https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84037
--- Comment #12 from Richard Biener <rguenth at gcc dot gnu.org> ---
I have opened PR84102 for the missed optimizations in this particular loop. I
believe now the interesting one is the other.
30.25% a.out a.out [.] __solv_cap_MOD_fourir2d
24.83% a.out a.out [.] __solv_cap_MOD_fourir
24.83% a.out a.out [.] __solv_cap_MOD_fourir2dx
18.78% a.out [unknown] [k] 0xffffffff813366e7
and the 551 loops are in the innermost nest of fourir() which is called/inlined
to fourir*. It's actually not array expressions but the different inline
copies we vectorize. Cost model for that one:
capacita2.f90:551:0: note: Cost model analysis:
Vector inside of loop cost: 3756
Vector prologue cost: 64
Vector epilogue cost: 1712
Scalar iteration cost: 516
Scalar outside cost: 4
Vector outside cost: 1776
prologue iterations: 0
epilogue iterations: 4
Calculated minimum iters for profitability: 0
capacita2.f90:551:0: note: Runtime profitability threshold = 8
capacita2.f90:551:0: note: Static estimate profitability threshold = 11545608
the loop is hybrid SLP, VF is 8
capacita2.f90:551:0: note: improved number of alias checks from 36 to 6
so here it's also dependence analysis breaking down because the step is
unknown:
(compute_affine_dependence
stmt_a: _170 = REALPART_EXPR <*a.0_107[_54]>;
stmt_b: _196 = REALPART_EXPR <*a.0_107[_73]>;
(analyze_overlapping_iterations
(chrec_a = 0)
(chrec_b = 0)
(overlap_iterations_a = [0])
(overlap_iterations_b = [0]))
(analyze_overlapping_iterations
(chrec_a = {((integer(kind=8)) j0_119 + 1) * iftmp.476_91, +,
iftmp.476_91}_3)
(chrec_b = {((integer(kind=8)) j1_124 + 1) * iftmp.476_91, +,
iftmp.476_91}_3)
(analyze_siv_subscript
siv test failed: unimplemented)
(overlap_iterations_a = not known)
(overlap_iterations_b = not known))
) -> dependence analysis failed
the only thing we know about iftmp.476_91 is that it isn't zero (but I'm sure
dependence analysis doesn't use that fact). I think this specific dependence
should be computable...? [by just dividing both chrecs by iftmp.476_91?]