Don't allocate on SimplifyCfg/Locals/Const on every MIR pass
Hey! 👋🏾 This is a first PR attempt to see if I could speed up some rustc internals.
Thought process:
```rust
pub struct SimplifyCfg {
label: String,
}
```
in [compiler/src/rustc_mir_transform/simplify.rs](7908a1d654/compiler/rustc_mir_transform/src/simplify.rs (L39)) fires multiple times per MIR analysis. This means that a likely string allocation is happening in each of these runs, which may add up, as they are not being lazily allocated or cached in between the different passes.
...yes, I know that adding a global static array is probably not the future-proof solution, but I wanted to lob this now as a proof of concept to see if it's worth shaving off a few cycles and then making more robust.
Encode hashes as bytes, not varint
In a few places, we store hashes as `u64` or `u128` and then apply `derive(Decodable, Encodable)` to the enclosing struct/enum. It is more efficient to encode hashes directly than try to apply some varint encoding. This PR adds two new types `Hash64` and `Hash128` which are produced by `StableHasher` and replace every use of storing a `u64` or `u128` that represents a hash.
Distribution of the byte lengths of leb128 encodings, from `x build --stage 2` with `incremental = true`
Before:
```
( 1) 373418203 (53.7%, 53.7%): 1
( 2) 196240113 (28.2%, 81.9%): 3
( 3) 108157958 (15.6%, 97.5%): 2
( 4) 17213120 ( 2.5%, 99.9%): 4
( 5) 223614 ( 0.0%,100.0%): 9
( 6) 216262 ( 0.0%,100.0%): 10
( 7) 15447 ( 0.0%,100.0%): 5
( 8) 3633 ( 0.0%,100.0%): 19
( 9) 3030 ( 0.0%,100.0%): 8
( 10) 1167 ( 0.0%,100.0%): 18
( 11) 1032 ( 0.0%,100.0%): 7
( 12) 1003 ( 0.0%,100.0%): 6
( 13) 10 ( 0.0%,100.0%): 16
( 14) 10 ( 0.0%,100.0%): 17
( 15) 5 ( 0.0%,100.0%): 12
( 16) 4 ( 0.0%,100.0%): 14
```
After:
```
( 1) 372939136 (53.7%, 53.7%): 1
( 2) 196240140 (28.3%, 82.0%): 3
( 3) 108014969 (15.6%, 97.5%): 2
( 4) 17192375 ( 2.5%,100.0%): 4
( 5) 435 ( 0.0%,100.0%): 5
( 6) 83 ( 0.0%,100.0%): 18
( 7) 79 ( 0.0%,100.0%): 10
( 8) 50 ( 0.0%,100.0%): 9
( 9) 6 ( 0.0%,100.0%): 19
```
The remaining 9 or 10 and 18 or 19 are `u64` and `u128` respectively that have the high bits set. As far as I can tell these are coming primarily from `SwitchTargets`.
Check freeze with right param-env in `deduced_param_attrs`
We're checking if a trait (`Freeze`) holds in a polymorphic function, but not using that function's own (reveal-all) param-env. This causes us to try to eagerly normalize a specializable projection type that has no default value, which causes an ICE.
Fixes#110171
Permit MIR inlining without #[inline]
I noticed that there are at least a handful of portable-simd functions that have no `#[inline]` but compile to an assign + return.
I locally benchmarked inlining thresholds between 0 and 50 in increments of 5, and 50 seems to be the best. Interesting. That didn't include check builds though, ~maybe perf will have something to say about that~.
Perf has little useful to say about this. We generally regress all the check builds, as best as I can tell, due to a number of small codegen changes in a particular hot function in the compiler. Probably this is because we've nudged the inlining outcomes all over, and uses of `#[inline(always)]`/`#[inline(never)]` might need to be adjusted.
This allows us to get rid of the `rustc_const_eval->rustc_borrowck`
dependency edge which was delaying the compilation of borrowck.
The added utils in `rustc_middle` are small and should not affect
compile times there.
