From f3fcac342ca3d6a216aa6b3dd4f4755b61da04c0 Mon Sep 17 00:00:00 2001 From: amit Date: Tue, 22 Aug 2017 13:48:13 +0900 Subject: [PATCH v48 3/4] Faster partition pruning This adds a new module partprune.c in the optimizer, which is meant as a replacement for using constraint exclusion to prune individual partitions. The new module performs partition pruning using the information contained in parent/partitioned table's boundinfo, after extracting clauses that involve partition keys. With the new module's functionality in place, set_append_rel_size() calls prune_append_rel_partitions() to get a Bitmapset of partitions that need to be scanned and processes only the partitions contained in the set. Authors: Amit Langote, David Rowley (david.rowley@2ndquadrant.com), Dilip Kumar (dilipbalaut@gmail.com) --- src/backend/catalog/partition.c | 1115 ++++++++++++++++ src/backend/nodes/copyfuncs.c | 37 + src/backend/nodes/nodeFuncs.c | 25 + src/backend/nodes/outfuncs.c | 28 + src/backend/nodes/readfuncs.c | 30 + src/backend/optimizer/path/allpaths.c | 28 + src/backend/optimizer/util/Makefile | 2 +- src/backend/optimizer/util/partprune.c | 1678 +++++++++++++++++++++++++ src/backend/optimizer/util/plancat.c | 44 +- src/backend/optimizer/util/relnode.c | 8 + src/include/catalog/partition.h | 25 + src/include/catalog/pg_opfamily.h | 3 + src/include/nodes/nodes.h | 3 + src/include/nodes/primnodes.h | 73 ++ src/include/nodes/relation.h | 4 + src/include/optimizer/partprune.h | 23 + src/test/regress/expected/inherit.out | 4 +- src/test/regress/expected/partition_prune.out | 282 ++++- src/test/regress/sql/partition_prune.sql | 39 +- src/tools/pgindent/typedefs.list | 7 + 20 files changed, 3385 insertions(+), 73 deletions(-) create mode 100644 src/backend/optimizer/util/partprune.c create mode 100644 src/include/optimizer/partprune.h diff --git a/src/backend/catalog/partition.c b/src/backend/catalog/partition.c index 39ee773d93..df45ce3b43 100644 --- a/src/backend/catalog/partition.c +++ b/src/backend/catalog/partition.c @@ -138,6 +138,23 @@ typedef struct PartitionRangeBound bool lower; /* this is the lower (vs upper) bound */ } PartitionRangeBound; +/* + * The following struct describes the result of performing one + * PartitionPruneStep. + */ +typedef struct PruneStepResult +{ + /* + * This contains the offsets of the bounds in a table's boundinfo, each of + * which is a bound whose corresponding partition is selected by a a given + * pruning step. + */ + Bitmapset *bound_offsets; + + /* Set if we need to scan the default and/or the null partition, resp. */ + bool scan_default; + bool scan_null; +} PruneStepResult; static Oid get_partition_parent_worker(Relation inhRel, Oid relid); static void get_partition_ancestors_worker(Relation inhRel, Oid relid, @@ -197,6 +214,23 @@ static int get_greatest_modulus(PartitionBoundInfo b); static uint64 compute_hash_value(int partnatts, FmgrInfo *partsupfunc, Datum *values, bool *isnull); +static PruneStepResult *perform_pruning_base_step(PartitionPruneContext *context, + PartitionPruneStepOp *opstep); +static PruneStepResult *perform_pruning_combine_step(PartitionPruneContext *context, + PartitionPruneStepCombine *cstep, + PruneStepResult **step_results); +static bool partkey_datum_from_expr(PartitionPruneContext *context, + Expr *expr, Datum *value); +static PruneStepResult *get_matching_hash_bounds(PartitionPruneContext *context, + int opstrategy, Datum *values, int nvalues, + FmgrInfo *partsupfunc, Bitmapset *nullkeys); +static PruneStepResult *get_matching_list_bounds(PartitionPruneContext *context, + int opstrategy, Datum value, int nvalues, + FmgrInfo *partsupfunc, Bitmapset *nullkeys); +static PruneStepResult *get_matching_range_bounds(PartitionPruneContext *context, + int opstrategy, Datum *values, int nvalues, + FmgrInfo *partsupfunc, Bitmapset *nullkeys); + /* * RelationBuildPartitionDesc * Form rel's partition descriptor @@ -1620,9 +1654,1090 @@ get_partition_qual_relid(Oid relid) return result; } +/* + * get_matching_partitions + * Determine partitions that survive partition pruning steps + * + * Returns a Bitmapset of indexes of surviving partitions. + */ +Bitmapset * +get_matching_partitions(PartitionPruneContext *context, + List *pruning_steps) +{ + Bitmapset *result; + int num_steps = list_length(pruning_steps), + i; + PruneStepResult **step_results, + *final_result; + ListCell *lc; + + /* If there are no pruning steps then all partitions match. */ + if (num_steps == 0) + return bms_add_range(NULL, 0, context->nparts - 1); + + /* + * Allocate space for individual pruning steps to store its result. Each + * slot will hold a PruneStepResult after performing a given pruning step. + * Later steps may use the result of one or more earlier steps. The + * result of applying all pruning steps is the value contained in the slot + * of the last pruning step. + */ + step_results = (PruneStepResult **) + palloc0(num_steps * sizeof(PruneStepResult *)); + foreach(lc, pruning_steps) + { + PartitionPruneStep *step = lfirst(lc); + + switch (nodeTag(step)) + { + case T_PartitionPruneStepOp: + step_results[step->step_id] = + perform_pruning_base_step(context, + (PartitionPruneStepOp *) step); + break; + + case T_PartitionPruneStepCombine: + step_results[step->step_id] = + perform_pruning_combine_step(context, + (PartitionPruneStepCombine *) step, + step_results); + break; + + default: + elog(ERROR, "invalid pruning step type: %d", + (int) nodeTag(step)); + } + } + + /* + * At this point we know the offsets of all the datums whose corresponding + * partitions need to be in the result, including special null-accepting + * and default partitions. Collect the actual partition indexes now. + */ + final_result = step_results[num_steps - 1]; + Assert(final_result != NULL); + i = -1; + result = NULL; + while ((i = bms_next_member(final_result->bound_offsets, i)) >= 0) + { + int partindex = context->boundinfo->indexes[i]; + + /* + * In range and hash partitioning cases, some slots may contain -1, + * indicating that no partition has been defined to accept a + * given range of data or for a given remainder, respectively. + * The default partition, if any, in case of range partitioning, will + * be added to the result, because the specified range still satisfies + * the query's conditions. + */ + if (partindex >= 0) + result = bms_add_member(result, partindex); + } + + /* Add the null and/or default partition if needed and if present. */ + if (final_result->scan_null) + { + Assert(context->strategy == PARTITION_STRATEGY_LIST); + if (partition_bound_accepts_nulls(context->boundinfo)) + result = bms_add_member(result, context->boundinfo->null_index); + } + if (final_result->scan_default) + { + Assert(context->strategy == PARTITION_STRATEGY_LIST || + context->strategy == PARTITION_STRATEGY_RANGE); + if (partition_bound_has_default(context->boundinfo)) + result = bms_add_member(result, + context->boundinfo->default_index); + } + + return result; +} + /* Module-local functions */ /* + * perform_pruning_base_step + * Determines the indexes of datums that satisfy conditions specified in + * 'opstep'. + * + * Result also contains whether special null-accepting and/or default + * partition need to be scanned. + */ +static PruneStepResult * +perform_pruning_base_step(PartitionPruneContext *context, + PartitionPruneStepOp *opstep) +{ + ListCell *lc1, + *lc2; + int keyno, + nvalues; + Datum values[PARTITION_MAX_KEYS]; + FmgrInfo partsupfunc[PARTITION_MAX_KEYS]; + + /* + * There better be the same number of expressions and compare functions. + */ + Assert(list_length(opstep->exprs) == list_length(opstep->cmpfns)); + + nvalues = 0; + lc1 = list_head(opstep->exprs); + lc2 = list_head(opstep->cmpfns); + + /* + * Generate the partition look-up key that will be used by one of + * the get_partitions_from_keys_* functions called below. + */ + for (keyno = 0; keyno < context->partnatts; keyno++) + { + /* + * For hash partitioning, it is possible that values of some keys are + * not provided in operator clauses, but instead the planner found + * that they appeared in a IS NULL clause. + */ + if (bms_is_member(keyno, opstep->nullkeys)) + continue; + + /* + * For range partitioning, we must only perform pruning with values + * for either all partition keys or a prefix thereof. + */ + if (keyno > nvalues && context->strategy == PARTITION_STRATEGY_RANGE) + break; + + if (lc1 != NULL) + { + Expr *expr; + Datum datum; + + expr = lfirst(lc1); + if (partkey_datum_from_expr(context, expr, &datum)) + { + Oid cmpfn; + + /* + * If we're going to need a different comparison function + * than the one cached in the PartitionKey, we'll need to + * look up the FmgrInfo. + */ + cmpfn = lfirst_oid(lc2); + Assert(OidIsValid(cmpfn)); + if (cmpfn != context->partsupfunc[keyno].fn_oid) + fmgr_info(cmpfn, &partsupfunc[keyno]); + else + fmgr_info_copy(&partsupfunc[keyno], + &context->partsupfunc[keyno], + CurrentMemoryContext); + + values[keyno] = datum; + nvalues++; + } + + lc1 = lnext(lc1); + lc2 = lnext(lc2); + } + } + + switch (context->strategy) + { + case PARTITION_STRATEGY_HASH: + return get_matching_hash_bounds(context, + opstep->opstrategy, + values, nvalues, + partsupfunc, + opstep->nullkeys); + + case PARTITION_STRATEGY_LIST: + return get_matching_list_bounds(context, + opstep->opstrategy, + values[0], nvalues, + &partsupfunc[0], + opstep->nullkeys); + + case PARTITION_STRATEGY_RANGE: + return get_matching_range_bounds(context, + opstep->opstrategy, + values, nvalues, + partsupfunc, + opstep->nullkeys); + + default: + elog(ERROR, "unexpected partition strategy: %d", + (int) context->strategy); + break; + } + + return NULL; +} + +/* + * perform_pruning_combine_step + * Determines the indexes of datums obtained by combining those given + * by the steps identified by cstep->source_stepids using the specified + * combination method + * + * Since cstep may refer to the result of earlier steps, we also receive + * step_results here. + */ +static PruneStepResult * +perform_pruning_combine_step(PartitionPruneContext *context, + PartitionPruneStepCombine *cstep, + PruneStepResult **step_results) +{ + ListCell *lc1; + PruneStepResult *result = NULL; + + /* + * In some cases, the planner generates a combine step that doesn't + * contain any argument steps, to signal us to not prune any partitions. + * So, return indexes of all datums in that case, including null and/or + * default partition, if any. + */ + result = (PruneStepResult *) palloc0(sizeof(PruneStepResult)); + if (list_length(cstep->source_stepids) == 0) + { + PartitionBoundInfo boundinfo = context->boundinfo; + + result->bound_offsets = bms_add_range(NULL, 0, boundinfo->ndatums - 1); + result->scan_default = partition_bound_has_default(boundinfo); + result->scan_null = partition_bound_accepts_nulls(boundinfo); + return result; + } + else + { + bool firststep; + + switch (cstep->combineOp) + { + case COMBINE_UNION: + foreach(lc1, cstep->source_stepids) + { + int step_id = lfirst_int(lc1); + PruneStepResult *step_result; + + /* + * step_results[step_id] must contain a valid result, + * which is confirmed by the fact that cstep's step_id is + * greater than step_id and the fact that results of the + * individual steps are evaluated in sequence of their + * step_ids. + */ + if (step_id >= cstep->step.step_id) + elog(ERROR, "invalid pruning combine step argument"); + step_result = step_results[step_id]; + Assert(step_result != NULL); + + /* Record any additional datum indexes from this step */ + result->bound_offsets = + bms_add_members(result->bound_offsets, + step_result->bound_offsets); + + /* Update whether to scan null and default partitions. */ + if (!result->scan_null) + result->scan_null = step_result->scan_null; + if (!result->scan_default) + result->scan_default = step_result->scan_default; + } + break; + + case COMBINE_INTERSECT: + firststep = true; + foreach(lc1, cstep->source_stepids) + { + int step_id = lfirst_int(lc1); + PruneStepResult *step_result; + + if (step_id >= cstep->step.step_id) + elog(ERROR, "invalid pruning combine step argument"); + step_result = step_results[step_id]; + Assert(step_result != NULL); + + if (firststep) + { + /* Copy step's result the first time. */ + result->bound_offsets = step_result->bound_offsets; + result->scan_null = step_result->scan_null; + result->scan_default = step_result->scan_default; + firststep = false; + } + else + { + /* Record datum indexes common to both steps */ + result->bound_offsets = + bms_int_members(result->bound_offsets, + step_result->bound_offsets); + + /* + * Update whether to scan null and default partitions. + */ + if (result->scan_null) + result->scan_null = step_result->scan_null; + if (result->scan_default) + result->scan_default = + step_result->scan_default; + } + } + break; + + default: + elog(ERROR, "invalid pruning combine op: %d", + (int) cstep->combineOp); + } + } + + return result; +} + +/* + * partkey_datum_from_expr + * Set *value to the constant value if 'expr' provides one + */ +static bool +partkey_datum_from_expr(PartitionPruneContext *context, + Expr *expr, Datum *value) +{ + switch (nodeTag(expr)) + { + case T_Const: + *value = ((Const *) expr)->constvalue; + return true; + + default: + break; + } + + return false; +} + +/* + * get_matching_hash_bounds + * Determine offset of the hash bound matching the specified value, + * considering that all the non-null values come from clauses containing + * a compatible hash eqaulity operator and any keys that are null come + * from an IS NULL clause + * + * In most cases, the result would contain just one bound's offset, although + * the set may be empty if the corresponding hash partition has not been + * created. Also, if insufficient number of values were provided, all bounds + * are returned. + * + * 'nvalues', if non-zero, denotes the number of values contained in 'values' + + * 'values' contains values to be used for pruning appearing in the array in + * respective partition key position. + + * 'opstrategy' if non-zero must be HTEqualStrategyNumber. + + * 'partsupfunc' contains partition hashing functions that can produce correct + * hash for the type of the values contained in 'values' + + * 'nullkeys' is the set of partition keys that are