Use QFontMetrics::elidedText over manual implementation

[Nerixyz]

Description

There is a major performance issue when displaying replies to a message with certain Arabic characters (measuring the width for them seems to be a slow path). This PR attempts to resolve parts of that issue.

Reproduction

If you're running a debug build, do the following:

1. Open !!! as a new split
2. Run the following command a few times:

   /fakemsg @reply-parent-user-login=medicwho;id=162f1550-4686-40a9-bec1-66aa5cf4fcc0;subscriber=1;reply-parent-msg-id=72733dad-a857-408d-ab16-e96834aadf73;first-msg=0;reply-parent-display-name=MedicWho;reply-thread-parent-msg-id=72733dad-a857-408d-ab16-e96834aadf73;badges=subscriber/24,no_audio/1;reply-thread-parent-user-login=medicwho;room-id=11148817;user-id=117691339;turbo=0;color=#DAA520;reply-parent-user-id=71651832;display-name=Mm2PL;badge-info=subscriber/27;reply-parent-msg-body=ࢿࣃࣄࣁࢾࣂࣄࣁࢿࣃࣄࢿࢾࣂࣃࢿࣄࣃࣁࢿࣆࣀࣃࣄࢿࣃࣆࣃࣀࣂࢾࣅࣀࣃࢿࣆࣄࢿࣂࣂࢾࣆࣁࣂࢾࣅࢿࢾࣁࣄࢿࣀࣁࣆࣁࣁࢿࣂࣆࣆࣄࣃࢿࣀࣁࣁࣃࣀࣅࣃࣃࣀࣂࣂࣀࢿࢾࣅࢾࣃࢿࢾࣅࣀࣂࣀࣁࣃࣃࣃࣀࢾࣄࣃࣆࢾࣃࣆࣄࣃࣅࣃࣂࣄࣁࣆࢾࣄࢿࣃࢾࢿࣅࣆࣃࢾࣄࣄࣂࢾࣂࣂࣁࣅࣅࢿࢾࣀࢾࣄࣅࣅࢿࣆࣄࣁࣄࣂࣁࣄࢾࣁࣅࣆࣂࢿࣄࣅࣆࢿࢿࣄࣀࣄࣅࣃࣄࢿࣀࣄࢿࣅࣆࣆࣆࢾࣀࢾࢿࣄࣂࣁࣆࣁࣅࣀࢿࣃࣀࣅࢿࣆࣃࣅࣂࢿࢿࢾࣀࣆࣆࣁࣂࣃࣄࣁࣀࣁࣆࢾࢾࣁࢾࣆࢿࣁࣆࣂࣃࢾࣄࢿࣂࣂࢿࣂࣃࣃࣄࣃࣂࣆࢿࢿࣄࣂࢾࣄࢾࢿࣃࢿࣁࣅࣆࣅࣁࢾࣀࣆࢾࣂࣂࣅࣃࣀࣂࢾࣁࢿࣅࣃࢿࣂࣄࢾࣆࢿࣂࣄࣁࣁࢿࣂࢿࣄࣆࣃࢾࣆࣀࣀࣀࢾࣃࣃࣀࣁࣃࣄࣀࣀࣆࣅࣁࣁࣂࣁࣅࣅࢾࣀࢿࣂࣀࣀࢿࢿࣃࣆࣆࣃࣄࣆࣆࣁࣁࣆࢾࢾࣃࣄࢿࢿࢾࣆࣆࣄࢿࣄࢾࣅࣂࣆࣀࣀࣅࣂࣁࣄࣅࢾࢾࣆࣀࣆࣆࣅࣀࣁࣅࣅࢿࣆࣀࣆࣂࢾࢿࢿࣃࢿࣃࢿࢾࢾࣄࣁࣁࢾࣁࣀࣅࣅࢾࣆࣄࣂࣄࣅࢿࢾࣄࣅࣄࣁࣅࣆࣄࢾࣀࣃࣄࢿࣂࣀࣁࣆࣆࣃࣁࣃࣆࣄࣆࢿࣁࢾࣂࣃࣃࣂࣆࣆࣂࣃࢿࣁࣆࣃࣆࣀࣅࣅࣃ;mod=0;returning-chatter=0;emotes=;flags=;tmi-sent-ts=1694024399198;user-type= :[email protected] PRIVMSG #!!! :@MedicWho _unicode
   

---

When running a release build:

[!WARNING]
Do this in a private channel! This will impact others!

