Implement ffs() / fls() and use the latter to compute the MSB.

configure.ac

@@ -21,19 +21,6 @@ AC_C_CONST
 AC_C_RESTRICT
 AC_C_VOLATILE
 AC_C_BIGENDIAN
-AC_TYPE_INT16_T
-AC_TYPE_INT32_T
-AC_TYPE_INT8_T
-AC_TYPE_INTMAX_T
-AC_TYPE_INTPTR_T
-AC_TYPE_OFF_T
-AC_TYPE_SIZE_T
-AC_TYPE_SSIZE_T
-AC_TYPE_UINT16_T
-AC_TYPE_UINT32_T
-AC_TYPE_UINT8_T
-AC_TYPE_UINTMAX_T
-AC_TYPE_UINTPTR_T
 
 # libtool
 LT_PREREQ([2.2.6])
@@ -46,15 +33,46 @@ PKG_INSTALLDIR
 # other programs
 AC_PROG_INSTALL
 
+############################################################################
+#
+# Types
+#
+AC_TYPE_INT16_T
+AC_TYPE_INT32_T
+AC_TYPE_INT8_T
+AC_TYPE_INTMAX_T
+AC_TYPE_INTPTR_T
+AC_TYPE_OFF_T
+AC_TYPE_SIZE_T
+AC_TYPE_SSIZE_T
+AC_TYPE_UINT16_T
+AC_TYPE_UINT32_T
+AC_TYPE_UINT8_T
+AC_TYPE_UINTMAX_T
+AC_TYPE_UINTPTR_T
+
 ############################################################################
 #
 # Headers and functions
 #
 
-AC_CHECK_HEADERS([endian.h sys/endian.h])
+AC_CHECK_HEADERS([endian.h sys/endian.h strings.h])
 AX_GCC_BUILTIN([__builtin_bswap16])
 AX_GCC_BUILTIN([__builtin_bswap32])
 AX_GCC_BUILTIN([__builtin_bswap64])
+AX_GCC_BUILTIN([__builtin_clz])
+AX_GCC_BUILTIN([__builtin_clzl])
+AX_GCC_BUILTIN([__builtin_clzll])
+AX_GCC_BUILTIN([__builtin_ctz])
+AX_GCC_BUILTIN([__builtin_ctzl])
+AX_GCC_BUILTIN([__builtin_ctzll])
+AX_GCC_BUILTIN([__builtin_ffs])
+AX_GCC_BUILTIN([__builtin_ffsl])
+AX_GCC_BUILTIN([__builtin_ffsll])
+# No compiler we know of has these
+#AX_GCC_BUILTIN([__builtin_fls])
+#AX_GCC_BUILTIN([__builtin_flsl])
+#AX_GCC_BUILTIN([__builtin_flsll])
 AC_CHECK_DECLS([
     bswap16, bswap32, bswap64,
     bswap16v, bswap32v, bswap64v,
@@ -76,6 +94,11 @@ AC_CHECK_DECLS([
 #include <endian.h>
 #endif
 ]])
+AC_CHECK_FUNCS([ffs ffsl ffsll fls flsl flsll], [], [], [[
+#if HAVE_STRINGS_H
+#include <strings.h>
+#endif
+]])
 AC_CHECK_FUNCS([strlcat strlcmp strlcpy])
 AC_CHECK_FUNCS([wcslcat wcslcmp wcslcpy])
 

include/cryb/bitwise.h

@@ -54,6 +54,109 @@ CRYB_ROL_ROR(64);
 
