Clarify which kinds of MIR are allowed during which phases.
This enhances documentation with these details and extends the validator to check these requirements more thoroughly. Most of these conditions were already being checked.
There was also some disagreement between the `MirPhase` docs and validator as to what it meant for the `body.phase` field to have a certain value. This PR resolves those disagreements in favor of the `MirPhase` docs (which is what the pass manager implemented), adjusting the validator accordingly. The result is now that the `DropLowering` phase begins with the end of the elaborate drops pass, and lasts until the beginning of the generator lowring pass. This doesn't feel entirely natural to me, but as long as it's documented accurately it should be ok.
r? rust-lang/mir-opt
This enhances documentation with these details and extends the validator to check these requirements
more thoroughly. As a part of this, we add a new `Deaggregated` phase, and rename other phases so
that their names more naturally correspond to what they represent.
rename LocalState::Uninitialized to Unallocated
This is to avoid confusion with `Uninit` as in `ScalarMaybeUninit`, which is very different.
r? `@oli-obk`
There are a few places were we have to construct it, though, and a few
places that are more invasive to change. To do this, we create a
constructor with a long obvious name.
Improve `AdtDef` interning.
This commit makes `AdtDef` use `Interned`. Much of the commit is tedious
changes to introduce getter functions. The interesting changes are in
`compiler/rustc_middle/src/ty/adt.rs`.
r? `@fee1-dead`
This commit makes `AdtDef` use `Interned`. Much the commit is tedious
changes to introduce getter functions. The interesting changes are in
`compiler/rustc_middle/src/ty/adt.rs`.
`Layout` is another type that is sometimes interned, sometimes not, and
we always use references to refer to it so we can't take any advantage
of the uniqueness properties for hashing or equality checks.
This commit renames `Layout` as `LayoutS`, and then introduces a new
`Layout` that is a newtype around an `Interned<LayoutS>`. It also
interns more layouts than before. Previously layouts within layouts
(via the `variants` field) were never interned, but now they are. Hence
the lifetime on the new `Layout` type.
Unlike other interned types, these ones are in `rustc_target` instead of
`rustc_middle`. This reflects the existing structure of the code, which
does layout-specific stuff in `rustc_target` while `TyAndLayout` is
generic over the `Ty`, allowing the type-specific stuff to occur in
`rustc_middle`.
The commit also adds a `HashStable` impl for `Interned`, which was
needed. It hashes the contents, unlike the `Hash` impl which hashes the
pointer.
Currently some `Allocation`s are interned, some are not, and it's very
hard to tell at a use point which is which.
This commit introduces `ConstAllocation` for the known-interned ones,
which makes the division much clearer. `ConstAllocation::inner()` is
used to get the underlying `Allocation`.
In some places it's natural to use an `Allocation`, in some it's natural
to use a `ConstAllocation`, and in some places there's no clear choice.
I've tried to make things look as nice as possible, while generally
favouring `ConstAllocation`, which is the type that embodies more
information. This does require quite a few calls to `inner()`.
The commit also tweaks how `PartialOrd` works for `Interned`. The
previous code was too clever by half, building on `T: Ord` to make the
code shorter. That caused problems with deriving `PartialOrd` and `Ord`
for `ConstAllocation`, so I changed it to build on `T: PartialOrd`,
which is slightly more verbose but much more standard and avoided the
problems.
Miri/CTFE: properly treat overflow in (signed) division/rem as UB
To my surprise, it looks like LLVM treats overflow of signed div/rem as UB. From what I can tell, MIR `Div`/`Rem` directly lowers to the corresponding LLVM operation, so to make that correct we also have to consider these overflows UB in the CTFE/Miri interpreter engine.
r? `@oli-obk`
Only create a single expansion for each inline integration.
The inlining integrator used to create one expansion for each span from the callee body.
This PR reverses the logic to create a single expansion for the whole call,
which is more consistent with how macro expansions work for macros.
This should remove the large memory regression in #91743.
Remove an unnecessary restriction in `dest_prop`
I had asked about this [on Zulip](https://rust-lang.zulipchat.com/#narrow/stream/131828-t-compiler/topic/Do.20unions.20have.20active.20fields.3F) but didn't receive a response, so putting up this PR that makes the change I think we can. If it turns out that this is wrong, hopefully I'll find out here. Reposting my Zulip comment:
> Not sure what channel to put this into, so using this as a fallback. The dest prop MIR opt has this comment:
>
> ```rust
> //! Subtle case: If `dest` is a, or projects through a union, then we have to make sure that there
> //! remains an assignment to it, since that sets the "active field" of the union. But if `src` is
> //! a ZST, it might not be initialized, so there might not be any use of it before the assignment,
> //! and performing the optimization would simply delete the assignment, leaving `dest`
> //! uninitialized.
