[PATCH, PR68715] Add missing single_pred_p test in scop_detection::merge_sese

[PATCH, PR68715] Add missing single_pred_p test in scop_detection::merge_sese

[Tom de Vries]

[-- Attachment #1: Type: text/plain, Size: 1546 bytes --]

Hi,

this patch fixes graphite PR68715, a 6 regression.

In scop_detection::merge_sese, we check if the exit bb of the merged 
sese region is dominated by the entry bb:
...
   if (...
       || !dominated_by_p (CDI_DOMINATORS, get_exit_bb (combined),
                           get_entry_bb (combined)))
     {
       ...
       return invalid_sese;
     }
...

Subsequently, we check for an empty exit bb, and if that one's not 
empty, we try to merge an empty successor block into the sese:
...
   /* FIXME: We should remove this piece of code once
      canonicalize_loop_closed_ssa has been removed, because that
      function adds a BB with single exit.  */
   if (!trivially_empty_bb_p (get_exit_bb (combined)))
     {
       /* Find the first empty succ (with single exit) of
          combined.exit.  */
       basic_block imm_succ = combined.exit->dest;
       if (single_succ_p (imm_succ)
           && trivially_empty_bb_p (imm_succ))
         combined.exit = single_succ_edge (imm_succ);
       else
         {
           ...
           return invalid_sese;
         }
     }
...

However, when the imm_succ block has more than one predecessor, merging 
the block into the sese region breaks the property that the exit bb is 
dominated by the entry bb. We then run into an ICE in 
harmful_loop_in_region a bit later, when re-checking that property.

The patch fixes this by adding a test for 'single_pred_p (imm_succ)' in 
the empty-block-merge condition.

Bootstrapped and reg-tested on x86_64.

OK for stage4 trunk?

Thanks,
- Tom

[-- Attachment #2: 0002-Add-missing-single_pred_p-test-in-scop_detection-merge_sese.patch --]
[-- Type: text/x-patch, Size: 3626 bytes --]

Add missing single_pred_p test in scop_detection::merge_sese

2016-03-15  Tom de Vries  <tom@codesourcery.com>

	PR tree-optimization/68715
	* graphite-scop-detection.c (scop_detection::merge_sese): Add missing
	single_pred_p test.

	* gcc.dg/graphite/pr68715-2.c: New test.
	* gcc.dg/graphite/pr68715.c: New test.
	* gfortran.dg/graphite/pr68715.f90: New test.

---
 gcc/graphite-scop-detection.c                  |  4 ++-
 gcc/testsuite/gcc.dg/graphite/pr68715-2.c      | 35 +++++++++++++++++++++++++
 gcc/testsuite/gcc.dg/graphite/pr68715.c        | 36 ++++++++++++++++++++++++++
 gcc/testsuite/gfortran.dg/graphite/pr68715.f90 | 31 ++++++++++++++++++++++
 4 files changed, 105 insertions(+), 1 deletion(-)

