Use perfect hashing, instead of binary search, for keyword lookup.

We've been speculating for a long time that hash-based keyword lookup
ought to be faster than binary search, but up to now we hadn't found
a suitable tool for generating the hash function.  Joerg Sonnenberger
provided the inspiration, and sample code, to show us that rolling our
own generator wasn't a ridiculous idea.  Hence, do that.

The method used here requires a lookup table of approximately 4 bytes
per keyword, but that's less than what we saved in the predecessor commit
afb0d071, so it's not a big problem.  The time savings is indeed
significant: preliminary testing suggests that the total time for raw
parsing (flex + bison phases) drops by ~20%.

Patch by me, but it owes its existence to Joerg Sonnenberger;
thanks also to John Naylor for review.

Discussion: https://postgr.es/m/20190103163340.GA15803@britannica.bec.de

src/common/Makefile

@@ -63,6 +63,11 @@ OBJS_FRONTEND = $(OBJS_COMMON) fe_memutils.o file_utils.o restricted_token.o
 OBJS_SHLIB = $(OBJS_FRONTEND:%.o=%_shlib.o)
 OBJS_SRV = $(OBJS_COMMON:%.o=%_srv.o)
 
+# where to find gen_keywordlist.pl and subsidiary files
+TOOLSDIR = $(top_srcdir)/src/tools
+GEN_KEYWORDLIST = $(PERL) -I $(TOOLSDIR) $(TOOLSDIR)/gen_keywordlist.pl
+GEN_KEYWORDLIST_DEPS = $(TOOLSDIR)/gen_keywordlist.pl $(TOOLSDIR)/PerfectHash.pm
+
 all: libpgcommon.a libpgcommon_shlib.a libpgcommon_srv.a
 
 distprep: kwlist_d.h
@@ -118,8 +123,8 @@ libpgcommon_srv.a: $(OBJS_SRV)
 	$(CC) $(CFLAGS) $(subst -DFRONTEND,, $(CPPFLAGS)) -c $< -o $@
 
 # generate SQL keyword lookup table to be included into keywords*.o.
-kwlist_d.h: $(top_srcdir)/src/include/parser/kwlist.h $(top_srcdir)/src/tools/gen_keywordlist.pl
-	$(PERL) $(top_srcdir)/src/tools/gen_keywordlist.pl --extern $<
+kwlist_d.h: $(top_srcdir)/src/include/parser/kwlist.h $(GEN_KEYWORDLIST_DEPS)
+	$(GEN_KEYWORDLIST) --extern $<
 
 # Dependencies of keywords*.o need to be managed explicitly to make sure
 # that you don't get broken parsing code, even in a non-enable-depend build.

src/common/kwlookup.c

@@ -35,60 +35,51 @@
  * receive a different case-normalization mapping.
  */
 int
-ScanKeywordLookup(const char *text,
+ScanKeywordLookup(const char *str,
 				  const ScanKeywordList *keywords)
 {
-	int			len,
-				i;
-	char		word[NAMEDATALEN];
-	const char *kw_string;
-	const uint16 *kw_offsets;
-	const uint16 *low;
-	const uint16 *high;
-
-	len = strlen(text);
+	size_t		len;
+	int			h;
+	const char *kw;
 
+	/*
+	 * Reject immediately if too long to be any keyword.  This saves useless
+	 * hashing and downcasing work on long strings.
+	 */
+	len = strlen(str);
 	if (len > keywords->max_kw_len)
-		return -1;				/* too long to be any keyword */
-
-	/* We assume all keywords are shorter than NAMEDATALEN. */
-	Assert(len < NAMEDATALEN);
+		return -1;
 
 	/*
-	 * Apply an ASCII-only downcasing.  We must not use tolower() since it may
-	 * produce the wrong translation in some locales (eg, Turkish).
+	 * Compute the hash function.  We assume it was generated to produce
+	 * case-insensitive results.  Since it's a perfect hash, we need only
+	 * match to the specific keyword it identifies.
 	 */
-	for (i = 0; i < len; i++)
-	{
-		char		ch = text[i];
+	h = keywords->hash(str, len);
 
-		if (ch >= 'A' && ch <= 'Z')
-			ch += 'a' - 'A';
-		word[i] = ch;
-	}
-	word[len] = '\0';
+	/* An out-of-range result implies no match */
+	if (h < 0 || h >= keywords->num_keywords)
+		return -1;
 
