Group some variables based on their sizes to reduce hole and avoid padding.
On x86_64, this shrinks the size of 'struct clk_fractional_divider'
from 72 to 64 bytes.
It saves a few bytes of memory when the structure is kzalloc()'ed.
Signed-off-by: Christophe JAILLET <[email protected]>
---
Using pahole
Before:
======
struct clk_fractional_divider {
struct clk_hw hw; /* 0 24 */
void * reg; /* 24 8 */
u8 mshift; /* 32 1 */
u8 mwidth; /* 33 1 */
/* XXX 2 bytes hole, try to pack */
u32 mmask; /* 36 4 */
u8 nshift; /* 40 1 */
u8 nwidth; /* 41 1 */
/* XXX 2 bytes hole, try to pack */
u32 nmask; /* 44 4 */
u8 flags; /* 48 1 */
/* XXX 7 bytes hole, try to pack */
void (*approximation)(struct clk_hw *, long unsigned int, long unsigned int *, long unsigned int *, long unsigned int *); /* 56 8 */
/* --- cacheline 1 boundary (64 bytes) --- */
spinlock_t * lock; /* 64 8 */
/* size: 72, cachelines: 2, members: 11 */
/* sum members: 61, holes: 3, sum holes: 11 */
/* last cacheline: 8 bytes */
};
After:
=====
struct clk_fractional_divider {
struct clk_hw hw; /* 0 24 */
void * reg; /* 24 8 */
u8 flags; /* 32 1 */
u8 mshift; /* 33 1 */
u8 mwidth; /* 34 1 */
/* XXX 1 byte hole, try to pack */
u32 mmask; /* 36 4 */
u8 nshift; /* 40 1 */
u8 nwidth; /* 41 1 */
/* XXX 2 bytes hole, try to pack */
u32 nmask; /* 44 4 */
void (*approximation)(struct clk_hw *, long unsigned int, long unsigned int *, long unsigned int *, long unsigned int *); /* 48 8 */
spinlock_t * lock; /* 56 8 */
/* size: 64, cachelines: 1, members: 11 */
/* sum members: 61, holes: 2, sum holes: 3 */
};
---
include/linux/clk-provider.h | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/include/linux/clk-provider.h b/include/linux/clk-provider.h
index cf1adfeaf257..8cec844f5dde 100644
--- a/include/linux/clk-provider.h
+++ b/include/linux/clk-provider.h
@@ -1134,13 +1134,13 @@ struct clk_hw *clk_hw_register_fixed_factor_parent_hw(struct device *dev,
struct clk_fractional_divider {
struct clk_hw hw;
void __iomem *reg;
+ u8 flags;
u8 mshift;
u8 mwidth;
u32 mmask;
u8 nshift;
u8 nwidth;
u32 nmask;
- u8 flags;
void (*approximation)(struct clk_hw *hw,
unsigned long rate, unsigned long *parent_rate,
unsigned long *m, unsigned long *n);
--
2.34.1
Quoting Christophe JAILLET (2023-02-17 07:55:55)
> diff --git a/include/linux/clk-provider.h b/include/linux/clk-provider.h
> index cf1adfeaf257..8cec844f5dde 100644
> --- a/include/linux/clk-provider.h
> +++ b/include/linux/clk-provider.h
> @@ -1134,13 +1134,13 @@ struct clk_hw *clk_hw_register_fixed_factor_parent_hw(struct device *dev,
> struct clk_fractional_divider {
> struct clk_hw hw;
> void __iomem *reg;
> + u8 flags;
> u8 mshift;
> u8 mwidth;
> u32 mmask;
This member ...
> u8 nshift;
> u8 nwidth;
> u32 nmask;
and this member can be calculated. Doing that would save more bytes when
combined with this patch. Can you make this into a patch series?
> - u8 flags;
> void (*approximation)(struct clk_hw *hw,
> unsigned long rate, unsigned long *parent_rate,
> unsigned long *m, unsigned long *n);
Le 29/03/2023 à 20:46, Stephen Boyd a écrit :
> Quoting Christophe JAILLET (2023-02-17 07:55:55)
>> diff --git a/include/linux/clk-provider.h b/include/linux/clk-provider.h
>> index cf1adfeaf257..8cec844f5dde 100644
>> --- a/include/linux/clk-provider.h
>> +++ b/include/linux/clk-provider.h
>> @@ -1134,13 +1134,13 @@ struct clk_hw *clk_hw_register_fixed_factor_parent_hw(struct device *dev,
>> struct clk_fractional_divider {
>> struct clk_hw hw;
>> void __iomem *reg;
>> + u8 flags;
>> u8 mshift;
>> u8 mwidth;
>> u32 mmask;
>
> This member ...
>
>> u8 nshift;
>> u8 nwidth;
>> u32 nmask;
>
> and this member can be calculated. Doing that would save more bytes when
> combined with this patch. Can you make this into a patch series?
Hi,
with your proposed change, there is no more need to move flags. All u8
are grouped and the struct is optimal.
I'll send a v2, but not sure it really worth it.
When dynamically allocated, even when the struct is 56 bytes, 64 are
still allocated.
And there are only 2 users of struct clk_fractional_divider embedded in
another struct that will be shrinked with your change.
On the other side, having theses masks pre-computed could save a few
cycles, at nearly no memory usage impact.
I'll let you decide if v1 or v2 is the preferred way to go.
CJ
>
>> - u8 flags;
>> void (*approximation)(struct clk_hw *hw,
>> unsigned long rate, unsigned long *parent_rate,
>> unsigned long *m, unsigned long *n);
>