Re: [PATCH 0/1] target/arm: Check NaN mode before silencing NaN

[Peter Maydell]

On Mon, 28 Jun 2021 at 16:05, Richard Henderson
<richard.henderson@linaro.org> wrote:
>
> On 6/28/21 7:54 AM, Peter Maydell wrote:
> > Richard, Alex: what is the assertion trying to achieve ? It doesn't
> > seem entirely obvious to me that because we're in default-NaN mode
> > (which is a property of the *output* of FPU insns) that we should
> > blow up on calling float*_silence_nan() (which is typically an action
> > performed on the *input* of FPU insns).
>
> This was in response to e9e5534ff30.
>
> My assumption in adding the assert is that it was probably a configuration 
> error.  If you
> disagree, I suppose we can revert it, as it's not critical.

I just ran across this again, in a different context. For MVE VMAXNMV
(which uses the "default FPSCR value"), I need to silence input SNaNs
before performing the max/min operation. The logical way to do that is
to call float*_silence_nan(). Except that that barfs on this assertion.

So I think that having run into this assertion() twice now it's
more awkward than helpful and I intend to put a patch deleting it
in the appropriate part of my next MVE series.

(In theory I could work around it by deliberately squashing the
"use the default NaN value flag" in a local copy of the fp_status,
but that seems like unnecessary work.)

thanks
-- PMM