This is the supported way rocm is tested.
It makes packaging in nix a *lot* easier (see the code size).
An important change is the dontLink detection in the clang/clang++
wrapper script: When compiling with --cuda-device-only,
the linker must not be set, otherwise e.g. the blender kernels fail to
compile.
It is possible to both be bare metal and have a libc (newlib).
This libc doesn't provide very much --- not enough for CMake to think
the C toolchain works. We therefore adjust our logic so we hit the "bare
metal" case with or without libc.
The "use LLVM" bootstrap is intentionally not affected.
GCC 10 sets -fno-common by default. This broke some packages, so
when moving to GCC 10 we initially disabled this behavior. This
commit reverts that, bringing us closer to the standard and
upstream.
Co-authored-by: Sergei Trofimovich <slyich@gmail.com>
Reverts #162607 / 1748887ff2.
Reason for revert: This change caused llvm-config{,-native} to be unable
to find static archives bundled with LLVM, as has been [reported]. Ever
since #152944 using moveToOutput in LLVM is _evil_ because llvm-config
obtains it knowledge about the installation locations from the CMake
configure step.
Consequently a change like #162607 will need to be implemented by making
LLVM itself install the static archives to the correct location or by
adding yet another patch which updates llvm-config's knowledge of the
location. The latter is not desireable in my opinion, though, since it
is just asking for this sort of trouble: Before #152944 we had an
outputs.patch that did this sort of things which broke spectacularly in
edge cases.
Fixes#148117.
[reported]: https://github.com/NixOS/nixpkgs/issues/148117#issuecomment-1158245576
Fixes#166833.
The build creates a symlink for this assuming it's present,
so be sure it's there when filtering the source for clang.
Alternatively we could use LLVM_EXTERNAL_CLANG_TOOLS_EXTRA_SOURCE_DIR.
lld package provides an unwrapped lld. It doesn't always work on NixOS
(eg, it doesn't set rpath), and so dosen't always work.
What one should be using instead is the `lld` from
`llvmPackages.bintools` package. This super counterintutive.
One incremental step we can take here is to clarify that the `lld`
package is unrwapped -- right now, it looks like 100% legit thing one
should be using!
This seems to fix the notorious "CommandLine Error: Option 'xxxxx'
registered more than once!" error in applications that use both Mesa and
ROCm.
Since Mesa is built with llvmPackages_latest and ROCm stuff is built
with llvmPackages_rocm, applications that use both (such as Blender) end
up with two different `libLLVM*.so`s loaded, which breaks things.
This seems like a straightforward way to fix the problem, and since the
ROCm stack seems to be the only thing in Nixpkgs that uses
llvmPackages_rocm this hopefully shouldn't break anything.
While there might be another way to fix this problem that doesn't
require disabling the shared libraries, I haven't been able to find it
yet, and since this issue seems to affect a lot of people I think it
might make sense to merge this fix for now and revisit it later if a
better solution is found.
This also removes a small patch to rocm-comgr since there are no longer
LLVM shared libraries for it to link against.
