Unreliability in process output

[Richard Stallman <rms@gnu.org>]

After emacs -Q, visiting src/window.c and then doing C-x v l gives
unreliable results.  Large chunks of the buffer are missing.  Which
parts are missing varies each time, but the first missing chunk seems
to start on a 4096-character boundary.

The problem still happens if process-adaptive-read-buffering is nil
and coding-system-for-read is no-conversion.  I've determined that the
variable carryover in read_process_output is always zero in that mode.

When I tried this with a breakpoint at read_process_output, and paused
for some time in that breakpoint, the text obtained consisted of the
first chunk plus a little more; nearly all was lost.

Stepping through read_process_output showed that the only calls
to emacs_read were from read_process_output, and that the text they
read was the text that appeared in the buffer.  Somehow the rest
of the text is disappearing.

Similar problems happen with an executable I compiled almost 3 years
ago.  It is not a newly introduced problem.

Can anyone else observe this problem?  Can anyone make more headway
debugging this?