exp-w (used to crash)

Percentage Accurate: 99.3% → 99.3%

Time: 14.2s

Alternatives: 8

Speedup: 1.0×

• could not determine a ground truth

  1. w = 8.98941569039849e+74
  2. l = 3.355091021728e+220

Specification

Math

\[\begin{array}{l} \\ e^{-w} \cdot {\ell}^{\left(e^{w}\right)} \end{array} \]

FPCore

(FPCore (w l) :precision binary64 (* (exp (- w)) (pow l (exp w))))

C

double code(double w, double l) {
	return exp(-w) * pow(l, exp(w));
}

Fortran

module fmin_fmax_functions
    implicit none
    private
    public fmax
    public fmin

    interface fmax
        module procedure fmax88
        module procedure fmax44
        module procedure fmax84
        module procedure fmax48
    end interface
    interface fmin
        module procedure fmin88
        module procedure fmin44
        module procedure fmin84
        module procedure fmin48
    end interface
contains
    real(8) function fmax88(x, y) result (res)
        real(8), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(x, max(x, y), y /= y), x /= x)
    end function
    real(4) function fmax44(x, y) result (res)
        real(4), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(y, merge(x, max(x, y), y /= y), x /= x)
    end function
    real(8) function fmax84(x, y) result(res)
        real(8), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(dble(y), merge(x, max(x, dble(y)), y /= y), x /= x)
    end function
    real(8) function fmax48(x, y) result(res)
        real(4), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(dble(x), max(dble(x), y), y /= y), x /= x)
    end function
    real(8) function fmin88(x, y) result (res)
        real(8), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(x, min(x, y), y /= y), x /= x)
    end function
    real(4) function fmin44(x, y) result (res)
        real(4), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(y, merge(x, min(x, y), y /= y), x /= x)
    end function
    real(8) function fmin84(x, y) result(res)
        real(8), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(dble(y), merge(x, min(x, dble(y)), y /= y), x /= x)
    end function
    real(8) function fmin48(x, y) result(res)
        real(4), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(dble(x), min(dble(x), y), y /= y), x /= x)
    end function
end module

real(8) function code(w, l)
use fmin_fmax_functions
    real(8), intent (in) :: w
    real(8), intent (in) :: l
    code = exp(-w) * (l ** exp(w))
end function

Java

public static double code(double w, double l) {
	return Math.exp(-w) * Math.pow(l, Math.exp(w));
}

Python

def code(w, l):
	return math.exp(-w) * math.pow(l, math.exp(w))

Julia

function code(w, l)
	return Float64(exp(Float64(-w)) * (l ^ exp(w)))
end

MATLAB

function tmp = code(w, l)
	tmp = exp(-w) * (l ^ exp(w));
end

Wolfram

code[w_, l_] := N[(N[Exp[(-w)], $MachinePrecision] * N[Power[l, N[Exp[w], $MachinePrecision]], $MachinePrecision]), $MachinePrecision]

Initial Program: 99.3% accurate, 1.0× speedup

Math

\[\begin{array}{l} \\ e^{-w} \cdot {\ell}^{\left(e^{w}\right)} \end{array} \]

FPCore

(FPCore (w l) :precision binary64 (* (exp (- w)) (pow l (exp w))))

C

double code(double w, double l) {
	return exp(-w) * pow(l, exp(w));
}

Fortran

module fmin_fmax_functions
    implicit none
    private
    public fmax
    public fmin

    interface fmax
        module procedure fmax88
        module procedure fmax44
        module procedure fmax84
        module procedure fmax48
    end interface
    interface fmin
        module procedure fmin88
        module procedure fmin44
        module procedure fmin84
        module procedure fmin48
    end interface
contains
    real(8) function fmax88(x, y) result (res)
        real(8), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(x, max(x, y), y /= y), x /= x)
    end function
    real(4) function fmax44(x, y) result (res)
        real(4), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(y, merge(x, max(x, y), y /= y), x /= x)
    end function
    real(8) function fmax84(x, y) result(res)
        real(8), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(dble(y), merge(x, max(x, dble(y)), y /= y), x /= x)
    end function
    real(8) function fmax48(x, y) result(res)
        real(4), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(dble(x), max(dble(x), y), y /= y), x /= x)
    end function
    real(8) function fmin88(x, y) result (res)
        real(8), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(x, min(x, y), y /= y), x /= x)
    end function
    real(4) function fmin44(x, y) result (res)
        real(4), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(y, merge(x, min(x, y), y /= y), x /= x)
    end function
    real(8) function fmin84(x, y) result(res)
        real(8), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(dble(y), merge(x, min(x, dble(y)), y /= y), x /= x)
    end function
    real(8) function fmin48(x, y) result(res)
        real(4), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(dble(x), min(dble(x), y), y /= y), x /= x)
    end function
end module

real(8) function code(w, l)
use fmin_fmax_functions
    real(8), intent (in) :: w
    real(8), intent (in) :: l
    code = exp(-w) * (l ** exp(w))
end function

Java

public static double code(double w, double l) {
	return Math.exp(-w) * Math.pow(l, Math.exp(w));
}

