EHABILowering.cpp

Go to the documentation of this file.

//===- EHABILowering.cpp - Lower flattened CIR EH ops to ABI-specific form ===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// This file implements a pass that lowers ABI-agnostic flattened CIR exception
// handling operations into an ABI-specific form. Currently only the Itanium
// C++ ABI is supported.
//
// The Itanium ABI lowering performs these transformations:
//   - cir.eh.initiate            → cir.eh.inflight_exception (landing pad)
//   - cir.eh.dispatch            → cir.eh.typeid + cir.cmp + cir.brcond chains
//   - cir.begin_cleanup          → (removed)
//   - cir.end_cleanup            → (removed)
//   - cir.begin_catch            → call to __cxa_begin_catch
//   - cir.end_catch              → call to __cxa_end_catch
//   - cir.eh.terminate           → call to __clang_call_terminate + unreachable
//   - cir.eh.unexpected          → call to __cxa_call_unexpected + unreachable
//   - cir.resume                 → cir.resume.flat
//   - !cir.eh_token values       → (!cir.ptr<!void>, !u32i) value pairs
//   - cir.construct_catch_param  → __cxa_get_exception_ptr + inlined
//                                  catch-copy thunk body, or a store of the
//                                  exception object address for
//                                  reference_to_pointer
//   - personality function set on functions requiring EH and removed from
//     functions without landing pads
//
//===----------------------------------------------------------------------===//
 
#include "PassDetail.h"
#include "mlir/IR/Builders.h"
#include "mlir/IR/IRMapping.h"
#include "mlir/IR/PatternMatch.h"
#include "clang/CIR/Dialect/IR/CIRDataLayout.h"
#include "clang/CIR/Dialect/IR/CIRDialect.h"
#include "clang/CIR/Dialect/IR/CIROpsEnums.h"
#include "clang/CIR/Dialect/IR/CIRTypes.h"
#include "clang/CIR/Dialect/Passes.h"
#include "clang/CIR/Dialect/Transforms/CIRTransformUtils.h"
#include "clang/CIR/MissingFeatures.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/SetVector.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/TargetParser/Triple.h"
 
using namespace mlir;
using namespace cir;
 
namespace mlir {
#define GEN_PASS_DEF_CIREHABILOWERING
#include "clang/CIR/Dialect/Passes.h.inc"
} // namespace mlir
 
namespace {
 
//===----------------------------------------------------------------------===//
// Shared utilities
//===----------------------------------------------------------------------===//
 
/// Ensure a function with the given name and type exists in the module. If it
/// does not exist, create a private external declaration.
static cir::FuncOp getOrCreateRuntimeFuncDecl(mlir::ModuleOp mod,
                                              mlir::Location loc,
                                              StringRef name,
                                              cir::FuncType funcTy) {
  if (auto existing = mod.lookupSymbol<cir::FuncOp>(name))
    return existing;
 
  mlir::OpBuilder builder(mod.getContext());
  builder.setInsertionPointToEnd(mod.getBody());
  auto funcOp = cir::FuncOp::create(builder, loc, name, funcTy);
  funcOp.setLinkage(cir::GlobalLinkageKind::ExternalLinkage);
  funcOp.setPrivate();
  // This pass runs after CallConvLowering and builds the declaration in its
  // target ABI form already.
  funcOp.setRequiresCallConvLowering(false);
  return funcOp;
}
 
//===----------------------------------------------------------------------===//
// EH ABI Lowering Base Class
//===----------------------------------------------------------------------===//
 
/// Abstract base class for exception-handling ABI lowering.
/// Each supported ABI (Itanium, Microsoft, etc.) provides a concrete subclass.
class EHABILowering {
public:
  explicit EHABILowering(mlir::ModuleOp mod)
      : mod(mod), ctx(mod.getContext()), builder(ctx) {}
  virtual ~EHABILowering() = default;
 
  /// Lower all EH operations in the module to an ABI-specific form.
  virtual mlir::LogicalResult run() = 0;
 
protected:
  mlir::ModuleOp mod;
  mlir::MLIRContext *ctx;
  mlir::OpBuilder builder;
};
 
//===----------------------------------------------------------------------===//
// Itanium EH ABI Lowering
//===----------------------------------------------------------------------===//
 
/// Lowers flattened CIR EH operations to the Itanium C++ ABI form.
///
/// The entry point is run(), which iterates over all functions and
/// calls lowerFunc() for each. lowerFunc() drives all lowering from
/// cir.eh.initiate operations: every other EH op (begin/end_cleanup,
/// eh.dispatch, begin/end_catch, resume) is reachable by tracing the
/// eh_token produced by the initiate through its users.
class ItaniumEHLowering : public EHABILowering {
public:
  using EHABILowering::EHABILowering;
  mlir::LogicalResult run() override;
 
private:
  /// Maps a !cir.eh_token value to its Itanium ABI replacement pair:
  /// an exception pointer (!cir.ptr<!void>) and a type id (!u32i).
  using EhTokenMap = DenseMap<mlir::Value, std::pair<mlir::Value, mlir::Value>>;
 
  cir::VoidType voidType;
  cir::PointerType voidPtrType;
  cir::PointerType u8PtrType;
  cir::IntType u32Type;
  cir::IntType s32Type;
 
  // Cached runtime function declarations, initialized when needed by
  // ensureRuntimeDecls().
  cir::FuncOp personalityFunc;
  cir::FuncOp beginCatchFunc;
  cir::FuncOp endCatchFunc;
  cir::FuncOp getExceptionPtrFunc;
  cir::FuncOp clangCallTerminateFunc;
  cir::FuncOp cxaThrowFunc;
  cir::FuncOp cxaRethrowFunc;
  cir::FuncOp cxaCallUnexpectedFunc;
 
  DenseMap<mlir::StringAttr, cir::FuncOp> catchCopyThunks;
 
  llvm::SmallSetVector<mlir::StringAttr, 1> droppedPersonalities;
 
  constexpr const static ::llvm::StringLiteral kGxxPersonality =
      "__gxx_personality_v0";
 
  void ensureRuntimeDecls(mlir::Location loc);
  void ensureClangCallTerminate(mlir::Location loc);
  void ensureCxaThrowDecl(mlir::Location loc);
  void ensureCxaRethrowDecl(mlir::Location loc);
  void ensureCxaCallUnexpectedDecl(mlir::Location loc);
  mlir::Block *buildTerminateBlock(cir::FuncOp funcOp, mlir::Location loc);
  mlir::FailureOr<cir::FuncOp>
  resolveCatchCopyThunk(cir::ConstructCatchParamOp op);
  mlir::LogicalResult lowerFunc(cir::FuncOp funcOp);
  mlir::LogicalResult
  lowerEhInitiate(cir::EhInitiateOp initiateOp,
                  llvm::ArrayRef<cir::EhDispatchOp> reachedDispatches,
                  bool reachesCleanup, EhTokenMap &ehTokenMap);
  void lowerDispatch(cir::EhDispatchOp dispatch, mlir::Value exnPtr,
                     mlir::Value typeId);
  mlir::LogicalResult lowerConstructCatchParam(cir::ConstructCatchParamOp op,
                                               mlir::Value exnPtr);
  void lowerInitCatchParam(cir::InitCatchParamOp op);
  mlir::LogicalResult lowerTryThrow(cir::TryThrowOp op);
};
 
