World: r4wp

foo: func [[catch]][
	print "2"
	throw 10
	print "KO"
]

print "1"
foo
print "3"
print ["^/thrown value: " system/thrown lf]

will output:

123
thrown value: 10

A deeper nested example:

foo2: does [
	print "3"
	throw 100
	print "KO"
]

bar: func [[catch]][
	print "2"
	foo2
	print "KO"
]

print "1"
bar
print ["^/thrown value: " system/thrown lf]

will output:

123
thrown value: 100

The implementation is done and I will push it in a few minutes. It 
required about 25 additional LOC to implement all such simple exception 
system (right, 25!). ;-)

So, how does it work? When you need to interrupt the flow of code 
in a function in Red/System, currently you can just use EXIT/RETURN 
to make an early exit. But, sometimes, you need to go up through 
several nested calls, that's where the new THROW function comes handy. 
It will interrupt the execution and go up the call tree to find the 
first function that has the CATCH attribut set. It will then just 
resume execution after the last function call (from which the exception 
has been generated).


If no CATCH attribut is found, it will go up to global code and resume 
from there.

THROW requires an integer! value. Such value represent the exception 
ID and is user-defined. After the resume from a caught exception, 
you can use SYSTEM/THROWN to read the passed exception ID and act 
accordingly (usually using a SWITCH dispatcher).

Important thing to note: system/thrown needs to be manually reset, 
as the last thrown value will stay there if no exception occured. 
Such reset could be done before each call to a function that could 
generate an exception or after processing the thrown value.

I could have added a much more sophisticated system with a true CATCH 
function, but this would have made the implementation way more complex 
and would have taken a lot more time. As I need it only for Red's 
interpreter, I think this way should be enough and will be usable 
by other Red/System programmers to enhance their own code.

Thinking about it, it might be possible to allow an extended catch 
attribut with a integer value to specify the barrier value for catching 
exceptions (and avoid manual re-throwning), something like:

foo: func [[catch 100]][...]


would catch all thrown exceptions with a value <= 100 and let others 
pass up to caller.

What do you think about all this?

BTW, this exception system relies on stack frames unwinding, so it 
won't work for callbacks (CATCH attribut won't be accepted in callback 
functions anyway).

I've considered using a continuation approach for constant-time exception 
throwing (would also work from callbacks), but I didn't find a simple 
enough syntax for defining/using it without having to rely on CPS 
(Continuation-Passing Style) for every functions...

See http://en.wikipedia.org/wiki/Continuation-passing_style

Another note: messing up with the stack (using push/pop) in a user-function 
is fine as long as you returned the stack clean before calling a 
function that could generate an exception. Failure to do so will 
result in a crash. It might be possible to make stack-unwinding for 
exception resistant to such cases, I will investigate that.

Exception handling code pushed. Have fun! ;)

(only IA-32 for now, will add ARM support in the next days)

That sounds great Doc. I collected your comments into a file in %Red/.

Hmm, shouldn't bar print KO in the second example, from your description?

Nope, see first example. This might be a bit non/counter-intuitive, 
I will see if I can improve that.

Ah, I get it now. Yes, that wasn't obvious at first.

I won't be online tomorrow most of the day, will look into that and 
will upgrade the documentation tomorrow night.

So you don't check system/thrown in the func with [catch], correct?

But in the func that *called* the func with [catch].

Right.

I might adjust that tomorrow if possible to make it more intuitive.

Absolutely fabulous Nenad!!!

Good progress. Nice of you to sum it up Gregg, makes the comments 
a whole lot more readable. Filename is a bit awkward though: "esceptions"

Fixed. Thanks Arnold.

I had a Spanish accent last night.

Gregg, I've reconsidered the behavior of the [catch] function attribut, 
if I change it, some use cases become even less intuitive. For example, 
the single function use case (my first example above) performs quite 
intuitively. I think the  vague non-intuitive feeling at the beginning 
comes from our habits with CATCH native in Rebol, after using Red/System's 
way a bit, that feeling just faded away. Anyway, if you have suggestions 
for improvements, it's the right time.

Doc, any idea how I can convert a string! passed into a routine! 
to UTF-8, or access a cached UTF-8 value?

Kaj, can you use string/mold inside the routine to create a c-string! 
?

The answer is not you can't as mold doesnt output a UTF-8 string.

Kaj: cached UTF-8 string is available using str/cache if str is a 
red-string! value.

Only small strings (< 64 bytes) are cached.

We haven't yet implemented UTF-8 encoding functions in the standard 
library. It will be done during the I/O implementation (unless you 
have a strong need for it, then I'll have a look at it).

After trying to use the new exception handling in the interpreter 
to implement EXIT/RETURN, I realized that the current behavior of 
the [catch] attribut is not suitable in a intensively recursive environment. 
So I had to change the moment the exceptions are caught to an earlier 
time. Now, it works like this:

	foo: does [
		print "3"
		throw 100
		print "KO"
	]

	bar: func [[catch]][
		print "2"
		foo
		print "4"
	]

	print "1"
	bar

will output: 1234


So, now the [catch] attribut resumes the execution at the current 
level, instead of the parent caller level like before.

The catch flags position on stack has been changed to a safer place, 
so it's now resistant to a "dirty" stack left by user code (unbalanced 
PUSH/POP actions at the exception raising point).

A [catch] flag will have no effect on exceptions launched from the 
same level (enables re-throwing exceptions).

EXIT and RETURN have been implemented for the interpreter too now. 
All related tests are now passing. An important fact to note, which 
differs from the Rebol way:


EXIT and RETURN are dialect keywords in Red, not native functions. 
Both the compiler and the interpreter are processing them as part 
of function's body dialect.

@Ladislav, I know you don't care much about Red, but you might be 
pleased to see that some of your old propositions for Rebol found 
their way into Red. ;-)

Can't speak for Ladislav, but maybe he is mostly busy with other 
stuff, so he'll wait, how Red turns out in the end :-)

So, now as exceptions are working, my typical - what's next? :-) 
Dyn-lib emmiter? Contexts? IO? :-)

Shared lib generation for Mac OS X and Linux.

Does that include ARM?

Not sure if you mean for exceptions or shared libs, so:
- Exceptions: I will push the upgraded ARM version in a few hours
- Shared lib: yes, we will support it on ARM too

I meant shared libs

...you might be pleased to see that some of your old propositions 
for Rebol found their way into Red
 - yes, that is great!

EXIT from a func exits the console. Is that intended?

N ope.