Re: [PATCH 0/6] Conversion of gimple types to C++ inheritance (v3)

[David Malcolm <dmalcolm@redhat.com>]

On Tue, 2013-11-05 at 14:18 -0700, Jeff Law wrote:
> On 10/31/13 10:26, David Malcolm wrote:
> > The gimple statement types are currently implemented using a hand-coded
> > C inheritance scheme, with a "union gimple_statement_d" holding the
> > various possible structs for a statement.
> >
> > The following series of patches convert it to a C++ hierarchy, using the
> > existing structs, eliminating the union. The "gimple" typedef changes
> > from being a
> >    (union gimple_statement_d *)
> > to being a:
> >    (struct gimple_statement_base *)
> >
> > There are no virtual functions in the new code: the sizes of the various
> > structs are unchanged.
> >
> > It makes use of "is-a.h", using the as_a <T> template function to
> > perform downcasts, which are checked (via gcc_checking_assert) in an
> > ENABLE_CHECKING build, and are simple casts in an unchecked build,
> > albeit it in an inlined function rather than a macro.
> >
> > For example, one can write:
> >
> >    gimple_statement_phi *phi =
> >      as_a <gimple_statement_phi> (gsi_stmt (gsi));
> >
> > and then directly access the fields of the phi, as a phi.  The existing
> > accessor functions in gimple.h become somewhat redundant in this
> > scheme, but are preserved.
> >
> > The earlier versions of the patches made all of the types GTY((user))
> > and provided hand-written implementations of the gc and pch marker
> > routines.  In this new version we rely on the support for simple
> > inheritance that I recently added to gengtype, by adding a "desc"
> > to the GTY marking for the base class, and a "tag" to the marking
> > for all of the concrete subclasses.  (I say "class", but all the types
> > remain structs since their fields are all publicly accessible).
> >
> > As noted in the earlier patch, I believe this is a superior scheme to
> > the C implementation:
> >
> >    * We can get closer to compile-time type-safety, checking the gimple
> >      code once and downcasting with an as_a, then directly accessing
> >      fields, rather than going through accessor functions that check
> >      each time.  In some places we may want to replace a "gimple" with
> >      a subclass e.g. phis are always of the phi subclass, to get full
> >      compile-time type-safety.
> >
> >    * This scheme is likely to be easier for newbies to understand.
> >
> >    * Currently in gdb, dereferencing a gimple leads to screenfuls of text,
> >      showing all the various union values.  With this, you get just the base
> >      class, and can cast it to the appropriate subclass.
> >
> >    * With this, we're working directly with the language constructs,
> >      rather than rolling our own, and thus other tools can better
> >      understand the code. (e.g. doxygen).
> >
> > Again, as noted in the earlier patch series, the names of the structs
> > are rather verbose.  I would prefer to also rename them all to eliminate
> > the "_statement" component:
> >    "gimple_statement_base" -> "gimple_base"
> >    "gimple_statement_phi"  -> "gimple_phi"
> >    "gimple_statement_omp"  -> "gimple_omp"
> > etc, but I didn't do this to mimimize the patch size.  But if the core
> > maintainers are up for that, I can redo the patch series with that
> > change also, or do that as a followup.
> >
> > The patch is in 6 parts; all of them are needed together.
> And that's part of the problem.  There's understandable resistance to 
> (for example) the as_a casting.
> 
> There's a bit of natural tension between the desire to keep patches 
> small and self-contained and the size/scope of the changes necessary to 
> do any serious reorganization work.  This set swings too far in the 
> latter direction :-)
> 
> Is there any way to go forward without the is_a/as_a stuff?  ie, is 
> there're a simpler step towards where we're trying to go that allows 
> most of this to go forward now rather than waiting?
> 
> >
> >    * Patch 1 of 6: This patch adds inheritance to the various gimple
> >      types, eliminating the initial baseclass fields, and eliminating the
> >      union gimple_statement_d.   All the types remain structs.  They
> >      become marked with GTY(()), gaining GSS_ tag values.
> >
> >    * Patch 2 of 6: This patch ports various accessor functions within
> >      gimple.h to the new scheme.
> >
> >    * Patch 3 of 6: This patch is autogenerated by "refactor_gimple.py"
> >      from https://github.com/davidmalcolm/gcc-refactoring-scripts
> >      There is a test suite "test_refactor_gimple.py" which may give a
> >      clearer idea of the changes that the script makes (and add
> >      confidence that it's doing the right thing).
> >      The patch converts code of the form:
> >        {
> >          GIMPLE_CHECK (gs, SOME_CODE);
> >          gimple_subclass_get/set_some_field (gs, value);
> >        }
> >      to code of this form:
> >        {
> >          some_subclass *stmt = as_a <some_subclass> (gs);
> >          stmt->some_field = value;
> >        }
> >      It also autogenerates specializations of
> >          is_a_helper <T>::test
> >      equivalent to a GIMPLE_CHECK() for use by is_a and as_a.
> Conceptually I'm fine with #1-#3.
> 
> >
> >    * Patch 4 of 6: This patch implement further specializations of
> >      is_a_helper <T>::test, for gimple_has_ops and gimple_has_mem_ops.
> Here's where I start to get more concerned.

