Bump LLVM version to llvmorg-21.1.0-rc2

.gitmodules

@@ -1,7 +1,7 @@
 [submodule "llvm"]
 	path = llvm
 	url = https://github.com/llvm/llvm-project.git
-	branch = main
+	branch = llvmorg-21.1.0-rc2
 	shallow = true
 [submodule "thirdparty/mimalloc"]
 	path = thirdparty/mimalloc

backend/llvm/lib/CodeGen/CMakeLists.txt

@@ -31,10 +31,10 @@ add_llvm_component_library(LLVMBuddyCodeGen
   ${LLVM_CodeGen_DIR}/DwarfEHPrepare.cpp
   ${LLVM_CodeGen_DIR}/EarlyIfConversion.cpp
   ${LLVM_CodeGen_DIR}/EdgeBundles.cpp
-  ${LLVM_CodeGen_DIR}/EHContGuardCatchret.cpp
+  ${LLVM_CodeGen_DIR}/EHContGuardTargets.cpp
   ${LLVM_CodeGen_DIR}/ExecutionDomainFix.cpp
   ${LLVM_CodeGen_DIR}/ExpandLargeDivRem.cpp
-  ${LLVM_CodeGen_DIR}/ExpandLargeFpConvert.cpp
+  ${LLVM_CodeGen_DIR}/ExpandFp.cpp
   ${LLVM_CodeGen_DIR}/ExpandMemCmp.cpp
   ${LLVM_CodeGen_DIR}/ExpandPostRAPseudos.cpp
   ${LLVM_CodeGen_DIR}/ExpandReductions.cpp

backend/llvm/lib/IR/CMakeLists.txt

@@ -74,7 +74,6 @@ add_llvm_component_library(LLVMBuddyCore
   ${LLVM_IR_DIR}/User.cpp
   ${LLVM_IR_DIR}/Value.cpp
   ${LLVM_IR_DIR}/ValueSymbolTable.cpp
-  ${LLVM_IR_DIR}/VectorBuilder.cpp
   ${LLVM_IR_DIR}/VectorTypeUtils.cpp
   ${LLVM_IR_DIR}/Verifier.cpp
   ${LLVM_IR_DIR}/VFABIDemangler.cpp

backend/llvm/lib/Target/CMakeLists.txt

@@ -6,7 +6,6 @@ set(LLVM_Target_DIR ${LLVM_MAIN_SRC_DIR}/lib/Target)
 
 add_llvm_component_library(LLVMBuddyTarget
   ${LLVM_Target_DIR}/Target.cpp
-  ${LLVM_Target_DIR}/TargetIntrinsicInfo.cpp
   ${LLVM_Target_DIR}/TargetLoweringObjectFile.cpp
   ${LLVM_Target_DIR}/TargetMachine.cpp
   ${LLVM_Target_DIR}/TargetMachineC.cpp

backend/llvm/lib/Target/RISCV/CMakeLists.txt

@@ -49,6 +49,7 @@ tablegen(LLVM RISCVGenRegisterInfo.inc -gen-register-info)
 tablegen(LLVM RISCVGenSearchableTables.inc -gen-searchable-tables)
 tablegen(LLVM RISCVGenSubtargetInfo.inc -gen-subtarget)
 tablegen(LLVM RISCVGenExegesis.inc -gen-exegesis)
+tablegen(LLVM RISCVGenSDNodeInfo.inc -gen-sd-node-info)
 
 set(LLVM_TARGET_DEFINITIONS ${LLVM_TARGET_RISCV_DIR}/RISCVGISel.td)
 tablegen(LLVM RISCVGenGlobalISel.inc -gen-global-isel)
@@ -189,42 +190,30 @@ endforeach()
 