1. Send

   ࢿࣃࣄࣁࢾࣂࣄࣁࢿࣃࣄࢿࢾࣂࣃࢿࣄࣃࣁࢿࣆࣀࣃࣄࢿࣃࣆࣃࣀࣂࢾࣅࣀࣃࢿࣆࣄࢿࣂࣂࢾࣆࣁࣂࢾࣅࢿࢾࣁࣄࢿࣀࣁࣆࣁࣁࢿࣂࣆࣆࣄࣃࢿࣀࣁࣁࣃࣀࣅࣃࣃࣀࣂࣂࣀࢿࢾࣅࢾࣃࢿࢾࣅࣀࣂࣀࣁࣃࣃࣃࣀࢾࣄࣃࣆࢾࣃࣆࣄࣃࣅࣃࣂࣄࣁࣆࢾࣄࢿࣃࢾࢿࣅࣆࣃࢾࣄࣄࣂࢾࣂࣂࣁࣅࣅࢿࢾࣀࢾࣄࣅࣅࢿࣆࣄࣁࣄࣂࣁࣄࢾࣁࣅࣆࣂࢿࣄࣅࣆࢿࢿࣄࣀࣄࣅࣃࣄࢿࣀࣄࢿࣅࣆࣆࣆࢾࣀࢾࢿࣄࣂࣁࣆࣁࣅࣀࢿࣃࣀࣅࢿࣆࣃࣅࣂࢿࢿࢾࣀࣆࣆࣁࣂࣃࣄࣁࣀࣁࣆࢾࢾࣁࢾࣆࢿࣁࣆࣂࣃࢾࣄࢿࣂࣂࢿࣂࣃࣃࣄࣃࣂࣆࢿࢿࣄࣂࢾࣄࢾࢿࣃࢿࣁࣅࣆࣅࣁࢾࣀࣆࢾࣂࣂࣅࣃࣀࣂࢾࣁࢿࣅࣃࢿࣂࣄࢾࣆࢿࣂࣄࣁࣁࢿࣂࢿࣄࣆࣃࢾࣆࣀࣀࣀࢾࣃࣃࣀࣁࣃࣄࣀࣀࣆࣅࣁࣁࣂࣁࣅࣅࢾࣀࢿࣂࣀࣀࢿࢿࣃࣆࣆࣃࣄࣆࣆࣁࣁࣆࢾࢾࣃࣄࢿࢿࢾࣆࣆࣄࢿࣄࢾࣅࣂࣆࣀࣀࣅࣂࣁࣄࣅࢾࢾࣆࣀࣆࣆࣅࣀࣁࣅࣅࢿࣆࣀࣆࣂࢾࢿࢿࣃࢿࣃࢿࢾࢾࣄࣁࣁࢾࣁࣀࣅࣅࢾࣆࣄࣂࣄࣅࢿࢾࣄࣅࣄࣁࣅࣆࣄࢾࣀࣃࣄࢿࣂࣀࣁࣆࣆࣃࣁࣃࣆࣄࣆࢿࣁࢾࣂࣃࣃࣂࣆࣆࣂࣃࢿࣁࣆࣃࣆࣀࣅࣅࣃ
   

2. Reply to the message

The manual text elision is really slow here, because it removes one character of the last overflowing word at a time. Qt has a native text elision algorithm, that seems to be faster (QFontMetrics::elidedText).

This isn't the exact same workload (as I couldn't interact with Chatterino before) but in both cases, the same channel was shown.

Before	After

Further Improvements

Even with this implementation, Chatterino will lag if many channels are open and one or more channels contain a reply to a bad message. This is because there are too many forced layouts. When many channels are open and an image gets loaded, all ChannelViews are forcefully laid out. Images are loaded when laying out messages (see #3929 and #3929).

A ChannelView that's not visible shouldn't need to be laid out. I attempted a fix in perf/layout-visible. This does have some consequences (e.g. switching channels isn't as smooth but not really noticeable - could be behind a setting [lazy vs. eager layout]). On this upside, this has the same effect as #3929 in that images won't be loaded if they're not needed, thus reducing memory usage and reducing lag when loading recent messages at the start for all open channels.

Another optimization one can make is to swap LimitedQueue for message layouts inside a ChannelView with a raw boost::circular_buffer, since it's only ever accessed by the GUI thread. I did that in perf/no-limited-queue. This will reduce a lot of unnecessary copies (for all the snapshots) resulting in less memory use and lower rendering times (copying the shared pointers requires incrementing the reference count).

Yet another optimization could be to remember in a MessageLayout if it contained any unloaded images. MessageLayout::layout could then ignore layout requests if an image was loaded, but the current element doesn't contain any unloaded image.

