public inbox for gcc-patches@gcc.gnu.org
 help / color / mirror / Atom feed
* [PATCH] Handle empty infinite loops in OpenACC for PR84955
@ 2018-04-06 13:49 Cesar Philippidis
  2018-04-06 15:35 ` Jakub Jelinek
  0 siblings, 1 reply; 12+ messages in thread
From: Cesar Philippidis @ 2018-04-06 13:49 UTC (permalink / raw)
  To: gcc-patches

[-- Attachment #1: Type: text/plain, Size: 1775 bytes --]

As shown in PR84955, GCC has problems empty infinite loops of the form

  #pragma acc loop tile (2, 2)
  for (...)
    for (...)
    {
      for (...)
        /* do nothing */
        ;
    }

I also observed that it generates bad code for the following loop

  #pragma acc loop
  for (...)
  {
    for (...)
      /* do nothing */
      ;
  }

There are two problems here. First, for the tiled loop, as
expand_oacc_for splits the first basic block of the loop body to add a
branch to perform the tiling control logic, it does not consider fact
that the body may not have an exit when it contains an infinite loop.
This situation can occur when region->cont is NULL. The fix here was to
add a dummy false edge to exit_bb.

The second problem involves fixing up the CFG. Unlike OpenMP which
defers a lot of the scheduling control to libgomp, OpenACC makes heavy
use of calls to internal functions. The problem here is that the
fixup_cfg pass was largely ignoring calls to internal functions when it
decides to set PROP_cleanup_cfg. When the CFG isn't cleaned up, the LTO
streamer will report that there are fewer loops than there actually are
and that causes an ICE because the empty infinite loop is never
accounted for. This patch resolves this issue by relaxing the fixup_cfg
pass to treat all functions calls, including those to internal
functions, the same. An alternative approach to resolve this issue would
have been to teach ipa_write_summaries to check if the loop structure
needs to be fixed up and call cleanup_cfg as necessary. But I wanted to
keep the OMP and OACC code paths similar, so I took the former approach.

I regression tested this patch on x86_64-linux using nvptx offloading.
Is this patch OK for trunk and GCC 7 (and probably GCC 6).

Thanks,
Cesar

[-- Attachment #2: trunk-pr84955.diff --]
[-- Type: text/x-patch, Size: 3218 bytes --]

Fix PR84955

2018-04-06  Cesar Philippidis  <cesar@codesourcery.com>

	PR middle-end/84955

	gcc/
	* cfgloop.c (flow_loops_find): Add assert.
	* omp-expand.c (expand_oacc_for): Add dummy false branch for
        tiled basic blocks without omp continue statements.
	* tree-cfg.c (execute_fixup_cfg): Handle calls to internal
        functions like regular functions.

	libgomp/
	* testsuite/libgomp.oacc-c-c++-common/pr84955.c: New test.
	* testsuite/libgomp.oacc-fortran/pr84955.f90: New test.


diff --git a/gcc/cfgloop.c b/gcc/cfgloop.c
index 8af793c6015..6e68639452c 100644
--- a/gcc/cfgloop.c
+++ b/gcc/cfgloop.c
@@ -462,6 +462,9 @@ flow_loops_find (struct loops *loops)
 	{
 	  struct loop *loop;
 
+	  if (!from_scratch)
+	    gcc_assert (header->loop_father != NULL);
+
 	  /* The current active loop tree has valid loop-fathers for
 	     header blocks.  */
 	  if (!from_scratch
diff --git a/gcc/omp-expand.c b/gcc/omp-expand.c
index bb204906ea6..1c7b68fbd8c 100644
--- a/gcc/omp-expand.c
+++ b/gcc/omp-expand.c
@@ -5439,6 +5439,13 @@ expand_oacc_for (struct omp_region *region, struct omp_for_data *fd)
 
 	  split->flags ^= EDGE_FALLTHRU | EDGE_TRUE_VALUE;
 
