[LLVMdev] Issues with clang-llvm debug info validity

[Adrian Prantl]

> On Jun 24, 2014, at 2:47 PM, David Blaikie <dblaikie at gmail.com> wrote:
> 
> On Tue, Jun 24, 2014 at 2:38 PM, Adrian Prantl <aprantl at apple.com> wrote:
>> I will take this.
>> 
>> !5 = metadata !{metadata !"./test.h", metadata !"/Volumes/Data/radar/17052973"}
>> !6 = metadata !{i32 786489, metadata !5, null, metadata !"test", i32 1} ; [ DW_TAG_namespace ] [test] [line 1]
>> !4 = metadata !{i32 786434, metadata !5, metadata !6, metadata !"Test<int>", i32 3, i64 8, i64 8, i32 0, i32 0, null, metadata !2, i32 0, null, metadata !7, metadata !"_ZTSN4test4TestIiEE"} ; [ DW_TAG_class_type ] [Test<int>] [line 3, size 8, align 8, offset 0] [def] [from ]
>> 
>> 
>> !15 = metadata !{metadata !"test2.cpp", metadata !"/Volumes/Data/radar/17052973”}
>> !18 = metadata !{i32 786489, metadata !15, null, metadata !"test", i32 1} ; [ DW_TAG_namespace ] [test] [line 1]
>> !17 = metadata !{i32 786434, metadata !5, metadata !18, metadata !"Test<int>", i32 3, i64 8, i64 8, i32 0, i32 0, null, metadata !2, i32 0, null, metadata !7, metadata !”_ZTSN4test4TestIiEE”}
>> 
>> I think we will need to have some kind of first-come-first-serve policy when it comes to the decl_file/decl_line of namespaces. The way to implement this would be to use DIRefs for namespaces.
> 
> I'm not sure what you're getting at/why this would be necessary. If
> you break the reference to the type itself, it won't matter that
> sometimes the type is defined in a different place, will it?
> 
> For example, the following patch (actually only the first part is
> needed for this specific test case) causes the assertion not to fire:
> 
> diff --git lib/CodeGen/CGDebugInfo.cpp lib/CodeGen/CGDebugInfo.cpp
> index 048c8f8..2bded23 100644
> --- lib/CodeGen/CGDebugInfo.cpp
> +++ lib/CodeGen/CGDebugInfo.cpp
> @@ -771,7 +771,7 @@ llvm::DIType CGDebugInfo::CreateType(const FunctionType *Ty,
>   SmallVector<llvm::Value *, 16> EltTys;
> 
>   // Add the result type at least.
> -  EltTys.push_back(getOrCreateType(Ty->getReturnType(), Unit));
> +  EltTys.push_back(getOrCreateType(Ty->getReturnType(), Unit).getRef());
> 
>   // Set up remainder of arguments if there is a prototype.
>   // otherwise emit it as a variadic function.
> @@ -779,7 +779,7 @@ llvm::DIType CGDebugInfo::CreateType(const FunctionType *Ty,
>     EltTys.push_back(DBuilder.createUnspecifiedParameter());
>   else if (const FunctionProtoType *FPT = dyn_cast<FunctionProtoType>(Ty)) {
>     for (unsigned i = 0, e = FPT->getNumParams(); i != e; ++i)
> -      EltTys.push_back(getOrCreateType(FPT->getParamType(i), Unit));
> +      EltTys.push_back(getOrCreateType(FPT->getParamType(i), Unit).getRef());
>     if (FPT->isVariadic())
>       EltTys.push_back(DBuilder.createUnspecifiedParameter());
>   }
> 
> Without needing to change how namespaces are handled.
> 
> Yes, using DIRefs for namespaces would cause the actual type nodes to
> be deduplicated when they aren't even with the patch above, and that
> would help reduce debug info metadata further if that's
> useful/necessary.

Exactly. While the above patch will cover up the problem, we will still have many duplicate type entries during LTO that will take up memory, even though only the last one to be entered into DIRefMap will actually be used for emitting DWARF. Since DIRefs are also more expensive in terms of lookup speed, I think I’d prefer to only use them only where necessary (=as far down the MDNode DAG as possible).

> 
>> 
>> Question to the C++ experts: Is the name (“test” in this case) a sufficient appropriate unique ID for a C++ namespace?
> 
> Potentially - assuming you don't do this for anything in an anonymous
> namespace. I'd probably see about using the standard mangling
Good point! We should avoid unifying the anonymous namespace :-)
> machinery as we do for other things, though, if possible. Might keep
> it more consistent.

I will do some experiments with nested namespaces, but it seems as if name mangling does not buy us much: this will merely turn into N4test.

