This is the mail archive of the
gcc-bugs@gcc.gnu.org
mailing list for the GCC project.
[Bug middle-end/81931] [8 regression] r251260 breaks many tests on powerpc64
- From: "rguenth at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>
- To: gcc-bugs at gcc dot gnu dot org
- Date: Wed, 23 Aug 2017 11:48:59 +0000
- Subject: [Bug middle-end/81931] [8 regression] r251260 breaks many tests on powerpc64
- Auto-submitted: auto-generated
- References: <bug-81931-4@http.gcc.gnu.org/bugzilla/>
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81931
--- Comment #7 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Aldy Hernandez from comment #6)
> The problem here is that get_nonzero_bits() is being called with an SSA
> whose precision is 0 (_Complex float). This is causing this code in
> get_nonzero_bits():
>
> range_info_def *ri = SSA_NAME_RANGE_INFO (name);
> if (!ri)
> return wi::shwi (-1, precision);
>
> ...to return a wide int of:
>
> (gdb) p nonzero_bits.dump()
> [0], precision = 0
>
> whereas previous to the aforementioned patch, we were returning:
>
> (gdb) p nonzero_bits.dump()
> [0xffffffffffffffff], precision = 0
>
> This discrepancy is causing the CCP code to think a value is a known
> constant, instead of a VARYING:
>
> if (flag_tree_bit_ccp
> && ((is_constant && TREE_CODE (val.value) == INTEGER_CST)
> || !is_constant)
> && gimple_get_lhs (stmt)
> && TREE_CODE (gimple_get_lhs (stmt)) == SSA_NAME)
> {
> tree lhs = gimple_get_lhs (stmt);
> wide_int nonzero_bits = get_nonzero_bits (lhs);
> if (nonzero_bits != -1)
> {
> if (!is_constant)
> ...
>
> nonzero_bits is no longer equal to -1, so we think we have a known value.
>
> IMO, a precision of 0 makes no sense. Perhaps we should even hard fail in
> get_nonzero_bits for this case, and have all callers fix their nonsense.
Yeah, we do assert we are not passed a POINTER_TYPE but in reality in addition
to that we only may look at INTEGRAL_TYPE_P types.
So -- assert that in *_range_info and *nonzero_bits and fix callers. Oh,
I think nonzero_bits can handle POINTER_TYPEs just fine.
I suppose the idea was to make nonzero_bits foolproof. And -1 being
sign-extended should be fine... unless precision is 0 ;)
So, in nonzero_bits use TYPE_PRECISION only for INTEGRAL_TYPE_P and
POINTER_TYPE_P but for other types use TYPE_SIZE (or some arbitrary
nonzero precision? BITS_PER_UNIT?).