Re: [avr-libc-dev] soliciting feedback/thoughts on potential value of a few gcc 4.0 tweaks.

[Paul Schlie]

Yes but:

    volatile f;
    f = f * f;

Produces:

        ldd    r14, Y+18    ; 0x12
        ldd    r15, Y+19    ; 0x13
        ldd    r16, Y+20    ; 0x14
        ldd    r17, Y+21    ; 0x15
        ldd    r24, Y+18    ; 0x12
        ldd    r25, Y+19    ; 0x13
        ldd    r26, Y+20    ; 0x14
        ldd    r27, Y+21    ; 0x15
        movw    r18, r24
        movw    r20, r26
        movw    r24, r16
        movw    r22, r14
        call    0x5a4
        movw    r26, r24
        movw    r24, r22
        std    Y+18, r24    ; 0x12
        std    Y+19, r25    ; 0x13
        std    Y+20, r26    ; 0x14
        std    Y+21, r27    ; 0x15

But would have expected:

        ldd    r22, Y+18    ; 0x12
        ldd    r23, Y+19    ; 0x13
        ldd    r24, Y+20    ; 0x14
        ldd    r25, Y+21    ; 0x15
        ldd    r18, Y+18    ; 0x12
        ldd    r19, Y+19    ; 0x13
        ldd    r20, Y+20    ; 0x14
        ldd    r21, Y+21    ; 0x15
        call    0x5a4
        std    Y+18, r22    ; 0x12
        std    Y+19, r23    ; 0x13
        std    Y+20, r24    ; 0x14
        std    Y+21, r25    ; 0x15

Anyone know why this is not being produced? (or simply an optimization bug?)


> From: "Dmitry K." <address@hidden>
> 
> Paul Schlie wrote:
> 
>> - any insight as to why the compiler (3.4.3 for the sake of argument) first
>>   load variable into one set of registers to just then passes them to the
>>   function's parameter registers, as opposed to loading them there
>>   initially, which seems like a waste of cycles? happens for example with:
>> 
>>   volatile float f;
>>   f = f * f; // or most any other multi-byte libgcc defined function.
> 
> `volatile' is the reason of double reading.
> 
> avr-gcc 3.3.4 with `-mmcu=avr4 -Os -fnew-ra': beautiful code.
> 
> See:
> 
>    float x;
>    float foo (void)
>    {
>        return x * x;
>    }
> 
> produce:
>         lds r22,x
>         lds r23,(x)+1
>         lds r24,(x)+2
>         lds r25,(x)+3
>         movw r18,r22
>         movw r20,r24
>         rcall __mulsf3
> 
> P.S. I do not say about the pair <rcall, ret>. In many cases gcc (x86 port,
> for example) reduce such sequences.
>