[clang] [llvm] [mlir] [OpenMP] Reserve the main thread's warp for generic mode kernels (PR #218790)

[Larry Meadows via cfe-commits]

lfmeadow wrote:

If the three lines are widening the clamp in `getEffectiveNumThreads` — it does
add the warp, and the `std::min(MaxNumThreads, ...)` on the next line takes it
back off — I tried that. It launches the bigger block, and then faults.

`MaxNumThreads` is `KernelEnvironment.Configuration.MaxThreads`, which is also
the number that became `amdgpu-flat-work-group-size`, and `.maxntid` on NVPTX,
when the kernel was compiled. Widening the clamp puts 65 threads into a kernel
whose metadata says `thread_limit(1)` means 1. The backend has already used that
number: `AMDGPULowerIntrinsics` demotes a workgroup barrier to a wave barrier
when the bound is no bigger than the wavefront, and `AMDGPUPromoteAlloca` sizes
an LDS-promoted alloca from it. CUDA rejects a launch over `.maxntid` outright.

So the runtime can't widen past the bound — the warp has to be in the number the
compiler writes.

If you meant something else, tell me which lines and I'll try it.


https://github.com/llvm/llvm-project/pull/218790