/// Lower all EH operations in the module to the Itanium-specific form.
mlir::LogicalResult ItaniumEHLowering::run() {
  // Pre-compute the common types used throughout all function lowerings.
  // TODO(cir): Move these to the base class if they are also needed for MSVC.
  voidType = cir::VoidType::get(ctx);
  voidPtrType = cir::PointerType::get(voidType);
  auto u8Type = cir::IntType::get(ctx, 8, /*isSigned=*/false);
  u8PtrType = cir::PointerType::get(u8Type);
  u32Type = cir::IntType::get(ctx, 32, /*isSigned=*/false);
  s32Type = cir::IntType::get(ctx, 32, /*isSigned=*/true);
 
  for (cir::FuncOp funcOp : mod.getOps<cir::FuncOp>()) {
    if (mlir::failed(lowerFunc(funcOp)))
      return mlir::failure();
  }
 
  for (mlir::StringAttr name : droppedPersonalities) {
    auto personalityDecl = mod.lookupSymbol<cir::FuncOp>(name);
    if (personalityDecl && personalityDecl.isDeclaration() &&
        mlir::SymbolTable::symbolKnownUseEmpty(personalityDecl, mod))
      personalityDecl.erase();
  }
  return mlir::success();
}
 
/// Ensure the necessary Itanium runtime function declarations exist in the
/// module.
void ItaniumEHLowering::ensureRuntimeDecls(mlir::Location loc) {
  // TODO(cir): Handle other personality functions. This probably isn't needed
  // here if we fix codegen to always set the personality function.
  if (!personalityFunc) {
    auto personalityFuncTy = cir::FuncType::get({}, s32Type, /*isVarArg=*/true);
    personalityFunc = getOrCreateRuntimeFuncDecl(mod, loc, kGxxPersonality,
                                                 personalityFuncTy);
  }
 
  if (!beginCatchFunc) {
    auto beginCatchFuncTy =
        cir::FuncType::get({voidPtrType}, u8PtrType, /*isVarArg=*/false);
    beginCatchFunc = getOrCreateRuntimeFuncDecl(mod, loc, "__cxa_begin_catch",
                                                beginCatchFuncTy);
  }
 
  if (!endCatchFunc) {
    auto endCatchFuncTy = cir::FuncType::get({}, voidType, /*isVarArg=*/false);
    endCatchFunc =
        getOrCreateRuntimeFuncDecl(mod, loc, "__cxa_end_catch", endCatchFuncTy);
  }
 
  if (!getExceptionPtrFunc) {
    auto getExceptionPtrFuncTy =
        cir::FuncType::get({voidPtrType}, u8PtrType, /*isVarArg=*/false);
    getExceptionPtrFunc = getOrCreateRuntimeFuncDecl(
        mod, loc, "__cxa_get_exception_ptr", getExceptionPtrFuncTy);
  }
}
 
/// Ensure the __clang_call_terminate function exists in the module. This
/// function is defined with a body that calls __cxa_begin_catch followed by
/// std::terminate, matching the behavior of Clang's LLVM IR codegen.
///
///   void __clang_call_terminate(void *exn) nounwind noreturn {
///     __cxa_begin_catch(exn);
///     std::terminate();
///     unreachable;
///   }
void ItaniumEHLowering::ensureClangCallTerminate(mlir::Location loc) {
  // TODO(cir): set the runtime calling convention on the runtime calls below.
  assert(!cir::MissingFeatures::opFuncCallingConv());
  if (clangCallTerminateFunc)
    return;
 
  ensureRuntimeDecls(loc);
 
  if (auto existing = mod.lookupSymbol<cir::FuncOp>("__clang_call_terminate")) {
    clangCallTerminateFunc = existing;
    return;
  }
 
  auto funcTy = cir::FuncType::get({voidPtrType}, voidType, /*isVarArg=*/false);
  builder.setInsertionPointToEnd(mod.getBody());
  auto funcOp =
      cir::FuncOp::create(builder, loc, "__clang_call_terminate", funcTy);
  funcOp.setLinkage(cir::GlobalLinkageKind::LinkOnceODRLinkage);
  funcOp.setGlobalVisibility(cir::VisibilityKind::Hidden);
  // This pass runs after CallConvLowering and builds the function in its
  // target ABI form already.
  funcOp.setRequiresCallConvLowering(false);
 
  mlir::Block *entryBlock = funcOp.addEntryBlock();
  builder.setInsertionPointToStart(entryBlock);
  mlir::Value exnArg = entryBlock->getArgument(0);
 
  auto catchCall = cir::CallOp::create(
      builder, loc, mlir::FlatSymbolRefAttr::get(beginCatchFunc), u8PtrType,
      mlir::ValueRange{exnArg});
  catchCall.setNothrowAttr(builder.getUnitAttr());
 
  auto terminateFuncDecl = getOrCreateRuntimeFuncDecl(
      mod, loc, "_ZSt9terminatev",
      cir::FuncType::get({}, voidType, /*isVarArg=*/false));
  terminateFuncDecl->setAttr(cir::CIRDialect::getNoReturnAttrName(),
                             builder.getUnitAttr());
  auto terminateCall = cir::CallOp::create(
      builder, loc, mlir::FlatSymbolRefAttr::get(terminateFuncDecl), voidType,
      mlir::ValueRange{});
  terminateCall.setNothrowAttr(builder.getUnitAttr());
  terminateCall->setAttr(cir::CIRDialect::getNoReturnAttrName(),
                         builder.getUnitAttr());
 
  cir::UnreachableOp::create(builder, loc);
 
  funcOp->setAttr(cir::CIRDialect::getNoReturnAttrName(),
                  builder.getUnitAttr());
  clangCallTerminateFunc = funcOp;
}
 
/// Ensure the __cxa_throw runtime function is declared in the module.
///
///   void __cxa_throw(void *exception, void *type_info, void *dtor);
void ItaniumEHLowering::ensureCxaThrowDecl(mlir::Location loc) {
  if (cxaThrowFunc)
    return;
  auto throwFuncTy = cir::FuncType::get({voidPtrType, voidPtrType, voidPtrType},
                                        voidType, /*isVarArg=*/false);
  cxaThrowFunc =
      getOrCreateRuntimeFuncDecl(mod, loc, "__cxa_throw", throwFuncTy);
}
 
/// Ensure the __cxa_rethrow runtime function is declared in the module.
///
///   void __cxa_rethrow();
void ItaniumEHLowering::ensureCxaRethrowDecl(mlir::Location loc) {
  if (cxaRethrowFunc)
    return;
  auto rethrowFuncTy = cir::FuncType::get({}, voidType, /*isVarArg=*/false);
  cxaRethrowFunc =
      getOrCreateRuntimeFuncDecl(mod, loc, "__cxa_rethrow", rethrowFuncTy);
}
 
