* Unify the two maps in memory to store the allocation and its kind together.
* Share the handling of statics between CTFE and miri: The miri engine always
uses "lazy" `AllocType::Static` when encountering a static. Acessing that
static invokes CTFE (no matter the machine). The machine only has any
influence when writing to a static, which CTFE outright rejects (but miri
makes a copy-on-write).
* Add an `AllocId` to by-ref consts so miri can use them as operands without
making copies.
* Move responsibilities around for the `eval_fn_call` machine hook: The hook
just has to find the MIR (or entirely take care of everything); pushing the
new stack frame is taken care of by the miri engine.
* Expose the intrinsics and lang items implemented by CTFE so miri does not
have to reimplement them.
fix NLL ICEs
Custom type-ops reuse some of the query machinery -- but while query results are canonicalized after they are constructed, custom type ops are not, and hence we have to resolve the type variables to avoid an ICE here.
Also, use the type-op machinery for implied outlives bounds.
Fixes#53568Fixes#52992Fixes#53680
In investigating [an issue][1] with `panic_implementation` defined in an
executable that's optimized I once again got to rethinking a bit about the
`rustc_std_internal_symbol` attribute as well as weak lang items. We've sort of
been non-stop tweaking these items ever since their inception, and this
continues to the trend.
The crux of the bug was that in the reachability we have a [different branch][2]
for non-library builds which meant that weak lang items (and std internal
symbols) weren't considered reachable, causing them to get eliminiated by
ThinLTO passes. The fix was to basically tweak that branch to consider these
symbols to ensure that they're propagated all the way to the linker.
Along the way I've attempted to erode the distinction between std internal
symbols and weak lang items by having weak lang items automatically configure
fields of `CodegenFnAttrs`. That way most code no longer even considers weak
lang items and they're simply considered normal functions with attributes about
the ABI.
In the end this fixes the final comment of #51342
[1]: https://github.com/rust-lang/rust/issues/51342#issuecomment-414368019
[2]: 35bf1ae257/src/librustc/middle/reachable.rs (L225-L238)
Structs and enums can both be non-exhaustive, with a very different
meaning. This PR splits `is_non_exhaustive` to 2 separate functions - 1
for structs, and another for enums, and fixes the places that got the
usage confused.
Fixes#53549.
MIR: support user-given type annotations on fns, structs, and enums
This branch adds tooling to track user-given type annotations on functions, structs, and enum variant expressions. The user-given types are passed onto NLL which then enforces them.
cc #47184 — not a complete fix, as there are more cases to cover
r? @eddyb
cc @rust-lang/wg-compiler-nll
Use optimized SmallVec implementation
This PR replaces current SmallVec implementation with the one from the Servo project.
Closes https://github.com/rust-lang/rust/issues/51640
r? @Mark-Simulacrum
Allow panicking with string literal messages inside constants
r? @eddyb
cc https://github.com/rust-lang/rust/issues/51999
we can't implement things like `panic!("foo: {}", x)` right now because we can't call trait methods (most notably `Display::fmt`) inside constants. Also most of these impls probably have loops and conditions, so it's messy anyway.
But hey `panic!("foo")` works at least.
cc @japaric got any test ideas for `#![no_std]`?
Rollup of 10 pull requests
Successful merges:
- #53418 (Mark some suggestions as MachineApplicable)
- #53431 (Moved some feature gate ui tests to correct location)
- #53442 (Update version of rls-data used with save-analysis)
- #53504 (Set applicability for more suggestions.)
- #53541 (Fix missing impl trait display as ret type)
- #53544 (Point at the trait argument when using unboxed closure)
- #53558 (Normalize source line and column numbers.)
- #53562 (Lament the invincibility of the Turbofish)
- #53574 (Suggest direct raw-pointer dereference)
- #53585 (Remove super old comment on function that parses items)
Failed merges:
- #53472 (Use FxHash{Map,Set} instead of the default Hash{Map,Set} everywhere in rustc.)
- #53563 (use String::new() instead of String::from(""), "".to_string(), "".to_owned() or "".into())
r? @ghost
Rename TyVariants and variants
- Rename `TypeVariants` to `TyKind`.
