Autodiff batching
Enzyme supports batching, which is especially known from the ML side when training neural networks.
There we would normally have a training loop, where in each iteration we would pass in some data (e.g. an image), and a target vector. Based on how close we are with our prediction we compute our loss, and then use backpropagation to compute the gradients and update our weights.
That's quite inefficient, so what you normally do is passing in a batch of 8/16/.. images and targets, and compute the gradients for those all at once, allowing better optimizations.
Enzyme supports batching in two ways, the first one (which I implemented here) just accepts a Batch size,
and then each Dual/Duplicated argument has not one, but N shadow arguments. So instead of
```rs
for i in 0..100 {
df(x[i], y[i], 1234);
}
```
You can now do
```rs
for i in 0..100.step_by(4) {
df(x[i+0],x[i+1],x[i+2],x[i+3], y[i+0], y[i+1], y[i+2], y[i+3], 1234);
}
```
which will give the same results, but allows better compiler optimizations. See the testcase for details.
There is a second variant, where we can mark certain arguments and instead of having to pass in N shadow arguments, Enzyme assumes that the argument is N times longer. I.e. instead of accepting 4 slices with 12 floats each, we would accept one slice with 48 floats. I'll implement this over the next days.
I will also add more tests for both modes.
For any one preferring some more interactive explanation, here's a video of Tim's llvm dev talk, where he presents his work. https://www.youtube.com/watch?v=edvaLAL5RqU
I'll also add some other docs to the dev guide and user docs in another PR.
r? ghost
Tracking:
- https://github.com/rust-lang/rust/issues/124509
- https://github.com/rust-lang/rust/issues/135283
add `TypingMode::Borrowck`
Shares the first commit with #138499, doesn't really matter which PR to land first 😊😁
Introduces `TypingMode::Borrowck` which unlike `TypingMode::Analysis`, uses the hidden type computed by HIR typeck as the initial value of opaques instead of an unconstrained infer var. This is a part of https://github.com/rust-lang/types-team/issues/129.
Using this new `TypingMode` is unfortunately a breaking change for now, see tests/ui/impl-trait/non-defining-uses/as-projection-term.rs. Using an inference variable as the initial value results in non-defining uses in the defining scope. We therefore only enable it if with `-Znext-solver=globally` or `-Ztyping-mode-borrowck`
To do that the PR contains the following changes:
- `TypeckResults::concrete_opaque_type` are already mapped to the definition of the opaque type
- writeback now checks that the non-lifetime parameters of the opaque are universal
- for this, `fn check_opaque_type_parameter_valid` is moved from `rustc_borrowck` to `rustc_trait_selection`
- we add a new `query type_of_opaque_hir_typeck` which, using the same visitors as MIR typeck, attempts to merge the hidden types from HIR typeck from all defining scopes
- done by adding a `DefiningScopeKind` flag to toggle between using borrowck and HIR typeck
- the visitors stop checking that the MIR type matches the HIR type. This is trivial as the HIR type are now used as the initial hidden types of the opaque. This check is useful as a safeguard when not using `TypingMode::Borrowck`, but adding it to the new structure is annoying and it's not soundness critical, so I intend to not add it back.
- add a `TypingMode::Borrowck` which behaves just like `TypingMode::Analysis` except when normalizing opaque types
- it uses `type_of_opaque_hir_typeck(opaque)` as the initial value after replacing its regions with new inference vars
- it uses structural lookup in the new solver
fixes#112201, fixes#132335, fixes#137751
r? `@compiler-errors` `@oli-obk`
Initial support for auto traits with default bounds
This PR is part of ["MCP: Low level components for async drop"](https://github.com/rust-lang/compiler-team/issues/727)
Tracking issue: #138781
Summary: https://github.com/rust-lang/rust/pull/120706#issuecomment-1934006762
### Intro
Sometimes we want to use type system to express specific behavior and provide safety guarantees. This behavior can be specified by various "marker" traits. For example, we use `Send` and `Sync` to keep track of which types are thread safe. As the language develops, there are more problems that could be solved by adding new marker traits:
- to forbid types with an async destructor to be dropped in a synchronous context a trait like `SyncDrop` could be used [Async destructors, async genericity and completion futures](https://sabrinajewson.org/blog/async-drop).
