Index: lib/Transform/ScheduleOptimizer.cpp
===================================================================
--- lib/Transform/ScheduleOptimizer.cpp
+++ lib/Transform/ScheduleOptimizer.cpp
@@ -671,12 +671,19 @@
 static bool isMatMulOperandAcc(__isl_keep isl_map *AccMap, int &FirstPos,
                                int &SecondPos) {
   int DimInPos[] = {FirstPos, SecondPos};
-  if (isl_map_foreach_basic_map(AccMap, isMatMulOperandBasicMap,
+  if (isl_map_dim(AccMap, isl_dim_out) != 2)
+    return false;
+  auto *NonPartialAccMap = isl_map_drop_constraints_not_involving_dims(
+      isl_map_copy(AccMap), isl_dim_out, 0, 2);
+  if (isl_map_foreach_basic_map(NonPartialAccMap, isMatMulOperandBasicMap,
                                 static_cast<void *>(DimInPos)) != isl_stat_ok ||
-      DimInPos[0] < 0 || DimInPos[1] < 0)
+      DimInPos[0] < 0 || DimInPos[1] < 0) {
+    isl_map_free(NonPartialAccMap);
     return false;
+  }
   FirstPos = DimInPos[0];
   SecondPos = DimInPos[1];
+  isl_map_free(NonPartialAccMap);
   return true;
 }
 
@@ -693,7 +700,7 @@
 ///                  false, otherwise.
 static bool isMatMulNonScalarReadAccess(MemoryAccess *MemAccess,
                                         MatMulInfoTy &MMI) {
-  if (!MemAccess->isArrayKind() || !MemAccess->isRead())
+  if (!MemAccess->isLatestArrayKind() || !MemAccess->isRead())
     return false;
   isl_map *AccMap = MemAccess->getAccessRelation();
   if (isMatMulOperandAcc(AccMap, MMI.i, MMI.j) && !MMI.ReadFromC &&
@@ -748,7 +755,7 @@
                                  MMI.k, OutDimNum - 1);
   for (auto *MemA = Stmt->begin(); MemA != Stmt->end() - 1; MemA++) {
     auto *MemAccessPtr = *MemA;
-    if (MemAccessPtr->isArrayKind() && MemAccessPtr != MMI.WriteToC &&
+    if (MemAccessPtr->isLatestArrayKind() && MemAccessPtr != MMI.WriteToC &&
         !isMatMulNonScalarReadAccess(MemAccessPtr, MMI) &&
         !(MemAccessPtr->isStrideZero(isl_map_copy(MapI)) &&
           MemAccessPtr->isStrideZero(isl_map_copy(MapJ)) &&
@@ -840,7 +847,7 @@
     return false;
   for (auto *MemA = Stmt->end() - 1; MemA != Stmt->begin(); MemA--) {
     auto *MemAccessPtr = *MemA;
-    if (!MemAccessPtr->isArrayKind())
+    if (!MemAccessPtr->isLatestArrayKind())
       continue;
     if (!MemAccessPtr->isWrite())
       return false;
Index: test/ScheduleOptimizer/pattern-matching-based-opts_11.ll
===================================================================
--- /dev/null
+++ test/ScheduleOptimizer/pattern-matching-based-opts_11.ll
@@ -0,0 +1,184 @@
+; RUN: opt %loadPolly -polly-invariant-load-hoisting \
+; RUN: -polly-delicm-overapproximate-writes -polly-opt-isl \
+; RUN: -polly-pattern-matching-based-opts=true \
+; RUN: -debug < %s 2>&1| FileCheck %s
+;
+; TODO: Add the description.
+;
+; CHECK: The matrix multiplication pattern was detected
+;
+%"class.boost::numeric::ublas::matrix" = type { i64, i64, %"class.boost::numeric::ublas::unbounded_array" }
+%"class.boost::numeric::ublas::unbounded_array" = type { %"class.std::allocator", i64, float* }
+%"class.std::allocator" = type { i8 }
+
+declare i32 @__gxx_personality_v0(...)
