[lldb] Assertion Crash when evaluating structures with bases with non-natural alignment

[hokein]

A small source code:

#include <coroutine>
#include <cstddef>


struct alignas(std::max_align_t) promise_base {
 int count = 0;
};

struct task{
  struct promise_type : promise_base {
    task get_return_object() { return std::coroutine_handle<promise_type>::from_promise(*this); }
    std::suspend_always initial_suspend() { return {}; }
    std::suspend_always final_suspend() noexcept { return {}; }
    void return_void() noexcept {}
    void unhandled_exception() noexcept {}
    
    int b;
  };

  void resume() noexcept {
    handle.resume();
  }

  task(std::coroutine_handle<promise_type> hdl) : handle(hdl) {}
  ~task() {
    if (handle)
      handle.destroy();
  }

  std::coroutine_handle<> handle;
};

 task coro_task(int v) {
  int a = v;
  __builtin_debugtrap();
  co_return;
}

int main() {
  auto k = coro_task(1);
  k.resume();
  return 0;
}

Reproduce steps:

$ ./bin/clang++ -std=c++20 -g /tmp/t.cpp -o /tmp/debugging-example
$ ./bin/lldb --one-line "run" --one-line "expr __promise" /tmp/debugging-example    

We hit an assertion, stack trace:

assert.h assertion failed at llvm-project/clang/lib/CodeGen/CGRecordLayoutBuilder.cpp:754 in void (anonymous namespace)::CGRecordLowering::clipTailPadding(): Prior->Kind == MemberInfo::Field && "Only storage fields have tail padding!"
*** Check failure stack trace: ***
    @   absl::log_internal::LogMessage::SendToLog()
    @   absl::log_internal::LogMessage::Flush()
    @   absl::log_internal::LogMessageFatal::~LogMessageFatal()
    @   __assert_fail
    @   (anonymous namespace)::CGRecordLowering::lower()
    @   clang::CodeGen::CodeGenTypes::ComputeRecordLayout()
    @   clang::CodeGen::CodeGenTypes::ConvertRecordDeclType()
    @   clang::CodeGen::CodeGenTypes::ConvertType()
    @   clang::CodeGen::CodeGenTypes::ConvertTypeForMem()
    @   clang::CodeGen::CodeGenFunction::EmitLoadOfReference()
    @   clang::CodeGen::CodeGenFunction::EmitLoadOfReferenceLValue()
    @   clang::CodeGen::CodeGenFunction::EmitDeclRefLValue()
    @   clang::CodeGen::CodeGenFunction::EmitLValueHelper()
    @   clang::CodeGen::CodeGenFunction::EmitLValue()
    @   (anonymous namespace)::ScalarExprEmitter::VisitUnaryAddrOf()
    @   clang::StmtVisitorBase<>::Visit()
    @   clang::CodeGen::CodeGenFunction::EmitScalarExpr()
    @   clang::CodeGen::CodeGenFunction::EmitScalarInit()
    @   clang::CodeGen::CodeGenFunction::EmitCXXGlobalVarDeclInit()
    @   (anonymous namespace)::ItaniumCXXABI::EmitGuardedInit()
    @   clang::CodeGen::CodeGenFunction::AddInitializerToStaticVarDecl()
    @   clang::CodeGen::CodeGenFunction::EmitStaticVarDecl()
    @   clang::CodeGen::CodeGenFunction::EmitDecl()
    @   clang::CodeGen::CodeGenFunction::EmitDeclStmt()
    @   clang::CodeGen::CodeGenFunction::EmitSimpleStmt()
    @   clang::CodeGen::CodeGenFunction::EmitStmt()
    @   clang::CodeGen::CodeGenFunction::EmitCompoundStmtWithoutScope()
    @   clang::CodeGen::CodeGenFunction::GenerateCode()
    @   clang::CodeGen::CodeGenModule::EmitGlobalFunctionDefinition()
    @   clang::CodeGen::CodeGenModule::EmitGlobalDefinition()
    @   clang::CodeGen::CodeGenModule::EmitGlobal()
    @   clang::CodeGen::CodeGenModule::EmitTopLevelDecl()
    @   (anonymous namespace)::CodeGeneratorImpl::HandleTopLevelDecl()
    @   lldb_private::ASTResultSynthesizer::HandleTopLevelDecl()
    @   clang::ParseAST()
    @   lldb_private::ClangExpressionParser::ParseInternal()
    @   lldb_private::ClangUserExpression::TryParse()
    @   lldb_private::ClangUserExpression::Parse()
    @   lldb_private::UserExpression::Evaluate()
    @   lldb_private::Target::EvaluateExpression()
    @   lldb_private::CommandObjectExpression::EvaluateExpression()
    @   lldb_private::CommandObjectExpression::DoExecute()
    @   lldb_private::CommandObjectRaw::Execute()
    @   lldb_private::CommandInterpreter::HandleCommand()
    @   lldb_private::CommandInterpreter::IOHandlerInputComplete()
    @   lldb_private::IOHandlerEditline::Run()
    @   lldb_private::Debugger::RunIOHandlers()
    @   lldb_private::CommandInterpreter::RunCommandInterpreter()
    @   lldb::SBDebugger::RunCommandInterpreter()
    @   Driver::MainLoop()
    @   main
    @   __libc_start_main
    @   _start
PLEASE submit a bug report to http://go/llvm-crash-bug and include the crash backtrace.
Stack dump:
0.      Program arguments: llvm/llvm-project/lldb/lldb /tmp/debugging-example --one-line run --one-line "expr __promise" /tmp/debugging-example
1.      HandleCommand(command = "expr __promise")
2.      <eof> parser at end of file
3.      <lldb wrapper prefix>:43:1: Generating code for declaration '$__lldb_expr'

