[CIR] Implement IndirectBrOp to support GCC's labels-as-values extension (#1945)

This PR introduces the `cir.indirectbr` operation to support GCC’s
labels-as-values extension, which allows `goto` statements to jump to
computed block addresses. The implementation creates a dedicated block
that hosts the `indirectbr`, where the target address is provided
through a PHI Node, similar to how classic code generation handles
indirect gotos.

Author: Andres-Salamanca

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

@@ -2297,6 +2297,67 @@ def CIR_BrCondOp : CIR_Op<"brcond", [
   }];
 }
 
+//===----------------------------------------------------------------------===//
+// IndirectBrOp
+//===----------------------------------------------------------------------===//
+
+def CIR_IndirectBrOp : CIR_Op<"indirectbr", [
+  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.alloca !cir.ptr<!void>, !cir.ptr<!cir.ptr<!void>>, ["ptr", init]
+      %1 = cir.blockaddress <@A, "A"> -> !cir.ptr<!void>
+      cir.store align(8) %1, %0 : !cir.ptr<!void>, !cir.ptr<!cir.ptr<!void>>
+      %2 = cir.load align(8) %0 : !cir.ptr<!cir.ptr<!void>>, !cir.ptr<!void>
+      cir.br ^bb1(%2 : !cir.ptr<!void>)
+    ^bb1(%3: !cir.ptr<!void>):
+      cir.indirectbr %3 : <!void>, [
+      ^bb2
+      ]
+    ```
+    or with a poison:
+
+    ```mlir
+      cir.indirectbr %0 poison : <!void>, [
+      ^bb3,
+      ^bb2
+      ]
+    ```
+  }];
+
+  let arguments = (ins
+    CIR_VoidPtrType:$addr,
+    UnitAttr:$poison,
+    VariadicOfVariadic<AnyType, "indbr_operand_segments">:$succOperands,
+    DenseI32ArrayAttr:$indbr_operand_segments
+    );
+
+  let successors = (successor VariadicSuccessor<AnySuccessor>:$successors);
+  let assemblyFormat = [{
+    $addr ( `poison` $poison^ )? `:` type($addr) `,`
+    custom<IndirectBrOpSucessors>(ref(type($addr)),
+                                  $successors,
+                                  $succOperands,
+                                  type($succOperands))
+    attr-dict
+  }];
+}
+
 //===----------------------------------------------------------------------===//
 // While & DoWhileOp
 //===----------------------------------------------------------------------===//

clang/include/clang/CIR/MissingFeatures.h

@@ -240,7 +240,6 @@ struct MissingFeatures {
   static bool emitScalarRangeCheck() { return false; }
   static bool stmtExprEvaluation() { return false; }
   static bool setCallingConv() { return false; }
-  static bool indirectBranch() { return false; }
   static bool escapedLocals() { return false; }
   static bool deferredReplacements() { return false; }
   static bool shouldInstrumentFunction() { return false; }

clang/lib/CIR/CodeGen/CIRGenExprScalar.cpp

@@ -231,10 +231,22 @@ class ScalarExprEmitter : public StmtVisitor<ScalarExprEmitter, mlir::Value> {
     auto func = cast<cir::FuncOp>(CGF.CurFn);
     auto 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);
+    auto 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.getIndirectGotoBlock(blockAddressOp);
+    return blockAddressOp;
   }
+
   mlir::Value VisitSizeOfPackExpr(SizeOfPackExpr *E) {
     llvm_unreachable("NYI");
   }

clang/lib/CIR/CodeGen/CIRGenFunction.cpp

@@ -529,6 +529,34 @@ cir::TryOp CIRGenFunction::LexicalScope::getClosestTryParent() {
   return nullptr;
 }
 
+void CIRGenFunction::resolveBlockAddresses() {
+
+  for (auto &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) {
   // CIRGen doesn't use a BreakContinueStack or evaluates OnlySimpleReturnStmts.
 
@@ -585,15 +613,23 @@ void CIRGenFunction::finishFunction(SourceLocation endLoc) {
 
   // If someone did an indirect goto, emit the indirect goto block at the end of
   // the function.
-  assert(!cir::MissingFeatures::indirectBranch() && "NYI");
 
+  // Resolve block address-to-label mappings, then emit the indirect branch
+  // with the corresponding targets.
+  resolveBlockAddresses();
+  finishIndirectBranch();
   // If some of our locals escaped, insert a call to llvm.localescape in the
   // entry block.
   assert(!cir::MissingFeatures::escapedLocals() && "NYI");
 
-  // If someone took the address of a label but never did an indirect goto, we
-  // made a zero entry PHI node, which is illegal, zap it now.
-  assert(!cir::MissingFeatures::indirectBranch() && "NYI");
+  // 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);
+  }
 
   // CIRGen doesn't need to emit EHResumeBlock, TerminateLandingPad,
   // TerminateHandler, UnreachableBlock, TerminateFunclets, NormalCleanupDest
@@ -1964,6 +2000,17 @@ CIRGenFunction::emitArrayLength(const clang::ArrayType *origArrayType,
   return numElements;
 }
 
+mlir::Block *CIRGenFunction::getIndirectGotoBlock(cir::BlockAddressOp op) {
+  // If we already made the indirect branch for indirect goto, return its block.
+  if (indirectGotoBlock)
