Re: [Boost-build] bjam memory usage

From: David Abrahams (dave_at_[hidden])
Date: 2007-10-03 12:53:13

on Mon Sep 17 2007, Markus Schöpflin <markus.schoepflin-AT-comsoft.de> wrote:

> Rene Rivera wrote:
>
>> Dave mentioned he might look at why it crashes when using boehmgc. But 
>> as we all know he's a busy person ;-) If any one else is brave enough, 
>> they can try it out with the svn bjam version (3.1.15), just by adding 
>> --gc when building bjam.
>> 
>> NOTE; Using boehmgc does help, when I run with it, it does use less 
>> memory. It just crashes at some point :-(
>
> I have built a version of bjam linked against boehm gc 7.0. When this 
> version is run, it get tons of  complaints from the gc:
>
> GC_debug_register_finalizer called with non-base-pointer 0x8168fe0
> GC_debug_register_finalizer called with non-base-pointer 0x8161f78
> GC_debug_register_finalizer called with non-base-pointer 0x8161f78
> ...
>
> Maybe this rings a bell with someone on what might be wrong?

Hans Boehm has this to say about it:

  The message means the GC_debug_register_finalizer was passed an
  address that at was at the wrong offset from a GC allocated object.
  It should normally be at a fixed nonzero offset, since it should be
  passed an object allocated by GC_debug_malloc, which adds a header
  with debug information.

  The call stack at the call to GC_err_printf might well be
  interesting.

  Typically, you would be calling GC_MALLOC and GC_REGISTER_FINALIZER
  with GC_DEBUG defined to get here.  An inconsistent GC_DEBUG setting
  for different compilation units might do this.  IObj and base in
  GC_debug_register_finalizer would probably have to have the same
  value for this to be the case.

  This error by itself is probably not
  fatal. GC_debug_register_finalizer computes the right object address
  anyway.  The real problem may be elsewhere.

Still, it seems worth investigating.

-- 
Dave Abrahams
Boost Consulting
http://www.boost-consulting.com