Julia Benchmark comparing the performance of `jl_eval_string` over `jl_get_global` and evaluating every frame. 648x (!) speed increase

julia-speed-increase.cpp

#include "/usr/include/julia/julia.h"
#include <chrono>
#include <cstdio>

int main() {
  jl_init();
  std::chrono::high_resolution_clock c;
  jl_eval_string("module Some module Nested module Module function f(a::Int64) "
                 "end end end end");
  auto t1 = c.now();
  for (int i = 0; i < 100000; ++i)
    jl_call1(jl_eval_string("Some.Nested.Module.f"), jl_box_int64(5));

  auto t2 = c.now();

  jl_module_t *some, *nested, *module;
  jl_function_t *f;
  JL_GC_PUSH4(some, nested, module, f);
  some = reinterpret_cast<jl_module_t *>(
      jl_get_global(jl_main_module, jl_symbol("Some")));
  nested =
      reinterpret_cast<jl_module_t *>(jl_get_global(some, jl_symbol("Nested")));
  module = reinterpret_cast<jl_module_t *>(
      jl_get_global(nested, jl_symbol("Module")));
  f = jl_get_function(module, "f");
  for (int i = 0; i < 100000; ++i)
    jl_call1(f, jl_box_int64(5));
  JL_GC_POP();

  auto t3 = c.now();

  long d1 = std::chrono::duration_cast<std::chrono::microseconds>(t2 - t1)
                .count(),
       d2 = std::chrono::duration_cast<std::chrono::microseconds>(t3 - t2)
                .count();
  printf("%ldms, %ldms, %Lfx speed increase\n", d1, d2, (long double)d1 / d2);
  jl_atexit_hook(0);
}

Compile with g++ -g -o eval -I/usr/include/julia -DJULIA_ENABLE_THREADING=1 -fPIC -L'/usr/lib' -Wl,--export-dynamic -Wl,-rpath,'/usr/lib' -Wl,-rpath,'/usr/lib/julia' juliaeval.cpp -ljulia