Denormals are not flushed to zero merely by `mov xmm,rax` - they can't be, since xmm is also used for integer arithmetic, and the CPU doesn't "know" whether it's an FP or integer value at the point of the mov.
Your gist demonstrates this: notice in the bottom right that xmm0 contains `v2_int64 = {0x1, 0x0}` so the bottom half containing the result is nonzero. (The top half is zero as it's unused).
The "info" command rounds the value to 0 for compactness, but if you run `p $xmm0.v2_double` you will also see it is nonzero when interpreted as a double (I get 4.94e-324).
Your gist demonstrates this: notice in the bottom right that xmm0 contains `v2_int64 = {0x1, 0x0}` so the bottom half containing the result is nonzero. (The top half is zero as it's unused).
The "info" command rounds the value to 0 for compactness, but if you run `p $xmm0.v2_double` you will also see it is nonzero when interpreted as a double (I get 4.94e-324).
A fun aside: that `mulss` instruction you stopped at is just there to raise an underflow FP exception; its result is discarded. https://github.com/bminor/glibc/blob/595c22ecd8e87a27fd19270...