[CIR] Upstream support for cir.indirectbr (#169967)

This PR upstreams support for the `cir.indirectBr` operation, which is
used to implement GCC’s labels-as-values `indirect goto`.
To ensure correct lowering, we introduce precise bookkeeping to
associate each `block_address` operation with its corresponding `label`
op. This is required because a `block_address` may be emitted before the
`label` it refers to. In such cases, the reference is deferred and later
resolved by `resolveBlockAddresses`, which guarantees that all
`indirectBr` successors are emitted in the correct and fully resolved
order.

Author: Andres-Salamanca

clang/include/clang/CIR/Dialect/IR/CIROps.td

@@ -1492,6 +1492,55 @@ def CIR_BrCondOp : CIR_Op<"brcond", [
   }];
 }
 
+//===----------------------------------------------------------------------===//
+// IndirectBrOp
+//===----------------------------------------------------------------------===//
+
+def CIR_IndirectBrOp : CIR_Op<"indirect_br", [
+  DeclareOpInterfaceMethods<BranchOpInterface>,
+  SameVariadicOperandSize, Terminator, Pure
+]> {
+  let summary = "Indirect branch";
+  let description = [{
+    The `cir.indirectbr` operation represents an indirect branch to one of
+    several possible successor blocks. The target block is computed from
+    the value of the given address operand.
+
+    This operation is typically generated when handling constructs like
+    the GCC extension `&&label` combined with an indirect `goto *ptr;`.
+
+    The `poison` attribute is used to mark an `indirectbr` that was created
+    but is known to be invalid, for instance when a label address was
+    taken but no indirect branch was ever emitted.
+
+    Example:
+
+    ```mlir
+      %0 = cir.block_address <@A, "A"> : !cir.ptr<!void>
+      cir.indirectbr %0 poison : <!void>, [
+      ^bb1
+      ]
+    ```
+  }];
+
+  let arguments = (ins
+    CIR_VoidPtrType:$addr,
+    UnitAttr:$poison,
+    VariadicOfVariadic<AnyType, "operand_segments">:$succ_operands,
+    DenseI32ArrayAttr:$operand_segments
+    );
+
+  let successors = (successor VariadicSuccessor<AnySuccessor>:$successors);
+  let assemblyFormat = [{
+    $addr ( `poison` $poison^ )? `:` qualified(type($addr)) `,`
+    custom<IndirectBrOpSucessors>(ref(type($addr)),
+                                  $successors,
+                                  $succ_operands,
+                                  type($succ_operands))
+    attr-dict
+  }];
+}
+
 //===----------------------------------------------------------------------===//
 // Common loop op definitions
 //===----------------------------------------------------------------------===//

clang/lib/CIR/CodeGen/CIRGenExprScalar.cpp

@@ -174,11 +174,22 @@ class ScalarExprEmitter : public StmtVisitor<ScalarExprEmitter, mlir::Value> {
 
   mlir::Value VisitAddrLabelExpr(const AddrLabelExpr *e) {
     auto func = cast<cir::FuncOp>(cgf.curFn);
-    auto blockInfoAttr = cir::BlockAddrInfoAttr::get(
+    cir::BlockAddrInfoAttr blockInfoAttr = cir::BlockAddrInfoAttr::get(
         &cgf.getMLIRContext(), func.getSymName(), e->getLabel()->getName());
-    return cir::BlockAddressOp::create(builder, cgf.getLoc(e->getSourceRange()),
-                                       cgf.convertType(e->getType()),
-                                       blockInfoAttr);
+    cir::BlockAddressOp blockAddressOp = cir::BlockAddressOp::create(
+        builder, cgf.getLoc(e->getSourceRange()), cgf.convertType(e->getType()),
+        blockInfoAttr);
+    cir::LabelOp resolvedLabel = cgf.cgm.lookupBlockAddressInfo(blockInfoAttr);
+    if (!resolvedLabel) {
+      cgf.cgm.mapUnresolvedBlockAddress(blockAddressOp);
+      // Still add the op to maintain insertion order it will be resolved in
+      // resolveBlockAddresses
+      cgf.cgm.mapResolvedBlockAddress(blockAddressOp, nullptr);
+    } else {
+      cgf.cgm.mapResolvedBlockAddress(blockAddressOp, resolvedLabel);
+    }
+    cgf.instantiateIndirectGotoBlock();
+    return blockAddressOp;
   }
 
