public inbox for gcc-patches@gcc.gnu.org
 help / color / mirror / Atom feed
* [PATCH, 2/2][PR66642] Add empty loop exit block in transform_to_exit_first_loop_alt
@ 2015-06-25  8:15 Tom de Vries
  2015-06-30 10:11 ` Tom de Vries
                   ` (2 more replies)
  0 siblings, 3 replies; 6+ messages in thread
From: Tom de Vries @ 2015-06-25  8:15 UTC (permalink / raw)
  To: Richard Biener; +Cc: gcc-patches

[-- Attachment #1: Type: text/plain, Size: 1320 bytes --]

Hi,

I ran into a failure with parloops for reduction loop testcase 
libgomp/testsuite/libgomp.c/parloops-exit-first-loop-alt-3.c.  When we 
exercise the low iteration count loop, the test-case fails.

To understand the problem, let's first look at what happens when we use 
transform_to_exit_first_loop (the original one) instead of 
transform_to_exit_first_loop_alt (the alternative one, which is 
currently used, and causing the failure).

Before transform_to_exit_first_loop, the low iteration count loop and 
the main loop share the loop exit block. After 
transform_to_exit_first_loop, that's not the case anymore, the main loop 
now has an exit block with a single predecessor. Subsequently, 
separate_decls_in_region inserts code in the main loop exit block, which 
is only triggered upon exit of the main loop.

However, transform_to_exit_first_loop_alt does not insert such an exit 
block, and the code inserted by separate_decls_in_region is also active 
for the low iteration count loop, which results in an incorrect 
reduction result when the low iteration count loop is used.


This patch fixes the problem by making sure 
transform_to_exit_first_loop_alt adds a new exit block inbetween the 
main loop header and the old exit block.


Bootstrapped and reg-tested on x86_64.

OK for trunk?

Thanks,
- Tom

[-- Attachment #2: 0002-Add-empty-loop-exit-block-in-transform_to_exit_first.patch --]
[-- Type: text/x-patch, Size: 6359 bytes --]

Add empty loop exit block in transform_to_exit_first_loop_alt

2015-06-24  Tom de Vries  <tom@codesourcery.com>

	PR tree-optimization/66642
	* tree-parloops.c (transform_to_exit_first_loop_alt): Update function
	header comment.  Rename split_edge variable to edge_at_split.  Split
	exit edge to create new loop exit bb.  Insert loop exit phis in new loop
	exit bb.

	* testsuite/libgomp.c/parloops-exit-first-loop-alt-3.c (main): Test low
	iteration count case.
	* testsuite/libgomp.c/parloops-exit-first-loop-alt.c (init): New
	function, factor out of ...
	(main): ... here.  Test low iteration count case.
---
 gcc/tree-parloops.c                                | 45 ++++++++++++++++------
 .../libgomp.c/parloops-exit-first-loop-alt-3.c     |  5 +++
 .../libgomp.c/parloops-exit-first-loop-alt.c       | 28 +++++++++++++-
 3 files changed, 64 insertions(+), 14 deletions(-)

diff --git a/gcc/tree-parloops.c b/gcc/tree-parloops.c
index df7c351..6c8aaab 100644
--- a/gcc/tree-parloops.c
+++ b/gcc/tree-parloops.c
@@ -1522,7 +1522,7 @@ replace_uses_in_bb_by (tree name, tree val, basic_block bb)
      goto <bb header>
 
      <bb exit>:
-     sum_z = PHI <sum_b (cond[1])>
+     sum_z = PHI <sum_b (cond[1]), ...>
 
      [1] Where <bb cond> is single_pred (bb latch); In the simplest case,
 	 that's <bb header>.
@@ -1549,14 +1549,17 @@ replace_uses_in_bb_by (tree name, tree val, basic_block bb)
      if (ivtmp_c < n + 1)
        goto <bb header>;
      else
-       goto <bb exit>;
+       goto <bb newexit>;
 
      <bb latch>:
      ivtmp_b = ivtmp_a + 1;
      goto <bb newheader>
 
+     <bb newexit>:
+     sum_y = PHI <sum_c (newheader)>
+
      <bb exit>:
-     sum_z = PHI <sum_c (newheader)>
+     sum_z = PHI <sum_y (newexit), ...>
 
 
    In unified diff format:
@@ -1593,9 +1596,12 @@ replace_uses_in_bb_by (tree name, tree val, basic_block bb)
 -     goto <bb header>
 +     goto <bb newheader>
 
++    <bb newexit>:
++    sum_y = PHI <sum_c (newheader)>
+
       <bb exit>:
--     sum_z = PHI <sum_b (cond[1])>
-+     sum_z = PHI <sum_c (newheader)>
+-     sum_z = PHI <sum_b (cond[1]), ...>
++     sum_z = PHI <sum_y (newexit), ...>
 
    Note: the example does not show any virtual phis, but these are handled more
    or less as reductions.
@@ -1626,7 +1632,7 @@ transform_to_exit_first_loop_alt (struct loop *loop,
 
   /* Create the new_header block.  */
   basic_block new_header = split_block_before_cond_jump (exit->src);
-  edge split_edge = single_pred_edge (new_header);
+  edge edge_at_split = single_pred_edge (new_header);
 
