The existing code in stackdepth_walk gives conservative, but not very tight, bounds for stack depth.  Some of this is simply bugs: for example,
  import a,b,c,d,e,f,...,z
will leak a stack slot for each import, and
  a[:],b[:],c[:],...,z[:]
leaks a stack slot for each SLICE+0 opcode.

Some of it is because stackdepth_walk assumes that the stack effect of jumps is the same at the target.  In particular, a sequence of for-loops will leak 2 stack slots for each for-loop.

The attached patch fixes both of these problems.  It corrects the stack effect for several opcodes that were simply incorrect (perhaps from an older version?).  It also changes stackdepth_walk to treat FOR_ITER, SETUP_EXCEPT, and SETUP_FINALLY specially when following their targets.

Note: this patch also introduces a dependency from stackdepth_walk to the compiler, because it assumes that END_FINALLY is always preceded by the target of a SETUP_EXCEPT or SETUP_FINALLY.