On 3/16/2022 8:33 PM, sean@clipperadams.com wrote:

I care because it makes “everything is a message” a lie! And I suspect (no proof and could be wrong) it’s an optimization that only made sense with the hardware constraints of 40+ years ago. Arguing against premature optimization is hardly something I just made up ;-) When I read that, I remember that programming languages are a category of software that cannot run fast enough. So they still make sense more often than not. Moreover the few messages that are inlined (or optimized in general) are heavily used in the core of Pharo. Also, the quote that you're citing, from Donald Knuth, is often if not always used incompletely. I could try to justify it, but I've learned that Scott Meyers is better at explaining things than i'll ever be [1] =)

Also, I have been overriding core messages that are optimized (#ifTrue and variant. ifNil and variant and many more) without any problem from the compiler, so I have had no need to disable this inlining. Regardless, you should be able to remove all inlining, it's a matter of tweaking the compiler. However the compiler is one thing that I have touched very little in the system. Pierre  [1] https://www.youtube.com/watch?v=3WBaY61c9sE

Richard, My _concern_ with inlining is that since it is designed to short-circuit dynamic method lookup, it is impossible to call a _different_ implementation. That is, you lose the opportunity to have the _receiver_ decide how to respond to the message. You may think of it as a message, but the caller is deciding how the receiver will respond—which largely defeats the purpose and role of it being a message. Yes, at the machine code level you are performing a branch instruction, but when comparing OOP to Procedural Programming we typically make a distinction between “messages” and "procedure calls." The distinction is that the receiver gets to decide how to respond to a message. In C++ this is the distinction between a “virtual" and "non-virtual" function. By inlining, you are converting the function from a virtual function to a non-virtual function, and this can make a difference (which is why virtual functions exist). How should a proxy (https://en.wikipedia.org/wiki/Proxy_pattern <https://en.wikipedia.org/wiki/Proxy_pattern>) to nil respond to the #’isNil’ message? How should the Null Object Pattern (https://en.wikipedia.org/wiki/Null_object_pattern <https://en.wikipedia.org/wiki/Null_object_pattern>) respond to #’isNil’? And, yes, I’m sure you can come up with benchmarks that show a measurable difference, but what is the impact in realistic code? When someone asked about inlining #’yourself’ in GemStone I believe I measured the performance as taking 2 nanoseconds per call (on a 2012 machine). A 10% speedup would make it 1.8 nanoseconds. Is that worth it? Maybe, maybe not. Note that I described my position as a “concern,” not an ideological objection. Mostly I’m giving a rationale for something that doesn’t seem to be explained very well for you. I accept that there may be a time for inlining, but I can “comprehend" another side to the issue. James

On Mar 16, 2022, at 9:42 PM, Richard O'Keefe <raoknz@gmail.com> wrote:

We're still not on the same page. You seem to have some ideological objection to inlining that I am completely failing to comprehend. Just because a procedure call (message send) is inlined doesn't in the least mean it *isn't* a procedure call (message send), just as compiling a procedure call (message send) as a jump (last-call optimisation) doesn't mean it *isn't* a procedure call (message send). By the way, forget about "40 years ago". I just did an experiment in Pharo 9, and found that using "_ ifNotNil: " instead of "_ izNil ifFalse: " -- where izNil is a non-inlined self == nil -- gave a 10% speedup, in a test code where real work was going on as well. As for turning off all inlining, what do you think that would do to #ifFalse:ifTrue: and its relatives?

On Thu, 17 Mar 2022 at 08:34, <sean@clipperadams.com <mailto:sean@clipperadams.com>> wrote:

To start with, why do you CARE whether a particular method is inlined or not?

I care because it makes “everything is a message” a lie! And I suspect (no proof and could be wrong) it’s an optimization that only made sense with the hardware constraints of 40+ years ago. Arguing against premature optimization is hardly something I just made up ;-)

This makes absolutely no sense to me. What makes you think that the combination "_ isNil ifFalse: [_]" will NOT be inlined?

I may have been unclear. My intent was to communicate: “I’d like to stop ALL* inlining of messages by default if possible”

*or as many as practical

The thing that rings loud alarm bells for me is there being "long chains" in the first place.

I agree that it is in general a smell, but long chains was tangential to the intention above

Can you give an example?

I don’t know if I can think of one that’s not contrived… Wrapping something external? Squeak’s AppleScript support used to mirror the underlying AS, which is pretty much exactly that.

In my own programming, I've generally found that nils turning up in the middle of a chain indicates a serious design error somewhere.

Agreed. See smell comment above.

My chief concern is that I am a bear of very little brain, and if you change the meaning of #isNil to anything at all other than "is the receiver identical to nil" you *WILL* (not may) confuse me. This extends to things that happen not to be inlined: if even a god-like Smalltalker like Andres Valloud overloads #, to something other than "combine the collection that is the receiver with the collection that is the argument to yield a new collection" than I *WILL* most certainly be confused and find the code unmaintainable. Smalltalk being Smalltalk, if you admit an inconsistent overload anywhere, I can no longer understand sends of that selector anywhere. One of the things to like about Traits is that you can say "this class doesn't just *happen* to have selectors x and y, it has them *because* it has this whole consistent bundle of selectors." There are more annotations documented for my Smalltalk compiler than are actually implemented. One that *is* implemented is <doNotOverride>, and it has caught more mistakes than I care to admit to. It's particularly important for a bundle of methods with varying arguments that are meant to be routed through a single method, which *is* meant to be overridden. It makes sure that I override the *right* method. (Take #= and #~= as an obvious example.) Once you start down the path of lying about things like #isNil you find that EITHER you have to go very far down that path and override #== and #instVarAt: and a whole lot of other things OR you are working with a semantically incoherent system. "How should a proxy (https://en.wikipedia.org/wiki/Proxy_pattern) to nil respond to the #’isNil’ message?" It SHOULD NOT LIE. A proxy *isn't* nil, and it doesn't *behave* like nil, even if it is a proxy for nil. A proxy, qua proxy, can do things that nil cannot. Use another selector, #isEffectivelyNil, or whatever reveals your intentions, and give it what semantics you find useful. "How should the Null Object Pattern ( https://en.wikipedia.org/wiki/Null_object_pattern) respond to #’isNil’?" It should answer false. Period. No ifs, buts, quibbles, or maybes. The whole *point* of the Null Object Pattern is to return something that *isn't* nil, that has quite a different protocol. If you call something that is supposed to return either a Foo or a NullFoo, and it gives you nil instead, there is a BUG in what you just called so the sooner you find out the better. What you want is Foo >> isNullFoo ^false NullFoo >> isNullFoo ^true and no other class defines #isNullFoo. That way, when you ask "tripeWorks grobblingMill lastPallet isNullPallet" (a) you make it clear to someone reading your code what you are expecting (b) if you DON'T get what you are expecting, Smalltalk will tell you. I must admit that on the few occasions when I've used Null Object I've used an all-purpose #isMissing instead of a task-appropriate #isNullFoo, but then I figured out what I was doing wrong. You look at the code, you say "there's *something* off here, but I don't know what." But when you ask "what, exactly, does this method MEAN?" you realise "oh, THAT'S what I was doing wrong." #isMissing told me it was a NullPerson or a NullAddress or a NullPartsList or ... but in this case I needed to know whether it was a NullSummary. And of course you run into E.F.Codd's lesson: "one NULL is never enough, information can be missing for more than one reason". Take the familiar example of a phone number: - I know that Fred's phone number is X - I know that Fred has a phone but I don't know what the number is - I don't know whether Fred has a phone or not - I know that Fred has no phone There's room for three *different* null objects there. Should we have UnknownNumberOfActualPhone to answer false to #isnil and NonNumberOfNonexistentPhone to answer true? Or what? By the way, you may have misunderstood my benchmark. It wasn't that #isNil or even _ ifNotNil: speeded up by 10%, it was the *whole* matrix-munching benchmark that speeded up. Certainly not a big deal, but it's not something I'd be happy to give up in order to get less maintainable code. On Thu, 17 Mar 2022 at 18:21, James Foster <smalltalk@jgfoster.net> wrote:

Richard,

My _concern_ with inlining is that since it is designed to short-circuit dynamic method lookup, it is impossible to call a _different_ implementation. That is, you lose the opportunity to have the _receiver_ decide how to respond to the message. You may think of it as a message, but the caller is deciding how the receiver will respond—which largely defeats the purpose and role of it being a message. Yes, at the machine code level you are performing a branch instruction, but when comparing OOP to Procedural Programming we typically make a distinction between “messages” and "procedure calls." The distinction is that the receiver gets to decide how to respond to a message. In C++ this is the distinction between a “virtual" and "non-virtual" function. By inlining, you are converting the function from a virtual function to a non-virtual function, and this can make a difference (which is why virtual functions exist).

