For example, a graph's edge have 2 property: weight & capacity. I need to
compute a path which have min weight and any edges in this path has bigger
capacity than 'x'.
Seems BGL can't do this uniformly because shortest_path function has no
place to accept others parameter(constraints). Some functors such as
distance_compare seems only work on weight.
That's not entirely true. distance_compare compares the results of lookups on the distance_map provided to the function. A property map is actually very much like a functor except that it supports reads and writes. This means you can build a custom property map that returns (for example) a pair containing the weight and capacity for each edge. The distance_compare function would then compare the resulting pairs.
I mean BGL cope with pure graph theory problem nicely, but when the
application gets complex(still can be reduced to pure graph data structure,
but more rules restricted on algorithm), BGL fails.
Am I right?
Not really.