https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119690
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
See Also| |https://gcc.gnu.org/bugzill
| |a/show_bug.cgi?id=114074
--- Comment #8 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Jakub Jelinek from comment #7)
> I think this is ivopts bug (or at least that pass introduces UB into the IL
> that wasn't there before).
> In particular on the c <= 0 guarded edge it adds
> <bb 8> [local count: 105119324]:
> + _26 = b.1_13 + -2147483648;
> + _1 = (unsigned int) _26;
> + _8 = (unsigned int) _14;
> + ivtmp.17_11 = _8 + 2147483648;
> where b.1_13 at runtime is -2147483647 and both b.1_13 and _26 have int type.
> So, the addition invokes UB. Given that previously all that was there was
> (a + b) + 2147483647, nothing was known about b's range, only about a + b,
> so I think it certainly should have been unsigned _26:
> _26 = (unsigned int) b.1_13;
> _1 = _26 + 2147483648;
I guess this is one of the more long-standing SCEV issues, we analyze the
evolution of the loop PHI as:
(get_scalar_evolution
(scalar = _18)
(scalar_evolution = {_14, +, b.1_13 + -2147483648(OVF)}_1))
and from that
(get_scalar_evolution
(scalar = _5)
(scalar_evolution = {_14 - -2147483648, +, b.1_13 + -2147483648(OVF)}_1))
)
IVOPTs in the end selects
Selected IV set for loop 1 at t.c:8, 10 avg niters, 1 IVs:
Candidate 8:
Depend on inv.exprs: 1
Var befor: ivtmp.17_22
Var after: ivtmp.17_21
Incr POS: before exit test
IV struct:
Type: unsigned int
Base: (unsigned int) _14 + 2147483648
Step: (unsigned int) (b.1_13 + -2147483648(OVF))
Biv: N
Overflowness wrto loop niter: Overflow
but it's really the association done by SCEV where things go downhill from.
I think this is another instance of PR114074 and friends.
