Storage buffers in binding arrays aren't writable

[jimblandy]

Since indexing a BufferArray never produces a pointer, it's impossible to write to the contents of buffers accessed through a buffer array.

Permitting pointers to buffer arrays seems like a bad idea. Perhaps indexing a BufferArray whose base type is not Image or Sampler should just always produce a pointer to the buffer's value, as if it were an ordinary var<storage> or var<uniform>, even when the base operand to Access or AccessIndex is not a pointer.

cc @cwfitzgerald

[cwfitzgerald]

Yeah I agree with the "just pointerize the result" solution. I was intentionally trying to avoid funky pointer stuff (like a pointer to an image in a binding array might lead to some funky consequences I don't understand).

[jimblandy]

I agree we want to avoid introducing pointers to new things. I think you'd say, for binding_array<texture> and binding_array<sampler>, indirection doesn't produce a pointer, but for everything else, it does.

[jimblandy]

There's also a language-level problem here. For buffers, we distinguish between read-only and read/write storage variables with:

var<storage> read_only_storage: i32;
var<storage, read> more_read_only_storage: i32;
var<storage, read_write> writable_storage: i32;

But since all buffer_array<Data> variables must go in the handle address space, there's no syntactic opportunity to say whether the variable is writable or not. (There's also no such opportunity in the IR.)

[kvark]

Note that this problem doesn't exist today because you can't actually use uniform/storage buffers in binding arrays.
#2282 makes it possible, and it uses the same address spaces as you'd use for non-arrays. So the problem is solved.