 # Build LLVMBuddyRISCVDesc target.
 add_llvm_component_library(LLVMBuddyRISCVDesc
-  ${CMAKE_CURRENT_BINARY_DIR}/RISCVAsmPrinter.cpp
-  ${CMAKE_CURRENT_BINARY_DIR}/RISCVCallingConv.cpp
-  ${CMAKE_CURRENT_BINARY_DIR}/RISCVCodeGenPrepare.cpp
-  ${CMAKE_CURRENT_BINARY_DIR}/RISCVConstantPoolValue.cpp
-  ${CMAKE_CURRENT_BINARY_DIR}/RISCVDeadRegisterDefinitions.cpp
-  ${CMAKE_CURRENT_BINARY_DIR}/RISCVExpandAtomicPseudoInsts.cpp
-  ${CMAKE_CURRENT_BINARY_DIR}/RISCVExpandPseudoInsts.cpp
-  ${CMAKE_CURRENT_BINARY_DIR}/RISCVFrameLowering.cpp
-  ${CMAKE_CURRENT_BINARY_DIR}/RISCVGatherScatterLowering.cpp
-  ${CMAKE_CURRENT_BINARY_DIR}/RISCVISelDAGToDAG.cpp
-  ${CMAKE_CURRENT_BINARY_DIR}/RISCVLandingPadSetup.cpp
-  ${CMAKE_CURRENT_BINARY_DIR}/RISCVMergeBaseOffset.cpp
-  ${CMAKE_CURRENT_BINARY_DIR}/RISCVMoveMerger.cpp
-  ${CMAKE_CURRENT_BINARY_DIR}/RISCVOptWInstrs.cpp
-  ${CMAKE_CURRENT_BINARY_DIR}/RISCVPostRAExpandPseudoInsts.cpp
-  ${CMAKE_CURRENT_BINARY_DIR}/RISCVPushPopOptimizer.cpp
-  ${CMAKE_CURRENT_BINARY_DIR}/RISCVRedundantCopyElimination.cpp
-  ${CMAKE_CURRENT_BINARY_DIR}/RISCVRegisterInfo.cpp
-  ${CMAKE_CURRENT_BINARY_DIR}/RISCVSelectionDAGInfo.cpp
-  ${CMAKE_CURRENT_BINARY_DIR}/RISCVSubtarget.cpp
-  ${CMAKE_CURRENT_BINARY_DIR}/RISCVTargetMachine.cpp
-  ${CMAKE_CURRENT_BINARY_DIR}/RISCVTargetObjectFile.cpp
-  ${CMAKE_CURRENT_BINARY_DIR}/RISCVTargetTransformInfo.cpp
-  ${CMAKE_CURRENT_BINARY_DIR}/RISCVVectorMaskDAGMutation.cpp
-  ${CMAKE_CURRENT_BINARY_DIR}/RISCVVectorPeephole.cpp
-  ${CMAKE_CURRENT_BINARY_DIR}/RISCVZacasABIFix.cpp
-
+  ${CMAKE_CURRENT_BINARY_DIR}/RISCVAsmBackend.cpp
+  ${CMAKE_CURRENT_BINARY_DIR}/RISCVBaseInfo.cpp
+  ${CMAKE_CURRENT_BINARY_DIR}/RISCVELFObjectWriter.cpp
+  ${CMAKE_CURRENT_BINARY_DIR}/RISCVInstPrinter.cpp
+  ${CMAKE_CURRENT_BINARY_DIR}/RISCVMCAsmInfo.cpp
+  ${CMAKE_CURRENT_BINARY_DIR}/RISCVMCCodeEmitter.cpp
+  ${CMAKE_CURRENT_BINARY_DIR}/RISCVMCExpr.cpp
+  ${CMAKE_CURRENT_BINARY_DIR}/RISCVMCObjectFileInfo.cpp
+  ${CMAKE_CURRENT_BINARY_DIR}/RISCVMCTargetDesc.cpp
+  ${CMAKE_CURRENT_BINARY_DIR}/RISCVMatInt.cpp
+  ${CMAKE_CURRENT_BINARY_DIR}/RISCVTargetStreamer.cpp
+  ${CMAKE_CURRENT_BINARY_DIR}/RISCVELFStreamer.cpp
 