 	/*
-	 * Now do a binary search using plain strcmp() comparison.
+	 * Compare character-by-character to see if we have a match, applying an
+	 * ASCII-only downcasing to the input characters.  We must not use
+	 * tolower() since it may produce the wrong translation in some locales
+	 * (eg, Turkish).
 	 */
-	kw_string = keywords->kw_string;
-	kw_offsets = keywords->kw_offsets;
-	low = kw_offsets;
-	high = kw_offsets + (keywords->num_keywords - 1);
-	while (low <= high)
+	kw = GetScanKeyword(h, keywords);
+	while (*str != '\0')
 	{
-		const uint16 *middle;
-		int			difference;
+		char		ch = *str++;
 
-		middle = low + (high - low) / 2;
-		difference = strcmp(kw_string + *middle, word);
-		if (difference == 0)
-			return middle - kw_offsets;
-		else if (difference < 0)
-			low = middle + 1;
-		else
-			high = middle - 1;
+		if (ch >= 'A' && ch <= 'Z')
+			ch += 'a' - 'A';
+		if (ch != *kw++)
+			return -1;
 	}
+	if (*kw != '\0')
+		return -1;
 
-	return -1;
+	/* Success! */
+	return h;
 }

src/include/common/kwlookup.h

@@ -14,6 +14,9 @@
 #ifndef KWLOOKUP_H
 #define KWLOOKUP_H
 
+/* Hash function used by ScanKeywordLookup */
+typedef int (*ScanKeywordHashFunc) (const void *key, size_t keylen);
+
 /*
  * This struct contains the data needed by ScanKeywordLookup to perform a
  * search within a set of keywords.  The contents are typically generated by
@@ -23,6 +26,7 @@ typedef struct ScanKeywordList
 {
 	const char *kw_string;		/* all keywords in order, separated by \0 */
 	const uint16 *kw_offsets;	/* offsets to the start of each keyword */
+	ScanKeywordHashFunc hash;	/* perfect hash function for keywords */
 	int			num_keywords;	/* number of keywords */
 	int			max_kw_len;		/* length of longest keyword */
 } ScanKeywordList;

src/include/parser/kwlist.h

@@ -21,8 +21,7 @@
 /*
  * List of keyword (name, token-value, category) entries.
  *
- * !!WARNING!!: This list must be sorted by ASCII name, because binary
- *		 search is used to locate entries.
+ * Note: gen_keywordlist.pl requires the entries to appear in ASCII order.
  */
 
 /* name, value, category */

src/interfaces/ecpg/preproc/Makefile

@@ -28,7 +28,10 @@ OBJS=	preproc.o pgc.o type.o ecpg.o output.o parser.o \
 	keywords.o c_keywords.o ecpg_keywords.o typename.o descriptor.o variable.o \
 	$(WIN32RES)
 
-GEN_KEYWORDLIST = $(top_srcdir)/src/tools/gen_keywordlist.pl
+# where to find gen_keywordlist.pl and subsidiary files
+TOOLSDIR = $(top_srcdir)/src/tools
+GEN_KEYWORDLIST = $(PERL) -I $(TOOLSDIR) $(TOOLSDIR)/gen_keywordlist.pl
+GEN_KEYWORDLIST_DEPS = $(TOOLSDIR)/gen_keywordlist.pl $(TOOLSDIR)/PerfectHash.pm
 
 # Suppress parallel build to avoid a bug in GNU make 3.82
 # (see comments in ../Makefile)
@@ -56,11 +59,11 @@ preproc.y: ../../../backend/parser/gram.y parse.pl ecpg.addons ecpg.header ecpg.
 	$(PERL) $(srcdir)/check_rules.pl $(srcdir) $<
 
 # generate keyword headers
-c_kwlist_d.h: c_kwlist.h $(GEN_KEYWORDLIST)
-	$(PERL) $(GEN_KEYWORDLIST) --varname ScanCKeywords $<
+c_kwlist_d.h: c_kwlist.h $(GEN_KEYWORDLIST_DEPS)
+	$(GEN_KEYWORDLIST) --varname ScanCKeywords --no-case-fold $<
 
-ecpg_kwlist_d.h: ecpg_kwlist.h $(GEN_KEYWORDLIST)
-	$(PERL) $(GEN_KEYWORDLIST) --varname ScanECPGKeywords $<
+ecpg_kwlist_d.h: ecpg_kwlist.h $(GEN_KEYWORDLIST_DEPS)
+	$(GEN_KEYWORDLIST) --varname ScanECPGKeywords $<
 
