Lisp2Perl - Lisp to perl compiler

[David]

Hi,

Is anyone here at all interested in programs to translate lisp into
perl? I see there's a lisp interpreter on CPAN but I have implemented
something a bit different.

Basically its a program which can translate lisp programs (written in a
rather scheme-like dialect - even has a single namespace) into efficient
perl scripts. Perl has enough lisp like features to be able to implement
most things which can be done in lisp (lexical closures - hurrah!).

This lisp2perl translator/compiler works rather well. I've used it for a
major project at work. It's also self hosting (used to compile itself).

I've put a bit more info, though not much, and the source on my website:-

http://www.hhdave.pwp.blueyonder.co.uk

Any thoughts or comments anyone?

-- David

 

[Uri Guttman]

D> This lisp2perl translator/compiler works rather well. I've used it
D> for a major project at work. It's also self hosting (used to
D> compile itself).

D> http://www.hhdave.pwp.blueyonder.co.uk

D> Any thoughts or comments anyone?

can it translate the entire emacs lisp library? will it integrate with
emacs? there is a perl in emacs port that sorta worked a bit. i think
the author dropped it a few years ago. if you integrate it with that and
make it work on lisp emacs, i bet more than a few emacs users would love
to break their lisp shackles.

uri

 

[Thomas F. Burdick]

Hi,

Is anyone here at all interested in programs to translate lisp into
perl? I see there's a lisp interpreter on CPAN but I have implemented
something a bit different.

Basically its a program which can translate lisp programs (written in a
rather scheme-like dialect - even has a single namespace) into efficient
perl scripts. Perl has enough lisp like features to be able to implement
most things which can be done in lisp (lexical closures - hurrah!).

I have a grungy Lisp->Perl compiler. The major goal was to produce
readable Perl code, that looked like it might as well have been
written by a human, and could be maintained by a normal Perl hacker.
The lowest-level Lisp dialect is pretty much Perl-in-Lisp, then there
are enough macros and functions built on top of it to make it fairly
Lispy.

I have to say, I'm horrified that you're thinking of a scheme-like
Lisp. Perl itself has 4 namespaces, and I preserved that at my Lisp
level (special-forms called FUNCTION, SCALAR, HASH, and ARRAY, but an
inferencing engine generally takes care of picking the correct
namespace for you). Of course, my Lisp->Perl is hosted on Common
Lisp, so it was a pretty natural choice.

This lisp2perl translator/compiler works rather well. I've used it for a
major project at work. It's also self hosting (used to compile itself).

I've put a bit more info, though not much, and the source on my website:-

http://www.hhdave.pwp.blueyonder.co.uk

Any thoughts or comments anyone?

- Since you're really compiling Scheme to Perl, I'd think scheme2perl
would be a better name. You might as well pick the most specific
name (otherwise someone might think you mean elisp, for example

- If you're at all interested in readability of the resulting Perl,
don't offer gensym, just make-symbol (I accompanied mine with a
with-gensyms macro that takes the symbol names from the vars that
are bound to the new symbols). The compiler has to pay attention
to scoping anyway, so you can have it resolve the names. Hmm, that
wasn't a very comprehensible sentance, was it? How about an
example:

(let ((a (make-symbol "a"))
(b (make-symbol "a"))
(c (make-symbol "a")))
`(let ((,a 0))
(let ((,b 100))
(setf ,a (+ ,b 1)))
(let ((,c "hi"))
(print ,c))
(print ,a)))
=> (let ((#1=#:a 0))
(let ((#2=#:a 100))
(setf #1# (+ #2# 1)))
(let ((#3=#:a "hi"))
(print #3#))
(print #1#))
=> { my $a = 0;
my $a2 = 100;
$a = $a2 + 100;
{
my $a = "hi";
print $a
}
print $a; }

- I'm not sure how you can get better undefined-function error
reporting if you use the scheme approach. If you use multiple
namespaces in your Lisp, the function namespace maps directly from
Lisp to Perl, so you get normal Perl error reporting.

