> You seem to disagree that tailcalls should be optimized everywhere possible
I don't disagree, assuming there's a way to turn that off for debugging. I love it when things run faster and take less memory.
I disagree that "there's already a syntax for tail call" (that can be relied upon if Python did TCE) because in Python there isn't. Therefore, even if Python had TCE, it would be brittle (a small change 50 lines away could make it into a non-tail call) and it would be bad style to rely on it as a programming mechanism.
I've used common lisp with tailcalls. They work way better there than they would in Python, although there are still interaction problems with conditions/restarts/continuations/macros, mostly as invisible as they would be in Python. (Though, it's been something like 20 years, my memory is hazy).
My claim can be summarized as follows:
1. TCO is nice and awesome. But it is an OPTIMIZATION, and shouldn't be a central language feature you rely on, unless it is easy to guarantee it happens.
2. If you want to rely on TCE as a general programming technique, you should be able to say "and if this thing isn't a tail call, don't compile" or "and if this thing isn't a tail call, give me a warning".
3. A "return from" in Python could be an excellent way to implement (2) for Python
4. I think a special form for Scheme that says "I rely on the following tail call being eliminated" would have been useful, as it is easy to shoot yourself in the foot relying on it while e.g. using macros - where something that looks like it is a simple call expands to a non-trivial expression. Luckily, in CL (and I assume in Scheme as well), you could do that at the library level with a relatively simple code walker.
The Python Library-Level implementations that I've seen use a decorator to add the trampoline and need you to replace "return f(a,b,c)" with "raise TailCall(f, a, b, c)" (or something to that effect - you can get a much nicer syntax with some limitations). Ugly, inefficient, but available if you really need it.
I don't disagree, assuming there's a way to turn that off for debugging. I love it when things run faster and take less memory.
I disagree that "there's already a syntax for tail call" (that can be relied upon if Python did TCE) because in Python there isn't. Therefore, even if Python had TCE, it would be brittle (a small change 50 lines away could make it into a non-tail call) and it would be bad style to rely on it as a programming mechanism.
I've used common lisp with tailcalls. They work way better there than they would in Python, although there are still interaction problems with conditions/restarts/continuations/macros, mostly as invisible as they would be in Python. (Though, it's been something like 20 years, my memory is hazy).
My claim can be summarized as follows:
1. TCO is nice and awesome. But it is an OPTIMIZATION, and shouldn't be a central language feature you rely on, unless it is easy to guarantee it happens.
2. If you want to rely on TCE as a general programming technique, you should be able to say "and if this thing isn't a tail call, don't compile" or "and if this thing isn't a tail call, give me a warning".
3. A "return from" in Python could be an excellent way to implement (2) for Python
4. I think a special form for Scheme that says "I rely on the following tail call being eliminated" would have been useful, as it is easy to shoot yourself in the foot relying on it while e.g. using macros - where something that looks like it is a simple call expands to a non-trivial expression. Luckily, in CL (and I assume in Scheme as well), you could do that at the library level with a relatively simple code walker.
The Python Library-Level implementations that I've seen use a decorator to add the trampoline and need you to replace "return f(a,b,c)" with "raise TailCall(f, a, b, c)" (or something to that effect - you can get a much nicer syntax with some limitations). Ugly, inefficient, but available if you really need it.