Investigate performance improvements
varungandhi-src opened this issue · comments
I did some rough benchmarking on scip-clang (based on Clang 16) performance, as well as some code digging to see if there were any easy wins. It seems like the perf overhead compared to lsif-clang (based on Clang 11) is nearly 80%-90%.
| 100 TUs | 500 TUs | 1000 TUs | |
|---|---|---|---|
| lsif-clang time | 4.3 | 32.8 | 74.8 |
| scip-clang time | 5.23 | 59.7 (+ 82%) | 144.33 (+ 93%) |
| lsif-clang index size | 80M | 544M | 1.2G |
| scip-clang index size | 11M | 80M | 144M |
When comparing scip-clang to type-checking only with Clang 14 (the one I have on my system), it seems like the indexer adds about 30% overhead over type-checking. Ideally IMO this would be at most 10%-15%. However, it's still much smaller than the 80%-90% number.
So somehow, lsif-clang is even faster than the baseline type-checking in Clang 14. Some potential hypotheses:
- The clangd code in lsif-clang skips a bunch of declarations. We don't do any of that in scip-clang because we want to index everything. It might be possible to skip some things maybe, not sure. I haven't looked at the code deeply, but it's possible that if the traversal in lsif-clang is fused with type-checking (not unreasonable), then type-checking work itself is getting skipped for a bunch of decls (which would explain the "negative overhead").
- There is some serious type-checking perf regression in Clang itself between Clang 11 and Clang 14 which hasn't yet been fixed in Clang 16, explaining the missing 50%-60%. <- Seems implausible, but can't rule it out.
I'm not going to spend more time on this right now, since there isn't any obvious low-hanging fruit. That said, I think that