One thing you might consider here is losing the Lisp1-ness, but
keeping the Scheme-like handling (normal evaluation) of the first
position in a form. IE:

(defun foo (x) ...)
(let ((foo (lambda () ...)))
((complement (function foo)) ((var foo))))

For an unadorned symbol in the first position, you could have the
compiler infer the namespace based on the innermost lexical
binding. EG:

(let ((x 1))
(foo x)
(let ((foo (lambda (x) ...)))
(foo x)))
<==>
(let ((x 1))
((function foo) x)
(let ((foo (lambda (x) ...)))
((var foo) x)))

--
/|_ .-----------------------.
,' .\ / | No to Imperialist war |
,--' _,' | Wage class war! |
/ / `-----------------------'
( -. |
| ) |
(`-. '--.)
`. )----'

 

[Ben Morrow]

D> This lisp2perl translator/compiler works rather well.

can it translate the entire emacs lisp library? will it integrate with
emacs? there is a perl in emacs port that sorta worked a bit. i think
the author dropped it a few years ago. if you integrate it with that and
make it work on lisp emacs, i bet more than a few emacs users would love
to break their lisp shackles.

PleasePleasePleasePleasePlease

I love the functional concepts in Perl (I really can't see how any
language manages without closures but Lisp just makes my eyes go
funny.

Ben

 

[David]

I have a grungy Lisp->Perl compiler. The major goal was to produce
readable Perl code, that looked like it might as well have been
written by a human, and could be maintained by a normal Perl hacker.
The lowest-level Lisp dialect is pretty much Perl-in-Lisp, then there
are enough macros and functions built on top of it to make it fairly
Lispy.

Producing readable, hackable perl code was not one of the goals of my
program (as you might have guessed if you've seen any of the output!). I
had sort of figured that when you started using macros of any degree of
complexity then, whether you are translating to perl or not, you
wouldn't want to see the entirety of the macro expanded code anyway.
This is certainly the case with one largeish program I've written - the
amount of expanded code is enormous compared to the un-macro-expanded
code. Also, using (cond) statements, especially nested ones, produces
horrible looking perl code. They compile down to lots of trinary ifs
(a?b:c) in perl. The fact is, if I were programming in perl my program
just wouldn't be structured like that, whereas in lisp it seems a
natural thing to do.

I guess readable code is nice to have if you are 'supposed to be'
programming in perl, but really want to use Lisp

I'd be very interested in seeing the source code to your program if I may.

I have to say, I'm horrified that you're thinking of a scheme-like
Lisp. Perl itself has 4 namespaces, and I preserved that at my Lisp
level (special-forms called FUNCTION, SCALAR, HASH, and ARRAY, but an
inferencing engine generally takes care of picking the correct
namespace for you). Of course, my Lisp->Perl is hosted on Common
Lisp, so it was a pretty natural choice.

I didn't intend it to be horrific! (well, not _too_ horrific
I know it seems an odd thing to use a 1 namespace language to translate
to a 4 namespace language. The reason I'm doing the scheme thing as
opposed to the common lisp thing is that I just kind of like the 1
namespace approach. It seems to be a lot simpler and remove the
necessity for ways of dealing with different kinds of bindings. I guess
that's just personal preference. It does produce rather odd looking perl
code (lots of '$fn->(...)'), but as I say, I don't really care about
that. As long as it executes fast enough. I don't know if $fn->()
executes any slower than &fn() - haven't checked.

The other thing I find, with hashes and arrays and such, is that half
the time in perl I end up using references to those things stored in
scalars anyway. Particularly when I need nested structures.

- Since you're really compiling Scheme to Perl, I'd think scheme2perl
would be a better name. You might as well pick the most specific
name (otherwise someone might think you mean elisp, for example

Yeah, I know. It's just that, although it is scheme-like, it certainly
isn't scheme. It doesn't conform to the standard. It has unschemish
concepts of truth, for example, and the fundamental datatypes don't
behave as they should. 'lisp' is a less specific term though,
encompassing a multitude of similar, but distinct thing. It's its own
peculiar dialect of lisp

- If you're at all interested in readability of the resulting Perl,
don't offer gensym, just make-symbol (I accompanied mine with a
with-gensyms macro that takes the symbol names from the vars that
are bound to the new symbols). The compiler has to pay attention
to scoping anyway, so you can have it resolve the names. Hmm, that
wasn't a very comprehensible sentance, was it? How about an
example:

