Re: [PATCH 1/3] xen: add no_instrument_function attributes

[Jan Beulich <jbeulich@xxxxxxxx>]

On 29.07.2024 20:56, Stewart Hildebrand wrote:
> On 7/29/24 10:55, Jan Beulich wrote:
>> On 29.07.2024 16:24, Stewart Hildebrand wrote:
>>> In preparation for using -finstrument-functions option, we need to tag a
>>> few functions that don't work well with such instrumentation. If we
>>> don't intervene, we would end up with linker errors such as undefined
>>> reference to __bad_cmpxchg.
>>
>> I can't spot mention of such a side effect from the documentation. Talk
>> there is of function calls being added at function entry and exit.
>> Nothing is being said that calls to other functions would also be
>> affected.
> 
> Oddly, it seems the compiler fails remove the implementations of the
> affected functions from the object files even though they are not called
> in the generated code (because they were inlined). I can mention this in
> the commit message.
> 
> For example, if no_instrument is omitted from __int_cmpxchg, feeding one
> of the object files through aarch64-none-linux-gnu-objdump -d yields for
> example:
> 
> 0000000000000048 <__int_cmpxchg>:
>      <snip>
>      16c:       94000000        bl      0 <__bad_cmpxchg>
>      <snip>
> 
> Yet, __int_cmpxchg is not called by anything.
> 
> Even more oddly, this appears to be the case for both clang and gcc.

Upon a 2nd, closer look this doesn't appear odd at all. The 1st
argument to the two instrumentation functions is a pointer to the
function, thus requiring instantiation somewhere. Question then is
whether adding the attribute is an appropriate approach to the issue.
Gcc doc kind of suggests to possibly use extern inline instead.

Jan