Make spirv an optional feature

[kvark]

Connections
gfx-rs/naga#940 shows how much SPIR-V parsing can be a pain.

Description
Keep it supported natively, but put it behind a feature flag. This allows to skip compilation of parts of Naga as well as dependencies like petgraph.
On my machine, compiling wgpu-core time is reduced from 40.87s to 35.36s, which is about 13% improvement.

Testing
Just compiling

[ElectronicRU]

Regarding #1520 -- should that still be enabled without spirv? I'd say yes because this feature is related to parsing priv, not using spirv, but the name could get confusing.

[kvark]

@ElectronicRU Yes, I was just thinking about the same. Perhaps, we should adjust the (run-time) feature name to spell it more clearly.
Something like: SPIRV_SHADER_PASSTHROUGH? If it's just SPIRV_SHADERS then it's easy to confuse with the effect of the compile-time feature in this PR.

[cwfitzgerald]

I think we should go ahead with this, the compile time improvement is nothing to shake a stick at

[grovesNL]

Interestingly if we did allow naga modules to be provided as mentioned on Matrix recently, we wouldn't need SPIR-V specific options here. i.e. the application/library could opt-in to the SPIR-V frontend in naga

[cwfitzgerald]

It does predicate that naga can actually parse the spirv, which is one of the things passing raw spirv is trying to get around in #1525

[kvark]

@grovesNL indeed. It would have to be passed by move though, because wgpu wants to own Naga's module, but these modules aren't cloneable (at least today).

[kvark]

@grovesNL @cwfitzgerald I added a commit that removes SPIR-V support entirely from wgc. The support on wgpu-rs side is unchanged.

One of the nasty trade-offs here is being able to API trace the run involving SPIR-V. Since it's not in wgc, we have to dump the Naga IR, which requires "serialize" feature to be enabled. I thought that would increase the build times, but I wasn't able to see any difference locally. It will not affect most users, supposedly, because trace isn't enabled by default.

[cwfitzgerald]

I'd be really hesitant to merge removing spirv entirely right now. I do think this is eventually where we want to be, but having things be predicated on naga understanding every shader means that a good chunk of our features are unusable because naga is yet unaware of them. spirv was basically our last way around any current limitations of naga. I think in some time we should reconsider.

[kvark]

spirv was basically our last way around any current limitations of naga. I think in some time we should reconsider.

This functionality is here, it stays, and it's no directly related to this PR. See #1525

Also, the current state of the PR is still keeping SPIRV parsing support, just at wgpu-rs level instead of wgpu-core level.

[cwfitzgerald]

Oh I see, so we're just removing spirv ingestion from wgpu-core in the safe path. That sounds reasonable to me. I've got some nits, but otherwise LGTM.

[kvark]

bors r=cwfitzgerald

[bors]

Build succeeded:

• Android Stable
• iOS Stable
• MacOS Nightly
• MacOS Stable
• Ubuntu Nightly
• Ubuntu Stable
• Web Assembly
• Windows Nightly
• Windows Stable