test failure: 1.3.1, swig 1.3.28, ruby 1.8, Linux / PowerPC

On Debian we've been seeing test suite failures from the ruby bindings,
Subversion 1.3.0 and 1.3.1, swig 1.3.28, on PowerPC. One user on i386
_may_ have tracked it down to a Ruby problem; see below.

Can anyone shed any light on this? I have been promised access to a
PowerPC shell in the near future, so I'll continue to investigate.

The PowerPC build daemon reports (paths trimmed with "[...]" for
readability):

    cd [...]/subversion-1.3.1/BUILD/subversion/bindings/swig/ruby; \
              /usr/bin/ruby1.8 -I [...]/subversion/bindings/swig/ruby \
                [...]/subversion/bindings/swig/ruby/test/run-test.rb \
                --verbose=normal
    Loaded suite test
    Started
    .................................................................E.....................................................
    Finished in 320.14096 seconds.
      1) Error:
    test_apply(SvnDeltaTest):
    ArgumentError: NULL pointer given
        [...]/bindings/swig/ruby/svn/util.rb:58:in `svn_txdelta_apply_wrapper'
        [...]/bindings/swig/ruby/svn/util.rb:58:in `txdelta_apply_wrapper'
        [...]/bindings/swig/ruby/svn/delta.rb:51:in `apply'
        [...]/bindings/swig/ruby/test/test_delta.rb:70:in `test_apply'
    119 tests, 617 assertions, 0 failures, 1 errors
    make[1]: *** [check-swig-rb] Error 1

The full log is 2.3 MB:

    http://buildd.debian.org/fetch.php?&pkg=subversion&ver=1.3.1-1&arch=powerpc&stamp=1144210796&file=log&as=raw

I also have a user who reproduced this on i386 (not PowerPC):

    http://bugs.debian.org/359909

But he wasn't on a purely Debian system - he had compiled his own
packages - and he indicates that recompiling ruby with
-fno-strict-aliasing made it go away. That suggests that it's actually
a Ruby problem, not a Subversion problem. However, this doesn't
explain why we get this failure on PowerPC but not on the other 9 or so
architectures that have compiled 1.3.0 or 1.3.1 on Debian.