[PATCH] Add more primes to hash_size_primes.

Alan Modra amodra@gmail.com
Fri Aug 14 08:24:46 GMT 2020


On Thu, Aug 13, 2020 at 02:26:46PM +0200, Martin Liška wrote:
> 	* bfd/hash.c (bfd_hash_set_default_size): Add more primes
> 	as biggest one is only 64K.

No, let's not do this.  There is already a larger table of primes in
bfd/hash.c.  Also, allowing an initial allocation of nearly 8G of
memory on a 64-bit system just for the hash table of pointers is
slightly crazy, and it's completely crazy to try to allocate 4G on a
32-bit system.

How does this look instead?

	* hash.c (bfd_hash_set_default_size): Use higher_prime_number
	rather than another copy of primes.  Increase maximum default
	size allowed.

diff --git a/bfd/hash.c b/bfd/hash.c
index 56d18ac317..ba6ff92517 100644
--- a/bfd/hash.c
+++ b/bfd/hash.c
@@ -664,19 +664,18 @@ bfd_hash_traverse (struct bfd_hash_table *table,
 unsigned long
 bfd_hash_set_default_size (unsigned long hash_size)
 {
-  /* Extend this prime list if you want more granularity of hash table size.  */
-  static const unsigned long hash_size_primes[] =
-    {
-      31, 61, 127, 251, 509, 1021, 2039, 4091, 8191, 16381, 32749, 65537
-    };
-  unsigned int _index;
-
-  /* Work out best prime number near the hash_size.  */
-  for (_index = 0; _index < ARRAY_SIZE (hash_size_primes) - 1; ++_index)
-    if (hash_size <= hash_size_primes[_index])
-      break;
-
-  bfd_default_hash_table_size = hash_size_primes[_index];
+  /* These silly_size values result in around 1G and 32M of memory
+     being allocated for the table of pointers.  Note that the number
+     of elements allocated will be almost twice the size of any power
+     of two chosen here.  */
+  unsigned long silly_size = sizeof (size_t) > 4 ? 0x4000000 : 0x400000;
+  if (hash_size > silly_size)
+    hash_size = silly_size;
+  else if (hash_size != 0)
+    hash_size--;
+  hash_size = higher_prime_number (hash_size);
+  BFD_ASSERT (hash_size != 0);
+  bfd_default_hash_table_size = hash_size;
   return bfd_default_hash_table_size;
 }
 


-- 
Alan Modra
Australia Development Lab, IBM


More information about the Binutils mailing list