+	  /* Add a dummy exit for the tiled block when cont_bb is missing.  */
+	  if (cont_bb == NULL)
+	    {
+	      edge e = make_edge (body_bb, exit_bb, EDGE_FALSE_VALUE);
+	      e->probability = profile_probability::even ();
+	    }
+
 	  /* Initialize the user's loop vars.  */
 	  gsi = gsi_start_bb (elem_body_bb);
 	  expand_oacc_collapse_vars (fd, true, &gsi, counts, e_offset);
diff --git a/gcc/tree-cfg.c b/gcc/tree-cfg.c
index 9485f73f341..cb676d78128 100644
--- a/gcc/tree-cfg.c
+++ b/gcc/tree-cfg.c
@@ -9586,10 +9586,7 @@ execute_fixup_cfg (void)
       for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi);)
 	{
 	  gimple *stmt = gsi_stmt (gsi);
-	  tree decl = is_gimple_call (stmt)
-		      ? gimple_call_fndecl (stmt)
-		      : NULL;
-	  if (decl)
+	  if (is_gimple_call (stmt))
 	    {
 	      int flags = gimple_call_flags (stmt);
 	      if (flags & (ECF_CONST | ECF_PURE | ECF_LOOPING_CONST_OR_PURE))
diff --git a/libgomp/testsuite/libgomp.oacc-c-c++-common/pr84955.c b/libgomp/testsuite/libgomp.oacc-c-c++-common/pr84955.c
new file mode 100644
index 00000000000..5910b57b68d
--- /dev/null
+++ b/libgomp/testsuite/libgomp.oacc-c-c++-common/pr84955.c
@@ -0,0 +1,20 @@
+/* { dg-do compile }  */
+
+int
+main ()
+{
+  int i, j;
+
+#pragma acc parallel loop tile(2,3)
+  for (i = 1; i < 10; i++)
+    for (j = 1; j < 10; j++)
+      for (;;)
+	;
+
+#pragma acc parallel loop
+  for (i = 1; i < 10; i++)
+    for (;;)
+      ;
+
+  return i + j;
+}
diff --git a/libgomp/testsuite/libgomp.oacc-fortran/pr84955.f90 b/libgomp/testsuite/libgomp.oacc-fortran/pr84955.f90
new file mode 100644
index 00000000000..878d8a89f41
--- /dev/null
+++ b/libgomp/testsuite/libgomp.oacc-fortran/pr84955.f90
@@ -0,0 +1,20 @@
+! { dg-do compile }
+
+subroutine s
+   integer :: i, j
+   !$acc parallel loop tile(2,3)
+   do i = 1, 10
+      do j = 1, 10
+         do
+         end do
+      end do
+   end do
+  !$acc end parallel loop
+
+  !$acc parallel loop
+   do i = 1, 10
+      do
+      end do
+   end do
+  !$acc end parallel loop
+end subroutine s

^ permalink raw reply	[flat|nested] 12+ messages in thread
end of thread, other threads:[~2018-05-01 13:31 UTC | newest]

Thread overview: 12+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2018-04-06 13:49 [PATCH] Handle empty infinite loops in OpenACC for PR84955 Cesar Philippidis
2018-04-06 15:35 ` Jakub Jelinek
2018-04-09 11:32   ` Richard Biener
2018-04-11 19:31     ` Cesar Philippidis
2018-04-12  8:02       ` Richard Biener
2018-04-12 18:27       ` H.J. Lu
2018-04-12 18:39         ` Cesar Philippidis
2018-04-12 20:35           ` Jakub Jelinek
2018-04-16 18:13             ` Tom de Vries
2018-04-23 12:07               ` [PATCH, lto, PR85422] Fixup loops before lto write-out Tom de Vries
2018-04-23 12:11                 ` Richard Biener
2018-05-01 13:31               ` [PATCH] Handle empty infinite loops in OpenACC for PR84955 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).