Previously, CFL-AA always says "ModRef" to any getModRefInfo() queries. This patch makes CFL-AA returns slightly better answers.
The precision increase is less than I originally expected. The biggest problem here is that we don't distinguish reads from writes in our analysis (it's also useless to make the distinction, since the value we read from and the value we write to are very likely to be merged together anyway, if there's a direct/indirect assignment). Therefore, all I can do here is to return NoModRef in certain special cases, and try to be conservative otherwise.