delay bug when adjusting `NeverToAny` twice during diagnostic code
Addresses Issue 96335 (https://github.com/rust-lang/rust/issues/96335) by using `delay_span_bug` instead of an assert and returning an error type from `check_expr_meets_expectation_or_error`.
Fixes#96335
Better handle too many `#` recovery in raw str
Point at all the unnecessary trailing `#`.
Better handle interaction with outer attributes when `;` is missing.
Fix#95030.
Display function path in unsafety violations - E0133
adds `DefId` to `UnsafetyViolationDetails`
this enables consumers to access the function definition that was reported to be unsafe and also changes the output for some E0133 diagnostics
Generate synthetic object file to ensure all exported and used symbols participate in the linking
Fix#50007 and #47384
This is the synthetic object file approach that I described in https://github.com/rust-lang/rust/pull/95363#issuecomment-1079932354, allowing all exported and used symbols to be linked while still allowing them to be GCed.
Related #93791, #95363
r? `@petrochenkov`
cc `@carbotaniuman`
Add `BoundKind` in `visit_param_bounds` to check questions in bounds
From the FIXME in the impl of `AstValidator`. Better bound checks by adding `BoundCtxt` type parameter to `visit_param_bound`
cc `@ecstatic-morse`
Drop support for legacy PM with LLVM 15
LLVM 15 already removes some of the legacy PM APIs we're using. This patch forces use of NewPM with LLVM 15 (with `-Z new-llvm-pass-manager=no` throwing a warning) and stubs out various FFI methods with a report_fatal_error on LLVM 15.
For LLVMPassManagerBuilderPopulateLTOPassManager() I went with adding our own wrapper, as the alternative would be to muck about with weak symbols, which seems to be non-trivial as far as cross-platform support is concerned (std has `weak!` for this purpose, but only as an internal utility.)
Fixes#96072.
Fixes#96362.
Implement MIR opt unit tests
This implements rust-lang/compiler-team#502 .
There's not much to say here, this implementation does everything as proposed. I also added the flag to a bunch of existing tests (mostly those to which I could add it without causing huge diffs due to changes in line numbers). Summarizing the changes to test outputs:
- Every time an `MirPatch` is created, it adds a cleanup block to the body if it did not exist already. If this block is unused (as is usually the case), it usually gets removed soon after by some pass calling `SimplifyCFG` for unrelated reasons (in many cases this cycle happens quite a few times for a single body). We now run `SimplifyCFG` less often, so those blocks end up in some of our outputs. I looked at changing `MirPatch` to not do this, but that seemed too complicated for this PR. I may still do that in a follow-up.
- The `InstCombine` test had set `-C opt-level=0` in its flags and so there were no storage markers. I don't really see a good motivation for doing this, so bringing it back in line with what everything else does seems correct.
- One of the `EarlyOtherwiseBranch` tests had `UnreachableProp` running on it. Preventing that kind of thing is the goal of this feature, so this seems fine.
For the remaining tests for which this feature might be useful, we can gradually migrate them as opportunities present themselves.
In terms of documentation, I plan on submitting a PR to the rustc dev guide in the near future documenting this and other recent changes to MIR. If there's any other places to update, do let me know
r? `@nagisa`
Improve span for `consider adding an explicit lifetime bound` suggestions under NLL
Because NLL borrowck is run after typeck, `in_progress_typeck_results` was always `None` which was preventing the retrieval of the span to which the suggestion is suppose to add the lifetime bound.
We now manually pass the `LocalDefId` owner to `construct_generic_bound_failure` so that under NLL, we give the owner id of the current body.
This helps with #96332
make `classify_drop_access_kind` iterate
This PR:
1. fixes the FIXME of `classify_drop_access_kind` func in the borrowck part. The process of obtaining `StorageDeadOrDrop` has been changed from recursive to iterative.
2. gets `place_ty` in each iteration, avoid repeatedly getting the `ty` of the same place (O(n^2) => O(n))
Because NLL borrowck is run after typeck, `in_progress_typeck_results`
was always `None` which was preventing the retrieval of the span to which
the suggestion is suppose to add the lifetime bound.
