------- Comment #15 from zadeck at naturalbridge dot com 2008-01-20 23:15 ------- There appears to be an design inconsistency in the way that we have specified the various dataflow problems with respect to the eq notes.
I hate eq notes. In the rd patch that just went in where we trim the solution based on the lr info, it seems to also trim the bits associated with the eq notes because these are not carried around in the lr problem. This screws ups the use-def chains that are used in fwprop, see and web (in this particular case, the problem is in web, but that is only a matter of bad luck). I can either blow away the rd patch for 34400, or we can teach lr to respect the DF_EQ_NOTES patch or just have lr always propagate the eq notes as if they were real uses. The latter actually sounds like the only correct solution but at this late point of stage 3, we may need to sacrifice correctness for stability. Kenny -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34884