GitLab

Calling HtmlEndParse at the end of html_get_next_token and setting up->context to NULL was wrong; fetch_fetcher would end up recalling html_get_next_token later which would then call HtmlParse even though the document parse had just finished! The HtmlEndParse call has now been moved to fetch_stop.

Altered reformatter to claim fonts inside a browser, rather than
claiming them for nothing and losing them afterwards.

Got rid of TAG == TABLE and ISBODY tests, replacing with tagno == TAG_TABLE.
Removed all references to TD, TH, TR and TABLE in the style word.

Big steps forward in vertical white space handling as a result of improvements in HTMLLib in this area ( -> all versions now 1.22 beta-2).

As a result, the 'last_space' field in browser_data struct has been
removed. Note that this relies quite heavily on setting of the PCDATA
bit in the 'style' field of an HStream and the automatic collapsing
of (for example) multiple P tags inside HTMLLib.

Trace.c improved to recognise various bits in the 'style' field of an
HStream structure.

Phoenix defaults altered to a more 'standard' set of choices; Trinity
as the serif font, with a slightly larger default font size. This is
because there's a good chance it might get released to a wider audience
than Acorn internal (though the animation and icon bar sprites will
have to change before then...).

MiscDefs updated for new SWI numbers in HTTP module; !Run[D] files thus
updated to require HTTP 0.42 or later. At this point, all earlier
modules are not backwards compatible in terms of direct calls to the
HTTP module, though this only affects cookies_process_cookie at present.
At the same time, checks for System$Path, InetDBase$Path, and setting
of Inet$MimeMappings if not already defined have been added to the Run
files along with RMEnsures of Resolver and MimeMap.

Object and PlugIn c/h pairs created to handle OBJECT, EMBED and APPLET,
and the RISC OS Plug-In interface respectively. Not part of the build
process yet. Addition of 'odata' field in browser_data struct and
definition of chunk CK_OBJB for memory_set_chunk_size() are in support
of this.

Fixed problem that affected some internal URLs, most noticably loading external images, where up->fetching was set to zero when it should have been left at 1 (see fetch shutdown code at about line 1670 in FetchHTML.c).

*This* is the source that will be sent to PSI, not the previous release as
stated in the log, though Customer were sent a browser based on this
previous code.

!Run[D] files updated to require FTP 0.14, and ensure that the Resolver and MimeMap modules are present. Paragraph tags can now affect images and other such objects - previously they only worked on text (this was noticed for a DIV element containing one image, which was preceeded by a P element but didn't have any white space above it).

This is the source that will be released to PSI, and builds the Customer
!Customer and internal !Phoenix releases at about 5:00pm on 10/03/97.

Redraw routines now allow tables to be indented with body text. Some page
layouts rely on this.

Dropping on the lower half of an open, empty directory will now drop
inside it rather than after it.

Updated Res files in appropriate builds to hold various (similar) Choices
designs.

Choices related menus were flagged as Shared, but none of the dialogues
(including ColourDBox) were - potential future problems, though shouldn't
cause any leaks at present. This has been sorted out anyway.

Encoding function encoding_init no longer returns an error from
toolbox_create_object, so the Encoding menu and all those attached
to it do not have to be present (e.g. the Customer build).

SUB, SUP, STRIKE and U supported. U underlines the baseline of the body text
font, whilst STRIKE will go through roughly the middle of the lower case
chars even if the font is SUP or SUB. Note that Navigator appears to shift
the underline point for SUB and SUP; it may be necessary to copy this
behaviour, but testing on real sites must proceed before that. There could
also be a problem with the automatic lowering of font size, which Navigator
doesn't do, so any FONT SIZE = -n commands could make it too small. Again,
this needs testing on real sites.

'http://' is added to URLs with no protocol specified, unless they start with
'ftp.', in which case the new behaviour is to add 'ftp://'.

Choices code altered to do less error checking on components! They should be
able to be missing without raising errors. Referencing of the subwindow
array changed from *(subwindows + number) to subwindows[number].

Made trace_tag_name code look pretty...

reformat_useless_token now checks tagno is non-zero.

User Agent string setting now done through URL_GetURL, on a per-session
basis.

Ellipsis character removed from all Messages files, replaced with '...'.
There's little difference between the two in an outline font, and in System
font the latter looks much better. Smart quotes left in, as they look
better in all cases.

Didn't redraw browsers which would be immediately affected by a change to the Choices when 'Set' was used. Appropriate redraw calls added. Reversed decision to toe in the pane window by 8 OS units relative to the placeholder icon sides.

Changed some error handling in Choices, a little bit of reformatting, removed unnecessary tempstring in Main, and played about with the choices Res file objects.

Added run time RMEnsuring type behaviour for ColourDbox to save RMA space on start up of the browser.

Also added additional deregistering of events which I forgot to do last time :(

Renamed Symbols to syms and ensured it is CDir'ed.
Made Trace.c show tag_no field textually.
Ungreyed CJK entries in Encoding menu.

Altered hotlist_add to create the new url after the first selected item or inside the first selected directory when an item is selected.

Text items with no VALUE specified in the HTML were not cleared on
hitting a Reset button as the text field of the token is NULL; fixed.
SELECT lists with no default selection item were not being changed
on hitting a Reset button; fixed. In radio groups with no default
selection specified, no item will be selected. This goes against
the HTML 2 spec but allows broken Navigator-esque forms behaviour.
In consequence, radios can be deselected by clicking on the same
one twice. I'd fixed the flickering experienced when doing this
recently, but that, now, is irrelevant, since the state of a radio
must always change.

Reformatter's new 'find width of a SELECT field' didn't account for
the width of <none> and <many> items, and now does (particularly
important for broken items with no OPTION contents).

Under certain circumstances, illegally named targets would open in the
ancestor rather than a new window - a deliberate decision in the code,
but now reversed due to a test suite failure; they'll open in a new
window instead.

Altered the left hand indent handling for lists so that headings and
body text follow the Navigator 48 pixel indent and the bullet point
items drop to the left of the left hand margin, rather than sitting
on the margin with the text indented to the right.

Tightened up the hotlist_load_directory checking of URLs, so that
only those with '://' in - i.e. look fully specified - are accepted.
HTML files are loaded into a new directory now.

Items without a descriptive title in hotlist_new_url will have the URL
substituted in instead. White space before and after descriptions and
after URLs is stripped. Directories with zero length names or null
strings (after white space is stripped) will still be created, with a
generic name (see Messages, token 'HotlistUntitled').

StrongHelp generator on it. Revealed a spurious toolbox_hide_object call
that would always fault and never do anything, and a few heinous uses
of show_error which have all been removed.

When dragging hotlist items, directories will now be highlighted whenever
the item would be added to them, rather than just over the sprite. In
practice this only changes if the item is open and has contents...

Drag box selections had broken in the hotlist; fixed (uninitialised
variable in hotlist_select_box). Adjust-drags now close the window on
completion. Minor problem with Shift+Drag (copy) reversing order of items
sorted out. Added hotlist_add_html_file and implemented loading of HTML.
DO NOT add broken HTML files, this has not been coded for and URLs of the
form http:///this/that (no host name) will hang the machine (bug in the
cookie code of the HTTP module up to at least version 0.39).

Made SELECT list items only as wide as the widest entry actually appears
(using fm_get_string_width), rather than as wide as the widest character
BBox in the font multiplied by the string length of the longest entry.

Updated !Run[D] files to require HTTP 0.39, FTP 0.12 and File 0.23.

Forgot to mention the important bit - Choices and Controls are now *not* loaded into a malloc block, and chf/cof have been removed. This is because the length of the message file is indeterminate when loaded by MessageTrans; precisely what MessageTrans does with it on loading is unspecified in the

PRMs. MessageTrans_FileInfo could imply that the size required for the file
may not be the same as the actual file size on disc... so the save choices
routine now loads a copy as it needs it.

New save choices function with supporting code in Save.c; 'Save choices'
option in the Choices menu is now implemented.

Choices and Controls files are now loaded into malloc space (use chf and cof to address them). Previously only in Main.c, task_dir is now globally visible.

Choices updated to allow 16ths of a point in the font size specification
and the System font face has been removed - it isn't used. The use of
serif and sans serif fonts has been rationalised across builds (e.g. Desktop
builds use what they say, but the Customer build uses Homerton as the
default body (serif, so to speak) font as it matches the box's look and
feel better).

Font loading altered a bit to work properly! Font library defines some
standard startup fonts without using Choices, which Choices can then
override.

Now working on source merged with Kevin Bracey's internationalisation support. UNIFONT is undefined in the Make File for now. All Res and

Choices files updated appropriately.

Having sorted out the old Choices and Messages to form Choices, Controls
and Messages, this build has had the same cleaning up done internally.
This includes greater consistency in naming schemes and the removal of
the inconsitent choices items - e.g. Choices file entries saying 'delay
images' and 'plain backgrounds' where internally all the flags say 'show
images' and 'show backgrounds'. ChoiceDefs.h and CtrlDefs.h added to
clarify the meaning of some fields, though usage of these is not 100%
in the source (there are cases where parameters are passed through to
functions as ints, and those functions still check these against hard
coded values rather than the #define stuff).

Fetcher status return bits (connected, sent request, etc.) now reflected
in status bar. Progress during fetchs to files are reported by %, where
the size of the object is known. Exceeding 100% drops back to a byte
counter, in case the estimated size was wrong. The progress counter
may be updated after specific delays, rather than 'as often as possible',
to reduce flicker (as requested by D.Brown some time ago).

I've done a small rewrite of the fetch prioritisation scheme in FetchPage.c;
how well this performs in general use across different processor speeds
remains to be tested, but certainly it has some advantages. For each small
fetch window before the rewrite, a 4cs tight loop was entered - this gave a
noticable and substantial drain to the Desktop performance if more than one
was opened. Now, several can be up at once with little hit. The actual file
fetch is on half the priority it was before, with all others taken back
just a bit - e.g. from 20cs per poll to 15cs per poll for flat out
reformatting. You don't seem to lose much time on the format in practice,
and the Desktop feels quite a bit lighter at the same time. There's the
potential for smoother frameset loading in this scheme, too.

When Shift+Clicking on a link meant you still fetched inside the main
browser window, several fetches could occur in a frameset - one per frame.
However, now that you can only do this by clicking on a link that leads to
non-displayable data - or by turning off the small fetch windows by
setting UseSmall to 'no' in Choices - a bug where fetchpage_preprocessed
would stop such fetches as new ones were started was revealed.
The API to frames_abort_fetching has now been extended to include a
'stop file spooling too' flag, allowing a fix to be made by having
fetchpage_preprocess's calls not set this (and it doesn't check the
savelink flag is unset before proceeding, since frames_abort_fetching
does that implicitly now).

Had left the RAM transfer buffer at 16 bytes (from testing) accidentally...
Oops. Upped it to 4K. In addition, when loading data by RAM transfer,
the browser didn't notice if a RAMFetch bounced during the transfer. It
would be treated as a 'first' RAMFetch bounce, basically, and try to go to
file transfer - oops. Fixed.

Also much faster drag box selections.

Encoding support added: if UNIFONT is defined then the browser will attempt to use the system font to render Unicode characters (new

International module required). If not defined, the browser will just
handle Acorn Latin-1 characters, but HTMLLib will convert other encodings
into Latin-1. Must keep the UNIFONT define in sync with HTMLLib (if
defined HTMLLib outputs UTF-8, otherwise Acorn Latin-1).

Still needed:

Reverse encoding (converting form input etc back into the server's
character set).
Reparse when encoding changes.
Setting of charset from an attribute on the <A> tag.

Altered Newdirectory/NewURL routines to create new directories at the pointer.

Image deletion routines rewritten; nasty animated GIF bug now gone
(it kept trying to update the cross-referenced image in the browser that
no longer existed, and I've no idea how animated GIFs ever managed to *not*
blow up if a cross referencing browser was shut down; and I've tested
this...). There are now as many image structures as there are total images
in the visible browsers, and no more; cross referenced images in the
to-be-deleted browser have their ownership moved to the cross
referencer before the cross referencee is wiped, with any required fetches
being transferred with them.

Reformatter sped up a bit by getting rid of strlen on the main text field
of tokens in all places bar one, where it is called extremely rarely. This
means direct app-to-app transfer of, say, a 359K text file to the browser
won't run like a drain as it tries to do strlen on a 359K string over and
over again! Most of the time the performance benefits won't really be
noticable, but on the above text file, formatting time went down from
290 seconds before the change to just under 7 seconds, a 41 fold speed
increase or thereabouts (SA110 228MHz).

MakeFile changed to force function name compilation on for debug builds
(-fn in the DD...FLAGS); useful for certain functions such as
register_null_handler, which can now output the name of the function
being registered.

...Which helped show up some nasties in the animation handler and animation
drift handler registration/deregistration process, which have been fixed
(callers of fetchpage_release_nulls were unaware that the drift handler
could be installed, and fragments of old code checking choices.anim_drift
had a value other than 1 or 0 were still hanging around - bit tricky for
a single bit item).

Added support for VALIGN in table cells. Spotted a bug or two in the
'what token is the pointer over' routines as a result, and fixed them - the
worst was in browser_line_at_y, which checked the y coordinate was below
the given one [the mouse], but didn't check the line height to see if the
given y coordinate was *within* the line, rather than just above it. Never
used to matter pre-tables, but a definite concern once multiple line
arrays can exist on one page.

Added 'save current location'. Res files updated appropriately. Version number taken to Beta-3; next will be 1.21, including K.Bracey's i18n code.

Got the deferred reformatter working properly. It doesn't do that 'OK, the page is fetched and reformatted, but just to annoy you, I'm going to wait 5 seconds and then suddenly reformat the whole thing again' trick anymore. The fetcher was calling the reformatter in a delayed form even when the reformatter was already running, so it would carry on past the reformat point or from below what had become and invalidated line, and some time later, get back to the delayed reformat. Now, reformatting is only delayed by the fetcher when the reformatter is not running. In practice this means body text reformats as it fetches, but large tables will show delayed reformatting - which was exactly the intended behaviour of the feature when it was originally thought of.

fetch_token_data_address removed; it was only needed in two places,
both of which already knew when to read tp->text and when to ignore
it. Its functionality is duplicated in an 'if' involving reformat_istext,
anyway.

!Run[D] files taken back to requiring HTTP 0.33, since 0.36 introduces
many weird and wonderful problems.

Typo in Messages files, 'All current images (sic.) fetches finished' -
'images' is now 'image'.

Shift+Click saving - you could save to an application. No problem. But
the equivalent (just clicking on a link that led to an unknown datatype
and getting the save dialogue that way) didn't work. It does now.
Another problem was saving to an application that didn't support the
datatype - oops, the dialogue would close but the fetch would sit there
waiting to be told where to save. It doesn't close now (as expected).

NB, doing several simultaneous fetches to a text editor may have problems
as the editors are too clever for their own good. Despite receiving
DataLoad messages for <Wimp$Scrap> for files of different types, sizes
and datestamps, the editors can decide it's still the same file and:

Zap      - Hats off, it gets it right, almost. You do get warned 'Multiple
           copies - one on disc is newer' as everything after the first
           text loads, but they do load, and in separate windows.
StrongED - Does not load the subsequent files, so the browser gives 'Data
           transfer failed' errors and opens up Scrap. Turning off 'Don't
           load same file twice' fixes it - each file is loaded in a new
           window with no warnings. At least in the first case, you don't
           lose data, since the files are kept in Scrap.
Edit     - Each time it loads the file, it *replaces* the other one in
           memory, using the same window for each. This is the worst
           behaviour as it isn't configurable (well, I don't know of a way
           to change it...) and results in data loss as successive texts
           get trounced by the new data.

I can't see how I can fix this in the browser as it's basically silly
behaviour on behalf of the editors. Other applications which don't try
to work out if it's a new file or not are fine!

When conducting image fetches, proxying is allowed unless reloading.
When conducting page fetches, proxying was never allowed - so web cache
stuff would have been, er, interesting. It now sets X-NoProxy: in the
request header when reloading, but otherwise this is not included.

AnimSpeed is, at last, independent of browser poll speed. They used to
be tied together. Guess how the animation code used to work ;-)

'Can't handle this datatype' - deprecated now that save dialogues can be
popped up. The 'can't save objects in full screen mode' error would never
be shown due to a bug, anyway; this now replaces 'can't handle', which
has been removed from all Messages files.

RefoWait, RefoHang and RefoTime moved from Controls back to Choices.

Trying to get rid of strlen in the reformatter - it can get very slow
(e.g. strlen of 8K chunks of text, or if a 330K text file is transferred
from a text editor straight to the browser, strlen of a 330K string...).
There will be unfinished bits of code in the reformatter that may seem
unnecessary - they've just not been plugged in yet (since they don't
actually work). Don't remove them!

As warned in the last log, pretty much all event codes and component IDs have now changed along with many of the names, to provide a consistent name and numberspace for events and components. This also minimises number clashes (e.g. as was, the Save File origin when opened from a

Hotlist menu with an already-used component ID). To get full details on
this, please carefully read through TBEvents.h.

Res files, Sprites files and Messages files have been updated again both
due to the above, and routine additions (e.g. message support for a few
Hotlist bits and pieces). !Run[D] files updated for FTP 0.11 and
HTTP 0.36.

In the Makefile, the Customer objects list was missing Save - must have
moved something when I should've copied it, when adding in SaveFile or
SaveObject in all probability. Now fixed. Couple of other bits and bobs
fixed in the build environment (e.g. stuff saying !Argo instead of
!Customer). Oh, and I've altered the MakeFile copy options again to the
best compromise I can come up with. Newer is turned on for everything
except !Run[D] and !RunImage, since both of those change between debug
builds - otherwise if you'd built debug and non-debug versions, it was
not possible to switch between them - one version would have the newer
timestamp and thus never get overwritten.

The default hotlist has had a few items added - that'll be about the end
of it, I think; there's more than enough stuff in there now.

Saving of the hotlist from the document menu and of URI files,
directories and selections from the hotlist menu tree is now implemented
- this new save system rocks... Oh, and you can save all images and
backgrounds as sprites.

Saving of items with Shift+Click to other applications directly now
works, and is robust. Unique Scrap filenames are used, with data load
bounces (e.g. if some pra - er, person quits the app they're sending
to) working correctly - that is, give an error, keep the file, rename
it to something safe, and open the directory it lies in. As opposed
to normal app-to-app bounces, where the scrap file is deleted (see
PRM 3-254). This means you can now look at README files in FTP
sites, say, without using a disc intermediate. Or you can send pictures
straight to ChangeFSI, fetch HTML links into editors, and so-on, and
so-on - it's all very funky. Known problems include the ambiguous
'invalid component ID' instead of 'file open' for *normal* (straight to
disc) shift+click saves, and I think I'll introduce a unique name
guarantee of some sort to stop 'file open' in the first place. Odd
that the really tricky part (app-to-app) should be least likely to
suffer from this!

Only the Browse resources are currently valid.

Added Utils.Icons - has a few archives inside containing the resources
(well, some of them) used to build various UI sprites for various builds.
Archived because these are unlikely to change much, and putting them on
CVS was a move to, well, archive the stuff...

SaveDBox objects vanquished and requirements in !Run[D] files removed. The
data save code fits much more neatly in amongst the data load protocol
stuff now (with the slight exception of having to split the SaveObject
source into SaveObject and SaveFile - the former handles multiple persistent
dialogues for Shift+Click on links and the like, the latter handles 'one at
a time' transient dialogues for save source and similar). Export Link is now
supported, too, and writes a 'proper' version URI file. You'll find that
double-clicking on old URI files will work as the URI handler picks them up,
whilst new version ones don't; however, dragging onto the browser will only
work with new version files. Note that support for saving and loading URL
files (ANT suite stuff) is present too, so old URI files can be typed as URL
files if you want to keep them working without modification - the URI
handler itself will hopefully support the defined URI file format soon;
double-clicking on old URI files will stop working at that point. Note
there are *lots* of changes in every Res file to support all this. This may
all seem a bit pointless to some, but the changes do in fact make it very
easy to add new save dialogues all over the place. Certainly much easier
than with the previous system, anyway. In fact, post script, image
'save as sprite' took about half an hour, which I hope proves the worth
of the new system.

Merged in newer hotlist code with support for drag cancelling with Escape
(all relevant Res files appropriately updated) and cancelling scrolling
when you've reached the window scroll limit. Had to move some of the
Wimp message handling stuff to the central Protocols source, as clashes
were occuring, and also the hotlist routines were using independent saving
code - a lot of duplicated effort. This was fair enough as at the time the
Hotlist code was written, the Save code couldn't be used in the way it is
now.

New Save Source and Print buttons on the toolbar of some builds.

Phoenix Sprites file made more efficient - the Acorn base section has been
split from the animated upper region. Browse build has a new grey fade
sprite at the back, which is less grainy than the previous one and only
uses 16 colours (with a 16 greyscale palette).

Not really a bug, bug the routine to start an image fetch for INPUT
TYPE=IMAGE forms items only did so if the src field (or equivalent, for
this tag type) was non-NULL. In fact, you should always call image_new_image
and let that handle the rest, otherwise other sections of the code will fail
as they try to obtain an image number for a given HStream and get -1 back.
This problem only generally manifested itself when loading an HTML file to
the browser straight from an application, as many src fields become NULL
when the relativisation routines find nothing to relativise to...

Authentication got broken somewhere along the line - this has been fixed
(in HTMLLib and the browser).

Ctrl+Click on a cross referenced image updates *all* copies, not just the
one with the image data attached.

Next big step: Rip up TBEvents.h and rebuild that whole approach somewhat.
To all those working on the code, my apologies but this means all Res files
will receive a very large number of alterations and there will be extensive
code changes too (mostly naming convention stuff), in more or less all
source files. I am endeavouring to ensure that the new numberspace
convention does not clash with the work being done by Kevin on
internationalisation.

File fetches now set DEADDEAD during the fetch, Data if it is aborted, or an appropriate filetype when finished. If the fetcher routines return a zero or data filetype, the browser looks at a set of hard coded filename extensions to try and determine if there's a better filetype to use

(urlutils_filetype_from_url) - since this is a centralised routine it can
use a Mime mapping system (or be removed entirely) as and when one becomes
available.

You can now drag URI / URL files to the Hotlist window to add them to the
list. They are added roughly where dragged to. (NB, note that I renamed
the function to return the window ID - just being picky; it matches other
similar functions now). Oh, and hotlist_add_position actually works now ;-)

The global history will save when titles are added, as well as when a URL
is added. Before, a browser crash could mean a title got dropped out
of the history file even if SaveHistory was set to 'always'.

SaveObject sources have been added but nothing references them or links
them in yet - they're not finished. Mostly checking this one in because
of the hotlist changes.