openmp: Make finalize_task_copyfn order reproduceable [PR104517]

The following testcase fails -fcompare-debug, because finalize_task_copyfn
was invoked from splay tree destruction, whose order can in some cases
depend on -g/-g0.  The fix is to queue the task stmts that need copyfn
in a vector and run finalize_task_copyfn on elements of that vector.

2022-02-15  Jakub Jelinek  <jakub@redhat.com>

	PR debug/104517
	* omp-low.cc (task_cpyfns): New variable.
	(delete_omp_context): Don't call finalize_task_copyfn from here.
	(create_task_copyfn): Push task_stmt into task_cpyfns.
	(execute_lower_omp): Call finalize_task_copyfn here on entries from
	task_cpyfns vector and release the vector.

	* gcc.dg/gomp/pr104517.c: New test.

2 files changed, 60 insertions, 3 deletions

diff --git a/gcc/omp-low.cc b/gcc/omp-low.cc
index 065208464b2..77176efe715 100644
--- a/gcc/omp-low.cc
+++ b/gcc/omp-low.cc
@@ -191,6 +191,7 @@ static int target_nesting_level;
 static bitmap task_shared_vars;
 static bitmap global_nonaddressable_vars;
 static vec<omp_context *> taskreg_contexts;
+static vec<gomp_task *> task_cpyfns;
 
 static void scan_omp (gimple_seq *, omp_context *);
 static tree scan_omp_1_op (tree *, int *, void *);
@@ -1082,9 +1083,6 @@ delete_omp_context (splay_tree_value value)
 	DECL_ABSTRACT_ORIGIN (t) = NULL;
     }
 
-  if (is_task_ctx (ctx))
-    finalize_task_copyfn (as_a <gomp_task *> (ctx->stmt));
-
   if (ctx->task_reduction_map)
     {
       ctx->task_reductions.release ();
@@ -11951,6 +11949,7 @@ create_task_copyfn (gomp_task *task_stmt, omp_context *ctx)
   size_t looptempno = 0;
 
   child_fn = gimple_omp_task_copy_fn (task_stmt);
+  task_cpyfns.safe_push (task_stmt);
   child_cfun = DECL_STRUCT_FUNCTION (child_fn);
   gcc_assert (child_cfun->cfg == NULL);
   DECL_SAVED_TREE (child_fn) = alloc_stmt_list ();
@@ -14475,6 +14474,10 @@ execute_lower_omp (void)
       && (TREE_CODE (TREE_TYPE (DECL_ARGUMENTS (current_function_decl)))
 	  == POINTER_TYPE))
     remove_member_access_dummy_vars (DECL_INITIAL (current_function_decl));
+
+  for (auto task_stmt : task_cpyfns)
+    finalize_task_copyfn (task_stmt);
+  task_cpyfns.release ();
   return 0;
 }
 
diff --git a/gcc/testsuite/gcc.dg/gomp/pr104517.c b/gcc/testsuite/gcc.dg/gomp/pr104517.c
new file mode 100644
index 00000000000..efb3175beb3
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/gomp/pr104517.c
@@ -0,0 +1,54 @@
+/* PR debug/104517 */
+/* { dg-do compile } */
+/* { dg-options "-O1 -fcompare-debug -fopenmp -fno-tree-ter -save-temps" } */
+
+enum {
+  omp_default_mem_alloc,
+  omp_large_cap_mem_alloc,
+  omp_const_mem_alloc,
+  omp_high_bw_mem_alloc
+} omp_allocator_handle_t;
+
+int t, bar_nte, bar_tl, bar_i3, bar_dd;
+
+#pragma omp threadprivate(t)
+#pragma omp declare target
+int f, l, ll, r, r2;
+#pragma omp end declare target
+
+void
+bar (int *idp, int s, int nth, int g, int nta, int fi, int pp, int *q,
+     int ntm)
+{
+  int p = 0, i2 = 0, i1 = 0, m = 0, d = 0;
+
+#pragma omp target parallel for                               \
+  device(p) firstprivate (f) allocate (f) in_reduction(+:r2)
+  for (int i = 0; i < 4; i++)
+    ll++;
+
+#pragma omp target parallel for                                         \
+  device(d) map (m)                                                     \
+  if (target: p) firstprivate (f) defaultmap(tofrom: scalar) is_device_ptr (idp) \
+  if (parallel: i2) reduction(+:r) num_threads (nth) linear (ll)        \
+  schedule(static) collapse(1) nowait depend(inout: d) allocate (f)     \
+  in_reduction(+:r2)
+  for (int i = 0; i < 4; i++)
+    ll++;
+
+#pragma omp taskloop simd firstprivate(f) lastprivate(s) grainsize(g) \
+  collapse(1) untied if (i1) final(fi) mergeable nogroup              \
+  priority(pp) linear(ll) aligned(q) allocate(f)
+  for (int i = 0; i < 4; i++)
+    ll++;
+
+#pragma omp taskloop simd firstprivate(f) lastprivate(s) num_tasks(nta) \
+  collapse(1) if (i1) final(fi) priority(pp) safelen(8) simdlen(4)      \
+  linear(ll) aligned(q) nontemporal(ntm) order(concurrent) allocate(f)
+  for (int i = 0; i < 4; i++)
+    ll++;
+
+#pragma omp parallel master firstprivate(f) shared(nth) proc_bind(spread) \
+  copyin(t) allocate(f)
+  ;
+}