[RFC PATCH] Fix middle-end/6963

[David Edelsohn]

	It turns out that this patch introduces new regressions for 64-bit
PowerPC Linux and 64-bit AIX.  With -O1, GCC produces invalid assembly
language, as shown by the following diff:

--- Lisp.s      Thu Jun 27 17:59:18 2002
+++ Lisp.s.bad  Thu Jun 27 17:59:14 2002
@@ -58,11 +58,11 @@
        bl .LispNewCons
        nop
        std 3,8(30)
-       stfd 31,120(31)
-       ld 0,120(31)
+       stfd 31,8()
+       ld 0,8()
        addic. 0,0,-1
-       std 0,120(31)
-       lfd 31,120(31)
+       std 0,8()
+       lfd 31,8()
        bne+ 0,L..20
 L..16:
        mr 3,31

"8()" is an invalid address -- this should be a displacement from a
register number inside the parentheses.

	At -O2, GCC ICEs:

Lisp.c: In function `Lisp_Makearray':
Lisp.c:32: Internal compiler error in sched_analyze_2, at sched-deps.c:758

	Both failure modes look like the function.c change causing
corrupted RTL, e.g., a wild write.

	The testcase does not fail in 32-bit mode, but I doubt this is a
bug in the 64-bit support, probably just different sequence of events.

	This needs to be fixed or reverted on both the trunk and GCC 3.1
release branch.

Thanks, David