Richard Henderson <richard.hender...@linaro.org> writes:
> On 6/7/21 8:35 AM, Alex Bennée wrote: >> So I'm guessing Emilio had the original flush code split was to avoid >> multiple checks against s->flush_inputs_to_zero in the code. The was >> possibly a good reason, comparing the before/after of float32_mul: > > I assumed that the most important thing now is that we test > floatN_is_denormal only once -- the test for flush_inputs_to_zero is > fairly trivial. > > If you've got a better ordering of operations for this, do tell. What I really want is to know which instructions translate into the if (s->flush_inputs_to_zero) and verifying that is only checked once. Maybe I'm just suspicious of compilers ability to optimise things away... -- Alex Bennée