The test that succeeded from this change was as a result of allowing the
end of the character format change be specified using NULL as the rest
of the text. Before, the end paragraph run at the end of the text was
not being set for this case, when all the text was supposed to have its
character format changed.
Plenty of places in the code find following or preceding runs, then
afterwards find the paragraph from the run. This is inefficient because
the same linked list is used for both runs and paragraphs, so changes in
paragraphs can be detected while returning the next or previous run.
The code was previously getting the same text in the loop from the
editor, and it was converting each of the prefixes to compare against
for each URL candidate.
Before a single length was used for the number of characters to retrieve
from the text, and to keep track of the size of the buffer. These are
not equivalent, since there is a possible end of line conversion.
Previously the only convenient way to get the start and end of the
selection was through offsets, which eventually need to get converted
back into items in the linked list storing the text. The new function
will help with eliminating these inefficiencies.
This function will make it easier to work with ME_Cursor objects, which
should be used in a lot of places instead of character offsets (which
often require seeking through the linked lists to perform operations
with).
Since the table row start is inserted after the rest of the table is
inserted, the cursor is saved, and temporarily moved to the start of the
row to insert the table row start paragraph. Unfortunately the
paragraph in saved cursor becomes invalid during this insertion and
needs to be updated, so this code introduced a regression once
paragraphs started to be stored in cursors.
Style objects are referenced counted in richedit controls, so I tried to
make sure styles were released properly. This can be checked using with
the all_refs global reference count to see if everything is cleaned up.
Previously inserting the object didn't result in the text being wrapped,
which would cause an assertion error when this is checked for during
repainting the text. It is also important to invalidate the affected
areas of text, update the scrollbar, and end the creation of undo
transactions for this insertion.
The application AutoGK was getting the length of the text with
WM_GETTEXTLENGTH to allocate an appropriate buffer size, but then
claimed the buffer was twice the size when sending WM_GETTEXTEX. This
caused the memcpy call to overflow the actual buffer since the count
is based on the size of the buffer alone, regardless of the amount of
text retrieved.
ME_GetTextEx directly handles EM_GETTEXTEX, and previously a NULL buffer
would be dereferenced, and a 0 buffer length would cause nCount an
underflow in the nCount value which would allow a buffer overflow to
occur.
The application Blitzin2 was sending WM_VSCROLL messages to the
richedit control directly, when normally this message is supposed to
be a notification sent after the scrollinfo is set. Native richedit
controls always use the 16-bit value passed to this message to set the
scroll position for SB_THUMBPOSITION, rather than trying to find a
32-bit value through GetScrollInfo like I had previously done.
The missing return was to prevent automatic vertically scrolling when
style for it is not set, but is set for the horizontal scrollbar. I
initially missed this by testing with no autoscrolling on either
scrollbar.
Rich text files have groupings of text, where styles are pushed onto
the stack when encountering a start of the group, then popped at the
end of the group. This was being handled improperly before, because a
single styleChanged flag was being stored to keep track of whether the
style needed to be restored at the end of a group. This fails to work
properly since the single flag isn't keeping track of all the levels
of the stack, so some styles are not restored properly.
When a colour table entry is empty, then the default colour is used.
For an incomplete colour table entry 0 is used for the missing colours.
Previously the -1 value used internally for missing colours was being
converted into white, where it should be using the default colour that
is normally black.
This bug could be seen by loading the following rich text into wordpad:
{\rtf{\colortbl;;}\cf1 text}
Previously after initial window creation of a richedit control with the
ES_DISABLENOSCROLL window style flag the scrollbar would be shown but
not disabled. This patch fixes this issue by explicitly disabling and
showing the scrollbar.
When the richedit control is created without the ES_AUTOHSCROLL or the
ES_AUTOVSCROLL, then scrolling isn't done automatically when moving the
cursor with the arrow keys, drag selecting with the mouse, or even from
using the EM_SCROLLCARET (based on testing using a modified version of
wordpad).
WM_SETTEXT seems to check for {\rtf or {\urtf to determine if it is an
ascii RTF string, even if it is a unicode message. So I removed the
check to see if it is a unicode message, and added a check for {\urtf.
