I wonder how hard the JVM folks looked into specialized weak references for solving this issue. The mark phase would treat all Strings with zero offset and ful

I wonder how hard the JVM folks looked into specialized weak references for solving this issue. The mark phase would treat all Strings with zero offset and full length as strongly referencing the byte[], and weakly otherwise. At the end of each full GC, you could iterate over all of your Strings (custom allocate/compact them to their own ranges of the heap for faster scanning), use some heuristics and probabilistic sampling to select some of the weakly reachable byte[]s for size reduction. A specialized copy/compact pass over the Strings could in-place replace byte[] references and fix up offsets. You'd probably also want to modify String.equals() to internally mutate equal strings to point to the same byte[], preferring smaller offsets, and when offsets are equal, preferring lower addresses. This is a light weight lazy version of the background String byte[] interning done by some JVMs.