"`const static`" function parameters

General Usage
1

I have a function that uses a constant parameter. The parameter never changes between function calls. For example:

function radtodeg(x)
    const PI = 3.14159
    return x * 180 / PI
end

Here PI is a constant that doesn’t change between function calls. Moreover, it is only used by the function (PI is a bad example in this regard), so it makes no sense to declare it as a global constant. Also, it makes no sense to have the caller pass the constant variable pre-allocated, because this is a parameter that only makes sense internally in the function.

A more complicated example could use a constant array instead of just a number. I don’t want to have to reallocate this variable every time the function is called.

In C++, you could mark such constant parameters as const static. I understand that a proposal for static variables was rejected:

https://github.com/JuliaLang/julia/issues/12627

So what is the proper way to deal with this scenario? Or maybe Julia’s compiler is smart enough to detect this condition, and allocates PI only once?

1 Like
2

I think Jameson’s suggestion from the end of the thread you linked should just work in your case:

let 
   const a = pi
   global f
   function f(x)
       x + a
   end
end

The let block ensures that the constant a doesn’t show up at global scope, and the global f ensures that the function f does show up at global scope.

1 Like
3

I saw that. I just think it is too convoluted. Specially compared to just saying const static as in C++.
I was wondering if there was a better and also more readable way.

2 Likes
4

Also, no const. <20 chars…>

5

If verbosity is a problem, you can go back to the good ol’ eval at compile time trick, which comes with its own set of restrictions:

macro static_var(x)
    anon = gensym()
    fun = eval(Main, :(const $anon = $x))
    :(Main.$anon)
end
function test()
    secret = @static_var begin
        println("Var, so static right now")
        "i did it!"
    end
    secret
end
>Var, so static right now # println at definition time

test() # no println no more
> i did it

julia> @code_warntype test()
Variables:
  #self#::#test
  secret::String

Body:
  begin
      secret::String = Main.##270 # line 6:
      return secret::String
  end::String

You could come up with quite a variety of macros to do similar things depending on your taste and goals.

6

@sdanisch This only works with compile-time constants, right? What if my static variable is run-time constant (meaning that the value is set once and for all at run time, but may change between runs)?

Maybe that use-case is too weird and need not be considered…

7

I don’t think that feature request was rejected so much as deemed “not important enough to work on now”. There are ways of getting what you need and a static variable feature could be added in the future. Using a global constant here seems fine to me since a static variable is just a global constant whose name is only visible to one function.

2 Likes
8

@StefanKarpinski It would be nice if the variable was visible only inside the function. It would not pollute global scope.

9

Can you explain what the problem with just using

PI = 3.14159

is?

AFAIK there are no performance gains to a const inside a function in Julia and this will be optimized. What is the problem you are trying to solve?

For example for the code above:

function radtodeg2=(x)
       PI = 3.14159
       return x * 180 / PI
end

Then code_llvm returns

define double @julia_radtodeg_71699(double) #0 { 
top:                                              
  %1 = fmul double %0, 1.800000e+02               
  %2 = fdiv double %1, 3.141590e+00               
  ret double %2                                   
}

i.e the constant is inline.

2 Likes
10

@jtravs I’m thinking of a constant that takes more space, like an array of coefficients. I don’t think that can be inlined, and I want to make sure it is allocated only once.

11

If it’s an array of constants, can’t you write it without the array? Maybe a macro would make it nicer, and then you could use a1 instead of a[1] etc., and it would all inline at compilation.

12

Sometimes you just want to put a fat array inside a function. The macro posted above is a bit of a hack but it does the job.

13

@ChrisRackauckas What if it is a constant matrix? You don’t want to expand it into components a1, ... because you want to exploit matrix multiplication.

1 Like
14

FWIW, you can boil this down to:

macro static(ex)
  eval(ex)
end

:slight_smile: (though you might strictly need a quotenode or something if ex could be an expression)

It is indeed pretty weird to want to do this, but just for fun:

using MacroTools

macro static(ex)
  @capture(ex, name_::T_ = val_) || error("invalid @static")
  ref = Ref{eval(T)}()
  set = Ref(false)
  :($(esc(name)) = if $set[]
      $ref[]
    else
      $ref[] = $(esc(ex))
      $set[] = true
      $ref[]
    end)
end

function test(n)
  @static pi::Float64 = 2n
  return pi
end

test(6) # 12.0
test(3) # also 12.0
4 Likes
15

@MikeInnes I propose that something like this be included in Base. Don’t you think?

16

@MikeInnes Why is the llvm code so messy?

@code_llvm test(6)

define double @julia_test_71387(i64) #0 {
top:
%1 = load i8, i8* inttoptr (i64 139707517794576 to i8*), align 16
%2 = and i8 %1, 1
%3 = icmp eq i8 %2, 0
br i1 %3, label %L, label %L1

L: ; preds = %top
%4 = shl i64 %0, 1
%5 = sitofp i64 %4 to double
store double %5, double* inttoptr (i64 139707517794560 to double*), align 256
store i8 1, i8* inttoptr (i64 139707517794576 to i8*), align 16
br label %L1

L1: ; preds = %top, %L
%“#temp#.sroa.0.02” = load double, double* inttoptr (i64 139707517794560 to double*), align 256
ret double %“#temp#.sroa.0.02”
}

17

I suspect it’s not widely useful enough to have in base; but the good news is that if it doesn’t require language support, there’s no real downside to having it in a package.

The LLVM code isn’t going to be as clean as the “really static” version because you have to do pointer lookups and branches to initially set the variable. Such is the cost of being dynamic, I’m afraid.

18

Depends on how big it is. It may be a good idea to skip BLAS if the constant matrix is small enough. But yes, the right approach all depends on the problem, which I do not know. I was just suggesting other approaches.

Another way you might want to handle this without polluting any good names in the global namespace is to use a more specific global name, and alias it in the function. Like

const GLOBAL_PI = ...

and then in the function, just

pi = GLOBAL_PI
19

I dont think it’s weird. This is how local static variables work in C. I think it would also reduce a lot of the use cases for global variables. IMHO, static local vars would be better than globals.

1 Like
20

This reminds me of a situation I encounted the other day in which some kind of global variable would be convenient. I had an array of objects that needed a common variable. In some languages, its possible to do something like this:

mutable myStruct
    global x::Int
    y::Int
end

M = [myStruct() for i in 1:100]

myStruct.x = 1

The field x would be the same across all elements in M and could be updated simultaneously. The field y, on the other hand, could be unique for each element in M. Is it possible to achieve this with a variation of your code (it does not appear to work as is)?