inherit `#[align]` from trait method prototypes
````@workingjubilee```` this seems straightforward enough. Now that we're planning to make `-Cmin-function-alignment` a target modifier, I don't think there are any cross-crate complications here?
````@Jules-Bertholet```` is this the behavior you had in mind? In particular the inheritance of the attribute of a default impl is maybe a bit unintuitive at first? (but I think it's ok if that behavior is explicitly documented).
r? ghost
fix `-Zmin-function-alignment` on functions without attributes
tracking issue: https://github.com/rust-lang/rust/issues/82232
related: https://github.com/rust-lang/rust/pull/142854
The minimum function alignment was skipped on functions without attributes (because the logic was in a loop that only runs if there is at least one attribute). The underlying reason we didn't catch this before is that in our testing we generally apply `#[no_mangle]` to functions that are tested. I've added a test now that deliberately has no attributes.
r? `@workingjubilee`
the minimum function alignment was skipped on functions without attributes. That is because in our testing we generally apply `#[no_mangle]` to functions that are tested. I've added a test now that deliberately has no attributes
centralize `-Zmin-function-alignment` logic
tracking issue: https://github.com/rust-lang/rust/issues/82232
discussed in: https://github.com/rust-lang/rust/pull/142824#discussion_r2160056244
Apply the `-Zmin-function-alignment` value to the alignment field of the function attributes when those are created, so that individual backends don't need to consider it.
The one exception right now is cranelift, because it can't yet set the alignment for individual functions, but it can (and does) set the global minimum function alignment.
cc ``@RalfJung`` I think this is an improvement regardless, is there anything else that should be done for miri?
Extract some shared code from codegen backend target feature handling
There's a bunch of code duplication between the GCC and LLVM backends in target feature handling. This moves that into new shared helper functions in `rustc_codegen_ssa`.
The first two commits should be purely refactoring. I am fairly sure the LLVM-side behavior stays the same; if the GCC side deliberately diverges from this then I may have missed that. I did account for one divergence, which I do not know is deliberate or not: GCC does not seem to use the `-Ctarget-feature` flag to populate `cfg(target_feature)`. That seems odd, since the `-Ctarget-feature` flag is used to populate the return value of `global_gcc_features` which controls the target features actually used by GCC. ``@GuillaumeGomez`` ``@antoyo`` is there a reason `target_config` ignores `-Ctarget-feature` but `global_gcc_features` does not? The second commit also cleans up a bunch of unneeded complexity added in https://github.com/rust-lang/rust/pull/135927.
The third commit extracts some shared logic out of the functions that populate `cfg(target_feature)` and the backend target feature set, respectively. This one actually has some slight functional changes:
- Before, with `-Ctarget-feature=-feat`, if there is some other feature `x` that implies `feat` we would *not* add `-x` to the backend target feature set. Now, we do. This fixesrust-lang/rust#134792.
- The logic that removes `x` from `cfg(target_feature)` in this case also changed a bit, avoiding a large number of calls to the (uncached) `sess.target.implied_target_features` (if there were a large number of positive features listed before a negative feature) but instead constructing a full inverse implication map when encountering the first negative feature. Ideally this would be done with queries but the backend target feature logic runs before `tcx` so we can't use that...
- Previously, if feature "a" implied "b" and "b" was unstable, then using `-Ctarget-feature=+a` would also emit a warning about `b`. I had to remove this since when accounting for negative implications, this emits a ton of warnings in a bunch of existing tests... I assume this was unintentional anyway.
The fourth commit increases consistency of the GCC backend with the LLVM backend.
The last commit does some further cleanup:
- Get rid of RUSTC_SPECIAL_FEATURES. It was only needed for s390x "backchain", but since LLVM 19 that is always a regular target feature so we don't need this hack any more. The hack also has various unintended side-effects so we don't want to keep it. Fixes https://github.com/rust-lang/rust/issues/142412.