Preserve argument indexes when inlining MIR
We store argument indexes on VarDebugInfo. Unlike the previous method of relying on the variable index to know whether a variable is an argument, this survives MIR inlining.
We also no longer check if var.source_info.scope is the outermost scope. When a function gets inlined, the arguments to the inner function will no longer be in the outermost scope. What we care about though is whether they were in the outermost scope prior to inlining, which we know by whether we assigned an argument index.
Fixes#83217
I considered using `Option<NonZeroU16>` instead of `Option<u16>` to store the index. I didn't because `TypeFoldable` isn't implemented for `NonZeroU16` and because it looks like due to padding, it currently wouldn't make any difference. But I indexed from 1 anyway because (a) it'll make it easier if later it becomes worthwhile to use a `NonZeroU16` and because the arguments were previously indexed from 1, so it made for a smaller change.
This is my first PR on rust-lang/rust, so apologies if I've gotten anything not quite right.
We store argument indexes on VarDebugInfo. Unlike the previous method of
relying on the variable index to know whether a variable is an argument,
this survives MIR inlining.
We also no longer check if var.source_info.scope is the outermost scope.
When a function gets inlined, the arguments to the inner function will
no longer be in the outermost scope. What we care about though is
whether they were in the outermost scope prior to inlining, which we
know by whether we assigned an argument index.
Make elaboration generic over input
Combines all the `elaborate_*` family of functions into just one, which is an iterator over the same type that you pass in (e.g. elaborating `Predicate` gives `Predicate`s, elaborating `Obligation`s gives `Obligation`s, etc.)
Refactor unwind in MIR
This makes unwinding from current `Option<BasicBlock>` into
```rust
enum UnwindAction {
Continue,
Cleanup(BasicBlock),
Unreachable,
Terminate,
}
```
cc `@JakobDegen` `@RalfJung` `@Amanieu`
Unify terminology used in unwind action and terminator, and reflect
the fact that a nounwind panic is triggered instead of an immediate
abort is triggered for this terminator.
Move a const-prop-lint specific hack from mir interpret to const-prop-lint and make it fallible
fixes#109743
This hack didn't need to live in the mir interpreter. For const-prop-lint it is entirely correct to avoid doing any const prop if normalization fails at this stage. Most likely we couldn't const propagate anything anyway, and if revealing was needed (so opaque types were involved), we wouldn't want to be too smart and leak the hidden type anyway.
Insert alignment checks for pointer dereferences when debug assertions are enabled
Closes https://github.com/rust-lang/rust/issues/54915
- [x] Jake tells me this sounds like a place to use `MirPatch`, but I can't figure out how to insert a new basic block with a new terminator in the middle of an existing basic block, using `MirPatch`. (if nobody else backs up this point I'm checking this as "not actually a good idea" because the code looks pretty clean to me after rearranging it a bit)
- [x] Using `CastKind::PointerExposeAddress` is definitely wrong, we don't want to expose. Calling a function to get the pointer address seems quite excessive. ~I'll see if I can add a new `CastKind`.~ `CastKind::Transmute` to the rescue!
- [x] Implement a more helpful panic message like slice bounds checking.
r? `@oli-obk`
Update `ty::VariantDef` to use `IndexVec<FieldIdx, FieldDef>`
And while doing the updates for that, also uses `FieldIdx` in `ProjectionKind::Field` and `TypeckResults::field_indices`.
There's more places that could use it (like `rustc_const_eval` and `LayoutS`), but I tried to keep this PR from exploding to *even more* places.
Part 2/? of https://github.com/rust-lang/compiler-team/issues/606
And while doing the updates for that, also uses `FieldIdx` in `ProjectionKind::Field` and `TypeckResults::field_indices`.
There's more places that could use it (like `rustc_const_eval` and `LayoutS`), but I tried to keep this PR from exploding to *even more* places.
