Bug #13078: with Integer a and b, a.fdiv(b) sometimes inaccurate than Rational(a, b).to_f

• Status: Closed
• Assignee: -
• Target version: -
• Backport: 2.2: REQUIRED, 2.3: REQUIRED, 2.4: DONE

[metanest (Makoto Kishimoto)]

For example,

a = 3
b = 0x20_0000_0000_0001
p a.fdiv(b)
p Rational(a, b).to_f

#=>
3.3306690738754696e-16
3.330669073875469e-16

I'm hacking to solve this problem, see diff view.
https://github.com/ruby/ruby/compare/trunk...metanest:fdiv_spike

[wanabe (_ wanabe)]

The accuracy of Rational(a, b).to_f seems to be due to reduction of fraction.
How about using Rational object, or its internal function f_gcd() and f_idiv()?

[nobu (Nobuyoshi Nakada)]

I thought the reduction would be slower, but actually not significant.

$ ruby -rbenchmark -e 'a = 3; b = -0x20_0000_0000_0001; n=10_000_000; Benchmark.bm(13) {|x|x.report("Integer#fdiv"){n.times{a.fdiv(b)}}; x.report("Rational#to_f"){n.times{Rational(a, b).to_f}}}'
                    user     system      total        real
Integer#fdiv    3.420000   0.000000   3.420000 (  3.446264)
Rational#to_f   3.450000   0.010000   3.460000 (  3.465592)

[nobu (Nobuyoshi Nakada)]

• Status changed from Open to Closed

Applied in changeset r57227.

---

numeric.c: reduce fdiv

• numeric.c (rb_int_fdiv_double): reduce first for more precise
  result. [ruby-core:78886] [Bug #13078]

[metanest (Makoto Kishimoto)]

pull requested
https://github.com/ruby/ruby/pull/1508

[naruse (Yui NARUSE)]

• Backport changed from 2.2: UNKNOWN, 2.3: UNKNOWN, 2.4: UNKNOWN to 2.2: UNKNOWN, 2.3: UNKNOWN, 2.4: DONE

ruby_2_4 r57840 merged revision(s) 57227.

[nagachika (Tomoyuki Chikanaga)]

• Backport changed from 2.2: UNKNOWN, 2.3: UNKNOWN, 2.4: DONE to 2.2: UNKNOWN, 2.3: REQUIRED, 2.4: DONE

[usa (Usaku NAKAMURA)]

• Backport changed from 2.2: UNKNOWN, 2.3: REQUIRED, 2.4: DONE to 2.2: REQUIRED, 2.3: REQUIRED, 2.4: DONE