[clang] [Clang][SYCL] Infer the SPIR-V target from an Intel offload architecture (PR #229248)
Srividya Sundaram via cfe-commits
cfe-commits at lists.llvm.org
Tue Nov 3 11:52:20 PST 2026
srividya-sundaram wrote:
> The `sycl-system-arch.cpp` test is breaking the pre-merge CI quite a bit, see e.g. https://github.com/llvm/llvm-project/actions/runs/41635790012/job/125916586515?pr=221794, can you please look into this, and if it's not something you can fix right away, revert the change or disable the test?
Since https://github.com/llvm/llvm-project/pull/229248 landed, `sycl-system-arch.cpp` seems to fail in about half of the macOS pre-merge runs: 29 of 53 runs that test clang, always with "Child timed out".
> 1: clang: error: cannot determine sycl architecture: /Users/runner/actions-runner/_work/llvm-project/llvm-project/build/tools/clang/test/Driver/Output/sycl-system-arch.cpp.tmp/offload_arch_sm_70_gfx906_gfx90a_xe_pvc: Child timed out: ; consider passing it via '--offload-arch'; environment variable CLANG_TOOLCHAIN_PROGRAM_TIMEOUT specifies the tool timeout (integer secs, <=0 is infinite)
On macOS CI, every launch of a fake GPU-detection tool is very slow, roughly 30 seconds.
The existing HIP, CUDA and OpenMP versions of this test already take 3-4 minutes there, but stay under clang's 60-second limit per launch.
Ours often goes over it. https://github.com/llvm/llvm-project/pull/238500 removes the limit for this test (CLANG_TOOLCHAIN_PROGRAM_TIMEOUT=0), so it passes, but it stays slow.
https://github.com/llvm/llvm-project/pull/229248
More information about the cfe-commits
mailing list