https://gcc.gnu.org/bugzilla/show_bug.cgi?id=61247
--- Comment #6 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #5)
> int foo (int *p, unsigned long sz)
> {
> int sum = 0;
> for (unsigned i = 0; i < sz; ++i)
> sum += p[i];
> return sum;
> }
>
> is a simpler variant. What works is that we set up to version the loop
> with sz_7(D) + 18446744073709551615 <= 4294967294 and do
> loop_constraint_set (loop, LOOP_C_FINITE), but that is not enough to make
> SCEV analysis succeed for the DR (so we get a gather) or analyze the
> index PHI in the induction defining the supposed gather index:
>
> t.c:4:26: note: Analyze phi: _15 = PHI <_4(6), 0(5)>
> t.c:4:26: missed: reduction used in loop.
> t.c:4:26: missed: Unknown def-use cycle pattern.
> ...
> t.c:4:26: note: worklist: examine stmt: _1 = _15 * 4;
> t.c:4:26: note: vect_is_simple_use: operand _15 = PHI <_4(6), 0(5)>, type
> of def: unknown
> t.c:4:26: missed: Unsupported pattern.
> t.c:5:13: missed: not vectorized: unsupported use in stmt.
> t.c:4:26: missed: unexpected pattern.
>
> I will see to fix this on the SCEV side.
So for this particular case where we have
<bb 3> [local count: 955630224]:
# _15 = PHI <_4(6), 0(5)>
# sum_16 = PHI <sum_10(6), 0(5)>
# i_18 = PHI <i_11(6), 0(5)>
_1 = _15 * 4;
_2 = p_9(D) + _1;
_3 = *_2;
sum_10 = _3 + sum_16;
i_11 = i_18 + 1;
_4 = (long unsigned int) i_11;
if (_4 < sz_7(D))
_15 is a recurrence that fails to analyze in SCEV DFS because of the
widening conversion. Eliding this IV as
<bb 3> [local count: 955630224]:
# sum_16 = PHI <sum_10(6), 0(5)>
# i_18 = PHI <i_11(6), 0(5)>
_15 = (long unsigned int) i_18;
_1 = _15 * 4;
_2 = p_9(D) + _1;
_3 = *_2;
sum_10 = _3 + sum_16;
i_11 = i_18 + 1;
_4 = (long unsigned int) i_11;
if (_4 < sz_7(D))
goto <bb 6>; [89.00%]
would avoid this particular issue.
But it's possible to enhance simplify_peeled_chrec to handle this situation.
This helps up to epilogue peeling which needs to handle "derived IVs" as well.