///   void __cxa_call_unexpected(void *exn);
void ItaniumEHLowering::ensureCxaCallUnexpectedDecl(mlir::Location loc) {
  if (cxaCallUnexpectedFunc)
    return;
  auto unexpectedFuncTy =
      cir::FuncType::get({voidPtrType}, voidType, /*isVarArg=*/false);
  cxaCallUnexpectedFunc = getOrCreateRuntimeFuncDecl(
      mod, loc, "__cxa_call_unexpected", unexpectedFuncTy);
}
 
/// Create a terminate landing pad block at the end of the specified function.
mlir::Block *ItaniumEHLowering::buildTerminateBlock(cir::FuncOp funcOp,
                                                    mlir::Location loc) {
  // TODO(cir): set the runtime calling convention on the runtime calls below.
  assert(!cir::MissingFeatures::opFuncCallingConv());
  assert(clangCallTerminateFunc &&
         "ensureClangCallTerminate must run before buildTerminateBlock");
  mlir::Region &body = funcOp.getRegion();
  mlir::Block *terminateBlock = builder.createBlock(&body, body.end());
  auto inflight = cir::EhInflightOp::create(
      builder, loc, /*cleanup=*/false, /*catch_all=*/true,
      /*catch_type_list=*/mlir::ArrayAttr{},
      /*filter_type_list=*/mlir::ArrayAttr{});
  auto terminateCall = cir::CallOp::create(
      builder, loc, mlir::FlatSymbolRefAttr::get(clangCallTerminateFunc),
      voidType, mlir::ValueRange{inflight.getExceptionPtr()});
  terminateCall.setNothrowAttr(builder.getUnitAttr());
  terminateCall->setAttr(cir::CIRDialect::getNoReturnAttrName(),
                         builder.getUnitAttr());
  cir::UnreachableOp::create(builder, loc);
  return terminateBlock;
}
 
/// Read-only walk of the eh_token graph from an initiate's root token,
/// collecting every cir.eh.dispatch its exception can reach, innermost first.
/// The token flows through cleanups to the innermost dispatch and then, via
/// that dispatch's continue-unwind edge, on to each enclosing dispatch (nested
/// try/catch).  The walk follows the token into catch-handler blocks too, but
/// it dead-ends there because cir.begin_catch consumes the eh_token (producing
/// a catch_token), so only the unwind chain yields further dispatches.
///
/// This is computed before any destructive lowering so a landing pad's catch
/// types -- which are a property of the EH graph -- do not depend on the order
/// in which the destructive per-initiate traversal tears down shared
/// token-graph edges.  \p dispatches is left empty for a cleanup-only initiate
/// that reaches no dispatch (e.g. a path that only resumes).
static void collectReachableDispatches(
    mlir::Value rootToken,
    llvm::SmallSetVector<cir::EhDispatchOp, 4> &dispatches,
    bool &reachesCleanup) {
  llvm::SmallVector<mlir::Value> worklist;
  llvm::SmallPtrSet<mlir::Value, 8> visited;
  worklist.push_back(rootToken);
  // Breadth-first (process in insertion order) so dispatches are discovered
  // innermost first, matching the order catch clauses must appear in the
  // landing pad.
  for (unsigned i = 0; i < worklist.size(); ++i) {
    mlir::Value current = worklist[i];
    if (!visited.insert(current).second)
      continue;
    for (mlir::OpOperand &use : current.getUses()) {
      mlir::Operation *user = use.getOwner();
      // A cleanup anywhere on the unwind path (this initiate's own cleanup or
      // an enclosing scope's) means the landing pad must carry the cleanup
      // clause so destructors still run when a foreign exception unwinds
      // through this frame.
      if (mlir::isa<cir::BeginCleanupOp>(user))
        reachesCleanup = true;
      if (auto dispatch = mlir::dyn_cast<cir::EhDispatchOp>(user))
        dispatches.insert(dispatch);
      // Follow the token into eh_token block arguments of successor blocks.
      for (unsigned s = 0, e = user->getNumSuccessors(); s < e; ++s)
        for (mlir::BlockArgument arg : user->getSuccessor(s)->getArguments())
          if (mlir::isa<cir::EhTokenType>(arg.getType()))
            worklist.push_back(arg);
    }
  }
}
 
/// Lower all EH operations in a single function.
mlir::LogicalResult ItaniumEHLowering::lowerFunc(cir::FuncOp funcOp) {
  if (funcOp.isDeclaration())
    return mlir::success();
 
  // All EH lowering follows from cir.eh.initiate operations. The token each
  // initiate produces connects it to every other EH op in the function
  // (begin/end_cleanup, eh.dispatch, begin/end_catch, resume) through the
  // token graph. A single walk to collect initiates is therefore sufficient.
  SmallVector<cir::EhInitiateOp> initiateOps;
  bool hasLandingPad = false;
  funcOp.walk([&](mlir::Operation *op) {
    if (auto initiateOp = mlir::dyn_cast<cir::EhInitiateOp>(op))
      initiateOps.push_back(initiateOp);
    else if (mlir::isa<cir::EhInflightOp>(op))
      hasLandingPad = true;
  });
  if (initiateOps.empty()) {
    // Without a landing pad the personality is dead, as in classic codegen.
    mlir::FlatSymbolRefAttr personality = funcOp.getPersonalityAttr();
    if (personality && !hasLandingPad) {
      droppedPersonalities.insert(personality.getAttr());
      funcOp.removePersonalityAttr();
    }
    return mlir::success();
  }
 
  ensureRuntimeDecls(funcOp.getLoc());
 
  // Set the personality function if it is not already set.
  // TODO(cir): The personality function should already have been set by this
  // point. If we've seen a try operation, it will have been set by
  // emitCXXTryStmt. If we only have cleanups, it may not have been set. We
  // need to fix that in CodeGen. This is a placeholder until that is done.
  if (!funcOp.getPersonality())
    funcOp.setPersonality(kGxxPersonality);
 
  // Compute, read-only and before any destructive lowering, the dispatches each
  // initiate's exception can reach (innermost first; more than one for nested
  // try/catch).  A landing pad's catch types are a property of the EH graph, so
  // deriving them here keeps them independent of the order in which the
  // destructive per-initiate traversal in lowerEhInitiate tears down shared
  // token-graph edges.  Otherwise a sibling or outer dispatch could be missed,
  // leaving a landing pad without its catch clause (it would resume past the
  // handler to std::terminate) or an un-lowered leftover dispatch.
  // Per initiate: the dispatches its exception can reach (innermost first) and
  // whether a cleanup lies on its unwind path.  Both are properties of the EH
  // graph and are always consumed together for the same initiate, so they live
  // in one map keyed by the initiate op.
  struct InitiateEHInfo {
    llvm::SmallSetVector<cir::EhDispatchOp, 4> reachedDispatches;
    bool reachesCleanup = false;
  };
  llvm::DenseMap<mlir::Operation *, InitiateEHInfo> initiateInfo;
  llvm::SmallSetVector<cir::EhDispatchOp, 4> dispatchesToLower;
  for (cir::EhInitiateOp initiateOp : initiateOps) {
    InitiateEHInfo &info = initiateInfo[initiateOp.getOperation()];
    collectReachableDispatches(initiateOp.getEhToken(), info.reachedDispatches,
                               info.reachesCleanup);
    dispatchesToLower.insert(info.reachedDispatches.begin(),
                             info.reachedDispatches.end());
  }
 
