On Linux/x86-64, I got FAIL: gfortran.dg/ldist-1.f90 -O scan-tree-dump-times ldist "distributed: split to 4 loops" 1 FAIL: gcc.dg/tree-ssa/ldist-4.c scan-tree-dump-times ldist "distributed: split to 2 loops" 0
Confirming on i686-apple-darwin9 with m64.
Both are testsuite bugs, apparently. ldist-1.f90 scans for: ! { dg-final { scan-tree-dump-times "distributed: split to 4 loops" 1 "ldist" } } and the dump output is: "Loop 1 distributed: split to 5 loops." And ldist-4.c is even xfailed in a weird way: /* We used to distribute also innermost loops, but these could produce too much code in the outer loop, degrading performance of scalar code. So this test is XFAILed because the cost model of the stand alone distribution pass has evolved. */ /* { dg-final { scan-tree-dump-times "distributed: split to 2 loops" 0 "ldist" } } */ and it works now: FIXME: Loop 1 not distributed: it has more than two basic blocks. Loop 2 distributed: split to 2 loops.
Also fails on hppa2.0w-hp-hpux11.11.
If it is just the test case I will attempt to fix it. If anyone else is already doing so, please let me know so I don't waste my time. :)
Subject: Bug 35406 Author: jvdelisle Date: Fri Mar 14 03:48:23 2008 New Revision: 133206 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=133206 Log: 2008-03-13 Jerry DeLisle <jvdelisle@gcc.gnu.org> PR fortran/35406 * gfortran.dg/ldist-1.f90: Update test to check for 5 loops distributed instead of 4. Modified: trunk/gcc/testsuite/ChangeLog trunk/gcc/testsuite/gfortran.dg/ldist-1.f90
Closing. Please reopen if I broke anything.
I forgot the C case with ldist-4.c, :)
Subject: Bug 35406 Author: jvdelisle Date: Fri Mar 14 04:59:39 2008 New Revision: 133207 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=133207 Log: 2008-03-13 Jerry DeLisle <jvdelisle@gcc.gnu.org> PR testsuite/35406 * gcc.dg/tree-ssa/ldist-4.c: Update expected scan match count from 0 to 1. Modified: trunk/gcc/testsuite/ChangeLog trunk/gcc/testsuite/gcc.dg/tree-ssa/ldist-4.c
Closing again.
The old gcc.dg/tree-ssa/ldist-4.c passed on Linux/ia32 and failed on Linux/Intel64. Revision 133206 changed it to fail on Linux/ia32 and pass on Linux/Intel64. I think you want "4|5" instead of 4 or 5. The same goes for gcc.dg/tree-ssa/ldist-4.c.
(In reply to comment #10) > The old gcc.dg/tree-ssa/ldist-4.c passed on Linux/ia32 and failed > on Linux/Intel64. Revision 133206 changed it to fail on Linux/ia32 > and pass on Linux/Intel64. I think you want "4|5" instead of 4 or > 5. The same goes for gcc.dg/tree-ssa/ldist-4.c. Actually, the test produces different results for 32bit vs 64bit targets. You can see that in addition to x86_64, it also fails for ppc64, I guess that the same way [1]. I think we need: /* { dg-final { scan-tree-dump-times "distributed: split to 2 loops" 0 "ldist" { target ilp32 } } } */ /* { dg-final { scan-tree-dump-times "distributed: split to 2 loops" 1 "ldist" { target lp64 } } } */ for ldist-4.c and something similar for fortran test. [1] http://gcc.gnu.org/ml/gcc-testresults/2008-03/msg01040.html
(In reply to comment #11) > Actually, the test produces different results for 32bit vs 64bit targets. Are you sure you don't mean 32 vs 64-bit *hosts*? My cris-elf (32-bit target) autotester runs on a x86_64 F8, and it tells me that between 133196:133207 these regressions were introduced, coincident with jvdelisle's changes and that there were no non-testsuite changes in this range: gcc.sum gcc.dg/tree-ssa/ldist-4.c gfortran.sum gfortran.dg/ldist-1.f90 and gcc.log and fortran.log says: FAIL: gcc.dg/tree-ssa/ldist-4.c scan-tree-dump-times ldist "distributed: split to 2 loops" 1 FAIL: gfortran.dg/ldist-1.f90 -O scan-tree-dump-times ldist "distributed: split to 5 loops" 1
(In reply to comment #12) > > Actually, the test produces different results for 32bit vs 64bit targets. > > Are you sure you don't mean 32 vs 64-bit *hosts*? Hm, I have to look into my {,-m32} testresults on x86_64 host. No, looking at [1] I don't think so. [1] http://gcc.gnu.org/ml/gcc-testresults/2008-03/msg01037.html
Created attachment 15315 [details] testsuite patch relative to r133207 Proposed patch, relative to r133207. HP, does this patch solve your problems?
(In reply to comment #14) > HP, does this patch solve your problems? Yes it does, thank you.
Subject: Bug 35406 Author: uros Date: Fri Mar 14 14:17:41 2008 New Revision: 133216 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=133216 Log: PR testsuite/35406 * gcc.dg/tree-ssa/ldist-4.c: Update expected scan match count from 0 to 1 for lp64 targets. * gfortran.dg/ldist-1.f90: Update test to check for 5 loops distributed instead of 4 for lp64 targets. Modified: trunk/gcc/testsuite/ChangeLog trunk/gcc/testsuite/gcc.dg/tree-ssa/ldist-4.c trunk/gcc/testsuite/gfortran.dg/ldist-1.f90
Fixed again.
The revised testcase still fails on hppa*-*-* (32-bit). The dump contains "Loop 1 distributed: split to 5 loops."
Removing milestone; no indication here that this is a regression.
Closing. Stale