- Remove the `Ty` prefix from each one of its variants (plus the identically-named variants of `PrimTy`).
- Rename `ty::Slice` to `ty::List`.
The new names look cleaner.
r? @eddyb
fix array drop glue: properly turn raw ptr into reference
Discovered while working on https://github.com/rust-lang/rust/pull/53424: The generated drop glue uses an assignment `ptr = cur` where `ptr` is a reference and `cur` a raw pointer. This is not well-formed MIR.
Do we have MIR sanity checks that run on the drop glue and should have caught this?
r? @eddyb
During the sanity check, we keep track of the path we are below in a `Vec`. We
avoid cloning that `Vec` unless we hit a pointer indirection. The `String`
representation is only computed when validation actually fails.
This is still roughly 45ns slower than the old state, because it now works with
an MPlaceTy and uses the appropriate abstractions, instead of working with a
ptr-align pair directly.
* Value gets renamed to Operand, so that now interpret::{Place, Operand} are the
"dynamic" versions of mir::{Place, Operand}.
* Operand and Place share the data for their "stuff is in memory"-base in a new
type, MemPlace. This also makes it possible to give some more precise types
in other areas. Both Operand and MemPlace have methods available to project
into fields (and other kinds of projections) without causing further
allocations.
* The type for "a Scalar or a ScalarPair" is called Value, and again used to
give some more precise types.
* All of these have versions with an attached layout, so that we can more often
drag the layout along instead of recomputing it. This lets us get rid of
`PlaceExtra::Downcast`. MPlaceTy and PlaceTy can only be constructed
in place.rs, making sure the layout is handled properly.
(The same should eventually be done for ValTy and OpTy.)
* All the high-level functions to write typed memory take a Place, and live in
place.rs. All the high-level typed functions to read typed memory take an
Operand, and live in operands.rs.
Exhaustive integer matching
This adds a new feature flag `exhaustive_integer_patterns` that enables exhaustive matching of integer types by their values. For example, the following is now accepted:
```rust
#![feature(exhaustive_integer_patterns)]
#![feature(exclusive_range_pattern)]
fn matcher(x: u8) {
match x { // ok
0 .. 32 => { /* foo */ }
32 => { /* bar */ }
33 ..= 255 => { /* baz */ }
}
}
```
This matching is permitted on all integer (signed/unsigned and char) types. Sensible error messages are also provided. For example:
```rust
fn matcher(x: u8) {
match x { //~ ERROR
0 .. 32 => { /* foo */ }
}
}
```
results in:
```
error[E0004]: non-exhaustive patterns: `32u8...255u8` not covered
--> matches.rs:3:9
|
6 | match x {
| ^ pattern `32u8...255u8` not covered
```
This implements https://github.com/rust-lang/rfcs/issues/1550 for https://github.com/rust-lang/rust/issues/50907. While there hasn't been a full RFC for this feature, it was suggested that this might be a feature that obviously complements the existing exhaustiveness checks (e.g. for `bool`) and so a feature gate would be sufficient for now.
This makes it more like `AllSets::{gen,kill}_set`, removes the need for
a bunch of bitset range computations, and removes the need for `Bits`.
It's marginally less efficient, because we have to allocate one bitset
per basic block instead of one large shared bitset, but the difference
is negligible in practice.
optimize reassignment immutable state
This is the "simple fix" when it comes to checking for reassignment. We just shoot for compatibility with the AST-based checker. Makes no attempt to solve #21232.
I opted for this simpler fix because I didn't want to think about complications [like the ones described here](https://github.com/rust-lang/rust/issues/21232#issuecomment-412219247).
Let's do some profiling measurements.
Fixes#53189
r? @pnkfelix
Implement Unsized Rvalues
This PR is the first step to implement RFC1909: unsized rvalues (#48055).
## Implemented
- `Sized` is removed for arguments and local bindings. (under `#![feature(unsized_locals)]`)
- Unsized locations are allowed in MIR
- Unsized places and operands are correctly translated at codegen
## Not implemented in this PR
- Additional `Sized` checks:
- tuple struct constructor (accidentally compiles now)
- closure arguments at closure generation (accidentally compiles now)
- upvars (ICEs now)
- Generating vtable for `fn method(self)` (ICEs now)
- VLAs: `[e; n]` where `n` isn't const
- Reduce unnecessary allocations
## Current status
- [x] Fix `__rust_probestack` (rust-lang-nursery/compiler-builtins#244)
- [x] Get the fix merged
- [x] `#![feature(unsized_locals)]`
- [x] Give it a tracking issue number
- [x] Lift sized checks in typeck and MIR-borrowck
- [ ] <del>Forbid `A(unsized-expr)`</del> will be another PR
- [x] Minimum working codegen
- [x] Add more examples and fill in unimplemented codegen paths
- [ ] <del>Loosen object-safety rules (will be another PR)</del>
- [ ] <del>Implement `Box<FnOnce>` (will be another PR)</del>
- [ ] <del>Reduce temporaries (will be another PR)</del>
[NLL] Returns are interesting for free regions
Based on #53088 - creating now to get feedback.
Closes#51175
* Make assigning to the return type interesting.
* Use "returning this value" instead of "return" in error messages.
* Prefer one of the explanations that we have a name for to a generic interesting cause in some cases.
* Treat causes that involve the destination of a call like assignments.
Speed up NLL with HybridIdxSetBuf.
It's a sparse-when-small but dense-when-large index set that is very
efficient for sets that (a) have few elements, (b) have large
universe_size values, and (c) are cleared frequently. Which makes it
perfect for the `gen_set` and `kill_set` sets used by the new borrow
checker.
This patch reduces `tuple-stress`'s NLL-check time by 40%, and up to 12%
for several other benchmarks. And it halves the max-rss for `keccak`,
and has smaller wins for `inflate` and `clap-rs`.
`HybridIdxSetBuf` is a sparse-when-small but dense-when-large index set
that is very efficient for sets that (a) have few elements, (b) have
large `universe_size` values, and (c) are cleared frequently. Which
makes it perfect for the `gen_set` and `kill_set` sets used by the new
borrow checker.
This patch reduces the execution time of the five slowest NLL benchmarks
by 55%, 21%, 16%, 10% and 9%. It also reduces the max-rss of three
benchmarks by 53%, 33%, and 9%.
optimize redundant borrows and escaping paths in NLL
This builds on https://github.com/rust-lang/rust/pull/53168 and adds a commit that addresses https://github.com/rust-lang/rust/issues/53176 -- or at least I think it does. I marked this as WIP because I want to see the test results (and measure the performance). I also want to double check we're not adding in any unsoundness here.
[NLL] Use span of the closure args in free region errors
Also adds a note when one of the free regions is BrEnv.
In a future PR I'll change these messages to say "return requires", which should improve them a bit more.
r? @nikomatsakis
rustc: Tweak visibility of some lang items
This commit tweaks the linker-level visibility of some lang items that rustc
uses and defines. Notably this means that `#[panic_implementation]` and
`#[alloc_error_handler]` functions are never marked as `internal`. It's up to
the linker to eliminate these, not rustc.
Additionally `#[global_allocator]` generated symbols are no longer forced to
`Default` visibility (fully exported), but rather they're relaxed to `Hidden`
visibility). This symbols are *not* needed across DLL boundaries, only as a
local implementation detail of the compiler-injected allocator symbols, so
`Hidden` should suffice.
Closes#51342Closes#52795
revert #52991
Reverts https://github.com/rust-lang/rust/pull/52991 which is flawed. I have an idea how to fix it but might as well revert first since it is so wildly flawed. That's what I get for opening PRs while on PTO =)
r? @pnkfelix
This commit tweaks the linker-level visibility of some lang items that rustc
uses and defines. Notably this means that `#[panic_implementation]` and
`#[alloc_error_handler]` functions are never marked as `internal`. It's up to
the linker to eliminate these, not rustc.
