Can we make computing the local variables in the debugger lazy?

Previous Topic Next Topic
 
classic Classic list List threaded Threaded
12 messages Options
Reply | Threaded
Open this post in threaded view
|

Can we make computing the local variables in the debugger lazy?

Eliot Miranda-2
Hi Marcel,

    can we try and reduce the frequency at which we compute the variables in the context inspector in the debugger?  It is a noticeable performance hit.  I really like the user interface, but the performance hit is making debugging difficult.

As an example use this:

| samples sineTable sound |
"1 second of A below middle C (220Hz). 16000 / 220 is 72.72 recurring"
sineTable := SoundPlayer sineTable: 73.
sineTable doWithIndex: "And let's not deafen anyone..."
[:sample :index| sineTable at: index put: sample // 4].
samples := SoundBuffer new: 16000.
1 to: samples size by: sineTable size do:
[:i| samples replaceFrom: i to: (i + sineTable size - 1 min: 16000) with: sineTable startingAt: 1].
1 to: 146 do:
[:i|
samples at: i put: ((samples at: i) * i / 146) asInteger.
samples at: 16001 - i put: ((samples at: 16001 - i) * i / 146) asInteger].
sound := SampledSound
samples: samples
samplingRate: 16000.
sound := MixedSound new
add: sound pan: 0.25;
add: sound pan: 0.75;
yourself.
sound computeSamplesForSeconds: sound duration


Open a workspace in e.g. a 64-bit image prior to Tools-mt.965 (I used an image containing Tools-mt.942).  Debug the above in the workspace.  Position the cursor at "sound computeSamplesForSeconds: sound duration" and do "run to here".  It is essentially instantaneous.

Now do the same in a contemporary trunk image.  It takes almost 6 seconds.

I used to be able to click step as fast as I could and the system would keep up with me.  Now I find that if I click too fast I can accumulate excess clicks and when I stp clicking the system will continue stepping and go too far.

I don't want to lose the feedback the new variables list gives us.  But I do find the performance hit tedious.  I wonder could we cache the list and only update it
- when Morphic updates, and
- when the context changes?


_,,,^..^,,,_
best, Eliot


Reply | Threaded
Open this post in threaded view
|

Re: Can we make computing the local variables in the debugger lazy?

Christoph Thiede

Hi Eliot,


very nice finding once again! I #timeProfile'd the menu button action and as I expected, the most expensive operation is the shout styling by the new inspectors, including the decompilation of every method:




First, when loading ShoutCore-ct.78 (Inbox), the speed doubles (but probably that's rather a problem with my sources file, see the thread about this commit).

Second, we do not redraw the world while running to the selection, so we do not need to update the inspectors at all. I think we could split up #doStep into some #basicDoStep (which would perform the actual stepping logic) + some #updateContextDuring: (which would update the stack list and the inspectors), then we would need to trigger the updates only once from #runToSelection:.
As an alternative, we could make this method a bit more complex but responsive by applying the same updating logic from #runUntil.

What do you think? :-)

Best,
Christoph

Von: Squeak-dev <[hidden email]> im Auftrag von Eliot Miranda <[hidden email]>
Gesendet: Samstag, 19. September 2020 20:17:12
An: The general-purpose Squeak developers list; Taeumel, Marcel
Betreff: [squeak-dev] Can we make computing the local variables in the debugger lazy?
 
Hi Marcel,

    can we try and reduce the frequency at which we compute the variables in the context inspector in the debugger?  It is a noticeable performance hit.  I really like the user interface, but the performance hit is making debugging difficult.

As an example use this:

| samples sineTable sound |
"1 second of A below middle C (220Hz). 16000 / 220 is 72.72 recurring"
sineTable := SoundPlayer sineTable: 73.
sineTable doWithIndex: "And let's not deafen anyone..."
[:sample :index| sineTable at: index put: sample // 4].
samples := SoundBuffer new: 16000.
1 to: samples size by: sineTable size do:
[:i| samples replaceFrom: i to: (i + sineTable size - 1 min: 16000) with: sineTable startingAt: 1].
1 to: 146 do:
[:i|
samples at: i put: ((samples at: i) * i / 146) asInteger.
samples at: 16001 - i put: ((samples at: 16001 - i) * i / 146) asInteger].
sound := SampledSound
samples: samples
samplingRate: 16000.
sound := MixedSound new
add: sound pan: 0.25;
add: sound pan: 0.75;
yourself.
sound computeSamplesForSeconds: sound duration


Open a workspace in e.g. a 64-bit image prior to Tools-mt.965 (I used an image containing Tools-mt.942).  Debug the above in the workspace.  Position the cursor at "sound computeSamplesForSeconds: sound duration" and do "run to here".  It is essentially instantaneous.

Now do the same in a contemporary trunk image.  It takes almost 6 seconds.

I used to be able to click step as fast as I could and the system would keep up with me.  Now I find that if I click too fast I can accumulate excess clicks and when I stp clicking the system will continue stepping and go too far.

I don't want to lose the feedback the new variables list gives us.  But I do find the performance hit tedious.  I wonder could we cache the list and only update it
- when Morphic updates, and
- when the context changes?


_,,,^..^,,,_
best, Eliot


Reply | Threaded
Open this post in threaded view
|

Re: Can we make computing the local variables in the debugger lazy?

Eliot Miranda-2
Hi Christoph,

On Sat, Sep 19, 2020 at 11:57 AM Thiede, Christoph <[hidden email]> wrote:

Hi Eliot,


very nice finding once again! I #timeProfile'd the menu button action and as I expected, the most expensive operation is the shout styling by the new inspectors, including the decompilation of every method:




First, when loading ShoutCore-ct.78 (Inbox), the speed doubles (but probably that's rather a problem with my sources file, see the thread about this commit).

Second, we do not redraw the world while running to the selection, so we do not need to update the inspectors at all. I think we could split up #doStep into some #basicDoStep (which would perform the actual stepping logic) + some #updateContextDuring: (which would update the stack list and the inspectors), then we would need to trigger the updates only once from #runToSelection:.
As an alternative, we could make this method a bit more complex but responsive by applying the same updating logic from #runUntil.

Nice.  What ever is TSTTCPW :-).  But it strikes me that one way is to use addDeferredUIMessage:.  The only tricky thing is not creating lots of these unnecessarily.  But that could be done by checking if the deferred action queue is empty or not.  E.g.  rewrite this in doStep:

self contextStackIndex > 1
ifTrue: [self resetContext: newContext]
ifFalse:
[newContext == currentContext
ifTrue: [self changed: #contentsSelection.
self updateInspectors]
ifFalse: [self resetContext: newContext]].

as, say, 

self contextStackIndex > 1
ifTrue: [self resetContext: newContext]
ifFalse:
[newContext == currentContext
ifTrue: [self scheduleUIUpdate]
ifFalse: [self resetContext: newContext]].

add an inst var to hold the last deferred action, and then do something like

Debugger>>scheduleUIUpdate
    (lastScheduledUpdate notNil
    and: [WorldState lastDeferredUIMessage == lastScheduledUpdate]) ifTrue:
        [^self].
    lastScheduledUpdate := [self changed: #contentsSelection. self updateInspectors].
    WorldState addDeferredUIMessage: lastScheduledUpdate

For this we have to add the lastDeferredUIMessage accessor to all the relevant places, but it seems a nice pattern to me.