(let ((a (make-symbol "a"))
(b (make-symbol "a"))
(c (make-symbol "a")))
`(let ((,a 0))
(let ((,b 100))
(setf ,a (+ ,b 1)))
(let ((,c "hi"))
(print ,c))
(print ,a)))
=> (let ((#1=#:a 0))
(let ((#2=#:a 100))
(setf #1# (+ #2# 1)))
(let ((#3=#:a "hi"))
(print #3#))
(print #1#))
=> { my $a = 0;
my $a2 = 100;
$a = $a2 + 100;
{
my $a = "hi";
print $a
}
print $a; }

I'm not quite sure I understand this at the moment. I'll have to think
about it some more. If the generated perl looked like that wouldn't it
clash with a variable called 'a'? I know I'm probably being thick here.

- I'm not sure how you can get better undefined-function error
reporting if you use the scheme approach. If you use multiple
namespaces in your Lisp, the function namespace maps directly from
Lisp to Perl, so you get normal Perl error reporting.

I know, and it would seem the obvious thing to do wouldn't it? I still
like the single namespace though, but fear not - I thought of a solution
[me: looks up solution in files of notes about the program...]
Ah, here we go: I'm planning to modify the compiler to keep track of
lexical scope. That way, when it compiles a reference to an undefined
variable it should know and generate a warning about it. This may be
related to the gensym issue above. I guess generating symbols can be
done if I keep track of scope. I'll have to think about that some more.

Incidentally, can you think of a good argument AGAINST the scheme single
namespace approach?

One thing you might consider here is losing the Lisp1-ness, but
keeping the Scheme-like handling (normal evaluation) of the first
position in a form. IE:

(defun foo (x) ...)
(let ((foo (lambda () ...)))
((complement (function foo)) ((var foo))))

For an unadorned symbol in the first position, you could have the
compiler infer the namespace based on the innermost lexical
binding. EG:

(let ((x 1))
(foo x)
(let ((foo (lambda (x) ...)))
(foo x)))
<==>
(let ((x 1))
((function foo) x)
(let ((foo (lambda (x) ...)))
((var foo) x)))

That is a thought. I suppose it would make it play nicer with 'normal'
perl code. It would be particularly useful for using built in functions.
At the moment I have to 'declare' those:-

(<perl-sub> print)
which expands to a (defmacro ...)

I guess the first thing to do in any case is to extend the compilartion
functions so that they keep track of lexical scope.

 

[Thomas F. Burdick]

Producing readable, hackable perl code was not one of the goals of my
program (as you might have guessed if you've seen any of the output!). I
had sort of figured that when you started using macros of any degree of
complexity then, whether you are translating to perl or not, you
wouldn't want to see the entirety of the macro expanded code anyway.
This is certainly the case with one largeish program I've written - the
amount of expanded code is enormous compared to the un-macro-expanded
code.

It does take some care when writing macros, but it's pretty much the
same as always being in "I might have to debug this macro" mode. If
you use meaningful variable names, and prune unused branches, it helps
a lot. Plus, if you allow the macros to insert comments into the
resulting code, the volume of Perl code isn't so daunting (kinda like
if you wrote it by hand).

Also, using (cond) statements, especially nested ones, produces
horrible looking perl code. They compile down to lots of trinary ifs
(a?b:c) in perl. The fact is, if I were programming in perl my program
just wouldn't be structured like that, whereas in lisp it seems a
natural thing to do.

Right, whereas in Lisp you might write (setf x (cond ...)), in Perl,
you'd write:

if (...) {... $x = 1;}
elsif (...) {... $x = 2;}
else {... $x = 3;}

In my compiler, the translators for compound expressions have three
modes: producing code for side-effect only, producing code used for
its value, and producing code to put the value in a specific location.
So, depending on context, cond might expand into if/elsif/else, or
trinary-if, or if it's complicated and/or nested, a call to an
anonymous lambda:

(sub { if (...) {... return 1;}
elsif (...) {... return 2;}
else {... return 3;} })->();

I guess readable code is nice to have if you are 'supposed to be'
programming in perl, but really want to use Lisp

Well, it wasn't so much "supposed to be", as much as the final product
had to be in Perl, so it would be easy to find someone later to
maintain it. No one had any problem with me using whatever expert
development tools I wanted, as long as the output was maintainable
as-is. But, pretty much, yeah

I'd be very interested in seeing the source code to your program if I may.