  EhTokenMap ehTokenMap;
  for (cir::EhInitiateOp initiateOp : initiateOps) {
    const InitiateEHInfo &info = initiateInfo[initiateOp.getOperation()];
    if (mlir::failed(lowerEhInitiate(initiateOp,
                                     info.reachedDispatches.getArrayRef(),
                                     info.reachesCleanup, ehTokenMap)))
      return mlir::failure();
  }
 
  // Lower each dispatch exactly once.  Every initiate's token block-argument
  // (ptr, u32) replacements are registered in ehTokenMap by now, so each
  // dispatch's own (exnPtr, typeId) pair is available.  lowerDispatch erases
  // the dispatch after building its comparison chain.
  for (cir::EhDispatchOp dispatch : dispatchesToLower) {
    auto [exnPtr, typeId] = ehTokenMap.lookup(dispatch.getEhToken());
    assert(exnPtr && typeId &&
           "dispatch eh_token must be registered in ehTokenMap");
    lowerDispatch(dispatch, exnPtr, typeId);
  }
 
  // Remove the !cir.eh_token block arguments that were replaced by (ptr, u32)
  // pairs. Iterate in reverse to preserve argument indices during removal.
  for (mlir::Block &block : funcOp.getBody()) {
    for (int i = block.getNumArguments() - 1; i >= 0; --i) {
      if (mlir::isa<cir::EhTokenType>(block.getArgument(i).getType()))
        block.eraseArgument(i);
    }
  }
 
  // Lower any cir.init_catch_param ops in this function. These materialize
  // the catch parameter local from the (already lowered) begin_catch result,
  // and are independent of the eh_token graph traversal above.
  SmallVector<cir::InitCatchParamOp> initCatchOps;
  funcOp.walk([&](cir::InitCatchParamOp op) { initCatchOps.push_back(op); });
  for (cir::InitCatchParamOp op : initCatchOps)
    lowerInitCatchParam(op);
 
  // Lower any cir.try_throw ops in this function to cir.try_call of
  // __cxa_throw / __cxa_rethrow. These are produced by FlattenCFG when a
  // cir.throw appears inside a cleanup scope or try region.
  SmallVector<cir::TryThrowOp> tryThrowOps;
  funcOp.walk([&](cir::TryThrowOp op) { tryThrowOps.push_back(op); });
  for (cir::TryThrowOp op : tryThrowOps)
    if (mlir::failed(lowerTryThrow(op)))
      return mlir::failure();
 
  return mlir::success();
}
 
/// Lower all EH operations connected to a single cir.eh.initiate.
///
/// The cir.eh.initiate is the root of a token graph. The token it produces
/// flows through branch edges to consuming operations:
///
///   cir.eh.initiate → (via cir.br) → cir.begin_cleanup
///                                     → cir.end_cleanup (via cleanup_token)
///                                   → (via cir.br) → cir.eh.dispatch
///                                                     → (successors) →
///                                                       cir.begin_catch
///                                                       → cir.end_catch
///                                                         (via catch_token)
///                                   → cir.resume
///
/// A single traversal of the token graph discovers and processes every
/// connected op inline. The inflight_exception is created up-front without
/// a catch_type_list; when the dispatch is encountered during traversal,
/// the catch types are read and set on the inflight op.
///
/// Dispatch ops are not lowered here; they are lowered once by the caller after
/// every initiate has been processed (a dispatch can be shared by sibling
/// initiates), so this traversal only registers the catch-handler block
/// arguments reachable through them.
///
/// \p ehTokenMap is shared across all initiates in the function so that block
/// arguments reachable from multiple sibling initiates are registered once.
mlir::LogicalResult ItaniumEHLowering::lowerEhInitiate(
    cir::EhInitiateOp initiateOp,
    llvm::ArrayRef<cir::EhDispatchOp> reachedDispatches, bool reachesCleanup,
    EhTokenMap &ehTokenMap) {
  // TODO(cir): set the runtime calling convention on the runtime calls below.
  assert(!cir::MissingFeatures::opFuncCallingConv());
  mlir::Value rootToken = initiateOp.getEhToken();
 
  // The catch clauses for this landing pad come from the dispatches its
  // exception reaches (computed read-only by the caller before any destructive
  // lowering).  For nested try/catch the exception can reach several dispatches
  // innermost first, so the landing pad lists their catch types in that order
  // (matching classic CodeGen), stopping at a catch-all since nothing escapes
  // it.  Deriving this from the original EH graph -- rather than during the
  // destructive token-graph traversal below -- keeps it correct regardless of
  // the order in which sibling/nested initiates are lowered.
  mlir::ArrayAttr catchTypeList;
  mlir::ArrayAttr filterTypeList;
  bool catchAll = false;
  SmallVector<mlir::Attribute> typeSymbols;
  for (cir::EhDispatchOp dispatch : reachedDispatches) {
    if (mlir::ArrayAttr catchTypes = dispatch.getCatchTypesAttr()) {
      for (mlir::Attribute attr : catchTypes) {
        if (auto filter = mlir::dyn_cast<cir::EhFilterAttr>(attr)) {
          SmallVector<mlir::Attribute> filterSymbols;
          // A filter terminates the landing-pad clause list the same way a
          // catch-all does. Nothing outside the specification is reachable.
          for (mlir::Attribute typeAttr : filter.getPermittedTypes()) {
            auto globalView = mlir::cast<cir::GlobalViewAttr>(typeAttr);
            filterSymbols.push_back(globalView.getSymbol());
          }
          filterTypeList = builder.getArrayAttr(filterSymbols);
          continue;
        }
        typeSymbols.push_back(
            mlir::cast<cir::GlobalViewAttr>(attr).getSymbol());
      }
    }
    if (dispatch.getDefaultIsCatchAll() || filterTypeList) {
      catchAll = dispatch.getDefaultIsCatchAll();
      // A catch-all or filter handles the remaining exceptions, so it
      // stops the unwind. No enclosing dispatch is reachable past it.
      assert(dispatch == reachedDispatches.back() &&
             "catch-all or filter must be the last reachable dispatch");
      break;
    }
  }
  if (!typeSymbols.empty())
    catchTypeList = builder.getArrayAttr(typeSymbols);
 
  builder.setInsertionPoint(initiateOp);
  auto inflightOp = cir::EhInflightOp::create(
      builder, initiateOp.getLoc(), initiateOp.getCleanup() || reachesCleanup,
      catchAll, catchTypeList, filterTypeList);
 
  ehTokenMap[rootToken] = {inflightOp.getExceptionPtr(),
                           inflightOp.getTypeId()};
 
  // Single traversal of the token graph. For each token value (the root token
  // or a block argument that carries it), we snapshot its users, register
  // (ptr, u32) replacement arguments on successor blocks, then process every
  // user inline. This avoids collecting ops into separate vectors.
  SmallVector<mlir::Value> worklist;
  SmallPtrSet<mlir::Value, 8> visited;
  worklist.push_back(rootToken);
 
  while (!worklist.empty()) {
    mlir::Value current = worklist.pop_back_val();
    if (!visited.insert(current).second)
      continue;
 
