GVN: stop hashing opaque values
GVN generates values that are not meant to be unified with any other. For instance `Opaque` (aka we don't know anything), non-deterministic constants and borrows.
The current algorithm generates a unique index, so the generated `Value` will be different from all the existing. This is wasteful, as we should not hash that `Value` at all.
This PR proposes to do this. This involves partially reimplementing a `FxIndexSet`, but yields a small but consistent perf improvement (https://github.com/rust-lang/rust/pull/145737#issuecomment-3276951054).
Clean up `ty::Dynamic`
1. As a follow-up to PR rust-lang/rust#143036, remove `DynKind` entirely.
2. Inside HIR ty lowering, consolidate modules `dyn_compatibility` and `lint` into `dyn_trait`
* `dyn_compatibility` wasn't about dyn compatibility itself, it's about lowering trait object types
* `lint` contained dyn-Trait-specific diagnostics+lints only
Enable DestinationPropagation by default
This PR proposes to perform destination propagation on MIR. Most of the pass was fully rewritten by `@JakobDegen` in rust-lang/rust#96451.
This pass is quite heavy, as it needs to perform and save the results of a full liveness dataflow analysis. This accounts for ~50% of the pass' runtime.
Perf sees a few decent savings in later llvm passes, but also sizeable régressions when there are no savings to balance this pass' runtime.
Remove Rvalue::Len again.
Now that we have `RawPtrKind::FakeForPtrMetadata`, we can reimplement `Rvalue::Len` using `PtrMetadata(&raw const (fake) place)`.
r? ``@scottmcm``
rename erase_regions to erase_and_anonymize_regions
I find it consistently confusing that `erase_regions` does more than replacing regions with `'erased`. it also makes some code look real goofy to be writing manual folders to erase regions with a comment saying "we cant use erase regions" :> or code that re-calls erase_regions on types with regions already erased just to anonymize all the bound regions.
r? lcnr
idk how i feel about the name being almost twice as long now
GVN: Ensure indirect is first projection in try_as_place.
I haven't found any report for this bug on existing code, but managed to trigger it with rust-lang/rust#143333
Ignore intrinsic calls in cross-crate-inlining cost model
I noticed in a side project that a function which just compares to `[u64; 2]` for equality is not cross-crate-inlinable. That was surprising to me because I didn't think that code contained a function call, but of course our array comparisons are lowered to an intrinsic. Intrinsic calls don't make a function no longer a leaf, so it makes sense to add this as an exception to the "only leaves" cross-crate-inline heuristic.
This is the useful compare link: https://perf.rust-lang.org/compare.html?start=7cb1a81145a739c4fd858abe3c624ce8e6e5f9cd&end=c3f0a64dbf9fba4722dacf8e39d2fe00069c995e&stat=instructions%3Au because it disables CGU merging in both commits, so effects that cause changes in the sysroot to perturb partitioning downstream are excluded. Perturbations to what is and isn't cross-crate-inlinable in the sysroot has chaotic effects on what items are in which CGUs after merging. It looks like before this PR by sheer luck some of the CGUs dirtied by the patch in eza incr-unchanged happened to be merged together, and with this PR they are not.
The perf runs on this PR point to a nice runtime performance improvement.
Reimplement DestinationPropagation according to live ranges.
This PR reimplements DestinationPropagation as a problem of merging live-ranges of locals. We merge locals that have disjoint live-ranges. This allows merging several locals in the same round by updating live range information.
Live ranges are mainly computed using the `MaybeLiveLocals` analysis. The subtlety is that we split each statement and terminator in 2 positions. The first position is the regular statement. The second position is a shadow, which is always more live. It encodes partial writes and dead writes as a local being live for half a statement. This half statement ensures that writes conflict with another local's writes and regular liveness.
r? `@Amanieu`
Use `Itertools::all_equal_value()` where applicable
Just a small cleanup.
We already have `itertools` as a dep in these crates, so might as well use another of its features.