   # Add *.h files to track the copies above.
-  ${CMAKE_CURRENT_BINARY_DIR}/RISCVCallingConv.h
-  ${CMAKE_CURRENT_BINARY_DIR}/RISCVInstrInfo.h
-  ${CMAKE_CURRENT_BINARY_DIR}/RISCVRegisterInfo.h
-  ${CMAKE_CURRENT_BINARY_DIR}/RISCVSelectionDAGInfo.h
-  ${CMAKE_CURRENT_BINARY_DIR}/RISCVSubtarget.h
-  ${CMAKE_CURRENT_BINARY_DIR}/RISCVTargetMachine.h
-  ${CMAKE_CURRENT_BINARY_DIR}/RISCVTargetObjectFile.h
+  ${CMAKE_CURRENT_BINARY_DIR}/RISCVAsmBackend.h
+  ${CMAKE_CURRENT_BINARY_DIR}/RISCVBaseInfo.h
+  ${CMAKE_CURRENT_BINARY_DIR}/RISCVELFStreamer.h
+  ${CMAKE_CURRENT_BINARY_DIR}/RISCVFixupKinds.h
+  ${CMAKE_CURRENT_BINARY_DIR}/RISCVInstPrinter.h
+  ${CMAKE_CURRENT_BINARY_DIR}/RISCVMatInt.h
+  ${CMAKE_CURRENT_BINARY_DIR}/RISCVMCAsmInfo.h
+  ${CMAKE_CURRENT_BINARY_DIR}/RISCVMCObjectFileInfo.h
+  ${CMAKE_CURRENT_BINARY_DIR}/RISCVMCTargetDesc.h
+  ${CMAKE_CURRENT_BINARY_DIR}/RISCVTargetStreamer.h
 
 
   LINK_COMPONENTS

backend/llvm/lib/Target/RISCV/RISCVInstrInfoBuddyExt.td

@@ -17,6 +17,22 @@
 // This is the instruction information file of RISC-V buddy extension.
 //
 //===----------------------------------------------------------------------===//
+//
+// IMPORTANT: Instruction Encoding Fix (LLVM 21.0.0 RC1 Upgrade)
+// ------------------------------------------------------------
+// The CONFIG_* instructions originally had identical encodings (func7=0b0000000),
+// which caused decoding conflicts in LLVM's TableGen disassembler.
+// 
+// Fixed encodings:
+// - CONFIG_LD:   func7 = 0b0010110 (was 0b0000000)
+// - CONFIG_ST:   func7 = 0b0010111 (was 0b0000000)
+// - CONFIG_EX:   func7 = 0b0011000 (was 0b0000000)
+// - CONFIG_NORM: func7 = 0b0011001 (was 0b0000000)
+//
+// All instructions retain func3=0b011 and use OPC_CUSTOM_3.
+// This ensures each instruction has a unique bit pattern for proper decoding.
+//
+//===----------------------------------------------------------------------===//
 
 include "llvm/IR/IntrinsicsRISCVBuddyExt.td"
 
@@ -60,25 +76,25 @@ def FLUSH : RVInstR<0b0000111, 0b011, OPC_CUSTOM_3, (outs),
 }
 
 let Predicates = [HasBuddyExt] in 
-def CONFIG_LD : RVInstR<0b0000000, 0b011, OPC_CUSTOM_3, (outs),
+def CONFIG_LD : RVInstR<0b0010110, 0b011, OPC_CUSTOM_3, (outs),
                     (ins GPR:$rs1, GPR:$rs2), "config_ld", "$rs1, $rs2"> {
   let rd = 0;
 }
 
 let Predicates = [HasBuddyExt] in 
-def CONFIG_ST : RVInstR<0b0000000, 0b011, OPC_CUSTOM_3, (outs),
+def CONFIG_ST : RVInstR<0b0010111, 0b011, OPC_CUSTOM_3, (outs),
                     (ins GPR:$rs1, GPR:$rs2), "config_st", "$rs1, $rs2"> {
   let rd = 0;
 }
 
