[InsertSync] scalar autosync + addptr alias-through (scalar-only scope)

include/PTO/IR/PTOOps.td

@@ -100,6 +100,7 @@ def AddPtrOp : PTO_Op<"addptr", [
 //===----------------------------------------------------------------------===//
 
 def LoadScalarOp : PTO_Op<"load_scalar", [
+    OpPipeInterface,
     DeclareOpInterfaceMethods<MemoryEffectsOpInterface>
   ]> {
   let summary = "Load a single scalar element from a pointer at offset.";
@@ -116,9 +117,14 @@ def LoadScalarOp : PTO_Op<"load_scalar", [
   let assemblyFormat = [{
     $ptr `[` $offset `]` attr-dict `:` type($ptr) `->` type($value)
   }];
+
+  let extraClassDeclaration = [{
+    ::mlir::pto::PIPE getPipe() { return ::mlir::pto::PIPE::PIPE_S; }
+  }];
 }
 
 def StoreScalarOp : PTO_Op<"store_scalar", [
+    OpPipeInterface,
     DeclareOpInterfaceMethods<MemoryEffectsOpInterface>
   ]> {
   let summary = "Store a single scalar element to a pointer at offset.";
@@ -136,6 +142,10 @@ def StoreScalarOp : PTO_Op<"store_scalar", [
   let assemblyFormat = [{
     $value `,` $ptr `[` $offset `]` attr-dict `:` type($ptr) `,` type($value)
   }];
+
+  let extraClassDeclaration = [{
+    ::mlir::pto::PIPE getPipe() { return ::mlir::pto::PIPE::PIPE_S; }
+  }];
 }
 
 def MakeTensorViewOp : PTO_Op<"make_tensor_view", [AttrSizedOperandSegments]> {

include/PTO/Transforms/InsertSync/PTOIRTranslator.h

@@ -8,6 +8,7 @@
 #include "mlir/Dialect/SCF/IR/SCF.h"
 #include "mlir/IR/BuiltinOps.h"
 #include "llvm/Support/raw_ostream.h"
+#include <vector>
 
 namespace mlir {
 namespace pto {
@@ -76,12 +77,19 @@ class PTOIRTranslator {
 
   // 根据 Values 填充 Def/Use 列表
   void UpdateDefUseVec(ValueRange values, SmallVector<const BaseMemInfo *> &vec);
+
+  // scalar 访问切片建模：按 ptr+offset 构建访问级依赖信息。
+  void UpdateScalarDefUseVec(Value ptr, Value offset, Type scalarType,
+                             SmallVector<const BaseMemInfo *> &vec);
 
   // 调试辅助
   std::string getPipelineName(PipelineType pipe);
   void printMemInfoList(llvm::raw_ostream &os, 
                         const SmallVector<const BaseMemInfo *> &list, 
                         AsmState &state);
+
+  // 持久化 scalar 访问切片，保证 def/use 指针在分析期间有效。
+  std::vector<std::unique_ptr<BaseMemInfo>> scalarAccessMemInfoPool_;
 };
 
 } // namespace pto

include/PTO/Transforms/InsertSync/SyncCommon.h

@@ -73,9 +73,11 @@ enum class TCoreType {
 struct BaseMemInfo {
   BaseMemInfo(
       Value baseBuffer, Value rootBuffer, pto::AddressSpace scope,
-      SmallVector<uint64_t> baseAddresses, uint64_t allocateSize)
+      SmallVector<uint64_t> baseAddresses, uint64_t allocateSize,
+      bool unknownRange = false)
       : baseBuffer(baseBuffer), rootBuffer(rootBuffer), scope(scope),
-        baseAddresses(std::move(baseAddresses)), allocateSize(allocateSize) {}
+        baseAddresses(std::move(baseAddresses)), allocateSize(allocateSize),
+        unknownRange(unknownRange) {}
 