> ```
>
> In particular, the claim seems to be that we can't take
> ```
> x = ();
> y.field = x;
> ```
> where `y` is a union having `field: ()` as one of its variants, and optimize the entire thing away (assuming `x` is unused otherwise). As far as I know though, Rust unions don't have active fields. Is this comment correct and am I missing something? Is there a worry about this interacting poorly with FFI code/C unions/LTO or something?
This PR just removes that comment and the associated code. Also it fixes one unrelated comment that did not match the code it was commenting on.
r? rust-lang/mir-opt
Convert `newtype_index` to a proc macro
The `macro_rules!` implementation was becomng excessively complicated,
and difficult to modify. The new proc macro implementation should make
it much easier to add new features (e.g. skipping certain `#[derive]`s)
rustc_errors: let `DiagnosticBuilder::emit` return a "guarantee of emission".
That is, `DiagnosticBuilder` is now generic over the return type of `.emit()`, so we'll now have:
* `DiagnosticBuilder<ErrorReported>` for error (incl. fatal/bug) diagnostics
* can only be created via a `const L: Level`-generic constructor, that limits allowed variants via a `where` clause, so not even `rustc_errors` can accidentally bypass this limitation
* asserts `diagnostic.is_error()` on emission, just in case the construction restriction was bypassed (e.g. by replacing the whole `Diagnostic` inside `DiagnosticBuilder`)
* `.emit()` returns `ErrorReported`, as a "proof" token that `.emit()` was called
(though note that this isn't a real guarantee until after completing the work on
#69426)
* `DiagnosticBuilder<()>` for everything else (warnings, notes, etc.)
* can also be obtained from other `DiagnosticBuilder`s by calling `.forget_guarantee()`
This PR is a companion to other ongoing work, namely:
* #69426
and it's ongoing implementation:
#93222
the API changes in this PR are needed to get statically-checked "only errors produce `ErrorReported` from `.emit()`", but doesn't itself provide any really strong guarantees without those other `ErrorReported` changes
* #93244
would make the choices of API changes (esp. naming) in this PR fit better overall
In order to be able to let `.emit()` return anything trustable, several changes had to be made:
* `Diagnostic`'s `level` field is now private to `rustc_errors`, to disallow arbitrary "downgrade"s from "some kind of error" to "warning" (or anything else that doesn't cause compilation to fail)
* it's still possible to replace the whole `Diagnostic` inside the `DiagnosticBuilder`, sadly, that's harder to fix, but it's unlikely enough that we can paper over it with asserts on `.emit()`
* `.cancel()` now consumes `DiagnosticBuilder`, preventing `.emit()` calls on a cancelled diagnostic
* it's also now done internally, through `DiagnosticBuilder`-private state, instead of having a `Level::Cancelled` variant that can be read (or worse, written) by the user
* this removes a hazard of calling `.cancel()` on an error then continuing to attach details to it, and even expect to be able to `.emit()` it
* warnings were switched to *only* `can_emit_warnings` on emission (instead of pre-cancelling early)
* `struct_dummy` was removed (as it relied on a pre-`Cancelled` `Diagnostic`)
* since `.emit()` doesn't consume the `DiagnosticBuilder` <sub>(I tried and gave up, it's much more work than this PR)</sub>,
we have to make `.emit()` idempotent wrt the guarantees it returns
* thankfully, `err.emit(); err.emit();` can return `ErrorReported` both times, as the second `.emit()` call has no side-effects *only* because the first one did do the appropriate emission
* `&mut Diagnostic` is now used in a lot of function signatures, which used to take `&mut DiagnosticBuilder` (in the interest of not having to make those functions generic)
* the APIs were already mostly identical, allowing for low-effort porting to this new setup
* only some of the suggestion methods needed some rework, to have the extra `DiagnosticBuilder` functionality on the `Diagnostic` methods themselves (that change is also present in #93259)
* `.emit()`/`.cancel()` aren't available, but IMO calling them from an "error decorator/annotator" function isn't a good practice, and can lead to strange behavior (from the caller's perspective)
* `.downgrade_to_delayed_bug()` was added, letting you convert any `.is_error()` diagnostic into a `delay_span_bug` one (which works because in both cases the guarantees available are the same)
This PR should ideally be reviewed commit-by-commit, since there is a lot of fallout in each.
r? `@estebank` cc `@Manishearth` `@nikomatsakis` `@mark-i-m`
These links never worked, but the lint was suppressed due to the fact
that the span was pointing into the macro. With the new macro
implementation, the span now points directly to the doc comment in the
macro invocation, so it's no longer suppressed.
