When loading the dependency graph during incremental compilation, we already have to compute the ICH of most (or all?) HIR nodes in the current codebase. These hashes could be cached instead of calculating them again later when saving the dep-graph back out.
cc @nikomatsakis