[PATCH] ARM: For thumb fixups store halfwords high first and low second

[scott douglass]

  These changes look reasonable to me and do improve things on our ABI test suite.

  But it should include one or more regression tests in llvm/test/MC/ARM along the lines of:

  thumb2be-b.w-encoding.s
    @ RUN: llvm-mc -triple=thumbebv7-none-linux-gnueabi -mcpu=cortex-a8 -filetype=obj < %s | llvm-objdump -arch=thumbeb -s - | FileCheck %s
      .syntax unified
      .code	16
      .thumb_func
    foo:
      b.w   bar

    @ CHECK: Contents of section .text:
    @ CHECK-NEXT: 0000 f7ffbffe

  I'd rather use 'llvm-objdump -arch=thumbeb -disassemble -' instead of '-s' but llvm-objdump seems broken for -arch=thumbeb (which would be good to fix too, separately).

http://reviews.llvm.org/D3380