-- adrian

> 
>> 
>> -- adrian
>> 
>> 
>>> On Jun 24, 2014, at 1:49 PM, David Blaikie <dblaikie at gmail.com> wrote:
>>> 
>>> +Adrian & Manman,
>>> 
>>> Looks like this is a case of non-DIRef references ending up in the IR,
>>> and thus the references not being deduplicated. This could lead to
>>> some of the IR bloat that you guys implemented the DIRef stuff to
>>> reduce/avoid.
>>> 
>>> The specific issue is that the type is slightly different in the two
>>> TUs (since the namespace scope it's contained within is different in
>>> the two TUs - note the namespace preceeding the header include in
>>> test2.cpp) and the actual DIRef-usage failure is when building
>>> subroutine types. The subroutine types are made with direct references
>>> to types, not with DIRef references to types. This means they won't be
>>> deduplicated/resolved to a single type during linking correctly.
>>> 
>>> It also leads to an assertion as per the original email. Figured you
>>> guys might want to look into/fix this (I assume the right fix is to
>>> use DIRef more pervasively in the frontend - I'm doing a little
>>> experimenting with clang now to see how that looks).
>>> 
>>> - David
>>> 
>>> On Fri, Jun 13, 2014 at 11:08 PM, David Blaikie <dblaikie at gmail.com> wrote:
>>>> Yep, looks like a legit bug to me. I'll try to reproduce it and reduce
>>>> a test case, etc.
>>>> 
>>>> (the ODR doesn't require that the two copies of the Test class
>>>> template appear on the same line in the same file in two distinct
>>>> translation units, it only requires that the two copies have the same
>>>> sequence of tokens - which they do in your example)
>>>> 
>>>> On Fri, Jun 13, 2014 at 5:33 PM, Kevin Modzelewski <kmod at dropbox.com> wrote:
>>>>> Hi all, sorry to post to both lists, but I'm running into an issue where
>>>>> clang-generated debug info is deemed to be invalid by LLVM tools (throws an
>>>>> assertion error in both llc and mcjit), and I'm not sure what the proper
>>>>> resolution is.
>>>>> 
>>>>> Here's a test case; I last tested it on revision r210953:
>>>>> 
>>>>> $ cat test1.cpp
>>>>> #include "test.h"
>>>>> test::Test<int> foo1() {
>>>>>   return test::Test<int>();
>>>>> }
>>>>> 
>>>>> $ cat test2.cpp
>>>>> namespace test {
>>>>> }
>>>>> #include "test.h"
>>>>> test::Test<int> foo2() {
>>>>>   return test::Test<int>();
>>>>> }
>>>>> 
>>>>> $ cat test.h
>>>>> namespace test {
>>>>> template <class T>
>>>>> class Test {
>>>>> };
>>>>> }
>>>>> 
>>>>> $ clang++ -g -emit-llvm test1.cpp -S -o test1.ll
>>>>> $ clang++ -g -emit-llvm test2.cpp -S -o test2.ll
>>>>> $ llvm-link test1.ll test2.ll -S -o linked.ll
>>>>> $ llc linked.ll
>>>>> 
>>>>> The last step raises an assertion error, assuming you have assertions
>>>>> enabled:
>>>>> llc: lib/CodeGen/AsmPrinter/DwarfUnit.cpp:953: llvm::DIE*
>>>>> llvm::DwarfUnit::getOrCreateTypeDIE(const llvm::MDNode*): Assertion `Ty ==
>>>>> resolve(Ty.getRef()) && "type was not uniqued, possible ODR violation."'
>>>>> failed
>>>>> 
>>>>> The initial introduction of the "test" namespace in test2.cpp seems to
>>>>> confuse clang -- test1.cpp and test2.cpp have non-compatible debug
>>>>> information for the test::Test type.  test1.cpp says that test::Test's
>>>>> namespace is defined in test.h, and test2.cpp says that test::Test's
>>>>> namespace is defined in test2.cpp (they both say that the class itself is
>>>>> defined in test.h).  Regardless of whether LLVM tools should allow this,
>>>>> this seems like the wrong output for these files?
>>>>> 
>>>>> Another related example, is if we simply defined the test::Test template in
>>>>> both files (ie, manually executed the #include "test.h") and got rid of the
>>>>> test2.cpp "namespace test {}" definition, we'd run into the same assertion,
>>>>> without hitting the weird clang behavior with the empty namespace
>>>>> pre-declaration.
>>>>> 
>>>>> It's probably also worth noting that compiling directly with clang, ie
>>>>> adding a "int main() {}" function and then doing
>>>>> $ clang++ -g test1.cpp test2.cpp -o test
>>>>> works fine.
>>>>> 
>>>>> I'm not a C++ expert and I don't know if these template issues really are
>>>>> violations of the ODR, but it feels like this should be allowed, and
>>>>> somewhere in the llvm toolchain (llvm-link? DwarfUnit.cpp?) this should be
>>>>> handled.  What do people think?
>>>>> 
>>>>> kmod
>>