[Darwin] Get rid of trampoline hack for lgamma function.

[rintaro]

What's in this pull request?

Since swiftc pass-through -Xcc arguments to ClangImporter,
we can unveil lgamma_r with -Xcc -D_REENTRANT.

Resolved bug number: N/A

---

Before merging this pull request to apple/swift repository:

• Test pull request on Swift continuous integration.

[gribozavr]

@jrose-apple Is this change safe for the downstream modules and apps?

[jrose-apple]

Maybe? Because lgamma is inlineable, its body must only reference things that a downstream app would see. But since lgamma has a simple signature, it's possible the forward-declaration in SIL would be self-contained and not reference anything imported.

Fortunately, compiling simple calls to lgamma (1) covering all the overloads, (2) without the macro defined, and (3) with optimizations on should be enough to verify whether it's a problem—if it is, the SIL verifier will fail. Unfortunately, I don't think the general case is or will ever be sound for inlineable code. It won't ever miscompile, though—either it'll work, or a debug compiler will trap and a release compiler will probably crash.

It's up to you whether you want to take it under those conditions.

[gribozavr]

it's possible the forward-declaration in SIL would be self-contained and not reference anything imported.

Based on #1724, I'm afraid that does not really work.

To be on the safe side, I'm inclined to reject this change.

[jrose-apple]

I'm confused. #1724 doesn't seem to offer any barriers to this at all.

[gribozavr]

For this change, my concern is that we won't pass -D_REENTRANT to swiftc invocations, which would mean that we would have inconsistent Clang importer state between invocations that serialize the Darwin module and those that import it. The ones that import it won't get the Clang declarations if they don't use -D_REENTRANT. Does this mean that cross-references in the module would be broken, for example?

[jrose-apple]

That's exactly the right concern, but because the only references are at the SIL level and there are no Clang types involved, there shouldn't be any cross-references to Clang decls. …I think. There are many ways that would no longer be true in the future (debug info, perhaps), but again we'd detect it very quickly.

[jrose-apple]

It also depends on what else _REENTRANT does, which might be enough reason to leave it alone for now.

[gribozavr]

@rintaro Sorry!

[rintaro]

To be on the safe side,

I agree.
Thank you for very thoughtful discussion @gribozavr @jrose-apple !