All 9 patches now apply cleanly with git am on Linux (git 2.43.0).
Root cause of previous failures: hunk offsets were systematically wrong
by 7-40 lines; macOS git fuzzy-matched them, Linux did not.
Patches regenerated via git format-patch after applying all changes.
Fixes applied:
- P0000: unbind_to on no-candidate fall-through path; hunk regenerated
- P0001: block_input + record_unwind_protect_void in ns_ax_buffer_text
- P0003: [trims release] MRC memory leak; block_input already present
- P0004: mojibake comment (--- not UTF-8 em-dash)
- P0006: texinfo dangling semicolons -> periods in GNUstep paragraph
- P0007: em-dash fixes removed (content was already --- from P0004/P0005)
- P0008: childFrameLastBuffer -> BVAR(b,name) for GC safety;
BUF_OVERLAY_MODIFF removed from ensureTextCache (hl-line-mode O(N)
rebuild regression); block_input in ns_ax_buffer_text (P0001 scope);
voiceoverSetPoint and childFrameLastBuffer explicit init in
initFrameFromEmacs:; cachedOverlayModiffForText ivar removed
VoiceOver's browse cursor was not following Emacs cursor moves because
FocusedUIElementChangedNotification was posted on the buffer element
(a custom NSObject subclass), which VoiceOver ignores for browse cursor
re-anchoring.
Three changes:
1. Post FocusedUIElementChanged on the NSWindow instead of self, causing
VoiceOver to call accessibilityFocusedUIElement on the window and
re-anchor its browse cursor at the returned buffer element.
2. Post NSAccessibilityLayoutChangedNotification on self.emacsView with
NSAccessibilityUIElementsKey pointing to the buffer element, causing
VoiceOver to re-examine the element's properties including
accessibilitySelectedTextRange.
3. Post SelectedTextChangedNotification from self.emacsView (an NSView)
instead of self (NSObject), with NSAccessibilityUIElementsKey, so
VoiceOver processes text-change notifications from a view-based
element in the hierarchy.
The notification was posted on self.emacsView (the NSView container).
VoiceOver called accessibilityFocusedUIElement on the view, got back
the same buffer element, and did not re-anchor its browse cursor.
Fix: post on self (the buffer element itself). When VoiceOver receives
NSAccessibilityFocusedUIElementChangedNotification from an element that
is already focused, it silently re-queries accessibilitySelectedTextRange
and re-anchors its browse cursor without re-announcing the element name.
Covers all emacsMovedCursor granularities (arrow keys, M-f/M-b, C-n/C-p
via the separate isCtrlNP path).
Bug 1 (VO cursor not following Emacs cursor):
- Remove FocusedUIElementChangedNotification on emacsView (was a no-op:
VO re-queried the same element)
- For Emacs-initiated char/word moves, keep natural next/previous
direction instead of forcing discontiguous; SelectedTextChanged with
direction=next advances VO browse cursor sequentially
- Only force discontiguous for line-boundary crossings and large jumps
Bug 2 (word double-read with punctuation):
- Root cause was FocusedUIElementChanged causing VO re-anchor speech
on top of the explicit word announcement
- Removing FocusedUIElementChanged eliminates the duplicate speech
- Add emacsInitiated parameter to postFocusedCursorNotification;
omit AXTextSelectionGranularity for Emacs-initiated moves so VO
does not auto-speak (only explicit announcements provide speech)
- isWordMove now triggers on emacsInitiated flag (Emacs-initiated
word moves always get explicit announcement)
Removing the evil-mode EQ conditions left the if() without its
closing paren. Fix:
if (EQ (cmd, Qns_ax_next_line)
|| EQ (cmd, Qns_ax_dired_next_line)) <- add ')'
Same for previous_line. Verified: all 9 patches apply and build.
Commit 7609922 changed patch 0003 comment from
'for evil block-cursor mode' to 'for block-cursor mode'.
Patch 0008 had a '-' (removal) line still referencing 'for evil
block-cursor mode', which git am could not find in the file ->
'patch does not apply'.
Verified locally: git am now applies all 9 patches cleanly on f0dbe25.
Adding 4 lines in patch 0003 (punctuation trim fix) shifted all
subsequent nsterm.m positions by 4. Update the @@ -NNNN offsets
for all 14 nsterm.m hunks at line >= 9060 in patch 0008.
patch 0002: Do not activate mark in setAccessibilitySelectedTextRange.
VoiceOver range.length is an internal word-boundary hint, not a text
selection. Activating the mark made accessibilitySelectedTextRange
return a non-zero length, causing VoiceOver to position its browse
cursor at the END of the selection instead of the START.
patch 0003: Fix word announcement double-read and punctuation.
- Explicit word announcement only for Emacs-initiated (discontiguous)
moves; VO-initiated (sequential) word navigation relies on VO
auto-speech from the granularity=word notification, preventing
double-read.