 let Predicates = [HasBuddyExt] in 
-def CONFIG_EX : RVInstR<0b0000000, 0b011, OPC_CUSTOM_3,(outs),
+def CONFIG_EX : RVInstR<0b0011000, 0b011, OPC_CUSTOM_3,(outs),
                     (ins GPR:$rs1, GPR:$rs2), "config_ex", "$rs1, $rs2"> {
   let rd = 0;
 }
 
 let Predicates = [HasBuddyExt] in 
-def CONFIG_NORM : RVInstR<0b0000000, 0b011, OPC_CUSTOM_3,(outs),
+def CONFIG_NORM : RVInstR<0b0011001, 0b011, OPC_CUSTOM_3,(outs),
                     (ins GPR:$rs1, GPR:$rs2), "config_norm", "$rs1, $rs2"> {
   let rd = 0;
 }

backend/llvm/lib/Transforms/Vectorize/CMakeLists.txt

@@ -8,7 +8,6 @@ add_llvm_component_library(LLVMBuddyVectorize
   ${LLVM_Vectorize_DIR}/Vectorize.cpp
   ${LLVM_Vectorize_DIR}/VectorCombine.cpp
   ${LLVM_Vectorize_DIR}/VPlan.cpp
-  ${LLVM_Vectorize_DIR}/VPlanHCFGBuilder.cpp
   ${LLVM_Vectorize_DIR}/VPlanRecipes.cpp
   ${LLVM_Vectorize_DIR}/VPlanSLP.cpp
   ${LLVM_Vectorize_DIR}/VPlanTransforms.cpp

examples/BuddyConvolution/conv2d-nhwc-fhwc-opt.mlir

@@ -4,7 +4,7 @@
 // RUN:   --one-shot-bufferize="bufferize-function-boundaries" \
 // RUN:   -convert-scf-to-cf \
 // RUN: 	-convert-cf-to-llvm \
-// RUN:   -convert-vector-to-llvm \
+// RUN:   -convert-vector-to-llvm -convert-ub-to-llvm \
 // RUN:   -convert-arith-to-llvm \
 // RUN:   -finalize-memref-to-llvm \
 // RUN:   -convert-func-to-llvm \

