LLVM fails to tail call when `ret void` is replaced with `unreachable`

The following llvm ir tail calls twice. The only difference is that one return known to be `unreachable`, which should only improve optimizations
```llvm
define void @base(ptr noundef nonnull readonly captures(none) %ptr) unnamed_addr {
start:
tail call void %ptr()
ret void
}

define void @worse(ptr noundef nonnull readonly captures(none) %ptr) unnamed_addr {
start:
tail call void %ptr() #2
unreachable
}
```
On `LLVM version: 22.1.0` produces the following for `x86_64-unknown-linux-gnu` (and also seemingly equivalent assembly for other targets)
```asm
base: # @base
jmp rdi # TAILCALL
worse: # @worse
push rax
call rdi
```
I think that it is perfectly safe and correct to just produce the `base` code in both cases?

Start by reproducing the two LLVM IR functions on the x86_64-unknown-linux-gnu target and compare their generated assembly. Trace the tail-call decision for the `ret void` and `unreachable` cases; done means both cases produce equivalent tail-call assembly without changing correctness.

Written by the indexing model from the issue text.