- Strip trailing/leading punctuation from word announcements using
punctuationCharacterSet so 'Ahoj,' is announced as 'Ahoj'.
patch 0008: Post FocusedUIElementChangedNotification on the EmacsView
(not just on line-granularity moves) for all Emacs-initiated cursor
movements. Posting on the view causes VoiceOver to re-query
accessibilityFocusedUIElement and re-anchor its browse cursor at the
current accessibilitySelectedTextRange. Also fixes a pre-existing
off-by-one in the macos.texi hunk header.
Bug 1 (crash): postEchoAreaAnnouncementIfNeeded called Fbuffer_string()
without block_input, allowing timer events to interleave with Lisp
calls and corrupt buffer state. Added block_input/unblock_input via
record_unwind_protect_void. Also removed unpaired unblock_input()
in postCompletionAnnouncementForBuffer (patch 0003) that became a
double-unblock after patch 0008 added block_input + unwind protect.
Bug 2 (cursor sync): VoiceOver browse cursor did not follow Emacs
keyboard cursor because SelectedTextChanged with discontiguous
direction alone is not sufficient for VoiceOver to re-anchor. Added
NSAccessibilityFocusedUIElementChangedNotification post when Emacs
moves the cursor across a line boundary, forcing VoiceOver to re-query
accessibilitySelectedTextRange.
Bug 3 (word off-by-one): Evil w (next-word) read the previous word
instead of the destination word. VO auto-speech from
SelectedTextChanged with direction=next+granularity=word reads the
traversed word, not the arrived-at word. Added explicit word
announcement (like char moves) that reads the word AT the new cursor
position using whitespace-delimited word boundary scan.
BUF_MODIFF(b) dereferences the struct buffer pointer unconditionally.
If the buffer was killed, this accesses freed memory and crashes.
Check BUFFER_LIVE_P first.
Use precise Python line-index swap instead of Edit tool to avoid
accidentally replacing other patch content.
Patch 0008 modifies postFocusedCursorNotification to add
'&& direction != discontiguous' to the isCharMove guard, which IS
the cursor sync fix. Changing patch 0003 independently broke
patch 0008's context match. Restore patch 0003 to original.
Patch 0008: Move BUFFER_LIVE_P check before BUF_MODIFF dereference
in announceChildFrameCompletion. Accessing BUF_MODIFF on a killed
buffer is a null/garbage dereference.
Patch 0003: Always include AXTextSelectionGranularity in
postFocusedCursorNotification, including for character-granularity
moves. Without granularity, VoiceOver leaves its browse cursor at
the previous position on C-f/C-b/arrow moves. The explicit
AnnouncementRequested (High priority) still overrides VO speech
for evil block-cursor correctness.
- Correct three pre-existing DEFVAR_BOOL doc strings (ns-use-native-
fullscreen, ns-use-mwheel-acceleration, ns-use-mwheel-momentum) that
were accidentally modified by an earlier rebase; restore original text
- Break @property line in nsterm.h to 79 chars
- Replace em-dash with --- and break 80-char comment line in patch 0004
- Issue 1: Add explicit ApplicationServices import for UAZoomEnabled/
UAZoomChangeFocus (was implicit via Carbon.h, now explicit)
- Issue 2: Rename FOR_EACH_FRAME variable 'frames' -> 'frame' (plural
was misleading; matches Emacs convention)
- Issue 3: Move unblock_input before ObjC calls in
postCompletionAnnouncementForBuffer: to avoid holding block_input
during @synchronized operations
- Issue 4: Fix DEFVAR_BOOL doc and Texinfo: initial value is nil,
not t; auto-detection sets it at startup
- Issue 5: Replace magic 10000 with NS_AX_MAX_COMPLETION_BUFFER_CHARS
constant with explanatory comment
- Issue 6: Add comment to lineStartOffsets loop explaining it is gated
on BUF_CHARS_MODIFF and never runs on the hot path
- Issue 8: Rewrite all 9 commit messages to GNU ChangeLog format with
'* file (symbol): description' entries
- Issue 9: Break 81-char @interface line in nsterm.h
- Issue 10: Add WINDOWP/BUFFERP guards before dereferencing
cf->selected_window and cw->contents in ns_zoom_find_child_frame_candidate
- Issue 11: Fix @pxref -> @xref at sentence start in macos.texi
accessibilityIndexForCharpos: walked composed character sequences
from run.ax_start up to the target charpos offset. For a run
covering an entire ASCII buffer, chars_in = pt - BUF_BEGV, making
each call O(cursor_position).
This method is called from ensureTextCache on EVERY redisplay frame
(as part of the cache validity check), making each frame O(position)
even when the buffer is completely unchanged. At line 34,000 of a
large file this is ~1,000,000 iterations per frame.
Fix: when ax_length == length for a run (all single-unit characters),
the ax_index is simply ax_start + chars_in. O(1) instead of O(N).
This is the symmetric counterpart to the charposForAccessibilityIndex:
fast path added in the previous commit. Both conversion directions
now run in O(1) for pure-ASCII buffers.
charposForAccessibilityIndex: walked composed character sequences
from the start of a visible run to the target AX index. For a run
covering an entire ASCII buffer, this is O(cursor_position): moving
to line 10,000 requires ~500,000 iterations per call.
