Re: op_alloc_counter.c

[William C. <wc...@re...>]

Philippe Elie wrote:
> Perhaps I missed a discussion but John applying the xml patch
> reverted the last commit by Will. If the revert was intended the
> ChangeLog says nothing about this, I'm speaking about:
> 
> --- op_alloc_counter.c	2006/11/17 18:20:38	1.7
> +++ op_alloc_counter.c	2006/11/17 23:47:29	1.8
> @@ -130,7 +130,7 @@
>  		counter_arc const * arc = list_entry(pos, counter_arc, next);
>  
>  		if (allocated_mask & (1 << arc->counter))
> -			continue;
> +			return 0;
>  
>  		counter_map[depth] = arc->counter;
>  
> anyway I'm unsure if Will patch is correct. Will have you a test case
> showing a failure which is now working with your patch applied ?
> 

Hi Philippe,

There are two cases where this patch is needed.

In August Jeramiah Lott encountered a case on mips/34K where op_help could not 
allocate the registers. This is due to some events being available on a subset 
of registers other events. INSTRUCTIONS and L2_MISSES available and 0,1,2,and 3, 
but BRANCH_MISPREDICTS and BRANCH_INSNS_LAUNCHED only available on 0 and 2. With 
patch get the following:

[test@timesys-y91ad5l oprofile]$ utils/ophelp -c mips/34K --check-events
INSTRUCTIONS:500:0:1:1 BRANCH_MISPREDICTS:500:0:1:1
BRANCH_INSNS_LAUNCHED:500:0:1:1 L2_MISSES:500:0:1:1
1 0 2 3

The other case is dealing with the changes in the 2.6.19-rc6 kernel where one of 
the registers is already reserved by the watchdog timer.

-Will