    // Snapshot users before modifying any of them (erasing ops during
    // iteration would invalidate the use-list iterator).
    SmallVector<mlir::Operation *> users;
    for (mlir::OpOperand &use : current.getUses())
      users.push_back(use.getOwner());
 
    // Register replacement block arguments on successor blocks (extending the
    // worklist), then lower the op itself.
    for (mlir::Operation *user : users) {
      // Trace into successor blocks to register (ptr, u32) replacement
      // arguments for any !cir.eh_token block arguments found there.  Even
      // if a block arg was already registered by a sibling initiate, it is
      // still added to the worklist so that the traversal can reach the
      // shared dispatch to read catch types.
      for (unsigned s = 0; s < user->getNumSuccessors(); ++s) {
        mlir::Block *succ = user->getSuccessor(s);
        for (mlir::BlockArgument arg : succ->getArguments()) {
          if (!mlir::isa<cir::EhTokenType>(arg.getType()))
            continue;
          if (!ehTokenMap.count(arg)) {
            mlir::Value ptrArg = succ->addArgument(voidPtrType, arg.getLoc());
            mlir::Value u32Arg = succ->addArgument(u32Type, arg.getLoc());
            ehTokenMap[arg] = {ptrArg, u32Arg};
          }
          worklist.push_back(arg);
        }
      }
 
      if (auto op = mlir::dyn_cast<cir::BeginCleanupOp>(user)) {
        // begin_cleanup / end_cleanup are no-ops for Itanium.  Erase the
        // end_cleanup first (drops the cleanup_token use) then the begin.
        for (auto &tokenUsers :
             llvm::make_early_inc_range(op.getCleanupToken().getUses())) {
          if (auto endOp =
                  mlir::dyn_cast<cir::EndCleanupOp>(tokenUsers.getOwner()))
            endOp.erase();
        }
        op.erase();
      } else if (auto op = mlir::dyn_cast<cir::BeginCatchOp>(user)) {
        // Replace end_catch → __cxa_end_catch (drops the catch_token use),
        // then replace begin_catch → __cxa_begin_catch.
        for (auto &tokenUsers :
             llvm::make_early_inc_range(op.getCatchToken().getUses())) {
          if (auto endOp =
                  mlir::dyn_cast<cir::EndCatchOp>(tokenUsers.getOwner())) {
            builder.setInsertionPoint(endOp);
            cir::CallOp::create(builder, endOp.getLoc(),
                                mlir::FlatSymbolRefAttr::get(endCatchFunc),
                                voidType, mlir::ValueRange{});
            endOp.erase();
          }
        }
 
        auto [exnPtr, typeId] = ehTokenMap.lookup(op.getEhToken());
        builder.setInsertionPoint(op);
        auto callOp = cir::CallOp::create(
            builder, op.getLoc(), mlir::FlatSymbolRefAttr::get(beginCatchFunc),
            u8PtrType, mlir::ValueRange{exnPtr});
        mlir::Value castResult = callOp.getResult();
        mlir::Type expectedPtrType = op.getExnPtr().getType();
        if (castResult.getType() != expectedPtrType)
          castResult =
              cir::CastOp::create(builder, op.getLoc(), expectedPtrType,
                                  cir::CastKind::bitcast, callOp.getResult());
        op.getExnPtr().replaceAllUsesWith(castResult);
        op.erase();
      } else if (auto op = mlir::dyn_cast<cir::ConstructCatchParamOp>(user)) {
        auto [exnPtr, typeId] = ehTokenMap.lookup(op.getEhToken());
        if (mlir::failed(lowerConstructCatchParam(op, exnPtr)))
          return mlir::failure();
      } else if (mlir::isa<cir::EhDispatchOp>(user)) {
        // The dispatch's catch types were already read into every reaching
        // landing pad (read-only, before this traversal).  The dispatch op
        // itself is lowered once by the caller after all initiates are
        // processed, so nothing is done here; the successor catch-handler
        // blocks are still reached via the block-argument registration above.
      } else if (auto op = mlir::dyn_cast<cir::EhTerminateOp>(user)) {
        auto [exnPtr, typeId] = ehTokenMap.lookup(op.getEhToken());
        ensureClangCallTerminate(op.getLoc());
        builder.setInsertionPoint(op);
        auto call = cir::CallOp::create(
            builder, op.getLoc(),
            mlir::FlatSymbolRefAttr::get(clangCallTerminateFunc), voidType,
            mlir::ValueRange{exnPtr});
        call.setNothrowAttr(builder.getUnitAttr());
        call->setAttr(cir::CIRDialect::getNoReturnAttrName(),
                      builder.getUnitAttr());
        cir::UnreachableOp::create(builder, op.getLoc());
        op.erase();
      } else if (auto op = mlir::dyn_cast<cir::EhUnexpectedOp>(user)) {
        auto [exnPtr, typeId] = ehTokenMap.lookup(op.getEhToken());
        ensureCxaCallUnexpectedDecl(op.getLoc());
        builder.setInsertionPoint(op);
        auto call = cir::CallOp::create(
            builder, op.getLoc(),
            mlir::FlatSymbolRefAttr::get(cxaCallUnexpectedFunc), voidType,
            mlir::ValueRange{exnPtr});
        call->setAttr(cir::CIRDialect::getNoReturnAttrName(),
                      builder.getUnitAttr());
        cir::UnreachableOp::create(builder, op.getLoc());
        op.erase();
      } else if (auto op = mlir::dyn_cast<cir::ResumeOp>(user)) {
        auto [exnPtr, typeId] = ehTokenMap.lookup(op.getEhToken());
        builder.setInsertionPoint(op);
        cir::ResumeFlatOp::create(builder, op.getLoc(), exnPtr, typeId);
        op.erase();
      } else if (auto op = mlir::dyn_cast<cir::BrOp>(user)) {
        // Replace eh_token operands with the (ptr, u32) pair.
        SmallVector<mlir::Value> newOperands;
        bool changed = false;
        for (mlir::Value operand : op.getDestOperands()) {
          auto it = ehTokenMap.find(operand);
          if (it != ehTokenMap.end()) {
            newOperands.push_back(it->second.first);
            newOperands.push_back(it->second.second);
            changed = true;
          } else {
            newOperands.push_back(operand);
          }
        }
        if (changed) {
          builder.setInsertionPoint(op);
          // We shouldn't get any loop information here.
          assert(!op.getLoopIdAttr() && !op.getLoopHintsAttr());
          cir::BrOp::create(builder, op.getLoc(), newOperands,
                            /*loop_id=*/nullptr, /*loop_hints=*/nullptr,
                            op.getDest());
          op.erase();
        }
      }
    }
  }
 
  initiateOp.erase();
  return mlir::success();
}
 
/// Lower a cir.eh.dispatch by creating a comparison chain in new blocks.
/// The dispatch itself is replaced with a branch to the first comparison
/// block and then erased.
void ItaniumEHLowering::lowerDispatch(cir::EhDispatchOp dispatch,
                                      mlir::Value exnPtr, mlir::Value typeId) {
  mlir::Location dispLoc = dispatch.getLoc();
  mlir::Block *defaultDest = dispatch.getDefaultDestination();
  mlir::Block *dispatchBlock = dispatch->getBlock();
 
  llvm::SmallVector<mlir::Attribute> catchAttrs;
  llvm::SmallVector<mlir::Block *> catchDests;
  cir::EhFilterAttr filterAttr;
  mlir::Block *filterDest = nullptr;
  if (mlir::ArrayAttr catchTypes = dispatch.getCatchTypesAttr()) {
    for (auto [attr, dest] :
         llvm::zip(catchTypes, dispatch.getCatchDestinations())) {
      if (auto filter = mlir::dyn_cast<cir::EhFilterAttr>(attr)) {
        assert(!filterAttr && "at most one filter handler");
        filterAttr = filter;
        filterDest = dest;
        continue;
      }
      catchAttrs.push_back(attr);
      catchDests.push_back(dest);
    }
  }
 
