Best Linux distribution for a MiniITX server 679

The x87 internal eight level floating point hardware stack is 80 bits wide, while otherwise the FP memory access is 32 or 64 bits wide. When you load the 32 or 64 bit value from memory it is converted to 80 bits and pushed on the hardware stack. The intermediate calculations are performed on the stack in the 80 bit format and when the expression has been evaluated, the 80 bit value is popped from the stack, converted to 32 or 64 bits and stored into memory.

The nasty things with many FP formats is that many use hidden bit normalisation, sign-magnitude mantissa and biased exponent offsets (and even different bias for 32 and 64 bits in IEEE FP). It would be natural to use the same internal representation for different FP types, use 2's complement exponent notation instead of offset notation for floating point, get rid of the hidden bit normalisation and have some extra bits for various rounding modes.

When doing the conversion only when loading and storing does not cost a lot and with the 8 level stack, you can have 7-8 levels of nested parenthesis without any reorganisations and the intermediate calculations are performed in the internal precision.

However, if the arithmetic expression contains a call for an external floating point function, the compiler does not know how much stack locations will be used by the external function, to avoid stack overflows, the intermediate values in the hardware stack may have to be saved before the call, which can cause rounding problems. Paul