[Bug tree-optimization/68835] [5/6 Regression] ICE in set_value_range, at tree-vrp.c:387, with __int128 bit field

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68835

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |rsandifo at gcc dot gnu.org
           Assignee|jakub at gcc dot gnu.org           |unassigned at gcc dot gnu.org

--- Comment #8 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
The > HOST_BITS_PER_WIDE_INT stuff is just weird.
p debug_tree (build_nonstandard_integer_type (66, 1))
 <integer_type 0x7ffff18170a8 public unsigned TI
    size <integer_cst 0x7ffff1713cf0 type <integer_type 0x7ffff17172a0
bitsizetype> constant 128>
    unit size <integer_cst 0x7ffff1713d08 type <integer_type 0x7ffff17171f8
sizetype> constant 16>
    align 128 symtab 0 alias set -1 canonical type 0x7ffff18170a8 precision 66
min <integer_cst 0x7ffff1827180 0> max <integer_cst 0x7ffff17302b8
0xffffffffffffffffffffffffffffffff>>
(gdb) p debug_tree (build_int_cst (0x7ffff18170a8, 1))
 <integer_cst 0x7ffff18271b0 type <integer_type 0x7ffff18170a8> constant 1>
(gdb) p debug_tree (int_const_binop (MINUS_EXPR, 0x7ffff17302b8,
0x7ffff18271b0))
 <integer_cst 0x7ffff1827198 type <integer_type 0x7ffff18170a8> constant
0xfffffffffffffffffffffffffffffffe>
(gdb) p debug_tree (build_int_cst (0x7ffff18170a8, 2))
 <integer_cst 0x7ffff18271e0 type <integer_type 0x7ffff18170a8> constant 2>
(gdb) p debug_tree (int_const_binop (TRUNC_DIV_EXPR, 0x7ffff17302b8,
0x7ffff18271e0))
 <integer_cst 0x7ffff1824ea0 type <integer_type 0x7ffff18170a8> constant
0x1ffffffffffffffff>

The max value of the unsigned type (for signed types both the min and max
values look ok) looks wrong, and so does the max - 1 value, but the max / 2
value is printed correctly.  This can be surely a pretty printer bug, or could
be a bug in wide_int_to_tree.  Depends on what should be the canonical
representation of precision > HOST_BITS_PER_WIDE_INT (and < 2 *
HOST_BITS_PER_WIDE_INT) unsigned values in between maximum and maximum -
LLONG_MAX.