Makes the code simpler IMHO :)
Allow `inline(always)` with a target feature behind a unstable feature `target_feature_inline_always`.
Rather than adding the inline always attribute to the function definition, we add it to the callsite. We can then check that the target features match and that the call would be safe to inline. If the function isn't inlined due to a mismatch, we emit a warning informing the user that the function can't be inlined due to the target feature mismatch.
See tracking issue rust-lang/rust#145574
This was done in #145740 and #145947. It is causing problems for people
using r-a on anything that uses the rustc-dev rustup package, e.g. Miri,
clippy.
This repository has lots of submodules and subtrees and various
different projects are carved out of pieces of it. It seems like
`[workspace.dependencies]` will just be more trouble than it's worth.
Rollup of 5 pull requests
Successful merges:
- rust-lang/rust#145468 (dedup recip, powi, to_degrees, and to_radians float tests)
- rust-lang/rust#145643 (coverage: Build an "expansion tree" and use it to unexpand raw spans)
- rust-lang/rust#145754 (fix(lexer): Don't require frontmatters to be escaped with indented fences)
- rust-lang/rust#146060 (fixup nix dev shell again)
- rust-lang/rust#146068 (compiletest: Capture panic messages via a custom panic hook)
r? `@ghost`
`@rustbot` modify labels: rollup
coverage: Build an "expansion tree" and use it to unexpand raw spans
Historically and currently, coverage instrumentation assumes that all of a function's spans are in the same file and have the same syntax context. The spans extracted directly from MIR don't satisfy that assumption, so there is an “unexpansion” step that walks up each span's expansion-call-site tree to find a suitable span in the same context as the function's body span.
(That unexpansion step is what allows us to have somewhat reasonable coverage instrumentation for macros like `println!`, and for syntax like `for` and `?` that undergo desugaring expansion.)
The current unexpansion code mostly works fine in that “flat” single-file single-context world. But it's not suitable for incremental work towards proper expansion-aware coverage instrumentation, which would allow a function's coverage spans to encompass multiple expansion contexts and multiple files.
This PR therefore replaces the current unexpansion code with a more sophisticated system that uses the raw MIR spans to reconstruct an “expansion tree”, and then uses that tree to help perform most of the unexpansion work.
Building the tree is “overkill” for current unexpansion needs (though it does give some minor edge-case improvements), but my hope is that having the explicit tree available will be a big help when taking the next steps towards proper expansion-region support.
MIR dumping is a mess. There are lots of functions and entry points,
e.g. `dump_mir`, `dump_mir_with_options`, `dump_polonius_mir`,
`dump_mir_to_writer`. Also, it's crucial that `create_dump_file` is
never called without `dump_enabled` first being checked, but there is no
mechanism for ensuring this and it's hard to tell if it is satisfied on
all paths. (`dump_enabled` is checked twice on some paths, however!)
This commit introduces `MirWriter`, which controls the MIR writing, and
encapsulates the `extra_data` closure and `options`. Two existing
functions are now methods of this type. It sets reasonable defaults,
allowing the removal of many `|_, _| Ok(())` closures.
The commit also introduces `MirDumper`, which is layered on top of
`MirWriter`, and which manages the creation of the dump files,
encapsulating pass names, disambiguators, etc. Four existing functions
are now methods of this type.
- `MirDumper::new` will only succeed if dumps are enabled, and will
return `None` otherwise, which makes it impossible to dump when you
shouldn't.
- It also sets reasonable defaults for various things like
disambiguators, which means you no longer need to specify them in many
cases. When they do need to be specified, it's now done via setter
methods.
- It avoids some repetition. E.g. `dump_nll_mir` previously specifed the
pass name `"nll"` four times and the disambiguator `&0` three times;
now it specifies them just once, to put them in the `MirDumper`.
- For Polonius, the `extra_data` closure can now be specified earlier,
which avoids having to pass some arguments through some functions.
The dynamic dispatch cost doesn't matter for MIR dumping, which is
perf-insensitive. And it's necessary for the next commit, which will
store some `extra_data` closures in a struct.
