Slightly optimize hash map stable hashing
I was profiling some of the `rustc-perf` benchmarks locally and noticed that quite some time is spent inside the stable hash of hashmaps. I tried to use a `SmallVec` instead of a `Vec` there, which helped very slightly.
Then I tried to remove the sorting, which was a bottleneck, and replaced it with insertion into a binary heap. Locally, it yielded nice improvements in instruction counts and RSS in several benchmarks for incremental builds. The implementation could probably be much nicer and possibly extended to other stable hashes, but first I wanted to test the perf impact properly.
Can I ask someone to do a perf run? Thank you!
This largely avoids remapping from and to the 'real' indices, with the exception
of predecessor lookup and the final merge back, and is conceptually better.
As the paper indicates, the unprocessed vertices in the DFS tree and processed
vertices are disjoint, and we can use them in the same space, tracking only the index
of the split.
This replaces the previous implementation with the simple variant of
Lengauer-Tarjan, which performs better in the general case. Performance on the
keccak benchmark is about equivalent between the two, but we don't see
regressions (and indeed see improvements) on other benchmarks, even on a
partially optimized implementation.
The implementation here follows that of the pseudocode in "Linear-Time
Algorithms for Dominators and Related Problems" thesis by Loukas Georgiadis. The
next few commits will optimize the implementation as suggested in the thesis.
Several related works are cited in the comments within the implementation, as
well.
Implement the simple Lengauer-Tarjan algorithm
This replaces the previous implementation (from #34169), which has not been
optimized since, with the simple variant of Lengauer-Tarjan which performs
better in the general case. A previous attempt -- not kept in commit history --
attempted a replacement with a bitset-based implementation, but this led to
regressions on perf.rust-lang.org benchmarks and equivalent wins for the keccak
benchmark, so was rejected.
The implementation here follows that of the pseudocode in "Linear-Time
Algorithms for Dominators and Related Problems" thesis by Loukas Georgiadis. The
next few commits will optimize the implementation as suggested in the thesis.
Several related works are cited in the comments within the implementation, as
well.
On the keccak benchmark, we were previously spending 15% of our cycles computing
the NCA / intersect function; this function is quite expensive, especially on
modern CPUs, as it chases pointers on every iteration in a tight loop. With this
commit, we spend ~0.05% of our time in dominator computation.
There's a conversation in the tracking issue about possibly unaccepting `in_band_lifetimes`, but it's used heavily in the compiler, and thus there'd need to be a bunch of PRs like this if that were to happen.
So here's one to see how much of an impact it has.
(Oh, and I removed `nll` while I was here too, since it didn't seem needed. Let me know if I should put that back.)
Implement write() method for Box<MaybeUninit<T>>
This adds method similar to `MaybeUninit::write` main difference being
it returns owned `Box`. This can be used to elide copy from stack
safely, however it's not currently tested that the optimization actually
occurs.
Analogous methods are not provided for `Rc` and `Arc` as those need to
handle the possibility of sharing. Some version of them may be added in
the future.
This was discussed in #63291 which this change extends.
This adds method similar to `MaybeUninit::write` main difference being
it returns owned `Box`. This can be used to elide copy from stack
safely, however it's not currently tested that the optimization actually
occurs.
Analogous methods are not provided for `Rc` and `Arc` as those need to
handle the possibility of sharing. Some version of them may be added in
the future.
This was discussed in #63291 which this change extends.
Revert "Add rustc lint, warning when iterating over hashmaps"
Fixes perf regressions introduced in https://github.com/rust-lang/rust/pull/90235 by temporarily reverting the relevant PR.
Particularly for ctfe-stress-4, the hashing of byte slices as part of the
MIR Allocation is quite hot. Previously, we were falling back on byte-by-byte
copying of the slice into the SipHash buffer (64 bytes long) before hashing a 64
byte chunk, and then doing that again and again.
This should hopefully be an improvement for that code.
