And now... drumroll please... The great answer!!! It is: YMMV

I know 'measure and see', but the point of an article like this is usually
to collate things like that which others have already done so that we can
get an idea of the magnitude of the benefit *before* we download it. No
article can replace real-world tests, but it could at least report on a
few of them.

You've provided me with more of a clue than the article :) I was wondering
how the 'early optimizers have no understanding of register pressure'
problem had affected this: I know it was part of the design intent of IRA
to try to make this problem less severe...