I'd really appreciate a sceptical eye on this. I'm not 100% sure this is correct.
The motivation is that unswitching a loop nest will produce conditions in the outer loop which are frozen. If that exit would otherwise be computable, we'd really like it to remain computable w/freeze.
My hesitation is that I'm not sure it's sound to propagate an analysis result (potentially based on UB) through freeze. What do folks think? Is this legal? Or not?
If not, suggestions for alternate approaches to recover knowledge in SCEV?