examples/BuddyNext/next-attention-fusion.mlir

@@ -29,18 +29,13 @@ module {
   memref.global "private" constant @__constant_32x128x40xf32 : memref<32x128x40xf32> = dense<2.000000e+00> {alignment = 64 : i64}
   memref.global "private" constant @__constant_32x40x128xf32 : memref<32x40x128xf32> = dense<3.000000e+00> {alignment = 64 : i64}
   memref.global "private" constant @__constant_1x32x40x40xf32 : memref<1x32x40x40xf32> = dense<11.3137083> {alignment = 64 : i64}
-  func.func @kenerl(%arg0: tensor<32x40x128xf32>, %arg1: tensor<32x128x40xf32>, %arg2: tensor<1x1x40x40xf32>, %arg3: tensor<1x32x40x128xf32>) {
+  func.func @kenerl(%arg0: memref<32x40x128xf32>, %arg1: memref<32x128x40xf32>, %arg2: memref<1x1x40x40xf32>, %arg3: memref<1x32x40x128xf32>) {
     %t_start = call @rtclock() : () -> f64
     %cst = arith.constant 0.0883883461 : f32
     %c0 = arith.constant 0 : index
     %cst_0 = arith.constant 0.000000e+00 : f32
     %cst_1 = arith.constant 1.000000e+00 : f32
     %cst_2 = arith.constant -3.40282347E+38 : f32
-    %0 = bufferization.to_memref %arg3 : tensor<1x32x40x128xf32> to memref<1x32x40x128xf32, strided<[?, ?, ?, ?], offset: ?>>
-    %1 = bufferization.to_memref %arg2 : tensor<1x1x40x40xf32> to memref<1x1x40x40xf32, strided<[?, ?, ?, ?], offset: ?>>
-    %2 = bufferization.to_memref %arg1 : tensor<32x128x40xf32> to memref<32x128x40xf32, strided<[?, ?, ?], offset: ?>>
-    %3 = bufferization.to_memref %arg0 : tensor<32x40x128xf32> to memref<32x40x128xf32, strided<[?, ?, ?], offset: ?>>
-
     // MatMul
     // %0 = tosa.matmul %t0, %t1 : (tensor<32x40x128xf32>, tensor<32x128x40xf32>) -> tensor<32x40x40xf32>
     // Initialize MatMul Output.
@@ -57,8 +52,8 @@ module {
       affine.for %arg5 = 0 to 40 {
         affine.for %arg6 = 0 to 40 {
           affine.for %arg7 = 0 to 128 {
-            %5 = affine.load %3[%arg4, %arg5, %arg7] : memref<32x40x128xf32, strided<[?, ?, ?], offset: ?>>
-            %6 = affine.load %2[%arg4, %arg7, %arg6] : memref<32x128x40xf32, strided<[?, ?, ?], offset: ?>>
+            %5 = affine.load %arg0[%arg4, %arg5, %arg7] : memref<32x40x128xf32>
+            %6 = affine.load %arg1[%arg4, %arg7, %arg6] : memref<32x128x40xf32>
             %7 = affine.load %alloc[%arg4, %arg5, %arg6] : memref<32x40x40xf32>
             %8 = arith.mulf %5, %6 : f32
             %9 = arith.addf %7, %8 : f32
@@ -72,7 +67,8 @@ module {
     // %1 = tosa.reshape %0 {new_shape = array<i64: 1, 32, 40, 40>} : (tensor<32x40x40xf32>) -> tensor<1x32x40x40xf32>
     // %2 = "tosa.const"() <{value = dense<11.3137083> : tensor<1x32x40x40xf32>}> : () -> tensor<1x32x40x40xf32>
     // %3 = tosa.reciprocal %2 : (tensor<1x32x40x40xf32>) -> tensor<1x32x40x40xf32>
-    // %4 = tosa.mul %1, %3 {shift = 0 : i8} : (tensor<1x32x40x40xf32>, tensor<1x32x40x40xf32>) -> tensor<1x32x40x40xf32>
+    //   %shift_4 = "tosa.const"() <{values = dense<0> : tensor<1xi8>}> : () -> tensor<1xi8>
+    // %4 = tosa.mul %1, %3, %shift_4 : (tensor<1x32x40x40xf32>, tensor<1x32x40x40xf32>, tensor<1xi8>) -> tensor<1x32x40x40xf32>
     // %5 = tosa.add %4, %t2 : (tensor<1x32x40x40xf32>, tensor<1x1x40x40xf32>) -> tensor<1x32x40x40xf32>
     // %6 = tosa.reduce_max %5 {axis = 3 : i32} : (tensor<1x32x40x40xf32>) -> tensor<1x32x40x1xf32>
     %expand_shape = memref.expand_shape %alloc [[0, 1], [2], [3]] output_shape [1, 32, 40, 40]: memref<32x40x40xf32> into memref<1x32x40x40xf32>
@@ -93,7 +89,7 @@ module {
             // Fusion point: reshape + constant + reciprocal -> %cst
             %6 = arith.mulf %5, %cst : f32
             // Fusion point: addition
-            %7 = affine.load %1[%c0, %c0, %arg6, %arg7] : memref<1x1x40x40xf32, strided<[?, ?, ?, ?], offset: ?>>
+            %7 = affine.load %arg2[%c0, %c0, %arg6, %arg7] : memref<1x1x40x40xf32>
             %8 = arith.addf %6, %7 : f32
             // Fusion point: reduce max
             %9 = affine.load %alloc_6[%arg4, %arg5, %arg6] : memref<1x32x40xf32>
@@ -142,7 +138,8 @@ module {
 
