diff --git a/include/os/osMemory.h b/include/os/osMemory.h
index 14d53a7a066f60b5f0d3c0f37526b2533b3b05af..4681ff66741d4781c88dba5ccc950094c96f6769 100644
--- a/include/os/osMemory.h
+++ b/include/os/osMemory.h
@@ -37,6 +37,7 @@ void    taosMemoryFree(void *ptr);
 int64_t taosMemorySize(void *ptr);
 void    taosPrintBackTrace();
 void    taosMemoryTrim(int32_t size);
+void   *taosMemoryMallocAlign(uint32_t alignment, int64_t size);
 
 #define taosMemoryFreeClear(ptr)   \
   do {                             \
diff --git a/source/common/src/tdatablock.c b/source/common/src/tdatablock.c
index f5dba354405855cdc5a308dce7db91d6ac9c63cb..c79910978a2adadac8e52f413ab42693f9159517 100644
--- a/source/common/src/tdatablock.c
+++ b/source/common/src/tdatablock.c
@@ -19,6 +19,8 @@
 #include "tlog.h"
 #include "tname.h"
 
+#define MALLOC_ALIGN_BYTES  32
+
 int32_t colDataGetLength(const SColumnInfoData* pColumnInfoData, int32_t numOfRows) {
   ASSERT(pColumnInfoData != NULL);
   if (IS_VAR_DATA_TYPE(pColumnInfoData->info.type)) {
@@ -1163,6 +1165,7 @@ static int32_t doEnsureCapacity(SColumnInfoData* pColumn, const SDataBlockInfo*
     pColumn->varmeta.offset = (int32_t*)tmp;
     memset(&pColumn->varmeta.offset[existedRows], 0, sizeof(int32_t) * (numOfRows - existedRows));
   } else {
+    // prepare for the null bitmap
     char* tmp = taosMemoryRealloc(pColumn->nullbitmap, BitmapLen(numOfRows));
     if (tmp == NULL) {
       return TSDB_CODE_OUT_OF_MEMORY;
@@ -1173,11 +1176,19 @@ static int32_t doEnsureCapacity(SColumnInfoData* pColumn, const SDataBlockInfo*
     memset(&pColumn->nullbitmap[oldLen], 0, BitmapLen(numOfRows) - oldLen);
 
     ASSERT(pColumn->info.bytes);
-    tmp = taosMemoryRealloc(pColumn->pData, numOfRows * pColumn->info.bytes);
+
+    // make sure the allocated memory is MALLOC_ALIGN_BYTES aligned
+    tmp = taosMemoryMallocAlign(MALLOC_ALIGN_BYTES, numOfRows * pColumn->info.bytes);
     if (tmp == NULL) {
       return TSDB_CODE_OUT_OF_MEMORY;
     }
 
+    // copy back the existed data
+    if (pColumn->pData != NULL) {
+      memcpy(tmp, pColumn->pData, existedRows * pColumn->info.bytes);
+      taosMemoryFreeClear(pColumn->pData);
+    }
+
     pColumn->pData = tmp;
     if (clearPayload) {
       memset(tmp + pColumn->info.bytes * existedRows, 0, pColumn->info.bytes * (numOfRows - existedRows));
diff --git a/source/libs/function/src/detail/tminmax.c b/source/libs/function/src/detail/tminmax.c
index ed297e2b66e67f045922be8f986ca5a74e7c7e0d..a98b17287327a03d5e69179494314d69259691cc 100644
--- a/source/libs/function/src/detail/tminmax.c
+++ b/source/libs/function/src/detail/tminmax.c
@@ -271,67 +271,6 @@ static int16_t i16VectorCmpAVX2(const int16_t* pData, int32_t numOfRows, bool is
   return v;
 }
 
-//static int64_t i64VectorCmpAVX2(const int64_t* pData, int32_t numOfRows, bool isMinFunc) {
-//  int64_t        v = 0;
-//  const int32_t  bitWidth = 256;
-//  const int64_t* p = pData;
-//
-//  int32_t width = (bitWidth>>3u) / sizeof(int64_t);
-//  int32_t remain = numOfRows % width;
-//  int32_t rounds = numOfRows / width;
-//
-//#if __AVX2__
-//  __m256i next;
-//  __m256i initialVal = _mm256_loadu_si256((__m256i*)p);
-//  p += width;
-//
-//  if (!isMinFunc) {  // max function
-//    for (int32_t i = 0; i < rounds; ++i) {
-//      next = _mm256_lddqu_si256((__m256i*)p);
-//      initialVal = _mm256_max_epi64(initialVal, next);
-//      p += width;
-//    }
-//
-//    // let sum up the final results
-//    const int64_t* q = (const int64_t*)&initialVal;
-//    v = TMAX(q[0], q[1]);
-//    for(int32_t k = 1; k < width; ++k) {
-//      v = TMAX(v, q[k]);
-//    }
-//
-//    // calculate the front and the reminder items in array list
-//    int32_t start = rounds * width;
-//    for (int32_t j = 0; j < remain; ++j) {
-//      if (v < p[j + start]) {
-//        v = p[j + start];
-//      }
-//    }
-//  } else {  // min function
-//    for (int32_t i = 0; i < rounds; ++i) {
-//      next = _mm256_lddqu_si256((__m256i*)p);
-//      initialVal = _mm256_min_epi64(initialVal, next);
-//      p += width;
-//    }
-//
-//    // let sum up the final results
-//    const int64_t* q = (const int64_t*)&initialVal;
-//    v = TMIN(q[0], q[1]);
-//    for(int32_t k = 1; k < width; ++k) {
-//      v = TMIN(v, q[k]);
-//    }
-//
-//    // calculate the front and the remainder items in array list
-//    int32_t start = rounds * width;
-//    for (int32_t j = 0; j < remain; ++j) {
-//      if (v > p[j + start]) {
-//        v = p[j + start];
-//      }
-//    }
-//  }
-//#endif
-//
-//  return v;
-//}
 
 static int32_t handleInt32Col(SColumnInfoData* pCol, int32_t start, int32_t numOfRows, SqlFunctionCtx* pCtx,
                               SMinmaxResInfo* pBuf, bool isMinFunc) {
diff --git a/source/os/src/osMemory.c b/source/os/src/osMemory.c
index 78fa362179051fafe4adc3f618a2fdd08ed5ec10..1facff1f3be1e2c61af98b9a022470dd3f231ba6 100644
--- a/source/os/src/osMemory.c
+++ b/source/os/src/osMemory.c
@@ -345,3 +345,11 @@ void taosMemoryTrim(int32_t size) {
   malloc_trim(size);
 #endif
 }
+
+void* taosMemoryMallocAlign(uint32_t alignment, int64_t size) {
+#ifdef USE_TD_MEMORY
+  ASSERT(0);
+#else
+  return memalign(alignment, size);
+#endif
+}