SVE `__SVInt32_t` parameters become a zero-sized opaque; rustc drops the argument

Input C/C++ Header

int take(__SVInt32_t a, int t);

Bindgen Invocation

$ bindgen input.h -- -std=gnu11 -march=armv9-a+sve

Actual Results

bindgen 0.72.1 and 0.73.1 both turn the sizeless SVE vector into a zero-sized opaque blob (exit 0, default layout tests rustc-ok):

// 0.73.1
pub fn take(a: __BindgenOpaqueArray16<[u8; 0]>, t: c_int) -> c_int;

// 0.72.1
pub fn take(a: __BindgenOpaqueArray<u128, 0>, t: c_int) -> c_int;

Clang's LLVM for the same header keeps the scalable vector argument:

define i32 @take(<vscale x 4 x i32> %0, i32 noundef %1)

rustc of the generated bindings drops the ZST from the FFI signature (twice on each bindgen):

declare i32 @take(i32)

This is not GNU vector_size / ext_vector_type / NEON (those become arrays of the lane type). __SVInt32_t is a sizeless ACLE builtin; libclang reports size 0, and bindgen's opaque fallback uses that layout as the FFI argument. It is also not #727 / #3475: those are calling-convention enums on functions / function pointers. Here the data type of a parameter is wrong.

from_clang_ty has no arm for SVE sizeless vector kinds; they fall through to the opaque layout path with size 0.

Expected Results

Map SVE vector parameters as a real scalable-vector ABI (or skip take with a warning), not a ZST that rustc deletes from the signature.

Environment

bindgen: 0.72.1 and 0.73.1 (rust-lang/rust-bindgen 77cbc723)
clang/libclang: Homebrew clang 21.1.8
rustc: 1.97.1
target: aarch64-apple-darwin
OS: macOS (Apple M4)

Reproduce the issue with the provided input.h and bindgen invocation, then inspect from_clang_ty handling for SVE sizeless vector kinds. The fix should preserve a real scalable-vector ABI in the generated signature or skip take with a warning, and should ensure rustc does not drop the parameter as a zero-sized type.