applet_tables: save space by removing applet name offsets
The array applet_nameofs consumes two bytes per applet. It encodes
nofork/noexec flags
suid flags
the offset of the applet name in the applet_name string
Change the applet_table build tool to store the flags in two separate
arrays (applet_flags and applet_suid). Replace applet_nameofs[] with a
smaller version that only stores a limited number of offsets.
This requires changes to the macros APPLET_IS_NOFORK, APPLET_IS_NOEXEC
and APPLET_SUID.
According to Valgrind the original find_applet_by_name required
353 cycles per call, averaged over all names. Adjusting the number
of known offsets allows space to be traded off against execution time:
KNOWN_OFFSETS cycles bytes (wrt KNOWN_OFFSETS = 0)
0 9057 -
2 4604 32
4 2407 75
8 1342 98
16 908 130
32 884 194
This patch uses KNOWN_OFFSETS = 8.
v2:
Remove some dead code from the applet_table tool;
Treat the applet in the middle of the table as a special case.
v3:
Use the middle applet to adjust the start of the linear search as
well as the last applet found.
v4:
Use an augmented linear search in find_applet_by_name.
Drop the special treatment of the middle name from get_applet_name:
most of the advantage now derives from the last stored value.
v5:
Don't store index in applet_nameofs, it can be calculated.
v6:
Tweaks by Denys
function old new delta
find_applet_by_name 25 125 +100
applet_suid - 92 +92
run_applet_no_and_exit 452 460 +8
run_applet_and_exit 695 697 +2
applet_name_compare 31 - -31
applet_nameofs 734 14 -720
------------------------------------------------------------------------------
(add/remove: 1/1 grow/shrink: 3/1 up/down: 202/-751) Total: -549 bytes
text data bss dec hex filename
925464 906 17160 943530 e65aa busybox_old
924915 906 17160 942981 e6385 busybox_unstripped
Signed-off-by: Ron Yorston <rmy@pobox.com>
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
4 files changed