[aarch64-port-dev ] aarch64: jdk9: 8078245: fails to build from source

[Edward Nevill]

Hi,

You are right, there is a better place for this,

in make/lib/Awt2dLibraries.gmk b/make/lib/Awt2dLibraries.gmk

is the line

  LIBSPLASHSCREEN_CFLAGS += -DSPLASHSCREEN -DPNG_NO_MMX_CODE \

so, someone has already done this here for x86, so I just need to add the equivalent for aarch64

  LIBSPLASHSCREEN_CFLAGS += -DSPLASHSCREEN -DPNG_NO_MMX_CODE -DPNG_ARM_NEON_OPT=0 \

Note: I have not conditionalised this on OPENJDK_TARGET_CPU_ARCH as the corresponding x86 code is not conditionalised and I really don't think the symbol PNG_ARM_NEON_OPT will have any effect on any other architectures.

Erik, Magnus & David, I have listed you as reviewers in the following webrev. If you are happy with this please let me know and I will push.

http://cr.openjdk.java.net/~enevill/8078245/webrev.02/

All the best,
Ed.

On Mon, 2015-05-04 at 17:03 +0200, Erik Joelsson wrote:
> Right, that does make sense. This isn't a toolchain global flag like 
> some others we've had to deal with lately, but rather seems pretty 
> specific to the png code, or am I missing something? If that's the case, 
> then it's better added for the specific library.
> 
> /Erik
> 
> On 2015-05-04 16:43, Magnus Ihse Bursie wrote:
> > Why not just for the library that needs it?
> >
> > /Magnus
> >
> >> 4 maj 2015 kl. 11:05 skrev Erik Joelsson <erik.joelsson at oracle.com>:
> >>
> >> Hello,
> >>
> >> I think this looks good enough. The flags handling is still a big mess so even if there were a better place for this, I couldn't say where.
> >>
> >> /Erik
> >>
> >>> On 2015-04-30 07:26, David Holmes wrote:
> >>> Hi Nevill,
> >>>
> >>> Just realized this was sent to hotspot-dev (attempting bcc) but is not a hotspot issue. With your new approach this is a build issue so cc'ing build-dev.
> >>>
> >>> The new approach seems better to me but build folk need to confirm the placement.
> >>>
> >>> Thanks,
> >>> David
> >>>
> >>>> On 29/04/2015 8:52 PM, Edward Nevill wrote:
> >>>>> On Fri, 2015-04-24 at 17:11 +1000, David Holmes wrote:
> >>>>> Hi Ed,
> >>>>>
> >>>>>> On 21/04/2015 7:16 PM, Edward Nevill wrote:
> >>>>>> Hi,
> >>>>>>
> >>>>>> The current jdk9 tip fails to build from source on aarch64 with the following error message
> >>>>>>
> >>>>>> /home/ed/build/1504/dev/build/linux-aarch64-normal-server-release/support/native/java.desktop/libsplashscreen/pngrutil.o: In function `png_init_filter_functions':
> >>>>>> /home/ed/build/1504/dev/jdk/src/java.desktop/share/native/libsplashscreen/libpng/pngrutil.c:3947: undefined reference to `png_init_filter_functions_neon'
> >>>>>> collect2: error: ld returned 1 exit status
> >>>>>>
> >>>>>> The following webrev gets it building again.
> >>>>>>
> >>>>>> http://cr.openjdk.java.net/~enevill/8078245/webrev.00/
> >>>>> Shouldn't the guard be Aarch64 specific rather than just __arm__ ? I'm
> >>>>> also wondering how we would get __ARM_NEON defined but not __arm__?
> >>>> On arm 32 bit gcc defines the symbol __ARM_NEON if the flag -mfpu=neon is specified. Since this is not specified as part of the OpenJDK build the symbol is not defined and the build succeeds.
> >>>>
> >>>> On aarch64 the symbol __ARM_NEON is always defined (the theory being that aarch64 always supports Neon). Personally I think this is borken as it causes builds to fail (and not just OpenJDK, several other projects have had the same build failure - try googling the above error message). But we are stuck with gcc as it is.
> >>>>
> >>>> However, I don't like the above fix because it not only modifies the jdk, it modifies an external component which is pulled into jdk, which means every time a new revision of linpng is pulled in, the patch will have to be applied again.
> >>>>
> >>>> A better approach I think is to define the symbol PNG_ARM_NEON_OPT=0 in the build (only if aarch64).
> >>>>
> >>>> The failing code in pngpriv.h reads
> >>>>
> >>>> #ifndef PNG_ARM_NEON_OPT
> >>>>
> >>>> ...
> >>>>
> >>>> #  if (defined(__ARM_NEON__) || defined(__ARM_NEON)) && \
> >>>>     defined(PNG_ALIGNED_MEMORY_SUPPORTED)
> >>>> #     define PNG_ARM_NEON_OPT 2
> >>>> #  else
> >>>> #     define PNG_ARM_NEON_OPT 0
> >>>> #  endif
> >>>> #endif
> >>>>
> >>>> So, if we just predefine PNG_ARM_NEON=0 in the build this will have the same effect as adding defined(__arm__) or !defined(__aarch64__) above.
> >>>>
> >>>> The following patch webrev does this:-
> >>>>
> >>>> http://cr.openjdk.java.net/~enevill/8078245/webrev.01/
> >>>>
> >>>> If you are happy with this and if I could have another reviewer I will push this.
> >>>>
> >>>> All the best,
> >>>> Ed.
>