On 1/24/24 12:04 AM, Kewen.Lin wrote: > on 2024/1/24 11:11, Peter Bergner wrote: >> But not with this. The -mdejagnu-cpu=power10 option already enables -mvsx. >> If the user explcitly forces -mno-vsx via RUNTESTFLAGS, then let them. >> The options set in RUNTESTFLAGS come after the options in the dg-options >> line, so even adding -mvsx like the above won't help the test case PASS > > But this is NOT true, at least on one of internal Power10 machine > (ltcden2-lp1). > > With the command below: > > make check-gcc-c RUNTESTFLAGS="--target_board=unix/-mno-vsx > powerpc.exp=pr112886.c" > > this test case fails without the explicit -mvsx in dg-options. > > From the verbose dumping, the compilation command looks like: > > /home/linkw/gcc/build/gcc-test-debug/gcc/xgcc > -B/home/linkw/gcc/build/gcc-test-debug/gcc/ > /home/linkw/gcc/gcc-test/gcc/testsuite/gcc.target/powerpc/pr112886.c > -mno-vsx > -fdiagnostics-plain-output -mdejagnu-cpu=power10 -O2 -ffat-lto-objects > -fno-ident -S > -o pr112886.s > > "-mno-vsx" comes **before** "-mdejagnu-cpu=power10 -O2" rather than **after**. > > I guess it might be due to different behaviors of different versions of > runtest framework?
That is confusing, unless as you say, the behavior changed. The whole reason we added -mdejagnu-cpu= (and the dg-skip usage before that) was due to encountering problems when the test case wanted a specific -mcpu= value and the user overrode it in their RUNTESTFLAGS and that can only happen when its options come last on the command line. Then again, why didn't the powerpc_vsx_ok test not save us here? > So there can be two cases with user explicitly specified -mno-vsx: > > 1) RUNTESTFLAGS comes after dg-options (assuming same order for -mvsx in > powerpc_vsx_ok) > > powerpc_vsx_ok test failed, so UNSUPPORTED > > // with explicit -mvsx does nothing as you said. > > 2) RUNTESTFLAGS comes before dg-options > > powerpc_vsx_ok test succeeds, but FAIL. > > // with suggested -mvsx, make it match the powerpc_vsx_ok checking and the > case not fail. > > As above I think we still need to append the "-mvsx" explicitly. As > tested/verified, it > does help the case not to fail on ltcden2-lp1. I'd like to verify that the behavior did change before we enforce adding that option. The problem is, there are a HUGE number of older test cases that would need updating to "fix" them too. ...and not just adding -mnsx, but -maltivec and basically any -mfoo option where the test case is expecting the feature foo to be used/tested. It would be a huge mess. Peter