   mlir::Value VisitIntegerLiteral(const IntegerLiteral *e) {

clang/lib/CIR/CodeGen/CIRGenFunction.cpp

@@ -531,7 +531,48 @@ void CIRGenFunction::startFunction(GlobalDecl gd, QualType returnType,
   }
 }
 
+void CIRGenFunction::resolveBlockAddresses() {
+  for (cir::BlockAddressOp &blockAddress : cgm.unresolvedBlockAddressToLabel) {
+    cir::LabelOp labelOp =
+        cgm.lookupBlockAddressInfo(blockAddress.getBlockAddrInfo());
+    assert(labelOp && "expected cir.labelOp to already be emitted");
+    cgm.updateResolvedBlockAddress(blockAddress, labelOp);
+  }
+  cgm.unresolvedBlockAddressToLabel.clear();
+}
+
+void CIRGenFunction::finishIndirectBranch() {
+  if (!indirectGotoBlock)
+    return;
+  llvm::SmallVector<mlir::Block *> succesors;
+  llvm::SmallVector<mlir::ValueRange> rangeOperands;
+  mlir::OpBuilder::InsertionGuard guard(builder);
+  builder.setInsertionPointToEnd(indirectGotoBlock);
+  for (auto &[blockAdd, labelOp] : cgm.blockAddressToLabel) {
+    succesors.push_back(labelOp->getBlock());
+    rangeOperands.push_back(labelOp->getBlock()->getArguments());
+  }
+  cir::IndirectBrOp::create(builder, builder.getUnknownLoc(),
+                            indirectGotoBlock->getArgument(0), false,
+                            rangeOperands, succesors);
+  cgm.blockAddressToLabel.clear();
+}
+
 void CIRGenFunction::finishFunction(SourceLocation endLoc) {
+  // Resolve block address-to-label mappings, then emit the indirect branch
+  // with the corresponding targets.
+  resolveBlockAddresses();
+  finishIndirectBranch();
+
+  // If a label address was taken but no indirect goto was used, we can't remove
+  // the block argument here. Instead, we mark the 'indirectbr' op
+  // as poison so that the cleanup can be deferred to lowering, since the
+  // verifier doesn't allow the 'indirectbr' target address to be null.
+  if (indirectGotoBlock && indirectGotoBlock->hasNoPredecessors()) {
+    auto indrBr = cast<cir::IndirectBrOp>(indirectGotoBlock->front());
+    indrBr.setPoison(true);
+  }
+
   // Pop any cleanups that might have been associated with the
   // parameters.  Do this in whatever block we're currently in; it's
   // important to do this before we enter the return block or return
@@ -1086,6 +1127,17 @@ CIRGenFunction::emitArrayLength(const clang::ArrayType *origArrayType,
   return builder.getConstInt(*currSrcLoc, sizeTy, countFromCLAs);
 }
 
+void CIRGenFunction::instantiateIndirectGotoBlock() {
+  // If we already made the indirect branch for indirect goto, return its block.
+  if (indirectGotoBlock)
+    return;
+
+  mlir::OpBuilder::InsertionGuard guard(builder);
+  indirectGotoBlock =
+      builder.createBlock(builder.getBlock()->getParent(), {}, {voidPtrTy},
+                          {builder.getUnknownLoc()});
+}
+
 mlir::Value CIRGenFunction::emitAlignmentAssumption(
     mlir::Value ptrValue, QualType ty, SourceLocation loc,
     SourceLocation assumptionLoc, int64_t alignment, mlir::Value offsetValue) {

clang/lib/CIR/CodeGen/CIRGenFunction.h

@@ -658,6 +658,14 @@ class CIRGenFunction : public CIRGenTypeCache {
     return JumpDest(target, ehStack.getInnermostNormalCleanup(),
                     nextCleanupDestIndex++);
   }
+  /// IndirectBranch - The first time an indirect goto is seen we create a block
+  /// reserved for the indirect branch. Unlike before,the actual 'indirectbr'
+  /// is emitted at the end of the function, once all block destinations have
+  /// been resolved.
+  mlir::Block *indirectGotoBlock = nullptr;
+
+  void resolveBlockAddresses();
+  void finishIndirectBranch();
 