And of course, there is no need to refresh lastScheduledUpdate.  It could be 

Debugger>>scheduleUIUpdate
    lastScheduledUpdate ifNil: [lastScheduledUpdate := [self changed: #contentsSelection. self updateInspectors]].
    WorldState lastDeferredUIMessage ~~ lastScheduledUpdate ifTrue:
        [WorldState addDeferredUIMessage: lastScheduledUpdate]

What do you think? :-)

What do you like?
Best,
Christoph

Von: Squeak-dev <[hidden email]> im Auftrag von Eliot Miranda <[hidden email]>
Gesendet: Samstag, 19. September 2020 20:17:12
An: The general-purpose Squeak developers list; Taeumel, Marcel
Betreff: [squeak-dev] Can we make computing the local variables in the debugger lazy?
 
Hi Marcel,

    can we try and reduce the frequency at which we compute the variables in the context inspector in the debugger?  It is a noticeable performance hit.  I really like the user interface, but the performance hit is making debugging difficult.

As an example use this:

| samples sineTable sound |
"1 second of A below middle C (220Hz). 16000 / 220 is 72.72 recurring"
sineTable := SoundPlayer sineTable: 73.
sineTable doWithIndex: "And let's not deafen anyone..."
[:sample :index| sineTable at: index put: sample // 4].
samples := SoundBuffer new: 16000.
1 to: samples size by: sineTable size do:
[:i| samples replaceFrom: i to: (i + sineTable size - 1 min: 16000) with: sineTable startingAt: 1].
1 to: 146 do:
[:i|
samples at: i put: ((samples at: i) * i / 146) asInteger.
samples at: 16001 - i put: ((samples at: 16001 - i) * i / 146) asInteger].
sound := SampledSound
samples: samples
samplingRate: 16000.
sound := MixedSound new
add: sound pan: 0.25;
add: sound pan: 0.75;
yourself.
sound computeSamplesForSeconds: sound duration


Open a workspace in e.g. a 64-bit image prior to Tools-mt.965 (I used an image containing Tools-mt.942).  Debug the above in the workspace.  Position the cursor at "sound computeSamplesForSeconds: sound duration" and do "run to here".  It is essentially instantaneous.

Now do the same in a contemporary trunk image.  It takes almost 6 seconds.

I used to be able to click step as fast as I could and the system would keep up with me.  Now I find that if I click too fast I can accumulate excess clicks and when I stp clicking the system will continue stepping and go too far.

I don't want to lose the feedback the new variables list gives us.  But I do find the performance hit tedious.  I wonder could we cache the list and only update it
- when Morphic updates, and
- when the context changes?


_,,,^..^,,,_
best, Eliot



--
_,,,^..^,,,_
best, Eliot


Reply | Threaded
Open this post in threaded view
|

Re: Can we make computing the local variables in the debugger lazy?

Christoph Thiede

Hi Eliot, these are interesting ideas! :-)


Your pattern is cool, it reminds me a little bit of the approach we are using for Shout, see SHTextStyler >> #styleInBackgroundProcess:.

It even more reminds me of Collection >> #do:displayingProgress:every:.

And last but not least, at a fundamental level, aren't we replicating the usual stepping logic from WorldState here?


In your approach, where do you process the scheduled messages? The problem is that we are running the simulation/execution code on the UI process. See this thread, too: http://forum.world.st/Debugging-of-other-processes-with-wait-tp5119965p5120049.html


I would not use #lastDeferredUIMessage unless we can rule out that any other client is communicating with the WorldState. Which we, afaik, never can rule out since the idea of the deferred messages is that you can have them scheduled from any background process.


It would be cool to develop some reusable logic that is applicable to different domains without expensive adaption. Maybe we should just add some static generic helper method for this?


Best,

Christoph



Von: Squeak-dev <[hidden email]> im Auftrag von Eliot Miranda <[hidden email]>
Gesendet: Samstag, 19. September 2020 21:36:57
An: The general-purpose Squeak developers list
Betreff: Re: [squeak-dev] Can we make computing the local variables in the debugger lazy?
 
Hi Christoph,

On Sat, Sep 19, 2020 at 11:57 AM Thiede, Christoph <[hidden email]> wrote:

Hi Eliot,


very nice finding once again! I #timeProfile'd the menu button action and as I expected, the most expensive operation is the shout styling by the new inspectors, including the decompilation of every method:




First, when loading ShoutCore-ct.78 (Inbox), the speed doubles (but probably that's rather a problem with my sources file, see the thread about this commit).

Second, we do not redraw the world while running to the selection, so we do not need to update the inspectors at all. I think we could split up #doStep into some #basicDoStep (which would perform the actual stepping logic) + some #updateContextDuring: (which would update the stack list and the inspectors), then we would need to trigger the updates only once from #runToSelection:.
As an alternative, we could make this method a bit more complex but responsive by applying the same updating logic from #runUntil.

Nice.  What ever is TSTTCPW :-).  But it strikes me that one way is to use addDeferredUIMessage:.  The only tricky thing is not creating lots of these unnecessarily.  But that could be done by checking if the deferred action queue is empty or not.  E.g.  rewrite this in doStep:

self contextStackIndex > 1
ifTrue: [self resetContext: newContext]
ifFalse:
[newContext == currentContext
ifTrue: [self changed: #contentsSelection.
self updateInspectors]
ifFalse: [self resetContext: newContext]].

as, say, 

self contextStackIndex > 1
ifTrue: [self resetContext: newContext]
ifFalse:
[newContext == currentContext
ifTrue: [self scheduleUIUpdate]
ifFalse: [self resetContext: newContext]].

add an inst var to hold the last deferred action, and then do something like

Debugger>>scheduleUIUpdate
    (lastScheduledUpdate notNil
    and: [WorldState lastDeferredUIMessage == lastScheduledUpdate]) ifTrue:
        [^self].
    lastScheduledUpdate := [self changed: #contentsSelection. self updateInspectors].
    WorldState addDeferredUIMessage: lastScheduledUpdate

For this we have to add the lastDeferredUIMessage accessor to all the relevant places, but it seems a nice pattern to me.

And of course, there is no need to refresh lastScheduledUpdate.  It could be 

Debugger>>scheduleUIUpdate
    lastScheduledUpdate ifNil: [lastScheduledUpdate := [self changed: #contentsSelection. self updateInspectors]].
    WorldState lastDeferredUIMessage ~~ lastScheduledUpdate ifTrue:
        [WorldState addDeferredUIMessage: lastScheduledUpdate]

What do you think? :-)

What do you like?
Best,
Christoph

Von: Squeak-dev <[hidden email]> im Auftrag von Eliot Miranda <[hidden email]>
Gesendet: Samstag, 19. September 2020 20:17:12
An: The general-purpose Squeak developers list; Taeumel, Marcel
Betreff: [squeak-dev] Can we make computing the local variables in the debugger lazy?
 
Hi Marcel,

    can we try and reduce the frequency at which we compute the variables in the context inspector in the debugger?  It is a noticeable performance hit.  I really like the user interface, but the performance hit is making debugging difficult.

