https://gcc.gnu.org/bugzilla/show_bug.cgi?id=124222
Bug ID: 124222
Summary: SLP patterns confuse scalar costing
Product: gcc
Version: 16.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: rguenth at gcc dot gnu.org
Target Milestone: ---
Split out from PR116979.
struct S { __complex__ float f; };
struct S
foo (const struct S *a, const struct S *b)
{
struct S r;
r.f = a->f * b->f;
return r;
}
with -O2 -mfma on x86-64 shows
t.c:8:10: note: Cost model analysis for part in loop 0:
Vector cost: 156
Scalar cost: 152
t.c:8:10: missed: not vectorized: vectorization is not profitable.
which is because the SLP pattern for FMADDSUB makes us only cost part of
the covered scalar stmts. This is because the scalar coverage is computed
from SLP_TREE_SCALAR_STMTS only.
