[v2,1/3] bitops: fix types

From: Blue Swirl <blauwirbel@gmail.com>

Use 'unsigned int' for bit numbers instead of 'unsigned long' or
'int'. Adjust asserts.

Signed-off-by: Blue Swirl <blauwirbel@gmail.com>
---
 bitops.h |   46 ++++++++++++++++++++++++++--------------------
 1 files changed, 26 insertions(+), 20 deletions(-)

blauwirbel@gmail.com writes:

> From: Blue Swirl <blauwirbel@gmail.com>
>
> Use 'unsigned int' for bit numbers instead of 'unsigned long' or
> 'int'. Adjust asserts.

I'd like to lodge a formal objection to this part.

There is no consensus.  I recognize the power of maintainers to force a
change even without consensus.  Use it wisely.

On Mon, Jul 9, 2012 at 7:49 AM, Markus Armbruster <armbru@redhat.com> wrote:
> blauwirbel@gmail.com writes:
>
>> From: Blue Swirl <blauwirbel@gmail.com>
>>
>> Use 'unsigned int' for bit numbers instead of 'unsigned long' or
>> 'int'. Adjust asserts.
>
> I'd like to lodge a formal objection to this part.
>
> There is no consensus.  I recognize the power of maintainers to force a
> change even without consensus.  Use it wisely.

I thought I refuted all concrete arguments except performance.
However, Avi kindly provided this part.

On 10 July 2012 20:18, Blue Swirl <blauwirbel@gmail.com> wrote:
> On Mon, Jul 9, 2012 at 7:49 AM, Markus Armbruster <armbru@redhat.com> wrote:
>> There is no consensus.  I recognize the power of maintainers to force a
>> change even without consensus.  Use it wisely.
>
> I thought I refuted all concrete arguments except performance.

No, you made various claims that Markus and I at least
disagreed with. (Conversely, we have made various claims
that you disagree with -- this is what "no consensus" means...)

> However, Avi kindly provided this part.

One extra instruction, which isn't a load/store or branch? This is the
worst kind of premature microoptimisation. If we're changing this
it should be rooted in arguments about maintainability or similar,
or alternatively if it's really a performance improvement I'm
sure you can produce the benchmarks that demonstrate that :-)

[Plus the extract/deposit ops aren't doing array accesses!]

-- PMM

On Tue, Jul 10, 2012 at 7:37 PM, Peter Maydell <peter.maydell@linaro.org> wrote:
> On 10 July 2012 20:18, Blue Swirl <blauwirbel@gmail.com> wrote:
>> On Mon, Jul 9, 2012 at 7:49 AM, Markus Armbruster <armbru@redhat.com> wrote:
>>> There is no consensus.  I recognize the power of maintainers to force a
>>> change even without consensus.  Use it wisely.
>>
>> I thought I refuted all concrete arguments except performance.
>
> No, you made various claims that Markus and I at least
> disagreed with. (Conversely, we have made various claims
> that you disagree with -- this is what "no consensus" means...)

You did not present any concrete arguments. In this review you have
pointed at bugs in the assert() expression, thanks.

>
>> However, Avi kindly provided this part.
>
> One extra instruction, which isn't a load/store or branch? This is the
> worst kind of premature microoptimisation. If we're changing this
> it should be rooted in arguments about maintainability or similar,

I think signed and unsigned are pretty much equal in this respect.
Both signed and unsigned have corner cases. Unsigned has the advantage
of being more natural range for values which can't be negative by
their nature, like bit numbers. There's also the security argument
like Avi mentioned.

> or alternatively if it's really a performance improvement I'm
> sure you can produce the benchmarks that demonstrate that :-)
>
> [Plus the extract/deposit ops aren't doing array accesses!]

It's a clear benefit which signed integers can't provide, especially
on x86_64 which is an important platform. For me this clearly tips the
balance.

