On 04/27/2015 12:29 PM, Richard Biener wrote:
On April 27, 2015 6:24:47 PM GMT+02:00, Jeff Law <l...@redhat.com>
n & -n will (of course) get computed into an SSA_NAME. We then
propagate that name for the use of "n" in the return statement
rather than using the effectively zero cost "n".
If we put some smarts into tree_swap_operands_p to order sensibly
in this kind of case, we end up regressing a different case that
I'll be looking at today.
In this case the temporary we propagate has a single use (in the
comparison). Might be interesting to disallow this case by extra
checks in record equality. I wouldn't change tree_swap_operands_p.
Yea, I keep going back and forth over whether or not the heuristic
should go in tree_swap_operand_p or in record_equality.
It's a fairly narrow issue, so maybe record_equality would be a better spot.
jeff
