[mlir][gpu] Eliminate redundant gpu.barrier ops

Author: Eric Eaton

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", [

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 eraseRedundantGpuBarrierOps(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(eraseRedundantGpuBarrierOps);
+}
+
 //===----------------------------------------------------------------------===//
 // GPUFuncOp
 //===----------------------------------------------------------------------===//

mlir/test/Dialect/GPU/canonicalize.mlir

@@ -11,6 +11,16 @@ func.func @fold_wait_op_test1() {
 }
 // CHECK-NOT: gpu.wait
 
+// Erase duplicate barriers.
+// CHECK-LABEL: func @erase_barriers
+//       CHECK-NEXT: gpu.barrier
+//       CHECK-NEXT: return
+func.func @erase_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>) {