nip2/TODO at master · libvips/nip2

Cannot retrieve contributors at this time

1202 lines (747 sloc) 30.7 KB

	- add tests for new stuff

	- seeing occasional "rewind image with active regions" message

	- should parse_float / _int etc. allow leading and trailing spaces?

	(split is_space x)?0

	- how about adding

	zip2 10 [1..5] == [[10, 1], [10, 2], [10, 3], ..

	should be harmless, and quite useful

	same for zip3 etc. as well

	check zip use, we probably have this code there already, in various places

	- sharpen should use new interface?

	- can we call affine from nip2 vips_call? do we need a double array?

	- hough_circle etc. don't get cached ... they use the vips8 API and the vips
	cache only works for vips8

	we can't turn on the vips8 cache since it does not know about invalidate

	- columns can move about on load in large workspaces

	- hide tabs if only one tab in window, though we'd need to allow tab drop
	anywhere in window for that

	- Matrix / New / Laplacian

	edit a cell, turns into a plain matrix

	need to override edit method

	same for gaussian

	rather tricky, compared to square / circle etc.



	- breadcrumb trail for prog window, so you can get back to where you were?

	- lambdas don't pattern match?

	(\[a, b] a + b)



	- load jpg, ^Q, no leaks
	load jpg, paint bar, paint one dot, ^Q, no leaks

	load jpg, extract area, paint bar, paint one dot, ^Q, leaks
	load jpg, extract area, paint bar, paint one dot, ^Z, ^Q, leaks
	load jpg, extract area, paint bar, paint one dot, close window, ^Q, leaks

	seems to leak:

	original image,
	one large regions on that image (full width)
	extract area operation
	extracted image
	region a bit bigger than paint action on that image

	0) VipsRegion (0x8b052a0)
	VipsRegion (object), base class VipsRegion: 0x8b052a0, im = 0x887aad0, left =
	192, top = 128, width = 45, height = 2 VipsRegion (0x8b052a0)
	1) VipsImage (0x887aad0)
	VipsImage (image), image class 237x202 uchar, 3 bands, srgb, partial VipsImage
	(0x887aad0)
	2) VipsImage (0x8016b30)
	VipsImage (image), image class 972x1296 uchar, 3 bands, srgb, openin VipsImage
	(0x8016b30)
	3) VipsRegion (0x8b05340)
	VipsRegion (object), base class VipsRegion: 0x8b05340, im = 0x8016b30, left =
	0, top = 30, width = 972, height = 283 VipsRegion (0x8b05340)
	4) VipsExtractArea (0x882a910)
	VipsExtractArea (extract_area), extract an area from an image - extract_area
	((VipsImage*) 0x8016b30) 142 158 237 202 VipsExtractArea (0x882a910)

	something to do with vips_image_wio_input() and the way it rewinds a PARTIAL
	image? called from im_rwcheck()


	- os x build reports missing jasper dylib?

	- draw_circle could extract, draw and insert for huge memuse reduction

	- section on compat mode for the docs

	see mail to MvGulick for some notes

	- expose more of the tone funcs in nip2

	- quite a few hist operations have no GUI ... histspec, for example?

	- nip2 should use zooming support, if possible

	- the windows setup .exe install a bazillion .png icons we will never use,
	then installs them all again as .svg, which we will certainly never use

	- add extract_volume, cf. "grid"

	record tile_size in meta somewhere? grid could set it, save a parameter off
	extract_volume

	also extract_sequence to get volume over time

	- image + complex constant would be nice




	- ban parameters which are built-ins, eg. "im", "re" etc., you get nasty
	crashes

	see also notes below: a new parser could fix this

	can only ban for new code? do we have duff code in compat?

	argh yes there are at least 15 like this, fix them or fix the parser? also
	need to disable this check for compat defs

	better to fix the parser, it can't be that hard

	need to fix up the list comp compiler too, sigh






	- we can't define local superclasses at the moment

	eg. consider:

	Fred = class {
	Jim a = class {
	value = a + 12;
	}

	Jennie x = class Jim x {
	value = 99;
	}
	}

	you can't do (Fred.Jennie 12) since Jim will have a secret 'this' param
	(because it is a class member) and superclass constructors can't have
	secrets

	don't automatically give all members 'this' as a secret, check that they
	make references to other class members that do need secrets first

	- turn on GM in prefs, have to restart before _stdenv.def:magick sees the
	change

	- try this:

	Workspaces.untitled
	has_member "A1" A1

	doesn't seem to work?