As an example use this:

| samples sineTable sound |
"1 second of A below middle C (220Hz). 16000 / 220 is 72.72 recurring"
sineTable := SoundPlayer sineTable: 73.
sineTable doWithIndex: "And let's not deafen anyone..."
[:sample :index| sineTable at: index put: sample // 4].
samples := SoundBuffer new: 16000.
1 to: samples size by: sineTable size do:
[:i| samples replaceFrom: i to: (i + sineTable size - 1 min: 16000) with: sineTable startingAt: 1].
1 to: 146 do:
[:i|
samples at: i put: ((samples at: i) * i / 146) asInteger.
samples at: 16001 - i put: ((samples at: 16001 - i) * i / 146) asInteger].
sound := SampledSound
samples: samples
samplingRate: 16000.
sound := MixedSound new
add: sound pan: 0.25;
add: sound pan: 0.75;
yourself.
sound computeSamplesForSeconds: sound duration


Open a workspace in e.g. a 64-bit image prior to Tools-mt.965 (I used an image containing Tools-mt.942).  Debug the above in the workspace.  Position the cursor at "sound computeSamplesForSeconds: sound duration" and do "run to here".  It is essentially instantaneous.

Now do the same in a contemporary trunk image.  It takes almost 6 seconds.

I used to be able to click step as fast as I could and the system would keep up with me.  Now I find that if I click too fast I can accumulate excess clicks and when I stp clicking the system will continue stepping and go too far.

I don't want to lose the feedback the new variables list gives us.  But I do find the performance hit tedious.  I wonder could we cache the list and only update it
- when Morphic updates, and
- when the context changes?


_,,,^..^,,,_
best, Eliot



--
_,,,^..^,,,_
best, Eliot


Reply | Threaded
Open this post in threaded view
|

Re: Can we make computing the local variables in the debugger lazy?

Eliot Miranda-2
Hi Christoph,

On Sat, Sep 19, 2020 at 1:17 PM Thiede, Christoph <[hidden email]> wrote:

Hi Eliot, these are interesting ideas! :-)


Your pattern is cool, it reminds me a little bit of the approach we are using for Shout, see SHTextStyler >> #styleInBackgroundProcess:.

It even more reminds me of Collection >> #do:displayingProgress:every:.

And last but not least, at a fundamental level, aren't we replicating the usual stepping logic from WorldState here?


I don't think so.  What we're doing is decouplng from it.  But you could be right.  It seems to me like an MVC/MVP pattern.  The model is changing, but it only makes sense to change the display when the observer can represent the model.  If one is in an animation system such as Morphic then the observer (V or V&P)  only needs display the model when rendering the next frame, if the model has changed.  So what we'd like is automatic filtering or buffering of the state the odel uses to announce it has changed.  But we don't have that.  We've just got SharedQueue.  So a poor man's solution is to use it.


In your approach, where do you process the scheduled messages? The problem is that we are running the simulation/execution code on the UI process. See this thread, too: http://forum.world.st/Debugging-of-other-processes-with-wait-tp5119965p5120049.html

I would not use #lastDeferredUIMessage unless we can rule out that any other client is communicating with the WorldState. Which we, afaik, never can rule out since the idea of the deferred messages is that you can have them scheduled from any background process.

It would be cool to develop some reusable logic that is applicable to different domains without expensive adaption. Maybe we should just add some static generic helper method for this?


Well, take a look at what I commited to inbox.  It soles the run to here problem, but does not fix fast clicking.  So I'm just pointing you to it because it may be useful.  I think you can focus on the problem better than I.  But I'm here if you need to bounce ideas off me.  Thanks for your energy here!!

Best,

Christoph



Von: Squeak-dev <[hidden email]> im Auftrag von Eliot Miranda <[hidden email]>
Gesendet: Samstag, 19. September 2020 21:36:57
An: The general-purpose Squeak developers list
Betreff: Re: [squeak-dev] Can we make computing the local variables in the debugger lazy?
 
Hi Christoph,

On Sat, Sep 19, 2020 at 11:57 AM Thiede, Christoph <[hidden email]> wrote:

Hi Eliot,


very nice finding once again! I #timeProfile'd the menu button action and as I expected, the most expensive operation is the shout styling by the new inspectors, including the decompilation of every method:




First, when loading ShoutCore-ct.78 (Inbox), the speed doubles (but probably that's rather a problem with my sources file, see the thread about this commit).

Second, we do not redraw the world while running to the selection, so we do not need to update the inspectors at all. I think we could split up #doStep into some #basicDoStep (which would perform the actual stepping logic) + some #updateContextDuring: (which would update the stack list and the inspectors), then we would need to trigger the updates only once from #runToSelection:.
As an alternative, we could make this method a bit more complex but responsive by applying the same updating logic from #runUntil.

Nice.  What ever is TSTTCPW :-).  But it strikes me that one way is to use addDeferredUIMessage:.  The only tricky thing is not creating lots of these unnecessarily.  But that could be done by checking if the deferred action queue is empty or not.  E.g.  rewrite this in doStep:

self contextStackIndex > 1
ifTrue: [self resetContext: newContext]
ifFalse:
[newContext == currentContext
ifTrue: [self changed: #contentsSelection.
self updateInspectors]
ifFalse: [self resetContext: newContext]].

as, say, 

self contextStackIndex > 1
ifTrue: [self resetContext: newContext]
ifFalse:
[newContext == currentContext
ifTrue: [self scheduleUIUpdate]
ifFalse: [self resetContext: newContext]].

add an inst var to hold the last deferred action, and then do something like

Debugger>>scheduleUIUpdate
    (lastScheduledUpdate notNil
    and: [WorldState lastDeferredUIMessage == lastScheduledUpdate]) ifTrue:
        [^self].
    lastScheduledUpdate := [self changed: #contentsSelection. self updateInspectors].
    WorldState addDeferredUIMessage: lastScheduledUpdate

For this we have to add the lastDeferredUIMessage accessor to all the relevant places, but it seems a nice pattern to me.

And of course, there is no need to refresh lastScheduledUpdate.  It could be 

Debugger>>scheduleUIUpdate
    lastScheduledUpdate ifNil: [lastScheduledUpdate := [self changed: #contentsSelection. self updateInspectors]].
    WorldState lastDeferredUIMessage ~~ lastScheduledUpdate ifTrue:
        [WorldState addDeferredUIMessage: lastScheduledUpdate]

What do you think? :-)

What do you like?
Best,
Christoph

Von: Squeak-dev <[hidden email]> im Auftrag von Eliot Miranda <[hidden email]>
Gesendet: Samstag, 19. September 2020 20:17:12
An: The general-purpose Squeak developers list; Taeumel, Marcel
Betreff: [squeak-dev] Can we make computing the local variables in the debugger lazy?
 
Hi Marcel,

    can we try and reduce the frequency at which we compute the variables in the context inspector in the debugger?  It is a noticeable performance hit.  I really like the user interface, but the performance hit is making debugging difficult.

