Reason about borrowed classes in CopyProp.
Fixes https://github.com/rust-lang/rust/issues/141122
The current implementation of `CopyProp` avoids unifying two borrowed locals, as this would change the result of address comparison.
However, the implementation was inconsistent with the general algorithm, which identifies equivalence classes of locals and then replaces all locals by a single representative of their equivalence class.
This PR fixes it by forbidding the unification of two *classes* if any of those contain a borrowed local.
mir-opt: Do not create storage marks in EarlyOtherwiseBranch
Fixes#141212.
The first commit add `StorageDead` by creating new indirect BB that makes CFG more complicated, but I think it's better to just not create storage marks.
r? mir-opt
`slice.get(i)` should use a slice projection in MIR, like `slice[i]` does
`slice[i]` is built-in magic, so ends up being quite different from `slice.get(i)` in MIR, even though they're both doing nearly identical operations -- checking the length of the slice then getting a ref/ptr to the element if it's in-bounds.
This PR adds a `slice_get_unchecked` intrinsic for `impl SliceIndex for usize` to use to fix that, so it no longer needs to do a bunch of lines of pointer math and instead just gets the obvious single statement. (This is *not* used for the range versions, since `slice[i..]` and `slice[..k]` can't use the mir Slice projection as they're using fenceposts, not indices.)
I originally tried to do this with some kind of GVN pattern, but realized that I'm pretty sure it's not legal to optimize `BinOp::Offset` to `PlaceElem::Index` without an extremely complicated condition. Basically, the problem is that the `Index` projection on a dereferenced slice pointer *cares about the metadata*, since it's UB to `PlaceElem::Index` outside the range described by the metadata. But then you cast the fat pointer to a thin pointer then offset it, that *ignores* the slice length metadata, so it's possible to write things that are legal with `Offset` but would be UB if translated in the obvious way to `Index`. Checking (or even determining) the necessary conditions for that would be complicated and error-prone, whereas this intrinsic-based approach is quite straight-forward.
Zero backend changes, because it just lowers to MIR, so it's already supported naturally by CTFE/Miri/cg_llvm/cg_clif.
Add fast path for maybe-initializedness in liveness
r? `@matthewjasper`
Correct me if I'm wrong Matthew, but my understanding is that
1. `MaybeInitializedPlaces` is currently eagerly computed, in `do_mir_borrowck`
2. but this data is only used in liveness
3. and `liveness::trace` actually only uses it for drop-liveness
This PR moves the computation to `liveness::trace` which looks to be its only use-site. We also add a fast path there, so that it's only computed by drop-liveness.
This is interesting because 1) liveness is only computed for relevant live locals, 2) drop-liveness is only computed for relevant live locals with >0 drop points; 0 is the common case from our benchmarks, as far as I can tell, so even just computing the entire data lazily helps.
It seems possible to also reduce the domain here, and speed up the analysis for the cases where it has to be computed -- so I've left a fixme for that, and may look into it soon.
(I've come upon this while doing implementation work for polonius, so don't be too enamored with possible wins: the goal is to reduce the eventual polonius overhead and make it more palatable 😓)
gvn: bail out unavoidable non-ssa locals in repeat
Fixes#141251.
We cannot transform `*elem` to `array[idx1]` in the following code, as `idx1` has already been modified.
```rust
mir! {
let array;
let elem;
{
array = [*val; 5];
elem = &array[idx1];
idx1 = idx2;
RET = *elem;
Return()
}
}
```
Perhaps I could transform it to `array[0]`, but I prefer the conservative approach.
r? mir-opt
Add some track_caller info to precondition panics
Currently, when you encounter a precondition check, you'll always get the caller location of the implementation of the precondition checks. But with this PR, you'll be told the location of the invalid call. Which is useful.
I thought of this while looking at https://github.com/rust-lang/rust/pull/129642#issuecomment-2311703898.
The changes to `tests/ui/const*` happen because the const-eval interpreter skips `#[track_caller]` frames in its backtraces.
The perf implications of this are:
* Increased debug binary sizes. The caller_location implementation requires that the additional data we want to display here be stored in const allocations, which are deduplicated but not across crates. There is no impact on optimized build sizes. The panic path and the caller location data get optimized out.
