[llvm] AMDGPU: Default to selecting frame indexes to SGPRs (PR #115060)

[via llvm-commits]

bcahoon wrote:

Hi @arsenm, can you take a look at the following case. It still fails with [#115747](https://github.com/llvm/llvm-project/pull/115747). Thanks.

```
%"struct.ck::Tuple.1296.524.974.1094.1109.1114.1144.1159.1174" = type { %"struct.ck::detail::TupleImpl.1297.523.973.1093.1108.1113.1143.1158.1173" }
%"struct.ck::detail::TupleImpl.1297.523.973.1093.1108.1113.1143.1158.1173" = type { %"struct.ck::detail::TupleElementKeyData.241.522.972.1092.1107.1112.1142.1157.1172" }
%"struct.ck::detail::TupleElementKeyData.241.522.972.1092.1107.1112.1142.1157.1172" = type { %"struct.ck::vector_type.7.521.971.1091.1106.1111.1141.1156.1171" }
%"struct.ck::vector_type.7.521.971.1091.1106.1111.1141.1156.1171" = type { %union.anon.8.520.970.1090.1105.1110.1140.1155.1170 }
%union.anon.8.520.970.1090.1105.1110.1140.1155.1170 = type { <4 x float> }

define amdgpu_kernel void @test() {
entry:
  %src_vectors.i1.i.i.i.i.i.i.i.i = alloca %"struct.ck::Tuple.1296.524.974.1094.1109.1114.1144.1159.1174", i32 0, align 16, addrspace(5)
  %src_vectors.ascast.i3.i.i.i.i.i.i.i.i = addrspacecast ptr addrspace(5) %src_vectors.i1.i.i.i.i.i.i.i.i to ptr
  %add.ptr.i.i.i.i.i.i.i.i.i.i.i.i12.i.i.i.i.i.i.i.i = getelementptr i8, ptr %src_vectors.ascast.i3.i.i.i.i.i.i.i.i, i64 4
  %vtable.i.i.i.i2.i.i.i15.i.i.i.i.i.i.i.i = load ptr addrspace(1), ptr null, align 8
  %0 = load ptr addrspace(1), ptr addrspace(1) %vtable.i.i.i.i2.i.i.i15.i.i.i.i.i.i.i.i, align 8
  call addrspace(1) void %0(ptr null, ptr null, ptr %add.ptr.i.i.i.i.i.i.i.i.i.i.i.i12.i.i.i.i.i.i.i.i)
  ret void
}
```
```
  llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx90a test.ll -o test.s 
```


https://github.com/llvm/llvm-project/pull/115060