How does MSVC's fetchRegressOpts() work at all?

The MSVC members of the buildfarm seem to not like my recent patch
b14cf229f4bd7238, which basically did this to contrib/hstore_plpython's
Makefile:
SHLIB_LINK += ../hstore/libhstore.a $(wildcard ../../src/pl/plpython/libpython*.a) $(wildcard
../../src/pl/plpython/libplpython*.a)endif
-REGRESS_OPTS = --load-extension=hstore
+REGRESS_OPTS += --load-extension=hstoreifeq ($(python_majorversion),2)REGRESS_OPTS += --load-extension=plpythonu
--load-extension=hstore_plpythonuendif

Now they're failing to load hstore before running the regression test, eg
http://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=bowerbird&dt=2015-05-18%2015%3A37%3A17

Looking at vcregress.pl's sub fetchRegressOpts, it seems that it only
notices "REGRESS_OPTS =" lines not "REGRESS_OPTS +=" lines, so that
isn't surprising --- but how is it that the tests *are* still loading
plpythonu and hstore_plpythonu?  I grant that my Perl is weak, but
I don't see how this code would ever have dealt with either lines
using +=, or multiple assignments to REGRESS_OPTS.  How come it worked
before?
        regards, tom lane