[llvmbot]

@llvm/issue-subscribers-coroutines

Author: Haojian Wu (hokein)

A small source code:

#include <coroutine>
#include <cstddef>


struct alignas(std::max_align_t) promise_base {
 int count = 0;
};

struct task{
  struct promise_type : promise_base {
    task get_return_object() { return std::coroutine_handle<promise_type>::from_promise(*this); }
    std::suspend_always initial_suspend() { return {}; }
    std::suspend_always final_suspend() noexcept { return {}; }
    void return_void() noexcept {}
    void unhandled_exception() noexcept {}
    
    int b;
  };

  void resume() noexcept {
    handle.resume();
  }

  task(std::coroutine_handle<promise_type> hdl) : handle(hdl) {}
  ~task() {
    if (handle)
      handle.destroy();
  }

  std::coroutine_handle<> handle;
};

 task coro_task(int v) {
  int a = v;
  __builtin_debugtrap();
  co_return;
}

int main() {
  auto k = coro_task(1);
  k.resume();
  return 0;
}

Reproduce steps:

$ ./bin/clang++ -std=c++20 -g /tmp/t.cpp -o /tmp/debugging-example
$ ./bin/lldb --one-line "run" --one-line "expr __promise" /tmp/debugging-example    

We hit an assertion, stack trace:

assert.h assertion failed at llvm-project/clang/lib/CodeGen/CGRecordLayoutBuilder.cpp:754 in void (anonymous namespace)::CGRecordLowering::clipTailPadding(): Prior->Kind == MemberInfo::Field && "Only storage fields have tail padding!"
*** Check failure stack trace: ***
    @   absl::log_internal::LogMessage::SendToLog()
    @   absl::log_internal::LogMessage::Flush()
    @   absl::log_internal::LogMessageFatal::~LogMessageFatal()
    @   __assert_fail
    @   (anonymous namespace)::CGRecordLowering::lower()
    @   clang::CodeGen::CodeGenTypes::ComputeRecordLayout()
    @   clang::CodeGen::CodeGenTypes::ConvertRecordDeclType()
    @   clang::CodeGen::CodeGenTypes::ConvertType()
    @   clang::CodeGen::CodeGenTypes::ConvertTypeForMem()
    @   clang::CodeGen::CodeGenFunction::EmitLoadOfReference()
    @   clang::CodeGen::CodeGenFunction::EmitLoadOfReferenceLValue()
    @   clang::CodeGen::CodeGenFunction::EmitDeclRefLValue()
    @   clang::CodeGen::CodeGenFunction::EmitLValueHelper()
    @   clang::CodeGen::CodeGenFunction::EmitLValue()
    @   (anonymous namespace)::ScalarExprEmitter::VisitUnaryAddrOf()
    @   clang::StmtVisitorBase<>::Visit()
    @   clang::CodeGen::CodeGenFunction::EmitScalarExpr()
    @   clang::CodeGen::CodeGenFunction::EmitScalarInit()
    @   clang::CodeGen::CodeGenFunction::EmitCXXGlobalVarDeclInit()
    @   (anonymous namespace)::ItaniumCXXABI::EmitGuardedInit()
    @   clang::CodeGen::CodeGenFunction::AddInitializerToStaticVarDecl()
    @   clang::CodeGen::CodeGenFunction::EmitStaticVarDecl()
    @   clang::CodeGen::CodeGenFunction::EmitDecl()
    @   clang::CodeGen::CodeGenFunction::EmitDeclStmt()
    @   clang::CodeGen::CodeGenFunction::EmitSimpleStmt()
    @   clang::CodeGen::CodeGenFunction::EmitStmt()
    @   clang::CodeGen::CodeGenFunction::EmitCompoundStmtWithoutScope()
    @   clang::CodeGen::CodeGenFunction::GenerateCode()
    @   clang::CodeGen::CodeGenModule::EmitGlobalFunctionDefinition()
    @   clang::CodeGen::CodeGenModule::EmitGlobalDefinition()
    @   clang::CodeGen::CodeGenModule::EmitGlobal()
    @   clang::CodeGen::CodeGenModule::EmitTopLevelDecl()
    @   (anonymous namespace)::CodeGeneratorImpl::HandleTopLevelDecl()
    @   lldb_private::ASTResultSynthesizer::HandleTopLevelDecl()
    @   clang::ParseAST()
    @   lldb_private::ClangExpressionParser::ParseInternal()
    @   lldb_private::ClangUserExpression::TryParse()
    @   lldb_private::ClangUserExpression::Parse()
    @   lldb_private::UserExpression::Evaluate()
    @   lldb_private::Target::EvaluateExpression()
    @   lldb_private::CommandObjectExpression::EvaluateExpression()
    @   lldb_private::CommandObjectExpression::DoExecute()
    @   lldb_private::CommandObjectRaw::Execute()
    @   lldb_private::CommandInterpreter::HandleCommand()
    @   lldb_private::CommandInterpreter::IOHandlerInputComplete()
    @   lldb_private::IOHandlerEditline::Run()
    @   lldb_private::Debugger::RunIOHandlers()
    @   lldb_private::CommandInterpreter::RunCommandInterpreter()
    @   lldb::SBDebugger::RunCommandInterpreter()
    @   Driver::MainLoop()
    @   main
    @   __libc_start_main
    @   _start
PLEASE submit a bug report to http://go/llvm-crash-bug and include the crash backtrace.
Stack dump:
0.      Program arguments: llvm/llvm-project/lldb/lldb /tmp/debugging-example --one-line run --one-line "expr __promise" /tmp/debugging-example
1.      HandleCommand(command = "expr __promise")
2.      <eof> parser at end of file
3.      <lldb wrapper prefix>:43:1: Generating code for declaration '$__lldb_expr'

[llvmbot]

@llvm/issue-subscribers-lldb

Author: Haojian Wu (hokein)

A small source code:

#include <coroutine>
#include <cstddef>


struct alignas(std::max_align_t) promise_base {
 int count = 0;
};

struct task{
  struct promise_type : promise_base {
    task get_return_object() { return std::coroutine_handle<promise_type>::from_promise(*this); }
    std::suspend_always initial_suspend() { return {}; }
    std::suspend_always final_suspend() noexcept { return {}; }
    void return_void() noexcept {}
    void unhandled_exception() noexcept {}
    
    int b;
  };

  void resume() noexcept {
    handle.resume();
  }

  task(std::coroutine_handle<promise_type> hdl) : handle(hdl) {}
  ~task() {
    if (handle)
      handle.destroy();
  }

  std::coroutine_handle<> handle;
};

 task coro_task(int v) {
  int a = v;
  __builtin_debugtrap();
  co_return;
}

int main() {
  auto k = coro_task(1);
  k.resume();
  return 0;
}

Reproduce steps:

$ ./bin/clang++ -std=c++20 -g /tmp/t.cpp -o /tmp/debugging-example
$ ./bin/lldb --one-line "run" --one-line "expr __promise" /tmp/debugging-example    

We hit an assertion, stack trace:

assert.h assertion failed at llvm-project/clang/lib/CodeGen/CGRecordLayoutBuilder.cpp:754 in void (anonymous namespace)::CGRecordLowering::clipTailPadding(): Prior->Kind == MemberInfo::Field && "Only storage fields have tail padding!"
*** Check failure stack trace: ***
    @   absl::log_internal::LogMessage::SendToLog()
    @   absl::log_internal::LogMessage::Flush()
    @   absl::log_internal::LogMessageFatal::~LogMessageFatal()
    @   __assert_fail
    @   (anonymous namespace)::CGRecordLowering::lower()
    @   clang::CodeGen::CodeGenTypes::ComputeRecordLayout()
    @   clang::CodeGen::CodeGenTypes::ConvertRecordDeclType()
    @   clang::CodeGen::CodeGenTypes::ConvertType()
    @   clang::CodeGen::CodeGenTypes::ConvertTypeForMem()
    @   clang::CodeGen::CodeGenFunction::EmitLoadOfReference()
    @   clang::CodeGen::CodeGenFunction::EmitLoadOfReferenceLValue()
    @   clang::CodeGen::CodeGenFunction::EmitDeclRefLValue()
    @   clang::CodeGen::CodeGenFunction::EmitLValueHelper()
    @   clang::CodeGen::CodeGenFunction::EmitLValue()
    @   (anonymous namespace)::ScalarExprEmitter::VisitUnaryAddrOf()
    @   clang::StmtVisitorBase<>::Visit()
    @   clang::CodeGen::CodeGenFunction::EmitScalarExpr()
    @   clang::CodeGen::CodeGenFunction::EmitScalarInit()
    @   clang::CodeGen::CodeGenFunction::EmitCXXGlobalVarDeclInit()
    @   (anonymous namespace)::ItaniumCXXABI::EmitGuardedInit()
    @   clang::CodeGen::CodeGenFunction::AddInitializerToStaticVarDecl()
    @   clang::CodeGen::CodeGenFunction::EmitStaticVarDecl()
    @   clang::CodeGen::CodeGenFunction::EmitDecl()
    @   clang::CodeGen::CodeGenFunction::EmitDeclStmt()
    @   clang::CodeGen::CodeGenFunction::EmitSimpleStmt()
    @   clang::CodeGen::CodeGenFunction::EmitStmt()
    @   clang::CodeGen::CodeGenFunction::EmitCompoundStmtWithoutScope()
    @   clang::CodeGen::CodeGenFunction::GenerateCode()
    @   clang::CodeGen::CodeGenModule::EmitGlobalFunctionDefinition()
    @   clang::CodeGen::CodeGenModule::EmitGlobalDefinition()
    @   clang::CodeGen::CodeGenModule::EmitGlobal()
    @   clang::CodeGen::CodeGenModule::EmitTopLevelDecl()
    @   (anonymous namespace)::CodeGeneratorImpl::HandleTopLevelDecl()
    @   lldb_private::ASTResultSynthesizer::HandleTopLevelDecl()
    @   clang::ParseAST()
    @   lldb_private::ClangExpressionParser::ParseInternal()
    @   lldb_private::ClangUserExpression::TryParse()
    @   lldb_private::ClangUserExpression::Parse()
    @   lldb_private::UserExpression::Evaluate()
    @   lldb_private::Target::EvaluateExpression()
    @   lldb_private::CommandObjectExpression::EvaluateExpression()
    @   lldb_private::CommandObjectExpression::DoExecute()
    @   lldb_private::CommandObjectRaw::Execute()
    @   lldb_private::CommandInterpreter::HandleCommand()
    @   lldb_private::CommandInterpreter::IOHandlerInputComplete()
    @   lldb_private::IOHandlerEditline::Run()
    @   lldb_private::Debugger::RunIOHandlers()
    @   lldb_private::CommandInterpreter::RunCommandInterpreter()
    @   lldb::SBDebugger::RunCommandInterpreter()
    @   Driver::MainLoop()
    @   main
    @   __libc_start_main
    @   _start
PLEASE submit a bug report to http://go/llvm-crash-bug and include the crash backtrace.
Stack dump:
0.      Program arguments: llvm/llvm-project/lldb/lldb /tmp/debugging-example --one-line run --one-line "expr __promise" /tmp/debugging-example
1.      HandleCommand(command = "expr __promise")
2.      <eof> parser at end of file
3.      <lldb wrapper prefix>:43:1: Generating code for declaration '$__lldb_expr'

[vogelsgesang]

#58397 might be vaguely related. Currently, clang misplaces the alignment within the coroutine frame.

This will also become an issue with lldb's std::coroutine_handle pretty-printer which assumes that the promise is at a fixed offset of 16 bytes, as mandated by the ABI. But afaik, clang does not follow the ABI when generating code, yet

[Michael137]

Doesn't seem coroutine related.

This also crashes:

struct alignas(8) bar {               
    int count = 0;                    
};                                    
                                      
struct foo : bar {                    
    int b = 0;                        
};                                    
                                      
int main() {                          
    foo p;                            
    __builtin_debugtrap();  // expr p 
    return 0;                         
}                                             

So it seems LLDB doesn't handle evaluating structures with bases with non-natural alignment

[hokein]

Thanks for the helpful comments. @vogelsgesang, @Michael137

Yes, this is irrelevant to coroutines (removed the wrong label).

I think it is unlikely a bug in clang's codeGen backend. My hypothesis is that the alignment information is somehow messed up when lldb builds the AST node RecordDecl. I will take a further look.

