https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119057
--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
So there's code in check_reduction_path that checks for additional uses of the
path defs that explicitly allows out-of-loop uses for the "tail":
/* Check there's only a single stmt the op is used on. For the
not value-changing tail and the last stmt allow out-of-loop uses.
??? We could relax this and handle arbitrary live stmts by
forcing a scalar epilogue for example. */
...
else if (!is_gimple_debug (op_use_stmt)
&& (*code != ERROR_MARK
|| flow_bb_inside_loop_p (loop,
gimple_bb (op_use_stmt))))
FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter)
cnt++;
we have
<bb 4> [local count: 955630224]:
# a.1_28 = PHI <_2(9), a.1_6(3)>
# b_lsm.15_27 = PHI <_25(9), b_lsm.15_33(3)> <---
# c_lsm.17_3 = PHI <_20(9), c_lsm.17_26(3)>
<bb 5> [local count: 7731917314]:
# d.6_31 = PHI <_14(10), 0(4)>
# b_lsm.15_13 = PHI <_12(10), b_lsm.15_27(4)> <---
# ivtmp_23 = PHI <ivtmp_22(10), 2(4)>
b.3_9 = (unsigned int) b_lsm.15_13; <---
_11 = b.3_9 | e.4_10; <--- (*)
_12 = (int) _11; <---
_14 = d.6_31 + 1;
ivtmp_22 = ivtmp_23 - 1;
if (ivtmp_22 != 0)
goto <bb 10>; [89.00%]
else
goto <bb 6>; [11.00%]
<bb 6> [local count: 955630224]:
# _51 = PHI <_11(5)> (*)
# _25 = PHI <_12(5)> <---
c.9_18 = (unsigned int) c_lsm.17_3;
_19 = _51 | c.9_18;
with the (*) def being used outside of (the inner) loop.
That's undesirable behavior for the double-reduction case. Testing a patch.