- to support [scoped tasks](https://without.boats/blog/the-scoped-task-trilemma/) or in a more general sense to provide a [destruction guarantee](https://zetanumbers.github.io/book/myosotis.html) there is a desire among some users to see a `Leak` (or `Forget`) trait.
- Withoutboats in his [post](https://without.boats/blog/changing-the-rules-of-rust/) reflected on the use of `Move` trait instead of a `Pin`.
All the traits proposed above are supposed to be auto traits implemented for most types, and usually implemented automatically by compiler.
For backward compatibility these traits have to be added implicitly to all bound lists in old code (see below). Adding new default bounds involves many difficulties: many standard library interfaces may need to opt out of those default bounds, and therefore be infected with confusing `?Trait` syntax, migration to a new edition may contain backward compatibility holes, supporting new traits in the compiler can be quite difficult and so forth. Anyway, it's hard to evaluate the complexity until we try the system on a practice.
In this PR we introduce new optional lang items for traits that are added to all bound lists by default, similarly to existing `Sized`. The examples of such traits could be `Leak`, `Move`, `SyncDrop` or something else, it doesn't matter much right now (further I will call them `DefaultAutoTrait`'s). We want to land this change into rustc under an option, so it becomes available in bootstrap compiler. Then we'll be able to do standard library experiments with the aforementioned traits without adding hundreds of `#[cfg(not(bootstrap))]`s. Based on the experiments, we can come up with some scheme for the next edition, in which such bounds are added in a more targeted way, and not just everywhere.
Most of the implementation is basically a refactoring that replaces hardcoded uses of `Sized` with iterating over a list of traits including both `Sized` and the new traits when `-Zexperimental-default-bounds` is enabled (or just `Sized` as before, if the option is not enabled).
### Default bounds for old editions
All existing types, including generic parameters, are considered `Leak`/`Move`/`SyncDrop` and can be forgotten, moved or destroyed in generic contexts without specifying any bounds. New types that cannot be, for example, forgotten and do not implement `Leak` can be added at some point, and they should not be usable in such generic contexts in existing code.
To both maintain this property and keep backward compatibility with existing code, the new traits should be added as default bounds _everywhere_ in previous editions. Besides the implicit `Sized` bound contexts that includes supertrait lists and trait lists in trait objects (`dyn Trait1 + ... + TraitN`). Compiler should also generate implicit `DefaultAutoTrait` implementations for foreign types (`extern { type Foo; }`) because they are also currently usable in generic contexts without any bounds.
#### Supertraits
Adding the new traits as supertraits to all existing traits is potentially necessary, because, for example, using a `Self` param in a trait's associated item may be a breaking change otherwise:
```rust
trait Foo: Sized {
fn new() -> Option<Self>; // ERROR: `Option` requires `DefaultAutoTrait`, but `Self` is not `DefaultAutoTrait`
}
// desugared `Option`
enum Option<T: DefaultAutoTrait + Sized> {
Some(T),
None,
}
```
However, default supertraits can significantly affect compiler performance. For example, if we know that `T: Trait`, the compiler would deduce that `T: DefaultAutoTrait`. It also implies proving `F: DefaultAutoTrait` for each field `F` of type `T` until an explicit impl is be provided.
If the standard library is not modified, then even traits like `Copy` or `Send` would get these supertraits.
