\echo Use "CREATE EXTENSION pg_sift" to load this file. \quit

CREATE FUNCTION sift.sift_handler(internal) RETURNS index_am_handler
	AS 'MODULE_PATHNAME', 'sift_handler' LANGUAGE C STRICT;

CREATE ACCESS METHOD sift TYPE INDEX HANDLER sift.sift_handler;

CREATE FUNCTION sift.match_support(internal) RETURNS internal
	AS 'MODULE_PATHNAME', 'sift_match_support' LANGUAGE C STRICT;

CREATE FUNCTION sift.matchsel(internal, oid, internal, integer) RETURNS float8
	AS 'MODULE_PATHNAME', 'sift_matchsel' LANGUAGE C STRICT STABLE PARALLEL SAFE;

-- Default analyzer; the planner support function rebinds calls on columns
-- whose sift index uses another analyzer to sift.match_idx.
CREATE FUNCTION sift.match(doc text, query text) RETURNS bool
	AS 'MODULE_PATHNAME', 'sift_match'
	LANGUAGE C STRICT IMMUTABLE PARALLEL SAFE COST 1000 SUPPORT sift.match_support;

CREATE FUNCTION sift.match_idx(doc text, query text, index oid) RETURNS bool
	AS 'MODULE_PATHNAME', 'sift_match_idx'
	LANGUAGE C STRICT STABLE PARALLEL SAFE COST 1000 SUPPORT sift.match_support;

-- In pg_catalog so `col @@> 'query'` works without changing search_path.
CREATE OPERATOR pg_catalog.@@> (
	LEFTARG = text,
	RIGHTARG = text,
	FUNCTION = sift.match,
	RESTRICT = sift.matchsel,
	JOIN = contjoinsel
);

CREATE OPERATOR CLASS sift.text_ops DEFAULT FOR TYPE text USING sift AS
	OPERATOR 1 pg_catalog.@@> (text, text);

-- Computed by the sift scan and projected through the plan; calling them
-- anywhere else raises an error.
CREATE FUNCTION sift.score(tid) RETURNS real
	AS 'MODULE_PATHNAME', 'sift_score' LANGUAGE C STRICT STABLE PARALLEL SAFE;

CREATE FUNCTION sift.max_score(tid) RETURNS real
	AS 'MODULE_PATHNAME', 'sift_max_score' LANGUAGE C STRICT STABLE PARALLEL SAFE;

CREATE FUNCTION sift.highlight(doc text, query text,
							   start_tag text DEFAULT '<b>', stop_tag text DEFAULT '</b>',
							   index regclass DEFAULT NULL) RETURNS text
	AS 'MODULE_PATHNAME', 'sift_highlight' LANGUAGE C STABLE PARALLEL SAFE;

-- Analyzer of index (or the default one), overridden by non-NULL options;
-- offsets are 0-based character offsets [start, end) into doc.
CREATE FUNCTION sift.tokenize(doc text, index regclass DEFAULT NULL,
							  stemmer text DEFAULT NULL, case_folding bool DEFAULT NULL,
							  accent_folding bool DEFAULT NULL,
							  OUT token text, OUT pos int4,
							  OUT start_offset int4, OUT end_offset int4)
	RETURNS SETOF record
	AS 'MODULE_PATHNAME', 'sift_tokenize' LANGUAGE C STABLE PARALLEL SAFE;

CREATE FUNCTION sift.parse_query(query text, index regclass DEFAULT NULL) RETURNS text
	AS 'MODULE_PATHNAME', 'sift_parse_query' LANGUAGE C STABLE PARALLEL SAFE;

CREATE FUNCTION sift.promote(index regclass) RETURNS int4
	AS 'MODULE_PATHNAME', 'sift_promote' LANGUAGE C STRICT VOLATILE;

CREATE FUNCTION sift.merge(index regclass, force bool DEFAULT false) RETURNS int4
	AS 'MODULE_PATHNAME', 'sift_merge' LANGUAGE C STRICT VOLATILE;

-- One row per segment; the section columns are NULL for write buffers (extents and dead_docs too).
-- dead_docs: documents of an immutable segment that VACUUM has marked dead (rows of rolled-back
-- transactions and deleted rows still count in ndocs, the document frequencies and avgdl until then).
CREATE FUNCTION sift.segments(index regclass,
							  OUT gen int8, OUT kind text, OUT ndocs int8, OUT sum_dl int8,
							  OUT avgdl float8, OUT bytes int8, OUT sb_lo int4, OUT sb_hi int4,
							  OUT dirty bool, OUT nterms int8, OUT avgdl_ref float8,
							  OUT docmap_bytes int8, OUT postings_bytes int8, OUT dict_bytes int8,
							  OUT live_bytes int8, OUT extents int4, OUT dead_docs int8)
	RETURNS SETOF record
	AS 'MODULE_PATHNAME', 'sift_segments' LANGUAGE C STRICT VOLATILE;

