`clang-cl /EHa` with MSVC `std::variant` emits linker error LNK2019: unresolved external symbol

[StephanTLavavej]

Repros with Clang 17.0.3 and VS 2022 17.11 Preview 1. MSVC accepts, but Clang rejects.

C:\Temp>"C:\Program Files\Microsoft Visual Studio\2022\Preview\VC\Auxiliary\Build\vcvarsall.bat" x64
**********************************************************************
** Visual Studio 2022 Developer Command Prompt v17.11.0-pre.1.0
** Copyright (c) 2022 Microsoft Corporation
**********************************************************************
[vcvarsall.bat] Environment initialized for: 'x64'

C:\Temp>clang-cl -v
clang version 17.0.3
Target: x86_64-pc-windows-msvc
Thread model: posix
InstalledDir: C:\Program Files\Microsoft Visual Studio\2022\Preview\VC\Tools\Llvm\x64\bin

C:\Temp>type meow.cpp

#include <variant>
using namespace std;

struct UDT {
    UDT() {}
    ~UDT() {}
};

int main() {
    using V = variant<int, double, UDT>;
    [[maybe_unused]] V a{1729};
    [[maybe_unused]] V b{3.14};
    [[maybe_unused]] V c{UDT{}};
}

C:\Temp>cl /EHsc /nologo /W4 /std:c++17 /MTd /Od meow.cpp
meow.cpp

C:\Temp>cl /EHa /nologo /W4 /std:c++17 /MTd /Od meow.cpp
meow.cpp

C:\Temp>clang-cl /EHsc /nologo /W4 /std:c++17 /MTd /Od meow.cpp

C:\Temp>clang-cl /EHa /nologo /W4 /std:c++17 /MTd /Od meow.cpp
meow-aeb2d1.obj : error LNK2019: unresolved external symbol "public: __cdecl std::_Variant_storage_<0,double,struct UDT>::~_Variant_storage_<0,double,struct UDT>(void)" (??1?$_Variant_storage_@$0A@NUUDT@@@std@@QEAA@XZ) referenced in function "int `public: __cdecl std::_Variant_storage_<0,int,double,struct UDT>::_Variant_storage_<0,int,double,struct UDT><1,double,0>(struct _Variant_storage_<0,int,double,struct UDT>::integral_constant<unsigned __int64,1>,double &&)'::`1'::dtor$3" (?dtor$3@?0???$?0$00N$0A@@?$_Variant_storage_@$0A@HNUUDT@@@std@@QEAA@U?$integral_constant@_K$00@1@$$QEAN@Z@4HA)
meow-aeb2d1.obj : error LNK2019: unresolved external symbol "public: __cdecl std::_Variant_storage_<0,struct UDT>::~_Variant_storage_<0,struct UDT>(void)" (??1?$_Variant_storage_@$0A@UUDT@@@std@@QEAA@XZ) referenced in function "int `public: __cdecl std::_Variant_storage_<0,double,struct UDT>::_Variant_storage_<0,double,struct UDT><1,struct UDT,0>(struct _Variant_storage_<0,double,struct UDT>::integral_constant<unsigned __int64,1>,struct UDT &&)'::`1'::dtor$3" (?dtor$3@?0???$?0$00UUDT@@$0A@@?$_Variant_storage_@$0A@NUUDT@@@std@@QEAA@U?$integral_constant@_K$00@1@$$QEAUUDT@@@Z@4HA)
meow.exe : fatal error LNK1120: 2 unresolved externals
clang-cl: error: linker command failed with exit code 1120 (use -v to see invocation)

Reduced from the original user report DevCom-10647850 "Linking bug of std::unordered_map with std::variant Value Type with latest MS Visual Studio Community 2022".

[jneuhaus20]

Has anyone had any luck working around this? I have this exact same problem in a larger codebase. We're converting to clang to move towards Linux, so we'll have to wean ourselves off of SEH eventually, but it'd be real nice to keep it on Windows for the transition.

