Re: Removing the TARGET_* layer or not ?

[Michael Koch]

Am Mittwoch, 4. August 2004 10:06 schrieb Ingo Prötel:
> Hi,
>
> Roman Kennke wrote:
> > Am Di, den 03.08.2004 schrieb Ingo Prötel um 15:39:
> >>Hi Michael,
> >>
> >>Michael Koch wrote:
> >>>Hi list, hi Ingo,
> >>>
> >>>
> >>>
> >>>I just started porting/testing GNU classpath to solaris (2.6).
> >>> Its not really different to other unices (we currently support
> >>> linux and *BSD) but some things are. When I wanted to do the
> >>> needed changes I stumbled over the TARGET_* layer which adds
> >>> (in my eyes) some unneeded complexity. It makes it more hard to
> >>> read the code and all the stuff seem to be implemented either
> >>> in target/generic or target/Linux with no real rule what have
> >>> to go where. It's just difficult to understand. In fact I try
> >>> to understand it since a long time and always fail because of
> >>> its "indirectness". Personally I think some AUTOCONF checks
> >>> would be more appropriate and would make the code much more
> >>> readable and bugfixable. There are some known bugs in it but
> >>> noone attacked them because none understands the code it seems.
> >>>
> >>>Ingo: Now my question are you really using the TARGET_* system
> >>> or is it only rotting around in GNU classpath ? I really wonder
> >>> if some AUTOCONF macros would be more helpful for you ?
> >>
> >>Yes, the TARGET-layer is in active use here.
> >>
> >>The TARGET-layer has two advantages:
> >>1. It makes JNI-code more readable,
> >>2. It makes porting simple.
> >>
> >>It makes JNI-code more readable because whenever a native
> >> function is called there is only one macro call. The other
> >> possibility would be that one has ifdefs for every system
> >> (possibly for every system+architecture) that one wants to
> >> support with this code. This would make the JNI-code difficult
> >> to read.
> >>
> >>It makes porting easier in that one can reuse generic macros and
> >> only needs to override specific macros that differ for a given
> >> new system. And this all can happen without modifing the actual
> >> JNI-code. So the main step in porting is creating a new
> >> subdirectory in tartget name it Solaris. Copy over all the files
> >> in the Linux target. Add this to the include path. Then just
> >> work on the error that the compiler will find and where
> >> necessary create Solaris specific macros.
> >>
> >>If there is interest we could provide our Solaris, Darwin, and
> >> MinGW TARGET-layer. This would probably help to understand how
> >> this layer is helpful. In any case we can help if anyone needs
> >> help understanding the code.
> >
> > I have thought about how I would design this stuff in order to
> > write portable code:
> > 1. first of all, try to stick to POSIX standard stuff ;)
> > 2. Of course there are situations where this won't do. In this
> > case I would split out the non-portable stuff on a case-by-case
> > basis in the following scheme: Suppose you have the following
> > expressions in a function (pseudocode):
> >
> > func() {
> >   exp_a;
> >   exp_b;
> >   exp_c;
> > }
> >
> > of which exp_b is not portable. Then make a function call out of
> > it:
> >
> > func() {
> >   exp_a;
> >   do_b();
>
>      TARGET_NATIVE_B();  // this is what it currently looks like
>
> >   exp_c;
> > }
> >
> > and define the function in a file in a file in an OS-specific
> > subdir.
>
> This is actually what the whole idea is about. Except that macros
> are used instead of functions. The problem with functions is that
> many embedded compilers cannot inline function calls. This would be
> ok if the overhead would really be only one function call but if
> you got through the JNI-code and replace every TARGET-macro with a
> function call you will see that there are more function calls
> involved. Maybe we could get some mixed mode where we leave the
> macros in place and the macros for Linux just point to functions.
> This would make the Linux code more debuggable.
>
> > the autotools should check for OSs/arches and include the correct
> > subdirs here. These functions must be kept minimal, in order to
> > avoid the effect that Michael mentioned, where you have to fix a
> > bug multiple times.
>
> If one fixes a bug in the generic part it will be automatically
> picked up in every system that uses the generic stuff. If a system
> does not use the generic stuff most likely there will be no system
> that will depend on this bug fix. Especially if one fixes a bug
> somewhere this might as a side effect break another system that
> depends on the bogous behavior. There is no way of testing it if
> one doesn't have a target setup to test it on. So sometimes this
> sort of isolation is good.
>
> > Note, that with this system we would avoid the use of macros
> > altogether (You see, I don't like them ;) I better go with a
> > well-thought-out design). Sure, there is the 'overhead' of 1
> > additional function call.
>
> That is basically on function call for every TARGET-macro.

I'm not a friend of this one-OS-one-dir approach. I would more test 
for features with autoconf and use the features if present. This 
makes it more easy to port to another arch. When the feature is 
available there, use it. If not, add some code to make it work.

The TARGET api is really not intutive and the very long MACRO names 
makes it even harder to understand. Sometimes the names just badly 
choosed. The "Linux" layer is full of bugs. And I suppose your other 
ports are too because they are copies of the "Linux" layer. Some bugs 
are fixed in the classpath version of "Linux". I wonder how many of 
them were ported to you other ports. When an autoconf approach you 
wouldnt even think about this. I know that with the autoconf approach 
the patches need a closed review and more testing before they go in 
but IMO this is much better then duplicatin bugs over several trees.

I was pretty surprised when ypu wrote yesterday, that you have more 
"ports" then just "Linux". Why don't you submitted them ? I think it 
would be a very good idea to support as much OSes upstream as 
possible, even Plan9 and QNX and ...


Michael