Re: [PATCH] Disable branch profiling macros when sparsed.

["Alexey Zaytsev" <alexey.zaytsev@gmail.com>]

On Sat, Jan 10, 2009 at 10:18, Harvey Harrison
<harvey.harrison@gmail.com> wrote:
> On Fri, 2009-01-09 at 22:13 -0800, David Miller wrote:
>> From: Alexey Zaytsev <alexey.zaytsev@gmail.com>
>> Date: Sat, 10 Jan 2009 08:57:28 +0300
>>
>> > The macros produce lots of unneeded warnings when
>> > recursive if(({ .. if() {..} ..})) {..} and such
>> > are substituted. And there is no point in sparsing
>> > them anyway. This is useful if someone decides to
>> > sparse an allyesconfig kernel.
>> >
>> > Signed-off-by: Alexey Zaytsev <alexey.zaytsev@gmail.com>
>>
>> If even sparse can't handle these things, it's no surprise
>> how many gcc bogus warning problems we've run into because
>> of this hairy if() macro.
>
> It's not that sparse can't handle it, the warning is valid,
> _____r and ______f are shadowed when these get nested.  It
> gets even worse when interacting with likely/unlikely tracing
> as that chose the same identifiers too.  So there the noise
> could be drastically reduced changing the different identifiers
> for the if () and __branch_check macros, but nesting will always
> warn.
>
> I've just been setting this to no in my allyesconfig sparse
> runs....just wait until kmemtrace gets to mainline, then it
> gets really bad :(
>

I don't really understand what is bad here. The 'unlikely' and 'if'
trace implementation looks quite elegant to me. Yes, they generate
10kbyte spaghetti monsters (in C) for a simple WARN_ON_ONCE(),
but probably we should just remove a few unlekely() from the WARN_*
code, and I'm not sure it's even worth it. There would be no direct
speedup.

And it took only one line to disable.