Hacker News Reader

Comment by oconnor663 | Hacker News Reader

Maybe that depends on what "state of the art" means? It sounds like it might mean https://en.wikipedia.org/wiki/Superoptimization, but that's so slow-to-compile that I don't think any production compiler does anything like it. It might be that what "state of the art" really means in practice is "the slowest-to-compile level of optimization that regular folks are willing to put up with on a day-to-day basis." In which case, it cannot be fast kind of by definition :)

View on HN

No, not superoptimization. When I say "state-of-the-art", I meant "up-to-date techniques currently practiced in the industry", not theoretically optimal. For example, "state-of-the-art" cryptography is TLS v1.3, AES-GCM, or the Signal protocol, while a theoretically optimal one might be an "one-time pad" (or superoptimization in your example).

In other words, "We all know standard compilers are slow. Does it mean that for a compiler to reach ICC/GCC/clang's -O2 performance in benchmarks require inherently expensive optimizations? Or is it simply that ICC/GCC/clang's optimizes are poorly implemented in terms of compile performance?"

But what is practised is often “the best that's not slow enough to affect development cycle”, so it is something of a recursive definition.

When I asked this question, I make the distinction between "currently used optimization techniques and algorithms vs their implementations in GCC/clang". How large is the gap? If the gap is small, it means the algorithms as currently used for optimizations are inherently expensive (i.e. a "good" compiler "must" be slow). if the gap is large, it means a much faster implementation is theoretically possible (i.e. a "good" compiler "can" be fast).

LLVM does indeed use known suboptimal algorithms for increased speed. Its register allocator is an example: a "perfect" register allocator is NP complete, and LLVM uses a relatively simple greedy algorithm instead.

I don’t understand why a fast debug build mode and a really slow, but super-optimizing (not necessarilly in the meaning you mentioned in your link) compiler is not more used (I think zig does that). It would be the best of both words, really fast write-compile-test phase, and a production ready, fast binary on release. (i know that -O0-3 is similar, but on the whole it is not as popular as it could)

Reply on news.ycombinator.com