jkorous added a comment. We did performance tests of alternative approach - just hashing the serialized bit code representation. There's a performance regression in the sense that while the current implementation costs approx. extra 2.2% in build time the alternative approach costs 3.8%.
We are not happy about the regression but the best way to fix the current implementation seems to be using the alternative approach as a temporary solution. The plan is to move on with upstreaming the rest of the index-while-building so we can optimize the hasher after i-w-b lands - possibly using the ODR violation hasher. This means the whole implementation will become just: hash_code RecordHash = hash_combine_range(State.Buffer.begin(), State.Buffer.end()); where SmallString<512> Buffer; Due to necessary change in interface of `IndexRecordWriter` I'll abandon this patch and land the new hasher later. CHANGES SINCE LAST ACTION https://reviews.llvm.org/D58749/new/ https://reviews.llvm.org/D58749 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