	- matrix_gaussian_blur could have an 'accuracy' or 'max error' param? expose
	in custom blur etc.



	- oo_binary etc. needs revising, we don't search down branches as we should

	for example, Matrix does:

	// compound ... don't do iteration
	[this.Matrix_base (op.fn this.value x.value),
	(is_Matrix x \|\| is_Real x \|\| is_Vector x) &&
	op.type == Operator_type.COMPOUND_REWRAP],

	which is stupid, we should not wire Real and Vector in there, it ought to be
	something like:

	[this.Matrix_base (op.fn this.value x),
	op.type == Operator_type.COMPOUND_REWRAP],

	ie. don't strip the .value off x and rely on op.fn to do that, but this
	breaks in various ways

	remove all of _Object and redo it, thinking about what we want operators to
	look like and what we want types to look like.

	Have a base class for operators that does most of the standard stuff

	get rid of the operator types rewrap / arithmetic / relational etc. etc.








	- try

	point re
	= x
	{
	(x, y) = re;
	}

	it's the 're' param, it stops x being bound to the
	get-real-part-of-complex builtin

	expands to

	point re
	= x
	{
	$$x = re;
	x
	= re $$x, is_complex $$x
	= error "bad match";
	}

	add secrets

	point point.re
	= x point.re
	{
	$$1 $$1.re = point.re;
	x x.re
	= x.re ($$x point.re), is_complex ($$x point.re)
	= error "bad match";
	}


	x compiles to

	if_then_else (<symbol "point.x.is_complex"> (<symbol "point.$$pattern_lhs78"> <symbol "point.re">))
	<symbol "point.x.re"> (<symbol "point.$$pattern_lhs78"> <symbol "point.re">)
	<symbol "point.x.error"> <Managed* 0x37c5a80>

	abstracting point.re

	after var abstract

	((S ((Sr (if_then_else <compileref "point.x">)) ((Sl ((Sr (&& <compileref "point.x">)) ((Sr <symbol "point.x.is_complex">) <symbol "point.$$pattern_lhs78">))) true))) ((Sl ((Sr SHARE0[(: <compileref "point.x">)]) ((Sr <symbol "point.x.re">) <symbol "point.$$pattern_lhs78">))) ((REF0 (<symbol "point.x.error"> <Managed* 0x3a55980>)) [ ])))

	reduce and get

	reduce_spine: (<symbol "point.x.re"> (<symbol "point.$$pattern_lhs0"> (I (1,2))))
	reduce_spine: (<symbol "point.re"> (<symbol "point.$$pattern_lhs0"> (I (1,2))))
	sym-param found, argh: point.re

	maybe fix this when we revise the parser

	would be a good time to add multiple definitions as well

	- redo name resolution in parser ... scrap the patch thing, instead have a
	separate 'resolve' step that runs after parsing a top-level

	don't create ZOMBIE symbols, instead make REF nodes in the tree

	this binds local references, but leaves external refs dangling

	we do a final link at load time when we copy into the heap

	do we need zombies at all now?

	make a fork for this



	- we have ws->window_width, can we use the one on Model now instead?



	- new inplace stuff needs a test suite

	- 'don't show this dialog again' on delete row dialog, also in prefs

	box_yesno() could take a string which is the name of a pref to check for
	ask-or-not



	- how about something that does:

	im' = operation_list [a, b, c, d] im

	it does a fold:

	im' = d (c (b (a im)))

	but the intermediate images are reused, so you can do in-place stuff with it

	we could get rid of lineset!






	- mac binary has a broken im_text() argh

	- do we allow

	[r, g, b] = Image_file "babe.jpg"

	since ? is band index and list index, it seems to make sense

	we have

	image ++ image
	image ++ [] == image

	for bandjoin, so that lines up too, I guess

	hmm

	reverse image

	to swap the bands over? heh need to be able to override hd and tl

	- Ackermann

	http://shootout.alioth.debian.org/great/benchmark.php?test=ackermann&lang=all&sort=cpu

	A x y
	= y + 1, x == 0
	= A (x - 1) 1, y == 0
	= A (x - 1) (A x (y - 1));

	correct test result: A 3 4 == 125

	A 3 10 is benchmark test ... we fail with a "C stack overflow" error

	A 3 9 == 4093 works OK

	we could make this a lot quicker with some arithmetic streamlining

	could we do tail-recursion elimination?

