Some statistics:
base + COMBRELOC
$ time (for i in `seq 100`;do noatun --help>/dev/null;done)
(; for i in `seq 100`; do; noatun --help > /dev/null; done; ) 148.64s
user 4.82s system 99% cpu 2:33.93 total
base + DF caching:
$ time (for i in `seq 100`;do noatun --help>/dev/null;done)
(; for i in `seq 100`; do; noatun --help > /dev/null; done; ) 151.15s
user 5.53s system 99% cpu 2:37.23 total
base:
$ time (for i in `seq 100`;do noatun --help>/dev/null;done)
(; for i in `seq 100`; do; noatun --help > /dev/null; done; ) 492.36s user
5.34s system 99% cpu 8:19.17 total
for set{u,g}id binaries, so that in case they are playing with set{u,g}id
and exec'ing other binaries they don't get affected by the
LD_{PRELOAD,DEBUG,LIBRARY_PATH} environment setup. We leave LD_BIND_NOW alone.
There are no binaries affected in the base system.
free() call, change the allocation policy to leave the responsibility for
allocation/freeing the pathname to _rtld_map_object(), instead of having
the caller allocate it and _rtld_map_object() free it. This simplifies the
code a lot and it is more efficient.
"Fix a bug that showed up when debugging dynamically linked programs.
References from GDB to "printf" and various other functions would
find the versions in the dynamic linker itself, rather than the
versions in the program's libc. This fix moves the GDB link map
entry for the dynamic linker to the end of the search list, where
its symbols will be found only if they are not found anywhere else."
"printf" isn't true for us, but various libc symbols are, e.g. "malloc".
Fixes PR 32074 as noted by uwe@
OK'd by christos@
declaration too.
32-bit SuperH can not pretend that _GLOBAL_OFFSET_TABLE_ is a normal
variable, because of the way PIC variable references is generated, but
as compiler arranges for _GLOBAL_OFFSET_TABLE_ address to be in r12
anyway, just use that value by declaring it as a global register
variable. Makes sh3 compile with RTLD_DEBUG.
_rtld_find_symdef use _rtld_symlook_default.
This reduces the code size and means that dlsym(RTLD_DEFAULT,...) has the
correct lookup order.
Reviewed by kleink. Thanks.
only the second part of the previous revision was needed.
put the logic for the initial icache invalidation of the PLT
back the way it was, but in a way that makes clearer what's going on,
and add some comments explaining it.
scw