+    return indirectGotoBlock;
+
+  mlir::OpBuilder::InsertionGuard guard(builder);
+  indirectGotoBlock = builder.createBlock(builder.getBlock()->getParent(), {},
+                                          {op.getType()}, {op.getLoc()});
+  return indirectGotoBlock;
+}
+
 mlir::Value CIRGenFunction::emitAlignmentAssumption(
     mlir::Value ptrValue, QualType ty, SourceLocation loc,
     SourceLocation assumptionLoc, mlir::IntegerAttr alignment,

clang/lib/CIR/CodeGen/CIRGenFunction.h

@@ -100,6 +100,15 @@ class CIRGenFunction : public CIRGenTypeCache {
   llvm::DenseMap<const clang::LabelDecl *, JumpDest> LabelMap;
   JumpDest &getJumpDestForLabel(const clang::LabelDecl *D);
 
+  /// 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();
+
   // ---------------------
   // Opaque value handling
   // ---------------------
@@ -688,6 +697,8 @@ class CIRGenFunction : public CIRGenTypeCache {
 
   int64_t getAccessedFieldNo(unsigned idx, const mlir::ArrayAttr elts);
 
+  mlir::Block *getIndirectGotoBlock(cir::BlockAddressOp op);
+
   void checkTargetFeatures(const CallExpr *E, const FunctionDecl *TargetDecl);
   void checkTargetFeatures(SourceLocation Loc, const FunctionDecl *TargetDecl);
 
@@ -2165,7 +2176,7 @@ class CIRGenFunction : public CIRGenTypeCache {
   mlir::LogicalResult emitFunctionBody(const clang::Stmt *Body);
 
   mlir::LogicalResult emitGotoStmt(const clang::GotoStmt &S);
-
+  mlir::LogicalResult emitIndirectGotoStmt(const IndirectGotoStmt &s);
   /// Emit an if on a boolean condition to the specified blocks.
   /// FIXME: Based on the condition, this might try to simplify the codegen of
   /// the conditional based on the branch. TrueCount should be the number of

clang/lib/CIR/CodeGen/CIRGenModule.cpp

@@ -4405,3 +4405,41 @@ CIRGenModule::mergeTBAAInfoForMemoryTransfer(TBAAAccessInfo destInfo,
     return TBAAAccessInfo();
   return tbaa->mergeTBAAInfoForConditionalOperator(destInfo, srcInfo);
 }
+
+void CIRGenModule::mapBlockAddress(cir::BlockAddrInfoAttr blockInfo,
+                                   cir::LabelOp label) {
+  auto result = blockAddressInfoToLabel.try_emplace(blockInfo, label);
+  (void)result;
+  assert(result.second &&
+         "attempting to map a blockaddress info that is already mapped");
+}
+
+void CIRGenModule::mapUnresolvedBlockAddress(cir::BlockAddressOp op) {
+  auto result = unresolvedBlockAddressToLabel.insert(op);
+  (void)result;
+  assert(result.second &&
+         "attempting to map a blockaddress operation that is already mapped");
+}
+
+void CIRGenModule::mapResolvedBlockAddress(cir::BlockAddressOp op,
+                                           cir::LabelOp label) {
+  auto result = blockAddressToLabel.try_emplace(op, label);
+  (void)result;
+  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

@@ -206,6 +206,24 @@ class CIRGenModule : public CIRGenTypeCache {
   /// 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);
+
   /// If the declaration has internal linkage but is inside an
   /// extern "C" linkage specification, prepare to emit an alias for it
   /// to the expected name.

clang/lib/CIR/CodeGen/CIRGenStmt.cpp

@@ -194,8 +194,9 @@ mlir::LogicalResult CIRGenFunction::emitStmt(const Stmt *S,
     return emitOMPTaskyieldDirective(cast<OMPTaskyieldDirective>(*S));
   case Stmt::OMPBarrierDirectiveClass:
     return emitOMPBarrierDirective(cast<OMPBarrierDirective>(*S));
-  // Unsupported AST nodes:
   case Stmt::IndirectGotoStmtClass:
+    return emitIndirectGotoStmt(cast<IndirectGotoStmt>(*S));
+  // Unsupported AST nodes:
   case Stmt::CapturedStmtClass:
   case Stmt::ObjCAtTryStmtClass:
   case Stmt::ObjCAtThrowStmtClass:
@@ -652,15 +653,31 @@ mlir::LogicalResult CIRGenFunction::emitLabel(const LabelDecl *D) {
   }
 
   builder.setInsertionPointToEnd(labelBlock);
-  builder.create<cir::LabelOp>(getLoc(D->getSourceRange()), D->getName());
+  auto label =
+      builder.create<cir::LabelOp>(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
   return mlir::success();
 }
 
-mlir::LogicalResult CIRGenFunction::emitContinueStmt(
-    const clang::ContinueStmt &S) {
+mlir::LogicalResult
+CIRGenFunction::emitIndirectGotoStmt(const IndirectGotoStmt &s) {
+  auto 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.getBeginLoc()));
 
   // Insert the new block to continue codegen after the continue statement.

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

@@ -1854,6 +1854,65 @@ Block *cir::BrCondOp::getSuccessorForOperands(ArrayRef<Attribute> operands) {
   return nullptr;
 }
 
+//===----------------------------------------------------------------------===//
+// 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 << "]";
+}
+
 //===----------------------------------------------------------------------===//
 // CaseOp
 //===----------------------------------------------------------------------===//