This was partially handled by ungetting the end group token (i.e. '}')
so that it was read again at the proper place. Unfortunately there is a
read hook that maintains a stackTop variable, which is decremented when
receiving this end group token. Therefore stackTop would get
decremented twice, and the rich text file would end prematurely.
This is fixed by incrementing the stackTop variable to compensate for it
being decremented twice.
Wine was not doing bounds checks for EM_GETTEXTRANGE, which was causing
a crash in Bug 17822. The added tests would cause a crash without the
added bounds checks in the richedit code.
The bounds checks I put in HandleMessage, since ME_GetTextRange is also
called for ME_GETSELTEXT which should not have bounds checks, since it
uses the selection range.
When the ME_GETTEXTRANGE message returns 0, no text is copied, not even
the NULL terminating charter. This differs from EM_GETSELTEXT which
will copy the NULL terminating character when no text is selected. This
behaviour is consistent with native richedit controls.
Images that are inserted into richedit controls store a space for the
text, since that is the character returned when getting the plain text
from the control.
When calculating the width of a line, the space character is skipped,
but images should not be skipped. This can be seen by inserting an
image into wordpad on a line by it's own, then centering the line. The
image will start from the center rather than being centered in the
control.
EM_SETCHARFORMAT can be used to make text links. Automatic URL
detection being enable would cause these links to be removed if the text
is not a URL, so this must be prevented.
Previously checks were made for AutoURLDetect_bEnable before calling
ME_UpdateSelectionLinkAttribute, or ME_UpdateLinkAttribute. This is
more error prone than checking for this within the function, so one call
was missing this check.
ME_SetCursor also didn't respect this behaviour, since it wouldn't set
the cursor to the hand when hovering over a link without automatic URL
detection disabled.
Before the code was modifying the format rect to compensate for space
being added or removed for the selection bar, but this should only
happen when the ECO_SELECTIONBAR setting bit changes.
When all the text fits on the screen, the scrollbars are not shown from
EM_SHOWSCROLLBAR. The message instead adds support for the specified
scrollbar when lParam is TRUE, so that the scrollbar will be shown when
sufficient text is in the document.
The scrollbar visibility can be changed from SetScrollRange or
SetScrollInfo, but the visiblity that is a result of these calls are
not consistent with the calculation made by richedit controls to
decide whether to show or hide the scrollbars.
The ME_TextBuffer structure is what is used to store the document (as
a linked list), so the ME_Document structure isn't being used. There
was also a document pointer in the ME_Paragraph structure that was
also unused, so I removed it because it is probably related to this
used structure.
This prevents some needless searching for the start of the paragraph
from a run stored in a cursor. Usually a pointer to the paragraph is
already available when the cursor is set anyway.
These functions were just being used for addition, so it was simpler to
remove the functions and modify the places it was used.
The ME_StrRelPos2 and ME_PosToVPos were just simple wrappers around
ME_StrRelPos, and ME_PosToVPos wasn't being used.
These two functions were being used for simple operations, to get the
first or last character when pre-computing flags for splitting runs.
The call to ME_GetCharBack wasn't even giving the correct result, it
would always return -1 since it is being called with nPos of 0.
This patch simplifies the code by removing the functions and getting the
characters directly from the string.
These functions were probably previously needed because of some wierd
special handling of backspace characters, but currently there is no
reason why the nLen field can't be accessed directly.
Having to functions that just access the string length field just causes
slightly more effort for someone to look at the code, because they need
to enter the function to find out what it actually is doing.
The function was just returning the second parameter. It had some
commented out code that indicated that previously backslashes weren't
included in the length. Native wordpad doesn't handle backspaces in a
special way, so this must have been an internal representation that
complicated finding the position of characters.
ME_GetCursorCoordinates had two conditions that were always taken. The
first condition was if(pCursor->pRun->type == diRun) was following an
assertion making the exact same check. The next one, if(row), should
always be taken, otherwise the richedit controls are in a corrupt state,
therefore an assertion is more appropriate.
I found that ME_MakeStringB was previously unused, and that the other
ME_MakeString functions repeated code that was already in ME_MakeStringB.
