public inbox for gcc-patches@gcc.gnu.org
 help / color / mirror / Atom feed
From: Jan Hubicka <hubicka@kam.mff.cuni.cz>
To: gcc-patches@gcc.gnu.org, rguenther@suse.de
Subject: Add EAF_NOT_RETURNED_DIRECTLY
Date: Mon, 1 Nov 2021 22:39:15 +0100	[thread overview]
Message-ID: <20211101213915.GD14166@kam.mff.cuni.cz> (raw)

Hi,
this patch adds EAF_NOT_RETURNED_DIRECTLY which works similarly as
EAF_NODIRECTESCAPE.  Values pointed to by a given argument may be returned but
not the argument itself.  This helps PTA quite noticeably because we mostly
care about tracking points to which given memory location can escape.

I think this is about last reasonable improvement we can get to EAF
flags.

cc1plus disambiguation counts change from:

Alias oracle query stats:
  refs_may_alias_p: 77976088 disambiguations, 98744590 queries
  ref_maybe_used_by_call_p: 572845 disambiguations, 79014622 queries
  call_may_clobber_ref_p: 340823 disambiguations, 344823 queries
  nonoverlapping_component_refs_p: 0 disambiguations, 26590 queries
  nonoverlapping_refs_since_match_p: 31626 disambiguations, 65379 must overlaps, 97963 queries
  aliasing_component_refs_p: 57414 disambiguations, 11434878 queries
  TBAA oracle: 27749649 disambiguations 91624184 queries
               14733408 are in alias set 0
               8847139 queries asked about the same object
               139 queries asked about the same alias set
               0 access volatile
               38412201 are dependent in the DAG
               1881648 are aritificially in conflict with void *

Modref stats:
  modref use: 23785 disambiguations, 702425 queries
  modref clobber: 2296391 disambiguations, 22690531 queries
  5260226 tbaa queries (0.231825 per modref query)
  731741 base compares (0.032249 per modref query)

PTA query stats:
  pt_solution_includes: 12580233 disambiguations, 35854408 queries
  pt_solutions_intersect: 1409041 disambiguations, 13496899 queries

To:

Alias oracle query stats:
  refs_may_alias_p: 78304485 disambiguations, 98830913 queries
  ref_maybe_used_by_call_p: 630360 disambiguations, 79308222 queries
  call_may_clobber_ref_p: 381549 disambiguations, 384627 queries
  nonoverlapping_component_refs_p: 0 disambiguations, 26299 queries
  nonoverlapping_refs_since_match_p: 29919 disambiguations, 64917 must overlaps, 95781 queries
  aliasing_component_refs_p: 57250 disambiguations, 11336880 queries
  TBAA oracle: 27835747 disambiguations 91534430 queries
               14884868 are in alias set 0
               8933627 queries asked about the same object
               123 queries asked about the same alias set
               0 access volatile
               37974723 are dependent in the DAG
               1905342 are aritificially in conflict with void *

Modref stats:
  modref use: 24929 disambiguations, 756294 queries
  modref clobber: 2334910 disambiguations, 23414495 queries
  5359212 tbaa queries (0.228884 per modref query)
  754642 base compares (0.032230 per modref query)

PTA query stats:
  pt_solution_includes: 13262256 disambiguations, 36306509 queries
  pt_solutions_intersect: 1574672 disambiguations, 13638933 queries

So about 5% more pt_solution_includes and 11% more pt_solutions_intersect
disambiguations.

Bootstrapped/regtested x86_64-linux, OK?