In this PR for optimization purposes instead of adding default supertraits, bounds are added to the associated items:
```rust
// Default bounds are generated in the following way:
trait Trait {
fn foo(&self) where Self: DefaultAutoTrait {}
}
// instead of this:
trait Trait: DefaultAutoTrait {
fn foo(&self) {}
}
```
It is not always possible to do this optimization because of backward compatibility:
```rust
pub trait Trait<Rhs = Self> {}
pub trait Trait1 : Trait {} // ERROR: `Rhs` requires `DefaultAutoTrait`, but `Self` is not `DefaultAutoTrait`
```
or
```rust
trait Trait {
type Type where Self: Sized;
}
trait Trait2<T> : Trait<Type = T> {} // ERROR: `???` requires `DefaultAutoTrait`, but `Self` is not `DefaultAutoTrait`
```
Therefore, `DefaultAutoTrait`'s are still being added to supertraits if the `Self` params or type bindings were found in the trait header.
#### Trait objects
Trait objects requires explicit `+ Trait` bound to implement corresponding trait which is not backward compatible:
```rust
fn use_trait_object(x: Box<dyn Trait>) {
foo(x) // ERROR: `foo` requires `DefaultAutoTrait`, but `dyn Trait` is not `DefaultAutoTrait`
}
// implicit T: DefaultAutoTrait here
fn foo<T>(_: T) {}
```
So, for a trait object `dyn Trait` we should add an implicit bound `dyn Trait + DefaultAutoTrait` to make it usable, and allow relaxing it with a question mark syntax `dyn Trait + ?DefaultAutoTrait` when it's not necessary.
#### Foreign types
If compiler doesn't generate auto trait implementations for a foreign type, then it's a breaking change if the default bounds are added everywhere else:
```rust
// implicit T: DefaultAutoTrait here
fn foo<T: ?Sized>(_: &T) {}
extern "C" {
type ExternTy;
}
fn forward_extern_ty(x: &ExternTy) {
foo(x); // ERROR: `foo` requires `DefaultAutoTrait`, but `ExternTy` is not `DefaultAutoTrait`
}
```
We'll have to enable implicit `DefaultAutoTrait` implementations for foreign types at least for previous editions:
```rust
// implicit T: DefaultAutoTrait here
fn foo<T: ?Sized>(_: &T) {}
extern "C" {
type ExternTy;
}
impl DefaultAutoTrait for ExternTy {} // implicit impl
fn forward_extern_ty(x: &ExternTy) {
foo(x); // OK
}
```
### Unresolved questions
New default bounds affect all existing Rust code complicating an already complex type system.
- Proving an auto trait predicate requires recursively traversing the type and proving the predicate for it's fields. This leads to a significant performance regression. Measurements for the stage 2 compiler build show up to 3x regression.
- We hope that fast path optimizations for well known traits could mitigate such regressions at least partially.
- New default bounds trigger some compiler bugs in both old and new trait solver.
- With new default bounds we encounter some trait solver cycle errors that break existing code.
- We hope that these cases are bugs that can be addressed in the new trait solver.
Also migration to a new edition could be quite ugly and enormous, but that's actually what we want to solve. For other issues there's a chance that they could be solved by a new solver.
Target modifiers fix for bool flags without value
Fixed support of boolean flags without values: `-Zbool-flag` is now consistent with `-Zbool-flag=true` in another crate.
When flag is explicitly set to default value, target modifier will not be set in crate metainfo (`-Zflag=false` when `false` is a default value for the flag).
Improved error notification when target modifier flag is absent in a crate ("-Zflag unset").
Example:
```
note: `-Zreg-struct-return=true` in this crate is incompatible with unset `-Zreg-struct-return` in dependency `default_reg_struct_return`
```
add FCW to warn about wasm ABI transition
See https://github.com/rust-lang/rust/issues/122532 for context: the "C" ABI on wasm32-unk-unk will change. The goal of this lint is to warn about any function definition and calls whose behavior will be affected by the change. My understanding is the following:
- scalar arguments are fine
- including 128 bit types, they get passed as two `i64` arguments in both ABIs
- `repr(C)` structs (recursively) wrapping a single scalar argument are fine (unless they have extra padding due to over-alignment attributes)
- all return values are fine
`@bjorn3` `@alexcrichton` `@Manishearth` is that correct?
