I attach my results for the same builds as yours. Tried to calibrate
the running time and I ended up running really long tests (quite
exhausting to move mouse every few minutes as my screensaver is broken
and always turns on after few minutes). Will try to calibrate again
later, and I also am running the same tests on a ubuntu netbook. Few
conclusions:

> Running [zsh-5.0.7]:                     string_test      48631,87
> Running [zsh-5.0.8]:                     string_test      86302,14

This says there something happened after 5.0.7 and reveals itself on
OS X, could try to track it with bisect.

> Running [zsh-5.1]:                       string_test      86698,42
> Running [zsh-head-8e9a68a]:              string_test      38887,54
> Running [zsh-36926]:                     string_test      86497,62

This says 36926 doesn't include string boost up (at cost of memory
use) that current head has.

> Running [zsh-5.1]:                       search_test      51321,70
> Running [zsh-head-8e9a68a]:              search_test      10071,26
> Running [zsh-36926]:                     search_test      10260,60

This says that pattern maching boost up is retained.

> Running [zsh-5.1]:                       function_test    49070,83
> Running [zsh-head-8e9a68a]:              function_test    56299,14
> Running [zsh-36926]:                     function_test    54456,34

This informs about slowdown caused by newheaps patch, and points that
36926 can make it smaller.


I also did memory tests:

> # Tests for zsh-head-8e9a68a
> # string_test 42, 182, 318, 457, 590, 720, last: 10
> string_test 317.8

This informs about the 36834 problem with memory.

> # Tests for zsh-36926
> # string_test 2.1, 6.6, 8.6, 10, 10, 11, 9.5, 11, 11, 11, 11, 11, last: 9.6
> string_test 9.3

This shows it's solved.

> # search_test 110, 215, last: 265
> search_test 110.0

This is the high memory usage when matching patterns in large arrays
you mentioned.

Best regards,
Sebastian Gniazdowski