Question

All the info I can find in documentation and the web for -ffinite-math-only is “Allow optimizations for floating-point arithmetic that assume that arguments and results are not NaNs or +-Infs.” This does not seem forthcoming to me. Does anyone know exactly what those optimizations are?
Thanks

Answer 1

Lots of little things can be optimized under that assumption, like:

• x == x –> 1
• x * 1 –> x
• x >= y –> !(x < y) and similar.
• x/x –> 1 if the compiler can prove x != 0.
• it may allow a compiler to use hardware max/min instructions for expressions like x > y ? x : y.
• … lots more

You often see this assumption together with assumptions like “sign of zero doesn’t matter”, which then allows things like:

• x - x –> 0
• 0 / x –> 0
• x * 0 –> 0