From: Paul Brannan <pbrannan@...>
Date: 2008-03-27T22:31:31+09:00
Subject: Re: some Enumerable methods slower in 1.9 on OS X after revision 15124

On Thu, Mar 27, 2008 at 02:26:51PM +0900, Nobuyoshi Nakada wrote:
> Should the priority be _setjmp > sigsetjmp > setjmp?

I think it may depend on the platform.  I ran across this the other day
while trying to decide whether I could afford the overhead of a call to
rb_protect():

http://bugs.opensolaris.org/view_bug.do;jsessionid=3681352170809a64f236f8b6ec?bug_id=4404612

Quote:

  One of the 'big wins' I discovered was to rebuild perl so that it used setjmp
  instead of the default sigsetjmp.  This gave about an 18% performance boost.
  It turns out that the perl interpreter calls sigsetjmp every time it enters a
  new block scope, and the libexacct module does this a lot, which is why the
  setjmp/sigsetjmp difference is so noticeable.  I wrote a micro benchmark to
  time how long setjmp and sigsetjmp took.

  Here are the results for sparc (450Mhz):

  Loop overhead: 22 ns/iter
  setjmp(jb): 37 ns/call
  sigsetjmp(sjb, 0): 11774 ns/call
  sigsetjmp(sjb, 1): 11891 ns/call

  and here are the results for i386 (600MHz):

  Loop overhead: 2 ns/iter
  setjmp(jb): 25 ns/call
  sigsetjmp(sjb, 0): 1963 ns/call
  sigsetjmp(sjb, 1): 1975 ns/call
  mongo$ 

  Linux (laptop 360MHz) gives:

  Loop overhead: 16 ns/iter
  setjmp(jb): 101 ns/call
  sigsetjmp(sjb, 0): 101 ns/call
  sigsetjmp(sjb, 1): 1288 ns/call

It seems there exist platforms where setjmp is preferable over
sigsetjmp with the second parameter set to 0.

Paul