From c084f884bd0f29b84999c05dcdd497ff42cd7000 Mon Sep 17 00:00:00 2001
From: Nick Downing <downing.nick@gmail.com>
Date: Thu, 15 Oct 2020 13:25:52 -0700
Subject: [PATCH] Change inner loops to use int not YY_CHAR, removing need for
 separate NUL table

---
 src/cpp-flex.skl | 11 ++++++++++-
 src/dfa.c        | 29 +++++------------------------
 2 files changed, 15 insertions(+), 25 deletions(-)

diff --git a/src/cpp-flex.skl b/src/cpp-flex.skl
index a1eb04c91..1fa3f0d18 100644
--- a/src/cpp-flex.skl
+++ b/src/cpp-flex.skl
@@ -1778,8 +1778,17 @@ m4_ifdef([[M4_MODE_NO_FULLSPD]], [[
 ]])
 ]])
 
+%# yy_c was formerly YY_CHAR, changed to int because table can now
+%# have up to 0x101 entries, since we no longer generate a separate
+%# NUL table.
+%#
+%# Note: on x86-64 architecture with gcc -O2, we save an instruction
+%# in the main loop, since the character can now be zero-extended in
+%# the process of retrieving it from the input stream or the yy_ec[]
+%# or yy_meta[] arrays, whereas previously it was zero-extended by a
+%# register-to-register move just prior to the yy_chk[] table lookup
 m4_define([[M4_GEN_NEXT_COMPRESSED_STATE]], [[
-	YY_CHAR yy_c = $1;
+	int yy_c = $1;
 	/* Save the backing-up info \before/ computing the next state
 	 * because we always compute one more state than needed - we
 	 * always proceed until we reach a jam state
diff --git a/src/dfa.c b/src/dfa.c
index 06bbd6320..baa42ed12 100644
--- a/src/dfa.c
+++ b/src/dfa.c
@@ -455,31 +455,12 @@ size_t ntod (void)
 
 	/* Note that the test for ecgroup[0] == numecs below accomplishes
 	 * both (1) and (2) above
+	 *
+	 * New way: we will only use NUL table for fulltbl, because the
+	 * scanner will use an integer instead of YY_CHAR as noted above
 	 */
-	if (!ctrl.fullspd && ecgroup[0] == numecs) {
-		/* NUL is alone in its equivalence class, which is the
-		 * last one.
-		 */
-		int     use_NUL_table = (numecs == ctrl.csize);
-
-		if (ctrl.fulltbl && !use_NUL_table) {
-			/* We still may want to use the table if numecs
-			 * is a power of 2.
-			 */
-			if (numecs <= ctrl.csize && is_power_of_2(numecs)) {
-				use_NUL_table = true;
-			}
-		}
-
-		if (use_NUL_table)
-			nultrans =
-				allocate_integer_array (current_max_dfas);
-
-		/* From now on, nultrans != nil indicates that we're
-		 * saving null transitions for later, separate encoding.
-		 */
-	}
-
+	if (ctrl.fulltbl && ecgroup[0] == numecs && is_power_of_2(numecs))
+		nultrans = allocate_integer_array (current_max_dfas);
 
 	if (ctrl.fullspd) {
 		for (i = 0; i <= numecs; ++i)