[LLVMdev] Metadata/Value split has landed
Zalman Stern
zalman at google.com
Thu Dec 11 14:41:20 PST 2014
The assertion no longer fires in Halide with top-of-tree llvm. Thank you
for the fix.
-Z-
On Thu, Dec 11, 2014 at 1:56 PM, Duncan P. N. Exon Smith <
dexonsmith at apple.com> wrote:
> I committed:
>
> r224058 = 966942da9e68b59c31ce770e7f94c55a63482c6b
> r224060 = da75f7277e3a129aed8ef8aa4e0d84de40b76fd4
> r224061 = f88e4c8e9171045454b2c8e05054c2af8da3fe4f
>
> Let me know if somehow you're still hitting the problem.
>
> r224061 removes leak detection entirely from `MachineInstr`. There aren't
> any leaks to be had there, since they're allocated in a custom allocator.
> They're just dropped away once `MachineFunction` is deleted.
>
> @Zalman, thanks again for your help digging into this.
>
> > On 2014 Dec 11, at 09:05, Tom Stellard <tom at stellard.net> wrote:
> >
> > On Thu, Dec 11, 2014 at 11:52:34AM -0500, Tom Stellard wrote:
> >> On Wed, Dec 10, 2014 at 05:27:45PM -0800, Duncan P. N. Exon Smith wrote:
> >>> +zalman at google.com
> >>>
> >>
> >> Hi Duncan,
> >>
> >> This patch plus another small change fixes the assertion failure for
> >> me. With the patch alone, the void* overload of addGarbageObject()
> >> was being used by MDNode::getTemporary(), so I had to cast the object as
> >> an MDNode*:
> >>
> >> diff --git a/lib/IR/Metadata.cpp b/lib/IR/Metadata.cpp
> >> index cd5edd2..916d216 100644
> >> --- a/lib/IR/Metadata.cpp
> >> +++ b/lib/IR/Metadata.cpp
> >> @@ -564,7 +564,7 @@ MDNode *MDNode::getMDNode(LLVMContext &Context,
> >> ArrayRef<Metadata *> MDs,
> >> MDNodeFwdDecl *MDNode::getTemporary(LLVMContext &Context,
> >> ArrayRef<Metadata *> MDs) {
> >> MDNodeFwdDecl *N = new (MDs.size()) MDNodeFwdDecl(Context, MDs);
> >> - LeakDetector::addGarbageObject(N);
> >> + LeakDetector::addGarbageObject((MDNode*)N);
> >> return N;
> >> }
> >
> > Sorry, after more extensive testing, this doesn't work. It looks like
> > you need to add const MDNode * overloads to addGarbageObject() adding
> > them for addGarbageObjectImpl() doesn't seem to work:
> >
> > diff --git a/include/llvm/IR/LeakDetector.h
> > b/include/llvm/IR/LeakDetector.h
> > index e0b131e..b272eaf 100644
> > --- a/include/llvm/IR/LeakDetector.h
> > +++ b/include/llvm/IR/LeakDetector.h
> > @@ -79,6 +79,17 @@ struct LeakDetector {
> > #endif
> > }
> >
> > + static void addGarbageObject(const MDNode *Object) {
> > +#ifndef NDEBUG
> > + addGarbageObjectImpl(Object);
> > +#endif
> > + }
> > + static void removeGarbageObject(const MDNode *Object) {
> > +#ifndef NDEBUG
> > + removeGarbageObjectImpl(Object);
> > +#endif
> > + }
> > +
> > private:
> > // If we are debugging, the actual implementations will be called...
> > static void addGarbageObjectImpl(const Value *Object);
> >>
> >>
> >> I'm in favor of committing this.
> >>
> >> -Tom
> >>
> >>
> >>>> On 2014 Dec 10, at 15:57, Duncan P. N. Exon Smith <
> dexonsmith at apple.com> wrote:
> >>>>
> >>>>>
> >>>>> On 2014 Dec 10, at 14:08, Tom Stellard <tom at stellard.net> wrote:
> >>>>>
> >>>>> 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?
