Jason, would you like to take a look at this patch before I commit it?
If I turn on -fdebug-types-section, I get a random failure in
g++.dg/pch/system-2.C. This failure is being caused by
non-deterministic order of the type stubs emitted by
optimize_external_refs, because it's using the address of the DIE as a
hash for the extern_map hash table. This patch makes the output order
deterministic by using either the die_symbol (as a string) or the type
signature for the hash value.
Bootstrapped on x86_64 with no regressions.
(I'm not planning to turn on -fdebug-types-section by default; I'm
just trying to make sure the test suite is clean with it on so that I
can turn it on in the google branches.)
-cary
2013-04-24 Cary Coutant <ccout...@google.com>
gcc/
* dwarf2out.c (hash_external_ref): Use die_symbol or signature
for hash so that hash table traversal order is deterministic.
Index: gcc/dwarf2out.c
===================================================================
--- gcc/dwarf2out.c (revision 198260)
+++ gcc/dwarf2out.c (working copy)
@@ -7385,7 +7385,22 @@ static hashval_t
hash_external_ref (const void *p)
{
const struct external_ref *r = (const struct external_ref *)p;
- return htab_hash_pointer (r->type);
+ dw_die_ref die = r->type;
+ hashval_t h = 0;
+
+ /* We can't use the address of the DIE for hashing, because
+ that will make the order of the stub DIEs non-deterministic. */
+ if (! die->comdat_type_p)
+ /* We have a symbol; use it to compute a hash. */
+ h = htab_hash_string (die->die_id.die_symbol);
+ else
+ {
+ /* We have a type signature; use a subset of the bits as the hash.
+ The 8-byte signature is at least as large as hashval_t. */
+ comdat_type_node_ref type_node = die->die_id.die_type_node;
+ memcpy (&h, type_node->signature, sizeof (h));
+ }
+ return h;
}
/* Compare external_refs. */
