@Kai771: Try with -mtune set, if it wasn't already done... It should be overriden by -march, but who knows...
Oh, then it's just different/new optimization flag because of the newer GCC version
I'm pretty sure Amazon's stuff is using ARM instructions, not thumb. (Thumb1 on the K3 isn't really interesting performance-wise anyway).
Yes, the vectorize stuff is probably only used on armv7-a/neon anyway
No idea about mtls-dialect, but gnu is the default anyway