Apply the "nodeAgg" optimization to more of the builtin transition

functions. This patch optimizes int2_sum(), int4_sum(), float4_accum()
and float8_accum() to avoid needing to copy the transition function's
state for each input tuple of the aggregate. In an extreme case
(e.g. SELECT sum(int2_col) FROM table where table has a single column),
it improves performance by about 20%. For more complex queries or tables
with wider rows, the relative performance improvement will not be as
significant.

diff --git a/src/backend/utils/adt/float.c b/src/backend/utils/adt/float.c
index e51babfebfb7f9cd801f2caac545faa09c3439ae..c943ee2c71d405ef5c9eb0a761f596ce2c6676dc 100644 (file)
--- a/src/backend/utils/adt/float.c
+++ b/src/backend/utils/adt/float.c
@@ -8,7 +8,7 @@
  *
  *
  * IDENTIFICATION
- *   $PostgreSQL: pgsql/src/backend/utils/adt/float.c,v 1.113 2005/02/11 04:08:58 neilc Exp $
+ *   $PostgreSQL: pgsql/src/backend/utils/adt/float.c,v 1.114 2005/04/06 23:56:07 neilc Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -1902,8 +1902,6 @@ float8_accum(PG_FUNCTION_ARGS)
    float8      N,
                sumX,
                sumX2;
-   Datum       transdatums[3];
-   ArrayType  *result;
 
    transvalues = check_float8_array(transarray, "float8_accum");
    N = transvalues[0];
@@ -1914,15 +1912,35 @@ float8_accum(PG_FUNCTION_ARGS)
    sumX += newval;
    sumX2 += newval * newval;
 
-   transdatums[0] = Float8GetDatumFast(N);
-   transdatums[1] = Float8GetDatumFast(sumX);
-   transdatums[2] = Float8GetDatumFast(sumX2);
+   /*
+    * If we're invoked by nodeAgg, we can cheat and modify our first
+    * parameter in-place to reduce palloc overhead. Otherwise we
+    * construct a new array with the updated transition data and
+    * return it.
+    */
+   if (fcinfo->context && IsA(fcinfo->context, AggState))
+   {
+       transvalues[0] = N;
+       transvalues[1] = sumX;
+       transvalues[2] = sumX2;
 
-   result = construct_array(transdatums, 3,
-                            FLOAT8OID,
-                        sizeof(float8), false /* float8 byval */ , 'd');
+       PG_RETURN_ARRAYTYPE_P(transarray);
+   }
+   else
+   {
+       Datum       transdatums[3];
+       ArrayType  *result;
+
+       transdatums[0] = Float8GetDatumFast(N);
+       transdatums[1] = Float8GetDatumFast(sumX);
+       transdatums[2] = Float8GetDatumFast(sumX2);
 
-   PG_RETURN_ARRAYTYPE_P(result);
+       result = construct_array(transdatums, 3,
+                                FLOAT8OID,
+                                sizeof(float8), false /* float8 byval */ , 'd');
+
+       PG_RETURN_ARRAYTYPE_P(result);
+   }
 }
 
 Datum
@@ -1935,8 +1953,6 @@ float4_accum(PG_FUNCTION_ARGS)
                sumX,
                sumX2,
                newval;
-   Datum       transdatums[3];
-   ArrayType  *result;
 
    transvalues = check_float8_array(transarray, "float4_accum");
    N = transvalues[0];
@@ -1950,15 +1966,35 @@ float4_accum(PG_FUNCTION_ARGS)
    sumX += newval;
    sumX2 += newval * newval;
 
-   transdatums[0] = Float8GetDatumFast(N);
-   transdatums[1] = Float8GetDatumFast(sumX);
-   transdatums[2] = Float8GetDatumFast(sumX2);
+   /*
+    * If we're invoked by nodeAgg, we can cheat and modify our first
+    * parameter in-place to reduce palloc overhead. Otherwise we
+    * construct a new array with the updated transition data and
+    * return it.
+    */
+   if (fcinfo->context && IsA(fcinfo->context, AggState))
+   {
+       transvalues[0] = N;
+       transvalues[1] = sumX;
+       transvalues[2] = sumX2;
 
-   result = construct_array(transdatums, 3,
-                            FLOAT8OID,
-                        sizeof(float8), false /* float8 byval */ , 'd');
+       PG_RETURN_ARRAYTYPE_P(transarray);
+   }
+   else
+   {
+       Datum       transdatums[3];
+       ArrayType  *result;
+
+       transdatums[0] = Float8GetDatumFast(N);
+       transdatums[1] = Float8GetDatumFast(sumX);
+       transdatums[2] = Float8GetDatumFast(sumX2);
 
-   PG_RETURN_ARRAYTYPE_P(result);
+       result = construct_array(transdatums, 3,
+                                FLOAT8OID,
+                                sizeof(float8), false /* float8 byval */ , 'd');
+
+       PG_RETURN_ARRAYTYPE_P(result);
+   }
 }
 
