> -----Original Message----- > From: Richard Henderson <richard.hender...@linaro.org> > Sent: Wednesday, February 16, 2022 4:53 PM > To: Taylor Simpson <tsimp...@quicinc.com>; qemu-devel@nongnu.org > Cc: f4...@amsat.org; a...@rev.ng; Brian Cain <bc...@quicinc.com>; Michael > Lambert <mlamb...@quicinc.com> > Subject: Re: [PATCH v3] Hexagon (target/hexagon) properly handle NaN in > dfmin/dfmax/sfmin/sfmax > > On 2/16/22 15:39, Taylor Simpson wrote: > > The float??_minnum implementation differs from Hexagon for SNaN, it > > returns NaN, but Hexagon returns the other input. So, we use > > float??_minimum_number. For double precision, we check for QNaN and > > raise the invalid flag. > > I'm surprised that the behaviour for double differs from single, but the docs > are light on the subject. Anyway, > > Reviewed-by: Richard Henderson <richard.hender...@linaro.org>
The dfmin/dfmax were added in a later version of the hardware than sfmin/sfmax. I was seeing the different behavior on the reference simulator. I was able to get my hands on the newer hardware to run the test and found that the reference simulator behaves differently. I'll respin this patch to match the behavior of the hardware. Thanks, Taylor