   /* Redirect entry edge to new_header.  */
   edge entry = loop_preheader_edge (loop);
@@ -1643,9 +1649,9 @@ transform_to_exit_first_loop_alt (struct loop *loop,
   e = redirect_edge_and_branch (post_cond_edge, header);
   gcc_assert (e == post_cond_edge);
 
-  /* Redirect split_edge to latch.  */
-  e = redirect_edge_and_branch (split_edge, latch);
-  gcc_assert (e == split_edge);
+  /* Redirect edge_at_split to latch.  */
+  e = redirect_edge_and_branch (edge_at_split, latch);
+  gcc_assert (e == edge_at_split);
 
   /* Set the new loop bound.  */
   gimple_cond_set_rhs (cond_stmt, bound);
@@ -1697,21 +1703,36 @@ transform_to_exit_first_loop_alt (struct loop *loop,
   /* Set the latch arguments of the new phis to ivtmp/sum_b.  */
   flush_pending_stmts (post_inc_edge);
 
-  /* Register the reduction exit phis.  */
+  /* Create a new empty exit block, inbetween the new loop header and the old
+     exit block.  The function separate_decls_in_region needs this block to
+     insert code that is active on loop exit, but not any other path.  */
+  basic_block new_exit_block = split_edge (exit);
+
+  /* Insert and register the reduction exit phis.  */
   for (gphi_iterator gsi = gsi_start_phis (exit_block);
        !gsi_end_p (gsi);
        gsi_next (&gsi))
     {
       gphi *phi = gsi.phi ();
       tree res_z = PHI_RESULT (phi);
+
+      /* Now that we have a new exit block, duplicate the phi of the old exit
+	 block in the new exit block to preserve loop-closed ssa.  */
+      edge succ_new_exit_block = single_succ_edge (new_exit_block);
+      edge pred_new_exit_block = single_pred_edge (new_exit_block);
+      tree res_y = copy_ssa_name (res_z, phi);
+      gphi *nphi = create_phi_node (res_y, new_exit_block);
+      tree res_c = PHI_ARG_DEF_FROM_EDGE (phi, succ_new_exit_block);
+      add_phi_arg (nphi, res_c, pred_new_exit_block, UNKNOWN_LOCATION);
+      add_phi_arg (phi, res_y, succ_new_exit_block, UNKNOWN_LOCATION);
+
       if (virtual_operand_p (res_z))
 	continue;
 
-      tree res_c = PHI_ARG_DEF_FROM_EDGE (phi, exit);
       gimple reduc_phi = SSA_NAME_DEF_STMT (res_c);
       struct reduction_info *red = reduction_phi (reduction_list, reduc_phi);
       if (red != NULL)
-	red->keep_res = phi;
+	red->keep_res = nphi;
     }
 
   /* We're going to cancel the loop at the end of gen_parallel_loop, but until
diff --git a/libgomp/testsuite/libgomp.c/parloops-exit-first-loop-alt-3.c b/libgomp/testsuite/libgomp.c/parloops-exit-first-loop-alt-3.c
index cb5bf9c..bf06e0e 100644
--- a/libgomp/testsuite/libgomp.c/parloops-exit-first-loop-alt-3.c
+++ b/libgomp/testsuite/libgomp.c/parloops-exit-first-loop-alt-3.c
@@ -36,5 +36,10 @@ main (void)
   if (res != 11995)
     abort ();
 
+  /* Test low iteration count case.  */
+  res = f (10);
+  if (res != 25)
+    abort ();
+
   return 0;
 }
diff --git a/libgomp/testsuite/libgomp.c/parloops-exit-first-loop-alt.c b/libgomp/testsuite/libgomp.c/parloops-exit-first-loop-alt.c
index 1c32ea3..35c47d7 100644
--- a/libgomp/testsuite/libgomp.c/parloops-exit-first-loop-alt.c
+++ b/libgomp/testsuite/libgomp.c/parloops-exit-first-loop-alt.c
@@ -21,8 +21,8 @@ f (unsigned int n)
     c[i] = a[i] + b[i];
 }
 
-int
-main (void)
+static void __attribute__((noclone,noinline))
+init (void)
 {
   int i, j;
 
@@ -35,6 +35,14 @@ main (void)
 	b[k] = (k * 3) % 7;
 	c[k] = k * 2;
       }
+}
+
+int
+main (void)
+{
+  int i;
+
+  init ();
 
   f (N);
 
@@ -46,5 +54,21 @@ main (void)
 	abort ();
     }
 
+  /* Test low iteration count case.  */
+
+  init ();
+
+  f (10);
+
+  for (i = 0; i < N; i++)
+    {
+      unsigned int actual = c[i];
+      unsigned int expected = (i < 10
+			       ? i + ((i * 3) % 7)
+			       : i * 2);
+      if (actual != expected)
+	abort ();
+    }
+
   return 0;
 }
-- 
1.9.1


^ permalink raw reply	[flat|nested] 6+ messages in thread
end of thread, other threads:[~2015-07-08 12:46 UTC | newest]

Thread overview: 6+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2015-06-25  8:15 [PATCH, 2/2][PR66642] Add empty loop exit block in transform_to_exit_first_loop_alt Tom de Vries
2015-06-30 10:11 ` Tom de Vries
2015-07-06 10:13 ` [PING][PATCH, " Tom de Vries
2015-07-06 13:45   ` Richard Biener
2015-07-08 10:40 ` [PATCH, " Andreas Schwab
2015-07-08 12:46   ` Tom de Vries

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).