fix(cuda-macros): obey #[inline] on generic device functions

[fxeqxmulfx]

Summary

#[device] added #[inline(never)] before the other attributes of a generic function. rustc uses only the first inline attribute. Because of this, rustc did not use an #[inline(always)] or #[inline] attribute on the function. rustc also gave an unused attribute warning. Now #[device] adds #[inline(never)] only if there is no inline attribute on the function.

Fixes #1471

Before → after

#[device]
#[inline(always)]
fn scale<T: Copy + core::ops::Mul<Output = T>>(value: T, factor: T) -> T {
    value * factor
}

#[kernel]
pub fn scale_all(input: &[f32], mut output: DisjointSlice<f32>) {
    let idx = thread::index_1d();
    let i = idx.get();
    if let Some(out) = output.get_mut(idx) {
        *out = scale(input[i], 2.0);
    }
}

	main	This PR
inline attributes of {DEVICE_PREFIX}scale	#[inline(never)] (from #[device]), #[inline(always)]	#[inline(always)]
rustc	warning: unused attribute for #[inline(always)]: "it will become a hard error in a future release"	No warning
Device LLVM IR before opt	scale_all calls scalefEB2_	No scalefEB2_: scale_all contains the code of scale
PTX	.entry scale_all and .visible .func scalefEB2_	.entry scale_all only

Changes

• cuda-macros/src/device.rs: #[device] adds #[inline(never)] to a generic function only if the function has no inline attribute. The check uses attr_path_ends_with. #[kernel] uses the same function for the attributes of generic kernels. rustc expands cfg_attr before it calls the macro. Because of this, the check also finds an inline attribute in cfg_attr.
• The expansion is now in device_function_tokens, as for generic kernels in generic_kernel_no_instantiation_tokens. Unit tests can examine it.
• Regression tests: the unit test generic_device_function_keeps_the_inline_attribute_of_the_user (#[inline], #[inline(always)], #[inline(never)]), and the trybuild pass test tests/pass/device_inline_policy.rs with #![deny(unused_attributes)]. Both fail when #[device] always adds #[inline(never)], as on main.
• Control tests: generic_device_function_is_inline_never_by_default and non_generic_device_function_gets_no_inline_attribute.
• Negative test: other_attributes_do_not_replace_the_default (#[cold], #[doc = "inline"]).
• standalone_device_fn: the comment about #[inline(never)] now includes "by default".

Testing

• just -f cuda-oxide/Justfile check passes (the local mirror of CI: fmt, clippy, tests, guards, docs)
• cargo oxide run <example> passes, or cuda-oxide/scripts/smoketest.sh -o '^<example>$'
• New example added (if applicable)

cargo oxide run standalone_device_fn --arch sm_80 gives SUCCESS: 18/18 tests passed.

On an RTX 3050 (sm_86), the example above gives correct results on main and with this change. The table shows the output of the two builds.

Checklist

• All commits signed off (git commit -s)
• SPDX headers on new source files

🤖 Generated with Claude Code

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.