Re: [PATCH v3 3/3] memory: Add Broadcom STB memory controller driver

[Florian Fainelli <f.fainelli@gmail.com>]

On 8/12/22 11:41, Krzysztof Kozlowski wrote:
> On 12/08/2022 20:52, Florian Fainelli wrote:
> 
>>>> unless you also implied enclosing those functions under an #if
>>>> IS_ENABLED(CONFIG_PM) or something which is IMHO less preferable.
>>>
>>> Are you sure you added also pm_ptr()? I don't see such warnings with W=1
>>> and final object does not have the functions (for a different driver but
>>> same principle).
>>
>> Yes I am sure I added pm_ptr() see the v4 I just submitted. I don't see
>> how the compiler cannot warn about the functions being unused the day
>> they stop being referenced by the pm_ops structure which is eliminated?
> 
> I don't have the answer how it exactly works (or which gcc version
> introduced it), but I tested it and the functions were not present in
> the object file, thus of course no warnings.
> 
> The driver change I am referring to is:
> https://lore.kernel.org/all/20220808174107.38676-15-paul@crapouillou.net/
> 
> I think the only difference against your v4 is:
> DEFINE_SIMPLE_DEV_PM_OPS
> and lack of __maybe_unused on the functions.
> 
> The DEFINE_SIMPLE_DEV_PM_OPS itself adds __maybe_unused for !CONFIG_PM
> case, but I don't think it is relevant.

It definitively is relevant here. SIMPLE_DEV_PM_OPS without 
__maybe_unused results in the following pre-processed code:

static int brcmstb_memc_suspend(struct device *dev)
{
  struct brcmstb_memc *memc = dev_get_drvdata(dev);
  void *cfg = memc->ddr_ctrl + memc->srpd_offset;
  u32 val;

  if (memc->timeout_cycles == 0)
   return 0;






  val = ({ u32 __r = (( __u32)(__le32)(( __le32) __raw_readl(cfg))); 
__r; });
  val &= ~((((1UL))) << (16));
  __raw_writel(( u32) (( __le32)(__u32)(val)),cfg);

  (void)({ u32 __r = (( __u32)(__le32)(( __le32) __raw_readl(cfg))); 
__r; });

  return 0;
}

static int brcmstb_memc_resume(struct device *dev)
{
  struct brcmstb_memc *memc = dev_get_drvdata(dev);

  if (memc->timeout_cycles == 0)
   return 0;

  return brcmstb_memc_srpd_config(memc, memc->timeout_cycles);
}

static const struct dev_pm_ops __attribute__((__unused__)) 
brcmstb_memc_pm_ops = { }
                         ;

static struct platform_driver brcmstb_memc_driver = {
  .probe = brcmstb_memc_probe,
  .remove = brcmstb_memc_remove,
  .driver = {
   .name = "brcmstb_memc",
   .of_match_table = brcmstb_memc_of_match,
   .pm = ((1) ? ((&brcmstb_memc_pm_ops)) : ((void *)0)),
  },
};

Now with DEFINE_SIMPLE_PM_OPS, we get the following instead:

static const struct dev_pm_ops brcmstb_memc_pm_ops = { .suspend = ((0) ? 
((brcmstb_memc_suspend)) : ((void *)0)), .resume = ((0) ? 
((brcmstb_memc_resume)) : ((void *)0)), .freeze = ((0) ? 
((brcmstb_memc_suspend)) : ((void *)0)), .thaw = ((0) ? 
((brcmstb_memc_resume)) : ((void *)0)), .poweroff = ((0) ? 
((brcmstb_memc_suspend)) : ((void *)0)), .restore = ((0) ? 
((brcmstb_memc_resume)) : ((void *)0)), .runtime_suspend = ((void *)0), 
.runtime_resume = ((void *)0), .runtime_idle = ((void *)0), }
                         ;

static struct platform_driver brcmstb_memc_driver = {
  .probe = brcmstb_memc_probe,
  .remove = brcmstb_memc_remove,
  .driver = {
   .name = "brcmstb_memc",
   .of_match_table = brcmstb_memc_of_match,
   .pm = ((1) ? ((&brcmstb_memc_pm_ops)) : ((void *)0)),
  },
};

so we will continue to reference the functions, but eventually we will 
eliminate those at the optimization stage by figuring out that this is 
dead code, therefore it can be eliminated. Note that in both cases the 
object does not contain brcmstb_memc_{resume,suspend} which makes sense 
because the warnings are generated at the time the code is processed, 
and the dead code elimination at one of the optimization stages.

I can spin a v5 with DEFINE_SIMPLE_PM_OPS if you prefer to get rid of 
the __maybe_unused.
-- 
Florian