| Age | Commit message (Collapse) | Author | Lines |
|---|
|
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
|
|
|
|
|
|
`ast::Item` has an `ident` field.
- It's always non-empty for these item kinds: `ExternCrate`, `Static`,
`Const`, `Fn`, `Mod`, `TyAlias`, `Enum`, `Struct`, `Union`,
`Trait`, `TraitAlias`, `MacroDef`, `Delegation`.
- It's always empty for these item kinds: `Use`, `ForeignMod`,
`GlobalAsm`, `Impl`, `MacCall`, `DelegationMac`.
There is a similar story for `AssocItemKind` and `ForeignItemKind`.
Some sites that handle items check for an empty ident, some don't. This
is a very C-like way of doing things, but this is Rust, we have sum
types, we can do this properly and never forget to check for the
exceptional case and never YOLO possibly empty identifiers (or possibly
dummy spans) around and hope that things will work out.
The commit is large but it's mostly obvious plumbing work. Some notable
things.
- `ast::Item` got 8 bytes bigger. This could be avoided by boxing the
fields within some of the `ast::ItemKind` variants (specifically:
`Struct`, `Union`, `Enum`). I might do that in a follow-up; this
commit is big enough already.
- For the visitors: `FnKind` no longer needs an `ident` field because
the `Fn` within how has one.
- In the parser, the `ItemInfo` typedef is no longer needed. It was used
in various places to return an `Ident` alongside an `ItemKind`, but
now the `Ident` (if present) is within the `ItemKind`.
- In a few places I renamed identifier variables called `name` (or
`foo_name`) as `ident` (or `foo_ident`), to better match the type, and
because `name` is normally used for `Symbol`s. It's confusing to see
something like `foo_name.name`.
|
|
`expand_test_case` looks for any item with a `#[test_case]` attribute
and adds a `test_path_symbol` attribute to it while also fiddling with
the item's ident's span.
This is pretty weird, because `#[test_case]` is only valid on
`fn`/`const`/`static` items, as far as I can tell. But you don't
currently get an error or warning if you use it on other kinds of items.
This commit changes things so that a `#[test_case]` item is modified
only if it is `fn`/`const`/`static`. This is relevant for moving idents
from `Item` to `ItemKind`, because some item kinds don't have an ident,
e.g. `impl` blocks.
The commit also does the following.
- Renames a local variable `test_id` as `test_ident`.
- Changes a `const` to `static` in
`tests/ui/custom_test_frameworks/full.rs` to give the `static` case
some test coverage.
- Adds a `struct` and `impl` to the same test to give some test coverage
to the non-affected item kinds. These have a `FIXME` comment
identifying the weirdness here. Hopefully this will be useful
breadcrumbs for somebody else in the future.
|
|
|
|
r=compiler-errors
Mostly parser: Eliminate code that's been dead / semi-dead since the removal of type ascription syntax
**Disclaimer**: This PR is intended to mostly clean up code as opposed to bringing about behavioral changes. Therefore it doesn't aim to address any of the 'FIXME: remove after a month [dated: 2023-05-02]: "type ascription syntax has been removed, see issue [#]101728"'.
---
By commit:
1. Removes truly dead code:
* Since 1.71 (#109128) `let _ = { f: x };` is a syntax error as opposed to a semantic error which allows the parse-time diagnostic (suggestion) "*struct literal body without path // you might have forgotten […]*" to kick in.
* The analysis-time diagnostic (suggestion) from <=1.70 "*cannot find value \`f\` in this scope // you might have forgotten […]*" is therefore no longer reachable.
2. Updates `is_certainly_not_a_block` to be in line with the current grammar:
* The seq. `{ ident:` is definitely not the start of a block. Before the removal of ty ascr, `{ ident: ty_start` would begin a block expr.
* This shouldn't make more code compile IINM, it should *ultimately* only affect diagnostics.
* For example, `if T { f: () } {}` will now be interpreted as an `if` with struct lit `T { f: () }` as its *condition* (which is banned in the parser anyway) as opposed to just `T` (with the *consequent* being `f : ()` which is also invalid (since 1.71)). The diagnostics are almost the same because we have two separate parse recovery procedures + diagnostics: `StructLiteralNeedingParens` (*invalid struct lit*) before and `StructLiteralNotAllowedHere` (*struct lits aren't allowed here*) now, as you can see from the diff.
* (As an aside, even before this PR, fn `maybe_suggest_struct_literal` should've just used the much older & clearer `StructLiteralNotAllowedHere`)
* NB: This does sadly regress the compiler output for `tests/ui/parser/type-ascription-in-pattern.rs` but that can be fixed in follow-up PRs. It's not super important IMO and a natural consequence.
