https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110222
Bug ID: 110222
Summary: Inefficient fully masked loop vectorization with
AVX512
Product: gcc
Version: 14.0
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: ---
gfortran.dg/matmul_2.f90 with -march=znver4 --param vect-partial-vector-usage=2
shows the
c(:,1:7:2) = matmul(a,b(:,1:7:2))
innermost loop vectorized with
note: vectorization_factor = 16, niters = 2
which means a statically known loop mask which is even power-of-two. This
should be instead vectorized without masking and V2SImode vectors. Similarly
for a theoretical niters = 3 or a niters < 4 this should use a smaller
(but masked) vector mode for vectorization, not the target preferred 512bit
size.
The x86 target currently chooses to not get costs with different vector modes
compared but in these cases statically selecting a better mode should be
possible.
