[LispM] bug in hash-table-rehash?

[Greg Gilley]

But hash-tables do have an array leader:

(defstruct (hash-table :named :ARRAY-LEADER  (:constructor make-hash-array-internal) (:conc-name nil)
		       (:callable-constructors nil) (:copier nil))


> On Mar 11, 2018, at 5:32 PM, Scott L. Burson <Scott at ergy.com> wrote:
> 
> Right.  I'm not certain, but I think an empty leader probably took zero space, so the pointers would have the same value.
> 
> -- Scott
> 
> 
> On Sun, Mar 11, 2018 at 5:28 PM, Greg Gilley <ggilley at gerg.org <mailto:ggilley at gerg.org>> wrote:
> But %structure-total-size doesn’t include the leader if it’s passed the array header pointer (in this case hash-table). So the blt-typed is being handed the pointers to the leader to start the copy, but the size doesn’t include the leader if I believe the microcode comment.
> 
> 	Greg
> 
> 
>> On Mar 11, 2018, at 5:18 PM, Scott L. Burson <Scott at ergy.com <mailto:Scott at ergy.com>> wrote:
>> 
>> You may have found a bug.  Since rehashing hash tables is presumably a frequent operation, and thus this code is likely to be well exercised, I strongly suspect that in practice, hash tables never had nonempty leaders.
>> 
>> -- Scott
>> 
>> 
>> On Sun, Mar 11, 2018 at 4:24 PM, Greg Gilley <ggilley at gerg.org <mailto:ggilley at gerg.org>> wrote:
>> 
>> This is in the MIT sys99 code and the Explorer code:
>> 
>> 	((null grow) (setq hash-table (follow-structure-forwarding hash-table))
>> 		     (setf (hash-table-lock new-hash-table) (hash-table-lock hash-table))
>> 		     (%blt-typed (%find-structure-leader new-hash-table)
>> 				 (%find-structure-leader hash-table)
>> 				 (%structure-total-size hash-table) 1)
>> 		     (return-array (prog1 new-hash-table (setf new-hash-table nil)))
>> 		     hash-table)
>> 
>> In the function hash-table-rehash, if it didn’t grow the table (after a gc perhaps), then it copies the re-hashed table back to the original hash table.
>> 
>> However, it doesn’t seem to copy the whole table. if I read the comments from the microcode (this is from sys99, but seems similar to the lmi microcode):
>> 
>> 
>> ;;; Given the address of the base of a structure, return information on its size.
>> ;;; Note that if given the address of an array header, the leader (if any) is
>> ;;; not counted, but if given the address of the leader, the leader is
>> ;;; counted.  I.e. nothing before the given address is counted.
>> ;;; In the case of an RPLACD-forwarded list, the 2 words pointed to by the
>> ;;; forwarding-pointer are counted, and the forwarding-pointer itself isn't.
>> 
>> It looks as though it’s not copying the complete hash table over because it’s passing in the pointer to the hash-table and the not the hash-table leader.
>> 
>> Anyone familiar enough with the code to confirm my diagnosis?
>> 
>> Thanks,
>> 
>> 	Greg
>> 
>> 
>> _______________________________________________
>> LispM mailing list
>> LispM at tunes.org <mailto:LispM at tunes.org>
>> http://lists.tunes.org/mailman/listinfo/lispm <http://lists.tunes.org/mailman/listinfo/lispm>
>> 
>> 
> 
> 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: </archives/lispm/attachments/20180311/5eb23138/attachment.html>