rjmccall added inline comments.
================
Comment at: lib/CodeGen/CGBlocks.cpp:1276
+ InitVar->setSection(".CRT$XCLa");
+ CGM.addUsedGlobal(InitVar);
+ }
----------------
DHowett-MSFT wrote:
> rjmccall wrote:
> > Is the priority system not good enough?
> My reading of the LLVM language reference leads me to believe it’s only
> ordered per-module. If that’s the case, the benefit of emitting into `XC*` is
> that it provides guaranteed order over all linker input.
>
> `llvm.global_ctors` excerpt:
>
> > The functions referenced by this array will be called in ascending order of
> > priority (i.e. lowest first) when the module is loaded.
>
> Now if the priority system _is_ guaranteed over all linker input, will that
> guarantee hold for mixed Clang and CL objects?
Init priorities on ELF are preserved in the section name like
`.init_array.200`, and the sections get sorted by that in the image — it's
really a very similar trick to how it works with `.CRT$XC`. Of the major
formats, I think it's just Mach-O that doesn't have any built-in prioritization
mechanism across object files. But I don't know if LLVM actually tries to
translate init priorities over into `.CRT$XC` suffices when targeting PE/COFF,
and arguably that's good: init priorities as presented in LLVM right now are
pretty specific to the ELF mechanism. Long-term, maybe `llvm.global_ctors`
should be generalized so that on ELF targets it takes an integer priority, on
PE/COFF targets it takes a string, and on Mach-O it doesn't take anything at
all; but I won't hold up this patch for that.
On the other hand, I tend to agree that maybe the best solution is for the
backend to just take care of this and automatically create a global initializer
to install non-function (or maybe non-function & `unnamed_addr`) `dllimport`ed
symbols in global data.
Repository:
rC Clang
https://reviews.llvm.org/D50144
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