I'm sitting on it, pending my thinking about how much time/effort it
would take to make it useful to the general public, and if there's a
market for it or not. And it's a mess of unfactored hacks, because I
was concentrating on the systems I was supposed to be writing, not the
compiler itself.

I didn't intend it to be horrific! (well, not _too_ horrific

Horrific because it's a Lisp->Perl compiler, but not because of the
namespace issue?

I know it seems an odd thing to use a 1 namespace language to translate
to a 4 namespace language. The reason I'm doing the scheme thing as
opposed to the common lisp thing is that I just kind of like the 1
namespace approach. It seems to be a lot simpler and remove the
necessity for ways of dealing with different kinds of bindings. I guess
that's just personal preference. It does produce rather odd looking perl
code (lots of '$fn->(...)'), but as I say, I don't really care about
that. As long as it executes fast enough. I don't know if $fn->()
executes any slower than &fn() - haven't checked.

I'd imagine it is, but I wouldn't sweat an added indirection when
you're talking about a bytecode interpreter.

The other thing I find, with hashes and arrays and such, is that half
the time in perl I end up using references to those things stored in
scalars anyway. Particularly when I need nested structures.

Certainly, references to hashes especially are important, in
particular for supporting defstruct. But if you want to interact with
Perl builtins, being able to spread arrays is important. I guess you
don't need all 4 namespaces for that, it just makes the resulting Perl
less crazy-looking.

Oooh, just noticed this. I'm glad I didn't try to go that route, I
was happy to have all of Common Lisp at my disposal when writing my
compiler. You might want to reconsider this decision, if you find
yourself having implementation difficulties -- compilers are a lot
easier to write in big languages (like CL, or one of the big scheme
implementations' dialects with all the add-ons).

I'm not quite sure I understand this at the moment. I'll have to think
about it some more. If the generated perl looked like that wouldn't it
clash with a variable called 'a'? I know I'm probably being thick here.

The point is that a naive translation would be:

{ my $a1 = 0;
{ my $a2 = 100;
$a1 = $a2 + 100; # clash
{ my $a3 = "hi";
print $a3;
}
}
print $a1;
}

If you named $a1, $a2, and $a3 all just $a, it would work, except for
the line labeld "clash", which refers to an $a from two different
scoping levels. So you can name $a1 and $a3 plain old $a, and only
need to give $a2 a distinct name. In $a3's scope, it is the only $a
variable used.

- I'm not sure how you can get better undefined-function error
reporting if you use the scheme approach. If you use multiple
namespaces in your Lisp, the function namespace maps directly from
Lisp to Perl, so you get normal Perl error reporting.

I know, and it would seem the obvious thing to do wouldn't it? I still
like the single namespace though, but fear not - I thought of a solution
[me: looks up solution in files of notes about the program...]
Ah, here we go: I'm planning to modify the compiler to keep track of
lexical scope. That way, when it compiles a reference to an undefined
variable it should know and generate a warning about it. This may be
related to the gensym issue above. I guess generating symbols can be
done if I keep track of scope. I'll have to think about that some more.

Yeah, they're def related.

Incidentally, can you think of a good argument AGAINST the scheme single
namespace approach?

You get to it yourself in a second

One thing you might consider here is losing the Lisp1-ness, but
keeping the Scheme-like handling (normal evaluation) of the first
position in a form.

[snip]
That is a thought. I suppose it would make it play nicer with 'normal'
perl code. It would be particularly useful for using built in functions.
At the moment I have to 'declare' those:-

(<perl-sub> print)
which expands to a (defmacro ...)

Yeah, that's a benefit of recognizing at least the function and
variable namespaces. That way, you can easily use normal Perl
functions, and your functions aren't second-class citizens (eg, you
can write a module that Perl coders can use directly, normally).

I guess the first thing to do in any case is to extend the compilartion
functions so that they keep track of lexical scope.

