diff --git a/coregrind/m_deduppoolalloc.c b/coregrind/m_deduppoolalloc.c
index 2456e1b65..bf6d1e3c8 100644
--- a/coregrind/m_deduppoolalloc.c
+++ b/coregrind/m_deduppoolalloc.c
@@ -231,6 +231,19 @@ void VG_(freezeDedupPA) (DedupPoolAlloc *ddpa,
    ddpa->ht_node_pa = NULL;
 }
 
+
+// hash function used by gawk and SDBM.
+static UInt sdbm_hash (const UChar* buf, UInt len )
+{
+  UInt h;
+  UInt i;
+
+  h = 0;
+  for (i = 0; i < len; i++)
+    h = *buf++ + (h<<6) + (h<<16) - h;
+  return h;
+}
+
 const void* VG_(allocEltDedupPA) (DedupPoolAlloc *ddpa, SizeT eltSzB,
                                   const void *elt)
 {
@@ -243,14 +256,7 @@ const void* VG_(allocEltDedupPA) (DedupPoolAlloc *ddpa, SizeT eltSzB,
 
    ddpa->nr_alloc_calls++;
 
-   // Currently using adler32 as hash function.
-   // Many references tells adler32 is bad as a hash function.
-   // And effectively, some tests on dwarf debug string shows
-   // a lot of collisions (at least for short elements).
-   // (A lot can be 10% of the elements colliding, even on
-   // small nr of elements such as 10_000).
-   ht_elt.key = VG_(adler32) (0, NULL, 0);
-   ht_elt.key = VG_(adler32) (ht_elt.key, elt, eltSzB);
+   ht_elt.key = sdbm_hash (elt, eltSzB);
 
    ht_elt.eltSzB = eltSzB;
    ht_elt.elt = elt;