# "NOTES"
#	$Id: NOTES,v 1.19 2007/07/18 23:28:13 fang Exp $

Developer's Notes (transcripts of fangism muttering to himself...)

//=============================================================================
2007-07-10:
	on relaxed template types: discussion
	Q: should relaxed template types be allowed in port declarations?
		In other-words, can the type of a port alias be delayed
		from the top-down?
	A: No.  There is a fundamental problem with delaying the type binding
		on public ports.  
		Since each complete type needs to define a fixed footprint, 
		the types of the ports cannot be further delayed.  
		Connections: delaying type completion of ports would
			result in a mutual dependence between levels of
			hierarchy.  
		We will, however, allow relaxed typed ports whose parameters
			are bound by the definition (internals).  
		Thus, ports, like any internal subinstances must be fully 
			type-bound in well-formed definitions.  
		Note that this means, we never need member expressions as 
			references in type-completion declarations, since
			subinstances of hierarchical types are guaranteed
			to have complete type.  Nifty.  

	Phase ordering between relaxed type-binding and aliasing:
		presents difficulties in instantiations and connections
		since instantiations can now happen *after* aliases
		are formed.  Resort to hackish patches to address.  :(

//=============================================================================
2006-11-14:
	more on channels:
	can directional (send-only, receive-only) channels be declared
		in any scope or only in ports?  Doesn't make sense for
		ports to have direction in scope, since channels
		cannot dangle, must eventually connect up.  
		Proposal: reject scope-local directional channel decls.  

	Defect report: original grammar never allowed arrays in channel's
		data type specifiers, e.g. chan(bool[2]) rejected. 
		This was accidental, and arrays should be allowed.
	enqueue TODO list.  

//- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
2006-11-14:
	channels: for now, enforce 1-to-1 sender-receiver connectivity
		and reject shared sending/receiving, and dangling channels.
	Future: provisions for shared channels (requiring use to 
		guarantee exclusive use)
		semantics: allow multiple senders to connect iff all
		participating senders (aware) agree to share.
		Connecting to a non-sharing terminal is rejected.
	Syntax proposal: something like ?? and !!

	connection checking: a channel send-only port are considered produced
		if it appears as a sent-literal in some CHP action
		in the CHP body.  In other words, the CHP body is like a
		process whose interface is given in the way channels
		are referenced (send vs. receive).  Connection checking
		should account for the CHP definition.  

	CHP: should there be a top-level CHP body?  Currently, syntax 
		forbids, but internal representation allows. 

//=============================================================================
2006-11-12:
	CHP: bool vs. bool (discussion with Rajit)
	Q: need to distinguish betwen physical 'bool' and abstract datatype
		bool, which is synonymous with int<1>.  
		e.g. ambiguity of "defproc blah(bool a) ..."
	Considering a 'dbool' typedef to int<1>.  
	However, would like CHP to evaluate expressions and manipulate
		physical bools directly, since there's no interface-to-
		implementation mechanism in-place yet.  Eventually, will be 
		able to "deftype dual_rail <: bool" to describe 
		physical implementations.  
	Semantics for CHP operations on physical bools is as expected,
		nothing tricky.  
	Rationale: want to punt implementation/interface work, but still
		allow CHP to handle physical bools directly.  

//=============================================================================
2006-05-14:
	CHP footprint implementation:
	is it mostly a copy of the source representation?
		Yes, so we use same class hierarchy, but unrolling
		results in substituting meta-values with constants.  
	Does CHP have any parameter dependents that aren't already 
		taken care of in type-unrolling and object-instantiation?
		Proposed a metaloop-selection, which could be expanded, 
		but would rather punt on it.
	The purpose of such footprint unrolling is to finish compile-time
		evaluation of meta-parameter dependents to fixed constants.  
		e.g. x[i].  'i' could be a meta parameter (constant).  
	Status: done (2006-06-25)

//=============================================================================
2006-05-11:
	Re: Global parameter value lookup
	Temporary workaround to allow global meta parameters to be looked
		up during unroll value resolution.
	But what namespace is chosen as the start for parent lookups!?
	Currently, the namespace that is parent of the TYPE being
		instantiated is looked up, NOT the namespace in which
		the type is INSTANTIATED.  
		See implementation in 
		process_definition::unroll_complete_type and 
		::create_complete_type.
	Read that last statement one more time.  
	This resembles argument-dependent lookup, a.k.a. ADL.  
	The underlying problem is that unroll_context doesn't really
		support namespace lookup yet.  
	TODO: need well-defined lookup semantics and bindings.  

