[cozzyd]

Well in most cases I've looked at generated assembly (not that often), the xmm registers are used even for scalar operations, which I thought was the default option for gcc on x86-64, but I suppose it might differ on different systems (or perhaps 32-bit mode was used for some reason).

[cperciva]

Right, if you're using xmm regs you're getting double or single precision. Sometimes the x87 regs get used and that's when "accidentally computed with extra precision and then double-rounded" comes up.