Backout of r59196, plus a new ICE test. Sorry if this is a

little rude; I figure it's cleaner to just back this out now so 
it doesn't get forgotten or mixed up with other checkins.

The modification to isICE is simply wrong; I've added a test that the 
change to isICE breaks.

I'm pretty sure the modification to tryEvaluate is also wrong.  
At the very least, there's some serious miscommunication going on here, 
as this is going in exactly the opposite direction of r59105.  My 
understanding is that tryEvaluate is not supposed to care about side 
effects.  That said, a lot of the clients to tryEvaluate are 
expecting it to enforce a no-side-effects policy, so we probably need 
another method that provides that guarantee.

llvm-svn: 59212