I am making this a "show up in future compat reports" lint to maximize the chances people become aware of this. OTOH this likely means warnings for most users of Diplomat so maybe we shouldn't do this?
IIUC, wasm-bindgen should be unaffected by this lint as they only pass scalar types as arguments.
Tracking issue: https://github.com/rust-lang/rust/issues/138762
Transition plan blog post: https://github.com/rust-lang/blog.rust-lang.org/pull/1531
try-job: dist-various-2
Cache current_dll_path output
Computing the current dll path is somewhat expensive relative to other work when compiling `fn main() {}` as `dladdr` needs to iterate over the symbol table of librustc_driver.so until it finds a match.
Computing the current dll path is somewhat expensive relative to other
work when compiling `fn main() {}` as `dladdr` needs to iterate over the
symbol table of librustc_driver.so until it finds a match.
Represent diagnostic side effects as dep nodes
This changes diagnostic to be tracked as a special dep node (`SideEffect`) instead of having a list of side effects associated with each dep node. `SideEffect` is always red and when forced, it emits the diagnostic and marks itself green. Each emitted diagnostic generates a new `SideEffect` with an unique dep node index.
Some implications of this:
- Diagnostic may now be emitted more than once as they can be emitted once when the `SideEffect` gets marked green and again if the task it depends on needs to be re-executed due to another node being red. It relies on deduplicating of diagnostics to avoid that.
- Anon tasks which emits diagnostics will no longer *incorrectly* be merged with other anon tasks.
- Reusing a CGU will now emit diagnostics from the task generating it.
To correspond to their actual print request names, `target-spec-json`
and `all-target-specs-json`, and for consistency with other print name
<-> print kind mappings.
Revert <https://github.com/rust-lang/rust/pull/138084> to buy time to
consider options that avoids breaking downstream usages of cargo on
distributed `rustc-src` artifacts, where such cargo invocations fail due
to inability to inherit `lints` from workspace root manifest's
`workspace.lints` (this is only valid for the source rust-lang/rust
workspace, but not really the distributed `rustc-src` artifacts).
This breakage was reported in
<https://github.com/rust-lang/rust/issues/138304>.
This reverts commit 48caf81484, reversing
changes made to c6662879b2.
compiler: Use `size_of` from the prelude instead of imported
Use `std::mem::{size_of, size_of_val, align_of, align_of_val}` from the prelude instead of importing or qualifying them. Apply this change across the compiler.
These functions were added to all preludes in Rust 1.80.
r? ``@compiler-errors``
Use `std::mem::{size_of, size_of_val, align_of, align_of_val}` from the
prelude instead of importing or qualifying them.
These functions were added to all preludes in Rust 1.80.
`-Zteach` is perma-unstable, barely used, the highlighting logic buggy and the flag being passed around is tech-debt. We should likely remove `-Zteach` in its entirely.
Use `default_field_values` for `rustc_errors::Context`, `rustc_session::config::NextSolverConfig` and `rustc_session::config::ErrorOutputType`
Wanted to see where `#![feature(default_field_values)]` could be used in the codebase. These three seemed like no-brainers. There are a bunch of more places where we could remove manual `Default` impls, but they `derive` other traits that rely on `syn`, which [doesn't yet support `default_field_values`](https://github.com/dtolnay/syn/issues/1774).
Remove unused `PpMode::needs_hir`
This method was added in #99360 to avoid an overzealous `span_delayed_bug` ICE in specific circumstances, but the only caller was subsequently removed in #136603, which presumably avoids the problem in a more principled way.
Support raw-dylib link kind on ELF
raw-dylib is a link kind that allows rustc to link against a library without having any library files present.
This currently only exists on Windows. rustc will take all the symbols from raw-dylib link blocks and put them in an import library, where they can then be resolved by the linker.