That is the traditional thing to do when writing scheme compilers

--
/|_ .-----------------------.
,' .\ / | No to Imperialist war |
,--' _,' | Wage class war! |
/ / `-----------------------'
( -. |
| ) |
(`-. '--.)
`. )----'

 

[David]

It does take some care when writing macros, but it's pretty much the
same as always being in "I might have to debug this macro" mode. If
you use meaningful variable names, and prune unused branches, it helps
a lot. Plus, if you allow the macros to insert comments into the
resulting code, the volume of Perl code isn't so daunting (kinda like
if you wrote it by hand).




Right, whereas in Lisp you might write (setf x (cond ...)), in Perl,
you'd write:

if (...) {... $x = 1;}
elsif (...) {... $x = 2;}
else {... $x = 3;}

In my compiler, the translators for compound expressions have three
modes: producing code for side-effect only, producing code used for
its value, and producing code to put the value in a specific location.
So, depending on context, cond might expand into if/elsif/else, or
trinary-if, or if it's complicated and/or nested, a call to an
anonymous lambda:

(sub { if (...) {... return 1;}
elsif (...) {... return 2;}
else {... return 3;} })->();

Sounds much like what I did except that I had 2 modes, not 3. I did have
to create an anonymous sub (lambda) in the perl in certain cases and
immediately eval it. That was cunning I thought. My compiler doesn't do
that in the case above though. I actually noticed (I think) that ifs
which were forced to compile to the trinary operator (since the value
was used) were somewhat faster than the expression (side effect only)
ones too. I suppose that's due to not creating a new scope in perl.

Well, it wasn't so much "supposed to be", as much as the final product
had to be in Perl, so it would be easy to find someone later to
maintain it. No one had any problem with me using whatever expert
development tools I wanted, as long as the output was maintainable
as-is. But, pretty much, yeah

Fair enough. It was good that you managed to find a way of still using
lisp despite the requirement for maintainability by a perl programmer.

I'm sitting on it, pending my thinking about how much time/effort it
would take to make it useful to the general public, and if there's a
market for it or not. And it's a mess of unfactored hacks, because I
was concentrating on the systems I was supposed to be writing, not the
compiler itself.

I know the feeling. There are lots of improvements I want to make to
mine as well. For one thing I want to change the way the whole thing
works as I've used a bit of a hack to get things like this to work:-

(define (f x)
...some function of x...)

(defmacro my-macro (a b)
(list 'foo a (f b)))

Basically, in order that the macro can use the function defined before
it the function must be evaluated at compile time as well as runtime. I
think the root of this problem is that I want to be able to translate
the lisp code into a perl script which can be executed in the normal
way. I don't want to have to read in the lisp code and translate and
execute each form one by one. This means that the macro definitions must
executed at compile time (obviously) hence function definitions must be
executed at compile time. I don't think there should really be such a
distinction between compile time and runtime. I have found a solution
though, which lies in having an intermediate lisp representation which
is just fully macro expanded lisp code which is generated as a side
effect of 'running' a lisp file. This means that you can't just compile
the lisp to perl as such, you have to run it, and it gets compiled
(partially) as a side effect. Separate modules could then be linked, and
fully translated to perl, later. Of course, running a file is not a
problem if all it does is define things (functions and macros).

Does this seem sane?

Even though lisp2perl isn't nearly where I'd like it to be I decided to
just release it anyway. Where I work, even though I use this and people
see its benefits, I think it would have been frowned upon if I'd spent
loads of time on it instead of what I was using it for (what a
surprise). I wrote much of it at home.

Horrific because it's a Lisp->Perl compiler, but not because of the
namespace issue?

Well, quite. There is something 'unsettling' about the concept. Really
I'm looking forward to the release of Perl 6, because then I could just
compile lisp to parrot vm code and still get the benefits of using perl
code from lisp. That doesn't help your perl generation of course. I'm
surprised to find that someone else has done something similar.

I'd imagine it is, but I wouldn't sweat an added indirection when
you're talking about a bytecode interpreter.




Certainly, references to hashes especially are important, in
particular for supporting defstruct. But if you want to interact with
Perl builtins, being able to spread arrays is important. I guess you
don't need all 4 namespaces for that, it just makes the resulting Perl
less crazy-looking.

I know what you mean. Interacting with built ins is something I haven't
really solved nicely yet.

