[PATCH] D43762: [IPSCCP] Use constant range information for comparisons of parameters.

[Eric Christopher via llvm-commits]

On Fri, Jul 20, 2018 at 6:57 AM Florian Hahn <florian.hahn at arm.com> wrote:

>
>
> On 12/07/2018 03:12, Eric Christopher wrote:
> > I've gone ahead and reverted this for now in r336877. Happy to help if
> > we need anything more here.
> >
> > -eric
> >
> > On Wed, Jul 11, 2018 at 3:39 PM Jordan Rupprecht via llvm-commits
> > <llvm-commits at lists.llvm.org <mailto:llvm-commits at lists.llvm.org>>
> wrote:
> >
> >     It looks like this commit caused a miscompile regression (with -O1
> >     or greater), reduced to this example:
> >
> >     namespace {
> >     template <typename T1, typename T2>
> >     bool Same(const T1& v1, const T2& v2) {
> >        return v1 == v2;
> >     }
> >
> >     class Thing {
> >       public:
> >        enum Color { kPurple, kYellow };
> >        enum Numbers { kZero };
> >
> >        static bool CheckFoo(int x, Color color) {
> >          if (!Same(kZero, x)) __builtin_abort();
> >          if (color == kPurple) __builtin_printf("color is purple\n");
> >          return true;
> >        }
> >     };
> >
> >     class SomeThing : public Thing {
> >       public:
> >        static void DoBar() {
> >          if (!CheckFoo(0, kPurple)) {
> >            __builtin_printf("This should never happen\n");
> >          }
> >        }
> >     };
> >
> >     class OtherThing : public Thing {
> >       public:
> >        static void DoBar() {
> >          if (!CheckFoo(0, kYellow)) {
> >            __builtin_printf("This should never happen\n");
> >          }
> >        }
> >     };
> >     }  // namespace
> >
> >     int main(int argc, char** argv) {
> >        SomeThing::DoBar();
> >        OtherThing::DoBar();
> >        return 0;
> >     }
> >
> >     Output:
> >     $ <clang built @r336098> -O1 regression.cc && ./a.out
> >     color is purple
> >     This should never happen
> >     This should never happen
> >     $ <clang built @r336097> -O1 regression.cc && ./a.out
> >     color is purple
> >     It looks like this might just be uncovering a bug in an earlier
> >     revision (r335588), but we're not sure yet. Mind if we revert this
> >     change until we figure out what's going on?
> >
> >     (I also left this comment on https://reviews.llvm.org/rL336098,
> >     reposting here for visibility)
>
> I've recommitted the patch again as r337548. The underlying issue was
> just exposed by the commit and was fixed in rL337507.
>
> Thanks again for the excellent reproducer and please let me know if
> there are any more problems with the non-reduced code.
>
>
Thanks a ton. We'll give it a shot soon.

-eric
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20180720/24436d8c/attachment.html>