RFR: JDK-8235903: GCC default -fno-common exposes "multiple definition" link errors
Patrick Zhang OS
patrick at os.amperecomputing.com
Tue Dec 17 03:00:30 UTC 2019
Hi net-dev, and/or security-dev Reviewers,
Please help review and sponsor this patch if acceptable.
It does not tend to bring any functionality changes, instead to propose an early fix, for the build (linking) errors with further toolchain (GCC 10).
From: Martin Buchholz <martinrb at google.com>
Sent: Monday, December 16, 2019 10:44 AM
To: Patrick Zhang OS <patrick at os.amperecomputing.com>; net-dev <net-dev at openjdk.java.net>; OpenJDK <security-dev at openjdk.java.net>
Cc: core-libs-dev <core-libs-dev at openjdk.java.net>
Subject: Re: RFR: JDK-8235903: GCC default -fno-common exposes "multiple definition" link errors
forwarded to other teams for review.
On Fri, Dec 13, 2019 at 3:14 AM Patrick Zhang OS <patrick at os.amperecomputing.com<mailto:patrick at os.amperecomputing.com>> wrote:
Please review this patch, if it should be reviewed by any group other than core-libs, please help forward it. Thanks.
A recent GCC patch (supposed to be in GCC 10) exposes a couple of "multiple definition" link errors when building the jdk tip.
[PATCH] PR85678: Change default to -fno-common
For example, the error message looks like:
* For target support_native_java.base_libjava_BUILD_LIBJAVA_link:
build/support/native/java.base/libjava/childproc.o:(.bss+0x0): multiple definition of `parentPathv'
build/support/native/java.base/libjava/ProcessImpl_md.o:(.bss+0x0): first defined here
collect2: error: ld returned 1 exit status
This was not an issue because the original default -fcommon allowed "global variables defined without an initializer" be handled as COMMON symbols, so it would not warn the problem like "same variable is accidentally defined in more than one compilation unit".
About -fcommon vs -fno-cmmon:
Moving forward, building jdk with latest versions of GCC will trigger this error. Specifying "--with-extra-cflags='-fcommon'" can make it work, but it just got things hidden again.
In addition, -fcommon's behavior "is inconsistent with C++, and on many targets implies a speed and code size penalty on global variable references. It is mainly useful to enable legacy code to link without errors."
Last, in case that other jdk developers would revisit this problem once again, I suggest fixing the error explicitly instead of using "-fcommon"
More information about the core-libs-dev