[spv-out] decorate non-uniform image/sampler access chains

[teoxoy]

See #4762 (comment).

Fixes #4762.
Reverts c3e35df#diff-42823a07fbb2862baa61487d0fe17c6ba02e2db32560e884875449114897d975L1459-R1480.

[Elabajaba]

Tested this and it fixed the issue on windows+amd+vulkan for me, and didn't regress linux (radv)+amd+vulkan.

[jimblandy]

I'm confused - the patch doesn't match the commit message or the discussion.

The code looks like it has always been decorating the results of OpAccessChain, just not for binding arrays in the Handle address space.

The discussion suggests that both the pointer and the result of the load need to be decorated, but we are not decorating the result of the load / sample at all, as far as I can see.

Can't we do better at helping ourselves understand what's going on the next time we have to look at this?

[teoxoy]

The code looks like it has always been decorating the results of OpAccessChain, just not for binding arrays in the Handle address space.

The code originally decorated all access chains, c3e35df changed it so that it only happened for storage and uniform buffers since I thought that decorating the image/sampler access chains was not necessary. We now know this was necessary and is what this PR is reverting.

The discussion suggests that both the pointer and the result of the load need to be decorated, but we are not decorating the result of the load / sample at all, as far as I can see.

Since c3e35df we decorate the load result of image/sampler binding arrays (the load result being an image/sampler); this is part of the VUID requirement.

The results of texture load/sampling operations don't have to be decorated.

[jimblandy]

Okay, I understand this better now. But please update the commit message.

[teoxoy]

How should the commit message be updated?

[jimblandy]

@teoxoy Here's my suggestion (summary line and detailed explanation):

[spv-out] Decorate all non-uniform binding array access chains as NonUniform, not just buffer binding arrays.

Apply the NonUniform decoration to the results of all access chains rooted in binding arrays that use non-uniform values as indices, regardless of the binding array's element type and address space. Previously, Naga only decorated non-uniform access chains for binding arrays of buffers.

The important difference is that this message doesn't mention load results, which this PR has no effect on (and which Naga has never decorated as far as I can tell).