In reply to rjek:
I believe we are talking here of an optimisation achieved by giving the compiler licence to use a wider set of instructions (ie including those found only on the XScale). Such an optimisation is quite different to one achieved by changes to the program logic.
Optimisation through logic or algorithm changes can indeed be unwise if done early in the development cycle. I can see no reason why using the compiler's optimisations is a problem for any release though and I think that's what Peter did. I stand by my statement that using all the speed related optimisations is a reasonable approach; indeed, the A9 incompatibility did bring a performance improvement, albeit a tiny one.
My speculation (and I'm no better informed than anyone else here) is that the gain from that particular optimisation was disappointing; the trade-off between improved performance and improved compatibility then became obvious, leading to the A9 compatible release that quickly followed. I could be completely wrong, of course, but I still think rational reasons are more likely than spite from Peter.