[incremental] Patch: FYI: delay gimplification + partial cgraph revert

[incremental] Patch: FYI: delay gimplification + partial cgraph revert

[Tom Tromey]

I'm checking this in on the incremental-compiler branch.

The cgraph change I made the other day didn't really do that well in
all cases; it ICEd in some inlining scenarios.  So rather than
continue down this path I thought it would be better to defer
gimplification until after the server forks for code generation.  This
should save memory in the server as well.

This patch is based on a patch that Michael Matz wrote, hence his name
in the ChangeLog.  Complain to me if you hate it, though.

The cgraph_finalize_compilation_unit part of this patch will need a
bit of work... some in the other front ends, and then (maybe) some for
--combine.

Though the current patch is flawed, I think this is a good direction
for the compiler to go.  It helps to further separate the front and
back ends.  Michael pointed out that c_gimple_diagnostics_recursively
needs to be done in an FE-independent way, but from what I can tell
both c_warn_unused_result and diagnose_omp_structured_block_errors are
already language independent.

Tom

ChangeLog:
2007-09-28  Tom Tromey  <tromey@redhat.com>
	    Michael Matz  <matz@suse.de>

	* c-gimplify.c (c_genericize): Don't call gimplify_function_tree.
	(c_gimplify_expr): Handle CALL_EXPR, FDESC_EXPR, and ADDR_EXPR.
	* tree-optimize.c (tree_lowering_passes): Call
	gimplify_function_tree.
	* c-decl.c (pop_file_scope): Don't call
	cgraph_finalize_compilation_unit.
	(finish_function): Don't call c_gimple_diagnostics_recursively.
	* toplev.c (compile_file): Call cgraph_finalize_compilation_unit.
	* cgraphunit.c (verify_cgraph_node): Reverted.
	* cgraph.h (cgraph_get_callee_fndecl): Removed.
	* cgraphbuild.c (build_cgraph_edges): Reverted.
	(rebuild_cgraph_edges): Likewise.
	* cgraph.c (cgraph_note_duplicate): Clean up.
	(cgraph_get_callee_fndecl): Removed.

Index: cgraphbuild.c
===================================================================
--- cgraphbuild.c	(revision 128823)
+++ cgraphbuild.c	(working copy)
@@ -130,7 +130,7 @@
 	tree call = get_call_expr_in (stmt);
 	tree decl;
 
-	if (call && (decl = cgraph_get_callee_fndecl (call)))
+	if (call && (decl = get_callee_fndecl (call)))
 	  {
 	    int i;
 	    int n = call_expr_nargs (call);
@@ -224,7 +224,7 @@
 	tree call = get_call_expr_in (stmt);
 	tree decl;
 
-	if (call && (decl = cgraph_get_callee_fndecl (call)))
+	if (call && (decl = get_callee_fndecl (call)))
 	  {
 	    int freq = (!bb->frequency && !entry_freq ? CGRAPH_FREQ_BASE
 			: bb->frequency * CGRAPH_FREQ_BASE / entry_freq);
Index: cgraph.c
===================================================================
--- cgraph.c	(revision 128823)
+++ cgraph.c	(working copy)
@@ -270,33 +270,12 @@
   if (!duplicate_map)
     duplicate_map = pointer_map_create ();
 
-  /* FIXME: this seems bogus & lame.  */
-
-  if (DECL_INITIAL (duplicate))
-    {
-      /* FIXME: shouldn't be able to see a DECL_INITIAL here.
-	 Something is decl-smashing again.  */
-      gcc_assert (!DECL_INITIAL (decl)
-		  || DECL_INITIAL (decl) == DECL_INITIAL (duplicate));
-      slot = pointer_map_insert (duplicate_map, decl);
-      *slot = duplicate;
-    }
-  else
-    {
-      pointer_map_insert (duplicate_map, duplicate);
-      *slot = decl;
-    }
+  /* Due to the way decl smashing works, the most recent decl is
+     always canonical.  */
+  slot = pointer_map_insert (duplicate_map, decl);
+  *slot = duplicate;
 }
 
-/* Like get_callee_fndecl, but returns the canonical target of the
-   call.  */
-tree
-cgraph_get_callee_fndecl (tree expr)
-{
-  tree result = get_callee_fndecl (expr);
-  return cgraph_canonical_decl (result);
-}
-
 /* Insert already constructed node into hashtable.  */
 
 void
Index: cgraph.h
===================================================================
--- cgraph.h	(revision 128823)
+++ cgraph.h	(working copy)
@@ -290,7 +290,6 @@
 void cgraph_reset (void);
 void cgraph_note_duplicate (tree, tree);
 tree cgraph_canonical_decl (tree);
-tree cgraph_get_callee_fndecl (tree);
 void dump_cgraph (FILE *);
 void dump_cgraph_node (FILE *, struct cgraph_node *);
 void cgraph_insert_node_to_hashtable (struct cgraph_node *node);
Index: toplev.c
===================================================================
--- toplev.c	(revision 128735)
+++ toplev.c	(working copy)
@@ -1067,6 +1067,8 @@
   if (server_mode && server_start_back_end ())
     return false;
 
+  cgraph_finalize_compilation_unit ();
+
   lang_hooks.decls.final_write_globals ();
 
   if (errorcount || sorrycount)
Index: cgraphunit.c
===================================================================
--- cgraphunit.c	(revision 128823)
+++ cgraphunit.c	(working copy)
@@ -689,7 +689,7 @@
 		tree stmt = bsi_stmt (bsi);
 		tree call = get_call_expr_in (stmt);
 		tree decl;
-		if (call && (decl = cgraph_get_callee_fndecl (call)))
+		if (call && (decl = get_callee_fndecl (call)))
 		  {
 		    struct cgraph_edge *e = cgraph_edge (node, stmt);
 		    if (e)
Index: c-decl.c
===================================================================
--- c-decl.c	(revision 128823)
+++ c-decl.c	(working copy)
@@ -983,7 +983,6 @@
   file_scope = 0;
 
   maybe_apply_pending_pragma_weaks ();
-  cgraph_finalize_compilation_unit ();
 }
 
 /* Insert BLOCK at the end of the list of subblocks of the current
@@ -7024,7 +7023,8 @@
       if (!decl_function_context (fndecl))
 	{
 	  c_genericize (fndecl);
-	  c_gimple_diagnostics_recursively (fndecl);
+	  /* FIXME: this should be a generic pass.  */
+/* 	  c_gimple_diagnostics_recursively (fndecl); */
 
 	  /* ??? Objc emits functions after finalizing the compilation unit.
 	     This should be cleaned up later and this conditional removed.  */
Index: tree-optimize.c
===================================================================
--- tree-optimize.c	(revision 127650)
+++ tree-optimize.c	(working copy)
@@ -355,6 +355,7 @@
   current_function_decl = fn;
   push_cfun (DECL_STRUCT_FUNCTION (fn));
   tree_register_cfg_hooks ();
+  gimplify_function_tree (fn);
   bitmap_obstack_initialize (NULL);
   execute_pass_list (all_lowering_passes);
   if (optimize && cgraph_global_info_ready)
Index: c-gimplify.c
===================================================================
--- c-gimplify.c	(revision 127650)
+++ c-gimplify.c	(working copy)
@@ -101,9 +101,6 @@
       dump_end (TDI_original, dump_orig);
     }
 
-  /* Go ahead and gimplify for now.  */
-  gimplify_function_tree (fndecl);
-
   /* Dump the genericized tree IR.  */
   dump_function (TDI_generic, fndecl);
 
@@ -230,6 +227,23 @@
 	TREE_NO_WARNING (DECL_EXPR_DECL (*expr_p)) = 1;
       return GS_UNHANDLED;
 
+    case CALL_EXPR:
+      {
+ 	tree callee = TREE_OPERAND (*expr_p, 1);
+	if (TREE_CODE (callee) == FUNCTION_DECL)
+	  TREE_OPERAND (*expr_p, 1) = cgraph_canonical_decl (callee);
+      }
+      return GS_UNHANDLED;
+
+    case FDESC_EXPR:
+    case ADDR_EXPR:
+      {
+ 	tree decl = TREE_OPERAND (*expr_p, 0);
+	if (TREE_CODE (decl) == FUNCTION_DECL)
+	  TREE_OPERAND (*expr_p, 0) = cgraph_canonical_decl (decl);
+      }
+      return GS_UNHANDLED;
+
     case COMPOUND_LITERAL_EXPR:
       return gimplify_compound_literal_expr (expr_p, pre_p);
 

Re: [incremental] Patch: FYI: delay gimplification + partial cgraph revert

[Diego Novillo]

On 9/28/07, Tom Tromey <tromey@redhat.com> wrote:

> Though the current patch is flawed, I think this is a good direction
> for the compiler to go.

I agree.  Would you consider moving this into mainline when stage1
opens again?  Delaying gimplification on the C/C++ FEs is a Good
Thing.

Thanks.

Re: [incremental] Patch: FYI: delay gimplification + partial cgraph revert

[Tom Tromey]

>>>>> "Diego" == Diego Novillo <dnovillo@google.com> writes:

Diego> On 9/28/07, Tom Tromey <tromey@redhat.com> wrote:
>> Though the current patch is flawed, I think this is a good direction
>> for the compiler to go.

Diego> I agree.  Would you consider moving this into mainline when stage1
Diego> opens again?  Delaying gimplification on the C/C++ FEs is a Good
Diego> Thing.

Sure thing.  Though perhaps Michael will beat me to it :-)

Tom

Re: [incremental] Patch: FYI: delay gimplification + partial cgraph revert

[Richard Guenther]

On 9/29/07, Tom Tromey <tromey@redhat.com> wrote:
> >>>>> "Diego" == Diego Novillo <dnovillo@google.com> writes:
>
> Diego> On 9/28/07, Tom Tromey <tromey@redhat.com> wrote:
> >> Though the current patch is flawed, I think this is a good direction
> >> for the compiler to go.
>
> Diego> I agree.  Would you consider moving this into mainline when stage1
> Diego> opens again?  Delaying gimplification on the C/C++ FEs is a Good
> Diego> Thing.
>
> Sure thing.  Though perhaps Michael will beat me to it :-)

There were concerns about memory usage regressions (it's not clear
whether the un-gimplified or the gimplified variant takes more memory).

Richard.

Re: [incremental] Patch: FYI: delay gimplification + partial cgraph revert

[Tom Tromey]

>>>>> "Richard" == Richard Guenther <richard.guenther@gmail.com> writes:

Richard> There were concerns about memory usage regressions (it's not clear
Richard> whether the un-gimplified or the gimplified variant takes more memory).

Thanks.  I made a note to look at this.

Tom