VC++5 changes the floating point unit to use less resolution it appears.
In the 3DFX 3D chip programming  newsgroups they talk about how VC++5
reduces Floating point bit resolution to something like 24bits and has to be
fpu has to manualy be told to use more resolution in assembly code or
by some other means.
This is just speculation but it appears that some 3D programmers have
are getting the same not so accurate results and other float related problems.

----------
> From: Richard Dobson 
> On compiling the file lptrkfns.c I hit an Internal Compiler Error, which
> seems to relate to the floating-point optimiser. A get-around is to go
> to the Code-Generation part of the Setting Dialogs, and use Custom settings
> rather than the blanket 'Optimize for Speed'.

> More insidiously, when compiling the standard CARL implementation of the 
> Phase Vocoder (our CDP port of the original Fortran code to C), the 
> compilation completed normally, but pvoc generated an obviously incorrect
> output. Changing the Settings as described above restored pvoc to correct
> operation.