@@ -3971,15 +3971,20 @@ get_vector_binary_rtx_cost (rtx x, int
scalar2vr_cost)
rtx op_0;
rtx op_1;
- if (GET_CODE (x) == UNSPEC)
- {
- op_0 = XVECEXP (x, 0, 0);
- op_1 = XVECEXP (x, 0, 1);
- }
- else
+ switch (GET_CODE (x))
{
- op_0 = XEXP (x, 0);
- op_1 = XEXP (x, 1);
+ case REG:
+ return COSTS_N_INSNS (1);
+ case VEC_DUPLICATE:
+ return (scalar2vr_cost + 1) * COSTS_N_INSNS (1);
+ case UNSPEC:
+ op_0 = XVECEXP (x, 0, 0);
+ op_1 = XVECEXP (x, 0, 1);
+ break;
+ default:
+ op_0 = XEXP (x, 0);
+ op_1 = XEXP (x, 1);
+ break;
Generally, the patch looks reasonable to me but costing gets more confusing by
the day :)
Originally we used get_vector_binary_rtx_cost just for binary ops but now it
also takes regs, vec_duplicates etc.? And merge is not really a binary op
either, is it?
The whole idea of rtx costing is to use the recursive properties and build a
"cost tree". Generally, isn't it sufficient to find a (vec_duplicate: scalar)
anywhere and then be done? The only distinction we need is whether we're
costing a real vmv.v.x or any other instruction.
As we're not recursively traversing, does something like
FOR_EACH_SUBRTX (...)
if (GET_CODE (*iter) == VEC_DUPLICATE)
help, regardless of the actual operation?
--
Regards
Robin
