Re: RFC/Ada: removal of $(ADAC) support

Re: RFC/Ada: removal of $(ADAC) support

[Richard Kenner]

    The systems under discussion ship a "gnatgcc" based on GCC 2.8, but a
    "gcc" based on GCC 2.9x.  This was done for very good reason, that GCC
    2.8 worked fine for Ada but silently miscompiled C source code on at
    least one architecture that that system supports.  Second-guessing the
    system integrator in that scenario, by forcing CC to gnatgcc, would
    raise the risk of bootstrap failure due to a bug in the host compiler.

This isn't something we have to "guess" about.  Has anybody tried the
experiment of bootstrapping starting with that compiler? 

It seems likely it would work since lots of people bootstrapped GCC
starting with 2.8.1.

Re: RFC/Ada: removal of $(ADAC) support

[Zack Weinberg]

kenner@vlsi1.ultra.nyu.edu (Richard Kenner) writes:

>     The systems under discussion ship a "gnatgcc" based on GCC 2.8, but a
>     "gcc" based on GCC 2.9x.  This was done for very good reason, that GCC
>     2.8 worked fine for Ada but silently miscompiled C source code on at
>     least one architecture that that system supports.  Second-guessing the
>     system integrator in that scenario, by forcing CC to gnatgcc, would
>     raise the risk of bootstrap failure due to a bug in the host compiler.
>
> This isn't something we have to "guess" about.  Has anybody tried the
> experiment of bootstrapping starting with that compiler? 

I can't answer this, I only have convenient access to boring platforms
like x86 and sparc.  The problems with 2.8 that I know about were with
arm, mips, etc.

zw

Re: RFC/Ada: removal of $(ADAC) support

[Michael S. Zick]

On Monday 12 May 2003 06:28 pm, Daniel Jacobowitz wrote:
> On Mon, May 12, 2003 at 07:16:01PM -0400, Richard Kenner wrote:
> >     On some popular architectures it can.  Debian did experience a number
> >     of bootstrap failures on other architectures when we frst
> > transitioned to GCC 3.2 however.  It was also particularly sensitive to
> > small perturbations in the compiler being bootstrapped, making it
> >     particularly frustrating.
> >
> >     Sorry, I no longer have the logs, so consider this only informed
> > hearsay.
> >
> > Do you remember which architectures?  Were they ones on which GNAT is
> > supported?
>
> I don't know.  I believe that ARM and/or MIPS were included though.
I had many problems with alpha-freeBSD ...
Never did solve them, so I can't say if it was all cockpit error or that
alpha-freeBSD just didn't work.

Mike

Re: RFC/Ada: removal of $(ADAC) support

[Richard Kenner]

    I don't know.  I believe that ARM and/or MIPS were included though.

I don't know of any ARM Ada host, so the only issue is MIPS, which is
not that common now.  So the only cost of having a common compiler for
front ends and backends is that a bootstrap from MIPS *might* not work
without a more recent compiler.  I think that's a small price to pay for
consistency: remember that if we use $(ADAC) for some C files, it complicates
the Makefiles.

Re: RFC/Ada: removal of $(ADAC) support

[Daniel Jacobowitz]

On Mon, May 12, 2003 at 07:16:01PM -0400, Richard Kenner wrote:
>     On some popular architectures it can.  Debian did experience a number
>     of bootstrap failures on other architectures when we frst transitioned
>     to GCC 3.2 however.  It was also particularly sensitive to small
>     perturbations in the compiler being bootstrapped, making it
>     particularly frustrating.
> 
>     Sorry, I no longer have the logs, so consider this only informed hearsay.
> 
> Do you remember which architectures?  Were they ones on which GNAT is
> supported?

I don't know.  I believe that ARM and/or MIPS were included though.

-- 
Daniel Jacobowitz
MontaVista Software                         Debian GNU/Linux Developer

Re: RFC/Ada: removal of $(ADAC) support

[Richard Kenner]

    On some popular architectures it can.  Debian did experience a number
    of bootstrap failures on other architectures when we frst transitioned
    to GCC 3.2 however.  It was also particularly sensitive to small
    perturbations in the compiler being bootstrapped, making it
    particularly frustrating.

