This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [RFA][PATCH] Provide a class interface to ssa_propagate
- From: David Malcolm <dmalcolm at redhat dot com>
- To: Jeff Law <law at redhat dot com>, gcc-patches <gcc-patches at gcc dot gnu dot org>
- Date: Tue, 24 Oct 2017 16:49:33 -0400
- Subject: Re: [RFA][PATCH] Provide a class interface to ssa_propagate
- Authentication-results: sourceware.org; auth=none
- Authentication-results: ext-mx03.extmail.prod.ext.phx2.redhat.com; dmarc=none (p=none dis=none) header.from=redhat.com
- Authentication-results: ext-mx03.extmail.prod.ext.phx2.redhat.com; spf=fail smtp.mailfrom=dmalcolm at redhat dot com
- Dmarc-filter: OpenDMARC Filter v1.3.2 mx1.redhat.com 370686A7D8
- References: <4b766dde-648f-97a4-7366-e38e0154f29d@redhat.com>
On Tue, 2017-10-24 at 11:40 -0600, Jeff Law wrote:
> tree-ssa-propagate.c provides a fairly generic engine to propagate
> values through a lattice while in SSA form. The engine uses two
> callbacks to allow passes to provide pass specific handling of
> statements and phi nodes.
>
> The callback mechanism served us well in a C world. It is however
> somewhat painful to have state in those callbacks without resorting
> to
> global variables or passing around void * objects which contain the
> class instance pointer.
>
> For example, tree-vrp uses the propagation engine to compute global
> range information. Its callbacks vrp_visit_stmt and vrp_visit_phi
> and
> their children read/modify a variety of tree-vrp.c statics such as
> vr_data.
>
> In some changes I'm working on I'd really like to move vr_data into a
> distinct class and avoid having direct accesses to the underlying
> array.
>
> So the problem is how are routines like vrp_visit_stmt and
> vrp_visit_phi
> and their children supposed to access the class instance?
>
> One way would be to just add a void * argument to them and pass the
> class instance around. Alternately we could leave the global
> variable
> in place and have it set up, checked and wiped clean by the vr_data
> class's ctor/dtor. Both are valid and would work, but they're a bit
> ugly IMHO.
>
> This patch takes another approach. It builds a simple little class
> around ssa_propagate where the statement and phi visitors are virtual
> functions. Thus clients can override the visitors *and* they'll get
> a
> class instance pointer.
>
> I haven't gone hog wild with C++-ification, basically just enough to
> get
> the class around ssa_propagate and its children which are going to
> need
> to pass down the class instance to the virtual functions. There's a
> lot
> more that could be done here.
>
> As you can see the client side changes are pretty minimal. They just
> derive a new class from ssa_propagation_engine to provide their
> implementations of the statement and phi visitor. More importantly,
> they can hang data off that derived class which we'll exploit later.
>
> There will be a similar patch for the substitute_and_fold which has
> callbacks of its own.
>
> Bootstrapped and regression tested on x86_64.
>
> The ChangeLog makes the patch look huge. But it's actually
> relatively
> small and the client side bits are repetitive.
>
> OK for the trunk?
>
> Jeff
Looks like you forgot to attach the patch.
Dave