   /// Perform the usual unary conversions on the specified expression and
   /// compare the result against zero, returning an Int1Ty value.
@@ -1392,6 +1400,8 @@ class CIRGenFunction : public CIRGenTypeCache {
 
   int64_t getAccessedFieldNo(unsigned idx, mlir::ArrayAttr elts);
 
+  void instantiateIndirectGotoBlock();
+
   RValue emitCall(const CIRGenFunctionInfo &funcInfo,
                   const CIRGenCallee &callee, ReturnValueSlot returnValue,
                   const CallArgList &args, cir::CIRCallOpInterface *callOp,
@@ -1575,6 +1585,8 @@ class CIRGenFunction : public CIRGenTypeCache {
 
   mlir::LogicalResult emitGotoStmt(const clang::GotoStmt &s);
 
+  mlir::LogicalResult emitIndirectGotoStmt(const IndirectGotoStmt &s);
+
   void emitImplicitAssignmentOperatorBody(FunctionArgList &args);
 
   void emitInitializerForField(clang::FieldDecl *field, LValue lhs,

clang/lib/CIR/CodeGen/CIRGenModule.cpp

@@ -2498,3 +2498,39 @@ DiagnosticBuilder CIRGenModule::errorNYI(SourceRange loc,
                                          llvm::StringRef feature) {
   return errorNYI(loc.getBegin(), feature) << loc;
 }
+
+void CIRGenModule::mapBlockAddress(cir::BlockAddrInfoAttr blockInfo,
+                                   cir::LabelOp label) {
+  [[maybe_unused]] auto result =
+      blockAddressInfoToLabel.try_emplace(blockInfo, label);
+  assert(result.second &&
+         "attempting to map a blockaddress info that is already mapped");
+}
+
+void CIRGenModule::mapUnresolvedBlockAddress(cir::BlockAddressOp op) {
+  [[maybe_unused]] auto result = unresolvedBlockAddressToLabel.insert(op);
+  assert(result.second &&
+         "attempting to map a blockaddress operation that is already mapped");
+}
+
+void CIRGenModule::mapResolvedBlockAddress(cir::BlockAddressOp op,
+                                           cir::LabelOp label) {
+  [[maybe_unused]] auto result = blockAddressToLabel.try_emplace(op, label);
+  assert(result.second &&
+         "attempting to map a blockaddress operation that is already mapped");
+}
+
+void CIRGenModule::updateResolvedBlockAddress(cir::BlockAddressOp op,
+                                              cir::LabelOp newLabel) {
+  auto *it = blockAddressToLabel.find(op);
+  assert(it != blockAddressToLabel.end() &&
+         "trying to update a blockaddress not previously mapped");
+  assert(!it->second && "blockaddress already has a resolved label");
+
+  it->second = newLabel;
+}
+
+cir::LabelOp
+CIRGenModule::lookupBlockAddressInfo(cir::BlockAddrInfoAttr blockInfo) {
+  return blockAddressInfoToLabel.lookup(blockInfo);
+}

