Can't use fn() for C binding if C uses sentinel value

[karolherbst]

I've tried to implement the cl_khr_icd 2.0.0 OpenCL extension in Rust. This extension requires the implementers to set constant values to certain function pointers. rustc told me to file an issue, so here I am.

I don't know how widespread this kind of pattern is, and if it could become a problem implementing other C APIs in Rust.

As a side note, constructing the same value at runtime instead of at compile time makes the code compile. Obviously actually calling the function will be undefined behavior. The constant value just is just a marker to loaders to advertise that version 2.0.0 of cl_khr_icd is supported. It is not expected to be called at all by anything.

Below is a simplified version of the code:

#![allow(non_camel_case_types, non_snake_case)]

use std::mem;

pub type cl_int = i32;
pub type cl_uint = u32;
pub type cl_platform_id = *mut _cl_platform_id;
pub type cl_icd_dispatch = _cl_icd_dispatch;

pub struct _cl_platform_id {
    pub dispatch: *const cl_icd_dispatch,
}

pub type clGetPlatformIDs_t = ::std::option::Option<
    unsafe extern "C" fn(
        num_entries: cl_uint,
        platforms: *mut cl_platform_id,
        num_platforms: *mut cl_uint,
    ) -> cl_int,
>;

pub struct _cl_icd_dispatch {
    pub clGetPlatformIDs: clGetPlatformIDs_t,
}

#[cfg(target_pointer_width = "32")]
const CL_ICD2_TAG_KHR: usize = 0x434C3331;

#[cfg(target_pointer_width = "64")]
const CL_ICD2_TAG_KHR: usize = 0x4F50454E434C3331;

pub static DISPATCH: cl_icd_dispatch = cl_icd_dispatch {
    clGetPlatformIDs: unsafe { mem::transmute_copy(&CL_ICD2_TAG_KHR) },
};

I expected to see this happen: It should compile and place the constant values in the static data.

Instead, this happened:

error[E0080]: it is undefined behavior to use this value
  --> test.rs:32:1
   |
32 | pub static DISPATCH: cl_icd_dispatch = cl_icd_dispatch {
   | ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ constructing invalid value at .clGetPlatformIDs.<enum-variant(Some)>.0: encountered 0x4f50454e434c3331[noalloc], but expected a function pointer
   |
   = note: The rules on what exactly is undefined behavior aren't clear, so this check might be overzealous. Please open an issue on the rustc repository if you believe it should not be considered undefined behavior.
   = note: the raw bytes of the constant (size: 8, align: 8) {
               31 33 4c 43 4e 45 50 4f                         │ 13LCNEPO
           }

Meta

rustc --version --verbose:

rustc 1.85.0 (4d91de4e4 2025-02-17)
binary: rustc
commit-hash: 4d91de4e48198da2e33413efdcd9cd2cc0c46688
commit-date: 2025-02-17
host: x86_64-unknown-linux-gnu
release: 1.85.0
LLVM version: 19.1.7

[the8472]

The constant value just is just a marker to loaders to advertise that version 2.0.0 of cl_khr_icd is supported

Would a union work then?

[karolherbst]

The constant value just is just a marker to loaders to advertise that version 2.0.0 of cl_khr_icd is supported

Would a union work then?

Ohh. I forgot to mention. This is all bindgen generated code from the official OpenCL headers. Just the last code block (assigning the static value) is my own code.

[karolherbst]

So the Working group could be convinced to change the headers, but it's not really in my control if the field could be a union. But I'll try to see if that would compile or not.

[karolherbst]

yeah, this code compiles successfully. Also gets rid of the unsafe and casts. Maybe I can bring it up and get it changed to an anonymous union on the C side to not break existing code.

#![allow(non_camel_case_types, non_snake_case)]

use std::mem;

pub type cl_int = i32;
pub type cl_uint = u32;
pub type cl_platform_id = *mut _cl_platform_id;
pub type cl_icd_dispatch = _cl_icd_dispatch;

pub struct _cl_platform_id {
    pub dispatch: *const cl_icd_dispatch,
}

pub type clGetPlatformIDs_t = ::std::option::Option<
    unsafe extern "C" fn(
        num_entries: cl_uint,
        platforms: *mut cl_platform_id,
        num_platforms: *mut cl_uint,
    ) -> cl_int,
>;

pub union _cl_icd_dispatch__bindgen_ty_1 {
    pub clGetPlatformIDs: clGetPlatformIDs_t,
    pub clGetPlatformIDs_icd2_tag: isize,
}

pub struct _cl_icd_dispatch {
    pub anon_1: _cl_icd_dispatch__bindgen_ty_1,
}

#[cfg(target_pointer_width = "32")]
const CL_ICD2_TAG_KHR: isize = 0x434C3331;

#[cfg(target_pointer_width = "64")]
const CL_ICD2_TAG_KHR: isize = 0x4F50454E434C3331;

pub static DISPATCH: cl_icd_dispatch = cl_icd_dispatch {
    anon_1: _cl_icd_dispatch__bindgen_ty_1 {
        clGetPlatformIDs_icd2_tag: CL_ICD2_TAG_KHR,
    },
};

[workingjubilee]

Bindgen allows targeted rewriting in some cases, most notably by excluding certain types or values from bindings generation and using your own definitions for them with a bit of twiddling.

...though obviously that struct is a huge dispatch table, so I understand any refusal to actually do that.

[saethlin]

I agree that the problem here is the binding generation. If magic sentinel values are permitted, Option<fn()> is simply the wrong type to represent the C API in terms of Rust types.

[karolherbst]

I think the one argument one could make here that it's strange that for a static initializer rustc complains about this, but if you have the same code assigning to a local variable it magically works.

However personally I'm fine with pushing the responsibility dealing with this upwards and as long as this isn't widespread I think it's fair to leave it as is until more people actually need a more streamlined solutions for this.

[workingjubilee]

The initializer of a static is a "const context" and thus is evaluated via CTFE. That means it adheres to the stricter rules required for such.