I just wanted to post an update, mainly that I have solved my problem
:-) A bit more on this follows below...
On Sat, Feb 2, 2013 at 12:01 PM, Matt Davis <mattdav...@gmail.com> wrote:
> Thanks Ian, Richard.
> I have some modified code which seems to be along the same lines as
> what you all suggested. However, I am still having troubles. Mainly,
> I see the array in the callee but the contents are still empty, and I
> verified by looking at the asm dump of the resulting code.
>
> On Fri, Feb 1, 2013 at 8:09 PM, Richard Biener
> <richard.guent...@gmail.com> wrote:
>> On Fri, Feb 1, 2013 at 5:03 AM, Matt Davis <mattdav...@gmail.com> wrote:
>>> Hello,
>>> I have a routine that creates a local array containing pointers to
>>> global data. At runtime, when this array is passed to a function, I
>>> do not see the pointers to the global objects. The GIMPLE does show
>>> that the array is declared with the addresses of the globals as the
>>> elements to the array, and that looks fine to me. But at runtime,
>>> when this array is passed to a callee function, the callee receives
>>> the array, but its contents are not the addresses of the globals.
>>>
>>> After looking at the corresponding assembly, this makes sense, as I do
>>> not see the assembly building the array before it passes the array to
>>> the callee. I do not see the asm code assigning the elements of the
>>> array to be that of the global pointers. I suppose there is a flag I
>>> need to set? If not, I suppose I can always build the array as a
>>> series of assignments.
>>>
>>> Here is how I build the local array. Note, if I build this array as a
>>> global with static-linkage, then everything works fine:
>>>
>>> tree create_array(const region_t reg, gimple stmt)
>>> {
>>> unsigned i, n_elts;
>>> const type_info_t *ti;
>>> tree type, unique, decl;
>>> VEC(constructor_elt,gc) *entries = NULL;
>>>
>>> n_elts = VEC_length(tree, reg->unique_types);
>>> type = build_array_type_nelts(ptr_type_node, n_elts);
>>>
>>> FOR_EACH_VEC_ELT(tree, reg->unique_types, i, unique)
>>> {
>>> ti = get_type_info(reg, unique);
>>> CONSTRUCTOR_APPEND_ELT(
>>> entries, NULL, build1(ADDR_EXPR, ptr_type_node, ti->decl));
>>> }
>>>
>>> decl = create_tmp_var(type, "testarray");
>>> DECL_INITIAL(decl) = build_constructor(type, entries);
>>>
>>> return decl;
>>> }
>>>
>>> Do I have to explicitly create assignment statements for each element,
>>> since my array is local? As I mention above, if I make my array
>>> global, everything is fine.
>>
>> Locals with DECL_INITIAL need to be lowered in GIMPLE to make the
>> initialization explicit. What you can do is output the constructor as
>> constant (supposed all elements are constant) using tree_output_constant_def.
>> See how gimplification handles initializers of locals.
>
> I did look through the gimplification stuff. And I added a
> "tree_output_constant_def" on my DECL_INITIAL as suggested by Richard.
> As Ian suggested, I created a DECL_EXPR for this array that I am generating.
>
> The new code is the following:
>
> /* Build the constructor for the array */
> FOR_EACH_VEC_ELT(tree, reg->unique_types, i, unique)
> {
> ti = get_type_info(reg, unique);
> CONSTRUCTOR_APPEND_ELT(
> entries, NULL, build1(ADDR_EXPR, ptr_type_node, ti->decl));
> }
>
> /* Create the VAR_DECL for the array, set the constructor and force it to
> * be generated at local scope.
> */
> decl = create_tmp_var(type, "testarray");
> DECL_INITIAL(decl) =
> tree_output_constant_def(build_constructor(type, entries));
> return build1_loc(gimple_location(stmt), DECL_EXPR, type, decl);
>
>
> I use the DECL_EXPR_DECL() of the returned value as I would have
> normally used 'decl' in my previous listing. Looking at the gimple
> dump and assembly, I see the gimple code now containing a pointer to
> a label for the declaration (thanks to tree_output_constant_def). In
> the resulting asm, I do not see this label anywhere.
I fixed my code up, removed the tree_output_constant_def, and left the
DECL_EXPR in place. I traced through the gimple code using
"gimple_decl_expr" as a hint as to what I needed gcc to accomplish for
me. Anyways, I just called the gimplifier on the DECL_EXPR, and that
seemed to do the trick. Thanks again Ian and Richard.
-Matt
