shell exec is about 30 times slower than raw exec syscall

[X512]

This is hrev57820.

Shell (Bash/Zsh etc.) execution time is very slow compared Linux/*BSD that significantly slow down Autotools based projects compilation and other tasks that heavily use shell scripts. Raw exec syscall is significantly faster (30x). I am not sure is it kernel, libroot or shell port problem, it need more investigation.

UPDATE: problem is related to POSIX locale support. Setting LC_ALL=C significantly speed up shell execution.

loop.sh:

#!sh

if (( $1 <= 0 ))
then
    exit
fi

exec loop.sh $(($1 - 1))

loop.c:

#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>


int main(int argc, char **argv)
{
    if (argc < 2)
        return 1;

    int n = atoi(argv[1]);
    if (n <= 0)
        return 0;
    
    char buf[32];
    sprintf(buf, "%d", n - 1);
    char *newArgs[3] = {argv[0], buf, NULL};
    execvp(newArgs[0], newArgs);
    
    return 0;
}

time ./loop.sh 1000:

real    0m13.775s
user    0m11.713s
sys     0m1.875s

time ./loop 1000:

real    0m0.409s
user    0m0.144s
sys     0m0.240s

export LC_ALL=C time ./loop.sh 1000:

real    0m1.470s
user    0m0.878s
sys     0m0.546s

[waddlesplash]

For reference, here's what I get for both these benchmarks on Linux (in a VM):

time ./loop.sh 1000

real	0m0.930s
user	0m0.111s
sys	0m0.804s

time ./loop 1000

real	0m0.432s
user	0m0.047s
sys	0m0.373s

[waddlesplash]

And on WSL1 (on the host):

time ./loop.sh 1000

real    0m4.247s
user    0m0.422s
sys     0m2.875s

time ./loop 1000

real    0m1.869s
user    0m0.031s
sys     0m1.031s

[waddlesplash]

hrev57827 may help a little bit here.

[waddlesplash]

After fixing the profiler (though admittedly we may still be missing ticks...), we can now clearly see that most of the time is spent in loading the locale backend, over half of it in runtime_loader ELF symbol resolution and the rest in the locale initialization.

The only critical reason we absolutely must load the locale backend so early is because we need to initialize a lot of static data (like CTYPE and others). If we had a way to share this data between processes, we could avoid loading the locale backend until the first invocation of an actual locale-dependent method, speeding things up a lot.

[waddlesplash]

​https://review.haiku-os.org/c/haiku/+/7855 does help a bit here but not as much as I'd hoped it would.

[waddlesplash]

DT_GNU_HASH implemented in hrev57868. After rebuilding ICU this does improve time somewhat. The change to adopt the optimized SVR4 hash routine also improved things by a few percent even without rebuilding anything.

[waddlesplash]

It looks like we could defer loading of the locale until it is actually used. This would be a bit tricky but would have the biggest performance advantage, of course.

[waddlesplash]

hrev57870 should improve things a bit more here.