Introduce `ChunkedBitSet` and use it for some dataflow analyses.
This reduces peak memory usage significantly for some programs with very
large functions.
r? `@ghost`
This reduces peak memory usage significantly for some programs with very
large functions, such as:
- `keccak`, `unicode_normalization`, and `match-stress-enum`, from
the `rustc-perf` benchmark suite;
- `http-0.2.6` from crates.io.
The new type is used in the analyses where the bitsets can get huge
(e.g. 10s of thousands of bits): `MaybeInitializedPlaces`,
`MaybeUninitializedPlaces`, and `EverInitializedPlaces`.
Some refactoring was required in `rustc_mir_dataflow`. All existing
analysis domains are either `BitSet` or a trivial wrapper around
`BitSet`, and access in a few places is done via `Borrow<BitSet>` or
`BorrowMut<BitSet>`. Now that some of these domains are `ClusterBitSet`,
that no longer works. So this commit replaces the `Borrow`/`BorrowMut`
usage with a new trait `BitSetExt` containing the needed bitset
operations. The impls just forward these to the underlying bitset type.
This required fiddling with trait bounds in a few places.
The commit also:
- Moves `static_assert_size` from `rustc_data_structures` to
`rustc_index` so it can be used in the latter; the former now
re-exports it so existing users are unaffected.
- Factors out some common "clear excess bits in the final word"
functionality in `bit_set.rs`.
- Uses `fill` in a few places instead of loops.
Main motivation: Fixes some issues with the current behavior. This PR is
more-or-less completely re-implementing the unused_unsafe lint; it’s also only
done in the MIR-version of the lint, the set of tests for the `-Zthir-unsafeck`
version no longer succeeds (and is thus disabled, see `lint-unused-unsafe.rs`).
On current nightly,
```rs
unsafe fn unsf() {}
fn inner_ignored() {
unsafe {
#[allow(unused_unsafe)]
unsafe {
unsf()
}
}
}
```
doesn’t create any warnings. This situation is not unrealistic to come by, the
inner `unsafe` block could e.g. come from a macro. Actually, this PR even
includes removal of one unused `unsafe` in the standard library that was missed
in a similar situation. (The inner `unsafe` coming from an external macro hides
the warning, too.)
The reason behind this problem is how the check currently works:
* While generating MIR, it already skips nested unsafe blocks (i.e. unsafe
nested in other unsafe) so that the inner one is always the one considered
unused
* To differentiate the cases of no unsafe operations inside the `unsafe` vs.
a surrounding `unsafe` block, there’s some ad-hoc magic walking up the HIR to
look for surrounding used `unsafe` blocks.
There’s a lot of problems with this approach besides the one presented above.
E.g. the MIR-building uses checks for `unsafe_op_in_unsafe_fn` lint to decide
early whether or not `unsafe` blocks in an `unsafe fn` are redundant and ought
to be removed.
```rs
unsafe fn granular_disallow_op_in_unsafe_fn() {
unsafe {
#[deny(unsafe_op_in_unsafe_fn)]
{
unsf();
}
}
}
```
```
error: call to unsafe function is unsafe and requires unsafe block (error E0133)
--> src/main.rs:13:13
|
13 | unsf();
| ^^^^^^ call to unsafe function
|
note: the lint level is defined here
--> src/main.rs:11:16
|
11 | #[deny(unsafe_op_in_unsafe_fn)]
| ^^^^^^^^^^^^^^^^^^^^^^
= note: consult the function's documentation for information on how to avoid undefined behavior
warning: unnecessary `unsafe` block
--> src/main.rs:10:5
|
9 | unsafe fn granular_disallow_op_in_unsafe_fn() {
| --------------------------------------------- because it's nested under this `unsafe` fn
10 | unsafe {
| ^^^^^^ unnecessary `unsafe` block
|
= note: `#[warn(unused_unsafe)]` on by default
```
Here, the intermediate `unsafe` was ignored, even though it contains a unsafe
operation that is not allowed to happen in an `unsafe fn` without an additional `unsafe` block.
