I think this just means not to eliminate it in dead function elimination, which seems reasonable. (It works in non-WMO -O and in -Onone.) @eeckstein, who works on dead function elimination these days?

It's pretty easy to add that in dead function elimination: Just add an additional check in

bool isAnchorFunction(SILFunction *F)

That turns out not to work: isAnchorFunction() is never called during the stdlib's build (at least not a debug-stdlib build), presumably because DFE is not running at all. Something else must be killing the function.

There is another thing: IRGen only emits functions lazily. The check for this is done in IRGenerator::emitGlobalTopLevel().

I think the best solution is to check for @_silgen_name in SILFunction::isPossiblyUsedExternally(). Unfortunately in some places (like IRGenerator::emitGlobalTopLevel) doesn't call this SILFunction member, but instead call the global function isPossiblyUsedExternally. This has to be changed.

Original isAnchorFunction WIP is https://github.com/gparker42/swift/tree/keep-silgenname .

Changing SILFunction::isPossiblyUsedExternally() and SILFunction::isExternallyUsed() was insufficient.

Changing the global isPossiblyUsedExternally() function in SILLinkage.h to return true did work.

One conceptually correct fix would be to introduce an additional SILLinkage value between Public and Hidden which I shall call MostlyHidden.
Public: visible outside the current Swift module and/or linkage unit.
MostlyHidden: visible only to the current Swift module and to non-Swift code in the same linkage unit.
Hidden: visible only to the current Swift module.

> Changing SILFunction::isPossiblyUsedExternally() and SILFunction::isExternallyUsed() was insufficient.

As I said earlier, there are some uses of the global isPossiblyUsedExternally function, which are called with the function's linkage (like in IRGen). You have to replace this with the SILFunction::isPossiblyUsedExternally.

That's the cleanest and easiest fix.

> Changing the global isPossiblyUsedExternally() function in SILLinkage.h to return true did work.

That would be bad for the optimizer.

> One conceptually correct fix would be to introduce an additional SILLinkage value between Public and Hidden

Please don't. It would make our already complicated linkage model even more not-understandable.

#12696

This uses the SILFunction::isPossiblyUsedExternally() approach. It also covers @_cdecl to fix rdar://33924873.

There is residual ugliness involving this and isKeepAsPublic() and isExternallyUsedSymbol(). Future cleanup ought to get these three special cases to share more code in common.

This is done now.