Test 183.equake from SPEC CPU2000 gets wrong results on powerpc64-linux (-m32 or -m64) when compiled with either "-O2 -fgraphite-identity" or "-O2 -floop-parallelize-all". I'll attach an executable testcase that demonstrates the failure. Test 178.galgel fails with the same options but I haven't investigated it. The failure starts with the merge from the Graphite branch for r150301. With options "-O2 -fgraphite-identity" the minimized testcase and the two CPU2000 tests pass with r150248, the last revision before that merge that builds on powerpc64-linux. They also pass with GCC 4.4.2 with "-O2 -fgraphite-identity".
Created attachment 19495 [details] minimized executable testcase
This might be a duplicate of PR42130.
Created attachment 19666 [details] Reduced test case for this PR We (Tobias, Li and Ramakrishna) have reduced the test case further. This file shows the problem.
Fixed by http://gcc.gnu.org/viewcvs?view=revision&revision=157286