Bug? Fixnum#** is slower than Bignum#**

[murphy]

(original text by Wolfgang Nádasi-Donner)
The German Ruby forum (http://rubyforen.de) recognized a strange
behaviour of the Fixnum class.

In exponentation, Fixnum is much slower than Bignum, which lets us
assume that some kind of a problem is
in the implementation.

Example:

require 'benchmark'

N = 10**6
B = 65891264863465298234965902602612457060348499377
F = 4628

Benchmark.bmbm(10) do |bm|
bm.report('Bignum') do N.times do
B ** 0.5
end end

bm.report('Fixnum') do N.times do
F ** 0.5
end end
end
user system total real

Bignum 3.856000 0.000000 3.856000 ( 3.946000)
Fixnum 7.861000 0.020000 7.881000 ( 8.081000)

Further investigations lead to the responsible C code:

// in numeric.c
static VALUE
fix_pow(x, y)
VALUE x, y;
{
if (FIXNUM_P(y)) {
long a, b;

b = FIX2LONG(y);
if (b == 0) return INT2FIX(1);
if (b == 1) return x;
a = FIX2LONG(x);
if (b > 0) {
return rb_big_pow(rb_int2big(a), y);
}
return rb_float_new(pow((double)a, (double)b));
}
return rb_num_coerce_bin(x, y); // <-- this take a lot of time
}

// ...versus...

// in bignum.c
VALUE
rb_big_pow(x, y)
VALUE x, y;
{
double d;
long yy;

if (y == INT2FIX(0)) return INT2FIX(1);
switch (TYPE(y)) {
case T_FLOAT: // <-- this is fast
d = RFLOAT(y)->value;
break;

case T_BIGNUM:
...

case T_FIXNUM:
...

default:
return rb_num_coerce_bin(x, y);
}
return rb_float_new(pow(rb_big2dbl(x), d));
}

If it's a feature, maybe someone can explain.
If it's a bug, maybe someone can fix

German readers may be interested in the original thread:
http://www.rubyforen.de/topic,749.html