Oooh, just noticed this. I'm glad I didn't try to go that route, I
was happy to have all of Common Lisp at my disposal when writing my
compiler. You might want to reconsider this decision, if you find
yourself having implementation difficulties -- compilers are a lot
easier to write in big languages (like CL, or one of the big scheme
implementations' dialects with all the add-ons).

I'm not quite sure I understand this at the moment. I'll have to think
about it some more. If the generated perl looked like that wouldn't it
clash with a variable called 'a'? I know I'm probably being thick here.

The point is that a naive translation would be:

{ my $a1 = 0;
{ my $a2 = 100;
$a1 = $a2 + 100; # clash
{ my $a3 = "hi";
print $a3;
}
}
print $a1;
}

If you named $a1, $a2, and $a3 all just $a, it would work, except for
the line labeld "clash", which refers to an $a from two different
scoping levels. So you can name $a1 and $a3 plain old $a, and only
need to give $a2 a distinct name. In $a3's scope, it is the only $a
variable used.

- I'm not sure how you can get better undefined-function error
reporting if you use the scheme approach. If you use multiple
namespaces in your Lisp, the function namespace maps directly from
Lisp to Perl, so you get normal Perl error reporting.

I know, and it would seem the obvious thing to do wouldn't it? I still
like the single namespace though, but fear not - I thought of a solution
[me: looks up solution in files of notes about the program...]
Ah, here we go: I'm planning to modify the compiler to keep track of
lexical scope. That way, when it compiles a reference to an undefined
variable it should know and generate a warning about it. This may be
related to the gensym issue above. I guess generating symbols can be
done if I keep track of scope. I'll have to think about that some more.

Yeah, they're def related.

Incidentally, can you think of a good argument AGAINST the scheme single
namespace approach?

You get to it yourself in a second

One thing you might consider here is losing the Lisp1-ness, but
keeping the Scheme-like handling (normal evaluation) of the first
position in a form.

[snip]

That is a thought. I suppose it would make it play nicer with 'normal'
perl code. It would be particularly useful for using built in functions.
At the moment I have to 'declare' those:-

(<perl-sub> print)
which expands to a (defmacro ...)

Yeah, that's a benefit of recognizing at least the function and
variable namespaces. That way, you can easily use normal Perl
functions, and your functions aren't second-class citizens (eg, you
can write a module that Perl coders can use directly, normally).

Well yes, I know _that_ benefit of the 2 namespace approach. But I mean,
forgetting about perl (ie if we were compiling to something else) can
you think of any benefit of 2 namespaces?

More sane integration with normal perl code would definitely be good.
I'll do something about it if/when I get chance. I just wish I had more
time to work on it. Ho hum.

That is the traditional thing to do when writing scheme compilers

Yeah, sounds like a good idea. I'm new at this Its a slightly unusual
compilation target too.

 

[Jens Axel Søgaard]

Well, quite. There is something 'unsettling' about the concept. Really
I'm looking forward to the release of Perl 6, because then I could just
compile lisp to parrot vm code and still get the benefits of using perl
code from lisp. That doesn't help your perl generation of course. I'm
surprised to find that someone else has done something similar.

Never say never:

<http://okmij.org/ftp/Scheme/Scheme-in-Perl.txt>

<http://www.venge.net/graydon/scm2p.html>

 

[Lars Brinkhoff]

forgetting about perl (ie if we were compiling to something else)
can you think of any benefit of 2 namespaces?

Suggested reading:
http://www.nhplace.com/kent/Papers/Technical-Issues.html

 

[Thomas F. Burdick]

I know the feeling. There are lots of improvements I want to make to
mine as well. For one thing I want to change the way the whole thing
works as I've used a bit of a hack to get things like this to work:-

(define (f x)
...some function of x...)

(defmacro my-macro (a b)
(list 'foo a (f b)))

Basically, in order that the macro can use the function defined before
it the function must be evaluated at compile time as well as runtime.

Since your system is self-hosting, that shouldn't be a problem. Do
you have something like Common Lisp's eval-when? If not, you should
read the page in the spec

http://www.lispworks.com/reference/HyperSpec/Body/s_eval_w.htm#eval-when

especially the Notes section at the bottom. In CL, forms like defun
and defmacro (when they're toplevel) cause the function to be known at
compile time by using eval-when. It also lets you write your own
forms of this type.

