https://gcc.gnu.org/bugzilla/show_bug.cgi?id=89691
--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
It doesn't handle negative offsets, we're doing
/* Optional constant offset. */
if (c_parser_next_token_is (parser, CPP_PLUS))
{
c_parser_consume_token (parser);
alias_off
= c_parser_gimple_postfix_expression (parser).value;
alias_off = fold_convert (alias_type, alias_off);
}
I tried to parse what -gimple dumps show (which indeed is the "ugly"
hex number given the INTEGER_CSTs have pointer type). It works when
you write
__MEM <int, 16> ((char *)&x + _Literal (int) -2) = i;
the magic _Literal forces (some) constant foldings/type conversions to
be done so we get a single value rather than the syntactical negate of
two. This shows elsewhere, too. Note it is required to not always parse
-2 as literal since else you couldn't write an unfolded unary minus operation
of a constant.
OTOH I just figured we -gimple dump
int foo (int i)
{
int j = -1;
return j + i;
}
as
int __GIMPLE ()
foo (int i)
{
int D_1910;
int j;
j = -1;
D_1910 = j + i;
return D_1910;
}
which is then wrong.