- Move RUSTC_SPECIFIC_FEATURES handling into the shared parse_rust_feature_flag helper so all consumers of `-Ctarget-feature` that only care about actual target features (and not "crt-static") have it. Previously, we actually set `cfg(target_feature = "crt-static")` twice: once in the backend target feature logic, and once specifically for that one feature. IIUC, some targets are meant to ignore `-Ctarget-feature=+crt-static`, it seems like before this PR that flag still incorrectly enabled `cfg(target_feature = "crt-static")` (but I didn't test this).
- Move fixed_x18 handling together with retpoline handling.
- Forbid setting fixed_x18 as a regular target feature, even unstably. It must be set via the `-Z` flag.
``@bjorn3`` I did not touch the cranelift backend here, since AFAIK it doesn't really support target features. But if you ever do, please use the new helpers. :)
Cc ``@workingjubilee``
rewrite `optimize` attribute to use new attribute parsing infrastructure
r? ```@oli-obk```
I'm afraid we'll get quite a few of these PRs in the future. If we get a lot of trivial changes I'll start merging multiple into one PR. They should be easy to review :)
Waiting on #138165 first
#[used] currently is an alias for #[used(linker)] on all platforms
except ELF based ones where it is an alias for #[used(compiler)]. The
latter has surprising behavior and the LLVM LangRef explicitly states
that it "should only be used in rare circumstances, and should not be
exposed to source languages."
The reason #[used] still was an alias to #[used(compiler)] on ELF is
because the gold linker has issues with it. Luckily gold has been
deprecated with GCC 15 and seems to be unable to bootstrap rustc anyway.
As such we shouldn't really care about supporting gold.
- Rename `USED` to `USED_COMPILER` to better reflect its behavior.
- Reorder some items to group the used and allocator flags together
- Renumber them without gaps
make `rustc_attr_parsing` less dominant in the rustc crate graph
It has/had a glob re-export of `rustc_attr_data_structures`, which is a crate much lower in the graph, and a lot of crates were using it *just* (or *mostly*) for that re-export, while they can rely on `rustc_attr_data_structures` directly.
Previous graph:
Graph with this PR:
The first commit keeps the re-export, and just changes the dependency if possible. The second commit is the "breaking change" which removes the re-export, and "explicitly" adds the `rustc_attr_data_structures` dependency where needed. It also switches over some src/tools/*.
The second commit is actually a lot more involved than I expected. Please let me know if it's a better idea to back it out and just keep the first commit.
add suggestion on how to add a panic breakpoint
Co-authored-by: Pat Pannuto <pat.pannuto@gmail.com>
delete no_mangle from ui/panic-handler/panic-handler-wrong-location test
issue an error for the usage of #[no_mangle] on internal language items
delete the comments
add newline
rephrase note
Co-authored-by: bjorn3 <17426603+bjorn3@users.noreply.github.com>
update error not to leak implementation details
delete no_mangle_span
Co-authored-by: bjorn3 <17426603+bjorn3@users.noreply.github.com>
delete commented code
I'm removing empty identifiers everywhere, because in practice they
always mean "no identifier" rather than "empty identifier". (An empty
identifier is impossible.) It's better to use `Option` to mean "no
identifier" because you then can't forget about the "no identifier"
possibility.
Some specifics:
- When testing an attribute for a single name, the commit uses the
`has_name` method.
- When testing an attribute for multiple names, the commit uses the new
`has_any_name` method.
- When using `match` on an attribute, the match arms now have `Some` on
them.
In the tests, we now avoid printing empty identifiers by not printing
the identifier in the `error:` line at all, instead letting the carets
point out the problem.
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
Rename `is_like_osx` to `is_like_darwin`
Replace `is_like_osx` with `is_like_darwin`, which more closely describes reality (OS X is the pre-2016 name for macOS, and is by now quite outdated; Darwin is the overall name for the OS underlying Apple's macOS, iOS, etc.).
``@rustbot`` label O-apple
r? compiler
Continuing the work from #137350.
Removes the unused methods: `expect_variant`, `expect_field`,
`expect_foreign_item`.
Every method gains a `hir_` prefix.
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``
