RE: weird sb1250 behavior

RE: weird sb1250 behavior

[Adam Kiepul]

Jamal,

There is a possible cache line read-after-write pseudo-dependency that, along with the code alignment in terms of the instruction pair doublewords, may do something weird to the sb1250 pipeline. Just my guess.

Have a great weekend,

Adam


-----Original Message-----
From: jamal [mailto:hadi@cyberus.ca]
Sent: Friday, May 28, 2004 12:37 PM
To: linux-mips@linux-mips.org
Cc: sibyte-users@bitmover.com
Subject: weird sb1250 behavior



found some very strange behavior with sb1250.
Gcc 3.2.3 with sibyte mods. Running Linux 2.4.21 with whatever
mods off sibyte.

Testcase:
sending a large amount of traffic 
-->eth0-->someprocessing-->eth1

given the nature of processing, say i was getting 100Kpps throughput.
Now i fire a very basic program that has just loops and forever
sums up two numbers.

---
      1 #include <stdlib.h>
      2 
      3 int main ()
      4 {
      5         int a = 1;
      6         int b = 2;
      7         int c = 0; 
      8         // int c;
      9         while (1) {
     10                 c = a + b;
     11         }
     12 }
--------

I see very little drop in throughput - probably around 0.01%.

Now comment line 7 then uncomment line 8. Hallelujah.
Perfomance drops to about 100pps. Thats about a factor of 1000 down!

Interesting thing is if you add a nop (__asm__ __volatile__("nop");)
in the second version just before the while loop, we get back the same
performance as in the earlier version.
Apologies in advance for attaching objdumps (since there maybe folks who
dont have access to the sibyte tools)
1) while-init-dis is for case 1 where c is initialized
2) while-noinit-dis is for case 2 where c is not initialize
3) while-nop-dis is for case 3 when you have nop thrown in.


cheers,
jamal

Re: [SiByte] RE: weird sb1250 behavior

[Ralf Baechle]

On Fri, May 28, 2004 at 01:11:38PM -0700, Adam Kiepul wrote:

> There is a possible cache line read-after-write pseudo-dependency that, along with the code alignment in terms of the instruction pair doublewords, may do something weird to the sb1250 pipeline. Just my guess.

memcpy's source deals with what probably is another instance of the same
effect:

#ifdef CONFIG_CPU_SB1
        nop                             # improves slotting
#endif

  Ralf