[phoebewang]

I did an experiment, with [[maybe_unused]] V a{1729}; only, we can find a

define linkonce_odr dso_local void @"??1?$_Variant_storage_@$0A@HNUUDT@@@std@@QEAA@XZ" ...

in dumped IR.
With

    [[maybe_unused]] V a{1729};
    [[maybe_unused]] V b{3.14};

There are

define linkonce_odr dso_local void @"??1?$_Variant_storage_@$0A@HNUUDT@@@std@@QEAA@XZ" ...
declare dso_local void @"??1?$_Variant_storage_@$0A@NUUDT@@@std@@QEAA@XZ" ...

In which, the declare of "??1?$_Variant_storage_@$0A@NUUDT@@@std@@QEAA@XZ" causes unresolved external error.
In the original case, there are

define linkonce_odr dso_local void @"??1?$_Variant_storage_@$0A@HNUUDT@@@std@@QEAA@XZ" ...
declare dso_local void @"??1?$_Variant_storage_@$0A@NUUDT@@@std@@QEAA@XZ" ...
declare dso_local void @"??1?$_Variant_storage_@$0A@UUDT@@@std@@QEAA@XZ" ...

For

"??1?$_Variant_storage_@$0A@HNUUDT@@@std@@QEAA@XZ"

is called once by

"??1?$_Variant_base@HNUUDT@@@std@@QEAA@XZ"

in IR built with /EHsc. But with /EHa, it's called 4 more times in ehcleanup by

"??1?$_Variant_base@HNUUDT@@@std@@QEAA@XZ"
"??$?0$0A@H$0A@@?$_Variant_base@HNUUDT@@@std@@QEAA@U?$in_place_index_t@$0A@@1@$$QEAH@Z"
"??$?0$00N$0A@@?$_Variant_base@HNUUDT@@@std@@QEAA@U?$in_place_index_t@$00@1@$$QEAN@Z"
"??$?0$01UUDT@@$0A@@?$_Variant_base@HNUUDT@@@std@@QEAA@U?$in_place_index_t@$01@1@$$QEAUUDT@@@Z"

Similarly, in /EHa case

"??1?$_Variant_storage_@$0A@NUUDT@@@std@@QEAA@XZ"

is called twice by

"??$?0$00N$0A@@?$_Variant_storage_@$0A@HNUUDT@@@std@@QEAA@U?$integral_constant@_K$00@1@$$QEAN@Z"
"??$?0$01UUDT@@$0A@@?$_Variant_storage_@$0A@HNUUDT@@@std@@QEAA@U?$integral_constant@_K$01@1@$$QEAUUDT@@@Z"

in ehcleanup block only.

I think this explains the difference between define and declare, i.e., Clang won't generate function defination if it's only used by ehcleanup. It is true even if specify -Xclang -disable-llvm-passes. So I believe it is a Clang front end issue.

CC @AaronBallman

[llvmbot]

@llvm/issue-subscribers-clang-codegen

Author: Stephan T. Lavavej (StephanTLavavej)

Repros with Clang 17.0.3 and VS 2022 17.11 Preview 1. MSVC accepts, but Clang rejects.

C:\Temp>"C:\Program Files\Microsoft Visual Studio\2022\Preview\VC\Auxiliary\Build\vcvarsall.bat" x64
**********************************************************************
** Visual Studio 2022 Developer Command Prompt v17.11.0-pre.1.0
** Copyright (c) 2022 Microsoft Corporation
**********************************************************************
[vcvarsall.bat] Environment initialized for: 'x64'

C:\Temp>clang-cl -v
clang version 17.0.3
Target: x86_64-pc-windows-msvc
Thread model: posix
InstalledDir: C:\Program Files\Microsoft Visual Studio\2022\Preview\VC\Tools\Llvm\x64\bin

C:\Temp>type meow.cpp

#include <variant>
using namespace std;

struct UDT {
    UDT() {}
    ~UDT() {}
};

int main() {
    using V = variant<int, double, UDT>;
    [[maybe_unused]] V a{1729};
    [[maybe_unused]] V b{3.14};
    [[maybe_unused]] V c{UDT{}};
}

