This is text comes from Mark Adler


The problem was that my previous check for an incomplete code set was  much
too liberal.  It allowed the passage of pathological code  descriptions,
such as a single code of length 10, like what's in  zlib-testcase.gz
(twice).   The deflate spec allows for distance codes  with a single code,
but in that case that single code must have length  1.  The patch below now
permits incomplete codes only if they consist  of a single code of length
1.   Though the spec only mentions distance  codes in this regard, my patch
also allows literal/length codes with a  single code, which would permit a
dynamic block that only defines the  end-of-block code.  This should be
permitted by the spec, and it won't  crash the code, so I'm letting it
through.

The consequence of letting pathological code descriptions through was  that
filling out the decoding table with invalid code markers overran  the
ENOUGH sized array and overwrote some other stuff in memory  (whatever
happened to follow the inflate state).  I thought that I had  prevented
that with a check on how much of ENOUGH was being used by the
literal/length code, and knowing the maximum that could possibly be  used
by a distance code based on an exhaustive search (MAXD).  However  my
exhaustive search only included valid distance code sets, and so
zlib-testcase.gz overran MAXD, after a similar pathological  literal/length
code set had used most of ENOUGH - MAXD.  The patch now  only allows valid
code sets through, which are either exactly complete  code sets or code
sets that consist of a single code of length 1.