What could the Julia compiler learn from the LPython compiler?

xiaoxi · July 29, 2023, 7:45am

LPython is a Python compiler that can compile Python code with type annotations into optimized machine code, and offers Just-In-Time (JIT) compilation and seamless interoperability with CPython.

jdad · July 29, 2023, 11:42am

Note, Julia is mentioned in the announcement as a possible backend, but after a short exploration of both LPython and LFortran websites (both .org and GitHub) I could not find any other trace of it …

longemen3000 · July 30, 2023, 4:32am

github.com

lcompilers/lpython/blob/main/src/libasr/codegen/asr_to_julia.cpp

#include <libasr/asr.h>
#include <libasr/asr_utils.h>
#include <libasr/pass/intrinsic_function_registry.h>
#include <libasr/diagnostics.h>
#include <libasr/codegen/asr_to_julia.h>

namespace LCompilers {

/*
Julia operator precedence:
https://docs.julialang.org/en/v1/manual/mathematical-operations/#Operator-Precedence-and-Associativity

Can also be queried by `Base.operator_precedence()`.

Different from C++, the larger the number, the higher the precedence. To follow LFortran's
convention, we need to revert the precedence table.
*/
enum julia_prec {
    Base = 2,
    Pow,         // ^

This file has been truncated. show original

jdad · July 30, 2023, 9:44am

So it was in libasr. Thanks for the ref. I wonder if somebody could try (I may be doing that later) / have an opinion of converting python to Julia ?

stevengj · July 30, 2023, 12:02pm

This is a FAQ.

lmiq · July 30, 2023, 12:04pm

From what I can see, LPython is compiling code where every variable type is declared and concrete. Like a Fortran code. It doesn’t include a type inference machinery. Julia does not have a declaration syntax that guarantees that a label is assigned always to the same type. Thus that conversion is probably limited.

jdad · July 30, 2023, 12:20pm

Yes, now I remember it. The FAQ item explains very clearly why Julia as a compiler is needed/much better solution than thinking of it as a front end.

My asking/wondering was more on the two following other points , 1) why do a Julia backend, is there any interest for those LPython users - or developers-?, and 2) could it help to pushed some Python users to Julia?

That was only my 2cts of contribution inspired by the news,

PS. For 1), I was (notably) thinking of the possible usefulness of Julia stack of GPU, that is CUDA.jl, AMDGpu.jl, Metal.jl, oneAPI.jl, …