Modify MIR building to drop repeat expressions with length zero
Closes#74836 .
Previously, when a user wrote `[foo; 0]` we used to simply leak `foo`. The goal is to fix that. This PR changes MIR building to make `[foo; 0]` equivalent to `{ drop(foo); [] }` in all cases. Of course, this is a breaking change (see below). A crater run did not indicate any regressions though, and given that the previous behavior was almost definitely not what any user wanted, it seems unlikely that anyone was relying on this.
Note that const generics are in general unaffected by this. Inserting the extra `drop` is only meaningful/necessary when `foo` is of a non-`Copy` type, and array repeat expressions with const generic repetition count must always be `Copy`.
Besides the obvious change to behavior associated with the additional drop, there are three categories of examples where this also changes observable behavior. In all of these cases, the new behavior is consistent with what you would get by replacing `[foo; 0]` with `{ drop(foo); [] }`. As such, none of these give the user new powers to express more things.
**No longer allowed in const (breaking)**:
```rust
const _: [String; 0] = [String::new(); 0];
```
This compiles on stable today. Because we now introduce the drop of `String`, this no longer compiles as `String` may not be dropped in a const context.
**Reduced dataflow (non-breaking)**:
```rust
let mut x: i32 = 0;
let r = &x;
let a = [r; 0];
x = 5;
let _b = a;
```
Borrowck rejects this code on stable because it believes there is dataflow between `a` and `r`, and so the lifetime of `r` has to extend to the last statement. This change removes the dataflow and the above code is allowed to compile.
**More const promotion (non-breaking)**:
```rust
let _v: &'static [String; 0] = &[String::new(); 0];
```
This does not compile today because `String` having drop glue keeps it from being const promoted (despite that drop glue never being executed). After this change, this is allowed to compile.
### Alternatives
A previous attempt at this tried to reduce breakage by various tricks. This is still a possibility, but given that crater showed no regressions it seems unclear why we would want to introduce this complexity.
Disallowing `[foo; 0]` completely is also an option, but obviously this is more of a breaking change. I do not know how often this is actually used though.
r? `@oli-obk`
correctly deal with user type ascriptions in pat
supersedes #93856
`thir::PatKind::AscribeUserType` previously resulted in `CanonicalUserTypeAnnotations` where the inferred type already had a subtyping relation according to `variance` to the `user_ty`.
The bug can pretty much be summarized as follows:
- during mir building
- `user_ty -> inferred_ty`: considers variance
- `StatementKind::AscribeUserType`: `inferred_ty` is the type of the place, so no variance needed
- during mir borrowck
- `user_ty -> inferred_ty`: does not consider variance
- `StatementKind::AscribeUserType`: applies variance
This mostly worked fine. The lifetimes in `inferred_ty` were only bound by its relation to `user_ty` and to the `place` of `StatementKind::AscribeUserType`, so it doesn't matter where exactly the subtyping happens.
It does however matter when having higher ranked subtying. At this point the place where the subtyping happens is forced, causing this mismatch between building and borrowck to result in unintended errors.
cc #96514 which is pretty much the same issue
r? `@nikomatsakis`
Add EarlyBinder
Chalk has no concept of `Param` (e0ade19d13/chalk-ir/src/lib.rs (L579)) or `ReEarlyBound` (e0ade19d13/chalk-ir/src/lib.rs (L1308)). Everything is just "bound" - the equivalent of rustc's late-bound. It's not completely clear yet whether to move everything to the same time of binder in rustc or add `Param` and `ReEarlyBound` in Chalk.
Either way, tracking when we have or haven't already substituted out these in rustc can be helpful.
As a first step, I'm just adding a `EarlyBinder` newtype that is required to call `subst`. I also add a couple "transparent" `bound_*` wrappers around a couple query that are often immediately substituted.
r? `@nikomatsakis`
don't encode only locally used attrs
Part of https://github.com/rust-lang/compiler-team/issues/505.
We now filter builtin attributes before encoding them in the crate metadata in case they should only be used in the local crate. To prevent accidental misuse `get_attrs` now requires the caller to state which attribute they are interested in. For places where that isn't trivially possible, I've added a method `fn get_attrs_unchecked` which I intend to remove in a followup PR.
After this pull request landed, we can then slowly move all attributes to only be used in the local crate while being certain that we don't accidentally try to access them from extern crates.
cc https://github.com/rust-lang/rust/pull/94963#issuecomment-1082924289
Use `FxIndexSet` to avoid sorting fake borrows
This fixes#96449, but I haven't yet been able to
make the reproducer work using `#[cfg]` attributes,
so we can't use the 'revision' infra to write a test
The previous implementation relied on sorting by `PlaceRef`.
