In fact the answer was 5 bits for 32bit and 7 bits for 64bit[1]. x86-64 uses 5 bits for 32bit and for 64bit (shld/shrd can't shift more than 31 bits).
7 bits is problematic because some code may assume that x << -y is an optimization for x << (32-y). I thought there was some architecture where this assumption didn't work on 32bit either, which led to my post above.
