Issue #7724 has been updated by mame (Yusuke Endoh).

Assignee changed from mame (Yusuke Endoh) to marcandre (Marc-Andre Lafortune)

Thanks Marc-Andre!

But I'm very sorry that I'll have no enough time to pursue this issues.
Could you please create a patch for C implementation?

Notice that the current implementation does NOT assume that the internal
representation of double is IEEE 754.  Please do not depend on it but
use C's constants about double, such as FLT_RADIX and DBL_MANT_DIG.

-- 
Yusuke Endoh <mame / tsg.ne.jp>
----------------------------------------
Bug #7724: 6 bugs with Range#bsearch over floats
https://bugs.ruby-lang.org/issues/7724#change-35546

Author: marcandre (Marc-Andre Lafortune)
Status: Assigned
Priority: Normal
Assignee: marcandre (Marc-Andre Lafortune)
Category: core
Target version: 2.0.0
ruby -v: r38825


Take the following code, with from, to and search all Integer or all Float:

  (from...to).bsearch{|f| f >= search}

I expected it to:
0) never yield and return nil if range is empty, i.e. from >= to ("trivial test")
1) never yield more than 65 times for floats
   or Math.log(to-from, 2)+1 for Integer ("log test")
2) return search if from <= search < to, nil otherwise ("coverage test")
3) never yield the same `f` ("uniqueness test")
4) if range is exclusive, never yield `to` nor return `to`; if range is inclusive and block returns false always yield `to` ("end of range test")
5) never yield a number < from or > to ("out of range test")

These conditions are all respected for Integers but all can be broken for Floats
Test 0, 1, 3, 4 & 5 fail even for small ordinary float values, while test 2 requires some larger floats, say 1e100 to fail.
For example bsearch can yield over 2000 times (instead of at most 65).


These tests and a correct Ruby implementation of bsearch can be found here: https://github.com/marcandre/backports/compare/marcandre:master...marcandre:bsearch
My implementation also passes the MRI tests.


-- 
http://bugs.ruby-lang.org/