Use unambiguous inline assembly to load a float variable. GNU as
silently converts 'fld' to 'flds', without taking the actual variable
type into account (!), but clang's integrated assembler rightfully
complains about it.

Discussed with: cperciva