RFR (S): 7178667: ALT_EXPORT_PATH does not export server jvm on macosx

[Staffan Larsen]

On 21 jun 2012, at 19:12, Scott Kovatch wrote:

>  54 # Package built libraries in a universal binary
>  55 $(UNIVERSAL_LIPO_LIST):
>  56         BUILT_LIPO_FILES="`find $(EXPORT_JRE_LIB_DIR)/{i386,amd64}/$(subst $(EXPORT_JRE_LIB_DIR)/,,$@) 2>/dev/null`"; \
>  57         if [ -n "$${BUILT_LIPO_FILES}" ]; then \
>  58           $(MKDIR) -p $(shell dirname $@); \
>  59           lipo -create -output $@ $${BUILT_LIPO_FILES}; \
>  60         fi      
>  61 
>  62 
>  63 # Copy built non-universal binaries in place
>  64 $(UNIVERSAL_COPY_LIST):
>  65         BUILT_COPY_FILES="`find $(EXPORT_JRE_LIB_DIR)/{i386,amd64}/$(subst $(EXPORT_JRE_LIB_DIR)/,,$@) 2>/dev/null`"; \
>  66         if [ -n "$${BUILT_COPY_FILES}" ]; then \
>  67           for i in $${BUILT_COPY_FILES}; do \
>  68             if [ -f $${i} ]; then \
>  69               $(MKDIR) -p $(shell dirname $@); \
>  70               $(CP) $${i} $@; \
>  71             fi; \
>  72           done; \
>  73         fi
>  74 
> 
> This first item will find all object files that were built separately for i386 and amd64 architectures, and then use 'lipo -create -output …' to create a single universal binary.  
> 
> The next phase copies those combined, universal binaries into EXPORT_JRE_LIB_DIR, since the Mac has never had/needed to break out libraries by architecture.
> 
> So, it sounds like when you rebuilt, everything was built into jre/lib/i386 and jre/lib/amd64, but never combined (or, in this case, just copied) into jre/lib, and therefore not found. 

Yes. Or rather, only the client jvm was combined, but the client jvm isn't copied into the j2sdk-image on mac, so nothing was copied.

/Staffan

> 
> Since we only build x86_64 the lipo command is effectively a cp.
> 
> -- Scott
> 
> 
> On Jun 21, 2012, at 9:54 AM, Staffan Larsen <staffan.larsen at oracle.com> wrote:
> 
>> At least for me, MACOSX_UNIVERSAL ends up being set to true in hotspot/make/bsd/makefiles/defs.make. 
>> 
>> Line 186 and forward:
>> 
>> # Universal build settings
>> ifeq ($(OS_VENDOR), Darwin)
>> # Build universal binaries by default on Mac OS X
>> MACOSX_UNIVERSAL = true
>> 
>> If this isn't intentional, then the fix for my problem is something else.
>> 
>> /Staffan
>> 
>> On 21 jun 2012, at 17:53, Henri Gomez wrote:
>> 
>>> universal build, on OSX ? Happy to see that some works/fixes around it :)
>>> 
>>> BTW, how did you get in trouble since universal build is disabled for
>>> now unless some code is added to :
>>> 
>>> There was an old thread on jdk7u-dev list and a proposed patch for
>>> review (http://openjdk-osx-build.googlecode.com/svn/trunk/patches-jdk7u-osx/universal-build.patch)
>>> 
>>> 
>>> 
>>> 
>>> 
>>> 2012/6/21 Staffan Larsen <staffan.larsen at oracle.com>:
>>>> [adding build-dev and macosx-port-dev]
>>>> 
>>>> On 21 jun 2012, at 14:43, David Holmes wrote:
>>>> 
>>>>> On 21/06/2012 10:30 PM, Staffan Larsen wrote:
>>>>>> Do you mean:
>>>>>> 
>>>>>> .PHONY: $(UNIVERSAL_LIPO_LIST) $(UNIVERSAL_COPY_LIST)
>>>>> 
>>>>> Yes. Now they will always be rebuilt.
>>>>> 
>>>>>> Yes, that seems to have the same effect. Probably a better solution.
>>>>> 
>>>>> I think both of these simply mask the real problem. I still don't understand how only some of the list items get "rebuilt". The CR says
>>>>> 
>>>>> "These targets will only be run for the last item in the xxx_LIST variables (which happens to be the client jvm)"
>>>>> 
>>>>> but I don't understand why that is?
>>>> 
>>>> Neither do I. Makefiles is black magic to me. I only discovered that building the complete JDK from the top-level directory did not update the hotspot bits in the j2sdk-image and this was the ultimate cause.
>>>> 
>>>> Here is an updated webrev: http://cr.openjdk.java.net/~sla/7178667/webrev.02/
>>>> 
>>>> Thanks,
>>>> /Staffan
>>>> 
>>>> 
>>>>> But I also don't understand this universalization process.
>>>>> 
>>>>> BTW you might want to run this past the bsd-port folks (don't recall the exact alias) and/or build-dev. I seem to recall that last time we changed something to do with universal builds it actually broke something.
>>>>> 
>>>>> David
>>>>> 
>>>>>> Thanks,
>>>>>> /Staffan
>>>>>> 
>>>>>> On 21 jun 2012, at 14:12, David Holmes wrote:
>>>>>> 
>>>>>>> Hi Staffan,
>>>>>>> 
>>>>>>> On 21/06/2012 6:33 PM, Staffan Larsen wrote:
>>>>>>>> Please review the following fix to makefiles for universal binaries on
>>>>>>>> max os x. The idea is to force the target to be executed for all items
>>>>>>>> in the list.
>>>>>>>> 
>>>>>>>> Fix contributed by Rickard Bäckman (rbackman).
>>>>>>>> 
>>>>>>>> webrev: http://cr.openjdk.java.net/~sla/7178667/webrev.01/
>>>>>>> 
>>>>>>> I don't understand the problem that this addresses but wouldn't you get the same affect by declaring those targets as PHONY ?
>>>>>>> 
>>>>>>> David
>>>>>>> 
>>>>>>> PS. Unrelated but I was astounded to see that bsd/Makefile and linux/Makefile both have a chunk of code conditional on "ifeq ($(OSNAME),solaris)" Huh!
>>>>>> 
>>>> 
>> 
>