[LLVMdev] Metadata/Value split has landed

[Tom Stellard]

On Wed, Dec 10, 2014 at 11:21:08AM -0800, Duncan P. N. Exon Smith wrote:
> 
> > On 2014 Dec 10, at 08:40, Tom Stellard <tom at stellard.net> wrote:
> > 
> > On Tue, Dec 09, 2014 at 09:22:16PM -0800, Duncan P. N. Exon Smith wrote:
> >> The `Metadata`/`Value` split (PR21532) landed in r223802 -- at least, the
> >> C++ side of it.  This was a rocky day, but I suppose that's what I get
> >> for failing to stage the change in smaller pieces.
> >> 
> >> As of r223916 (lldb), I'm not aware of any remaining (in-tree) breakage,
> >> so if I've missed some problem in the sea of buildbot errors, please
> >> flag me down.
> >> 
> >> I'll follow up soon with bitcode and assembly changes!
> > 
> > Hi Duncan,
> > 
> > I started getting random assertion failures in some tests yesterday, and I think
> > it may be related to this change.  Here is the stack trace:
> > 
> > #0  0x00007ffff59f4c39 in raise () from /lib64/libc.so.6
> > #1  0x00007ffff59f6348 in abort () from /lib64/libc.so.6
> > #2  0x00007ffff59edb96 in __assert_fail_base () from /lib64/libc.so.6
> > #3  0x00007ffff59edc42 in __assert_fail () from /lib64/libc.so.6
> > #4  0x00007ffff3a30e92 in llvm::LeakDetectorImpl<void>::addGarbage(void const*) [clone .part.19] () from /opt/buildbot/lib/libLLVM-3.6svn.so
> > #5  0x00007ffff3a30fd3 in llvm::LeakDetector::addGarbageObjectImpl(void*) () from /opt/buildbot/lib/libLLVM-3.6svn.so
> > #6  0x00007ffff3a40eed in llvm::MDNode::getTemporary(llvm::LLVMContext&, llvm::ArrayRef<llvm::Metadata*>) () from /opt/buildbot/lib/libLLVM-3.6svn.so
> > #7  0x00007ffff3426b3f in MapValueImpl(llvm::Metadata const*, llvm::ValueMap<llvm::Value const*, llvm::WeakVH, llvm::ValueMapConfig<llvm::Value const*, llvm::sys::SmartMutex<false> > >&, llvm::RemapFlags, llvm::ValueMapTypeRemapper*, llvm::ValueMaterializer*) ()
> >   from /opt/buildbot/lib/libLLVM-3.6svn.so
> > #8  0x00007ffff3426bd6 in MapValueImpl(llvm::Metadata const*, llvm::ValueMap<llvm::Value const*, llvm::WeakVH, llvm::ValueMapConfig<llvm::Value const*, llvm::sys::SmartMutex<false> > >&, llvm::RemapFlags, llvm::ValueMapTypeRemapper*, llvm::ValueMaterializer*) ()
> >   from /opt/buildbot/lib/libLLVM-3.6svn.so
> > #9  0x00007ffff3426bd6 in MapValueImpl(llvm::Metadata const*, llvm::ValueMap<llvm::Value const*, llvm::WeakVH, llvm::ValueMapConfig<llvm::Value const*, llvm::sys::SmartMutex<false> > >&, llvm::RemapFlags, llvm::ValueMapTypeRemapper*, llvm::ValueMaterializer*) ()
> >   from /opt/buildbot/lib/libLLVM-3.6svn.so
> > #10 0x00007ffff3426eed in llvm::MapValue(llvm::Metadata const*, llvm::ValueMap<llvm::Value const*, llvm::WeakVH, llvm::ValueMapConfig<llvm::Value const*, llvm::sys::SmartMutex<false> > >&, llvm::RemapFlags, llvm::ValueMapTypeRemapper*, llvm::ValueMaterializer*) ()
> >   from /opt/buildbot/lib/libLLVM-3.6svn.so
> > #11 0x00007ffff3426f39 in llvm::MapValue(llvm::MDNode const*, llvm::ValueMap<llvm::Value const*, llvm::WeakVH, llvm::ValueMapConfig<llvm::Value const*, llvm::sys::SmartMutex<false> > >&, llvm::RemapFlags, llvm::ValueMapTypeRemapper*, llvm::ValueMaterializer*) ()
> >   from /opt/buildbot/lib/libLLVM-3.6svn.so
