From: W van den Bos <redmine@...>
Date: 2010-12-01T21:46:11+09:00
Subject: [ruby-core:33503] [Ruby 1.9-Bug#4105][Open] Bug in numeric.c for Borland compiler

Bug #4105: Bug in numeric.c for Borland compiler
http://redmine.ruby-lang.org/issues/show/4105

Author: W van den Bos
Status: Open, Priority: Low
Category: core, Target version: 1.9.1
ruby -v: ruby 1.9.1p376 (2009-12-07 revision 26041) [i386-mswin32]

There is a potential bug in Init_Numeric()

#elif defined(__BORLANDC__)
    /* Turn off floating point exceptions for overflow, etc. */
    _control87(MCW_EM, MCW_EM);
#endif

This call wil set the coprocessor in a state that double floating point will become single float when an exception is thrown

See http://qc.embarcadero.com/wc/qcmain.aspx?d=32483 for the workaround.

The next code should eleminate this

#elif defined(__BORLANDC__)
    /* Turn off floating point exceptions for overflow, etc. */
    _control87(MCW_EM, MCW_EM);
    _control87(_control87(0,0),0x1FFF);
#endif

This will only appear when borland and for example microsoft dll's are mixed.
This can happend when OpenGl dll's are used

Btw. This source is also present in 1.8

I found this problem because I created a borland dll calling the ruby mswin32 version dll
It generated strange errors. A 'puts 0.0' generated a devide by zero in the ruby sprintf implementation
Putting the two _control87 calls in my dll solved this.
I found the same construction in the Ruby source without the second _control87 call :-)


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