Re: generic strncpy - off-by-one error

[Timothy Miller <miller@techsource.com>]

Daniel Forrest wrote:
> On Sat, Aug 16, 2003 at 10:15:14AM +0200, Peter Kjellerstedt wrote:


> Shouldn't this be:
> 
> 		while (tmp & (sizeof(long) - 1)) {
> 
> 
>>			*tmp++ = '\0';
>>			count--;
>>		}

Oh, yeah!  That's right.  We need to check the address.  Also need to 
cast tmp to (int) or something (doesn't matter what it's cast to, 
because we only care about the lower 2 or 3 bits).

Peter, please see if this makes any speed difference.  But it definately 
needs this fix.


Frankly, I'm surprised it works.  In fact, it might not, but it's hard 
to tell from the tests just benchmarks.


Also, if you're doing dense addressing on Alpha, and you do byte 
accesses the addresses for char are byte addresses, but the code does 
read-modify-write to memory for byte accesses, because in that mode, you 
can only do 32-bit and 64-bit accesses.  The performance increase could 
be even greater for Alpha than for x86.


For Sparc, we might be able to do something with VIS instructions, 
although I don't know what the setup overhead is.  Sun's memcpy and 
memset only use VIS when the size is greater than 512, because 
otherwise, it's not worth it.


I don't know enough about PowerPC other than the proper use of the 
"eieio" instruction. :)