Performance

[migueldeicaza]

Currently about 12% of our performance when outputting characters is being consumed by swift_beginAccess when the insertCharacter accesses the buffer properties, and then when the buffer itself accesses

Purely for the insertCharacter call, this accounts for 43.2% and retain/release for another 12%

Started at throughput 94,983, the way I measure this is by using the "Profile" option on the PerformanceTest class, like this:

Branch is 'performance-experiment'

All my measurements are on a 2021 MacBook Pro with an M1 Max.

• Validate that the syncValues in terminal for some variables were only needed when going from alt to normal, and not on the other way around (what the code in this branch does).
• Implement lastBufferStorage for combing characters
• Sync the wrap modes

[migueldeicaza]

Before moving the Terminal code that handles insertCharacter to Buffer itself:

the throughput for a profile build was 121,669

After the changes the throughput is: 138,113

[migueldeicaza]

Caching the array.count in CircularList<> to avoid the __swift_instantiateCanonicalPrespecializedGenericMetadata brings the throughput to 160,690:

https://gist.github.com/migueldeicaza/e6b93fb4b53d433048fcf5392369b534

But there is still significant stuff here, I think I will inline CircularBuffer for each type that we use, the ViewLineInfo and BufferLine, and also inline the mechanism to create the new empty slot rather than having a callback that might be triggering the problem.

[migueldeicaza]

Replacing the CircularList with a dedicated non-generic version (brute force, because I have no clue why we are triggering that code in Swift), throughput goes to 182,602. A small bump by adding a mechanism that prevents the three calls to swift_beginAccess on the bufferLine by providing a callback.

https://gist.github.com/migueldeicaza/03c8709dcd09e681573871296e8d1eee

[migueldeicaza]

[migueldeicaza]

Plain: 184,523
Borrow: 185,215
eliminate additional beginAccess: 188,985

https://gist.github.com/migueldeicaza/7d891a7fa5932986b4142a7d50c0c086

[migueldeicaza]

https://gist.github.com/migueldeicaza/90c6ece5c418f23c822f99ef1bbf5ec0

221,525

[migueldeicaza]

Improving upon the dedicated codepath for ascii to avoid all the slow code paths:

251,516

https://gist.github.com/migueldeicaza/665648e3c0ad10f43c8dec5c18ede078

[migueldeicaza]

Eliminating the BufferSet, and instead inlining its behavior into the Terminal:

https://gist.github.com/migueldeicaza/3c20e2307c42208630702a47f9f88e91

279,591

https://gist.github.com/migueldeicaza/93b8bae1df29e22d9dd4170d0f554677

Replacing the checked operation with an unchecked one:

283,264

[migueldeicaza]

293,496: avoid calling array.count in getCyclicIndex
https://gist.github.com/migueldeicaza/06c495be1798cde1ffe29d53b2279113

But now, my time is dominated by swift_beginAccess, and I do not know how to get rid of this, the previous hack did not seem to be enough (where I used a callback), with the 'with(line:)' thing:

[migueldeicaza]

Wow, making CircularBufferList members private, and avoiding the getter, makes the call to the getter go from 11% to 3% of the runtime, which was previously dominated by swift_beginAccess:

https://gist.github.com/migueldeicaza/661c91800cfc1954dc8a40d577b346a9

The patch above is not ready, as I took a shortcut just to measure (the removal of the GC for atoms), but I go 327,258

[migueldeicaza]

Make the ReadingBuffer private, and make CircularBufferLineList internal and the variable declaration for _lines private, this bumps the performance to 353,599

https://gist.github.com/migueldeicaza/40347c7bde4ce63875b9e8970fd67fc5

[migueldeicaza]

Making a few static fields private in UnicodeUtil, which eliminates the cost of swift_beingAccess in expectedSizeFromFirstByte

That brings us to 368,724

[migueldeicaza]

Making Terminal.curAttr private brings us to 378.192 which avoids the swift_beginAccess, I am starting to get a feeling for what I need to do ;-)

96bc4df