Re: [avr-gcc-list] movw and asm

[Bruce D. Lightner]

James

> Thanks Rune!  These links are helpful, and they answered my next question.
> But I am still wondering about parameters passed into asm blocks.  For
> instance..
>
> inline unsigned long Test(unsigned long Input)
> {
>     unsigned long Output;
>
>     Input = ((Input << 4) ^ (Input >> 5)) + Input;
>
>     asm
>     (
>         "mov %A0,%A1\n\t"
>         "mov %B0,%B1\n\t"
>         "mov %C0,%C1\n\t"
>         "mov %D0,%D1\n\t"
>         :"=r"(Output)
>         :"r"(Input)
>     );
>
>     Output = Output << 4;
>
>     return Output;
> }
>
> Is there is any determinism in how the compiler arranges the 8 registers
> used to hold "Output" and "Input" before it falls into the asm block?  I'm
> also looking to see if there is a trick to emit conditional statements to
> gas, like "if A0 and A1 are even registers and B0 and B1 are adjacent to
> them then use movd" or whatever.
>
> At this point, we met the benchmark by writing the whole routine over in
> assembler.  However, I'd still like to know.  Atmel added movd to their
> enhanced core.. It would be nice to be able to make use of it.

This works for me using avr-gcc version 3.3 20030421 (prerelease), 
targeted to the ATmega162, which of course has the "movw" instruction...

 inline unsigned long Test(unsigned long Input)
 {
    unsigned long Output;

    Input = ((Input << 4) ^ (Input >> 5)) + Input;

    asm
    (
        "movw %w0,%w1\n\t"
        "movw %C0,%C1\n\t"
        :"=r"(Output)
        :"r"(Input)
    );

    Output = Output << 4;

    return Output;
 }

Note the use of "w", which specifies one of the "special upper register 
pairs": r24/r25, r26/r27, r28/r29, or r30/r31.  This gives the code 
generator only 4 of the technically possible 16 register-pairs to work 
with, but it may be fine for your purposes.

Until such time as the syntax of the avr-gcc inline assembler is augmented 
to allow the specification of *even* register pairs, this is likely the 
best that one can do with inline assembly.  It would indeed be a nice 
option.

I just started using the ATmega162 a couple of days ago and have not 
looked at the avr-gcc generated code for copying "longs".  Is "movw" not 
being used when appropriate?

Best regards,

Bruce

P.S.- I note that Harald Kipp's wonderful in-line assembly "HOWTO" 
document at...

http://savannah.nongnu.org/download/avr-libc/doc/avr-libc-user-manual/inline_asm.html

...is missing the assembly "mnemonic" for "movw".  Clearly it works. 
Also, the link there to Harald's Web site has a problem: 
"http://www.ethernut.de."; should be something like 
"http://www.ethernut.de/"; instead.

--
 Bruce D. Lightner
 Lightner Engineering
 La Jolla, California
 Voice: +1-858-551-4011
 FAX: +1-858-551-0777
 Email: address@hidden
 URL: http://www.lightner.net/lightner/bruce/