 Datum

diff --git a/src/backend/utils/adt/numeric.c b/src/backend/utils/adt/numeric.c
index a459ca2d6e42975699e52b26c6bc16e0ea8d25bb..214f9ff7a347919008342a94d3724a5e7d68653c 100644 (file)
--- a/src/backend/utils/adt/numeric.c
+++ b/src/backend/utils/adt/numeric.c
@@ -14,7 +14,7 @@
  * Copyright (c) 1998-2005, PostgreSQL Global Development Group
  *
  * IDENTIFICATION
- *   $PostgreSQL: pgsql/src/backend/utils/adt/numeric.c,v 1.82 2005/04/04 23:50:27 neilc Exp $
+ *   $PostgreSQL: pgsql/src/backend/utils/adt/numeric.c,v 1.83 2005/04/06 23:56:07 neilc Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -2357,7 +2357,6 @@ numeric_stddev(PG_FUNCTION_ARGS)
 Datum
 int2_sum(PG_FUNCTION_ARGS)
 {
-   int64       oldsum;
    int64       newval;
 
    if (PG_ARGISNULL(0))
@@ -2370,22 +2369,39 @@ int2_sum(PG_FUNCTION_ARGS)
        PG_RETURN_INT64(newval);
    }
 
-   oldsum = PG_GETARG_INT64(0);
+   /*
+    * If we're invoked by nodeAgg, we can cheat and modify out first
+    * parameter in-place to avoid palloc overhead. If not, we need to
+    * return the new value of the transition variable.
+    */
+   if (fcinfo->context && IsA(fcinfo->context, AggState))
+   {
+       int64 *oldsum = (int64 *) PG_GETARG_POINTER(0);
 
-   /* Leave sum unchanged if new input is null. */
-   if (PG_ARGISNULL(1))
-       PG_RETURN_INT64(oldsum);
+       /* Leave the running sum unchanged in the new input is null */
+       if (!PG_ARGISNULL(1))
+           *oldsum = *oldsum + (int64) PG_GETARG_INT16(1);
 
-   /* OK to do the addition. */
-   newval = oldsum + (int64) PG_GETARG_INT16(1);
+       PG_RETURN_POINTER(oldsum);
+   }
+   else
+   {
+       int64       oldsum = PG_GETARG_INT64(0);
+
+       /* Leave sum unchanged if new input is null. */
+       if (PG_ARGISNULL(1))
+           PG_RETURN_INT64(oldsum);
+
+       /* OK to do the addition. */
+       newval = oldsum + (int64) PG_GETARG_INT16(1);
 
-   PG_RETURN_INT64(newval);
+       PG_RETURN_INT64(newval);
+   }
 }
 
 Datum
 int4_sum(PG_FUNCTION_ARGS)
 {
-   int64       oldsum;
    int64       newval;
 
    if (PG_ARGISNULL(0))
@@ -2398,16 +2414,34 @@ int4_sum(PG_FUNCTION_ARGS)
        PG_RETURN_INT64(newval);
    }
 
-   oldsum = PG_GETARG_INT64(0);
+   /*
+    * If we're invoked by nodeAgg, we can cheat and modify out first
+    * parameter in-place to avoid palloc overhead. If not, we need to
+    * return the new value of the transition variable.
+    */
+   if (fcinfo->context && IsA(fcinfo->context, AggState))
+   {
+       int64 *oldsum = (int64 *) PG_GETARG_POINTER(0);
 
-   /* Leave sum unchanged if new input is null. */
-   if (PG_ARGISNULL(1))
-       PG_RETURN_INT64(oldsum);
+       /* Leave the running sum unchanged in the new input is null */
+       if (!PG_ARGISNULL(1))
+           *oldsum = *oldsum + (int64) PG_GETARG_INT32(1);
 
-   /* OK to do the addition. */
-   newval = oldsum + (int64) PG_GETARG_INT32(1);
+       PG_RETURN_POINTER(oldsum);
+   }
+   else
+   {
+       int64       oldsum = PG_GETARG_INT64(0);
 
-   PG_RETURN_INT64(newval);
+       /* Leave sum unchanged if new input is null. */
+       if (PG_ARGISNULL(1))
+           PG_RETURN_INT64(oldsum);
+
+       /* OK to do the addition. */
+       newval = oldsum + (int64) PG_GETARG_INT32(1);
+
+       PG_RETURN_INT64(newval);
+   }
 }
 
 Datum
@@ -2426,6 +2460,12 @@ int8_sum(PG_FUNCTION_ARGS)
        PG_RETURN_DATUM(newval);
    }
 
+   /*
+    * Note that we cannot special-case the nodeAgg case here, as we
+    * do for int2_sum and int4_sum: numeric is of variable size, so
+    * we cannot modify our first parameter in-place.
+    */
+
    oldsum = PG_GETARG_NUMERIC(0);
 
    /* Leave sum unchanged if new input is null. */