On Tue, Aug 27, 2019 at 2:27 PM Nick Desaulniers <ndesaulniers@xxxxxxxxxx> wrote: > > On Tue, Aug 27, 2019 at 12:23 PM Josh Poimboeuf <jpoimboe@xxxxxxxxxx> wrote: > > > > On Tue, Aug 27, 2019 at 08:00:18PM +0300, Leon Romanovsky wrote: > > > On Tue, Aug 27, 2019 at 10:08:30AM -0500, Josh Poimboeuf wrote: > > > > On Mon, Aug 26, 2019 at 04:38:29PM -0700, Nathan Chancellor wrote: > > > > > Looks like that comes from tune_qsfp, which gets inlined into > > > > > tune_serdes but I am far from an objtool expert so I am not > > > > > really sure what kind of issues I am looking for. Adding Josh > > > > > and Peter for a little more visibility. > > > > > > > > > > Here is the original .o file as well: > > > > > > > > > > https://github.com/nathanchance/creduce-files/raw/4e252c0ca19742c90be1445e6c722a43ae561144/rdma-objtool/platform.o.orig > > > > > > > > 574: 0f 87 00 0c 00 00 ja 117a <tune_serdes+0xdfa> > > > > > > > > It's jumping to la-la-land past the end of the function. > > > > > > How is it possible? > > > > Looks like a compiler bug. > > Nathan, > Thanks for the reduced test case. I modified it slightly: > https://godbolt.org/z/15xejg > > You can see that the label LBB0_5 seemingly points off into space. > Let me play with this one more a bit, then I will file a bug and > report back. Something funny going on in one of the earliest optimizations. Seems related to an analysis that's deducing that the case statement is exhaustive (so the GNU C case range is unrelated), but it's keeping the default case and its comparison around. The analysis is correct; the value should never be > 0xF so there shouldn't be any runtime bugs, but this would avoid an unnecessary comparison for exhaustive switch statements and save a few bytes of object code in such cases. Filed: https://bugs.llvm.org/show_bug.cgi?id=43129 -- Thanks, ~Nick Desaulniers
