egcs-19980803 (pre-1.1) powerpc-ibm-aix4.1.4.0 regressions

[David Edelsohn]

>>>>> "Kaveh R Ghazi" writes:

Kaveh> I'm sorry I was unclear in saying this, I don't know exactly
Kaveh> which snapshot it started to break.  Its certainly not the fault of
Kaveh> the most recent changes.  I previously reported this mid-July.

Kaveh> http://www.cygnus.com/ml/egcs-bugs/1998-Jul/0510.html

Kaveh> But sometime prior to that and certainly in gcc-2.8.1 and
Kaveh> egcs-1.0.3 it worked. 

	If this truly appeared in July, nothing changed in the PowerPC
configuration which directly should have caused this.  It is possible that
some changes now are tickling a common backend bug.  The only real change
in early July was converting the gen_rtx macros to gen_rtx_FOO, especially
needed for gen_rtx_MEM and the new aliasing code.  If anything, this
should have prevented code-generation bugs, not created them.

	Can you distinguish between -O3 and -funroll-all-loops as the
culprit? 

	I could believe that the LEGITIMZE_RELOAD_ADDRESS macro could be
causing problems as there were some comments from Richard Henderson
recently that it might not be complete enough to function properly.  I
never heard a resolution to that.  That change was made in early June, not
July.  You could try commenting out that macro and bootstrapping again if
the failure could have been introduced in June instead of July.  Otherwise
this would seem to be some common backend bug that PowerPC now happens to
tickle. 

David