Python

def code(w, l):
	return math.exp(-w) * math.pow(l, math.exp(w))

Julia

function code(w, l)
	return Float64(exp(Float64(-w)) * (l ^ exp(w)))
end

MATLAB

function tmp = code(w, l)
	tmp = exp(-w) * (l ^ exp(w));
end

Wolfram

code[w_, l_] := N[(N[Exp[(-w)], $MachinePrecision] * N[Power[l, N[Exp[w], $MachinePrecision]], $MachinePrecision]), $MachinePrecision]

Alternative 1: 99.3% accurate, 1.0× speedup

Math

\[\begin{array}{l} \\ e^{-w} \cdot {\ell}^{\left(e^{w}\right)} \end{array} \]

FPCore

(FPCore (w l) :precision binary64 (* (exp (- w)) (pow l (exp w))))

C

double code(double w, double l) {
	return exp(-w) * pow(l, exp(w));
}

Fortran

module fmin_fmax_functions
    implicit none
    private
    public fmax
    public fmin

    interface fmax
        module procedure fmax88
        module procedure fmax44
        module procedure fmax84
        module procedure fmax48
    end interface
    interface fmin
        module procedure fmin88
        module procedure fmin44
        module procedure fmin84
        module procedure fmin48
    end interface
contains
    real(8) function fmax88(x, y) result (res)
        real(8), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(x, max(x, y), y /= y), x /= x)
    end function
    real(4) function fmax44(x, y) result (res)
        real(4), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(y, merge(x, max(x, y), y /= y), x /= x)
    end function
    real(8) function fmax84(x, y) result(res)
        real(8), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(dble(y), merge(x, max(x, dble(y)), y /= y), x /= x)
    end function
    real(8) function fmax48(x, y) result(res)
        real(4), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(dble(x), max(dble(x), y), y /= y), x /= x)
    end function
    real(8) function fmin88(x, y) result (res)
        real(8), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(x, min(x, y), y /= y), x /= x)
    end function
    real(4) function fmin44(x, y) result (res)
        real(4), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(y, merge(x, min(x, y), y /= y), x /= x)
    end function
    real(8) function fmin84(x, y) result(res)
        real(8), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(dble(y), merge(x, min(x, dble(y)), y /= y), x /= x)
    end function
    real(8) function fmin48(x, y) result(res)
        real(4), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(dble(x), min(dble(x), y), y /= y), x /= x)
    end function
end module

real(8) function code(w, l)
use fmin_fmax_functions
    real(8), intent (in) :: w
    real(8), intent (in) :: l
    code = exp(-w) * (l ** exp(w))
end function

Java

public static double code(double w, double l) {
	return Math.exp(-w) * Math.pow(l, Math.exp(w));
}

Python

def code(w, l):
	return math.exp(-w) * math.pow(l, math.exp(w))

Julia

function code(w, l)
	return Float64(exp(Float64(-w)) * (l ^ exp(w)))
end

MATLAB

function tmp = code(w, l)
	tmp = exp(-w) * (l ^ exp(w));
end

Wolfram

code[w_, l_] := N[(N[Exp[(-w)], $MachinePrecision] * N[Power[l, N[Exp[w], $MachinePrecision]], $MachinePrecision]), $MachinePrecision]

Derivation

1. Initial program 99.7%

   \[e^{-w} \cdot {\ell}^{\left(e^{w}\right)} \]
2. Add Preprocessing
3. Add Preprocessing

Alternative 2: 97.7% accurate, 2.7× speedup

Math

\[\begin{array}{l} \\ \begin{array}{l} \mathbf{if}\;w \leq -0.68 \lor \neg \left(w \leq 1\right):\\ \;\;\;\;e^{-w}\\ \mathbf{else}:\\ \;\;\;\;\left(\left(-w\right) + 1\right) \cdot \ell\\ \end{array} \end{array} \]

FPCore

(FPCore (w l)
 :precision binary64
 (if (or (<= w -0.68) (not (<= w 1.0))) (exp (- w)) (* (+ (- w) 1.0) l)))

C

double code(double w, double l) {
	double tmp;
	if ((w <= -0.68) || !(w <= 1.0)) {
		tmp = exp(-w);
	} else {
		tmp = (-w + 1.0) * l;
	}
	return tmp;
}

Fortran

module fmin_fmax_functions
    implicit none
    private
    public fmax
    public fmin

    interface fmax
        module procedure fmax88
        module procedure fmax44
        module procedure fmax84
        module procedure fmax48
    end interface
    interface fmin
        module procedure fmin88
        module procedure fmin44
        module procedure fmin84
        module procedure fmin48
    end interface
contains
    real(8) function fmax88(x, y) result (res)
        real(8), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(x, max(x, y), y /= y), x /= x)
    end function
    real(4) function fmax44(x, y) result (res)
        real(4), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(y, merge(x, max(x, y), y /= y), x /= x)
    end function
    real(8) function fmax84(x, y) result(res)
        real(8), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(dble(y), merge(x, max(x, dble(y)), y /= y), x /= x)
    end function
    real(8) function fmax48(x, y) result(res)
        real(4), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(dble(x), max(dble(x), y), y /= y), x /= x)
    end function
    real(8) function fmin88(x, y) result (res)
        real(8), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(x, min(x, y), y /= y), x /= x)
    end function
    real(4) function fmin44(x, y) result (res)
        real(4), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(y, merge(x, min(x, y), y /= y), x /= x)
    end function
    real(8) function fmin84(x, y) result(res)
        real(8), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(dble(y), merge(x, min(x, dble(y)), y /= y), x /= x)
    end function
    real(8) function fmin48(x, y) result(res)
        real(4), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(dble(x), min(dble(x), y), y /= y), x /= x)
    end function
end module