Part 2/? of https://github.com/rust-lang/compiler-team/issues/606
Partial stabilization of `once_cell`
This PR aims to stabilize a portion of the `once_cell` feature:
- `core::cell::OnceCell`
- `std::cell::OnceCell` (re-export of the above)
- `std::sync::OnceLock`
This will leave `LazyCell` and `LazyLock` unstabilized, which have been moved to the `lazy_cell` feature flag.
Tracking issue: https://github.com/rust-lang/rust/issues/74465 (does not fully close, but it may make sense to move to a new issue)
Future steps for separate PRs:
- ~~Add `#[inline]` to many methods~~ #105651
- Update cranelift usage of the `once_cell` crate
- Update rust-analyzer usage of the `once_cell` crate
- Update error messages discussing once_cell
## To be stabilized API summary
```rust
// core::cell (in core/cell/once.rs)
pub struct OnceCell<T> { .. }
impl<T> OnceCell<T> {
pub const fn new() -> OnceCell<T>;
pub fn get(&self) -> Option<&T>;
pub fn get_mut(&mut self) -> Option<&mut T>;
pub fn set(&self, value: T) -> Result<(), T>;
pub fn get_or_init<F>(&self, f: F) -> &T where F: FnOnce() -> T;
pub fn into_inner(self) -> Option<T>;
pub fn take(&mut self) -> Option<T>;
}
impl<T: Clone> Clone for OnceCell<T>;
impl<T: Debug> Debug for OnceCell<T>
impl<T> Default for OnceCell<T>;
impl<T> From<T> for OnceCell<T>;
impl<T: PartialEq> PartialEq for OnceCell<T>;
impl<T: Eq> Eq for OnceCell<T>;
```
```rust
// std::sync (in std/sync/once_lock.rs)
impl<T> OnceLock<T> {
pub const fn new() -> OnceLock<T>;
pub fn get(&self) -> Option<&T>;
pub fn get_mut(&mut self) -> Option<&mut T>;
pub fn set(&self, value: T) -> Result<(), T>;
pub fn get_or_init<F>(&self, f: F) -> &T where F: FnOnce() -> T;
pub fn into_inner(self) -> Option<T>;
pub fn take(&mut self) -> Option<T>;
}
impl<T: Clone> Clone for OnceLock<T>;
impl<T: Debug> Debug for OnceLock<T>;
impl<T> Default for OnceLock<T>;
impl<#[may_dangle] T> Drop for OnceLock<T>;
impl<T> From<T> for OnceLock<T>;
impl<T: PartialEq> PartialEq for OnceLock<T>
impl<T: Eq> Eq for OnceLock<T>;
impl<T: RefUnwindSafe + UnwindSafe> RefUnwindSafe for OnceLock<T>;
unsafe impl<T: Send> Send for OnceLock<T>;
unsafe impl<T: Sync + Send> Sync for OnceLock<T>;
impl<T: UnwindSafe> UnwindSafe for OnceLock<T>;
```
No longer planned as part of this PR, and moved to the `rust_cell_try` feature gate:
```rust
impl<T> OnceCell<T> {
pub fn get_or_try_init<F, E>(&self, f: F) -> Result<&T, E> where F: FnOnce() -> Result<T, E>;
}
impl<T> OnceLock<T> {
pub fn get_or_try_init<F, E>(&self, f: F) -> Result<&T, E> where F: FnOnce() -> Result<T, E>;
}
```
I am new to this process so would appreciate mentorship wherever needed.
Move `mir::Field` → `abi::FieldIdx`
The first PR for https://github.com/rust-lang/compiler-team/issues/606
This is just the move-and-rename, because it's plenty big already. Future PRs will start using `FieldIdx` more broadly, and concomitantly removing `FieldIdx::new`s.
Support TLS access into dylibs on Windows
This allows access to `#[thread_local]` in upstream dylibs on Windows by introducing a MIR shim to return the address of the thread local. Accesses that go into an upstream dylib will call the MIR shim to get the address of it.
`convert_tls_rvalues` is introduced in `rustc_codegen_ssa` which rewrites MIR TLS accesses to dummy calls which are replaced with calls to the MIR shims when the dummy calls are lowered to backend calls.
