id	summary	reporter	owner	description	type	status	priority	milestone	component	version	resolution	keywords	cc
2456	Change the garbage collector	Martin Sjölund	Martin Sjölund	"We need to change our data structures slightly in order to achieve a better performance and perhaps remove the need for boxptr-functions (marshalling, box/unbox).

(16:22:18) Martin Sjölund: I think in general the way we tag pointers is wrong
(16:22:32) Martin Sjölund: we should tag integers instead
(16:22:33) Adrian Pop: yes, probably tagging pointers is confusing the hell out of GC
(16:22:49) Adrian Pop: we could probably not tag anything anymore
(16:23:03) Martin Sjölund: maybe
(16:23:11) Adrian Pop: as that is needed only for the previous RML gc
(16:23:19) Adrian Pop: and have 32 bit integers
(16:23:49) Martin Sjölund: !!!
(16:23:55) Martin Sjölund: that would be great actually
(16:24:19) Martin Sjölund: would probably be a good performance boost, too
(16:24:23) Adrian Pop: the only problem i think it might be with debugging
(16:24:40) Martin Sjölund: only problem is knowing what is integer and what is not
(16:25:15) Martin Sjölund: or we box integers too
(16:28:56) Martin Sjölund: I mean... allocating data could be as simple as:
void*[2] = {header,double_word}; // 64-bit
void*[3] = {header,double_word1,double_word2}; // 32-bit
void*[2] = {header,word}; // 32-bit, 64-bit
char*[2] = {headerbyte1,headerbyte2,headerbyte3,headerbyte4,char1,...,charn}; // 32-bit, 64-bit
(16:30:28) Martin Sjölund: since GC is simplified, we need less bits for book-keeping too
(16:31:22) Martin Sjölund: could just be a bitset for int|double|enum|string|list|option|record = 3 bits to tag what data it is
(16:31:32) Adrian Pop: true
(16:32:21) Martin Sjölund: runtime could be quite simple with that
(16:32:47) Martin Sjölund: oh, I forgot array as type (of course)
(16:33:14) Martin Sjölund: but still 3 bits... and arrays could have 3 bits to say what is stored in it (same 8 types)
(16:33:51) Martin Sjölund: and same for lists
(16:34:09) Martin Sjölund: lists of unboxed data could be cool
(16:35:06) Martin Sjölund: and I guess 28-bit arrays still fine length
(16:37:03) Martin Sjölund: and uniontypes with only atomic members (many of them) could be allocated atomic too...
(16:37:10) Martin Sjölund: so many things that can be done... 
(16:37:20) Adrian Pop: :)
(16:37:35) Martin Sjölund: do you think we should move in this direction?
(16:38:09) Martin Sjölund: before we rely too much on the eclipse debugger :D
(16:38:26) Adrian Pop: i think we should not do that yet
(16:38:38) Adrian Pop: wait until we move to the bootstrapped compiler
(16:38:46) Adrian Pop: and then we can do changes"	defect	accepted	normal	Future	Code Generation	trunk			Adrian Pop Adeel Asghar Peter Aronsson