https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98526
Bug ID: 98526
Summary: [11 Regression] Double-counting of reduction cost
Product: gcc
Version: 11.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: rsandifo at gcc dot gnu.org
CC: rguenth at gcc dot gnu.org
Target Milestone: ---
The vect_body cost of most reductions is added twice,
once in vect_model_reduction_cost and once in whichever
vectorizable_* routine handles the reduction statement.
E.g. for:
long
f (long *x, int n)
{
long res = 0;
for (int i = 0; i < n; ++i)
res += x[i];
return res;
}
we have:
0x3c40a60 _4 + res_14 1 times scalar_to_vec costs 1 in prologue
0x3c40a60 _4 + res_14 1 times vector_stmt costs 1 in body <----
0x3c40a60 _4 + res_14 1 times vector_stmt costs 1 in epilogue
0x3c40a60 _4 + res_14 1 times vec_to_scalar costs 2 in epilogue
0x3c40a60 *_3 1 times vector_load costs 1 in body
0x3c40a60 _4 + res_14 1 times vector_stmt costs 1 in body <----
I don't yet have a case where this tips the balance though.