    Sorry, I no longer have the logs, so consider this only informed hearsay.

Do you remember which architectures?  Were they ones on which GNAT is
supported?

Re: RFC/Ada: removal of $(ADAC) support

[Daniel Jacobowitz]

On Mon, May 12, 2003 at 06:25:58PM -0400, Richard Kenner wrote:
>     The difference is that major Linux distributors, and all three of the
>     *BSDs, deliberately shipped a 2.95 "gcc" but a 2.8 "gnatgcc".
>     Precisely because of (1).  That's enough evidence as far as I care.
> 
> That's not my understanding.  They shipped a "2.95" gcc because they wanted
> "the latest and greatest" gcc.  But there was no GNAT for 2.95 so they had no
> choice but to distribute one based on 2.8 since they wanted GNAT there.
> 
> I have seen no evidence that "gnatgcc" can't bootstrap GCC 3.x and indeed
> have seen plenty of evidence that it can.

On some popular architectures it can.  Debian did experience a number
of bootstrap failures on other architectures when we frst transitioned
to GCC 3.2 however.  It was also particularly sensitive to small
perturbations in the compiler being bootstrapped, making it
particularly frustrating.

Sorry, I no longer have the logs, so consider this only informed hearsay.

-- 
Daniel Jacobowitz
MontaVista Software                         Debian GNU/Linux Developer

Re: RFC/Ada: removal of $(ADAC) support

[Richard Kenner]

    The difference is that major Linux distributors, and all three of the
    *BSDs, deliberately shipped a 2.95 "gcc" but a 2.8 "gnatgcc".
    Precisely because of (1).  That's enough evidence as far as I care.

That's not my understanding.  They shipped a "2.95" gcc because they wanted
"the latest and greatest" gcc.  But there was no GNAT for 2.95 so they had no
choice but to distribute one based on 2.8 since they wanted GNAT there.

I have seen no evidence that "gnatgcc" can't bootstrap GCC 3.x and indeed
have seen plenty of evidence that it can.

Re: RFC/Ada: removal of $(ADAC) support

[Zack Weinberg]

kenner@vlsi1.ultra.nyu.edu (Richard Kenner) writes:

>     I reiterate that there is no known problem in this domain other than
>     the one affecting __gnat_raise_exception; you are spreading FUD.
>
> You are doing exactly the same thing by claiming that "gnatgcc" won't
> successfully bootstrap the compiler.
>
> There are two possible sources of problems:
>
> (1) The "gnatgcc" may only compile Ada correctly and not C.
> (2) There may be a layout difference between "gnatgcc" and "gcc" in
>     some structure.

The difference is that major Linux distributors, and all three of the
*BSDs, deliberately shipped a 2.95 "gcc" but a 2.8 "gnatgcc".
Precisely because of (1).  That's enough evidence as far as I care.

zw

Re: RFC/Ada: removal of $(ADAC) support

[Robert Dewar]

> I reiterate that there is no known problem in this domain other than
> the one affecting __gnat_raise_exception; you are spreading FUD.

If you try to use incompatible Ada and C compilers, it might work, but it is
quite possible there wil be problems, and you will be using a combination
that has likely not been tested. The fact that something has not been tested,
does not mean it will not work, but it seems to me better to rely on a tested
combination!

> I disagree that it is unnecessary.  See my response to Kenner.

I still don't understand your argument of necessity here.

Re: RFC/Ada: removal of $(ADAC) support

[Richard Kenner]

    I reiterate that there is no known problem in this domain other than
    the one affecting __gnat_raise_exception; you are spreading FUD.

You are doing exactly the same thing by claiming that "gnatgcc" won't
successfully bootstrap the compiler.

There are two possible sources of problems:

(1) The "gnatgcc" may only compile Ada correctly and not C.
(2) There may be a layout difference between "gnatgcc" and "gcc" in
    some structure.

We have zero evidence for *either* of these things happening and I'd argue
that neither is likely.

That argues for the simpler approach, which is to use the same compiler
on all sources.

