[llvm-dev] Ubuntu LLVM packages incompatible with clang built projects?

Mon Oct 1 13:29:18 PDT 2018

I can confirm that Alastair's patch works for me. Steps taken:

1. git checkout https://git.llvm.org/git/llvm.git/ @ 65ce2e56889
(release_70)
2. configure with "CC=gcc-8 CXX=g++-8 cmake -GNinja
-DCMAKE_BUILD_TYPE=RelWithDebInfo .. -DLLVM_TOOL_CLANG_BUILD=OFF
-DCMAKE_INSTALL_PREFIX=$HOME/code/llvm-install/7/gcc/RelWithDebInfo"
3. build with "ninja install"
4. build my own project and link against new custom llvm build
5. run offending executable and notice SEGFAULT
6. apply patch (below)
7. step #2, but with install prefix changed to
$HOME/code/llvm-install/7/gcc-patch/RelWithDebInfo
8. step #3 and #4
9. run offending executable and notice successful run

If this patch can't make it in for LLVM 7.0.1, could it make 7.1? If so,
what kind of timeline would you expect?

HTH,
Kern

------------------------------------------

Patch:

diff --git a/include/llvm/ADT/Optional.h b/include/llvm/ADT/Optional.h
index 353e5d0ec9d..150915c5b9e 100644
--- a/include/llvm/ADT/Optional.h
+++ b/include/llvm/ADT/Optional.h
@@ -108,7 +108,6 @@ template <typename T, bool IsPodLike> struct
OptionalStorage {
   }
 };

-#if !defined(__GNUC__) || defined(__clang__) // GCC up to GCC7 miscompiles
this.
 /// Storage for trivially copyable types only.
 template <typename T> struct OptionalStorage<T, true> {
   AlignedCharArrayUnion<T> storage;
@@ -118,14 +117,20 @@ template <typename T> struct OptionalStorage<T, true>
{

   OptionalStorage(const T &y) : hasVal(true) { new (storage.buffer) T(y); }
   OptionalStorage &operator=(const T &y) {
-    *reinterpret_cast<T *>(storage.buffer) = y;
-    hasVal = true;
+   if (hasVal) {
+     // Use std::addressof to silence strict-aliasing warnings from gcc.
+     *reinterpret_cast<T *>(std::addressof(storage.buffer)) = y;
+   } else {
+     // Copy assignment to uninitialized storage is undefined behavior so
copy
+     // construct with placement new when no value is held.
+     new (storage.buffer) T(y);
+     hasVal = true;
+   }
     return *this;
   }

   void reset() { hasVal = false; }
 };
-#endif
 } // namespace optional_detail

 template <typename T> class Optional {


------------------------------------------




On Mon, Oct 1, 2018 at 10:47 AM Tom Stellard <tstellar at redhat.com> wrote:

> On 09/29/2018 01:09 AM, Hans Wennborg via llvm-dev wrote:
> > Trunk still has the different gcc and clang versions.
> >
> > What's worse, the 7.0.0 release has them too :-( I completely missed
> > this and we can't fix it for 7.0.1 since that would also be an ABI
> > break.
> >
> Is this something we could fix by adding a symbol alias to the
> linker script for libLLVM.so?
>
> -Tom
>
> > Alastair, if you noticed this during the release testing, did you
> > surface it anywhere?
> >
> >
> > On Fri, Sep 28, 2018 at 6:07 PM, Reid Kleckner <rnk at google.com> wrote:
> >> Just be aware that those ifdefs were recommitted and reverted several
> times,
> >> so I'm not sure what the state is.
> >>
> >> On Fri, Sep 28, 2018 at 8:48 AM Alastair Murray via llvm-dev
> >> <llvm-dev at lists.llvm.org> wrote:
> >>>
> >>> Hi Kern,
> >>>
> >>> We also had issues with mixing GCC/Clang builds when testing the 7.0
> >>> release branch.
> >>>
> >>> My colleague submitted a patch that fixed the issue for us:
> >>> * https://reviews.llvm.org/D50710
> >>>
> >>> I no longer have a copy of the error we were seeing to see if it was
> the
> >>> same, but the fix is to llvm::OptionalStorage and your error message
> >>> involves llvm::Optional.  The patch removes a GCC specific ifdef, and
> >>> hopefully fixes the issue that made that necessary in the first place,
> but
> >>> we never really knew how the GCC miscompilation was expressed.
> >>>
> >>> It would be interesting to know whether this resolves your issue, but
> >>> regardless we should give the patch a prod.
> >>>
> >>> Regards,
> >>> Alastair Murray.
> >>>
> >>>
> >>> On 27/09/18 08:46, Kern Handa via llvm-dev wrote:
> >>>
> >>> Hi folks,
> >>>
> >>> Not sure if this is the right mailing list target, but I'm trying out
> the
> >>> new LLVM 7.0 packages found at http://apt.llvm.org by porting over an
> >>> existing LLVM 6.0 project of ours to the new version. In doing so, I
> found
> >>> that the executable always segfaulted at the same spot with no
> explanation:
> >>>
> >>> 0x0000000000fefe33 in
> >>>
> llvm::RegisterTargetMachine<llvm::X86TargetMachine>::Allocator(llvm::Target
> >>> const&, llvm::Triple const&, llvm::StringRef, llvm::StringRef,
> >>> llvm::TargetOptions const&, llvm::Optional<llvm::R
> >>> eloc::Model>, llvm::Optional<llvm::CodeModel::Model>,
> >>> llvm::CodeGenOpt::Level, bool) ()
> >>>
> >>> This happens if I build my project and link it against LLVM 7.0 with
> >>> either clang++ 6.1 or clang++ 7.0. This does not happen if I build my
> >>> project with g++ 8.
> >>>
> >>> I have also confirmed that building LLVM 7.0 with clang++ and then
> using
> >>> that private build of the libraries fixes the issue.
> >>>
> >>> To summarize, it seems that there's an ABI incompatibility introduced
> >>> somewhere, which means LLVM 7.0 compiled with gcc can only be used by
> other
> >>> projects also built with gcc. Is this something that's being
> investigated
> >>> and to be resolved as a fix in 7.1? If not, is there any official
> release
> >>> note that's remarking on this incompatibility?
> >>>
> >>> Thanks,
> >>> Kern
> > _______________________________________________
> > LLVM Developers mailing list
> > llvm-dev at lists.llvm.org
> > http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
> >
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20181001/1ad7de65/attachment.html>