Thanks for looking through this.

Both you and Andrew objected to my use of the is-a.h stuff.  Is this due
to the use of C++ templates in that code?   If I were to rewrite things
in a more C idiom, would that be acceptable?

For instance, rather than, say:

  p = as_a <gimple_statement_asm> (
        gimple_build_with_ops (GIMPLE_ASM, ERROR_MARK,
                              ninputs + noutputs + nclobbers + nlabels));

we could have an inlined as_a equivalent in C syntax:

  p = gimple_as_a_gimple_asm (
        gimple_build_with_ops (GIMPLE_ASM, ERROR_MARK,
                              ninputs + noutputs + nclobbers + nlabels));

where there could be, say, a pair of functions like this (to handle
const vs non-const):

inline gimple_asm
gimple_as_a_gimple_asm (gimple gs)
{
  GIMPLE_CHECK (gs->code == GIMPLE_ASM);
  return (gimple_asm)gs;
}

inline const_gimple_asm
gimple_as_a_gimple_asm (const_gimple gs)
{
  GIMPLE_CHECK (gs->code == GIMPLE_ASM);
  return (const_gimple_asm)gs;
}

(where typedef gimple_statement_asm *gimple_asm)

That would avoid template usage within the patch, leaving the use of C++
inheritance as the only overtly C++ish aspect.

We could do the above using preprocessor magic, but I'd prefer to have
actual code to do it.

Similarly, instead of:

  const gimple_statement_with_ops *ops_stmt =
    dyn_cast <const gimple_statement_with_ops> (g);
  if (!ops_stmt)
    return NULL;

we could have:

  const_gimple_with_ops ops_stmt =
    gimple_dyn_cast_gimple_with_ops (g);
  if (!ops_stmt)
    return NULL;

> >    * Patch 5 of 6: This patch does the rest of porting from union access
> >      to subclass access (all the fiddly places that the script in patch 3
> >      couldn't handle).
> >
> >    * Patch 6 of 6: This patch updates the gdb python pretty-printing
> >      hook.
> Conceptually #5 and #6 shouldn't be terribly controversial.

(...though they're implicitly using the template specializations from #3
and #4)

> THe question is can we move forward without patch #4, even if that means 
> we aren't getting the static typechecking we want?

Maybe.  If the above idea is still too far, we could keep the
GIMPLE_CHECK checking, and cast by hand.  I suspect the results would be
more ugly (though it's clear that beauty is in the eye of the beholder
here :))

BTW, how do you feel about static_cast<> vs C-style casts?

Thanks
Dave