On Tue, Jul 9, 2013 at 1:27 AM, Igor Stasenko siguctua@gmail.com wrote:
On 9 July 2013 10:15, Clément Bera bera.clement@gmail.com wrote:
Hello pharoers,
Recently I have been looking at the Pharo thisContext capabilities, in
order
to perhaps one day in the far future edit its implementation with the
Pharo
dev team. Nothing is planned or confirmed, it is just to discuss.
A context has instances variables (method closureOrNil receiver stackp sender pc) and holds the arguments and temporaries.
Now I'd like to know what context's state do we modify, and what states
are
just internal representations ?
For example, it seems that you can do 'thisContext receiver: #foo', but
you
cannot with Cog.
SomeClass>>foo thisContext receiver: #foo. ^ self
In workspace, evaluating: 1 to: 5 do: [:i | Transcript show: SomeClass new foo ]
Transcript result (with Cog): foo foo a SomeClass a SomeClass a SomeClass
Transcript result (with Stack or Vanilla): foo foo foo foo foo
Now as no one has ever complained, I guess this feature is not used.
As far as I know, the real use cases of the context seems to be:
- setting and gettings temporaries
- setting and getting the sender of a context
- setting and getting the pc
- getting method, closureOrNil, receiver, stackp, arguments but NOT
setting
them
Now setting the sender of a context seems to be used only in two cases:
- continuations (as seaside continuations)
- exception implementation
So imagine that in the future you would have a context that can be
accessed
in read-only, where you could only:
- set the temporaries (but not arguments)
- set the pc (or something equivalent, as set the currently executed ast
node)
- use continuations (exceptions can be implemented on top of
continuations)
I would like to know if there are things that you do now and that you
would
not be able to do with a context like that. For non meta developer (like enterprise app developer) I guess it will not change anything, but I
want to
know if you implemented a framework as seaside, does it requires other things from the context and why ?
To my thinking playing with pc is awfully evil, it should be read-only, only to allow debugger to map it to source code. Same goes for stack pointer.
There are good uses for this I just used it to implement basic-block coverage for methods. One can use it for exception handling (mustBeBoolean etc). I'm always wary of introducing restrictions when the system has worked fine without them for years. One exception recently was in adding bounds checking to CompiledMethod>>at:put: so one could not change arbitrary bytes in the literals via at:put:. But this was adding safety, not adding a restriction. I see no harm in allowing one to assign the pc. It is unsafe, and the system will likely crash if you get it wrong. But it is also potentially useful (e.g. I've used it to do a prototype of tail-recursion elimination). So let's be laissez faire, unless the freedoms in question really do only do harm.
Setting receiver is less evil and it may work, if receiver belongs to same inheritance chain as compiled method's class. But it should not be set without checking this.
And for method's temps, it is fine.
Thanks for answering,
-- Best regards, Igor Stasenko.
On 9 July 2013 23:03, Eliot Miranda eliot.miranda@gmail.com wrote:
On Tue, Jul 9, 2013 at 1:27 AM, Igor Stasenko siguctua@gmail.com wrote:
On 9 July 2013 10:15, Clément Bera bera.clement@gmail.com wrote:
Hello pharoers,
Recently I have been looking at the Pharo thisContext capabilities, in order to perhaps one day in the far future edit its implementation with the Pharo dev team. Nothing is planned or confirmed, it is just to discuss.
A context has instances variables (method closureOrNil receiver stackp sender pc) and holds the arguments and temporaries.
Now I'd like to know what context's state do we modify, and what states are just internal representations ?
For example, it seems that you can do 'thisContext receiver: #foo', but you cannot with Cog.
SomeClass>>foo thisContext receiver: #foo. ^ self
In workspace, evaluating: 1 to: 5 do: [:i | Transcript show: SomeClass new foo ]
Transcript result (with Cog): foo foo a SomeClass a SomeClass a SomeClass
Transcript result (with Stack or Vanilla): foo foo foo foo foo
Now as no one has ever complained, I guess this feature is not used.
As far as I know, the real use cases of the context seems to be:
- setting and gettings temporaries
- setting and getting the sender of a context
- setting and getting the pc
- getting method, closureOrNil, receiver, stackp, arguments but NOT setting
them
Now setting the sender of a context seems to be used only in two cases:
- continuations (as seaside continuations)
- exception implementation
So imagine that in the future you would have a context that can be accessed in read-only, where you could only:
- set the temporaries (but not arguments)
- set the pc (or something equivalent, as set the currently executed ast
node)
- use continuations (exceptions can be implemented on top of continuations)
I would like to know if there are things that you do now and that you would not be able to do with a context like that. For non meta developer (like enterprise app developer) I guess it will not change anything, but I want to know if you implemented a framework as seaside, does it requires other things from the context and why ?