Re: RFC/Ada: removal of $(ADAC) support

[Zack Weinberg]

dewar@gnat.com (Robert Dewar) writes:

>> No. There is only one C source file that needs to be compiled
>> with the compiler that is used for the Ada source code: the one that
>> defines __gnat_raise_exception.  Which is what the patch you rejected
>> did.
>
> That's just with respect to one particular problem. The connection
> between the Ada parts of the code and the C parts of the code is
> intimate and extensive, and even the slightest change in layout can
> disturb this connection as well as other things.

I reiterate that there is no known problem in this domain other than
the one affecting __gnat_raise_exception; you are spreading FUD.

> Why encourage people to try compiling GNAT with non-matching
> compilers when this is known to be a potential source of problems
> and is completely unnecessary.

I disagree that it is unnecessary.  See my response to Kenner.

zw

Re: RFC/Ada: removal of $(ADAC) support

[Robert Dewar]

> No. There is only one C source file that needs to be compiled
> with the compiler that is used for the Ada source code: the one that
> defines __gnat_raise_exception.  Which is what the patch you rejected
> did.

That's just with respect to one particular problem. The connection between
the Ada parts of the code and the C parts of the code is intimate and
extensive, and even the slightest change in layout can disturb this connection
as well as other things. Why encourage people to try compiling GNAT with
non-matching compilers when this is known to be a potential source of
problems and is completely unnecessary.

Re: RFC/Ada: removal of $(ADAC) support

[Robert Dewar]

> You rejected a patch which would have fixed this.


You fixed one particular aspect, but there is no guarantee that changes made
to GNAT will be compatible with various ill-matched GNAT/C compilers, we don't
test for that, and indeed how can you test over some ill-defined universe.

I still don't understand why you would want to make additional difficulties
here.

Re: RFC/Ada: removal of $(ADAC) support

[Zack Weinberg]

kenner@vlsi1.ultra.nyu.edu (Richard Kenner) writes:

>     No. There is only one C source file that needs to be compiled
>     with the compiler that is used for the Ada source code: the one that
>     defines __gnat_raise_exception.  Which is what the patch you rejected
>     did.
>
> It's certainly true that that file does need to be compiled with the same
> compiler as the Ada files, but it's not at all clear why it isn't simpler to
> compile *all* of them with that same compiler?  What's the problem with
> simply setting CC to be "gnatgcc" for such systems?  This is what I never
> understood.

The systems under discussion ship a "gnatgcc" based on GCC 2.8, but a
"gcc" based on GCC 2.9x.  This was done for very good reason, that GCC
2.8 worked fine for Ada but silently miscompiled C source code on at
least one architecture that that system supports.  Second-guessing the
system integrator in that scenario, by forcing CC to gnatgcc, would
raise the risk of bootstrap failure due to a bug in the host compiler.
It is safer to compile only the Ada source code with gnatgcc, plus
whatever C modules absolutely have to be compiled with that compiler
for ABI compatibility.  

There is only one known C-and-Ada ABI incompatibility issue between
2.8 and 2.9x.  It affects __gnat_raise_exception, so that one file
needs to be compiled with gnatgcc.  But none of the others do.

zw

Re: RFC/Ada: removal of $(ADAC) support

[Richard Kenner]

    No. There is only one C source file that needs to be compiled
    with the compiler that is used for the Ada source code: the one that
    defines __gnat_raise_exception.  Which is what the patch you rejected
    did.

It's certainly true that that file does need to be compiled with the same
compiler as the Ada files, but it's not at all clear why it isn't simpler to
compile *all* of them with that same compiler?  What's the problem with
simply setting CC to be "gnatgcc" for such systems?  This is what I never
understood.

Re: RFC/Ada: removal of $(ADAC) support

[Zack Weinberg]

kenner@vlsi1.ultra.nyu.edu (Richard Kenner) writes:

>     It achieves the goal I had when I wrote it, which was to enable
>     bootstrapping with no special effort on common free operating system
>     distributions that ship compatible Ada and C compilers, but with the
>     Ada compiler accessible only as 'gnatgcc' or similar.
>
> Unfortunately, it didn't work and we got lots of complaints from users
> of those systems.