Making ME_MakeStringB static and using it to avoid duplicate code seemed
like a better idea than removing the function.
Previously it found the start or end by traversing the linked lists of
run, rows, paragraphs, and cells from the current position of the
cursors. Clearly it is better to get the start or end directly to make
it a constant time operation.
Wrapping is needed to be done even when repainting isn't done since
later messages expect line breaks to reflect the current text. Some
message can specify not to paint the sceen, but this should prevent
wrapping from being done.
- A HWND can be safely marshaled over a LONG as its payload is not
a pointer but a user handle.
- Use GetWindowLongPtr instead of GetWindowLong to retrieve a pointer.
When finding an adjacent paragraph, the next_para and prev_para pointers
should be used because they are direct pointers, a constant time
operation. Instead I found some places in the code that searched through
the general linked list to get to an adjacent paragraph, which is a linear
time operation, depending on the number of rows and runs in between
paragraphs.
The fixme comment is suggesting wrapping a paragraph within a function
that is for moving the selection cursor up or down a line when the up
or down keys are pressed. The contents fo paragraph aren't being
changed, so there is no need to wrap the paragraph.
More case of searching for the paragraph through the linked list when
is was already previously available. Since each wrap context is used
for wrapping each paragraph, I decided to add the reference to the
paragarph in the structure.
Rather than get the paragraph from the run, the function allows the
caller to provide the paragraph, since it is already available. This
reduces unnecessary traversals of lists that take longer as more runs
and rows are in the paragraph.
The ME_RunOfsFromCharOfs function finds the paragraph before finding the
run and offset within the run, so the function may as well be able to
return this paragraph to the caller. Many callers to the function
instead find the paragraph from the run, which ends up unnecessarily
traversing a linked list of runs within the paragraph.
Whenever ME_InitContext is called, ME_DestroyContext should be used to
clean it up. This way the context can be extended easily by modifying
those two functions. Instead, these two places of code just released
the DC, without using ME_DestroyContext, so the created brush for the
margin was not deleted.
These calls to ME_WrapMarkedParagraphs never do anything, and don't make
sense to be called in these places. These places are for ME_MoveCaret,
and ME_ArrowHome, which both don't involve any text being modified, and
all (direct and indirect) calls to these functions are done after the
text has already been wrapped.
There was a bug in ME_FindText which would cause the final caracter
offset to be incorrect when a paragraph was crossed while matching
characters. The problem was the character offset of the wrong
paragraph was used in the calculation of the start offset of the
match.
The text mode is already stored, and EM_SETTEXTMODE already exists.
There was however a bug in EM_MakeEditor that could cause TM_PLAINTEXT
and TM_RICHEDIT to be set at the same time. This was corrected to ensure
EM_GETTEXTMODE returned the proper mode being used.
The width for EM_SETTARGETDEVICE is used by some applications to set the
wrapping width to a certain distance in twips. This can be used even
though the target device is ignored.
The internal style flags are used to determine whether to show or hide
the scrollbar when ME_UpdateScrollBar is called. EM_SHOWSCROLLBAR seems
to update this state in native richedit controls.
If the scrollbar style isn't initially used, then the scrollbar should
be shown. Otherwise this can be a problem when the horizontal scrollbar
is shown for a single line richedit control, since it will cover all the
text (See bug 12088).
Previously a count of the carraige returns and line feeds were stored
for end of paragraph runs, and a paragraph sign was stored as the actual
string. This was causing many special cases where the length of the
run needed to be determined differently if the run was or wasn't an
end of paragraph run.
There wasn't any use for storing the paragraph sign unless some drawing
code gets commented out to allow the end paragraphs to be shown,
therefore I changed the code to store the actual string that gets
retrieved by WM_GETTEXT.
The two functions ME_FindItemAtOffset and ME_RunOfsFromCharOfs were almost
identically used, since ME_FindItemAtOffset was always used to find a run.
The only difference was how they returned the offset within the run for an
end of paragraph run.
