[clang] [CIR] Record if c++ ABI was overridden via -fc++-abi flag (PR #229494)
Bruno Cardoso Lopes via cfe-commits
cfe-commits at lists.llvm.org
Wed Oct 7 14:50:13 PDT 2026
================
@@ -257,6 +258,42 @@ cir::CIRDialect::verifyOperationAttribute(mlir::Operation *op,
if (isOpenCLVersionAttrName(attrName))
return verifyOpenCLVersionAttr(op, attr);
+ if (attrName == getCXXABIAttrName()) {
+ if (!mlir::isa<mlir::ModuleOp>(op))
+ return op->emitOpError() << "expects '" << getCXXABIAttrName()
+ << "' attribute to be attached to '"
+ << mlir::ModuleOp::getOperationName() << "'";
+ auto abi = mlir::dyn_cast<mlir::StringAttr>(attr.getValue());
+ if (!abi)
+ return op->emitOpError()
+ << "expects '" << getCXXABIAttrName() << "' to be a string";
+
+ StringRef abival = abi.getValue();
+ if (!clang::TargetCXXABI::isABI(abival))
+ return op->emitOpError() << "unknown C++ ABI '" << abi.getValue() << "'";
+
+ clang::TargetCXXABI::Kind kind =
+ clang::TargetCXXABI::getKind(abival);
+ if (auto triple = op->getAttrOfType<mlir::StringAttr>(getTripleAttrName()))
+ if (!clang::TargetCXXABI::isSupportedCXXABI(
+ llvm::Triple(triple.getValue()), kind))
+ return op->emitOpError()
+ << "C++ ABI '" << abi.getValue()
+ << "' is not supported on target '" << triple.getValue() << "'";
+
+ switch (kind) {
----------------
bcardosolopes wrote:
The unknown-ABI and triple checks belong in the verifier, but this switch doesn't. `cir.cxx_abi = "fuchsia"` on a Fuchsia triple is well-formed IR. It just isn't handled by TargetLowering yet, and `createItaniumCXXABI` already makes that decision, with `default: llvm_unreachable("Other Itanium ABI?")`. Keeping the list here means two copies of the same policy that will drift, and the IR stops parsing for a reason unrelated to the IR.
Please drop the switch and turn that `llvm_unreachable` into an NYI emit error in the lowering instead. That also covers ABIs that come from the triple without the attribute, which this check never sees.
https://github.com/llvm/llvm-project/pull/229494
More information about the cfe-commits
mailing list