Optimization: keeping pointer all the time, or defining getter?

Optimization: keeping pointer all the time, or defining getter?

[Rafał Miłecki]

I try to abstract bus in b43 driver. For frequently used fields
solution is obvious:
At init time I get info from bus-specific stuct and put it in generic field.
Example: core revision. We often refer to it, I put it in abstraction struct.

My question: what is the policy for cases with less frequently used fields?
Example: I need to use "struct device" for registering some low lever stuff.

I have two solutions:

1) Realtime:
struct device *get_device(...)
{
if (dev->bus_type = B43_BUS_SSB)
return dev->ssbdev->device;
else if (dev->bus_type = B43_BUS_BCMA)
return dev->bcmadev->device;
}

2) Init time:
void b43_ssb_init(struct ssb_device *ssbdev)
{
dev->struct = ssbdev->device;
}

The first one is slower but we don't keep "struct device" pointer in
abstraction struct. Situation is opposite for the second one.

Does anyone care? ;)

-- 
Rafa?

Optimization: keeping pointer all the time, or defining getter?

[Rafał Miłecki]

I try to abstract bus in b43 driver. For frequently used fields
solution is obvious:
At init time I get info from bus-specific stuct and put it in generic field.
Example: core revision. We often refer to it, I put it in abstraction struct.

My question: what is the policy for cases with less frequently used fields?
Example: I need to use "struct device" for registering some low lever stuff.

I have two solutions:

1) Realtime:
struct device *get_device(...)
{
if (dev->bus_type = B43_BUS_SSB)
return dev->ssbdev->device;
else if (dev->bus_type = B43_BUS_BCMA)
return dev->bcmadev->device;
}

2) Init time:
void b43_ssb_init(struct ssb_device *ssbdev)
{
dev->struct = ssbdev->device;
}

The first one is slower but we don't keep "struct device" pointer in
abstraction struct. Situation is opposite for the second one.

Does anyone care? ;)

-- 
Rafał

Optimization: keeping pointer all the time, or defining getter?

[Michael Büsch]

On Wed, 2011-05-11 at 18:22 +0200, Rafa? Mi?ecki wrote: 
> I try to abstract bus in b43 driver. For frequently used fields
> solution is obvious:
> At init time I get info from bus-specific stuct and put it in generic field.
> Example: core revision. We often refer to it, I put it in abstraction struct.
> 
> My question: what is the policy for cases with less frequently used fields?
> Example: I need to use "struct device" for registering some low lever stuff.
> 
> I have two solutions:
> 
> 1) Realtime:
> struct device *get_device(...)
> {
> if (dev->bus_type = B43_BUS_SSB)
> return dev->ssbdev->device;
> else if (dev->bus_type = B43_BUS_BCMA)
> return dev->bcmadev->device;
> }
> 
> 2) Init time:
> void b43_ssb_init(struct ssb_device *ssbdev)
> {
> dev->struct = ssbdev->device;
> }
> 
> The first one is slower but we don't keep "struct device" pointer in
> abstraction struct. Situation is opposite for the second one.
> 
> Does anyone care? ;)

Put a struct device pointer into b43_wldev.
The pointer is needed on every DMA transaction several times, so this
seems worth it.

-- 
Greetings Michael.

Re: Optimization: keeping pointer all the time, or defining getter?

[Michael Büsch]

On Wed, 2011-05-11 at 18:22 +0200, Rafał Miłecki wrote: 
> I try to abstract bus in b43 driver. For frequently used fields
> solution is obvious:
> At init time I get info from bus-specific stuct and put it in generic field.
> Example: core revision. We often refer to it, I put it in abstraction struct.
> 
> My question: what is the policy for cases with less frequently used fields?
> Example: I need to use "struct device" for registering some low lever stuff.
> 
> I have two solutions:
> 
> 1) Realtime:
> struct device *get_device(...)
> {
> if (dev->bus_type = B43_BUS_SSB)
> return dev->ssbdev->device;
> else if (dev->bus_type = B43_BUS_BCMA)
> return dev->bcmadev->device;
> }
> 
> 2) Init time:
> void b43_ssb_init(struct ssb_device *ssbdev)
> {
> dev->struct = ssbdev->device;
> }
> 
> The first one is slower but we don't keep "struct device" pointer in
> abstraction struct. Situation is opposite for the second one.
> 
> Does anyone care? ;)

Put a struct device pointer into b43_wldev.
The pointer is needed on every DMA transaction several times, so this
seems worth it.

-- 
Greetings Michael.

Optimization: keeping pointer all the time, or defining getter?

[Rafał Miłecki]

W dniu 11 maja 2011 18:27 u?ytkownik Michael B?sch <m@bues.ch> napisa?:
> On Wed, 2011-05-11 at 18:22 +0200, Rafa? Mi?ecki wrote:
>> I try to abstract bus in b43 driver. For frequently used fields
>> solution is obvious:
>> At init time I get info from bus-specific stuct and put it in generic field.
>> Example: core revision. We often refer to it, I put it in abstraction struct.
>>
>> My question: what is the policy for cases with less frequently used fields?
>> Example: I need to use "struct device" for registering some low lever stuff.
>>
>> I have two solutions:
>>
>> 1) Realtime:
>> struct device *get_device(...)
>> {
>> if (dev->bus_type = B43_BUS_SSB)
>> return dev->ssbdev->device;
>> else if (dev->bus_type = B43_BUS_BCMA)
>> return dev->bcmadev->device;
>> }
>>
>> 2) Init time:
>> void b43_ssb_init(struct ssb_device *ssbdev)
>> {
>> dev->struct = ssbdev->device;
>> }
>>
>> The first one is slower but we don't keep "struct device" pointer in
>> abstraction struct. Situation is opposite for the second one.
>>
>> Does anyone care? ;)
>
> Put a struct device pointer into b43_wldev.
> The pointer is needed on every DMA transaction several times, so this
> seems worth it.