As an example use this:

| samples sineTable sound |
"1 second of A below middle C (220Hz). 16000 / 220 is 72.72 recurring"
sineTable := SoundPlayer sineTable: 73.
sineTable doWithIndex: "And let's not deafen anyone..."
[:sample :index| sineTable at: index put: sample // 4].
samples := SoundBuffer new: 16000.
1 to: samples size by: sineTable size do:
[:i| samples replaceFrom: i to: (i + sineTable size - 1 min: 16000) with: sineTable startingAt: 1].
1 to: 146 do:
[:i|
samples at: i put: ((samples at: i) * i / 146) asInteger.
samples at: 16001 - i put: ((samples at: 16001 - i) * i / 146) asInteger].
sound := SampledSound
samples: samples
samplingRate: 16000.
sound := MixedSound new
add: sound pan: 0.25;
add: sound pan: 0.75;
yourself.
sound computeSamplesForSeconds: sound duration


Open a workspace in e.g. a 64-bit image prior to Tools-mt.965 (I used an image containing Tools-mt.942).  Debug the above in the workspace.  Position the cursor at "sound computeSamplesForSeconds: sound duration" and do "run to here".  It is essentially instantaneous.

Now do the same in a contemporary trunk image.  It takes almost 6 seconds.

I used to be able to click step as fast as I could and the system would keep up with me.  Now I find that if I click too fast I can accumulate excess clicks and when I stp clicking the system will continue stepping and go too far.

I don't want to lose the feedback the new variables list gives us.  But I do find the performance hit tedious.  I wonder could we cache the list and only update it
- when Morphic updates, and
- when the context changes?


_,,,^..^,,,_
best, Eliot



--
_,,,^..^,,,_
best, Eliot



--
_,,,^..^,,,_
best, Eliot


Reply | Threaded
Open this post in threaded view
|

Re: Can we make computing the local variables in the debugger lazy?

Levente Uzonyi
In reply to this post by Christoph Thiede
Hi Christoph,

On Sat, 19 Sep 2020, Thiede, Christoph wrote:

>
> Hi Eliot,
>
>
> very nice finding once again! I #timeProfile'd the menu button action and as I expected, the most expensive operation is the shout styling by the new inspectors, including the decompilation of every method:

What was it exactly that you profiled?

The screenshot shows that 76.9% was spent in #initializeVariablesFromContext,
of which 52.5% of the time was spent in CompiledMethod(CompiledCode) >>
#getSource. The rest of the tree is not visible, but these methods have
nothing to do with parsing or styling, they initialize the parser and
normally should take <1 ms.

Also, why is the decompiler producing the source code?

>
>
> [IMAGE]
>
>
> First, when loading ShoutCore-ct.78 (Inbox), the speed doubles (but probably that's rather a problem with my sources file, see the thread about this commit).

You may want to try compiling a VM where FilePlugin's CreateFile does not
set the FILE_FLAG_SEQUENTIAL_SCAN flag, and see if it helps with file
reading performance.


Levente

>
> Second, we do not redraw the world while running to the selection, so we do not need to update the inspectors at all. I think we could split up #doStep into some #basicDoStep (which would perform the actual stepping logic) +
> some #updateContextDuring: (which would update the stack list and the inspectors), then we would need to trigger the updates only once from #runToSelection:.
> As an alternative, we could make this method a bit more complex but responsive by applying the same updating logic from #runUntil.
>
> What do you think? :-)
>
> Best,
> Christoph
>
> _________________________________________________________________________________________________________________________________________________________________________________________________________________________________
> Von: Squeak-dev <[hidden email]> im Auftrag von Eliot Miranda <[hidden email]>
> Gesendet: Samstag, 19. September 2020 20:17:12
> An: The general-purpose Squeak developers list; Taeumel, Marcel
> Betreff: [squeak-dev] Can we make computing the local variables in the debugger lazy?  
> Hi Marcel,
>
>     can we try and reduce the frequency at which we compute the variables in the context inspector in the debugger?  It is a noticeable performance hit.  I really like the user interface, but the performance hit is making
> debugging difficult.
>
> As an example use this:
>
> | samples sineTable sound |
> "1 second of A below middle C (220Hz). 16000 / 220 is 72.72 recurring"
> sineTable := SoundPlayer sineTable: 73.
> sineTable doWithIndex: "And let's not deafen anyone..."
> [:sample :index| sineTable at: index put: sample // 4].
> samples := SoundBuffer new: 16000.
> 1 to: samples size by: sineTable size do:
> [:i| samples replaceFrom: i to: (i + sineTable size - 1 min: 16000) with: sineTable startingAt: 1].
> 1 to: 146 do:
> [:i|
> samples at: i put: ((samples at: i) * i / 146) asInteger.
> samples at: 16001 - i put: ((samples at: 16001 - i) * i / 146) asInteger].
> sound := SampledSound
> samples: samples
> samplingRate: 16000.
> sound := MixedSound new
> add: sound pan: 0.25;
> add: sound pan: 0.75;
> yourself.
> sound computeSamplesForSeconds: sound duration
>
>
> Open a workspace in e.g. a 64-bit image prior to Tools-mt.965 (I used an image containing Tools-mt.942).  Debug the above in the workspace.  Position the cursor at "sound computeSamplesForSeconds: sound duration" and do "run
> to here".  It is essentially instantaneous.
>
> Now do the same in a contemporary trunk image.  It takes almost 6 seconds.
>
> I used to be able to click step as fast as I could and the system would keep up with me.  Now I find that if I click too fast I can accumulate excess clicks and when I stp clicking the system will continue stepping and go too
> far.
>
> I don't want to lose the feedback the new variables list gives us.  But I do find the performance hit tedious.  I wonder could we cache the list and only update it
> - when Morphic updates, and
> - when the context changes?
>
>
> _,,,^..^,,,_
> best, Eliot
>
>

Reply | Threaded
Open this post in threaded view
|

Re: Can we make computing the local variables in the debugger lazy?

marcel.taeumel
Hi Eliot, hi all!

I fixed the issue with Tools-mt.989. The logic was already there in #runUntil.

Best,
Marcel

Am 19.09.2020 23:00:33 schrieb Levente Uzonyi <[hidden email]>:

Hi Christoph,

On Sat, 19 Sep 2020, Thiede, Christoph wrote:

>
> Hi Eliot,
>
>
> very nice finding once again! I #timeProfile'd the menu button action and as I expected, the most expensive operation is the shout styling by the new inspectors, including the decompilation of every method:

What was it exactly that you profiled?

The screenshot shows that 76.9% was spent in #initializeVariablesFromContext,
of which 52.5% of the time was spent in CompiledMethod(CompiledCode) >>
#getSource. The rest of the tree is not visible, but these methods have
nothing to do with parsing or styling, they initialize the parser and
normally should take <1 ms.

Also, why is the decompiler producing the source code?

>
>
> [IMAGE]
>
>
> First, when loading ShoutCore-ct.78 (Inbox), the speed doubles (but probably that's rather a problem with my sources file, see the thread about this commit).

You may want to try compiling a VM where FilePlugin's CreateFile does not
set the FILE_FLAG_SEQUENTIAL_SCAN flag, and see if it helps with file
reading performance.


