oprofile fix for hashing backtraces.
Oprofile's hash function uses a 64-bit value for the
key where the high 32-bit are the 'from' address in a backtrace,
and the low 32-bits are the current PC. However, the hash
function was:
uint32_t temp = (value >> 32) ^ value;
return ((temp << 0) ^ (temp >> 8)) & data->hash_mask;
If 'from' and 'to' are the same (recursive function), this hashes
to 0, and you end up with lots of collisions, turning the
theoretically O(1) operation into O(n).
To fix it, I just drop the high 32-bits from the hash:
uint32_t temp = value & 0xffffffff;
return ((temp << 0) ^ (temp >> 8)) & data->hash_mask;
In testing, this drastically reduces the oprofile overhead
for some tracing.
Change-Id: I8ae65a8a73771c89b576c895f135efd7b730eaf5
diff --git a/.gitignore b/.gitignore
new file mode 100644
index 0000000..86b18f5
--- /dev/null
+++ b/.gitignore
@@ -0,0 +1,2 @@
+.cproject
+.project
diff --git a/libdb/odb.h b/libdb/odb.h
index 9ad1da2..1e777da 100644
--- a/libdb/odb.h
+++ b/libdb/odb.h
@@ -228,7 +228,7 @@
* files avoiding to rebuilding them at profiling re-start so
* on changing do_hash() change the file format!
*/
- uint32_t temp = (value >> 32) ^ value;
+ uint32_t temp = value & 0xffffffff;
return ((temp << 0) ^ (temp >> 8)) & data->hash_mask;
}