You never stop learning, I just explored how I can force a compiler into producing the sbb op on x86. It’s quite difficult actually, most of them produce for example cmp + seta + neg for the term -(a > b). In fact the only deterministic behaviour is with a very special ternery term (a > b ? -1 : 0).
So the function I tried to optimize for the optimzer to be optimized was:

  // max(l, t)
  unsigned char L = (l - t); t += (L & (L < l ? -1 : 0));

The produced code, even though almost cool, was terrifying at the same time:

  0002c	8a 54 24 16	 	mov	 dl, ...
  00030	8a 44 24 04	 	mov	 al, ...
  00034	8a ca		 	mov	 cl, dl
  00036	2a c8		 	sub	 cl, al
  00038	3a ca		 	cmp	 cl, dl
  0003a	1a d2		 	sbb	 dl, dl
  00047	81 e2 ff 00 00 00	and	 edx, 255
  0004d	22 d1		 	and	 dl, cl
  0004f	02 c2		 	add	 al, dl

I marked the minor redundancy in rose, and the real offending ones red. And the worst is, I couldn’t force him to quit then. In this case it was the VC7, maybe you don’t want to expect too much from that one, I though and threw the ball to some others.

The result was VC8, VC9 and the gcc all at least dropped the and, but none, not one removed the cmp + mov (the cmp does not affect the borrow-bit, which is successively used by the sbb, thus is a nop, because it’s the one op using the backupped variable, the mov is also a nop).

The correct minimal code would be:

  0002c	8a 54 24 16	 	mov	 cl, ...
  00030	8a 44 24 04	 	mov	 al, ...
  00036	2a c8		 	sub	 cl, al
  0003a	1a d2		 	sbb	 dl, dl
  0004d	22 d1		 	and	 dl, cl
  0004f	02 c2		 	add	 al, dl

The morale: compilers still don’t do flag-aware peephole-optimization. That’s too bad after the computers became fast enough to profit from almost 40 years of compiler-development …