Re: Native code function call overhead vs webassembly

Thank you very much for your fast reply.

I think my question was a bit misleading.
I am interested in knowing the function call overhead of JS/native (e.g. with functions defined in C++ that are statically linked with the SM runtime with JS_DefineFunction) and how does it compare with the JS/wasm calls (with emscripten-generated wasm).
Do JS->native function calls have a higher overhead than JS->wasm or the other way around? Are there significant differences?

Andrea

On Monday, March 8, 2021 at 9:26:58 AM UTC+1, Lars Hansen wrote:
> In general, you get pretty good JS/wasm and wasm/JS intercalling 
> performance if the JS code has been jitted and accepts known types (as will 
> happen once you've run the code a few times), and if the JS compiler can 
> determine the types you'll also get good performance of the JS code. (And 
> if performance remains poor we'd probably like to hear about it.) If you 
> have a large library already and don't want to port it to C++, I would 
> probably stick with this code until profiling shows that the overheads are 
> unaffordable. 
> 
> However, you're likely to see better performance by converting the code to 
> C++ and compiling that to wasm. Procedure call overheads will be 
> significantly lower - it's not cheap to cross the wasm/JS boundary even in 
> the best case - and the C++ compiler is likely to have more sophisticated 
> optimizations than the JS jit, as well as better type information to work 
> from, and it will need to insert no run-time type guards. 
> 
> (A final possibility is to port to C++ and then compile the C++ into the 
> wasm engine, and invoke it using custom instructions, but this is rather a 
> large project.) 
> 
> --lars
> On Mon, Mar 8, 2021 at 9:00 AM Bellaz89 <[email protected]> wrote: 
> 
> > Hi everyone, 
> > 
> > I am working with an application that uses SM78 as its scripting engine. 
> > This application has a bunch of mathematical libraries in pure JS that I 
> > want to port to emscripten/webassembly. I see that some years ago progress 
> > was made in improving the call overhead between JS <-> webassembly ( 
> > https://hacks.mozilla.org/2018/10/calls-between-javascript-and-webassembly-are-finally-fast-%F0%9F%8E%89/). 
> > This is highly beneficial for me. However, I wonder how this compares to 
> > functions/modules implemented using the native code interface in C++. How 
> > much is the native code call overhead? How does it compare to webassembly? 
> > Is there an advantage in using one solution (wasm or native c++) for 
> > 'short' functions? 
> > 
> > Thanks in advantage and sorry if the question is a bit vague. 
> > 
> > Andrea
> > _______________________________________________ 
> > dev-tech-js-engine mailing list 
> > [email protected] 
> > https://lists.mozilla.org/listinfo/dev-tech-js-engine 
> >