vgvassilev wrote:

> I cannot check the actual bot since it just started a new run, but in another 
> `projects` build it's in `clang/21/lib/sunos/liborc_rt-x86_64.a`. 

Ok, that exposed a bug somewhere in the particular toolchain when we called 
`TC.getCompilerRT(C.getArgs(), "orc_rt");`. That means that clang does not 
detect that's the right path. The issue is very likely here:
https://github.com/llvm/llvm-project/blob/46de3200388c300ee871f5b4941dff0d97ba9480/clang/lib/Driver/ToolChain.cpp#L682-L703

We will need to provide a special case like we do here:

https://github.com/llvm/llvm-project/blob/46de3200388c300ee871f5b4941dff0d97ba9480/clang/lib/Driver/ToolChain.cpp#L717

It's late on my end -- do you mind giving this a shot?

> There **must** be some way to locate runtime libs without every command that 
> needs that second-guessing...

This is a very long conversation. The second guessing happens because of the 
fact that clang computes the default resource-dir relative to the binary 
directory. However, if you link clang as a library against a test that relative 
path is wrong. This is a bit of grey area when using clang as a library and in 
the tests we have chosen to write tests in a such a way that we don't rely on 
that. If we want to test the out-of-process execution in clang-repl we need to 
provide such path either on the test side or here. The upshot of the current 
approach is that it will be useful for users and deployments outside of that 
one single test. If you have good ideas how to make this better I am all ears...



https://github.com/llvm/llvm-project/pull/175322
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits

Reply via email to