Interface with Mjolnir about symbolicutils.jl HOT 4 OPEN

It might also be a good idea to use Mjolnir to extract symbolic expressions from Julia code, over the current operator overloading approach. This would be better able to support type constraints, intrinsics (ifelse), higher-order functions (map/reduce/sum etc.) and even simple control flow (the branches could be lowered to a functional or ifelse representation if that's easier for SymbolicUtils to work with). I plan to throw up a prototype of that idea soon.

from symbolicutils.jl.

the branches could be lowered to a functional or ifelse representation

By the way, the XLA work effectively has a proof of concept for this:

julia> using XLA

julia> relu(x) = x > 0 ? x : 0
relu (generic function with 1 method)

julia> ir = XLA.@code_xla relu(5)
1: (%1 :: Int64)
  %2 = (XLA.Gt())(%1, 0)
  br 2 unless %2
  br 3 (%1)
2:
  br 3 (0)
3: (%3 :: Int64)
  return %3

julia> XLA.controlflow(ir)
1: (%1 :: Int64)
  %2 = (XLA.Gt())(%1, 0)
  %4 = (XLA.Not())(%2)
  %5 = 
    1: (%1)
      %2 = (XLA.XTuple())(0)
  %6 = 
    1: (%1)
  %7 = (XLA.XTuple())()
  %8 = (XLA.XTuple())(%1)
  %9 = (XLA.Conditional())(%4, %7, %5, %8, %6)
  %10 = (XLA.GetTupleElement(0))(%9)

The XLA representation is a bit fiddly, but the controlflow transformed IR is equivalent to

julia> cond(c, t, f) = c ? t() : f() # XLA's `cond`, roughly

julia> relu(x) = cond(x > 0, () -> x, () -> 0)

Which is equivalent to a piecewise function definition. I'm not sure if you support things like lambdas, piecewise functions, list comprehensions etc., or if you plan to, but at any rate there's no issue using pulling those things out of Julia code.

from symbolicutils.jl.

Awesome! Yes this could be a fix for #16 cc @MasonProtter

from symbolicutils.jl.

the branches could be lowered to a functional or ifelse representation if that's easier for SymbolicUtils to work with

Just added cond in #75

from symbolicutils.jl.