> ... here we probably get PLUS_EXPR for MINUS_EXPR above but IIRC
> for MINUS_EXPR the !as_initial case should return positive zero.
>
> Can you double-check?
You're referring to the canonicalization from a - CST to a + -CST so
that the neutral op would need to change with it? Argh, good point.
>From what I can tell the only difference for MINUS_EXPR is that we
negate the reduction operand and then just continue as if it were
a PLUS_EXPR (which is the right thing to do also for +-0.0?).
At least I didn't observe a canonicalization and we don't call
neutral_op_for_reduction in between.
What we do have, though, is for the fully-masked case (you added
that recently):
if (LOOP_VINFO_FULLY_MASKED_P (loop_vinfo))
{
vector_identity = build_zero_cst (vectype_out);
if (!HONOR_SIGNED_ZEROS (vectype_out))
;
else
{
gcc_assert (!HONOR_SIGN_DEPENDENT_ROUNDING (vectype_out));
vector_identity = const_unop (NEGATE_EXPR, vectype_out,
vector_identity);
}
}
So for
/* Handle MINUS by adding the negative. */
if (reduc_fn != IFN_LAST && code == MINUS_EXPR)
{
tree negated = make_ssa_name (vectype_out);
We might need a similar assert
gcc_assert (HONOR_SIGNED_ZEROS (vectype_out)
&& !HONOR_SIGN_DEPENDENT_ROUNDING (vectype_out));?
Apart from that the only call with !as_inital is in
vect_create_epilog_for_reduction. I just instrumented it with an
assert (false) but i386.exp doesn't trigger it at all.
Regards
Robin