null. + */ +static PruneStepResult * +get_matching_hash_bounds(PartitionPruneContext *context, + int opstrategy, Datum *values, int nvalues, + FmgrInfo *partsupfunc, Bitmapset *nullkeys) +{ + PruneStepResult *result = palloc0(sizeof(PruneStepResult)); + PartitionBoundInfo boundinfo = context->boundinfo; + int *partindices = boundinfo->indexes; + int partnatts = context->partnatts; + bool isnull[PARTITION_MAX_KEYS]; + int i; + uint64 rowHash; + int greatest_modulus; + + Assert(context->strategy == PARTITION_STRATEGY_HASH); + + /* + * For hash partitioning we can only perform pruning based on equality + * clauses to the partition key or IS NULL clauses. We also can only + * prune if we got values for all keys. + */ + if (nvalues + bms_num_members(nullkeys) == partnatts) + { + /* + * If there are any values, they must have come from clauses + * containing an equality operator compatible with hash partitioning. + */ + Assert(opstrategy == HTEqualStrategyNumber || nvalues == 0); + + for (i = 0; i < partnatts; i++) + isnull[i] = bms_is_member(i, nullkeys); + + greatest_modulus = get_greatest_modulus(boundinfo); + rowHash = compute_hash_value(partnatts, partsupfunc, values, isnull); + + if (partindices[rowHash % greatest_modulus] >= 0) + result->bound_offsets = + bms_make_singleton(rowHash % greatest_modulus); + } + else + result->bound_offsets = bms_add_range(NULL, 0, + boundinfo->ndatums - 1); + + /* + * There is neither a special hash null partition or the default hash + * partition. + */ + result->scan_null = result->scan_default = false; + + return result; +} + +/* + * get_matching_list_bounds + * Determine the offsets of list bounds matching the specified value, + * according to the semantics of the given operator strategy + * + * 'nvalues', if non-zero, should be exactly 1, because of list partitioning. + + * 'value' contains the value to use for pruning + + * 'opstrategy' if non-zero must be a btree strategy number + + * 'partsupfunc' contains the list partitioning comparison function to be used + * to perform partition_list_bsearch + */ +static PruneStepResult * +get_matching_list_bounds(PartitionPruneContext *context, + int opstrategy, Datum value, int nvalues, + FmgrInfo *partsupfunc, Bitmapset *nullkeys) +{ + PruneStepResult *result = palloc0(sizeof(PruneStepResult)); + PartitionBoundInfo boundinfo = context->boundinfo; + int off, + minoff, + maxoff; + bool is_equal; + bool inclusive = false; + Oid *partcollation = context->partcollation; + + Assert(context->strategy == PARTITION_STRATEGY_LIST); + Assert(context->partnatts == 1); + + result->scan_null = result->scan_default = false; + + if (!bms_is_empty(nullkeys)) + { + /* + * Nulls may exist in only one partition - the partition whose + * accepted set of values includes null or the default partition if + * the former doesn't exist. + */ + if (partition_bound_accepts_nulls(boundinfo)) + result->scan_null = true; + else + result->scan_default = partition_bound_has_default(boundinfo); + return result; + } + + /* + * If there are no datums to compare keys with, but there are partitions, + * just return the default partition if one exists. + */ + if (boundinfo->ndatums == 0) + { + result->scan_default = partition_bound_has_default(boundinfo); + return result; + } + + minoff = 0; + maxoff = boundinfo->ndatums - 1; + + /* + * If there are no values to compare with the datums in boundinfo, it + * means the caller asked for partitions for all non-null datums. Add + * indexes of *all* partitions, including the default if any. + */ + if (nvalues == 0) + { + result->bound_offsets = bms_add_range(NULL, 0, + boundinfo->ndatums - 1); + result->scan_default = partition_bound_has_default(boundinfo); + return result; + } + + /* Speical case handling of values coming from a <> operator clause. */ + if (opstrategy == InvalidStrategy) + { + result->bound_offsets = bms_add_range(NULL, 0, + boundinfo->ndatums - 1);; + + off = partition_list_bsearch(partsupfunc, partcollation, boundinfo, + value, &is_equal); + if (off >= 0 && is_equal) + { + + /* All bounds except this one qualify. */ + Assert(boundinfo->indexes[off] >= 0); + result->bound_offsets = bms_del_member(result->bound_offsets, + off); + } + + /* Always include the default partition if any. */ + result->scan_default = partition_bound_has_default(boundinfo); + + return result; + } + + /* + * With range queries, always include the default list partition, because + * list partitions divide the key space in a discontinuous manner, not all + * values in the given range will have a partition assigned. This may not + * technically be true for some data types (e.g. integer types), however, + * we currently lack any sort of infrastructure to provide us with proofs + * that would allow us to do anything smarter here. + */ + if (opstrategy != BTEqualStrategyNumber) + result->scan_default = partition_bound_has_default(boundinfo); + + switch (opstrategy) + { + case BTEqualStrategyNumber: + off = partition_list_bsearch(partsupfunc, + partcollation, + boundinfo, value, + &is_equal); + if (off >= 0 && is_equal) + { + Assert(boundinfo->indexes[off] >= 0); + result->bound_offsets = bms_make_singleton(off); + } + else + result->scan_default = partition_bound_has_default(boundinfo); + return result; + + case BTGreaterEqualStrategyNumber: + inclusive = true; + /* fall through */ + case BTGreaterStrategyNumber: + off = partition_list_bsearch(partsupfunc, + partcollation, + boundinfo, value, + &is_equal); + if (off >= 0) + { + /* We don't want the matched datum to be in the result. */ + if (!is_equal || !inclusive) + off++; + } + else + { + /* + * This case means all partition bounds are greater, which in + * turn means that all partitions satisfy this key. + */ + off = 0; + } + + /* + * off is greater than the numbers of datums we have partitions + * for. The only possible partition that could contain a match is + * the default partition, but we must've set context->scan_default + * above anyway if one exists. + */ + if (off > boundinfo->ndatums - 1) + return result; + + minoff = off; + break; + + case BTLessEqualStrategyNumber: + inclusive = true; + /* fall through */ + case BTLessStrategyNumber: + off = partition_list_bsearch(partsupfunc, + partcollation, + boundinfo, value, + &is_equal); + if (off >= 0 && is_equal && !inclusive) + off--; + + /* + * off is smaller than the datums of all non-default partitions. + * The only possible partition that could contain a match is the + * default partition, but we must've set context->scan_default + * above anyway if one exists. + */ + if (off < 0) + return result; + + maxoff = off; + break; + + default: + elog(ERROR, "invalid strategy number %d", opstrategy); + break; + } + + result->bound_offsets = bms_add_range(NULL, minoff, maxoff); + return result; +} + +/* + * get_matching_range_datums + * Determine the offsets of range bounds matching the specified values, + * according to the semantics of the given operator strategy + * + * Each datum whose offset is in result is to be treated as the upper bound of + * the partition that will contain the desired values. + * + * If default partition needs to be scanned for given values, set scan_default + * in result if present. + * + * 'nvalues', if non-zero, should be <= context->partnatts - 1 + + * 'values' contains values for partition keys (or a prefix) to be used for + * pruning + + * 'opstrategy' if non-zero must be a btree strategy number + + * 'partsupfunc' contains the range partitioning comparison functions to be + * used to perform partition_range_datum_bsearch or partition_rbound_datum_cmp + * using. + * + * 'nullkeys' is the set of partition keys that are null. + */ +static PruneStepResult * +get_matching_range_bounds(PartitionPruneContext *context, + int opstrategy, Datum *values, int nvalues, + FmgrInfo *partsupfunc, Bitmapset *nullkeys) +{ + PruneStepResult *result = palloc0(sizeof(PruneStepResult)); + PartitionBoundInfo boundinfo = context->boundinfo; + Oid *partcollation = context->partcollation; + int partnatts = context->partnatts; + int *partindices = boundinfo->indexes; + int off, + minoff, + maxoff, + i; + bool is_equal; + bool inclusive = false; + + Assert(context->strategy == PARTITION_STRATEGY_RANGE); + Assert(nvalues <= partnatts); + + result->scan_null = result->scan_default = false; + + /* + * If there are no datums to compare keys with, or if we got an IS NULL + * clause just return the default partition, if it exists. + */ + if (boundinfo->ndatums == 0 || !bms_is_empty(nullkeys)) + { + result->scan_default = partition_bound_has_default(boundinfo); + return result; + } + + minoff = 0; + maxoff = boundinfo->ndatums; + + /* + * If there are no values to compare with the datums in boundinfo, it + * means the caller asked for partitions for all non-null datums. Add + * indexes of *all* partitions, including the default partition if one + * exists. + */ + if (nvalues == 0) + { + if (partindices[minoff] < 0) + minoff++; + if (partindices[maxoff] < 0) + maxoff--; + + result->scan_default = partition_bound_has_default(boundinfo); + result->bound_offsets = bms_add_range(NULL, minoff, maxoff); + + return result; + } + + /* + * If the query does not constrain all key columns, we'll need to scan the + * the default partition, if any. + */ + if (nvalues < partnatts) + result->scan_default = partition_bound_has_default(boundinfo); + + switch (opstrategy) + { + case BTEqualStrategyNumber: + /* + * Look for the smallest bound that is = look-up value. + */ + off = partition_range_datum_bsearch(partsupfunc, + partcollation, + boundinfo, + nvalues, values, + &is_equal); + + if (off >= 0 && is_equal) + { + if (nvalues == partnatts) + { + /* There can only be zero or one matching partition. */ + if (partindices[off + 1] >= 0) + result->bound_offsets = bms_make_singleton(off + 1); + else + result->scan_default = + partition_bound_has_default(boundinfo); + return result; + } + else + { + int saved_off = off; + + /* + * Since the look-up value contains only a prefix of keys, + * we must find other bounds that may also match the + * prefix. partition_range_datum_bsearch() returns the + * offset of one of them, find others by checking adjacent + * bounds. + */ + + /* + * First find greatest bound that's smaller than the + * look-up value. + */ + while (off >= 1) + { + int32 cmpval; + + cmpval = + partition_rbound_datum_cmp(partsupfunc, + partcollation, + boundinfo->datums[off - 1], + boundinfo->kind[off - 1], + values, nvalues); + if (cmpval != 0) + break; + off--; + } + + Assert(0 == + partition_rbound_datum_cmp(partsupfunc, + partcollation, + boundinfo->datums[off], + boundinfo->kind[off], + values, nvalues)); + /* + * We can treat off as the offset of the smallest bound to + * be included in the result, if we know it is the upper + * bound of the partition in which the look-up value could + * possibly exist. One case it couldn't is if the bound, + * or precisely the matched portion of its prefix, is not + * inclusive. + */ + if (boundinfo->kind[off][nvalues] == + PARTITION_RANGE_DATUM_MINVALUE) + off++; + + minoff = off; + + /* + * Now find smallest bound that's greater than the look-up + * value. + */ + off = saved_off; + while (off < boundinfo->ndatums - 1) + { + int32 cmpval; + + cmpval = partition_rbound_datum_cmp(partsupfunc, + partcollation, + boundinfo->datums[off + 1], + boundinfo->kind[off + 1], + values, nvalues); + if (cmpval != 0) + break; + off++; + } + + Assert(0 == + partition_rbound_datum_cmp(partsupfunc, + partcollation, + boundinfo->datums[off], + boundinfo->kind[off], + values, nvalues)); + + /* + * off + 1, then would be the offset of the greatest bound + * to be included in the result. + */ + maxoff = off + 1; + } + + /* + * Skip if minoff/maxoff are actually the upper bound of a + * un-assigned portion of values. + */ + if (partindices[minoff] < 0 && minoff < boundinfo->ndatums) + minoff++; + if (partindices[maxoff] < 0 && maxoff >= 1) + maxoff--; + + /* + * There may exist a range of values unassigned to any + * non-default partition between the datums at minoff and + * maxoff. Add the default partition in that case. + */ + if (partition_bound_has_default(boundinfo)) + { + for (i = minoff; i <= maxoff; i++) + { + if (partindices[i] < 0) + { + result->scan_default = true; + break; + } + } + } + + Assert(minoff >= 0 && maxoff >= 0); + result->bound_offsets = bms_add_range(NULL, minoff, maxoff); + return result; + } + else if (off >= 0) /* !is_equal */ + { + /* + * The look-up value falls in the range between some bounds in + * boundinfo. 