Don't run const propagation on items with inconsistent bounds
Fixes#67696
Using `#![feature(trivial_bounds)]`, it's possible to write functions
with unsatisfiable 'where' clauses, making them uncallable. However, the
user can act as if these 'where' clauses are true inside the body of the
function, leading to code that would normally be impossible to write.
Since const propgation can run even without any user-written calls to a
function, we need to explcitly check for these uncallable functions.
Add unreachable propagation mir optimization pass
@oli-obk suggested we create a MIR pass that optimizes away basic blocks that lead only to basic blocks with terminator kind **unreachable**. This is a first take on this, which we started with @gilescope at RustFest Impl Days.
The test currently fails when the compiled program runs (undefined behaviour). Is there a way to avoid running the compiled program?
perf: Eagerly convert literals to consts
Previousely even literal constants were being converted to an `Unevaluted` constant for evaluation later. This seems unecessary as no more information is needed to be able to convert the literal to a mir constant.
Hopefully this will also minimise the performance impact of #67717, as far less constant evaluations are needed.
Fix memory leak if C++ catches a Rust panic and discards it
If C++ catches a Rust panic using `catch (...)` and then chooses not to rethrow it, the `Box<dyn Any>` in the exception may be leaked. This PR fixes this by adding the necessary destructors to the exception object.
r? @Mark-Simulacrum
Using `#![feature(trivial_bounds)]`, it's possible to write functions
with unsatisfiable 'where' clauses, making them uncallable. However, the
user can act as if these 'where' clauses are true inside the body of the
function, leading to code that would normally be impossible to write.
Since const propgation can run even without any user-written calls to a
function, we need to explcitly check for these uncallable functions.
Prepare for LLVM 10 upgrade
Split off from #67759, this just adds the necessary compatibility bits and updates codegen tests, without performing the actual LLVM upgrade.
r? @alexcrichton
Split MIR building into its own crate
This moves `rustc_mir::{build, hair, lints}` to `rustc_mir_build`.
The new crate only has a `provide` function as it's public API.
Based on #67898
cc @Centril @rust-lang/compiler
r? @oli-obk
Add suggestions when encountering chained comparisons
Ideally, we'd also prevent the type error, which is just extra noise, but that will require moving the error from the parser, and I think the suggestion makes things clear enough for now.
Fixes https://github.com/rust-lang/rust/issues/65659.
Move more of `rustc::lint` into `rustc_lint`
Based on https://github.com/rust-lang/rust/pull/67806.
Here we try to consolidate more of the linting infra into `rustc::lint`. Some high-level notes:
- We now store an `Lrc<dyn Any + Send + Sync>` as opposed to `Lrc<LintStore>` in the `GlobalCtxt`. This enables us to avoid referring to the type, breaking a cyclic dependency, and so we can move things from `rustc::lint` to `rustc_lint`.
- `in_derive_expansion` is, and needs to, be moved as a method on `Span`.
- We reduce the number of ways on `tcx` to emit a lint so that the developer UX is more streamlined.
- `LintLevelsBuilder` is moved to `rustc_lint::levels`, leaving behind `LintLevelMap/Set` in a purified form due to current constraints (hopefully fixable in the future after https://github.com/rust-lang/rust/pull/68133).
- `struct_lint_level` is moved to `rustc::lint` due to current dependency constraints.
- `rustc::lint::context` is moved to `rustc_lint::context`.
- The visitors in `rustc::lint` are moved to `rustc_lint::passes`.
Don't require `allow_internal_unstable` unless `staged_api` is enabled.
#63770 changed `qualify_min_const_fn` to require `allow_internal_unstable` for *all* crates that used an unstable feature, regardless of whether `staged_api` was enabled or the `fn` that used that feature was stably const. In practice, this meant that every crate in the ecosystem that wanted to use nightly features added `#![feature(const_fn)]`, which skips `qualify_min_const_fn` entirely.
After this PR, crates that do not have `#![feature(staged_api)]` will only need to enable the feature they are interested in. For example, `#![feature(const_if_match)]` will be enough to enable `if` and `match` in constants. Crates with `staged_api` (e.g., `libstd`) require `#[allow_internal_unstable]` to be added to a function if it uses nightly features unless that function is also marked `#[rustc_const_unstable]`. This prevents proliferation of `#[allow_internal_unstable]` into functions that are not callable in a `const` context on stable.
r? @oli-obk (author of #63770)
cc @Centril
Extract `rustc_ast_passes`, move gating, & refactor linting
Based on https://github.com/rust-lang/rust/pull/67770.
This PR extracts a crate `rustc_ast_passes`:
- `ast_validation.rs`, which is contributed by `rustc_passes` (now only has HIR based passes) -- the goal here is to improve recompilation of the parser,
- `feature_gate.rs`, which is contributed by `syntax` and performs post-expansion-gating & final erroring for pre-expansion gating,
- `show_span`, which is contributed by `syntax`.
To facilitate this, we first have to also:
- Move `{leveled_}feature_err{_err}` from `syntax::feature_gate::check` into `rustc_session::parse`.
- Move `get_features` into `rustc_parse::config`, the only place it is used.
- Move some some lint datatypes and traits, e.g. `LintBuffer`, `BufferedEarlyLint`, `BuiltinLintDiagnostics`, `LintPass`, and `LintArray` into `rustc_session::lint`.
- Move all the hard-wired lint `static`s into `rustc_session::lint::builtin`.
This flag opts out of the min-const-fn checks entirely, which is usually
not what we want. The few cases where the flag is still necessary have
been annotated.
build-std compatible sanitizer support
### Motivation
When using `-Z sanitizer=*` feature it is essential that both user code and
standard library is instrumented. Otherwise the utility of sanitizer will be
limited, or its use will be impractical like in the case of memory sanitizer.
The recently introduced cargo feature build-std makes it possible to rebuild
standard library with arbitrary rustc flags. Unfortunately, those changes alone
do not make it easy to rebuild standard library with sanitizers, since runtimes
are dependencies of std that have to be build in specific environment,
generally not available outside rustbuild process. Additionally rebuilding them
requires presence of llvm-config and compiler-rt sources.
The goal of changes proposed here is to make it possible to avoid rebuilding
sanitizer runtimes when rebuilding the std, thus making it possible to
instrument standard library for use with sanitizer with simple, although
verbose command:
```
env CARGO_TARGET_X86_64_UNKNOWN_LINUX_GNU_RUSTFLAGS=-Zsanitizer=thread cargo test -Zbuild-std --target x86_64-unknown-linux-gnu
```
### Implementation
* Sanitizer runtimes are no long packed into crates. Instead, libraries build
from compiler-rt are used as is, after renaming them into `librusc_rt.*`.
* rustc obtains runtimes from target libdir for default sysroot, so that
they are not required in custom build sysroots created with build-std.
* The runtimes are only linked-in into executables to address issue #64629.
(in previous design it was hard to avoid linking runtimes into static
libraries produced by rustc as demonstrated by sanitizer-staticlib-link
test, which still passes despite changes made in #64780).
cc @kennytm, @japaric, @firstyear, @choller
