It’s absolutely harmful. It turns computations that would be guaranteed to be exact (e.g. head-tail arithmetic primitives used in computational geometry) into “maybe it’s exact and maybe it’s not, it’s at the compiler’s whim” and suddenly your tests for triangle orientation do not work correctly and your mesh-generation produces inadmissible meshes, so your PDE solver fails.
Thank you, I found this hint very interesting. Is there a source you wouldn't mind pointing me to for those "head, tail" methods?
I am assuming it relates to the kinds of "variable precision floating point with bounds" methods used in CGAL and the like; Googling turns up this survey paper:
