[clang] 0cb237d - [CIR] Add AtomicFenceOp and signal/thread fence builtins and required helpers (#168346)
via cfe-commits
cfe-commits at lists.llvm.org
Fri Dec 5 14:12:25 PST 2025
Author: Hendrik Hübner
Date: 2025-12-05T23:12:21+01:00
New Revision: 0cb237d69e16a8c5d234a5c5ab1b1ebcf6f94cd3
URL: https://github.com/llvm/llvm-project/commit/0cb237d69e16a8c5d234a5c5ab1b1ebcf6f94cd3
DIFF: https://github.com/llvm/llvm-project/commit/0cb237d69e16a8c5d234a5c5ab1b1ebcf6f94cd3.diff
LOG: [CIR] Add AtomicFenceOp and signal/thread fence builtins and required helpers (#168346)
This PR adds the AtomicFenceOp and signal/thread fence builtins.
Added:
clang/test/CIR/CodeGen/atomic-thread-fence.c
Modified:
clang/include/clang/CIR/Dialect/IR/CIROps.td
clang/lib/CIR/CodeGen/CIRGenBuiltin.cpp
clang/lib/CIR/Lowering/DirectToLLVM/LowerToLLVM.cpp
Removed:
################################################################################
diff --git a/clang/include/clang/CIR/Dialect/IR/CIROps.td b/clang/include/clang/CIR/Dialect/IR/CIROps.td
index 80c184e4f478c..bb89d3cd4592a 100644
--- a/clang/include/clang/CIR/Dialect/IR/CIROps.td
+++ b/clang/include/clang/CIR/Dialect/IR/CIROps.td
@@ -5188,6 +5188,11 @@ def CIR_AtomicFetchKind : CIR_I32EnumAttr<
I32EnumAttrCase<"Min", 7, "min">
]>;
+def CIR_SyncScopeKind : CIR_I32EnumAttr<"SyncScopeKind", "sync scope kind", [
+ I32EnumAttrCase<"SingleThread", 0, "single_thread">,
+ I32EnumAttrCase<"System", 1, "system">
+]>;
+
def CIR_AtomicFetchOp : CIR_Op<"atomic.fetch", [
AllTypesMatch<["result", "val"]>,
TypesMatchWith<"type of 'val' must match the pointee type of 'ptr'",
@@ -5411,6 +5416,36 @@ def CIR_AtomicClearOp : CIR_Op<"atomic.clear"> {
}];
}
+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">:$ordering,
+ OptionalAttr<CIR_SyncScopeKind>:$syncscope
+ );
+
+ let assemblyFormat = [{
+ (`syncscope` `(` $syncscope^ `)`)? $ordering attr-dict
+ }];
+}
+
//===----------------------------------------------------------------------===//
// BlockAddressOp
//===----------------------------------------------------------------------===//
diff --git a/clang/lib/CIR/CodeGen/CIRGenBuiltin.cpp b/clang/lib/CIR/CodeGen/CIRGenBuiltin.cpp
index 72495270b11ed..16c006df6853e 100644
--- a/clang/lib/CIR/CodeGen/CIRGenBuiltin.cpp
+++ b/clang/lib/CIR/CodeGen/CIRGenBuiltin.cpp
@@ -60,6 +60,33 @@ static RValue emitBuiltinBitOp(CIRGenFunction &cgf, const CallExpr *e,
return RValue::get(result);
}
+static mlir::Value makeAtomicFenceValue(CIRGenFunction &cgf,
+ const CallExpr *expr,
+ cir::SyncScopeKind syncScope) {
+ CIRGenBuilderTy &builder = cgf.getBuilder();
+ mlir::Value orderingVal = cgf.emitScalarExpr(expr->getArg(0));
+
+ auto constOrdering = orderingVal.getDefiningOp<cir::ConstantOp>();
+
+ if (!constOrdering) {
+ // TODO(cir): Emit code to switch on `orderingVal`,
+ // and creating the fence op for valid values.
