On 8/7/24 1:25 AM, Manolis Tsamis wrote:
On Wed, Aug 7, 2024 at 8:57 AM Sam James <s...@gentoo.org> wrote:
Philipp Tomsich <philipp.toms...@vrull.eu> writes:
Hi,
Sam, Jakub & Robin,
We had an "OK for trunk" from Jeff for v4 (see
https://gcc.gnu.org/pipermail/gcc-patches/2024-July/656907.html) and
it has been two more weeks for this RESEND.
I'll push this by end of this week unless I hear otherwise.
I'd ping Jeff for a quick re-review of v5 to make sure he's happy with
the changes to address
https://gcc.gnu.org/pipermail/gcc-patches/2024-July/656908.html.
To answer Jeff's question, although that patch removes the checks for
particular SRC codes, it adds noce_operand_ok (src) which checks
side_effects_p (src) and may_trap_p (src). This handles the asm
related things and is in line with what the rest of ifcvt does, so
there was no change needed for that comment.
Excellent. Thanks.
As noted elsewhere, I see some evidence that the conditions we're using
for "noce_can_force_operand" aren't really sufficient to tell us if we
can force an operand into a register. Case in point is the v850 port
where we can't force certain rotates into a register using force_operand.
I've got a workaround as the port can be slightly improved and that
improvement will make the problem go away. But I do suspect we're going
to see additional issues in this space.
jeff