3. Removes code that's become dead due to the prior commit.
* Basically reverts #106620 + #112475 (without regressing rustc's output!).
* Now the older & more robust parse recovery procedure (cc `StructLiteralNotAllowedHere`) takes care of the cases the removed code used to handle.
* This automatically fixes the suggestions for \[[playground](https://play.rust-lang.org/?version=stable&mode=debug&edition=2024&gist=7e2030163b11ee96d17adc3325b01780)\]:
* `if Ty::<i32> { f: K }.m() {}`: `if Ty::<i32> { SomeStruct { f: K } }.m() {}` (broken) → ` if (Ty::<i32> { f: K }).m() {}`
* `if <T as Trait>::Out { f: K::<> }.m() {}`: `if <T as Trait>(::Out { f: K::<> }).m() {}` (broken) → `if (<T as Trait>::Out { f: K::<> }).m() {}`
4. Merge and simplify UI tests pertaining to this issue, so it's easier to add more regression tests like for the two cases mentioned above.
5. Merge UI tests and add the two regression tests.
Best reviewed commit by commit (on request I'll partially squash after approval).
|
|
Allow defining opaques in statics and consts
r? oli-obk
Fixes https://github.com/rust-lang/rust/issues/138902
|
|
Visitors track whether an assoc item is in a trait impl or an inherent impl
`AssocCtxt::Impl` now contains an `of_trait` field. This allows ast lowering and nameres to not have to track whether we're in a trait impl or an inherent impl.
|
|
|
|
|
|
|
|
|
|
This is exactly the kind of case `Ident::dummy()` is for.
|
|
Since `{ ident: ident }` is a parse error, these fields are dead.
|
|
Autodiff cleanups
Splitting out some cleanups to reduce the size of my batching PR and simplify ``@haenoe`` 's [PR](https://github.com/rust-lang/rust/pull/138314).
r? ``@oli-obk``
Tracking:
- https://github.com/rust-lang/rust/issues/124509
|
|
This removes E0773 "A builtin-macro was defined more than once."
|
|
|
|
|
|
|
|
|
|
Rollup of 7 pull requests
Successful merges:
- #137314 (change definitely unproductive cycles to error)
- #137701 (Convert `ShardedHashMap` to use `hashbrown::HashTable`)
- #138269 (uefi: fs: Implement FileType, FilePermissions and FileAttr)
- #138331 (Use `RUSTC_LINT_FLAGS` more)
- #138345 (Some autodiff cleanups)
- #138387 (intrinsics: remove unnecessary leading underscore from argument names)
- #138390 (fix incorrect tracing log)
r? `@ghost`
`@rustbot` modify labels: rollup
|
|
r=onur-ozkan,jieyouxu
Use `RUSTC_LINT_FLAGS` more
An alternative to the failed #138084.
Fixes #138106.
r? ````@jieyouxu````
|
|
Remove `NtItem` and `NtStmt`
Another piece of #124141.
r? `@petrochenkov`
|
|
Rollup of 10 pull requests
Successful merges:
- #137715 (Allow int literals for pattern types with int base types)
- #138002 (Disable CFI for weakly linked syscalls)
- #138051 (Add support for downloading GCC from CI)
- #138231 (Prevent ICE in autodiff validation by emitting user-friendly errors)
- #138245 (stabilize `ci_rustc_if_unchanged_logic` test for local environments)
- #138256 (Do not feed anon const a type that references generics that it does not have)
- #138284 (Do not write user type annotation for const param value path)
- #138296 (Remove `AdtFlags::IS_ANONYMOUS` and `Copy`/`Clone` condition for anonymous ADT)
- #138352 (miri native_calls: ensure we actually expose *mutable* provenance to the memory FFI can access)
- #138354 (remove redundant `body` arguments)
r? `@ghost`
`@rustbot` modify labels: rollup
|
|
Prevent ICE in autodiff validation by emitting user-friendly errors
This PR moves `valid_ret_activity` and `valid_input_activity` checks to the macro expansion phase in compiler/rustc_builtin_macros/src/autodiff.rs, replacing the following internal compiler error (ICE):
```
error: internal compiler error:
compiler/rustc_codegen_ssa/src/codegen_attrs.rs:935:13:
Invalid input activity Dual for Reverse mode
```
with a more user-friendly message.
The issue specifically affected the test file `tests/ui/autodiff/autodiff_illegal.rs`, impacting the functions `f5` and `f6`.
The ICE can be reproduced by following [Enzyme's Rustbook](https://enzymead.github.io/rustbook/installation.html) installation guide.