For ME_FindItemAtOffset it would return the next run if it was in between \r
and \n. ME_RunOfsFromCharOfs would instead return an nOffset of 0 for end
paragraph runs. This subtle difference introduced bugs, so I decided to
avoid having special case in this function when creating this patch, and
instead let the caller handle this case.
EM_GETTEXTRANGE allows the start character offset and end characters
offset to be used to specify the range of text to retrieve. If the
start offset is in the middle of an end of paragraph run (i.e. \r\n),
then it should only retrieve the characters after the specified
character offset.
I found that ME_FindItemAtOffset and ME_CursorFromCharOfs are used
almost identically, except for how they handle a character offset that
is between a carriage return and line feed. In this case
ME_CursorFromCharOfs sets the cursor's run offset to 0, but
ME_FindItemAtOffset instead returns the next run which is what was
causing ME_LINELENGTH to incorrectly return the length of the next
line.
riched32.dll does preserve the carriage returns and line feeds unlike
later versions of the richedit control, however the tests previously
missed the fact that a sequence of carriage returns followed by a line
feed (e.g. \r\r\r\n) can actually cause multiple paragraph breaks.
I noticed a while ago that on Windows XP richedit controls ignored
characters typed while the mouse is captured (e.g. from holding the left
or middle button down). Arrow keys, delete, and backspace, copying,
cutting, pasting, and everything else handled on WM_CHAR and WM_KEYDOWN
messages are also ignored.
Certain operations will simply not be done for windowless richedit
controls, such as WM_PAINT which isn't done for windowless richedit
controls since ITextServices provides a TxDraw method.
The methods in ITextHost are mostly thin wrappers around functions that
take a handle to a window as their first parameter. This patch just
uses the wrapper functions provided by ITextHost instead of using the
functions that require a handle to a window that the editor might now
have (for windowless richedit controls).
EM_GETPARAFORMAT previously would overwrite the cbSize field with the
size of PARAFORMAT2, would read past the end of the struct, and might
indicate that PARAFORMAT2 fields are valid using the mask regardless
of the value of cbSize.
If a EM_SETPARAFORMAT message is sent to the richedit control with bits
in the dwMask field that correspond to PARAFORMAT2 fields, then these
fields should be ignored. Instead data was copied from outside of the
structure.
Using the WS_VSCROLL style causes the ES_AUTOVSCROLL option to be set,
and using the WS_HSCROLL style causes the ES_AUTOHSCROLL flag to be
set (except with richedit v1.0).
The contents of the text can be zoomed in with EM_SETZOOM, or with the
mouse wheel. EM_SETZOOM is implemented, but these tests show bugs in
the implementation, and zooming using the mouse wheel isn't
implemented at all yet.
Windowless richedit control will not be able to call GetCapture without
a handle to the host window (and there is no ITextHost_TxGetCapture
method), but there is a ITextHost_TxSetCapture method available for
setting and releasing the capture on the mouse. This means that the
richedit control will need to keep track of whether it has captured the
mouse or not to implement windowless richedit controls.
Previously the WM_NCCREATE was handled by the as if it was always for
later versions, then the window proc for version 1.0 would make
appropriate changes afterwards. Instead both versions should call the
same function (e.g. ME_MakeEditor) and provide the value for
bEmulateVersion10 to make the code clearer.
This fixes inconsistencies shown in the tests I added for the
WM_GETDLGCODE. The tests covered different cases handled by the
current implementation in order to show that the native implementation
is simpler for all these cases.
When the character or paragraph format is changed the paragraph that
is changed is already marked to be rewrapped, so ME_MarkAllForWrapping
shouldn't be called. Since ME_RewrapRepaint uses this function, it
shouldn't be called in these circumstances, since rewrapping all the
text can cause noticable delays when working with a lot of text.
A common case for richedit controls are that a large amount of text is
set initially with word wrap enabled. This causes the initially
wrapping of the text, which also calculates the text length. After
this the vertical scrollbar will be shown, which causes the text to be
rewrapped again. After this there are two redundant rewraps that are
done which this patch eliminates.
On WM_DESTROY the editor was getting freed, then it was used to obtain
a handle to the editor. This patch moves it just before the editor is
freed within ME_DestroyEditor.
