You can also look into _controlfp_s to fiddle with the SSE floating point behaviour.

Are you actually using SSE instructions in your filter though? Setting the SSE register to round to zero won't help if you're not...

If you're using regular floats (not SSE value), then you're probably looking for something like fesetenv to set the flush/round mode...
This will depend on your compiler...
e.g. on MSVC you'd use https://msdn.microsoft.com/en-us/library/e9b52ceh.aspx

To verify your hypthesis that denormals are actually your problem, you can check for them like this:

#include <cmath>
if ( std::fpclassify( your_float_here ) == FP_SUBNORMAL ) printf("found a denormal!\n");

If your compiler is very old, it might provide it's own custom fpclassify function. Check it's documentation.

Also read this:
http://stackoverflow.com/questions/2219244/dealing-with-floating-point-exceptions

You probably want to enable the "underflow exception", which will trigger if any of your math ever produces a denormal float. You can then fix your code so that it doesn't happen.
In my experience with games (where performance is also a large concern), this is often used to find causes of denormal numbers and fix that code before shipping the game.

I don't know anything about Borland but I see no reason why denormalized numbers would be a problem.

CPU's often decide to run 100 to 1000 times slower when you give them a denormalized float. Their existence in your data sets can be a massive performance pitfall.

[edit]

p.s. can everyone in this thread please try not to be so damn rude to each other. Yes you. You're being rude. Stahp.