grok_bin_oct_hex: Minor improvements to precision, speed

[khwilliamson]

These commits

• add a calculation before the loop that saves calculations each iteration
• replace floating multiplications each loop iteration with an integer increment, plus after the loop a call to ldexp

One question I have is can I just assume that Perl_ldexp exists and not have a fallback to Perl_pow? I think C99 requires all versions we need to be available, but it's not clear to me that perl.h guarantees one.

• This set of changes does not require a perldelta entry.

[t-a-k]

I'm suspicious whether replacing floating point multiplication to ldexp improves precision and/or speed:

• in this case multiplicand (factor) is exact power of 2, so multiplication is always accurate without losing precision, as long as floating point is radix-2ⁿ.
• many modern CPUs can do floating point multiplications with single instruction, while ldexp typically requires a library call or lengthy inlined code. So I'm afraid that this change may slow things down.

[tonycoz]

I think t-a-k is right.

[tonycoz]

Though I could see the ldexp being faster with -Dusequadmath or other soft floats (which are unlikely in modern general computing systems), but with doubles and probably long doubles the multiplication will likely be faster.

[khwilliamson]

I'm unsure of this, but I thought that as the mantissa gets overflowed the ldexp could save one or two floating point operations that each lose precision.

[tonycoz]

I'm unsure of this, but I thought that as the mantissa gets overflowed the ldexp could save one or two floating point operations that each lose precision.

Multiplying by a positive power of two (after conversion to double that's an exponent and a implicit 1 mantissa) doesn't lose any more precision than the step of converting a 64-bit integer to a 53-bit mantissa in the first place.