+
+declare void @_ZSt17__throw_bad_allocv()
+
+declare noalias nonnull i8* @_Znwm(i64)
+
+define void @_Z4gemmRN5boost7numeric5ublas6matrixIfNS1_15basic_row_majorImlEENS1_15unbounded_arrayIfSaIfEEEEES9_S9_(%"class.boost::numeric::ublas::matrix"* dereferenceable(40) %C, %"class.boost::numeric::ublas::matrix"* dereferenceable(40) %A, %"class.boost::numeric::ublas::matrix"* dereferenceable(40) %B) personality i8* bitcast (i32 (...)* @__gxx_personality_v0 to i8*) {
+entry:
+  %temporary.i = alloca %"class.boost::numeric::ublas::matrix", align 8
+  %tmp2 = bitcast %"class.boost::numeric::ublas::matrix"* %temporary.i to i8*
+  %size1_.i.i = getelementptr inbounds %"class.boost::numeric::ublas::matrix", %"class.boost::numeric::ublas::matrix"* %temporary.i, i64 0, i32 0
+  %size1_.i.i.i.i.i = getelementptr inbounds %"class.boost::numeric::ublas::matrix", %"class.boost::numeric::ublas::matrix"* %A, i64 0, i32 0
+  %Asize1 = load i64, i64* %size1_.i.i.i.i.i, align 8
+  store i64 %Asize1, i64* %size1_.i.i, align 8
+  %size2_.i.i.i.i.i = getelementptr inbounds %"class.boost::numeric::ublas::matrix", %"class.boost::numeric::ublas::matrix"* %B, i64 0, i32 1
+  %Bsize2 = load i64, i64* %size2_.i.i.i.i.i, align 8
+  %size2_.i.i = getelementptr inbounds %"class.boost::numeric::ublas::matrix", %"class.boost::numeric::ublas::matrix"* %temporary.i, i64 0, i32 1
+  store i64 %Bsize2, i64* %size2_.i.i, align 8
+  %mul.i.i.i = mul i64 %Bsize2, %Asize1
+  %size_.i21.i.i = getelementptr inbounds %"class.boost::numeric::ublas::matrix", %"class.boost::numeric::ublas::matrix"* %temporary.i, i64 0, i32 2, i32 1
+  store i64 %mul.i.i.i, i64* %size_.i21.i.i, align 8
+  %tobool.i22.i.i = icmp eq i64 %mul.i.i.i, 0
+  br i1 %tobool.i22.i.i, label %_ZN5boost7numeric5ublas15unbounded_arrayIfSaIfEEC2EmRKS3_.exit.i.i, label %if.then.i23.i.i
+
+if.then.i23.i.i:                                  ; preds = %entry
+  %cmp.i.i.i.i = icmp ugt i64 %mul.i.i.i, 4611686018427387903
+  br i1 %cmp.i.i.i.i, label %if.then.i.i.i.i, label %if.end.i.i.i.i
+
+if.then.i.i.i.i:                                  ; preds = %if.then.i23.i.i
+  invoke void @_ZSt17__throw_bad_allocv()
+          to label %.noexc.i.i.i unwind label %lpad.i.i.i
+
+.noexc.i.i.i:                                     ; preds = %if.then.i.i.i.i
+  unreachable
+
+if.end.i.i.i.i:                                   ; preds = %if.then.i23.i.i
+  %mul.i.i.i.i = shl i64 %mul.i.i.i, 2
+  %call2.i8.i.i.i = invoke i8* @_Znwm(i64 %mul.i.i.i.i)
+          to label %_ZN9__gnu_cxx13new_allocatorIfE8allocateEmPKv.exit.i.i.i unwind label %lpad.i.i.i
+
+_ZN9__gnu_cxx13new_allocatorIfE8allocateEmPKv.exit.i.i.i: ; preds = %if.end.i.i.i.i
+  %tmp7 = bitcast i8* %call2.i8.i.i.i to float*