real(8) function code(w, l)
use fmin_fmax_functions
    real(8), intent (in) :: w
    real(8), intent (in) :: l
    real(8) :: tmp
    if ((w <= (-0.68d0)) .or. (.not. (w <= 1.0d0))) then
        tmp = exp(-w)
    else
        tmp = (-w + 1.0d0) * l
    end if
    code = tmp
end function

Java

public static double code(double w, double l) {
	double tmp;
	if ((w <= -0.68) || !(w <= 1.0)) {
		tmp = Math.exp(-w);
	} else {
		tmp = (-w + 1.0) * l;
	}
	return tmp;
}

Python

def code(w, l):
	tmp = 0
	if (w <= -0.68) or not (w <= 1.0):
		tmp = math.exp(-w)
	else:
		tmp = (-w + 1.0) * l
	return tmp

Julia

function code(w, l)
	tmp = 0.0
	if ((w <= -0.68) || !(w <= 1.0))
		tmp = exp(Float64(-w));
	else
		tmp = Float64(Float64(Float64(-w) + 1.0) * l);
	end
	return tmp
end

MATLAB

function tmp_2 = code(w, l)
	tmp = 0.0;
	if ((w <= -0.68) || ~((w <= 1.0)))
		tmp = exp(-w);
	else
		tmp = (-w + 1.0) * l;
	end
	tmp_2 = tmp;
end

Wolfram

code[w_, l_] := If[Or[LessEqual[w, -0.68], N[Not[LessEqual[w, 1.0]], $MachinePrecision]], N[Exp[(-w)], $MachinePrecision], N[(N[((-w) + 1.0), $MachinePrecision] * l), $MachinePrecision]]

Derivation

1. Split input into 2 regimes

2. if w < -0.680000000000000049 or 1 < w

   1. Initial program 100.0%

      \[e^{-w} \cdot {\ell}^{\left(e^{w}\right)} \]
   2. Add Preprocessing
   3. Step-by-step derivation

      1. lift-*.f64 N/A

         \[\leadsto \color{blue}{e^{-w} \cdot {\ell}^{\left(e^{w}\right)}} \]

      2. lift-neg.f64 N/A

         \[\leadsto e^{\color{blue}{\mathsf{neg}\left(w\right)}} \cdot {\ell}^{\left(e^{w}\right)} \]

      3. lift-exp.f64 N/A

         \[\leadsto \color{blue}{e^{\mathsf{neg}\left(w\right)}} \cdot {\ell}^{\left(e^{w}\right)} \]

      4. lift-exp.f64 N/A

         \[\leadsto e^{\mathsf{neg}\left(w\right)} \cdot {\ell}^{\color{blue}{\left(e^{w}\right)}} \]

      5. lift-pow.f64 N/A

         \[\leadsto e^{\mathsf{neg}\left(w\right)} \cdot \color{blue}{{\ell}^{\left(e^{w}\right)}} \]

      6. *-commutative N/A

         \[\leadsto \color{blue}{{\ell}^{\left(e^{w}\right)} \cdot e^{\mathsf{neg}\left(w\right)}} \]

      7. pow-to-exp N/A

         \[\leadsto \color{blue}{e^{\log \ell \cdot e^{w}}} \cdot e^{\mathsf{neg}\left(w\right)} \]

      8. prod-exp N/A

         \[\leadsto \color{blue}{e^{\log \ell \cdot e^{w} + \left(\mathsf{neg}\left(w\right)\right)}} \]

      9. lower-exp.f64 N/A

         \[\leadsto \color{blue}{e^{\log \ell \cdot e^{w} + \left(\mathsf{neg}\left(w\right)\right)}} \]

      10. lower-fma.f64 N/A

          \[\leadsto e^{\color{blue}{\mathsf{fma}\left(\log \ell, e^{w}, \mathsf{neg}\left(w\right)\right)}} \]

      11. lower-log.f64 N/A

          \[\leadsto e^{\mathsf{fma}\left(\color{blue}{\log \ell}, e^{w}, \mathsf{neg}\left(w\right)\right)} \]

      12. lift-exp.f64 N/A

          \[\leadsto e^{\mathsf{fma}\left(\log \ell, \color{blue}{e^{w}}, \mathsf{neg}\left(w\right)\right)} \]

      13. lift-neg.f64 100.0

          \[\leadsto e^{\mathsf{fma}\left(\log \ell, e^{w}, \color{blue}{-w}\right)} \]

   4. Applied rewrites 100.0%

      \[\leadsto \color{blue}{e^{\mathsf{fma}\left(\log \ell, e^{w}, -w\right)}} \]

