3 ms·

Yeah, just one that will take a decade of compiler optimizations to solve. C and C++ didn't become what they are today overnight.

by fooker 9y ago

steveklabnik 9y ago
We share an optimizer with clang thanks to LLVM; we get to take advantages of those decades already.
- fooker 9y ago
  Clang and LLVM have evolved together for a while. It is pretty difficult to get similar results with other LLVM frontends. You don't have to believe me, just write a reasonably complicated algorithm (say.. a naive SAT solver) with both C++ and Rust (or Swift or $NEW_LANG) and compare the speeds.
  - steveklabnik 9y ago
    You don't have to believe me either; if you try it, you should generally see that they have the same performance.
    - javajosh 9y ago
      When you say "naive SAT solver" do you mean one where you hard-code a boolean formula with N variables and then check every possible solution (from 0 to 2^N)? If so, does this kind of program accurately characterize the kind of work a kernel does?
      - fooker 9y ago
        No, something like the CDCL algorithm without domain specific heuristics and tweaks. I expect that is fairly representative in terms of ..say.. common data structures needed in most systems.
    - fooker 9y ago
      Here are some fairly realistic examples. 1. Rust : http://benchmarksgame.alioth.debian.org/u64q/compare.php?lang=rust&lang2=gcc http://benchmarksgame.alioth.debian.org/u64q/compare.php?lan... 2. Swift : http://benchmarksgame.alioth.debian.org/u64q/compare.php?lang=swift&lang2=gcc http://benchmarksgame.alioth.debian.org/u64q/compare.php?lan... As for a SAT solver, I was not speaking out of my ass. Here is one try that got a fair bit of attention. https://github.com/mishun/minisat-rust https://github.com/mishun/minisat-rust
      - steveklabnik 9y ago
        Yeah, we've had some small regressions recently; a few months ago we were usually faster, now we're slightly slower. Still generally close, except for the ones that need explicit SIMD, as that's still being designed.

Yeah, just one that will take a decade of compiler optimizations to solve. C and C++ didn't become what they are today overnight.

by fooker 9y ago

Yeah, just one that will take a decade of compiler optimizations to solve. C and C++ didn't become what they are today overnight.

steveklabnik 9y ago

We share an optimizer with clang thanks to LLVM; we get to take advantages of those decades already.

fooker 9y ago
Clang and LLVM have evolved together for a while. It is pretty difficult to get similar results with other LLVM frontends. You don't have to believe me, just write a reasonably complicated algorithm (say.. a naive SAT solver) with both C++ and Rust (or Swift or $NEW_LANG) and compare the speeds.
- steveklabnik 9y ago
  You don't have to believe me either; if you try it, you should generally see that they have the same performance.
  - javajosh 9y ago
    When you say "naive SAT solver" do you mean one where you hard-code a boolean formula with N variables and then check every possible solution (from 0 to 2^N)? If so, does this kind of program accurately characterize the kind of work a kernel does?
    - fooker 9y ago
      No, something like the CDCL algorithm without domain specific heuristics and tweaks. I expect that is fairly representative in terms of ..say.. common data structures needed in most systems.
  - fooker 9y ago
    Here are some fairly realistic examples. 1. Rust : http://benchmarksgame.alioth.debian.org/u64q/compare.php?lang=rust&lang2=gcc http://benchmarksgame.alioth.debian.org/u64q/compare.php?lan... 2. Swift : http://benchmarksgame.alioth.debian.org/u64q/compare.php?lang=swift&lang2=gcc http://benchmarksgame.alioth.debian.org/u64q/compare.php?lan... As for a SAT solver, I was not speaking out of my ass. Here is one try that got a fair bit of attention. https://github.com/mishun/minisat-rust https://github.com/mishun/minisat-rust
    - steveklabnik 9y ago
      Yeah, we've had some small regressions recently; a few months ago we were usually faster, now we're slightly slower. Still generally close, except for the ones that need explicit SIMD, as that's still being designed.