Stabilize `const_panic`
Closes#51999
FCP completed in #89006
```@rustbot``` label +A-const-eval +A-const-fn +T-lang
cc ```@oli-obk``` for review (not `r?`'ing as not on lang team)
Rollup of 12 pull requests
Successful merges:
- #88795 (Print a note if a character literal contains a variation selector)
- #89015 (core::ascii::escape_default: reduce struct size)
- #89078 (Cleanup: Remove needless reference in ParentHirIterator)
- #89086 (Stabilize `Iterator::map_while`)
- #89096 ([bootstrap] Improve the error message when `ninja` is not found to link to installation instructions)
- #89113 (dont `.ensure()` the `thir_abstract_const` query call in `mir_build`)
- #89114 (Fixes a technicality regarding the size of C's `char` type)
- #89115 (⬆️ rust-analyzer)
- #89126 (Fix ICE when `indirect_structural_match` is allowed)
- #89141 (Impl `Error` for `FromSecsError` without foreign type)
- #89142 (Fix match for placeholder region)
- #89147 (add case for checking const refs in check_const_value_eq)
Failed merges:
r? `@ghost`
`@rustbot` modify labels: rollup
Migrate in-tree crates to 2021
This replaces #89075 (cherry picking some of the commits from there), and closes#88637 and fixes#89074.
It excludes a migration of the library crates for now (see tidy diff) because we have some pending bugs around macro spans to fix there.
I instrumented bootstrap during the migration to make sure all crates moved from 2018 to 2021 had the compatibility warnings applied first.
Originally, the intent was to support cargo fix --edition within bootstrap, but this proved fairly difficult to pull off. We'd need to architect the check functionality to support running cargo check and cargo fix within the same x.py invocation, and only resetting sysroots on check. Further, it was found that cargo fix doesn't behave too well with "not quite workspaces", such as Clippy which has several crates. Bootstrap runs with --manifest-path ... for all the tools, and this makes cargo fix only attempt migration for that crate. We can't use e.g. --workspace due to needing to maintain sysroots for different phases of compilation appropriately.
It is recommended to skip the mass migration of Cargo.toml's to 2021 for review purposes; you can also use `git diff d6cd2c6c87 -I'^edition = .20...$'` to ignore the edition = 2018/21 lines in the diff.
Rework DepthFirstSearch API
This expands the API to be more flexible, allowing for more visitation patterns
on graphs. This will be useful to avoid extra datasets (and allocations) in
cases where the expanded DFS API is sufficient.
This also fixes a bug with the previous DFS constructor, which left the start
node not marked as visited (even though it was immediately returned).
Commit written by ```@nikomatsakis``` originally, cherry picked from several commits in work on never type stabilization, but stands alone.
This expands the API to be more flexible, allowing for more visitation patterns
on graphs. This will be useful to avoid extra datasets (and allocations) in
cases where the expanded DFS API is sufficient.
This also fixes a bug with the previous DFS constructor, which left the start
node not marked as visited (even though it was immediately returned).
Remove SmallVector mention
SmallVector is long gone, as it's been first replaced
by OneVector in commit e5e6375352,
which then has been removed entirely in favour of SmallVec in
commit 130a32fa72.
SmallVector is long gone, as it's been first replaced
by OneVector in commit e5e6375352,
which then has been removed entirely in favour of SmallVec in
commit 130a32fa72.
Since RFC 3052 soft deprecated the authors field anyway, hiding it from
crates.io, docs.rs, and making Cargo not add it by default, and it is
not generally up to date/useful information, we should remove it from
crates in this repo.
Make mir borrowck's use of opaque types independent of the typeck query's result
fixes#87218fixes#86465
we used to use the typeck results only to generate an obligation for the mir borrowck type to be equal to the typeck result.
When i removed the `fixup_opaque_types` function in #87200, I exposed a bug that showed that mir borrowck can't doesn't get enough information from typeck in order to build the correct lifetime mapping from opaque type usage to the actual concrete type. We therefor now fully compute the information within mir borrowck (we already did that, but we only used it to verify the typeck result) and stop using the typeck information.
We will likely be able to remove most opaque type information from the borrowck results in the future and just have all current callers use the mir borrowck result instead.
r? `@spastorino`
Profile incremental compilation hashing fingerprints
Adds profiling instrumentation for the hashing of incremental compilation fingerprints per query.
This will eventually feed into the `measureme` and `rustc-perf` infrastructure for tracking if computing hashes changes over time.
TODOs:
* [x] Address the FIXME where we are including node interning in the hash timing.
* [ ] Update measureme/summarize to handle this new data: https://github.com/rust-lang/measureme/pull/166
* [ ] ~Update rustc-perf to handle the new data from measureme~ (will be done at a later time)
r? `@ghost`
cc `@michaelwoerister`
rustc_data_structures has a dependency on crossbeam-utils but never uses
it. It appears to have originally had this dependency in order to set
the "nightly" feature; however, its other dependencies use a different
version of crossbeam-utils, so this doesn't actually affect anything.