  // Build the comparison chain in new blocks inserted after the dispatch's
  // block. The dispatch itself is replaced with a branch to the first
  // comparison block and erased below.
  mlir::Block *insertBefore = dispatchBlock->getNextNode();
  mlir::Block *falseDest = defaultDest;
  if (filterAttr) {
    assert(filterDest && "filter handler requires a destination");
    // An empty permitted-type list means that every exception violates the
    // specification, so the filter destination is taken unconditionally.
    if (filterAttr.getPermittedTypes().empty()) {
      falseDest = filterDest;
    } else {
      // The personality reports a filter failure with a negative selector.
      // Type ids are still !u32i on the eh_token replacement pair, so recast
      // to !s32i for the signed comparison. TODO: produce !s32i throughout.
      auto *cmpBlock = builder.createBlock(insertBefore, {voidPtrType, u32Type},
                                           {dispLoc, dispLoc});
      mlir::Value cmpExnPtr = cmpBlock->getArgument(0);
      mlir::Value cmpTypeId = cmpBlock->getArgument(1);
      mlir::Value signedTypeId = cir::CastOp::create(
          builder, dispLoc, s32Type, cir::CastKind::integral, cmpTypeId);
      mlir::Value zero = cir::ConstantOp::create(builder, dispLoc,
                                                 cir::IntAttr::get(s32Type, 0));
      auto cmpOp = cir::CmpOp::create(builder, dispLoc, cir::CmpOpKind::lt,
                                      signedTypeId, zero);
      cir::BrCondOp::create(builder, dispLoc, cmpOp, filterDest, defaultDest,
                            mlir::ValueRange{cmpExnPtr, cmpTypeId},
                            mlir::ValueRange{cmpExnPtr, cmpTypeId});
      insertBefore = cmpBlock;
      falseDest = cmpBlock;
    }
  }
 
  mlir::Block *firstCmpBlock = nullptr;
  if (!catchAttrs.empty()) {
    unsigned numCatches = catchAttrs.size();
    for (int i = numCatches - 1; i >= 0; --i) {
      auto *cmpBlock = builder.createBlock(insertBefore, {voidPtrType, u32Type},
                                           {dispLoc, dispLoc});
 
      mlir::Value cmpExnPtr = cmpBlock->getArgument(0);
      mlir::Value cmpTypeId = cmpBlock->getArgument(1);
 
      auto globalView = mlir::cast<cir::GlobalViewAttr>(catchAttrs[i]);
      auto ehTypeIdOp =
          cir::EhTypeIdOp::create(builder, dispLoc, globalView.getSymbol());
      auto cmpOp = cir::CmpOp::create(builder, dispLoc, cir::CmpOpKind::eq,
                                      cmpTypeId, ehTypeIdOp.getTypeId());
 
      cir::BrCondOp::create(builder, dispLoc, cmpOp, catchDests[i], falseDest,
                            mlir::ValueRange{cmpExnPtr, cmpTypeId},
                            mlir::ValueRange{cmpExnPtr, cmpTypeId});
 
      insertBefore = cmpBlock;
      falseDest = cmpBlock;
      firstCmpBlock = cmpBlock;
    }
  } else {
    firstCmpBlock = falseDest;
  }
 
  builder.setInsertionPoint(dispatch);
  cir::BrOp::create(builder, dispLoc, firstCmpBlock,
                    mlir::ValueRange{exnPtr, typeId});
 
  // The caller lowers each dispatch exactly once after every initiate has been
  // processed, so no sibling still needs it; erase it now.
  dispatch.erase();
}
 
mlir::FailureOr<cir::FuncOp>
ItaniumEHLowering::resolveCatchCopyThunk(cir::ConstructCatchParamOp op) {
  mlir::FlatSymbolRefAttr thunkRef = op.getCopyFnAttr();
  mlir::StringAttr thunkName = thunkRef.getAttr();
  auto cached = catchCopyThunks.find(thunkName);
  if (cached != catchCopyThunks.end())
    return cached->second;
 
  cir::FuncOp thunk = mod.lookupSymbol<cir::FuncOp>(thunkRef);
  if (!thunk)
    return op.emitError("could not resolve catch-copy thunk symbol");
  assert(thunk->hasAttr(cir::CIRDialect::getCatchCopyThunkAttrName()) &&
         "verifier should have rejected non-thunk catch-copy reference");
  if (thunk.isDeclaration())
    return op.emitError("catch-copy thunk has no body to inline");
 
  mlir::Region &thunkRegion = thunk.getRegion();
  if (!llvm::hasSingleElement(thunkRegion))
    return op.emitError("multi-block catch-copy thunks are NYI");
 
  mlir::Block &thunkEntry = thunkRegion.front();
  assert(thunkEntry.getNumArguments() == 2 &&
         "catch-copy thunk must have exactly two parameters");
  if (!mlir::isa<cir::ReturnOp>(thunkEntry.getTerminator()))
    return op.emitError("catch-copy thunk must end in cir.return");
 
  catchCopyThunks[thunkName] = thunk;
  return thunk;
}
 
/// Lower a cir.construct_catch_param into the Itanium-specific sequence
/// that runs before `__cxa_begin_catch` to bind the catch parameter to the
/// in-flight exception.
mlir::LogicalResult
ItaniumEHLowering::lowerConstructCatchParam(cir::ConstructCatchParamOp op,
                                            mlir::Value exnPtr) {
  // TODO(cir): set the runtime calling convention on the runtime calls below.
  assert(!cir::MissingFeatures::opFuncCallingConv());
  mlir::Location loc = op.getLoc();
  mlir::Value paramAddr = op.getParamAddr();
  cir::PointerType paramAddrType =
      mlir::cast<cir::PointerType>(paramAddr.getType());
 
  if (op.getKind() == cir::InitCatchKind::ReferenceToPointer) {
    assert(!MissingFeatures::sizeOfUnwindException());
    constexpr unsigned headerSize = 32;
 
    builder.setInsertionPoint(op);
    auto index = cir::ConstantOp::create(
        builder, loc, cir::IntAttr::get(u32Type, headerSize));
    assert((exnPtr.getType() == voidPtrType || exnPtr.getType() == u8PtrType) &&
           "lowerConstructCatchParam exn ptr not void* or i8*");
    auto exnObj =
        cir::PtrStrideOp::create(builder, loc, exnPtr.getType(), exnPtr, index);
    mlir::Value casted =
        cir::CastOp::create(builder, loc, paramAddrType.getPointee(),
                            cir::CastKind::bitcast, exnObj);
    cir::StoreOp::create(builder, loc, casted, paramAddr, {}, {}, {}, {}, {});
    op.erase();
    return success();
  }
 
  if (op.getKind() != cir::InitCatchKind::NonTrivialCopy)
    return op.emitOpError("only reference_to_pointer and non_trivial_copy "
                          "kinds are supported");
 
  ensureRuntimeDecls(loc);
  ensureClangCallTerminate(loc);
 
