Using IEEE 754 algorithm i assume, that it can be implemented in a branchless way.

But how does CPU determine special cases (Reserved Exponent values):

Exponent Significand is
11111111 000000000... Inf
11111111 000001000... NaN
00000000 000000000... 0
00000000 000001000... Subnormal

Without any tricks which i don't know / understand it should be expensive.


It was too big for comment, so editing source question:

I'm currently reading ARM spec/documentation https://developer.arm.com/documentation/ddi0403/d/Application-Level-Architecture/Application-Level-Programmers--Model/The-optional-Floating-point-extension/Floating-point-data-types-and-arithmetic?lang=en#BEIBFIBJ and especially interesting part is FPUnpack() pseudocode. If i understand correctly, CPU doesn't have intrinsics / instructions for special cases, but the compiler, that produce machine code should consider to validate result from registers (FVP for old ones and NEON for new). The example i found from ARM team https://github.com/ARM-software/ComputeLibrary/blob/8f587de9214dbc3aee4ff4eeb2ede66747769b19/include/CL/cl_half.h#L135. Am i right ?


Nope, i'm not right, according to this answer https://stackoverflow.com/questions/61646510/how-does-the-cpu-cast-a-floating-point-x87-i-think-value.
But it's still the question how does CPU registers handle special cases...


1 Answer 1


Not really. There are eight bits to check in two numbers usually to see there is a special case. For infinity/NaN you don’t need to do the normal operation anymore, so the rest is not time critical. More time critical is normalising the mantissa for a zero exponent.

But there are processors that assumed these cases are so rare that many can be handled by an interrupt (not directly in hardware, but by executing code).


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.