-- Postings by segment and df class (decodes every term): representation
-- counts, codec and sidecar (tf, bounds) bytes, bits per posting.
CREATE FUNCTION sift.segment_postings(index regclass,
									  OUT gen int8, OUT df_class text, OUT terms int8,
									  OUT postings int8, OUT inline_terms int8, OUT list_terms int8,
									  OUT grouped_terms int8, OUT codec_bytes int8,
									  OUT sidecar_bytes int8, OUT bits_per_posting float8)
	RETURNS SETOF record
	AS 'MODULE_PATHNAME', 'sift_segment_postings' LANGUAGE C STRICT VOLATILE;

CREATE FUNCTION sift.verify(index regclass, heapallindexed bool DEFAULT false) RETURNS void
	AS 'MODULE_PATHNAME', 'sift_verify' LANGUAGE C STRICT VOLATILE;

-- Keyword facets (DESIGN.md item 11): the number of matches of a query per value of a keyword index.
-- A keyword index is CREATE INDEX ... USING sift (col) WITH (analyzer = keyword).

-- Counts per value of facet_index among the visible rows that index matches query for, count
-- descending then value; lim NULL = every value, nulls = also a row with a NULL value counting the
-- matches without one (it makes every value be counted, so lim cannot end the search early).
CREATE FUNCTION sift.facets(index regclass, facet_index regclass, query text,
							lim int DEFAULT 10, nulls bool DEFAULT false,
							OUT value text, OUT count int8)
	RETURNS SETOF record
	AS 'MODULE_PATHNAME', 'sift_facet_counts' LANGUAGE C STABLE;

-- The space and directory state of an index. relation_blocks: size of the relation; high_water: blocks
-- ever allocated; free_blocks: blocks on the free list, of which reusable_blocks can be taken now (the
-- rest wait for the snapshots that may still read them); maintenance_busy: a promotion, merge, VACUUM or
-- verify holds the maintenance lock (NULL on a standby).
CREATE FUNCTION sift.index_info(index regclass,
								OUT relation_blocks int8, OUT high_water int8, OUT free_blocks int8,
								OUT reusable_blocks int8, OUT free_extents int8, OUT segments int4,
								OUT max_segments int4, OUT ndocs int8, OUT sum_dl int8,
								OUT maintenance_busy bool)
	RETURNS record
	AS 'MODULE_PATHNAME', 'sift_index_info' LANGUAGE C STRICT VOLATILE;

-- The version of the loaded library; the installed extension's is in pg_extension.
CREATE FUNCTION sift.version() RETURNS text
	AS 'MODULE_PATHNAME', 'sift_version' LANGUAGE C STABLE;

-- The sift_col index access method (docs/COL.md): a columnar index over an existing heap table.

CREATE FUNCTION sift.col_handler(internal) RETURNS index_am_handler
	AS 'MODULE_PATHNAME', 'sift_col_handler' LANGUAGE C STRICT;

CREATE ACCESS METHOD sift_col TYPE INDEX HANDLER sift.col_handler;

-- No operators: the index is read by the column scans, not through search operators. An operator class needs one member, which
-- the access method never calls.
CREATE FUNCTION sift.col_ops_unused(anyelement) RETURNS void
	LANGUAGE sql IMMUTABLE PARALLEL SAFE AS 'SELECT NULL::void';

CREATE OPERATOR CLASS sift.col_ops DEFAULT FOR TYPE anyelement USING sift_col AS FUNCTION 1 sift.col_ops_unused(anyelement);

-- Rows are the live rows of the segments; units are (segment, 256-page superblock) pairs. simd: the kernel level the scans use.
CREATE FUNCTION sift.col_index_info(index regclass,
									OUT relation_blocks int8, OUT high_water int8, OUT free_blocks int8,
									OUT reusable_blocks int8, OUT free_extents int8, OUT segments int4,
									OUT max_segments int4, OUT rows int8, OUT bytes int8, OUT units int8,
									OUT buffers int4, OUT leaves int4, OUT maintenance_busy bool,
									OUT simd text, OUT preagg_bytes int8, OUT preagg_units int8)
	RETURNS record
	AS 'MODULE_PATHNAME', 'sift_col_index_info' LANGUAGE C STRICT VOLATILE;

-- The pre-aggregates of every stored unit (docs/COL-PREAGG.md): the aggregate of a leaf, and for a dimension the histogram of its groups with the
-- count and sum of one measure. exact: the unit may answer from them (clean, every page all-visible). Debug and test hooks.
CREATE FUNCTION sift.col_unit_agg(index regclass, leaf text,
								  OUT segment int8, OUT sb int4, OUT rows int4, OUT exact bool, OUT nvals int8, OUT nnull int8,
								  OUT min text, OUT max text, OUT sum numeric, OUT sumsq numeric, OUT sum_float float8)
	RETURNS SETOF record
	AS 'MODULE_PATHNAME', 'sift_col_unit_aggs_sql' LANGUAGE C VOLATILE;

