[llvm] Add a pass "SinkGEPConstOffset" (PR #140657)

[via llvm-commits]

================
@@ -829,26 +829,26 @@ define amdgpu_kernel void @test_sched_group_barrier_pipeline_MFMA_interleave_spl
 ; GCN-ILP-NEXT:    ds_read_b128 a[28:31], v3 offset:49264
 ; GCN-ILP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN-ILP-NEXT:    v_mfma_f32_32x32x1f32 a[0:31], v0, v1, a[0:31]
-; GCN-ILP-NEXT:    v_add_u32_e32 v3, 0x6000, v3
+; GCN-ILP-NEXT:    v_add_u32_e32 v3, 0x14000, v3
----------------
StevenYangCC wrote:

I have made changes based on the suggestions you made, please verify the results.

https://github.com/llvm/llvm-project/pull/140657