Re: [PATCH v2] sched/core: silence a warning in sched_init()

[Peter Zijlstra]

On Tue, Jun 25, 2019 at 10:04:19AM -0400, Qian Cai wrote:
> On Tue, 2019-06-25 at 15:52 +0200, Peter Zijlstra wrote:

> Yes, -Wmissing-prototype makes no sense, but "-Wunused-but-set-variable" is
> pretty valid to catch certain developer errors. For example,
> 
> https://lists.linuxfoundation.org/pipermail/iommu/2019-May/035680.html
> 
> > 
> > As to this one, ideally the compiler would not be stupid, and understand
> > the below, but alas.
> 
> Pretty sure that won't work, as the compiler will complain something like,
> 
> ISO C90 forbids mixed declarations and code

No, it builds just fine, it's a new block and C allows new variables at
every block start -- with the scope of that block.

And for our config, alloc_size is an unconditional 0, so it should DCE
the whole block and with that our variable. But clearly the passes are
the other way around :/

> > diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> > index fa43ce3962e7..cb652e165570 100644
> > --- a/kernel/sched/core.c
> > +++ b/kernel/sched/core.c
> > @@ -6369,7 +6369,7 @@ DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
> >  
> >  void __init sched_init(void)
> >  {
> > -	unsigned long alloc_size = 0, ptr;
> > +	unsigned long alloc_size = 0;
> >  	int i;
> >  
> >  	wait_bit_init();
> > @@ -6381,7 +6381,7 @@ void __init sched_init(void)
> >  	alloc_size += 2 * nr_cpu_ids * sizeof(void **);
> >  #endif
> >  	if (alloc_size) {
> > -		ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
> > +		unsigned long ptr = (unsigned long)kzalloc(alloc_size,
> > GFP_NOWAIT);
> >  
> >  #ifdef CONFIG_FAIR_GROUP_SCHED
> >  		root_task_group.se = (struct sched_entity **)ptr;