Re: [PATCH v2 3/4] gdb, typeprint: workaround icc/icpc/ifort pointer/reference DWARF

[Tom Tromey <tom@tromey.com>]

>>>>> Nils-Christian Kempke via Gdb-patches <gdb-patches@sourceware.org> writes:

> Intel classic compilers (icc/icpc/ifort) for references/pointers
> to arrays generate DWARF that looks like

Thank you for the patch.

> +  if (!top_level && icc_pointer_or_reference_type (type))
> +    {
> +      /* Icc/ifort emit the DW_AT_associated for pointers and references.  To
> +	 not mark such types as dynamic further down, which would lead to
> +	 infinite resolution loops for, e.g., cyclic dynamic pointers, we
> +	 return here already.  */
> +      return 0;

I suspect this is not the best spot to do this kind of check.

> +bool
> +icc_pointer_or_reference_type (const struct type *type)
> +{
> +  return (type->code () == TYPE_CODE_PTR || type->code () == TYPE_CODE_REF)
> +	 && type->is_objfile_owned ()
> +	 && std::any_of (type->objfile_owner ()->compunits ().begin (),
> +			 type->objfile_owner ()->compunits ().end (),
> +			 [] (const compunit_symtab *cu)
> +			 {
> +			   return producer_is_icc (cu->producer (), nullptr,
> +						   nullptr);
> +			 });

And I really don't like this, because it's over-broad.  If an objfile
has objects from two different compilers, this code will erroneously
trigger.

Instead, it seems to me that a better approach would be to recognize
the oddities in the DWARF reader, and then perhaps either introduce new
types or mark the type somehow for later processing.

If you search for "quirk" it the reader, you'll see other examples of
this technique.  E.g., the Rust compiler used to emit a custom format
for Rust enums, and the code in the DWARF reader converts these types
into the internal representation used in the rest of gdb.

I'm not sure if this can completely be done in your case or not, but I
think it would be better to try.  If it fails, adding a new flag to the
type would be better than searching all the objfile's compunits.

Tom