Offload device
LLVM's offload functionality usually expects an extra dyn_ptr argument. We could avoid it,b ut likely gonna need it very soon in one of the follow-up PRs (e.g. to request shared memory). So we might as well already add it.
This PR adds a %dyn_ptr ptr to GPUKernel ABI functions, if the offload feature is enabled.
WIP
r? ```@ghost```
Add default sanitizers to TargetOptions
Some sanitizers are part of a system's ABI, like the shadow call stack on Aarch64 and RISC-V Fuchsia. Typically ABI options have other spellings, but LLVM has, for historical reasons, marked this as a sanitizer instead of an alternate ABI option. As a result, Fuchsia targets may not be compiled against the correct ABI unless this option is set. This hasn't caused correctness problems, since the backend reserves the SCS register, and thus preserves its value. But this is an issue for unwinding, as the SCS will not be an array of PCs describing the call complete call chain, and will have gaps from callers that don't use the correct ABI.
In the long term, I'd like to see all the sanitizer configs that all frontends copy from clang moved into llvm's libFrontend, and exposed so that frontend consumers can use a small set of simple APIs to use sanitizers in a consistent way across the LLVM ecosystem, but that work is not yet ready today.
Some sanitizers are part of a system's ABI, like the shadow call stack
on Aarch64 and RISC-V Fuchsia. Typically ABI options have other
spellings, but LLVM has, for historical reasons, marked this as a
sanitizer instead of an alternate ABI option. As a result, Fuchsia
targets may not be compiled against the correct ABI unless this option
is set. This hasn't caused correctness problems, since the backend
reserves the SCS register, and thus preserves its value. But this is an
issue for unwinding, as the SCS will not be an array of PCs describing
the call complete call chain, and will have gaps from callers that don't
use the correct ABI.
In the long term, I'd like to see all the sanitizer configs that all
frontends copy from clang moved into llvm's libFrontend, and exposed so
that frontend consumers can use a small set of simple APIs to use
sanitizers in a consistent way across the LLVM ecosystem, but that work
is not yet ready today.
Remove a special case and move another one out of reachable_non_generics
One is no longer necessary, the other is best placed in cg_llvm as it works around a cg_llvm bug.
Allow codegen backends to indicate which crate types they support
This way cargo will drop the unsupported crate types for crates that
specify multiple crate types.
This is a prerequisite for https://github.com/rust-lang/miri/pull/4648.
Skip codegen_crate call in check mode
This way we don't have to spawn the coordinator thread. Some errors will no longer be emitted with this in check mode. For example the check that `-Ctarget-cpu` is passed on targets that need this.
Suggested by `@saethlin`
Forbid ShallowInitBox after box deref elaboration.
MIR currently contains a `ShallowInitBox` rvalue. Its principal usage is to allow for in-place initialization of boxes. Having it is necessary for drop elaboration to be correct with that in-place initialization.
As part of analysis->runtime MIR lowering, we canonicalize deref of boxes to use the stored raw pointer. But we did not perform the same change to the construction of the box.
This PR replaces `ShallowInitBox` by the pointer manipulation it represents.
Alternatives:
- fully remove `ShallowInitBox` and implement `Box` in-place initialization differently;
- remove the `ElaborateBoxDeref` pass and keep dereferencing `Box` in runtime MIR.
Add NonNull pattern types
These are the final piece missing for
* https://github.com/rust-lang/rust/pull/136006
We cannot use the previous scheme of using an integer range for raw pointers, as we're not just changing the layout of raw pointers anymore, but also the type representation. And we can't represent "any provenance or NonZero<usize>" natively as patterns. So I created a new `!null` pattern. Since this is all unstable representation stuff for replacing rustc_layout_scalar_range_start with pattern types, the divergence from normal patterns is fine, especially since T-lang seems interested in exploring general negation patterns
r? `@BoxyUwU`
the `#[track_caller]` shim should not inherit `#[no_mangle]`
fixes https://github.com/rust-lang/rust/issues/143162
builds on https://github.com/rust-lang/rust/pull/143293 which introduced a mechanism to strip attributes from shims.
cc `@Jules-Bertholet` `@workingjubilee` `@bjorn3`
---
Summary:
This PR fixes an interaction between `#[track_caller]`, `#[no_mangle]`, and casting to a function pointer.