Oops, OK, that was example. I thought DMA uses "dma_dev" frequently, not "dev".

But if we assume there is some not frequently used struct. What
solution should be used then?

-- 
Rafa?

Re: Optimization: keeping pointer all the time, or defining getter?

[Rafał Miłecki]

W dniu 11 maja 2011 18:27 użytkownik Michael Büsch <m@bues.ch> napisał:
> On Wed, 2011-05-11 at 18:22 +0200, Rafał Miłecki wrote:
>> I try to abstract bus in b43 driver. For frequently used fields
>> solution is obvious:
>> At init time I get info from bus-specific stuct and put it in generic field.
>> Example: core revision. We often refer to it, I put it in abstraction struct.
>>
>> My question: what is the policy for cases with less frequently used fields?
>> Example: I need to use "struct device" for registering some low lever stuff.
>>
>> I have two solutions:
>>
>> 1) Realtime:
>> struct device *get_device(...)
>> {
>> if (dev->bus_type = B43_BUS_SSB)
>> return dev->ssbdev->device;
>> else if (dev->bus_type = B43_BUS_BCMA)
>> return dev->bcmadev->device;
>> }
>>
>> 2) Init time:
>> void b43_ssb_init(struct ssb_device *ssbdev)
>> {
>> dev->struct = ssbdev->device;
>> }
>>
>> The first one is slower but we don't keep "struct device" pointer in
>> abstraction struct. Situation is opposite for the second one.
>>
>> Does anyone care? ;)
>
> Put a struct device pointer into b43_wldev.
> The pointer is needed on every DMA transaction several times, so this
> seems worth it.

Oops, OK, that was example. I thought DMA uses "dma_dev" frequently, not "dev".

But if we assume there is some not frequently used struct. What
solution should be used then?

-- 
Rafał

Optimization: keeping pointer all the time, or defining getter?

[Michael Büsch]

On Wed, 2011-05-11 at 18:34 +0200, Rafa? Mi?ecki wrote: 
> W dniu 11 maja 2011 18:27 u?ytkownik Michael B?sch <m@bues.ch> napisa?:
> > On Wed, 2011-05-11 at 18:22 +0200, Rafa? Mi?ecki wrote:
> >> I try to abstract bus in b43 driver. For frequently used fields
> >> solution is obvious:
> >> At init time I get info from bus-specific stuct and put it in generic field.
> >> Example: core revision. We often refer to it, I put it in abstraction struct.
> >>
> >> My question: what is the policy for cases with less frequently used fields?
> >> Example: I need to use "struct device" for registering some low lever stuff.
> >>
> >> I have two solutions:
> >>
> >> 1) Realtime:
> >> struct device *get_device(...)
> >> {
> >> if (dev->bus_type = B43_BUS_SSB)
> >> return dev->ssbdev->device;
> >> else if (dev->bus_type = B43_BUS_BCMA)
> >> return dev->bcmadev->device;
> >> }
> >>
> >> 2) Init time:
> >> void b43_ssb_init(struct ssb_device *ssbdev)
> >> {
> >> dev->struct = ssbdev->device;
> >> }
> >>
> >> The first one is slower but we don't keep "struct device" pointer in
> >> abstraction struct. Situation is opposite for the second one.
> >>
> >> Does anyone care? ;)
> >
> > Put a struct device pointer into b43_wldev.
> > The pointer is needed on every DMA transaction several times, so this
> > seems worth it.
> 
> Oops, OK, that was example. I thought DMA uses "dma_dev" frequently, not "dev".

Ah well. I don't know. Isn't dma_dev part of some SSB struct?

> But if we assume there is some not frequently used struct. What
> solution should be used then?

Apply common sense. ;)

-- 
Greetings Michael.

Re: Optimization: keeping pointer all the time, or defining getter?

[Michael Büsch]

On Wed, 2011-05-11 at 18:34 +0200, Rafał Miłecki wrote: 
> W dniu 11 maja 2011 18:27 użytkownik Michael Büsch <m@bues.ch> napisał:
> > On Wed, 2011-05-11 at 18:22 +0200, Rafał Miłecki wrote:
> >> I try to abstract bus in b43 driver. For frequently used fields
> >> solution is obvious:
> >> At init time I get info from bus-specific stuct and put it in generic field.
> >> Example: core revision. We often refer to it, I put it in abstraction struct.
> >>
> >> My question: what is the policy for cases with less frequently used fields?
> >> Example: I need to use "struct device" for registering some low lever stuff.
> >>
> >> I have two solutions:
> >>
> >> 1) Realtime:
> >> struct device *get_device(...)
> >> {
> >> if (dev->bus_type = B43_BUS_SSB)
> >> return dev->ssbdev->device;
> >> else if (dev->bus_type = B43_BUS_BCMA)
> >> return dev->bcmadev->device;
> >> }
> >>
> >> 2) Init time:
> >> void b43_ssb_init(struct ssb_device *ssbdev)
> >> {
> >> dev->struct = ssbdev->device;
> >> }
> >>
> >> The first one is slower but we don't keep "struct device" pointer in
> >> abstraction struct. Situation is opposite for the second one.
> >>
> >> Does anyone care? ;)
> >
> > Put a struct device pointer into b43_wldev.
> > The pointer is needed on every DMA transaction several times, so this
> > seems worth it.
> 
> Oops, OK, that was example. I thought DMA uses "dma_dev" frequently, not "dev".

Ah well. I don't know. Isn't dma_dev part of some SSB struct?

> But if we assume there is some not frequently used struct. What
> solution should be used then?

Apply common sense. ;)

-- 
Greetings Michael.