Extended Description

Compiling the following C code with:

$ clang -cc1 -triple thumbv5-u-u -S -O2 test.c -o -

typedef unsigned long long t;
t foo(t a, t b)
{
t a4 = b^a^18;
return a4;
}

emits:

eors	r1, r3
mov	r3, r0
eors	r3, r2
movs	r0, #​18
eors	r0, r3
bx	lr

instead of

eors	r1, r3
eors	r0, r2
movs	r4, #​18
eors	r0, r4
bx	lr

IR:
define arm_aapcscc i64 @​foo(i64 %a, i64 %b) nounwind readnone {
entry:
%xor = xor i64 %a, 18 ; [#uses=1]
%xor2 = xor i64 %xor, %b ; [#uses=1]
ret i64 %xor2
}

The first move could be eliminated if the imm value is moved to a scratch register instead of moving it to r0 which is used as a return value register.

I noticed this while developing a new backend, however it can be reproduced in ARM aswell (shown here), so it's not a backend issue.
Jakob Stoklund did a great analysis explaining where the problem is that can be found here:
http://lists.cs.uiuc.edu/pipermail/llvmdev/2010-September/034430.html