We now manually pass the `LocalDefId` owner to `construct_generic_bound_failure`
so that under NLL, we give the owner id of the current body.
This adds `nounused` to the set of extern flags:
`--extern nounused:core=/path/to/core/libcore.rlib`.
The effect of this flag is to suppress `unused-crate-dependencies`
warnings relating to the crate.
Optimize `const_prop` mir-opt by accessing `local_decls` through `ecx`
From the FIXME in the impl of `ConstPropagator`. Accessing `local_decls` and `scource_scopes` from `ecx` can reduce `clone` calls and save compile time.
Besides, according to #96213 , the FIXME about writing `layouts` to `ecx` in advance can also be removed.
Remove visibility information from HIR
The resolver exports all the necessary visibility information through the `tcx.visibility` query.
This PR stops having a dedicated visibility field in HIR, in order to use this query.
We keep a `vis_span` field for diagnostic purposes.
Sync rustc_codegen_cranelift
Mostly fixing bugs this time, but also a Cranelift update.
r? `@ghost`
`@rustbot` label +A-codegen +A-cranelift +T-compiler
debuginfo: Emit ZST struct debuginfo for unit type when CPP-like debuginfo is enabled
As already discovered in 24a728a8eb, PDB does not play well with custom basic types. This PR extends to the fix to `()`: Instead of a custom basic type, we treat it like an empty tuple (i.e. it is described as a struct which happens to have no fields).
Before this change anything with a `()` in it would cause trouble, which is especially bad for `*const ()` and `*mut ()` which are often used for opaque pointers. E.g. the test case added in this PR would look like:
```
0:000> dx _ref
Error: Unable to bind name '_ref'
0:000> dx _ptr
Error: Unable to bind name '_ptr'
0:000> dx _local
Error: Unable to bind name '_local'
0:000> dx _field,d
_field,d [Type: unit_type::_TypeContainingUnitField]
[+0x008] _a : 123 [Type: unsigned int]
[+0x000] _unit : Unexpected failure to dereference object
[+0x000] _b : 456 [Type: unsigned __int64]
0:000> dx ((__int64 *)_ptr),x
Error: Unable to bind name '_ptr'
```
With the PR it produces the expected output:
```
0:000> dx _ref
_ref : 0x7ff6f2012230 : () [Type: tuple$<> *]
0:000> dx _ptr
_ptr : 0x7e8ddffc20 : () [Type: tuple$<> *]
0:000> dx _local
_local : () [Type: tuple$<>]
0:000> dx _field,d
_field,d [Type: unit_type::_TypeContainingUnitField]
[+0x008] _a : 123 [Type: unsigned int]
[+0x000] _unit : () [Type: tuple$<>]
[+0x000] _b : 456 [Type: unsigned __int64]
0:000> dx ((__int64 *)_ptr),x
((__int64 *)_ptr),x : 0x7e8ddffc20 : 0x1122334455667788 [Type: __int64 *]
```
r? `@wesleywiser`
self-profiler: record spans for proc-macro expansions
This PR is a follow-up to #95473, using the arg recorder feature from #95689:
- it adds support code to easily record spans in the event's arguments, when using `generic_activity_with_arg_recorder`.
- uses that to record the spans where proc-macro expansions happen in addition to their name.
As for the other 2 PRs, the goal here is to provide visibility into proc-macro expansion performance, so that users can diagnose which uses of proc-macros in their code could be causing compile time issues.
Some areas where I'd love feedback:
- [x] the API and names: the `SpannedEventArgRecorder` trait and its method, much like #95689 had the same question about the `EventArgRecorder` naming
- [x] we don't currently have a way to record the names of the event arguments, so should `record_arg_spanned` record the span as "location: {}" or similar ?
Make the lifetime accurate which is used in the region constraints part
This PR fixes the FIXME about lifetime using in the region constraints part.
We cannot write `<'graph, 'tcx, D>` because the definition of `Successors<'0, '1, D>` requires `'1 : '0`.
