aaronpuchert wrote: > Does this have measurable performance impacts? Or was there some other > benefit driving the change?
Thread Safety Analysis is not very expensive overall. I just benchmarked compilation of `clang/test/SemaCXX/warn-thread-safety-analysis.cpp` and it comes out at around 1% of the overall compilation time. (Although it is also just one analysis of many.) Overall compilation before the change takes 327,801,317 instructions, and 327,433,878 after the change, according to `perf stat -r 100` with an error of 0.01%. Which makes it 0.1% faster, or 10% of the time for Thread Safety Analysis. So in that sense it is measurable. Will an end user notice? Probably not. (Release build with ThinLTO but shared libs for faster recompilation. A production build might behave slightly different. The profile showed 15–20% in `ld-linux-x86-64.so.2`.) But the real motivation was conceptual: * We analyze one function after the other, so for every function, we create a bunch of nodes and then throw them all away. We never throw away old facts before we're done with the function. So lifetimes already behave as if we had an arena: we allocate a bunch of entries and free everything at the end. * We already use the arena for TIL expressions referenced by the `FactEntry`s. With this change we no longer mix "persistent" allocations and arena allocations. Everything lives in the arena, and when we're done we throw the arena away. https://github.com/llvm/llvm-project/pull/149660 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
