spirv-val: Disallow stores according to VUID 06924

[svenvh]

Ensure that the validator rejects stores to objects of types
OpTypeImage, OpTypeSampler, OpTypeSampledImage,
OpTypeAccelerationStructureKHR, and arrays of these types, according
to VUID-StandaloneSpirv-OpTypeImage-06924.

Guard the check behind the before_hlsl_legalization option, as
sometimes we may have temporaries or local variables that are expected
to get optimized away.

Fixes #4796

[svenvh]

If I understand the CI failures correctly, HLSL-to-SPIR-V translation by glslang and shaderc seems to result in some violations of this VUID, as HLSL allows the construct. Any suggestions what to do here?

[alan-baker]

What does glslang do in the GLSL path when you have a local acceleration structure variable?

[svenvh]

GLSL allows acceleration structures (and images/samplers) only as uniforms or function parameters, and glslang diagnoses incorrect uses:

ERROR: 0:10: 'accelerationStructureNV' : accelerationStructureNV can only be used in uniform variables or function parameters

[alan-baker]

Does DXC handle this? I expect this is a case where glslang is missing some legalization optimizations.

[spencer-lunarg]

Does this check cases where there is a OpImageWrite through an OpLoad or and Atomic instructions?

[svenvh]

Does this check cases where there is a OpImageWrite through an OpLoad or and Atomic instructions?

OpImageWrite is checked elsewhere (in ValidateImageWrite) as far as I can tell.

[svenvh]

To progress on this, I've excluded validation of stores to objects of types OpTypeImage and OpTypeSampler.

[svenvh]

Rebased onto latest main. @spencer-lunarg or @alan-baker would you be happy to give this a look again?

[alan-baker]

Just some updates to the message for now.

[alan-baker]

@s-perron Coming from a DXC perspective, should this check be gated by a pre legalization guard?

[s-perron]

I'm not sure. The OpTypeImage and OpTypeSampler should be guarded. We do not currently generate OpTypeSampledImage in a way that would require it, but I think that could happen in the near future.

I would say yes it we should check the flag. In general, the compiler may add temporaries or the user may have local variables that are expect to get optimized away.

This would be to solution to the issue with OpTypeImage and OpTypeSampler mentioned above. The output from the glslang HLSL->SPIRV should be validated with the pre-legalization option. If the checks are guarded by that option, then they should pass.

[svenvh]

This would be to solution to the issue with OpTypeImage and OpTypeSampler mentioned above. The output from the glslang HLSL->SPIRV should be validated with the pre-legalization option. If the checks are guarded by that option, then they should pass.

@s-perron thanks for the suggestion, I'd much rather land the complete validation including OpTypeImage and OpTypeSampler. Is it options()->before_hlsl_legalization that you're referring to?

[svenvh]

I've reinstated the full checks, including OpTypeImage and OpTypeSampler, now all guarded behind before_hlsl_legalization.