On Fri, 11 Aug 2023, Richard Biener wrote:
> > I think it converts SNan to QNan (when the partial vector has just one
> > element which is SNan), so is a test for -fsignaling-nans missing?
>
> Hm, I guess that's a corner case that could happen when there's no
> runtime profitability check on more than one element and when the
> element accumulated is directly loaded from memory. OTOH the
> loop vectorizer always expects an initial value for the reduction
> and thus we perform either no add (when the loop isn't entered)
> or at least a single add (when it is). So I think this particular
> situation cannot occur?
Yes, that makes sense, thanks for the elaboration.
(it's a bit subtle so maybe worth a comment? not sure)
> > In the defaut -fno-rounding-math -fno-signaling-nans mode I think we
> > can do the reduction by substituting negative zero for masked-off
> > elements ? maybe it's worth diagnosing that case separately (i.e.
> > as "not yet implemented", not an incorrect transform)?
>
> Ah, that's interesting. So the only case we can't handle is
> -frounding-math -fsigned-zeros then. I'll see to adjust the patch
> accordingly, like the following incremental patch:
Yeah, nice!
> > (note that in avx512 it's possible to materialize negative zeroes
> > by mask with a single vpternlog instruction, which is cheap)
>
> It ends up loading the { -0.0, ... } constant from memory, the
> { 0.0, ... } mask is handled by using a zero-masked load, so
> indeed cheaper.
I was thinking it could be easily done without a memory load,
but got confused, sorry.
Alexander
