[Mlir-commits] [clang] [llvm] [mlir] [openmp] [OpenMP][offload] Cross-team reductions with variable number of teams (PR #195102)

[Dhruva Chakrabarti]

dhruvachak wrote:

` For a dot reduction using double type, for example, we previously had something like
LDS Usage: 540B #SGPRs/VGPRs: 106/45 #SGPR/VGPR Spills: 34/0 Tripcount: 177777777, which now became
LDS Usage: 280B #SGPRs/VGPRs: 48/30 #SGPR/VGPR Spills: 0/0 Tripcount: 177777777`

Can you add a test in-tree or point out an existing test for which you see the above behavior? I did not think that the current compiler would spill for a simple dot reduction.

https://github.com/llvm/llvm-project/pull/195102