https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116684
--- Comment #4 from Feng Xue <fxue at os dot amperecomputing.com> ---
(In reply to Richard Biener from comment #3)
> Since the reduction opportunity is in the unrolled scalar inner loop we'd
> have
> to know how DOT_PROD combines lanes which we do not specify but instead
> expect the whole vector to be reduced to a single lane.
Yes, the VeGen's scheme relies on target-specific lane combining strategy
(every 4-element to 1) of dot-product, which is not described in middle-end.
>
> I think Feng works on related areas, not sure whether exactly covering this
> one.
Could not.
For this case, gcc is in all possibility to unroll the inner loop, while retain
the outer. One complication is that widen_op_pattern may break isomorphism of
inner 4 add-mult statements, which would impact discovery SLP reduction.
An alternative is to match inner 4 add-mul statements as a compound operation
which exactly corresponds to semantic of combining every 4-element to 1, as:
for (int = 0; i < 16; i++) {
output[i] += COMBINE_LANE_4x1(data[0-3], kernel[i][0-3])
}
and then recognize it as a dot-product pattern in tree-vect-pattern stage, both
happens before slp.
>
> Implementation wise this is a (SLP) pattern to recognize.
