Faster than bit shifting macros?

[RbgoWeb]

I'm experimenting with macros that can read and write parts of basic types. Normal bit shifts and AND masks are used a lot. But I thought that with help of a little structure operants are only casted and assigned and things may run a little faster when loads of data need processing.
Can someone tell me if this approach would have higher performance than the usuall approach? Do type casts cost performance?

typedef struct _B2 { BYTE _0, _1; } B2;

//packs two bytes into a word
#define B2PK(w, b1, b0) { (((B2 *)&w)->_1 = (b1)); (((B2 *)&w)->_0 = (b0)); }

//Reads words
#define B2R0(w) (((B2 *)&w)->_0)
#define B2R1(w) (((B2 *)&w)->_1)
//Writes words
#define B2W0(w, b0) (((B2 *)&w)->_0 = (b0))
#define B2W1(w, b1) (((B2 *)&w)->_1 = b1))

 

[Salem]

> Can someone tell me if this approach would have higher performance than the usuall approach?
Seems doubtful - all those pointer accesses look expensive

Why are you using a structure in the first place? I would be using an unsigned short to store two bytes, and using something like &, |, << and >> to extract and insert bytes

> Do type casts cost performance?
Depends - for example and int to a float takes some work.
Casting from one pointer type to another is fairly transparent on the usual desktop machines (I think). However, it is generally best avoided - casting Type1* into Type2* can lead to things like bus errors (trying to access mis-aligned data).

> #define B2R0(w)
What sort of 'w' are you passing to this function?
If w is of type B2, why bother with all the pointer / cast / dereference.

 

[lionelhill]

casting words into bytes and vice versa should cost nothing. It's just a matter of addressing, and the compiler should be able to handle that without any extra work. Shifting definitely costs (though not a lot).
On a 32 bit system if you are worried about performance, though, you should not be using words at all. Direct manipulation of a word in 32 bit code needs an operand size prefix on the op code, and this is a costly thing.
Masking is not free, but it is dirt cheap, ands being amongst the fastest and least offensive of all processor operations.

 

[RbgoWeb]

I submitted this thread in two forums at once. Sorry, but I'll continue on the other one in C++: Microsoft.
I won't do it again. =)