Furthermore, in current crossbeam-utils, the "nightly" feature has
become a no-op.
Don't use a generator for BoxedResolver
The generator is non-trivial and requires unsafe code anyway. Using regular unsafe code without a generator is much easier to follow.
Based on #85810 as it touches rustc_interface too.
Remove unused feature gates
The first commit removes a usage of a feature gate, but I don't expect it to be controversial as the feature gate was only used to workaround a limitation of rust in the past. (closures never being `Clone`)
The second commit uses `#[allow_internal_unstable]` to avoid leaking the `trusted_step` feature gate usage from inside the index newtype macro. It didn't work for the `min_specialization` feature gate though.
The third commit removes (almost) all feature gates from the compiler that weren't used anyway.
don't enable parking_lot nightly features
Having the compiler itself depend on external libraries that use nightly features can lead to "fun" bootstrap situations. Within the rustc repo we use `cfg(bootstrap)` to resolve those, but that is not a reasonable option for external dependencies.
So I propose we stop enabling the "nightly" feature of `parking_lot` here. In my experiments, this then indeed leads to the feature not being enabled (i.e., nothing else enables it), and everything still builds. However, this means parking_lot's `RwLock` will no longer have hardware lock elision for readers -- I hope that is okay to lose in exchange for less bootstrap brain twisting. ;)
Cc `@Amanieu`
With the arrival of min const generics, many alt-vec libraries have
updated to use it in some way and arrayvec is no exception. Use the
latest with minor refactoring.
Also, rustc_workspace_hack is the only user of smallvec 0.6 in the
entire tree, so drop it.
Add `FromIterator` and `IntoIterator` impls for `ThinVec`
These should make using `ThinVec` feel much more like using `Vec`.
They will allow users of `Vec` to switch to `ThinVec` while continuing
to use `collect()`, `for` loops, and other parts of the iterator API.
I don't know if there were use cases before for using the iterator API
with `ThinVec`, but I would like to start using `ThinVec` in rustdoc,
and having it conform to the iterator API would make the transition
*a lot* easier.
I added a `FromIterator` impl, an `IntoIterator` impl that yields owned
elements, and `IntoIterator` impls that yield immutable or mutable
references to elements. I also added some unit tests for `ThinVec`.
These should make using `ThinVec` feel much more like using `Vec`.
They will allow users of `Vec` to switch to `ThinVec` while continuing
to use `collect()`, `for` loops, and other parts of the iterator API.
I don't know if there were use cases before for using the iterator API
with `ThinVec`, but I would like to start using `ThinVec` in rustdoc,
and having it conform to the iterator API would make the transition
*a lot* easier.
I added a `FromIterator` impl, an `IntoIterator` impl that yields owned
elements, and `IntoIterator` impls that yield immutable or mutable
references to elements. I also added some unit tests for `ThinVec`.
Add an Mmap wrapper to rustc_data_structures
This wrapper implements StableAddress and falls back to directly reading the file on wasm32.
Taken from #83640, which I will close due to the perf regression.
There isn't currently a good reviewer for these, and I don't want to
remove things that will just be added again. I plan to make a separate
PR for these changes so the rest of the cleanup can land.
- Add back `HirIdVec`, with a comment that it will soon be used.
- Add back `*_region` functions, with a comment they may soon be used.
- Remove `-Z borrowck_stats` completely. It didn't do anything.
- Remove `make_nop` completely.
- Add back `current_loc`, which is used by an out-of-tree tool.
- Fix style nits
- Remove `AtomicCell` with `cfg(parallel_compiler)` for consistency.
Found with https://github.com/est31/warnalyzer.
Dubious changes:
- Is anyone else using rustc_apfloat? I feel weird completely deleting
x87 support.
- Maybe some of the dead code in rustc_data_structures, in case someone
wants to use it in the future?
- Don't change rustc_serialize
I plan to scrap most of the json module in the near future (see
https://github.com/rust-lang/compiler-team/issues/418) and fixing the
tests needed more work than I expected.
TODO: check if any of the comments on the deleted code should be kept.