>
> -- PMM

On 10 July 2012 21:01, Blue Swirl <blauwirbel@gmail.com> wrote:
> On Tue, Jul 10, 2012 at 7:37 PM, Peter Maydell <peter.maydell@linaro.org> wrote:
>> On 10 July 2012 20:18, Blue Swirl <blauwirbel@gmail.com> wrote:
>>> On Mon, Jul 9, 2012 at 7:49 AM, Markus Armbruster <armbru@redhat.com> wrote:
>>>> There is no consensus.  I recognize the power of maintainers to force a
>>>> change even without consensus.  Use it wisely.
>>>
>>> I thought I refuted all concrete arguments except performance.
>>
>> No, you made various claims that Markus and I at least
>> disagreed with. (Conversely, we have made various claims
>> that you disagree with -- this is what "no consensus" means...)
>
> You did not present any concrete arguments. In this review you have
> pointed at bugs in the assert() expression, thanks.

No, a bug in your assert is an indication of why there are
downsides to making random changes, not a concrete argument
against making the changes. The major problem here
is (a) there is no really good reason to make this change
(b) it's moving away from the existing tested code we have
(and that the kernel has). Basically 'int' has more natural
behaviour for reasoning about than 'unsigned' in ranges
where it's usually used (ie small ones).

-- PMM

Am 08.07.2012 21:22, schrieb blauwirbel@gmail.com:
> From: Blue Swirl <blauwirbel@gmail.com>
> 
> Use 'unsigned int' for bit numbers instead of 'unsigned long' or
> 'int'. Adjust asserts.
> 
> Signed-off-by: Blue Swirl <blauwirbel@gmail.com>

I haven't followed the original discussion and therefore don't know what
the controversy is about (nor do I feel like reading it up), but if
there is no consensus, I would expect even more than already for normal
patches that the commit message doesn't only state the obvious change,
but also the reasons for the change.

This message isn't much different from the famous "i++; /* increase i by
one */" code comment.

Kevin

On Tue, Jul 10, 2012 at 9:36 PM, Peter Maydell <peter.maydell@linaro.org> wrote:
> On 10 July 2012 21:01, Blue Swirl <blauwirbel@gmail.com> wrote:
>> On Tue, Jul 10, 2012 at 7:37 PM, Peter Maydell <peter.maydell@linaro.org> wrote:
>>> On 10 July 2012 20:18, Blue Swirl <blauwirbel@gmail.com> wrote:
>>>> On Mon, Jul 9, 2012 at 7:49 AM, Markus Armbruster <armbru@redhat.com> wrote:
>>>>> There is no consensus.  I recognize the power of maintainers to force a
>>>>> change even without consensus.  Use it wisely.
>>>>
>>>> I thought I refuted all concrete arguments except performance.
>>>
>>> No, you made various claims that Markus and I at least
>>> disagreed with. (Conversely, we have made various claims
>>> that you disagree with -- this is what "no consensus" means...)
>>
>> You did not present any concrete arguments. In this review you have
>> pointed at bugs in the assert() expression, thanks.
>
> No, a bug in your assert is an indication of why there are
> downsides to making random changes, not a concrete argument
> against making the changes. The major problem here
> is (a) there is no really good reason to make this change

There are inconsistencies in the API that need fixing (unsigned long
vs. int) and a technically improved solution (unsigned int in x86_64)
exists. These are good and valid reasons.

> (b) it's moving away from the existing tested code we have
> (and that the kernel has).

Keeping the code intact should not stop progress: fix inconsistencies
and use a better solution. The distance from kernel is questionable
value BTW: have you checked how little bitops.h resembles the kernel
versions?

> Basically 'int' has more natural
> behaviour for reasoning about than 'unsigned' in ranges
> where it's usually used (ie small ones).

But 'unsigned' is much more naturally suited to values that can't be
negative. This part is bikeshedding, your point of view against mine.
Currently some functions use 'unsigned long' for bit numbers,
shouldn't they be changed to 'int' in your logic?

