[tree-ssa] Bootstrap fails with undefined reference to find_reachable_lable

[Tim Josling]

> On Thursday 24 April 2003 22:28, Diego Novillo wrote:
> 
>> On Thu, 2003-04-24 at 16:06, law@redhat.com wrote:
> 
>> > In message <1051213400.15231.85.camel@frodo.toronto.redhat.com>, Diego
>> > Novillo
>> >
>> > writes:
>> >  >On Thu, 2003-04-24 at 15:17, Andreas Jaeger wrote:
>> >  >> -----BEGIN PGP SIGNED MESSAGE-----
>> >  >> Hash: SHA1
>> >  >>
>> >  >>
>> >  >> bootstrap of current tree-ssa fails on i686-linux-gnu for me with:
>> >  >
>> >  >Ah, yes.  I've got a local fix that I still haven't flushed out.  Why
> 
> 
> Can you send out the fix?  I'm disabling tree-lang in my tree-ssa testing now.
> 
> 
>> >  > is it that treelang uses parts of the C front end, but not all?  It's
>> >  > quite annoying and easy to break.
>> >
>> > You're kidding, right?  [ Clearly you're not, but I couldn't think of a
>> > better way to express my surprise that something used c-simplify, but
>> > did not use c-semantics. ]
> 
>>
>> I wish I was :(  It may be a misunderstanding on my side, but I have
>> seen function duplication in treelang/treetree.c that merely call
>> abort.  Previously I had to fix other link errors for treelang's front
>> end that essentially did the same thing: create an empty function in
>> treetree.c.
>>
> 
>> > This is a problem with treelang's implementation that the treelang folks
>> > really need to fix.
> 
>>
>> There was a thread discussing this issue recently:
>> http://gcc.gnu.org/ml/gcc-patches/2003-03/msg01444.html
>>
>> Perhaps treelang should be changed to use language hooks or generate
>> GIMPLE or ... something, I don't know.
> 
> 
> Tim, could you look into these issues and design a cleaner interface for 
> treelang?  With every major change treelang breaks because the interface is 
> not well-defined,
> 
> Andreas
> - -- 
>  Andreas Jaeger
>   SuSE Labs aj@suse.de
>    private aj@arthur.inka.de
>     http://www.suse.de/~aj

The problem is that the interface between the back end and the front 
ends is not clean. This is not a treelang specific problem. Current 
languages solve the problem in two ways

1. Objective C, C++ - by sharing code with C though the code is not 
properly abstracted out into a language independent layer.
2. Fortran, etc - by duplicating code.

Both these are bad solutions. Treelang uses option (1).

A lot of code that is common to numerous languages is in c-* files. 
There is c-specific stuff in the back end as well. Example: The rules on 
deferred definition are straight C. Example: PCH spews a whole lot of C 
specific PCH stuff into every front end, whether they have restartable 
compilation / PCH or not.

The choices I have are

- Duplicate a whole lot of code that's in c-*, and manage the fractal 
interface of that code into the back end. I did try that by the way, and 
it was a lot of work and very painful.

- The current solution. Reuse some c-* code. This is not ideal but it 
involves far fewer patches.I you have a look at the changelog there is 
not a huge lot of this. Personally I think that this is not a big 
problem, it is just that the idea of it does not appeal.

- Restructure the c-* code and the back end to make a clean interface

Option 3 is the *right thing* but it is a huge job, particularly without 
having a mindset change about GCC being a C compiler. I am also of the 
view that any attempt by me to do this would not succeed and I have 
other priorities.

I am more than happy for someone to implement option (1) as long as they 
commit to keep it alive. That would be a lot of work.

Perhaps someone can find a better solution but at this stage I am stuck.

Tim Josling