+  br label %_ZN5boost7numeric5ublas15unbounded_arrayIfSaIfEEC2EmRKS3_.exit.i.i
+
+lpad.i.i.i:                                       ; preds = %if.end.i.i.i.i, %if.then.i.i.i.i
+  %tmp8 = landingpad { i8*, i32 }
+          cleanup
+  resume { i8*, i32 } %tmp8
+
+_ZN5boost7numeric5ublas15unbounded_arrayIfSaIfEEC2EmRKS3_.exit.i.i: ; preds = %_ZN9__gnu_cxx13new_allocatorIfE8allocateEmPKv.exit.i.i.i, %entry
+  %.sink.i.i.i = phi float* [ null, %entry ], [ %tmp7, %_ZN9__gnu_cxx13new_allocatorIfE8allocateEmPKv.exit.i.i.i ]
+  %data_5.i.i.i = getelementptr inbounds %"class.boost::numeric::ublas::matrix", %"class.boost::numeric::ublas::matrix"* %temporary.i, i64 0, i32 2, i32 2
+  store float* %.sink.i.i.i, float** %data_5.i.i.i, align 8
+  %cmp26.i.i.i.i.i = icmp eq i64 %Asize1, 0
+  br i1 %cmp26.i.i.i.i.i, label %_ZN5boost7numeric5ublas6matrixIfNS1_15basic_row_majorImlEENS1_15unbounded_arrayIfSaIfEEEEC2INS1_20matrix_matrix_binaryIS8_S8_NS1_18matrix_matrix_prodIS8_S8_fEEEEEERKNS1_17matrix_expressionIT_EE.exit.i, label %for.cond2.preheader.i.i.i.i.i.preheader
+
+for.cond2.preheader.i.i.i.i.i.preheader:          ; preds = %_ZN5boost7numeric5ublas15unbounded_arrayIfSaIfEEC2EmRKS3_.exit.i.i
+  br label %for.cond2.preheader.i.i.i.i.i
+
+for.cond2.preheader.i.i.i.i.i:                    ; preds = %for.cond.cleanup4.i.i.i.i.i, %for.cond2.preheader.i.i.i.i.i.preheader
+  %i.027.i.i.i.i.i = phi i64 [ %inc10.i.i.i.i.i, %for.cond.cleanup4.i.i.i.i.i ], [ 0, %for.cond2.preheader.i.i.i.i.i.preheader ]
+  %cmp324.i.i.i.i.i = icmp eq i64 %Bsize2, 0
+  br i1 %cmp324.i.i.i.i.i, label %for.cond.cleanup4.i.i.i.i.i, label %for.body5.i.i.i.i.i.preheader
+
+for.body5.i.i.i.i.i.preheader:                    ; preds = %for.cond2.preheader.i.i.i.i.i
+  br label %for.body5.i.i.i.i.i
+
+for.cond.cleanup4.i.i.i.i.i.loopexit:             ; preds = %ExitStmt
+  br label %for.cond.cleanup4.i.i.i.i.i
+
+for.cond.cleanup4.i.i.i.i.i:                      ; preds = %for.cond.cleanup4.i.i.i.i.i.loopexit, %for.cond2.preheader.i.i.i.i.i
+  %inc10.i.i.i.i.i = add nuw i64 %i.027.i.i.i.i.i, 1
+  %exitcond5 = icmp ne i64 %inc10.i.i.i.i.i, %Asize1
+  br i1 %exitcond5, label %for.cond2.preheader.i.i.i.i.i, label %_ZN5boost7numeric5ublas6matrixIfNS1_15basic_row_majorImlEENS1_15unbounded_arrayIfSaIfEEEEC2INS1_20matrix_matrix_binaryIS8_S8_NS1_18matrix_matrix_prodIS8_S8_fEEEEEERKNS1_17matrix_expressionIT_EE.exit.i.loopexit
+
+for.body5.i.i.i.i.i:                              ; preds = %ExitStmt, %for.body5.i.i.i.i.i.preheader
+  %j.025.i.i.i.i.i = phi i64 [ %inc.i.i.i.i.i, %ExitStmt ], [ 0, %for.body5.i.i.i.i.i.preheader ]
+  %tmp11 = load i64, i64* %size2_.i.i, align 8
+  %mul.i.i.i.i.i.i.i.i = mul i64 %tmp11, %i.027.i.i.i.i.i