 #undef CRYB_ROL_ROR
 
+#if !HAVE_FFS
+#define ffs	cryb_ffs
+#endif
+#if !HAVE_FFSL
+#define ffsl	cryb_ffsl
+#endif
+#if !HAVE_FFSLL
+#define ffsll	cryb_ffsll
+#endif
+#if !HAVE_FLS
+#define fls	cryb_fls
+#endif
+#if !HAVE_FLSL
+#define flsl	cryb_flsl
+#endif
+#if !HAVE_FLSLL
+#define flsll	cryb_flsll
+#endif
+
+static inline int cryb_ffs(int n) {
+#if HAVE___BUILTIN_FFS
+	return (__builtin_ffs(n));
+#elif HAVE___BUILTIN_CTZ
+	return (n ? __builtin_ctz(n) : 0);
+#else
+	int i = 8 * sizeof n - 1;
+	for (i > 0)
+		if (n & (1 << --i))
+			break;
+	return (i);
+#endif
+}
+
+static inline int cryb_ffsl(long int n) {
+#if HAVE___BUILTIN_FFSL
+	return (__builtin_ffsl(n));
+#elif HAVE___BUILTIN_CLZ
+	return (n ? __builtin_ctz(n) : 0);
+#else
+	int i = 8 * sizeof n - 1;
+	for (i > 0)
+		if (n & (1 << --i))
+			break;
+	return (i);
+#endif
+}
+
+static inline int cryb_ffsll(long long int n) {
+#if HAVE___BUILTIN_FFSLL
+	return (__builtin_ffsll(n));
+#elif HAVE___BUILTIN_CLZ
+	return (n ? __builtin_ctz(n) : 0);
+#else
+	int i = 8 * sizeof n - 1;
+	for (i > 0)
+		if (n & (1 << --i))
+			break;
+	return (i);
+#endif
+}
+
+static inline int cryb_fls(int n) {
+#if HAVE___BUILTIN_FLS
+	return (__builtin_fls(n));
+#elif HAVE___BUILTIN_CLZ
+	return (n ? (8 * sizeof n) - __builtin_clz(n) : 0);
+#else
+	int i = 8 * sizeof n - 1;
+	for (i > 0)
+		if (n & (1 << --i))
+			break;
+	return (i);
+#endif
+}
+
+static inline int cryb_flsl(long int n) {
+#if HAVE___BUILTIN_FLSL
+	return (__builtin_flsl(n));
+#elif HAVE___BUILTIN_CLZ
+	return (n ? (8 * sizeof n) - __builtin_clzl(n) : 0);
+#else
+	int i = 8 * sizeof n - 1;
+	for (i > 0)
+		if (n & (1 << --i))
+			break;
+	return (i);
+#endif
+}
+
+static inline int cryb_flsll(long long int n) {
+#if HAVE___BUILTIN_FLSLL
+	return (__builtin_flsll(n));
+#elif HAVE___BUILTIN_CLZ
+	return (n ? (8 * sizeof n) - __builtin_clzll(n) : 0);
+#else
+	int i = 8 * sizeof n - 1;
+	for (i > 0)
+		if (n & (1 << --i))
+			break;
+	return (i);
+#endif
+}
+
 CRYB_END
 
 #endif

lib/mpi/cryb_mpi_add_abs.c

@@ -31,7 +31,9 @@
 
 #include <stddef.h>
 #include <stdint.h>
+#include <strings.h>
 
+#include <cryb/bitwise.h>
 #include <cryb/mpi.h>
 
 #include "cryb_mpi_impl.h"
@@ -104,13 +106,7 @@ mpi_add_abs(cryb_mpi *X, const cryb_mpi *A, const cryb_mpi *B)
 	}
 	if (X->words[i] == 0)
 		--i;
-	/* compute msb of msw */
+	X->msb = i * 32 + flsl(X->words[i]);
-	/* XXX should use flsl() */
-	for (X->msb = 31; X->msb > 0; --X->msb)
-		if (X->words[i] & (1 << X->msb))
-			break;
-	/* add msw offset */
-	X->msb += i * 32 + 1;
 	X->neg = 0;
 	return (0);
 }

lib/mpi/cryb_mpi_load.c

@@ -31,7 +31,9 @@
 
 #include <stddef.h>
 #include <stdint.h>
+#include <strings.h>
 
+#include <cryb/bitwise.h>
 #include <cryb/endian.h>
 #include <cryb/mpi.h>
 
@@ -70,13 +72,6 @@ mpi_load(cryb_mpi *X, const uint8_t *a, size_t len)
 		--i;
 	CRYB_NO_DEFAULT_CASE;
 	}
-	/* i now points to the msw */
+	X->msb = i * 32 + flsl(X->words[i]);
-	/* compute msb of msw */
-	/* XXX use flsl() */
-	for (X->msb = 31; X->msb > 0; --X->msb)
-		if (X->words[i] & (1 << X->msb))
-			break;
-	/* add msw offset */
-	X->msb += i * 32 + 1;
 	return (0);
 }

lib/mpi/cryb_mpi_sub_abs.c

@@ -31,7 +31,9 @@
 
 #include <stddef.h>
 #include <stdint.h>
+#include <strings.h>
 
+#include <cryb/bitwise.h>
 #include <cryb/mpi.h>
 
 #include "cryb_mpi_impl.h"
@@ -92,15 +94,9 @@ mpi_sub_abs(cryb_mpi *X, const cryb_mpi *A, const cryb_mpi *B)
 		c = cn;
 		++i;
 	}
-	while (i > 0 && X->words[i] == 0)
+	while (X->words[i] == 0)
 		--i;
-	/* compute msb of msw */
+	X->msb = i * 32 + flsl(X->words[i]);
-	/* XXX use flsl() */
-	for (X->msb = 31; X->msb > 0; --X->msb)
-		if (X->words[i] & (1 << X->msb))
-			break;
-	/* add msw offset */
-	X->msb += i * 32 + 1;
 	X->neg = 0;
 	return (0);
 }