Re: [PATCH] oprofile per-cpu buffer overrun

[Anton Blanchard]

>   It helps if the buffer size is a power of two, of course, but integer
>   modulus is pretty quick.

If its something thats hit real hard, the way the modulus is done can
make a difference. We've seen it in various places (eg disabling tigon 1
support in the acenic changes the driver from using a variable to a
compile time constant and you can actually see it in the profile):

quickest == power of 2 compile time constant (results in a mask)
quickish == compile time constant (results in the multiplication by
		an inverse trick)
slow == run time (results in a divide)

Of course you can do like the printk stuff and use a variable but
enforce a power of 2 value and & it.

Anton

--

unsigned long i, j, k;

int quickest()
{
        j = i % 2;
}

int quickish()
{
        j = i % 3;
}

int slow()
{
        j = i % k;
}

--

quickest:
        lis 9,i@ha
        lwz 0,i@l(9)
        lis 9,j@ha
        rlwinm 0,0,0,31,31
        stw 0,j@l(9)
        blr
quickish:
        lis 9,i@ha
        lis 0,0xaaaa
        lwz 11,i@l(9)
        ori 0,0,43691
        lis 9,j@ha
        mulhwu 0,11,0
        srwi 0,0,1
        mulli 0,0,3
        subf 11,0,11
        stw 11,j@l(9)
        blr
slow:
        lis 9,i@ha
        lis 11,k@ha
        lwz 10,i@l(9)
        lwz 9,k@l(11)
        divwu 0,10,9
        mullw 0,0,9
        lis 9,j@ha
        subf 10,0,10
        stw 10,j@l(9)
        blr
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/