>
> -- PMM

On Wed, Jul 11, 2012 at 12:49 PM, Kevin Wolf <kwolf@redhat.com> wrote:
> Am 08.07.2012 21:22, schrieb blauwirbel@gmail.com:
>> From: Blue Swirl <blauwirbel@gmail.com>
>>
>> Use 'unsigned int' for bit numbers instead of 'unsigned long' or
>> 'int'. Adjust asserts.
>>
>> Signed-off-by: Blue Swirl <blauwirbel@gmail.com>
>
> I haven't followed the original discussion and therefore don't know what
> the controversy is about (nor do I feel like reading it up), but if
> there is no consensus, I would expect even more than already for normal
> patches that the commit message doesn't only state the obvious change,
> but also the reasons for the change.
>
> This message isn't much different from the famous "i++; /* increase i by
> one */" code comment.

The message could be improved by vast amounts, but in my view it is
sufficient for such a simple change.

>
> Kevin

On 12 July 2012 21:18, Blue Swirl <blauwirbel@gmail.com> wrote:
> On Tue, Jul 10, 2012 at 9:36 PM, Peter Maydell <peter.maydell@linaro.org> wrote:
>> Basically 'int' has more natural
>> behaviour for reasoning about than 'unsigned' in ranges
>> where it's usually used (ie small ones).
>
> But 'unsigned' is much more naturally suited to values that can't be
> negative. This part is bikeshedding, your point of view against mine.

I agree that this is all veering quite close to bikeshedding, so
this email is my last on the subject.

So, both 'unsigned' and 'int' have two discontinuities, at the extreme
ends of their ranges. For 'unsigned' these are at 0 and UINT_MAX;
for 'int' they're at INT_MIN and INT_MAX. Discontinuities are bad
because they break our intuitive reasoning about the behaviour of
integers and result in bugs which can easily slip through code review.
Any time you might be near a discontinuity you have to think carefully
about what is actually going to happen. Sometimes that's unavoidable
whatever type you pick (eg in situations where values are coming from
an untrusted source which might be trying to find a security hole).
But often values are from another trusted area of the code (or even
obviously small by inspection of the immediate area). In this case
'int' is good because it keeps the discontinuities well away from the
range you'll be working in. Because 'unsigned' has a discontinuity
near 0 it's much easier to bump into it even for numbers in the
expected use range. So when you're really just dealing with "small
numbers that happen to be positive" and don't need the unsigned
arithmetic semantics, 'int' is the better choice.
I think the maintainability/reduced-bugginess argument hugely
outweighs minor details of one insn more or less in some operations.

(If we care about that, we should be much more in favour of using 'int'
returns rather than 'bool', since the bool return forces the compiler
to insert code which normalises a zero-or-nonzero value to zero-or-one.
But there too the readability and maintainability benefits are worth
taking the unmeasurably tiny extra execution time.)

> Currently some functions use 'unsigned long' for bit numbers,
> shouldn't they be changed to 'int' in your logic?

I think that falls into the category of things I would recommend
changing in code review, or if fixing some related kind of
overflow bug, but would not change existing working code for.

-- PMM

Am 12.07.2012 22:21, schrieb Blue Swirl:
> On Wed, Jul 11, 2012 at 12:49 PM, Kevin Wolf <kwolf@redhat.com> wrote:
>> Am 08.07.2012 21:22, schrieb blauwirbel@gmail.com:
>>> From: Blue Swirl <blauwirbel@gmail.com>
>>>
>>> Use 'unsigned int' for bit numbers instead of 'unsigned long' or
>>> 'int'. Adjust asserts.
>>>
>>> Signed-off-by: Blue Swirl <blauwirbel@gmail.com>
>>
>> I haven't followed the original discussion and therefore don't know what
>> the controversy is about (nor do I feel like reading it up), but if
>> there is no consensus, I would expect even more than already for normal
>> patches that the commit message doesn't only state the obvious change,
>> but also the reasons for the change.
>>
>> This message isn't much different from the famous "i++; /* increase i by
>> one */" code comment.
> 
> The message could be improved by vast amounts, but in my view it is
> sufficient for such a simple change.

No, it's not. The change is simple (so you don't necessarily need to
repeat what has changed, I see it in the diff), but the reasons aren't
obvious. So please state them even for totally simple mechanical changes.

Kevin