Skip to content

Problems view on the viewport: search text on the lanes, JFace delta inserts and removes under filters and limits, proof harness - #55

Merged
hsoliwal merged 27 commits into
masterfrom
m3/jface-viewport-20261019
Oct 28, 2026
Merged

hsoliwal merged 27 commits into
masterfrom
m3/jface-viewport-20261019

Conversation

@hsoliwal

@hsoliwal hsoliwal commented Oct 24, 2026 •

Copy link
Copy Markdown
Owner

Problems view on the viewport: search text on the lanes, JFace indexed insert under a filter, proof harness

Branch m3/jface-viewport-20261019 forward-merged with master (the superset of PR #45, #48, #49, #44); this PR carries
the work of 2026-10-24 on top of it.

What changes

  • Search text on the lanes (M3TextMatch, M3MarkerLanes.applyText, MarkerPatternFilter.visibleFields/matchesText/matchesAny):
    the value a standard column shows is an interned label (message, resource, path, location) or a number (line,
    creation time, id), so the pattern's word match runs once per distinct label and is remembered per label id instead
    of once per row through a MarkerEntry facade (10^5 facades per keystroke at 10^5 markers before). Fields of other
    classes keep the facade route, evaluated only when no lane column matched. Design: docs/m3/jface-viewport/PROBLEMS_VIEW.md 3.1.
  • JFace viewport: an indexed delta insert keeps its position under a filter (AbstractTreeViewer.insertAt,
    StructuredViewerInternals.insertAt, M3Viewers): the stock insert turns an indexed insert into add() whenever a
    viewer filter is installed, and add() appends without a comparator; a filtered tree always has its pattern filter
    installed, so every viewport delta row landed at the end of its parent (category order after builds and filter
    toggles, rows added under a search text). Found by the 10^4 differential of this PR.
  • Proof harness: MarkerPatternFilter joins the overlay compile set (the overlay had carried the stock class);
    the bench reaches the e4 FilteredTree search box reflectively; the Windows Defender prompt is skipped on the hidden
    desktop; settle() logs its state every 10 s and gives up after 300 s; result rows are read with awk.
  • Docs: DELTA_GATE partVisible row (hidden views accumulate, shown patches); UPSTREAM_DEFECTS MV-8;
    donor sweep of the user's TV_5.0 structured viewers and the x86core jsonviewer (DONORS_TV5_X86CORE.md) with four
    queued leaves; HANDOVER log.

Proof

  • Headless: M3MarkerLanesAlgorithmsTest, M3SparseExtentsTest, M3TextMatchTest (lane match against the stock
    derivation over 5,000 random rows and 8 tokens): 15/15.
  • UI (hidden desktop, viewport mode): M3ViewersInsertUnderFilterTest: 2/2.
  • JFace layer A (run_differential.sh, 3 seeds x 150 steps, stock vs original vs viewport over list, table, tree,
    hashed and sorted-filtered variants): 0 mismatches outside the pixel hash with the insert fix (18/18 logs, both
    modes); the pixel hashes of some update and reveal steps differed today under memory starvation, and differed
    the same way in a pre-merge build (bfe9344) run alongside, so they are load noise (the 2026-10-20 run la4 was
    identical in both modes); to be rerun on a quiet machine before merging.
  • Differential stock vs m3 (seed 7, 30 steps; docs/m3/markers-view/tools/markers_round.sh 7 30 <n>):
    • 10^4 before the insert fix: 52 of 69 steps differed (category order, one misplaced row, MV-8).
    • 10^4 after the insert and remove fixes (run of 05:54-06:03 UTC, c499190): IDENTICAL, 69 of 69 steps,
      m3 counters applied=76 fallbacks=0 builds=2 shown-patches=1.
    • 10^5: IDENTICAL, 69 of 69 steps (stock 06:04-06:10 UTC; m3 06:59-07:04 UTC after a first start that came up without its extension registry on a starved machine, see the 07:05 log entry). m3 counters applied=76 fallbacks=0 builds=2 shown-patches=1; 0 rebuilds for 30 edits, 9 search and 4 scope steps (stock 30, 9 and 4 rebuilds); create 1119 ms vs 856, initial build 1575 ms vs 1514.
    • Selection lane (3.2 step 2, -Dorg.eclipse.ui.ide.markers.selection=lane, c28418d): 10^4 IDENTICAL, 69 of 69 steps, every one of
      the 76 patches restored its selection from the lane by position (lane-selections=76, selection-lane-mismatches=0); 10^5
      IDENTICAL as well; the default since 07:50 UTC (-Dorg.eclipse.ui.ide.markers.selection=jface keeps JFace preservingSelection).
    • 10^6 (m3 rerun 07:50-07:57 UTC with the bench waiting for the M3 jobs): IDENTICAL modulo the e4 FilteredTree expansion budget:
      65 of 69 steps identical, the 4 search steps differ only in a category expansion flag (MV-9, SOFT_MAX_EXPAND_TIME 200 ms;
      tools/diff_modulo_expansion.awk), 0 other differences; m3 applied=76 fallbacks=0 lane-selections=76, 0 rebuilds against
      stock 30/9/4; create 7061 vs 6141 ms, initial build 10929 vs 10747 ms.
    • JFace layer-A la5 (08:06 UTC, 3 seeds x 150, original and viewport, 18 of 18 logs): 0 non-pixel mismatches in both modes;
      px (paint hash) differs on timing under load only (unpainted first frame, paint runs); the quiet-machine all-fields-zero
      run of 2026-10-20 (la4) stands as the strict gate.
    • Counters (m3, 10^4): 76 deltas applied, 0 fallbacks, 1 shown patch; stock: 30 full refreshes and 30 rebuilds
      for 30 edits, 6 for 9 search steps, 16 for 4 scope steps.