How should a proxy (https://en.wikipedia.org/wiki/Proxy_pattern) to nil respond to the #’isNil’ message? How should the Null Object Pattern ( https://en.wikipedia.org/wiki/Null_object_pattern) respond to #’isNil’?

And, yes, I’m sure you can come up with benchmarks that show a measurable difference, but what is the impact in realistic code? When someone asked about inlining #’yourself’ in GemStone I believe I measured the performance as taking 2 nanoseconds per call (on a 2012 machine). A 10% speedup would make it 1.8 nanoseconds. Is that worth it? Maybe, maybe not.

Note that I described my position as a “concern,” not an ideological objection. Mostly I’m giving a rationale for something that doesn’t seem to be explained very well for you. I accept that there may be a time for inlining, but I can “comprehend" another side to the issue.

James

On Mar 16, 2022, at 9:42 PM, Richard O'Keefe <raoknz@gmail.com> wrote:

We're still not on the same page. You seem to have some ideological objection to inlining that I am completely failing to comprehend. Just because a procedure call (message send) is inlined doesn't in the least mean it *isn't* a procedure call (message send), just as compiling a procedure call (message send) as a jump (last-call optimisation) doesn't mean it *isn't* a procedure call (message send). By the way, forget about "40 years ago". I just did an experiment in Pharo 9, and found that using "_ ifNotNil: " instead of "_ izNil ifFalse: " -- where izNil is a non-inlined self == nil -- gave a 10% speedup, in a test code where real work was going on as well. As for turning off all inlining, what do you think that would do to #ifFalse:ifTrue: and its relatives?

On Thu, 17 Mar 2022 at 08:34, <sean@clipperadams.com> wrote:

...

To start with, why do you CARE whether a particular method is inlined or not?

I care because it makes “everything is a message” a lie! And I suspect (no proof and could be wrong) it’s an optimization that only made sense with the hardware constraints of 40+ years ago. Arguing against premature optimization is hardly something I just made up ;-)

This makes absolutely no sense to me. What makes you think that the combination "_ isNil ifFalse: [_]" will NOT be inlined?

I may have been unclear. My intent was to communicate: “I’d like to stop ALL* inlining of messages by default if possible”

*or as many as practical

The thing that rings loud alarm bells for me is there being "long chains" in the first place.

I agree that it is in general a smell, but long chains was tangential to the intention above

Can you give an example?

I don’t know if I can think of one that’s not contrived… Wrapping something external? Squeak’s AppleScript support used to mirror the underlying AS, which is pretty much exactly that.

In my own programming, I've generally found that nils turning up in the middle of a chain indicates a serious design error somewhere.

Agreed. See smell comment above.

Just to add another argument in favor of not inlining some message sends, is for instance in things like Glorp you need to use a special selector for #and:/#or: message sends, breaking not only the naming conventions but also the polymorphism. E.g. you have to use #AND: and #OR: to avoid inlining. db read: YourClass where: [:inst | inst name = 'John' AND: [inst color = 'blue' OR: [inst color = 'red']]]. This is not only annoying, it also causes that if you want to to use something else as a backend (instead of Glorp), your block doesn't work anymore, which would perfectly work for a regular collection of elements using their underscore #and:/#or: equivalents. The important thing, IMO, is that the core of OO is message sending, and enabling the receiver what to do with a message, which in the worst case would mean "I don't understand this message". A procedure call inverts the responsibility. A great trade off would be to be able to enable/disable for certain things instead of it being a global setting. Regards, Esteban A. Maringolo On Thu, Mar 17, 2022 at 11:53 AM James Foster <smalltalk@jgfoster.net> wrote:

Richard,

I very much admire Dijkstra’s admonition regarding “The Humble Programmer” and was pointing a student to that article just this week.

In any case, I think you’ve demonstrated that you now comprehend the argument against inlining—you just don’t agree. That’s fair and I think the discussion has been clarified. Would it be fair to say that you have an “ideological objection” to allowing a Proxy or Stub to transparently stand in for another object (say, in a two-object-space environment such as Pharo and GemStone)? That is, a domain object can’t be replaced by a Proxy or Stub without a wholesale rewrite of the rest of the application? I respect that as a reasonable position (demanding perfect clarity), but I see a cost to that position as well.

Of course, if you really want to avoid allowing the receiver to chose its response to a message, you can use other messages. So if you want to find out if an object is identical to nil you should use `nil == myObject` to ensure that there was not an override of #’isNil’ or #’==‘ by the object’s class.

James

On Mar 17, 2022, at 2:27 AM, Richard O'Keefe <raoknz@gmail.com> wrote:

My chief concern is that I am a bear of very little brain, and if you change the meaning of #isNil to anything at all other than "is the receiver identical to nil" you *WILL* (not may) confuse me. This extends to things that happen not to be inlined: if even a god-like Smalltalker like Andres Valloud overloads #, to something other than "combine the collection that is the receiver with the collection that is the argument to yield a new collection" than I *WILL* most certainly be confused and find the code unmaintainable. Smalltalk being Smalltalk, if you admit an inconsistent overload anywhere, I can no longer understand sends of that selector anywhere. One of the things to like about Traits is that you can say "this class doesn't just *happen* to have selectors x and y, it has them *because* it has this whole consistent bundle of selectors."

There are more annotations documented for my Smalltalk compiler than are actually implemented. One that *is* implemented is <doNotOverride>, and it has caught more mistakes than I care to admit to. It's particularly important for a bundle of methods with varying arguments that are meant to be routed through a single method, which *is* meant to be overridden. It makes sure that I override the *right* method. (Take #= and #~= as an obvious example.)

Once you start down the path of lying about things like #isNil you find that EITHER you have to go very far down that path and override #== and #instVarAt: and a whole lot of other things OR you are working with a semantically incoherent system.

"How should a proxy (https://en.wikipedia.org/wiki/Proxy_pattern) to nil respond to the #’isNil’ message?"

It SHOULD NOT LIE. A proxy *isn't* nil, and it doesn't *behave* like nil, even if it is a proxy for nil. A proxy, qua proxy, can do things that nil cannot. Use another selector, #isEffectivelyNil, or whatever reveals your intentions, and give it what semantics you find useful.

"How should the Null Object Pattern (https://en.wikipedia.org/wiki/Null_object_pattern) respond to #’isNil’?"

It should answer false. Period. No ifs, buts, quibbles, or maybes. The whole *point* of the Null Object Pattern is to return something that *isn't* nil, that has quite a different protocol. If you call something that is supposed to return either a Foo or a NullFoo, and it gives you nil instead, there is a BUG in what you just called so the sooner you find out the better. What you want is

Foo >> isNullFoo ^false NullFoo >> isNullFoo ^true and no other class defines #isNullFoo.

That way, when you ask "tripeWorks grobblingMill lastPallet isNullPallet" (a) you make it clear to someone reading your code what you are expecting (b) if you DON'T get what you are expecting, Smalltalk will tell you.

I must admit that on the few occasions when I've used Null Object I've used an all-purpose #isMissing instead of a task-appropriate #isNullFoo, but then I figured out what I was doing wrong. You look at the code, you say "there's *something* off here, but I don't know what." But when you ask "what, exactly, does this method MEAN?" you realise "oh, THAT'S what I was doing wrong." #isMissing told me it was a NullPerson or a NullAddress or a NullPartsList or ... but in this case I needed to know whether it was a NullSummary.

And of course you run into E.F.Codd's lesson: "one NULL is never enough, information can be missing for more than one reason". Take the familiar example of a phone number: - I know that Fred's phone number is X - I know that Fred has a phone but I don't know what the number is - I don't know whether Fred has a phone or not - I know that Fred has no phone There's room for three *different* null objects there. Should we have UnknownNumberOfActualPhone to answer false to #isnil and NonNumberOfNonexistentPhone to answer true? Or what?