The method is called on every SelectedTextChanged notification
response (accessibilityBoundsForRange: from the AX server for cursor
tracking), making cursor movement O(position) in large files.
Fix: when ax_length == length for a run (all characters are single
AX index units — true for all ASCII/Latin text), the charpos offset
is simply ax_idx - run.ax_start. O(1) instead of O(position).
Multi-byte runs (emoji, CJK, non-BMP) fall back to the sequence walk,
bounded by run length (visible window size), not total buffer size.
postAccessibilityNotificationsForFrame: was calling NSString
lineRangeForRange: to detect line-crossing cursor moves. That
method scans from the start of the string, making it O(cursor_offset).
In large buffers with the cursor near the end, this executes on every
redisplay cycle — causing progressive slowdown proportional to cursor
position.
patch 0002 already builds lineStartOffsets in ensureTextCache (O(N)
once per text change) and exposes lineForAXIndex: (O(log L) binary
search). Use it instead.
Also remove the now-redundant tlen clamping that existed solely
to prevent lineRangeForRange: from receiving an out-of-range index.
lineForAXIndex: handles out-of-range inputs safely.
Root cause: the 50ms rate limit broke child-frame (Corfu) tracking.
When the Corfu child frame redraws, its ns_update_end fires first and
resets the rate-limit timer. When the parent frame's ns_update_end
fires immediately after, the timer has not expired, so
ns_zoom_track_completion returns early without scanning child frames.
Zoom focus stays on the first candidate.
Fix: remove the rate limit; add a FRAME_PARENT_FRAME(f) guard instead.
Child frames have no completion children to scan; their parent's
ns_update_end does the scan via FOR_EACH_FRAME. Returning early on
child-frame calls avoids the redundant scan and leaves the timer
problem moot. Overhead without the rate limit is ~40 Lisp evaluations
per redisplay (~5-20 µs), acceptable given ns_zoom_enabled_p() already
caches the UAZoomEnabled() IPC call.
Fget_char_property with a buffer as OBJECT checks text properties
only. Corfu highlights the selected candidate (corfu-current) via
an overlay, not a text property, so the scan always returned -1 and
Zoom focus stayed on the first line.
Pass cf->selected_window instead of cw->contents so that overlays
are included in the property lookup. Vertico uses text properties
and is unaffected; child-frame completion frameworks that use overlays
(Corfu, Company-box) now track correctly.
Changes:
- EmacsApp gets ns_update_accessibility_state and
ns_accessibility_did_change: methods (patch 0005)
- At startup: UAZoomEnabled() + AXIsProcessTrustedWithOptions()
determine initial ns_accessibility_enabled state
- com.apple.accessibility.api distributed notification updates it
whenever any AT connects or disconnects
- All Zoom call sites (UAZoomChangeFocus) now gated by
ns_accessibility_enabled in addition to ns_zoom_enabled_p()
- ns-accessibility-enabled docstring updated to describe auto-detect
Result: zero config needed; zero overhead when no AT is active;
single variable overrides auto-detection when needed.
500ms (2 Hz) was too aggressive — Zoom focus stopped updating during
keyboard navigation in Vertico/Corfu lists. 50ms (20 Hz) tracks
fast arrow-key navigation while still avoiding per-frame overhead.
UAZoomEnabled() is already cached so the main cost is the overlay
scan, which is cheap.
The Qnil initialization was in patch 0000 (Zoom) but the ivar
declaration is in patch 0008 (child frame tracking). Moved the
init to patch 0008 so each patch compiles independently.
Three fixes:
1. Patch 0000 now compiles standalone: replaced forward declaration
of ns_ax_face_is_selected (defined in VoiceOver patches) with
self-contained ns_zoom_face_is_selected in the Zoom patch.
2. ns_accessibility_enabled defaults to nil: eliminates ALL VoiceOver
overhead (text cache rebuild, AX notifications, Mach IPC to AX
server) when VoiceOver is not in use. Zero per-redisplay cost.
Enable with (setq ns-accessibility-enabled t).
3. UAZoomEnabled() cached for 1s + ns_zoom_track_completion rate-
limited to 2Hz: eliminates 150-600µs/frame of IPC overhead.
Root cause (per Opus analysis): UAZoomEnabled() is a synchronous
Mach IPC roundtrip to macOS Accessibility server, called 3x per
redisplay cycle. At 60fps = 180 IPC roundtrips/second blocking the
main thread. Combined with Emacs's inherent O(position) redisplay
cost, this compounded into progressive choppy behavior.
Fix 1: ns_zoom_enabled_p() caches UAZoomEnabled() for 1 second.
Fix 2: ns_zoom_track_completion() rate-limited to 2 Hz.
Also includes BUF_CHARS_MODIFF fix (patch 0009) for VoiceOver cache.
BUF_CHARS_MODIFF fix — the core performance regression:
ensureTextCache checked BUF_MODIFF which font-lock bumps on every
redisplay. Each cursor movement in a large file triggered full buffer
rebuild. Now uses BUF_CHARS_MODIFF (changes only on char insert/delete).