Levente

>
> Second, we do not redraw the world while running to the selection, so we do not need to update the inspectors at all. I think we could split up #doStep into some #basicDoStep (which would perform the actual stepping logic) +
> some #updateContextDuring: (which would update the stack list and the inspectors), then we would need to trigger the updates only once from #runToSelection:.
> As an alternative, we could make this method a bit more complex but responsive by applying the same updating logic from #runUntil.
>
> What do you think? :-)
>
> Best,
> Christoph
>
> _________________________________________________________________________________________________________________________________________________________________________________________________________________________________
> Von: Squeak-dev im Auftrag von Eliot Miranda
> Gesendet: Samstag, 19. September 2020 20:17:12
> An: The general-purpose Squeak developers list; Taeumel, Marcel
> Betreff: [squeak-dev] Can we make computing the local variables in the debugger lazy?  
> Hi Marcel,
>
>     can we try and reduce the frequency at which we compute the variables in the context inspector in the debugger?  It is a noticeable performance hit.  I really like the user interface, but the performance hit is making
> debugging difficult.
>
> As an example use this:
>
> | samples sineTable sound |
> "1 second of A below middle C (220Hz). 16000 / 220 is 72.72 recurring"
> sineTable := SoundPlayer sineTable: 73.
> sineTable doWithIndex: "And let's not deafen anyone..."
> [:sample :index| sineTable at: index put: sample // 4].
> samples := SoundBuffer new: 16000.
> 1 to: samples size by: sineTable size do:
> [:i| samples replaceFrom: i to: (i + sineTable size - 1 min: 16000) with: sineTable startingAt: 1].
> 1 to: 146 do:
> [:i|
> samples at: i put: ((samples at: i) * i / 146) asInteger.
> samples at: 16001 - i put: ((samples at: 16001 - i) * i / 146) asInteger].
> sound := SampledSound
> samples: samples
> samplingRate: 16000.
> sound := MixedSound new
> add: sound pan: 0.25;
> add: sound pan: 0.75;
> yourself.
> sound computeSamplesForSeconds: sound duration
>
>
> Open a workspace in e.g. a 64-bit image prior to Tools-mt.965 (I used an image containing Tools-mt.942).  Debug the above in the workspace.  Position the cursor at "sound computeSamplesForSeconds: sound duration" and do "run
> to here".  It is essentially instantaneous.
>
> Now do the same in a contemporary trunk image.  It takes almost 6 seconds.
>
> I used to be able to click step as fast as I could and the system would keep up with me.  Now I find that if I click too fast I can accumulate excess clicks and when I stp clicking the system will continue stepping and go too
> far.
>
> I don't want to lose the feedback the new variables list gives us.  But I do find the performance hit tedious.  I wonder could we cache the list and only update it
> - when Morphic updates, and
> - when the context changes?
>
>
> _,,,^..^,,,_
> best, Eliot
>
>

Reply | Threaded
Open this post in threaded view
|

Re: Can we make computing the local variables in the debugger lazy?

Eliot Miranda-2


On Sep 29, 2020, at 1:07 AM, Marcel Taeumel <[hidden email]> wrote:


Hi Eliot, hi all!

I fixed the issue with Tools-mt.989. The logic was already there in #runUntil.

<3 thank you, thank you, thank you!!


Best,
Marcel

Am 19.09.2020 23:00:33 schrieb Levente Uzonyi <[hidden email]>:

Hi Christoph,

On Sat, 19 Sep 2020, Thiede, Christoph wrote:

>
> Hi Eliot,
>
>
> very nice finding once again! I #timeProfile'd the menu button action and as I expected, the most expensive operation is the shout styling by the new inspectors, including the decompilation of every method:

What was it exactly that you profiled?

The screenshot shows that 76.9% was spent in #initializeVariablesFromContext,
of which 52.5% of the time was spent in CompiledMethod(CompiledCode) >>
#getSource. The rest of the tree is not visible, but these methods have
nothing to do with parsing or styling, they initialize the parser and
normally should take <1 ms.

Also, why is the decompiler producing the source code?

>
>
> [IMAGE]
>
>
> First, when loading ShoutCore-ct.78 (Inbox), the speed doubles (but probably that's rather a problem with my sources file, see the thread about this commit).

You may want to try compiling a VM where FilePlugin's CreateFile does not
set the FILE_FLAG_SEQUENTIAL_SCAN flag, and see if it helps with file
reading performance.


Levente

>
> Second, we do not redraw the world while running to the selection, so we do not need to update the inspectors at all. I think we could split up #doStep into some #basicDoStep (which would perform the actual stepping logic) +
> some #updateContextDuring: (which would update the stack list and the inspectors), then we would need to trigger the updates only once from #runToSelection:.
> As an alternative, we could make this method a bit more complex but responsive by applying the same updating logic from #runUntil.
>
> What do you think? :-)
>
> Best,
> Christoph
>
> _________________________________________________________________________________________________________________________________________________________________________________________________________________________________
> Von: Squeak-dev im Auftrag von Eliot Miranda
> Gesendet: Samstag, 19. September 2020 20:17:12
> An: The general-purpose Squeak developers list; Taeumel, Marcel
> Betreff: [squeak-dev] Can we make computing the local variables in the debugger lazy?  
> Hi Marcel,
>
>     can we try and reduce the frequency at which we compute the variables in the context inspector in the debugger?  It is a noticeable performance hit.  I really like the user interface, but the performance hit is making
> debugging difficult.
>
> As an example use this:
>
> | samples sineTable sound |
> "1 second of A below middle C (220Hz). 16000 / 220 is 72.72 recurring"
> sineTable := SoundPlayer sineTable: 73.
> sineTable doWithIndex: "And let's not deafen anyone..."
> [:sample :index| sineTable at: index put: sample // 4].
> samples := SoundBuffer new: 16000.
> 1 to: samples size by: sineTable size do:
> [:i| samples replaceFrom: i to: (i + sineTable size - 1 min: 16000) with: sineTable startingAt: 1].
> 1 to: 146 do:
> [:i|
> samples at: i put: ((samples at: i) * i / 146) asInteger.
> samples at: 16001 - i put: ((samples at: 16001 - i) * i / 146) asInteger].
> sound := SampledSound
> samples: samples
> samplingRate: 16000.
> sound := MixedSound new
> add: sound pan: 0.25;
> add: sound pan: 0.75;
> yourself.
> sound computeSamplesForSeconds: sound duration
>
>
> Open a workspace in e.g. a 64-bit image prior to Tools-mt.965 (I used an image containing Tools-mt.942).  Debug the above in the workspace.  Position the cursor at "sound computeSamplesForSeconds: sound duration" and do "run
> to here".  It is essentially instantaneous.
>
> Now do the same in a contemporary trunk image.  It takes almost 6 seconds.
>
> I used to be able to click step as fast as I could and the system would keep up with me.  Now I find that if I click too fast I can accumulate excess clicks and when I stp clicking the system will continue stepping and go too
> far.
>
> I don't want to lose the feedback the new variables list gives us.  But I do find the performance hit tedious.  I wonder could we cache the list and only update it
> - when Morphic updates, and
> - when the context changes?
>
>
> _,,,^..^,,,_
> best, Eliot
>
>


Reply | Threaded
Open this post in threaded view
|

Re: Can we make computing the local variables in the debugger lazy?

Eliot Miranda-2
In reply to this post by marcel.taeumel
Hi Marcel,

On Tue, Sep 29, 2020 at 1:07 AM Marcel Taeumel <[hidden email]> wrote:
Hi Eliot, hi all!

I fixed the issue with Tools-mt.989. The logic was already there in #runUntil.

Can you point me to where in ContextInspector the emphasis for the temp vars is chosen?  I want to modify ContextInspector to cache the temp names for a given method because I believe this will speed up stepping a lot.  However I couldn't find where the emphasis is applied so I'm worried that my cache may break something.  If I can see where that is being done I have a better chance at avoiding breaking things.