By the way, you may have misunderstood my benchmark. It wasn't that #isNil or even _ ifNotNil: speeded up by 10%, it was the *whole* matrix-munching benchmark that speeded up. Certainly not a big deal, but it's not something I'd be happy to give up in order to get less maintainable code.

On Thu, 17 Mar 2022 at 18:21, James Foster <smalltalk@jgfoster.net> wrote:

...

Richard,

My _concern_ with inlining is that since it is designed to short-circuit dynamic method lookup, it is impossible to call a _different_ implementation. That is, you lose the opportunity to have the _receiver_ decide how to respond to the message. You may think of it as a message, but the caller is deciding how the receiver will respond—which largely defeats the purpose and role of it being a message. Yes, at the machine code level you are performing a branch instruction, but when comparing OOP to Procedural Programming we typically make a distinction between “messages” and "procedure calls." The distinction is that the receiver gets to decide how to respond to a message. In C++ this is the distinction between a “virtual" and "non-virtual" function. By inlining, you are converting the function from a virtual function to a non-virtual function, and this can make a difference (which is why virtual functions exist).

How should a proxy (https://en.wikipedia.org/wiki/Proxy_pattern) to nil respond to the #’isNil’ message? How should the Null Object Pattern (https://en.wikipedia.org/wiki/Null_object_pattern) respond to #’isNil’?

And, yes, I’m sure you can come up with benchmarks that show a measurable difference, but what is the impact in realistic code? When someone asked about inlining #’yourself’ in GemStone I believe I measured the performance as taking 2 nanoseconds per call (on a 2012 machine). A 10% speedup would make it 1.8 nanoseconds. Is that worth it? Maybe, maybe not.

Note that I described my position as a “concern,” not an ideological objection. Mostly I’m giving a rationale for something that doesn’t seem to be explained very well for you. I accept that there may be a time for inlining, but I can “comprehend" another side to the issue.

James

On Mar 16, 2022, at 9:42 PM, Richard O'Keefe <raoknz@gmail.com> wrote:

We're still not on the same page. You seem to have some ideological objection to inlining that I am completely failing to comprehend. Just because a procedure call (message send) is inlined doesn't in the least mean it *isn't* a procedure call (message send), just as compiling a procedure call (message send) as a jump (last-call optimisation) doesn't mean it *isn't* a procedure call (message send). By the way, forget about "40 years ago". I just did an experiment in Pharo 9, and found that using "_ ifNotNil: " instead of "_ izNil ifFalse: " -- where izNil is a non-inlined self == nil -- gave a 10% speedup, in a test code where real work was going on as well. As for turning off all inlining, what do you think that would do to #ifFalse:ifTrue: and its relatives?

On Thu, 17 Mar 2022 at 08:34, <sean@clipperadams.com> wrote:

...

To start with, why do you CARE whether a particular method is inlined or not?

I care because it makes “everything is a message” a lie! And I suspect (no proof and could be wrong) it’s an optimization that only made sense with the hardware constraints of 40+ years ago. Arguing against premature optimization is hardly something I just made up ;-)

This makes absolutely no sense to me. What makes you think that the combination "_ isNil ifFalse: [_]" will NOT be inlined?

I may have been unclear. My intent was to communicate: “I’d like to stop ALL* inlining of messages by default if possible”

*or as many as practical

The thing that rings loud alarm bells for me is there being "long chains" in the first place.

I agree that it is in general a smell, but long chains was tangential to the intention above

Can you give an example?

I don’t know if I can think of one that’s not contrived… Wrapping something external? Squeak’s AppleScript support used to mirror the underlying AS, which is pretty much exactly that.

In my own programming, I've generally found that nils turning up in the middle of a chain indicates a serious design error somewhere.

Agreed. See smell comment above.

On 17 Mar 2022, at 17:17, Richard Sargent <richard.sargent@gemtalksystems.com> wrote:

A pragma might be the way to go.

<noInline: #and:> for example.

There is already the #compilerOptions: Pragma, e.g.: <compilerOptions: #(- optionInlineAndOr )> For example this method compiled the and: as a real send: tt <compilerOptions: #(- optionInlineAndOr)> true and: [ false ]

On Thu, Mar 17, 2022 at 8:17 AM Esteban Maringolo <emaringolo@gmail.com <mailto:emaringolo@gmail.com>> wrote: Just to add another argument in favor of not inlining some message sends, is for instance in things like Glorp you need to use a special selector for #and:/#or: message sends, breaking not only the naming conventions but also the polymorphism.

E.g. you have to use #AND: and #OR: to avoid inlining. db read: YourClass where: [:inst | inst name = 'John' AND: [inst color = 'blue' OR: [inst color = 'red']]].

This is not only annoying, it also causes that if you want to to use something else as a backend (instead of Glorp), your block doesn't work anymore, which would perfectly work for a regular collection of elements using their underscore #and:/#or: equivalents.

The important thing, IMO, is that the core of OO is message sending, and enabling the receiver what to do with a message, which in the worst case would mean "I don't understand this message". A procedure call inverts the responsibility. A great trade off would be to be able to enable/disable for certain things instead of it being a global setting.

Regards,

Esteban A. Maringolo

On Thu, Mar 17, 2022 at 11:53 AM James Foster <smalltalk@jgfoster.net <mailto:smalltalk@jgfoster.net>> wrote:

...

Richard,

I very much admire Dijkstra’s admonition regarding “The Humble Programmer” and was pointing a student to that article just this week.

In any case, I think you’ve demonstrated that you now comprehend the argument against inlining—you just don’t agree. That’s fair and I think the discussion has been clarified. Would it be fair to say that you have an “ideological objection” to allowing a Proxy or Stub to transparently stand in for another object (say, in a two-object-space environment such as Pharo and GemStone)? That is, a domain object can’t be replaced by a Proxy or Stub without a wholesale rewrite of the rest of the application? I respect that as a reasonable position (demanding perfect clarity), but I see a cost to that position as well.

Of course, if you really want to avoid allowing the receiver to chose its response to a message, you can use other messages. So if you want to find out if an object is identical to nil you should use `nil == myObject` to ensure that there was not an override of #’isNil’ or #’==‘ by the object’s class.

James

On Mar 17, 2022, at 2:27 AM, Richard O'Keefe <raoknz@gmail.com <mailto:raoknz@gmail.com>> wrote:

My chief concern is that I am a bear of very little brain, and if you change the meaning of #isNil to anything at all other than "is the receiver identical to nil" you *WILL* (not may) confuse me. This extends to things that happen not to be inlined: if even a god-like Smalltalker like Andres Valloud overloads #, to something other than "combine the collection that is the receiver with the collection that is the argument to yield a new collection" than I *WILL* most certainly be confused and find the code unmaintainable. Smalltalk being Smalltalk, if you admit an inconsistent overload anywhere, I can no longer understand sends of that selector anywhere. One of the things to like about Traits is that you can say "this class doesn't just *happen* to have selectors x and y, it has them *because* it has this whole consistent bundle of selectors."

There are more annotations documented for my Smalltalk compiler than are actually implemented. One that *is* implemented is <doNotOverride>, and it has caught more mistakes than I care to admit to. It's particularly important for a bundle of methods with varying arguments that are meant to be routed through a single method, which *is* meant to be overridden. It makes sure that I override the *right* method. (Take #= and #~= as an obvious example.)

Once you start down the path of lying about things like #isNil you find that EITHER you have to go very far down that path and override #== and #instVarAt: and a whole lot of other things OR you are working with a semantically incoherent system.

"How should a proxy (https://en.wikipedia.org/wiki/Proxy_pattern <https://en.wikipedia.org/wiki/Proxy_pattern>) to nil respond to the #’isNil’ message?"

It SHOULD NOT LIE. A proxy *isn't* nil, and it doesn't *behave* like nil, even if it is a proxy for nil. A proxy, qua proxy, can do things that nil cannot. Use another selector, #isEffectivelyNil, or whatever reveals your intentions, and give it what semantics you find useful.

"How should the Null Object Pattern (https://en.wikipedia.org/wiki/Null_object_pattern <https://en.wikipedia.org/wiki/Null_object_pattern>) respond to #’isNil’?"