//=============================================================================
2006-05-10:
	Brief discussion about supporting nonmeta subscripts in CHP.
	Conclusion: must have nonmeta-value (int<>) subscripts.
		subscripts are an implicit decode (one-of-many).
	Status: CHP sublanguage current constrains instances references
		to be meta-references (compile-time only).
	However, CHP intermediate representation might be more restrictive,
		and may benefit from having static selection graphs.  
	Strategies: internal expansion of nonmeta selections
		into meta-selection equivalents.  How?
	Concerns: range checking
	Update (2006-06-xx): no longer restricted to meta indices!

//=============================================================================
// forgot about this file for several months...
//=============================================================================
2005-09-01: ARTXX-00-01-04-main-00-55-footprint-01-32:
	Notes on algorithm for replaying internal port aliases:
	had: detect whether or not instance in question is a 
		a direct member of a port, and if so, replay the connection.
		(in instance_alias_info::replay_internal_aliases)
		WRONG: this misses *aliases* of ports.  
	then: tried to replace with instance_alias_info::retrace_port_alias, 
		if the instance in question (or any of its parents)
		are aliased to a port, then replay its connection.
		WRONG: this only finds the first port connection, 
			but an instance may be a public port to *multiple*
			types, and thus ALL port-aliases must be replayed.  
	SOLUTION: replay all port-aliases at all levels of instance 
		hierarchy, i.e. given x.y.z, replay aliases
		of the .z level and those of the .y, and then the .x level.  
		Is upward recursive in hierarchy.  
	Q: is there a top-down only solution, since replay already traverses
		top-down?  it sucks to have to inspect upward...
		A top-down solution would involve playing aliases
		between subinstances of aliases explicitly because
		x.y (x port) could alias to z.y (z internal), which
		could then alias to another port -- only discovered 
		by upward inspection.  
	TODO: save a compact record of internal aliases for each
		complete type and replay from that without having
		to check the same hierarchies over and over.  
		(like instance reference connections)
	UPDATE (2005-09-02):
		above is STILL insufficient.
		consider w.x.y whose formal alias is W::x.y 
		and aliased to port W::p,
		X (y, z) has X::y = X::z, and W contains
		X x,a,b; with x.z = a.y, a.z = b.z;
		W::q = W::b.y;
		Then p is alias to q, but only discovered through
		visiting up and down aliases...
		Thus we need a traversal to visit ALL equivalent aliases
		that walks both up and down hierarchies because
		connections are not flattened.  

//=============================================================================
2005-08-25:
	In final phase of creation checking:
		every unique instance whose collection type is relaxed
		must have relaxed actuals.  
		Definitions however are allowed to defer this because
		relaxed actuals may be passed in from the public ports.  
		Top-level instances however, MUST meet this criterion.  
		Intelligent: if instance placeholder is not aliased
		to a public port, and it is missing, then can complain.  


//=============================================================================
2005-08-20:
	Considerations for unrolling, in the grand scheme of things.  
	Top-level unrolling alters the top-level instance_collections
		that are scattered throughout the namespace hierarchy.  
	Definition-level unrolling CANNOT alter the instance_collections
		that reside in the definition's scopespace because they may be 
		shared by types with different template parameters.  
	Each definition contains a map of footprints.  
		Unrolling should work with instance_collections that belong to
		a particular footprint.  
		This means, we need to replicate the instance collections
		in each footprint??  (The ones in the definitions are
		really skeletal placeholders for unrolling.)
		But what about the top-level?  It references globally 
		accessible instances, that are NOT in the same namespace.  
	Now it seems we may need to extend *unroll_context* to point 
		to the correct footprint-scope:
		if NULL, then we are at the global scope, no further 
		resolution is needed.  
		otherwise, we are in some definition scope, so
		all instance_references to the base definition's 
		instance_collections must be resolved to the 
		particular complete_type's footprint instance_collection.  
		The top-level (modules) footprint will just have an 
		empty instance_collection map.  
	Will need to modify simple_met_instance_reference::unroll_*
		their inst_collection_ref member (instance_base reference)
		will need one additional level of translation.  
	NOTE: this idea was noted in the comments of class footprint
		some revisions ago, though I don't recall having written it.  
		Already contains an instance_collection_map_type
		for the purpose just described.  
	LATER: what about references to global instances in local scopes?