Additionally `#[global_allocator]` generated symbols are no longer forced to
`Default` visibility (fully exported), but rather they're relaxed to `Hidden`
visibility). This symbols are *not* needed across DLL boundaries, only as a
local implementation detail of the compiler-injected allocator symbols, so
`Hidden` should suffice.
Closes#51342Closes#52795
The previous iteration was a large `match` which was quite heavily indented,
making it slightly difficult to read and see what was going on. This iteration
is a refactoring (no functional change intended) to make it a bit easier on the
eyes and a bit easier to modify over time.
Place unions, pointer casts and pointer derefs behind extra feature gates
To ensure we don't stabilize these things together with const fn stabilization (or any other stabilization)
This PR moves union field accesses inside `const fn` behind a feature gate. It was possible without a feature gate before, but since `const fn` was behind a feature gate we can do this change.
While "dereferencing raw pointers" and "casting raw pointers to usize" were hard errors before this PR, one could work around them by abusing unions:
```rust
// deref
union Foo<T> {
x: &'static T,
y: *const T,
}
const FOO: u32 = unsafe { *Foo { y: 42 as *const T }.x };
// as usize cast
union Bar<T> {
x: usize,
y: *const T,
}
const BAR: usize = unsafe { Bar { y: &1u8 }.x };
```
r? @eddyb
cc @nikomatsakis
Remove references to `StaticMutex` which got removed a while ago
`StaticMutex` got removed two years ago with https://github.com/rust-lang/rust/pull/34705, but still got referenced in some comments and even an error explanation.
[NLL] Use smaller spans for errors involving closure captures
Closes#51170Closes#46599
Error messages involving closures now point to the captured variable/closure args.
r? @pnkfelix
avoid computing liveness for locals that escape into statics
Fixes#52713
I poked at this on the plane and I think it's working -- but I want to do a bit more investigation and double check. The idea is to identify those local variables where the entire value will "escape" into the return -- for them, we don't need to compute liveness, since we know that the outlives relations from the return type will force those regions to be equal to free regions. This should help with html5ever in particular.
- [x] test performance
- [x] verify correctness
- [x] add comments
r? @pnkfelix
cc @lqd
Reintroduce `Undef` and properly check constant value sizes
r? @RalfJung
cc @eddyb
basically all kinds of silent failures that never occurred are assertions now
[NLL] Dangly paths for box
Special-case `Box` in `rustc_mir::borrow_check`.
Since we know dropping a box will not access any `&mut` or `&` references, it is safe to model its destructor as only touching the contents *owned* by the box.
----
There are three main things going on here:
1. The first main thing, this PR is fixing a bug in NLL where `rustc` previously would issue a diagnostic error in a case like this:
```rust
fn foo(x: Box<&mut i32>) -> &mut i32 { &mut **x }
```
such code was accepted by the AST-borrowck in the past, but NLL was rejecting it with the following message ([playground](https://play.rust-lang.org/?gist=13c5560f73bfb16d6dab3ceaad44c0f8&version=nightly&mode=release&edition=2015))
```
error[E0597]: `**x` does not live long enough
--> src/main.rs:3:40
|
3 | fn foo(x: Box<&mut i32>) -> &mut i32 { &mut **x }
| ^^^^^^^^ - `**x` dropped here while still borrowed
| |
| borrowed value does not live long enough
|
note: borrowed value must be valid for the anonymous lifetime #1 defined on the function body at 3:1...
--> src/main.rs:3:1
|
3 | fn foo(x: Box<&mut i32>) -> &mut i32 { &mut **x }
| ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
error: aborting due to previous error
```
2. The second main thing: The reason such code was previously rejected was because NLL (MIR-borrowck) incorporates a fix for issue #31567, where it models a destructor's execution as potentially accessing any borrows held by the thing being destructed. The tests with `Scribble` model this, showing that the compiler now catches such unsoundness.
However, that fix for issue #31567 is too strong, in that NLL (MIR-borrowck) includes `Box` as one of the types with a destructor that potentially accesses any borrows held by the box. This thus was the cause of the main remaining discrepancy between AST-borrowck and MIR-borrowck, as documented in issue #45696, specifically in [the last example of this comment](https://github.com/rust-lang/rust/issues/45696#issuecomment-345367873), which I have adapted into the `fn foo` shown above.