To my thinking playing with pc is awfully evil, it should be read-only, only to allow debugger to map it to source code. Same goes for stack pointer.
There are good uses for this I just used it to implement basic-block coverage for methods. One can use it for exception handling (mustBeBoolean etc). I'm always wary of introducing restrictions when the system has worked fine without them for years. One exception recently was in adding bounds checking to CompiledMethod>>at:put: so one could not change arbitrary bytes in the literals via at:put:. But this was adding safety, not adding a restriction. I see no harm in allowing one to assign the pc. It is unsafe, and the system will likely crash if you get it wrong. But it is also potentially useful (e.g. I've used it to do a prototype of tail-recursion elimination).
Do you mean just eliminating tail calls in some easy places, or proper tail calls? (I keep saying "proper tail call" and not "tail call optimisation/elimination" because implementing proper tail calls is _language design_, not optimising. Or, having proper tail calls lets you do things that are impossible without a global program transformation. Handily, we can often implement language design in our own language!)
frank
So let's be laissez faire, unless the freedoms in question really do only do harm.
Setting receiver is less evil and it may work, if receiver belongs to same inheritance chain as compiled method's class. But it should not be set without checking this.
And for method's temps, it is fine.
Thanks for answering,
-- Best regards, Igor Stasenko.
-- best, Eliot
On Tue, Jul 9, 2013 at 11:29 PM, Frank Shearar frank.shearar@gmail.comwrote:
On 9 July 2013 23:03, Eliot Miranda eliot.miranda@gmail.com wrote:
On Tue, Jul 9, 2013 at 1:27 AM, Igor Stasenko siguctua@gmail.com
wrote:
On 9 July 2013 10:15, Clément Bera bera.clement@gmail.com wrote:
Hello pharoers,
Recently I have been looking at the Pharo thisContext capabilities,
in order
to perhaps one day in the far future edit its implementation with the
Pharo
dev team. Nothing is planned or confirmed, it is just to discuss.
A context has instances variables (method closureOrNil receiver stackp sender pc) and holds the arguments and temporaries.
Now I'd like to know what context's state do we modify, and what
states are
just internal representations ?
For example, it seems that you can do 'thisContext receiver: #foo',
but you
cannot with Cog.
SomeClass>>foo thisContext receiver: #foo. ^ self
In workspace, evaluating: 1 to: 5 do: [:i | Transcript show: SomeClass new foo ]
Transcript result (with Cog): foo foo a SomeClass a SomeClass a SomeClass
Transcript result (with Stack or Vanilla): foo foo foo foo foo
Now as no one has ever complained, I guess this feature is not used.
As far as I know, the real use cases of the context seems to be:
- setting and gettings temporaries
- setting and getting the sender of a context
- setting and getting the pc
- getting method, closureOrNil, receiver, stackp, arguments but NOT
setting
them
Now setting the sender of a context seems to be used only in two
cases:
- continuations (as seaside continuations)
- exception implementation
So imagine that in the future you would have a context that can be
accessed
in read-only, where you could only:
- set the temporaries (but not arguments)
- set the pc (or something equivalent, as set the currently executed
ast
node)
- use continuations (exceptions can be implemented on top of
continuations)
I would like to know if there are things that you do now and that you
would
not be able to do with a context like that. For non meta developer
(like
enterprise app developer) I guess it will not change anything, but I
want to
know if you implemented a framework as seaside, does it requires other things from the context and why ?
To my thinking playing with pc is awfully evil, it should be read-only, only to allow debugger to map it to source code. Same goes for stack pointer.
There are good uses for this I just used it to implement basic-block
coverage for methods. One can use it for exception handling (mustBeBoolean etc). I'm always wary of introducing restrictions when the system has worked fine without them for years. One exception recently was in adding bounds checking to CompiledMethod>>at:put: so one could not change arbitrary bytes in the literals via at:put:. But this was adding safety, not adding a restriction. I see no harm in allowing one to assign the pc. It is unsafe, and the system will likely crash if you get it wrong. But it is also potentially useful (e.g. I've used it to do a prototype of tail-recursion elimination).
Do you mean just eliminating tail calls in some easy places, or proper tail calls? (I keep saying "proper tail call" and not "tail call optimisation/elimination" because implementing proper tail calls is _language design_, not optimising. Or, having proper tail calls lets you do things that are impossible without a global program transformation. Handily, we can often implement language design in our own language!)
I mean a hack that allows a context to restart itself with any of different receiver/arguments/method. e.g. thisContext tailSend: selector to: receiver with: arg with: arg. More a test of contexts than a useful facility. But a fun experiment all the same.
frank
So let's be laissez faire, unless the freedoms in question really do
only do harm.
