| Description: |
When a tag — including the implicit `sel` tag — splits a display line containing
RTL text into several chunks, `LayoutDLine` places the chunks left-to-right in
logical order while the font backend shapes each individual chunk right-to-left.
The result is a scrambled line: reading order jumps around, and while dragging
the mouse the moving selection boundary keeps re-splitting the line, so the
layout changes under the pointer at every index step. Visually, the selection
"dances".
Reproducer (X11, Tk trunk built with `--enable-bidi`):
```tcl
package require Tk
text .t -font {DejaVuSans 16} -wrap none
.t insert 1.0 "שלום עליכם, זהו טקסט מימין לשמאל"
pack .t
update
# record the x of each character bbox, then apply the sel tag:
.t tag add sel 1.4 1.8
update
# record the x of each character bbox again: every index has moved
```
Expected: the visual position of every character is unchanged by the tag.
Observed (trunk): all 32 character positions move when the tag is applied; the
first four characters even end up on the wrong side of the line. Sample data
(x per char, before → after): `949 938 933 920 …` → `684 673 668 655 725 713 …`
**Root cause:**
The bidi resolution happens inside the font backend for a single contiguous
run, but nothing reorders the *chunks* of the display line when a tag splits
that run. This is the chunk-granularity counterpart of what the revised text
widget already solved ([e7c85cd941]).
**Proposed fix** (attached: `tkTextDisp-bidi-reorder.patch` against trunk
`generic/tkTextDisp.c`):
1. `BidiReorderChunks()`, called at the end of `LayoutDLine()`: resolves the
strong bidi direction of each char chunk (`ChunkBidiDir()`, a tri-state
variant of `ChunkIsRtl` using the same Unicode ranges — Hebrew, Arabic,
Syriac, NKo, Thaana, Samaritan, presentation forms, historical RTL scripts),
resolves neutrals (simplified N1/N2), then applies rule L2 by reversing each
maximal run of RTL chunks, reassigning their x positions in visual order.
The run keeps its left edge and its total width, so the line length and all
downstream geometry are unaffected. Non-character chunks never move.
2. `CharBboxProc()`: with `TK_LAYOUT_WITH_BASE_CHUNKS`, `CharChunkMeasureChars`
always anchors the advance at the *base* chunk, so the mirrored bbox of an
RTL chunk must be computed from the base chunk, not from the chunk itself:
mirror around the right edge of the run (the right edge of the base chunk,
which the reorder puts first-logical → rightmost). Without this, every char
of the non-base chunks collapses to a single x. The mirrored x is *not*
clamped to the chunk box: in base-chunk space it is expressed in the line's
coordinate system and legitimately falls outside the chunk that merely
carries the character's style (e.g. a `sel` chunk placed at the right edge
of an RTL line while its characters render further left) — clamping it
there re-created the dancing selection under Windows.
3. `TkTextDLineInfo()`: after the reorder the first chunk of the list is no
longer the leftmost one, so the left edge of the line is the minimum chunk x.
4. Windows only (`#ifdef _WIN32`, base-chunk block of
`CharChunkMeasureChars`): clamp the value returned by `MeasureChars` to be
no less than the span start. The Uniscribe backend returns non-monotonic
positions (notably when the measured range ends with `\n`, see Ticket 3);
without the clamp the newline chunk gets a negative width that corrupts the
DLine length and, with `-justify right`, shifts RTL lines out of the
widget.
Deliberate simplifications (documented in the code): the paragraph base level
is always LTR; European numbers are treated as neutrals; reordering is at chunk
granularity, not character granularity.
**Evidence:**
- Branch: https://github.com/Surcouf01/tk/tree/bugfix/dancing-selection-bidi
- CI (headless Xvfb, repro above scripted): `PASS: RTL layout is stable under
the sel tag` — all 32 character x positions identical before/after the tag,
`dlineinfo` unchanged (655 3 308 24 19). Logs:
https://github.com/Surcouf01/tk/tree/ci-results/bugfix/dancing-selection-bidi
(`.ci-logs/test1.log`); baseline on trunk-aligned `main` FAILs the same test.
- Windows (manual build, MSVC `nmake -f makefile.vc`, Arial 14, widget
884x532, `-wrap none`, same scripted test with `-justify right`): all 32
character x positions identical before/after the tag,
`indices whose x moved:` empty — `PASS`. Diagnostic branch with
instrumentation: https://github.com/Surcouf01/tk/tree/win-debug-bidi
(probes are compiled out in the attached patch; nothing is gated on
environment variables in the submitted code).
- Manual drag-testing of an interactive selection on a pointed-Hebrew line is
still recommended as final confirmation (CI is headless).
|