https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113210
--- Comment #14 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #13)
> (In reply to Richard Biener from comment #12)
> > (In reply to Jakub Jelinek from comment #11)
> > > (In reply to Richard Biener from comment #9)
> > > > That is, another fix might be to adjust NITERSM1 to NITERS - 1 when
> > > > NITERS went constant ... (btw, I want to get rid of _NITERS and only
> > >
> > > Or we could only use fold_build2 for the PLUS_EXPR 1 computation if
> > > NITERSM1
> > > is INTEGER_CST, otherwise use build2...
> >
> > I think we should see where the original expression is built but not folded.
>
> Hmm, probably in estimate_numbers_of_iterations,
>
> if (TREE_CODE (niter_desc.may_be_zero) != INTEGER_CST)
> niter = build3 (COND_EXPR, type, niter_desc.may_be_zero,
> build_int_cst (type, 0),
> niter);
>
> I vaguely remember code trying to pattern match the COND_EXPR created
> by this (though it should instead use number_of_iterations_exit). It
> should be safe to replace the above with fold_build3.
No, it's not that. The COND_EXPR is folded, but we see
(short unsigned int) (a.0_1 + 255) + 1 > 256 ? ~(short unsigned int) (a.0_1 +
255) : 0
and that isn't catched.
#0 0x00000000014d3464 in fold_build3_loc (loc=0, code=COND_EXPR,
type=<integer_type 0x7ffff704b540 short unsigned int>,
op0=<gt_expr 0x7ffff7207f50>, op1=<bit_not_expr 0x7ffff6e13540>,
op2=<integer_cst 0x7ffff7047378>)
at /space/rguenther/src/gcc/gcc/fold-const.cc:14174
#1 0x00000000014d0513 in fold_ternary_loc (loc=0, code=COND_EXPR,
type=<integer_type 0x7ffff704b540 short unsigned int>,
op0=<le_expr 0x7ffff7207f00>, op1=<integer_cst 0x7ffff7047378>,
op2=<bit_not_expr 0x7ffff6e13540>)
at /space/rguenther/src/gcc/gcc/fold-const.cc:13238
#2 0x00000000014d3436 in fold_build3_loc (loc=0, code=COND_EXPR,
type=<integer_type 0x7ffff704b540 short unsigned int>,
op0=<le_expr 0x7ffff7207f00>, op1=<integer_cst 0x7ffff7047378>,
op2=<bit_not_expr 0x7ffff6e13540>)
at /space/rguenther/src/gcc/gcc/fold-const.cc:14172
#3 0x0000000001d4de50 in vect_get_loop_niters (loop=0x7ffff71e3600,
main_exit=0x7ffff7203990, assumptions=0x7fffffffd200,
number_of_iterations=0x7fffffffd1f0, number_of_iterationsm1=0x7fffffffd1f8)
at /space/rguenther/src/gcc/gcc/tree-vect-loop.cc:919
So I suggest to either try fixing that or, when adding one folds to a constant
make sure M1 is constant as well.
