https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85563
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |amacleod at redhat dot com,
| |rguenth at gcc dot gnu.org
Last reconfirmed|2021-04-15 00:00:00 |2023-3-30
--- Comment #23 from Richard Biener <rguenth at gcc dot gnu.org> ---
Re-confirmed. On the testcase in comment#6 we have before EVRP
<bb 2> :
Vframe_list.0_1 = Vframe_list;
a.1_17 = (long int) Vframe_list.0_1;
_15 = (unsigned int) a.1_17;
_13 = _15 & 7;
_10 = _13 == 3;
if (_13 != 3)
goto <bb 3>; [INV]
else
goto <bb 4>; [INV]
<bb 3> :
__builtin_unreachable ();
<bb 4> :
# frame1_6 = PHI <frame1_12(D)(2), frame1_16(5)>
# tail_7 = PHI <Vframe_list.0_1(2), tail_18(5)>
a.1_19 = (long int) tail_7;
_20 = (unsigned int) a.1_19;
_21 = _20 & 7;
_22 = _21 == 3;
if (_21 == 3)
goto <bb 5>; [INV]
else
goto <bb 6>; [INV]
<bb 5> :
_25 = tail_7 + 18446744073709551613;
_26 = __builtin_assume_aligned (_25, 8);
frame1_16 = _26->u.s.car;
tail_18 = _26->u.s.u.cdr;
goto <bb 4>; [INV]
<bb 6> :
do_switch_frame (frame1_6);
and we fail to realize we are never exiting the loop in the first iteration.
The diagnostic happens because we run into loop header copying which
then results in a jump around the loop with an obviously uninitialized
argument to frame1_6. And that's because we have thrown away the
__builtin_unreachable before loop header copying (in VRP1, after the first
threadfull).
Re-ordering CH before VRP fails to optimize even when I also put a copyprop
between CH and VRP. We then see
<bb 2> [local count: 118111600]:
# PT = nonlocal escaped null
Vframe_list.0_1 = Vframe_list;
a.1_12 = (long int) Vframe_list.0_1;
_10 = (unsigned int) a.1_12;
# RANGE [irange] Lisp_Type [0, 7] NONZERO 0x7
_8 = _10 & 7;
if (_8 != 3)
goto <bb 3>; [0.00%]
else
goto <bb 4>; [100.00%]
<bb 3> [count: 0]:
__builtin_unreachable ();
<bb 4> [local count: 118111600]:
a.1_21 = (long int) Vframe_list.0_1;
_22 = (unsigned int) a.1_21;
# RANGE [irange] Lisp_Type [0, 7] NONZERO 0x7
_23 = _22 & 7;
if (_23 == 3)
but for some reason VRP doesn't optimize the second compare. We'd probably
need to value-number here (loop header copying does, but only parts of the
function and not including this leading block).
So somewhat of a pass ordering issue. The next CSE is DOM and then PRE.