+  %add.i.i.i.i.i.i.i.i = add i64 %mul.i.i.i.i.i.i.i.i, %j.025.i.i.i.i.i
+  %tmp12 = load float*, float** %data_5.i.i.i, align 8
+  %arrayidx.i.i.i.i.i.i.i.i = getelementptr inbounds float, float* %tmp12, i64 %add.i.i.i.i.i.i.i.i
+  %size2_.i.i24.i.i.i.i.i.i.i = getelementptr inbounds %"class.boost::numeric::ublas::matrix", %"class.boost::numeric::ublas::matrix"* %A, i64 0, i32 1
+  %Asize2 = load i64, i64* %size2_.i.i24.i.i.i.i.i.i.i, align 8
+  %cmp14.i.i.i.i.i.i.i = icmp eq i64 %Asize2, 0
+  br i1 %cmp14.i.i.i.i.i.i.i, label %ExitStmt, label %KernelStmt.preheader
+
+KernelStmt.preheader:                             ; preds = %for.body5.i.i.i.i.i
+  %data_.i4.i.i22.i.i.i.i.i.i.i.phi.trans.insert = getelementptr inbounds %"class.boost::numeric::ublas::matrix", %"class.boost::numeric::ublas::matrix"* %A, i64 0, i32 2, i32 2
+  %tmp15.pre = load float*, float** %data_.i4.i.i22.i.i.i.i.i.i.i.phi.trans.insert, align 8
+  %tmp18.pre = load i64, i64* %size2_.i.i.i.i.i, align 8
+  %data_.i4.i.i.i.i.i.i.i.i.i.phi.trans.insert = getelementptr inbounds %"class.boost::numeric::ublas::matrix", %"class.boost::numeric::ublas::matrix"* %B, i64 0, i32 2, i32 2
+  %tmp19.pre = load float*, float** %data_.i4.i.i.i.i.i.i.i.i.i.phi.trans.insert, align 8
+  br label %KernelStmt
+
+KernelStmt:                                       ; preds = %KernelStmt, %KernelStmt.preheader
+  %k.016.i.i.i.i.i.i.i = phi i64 [ %inc.i.i.i.i.i.i.i, %KernelStmt ], [ 0, %KernelStmt.preheader ]
+  %t.015.i.i.i.i.i.i.i = phi float [ %add.i.i.i.i.i.i.i, %KernelStmt ], [ 0.000000e+00, %KernelStmt.preheader ]
+  %mul.i.i.i20.i.i.i.i.i.i.i = mul i64 %Asize2, %i.027.i.i.i.i.i
+  %add.i.i.i21.i.i.i.i.i.i.i = add i64 %mul.i.i.i20.i.i.i.i.i.i.i, %k.016.i.i.i.i.i.i.i
+  %arrayidx.i.i.i23.i.i.i.i.i.i.i = getelementptr inbounds float, float* %tmp15.pre, i64 %add.i.i.i21.i.i.i.i.i.i.i
+  %tmp16 = load float, float* %arrayidx.i.i.i23.i.i.i.i.i.i.i, align 4
+  %mul.i.i.i.i.i.i.i.i.i.i = mul i64 %tmp18.pre, %k.016.i.i.i.i.i.i.i
+  %add.i.i.i.i.i.i.i.i.i.i = add i64 %mul.i.i.i.i.i.i.i.i.i.i, %j.025.i.i.i.i.i
+  %arrayidx.i.i.i.i.i.i.i.i.i.i = getelementptr inbounds float, float* %tmp19.pre, i64 %add.i.i.i.i.i.i.i.i.i.i
+  %tmp20 = load float, float* %arrayidx.i.i.i.i.i.i.i.i.i.i, align 4
+  %mul.i.i.i.i.i.i.i = fmul float %tmp16, %tmp20
+  %add.i.i.i.i.i.i.i = fadd float %t.015.i.i.i.i.i.i.i, %mul.i.i.i.i.i.i.i
+  %inc.i.i.i.i.i.i.i = add nuw i64 %k.016.i.i.i.i.i.i.i, 1
+  %exitcond = icmp ne i64 %inc.i.i.i.i.i.i.i, %Asize2
+  br i1 %exitcond, label %KernelStmt, label %LoopexitStmt
+
+LoopexitStmt: ; preds = %KernelStmt
+  br label %ExitStmt
+
+ExitStmt: ; preds = %LoopexitStmt, %for.body5.i.i.i.i.i
