On 02.01.2019 08:42, Joerg Sonnenberger wrote: > On Tue, Jan 01, 2019 at 10:30:38PM +0100, Kamil Rytarowski wrote: >> Joerg seems to just want to play with Clang out of the LLVM projects, >> deteriorating experience with the rest blocking this patch. Other people >> need more than that. > > Can we please keep down on the ad hominem, thanks? > I'm sorry, but I didn't have intention of this interpretation. >>> We provide abi compatibility, so does this mean we just need to bounce >>> the llvm libs major version on every import or is it worse? >>> >> >> We will need to bump major version with each LLVM upgrade. >> >> We need to keep building a selection of LLVM tools to generate or enable >> building neded LLVM files. >> >> One thing that might tricky and considered as worse.. LLVM profile >> (--coverage option) for some reason has been pushed into NetBSD libc >> (it's part of compiler-rt) in a legacy version ABI v2. Today LLVM uses >> ABI v4 and not sure if we can do anything with it without formal libc >> major bump. > > It's the v2 ABI because that's what compiler-rt provided at the time. As > long as the interface of v4 is a superset of v2, that's perfectly fine. > If it is, please apply the complains to the origin of the problem... > LLVM profile tests fail with our unmodified libc, as in part of the tests the runtime from libc is picked up (instead of a locally build version linked with a test program) and it causes abort. "LLVM Profile Error: Runtime and instrumentation version mismatch : expected 4, but get 2" For the process of porting LLVM profile to NetBSD (in HEAD of LLVM compiler-rt), I had to eject the existing copy from libc. I don't intend to do anything with it myself for now. I've just submitted a request to Itanium C++ ABI team to make it clearer how to handle the dlclose(3) + atexit(3) behavior. > Joerg >
Attachment:
signature.asc
Description: OpenPGP digital signature