Also closures were problematic and the workaround/algorithms used on current
nightly didn’t work properly. (I skipped trying to fully understand what it was
supposed to do, because this PR uses a completely different approach.)
```rs
fn nested() {
unsafe {
unsafe { unsf() }
}
}
```
```
warning: unnecessary `unsafe` block
--> src/main.rs:10:9
|
9 | unsafe {
| ------ because it's nested under this `unsafe` block
10 | unsafe { unsf() }
| ^^^^^^ unnecessary `unsafe` block
|
= note: `#[warn(unused_unsafe)]` on by default
```
vs
```rs
fn nested() {
let _ = || unsafe {
let _ = || unsafe { unsf() };
};
}
```
```
warning: unnecessary `unsafe` block
--> src/main.rs:9:16
|
9 | let _ = || unsafe {
| ^^^^^^ unnecessary `unsafe` block
|
= note: `#[warn(unused_unsafe)]` on by default
warning: unnecessary `unsafe` block
--> src/main.rs:10:20
|
10 | let _ = || unsafe { unsf() };
| ^^^^^^ unnecessary `unsafe` block
```
*note that this warning kind-of suggests that **both** unsafe blocks are redundant*
--------------------------------------------------------------------------------
I also dislike the fact that it always suggests keeping the outermost `unsafe`.
E.g. for
```rs
fn granularity() {
unsafe {
unsafe { unsf() }
unsafe { unsf() }
unsafe { unsf() }
}
}
```
I prefer if `rustc` suggests removing the more-course outer-level `unsafe`
instead of the fine-grained inner `unsafe` blocks, which it currently does on nightly:
```
warning: unnecessary `unsafe` block
--> src/main.rs:10:9
|
9 | unsafe {
| ------ because it's nested under this `unsafe` block
10 | unsafe { unsf() }
| ^^^^^^ unnecessary `unsafe` block
|
= note: `#[warn(unused_unsafe)]` on by default
warning: unnecessary `unsafe` block
--> src/main.rs:11:9
|
9 | unsafe {
| ------ because it's nested under this `unsafe` block
10 | unsafe { unsf() }
11 | unsafe { unsf() }
| ^^^^^^ unnecessary `unsafe` block
warning: unnecessary `unsafe` block
--> src/main.rs:12:9
|
9 | unsafe {
| ------ because it's nested under this `unsafe` block
...
12 | unsafe { unsf() }
| ^^^^^^ unnecessary `unsafe` block
```
--------------------------------------------------------------------------------
Needless to say, this PR addresses all these points. For context, as far as my
understanding goes, the main advantage of skipping inner unsafe blocks was that
a test case like
```rs
fn top_level_used() {
unsafe {
unsf();
unsafe { unsf() }
unsafe { unsf() }
unsafe { unsf() }
}
}
```
should generate some warning because there’s redundant nested `unsafe`, however
every single `unsafe` block _does_ contain some statement that uses it. Of course
this PR doesn’t aim change the warnings on this kind of code example, because
the current behavior, warning on all the inner `unsafe` blocks, makes sense in this case.
As mentioned, during MIR building all the unsafe blocks *are* kept now, and usage
is attributed to them. The way to still generate a warning like
```
warning: unnecessary `unsafe` block
--> src/main.rs:11:9
|
9 | unsafe {
| ------ because it's nested under this `unsafe` block
10 | unsf();
11 | unsafe { unsf() }
| ^^^^^^ unnecessary `unsafe` block
|
= note: `#[warn(unused_unsafe)]` on by default
warning: unnecessary `unsafe` block
--> src/main.rs:12:9
|
9 | unsafe {
| ------ because it's nested under this `unsafe` block
...
12 | unsafe { unsf() }
| ^^^^^^ unnecessary `unsafe` block
warning: unnecessary `unsafe` block
--> src/main.rs:13:9
|
9 | unsafe {
| ------ because it's nested under this `unsafe` block
...
13 | unsafe { unsf() }
| ^^^^^^ unnecessary `unsafe` block
```
in this case is by emitting a `unused_unsafe` warning for all of the `unsafe`
blocks that are _within a **used** unsafe block_.
The previous code had a little HIR traversal already anyways to collect a set of
all the unsafe blocks (in order to afterwards determine which ones are unused
afterwards). This PR uses such a traversal to do additional things including logic
like _always_ warn for an `unsafe` block that’s inside of another **used**
unsafe block. The traversal is expanded to include nested closures in the same go,
this simplifies a lot of things.
