[Mlir-commits] [mlir] b4d7d3f - [mlir][NVVM] Add nvvm.membar operation (#166698)

[llvmlistbot at llvm.org]

Author: Valentin Clement (バレンタイン クレメン)
Date: 2025-11-07T10:39:01-08:00
New Revision: b4d7d3f745aa7647cd7085b3a03ba15e7c21aa8b

URL: https://github.com/llvm/llvm-project/commit/b4d7d3f745aa7647cd7085b3a03ba15e7c21aa8b
DIFF: https://github.com/llvm/llvm-project/commit/b4d7d3f745aa7647cd7085b3a03ba15e7c21aa8b.diff

LOG: [mlir][NVVM] Add nvvm.membar operation (#166698)

Add nvvm.membar operation with level as defined in
https://docs.nvidia.com/cuda/parallel-thread-execution/#parallel-synchronization-and-communication-instructions-membar

This will be used to replace direct intrinsic call in CUDA Fortran for
`threadfence()`, `threadfence_block` and `thread fence_system()`
currently lowered here:
https://github.com/llvm/llvm-project/blob/e700f157026bf8b4d58f936c5db8f152e269d77f/flang/lib/Optimizer/Builder/CUDAIntrinsicCall.cpp#L1310

The nvvm membar intrsinsic are also used in CUDA C/C++
(https://github.com/llvm/llvm-project/blob/49f55f4991227f3c7a2b8161bbf45c74b7023944/clang/lib/Headers/__clang_cuda_device_functions.h#L528)

Added: 
    mlir/test/Target/LLVMIR/nvvm/membar.mlir

Modified: 
    mlir/include/mlir/Dialect/LLVMIR/NVVMOps.td
    mlir/lib/Target/LLVMIR/Dialect/NVVM/NVVMToLLVMIRTranslation.cpp

Removed: 
    


################################################################################
diff  --git a/mlir/include/mlir/Dialect/LLVMIR/NVVMOps.td b/mlir/include/mlir/Dialect/LLVMIR/NVVMOps.td
index 80bc0e5986e51..1cc5b74a3cb67 100644
--- a/mlir/include/mlir/Dialect/LLVMIR/NVVMOps.td
+++ b/mlir/include/mlir/Dialect/LLVMIR/NVVMOps.td
@@ -1236,6 +1236,23 @@ def NVVM_FenceProxyAcquireOp : NVVM_Op<"fence.proxy.acquire">,
   let hasVerifier = 1;
 }
 
+def NVVM_MembarOp : NVVM_Op<"memory.barrier">,
+                    Arguments<(ins MemScopeKindAttr:$scope)> {
+  let summary = "Memory barrier operation";
+  let description = [{
+    `membar` operation guarantees that prior memory accesses requested by this
+    thread are performed at the specified `scope`, before later memory
+    operations requested by this thread following the membar instruction.
+
+    [For more information, see PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#parallel-synchronization-and-communication-instructions-membar)
+  }];
+
+  let assemblyFormat = "$scope attr-dict";
+  let llvmBuilder = [{
+    createIntrinsicCall(builder, getMembarIntrinsicID($scope), {});
+  }];
+}
+
 def NVVM_FenceProxyReleaseOp : NVVM_Op<"fence.proxy.release">,
       Arguments<(ins MemScopeKindAttr:$scope,
                      DefaultValuedAttr<ProxyKindAttr,

diff  --git a/mlir/lib/Target/LLVMIR/Dialect/NVVM/NVVMToLLVMIRTranslation.cpp b/mlir/lib/Target/LLVMIR/Dialect/NVVM/NVVMToLLVMIRTranslation.cpp
index 0964e1b8c5ef3..cecff51e637a5 100644
--- a/mlir/lib/Target/LLVMIR/Dialect/NVVM/NVVMToLLVMIRTranslation.cpp
+++ b/mlir/lib/Target/LLVMIR/Dialect/NVVM/NVVMToLLVMIRTranslation.cpp
@@ -291,6 +291,20 @@ static unsigned getUnidirectionalFenceProxyID(NVVM::ProxyKind fromProxy,
   llvm_unreachable("Unsupported proxy kinds");
 }
 
+static unsigned getMembarIntrinsicID(NVVM::MemScopeKind scope) {
+  switch (scope) {
+  case NVVM::MemScopeKind::CTA:
+    return llvm::Intrinsic::nvvm_membar_cta;
+  case NVVM::MemScopeKind::CLUSTER:
+    return llvm::Intrinsic::nvvm_fence_sc_cluster;
+  case NVVM::MemScopeKind::GPU:
+    return llvm::Intrinsic::nvvm_membar_gl;
+  case NVVM::MemScopeKind::SYS:
+    return llvm::Intrinsic::nvvm_membar_sys;
+  }
+  llvm_unreachable("Unknown scope for memory barrier");
+}
+
 #define TCGEN05LD(SHAPE, NUM) llvm::Intrinsic::nvvm_tcgen05_ld_##SHAPE##_##NUM
 
 static llvm::Intrinsic::ID

diff  --git a/mlir/test/Target/LLVMIR/nvvm/membar.mlir b/mlir/test/Target/LLVMIR/nvvm/membar.mlir
new file mode 100644
index 0000000000000..1b794f663b573
--- /dev/null
+++ b/mlir/test/Target/LLVMIR/nvvm/membar.mlir
@@ -0,0 +1,14 @@
+// RUN: mlir-translate -mlir-to-llvmir %s  -split-input-file --verify-diagnostics | FileCheck %s
+
+// CHECK-lABEL: @memorybarrier()
+llvm.func @memorybarrier() {
+  // CHECK: call void @llvm.nvvm.membar.cta()
+  nvvm.memory.barrier #nvvm.mem_scope<cta>
+  // CHECK: call void @llvm.nvvm.fence.sc.cluster()
+  nvvm.memory.barrier #nvvm.mem_scope<cluster>
+  // CHECK: call void @llvm.nvvm.membar.gl()
+  nvvm.memory.barrier #nvvm.mem_scope<gpu>
+  // CHECK: call void @llvm.nvvm.membar.sys()
+  nvvm.memory.barrier #nvvm.mem_scope<sys>
+  llvm.return
+}