[llvm] [SDAG] Avoid creating redundant stack slots when lowering FSINCOS (PR #108401)

[Benjamin Maxwell via llvm-commits]

MacDue wrote:

> I think you can drop all of the chain logic and just use DAG.getEntryNode. FSINCOS is derived from no-memory/errno operations, and thus assumed to not care about the state of errno. You don't need all of this code trying to figure out the suitable chain

I don't think that's safe with the stores. Given this IR (which contains a sincos intrinsic with lowers to FSINCOS):

```llvm
define void @test(float %v, ptr %x, ptr %y) {
  call void @foo(ptr %x, ptr %y)

  %sret = call { float, float } @llvm.sincos.f32(float %v)
  %ret0 = extractvalue { float, float } %sret, 0
  %ret1 = extractvalue { float, float } %sret, 1
  store float %ret0, ptr %x, align 4
  store float %ret1, ptr %y, align 4

  ret void
}

declare void @foo(ptr, ptr)
```

The stores for the results of `FSINCOS` take the chain from the call to `@foo` as the input. With the current logic, this lowers correctly (a call to `foo` followed by a call to `sincosf`). If I instead make the input chain `DAG.getEntryNode()` the call to `foo` is lost.  


https://github.com/llvm/llvm-project/pull/108401