This largely works as expected, except that as far as I can tell,
without explicit annotations added to the caller, (or a
`#force_inline`), the LLVM `target-features` function attribute
gets ignored by the inliner under the rationale of `-sse,-avx,-avx2`
is a subset of `+sse,+avx,+avx2`.
With `#force_no_inline` the correct code does get generated, but in
the regression I am trying to fix, the caller gratuitously also uses
SIMD, leading to horrific performance.
On Windows with LTO, required procedures with external linkage need to
be added to @llvm.used to survive linker-level dead code elimination.
LLVM may generate implicit calls to runtime builtins (e.g., __extendhfsf2
for f16 conversions) during instruction lowering, after the IR is
finalized. Without @llvm.used, the linker discards these procedures
before the implicit calls are generated.
This adds required procedures to @llvm.used at creation time. The fix
is Windows-specific; other platforms handle this correctly.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add intrinsics.objc_super()
- Emit objc_msgSendSuper2 calls when an objc method call is combined with objc_super(self)
- Fix objc_block return value ABI for large struct returns
- Fix objc_implement method wrappers bad ABI for large struct returns and indirect args
- Simplify parameter forwarding for objc_imlpement methods
- Add intrinsics.objc_instancetype to mimi Objective-C instancetype* returns
This facilitates returning the correct type on subclasses when calling mehtods
such as `alloc`, `init`, `retain`, etc.
- Refactor Objective-C class implementations generation so that hierarchies are properly initialized
- Better codegen for context passing with ivar-based autocontext
- Allow @superclass on imported objc-c objects
- Better codegen for block forwarding invoker, arguments are forwarded directly
