On 10/11/2021 20:48, Ihor Radchenko wrote:
I pushed the patch upstream.
P.S. I am puzzled that `org-element-at-point' returns paragraph even
when patch is not applied. After such call `org-store-link' does what I
expect. However "after undo" erroneous state survives emacs restart,
thanks to org-persist, I believe.
The bug was in very specific case. org-element-at-point now has a
special fast mode of operation when CACHED-ONLY argument is non-nil. In
this mode, it does not try to parse anything, but only checks the
nearest cached element and returns it if point is at the element
beginning, within +begin/end line of the element, or inside a
non-greater element. The error happened when the cached element was a
section and point was within blank lines at the end of the section
exactly one symbol before the following headline. In such cases,
org-element-at-point must follow some special rules (see comments in
org-element--parse-to), but it did not in this specific scenario.
On top of that, org-element-parse-to cannot return section and the
current code assumes that we ever attempt to return section when point
is at the beginning of section. So, it returns
(org-element-at-point (1+ pos)) instead. Hence, the next headline
was wrongly returned when (1) point is 1- beginning of next headline
(2) we are within blank lines at the end of section (3) some edit in
buffer just cleared everything in cache inside current section
(4) org-element-at-point is called with non-nil CACHED-ONLY argument.
Thank you for the fix and for the detailed explanation, Ihor. I was
naively suspecting some weird issue with undo, maybe similar to the one
found with interference of tables and non-latin input methods.
