This document describes the state of text layout in librsvg as of version 2.55.90, and how I want to overhaul it completely for SVG2.
Status as of librsvg 2.55.90
Basic supported features:
Librsvg supports the elements
ainside text, and
tref(deprecated in SVG2, but kept around for SVG1.1 compatibility). See below for the
x/y/dx/dyattributes; librsvg supports single-number values in these.
SVG1.1 values for
writing-mode. Non-LTR or vertical text layout is very much untested.
SVG1.1 values for
direction. This is done for each text
span, but not for the whole
<text>element yet. See below for details.
Major missing features:
glyph-orientation-verticalfallbacks, SVG2 values for
white-spacehandling. This deprecates
Support for multiple values in each of the attributes
tspanelements. Librsvg supports a single value for each attribute, whereas SVG allows for multiple values — these then get used to individually position “typographic characters” (Pango clusters). In effect, librsvg’s single values for each of those attributes mean that each text span can be positioned independently, but not each character.
rotateattribute is not supported. In SVG it also allows multiple values, one for each character.
glyph-orientation-horizontalis deprecated in SVG2).
textPathis not supported at all. This will be made much easier by implementing
x/y/dx/dy/rotationfirst, since each character needs to be positioned and oriented individually.
@font-faceand WOFF fonts.
Emoji got inadvertently broken; see the “Emoji” section below.
Other missing features:
visibilityare not very well tested for the sub-elements of
SVG2 text with a content area / multi-line / wrapped text:
shape-padding. This is lower priority than the features above. Also the related properties
baseline-source, and SVG2 values for
baseline-shift. Note that Pango doesn’t provide baseline information yet.
line-height(parsed, but not processed).
text-decoration, which translates to
CSS Text 3/4 features not mentioned here.
Features that will not be implemented:
SVG1.1 features like
glyph-orientation-horizontalproperty, that were deprecated for SVG2.
Since librsvg 2.52.1 I’ve started to systematically improve text support. Many thanks to Behdad Esfahbod, Khaled Ghetas, Matthias Clasen for their advice and inspiration.
First, I want to get bidi to a state where it is reliable, at least as much as LTR languages with Latin text are reliable right now:
Add tests for the different combinations of
direction; right now there are only a few tested combinations.
Test and implement multiply-nested changes of direction. I think only a single level works right now.
Even if white-space handling remains semi-broken, I think it’s more important to have “mostly working” bidi than completely accurate white-space handling and layout.
Second, actually overhaul librsvg’s text engine by implementing the SVG2 text layout algorithm:
text-orientationproperty, and implement fallbacks from the deprecated
glyph-orientation-verticalto it. If this turns out to be hard with the current state of the code, I will defer it until the SVG2 text layout algorithm below.
Implement the SVG2 text layout algorithm and
white-spacehandling at the same time. See the detailed roadmap below.
Third, implement all the properties that are not critical for the text
layout algorithm, and things like
@font-face. Those can be done
gradually, but I feel the text layout algorithm has to be done all in a
A common theme over the next subsections is, “we need a single
<text> element”. Keep that in mind as the
main goal of initial refactoring.
Chunks and spans
Librsvg implements a limited subset the text layout as per SVG1.1, which was feasible to implement in terms of chunks and spans.
A span is an
<tspan> element, or some character content inside
tspan explicitly lists
y attributes, it
creates a new chunk. A text chunk defines an absolutely-positioned
sequence of spans.
This is why you’ll see that the code does this; start at
Start with an empty list of chunks (Text::make_chunks). Push an empty initial chunk defined by the
ycoordinates of the
Recursively call children_to_chunks on the children of the
<text>element, to create chunks and spans for them.
TSpan::to_chunks sees if the span has
yattributes; if so, it pushes a new empty chunk with those coordinates. Then it recursively calls
children_to_chunksto grab its character content and children.
Text::drawtakes the list of chunks and their spans, and converts them into a list of
MeasuredChunk`. This process turns each span into a ``MeasuredSpan. The key element here is to create a
pango::Layoutfor each span, and ask it for its size.
