On Tue, Apr 10, 2012 at 2:28 PM, Dave Anglin <dave.ang...@bell.net> wrote:
> On 10-Apr-12, at 7:06 AM, Richard Guenther wrote:
>> I can't immediately see how your description of "the list of pending
>> and the vector" is deleted. pa.c keeps its own vector which references
>> decls and the only issue I see is that if you call assemble_external after
>> processing externals you ICE because the pointer-set is not initialized?
>> Why does the pa backend end up calling assemble_external at final time?
> The backend calls assemble_integer to output function descriptors at final
> time. This indirectly calls assemble_external from output_addr_const.
> This can be seen in comment #4 in the PR. This occurs after the pending
> externals are processed.
> We have to ensure that we only output function descriptors that are
> and we can only determine this at final.
You mean function destcriptors for functions that are output? You already
know this at cgraph clone materialization time.
> As a result, assemble_external is called after the VEC has been deleted.
> The compiler doesn't ICE. It goes into an infinite loop when a call to
> pointer_set_insert tries to extend the deleted VEC. It does this even if
> pointer is already in the VEC hash.
Isn't the bug then that the backend deletes the VEC too early?
> Previously, the storage wasn't deleted. I suspect the ineffective calls on
> weren't noticed because assemble_external had already been called for the
> function descriptors earlier, so the list in the backend was complete.
> The problem was first seen with the Linux target. There is no backend
> list for this target because it doesn't use or need assemble_external.
> I personally think it was wrong to add the deferral in assemble_external,
> but that's another issue.
I think the best thing would be to revert the offending patch on old branches
(4.6 and 4.5) and see if the deferal can be fixed properly - though I didn't
look at the pa issue in detail. Maybe Steven can do this.