MTLBuffer contents pointer is nullable

[K0bin]

objc2_metal:

        /// Returns the data pointer of this buffer's shared copy.
        #[unsafe(method(contents))]
        #[unsafe(method_family = none)]
        fn contents(&self) -> NonNull<c_void>;

Apple documentation:

A pointer to the shared copy of the buffer data, or NULL for buffers allocated with a private resource storage mode (MTLStorageModePrivate).

[madsmtm]

Damn. This is a bug in Apple's headers (the header is marked NS_ASSUME_NONNULL_BEGIN, so contents should have been opted out with some sort of __nullable). Swift is actually also affected by this, they map the pointer as UnsafeMutableRawPointer instead of UnsafeMutableRawPointer?.

It's unfortunately a breaking change for me to properly fix, and while you can sometimes argue that it's fine when it's done to fix soundness issues, this isn't really a soundness issue for consumers unless they use MTLStorageModePrivate.

I'll probably fix for now by checking internally on access whether the pointer is NULL.

[madsmtm]

Hmm, actually, I know that's what the docs say, but how are you observing this?

The small test below seems to indicate that the docs for contents are actually incorrect?

#!/usr/bin/env cargo -Z script
---
[package]
edition = "2024"
[dependencies]
objc2-metal = "0.3.2"
---
use objc2_metal::{MTLBuffer, MTLCreateSystemDefaultDevice, MTLDevice, MTLResourceOptions};

#[link(name = "CoreGraphics", kind = "framework")]
unsafe extern "C" {}

fn main() {
    let device = MTLCreateSystemDefaultDevice().unwrap();
    let buffer = device.newBufferWithLength_options(100, MTLResourceOptions::StorageModePrivate).unwrap();
    println!("{:?}", buffer.contents());
}

[madsmtm]

See also gfx-rs/wgpu#7357.

[madsmtm]

The small test below

I tried running this on both a Macbook Pro M2 running macOS 15.6.1, and on an older 2013-something Intel Macbook (integrated graphics card) running Mac OS X 10.12.6, and both return a pointer (and not NULL).

But maybe it's NULL when using an external graphics card?

[madsmtm]

I got a friend to test the snippet, it does indeed return NULL on systems with discrete GPUs.