[llvm] [SROA] Improve select folding around casts (PR #222466)

[Yaxun Liu via llvm-commits]

yxsamliu wrote:

Thanks for the volatile-case fix. Here is the complete reproducer (16 lines):

```llvm
%struct.T = type { i32, i32 }

define i32 @test_select_fold_split_volatile(i1 %cond) {
  %alloc0 = alloca %struct.T, align 8, addrspace(5)
  %alloc1 = alloca %struct.T, align 8, addrspace(5)
  store %struct.T { i32 0, i32 1 }, ptr addrspace(5) %alloc0
  store %struct.T { i32 2, i32 3 }, ptr addrspace(5) %alloc1
  %sel = select i1 %cond, ptr addrspace(5) %alloc0, ptr addrspace(5) %alloc1
  %cast = addrspacecast ptr addrspace(5) %sel to ptr
  %gep0 = getelementptr inbounds %struct.T, ptr %cast, i32 0, i32 0
  %gep1 = getelementptr inbounds %struct.T, ptr %cast, i32 0, i32 1
  %val1 = load volatile i32, ptr %gep0
  %val2 = load volatile i32, ptr %gep1
  %val3 = add i32 %val1, %val2
  ret i32 %val3
}
```

I ran `opt -S -passes='default<O2>'` followed by `llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -O2`, using the same LLVM build with only `SROA.cpp` changed between revisions:

| Revision | Code bytes | VGPRs |
| --- | ---: | ---: |
| PR base (`4cc7b28c`) | 124 | 4 |
| Earlier PR head (`cf93fc85`) | 156 | 6 |
| Current head (`639ece04`) | 124 | 4 |

The current head preserves the single selected address and fixes this reproducer. Thanks!


https://github.com/llvm/llvm-project/pull/222466