Not sure if a C version of the jmcnamara idea got benchmarked. I think such a thing would look like below. I didn't look at the ASM code, but even at a low optimize level, this probably results in pretty good code.

#include <stdio.h>

#define MASK05_00 (0x3f)            /*6 bits at a time*/
#define MASK11_06 (MASK05_00 << 6)
#define MASK17_12 (MASK11_06 << 6)
#define MASK23_18 (MASK17_12 << 6)


int main (int argc, char **argv)
{

   unsigned char in  [24] =
          { 00, 16, 131, 16, 81, 135, 32, 146, 139,  48, 211, 143,
            65, 20, 147, 81, 85, 151, 97, 150, 155, 113, 215, 159};
   unsigned char out [32];

   void expand6 (unsigned char *in, unsigned char* out);
   void hexdump (unsigned char *p, int n);

   expand6(in,  out);
   hexdump(in,  24);
   hexdump(out, 32);

   return(0);
}

void expand6 (unsigned char *in, unsigned char* out)
{
    unsigned int temp;
    int i;

    for (i=0; i<8; i++)    /*8 24 bit sequences */
    {
        temp  = *in++ << 16;
        temp |= *in++ << 8;
        temp |= *in++;

        *out++ = (temp & MASK23_18)>>18;
        *out++ = (temp & MASK17_12)>>12;
        *out++ = (temp & MASK11_06)>>6;
        *out++ = (temp & MASK05_00);
    }
}

void hexdump(unsigned char *p, int n)
{
    int i;
    for (i=0; i<n; i++)
        printf("%02x ", *p++);

    puts("\n");
}
/* ouput

00 10 83 10 51 87 20 92  **in buf
8b 30 d3 8f 41 14 93 51 
55 97 61 96 9b 71 d7 9f 

00 01 02 03 04 05 06 07  **out buf
08 09 0a 0b 0c 0d 0e 0f 
10 11 12 13 14 15 16 17 
18 19 1a 1b 1c 1d 1e 1f 
*/
[download]