   5. Taylor expanded in w around inf

      \[\leadsto e^{\color{blue}{-1 \cdot w}} \]
   6. Step-by-step derivation

      1. mul-1-neg N/A

         \[\leadsto e^{\mathsf{neg}\left(w\right)} \]

      2. lift-neg.f64 100.0

         \[\leadsto e^{-w} \]

   7. Applied rewrites 100.0%

      \[\leadsto e^{\color{blue}{-w}} \]

   if -0.680000000000000049 < w < 1

   1. Initial program 99.5%

      \[e^{-w} \cdot {\ell}^{\left(e^{w}\right)} \]
   2. Add Preprocessing

   3. Taylor expanded in w around 0

      \[\leadsto e^{-w} \cdot \color{blue}{\ell} \]
   4. Step-by-step derivation

   5. Applied rewrites 97.1%

      \[\leadsto e^{-w} \cdot \color{blue}{\ell} \]

   6. Taylor expanded in w around 0

      \[\leadsto \color{blue}{\left(1 + -1 \cdot w\right)} \cdot \ell \]
   7. Step-by-step derivation

      1. mul-1-neg N/A

         \[\leadsto \left(1 + \left(\mathsf{neg}\left(w\right)\right)\right) \cdot \ell \]

      2. +-commutative N/A

         \[\leadsto \left(\left(\mathsf{neg}\left(w\right)\right) + \color{blue}{1}\right) \cdot \ell \]

      3. mul-1-neg N/A

         \[\leadsto \left(-1 \cdot w + 1\right) \cdot \ell \]

      4. lower-fma.f64 97.1

         \[\leadsto \mathsf{fma}\left(-1, \color{blue}{w}, 1\right) \cdot \ell \]

   8. Applied rewrites 97.1%

      \[\leadsto \color{blue}{\mathsf{fma}\left(-1, w, 1\right)} \cdot \ell \]
   9. Step-by-step derivation

      1. lift-fma.f64 N/A

         \[\leadsto \left(-1 \cdot w + \color{blue}{1}\right) \cdot \ell \]

      2. mul-1-neg N/A

         \[\leadsto \left(\left(\mathsf{neg}\left(w\right)\right) + 1\right) \cdot \ell \]

      3. lower-+.f64 N/A

         \[\leadsto \left(\left(\mathsf{neg}\left(w\right)\right) + \color{blue}{1}\right) \cdot \ell \]

      4. lift-neg.f64 97.1

         \[\leadsto \left(\left(-w\right) + 1\right) \cdot \ell \]

   10. Applied rewrites 97.1%

       \[\leadsto \left(\left(-w\right) + \color{blue}{1}\right) \cdot \ell \]
3. Recombined 2 regimes into one program.

4. Final simplification 98.1%

   \[\leadsto \begin{array}{l} \mathbf{if}\;w \leq -0.68 \lor \neg \left(w \leq 1\right):\\ \;\;\;\;e^{-w}\\ \mathbf{else}:\\ \;\;\;\;\left(\left(-w\right) + 1\right) \cdot \ell\\ \end{array} \]
5. Add Preprocessing

Alternative 3: 97.6% accurate, 2.9× speedup

Math

\[\begin{array}{l} \\ e^{-w} \cdot \ell \end{array} \]

FPCore

(FPCore (w l) :precision binary64 (* (exp (- w)) l))

C

double code(double w, double l) {
	return exp(-w) * l;
}

Fortran

module fmin_fmax_functions
    implicit none
    private
    public fmax
    public fmin

    interface fmax
        module procedure fmax88
        module procedure fmax44
        module procedure fmax84
        module procedure fmax48
    end interface
    interface fmin
        module procedure fmin88
        module procedure fmin44
        module procedure fmin84
        module procedure fmin48
    end interface
contains
    real(8) function fmax88(x, y) result (res)
        real(8), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(x, max(x, y), y /= y), x /= x)
    end function
    real(4) function fmax44(x, y) result (res)
        real(4), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(y, merge(x, max(x, y), y /= y), x /= x)
    end function
    real(8) function fmax84(x, y) result(res)
        real(8), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(dble(y), merge(x, max(x, dble(y)), y /= y), x /= x)
    end function
    real(8) function fmax48(x, y) result(res)
        real(4), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(dble(x), max(dble(x), y), y /= y), x /= x)
    end function
    real(8) function fmin88(x, y) result (res)
        real(8), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(x, min(x, y), y /= y), x /= x)
    end function
    real(4) function fmin44(x, y) result (res)
        real(4), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(y, merge(x, min(x, y), y /= y), x /= x)
    end function
    real(8) function fmin84(x, y) result(res)
        real(8), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(dble(y), merge(x, min(x, dble(y)), y /= y), x /= x)
    end function
    real(8) function fmin48(x, y) result(res)
        real(4), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(dble(x), min(dble(x), y), y /= y), x /= x)
    end function
end module

real(8) function code(w, l)
use fmin_fmax_functions
    real(8), intent (in) :: w
    real(8), intent (in) :: l
    code = exp(-w) * l
end function

