public inbox for gcc-patches@gcc.gnu.org
 help / color / mirror / Atom feed
From: Jakub Jelinek <jakub@redhat.com>
To: Andrew MacLeod <amacleod@redhat.com>
Cc: Jan Hubicka <jh@suse.cz>, Richard Biener <rguenther@suse.de>,
	gcc-patches@gcc.gnu.org
Subject: Re: [PATCH] middle-end, v4: IFN_ASSUME support [PR106654]
Date: Wed, 19 Oct 2022 19:39:05 +0200	[thread overview]
Message-ID: <Y1A2Ofdwdcm93s03@tucnak> (raw)
In-Reply-To: <70311a20-1a94-21a1-e897-37289c9e0f0c@redhat.com>

On Wed, Oct 19, 2022 at 12:55:12PM -0400, Andrew MacLeod via Gcc-patches wrote:
> > Not sure I understand this part.  op is whatever we pass as the ith
> > argument to IFN_ASSUME.  I'd expect that at this point one needs to
> > remap that to the (i-1)th PARM_DECL of assume_id (so e.g. when you
> > have the above loop you could as well start with DECL_ARGUMENTS and move
> > that to DECL_CHAIN at the end of every iteration.  And then
> > query ssa_default_def (DECL_STRUCT_FUNCTION (assume_id), parm)
> > in each case and get global range of what that returns.
> 
> OK, this is the bit of code I dont know how to write :-)
> 
> yes, op is the name of the value within this current function, and yes, that
> needs to be mapped to the argument decl in the assume function.   Then we
> need to query what range was given to that name during the assume pass. 
> when that is returned, the add_range (op, range) will inject it as a side
> effect.
> 
> Can you write that loop?

I meant something like (untested code):
       && gimple_call_internal_fn (s) == IFN_ASSUME)
     {
       tree assume_id = gimple_call_arg (s, 0);
-      for (unsigned i = 1; i < gimple_call_num_args (s); i++)
+      tree parm = DECL_ARGUMENTS (assume_id);
+      struct function *fun = DECL_STRUCT_FUNCTION (assume_id);
+      for (unsigned i = 1;
+	   i < gimple_call_num_args (s) && parm;
+	   i++, parm = DECL_CHAIN (parm))
 	{
 	  tree op = gimple_call_arg (s, i);
 	  tree type = TREE_TYPE (op);
+	  tree arg = ssa_default_def (fun, parm);
+	  if (arg == NULL_TREE)
+	    continue;
 	  if (gimple_range_ssa_p (op) && Value_Range::supports_type_p (type))
 	    {
 	      Value_Range assume_range (type);
and querying SSA_NAME_RANGE_INFO of arg rather than op.
> 
> > 
> > > +  for (unsigned  x= 0; x < gimple_phi_num_args (phi); x++)
> >    for (unsigned x = 0; ...
> > ?
> > 
> > > @@ -4345,6 +4345,30 @@ execute_ranger_vrp (struct function *fun, bool warn_array_bounds_p)
> > >     scev_initialize ();
> > >     calculate_dominance_info (CDI_DOMINATORS);
> > > +  assume_query *r2 = new assume_query ();
> > > +  for (unsigned i = 0; i < num_ssa_names; i++)
> > > +    {
> > > +      tree name = ssa_name (i);
> > > +      if (!name || !gimple_range_ssa_p (name))
> > > +	continue;
> > > +      tree type = TREE_TYPE (name);
> > > +      if (!Value_Range::supports_type_p (type))
> > > +	continue;
> > > +      Value_Range assume_range (type);
> > > +      if (r2->assume_range_p (assume_range, name))
> > > +	{
> > > +	  if (dump_file)
> > > +	    {
> > > +	      fprintf (dump_file, "for an assume function, ");
> > > +	      print_generic_expr (dump_file, name, TDF_SLIM);
> > > +	      fprintf (dump_file, " would have a range of ");
> > > +	      assume_range.dump (dump_file);
> > > +	      fputc ('\n', dump_file);
> > > +	    }
> > > +	}
> > > +    }
> > > +  delete r2;
> > I have expected (but tell me if that isn't possible) this could be something
> > done in the new pass_assumptions::execute () rather than vrp and you'd
> > create the assume_query there (i.e. just for assume_functions) and then
> > query it solely for ssa_default_def of the parameters and save in
> > SSA_NAME_RANGE_INFO.
> 
> I just discovered the assumption pass, and I have moved it to there.
> 
> I dont know much about managing the parameters, but presumably yes, we'd
> only query it for the parameters...........  I was showing the query for
> every name just to show what its producing.

