From: Eric Botcazou Date: Mon, 19 Jun 2017 06:47:47 +0000 Subject: Re: sparc gcc 7.1 compile issue Message-Id: <1577186.YNi7BlrOG7@polaris> List-Id: References: <20170531193431.GA26922@waldemar-brodkorb.de> In-Reply-To: <20170531193431.GA26922@waldemar-brodkorb.de> MIME-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit To: sparclinux@vger.kernel.org > > All of my testing is being done with gcc-6.3 vanilla and current > > kernels, and in fact I'm doing parallel "make -j128" gcc and glibc > > testsuite runs and not hitting any problems at all. > > According to Eric Botcazou from gcc upstream, we most likely know the > culprit now. Eric noticed that the issue occurred within the cilk-plus > testsuite and by passing --disable-libcilkrts to configure, the testsuite > finished without problems, even with -j32. The Solaris maintainer ported Cilk++ (https://en.wikipedia.org/wiki/Cilk) to the SPARC architecture for GCC 7 and later, which explains why the issue doesn't show up with GCC 6. Cilk++ is a multithreading layer on top of C and C++ and its testsuite does a lot of thread manipulation. Note that, even when the machine doesn't freeze, the Cilk++ testsuite reports several timeouts (there are no such timeouts on Solaris): https://gcc.gnu.org/ml/gcc-testresults/2017-06/msg00462.html This is probably the most promising angle of attack: someone should look into the timeouts by running the Cilk++ testsuite with low/no parallelism and find out where they come from (compiler, Cilk++ runtime or system/kernel). I can do it, but I'm essentially a compiler guy and I don't have much expertise in multithreading at the system/kernel level. -- Eric Botcazou