In order to make the message handling available to windowless richedit
controls, the message handling must be in a function that can be
called from the ITextServices_TxSendMessage method. This method will
never have a handle to a window to pass to RichEditWndProc_common in
order to get the editor with GetWindowLongPtrW, but passing the editor
will work (even if hWnd is NULL).
When the text is wrapped, the positions for all the runs, paragraphs,
and cells are already calculated and stored. The only thing left to do
for painting is to offset them by the formatting rectangle and the
scroll position.
During wrapping there were three different heights that were being
stored, with only one of them being done correctly. The other ones
failed to incorporate the height of the paragraph or row, so ended up
being incorrect.
The formatting rectangle is set with EM_SETRECT, and retrieved with
EM_GETRECT, so it corresponds to rcFormat in the code. This defines the
area that the richedit control should draw the text so that it is
offset by the top-left corner of the formatting rectangle, and clipped
so that it doesn't draw past the bottom or right hand side. Thus this
is important for implementing windowless richedit controls to not
interfere with the rest of the window.
There were several methods that do not have a HRESULT for a return
value, so returning E_NOTIMPL is not appropriate. For all the BOOL
return values FALSE was returned to indicate the operation was not
performed.
The RichEditWndProc_common function is big enough already by handling
all the window messages, so moving code to handle a message to its own
function makes the code more readable.
There is no reason for the rich text format parser to need a handle to
the window, and even if there were it has a handle to the editor which
contains a handle to the window. It is better to remove this
considering we need to cut down on the use of window handles to
implement windowless richedit controls.
The vertical scrollbar state is stored internally within the control,
so should be used when possible. This will become more necessary when
windowless richedit controls are implemented, and there will no hWnd
to pass to GetScrollInfo.
Comparing the editor as apposed to the handle to the window will work
just as well right now, but will also work when there is no window
handle to make a comparison with, which will be the case with
windowless richedit controls.
The code for the ME_EnsureVisible function does exactly what
EM_SCROLLCARET does, yet this code is duplicated in order to handle
this message. It is simpler to just use the existing function to
implement the message, and avoid internally sending the EM_SCROLLCARET
when this function is available.
EditWordBreakProc documented the third parameter as being the number
of unicode characters in the string for richedit 2.0 and up. It turns
out that it should actually be the number of bytes in the string.
For some reason EM_POSFROMCHAR was returning 0 when the position was
equal to the end of the text, or beyond the end of the text. Instead
it should use the position at the end of the text for both these
cases. The x value was also seen to be offset by 1 according to the
tests.
Previously the shortcuts for cut, copy, paste, undo, redo, and select
all were being handled during the WM_CHAR message. These tests show
that these shortcuts should be handled with the WM_KEYDOWN message
instead.
There were some notifications that weren't sent in ME_UpdateRepaint
while redraw was disabled, so this verifies that they are not sent
with redraw disabled.
The assertion was not valid, because it neglected to take into account
the situation where a line break is forced with a MERF_ENDROW run
(caused by \line control word or pressing Shift-Enter). This means
that spaces can cause a line wrap after a forced line break as well as
after a paragraph break, so we cannot assert that it is the first row
in the paragraph.
The test for EM_GETLINE was testing to make sure the null terminating
character was written at the end of the text as long as the buffer was
long enough, and also tested to make sure that no other bytes were
written after this null terminating character. This is consistent with
Windows 2000 and up, but not for previous versions of Windows.
The table properties are streamed out at the start of the table for
non-nested tables, and at the end of the table for nested tables. The
assertion caught the fact that I didn't get the start of the table row
for nested tables before trying to stream out the properties.
The call to ME_GetTableRowStart will handle both of these cases by
getting the table row start paragraph and asserting that it is found.
This call was also the reason for removing the const qualifier on one
of the parameters.
Previously the paragraph and cell border properties were lost when
deleting the text, then undoing the deletion. This would cause tables
to lose the colour and width of the table border.
Previously the control words in skipped groups were being processed by
the read hook on the RTF parser. By moving this code into the class
callbacks for the parser, the skipped groups actually remain skipped.
Dylan Smith