It should answer false. Period. No ifs, buts, quibbles, or maybes. The whole *point* of the Null Object Pattern is to return something that *isn't* nil, that has quite a different protocol. If you call something that is supposed to return either a Foo or a NullFoo, and it gives you nil instead, there is a BUG in what you just called so the sooner you find out the better. What you want is

Foo >> isNullFoo ^false NullFoo >> isNullFoo ^true and no other class defines #isNullFoo.

That way, when you ask "tripeWorks grobblingMill lastPallet isNullPallet" (a) you make it clear to someone reading your code what you are expecting (b) if you DON'T get what you are expecting, Smalltalk will tell you.

I must admit that on the few occasions when I've used Null Object I've used an all-purpose #isMissing instead of a task-appropriate #isNullFoo, but then I figured out what I was doing wrong. You look at the code, you say "there's *something* off here, but I don't know what." But when you ask "what, exactly, does this method MEAN?" you realise "oh, THAT'S what I was doing wrong." #isMissing told me it was a NullPerson or a NullAddress or a NullPartsList or ... but in this case I needed to know whether it was a NullSummary.

And of course you run into E.F.Codd's lesson: "one NULL is never enough, information can be missing for more than one reason". Take the familiar example of a phone number: - I know that Fred's phone number is X - I know that Fred has a phone but I don't know what the number is - I don't know whether Fred has a phone or not - I know that Fred has no phone There's room for three *different* null objects there. Should we have UnknownNumberOfActualPhone to answer false to #isnil and NonNumberOfNonexistentPhone to answer true? Or what?

By the way, you may have misunderstood my benchmark. It wasn't that #isNil or even _ ifNotNil: speeded up by 10%, it was the *whole* matrix-munching benchmark that speeded up. Certainly not a big deal, but it's not something I'd be happy to give up in order to get less maintainable code.

On Thu, 17 Mar 2022 at 18:21, James Foster <smalltalk@jgfoster.net <mailto:smalltalk@jgfoster.net>> wrote:

...

Richard,

My _concern_ with inlining is that since it is designed to short-circuit dynamic method lookup, it is impossible to call a _different_ implementation. That is, you lose the opportunity to have the _receiver_ decide how to respond to the message. You may think of it as a message, but the caller is deciding how the receiver will respond—which largely defeats the purpose and role of it being a message. Yes, at the machine code level you are performing a branch instruction, but when comparing OOP to Procedural Programming we typically make a distinction between “messages” and "procedure calls." The distinction is that the receiver gets to decide how to respond to a message. In C++ this is the distinction between a “virtual" and "non-virtual" function. By inlining, you are converting the function from a virtual function to a non-virtual function, and this can make a difference (which is why virtual functions exist).

How should a proxy (https://en.wikipedia.org/wiki/Proxy_pattern <https://en.wikipedia.org/wiki/Proxy_pattern>) to nil respond to the #’isNil’ message? How should the Null Object Pattern (https://en.wikipedia.org/wiki/Null_object_pattern <https://en.wikipedia.org/wiki/Null_object_pattern>) respond to #’isNil’?

And, yes, I’m sure you can come up with benchmarks that show a measurable difference, but what is the impact in realistic code? When someone asked about inlining #’yourself’ in GemStone I believe I measured the performance as taking 2 nanoseconds per call (on a 2012 machine). A 10% speedup would make it 1.8 nanoseconds. Is that worth it? Maybe, maybe not.

Note that I described my position as a “concern,” not an ideological objection. Mostly I’m giving a rationale for something that doesn’t seem to be explained very well for you. I accept that there may be a time for inlining, but I can “comprehend" another side to the issue.

James

On Mar 16, 2022, at 9:42 PM, Richard O'Keefe <raoknz@gmail.com <mailto:raoknz@gmail.com>> wrote:

We're still not on the same page. You seem to have some ideological objection to inlining that I am completely failing to comprehend. Just because a procedure call (message send) is inlined doesn't in the least mean it *isn't* a procedure call (message send), just as compiling a procedure call (message send) as a jump (last-call optimisation) doesn't mean it *isn't* a procedure call (message send). By the way, forget about "40 years ago". I just did an experiment in Pharo 9, and found that using "_ ifNotNil: " instead of "_ izNil ifFalse: " -- where izNil is a non-inlined self == nil -- gave a 10% speedup, in a test code where real work was going on as well. As for turning off all inlining, what do you think that would do to #ifFalse:ifTrue: and its relatives?

On Thu, 17 Mar 2022 at 08:34, <sean@clipperadams.com <mailto:sean@clipperadams.com>> wrote:

...

To start with, why do you CARE whether a particular method is inlined or not?

I care because it makes “everything is a message” a lie! And I suspect (no proof and could be wrong) it’s an optimization that only made sense with the hardware constraints of 40+ years ago. Arguing against premature optimization is hardly something I just made up ;-)

This makes absolutely no sense to me. What makes you think that the combination "_ isNil ifFalse: [_]" will NOT be inlined?

I may have been unclear. My intent was to communicate: “I’d like to stop ALL* inlining of messages by default if possible”

*or as many as practical

The thing that rings loud alarm bells for me is there being "long chains" in the first place.

I agree that it is in general a smell, but long chains was tangential to the intention above

Can you give an example?

I don’t know if I can think of one that’s not contrived… Wrapping something external? Squeak’s AppleScript support used to mirror the underlying AS, which is pretty much exactly that.

In my own programming, I've generally found that nils turning up in the middle of a chain indicates a serious design error somewhere.

Agreed. See smell comment above.

On 17 Mar 2022, at 18:00, Esteban Maringolo <emaringolo@gmail.com> wrote:

That is great!

Is it possible to disable it for doits and other playground expressions?

No, there is no hook yet… it would be interesting to see how to do that. (e.g. per playground compiler parameters, parameters for doit independent of the compiler used for compiling methods…) What you can do is turning it off globally in the setting (all compiler option are listed there) But take care: we can not recompile the image without optimzations as this adds intererrupt possibilities to code that was not interruptable before, which breaks the code in for process switching itself, I think.

Regards!

Esteban A. Maringolo

On Thu, Mar 17, 2022 at 1:23 PM Marcus Denker <marcus.denker@inria.fr> wrote:

...

On 17 Mar 2022, at 17:17, Richard Sargent <richard.sargent@gemtalksystems.com> wrote: A pragma might be the way to go. <noInline: #and:> for example.

There is already the #compilerOptions: Pragma, e.g.:

<compilerOptions: #(- optionInlineAndOr )> For example this method compiled the and: as a real send: tt <compilerOptions: #(- optionInlineAndOr)> true and: [ false ]

On Thu, Mar 17, 2022 at 8:17 AM Esteban Maringolo <emaringolo@gmail.com> wrote:

...

...

A great trade off would be to be able to enable/disable for certain things instead of it being a global setting.

I feel like you’ve latched onto something that is genuinely a non problem and if you get your way we all suffer reduced performance and a guarantee that a nil test is universally the same thing. If you want an overridable nil test, write your own nil test like isNilish and make it mean what you like without disrupting the rest of the system. Problem solved and you’re not violating the open closed principle. Sent from my iPhone

On Mar 18, 2022, at 8:39 AM, sean@clipperadams.com wrote:

 What you can do is turning it off globally in the setting (all compiler option are listed there)

But take care: we can not recompile the image without optimzations as this adds intererrupt possibilities to code that was not interruptable before, which breaks the code in for process switching itself, I think.

So it sounds like my daydream that inlining be turned off by default is a heavier lift than just flipping a switch, correct?

In principle (ignoring the amount of work involved) what are your thoughts RE my musing: “After 40+ years of Moore’s Law, can we turn off these inlines by default?” and my hypothesis that this is an optimization that made sense at the time but now adds complexity for questionable payoff?

What, exactly, is inconsistent about a key message isNil being allowed to be overridden? There is a VERY simple solution to your problem that does not involve disrupting the platform. But you'd rather change the platform. Part of the reason Smalltalk has not been the greatest platform to build software on is its extreme malleability. It is a double edged sword for sure. It is great for research. Not so hot for actual systems development. "The key in making great and growable systems is much more to design how its modules communicate rather than what their internal properties and behaviors should be. Think of the internet – to live, it (a) has to allow many different kinds of ideas and realizations that are beyond any single standard and (b) to allow varying degrees of safe interoperability between these ideas." -Alan Kay The inlined methods implement key operations that I believe *should* be immutable. Because the system is implemented in terms of itself, it is not very hard to completely destroy an image and render it unusable by overriding the wrong message in the wrong class (try overriding #environment at the class level and let me know how long before your image completely packs up under a torrent of walkbacks). If anything, I would like to see more immutable structure in the core classes. Add methods? Always. Casually override anything? There be dragons there.