+ cgf.cgm.errorNYI("Variable atomic fence ordering");
+ return {};
+ }
+
+ auto constOrderingAttr = constOrdering.getValueAttr<cir::IntAttr>();
+ assert(constOrderingAttr && "Expected integer constant for ordering");
+
+ auto ordering = static_cast<cir::MemOrder>(constOrderingAttr.getUInt());
+
+ cir::AtomicFence::create(
+ builder, cgf.getLoc(expr->getSourceRange()), ordering,
+ cir::SyncScopeKindAttr::get(&cgf.getMLIRContext(), syncScope));
+
+ return {};
+}
+
namespace {
struct WidthAndSignedness {
unsigned width;
@@ -982,8 +1009,13 @@ RValue CIRGenFunction::emitBuiltinExpr(const GlobalDecl &gd, unsigned builtinID,
case Builtin::BI__atomic_is_lock_free:
case Builtin::BI__atomic_test_and_set:
case Builtin::BI__atomic_clear:
+ return errorBuiltinNYI(*this, e, builtinID);
case Builtin::BI__atomic_thread_fence:
+ return RValue::get(
+ makeAtomicFenceValue(*this, e, cir::SyncScopeKind::System));
case Builtin::BI__atomic_signal_fence:
+ return RValue::get(
+ makeAtomicFenceValue(*this, e, cir::SyncScopeKind::SingleThread));
case Builtin::BI__c11_atomic_thread_fence:
case Builtin::BI__c11_atomic_signal_fence:
case Builtin::BI__scoped_atomic_thread_fence:
diff --git a/clang/lib/CIR/Lowering/DirectToLLVM/LowerToLLVM.cpp b/clang/lib/CIR/Lowering/DirectToLLVM/LowerToLLVM.cpp
index 40e14474890dc..5ef0c73a99a4d 100644
--- a/clang/lib/CIR/Lowering/DirectToLLVM/LowerToLLVM.cpp
+++ b/clang/lib/CIR/Lowering/DirectToLLVM/LowerToLLVM.cpp
@@ -775,6 +775,15 @@ getLLVMMemOrder(std::optional<cir::MemOrder> memorder) {
llvm_unreachable("unknown memory order");
}
+static std::optional<llvm::StringRef>
+getLLVMSyncScope(std::optional<cir::SyncScopeKind> syncScope) {
+ if (syncScope.has_value())
+ return syncScope.value() == cir::SyncScopeKind::SingleThread
+ ? "singlethread"
+ : "";
+ return std::nullopt;
+}
+
mlir::LogicalResult CIRToLLVMAtomicCmpXchgOpLowering::matchAndRewrite(
cir::AtomicCmpXchgOp op, OpAdaptor adaptor,
mlir::ConversionPatternRewriter &rewriter) const {
@@ -851,6 +860,19 @@ 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) {
diff --git a/clang/test/CIR/CodeGen/atomic-thread-fence.c b/clang/test/CIR/CodeGen/atomic-thread-fence.c
new file mode 100644
index 0000000000000..f28bc6808cbfa
--- /dev/null
+++ b/clang/test/CIR/CodeGen/atomic-thread-fence.c
@@ -0,0 +1,181 @@
+// RUN: %clang_cc1 -triple aarch64-none-linux-android21 -fclangir -emit-cir %s -o %t.cir
+// RUN: FileCheck --check-prefix=CIR --input-file=%t.cir %s
+// RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -fclangir -emit-llvm %s -o %t.ll
+// RUN: FileCheck --check-prefix=LLVM --input-file=%t.ll %s
+// RUN: %clang_cc1 -triple aarch64-none-linux-android21 -emit-llvm %s -o %t.ll
+// RUN: FileCheck --check-prefix=OGCG --input-file=%t.ll %s
+
+struct Data {
+ int value;
+ void *ptr;
+};
+
+typedef struct Data *DataPtr;
+
+void applyThreadFence() {
+ __atomic_thread_fence(__ATOMIC_SEQ_CST);
+ // CIR-LABEL: @applyThreadFence
