Re: [avr-gcc-list] Re: sprintf

[Georg-Johann Lay]

David Brown schrieb:
> However, the compiler has a much 
> better chance of doing bounds checking, alias checking, and other 
> optimisations on the array expression.  It is even free to reorganise 
> the array, or keep it all in registers (I don't think any compilers do 
> that, but maybe one day...), whereas working with raw addresses will 
> often cause the compiler to lose the extra knowledge of the array. 

Taking addresses of auto and params forces them into memory, because you 
cannot take an address of a register. Even if it is possible in a 
specific hardware (AVR is an example), addresses on regs cannot work as 
intended. a=b[i] with i unknown at compiletime is the same: addresses 
must be calculated at runtime.

If no addresses are taken and the array has some comfortable size, gcc 
can hold it in register:

typedef union
{
    uint8_t  asByte[4];
    uint16_t asWord[2];
    int16_t  asShort[2];
    uint32_t asULong;
    int32_t  asLong;
} data32_t;

avr-gcc can hold that in a register, thus I am using such types for data 
conversion and byte extraction/insertion.

Yust try this:

data32_t bswap (data32_t a)
{
    data32_t b;
    b.asByte[0] = a.asByte[3];
    b.asByte[1] = a.asByte[2];
    b.asByte[2] = a.asByte[1];
    b.asByte[3] = a.asByte[0];

    return b;
}

data32_t bswap2 (data32_t a)
{
    return (data32_t) {.asByte = {a.asByte[3], a.asByte[2], 
a.asByte[1], a.asByte[0]}};
}

Depending if you allow for wide types to split, the performance changes 
slightly. The first version with -Os compiles to optimum.

Georg-Johann