C:\Temp>cl /EHsc /nologo /W4 /std:c++17 /MTd /Od meow.cpp
meow.cpp

C:\Temp>cl /EHa /nologo /W4 /std:c++17 /MTd /Od meow.cpp
meow.cpp

C:\Temp>clang-cl /EHsc /nologo /W4 /std:c++17 /MTd /Od meow.cpp

C:\Temp>clang-cl /EHa /nologo /W4 /std:c++17 /MTd /Od meow.cpp
meow-aeb2d1.obj : error LNK2019: unresolved external symbol "public: __cdecl std::_Variant_storage_&lt;0,double,struct UDT&gt;::~_Variant_storage_&lt;0,double,struct UDT&gt;(void)" (??1?$_Variant_storage_@$0A@<!-- -->NUUDT@@@<!-- -->std@@<!-- -->QEAA@<!-- -->XZ) referenced in function "int `public: __cdecl std::_Variant_storage_&lt;0,int,double,struct UDT&gt;::_Variant_storage_&lt;0,int,double,struct UDT&gt;&lt;1,double,0&gt;(struct _Variant_storage_&lt;0,int,double,struct UDT&gt;::integral_constant&lt;unsigned __int64,1&gt;,double &amp;&amp;)'::`1'::dtor$3" (?dtor$3@?0???$?0$00N$0A@@?$_Variant_storage_@$0A@<!-- -->HNUUDT@@@<!-- -->std@@<!-- -->QEAA@<!-- -->U?$integral_constant@<!-- -->_K$00@<!-- -->1@$$QEAN@<!-- -->Z@<!-- -->4HA)
meow-aeb2d1.obj : error LNK2019: unresolved external symbol "public: __cdecl std::_Variant_storage_&lt;0,struct UDT&gt;::~_Variant_storage_&lt;0,struct UDT&gt;(void)" (??1?$_Variant_storage_@$0A@<!-- -->UUDT@@@<!-- -->std@@<!-- -->QEAA@<!-- -->XZ) referenced in function "int `public: __cdecl std::_Variant_storage_&lt;0,double,struct UDT&gt;::_Variant_storage_&lt;0,double,struct UDT&gt;&lt;1,struct UDT,0&gt;(struct _Variant_storage_&lt;0,double,struct UDT&gt;::integral_constant&lt;unsigned __int64,1&gt;,struct UDT &amp;&amp;)'::`1'::dtor$3" (?dtor$3@?0???$?0$00UUDT@@$0A@@?$_Variant_storage_@$0A@<!-- -->NUUDT@@@<!-- -->std@@<!-- -->QEAA@<!-- -->U?$integral_constant@<!-- -->_K$00@<!-- -->1@$$QEAUUDT@@@<!-- -->Z@<!-- -->4HA)
meow.exe : fatal error LNK1120: 2 unresolved externals
clang-cl: error: linker command failed with exit code 1120 (use -v to see invocation)

Reduced from the original user report DevCom-10647850 "Linking bug of std::unordered_map with std::variant Value Type with latest MS Visual Studio Community 2022".

[AaronBallman]

I think this explains the difference between define and declare, i.e., Clang won't generate function defination if it's only used by ehcleanup. It is true even if specify -Xclang -disable-llvm-passes.

Thank you for the investigation!

So I believe it is a Clang front end issue.

CC @rjmccall @efriedma-quic @rnk for awareness (or to see if you have any bandwidth for addressing this one, as it seems pretty important)

[efriedma-quic]

At first glance, this doesn't really make sense to me. The only way I can see that we wouldn't emit the definition of a destructor is if the destructor isn't odr-used. I have no idea how you could declare a variable in a way that would make CodeGen try to call a destructor that isn't odr-used. (If the destructor is in fact odr-used, some core infrastructure would have to be very fundamentally broken to avoid triggering deferred emission.)

If someone could reduce a testcase that doesn't require including all of std::variant, that would be helpful.

[AaronBallman]

CC @Endilll for help with reducing, in case he's got bandwidth