How many syscalls on Linux can you have?

test-syscalls-musl-static.sh

$ lscpu -J | jq '.lscpu.[] | select(.field == "Model name:").data' -r
AMD Ryzen Threadripper PRO 5975WX 32-Cores
$ musl-clang -static -O3 -o test-syscalls test-syscalls.cpp
$ ldd test-syscalls
        not a dynamic executable
$ ./test-syscalls
gettid: 48735.848us, 20518777.061 cps, 48.736 ns per call
clock_gettime (VDSO): 20660.516us, 48401501.686 cps, 20.661 ns per call
nothing_indirect_call: 0.020us, 50000000000000.000 cps, 0.000 ns per call
nothing: 0.020us, 50000000000000.000 cps, 0.000 ns per call
nothing_no_optimize: 220.223us, 4540851772.976 cps, 0.220 ns per call

test-syscalls-musl.sh

$ lscpu -J | jq '.lscpu.[] | select(.field == "Model name:").data' -r
AMD Ryzen Threadripper PRO 5975WX 32-Cores
$ musl-clang -O3 -o test-syscalls test-syscalls.cpp
$ file test-syscalls
test-syscalls: ELF 64-bit LSB pie executable, x86-64, version 1 (SYSV), dynamically linked, interpreter /lib/ld-musl-x86_64.so.1, BuildID[sha1]=f0af0be6a9a0a426059fc81a786541891f711335, not stripped
$ ./test-syscalls
gettid: 45955.213us, 21760316.942 cps, 45.955 ns per call
clock_gettime (VDSO): 20291.234us, 49282364.986 cps, 20.291 ns per call
nothing_indirect_call: 0.020us, 50000000000000.000 cps, 0.000 ns per call
nothing: 0.020us, 50000000000000.000 cps, 0.000 ns per call
nothing_no_optimize: 225.633us, 4431975819.140 cps, 0.226 ns per call

test-syscalls-static.sh

$ lscpu -J | jq '.lscpu.[] | select(.field == "Model name:").data' -r
AMD Ryzen Threadripper PRO 5975WX 32-Cores
$ clang -static -O3 -o test-syscalls test-syscalls.cpp
$ ldd test-syscalls
        not a dynamic executable
$ ./test-syscalls
gettid: 44307.483us, 22569551.062 cps, 44.307 ns per call
clock_gettime (VDSO): 19663.977us, 50854412.614 cps, 19.664 ns per call
nothing_indirect_call: 0.020us, 50000000000000.000 cps, 0.000 ns per call
nothing: 0.020us, 50000000000000.000 cps, 0.000 ns per call
nothing_no_optimize: 220.233us, 4540645588.990 cps, 0.220 ns per call

test-syscalls.cpp

#include <stdint.h>
#include <stdio.h>
#include <time.h>
#include <syscall.h>
#include <unistd.h>

uint64_t now_ns()
{
    struct timespec ts;
    clock_gettime(CLOCK_MONOTONIC_RAW, &ts);
    return ts.tv_sec * (uint64_t)1e9 + ts.tv_nsec;
}

uint64_t getThreadId()
{
    return static_cast<uint64_t>(syscall(SYS_gettid));
}
void getClockGetTime()
{
    struct timespec ts;
    clock_gettime(CLOCK_MONOTONIC_RAW, &ts);
}
void nothing_indirect()
{
}
struct Nothing
{
    void operator()() const {}
};
struct NothingNoOptimize
{
    void operator()() const
    {
        /// Idea from old version of folly
        int r;
        asm volatile ("" : "+r" (r));
    }
};

template <class F>
uint64_t run_bench(const char * name, const F & f, size_t iterations = 1'000'000)
{
    uint64_t start = now_ns();
    for (size_t i = 0; i < iterations; ++i)
        f();
    uint64_t end = now_ns();
    uint64_t elapsed = end - start;
    printf("%s: %.3fus, %.3f cps, %.3f ns per call\n",
        name,
        elapsed/1e3,
        iterations/(elapsed/1e9),
        (double)elapsed/iterations);
    return elapsed;
}

int main()
{
    run_bench("gettid", getThreadId);
    /// clock_gettime() in VDSO (so it is 2x faster then regular syscall)
    run_bench("clock_gettime (VDSO)", getClockGetTime);
    run_bench("nothing_indirect_call", nothing_indirect);
    run_bench("nothing", Nothing());
    run_bench("nothing_no_optimize", NothingNoOptimize());
    return 0;
}

test-syscalls.sh

$ lscpu -J | jq '.lscpu.[] | select(.field == "Model name:").data' -r
AMD Ryzen Threadripper PRO 5975WX 32-Cores
$ clang++ test-syscalls.cpp -o test-syscalls -O3 -g3
$ ./test-syscalls
gettid: 45611.8us, 2.19241e+07cps, 45.6118ns per call
clock_gettime: 21192.4us, 4.71867e+07cps, 21.1924ns per call
nothing_indirect_call: 1165.02us, 8.58358e+08cps, 1.16501ns per call
nothing: 0.02us, 5e+13cps, 2e-05ns per call
nothing_no_optimize: 233.818us, 4.27683e+09cps, 0.233818ns per call

https://godbolt.org/z/M47hjq4v3