Allow for reading raw bytes from rustc_serialize::Decoder without unsafe code
The current `read_raw_bytes` method requires using `MaybeUninit` and `unsafe`. I don't think this is necessary. Let's see if a safe interface has any performance drawbacks.
This is a followup to #83273 and will make it easier to rebase #82183.
r? `@cjgillot`
Update to rustc-rayon 0.3.1
This pulls in rust-lang/rustc-rayon#8 to fix#81425. (h/t `@ammaraskar)`
That revealed weak constraints on `rustc_arena::DropArena`, because its
`DropType` was holding type-erased raw pointers to generic `T`. We can
implement `Send` for `DropType` (under `cfg(parallel_compiler)`) by
requiring all `T: Send` before they're type-erased.
This pulls in rust-lang/rustc-rayon#8 to fix#81425. (h/t @ammaraskar)
That revealed weak constraints on `rustc_arena::DropArena`, because its
`DropType` was holding type-erased raw pointers to generic `T`. We can
implement `Send` for `DropType` (under `cfg(parallel_compiler)`) by
requiring all `T: Send` before they're type-erased.
Let a portion of DefPathHash uniquely identify the DefPath's crate.
This allows to directly map from a `DefPathHash` to the crate it originates from, without constructing side tables to do that mapping -- something that is useful for incremental compilation where we deal with `DefPathHash` instead of `DefId` a lot.
It also allows to reliably and cheaply check for `DefPathHash` collisions which allows the compiler to gracefully abort compilation instead of running into a subsequent ICE at some random place in the code.
The following new piece of documentation describes the most interesting aspects of the changes:
```rust
/// A `DefPathHash` is a fixed-size representation of a `DefPath` that is
/// stable across crate and compilation session boundaries. It consists of two
/// separate 64-bit hashes. The first uniquely identifies the crate this
/// `DefPathHash` originates from (see [StableCrateId]), and the second
/// uniquely identifies the corresponding `DefPath` within that crate. Together
/// they form a unique identifier within an entire crate graph.
///
/// There is a very small chance of hash collisions, which would mean that two
/// different `DefPath`s map to the same `DefPathHash`. Proceeding compilation
/// with such a hash collision would very probably lead to an ICE and, in the
/// worst case, to a silent mis-compilation. The compiler therefore actively
/// and exhaustively checks for such hash collisions and aborts compilation if
/// it finds one.
///
/// `DefPathHash` uses 64-bit hashes for both the crate-id part and the
/// crate-internal part, even though it is likely that there are many more
/// `LocalDefId`s in a single crate than there are individual crates in a crate
/// graph. Since we use the same number of bits in both cases, the collision
/// probability for the crate-local part will be quite a bit higher (though
/// still very small).
///
/// This imbalance is not by accident: A hash collision in the
/// crate-local part of a `DefPathHash` will be detected and reported while
/// compiling the crate in question. Such a collision does not depend on
/// outside factors and can be easily fixed by the crate maintainer (e.g. by
/// renaming the item in question or by bumping the crate version in a harmless
/// way).
///
/// A collision between crate-id hashes on the other hand is harder to fix
/// because it depends on the set of crates in the entire crate graph of a
/// compilation session. Again, using the same crate with a different version
/// number would fix the issue with a high probability -- but that might be
/// easier said then done if the crates in questions are dependencies of
/// third-party crates.
///
/// That being said, given a high quality hash function, the collision
/// probabilities in question are very small. For example, for a big crate like
/// `rustc_middle` (with ~50000 `LocalDefId`s as of the time of writing) there
/// is a probability of roughly 1 in 14,750,000,000 of a crate-internal
/// collision occurring. For a big crate graph with 1000 crates in it, there is
/// a probability of 1 in 36,890,000,000,000 of a `StableCrateId` collision.
```
Given the probabilities involved I hope that no one will ever actually see the error messages. Nonetheless, I'd be glad about some feedback on how to improve them. Should we create a GH issue describing the problem and possible solutions to point to? Or a page in the rustc book?
r? `@pnkfelix` (feel free to re-assign)
Update measureme dependency to the latest version
This version adds the ability to use `rdpmc` hardware-based performance
counters instead of wall-clock time for measuring duration. This also
introduces a dependency on the `perf-event-open-sys` crate on Linux
which is used when using hardware counters.
r? ```@oli-obk```
Replace const_cstr with cstr crate
This PR replaces the `const_cstr` macro inside `rustc_data_structures` with `cstr` macro from [cstr](https://crates.io/crates/cstr) crate.
