
    Copyright (C) 2001-2023 Artifex Software, Inc.
    All Rights Reserved.

    This software is provided AS-IS with no warranty, either express or
   implied.

    This software is distributed under license and may not be copied,
    modified or distributed except as expressly authorized under the terms
    of the license contained in the file LICENSE in this distribution.

    Refer to licensing information at http://www.artifex.com or contact
    Artifex Software, Inc.,  39 Mesa Street, Suite 108A, San Francisco,
    CA 94129, USA, for further information.

This document presents the results of Artifex's investigation of the PCL XL
Feature Reference Protocol Class 1.1 specification and of its relationship
to the implementation in the H-P LaserJet 6MP printer.

		Report on PCL XL and LaserJet 6MP

Introduction
============

In order to implement a fully test-suite-compliant PCL XL interpreter, we
had to compare the output of our own code with the output printed by an H-P
printer for the same input data.  These comparisons uncovered a number of
disagreements between the specification and the implementation in the
printer we used (the LJ 6MP), as well as omission of many details necessary
to create compatible implementations.  In some cases, it seemed likely to us
that the published specification did not properly describe the intended
behavior; in other cases, it seemed likely that the observed behavior was
the result of a printer firmware bug.  Of course, only H-P can answer
authoritatively the question of which alternative is correct.

In presenting test cases below, we have used an imagined source syntax for
PCL XL.  This code is meant to be executed in an environment where one user
space unit is 1/300" and where all elements of the graphics state have their
default values.

This document only addresses a very few of the many typographical and
grammatical errors and internal inconsistencies in the published PCL XL
specification.  In most cases, we found the intention clear.

Changes in a given revision of this document are marked with the revision
number in [brackets].  Revision history:
		first issued December 6, 1996
		rev. [1] December 12, 1996
		rev. [2] December 13, 1996
		rev. [3] December 19, 1996
		rev. [4] December 31, 1996
		rev. [5] January 7, 1997
		rev. [6] January 17, 1997
		rev. [7] February 6, 1997

Miscellaneous
=============

[7] Embedded data streams
-------------------------

H-P printers require that if a command reads data from the data source, then
all the data required by the command must follow the command in a *single*
data block; it cannot be divided up into multiple blocks whose total length
is the amount of data required.

Clipping
--------

The overall discussion of clipping mode and clipping region is inconsistent
in many places both with itself and with the implementation.  Based on
experiments with the printer, we believe the following is the intended
behavior:

	- The ClipMode in the graphics state only controls which rule is
used to determine the inside of the region defined by the *newly presented
path*.  The inside of that region is then intersected with the existing
clipping region (which is simply an abstract region of the plane,
independent of being defined by a particular path) to define the new
clipping region, again as an abstract region of the plane.

	- The ClipRegion parameter of the SetClip operators only controls
whether the interior or exterior of the region defined by the *newly
presented path* should be used to intersect with the existing clipping
region.  The result of the operator is an abstract region of the plane.

This interpretation is consistent with the PostScript clipping model and is
much simpler to understand (and implement) than what the specification
attempts to describe.  Unfortunately, the sample code to illustrate these
conclusions is too large to present here.

On top of all this, there appears to be a firmware bug in the implementation
of SetClipIntersect.  See below.

[5] Line joins
--------------

Unlike PostScript, which applies a line join at the ends of every line
segment (including the segments produced by flattening curves), PCL XL [7]
does not apply the line join within an individual arc or Bezier (including
curves that are part of TrueType characters).  This produces a "smooth"
rather than a "bristly" effect when a null line join is selected.

Session operators
=================

[3] BeginPage
-------------

Apparently illegal values for Orientation, like MediaSize and MediaSource,
only produce warnings, not errors.

Font control operators
======================

[7] RemoveFont
--------------

[2] If RemoveFont provokes multiple warnings within a single page,
apparently only the last occurrence of each warning is remembered.  Test
case:

	(Bogus1) ba @FontName RemoveFont
	(Bogus2) ba @FontName RemoveFont
	(Arial           ) ba @FontName RemoveFont
	(Bogus3) ba @FontName RemoveFont
	(CG Times        ) ba @FontName RemoveFont

Only Bogus3 (UndefinedFontNotRemoved) and CG Times (InternalFontNotRemoved)
will appear on the error page.

