------- Comment #8 from joseph at codesourcery dot com 2006-01-15 23:52 -------
Subject: Re: [4.0/4.1/4.2 Regression] Internal compiler error
(segfault) instead of error message
On Sun, 8 Jan 2006, steven at gcc dot gnu dot org wrote:
> My hack-around is to deal with label_context_stack_vm==NULL in more places.
> Sadly I have no idea what this variable is for, and Joseph did not add any
> comment or documentation of any form for these global variables.
The documentation is on the structure definition c_label_context_vm (and
similarly c_label_context_se).
I think fixing pushdecl not to call c_begin_vm_scope if current_scope ==
file_scope should suffice.
> If there is a way to reject this non-constant _before_ we call
> c_begin_vm_scope, that would be the "right" fix AFAICT.
The right fix there would perhaps be in grokdeclarator,
if (TREE_CODE (size) == INTEGER_CST)
{
constant_expression_warning (size);
if (tree_int_cst_sgn (size) < 0)
{
error ("size of array %qs is negative", name);
size = integer_one_node;
}
}
else
{
/* Make sure the array size remains visibly
nonconstant even if it is (eg) a const variable
with known value. */
size_varies = 1;
before the "else" have an "else if" case that checks for NORMAL or FIELD
at file scope, gives an error and uses integer_one_node instead. This
should also cause certain other cases we silently accept to be
unconditionally rejected instead:
int a;
int (*b)[(__SIZE_TYPE__)&a];
(ICEs at present because of this bug, older GCC accepts silently)
or
int a;
struct s { int b[(__SIZE_TYPE__)&a]; };
or
int a;
struct s { int (*x)[(__SIZE_TYPE__)&a]; };
That way we don't rely on variable_size to give an error (it's not a good
place to do so anyway, this error belongs in the front end), which won't
work for TREE_CONSTANT cases such as these addresses. (All examples
including the original test ((int) changed to (__SIZE_TYPE__)) should of
course be added as testcases.)
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=25161