	strictness analysis would help too

	- Fib

	http://shootout.alioth.debian.org/great/benchmark.php?test=fibo&lang=all&sort=cpu

	F x
	= 1, x == 0
	= 1, x == 1
	= F (x - 2) + F (x - 1);

	correct output F 32 == 3524578

	cima is ~370s for this (!!)

	work machine is about 50s

	dell vostro laptop is 45s without optimiser

	- turn on DEBUG in heap.c, run the fibonacci benchmark

	we heap_copy F every time it recurses! because when we heap_copy we don't
	link recursive references

	try fixing this ... but remember the problems we had with shared classes
	when we did link-to-value rather than link-to-copy

	we also have a huge amount of stuff in the heap, could we trim this down?
	how does it all get pulled in? is it preferences?

	in nip1, F is about 4x faster

	WONTFIX for 7.20
	================

	- look at:

	http://www.eggheadcafe.com/software/aspnet/35467981/programmatic-console-appl.aspx

	possibly better than nip2-cli.exe

	- turning a column from Many Stats into a vector for doing arithmetic is very
	tricky argh

	add a matrix->vector converter? or maybe a one column or 1 row matrix should
	also be a vector

	- try:

	nip2 Workspaces.uniformshapes2.A1.start_line=21 uniformshapes2.ws

	--set does not work for non-toplevels

	- try:

	A1 = [1]
	A2 = [x :: x <- A1]

	change A1, A2 does not update, argh

	we get:

	link_expr_new: expr A2.$$lcomp0 references link->child = A1

	so perhaps we are updating the local of A2, but not A2?

	A2 is certainly being marked dirty ... on change of A1 we get:

	row_dirty_set_single: A1 clear_error = true
	symbol_dirty_set: A1 (0x1a59480)
	symbol_dirty_set: A2 (0x1a595a0)
	symbol_recalculate_check: untitled.A1
	row_dirty_clear: A1
	row_recomp_all: done row A1 - 0.000143873s
	row_dirty_set_single: A1 clear_error = false
	row_dirty_clear: A1
	symbol_dirty_clear: A1 (0x1a59480)
	success: [2]
	symbol_recalculate_check: untitled.A2
	symbol_dirty_clear: A2 (0x1a595a0)
	success: [1]

	so maybe A2.something is being updated, but the row is not

	we now mark a row dirty if a sub-expr is dirty. but in row_renerate(), we
	don't build subexprs

	should we mark the subexpr dirty? (or maybe we do?)

	or should we always copy all subexprs when we copy an expr

	or only subexprs with no row?

	do we calc rows outside-in or inside-out? does this affect copying subrows?

	when do we copy now, the first time a row is made?

	- we destroy and rebuild all links during recomp (eg. turn on DEBUG in
	link.c), why is this? can't we only rebuild on a change of source text?

	- fix the FIXME in itext_clear_edited() or wherever it is

	- try:

	start nip2
	dir untitled
	create A2, A3, etc.
	A1 does not update

	when we add/remove a def to workspace, should we mark the ws dirty?

	- lambdas should allow patterns? eg.:

	map (\[x, y] x + y) [[1, 2], [3, 4]] == [3, 7]

	- OS X bundler:

	http://sourceforge.net/apps/trac/gtk-osx/wiki/Bundle

	test?

	- imageinfo_make_paintable() no longer copies to a file, since this used to
	cause problems with dangling pointers because of the im_close()s we had to
	do

	however, this means we now do all painting in memory :-( do we need to add
	API to change a memory image (eg. a "p") into a file?

	- test Joe's layout thing, compare to the thing we do in study2 to make the
	diagnostic image

	- im_blend(), im_ifthenelse(), im_add() etc. now do bandalike/formatalike

	where do we use our bandalike/formatalike stuff? remove our stuff, though
	make sure we have equivalents in vips now

	- outline text example

	- needs a custom convol menu item which can loop over a group of matricies
	with a single image

	actually, we need to nail this down, otherwise when we pass in a list of
	sample texts the loops don't nest

	- right-click menu on row button should have items for "Jump to referrer / WC1
	/ JC1 ..." and "Jump to referred / ..."

