question on invalid_flag for bulknormaldot

[Vivvvvvvv]

Hi,

I got a question on invalid flag calculation in the bulknormaldot.h file. Normally, a pos/neg infinite number should be non-NAN and non-Zero. Currently, I am performing instruction vfwldot.vv with sew=16 and lmul=1, and a=0xffd8 b=0x7f80 as one pair of the src。 For the src B, both any_neg_inf and nan are set to 1, which later resulting in invalid_flag being set(as there are other pos_inf srcs ). In my opinion, !NAN should be checked when setting any_pos/neg_inf.

I am not sure if I misunderstood this part. Please correct me if I am wrong.

[Vivvvvvvv]

bool either_inf = a[i].inf() || b[i].inf();
any_pos_inf |= either_inf && a[i].sign() == b[i].sign();
any_neg_inf |= either_inf && a[i].sign() != b[i].sign();

This is the part of code I am questioning about.

[aswaterman]

@nibrunieAtSi5 I agree this looks sketchy; I think we need something like

bool either_inf = a[i].inf() || b[i].inf();
bool either_nan = a[i].nan() || b[i].nan();
any_pos_inf |= either_inf && !either_nan && a[i].sign() == b[i].sign();
any_neg_inf |= either_inf && !either_nan && a[i].sign() != b[i].sign();

Can you confirm and file a PR?

[nibrunieAtSi5]

yes this does not look correct.

I believe the C code is incorrect indeed and will raise spurious invalid flag in the described cases: inf * quietNaN.
The expected result will be NaN in any case, so I believe this issue only affect the flags.