https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92266
Bug ID: 92266
Summary: Duplicate code generation for live stmts from SLP
Product: gcc
Version: 10.0
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: ---
int foo (int * __restrict__ a, int * b, int n)
{
int tem1, tem2;
for (int i = 0; i < n; ++i)
{
tem1 = a[i*2 + 0] * 2;
tem2 = a[i*2 + 1] * 2;
b[i*4 + 0] = tem1;
b[i*4 + 1] = tem1;
b[i*4 + 2] = tem2;
b[i*4 + 3] = tem2;
}
return tem1 + tem2;
}
shows
<bb 10> [local count: 105119324]:
# tem1_55 = PHI <tem1_35(3)>
# tem2_54 = PHI <tem2_36(3)>
# vect_tem1_35.9_53 = PHI <vect_tem1_35.9_65(3)>
_61 = BIT_FIELD_REF <vect_tem1_35.9_53, 32, 96>;
_62 = BIT_FIELD_REF <vect_tem1_35.9_53, 32, 64>;
_63 = BIT_FIELD_REF <vect_tem1_35.9_53, 32, 32>;
_64 = BIT_FIELD_REF <vect_tem1_35.9_53, 32, 0>;
goto <bb 4>; [100.00%]
<bb 6> [local count: 850510900]:
goto <bb 3>; [100.00%]
<bb 4> [local count: 118111601]:
# tem1_45 = PHI <tem1_29(D)(2), _64(10)>
# tem2_46 = PHI <tem2_30(D)(2), _62(10)>
_33 = tem1_45 + tem2_46;
which is because we iterate like
FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (slp_node), i, slp_stmt_info)
{
if (STMT_VINFO_LIVE_P (slp_stmt_info)
&& !vectorizable_live_operation (slp_stmt_info, gsi, slp_node,
slp_node_instance, i,
vec_stmt_p, cost_vec))
return false;
}
so for stmts appearing multiple times we code-gen the live operation multiple
times. This is even worse for stmts appearing in multiple SLP nodes.
Luckily the code is all dead in the end.