Graphics state operators
========================

Miscellaneous
-------------

The specification fails to state (section 5.0) that CharBoldValue and
CharSubMode are elements of the graphics state, and that their default
values are 0 and eNoSubstitution respectively.

SetColorSpace
-------------

The specification says that this operator sets both the pen and brush to
paint black.  In fact, [3] the operator does no such thing.  Test case:

	20 us @PenWidth SetPenWidth
	4 4 usp @PageScale SetPageScale

	1 b @ColorSpace SetColorSpace		% eGray
	0.5 r @GrayLevel SetPenSource
	0.8 r @GrayLevel SetBrushSource
	100 100 200 200 usq @BoundingBox Rectangle

	2 b @ColorSpace SetColorSpace		% eRGB
	300 100 400 200 usq @BoundingBox Rectangle

The two rectangles are identical.  We think this is probably an error in the
specification, since the observed behavior seems reasonable.

SetBrushSource, SetPenSource
----------------------------

The statement "The paint source identified in the attribute list is
compatible with the current color space" is misleading, since it is
acceptable to set the brush source to a raster pattern defined in a color
space different from the current one.

[6] SetCharAngle, SetCharScale, SetCharShear
--------------------------------------------

The statement that these are "not cumulative" is somewhat misleading.  What
the H-P printers apparently do is remember only the most recent invocation
of these commands, including the order in which they were received.  Thus,
for example, the sequences <Angle1, Scale, Angle2> and <Scale, Angle1,
Angle2> are both equivalent to <Scale, Angle2>, but <Angle1, Angle2, Scale>
is equivalent to <Angle2, Scale>, which is different.

SetCharSubMode
--------------

Character substitution is not defined or discussed anywhere else in this
document.  However, from other reading, we are quite certain that this
operator requires an array with exactly 1 element, and that it turns on or
off vertical substitution using the VT segment of TrueType fonts, as for
PCL5.

SetMiterLimit
-------------

Setting a miter limit of 0 is [4] apparently equivalent to setting the miter
limit to its default value of 10.  We believe this behavior is deliberate
and that the specification omitted this point accidentally.

SetLineDash
-----------

The last sentence, stating that "the dash style of a line is not scaled when
a line is scaled", is simply wrong.  Test case:

	[40 20] @LineDashStyle SetLineDash
	[100 100] @Point SetCursor
	[300 0] @EndPoint LineRelPath
	PaintPath NewPath
	[100 200] @Point SetCursor
	[2 2] @Scale SetPageScale
	[300 0] @EndPoint LineRelPath
	PaintPath

In the second line, the dashes and gaps are twice as long.  We think the
text in the specification may be a holdover from an earlier version of the
design, and that the scaling behavior is the one that is actually intended,
since, for example, it matches the behavior of PostScript's setdash
operator.

SetClip*
--------

The discussion of ClipMode and ClipRegion is wrong almost everywhere.  See
the "Graphics State" section above.

SetClipIntersect
----------------

It appears that SetClipIntersect disregards the ClipRegion attribute if any
intersection actually occurs.  Test cases ("region1" and "region2" are
parameters, eInterior or eExterior):

	eEvenOdd @ClipMode SetClipMode
	[120 120] @Point SetCursor
	[100 0] @EndPoint LineRelPath
	[-100 100] @EndPoint LineRelPath
	CloseSubPath
	region1 @ClipRegion SetClipIntersect
	[120 120] @Point SetCursor
	[100 0] @EndPoint LineRelPath
	[0 100] @EndPoint LineRelPath
	CloseSubPath
	region2 @ClipRegion SetClipIntersect
	[100 100 440 440] @BoundingBox Rectangle

The outputs with region2 = eInterior are what one would expect (triangles
pointing "south" and "west"), but the outputs with region2 = eExterior are
the same (a square with a triangle cut-out pointing "northeast") whether
region1 = eInterior or region1 = eExterior.  We were unable to come up with
an interpretation of the specification that would make this the correct
behavior, so we think this is a firmware bug.

