[llvm] [CodeGen] Use first EHLabel as a stop gate for live range shrinking (PR #114195)

[via llvm-commits]

MuellerMP wrote:

> What is the context for this pass? I didn't know it existed, and I see it's only added by x86. Why does it exist? This is just an extra machine scheduler?

That's a good question, because the machine scheduler actually doesn't run into this problem due to running after phi elimination.
So I read the original [MR](https://reviews.llvm.org/D32563) and this pass seems to relieve register pressure in some special cases: 
1) when reassociation reorders ir instructions in a bad way
2) when user input introduces live-range issues
It also seemd to require insight of the whole basicblock, which is why it couldn't be solved with exisiting generic schdeduling.

I wonder if these assumptions still hold in the current version of misched and resassociate.
Since I'm not to familiar with either of them I propose to fix the current version of the pass for now and leave the analysis regarding the redundancy of the pass for followups.

https://github.com/llvm/llvm-project/pull/114195