+ // CIR: cir.atomic.fence syncscope(system) seq_cst
+ // CIR: cir.return
+
+ // LLVM-LABEL: @applyThreadFence
+ // LLVM: fence seq_cst
+ // LLVM: ret void
+
+ // OGCG-LABEL: @applyThreadFence
+ // OGCG: fence seq_cst
+ // OGCG: ret void
+}
+
+void applySignalFence() {
+ __atomic_signal_fence(__ATOMIC_SEQ_CST);
+ // CIR-LABEL: @applySignalFence
+ // CIR: cir.atomic.fence syncscope(single_thread) seq_cst
+ // CIR: cir.return
+
+ // LLVM-LABEL: @applySignalFence
+ // LLVM: fence syncscope("singlethread") seq_cst
+ // LLVM: ret void
+
+ // OGCG-LABEL: @applySignalFence
+ // OGCG: fence syncscope("singlethread") seq_cst
+ // OGCG: ret void
+}
+
+void modifyWithThreadFence(DataPtr d) {
+ __atomic_thread_fence(__ATOMIC_SEQ_CST);
+ d->value = 42;
+ // CIR-LABEL: @modifyWithThreadFence
+ // CIR: %[[DATA:.*]] = cir.alloca !cir.ptr<!rec_Data>, !cir.ptr<!cir.ptr<!rec_Data>>, ["d", init] {alignment = 8 : i64}
+ // CIR: cir.atomic.fence syncscope(system) seq_cst
+ // CIR: %[[VAL_42:.*]] = cir.const #cir.int<42> : !s32i
+ // CIR: %[[LOAD_DATA:.*]] = cir.load{{.*}} %[[DATA]] : !cir.ptr<!cir.ptr<!rec_Data>>, !cir.ptr<!rec_Data>
+ // CIR: %[[DATA_VALUE:.*]] = cir.get_member %[[LOAD_DATA]][0] {name = "value"} : !cir.ptr<!rec_Data> -> !cir.ptr<!s32i>
+ // CIR: cir.store{{.*}} %[[VAL_42]], %[[DATA_VALUE]] : !s32i, !cir.ptr<!s32i>
+ // CIR: cir.return
+
+ // LLVM-LABEL: @modifyWithThreadFence
+ // LLVM: %[[DATA:.*]] = alloca ptr, i64 1, align 8
+ // LLVM: fence seq_cst
+ // LLVM: %[[DATA_PTR:.*]] = load ptr, ptr %[[DATA]], align 8
+ // LLVM: %[[DATA_VALUE:.*]] = getelementptr %struct.Data, ptr %[[DATA_PTR]], i32 0, i32 0
+ // LLVM: store i32 42, ptr %[[DATA_VALUE]], align 8
+ // LLVM: ret void
+
+ // OGCG-LABEL: @modifyWithThreadFence
+ // OGCG: %[[DATA:.*]] = alloca ptr, align 8
+ // OGCG: fence seq_cst
+ // OGCG: %[[DATA_PTR:.*]] = load ptr, ptr %[[DATA]], align 8
+ // OGCG: %[[DATA_VALUE:.*]] = getelementptr inbounds nuw %struct.Data, ptr %[[DATA_PTR]], i32 0, i32 0
+ // OGCG: store i32 42, ptr %[[DATA_VALUE]], align 8
+ // OGCG: ret void
+}
+
+void modifyWithSignalFence(DataPtr d) {
+ __atomic_signal_fence(__ATOMIC_SEQ_CST);
+ d->value = 24;
+ // CIR-LABEL: @modifyWithSignalFence
+ // CIR: %[[DATA:.*]] = cir.alloca !cir.ptr<!rec_Data>, !cir.ptr<!cir.ptr<!rec_Data>>, ["d", init] {alignment = 8 : i64}
+ // CIR: cir.atomic.fence syncscope(single_thread) seq_cst
+ // CIR: %[[VAL_42:.*]] = cir.const #cir.int<24> : !s32i
+ // CIR: %[[LOAD_DATA:.*]] = cir.load{{.*}} %[[DATA]] : !cir.ptr<!cir.ptr<!rec_Data>>, !cir.ptr<!rec_Data>
+ // CIR: %[[DATA_VALUE:.*]] = cir.get_member %[[LOAD_DATA]][0] {name = "value"} : !cir.ptr<!rec_Data> -> !cir.ptr<!s32i>
+ // CIR: cir.store{{.*}} %[[VAL_42]], %[[DATA_VALUE]] : !s32i, !cir.ptr<!s32i>
+ // CIR: cir.return
+
+ // LLVM-LABEL: @modifyWithSignalFence
+ // LLVM: %[[DATA:.*]] = alloca ptr, i64 1, align 8
+ // LLVM: fence syncscope("singlethread") seq_cst