A function annotated with `#[track_caller]` internally has a hidden extra argument for the panic location. The `#[track_caller]` attribute is only allowed on `extern "Rust"` functions. When a function is annotated with both `#[no_mangle]` and `#[track_caller]`, the exported symbol has the signature that includes the extra panic location argument. This works on stable rust today:
```rust
extern "Rust" {
#[track_caller]
fn rust_track_caller_ffi_test_tracked() -> &'static Location<'static>;
}
mod provides {
use std::panic::Location;
#[track_caller] // UB if we did not have this!
#[no_mangle]
fn rust_track_caller_ffi_test_tracked() -> &'static Location<'static> {
Location::caller()
}
}
```
When a `#[track_caller]` function is converted to a function pointer, a shim is added to drop the additional argument. So this is a valid program:
```rust
#[track_caller]
fn foo() {}
fn main() {
let f = foo as fn();
f();
}
```
The issue arises when `foo` is additionally annotated with `#[no_mangle]`, the generated shim currently inherits this attribute, also exporting a symbol named `foo`, but one without the hidden panic location argument. The linker rightfully complains about a duplicate symbol.
The solution of this PR is to have the generated shim drop the `#[no_mangle]` attribute.
Move computation of allocator shim contents to cg_ssa
In the future this should make it easier to use weak symbols for the allocator shim on platforms that properly support weak symbols. And it would allow reusing the allocator shim code for handling default implementations of the upcoming externally implementable items feature on platforms that don't properly support weak symbols.
In addition to make this possible, the alloc error handler is now handled in a way such that it is possible to avoid using the allocator shim when liballoc is compiled without `no_global_oom_handling` if you use `#[alloc_error_handler]`. Previously this was only possible if you avoided liballoc entirely or compiled it with `no_global_oom_handling`. You still need to avoid libstd and to define the symbol that indicates that avoiding the allocator shim is unstable.
Validate CopyForDeref and DerefTemps better and remove them from runtime MIR
(split from my WIP rust-lang/rust#145344)
This PR:
- Removes `Rvalue::CopyForDeref` and `LocalInfo::DerefTemp` from runtime MIR
- Using a new mir pass `EraseDerefTemps`
- `CopyForDeref(x)` is turned into `Use(Copy(x))`
- `DerefTemp` is turned into `Boring`
- Not sure if this part is actually necessary, it made more sense in rust-lang/rust#145344 with `DerefTemp` storing actual data that I wanted to keep from having to be kept in sync with the rest of the body in runtime MIR
- Checks in validation that `CopyForDeref` and `DerefTemp` are only used together
- Removes special handling for `CopyForDeref` from many places
- Removes `CopyForDeref` from `custom_mir` reverting rust-lang/rust#111587
- In runtime MIR simple copies can be used instead
- In post cleanup analysis MIR it was already wrong to use due to the lack of support for creating `DerefTemp` locals
- Possibly this should be its own PR?
- Adds an argument to `deref_finder` to avoid creating new `DerefTemp`s and `CopyForDeref` in runtime MIR.
- Ideally we would just avoid making intermediate derefs instead of fixing it at the end of a pass / during shim building
- Removes some usages of `deref_finder` that I found out don't actually do anything
r? oli-obk
In the future this should make it easier to use weak symbols for the
allocator shim on platforms that properly support weak symbols. And it
would allow reusing the allocator shim code for handling default
implementations of the upcoming externally implementable items feature
on platforms that don't properly support weak symbols.
Currently it is possible to avoid linking the allocator shim when
__rust_no_alloc_shim_is_unstable_v2 is defined when linking rlibs
directly as some build systems need. However this requires liballoc to
be compiled with --cfg no_global_oom_handling, which places huge
restrictions on what functions you can call and makes it impossible to
use libstd. Or alternatively you have to define
__rust_alloc_error_handler and (when using libstd)
__rust_alloc_error_handler_should_panic
using #[rustc_std_internal_symbol]. With this commit you can either use
libstd and define __rust_alloc_error_handler_should_panic or not use
libstd and use #[alloc_error_handler] instead. Both options are still
unstable though.
Eventually the alloc_error_handler may either be removed entirely
(though the PR for that has been stale for years now) or we may start
using weak symbols for it instead. For the latter case this commit is a
prerequisite anyway.