On Mar 18, 2022, at 11:06 AM, sean@clipperadams.com wrote:

I feel like you’ve latched onto something that is genuinely a non problem…

I wouldn’t call complexity and lack of consistency a “non problem”, but it sounds like for you the practical implications outweigh my seemingly-somewhat-ideological/niche concerns. Is that a fair summary? In any case, I appreciate your perspective.

On Mar 18, 2022, at 11:39 AM, Esteban Maringolo <emaringolo@gmail.com> wrote:

You say that Smalltalk is not so hot for system developments because it's extremely malleable? What are you measuring it against?

Lots of things but to keep things simple lets go with Objective C since it is quite similar in concept and style but the library classes are closed to modification without extraordinary effort (you can replace a method but you're gonna have to work to do it because you don't have the source code for the core classes).

It is still easy to break almost anything, without having to resort to compiler pragmas. :-)

Yeah, that's kind of my point. I have junked a lot of images in my wake.

...

If anything, I would like to see more immutable structure in the core classes. Add methods? Always. Casually override anything? There be dragons there.

Something like "final" classes and "const" globals or not being able to extend classes such as [Proto]Object?

IDK exactly. Yeah, maybe a pragma that was part of the core library that would prevent an override of a key method that say, the browser, depends on. Smalltalk is great for building things, but unlike when I build, say, a car, I don't have to drag around the entire factory I used to build the car behind it everywhere. Toying with the idea of using one image to host remote developer tools that we do not change and using it to operate on a second image that has minimal dev tools but will be the production version of my app. That would provide the separation that would keep me from destroying toolkit by accident while trying to tweak the app. Just musing here. I've become kind of dissatisfied with Pharo development of late. Hard to build anything that lasts very long.

...

On Mar 18, 2022, at 11:06 AM, sean@clipperadams.com wrote:

I feel like you’ve latched onto something that is genuinely a non problem…

I wouldn’t call complexity and lack of consistency a “non problem”, but it sounds like for you the practical implications outweigh my seemingly-somewhat-ideological/niche concerns. Is that a fair summary? In any case, I appreciate your perspective.

I don’t understand. Wouldn’t an override of #'doesNotUnderstand:’ solve this problem? The proxies I’ve seen subclass from nil or ProtoObject and forward almost everything to the target. It’s really very easy.

On Mar 19, 2022, at 3:14 AM, Richard O'Keefe <raoknz@gmail.com> wrote:

An object should be a Proxy or Stub only with reference to a specific protocol, which should be kept narrow.

Being a Proxy is a form of coupling. Proxying a wide interface creates maintenance problems: Squeak 5.3 : Object selectors size => 485 Pharo 9.0 : Object selectors size => 435 astc : Object selectors size => 325 VW 8.3PUL : Object selectors size => 304

The interface of Object is HUGE. You want to bet that your Proxy got *all* of the methods right? This interface didn't get that way all at once; it grew. The number was 78 in Smalltalk-80. At a minimum, then, Smalltalk systems have accreted one extra Object method every two months.

So you set up your proxy to *flawlesly* mirror Object, and then, WHOOPS, upgrade Smalltalk and now there is a method that Object has and your Proxy either lacks (if it descends from ProtoObject but not Object) or inherits an inappropriate version of (if it descends from Object).

What this means is that nobody ever *does* flawlessly mock everything in the public interface of an object they are Proxying. They proxy a *limited* protocol. Because that is all they *can* do.

Look, I know that people who have been trained to work with the stuff can use C4 as cooking fuel. But I haven't had that training, so I won't touch the stuff. In the same way, I dare say there are things *you* can safely do in Smalltalk that fumblefingers here would be burnt badly by. There are many things that *can* be done that I *won't* do. In a chemistry lab, I would not work with ClF3 let alone O2F2. In Smalltalk, I don't monkey with #isNil.

On Fri, 18 Mar 2022 at 03:52, James Foster <smalltalk@jgfoster.net <mailto:smalltalk@jgfoster.net>> wrote: Richard,

I very much admire Dijkstra’s admonition regarding “The Humble Programmer” and was pointing a student to that article just this week.

In any case, I think you’ve demonstrated that you now comprehend the argument against inlining—you just don’t agree. That’s fair and I think the discussion has been clarified. Would it be fair to say that you have an “ideological objection” to allowing a Proxy or Stub to transparently stand in for another object (say, in a two-object-space environment such as Pharo and GemStone)? That is, a domain object can’t be replaced by a Proxy or Stub without a wholesale rewrite of the rest of the application? I respect that as a reasonable position (demanding perfect clarity), but I see a cost to that position as well.

Of course, if you really want to avoid allowing the receiver to chose its response to a message, you can use other messages. So if you want to find out if an object is identical to nil you should use `nil == myObject` to ensure that there was not an override of #’isNil’ or #’==‘ by the object’s class.

James

...

On Mar 17, 2022, at 2:27 AM, Richard O'Keefe <raoknz@gmail.com <mailto:raoknz@gmail.com>> wrote:

My chief concern is that I am a bear of very little brain, and if you change the meaning of #isNil to anything at all other than "is the receiver identical to nil" you *WILL* (not may) confuse me. This extends to things that happen not to be inlined: if even a god-like Smalltalker like Andres Valloud overloads #, to something other than "combine the collection that is the receiver with the collection that is the argument to yield a new collection" than I *WILL* most certainly be confused and find the code unmaintainable. Smalltalk being Smalltalk, if you admit an inconsistent overload anywhere, I can no longer understand sends of that selector anywhere. One of the things to like about Traits is that you can say "this class doesn't just *happen* to have selectors x and y, it has them *because* it has this whole consistent bundle of selectors."

There are more annotations documented for my Smalltalk compiler than are actually implemented. One that *is* implemented is <doNotOverride>, and it has caught more mistakes than I care to admit to. It's particularly important for a bundle of methods with varying arguments that are meant to be routed through a single method, which *is* meant to be overridden. It makes sure that I override the *right* method. (Take #= and #~= as an obvious example.)

Once you start down the path of lying about things like #isNil you find that EITHER you have to go very far down that path and override #== and #instVarAt: and a whole lot of other things OR you are working with a semantically incoherent system.

"How should a proxy (https://en.wikipedia.org/wiki/Proxy_pattern <https://en.wikipedia.org/wiki/Proxy_pattern>) to nil respond to the #’isNil’ message?"

It SHOULD NOT LIE. A proxy *isn't* nil, and it doesn't *behave* like nil, even if it is a proxy for nil. A proxy, qua proxy, can do things that nil cannot. Use another selector, #isEffectivelyNil, or whatever reveals your intentions, and give it what semantics you find useful.

"How should the Null Object Pattern (https://en.wikipedia.org/wiki/Null_object_pattern <https://en.wikipedia.org/wiki/Null_object_pattern>) respond to #’isNil’?"

It should answer false. Period. No ifs, buts, quibbles, or maybes. The whole *point* of the Null Object Pattern is to return something that *isn't* nil, that has quite a different protocol. If you call something that is supposed to return either a Foo or a NullFoo, and it gives you nil instead, there is a BUG in what you just called so the sooner you find out the better. What you want is

Foo >> isNullFoo ^false NullFoo >> isNullFoo ^true and no other class defines #isNullFoo.

That way, when you ask "tripeWorks grobblingMill lastPallet isNullPallet" (a) you make it clear to someone reading your code what you are expecting (b) if you DON'T get what you are expecting, Smalltalk will tell you.

I must admit that on the few occasions when I've used Null Object I've used an all-purpose #isMissing instead of a task-appropriate #isNullFoo, but then I figured out what I was doing wrong. You look at the code, you say "there's *something* off here, but I don't know what." But when you ask "what, exactly, does this method MEAN?" you realise "oh, THAT'S what I was doing wrong." #isMissing told me it was a NullPerson or a NullAddress or a NullPartsList or ... but in this case I needed to know whether it was a NullSummary.

