Thanks Robin for fixing it.
- : cond (cond_in), else_value (else_value_in)
+ : cond (cond_in), else_value (else_value_in), len (NULL_TREE),
+ bias (NULL_TREE)It seems that you shouldn't include this fix in the patch?
+
+ if (len)
+ {
+ /* If we had a COND_LEN before we need to ensure that it stays that
+ way. */
+ gimple_match_op old_op = *res_op;
+ *res_op = cond_op;
+ maybe_resimplify_conditional_op (seq, res_op, valueize);
+
+ auto cfn = combined_fn (res_op->code);
+ if (internal_fn_p (cfn)
+ && internal_fn_len_index (as_internal_fn (cfn)) != -1)
+ return true;
+
+ *res_op = old_op;
+ return false;
+ }
+ else
+ {
+ *res_op = cond_op;
+ maybe_resimplify_conditional_op (seq, res_op, valueize);
+ return true;
+ }
This looks odd to me.
Currently, we never has cond_len_xxx with dummy length (length = VF) and we
always use cond_xxx if we don't have a loop mask.
So, the length of cond_len_xxx is always generated by MIN or SELET_VL.
I think we don't need the gimple simplification like cond_len -> into argument
value.
But we need this following optimization:
negate + cond_len_fma -> cond_len_fnma/cond_len_fms/cond_len_fnms.
That's what I want to support in gimple fold.
Let's see more comments from Richard and Richi.
juzhe.zh...@rivai.ai
