1 - Behdad's MMX issue - see list
5 - Commented-out uses of fbCompositeCopyAreasse2()
7 - Use MM_HINT_NTA instead of MM_HINT_T0
9 - Use of fbCompositeOver_x888x8x8888sse2()
11 - Use pixmanFillsse2 and pixmanBltsse2
13 - Update the RLEASING file
15 - Things to keep in mind if breaking ABI:
17 - There should be a guard #ifndef I_AM_EITHER_CAIRO_OR_THE_X_SERVER
19 - X server will require 16.16 essentially forever. Can we get
20 the required precision by simply adding offset_x/y to the
21 relevant rendering API?
23 - pixman_image_set_indexed() should copy its argument, and X
24 should be ported over to use a pixman_image as the
25 representation of a Picture, rather than creating one on each
28 - We should get rid of pixman_set_static_pointers()
30 - We should get rid of the various trapezoid helper functions().
31 (They only exist because they are theoretically available to
34 - Regions should become 32 bit
36 - There should only be one trap rasterization API.
38 - The PIXMAN_g8/c8/etc formats should use the A channel
39 to indicate the actual depth. That way PIXMAN_x4c4 and PIXMAN_c8
42 - Make pixman_region_point_in() survive a NULL box, then fix up
45 - Possibly look into inlining the fetch functions
49 - Add a general way of dealing with architecture specific
50 fast-paths. The current idea is to have each operation that can
51 be optimized is called through a function pointer that is
52 initially set to an initialization function that is responsible for
53 setting the function pointer to the appropriate fast-path.
55 - Go through things marked FIXME
57 - Add calls to prepare and finish access where necessary. grep for
58 ACCESS_MEM, and make sure they are correctly wrapped in prepare
61 - restore READ/WRITE in the fbcompose combiners since they sometimes
62 store directly to destination drawables.
64 - It probably makes sense to move the more strange X region API
65 into pixman as well, but guarded with PIXMAN_XORG_COMPATIBILITY
67 - Reinstate the FbBits typedef? At the moment we don't
68 even have the FbBits type; we just use uint32_t everywhere.
70 Keith says in bug 2335:
72 The 64-bit code in fb (pixman) is probably broken; it hasn't been
73 used in quite some time as PCI (and AGP) is 32-bits wide, so
74 doing things 64-bits at a time is a net loss. To quickly fix
75 this, I suggest just using 32-bit datatypes by setting
76 IC_SHIFT to 5 for all machines.
78 - Consider whether calling regions region16 is really such a great
79 idea. Vlad wants 32 bit regions for Cairo. This will break X server
80 ABI, but should otherwise be mostly harmless, though a
81 pixman_region_get_boxes16() may be useful.
83 - Consider optimizing the 8/16 bit solid fills in pixman-util.c by
84 storing more than one value at a time.
86 - Add an image cache to prevent excessive malloc/free. Note that pixman
87 needs to be thread safe when used from cairo.
89 - Moving to 24.8 coordinates. This is tricky because X is still
90 defined as 16.16 and will be basically forever. It's possible we
91 could do this by adding extra offset_x/y parameters to the
92 trapezoid calls. The X server could then just call the API with
93 (0, 0). Cairo would have to make sure that the delta *within* a
94 batch of trapezoids does not exceed 16 bit.
96 - Review the pixman_format_code_t enum to make sure it will support
97 future formats. Some formats we will probably need:
99 ARGB/ABGR with 16/32/64 bit integer/floating channels
103 Also we may need the ability to distinguish between PICT_c8 and
104 PICT_x4c4. (This could be done by interpreting the A channel as
105 the depth for TYPE_COLOR and TYPE_GRAY formats).
107 A possibility may be to reserve the two top bits and make them
108 encode "number of places to shift the channel widths given" Since
109 these bits are 00 at the moment everything will continue to work,
110 but these additional widths will be allowed:
112 All even widths between 18-32
113 All multiples of four widths between 33 and 64
114 All multiples of eight between 64 and 128
116 This means things like r21g22b21 won't work - is that worth
117 worrying about? I don't think so. And of course the bpp field
118 can't handle a depth of over 256, so > 64 bit channels arent'
119 really all that useful.
121 We could reserve one extra bit to indicate floating point, but
124 PIXMAN_TYPE_ARGB_FLOAT
125 PIXMAN_TYPE_BGRA_FLOAT
128 image types. With five bits we can support up to 32 different
129 format types, which should be enough for everybody, even if we
130 decide to support all the various video formats here:
132 http://www.fourcc.org/yuv.php
134 It may make sense to have a PIXMAN_TYPE_YUV, and then use the
135 channel bits to specify the exact subtype.
137 Another possibility is to add
142 where the channel widths would get 16 and 32 added to them,
145 What about color spaces such a linear vs. srGB etc.?
150 - Crashes in Mozilla because of unaligned stack. Possible fixes
151 - Make use of gcc 4.2 feature to align the stack
152 - Write some sort of trampoline that aligns the stack
153 before calling SSE functions.
155 - Get rid of the switch-of-doom; replace it with a big table
156 describing the various fast paths.
158 - Make source clipping optional.
159 - done: source clipping happens through an indirection.
160 still needs to make the indirection settable. (And call it
163 - Run cairo test suite; fix bugs
164 - one bug in source-scale-clip
166 - Remove the warning suppression in the ACCESS_MEM macro and fix the
167 warnings that are real
170 - make the wrapper functions global instead of image specific
171 - this won't work since pixman is linked to both fb and wfb
173 - Add non-mmx solid fill
175 - Make sure the endian-ness macros are defined correctly.
177 - The rectangles in a region probably shouldn't be returned const as
178 the X server will be changing them.
180 - Right now we _always_ have a clip region, which is empty by default.
181 Why does this work at all? It probably doesn't. The server
182 distinguishes two cases, one where nothing is clipped (CT_NONE), and
183 one where there is a clip region (CT_REGION).
185 - Default clip region should be the full image
187 - Test if pseudo color still works. It does, but it also shows that
188 copying a pixman_indexed_t on every composite operation is not
189 going to fly. So, for now set_indexed() does not copy the
192 Also just the malloc() to allocate a pixman image shows up pretty
197 - Make all the setters not copy their arguments
199 - Possibly combined with going back to the stack allocated
200 approach that we already use for regions.
202 - Keep a cached pixman_image_t around for every picture. It would
203 have to be kept uptodate every time something changes about the
206 - Break the X server ABI and simply have the relevant parameter
207 stored in the pixman image. This would have the additional benefits
210 - We can get rid of the annoying repeat field which is duplicated
213 - We can use pixman_color_t and pixman_gradient_stop_t
214 etc. instead of the types that are defined in