InsertSync: ping/pong multibuffer support

include/PTO/IR/PTOOps.td

@@ -1165,6 +1165,32 @@ def WaitFlagOp : PTO_Op<"wait_flag"> {
   }];
 }
 
+def SetFlagDynOp : PTO_Op<"set_flag_dyn"> {
+  let summary = "Set synchronization flag between pipes (dynamic event id)";
+  let arguments = (ins
+    PTO_PipeAttr:$src_pipe,
+    PTO_PipeAttr:$dst_pipe,
+    Index:$event_id
+  );
+  let results = (outs);
+  let assemblyFormat = [{
+    `[` $src_pipe `,` $dst_pipe `,` $event_id `]` attr-dict
+  }];
+}
+
+def WaitFlagDynOp : PTO_Op<"wait_flag_dyn"> {
+  let summary = "Wait for synchronization flag (dynamic event id)";
+  let arguments = (ins
+    PTO_PipeAttr:$src_pipe,
+    PTO_PipeAttr:$dst_pipe,
+    Index:$event_id
+  );
+  let results = (outs);
+  let assemblyFormat = [{
+    `[` $src_pipe `,` $dst_pipe `,` $event_id `]` attr-dict
+  }];
+}
+
 //===----------------------------------------------------------------------===//
 // Buffer-ID Synchronization (A5)
 //===----------------------------------------------------------------------===//

include/PTO/Transforms/InsertSync/MultiBufferSelector.h

@@ -0,0 +1,22 @@
+#ifndef MLIR_DIALECT_PTO_TRANSFORMS_INSERTSYNC_MULTIBUFFERSELECTOR_H
+#define MLIR_DIALECT_PTO_TRANSFORMS_INSERTSYNC_MULTIBUFFERSELECTOR_H
+
+#include "mlir/Dialect/SCF/IR/SCF.h"
+#include "mlir/IR/PatternMatch.h"
+
+namespace mlir {
+namespace pto {
+
+/// Build a boolean `cond` that flips between even/odd iterations across a loop
+/// nest.
+///
+/// - The condition is inserted at the beginning of `baseLoop`'s body.
+/// - The computed parity is based on a flattened linear index across `baseLoop`
+///   and all its parent `scf.for` loops, supporting non-unit steps.
+/// - Returns a null Value if `baseLoop` is invalid.
+Value buildLoopNestParityCond(IRRewriter &rewriter, scf::ForOp baseLoop);
+
+} // namespace pto
+} // namespace mlir
+
+#endif // MLIR_DIALECT_PTO_TRANSFORMS_INSERTSYNC_MULTIBUFFERSELECTOR_H

include/PTO/Transforms/InsertSync/SyncCodegen.h

@@ -57,9 +57,6 @@ class SyncCodegen {
   Value GetBufferSelected(IRRewriter &rewriter, Operation *op,
                           SyncOperation *sync);
 
-  // 生成嵌套循环的计数器 (用于多缓冲切换)
-  Value createNestedIndexModular(IRRewriter &rewriter, Operation *defineOp);
-
 private:
   SyncIRs &syncIR_;
   func::FuncOp func_;

include/PTO/Transforms/Passes.h

@@ -41,6 +41,7 @@ std::unique_ptr<Pass> createLoweringSyncToPipePass();
 
 // Creates a pass for ...
 std::unique_ptr<Pass> createPTOInsertSyncPass();
+std::unique_ptr<Pass> createPTOEnableMultiBufferPass();
 // Default arch is A3 unless overridden by callers.
 std::unique_ptr<Pass> createEmitPTOManualPass();
 // Explicitly select target arch for codegen.

include/PTO/Transforms/Passes.td

@@ -50,6 +50,23 @@ def PTOInsertSync : Pass<"pto-insert-sync", "func::FuncOp"> {
   ];
 }
 
+def PTOEnableMultiBuffer : Pass<"pto-enable-multibuffer", "func::FuncOp"> {
+  let summary = "Materialize multi-buffer (ping/pong) selection";
+  let description = [{
+    Rewrites `pto.pointer_cast(addrs=[ping,pong])` into a loop-local selector
+    that dynamically picks the active address. This enables planned ping/pong
+    buffers to take effect in the emitted C++.
+  }];
+
+  let constructor = "mlir::pto::createPTOEnableMultiBufferPass()";
+
+  let dependentDialects = [
+    "mlir::pto::PTODialect",
+    "mlir::arith::ArithDialect",
+    "mlir::scf::SCFDialect"
+  ];
+}
+
 def ConvertToPTOOp : Pass<"convert-to-pto-op"> {
   let summary = "Convert Ops from other dialects to PTO Ops";
   let constructor = "mlir::pto::createConvertToPTOOpPass()";

lib/PTO/Transforms/CMakeLists.txt

@@ -11,6 +11,7 @@ add_mlir_dialect_library(PTOTransforms
   AllocToPointerCast.cpp
   InferPTOMemScope.cpp
   PTOPlanMemory.cpp
+  EnableMultiBuffer.cpp
   PTORemoveRedundantBarrier.cpp
   InferPTOLayout.cpp
   BufferizableOpInterfaceImpl.cpp
@@ -22,6 +23,7 @@ add_mlir_dialect_library(PTOTransforms
   InsertSync/InsertSyncAnalysis.cpp
   InsertSync/MemoryDependentAnalyzer.cpp
   InsertSync/MoveSyncState.cpp
+  InsertSync/MultiBufferSelector.cpp
   InsertSync/RemoveRedundantSync.cpp
   InsertSync/SyncEventIdAllocation.cpp
   InsertSync/SyncCodegen.cpp