[Bazel] LLVM builds fail unless NVPTX backend is enabled in llvm_configure

[j2kun]

When I set up the build rules for a project (https://github.com/google/heir/tree/2054509f517a1b2abaaeff0e592f99d223f383a6), I decided I only wanted to compile the X86 backend in llvm_configure.

The result was a build error like this when building most targets.

ERROR: /home/j2kun/.cache/bazel/_bazel_j2kun/fc8ffaa09c93321753c7c87483153cea/external/llvm-project/mlir/test/BUILD.bazel:591:11: 
no such target '@llvm-project//llvm:NVPTXCodeGen': target 'NVPTXCodeGen' not declared in package 'llvm' defined by 
/home/j2kun/.cache/bazel/_bazel_j2kun/fc8ffaa09c93321753c7c87483153cea/external/llvm-project/llvm/BUILD.bazel 
(Tip: use `query "@llvm-project//llvm:*"` to see all the targets in that package) 
and referenced by '@llvm-project//mlir/test:TestGPU'

This can be reproduced by checking out the above repository at 2054509f517a1b2abaaeff0e592f99d223f383a6 and commenting out this line, and then running

bazel build @llvm-project//mlir/...:all

It seems like these GPU-related targets should be conditionally included when the relevant backend targets are present.

I asked about this on discord and got no response. Seems like a minor bug, and I understand bazel support is best effort, so for now the workaround is just to compile NVPTX as well.

[aaronmondal]

Thanks for tracking this.

I've also encountered this, but it never seemed to matter enough to further investigate, though it might be worth looking into.

I believe that this actually affects all targets, i.e. the configure rule is broken/doesn't properly gate targets and config flags. In practice this means that it's not really possible to disable any target. At least if you want to link against bazel-built LLVM, you'll have to link against every backend.

This is probably not intended behavior and adds a lot of unnecessary compile time. I can try looking into this as it could potentially mean a big compiletime reduction for e.g. CPU-only projects using the overlay.

[vimarsh6739]

@j2kun @aaronmondal Was this ever resolved? I still have to build llvm with NVPTX (when I just need the X86 target)

[j2kun]

I haven't personally looked at it. I filed it when I noticed it during the writing of https://github.com/j2kun/mlir-tutorial