RFA: 8031668: TOOLCHAIN_FIND_COMPILER unexpectedly resolves symbolic links AND 8148351: Only display resolved symlink for compiler, do not change path

Tue Jun 19 14:17:08 UTC 2018

Hi Erik,

On Tue, 2018-06-19 at 07:06 -0700, Erik Joelsson wrote:
> Hello,
> 
> On 2018-06-19 06:30, Severin Gehwolf wrote:
> > Hi Kevin,
> > 
> > Thanks for your help! Adding in build-dev for additional input.
> > 
> > On Tue, 2018-06-19 at 13:51 +0100, Kevin Walls wrote:
> > > Hi Severin -
> > > 
> > > Ah, actually there's a conflict with the way we build, where we use
> > > ccache in the path, and it no longer likes it with this change:
> > > 
> > > Our build system jprt fails, with the configure output:
> > > ...
> > > checking flags for boot jdk java command for small workloads...
> > > -XX:+UseSerialGC -Xms32M -Xmx512M
> > > configure: Using default toolchain gcc (GNU Compiler Collection)
> > > checking for gcc... /opt/jprt/products/P1/ccache2.4/bin/gcc
> > > checking resolved symbolic links for CC...
> > > /opt/jprt/products/P1/ccache2.4/bin/ccache
> > > configure: error: /opt/jprt/products/P1/ccache2.4/bin/gcc is a symbolic
> > > link to ccache. This is not supported.
> > > configure: Please use --enable-ccache instead of providing a wrapped
> > > compiler.
> > > configure exiting with result code 1
> > 
> > Erik, Magnus, how do you do this for your JDK 9+ builds? Is there a way
> > to use "--enable-ccache" for JDK 8 builds?
> 
> We don't. The problem here is that the build machines we have for 8 have 
> ccache installed as a gcc wrapper. This was done back in 7 with the 
> intention of transparently improving build performance (but in reality 
> it didn't). Because of this situation, we implemented a workaround for 8 
> where we explicitly don't accept ccache as gcc wrapper and look for gcc 
> elsewhere instead. Our build machines for 9 do not have cache installed 
> at all (it's still in the devkit, but not as a wrapper, and since we saw 
> no benefit for our use case we don't use ccache anyway).
> 
> The change being backported here removes the workaround that allows the 
> build to function with the ccache wrapper on the system. I don't think 
> that's a good idea for 8. We need to retain that workaround so the patch 
> needs to be modified before going into 8u.

Sigh. Back to the drawing board ;-)

> Reproducing this issue locally should be pretty simple. Create a dummy 
> script named ccache somewhere. Put a symlink named gcc, pointing to that 
> ccache script first in your path. Then run configure.

Thanks, yes. I've got that reproduced.

Cheers,
Severin

> /Erik
> > > It's the second change, 8148351, which triggers this failure, as it
> > > removes lines at 549 onwards which previously tried to handle the ccache
> > > in the path case.
> > 
> > Right. It's 8148351 which we'd be interested in.
> > 
> > > Yes, in 9 onward we abort the build if ccache is found at the end of a
> > > link, we haven't done that in jdk 8 before... Would retaining the ccache
> > > handling code in toolchain.m4 still give you the behaviour you want from
> > > doing the backport?
> > 
> > Thanks,
> > Severin
> > 
> > > Thanks
> > > Kevin
> > > 
> > > 
> > > 
> > > 
> > > 
> > > On 19/06/2018 07:23, Severin Gehwolf wrote:
> > > > Hi Kevin,
> > > > 
> > > > On Mon, 2018-06-18 at 21:11 +0100, Kevin Walls wrote:
> > > > > Hi -- I've been doing various 8u build changes recently -- would you
> > > > > like me to push this, with the regenerated autogen-generated file?
> > > > 
> > > > It would be very much appreciated!
> > > > 
> > > > Thanks,
> > > > Severin
> > > > 
> > > > > On 18/06/2018 17:45, Rob McKenna wrote:
> > > > > > Approved. Please work with the appropriate team to find a sponsor.
> > > > > > 
> > > > > >        -Rob
> > > > > > 
> > > > > > On 18/06/18 16:36, Severin Gehwolf wrote:
> > > > > > > Hi,
> > > > > > > 
> > > > > > > Please approve these two backports to JDK 8u-dev which are already in
> > > > > > > JDK 9 and better.
> > > > > > > 
> > > > > > > 8031668: TOOLCHAIN_FIND_COMPILER unexpectedly resolves symbolic links
> > > > > > > webrev: http://cr.openjdk.java.net/~sgehwolf/webrevs/JDK-8031668/webrev.01/
> > > > > > > hg export: http://cr.openjdk.java.net/~sgehwolf/webrevs/JDK-8031668/JDK-8031668.jdk8.export.patch
> > > > > > > 
> > > > > > > 8148351: Only display resolved symlink for compiler, do not change path
> > > > > > > webrev: http://cr.openjdk.java.net/~sgehwolf/webrevs/JDK-8148351/webrev.01/
> > > > > > > hg export: http://cr.openjdk.java.net/~sgehwolf/webrevs/JDK-8148351/JDK-8148351.jdk8.export.patch
> > > > > > > 
> > > > > > > Bug 8031668 is a dependency for 8148351 which actually fixes the
> > > > > > > wrapped compiler issue on the JDK 8 tree. The JDK 8 patch of 8031668 is
> > > > > > > the same as for JDK 9, modulo some context changes. After the JDK 8
> > > > > > > patch of 8031668, the JDK 9 patch of 8148351 imports as is.
> > > > > > > 
> > > > > > > The issue at hand is that one cannot build the JDK 8 tree with certain
> > > > > > > compiler wrappers such as cscppc. It currently fails with:
> > > > > > > 
> > > > > > > configure: Using default toolchain gcc (GNU Compiler Collection)
> > > > > > > checking for gcc... /usr/lib64/cscppc/gcc
> > > > > > > checking resolved symbolic links for CC... /usr/bin/cscppc
> > > > > > > checking if CC is disguised ccache... no, keeping CC
> > > > > > > configure: The C compiler (located as /usr/bin/cscppc) does not seem to be the required gcc compiler.
> > > > > > > configure: The result from running with --version was: ""
> > > > > > > configure: error: A gcc compiler is required. Try setting --with-tools-dir.
> > > > > > > configure exiting with result code 1
> > > > > > > 
> > > > > > > After both backports configure with a wrapped GCC succeeds.
> > > > > > > 
> > > > > > > Please note that I'd need an Oracle sponsor for getting this pushed to
> > > > > > > jdk8u-dev since both changes require re-generation of generated-
> > > > > > > configure.sh
> > > > > > > 
> > > > > > > Thanks,
> > > > > > > Severin
> 
> 