Honza

gcc/ChangeLog:

	* tree-core.h (EAF_NOT_RETURNED_DIRECTLY): New flag.
	(EAF_NOREAD): Renumber.
	* ipa-modref.c (dump_eaf_flags): Dump EAF_NOT_RETURNED_DIRECTLY.
	(remove_useless_eaf_flags): Handle EAF_NOT_RETURNED_DIRECTLY
	(deref_flags): Likewise.
	(modref_lattice::init): Likewise.
	(modref_lattice::merge): Likewise.
	(merge_call_lhs_flags): Likewise.
	(analyze_ssa_name_flags): Likewise.
	(modref_merge_call_site_flags): Likewise.
	* tree-ssa-structalias.c (handle_call_arg): Likewise.

gcc/testsuite/ChangeLog:

	* gcc.dg/ipa/modref-3.c: New test.
	* gcc.dg/tree-ssa/modref-10.c: New test.

diff --git a/gcc/ipa-modref.c b/gcc/ipa-modref.c
index d866d9ed6b3..c0aae084dbd 100644
--- a/gcc/ipa-modref.c
+++ b/gcc/ipa-modref.c
@@ -160,6 +160,8 @@ dump_eaf_flags (FILE *out, int flags, bool newline = true)
     fprintf (out, " unused");
   if (flags & EAF_NOT_RETURNED)
     fprintf (out, " not_returned");
+  if (flags & EAF_NOT_RETURNED_DIRECTLY)
+    fprintf (out, " not_returned_directly");
   if (flags & EAF_NOREAD)
     fprintf (out, " noread");
   if (newline)
@@ -295,7 +297,7 @@ remove_useless_eaf_flags (int eaf_flags, int ecf_flags, bool returns_void)
   else if (ecf_flags & ECF_PURE)
     eaf_flags &= ~implicit_pure_eaf_flags;
   else if ((ecf_flags & ECF_NORETURN) || returns_void)
-    eaf_flags &= ~EAF_NOT_RETURNED;
+    eaf_flags &= ~(EAF_NOT_RETURNED | EAF_NOT_RETURNED_DIRECTLY);
   return eaf_flags;
 }
 
@@ -1373,7 +1375,7 @@ memory_access_to (tree op, tree ssa_name)
 static int
 deref_flags (int flags, bool ignore_stores)
 {
-  int ret = EAF_NODIRECTESCAPE;
+  int ret = EAF_NODIRECTESCAPE | EAF_NOT_RETURNED_DIRECTLY;
   /* If argument is unused just account for
      the read involved in dereference.  */
   if (flags & EAF_UNUSED)
@@ -1458,7 +1460,8 @@ modref_lattice::init ()
 {
   /* All flags we track.  */
   int f = EAF_DIRECT | EAF_NOCLOBBER | EAF_NOESCAPE | EAF_UNUSED
-	  | EAF_NODIRECTESCAPE | EAF_NOT_RETURNED | EAF_NOREAD;
+	  | EAF_NODIRECTESCAPE | EAF_NOT_RETURNED |
+	  EAF_NOT_RETURNED_DIRECTLY | EAF_NOREAD;
   flags = f;
   /* Check that eaf_flags_t is wide enough to hold all flags.  */
   gcc_checking_assert (f == flags);
@@ -1540,6 +1543,8 @@ modref_lattice::merge (int f)
      Fnspec machinery does set both so compensate for this.  */
   if (f & EAF_NOESCAPE)
     f |= EAF_NODIRECTESCAPE;
+  if (f & EAF_NOT_RETURNED)
+    f |= EAF_NOT_RETURNED_DIRECTLY;
   if ((flags & f) != flags)
     {
       flags &= f;
@@ -1647,7 +1652,9 @@ merge_call_lhs_flags (gcall *call, int arg, int index, bool deref,
       && (flags & ERF_RETURN_ARG_MASK) != arg)
     return;
 
-  if (gimple_call_arg_flags (call, arg) & (EAF_NOT_RETURNED | EAF_UNUSED))
+  int eaf_flags = gimple_call_arg_flags (call, arg);
+
+  if (eaf_flags & (EAF_NOT_RETURNED | EAF_UNUSED))
     return;
 
