[RFC] staging: media: atomisp: Simplyfy masking bit logic

[RFC] staging: media: atomisp: Simplyfy masking bit logic

[Adrian Barnaś]

Simplified masking logic in pci/hive_isp_css_common/host/vmem.c.

---

I have tested this change on whole range of *valid* inputs, and it gives
the same results as before, but this function seems to be little
counter-intuitive as far as start is a (bit index) but end is
(bit index + 1).

It is follow up to: https://lore.kernel.org/linux-staging/20250901091050.1935505-1-abarnas@google.com/
---
 .../staging/media/atomisp/pci/hive_isp_css_common/host/vmem.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/drivers/staging/media/atomisp/pci/hive_isp_css_common/host/vmem.c b/drivers/staging/media/atomisp/pci/hive_isp_css_common/host/vmem.c
index 722b684fbc37..9703e39b7497 100644
--- a/drivers/staging/media/atomisp/pci/hive_isp_css_common/host/vmem.c
+++ b/drivers/staging/media/atomisp/pci/hive_isp_css_common/host/vmem.c
@@ -22,14 +22,14 @@ typedef hive_uedge *hive_wide;
 static inline hive_uedge
 subword(hive_uedge w, unsigned int start, unsigned int end)
 {
-	return (w & (((1ULL << (end - 1)) - 1) << 1 | 1)) >> start;
+	return (w & __GENMASK_ULL(end-1, 0)) >> start;
 }
 
 /* inverse subword bits move like this: MSB[xxxx____xxxx]LSB -> MSB[xxxx0000xxxx]LSB */
 static inline hive_uedge
 inv_subword(hive_uedge w, unsigned int start, unsigned int end)
 {
-	return w & (~(((1ULL << (end - 1)) - 1) << 1 | 1) | ((1ULL << start) - 1));
+	return w & (~__GENMASK_ULL(end-1, start));
 }
 
 #define uedge_bits (8 * sizeof(hive_uedge))
-- 
2.51.0.318.gd7df087d1a-goog

Re: [RFC] staging: media: atomisp: Simplyfy masking bit logic

[Dan Carpenter]

On Tue, Sep 02, 2025 at 07:38:40AM +0000, Adrian Barnaś wrote:
> Simplified masking logic in pci/hive_isp_css_common/host/vmem.c.
> 
> ---
> 
> I have tested this change on whole range of *valid* inputs, and it gives
> the same results as before, but this function seems to be little
> counter-intuitive as far as start is a (bit index) but end is
> (bit index + 1).

Yeah.  The "end - 1" is unfortunate.  I guess we accidentally started
counting from 1...  I looked at how to remove it but got lost.

> 
> It is follow up to: https://lore.kernel.org/linux-staging/20250901091050.1935505-1-abarnas@google.com/
> ---
>  .../staging/media/atomisp/pci/hive_isp_css_common/host/vmem.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/staging/media/atomisp/pci/hive_isp_css_common/host/vmem.c b/drivers/staging/media/atomisp/pci/hive_isp_css_common/host/vmem.c
> index 722b684fbc37..9703e39b7497 100644
> --- a/drivers/staging/media/atomisp/pci/hive_isp_css_common/host/vmem.c
> +++ b/drivers/staging/media/atomisp/pci/hive_isp_css_common/host/vmem.c
> @@ -22,14 +22,14 @@ typedef hive_uedge *hive_wide;
>  static inline hive_uedge
>  subword(hive_uedge w, unsigned int start, unsigned int end)
>  {
> -	return (w & (((1ULL << (end - 1)) - 1) << 1 | 1)) >> start;
> +	return (w & __GENMASK_ULL(end-1, 0)) >> start;
>  }
>  
>  /* inverse subword bits move like this: MSB[xxxx____xxxx]LSB -> MSB[xxxx0000xxxx]LSB */
>  static inline hive_uedge
>  inv_subword(hive_uedge w, unsigned int start, unsigned int end)
>  {
> -	return w & (~(((1ULL << (end - 1)) - 1) << 1 | 1) | ((1ULL << start) - 1));
> +	return w & (~__GENMASK_ULL(end-1, start));
>  }

nit: white space.  Add spaces.  Remove parentheses.

These are supposed to be opposites, right?  Subword and inverse Subword.
You could dress them up to make them look more opposite.

	return (w & __GENMASK_ULL(end - 1, start)) >> start;
	return w & ~__GENMASK_ULL(end - 1, start);

regards,
dan carpenter

Re: [RFC] staging: media: atomisp: Simplyfy masking bit logic

[Andy Shevchenko]

On Tue, Sep 02, 2025 at 07:38:40AM +0000, Adrian Barnaś wrote:
> Simplified masking logic in pci/hive_isp_css_common/host/vmem.c.

...

> ---
> 
> I have tested this change on whole range of *valid* inputs, and it gives
> the same results as before, but this function seems to be little
> counter-intuitive as far as start is a (bit index) but end is
> (bit index + 1).

