ELF: Emit (most) instantiated symbols in COMDATs #4748
Conversation
kinke
changed the title
| const bool inCOMDAT = needsCOMDAT() || | ||
| // ELF needs some help for ODR linkages: | ||
| // https://github.com/ldc-developers/ldc/issues/3589 | ||
| (linkage == templateLinkage && | ||
| global.params.targetTriple->isOSBinFormatELF()); | ||
| return {linkage, inCOMDAT}; |
There was a problem hiding this comment.
I don't think this works, I'll double confirm on weka source code, but we have been getting too much linker issues by applying COMDAT to most of the symbols. e.g.:
ld.lld: error: relocation refers to a discarded section: .text._D5mecca3lib10reflection__T2asVAyaa13_6e6f7468726f7720406e6f6763TDFNaNbNiNfZvZQBzFNcMQuZ9__lambda2MFNaNbNiNeZQBs
>>> defined in core/3rdparty/build/mecca/libmecca-notrace.a(.._.._mecca_src_mecca_lib_exception.d.o)
>>> section group signature: _D5mecca3lib10reflection__T2asVAyaa13_6e6f7468726f7720406e6f6763TDFNaNbNiNfZvZQBzFNcMQuZ9__lambda2MFNaNbNiNeZQBs
>>> prevailing definition is in core/3rdparty/build/mecca/libmecca-notrace.a(.._.._mecca_src_mecca_reactor_io_inotify.d.o)
>>> referenced by reflection.d:556 (../core/3rdparty/mecca/src/mecca/lib/reflection.d:556)
>>> .._.._mecca_src_mecca_lib_exception.d.o:(_D5mecca3lib9exception6ExcBuf__T15constructHelperHTC4coreQBt11AssertErrorTAyaZQBuMFNbNiNeQpmbKQuZQBu) in archive core/3rdparty/build/mecca/libmecca-notrace.a
This verifies on symbols with InternalLinkage and only after upgrading the compiler (symbols with internal linkage are the ones that are relocations to discarded symbols). I haven't tested with the same LLVM version tho, could be an upstream regression.
Bottom line: weka@e1cca7a with --enable-wekamods=true yields this linker error, and, without the mods, data culling doesn't work (we start to get duplicate sections, as described in the issue). The --enable-wekamods=true works on older versions. So we are on this middle ground where none works unless we only mark specific linkage with comdat.
I'll try to create a test to demonstrate this.
There was a problem hiding this comment.
Tested, and it works! I see, this works if we guarantee that templateLinkage is not ever InternalLinkage, which probably is just a derivative of _odr attributes? What about --fvisibility=hidden, this doesn't influence templateLinkage, right? If that's all true, I'm good with this change.
There was a problem hiding this comment.
templateLinkage is always an ODR linkage:
Lines 470 to 472 in 580ff68
ELF: Emit (most) instantiated symbols in COMDATs
No description provided.