While import libraries don't exist on ELF, it would still be convenient to have this same functionality. Not having the libraries present at build-time can be convenient for several reasons, especially cross-compilation. With raw-dylib, code linking against a library can be cross-compiled without needing to have these libraries available on the build machine. If the libc crate makes use of this, it would allow cross-compilation without having any libc available on the build machine. This is not yet possible with this implementation, at least against libc's like glibc that use symbol versioning. The raw-dylib kind could be extended with support for symbol versioning in the future.
This implementation is very experimental and I have not tested it very well. I have tested it for a toy example and the lz4-sys crate, where it was able to successfully link a binary despite not having a corresponding library at build-time.
I was inspired by Björn's comments in https://internals.rust-lang.org/t/bundle-zig-cc-in-rustup-by-default/22096/27
Tracking issue: #135694
r? bjorn3
try-job: aarch64-apple
try-job: x86_64-msvc-1
try-job: x86_64-msvc-2
try-job: test-various
The embedded bitcode should always be prepared for LTO/ThinLTO
Fixes#115344. Fixes#117220.
There are currently two methods for generating bitcode that used for LTO. One method involves using `-C linker-plugin-lto` to emit object files as bitcode, which is the typical setting used by cargo. The other method is through `-C embed-bitcode=yes`.
When using with `-C embed-bitcode=yes -C lto=no`, we run a complete non-LTO LLVM pipeline to obtain bitcode, then the bitcode is used for LTO. We run the Call Graph Profile Pass twice on the same module.
This PR is doing something similar to LLVM's `buildFatLTODefaultPipeline`, obtaining the bitcode for embedding after running `buildThinLTOPreLinkDefaultPipeline`.
r? nikic
raw-dylib is a link kind that allows rustc to link against a library
without having any library files present.
This currently only exists on Windows. rustc will take all the symbols
from raw-dylib link blocks and put them in an import library, where they
can then be resolved by the linker.
While import libraries don't exist on ELF, it would still be convenient
to have this same functionality. Not having the libraries present at
build-time can be convenient for several reasons, especially
cross-compilation. With raw-dylib, code linking against a library can be
cross-compiled without needing to have these libraries available on the
build machine. If the libc crate makes use of this, it would allow
cross-compilation without having any libc available on the build
machine. This is not yet possible with this implementation, at least
against libc's like glibc that use symbol versioning.
The raw-dylib kind could be extended with support for symbol versioning
in the future.
This implementation is very experimental and I have not tested it very
well. I have tested it for a toy example and the lz4-sys crate, where it
was able to successfully link a binary despite not having a
corresponding library at build-time.
Make it so that every structured error annotated with `#[derive(Diagnostic)]` that has a field of type `Ty<'_>`, the printing of that value into a `String` will look at the thread-local storage `TyCtxt` in order to shorten to a length appropriate with the terminal width. When this happen, the resulting error will have a note with the file where the full type name was written to.
```
error[E0618]: expected function, found `((..., ..., ..., ...), ..., ..., ...)``
--> long.rs:7:5
|
6 | fn foo(x: D) { //~ `x` has type `(...
| - `x` has type `((..., ..., ..., ...), ..., ..., ...)`
7 | x(); //~ ERROR expected function, found `(...
| ^--
| |
| call expression requires function
|
= note: the full name for the type has been written to 'long.long-type-14182675702747116984.txt'
= note: consider using `--verbose` to print the full type name to the console
```
Use StableHasher + Hash64 for dep_tracking_hash
This is similar to https://github.com/rust-lang/rust/pull/137095. We currently have a +/- 1 byte jitter in the size of dep graphs reported on perf.rust-lang.org. I think this fixes that jitter.
When I introduced `Hash64`, I wired it through most of the compiler by making it an output of `StableHasher::finalize` then fixing the compile errors. I missed this case because the `u64` hash in this function is being produced by `DefaultHasher` instead. That seems pretty sketchy because the code seems confident that the hash needs to be stable, and we have a mechanism for stable hashing that we weren't using here.
