GDB test coverage

GDB test coverage

[Yao Qi]

Hi,
We had the doc on wiki about doing coverage tests to GDB,
https://sourceware.org/gdb/wiki/TestingGDB#Coverage_Testing
and I give it a try last week.  Get some interesting results to share
now,

http://people.linaro.org/~yao.qi/gdb/coverage/20161102/

In general, the coverage is better than I expected :-).  Some initial
analysis to the coverage data,

 - the coverage in linux-record.c is low, because we don't test many
   syscalls in gdb.reverse,
 - the coverage in *-lang.c is low.  Looks we are still lack of language
   related tests.
 - we don't test all the start options in gdb, see the coverage in
   main.c:captured_main_1
 - we don't test "set cp-abi", "set extension-language", "delete checkpoint"
   "detach checkpopint",
 - here is a list of functions are never used,
   main.c:get_gdb_program_name,
   corefile.c:read_stack,
   valarith.c:value_in,
 - linux_gdb_signal_from_target and linux_gdb_signal_to_target are not
   well covered,

I am sure we can still get more from these coverage results.  How do we
make full use of coverage tests?  What I can think of are
 - we can find some never-used functions, and remove some of them,
 - add more test cases to address some low-coverage areas,

-- 
Yao (齐尧)

Re: GDB test coverage

[Matt Rice]

On Tue, Nov 8, 2016 at 9:38 AM, Yao Qi <qiyaoltc@gmail.com> wrote:
>
> I am sure we can still get more from these coverage results.  How do we
> make full use of coverage tests?  What I can think of are
>  - we can find some never-used functions, and remove some of them,
>  - add more test cases to address some low-coverage areas,

A while back I would run a per-commit coverage under something like
the build bot, and commit the results to a git repository, (one of the
output formats of gcov does compress fairly well...)

One thing that I noticed was that coverage would swing fairly wildly
(even when running the coverage under the same checkout of a commit
without changes) due to the races in the testsuite.

I'm not sure if that would still be the case or not, this made it
fairly difficult to isolate the particular effect that any given
commit had on the coverage.
which I thought would have made a nice barometer of sorta...