 # Force these dependencies to be known even without dependency info built:
 ecpg_keywords.o c_keywords.o keywords.o preproc.o pgc.o parser.o: preproc.h

src/interfaces/ecpg/preproc/c_keywords.c

@@ -9,8 +9,6 @@
  */
 #include "postgres_fe.h"
 
-#include <ctype.h>
-
 #include "preproc_extern.h"
 #include "preproc.h"
 
@@ -32,39 +30,38 @@ static const uint16 ScanCKeywordTokens[] = {
  *
  * Returns the token value of the keyword, or -1 if no match.
  *
- * Do a binary search using plain strcmp() comparison.  This is much like
+ * Do a hash search using plain strcmp() comparison.  This is much like
  * ScanKeywordLookup(), except we want case-sensitive matching.
  */
 int
-ScanCKeywordLookup(const char *text)
+ScanCKeywordLookup(const char *str)
 {
-	const char *kw_string;
-	const uint16 *kw_offsets;
-	const uint16 *low;
-	const uint16 *high;
+	size_t		len;
+	int			h;
+	const char *kw;
+
+	/*
+	 * Reject immediately if too long to be any keyword.  This saves useless
+	 * hashing work on long strings.
+	 */
+	len = strlen(str);
+	if (len > ScanCKeywords.max_kw_len)
+		return -1;
 
-	if (strlen(text) > ScanCKeywords.max_kw_len)
-		return -1;				/* too long to be any keyword */
+	/*
+	 * Compute the hash function.  Since it's a perfect hash, we need only
+	 * match to the specific keyword it identifies.
+	 */
+	h = ScanCKeywords_hash_func(str, len);
 
-	kw_string = ScanCKeywords.kw_string;
-	kw_offsets = ScanCKeywords.kw_offsets;
-	low = kw_offsets;
-	high = kw_offsets + (ScanCKeywords.num_keywords - 1);
+	/* An out-of-range result implies no match */
+	if (h < 0 || h >= ScanCKeywords.num_keywords)
+		return -1;
 
-	while (low <= high)
-	{
-		const uint16 *middle;
-		int			difference;
+	kw = GetScanKeyword(h, &ScanCKeywords);
 
-		middle = low + (high - low) / 2;
-		difference = strcmp(kw_string + *middle, text);
-		if (difference == 0)
-			return ScanCKeywordTokens[middle - kw_offsets];
-		else if (difference < 0)
-			low = middle + 1;
-		else
-			high = middle - 1;
-	}
+	if (strcmp(kw, str) == 0)
+		return ScanCKeywordTokens[h];
 
 	return -1;
 }

src/interfaces/ecpg/preproc/c_kwlist.h

@@ -20,8 +20,7 @@
 /*
  * List of (keyword-name, keyword-token-value) pairs.
  *
- * !!WARNING!!: This list must be sorted by ASCII name, because binary
- *		 search is used to locate entries.
+ * Note: gen_keywordlist.pl requires the entries to appear in ASCII order.
  */
 
 /* name, value */

src/interfaces/ecpg/preproc/ecpg_kwlist.h

@@ -20,8 +20,7 @@
 /*
  * List of (keyword-name, keyword-token-value) pairs.
  *
- * !!WARNING!!: This list must be sorted by ASCII name, because binary
- *		 search is used to locate entries.
+ * Note: gen_keywordlist.pl requires the entries to appear in ASCII order.
  */
 
 /* name, value */

src/pl/plpgsql/src/Makefile

@@ -29,7 +29,10 @@ REGRESS_OPTS = --dbname=$(PL_TESTDB)
 REGRESS = plpgsql_call plpgsql_control plpgsql_domain plpgsql_record \
 	plpgsql_cache plpgsql_transaction plpgsql_trigger plpgsql_varprops
 
-GEN_KEYWORDLIST = $(top_srcdir)/src/tools/gen_keywordlist.pl
+# where to find gen_keywordlist.pl and subsidiary files
+TOOLSDIR = $(top_srcdir)/src/tools
+GEN_KEYWORDLIST = $(PERL) -I $(TOOLSDIR) $(TOOLSDIR)/gen_keywordlist.pl
+GEN_KEYWORDLIST_DEPS = $(TOOLSDIR)/gen_keywordlist.pl $(TOOLSDIR)/PerfectHash.pm
 
 all: all-lib
 
@@ -76,11 +79,11 @@ plerrcodes.h: $(top_srcdir)/src/backend/utils/errcodes.txt generate-plerrcodes.p
 	$(PERL) $(srcdir)/generate-plerrcodes.pl $< > $@
 
