extract() and implement() have brain damaged attempts to handle 32-bit wide
"fields".

The problem is the index math in the original code didn't clear all the
relevant bits.  (offset >> 5) only compensated for 32-bit index.  We need
(offset >> 6) if we want to use 64-bit loads.

But it was also wrong in that it tried to use quasi-aligned loads.  Ie
"report" was only incremented in multiples of 4 bytes and then the offset
was masked off for values greater than 4 bytes.  The right way is to
pretend "report" points at a byte array.  And offset is then only minor
adjustment for < 8 bits of offset.  "n" (field width) can then be as big as
24 (assuming 32-bit loads) since "offset" will never be bigger than 7.

If someone needs either function to handle more than 24-bits, please
document why - point at a specification or specific USB hid device - in
comments in the code.

extract/implement() are also an eyesore to read.  Please banish whoever
wrote it to read CodingStyle 3 times in a row to a classroom full of 1st
graders armed with rubberbands.  Or just flame them.  Whatever.  Globbing
all the code together on two lines does NOT make it faster and is Just
Wrong.

I've tested this patch on j6000 (dual 750Mhz PA-RISC, 32-bit 2.6.12-rc5). 
Kyle McMartin tested on c3000 (up 400Mhz PA-RISC, same kernel).  "p2-mate"
(Peter De Schrijver?) tested on sb1250 (dual core Mips, broadcom "swarm"
eval board).
Signed-off-by: Grant Grundler <grundler@parisc-linux.org>
Signed-off-by: Matthew Wilcox <matthew@wil.cx>
Cc: Vojtech Pavlik <vojtech@suse.cz>
Cc: Dmitry Torokhov <dtor@mail.ru>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>