The two macros basically serve the same purpose, which is to generate `&'static CStr` from a string literal. `cstr` is better because it validates the literal at compile time, while the existing `const_cstr` does it at runtime when `debug_assertions` is enabled. In addition, the value `cstr` generates can be used in constant context (which is seemingly not needed anywhere currently, though).
This version adds the ability to use `rdpmc` hardware-based performance
counters instead of wall-clock time for measuring duration. This also
introduces a dependency on the `perf-event-open-sys` crate on Linux
which is used when using hardware counters.
Check the result cache before the DepGraph when ensuring queries
Split out of https://github.com/rust-lang/rust/pull/70951
Calling `ensure` on already forced queries is a common operation.
Looking at the results cache first is faster than checking the DepGraph for a green node.
Indicate change in RSS from start to end of pass in time-passes output
Previously, this was omitted because it could be misleading, but the
functionality seems too useful not to include.
r? ``@oli-obk``
This allows to directly map from a DefPathHash to the crate it
originates from, without constructing side tables to do that mapping.
It also allows to reliably and cheaply check for DefPathHash collisions.
Indicate both start and end of pass RSS in time-passes output
Previously, only the end of pass RSS was indicated. This could easily
lead one to believe that the change in RSS from one pass to the next was
attributable to the second pass, when in fact it occurred between the
end of the first pass and the start of the second.
Also, improve alignment of columns.
Sample of output:
```
time: 0.739; rss: 607MB -> 637MB item_types_checking
time: 8.429; rss: 637MB -> 775MB item_bodies_checking
time: 11.063; rss: 470MB -> 775MB type_check_crate
time: 0.232; rss: 775MB -> 777MB match_checking
time: 0.139; rss: 777MB -> 779MB liveness_and_intrinsic_checking
time: 0.372; rss: 775MB -> 779MB misc_checking_2
time: 8.188; rss: 779MB -> 1019MB MIR_borrow_checking
time: 0.062; rss: 1019MB -> 1021MB MIR_effect_checking
```
Previously, only the end of pass RSS was indicated. This could easily
lead one to believe that the change in RSS from one pass to the next was
attributable to the second pass, when in fact it occurred between the
end of the first pass and the start of the second.
Also, improve alignment of columns.
Enforce that query results implement Debug
Currently, we require that query keys implement `Debug`, but we do not do the same for query values. This can make incremental compilation bugs difficult to debug - there isn't a good place to print out the result loaded from disk.
This PR adds `Debug` bounds to several query-related functions, allowing us to debug-print the query value when an 'unstable fingerprint' error occurs. This required adding `#[derive(Debug)]` to a fairly large number of types - hopefully, this doesn't have much of an impact on compiler bootstrapping times.
Before:
```
thread 'rustc' panicked at 'attempt to read from stolen value', /home/joshua/rustc/compiler/rustc_data_structures/src/steal.rs:43:15
```
After:
```
thread 'rustc' panicked at 'attempt to steal from stolen value', compiler/rustc_mir/src/transform/mod.rs:423:25
```
Reduce a large memory spike that happens during serialization by writing
the incr comp structures to file by way of a fixed-size buffer, rather
than an unbounded vector.
Effort was made to keep the instruction count close to that of the
previous implementation. However, buffered writing to a file inherently
has more overhead than writing to a vector, because each write may
result in a handleable error. To reduce this overhead, arrangements are
made so that each LEB128-encoded integer can be written to the buffer
with only one capacity and error check. Higher-level optimizations in
which entire composite structures can be written with one capacity and
error check are possible, but would require much more work.
The performance is mostly on par with the previous implementation, with
small to moderate instruction count regressions. The memory reduction is
significant, however, so it seems like a worth-while trade-off.
Stabilize or_insert_with_key
Stabilizes the `or_insert_with_key` feature from https://github.com/rust-lang/rust/issues/71024. This allows inserting key-derived values when a `HashMap`/`BTreeMap` entry is vacant.
The difference between this and `.or_insert_with(|| ... )` is that this provides a reference to the key to the closure after it is moved with `.entry(key_being_moved)`, avoiding the need to copy or clone the key.
They were originally called "opt-in, built-in traits" (OIBITs), but
people realized that the name was too confusing and a mouthful, and so
they were renamed to just "auto traits". The feature flag's name wasn't
updated, though, so that's what this PR does.
