Miscompilation of signed load by ARMv8 NCG

This Cmm program is miscompiled by the ARMv8 NCG.

test(W_ buffer) {
  return (%zx64(%shrl(bits16[buffer + (128 :: bits64)], (1 :: bits64))) & (64711 :: bits64));
}

With -fllvm (and under the bgamari/test-primops> interpreter) this evaluates to 0x40c0. With the NCG it evaluates to 0xc0c0.