Open

  • MV-8 withdrawn: the stock-side Warnings selection on a text change appeared only while the Windows Defender
    prompt was open on the hidden desktop (the bench then ran inside the dialog's nested event loop); with the prompt
    skipped under its real preference qualifier both sides agree on every selection segment.
  • The >50k pending-deltas fallback stays FAIL in DELTA_GATE until measured against the merge path.

… of 2026-10-23 (preserved on a side branch; run_ab.sh keeps the published version, the local version is run_ab.local-20261023.sh)
…ollections copy (M3SparseExtents, PR #49) that M3SlotOrder imports
The lane version of the search box (2026-10-20) ran the stock per-child test on a transient MarkerEntry per
slot: 10^5 facades per keystroke at 10^5 markers. M3TextMatch evaluates the standard columns on the lanes:
the value a column shows is an interned label (message, resource, path, location) or a number (line,
creation time, id), so the pattern's word match runs once per distinct label and is remembered per label
id for the pass. Visible fields map to lanes by exact class; any other field keeps the facade route,
evaluated only for rows no lane column matched (the stock OR over the visible fields). Derivations are
the stock ones (MarkerEntry.computeLocation, MarkerPathField bidi, creation time, id).

MarkerPatternFilter: visibleFields(), matchesText(String) (the stock wordMatches), matchesAny(item, fields).
M3TextMatchTest: headless, lane match against the stock derivation over 5,000 random rows and 8 tokens.
Docs: PROBLEMS_VIEW.md 3.1; HANDOVER open item 3 verified done (facades, id table).
…, e4 FilteredTree in the bench, Defender prompt skipped

apply_recipe.sh lists MarkerPatternFilter.java so build_overlay.sh compiles it (the overlay had carried the
stock class: the setPattern hook never reached a run). M3MarkersBench.filterText reaches the search box
through getFilterControl() reflectively (MarkersFilteredTree extends the e4 FilteredTree; the cast to
org.eclipse.ui.dialogs.FilteredTree failed both sides of the 10^4 round). The bench configuration sets
org.eclipse.ui.workbench/windows.defender.startup.check.skip=true (a modal prompt opened on the hidden
desktop during startup).
…ew patches, the text lane and the scope diff; handover log
…re jsonviewer swept against the programme (state providers, prioritised per-entry updater, filter algebra, placeholder entries); four leaves queued
…300 s; row grep locale-proof

A stock side of the 10^4 round settled for the whole 1800 s timeout without a line in the run log; settle()
now logs the elapsed time, the view description and the pending markers jobs every 10 s and throws after
300 s with the job names, so a hang names its step. run_markers.sh reads the result rows with awk (grep -P
fails under LC_ALL=C). Donor note: the relocate check of TV_5.0 verified in M3MarkerLanes.apply.
…lter (insertAt)

AbstractTreeViewer.insert turns an indexed insert into add() whenever a viewer filter is installed,
and add() appends when the viewer has no comparator. A filtered tree always has its pattern filter
installed, so every viewport delta row of the Problems view landed at the end of its parent: the
category order after the first build and after filter toggles (Warnings, Errors, Infos instead of
Errors, Warnings, Infos) and a row added while a search text was active (10^4 differential, seed 7,
52 of 69 steps). insertAt (package-private) runs the stock insert body without that check;
StructuredViewerInternals.insertAt exposes it to the M3 deltas in viewport mode and without a viewer
comparator, else the stock insert runs; M3Viewers uses it for tree deltas.