'off' would be the offset of the greatest + * bound that is <= look-up value, so add off + 1 to the + * result instead as the offset of the upper bound of the + * only partition that may contain the look-up value. + */ + if (partindices[off + 1] >= 0) + result->bound_offsets = bms_make_singleton(off + 1); + else + result->scan_default = + partition_bound_has_default(boundinfo); + return result; + } + /* + * off < 0, meaning the look-up value is smaller that all bounds, + * so only the default partition, if any, qualifies. + */ + else + result->scan_default = partition_bound_has_default(boundinfo); + return result; + + case BTGreaterEqualStrategyNumber: + inclusive = true; + /* fall through */ + case BTGreaterStrategyNumber: + /* + * Look for the smallest bound that is > or >= look-up value + * and set minoff to its offset. + */ + off = partition_range_datum_bsearch(partsupfunc, + partcollation, + boundinfo, + nvalues, values, + &is_equal); + if (off < 0) + { + /* + * All bounds are greater than the look-up value, so include + * all of them in the result. + */ + minoff = 0; + } + else + { + if (is_equal && nvalues < partnatts) + { + /* + * Since the look-up value contains only a prefix of keys, + * we must find other bounds that may also match the + * prefix. partition_range_datum_bsearch() returns the + * offset of one of them, find others by checking adjacent + * bounds. + * + * Based on whether the look-up values is inclusive or + * not, we must either include the indexes of all such + * bounds in the result (that is, set minoff to the index + * of smallest such bound) or find the smallest one that's + * greater than the look-up value and set minoff to that. + */ + while (off >= 1 && off < boundinfo->ndatums - 1) + { + int32 cmpval; + int nextoff; + + nextoff = inclusive ? off - 1 : off + 1; + cmpval = + partition_rbound_datum_cmp(partsupfunc, + partcollation, + boundinfo->datums[nextoff], + boundinfo->kind[nextoff], + values, nvalues); + if (cmpval != 0) + break; + + off = nextoff; + } + + Assert(0 == + partition_rbound_datum_cmp(partsupfunc, + partcollation, + boundinfo->datums[off], + boundinfo->kind[off], + values, nvalues)); + + minoff = inclusive ? off : off + 1; + } + /* + * Look-up value falls in the range between some bounds in + * boundinfo. off would be the offset of the greatest + * bound that is <= look-up value, so add off + 1 to the + * result instead as the offset of the upper bound of the + * smallest partition that may contain the look-up value. + */ + else + minoff = off + 1; + } + break; + + case BTLessEqualStrategyNumber: + inclusive = true; + /* fall through */ + case BTLessStrategyNumber: + /* + * Look for the greatest bound that is < or <= look-up value + * and set minoff to its offset. + */ + off = partition_range_datum_bsearch(partsupfunc, + partcollation, + boundinfo, + nvalues, values, + &is_equal); + if (off < 0) + { + /* + * All bounds are greater than the key, so we could only + * expect to find the look-up key in the default partition. + */ + result->scan_default = partition_bound_has_default(boundinfo); + return result; + } + else + { + /* + * See the comment above. + */ + if (is_equal && nvalues < partnatts) + { + while (off >= 1 && off < boundinfo->ndatums - 1) + { + int32 cmpval; + int nextoff; + + nextoff = inclusive ? off + 1 : off - 1; + cmpval = partition_rbound_datum_cmp(partsupfunc, + partcollation, + boundinfo->datums[nextoff], + boundinfo->kind[nextoff], + values, nvalues); + if (cmpval != 0) + break; + + off = nextoff; + } + + Assert(0 == + partition_rbound_datum_cmp(partsupfunc, + partcollation, + boundinfo->datums[off], + boundinfo->kind[off], + values, nvalues)); + + maxoff = inclusive ? off + 1: off; + } + /* + * The look-up value falls in the range between some bounds in + * boundinfo. 'off' would be the offset of the greatest + * bound that is <= look-up value, so add off + 1 to the + * result instead as the offset of the upper bound of the + * greatest partition that may contain look-up value. If + * the look-up value had exactly matched the bound, but it + * isn't inclusive, no need add the adjacent partition. + */ + else if (!is_equal || inclusive) + maxoff = off + 1; + else + maxoff = off; + } + break; + + default: + elog(ERROR, "invalid strategy number %d", opstrategy); + break; + } + + /* + * Skip a gap and when doing so, check if the bound contains a finite + * value to decide if we need to add the default partition. If it's an + * infinite bound, we need not add the default partition, as having an + * infinite bound means the partition in question catches any values + * that would otherwise be in the default partition. + */ + if (partindices[minoff] < 0) + { + int lastkey = nvalues - 1; + + if (minoff >= 0 && + minoff < boundinfo->ndatums && + boundinfo->kind[minoff][lastkey] == + PARTITION_RANGE_DATUM_VALUE) + result->scan_default = partition_bound_has_default(boundinfo); + + minoff++; + } + + /* + * Skip a gap. See the above comment about how we decide whether or + * or not to scan the default partition based whether the datum that + * will become the maximum datum is finite or not. + */ + if (maxoff >= 1 && partindices[maxoff] < 0) + { + int lastkey = nvalues - 1; + + if (maxoff >= 0 && + maxoff <= boundinfo->ndatums && + boundinfo->kind[maxoff - 1][lastkey] == + PARTITION_RANGE_DATUM_VALUE) + result->scan_default = partition_bound_has_default(boundinfo); + + maxoff--; + } + + if (partition_bound_has_default(boundinfo)) + { + /* + * There may exist a range of values unassigned to any non-default + * partition between the datums at minoff and maxoff. Add the default + * partition in that case. + */ + for (i = minoff; i <= maxoff; i++) + { + if (partindices[i] < 0) + { + result->scan_default = true; + break; + } + } + } + + Assert(minoff >= 0 && maxoff >= 0); + if (minoff > maxoff) + return result; + result->bound_offsets = bms_add_range(NULL, minoff, maxoff); + + return result; +} + +/* * get_partition_operator * * Return oid of the operator of given strategy for a given partition key diff --git a/src/backend/nodes/copyfuncs.c b/src/backend/nodes/copyfuncs.c index c3efca3c45..450c64d6fc 100644 --- a/src/backend/nodes/copyfuncs.c +++ b/src/backend/nodes/copyfuncs.c @@ -2136,6 +2136,37 @@ _copyOnConflictExpr(const OnConflictExpr *from) return newnode; } +/* + * _copyPartitionPruneStepOp + */ +static PartitionPruneStepOp * +_copyPartitionPruneStepOp(const PartitionPruneStepOp *from) +{ + PartitionPruneStepOp *newnode = makeNode(PartitionPruneStepOp); + + COPY_SCALAR_FIELD(step.step_id); + COPY_SCALAR_FIELD(opstrategy); + COPY_NODE_FIELD(exprs); + COPY_NODE_FIELD(cmpfns); + COPY_BITMAPSET_FIELD(nullkeys); + + return newnode; +} + +/* + * _copyPartitionPruneStepCombine + */ +static PartitionPruneStepCombine * +_copyPartitionPruneStepCombine(const PartitionPruneStepCombine *from) +{ + PartitionPruneStepCombine *newnode = makeNode(PartitionPruneStepCombine); + COPY_SCALAR_FIELD(step.step_id); + COPY_SCALAR_FIELD(combineOp); + COPY_NODE_FIELD(source_stepids); + + return newnode; +} + /* **************************************************************** * relation.h copy functions * @@ -5059,6 +5090,12 @@ copyObjectImpl(const void *from) case T_OnConflictExpr: retval = _copyOnConflictExpr(from); break; + case T_PartitionPruneStepOp: + retval = _copyPartitionPruneStepOp(from); + break; + case T_PartitionPruneStepCombine: + retval = _copyPartitionPruneStepCombine(from); + break; /* * RELATION NODES diff --git a/src/backend/nodes/nodeFuncs.c b/src/backend/nodes/nodeFuncs.c index 4157e7eb9a..c3f1789ce2 100644 --- a/src/backend/nodes/nodeFuncs.c +++ b/src/backend/nodes/nodeFuncs.c @@ -2156,6 +2156,17 @@ expression_tree_walker(Node *node, return true; } break; + case T_PartitionPruneStepOp: + { + PartitionPruneStepOp *opstep = (PartitionPruneStepOp *) node; + + if (walker((Node *) opstep->exprs, context)) + return true; + } + break; + case T_PartitionPruneStepCombine: + /* no expression subnodes */ + break; case T_JoinExpr: { JoinExpr *join = (JoinExpr *) node; @@ -2958,6 +2969,20 @@ expression_tree_mutator(Node *node, return (Node *) newnode; } break; + case T_PartitionPruneStepOp: + { + PartitionPruneStepOp *opstep = (PartitionPruneStepOp *) node; + PartitionPruneStepOp *newnode; + + FLATCOPY(newnode, opstep, PartitionPruneStepOp); + MUTATE(newnode->exprs, opstep->exprs, List *); + + return (Node *) newnode; + } + break; + case T_PartitionPruneStepCombine: + /* no expression sub-nodes */ + return (Node *) copyObject(node); case T_JoinExpr: { JoinExpr *join = (JoinExpr *) node; diff --git a/src/backend/nodes/outfuncs.c b/src/backend/nodes/outfuncs.c index c8d962670e..efd0a71a2c 100644 --- a/src/backend/nodes/outfuncs.c +++ b/src/backend/nodes/outfuncs.c @@ -1710,6 +1710,28 @@ _outFromExpr(StringInfo str, const FromExpr *node) } static void +_outPartitionPruneStepOp(StringInfo str, const PartitionPruneStepOp *node) +{ + WRITE_NODE_TYPE("PARTITIONPRUNESTEPOP"); + + WRITE_INT_FIELD(step.step_id); + WRITE_INT_FIELD(opstrategy); + WRITE_NODE_FIELD(exprs); + WRITE_NODE_FIELD(cmpfns); + WRITE_BITMAPSET_FIELD(nullkeys); +} + +static void +_outPartitionPruneStepCombine(StringInfo str, const PartitionPruneStepCombine *node) +{ + WRITE_NODE_TYPE("PARTITIONPRUNESTEPCOMBINE"); + + WRITE_INT_FIELD(step.step_id); + WRITE_ENUM_FIELD(combineOp, PartitionPruneCombineOp); + WRITE_NODE_FIELD(source_stepids); +} + +static void _outOnConflictExpr(StringInfo str, const OnConflictExpr *node) { WRITE_NODE_TYPE("ONCONFLICTEXPR"); @@ -3958,6 +3980,12 @@ outNode(StringInfo str, const void *obj) case T_OnConflictExpr: _outOnConflictExpr(str, obj); break; + case T_PartitionPruneStepOp: + _outPartitionPruneStepOp(str, obj); + break; + case T_PartitionPruneStepCombine: + _outPartitionPruneStepCombine(str, obj); + break; case T_Path: _outPath(str, obj); break; diff --git a/src/backend/nodes/readfuncs.c b/src/backend/nodes/readfuncs.c index 4518fa0cdb..25874074a0 100644 --- a/src/backend/nodes/readfuncs.c +++ b/src/backend/nodes/readfuncs.c @@ -1331,6 +1331,32 @@ _readOnConflictExpr(void) READ_DONE(); } +static PartitionPruneStepOp * +_readPartitionPruneStepOp(void) +{ + READ_LOCALS(PartitionPruneStepOp); + + READ_INT_FIELD(step.step_id); + READ_INT_FIELD(opstrategy); + READ_NODE_FIELD(exprs); + READ_NODE_FIELD(cmpfns); + READ_BITMAPSET_FIELD(nullkeys); + + READ_DONE(); +} + +static PartitionPruneStepCombine * +_readPartitionPruneStepCombine(void) +{ + READ_LOCALS(PartitionPruneStepCombine); + + READ_INT_FIELD(step.step_id); + READ_ENUM_FIELD(combineOp, PartitionPruneCombineOp); + READ_NODE_FIELD(source_stepids); + + READ_DONE(); +} + /* * Stuff from parsenodes.h. */ @@ -2596,6 +2622,10 @@ parseNodeString(void) return_value = _readFromExpr(); else if (MATCH("ONCONFLICTEXPR", 14)) return_value = _readOnConflictExpr(); + else if (MATCH("PARTITIONPRUNESTEPOP", 20)) + return_value = _readPartitionPruneStepOp(); + else if (MATCH("PARTITIONPRUNESTEPCOMBINE", 25)) + return_value = _readPartitionPruneStepCombine(); else if (MATCH("RTE", 3)) return_value = _readRangeTblEntry(); else if (MATCH("RANGETBLFUNCTION", 16)) diff --git a/src/backend/optimizer/path/allpaths.c b/src/backend/optimizer/path/allpaths.c index c4e4db15a6..fd89c7cfee 100644 --- a/src/backend/optimizer/path/allpaths.c +++ b/src/backend/optimizer/path/allpaths.c @@ -33,6 +33,7 @@ #include "optimizer/clauses.h" #include "optimizer/cost.h" #include "optimizer/geqo.h" +#include "optimizer/partprune.h" #include "optimizer/pathnode.h" #include "optimizer/paths.h" #include "optimizer/plancat.h" @@ -874,6 +875,8 @@ set_append_rel_size(PlannerInfo *root, RelOptInfo *rel, double *parent_attrsizes; int nattrs; ListCell *l; + Relids live_children = NULL; + bool did_pruning = false; /* Guard against stack overflow due to overly deep inheritance tree. */ check_stack_depth(); @@ -881,6 +884,20 @@ set_append_rel_size(PlannerInfo *root, RelOptInfo *rel, Assert(IS_SIMPLE_REL(rel)); /* + * If the partitioned relation has any baserestrictinfo quals then we + * attempt to use these quals to prune away partitions that cannot + * possibly contain any tuples matching these quals. In this case we'll + * store the relids of all partitions which could possibly contain a + * matching tuple, and skip anything else in the loop below. + */ + if (rte->relkind == RELKIND_PARTITIONED_TABLE && + rel->baserestrictinfo != NIL) + { + live_children = prune_append_rel_partitions(rel); + did_pruning = true; + } + + /* * Initialize to compute size estimates for whole append relation. * * We handle width estimates by weighting the widths of different child @@ -1128,6 +1145,17 @@ set_append_rel_size(PlannerInfo *root, RelOptInfo *rel, continue; } + if (did_pruning && + !bms_is_member(appinfo->child_relid, live_children)) + { + /* + * Partition pruning determined this partition cannot possibly + * contain any tuples matching the baserestrictinfo, so skip it. + */ + set_dummy_rel_pathlist(childrel); + continue; + } + if (relation_excluded_by_constraints(root, childrel, childRTE)) { /* diff --git a/src/backend/optimizer/util/Makefile b/src/backend/optimizer/util/Makefile index c54d0a690d..aebd98875e 100644 --- a/src/backend/optimizer/util/Makefile +++ b/src/backend/optimizer/util/Makefile @@ -12,7 +12,7 @@ subdir = src/backend/optimizer/util top_builddir = ../../../.. include $(top_builddir)/src/Makefile.global -OBJS = clauses.o joininfo.o orclauses.o pathnode.o placeholder.o \ +OBJS = clauses.o joininfo.o orclauses.o partprune.o pathnode.o placeholder.o \ plancat.o predtest.o relnode.o restrictinfo.o tlist.o var.o include $(top_srcdir)/src/backend/common.mk diff --git a/src/backend/optimizer/util/partprune.c b/src/backend/optimizer/util/partprune.c new file mode 100644 index 0000000000..ab6390234b --- /dev/null +++ b/src/backend/optimizer/util/partprune.c @@ -0,0 +1,1678 @@ +/*------------------------------------------------------------------------- + * + * partprune.c + * Provides the functionality to match the provided set of clauses with + * the partition key to partition pruning "steps" + * + * If the "steps" contain enough information, partitions can be pruned right + * away in this module, which is called "static pruning", as all the needed + * information is statically available in the query being planned. Otherwise, + * they'd need to be delivered to the executor where the missing information + * can be filled and pruning tried one more time, which would be called + * "dynamic pruning". + * + * Portions Copyright (c) 1996-2018, PostgreSQL Global Development Group + * Portions Copyright (c) 1994, Regents of the University of California + * + * IDENTIFICATION + * src/backend/optimizer/util/partprune.c + * + *------------------------------------------------------------------------- +*/ + +#include "postgres.h" + +#include "access/hash.h" +#include "access/nbtree.h" +#include "catalog/pg_operator.h" +#include "catalog/pg_opfamily.h" +#include "catalog/pg_type.h" +#include "miscadmin.h" +#include "nodes/makefuncs.h" +#include "nodes/nodeFuncs.h" +#include "optimizer/clauses.h" +#include "optimizer/partprune.h" +#include "optimizer/planner.h" +#include "optimizer/predtest.h" +#include "optimizer/prep.h" +#include "parser/parse_coerce.h" +#include "parser/parsetree.h" +#include "rewrite/rewriteManip.h" +#include "utils/lsyscache.h" + +/* + * Information about a clause matched with a partition key. + */ +typedef struct PartClauseInfo +{ + int keyno; /* Partition key number (0 to partnatts - 1) */ + Oid opno; /* operator used to compare partkey to 'expr' */ + bool isopne; /* is clause's original operator <> ? */ + Expr *expr; /* The expr the partition key is being + * compared to */ + Oid cmpfn; /* Oid of function to compare 'expr' to the + * partition key */ + + /* cached info. */ + int op_strategy; +} PartClauseInfo; + +/* + * PartClauseMatchStatus + * Describes the result match_clause_to_partition_key produces for a + * given clause and the partition