A new `dll_tls_export` target option enables this behavior with a `false` value which is set for Windows platforms.
This fixes https://github.com/rust-lang/rust/issues/84933.
As I've been trying to replace a `Vec` with an `IndexVec`, having `last` exist on both but returning very different types makes the transition a bit awkward -- the errors are later, where you get things like "there's no `ty` method on `mir::Field`" rather than a more localized error like "hey, there's no `last` on `IndexVec`".
So I propose renaming `last` to `last_index` to help distinguish `Vec::last`, which returns an element, and `IndexVec::last_index`, which returns an index.
(Similarly, `Iterator::last` also returns an element, not an index.)
The first PR for https://github.com/rust-lang/compiler-team/issues/606
This is just the move-and-rename, because it's plenty big-and-bitrotty already. Future PRs will start using `FieldIdx` more broadly, and concomitantly removing `FieldIdx::new`s.
Thanks to the combination of #108246 and #108442 it could already remove identity transmutes.
With this PR, it can also simplify them to `IntToInt` casts, `Discriminant` reads, or `Field` projections.
Add a builtin `FnPtr` trait that is implemented for all function pointers
r? `@ghost`
Rebased version of https://github.com/rust-lang/rust/pull/99531 (plus adjustments mentioned in the PR).
If perf is happy with this version, I would like to land it, even if the diagnostics fix in 9df8e1befb5031a5bf9d8dfe25170620642d3c59 only works for `FnPtr` specifically, and does not generally improve blanket impls.
Don't elaborate non-obligations into obligations
It's suspicious to elaborate a `PolyTraitRef` or `Predicate` into an `Obligation`, since the former does not have a param-env associated with it, but the latter does. This is a footgun that, while not being misused *currently* in the compiler, easily could be misused by someone less familiar with the elaborator's inner workings.
This PR just changes the API -- ideally, the elaborator wouldn't even have to deal with obligations if we're not elaborating obligations, but that would require a bit more abstraction than I could be bothered with today.
Permit the MIR inliner to inline diverging functions
This heuristic prevents inlining of `hint::unreachable_unchecked`, which in turn makes `Option/Result::unwrap_unchecked` a bad inlining candidate. I looked through the changes to `core`, `alloc`, `std`, and `hashbrown` by hand and they all seem reasonable. Let's see how this looks in perf...
---
Based on rustc-perf it looks like this regresses ctfe-stress, and the cachegrind diff indicates that this regression is in `InterpCx::statement`. I don't know how to do any deeper analysis because that function is _enormous_ in the try toolchain, which has no debuginfo in it. And a local build produces significantly different codegen for that function, even with LTO.
Since structs are always `VariantIdx(0)`, there's a bunch of files where the only reason they had `VariantIdx` or `vec::Idx` imported at all was to get the first variant.
So this uses a constant for that, and adds some doc-comments to `VariantIdx` while I'm there, since it doesn't have any today.
Updates `interpret`, `codegen_ssa`, and `codegen_cranelift` to consume the new cast instead of the intrinsic.
Includes `CastTransmute` for custom MIR building, to be able to test the extra UB.
Detect uninhabited types early in const eval
r? `@RalfJung`
implements https://github.com/rust-lang/rust/pull/108442#discussion_r1143003840
this is a breaking change, as some UB during const eval is now detected instead of silently being ignored. Users can see this and other UB that may cause future breakage with `-Zextra-const-ub-checks` or just by running miri on their code, which sets that flag by default.
move Option::as_slice to intrinsic
````@scottmcm```` suggested on #109095 I use a direct approach of unpacking the operation in MIR lowering, so here's the implementation.
cc ````@nikic```` as this should hopefully unblock #107224 (though perhaps other changes to the prior implementation, which I left for bootstrapping, are needed).
Only clear written-to locals in ConstProp
This aims to revert the regression in https://github.com/rust-lang/rust/pull/108872
Clearing all locals at the end of each bb is very costly.
