RFR(S): 8141290: AIX: fix build after '8140661: Rename LDFLAGS_SUFFIX to LIBS'

[Magnus Ihse Bursie]

On 2015-11-03 17:20, Volker Simonis wrote:
> Hi,
>
> unfortunately, change '8140661: Rename LDFLAGS_SUFFIX to LIBS' broke
> the AIX build.

I'm sorry for that. :-(
>
> Can somebody please review this small change which fixes the build?
>
> http://cr.openjdk.java.net/~simonis/webrevs/2015/8141290/
> https://bugs.openjdk.java.net/browse/JDK-8141290

The fix looks good to me. However, just to be sure, I've put it to a 
spin on JPRT, so it does not break any other platforms. I'll let you 
know the results when I have them.

>
>
> Besides this immediate build fix I have some general questions/remarks
> regarding change '8140661: Rename LDFLAGS_SUFFIX to LIBS':
>
> First of all I think it is always good to further cleanup and simplify
> the build and the idea to separate linker options from library
> dependencies is yet another good step into this direction.
Thanks. :)
>
> But I have some problems understanding the semantics of the new
> JDKLIB_LIBS variable. As far as I understand it, JDKLIB_LIBS holds the
> native library dependencies for the native libraries build within the
> jdk build (is this true?). As such I'd expect that every call to
> SetupNativeCompilation in jdk/make/lib/ should contain a line like
> "LIBS := $(JDKLIB_LIBS)". If that would have been true, we wouldn't
> have seen any problems on AIX.
>
> Before change 8140661, LDFLAGS_JDKLIB contained the library dependency
> "-liconv" and because every call to SetupNativeCompilation in
> jdk/make/lib/ contained a line like "LDFLAGS := $(LDFLAGS_JDKLIB)" the
> dependency to libiconv.so was always satisfied. Now I admit, that
> having "-liconv" in the LDFLAGS is probably not very elegant and it is
> probably also not strictly needed, because potentially not every
> native jdk library really depends on libiconv.so. So in general the
> cleanup done by 8140661 is good (if it is done right :)
>
> But until now, not every call to SetupNativeCompilation contains the
> line "LIBS := $(JDKLIB_LIBS)". Instead, a lot of the calls still
> contain lines like "LIBS_unix := -ljava" although JDKLIB_LIBS already
> contains "-ljava -ljvm" by default on all Unix platforms. Also, there
> are a lot of lines like "LIBS_solaris := -lc" although JDKLIB_LIBS
> already contains "-lc" by default on Solaris.
>
> So my question: are there plans to further improve the build system by
> adding "LIBS := $(JDKLIB_LIBS)" to all the calls to
> SetupNativeCompilation in jdk/make/lib/ and remove the then redundant
> platform specific settings?

Yes, the plan is to continue cleaning up how this is handled. Let me 
start with some background here:

The new build-infra based hotspot build system has a more or less 
working prototype. The next step is to integrate this properly in the 
build-infra family. For this to work, we need to re-think some of the 
assumptions in build-infra, mostly relating to how we handle flags. 
Right now, we have not really been clear on what flags are needed where, 
and why. For instance, -shared is needed to compile shared libraries on 
gcc, and should really belong in SetupNativeCompilation, but is now set 
as a part of the JDK flags.

So the whole area of flags needs an overhaul. My general idea is to 
break the flags into separate units, somewhat like FREETYPE_LIBS and 
FREETYPE_CFLAGS, so you only need to add the flags you actually need for 
every library. (Maybe increase the abstraction so you just need to 
declare a dependency on freetype to get both FREETYPE_LIBS and 
FREETYPE_CFLAGS automatically?) But the whole project is of the kind 
where you have to start clean up the mess and then see where it lands.

As for JDKLIB_LIBS, I'm not really sure of the semantics myself. :-) As 
you say, it is included in most libs, but not all, but the contents are 
still duplicated in some places. I think it should just go away. Or 
maybe be split up in a "i need libjava/libjvm" (-ljava -ljvm) and 
"really, all libraries need this" (-lc on solaris). And make sure that 
all libraries really get the "really, all libraries need this" part.

> I think that would be yet another nice
> cleanup and in that case I'd also prefer to remove '-liconv' from
> JDKLIB_LIBS on AIX and only add it to the required libraries, in the
> same way as this is done for MacOS X.

That's something I think you can do already. :-) If it's just a few 
libraries that need -liconv, I'd suggest using LIBS_aix for those. If 
it's like almost every library, you should probably wait a while and see 
where the library handling ends up.

/Magnus