• John Jameson (19/19) Apr 04 2006 Hi All

John Jameson <John_member pathlink.com> writes:

Hi All

Here's a nasty little bug in the runtime library:

In the DM C++ startup code, the 80x87 control word is intialized to 0x137F such
that arithmetic is performed with 80 bit precision.

But when a new thread is started, it is initialized to 0x027F which specifies
only 64 bit precision. This is probably a default set by Windows. 

So the results of floating opertions can differ according to whether they are
performed in the main application thread, or in a worker thread!!!

The code in thread.c seems to initialize a t->f_env structure, but this isn't
the same as actually loading the value in the 80x87 control word; I'm a little
loath to modify this code as I don't know exactly the purpose of t->fenv i.e.
when it is loaded into the FCPU.

Meanwhile you can do

WORD newcw 0x137F
asm fldcw newcw;

at the start of worker threads to correctly initialize them for 80 bit
precision.

Best Regards 
John