Best,
Marcel

Am 19.09.2020 23:00:33 schrieb Levente Uzonyi <[hidden email]>:

Hi Christoph,

On Sat, 19 Sep 2020, Thiede, Christoph wrote:

>
> Hi Eliot,
>
>
> very nice finding once again! I #timeProfile'd the menu button action and as I expected, the most expensive operation is the shout styling by the new inspectors, including the decompilation of every method:

What was it exactly that you profiled?

The screenshot shows that 76.9% was spent in #initializeVariablesFromContext,
of which 52.5% of the time was spent in CompiledMethod(CompiledCode) >>
#getSource. The rest of the tree is not visible, but these methods have
nothing to do with parsing or styling, they initialize the parser and
normally should take <1 ms.

Also, why is the decompiler producing the source code?

>
>
> [IMAGE]
>
>
> First, when loading ShoutCore-ct.78 (Inbox), the speed doubles (but probably that's rather a problem with my sources file, see the thread about this commit).

You may want to try compiling a VM where FilePlugin's CreateFile does not
set the FILE_FLAG_SEQUENTIAL_SCAN flag, and see if it helps with file
reading performance.


Levente

>
> Second, we do not redraw the world while running to the selection, so we do not need to update the inspectors at all. I think we could split up #doStep into some #basicDoStep (which would perform the actual stepping logic) +
> some #updateContextDuring: (which would update the stack list and the inspectors), then we would need to trigger the updates only once from #runToSelection:.
> As an alternative, we could make this method a bit more complex but responsive by applying the same updating logic from #runUntil.
>
> What do you think? :-)
>
> Best,
> Christoph
>
> _________________________________________________________________________________________________________________________________________________________________________________________________________________________________
> Von: Squeak-dev im Auftrag von Eliot Miranda
> Gesendet: Samstag, 19. September 2020 20:17:12
> An: The general-purpose Squeak developers list; Taeumel, Marcel
> Betreff: [squeak-dev] Can we make computing the local variables in the debugger lazy?  
> Hi Marcel,
>
>     can we try and reduce the frequency at which we compute the variables in the context inspector in the debugger?  It is a noticeable performance hit.  I really like the user interface, but the performance hit is making
> debugging difficult.
>
> As an example use this:
>
> | samples sineTable sound |
> "1 second of A below middle C (220Hz). 16000 / 220 is 72.72 recurring"
> sineTable := SoundPlayer sineTable: 73.
> sineTable doWithIndex: "And let's not deafen anyone..."
> [:sample :index| sineTable at: index put: sample // 4].
> samples := SoundBuffer new: 16000.
> 1 to: samples size by: sineTable size do:
> [:i| samples replaceFrom: i to: (i + sineTable size - 1 min: 16000) with: sineTable startingAt: 1].
> 1 to: 146 do:
> [:i|
> samples at: i put: ((samples at: i) * i / 146) asInteger.
> samples at: 16001 - i put: ((samples at: 16001 - i) * i / 146) asInteger].
> sound := SampledSound
> samples: samples
> samplingRate: 16000.
> sound := MixedSound new
> add: sound pan: 0.25;
> add: sound pan: 0.75;
> yourself.
> sound computeSamplesForSeconds: sound duration
>
>
> Open a workspace in e.g. a 64-bit image prior to Tools-mt.965 (I used an image containing Tools-mt.942).  Debug the above in the workspace.  Position the cursor at "sound computeSamplesForSeconds: sound duration" and do "run
> to here".  It is essentially instantaneous.
>
> Now do the same in a contemporary trunk image.  It takes almost 6 seconds.
>
> I used to be able to click step as fast as I could and the system would keep up with me.  Now I find that if I click too fast I can accumulate excess clicks and when I stp clicking the system will continue stepping and go too
> far.
>
> I don't want to lose the feedback the new variables list gives us.  But I do find the performance hit tedious.  I wonder could we cache the list and only update it
> - when Morphic updates, and
> - when the context changes?
>
>
> _,,,^..^,,,_
> best, Eliot
>
>


--
_,,,^..^,,,_
best, Eliot


Reply | Threaded
Open this post in threaded view
|

Re: Can we make computing the local variables in the debugger lazy?

Christoph Thiede

Hi Eliot,


Can you point me to where in ContextInspector the emphasis for the temp vars is chosen?


By convention, inspector fields are defined in the category "fields - streaming". So you are probably searching for ContextInspector >> #streamTemporaryVariablesOn: or ContextVariablesInspector >> #streamTemporaryVariablesOn:?

Looking forward to your optimization! :-)


Best,

Christoph


Von: Squeak-dev <[hidden email]> im Auftrag von Eliot Miranda <[hidden email]>
Gesendet: Mittwoch, 30. September 2020 09:09:27
An: The general-purpose Squeak developers list
Betreff: Re: [squeak-dev] Can we make computing the local variables in the debugger lazy?
 
Hi Marcel,

On Tue, Sep 29, 2020 at 1:07 AM Marcel Taeumel <[hidden email]> wrote:
Hi Eliot, hi all!

I fixed the issue with Tools-mt.989. The logic was already there in #runUntil.

Can you point me to where in ContextInspector the emphasis for the temp vars is chosen?  I want to modify ContextInspector to cache the temp names for a given method because I believe this will speed up stepping a lot.  However I couldn't find where the emphasis is applied so I'm worried that my cache may break something.  If I can see where that is being done I have a better chance at avoiding breaking things.


Best,
Marcel

Am 19.09.2020 23:00:33 schrieb Levente Uzonyi <[hidden email]>:

Hi Christoph,

On Sat, 19 Sep 2020, Thiede, Christoph wrote:

>
> Hi Eliot,
>
>
> very nice finding once again! I #timeProfile'd the menu button action and as I expected, the most expensive operation is the shout styling by the new inspectors, including the decompilation of every method:

What was it exactly that you profiled?

The screenshot shows that 76.9% was spent in #initializeVariablesFromContext,
of which 52.5% of the time was spent in CompiledMethod(CompiledCode) >>
#getSource. The rest of the tree is not visible, but these methods have
nothing to do with parsing or styling, they initialize the parser and
normally should take <1 ms.

Also, why is the decompiler producing the source code?

>
>
> [IMAGE]
>
>
> First, when loading ShoutCore-ct.78 (Inbox), the speed doubles (but probably that's rather a problem with my sources file, see the thread about this commit).

You may want to try compiling a VM where FilePlugin's CreateFile does not
set the FILE_FLAG_SEQUENTIAL_SCAN flag, and see if it helps with file
reading performance.