key to match with that are passed to it + */ +typedef enum PartClauseMatchStatus +{ + PARTCLAUSE_NOMATCH, + PARTCLAUSE_MATCH_CLAUSE, + PARTCLAUSE_MATCH_NULLNESS, + PARTCLAUSE_MATCH_STEPS, + PARTCLAUSE_MATCH_CONTRADICT, + PARTCLAUSE_UNSUPPORTED +} PartClauseMatchStatus; + +/* + * GeneratePruningStepsContext + * Information about the current state of generation of "pruning steps" + * for a given set of clauses + * + * generate_partition_pruning_steps() initializes an instance of this struct, + * which is used throughout the step generation process. + */ +typedef struct GeneratePruningStepsContext +{ + bool static_pruning; + int next_step_id; + List *steps; +} GeneratePruningStepsContext; + +static List *generate_partition_pruning_steps_internal(RelOptInfo *rel, + GeneratePruningStepsContext *context, + List *clauses, + bool *constfalse); +static PartClauseMatchStatus match_clause_to_partition_key(RelOptInfo *rel, + GeneratePruningStepsContext *context, + Expr *clause, Expr *partkey, int partkeyidx, + bool *key_is_null, bool *key_is_not_null, + PartClauseInfo **pc, List **clause_steps); +static bool match_boolean_partition_clause(Oid partopfamily, Expr *clause, + Expr *partkey, Expr **rightop); +static PartitionPruneStep *generate_pruning_steps_from_opexprs( + PartitionScheme part_scheme, + GeneratePruningStepsContext *context, + List **keyclauses, + Bitmapset *nullkeys); +static List *get_steps_using_prefix(GeneratePruningStepsContext *context, + int step_opstrategy, + bool step_isopne, + Expr *step_lastexpr, + Oid step_lastcmpfn, + int step_lastkeyno, + Bitmapset *step_nullkeys, + List *prefix); +static List *get_steps_using_prefix_recurse(GeneratePruningStepsContext *context, + int step_opstrategy, + bool step_isopne, + Expr *step_lastexpr, + Oid step_lastcmpfn, + int step_lastkeyno, + Bitmapset *step_nullkeys, + ListCell *start, + List *step_exprs, + List *step_cmpfns); +static PartitionPruneStep *generate_pruning_step_op(GeneratePruningStepsContext *context, + int opstrategy, bool isopne, + List *exprs, List *cmpfns, Bitmapset *nullkeys); +static PartitionPruneStep *generate_pruning_step_combine(GeneratePruningStepsContext *context, + List *source_stepids, + PartitionPruneCombineOp combineOp); + +/* + * prune_append_rel_partitions + * Returns RT indexes of the minimum set of child partitions which must + * be scanned to satisfy rel's baserestrictinfo quals. + * + * Callers must ensure that 'rel' is a partitioned table. + */ +Relids +prune_append_rel_partitions(RelOptInfo *rel) +{ + Relids result = NULL; + List *clauses = rel->baserestrictinfo; + List *pruning_steps; + bool constfalse; + + Assert(clauses != NIL); + Assert(rel->part_scheme != NULL); + + /* Quick exit. */ + if (rel->nparts == 0) + return NULL; + + /* process clauses */ + pruning_steps = generate_partition_pruning_steps(rel, clauses, true, + &constfalse); + + if (!constfalse) + { + /* Actual pruning happens here. */ + PartitionPruneContext context; + Bitmapset *partindexes; + int i; + + /* Initiate partition pruning using clauses. */ + memset(&context, 0, sizeof(context)); + context.strategy = rel->part_scheme->strategy; + context.partnatts = rel->part_scheme->partnatts; + context.partopfamily = rel->part_scheme->partopfamily; + context.partopcintype = rel->part_scheme->partopcintype; + context.partcollation = rel->part_scheme->partcollation; + context.partsupfunc = rel->part_scheme->partsupfunc; + context.nparts = rel->nparts; + context.boundinfo = rel->boundinfo; + + partindexes = get_matching_partitions(&context, pruning_steps); + + /* Add selected partitions' RT indexes to result. */ + i = -1; + while ((i = bms_next_member(partindexes, i)) >= 0) + result = bms_add_member(result, rel->part_rels[i]->relid); + } + + return result; +} + +/* + * generate_partition_pruning_steps + * Processes 'clauses' and returns a list of "partition pruning steps" + * + * If static_pruning is true, include in the result only steps that contain at + * least one Const. If any of the clause in the input list is a + * pseudo-constant "false", *constfalse is set to true upon return. + */ +List * +generate_partition_pruning_steps(RelOptInfo *rel, List *clauses, + bool static_pruning, bool *constfalse) +{ + GeneratePruningStepsContext context; + + context.static_pruning = static_pruning; + context.next_step_id = 0; + context.steps = NIL; + + /* The clauses list may be modified below, so better make a copy. */ + clauses = list_copy(clauses); + + /* + * For sub-partitioned tables there's a corner case where if the + * sub-partitioned table shares any partition keys with its parent, then + * it's possible that the partitioning hierarchy allows the parent + * partition to only contain a narrower range of values than the + * sub-partitioned table does. In this case it is possible that we'd + * include partitions that could not possibly have any tuples matching + * 'clauses'. The possibility of such a partition arrangement is perhaps + * unlikely for non-default partitions, but it may be more likely in the + * case of default partitions, so we'll add the parent partition table's + * partition qual to the clause list in this case only. This may result + * in the default partition being eliminated. + */ + if (rel->has_default_part && rel->partition_qual != NIL) + { + List *partqual = rel->partition_qual; + + partqual = (List *) expression_planner((Expr *) partqual); + + /* Fix Vars to have the desired varno */ + if (rel->relid != 1) + ChangeVarNodes((Node *) partqual, 1, rel->relid, 0); + + clauses = list_concat(clauses, partqual); + } + + /* Down into the rabbit-hole. */ + (void) generate_partition_pruning_steps_internal(rel, &context, clauses, + constfalse); + + return context.steps; +} + +/* Module-local functions */ + +/* + * generate_partition_pruning_steps_internal + * Processes 'clauses' to generate partition pruning steps. + * + * From OpExpr clauses that are mutually AND'd, we find combinations of those + * that match to the partition key columns and for every such combination, + * we emit a PartitionPruneStepOp containing a vector of expressions whose + * values are used as a look up key to search partitions by comparing the + * values with partition bounds. Relevant details of the operator and a + * vector of (possibly cross-type) comparison functions is also included with + * each step. + * + * For BoolExpr clauses, we recursively generate steps for each of its + * arguments and generate PartitionPruneStepCombine step that will combine + * results of those steps. + * + * All of the generated steps are added to the global array context->steps and + * each one gets an identifier which is unique across all recursive + * invocations. + * + * If when going through clauses, we find any that are marked as pseudoconstant + * and contains a constant false value, we stop generating any further steps + * and simply return NIL (that is, no pruning steps) after setting *constfalse + * to true. The caller should consider all partitions as pruned in that case. + * We may do the same if we find that mutually contradictory clauses are + * present, but were not turned into a pseudoconstant at higher levels. + * + * Note: the 'clauses' List may be modified inside this function. Callers may + * like to make a copy of it before passing them to this function. + */ +static List * +generate_partition_pruning_steps_internal(RelOptInfo *rel, + GeneratePruningStepsContext *context, + List *clauses, + bool *constfalse) +{ + PartitionScheme part_scheme = rel->part_scheme; + List *keyclauses[PARTITION_MAX_KEYS]; + Bitmapset *nullkeys = NULL, + *notnullkeys = NULL; + bool generate_opsteps = false; + List *result = NIL; + ListCell *lc; + + *constfalse = false; + memset(keyclauses, 0, sizeof(keyclauses)); + foreach(lc, clauses) + { + Expr *clause = (Expr *) lfirst(lc); + int i; + + if (IsA(clause, RestrictInfo)) + { + RestrictInfo *rinfo = (RestrictInfo *) clause; + + clause = rinfo->clause; + if (rinfo->pseudoconstant && + !DatumGetBool(((Const *) clause)->constvalue)) + { + *constfalse = true; + return NIL; + } + } + + /* Get the BoolExpr's out of the way. */ + if (IsA(clause, BoolExpr)) + { + /* + * Generate steps for arguments. + * + * While steps generated for the arguments themselves will be + * added to context->steps during recursion and will be evaluated + * indepdently, collect their step IDs to be stored in the combine + * step we'll be creating. + */ + if (or_clause((Node *) clause)) + { + List *arg_stepids = NIL; + bool all_args_constfalse = true; + ListCell *lc1; + + /* + * Get pruning step for each arg. If we get constfalse for + * all args, it means the OR expression is false as a whole. + */ + foreach(lc1, ((BoolExpr *) clause)->args) + { + Expr *arg = lfirst(lc1); + bool arg_constfalse; + List *argsteps; + + argsteps = + generate_partition_pruning_steps_internal(rel, context, + list_make1(arg), + &arg_constfalse); + if (!arg_constfalse) + all_args_constfalse = false; + + if (argsteps != NIL) + { + PartitionPruneStep *step; + + Assert(list_length(argsteps) == 1); + step = (PartitionPruneStep *) linitial(argsteps); + arg_stepids = lappend_int(arg_stepids, step->step_id); + } + else + { + /* + * No steps either means that arg_constfalse is true + * or the arg didn't contain a clause matching this + * partition key. + * + * In case of the latter, we cannot prune using such + * an arg. To indicate that to the pruning code, we + * must construct a dummy PartitionPruneStepCombine + * whose source_stepids is set to to an empty List. + * However, if we can prove using constraint exclusion + * that the clause refutes the table's partition + * constraint (if it's sub-partitioned), we need not + * bother with that. That is, we effectively ignore + * this OR arm. + */ + List *partconstr = rel->partition_qual; + PartitionPruneStep *orstep; + + /* Just ignore this argument. */ + if (arg_constfalse) + continue; + + if (partconstr) + { + partconstr = (List *) + expression_planner((Expr *) partconstr); + if (rel->relid != 1) + ChangeVarNodes((Node *) partconstr, 1, + rel->relid, 0); + if (predicate_refuted_by(partconstr, + list_make1(arg), + false)) + continue; + } + + orstep = generate_pruning_step_combine(context, + NIL, + COMBINE_UNION); + arg_stepids = lappend_int(arg_stepids, + orstep->step_id); + } + } + + *constfalse = all_args_constfalse; + + /* Check if any contradicting clauses were found */ + if (*constfalse) + return NIL; + + if (arg_stepids != NIL) + result = + lappend(result, + generate_pruning_step_combine(context, + arg_stepids, + COMBINE_UNION)); + continue; + } + else if (and_clause((Node *) clause)) + { + List *args = ((BoolExpr *) clause)->args; + List *argsteps, + *arg_stepids = NIL; + ListCell *lc1; + + /* + * args may itself contain clauses of arbitrary type, so just + * recurse and later combine the component partitions sets + * using a combine step. + */ + argsteps = + generate_partition_pruning_steps_internal(rel, + context, + args, + constfalse); + if (*constfalse) + return NIL; + + foreach (lc1, argsteps) + { + PartitionPruneStep *step = lfirst(lc1); + + arg_stepids = lappend_int(arg_stepids, step->step_id); + } + + if (arg_stepids) + result = + lappend(result, + generate_pruning_step_combine(context, + arg_stepids, + COMBINE_INTERSECT)); + continue; + } + + /* + * Fall-through for a NOT clause, which if it's a Boolean clause + * clause, will be handled in match_clause_to_partition_key(). + */ + } + + /* + * Must be a clause for which we can check if one of its args matches + * the partition key. + */ + for (i = 0; i < part_scheme->partnatts; i++) + { + Expr *partkey = linitial(rel->partexprs[i]); + bool unsupported_clause = false, + key_is_null = false, + key_is_not_null = false; + PartClauseInfo *pc = NULL; + List *clause_steps = NIL; + + switch (match_clause_to_partition_key(rel, context, + clause, partkey, i, + &key_is_null, + &key_is_not_null, + &pc, &clause_steps)) + { + case PARTCLAUSE_MATCH_CLAUSE: + Assert(pc != NULL); + + /* + * Since we only allow strict operators, check for any + * contradicting IS NULL. + */ + if (bms_is_member(i, nullkeys)) + { + *constfalse = true; + return NIL; + } + generate_opsteps = true; + keyclauses[i] = lappend(keyclauses[i], pc); + break; + + case PARTCLAUSE_MATCH_NULLNESS: + if (key_is_null) + { + /* check for conflicting IS NOT NULL */ + if (bms_is_member(i, notnullkeys)) + { + *constfalse = true; + return NIL; + } + nullkeys = bms_add_member(nullkeys, i); + } + else if (key_is_not_null) + { + /* check for conflicting IS NULL */ + if (bms_is_member(i, nullkeys)) + { + *constfalse = true; + return NIL; + } + notnullkeys = bms_add_member(notnullkeys, i); + } + else + Assert(false); + break; + + case PARTCLAUSE_MATCH_STEPS: + Assert(clause_steps != NIL); + result = list_concat(result, clause_steps); + break; + + case PARTCLAUSE_MATCH_CONTRADICT: + /* We've nothing more to do if a contradiction was found. */ + *constfalse = true; + return NIL; + + case PARTCLAUSE_NOMATCH: + /* + * Clause didn't match this key, but it might match the + * next one. + */ + continue; + + case PARTCLAUSE_UNSUPPORTED: + /* This clause cannot be used for pruning. */ + unsupported_clause = true; + break; + + default: + Assert(false); + break; + } + + /* go check the next clause. */ + if (unsupported_clause) + break; + } + } + + /* + * generate_opsteps set to false means no OpExprs were directly present in + * the input list. + */ + if (!generate_opsteps) + { + /* + * Generate one prune step for the information derived from IS NULL, if + * any. To prune hash partitions, we must have found IS NULL clauses + * for all partition keys. + */ + if (!bms_is_empty(nullkeys) && + (part_scheme->strategy != PARTITION_STRATEGY_HASH || + bms_num_members(nullkeys) == part_scheme->partnatts)) + { + PartitionPruneStep *step; + + step = generate_pruning_step_op(context, 0, false, NIL, NIL, + nullkeys); + if (step != NULL) + result = lappend(result, step); + } + + /* + * Note that for IS NOT NULL clauses, simply having step suffices; + * there is no need to propagate the exact details of which keys are + * required to be NOT NULL. Hash partitioning expects to see actual + * values to perform any pruning. + */ + if (!bms_is_empty(notnullkeys) && + part_scheme->strategy != PARTITION_STRATEGY_HASH) + { + PartitionPruneStep *step; + + step = generate_pruning_step_op(context, 0, false, + NIL, NIL, NULL); + if (step != NULL) + result = lappend(result, step); + } + } + else + { + PartitionPruneStep *step; + + /* Generate pruning steps from OpExpr clauses in keyclauses. */ + step = generate_pruning_steps_from_opexprs(part_scheme, context, + keyclauses, nullkeys); + if (step != NULL) + result = lappend(result, step); + } + + /* + * Finally, results from all entries appearing in result should be + * combined using an AND combine step, if there are more than 1. + */ + if (list_length(result) > 1) + { + List *step_ids = NIL; + + foreach(lc, result) + { + PartitionPruneStep *step = lfirst(lc); + + step_ids = lappend_int(step_ids, step->step_id); + } + + if (step_ids != NIL) + result = lappend(result, + generate_pruning_step_combine(context, step_ids, + COMBINE_INTERSECT)); + } + + return result; +} + +/* + * If the partition key has a collation, then the clause must have the same + * input collation. If the partition key is non-collatable, we assume the + * collation doesn't matter, because while collation wasn't considered when + * performing partitioning, the clause still may have a collation assigned + * due to the other input being of a collatable type. + */ +#define PartCollMatchesExprColl(partcoll, exprcoll) \ + ((partcoll) == InvalidOid || (partcoll) == (exprcoll)) + +/* + * match_clause_to_partition_key + * Match a given clause with the specified partition key + * + * Return value: + * + * One of PARTCLAUSE_MATCH_* enum values if the clause is successfully + * matched to the partition key. If it is PARTCLAUSE_MATCH_CONTRADICT, then + * this means the clause is self-contradictory (which can happen only if it's + * a BoolExpr whose arguments may be self-contradictory) + * + * PARTCLAUSE_NOMATCH if the clause doesn't match *this* partition key but + * the caller should continue trying because it may match a subsequent key + * + * PARTCLAUSE_UNSUPPORTED if the clause cannot be used for pruning at all, + * even if it may have been matched with a key, due to one of its properties, + * such as volatility of the arguments + * + * Based on the returned enum value, different output arguments are set as + * follows: + * + * PARTCLAUSE_UNSUPPORTED or + * PARTCLAUSE_NOMATCH or + * PARTCLAUSE_MATCH_CONTRADICT: None set (caller shouldn't rely on any of + * them being set) + * + * PARTCLAUSE_MATCH_CLAUSE: *pc set to PartClauseInfo constructed for the + * matched clause + * + * PARTCLAUSE_MATCH_NULLNESS: either *key_is_null or *key_is_not_null set + * based on whether the matched clause was a IS NULL or IS NOT NULL clause, + * respectively + * + * PARTCLAUSE_MATCH_STEPS: *clause_steps set to list of "partition pruning + * step(s)" generated for the clause due to it being a BoolExpr or a + * ScalarArrayOpExpr that's turned into one + */ +static PartClauseMatchStatus +match_clause_to_partition_key(RelOptInfo *rel, + GeneratePruningStepsContext *context, + Expr *clause, Expr *partkey, int partkeyidx, + bool *key_is_null, bool *key_is_not_null, + PartClauseInfo **pc, List **clause_steps) +{ + PartitionScheme part_scheme = rel->part_scheme; + Expr *expr; + Oid partopfamily = part_scheme->partopfamily[partkeyidx], + partcoll = part_scheme->partcollation[partkeyidx]; + + /* + * Recognize specially shaped clauses that match with the Boolean + * partition key. + */ + if (match_boolean_partition_clause(partopfamily, clause, partkey, &expr)) + { + *pc = palloc0(sizeof(PartClauseInfo)); + (*pc)->keyno = partkeyidx; + /* Do pruning with the Boolean equality operator. */ + (*pc)->opno = BooleanEqualOperator; + (*pc)->expr = expr; + /* We know that expr is of Boolean type. */ + (*pc)->cmpfn = rel->part_scheme->partsupfunc[partkeyidx].fn_oid; + + return PARTCLAUSE_MATCH_CLAUSE; + } + else if (IsA(clause, OpExpr) && + list_length(((OpExpr *) clause)->args) == 2) + { + OpExpr *opclause = (OpExpr *) clause; + Expr *leftop, + *rightop; + Oid commutator = InvalidOid, + negator = InvalidOid; + Oid cmpfn; + Oid exprtype; + bool is_opne_listp = false; + + leftop = (Expr *) get_leftop(clause); + if (IsA(leftop, RelabelType)) + leftop = ((RelabelType *) leftop)->arg; + rightop = (Expr *) get_rightop(clause); + if (IsA(rightop, RelabelType)) + rightop = ((RelabelType *) rightop)->arg; + + /* check if the clause matches this partition key */ + if (equal(leftop, partkey)) + expr = rightop; + else if (equal(rightop, partkey)) + { + expr = leftop; + commutator = get_commutator(opclause->opno); + + /* nothing we can do unless we can swap the operands */ + if (!OidIsValid(commutator)) + return PARTCLAUSE_UNSUPPORTED; + } + else + /* clause does not match this partition key, but perhaps next. */ + return PARTCLAUSE_NOMATCH; + + /* + * Partition key also consists of a collation that's specified for it, + * so try to match it too. There may be multiple keys with the same + * expression but different collations. + */ + if (!PartCollMatchesExprColl(partcoll, opclause->inputcollid)) + return PARTCLAUSE_NOMATCH; + + /* + * Matched with this key. Now check various properties of the clause + * to see if it's sane to use it for pruning. If any of the + * properties makes it unsuitable for pruning, then the clause is + * useless no matter which key it's matched to. + */ + + /* + * Only allow strict operators. This will guarantee nulls are + * filtered. + */ + if (!op_strict(opclause->opno)) + return PARTCLAUSE_UNSUPPORTED; + + /* We can't use any volatile expressions to prune partitions. */ + if (contain_volatile_functions((Node *) expr)) + return PARTCLAUSE_UNSUPPORTED; + + /* + * Normally we only bother with operators that are listed as being + * part of the partitioning operator family. But we make an exception + * in one case -- operators named '<>' are not listed in any operator + * family whatsoever, in which case, we try to perform partition + * pruning with it only if list partitioning is in use. + */ + if (!op_in_opfamily(opclause->opno, partopfamily)) + { + if (part_scheme->strategy != PARTITION_STRATEGY_LIST) + return PARTCLAUSE_UNSUPPORTED; + + /* + * To confirm if the operator is really '<>', check if its negator + * is a btree equality operator. + */ + negator = get_negator(opclause->opno); + if (OidIsValid(negator) && op_in_opfamily(negator, partopfamily)) + { + Oid lefttype; + Oid righttype; + int strategy; + + get_op_opfamily_properties(negator, partopfamily, false, + &strategy, &lefttype, &righttype); + + if (strategy == BTEqualStrategyNumber) + is_opne_listp = true; + } + + /* Operator isn't really what we were hoping it'd be. */ + if (!is_opne_listp) + return PARTCLAUSE_UNSUPPORTED; + } + + /* Check if we're going to need a cross-type comparison function. */ + exprtype = exprType((Node *) expr); + if (exprtype != part_scheme->partopcintype[partkeyidx]) + { + switch (part_scheme->strategy) + { + case PARTITION_STRATEGY_LIST: + case PARTITION_STRATEGY_RANGE: + cmpfn = + get_opfamily_proc(part_scheme->partopfamily[partkeyidx], + part_scheme->partopcintype[partkeyidx], + exprtype, BTORDER_PROC); + break; + + case PARTITION_STRATEGY_HASH: + cmpfn = + get_opfamily_proc(part_scheme->partopfamily[partkeyidx], + exprtype, exprtype, HASHEXTENDED_PROC); + break; + + default: + elog(ERROR, "invalid partition strategy: %c", + part_scheme->strategy); + break; + } + + /* If we couldn't find one, we cannot use this expression. */ + if (!OidIsValid(cmpfn)) + return PARTCLAUSE_UNSUPPORTED; + } + else + cmpfn = part_scheme->partsupfunc[partkeyidx].fn_oid; + + *pc = palloc0(sizeof(PartClauseInfo)); + (*pc)->keyno = partkeyidx; + + /* For <> operator clauses, pass on the negator. */ + (*pc)->isopne = false; + if (is_opne_listp) + { + Assert(OidIsValid(negator)); + (*pc)->opno = negator; + (*pc)->isopne = true; + } + /* And if commuted before matching, pass on the commutator */ + else if (OidIsValid(commutator)) + (*pc)->opno = commutator; + else + (*pc)->opno = opclause->opno; + + (*pc)->expr = expr; + (*pc)->cmpfn = cmpfn; + + return PARTCLAUSE_MATCH_CLAUSE; + } + else if (IsA(clause, ScalarArrayOpExpr)) + { + ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) clause; + Oid saop_op = saop->opno; + Oid saop_coll = saop->inputcollid; + Expr *leftop = (Expr *) linitial(saop->args), + *rightop = (Expr *) lsecond(saop->args); + List *elem_exprs, + *elem_clauses; + ListCell *lc1; + + if (IsA(leftop, RelabelType)) + leftop = ((RelabelType *) leftop)->arg; + + /* Check it matches this partition key */ + if (!equal(leftop, partkey) || + !PartCollMatchesExprColl(partcoll, saop->inputcollid)) + return PARTCLAUSE_NOMATCH; + + /* + * Matched with this key. Check various properties of the clause to + * see if it can sanely be used for partition pruning. + */ + + /* + * Only allow strict operators. This will guarantee nulls are + * filtered. + */ + if (!op_strict(saop->opno)) + return PARTCLAUSE_UNSUPPORTED; + + /* Useless if the array has any volatile functions. */ + if (contain_volatile_functions((Node *) rightop)) + return PARTCLAUSE_UNSUPPORTED; + + /* + * In case of NOT IN (..), we get a '<>', which we handle if list + * partitioning is in use and we're able to confirm that it's negator + * is a btree equality operator belonging to the partitioning operator + * family. + */ + if (!op_in_opfamily(saop_op, partopfamily)) + { + Oid negator; + + if (part_scheme->strategy != PARTITION_STRATEGY_LIST) + return PARTCLAUSE_UNSUPPORTED; + + negator = get_negator(saop_op); + if (OidIsValid(negator) && op_in_opfamily(negator, partopfamily)) + { + int strategy; + Oid lefttype, + righttype; + + get_op_opfamily_properties(negator, partopfamily, + false, &strategy, + &lefttype, &righttype); + if (strategy != BTEqualStrategyNumber) + return PARTCLAUSE_UNSUPPORTED; + } + } + + /* + * First generate a list of Const nodes, one for each array element. + */ + elem_exprs = NIL; + if (IsA(rightop, Const)) + { + Const *arr = (Const *) lsecond(saop->args); + ArrayType *arrval = DatumGetArrayTypeP(arr->constvalue); + int16 elemlen; + bool elembyval; + char elemalign; + Datum *elem_values; + bool *elem_nulls; + int num_elems, + i; + + get_typlenbyvalalign(ARR_ELEMTYPE(arrval), + &elemlen, &elembyval, &elemalign); + deconstruct_array(arrval, + ARR_ELEMTYPE(arrval), + elemlen, elembyval, elemalign, + &elem_values, &elem_nulls, + &num_elems); + for (i = 0; i < num_elems; i++) + { + /* Only consider non-null values. */ + if (!elem_nulls[i]) + { + Const *elem_expr = makeConst(ARR_ELEMTYPE(arrval), + -1, arr->constcollid, + elemlen, + elem_values[i], + false, elembyval); + + elem_exprs = lappend(elem_exprs, elem_expr); + } + } + } + else + { + ArrayExpr *arrexpr = castNode(ArrayExpr, rightop); + + /* + * For a nested ArrayExpr, we don't know how to get the actual + * scalar values out into a flat list, so we give up doing + * anything with this ScalarArrayOpExpr. + */ + if (arrexpr->multidims) + return PARTCLAUSE_UNSUPPORTED; + + elem_exprs = arrexpr->elements; + } + + /* + * Now generate a list of clauses, one for each array element, of the + * form saop_leftop saop_op elem_expr + */ + elem_clauses = NIL; + foreach(lc1, elem_exprs) + { + Expr *rightop = (Expr *) lfirst(lc1), + *elem_clause; + + elem_clause = (Expr *) make_opclause(saop_op, BOOLOID, + false, + leftop, rightop, + InvalidOid, + saop_coll); + elem_clauses = lappend(elem_clauses, elem_clause); + } + + /* + * Build a combine step as if for an OR clause or add the clauses to + * the end of the list that's being processed currently. + */ + if (saop->useOr && list_length(elem_clauses) > 1) + { + Expr *orexpr; + bool constfalse; + + orexpr = makeBoolExpr(OR_EXPR, elem_clauses, -1); + *clause_steps = + generate_partition_pruning_steps_internal(rel, context, + list_make1(orexpr), + &constfalse); + if (constfalse) + return PARTCLAUSE_MATCH_CONTRADICT; + Assert(list_length(*clause_steps) == 1); + return PARTCLAUSE_MATCH_STEPS; + } + else + { + bool constfalse; + + *clause_steps = + generate_partition_pruning_steps_internal(rel, context, + elem_clauses, + &constfalse); + if (constfalse) + return PARTCLAUSE_MATCH_CONTRADICT; + Assert(list_length(*clause_steps) >= 1); + return PARTCLAUSE_MATCH_STEPS; + } + } + else if (IsA(clause, NullTest)) + { + NullTest *nulltest = (NullTest *) clause; + Expr *arg = nulltest->arg; + + if (IsA(arg, RelabelType)) + arg = ((RelabelType *) arg)->arg; + + /* Does arg match with this partition key column? */ + if (!equal(arg, partkey)) + return PARTCLAUSE_NOMATCH; + + if (nulltest->nulltesttype == IS_NULL) + *key_is_null = true; + else + *key_is_not_null = true; + + return PARTCLAUSE_MATCH_NULLNESS; + } + + return PARTCLAUSE_UNSUPPORTED; +} + +/* + * match_boolean_partition_clause + * + * Sets *rightop to a Const containing true or false value and returns true if + * we're able to match the clause to the partition key as specially-shaped + * Boolean clause. Returns false otherwise with *rightop set to NULL. + */ +static bool +match_boolean_partition_clause(Oid partopfamily, Expr *clause, Expr *partkey, + Expr **rightop) +{ + Expr *leftop; + + if (!IsBooleanOpfamily(partopfamily)) + return false; + + *rightop = NULL; + if (IsA(clause, BooleanTest)) + { + BooleanTest *btest = (BooleanTest *) clause; + + /* Only IS [NOT] TRUE/FALSE are any good to us */ + if (btest->booltesttype == IS_UNKNOWN || + btest->booltesttype == IS_NOT_UNKNOWN) + return false; + + leftop = btest->arg; + if (IsA(leftop, RelabelType)) + leftop = ((RelabelType *) leftop)->arg; + + if (equal(leftop, partkey)) + *rightop = (btest->booltesttype == IS_TRUE || + btest->booltesttype == IS_NOT_FALSE) + ? (Expr *) makeBoolConst(true, false) + : (Expr *) makeBoolConst(false, false); + + if (*rightop) + return true; + } + else + { + leftop = not_clause((Node *) clause) + ? get_notclausearg(clause) + : clause; + + if (IsA(leftop, RelabelType)) + leftop = ((RelabelType *) leftop)->arg; + + /* Clause does not match this partition key. */ + if (equal(leftop, partkey)) + *rightop = not_clause((Node *) clause) + ? (Expr *) makeBoolConst(false, false) + : (Expr *) makeBoolConst(true, false); + else if (equal(negate_clause((Node *) leftop), partkey)) + *rightop = (Expr *) makeBoolConst(false, false); + + if (*rightop) + return