GDB testsuite overrides default_target_compile and breaks

[Jacob Bachmeyer]

[previous Subject:  Re: dejagnu version update? [CORRECTION: not a 
regression in DejaGnu; GDB testsuite bug] ]
[adding Tom Tromey to CC list per advice from Rob Savoye]
[adding main DejaGnu mailing list to CC list]

In brief for those new to this, the GDB testsuite is currently broken 
when run with the current DejaGnu Git master.  The problem was initially 
reported off-list and traced to a copy of default_target_compile in the 
GDB testsuite.  A recent change to DejaGnu internals included a change 
to that procedure upstream.  Cleanup enabled by that change breaks the 
old copy of the code in the GDB testsuite.

What is needed in brief:
(1) Merge the features of default_target_compile that the GDB testsuite 
depends on upstream for 1.6.3.
(2) Find the actual extensibility that GDB needs here and add that 
support to the default_target_compile rewrite slated for 1.6.4.

Maciej W. Rozycki wrote:
> [cc-ing Joel as the originator, and <gdb@sourceware.org>]
>
> On Tue, 9 Jun 2020, Jacob Bachmeyer wrote:
>   
>>>> I ran a quick bisection and the culprit turned out to be:
>>>>
>>>> ba60272a5ac6f6a7012acca03f596a6ed003f044 is the first bad commit
>>>> commit ba60272a5ac6f6a7012acca03f596a6ed003f044
>>>> Author: Jacob Bachmeyer <jcb62281+dev@gmail.com>
>>>> Date:   Mon May 25 08:40:46 2020 -0600
>>>>
>>>>      Establish a default C compiler by evaluating [find_gcc] if no 
>>>> other compiler is given.
>>>>
>>>>  So this regression has to be fixed before any new release is made.
>>>>         
>>> I will look into this.  So far, I have confirmed that the problem does 
>>> occur and that setting the "compiler" board_info key in 
>>> baseboards/unix.exp seems to avoid it.  It looks like the default is 
>>> not being selected in all cases where it should be used.  I still need 
>>> to find the exact problem to write a regression test to isolate this 
>>> bug and make it stay squashed.
>>>       
>> After further efforts, and a few hours wasted trying to figure out why 
>> additional tracing code added to default_target_compile was not 
>> producing output, I eventually decided to just make 
>> default_target_compile throw a Tcl error -- and I did not get a Tcl 
>> error when running the tests...
>>
>> That is "very interesting" and a quick grep -R reveals the culprit:  the 
>> GDB testsuite is overriding default_target_compile in 
>> gdb/testsuite/lib/future.exp.  Comments indicate that that file was 
>> intended to temporarily bundle certain features with the GDB testsuite 
>> that had not yet been merged into upstream DejaGnu -- in 2004.  Now 
>> DejaGnu core is continuing to develop, leaving the old code copied into 
>> the GDB testsuite broken, and making this NOTOURBUG.
>>     
>
>  Well, as the name suggests (regrettably there's no adequate documentation 
> there) this procedure is there to be overridden.

No, *_hook indicates a procedure that is there to be overridden, but not 
by testsuites.  Overriding procedures inside DejaGnu is generally for 
testing lab customization, because a site can update their init files 
when they update DejaGnu, but testsuites need to work with multiple 
versions of DejaGnu.

>   The way it's done in GDB 
> might perhaps be non-standard, as the standard way AFAICS would be by 
> providing a `gdb_compile' handler, but I think this is tangential, and the 
> chosen solution is there possibly because DejaGnu had no `${tool}_compile' 
> knob back then (I haven't checked).  That does not really matter though, 
> as the net effect would be the same.
>   

DejaGnu does not have a ${tool}_compile -- that would be a legitimate 
procedure for a testsuite to define and use, and I believe that the GDB 
testsuite does so.  There is a ${target}_compile customization point 
that is intended for testing labs to override for specific target 
boards.  (It is not quite correctly handled in target_compile at the 
moment, but this is slated to be fixed for 1.6.4.)  The 
default_target_compile procedure is intended to be the default handler 
for target_compile for targets that do not have their own special 
compilation procedures.

>> In short, this is not a regression in DejaGnu; this is code duplicated 
>> into the GDB testsuite that was slated for removal from that location 
>> years ago and needs to be removed from the GDB testsuite, or at least 
>> made conditional on running under a version of DejaGnu old enough to 
>> require it, if such versions are still supported for running the GDB 
>> testsuite.  If that code has added features not present in upstream 
>> DejaGnu over the years, now is the time to get those patches in.
>>     
>
>  Well, it clearly is a functional regression as the interface has changed, 
> even if not a formal one, and I would consider it a release stopper.
>   

The changes were to an internal component and broke an out-of-tree copy 
of same.  I have to draw a line somewhere, and "monkeypatching the core 
is not supported and can break" is necessary for DejaGnu to develop at 
all.  Add that comments in gdb/testsuite/lib/future.exp clearly state 
(from 2004! 16 years ago!) that that entire file was supposed to only 
temporarily exist until its features were merged upstream and I say that 
now is the time to pick that dropped ball back up and get those features 
merged upstream.

>  As it happens there are about 2 users of DejaGnu there, which means any 
> fatal regression would have been easily caught in the course of change 
> verification (and running binutils/GDB and GCC test suites natively is 
> about as easy as it can be: `configure && make && make check'), and indeed 
> should have, and then sorted with the respective communities if indeed the 
> interface change is unavoidable, with a transition period so that the 
> users can prepare changes to their frameworks, including backports to 
> various release branches if applicable, before the old interface is 
> removed from DejaGnu.

There will certainly be a lengthy deprecation process for the planned 
future API cleanups; indeed, the difference between the last release in 
1.x and 2.0 is planned to be the removal of compatibility shims, but 
that is still very far off.

>   My suggestion would be to revert the change, get 
> the details sorted with GDB people and only reapply it when all parties 
> are ready.
>   

The actual change that broke the out-of-tree copy of 
default_target_compile was cleaning up the board files to remove the 
now-optional definition of the "compiler" board_info key.  (Some 
testsuites do not need to compile code and running find_gcc emits 
useless "verbose noise" in those cases.)  A patch to revert this clean 
up (but not the addition of [find_gcc] as default in 
default_target_compile) may be reasonable on the release branches for 
the remaining 1.6.x releases, since it would not significantly affect 
the core, but the change stays on the master branch.  Monkeypatching the 
DejaGnu core is not supported and cannot be supported long-term.  (In 
the future, default_target_compile may eventually disappear from the 
global namespace as seen by testsuites.  If GDB continues to override 
it, the rename will end up deleting the compatibility shim while the 
internal calls still go to the internal copy of the procedure.  Then it 
all blows up with a Tcl error in DejaGnu 2.0 when the compatibility shim 
is removed.)

GDB seems to have also extended default_target_compile, so a discussion 
on this is needed because there seems to be a need for some kind of 
language extensibility feature to allow new language-default-selector 
options to be defined without overriding the entire 
default_target_compile procedure (which could also break in a test lab 
that uses a custom ${target}_compile procedure for a certain board).  
Since default_target_compile is slated to be rewritten between 1.6.3 and 
1.6.4, now is a very good time to get additional features that it needs 
to have into the DejaGnu core.  The API is now documented as 
"target_compile" in Git master.  Backwards-compatible extensions (like 
adding new language default options) are possible, of course.

If I understand correctly, merging the features currently carried in 
gdb/testsuite/lib/future.exp upstream will cause all existing GDB 
releases to correctly use the upstream version of 
default_target_compile, making the partial reversion patch unnecessary.

>  Anyway, I have completed the verification I have volunteered to do and 
> therefore I'm done with my part.

Thank you for your help.  [To Maciej W. Rozycki:  The requests for 
further efforts in this email are for other people who work more closely 
with GDB.]



-- Jacob