[committed] openmp: Fix up omp_declare_variant{s, _alt} htab handling [PR95315]

Jakub Jelinek jakub@redhat.com
Wed May 27 08:43:48 GMT 2020


Hi!

This patch fixes a GC ICE.  During debugging, I've found that during
gimplification we can actually call omp_resolve_declare_variant multiple
times and it would create a new magic declare_variant_alt FUNCTION_DECL
each time, which is undesirable, once we have such a decl, we should just
use that.  The other problem is that there was no cgraph node removal hook.
As the omp_declare_variants htab is used just early during gimplification,
we can just clear the whole htab, rather than trying to lookup and remove
a particular entry.  The other hash table is used later as well and that
one uses just DECL_UID as hash, so in that case the patch removes the elt.

Bootstrapped/regtested on x86_64-linux and i686-linux, committed to trunk.

2020-05-27  Jakub Jelinek  <jakub@redhat.com>

	PR middle-end/95315
	* omp-general.c (omp_declare_variant_remove_hook): New function.
	(omp_resolve_declare_variant): Always return base if it is already
	declare_variant_alt magic decl itself.  Register
	omp_declare_variant_remove_hook as cgraph node removal hook.

	* gcc.dg/gomp/pr95315.c: New test.

--- gcc/omp-general.c.jj	2020-05-26 09:35:12.222438412 +0200
+++ gcc/omp-general.c	2020-05-26 13:06:03.381443260 +0200
@@ -1695,6 +1695,28 @@ omp_resolve_late_declare_variant (tree a
   return varentry2->variant->decl;
 }
 
+/* Hook to adjust hash tables on cgraph_node removal.  */
+
+static void
+omp_declare_variant_remove_hook (struct cgraph_node *node, void *)
+{
+  if (!node->declare_variant_alt)
+    return;
+
+  /* Drop this hash table completely.  */
+  omp_declare_variants = NULL;
+  /* And remove node from the other hash table.  */
+  if (omp_declare_variant_alt)
+    {
+      omp_declare_variant_base_entry entry;
+      entry.base = NULL;
+      entry.node = node;
+      entry.variants = NULL;
+      omp_declare_variant_alt->remove_elt_with_hash (&entry,
+						     DECL_UID (node->decl));
+    }
+}
+
 /* Try to resolve declare variant, return the variant decl if it should
    be used instead of base, or base otherwise.  */
 
@@ -1715,6 +1737,11 @@ omp_resolve_declare_variant (tree base)
 	break;
       if (TREE_CODE (TREE_PURPOSE (TREE_VALUE (attr))) != FUNCTION_DECL)
 	continue;
+      cgraph_node *node = cgraph_node::get (base);
+      /* If this is already a magic decl created by this function,
+	 don't process it again.  */
+      if (node && node->declare_variant_alt)
+	return base;
       switch (omp_context_selector_matches (TREE_VALUE (TREE_VALUE (attr))))
 	{
 	case 0:
@@ -1823,6 +1850,12 @@ omp_resolve_declare_variant (tree base)
 	    }
 	}
 
+      static struct cgraph_node_hook_list *node_removal_hook_holder;
+      if (node_removal_hook_holder)
+	node_removal_hook_holder
+	  = symtab->add_cgraph_removal_hook (omp_declare_variant_remove_hook,
+					     NULL);
+
       if (omp_declare_variants == NULL)
 	omp_declare_variants
 	  = hash_table<omp_declare_variant_hasher>::create_ggc (64);
--- gcc/testsuite/gcc.dg/gomp/pr95315.c.jj	2020-05-26 13:42:42.990567525 +0200
+++ gcc/testsuite/gcc.dg/gomp/pr95315.c	2020-05-26 13:42:35.748672050 +0200
@@ -0,0 +1,5 @@
+/* PR middle-end/95315 */
+/* { dg-do compile } */
+/* { dg-options "-O2 -fopenmp --param ggc-min-heapsize=0" } */
+
+#include "../../c-c++-common/gomp/declare-variant-5.c"

	Jakub



More information about the Gcc-patches mailing list