I
think the root of this problem is that I want to be able to translate
the lisp code into a perl script which can be executed in the normal
way. I don't want to have to read in the lisp code and translate and
execute each form one by one. This means that the macro definitions must
executed at compile time (obviously) hence function definitions must be
executed at compile time. I don't think there should really be such a
distinction between compile time and runtime.

Actually, I think the answer is probably having a clearer concept of
time(s) in your Lisp dialect. When the compiler sees a toplevel form
that macroexpands into (eval-when (compile) ...), it should
recursively invoke itself, to deal with the eval-when, evaluate that
in the Perl interpreter, then get back to the job at hand.

I have found a solution
though, which lies in having an intermediate lisp representation which
is just fully macro expanded lisp code which is generated as a side
effect of 'running' a lisp file. This means that you can't just compile
the lisp to perl as such, you have to run it, and it gets compiled
(partially) as a side effect. Separate modules could then be linked, and
fully translated to perl, later. Of course, running a file is not a
problem if all it does is define things (functions and macros).

Does this seem sane?

It seems messy. You should be able to start Perl, load all the
functions that your macros use, compile the files defining and using
them, then quit Perl and load the resulting .pl files. Using a
CL-like concept of times (compile/macroexpand, load, and eval) would
help keep things cleaner.

(In my system, the macros are run in the hosting Common Lisp, so my
issues were different. You define Perl functions with perlisp:defun,
but they're not available at compile-time. Macros can use functions
defined with common-lisp:defun).

Well, quite. There is something 'unsettling' about the concept. Really
I'm looking forward to the release of Perl 6

I wouldn't hold my breath. And personally, given the
backwards-incompatibilities of Perl's past, I hope it never comes.

Well yes, I know _that_ benefit of the 2 namespace approach. But I mean,
forgetting about perl (ie if we were compiling to something else) can
you think of any benefit of 2 namespaces?

Oh, certainly. In the realm of purely style issues, there's the
nicety of being able to have a type (eg, cons) whose constructor
function is the same as the type's name (cons), and being able to
stick an instance of this type in a variable of the same name, while
still being able to use the constructor function. Eg:

(let ((cons (assoc 'foo alist)))
(if cons
(setf (cdr cons) 'bar)
(setf alist (cons (cons 'foo 'bar) alist))))

There are a lot of style issues like this. But the big reason I bring
up something that I know Schemers would prefer to avoid, is that it's
important for supporting defmacro-style macros. In a Lisp-1, you
worry about binding some name with a let, and in the body of that let,
having a macro expand into a call to that name. This is the problem
that syntactic closures, and Scheme's pattern-matching macros were
invented to solve. In a Lisp-2, there is still a *posibility* of
doing this, using flet/labels, but it's much less likely to happen.
If you add a package system, like in CL, the likelyhood of a user
accidentally shadowing a global function definition is almost nil.
It's essentially a non-issue.

And remember, Lisp-2'ness is orthogonal to normal evaluation of the
first element in forms. You could have a perfectly legit Lisp-2 that
would allow

(flet ((f (x) (lambda (y) (+ x y))))
((f 10) 5))

Yeah, sounds like a good idea. I'm new at this Its a slightly unusual
compilation target too.

No kidding. I definately get a kick out of the fact that someone else
is doing it too. BTW, you might be interested to know about Linj, a
Lisp->Java compiler that produces readable Java from a CL-like dialect
of Lisp. It's not currently available, but its developer has given
presentations/demos at the last two International Lisp Conferences.

--
/|_ .-----------------------.
,' .\ / | No to Imperialist war |
,--' _,' | Wage class war! |
/ / `-----------------------'
( -. |
| ) |
(`-. '--.)
`. )----'

 

[Marcin 'Qrczak' Kowalczyk]

This means that the macro definitions must executed at compile time
(obviously) hence function definitions must be executed at compile time.
I don't think there should really be such a distinction between compile
time and runtime.

I think there should. Please read
<http://www.cs.utah.edu/plt/publications/macromod.pdf>

 

[Pascal Costanza]

I think there should. Please read
<http://www.cs.utah.edu/plt/publications/macromod.pdf>

Oh dear. This is just another one of those bad examples of unbreakable
abstractions. Unbreakability sucks.

