On 02/23/2018 03:59 AM, Alex Bennée wrote: >> Not using float16_eq etc? > > These don't actually exist.
Ah. > But I guess we could make stubs for them > based on the generic float_compare support. But would it buy us much? ... >>> + return ADVSIMD_CMPRES(compare == float_relation_greater || >>> + compare == float_relation_equal); >> >> Especially float16_le(b, a, fpst). It buys us knowledge of the float_relation_* values, such that instead of the two comparisons above you can use <= 0 (note that this only works for le not ge, because of float_relation_unordered == 2). I'll grant you that two compares vs one isn't much, but it is simpler... r~