Fortran double precision scalars get evaluated in single precision

When double precision expressions are assigned to double precision variables, the result has roundoff error at the single-precision level.

E.g. for a variable declared as real*8 a:

• a = 1.1 -> roundoff at single-precision level (as expected)
• a = 11 followed by a = a / 10 -> roundoff at double-precision level (as expected)
• a = 1.1d0 -> roundoff at single-precision level (expected roundoff at double-precision level instead)

Some simple test cases to demonstrate each of these scenarios are at my forked repo here.