We cannot add bound to `'graph` either because `'graph` is required to be an arbitrary value in the definition of `WithSuccessors`
So the most accurate way is to use `<'s, 'tcx, D>`.
cc `@Aaron1011` who added this FIXME in #85343
The self-profiler's `EventArgRecorder` is general-purpose in its ability to record Strings (and `rustc_span` depends on the crate its defined in, `rustc_data_structure`).
Some generic activities could use recording locations where they happen in the user's code: to allow e.g. to track macro expansions and diagnose performance issues there.
This adds a `SpannedEventArgRecorder` that can record an argument given as a span, rather than a String, since turning spans into Strings can be tricky if you're not happy with its default Debug output. This way the recorder can have a `record_arg_spanned` method which will do that.
Make all thir types implement clone
This PR adds `Clone` impl to all of the `Thir<'tcx>` types.
I would like to be able to clone a `Thir` body so that I can make a copy in my rustc driver without breaking further compilation. Without this my driver is forced to run in the `after_expansion` callback and thus doesn't benefit from running all the safety checks that `rustc` usually does, instead i need to do them all myself.
errors: minor translation-related changes
- For one error in typeck, specifying "suggestion" as the attribute for the only suggestion is unnecessary, it's the default of the derive.
- The documentation comment for the `SessionDiagnostic` derive is out-of-date, it should have been updated in #95512.
r? `@oli-obk`
Only output DepKind in dump-dep-graph.
When printing the whole DepNode, the output file is simply too massive to
be actually useful for profiling.
This trimmed down version mixes a lot of information together, but it also
allows to ask questions such that "why does this query ever access HIR?".
bootstrap: add split-debuginfo config
Replace `run-dysutil` option with more general `split-debuginfo` option that works on all platforms.
r? `@Mark-Simulacrum`
This lets us clone just the parts within a `TokenTree` that need
cloning, rather than the entire thing. This is a surprisingly large
performance win, up to 4% on `async-std-1.10.0`.
Specifying "suggestion" as the attribute for the only suggestion is
unnecessary, it's the default of the derive.
Signed-off-by: David Wood <david.wood@huawei.com>
This makes `CloseDelim` handling more like `OpenDelim` handling, which
produces `OpenDelim` and pushes the stack at the same time. It requires
some adjustment to `parse_token_tree` now that we don't remain within
the frame after getting the `CloseDelim`.
Add an explicit `Span` field to `OutlivesConstraint`
Previously, we would retrieve the span from the `Body` using
the `locations` field. However, we may end up changing the
`locations` field when moving a constraint from a promoted
to a different body.
We now store the original `Span` in a dedication field, so that
changes to the `locations` do not affect the quality of our
diagnostics.
Check if call return type is visibly uninhabited when building MIR
The main motivation behind the change is to expose information about diverging
calls to the generator transform and match the precision of drop range tracking
which already understands that call expressions with visibly uninhabited types
diverges.
This change should also accept strictly more programs than before. That is
programs that were previously rejected due to errors raised by control-flow
sensitive checks in a code that is no longer considered reachable.
Fixes#93161.
A Google search of the error message fails to return any relevant
resuts, suggesting this has never occurred in practice. And removeing it
reduces instruction counts by up to 2% on some benchmarks.
Previously, we would retrieve the span from the `Body` using
the `locations` field. However, we may end up changing the
`locations` field when moving a constraint from a promoted
to a different body.
We now store the original `Span` in a dedication field, so that
changes to the `locations` do not affect the quality of our
diagnostics.
The loop is there to handle a `NoDelim` open/close token. This commit
changes `TokenCursor::inlined_next` so it never returns such a token.
This is a performance win because the conditional test in `bump()` is
removed.
If the parser needs changing in the future to handle `NoDelim` tokens,
then `inlined_next()` can easily be changed to return them.
The `DelimToken` here is `NoDelim`, which means the returned delim
tokens will just be ignored by `Parser::bump()`. This commit changes
things so the delim tokens won't be returned.
