I am not sure, if this works as you expect... maybe the macros and wrappers generate a copy under the hood?

use Inline C;
use strict;
use warnings;

my $x = "012\xf0ABC\x0f\x00\x01\xff";

print unpack("H*", $x ), "\n";    #  303132f04142430f0001ff

flipbits( $x, length $x );

print unpack("H*", $x ), "\n";    #  cfcecd0fbebdbcf0fffe00

__END__

__C__

void flipbits(char* a, int len ) {

  for ( ; len--  ; ++a ) *a ^= 0xff;

}
[download]

Update: Switching to 32bit blocks gave a huge improvement on my 32bit setup...

void flipbits32(char* a, int len ) {

  int blocks = len >> 2;

  if ( sizeof blocks == 4 ) { /* try 32bit blocks */
    for ( ; blocks-- ; a += 4 )  *( (unsigned int*) a ) ^= 0xffffffffU
+;
    /* TODO: add treatment for last 1..3 bytes if len is not a multipl
+e of 4 */
  } else { /* fallback to 8bit */
    for ( ; len--   ; ++a    )   *a ^= 0xff;
  }

}
[download]

            Rate      str   inline inline32
str        69.1/s       --     -37%     -69%
flipbits    110/s      59%       --     -51%
flipbits32  223/s     222%     103%       --
[download]

Potential next steps which are also less general, but that's the price to pay:

• flipbits64()
• SSE
• CUDA