The changing of legend properties, particularly with set(), is now much faster.  It is the code that uses subplots that is particularly slow.  I extracted a benchmark test where I turn on profiling.

profile clear
clf;
x = (1:5)';
subplot (2,2,1);
 plot (x, rand (numel (x)));
 profile resume
 legend (cellstr (num2str (x)), "location", "northwestoutside");
 profile off
subplot (2,2,2);
 plot (x, rand (numel (x)));
 profile resume
 legend (cellstr (num2str (x)), "location", "northeastoutside");
 profile off
subplot (2,2,3);
 plot (x, rand (numel (x)));
 profile resume
 legend (cellstr (num2str (x)), "location", "southwestoutside");
 profile off
subplot (2,2,4);
 plot (x, rand (numel (x)));
 profile resume
 legend (cellstr (num2str (x)), "location", "southeastoutside");
 profile off

Using profexplore, it is clear that only two functions (update_layout_cb and axes) are responsible for most of the time consumed.

Top
  legend: 4 calls, 0.545 total, 0.006 self
    1) legend>update_layout_cb: 4 calls, 0.354 total, 0.004 self
    2) axes: 4 calls, 0.118 total, 0.001 self
    3) __go_text__: 4 calls, 0.021 total, 0.021 self
    4) legend>add_safe_listener: 28 calls, 0.014 total, 0.002 self
    5) legend>parse_opts: 4 calls, 0.010 total, 0.003 self
    6) legend>addproperties: 4 calls, 0.009 total, 0.002 self
    7) legend>update_numchild_cb: 4 calls, 0.003 total, 0.001 self
    8) setappdata: 8 calls, 0.003 total, 0.002 self
    9) addlistener: 64 calls, 0.002 total, 0.002 self
    10) ancestor: 4 calls, 0.001 total, 0.001 self

Within update_layout_cb, it is the function textitem_objects and set() that take the most time.  It might be possible to coalesce the set() calls in to fewer invocations (< 56 currently) which might help.

  legend: 4 calls, 0.545 total, 0.006 self
    legend>update_layout_cb: 4 calls, 0.354 total, 0.004 self
      1) legend>textitem_objects: 4 calls, 0.177 total, 0.019 self
      2) set: 56 calls, 0.121 total, 0.087 self
      3) legend>update_legend_position: 4 calls, 0.033 total, 0.003 self
      4) legend>textitem_data: 4 calls, 0.013 total, 0.004 self
      5) setappdata: 8 calls, 0.002 total, 0.001 self

Within textitem_objects, I see

      legend>textitem_objects: 4 calls, 0.177 total, 0.019 self
        1) __go_line__: 40 calls, 0.070 total, 0.070 self
        2) linkprop: 40 calls, 0.038 total, 0.026 self
        3) __go_text__: 20 calls, 0.030 total, 0.030 self
        4) setappdata: 40 calls, 0.009 total, 0.006 self
        5) get: 76 calls, 0.004 total, 0.004 self
        6) legend>update_marker_cb: 20 calls, 0.001 total, 0.001 self

There are 10 calls to _go_line_ per invocation even though there are only 5 legend objects.  Maybe we don't need to place a marker object if there is no marker to begin with?


(file #47902)

This bug is still present in the current dev 4.3.0 (39f39eb4e476).

I filed task #14243 about rewriting legend.m

What graphics toolkit was used for benchmarking?