true; + } + + return false; +} + +/* + * generate_pruning_steps_from_opexprs + * + * 'keyclauses' contains one list of clauses per partition key. We check here + * if we have found clauses for a valid subset of the partition key. In some + * cases, (depending on the type of partitioning being used) if we didn't + * find clauses for a given key, we discard clauses that may have been + * found for any subsequent keys; see specific notes below. + */ +static PartitionPruneStep * +generate_pruning_steps_from_opexprs(PartitionScheme part_scheme, + GeneratePruningStepsContext *context, + List **keyclauses, + Bitmapset *nullkeys) +{ + ListCell *lc; + List *opsteps = NIL; + List *btree_clauses[BTMaxStrategyNumber], + *hash_clauses[HTMaxStrategyNumber]; + bool need_next_less, + need_next_eq, + need_next_greater; + int i; + + memset(btree_clauses, 0, sizeof(btree_clauses)); + memset(hash_clauses, 0, sizeof(hash_clauses)); + for (i = 0; i < part_scheme->partnatts; i++) + { + List *clauselist = keyclauses[i]; + bool consider_next_key = true; + + /* + * To be useful for pruning, we must have clauses for a prefix of + * partition keys in the case of range partitioning. So, ignore + * clauses for keys after this one. + */ + if (part_scheme->strategy == PARTITION_STRATEGY_RANGE && + clauselist == NIL) + break; + + /* + * For hash partitioning, if a column doesn't have necessary equality + * clause, there should be an IS NULL clause, otherwise pruning is not + * possible. + */ + if (part_scheme->strategy == PARTITION_STRATEGY_HASH && + clauselist == NIL && !bms_is_member(i, nullkeys)) + return NULL; + + need_next_eq = need_next_less = need_next_greater = true; + foreach(lc, clauselist) + { + PartClauseInfo *pc = (PartClauseInfo *) lfirst(lc); + Oid lefttype, + righttype; + + /* Look up the operator's btree/hash strategy number. */ + if (pc->op_strategy == InvalidStrategy) + get_op_opfamily_properties(pc->opno, + part_scheme->partopfamily[i], + false, + &pc->op_strategy, + &lefttype, + &righttype); + + switch (part_scheme->strategy) + { + case PARTITION_STRATEGY_LIST: + case PARTITION_STRATEGY_RANGE: + { + PartClauseInfo *last = NULL; + bool inclusive = false; + + /* + * Add this clause to the list of clauses to be used + * for pruning if this is the first such key for this + * operator strategy or if it is consecutively next to + * the last column for which a clause with this + * operator strategy was matched. + */ + if (btree_clauses[pc->op_strategy - 1] != NIL) + last = llast(btree_clauses[pc->op_strategy - 1]); + + if (last == NULL || + i == last->keyno || i == last->keyno + 1) + btree_clauses[pc->op_strategy - 1] = + lappend(btree_clauses[pc->op_strategy - 1], pc); + + /* + * We may not need the next clause if they're of + * certain strategy. + */ + switch (pc->op_strategy) + { + case BTLessEqualStrategyNumber: + inclusive = true; + /* fall through */ + case BTLessStrategyNumber: + if (!inclusive) + need_next_eq = need_next_less = false; + break; + case BTEqualStrategyNumber: + /* always accept clauses for the next key. */ + break; + case BTGreaterEqualStrategyNumber: + inclusive = true; + /* fall through */ + case BTGreaterStrategyNumber: + if (!inclusive) + need_next_eq = need_next_greater = false; + break; + } + + /* We may want to change our mind. */ + if (consider_next_key) + consider_next_key = (need_next_eq || + need_next_less || + need_next_greater); + break; + } + + case PARTITION_STRATEGY_HASH: + if (pc->op_strategy != HTEqualStrategyNumber) + elog(ERROR, "invalid clause for hash partitioning"); + hash_clauses[pc->op_strategy - 1] = + lappend(hash_clauses[pc->op_strategy - 1], pc); + break; + + default: + elog(ERROR, "invalid partition strategy: %c", + part_scheme->strategy); + break; + } + } + + /* + * If we've decided that clauses for subsequent partition keys + * wouldn't be useful for pruning, don't look. + */ + if (!consider_next_key) + break; + } + + /* + * Now, we have divided clauses according to their operator strategies. + * Check for each strategy if we can generate pruning step(s) by + * collecting a list of expressions whose values will constitute a vector + * that can be used as a look-up key by a partition bound searching + * function. + */ + switch (part_scheme->strategy) + { + case PARTITION_STRATEGY_LIST: + case PARTITION_STRATEGY_RANGE: + { + List *eq_clauses = btree_clauses[BTEqualStrategyNumber - 1]; + List *le_clauses = btree_clauses[BTLessEqualStrategyNumber - 1]; + List *ge_clauses = btree_clauses[BTGreaterEqualStrategyNumber - 1]; + + /* + * For each clause under consideration for a given strategy, + * we collect expressions from clauses for earlier keys, whose + * operator strategy is inclusive, into a list called 'prefix'. + * By appending the clause's own expression to the 'prefix', + * we'll generate one step using the so generated vector and + * assign the current strategy to it. Actually, 'prefix' might + * contain multiple clauses for the same key, in which case, + * we must generate steps for various combinations of + * expressions of different keys, which get_steps_using_prefix + * takes care of doing. + */ + for (i = 0; i < BTMaxStrategyNumber; i++) + { + PartClauseInfo *pc; + List *pc_steps; + + foreach(lc, btree_clauses[i]) + { + ListCell *lc1; + List *prefix = NIL; + + /* Clause under consideration. */ + pc = lfirst(lc); + + /* + * Expressions from = clauses can always be in the + * prefix, provided they're from an earlier key. + */ + foreach(lc1, eq_clauses) + { + PartClauseInfo *eqpc = lfirst(lc1); + + if (eqpc->keyno == pc->keyno) + break; + if (eqpc->keyno < pc->keyno) + prefix = lappend(prefix, eqpc); + } + + /* + * If we're generating steps for keyno == pc->keyno) + break; + if (lepc->keyno < pc->keyno) + prefix = lappend(prefix, lepc); + } + } + + /* + * If we're generating steps for >/>= strategy, we can + * add other >= clauses to the prefix, provided they're + * from an earlier key. + */ + if (i == BTGreaterStrategyNumber - 1 || + i == BTGreaterEqualStrategyNumber - 1) + { + foreach(lc1, ge_clauses) + { + PartClauseInfo *gepc = lfirst(lc1); + + if (gepc->keyno == pc->keyno) + break; + if (gepc->keyno < pc->keyno) + prefix = lappend(prefix, gepc); + } + } + + /* + * As mentioned above, if 'prefix' contains multiple + * expressions for the same key, the following will + * generate multiple steps, one for each combination + * of the expressions for different keys. + * + * Note that we pass NULL for step_nullkeys, because + * we don't search list/range partition bounds where + * some keys are NULL. + */ + Assert(pc->op_strategy == i + 1); + pc_steps = get_steps_using_prefix(context, i + 1, + pc->isopne, + pc->expr, + pc->cmpfn, + pc->keyno, + NULL, + prefix); + opsteps = list_concat(opsteps, list_copy(pc_steps)); + } + } + break; + } + + case PARTITION_STRATEGY_HASH: + { + List *eq_clauses = hash_clauses[HTEqualStrategyNumber - 1]; + + /* For hash partitioning, we have just the = strategy. */ + if (eq_clauses != NIL) + { + PartClauseInfo *pc; + List *pc_steps; + List *prefix = NIL; + int last_keyno; + ListCell *lc1; + + /* + * Locate the clause for the greatest column (which may + * not be the last partition key column). Actually, the + * last element of eq_clauses must give us what we need. + */ + pc = llast(eq_clauses); + + /* + * But there might be multiple clauses that we matched to + * that column; go to the first such clause. While at it, + * add all the clauses before that one to 'prefix'. + */ + last_keyno = pc->keyno; + foreach(lc, eq_clauses) + { + pc = lfirst(lc); + if (pc->keyno == last_keyno) + break; + prefix = lappend(prefix, pc); + } + + /* + * For each clause for the "last" column, after appending + * the clause's own expression to the 'prefix', we'll + * generate one step using the so generated vector and + * and assign = as its strategy. Actually, 'prefix' might + * contain multiple clauses for the same key, in which + * case, we must generate steps for various combinations + * of expressions of different keys, which + * get_steps_using_prefix takes care of doing. + */ + for_each_cell(lc1, lc) + { + pc = lfirst(lc1); + + /* + * Note that we pass nullkeys for step_nullkeys, + * because we need to tell hash partition bound search + * function which of the keys are NULL. + */ + Assert(pc->op_strategy == HTEqualStrategyNumber); + pc_steps = + get_steps_using_prefix(context, + HTEqualStrategyNumber, + false, + pc->expr, + pc->cmpfn, + pc->keyno, + nullkeys, + prefix); + opsteps = list_concat(opsteps, list_copy(pc_steps)); + } + } + break; + } + + default: + elog(ERROR, "invalid partition strategy: %c", + part_scheme->strategy); + break; + } + + /* Finally, add a combine step to mutualy AND opsteps, if needed. */ + if (list_length(opsteps) > 1) + { + List *opstep_ids = NIL; + + foreach(lc, opsteps) + { + PartitionPruneStep *step = lfirst(lc); + + opstep_ids = lappend_int(opstep_ids, step->step_id); + } + + if (opstep_ids != NIL) + return generate_pruning_step_combine(context, opstep_ids, + COMBINE_INTERSECT); + return NULL; + } + else if (opsteps != NIL) + return linitial(opsteps); + + return NULL; +} + +/* + * get_steps_using_prefix + * Generate list of PartitionPruneStepOp steps each consisting of given + * opstrategy + * + * To generate steps, step_lastexpr and step_lastcmpfn are appended to + * expressions and cmpfns, respectively, extracted from the clauses in + * 'prefix'. Actually, since 'prefix' may contain multiple clauses for the + * same partition key column, we must generate steps for various combinations + * of the clauses of different keys. + */ +static List * +get_steps_using_prefix(GeneratePruningStepsContext *context, + int step_opstrategy, + bool step_isopne, + Expr *step_lastexpr, + Oid step_lastcmpfn, + int step_lastkeyno, + Bitmapset *step_nullkeys, + List *prefix) +{ + /* Quick exit if there are no values to prefix with. */ + if (list_length(prefix) == 0) + { + PartitionPruneStep *step; + + step = generate_pruning_step_op(context, + step_opstrategy, step_isopne, + list_make1(step_lastexpr), + list_make1_oid(step_lastcmpfn), + step_nullkeys); + return step != NULL ? list_make1(step) : NIL; + } + + /* Recurse to generate steps for various combinations. */ + return get_steps_using_prefix_recurse(context, + step_opstrategy, + step_isopne, + step_lastexpr, + step_lastcmpfn, + step_lastkeyno, + step_nullkeys, + list_head(prefix), + NIL, NIL); +} + +/* + * get_steps_using_prefix_recurse + * Recursively generate combinations of clauses for different partition + * keys and start generating steps upon reaching clauses for the greatest + * column that is less than the one for which we're currently generating + * steps (that is, step_lastkeyno) + * + * 'start' is where we should start iterating for the current invocation. + * 'step_exprs' and 'step_cmpfns' each contains the expressions and cmpfns + * we've generated so far from the clauses for the still earlier columns. + */ +static List * +get_steps_using_prefix_recurse(GeneratePruningStepsContext *context, + int step_opstrategy, + bool step_isopne, + Expr *step_lastexpr, + Oid step_lastcmpfn, + int step_lastkeyno, + Bitmapset *step_nullkeys, + ListCell *start, + List *step_exprs, + List *step_cmpfns) +{ + List *result = NIL; + ListCell *lc; + int cur_keyno; + + /* Actually, recursion would be limited by PARTITION_MAX_KEYS. */ + check_stack_depth(); + + /* Check if we need to recurse. */ + Assert(start != NULL); + cur_keyno = ((PartClauseInfo *) lfirst(start))->keyno; + if (cur_keyno < step_lastkeyno - 1) + { + PartClauseInfo *pc; + ListCell *next_start; + + /* + * For each clause with cur_keyno, adds its expr and cmpfn to + * step_exprs and step_cmpfns, respectively, and recurse after setting + * next_start to the ListCell of the first clause for the next + * partition key. + */ + for_each_cell(lc, start) + { + pc = lfirst(lc); + + if (pc->keyno > cur_keyno) + break; + } + next_start = lc; + + for_each_cell(lc, start) + { + pc = lfirst(lc); + if (pc->keyno == cur_keyno) + { + /* clean up before starting a new recursion cycle. */ + if (cur_keyno == 0) + { + list_free(step_exprs); + list_free(step_cmpfns); + step_exprs = list_make1(pc->expr); + step_cmpfns = list_make1_oid(pc->cmpfn); + } + else + { + step_exprs = lappend(step_exprs, pc->expr); + step_cmpfns = lappend_oid(step_cmpfns, pc->cmpfn); + } + } + else + { + Assert(pc->keyno > cur_keyno); + break; + } + + result = + list_concat(result, + get_steps_using_prefix_recurse(context, + step_opstrategy, + step_isopne, + step_lastexpr, + step_lastcmpfn, + step_lastkeyno, + step_nullkeys, + next_start, + step_exprs, + step_cmpfns)); + } + } + else + { + /* + * End the current recursion cycle and start generating steps, one + * for each clause with cur_keyno, which is all clauses from here + * onward till the end of the list. + */ + Assert(list_length(step_exprs) == cur_keyno); + for_each_cell(lc, start) + { + PartClauseInfo *pc = lfirst(lc); + List *step_exprs1, + *step_cmpfns1; + PartitionPruneStep *step; + + Assert(pc->keyno == cur_keyno); + + /* Leave the original step_exprs unmodified. */ + step_exprs1 = list_copy(step_exprs); + step_exprs1 = lappend(step_exprs1, pc->expr); + step_exprs1 = lappend(step_exprs1, step_lastexpr); + + /* Leave the original step_cmpfns unmodified. */ + step_cmpfns1 = list_copy(step_cmpfns); + step_cmpfns1 = lappend_oid(step_cmpfns1, pc->cmpfn); + step_cmpfns1 = lappend_oid(step_cmpfns1, step_lastcmpfn); + + step = generate_pruning_step_op(context, + step_opstrategy, step_isopne, + step_exprs1, step_cmpfns1, + step_nullkeys); + if (step != NULL) + result = lappend(result, step); + } + } + + return result; +} + +/* + * Following functions generate pruning steps of various types. Each step + * that's created is added to a global context->steps and receive a globally + * unique identifier that's sourced from context->next_step_id. + */ + +static PartitionPruneStep * +generate_pruning_step_op(GeneratePruningStepsContext *context, + int opstrategy, bool isopne, + List *exprs, List *cmpfns, + Bitmapset *nullkeys) +{ + PartitionPruneStepOp *opstep = makeNode(PartitionPruneStepOp); + + opstep->step.step_id = context->next_step_id++; + + if (!isopne) + opstep->opstrategy = opstrategy; + else + { + Assert(opstrategy == BTEqualStrategyNumber); + opstep->opstrategy = InvalidStrategy; + } + Assert(list_length(exprs) == list_length(cmpfns)); + opstep->exprs = exprs; + opstep->cmpfns = cmpfns; + opstep->nullkeys = nullkeys; + + context->steps = lappend(context->steps, opstep); + + return (PartitionPruneStep *) opstep; +} + +static PartitionPruneStep * +generate_pruning_step_combine(GeneratePruningStepsContext *context, + List *source_stepids, + PartitionPruneCombineOp combineOp) +{ + PartitionPruneStepCombine *cstep = makeNode(PartitionPruneStepCombine); + + cstep->step.step_id = context->next_step_id++; + cstep->combineOp = combineOp; + cstep->source_stepids = source_stepids; + + context->steps = lappend(context->steps, cstep); + + return (PartitionPruneStep *) cstep; +} diff --git a/src/backend/optimizer/util/plancat.c b/src/backend/optimizer/util/plancat.c index b46b33d4f7..32e973385d 100644 --- a/src/backend/optimizer/util/plancat.c +++ b/src/backend/optimizer/util/plancat.c @@ -1171,7 +1171,6 @@ get_relation_constraints(PlannerInfo *root, Index varno = rel->relid; Relation relation; TupleConstr *constr; - List *pcqual; /* * We assume the relation has already been safely locked. @@ -1257,24 +1256,34 @@ get_relation_constraints(PlannerInfo *root, } } - /* Append partition predicates, if any */ - pcqual = RelationGetPartitionQual(relation); - if (pcqual) + /* + * Append partition predicates, if any. + * + * For selects, partition pruning uses the parent table's partition bound + * descriptor, instead of constraint exclusion which is driven by the + * individual partition's partition constraint. + */ + if (root->parse->commandType != CMD_SELECT) { - /* - * Run the partition quals through const-simplification similar to - * check constraints. We skip canonicalize_qual, though, because - * partition quals should be in canonical form already; also, since - * the qual is in implicit-AND format, we'd have to explicitly convert - * it to explicit-AND format and back again. - */ - pcqual = (List *) eval_const_expressions(root, (Node *) pcqual); + List *pcqual = RelationGetPartitionQual(relation); - /* Fix Vars to have the desired varno */ - if (varno != 1) - ChangeVarNodes((Node *) pcqual, 1, varno, 0); + if (pcqual) + { + /* + * Run the partition quals through const-simplification similar to + * check constraints. We skip canonicalize_qual, though, because + * partition quals should be in canonical form already; also, + * since the qual is in implicit-AND format, we'd have to + * explicitly convert it to explicit-AND format and back again. + */ + pcqual = (List *) eval_const_expressions(root, (Node *) pcqual); - result = list_concat(result, pcqual); + /* Fix Vars to have the desired varno */ + if (varno != 1) + ChangeVarNodes((Node *) pcqual, 1, varno, 0); + + result = list_concat(result, pcqual); + } } heap_close(relation, NoLock); @@ -1869,6 +1878,9 @@ set_relation_partition_info(PlannerInfo *root, RelOptInfo *rel, rel->boundinfo = partition_bounds_copy(partdesc->boundinfo, partkey); rel->nparts = partdesc->nparts; set_baserel_partition_key_exprs(relation, rel); + rel->has_default_part = + OidIsValid(get_default_oid_from_partdesc(partdesc)); + rel->partition_qual = RelationGetPartitionQual(relation); } /* diff --git a/src/backend/optimizer/util/relnode.c b/src/backend/optimizer/util/relnode.c index da8f0f93fc..7f1428b8d8 100644 --- a/src/backend/optimizer/util/relnode.c +++ b/src/backend/optimizer/util/relnode.c @@ -154,6 +154,8 @@ build_simple_rel(PlannerInfo *root, int relid, RelOptInfo *parent) rel->part_scheme = NULL; rel->nparts = 0; rel->boundinfo = NULL; + rel->has_default_part = false; + rel->partition_qual = NIL; rel->part_rels = NULL; rel->partexprs = NULL; rel->nullable_partexprs = NULL; @@ -567,6 +569,8 @@ build_join_rel(PlannerInfo *root, joinrel->part_scheme = NULL; joinrel->nparts = 0; joinrel->boundinfo = NULL; + joinrel->has_default_part = false; + joinrel->partition_qual = NIL; joinrel->part_rels = NULL; joinrel->partexprs = NULL; joinrel->nullable_partexprs = NULL; @@ -734,6 +738,10 @@ build_child_join_rel(PlannerInfo *root, RelOptInfo *outer_rel, joinrel->has_eclass_joins = false; joinrel->top_parent_relids = NULL; joinrel->part_scheme = NULL; + joinrel->nparts = 0; + joinrel->boundinfo = NULL; + joinrel->has_default_part = false; + joinrel->partition_qual = NIL; joinrel->part_rels = NULL; joinrel->partexprs = NULL; joinrel->nullable_partexprs = NULL; diff --git a/src/include/catalog/partition.h b/src/include/catalog/partition.h index cd15faa7a1..8981901272 100644 --- a/src/include/catalog/partition.h +++ b/src/include/catalog/partition.h @@ -42,6 +42,28 @@ typedef struct PartitionDescData typedef struct PartitionDescData *PartitionDesc; +/* + * PartitionPruneContext + * + * Information about a partitioned table needed to perform partition pruning. + */ +typedef struct PartitionPruneContext +{ + /* Partition key information */ + char strategy; + int partnatts; + Oid *partopfamily; + Oid *partopcintype; + Oid *partcollation; + FmgrInfo *partsupfunc; + + /* Number of partitions */ + int nparts; + + /* Partition boundary info */ + PartitionBoundInfo boundinfo; +} PartitionPruneContext; + extern void RelationBuildPartitionDesc(Relation relation); extern bool partition_bounds_equal(int partnatts, int16 *parttyplen, bool *parttypbyval, PartitionBoundInfo b1, @@ -74,4 +96,7 @@ extern List *get_proposed_default_constraint(List *new_part_constaints); extern int get_partition_for_tuple(Relation relation, Datum *values, bool *isnull); +/* For partition-pruning */ +extern Bitmapset *get_matching_partitions(PartitionPruneContext *context, + List *pruning_steps); #endif /* PARTITION_H */ diff --git a/src/include/catalog/pg_opfamily.h b/src/include/catalog/pg_opfamily.h index b544474254..0847df97ff 100644 --- a/src/include/catalog/pg_opfamily.h +++ b/src/include/catalog/pg_opfamily.h @@ -188,4 +188,7 @@ DATA(insert OID = 4104 ( 3580 box_inclusion_ops PGNSP PGUID )); DATA(insert OID = 5000 ( 4000 box_ops PGNSP PGUID )); DATA(insert OID = 5008 ( 4000 poly_ops PGNSP PGUID )); +#define IsBooleanOpfamily(opfamily) \ + ((opfamily) == BOOL_BTREE_FAM_OID || (opfamily) == BOOL_HASH_FAM_OID) + #endif /* PG_OPFAMILY_H */ diff --git a/src/include/nodes/nodes.h b/src/include/nodes/nodes.h index fce48026b6..4df979e9eb 100644 --- a/src/include/nodes/nodes.h +++ b/src/include/nodes/nodes.h @@ -193,6 +193,9 @@ typedef enum NodeTag T_FromExpr, T_OnConflictExpr, T_IntoClause, + T_PartitionPruneStep, + T_PartitionPruneStepOp, + T_PartitionPruneStepCombine, /* * TAGS FOR EXPRESSION STATE NODES (execnodes.h) diff --git a/src/include/nodes/primnodes.h b/src/include/nodes/primnodes.h index 1b4b0d75af..c9d2187631 100644 --- a/src/include/nodes/primnodes.h +++ b/src/include/nodes/primnodes.h @@ -1506,4 +1506,77 @@ typedef struct OnConflictExpr List *exclRelTlist; /* tlist of the EXCLUDED pseudo relation */ } OnConflictExpr; +/* + * Node types to represent a partition pruning step + */ + +/* + * The base Node type. step_id is the global identifier of a given step + * within a given pruning context. + */ +typedef struct PartitionPruneStep +{ + NodeTag type; + int step_id; +} PartitionPruneStep; + +/*---------- + * PartitionPruneStepOp - Information to prune using a set of mutually AND'd + * OpExpr clauses + * + * This contains information extracted from up to partnatts OpExpr clauses, + * where partnatts is the number of partition key columns. 'opstrategy' is the + * strategy of the operator in the clause matched to the last partition key. + * 'exprs' contains expressions which comprise the look-up key to be passed to + * the partition bound search function. 'cmpfns' contains the OIDs of + * comparison function used to compare aforementioned expressions with + * partition bounds. Both 'exprs' and 'cmpfns' contain the same number of + * items up to partnatts items. + * + * Once we find the offset of a partition bound using the look-up key, we + * determine which partitions to include in the result based on the value of + * 'opstrategy'. For example, if it were equality, we'd return just the + * partition that would contain that key or a set of partitions if the key + * didn't consist of all partitioning columns. For non-equality strategies, + * we'd need to include other partitions as appropriate. + * + * 'nullkeys' is the set containing the offset of the partition keys (0 to + * partnatts - 1) that were matched to an IS NULL clause. This is only + * considered for hash partitioning as we need to pass which keys are null + * to the hash partition bound search function. It is never possible to + * have an expression be present in 'exprs' for a given partition key and + * the corresponding bit set in 'nullkeys'. + *---------- + */ +typedef struct PartitionPruneStepOp +{ + PartitionPruneStep step; + + int opstrategy; + List *exprs; + List *cmpfns; + Bitmapset *nullkeys; +} PartitionPruneStepOp; + +/*---------- + * PartitionPruneStepCombine - Information to prune using a BoolExpr clause + * + * For BoolExpr clauses, we combine the set of partitions determined for each + * of its argument clauses. + *---------- + */ +typedef enum PartitionPruneCombineOp +{ + COMBINE_UNION, + COMBINE_INTERSECT +} PartitionPruneCombineOp; + +typedef struct PartitionPruneStepCombine +{ + PartitionPruneStep step; + + PartitionPruneCombineOp combineOp; + List *source_stepids; +} PartitionPruneStepCombine; + #endif /* PRIMNODES_H */ diff --git a/src/include/nodes/relation.h b/src/include/nodes/relation.h index 83b03b41e4..9b9aabddef 100644 --- a/src/include/nodes/relation.h +++ b/src/include/nodes/relation.h @@ -535,6 +535,8 @@ typedef struct PartitionSchemeData *PartitionScheme; * part_scheme - Partitioning scheme of the relation * boundinfo - Partition bounds * nparts - Number of partitions + * has_default_part - Whether the table has a default partition + * partition_qual - Partition constraint if not the root * part_rels - RelOptInfos for each partition * partexprs, nullable_partexprs - Partition key expressions * @@ -667,6 +669,8 @@ typedef struct RelOptInfo PartitionScheme part_scheme; /* Partitioning scheme. */ int nparts; /* number of partitions */ struct PartitionBoundInfoData *boundinfo; /* Partition bounds */ + bool has_default_part; /* does it have a default partition? */ + List *partition_qual; /* partition constraint */ struct RelOptInfo **part_rels; /* Array of RelOptInfos of partitions, * stored in the same order of bounds */ List **partexprs; /* Non-nullable partition key expressions. */ diff --git a/src/include/optimizer/partprune.h b/src/include/optimizer/partprune.h new file mode 100644 index 0000000000..fb2f4b80fc --- /dev/null +++ b/src/include/optimizer/partprune.h @@ -0,0 +1,23 @@ +/*------------------------------------------------------------------------- + * + * partprune.h + * prototypes for partprune.c + * + * + * Portions Copyright (c) 1996-2018, PostgreSQL Global Development Group + * Portions Copyright (c) 1994, Regents of the University of California + * + * src/include/optimizer/partprune.h + * + *------------------------------------------------------------------------- + */ +#ifndef PARTPRUNE_H +#define PARTPRUNE_H + +#include "catalog/partition.h" + +extern Relids prune_append_rel_partitions(RelOptInfo *rel); +extern List *generate_partition_pruning_steps(RelOptInfo *rel, List *clauses, + bool static_pruning, bool *constfalse); + +#endif /* PARTPRUNE_H */ diff --git a/src/test/regress/expected/inherit.out b/src/test/regress/expected/inherit.out index f56151fc1e..d799acb91f 100644 --- a/src/test/regress/expected/inherit.out +++ b/src/test/regress/expected/inherit.out @@ -1926,11 +1926,13 @@ explain (costs off) select * from mcrparted where abs(b) = 5; -- scans all parti Filter: (abs(b) = 5) -> Seq Scan on mcrparted3 Filter: (abs(b) = 5) + -> Seq Scan on mcrparted4 + Filter: (abs(b) = 5) -> Seq Scan on mcrparted5 Filter: (abs(b) = 5) -> Seq Scan on mcrparted_def Filter: (abs(b) = 5) -(13 rows) +(15 rows) explain (costs off) select * from mcrparted where a > -1; -- scans all partitions QUERY PLAN diff --git a/src/test/regress/expected/partition_prune.out b/src/test/regress/expected/partition_prune.out index a0edba291f..0be31cce7e 100644 --- a/src/test/regress/expected/partition_prune.out +++ b/src/test/regress/expected/partition_prune.out @@ -208,16 +208,14 @@ explain (costs off) select * from rlp where 1 > a; /* commuted */ (3 rows) explain (costs off) select * from rlp where a <= 1; - QUERY PLAN ---------------------------------------- + QUERY PLAN +-------------------------- Append -> Seq Scan on rlp1 Filter: (a <= 1) -> Seq Scan on rlp2 Filter: (a <= 1) - -> Seq Scan on rlp_default_default - Filter: (a <= 1) -(7 rows) +(5 rows) explain (costs off) select * from rlp where a = 1; QUERY PLAN @@ -235,7 +233,7 @@ explain (costs off) select * from rlp where a = 1::bigint; /* same as above */ Filter: (a = '1'::bigint) (3 rows) -explain (costs off) select * from rlp where a = 1::numeric; /* only null can be pruned */ +explain (costs off) select * from rlp where a = 1::numeric; /* no pruning */ QUERY PLAN ----------------------------------------------- Append @@ -265,9 +263,11 @@ explain (costs off) select * from rlp where a = 1::numeric; /* only null can be Filter: ((a)::numeric = '1'::numeric) -> Seq Scan on rlp_default_30 Filter: ((a)::numeric = '1'::numeric) + -> Seq Scan on rlp_default_null + Filter: ((a)::numeric = '1'::numeric) -> Seq Scan on rlp_default_default Filter: ((a)::numeric = '1'::numeric) -(29 rows) +(31 rows) explain (costs off) select * from rlp where a <= 10; QUERY PLAN @@ -575,7 +575,9 @@ explain (costs off) select * from rlp where a > 20 and a < 27; Filter: ((a > 20) AND (a < 27)) -> Seq Scan on rlp4_default Filter: ((a > 20) AND (a < 27)) -(7 rows) + -> Seq Scan on rlp_default_default + Filter: ((a > 20) AND (a < 27)) +(9 rows) explain (costs off) select * from rlp where a = 29; QUERY PLAN @@ -714,9 +716,7 @@ explain (costs off) select * from mc3p where a = 1 and abs(b) = 1 and c < 8; Filter: ((c < 8) AND (a = 1) AND (abs(b) = 1)) -> Seq Scan on mc3p1 Filter: ((c < 8) AND (a = 1) AND (abs(b) = 1)) - -> Seq Scan on mc3p_default - Filter: ((c < 8) AND (a = 1) AND (abs(b) = 1)) -(7 rows) +(5 rows) explain (costs off) select * from mc3p where a = 10 and abs(b) between 5 and 35; QUERY PLAN @@ -892,6 +892,8 @@ explain (costs off) select * from mc3p where a = 1 or abs(b) = 1 or c = 1; Filter: ((a = 1) OR (abs(b) = 1) OR (c = 1)) -> Seq Scan on mc3p2 Filter: ((a = 1) OR (abs(b) = 1) OR (c = 1)) + -> Seq Scan on mc3p3 + Filter: ((a = 1) OR (abs(b) = 1) OR (c = 1)) -> Seq Scan on mc3p4 Filter: ((a = 1) OR (abs(b) = 1) OR (c = 1)) -> Seq Scan on mc3p5 @@ -902,7 +904,7 @@ explain (costs off) select * from mc3p where a = 1 or abs(b) = 1 or c = 1; Filter: ((a = 1) OR (abs(b) = 1) OR (c = 1)) -> Seq Scan on mc3p_default Filter: ((a = 1) OR (abs(b) = 1) OR (c = 1)) -(17 rows) +(19 rows) explain (costs off) select * from mc3p where (a = 1 and abs(b) = 1) or (a = 10 and abs(b) = 10); QUERY PLAN @@ -1007,24 +1009,20 @@ explain (costs off) select * from boolpart where a in (true, false); (5 rows) explain (costs off) select * from boolpart where a = false; - QUERY PLAN ------------------------------------- + QUERY PLAN +------------------------------ Append -> Seq Scan on boolpart_f Filter: (NOT a) - -> Seq Scan on boolpart_default - Filter: (NOT a) -(5 rows) +(3 rows) explain (costs off) select * from boolpart where not a = false; - QUERY PLAN ------------------------------------- + QUERY PLAN +------------------------------ Append -> Seq Scan on boolpart_t Filter: a - -> Seq Scan on boolpart_default - Filter: a -(5 rows) +(3 rows) explain (costs off) select * from boolpart where a is true or a is not true; QUERY PLAN @@ -1034,33 +1032,22 @@ explain (costs off) select * from boolpart where a is true or a is not true; Filter: ((a IS TRUE) OR (a IS NOT TRUE)) -> Seq Scan on boolpart_t Filter: ((a IS TRUE) OR (a IS NOT TRUE)) - -> Seq Scan on boolpart_default - Filter: ((a IS TRUE) OR (a IS NOT TRUE)) -(7 rows) +(5 rows) explain (costs off) select * from boolpart where a is not true; - QUERY PLAN ------------------------------------- + QUERY PLAN +--------------------------------- Append -> Seq Scan on boolpart_f Filter: (a IS NOT TRUE) - -> Seq Scan on boolpart_t - Filter: (a IS NOT TRUE) - -> Seq Scan on boolpart_default - Filter: (a IS NOT TRUE) -(7 rows) +(3 rows) explain (costs off) select * from boolpart where a is not true and a is not false; - QUERY PLAN --------------------------------------------------------- - Append - -> Seq Scan on boolpart_f - Filter: ((a IS NOT TRUE) AND (a IS NOT FALSE)) - -> Seq Scan on boolpart_t - Filter: ((a IS NOT TRUE) AND (a IS NOT FALSE)) - -> Seq Scan on boolpart_default - Filter: ((a IS NOT TRUE) AND (a IS NOT FALSE)) -(7 rows) + QUERY PLAN +-------------------------- + Result + One-Time Filter: false +(2 rows) explain (costs off) select * from boolpart where a is unknown; QUERY PLAN @@ -1111,13 +1098,21 @@ explain (costs off) select * from mc2p t1, lateral (select count(*) from mc3p t2 Filter: ((a = t1.b) AND (c = 1) AND (abs(b) = 1)) -> Seq Scan on mc3p1 t2_1 Filter: ((a = t1.b) AND (c = 1) AND (abs(b) = 1)) - -> Seq Scan on mc3p5 t2_2 + -> Seq Scan on mc3p2 t2_2 Filter: ((a = t1.b) AND (c = 1) AND (abs(b) = 1)) - -> Seq Scan on mc3p7 t2_3 + -> Seq Scan on mc3p3 t2_3 Filter: ((a = t1.b) AND (c = 1) AND (abs(b) = 1)) - -> Seq Scan on mc3p_default t2_4 + -> Seq Scan on mc3p4 t2_4 Filter: ((a = t1.b) AND (c = 1) AND (abs(b) = 1)) -(20 rows) + -> Seq Scan on mc3p5 t2_5 + Filter: ((a = t1.b) AND (c = 1) AND (abs(b) = 1)) + -> Seq Scan on mc3p6 t2_6 + Filter: ((a = t1.b) AND (c = 1) AND (abs(b) = 1)) + -> Seq Scan on mc3p7 t2_7 + Filter: ((a = t1.b) AND (c = 1) AND (abs(b) = 1)) + -> Seq Scan on mc3p_default t2_8 + Filter: ((a = t1.b) AND (c = 1) AND (abs(b) = 1)) +(28 rows) -- pruning should work fine, because values for a prefix of keys (a, b) are -- available @@ -1275,22 +1270,16 @@ explain (costs off) select * from coll_pruning_multi where substr(a, 1) = 'a' co Filter: (substr(a, 1) = 'a'::text COLLATE "POSIX") -> Seq Scan on coll_pruning_multi2 Filter: (substr(a, 1) = 'a'::text COLLATE "POSIX") - -> Seq Scan on coll_pruning_multi3 - Filter: (substr(a, 1) = 'a'::text COLLATE "POSIX") -(7 rows) +(5 rows) -- pruning, with values provided for both keys explain (costs off) select * from coll_pruning_multi where substr(a, 1) = 'e' collate "C" and substr(a, 1) = 'a' collate "POSIX"; QUERY PLAN --------------------------------------------------------------------------------------------------------- Append - -> Seq Scan on coll_pruning_multi1 - Filter: ((substr(a, 1) = 'e'::text COLLATE "C") AND (substr(a, 1) = 'a'::text COLLATE "POSIX")) -> Seq Scan on coll_pruning_multi2 Filter: ((substr(a, 1) = 'e'::text COLLATE "C") AND (substr(a, 1) = 'a'::text COLLATE "POSIX")) - -> Seq Scan on coll_pruning_multi3 - Filter: ((substr(a, 1) = 'e'::text COLLATE "C") AND (substr(a, 1) = 'a'::text COLLATE "POSIX")) -(7 rows) +(3 rows) -- -- LIKE operators don't prune @@ -1343,3 +1332,188 @@ explain (costs off) select * from rparted_by_int2 where a > 100000000000000; (3 rows) drop table lp, coll_pruning, rlp, mc3p, mc2p, boolpart, rp, coll_pruning_multi, like_op_noprune, lparted_by_int2, rparted_by_int2; +-- hash partitioning +create table hp (a int, b text) partition by hash (a, b); +create table hp0 partition of hp for values with (modulus 4, remainder 0); +create table hp3 partition of hp for values with (modulus 4, remainder 3); +create table hp1 partition of hp for values with (modulus 4, remainder 1); +create table hp2 partition of hp for values with (modulus 4, remainder 2); +insert into hp values (null, null); +insert into hp values (1, null); +insert into hp values (1, 'xxx'); +insert into hp values (null, 'xxx'); +insert into hp values (10, 'xxx'); +insert into hp values (10, 'yyy'); +select tableoid::regclass, * from hp order by 1; + tableoid | a | b +----------+----+----- + hp0 | | + hp0 | 1 | + hp0 | 1 | xxx + hp3 | 10 | yyy + hp1 | | xxx + hp2 | 10 | xxx +(6 rows) + +-- partial keys won't prune, nor would non-equality conditions +explain (costs off) select * from hp where a = 1; + QUERY PLAN +------------------------- + Append + -> Seq Scan on hp0 + Filter: (a = 1) + -> Seq Scan on hp1 + Filter: (a = 1) + -> Seq Scan on hp2 + Filter: (a = 1) + -> Seq Scan on hp3 + Filter: (a = 1) +(9 rows) + +explain (costs off) select * from hp where b = 'xxx'; + QUERY PLAN +----------------------------------- + Append + -> Seq Scan on hp0 + Filter: (b = 'xxx'::text) + -> Seq Scan on hp1 + Filter: (b = 'xxx'::text) + -> Seq Scan on hp2 + Filter: (b = 'xxx'::text) + -> Seq Scan on hp3 + Filter: (b = 'xxx'::text) +(9 rows) + +explain (costs off) select * from hp where a is null; + QUERY PLAN +----------------------------- + Append + -> Seq Scan on hp0 + Filter: (a IS NULL) + -> Seq Scan on hp1 + Filter: (a IS NULL) + -> Seq Scan on hp2 + Filter: (a IS NULL) + -> Seq Scan on hp3 + Filter: (a IS NULL) +(9 rows) + +explain (costs off) select * from hp where b is null; + QUERY PLAN +----------------------------- + Append + -> Seq Scan on hp0 + Filter: (b IS NULL) + -> Seq Scan on hp1 + Filter: (b IS NULL) + -> Seq Scan on hp2 + Filter: (b IS NULL) + -> Seq Scan on hp3 + Filter: (b IS NULL) +(9 rows) + +explain (costs off) select * from hp where a < 1 and b = 'xxx'; + QUERY PLAN +------------------------------------------------- + Append + -> Seq Scan on hp0 + Filter: ((a < 1) AND (b = 'xxx'::text)) + -> Seq Scan on hp1 + Filter: ((a < 1) AND (b = 'xxx'::text)) + -> Seq Scan on hp2 + Filter: ((a < 1) AND (b = 'xxx'::text)) + -> Seq Scan on hp3 + Filter: ((a < 1) AND (b = 'xxx'::text)) +(9 rows) + +explain (costs off) select * from hp where a <> 1 and b = 'yyy'; + QUERY PLAN +-------------------------------------------------- + Append + -> Seq Scan on hp0 + Filter: ((a <> 1) AND (b = 'yyy'::text)) + -> Seq Scan on hp1 + Filter: ((a <> 1) AND (b = 'yyy'::text)) + -> Seq Scan on hp2 + Filter: ((a <> 1) AND (b = 'yyy'::text)) + -> Seq Scan on hp3 + Filter: ((a <> 1) AND (b = 'yyy'::text)) +(9 rows) + +-- pruning should work if non-null values are provided for all the keys +explain (costs off) select * from hp where a is null and b is null; + QUERY PLAN +----------------------------------------------- + Append + -> Seq Scan on hp0 + Filter: ((a IS NULL) AND (b IS NULL)) +(3 rows) + +explain (costs off) select * from hp where a = 1 and b is null; + QUERY PLAN +------------------------------------------- + Append + -> Seq Scan on hp0 + Filter: ((b IS NULL) AND (a = 1)) +(3 rows) + +explain (costs off) select * from hp where a = 1 and b = 'xxx'; + QUERY PLAN +------------------------------------------------- + Append + -> Seq Scan on hp0 + Filter: ((a = 1) AND (b = 'xxx'::text)) +(3 rows) + +explain (costs off) select * from hp where a is null and b = 'xxx'; + QUERY PLAN +----------------------------------------------------- + Append + -> Seq Scan on hp1 + Filter: ((a IS NULL) AND (b = 'xxx'::text)) +(3 rows) + +explain (costs off) select * from hp where a = 10 and b = 'xxx'; + QUERY PLAN +-------------------------------------------------- + Append + -> Seq Scan on hp2 + Filter: ((a = 10) AND (b = 'xxx'::text)) +(3 rows) + +explain (costs off) select * from hp where a = 10 and b = 'yyy'; + QUERY PLAN +-------------------------------------------------- + Append + -> Seq Scan on hp3 + Filter: ((a = 10) AND (b = 'yyy'::text)) +(3 rows) + +explain (costs off) select * from hp where (a = 10 and b = 'yyy') or (a = 10 and b = 'xxx') or (a is null and b is null); + QUERY PLAN +------------------------------------------------------------------------------------------------------------------------- + Append + -> Seq Scan on hp0 + Filter: (((a = 10) AND (b = 'yyy'::text)) OR ((a = 10) AND (b = 'xxx'::text)) OR ((a IS NULL) AND (b IS NULL))) + -> Seq Scan on hp2 + Filter: (((a = 10) AND (b = 'yyy'::text)) OR ((a = 10) AND (b = 'xxx'::text)) OR ((a IS NULL) AND (b IS NULL))) + -> Seq Scan on hp3 + Filter: (((a = 10) AND (b = 'yyy'::text)) OR ((a = 10) AND (b = 'xxx'::text)) OR ((a IS NULL) AND (b IS NULL))) +(7 rows) + +-- hash partitiong pruning doesn't occur with <> operator clauses +explain (costs off) select * from hp where a <> 1 and b <> 'xxx'; + QUERY PLAN +--------------------------------------------------- + Append + -> Seq Scan on hp0 + Filter: ((a <> 1) AND (b <> 'xxx'::text)) + -> Seq Scan on hp1 + Filter: ((a <> 1) AND (b <> 'xxx'::text)) + -> Seq Scan on hp2 + Filter: ((a <> 1) AND (b <> 'xxx'::text)) + -> Seq Scan on hp3 + Filter: ((a <> 1) AND (b <> 'xxx'::text)) +(9 rows) + +drop table hp; diff --git a/src/test/regress/sql/partition_prune.sql b/src/test/regress/sql/partition_prune.sql index d2b4561530..8377671cde 100644 --- a/src/test/regress/sql/partition_prune.sql +++ b/src/test/regress/sql/partition_prune.sql @@ -60,7 +60,7 @@ explain (costs off) select * from rlp where 1 > a; /* commuted */ explain (costs off) select * from rlp where a <= 1; explain (costs off) select * from rlp where a = 1; explain (costs off) select * from rlp where a = 1::bigint; /* same as above */ -explain (costs off) select * from rlp where a = 1::numeric; /* only null can be pruned */ +explain (costs off) select * from rlp where a = 1::numeric; /* no pruning */ explain (costs off) select * from rlp where a <= 10; explain (costs off) select * from rlp where a > 10; explain (costs off) select * from rlp where a < 15; @@ -239,3 +239,40 @@ create table rparted_by_int2_maxvalue partition of rparted_by_int2 for values fr explain (costs off) select * from rparted_by_int2 where a > 100000000000000; drop table lp, coll_pruning, rlp, mc3p, mc2p, boolpart, rp, coll_pruning_multi, like_op_noprune, lparted_by_int2, rparted_by_int2; + +-- hash partitioning +create table hp (a int, b text) partition by hash (a, b); +create table hp0 partition of hp for values with (modulus 4, remainder 0); +create table hp3 partition of hp for values with (modulus 4, remainder 3); +create table hp1 partition of hp for values with (modulus 4, remainder 1); +create table hp2 partition of hp for values with (modulus 4, remainder 2); + +insert into hp values (null, null); +insert into hp values (1, null); +insert into hp values (1, 'xxx'); +insert into hp values (null, 'xxx'); +insert into hp values (10, 'xxx'); +insert into hp values (10, 'yyy'); +select tableoid::regclass, * from hp order by 1; + +-- partial keys won't prune, nor would non-equality conditions +explain (costs off) select * from hp where a = 1; +explain (costs off) select * from hp where b = 'xxx'; +explain (costs off) select * from hp where a is null; +explain (costs off) select * from hp where b is null; +explain (costs off) select * from hp where a < 1 and b = 'xxx'; +explain (costs off) select * from hp where a <> 1 and b = 'yyy'; + +-- pruning should work if non-null values are provided for all the keys +explain (costs off) select * from hp where a is null and b is null; +explain (costs off) select * from hp where a = 1 and b is null; +explain (costs off) select * from hp where a = 1 and b = 'xxx'; +explain (costs off) select * from hp where a is null and b = 'xxx'; +explain (costs off) select * from hp where a = 10 and b = 'xxx'; +explain (costs off) select * from hp where a = 10 and b = 'yyy'; +explain (costs off) select * from hp where (a = 10 and b = 'yyy') or (a = 10 and b = 'xxx') or (a is null and b is null); + +-- hash partitiong pruning doesn't occur with <> operator clauses +explain (costs off) select * from hp where a <> 1 and b <> 'xxx'; + +drop table hp; diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list index 6d8a44cd9e..aa2ec281c4 100644 --- a/src/tools/pgindent/typedefs.list +++ b/src/tools/pgindent/typedefs.list @@ -830,6 +830,7 @@ GatherMergeState GatherPath GatherState Gene +GeneratePruningStepsContext GenerationBlock GenerationChunk GenerationContext @@ -1587,6 +1588,7 @@ ParsedText ParsedWord ParserSetupHook ParserState +PartClauseInfo PartitionBoundInfo PartitionBoundInfoData PartitionBoundSpec @@ -1599,6 +1601,10 @@ PartitionElem PartitionHashBound PartitionKey PartitionListValue +PartitionPruneContext +PartitionPruneStep +PartitionPruneStepCombine +PartitionPruneStepOp PartitionRangeBound PartitionRangeDatum PartitionRangeDatumKind @@ -1752,6 +1758,7 @@ ProjectionPath ProtocolVersion PrsStorage PruneState +PruneStepResult PsqlScanCallbacks PsqlScanQuoteType PsqlScanResult -- 2.11.0