Missed loop-invariant PHI hoisting in LICM

**Godbolt:** https://godbolt.org/z/Mb4oE85x5

LLVM leaves loop-invariant control flow and a PHI inside the loop in this testcase.

I found this testcase in Folly's MurmurHash implementation, which uses the `partialLoadUnaligned` function (the Godbolt testcase above is a reduced version):
https://github.com/facebook/folly/blob/1e0a1b4dd4cebceef64c76bf63ed8be1984474fa/folly/lang/Bits.h#L517

GCC hoists the corresponding address/shift calculations out of the loop, while Clang recomputes them on every iteration.

Enabling:

```text
-licm-control-flow-hoisting
```

allows LICM to hoist the relevant control flow/PHI. However, this option is currently disabled by default.

From the commit history, it appears that enabling control-flow hoisting by default originally exposed a significant compile-time regression. However, that regression seems to have been fixed:
* https://bugs.llvm.org/show_bug.cgi?id=39836
* https://github.com/llvm/llvm-project/commit/a7eb2c863fa9c55ee60f37a4ad6846cd2bbeec67

Is there a remaining reason for keeping `licm-control-flow-hoisting` disabled by default?
If so, what would be needed to enable it by default?

Start with the reduced Godbolt testcase and the LICM handling controlled by -licm-control-flow-hoisting. Review the linked Bug 39836 and LLVM commit a7eb2c863fa9c55ee60f37a4ad6846cd2bbeec67 to determine whether the compile-time regression remains. Done means identifying any remaining reason the option is disabled and the requirements for enabling it by default.

Written by the indexing model from the issue text.