On 28/02/2013 4:31 AM, Ioi Lam wrote:
While we are cleaning this up, would it make sense to replace the
maddening \ and $ with something like this?
sun/awt/motif/X11GB2312${DOLLAR}Decoder.class
Perhaps. Though there would not be anything to tell people that this
variable exists and that they should use it.
I suspect that what we see now is that a variable that is subject to
secondary expansion requires the \$$$$, while a variable not subject to
secondary expansion requires \$$. Though I'm unsure how anything
recently changed would have affected this aspect.
David
- Ioi
On 02/26/2013 06:47 PM, David Holmes wrote:
So ... nashorn pushed this change:
http://hg.openjdk.java.net/jdk8/tl/diff/5b0b6ef58dbf/common/makefiles/MakeBase.gmk
which modified the way $ gets handled in a file name, and in doing so
completely broke many of the pre-existing escape sequences used for
dealing with nested classes (those with $ in their name). I'm not 100%
convinced that this actually relates to secondary expansion but ...
Alan found and fixed both:
8008977: profiles build broken by Nashorn build changes
and
8009029: SunEC provider classes ending up in rt.jar after Nashorn
build changes
However we have now ended up in a situation where nobody knows how to
write the name of a nested class, without knowing how that name will
actually get processed. If you look in CreateJars.gmk you will find:
RT_JAR_EXCLUDES += \
...
sun/awt/motif/X11GB2312.class \
sun/awt/motif/X11GB2312\$$Decoder.class \
sun/awt/motif/X11GB2312\$$Encoder.class \
sun/awt/motif/X11GBK.class \
sun/awt/motif/X11GBK\$$Encoder.class \
sun/awt/motif/X11KSC5601.class \
sun/awt/motif/X11KSC5601\$$Decoder.class \
sun/awt/motif/X11KSC5601\$$Encoder.class \
(which were modified to fix 8009029) and later:
ifneq ($(OPENJDK_TARGET_OS), windows)
CHARSETS_EXTRA_FILES:=sun/awt/motif/X11GBK.class \
sun/awt/motif/X11GB2312\$$$$Decoder.class \
sun/awt/motif/X11GB2312.class \
sun/awt/motif/X11KSC5601\$$$$Decoder.class \
sun/awt/motif/X11KSC5601\$$$$Encoder.class \
sun/awt/motif/X11GB2312\$$$$Encoder.class \
sun/awt/motif/X11GBK\$$$$Encoder.class \
sun/awt/motif/X11KSC5601.class
endif
So the same file names are listed once with \$$ and once with \$$$$,
and they both have to be that way to work!
This is untenable. There should only be one way to write the name of a
nested class file inside the makefile.
FYI in Profiles.gmk when expanding foo/*.class I already had to do a
similar substitution as is now in ListPathsSafely:
# Function to expand foo/*.class into the set of classes
# NOTE: Classfiles with $ in their name are problematic as that is the
# meta-character for both make and the shell! Hence the \$$$$
substitution.
# But note that if you echo these values they will NOT display as
expected.
class_list = $(patsubst $(JDK_OUTPUTDIR)/classes/%,%,\
$(foreach i,$(1), $(subst $$,\$$$$, $(wildcard
$(JDK_OUTPUTDIR)/classes/$i))))
So I'd like to understand why the nashorn change was made so that we
can determine how to get back to only having one way to specify file
names containing $
David
-----
