================
@@ -5448,12 +5448,13 @@ CodeGenFunction::getBundlesForFunclet(llvm::Value
*Callee) {
llvm::CallInst *CodeGenFunction::EmitRuntimeCall(llvm::FunctionCallee callee,
ArrayRef<llvm::Value *> args,
const llvm::Twine &name) {
- llvm::CallInst *call = Builder.CreateCall(
- callee, args, getBundlesForFunclet(callee.getCallee()), name);
+ SmallVector<llvm::OperandBundleDef> BundleList =
+ getBundlesForFunclet(callee.getCallee());
+ if (auto *F = dyn_cast<llvm::Function>(callee.getCallee());
+ F && F->isConvergent())
----------------
efriedma-quic wrote:
You're saying that indirect calls shouldn't go through this codepath, or
CodeGenFunction::GenerateCXXGlobalCleanUpFunc?
That's... maybe mostly correct? GenerateCXXGlobalCleanUpFunc should only be
calling either helper functions explicitly created to be called by
GenerateCXXGlobalCleanUpFunc, or actual destructor declarations. And
EmitRuntimeCall gets used for a bunch of stuff, but mostly not calls written by
the user.
I'm not sure we consistently end up with a literal "llvm::Function", though;
you could easily end up with a GlobalAlias.
Whatever we actually get here, we need to more aggressively assert that the
callee actually has the form we're expecting, or else we'll end up obscure
miscompiles or crashes.
https://github.com/llvm/llvm-project/pull/230765
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
