[llvm] [DebugInfo][InstCombine] Do not overwrite prior DILocation for new Insts (PR #108565)

[Jeremy Morse via llvm-commits]

jmorse wrote:

(the "unread" list in my inbox is looong,)

I'm not fully understanding the second paragraph about the bad case: in one mode the DILocation of a PHI is given to the new load instruction, and that's considered bad. In the second mode, there's a risk the new load "...will thus receive the DILocation from the PHI" -- so the same location as before, so there's no regression?

In principle I think this is a good patch that tries that keep source locations as specific as possible, I just want to understand the use case more.

https://github.com/llvm/llvm-project/pull/108565