Levente

>
> Second, we do not redraw the world while running to the selection, so we do not need to update the inspectors at all. I think we could split up #doStep into some #basicDoStep (which would perform the actual stepping logic) +
> some #updateContextDuring: (which would update the stack list and the inspectors), then we would need to trigger the updates only once from #runToSelection:.
> As an alternative, we could make this method a bit more complex but responsive by applying the same updating logic from #runUntil.
>
> What do you think? :-)
>
> Best,
> Christoph
>
> _________________________________________________________________________________________________________________________________________________________________________________________________________________________________
> Von: Squeak-dev im Auftrag von Eliot Miranda
> Gesendet: Samstag, 19. September 2020 20:17:12
> An: The general-purpose Squeak developers list; Taeumel, Marcel
> Betreff: [squeak-dev] Can we make computing the local variables in the debugger lazy?  
> Hi Marcel,
>
>     can we try and reduce the frequency at which we compute the variables in the context inspector in the debugger?  It is a noticeable performance hit.  I really like the user interface, but the performance hit is making
> debugging difficult.
>
> As an example use this:
>
> | samples sineTable sound |
> "1 second of A below middle C (220Hz). 16000 / 220 is 72.72 recurring"
> sineTable := SoundPlayer sineTable: 73.
> sineTable doWithIndex: "And let's not deafen anyone..."
> [:sample :index| sineTable at: index put: sample // 4].
> samples := SoundBuffer new: 16000.
> 1 to: samples size by: sineTable size do:
> [:i| samples replaceFrom: i to: (i + sineTable size - 1 min: 16000) with: sineTable startingAt: 1].
> 1 to: 146 do:
> [:i|
> samples at: i put: ((samples at: i) * i / 146) asInteger.
> samples at: 16001 - i put: ((samples at: 16001 - i) * i / 146) asInteger].
> sound := SampledSound
> samples: samples
> samplingRate: 16000.
> sound := MixedSound new
> add: sound pan: 0.25;
> add: sound pan: 0.75;
> yourself.
> sound computeSamplesForSeconds: sound duration
>
>
> Open a workspace in e.g. a 64-bit image prior to Tools-mt.965 (I used an image containing Tools-mt.942).  Debug the above in the workspace.  Position the cursor at "sound computeSamplesForSeconds: sound duration" and do "run
> to here".  It is essentially instantaneous.
>
> Now do the same in a contemporary trunk image.  It takes almost 6 seconds.
>
> I used to be able to click step as fast as I could and the system would keep up with me.  Now I find that if I click too fast I can accumulate excess clicks and when I stp clicking the system will continue stepping and go too
> far.
>
> I don't want to lose the feedback the new variables list gives us.  But I do find the performance hit tedious.  I wonder could we cache the list and only update it
> - when Morphic updates, and
> - when the context changes?
>
>
> _,,,^..^,,,_
> best, Eliot
>
>



--
_,,,^..^,,,_
best, Eliot


Reply | Threaded
Open this post in threaded view
|

Re: Can we make computing the local variables in the debugger lazy?

Eliot Miranda-2
Hi Christoph,

On Wed, Sep 30, 2020 at 2:27 AM Thiede, Christoph <[hidden email]> wrote:

Hi Eliot,


Can you point me to where in ContextInspector the emphasis for the temp vars is chosen?


By convention, inspector fields are defined in the category "fields - streaming". So you are probably searching for ContextInspector >> #streamTemporaryVariablesOn: or ContextVariablesInspector >> #streamTemporaryVariablesOn:?


I found this code; I don't see any text emphasis going on.  Where is the code that turns temp variables red when they're out of scope?  Marcel?

Looking forward to your optimization! :-)


Best,

Christoph


Von: Squeak-dev <[hidden email]> im Auftrag von Eliot Miranda <[hidden email]>
Gesendet: Mittwoch, 30. September 2020 09:09:27
An: The general-purpose Squeak developers list
Betreff: Re: [squeak-dev] Can we make computing the local variables in the debugger lazy?
 
Hi Marcel,

On Tue, Sep 29, 2020 at 1:07 AM Marcel Taeumel <[hidden email]> wrote:
Hi Eliot, hi all!

I fixed the issue with Tools-mt.989. The logic was already there in #runUntil.

Can you point me to where in ContextInspector the emphasis for the temp vars is chosen?  I want to modify ContextInspector to cache the temp names for a given method because I believe this will speed up stepping a lot.  However I couldn't find where the emphasis is applied so I'm worried that my cache may break something.  If I can see where that is being done I have a better chance at avoiding breaking things.


Best,
Marcel

Am 19.09.2020 23:00:33 schrieb Levente Uzonyi <[hidden email]>:

Hi Christoph,

On Sat, 19 Sep 2020, Thiede, Christoph wrote:

>
> Hi Eliot,
>
>
> very nice finding once again! I #timeProfile'd the menu button action and as I expected, the most expensive operation is the shout styling by the new inspectors, including the decompilation of every method:

What was it exactly that you profiled?

The screenshot shows that 76.9% was spent in #initializeVariablesFromContext,
of which 52.5% of the time was spent in CompiledMethod(CompiledCode) >>
#getSource. The rest of the tree is not visible, but these methods have
nothing to do with parsing or styling, they initialize the parser and
normally should take <1 ms.

Also, why is the decompiler producing the source code?

>
>
> [IMAGE]
>
>
> First, when loading ShoutCore-ct.78 (Inbox), the speed doubles (but probably that's rather a problem with my sources file, see the thread about this commit).

You may want to try compiling a VM where FilePlugin's CreateFile does not
set the FILE_FLAG_SEQUENTIAL_SCAN flag, and see if it helps with file
reading performance.


Levente

>
> Second, we do not redraw the world while running to the selection, so we do not need to update the inspectors at all. I think we could split up #doStep into some #basicDoStep (which would perform the actual stepping logic) +
> some #updateContextDuring: (which would update the stack list and the inspectors), then we would need to trigger the updates only once from #runToSelection:.
> As an alternative, we could make this method a bit more complex but responsive by applying the same updating logic from #runUntil.
>
> What do you think? :-)
>
> Best,
> Christoph
>
> _________________________________________________________________________________________________________________________________________________________________________________________________________________________________
> Von: Squeak-dev im Auftrag von Eliot Miranda
> Gesendet: Samstag, 19. September 2020 20:17:12
> An: The general-purpose Squeak developers list; Taeumel, Marcel
> Betreff: [squeak-dev] Can we make computing the local variables in the debugger lazy?  
> Hi Marcel,
>
>     can we try and reduce the frequency at which we compute the variables in the context inspector in the debugger?  It is a noticeable performance hit.  I really like the user interface, but the performance hit is making
> debugging difficult.
>
> As an example use this:
>
> | samples sineTable sound |
> "1 second of A below middle C (220Hz). 16000 / 220 is 72.72 recurring"
> sineTable := SoundPlayer sineTable: 73.
> sineTable doWithIndex: "And let's not deafen anyone..."
> [:sample :index| sineTable at: index put: sample // 4].
> samples := SoundBuffer new: 16000.
> 1 to: samples size by: sineTable size do:
> [:i| samples replaceFrom: i to: (i + sineTable size - 1 min: 16000) with: sineTable startingAt: 1].
> 1 to: 146 do:
> [:i|
> samples at: i put: ((samples at: i) * i / 146) asInteger.
> samples at: 16001 - i put: ((samples at: 16001 - i) * i / 146) asInteger].
> sound := SampledSound
> samples: samples
> samplingRate: 16000.
> sound := MixedSound new
> add: sound pan: 0.25;
> add: sound pan: 0.75;
> yourself.
> sound computeSamplesForSeconds: sound duration
>
>
> Open a workspace in e.g. a 64-bit image prior to Tools-mt.965 (I used an image containing Tools-mt.942).  Debug the above in the workspace.  Position the cursor at "sound computeSamplesForSeconds: sound duration" and do "run
> to here".  It is essentially instantaneous.
>
> Now do the same in a contemporary trunk image.  It takes almost 6 seconds.
>
> I used to be able to click step as fast as I could and the system would keep up with me.  Now I find that if I click too fast I can accumulate excess clicks and when I stp clicking the system will continue stepping and go too
> far.
>
> I don't want to lose the feedback the new variables list gives us.  But I do find the performance hit tedious.  I wonder could we cache the list and only update it
> - when Morphic updates, and
> - when the context changes?
>
>
> _,,,^..^,,,_
> best, Eliot
>
>



--
_,,,^..^,,,_
best, Eliot



--
_,,,^..^,,,_
best, Eliot


Reply | Threaded
Open this post in threaded view
|

Re: Can we make computing the local variables in the debugger lazy?

