[Distributed] Survive generics with Sendable requirement in executeDistributedTarget #41854
Conversation
ktoso
force-pushed
the
wip-generics-missing-witness
branch
from
7a2e0cf to
ae36adb
Compare
…stTarget resolves rdar://90293494
ktoso
force-pushed
the
wip-generics-missing-witness
branch
from
ae36adb to
a249ccd
Compare
| for (const auto &req : requirements) { | ||
| // Make sure we understand the requirement we're dealing with. | ||
| if (!req.hasKnownKind()) | ||
| return TypeLookupError("unknown kind"); | ||
| if (!req.hasKnownKind()) { |
There was a problem hiding this comment.
so the kind here ends up being some weird 24 for the Sendable requirement...
Since it has an unknown kind we can't really get the mangled name or protocol etc to check it in detail if it is the exact sendable.
But the same issue would happen with any other @_marker protocol I suppose...
ktoso
added
AArch64 Linux
distributed
|
@swift-ci please smoke test |
| /// \param skipUnknownKinds If true, unknown requirement kinds will not be | ||
| /// reported as errors. This should used with care, and e.g. only to ignore | ||
| /// @_marker protocols at runtime, which cannot be checked at runtime. |
There was a problem hiding this comment.
In the future, should we consider a runtime representation that indicates the requirement came from a marker protocol? Having any representation at runtime sounds like it goes against the purpose of a marker protocol, but I guess that was unavoidable in this case?
There was a problem hiding this comment.
Ignoring/skipping Sendable requirements at this level shouldn't be necessary. Rather, we shouldn't mangle the Sendable requirements at all, so we don't need to demangle them. ASTMangler has an option, AllowMarkerProtocols, that you can set to false so that the mangler skips over any marker protocols when it mangles.
|
@DougGregor This is not going to be as easy as mangling, generic environment descriptor @ktoso If you don't mind I'd like to take over fixing this issue. |
|
I see... thanks a lot, that makes sense @DougGregor . Thanks for the help @xedin ! |
|
Replaced by #41871 |
Since
Sendableas any other @_marker protocol does not exist at runtime, we're not able (and DON'T HAVE TO) find a witness for it. The function we're to invoke does not expect a witness for those anyway.I did not find a way to detect "oh, this was a marker protocol" so instead I provide a mode to ignore unknown kinds. We only use it in our code so there should be no risk to other existing code.
resolves rdar://90293494