Re: [PATCH 0/5] use mask operations instead of test_bit()

[Juergen Gross <jgross@suse.com>]

On 10/02/2015 11:03 AM, Dario Faggioli wrote:
> On Fri, 2015-10-02 at 06:40 +0200, Juergen Gross wrote:
>> Instead of using test_bit() which is an atomic operation and limits
>> the compiler's choices to do optimization, use logical ANDs with
>> bitmasks where possible.
>>
> That's a good idea, I think.

It's a fallout from a cleanup patch:

http://lists.xen.org/archives/html/xen-devel/2015-09/msg03184.html

>
>> The possible candidates have been detected by searching definitions
>> of bitmasks in the form:
>>
>> #define MASK  1 << _MASK
>>
> Right.
>
>> On x86 the resulting code is slightly smaller (about 2 bytes for each
>> case, checked via disassembly in few examples).
>>
>> I'm quite sure I didn't replace a test_bit() call required to be
>> atomic, but I'd be grateful for a thorough review especially in the
>> scheduler.
>>
> I'll have a deep look.

Thanks.

> One question, can we introduce a __test_bit() macro/inline function,
> like Jan did with __set_bit?

In the thread mentioned above you'll find a discussion about exactly
this idea between Jan and me.

> I've quickly-&-dirtily tested this:
>
> #define __test_bit(nr, addr) ({  \
>      unsigned _flags = 1 << nr;   \
>      addr & _flags; \
> })
>
> and the result (I've checked a couple of cases) seems the same to me.

The problem is the limited scope where this scheme is really working and
is a better solution at the same time (nr must be a constant less than
the numbers of bits of *addr).


Juergen