+  %t.0.lcssa.i.i.i.i.i.i.i = phi float [ 0.000000e+00, %for.body5.i.i.i.i.i ], [ %add.i.i.i.i.i.i.i, %LoopexitStmt ]
+  store float %t.0.lcssa.i.i.i.i.i.i.i, float* %arrayidx.i.i.i.i.i.i.i.i, align 4
+  %inc.i.i.i.i.i = add nuw i64 %j.025.i.i.i.i.i, 1
+  %exitcond4 = icmp ne i64 %inc.i.i.i.i.i, %Bsize2
+  br i1 %exitcond4, label %for.body5.i.i.i.i.i, label %for.cond.cleanup4.i.i.i.i.i.loopexit
+
+_ZN5boost7numeric5ublas6matrixIfNS1_15basic_row_majorImlEENS1_15unbounded_arrayIfSaIfEEEEC2INS1_20matrix_matrix_binaryIS8_S8_NS1_18matrix_matrix_prodIS8_S8_fEEEEEERKNS1_17matrix_expressionIT_EE.exit.i.loopexit: ; preds = %for.cond.cleanup4.i.i.i.i.i
+  br label %_ZN5boost7numeric5ublas6matrixIfNS1_15basic_row_majorImlEENS1_15unbounded_arrayIfSaIfEEEEC2INS1_20matrix_matrix_binaryIS8_S8_NS1_18matrix_matrix_prodIS8_S8_fEEEEEERKNS1_17matrix_expressionIT_EE.exit.i
+
+_ZN5boost7numeric5ublas6matrixIfNS1_15basic_row_majorImlEENS1_15unbounded_arrayIfSaIfEEEEC2INS1_20matrix_matrix_binaryIS8_S8_NS1_18matrix_matrix_prodIS8_S8_fEEEEEERKNS1_17matrix_expressionIT_EE.exit.i: ; preds = %_ZN5boost7numeric5ublas6matrixIfNS1_15basic_row_majorImlEENS1_15unbounded_arrayIfSaIfEEEEC2INS1_20matrix_matrix_binaryIS8_S8_NS1_18matrix_matrix_prodIS8_S8_fEEEEEERKNS1_17matrix_expressionIT_EE.exit.i.loopexit, %_ZN5boost7numeric5ublas15unbounded_arrayIfSaIfEEC2EmRKS3_.exit.i.i
+  %cmp.i.i.i = icmp eq %"class.boost::numeric::ublas::matrix"* %temporary.i, %C
+  br i1 %cmp.i.i.i, label %_ZN5boost7numeric5ublas6matrixIfNS1_15basic_row_majorImlEENS1_15unbounded_arrayIfSaIfEEEEC2INS1_20matrix_matrix_binaryIS8_S8_NS1_18matrix_matrix_prodIS8_S8_fEEEEEERKNS1_17matrix_expressionIT_EE.exit.i.invoke.cont.i_crit_edge, label %if.then.i.i.i
+
+_ZN5boost7numeric5ublas6matrixIfNS1_15basic_row_majorImlEENS1_15unbounded_arrayIfSaIfEEEEC2INS1_20matrix_matrix_binaryIS8_S8_NS1_18matrix_matrix_prodIS8_S8_fEEEEEERKNS1_17matrix_expressionIT_EE.exit.i.invoke.cont.i_crit_edge: ; preds = %_ZN5boost7numeric5ublas6matrixIfNS1_15basic_row_majorImlEENS1_15unbounded_arrayIfSaIfEEEEC2INS1_20matrix_matrix_binaryIS8_S8_NS1_18matrix_matrix_prodIS8_S8_fEEEEEERKNS1_17matrix_expressionIT_EE.exit.i
+  %Asize13.pre = load i64, i64* %size_.i21.i.i, align 8
+  br label %invoke.cont.i
+
+if.then.i.i.i:                                    ; preds = %_ZN5boost7numeric5ublas6matrixIfNS1_15basic_row_majorImlEENS1_15unbounded_arrayIfSaIfEEEEC2INS1_20matrix_matrix_binaryIS8_S8_NS1_18matrix_matrix_prodIS8_S8_fEEEEEERKNS1_17matrix_expressionIT_EE.exit.i
+  %size1_.i.i.i = getelementptr inbounds %"class.boost::numeric::ublas::matrix", %"class.boost::numeric::ublas::matrix"* %C, i64 0, i32 0
+  %tmp21 = load i64, i64* %size1_.i.i.i, align 8
+  %tmp22 = load i64, i64* %size1_.i.i, align 8
+  store i64 %tmp22, i64* %size1_.i.i.i, align 8