Additionally, this PR adds tests for invalid return activity in `autodiff_illegal.rs`, which previously triggered an unnoticed ICE before these fixes.
r? ``@oli-obk``
|
|
Add `#[define_opaques]` attribute and require it for all type-alias-impl-trait sites that register a hidden type
Instead of relying on the signature of items to decide whether they are constraining an opaque type, the opaque types that the item constrains must be explicitly listed.
A previous version of this PR used an actual attribute, but had to keep the resolved `DefId`s in a side table.
Now we just lower to fields in the AST that have no surface syntax, instead a builtin attribute macro fills in those fields where applicable.
Note that for convenience referencing opaque types in associated types from associated methods on the same impl will not require an attribute. If that causes problems `#[defines()]` can be used to overwrite the default of searching for opaques in the signature.
One wart of this design is that closures and static items do not have generics. So since I stored the opaques in the generics of functions, consts and methods, I would need to add a custom field to closures and statics to track this information. During a T-types discussion we decided to just not do this for now.
fixes #131298
|
|
|
|
|
|
|
|
|
|
It's no longer necessary now that `-Wunreachable_pub` is being passed.
|
|
`bootstrap` and `llvm_enzyme` are now both in the extra `check-cfg` list
in bootstrap, so they doesn't need to be handled explicitly in
`Cargo.toml` files.
|
|
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 48caf81484b50dca5a5cebb614899a3df81ca898, reversing
changes made to c6662879b27f5161e95f39395e3c9513a7b97028.
|
|
Use workspace lints for crates in `compiler/`
This is nicer and hopefully less error prone than specifying lints via bootstrap.
r? ``@jieyouxu``
|
|
remove clones
|
|
add a "future" edition
This idea has been discussed previously [on Zulip](https://rust-lang.zulipchat.com/#narrow/channel/213817-t-lang/topic/Continuous.20edition-like.20changes.3F/near/432559262) (though what I've implemented isn't exactly the "next"/"future" editions proposed in that message, just the "future" edition). I've found myself prototyping changes that involve edition migrations and wanting to target an upcoming edition for those migrations, but none exists. This should be permanently unstable and not removed.
|
|
(Except for `rustc_codegen_cranelift`.)
It's no longer necessary now that `unreachable_pub` is in the workspace
lints.
|
|
By naming them in `[workspace.lints.rust]` in the top-level
`Cargo.toml`, and then making all `compiler/` crates inherit them with
`[lints] workspace = true`. (I omitted `rustc_codegen_{cranelift,gcc}`,
because they're a bit different.)
The advantages of this over the current approach:
- It uses a standard Cargo feature, rather than special handling in
bootstrap. So, easier to understand, and less likely to get
accidentally broken in the future.
- It works for proc macro crates.
It's a shame it doesn't work for rustc-specific lints, as the comments
explain.
|
|
`llvm_enzyme` is now in the extra `check-cfg` list in bootstrap, so it
doesn't need to be handled explicitly here.
|
|
|
|
Ergonomic ref counting
This is an experimental first version of ergonomic ref counting.
This first version implements most of the RFC but doesn't implement any of the optimizations. This was left for following iterations.
RFC: https://github.com/rust-lang/rfcs/pull/3680
Tracking issue: https://github.com/rust-lang/rust/issues/132290
Project goal: https://github.com/rust-lang/rust-project-goals/issues/107
r? ```@nikomatsakis```
|
|
This is temporarily needed for `x doc compiler` to work. They can be
removed once the `Nonterminal` is removed (#124141).
|
|
This involves replacing `nt_pretty_printing_compatibility_hack` with
`stream_pretty_printing_compatibility_hack`.
The handling of statements in `transcribe` is slightly different to
other nonterminal kinds, due to the lack of `from_ast` implementation
for empty statements.
Notable test changes:
- `tests/ui/proc-macro/expand-to-derive.rs`: the diff looks large but
the only difference is the insertion of a single invisible-delimited
group around a metavar.
|
|
|
|
|
|
When `#![feature(min_generic_const_args)]` is enabled, we now lower all
const paths in generic arg position to `hir::ConstArgKind::Path`. We
then lower assoc const paths to `ty::ConstKind::Unevaluated` since we
can no longer use the anon const expression lowering machinery. In the
process of implementing this, I factored out `hir_ty_lowering` code that
is now shared between lowering assoc types and assoc consts.
This PR also introduces a `#[type_const]` attribute for trait assoc
consts that are allowed as const args. However, we still need to
implement code to check that assoc const definitions satisfy
`#[type_const]` if present (basically is it a const path or a
monomorphic anon const).
|
|
|
|
|
|
|