And of course you run into E.F.Codd's lesson: "one NULL is never enough, information can be missing for more than one reason". Take the familiar example of a phone number: - I know that Fred's phone number is X - I know that Fred has a phone but I don't know what the number is - I don't know whether Fred has a phone or not - I know that Fred has no phone There's room for three *different* null objects there. Should we have UnknownNumberOfActualPhone to answer false to #isnil and NonNumberOfNonexistentPhone to answer true? Or what?

By the way, you may have misunderstood my benchmark. It wasn't that #isNil or even _ ifNotNil: speeded up by 10%, it was the *whole* matrix-munching benchmark that speeded up. Certainly not a big deal, but it's not something I'd be happy to give up in order to get less maintainable code.

On Thu, 17 Mar 2022 at 18:21, James Foster <smalltalk@jgfoster.net <mailto:smalltalk@jgfoster.net>> wrote: Richard,

My _concern_ with inlining is that since it is designed to short-circuit dynamic method lookup, it is impossible to call a _different_ implementation. That is, you lose the opportunity to have the _receiver_ decide how to respond to the message. You may think of it as a message, but the caller is deciding how the receiver will respond—which largely defeats the purpose and role of it being a message. Yes, at the machine code level you are performing a branch instruction, but when comparing OOP to Procedural Programming we typically make a distinction between “messages” and "procedure calls." The distinction is that the receiver gets to decide how to respond to a message. In C++ this is the distinction between a “virtual" and "non-virtual" function. By inlining, you are converting the function from a virtual function to a non-virtual function, and this can make a difference (which is why virtual functions exist).

How should a proxy (https://en.wikipedia.org/wiki/Proxy_pattern <https://en.wikipedia.org/wiki/Proxy_pattern>) to nil respond to the #’isNil’ message? How should the Null Object Pattern (https://en.wikipedia.org/wiki/Null_object_pattern <https://en.wikipedia.org/wiki/Null_object_pattern>) respond to #’isNil’?

And, yes, I’m sure you can come up with benchmarks that show a measurable difference, but what is the impact in realistic code? When someone asked about inlining #’yourself’ in GemStone I believe I measured the performance as taking 2 nanoseconds per call (on a 2012 machine). A 10% speedup would make it 1.8 nanoseconds. Is that worth it? Maybe, maybe not.

Note that I described my position as a “concern,” not an ideological objection. Mostly I’m giving a rationale for something that doesn’t seem to be explained very well for you. I accept that there may be a time for inlining, but I can “comprehend" another side to the issue.

James

...

On Mar 16, 2022, at 9:42 PM, Richard O'Keefe <raoknz@gmail.com <mailto:raoknz@gmail.com>> wrote:

We're still not on the same page. You seem to have some ideological objection to inlining that I am completely failing to comprehend. Just because a procedure call (message send) is inlined doesn't in the least mean it *isn't* a procedure call (message send), just as compiling a procedure call (message send) as a jump (last-call optimisation) doesn't mean it *isn't* a procedure call (message send). By the way, forget about "40 years ago". I just did an experiment in Pharo 9, and found that using "_ ifNotNil: " instead of "_ izNil ifFalse: " -- where izNil is a non-inlined self == nil -- gave a 10% speedup, in a test code where real work was going on as well. As for turning off all inlining, what do you think that would do to #ifFalse:ifTrue: and its relatives?

On Thu, 17 Mar 2022 at 08:34, <sean@clipperadams.com <mailto:sean@clipperadams.com>> wrote:

To start with, why do you CARE whether a particular method is inlined or not?

I care because it makes “everything is a message” a lie! And I suspect (no proof and could be wrong) it’s an optimization that only made sense with the hardware constraints of 40+ years ago. Arguing against premature optimization is hardly something I just made up ;-)

This makes absolutely no sense to me. What makes you think that the combination "_ isNil ifFalse: [_]" will NOT be inlined?

I may have been unclear. My intent was to communicate: “I’d like to stop ALL* inlining of messages by default if possible”

*or as many as practical

The thing that rings loud alarm bells for me is there being "long chains" in the first place.

I agree that it is in general a smell, but long chains was tangential to the intention above

Can you give an example?

I don’t know if I can think of one that’s not contrived… Wrapping something external? Squeak’s AppleScript support used to mirror the underlying AS, which is pretty much exactly that.

In my own programming, I've generally found that nils turning up in the middle of a chain indicates a serious design error somewhere.

Agreed. See smell comment above.

Hello Richard,

What part of … is hard to understand?

Did you mean for this to come off as condescending? Or are you are honestly wondering how effectively you are communicating? In either case, all of it is hard for me to understand since it doesn’t describe the Proxy Pattern which requires that "Clients can't tell whether they work with a subject or its proxy” (https://en.wikipedia.org/wiki/Proxy_pattern <https://en.wikipedia.org/wiki/Proxy_pattern>). My understanding of a proxy is that it will rarely/never happen that "there will be *more* selectors you needed to think about but didn’t." So, far from being (an instance of) the problem, an override of #’doesNotUnderstand:’ is the elegant solution.

I wanted something that was like an OrderedCollection but could only hold elements of a specific class.

It seems what you want is not a Proxy but a Facade (https://en.wikipedia.org/wiki/Facade_pattern <https://en.wikipedia.org/wiki/Facade_pattern>) in which you "may perform additional functionality [such as type checking] before ... forwarding a request” (https://en.wikipedia.org/wiki/Facade_pattern <https://en.wikipedia.org/wiki/Facade_pattern>).

I *could* make a proxy by going through #doesNotUnderstand: but I am not going to.

That’s fine, and a proxy is probably inappropriate for the things you do. But there are a few usage scenarios <https://en.wikipedia.org/wiki/Proxy_pattern#Possible_usage_scenarios> where they are appropriate. My primary experience with them is with distributed object communication <https://en.wikipedia.org/wiki/Distributed_object_communication> where they are appropriate.

I *could* make a batch of ricin and store it in my granddaughter's bedroom, but I'm not going to.

I’m glad to hear it. To take a more realistic example, I could try to fly a 747 but I’m not going to (since my commercial pilot’s license doesn’t include that rating). That doesn’t mean that no one should fly a 747 (or implement a Proxy—the people who make GemBuilder for Smalltalk <https://gemtalksystems.com/products/gbs-vw/> should certainly use proxies). I remember hearing a joke about Kent Beck holding a sign saying “Will override #doesNotUnderstand: for food”! I took it as not saying that only a genius can understand DNU but that you should think carefully about how it will be used (and in many cases something else would be appropriate). Everything else about Kent Beck’s work makes it clear that leaving behind clear maintainable code is a top priority. James

On Mar 19, 2022, at 11:11 PM, Richard O'Keefe <raoknz@gmail.com> wrote:

An override of #doesNotUnderstand: *is* (an instance of) the problem. What part of "you may think you know what to forward now, but just wait a couple of months or install an additional package and there will be *more* selectors you needed to think about but didn't" is hard to understand?

Proxies done right are proxies *relative to a specified protocol*.

Let me give you a personal example. I wanted something that was like an OrderedCollection but could only hold elements of a specific class. No worries, just make a proxy. Make a bare-bones object, define all the #add... methods to check their argument, and use #doesNotUnderstand to forward everything else to the underlying OrderedCollection.

The first problem is the selectors that the bare-bones object *does* understand by virtue of being an object. (Remember there are *hundreds* of these selectors in Squeak, Pharo, VisualWorks, even in gst there are more than 120). You must, for example, ensure that #instVarAt: is forwarded, not handled locally, BUT you must also ensure that #instVarAt:put: is handled locally, not forwarded. And then one day you find that your program has broken, because mutable collections now have aMutableCollection inPlaceCollect: collectBlock and that can put unacceptable results in the collection *without* going through any #add... method.

Eventually you realise "handing out a wrapper for this collection constrained to only allow adding certain things" is the wrong way to do it, and you do something like

constrainedAdder: constraint ^[:x | (constraint value: x) ifTrue: [self add: x] ifFalse: [DomainError receiver: self selector: #add:]]

or you do something else entirely, like handing out a wrapper with a *narrow* interface that doesn't make the slightest pretence of *being* the other object.

I *could* make a batch of ricin and store it in my granddaughter's bedroom, but I'm not going to. I *could* make a proxy by going through #doesNotUnderstand: but I am not going to. That would be a textbook example of maxim 38: "Just because it's easy for you doesn't mean it can't be hard on your clients."

