While playing with single-precision floating-point arith in a T1040, it seemed to me that my algorithms were a little slower when subnormal values were involved. However, I could not find any specifics on subnormal instruction timing in the e5500 reference manual - is this information available anywhere else? Or timing is the same for normalized or subnormal numbers and there is a flaw in my experiment?
For FPU it says, Counts extra cycles delay due to denormalized inputs. If there is one, this is incremented 4 times, Two operands increments it 5 times. This shows the real penalty due to denorms, not just how often they occur.
