Emulate @weak functions on Windows and don't emit COMDATs for ELF anymore

[kinke]

The test adaptations are mostly a revert of 3893840.

@weak global variables are still left out but could be emulated exactly the same way.

[kinke]

I've looked into this because it might be a solution for #3158, at least for LDC - only emitting weak extern(C++) std::exception::what() etc. in druntime, so that they are ABI-compatible with the C++ ones and don't conflict with a potentially linked in C++ runtime.

[kinke]

Is there any point in emitting Comdats for ELF? It prevents @weak from working properly (linker apparently prefers the version in the 1st object file, independent of whether it's weak or not), and clang doesn't emit them for Linux (well, for MSVC neither, but AFAIK, there the Comdats play the role of -f{function,data}-sections for linker stripping)...

[JohanEngelen]

This does not do the same thing. What we want is the function to be opaque. Disabling inlining still allows IPO to learn about what is written to foo. I think the same holds for optStrategy(none).
Edit: but applied to the hot function it does not matter here because there indeed we don't want inlining.

[kinke]

I just wanted to avoid additional complexity for string matching; for hot, disabling inlining wasn't enough, as the call was still elided. What worked was pragma(inline, false); llvm_sideeffect(); in the hot body, but optStrategy seemed nicer and had the same effect.

For select_func here, I thought disabling inlining would be enough (for string matching in the main body wrt. the foo function ptr); re-adding @weak wouldn't affect any strings though, and might even fix the regression with LLVM 4.0.

[kinke]

@rainers: Any suggestions wrt. how to 'best' adapt the mangled name (for backtrace/debugger/...)?

[rainers]

@rainers: Any suggestions wrt. how to 'best' adapt the mangled name (for backtrace/debugger/...)?

If you want it to be demangled with some modifications, prepending a package weak by changing the _D prefix to _D4weak should work.

Your use case in #3158 probably needs C++ symbols, though. VC mangling doesn't encode identifier length, so patching the first identifier might work, but it can be back referenced so it appears multiple times in the demangled name.

[kinke]

Cleaned up & put a little more effort into the renamed mangle.

clang emits COMDATs for templates and inline functions only (with linkonce_odr), not for regular functions, for both MSVC and Linux targets. I have no idea whether ELF COMDATs provide any benefit and so whether I should try harder and only get rid of them for regular functions, to make @weak work properly.

[JohanEngelen]

You could ask on the forums about the COMDATs on Linux, but I doubt you'll get useful feedback.
If Clang does not emit it, then I think it is fairly safe to omit them.

[kinke]

Yeah I wanted to know if you guys know more about this, otherwise I'd propose to include this in the upcoming beta and see how it goes.

[kinke]

Interestingly, the produced static libraries are now noticeably smaller for the Azure Linux x64 job (compared against latest master run):

• libdruntime-ldc.a: 2,898 vs. 3,210 KB (-9.7%)
• libphobos2-ldc.a: 9,466 vs. 10,393 KB (-8.9%)

The shared libs are minimally larger:

• libdruntime-ldc-shared.so: 1,625 vs. 1,619 KB
• libphobos2-ldc-shared.so: 6,202 vs. 6,173 KB (+0.5%)

[kinke]

FWIW, I've also tried setting the COMDAT's selection kind to 'exact match' (instead of 'any'), but that only led back to the multiple definition errors.

[JohanEngelen]

I just found out that the ELF COMDAT change is causing trouble at Weka. It prevents merging of symbols, leaving excess data around in a section where Weka's tooling wants there to be no garbage (or well, it just exceeds the amount of data that the system can deal with (ushort...)). Re-enabling COMDAT fixes it.
I guess we have to start doing what @kinke noted about clang (I think that would fix Weka's problem):

clang emits COMDATs for templates and inline functions only (with linkonce_odr), not for regular functions,

I'll work on a small testcase, so you can see what the problem is.

[JohanEngelen]

Opened an issue for it with testcase: #3589

[JohanEngelen]

LLVM 11 release notes:
"Windows target: Produce COFF weak external symbols for IR level weak symbols without a comdat (e.g. for attribute((weak)) in C)"
https://releases.llvm.org/11.0.0/docs/ReleaseNotes.html#changes-to-the-windows-target

@kinke Does that remove the need for the "emulation" of this PR ?

[kinke]

I've read that, found it a nice conincidence, but haven't tested whether it works similar to this manual emulation here. Anyway, it's not restricted to LLVM11+, and we always emit comdats on Windows.