The previous code used casts of address of int to float pointer.
This is undefined behavior in C and may cause a compiler to
optimize code incorrectly.
I have replaced it with a union of int and float types,
eliminating the cast problem.
Signed-off-by: Lars-Ove Karlsson <lars-ove.karlsson@iar.com>
Changed initial guess from a simple x/3 to dividing the exponent by 2.
This makes large or small numbers like 10e10 and 01e-10 converge in a few
loops.
Added a loop counter to ensure that the algorithm breaks out of the loop in
the case that the algorithm doesn't converge (toggling between two
numbers).
Added test cases for sqrt and sqrtf in libc. Tested with a range of numbers
between 10e10 and 10e-10. Verify good accuracy in test case.
Closes: #55962
Signed-off-by: Lawrence King <lawrencek52@gmail.com>
The lmp90100_evb sample included an implementation of double sqrt, and the
on_off_level_lighting_vnd_app sample included an implementation of float
sqrtf. Move that code into minimal libc instead of requiring applications
to hand-roll their own version.
Signed-off-by: Keith Packard <keithp@keithp.com>