We can't change it without changing the callers at the same time.
So, better to change just helpers and later on, if working, change
the semantics of the parameter.

> ---

...

>  subword(hive_uedge w, unsigned int start, unsigned int end)
>  {
> -	return (w & (((1ULL << (end - 1)) - 1) << 1 | 1)) >> start;
> +	return (w & __GENMASK_ULL(end-1, 0)) >> start;

Why __ variant?

>  }

...

>  inv_subword(hive_uedge w, unsigned int start, unsigned int end)
>  {
> -	return w & (~(((1ULL << (end - 1)) - 1) << 1 | 1) | ((1ULL << start) - 1));
> +	return w & (~__GENMASK_ULL(end-1, start));
>  }

Ditto.

-- 
With Best Regards,
Andy Shevchenko

Re: [RFC] staging: media: atomisp: Simplyfy masking bit logic

[Andy Shevchenko]

On Tue, Sep 02, 2025 at 12:02:29PM +0300, Dan Carpenter wrote:
> On Tue, Sep 02, 2025 at 07:38:40AM +0000, Adrian Barnaś wrote:

...

> >  static inline hive_uedge
> >  inv_subword(hive_uedge w, unsigned int start, unsigned int end)
> >  {
> > -	return w & (~(((1ULL << (end - 1)) - 1) << 1 | 1) | ((1ULL << start) - 1));
> > +	return w & (~__GENMASK_ULL(end-1, start));
> >  }
> 
> nit: white space.  Add spaces.  Remove parentheses.
> 
> These are supposed to be opposites, right?  Subword and inverse Subword.
> You could dress them up to make them look more opposite.
> 
> 	return (w & __GENMASK_ULL(end - 1, start)) >> start;
> 	return w & ~__GENMASK_ULL(end - 1, start);

The problem is (and actually with the (end-1, start) above that it might
generate a really bad code on some CPUs, so I really, really prefer the way
when _at least_ one of the parameters is constant.
That said, using

	GENMASK_ULL(end - 1, 0)

in both cases makes also them look more similar (and opposition comes on how
start is being used).

-- 
With Best Regards,
Andy Shevchenko

Re: [RFC] staging: media: atomisp: Simplyfy masking bit logic

[Dan Carpenter]

On Tue, Sep 02, 2025 at 12:54:28PM +0300, Andy Shevchenko wrote:
> On Tue, Sep 02, 2025 at 12:02:29PM +0300, Dan Carpenter wrote:
> > On Tue, Sep 02, 2025 at 07:38:40AM +0000, Adrian Barnaś wrote:
> 
> ...
> 
> > >  static inline hive_uedge
> > >  inv_subword(hive_uedge w, unsigned int start, unsigned int end)
> > >  {
> > > -	return w & (~(((1ULL << (end - 1)) - 1) << 1 | 1) | ((1ULL << start) - 1));
> > > +	return w & (~__GENMASK_ULL(end-1, start));
> > >  }
> > 
> > nit: white space.  Add spaces.  Remove parentheses.
> > 
> > These are supposed to be opposites, right?  Subword and inverse Subword.
> > You could dress them up to make them look more opposite.
> > 
> > 	return (w & __GENMASK_ULL(end - 1, start)) >> start;
> > 	return w & ~__GENMASK_ULL(end - 1, start);
> 
> The problem is (and actually with the (end-1, start) above that it might
> generate a really bad code on some CPUs, so I really, really prefer the way
> when _at least_ one of the parameters is constant.

Ah.  Okay.  That makes sense.  Thanks.

regards,
dan carpenter

Re: [RFC] staging: media: atomisp: Simplyfy masking bit logic

[Adrian Barnaś]

> These are supposed to be opposites, right?  Subword and inverse Subword.
> You could dress them up to make them look more opposite.

In fact this name is misleading as well. For me it should be named
`clear_subword()`. It is zero-ing the "subword" provided with low and
high bit boundary (start, end), to be then easily replaced with binary
or operation.
And this operation is not an inverse of extracting a subword using the
subword function.

> The problem is (and actually with the (end-1, start) above that it might
> generate a really bad code on some CPUs, so I really, really prefer the way
> when _at least_ one of the parameters is constant.

It would be  easier to create a bitmask sing GENMASK_ULL and just
reverse it but if it is not safe, this is what looks fine for me and
works the same as previous function:

// Added a helper to create a mask
static inline unsigned long long subword_bitmask(unsigned int start,
unsigned int end)
{
        return GENMASK_ULL(end - 1 , 0) & ~(GENMASK_ULL(start, 0) >> 1);
}

/* inverse subword bits move like this: MSB[xxxx____xxxx]LSB ->
MSB[xxxx0000xxxx]LSB */
static inline unsigned long long
clear_subword(unsigned long long w, unsigned int start, unsigned int end)
{
        return w & ~subword_bitmask(start, end);
}

Best regards
Adrian Barnaś