   /* If return value is SSA name determine its flags.  */
@@ -1655,12 +1662,14 @@ merge_call_lhs_flags (gcall *call, int arg, int index, bool deref,
     {
       tree lhs = gimple_call_lhs (call);
       analyze_ssa_name_flags (lhs, lattice, depth + 1, ipa);
-      if (deref)
+      if (deref || (eaf_flags & EAF_NOT_RETURNED_DIRECTLY))
 	lattice[index].merge_deref (lattice[SSA_NAME_VERSION (lhs)], false);
       else
 	lattice[index].merge (lattice[SSA_NAME_VERSION (lhs)]);
     }
   /* In the case of memory store we can do nothing.  */
+  else if (eaf_flags & EAF_NOT_RETURNED_DIRECTLY)
+    lattice[index].merge (deref_flags (0, false));
   else
     lattice[index].merge (0);
 }
@@ -1731,11 +1740,13 @@ analyze_ssa_name_flags (tree name, vec<modref_lattice> &lattice, int depth,
 	      && DECL_BY_REFERENCE (DECL_RESULT (current_function_decl)))
 	    ;
 	  else if (gimple_return_retval (ret) == name)
-	    lattice[index].merge (~(EAF_UNUSED | EAF_NOT_RETURNED));
+	    lattice[index].merge (~(EAF_UNUSED | EAF_NOT_RETURNED
+				    | EAF_NOT_RETURNED_DIRECTLY));
 	  else if (memory_access_to (gimple_return_retval (ret), name))
 	    {
 	      lattice[index].merge_direct_load ();
-	      lattice[index].merge (~(EAF_UNUSED | EAF_NOT_RETURNED));
+	      lattice[index].merge (~(EAF_UNUSED | EAF_NOT_RETURNED
+				      | EAF_NOT_RETURNED_DIRECTLY));
 	    }
 	}
       /* Account for LHS store, arg loads and flags from callee function.  */
@@ -1794,7 +1805,8 @@ analyze_ssa_name_flags (tree name, vec<modref_lattice> &lattice, int depth,
 		    if (!(ecf_flags & (ECF_CONST | ECF_NOVOPS)))
 		      {
 			int call_flags = gimple_call_arg_flags (call, i)
-					 | EAF_NOT_RETURNED;
+					 | EAF_NOT_RETURNED
+					 | EAF_NOT_RETURNED_DIRECTLY;
 			if (ignore_stores)
 			  call_flags |= ignore_stores_eaf_flags;
 
