Optimizing Clang: A Practical Example of Applying BOLT | Hacker News Reader

Optimizing Clang: A Practical Example of Applying BOLT

github.com

48 pointsingve10 comments

Wow, that's an impressive result. I wouldn't have guessed that L1-icache-misses were that significant of a time sink such that reducing them by 1/3 would be enough to knock 15% off the total runtime.

Most of the time when I read about something that provides this much improvement on a major application, it turns out that the comparison isn't quite what it seems (as in, they are using -O0 as a baseline or some such). But from what I can tell, this is a genuine improvement over the best alternative.

Congratulations to the authors on a great job!

Yeah; wow thats a great result!

Is BOLT used much in the wild? I'd love to see it applied to something like Chrome or Firefox. A double digit performance increase would be huge improvement for a browser's html/css parsers and rendering engine.

I doubt the improvement comes from icache misses. Rather, the improvement comes from avoiding pipeline flushes by reducing mispredicts, and said mispredict reduction also reduces icache misses.

I'm a little doubtful that reduced prediction error can be the main driver here. Hardware branch predictors are really good for anything that's predictable after a short learning period, so layout changes are only going to have an effect on prediction errors for the first couple iterations on a cold start. Can there really be hundreds of billions of cold starts? Maybe, but I suspect that optimizing the straight line code path to match the most common branch is the greater effect, since consecutive instructions are almost always going to be in cache. I'm still surprised the effect is this large, but I think it's a little more likely, and it does match the blog post.

MySQL isn't the best possible test. Compile Android instead. That builds Chromium (twice, IIUC) and Java.

For bonus points, collect discrete perf data from each component build.

Apparently the conversion tool can fold in multiple datasets into a profile. It would be interesting to see whether an aggregate profile would produce the same level of efficiency as opposed to dedicated profiles for each major tool.

They should integrate BOLT into Clang/LLVM. Any BOLT dev here?

Bolt hasn’t much to do with clang/LLVM I think, unless you’re thinking it could be integrated with lld, the linker that is developed under the LLVM umbrella?

I'm still reading the paper now, but I had the impression that part of its pipeline involves disassembling binary artifacts back into IR that resembles what the compiler used to produce the binary in the first place, which could suggest that that step could benefit from tighter integration with LLVM.

Maybe they twern't suggesting integrating with the compiler.

Yeah, lld with bolt

Reply on news.ycombinator.com