clang/lib/CIR/CodeGen/CIRGenModule.h

@@ -126,7 +126,23 @@ class CIRGenModule : public CIRGenTypeCache {
   /// the pointers are supposed to be uniqued, should be fine. Revisit this if
   /// it ends up taking too much memory.
   llvm::DenseMap<const clang::FieldDecl *, llvm::StringRef> lambdaFieldToName;
-
+  /// Map BlockAddrInfoAttr (function name, label name) to the corresponding CIR
+  /// LabelOp. This provides the main lookup table used to resolve block
+  /// addresses into their label operations.
+  llvm::DenseMap<cir::BlockAddrInfoAttr, cir::LabelOp> blockAddressInfoToLabel;
+  /// Map CIR BlockAddressOps directly to their resolved LabelOps.
+  /// Used once a block address has been successfully lowered to a label.
+  llvm::MapVector<cir::BlockAddressOp, cir::LabelOp> blockAddressToLabel;
+  /// Track CIR BlockAddressOps that cannot be resolved immediately
+  /// because their LabelOp has not yet been emitted. These entries
+  /// are solved later once the corresponding label is available.
+  llvm::DenseSet<cir::BlockAddressOp> unresolvedBlockAddressToLabel;
+  cir::LabelOp lookupBlockAddressInfo(cir::BlockAddrInfoAttr blockInfo);
+  void mapBlockAddress(cir::BlockAddrInfoAttr blockInfo, cir::LabelOp label);
+  void mapUnresolvedBlockAddress(cir::BlockAddressOp op);
+  void mapResolvedBlockAddress(cir::BlockAddressOp op, cir::LabelOp);
+  void updateResolvedBlockAddress(cir::BlockAddressOp op,
+                                  cir::LabelOp newLabel);
   /// Tell the consumer that this variable has been instantiated.
   void handleCXXStaticMemberVarInstantiation(VarDecl *vd);
 

clang/lib/CIR/CodeGen/CIRGenStmt.cpp

@@ -203,6 +203,7 @@ mlir::LogicalResult CIRGenFunction::emitStmt(const Stmt *s,
     return emitCoroutineBody(cast<CoroutineBodyStmt>(*s));
   case Stmt::CoreturnStmtClass:
   case Stmt::IndirectGotoStmtClass:
+    return emitIndirectGotoStmt(cast<IndirectGotoStmt>(*s));
   case Stmt::OMPParallelDirectiveClass:
   case Stmt::OMPTaskwaitDirectiveClass:
   case Stmt::OMPTaskyieldDirectiveClass:
@@ -562,6 +563,17 @@ mlir::LogicalResult CIRGenFunction::emitGotoStmt(const clang::GotoStmt &s) {
   return mlir::success();
 }
 
+mlir::LogicalResult
+CIRGenFunction::emitIndirectGotoStmt(const IndirectGotoStmt &s) {
+  mlir::Value val = emitScalarExpr(s.getTarget());
+  assert(indirectGotoBlock &&
+         "If you jumping to a indirect branch should be alareadye emitted");
+  cir::BrOp::create(builder, getLoc(s.getSourceRange()), indirectGotoBlock,
+                    val);
+  builder.createBlock(builder.getBlock()->getParent());
+  return mlir::success();
+}
+
 mlir::LogicalResult
 CIRGenFunction::emitContinueStmt(const clang::ContinueStmt &s) {
   builder.createContinue(getLoc(s.getKwLoc()));
@@ -588,9 +600,14 @@ mlir::LogicalResult CIRGenFunction::emitLabel(const clang::LabelDecl &d) {
   }
 
   builder.setInsertionPointToEnd(labelBlock);
-  cir::LabelOp::create(builder, getLoc(d.getSourceRange()), d.getName());
+  cir::LabelOp label =
+      cir::LabelOp::create(builder, getLoc(d.getSourceRange()), d.getName());
   builder.setInsertionPointToEnd(labelBlock);
-
+  auto func = cast<cir::FuncOp>(curFn);
+  cgm.mapBlockAddress(cir::BlockAddrInfoAttr::get(builder.getContext(),
+                                                  func.getSymNameAttr(),
+                                                  label.getLabelAttr()),
+                      label);
   //  FIXME: emit debug info for labels, incrementProfileCounter
   assert(!cir::MissingFeatures::ehstackBranches());
   assert(!cir::MissingFeatures::incrementProfileCounter());

clang/lib/CIR/Dialect/IR/CIRDialect.cpp

@@ -1247,6 +1247,65 @@ Block *cir::BrOp::getSuccessorForOperands(ArrayRef<Attribute>) {
   return getDest();
 }
 
