Naga features for platform dependent compilation of msl & hlsl

[Wumpf]

Connections

• Quite related to Add a default feature for wgpu that enables Vulkan & GLES on Windows & Linux #3514

Description

Adds cfg_alias indirection to Naga's x_out features which allows us to introduce new features for platform dependent compilation. This in turn allows wgpu-hal to request shading language output more conditionally.

Drawbacks:

• build.rs also on Naga
• new feature names are kinda clunky

Testing
Eyeballing on Windows. Single sample of cargo clean && cargo build -p wgpu --timings looked even more promising than I would have imagined:

Before:

After:

1.4s cpu time saved on this setup for not actually touching anything 😮. When building wgpu this is even on the hot path, meaning the the wgpu build as a whole got that much faster!
Edit: as expected lots of noise here. Running again on trunk I got 7.0s and 6.2s here, so above trunk compilation is likely with cold file caches. ymmv.

TODO:

• changelog entry
• test this manually on mac or linux and report impact

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.

[teoxoy]

I thought the backends won't be enabled if they won't be used on the target but it seems not because the conditional target compilation is setup in wgpu-core and happens after feature selection. It doesn't feel right that naga gets those new cfgs though; I would say that wgpu-hal is at fault here for not conditionally enabling naga's features based on the target. We already have some target cfgs in wgpu-hal/Cargo.toml, I think we should add naga in those instead of adding new features to naga. Thoughts?

[Wumpf]

I would say that wgpu-hal is at fault here for not conditionally enabling naga's features based on the target.

agreed, if possible we should solve it on that level instead 🤔
The problem though is that it's not a pure platform conditional: If I build on Windows without dx12 I expect HLSL to be disabled. If we were to put the naga dependency with hlsl-out under a windows target conditional, it would always enable hlsl-out though even the user left didn't specify dx12.
Putting the hlsl-out feature inside dx12 = [....] gets us back to status quo where we would enable hlsl-out if a user on linux/mac enabled dx12.
Any ideas how to get around this?

[teoxoy]

Any ideas how to get around this?

wgpu-hal/Cargo.toml has some deps behind target cfgs, I thought we could move the naga features under those as well.

[Wumpf]

it works if a feature should be dragged in or not on a per-target basis. The problem here is that it's a combination of feature plus target condition that we want this to be active upon. I.e. what we want is:

iff BOTH feature=dx12 is active AND target=windows enable naga's hlsl-out

The left hand side is easy since cargo supports that directly: my_feature = [other_crate/feature]
The right hand side is harder but doable with deps behind target cfg. E.g. khronos-egl is used with different features depending on the platform in wgpu-hal

[target.'cfg(not(target_arch = "wasm32"))'.dependencies]
khronos-egl = { version = "6", features = ["dynamic"], optional = true }
#...
[target.'cfg(target_os = "emscripten")'.dependencies]
khronos-egl = { version = "6", features = ["static", "no-pkg-config"] }

The thing I'm struggling here with is to evaluate both conditions at the same time

[teoxoy]

Ah yeah, that's problematic. I previously didn't fully grasp what you meant in your previous comment (#5919 (comment)).

I can't come up with a better solution either. I will take another look at this tomorrow.