This fixes problems where printing floating point numbers less than
10^-6 output as "0.0". Such numbers now use using scientific notation
and preserve the value's actual floating point representation.
The modp_dtoa/modp_dtoa2 functions aren't capable of handling double
values larger than INT_MAX and fallback on using sprintf() in that
situation. Previously, the format string to that sprintf() was "%e",
defaulting to a precision of 6, which is already too few digits to
represent a number known to be larger than INT_MAX. Now, an sprintf()
is still performed for values larger than INT_MAX and still uses a
scientific notation format, but in a way that uses as many decimal
digits as needed to preserve information.
Those methods already had a fallback to use sprintf() for large values
except:
* The check-for-large-value was unnecessarily done after many
operations that aren't relevant to the check and those operations can
result in a conversion overflow (undefined behavior).
* The check-for-large-value was using the literal value for a
32-bit INT_MAX instead of just using INT_MAX. For a platform where
`int` is less than 32-bits, the same conversion overflow from the
previous point could still occur (undefined behavior).
* The check-for-large-value was not inclusive of INT_MAX.
In a case where the conversion of INT_MAX itself to a double
can't be represented exactly, it's implementation-defined whether
the closest higher or closest lower representable-value is selected.
If the higher value is selected, then a `double` value comparing equal
to INT_MAX-as-converted-to-double would cause an overflow of an `int`
upon conversion (undefined behavior).
* origin/fastpath:
Remove automatic use of gperftools on non-Linux systems.
Fix uninitialized value for 'is_partial' in TCP analyzer.
parse 64-bit consts correctly.
on 32-bit machines only unsigned long longs are 64-bits long. Not just unsigned longs...
