https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97064
Bug ID: 97064
Summary: BB vectorization behaves sub-optimal
Product: gcc
Version: 11.0
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: ---
The testcase g++.dg/vect/slp-pr87105.cc ends in
_64 = MIN_EXPR <_32, _87>;
bBox_6(D)->x0 = _64;
_67 = MIN_EXPR <_33, _86>;
bBox_6(D)->y0 = _67;
_70 = MAX_EXPR <_36, _87>;
bBox_6(D)->x1 = _70;
_73 = MAX_EXPR <_39, _86>;
bBox_6(D)->y1 = _73;
thus feeding a 4 element store with a non-uniform SLP opportunity
starting with { MIN, MIN, MAX, MAX }. With 2-element vector type
vectorization this eventually gets vectorized by splitting the group
which is prioritized over just building the { MIN..., MAX } vector
from scalars but with 4-element vector type vectorization no splitting
is considered and we end up successfully vectorizing just the store
with never considering the smaller vector size.
So at the moment the testcase PASSes with SSE but fails with AVX.