crunch.c: Speed up by 20%, if endianness can be determined at compile time.

2 files changed, 95 insertions, 2 deletions

diff --git a/src/bytorder.h b/src/bytorder.h
new file mode 100644
index 0000000..cb8231e
--- /dev/null
+++ b/src/bytorder.h
@@ -0,0 +1,52 @@
+#if defined(ALF_LSB_FIRST) && defined(ALF_MSB_FIRST)
+#  error Cannot define both ALF_LSB_FIRST and ALF_MSB_FIRST
+#endif
+
+#if defined(ALF_LSB_FIRST) || defined(ALF_MSB_FIRST)
+#  define ALF_ENDIAN_OK
+#endif
+
+/* try to get the byte order on this platform.
+   if we succeed, either ALF_LSB_FIRST or ALF_MSB_FIRST will
+   defined, and so will ALF_ENDIAN_OK. */
+
+#ifndef ALF_ENDIAN_OK
+#  if defined(__BYTE_ORDER__)
+#    if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
+#      define ALF_LSB_FIRST
+#      define ALF_ENDIAN_OK
+#    else
+#      if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
+#        define ALF_MSB_FIRST
+#        define ALF_ENDIAN_OK
+#      endif
+#    endif
+#  endif
+#endif
+
+#ifndef ALF_ENDIAN_OK
+#  if defined(__LITTLE_ENDIAN__)
+#    define ALF_LSB_FIRST
+#    define ALF_ENDIAN_OK
+#  else
+#    if defined(__BIG_ENDIAN__)
+#      define ALF_MSB_FIRST
+#      define ALF_ENDIAN_OK
+#    endif
+#  endif
+#endif
+
+#ifdef ALF_ENDIAN_OK
+#  ifdef ALF_LSB_FIRST
+#    define HIBYTE 2
+#    define MIDBYTE 1
+#    define LOBYTE 0
+#  else
+#    define HIBYTE 1
+#    define MIDBYTE 2
+#    define LOBYTE 3
+#  endif
+#else
+#  warning Cannot determine endianness, falling back to safe but slow append_bit().
+#  warning Recompile with -DALF_LSB_FIRST or -DALF_MSB_FIRST.
+#endif
diff --git a/src/crunch.c b/src/crunch.c
index cb1ce5f..82cc095 100644
--- a/src/crunch.c
+++ b/src/crunch.c
@@ -9,6 +9,7 @@
 #include "u816.h"
 #include "crunch.h"
 #include "self.h"
+#include "bytorder.h"
 
 #define INITIAL_BITS 9
 #define MAX_BITS 12
@@ -122,14 +123,52 @@ void init_table(void) {
 	curr_token = INIT_TOKEN;
 }
 
-void inc_output_len(void) {
-	if(++output_len == MAX_INPUT_SIZE) {
+void check_output_len(void) {
+	if(output_len >= MAX_INPUT_SIZE) {
 		fprintf(stderr, "%s: fatal: compressed file would be >16MB.\n", self);
 		exit(1);
 	}
+}
+
+void inc_output_len(void) {
+	output_len++;
+	check_output_len();
 	output_buf[output_len] = 0;
 }
 
+#ifdef ALF_ENDIAN_OK
+
+/* This is 20% faster, but it requires knowing the endianness of
+   the platform at compile time. See bytorder.h for gory details. */
+union { unsigned int ui; u8 bytes[4]; } ui2bytes;
+
+void store_token(int tok) {
+	if(opt_verbose > 1) {
+		printf("<%d >%d:%d #%d", in_pos, output_len, out_bitpos, tok);
+		if(tok == TOK_RESET)
+			fputs(" RESET", stdout);
+		else if(tok == TOK_END)
+			fputs(" END", stdout);
+		else if(tok < 256)
+			printf(" %s", fmt_chr(tok));
+		putchar('\n');
+	}
+
+	tok <<= (MAX_BITS - token_bits);
+	ui2bytes.ui = tok << (12 - out_bitpos);
+	output_buf[output_len] |= ui2bytes.bytes[HIBYTE];
+	output_buf[output_len + 1] = ui2bytes.bytes[MIDBYTE];
+	output_buf[output_len + 2] = ui2bytes.bytes[LOBYTE];
+	out_bitpos += token_bits;
+	output_len += out_bitpos / 8;
+	out_bitpos %= 8;
+
+	check_output_len();
+}
+
+#else
+
+/* Slower but portable. */
 void append_bit(int bit) {
 	output_buf[output_len] |= (bit << (7 - out_bitpos));
 	out_bitpos++;
@@ -157,6 +196,7 @@ void store_token(int tok) {
 		append_bit(tok & mask ? 1 : 0);
 	}
 }
+#endif
 
 short match_token(void) {
 	short t, nt;
@@ -211,6 +251,7 @@ void crunch(void) {
 	init_table();
 	out_bitpos = 0;
 	in_pos = 0;
+	output_buf[output_len] = 0; /* just in case */
 
 	/* 0-byte input files don't get a TOK_RESET */
 	if(!input_len) {