As Alina mentioned, there are several issues (miscompilations, termination problems, and the lack of a PRE stage). In my GSoC project, I upstreamed fixes for some of the miscompilation and termination issues; however, others still require further testing. Additionally, I implemented PRE, which is one of the main blockers for adopting NewGVN. The code still needs some refinement and testing before it can be upstreamed.
One of the advantages of using NewGVN over GVN is its ability to detect cyclic equivalences—that is, two values that are equivalent in any given loop iteration (issue). This is due to the optimistic assumption, that allows NewGVN to reason inductively about cyclic values. Retrofitting this into GVN is very impractical. For more details, feel free to ask or check my report: GSoC Report 2024.