The clang-i686-freebsd-selfhost-rel buildbot is failing APIntTest.i33_Count on the first run, but the stage 2 test is passing. It looks like it could be a miscompilation by the system gcc. We should upgrade the system gcc on that machine, or stop testing the stage 1 llvm. The tester is not useful as is. It is failing constantly. Who owns this machine? Can the system be upgraded? /jakob
On 2010-09-09 23:44, Jakob Stoklund Olesen wrote:> The clang-i686-freebsd-selfhost-rel buildbot is failing APIntTest.i33_Count on the first run, but the stage 2 test is passing. > > It looks like it could be a miscompilation by the system gcc.Yes, unfortunately on FreeBSD the first stage should be compiled using "-O2 -fno-strict-aliasing" instead of the default optimization flags, or this particular test will always fail. Stage 2 (which is compiled by stage 1 clang) has no problems with it, and does not require -fno-strict-aliasing. This is some obscure gcc 4.2.1 problem, but it is extremely hard to find out whether it really is an aliasing issue, or plainly a bug. It is most likely the latter, since compiling with e.g. gcc 4.5 and -O3 did not give such failures, the last time I tried it (about two months ago, approximately).> We should upgrade the system gcc on that machine, or stop testing the stage 1 llvm. > > The tester is not useful as is. It is failing constantly. > > Who owns this machine? Can the system be upgraded?FreeBSD's gcc will probably never be upgraded past 4.2.1, because that was the last GPLv2 version. This is one of the main reasons for getting clang in the tree... ;)
On Sep 9, 2010, at 3:26 PM, Dimitry Andric wrote:> On 2010-09-09 23:44, Jakob Stoklund Olesen wrote: >> The clang-i686-freebsd-selfhost-rel buildbot is failing APIntTest.i33_Count on the first run, but the stage 2 test is passing. >> >> It looks like it could be a miscompilation by the system gcc. > > Yes, unfortunately on FreeBSD the first stage should be compiled using > "-O2 -fno-strict-aliasing" instead of the default optimization flags, or > this particular test will always fail. Stage 2 (which is compiled by > stage 1 clang) has no problems with it, and does not require > -fno-strict-aliasing. > > This is some obscure gcc 4.2.1 problem, but it is extremely hard to find > out whether it really is an aliasing issue, or plainly a bug. It is > most likely the latter, since compiling with e.g. gcc 4.5 and -O3 did > not give such failures, the last time I tried it (about two months ago, > approximately).Thanks, Daniel, could this bot configure stage 1 with -fno-strict-aliasing? Alternatively, if anybody knows a source workaround for this bug and/or aliasing issue...> We should upgrade the system gcc on that machine, or stop testing the stage 1 llvm. >> >> The tester is not useful as is. It is failing constantly. >> >> Who owns this machine? Can the system be upgraded? > > FreeBSD's gcc will probably never be upgraded past 4.2.1, because that > was the last GPLv2 version. This is one of the main reasons for getting > clang in the tree... ;)Makes sense. /jakob