[clang] Remove Linux search paths on Windows (PR #113628)

[Joseph Huber via cfe-commits]

jhuber6 wrote:

> > Is there an issue with simply using the `HostTC` for everything? I feel like that's the solution to this mess, since the `HostTC` would always know whether or not the target is Windows without us needing to forward a bunch of stuff.
> 
> Yes, that would work too. But currently the HostTC is only available/accessible in HIPAMDToolChain. With the way we instantiate the RID, we need to tell the RID ctor to hold off on creating the Search Paths if the HostTC is Windows. And the RID ctor is called all the way up in the Generic_GCC ctor. We would still have to pass the HostTC all the way through class hierarchy (HIPAMDToolChain -> ROCMToolChin -> AMDGPUToolChain -> Generic_ELF). My thinking was that passing a bool would be a little less intrusive/heavy of a change. Though, we would only be passing a reference to the HostTC.

Maybe I'm missing something, both the Host / Device cases will initialize the ROCm detector, however I think for the GPU case this should be a trivial exit without doing any work. The *uses* of the ROCm detector should then be localized inside the ROCm tools I would think, at which case it's just a matter of reaching into the host toolchain. The AMDGPUToolChain does not have a host toolchain because it's the base class without ROCm information so it shouldn't have any dependencies on it.

https://github.com/llvm/llvm-project/pull/113628