CREATE FUNCTION sift.col_unit_hist(index regclass, dim text, measure text DEFAULT NULL,
								   OUT segment int8, OUT sb int4, OUT exact bool, OUT key text, OUT n int8,
								   OUT nn int8, OUT sum numeric, OUT sum_float float8)
	RETURNS SETOF record
	AS 'MODULE_PATHNAME', 'sift_col_unit_hists_sql' LANGUAGE C VOLATILE;

-- One row per segment; the immutable-only columns are NULL for write buffers.
CREATE FUNCTION sift.col_segments(index regclass,
								  OUT gen int8, OUT kind text, OUT rows int8, OUT bytes int8,
								  OUT sb_lo int4, OUT sb_hi int4, OUT units int4, OUT dirty bool,
								  OUT dead_rows int8, OUT extents int4)
	RETURNS SETOF record
	AS 'MODULE_PATHNAME', 'sift_col_segments' LANGUAGE C STRICT VOLATILE;

-- One row per leaf: its chunks by encoding (const, for, pack, other, text stored inline), their bytes, and the bytes of its dictionaries (text leaves).
CREATE FUNCTION sift.col_leaves(index regclass,
								OUT leaf int4, OUT path text, OUT kind text, OUT type text,
								OUT const_chunks int8, OUT for_chunks int8, OUT pack_chunks int8, OUT codec_chunks int8,
								OUT inline_chunks int8, OUT chunk_bytes int8, OUT dict_bytes int8)
	RETURNS SETOF record
	AS 'MODULE_PATHNAME', 'sift_col_leaves' LANGUAGE C STRICT VOLATILE;

-- Where the bytes of a sift_col index or a sift_table go, one row per kind: chunks and dict per leaf; unit_dir, dict_other, zones, preagg,
-- sbdir_schema, page_overhead, headers (segment structure), free (blocks on the free list) and other (metapage, directory, write buffer, tombstones).
-- The bytes add up to pg_relation_size.
CREATE FUNCTION sift.col_sections(rel regclass, OUT section text, OUT leaf text, OUT bytes int8)
	RETURNS SETOF record
	AS 'MODULE_PATHNAME', 'sift_col_sections' LANGUAGE C STRICT VOLATILE;

CREATE FUNCTION sift.col_verify(index regclass, heapallindexed bool DEFAULT false) RETURNS void
	AS 'MODULE_PATHNAME', 'sift_col_verify' LANGUAGE C STRICT VOLATILE;

CREATE FUNCTION sift.col_promote(index regclass) RETURNS int4
	AS 'MODULE_PATHNAME', 'sift_col_promote' LANGUAGE C STRICT VOLATILE;

CREATE FUNCTION sift.col_merge(index regclass, force bool DEFAULT false) RETURNS int4
	AS 'MODULE_PATHNAME', 'sift_col_merge' LANGUAGE C STRICT VOLATILE;

-- The sift_table table access method (docs/COL-TAM.md): CREATE TABLE ... USING sift_table.
CREATE FUNCTION sift.table_handler(internal) RETURNS table_am_handler
	AS 'MODULE_PATHNAME', 'sift_table_handler' LANGUAGE C STRICT;

CREATE ACCESS METHOD sift_table TYPE TABLE HANDLER sift.table_handler;

-- One row per segment of a sift_table: kind buffer (the row format write buffer; bytes = its pages), sealed buffer (no more rows,
-- waiting to become columns), columns (a column segment: rows including dead ones, units, dead rows) or tombstones.
CREATE FUNCTION sift.table_segments(tbl regclass,
									OUT gen int8, OUT kind text, OUT rows int8, OUT bytes int8,
									OUT sb_lo int4, OUT sb_hi int4, OUT units int4, OUT dead_rows int8)
	RETURNS SETOF record
	AS 'MODULE_PATHNAME', 'sift_table_segments' LANGUAGE C STRICT VOLATILE;

-- The space of a sift_table: relation_blocks (the file), high_water (blocks ever used), free_blocks (on the free list, of which
-- reusable_blocks can be taken now, the rest wait for snapshots that may still read them), free_extents.
CREATE FUNCTION sift.table_space(tbl regclass,
								 OUT relation_blocks int8, OUT high_water int8, OUT free_blocks int8, OUT reusable_blocks int8,
								 OUT free_extents int8)
	RETURNS record
	AS 'MODULE_PATHNAME', 'sift_table_space' LANGUAGE C STRICT VOLATILE;