[github-actions]

clang-tidy made some suggestions

[Mm2PL]

Alternate reproduction steps that don't impact others:

1. Open !!! as a new split
2. Run this command a bunch:

/fakemsg @reply-parent-user-login=medicwho;id=162f1550-4686-40a9-bec1-66aa5cf4fcc0;subscriber=1;reply-parent-msg-id=72733dad-a857-408d-ab16-e96834aadf73;first-msg=0;reply-parent-display-name=MedicWho;reply-thread-parent-msg-id=72733dad-a857-408d-ab16-e96834aadf73;badges=subscriber/24,no_audio/1;reply-thread-parent-user-login=medicwho;room-id=11148817;user-id=117691339;turbo=0;color=#DAA520;reply-parent-user-id=71651832;display-name=Mm2PL;badge-info=subscriber/27;reply-parent-msg-body=ࢿࣃࣄࣁࢾࣂࣄࣁࢿࣃࣄࢿࢾࣂࣃࢿࣄࣃࣁࢿࣆࣀࣃࣄࢿࣃࣆࣃࣀࣂࢾࣅࣀࣃࢿࣆࣄࢿࣂࣂࢾࣆࣁࣂࢾࣅࢿࢾࣁࣄࢿࣀࣁࣆࣁࣁࢿࣂࣆࣆࣄࣃࢿࣀࣁࣁࣃࣀࣅࣃࣃࣀࣂࣂࣀࢿࢾࣅࢾࣃࢿࢾࣅࣀࣂࣀࣁࣃࣃࣃࣀࢾࣄࣃࣆࢾࣃࣆࣄࣃࣅࣃࣂࣄࣁࣆࢾࣄࢿࣃࢾࢿࣅࣆࣃࢾࣄࣄࣂࢾࣂࣂࣁࣅࣅࢿࢾࣀࢾࣄࣅࣅࢿࣆࣄࣁࣄࣂࣁࣄࢾࣁࣅࣆࣂࢿࣄࣅࣆࢿࢿࣄࣀࣄࣅࣃࣄࢿࣀࣄࢿࣅࣆࣆࣆࢾࣀࢾࢿࣄࣂࣁࣆࣁࣅࣀࢿࣃࣀࣅࢿࣆࣃࣅࣂࢿࢿࢾࣀࣆࣆࣁࣂࣃࣄࣁࣀࣁࣆࢾࢾࣁࢾࣆࢿࣁࣆࣂࣃࢾࣄࢿࣂࣂࢿࣂࣃࣃࣄࣃࣂࣆࢿࢿࣄࣂࢾࣄࢾࢿࣃࢿࣁࣅࣆࣅࣁࢾࣀࣆࢾࣂࣂࣅࣃࣀࣂࢾࣁࢿࣅࣃࢿࣂࣄࢾࣆࢿࣂࣄࣁࣁࢿࣂࢿࣄࣆࣃࢾࣆࣀࣀࣀࢾࣃࣃࣀࣁࣃࣄࣀࣀࣆࣅࣁࣁࣂࣁࣅࣅࢾࣀࢿࣂࣀࣀࢿࢿࣃࣆࣆࣃࣄࣆࣆࣁࣁࣆࢾࢾࣃࣄࢿࢿࢾࣆࣆࣄࢿࣄࢾࣅࣂࣆࣀࣀࣅࣂࣁࣄࣅࢾࢾࣆࣀࣆࣆࣅࣀࣁࣅࣅࢿࣆࣀࣆࣂࢾࢿࢿࣃࢿࣃࢿࢾࢾࣄࣁࣁࢾࣁࣀࣅࣅࢾࣆࣄࣂࣄࣅࢿࢾࣄࣅࣄࣁࣅࣆࣄࢾࣀࣃࣄࢿࣂࣀࣁࣆࣆࣃࣁࣃࣆࣄࣆࢿࣁࢾࣂࣃࣃࣂࣆࣆࣂࣃࢿࣁࣆࣃࣆࣀࣅࣅࣃ;mod=0;returning-chatter=0;emotes=;flags=;tmi-sent-ts=1694024399198;user-type= :[email protected] PRIVMSG #!!! :@MedicWho _unicode

Requires debug build

[Nerixyz]

Alternate reproduction steps that don't impact others:

Thank you for mentioning that! I added it to the reproduction.

[Nerixyz]

Just realized that SingleLineTextElement doesn't need to split the message into words. Emoji::parse can handle spaces. This would also reduce the amount of measurements when the text contains spaces. Furthermore, once Qt 5.12 support is dropped, Emoji::parse can return QStringViews into the original string.

[github-actions]

clang-tidy made some suggestions