Grabbed with
wget 'http://sources.redhat.com/cgi-bin/cvsweb.cgi/src/opcodes/ppc-opc.c.diff?r1=1.70&r2=1.71&cvsroot=src'
See 
http://mail.gnu.org/archive/html/bug-binutils/2004-05/msg00071.html
and
http://sources.redhat.com/ml/binutils-cvs/2004-05/msg00111.html

This fixes problems like
{standard input}: Assembler messages:
{standard input}:1: Error: Unrecognized opcode: `dssall' 
which show up in binutils-2.15 when building the Linux kernel,
or possibly failed compilations when building setjmp/longjmp in glibc.

An alternative fix would be to pass -many to the assembler.
A patch to do that for glibc is at
wget 'http://sources.redhat.com/cgi-bin/get-raw-msg?listname=libc-alpha&date=2004-05&msgid=40B36E8C.9030609%40us.ibm.com'
Yet another, better, fix would be for gcc to generate assembly
that told the assembler which processor type to use.
Presumably the Linux kernel sources would need a fix, too.
Probably better to just fix binutils to accept sources that it used to.


===================================================================
RCS file: /cvs/src/src/opcodes/ppc-opc.c,v
retrieving revision 1.70
retrieving revision 1.71
diff -u -r1.70 -r1.71
--- src/opcodes/ppc-opc.c	2004/05/05 13:43:36	1.70
+++ src/opcodes/ppc-opc.c	2004/05/19 05:11:48	1.71
@@ -1004,8 +1004,13 @@
 
   /* If only one bit of the FXM field is set, we can use the new form
      of the instruction, which is faster.  Unlike the Power4 branch hint
-     encoding, this is not backward compatible.  */
-  else if ((dialect & PPC_OPCODE_POWER4) != 0 && (value & -value) == value)
+     encoding, this is not backward compatible.  Do not generate the
+     new form unless -mpower4 has been given, or -many and the two
+     operand form of mfcr was used.  */
+  else if ((value & -value) == value
+	   && ((dialect & PPC_OPCODE_POWER4) != 0
+	       || ((dialect & PPC_OPCODE_ANY) != 0
+		   && (insn & (0x3ff << 1)) == 19 << 1)))
     insn |= 1 << 20;
 
   /* Any other value on mfcr is an error.  */