Java

public static double code(double w, double l) {
	return Math.exp(-w) * l;
}

Python

def code(w, l):
	return math.exp(-w) * l

Julia

function code(w, l)
	return Float64(exp(Float64(-w)) * l)
end

MATLAB

function tmp = code(w, l)
	tmp = exp(-w) * l;
end

Wolfram

code[w_, l_] := N[(N[Exp[(-w)], $MachinePrecision] * l), $MachinePrecision]

Derivation

1. Initial program 99.7%

   \[e^{-w} \cdot {\ell}^{\left(e^{w}\right)} \]
2. Add Preprocessing

3. Taylor expanded in w around 0

   \[\leadsto e^{-w} \cdot \color{blue}{\ell} \]
4. Step-by-step derivation

5. Applied rewrites 98.1%

   \[\leadsto e^{-w} \cdot \color{blue}{\ell} \]
6. Add Preprocessing

Alternative 4: 76.7% accurate, 11.9× speedup

Math

\[\begin{array}{l} \\ \mathsf{fma}\left(\mathsf{fma}\left(-0.16666666666666666, w, 0.5\right) \cdot w - 1, w, 1\right) \cdot \ell \end{array} \]

FPCore

(FPCore (w l)
 :precision binary64
 (* (fma (- (* (fma -0.16666666666666666 w 0.5) w) 1.0) w 1.0) l))

C

double code(double w, double l) {
	return fma(((fma(-0.16666666666666666, w, 0.5) * w) - 1.0), w, 1.0) * l;
}

Julia

function code(w, l)
	return Float64(fma(Float64(Float64(fma(-0.16666666666666666, w, 0.5) * w) - 1.0), w, 1.0) * l)
end

Wolfram

code[w_, l_] := N[(N[(N[(N[(N[(-0.16666666666666666 * w + 0.5), $MachinePrecision] * w), $MachinePrecision] - 1.0), $MachinePrecision] * w + 1.0), $MachinePrecision] * l), $MachinePrecision]

Derivation

1. Initial program 99.7%

   \[e^{-w} \cdot {\ell}^{\left(e^{w}\right)} \]
2. Add Preprocessing

3. Taylor expanded in w around 0

   \[\leadsto e^{-w} \cdot \color{blue}{\ell} \]
4. Step-by-step derivation

5. Applied rewrites 98.1%

   \[\leadsto e^{-w} \cdot \color{blue}{\ell} \]

6. Taylor expanded in w around 0

   \[\leadsto \color{blue}{\left(1 + w \cdot \left(w \cdot \left(\frac{1}{2} + \frac{-1}{6} \cdot w\right) - 1\right)\right)} \cdot \ell \]
7. Step-by-step derivation

   1. +-commutative N/A

      \[\leadsto \left(w \cdot \left(w \cdot \left(\frac{1}{2} + \frac{-1}{6} \cdot w\right) - 1\right) + \color{blue}{1}\right) \cdot \ell \]

   2. *-commutative N/A

      \[\leadsto \left(\left(w \cdot \left(\frac{1}{2} + \frac{-1}{6} \cdot w\right) - 1\right) \cdot w + 1\right) \cdot \ell \]

   3. lower-fma.f64 N/A

      \[\leadsto \mathsf{fma}\left(w \cdot \left(\frac{1}{2} + \frac{-1}{6} \cdot w\right) - 1, \color{blue}{w}, 1\right) \cdot \ell \]

   4. lower--.f64 N/A

      \[\leadsto \mathsf{fma}\left(w \cdot \left(\frac{1}{2} + \frac{-1}{6} \cdot w\right) - 1, w, 1\right) \cdot \ell \]

   5. *-commutative N/A

      \[\leadsto \mathsf{fma}\left(\left(\frac{1}{2} + \frac{-1}{6} \cdot w\right) \cdot w - 1, w, 1\right) \cdot \ell \]

   6. lower-*.f64 N/A

      \[\leadsto \mathsf{fma}\left(\left(\frac{1}{2} + \frac{-1}{6} \cdot w\right) \cdot w - 1, w, 1\right) \cdot \ell \]

   7. +-commutative N/A

      \[\leadsto \mathsf{fma}\left(\left(\frac{-1}{6} \cdot w + \frac{1}{2}\right) \cdot w - 1, w, 1\right) \cdot \ell \]

   8. lower-fma.f64 81.8

      \[\leadsto \mathsf{fma}\left(\mathsf{fma}\left(-0.16666666666666666, w, 0.5\right) \cdot w - 1, w, 1\right) \cdot \ell \]

8. Applied rewrites 81.8%

   \[\leadsto \color{blue}{\mathsf{fma}\left(\mathsf{fma}\left(-0.16666666666666666, w, 0.5\right) \cdot w - 1, w, 1\right)} \cdot \ell \]
9. Add Preprocessing

Alternative 5: 73.6% accurate, 15.5× speedup

Math

\[\begin{array}{l} \\ \mathsf{fma}\left(0.5 \cdot w - 1, w, 1\right) \cdot \ell \end{array} \]

FPCore

(FPCore (w l) :precision binary64 (* (fma (- (* 0.5 w) 1.0) w 1.0) l))

