to_index` and instead makes use of the `DepNodeColorMap`.
To avoid racing when promoting a node from the previous session, the encoder lock is now used to ensure only one thread encodes the promoted node. This was previously done by the lock in `prev_index_to_index`.
This also changes `nodes_newly_allocated_in_current_session` used to detect duplicate dep nodes to contain both new and previous nodes, which is simpler and can better catch duplicates.
The dep node index encoding used in `DepNodeColorMap` is tweak to avoid subtraction / addition to optimize accessing the current equivalent of a previous index.
r? `@oli-obk`
|
|
|
|
|
|
|
|
Batch mark waiters as unblocked when resuming in the deadlock handler
This fixes a race when resuming multiple threads to resolve query cycles. This now marks all threads as unblocked before resuming any of them. Previously if one was resumed and marked as unblocked at a time. The first thread resumed could fall asleep then Rayon would detect a second false deadlock. Later the initial deadlock handler thread would resume further threads.
This also reverts the workaround added in https://github.com/rust-lang/rust/pull/137731.
cc `@SparrowLii` `@lqd`
|
|
Abort in deadlock handler if we fail to get a query map
Resolving query cycles requires the complete active query map, or it may miss query cycles. We did not check that the map is completely constructed before. If there is some error collecting the map, something has gone wrong already. This adds a check to abort/panic if we fail to construct the complete map.
This can help differentiate errors from the `deadlock detected` case if constructing query map has errors in practice.
An `Option` is not used for `collect_active_jobs` as the panic handler can still make use of a partial map.
|
|
|
|
Only use the new node hashmap for anonymous nodes
This is a rebase of https://github.com/rust-lang/rust/pull/112469.
cc `@cjgillot`
|
|
|
|
|
|
This reverts commit cc1e4ede9388d87750c3751f41e8c6c4f6cae995.
|
|
|
