https://gcc.gnu.org/bugzilla/show_bug.cgi?id=127909
Bug ID: 127909
Summary: Arm should detect sdot_prod and use smlad/smlald for
32/64 bit accumulation.
Product: gcc
Version: 17.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: rearnsha at gcc dot gnu.org
Target Milestone: ---
The function
typedef short v2hi __attribute__((vector_size(4)));
typedef int T;
T foo(v2hi a, v2hi b, T acc)
{
return acc + a[0]*b[0] + a[1]*b[1];
}
(and similarly if T is 'long long')
is a vector length 2 sdot_prod idiom and should optimise to use smlad (smlald).
But the idiom is never recognised during SLP.
Note that despite this there are further missing components in the back-end,
but the start of the problem is failing to detect the idiom.
