WGSL: Implement diagnostic(…); directives and derivative_uniformity triggering rule

[ErichDonGubler]

Connections

Resolution:

• Partially resolves Support diagnostic filters in WGSL (i.e., diagnostic(…)) #5320.
• Resolves Feature request: disable uniformity analysis on shader creation #3135.

Dependencies:

• Hard dependency on Add a skeleton for directives that reports nice errors #6352.
• Soft dependency on Unbreak rustfmt in some places #6349.

Description

Recommended review experience is commit-by-commit. Each individual commit should pass CI!

This change implements a meaningful first iteration of support for diagnostic(…); directives; to wit:

1. Define the essential state that Naga needs to track for compilation messages controlled by diagnostic(…) rules.
2. Control Naga's current uniformity analysis errors via the derivative_uniformity rule. The current analysis is incorrect, and blocks users in issues like Feature request: disable uniformity analysis on shader creation #3135. This can give them a way forward.

Several things are out-of-scope:

• There are more grammar rules where we should also @diagnostic(…) attributes, in order to conform fully with the WGSL spec. This PR does not attempt to handle all of these cases, to limit the complexity of this PR.

  The exhaustive list of standards-mandated parse sites for @diagnostic(…) not implemented in this PR is tracked in Support diagnostic filters in WGSL (i.e., diagnostic(…)) #5320.

• Naga's current diagnostic model for shader compilation is too limited for us to comply with the spec.-mandated behavior with warnings. Naga offers only a Result to indicate operation success, or a single fatal error message. This conflicts with the Severity::Warning model introduced in this PR, because standard WGSL compilation might report multiple diagnostics in cases where parsing the shader module…

  1. …emits multiple warning items.
  2. …emits warnings in addition to a fatal error-level item.

  Because of this limitation, and because we want to limit complexity in this PR, diagnostic(warn, …) will not report a standard warning-level compilation item in cases that the standard dictates a warning-level compilation item. Instead, we will opt for log::warn!(…) entries. This is tracked as follow-up issues:

  • Support multiple diagnostic items in shader compilation #5295
  • Naga: support all standard severities for diagnostics #6458

  New cases where the above is relevant:

  • unrecognized diagnostic rule names
  • uniformity analysis diagnostics emitted at a warning level

Testing

Current test plan:

• Demonstrate that the following CTS cases work:
  • webgpu:shader,validation,parse,diagnostic:valid_locations:*
    • …:type="directive";location="module"
  • webgpu:shader,validation,parse,diagnostic:conflicting_directive:*
• Port a meaningful subset of the above to our own Naga test suite, to prevent the most impactful regressions. Punted to test(naga): add coverage for diagnostic directive parsing #6457.

Checklist

• Run cargo fmt.
• Run cargo clippy. If applicable, add:
  • --target wasm32-unknown-unknown
  • --target wasm32-unknown-emscripten
• Run cargo xtask test to run tests.
• Add change to CHANGELOG.md. See simple instructions inside file.

[ErichDonGubler]

@jimblandy: This is ready for a full review!

[ErichDonGubler]

Just validated that tests written in the (revised) Testing portion of the OP work as expected when consumed by Firefox. 👏🏻

[jimblandy]

As far as I can see, this never actually populates naga::Module::diagnostic_filters or diagnostic_filter_leaf with what was parsed by the WGSL front end. The PR description says this is supposed to control how validation reports FunctionError::NonUniformControlFlow, so I would have expected the diagnostics to have been passed through to the Module.

If this is correct, then that also indicates some missing tests.

[ErichDonGubler]

@jimblandy: Ah, it seems I'd missed a couple of things:

1. Needed to remove DISABLE_UNIFORMITY_REQ_FOR_FRAGMENT_STAGE, which I've squashed into 2841f18. I added coverage to the existing uniform_control_flow test to check that one of the offending

   It would be much better if I had a WGSL snippet in hand that definitively fails, both for an end-to-end test and something for the CHANGELOG, but...we've struggled to find one, and I think the two hours I spent trying to figure it out needs to be price enough.

2. Needed to wire up the transfer of state from a TranslationUnit into a Module, which I've squashed in f950c59. I added a test to ensure that a module's IR actually contains a parsed diagnostic rule.

I fixed it, I think!

[jimblandy]

Okay, this looks good. Just a few comments.