//=============================================================================
2005-08-16:
	extension of 2005-08-07 entry:
	We're ready to unroll and create types encountered during hierarchical
		creation and allocation.  
	Some issues stand in the way:
		Relaxed template actuals may only exist in one place in the
		connection hierarchy, and may not even be in the same alias
		ring.  
		We need to guarantee that: one of the equivalent aliases
		(and only one) contributes relaxed alias actuals
		to the allocated instance.  We *must* detect conflicts
		across the hierarchy.  (This will get complicated once we
		glue instances and ports.)
	Proposal: process/chan/struct instance state will contain an optional
		relaxed const_param_expr_list ptr set from alias_actuals.  
		Need one pass over all instances with the same ID to collect 
		these relaxed actuals and check for equivalence and conflicts.  
		(even replace equivalent pointers as an opt.!)
		(No need to back-copy the established pointers to the 
		alias, the alias will just reference the state's
		actual parameters once its state ID is set.  
		No need to clobber the pointer either, it's just a reference
		count that doesn't take any additional storage.)
		(get_type: if ID set, use actuals if exists, else use own)
		Note: at each level of the hierarchy, psuedo-allocated 
		(placeholder) instances are allowed
		to be missing relaxed actuals, because they may be connected
		to another instance in a different level.  
		TODO: run this check first.
		Need one final check that combines the hierarchical
		footprints into flat global footprint.  
		At THAT point should we guarantee that every instance
		as satisfied (exactly once) unique relaxed actual values.  
	Hierarchical complications: additional aliases may occur due to 
		internal connections of the instantiated type --
		those checks must be deferred.  
		At each level of the hierarchy, just check for conflicts.  
		Call this from canonical_type and module,
			passing in the footprint&.  
	First, add relaxed actuals pointer to process/chan/struct
		state instances.  
	Q: if/when should relaxed actuals be propagated around
		rings of aliases?
	A: Earliest possible time is during 
		instance_reference_connection::unroll.
		The lastest point should be when state is allocated
		during the creation phase.  How to do this efficiently?

	Important comment (as if the others aren't):
		When comparing and propagating actuals during connections:
		if the two entities being compared have NULL actuals
		(where they are eventually needed, as required by the type), 
		then we should merge the two NULL-actual alias rings together.
		Even if they are connected elsewhere hierarchically.  
		WHY?  Because when we discover hierarchical connections, 
		we need to propagate known actuals to all known aliases.  
		If 'connected' null-actual-rings were kept separate, 
		then when they are merged with known actuals, 
		they will not propagate maximally; there will be isolated
		rings of NULL actuals that should NOT be so.  
		See CRITICAL comment in "Object/inst/alias_actuals.tcc":
		instance_alias_info_actuals::__symmetric_synchronize.

//=============================================================================
2005-08-08:
	In creating unique state, need to process the definition
	of the thing being instantiated to determine internal connections.
	Each complete type will have a pseudo state template
		used to check internal connections when creating
		external state.  

	CFLAT format notes:
	cflat -connect:
		"x" "y"
		where x != y, and x is the canonical name
		quotes requires to be safe.  
	cflat -prsim:
		adds production rules, literals also quoted
		parenthesization optional

//=============================================================================
2005-08-07:
	KNOWN ISSUE: need to check that all connected aliases (during or after
		unique-creation phase) have valid and equivalent
		relaxed template actuals.  Unrolling is insufficient to
		catch errors, mismatches that occur across the hierarchy.  
	TODO: fix crash when loading empty object file (easy).