 # generate keyword headers for the scanner
-pl_reserved_kwlist_d.h: pl_reserved_kwlist.h $(GEN_KEYWORDLIST)
-	$(PERL) $(GEN_KEYWORDLIST) --varname ReservedPLKeywords $<
+pl_reserved_kwlist_d.h: pl_reserved_kwlist.h $(GEN_KEYWORDLIST_DEPS)
+	$(GEN_KEYWORDLIST) --varname ReservedPLKeywords $<
 
-pl_unreserved_kwlist_d.h: pl_unreserved_kwlist.h $(GEN_KEYWORDLIST)
-	$(PERL) $(GEN_KEYWORDLIST) --varname UnreservedPLKeywords $<
+pl_unreserved_kwlist_d.h: pl_unreserved_kwlist.h $(GEN_KEYWORDLIST_DEPS)
+	$(GEN_KEYWORDLIST) --varname UnreservedPLKeywords $<
 
 
 check: submake

src/pl/plpgsql/src/pl_reserved_kwlist.h

@@ -20,10 +20,9 @@
 /*
  * List of (keyword-name, keyword-token-value) pairs.
  *
- * Be careful not to put the same word in both lists.
+ * Be careful not to put the same word into pl_unreserved_kwlist.h.
  *
- * !!WARNING!!: This list must be sorted by ASCII name, because binary
- *		 search is used to locate entries.
+ * Note: gen_keywordlist.pl requires the entries to appear in ASCII order.
  */
 
 /* name, value */

src/pl/plpgsql/src/pl_unreserved_kwlist.h

@@ -20,11 +20,10 @@
 /*
  * List of (keyword-name, keyword-token-value) pairs.
  *
- * Be careful not to put the same word in both lists.  Also be sure that
- * pl_gram.y's unreserved_keyword production agrees with this list.
+ * Be careful not to put the same word into pl_reserved_kwlist.h.  Also be
+ * sure that pl_gram.y's unreserved_keyword production agrees with this list.
  *
- * !!WARNING!!: This list must be sorted by ASCII name, because binary
- *		 search is used to locate entries.
+ * Note: gen_keywordlist.pl requires the entries to appear in ASCII order.
  */
 
 /* name, value */

src/tools/gen_keywordlist.pl

@@ -14,6 +14,12 @@
 # variable named according to the -v switch ("ScanKeywords" by default).
 # The variable is marked "static" unless the -e switch is given.
 #
+# ScanKeywordList uses hash-based lookup, so this script also selects
+# a minimal perfect hash function for the keyword set, and emits a
+# static hash function that is referenced in the ScanKeywordList struct.
+# The hash function is case-insensitive unless --no-case-fold is specified.
+# Note that case folding works correctly only for all-ASCII keywords!
+#
 #
 # Portions Copyright (c) 1996-2019, PostgreSQL Global Development Group
 # Portions Copyright (c) 1994, Regents of the University of California
@@ -25,15 +31,18 @@
 use strict;
 use warnings;
 use Getopt::Long;
+use PerfectHash;
 
 my $output_path = '';
 my $extern = 0;
+my $case_fold = 1;
 my $varname = 'ScanKeywords';
 
 GetOptions(
-	'output:s' => \$output_path,
-	'extern'   => \$extern,
-	'varname:s' => \$varname) || usage();
+	'output:s'   => \$output_path,
+	'extern'     => \$extern,
+	'case-fold!' => \$case_fold,
+	'varname:s'  => \$varname) || usage();
 
 my $kw_input_file = shift @ARGV || die "No input file.\n";
 
@@ -87,7 +96,22 @@ while (<$kif>)
 	}
 }
 
+# When being case-insensitive, insist that the input be all-lower-case.
+if ($case_fold)
+{
+	foreach my $kw (@keywords)
+	{
+		die qq|The keyword "$kw" is not lower-case in $kw_input_file\n|
+		  if ($kw ne lc $kw);
+	}
+}
+
 # Error out if the keyword names are not in ASCII order.
+#
+# While this isn't really necessary with hash-based lookup, it's still
+# helpful because it provides a cheap way to reject duplicate keywords.
+# Also, insisting on sorted order ensures that code that scans the keyword
+# table linearly will see the keywords in a canonical order.
 for my $i (0..$#keywords - 1)
 {
 	die qq|The keyword "$keywords[$i + 1]" is out of order in $kw_input_file\n|
@@ -128,15 +152,25 @@ print $kwdef "};\n\n";
 
 printf $kwdef "#define %s_NUM_KEYWORDS %d\n\n", uc $varname, scalar @keywords;
 
+# Emit the definition of the hash function.
+
+my $funcname = $varname . "_hash_func";
+
+my $f = PerfectHash::generate_hash_function(\@keywords, $funcname,
+	case_fold => $case_fold);
+
+printf $kwdef qq|static %s\n|, $f;
+
 # Emit the struct that wraps all this lookup info into one variable.
 
