Optimize poor benchmark result

[brson]

http://blog.cdleary.com/2012/06/simple-selfish-and-unscientific-shootout/

We had a particularly poor showing in this comparison.

[brson]

Thanks @marijnh for pointing it out

[brson]

Here's what perf says with an optimized runtime and corelib:

 10.64%  test  libcore-d27e4777a53c3e50-0.2.so  [.] uint::to_str::_f7f4236dcee4291a::_02
  9.34%  test  libc-2.15.so                     [.] _int_free
  7.09%  test  libc-2.15.so                     [.] __memset_x86_64
  6.90%  test  librustrt.so                     [.] upcall_s_vec_grow
  6.01%  test  librustrt.so                     [.] upcall_vec_grow
  5.07%  test  librustrt.so                     [.] exchange_malloc
  4.92%  test  librustrt.so                     [.] upcall_exchange_free
  4.72%  test  librustrt.so                     [.] upcall_exchange_malloc_dyn
  4.43%  test  librustrt.so                     [.] check_stack_canary(stk_seg*)
  4.19%  test  libc-2.15.so                     [.] malloc
  4.11%  test  libc-2.15.so                     [.] _int_malloc
  3.07%  test  librustrt.so                     [.] upcall_s_exchange_free
  3.03%  test  librustrt.so                     [.] check_stack_alignment
  3.00%  test  librustrt.so                     [.] upcall_s_exchange_malloc_dyn
  2.18%  test  librustrt.so                     [.] memory_region::malloc(unsigned long, char const*, bool)
  2.15%  test  libc-2.15.so                     [.] realloc
  1.97%  test  [kernel.kallsyms]                [k] 0xffffffff8103d0ca
  1.91%  test  librustrt.so                     [.] get_sp_limit
  1.82%  test  librustrt.so                     [.] memory_region::free(void*)
  1.53%  test  libc-2.15.so                     [.] _int_realloc
  1.33%  test  librustrt.so                     [.] __morestack
  1.31%  test  librustrt.so                     [.] get_sp
  1.19%  test  librustrt.so                     [.] upcall_call_shim_on_c_stack
  0.95%  test  libc-2.15.so                     [.] free
  0.89%  test  librustrt.so                     [.] memory_region::add_alloc()
  0.68%  test  librustrt.so                     [.] upcall_str_new_uniq

[brson]

It looks like there's a lot to be gained just in optimizing uint::to_str. It is very inefficient.

[kud1ing]

Often in such benchmarks the random number generator is the bottleneck. Can someone point out where this shows up in the analysis above?

[eholk]

I recently added the xorshift random number generator. It's not on by default, but it should be significantly faster than the default ISAAC generator.

https://github.com/mozilla/rust/commit/ad292a8c73a0cceddfa9618a4d6eea577897bae8

[pcwalton]

The biggest win is going to be to write a uint::write or something that directly writes to a file instead of allocating.

[brson]

6e0085210c54150f794d20791b2e9c1fda6049fc makes uint::to_str not allocate so much, though it still does one extra allocation when it creates the initial empty vector.

[brson]

That commit makes the time for time ./test 10000000 go from 32s to 6s.

[brson]

Graydon made another commit to improve it further. I believe we are competitive with the other languages now.

[kud1ing]

Somewhat related is #2105