+ - Testing
+ - Test implementations against each other
+ - Test both with and without the operator strength reduction.
+ They shold be identical.
+
+ - SSE 2 issues:
+
+ - Use MM_HINT_NTA instead of MM_HINT_T0
+
+ - Use of fbCompositeOver_x888x8x8888sse2()
+
+ - Update the RLEASING file
+
+ - Things to keep in mind if breaking ABI:
+
+ - There should be a guard #ifndef I_AM_EITHER_CAIRO_OR_THE_X_SERVER
+
+ - X server will require 16.16 essentially forever. Can we get
+ the required precision by simply adding offset_x/y to the
+ relevant rendering API?
+
+ - Get rid of workaround for X server bug.
+
+ - pixman_image_set_indexed() should copy its argument, and X
+ should be ported over to use a pixman_image as the
+ representation of a Picture, rather than creating one on each
+ operation.
+
+ - We should get rid of pixman_set_static_pointers()
+
+ - We should get rid of the various trapezoid helper functions().
+ (They only exist because they are theoretically available to
+ drivers).
+
+ - 16 bit regions should be deleted
+
+ - There should only be one trap rasterization API.
+
+ - The PIXMAN_g8/c8/etc formats should use the A channel
+ to indicate the actual depth. That way PIXMAN_x4c4 and PIXMAN_c8
+ won't collide.
+
+ - Maybe bite the bullet and make configure.ac generate a pixman-types.h
+ file that can be included from pixman.h to avoid the #ifdef magic
+ in pixman.h
+
+ - Make pixman_region_point_in() survive a NULL box, then fix up
+ pixman-compose.c
+
+ - Possibly look into inlining the fetch functions
+
+ - There is a bug with source clipping demonstrated by clip-test in the
+ test directory. If we interprete source clipping as given in
+ destination coordinates, which is probably the only sane choice,
+ then the result should have two red bars down the sides.
+
+ - Test suite
+
- Add a general way of dealing with architecture specific
fast-paths. The current idea is to have each operation that can
be optimized is called through a function pointer that is
initially set to an initialization function that is responsible for
setting the function pointer to the appropriate fast-path.
- - Get rid of the switch-of-doom; replace it with a big table
- describing the various fast paths.
-
- Go through things marked FIXME
- Add calls to prepare and finish access where necessary. grep for
this, I suggest just using 32-bit datatypes by setting
IC_SHIFT to 5 for all machines.
- - Consider whether calling regions region16 is really such a great
- idea Vlad wants 32 bit regions for Cairo. This will break X server
- ABI, but should otherwise be mostly harmless, though a
- pixman_region_get_boxes16() may be useful.
-
- Consider optimizing the 8/16 bit solid fills in pixman-util.c by
storing more than one value at a time.
- Add an image cache to prevent excessive malloc/free. Note that pixman
needs to be thread safe when used from cairo.
+ - Moving to 24.8 coordinates. This is tricky because X is still
+ defined as 16.16 and will be basically forever. It's possible we
+ could do this by adding extra offset_x/y parameters to the
+ trapezoid calls. The X server could then just call the API with
+ (0, 0). Cairo would have to make sure that the delta *within* a
+ batch of trapezoids does not exceed 16 bit.
+
+ - Consider adding actual backends. Brain dump:
+
+ A backend is something that knows how to
+
+ - Create images
+ - Composite three images
+ - Rasterize trapezoids
+ - Do solid fills and blits
+
+ These operations are provided by a vtable that the backend will
+ create when it is initialized. Initial backends:
+
+ - VMX
+ - SSE2
+ - MMX
+ - Plain Old C
+
+ When the SIMD backends are initialized, they will be passed a
+ pointer to the Plain Old C backend that they can use for fallback
+ purposes.
+
+ Images would gain a vtable as well that would contain things like
+
+ - Read scanline
+ - Write scanline
+
+ (Or even read_patch/write_patch as suggested by Keith a while
+ back).
+
+ This could simplify the compositing code considerably.
+
- Review the pixman_format_code_t enum to make sure it will support
future formats. Some formats we will probably need:
What about color spaces such a linear vs. srGB etc.?
+
done:
+- Use pixmanFillsse2 and pixmanBltsse2
+
+- Be consistent about calling sse2 sse2
+
+- Rename "SSE" to "MMX_EXTENSIONS". (Deleted mmx extensions).
+
+- Commented-out uses of fbCompositeCopyAreasse2()
+
+- Consider whether calling regions region16 is really such a great
+ idea. Vlad wants 32 bit regions for Cairo. This will break X server
+ ABI, but should otherwise be mostly harmless, though a
+ pixman_region_get_boxes16() may be useful.
+
+- Altivec signal issue (Company has fix, there is also a patch by
+ dwmw2 in rawhide).
+
+- Behdad's MMX issue - see list
+
+- SSE2 issues:
+ - Crashes in Mozilla because of unaligned stack. Possible fixes
+ - Make use of gcc 4.2 feature to align the stack
+ - Write some sort of trampoline that aligns the stack
+ before calling SSE functions.
+
+- Get rid of the switch-of-doom; replace it with a big table
+ describing the various fast paths.
+
- Make source clipping optional.
- done: source clipping happens through an indirection.
still needs to make the indirection settable. (And call it