MPILIB: avoid assembler warning

A wrapper around the umull assembly instruction might reuse the input register as an output, which is undefined on some ARM machines, as pointed out by this assembler warning:

CC lib/mpi/generic_mpih-mul1.o
/tmp/ccxJuxIy.s: Assembler messages:
/tmp/ccxJuxIy.s:53: rdhi, rdlo and rm must all be different
CC lib/mpi/generic_mpih-mul2.o
/tmp/ccI0scAD.s: Assembler messages:
/tmp/ccI0scAD.s:53: rdhi, rdlo and rm must all be different
CC lib/mpi/generic_mpih-mul3.o
/tmp/ccMvVQcp.s: Assembler messages:
/tmp/ccMvVQcp.s:53: rdhi, rdlo and rm must all be different

This changes the constraints to force different registers to be used as output.

---
Not sure who would be the best person to pick this patch up.

index aac511417ad1..c4dbdb506174 100644
@@ -216,7 +216,7 @@ extern UDItype __udiv_qrnnd(UDItype *, UDItype, UDItype, UDItype);
 	__asm__ ("%@ Inlined umul_ppmm\n" \
 		"umull %r1, %r0, %r2, %r3" \
 	: "=&r" ((USItype)(xh)), \
-			"=r" ((USItype)(xl)) \
+			"=&r" ((USItype)(xl)) \
 	: "r" ((USItype)(a)), \
 			"r" ((USItype)(b)) \
 	: "r0", "r1")

I can pick it up, but I wonder if the MPI asm bits should be distributed to asm/mpilib.h headers rather than being in the lib/ directory.

David