* The compile time hit to opt-incr-patched bitmaps happens because the patch changes the line number of some function calls with precondition checks, causing us to go from 0 dirty CGUs to 1 dirty CGU.
* The other compile time hits are marginal but real, and due to doing a handful of new queries. Adding more useful data isn't completely free.
We cannot transform `*elem` to `array[idx1]` in the following code,
as `idx1` has already been modified.
```rust
mir! {
let array;
let elem;
{
array = [*val; 5];
elem = &array[idx1];
idx1 = idx2;
RET = *elem;
Return()
}
}
```
coverage: Detect unused local file IDs to avoid an LLVM assertion
Each function's coverage metadata contains a *local file table* that maps local file IDs (used by the function's mapping regions) to global file IDs (shared by all functions in the same CGU).
LLVM requires all local file IDs to have at least one mapping region, and has an assertion that will fail if it detects a local file ID with no regions. To make sure that assertion doesn't fire, we need to detect and skip functions whose metadata would trigger it.
(This can't actually happen yet, because currently all of a function's spans must belong to the same file and expansion. But this will be an important edge case when adding expansion region support.)
remove intrinsics::drop_in_place
This was only ever accidentally stable, and has been marked as deprecated since Rust 1.52, released almost 4 years ago. We've removed the old serialization `derive`s, maybe we can remove this one as well?
As suggested by ``@jhpratt,`` let's see what crater says for this one.
mir-opt: execute MatchBranchSimplification after GVN
This can provide more opportunities for MatchBranchSimplification.
Currently, rustc does not optimize the following code into a single statement at mir-opt, and this PR fixes the first case.
```rust
pub fn match1(c: bool, v1: i32, v2: i32) -> i32 {
if c { v1 - v2 } else { v1 - v2 }
}
pub fn match2(c: bool, v1: i32) -> i32 {
if c { v1 - 1 } else { v1 - 1 }
}
```
https://rust.godbolt.org/z/Y8xPMjrfM
r? mir-opt
Do not remove trivial `SwitchInt` in analysis MIR
This PR ensures that we don't prematurely remove trivial `SwitchInt` terminators which affects both the borrow-checking and runtime semantics (i.e. UB) of the code. Previously the `SimplifyCfg` optimization was removing `SwitchInt` terminators when they was "trivial", i.e. when all arms branched to the same basic block, even if that `SwitchInt` terminator had the side-effect of reading an operand which (for example) may not be initialized or may point to an invalid place in memory.
This behavior is unlike all other optimizations, which are only applied after "analysis" (i.e. borrow-checking) is finished, and which Miri disables to make sure the compiler doesn't silently remove UB.
Fixing this code "breaks" (i.e. unmasks) code that used to borrow-check but no longer does, like:
```rust
fn foo() {
let x;
let (0 | _) = x;
}
```
This match expression should perform a read because `_` does not shadow the `0` literal pattern, and the compiler should have to read the match scrutinee to compare it to 0. I've checked that this behavior does not actually manifest in practice via a crater run which came back clean: https://github.com/rust-lang/rust/pull/139042#issuecomment-2767436367
As a side-note, it may be tempting to suggest that this is actually a good thing or that we should preserve this behavior. If we wanted to make this work (i.e. trivially optimize out reads from matches that are redundant like `0 | _`), then we should be enabling this behavior *after* fixing this. However, I think it's kinda unprincipled, and for example other variations of the code don't even work today, e.g.:
```rust
fn foo() {
let x;
let (0.. | _) = x;
}
```
Visit place in `BackwardIncompatibleDropHint` statement
Remove a weird hack from the `LocalUpdater` where we were manually visiting the place stored in a `StatementKind::BackwardIncompatibleDropHint` because the MIR visitor impls weren't doing so.
Also, clean up `BackwardIncompatibleDropHint`s in `CleanupPostBorrowck`, since they're not needed for runtime MIR.
Allow GVN to produce places and not just locals.
That may be too big of a hammer, as we may introduce new deref projections (possible UB footgun + probably not good for perf).
The second commit opts out of introducing projections that don't have a stable offset, which is probably what we want. Hence no new Deref and no new Index projections.
Fixes https://github.com/rust-lang/rust/issues/138936
cc `@scottmcm` `@dianqk`
