On Fri, Aug 24, 2001 at 03:39:50PM -0400, Ken Wellsch wrote:
>
> index % time    self  children    called     name
> -----------------------------------------------------------------
> [10]    19.7  249.70    0.00 2023108         memcpy [10]
> -----------------------------------------------
> [16]     7.8    0.36   98.55  796813         microtime <cycle 8> [16]
>                98.35    0.00  796813/2023108     memcpy [10]
> -----------------------------------------------
>	...

I kept looking at this wondering, where the heck does microtime()
ever call memcpy()?

Well, I pondered this for a while and in looking at the code in
pmax/machdep.c I notice two "structure copies."

Sure enough, when I look at the assembler produced by "gcc" I see:

	 ...
        move    $4,$16
        move    $17,$2
        la      $5,time
        .set    noreorder
        .set    nomacro
        jal     memcpy
        li      $6,8                    # 0x8
        .set    macro
        .set    reorder
	 ...

which matches up with the line "*tvp = time;" below:

  microtime(tvp)
        struct timeval *tvp;
  { 
        int s = splclock();
        static struct timeval lasttime; 
 
        *tvp = time;

	...

Apparently "time" is treated as volatile, while the closing
structure copy "lasttime = *tvp;" does not involve memcpy().

So given that this code is executing with splclock(), is it
possible to replace

        *tvp = time;

by

        tvp->tv_sec  = time.tv_sec;
        tvp->tv_usec = time.tv_usec;

and thus avoid this expensive and ever so frequent call?

Cheers,

-- Ken