C

double code(double w, double l) {
	return fma(((0.5 * w) - 1.0), w, 1.0) * l;
}

Julia

function code(w, l)
	return Float64(fma(Float64(Float64(0.5 * w) - 1.0), w, 1.0) * l)
end

Wolfram

code[w_, l_] := N[(N[(N[(N[(0.5 * w), $MachinePrecision] - 1.0), $MachinePrecision] * w + 1.0), $MachinePrecision] * l), $MachinePrecision]

Derivation

1. Initial program 99.7%

   \[e^{-w} \cdot {\ell}^{\left(e^{w}\right)} \]
2. Add Preprocessing

3. Taylor expanded in w around 0

   \[\leadsto e^{-w} \cdot \color{blue}{\ell} \]
4. Step-by-step derivation

5. Applied rewrites 98.1%

   \[\leadsto e^{-w} \cdot \color{blue}{\ell} \]

6. Taylor expanded in w around 0

   \[\leadsto \color{blue}{\left(1 + w \cdot \left(\frac{1}{2} \cdot w - 1\right)\right)} \cdot \ell \]
7. Step-by-step derivation

   1. +-commutative N/A

      \[\leadsto \left(w \cdot \left(\frac{1}{2} \cdot w - 1\right) + \color{blue}{1}\right) \cdot \ell \]

   2. *-commutative N/A

      \[\leadsto \left(\left(\frac{1}{2} \cdot w - 1\right) \cdot w + 1\right) \cdot \ell \]

   3. lower-fma.f64 N/A

      \[\leadsto \mathsf{fma}\left(\frac{1}{2} \cdot w - 1, \color{blue}{w}, 1\right) \cdot \ell \]

   4. lower--.f64 N/A

      \[\leadsto \mathsf{fma}\left(\frac{1}{2} \cdot w - 1, w, 1\right) \cdot \ell \]

   5. lift-*.f64 79.6

      \[\leadsto \mathsf{fma}\left(0.5 \cdot w - 1, w, 1\right) \cdot \ell \]

8. Applied rewrites 79.6%

   \[\leadsto \color{blue}{\mathsf{fma}\left(0.5 \cdot w - 1, w, 1\right)} \cdot \ell \]
9. Add Preprocessing

Alternative 6: 64.0% accurate, 22.0× speedup

Math

\[\begin{array}{l} \\ \begin{array}{l} \mathbf{if}\;w \leq -63:\\ \;\;\;\;\left(-w\right) \cdot \ell\\ \mathbf{else}:\\ \;\;\;\;\ell\\ \end{array} \end{array} \]

FPCore

(FPCore (w l) :precision binary64 (if (<= w -63.0) (* (- w) l) l))

C

double code(double w, double l) {
	double tmp;
	if (w <= -63.0) {
		tmp = -w * l;
	} else {
		tmp = l;
	}
	return tmp;
}

Fortran

module fmin_fmax_functions
    implicit none
    private
    public fmax
    public fmin

    interface fmax
        module procedure fmax88
        module procedure fmax44
        module procedure fmax84
        module procedure fmax48
    end interface
    interface fmin
        module procedure fmin88
        module procedure fmin44
        module procedure fmin84
        module procedure fmin48
    end interface
contains
    real(8) function fmax88(x, y) result (res)
        real(8), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(x, max(x, y), y /= y), x /= x)
    end function
    real(4) function fmax44(x, y) result (res)
        real(4), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(y, merge(x, max(x, y), y /= y), x /= x)
    end function
    real(8) function fmax84(x, y) result(res)
        real(8), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(dble(y), merge(x, max(x, dble(y)), y /= y), x /= x)
    end function
    real(8) function fmax48(x, y) result(res)
        real(4), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(dble(x), max(dble(x), y), y /= y), x /= x)
    end function
    real(8) function fmin88(x, y) result (res)
        real(8), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(x, min(x, y), y /= y), x /= x)
    end function
    real(4) function fmin44(x, y) result (res)
        real(4), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(y, merge(x, min(x, y), y /= y), x /= x)
    end function
    real(8) function fmin84(x, y) result(res)
        real(8), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(dble(y), merge(x, min(x, dble(y)), y /= y), x /= x)
    end function
    real(8) function fmin48(x, y) result(res)
        real(4), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(dble(x), min(dble(x), y), y /= y), x /= x)
    end function
end module

real(8) function code(w, l)
use fmin_fmax_functions
    real(8), intent (in) :: w
    real(8), intent (in) :: l
    real(8) :: tmp
    if (w <= (-63.0d0)) then
        tmp = -w * l
    else
        tmp = l
    end if
    code = tmp
end function

Java

public static double code(double w, double l) {
	double tmp;
	if (w <= -63.0) {
		tmp = -w * l;
	} else {
		tmp = l;
	}
	return tmp;
}

Python

def code(w, l):
	tmp = 0
	if w <= -63.0:
		tmp = -w * l
	else:
		tmp = l
	return tmp

Julia

function code(w, l)
	tmp = 0.0
	if (w <= -63.0)
		tmp = Float64(Float64(-w) * l);
	else
		tmp = l;
	end
	return tmp
end