You must be talking to a disjoint set of people from the people I am
talking to.

>     You rejected a patch which would have fixed this.
>
> I don't believe it would have.
>
> The only way to "fix" it is to use the same compiler for everything.  It
> can be called either "gnatgcc" or "gcc", but the point is that you must
> use the same one to compile both C and Ada.  

No. There is only one C source file that needs to be compiled
with the compiler that is used for the Ada source code: the one that
defines __gnat_raise_exception.  Which is what the patch you rejected
did.

... Whatever.  You lot are the Ada maintainers, it's your decision
whether or not to put this in.  I won't be personally affected anymore
so I am not going to stop you.  But you are doing our users a
disservice.

zw

Re: RFC/Ada: removal of $(ADAC) support

[Richard Kenner]

    It achieves the goal I had when I wrote it, which was to enable
    bootstrapping with no special effort on common free operating system
    distributions that ship compatible Ada and C compilers, but with the
    Ada compiler accessible only as 'gnatgcc' or similar.

Unfortunately, it didn't work and we got lots of complaints from users
of those systems.

    You rejected a patch which would have fixed this.

I don't believe it would have.

The only way to "fix" it is to use the same compiler for everything.  It
can be called either "gnatgcc" or "gcc", but the point is that you must
use the same one to compile both C and Ada.  

My objection was not that it was looking for a "gnatgcc", which is
fine, but using "gcc" to compile some files and "gnatgcc" to compile others.

Re: RFC/Ada: removal of $(ADAC) support

[Zack Weinberg]

kenner@vlsi1.ultra.nyu.edu (Richard Kenner) writes:

> But the problem is that the $(ADAC) mechanism basically *doesn't
> work*.

It achieves the goal I had when I wrote it, which was to enable
bootstrapping with no special effort on common free operating system
distributions that ship compatible Ada and C compilers, but with the
Ada compiler accessible only as 'gnatgcc' or similar.

> In general, you can't compile the Ada sources with a different
> version of GCC from the C files due to the potential for
> incompatible exception handling.  Nearly all of the build problems
> with Ada have been due to this.

You rejected a patch which would have fixed this.

zw

Re: RFC/Ada: removal of $(ADAC) support

[Richard Kenner]

    I don't want to have the huge argument again that we had when this
    went in, but I'll reiterate what I said at the time: This mechanism
    should be removed no earlier than when there is a released stable
    version of Debian GNU/Linux for which GCC 3.2 or later is the default
    compiler.  Realistically speaking, that means sometime in mid-2004.
    That is also the earliest time at which features of GNU Ada present
    only in GCC 3.2 can be used in GNAT.

But the problem is that the $(ADAC) mechanism basically *doesn't
work*.  In general, you can't compile the Ada sources with a different
version of GCC from the C files due to the potential for incompatible
exception handling.  Nearly all of the build problems with Ada have been
due to this. 

Re: RFC/Ada: removal of $(ADAC) support

[Zack Weinberg]

Laurent GUERBY <guerby@acm.org> writes:

> Hi, now that we have an official FSF release with Ada included (and
> soon a second one), I propose we drop support for older versions not
> supporting C and Ada with the same driver in 3.4 (this is achived by
> the ADAC Makefile variable with 3.2 and 3.3).

It's too soon to do this.

I don't want to have the huge argument again that we had when this
went in, but I'll reiterate what I said at the time: This mechanism
should be removed no earlier than when there is a released stable
version of Debian GNU/Linux for which GCC 3.2 or later is the default
compiler.  Realistically speaking, that means sometime in mid-2004.
That is also the earliest time at which features of GNU Ada present
only in GCC 3.2 can be used in GNAT.

zw

RFC/Ada: removal of $(ADAC) support

[Laurent GUERBY]

Hi, now that we have an official FSF release
with Ada included (and soon a second one), I
propose we drop support for older versions
not supporting C and Ada with the same driver in 3.4
(this is achived by the ADAC Makefile variable with 3.2 and 3.3).

This will limit the amount of variation in the Ada build 
path and remove some configure/Makefile complexity.

Any objection?

Laurent