Consider four executables, a.out and b.out a hello world example, and c.out and d.out copied from the dwz executable itself:
...
$ gcc hello.c -g ; cp a.out b.out
$ cp dwz c.out ; cp c.out d.out
...

This gives us two executables with 130 DIEs, and two executables with 5356 DIEs:
...
$ readelf -w a.out | grep '(DW_TAG' | wc -l
130
$ readelf -w b.out | grep '(DW_TAG' | wc -l
130
$ readelf -w c.out | grep '(DW_TAG' | wc -l
5356
$ readelf -w d.out | grep '(DW_TAG' | wc -l
5356
...

Now consider a gdb script that traces the dwz invocations:
...
$ cat gdb.script
b dwz
commands
continue
end
run
...

We run in multifile mode, with a low-mem limit of 1000 dies, and trace into LOG:
...
$ gdb \
    -batch \
    -x gdb.script \
    --args dwz -m3 -l1000 a.out b.out c.out d.out \
    > LOG 2>&1
...

which we then summarize as follows:
...
$ grep 'dwz (' LOG | awk '{print $5}'
"a.out",
"b.out",
"c.out",
"c.out",
"d.out",
"d.out",
"a.out",
"b.out",
"c.out",
"d.out",
...

The first 6 invocations are according to plan:
- a.out and b.out have fewer DIEs than 1000, and are processed.
- c.out and d.out have more DIEs than 1000, and are each processed twice:
  - once in regular mode (where dwz stops at a 1000 processed DIEs and returns
    2), and 
  - once in low-mem mode.

However, then all 4 files are once more processed in fi_multifile mode, while the intention is that c.out and d.out (being bigger than the low-mem limit) are not processed anymore.