serious subprocess output bug in 21.4, 21.5

Hi, In all of 21.4.6, 21.4.12, and 21.5-b11, I see a serious bug in capturing the process output in pcl-cvs. Running "xemacs -vanilla" if I run M-x cvs-examine in a CVS checkout directory (cvs is using the :ext: method if that matters) then hit "=" to diff a file with many hundred or thousand lines of changes, there will typically be (4096 + offset to next newline) byte chunks missing in the diff output starting at byte offsets that are multiples of 4096. The bug did not exist in 21.1.9, even running the same exact version of pcl-cvs. The problem is much worse when the machine is loaded. I am running RH 7.3 on a dual 2.8GHz P4 box with 4Gb ram. I am trying to isolate this to a simple reproducer, but wanted to send this out now in case this report is enough for whoever has been hacking on the subprocess code since the 21.1 days to figure out the bug. thanks greg