-print $kwdef "static " if !$extern;
+printf $kwdef "static " if !$extern;
 printf $kwdef "const ScanKeywordList %s = {\n", $varname;
 printf $kwdef qq|\t%s_kw_string,\n|, $varname;
 printf $kwdef qq|\t%s_kw_offsets,\n|, $varname;
+printf $kwdef qq|\t%s,\n|, $funcname;
 printf $kwdef qq|\t%s_NUM_KEYWORDS,\n|, uc $varname;
 printf $kwdef qq|\t%d\n|, $max_len;
-print $kwdef "};\n\n";
+printf $kwdef "};\n\n";
 
 printf $kwdef "#endif\t\t\t\t\t\t\t/* %s_H */\n", uc $base_filename;
 
@@ -144,10 +178,11 @@ printf $kwdef "#endif\t\t\t\t\t\t\t/* %s_H */\n", uc $base_filename;
 sub usage
 {
 	die <<EOM;
-Usage: gen_keywordlist.pl [--output/-o <path>] [--varname/-v <varname>] [--extern/-e] input_file
-    --output   Output directory (default '.')
-    --varname  Name for ScanKeywordList variable (default 'ScanKeywords')
-    --extern   Allow the ScanKeywordList variable to be globally visible
+Usage: gen_keywordlist.pl [--output/-o <path>] [--varname/-v <varname>] [--extern/-e] [--[no-]case-fold] input_file
+    --output        Output directory (default '.')
+    --varname       Name for ScanKeywordList variable (default 'ScanKeywords')
+    --extern        Allow the ScanKeywordList variable to be globally visible
+    --no-case-fold  Keyword matching is to be case-sensitive
 
 gen_keywordlist.pl transforms a list of keywords into a ScanKeywordList.
 The output filename is derived from the input file by inserting _d,

src/tools/msvc/Solution.pm

@@ -414,7 +414,7 @@ sub GenerateFiles
 			'src/include/parser/kwlist.h'))
 	{
 		print "Generating kwlist_d.h...\n";
-		system('perl src/tools/gen_keywordlist.pl --extern -o src/common src/include/parser/kwlist.h');
+		system('perl -I src/tools src/tools/gen_keywordlist.pl --extern -o src/common src/include/parser/kwlist.h');
 	}
 
 	if (IsNewer(
@@ -426,8 +426,8 @@ sub GenerateFiles
 	{
 		print "Generating pl_reserved_kwlist_d.h and pl_unreserved_kwlist_d.h...\n";
 		chdir('src/pl/plpgsql/src');
-		system('perl ../../../tools/gen_keywordlist.pl --varname ReservedPLKeywords pl_reserved_kwlist.h');
-		system('perl ../../../tools/gen_keywordlist.pl --varname UnreservedPLKeywords pl_unreserved_kwlist.h');
+		system('perl -I ../../../tools ../../../tools/gen_keywordlist.pl --varname ReservedPLKeywords pl_reserved_kwlist.h');
+		system('perl -I ../../../tools ../../../tools/gen_keywordlist.pl --varname UnreservedPLKeywords pl_unreserved_kwlist.h');
 		chdir('../../../..');
 	}
 
@@ -440,8 +440,8 @@ sub GenerateFiles
 	{
 		print "Generating c_kwlist_d.h and ecpg_kwlist_d.h...\n";
 		chdir('src/interfaces/ecpg/preproc');
-		system('perl ../../../tools/gen_keywordlist.pl --varname ScanCKeywords c_kwlist.h');
-		system('perl ../../../tools/gen_keywordlist.pl --varname ScanECPGKeywords ecpg_kwlist.h');
+		system('perl -I ../../../tools ../../../tools/gen_keywordlist.pl --varname ScanCKeywords --no-case-fold c_kwlist.h');
+		system('perl -I ../../../tools ../../../tools/gen_keywordlist.pl --varname ScanECPGKeywords ecpg_kwlist.h');
 		chdir('../../../..');
 	}