We did close issue #45696 back in December of 2017, but AFAICT that example was not fixed by PR #46268. (And we did not include a test, etc etc.)
This PR fixes that case, by trying to model the so-called `DerefPure` semantics of `Box<T>` when we traverse the type of the input to `visit_terminator_drop`.
3. The third main thing is that during a review of the first draft of this PR, @matthewjasper pointed out that the new traversal of `Box<T>` could cause the compiler to infinite loop. I have adjusted the PR to avoid this (by tracking what types we have previously seen), and added a much needed test of this somewhat odd scenario. (Its an odd scenario because the particular case only arises for things like `struct A(Box<A>);`, something which cannot be constructed in practice.)
Fix#45696.
[NLL] Allow conflicting borrows of promoted length zero arrays
This is currently overkill as there's no way to create two conflicting borrows of any promoted.
It is possible that the following code might not fail due to const eval in the future (@oli-obk?). In which case either the array marked needs to not be promoted, or to be checked for conflicts
```rust
static mut A: () = {
let mut y = None;
let z;
let mut done_y = false;
loop {
let x = &mut [1]; // < this array
if done_y {
z = x;
break;
}
y = Some(x);
done_y = true;
}
some_const_fn(y, z); // some_const_fn expects that y to not alias z.
};
```
r? @pnkfelix @nikomatsakis
closes#52671
cc #51823
This should address issue 45696.
Since we know dropping a box will not access any `&mut` or `&`
references, it is safe to model its destructor as only touching the
contents *owned* by the box.
Note: At some point we may want to generalize this machinery to other
reference and collection types that are "pure" in the same sense as
box. If we add a `&move` reference type, it would probably also fall
into this branch of code. But for the short term, we will be
conservative and restrict this change to `Box<T>` alone.
The code works by recursively descending a deref of the `Box`. We
prevent `visit_terminator_drop` infinite-loop (which can arise in a
very obscure scenario) via a linked-list of seen types.
Note: A similar style stack-only linked-list definition can be found
in `rustc_mir::borrow_check::places_conflict`. It might be good at
some point in the future to unify the two types and put the resulting
definition into `librustc_data_structures/`.
----
One final note: Review feedback led to significant simplification of
logic here.
During review, eddyb RalfJung and I uncovered the heart of why I
needed a so-called "step 2" aka the Shallow Write to the Deref of the
box. It was because the `visit_terminator_drop`, in its base case,
will not emit any write at all (shallow or deep) to a place unless
that place has a need_drop.
So I was encoding a Shallow Write by hand for a `Box<T>`, as a
separate step from recursively descending through `*a_box` (which was
at the time known as "step 1"; it is now the *only* step, apart from
the change to the base case for `visit_terminator_drop` that this
commit now has encoded).
eddyb aruged that *something* should be emitting some sort of write in
the base case here (even a shallow one), of the dropped place, since
by analogy we also emit a write when you *move* a place. That led
to the revision here in this commit.
* (Its possible that this desired write should be attached in some
manner to StorageDead instead of Drop. But in this PR, I tried to
leave the StorageDead logic alone and focus my attention solely on
how Drop(x) is modelled in MIR-borrowck.)
NLL: On "cannot move out of type" error, print original before rewrite
NLL: On "cannot move out of type" error, print original source before rewrite.
* Arguably this change is sometimes injecting noise into the output (namely in the cases where the suggested rewrite is inline with the suggestion and we end up highlighting the original source code). I would not be opposed to something more aggressive/dynamic, like revising the suggestion code to automatically print the original source when necessary (e.g. when the error does not have a span that includes the span of the suggestion).
* Also, as another note on this change: The doc comment for `Diagnostic::span_suggestion` says:
```rust
/// The message
///
/// * should not end in any punctuation (a `:` is added automatically)
/// * should not be a question
/// * should not contain any parts like "the following", "as shown"
```
* but the `:` is *not* added when the emitted line appears out-of-line relative to the suggestion. I find that to be an unfortunate UI experience.