SetROP
------

The specification, read literally, would require keeping an internal
representation of the page in which every pixel was represented as a 24-bit
RGB value.  The implementation in the printer does no such thing: it applies
the given RasterOp/transparency algorithm to the physical device pixels
*after halftoning*, with white pixels resulting from halftoning being
treated as transparent if the corresponding transparency mode is set.  This
is tremendously simpler and cheaper to implement than what is in the
specification, but it is also produces very different results.  Test case:

	1 @ColorSpace SetColorSpace
	15 @GrayLevel SetBrushSource
	[100 100 200 200] @BoundingBox Rectangle
	86 @ROP3 SetROP			% D ^ (T | S)
	240 @GrayLevel SetBrushSource
	[100 100 200 200] @BoundingBox Rectangle

The specification requires painting the interior of the rectangle with the
XOR of the 15 and the 240, i.e., 255, i.e., white.  In fact, this produces a
dark gray shade resulting from XORing the two halftone masks together.

The equations for transparency processing (section 5.7.5) are badly
presented: Src and Paint in the first equation of each case (the computation
of Temporary_ROP3) refer to the actual pixels, while Src and Paint in the
other equations refer to masks that have 1s where the corresponding pixel is
not white.  For black-and-white printers with black = 1, the two are
equivalent, but for color printers, they are quite different.

In order to make ORing different gray shades together produce a result
approximating the sum of the shade values rather than the maximum, H-P has
apparently used very carefully designed default halftone screens, and
*different* screens for the source and paint operands of RasterOp.  There is
no way to achieve this effect with user-defined dither matrices, because the
same matrix is used for all cases.  Test case:

	<< optionally download a dither matrix >>
	eRGB @ColorSpace SetColorSpace
	For values of x from 0 by 5 to 255
	    For values of y from 0 to 5 by 255
		Let X = x * 10 + 100, Y = y * 10 + 100
		252 @ROP3 SetROP
		x @GrayValue SetBrushSource
		[X Y X+10 Y+10] @BoundingBox Rectangle
		238 @ROP3 SetROP
		0 @GrayValue SetBrushSource
		[X Y] @Point SetCursor
			0 @ColorMapping 2 @ColorDepth
			1 @SourceWidth 1 @SourceHeight
			[10 10] @DestinationSize
		BeginImage
			0 @StartLine 1 @BlockHeight 0 @CompressMode
		ReadImage
		<fb 04>		% stream preamble
		<y 00 00 00>	% stream data
		EndImage

The output for the downloaded matrix is very different from the output for
the default screen; a careful examination of the output with the default
screen will reveal that the square at (x,y) is different from the square at
(y,x), confirming that different screens are used for source and paint.

[3] SetHalftoneMethod
---------------------

[4] Setting a new halftone method does not affect the current brush or pen:
apparently SetBrushSource and SetPenSource immediately render the color
using the current halftone method, and PaintPath uses that rendering.  To
verify this:

		<< SetBrushSource with a gray level >>
		... SetHalftoneMethod ...
		<< construct a path >>
		PaintPath

The path will be painted with a brush that uses the old halftone method, not
the new one.

[7] The DitherOrigin is apparently relative not to the current user
coordinate system (as documented), but to the default user coordinate system
in the current orientation.  Here is a test file:

	150 600 usp @PageOrigin SetPageOrigin
	0 b @DitherMatrixDataType
	32 32 usp @DitherMatrixSize
	2 b @DitherMatrixDepth
	SetHalftoneMethod
	<< 1024 bytes of distinctive-pattern matrix omitted >>
	2 b @ColorSpace SetColorSpace
	[100 100 100] ba @RGBColor SetBrushSource
	0 0 32 32 usq @BoundingBox Rectangle

If the DitherOrigin were taken correctly, the output would consist of a
single, unshifted copy of the halftone tile.  However, the tile is shifted.