Christoph Thiede

Hi Eliot,


you can have an inspector field's title be styled by sending "shouldStyleName: true" to it. The styling itself, then, is performed by the Inspector in #fieldList (or just browse the senders of #shouldStyleName), where the styler is invoked. Before styling the field list, the styler is configured in #aboutToStyle:requestor:/#updateStyler:requestor: where self doItContext is passed to the styler.


So tl;dr: Styling is always applied late (after the field list is constructed), you cannot break it by caching the temp names earlier. :-)


Best,

Christoph


Von: Squeak-dev <[hidden email]> im Auftrag von Eliot Miranda <[hidden email]>
Gesendet: Donnerstag, 1. Oktober 2020 20:33:08
An: The general-purpose Squeak developers list; Taeumel, Marcel
Betreff: Re: [squeak-dev] Can we make computing the local variables in the debugger lazy?
 
Hi Christoph,

On Wed, Sep 30, 2020 at 2:27 AM Thiede, Christoph <[hidden email]> wrote:

Hi Eliot,


Can you point me to where in ContextInspector the emphasis for the temp vars is chosen?


By convention, inspector fields are defined in the category "fields - streaming". So you are probably searching for ContextInspector >> #streamTemporaryVariablesOn: or ContextVariablesInspector >> #streamTemporaryVariablesOn:?


I found this code; I don't see any text emphasis going on.  Where is the code that turns temp variables red when they're out of scope?  Marcel?

Looking forward to your optimization! :-)


Best,

Christoph


Von: Squeak-dev <[hidden email]> im Auftrag von Eliot Miranda <[hidden email]>
Gesendet: Mittwoch, 30. September 2020 09:09:27
An: The general-purpose Squeak developers list
Betreff: Re: [squeak-dev] Can we make computing the local variables in the debugger lazy?
 
Hi Marcel,

On Tue, Sep 29, 2020 at 1:07 AM Marcel Taeumel <[hidden email]> wrote:
Hi Eliot, hi all!

I fixed the issue with Tools-mt.989. The logic was already there in #runUntil.

Can you point me to where in ContextInspector the emphasis for the temp vars is chosen?  I want to modify ContextInspector to cache the temp names for a given method because I believe this will speed up stepping a lot.  However I couldn't find where the emphasis is applied so I'm worried that my cache may break something.  If I can see where that is being done I have a better chance at avoiding breaking things.


Best,
Marcel

Am 19.09.2020 23:00:33 schrieb Levente Uzonyi <[hidden email]>:

Hi Christoph,

On Sat, 19 Sep 2020, Thiede, Christoph wrote:

>
> Hi Eliot,
>
>
> very nice finding once again! I #timeProfile'd the menu button action and as I expected, the most expensive operation is the shout styling by the new inspectors, including the decompilation of every method:

What was it exactly that you profiled?

The screenshot shows that 76.9% was spent in #initializeVariablesFromContext,
of which 52.5% of the time was spent in CompiledMethod(CompiledCode) >>
#getSource. The rest of the tree is not visible, but these methods have
nothing to do with parsing or styling, they initialize the parser and
normally should take <1 ms.

Also, why is the decompiler producing the source code?

>
>
> [IMAGE]
>
>
> First, when loading ShoutCore-ct.78 (Inbox), the speed doubles (but probably that's rather a problem with my sources file, see the thread about this commit).

You may want to try compiling a VM where FilePlugin's CreateFile does not
set the FILE_FLAG_SEQUENTIAL_SCAN flag, and see if it helps with file
reading performance.


Levente

>
> Second, we do not redraw the world while running to the selection, so we do not need to update the inspectors at all. I think we could split up #doStep into some #basicDoStep (which would perform the actual stepping logic) +
> some #updateContextDuring: (which would update the stack list and the inspectors), then we would need to trigger the updates only once from #runToSelection:.
> As an alternative, we could make this method a bit more complex but responsive by applying the same updating logic from #runUntil.
>
> What do you think? :-)
>
> Best,
> Christoph
>
> _________________________________________________________________________________________________________________________________________________________________________________________________________________________________
> Von: Squeak-dev im Auftrag von Eliot Miranda
> Gesendet: Samstag, 19. September 2020 20:17:12
> An: The general-purpose Squeak developers list; Taeumel, Marcel
> Betreff: [squeak-dev] Can we make computing the local variables in the debugger lazy?  
> Hi Marcel,
>
>     can we try and reduce the frequency at which we compute the variables in the context inspector in the debugger?  It is a noticeable performance hit.  I really like the user interface, but the performance hit is making
> debugging difficult.
>
> As an example use this:
>
> | samples sineTable sound |
> "1 second of A below middle C (220Hz). 16000 / 220 is 72.72 recurring"
> sineTable := SoundPlayer sineTable: 73.
> sineTable doWithIndex: "And let's not deafen anyone..."
> [:sample :index| sineTable at: index put: sample // 4].
> samples := SoundBuffer new: 16000.
> 1 to: samples size by: sineTable size do:
> [:i| samples replaceFrom: i to: (i + sineTable size - 1 min: 16000) with: sineTable startingAt: 1].
> 1 to: 146 do:
> [:i|
> samples at: i put: ((samples at: i) * i / 146) asInteger.
> samples at: 16001 - i put: ((samples at: 16001 - i) * i / 146) asInteger].
> sound := SampledSound
> samples: samples
> samplingRate: 16000.
> sound := MixedSound new
> add: sound pan: 0.25;
> add: sound pan: 0.75;
> yourself.
> sound computeSamplesForSeconds: sound duration
>
>
> Open a workspace in e.g. a 64-bit image prior to Tools-mt.965 (I used an image containing Tools-mt.942).  Debug the above in the workspace.  Position the cursor at "sound computeSamplesForSeconds: sound duration" and do "run
> to here".  It is essentially instantaneous.
>
> Now do the same in a contemporary trunk image.  It takes almost 6 seconds.
>
> I used to be able to click step as fast as I could and the system would keep up with me.  Now I find that if I click too fast I can accumulate excess clicks and when I stp clicking the system will continue stepping and go too
> far.
>
> I don't want to lose the feedback the new variables list gives us.  But I do find the performance hit tedious.  I wonder could we cache the list and only update it
> - when Morphic updates, and
> - when the context changes?
>
>
> _,,,^..^,,,_
> best, Eliot
>
>



--
_,,,^..^,,,_
best, Eliot



--
_,,,^..^,,,_
best, Eliot