From 6d4afdfd8af7215fd44f0d2173dca801a3123e00 Mon Sep 17 00:00:00 2001
From: Evgeny Leviant <eleviant@accesssoftek.com>
Date: Thu, 30 Apr 2026 15:30:59 +0200
Subject: [PATCH] [clang] Make MS atomics volatile in kernel mode

This prevents LLVM backend from folding idempotent atomic instructions to a
memory barrier.
---
 clang/lib/CodeGen/CGBuiltin.cpp             |  1 +
 clang/test/CodeGen/MSKernel/interlocked.c   | 15 +++++++++++++++
 llvm/lib/CodeGen/AtomicExpandPass.cpp       |  3 ++-
 llvm/test/CodeGen/X86/volatile-atomicrmw.ll | 13 +++++++++++++
 4 files changed, 31 insertions(+), 1 deletion(-)
 create mode 100644 clang/test/CodeGen/MSKernel/interlocked.c
 create mode 100644 llvm/test/CodeGen/X86/volatile-atomicrmw.ll

diff --git a/clang/lib/CodeGen/CGBuiltin.cpp b/clang/lib/CodeGen/CGBuiltin.cpp
index a048a566a092..28bc7866250a 100644
--- a/clang/lib/CodeGen/CGBuiltin.cpp
+++ b/clang/lib/CodeGen/CGBuiltin.cpp
@@ -232,6 +232,7 @@ static Value *MakeBinaryAtomicValue(
 
   llvm::Value *Result =
       CGF.Builder.CreateAtomicRMW(Kind, DestAddr, Val, Ordering);
+  cast<llvm::AtomicRMWInst>(Result)->setVolatile(CGF.CGM.getLangOpts().Kernel);
   return EmitFromInt(CGF, Result, T, ValueType);
 }
 
diff --git a/clang/test/CodeGen/MSKernel/interlocked.c b/clang/test/CodeGen/MSKernel/interlocked.c
new file mode 100644
index 000000000000..2ce1012ffa43
--- /dev/null
+++ b/clang/test/CodeGen/MSKernel/interlocked.c
@@ -0,0 +1,15 @@
+// Check that we don't fold no-op andl to memory barrier
+// REQUIRES: x86-registered-target,aarch64-registered-target
+// RUN: %clang_cc1 -fms-kernel -fms-extensions -Wno-implicit-function-declaration  -triple x86_64-pc-win32 -O2 -S -o - %s | FileCheck %s --check-prefix=X86
+// RUN: %clang_cc1 -fms-kernel -fms-extensions -Wno-implicit-function-declaration  -triple aarch64-pc-win32 -O2 -S -o - %s | FileCheck %s --check-prefix=ARM64
+
+// X86:      lock andl $-1, (%rcx)
+// X86-NEXT: retq
+
+// ARM64:      ldaxr
+// ARM64-NEXT: stlxr
+// ARM64-NEXT: cbnz
+
+void access_via_interlocked(long volatile* addr) {
+    _InterlockedAnd(addr, (long)-1);
+}
diff --git a/llvm/lib/CodeGen/AtomicExpandPass.cpp b/llvm/lib/CodeGen/AtomicExpandPass.cpp
index 3d4e2cb196a1..14042097c317 100644
--- a/llvm/lib/CodeGen/AtomicExpandPass.cpp
+++ b/llvm/lib/CodeGen/AtomicExpandPass.cpp
@@ -1563,7 +1563,8 @@ bool AtomicExpandImpl::isIdempotentRMW(AtomicRMWInst *RMWI) {
   auto C = dyn_cast<ConstantInt>(RMWI->getValOperand());
   if (!C)
     return false;
-
+  if (RMWI->isVolatile())
+    return false;
   AtomicRMWInst::BinOp Op = RMWI->getOperation();
   switch (Op) {
   case AtomicRMWInst::Add:
diff --git a/llvm/test/CodeGen/X86/volatile-atomicrmw.ll b/llvm/test/CodeGen/X86/volatile-atomicrmw.ll
new file mode 100644
index 000000000000..dcdd8ea5f5e8
--- /dev/null
+++ b/llvm/test/CodeGen/X86/volatile-atomicrmw.ll
@@ -0,0 +1,13 @@
+; RUN: opt -S -passes='require<libcall-lowering-info>,expand-ir-insts,atomic-expand' %s -o - | FileCheck %s
+
+; volatile atomicrmw shouldn't be converted to a fence
+; CHECK:  %0 = atomicrmw volatile and ptr %addr, i32 -1 seq_cst
+
+target triple = "x86_64-pc-windows-msvc"
+
+define dso_local void @access_via_interlocked(ptr noundef %addr) {
+entry:
+  %0 = atomicrmw volatile and ptr %addr, i32 -1 seq_cst, align 4
+  ret void
+}
+