Soundness of BufferViewMut

[agausmann]

Although attempting to get a &[u8] from a write-only BufferViewMut will panic, getting a &mut [u8] is considered perfectly fine, which seems suspicious to me, because &mut [u8] is still readable.

If we try to perform a read, what is going to happen, is the behavior defined? And even if it is defined soundly, if it crashes the program (e.g. it generates an access violation), I'd still consider it a hazardous interface because we can't guarantee that the compiler won't emit a load, even if we never explicitly access it in the source code.

[cwfitzgerald]

It's perfectly defined, it's just going to be potentially slow as hell due to the kind of caching behavior inherent with write-optimized mappings. The whole "getting a read only view panics" is just an attempt to discourage reading from it, as ultimately rust has no concept of a "write only" type.

[agausmann]

I forgot to mention, the reason why I mention and assume that the view is write-only is because that's implied in the documentation for MapMode::Write.

But based on your answer, I understand it's just heavily write-optimized and not intended to be used for reading, however reads are still allowed. Would that be a better summary?

[agausmann]

And I'd be curious to know, are there any coherency guarantees about reading the same value after writing?

[cwfitzgerald]

I understand it's just heavily write-optimized and not intended to be used for reading, however reads are still allowed. Would that be a better summary?

Correct.

are there any coherency guarantees about reading the same value after writing?

I believe it will be the same coherence as any other store.