[llvm] [SROA] Unfold gep of index phi (round 2) (PR #83494)

[Arthur Eubanks via llvm-commits]

aeubanks wrote:

> Q: why not put those GEPs into appropriate incoming basic blocks. That would be a smaller-scope unfolding. I suspect those geps will be sunk into appropriate basic blocks, but it may be better if we don't have to do that.

It doesn't really matter where the GEP is, later optimizations will place them wherever optimal.

If we want to support this unfolding for GEPs on top of allocas/constants/args, it's trickier
```
bb1:
bb2:
bb3:
  %i = phi [ 4, %bb1 ], [ 8, %bb2 ]
  %g1 = gep ...
  %g2 = gep %g1, %i
```
can't become
```
bb1:
  %g_bb1 = gep %g1, 4
bb2:
  %g_bb2 = gep %g1, 8
bb3:
  %g2 = phi [ %g_bb1, %bb1 ], [ %g_bb2, %bb2 ]
  %g1 = gep ...
```
since `%g1` is defined later than its uses. So I've been trying to keep the insertion point simple.

https://github.com/llvm/llvm-project/pull/83494