----
As a drive-by fix, also changed code to combine multiple suggestions for a pattern into a single multipart suggestion (which vastly improves user experience IMO).
----
Includes the updates to expected NLL diagnostics.
Fix#52877
NLL: sort diagnostics by span
Sorting the output diagnostics by span is a long planned revision to the NLL diagnostics that we hope will yield a less surprising user experience in some case.
Once we got them buffered, it was trivial to implement. (The hard part is skimming the resulting changes to the diagnostics to make sure nothing broke... Note that I largely rubber-stamped the `#[rustc_regions]` output change.)
Fix#51167
[NLL] Don't make "fake" match variables mutable
These variables can't be mutated by the user, but since they have names the unused-mut lint thinks that it should check them.
* Arguably this change is sometimes injecting noise into the output
(namely in the cases where the suggested rewrite is inline with the
suggestion and we end up highlighting the original source code).
I would not be opposed to something more aggressive/dynamic, like
revising the suggestion code to automatically print the original
source when necessary (e.g. when the error does not have a span
that includes the span of the suggestion).
* Also, as another note on this change: The doc comment for `Diagnostic::span_suggestion`
says:
/// The message
///
/// * should not end in any punctuation (a `:` is added automatically)
/// * should not be a question
/// * should not contain any parts like "the following", "as shown"
but the `:` is *not* added when the emitted line appears
out-of-line relative to the suggestion. I find that to be an
unfortunate UI experience.
----
As a drive-by fix, also changed code to combine multiple suggestions
for a pattern into a single multipart suggestion (which vastly
improves user experience IMO).
----
Includes the updates to expected NLL diagnostics.
[NLL] Fix some things for bootstrap
Some changes that are required when bootstrapping rustc with NLL enabled.
* Remove a bunch of unused `mut`s that aren't needed, but the existing lint doesn't catch.
* Rewrite a function call to satisfy NLL borrowck. Note that the borrow is two-phase, but gets activated immediately by an unsizing coercion.
cc #51823
Don't format!() string literals
Prefer `to_string()` to `format!()` take 2, this time targetting string literals. In some cases (`&format!("...")` -> `"..."`) also removes allocations. Occurences of `format!("")` are changed to `String::new()`.
Replace push loops with extend() where possible
Or set the vector capacity where I couldn't do it.
According to my [simple benchmark](https://gist.github.com/ljedrz/568e97621b749849684c1da71c27dceb) `extend`ing a vector can be over **10 times** faster than `push`ing to it in a loop:
10 elements (6.1 times faster):
```
test bench_extension ... bench: 75 ns/iter (+/- 23)
test bench_push_loop ... bench: 458 ns/iter (+/- 142)
```
100 elements (11.12 times faster):
```
test bench_extension ... bench: 87 ns/iter (+/- 26)
test bench_push_loop ... bench: 968 ns/iter (+/- 3,528)
```
1000 elements (11.04 times faster):
```
test bench_extension ... bench: 311 ns/iter (+/- 9)
test bench_push_loop ... bench: 3,436 ns/iter (+/- 233)
```
Seems like a good idea to use `extend` as much as possible.
Rollup of 11 pull requests
Successful merges:
- #52702 (Suggest fix when encountering different mutability from impl to trait)
- #52703 (Improve a few vectors - calculate capacity or build from iterators)
- #52740 (Suggest underscore when using dashes in crate namet push fork)
- #52759 (Impl Send & Sync for JoinHandle)
- #52760 (rustc_metadata: test loading atoi instead of cos)
- #52763 (Omit the vendor component in Fuchsia triple)
- #52765 (Remove unused "-Zenable_nonzeroing_move_hints" flag)
- #52769 (Incorporate a stray test)
- #52777 (Fix doc comment for 'ptr::copy_to' method)
- #52779 (revert accidental atty downgrade)
- #52781 (Use a slice where a vector is not necessary)
Failed merges:
r? @ghost
[NLL] Use better spans in some errors
* Use the span of the discriminant and patterns for "fake" statements created to properly check matches. I plan to special case these soon, but this felt like a good first step
* Use the span of the statement, rather than the initialization, when reporting move errors for `let x = ...`, which avoids giving an unhelpful suggestion to use `&{ }`.
r? @nikomatsakis cc @pnkfelix
Rollup of bare_trait_objects PRs
All deny attributes were moved into bootstrap so they can be disabled with a line of config.