	- why didn't im_copy_file() work? mysterious

	- line colours are wrong, argh, very mysterious, see

	plot_new_gplot()

	- gtk3.0 tests:

	build with

	#define G_DISABLE_DEPRECATED
	#define G_DISABLE_SINGLE_INCLUDES
	#define GDK_DISABLE_DEPRECATED
	#define GTK_DISABLE_DEPRECATED
	#define GDK_DISABLE_SINGLE_INCLUDES
	#define GTK_DISABLE_SINGLE_INCLUDES

	paste into config.h, somehow

	need to remove:

	GtkType
	gtk_type_new
	gtk_signal_connect
	GTK_SIGNAL_FUNC
	gtk_signal_handler_block_by_data
	` gtk_signal_connect_object
	GTK_CHECK_CAST
	GTK_CHECK_TYPE
	GtkSignalFunc

	- add Set menu to Math with mkset/union/intersection/difference ?

	could do bit operations on images?

	- lcomps like:

	argh = [x :: x <- poop]

	the 'x' gets copied inside the lcomp, leaving a zombie 'x' attached to argh,
	which Program / View / Errors then reports

	fix: don't resolve names as we parse and junk the ugly patch list thing

	instead, have a separate resolve stage that runs after we've moved scraps
	of graph to their final home

	- if we want full VipsObject introspection we will need a lot more

	vips_object_arguments (name2gtype "VipsInterpolateYafrsmooth")
	-> ["sharpness"]

	need some equivalent for GParamSpec / VipsArgument

	VipsArgument name type .... = class {}

	return a list of these from vips_object_arguments()?

	- heap_map_dict() should be reduce_map_dict(), since it does reduction, argh

	redo heap_map_dict() in terms of reduce_map_dict()

	actually, remove all the reduce_ stuff, it's daft to have a distinction

	something to do when we break Snip out into libsnip

	- Plot window should have image header menu item?

	not trivial, image header needs a conversion to watch

	we'd need to make a conversion in plotmodel

	- filesel guess-file-type-from-suffix needs fixing

	copy the vips model of having a user_name which is just "Workspace" or
	somesuch, and making "Workspace file (*.ws)" string at runtime

	use this to identity file types in util.c as well: get_image_info() needs it

	- for rows made by typing stuff, always show the formula as well as the value

	by default anyway?

	we'd need to always show the up/down arrows, not just for classes

	- drag from an image thumbnail to the ws background and you get a new column
	with "A2" or whgatever in

	does not work for plot thumbnails! how annoying

	- right-click on image background to get a context menu with
	save/replace/header? same as row thumbnail context?

	- look at using goffice instead of gtkplot for graphs

	http://ftp.gnome.org/pub/gnome/sources/goffice/

	also in synaptic

	there's also a new cairo-based gtkplot in SVN, apparently

	- try

	last [1..]

	then CTRL-W ... we can quit the app, but it's still evaling and the prompt
	never comes back

	- Math / Cluster is a bit useless, will it work for complex numbers? vectors?
	colours? groups?

	what would we have to do to get it to work for these other types?

	- toolkit load is where compiled code would go in

	need to make load / parse / compile self-contained ... the only output is a
	list of symbols, each with code and sub-symbols; there are no toolkits or
	whatever made

	after load / parse / compile, we need to walk the symbol list building tools
	and all that stuff

	we do:

	load toolkit:
	get filesize, date last modified, md5sum
	look in ~/.nip2-7.x.x/cache for a file named with that md5sum
	if present, open and check first two fields: filesize and date
	if match, load compiled code
	if no match, load / parse / compile toolkit, then save compiled
	code to ~/.nip2-7.x.x./cache
	walk symbol list building tools and all that stuff

	how much time will this really save? can we easily get an estimate?

	steps to follow:

	1. make load / parse / compile self-contained, with separate pass to
	build tools etc.

	this is a useful cleanup whatever else we do

	2. now we know exactly what the output of load / parse / compile is, we
	should be able to write to a file

	make our own binary format, don't use XML ... we want speed

	3. try loading and benchmarking

	4. if the benchmarks look promising, harden and polish


	- numbering of group of group save seems to skip one at end of line?

	- Math / Cluster is a bit useless, will it work for complex numbers? vectors?
	colours? groups?

	what would we have to do to get it to work for these other types?