+  store i64 %tmp21, i64* %size1_.i.i, align 8
+  %size2_.i.i.i = getelementptr inbounds %"class.boost::numeric::ublas::matrix", %"class.boost::numeric::ublas::matrix"* %C, i64 0, i32 1
+  %tmp23 = load i64, i64* %size2_.i.i.i, align 8
+  %tmp24 = load i64, i64* %size2_.i.i, align 8
+  store i64 %tmp24, i64* %size2_.i.i.i, align 8
+  store i64 %tmp23, i64* %size2_.i.i, align 8
+  %size_.i.i.i.i = getelementptr inbounds %"class.boost::numeric::ublas::matrix", %"class.boost::numeric::ublas::matrix"* %C, i64 0, i32 2, i32 1
+  %tmp25 = load i64, i64* %size_.i.i.i.i, align 8
+  %tmp26 = load i64, i64* %size_.i21.i.i, align 8
+  store i64 %tmp26, i64* %size_.i.i.i.i, align 8
+  store i64 %tmp25, i64* %size_.i21.i.i, align 8
+  %data_.i.i.i.i = getelementptr inbounds %"class.boost::numeric::ublas::matrix", %"class.boost::numeric::ublas::matrix"* %C, i64 0, i32 2, i32 2
+  %tmp27 = bitcast float** %data_.i.i.i.i to i64*
+  %tmp28 = load i64, i64* %tmp27, align 8
+  %tmp29 = bitcast float** %data_5.i.i.i to i64*
+  %Asize10 = load i64, i64* %tmp29, align 8
+  store i64 %Asize10, i64* %tmp27, align 8
+  store i64 %tmp28, i64* %tmp29, align 8
+  br label %invoke.cont.i
+
+invoke.cont.i:                                    ; preds = %_ZN5boost7numeric5ublas6matrixIfNS1_15basic_row_majorImlEENS1_15unbounded_arrayIfSaIfEEEEC2INS1_20matrix_matrix_binaryIS8_S8_NS1_18matrix_matrix_prodIS8_S8_fEEEEEERKNS1_17matrix_expressionIT_EE.exit.i.invoke.cont.i_crit_edge, %if.then.i.i.i
+  %Asize13 = phi i64 [ %Asize13.pre, %_ZN5boost7numeric5ublas6matrixIfNS1_15basic_row_majorImlEENS1_15unbounded_arrayIfSaIfEEEEC2INS1_20matrix_matrix_binaryIS8_S8_NS1_18matrix_matrix_prodIS8_S8_fEEEEEERKNS1_17matrix_expressionIT_EE.exit.i.invoke.cont.i_crit_edge ], [ %tmp25, %if.then.i.i.i ]
+  %tobool.i.i.i = icmp eq i64 %Asize13, 0
+  br i1 %tobool.i.i.i, label %_ZN5boost7numeric5ublas6matrixIfNS1_15basic_row_majorImlEENS1_15unbounded_arrayIfSaIfEEEEaSINS1_20matrix_matrix_binaryIS8_S8_NS1_18matrix_matrix_prodIS8_S8_fEEEEEERS8_RKNS1_17matrix_expressionIT_EE.exit, label %if.then.i.i4.i
+
+if.then.i.i4.i:                                   ; preds = %invoke.cont.i
+  %Asize14 = bitcast float** %data_5.i.i.i to i8**
+  %Asize15 = load i8*, i8** %Asize14, align 8
+  br label %_ZN5boost7numeric5ublas6matrixIfNS1_15basic_row_majorImlEENS1_15unbounded_arrayIfSaIfEEEEaSINS1_20matrix_matrix_binaryIS8_S8_NS1_18matrix_matrix_prodIS8_S8_fEEEEEERS8_RKNS1_17matrix_expressionIT_EE.exit
+
+_ZN5boost7numeric5ublas6matrixIfNS1_15basic_row_majorImlEENS1_15unbounded_arrayIfSaIfEEEEaSINS1_20matrix_matrix_binaryIS8_S8_NS1_18matrix_matrix_prodIS8_S8_fEEEEEERS8_RKNS1_17matrix_expressionIT_EE.exit: ; preds = %if.then.i.i4.i, %invoke.cont.i
+  ret void
+}