> > #12 0x00007ffff3427174 in llvm::RemapInstruction(llvm::Instruction*, llvm::ValueMap<llvm::Value const*, llvm::WeakVH, llvm::ValueMapConfig<llvm::Value const*, llvm::sys::SmartMutex<false> > >&, llvm::RemapFlags, llvm::ValueMapTypeRemapper*, llvm::ValueMaterializer*) ()
> >   from /opt/buildbot/lib/libLLVM-3.6svn.so
> > #13 0x00007ffff3755786 in (anonymous namespace)::ModuleLinker::linkGlobalValueBody(llvm::GlobalValue&) () from /opt/buildbot/lib/libLLVM-3.6svn.so
> > #14 0x00007ffff375767f in llvm::Linker::linkInModule(llvm::Module*) () from /opt/buildbot/lib/libLLVM-3.6svn.so
> > #15 0x00007ffff3758cfb in llvm::Linker::LinkModules(llvm::Module*, llvm::Module*, std::function<void (llvm::DiagnosticInfo const&)>) ()
> >   from /opt/buildbot/lib/libLLVM-3.6svn.so
> > #16 0x00007ffff6c9d8cf in clang::BackendConsumer::HandleTranslationUnit(clang::ASTContext&) () from /opt/buildbot/lib/libOpenCL.so.1
> > #17 0x00007ffff6e61f23 in clang::ParseAST(clang::Sema&, bool, bool) () from /opt/buildbot/lib/libOpenCL.so.1
> > #18 0x00007ffff6c9e6bb in clang::CodeGenAction::ExecuteAction() () from /opt/buildbot/lib/libOpenCL.so.1
> > #19 0x00007ffff6b7ead6 in clang::FrontendAction::Execute() () from /opt/buildbot/lib/libOpenCL.so.1
> > #20 0x00007ffff6b5d179 in clang::CompilerInstance::ExecuteAction(clang::FrontendAction&) () from /opt/buildbot/lib/libOpenCL.so.1
> > #21 0x00007ffff6b1282c in (anonymous namespace)::compile_llvm (llvm_ctx=..., 
> >    source="\n__kernel void test_fn(__local float *sSharedStorage, __global float *srcValues, __global uint *offsets, __global float *destBuffer, uint alignmentOffset )\n{\n    int tid = get_global_id( 0 );\n    sSha"..., headers=..., name="input.cl", triple="r600--", processor="verde", opts="", 
> >    address_spaces=..., optimization_level=@0x7fffffff21cc: 2, r_log=...) at llvm/invocation.cpp:255
> > #22 0x00007ffff6b140c8 in clover::compile_program_llvm (source=..., headers=..., ir=ir at entry=PIPE_SHADER_IR_NATIVE, target=..., opts=..., r_log=...)
> >    at llvm/invocation.cpp:710
> > #23 0x00007ffff6b0a371 in clover::program::build (this=this at entry=0x23a0530, devs=..., opts=opts at entry=0x7ffff793dc0d "", headers=...)
> >    at core/program.cpp:63
> > #24 0x00007ffff6af31c4 in clBuildProgram (d_prog=0x23a0538, num_devs=0, d_devs=0x0, p_opts=<optimized out>, pfn_notify=0x0, user_data=0x0)
> >    at api/program.cpp:182
> > 
> > Does this look related?  If so, let me know what other information you need to
> > try to debug this issue.
> 
> This could be related; I'm not sure.
> 

I'm pretty sure that this commit is the cause of the regression.

r223801 works and r223810 does not, and I don't think any of the other
commits in that range could cause this.

> It looks like a leak detection assertion, and I didn't need to change
> that logic at all.  `ValueMap` calls `MDNode::getTemporary()` and
> `MDNode::deleteTemporary()` in the same ways it used to (and I didn't
> touch the implementation of those).
> 
> Can you reproduce this with `llvm-link`?  If so, that sounds like the
> best place to start.

I can't reproduce this using llvm-link unfortunately.  Any other ideas?

-Tom