Re: [avr-gcc-list] Time-efficient read of high byte in two-bytevariable

[Wallace White]

David Brown wrote:
> You are on the right track with unions, but I think it could be a little
> neater.
> First off, if you want to make a type that combines a 16-bit and two 8-bit
> data, you can do it a bit neater with anonymous unions:
>
> typedef union {
>     uint16_t w;
>     struct {
>         unsigned char lo, hi;
>     };
> } combo;
>
> It is best to avoid making your normal variables of type "combo" - you'll
> make a mess of the rest of your program just for the sake of a couple of
> clock cycles (changing "encPosDetentsNow ++" into "encPosDetentsNow.w ++" is
> not good).
>
> I found the most efficient method is to use a typecast from int to a pair of
> bytes (or a "combo").  However, as far as I know it is impossible to force
> the compiler to do that directly - we can only force a convertion of pointer
> types.  As long as you want to take the high or low byte of something that
> is not a local variable in a register, the best arrangement seems to be:
>
> typedef struct { unsigned char lo, hi; } bytePair;
> #define LoByte(x) ((*(bytePair*) &(x)).lo)
> #define HiByte(x) ((*(bytePair*) &(x)).hi)
>
> signed int encPosDetentsNow;
> UDR = HiByte(encPosDetentsNow);

Ah, that's perfect for what I'm doing--using a global variable that, in 
this one place, needs to have its high byte read. (And indeed, it 
compiles into the pair of LDS & OUT instructions I was hoping for.)

> This produces optimal code under all optomisations.  But it forces the
> compiler to generate addressable objects, so code like:
>
> void test(void) {
>     int encPosDetentsNow = 0x123;
>     UDR = HiByte(encPosDetentsNow);
> }
>
> is going to be awful - encPosDetentsNow gets put on the stack.  Of course,
> if gcc knew that registers are also directly addressable on the AVR, then
> that would not be a problem.

Good point. Thanks for the warning.

> So for local variables, the best code is using a local "combo" variable to
> do the convertion.  If you want to wrap that in an inline function, remember
> to use "static inline" to avoid an extra copy of the function being
> generated - "static inline" can generally be used in header files just like
> #define macros.  Also, don't worry about signed or unsigned integers - as
> long as you are not using comparisons, shifts, multiplies or divides, you
> can freely mix them.

Right. Thanks very much, David.

- Wallace