Rollup of 7 pull requests
Successful merges:
- #137095 (Replace some u64 hashes with Hash64)
- #137100 (HIR analysis: Remove unnecessary abstraction over list of clauses)
- #137105 (Restrict DerefPure for Cow<T> impl to T = impl Clone, [impl Clone], str.)
- #137120 (Enable `relative-path-include-bytes-132203` rustdoc-ui test on Windows)
- #137125 (Re-add missing empty lines in the releases notes)
- #137145 (use add-core-stubs / minicore for a few more tests)
- #137149 (Remove SSE ABI from i586-pc-windows-msvc)
r? `@ghost`
`@rustbot` modify labels: rollup
Replace some u64 hashes with Hash64
I introduced the Hash64 and Hash128 types in https://github.com/rust-lang/rust/pull/110083, essentially as a mechanism to prevent hashes from landing in our leb128 encoding paths. If you just have a u64 or u128 field in a struct then derive Encodable/Decodable, that number gets leb128 encoding. So if you need to store a hash or some other value which behaves very close to a hash, don't store it as a u64.
This reverts part of https://github.com/rust-lang/rust/pull/117603, which turned an encoded Hash64 into a u64.
Based on https://github.com/rust-lang/rust/pull/110083, I don't expect this to be perf-sensitive on its own, though I expect that it may help stabilize some of the small rmeta size fluctuations we currently see in perf reports.
Overhaul `rustc_middle::limits`
In particular, to make `pattern_complexity` work more like other limits, which then enables some other simplifications.
r? ``@Nadrieril``
It's similar to the other limits, e.g. obtained via `get_limit`. So it
makes sense to handle it consistently with the other limits. We now use
`Limit`/`usize` in most places instead of `Option<usize>`, so we use
`Limit::new(usize::MAX)`/`usize::MAX` to emulate how `None` used to work.
The commit also adds `Limit::unlimited`.
Load all builtin targets at once instead of one by one in check-cfg
This PR adds a method on `rustc_target::Target` to load all the builtin targets at once, and then uses that method when constructing the `target_*` values in check-cfg instead of load loading each target one by one by their name, which requires a lookup and was more of a hack anyway.
This may give us some performance improvements as we won't need to do the lookup for the _currently_ 287 targets we have.
DWARF 1 is very different than DWARF 2+ (see the commentary in
https://gcc.gnu.org/onlinedocs/gcc/Debugging-Options.html#index-gdwarf)
and LLVM does not really seem to support DWARF 1 as Clang does not offer
a `-gdwarf-1` flag and `llc` will just generate DWARF 2 with the version
set to 1: https://godbolt.org/z/s85d87n3a.
Since this isn't actually supported (and it's not clear it would be
useful anyway), report that DWARF 1 is not supported if it is requested.
Also add a help message to the error saying which versions are supported.
tree-wide: parallel: Fully removed all `Lrc`, replaced with `Arc`
tree-wide: parallel: Fully removed all `Lrc`, replaced with `Arc`
This is continuation of https://github.com/rust-lang/rust/pull/132282 .
I'm pretty sure I did everything right. In particular, I searched all occurrences of `Lrc` in submodules and made sure that they don't need replacement.
There are other possibilities, through.
We can define `enum Lrc<T> { Rc(Rc<T>), Arc(Arc<T>) }`. Or we can make `Lrc` a union and on every clone we can read from special thread-local variable. Or we can add a generic parameter to `Lrc` and, yes, this parameter will be everywhere across all codebase.
So, if you think we should take some alternative approach, then don't merge this PR. But if it is decided to stick with `Arc`, then, please, merge.
cc "Parallel Rustc Front-end" ( https://github.com/rust-lang/rust/issues/113349 )
r? SparrowLii
`@rustbot` label WG-compiler-parallel
The extended syntax for function signature that includes contract clauses
should never be user exposed versus the interface we want to ship
externally eventually.
Target modifiers (special marked options) are recorded in metainfo
Target modifiers (special marked options) are recorded in metainfo and compared to be equal in different linked crates.
