public inbox for glibc-cvs@sourceware.org
help / color / mirror / Atom feed
* [glibc/fw/dl-bind-performance] elf: Optimize symbol binding by pre-computing divisions
@ 2022-04-11 14:07 Florian Weimer
0 siblings, 0 replies; only message in thread
From: Florian Weimer @ 2022-04-11 14:07 UTC (permalink / raw)
To: glibc-cvs
https://sourceware.org/git/gitweb.cgi?p=glibc.git;h=0fa7402cbb4c88ab230eb4447f3fc2a182b17ae7
commit 0fa7402cbb4c88ab230eb4447f3fc2a182b17ae7
Author: Florian Weimer <fweimer@redhat.com>
Date: Mon Nov 4 18:08:50 2019 +0100
elf: Optimize symbol binding by pre-computing divisions
The division for the hash table lookup shows up in profiles. We can
use a standard compiler optimization technique to speed up the hash
table lookup. The speedup is most pronounced when a symbol lookup
succeeds early in the scope array, otherwise the bitmap check
dominates the profiles.
Change-Id: I898b7c711979447d4756b3f7b567c49a8d33187b
Diff:
---
elf/dl-lookup.c | 20 ++++++++++++++++++--
include/link.h | 2 ++
2 files changed, 20 insertions(+), 2 deletions(-)
diff --git a/elf/dl-lookup.c b/elf/dl-lookup.c
index fd44cd4101..205d043717 100644
--- a/elf/dl-lookup.c
+++ b/elf/dl-lookup.c
@@ -28,6 +28,7 @@
#include <libc-lock.h>
#include <tls.h>
#include <atomic.h>
+#include <divopt.h>
#include <assert.h>
@@ -394,8 +395,18 @@ do_lookup_x (const char *undef_name, uint_fast32_t new_hash,
if (__glibc_unlikely ((bitmask_word >> hashbit1)
& (bitmask_word >> hashbit2) & 1))
{
- Elf32_Word bucket = map->l_gnu_buckets[new_hash
- % map->l_nbuckets];
+ Elf32_Word bucket;
+ if (map->l_nbuckets > 1)
+ {
+ uint32_t quotient
+ = divopt_32 (new_hash, map->l_nbuckets_multiplier,
+ map->l_nbuckets_multiplier_shift);
+ uint32_t remainder = new_hash - map->l_nbuckets * quotient;
+ bucket = map->l_gnu_buckets[remainder];
+ }
+ else
+ bucket = map->l_gnu_buckets[0];
+
if (bucket != 0)
{
const Elf32_Word *hasharr = &map->l_gnu_chain_zero[bucket];
@@ -931,6 +942,11 @@ _dl_setup_hash (struct link_map *map)
/* Initialize MIPS xhash translation table. */
ELF_MACHINE_XHASH_SETUP (hash32, symbias, map);
+ if (map->l_nbuckets >= 2)
+ map->l_nbuckets_multiplier_shift
+ = precompute_divopt_32 (map->l_nbuckets,
+ &map->l_nbuckets_multiplier);
+
return;
}
diff --git a/include/link.h b/include/link.h
index 1184201f91..b09aa81bb4 100644
--- a/include/link.h
+++ b/include/link.h
@@ -153,6 +153,8 @@ struct link_map
/* Symbol hash table. */
Elf_Symndx l_nbuckets;
+ uint32_t l_nbuckets_multiplier;
+ int l_nbuckets_multiplier_shift;
Elf32_Word l_gnu_bitmask_idxbits;
Elf32_Word l_gnu_shift;
const ElfW(Addr) *l_gnu_bitmask;
^ permalink raw reply [flat|nested] only message in thread
only message in thread, other threads:[~2022-04-11 14:07 UTC | newest]
Thread overview: (only message) (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2022-04-11 14:07 [glibc/fw/dl-bind-performance] elf: Optimize symbol binding by pre-computing divisions Florian Weimer
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).