//=============================================================================
2005-08-05:
	Unique creation: allocating unique state to aliases.  
	Considerations of implementation:
	Want state to be persistent (save/restorable)
		yet don't want each state instance to be managed
		by persistent object manager (each entry is costly).  
		Resolution: use a list_vector as an indexed pool of state
			(per type).
	Instead of keeping a reference count pointer per alias, 
		just keep an index into the pool of the appropriate type.  
		Instance-entries still have back-reference to canonical
		hierarchical instance alias.  
		(Reserve 0 as 'uninitialized')
	But who owns these state pools?
		Shouldn't be global UNLESS there is a way to
			save and restore all the state through module.  
			list_vectors (and their) instances are
			copy-constructible, so this is feasible.  
		Could belong to the module, since that needs to be 
			kept persistent.  Create a 'state' member.  
		Could belong to particular definition, since those
			must outlive their instances.  
		Each type could have its own pool.  
			relaxed or strict?
		pool could be public static member of instance_type
			and referenced by the module class.  
		Clear pool upon expiration of module?
	Tentative resolution: global static member pools of instance_types.  

//=============================================================================
2005-07-18:
	Minutes of meeting.
	cflat (backwards compatibility) has some constraints:
		some programs require prior canonicalization of
		unique instances, must use the same name to refer
		to same node.  
		Name canonicalization may only occur after unique
		creation (instantiation).  
		Therefore, creation must be done first.  
	For unique creation, need to play back instantiations
		after connection pass (unroll).  
		Visit each top-level instance and check whether state 
		has been created/allocated.  If not, then allocate.
		Then recursively do for all ports and members.  
		When check finds already created, establish
		the pointer to the same memory.  

//=============================================================================
2005-07-17:
	BUG to fix:
	Passing template parameters down to ports:
	e.g. process/059,060.in
	TERMINATED. (2005-07-18)

//=============================================================================
2005-07-14:
	Do built-in channel types have ports?  Should they be expanded
		recursively like other structured types?

//=============================================================================
RESOLUTIONS (minutes of 2005-07-12 meeting):
	(addressing  2005-07-11 entry)
UNROLL (connect) phase:
	Queries:
	Are x and y connected? (somewhere in hierarchy)
		Lemma: If super instances a, and b are explicitly aliased, 
			and subinstances a....x, b....y are somehow
			connected, then then all intermediate types
			from top to bottom must match.  
		Result: checking for connection requires upward checking
			of x and y (and their aliases') parents WHILE
			their types match.  As soon as parents' types
			mismatch, can deduce non-connectivity.  
	What is x connected to (all aliases)?
		print aliases of itself and all parents' aliases.
		(potentially multiplicative in size)
	Graph connectivity: are two instances connected by anything?
		Are any public ports of the two instances connected?
		(recursive in public ports)
	Subinstance implementation:
		Upon instantiation, can recursively expand sub-instances
		of ports down to leaf types, or defer until
		member is actually referenced (or ports connected), 
		a mere demand-driven optimization.
		Should be able to expand ports recursively regardless
		of relaxed template actuals because no ports may depend
		on relaxed actuals.  
		Private internal instances will not come into play until
		unique creation.  
	Any need to "play" local internal connections in this phase?
		NO.  Only needed during unique creation.  

cflat: will be purely text dump traversal, as originally planned.  

UNIQUE (create) phase:
	NOTE: will probably need one final type check for references
		and aliases that involve relaxed actual parameters
		because some checks are deferred during the unroll stage.  
	Recursive traversal of aliases to establish object uniqueness, 
		can be identified by e.g. a memory address of residence.  
	Allocation of memory for state information. 

SIMULATE operations and issues:
	Establish fan-in and fan-out relationships.  
	Don't want to replicate expression objects, instead have unique
		instance reference expression of definitions.
	(prsim) expression memoization
	expression templates for prs in definitions
		each production rule will map literals to 
		positions to be used in the...
	translation tables from local references to global unique IDs:
		unique instances will point to (OR-combination)
		of expression templates.
	