Test: M3ViewersInsertUnderFilterTest (UI, viewport mode): top-level and child inserts keep the model
position under a filter; a viewer comparator still orders the insert. 2/2 green on the hidden desktop.
Docs: HANDOVER log; UPSTREAM_DEFECTS MV-8 (stock gains a native selection on a text change).
…itial build, differential); matrix differential pass; PR body draft; handover log
…mismatches; pixel noise under load), handover
…TV_5.0 state providers): bit lanes over slots and categories, window-indexed restore, one selection event per patch, proof plan
…s; bench run script fixes

Every workbench view carries the large-view items limit (WorkbenchViewerSetup.setupViewer), so every
viewport delta remove of the Problems view went through the items-limit branch of
AbstractTreeViewer.internalRemove: one parent refresh per removed element, which on the lazy tree disposed
the items of the other elements of the batch without unmapping them and then threw "Widget is disposed";
the delta path fell back to ten clean rebuilds in one 10^4 round and the Warnings window kept 101-102 rows.
In viewport mode the branch now runs only when the parent shows an expandable node (a partial display);
disposed items are skipped and their stale mapping dropped. Stock mode is unchanged (UPSTREAM_DEFECTS JV-2).

Test: M3ViewersRemoveUnderItemsLimitTest (UI, viewport mode): two siblings removed in one batch under a
limit of 1000, children and top level.

run_markers.sh: the Windows Defender skip preference under its real qualifier (org.eclipse.ui:
WorkbenchPlugin.PI_WORKBENCH is PlatformUI.PLUGIN_ID), and a wait for 6 GB of free RAM before a run.
…; MV-8 withdrawn; API matrix differential cells filled; PR body
M3SlotBits: one bit per slot, packed long words, O(words) count, slot-ordered iteration, window positions
(headless M3SlotBitsTest against BitSet over 20,000 random operations). M3MarkerLanes keeps a selected lane
written from the view's selection listener (M3MarkerDeltas.selectionChanged: the marker ids of the selected
rows; a freed slot clears its bit) and answers isSelected, selectedCount and selectedPositions for a restore
by index. This step is a shadow: the viewer still preserves its selection through JFace; after every patch
the lane and the viewer's row selection are compared and a disagreement is counted
(diff-counters: selection-lane-mismatches), the proof that the lane follows the widget before it drives it.
Design: docs/m3/jface-viewport/PROBLEMS_VIEW.md 3.2. Dry-compiled against the overlay class path; the overlay
is rebuilt after the 10^5 round.
A workbench started on a starved machine (free RAM 941 MB to 5.7 GB for three
minutes before the run) came up without its extension registry: "Product
org.eclipse.platform.ide could not be found", the default theme missing, no
bench thread, and forty minutes to the hidden desktop timeout (m3-100000,
2026-10-24 06:15 UTC; the stock side of the same round ran in four minutes).

run_markers.sh now runs the hidden desktop runner in the background, watches
the run log for the bench's first line ("bench start") for MARKERS_STARTUP
seconds (300), stops the bench JVM when it never appears, and retries a run
whose JVM did not exit 0 (the bench halts with 0 after its last row)
MARKERS_RETRIES times (1), with the workspace, user area, log and the
label's result rows cleared before each attempt.
…rebuilds, 0 fallbacks); the starved first start and the watchdog in the handover log
…ction-lane-mismatches=0 over 76 patches; headless 17/17)
…(3.2 step 2, behind a property)

With -Dorg.eclipse.ui.ide.markers.selection=lane the window patch of M3MarkerDeltas
no longer runs inside JFace preservingSelection (the selection as an element list
before the patch, setSelectionToWidget with an item lookup per element after). The
patch runs bare; the selected items of the untouched windows and the category items
are kept, the rows of every touched window come from the selection lane by position
(M3MarkerLanes.selectedPositions over the window slots), the tree's selection is set
once, and one selection event follows when a selected row left the window or the
model (MarkersTreeViewer.m3SelectionPatched). O(selected + touched windows) item
references, no facade materialised, no element lookup. Counter lane-selections in
the differential counters. Default unchanged until the differential in lane mode is
IDENTICAL; the shadow lane it relies on was proven at 10^4
(selection-lane-mismatches=0 over 76 patches).
… M3 jobs

The window patch restores the selection from the lane unless
-Dorg.eclipse.ui.ide.markers.selection=jface (the differential in lane mode was
IDENTICAL at 10^4 and 10^5 with every one of the 76 patches restored from the lane).

Bench: settle() waited only for jobs whose class is in
org.eclipse.ui.internal.views.markers; the M3 jobs are Job.create instances
(anonymous classes of org.eclipse.core.runtime.jobs), so at 10^6 the scope job
outlived the 800 ms settle window and the m3 side was captured one step behind on
every scope and filter change (13 of 69 steps). isMarkersJob now matches the M3 jobs
by name; the settle log names them.
… time-bounded expansion (SOFT_MAX_EXPAND_TIME 200 ms after a search text: at 10^6 the slower side expands fewer categories with the same model)
…on after a search (NOISE, read through diff_modulo_expansion.awk)
…e expansion budget (65 of 69 identical, 4 expansion-only, 0 other; lane-selections=76)
…ion on DIFFERENT and honours OUT (one-side reruns in an isolated tree)
…on-pixel mismatches; px differs on timing under load only
@hsoliwal
hsoliwal merged commit 12a83cf into master Oct 28, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant