[CIR] Upstream AtomicFenceOp and signal/thread fence builtins

[HendrikHuebner]

This PR upstreams the AtomicFence operation from the incubator repo.

[llvmbot]

@llvm/pr-subscribers-clang

@llvm/pr-subscribers-clangir

Author: Hendrik Hübner (HendrikHuebner)

Changes

This PR upstreams the AtomicFence operation from the incubator repo.

---

Full diff: https://github.com/llvm/llvm-project/pull/168346.diff

3 Files Affected:

• (modified) clang/include/clang/CIR/Dialect/IR/CIROps.td (+35)
• (modified) clang/lib/CIR/CodeGen/CIRGenBuiltin.cpp (+39)
• (modified) clang/lib/CIR/Lowering/DirectToLLVM/LowerToLLVM.cpp (+21)

diff --git a/clang/include/clang/CIR/Dialect/IR/CIROps.td b/clang/include/clang/CIR/Dialect/IR/CIROps.td
index 2124b1dc62a81..ee63ac39c5ba4 100644
--- a/clang/include/clang/CIR/Dialect/IR/CIROps.td
+++ b/clang/include/clang/CIR/Dialect/IR/CIROps.td
@@ -4845,4 +4845,39 @@ def CIR_AtomicClearOp : CIR_Op<"atomic.clear"> {
   }];
 }
 
+def CIR_MemScopeKind : CIR_I32EnumAttr<"MemScopeKind", "memory scope kind", [
+  I32EnumAttrCase<"SingleThread", 0, "single_thread">,
+  I32EnumAttrCase<"System", 1, "system">
+]>;
+
+def CIR_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:
+    ```mlir
+      cir.atomic.fence syncscope(system) seq_cst
+      cir.atomic.fence syncscope(single_thread) seq_cst
+    ```
+  }];
+
+  let arguments = (ins
+    Arg<CIR_MemOrder, "memory order">:$ordering,
+    OptionalAttr<CIR_MemScopeKind>:$syncscope
+  );
+
+  let assemblyFormat = [{
+    (`syncscope` `(` $syncscope^ `)`)? $ordering attr-dict
+  }];
+}
+
 #endif // CLANG_CIR_DIALECT_IR_CIROPS_TD
diff --git a/clang/lib/CIR/CodeGen/CIRGenBuiltin.cpp b/clang/lib/CIR/CodeGen/CIRGenBuiltin.cpp
index 77f19343653db..1223669a76ab1 100644
--- a/clang/lib/CIR/CodeGen/CIRGenBuiltin.cpp
+++ b/clang/lib/CIR/CodeGen/CIRGenBuiltin.cpp
@@ -58,6 +58,28 @@ static RValue emitBuiltinBitOp(CIRGenFunction &cgf, const CallExpr *e,
   return RValue::get(result);
 }
 
