Add clobber-only register classes for asm!
These are needed to properly express a function call ABI using a clobber
list, even though we don't support passing actual values into/out of
these registers.
Improve opaque pointers support
Opaque pointers are coming, and rustc is not ready.
This adds partial support by passing an explicit load type to LLVM. Two issues I've encountered:
* The necessary type was not available at the point where non-temporal copies were generated. I've pushed the code for that upwards out of the memcpy implementation and moved the position of a cast to make do with the types we have available. (I'm not sure that cast is needed at all, but have retained it in the interest of conservativeness.)
* The `PlaceRef::project_deref()` function used during debuginfo generation seems to be buggy in some way -- though I haven't figured out specifically what it does wrong. Replacing it with `load_operand().deref()` did the trick, but I don't really know what I'm doing here.
These are needed to properly express a function call ABI using a clobber
list, even though we don't support passing actual values into/out of
these registers.
Add support for raw-dylib with stdcall, fastcall functions
Next stage of work for #58713: allow `extern "stdcall"` and `extern "fastcall"` with `#[link(kind = "raw-dylib")]`.
I've deliberately omitted support for vectorcall, as that doesn't currently work, and I wanted to get this out for review. (I haven't really investigated the vectorcall failure much yet, but at first (very cursory) glance it appears that the problem is elsewhere.)
This makes load generation compatible with opaque pointers.
The generation of nontemporal copies still accesses the pointer
element type, as fixing this requires more movement.
Refactor linker code
This merges `LinkerInfo` into `CrateInfo` as there is no reason to keep them separate. `LinkerInfo::to_linker` is merged into `get_linker` as both have different logic for each linker type and `to_linker` is directly called after `get_linker`. Also contains a couple of small cleanups.
See the individual commits for all changes.
Previously, only the fields would be displayed with no indication of the
variant name. If you already knew the enum was univariant, this was ok
but if the enum was univariant because of layout, for example, a
`Result<T, !>` then it could be very confusing which variant was the
active one.
Previously, directly tagged enums had a `variant$` field which would
show the name of the active variant. We now show the variant using a
`[variant]` synthetic item just like we do for niche-layout enums.
Improve debug symbol names to avoid ambiguity and work better with MSVC's debugger
There are several cases where names of types and functions in the debug info are either ambiguous, or not helpful, such as including ambiguous placeholders (e.g., `{{impl}}`, `{{closure}}` or `dyn _'`) or dropping qualifications (e.g., for dynamic types).
Instead, each debug symbol name should be unique and useful:
* Include disambiguators for anonymous `DefPathDataName` (closures and generators), and unify their formatting when used as a path-qualifier vs item being qualified.
* Qualify the principal trait for dynamic types.
* If there is no principal trait for a dynamic type, emit all other traits instead.
* Respect the `qualified` argument when emitting ref and pointer types.
* For implementations, emit the disambiguator.
* Print const generics when emitting generic parameters or arguments.
Additionally, when targeting MSVC, its debugger treats many command arguments as C++ expressions, even when the argument is defined to be a symbol name. As such names in the debug info need to be more C++-like to be parsed correctly:
* Avoid characters with special meaning (`#`, `[`, `"`, `+`).
* Never start a name with `<` or `{` as this is treated as an operator.
* `>>` is always treated as a right-shift, even when parsing generic arguments (so add a space to avoid this).
* Emit function declarations using C/C++ style syntax (e.g., leading return type).
* Emit arrays as a synthetic `array$<type, size>` type.
* Include a `$` in all synthetic types as this is a legal character for C++, but not Rust (thus we avoid collisions with user types).
Add support for leaf function frame pointer elimination
This PR adds ability for the target specifications to specify frame
pointer emission type that's not just “always” or “whatever cg decides”.
In particular there's a new mode that allows omission of the frame
pointer for leaf functions (those that don't call any other functions).
We then set this new mode for Aarch64-based Apple targets.
Fixes#86196
Remove unused dependencies from compiler crates
Various compiler crates have dependencies that they don't appear to use. I used some scripting to detect such dependencies, filtered them based on some manual review, and removed those that do indeed appear to be entirely unused.
There are several cases where names of types and functions in the debug info are either ambiguous, or not helpful, such as including ambiguous placeholders (e.g., `{{impl}}`, `{{closure}}` or `dyn _'`) or dropping qualifications (e.g., for dynamic types).
