i915: fix shift warning

Commit Message

Arnd Bergmann Dec. 30, 2020, 3:39 p.m. UTC

From: Arnd Bergmann <[e-mail]>

Randconfig builds on 32-bit machines show lots of warnings for
the i915 driver for incorrect bit masks like:

drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c:2584:9: error: shift count >= width of type [-Werror,-Wshift-count-overflow]
        return hweight64(VDBOX_MASK(&i915->gt));
               ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
include/asm-generic/bitops/const_hweight.h:29:49: note: expanded from macro 'hweight64'
 #define hweight64(w) (__builtin_constant_p(w) ? __const_hweight64(w) : __arch_hweight64(w))

Since this is a 64-bit mask, use GENMASK_ULL instead of GENMASK.

Signed-off-by: Arnd Bergmann <[e-mail]>
---
 drivers/gpu/drm/i915/i915_drv.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

Comments

Chris Wilson Dec. 30, 2020, 3:56 p.m. UTC | #1

Quoting Arnd Bergmann (2020-12-30 15:39:14)
> From: Arnd Bergmann <[e-mail]>
> 
> Randconfig builds on 32-bit machines show lots of warnings for
> the i915 driver for incorrect bit masks like:

mask is a u8.

VCS0 is 2, I915_MAX_VCS 4

(u8 & GENMASK(5, 2)) >> 2

> drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c:2584:9: error: shift count >= width of type [-Werror,-Wshift-count-overflow]
>         return hweight64(VDBOX_MASK(&i915->gt));
>                ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> include/asm-generic/bitops/const_hweight.h:29:49: note: expanded from macro 'hweight64'
>  #define hweight64(w) (__builtin_constant_p(w) ? __const_hweight64(w) : __arch_hweight64(w))

So it's upset by hweight64() on the unsigned long?
So hweight_long?

Or use a cast, hweight8((intel_engine_mask_t)VDMASK())?

static __always_inline int engine_count(intel_engine_mask_t mask)
{
	return sizeof(mask) == 1 ? hweight8(mask) :
		sizeof(mask) == 2 ? hweight16(mask) :
		sizeof(mask) == 4 ? hweight32(mask) :
		hweight64(mask);
}
-Chris

Arnd Bergmann Jan. 2, 2021, 11:23 a.m. UTC | #2

On Wed, Dec 30, 2020 at 4:56 PM Chris Wilson <[e-mail]> wrote:
>
> Quoting Arnd Bergmann (2020-12-30 15:39:14)
> > From: Arnd Bergmann <[e-mail]>
> >
> > Randconfig builds on 32-bit machines show lots of warnings for
> > the i915 driver for incorrect bit masks like:
>
> mask is a u8.
>
> VCS0 is 2, I915_MAX_VCS 4
>
> (u8 & GENMASK(5, 2)) >> 2

Ah right, I misread the warning then.

> > drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c:2584:9: error: shift count >= width of type [-Werror,-Wshift-count-overflow]
> >         return hweight64(VDBOX_MASK(&i915->gt));
> >                ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> > include/asm-generic/bitops/const_hweight.h:29:49: note: expanded from macro 'hweight64'
> >  #define hweight64(w) (__builtin_constant_p(w) ? __const_hweight64(w) : __arch_hweight64(w))
>
> So it's upset by hweight64() on the unsigned long?

I suspect what is going on is that clang once again warns because it performs
more code checks before dead-code elimination than gcc does. The warning is
for the __const_hweight64() case, which is not actually used here because the
input is not a compile-time constant.

> So hweight_long?

That seems to work, I'll send a new version with that.

> Or use a cast, hweight8((intel_engine_mask_t)VDMASK())?
>
> static __always_inline int engine_count(intel_engine_mask_t mask)
> {
>         return sizeof(mask) == 1 ? hweight8(mask) :
>                 sizeof(mask) == 2 ? hweight16(mask) :
>                 sizeof(mask) == 4 ? hweight32(mask) :
>                 hweight64(mask);
> }

Fine with me as well. If you prefer that way, I'll let you handle that.

        Arnd

Chris Wilson Jan. 2, 2021, 11:32 a.m. UTC | #3

Quoting Arnd Bergmann (2021-01-02 11:23:20)
> On Wed, Dec 30, 2020 at 4:56 PM Chris Wilson <[e-mail]> wrote:
> >
> > Quoting Arnd Bergmann (2020-12-30 15:39:14)
> > > From: Arnd Bergmann <[e-mail]>
> > >
> > > Randconfig builds on 32-bit machines show lots of warnings for
> > > the i915 driver for incorrect bit masks like:
> >
> > mask is a u8.
> >
> > VCS0 is 2, I915_MAX_VCS 4
> >
> > (u8 & GENMASK(5, 2)) >> 2
> 
> Ah right, I misread the warning then.
> 
> > > drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c:2584:9: error: shift count >= width of type [-Werror,-Wshift-count-overflow]
> > >         return hweight64(VDBOX_MASK(&i915->gt));
> > >                ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> > > include/asm-generic/bitops/const_hweight.h:29:49: note: expanded from macro 'hweight64'
> > >  #define hweight64(w) (__builtin_constant_p(w) ? __const_hweight64(w) : __arch_hweight64(w))
> >
> > So it's upset by hweight64() on the unsigned long?
> 
> I suspect what is going on is that clang once again warns because it performs
> more code checks before dead-code elimination than gcc does. The warning is
> for the __const_hweight64() case, which is not actually used here because the
> input is not a compile-time constant.
> 
> > So hweight_long?
> 
> That seems to work, I'll send a new version with that.
> 
> > Or use a cast, hweight8((intel_engine_mask_t)VDMASK())?
> >
> > static __always_inline int engine_count(intel_engine_mask_t mask)
> > {
> >         return sizeof(mask) == 1 ? hweight8(mask) :
> >                 sizeof(mask) == 2 ? hweight16(mask) :
> >                 sizeof(mask) == 4 ? hweight32(mask) :
> >                 hweight64(mask);
> > }
> 
> Fine with me as well. If you prefer that way, I'll let you handle that.

While we wait for a generic hweight(), lets use hweight_long() here.
-Chris

Patch

diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
index 0a3ee4f9dc0a..ca32fa0d6a57 100644
--- a/drivers/gpu/drm/i915/i915_drv.h
+++ b/drivers/gpu/drm/i915/i915_drv.h
@@ -1624,7 +1624,7 @@  tgl_revids_get(struct drm_i915_private *dev_priv)
 	unsigned int first__ = (first);					\
 	unsigned int count__ = (count);					\
 	((gt)->info.engine_mask &						\
-	 GENMASK(first__ + count__ - 1, first__)) >> first__;		\
+	 GENMASK_ULL(first__ + count__ - 1, first__)) >> first__;		\
 })
 #define VDBOX_MASK(gt) \
 	ENGINE_INSTANCES_MASK(gt, VCS0, I915_MAX_VCS)