[llvm] [AMDGPU] Remove the AnnotateKernelFeatures pass (PR #130198)

[Jun Wang via llvm-commits]

================
@@ -35,15 +35,15 @@ define amdgpu_kernel void @stack_object_addrspacecast_in_kernel_no_calls() {
 ; RO-FLAT:     scratch_store_dword
 ; RW-FLAT:     .amdhsa_user_sgpr_private_segment_buffer 1
 ; RO-FLAT-NOT: .amdhsa_user_sgpr_private_segment_buffer
-; RW-FLAT:     .amdhsa_user_sgpr_flat_scratch_init 1
+; RW-FLAT:     .amdhsa_user_sgpr_flat_scratch_init 0
----------------
jwanggit86 wrote:

Previouly the AnnotateKernelFeatures pass checks for calls and allocas. With the change to AMDGPUAttributor, alloca is no longer being checked. For this particular case, opt would set the `no-flat-scratch-init` attribute. So after llc, flat_scratch_init in line 38 would be 0.

https://github.com/llvm/llvm-project/pull/130198