CGCleanup.h:583: void clang::CodeGen::EHScopeStack::popTerminate(): Assertion `!empty() && "popping exception stack when not empty"' failed

[jsonn]

Bugzilla Link	44987
Resolution	FIXED
Resolved on	Apr 20, 2020 07:29
Version	trunk
OS	Linux
Blocks	#44654
Attachments	Reduced test case
CC	@berolinux,@zmodem,@zygoloid,@rnk,@yurivict
Fixed by commit(s)	ce5173c 68cd4f7

Extended Description

Compile the attached test case with -O1 or higher.

[jsonn]

Reid, bisect points to 28328c3 (D67028) as origin of this regression. Can you take a look please?

[rnk]

A bit more reduced:

struct a {
virtual void primary();
};
struct d {
virtual void e(int, ...);
};
class f : a, d {
virtual void primary();
void e(int, ...) noexcept;
};
void f::primary() {}

There are two vtable slots for f::e, one for a's vtable and one for d's. I'm not sure exactly which thunk causes the issue, but after my change, we use a musttail call so that we don't have to clone the body of f::e. We can't use a regular call like normal because of the variadic arguments.

But, the callee is actually noexcept, so we don't need the cleanup in the thunk in the first place. That seems like a generally beneficial change.

[rnk]

*** Bug llvm/llvm-bugzilla-archive#45251 has been marked as a duplicate of this bug. ***

[rnk]

The issue is the way that StartThunk calls StartFunction with a null Decl. In my musttail code path, I called FinishFunction instead of FinishThunk. FinishThunk zeroes out CurCodeDecl before finishing the function, which disables the EH spec codepath.

So the fix is just:

$ git diff
diff --git a/clang/lib/CodeGen/CGVTables.cpp b/clang/lib/CodeGen/CGVTables.cpp
index 403b9e25f7a..412fc693c5b 100644
--- a/clang/lib/CodeGen/CGVTables.cpp
+++ b/clang/lib/CodeGen/CGVTables.cpp
@@ -437,7 +437,8 @@ void CodeGenFunction::EmitMustTailThunk(GlobalDecl GD,
// Finish the function to maintain CodeGenFunction invariants.
// FIXME: Don't emit unreachable code.
EmitBlock(createBasicBlock());

• FinishFunction();

• FinishThunk();
  }

void CodeGenFunction::generateThunk(llvm::Function *Fn,

Plus tests.

[rnk]

Interestingly, if you remove the zeroing from FinishThunk, only two tests fail, and they are not directly related to thunks, they have to do with destructors, which often use thunks and are usually noexcept:

Failing Tests (2):
Clang :: CodeGenCXX/cxx0x-delegating-ctors.cpp
Clang :: CodeGenCXX/destructors.cpp

[rnk]

There is a related issue here on i686-windows-msvc with inalloca, the original reason for this codepath:

$ cat t.cpp
struct NonTrivial {
NonTrivial();
NonTrivial(const NonTrivial &o);
~NonTrivial();
};
struct A { virtual void f(NonTrivial o) noexcept; };
struct B { virtual void f(NonTrivial o) noexcept; };
class C : A, B { virtual void f(NonTrivial o) noexcept; };
C c;

$ clang --target=i686-windows-msvc -c t.cpp |& head -2
clang: /usr/local/google/home/rnk/llvm-project/clang/lib/CodeGen/CGCleanup.h:583: void clang::CodeGen::EHScopeStack::popTerminate(): Assertion `!empty() && "popping exception stack when not empty"' failed.

This bug has been present since the invariant of setting and unsetting CurCodeDecl was added in 7f416cc (2015).

[rnk]

Pushed: http://github.com/llvm/llvm-project/commit/ce5173c0e174870934d1b3a026f631d996136191

I assume you want to merge it, so I'll leave it open until we do that.

[zmodem]

Pushed:
http://github.com/llvm/llvm-project/commit/
ce5173c

I assume you want to merge it, so I'll leave it open until we do that.

I think it's too late to merge this for 10.0.0. I'll add it to the list for 10.0.1.

[zmodem]

Re-opening to track for 10.0.1.

[tstellar]

Merged: 68cd4f7

[DimitryAndric]

*** Bug llvm/llvm-bugzilla-archive#45611 has been marked as a duplicate of this bug. ***

[rnk]

mentioned in issue llvm/llvm-bugzilla-archive#45251

[tstellar]

mentioned in issue #44654

[DimitryAndric]

mentioned in issue llvm/llvm-bugzilla-archive#45611