IRGen: internalise well known types with static linking

[compnerd]

When statically linking the standard library ensure that we emit the well known metadata and accessors with internal linkage. This fixes a number of warnings about incorrect DLL storage when building Foundation on Windows for static linking.

[compnerd]

@swift-ci please smoke test

[compnerd]

@swift-ci please smoke test

[compnerd]

@swift-ci please smoke test macOS platform

[compnerd]

@swift-ci please smoke test Linux platform

[compnerd]

@asl would you happen to have any ideas of why this might be causing the Differentiation build to fail only on Linux?

[asl]

Interesting. Tagging @kovdan01 – he is on Linux doing stuff around autodiff.

[kovdan01]

@asl would you happen to have any ideas of why this might be causing the Differentiation build to fail only on Linux?

@compnerd @asl TLDR: I've submitted #82412 fixing that.

Details. It looks like that we have a reinterpret_cast of SILDifferentiabilityWitness * to TypeBase *, which, to the best of my knowledge, looks totally incorrect - TypeBase is not a base class for SILDifferentiabilityWitness, and I've not found any evidence of TypeBase being stored at the beginning of memory layout of SILDifferentiabilityWitness via some dirty hacks or smth.

Such an unsafe incorrect casts explains why we see the issue only on Linux. We just try to run isCanonical on smth which is just not TypeBase, so we can get any result depending on the memory contents (and it just happens that on Mac and Windows builds we had this assertion passing, and on Linux builds - not). Also note that when investigating this issue locally, I had linux build finishing successfully or failing randomly - such non-determinism in builds passing or non-passing seems to have the same root cause.

#82412 contains a fix for that. In LinkEntity, we have Pointer and SecondaryPointer for storing payload pointers. I believe that for SIL differentiability witness, the latter should be used, while leaving Pointer as nullptr. This way, we'll not need any special checks against Kind::DifferentiabilityWitness.

[compnerd]

Please test with following PRs:
#82412

@swift-ci please smoke test Linux platform

[aschwaighofer]

lgtm

[asl]

@compnerd I'm not sure, but it seems #82412 is not pulled into your test. I was under impression that this only works for cross-repo PRs?

[compnerd]

@swift-ci please smoke test Linux platform