Patchwork Fix combiner ICEs after my recent patch (PR rtl-optimization/48549)

login
register
mail settings
Submitter Jakub Jelinek
Date April 11, 2011, 10:49 a.m.
Message ID <20110411104958.GY17079@tyan-ft48-01.lab.bos.redhat.com>
Download mbox | patch
Permalink /patch/90585/
State New
Headers show

Comments

Jakub Jelinek - April 11, 2011, 10:49 a.m.
Hi!

On the testcase below we ICE since my PR48343 patch.
The problem is that update_cfg_for_uncondjump delete_insns the noop move,
if the insn to be deleted is last_combined_insn (which can be both
if it is i3 or when retrying and undobuf.other_insn happens to be
last_combined_insn), next propagate_for_debug won't find the deleted
insn in the stream and thus will search through the whole rest of the
function and crash on dereferencing NULL NEXT_INSN.

The following patch fixes it, bootstrapped/regtested on x86_64-linux
and i686-linux (also with the patch attached after it), but it is
mainly for discussion what to do.

The propagate_for_debug change alone could fix it, we should never
fall through into next basic block.  We are unforuntately not deleting
just jumps (which ought to appear at the end of bbs), but also
any other noop moves, which I think is unintentional, we have
delete_noop_moves that should clean that up instead (see the second patch).
With the second patch, bootstrap succeeded even with gcc_assert (at_end);
in update_cfg_for_uncondjump, so either it is always at the end, or at
least very often, thus perhaps with the second patch being applied
in combine_instructions we could just for INSN_DELETED_P clear
last_combined_insn right away, meaning we want to propagate till end of
current bb.

2011-04-11  Jakub Jelinek  <jakub@redhat.com>

	PR rtl-optimization/48549
	* combine.c (propagate_for_debug): Also stop after BB_END of
	this_basic_block.
	(combine_instructions): If last_combined_insn has been deleted,
	set last_combined_insn to the next nondebug insn, or NULL
	if after the end of BB.
	(try_combine): Handle last_combined_insn being NULL.

	* g++.dg/opt/pr48549.C: New test.


	Jakub
2011-04-11  Jakub Jelinek  <jakub@redhat.com>

	* combine.c (try_combine): Don't call update_cfg_for_uncondjump
	for noop non-jump moves.

--- gcc/combine.c.jj	2011-04-09 19:29:19.083421147 +0200
+++ gcc/combine.c	2011-04-11 10:27:41.769671151 +0200
@@ -4402,7 +4421,8 @@ try_combine (rtx i3, rtx i2, rtx i1, rtx
 
   /* A noop might also need cleaning up of CFG, if it comes from the
      simplification of a jump.  */
-  if (GET_CODE (newpat) == SET
+  if (JUMP_P (i3)
+      && GET_CODE (newpat) == SET
       && SET_SRC (newpat) == pc_rtx
       && SET_DEST (newpat) == pc_rtx)
     {
@@ -4411,6 +4431,7 @@ try_combine (rtx i3, rtx i2, rtx i1, rtx
     }
 
   if (undobuf.other_insn != NULL_RTX
+      && JUMP_P (undobuf.other_insn)
       && GET_CODE (PATTERN (undobuf.other_insn)) == SET
       && SET_SRC (PATTERN (undobuf.other_insn)) == pc_rtx
       && SET_DEST (PATTERN (undobuf.other_insn)) == pc_rtx)
Eric Botcazou - April 12, 2011, 7:18 a.m.
> The propagate_for_debug change alone could fix it, we should never
> fall through into next basic block.  We are unforuntately not deleting
> just jumps (which ought to appear at the end of bbs), but also
> any other noop moves, which I think is unintentional, we have
> delete_noop_moves that should clean that up instead (see the second patch).

Yes, I wondered several times about that.  This clearly looks unintentional and 
I agree that it is worth fixing on the mainline, adding gcc_assert (at_end); to 
update_cfg_for_uncondjump in the process.

> With the second patch, bootstrap succeeded even with gcc_assert (at_end);
> in update_cfg_for_uncondjump, so either it is always at the end, or at
> least very often, thus perhaps with the second patch being applied
> in combine_instructions we could just for INSN_DELETED_P clear
> last_combined_insn right away, meaning we want to propagate till end of
> current bb.

Something I don't understand at the moment: why moving last_combined_insn down 
to the next insn instead of up to the previous one when it has been deleted?
Wouldn't the latter simplify things, in particular avoid setting it to NULL_RTX 
which looks rather awkward to me?