diff --git a/gcc/graphite-scop-detection.c b/gcc/graphite-scop-detection.c
index 9161cb7..f0c13ee 100644
--- a/gcc/graphite-scop-detection.c
+++ b/gcc/graphite-scop-detection.c
@@ -836,7 +836,9 @@ scop_detection::merge_sese (sese_l first, sese_l second) const
     {
       /* Find the first empty succ (with single exit) of combined.exit.  */
       basic_block imm_succ = combined.exit->dest;
-      if (single_succ_p (imm_succ) && trivially_empty_bb_p (imm_succ))
+      if (single_succ_p (imm_succ)
+	  && single_pred_p (imm_succ)
+	  && trivially_empty_bb_p (imm_succ))
 	combined.exit = single_succ_edge (imm_succ);
       else
 	{
diff --git a/gcc/testsuite/gcc.dg/graphite/pr68715-2.c b/gcc/testsuite/gcc.dg/graphite/pr68715-2.c
new file mode 100644
index 0000000..270d948
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/graphite/pr68715-2.c
@@ -0,0 +1,35 @@
+/* { dg-do compile } */
+/* { dg-options "-Ofast -floop-interchange" } */
+
+int a, b, c, d, f, g;
+int e[1], h[1];
+void fn2 ();
+void fn3 ();
+void
+fn1 ()
+{
+  fn2 ();
+  b = 0;
+  for (; b < 10; b++)
+    ;
+}
+
+void
+fn2 ()
+{
+  if (a)
+    {
+      fn3 ();
+      c = d;
+    }
+}
+
+void
+fn3 ()
+{
+  for (; g; g++)
+    e[g] = 2;
+  if (f)
+    for (; g; g++)
+      h[g] = 5;
+}
diff --git a/gcc/testsuite/gcc.dg/graphite/pr68715.c b/gcc/testsuite/gcc.dg/graphite/pr68715.c
new file mode 100644
index 0000000..14da2fb
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/graphite/pr68715.c
@@ -0,0 +1,36 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -floop-interchange" } */
+
+int a[1], c[1];
+int b, d, e;
+
+void
+fn1 (int p1)
+{
+  for (;;)
+    ;
+}
+
+int
+fn3 ()
+{
+  for (; e; e++)
+    c[e] = 2;
+  for (; d; d--)
+    a[d] = 8;
+  return 0;
+}
+
+int fn5 (int);
+
+int
+fn2 ()
+{
+  fn3 ();
+}
+
+void
+fn4 ()
+{
+  fn1 (b || fn5 (fn2 ()));
+}
diff --git a/gcc/testsuite/gfortran.dg/graphite/pr68715.f90 b/gcc/testsuite/gfortran.dg/graphite/pr68715.f90
new file mode 100644
index 0000000..c011756
--- /dev/null
+++ b/gcc/testsuite/gfortran.dg/graphite/pr68715.f90
@@ -0,0 +1,31 @@
+! { dg-do compile }
+! { dg-options "-floop-nest-optimize -O1" }
+
+SUBROUTINE se_core_core_interaction(calculate_forces)
+  INTEGER, PARAMETER :: dp=8
+  LOGICAL, INTENT(in)		 :: calculate_forces
+  REAL(KIND=dp), DIMENSION(3)	 :: force_ab, rij
+  LOGICAL :: lfoo,kfoo,mfoo,nfoo,ffoo
+  INTEGER, PARAMETER :: mi2=42
+  CALL dummy(lfoo,kfoo,mfoo,nfoo,method_id,core_core)
+  IF (lfoo) THEN
+     DO WHILE (ffoo())
+	IF (lfoo) CYCLE
+	IF (kfoo) CYCLE
+	dr1 = DOT_PRODUCT(rij,rij)
+	IF (dr1 > rij_threshold) THEN
+	   SELECT CASE (method_id)
+	   CASE (mi2)
+	      IF (calculate_forces) THEN
+		 CALL dummy2(force_ab)
+		 IF (nfoo) THEN
+		    force_ab = force_ab + core_core*dr3inv
+		 END IF
+	      END IF
+	   END SELECT
+	END IF
+	enuclear = enuclear + enucij
+     END DO
+     CALL dummy3(enuclear)
+  END IF
+END SUBROUTINE se_core_core_interaction

Re: [PATCH, PR68715] Add missing single_pred_p test in scop_detection::merge_sese

[Richard Biener]

On Wed, 16 Mar 2016, Tom de Vries wrote:

> Hi,
> 
> this patch fixes graphite PR68715, a 6 regression.
> 
> In scop_detection::merge_sese, we check if the exit bb of the merged sese
> region is dominated by the entry bb:
> ...
>   if (...
>       || !dominated_by_p (CDI_DOMINATORS, get_exit_bb (combined),
>                           get_entry_bb (combined)))
>     {
>       ...
>       return invalid_sese;
>     }
> ...
> 
> Subsequently, we check for an empty exit bb, and if that one's not empty, we
> try to merge an empty successor block into the sese:
> ...
>   /* FIXME: We should remove this piece of code once
>      canonicalize_loop_closed_ssa has been removed, because that
>      function adds a BB with single exit.  */
>   if (!trivially_empty_bb_p (get_exit_bb (combined)))
>     {
>       /* Find the first empty succ (with single exit) of
>          combined.exit.  */
>       basic_block imm_succ = combined.exit->dest;
>       if (single_succ_p (imm_succ)
>           && trivially_empty_bb_p (imm_succ))
>         combined.exit = single_succ_edge (imm_succ);
>       else
>         {
>           ...
>           return invalid_sese;
>         }
>     }
> ...
> 
> However, when the imm_succ block has more than one predecessor, merging the
> block into the sese region breaks the property that the exit bb is dominated
> by the entry bb. We then run into an ICE in harmful_loop_in_region a bit
> later, when re-checking that property.
> 
> The patch fixes this by adding a test for 'single_pred_p (imm_succ)' in the
> empty-block-merge condition.
> 
> Bootstrapped and reg-tested on x86_64.
> 
> OK for stage4 trunk?

