Repository navigation
Edit and save MAT-files - #88
Merged
Merged
Conversation
The watcher and its settle, the reload of a clean document, the save-time conflict check and the Overwrite / Revert prompt were private methods of BinarySlddEditorProvider. They are now diskFollow.ts, over a FollowedDocument interface the binary dictionary document implements, so a second writable byte-backed editor can answer an outside write and a save over one by the same rules. What each format reads off the disk and how it replaces its document stay with its provider, handed in as read and apply. No behaviour changes.
A .mat now opens in its own writable editor (dataExplorer.matView, the default for *.mat) instead of the read-only view, which hands on a .mat that lands in it. Values, renames, Add (a variable named var, holding 0, with its name editor open), Add Child and Delete all go through core's session, whose gate offers no edit a save would refuse or write as something else; a refused edit records nothing and leaves the file clean. Undo and redo are core's: each gesture pushes one VS Code edit that replays the steps core recorded for it. A save is core's serializeSource: every variable nobody edited is the bytes it was read from, -v6 stays uncompressed, and a save core refuses writes nothing, keeps the edits and says why, naming the file. The bytes replace the file by a temporary file beside it, flushed and renamed over it, keeping its mode, following a symbolic link, writing in place where the file has another hard link, and refusing a read-only file; any file system but the local disk is written as VS Code writes it. A save keeps the tree, so an edit can still be undone after the save that wrote it, auto-save included. The Status column says what differs from the file as last saved, kept beside VS Code's stack. Revert, hot- exit backup and restore, and an outside change follow the binary dictionary editor's rules (diskFollow.ts). A file core cannot save opens read-only with the reason in the banner; a -v7.3 file shows why it cannot be opened. The table's setRows gains add: 'variable' and clipboard: false: the Add button and its key add a variable at once, and Copy, Cut, Paste and row drags are not offered, since a variable has no dictionary entry form. MAT rows carry the menu's capability flags. The integration suite opens copies of three MATLAB-written files from core's test corpus (test-integration/fixtures/mat) and compares each saved file with the original record by record.
…e a replace cannot be made A save's fallbacks wrote with fs.writeFile, which truncates first, so a full disk or a crash during a hard-linked file's save left it short. The write in place now opens the file for update, writes the new bytes over the old from the start, and only then cuts it to their length. It is also the fallback wherever a replace by rename cannot be made and VS Code's own save would still succeed: a hard link, a directory that takes no new file, a rename the file system refuses (EPERM, EBUSY, EXDEV, EACCES, ENOTSUP: OneDrive or a scanner on Windows, a bind-mounted file, some network and FUSE mounts), and a temporary file that cannot be given the original's mode, owner or group. A refused chmod no longer fails the save, and a replace never changes who may read or write the file. A link to nothing is written through, and stays a link. The write answers the stamp the written file has, read off it before it takes the name, so an outside write landing just after a save is not taken for the save. A temporary file a crash leaves behind is removed by the next save of that file; its name is random from crypto. The header says what a replace drops (extended attributes, ACLs, Finder tags) and what a write through another file system cannot promise.
…ackup against its own stamp - A gesture that throws after core recorded some of its steps is still one edit VS Code knows of, so the next Undo undoes it and not the edit before it. The table is repainted whole, then the error shown. - A hot-exit backup's id records the stamp of the file the edits were made over; a restore's first save asks before writing over a file another program changed since, instead of taking its current stamp. - A revert reads the disk before it marks the document clean, and one that cannot read it puts the tree back to the bytes the file held as last read or saved, so a clean document never holds dropped edits. - A backup core refuses keeps the last backup that could be written standing (each handle holds a reference, so the extension host's delete of the old handle does not take the file), and says so once. Measured: a backup of a 91 MB file with one edit serializes in 3 ms and writes in 15 ms; of a 100,000-variable file, 173 ms and 2 ms. - A save's stamp is read off the written file. Save As onto the file itself, or onto another link to it, is a save. - Every message shown of core's names the file, not the srcId, the parse-warning banner included. - After a revert or a reload, the Property Inspector shows its node again from the new tree, or clears. - Said in the code: a stale undo leaves the document dirty while it equals the disk, and a MAT-file opens in one panel, as a binary dictionary does.
…when its menu check throws With `clipboard: false` (a MAT-file) rows are no longer draggable, and a drag from another table draws no drop target over them. The row builder's menu flags are computed defensively: a check that throws offers nothing and the row still shows, where it used to fail the whole payload. Rows are copied once instead of twice; the flags are stamped on every MAT row, a read-only file's too, which the code now says. Timed against a 19 MB file of eleven 183,000-cell variables and a file of 100,000 variables (best of 5): expanding one cell takes 514 ms without the flags, 567 ms with them; the 100,000-row payload 285 ms without, 309 ms with. Copying once instead of twice is not measurable. The read-only view's dead .mat branches (its Usage filler, two comments) are gone: a .mat is handed on before they run.
The headers of diskConflict.ts and diskFollow.ts, and one comment in the binary dictionary editor, named a ticket and its ruling; they now say the behaviour. diskConflict.ts also names both writable byte-backed views, and diskFollow.ts as its vscode half.
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
1 participant
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.
.matfiles now open in an editable view and save, throughdata-explorer-core1.44's MAT-file save..slx,.mdland.prjstay read-only.What a user can do
All of these come with undo and redo, including undo past a save. Only edits core can write are offered. A file core cannot save opens read-only, with the reason in a banner.
How the save keeps the file
files.saveConflictResolution. That logic now lives indiskFollow.ts, shared by both editors with no behaviour change. A crash-recovery backup remembers the file version its edits were made over, so a restore asks before overwriting a file changed since.Tests
matEdit.test.tscover:npm run verifypasses (3598 tests), andnpm run test:integrationpasses (311 passing, 1 pending, 0 failing).-v6file and undo past a save.Two reviews found no blocker. Their findings, mostly about the safe write and edge cases of the document's lifecycle, are fixed in the last four commits.