Repository navigation
Conversation
 [Medium impact] The PR appears safe to merge; no new actionable issue was found. SummaryThis PR reduces memory held by large folder listings and grid text previews.
No new actionable issues were found. The earlier unnumbered finding about cancelled reads has been addressed by clearing pending requests and keeping owed rows available for retry. This review used source and test inspection; tests were not run.
|
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
thisisgm
force-pushed
the
memory
branch
2 times, most recently
from
November 12, 2026 13:03
45307c1 to
63917a5
Compare
Building a listing asked the disk about every file a second time, because Qt stats a file again before it reads the name. A regular file is now typed from the name the listing already read; a link, a pipe or a device keeps the full lookup.
A selection, a folder count or a second view of the same folder each made the model copy its whole listing. They now read it in place, so a large folder is held once however many views show it.
The preview's folder summary sorted every name to count them and to pick two pictures. It now keeps the first two pictures in one pass, so summarizing a large folder no longer holds every name.
Every entry kept its own copy of its type, kind and extension, and the listing kept spare room for entries that never came. The entries of a listing now share those strings, and the spare room is given back.
Opening a folder typed every file before its first row showed. The listing now leaves a regular file's type, kind and extension unread until a role asks, which touches no disk; a sort by kind still types every entry in the worker.
Every refresh of Places, the 15 second poll included, told all four sidebar lists they had changed. Each list now has its own signal and is announced only when its contents differ.
While active, Folio ran hyprctl once a second to read the corner rounding. It now reads it at start, on a palette change, on return to the window and when Hyprland's event socket reports a reloaded config.
Sorting 100,000 names collated both strings on every compare. A listing that arrives in a few ordered runs is now merged, and a name or kind sort makes one sort key an entry first. The comparator is the same, so every order and tie stays; the C locale compares as before.
A name or kind listing read every entry's size and date before its first row showed. It now reads the first 320 rows, and a later row when a view asks, one batch a turn, off the UI thread. A row not read yet shows no size and no date, never 0 B.
A re-read that found the same names kept each folder's item count, even when the folder's own date had moved. Such a folder's count is now dropped and counted again, as it was before the listing learned to keep its rows through a re-read.
The event socket was opened once. Hyprland hangs up on a reader that fell behind, and after that a config reload no longer reached the corners. The watch is now taken up again when Folio is back in use.
A text tile past the rows read at first asked for its page with no date, then again when the date arrived, and a page has no cache on disk, so the file was read and drawn twice. Like a thumbnail, a page is now asked for only once its date is known.
A folder past the rows read at first is counted without its date being read, so a re-read had nothing to compare and kept its old count. Its date is now read just before it is counted and kept with the count, and the re-read sees the folder changed since.
A burst of events was read whole into memory before any of it was looked at, on the UI thread. One turn of the event loop now reads 64 KiB at most and looks at each piece as it comes, and a line too long to keep is dropped to its end, its last piece included.
Hung up on by Hyprland, the theme took its watch up again only when Folio was left and come back to, so a config reload in between left the corners as they were. While in use it now connects again after a pause, and reads the rounding for a reload that went by untold.
Hung up on while Folio stayed in use, the theme tried Hyprland's event socket once more after a pause. A try that found no socket was the last, and reloads went untold until Folio was left and come back to. Each failed try now starts the pause again.
A re-read drops the answers on their way and reads the rows asked for itself. A row it could not read either stayed marked as being read, so no view could ask for it again. Such a row is now owed: a view may ask again, the folder's next re-read tries it again, and a view is told when another view read it.
A view asks for a row's kind, MIME type and extension each time it makes the row, and each ask read them from the name again: 48 us a row in /usr/bin. The model now keeps up to 4,096 names' types, until a folder is found changed.
A tile fits a text file's page into 128 by 92, and the page was kept as drawn, at 256 pixels: nine times the pixels shown, in memory and as a texture. The tile now says the most it draws, twice that once zoomed in, and the page is handed over no larger, scaled smoothly.
Every row held room for a type, an extension, a plugin item's glyph and details and a folder's first pictures: 224 bytes, mostly unused. They are kept aside in one block, shared by the rows of one type, and a row is 80 bytes: 14 MiB less with a folder of 100,000 files open.
Every row held its whole path beside its name, a string of its own that repeats the folder's. A row now shares one string for its folder and its path is put together when asked for: about 14 MiB less with a folder of 100,000 files open. A plugin item keeps its id whole.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Stacked on #26: its eighteen commits come first here, the last three are this pull request's, and I will rebase it as soon as #26 lands.
Three changes to what a large folder costs in memory, one commit each with its own tests. A row of a listing was 224 bytes and held its whole path; it is now 80 bytes and shares its folder's path with the rows beside it. And the grid keeps a text file's page at the size its tile draws it, where it kept the whole page it was made at. From a cold start into a folder of 100,000 files the process holds less (about 64 MiB in the grid, 29 in the list and 32 in the columns by the medians), and the time to the first window, the time to settle and the CPU time stay where they were. The suite goes from 236 to 241 tests. No library,
dependsoroptdependsentry is added.What you'd notice
Image.sourceSize, which Qt hands to the provider in the screen's own pixels, and the provider hands back a picture no larger and never an enlarged one. Zoomed in, the page is kept at the zoom's whole step. Photos, videos, PDFs and every other thumbnail are left as they are made.Numbers
Measured on a mini PC, cold start into a folder of 100,000 files, twelve runs a view, #26's head (
10f84a6) against this branch's. Each cell is the median, with the lowest and highest run in brackets: the timings are the same within their spread, the memory is what moved. Settled is the time until the window is up and the process has gone quiet. Both are the CMake build.Screenshots
The grid's text pages, before above and after below. They are drawn at the same size, from a picture with an eighth of the pixels.
Tests
Five new cases, each failing without its commit: a row is 80 bytes, a row listed by name keeps nothing aside, the rows of one type share one block and a write to one leaves the other alone, a counted folder without pictures stays small, and a name without an extension shares no block with a name that ends in a dot; the rows of a listing and the matches of a search share one string for their folder, each path is the file's own, and a plugin item's id survives whole; a row of the folder come to is not read because a row of the same name was read in the folder left; the provider hands a picture over no larger than the view draws and never enlarges one; and the grid keeps a text page at the tile's size, twice that once zoomed in, while a photo's thumbnail keeps all it has. The two picture cases pass at a display scale of 2 as well as 1.
Limits