view src/README.kkcc @ 4686:cdabd56ce1b5

Fix various small issues with the multiple-value implementation. lisp/ChangeLog addition: 2009-08-31 Aidan Kehoe <kehoea@parhasard.net> * byte-optimize.el (byte-optimize-form-code-walker): Be careful about discarding multiple values when optimising #'prog1 calls. (byte-optimize-or): Preserve any trailing nil, as this is a supported way to explicitly discard multiple values. (byte-optimize-cond-1): Discard multiple values with a singleton followed by no more clauses. * bytecomp.el (progn): (prog1): (prog2): Be careful about discarding multiple values in the byte-hunk handler of these three forms. * bytecomp.el (byte-compile-prog1, byte-compile-prog2): Don't call #'values explicitly, use `(or ,(pop form) nil) instead, since that compiles to bytecode, not a funcall. * bytecomp.el (byte-compile-values): With one non-const argument, byte-compile to `(or ,(second form) nil), not an explicit #'values call. * bytecomp.el (byte-compile-insert-header): Be nicer in the error message to emacs versions that don't understand our bytecode. src/ChangeLog addition: 2009-08-31 Aidan Kehoe <kehoea@parhasard.net> * eval.c (For, Fand): Don't declare val as REGISTER in these functions, for some reason it breaks the non-DEBUG union build. These functions are only called from interpreted code, the performance implication doesn't matter. Thank you Robert Delius Royar! * eval.c (Fmultiple_value_list_internal): Error on too many arguments. tests/ChangeLog addition: 2009-08-31 Aidan Kehoe <kehoea@parhasard.net> * automated/lisp-tests.el (Assert-rounding): Remove an overly-verbose failure message here. Correct a couple of tests which were buggy in themselves. Add three new tests, checking the behaviour of #'or and #'and when passed zero arguments, and a Known-Bug-Expect-Failure call involving letf and values. (The bug predates the C-level multiple-value implementation.)
author Aidan Kehoe <kehoea@parhasard.net>
date Sun, 06 Sep 2009 19:36:02 +0100
parents ac1be85b4a5f
children 3889ef128488
line wrap: on
line source

2002-07-17  Marcus Crestani  <crestani@informatik.uni-tuebingen.de>
	    Markus Kaltenbach  <makalten@informatik.uni-tuebingen.de>
	    Mike Sperber <mike@xemacs.org>

	updated 2003-07-29

	New KKCC-GC mark algorithm:
	configure flag : --use-kkcc

	For better understanding, first a few words about the mark algorithm 
	up to now:
	Every Lisp_Object has its own mark method, which calls mark_object
	with the stuff to be marked.
	Also, many Lisp_Objects have pdump descriptions memory_descriptions, 
	which are used by the portable dumper. The dumper gets all the 
	information it needs about the Lisp_Object from the descriptions.

	Also the garbage collector can use the information in the pdump
	descriptions, so we can get rid of the mark methods.
	That is what we have been doing.

	
	DUMPABLE FLAG
	-------------
	First we added a dumpable flag to lrecord_implementation. It shows,
	if the object is dumpable and should be processed by the dumper.
	The dumpable flag is the third argument of a lrecord_implementation
	definition (DEFINE_LRECORD_IMPLEMENTATION).
	If it is set to 1, the dumper processes the descriptions and dumps
	the Object, if it is set to 0, the dumper does not care about it.
		

	KKCC MARKING
	------------
	All Lisp_Objects have memory_descriptions now, so we could get
	rid of the mark_object calls.
	The KKCC algorithm manages its own stack. Instead of calling 
	mark_object, all the alive Lisp_Objects are pushed on the 
	kkcc_gc_stack. Then these elements on the stack  are processed 
	according to their descriptions.


	TODO
	----
	- For weakness use weak datatypes instead of XD_FLAG_NO_KKCC.
	  XD_FLAG_NO_KKCC occurs in:
		* elhash.c: htentry
		* extents.c: lispobject_gap_array, extent_list, extent_info
		* marker.c: marker     
	  Not everything has to be rewritten. See Ben's comment in lrecord.h.
	- Clean up special case marking (weak_hash_tables, weak_lists,
	  ephemerons).
	- Stack optimization (have one stack during runtime instead of 
	  malloc/free it for every garbage collect)

	There are a few Lisp_Objects, where there occured differences and
	inexactness between the mark-method and the pdump description.  All
	these Lisp_Objects get dumped (except image instances), so their
	descriptions have been written, before we started our work:
	* alloc.c: string
	description: size_, data_, and plist is described
	mark: only plist is marked, but flush_cached_extent_info is called.
	      flush_cached_extent_info ->
		free_soe ->
		  free_extent_list ->
		    free_gap_array ->
		      gap_array_delete_all_markers ->
			Add gap_array to the gap_array_marker_freelist

	* glyphs.c: image_instance
	description: device is not set to nil
	mark: mark method sets device to nil if dead
	See comment above the description.