Re: First kernel patch (optimization)

[Jaime Arrocha <jarr@innercoder.com>]

On 09/16/2015 07:56 AM, David Laight wrote:
> From: Austin S Hemmelgarn
>> Sent: 16 September 2015 12:46
>> On 2015-09-15 20:09, Steve Calfee wrote:
>>> On Tue, Sep 15, 2015 at 12:53 PM, Eric Curtin <ericcurtin17@gmail.com> wrote:
>>>> Signed-off-by: Eric Curtin <ericcurtin17@gmail.com>
>>>>
>>>> diff --git a/tools/usb/usbip/src/usbip_detach.c b/tools/usb/usbip/src/usbip_detach.c
>>>> index 05c6d15..9db9d21 100644
>>>> --- a/tools/usb/usbip/src/usbip_detach.c
>>>> +++ b/tools/usb/usbip/src/usbip_detach.c
>>>> @@ -47,7 +47,9 @@ static int detach_port(char *port)
>>>>           uint8_t portnum;
>>>>           char path[PATH_MAX+1];
>>>>
>>>> -
>>>> +       unsigned int port_len = strlen(port);
>>>> +
>>>> +       for (unsigned int i = 0; i < port_len; i++)
>>>>                   if (!isdigit(port[i])) {
>>>>                           err("invalid port %s", port);
>>>>                           return -1;
>>>>
>>>> --
>>> Hi Eric,
>>>
>>> This is fine, but what kind of wimpy compiler optimizer will not move
>>> the constant initializer out of the loop? I bet if you compare binary
>>> sizes/code it will be exactly the same, and you added some characters
>>> of code. Reorganizing code for readability is fine, but for compiler
>>> (in)efficiency seems like a bad idea.
>> While I agree with your argument, I would like to point out that it is a
>> well established fact that GCC's optimizers are kind of brain-dead at
>> times and need their hands held.
>>
>> I'd be willing to bet that the code will be marginally larger (because
>> of adding another variable), but might run slightly faster too (because
>> in my experience, GCC doesn't always catch things like this), and should
>> compile a little faster (because the optimizers don't have to do as much
>> work).
> The compiler probably can't optimise the strlen().
> If isdigit() is a real function (the locale specific one probably is)
> then the compile cannot assume that port[n] isn't changed by the call
> to isdigit.
>
> A simpler change would be:
> 	for (unsigned int i = 0; port[i] != 0; i++)
>
> Much better would be to use strtoul() instead of atoi().
>
> 	David
>
I actually took some time to verify this. GCC makes this optimization 
with -O2 at least on gcc 4.7.2.
One interesting observation I found was that in O0 and O2, it does make 
a call to strlen while in O1 it calculates
the length of the string using:

repnz scas    %es:(%rdi),%al
not                %rcx
sub               $0x2,%rcx

Why does it do that? Is the code above faster? If yes, why not do it in 
O2 too?
Is this still a topic for this forum?


gcc version 4.7.2 (Debian 4.7.2-5)
code

void conv_input(char *port)
{
     int portnum;

     for(int i = 0; i <strlen(port); i++)
         if(!isdigit(port[i])) {
             printf("invalid port %s", port);
             exit (1);
         }

     portnum = atoi(port);
     printf("Port number: %d\n", portnum);
}

Optimization done?
              O0    O1    O2
x86        No    No    Yes
amd64   No   No    Yes