[Bug middle-end/78684] [7 Regression] ICE in create_intersect_range_checks_index, at tree-vect-loop-manip.c:2074

[amker at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78684

--- Comment #8 from amker at gcc dot gnu.org ---
(In reply to Jakub Jelinek from comment #7)
> (In reply to amker from comment #6)
> > Ok, in this case I need to check idx_step is negative if it's signed type;
> > or idx_step has sign bit if it's unsigned type (negative step in effect). 
> > tree_int_cst_sgn always returning 1 for (sizetype)-1 looks not working here?
> 
> tree_int_cst_compare (idx_step, size_zero_node) < 0 would be also always
> false for unsigned types.  Perhaps you can use wi::neg_p (idx_step) then,
Yes, I made this mistake which causes this ICE.

> that will return true whenever idx_step in corresponding signed type (with
> the same precision as idx_step's type) is negative.
> 
> But, is it right to consider those as negative steps?
> I mean, say
> for (unsigned long long c = 0; c < 0x8000000000000005ULL; c +=
> 0x8000000000000001ULL)
> is actually positive step, not negative one.
It's safe?  The loop is equal to below one, right?
for (unsigned long long c = 0; c < 0x8000000000000005ULL; c -=
0x7fffffffffffffffULL)