Speedup of the Cappuccino framework by more efficient objj_msgSend

[mrcarlberg]

This pull request will make Objective-J run faster by implementing a more efficient objj_msgSend function and objj_method implementation. The compiler can also optionally inline the function to increase the speed even more.

Initial tests show that the speed increase is more than 50% in special cases. For a general Objective-J application the speed increase is around 20%.

The speed increase is possible by the following improvements:

1. More efficient objj_method implemention. That bad thing here is that the name property on objj_method has to change to method_name.
2. objj_msgSend and objj_msgSendSuper functions has a more efficient implementation.
3. Array and Dictionary Literals generate more efficient code.
4. The objj_msgSend and objj_msgSendSuper functions are optionally inlined by the compiler.

How to use:

• The Release version of the Cappuccino framework will have inlined objj_msgSend and objj_msgSendSuper functions. The Debugversion will use the old obj_msgSend functions.
• To generate inlined code when compiling in the browser add the below line in the index[-debug].html file.

OBJJ_COMPILER_FLAGS = ["IncludeDebugSymbols", "IncludeTypeSignatures", "InlineMsgSend"];

• To generate inline code for the Release version of your project when building from command line change task.setCompilerFlags("-O"); to task.setCompilerFlags("-O2"); in the project Jakefile. This is needed for all your subprojects and frameworks too.

A new version of OJTest framework (https://github.com/cappuccino/OJTest) is needed to run the test cases. Copy it to the package folder in the Narwhal directory.

[cappbot]

Milestone: Someday. Label: #new. What's next? A reviewer should examine this issue.

[aljungberg]

The objj_msgSend and objj_msgSendSuper functions are optionally inlined by the compiler.

How does that handle method swizzling?

[mrcarlberg]

The same code is just inlined so it works fine.
The thing you can't do is use decorators on obj_msgSend.
The compiler will only optionally inline the functions on the release version when building from command line.

[aljungberg]

Ah I get it. Yeah the most common objj_msgSend decorators are all about debugging so enabling this option for release code makes sense.

Shouldn't this PR have some code generation unit tests?

[mrcarlberg]

Yes, it definitively needs some code generation unit tests. I have plans to do the OutputTest in two different versions, one standard and one inlined. But I will add those when I know that the new code works for everyone.

[boucher]

Very cool!

On Tue, Oct 6, 2015 at 6:39 AM Martin Carlberg notifications@github.com
wrote:

Yes, it definitively needs some code generation unit tests. I have plans
to do the OutputTest in two different versions, one standard and one
inlined. But I will add those when I know that the new code works for
everyone.

—
Reply to this email directly or view it on GitHub
#2384 (comment)
.

[primalmotion]

That doesn't work on our built application either in release, release+press or release+press+flatten

The page just hangs. If I make the browser to stop on all exceptions I get an exception, but no message is logged anywhere

[primalmotion]

edit: in release, it doesn't load, but actually logs an exception:

TypeError: undefined is not an object (evaluating 'CPDictionary.isa.method_msgSend')

[mrcarlberg]

Hmm, this is strange. Are you sure it is running against a new Objective-J Runtime?

It feels like you are running on an old Runtime as it gets undefined when evaluating CPDictionary.isa.method_msgSend All objj_classobjects should have the attribute method_msgSend

[primalmotion]

my bad actually :)

I did not realize I had to activate the -O2 flag for all libraries

It's working great, we'll do some testing and report any issue.

Great Job!

[cappbot]

Milestone: Someday. Labels: #accepted, #needs-unit-test, Objective-J. What's next? This issue needs a volunteer to write and submit one or more unit tests execercising the changes and/or the relevant parts of the original problem.

[mrcarlberg]

I have now added code generation unit test cases with inlined objj_msgSend functions for all tests in OutputTest

[cappbot]

Milestone: Someday. Labels: #accepted, Objective-J. What's next? A reviewer should examine this issue.

[cappbot]

Milestone: Someday. Labels: #accepted, Objective-J. What's next? A reviewer should examine this issue.

[mrcarlberg]

I have now fixed a problem with compiler flags when compiling in the browser.

[mrcarlberg]

I don't have any more outstanding issues. We are now using it in our project and are happy with it.

[aljungberg]

Looks good so far. Some minor points we should be aware of:

• larger build output since each message send is more code
• _objj_forward becomes a keyword
• uses (very slightly) more memory to add method_msgSend to the isas after class initialisation

Did you compare the size of your build output for your app with and without this? What about memory usage in the browser? If it's 20-50% faster I imagine it can't be too significant.

[mrcarlberg]

The code size does increase:
Foundationgoes from 987,634 bytes to 1,151,583 bytes. (+17%)
AppKitgoes from 3,587,337 bytes to 4,343,767 bytes. (+21%)
This is not compressed and not zipped.

I can't see any difference in memory usage in the Browser but as you say it should be a small increase.

[zittix]

This generates a Uncaught ReferenceError: OBJJ_COMPILER_FLAGS is not defined on Chrome.

typeof OBJJ_COMPILER_FLAGS !== 'undefined'

would be better.

[mrcarlberg]

Thanks for finding this problem. I'll make a fix for this.

[Dogild]

Are we good with this one @mrcarlberg ?

[cappbot]

Assignee: mrcarlberg. Milestone: Someday. Labels: #works-for-me, Objective-J. What's next? Attempts to reproduce the problem described by this issue have failed to reveal any erroneous situation.