[naga] Write only the current entrypoint

[andyleiserson]

Changes the MSL and HLSL backends to support writing only a single entry point, and uses them that way in wgpu-hal. (The GLSL and SPIR-V backends already provide an entry point selector.)

This is working towards a fix for #5885, but this change doesn't alter override processing or validation, so we will still reject shaders for missing overrides whether or not they are used.

Testing
Both the write-one-entrypoint and write-all-entrypoints cases are exercised either by existing or changed code. Existing tests cover the change in behavior, in that the tests still work, even when no longer writing all entrypoints to the output. Additional partial-output test cases will be added with the complete fix for #5885.

Squash or Rebase? Squash

Checklist

• Run cargo fmt.
• Run taplo format.
• Run cargo clippy --tests. If applicable, add:
  • --target wasm32-unknown-unknown
• Run cargo xtask test to run tests.
• Added a changelog entry.

[cwfitzgerald]

+1 this for a different reason - it's less code to hand to the driver (for it to just throw out) and less for people looking in shader debuggers to wade through.

[teoxoy]

Entry points are uniquely identified by (name, stage), that's what the SPIR-V & GLSL backends use as well, I think the MSL & HLSL backends should follow that model to ensure only a single entry point is written.

[andyleiserson]

@teoxoy Currently I have entry_point: Option<String> for Metal and HLSL to provide backwards compatibility with the current behavior. Thoughts on how important backwards compatibility is? If adding the stage I could have shader_stage and entry_point as separate optional values, separate non-optional values, or as a single Option<(ShaderStage, String)> which seems cleaner than two separate options but would be inconsistent with GLSL and SPIR-V.

[teoxoy]

I think we should go with Option<(ShaderStage, String)> (or something like Option<EntryPointIdent> where EntryPointIdent is a struct that contains the name and stage). SPIR-V already uses this model by making its PipelineOptions optional in write_vec. GLSL is the only target that doesn't support multiple entry points per file and requires passing the name and stage.

[teoxoy]

Nice!

[teoxoy]

Looks like the msl test needs a bump in stack space?

[andyleiserson]

Looks like the msl test needs a bump in stack space?

Yes, I assume this is because of the increase in the size of back::msl::Error, because the new EntryPointNotFound variant has two members. I also had to increase the target in test_error_size from 32 to 40.

Although in the case of test_stack_size, I think it was already on the edge of failing -- it has been failing for me locally since the rust version was set to 1.84. But only rust 1.84 seems to have that issue, so I figured I could live with the local failures until we bump the rust version again.