On Fri, 3 Oct 2025, Jakub Jelinek wrote:
> On Wed, Oct 01, 2025 at 01:12:38PM +0200, Richard Biener wrote:
> > The finite_loop_p query should already do this (if scev_initialized (),
> > which it is, for CD-DCE).
>
> It does partially.
>
> I can reproduce e.g. on
> -std=c++98 Wstringop-overflow-4.ii -O2 -Wall -Wno-array-bounds
> with Jason's patch which prints 4 warnings instead of the expected
> two (attached).
>
> finite_loop_p is called on 8 loops during cddce, on the 4 inner ones
> it succeeds (despite them not having loop->finite_p set), on the 4 outer
> ones it doesn't.
>
> E.g. below is a loop on which it fails.
> The recorded value range (alread in dse1 and still after cddce1) is
> # RANGE [irange] long int [-1, 0]
> _41 = (long int) _1;
> In any case, the form of the loop with header containing
> # _18 = PHI <_41(19), -1(22)>
> if (_18 == 0)
> exit
> else
> latch
> doesn't really care what _41 is, even if it would be VARYING, the
> loop still has 0 or 1 latch executions, never more.
> Perhaps this is too weird case on which nbiter just punts?
I guess so. It was probably expected jump threading might thread
the latch to exit? But that probably happens too late (it should
be good early I think - it makes the loop a non-loop?)
Trying to pattern-match the static-computable exit condition from
latch in niter analysis would be OK I guess. Not sure where this
would fit best, technically we iterate once, so it's only an
upper bound candidate. Usually header copying would fix this up
but the inner loop likely prevents this due to cost.
Richard.
> loop_4 (header = 23, latch = 22)
> {
> bb_22 (preds = {bb_21 }, succs = {bb_23 })
> {
> <bb 22> :
> # .MEM_22 = PHI <.MEM_29(21)>
> _91 = _17 + 2;
>
> }
> bb_23 (preds = {bb_19 bb_22 }, succs = {bb_25 bb_24 })
> {
> <bb 23> :
> # _17 = PHI <_77(19), _91(22)>
> # _18 = PHI <_41(19), -1(22)>
> # .MEM_30 = PHI <.MEM_76(19), .MEM_22(22)>
> if (_18 == 0)
> goto <bb 25>; [INV]
> else
> goto <bb 24>; [INV]
>
> }
> bb_25 (preds = {bb_23 }, succs = {bb_21 })
> {
> <bb 25> :
> goto <bb 21>; [100.00%]
>
> }
> loop_5 (header = 21, latch = 20
> upper_bound 2
> likely_upper_bound 2
> estimate 2)
> {
> bb_20 (preds = {bb_21 }, succs = {bb_21 })
> {
> <bb 20> :
> # .MEM_92 = VDEF <.MEM_29>
> *_19 ={v} {CLOBBER(bob)};
> _93 = _20 + -1;
> _94 = _19 + 1;
>
> }
> bb_21 (preds = {bb_25 bb_20 }, succs = {bb_20 bb_22 })
> {
> <bb 21> :
> # _19 = PHI <_17(25), _94(20)>
> # _20 = PHI <1(25), _93(20)>
> # .MEM_29 = PHI <.MEM_30(25), .MEM_92(20)>
> if (_20 != -1)
> goto <bb 20>; [INV]
> else
> goto <bb 22>; [INV]
>
> }
> }
> }
>
>
> Jakub
>
--
Richard Biener <[email protected]>
SUSE Software Solutions Germany GmbH,
Frankenstrasse 146, 90461 Nuernberg, Germany;
GF: Ivo Totev, Andrew McDonald, Werner Knoblich; (HRB 36809, AG Nuernberg)