+//===----------------------------------------------------------------------===//
+// IndirectBrCondOp
+//===----------------------------------------------------------------------===//
+
+mlir::SuccessorOperands
+cir::IndirectBrOp::getSuccessorOperands(unsigned index) {
+  assert(index < getNumSuccessors() && "invalid successor index");
+  return mlir::SuccessorOperands(getSuccOperandsMutable()[index]);
+}
+
+ParseResult parseIndirectBrOpSucessors(
+    OpAsmParser &parser, Type &flagType,
+    SmallVectorImpl<Block *> &succOperandBlocks,
+    SmallVectorImpl<SmallVector<OpAsmParser::UnresolvedOperand>> &succOperands,
+    SmallVectorImpl<SmallVector<Type>> &succOperandsTypes) {
+  if (failed(parser.parseCommaSeparatedList(
+          OpAsmParser::Delimiter::Square,
+          [&]() {
+            Block *destination = nullptr;
+            SmallVector<OpAsmParser::UnresolvedOperand> operands;
+            SmallVector<Type> operandTypes;
+
+            if (parser.parseSuccessor(destination).failed())
+              return failure();
+
+            if (succeeded(parser.parseOptionalLParen())) {
+              if (failed(parser.parseOperandList(
+                      operands, OpAsmParser::Delimiter::None)) ||
+                  failed(parser.parseColonTypeList(operandTypes)) ||
+                  failed(parser.parseRParen()))
+                return failure();
+            }
+            succOperandBlocks.push_back(destination);
+            succOperands.emplace_back(operands);
+            succOperandsTypes.emplace_back(operandTypes);
+            return success();
+          },
+          "successor blocks")))
+    return failure();
+  return success();
+}
+
+void printIndirectBrOpSucessors(OpAsmPrinter &p, cir::IndirectBrOp op,
+                                Type flagType, SuccessorRange succs,
+                                OperandRangeRange succOperands,
+                                const TypeRangeRange &succOperandsTypes) {
+  p << "[";
+  llvm::interleave(
+      llvm::zip(succs, succOperands),
+      [&](auto i) {
+        p.printNewline();
+        p.printSuccessorAndUseList(std::get<0>(i), std::get<1>(i));
+      },
+      [&] { p << ','; });
+  if (!succOperands.empty())
+    p.printNewline();
+  p << "]";
+}
+
 //===----------------------------------------------------------------------===//
 // BrCondOp
 //===----------------------------------------------------------------------===//

clang/lib/CIR/Dialect/Transforms/CIRCanonicalize.cpp

@@ -52,7 +52,7 @@ struct RemoveRedundantBranches : public OpRewritePattern<BrOp> {
     Block *block = op.getOperation()->getBlock();
     Block *dest = op.getDest();
 
-    if (isa<cir::LabelOp>(dest->front()))
+    if (isa<cir::LabelOp, cir::IndirectBrOp>(dest->front()))
       return failure();
     // Single edge between blocks: merge it.
     if (block->getNumSuccessors() == 1 &&

clang/lib/CIR/Lowering/DirectToLLVM/LowerToLLVM.cpp

@@ -4125,6 +4125,12 @@ mlir::LogicalResult CIRToLLVMBlockAddressOpLowering::matchAndRewrite(
   return mlir::failure();
 }
 
+mlir::LogicalResult CIRToLLVMIndirectBrOpLowering::matchAndRewrite(
+    cir::IndirectBrOp op, OpAdaptor adaptor,
+    mlir::ConversionPatternRewriter &rewriter) const {
+  return mlir::failure();
+}
+
 mlir::LogicalResult CIRToLLVMAwaitOpLowering::matchAndRewrite(
     cir::AwaitOp op, OpAdaptor adaptor,
     mlir::ConversionPatternRewriter &rewriter) const {