     // Fusion: Reciprocal + Multiplication
     // %10 = tosa.reciprocal %9 : (tensor<1x32x40x1xf32>) -> tensor<1x32x40x1xf32>
-    // %11 = tosa.mul %8, %10 {shift = 0 : i8} : (tensor<1x32x40x40xf32>, tensor<1x32x40x1xf32>) -> tensor<1x32x40x40xf32>
+    //   %shift_11 = "tosa.const"() <{values = dense<0> : tensor<1xi8>}> : () -> tensor<1xi8>
+    // %11 = tosa.mul %8, %10, %shift_11 : (tensor<1x32x40x40xf32>, tensor<1x32x40x1xf32>, tensor<1xi8>) -> tensor<1x32x40x40xf32>
     %expand_shape_11 = memref.expand_shape %alloc_10 [[0], [1], [2, 3]] output_shape [1, 32, 40, 1]: memref<1x32x40xf32> into memref<1x32x40x1xf32>
     %alloc_13 = memref.alloc() {alignment = 64 : i64} : memref<1x32x40x40xf32>
     affine.for %arg4 = 0 to 1 {
@@ -171,7 +168,7 @@ module {
     %collapse_shape = memref.collapse_shape %alloc_13 [[0, 1], [2], [3]] : memref<1x32x40x40xf32> into memref<32x40x40xf32>
     %alloc_14 = memref.alloc() {alignment = 64 : i64} : memref<1x32x40x128xf32>
     // SSA value %0 is from %arg3
-    memref.copy %0, %alloc_14 : memref<1x32x40x128xf32, strided<[?, ?, ?, ?], offset: ?>> to memref<1x32x40x128xf32>
+    memref.copy %arg3, %alloc_14 : memref<1x32x40x128xf32> to memref<1x32x40x128xf32>
     %collapse_shape_15 = memref.collapse_shape %alloc_14 [[0, 1], [2], [3]] : memref<1x32x40x128xf32> into memref<32x40x128xf32>
 
     // MatMul
@@ -222,14 +219,10 @@ module {
   }
   func.func @main() {
     %0 = memref.get_global @__constant_32x40x128xf32 : memref<32x40x128xf32>
-    %1 = bufferization.to_tensor %0 restrict: memref<32x40x128xf32> to tensor<32x40x128xf32>
-    %2 = memref.get_global @__constant_32x128x40xf32 : memref<32x128x40xf32>
-    %3 = bufferization.to_tensor %2 restrict: memref<32x128x40xf32> to tensor<32x128x40xf32>
-    %4 = memref.get_global @__constant_1x1x40x40xf32 : memref<1x1x40x40xf32>
-    %5 = bufferization.to_tensor %4 restrict: memref<1x1x40x40xf32> to tensor<1x1x40x40xf32>
-    %6 = memref.get_global @__constant_1x32x40x128xf32 : memref<1x32x40x128xf32>
-    %7 = bufferization.to_tensor %6 restrict: memref<1x32x40x128xf32> to tensor<1x32x40x128xf32>
-    call @kenerl(%1, %3, %5, %7) : (tensor<32x40x128xf32>, tensor<32x128x40xf32>, tensor<1x1x40x40xf32>, tensor<1x32x40x128xf32>) -> ()
+    %1 = memref.get_global @__constant_32x128x40xf32 : memref<32x128x40xf32>
+    %2 = memref.get_global @__constant_1x1x40x40xf32 : memref<1x1x40x40xf32>
+    %3 = memref.get_global @__constant_1x32x40x128xf32 : memref<1x32x40x128xf32>
+    call @kenerl(%0, %1, %2, %3) : (memref<32x40x128xf32>, memref<32x128x40xf32>, memref<1x1x40x40xf32>, memref<1x32x40x128xf32>) -> ()
     return
   }
   func.func private @printMemrefF32(memref<*xf32>)

