Error in max_double implementation

[eschnett]

I see this strange behaviour in Julia 1.12-dev (todays nightly build), first reported at MakieOrg/Makie.jl#4729:

julia> using IntervalSets

julia> f(x) = (iv=0.0..1.0; mod(x,iv))
f (generic function with 1 method)

julia> f(0.0)
NaN

The correct result is 0.0, and this works fine in Julia 1.11, and also when mod is called directly instead of in a function (mod(0.0, 0.0..1.0)).

I believe this is a bug in Julia's constant propagation mechanism. I have posted more details in the Makie.jl issue.

I see that Julia determines already at compile time that the call to mod will return NaN:

julia> @code_warntype f(0.0)
MethodInstance for f(::Float64)
  from f(x) @ Main REPL[1]:1
Arguments
  #self#::Core.Const(Main.f)
  x::Float64
Locals
  iv::ClosedInterval{Float64}
Body::Float64
1 ─ %1 = Main.:..::Core.Const(IntervalSets...)
│        (iv = (%1)(0.0, 1.0))
│   %3 = Main.mod::Core.Const(mod)
│   %4 = iv::Core.Const(0.0 .. 1.0)
│   %5 = (%3)(x, %4)::Core.Const(NaN)
└──      return %5

IntervalSets' mod function is

mod(x, i::TypedEndpointsInterval{:closed,:closed}) = mod(x - leftendpoint(i), width(i)) + leftendpoint(i)

and its width function is

function width(A::AbstractInterval)
    _width = rightendpoint(A) - leftendpoint(A)
    max(zero(_width), _width)   # this works when T is a Date
end

I find that removing the line calling max avoids the problem. It seems to me that this line somehow confuses Julia into assuming that the second argument to mod is 0.

[eschnett]

Here is a stand-alone test case:

struct IV
    a::Float64
    b::Float64
end
width(iv::IV) = (w=iv.b-iv.a; max(zero(w),w))
Base.mod(x,iv::IV) = mod(x-iv.a, width(iv)) + iv.a
f(x) = (iv=IV(0.0,1.0); mod(x,iv))
f(0.0)

The call to f(0.0) returns NaN. The correct result would be 0.0.

[giordano]

Any reference for when this last worked? It'd help for a bisect.

[eschnett]

Sorry, no – 1.11 works for me.

[Seelengrab]

It works correctly for me on 195c4d2, which was from #56784, so quite recent.

[adienes]

on my machine this bisects to a861a551b764062d47ab3c2a3748bcf540221bae aka #56371 @gbaraldi

julia> versioninfo()
Julia Version 1.12.0-DEV.1885
Commit a861a551b7 (2025-01-14 04:14 UTC)
Platform Info:
  OS: macOS (arm64-apple-darwin24.2.0)
  CPU: 10 × Apple M1 Max
  WORD_SIZE: 64
  LLVM: libLLVM-18.1.7 (ORCJIT, apple-m1)
  GC: Built with stock GC
Threads: 1 default, 0 interactive, 1 GC (on 8 virtual cores)

[eschnett]

Look at these two functions from runtime_intrinsics.c:

float max_float(float x, float y) JL_NOTSAFEPOINT
{
    float diff = x - y;
    float argmin = signbit(diff) ? y : x;
    int is_nan = isnan(x) || isnan(y);
    return is_nan ? diff : argmin;
}

double max_double(double x, double y) JL_NOTSAFEPOINT
{
    double diff = x - y;
    double argmin = signbit(diff) ? x : y;
    int is_nan = isnan(x) || isnan(y);
    return is_nan ? diff : argmin;
}

Do you spot the difference? Only one of them is correct...

[eschnett]

Spoiler alert: the variable should be called argmax.

[gbaraldi]

This function didn't exist in 1.10/1.11 so no backports needed

[vtjnash]

Ah okay, I didn't realize #56371 was new. All new intrinsics added should have a test added to exercise them in test/intrinsics.jl

[eschnett]

@gbaraldi This is still a regression in 1.12...

[gbaraldi]

Too many tags 😕 yeh