+ // LLVM: %[[DATA_PTR:.*]] = load ptr, ptr %[[DATA]], align 8
+ // LLVM: %[[DATA_VALUE:.*]] = getelementptr %struct.Data, ptr %[[DATA_PTR]], i32 0, i32 0
+ // LLVM: store i32 24, ptr %[[DATA_VALUE]], align 8
+ // LLVM: ret void
+
+ // OGCG-LABEL: @modifyWithSignalFence
+ // OGCG: %[[DATA:.*]] = alloca ptr, align 8
+ // OGCG: fence syncscope("singlethread") seq_cst
+ // OGCG: %[[DATA_PTR:.*]] = load ptr, ptr %[[DATA]], align 8
+ // OGCG: %[[DATA_VALUE:.*]] = getelementptr inbounds nuw %struct.Data, ptr %[[DATA_PTR]], i32 0, i32 0
+ // OGCG: store i32 24, ptr %[[DATA_VALUE]], align 8
+ // OGCG: ret void
+}
+
+void loadWithThreadFence(DataPtr d) {
+ __atomic_thread_fence(__ATOMIC_SEQ_CST);
+ __atomic_load_n(&d->ptr, __ATOMIC_SEQ_CST);
+ // CIR-LABEL: @loadWithThreadFence
+ // CIR: %[[DATA:.*]] = cir.alloca !cir.ptr<!rec_Data>, !cir.ptr<!cir.ptr<!rec_Data>>, ["d", init] {alignment = 8 : i64}
+ // CIR: %[[ATOMIC_TEMP:.*]] = cir.alloca !cir.ptr<!void>, !cir.ptr<!cir.ptr<!void>>, ["atomic-temp"] {alignment = 8 : i64}
+ // CIR: cir.atomic.fence syncscope(system) seq_cst
+ // CIR: %[[LOAD_DATA:.*]] = cir.load{{.*}} %[[DATA]] : !cir.ptr<!cir.ptr<!rec_Data>>, !cir.ptr<!rec_Data>
+ // CIR: %[[DATA_VALUE:.*]] = cir.get_member %[[LOAD_DATA]][1] {name = "ptr"} : !cir.ptr<!rec_Data> -> !cir.ptr<!cir.ptr<!void>>
+ // CIR: %[[CASTED_DATA_VALUE:.*]] = cir.cast bitcast %[[DATA_VALUE]] : !cir.ptr<!cir.ptr<!void>> -> !cir.ptr<!u64i>
+ // CIR: %[[CASTED_ATOMIC_TEMP:.*]] = cir.cast bitcast %[[ATOMIC_TEMP]] : !cir.ptr<!cir.ptr<!void>> -> !cir.ptr<!u64i>
+ // CIR: %[[ATOMIC_LOAD:.*]] = cir.load{{.*}} atomic(seq_cst) %[[CASTED_DATA_VALUE]] : !cir.ptr<!u64i>, !u64i
+ // CIR: cir.store{{.*}} %[[ATOMIC_LOAD]], %[[CASTED_ATOMIC_TEMP]] : !u64i, !cir.ptr<!u64i>
+ // CIR: %[[DOUBLE_CASTED_ATOMIC_TEMP:.*]] = cir.cast bitcast %[[CASTED_ATOMIC_TEMP]] : !cir.ptr<!u64i> -> !cir.ptr<!cir.ptr<!void>>
+ // CIR: %[[ATOMIC_LOAD_PTR:.*]] = cir.load{{.*}} %[[DOUBLE_CASTED_ATOMIC_TEMP]] : !cir.ptr<!cir.ptr<!void>>, !cir.ptr<!void>
+ // CIR: cir.return
+
+ // LLVM-LABEL: @loadWithThreadFence
+ // LLVM: %[[DATA:.*]] = alloca ptr, i64 1, align 8
+ // LLVM: %[[DATA_TEMP:.*]] = alloca ptr, i64 1, align 8
+ // LLVM: fence seq_cst
+ // LLVM: %[[DATA_PTR:.*]] = load ptr, ptr %[[DATA]], align 8
+ // LLVM: %[[DATA_VALUE:.*]] = getelementptr %struct.Data, ptr %[[DATA_PTR]], i32 0, i32 1
+ // LLVM: %[[ATOMIC_LOAD:.*]] = load atomic i64, ptr %[[DATA_VALUE]] seq_cst, align 8
+ // LLVM: store i64 %[[ATOMIC_LOAD]], ptr %[[DATA_TEMP]], align 8
+ // LLVM: %[[DATA_TEMP_LOAD:.*]] = load ptr, ptr %[[DATA_TEMP]], align 8
+ // LLVM: ret void
+
+ // OGCG-LABEL: @loadWithThreadFence
+ // OGCG: %[[DATA:.*]] = alloca ptr, align 8
+ // OGCG: %[[DATA_TEMP:.*]] = alloca ptr, align 8
+ // OGCG: fence seq_cst
+ // OGCG: %[[DATA_PTR:.*]] = load ptr, ptr %[[DATA]], align 8
+ // OGCG: %[[DATA_VALUE:.*]] = getelementptr inbounds nuw %struct.Data, ptr %[[DATA_PTR]], i32 0, i32 1