Instead, each debug symbol name should be unique and useful:
* Include disambiguators for anonymous `DefPathDataName` (closures and generators), and unify their formatting when used as a path-qualifier vs item being qualified.
* Qualify the principal trait for dynamic types.
* If there is no principal trait for a dynamic type, emit all other traits instead.
* Respect the `qualified` argument when emitting ref and pointer types.
* For implementations, emit the disambiguator.
* Print const generics when emitting generic parameters or arguments.
Additionally, when targeting MSVC, its debugger treats many command arguments as C++ expressions, even when the argument is defined to be a symbol name. As such names in the debug info need to be more C++-like to be parsed correctly:
* Avoid characters with special meaning (`#`, `[`, `"`, `+`).
* Never start a name with `<` or `{` as this is treated as an operator.
* `>>` is always treated as a right-shift, even when parsing generic arguments (so add a space to avoid this).
* Emit function declarations using C/C++ style syntax (e.g., leading return type).
* Emit arrays as a synthetic `array$<type, size>` type.
* Include a `$` in all synthetic types as this is a legal character for C++, but not Rust (thus we avoid collisions with user types).
This PR adds ability for the target specifications to specify frame
pointer emission type that's not just “always” or “whatever cg decides”.
In particular there's a new mode that allows omission of the frame
pointer for leaf functions (those that don't call any other functions).
We then set this new mode for Aarch64-based Apple targets.
Fixes#86196
Allow loading of llvm plugins on nightly
Based on a discussion in #82734 / with `@wsmoses.`
Mainly moves [this](0149bc4e7e) behind a -Z flag, so it can only be used on nightly,
as requested by `@nagisa` in https://github.com/rust-lang/rust/issues/82734#issuecomment-835863940
This change allows loading of llvm plugins like Enzyme.
Right now it also requires a shared library LLVM build of rustc for symbol resolution.
```rust
// test.rs
extern { fn __enzyme_autodiff(_: usize, ...) -> f64; }
fn square(x : f64) -> f64 {
return x * x;
}
fn main() {
unsafe {
println!("Hello, world {} {}!", square(3.0), __enzyme_autodiff(square as usize, 3.0));
}
}
```
```
./rustc test.rs -Z llvm-plugins="./LLVMEnzyme-12.so" -C passes="enzyme"
./test
Hello, world 9 6!
```
I will try to figure out how to simplify the usage and get this into stable in a later iteration,
but having this on nightly will already help testing further steps.
Disable the machine outliner by default
This addresses a codegen-issue that needs to be fixed upstream in LLVM.
While we wait for the fix, we can disable it.
Verified manually that the outliner is no longer run when
`-Copt-level=z` is specified, and also that you can override this with
`-Cllvm-args=-enable-machine-outliner` if you need it anyway.
A regression test is not really feasible in this instance, given that we
do not have any minimal reproducers.
Fixes#85351
cc `@pnkfelix`
Driver improvements
This PR contains a couple of cleanups for the driver and a few small improvements for the custom codegen backend interface. It also implements `--version` and `-Cpasses=list` support for custom codegen backends.
Partial support for raw-dylib linkage
First cut of functionality for issue #58713: add support for `#[link(kind = "raw-dylib")]` on `extern` blocks in lib crates compiled to .rlib files. Does not yet support `#[link_name]` attributes on functions, or the `#[link_ordinal]` attribute, or `#[link(kind = "raw-dylib")]` on `extern` blocks in bin crates; I intend to publish subsequent PRs to fill those gaps. It's also not yet clear whether this works for functions in `extern "stdcall"` blocks; I also intend to investigate that shortly and make any necessary changes as a follow-on PR.
This implementation calls out to an LLVM function to construct the actual `.idata` sections as temporary `.lib` files on disk and then links those into the generated .rlib.
BPF target support
This adds `bpfel-unknown-none` and `bpfeb-unknown-none`, two new no_std targets that generate little and big endian BPF. The approach taken is very similar to the cuda target, where `TargetOptions::obj_is_bitcode` is enabled and code generation is done by the linker.
I added the targets to `dist-various-2`. There are [some tests](https://github.com/alessandrod/bpf-linker/tree/main/tests/assembly) in bpf-linker and I'm planning to add more. Those are currently not ran as part of rust CI.
This addresses a codegen-issue that needs to be fixed upstream in LLVM.
While we wait for the fix, we can disable it.
Verified manually that the outliner is no longer run when
`-Copt-level=z` is specified, and also that you can override this with
`-Cllvm-args=-enable-machine-outliner` if you need it anyway.
