Hacker News Reader

And I believe he should also give flags to target the specific hardware and instruction set that he has.

In extreme cases, GCC/Intel/MSVC are able to optimize pure C code with auto-vectorization and SSE/AVX instructions, possibly better than hand crafted ASM... but the compiler will never do that if not given the flags to allow it.

View on HN

Autovectorization is okay but not that good. Its advantage is just that it can transform any part of the program, when your limited dev time won't touch it all. But it can't focus on the hot loops or know about cache efficiency etc. even if you use PGO. And it's sensitive and unstable wrt code changes which is never good.

Yeah, and when autovectorization fails it can be really hard to tell why. The diagnostic messages aren't the best (or, at least, they weren't when I was doing this ~4 years ago)

EG: ICC has a AVX version of atan2(), which you can use through intrinsics[0]. But it refused to vectorize even the simplest loop with the function call in it. Even with every possible fast-math flag or every #pragma(please vectorize), it wouldn't do it. I assume due to reduced precision? Nothing I could find online had the answer (again, 4 years ago)

[0] https://software.intel.com/en-us/node/524364

Reply on news.ycombinator.com