	- segv in test_toolkits on laptop (inside fftw3) ???!? valgrinds cleanly on
	work machine

	- try

	last [1..]

	then CTRL-W ... we can quit the app, but it's still evaling and the prompt
	never comes back

	- configure no longers sets GMSGFMT, is this OK? test on OS X

	- for rows made by typing stuff, always show the formula as well as the value

	by default anyway?

	we'd need to always show the up/down arrows, not just for classes

	- drag from an image thumbnail to the ws background and you get a new column
	with "A2" or whgatever in

	does not work for plot thumbnails! how annoying

	- right-click on image background to get a context menu with
	save/replace/header? same as row thumbnail context?

	- look at using goffice instead of gtkplot for graphs

	http://ftp.gnome.org/pub/gnome/sources/goffice/

	also in synaptic

	WONTFIX for 7.14
	================

	- quit while a thumbnail is painting: IMAGEs are leaked? seems esp. bad with
	greyc, perhaps just because it's so slow

	- do we enforce no-args-to-LHS-pattern anywhere? try

	[a,b] a b = 12;

	- use destroy_if_destroyed() in more places?

	grep destroy_ *.h

	- after pressing "Process" in the edit window, we always select last_sym,
	which is often not what we want

	make it jump less after a process ... eg. try editing something in the
	middle of Image/Transform, very annoying

	- use compile->last_sym to spot chains of defs

	multiple definitions of the same function are allowed, provided they all
	multiple definitions of the same function are allowed, provided they all
	have the same number of arguments, and provided only one of them has no
	argument pattern matching

	example:

	fred [a, b, 1] = a * b;
	fred (Image x) = rot90 x;
	fred z = error "what";

	any of these can have locals, those locals just apply to that one
	definition

	this compiles to:

	fred $$a4
	= $$fred1, is_list $$a4 && len $$a4 == 2 && $$a4?2 == 1
	= $$fred2, is_instance_of "Image" $$a4
	= $$fred3
	{
	$$fred1
	= a * b
	{
	a = $$a4?0;
	b = $$a4?1;
	}
	$$fred2
	= rot90 x
	{
	x = $$a4;
	}
	$$fred3
	= error "what";
	{
	z = $$a4;
	}
	}

	so each pattern-matching definition generates a condition and an action

	constants in patterns become part of the condition test

	the action goes into a private function, the conditions are joined together
	in the function wrapper

	the no-pattern case (if present) becomes the action for the "otherwise"
	clause in the wrapper

	if not present, we generate (error "pattern match failed") or somesuch for
	the default case

	we will need to regenerate the wrapper function every time a definition of
	fred is added or removed ... can we do this easily?

	when are two definitions considered equal? should we warn about this? "fred
	x" could occur in two files, for example

	process:

	* we see a "fred arg-list" incoming
	* does the arg list contain any patterns?
	yes:
	* do we already have a fred in this scope?
	yes:
	* the existing fred must be the wrapper, this must be
	a new possible RHS
	* check that the number of args matches
	no:
	* generate a fred to act as the wrapper
	* add args called $$arg1 etc. to the main fred
	* add this new fred as a $$fredn local to the current
	fred
	* expand the patterns as local references to the main
	fred's arguments
	* parse the rest of the def in that context
	* keep the pattern list around, we'll need it to generate the
	ifs for the wrapper later
	no:
	* do we have a fred in this scope?
	yes:
	* check the previous fred was a pattern matcher
	* check the number of args matches
	* check there isn't already a default case
	* add as above
	no:
	* add as a regular non-pattern definition

	issues:

	* where do we store the pattern lists? we can't expand them at
	parse-time, since we need them to make the wrapper (which we can't
	make until we've seen all the candidate RHS)
	* when one of the RHS is changed, we need to regenerate the wrapper,
	how do we do this? (could nuke the generated code in the compile
	when we see a new RHS, then rebuild the wrapper in compile on the
	next heap_copy?)
	* our current condition generator won't work ... we need to test
	consts as well, and it'll be rather inefficient as we'll repeatedly
	test the trunk as we loop over the leaves --- instead, walk the
	pattern recursively top-down testing each node

	process:

	* see "fred" (as opposed to simple_pattern)
	* is there already a fred in scope?
	yes:
	* was current_compile->last_sym also a "fred"?
	yes:
	* another definition
	yes:
	* this must be an alternative definition


	- we put the 2nd fred in as a local of the first, but then the 2nd can see all
	the stuff the first has as locals

	z = 42;
	fred 1 = 12 { z = 99; }
	fred 2 = z;

	"fred 2" will return 99 :(


	we need to make "fred 2 = z" into another fred at the same level, eg

	$$alternate42$fred 2 = z;


	Nope, then how do we link the freds together for remove etc.?