  // Call __cxa_get_exception_ptr to get the in-flight exception.
  builder.setInsertionPoint(op);
  cir::CallOp getExnCall = cir::CallOp::create(
      builder, loc, mlir::FlatSymbolRefAttr::get(getExceptionPtrFunc),
      u8PtrType, mlir::ValueRange{exnPtr});
  getExnCall.setNothrowAttr(builder.getUnitAttr());
  mlir::Value adjusted =
      cir::CastOp::create(builder, loc, paramAddrType, cir::CastKind::bitcast,
                          getExnCall.getResult());
 
  // Get the thunk function definition.
  mlir::FailureOr<cir::FuncOp> thunkOr = resolveCatchCopyThunk(op);
  if (mlir::failed(thunkOr))
    return mlir::failure();
  cir::FuncOp thunk = *thunkOr;
 
  // This is also verified by resolveCatchCopyThunk, but the loop below is
  // where the constraint is required so let's assert it again here.
  assert(llvm::hasSingleElement(thunk.getRegion()) &&
         "multi-block catch-copy thunks are NYI");
 
  // Clone the thunk function to perform the copy.
  mlir::Block &thunkEntry = thunk.getRegion().front();
  mlir::IRMapping mapping;
  mapping.map(thunkEntry.getArgument(0), paramAddr);
  mapping.map(thunkEntry.getArgument(1), adjusted);
  llvm::SmallVector<cir::CallOp> throwingCalls;
  for (mlir::Operation &thunkOp : thunkEntry.without_terminator()) {
    mlir::Operation *cloned = builder.clone(thunkOp, mapping);
    if (cir::CallOp callOp = mlir::dyn_cast<cir::CallOp>(cloned))
      if (!callOp.getNothrow())
        throwingCalls.push_back(callOp);
  }
  op.erase();
 
  if (throwingCalls.empty())
    return mlir::success();
 
  // All calls in the copy (which is usually just a single call) need to
  // unwind to a terminate block if it throws an exception.
  mlir::IRRewriter rewriter(builder);
  mlir::Block *terminateBlock = nullptr;
  for (cir::CallOp call : throwingCalls) {
    if (!terminateBlock)
      terminateBlock = buildTerminateBlock(call->getParentOfType<cir::FuncOp>(),
                                           call.getLoc());
    cir::replaceCallWithTryCall(call, terminateBlock, call.getLoc(), rewriter);
  }
  return mlir::success();
}
 
/// Lower a cir.try_throw to a cir.try_call of __cxa_throw (or
/// __cxa_rethrow for the no-operand rethrow form). Materializes the
/// type_info and dtor pointers from their symbol attributes, bitcasting
/// each to !cir.ptr<!void> as required by the runtime function signature.
mlir::LogicalResult ItaniumEHLowering::lowerTryThrow(cir::TryThrowOp op) {
  // TODO(cir): set the runtime calling convention on the runtime calls below.
  assert(!cir::MissingFeatures::opFuncCallingConv());
  mlir::Location loc = op.getLoc();
  mlir::Block *normalDest = op.getNormalDest();
  mlir::Block *unwindDest = op.getUnwindDest();
  builder.setInsertionPoint(op);
 
  if (op.rethrows()) {
    ensureCxaRethrowDecl(loc);
    cir::TryCallOp::create(
        builder, loc, mlir::FlatSymbolRefAttr::get(cxaRethrowFunc), voidType,
        normalDest, unwindDest, mlir::ValueRange{});
    op.erase();
    return mlir::success();
  }
 
  ensureCxaThrowDecl(loc);
 
  // Bitcast the exception pointer to void* if necessary.
  mlir::Value exnPtr = op.getExceptionPtr();
  if (exnPtr.getType() != voidPtrType)
    exnPtr = cir::CastOp::create(builder, loc, voidPtrType,
                                 cir::CastKind::bitcast, exnPtr);
 
  // Materialize the type_info pointer, looking up the typed symbol in the
  // module so we get the correct pointer type for cir.get_global, then
  // bitcasting to void* to match the runtime signature.
  mlir::FlatSymbolRefAttr typeInfoAttr = op.getTypeInfoAttr();
  auto typeInfoGlobal = mod.lookupSymbol<cir::GlobalOp>(typeInfoAttr);
  if (!typeInfoGlobal)
    return op.emitError("type_info symbol not found in module");
  auto typeInfoPtrTy = cir::PointerType::get(typeInfoGlobal.getSymType());
  mlir::Value typeInfo = cir::GetGlobalOp::create(builder, loc, typeInfoPtrTy,
                                                  typeInfoAttr.getValue());
  if (typeInfo.getType() != voidPtrType)
    typeInfo = cir::CastOp::create(builder, loc, voidPtrType,
                                   cir::CastKind::bitcast, typeInfo);
 
  // Materialize the dtor pointer (or null if no dtor).
  mlir::Value dtor;
  if (mlir::FlatSymbolRefAttr dtorAttr = op.getDtorAttr()) {
    auto dtorFunc = mod.lookupSymbol<cir::FuncOp>(dtorAttr);
    if (!dtorFunc)
      return op.emitError("dtor symbol not found in module");
    auto dtorPtrTy = cir::PointerType::get(dtorFunc.getFunctionType());
    dtor =
        cir::GetGlobalOp::create(builder, loc, dtorPtrTy, dtorAttr.getValue());
    if (dtor.getType() != voidPtrType)
      dtor = cir::CastOp::create(builder, loc, voidPtrType,
                                 cir::CastKind::bitcast, dtor);
  } else {
    dtor = cir::ConstantOp::create(
        builder, loc,
        cir::ConstPtrAttr::get(voidPtrType, builder.getI64IntegerAttr(0)));
  }
 
  cir::TryCallOp::create(
      builder, loc, mlir::FlatSymbolRefAttr::get(cxaThrowFunc), voidType,
      normalDest, unwindDest, mlir::ValueRange{exnPtr, typeInfo, dtor});
  op.erase();
  return mlir::success();
}
 
