Web lists-archives.com

Re: Strange errors running gcc tests on Cygwin




On 03/07/2017 06:36 PM, David Billinghurst wrote:
On 8/03/2017 10:25, Daniel Santos wrote:

My concern is with the dynamic portion of this behavior -- what is affected by environment variables.

Many years ago I ran a nightly build/test of gcc under cygwin and reported the results to gcc-testresults. There may be is discussion on the gcc mailing lists from c2000-2005. If you search "site:gcc.gnu.org David Billinghurst cygwin" you ??might?? find something relevant.

From memory, I got it all working by

 * building gcc and friends
 * using find to locate all the .exe and .dll files in the build tree
 * worked out by trial and error which files were needed at run time by
   the test suite.
 * setting PATH when running the testsuite so that the directories
   containing (new) required .exe and .dll were in front of any system
   directories
 * making sure that PATH wasn't reset by the testsuite
 * looking at places where LD_LIBRARY_PATH was set/modified by the
   testsuite and checking if cygwin needed PATH to match
 * (submitting patched to fix gcc testsuite under cygwin)

Once that was done it all "just worked" until it broke again. Good luck.


Thank you very much for this. This is the path that I was kind-of setting off on, I just wanted to try one more time to run the tests as-is, this time with only one make job and see if I could get the mass of failures to match so that I could say that my patches cause "no additional errors" (ignoring the fact that there's 16k total failures), but I can't even get the breakages to match up. (This is another topic, when I run tests in parallel I eventually end up with some "broken pipe" errors and that make job hangs).

I found some of your patches, pity it got re-broken. I have a bug open for this here: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79867. In the end, I think that this should be fixed by adding some new "black box" interfaces to DejaGnu that manage the executable and library search paths. Then gcc's testsuite should deprecate ANY direct access to any of the *PATH environment variables in favor of this new interface in DejaGnu. As it is, the gcc code already changes both LD_LIBRARY_PATH and SHLIB_PATH to support HP-UX (not sure if any other systems use the latter), so it's already gotten a little hairy.

In order to facilitate this cleanly, I think that libgcc needs to be moved out of /usr/bin and into /usr/lib/gcc/<triplet>/<version>/ and then have that added to the PATH and LD_LIBRARY_PATH somewhere (autoexec.bat? registry? I have no idea). Having an /sbin/ldconfig would be the most ideal mechanism of managing this. (Anybody want a little project? :) The test harness regularly toggles in between the host and target compiler.

We need a clean and reproducible set of steps for running tests on Cygwin. Somewhere, these tests should be run regularly, maybe on a server/compiler farm somewhere under a VM, so that breakages can be addressed as soon as they appear rather than when somebody wants to touch the ms_abi code and has to test on Cygwin -- how I ended up here. :)

Thanks again for the help!
Daniel

--
Problem reports:       http://cygwin.com/problems.html
FAQ:                   http://cygwin.com/faq/
Documentation:         http://cygwin.com/docs.html
Unsubscribe info:      http://cygwin.com/ml/#unsubscribe-simple