rs6000 SPE breakage

[David Edelsohn]

	I'm not sure how to debug this.  The problem only occurs in the
stage1 compiler (compiled without optimization) linked with libiberty
(compiled with optimization).  I tried using gcc-3.2.1, gcc-3.2 branch,
and gcc-3.3 branch for bootstrap and all failed in the exact same way.

	When I compiled the libiberty change into an optimized GCC, the
fold-const.c testcase did not fail.  When I ran cc1 which ICEs in the
debugger, it did not fail (debugger changes address mappings).

	This appears to be an unstable bug that is address-dependent and
anything could perturb it to make it disappear.  I suspect that the hash
table compaction is breaking something.

David