	Better:

	see a new sym (fred), create it

	parse args with simple names becoming params, patterns becoming
	$$arg42 plus a local $$patt42 holding the pattern source

	expand the pattern to an access def as well, so our children can bind
	to it

	at the = sign, test for any pattern args present .. if there are none,
	carry on as before

	otherwise, make a new local called $$alternate42 or whatever and parse
	the RHS into that

	at the end of parse, need to resolve outwards twice, since we nest in
	twice

	now if we see another fred, check that the number of args matches and
	then parse in as $$alternate99

	what abut

	fred 2 a = 12;
	fred 1 b = 32;

	the first fred will make a top-level with

	fred $$arg12 a
	=
	{
	$$alernate42 = 12;
	$$patt12 = 2;
	}

	then when we parse the 2nd fred the name of the 2nd param is wrong :(


	Even betterer:

	use GLR to split off the four cases for us

	ident =
	pattern =
	ident ident_list =
	ident pattern_list =

	change pattern syntax so that ident is not part of simple_pattern

	need to change lcomp as well

	so we need to check why PARSE_PARAMS gets used: can we do without the
	params part? yes, it's used so we can edit functions, but we no longer
	do this

	these days all we need is expr I think, but we'd need a small action
	wrapper around it to wipe out any existing tree and locals



	Find_item = class
	Menuaction "_Find"
	("find a transform which will map sample image onto " ++
	"reference") {
	action reference sample = class
	Transform b reference.width reference.height {
	_vislevel = 3;

	// controls
	order = rubber_order;
	interp = rubber_interp;
	wrap = rubber_wrap;
	max_err = Expression "Maximum error" 0.3;
	max_iter = Expression "Maximum iterations" 10;

	// transform
	[a,b,c] = transform_search max_err max_iter order interp wrap
	sample reference;

	transformed_image = Image a;
	final_error = c;
	}
	}

	fails with

	Bad superclass.

	Superclass constructor
	"Image_transform_item.Image_rubber_item.Transform"
	should have no secret arguments.

	but this:

	Find_item = class
	Menuaction "_Find"
	("find a transform which will map sample image onto " ++
	"reference") {
	action reference sample = class
	_t {
	_vislevel = 3;

	// controls
	order = rubber_order;
	interp = rubber_interp;
	wrap = rubber_wrap;
	max_err = Expression "Maximum error" 0.3;
	max_iter = Expression "Maximum iterations" 10;

	// transform
	[a,b,c] = transform_search max_err max_iter order interp wrap
	sample reference;

	transformed_image = Image a;
	_t = Transform b reference.width reference.height;
	final_error = c;
	}
	}

	(ie. make the superclass constructor into a member) works fine



	- try using bison's location system

	http://www.gnu.org/software/bison/manual/html_mono/bison.html#Locations

	- add something to Symbol:

	Symbol *access_for;

	links generated access def to the $$thing4 which holds the RHS

	handy for the program window, also maybe for lcomp code gen?

	also for row edits

	- don't offer to clear temps if there's been a crash

	need to be able to test for process-still-running by PID

	try

	http://msdn2.microsoft.com/en-us/library/ms886766.aspx

	gboolean
	process_running( int pid )
	{
	HANDLE handle;

	if( (handle = OpenProcess( 0, FALSE, pid )) ) {
	CloseHandle( handle );
	return( TRUE );
	}

	return( FALSE );
	}

	- autoarrange after every column resize or move

	animate movement, so columns slither out of the way and in to place when
	you drop

	duplicate column placement would be odd: maybe place duplicate below? also
	new column? what about dropping an image on to the ws background?

	- there's some left-recursion in the parser, eg. comma_list, is this easily
	fixable?

	- add (*) (operator sections) ... need a binup / uop production? can't
	do this without losing precedence stuff, it'll need a separate production
	for sections

	- magic definition maker could make a workspace-local def, rather cool

	- test classmodel_dict_new() ... part of classmodel member automation

	need to implement member edit for OPTION groups

	classmodel_done_member() ... read widget set -> model
	classmodel_buildedit_member( ... model -> build widget set

	part of the [["key",value]] arg type

	- can we get VIPS errors reported in Error too?

	we'd need to add logging to vips I think

	- toolkits / find doesn't find builtins on their name ... eg. search for
	"im_add"

	too hard to fix with the way searching is done now

	- turn on update regions during drag, fix the x pos, try dragging, horrible
	flickering as we update twice, once after the drag motion and once after the
	recomp

	if you comment out the explicit vobject_refresh() in
	regionview_model_update() the flickering goes, but region dragging is then
	very unresponsive

	fix this when we get fast recomp back again

	- have a test_types.ws ... test arithmetic on all combinations of _types?

