Re: [PATCH][gdb/testsuite] Fail in gdb_compile if nopie results in PIE executable

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

On 16-08-19 20:48, Pedro Alves wrote:
> On 7/23/19 8:22 AM, Tom de Vries wrote:
>> When running gdb.base/dump.exp with --target_board=unix/-fPIE/-pie, we get:
>> ...
>> Running gdb/testsuite/gdb.base/dump.exp ...
>> FAIL: gdb.base/dump.exp: dump array as value, intel hex
>> ...
>>
>> The FAIL happens because although the test specifies nopie, the exec is
>> in fact compiled as PIE.  The "-fPIE -pie" options specified using the
>> target_board are interpreted by dejagnu as multilib_flags, and end up
>> overriding the nopie flags.
> 
> I'd think it would be better to temporarily strip out -fPIE/-pie (*)
> from multilib_flags if nopie is set?
> 
> (*) - and/or the contents of gdb,pie_ldflag gdb,pie_flag.
>  

I'm not sure about that.

I think the rationale you're applying here is to test as much as
possible in a single run: testing more is better.

But AFAIU, the implicit assumption about multilib flags is that a range
of multilib flags is tested (because multilib flags select different
libraries, so in order to claim complete testing you'd have to test all
libraries). In other words, taking the -m32/-m64 example, that you'd
test both with --target_board='unix/-m64' and
--target_board='unix/-m32'. So, if a test-case only works for -m32, then
forcing it to -m32 for unix/-m64 only makes sure you run the test twice
in identical fashion. In which case testing more is not better, just longer.

[ The same approach is used in gcc testing: a testcase can require a
certain effective target, and multilib flags influence whether the
effective target is available or not, and if not, the test is skipped as
unsupported. ]

Thanks,
- Tom