Re: Test failure of cxxclass

[Mark Wielaard <mjw at redhat dot com>]

Hi Wenji,

On Mon, 2009-12-14 at 10:51 +0800, Wenji Huang wrote:
> I got failed when I ran tests on FC11 32-bits.
> 
> $ stap ./systemtap.base/cxxclass.stp cxxclass.exe -c ./cxxclass.exe
> main_enter
> call: 64
> cons call: 64
> cons call: 64
> meth call: 64 24
> meth call: 64 24
> dest call: 42
> dest call: 42
> call2: 24
> main_exit
> 
> Seems there is double call on process.mark.

Yes, I am seeing the same. Wanted to file a bug report, but bugzilla was
down this weekend. Will do so now.

The issue seems to be that some of the static probe points get
duplicated in the code. This shouldn't be a problem, and the .probes
section does contain both places with different addresses. You can also
see (with enough -vvvvv) that stap finds both locations while resolving
the mark("cons") probe. But, only the first is registered twice, as can
be seen in the generated code:

  { .address=(unsigned
long)0x8048476ULL, .pp="process(\"/home/mark/src/systemtap/testsuite/cxxclass.exe\").statement(134513782)", .ph=&probe_1895, },
  { .address=(unsigned
long)0x8048476ULL, .pp="process(\"/home/mark/src/systemtap/testsuite/cxxclass.exe\").statement(134513782)", .ph=&probe_1895, },

This means the first call of the constructor is triggered as probe twice
(as is the method, and destructor call) but the second call isn't
triggered as probe at all.

It looks like this is a problem in how sdt_query::convert_location() is
called (it seems to do duplication detection based on probe name, but
not address. I haven't looked very deeply into it yet though.

Cheers,

Mark