[CIR][CIRGen] Support for builtin __atomic_thread_fence

Resolves #1274

Implements atomic thread fence synchronization primitive
corresponding to `atomic.thread_fence` CIR.

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

@@ -5409,6 +5409,47 @@ def AtomicCmpXchg : CIR_Op<"atomic.cmp_xchg",
   let hasVerifier = 0;
 }
 
+def MemScope_SingleThread : I32EnumAttrCase<"MemScope_SingleThread",
+                                            0, "single_thread">;
+def MemScope_System       : I32EnumAttrCase<"MemScope_System",
+                                            1, "system">;
+
+def MemScopeKind : I32EnumAttr<
+    "MemScopeKind",
+    "Memory Scope Enumeration",
+    [MemScope_SingleThread, MemScope_System]> {
+  let cppNamespace = "::cir";
+}
+
+def AtomicFence : CIR_Op<"atomic.fence"> {
+  let summary = "Atomic thread fence";
+  let description = [{
+    C/C++ Atomic thread fence synchronization primitive. Implements the builtin
+    `__atomic_thread_fence` which enforces memory ordering constraints across
+    threads within the specified synchronization scope.
+
+    This handles all variations including:
+      - `__atomic_thread_fence`
+      - `__atomic_signal_fence`
+      - `__c11_atomic_thread_fence`
+      - `__c11_atomic_signal_fence`
+
+    Example:
+
+
+  }];
+  let results = (outs);
+  let arguments = (ins Arg<MemScopeKind, "sync scope">:$sync_scope,
+                       Arg<MemOrder, "memory order">:$ordering);
+
+  let assemblyFormat = [{
+    `(` `sync_scope` `=` $sync_scope `,`
+        `ordering` `=` $ordering `)` attr-dict
+  }];
+
+  let hasVerifier = 0;
+}
+
 def SignBitOp : CIR_Op<"signbit", [Pure]> {
   let summary = "Checks the sign of a floating-point number";
   let description = [{

clang/lib/CIR/CodeGen/CIRGenBuiltin.cpp

@@ -16,7 +16,10 @@
 #include "CIRGenCstEmitter.h"
 #include "CIRGenFunction.h"
 #include "CIRGenModule.h"
+#include "CIRGenValue.h"
 #include "TargetInfo.h"
+#include "clang/AST/Expr.h"
+#include "clang/CIR/Dialect/IR/CIROpsEnums.h"
 #include "clang/CIR/MissingFeatures.h"
 
 // TODO(cir): we shouldn't need this but we currently reuse intrinsic IDs for
@@ -30,6 +33,7 @@
 #include "clang/Frontend/FrontendDiagnostic.h"
 
 #include "mlir/Dialect/Func/IR/FuncOps.h"
+#include "mlir/IR/BuiltinAttributes.h"
 #include "mlir/IR/Value.h"
 #include "clang/CIR/Dialect/IR/CIRDialect.h"
 #include "llvm/Support/ErrorHandling.h"
@@ -330,6 +334,30 @@ static mlir::Value MakeAtomicCmpXchgValue(CIRGenFunction &cgf,
   return returnBool ? op.getResult(1) : op.getResult(0);
 }
 
+static mlir::Value MakeAtomicFenceValue(CIRGenFunction &cgf,
+                                        const CallExpr *expr,
+                                        cir::MemScopeKind syncScope) {
+  QualType typ = expr->getType();
+  auto &builder = cgf.getBuilder();
+
+  auto intType =
+      expr->getArg(0)->getType()->getPointeeType()->isUnsignedIntegerType()
+          ? builder.getUIntNTy(cgf.getContext().getTypeSize(typ))
+          : builder.getSIntNTy(cgf.getContext().getTypeSize(typ));
+
+  auto orderingVal =
+      emitToInt(cgf, cgf.emitScalarExpr(expr->getArg(0)), typ, intType);
+  auto orderingAttr =
+      orderingVal.getDefiningOp()->getAttrOfType<mlir::IntegerAttr>("value");
+
+  cir::MemOrder ordering = static_cast<cir::MemOrder>(orderingAttr.getInt());
+
+  builder.create<cir::AtomicFence>(cgf.getLoc(expr->getSourceRange()),
+                                   syncScope, ordering);
+
+  return mlir::Value();
+}
+
 static bool
 typeRequiresBuiltinLaunderImp(const ASTContext &astContext, QualType ty,
                               llvm::SmallPtrSetImpl<const Decl *> &seen) {
@@ -1840,10 +1868,14 @@ RValue CIRGenFunction::emitBuiltinExpr(const GlobalDecl GD, unsigned BuiltinID,
     llvm_unreachable("BI__atomic_clear NYI");
 
   case Builtin::BI__atomic_thread_fence:
+    return RValue::get(
+        MakeAtomicFenceValue(*this, E, cir::MemScopeKind::MemScope_System));
   case Builtin::BI__atomic_signal_fence:
+    return RValue::get(MakeAtomicFenceValue(
+        *this, E, cir::MemScopeKind::MemScope_SingleThread));
   case Builtin::BI__c11_atomic_thread_fence:
   case Builtin::BI__c11_atomic_signal_fence:
-    llvm_unreachable("BI__atomic_thread_fence like NYI");
+    llvm_unreachable("BI__c11_atomic_thread_fence like NYI");
 
   case Builtin::BI__builtin_signbit:
   case Builtin::BI__builtin_signbitf:

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

@@ -18,6 +18,7 @@
 #include "mlir/Conversion/SCFToControlFlow/SCFToControlFlow.h"
 #include "mlir/Dialect/DLTI/DLTI.h"
 #include "mlir/Dialect/Func/IR/FuncOps.h"
+#include "mlir/Dialect/LLVMIR/LLVMAttrs.h"
 #include "mlir/Dialect/LLVMIR/Transforms/Passes.h"
 #include "mlir/IR/Attributes.h"
 #include "mlir/IR/Builders.h"
@@ -41,6 +42,8 @@
 #include "mlir/Target/LLVMIR/Dialect/LLVMIR/LLVMToLLVMIRTranslation.h"
 #include "mlir/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.h"
 #include "mlir/Target/LLVMIR/Export.h"
+#include "clang/CIR/Dialect/IR/CIRDialect.h"
+#include "clang/CIR/Dialect/IR/CIROpsEnums.h"
 #include "clang/CIR/Dialect/Passes.h"
 #include "clang/CIR/LoweringHelpers.h"
 #include "clang/CIR/MissingFeatures.h"
@@ -3399,6 +3402,12 @@ mlir::LogicalResult CIRToLLVMAtomicFetchLowering::matchAndRewrite(
   return mlir::success();
 }
 
+// mlir::LogicalResult CIRToLLVMAtomicFenceLowering::matchAndRewrite(
+//     cir::AtomicFence op, OpAdaptor adaptor,
+//     mlir::ConversionPatternRewriter &rewriter) const {
+//   return mlir::success();
+// }
+
 mlir::LogicalResult CIRToLLVMByteswapOpLowering::matchAndRewrite(
     cir::ByteswapOp op, OpAdaptor adaptor,
     mlir::ConversionPatternRewriter &rewriter) const {

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

@@ -16,6 +16,7 @@
 #include "mlir/IR/MLIRContext.h"
 #include "mlir/Interfaces/DataLayoutInterfaces.h"
 #include "mlir/Transforms/DialectConversion.h"
+#include "clang/CIR/Dialect/IR/CIRDialect.h"
 
 namespace cir {
 namespace direct {
@@ -820,6 +821,16 @@ class CIRToLLVMAtomicFetchLowering
                   mlir::ConversionPatternRewriter &) const override;
 };
 
+// class CIRToLLVMAtomicFenceLowering
+//     : public mlir::OpConversionPattern<cir::AtomicFence> {
+// public:
+//   using mlir::OpConversionPattern<cir::AtomicFence>::OpConversionPattern;
+//
+//   mlir::LogicalResult
+//   matchAndRewrite(cir::AtomicFence op, OpAdaptor,
+//                   mlir::ConversionPatternRewriter &) const override;
+// };
+
 class CIRToLLVMByteswapOpLowering
     : public mlir::OpConversionPattern<cir::ByteswapOp> {
 public:

clang/test/CIR/CodeGen/atomic-thread-fence.c

@@ -0,0 +1,40 @@
+// RUN: %clang_cc1 -triple aarch64-none-linux-android21 -fclangir -emit-cir %s -o %t.cir
+// RUN: FileCheck --input-file=%t.cir %s
+// UN: %clang_cc1 -triple x86_64-unknown-linux-gnu -fclangir -emit-llvm %s -o %t.ll
+// UN: FileCheck --check-prefix=LLVM --input-file=%t.ll %s
+
+
+struct Data {
+  int value;
+  void *ptr;
+};
+
+typedef struct Data *DataPtr;
+
+void applyThreadFence() {
+  __atomic_thread_fence(5);
+}
+
+void applySignalFence() {
+  __atomic_signal_fence(5);
+}
+
+void modifyWithThreadFence(DataPtr d) {
+  __atomic_thread_fence(5);
+  d->value = 42;
+}
+
+void modifyWithSignalFence(DataPtr d) {
+  __atomic_signal_fence(5);
+  d->value = 24;
+}
+
+void loadWithThreadFence(DataPtr d) {
+  __atomic_thread_fence(5);
+  __atomic_load_n(&d->ptr, 5);
+}
+
+void loadWithSignalFence(DataPtr d) {
+  __atomic_signal_fence(5);
+  __atomic_load_n(&d->ptr, 5);
+}