   /// baseBuffer: 当前操作直接使用的 Buffer (可能是 View 或 Alias)
   Value baseBuffer;
@@ -85,6 +87,8 @@ struct BaseMemInfo {
   pto::AddressSpace scope;
   SmallVector<uint64_t> baseAddresses; // 用于 Offset 分析
   uint64_t allocateSize;
+  // True means alias range is unknown and must be treated conservatively.
+  bool unknownRange{false};
 
   bool areVectorEqual(const SmallVector<uint64_t>& vec1,
                       const SmallVector<uint64_t>& vec2) const {
@@ -99,6 +103,7 @@ struct BaseMemInfo {
     if (!areVectorEqual(baseAddresses, other.baseAddresses)) return false;
     if (rootBuffer != other.rootBuffer) return false;
     if (scope != other.scope) return false;
+    if (unknownRange != other.unknownRange) return false;
     // allocateSize 和 baseBuffer 的严格相等性在某些别名分析中可能太强了，
     // 但为了保持原有逻辑，先保留。重点是 rootBuffer 必须一致。
     if (allocateSize != other.allocateSize) return false;
@@ -108,12 +113,14 @@ struct BaseMemInfo {
 
   std::unique_ptr<BaseMemInfo> clone() const {
     return std::make_unique<BaseMemInfo>(
-        baseBuffer, rootBuffer, scope, baseAddresses, allocateSize);
+        baseBuffer, rootBuffer, scope, baseAddresses, allocateSize,
+        unknownRange);
   }
- 
+
   std::unique_ptr<BaseMemInfo> clone(Value cloneBaseBuffer) const {
     return std::make_unique<BaseMemInfo>(
-        cloneBaseBuffer, rootBuffer, scope, baseAddresses, allocateSize);
+        cloneBaseBuffer, rootBuffer, scope, baseAddresses, allocateSize,
+        unknownRange);
   }
 };
 
@@ -355,4 +362,4 @@ void checkCondition(bool condition, const std::string &message);
 } // namespace pto
 } // namespace mlir
 
-#endif // MLIR_DIALECT_PTO_TRANSFORMS_SYNC_COMMON_H
+#endif // MLIR_DIALECT_PTO_TRANSFORMS_SYNC_COMMON_H

lib/PTO/Transforms/InsertSync/InsertSyncAnalysis.cpp

@@ -1,4 +1,5 @@
 #include "PTO/Transforms/InsertSync/InsertSyncAnalysis.h"
+#include "PTO/IR/PTO.h"
 #include "PTO/Transforms/InsertSync/SyncCommon.h"
 #include "mlir/Dialect/MemRef/IR/MemRef.h"
 #include "mlir/Dialect/SCF/IR/SCF.h"
@@ -121,9 +122,9 @@ bool InsertSyncAnalysis::IsNoNeedToInsertSync(
   const PipelineType frontPipe = frontCompound->kPipeValue;
   const PipelineType nowPipe = nowCompound->kPipeValue;
 
-  if (frontPipe == nowPipe && frontPipe == PipelineType::PIPE_S) {
+  // Scalar pipe is in-order on target hardware; skip same-pipe sync.
+  if (frontPipe == nowPipe && frontPipe == PipelineType::PIPE_S)
     return true;
-  }
 
   if (nowCompound->elementOp == frontCompound->elementOp && !isBackwardDep) {
     return true;

lib/PTO/Transforms/InsertSync/MemoryDependentAnalyzer.cpp

@@ -144,7 +144,9 @@ bool MemoryDependentAnalyzer::MemAlias(const BaseMemInfo *a,
   // 2. Local Memory (UB/L1)
 
   if (a->rootBuffer == b->rootBuffer) {
+    if (a->unknownRange || b->unknownRange) return true;
     if (a->baseAddresses.empty() || b->baseAddresses.empty()) return true;
+    if (a->allocateSize == 0 || b->allocateSize == 0) return true;
     return isBufferAddressRangeOverlap(a, b);
   }
 
@@ -182,6 +184,7 @@ bool MemoryDependentAnalyzer::isGMBufferOverlap(const BaseMemInfo *a,
     return true; 
   }
 
+  if (a->unknownRange || b->unknownRange) return true;
   if (a->baseAddresses.empty() || b->baseAddresses.empty()) return true; 
   if (a->allocateSize == 0 || b->allocateSize == 0) return true;