Setting receiver is less evil and it may work, if receiver belongs to same inheritance chain as compiled method's class. But it should not be set without checking
this.
And for method's temps, it is fine.
Thanks for answering,
-- Best regards, Igor Stasenko.
-- best, Eliot
On 10 July 2013 00:03, Eliot Miranda eliot.miranda@gmail.com wrote:
On Tue, Jul 9, 2013 at 1:27 AM, Igor Stasenko siguctua@gmail.com wrote:
On 9 July 2013 10:15, Clément Bera bera.clement@gmail.com wrote:
Hello pharoers,
Recently I have been looking at the Pharo thisContext capabilities, in order to perhaps one day in the far future edit its implementation with the Pharo dev team. Nothing is planned or confirmed, it is just to discuss.
A context has instances variables (method closureOrNil receiver stackp sender pc) and holds the arguments and temporaries.
Now I'd like to know what context's state do we modify, and what states are just internal representations ?
For example, it seems that you can do 'thisContext receiver: #foo', but you cannot with Cog.
SomeClass>>foo thisContext receiver: #foo. ^ self
In workspace, evaluating: 1 to: 5 do: [:i | Transcript show: SomeClass new foo ]
Transcript result (with Cog): foo foo a SomeClass a SomeClass a SomeClass
Transcript result (with Stack or Vanilla): foo foo foo foo foo
Now as no one has ever complained, I guess this feature is not used.
As far as I know, the real use cases of the context seems to be:
- setting and gettings temporaries
- setting and getting the sender of a context
- setting and getting the pc
- getting method, closureOrNil, receiver, stackp, arguments but NOT setting
them
Now setting the sender of a context seems to be used only in two cases:
- continuations (as seaside continuations)
- exception implementation
So imagine that in the future you would have a context that can be accessed in read-only, where you could only:
- set the temporaries (but not arguments)
- set the pc (or something equivalent, as set the currently executed ast
node)
- use continuations (exceptions can be implemented on top of continuations)
I would like to know if there are things that you do now and that you would not be able to do with a context like that. For non meta developer (like enterprise app developer) I guess it will not change anything, but I want to know if you implemented a framework as seaside, does it requires other things from the context and why ?
To my thinking playing with pc is awfully evil, it should be read-only, only to allow debugger to map it to source code. Same goes for stack pointer.
There are good uses for this I just used it to implement basic-block coverage for methods. One can use it for exception handling (mustBeBoolean etc). I'm always wary of introducing restrictions when the system has worked fine without them for years. One exception recently was in adding bounds checking to CompiledMethod>>at:put: so one could not change arbitrary bytes in the literals via at:put:. But this was adding safety, not adding a restriction. I see no harm in allowing one to assign the pc. It is unsafe, and the system will likely crash if you get it wrong. But it is also potentially useful (e.g. I've used it to do a prototype of tail-recursion elimination). So let's be laissez faire, unless the freedoms in question really do only do harm.
Apparently one should know what he doing, when assigning to context's pc. And that's the reason why it considered 'evil' :)
To me this is strange way to program: context just reflecting activation of your program, so if you changing pc, you changing the flow of your program. Isn't it would be wiser in such case to just change the original program to make it behave like you want, instead of hacking it via setting context's pc?
Of course, there is exceptions like tail-recursion elimination, which you cannot implement by changing code in your program.
Setting receiver is less evil and it may work, if receiver belongs to same inheritance chain as compiled method's class. But it should not be set without checking this.
And for method's temps, it is fine.
Thanks for answering,
-- Best regards, Igor Stasenko.
-- best, Eliot
On 2013-07-10, at 11:37, Igor Stasenko siguctua@gmail.com wrote:
Of course, there is exceptions like tail-recursion elimination, which you cannot implement by changing code in your program.
Or implementing GOTO. You've got to be able to have GOTO ;)
- Bert -
On 10 July 2013 11:43, Bert Freudenberg bert@freudenbergs.de wrote:
On 2013-07-10, at 11:37, Igor Stasenko siguctua@gmail.com wrote:
Of course, there is exceptions like tail-recursion elimination, which you cannot implement by changing code in your program.
Or implementing GOTO. You've got to be able to have GOTO ;)
But at which level? if at level of compiler , then it is completely acceptable. But not at level of direct manipulation. What i mean that for all such cases there should be a higher abstraction layer, allowing user to do what it does.. but not allow user to code: context pc: 10.
- Bert -
vm-dev@lists.squeakfoundation.org
-
Bert Freudenberg -
Eliot Miranda -
Frank Shearar -
Igor Stasenko