This requires sorting by a `DefId`, which uses untracked state
(see #93315)
This fixes#96449, but I haven't yet been able to
make the reproducer work using `#[cfg]` attributes,
so we can't use the 'revision' infra to write a test
The previous implementation relied on sorting by `PlaceRef`.
This requires sorting by a `DefId`, which uses untracked state
(see #93315)
Begin fixing all the broken doctests in `compiler/`
Begins to fix#95994.
All of them pass now but 24 of them I've marked with `ignore HELP (<explanation>)` (asking for help) as I'm unsure how to get them to work / if we should leave them as they are.
There are also a few that I marked `ignore` that could maybe be made to work but seem less important.
Each `ignore` has a rough "reason" for ignoring after it parentheses, with
- `(pseudo-rust)` meaning "mostly rust-like but contains foreign syntax"
- `(illustrative)` a somewhat catchall for either a fragment of rust that doesn't stand on its own (like a lone type), or abbreviated rust with ellipses and undeclared types that would get too cluttered if made compile-worthy.
- `(not-rust)` stuff that isn't rust but benefits from the syntax highlighting, like MIR.
- `(internal)` uses `rustc_*` code which would be difficult to make work with the testing setup.
Those reason notes are a bit inconsistently applied and messy though. If that's important I can go through them again and try a more principled approach. When I run `rg '```ignore \(' .` on the repo, there look to be lots of different conventions other people have used for this sort of thing. I could try unifying them all if that would be helpful.
I'm not sure if there was a better existing way to do this but I wrote my own script to help me run all the doctests and wade through the output. If that would be useful to anyone else, I put it here: https://github.com/Elliot-Roberts/rust_doctest_fixing_tool
Only crate root def-ids don't have a parent, and in majority of cases the argument of `DefIdTree::parent` cannot be a crate root.
So we now panic by default in `parent` and introduce a new non-panicing function `opt_parent` for cases where the argument can be a crate root.
Same applies to `local_parent`/`opt_local_parent`.
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.
Implement sym operands for global_asm!
Tracking issue: #93333
This PR is pretty much a complete rewrite of `sym` operand support for inline assembly so that the same implementation can be shared by `asm!` and `global_asm!`. The main changes are:
- At the AST level, `sym` is represented as a special `InlineAsmSym` AST node containing a path instead of an `Expr`.
- At the HIR level, `sym` is split into `SymStatic` and `SymFn` depending on whether the path resolves to a static during AST lowering (defaults to `SynFn` if `get_early_res` fails).
- `SymFn` is just an `AnonConst`. It runs through typeck and we just collect the resulting type at the end. An error is emitted if the type is not a `FnDef`.
- `SymStatic` directly holds a path and the `DefId` of the `static` that it is pointing to.
- The representation at the MIR level is mostly unchanged. There is a minor change to THIR where `SymFn` is a constant instead of an expression.
- At the codegen level we need to apply the target's symbol mangling to the result of `tcx.symbol_name()` depending on the target. This is done by calling the LLVM name mangler, which handles all of the details.
- On Mach-O, all symbols have a leading underscore.
- On x86 Windows, different mangling is used for cdecl, stdcall, fastcall and vectorcall.
- No mangling is needed on other platforms.
r? `@nagisa`
cc `@eddyb`
Use mir constant in thir instead of ty::Const
This is blocked on https://github.com/rust-lang/rust/pull/94059 (does include its changes, the first two commits in this PR correspond to those changes) and https://github.com/rust-lang/rust/pull/93800 being reinstated (which had to be reverted). Mainly opening since `@lcnr` offered to give some feedback and maybe also for a perf-run (if necessary).
This currently contains a lot of duplication since some of the logic of `ty::Const` had to be copied to `mir::ConstantKind`, but with the introduction of valtrees a lot of that functionality will disappear from `ty::Const`.
Only the last commit contains changes that need to be reviewed here. Did leave some `FIXME` comments regarding future implementation decisions and some things that might be incorrectly implemented.
r? `@oli-obk`
`MultiSpan` contains labels, which are more complicated with the
introduction of diagnostic translation and will use types from
`rustc_errors` - however, `rustc_errors` depends on `rustc_span` so
`rustc_span` cannot use types like `DiagnosticMessage` without
dependency cycles. Introduce a new `rustc_error_messages` crate that can
contain `DiagnosticMessage` and `MultiSpan`.
Signed-off-by: David Wood <david.wood@huawei.com>