[7] When using the default dither matrix, the X component of the
DitherOrigin is ignored.  For example:

	eGray @ColorSpace SetColorSpace
	0 b @NullPen SetPenSource
	90 b @ROP3 SetROP		% D ^ T
	175 b @GrayLevel SetBrushSource
	0 0 100 100 usq Rectangle
	1 0 usp @DitherOrigin
	eDeviceBest @DeviceMatrix
	SetHalftoneMethod
	175 b @GrayLevel SetBrushSource
	0 0 100 100 usq Rectangle

The result is white: the dither matrix was not translated, and the gray
pattern cancelled itself out.  If one replaces the 6th line with

	0 1 usp @DitherOrigin

the result is a gray square, showing that the matrix was translated.

Painting operators
==================

[3] ArcPath
-----------

If the corners of the BoundingBox are specified with x1 > x2 or y1 > y2, the
following peculiar changes occur in the output:

	- x1 < x2, y1 < y2: the arc is drawn counter-clockwise from
	StartPoint to EndPoint, per the specification.

	- x1 < x2, y1 > y2: the arc is drawn clockwise from EndPoint to
	StartPoint.

	- x1 > x2, y1 < y2: the arc is drawn clockwise from the point
	opposite EndPoint to the point opposite StartPoint.

	- x1 > x2, y1 > y2: the arc is drawn counter-clockwise from
	point opposite StartPoint to the point opposite EndPoint.

There is probably some simple way of characterizing these changes
mathematically, but we haven't found it.  We think this is probably a
firmware bug, but it could also be a specification error.

[3] Chord
---------

The arc of the chord is drawn before the line.  This is not documented; it
matters when a dash pattern is being used.

Chord behaves like ArcPath for BoundingBox values with x1 < x2 or y1 < y2.

Chord probably clears the path, rather than leaving it set to the shape.
(We didn't verify this, but it seems likely given that Ellipse and Pie do
it.)  See Rectangle below.

[3] ChordPath
-------------

The arc of the chord is drawn before the line.  This is not documented; it
matters when a dash pattern is being used.

ChordPath behaves like ArcPath for BoundingBox values with x1 < x2 or y1 <
y2.

[1] Ellipse
-----------

[1] The specification does not say what the starting point of an ellipse is,
or in which direction the ellipse is drawn; this matters when a dash pattern
is being used.  [7] The starting point and drawing direction for ellipses
are as follows:

	- x1 < x2, y1 < y2: starts at 180 degree point (on the ellipse's X
	axis at minimum X), draws counter-clockwise.

	- x1 < x2, y1 > y2: starts at 180 degree point, draws clockwise.

	- x1 > x2, y1 < y2: starts at 0 degree point, draws clockwise.

	- x1 > x2, y1 > y2: starts at 0 degree point, draws
	counter-clockwise.

Ellipse clears the path, rather than leaving it set to the shape.  See
Rectangle below.

[3] Pie
-------

The line from the center to the starting point of the arc is drawn first,
then the arc, then the line back to the center.  This is not documented; it
matters when a dash pattern is being used.

Pie behaves like ArcPath for BoundingBox values with x1 < x2 or y1 < y2.

Pie clears the path, rather than leaving it set to the shape.  See Rectangle
below.

[3] PiePath
-----------

PiePath draws in the same order as Pie.

PiePath behaves like ArcPath for BoundingBox values with x1 < x2 or y1 < y2.

