[Bug lto/40758] [LTO] ICE in partition_view_bitmap, at tree-ssa-live.c:331

[rguenth at gcc dot gnu dot org]

------- Comment #2 from rguenth at gcc dot gnu dot org  2009-07-15 10:54 -------
For me

./xgcc -B.  -shared  ice.o  -flto

takes ages because iterative_hash_gimple_type seems to be exponential in time!?
Called from gimple_register_type on

 <pointer_type 0xb7becb60
    type <function_type 0xb7becbd0
        type <void_type 0xb7cb69a0 VOID
            align 8 symtab 0 alias set -1 canonical type 0xb7cb69a0
            pointer_to_this <pointer_type 0xb7cb6a10>>
        QI
        size <integer_cst 0xb7ca9310 constant 8>
        unit size <integer_cst 0xb7ca932c constant 1>
        align 8 symtab 0 alias set -1 canonical type 0xb7becc40
        arg-types <tree_list 0xb7be99a0 value <pointer_type 0xb7bec000>
            chain <tree_list 0xb7be99bc value <integer_type 0xb7bd50e0
disp_color_t>
                chain <tree_list 0xb7be99d8 value <integer_type 0xb7bd50e0
disp_color_t>
                    chain <tree_list 0xb7be99f4 value <integer_type 0xb7cb6310>
                        chain <tree_list 0xb7be9a10 value <integer_type
0xb7cb6310> chain <tree_list 0xb7be9a2c>>>>>>
        pointer_to_this <pointer_type 0xb7becb60>>
    unsigned SI
    size <integer_cst 0xb7ca9498 type <integer_type 0xb7cb6070 bit_size_type>
constant 32>
    unit size <integer_cst 0xb7ca9284 type <integer_type 0xb7cb6000> constant
4>
    align 32 symtab 0 alias set -1 canonical type 0xb7beccb0>

I guess we should try hashing without duplicates or store the hash somewhere...

I cannot reproduce the ICE btw.  A patch to fix the slowness would be

Index: gimple.c
===================================================================
--- gimple.c    (revision 149668)
+++ gimple.c    (working copy)
@@ -3537,13 +3537,17 @@ gimple_types_compatible_p (tree t1, tree
   return same_p;
 }

-
 /* Returning a hash value for gimple type TYPE combined with VAL.  */

 static hashval_t
 iterative_hash_gimple_type (const_tree type, hashval_t val, unsigned level)
 {
   hashval_t v;
+  void **slot;
+
+  if ((slot = pointer_map_contains (type_hash_cache, type)) != NULL)
+    return iterative_hash_hashval_t (TREE_ADDRESSABLE (type),
+                                    (hashval_t) (size_t) *slot);

   /* Combine a few common features of types so that types are grouped into
      smaller sets; when searching for existing matching types to merge,
@@ -3604,6 +3608,8 @@ iterative_hash_gimple_type (const_tree t
       v = iterative_hash_hashval_t (nf, v);
     }

+  *pointer_map_insert (type_hash_cache, type) = (void *) (size_t) v;
+
   return v;
 }

@@ -3620,28 +3626,13 @@ static hashval_t
 gimple_type_hash (const void *p)
 {
   const_tree type;
-  void **slot;
-  hashval_t v;

   type = (const_tree) p;

   if (type_hash_cache == NULL)
-    {
-      type_hash_cache = pointer_map_create ();
-      slot = NULL;
-    }
-  else
-    slot = pointer_map_contains (type_hash_cache, type);
+    type_hash_cache = pointer_map_create ();

-  if (slot)
-    v = (hashval_t) (size_t) *slot;
-  else
-    {
-      v = iterative_hash_gimple_type (type, 0, 0);
-      *pointer_map_insert (type_hash_cache, type) = (void *) (size_t) v;
-    }
-
-  return v;
+  return iterative_hash_gimple_type (type, 0, 0);
 }


I am going to bootstrap and test a variant of that.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40758