> >>>>
> >>>> (Continuing via email, since Tom stepped away IRC.)
> >>>>
> >>>> Tom, from the trace [1], it the problematic pointer (0x27e4c80) only
> shows
> >>>> up once.
> >>>>
> >>>> [1]: http://people.freedesktop.org/~tstellar/md-crash.out
> >>>>
> >>>> That means that something *else* -- other than
> `MDNode::getTemporary()`
> >>>> -- must be adding that address to garbage and failing to remove it.
> >>>>
> >>>> I just dug into `LeakDetector::addGarbageObject()` and it stores *all*
> >>>> calls to `addGarbage()` in the same place. There are a fair number of
> >>>> these in the IR:
> >>>>
> >>>> $ git grep -e addGarbageObject -w -- lib/IR/
> >>>> lib/IR/BasicBlock.cpp: LeakDetector::addGarbageObject(this);
> >>>> lib/IR/BasicBlock.cpp: LeakDetector::addGarbageObject(this);
> >>>> lib/IR/Function.cpp: LeakDetector::addGarbageObject(this);
> >>>> lib/IR/Function.cpp: LeakDetector::addGarbageObject(this);
> >>>> lib/IR/Function.cpp: LeakDetector::addGarbageObject(this);
> >>>> lib/IR/Function.cpp: LeakDetector::addGarbageObject(this);
> >>>> lib/IR/Globals.cpp: LeakDetector::addGarbageObject(this);
> >>>> lib/IR/Globals.cpp: LeakDetector::addGarbageObject(this);
> >>>> lib/IR/Globals.cpp: LeakDetector::addGarbageObject(this);
> >>>> lib/IR/Globals.cpp: LeakDetector::addGarbageObject(this);
> >>>> lib/IR/Globals.cpp: LeakDetector::addGarbageObject(this);
> >>>> lib/IR/Instruction.cpp: LeakDetector::addGarbageObject(this);
> >>>> lib/IR/Instruction.cpp: LeakDetector::addGarbageObject(this);
> >>>> lib/IR/Instruction.cpp: if (!P)
> LeakDetector::addGarbageObject(this);
> >>>> lib/IR/Metadata.cpp: LeakDetector::addGarbageObject(N);
> >>>>
> >>>> I think the next step is to identify who called `addGarbageObject()`
> with
> >>>> the problematic address, and what the stack trace was.
> >>>>
> >>>> The weird thing is, I also noticed a semantic change I made here
> >>>> accidentally. `addGarbageObject()` has two overloads: `void*` and
> >>>> `const Value*`. `MDNode::getTemporary()` used to match the latter,
> but
> >>>> now it matches the former.
> >>>>
> >>>> The weird part: all the other calls to `addGarbageObject()` look like
> they
> >>>> send in a `Value *`.
> >>>>
> >>>> Do you have any other calls to `addGarbageObject()`? Do they match
> >>>> `void *`?
> >>>>
> >>>> Also, what happens with the attached patch? (If this fixes your
> problem,
> >>>> I think it's just papering over something...)
> >>>>
> >>>> <0001-IR-Detect-Metadata-leaks-separately-from-generic-obj.patch>
> >>>
> >>> Zalman also had a reproduction, and he's been able to track it down to
> >>> an `addGarbageObject()` call from `MachineBasicBlock`. It looks like
> >>> the MBB gets deallocated but `removeGarbageObject()` isn't called yet.
> >>>
> >>> CC'ing him here so he can join the thread once he's gotten a little
> >>> further.
> >>>
> >>> BTW, if this is blocking anyone, I can commit the patch I attached to
> >>> my previous email (or you can apply it locally). I think it's probably
> >>> the right thing eventually -- it improves the output when there *is* an
> >>> issue -- but I haven't committed it yet since it'll cover up the
> >>> problem.
> >> _______________________________________________
> >> LLVM Developers mailing list
> >> LLVMdev at cs.uiuc.edu http://llvm.cs.uiuc.edu
> >> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20141211/0b45819b/attachment.html>
More information about the llvm-dev
mailing list