[Mlir-commits] [mlir] [mlir][gpu] Eliminate redundant gpu.barrier ops (PR #71575)

[llvmlistbot at llvm.org]

https://github.com/spaceotter updated https://github.com/llvm/llvm-project/pull/71575

>From 00886128c550fac77abff37db6a04c633754a6c2 Mon Sep 17 00:00:00 2001
From: Eric Eaton <eric at nod-labs.com>
Date: Tue, 7 Nov 2023 11:07:23 -0800
Subject: [PATCH] [mlir][gpu] Eliminate redundant gpu.barrier ops

---
 mlir/include/mlir/Dialect/GPU/IR/GPUOps.td |  1 +
 mlir/lib/Dialect/GPU/IR/GPUDialect.cpp     | 23 ++++++++++++++++++++++
 mlir/test/Dialect/GPU/canonicalize.mlir    | 10 ++++++++++
 3 files changed, 34 insertions(+)

diff --git a/mlir/include/mlir/Dialect/GPU/IR/GPUOps.td b/mlir/include/mlir/Dialect/GPU/IR/GPUOps.td
index 6375d35f4311295..632cdd96c6d4c2b 100644
--- a/mlir/include/mlir/Dialect/GPU/IR/GPUOps.td
+++ b/mlir/include/mlir/Dialect/GPU/IR/GPUOps.td
@@ -1010,6 +1010,7 @@ def GPU_BarrierOp : GPU_Op<"barrier"> {
     in convergence.
   }];
   let assemblyFormat = "attr-dict";
+  let hasCanonicalizer = 1;
 }
 
 def GPU_GPUModuleOp : GPU_Op<"module", [
diff --git a/mlir/lib/Dialect/GPU/IR/GPUDialect.cpp b/mlir/lib/Dialect/GPU/IR/GPUDialect.cpp
index 5eb2cadc884e151..7f5b70a4ff084d6 100644
--- a/mlir/lib/Dialect/GPU/IR/GPUDialect.cpp
+++ b/mlir/lib/Dialect/GPU/IR/GPUDialect.cpp
@@ -1139,6 +1139,29 @@ void ShuffleOp::build(OpBuilder &builder, OperationState &result, Value value,
         mode);
 }
 
+//===----------------------------------------------------------------------===//
+// BarrierOp
+//===----------------------------------------------------------------------===//
+
+namespace {
+
+/// Remove gpu.barrier after gpu.barrier, the threads are already synchronized!
+LogicalResult foldRedundantGpuBarrierOps(BarrierOp op,
+                                         PatternRewriter &rewriter) {
+  if (isa_and_nonnull<BarrierOp>(op->getNextNode())) {
+    rewriter.eraseOp(op);
+    return success();
+  }
+  return failure();
+}
+
+} // end anonymous namespace
+
+void BarrierOp::getCanonicalizationPatterns(RewritePatternSet &results,
+                                            MLIRContext *context) {
+  results.add(foldRedundantGpuBarrierOps);
+}
+
 //===----------------------------------------------------------------------===//
 // GPUFuncOp
 //===----------------------------------------------------------------------===//
diff --git a/mlir/test/Dialect/GPU/canonicalize.mlir b/mlir/test/Dialect/GPU/canonicalize.mlir
index 1cf582e3822dae5..b370368e7faded2 100644
--- a/mlir/test/Dialect/GPU/canonicalize.mlir
+++ b/mlir/test/Dialect/GPU/canonicalize.mlir
@@ -11,6 +11,16 @@ func.func @fold_wait_op_test1() {
 }
 // CHECK-NOT: gpu.wait
 
+// Fold duplicate barriers.
+// CHECK-LABEL: func @fold_barriers
+//       CHECK-NEXT: gpu.barrier
+//       CHECK-NEXT: return
+func.func @fold_barriers() {
+  gpu.barrier
+  gpu.barrier
+  return
+}
+
 // Replace uses of gpu.wait op with its async dependency.
 // CHECK-LABEL: func @fold_wait_op_test2
 func.func @fold_wait_op_test2(%arg0: i1) -> (memref<5xf16>, memref<5xf16>) {