Hmm, it looks like for all what this function does this effectively
pessimizes scop merging and it would be easier to split 'exit'
in case its destination is unsuitable (not trivially empty).

The

  /* For now we just want to bail out when exit does not post-dominate 
entry.
     TODO: We might just add a basic_block at the exit to make exit
     post-dominate entry (the entire region).  */
  if (!dominated_by_p (CDI_POST_DOMINATORS, get_entry_bb (combined),
                       get_exit_bb (combined))
      || !dominated_by_p (CDI_DOMINATORS, get_exit_bb (combined),

comment also suggests that splitting the get_nearest_pdom_with_single_exit
edge and including the new BB in the combined region would also always fix
the dominance relation (though I don't see how it could do that and
the comment looks wrong and by construction the check should never
trigger).

Otherwise the patch is certainly fine of course.

Thanks,
Richard.

Re: [PATCH, PR68715] Add missing single_pred_p test in scop_detection::merge_sese

[Tom de Vries]

On 16/03/16 09:53, Richard Biener wrote:
> Hmm, it looks like for all what this function does this effectively
> pessimizes scop merging and it would be easier to split 'exit'
> in case its destination is unsuitable (not trivially empty).
>

Agreed.

> The
>
>    /* For now we just want to bail out when exit does not post-dominate
> entry.
>       TODO: We might just add a basic_block at the exit to make exit
>       post-dominate entry (the entire region).  */
>    if (!dominated_by_p (CDI_POST_DOMINATORS, get_entry_bb (combined),
>                         get_exit_bb (combined))
>        || !dominated_by_p (CDI_DOMINATORS, get_exit_bb (combined),
>
> comment also suggests that splitting the get_nearest_pdom_with_single_exit
> edge and including the new BB in the combined region would also always fix
> the dominance relation (though I don't see how it could do that and
> the comment looks wrong and by construction the check should never
> trigger).
>

I've replaced the entire condition with two asserts:
...
diff --git {a,b}/gcc/graphite-scop-detection.c
index 7615842..762a248 100644
--- a/gcc/graphite-scop-detection.c
+++ b/gcc/graphite-scop-detection.c
@@ -820,14 +820,10 @@ scop_detection::merge_sese
    /* For now we just want to bail out when exit does not post-dominate
       entry.
       TODO: We might just add a basic_block at the exit to make exit
       post-dominate entry (the entire region).  */
-  if (!dominated_by_p (CDI_POST_DOMINATORS, get_entry_bb (combined),
-                      get_exit_bb (combined))
-      || !dominated_by_p (CDI_DOMINATORS, get_exit_bb (combined),
-                         get_entry_bb (combined)))
-    {
-      DEBUG_PRINT (dp << "[scop-detection-fail] cannot merge "
-                   "seses.\n");
-      return invalid_sese;
-    }
+  gcc_assert (dominated_by_p (CDI_POST_DOMINATORS,
+                              get_entry_bb (combined),
+                              get_exit_bb (combined)));
+  gcc_assert (dominated_by_p (CDI_DOMINATORS,
+                              get_exit_bb (combined),
+                              get_entry_bb (combined)));

    /* FIXME: We should remove this piece of code once
       canonicalize_loop_closed_ssa has been removed, because that function
...
and ran graphite.exp.

I ran into an ICE for pr68693.f90, for the second assert. The exit bb 
has two incoming edges from within the sese, and one from outside. Given 
that the exit bb is empty, we could split the exit edge and redirect the 
edge from outside the sese to the new bb, which would fix the sese.

Perhaps that is what is meant in the comment, I'm not sure.

Thanks,
- Tom