[llvm] [AMDGPU] Introduce ABI occupancy for object linking (PR #199475)

Krzysztof Drewniak via llvm-commits llvm-commits at lists.llvm.org
Mon Jun 1 08:48:32 PDT 2026


================
@@ -2583,6 +2585,48 @@ unit's worst case (i.e, maxima) ``num_vgpr``, ``num_agpr``, and
 symbolic expressions. These three symbols are ``amdgcn.max_num_vgpr``,
 ``amdgcn.max_num_agpr``, and ``amdgcn.max_num_sgpr``.
 
+.. _amdgpu-abi-occupancy:
+
+ABI Occupancy (Object Linking)
----------------
krzysz00 wrote:

I claim that `!reqd_work_group_size { X, Y, Z }` implies `amdgpu-flat-work-group-size=f"{X * Y * Z},{X * Y * Z}` (abusing Python f-string syntax there) and we should either verify this or propagate `!reqd_work_group_size` onto flat-work-group-size.

https://github.com/llvm/llvm-project/pull/199475


More information about the llvm-commits mailing list