Non-package module should load swiftinterface instead of binary module with PackageCMO.

[elsh]

If the binary module optimized with Package CMO is loaded for a non-package client module, the client module does not understand the serialized content of the optimized binary module because it contains attributes and instructions specific to the optimization, i.e. [serialized_for_package] definition that contains operations on non-loadable types that are normally not allowed. This can cause an assert or a crash in SIL verifier.

This PR checks whether the optimized binary module and the client module are in the same package, and require loading swiftinterface module if not.

Resolves rdar://134584629

[elsh]

@swift-ci smoke test

[elsh]

@swift-ci smoke test

[xymus]

Can you move the MODULE_PACKAGE_NAME and SERIALIZE_PACKAGE_ENABLED fields from the options_block to the control_block as part of this PR? The control_block should have all the core information to tell if a module can be loaded or not, whereas the options_block is for additional information that is not always read depending on the client. This change should also streamline the logic as there's existing logic to reject the module. For an example you can take a look at the CHANNEL field and how it’s used to throw an ChannelIncompatible error.

The alternative would be like IS_TESTABLE, allowing to load the swiftmodule and report an error on the import. That's more appropriate in the case of a misconfiguration, the control_block approach is better for silently rejecting a swiftmodule file to prefer the swiftinterface.

[elsh]

The correct approach is to load the same optimized binary module for use by external clients of the package. Therefore, I prefer to keep MODULE_PACKAGE_NAME and SERIALIZE_PACKAGE_ENABLED within the options block, as it wouldn't make sense to move them to the control block and then back again. I've also updated the code comments to reflect this. Regarding diagnostics, serialization::Status::PackageMismatch has been added, along with the RebuildInfo::emitRemarks.

[elsh]

@swift-ci smoke test

[elsh]

@swift-ci smoke test

[elsh]

After further consideration, it seems more reasonable to implement the proper long-term fix stated above instead of relying on this interim solution, especially since there's already a workaround: pass -Xfrontend -module-load-mode -Xfrontend prefer-interface. Thus will be closing this PR.