On 10/17/24 11:18, Peter Zijlstra wrote:
> On Wed, Oct 16, 2024 at 12:44:18PM -0700, Dave Hansen wrote:
> 
>> Would anybody hate if we broke this up a bit, like:
>>
>>         const typeof(var) _val = val;
>>         const int paoconst = __builtin_constant_p(val);
>>         const int paoinc   = paoconst && ((_val) == 1);
>>         const int paodec   = paoconst && ((_val) == (typeof(var))-1);
>>
>> and then did
>>
>> 	if (paoinc)
>> 		percpu_unary_op(size, qual, "inc", var);
>> 	...
> I think that is an overall improvement. Proceed! 🙂

I poked at this a bit:

> https://git.kernel.org/pub/scm/linux/kernel/git/daveh/devel.git/commit/?h=testme&id=30e0899c6ab7fe1134e4b96db963f0be89b1dd5a

I believe it functions fine.  But it surprised me with a few things.
Here's one.  I assumed that doing an add((unsigned)-1) would be rare.
It's not.  It's actually pretty common because this:

#define this_cpu_sub(pcp, val)  this_cpu_add(pcp, -(typeof(pcp))(val))

ends up causing problems when 'pcp' is an unsigned type.  For example,
in this chain:

	mem_cgroup_exit ->
	obj_cgroup_put ->
	percpu_ref_put ->
	percpu_ref_put_many(ref, 1) ->
	this_cpu_sub

the compiler can see the '1' constant.  It effectively expands to:

	this_cpu_add(pcp, -(unsigned long)(1))

With the old code, gcc manages to generate a 'dec'.  Clang generates an
'add'.  With my hack above both compilers generate an 'add'.  This
actually matters in some code that seems potentially rather performance
sensitive:

add/remove: 0/0 grow/shrink: 219/9 up/down: 755/-141 (614)
Function                                     old     new   delta
flush_end_io                                 905    1070    +165
x86_pmu_cancel_txn                           242     338     +96
lru_add                                      554     594     +40
mlock_folio_batch                           3264    3300     +36
compaction_alloc                            3813    3838     +25
tcp_leave_memory_pressure                     86     110     +24
account_guest_time                           270     287     +17
...

So I think Peter's version was the best.  It shuts up clang and also
preserves the existing (good) gcc 'sub' behavior.  I'll send it out for
real in a bit, but I'm thinking of something like the attached patch.