	- panner would be cool

	- tooltips on Expression rows always show unedited formula

	could special-case formula for things with expression RHS?

	- what about iimage, iregion, iarrow ... can we member automate these? why are
	they different?

	- can't see error indications in noedit mode

	should set a red background for display area as well as for rowview button?

	- need to be able to override cons to be able to make a List class :-(

	see reduce.c:1710

	this will change the strictness of cons ... how much breakage will this
	cause? very unclear

	try this as a quick hack

	need to do this before we can finish List

	need List to make gamma easy

	- unselected column headers are too like the bg colour on windows?

	- python uses z.real and z.imag to extract real/image, should we add this too?

	we don't really have complex as a true class, so it would be rather odd

	need to add "[a].head" etc as well

	- python blocks complex->real with casts ... insists you use .imag/.real or
	abs()

	- if nip sees a IM_RW_IMAGE argument, it could automatically do this:

	int
	im_flood_blob_copy( IMAGE in, IMAGE out, int x, int y, PEL *ink )
	{
	IMAGE *t;

	if( !(t = im_open_local( out, "im_flood_blob_copy", "t" )) \|\|
	im_copy( in, t ) \|\|
	im_flood_blob( t, x, y, ink, NULL ) \|\|
	im_copy( t, out ) )
	return( -1 );

	return( 0 );
	}

	so it would turn a single IM_RW_IMAGE arg into a paired input and output
	arg

	could make im_lineset() into a regular inplace func and rely on nip to wrap
	and unwrap

	junk flood_blob_copy

	nip could do this lazilly ... if we see the user doing

	im_line (im_line ...) ...

	then we could make one memory image and call im_line twice on it
	destructively ... cool! we'd need to check refcounts to make sure the
	intermediate wasn't being used anywhere else

	hmm! might actually be very hard, we don't have true refcounts for things in
	the heap

	need to do it on read instead:

	- for image i

	- use as an IM_RW_IMAGE arg ... copy to a memory area and pass in memory
	handle

	- return memory area IMAGE j, and set a flag saying "can operate on
	destructively"

	- if we use j as an IM_RW_IMAGE arg, skip the copy and just pass memory
	area in destructively ... we now have two ImageInfo sharing a single
	IMAGE

	- !!!!

	- does ImageInfo allow IMAGE sharing? not sure it does

	- maybe this needs to be a vips8 feature when we'll have refcounts on
	IMAGE

	- tooltip on column says which other columns items in this column refer to,
	and which columns refer to items in this column

	- how about a nip start folder common to all versions

	so nip2-7.11.14 tries

	.nip2-7.11.14/start
	.nip2-7.11/start
	.nip2-7/start
	.nip2/start

	or maybe

	.nip2/7.11.14/start
	.nip2/7.11/start
	.nip2/7/start
	.nip2/start

	bit less cluttered

	also, we could have

	.nip2/tmp

	and not have multiple nip2 tmp areas

	workspace recover after crash could break though ... maybe keep ws saves in
	.nip2/7.11.4/tmp?

	- think again about class arg checks

	is there some way we can avoid the _check overhead? or at least check less
	often

	- plotpresent/imagepresent could have a common base class with the focus stuff
	in? also kb nav, zoom, drag-scroll

	a bit difficult, because we want two different policies on window resize:
	plot should change the object to match the window

	- photographic negative should also be in image/levels ?

	no, it does ->sRGB, (255-) etc., so it's better as a filter

	- gtk+ 2.12 has a treeview widget with rectangular select and grid lines

	use instead of gtksheet?

	- stop image flickering on clock recomp?

	want background pattern to be a property of the image display widget, not
	the image?

	so we fade in tiles when that section of the image has never been displayed
	before (eg. on scroll or zoom)

	we don't fade when that section has been painted and we are just changing
	the image (eg. on recalc)

	if fadesteps == 1, only paint the sections of the tile for which mask == 255

	this way we will never paint the bg pattern

	need some hack for scroll/zoom ; test for mask == 255 would be slow :(