From 5b9689ce553990517ab90aa23cb460980cb45ce2 Mon Sep 17 00:00:00 2001
From: Roman Nozdrin <roman.nozdrin@mariadb.com>
Date: Wed, 30 Dec 2020 11:35:51 +0000
Subject: [PATCH] MCOL-4478 MCS now rounds the last digits of an avg() result
 for wide-DECIMAL argument

---
 datatypes/mcs_decimal.cpp         | 21 +--------------------
 datatypes/mcs_decimal.h           | 17 +++++++++++++++++
 datatypes/mcs_int128.h            |  5 +++++
 utils/rowgroup/rowaggregation.cpp | 20 +++++++++++++++++---
 4 files changed, 40 insertions(+), 23 deletions(-)

diff --git a/datatypes/mcs_decimal.cpp b/datatypes/mcs_decimal.cpp
index 6df2061fa..29bc9cda2 100644
--- a/datatypes/mcs_decimal.cpp
+++ b/datatypes/mcs_decimal.cpp
@@ -22,24 +22,6 @@
 
 namespace datatypes
 {
-
-    struct lldiv_t_128
-    {
-        int128_t quot;
-        int128_t rem;
-        lldiv_t_128() : quot(0), rem(0) {}
-        lldiv_t_128(const int128_t& a_quot, const int128_t& a_rem)
-            : quot(a_quot), rem(a_rem) {}
-    };
-
-    inline lldiv_t_128 lldiv128(const int128_t& dividend, const int128_t& divisor)
-    {
-        if (UNLIKELY(divisor == 0) || UNLIKELY(dividend == 0))
-            return lldiv_t_128();
-
-        return lldiv_t_128(dividend / divisor, dividend % divisor);
-    }
-
     template<typename BinaryOperation,
         typename OpOverflowCheck,
         typename MultiplicationOverflowCheck>
@@ -115,7 +97,6 @@ namespace datatypes
             ? r.s128Value : r.value;
 
         opOverflowCheck(lValue, rValue);
-
         if (result.scale >= l.scale - r.scale)
         {
             int128_t scaleMultiplier;
@@ -213,7 +194,7 @@ namespace datatypes
             // rem carries the value's sign, but needs to be normalized.
             int64_t s = l.scale - r.scale;
             int128_t divisor;
-            getScaleDivisor(divisor, abs(s));
+            getScaleDivisor(divisor, std::abs(s));
 
             if (s < 0)
             {
diff --git a/datatypes/mcs_decimal.h b/datatypes/mcs_decimal.h
index 712bec55e..2d49e667a 100644
--- a/datatypes/mcs_decimal.h
+++ b/datatypes/mcs_decimal.h
@@ -161,6 +161,23 @@ inline void getScaleDivisor(T& divisor, const int8_t scale)
     }
 }
 
+struct lldiv_t_128
+{
+    int128_t quot;
+    int128_t rem;
+    lldiv_t_128() : quot(0), rem(0) {}
+    lldiv_t_128(const int128_t& a_quot, const int128_t& a_rem)
+        : quot(a_quot), rem(a_rem) {}
+};
+
+inline lldiv_t_128 lldiv128(const int128_t& dividend, const int128_t& divisor)
+{
+    if (UNLIKELY(divisor == 0) || UNLIKELY(dividend == 0))
+        return lldiv_t_128();
+
+    return lldiv_t_128(dividend / divisor, dividend % divisor);
+}
+
 // @brief The class for Decimal related operations
 // The class contains Decimal related operations are scale and
 // precision aware.
diff --git a/datatypes/mcs_int128.h b/datatypes/mcs_int128.h
index 154342952..d8bacd8a8 100644
--- a/datatypes/mcs_int128.h
+++ b/datatypes/mcs_int128.h
@@ -111,6 +111,11 @@ template<>
 struct is_uint128_t<uint128_t> {
   static const bool value = true;
 };
+
+inline int128_t abs(int128_t& x)
+{
+  return (x >= 0) ? x : -x;
+}
  
 class TSInt128
 {
diff --git a/utils/rowgroup/rowaggregation.cpp b/utils/rowgroup/rowaggregation.cpp
index 48b2e434f..d624c5cab 100755
--- a/utils/rowgroup/rowaggregation.cpp
+++ b/utils/rowgroup/rowaggregation.cpp
@@ -2788,7 +2788,7 @@ void RowAggregationUM::calculateAvgColumns()
                     uint32_t offset = fRow.getOffset(colOut);
                     uint32_t scale = fRow.getScale(colOut);
                     // Get multiplied to deliver AVG with the scale closest
-                    // to the expected original scale + 4. 
+                    // to the expected original scale + 4.
                     // There is a counterpart in buildAggregateColumn.
                     datatypes::Decimal::setScalePrecision4Avg(precision, scale);
                     int128_t* sumPnt = fRow.getBinaryField_offset<int128_t>(offset);
@@ -2800,8 +2800,22 @@ void RowAggregationUM::calculateAvgColumns()
                         multOp(*sumPnt, datatypes::mcs_pow_10[scaleDiff], sum);
                     else
                         sum = *sumPnt;
-                    int128_t avg = sum / cnt;
-                    fRow.setBinaryField_offset(&avg, sizeof(avg), offset);
+                    datatypes::lldiv_t_128 avgAndRem = datatypes::lldiv128(sum, cnt);
+                    // Round the last digit
+                    if (datatypes::abs(avgAndRem.rem) * 2 >= (int128_t)cnt)
+                    {
+                        if (utils::is_negative(avgAndRem.rem))
+                        {
+                            avgAndRem.quot--;
+                        }
+                        else
+                        {
+                            avgAndRem.quot++;
+                        }
+                    }
+                    fRow.setBinaryField_offset(&avgAndRem.quot,
+                                               sizeof(avgAndRem.quot),
+                                               offset);
                 }
             }
         }