Patch

--- gcc/combine.c.jj	2011-04-10 19:05:25.000000000 +0200
+++ gcc/combine.c	2011-04-11 09:13:04.000000000 +0200
@@ -1179,7 +1179,7 @@  combine_instructions (rtx f, unsigned in
 
   FOR_EACH_BB (this_basic_block)
     {
-      rtx last_combined_insn = NULL_RTX;
+      rtx last_combined_insn = pc_rtx;
       optimize_this_for_speed_p = optimize_bb_for_speed_p (this_basic_block);
       last_call_luid = 0;
       mem_last_set = -1;
@@ -1198,9 +1198,26 @@  combine_instructions (rtx f, unsigned in
 	  next = 0;
 	  if (NONDEBUG_INSN_P (insn))
 	    {
-	      if (last_combined_insn == NULL_RTX
-		  || DF_INSN_LUID (last_combined_insn) < DF_INSN_LUID (insn))
+	      if (last_combined_insn == pc_rtx)
 		last_combined_insn = insn;
+	      else if (last_combined_insn)
+		{
+		  if (INSN_DELETED_P (last_combined_insn))
+		    {
+		      while (last_combined_insn
+			     != NEXT_INSN (BB_END (this_basic_block))
+			     && (!NONDEBUG_INSN_P (last_combined_insn)
+				 || INSN_DELETED_P (last_combined_insn)))
+			last_combined_insn = NEXT_INSN (last_combined_insn);
+		      if (last_combined_insn
+			  == NEXT_INSN (BB_END (this_basic_block)))
+			last_combined_insn = NULL_RTX;
+		    }
+		  if (last_combined_insn
+		      && DF_INSN_LUID (last_combined_insn)
+			 <= DF_INSN_LUID (insn))
+		    last_combined_insn = insn;
+		}
 
 	      /* See if we know about function return values before this
 		 insn based upon SUBREG flags.  */
@@ -2451,14 +2468,14 @@  propagate_for_debug_subst (rtx from, con
 static void
 propagate_for_debug (rtx insn, rtx last, rtx dest, rtx src)
 {
-  rtx next, loc;
+  rtx next, loc, end = NEXT_INSN (BB_END (this_basic_block));
 
   struct rtx_subst_pair p;
   p.to = src;
   p.adjusted = false;
 
   next = NEXT_INSN (insn);
-  while (next != last)
+  while (next != last && next != end)
     {
       insn = next;
       next = NEXT_INSN (insn);
@@ -3904,8 +3921,9 @@  try_combine (rtx i3, rtx i2, rtx i1, rtx
 		    first = i3;
 		    last = undobuf.other_insn;
 		    gcc_assert (last);
-		    if (DF_INSN_LUID (last)
-			< DF_INSN_LUID (last_combined_insn))
+		    if (last_combined_insn == NULL_RTX
+			|| (DF_INSN_LUID (last)
+			    < DF_INSN_LUID (last_combined_insn)))
 		      last = last_combined_insn;
 		  }
 
--- gcc/testsuite/g++.dg/opt/pr48549.C.jj	2011-04-11 08:35:03.000000000 +0200
+++ gcc/testsuite/g++.dg/opt/pr48549.C	2011-04-11 08:34:19.000000000 +0200
@@ -0,0 +1,63 @@ 
+// PR rtl-optimization/48549
+// { dg-do compile }
+// { dg-options "-fcompare-debug -O2" }
+
+void
+foo (void *from, void *to)
+{
+  long offset = reinterpret_cast <long>(to) - reinterpret_cast <long>(from);
+  if (offset != static_cast <int>(offset))
+    *(int *) 0xC0DE = 0;
+  reinterpret_cast <int *>(from)[1] = offset;
+}
+struct A
+{
+  A () : a () {}
+  A (void *x) : a (x) {}
+  void *bar () { return a; }
+  void *a;
+};
+struct C;
+struct D;
+struct E : public A
+{
+  C m1 (int);
+  D m2 ();
+  E () {}
+  E (A x) : A (x) {}
+};
+struct C : public E
+{
+  C () {}
+  C (void *x) : E (x) {}
+};
+struct D : public E
+{
+  D (void *x) : E (x) {}
+};
+C
+E::m1 (int x)
+{
+  return (reinterpret_cast <char *>(bar ()) + x);
+}
+D
+E::m2 ()
+{
+  return reinterpret_cast <char *>(bar ());
+}
+struct B
+{
+  E a;
+  unsigned b : 16;
+  unsigned c : 1;
+};
+void
+baz (B *x)
+{
+  for (unsigned i = 0; i < 64; i++)
+    {
+      D d = x[i].a.m2 ();
+      C c = x[i].a.m1 (x[i].c);
+      foo (d.bar (), c.bar ());
+    }
+}