examples/BuddyNext/next-attention-loop.mlir

@@ -29,18 +29,13 @@ module {
   memref.global "private" constant @__constant_32x128x40xf32 : memref<32x128x40xf32> = dense<2.000000e+00> {alignment = 64 : i64}
   memref.global "private" constant @__constant_32x40x128xf32 : memref<32x40x128xf32> = dense<3.000000e+00> {alignment = 64 : i64}
   memref.global "private" constant @__constant_1x32x40x40xf32 : memref<1x32x40x40xf32> = dense<11.3137083> {alignment = 64 : i64}
-  func.func @kenerl(%arg0: tensor<32x40x128xf32>, %arg1: tensor<32x128x40xf32>, %arg2: tensor<1x1x40x40xf32>, %arg3: tensor<1x32x40x128xf32>) {
+  func.func @kenerl(%arg0: memref<32x40x128xf32>, %arg1: memref<32x128x40xf32>, %arg2: memref<1x1x40x40xf32>, %arg3: memref<1x32x40x128xf32>) {
     %t_start = call @rtclock() : () -> f64
     %cst = arith.constant 0.0883883461 : f32
     %c0 = arith.constant 0 : index
     %cst_0 = arith.constant 0.000000e+00 : f32
     %cst_1 = arith.constant 1.000000e+00 : f32
     %cst_2 = arith.constant -3.40282347E+38 : f32
-    %0 = bufferization.to_memref %arg3 : tensor<1x32x40x128xf32> to memref<1x32x40x128xf32, strided<[?, ?, ?, ?], offset: ?>>
-    %1 = bufferization.to_memref %arg2 : tensor<1x1x40x40xf32> to memref<1x1x40x40xf32, strided<[?, ?, ?, ?], offset: ?>>
-    %2 = bufferization.to_memref %arg1 : tensor<32x128x40xf32> to memref<32x128x40xf32, strided<[?, ?, ?], offset: ?>>
-    %3 = bufferization.to_memref %arg0 : tensor<32x40x128xf32> to memref<32x40x128xf32, strided<[?, ?, ?], offset: ?>>
-
     // MatMul
     // %0 = tosa.matmul %t0, %t1 : (tensor<32x40x128xf32>, tensor<32x128x40xf32>) -> tensor<32x40x40xf32>
     // Initialize MatMul Output.
@@ -57,8 +52,8 @@ module {
       affine.for %arg5 = 0 to 40 {
         affine.for %arg6 = 0 to 40 {
           affine.for %arg7 = 0 to 128 {
-            %5 = affine.load %3[%arg4, %arg5, %arg7] : memref<32x40x128xf32, strided<[?, ?, ?], offset: ?>>
-            %6 = affine.load %2[%arg4, %arg7, %arg6] : memref<32x128x40xf32, strided<[?, ?, ?], offset: ?>>
+            %5 = affine.load %arg0[%arg4, %arg5, %arg7] : memref<32x40x128xf32>
+            %6 = affine.load %arg1[%arg4, %arg7, %arg6] : memref<32x128x40xf32>
             %7 = affine.load %alloc[%arg4, %arg5, %arg6] : memref<32x40x40xf32>
             %8 = arith.mulf %5, %6 : f32
             %9 = arith.addf %7, %8 : f32
@@ -85,7 +80,8 @@ module {
     }
 
     // Multiplication
-    // %4 = tosa.mul %1, %3 {shift = 0 : i8} : (tensor<1x32x40x40xf32>, tensor<1x32x40x40xf32>) -> tensor<1x32x40x40xf32>
+    //   %shift_4 = "tosa.const"() <{values = dense<0> : tensor<1xi8>}> : () -> tensor<1xi8>
+  // %4 = tosa.mul %1, %3, %shift_4 : (tensor<1x32x40x40xf32>, tensor<1x32x40x40xf32>, tensor<1xi8>) -> tensor<1x32x40x40xf32>
     %alloc_4 = memref.alloc() {alignment = 64 : i64} : memref<1x32x40x40xf32>
     affine.for %arg4 = 0 to 1 {
       affine.for %arg5 = 0 to 32 {
@@ -108,7 +104,7 @@ module {
         affine.for %arg6 = 0 to 40 {
           affine.for %arg7 = 0 to 40 {
             %5 = affine.load %alloc_4[%c0, %arg5, %arg6, %arg7] : memref<1x32x40x40xf32>
-            %6 = affine.load %1[%c0, %c0, %arg6, %arg7] : memref<1x1x40x40xf32, strided<[?, ?, ?, ?], offset: ?>>
+            %6 = affine.load %arg2[%c0, %c0, %arg6, %arg7] : memref<1x1x40x40xf32>
             %7 = arith.addf %5, %6 : f32
             affine.store %7, %alloc_5[%arg4, %arg5, %arg6, %arg7] : memref<1x32x40x40xf32>
           }
@@ -220,7 +216,8 @@ module {
     }
 