Text::drawtakes the list of
MeasuredChunkand turns them into a list of
PositionedChunk. Each of those builds a list of
PositionedSpanbased on the span’s own text advance, plus the span’s
- Note about SVG2: The `text layout algorithm for SVG2
<https://www.w3.org/TR/SVG2/text.html#TextLayoutAlgorithm>`_ is very different from the above. It mostly dispenses with explicit computation of chunks with spans, and instead, for each glyph it stores a flag that says whether the glyph is at the beginning of a chunk.
Layouts and spans
Librsvg creates a
pango::Layout for each
text span in a
<text> element, whether it comes from a
or not. For example,
<text>A <tspan>B</tspan> C</text> has three
spans, and three Pango layouts created for it. Each span’s
pango::Layout gets configured via
pango::AttrList with the
styles it needs (bold/italic, font size, etc.).
pango::AttrList gets created, each individual attribute has
a start/end index based on the byte offsets for the corresponding
characters. Currently, all the attributes for a span occupy the whole text span. So, for something like
<text> Hello <tspan font-weight="bold"> BOLD </tspan> World </text>
pango::Layout objects get created, with
World, and the second one has a
pango::AttrList that spans
its entire 4 bytes. (There’s probably some whitespace in the span,
and the attribute list would include it — I’m saying “4” since it is easy
to visualize for example purposes.)
However, this is sub-optimal. Ideally there should be a single
pango::Layout for a single string,
Hello BOLD World, and the
attribute list should have a boldface attribute just for the word in
Why? Two reasons: shaping needs to happen across spans (it doesn’t
right now), and the handling for
need to be able to work across nested spans (they work with a single
level of nesting right now). Read the “Bidi handling” section below
for more info.
add_pango_attributes function is already able to handle
substrings of a
pango::Layout; it’s just that it is always called
with the whole layout right now.
The initial refactoring: Change the text handling code to first
gather all the character content inside a
<text> into a single
string, while keeping track of the offsets of each span. Make the
pango::AttrList taking those offsets into account. Then, feed
that single string to a
pango::Layout, with the attributes. Due
to the current code’s use of
MeasuredChunk, etc., it
may be better to create a
pango::Layout for each chunk, instead of
<text> (i.e. one layout for each absolutely-positioned
sequence of spans). The SVG2 text layout algorithm will compute
chunks completely differently, but it will still require per-span
offsets and cross-span shaping.
Further work: Don’t just paint the layout, but iterate it / break
it up into individual
pango::GlyphString, so librsvg can lay out
each individual glyph itself using the SVG2 layout algorithm.
Be careful with PDF output when handling individual glyphs: grep for
direction properties get handled
together. The BidiControl
struct computes which Unicode control characters need to be inserted
at the start and end of a
<tspan>’s text; SVG authors use these
properties to override text direction when inserting LTR or RTL text
within each other.
Unfortunately, these control characters can only really work for nested levels of embedding if the whole text is in a single ``pango::Layout``. Per the previous section, librsvg doesn’t do this yet.
implemented the SVG2 values for the
unicode-bidi property. You
may want to read the detailed commit messages there, and the
discussion in the merge request, to see details of future development.
Add tests for combinations of
These are easy to add now that librsvg’s tests make use of the Ahem font, in which each glyph is a 1x1 em square.
This may just be the property parser and hooking it up to the machinery for properties. Actual processing may be easier to do in the SVG2 text layout algorithm, detailed below.
Implement the SVG2 text layout algorithm and
Shaping: One thing librsvg does wrong is that for each
or for each synthesized text span from a
<text> element, it creates
pango::Layout. This means that text shaping is not done
across element boundaries (SVG2 requirement). Implementing this can be
done by creating a string by recursively concatenating the character
content of each
<text> element and its children, and adding
pango::Attributes with the proper indexes based on each child’s
character length. This creates an un-shaped string in logical order with
all the characters inside the
<text>, to be used in the next steps.