On Sun, 20 Mar 2022 at 15:18, James Foster <smalltalk@jgfoster.net <mailto:smalltalk@jgfoster.net>> wrote: I don’t understand. Wouldn’t an override of #'doesNotUnderstand:’ solve this problem? The proxies I’ve seen subclass from nil or ProtoObject and forward almost everything to the target. It’s really very easy.

...

On Mar 19, 2022, at 3:14 AM, Richard O'Keefe <raoknz@gmail.com <mailto:raoknz@gmail.com>> wrote:

An object should be a Proxy or Stub only with reference to a specific protocol, which should be kept narrow.

Being a Proxy is a form of coupling. Proxying a wide interface creates maintenance problems: Squeak 5.3 : Object selectors size => 485 Pharo 9.0 : Object selectors size => 435 astc : Object selectors size => 325 VW 8.3PUL : Object selectors size => 304

The interface of Object is HUGE. You want to bet that your Proxy got *all* of the methods right? This interface didn't get that way all at once; it grew. The number was 78 in Smalltalk-80. At a minimum, then, Smalltalk systems have accreted one extra Object method every two months.

So you set up your proxy to *flawlesly* mirror Object, and then, WHOOPS, upgrade Smalltalk and now there is a method that Object has and your Proxy either lacks (if it descends from ProtoObject but not Object) or inherits an inappropriate version of (if it descends from Object).

What this means is that nobody ever *does* flawlessly mock everything in the public interface of an object they are Proxying. They proxy a *limited* protocol. Because that is all they *can* do.

Look, I know that people who have been trained to work with the stuff can use C4 as cooking fuel. But I haven't had that training, so I won't touch the stuff. In the same way, I dare say there are things *you* can safely do in Smalltalk that fumblefingers here would be burnt badly by. There are many things that *can* be done that I *won't* do. In a chemistry lab, I would not work with ClF3 let alone O2F2. In Smalltalk, I don't monkey with #isNil.

On Fri, 18 Mar 2022 at 03:52, James Foster <smalltalk@jgfoster.net <mailto:smalltalk@jgfoster.net>> wrote: Richard,

I very much admire Dijkstra’s admonition regarding “The Humble Programmer” and was pointing a student to that article just this week.

In any case, I think you’ve demonstrated that you now comprehend the argument against inlining—you just don’t agree. That’s fair and I think the discussion has been clarified. Would it be fair to say that you have an “ideological objection” to allowing a Proxy or Stub to transparently stand in for another object (say, in a two-object-space environment such as Pharo and GemStone)? That is, a domain object can’t be replaced by a Proxy or Stub without a wholesale rewrite of the rest of the application? I respect that as a reasonable position (demanding perfect clarity), but I see a cost to that position as well.

Of course, if you really want to avoid allowing the receiver to chose its response to a message, you can use other messages. So if you want to find out if an object is identical to nil you should use `nil == myObject` to ensure that there was not an override of #’isNil’ or #’==‘ by the object’s class.

James

...

On Mar 17, 2022, at 2:27 AM, Richard O'Keefe <raoknz@gmail.com <mailto:raoknz@gmail.com>> wrote:

My chief concern is that I am a bear of very little brain, and if you change the meaning of #isNil to anything at all other than "is the receiver identical to nil" you *WILL* (not may) confuse me. This extends to things that happen not to be inlined: if even a god-like Smalltalker like Andres Valloud overloads #, to something other than "combine the collection that is the receiver with the collection that is the argument to yield a new collection" than I *WILL* most certainly be confused and find the code unmaintainable. Smalltalk being Smalltalk, if you admit an inconsistent overload anywhere, I can no longer understand sends of that selector anywhere. One of the things to like about Traits is that you can say "this class doesn't just *happen* to have selectors x and y, it has them *because* it has this whole consistent bundle of selectors."

There are more annotations documented for my Smalltalk compiler than are actually implemented. One that *is* implemented is <doNotOverride>, and it has caught more mistakes than I care to admit to. It's particularly important for a bundle of methods with varying arguments that are meant to be routed through a single method, which *is* meant to be overridden. It makes sure that I override the *right* method. (Take #= and #~= as an obvious example.)

Once you start down the path of lying about things like #isNil you find that EITHER you have to go very far down that path and override #== and #instVarAt: and a whole lot of other things OR you are working with a semantically incoherent system.

"How should a proxy (https://en.wikipedia.org/wiki/Proxy_pattern <https://en.wikipedia.org/wiki/Proxy_pattern>) to nil respond to the #’isNil’ message?"

It SHOULD NOT LIE. A proxy *isn't* nil, and it doesn't *behave* like nil, even if it is a proxy for nil. A proxy, qua proxy, can do things that nil cannot. Use another selector, #isEffectivelyNil, or whatever reveals your intentions, and give it what semantics you find useful.

"How should the Null Object Pattern (https://en.wikipedia.org/wiki/Null_object_pattern <https://en.wikipedia.org/wiki/Null_object_pattern>) respond to #’isNil’?"

It should answer false. Period. No ifs, buts, quibbles, or maybes. The whole *point* of the Null Object Pattern is to return something that *isn't* nil, that has quite a different protocol. If you call something that is supposed to return either a Foo or a NullFoo, and it gives you nil instead, there is a BUG in what you just called so the sooner you find out the better. What you want is

Foo >> isNullFoo ^false NullFoo >> isNullFoo ^true and no other class defines #isNullFoo.

That way, when you ask "tripeWorks grobblingMill lastPallet isNullPallet" (a) you make it clear to someone reading your code what you are expecting (b) if you DON'T get what you are expecting, Smalltalk will tell you.

I must admit that on the few occasions when I've used Null Object I've used an all-purpose #isMissing instead of a task-appropriate #isNullFoo, but then I figured out what I was doing wrong. You look at the code, you say "there's *something* off here, but I don't know what." But when you ask "what, exactly, does this method MEAN?" you realise "oh, THAT'S what I was doing wrong." #isMissing told me it was a NullPerson or a NullAddress or a NullPartsList or ... but in this case I needed to know whether it was a NullSummary.

And of course you run into E.F.Codd's lesson: "one NULL is never enough, information can be missing for more than one reason". Take the familiar example of a phone number: - I know that Fred's phone number is X - I know that Fred has a phone but I don't know what the number is - I don't know whether Fred has a phone or not - I know that Fred has no phone There's room for three *different* null objects there. Should we have UnknownNumberOfActualPhone to answer false to #isnil and NonNumberOfNonexistentPhone to answer true? Or what?

By the way, you may have misunderstood my benchmark. It wasn't that #isNil or even _ ifNotNil: speeded up by 10%, it was the *whole* matrix-munching benchmark that speeded up. Certainly not a big deal, but it's not something I'd be happy to give up in order to get less maintainable code.

On Thu, 17 Mar 2022 at 18:21, James Foster <smalltalk@jgfoster.net <mailto:smalltalk@jgfoster.net>> wrote: Richard,

My _concern_ with inlining is that since it is designed to short-circuit dynamic method lookup, it is impossible to call a _different_ implementation. That is, you lose the opportunity to have the _receiver_ decide how to respond to the message. You may think of it as a message, but the caller is deciding how the receiver will respond—which largely defeats the purpose and role of it being a message. Yes, at the machine code level you are performing a branch instruction, but when comparing OOP to Procedural Programming we typically make a distinction between “messages” and "procedure calls." The distinction is that the receiver gets to decide how to respond to a message. In C++ this is the distinction between a “virtual" and "non-virtual" function. By inlining, you are converting the function from a virtual function to a non-virtual function, and this can make a difference (which is why virtual functions exist).

How should a proxy (https://en.wikipedia.org/wiki/Proxy_pattern <https://en.wikipedia.org/wiki/Proxy_pattern>) to nil respond to the #’isNil’ message? How should the Null Object Pattern (https://en.wikipedia.org/wiki/Null_object_pattern <https://en.wikipedia.org/wiki/Null_object_pattern>) respond to #’isNil’?

And, yes, I’m sure you can come up with benchmarks that show a measurable difference, but what is the impact in realistic code? When someone asked about inlining #’yourself’ in GemStone I believe I measured the performance as taking 2 nanoseconds per call (on a 2012 machine). A 10% speedup would make it 1.8 nanoseconds. Is that worth it? Maybe, maybe not.