     // Multiplication
-    // %11 = tosa.mul %8, %10 {shift = 0 : i8} : (tensor<1x32x40x40xf32>, tensor<1x32x40x1xf32>) -> tensor<1x32x40x40xf32>
+    //   %shift_11 = "tosa.const"() <{values = dense<0> : tensor<1xi8>}> : () -> tensor<1xi8>
+  // %11 = tosa.mul %8, %10, %shift_11 : (tensor<1x32x40x40xf32>, tensor<1x32x40x1xf32>, tensor<1xi8>) -> tensor<1x32x40x40xf32>
     %alloc_13 = memref.alloc() {alignment = 64 : i64} : memref<1x32x40x40xf32>
     affine.for %arg4 = 0 to 1 {
       affine.for %arg5 = 0 to 32 {
@@ -245,7 +242,7 @@ module {
     %collapse_shape = memref.collapse_shape %alloc_13 [[0, 1], [2], [3]] : memref<1x32x40x40xf32> into memref<32x40x40xf32>
     %alloc_14 = memref.alloc() {alignment = 64 : i64} : memref<1x32x40x128xf32>
     // SSA value %0 is from %arg3
-    memref.copy %0, %alloc_14 : memref<1x32x40x128xf32, strided<[?, ?, ?, ?], offset: ?>> to memref<1x32x40x128xf32>
+    memref.copy %arg3, %alloc_14 : memref<1x32x40x128xf32> to memref<1x32x40x128xf32>
     %collapse_shape_15 = memref.collapse_shape %alloc_14 [[0, 1], [2], [3]] : memref<1x32x40x128xf32> into memref<32x40x128xf32>
 
     // MatMul
@@ -297,14 +294,10 @@ module {
   }
   func.func @main() {
     %0 = memref.get_global @__constant_32x40x128xf32 : memref<32x40x128xf32>
-    %1 = bufferization.to_tensor %0 restrict: memref<32x40x128xf32> to tensor<32x40x128xf32>
-    %2 = memref.get_global @__constant_32x128x40xf32 : memref<32x128x40xf32>
-    %3 = bufferization.to_tensor %2 restrict: memref<32x128x40xf32> to tensor<32x128x40xf32>
-    %4 = memref.get_global @__constant_1x1x40x40xf32 : memref<1x1x40x40xf32>
-    %5 = bufferization.to_tensor %4 restrict: memref<1x1x40x40xf32> to tensor<1x1x40x40xf32>
-    %6 = memref.get_global @__constant_1x32x40x128xf32 : memref<1x32x40x128xf32>
-    %7 = bufferization.to_tensor %6 restrict: memref<1x32x40x128xf32> to tensor<1x32x40x128xf32>
-    call @kenerl(%1, %3, %5, %7) : (tensor<32x40x128xf32>, tensor<32x128x40xf32>, tensor<1x1x40x40xf32>, tensor<1x32x40x128xf32>) -> ()
+    %1 = memref.get_global @__constant_32x128x40xf32 : memref<32x128x40xf32>
+    %2 = memref.get_global @__constant_1x1x40x40xf32 : memref<1x1x40x40xf32>
+    %3 = memref.get_global @__constant_1x32x40x128xf32 : memref<1x32x40x128xf32>
+    call @kenerl(%0, %1, %2, %3) : (memref<32x40x128xf32>, memref<32x128x40xf32>, memref<1x1x40x40xf32>, memref<1x32x40x128xf32>) -> ()
     return
   }
   func.func private @printMemrefF32(memref<*xf32>)