PR for this RFC: https://github.com/rust-lang/rfcs/pull/3716
Option may be marked as `TARGET_MODIFIER`, example: `regparm: Option<u32> = (None, parse_opt_number, [TRACKED TARGET_MODIFIER]`.
If an TARGET_MODIFIER-marked option has non-default value, it will be recorded in crate metainfo as a `Vec<TargetModifier>`:
```
pub struct TargetModifier {
pub opt: OptionsTargetModifiers,
pub value_name: String,
}
```
OptionsTargetModifiers is a macro-generated enum.
Option value code (for comparison) is generated using `Debug` trait.
Error example:
```
error: mixing `-Zregparm` will cause an ABI mismatch in crate `incompatible_regparm`
--> $DIR/incompatible_regparm.rs:10:1
|
LL | #![crate_type = "lib"]
| ^
|
= help: the `-Zregparm` flag modifies the ABI so Rust crates compiled with different values of this flag cannot be used together safely
= note: `-Zregparm=1` in this crate is incompatible with `-Zregparm=2` in dependency `wrong_regparm`
= help: set `-Zregparm=2` in this crate or `-Zregparm=1` in `wrong_regparm`
= help: if you are sure this will not cause problems, use `-Cunsafe-allow-abi-mismatch=regparm` to silence this error
error: aborting due to 1 previous error
```
`-Cunsafe-allow-abi-mismatch=regparm,reg-struct-return` to disable list of flags.
Autodiff Upstreaming - rustc_codegen_ssa, rustc_middle
This PR should not be merged until the rustc_codegen_llvm part is merged.
I will also alter it a little based on what get's shaved off from the cg_llvm PR,
and address some of the feedback I received in the other PR (including cleanups).
I am putting it already up to
1) Discuss with `@jieyouxu` if there is more work needed to add tests to this and
2) Pray that there is someone reviewing who can tell me why some of my autodiff invocations get lost.
Re 1: My test require fat-lto. I also modify the compilation pipeline. So if there are any other llvm-ir tests in the same compilation unit then I will likely break them. Luckily there are two groups who currently have the same fat-lto requirement for their GPU code which I have for my autodiff code and both groups have some plans to enable support for thin-lto. Once either that work pans out, I'll copy it over for this feature. I will also work on not changing the optimization pipeline for functions not differentiated, but that will require some thoughts and engineering, so I think it would be good to be able to run the autodiff tests isolated from the rest for now. Can you guide me here please?
For context, here are some of my tests in the samples folder: https://github.com/EnzymeAD/rustbook
Re 2: This is a pretty serious issue, since it effectively prevents publishing libraries making use of autodiff: https://github.com/EnzymeAD/rust/issues/173. For some reason my dummy code persists till the end, so the code which calls autodiff, deletes the dummy, and inserts the code to compute the derivative never gets executed. To me it looks like the rustc_autodiff attribute just get's dropped, but I don't know WHY? Any help would be super appreciated, as rustc queries look a bit voodoo to me.
Tracking:
- https://github.com/rust-lang/rust/issues/124509
r? `@jieyouxu`
Fix deduplication mismatches in vtables leading to upcasting unsoundness
We currently have two cases where subtleties in supertraits can trigger disagreements in the vtable layout, e.g. leading to a different vtable layout being accessed at a callsite compared to what was prepared during unsizing. Namely:
### #135315
In this example, we were not normalizing supertraits when preparing vtables. In the example,
```
trait Supertrait<T> {
fn _print_numbers(&self, mem: &[usize; 100]) {
println!("{mem:?}");
}
}
impl<T> Supertrait<T> for () {}
trait Identity {
type Selff;
}
impl<Selff> Identity for Selff {
type Selff = Selff;
}
trait Middle<T>: Supertrait<()> + Supertrait<T> {
fn say_hello(&self, _: &usize) {
println!("Hello!");
}
}
impl<T> Middle<T> for () {}
trait Trait: Middle<<() as Identity>::Selff> {}
impl Trait for () {}
fn main() {
(&() as &dyn Trait as &dyn Middle<()>).say_hello(&0);
}
```
When we prepare `dyn Trait`, we see a supertrait of `Middle<<() as Identity>::Selff>`, which itself has two supertraits `Supertrait<()>` and `Supertrait<<() as Identity>::Selff>`. These two supertraits are identical, but they are not duplicated because we were using structural equality and *not* considering normalization. This leads to a vtable layout with two trait pointers.