+ // OGCG: %[[ATOMIC_LOAD:.*]] = load atomic i64, ptr %[[DATA_VALUE]] seq_cst, align 8
+ // OGCG: store i64 %[[ATOMIC_LOAD]], ptr %[[DATA_TEMP]], align 8
+ // OGCG: %[[DATA_TEMP_LOAD:.*]] = load ptr, ptr %[[DATA_TEMP]], align 8
+ // OGCG: ret void
+}
+
+void loadWithSignalFence(DataPtr d) {
+ __atomic_signal_fence(__ATOMIC_SEQ_CST);
+ __atomic_load_n(&d->ptr, __ATOMIC_SEQ_CST);
+ // CIR-LABEL: @loadWithSignalFence
+ // CIR: %[[DATA:.*]] = cir.alloca !cir.ptr<!rec_Data>, !cir.ptr<!cir.ptr<!rec_Data>>, ["d", init] {alignment = 8 : i64}
+ // CIR: %[[ATOMIC_TEMP:.*]] = cir.alloca !cir.ptr<!void>, !cir.ptr<!cir.ptr<!void>>, ["atomic-temp"] {alignment = 8 : i64}
+ // CIR: cir.atomic.fence syncscope(single_thread) seq_cst
+ // CIR: %[[LOAD_DATA:.*]] = cir.load{{.*}} %[[DATA]] : !cir.ptr<!cir.ptr<!rec_Data>>, !cir.ptr<!rec_Data>
+ // CIR: %[[DATA_PTR:.*]] = cir.get_member %[[LOAD_DATA]][1] {name = "ptr"} : !cir.ptr<!rec_Data> -> !cir.ptr<!cir.ptr<!void>>
+ // CIR: %[[CASTED_DATA_PTR:.*]] = cir.cast bitcast %[[DATA_PTR]] : !cir.ptr<!cir.ptr<!void>> -> !cir.ptr<!u64i>
+ // CIR: %[[CASTED_ATOMIC_TEMP:.*]] = cir.cast bitcast %[[ATOMIC_TEMP]] : !cir.ptr<!cir.ptr<!void>> -> !cir.ptr<!u64i>
+ // CIR: %[[ATOMIC_LOAD:.*]] = cir.load{{.*}} atomic(seq_cst) %[[CASTED_DATA_PTR]] : !cir.ptr<!u64i>, !u64i
+ // CIR: cir.store{{.*}} %[[ATOMIC_LOAD]], %[[CASTED_ATOMIC_TEMP]] : !u64i, !cir.ptr<!u64i>
+ // CIR: %[[DOUBLE_CASTED_ATOMIC_TEMP:.*]] = cir.cast bitcast %[[CASTED_ATOMIC_TEMP]] : !cir.ptr<!u64i> -> !cir.ptr<!cir.ptr<!void>>
+ // CIR: %[[LOAD_ATOMIC_TEMP:.*]] = cir.load{{.*}} %[[DOUBLE_CASTED_ATOMIC_TEMP]] : !cir.ptr<!cir.ptr<!void>>, !cir.ptr<!void>
+ // CIR: cir.return
+
+ // LLVM-LABEL: @loadWithSignalFence
+ // LLVM: %[[DATA:.*]] = alloca ptr, i64 1, align 8
+ // LLVM: %[[DATA_TEMP:.*]] = alloca ptr, i64 1, align 8
+ // LLVM: fence syncscope("singlethread") seq_cst
+ // LLVM: %[[DATA_PTR:.*]] = load ptr, ptr %[[DATA]], align 8
+ // LLVM: %[[DATA_VALUE:.*]] = getelementptr %struct.Data, ptr %[[DATA_PTR]], i32 0, i32 1
+ // LLVM: %[[ATOMIC_LOAD:.*]] = load atomic i64, ptr %[[DATA_VALUE]] seq_cst, align 8
+ // LLVM: store i64 %[[ATOMIC_LOAD]], ptr %[[DATA_TEMP]], align 8
+ // LLVM: %[[DATA_TEMP_LOAD]] = load ptr, ptr %[[DATA_TEMP]], align 8
+ // LLVM: ret void
+
+ // OGCG-LABEL: @loadWithSignalFence
+ // OGCG: %[[DATA:.*]] = alloca ptr, align 8
+ // OGCG: %[[DATA_TEMP:.*]] = alloca ptr, align 8
+ // OGCG: fence syncscope("singlethread") seq_cst
+ // OGCG: %[[DATA_PTR:.*]] = load ptr, ptr %[[DATA]], align 8
+ // OGCG: %[[DATA_VALUE:.*]] = getelementptr inbounds nuw %struct.Data, ptr %[[DATA_PTR]], i32 0, i32 1
+ // OGCG: %[[ATOMIC_LOAD:.*]] = load atomic i64, ptr %[[DATA_VALUE]] seq_cst, align 8
+ // OGCG: store i64 %[[ATOMIC_LOAD]], ptr %[[DATA_TEMP]], align 8
+ // OGCG: %[[DATA_TEMP_LOAD]] = load ptr, ptr %[[DATA_TEMP]], align 8
+ // OGCG: ret void
+}
More information about the cfe-commits
mailing list