Fix TileLang ST CI failures

include/PTO/IR/PTOOps.td

@@ -1465,10 +1465,16 @@ def TMovOp  : PTO_TOp<"tmov", [
         return ::mlir::pto::PIPE::PIPE_V;
       }
 
-      // MAT -> L0 (Left/Right/Bias/Scaling) are MTE1 moves.
+      // MAT -> Scaling moves L1 data into the fixpipe parameter buffer.
+      if (s == ::mlir::pto::AddressSpace::MAT &&
+          d == ::mlir::pto::AddressSpace::SCALING) {
+        return ::mlir::pto::PIPE::PIPE_FIX;
+      }
+
+      // MAT -> L0 (Left/Right/Bias) moves are executed by MTE1.
       if ((s == ::mlir::pto::AddressSpace::MAT &&
            (d == ::mlir::pto::AddressSpace::LEFT || d == ::mlir::pto::AddressSpace::RIGHT ||
-            d == ::mlir::pto::AddressSpace::BIAS || d == ::mlir::pto::AddressSpace::SCALING))) {
+            d == ::mlir::pto::AddressSpace::BIAS))) {
         return ::mlir::pto::PIPE::PIPE_MTE1;
       }
 
@@ -6576,6 +6582,7 @@ def TSqrtOp: PTO_TOp<"tsqrt", [
 def TStoreFPOp: PTO_TOp<"tstore_fp", [
   PTO_DpsInitOpInterface,
   OpPipeInterface,
+  DeclareOpInterfaceMethods<MemoryEffectsOpInterface>
 ]> {
   let summary = "TSTORE_FP: Store an accumulator tile into global memory using a scaling (fp) tile for vector quantization parameters.";
 

lib/PTO/IR/PTO.cpp

@@ -13693,6 +13693,14 @@ void TStoreOp::getEffects(SmallVectorImpl<SideEffects::EffectInstance<MemoryEffe
   addEffect(effects, &getDstMutable(), MemoryEffects::Write::get());
 }
 
+// === TStoreFPOp ===
+// Read: src (ACC), fp (scaling), Write: dst (GM)
+void TStoreFPOp::getEffects(SmallVectorImpl<SideEffects::EffectInstance<MemoryEffects::Effect>> &effects) {
+  addEffect(effects, &getSrcMutable(), MemoryEffects::Read::get());
+  addEffect(effects, &getFpMutable(), MemoryEffects::Read::get());
+  addEffect(effects, &getDstMutable(), MemoryEffects::Write::get());
+}
+
 // === TMovOp ===
 // Read: src, Write: dst
 void TMovOp::getEffects(SmallVectorImpl<SideEffects::EffectInstance<MemoryEffects::Effect>> &effects) {

lib/PTO/IR/VPTO.cpp

@@ -2830,9 +2830,11 @@ static LogicalResult verifyStructuredAccStoreLike(
       if (!isStructuredAccStoreScalingPayload(preQuant))
         return op->emitOpError("vector pre_quant mode requires scaling pointer payload");
       if (!isStructuredAccStoreFloatScalarPayloadType(
-              getStructuredAccStoreScalingElementType(preQuant)))
+              getStructuredAccStoreScalingElementType(preQuant)) &&
+          getStructuredAccStoreScalingElementType(preQuant) !=
+              IntegerType::get(op->getContext(), 64, IntegerType::Unsigned))
         return op->emitOpError(
-            "vector pre_quant mode requires scaling pointer element type to be f16, bf16, or f32");
+            "vector pre_quant mode requires scaling pointer element type to be f16, bf16, f32, or packed ui64 Quant_PRE entries");
     } else if (!isStructuredAccStoreFloatScalarPayload(preQuant)) {
       return op->emitOpError(
           "scalar pre_quant mode requires f16/bf16/f32 payload");

lib/PTO/Transforms/ExpandTileOp.cpp

@@ -1048,6 +1048,14 @@ func::FuncOp ExpandState::invokeTilelangDaemon(const SpecKey &key,
 func::FuncOp ExpandState::invokeTilelangDSL(const SpecKey &key,
                                               Operation *tileOp,
                                               ModuleOp mod, MLIRContext *ctx) {
+  std::string uniqueName = buildUniqueFunctionBaseName(key);
+  SymbolTable targetSymTable(mod);
+  if (auto existingFunc = targetSymTable.lookup(uniqueName)) {
+    llvm::errs() << "ExpandTileOp: reuse existing function @" << uniqueName
+                 << "\n";
+    return cast<func::FuncOp>(existingFunc);
+  }
+
   // Try daemon first if daemon socket path is provided.
   if (!daemonSocketPath.empty()) {
     func::FuncOp daemonResult = invokeTilelangDaemon(key, tileOp, mod, ctx);
@@ -1189,16 +1197,6 @@ func::FuncOp ExpandState::invokeTilelangDSL(const SpecKey &key,
   SmallVector<func::FuncOp, 4> clonedFuncs;
   llvm::StringMap<std::string> renamedSymbols;
 
-  std::string uniqueName = buildUniqueFunctionBaseName(key);
-
-  // Check if function already exists in module (deduplication)
-  SymbolTable targetSymTable(mod);
-  if (auto existingFunc = targetSymTable.lookup(uniqueName)) {
-    // Function already exists, return it directly (avoid redefinition)
-    llvm::errs() << "ExpandTileOp: reuse existing function @" << uniqueName << "\n";
-    return cast<func::FuncOp>(existingFunc);
-  }
-
   std::vector<std::string> newNameStorage;
   for (auto [index, fn] : llvm::enumerate(parsedFuncs)) {
     IRMapping mapping;

lib/PTO/Transforms/VPTOCANN900LLVMEmitter.cpp

@@ -831,6 +831,8 @@ static std::string getCopyElementFragment(Type elementType) {
       return intType.isUnsigned() ? "u16" : "s16";
     case 32:
       return intType.isUnsigned() ? "u32" : "s32";
+    case 64:
+      return "s32";
     default:
       return {};
     }

lib/PTO/Transforms/VPTOLLVMEmitter.cpp

@@ -799,6 +799,8 @@ static std::string getCopyElementFragment(Type elementType) {
       return intType.isUnsigned() ? "u16" : "s16";
     case 32:
       return intType.isUnsigned() ? "u32" : "s32";
+    case 64:
+      return "s32";
     default:
       return {};
     }