There are some other spots in the compiler that still refer to OIBITs,
but I don't think changing those now is worth it since they are internal
and not particularly relevant to this PR.
Also see <https://rust-lang.zulipchat.com/#narrow/stream/131828-t-compiler/topic/opt-in.2C.20built-in.20traits.20(auto.20traits).20feature.20name>.
Reworks Sccc computation to iteration instead of recursion
Linear graphs, producing as many scc's as nodes, would recurse once for every node when entered from the start of the list. This adds a test that exhausted the stack at least on my machine with error:
```
thread 'graph::scc::tests::test_deep_linear' has overflowed its stack
fatal runtime error: stack overflow
```
This may or may not be connected to #78567. I was only reminded that I started this rework some time ago. It might be plausible as borrow checking a long function with many borrow regions around each other—((((((…))))))— may produce the linear list setup to trigger this stack overflow ? I don't know enough about borrow check to say for sure.
This is best read in two separate commits. The first addresses only `find_state` internally. This is classical union phase from union-find. There's also a common solution of using the parent pointers in the (virtual) linked list to track the backreferences while traversing upwards and then following them backwards in a second path compression phase.
The second is more involved as it rewrites the mutually recursive `walk_node` and `walk_unvisited_node`. Firstly, the caller is required to handle the unvisited case of `walk_node` so a new `start_walk_from` method is added to handle that by walking the unvisited node if necessary. Then `walk_unvisited_node`, where we would previously recurse into in the missing case, is rewritten to construct a manual stack of its frames. The state fields consist of the previous stack slots.
[self-profiling] Include the estimated size of each cgu in the profile
This is helpful when looking for CGUs where the size estimate isn't a
good indicator of compilation time.
I verified that moving the profiling timer call doesn't affect the
results.
Results:
<img width="297" alt="Screen Shot 2020-11-03 at 7 25 04 AM" src="https://user-images.githubusercontent.com/831192/97985503-5901d100-1da6-11eb-9f10-f3e399702952.png">
`measureme` doesn't have support for custom arg names yet so `arg0` is the CGU name and `arg1` is the estimated size.
Move likely/unlikely argument outside of invisible unsafe block
The previous `likely!`/`unlikely!` macros were unsound because it permits the caller's expr to contain arbitrary unsafe code.
```rust
pub fn huh() -> bool {
likely!(std::ptr::read(&() as *const () as *const bool))
}
```
**Before:** compiles cleanly.
**After:**
```console
error[E0133]: call to unsafe function is unsafe and requires unsafe function or block
|
70 | likely!(std::ptr::read(&() as *const () as *const bool))
| ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ call to unsafe function
|
= note: consult the function's documentation for information on how to avoid undefined behavior
```
The previous `likely!`/`unlikely!` macros were unsound because it
permits the caller's expr to contain arbitrary unsafe code.
pub fn huh() -> bool {
likely!(std::ptr::read(&() as *const () as *const bool))
}
Before: compiles cleanly.
After:
error[E0133]: call to unsafe function is unsafe and requires unsafe function or block
|
70 | likely!(std::ptr::read(&() as *const () as *const bool))
| ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ call to unsafe function
|
= note: consult the function's documentation for information on how to avoid undefined behavior
This allows constructing the sccc for large that visit many nodes before
finding a single cycle of sccc, for example lists. When used to find
dependencies in borrow checking the list case is what occurs in very
long functions.
The basic conversion is a straightforward conversion of the linear
recursion to a loop forwards and backwards propagation of the result.
But this uses an optimization to avoid the need for extra space that
would otherwise be necessary to store the stack of unfinished states as
the function is not tail recursive.
Observe that only non-root-nodes in cycles have a recursive call and
that every such call overwrites their own node state. Thus we reuse the
node state itself as temporary storage for the stack of unfinished
states by inverting the links to a chain back to the previous state
update. When we hit the root or end of the full explored chain we
propagate the node state update backwards by following the chain until
a node with a link to itself.
This is helpful when looking for CGUs where the size estimate isn't a
good indicator of compilation time.
I verified that moving the profiling timer call doesn't affect the
results.
The previous recursive approach might overflow the stack when walking a
particularly deep, list-like, graph. In particular, dominator
calculation for borrow checking does such a traversal and very long
functions might lead to a region dependency graph with in this
problematic structure.
