> On May 5, 2015, at 8:07 PM, David Blaikie <[email protected]> wrote: > > If we believe hash collisions to be an issue - then type units are broken and > I'd like to figure out how to fix them, then use the same feature here. >
This led me to perform a little experiment: I had a python script calculate the MD5 sum of ~60.000 unique mangled type names that I extracted from the IR of a clang LTO build. There was not a single collision in the abbreviated 64-bit hashes, and I had to truncate the length of the hash values to 32 bits before I was able to produced even a single collision. I think now that hash collisions might be a red herring and a calculable risk that we can decide to ignore. The only thing that we cannot ignore here is that for ODR-less languages two types may legally have the same mangled name (and thus hash) when they are defined in two different modules. This means we can’t use type units for these languages and external type reference must be a pair of mangled name and module. -- adrian _______________________________________________ cfe-commits mailing list [email protected] http://lists.cs.uiuc.edu/mailman/listinfo/cfe-commits
