This improves the Windows self-link of LLD by 4.4% (averaged over 10
runs, with well under 1% of variance on each).
There is still some room to improve here. Two things I clearly see in
the profile:
- This is one of the biggest stress tests for the LLVM hashing code. It actually consumes something like 3-4% of the link time after the change.
- The way that StringRef keys are handled in the DenseMap interface is pretty suboptimal. We pay the price of checking for empty and tombstone keys when we could only possibly be looking for a normal key. But fixing this requires invasive API changes.
So there is still some headroom here.