A regression test is not really feasible in this instance, given that we
do not have any minimal reproducers.
Fixes#85351
This does not yet support #[link_name] attributes on functions, the #[link_ordinal]
attribute, #[link(kind = "raw-dylib")] on extern blocks in bin crates, or
stdcall functions on 32-bit x86.
rustc: Store metadata-in-rlibs in object files
This commit updates how rustc compiler metadata is stored in rlibs.
Previously metadata was stored as a raw file that has the same format as
`--emit metadata`. After this commit, however, the metadata is encoded
into a small object file which has one section which is the contents of
the metadata.
The motivation for this commit is to fix a common case where #83730
arises. The problem is that when rustc crates a `dylib` crate type it
needs to include entire rlib files into the dylib, so it passes
`--whole-archive` (or the equivalent) to the linker. The problem with
this, though, is that the linker will attempt to read all files in the
archive. If the metadata file were left as-is (today) then the linker
would generate an error saying it can't read the file. The previous
solution was to alter the rlib just before linking, creating a new
archive in a temporary directory which has the metadata file removed.
This problem from before this commit is now removed if the metadata file
is stored in an object file that the linker can read. The only caveat we
have to take care of is to ensure that the linker never actually
includes the contents of the object file into the final output. We apply
similar tricks as the `.llvmbc` bytecode sections to do this.
This involved changing the metadata loading code a bit, namely updating
some of the LLVM C APIs used to use non-deprecated ones and fiddling
with the lifetimes a bit to get everything to work out. Otherwise though
this isn't intended to be a functional change really, only that metadata
is stored differently in archives now.
This should end up fixing #83730 because by default dylibs will no
longer have their rlib dependencies "altered" meaning that
split-debuginfo will continue to have valid paths pointing at the
original rlibs. (note that we still "alter" rlibs if LTO is enabled to
remove Rust object files and we also "alter" for the #[link(cfg)]
feature, but that's rarely used).
Closes#83730
This commit updates how rustc compiler metadata is stored in rlibs.
Previously metadata was stored as a raw file that has the same format as
`--emit metadata`. After this commit, however, the metadata is encoded
into a small object file which has one section which is the contents of
the metadata.
The motivation for this commit is to fix a common case where #83730
arises. The problem is that when rustc crates a `dylib` crate type it
needs to include entire rlib files into the dylib, so it passes
`--whole-archive` (or the equivalent) to the linker. The problem with
this, though, is that the linker will attempt to read all files in the
archive. If the metadata file were left as-is (today) then the linker
would generate an error saying it can't read the file. The previous
solution was to alter the rlib just before linking, creating a new
archive in a temporary directory which has the metadata file removed.
This problem from before this commit is now removed if the metadata file
is stored in an object file that the linker can read. The only caveat we
have to take care of is to ensure that the linker never actually
includes the contents of the object file into the final output. We apply
similar tricks as the `.llvmbc` bytecode sections to do this.
This involved changing the metadata loading code a bit, namely updating
some of the LLVM C APIs used to use non-deprecated ones and fiddling
with the lifetimes a bit to get everything to work out. Otherwise though
this isn't intended to be a functional change really, only that metadata
is stored differently in archives now.
This should end up fixing #83730 because by default dylibs will no
longer have their rlib dependencies "altered" meaning that
split-debuginfo will continue to have valid paths pointing at the
original rlibs. (note that we still "alter" rlibs if LTO is enabled to
remove Rust object files and we also "alter" for the #[link(cfg)]
feature, but that's rarely used).
Closes#83730
Previously, we would generate a single struct with the layout of the
dataful variant plus an extra field whose name contained the value of
the niche (this would only really work for things like `Option<&_>`
where we can determine that the `None` case maps to `0` but for enums
that have multiple tag only variants, this doesn't work).
Now, we generate a union of two structs, one which is the layout of the
dataful variant and one which just has a way of reading the
discriminant. We also generate an enum which maps the discriminant value
to the tag only variants.
We also encode information about the range of values which correspond to
the dataful variant in the type name and then use natvis to determine
which union field we should display to the user.
As a result of this change, all niche-layout enums render correctly in
WinDbg and Visual Studio!
This wasn't necessary for msvc and caused issues where different types
with the same name such as different instantiations of `Option<T>` would
have colliding debuginfo. This confused the debugger which would pick
one of the type definitions and use for all types with that name even
though they had different layout.
