Correction factor in computing exp()?

I have recently developed an interest in Cody's approximations of
erf() and erfc(). Here is a snippet of the Fortran code from
http://www.netlib.org/specfun/erf:

YSQ = AINT(Y*SIXTEN)/SIXTEN
DEL = (Y-YSQ)*(Y+YSQ)
RESULT = EXP(-YSQ*YSQ) * EXP(-DEL) * RESULT

The point of this is to simply compute exp(-y*y) and multiply to the
result of a rational approximation calculation. Instead of doing it
straight ahead, the code multiplies y by 16, takes the integer part,
multiplies the 16 back, calls it YSQ, then computes some curious factor
DEL. Then the routine takes exp() of these separated pieces and
multiplies the whole thing together.

Does anyone know the computational point of doing this? Is or was there
some bug in the Fortran implementation of exp() that required some sort
of correction? Someone familiar with the GSL in C advises me that this
turns up in that code too. I have Cody's original 1969 paper on which
this code is based (indeed, Cody is the author of the code too), and
there is no mention of this little adjustment, or correction, or
whatever it is.

Many thanks in advance. This is driving me nuts with head
scratching....

Les

.