Re: [Patch] Shut up warnings from files under drivers/

[Sam Ravnborg <sam@ravnborg.org>]

On Sat, Jan 26, 2008 at 04:55:58AM -0500, Jeff Garzik wrote:
> WANG Cong wrote:
> >diff --git a/drivers/video/kyro/fbdev.c b/drivers/video/kyro/fbdev.c
> >index acb9370..437ebd0 100644
> >--- a/drivers/video/kyro/fbdev.c
> >+++ b/drivers/video/kyro/fbdev.c
> >@@ -90,7 +90,9 @@ static int nomtrr __devinitdata = 0;
> > 
> > /* PCI driver prototypes */
> > static int kyrofb_probe(struct pci_dev *pdev, const struct pci_device_id 
> > *ent);
> >+#if defined(MODULE) || defined(CONFIG_HOTPLUG)
> > static void kyrofb_remove(struct pci_dev *pdev);
> >+#endif
> > 
> > static struct fb_videomode kyro_modedb[] __devinitdata = {
> > 	{
> >@@ -754,6 +756,7 @@ out_unmap:
> > 	return -EINVAL;
> > }
> > 
> >+#if defined(MODULE) || defined(CONFIG_HOTPLUG)
> > static void __devexit kyrofb_remove(struct pci_dev *pdev)
> > {
> > 	struct fb_info *info = pci_get_drvdata(pdev);
> >@@ -783,6 +786,7 @@ static void __devexit kyrofb_remove(struct pci_dev 
> >*pdev)
> > 	pci_set_drvdata(pdev, NULL);
> > 	framebuffer_release(info);
> > }
> >+#endif
> 
> 
> Quite strange -- due to __devexit_p() and the __devexit marker, ifdefs 
> should not be needed.
> 
> I would look into why that isn't working as designed in these cases...

I checked up on the synclink.c warning.
We have the following code:

static void synclink_remove_one (struct pci_dev *dev);

...

static struct pci_driver synclink_pci_driver = {
        .remove         = __devexit_p(synclink_remove_one),
};

...

static void __devexit synclink_remove_one (struct pci_dev *dev)
{
}

And I double checked the preprocessed source to check
that we applied the __attribute__((__used__)) to the function.

Investigating a bit more I realized that gcc looses the
__used__ attribution due to the prototype.
So there are two correct fixes:
a) move the function up so we do not need the forward
   declaration
b) add a __devexit to the forward decalration too.

I strongly prefer the first version and this is the
correct fix for these cases.

Do we have a gcc bug here - I did not see a definitive answer in gcc docs?

	Sam