Then you aren't multiplying doubles, it would be a shame to get a completely wrong answer on an ill-conditioned problem because the implementation disregarded the spec, cf. "the fastest way to process UTF-8 is to truncate it to ASCII", "the fastest RSA encryption is ROT-13". Also note that floats are only faster with packed instructions (often not possible without changing APIs) or when memory bandwidth is an issue.
(Although it seems like clang does something special here, as I don't see many double-to-float conversions in its output.)