Pango details: create a single
pango::Attribute for each text span. Set the layout to
set_single_paragraph_mode() so it does not break newlines. Pango
will then translate them to characters in the
Layout, and the
white-space handling and SVG2 text layout algorithm below can detect
White-space handling: SVG2 has a new
white-space property that
xml:space from SVG1.1. Implementing this depends on the
concatenated string from the steps above, so that white-space can be
collapsed on the result. Maybe this needs to be done before inserting
bidi control characters, or maybe not, if the state machine is adjusted
to ignore the control characters.
SVG2 text layout algorithm: This is the big one. The spec has
pseudocode. It depends on the shaping results from Pango, and involves
correlating “typographic characters” (Pango clusters) with the un-shaped
string in logical order from the “Shaping”, and the information about
discarded white-space characters. The complete text layout algorithm
would take care of supporting multi-valued
textPath (see below), plus bidi and vertical text.
Do look at the issues in the svgwg repository at GitHub - there are a couple that mention bugs in the spec’s pseudocode for the text layout algorithm.
Librsvg is moving towards a “render tree” or “display list” model, instead of just rendering everything directly while traversing the DOM tree.
Currently, the text layout process generates a
which is basically an array of
pango::Layout with extra information.
It should be possible to explode these into
pango::GlyphString and annotate these with
information, which will be the actual results of the SVG2 text layout
Although currently Pango deals with underlining, it may be necessary to
do that in librsvg instead - I am not sure yet how
x/y/dx/dy/rotate interact with underlining.
/** * pango_renderer_draw_glyph_item: * @renderer: a `PangoRenderer` * @text: (nullable): the UTF-8 text that @glyph_item refers to * @glyph_item: a `PangoGlyphItem` * @x: X position of left edge of baseline, in user space coordinates * in Pango units * @y: Y position of left edge of baseline, in user space coordinates * in Pango units * * Draws the glyphs in @glyph_item with the specified `PangoRenderer`, * embedding the text associated with the glyphs in the output if the * output format supports it. * * This is useful for rendering text in PDF. * ... */
Note that embedding text in PDF to make it selectable involves passing a
text to pango_renderer_draw_glyph_item(). We’ll have to
implement this by hand, probably.
Wrapped text in a content area
This roadmap does not consider the implementation fo wrapped text yet.
@font-face and WOFF
This involves changes to the CSS machinery, to parse the
at-rule. Librsvg would also have to obtain the font and feed it to
FontConfig. I am not sure if FontConfig can deal with WOFF just like
See the issue on the Future of the pango dependency for lots of goodies which may come in handy.
Emoji is broken
#599 is a terrible bug in Pango, which causes it to report incorrect metrics when text is scaled non-proportionally (e.g. different scale factors for the X/Y dimensions). Librsvg works around this by converting all text to Bézier paths, then scaling the paths, and then stroking/filling them.
However, this breaks emoji - #911, since converting its glyphs to paths loses the color information.
Two strategies to fix this; there may be more:
Detect if the text is scaled proportionally (this is the common case), and use the old code for that, without converting text to paths. This may be easy to do? Grep for
drawing_ctx.rswhich already has some of the logic but for handling PDF output.
Do the whole “split a
pango::Layoutinto glyphs” from above; keep handling individual glyphs as paths, and special-case emoji to render them via Cairo.
https://gitlab.gnome.org/GNOME/librsvg/-/issues/795 - Implement SVG2 white-space behavior.
Issues that have not been filed yet
From the spec: “It is possible to apply a gradient, pattern, clipping
path, mask or filter to text.” We need better tests for the
objectBoundingBox of the whole
<text>; I think they are wrong for
and this shows up when filling its spans with gradients or
Clip/mask/filter do not work on individual spans yet. I am not sure
if their objectBoundingBox refers to the whole
<text> or just
Multiply-nested changes of text direction / bidi overrides; see the “Bidi handling” section above.
Glossary so I don’t have to check the Pango docs every time
PangoItem - A range within the user’s string that has the same language/script/direction/level/etc. (Logical order).
PangoLayoutRun - same as PangoGlyphItem - a pair of PangoItem and the PangoGlyphString it generated during shaping. (Visual order).
PangoGlyphString - The glyphs generated for a single PangoItem.
PangoGravityHint - Defines how horizontal scripts should behave in a vertical context.