Bug #7142: mingw TestFloat#test_round_with_precision failure

• Status: Closed
• Assignee: -
• Target version: 2.0.0
• ruby -v: ruby 2.0.0dev (2012-10-09 trunk 37127) [i386-mingw32]

[h.shirosaki (Hiroshi Shirosaki)]

=begin
Trunk ruby on Windows XP x86 with mingw-w64 gcc 4.7.2 has the following test failure.

1. Failure:
   test_round_with_precision(TestFloat) [C:/Users/Worker/Jenkins/workspace/ruby-trunk-x86-build/test/ruby/test_float.rb:389]:
   <1.0e-300> expected but was
   <9.999999999999994e-301>.

I get the failure with make test-all TESTS="-qv -n test_round_with_precision".

But it's strange that I don't get the failure with make test-all TESTS="-qv ruby/test_float.rb".
And This failure doesn't occur on Win7.

Floating-point precision seems to be changed for some reason. Mingw-w64 has own pow() implementation and the precision of pow(10, ndigits) is not proper.
I found calling (({_controlfp(_PC_64, _MCW_PC)})) before pow() improves precision and fixes this failure.

I'll commit this patch if there is no other better fix.

Index: include/ruby/win32.h

--- include/ruby/win32.h (revision 37136)
+++ include/ruby/win32.h (working copy)
@@ -764,7 +764,7 @@
} /* extern "C" { */
#endif

-#ifdef MINGW64
+#if defined(MINGW64)
/*

• Use powl() instead of broken pow() of x86_64-w64-mingw32.
• This workaround will fix test failures in test_bignum.rb,
  @@ -775,6 +775,24 @@
  {
  return powl(x, y);
  }
  +#elif defined(__MINGW64_VERSION_MAJOR)
  +/*

• • Set floating point precision for pow() of mingw-w64 x86.
• • With default precision the result is not proper on WinXP.
• */
  +static inline double
  +rb_w32_pow(double x, double y)
  +{
• double r;
• unsigned int default_control = _controlfp(0, 0);
• _controlfp(_PC_64, _MCW_PC);
• r = pow(x, y);
• /* Restore setting */
• _controlfp(default_control, _MCW_PC);
• return r;
  +}
  +#endif
  +#if defined(__MINGW64_VERSION_MAJOR) || defined(MINGW64)
  #define pow rb_w32_pow
  #endif

=end

• Status changed from Open to Closed
• % Done changed from 0 to 100

This issue was solved with changeset r37168.
Hiroshi, thank you for reporting this issue.
Your contribution to Ruby is greatly appreciated.
May Ruby be with you.

---

win32.h: set floating point precision for pow()

• include/ruby/win32.h (rb_w32_pow): set floating point precision
  for mingw-w64 x86 pow(). This improves the precision of pow() on
  Windows XP for TestFloat#test_round_with_precision failure.
  [ruby-core:47911] [Bug #7142]