@@ -1817,7 +1829,8 @@ analyze_ssa_name_flags (tree name, vec<modref_lattice> &lattice, int depth,
 		      {
 			int call_flags = deref_flags
 			   (gimple_call_arg_flags (call, i)
-			    | EAF_NOT_RETURNED, ignore_stores);
+			    | EAF_NOT_RETURNED
+			    | EAF_NOT_RETURNED_DIRECTLY, ignore_stores);
 			if (!record_ipa)
 			  lattice[index].merge (call_flags);
 			else
@@ -3858,14 +3871,18 @@ modref_merge_call_site_flags (escape_summary *sum,
 	  flags_lto |= ignore_stores_eaf_flags;
 	}
       /* Returning the value is already accounted to at local propagation.  */
-      flags |= ee->min_flags | EAF_NOT_RETURNED;
-      flags_lto |= ee->min_flags | EAF_NOT_RETURNED;
+      flags |= ee->min_flags | EAF_NOT_RETURNED | EAF_NOT_RETURNED_DIRECTLY;
+      flags_lto |= ee->min_flags | EAF_NOT_RETURNED | EAF_NOT_RETURNED_DIRECTLY;
       /* Noescape implies that value also does not escape directly.
 	 Fnspec machinery does set both so compensate for this.  */
       if (flags & EAF_NOESCAPE)
 	flags |= EAF_NODIRECTESCAPE;
       if (flags_lto & EAF_NOESCAPE)
 	flags_lto |= EAF_NODIRECTESCAPE;
+      if (flags & EAF_NOT_RETURNED)
+	flags |= EAF_NOT_RETURNED_DIRECTLY;
+      if (flags_lto & EAF_NOT_RETURNED)
+	flags_lto |= EAF_NOT_RETURNED_DIRECTLY;
       if (!(flags & EAF_UNUSED)
 	  && cur_summary && ee->parm_index < (int)cur_summary->arg_flags.length ())
 	{
diff --git a/gcc/testsuite/gcc.dg/ipa/modref-3.c b/gcc/testsuite/gcc.dg/ipa/modref-3.c
new file mode 100644
index 00000000000..c90683bc962
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/ipa/modref-3.c
@@ -0,0 +1,20 @@
+/* { dg-options "-O2 -fdump-ipa-modref"  } */
+/* { dg-do link } */
+int *ptr;
+void linker_error ();
+int
+main ()
+{
+  int a;
+  __attribute__((noinline)) int test2 ()
+  {
+    ptr = 0;
+    return a;
+  }
+  a = 1;
+  test2 ();
+  if (a != 1)
+    linker_error ();
+  return 0;
+}
+/* { dg-final { scan-ipa-dump "Static chain flags: noclobber noescape nodirectescape" "modref" } */
diff --git a/gcc/testsuite/gcc.dg/tree-ssa/modref-10.c b/gcc/testsuite/gcc.dg/tree-ssa/modref-10.c
new file mode 100644
index 00000000000..c608408809d
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/tree-ssa/modref-10.c
@@ -0,0 +1,20 @@
+/* { dg-do link } */
+/* { dg-options "-O2 -fdump-tree-modref1"  } */
+void linker_error ();
+__attribute__((noinline))
+int *
+test (int **a)
+{
+  return *a;
+}
+int
+main()
+{
+	int val;
+	int *a=&val;
+	int *b = test (&a);
+	if (b == (int *)&a)
+		linker_error ();
+	return 0;
+}
+/* { dg-final { scan-tree-dump "parm 0 flags: noclobber noescape nodirectescape not_returned_directly" "modref1"} } */
diff --git a/gcc/tree-core.h b/gcc/tree-core.h
index d3d2a8d812f..8a55db2b3f5 100644
--- a/gcc/tree-core.h
+++ b/gcc/tree-core.h
@@ -117,8 +117,12 @@ struct die_struct;
 /* Nonzero if the argument does not escape to return value.  */
 #define EAF_NOT_RETURNED	(1 << 5)
 
+/* Nonzero if the argument argument itself does not escape
+   to return value but memory referenced by it may escape.  */
+#define EAF_NOT_RETURNED_DIRECTLY (1 << 6)
+
 /* Nonzero if the argument is not read.  */
-#define EAF_NOREAD		(1 << 6)
+#define EAF_NOREAD		(1 << 7)
 
 /* Call return flags.  */
 /* Mask for the argument number that is returned.  Lower two bits of
diff --git a/gcc/tree-ssa-structalias.c b/gcc/tree-ssa-structalias.c
index 99072df0768..5f24c0171f0 100644
--- a/gcc/tree-ssa-structalias.c
+++ b/gcc/tree-ssa-structalias.c
@@ -4082,8 +4082,16 @@ handle_call_arg (gcall *stmt, tree arg, vec<ce_s> *results, int flags,
     {
       struct constraint_expr cexpr;
       cexpr.var = tem->id;
-      cexpr.type = SCALAR;
-      cexpr.offset = 0;
+      if (flags & EAF_NOT_RETURNED_DIRECTLY)
+	{
+	  cexpr.type = DEREF;
+	  cexpr.offset = UNKNOWN_OFFSET;
+	}
+      else
+	{
+	  cexpr.type = SCALAR;
+	  cexpr.offset = 0;
+	}
       results->safe_push (cexpr);
     }
 

             reply	other threads:[~2021-11-01 21:39 UTC|newest]

Thread overview: 2+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2021-11-01 21:39 Jan Hubicka [this message]
2021-11-02 13:45 ` Richard Biener

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=20211101213915.GD14166@kam.mff.cuni.cz \
    --to=hubicka@kam.mff.cuni.cz \
    --cc=gcc-patches@gcc.gnu.org \
    --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).