/// Lower a cir.init_catch_param into the Itanium-specific sequence that
/// materializes the catch parameter's local variable from the exception
/// pointer returned by __cxa_begin_catch. The shape of the lowering
/// depends on the init catch kind:
///
///   - Reference: the begin_catch result is the adjusted exception object
///     pointer, so just bitcast and store it into the alloca.
///   - ReferenceToPointer: the companion `cir.construct_catch_param`
///     already bound the reference to the exception object, so this lowering
///     is a no-op.
///   - ReferenceToRecordPointer: the begin_catch result is the adjusted
///     pointer value itself, so store it into a temporary and bind the
///     reference to the temporary.
///   - Pointer: the begin_catch result is
///     the pointer value itself, so just bitcast and store it into the
///     alloca.
///   - Scalar (any other by-value catch): treat the begin_catch result as a
///     pointer to the value, load it, and store it into the alloca.
///   - Objc: Handle pointer representation with ObjCLifetime.
///   - TrivialCopy: copy the exception
///     object's bytes into the alloca via cir.copy.
///   - NonTrivialCopy: the construction was already performed by the
///     companion `cir.construct_catch_param` before `cir.begin_catch`, so
///     this lowering is a no-op.
///
void ItaniumEHLowering::lowerInitCatchParam(cir::InitCatchParamOp op) {
  builder.setInsertionPoint(op);
  mlir::Location loc = op.getLoc();
  mlir::Value exnPtr = op.getExnPtr();
  mlir::Value paramAddr = op.getParamAddr();
  auto paramAddrType = mlir::cast<cir::PointerType>(paramAddr.getType());
  mlir::Type elementType = paramAddrType.getPointee();
  cir::InitCatchKind kind = op.getKind();
 
  switch (kind) {
  case InitCatchKind::Reference: {
    mlir::Value casted = cir::CastOp::create(builder, loc, elementType,
                                             cir::CastKind::bitcast, exnPtr);
    cir::StoreOp::create(builder, loc, casted, paramAddr, {}, {}, {}, {}, {});
    break;
  }
  case InitCatchKind::ReferenceToPointer:
    // The reference was bound to the exception object by the matching
    // cir.construct_catch_param before cir.begin_catch.
    break;
  case InitCatchKind::ReferenceToRecordPointer: {
    // __cxa_begin_catch returns the caught pointer by value, and the
    // personality function may have adjusted it.  Store the adjusted pointer
    // into a temporary and bind the reference to it.  Assigning to the catch
    // parameter then changes the temporary, not the exception object, which
    // departs from the language semantics.
    auto caughtType = mlir::cast<cir::PointerType>(
        mlir::cast<cir::PointerType>(elementType).getPointee());
    cir::AllocaOp byRefTmp;
    {
      mlir::OpBuilder::InsertionGuard guard(builder);
      builder.setInsertionPointToStart(
          &op->getParentOfType<cir::FuncOp>().getBody().front());
      cir::CIRDataLayout dataLayout(mod);
      uint64_t alignment =
          dataLayout.getAlignment(caughtType, /*useABIAlign=*/true).value();
      byRefTmp =
          cir::AllocaOp::create(builder, loc, elementType, "exn.byref.tmp",
                                builder.getI64IntegerAttr(alignment));
    }
    mlir::Value casted = cir::CastOp::create(builder, loc, caughtType,
                                             cir::CastKind::bitcast, exnPtr);
    cir::StoreOp::create(builder, loc, casted, byRefTmp, {}, {}, {}, {}, {});
    cir::StoreOp::create(builder, loc, byRefTmp, paramAddr, {}, {}, {}, {}, {});
    break;
  }
  case InitCatchKind::TrivialCopy: {
    mlir::Value srcPtr = cir::CastOp::create(builder, loc, paramAddrType,
                                             cir::CastKind::bitcast, exnPtr);
    cir::CopyOp::create(builder, loc, paramAddr, srcPtr, {}, {});
    break;
  }
  case InitCatchKind::NonTrivialCopy:
    // The non-trivial copy was performed by the matching
    // cir.construct_catch_param before cir.begin_catch.
    break;
  case InitCatchKind::Scalar: {
    // Scalar by-value catch (integer, float, complex, etc.). The begin_catch
    // result points into the exception object; load the value through a
    // typed pointer and store it into the alloca.
    mlir::Value srcPtr = cir::CastOp::create(builder, loc, paramAddrType,
                                             cir::CastKind::bitcast, exnPtr);
    auto loadOp = cir::LoadOp::create(builder, loc, elementType, srcPtr);
    cir::StoreOp::create(builder, loc, loadOp.getResult(), paramAddr, {}, {},
                         {}, {}, {});
    break;
  }
  case InitCatchKind::Pointer: {
    mlir::Value casted = cir::CastOp::create(builder, loc, elementType,
                                             cir::CastKind::bitcast, exnPtr);
    cir::StoreOp::create(builder, loc, casted, paramAddr, {}, {}, {}, {}, {});
    break;
  }
  case InitCatchKind::Objc:
    llvm_unreachable("InitCatchParam: ObjCLifetime is NYI");
    break;
  }
 
  op.erase();
}
 
//===----------------------------------------------------------------------===//
// The Pass
//===----------------------------------------------------------------------===//
 
struct CIREHABILoweringPass
    : public impl::CIREHABILoweringBase<CIREHABILoweringPass> {
  CIREHABILoweringPass() = default;
  void runOnOperation() override;
};
 
/// Erase all catch-init thunks after the EHABI lowering. CIRGen emits a thunk
/// for every `cir.construct_catch_param` op, but those uses should all have
/// been replaced during the lowering.
static void eraseCatchCopyThunks(mlir::ModuleOp mod) {
  llvm::StringRef catchHelperAttr =
      cir::CIRDialect::getCatchCopyThunkAttrName();
  for (cir::FuncOp f : llvm::make_early_inc_range(mod.getOps<cir::FuncOp>())) {
    if (!f->hasAttr(catchHelperAttr))
      continue;
    // This is an expensive check, so we need to rely on the implementation
    // to have done the right thing.
    assert(mlir::SymbolTable::symbolKnownUseEmpty(f, mod) &&
           "catch-init helper has remaining users");
    f.erase();
  }
}
 
void CIREHABILoweringPass::runOnOperation() {
  auto mod = mlir::cast<mlir::ModuleOp>(getOperation());
 
  // The target triple is attached to the module as the "cir.triple"
  // attribute. If it is absent (e.g. a CIR module parsed from text without a
  // triple) we cannot determine the ABI and must skip the pass.
  auto tripleAttr = mlir::dyn_cast_if_present<mlir::StringAttr>(
      mod->getAttr(cir::CIRDialect::getTripleAttrName()));
  if (!tripleAttr) {
    mod.emitError("Module has no target triple");
    return;
  }
 
  // Select the ABI-specific lowering handler from the triple. The Microsoft
  // C++ ABI targets a Windows MSVC environment; everything else uses Itanium.
  // Extend this when Microsoft ABI lowering is added.
  llvm::Triple triple(tripleAttr.getValue());
  std::unique_ptr<EHABILowering> lowering;
  if (triple.isWindowsMSVCEnvironment()) {
    mod.emitError(
        "EH ABI lowering is not yet implemented for the Microsoft ABI");
    return signalPassFailure();
  } else {
    lowering = std::make_unique<ItaniumEHLowering>(mod);
  }
 
  if (mlir::failed(lowering->run()))
    return signalPassFailure();
 
  // Sweep away any the thunk functions. They've been inlined to all users now.
  eraseCatchCopyThunks(mod);
}
 
} // namespace
 
std::unique_ptr<Pass> mlir::createCIREHABILoweringPass() {
  return std::make_unique<CIREHABILoweringPass>();
}