Since about 3 weeks bootstrap is failing due to comparison failure on java/decl.o (no other object file is affected). It first started with r115810, then went away with r115850, but reoccurred with r115851. Since then it reoccurred often but irregularily. Bootstrap is currently failing for three days now. When running the compilation command for java/decl.o manually the resulting object file is identical to the stage2 object file, which means that there is an environmental dependency in the generated code. Bootstrapping with --enable-checking=all didn't show any error, the comparison failure remained.
Created attachment 12103 [details] Diff of "objdump -d" output
Created attachment 12104 [details] Diff of compiler dumps All dumps before mach are identical.
Created attachment 12105 [details] stage3 mach dump
The bug is very sensitive on the environment size. Adding "-da" makes it go away, adding "-da -da" makes it reappear. Looks like some decision is made on the low bits of an address.
I really don't think this is a regression, this is most likely related to PR 20586 or PR 28752.
I believe I have a patch. I'm just waiting for the fix for PR28672 (which I've just approved) to be applied, so I can complete bootstrap and regression test to confirm there are no unexpected side-effects.
Subject: Bug 28784 Author: sayle Date: Tue Sep 12 16:02:31 2006 New Revision: 116891 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=116891 Log: PR middle-end/4520 PR bootstrap/28784 * cselib.c (cselib_hash_rtx): Avoid hashing on the address of labels and symbols. Instead use the implementation from cse.c's hash_rtx. Modified: trunk/gcc/ChangeLog trunk/gcc/cselib.c
Fixed.