Your suggested improvement was broken for empty thread objects - 
array_replace requires at least one array.

Additionally, there's significant overhead involved with 
call_user_func_array() and array_replace().  Better just to use the 
array union operator - benchmarked at 25% faster on a 500,000 message 
thread object, with the same maximum memory usage.

Changes have been made in Git (master):

commit 7d4019642f279fc5471bd330e9654a976154a280
Author: Michael M Slusarz <slusarz@horde.org>
Date:   Wed Aug 5 01:12:01 2015 -0600

     [mms] Improved performance of Horde_Imap_Client_Data_Thread 
object when containing large number of messages (Request #14075).

  .../lib/Horde/Imap/Client/Data/Thread.php          |    7 +-
  framework/Imap_Client/package.xml                  |    8 ++-
  .../test/Horde/Imap/Client/Data/ThreadTest.php     |   66 
++++++++++++++++++++
  3 files changed, 75 insertions(+), 6 deletions(-)

http://github.com/horde/horde/commit/7d4019642f279fc5471bd330e9654a976154a280

This is not a bug report, since behavior is correct.

The library in question is Horde_Imap_Client.

The class where the performance problem reside is 
Horde_Imap_Client_Data_Thread.

And the method causing the high CPU usage is _getAllIndices().

This method uses a simple iteration through _thead array (in my case 
having 8088 elements) and merges the keys of each sub-array with the 
previous one. So it calls array_merge and array_keys 8088 times:

protected function _getAllIndices()
     {
         $out = array();
         reset($this->_thread);
         while (list(,$v) = each($this->_thread)) {
             $out = array_merge($out, array_keys($v));
         }

         return $out;
     }

I replaced this iteration with this and the performance issue were gone:

protected function _getAllIndices()
     {
             return array_keys(call_user_func_array('array_replace', 
$this->_thread));
     }

Hope it helps and best regards,

Cristiano da Cunha Duarte