ash
ash copied to clipboard
ash-rs
ash: Hide provisional extensions behind new `provisional` feature flag
Fixes #343
Provisional extensions are susceptible to breaking API changes at any point (with accompanying SPEC_VERSION bump). By hiding them behind a non-default feature flag, we can document that opting into this feature allows us to do breaking changes to these modules in non-breaking ash releases when driven by upstream Vulkan-Headers changes to vk.xml.
Moreover, upstream Khronos repositories that build ash in their CI can continue to build-test their changes without compile-testing any extension behind a provisional flag, which is especially useful if we've already wrapped a provisonal extension that is seeing breaking changes.
One important feature that is missing in this PR and the reason for its draft state is that we don't yet wrap any types <require>d by provisional changes. While these types are unlikely to be useful without enabling the extension, non-breaking releases with breaking changes to such types should be considered invalid. And there may be provisional extensions that solely consist of new types (extending existing types via pNext).
Secondly, feature flags may not technically exempt us from semver compatibility in releases, even if the feature flag is documented as such. We could "do this anyway" for simplicity, or move to something more aggressive like cfg definitions.
BTW, I don't know that it impacts anything you're doing here, but just to note that in the C bindings, the provisional types and commands are kept in a separate 'beta' header, but the provisional enums added to existing enumerated types have to be visible in the core API header as there is no way to define them separately from the type they are part of.
They could remain in the same file - it's only a handful - and individual enum variants or constants can also be hidden behind a #[cfg] (somewhat similar to a #if preprocessor directive).
Working on this PR is really showing that the generators' lack of building definitions etc. top-down while being aware of their enabling feature(s)/extension(s) is such a hassle to build back in. It takes quite a while to push that information back in and is messy while at it.
Secondly, feature flags may not technically exempt us from semver compatibility in releases, even if the feature flag is documented as such
I don't think we need to overthink this -- semver is ultimately a tool for humans, and we shouldn't make the human experience worse for its sake.
I don't think we need to overthink this -- semver is ultimately a tool for humans, and we shouldn't make the human experience worse for its sake.
Sure enough, keeping this behind a feature is adequate.
Semver isn't totally a "human" tool to me as it has a machine meaning to things like cargo, signifying backwards compatibility across patch revisions. A patch release with breaking changes would make implicit updates horrible in the end-user experience especially as a transitive dependency many layers down.