//-----------------------------------------------------------------------------
2005-07-11:
	Instance alias connection considerations:
	Suppose type A has public members b and c.
		A x, y;
		If b and c are connected (internally), then what exactly 
		happens when x = y?
	How to determine whether or not two instances are connected?
		Two instances referenced may be deep in the subinstance
		hierarchy, and they connect if *any* of their parents 
		are aliases.  But what an efficient way (<O(N^2)) of 
		checking for aliasing parents?
	WHEN are sub-instances expanded?
		Do they require type to be complete?
			(Theoretically, no, because instances should not 
			depend on relaxed actual parameters.)
		Sometimes types are never completed because they are 
			merely relaxed-typed aliases (in arrays, e.g.).  
		Create upon instantiation (regardless of relaxed actuals)?
		Expand public ports only? Defer private and internals?
	cflat -- want it to flatten connections in the hierarchy?
		Or just some sort of hierarchical traversal to 
		dump connections in some format?
	Preventing mutual recursion/inclusion between types requires
		some partial ordering between definitions.  
		Or simply require definition before use/instance declaration.  
	Procedure: check definition for existing type


//=============================================================================
2005-07-05:
	Enforcing that things that shouldn't depend on relaxed formal
		parameters... don't depend.  
	Difficulty: what if dependence is established through
		expression assignment? e.g. pint x = y, where y is relaxed.
	Difficulty: what if is done through arrays and collectives?
		(don't worry until complex aggregate expressions)
	Proposal: need mechanism for tracking origin... for EVERY
		value that may exist!  That could get expensive fast...
		Store extra 'poison' bit with each const_param, yuck???
		What about const-collection? pair<value_type, bool>?
		That's going to f*** up the binary format once more...
		But this will make room for more value attributes.  


//=============================================================================
2005-07-01:
	resolved CHP questions, answers inlined in the 2005-06-11 entry.

//=============================================================================
2005-06-27:
	TODO: clearly define standard lookup and name resolution rules.  
	Shadow warnings:
		If a template/port/loop parameter shadows another
		available identifier, issue warning (compiler option)?
	What about enum member names?  need special lookup rules?
		(application: CHP enum equality comparison)
		If id_expr used as an expression, then look-up as
		an expression or enum...


//=============================================================================
2005-06-23:
	Major goal: port and member instance connections (unroll-time)
	RESOLUTION (2005-07-01):
		use single unified pass, 
		don't worry about multifile instance linking, 
			we may drop support for that idea completely
	requires (meta) member_instance_reference unrolling, which
	requires complete-type reference/definition unrolling, which
	requires principle definitions to keep track of what template
		parameters have been used to complete type-references.
		(type-reference need non-const definition handles)
		Propose to keep some sort of map/repository of unrolled
		types and their substructure layouts.  
		Repeated instantations of same type should be able to
		trivially initialize with the same 'footprint'.
		(Far Future: each complete type should back-point to the 
		template (general, or specialized) used to generate it. 
		This is especially important for checking partial-order
		specialization consistency.)
	Substructure_manager / offset_map:
		(see also 2005-04-28 entry)
		map of member name to some index/offset.
		mapped type is some pointer to what?
		e.g.: how should "x.y.z" be resolved?
			lookup top-level instance x, 
			lookup member y, obtain offset to y.
			Use y's offset to get reference to the actual 
				sub-instance x.y.
			lookup member z of y, obtain offset in y.
			resolve instance z of y.  
		Need to have a public (port) section, 
			and a private internal section.  
			Basic meta language restricted to public ports,
			whereas tool-specific languages *may* use private.  (?)
			(Everything in top-level is default public.)
			e.g. params cannot be in ports, thus must be private
		entries of map contain pointers to index_maps for arrays.
		Index_maps are are similar to the implementation of the
			instance_array sparse collection class.  
			instance_array is currently constructed on the fly, 
				where offsets are not established.  
			Index maps will have established indices after
				a complete unroll of the instantiation
				statements in the scope.  
		Keep around reverse offset <-> key-index map?
		What kind of pointers?  count_ptr?
	Unrolling revisited: proposing 3-pass approach:
		1) meta evaluation only -- instantiation and value assignment
			of local meta parameters only.  Since meta-parameters
			may depend on other meta parameters, both instantiation
			and assignments have to be integrated into the 
			same evaluation pass.  Need to distinguish value
			instantiation statements from physical instantiations.  
		2) meta instantiation only -- creation of physical instances
			or sub-instances.  During this phase, meta types
			and values have been completed, so only
			completed types remain.  Recall, each 'type' refers
			back to a principle definition base with 
			template actuals.  The first time a specific
			type is instantiated, the completed definition
			goes through a definition-unroll to generate the
			substructure map (for the given set of parameters).  
			Subsequent instantiations of the same type will
			reuse those maps for fast instance creation.  
		3) meta connections only -- establishes physical connections
			between created instances.  
		advantages: can create static sparse collection map
			to translate offsets.  Provides static offset
			translation maps to the connection phase.  
		disadvantage: makes tracking proper instantiation before use
			more difficult because of decoupling.  
			(Should we not consider them errors?
			out-of-order declaration-uses?)
			Consider re-specifying language?
	What if type is relaxed, like a relaxed array (port) declaration?
		Technically, type is incomplete, unless we constrain
		relaxed types (in same family) to have same public ports.  
		But then their private internal members may be completely 
			different.  Tools that use private member references
			will require a separate instance checking phase...

