[compiler-rt: Builtins] Fix the incompatible assembly in ARM builtins library.

[sgundapa]

Summary

This patch address the PR20360. It has the following fixes

1. Make the assembly code in synh-ops.h to be UAL compliant
2. Convert the CBNZ backward branch instruction to CMP and BNE

[jroelofs]

LGTM, though I think you should check this in as two separate changes, first 1, then 2&3.

[sgundapa]

Sure. I will split in to two

[sgundapa]

I split the earlier patch in to two

[sgundapa]

The other half of the patch is at
http://reviews.llvm.org/D4614

[sgundapa]

I don't have commit access.
Can some one merge this ?

[rengolin]

r213685

lib/builtins/arm/sync-ops.h

  *
  *===----------------------------------------------------------------------===*/
 
 #include "../assembly.h"
 
 #define SYNC_OP_4(op) \
         .p2align 2 ; \
         .thumb ; \
+        .syntax unified ; \
         DEFINE_COMPILERRT_FUNCTION(__sync_fetch_and_ ## op) \
         dmb ; \
         mov r12, r0 ; \
         LOCAL_LABEL(tryatomic_ ## op): \
         ldrex r0, [r12] ; \
         op(r2, r0, r1) ; \
         strex r3, r2, [r12] ; \
-        cbnz r3, LOCAL_LABEL(tryatomic_ ## op) ; \
+        cmp r3, #0 ; \
+        bne LOCAL_LABEL(tryatomic_ ## op) ; \
         dmb ; \
         bx lr
 
 #define SYNC_OP_8(op) \
         .p2align 2 ; \
         .thumb ; \
+        .syntax unified ; \
         DEFINE_COMPILERRT_FUNCTION(__sync_fetch_and_ ## op) \
         push {r4, r5, r6, lr} ; \
         dmb ; \
         mov r12, r0 ; \
         LOCAL_LABEL(tryatomic_ ## op): \
         ldrexd r0, r1, [r12] ; \
         op(r4, r5, r0, r1, r2, r3) ; \
         strexd r6, r4, r5, [r12] ; \
-        cbnz r6, LOCAL_LABEL(tryatomic_ ## op) ; \
+        cmp r6, #0 ; \
+        bne LOCAL_LABEL(tryatomic_ ## op) ; \
         dmb ; \
         pop {r4, r5, r6, pc}
 
 #define MINMAX_4(rD, rN, rM, cmp_kind) \
         cmp rN, rM ; \
         mov rD, rM ; \
         it cmp_kind ; \
         mov##cmp_kind rD, rN