Re: [boost] [regex, xpressive] interesting(?) perf benchmark

Subject: Re: [boost] [regex, xpressive] interesting(?) perf benchmark
From: John Bytheway (jbytheway+boost_at_[hidden])
Date: 2010-06-08 16:56:16

On 08/06/10 16:16, Andy Venikov wrote:
> John Bytheway wrote:
> <snip>
>> The down side with this approach, of course, is the absolutely massive
>> compile-time cost.  Compiling the above example with g++ 4.4.3 takes
>> over 6 minutes and over 1.3GiB of memory (and this is after I've put in
>> effort to optimize it).  The time I could live with, but the memory
>> requirements are absurd.  Any lexer of real-world complexity would be
>> quite impossible to compile.
> 
> Could you try it with gcc 4.3?
> 4.4 has a known regression when it comes to compiling meta-c++ code.

Here are some numbers for compiling one of my tests which contains the
example I quoted and a couple of other relatively trivial examples.
These numbers are reported by /usr/bin/time.  The memory usage numbers
are not at all accurate, but I think they're proportional to the truth.

g++ 4.4.3:
  Elapsed (wall clock) time (h:mm:ss or m:ss): 7:01.17
  Maximum resident set size (kbytes): 5298256

(as I said before, this really needs about 1.3GiB, not the 5GB reported)

g++ 4.3.3:
  Elapsed (wall clock) time (h:mm:ss or m:ss): 9:38.92
  Maximum resident set size (kbytes): 5390064

So 4.3 is worse on both counts.

clang++ (svn HEAD):
  Elapsed (wall clock) time (h:mm:ss or m:ss): 0:15.59
  Maximum resident set size (kbytes): 5806528

So clang is *much* faster than either g++, but uses even more memory
(this number translates as a little over 1.5GiB in real terms).

John Bytheway