MATLAB

function tmp_2 = code(w, l)
	tmp = 0.0;
	if (w <= -63.0)
		tmp = -w * l;
	else
		tmp = l;
	end
	tmp_2 = tmp;
end

Wolfram

code[w_, l_] := If[LessEqual[w, -63.0], N[((-w) * l), $MachinePrecision], l]

Derivation

1. Split input into 2 regimes

2. if w < -63

   1. Initial program 100.0%

      \[e^{-w} \cdot {\ell}^{\left(e^{w}\right)} \]
   2. Add Preprocessing

   3. Taylor expanded in w around 0

      \[\leadsto e^{-w} \cdot \color{blue}{\ell} \]
   4. Step-by-step derivation

   5. Applied rewrites 100.0%

      \[\leadsto e^{-w} \cdot \color{blue}{\ell} \]

   6. Taylor expanded in w around 0

      \[\leadsto \color{blue}{\left(1 + -1 \cdot w\right)} \cdot \ell \]
   7. Step-by-step derivation

      1. mul-1-neg N/A

         \[\leadsto \left(1 + \left(\mathsf{neg}\left(w\right)\right)\right) \cdot \ell \]

      2. +-commutative N/A

         \[\leadsto \left(\left(\mathsf{neg}\left(w\right)\right) + \color{blue}{1}\right) \cdot \ell \]

      3. mul-1-neg N/A

         \[\leadsto \left(-1 \cdot w + 1\right) \cdot \ell \]

      4. lower-fma.f64 26.6

         \[\leadsto \mathsf{fma}\left(-1, \color{blue}{w}, 1\right) \cdot \ell \]

   8. Applied rewrites 26.6%

      \[\leadsto \color{blue}{\mathsf{fma}\left(-1, w, 1\right)} \cdot \ell \]

   9. Taylor expanded in w around inf

      \[\leadsto \left(-1 \cdot \color{blue}{w}\right) \cdot \ell \]
   10. Step-by-step derivation

       1. mul-1-neg N/A

          \[\leadsto \left(\mathsf{neg}\left(w\right)\right) \cdot \ell \]

       2. lift-neg.f64 26.6

          \[\leadsto \left(-w\right) \cdot \ell \]

   11. Applied rewrites 26.6%

       \[\leadsto \left(-w\right) \cdot \ell \]

   if -63 < w

   1. Initial program 99.6%

      \[e^{-w} \cdot {\ell}^{\left(e^{w}\right)} \]
   2. Add Preprocessing

   3. Taylor expanded in w around 0

      \[\leadsto \color{blue}{\ell} \]
   4. Step-by-step derivation

   5. Applied rewrites 83.1%

      \[\leadsto \color{blue}{\ell} \]
3. Recombined 2 regimes into one program.
4. Add Preprocessing

Alternative 7: 63.6% accurate, 28.1× speedup

Math

\[\begin{array}{l} \\ \left(\left(-w\right) + 1\right) \cdot \ell \end{array} \]

FPCore

(FPCore (w l) :precision binary64 (* (+ (- w) 1.0) l))

C

double code(double w, double l) {
	return (-w + 1.0) * l;
}

Fortran

module fmin_fmax_functions
    implicit none
    private
    public fmax
    public fmin

    interface fmax
        module procedure fmax88
        module procedure fmax44
        module procedure fmax84
        module procedure fmax48
    end interface
    interface fmin
        module procedure fmin88
        module procedure fmin44
        module procedure fmin84
        module procedure fmin48
    end interface
contains
    real(8) function fmax88(x, y) result (res)
        real(8), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(x, max(x, y), y /= y), x /= x)
    end function
    real(4) function fmax44(x, y) result (res)
        real(4), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(y, merge(x, max(x, y), y /= y), x /= x)
    end function
    real(8) function fmax84(x, y) result(res)
        real(8), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(dble(y), merge(x, max(x, dble(y)), y /= y), x /= x)
    end function
    real(8) function fmax48(x, y) result(res)
        real(4), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(dble(x), max(dble(x), y), y /= y), x /= x)
    end function
    real(8) function fmin88(x, y) result (res)
        real(8), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(x, min(x, y), y /= y), x /= x)
    end function
    real(4) function fmin44(x, y) result (res)
        real(4), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(y, merge(x, min(x, y), y /= y), x /= x)
    end function
    real(8) function fmin84(x, y) result(res)
        real(8), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(dble(y), merge(x, min(x, dble(y)), y /= y), x /= x)
    end function
    real(8) function fmin48(x, y) result(res)
        real(4), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(dble(x), min(dble(x), y), y /= y), x /= x)
    end function
end module

real(8) function code(w, l)
use fmin_fmax_functions
    real(8), intent (in) :: w
    real(8), intent (in) :: l
    code = (-w + 1.0d0) * l
end function

Java

public static double code(double w, double l) {
	return (-w + 1.0) * l;
}

Python

def code(w, l):
	return (-w + 1.0) * l

Julia

function code(w, l)
	return Float64(Float64(Float64(-w) + 1.0) * l)
end

MATLAB

function tmp = code(w, l)
	tmp = (-w + 1.0) * l;
end

Wolfram

code[w_, l_] := N[(N[((-w) + 1.0), $MachinePrecision] * l), $MachinePrecision]