[Michael137]

Clang attaches DW_AT_alignment to bar in my example, but LLDB doesn't ever actually use this. The fix is probably to parse said attribute and make sure we bubble it up to ExternalASTSource::layoutRecordType when the ItaniumRecordLayoutBuilder asks for it. We may have to also attach clang::AlignedAttr to the decl we create for bar, but not sure

[hokein]

Confirmed that the layout of the RecordDecl bar built in the lldb is wrong. The layout dump is

 0 | struct bar
 0 |   int count
   | [sizeof=8, dsize=8, align=4,
   |  nvsize=8, nvalign=4]

while a correct one emitted by clang is

  0 | struct bar
  0 |   int count
    | [sizeof=8, dsize=4, align=8,
    |  nvsize=4, nvalign=8]

Clang attaches DW_AT_alignment to bar in my example, but LLDB doesn't ever actually use this. The fix is probably to parse said attribute and make sure we bubble it up to ExternalASTSource::layoutRecordType when the ItaniumRecordLayoutBuilder asks for it. We may have to also attach clang::AlignedAttr to the decl we create for bar, but not sure

Thanks! This makes sense, will try to prepare a fix.

[Michael137]

Something I noticed that makes this trickier is that RecordDecl::isTrivial() returns true for bar when using LLDBs AST. Whereas with clang it's actually false (which is correct because bar has a default member initialiser, making the default constructor non-trivial). This is important for mustSkipTailPadding, to determine whether to round up the size to alignment or not. So that's a separate bug that needs fixing first

 0 | struct bar
 0 |   int count
   | [sizeof=8, dsize=8, align=4,
   |  nvsize=8, nvalign=4]

Thanks for checking. Yea so here alignment is wrong, but also dsize (due to the bug I described above).

[Michael137]

Something I noticed that makes this trickier is that RecordDecl::isTrivial() returns true for bar when using LLDBs AST. Whereas with clang it's actually false (which is correct because bar has a default member initialiser, making the default constructor non-trivial).

Don't actually see a good way of fixing this with the current DWARF output. We need to set the FieldDecl::StorageKind to ICIS_CopyInit, which LLDB can't know because it's not spelled out in DWARF anywhere. So we'd have to come up with some way to get that info into DWARF first. Couldn't find an existing attribute that's supposed to give us that info

[hokein]

Something I noticed that makes this trickier is that RecordDecl::isTrivial() returns true for bar when using LLDBs AST. Whereas with clang it's actually false (which is correct because bar has a default member initialiser, making the default constructor non-trivial). This is important for mustSkipTailPadding, to determine whether to round up the size to alignment or not. So that's a separate bug that needs fixing first

Good catch! I didn't pay attention to other fields. This looks like two separate bugs in lldb, they don't have a hard dependency. Fixing the original alignment issue can still make some cases work, which is better than current situation.

Don't actually see a good way of fixing this with the current DWARF output. We need to set the FieldDecl::StorageKind to ICIS_CopyInit, which LLDB can't know because it's not spelled out in DWARF anywhere. So we'd have to come up with some way to get that info into DWARF first. Couldn't find an existing attribute that's supposed to give us that info

Thanks for taking a look at this. Since we're mostly interested in whether a RecordDecl is trivial, is there an trivial-like attribute for DW_TAG_structure_type tag? I didn't find any related attribute from a quick search. If none of these attribute exist, would it be possible to implement an extension?

[Michael137]

Good catch! I didn't pay attention to other fields. This looks like two separate bugs in lldb, they don't have a hard dependency. Fixing the original alignment issue can still make some cases work, which is better than current situation.

Agreed. Fixing the alignment problem might be tricky because LLDB authoritatively tells clang about all the field/base offsets (see ExternalASTSource::layoutRecordType). So if a base class has a stricter alignment than the derived class, we'd have to make sure we reflect that in the layout calculations.

Don't actually see a good way of fixing this with the current DWARF output. We need to set the FieldDecl::StorageKind to ICIS_CopyInit, which LLDB can't know because it's not spelled out in DWARF anywhere. So we'd have to come up with some way to get that info into DWARF first. Couldn't find an existing attribute that's supposed to give us that info

Thanks for taking a look at this. Since we're mostly interested in whether a RecordDecl is trivial, is there an trivial-like attribute for DW_TAG_structure_type tag? I didn't find any related attribute from a quick search. If none of these attribute exist, would it be possible to implement an extension?

Perhaps @dwblaikie @adrian-prantl are aware attributes/past DWARF proposals that represent this. If motivation is compelling enough, such attribute would be easy to add