The whole logic around `unsafe_op_in_unsafe_fn` is a little complicated, there’s
some test cases of corner-cases in this PR. (The implementation involves
differentiating between whether a used unsafe block was used exclusively by
operations where `allow(unsafe_op_in_unsafe_fn)` was active.) The main goal was
to make sure that code should compile successfully if all the `unused_unsafe`-warnings
are addressed _simultaneously_ (by removing the respective `unsafe` blocks)
no matter how complicated the patterns of `unsafe_op_in_unsafe_fn` being
disallowed and allowed throughout the function are.
--------------------------------------------------------------------------------
One noteworthy design decision I took here: An `unsafe` block
with `allow(unused_unsafe)` **is considered used** for the purposes of
linting about redundant contained unsafe blocks. So while
```rs
fn granularity() {
unsafe { //~ ERROR: unnecessary `unsafe` block
unsafe { unsf() }
unsafe { unsf() }
unsafe { unsf() }
}
}
```
warns for the outer `unsafe` block,
```rs
fn top_level_ignored() {
#[allow(unused_unsafe)]
unsafe {
#[deny(unused_unsafe)]
{
unsafe { unsf() } //~ ERROR: unnecessary `unsafe` block
unsafe { unsf() } //~ ERROR: unnecessary `unsafe` block
unsafe { unsf() } //~ ERROR: unnecessary `unsafe` block
}
}
}
```
warns on the inner ones.
Extend uninhabited enum variant branch elimination to also affect fallthrough
The `uninhabited_enum_branching` mir opt eliminates branches on variants where the data is uninhabited. This change extends this pass to also ensure that the `otherwise` case points to a trivially unreachable bb if all inhabited variants are present in the non-otherwise branches.
I believe it was `@scottmcm` who said that LLVM eliminates some of this information in its SimplifyCFG pass. This is unfortunate, but this change should still be at least a small improvement in principle (I don't think it will show up on any benchmarks)
Adopt let else in more places
Continuation of #89933, #91018, #91481, #93046, #93590, #94011.
I have extended my clippy lint to also recognize tuple passing and match statements. The diff caused by fixing it is way above 1 thousand lines. Thus, I split it up into multiple pull requests to make reviewing easier. This is the biggest of these PRs and handles the changes outside of rustdoc, rustc_typeck, rustc_const_eval, rustc_trait_selection, which were handled in PRs #94139, #94142, #94143, #94144.
Do not ICE when inlining a function with un-satisfiable bounds
Fixes#93008
This is kinda a hack... but it's the fix I thought had the least blast-radius.
We use `normalize_param_env_or_error` to verify that the predicates in the param env are self-consistent, since with RevealAll, a bad predicate like `<&'static () as Clone>` will be evaluated with an empty ParamEnv (since it references no generics), and we'll raise an error for it.
The `uninhabited_enum_branch` miropt now also checks whether the fallthrough
case is inhabited, and if not will ensure that it points to an unreachable
block.
Specifically, rename the `Const` struct as `ConstS` and re-introduce `Const` as
this:
```
pub struct Const<'tcx>(&'tcx Interned<ConstS>);
```
This now matches `Ty` and `Predicate` more closely, including using
pointer-based `eq` and `hash`.
Notable changes:
- `mk_const` now takes a `ConstS`.
- `Const` was copy, despite being 48 bytes. Now `ConstS` is not, so need a
we need separate arena for it, because we can't use the `Dropless` one any
more.
- Many `&'tcx Const<'tcx>`/`&Const<'tcx>` to `Const<'tcx>` changes
- Many `ct.ty` to `ct.ty()` and `ct.val` to `ct.val()` changes.
- Lots of tedious sigil fiddling.
Specifically, change `Region` from this:
```
pub type Region<'tcx> = &'tcx RegionKind;
```
to this:
```
pub struct Region<'tcx>(&'tcx Interned<RegionKind>);
```
This now matches `Ty` and `Predicate` more closely.
Things to note
- Regions have always been interned, but we haven't been using pointer-based
`Eq` and `Hash`. This is now happening.
- I chose to impl `Deref` for `Region` because it makes pattern matching a lot
nicer, and `Region` can be viewed as just a smart wrapper for `RegionKind`.
- Various methods are moved from `RegionKind` to `Region`.
- There is a lot of tedious sigil changes.
- A couple of types like `HighlightBuilder`, `RegionHighlightMode` now have a
`'tcx` lifetime because they hold a `Ty<'tcx>`, so they can call `mk_region`.
- A couple of test outputs change slightly, I'm not sure why, but the new
outputs are a little better.
