In that case, you might want to include some threshold on array size which approximates the cost of the stack scan.
When you're working on small arrays, it sounds like the overhead of the check would exceed the time saved by the in-place mechanism. This heuristic doesn't need to be super accurate or anything, but should at least some of the time give an order-of-magnitude speedup (assuming it's worth doing at all). Thanks, -- Raul On Tue, Jun 21, 2016 at 4:52 PM, Henry Rich <henryhr...@gmail.com> wrote: > There is a usecount that indicates the number of names that refer to a > block. This is incremented when there is an assignment. > > There is also a stack (called the NVR stack) that is updated every time the > parser moves a name from its queue to its execution stack. The NVR stack is > used to defer freeing a block: consider > > (a =: 1) [ a > > right-to-left, the value of a is stacked, then a is reassigned. This would > free the value of a, which is disastrous because a pointer to it is on the > stack. So assignment always runs through the NVR stack to see if there is a > reference to the value about to be assigned. If so, the earliest such one > is tagged as representing a deferred free; the old value of a is then not > freed until the sentence containing the deferred free completes. Note that > the reference to a could have come in a higher-level sentence. > > I would fix the problem we are discussing by having the code that is about > to do an in-place assignment run a check on the NVR stack to see if the name > is in use in a higher sentence (it will always be in use in the sentence > containing the in-placeable assignment). If the name is in use, the > assignment would not be performed inplace. This is fast, since inplace > assignments are rare. > > Henry > > > On 6/21/2016 3:48 PM, Raul Miller wrote: >> >> Do you think it's faster to do a stack scan on each in-place update >> than to update an "in play" count each time that stack gets updated? >> >> Thanks, >> > > ---------------------------------------------------------------------- > For information about J forums see http://www.jsoftware.com/forums.htm ---------------------------------------------------------------------- For information about J forums see http://www.jsoftware.com/forums.htm
