at-evalpoly is broken

[stevengj]

The following works fine in 0.6:

julia> f(x) = @evalpoly x 1.0 2.0 3.0
f (generic function with 1 method)

julia> Base.Test.@inferred f(3.0)
ERROR: return type Float64 does not match inferred return type Any
Stacktrace:
 [1] error(::String) at ./error.jl:28

Probably the same as JuliaMath/SpecialFunctions.jl#42, where it was bisected to #22985.

cc @vtjnash

[yuyichao]

This is basically #23221. Right now it's impossible to write macros that calls other macros correctly without doing everything manually.

[stevengj]

As a workaround, we could inline the @horner macro; we could even eliminate that undocumented macro entirely in favor of @evalpoly.

However, in the long run I think we need to make it possible for macros to call macros.

[JeffBezanson]

From the code_lowered for f:

        Base.Math.t = tt
        SSAValue(0) = (Base.Math.muladd)(Base.Math.t, (Base.Math.muladd)(Base.Math.t, 3.0, 2.0), 1.0)
        # meta: pop location
        return SSAValue(0)

The horner macro returns a block with an assignment to a (hygienic) variable t, which somehow becomes a global. This seems different from the other recent macro issues, since t should be local no matter which scope it belongs to.

[JeffBezanson]

Reduced example:

julia> macro id(x)
        esc(x)
       end

julia> macro aa()
        quote
         a = 1
         @id b = 1
        end
       end

julia> @aa;

julia> a
ERROR: UndefVarError: a not defined

julia> b
1

I can see why this happens, but I find it really confusing.

[StefanKarpinski]

Related fallout: #23221, JuliaCI/BenchmarkTools.jl#69, JuliaMath/SpecialFunctions.jl#42.

[dlfivefifty]

Just came across what's presumably the same bug. On 0.6:

julia> W2 = @evalpoly(0.1729150690306449, @evalpoly(1.2151064605177766, -27.0, -84.0, -56.0),
                                         -11.614665966268532,
                                         163.17740776843848) 
-208.88143856876326

On 0.7:

julia> W2 = @evalpoly(0.1729150690306449, @evalpoly(1.2151064605177766, -27.0, -84.0, -56.0),
                                         -11.614665966268532,
                                         163.17740776843848) 
-191.5798454805012

[stevengj]

@dlfivefifty, your case is also fixed by #23247.

[dlfivefifty]

👍 FastGaussQuadrature.jl now passes all tests on 0.7