When compiling C/C++ code for the TMS320C28x only, the long double floating point type is now IEEE 64-bit double precision. No other floating-point types have changed formats. C28x floating point types are:
| Type | Format |
|---|---|
| float | IEEE 32-bit single precision |
| double | IEEE 32-bit single precision |
| long double | IEEE 64-bit double precision |
It is recommended that 32-bit floating point values be declared as float, not as double.
When you initialize a long double to a constant, you must use the l or L suffix. The constant is treated as a double type without the suffix and the run-time support double-to-long conversion routine is called for the initialization. This could result in the loss of precision. For example:
long double a = 12.34L; /* correctly initializes to double precision */
long double b = 56.78; /* converts single precision value to double precision */
The formatting rules for long doubles in C I/O require a capital ’L’ in the format string. For example:
printf("%Lg", 1.23L);
printf("%Le", 3.45L);
In response to the change in the long double type to 64-bit IEEE double-precision format, the C28x calling conventions have changed.
All long double arguments are passed by reference. A long double return value is returned by reference. The first two long double arguments will pass their addresses in XAR4 and XAR5. All other long double arguments will have their addresses passed on the stack.
If a function returns a long double, the function making that call will place the return address in XAR6. For example:
long double foo(long double a, long double b, long double c)
{
long double d = a + b + c;
return d;
}
long double a = 1.2L;
long double b = 2.2L;
long double c = 3.2L;
long double d;
void bar()
{
d = foo(a, b, c);
}
In function bar(), at the call to foo(), the register values are:
| Register | Equals |
|---|---|
| XAR4 | The address of a |
| XAR5 | The address of b |
| *−SP[2] | The address of c |
| XAR6 | The address of d |
The run-time-support library has been updated to include the necessary long double arithmetic operations and conversion functions. All C28x floating-point run-time-support routines have had their names updated. For example, a previous call to the floating point add routine was:
LCR F$$ADD
This has been updated to:
LCR FS$$ADD ; single-precision add
LCR FD$$ADD ; double-precision add
Any C28x routine that calls floating-point arithmetic or conversion routines will need to be recompiled.