[llvm] Avoid dbg.value between phi nodes when converting from DbgRecords (PR #83620)

[Jeremy Morse via llvm-commits]

jmorse wrote:

Thanks for chasing all that down,

> At first glance that doesn't look wrong but getFirstNonPHI() and getFirstNonPHIIt() return different positions despite the names implying they only differ by return type. Additionally, SetInsertPoint has overloads for both Instruction* and BasicBlock::iterator so it's not an error to use the wrong one.

Curses, yes, that's another set of API calls we'll need to work on. We got caught by all this late in the day and were hoping to make the transition completely transparent. Apologies for the disruption,

> Also, it's a lot easier to see what's going now that we have the record printing but one hack I still had to do to help me track it down was to stop the pass manager from converting back to the intrinsics so I could see the new records in -print-before-all/-print-after-all.

CC @SLTozer , IIRC print-before/after is due to print debug-records around about now (or last Friday or something)?

https://github.com/llvm/llvm-project/pull/83620