Inside of the assume function (cfun->assume_function being true)
one could again walk DECL_ARGUMENTS and for arguments with types
which ranger is able to cope with and they are reg types,
ssa_default_def (cfun, parm) to get the SSA_NAME of the default def (if
any).

> --- a/gcc/tree-vrp.cc
> +++ b/gcc/tree-vrp.cc
> @@ -4465,6 +4465,35 @@ public:
>    bool gate (function *fun) final override { return fun->assume_function; }

Regarding your second mail, see above gate, this pass is only run for
assume functions and nothing else.

>    unsigned int execute (function *) final override
>      {
> +      assume_query query;
> +      if (dump_file)
> +	fprintf (dump_file, "Assumptions :\n--------------\n");
> +      for (unsigned i = 0; i < num_ssa_names; i++)
> +	{
> +	  tree name = ssa_name (i);
> +	  if (!name || !gimple_range_ssa_p (name))
> +	    continue;
> +	  tree type = TREE_TYPE (name);
> +	  if (!Value_Range::supports_type_p (type))
> +	    continue;
> +	  Value_Range assume_range (type);
> +	  if (query.assume_range_p (assume_range, name))
> +	    {
> +	      set_range_info (name, assume_range);
> +	      if (dump_file)
> +		{
> +		  print_generic_expr (dump_file, name, TDF_SLIM);
> +		  fprintf (dump_file, " -> ");
> +		  assume_range.dump (dump_file);
> +		  fputc ('\n', dump_file);
> +		}
> +	    }
> +	}
> +      if (dump_file)
> +	{
> +	  fputc ('\n', dump_file);
> +	  gimple_dump_cfg (dump_file, dump_flags);
> +	}
>        return TODO_discard_function;
>      }

	Jakub


  reply	other threads:[~2022-10-19 17:39 UTC|newest]

Thread overview: 33+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2022-10-10  8:54 [PATCH] middle-end " Jakub Jelinek
2022-10-10 21:09 ` Jason Merrill
2022-10-10 21:19   ` Jakub Jelinek
2022-10-11 13:36     ` [PATCH] middle-end, v2: " Jakub Jelinek
2022-10-12 15:48       ` Jason Merrill
2022-10-13  6:50         ` [PATCH] middle-end, v3: " Jakub Jelinek
2022-10-14 11:27           ` Richard Biener
2022-10-14 18:33             ` Jakub Jelinek
2022-10-17  6:55               ` Richard Biener
2022-10-17 15:44             ` [PATCH] middle-end, v4: " Jakub Jelinek
2022-10-18  7:00               ` Richard Biener
2022-10-18 21:31               ` Andrew MacLeod
2022-10-19 16:06                 ` Jakub Jelinek
2022-10-19 16:55                   ` Andrew MacLeod
2022-10-19 17:39                     ` Jakub Jelinek [this message]
2022-10-19 17:41                       ` Jakub Jelinek
2022-10-19 18:25                         ` Andrew MacLeod
2022-10-19 17:14                   ` Andrew MacLeod
2022-10-11 18:05 ` [PATCH] middle-end " Andrew MacLeod
2022-10-12 10:15   ` Jakub Jelinek
2022-10-12 14:31     ` Andrew MacLeod
2022-10-12 14:39       ` Jakub Jelinek
2022-10-12 16:12         ` Andrew MacLeod
2022-10-13  8:11           ` Richard Biener
2022-10-13  9:53             ` Jakub Jelinek
2022-10-13 13:16               ` Andrew MacLeod
2022-10-13  9:57           ` Jakub Jelinek
2022-10-17 17:53     ` Andrew MacLeod
2022-10-14 20:43 ` Martin Uecker
2022-10-14 21:20   ` Jakub Jelinek
2022-10-15  8:07     ` Martin Uecker
2022-10-15  8:53       ` Jakub Jelinek
2022-10-17  5:52         ` Martin Uecker

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=Y1A2Ofdwdcm93s03@tucnak \
    --to=jakub@redhat.com \
    --cc=amacleod@redhat.com \
    --cc=gcc-patches@gcc.gnu.org \
    --cc=jh@suse.cz \
    --cc=rguenther@suse.de \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).