Hi Jakub, On Thu, Jan 26, 2017 at 02:29:51PM +0100, Jakub Jelinek wrote: > Hi! > > I've noticed that the file &1 in libgomp/testsuite/ directory after make > check and left over hsa_offloading_src*.exe files. > After looking more carefully, I found various issues. > 1) the way check_compile is invoked leaves result with values on which > lindex doesn't do what is expected to do > 2) "2>&1" doesn't do redirection of stderr to stdout for remote_load; but > apparently both stdout and stderr make up into [lindex $result 1], so it > isn't really needed > 3) if $status isn't pass, we wouldn't remote_file build delete the > executable > 4) otherwise neither, because while output variable contained the executable > filename initially, it was later set to the stdout+stderr of the proglet > > I don't have HSA offloading configured (no HSAIL offloading capable HW on > this WS, and on another WS where I have that I don't have the needed kernel > driver), so couldn't test with proper HSA offloading, but tested without it > and additionally tested with following hack that simulated in the proglet > what the plugin would do (in that case make check started compiling/running > libgomp.hsa.c/ tests where some of them obviously failed miserably with > actually NVPTX offloading, but that is to be expected when I cheated). > > Martin, can you please test this yourself, so that I can check it in?
I have just tested the patch and the resulting libgomp.sum is identical to one without it (modulo the line with time). Thanks for fixing the issues, Martin > > The attached patch is the hack I used, which obviously shouldn't be applied. > > 2017-01-26 Jakub Jelinek <ja...@redhat.com> > > * testsuite/lib/libgomp.exp > (check_effective_target_hsa_offloading_selected_nocache): Fix up > check_compile invocation. Fix up removal of executable. Drop > bogus "2>&1" argument. >
