Re: [committed][gdb/testsuite] Update psym-external-decl.exp for gcc-10/clang

[Tom de Vries <tdevries@suse.de>]

On 6/29/20 2:32 PM, Pedro Alves wrote:
> On 6/28/20 11:50 AM, Tom de Vries wrote:
>> On 6/26/20 11:37 AM, Gary Benson wrote:
>>> Tom de Vries wrote:
>>>> On 6/19/20 4:00 PM, Gary Benson wrote:
>>>>> Tom de Vries wrote:
>>>>>> On 6/18/20 6:10 PM, Gary Benson wrote:
>>>>>>> Tom de Vries wrote:
>>>>>>>> On 6/17/20 2:24 PM, Gary Benson wrote:
>>>>>>>>> Tom, I'd like this testcase to not fail silently.  Is the
>>>>>>>>> functionality under test something that isn't ever
>>>>>>>>> expected to work with clang, or is this a test that should
>>>>>>>>> pass with clang (but it currently doesn't, for whatever
>>>>>>>>> reason)?
>>>>>>>>
>>>>>>>> I'm not sure.  The test can pass with clang, provided it
>>>>>>>> generates the required debug info.  It currently doesn't.
>>>>>>>> Why that is the case, I have no idea.
>>>>>>>
>>>>>>> I think that means the test should work but it doesn't.  Would
>>>>>>> you object if I push a patch removing the test-skipping logic?
>>>>>>> It will mean an extra FAIL when tested using clang
>>>>>>
>>>>>> I don't think having a fail for a compiler bug/missing-feature
>>>>>> is a good idea.
>>>>>>
>>>>>> If this is due to a bug/missing-feature in clang, then we need to:
>>>>>> - xfail the test,
>>>>>> - file the PR in clang, and
>>>>>> - reference the PR at the xfail.
>>>>>
>>>>> Is this a bug/missing feature in clang though?
>>>>> How sure are you GDB isn't at fault?
>>>>
>>>> Clang emits less debug info than GCC.  Whether that's a bug, a
>>>> missing feature or an explicit unsupported feature in clang, I
>>>> don't known.
>>>>
>>>> I known that gdb isn't at fault.  It can't do anything without the
>>>> missing debug info.  The test was specifically written to use that
>>>> debug info.
>>>
>>> I'm not really sure what's the right thing to do here.
>>>
>>> On the one hand, my current task is ensuring GDB can debug
>>> clang-compiled with clang as well as it can debug GCC-compiled
>>> code.  From that perspective the skip-if-clang logic in this
>>> test is hiding a failure I need to investigate.
>>>
>>> On the other hand, I'm an engineer working on GDB, and from that
>>> perspective I want to be able to run the GDB testsuite and see
>>> 100% pass, on whatever setup I test it on.  And yes, I know it
>>> doesn't... but it *should*.
>>>
>>> Is there a way to pass a "don't skip clang failures" flag to the
>>> testcases, such that people running the testsuite normally would
>>> see tests like these return UNSUPPORTED, but I could run the
>>> testsuite with the flag so it'd not skip but FAIL wherever the
>>> problem is?
>>
>> I think the following is a good way of dealing with this.
>>
>> We introduce a proc in gdb.exp called debug_info_for_decl or some such,
>> that returns false by default for clang.
> 
> I think it would be useful to include an intro description to
> 
>  gdb.base/psym-external-decl.exp
> 
> since it currently doesn't describe what it is testing.  Looking
> at the commit log of the patch that introduced it helps, but
> one shouldn't have to do that.
> 
> So the difference between gcc and clang AFAIU is that gcc emits
> debug info for the extern variable declaration (and with newer GCCs,
> only if there's a reference to the variable in the CU, otherwise
> not even so), while seemingly clang would only emit debug info for
> the variable's definition, which isn't compiled with -g, so we
> end up with no debug info for the variable.
> 
> I would suggest filing a bug with clang, to confirm whether
> this is intentional, or whether they see it as a bug.  I would
> think it is a bug, but I'm not sure.  If indeed a bug, we would
> XFAIL the test.
> 

I've filed https://bugs.llvm.org/show_bug.cgi?id=46514 .

Thanks,
- Tom