+static mlir::Value makeAtomicFenceValue(CIRGenFunction &cgf,
+                                        const CallExpr *expr,
+                                        cir::MemScopeKind syncScope) {
+  auto &builder = cgf.getBuilder();
+  mlir::Value orderingVal = cgf.emitScalarExpr(expr->getArg(0));
+
+  auto constOrdering = orderingVal.getDefiningOp<cir::ConstantOp>();
+  if (!constOrdering)
+    llvm_unreachable("NYI: variable ordering not supported");
+
+  if (auto constOrderingAttr = constOrdering.getValueAttr<cir::IntAttr>()) {
+    cir::MemOrder ordering =
+        static_cast<cir::MemOrder>(constOrderingAttr.getUInt());
+
+    cir::AtomicFence::create(
+        builder, cgf.getLoc(expr->getSourceRange()), ordering,
+        cir::MemScopeKindAttr::get(&cgf.getMLIRContext(), syncScope));
+  }
+
+  return {};
+}
+
 RValue CIRGenFunction::emitRotate(const CallExpr *e, bool isRotateLeft) {
   mlir::Value input = emitScalarExpr(e->getArg(0));
   mlir::Value amount = emitScalarExpr(e->getArg(1));
@@ -520,6 +542,23 @@ RValue CIRGenFunction::emitBuiltinExpr(const GlobalDecl &gd, unsigned builtinID,
     cir::PrefetchOp::create(builder, loc, address, locality, isWrite);
     return RValue::get(nullptr);
   }
+  case Builtin::BI__c11_atomic_is_lock_free:
+    llvm_unreachable("BI__c11_atomic_is_lock_free NYI");
+  case Builtin::BI__atomic_is_lock_free:
+    llvm_unreachable("BI__atomic_is_lock_free NYI");
+  case Builtin::BI__atomic_test_and_set:
+    llvm_unreachable("BI__atomic_test_and_set NYI");
+  case Builtin::BI__atomic_clear:
+    llvm_unreachable("BI__atomic_clear NYI");
+  case Builtin::BI__atomic_thread_fence:
+    return RValue::get(
+        makeAtomicFenceValue(*this, e, cir::MemScopeKind::System));
+  case Builtin::BI__atomic_signal_fence:
+    return RValue::get(
+        makeAtomicFenceValue(*this, e, cir::MemScopeKind::SingleThread));
+  case Builtin::BI__c11_atomic_thread_fence:
+  case Builtin::BI__c11_atomic_signal_fence:
+    llvm_unreachable("BI__c11_atomic_thread_fence like NYI");
   }
 
   // If this is an alias for a lib function (e.g. __builtin_sin), emit
diff --git a/clang/lib/CIR/Lowering/DirectToLLVM/LowerToLLVM.cpp b/clang/lib/CIR/Lowering/DirectToLLVM/LowerToLLVM.cpp
index d88a4ad76f27b..755a9b572f3ca 100644
--- a/clang/lib/CIR/Lowering/DirectToLLVM/LowerToLLVM.cpp
+++ b/clang/lib/CIR/Lowering/DirectToLLVM/LowerToLLVM.cpp
@@ -732,6 +732,14 @@ getLLVMMemOrder(std::optional<cir::MemOrder> memorder) {
   llvm_unreachable("unknown memory order");
 }
 
+static std::optional<llvm::StringRef>
+getLLVMSyncScope(std::optional<cir::MemScopeKind> syncScope) {
+  if (syncScope.has_value())
+    return syncScope.value() == cir::MemScopeKind::SingleThread ? "singlethread"
+                                                                : "";
+  return std::nullopt;
+}
+
 mlir::LogicalResult CIRToLLVMAtomicCmpXchgOpLowering::matchAndRewrite(
     cir::AtomicCmpXchgOp op, OpAdaptor adaptor,
     mlir::ConversionPatternRewriter &rewriter) const {
@@ -808,6 +816,19 @@ mlir::LogicalResult CIRToLLVMAtomicClearOpLowering::matchAndRewrite(
   return mlir::success();
 }
 
+mlir::LogicalResult CIRToLLVMAtomicFenceLowering::matchAndRewrite(
+    cir::AtomicFence op, OpAdaptor adaptor,
+    mlir::ConversionPatternRewriter &rewriter) const {
+  mlir::LLVM::AtomicOrdering llvmOrder = getLLVMMemOrder(adaptor.getOrdering());
+
+  auto fence = mlir::LLVM::FenceOp::create(rewriter, op.getLoc(), llvmOrder);
+  fence.setSyncscope(getLLVMSyncScope(adaptor.getSyncscope()));
+
+  rewriter.replaceOp(op, fence);
+
+  return mlir::success();
+}
+
 static mlir::LLVM::AtomicBinOp
 getLLVMAtomicBinOp(cir::AtomicFetchKind k, bool isInt, bool isSignedInt) {
   switch (k) {

[bcardosolopes]

Seems like you are referring to an old version of incubator, we use CIR_SyncScopeKind there, please make sure this PR reflects the most updated version here and elsewhere in the PR.

[github-actions]

🐧 Linux x64 Test Results

• 112073 tests passed
• 4077 tests skipped

[bcardosolopes]

Move this attribute closer to CIR_AtomicFetchKind