[llvm] [SROA] Unfold gep of index phi (PR #83087)

[Arthur Eubanks via llvm-commits]

aeubanks wrote:

the issue causing the revert is that I assumed SROA only handled allocas at the very beginning of the entry block, but it actually handles all allocas in the entry block, so we might insert a GEP before an alloca if there is a non-alloca instruction between the alloca and the function entry

https://github.com/llvm/llvm-project/pull/83087