//=============================================================================
2005-06-11:
	Integer types: implicitly UNSIGNED.  
	Assignment semantics: x := y;
		impose the lvalue's type on y, i.e. infer what y's type
		should be.  
		This should take care of widths of integer constants
		in rvalues.  
		(What about constant indices?)
	Need well-defined semantics of data (int, bool) operators
		in the CHP context.  
		e.g. what does int<X> + int<X> return?
			whatever is being expected on the left.  
			TODO: optional warnings of truncation.  
		what does int<X> + int<Y> return?
			the max width, implicitly 0-padded because unsigned.  
		what does int<X> * int<X> return?
		Gut feeling: for now types must match, and return same type
			for arithmetic operators.
		What about comparisons?  operands must match?
			No, just pad to the wider.  
		Provide sign-extensions definition?  built-in?
			Since ints are unsigned, pad with 0's.
		Truncation operations?
			Possible warning.  
	Do pints implicitly have width 32?  (For now assume yes.)
		Moot.  Width is inferred from lvalue.  
		How is it possible for them to mix with other widths?
	"bits" type, bit fields of integers (not the same as bool).  
		Proposal: implement as public ports of the int built-in
			intrinsic definition, like:
			deftype int <: (bits b[W]);
		Resolution: Accepted.
	Logical and bitwise operations:
		overload '&' and '|' for both bools (logical) and 
		ints (bitwise) operations.  
	For data-ranges [x..y] where x,y are run-time values, 
		do we place any restrictions on the type (width) of x,y?
		must x <= y?
		Don't feel good about have size of expression be unknown...
			very ugly semantics, tons of run-time checks...
		Resolution: Reject range expressions in CHP.  
	Non-scalar data/expression references in CHP context?
		Currently, not allowed.  
		Resolution: agreed, not allowed.  
		Need to extend type-references (of built-in channel type)
			to include (unnamed) packed arrays.
		Not difficult, just low priority for now.  
	Support for dataless communications?  X!, Y? ?  just 'X'?
		Resolution: accepted, also be able to declare
			empty chan() types, allow dropping parens? ()
	CHP language:
		',' and ';'
		Can we allow ',' between actions other than communications?
			Currently grammar doesn't support it.  
		Resolution: accept generalized compositions, 
			and parenthesize precedence.  
	Grammar:
		dark corner: guard expressions starting with a + or - b
			must be wrapped in parens, after a transformation
			to eliminate the s/r conflict on x+.

//=============================================================================
2005-04-29:
	We had a VLSI group meeting on "the big picture of tools."
	Concern: certain synthesis-class tools will want to 
	automatically generate or modify definition and instance objects.  
	How will templates complicate things?
	Think about the hooks...

