RFR: JDK-8200178 Remove mapfiles for JDK native libraries

Magnus Ihse Bursie magnus.ihse.bursie at oracle.com
Fri Mar 23 13:56:34 UTC 2018

With modern compilers, we can use compiler directives (such as 
_attribute__((visibility("default"))), or __declspec(dllexport)) to 
control symbol visibility, directly in the source code. This has 
historically not been present on all compilers, so we had to resort to 
using mapfiles (also known as linker scripts).

This is no longer the case. Now all compilers we use support symbol 
visibility directives, in one form or another. We should start using 
this. Since this has been the only way to control symbol visibility on 
Windows, for most of the shared code, we already have proper JNIEXPORT 
decorations in place.

If we fix the remaining platform-specific files to have proper JNIEXPORT 
tagging, then we can finally get rid of mapfiles.

This fix removed mapfiles for all JDK libraries. It does not touch 
hotspot libraries nor JDK executables; they will have to wait for a 
future fix -- this was complex enough. This change will not have any 
impact on macosx, since we do not use mapfiles there, but instead export 
all symbols. (This is not a good idea, but I'll address that 
separately.) This change will also have a minimal impact on Windows. The 
only reason Windows is impacted at all, is that some changes needed by 
Solaris and Linux were simpler to fix for all platforms.

I have strived for this change to have no impact on the actual generated 
code. Unfortunately, this was not possible to fully achieve. I do not 
believe that these changes will have any actual impact on the product, 
though. I will present the differences more in detail further down. 
Those who are not interested can probably skip that.

The patch has passed tier1 testing and is currently running tier2 and 
tier3. Since the running code is more or less (see caveat below) 
unmodified, I don't expect any testing issues.

Bug: https://bugs.openjdk.java.net/browse/JDK-8200178

Details on changes:
Most of the source code changes are (unsurprisingly) in java.base and 
java.desktop. Remaining changes are in jdk.crypto.ucrypto, 
jdk.hotspot.agent, jdk.jdi and jdk.jdwp.agent.

Source code changes does almost to 100% consists in decorating an 
exported function with JNIEXPORT. I have also followed the long-standing 
convention of adding JNICALL. This is a no-op on non-Windows platforms, 
so for most of the changes this is purely cosmetic (and possibly adding 
in robustness, should the function ever be used on Windows in the 
future). I have also followed the stylistic convention of putting 
"JNIEXPORT <return type> JNICALL" on a separate line. For some 
functions, however, this might cause a change in calling convention on 
Windows. Since this can not apply to exported functions on Windows 
(otherwise they would already have had JNIEXPORT), I do not think this 
matters anything.

A few libraries did not have a mapfile, on Linux and/or Solaris. This 
actually meant that all symbols were exported. It is highly unclear if 
this was known and intended by the original make rule writer. I have 
emulated this by adding the flag $(EXPORT_ALL_SYMBOLS) to these 
libraries. Hopefully, we can remove this flag and fix proper exported 
symbols in the future.

I have run the complete build using COMPARE_BUILD, and made a thourough 
analysis of the differences for Linux and Solaris. All native libraries 
have symbol differences, but most of them are trivial and/or harmless. 
As a result, most libraries have disasm differences as well, but these 
too seem trivial and harmless. The differences in symbols that are 
common to all libraries include:
  * Internal symbols such as __bss_start, _edata, _end and _fini are now 
global. (They are imported as such from the compiler libraries/archives, 
and we have no linker script to override this behavior).
  * The versioning tag SUNWprivate_1.1 is not included, and thus neither 
the .gnu.version_d symbol.
  * There are a few differences in the symbol and/or mangling of some 
local functions. I'm not sure what's causing this,
but it's unlikely to have any effect on the product.

Another common source for change in symbols is due to previous platform 
differences. For instance, if we had "JNIEXPORT int JNICALL do_foo() { 
... }", but do_foo was not in the mapfile, the symbol was exported on 
Windows but not on Linux and Solaris. (Presumable since it was not 
needed there, even though it was compiled for those platforms as well.) 
Now, with the mapfiles gone, do_foo() will be exported on all platforms. 
And contrary, functions that are compiled on all platforms, and were 
exported in mapfiles, but now have gotten an JNIEXPORT decoration, will 
now be visible even on Windows. (This accounts for half of the noticed 
symbol differences on Windows.) I could have made the JNIEXPORT 
conditional on OS, but I didn't think the mess in source code were worth 
the keeping of binary confidence with the old build.

A third common source for change in symbols is due to exported functions 
"leaking" across library borders. For instance, some functions in 
java.desktop is compiled in both libawt_xawt and libawt_headless, but 
they were previously only included in the mapfile for one of these 
libraries. Now, since the visibility is determined by the source code 
itself, it gets exported in both libraries. A variant of this is when a 
library depends on another JDK library, and includes the header file 
from that other library, which in turn declares a function as JNIEXPORT. 
This will cause the including library to also export the function. This 
accounts for the other half of the changes on Windows. A typical example 
of this is that multiple libraries now re-export hotspot symbols from 
libjvm.so, like jio_fprintf. (I have not listed the libjvm re-exports 

Note that  Java_java_io_FileOutputStream_close0 in 
java.base/unix/native/libjava/FileOutputStream_md.c is no longer exported,
and can probably be removed.

Here is a detailed table showing and accounting for all the remaining 
differences found on Linux and Solaris:
java.base/unix/native/libjava: Java_java_io_FileOutputStream_close0 is 
now also exported on unix platforms due to JNIEXPORT.

java.base/jspawnlauncher: On solaris, we also include 
libjava/childproc.o, which
now exports less functions than it used to (it used to export all 
functions, now it is compiled with visibility=hidden).

java.base/java(w).exe: Is now also exporting the following symbols due 
to added JNIEXPORT in libjli on Windows:
(Yes, executables can export symbols on Windows. Confusing, I know.)

java.desktop:/libawt_xawt: The following symbols are now also exported 
on linux and solaris due to JNIEXPORT:

The following symbols are now also exported on linux and solaris due to 
JNIEXPORT (they were previously
  exported only in libawt):

java.desktop:/libawt_headless: The following symbols are now also 
exported due to JNIEXPORT (they were previously
  exported only in libawt_xawt and/or libawt):

java.desktop/libawt: The following symbols are now also exported on 
Windows, due to added

java.desktop/libawt on solaris:
A number of "#pragma weak" directives was previously overridden by the 
Now these directives are respected, so these symbols are now weak 
instead of local:

java.desktop/libawt on solaris: These are now also exported due to 
JNIEXPORT in libmlib_image.

java.desktop/libawt on solaris: This is now also exported due to JNIEXPORT:

java.desktop/libsplashscreen: JNI_OnLoad is now exported on linux and 
solaris due to JNIEXPORT.
libspashscreen also had JNIEXPORT (actually a pure _declspec(dllexport)) 
but no JNICALL, which I added as
a part of converting to JNIEXPORT. The same goes for libmlib_image .

jdk.sctp/libsctp: handleSocketError is now exported on linux and solaris 
due to JNIEXPORT in libnio.

java.instrument:/libinstrument: Agent_OnUnload is now also exported on 
linux and solaris platforms due to JNIEXPORT.
JLI_ManifestIterate is now also exported on Windows, due to added 
JNIEXPORT in libjli.

Java_com_sun_management_internal_Flag_setDoubleValue is now also 
exported on linux and solaris platforms due to JNIEXPORT.


More information about the build-dev mailing list