Generally speaking, layering or staging of approaches is probably a good
idea. But there might be situations in which I need to mix up the
layers/stages. If a proposed solution doesn't provide a back door for
these things, it sucks IMHO.

In general, computer scientists tend to confuse description and
prescription. I can describe a good solution and explain why it works
and what's good about it. But prescribing that everybody else should do
it exactly the same is the wrong conclusion. Especially when you don't
give them a way out.


Pascal

 

[Marcin 'Qrczak' Kowalczyk]

Generally speaking, layering or staging of approaches is probably a good
idea. But there might be situations in which I need to mix up the
layers/stages. If a proposed solution doesn't provide a back door for
these things, it sucks IMHO.

The same module can be independently imported to multiple stages,
so I don't see a problem.

 

[Joe Marshall]

Oh dear. This is just another one of those bad examples of unbreakable
abstractions. Unbreakability sucks.

Generally speaking, layering or staging of approaches is probably a
good idea. But there might be situations in which I need to mix up the
layers/stages. If a proposed solution doesn't provide a back door for
these things, it sucks IMHO.

In this case, you don't want to mix up the layers or stages. Really.

The macro/module system that Matthew is describing is one that allows
the system to determine what to load and when to load it in order to
ensure that you have the macro-expansion code in place *before* you
attempt to expand the macros that use it.

Mixing up the layers or stages means that you want to have circular
dependencies in your macro expansions, i.e., macro FOO needs the
QUASIQUOTE library to expand, but the QUASIQUOTE library is written
using the macro FOO.

 

[Pascal Costanza]

The same module can be independently imported to multiple stages,
so I don't see a problem.

I haven't analyzed the issues in all details, but someone reported to me
that my implementation of dynamically scoped functions wouldn't be
possible because of the separation of stages in that module system.

I don't see how my code would cause any real problems, so if a module
system doesn't allow me to write it, there's something wrong with that
module system, and not with my code.

See my paper about DSF at http://www.pascalcostanza.de/dynfun.pdf that
also includes the source code.


Pascal

 

[Matthias Blume]

In this case, you don't want to mix up the layers or stages. Really.

The macro/module system that Matthew is describing is one that allows
the system to determine what to load and when to load it in order to
ensure that you have the macro-expansion code in place *before* you
attempt to expand the macros that use it.

Mixing up the layers or stages means that you want to have circular
dependencies in your macro expansions, i.e., macro FOO needs the
QUASIQUOTE library to expand, but the QUASIQUOTE library is written
using the macro FOO.

I agree with Pascal: being able to go only forward in time really sucks.

 

[Joe Marshall]

I agree with Pascal: being able to go only forward in time really sucks.

For you, there's continuations.

 

[Joe Marshall]

I haven't analyzed the issues in all details, but someone reported to
me that my implementation of dynamically scoped functions wouldn't be
possible because of the separation of stages in that module system.

See my paper about DSF at http://www.pascalcostanza.de/dynfun.pdf that
also includes the source code.

From a quick perusal of your code, I see nothing in it that wouldn't
easily work with the module system.

 

[Matthias Blume]

For you, there's continuations.

I just hope you are trying to continue (no pun intended) my attempt at
being sarcastic...

 

[Pascal Costanza]

From a quick perusal of your code, I see nothing in it that wouldn't
easily work with the module system.

OK, I have checked my email archive to see what the issue was. Here it
is: In MzScheme, there is obviously no way to access run-time values at
compile time / macro expansion time. (I hope I have gotten the
terminology right here.)

This means that you can't say this:

(define a 1)
(define-macro (foo x) `(+ ,a ,x))
(foo 2)

This will result in a "reference to undefined identifier" error, because
the foo macro doesn't see the a reference.

In my implementation of dynamically scoped functions in Common Lisp, I
make use of this collapsing of stages in order to be able to redefine
existing functions as dynamically scoped functions, without changing
their defined behavior. See my paper for an example how I turn a CLOS
generic function into a dynamically scoped one.

Even if I can go only forward in time, I like the fact that I can change
the decisions I have made in the past. It escapes me why this should be
an evil thing to do in programs.


Pascal