https://github.com/kuhar approved this pull request. Thanks. Having to use a new flag is suboptimal, but overall this seems like a big improvement over using the gpu-to-spirv pass that accumulated unrelated patterns. https://github.com/llvm/llvm-project/pull/106082