//=============================================================================
2005-04-28:
	Tackling structure definition layout:
	Definition-identifiers are really references to
	concrete definition collections (in the case of templates), 
	corresponding to content layout.  
	Regarding specializations:
		I think it would be best if we avoid specializations
		at this point.  However, to plan for them in the future, 
		we should make sure that the shortcut we choose does
		not preclude partial and full specializations of definitions.  
	Regarding the hash_map of identifiers local to a definition:
		For now, without specialization is ok as is.
		Later, each specialization definition will need its own map.  
	Note however, that instantiation may be conditional, so
		each concrete definition should have its own layout map.  
	Generate these layout maps on-demand during unroll/instantiation.  
	When considering template actual parameters, 
		strict and relaxed parameters are considered equal
		and part of the distinction process.  
		It is possible for relaxed parameters to affect the
		layout of a concrete definition.  
		Consequence: many redundant and identical layouts kept...
			(this is unavoidable for semanctic correctness)
	TODO soon: update parser to accommodate relaxed and strict parameters.

//=============================================================================
2005-04-10:
	front-end:
	parser: multiple-parsers, possibly multiple-lexers
		use .yy.m4 or (cpp) for grammar snippets
	modular parser-functors
	import vs. include? support cpp directives such as #file line?
		file stacks

//=============================================================================
2005-02-17:
	TODO
	Target: 0.1.5
	After improving parser... (after second convergence -00-01-04-main-01)
	Must re-do object stack for type-checking!
		Make one stack for values and parameters, 
		another stack for physical instances, 
		and let the context restrict which one may be used!
	{chan, proc, bool, int, enum, struct, param}
	port-like connections require a generic stack, though...
	3 classes: alias, port, assign (expression)
		their uses should never mix!
	During alias-parse/type-checking, one type of list
		should "lock-out" the use of the others.  
	This will eliminate the need for any generic object list!
		And perhaps generic "objects" in general!
	In general, context interface could be more like stacktrace 
		interface: automatic stack-like management upon destruction.



//=============================================================================
2005-01-05:
CRITICAL BUG: util/memory/pointer_classes:
	some destructors (excl_ptr) use release instead of reset, 
	should be reset, HOWEVER, changing to reset() causes type-check
	phase (perhaps others) to die horrible deaths.  
Currently, leave it incorrect as release(), because in the middle of 
	other modifications (need isolation).  
	Meanwhile, littered with debug code...

Added FIX_DEBUG predefine-style switch to pointer_classes.h, 
	need to isolate the problem.  
	gdb stacktrace shows functions called with an impossible call stack!
	lookup_namespace inside check_null_template_arguments !?!?
	memory corruption?
	Worst case, may need to trace EVERY pointer_class object...
	FIXED: is safe to remove this conditional macro (2005-01-11)
	Even if ONLY sticky_ptr is fixed, there is still problem!
	Problem was failed ownership transfer leading to 
		premature deallocation, and implicit default copy-constructor, 
		which gives wrong behavior.  
	Produce (emulate) test cases?

Time to use stacktrace!

Auto-configure debug mode vs. develop mode vs. release mode

test suit intended for develop mode or release mode

//=============================================================================
2004-12-12: Regarding unroll-time parameter resolution.

Dependencies: unroll-time type-resolution, with templates, 
	simple example being built-in datatype int<width>

Issues: need to resolve into constants, or detect error
	What kind of parameter objects do template arguments need?
	Something that can hold list of arbitrary collections, 
		guaranteed to be densely packed, but need not be implemented so.
	Currently, const_param_expr_list can only hold scalars.  
	Need an abstract placeholder besides const_param, 
		perhaps const_param_collection?
	However, we already have param_instance_collections, 
		as sparse multidimensional maps.  
	Would seem a waste to implement yet another set of arrays...

Observations:
	Assignment was already implemented as passing around flattened 
		lists of values while retaining dimensions requirements.  
		Seemed like a quick short cut to get things to work.
		We might be able to reuse this idea when "assigning"
		type parameters.  

Prerequisites: (what do I need to do...)
	For typedefs, may need a parameter value context stack.

Idea:
	An extension to passing around lists, we can implement as a
	dense vector that's arithmetically addressed, and will be 
	extensible to any number of dimensions.
	Indexed with any multikey.
	Also make a reference-array for indirection.  

//=============================================================================
regarding multikey_qmap:

	Instead of deriving in diamond-inheritance, 
	specialize multikey_qmap as a special case of multikey_map.  

//=============================================================================