Note that I described my position as a “concern,” not an ideological objection. Mostly I’m giving a rationale for something that doesn’t seem to be explained very well for you. I accept that there may be a time for inlining, but I can “comprehend" another side to the issue.

James

...

On Mar 16, 2022, at 9:42 PM, Richard O'Keefe <raoknz@gmail.com <mailto:raoknz@gmail.com>> wrote:

We're still not on the same page. You seem to have some ideological objection to inlining that I am completely failing to comprehend. Just because a procedure call (message send) is inlined doesn't in the least mean it *isn't* a procedure call (message send), just as compiling a procedure call (message send) as a jump (last-call optimisation) doesn't mean it *isn't* a procedure call (message send). By the way, forget about "40 years ago". I just did an experiment in Pharo 9, and found that using "_ ifNotNil: " instead of "_ izNil ifFalse: " -- where izNil is a non-inlined self == nil -- gave a 10% speedup, in a test code where real work was going on as well. As for turning off all inlining, what do you think that would do to #ifFalse:ifTrue: and its relatives?

On Thu, 17 Mar 2022 at 08:34, <sean@clipperadams.com <mailto:sean@clipperadams.com>> wrote:

To start with, why do you CARE whether a particular method is inlined or not?

I care because it makes “everything is a message” a lie! And I suspect (no proof and could be wrong) it’s an optimization that only made sense with the hardware constraints of 40+ years ago. Arguing against premature optimization is hardly something I just made up ;-)

This makes absolutely no sense to me. What makes you think that the combination "_ isNil ifFalse: [_]" will NOT be inlined?

I may have been unclear. My intent was to communicate: “I’d like to stop ALL* inlining of messages by default if possible”

*or as many as practical

The thing that rings loud alarm bells for me is there being "long chains" in the first place.

I agree that it is in general a smell, but long chains was tangential to the intention above

Can you give an example?

I don’t know if I can think of one that’s not contrived… Wrapping something external? Squeak’s AppleScript support used to mirror the underlying AS, which is pretty much exactly that.

In my own programming, I've generally found that nils turning up in the middle of a chain indicates a serious design error somewhere.

Agreed. See smell comment above.

Richard, My primary reference for the Proxy Pattern is the classic "Design Patterns, Elements of Reusable Object-Oriented Software" by Gamma, Helm, Johnson, and Vlissides (Addison-Wesley 1995). In describing how to implement the pattern the “Gang of Four (or GOF)" advise “using doesNotUnderstand in Smalltalk” (p. 212). Furthermore, as to typing, they say that “Proxy doesn’t always have to know the type of real subject” (213). The GOF anticipated the problem with inlined selectors and caution that “a few special messages ... are handled directly by the virtual machine [and] do not cause the usual method look-up” (p. 212). So, I think it is reasonable, in the context of a discussion of the Proxy Pattern, to suggest that the Proxy Pattern would be more effective with fewer inlined selectors. As to where the Proxy Pattern was developed, the GOF describe “Known Uses” including remote objects in NEXTSTEP (1994) and in Smalltalk (1987). In the remote object context, a Proxy has no need to know the type of the target, it just needs to know how to forward messages and return results. While there is typically a way to find out if something is a proxy or not (e.g., #’_isProxy’) the beauty of the Proxy Pattern is that you don’t have to rewrite the entire application and use something like #’apparentClass’ or #’apparentlyNil’. Only the parts of the application that care if you have a proxy (which will typically not be the main domain code) need to look more deeply (with special selectors or reflection). It appears that you believe that a untyped proxy to a remote object is bad and that all the code in an application should be written to be aware of the possible “remoteness” of certain objects. It is my understanding that a transparent untyped proxy to a remote object is well-accepted, extremely useful, and can be implemented reliably. I don’t think I can contribute much more to the conversation than that so I’ll leave it there. Regards, James

On Mar 20, 2022, at 4:55 PM, Richard O'Keefe <raoknz@gmail.com> wrote:

Never forget that the Proxy pattern was developed in the context of a TYPED programming language. So "Clients can't tell whether they work with a subject or its proxy” means WITH REFERENCE TO A STATED PROTOCOL. Doing this in Ada or Java is fine, because you define an interface type, and that makes it safe. Here is a sentence from a typical description of the pattern: "Since the proxy implements the INTERFACE as the original class, it can be passed to any client that expects a real service object." Having a specified and *enforced* smallish interface prevents the maintenance issues that plague attempts to use this pattern in Smalltalk.

Remote objects and local objects have different semantics because remote messages and local messages have different semantics. Local messages cannot be lost, duplicated, or reordered, nor can communication with a local object drop out for no local reason. Version skew is a much bigger problem with distributed systems than local ones, so once again, it is proxying WITH REFERENCE TO A STATED PROTOCOL (and that not a large one!) that counts. There is no such thing as a 100% transparent remote proxy. So it is very useful to know what an object really is as well as what it pretends to be. That is, #class #apparentClass #isNil #apparentlyNil are *different*,

So the argument thus far is - a selector should not be inlined if it might need to be overridden - such as in an application of the Proxy pattern - but well-designed Proxies are such WITH REFERENCE TO STATED PROTOCOLS - meaning that selectors that are NOT in such protocols are fair game for inlining.

Being able to proxy to another object is such an important feature, that even in ES6 (aka "Javascript") there is a Proxy object [1] that allows you to intercept/redefine operations for that object. In the case of Smalltalk, and coming back to the original topic of this thread, that is even more powerful because you have first class messages that can be resolved via DNU handling. If anything, I would like to give the receiver of a message more control of the handling of such a message rather than having a DNU as the last resort, or worse, some messages that are written to be sent, but then are inlined during execution. Regards, [1] https://developer.mozilla.org/en-US/docs/Web/JavaScript/Reference/Global_Obj... Esteban A. Maringolo On Mon, Mar 21, 2022 at 5:47 PM Tim Mackinnon <tim@testit.works> wrote:

This has been an interesting thread to read on the side, and I appreciate the thought provoking conversion.

On Sun, 20 Mar 2022, at 6:11 AM, Richard O'Keefe wrote:

An override of #doesNotUnderstand: *is* (an instance of) the problem. What part of "you may think you know what to forward now, but just wait a couple of months or install an additional package and there will be *more* selectors you needed to think about but didn't" is hard to understand?

However I am confused by the above - as in Pharo and Dolphin (and I suspect other Smalltalks) - you don't have to make your proxy a subclass of Object right? If you want to forward as much behaviour as possible - you can subclass ProtoObject (or equivalent). The width of methods on that is much reduced (59 from a quick glance), and loading other packages doesn't tend to extend at that level, so keeping things much more stable. The downside, is that such objects are tricky to view in an inspector etc. There is a mocking framework in Pharo called Ghost that leverages this quite well from memory.

So I think that some of the side arguments on downsides might not be quite as bad as indicated.

Tim

My *concern* with inlining is that since it is designed to short-circuit dynamic method lookup, it is impossible to call a *different* implementation. That is, you lose the opportunity to have the *receiver* decide how to respond to the message. You may think of it as a message, but the caller is deciding how the receiver will respond—which largely defeats the purpose and role of it being a message.

Yes, this is exactly the point I was (apparently badly) trying to convey. Thanks!

if you are using the null object pattern then you should not write those checks (ifNil:ifNotNil:, isNil, isNotNil). you should send the message to an instance of the null object and that object should decide what to do. just an opinion.

I agree as a general rule, but it can get complicated, especially when crossing library boundaries. Once one uses a null object, no one anywhere can use those messages. This might, for example, make experimentation, refactoring and porting harder. At the risk of further distracting the conversation from the question of inlining, I was having lots of trouble applying the idea you mentioned to this (low level hack): self transcriptable ifNil: \[ self notify: 'Transcript subject not found' "We want this to be resumable because maybe e.g. the model just hasn't been loaded yet" \] ifNotNil: \[ "This is inside the guard because to get the line, we need a subject" self line ifNil: \[ "Brand new snippet" self initializeLine. ^ self \] \]. And this: self snippet line position ifNotNil: \[ :position | self firePlayerRequest: \[ :player | player currentPosition: position \] \] In the second code especially, the problem is that \`position\` could be many types that are not in the same hierarchy, so to send a message to the null object instead probably means further polluting Object for the non-null case.