Rectangle
---------

The specification says that Rectangle is equivalent to
	NewPath RectanglePath PaintPath

Since PaintPath does not reset the path, Rectangle should leave the path set
to the rectangle; the postcondition in the specification says this
explicitly.  However, the implementation resets the path after Rectangle.
Test case:

	25 @PenWidth SetPenWidth
	0 @NullBrush SetBrushSource
	eGray @ColorSpace SetColorSpace
	60 @GrayLevel SetPenSource
	100 100 @Point SetCursor
	200 0 @EndPoint LineRelPath
	PaintPath		% should not reset path
	120 @GrayLevel SetPenSource
	100 150 @Point SetCursor
	200 0 @EndPoint LineRelPath
	PaintPath		% should repaint first line
	180 @GrayLevel SetPenSource
	100 200 200 300 @BoundingBox Rectangle	% should leave path set to rectangle
	230 @GrayLevel SetPenSource
	100 400 @Point SetCursor
	200 0 @EndPoint LineRelPath
	PaintPath		% should repaint rectangle

The last line is painted a lighter gray than the rectangle, demonstrating
that Rectangle left the path empty.  We are not sure whether this is a
firmware bug or a change in the intended specification.

We verified that Ellipse and Pie behave the same way, but we did not test
whether this behavior extends to the other operators (Chord, RoundRectangle)
that one might expect to behave similarly.

Rectangle, RectanglePath
------------------------

[1] Even though the specification says rectangles are drawn
counter-clockwise, they are actually drawn clockwise, starting in the upper
left corner.  Test case:

	10 us @PenWidth SetPenWidth
	[130 10 40 10 30 10 20 10] ssa @LineDashStyle
	  0 ss @DashOffset SetLineDash
	0 b @NullBrush SetBrushSource
	200 200 700 400 ssq @BoundingBox Rectangle

The dash pattern clearly starts in the upper left corner and runs clockwise.

[7] Rectangles allow specifying the bounding box points in any order; the
rectangle is always drawn clockwise, starting with the point that has the
lesser user space coordinates.

RoundRectangle, RoundRectanglePath
----------------------------------

The BoundingBox attribute gives the bounding box for the rectangle, not the
ellipse.  This is just a typo, but a substantial one.

[1] Round rectangles, unlike rectangles, *are* drawn counter-clockwise.
starting at the top of the straight part of the left edge.  Test case:

	10 us @PenWidth SetPenWidth
	[130 10 40 10 30 10 20 10] ssa @LineDashStyle
	  0 ss @DashOffset SetLineDash
	0 b @NullBrush SetBrushSource
	200 200 700 400 ssq @BoundingBox
	  120 120 usp @EllipseDimension RoundRectangle

[7] Round rectangles give an IllegalAttributeValue error if the bounding box
points are not specified with x1 < x2, y1 < y2.

ScanLineRel
-----------

The description of x-pairs in section 6.12 is either misleading or wrong.
If the current X position is X and the values in the x-pair are U and V, the
x-pair causes a line to be drawn from X+U to X+U+V, not X+U to X+V.  (The
test case is too messy to present here.)

[3] Text, TextPath
------------------

[3] It appears that text transforms *objects*, whereas images and paths
transform *coordinates*.  Mathematically, this requires applying
transformations to text in the opposite order from graphics.  Test case:

	(TimesNewRmn     ) ba @FontName
	  200 us @CharSize 277 us @SymbolSet SetFont
	2500 2500 usp @PageOrigin SetPageOrigin
	4 1 usp @PageScale SetPageScale
	NewPath 0 0 usp @Point SetCursor

	<< repeat 2-4 times: >>

	PushGS
	(1) ba @TextData Text
	200 0 usp @Point SetCursor
	1 1.5 rp @PageScale SetPageScale
	(2) ba @TextData TextPath PaintPath NewPath
	400 0 usp @PageOrigin SetPageOrigin
	1 1.5 rp @PageScale SetPageScale
	90 us @PageAngle SetPageRotation
	0 0 usp @Point SetCursor
	(3) ba @TextData Text
	200 0 usp @Point SetCursor
	1 1.5 rp @PageScale SetPageScale
	(4) ba @TextData TextPath PaintPath NewPath
	PopGS
	90 us @PageAngle SetPageRotation

	<< end repeat >>

For example, the '1' characters are all the same shape, indicating that they
were scaled (as objects) before being rotated; if this transformation had
been applied to the coordinates, both the portrait and the landscape
characters would have been stretched in the page X direction.  This
interpretation of the specification is far from obvious, but it is not
unreasonable, so we think it is what is intended rather than a bug.
