[CodeGen] Make __clang_call_terminate have an unwind table entry

[smeenai]

Summary

This enables unwinders to step past that frame on architectures that
don't use DWARF unwinding (such as armv7), e.g. when debugging. The
problem should theoretically be architecture-agnostic, but according to
https://discourse.llvm.org/t/51633/2 it gets masked on architectures
that use DWARF unwind info.

Fixes https://github.com/llvm/llvm-project/issues/40696

[efriedma]

clang/lib/CodeGen/ItaniumCXXABI.cpp

4699

We probably want to call SetLLVMFunctionAttributesForDefinition() here instead. The end result isn't that much different, but it indicates the intent more clearly, and hopefully we avoid hitting similar issues in this code in the future.

[smeenai]

Call SetLLVMFunctionAttributesForDefinition instead

[rnk]

Makes sense to me too, but wait for approval by @efriedma

[efriedma]

LGTM

[MaskRay]

LGTM. Is there an existing test that can merge in the uwtable test?

[smeenai]

I searched for __clang_call_terminate in the Clang test directory when I was working on this diff. Most tests were testing calls to it, not definitions. There were a couple of tests checking the definition, but they seemed pretty targeted (e.g. https://github.com/llvm/llvm-project/blob/main/clang/test/CodeGenCXX/noexcept.cpp) or auto-generated, so I figured having a separate test case for the specific thing I cared about was best.

[MaskRay]

In D151393#4373764, @smeenai wrote:

I searched for __clang_call_terminate in the Clang test directory when I was working on this diff. Most tests were testing calls to it, not definitions. There were a couple of tests checking the definition, but they seemed pretty targeted (e.g. https://github.com/llvm/llvm-project/blob/main/clang/test/CodeGenCXX/noexcept.cpp) or auto-generated, so I figured having a separate test case for the specific thing I cared about was best.

Thank you for the research!

clang/lib/CodeGen/ItaniumCXXABI.cpp

const CGFunctionInfo &FI = CGM.getTypes().arrangeBuiltinFunctionDeclaration(
       C.VoidTy, {C.getPointerType(C.CharTy)});
   llvm::FunctionType *fnTy = CGM.getTypes().GetFunctionType(FI);
   llvm::FunctionCallee fnRef = CGM.CreateRuntimeFunction(
       fnTy, "__clang_call_terminate", llvm::AttributeList(), /*Local=*/true);
   llvm::Function *fn =
       cast<llvm::Function>(fnRef.getCallee()->stripPointerCasts());
   if (fn->empty()) {
     CGM.SetLLVMFunctionAttributes(GlobalDecl(), FI, fn, /*IsThunk=*/false);
+    CGM.SetLLVMFunctionAttributesForDefinition(nullptr, fn);
     fn->setDoesNotThrow();
     fn->setDoesNotReturn();
 
     // What we really want is to massively penalize inlining without
     // forbidding it completely.  The difference between that and
     // 'noinline' is negligible.
     fn->addFnAttr(llvm::Attribute::NoInline);
 
     // Allow this function to be shared across translation units, but
     // we don't want it to turn into an exported symbol.
     fn->setLinkage(llvm::Function::LinkOnceODRLinkage);
     fn->setVisibility(llvm::Function::HiddenVisibility);
     if (CGM.supportsCOMDAT())
       fn->setComdat(CGM.getModule().getOrInsertComdat(fn->getName()));
 

clang/test/CodeGenCXX/clang-call-terminate.uwtable.cpp

+// RUN: %clang_cc1 -triple=x86_64-linux-gnu -fexceptions -fcxx-exceptions -emit-llvm -o - %s | \
+// RUN:   FileCheck --check-prefixes=CHECK,NOUNWIND %s
+// RUN: %clang_cc1 -triple=x86_64-linux-gnu -fexceptions -fcxx-exceptions -funwind-tables=1 -emit-llvm -o - %s | \
+// RUN:   FileCheck --check-prefixes=CHECK,SYNCUNWIND %s
+// RUN: %clang_cc1 -triple=x86_64-linux-gnu -fexceptions -fcxx-exceptions -funwind-tables=2 -emit-llvm -o - %s | \
+// RUN:   FileCheck --check-prefixes=CHECK,ASYNCUNWIND %s
+
+void caller(void callback()) noexcept { callback(); }
+
+// CHECK: define {{.*}}void @__clang_call_terminate({{[^)]*}}) #[[#ATTRNUM:]]
+// CHECK: attributes #[[#ATTRNUM]] = {
+// NOUNWIND-NOT: uwtable
+// NOUNWIND-SAME: }
+// SYNCUNWIND-SAME: uwtable(sync)
+// ASYNCUNWIND-SAME: uwtable{{ }}