https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120398
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Component|tree-optimization |target
Target Milestone|--- |15.2
Keywords| |missed-optimization
--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
t.c:8:12: note: vect_is_simple_use: operand r1_25 = PHI <_11(7), 0.0(6)>,
type of def: reduction
t.c:8:12: missed: reduc op not supported by target.
t.c:8:12: missed: in-order unchained SLP reductions not supported.
so this is a special-case we do not support optimally. Note that in-order
reductions, esp with the high VF resulting, are necessarily "bad", but
costing still thinks it's profitable:
t.c:8:12: note: Cost model analysis:
Vector inside of loop cost: 376
Vector prologue cost: 0
Vector epilogue cost: 336
Scalar iteration cost: 80
Scalar outside cost: 32
Vector outside cost: 336
using -ffast-math you get nice code. aarch64 doesn't vectorize the loop
at -O2 for me unless enabling SVE which has in-order reductions.
IMO a costing issue only, if it is slower than not vectorizing (didn't try to
actually measure).