[CIR] Implement __sync_synchronize builtin

[erichkeane]

This showed up on a spec test, but is a very simple system-sequentially consistent fence instruction.

[llvmorg-github-actions]

@llvm/pr-subscribers-clang

@llvm/pr-subscribers-clangir

Author: Erich Keane (erichkeane)

Changes

This showed up on a spec test, but is a very simple system-sequentially consistent fence instruction.

---

Full diff: https://github.com/llvm/llvm-project/pull/200423.diff

2 Files Affected:

• (modified) clang/lib/CIR/CodeGen/CIRGenBuiltin.cpp (+8-1)
• (added) clang/test/CIR/CodeGenBuiltins/builtin-sync_synchronize.c (+9)

diff --git a/clang/lib/CIR/CodeGen/CIRGenBuiltin.cpp b/clang/lib/CIR/CodeGen/CIRGenBuiltin.cpp
index afa7e5b91251b..bfe29a3928961 100644
--- a/clang/lib/CIR/CodeGen/CIRGenBuiltin.cpp
+++ b/clang/lib/CIR/CodeGen/CIRGenBuiltin.cpp
@@ -2089,7 +2089,6 @@ RValue CIRGenFunction::emitBuiltinExpr(const GlobalDecl &gd, unsigned builtinID,
   case Builtin::BI__sync_lock_release_4:
   case Builtin::BI__sync_lock_release_8:
   case Builtin::BI__sync_lock_release_16:
-  case Builtin::BI__sync_synchronize:
   case Builtin::BI__builtin_nontemporal_load:
   case Builtin::BI__builtin_nontemporal_store:
   case Builtin::BI__c11_atomic_is_lock_free:
@@ -2107,6 +2106,14 @@ RValue CIRGenFunction::emitBuiltinExpr(const GlobalDecl &gd, unsigned builtinID,
     emitAtomicFenceOp(*this, e, cir::SyncScopeKind::SingleThread);
     return RValue::get(nullptr);
   }
+  case Builtin::BI__sync_synchronize: {
+    cir::AtomicFenceOp::create(
+        builder, getLoc(e->getSourceRange()),
+        cir::MemOrder::SequentiallyConsistent,
+        cir::SyncScopeKindAttr::get(&getMLIRContext(),
+                                    cir::SyncScopeKind::System));
+    return RValue::get(nullptr);
+  }
   case Builtin::BI__scoped_atomic_thread_fence:
     return errorBuiltinNYI(*this, e, builtinID);
   case Builtin::BI__builtin_signbit:
diff --git a/clang/test/CIR/CodeGenBuiltins/builtin-sync_synchronize.c b/clang/test/CIR/CodeGenBuiltins/builtin-sync_synchronize.c
new file mode 100644
index 0000000000000..e4047d44124a4
--- /dev/null
+++ b/clang/test/CIR/CodeGenBuiltins/builtin-sync_synchronize.c
@@ -0,0 +1,9 @@
+// RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -emit-cir %s -o - | FileCheck %s -check-prefix=CIR
+// RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -fclangir -emit-llvm %s -o - | FileCheck %s -check-prefix=LLVM
+// RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -emit-llvm %s -o - | FileCheck %s -check-prefix=LLVM
+
+void use() {
+  __sync_synchronize();
+  // CIR: cir.atomic.fence syncscope(system) seq_cst
+  // LLVM: fence seq_cst
+}

[andykaylor]

lgtm, with a nit about moving the case

[andykaylor]

We've been trying to keep these in the same order that they have in classic codegen. Can you leave it where it was and just introduce a new errorNYI call above it?

[andykaylor]

You didn't think it's worth copying the comment from classic codegen about how badly designed this intrinsic is?