Derivation

1. Initial program 99.7%

   \[e^{-w} \cdot {\ell}^{\left(e^{w}\right)} \]
2. Add Preprocessing

3. Taylor expanded in w around 0

   \[\leadsto e^{-w} \cdot \color{blue}{\ell} \]
4. Step-by-step derivation

5. Applied rewrites 98.1%

   \[\leadsto e^{-w} \cdot \color{blue}{\ell} \]

6. Taylor expanded in w around 0

   \[\leadsto \color{blue}{\left(1 + -1 \cdot w\right)} \cdot \ell \]
7. Step-by-step derivation

   1. mul-1-neg N/A

      \[\leadsto \left(1 + \left(\mathsf{neg}\left(w\right)\right)\right) \cdot \ell \]

   2. +-commutative N/A

      \[\leadsto \left(\left(\mathsf{neg}\left(w\right)\right) + \color{blue}{1}\right) \cdot \ell \]

   3. mul-1-neg N/A

      \[\leadsto \left(-1 \cdot w + 1\right) \cdot \ell \]

   4. lower-fma.f64 69.3

      \[\leadsto \mathsf{fma}\left(-1, \color{blue}{w}, 1\right) \cdot \ell \]

8. Applied rewrites 69.3%

   \[\leadsto \color{blue}{\mathsf{fma}\left(-1, w, 1\right)} \cdot \ell \]
9. Step-by-step derivation

   1. lift-fma.f64 N/A

      \[\leadsto \left(-1 \cdot w + \color{blue}{1}\right) \cdot \ell \]

   2. mul-1-neg N/A

      \[\leadsto \left(\left(\mathsf{neg}\left(w\right)\right) + 1\right) \cdot \ell \]

   3. lower-+.f64 N/A

      \[\leadsto \left(\left(\mathsf{neg}\left(w\right)\right) + \color{blue}{1}\right) \cdot \ell \]

   4. lift-neg.f64 69.3

      \[\leadsto \left(\left(-w\right) + 1\right) \cdot \ell \]

10. Applied rewrites 69.3%

    \[\leadsto \left(\left(-w\right) + \color{blue}{1}\right) \cdot \ell \]
11. Add Preprocessing

Alternative 8: 57.1% accurate, 309.0× speedup

Math

\[\begin{array}{l} \\ \ell \end{array} \]

FPCore

(FPCore (w l) :precision binary64 l)

C

double code(double w, double l) {
	return l;
}

Fortran

module fmin_fmax_functions
    implicit none
    private
    public fmax
    public fmin

    interface fmax
        module procedure fmax88
        module procedure fmax44
        module procedure fmax84
        module procedure fmax48
    end interface
    interface fmin
        module procedure fmin88
        module procedure fmin44
        module procedure fmin84
        module procedure fmin48
    end interface
contains
    real(8) function fmax88(x, y) result (res)
        real(8), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(x, max(x, y), y /= y), x /= x)
    end function
    real(4) function fmax44(x, y) result (res)
        real(4), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(y, merge(x, max(x, y), y /= y), x /= x)
    end function
    real(8) function fmax84(x, y) result(res)
        real(8), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(dble(y), merge(x, max(x, dble(y)), y /= y), x /= x)
    end function
    real(8) function fmax48(x, y) result(res)
        real(4), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(dble(x), max(dble(x), y), y /= y), x /= x)
    end function
    real(8) function fmin88(x, y) result (res)
        real(8), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(x, min(x, y), y /= y), x /= x)
    end function
    real(4) function fmin44(x, y) result (res)
        real(4), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(y, merge(x, min(x, y), y /= y), x /= x)
    end function
    real(8) function fmin84(x, y) result(res)
        real(8), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(dble(y), merge(x, min(x, dble(y)), y /= y), x /= x)
    end function
    real(8) function fmin48(x, y) result(res)
        real(4), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(dble(x), min(dble(x), y), y /= y), x /= x)
    end function
end module

real(8) function code(w, l)
use fmin_fmax_functions
    real(8), intent (in) :: w
    real(8), intent (in) :: l
    code = l
end function

Java

public static double code(double w, double l) {
	return l;
}

Python

def code(w, l):
	return l

Julia

function code(w, l)
	return l
end

MATLAB

function tmp = code(w, l)
	tmp = l;
end

Wolfram

code[w_, l_] := l

Derivation

1. Initial program 99.7%

   \[e^{-w} \cdot {\ell}^{\left(e^{w}\right)} \]
2. Add Preprocessing

3. Taylor expanded in w around 0

   \[\leadsto \color{blue}{\ell} \]
4. Step-by-step derivation

5. Applied rewrites 64.2%

   \[\leadsto \color{blue}{\ell} \]
6. Add Preprocessing

Reproduce

herbie shell --seed 2025064 
(FPCore (w l)
  :name "exp-w (used to crash)"
  :precision binary64
  (* (exp (- w)) (pow l (exp w))))