Warnings for external tools are allowed and it's up to the tool's maintainer to keep it warnings free.
r? @Mark-Simulacrum
cc @ljedrz @kennytm
[NLL] make temp for each candidate in `match` arm
In NLL, `ref mut` patterns leverage the two-phase borrow infrastructure to allow the shared borrows within a guard before the "activation" of the mutable borrow when we begin execution of the match arm's body. (There is further discussion of this on PR #50783.)
To accommodate the restrictions we impose on two-phase borrows (namely that there is a one-to-one mapping between each activation and the original initialization), this PR is making separate temps for each candidate pattern. So in an arm like this:
```rust
PatA(_, ref mut ident) |
PatB(ref mut ident) |
PatC(_, _, ref mut ident) |
PatD(ref mut ident) if guard_stuff(ident) => ...
```
instead of 3 temps (two for the guard and one for the arm body), we now have 4 + 2 temps associated with `ident`: one for each candidate plus the actual temp that the guard uses directly, and then the sixth is the temp used in the arm body.
Fix#51348
Add `-Z borrowck=migrate`
This adds `-Z borrowck=migrate`, which represents the way we want to migrate to NLL under Rust versions to come. It also hooks this new mode into `--edition 2018`, which means we're officially turning NLL on in the 2018 edition.
The basic idea of `-Z borrowck=migrate` that there are cases where NLL is fixing old soundness bugs in the borrow-checker, but in order to avoid just breaking code by immediately rejecting the programs that hit those soundness bugs, we instead use the following strategy:
If your code is accepted by NLL, then we accept it.
If your code is rejected by both NLL and the old AST-borrowck, then we reject it.
If your code is rejected by NLL but accepted by the old AST-borrowck, then we emit the new NLL errors as **warnings**.
These warnings will be turned into hard errors in the future, and they say so in these diagnostics.
Fix#46908
Rollup of 16 pull requests
Successful merges:
- #52558 (Add tests for ICEs which no longer repro)
- #52610 (Clarify what a task is)
- #52617 (Don't match on region kinds when reporting NLL errors)
- #52635 (Fix #[linkage] propagation though generic functions)
- #52647 (Suggest to take and ignore args while closure args count mismatching)
- #52649 (Point spans to inner elements of format strings)
- #52654 (Format linker args in a way that works for gcc and ld)
- #52667 (update the stdsimd submodule)
- #52674 (Impl Executor for Box<E: Executor>)
- #52690 (ARM: expose `rclass` and `dsp` target features)
- #52692 (Improve readability in a few sorts)
- #52695 (Hide some lints which are not quite right the way they are reported to the user)
- #52718 (State default capacity for BufReader/BufWriter)
- #52721 (std::ops::Try impl for std::task::Poll)
- #52723 (rustc: Register crates under their real names)
- #52734 (sparc ABI issue - structure returning from function is returned in 64bit registers (with tests))
Failed merges:
- #52678 ([NLL] Use better spans in some errors)
r? @ghost
Don't match on region kinds when reporting NLL errors
First half (by number of tests affected) of the changes to "does not live long enough".
Now that lexical MIR borrowck is gone, region kinds are always ReVar, so matching on them to change errors does nothing.
Changes "borrowed value only lives until here" to "`x` is dropped here while still borrowed".
r? @pnkfelix cc @nikomatsakis
Also convert an ICE that became reachable code under borrowck=migrate
into a normally reported error (which is then downgraded to a
warning). This actually has a nice side benefit of providing a
somewhat more useful error message, at least in the particular case of
the example from issue #27282.
This required a bit of plumbing to keep track of candidates. But I
took advantage of the hack session to try to improve the docs for the
relevant structs here.
(I also tried to simplify some of the related code in passing.)
