Uh, I don’t know how I may introduce this one. I don’t know if I should call it a horrific failure or just hilarious. This is what comes out of VS2008 for a signed short division:

  symSend = mergedNGram % numCases;
  mergedNGram /= numCases;

The produced code:

  00024	0f bf 86 2c 04
        00 00		 movsx	 eax, WORD PTR [esi+1068]
  0002b	99		 cdq
  0002c	b9 03 00 00 00	 mov	 ecx, 3
  00031	f7 f9		 idiv	 ecx
  00033	0f b7 ca	 movzx	 ecx, dx
  00036	66 89 86 2c 04
        00 00		 mov	 WORD PTR [esi+1068], ax

This little bugger knows exactly we have shorts (he sign extends them to longs), and he also knows, that the results must be shorts (he does not consider results over 65535)! I got courious and tried everything to make it produce non-long division, especially a char division. Nada, njiet.
I would forgive it it’s behaviour if it would be just a plain simple non-optimizing compiler, but this one is complex and it does have the entire boundary-check done right! And anyway he damn doesn’t produce a 6-7 clocks faster 8bit division!?