When we upcast to `dyn Middle<()>`, those two supertraits are now the same, leading to a vtable layout with only one trait pointer. This leads to an offset error, and we call the wrong method.
### #135316
This one is a bit more interesting, and is the bulk of the changes in this PR. It's a bit similar, except it uses binder equality instead of normalization to make the compiler get confused about two vtable layouts. In the example,
```
trait Supertrait<T> {
fn _print_numbers(&self, mem: &[usize; 100]) {
println!("{mem:?}");
}
}
impl<T> Supertrait<T> for () {}
trait Trait<T, U>: Supertrait<T> + Supertrait<U> {
fn say_hello(&self, _: &usize) {
println!("Hello!");
}
}
impl<T, U> Trait<T, U> for () {}
fn main() {
(&() as &'static dyn for<'a> Trait<&'static (), &'a ()>
as &'static dyn Trait<&'static (), &'static ()>)
.say_hello(&0);
}
```
When we prepare the vtable for `dyn for<'a> Trait<&'static (), &'a ()>`, we currently consider the PolyTraitRef of the vtable as the key for a supertrait. This leads two two supertraits -- `Supertrait<&'static ()>` and `for<'a> Supertrait<&'a ()>`.
However, we can upcast[^up] without offsetting the vtable from `dyn for<'a> Trait<&'static (), &'a ()>` to `dyn Trait<&'static (), &'static ()>`. This is just instantiating the principal trait ref for a specific `'a = 'static`. However, when considering those supertraits, we now have only one distinct supertrait -- `Supertrait<&'static ()>` (which is deduplicated since there are two supertraits with the same substitutions). This leads to similar offsetting issues, leading to the wrong method being called.
[^up]: I say upcast but this is a cast that is allowed on stable, since it's not changing the vtable at all, just instantiating the binder of the principal trait ref for some lifetime.
The solution here is to recognize that a vtable isn't really meaningfully higher ranked, and to just treat a vtable as corresponding to a `TraitRef` so we can do this deduplication more faithfully. That is to say, the vtable for `dyn for<'a> Tr<'a>` and `dyn Tr<'x>` are always identical, since they both would correspond to a set of free regions on an impl... Do note that `Tr<for<'a> fn(&'a ())>` and `Tr<fn(&'static ())>` are still distinct.
----
There's a bit more that can be cleaned up. In codegen, we can stop using `PolyExistentialTraitRef` basically everywhere. We can also fix SMIR to stop storing `PolyExistentialTraitRef` in its vtable allocations.
As for testing, it's difficult to actually turn this into something that can be tested with `rustc_dump_vtable`, since having multiple supertraits that are identical is a recipe for ambiguity errors. Maybe someone else is more creative with getting that attr to work, since the tests I added being run-pass tests is a bit unsatisfying. Miri also doesn't help here, since it doesn't really generate vtables that are offset by an index in the same way as codegen.
r? `@lcnr` for the vibe check? Or reassign, idk. Maybe let's talk about whether this makes sense.
<sup>(I guess an alternative would also be to not do any deduplication of vtable supertraits (or only a really conservative subset) rather than trying to normalize and deduplicate more faithfully here. Not sure if that works and is sufficient tho.)</sup>
cc `@steffahn` -- ty for the minimizations
cc `@WaffleLapkin` -- since you're overseeing the feature stabilization :3
Fixes#135315Fixes#135316
