std.math.pow produces incorrect results when used with std.parallelism

[Johannes Pfau]

Am Sun, 02 Jun 2013 07:17:12 +0200
schrieb "finalpatch" <fengli at gmail.com>:

> the following code compiled with GDC shows the bug
> (I'm using the gcc 4.8 windows mingw32 version from 
> https://bitbucket.org/goshawk/gdc/downloads)
> 
> import std.math;
> import std.stdio;
> import std.range;
> import std.parallelism;
> 
> void main()
> {
>      enum n = 1000;
>      float step = 1.0f/n;
> 
>      float[n] results_singlethreaded;
>      float[n] results_multithreaded;
> 
>      foreach(i; iota(n))
>          results_singlethreaded[i] = pow(step * i, 0.4545f);
>      foreach(i; parallel(iota(n)))
>          results_multithreaded[i] = pow(step * i, 0.4545f);
>      foreach(i; iota(n))
>          assert(results_singlethreaded[i] == 
> results_multithreaded[i]);
> }
> 
> ~/Desktop $ ./test.exe
> core.exception.AssertError at test.d(26): Assertion failure
> 
> The same code passes without assertion when compiled with DMD.

Please file a bug report on bugzilla.gdcproject.org

I can't reproduce this on linux/posix. Could you use feqrel to check
the number of equal bits of the single/multithreaded results?