This PR goes back to the original implementation of recording which locals are modified.
Wrap the whole LocalInfo in ClearCrossCrate.
MIR contains a lot of information about locals. The primary purpose of this information is the quality of borrowck diagnostics.
This PR aims to drop this information after MIR analyses are finished, ie. starting from post-cleanup runtime MIR.
Ensure `ptr::read` gets all the same LLVM `load` metadata that dereferencing does
I was looking into `array::IntoIter` optimization, and noticed that it wasn't annotating the loads with `noundef` for simple things like `array::IntoIter<i32, N>`. Trying to narrow it down, it seems that was because `MaybeUninit::assume_init_read` isn't marking the load as initialized (<https://rust.godbolt.org/z/Mxd8TPTnv>), which is unfortunate since that's basically its reason to exist.
The root cause is that `ptr::read` is currently implemented via the *untyped* `copy_nonoverlapping`, and thus the `load` doesn't get any type-aware metadata: no `noundef`, no `!range`. This PR solves that by lowering `ptr::read(p)` to `copy *p` in MIR, for which the backends already do the right thing.
Fortuitiously, this also improves the IR we give to LLVM for things like `mem::replace`, and fixes a couple of long-standing bugs where `ptr::read` on `Copy` types was worse than `*`ing them.
Zulip conversation: <https://rust-lang.zulipchat.com/#narrow/stream/219381-t-libs/topic/Move.20array.3A.3AIntoIter.20to.20ManuallyDrop/near/341189936>
cc `@erikdesjardins` `@JakobDegen` `@workingjubilee` `@the8472`
Fixes#106369Fixes#73258
Remove `box_syntax`
r? `@Nilstrieb`
This removes the feature `box_syntax`, which allows the use of `box <expr>` to create a Box, and finalises removing use of the feature from the compiler. `box_patterns` (allowing the use of `box <pat>` in a pattern) is unaffected.
It also removes `ast::ExprKind::Box` - the only way to create a 'box' expression now is with the rustc-internal `#[rustc_box]` attribute.
As a temporary measure to help users move away, `box <expr>` now parses the inner expression, and emits a `MachineApplicable` lint to replace it with `Box::new`
Closes#49733
Strengthen state tracking in const-prop
Some/many of the changes are replicated between both the const-prop lint and the const-prop optimization.
Behaviour changes:
- const-prop opt does not give a span to propagated values. This was useless as that span's primary purpose is to diagnose evaluation failure in codegen.
- we remove the `OnlyPropagateInto` mode. It was only used for function arguments, which are better modeled by a write before entry.
- the tracking of assignments and discriminants make clearer that we do nothing in `NoPropagation` mode or on indirect places.
Ensure value is on the on-disk cache before returning from `ensure()`.
The current logic for `ensure()` a query just checks that the node is green in the dependency graph.
However, a lot of places use `ensure()` to prevent the query from being called later. This is the case before stealing a query result.
If the query is actually green but the value is not available in the on-disk cache, `ensure` would return, but a subsequent call to the full query would run the code, and attempt to read from a stolen value.
This PR conforms the query system to the usage by checking whether the queried value is loadable from disk before returning.
Sadly, I can't manage to craft a proper test...
Should fix all instances of "attempted to read from stolen value".
Avoid unnecessary hashing
I noticed some stable hashing being done in a non-incremental build. It turns out that some of this is necessary to compute the crate hash, but some of it is not. Removing the unnecessary hashing is a perf win.
r? `@cjgillot`
I was looking into `array::IntoIter` optimization, and noticed that it wasn't annotating the loads with `noundef` for simple things like `array::IntoIter<i32, N>`.
Turned out to be a more general problem as `MaybeUninit::assume_init_read` isn't marking the load as initialized (<https://rust.godbolt.org/z/Mxd8TPTnv>), which is unfortunate since that's basically its reason to exist.
This PR lowers `ptr::read(p)` to `copy *p` in MIR, which fortuitiously also improves the IR we give to LLVM for things like `mem::replace`.
