Fix use of extraCPPFlags

[neonichu]

We should not use extraCPPFlags as a vehicle for specifying linker flags, instead just decide on which C++ library to link based on the target triple in BuildPlan.

Separately, it seems worthwhile to look into refactoring `Destination` to actually allow specifying linker flags since that seems to have been an accidental feature of `extraCPPFlags` (as long as one didn't pass compiler flags there as well).

rdar://100575898

[neonichu]

@swift-ci please smoke test

[neonichu]

Looks like we have some tests asserting the broken behavior since -lc++ happens to also be recognized by swiftc. I don't think we have to special case since -Xcc -lc++ is presumably identical to -lc++?

[abertelrud]

Wouldn't we need -Xlinker instead of -Xcc if we want to pass it to the linker?

[neonichu]

Wouldn't we need -Xlinker instead of -Xcc if we want to pass it to the linker?

Yah, looks like it, the question is why are we even passing extraCPPFlags to the linker at all. Seems like instead Destination should be changed to explicitly declare its linker flags separately or we could filter extraCPPFlags and only pass the linker relevant flags.

[abertelrud]

Yes, that seems more correct. Compilation flags and linking flags should generally be separate. There might be some leniency in that compiler drivers might accept and ignore a linker flag if only doing compilation, but it would better to not rely on that (also less confusing for users who are looking at build logs).

[neonichu]

I guess that would mean to also have per-language linker flags in the destination, because otherwise we won't be able to make -lc++ conditional on targets containing C++. I wonder if the right fix for this particular bug is to move this code into the build plan and then a follow up can explore how to restructure the Destination type.

[neonichu]

Updated to move the code for deciding on C++ library linkage, but I think we should be revisiting Destination separately as well.

[neonichu]

@swift-ci please smoke test

[abertelrud]

I guess that would mean to also have per-language linker flags in the destination, because otherwise we won't be able to make -lc++ conditional on targets containing C++. I wonder if the right fix for this particular bug is to move this code into the build plan and then a follow up can explore how to restructure the Destination type.

That seems reasonable. Separately, it seems to me that we should also factor out all of these platform specifics into separate protocol-conforming entities that provide the specifics. Maybe that is what you were saying about the Destination. But that, too, is something that can be done in the future. This change looks good for what it is.