in certain cases its really slow e.g. a x / 2 as discovered from #5650

TIMES = (ENV['TIMES'] || 10_000_000).to_i

Benchmark.bmbm do |bm|
  bm.report("/2 flt") {
    rad_per_deg = Math::PI/180
    loc1 = [46.3625, 15.114444]
    loc2 = [46.055556, 14.508333]
    val = ((loc2[1]-loc1[1]) * rad_per_deg)
    TIMES.times { val / 2 }
  }

  bm.report("/2 dec") {
    rad_per_deg = Math::PI/180
    loc1 = [BigDecimal('46.3625'), BigDecimal('15.114444')]
    loc2 = [BigDecimal('46.055556'), BigDecimal('14.508333')]
    val = ((loc2[1]-loc1[1]) * rad_per_deg)
    TIMES.times { val / 2 }
  }

  bm.report(" 1/1280") {
    one = BigDecimal(1)
    div = BigDecimal(1280)
    TIMES.times { one / div }
  }
end

here the "/2 dec" case is very slow - on JRuby 9.2.6 taking 400s, while MRI takes 10s :

                 user     system      total        real
/2 flt       0.670000   0.030000   0.700000 (  0.351421)
/2 dec     444.870000   0.930000 445.800000 (434.352299)
 1/1280     38.520000   0.100000  38.620000 ( 36.827996)

/2 flt       0.433256   0.000000   0.433256 (  0.433262)
/2 dec       9.910286   0.000000   9.910286 (  9.910299)
 1/1280      7.118558   0.000000   7.118558 (  7.118567)

... with this patch JRuby gets close to MRI's performance (over 30x improvement for the above / 2 case) :

/2 flt       0.740000   0.020000   0.760000 (  0.379111)
/2 dec      14.560000   0.240000  14.800000 ( 13.860334)
 1/1280      7.760000   0.030000   7.790000 (  6.865561)

Should be noted that only division (backed by java.math.BigDecimal#divide(another, context)) seemed slow, other decimal operations are either faster on on par with MRI. The 'new' division algorithm is based on Android's sources (ASF2 license) with a few tweaks and a major (in terms of performance) change of not down scaling exact results (remainder 0). Leaving that off to java.math.BigDecimal(BigInteger, int, MathContext) which will reduce the scale - n (doing BigInteger div 10^n) faster than we can using public API.