On 2018-03-14 21:44, Erik Joelsson wrote:
On 2018-03-14 13:22, Magnus Ihse Bursie wrote:
On 2018-03-14 15:59, Erik Joelsson wrote:
This change is unfortunately not correct for Linux and Solaris. We
cannot link libfontmanager explicitly against either libawt_xawt or
libawt_headless. See
https://bugs.openjdk.java.net/browse/JDK-8196516 for my suggestion
on a better fix than we currently have. I was hoping for Severin to
check it out and pick it up, but he is away for a bit so that hasn't
happened.
The reason we cannot link explicitly is that we need to decide at
runtime whether to pull in the headful or headless libraries. If one
or the other is already pulled in, and we explicitly load the other,
the runtime linker will lookup the common symbols in one or the
other in an unpredictable way. Some users get the correct behavior,
some get the wrong behavior. We recently had discussions around this
on this list if you want to dive deeper into it.
Also see https://bugs.openjdk.java.net/browse/JDK-8194870 for one of
the consequences of explicit linking here.
I think the mac part is ok though, but Phil has to have a look. For
Linux and Solaris, if you could remove the -lawt_headless and add
"-Xlinker --unresolved-symbols=ignore-all" to LDFLAGS for linux I
think we should be good.
Oh, this seems like a fine mess. :-(
I see we used to link with libawt_headless for solaris, but removed
it in JDK-8194870. Phil remarked in that bug: "We linked against
headless before then but allowed undefined symbols JDK 9 switched
this to libawt_headless when headless apps failed." but I'm not sure
I understand what this really means.
I think he means that we linked with xawt before, but switch to
headless at some point. If you run on a machine without X11 present,
then if you load xawt, you get a runtime linker error. That was the
reason for switching to linking with headless instead.
I agree it seems likely that the macosx changes is correct. I could
probably add -Wl,--unresolved-symbols=ignore-all for gcc on linux,
but that is unlikely to work on solstudio. :-( So I'm afraid we're
still stuck with the need to remove -z defs for some builds. :-(
On solaris we should be able to override -Wl,-z,defs with
-Wl,-z,nodefs later on the link line I think.
Unfortunately, that gives the error "-z defs and -z nodefs are mutually
exclusive", or something like that. :-(
Perhaps that works on Linux as well. I didn't find nodefs on the Linux
man page, but coming to think about it, the "no" prefix should be
universal.
Question to Phil: Would it be possible to rewrite the code to load
and reference the symbols dynamically, using libdl? The problem here
is that the functions are declared extern, instead of loaded into a
function pointer at runtime. That's how you typically use this kind
of runtime-determined dynamic loading.
That sounds like a better long term solution.
The problem here is the "long" in "long term". :-( What I'm trying to
accomplish here is to make sure that all JDK libraries use the
LDFLAGS_JDKLIB for linking (and possibly additional flags), not a
version with some parts subst:ed out. :(
I can't really find a way to accomplish this for solstudio short of
fixing this to use proper dynamic linking.
/Magnus
/Erik
/Magnus
/Erik
On 2018-03-14 04:45, Magnus Ihse Bursie wrote:
This is the final LDFLAGS cleanup, which required some more work to
resolve.
Libfontmanager had previously explicitely disabled -z defs, with
the result that linking did not complain about missing libraries.
To fix this, I had to provide the proper libraries to link with.
For linux and solaris, this was libawt_headless. For macosx, this
was libawt_lwawt, but also three system frameworks.
Note that this patch has a merge conflict with JDK-8199606. The end
result of both patches are shown in the patch (that is, with -lc
removed). I will make sure to resolve the conflicts properly when
committing this patch.
I have run COMPARE_BUILDS, with expected results. That is, no
changes for Windows, and a deps change for macosx, linux and
solaris. I also got a symbol change for linux, since the symbols
from libawt_headless changed from e.g. "AWTCharAdvance" to
"AWTCharAdvance@@SUNWprivate_1.1". And of course, when you do
linking changes, you also end up getting binary/disasm changes.
Bug: https://bugs.openjdk.java.net/browse/JDK-8199608
WebRev:
http://cr.openjdk.java.net/~ihse/JDK-8199608-clean-up-LDFLAGS-for-libfontmanager/webrev.01
