On 10/14/22 15:21, Koning, Paul wrote:
On Oct 14, 2022, at 5:15 PM, Jeff Law via Gcc-patches <gcc-patches@gcc.gnu.org>
wrote:
On 10/14/22 11:36, Koning, Paul wrote:
On Oct 14, 2022, at 1:10 PM, Jeff Law <jeffreya...@gmail.com> wrote:
On 10/14/22 10:37, Koning, Paul wrote:
...
But that approach falls down with reload/lra doing substitutions without
validating the result. I guess it might be possible to cobble together
something with secondary reloads, but it's way way way down on my todo list.
Aren't the constraints enforced? My experience is that I was getting these bad
addressing modes in some test programs, and that the constraints I created to
make the requirement explicit cured that. Maybe I'm expecting too much from
constraints, but my (admittedly inexperienced) understanding of them is that
they inform reload what sort of things it can construct, and what it cannot.
It's not really a constraint issue -- the pattern's condition would cause this
not to recognize, but LRA doesn't re-recognize the insn. We might be able to
hack something in the constraints to force a reload of the source operand in
this case. Ugly, but a possibility.
I find it hard to cope with constraints that don't constrain. Minimally it
should be clearly documented exactly what cases fail to obey the constraints
and what a target writer can do to deal with those failures.
Constraints have a purpose, but as I've noted, they really don't come into play here.
Had LRA tried to see if what it created as a valid move insn, the backend would have said
"nope, that's not valid". That's a stronger test than checking the
constraints. If the insn is not valid according to its condition, then the constraints
simply don't matter.
I'm not aware of a case where constraints are failing to be obeyed and
constraints simply aren't a viable solution here other than to paper over the
problem and hope it doesn't show up elsewhere.
Right now operand 0's constraint is "<" meaning pre-inc operand, operand 1 is
"r". How would you define a new constraint for operand 1 that disallows overlap with
operand 0 given that the H8 allows autoinc on any register operand? You can't look at operand 0
while processing the constraint for operand 1. Similarly if you try to define a new constraint for
operand0 without looking at operand1.
Easy but cumbersome: define constraints for "register N" (for each N) and another set for "autoinc on
any register other than N". In pdp11, I called these Z0, Z1... and Za, Zb... respectively. Then the insn gets
constraints that look like "Z0,Z1,Z2..." and "Za, Zb, Zc..." for the two operands. As I said, see
pdp11.md, the mov insn.
It generally looks sound, but golly gee, this runs into the "reload
doesn't validate insns problem" if it's done in a reload tree rather
than an lra tree. We've got an insn with a pre-inc destination and a
reg source. The source pseudo doesn't get a hard reg, reload replaces
the pseudo with a mem as expected. Reload finishes with something like this:
(insn 100 98 101 15 (set (mem/f:SI (pre_dec:SI (reg/f:SI 7 sp)) [8 S4 A32])
(mem/c:SI (plus:SI (reg/f:SI 7 sp)
(const_int 8 [0x8])) [9 %sfp+-24 S4 A32])) "j.c":62:11
19 {*movsix}
(expr_list:REG_ARGS_SIZE (const_int 4 [0x4])
(nil)))
Which, isn't a valid instruction on the H8. The insn's condition
verifies that one of the two operands must be a REG. But reload never
bothered to re-recognize the insn after makng the substitution, then
naturally it blows up in reload_cse with a constraint failure because
the pre-inc destination constraints require a reg for the source
operand. But the real culprit here is reload making the substitution
and not validing that the result is valid.
Arggh!
Which brings me back to pondering just removing the autoinc magic
checking in the H8 port :-) I've actually got that spinning in the
tester just to see if there's any obvious fallout. I've already spent
more time on this than I can reasonably justify.
Jeff
