Bootstrap comparison failures on i586 (mainline)

[Marcus G. Daniels]

NB> Am I the only one seeing these?  They've persisted for quite a
NB> while now:-

NB> Bootstrap comparison failure!
NB> cxxmain.o differs
NB> fold-const.o differs
NB> real.o differs
NB> tree.o differs
NB> f/com.o differs

GS> I believe this is caused this change by Bernds 

GS>      * cselib.c (hash_rtx): Don't do tail recursion elimination by hand.

BS> I wasn't able to reproduce the problem. 

I tried reverting it too, and it didn't help.