Re: [gomp4] Avoiding predication for certain blocks

[Tom de Vries <Tom_deVries at mentor dot com>]

On 29/05/15 18:23, Bernd Schmidt wrote:
> When predicating the code for OpenACC, we should avoid the entry block
> in an offloaded region, which contains setup code that should be run in
> every thread. The following patch adds a new marker statement that is
> used to identify this block. Currently, predication doesn't happen
> anyway due to an oversight in the algorithm, but I'll be fixing that in
> a followup patch.
>
> Committed on gomp-4_0-branch.

Hi Bernd,

this causes an ICE in kernels-eternal.c.

The GIMPLE_OMP_ENTRY_END is removed the first time (!ssa) we call 
expand_omp_target. The second time we call expand_omp_target (ssa), it's 
no longer there, and we call 'gimple_code (NULL)' here:
...
  basic_block entry_succ_bb = single_succ (entry_bb);
  gsi = gsi_last_bb (entry_succ_bb);
  if (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_ENTRY_END)
    gsi_remove (&gsi, true);
...

This patch fixes that by not attempting to remove GIMPLE_OMP_ENTRY_END 
when in ssa.

Committed to gomp-4_0-branch.

Thanks,
- Tom

Fix GIMPLE_OMP_ENTRY_END handling in expand_omp_target for ssa

2015-06-01  Tom de Vries  <tom@codesourcery.com>

	* omp-low.c (expand_omp_target): Fix GIMPLE_OMP_ENTRY_END handling for
	ssa.

---
 gcc/omp-low.c | 9 ++++++---
 1 file changed, 6 insertions(+), 3 deletions(-)

diff --git a/gcc/omp-low.c b/gcc/omp-low.c
index a74a2bea..2634051 100644
--- a/gcc/omp-low.c
+++ b/gcc/omp-low.c
@@ -9486,9 +9486,12 @@ expand_omp_target (struct omp_region *region)
     }
 
   basic_block entry_succ_bb = single_succ (entry_bb);
-  gsi = gsi_last_bb (entry_succ_bb);
-  if (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_ENTRY_END)
-    gsi_remove (&gsi, true);
+  if (!gimple_in_ssa_p (cfun))
+    {
+      gsi = gsi_last_bb (entry_succ_bb);
+      if (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_ENTRY_END)
+	gsi_remove (&gsi, true);
+    }
 
   if (offloaded
       && do_splitoff)
-- 
1.9.1