4 ms·

Transpile to C? Zig or Rust?

skohan 4y ago
Yeah I think Zig in particular is trying to be just exactly a "better C". I don't know if transpiling will get you there, because for instance if you're transpiling a dynamic language, you're going to have to output C that is essentially emulating all those dynamic language features, so it might be faster than say, the original Python, but it's not going to be as fast as a pure C implementation.
bruce343434 4y ago
You can't just "transpile to C" to get "C speed". C speed comes from low overhead. Naive transpilations will just include that overhead, like garbage collection or many layers of pointer indirection, but written in C. The unsavory answer is that you must use less abstractions if you want fast code. Compilers just aren't good enough to compile all the abstractions away.
- pjmlp 4y ago
  That includes C versus modern CPUs.
  - dahfizz 4y ago
    The CPU is the CPU regardless of what language you are running on it. C will still give you the best performance on a modern CPU. If anything, C has an even bigger advantage on modern CPUs because it has easier access to things like vectorize/SIMD intrinsics. It is also easier to tweak your data dependencies to help the branch predictor.
    - pjmlp 4y ago
      Nope, those are language extensions not defined by ISO C. Any language can have such intrinsics as extensions, for example D, Rust, C++, Swift, .NET, Java (as preview).
    - kaba0 4y ago
      > If anything, C has an even bigger advantage on modern CPUs because it has easier access to things like vectorize/SIMD intrinsics. It is also easier to tweak your data dependencies to help the branch predictor. Have you seen the ridiculously complex optimizations that C compilers do to maybe turn some shitty for loop into vector instructions? C is terribly bad fit for this use case and hardware tries to get closer to C than the reverse. C++, Rust but even C# and Java has much better SIMD support than C has. EDIT: It really doesn’t help C (and some of the listed languages as well) that they are very imperative. SIMD is exactly the place where pureness and some form of FP is much better at allowing these kind of optimizations (a map of a pure function can be “trivially” optimized into vector instructions, while it is really hard to decide whether this for loop is safe to convert, and it is really up to the heuristics of the compiler. A bit of rearrangement can cause a failure to optimize, resulting in a huge drop in performance)
mhaberl 4y ago
> Transpile to C? If you transpile something to C it does't mean it will be fast. You can write slow C code (or transpile something to C that will be slow). The compilers are not the issue here.
- gpderetta 4y ago
  In fact it might be slower than actually targeting a proper compiler intermediate language.

Yeah I think Zig in particular is trying to be just exactly a "better C". I don't know if transpiling will get you there, because for instance if you're transpiling a dynamic language, you're going to have to output C that is essentially emulating all those dynamic language features, so it might be faster than say, the original Python, but it's not going to be as fast as a pure C implementation.

bruce343434 4y ago

You can't just "transpile to C" to get "C speed". C speed comes from low overhead. Naive transpilations will just include that overhead, like garbage collection or many layers of pointer indirection, but written in C. The unsavory answer is that you must use less abstractions if you want fast code. Compilers just aren't good enough to compile all the abstractions away.

pjmlp 4y ago
That includes C versus modern CPUs.
- dahfizz 4y ago
  The CPU is the CPU regardless of what language you are running on it. C will still give you the best performance on a modern CPU. If anything, C has an even bigger advantage on modern CPUs because it has easier access to things like vectorize/SIMD intrinsics. It is also easier to tweak your data dependencies to help the branch predictor.
  - pjmlp 4y ago
    Nope, those are language extensions not defined by ISO C. Any language can have such intrinsics as extensions, for example D, Rust, C++, Swift, .NET, Java (as preview).
  - kaba0 4y ago
    > If anything, C has an even bigger advantage on modern CPUs because it has easier access to things like vectorize/SIMD intrinsics. It is also easier to tweak your data dependencies to help the branch predictor. Have you seen the ridiculously complex optimizations that C compilers do to maybe turn some shitty for loop into vector instructions? C is terribly bad fit for this use case and hardware tries to get closer to C than the reverse. C++, Rust but even C# and Java has much better SIMD support than C has. EDIT: It really doesn’t help C (and some of the listed languages as well) that they are very imperative. SIMD is exactly the place where pureness and some form of FP is much better at allowing these kind of optimizations (a map of a pure function can be “trivially” optimized into vector instructions, while it is really hard to decide whether this for loop is safe to convert, and it is really up to the heuristics of the compiler. A bit of rearrangement can cause a failure to optimize, resulting in a huge drop in performance)

mhaberl 4y ago

> Transpile to C? If you transpile something to C it does't mean it will be fast. You can write slow C code (or transpile something to C that will be slow). The compilers are not the issue here.

gpderetta 4y ago
In fact it might be slower than actually targeting a proper compiler intermediate language.