Optimisation for w=8,16,32

Signed-off-by: Mattias Andrée <maandree@kth.se>

9 files changed, 434 insertions, 69 deletions

diff --git a/TODO b/TODO
index ec8eda0..a498582 100644
--- a/TODO
+++ b/TODO
@@ -13,7 +13,3 @@ Add libkeccak_cshakesum_fd (TODO in libkeccak/cshake.h)
 Add KMAC and KMACXOF
 Add TupleHash and TupleHashXOF
 Add ParallelHash and ParallelHashXOF
-
-Optimise for wordsize 32
-Optimise for wordsize 16
-Optimise for wordsize 8
diff --git a/common.h b/common.h
index 96df9e7..44a0b5d 100644
--- a/common.h
+++ b/common.h
@@ -38,6 +38,33 @@
 #define COMMA ,
 
 /**
+ * X-macro-enabled listing of all intergers in [0, 0]
+ * 
+ * @param  X(int)  The macro to expand 4 times
+ * @param  D       Code to insert between each expansion of `X`
+ */
+#define LIST_1(X, D)\
+	X(0)
+
+/**
+ * X-macro-enabled listing of all intergers in [0, 1]
+ * 
+ * @param  X(int)  The macro to expand 2 times
+ * @param  D       Code to insert between each expansion of `X`
+ */
+#define LIST_2(X, D)\
+	X(0) D X(1)
+
+/**
+ * X-macro-enabled listing of all intergers in [0, 3]
+ * 
+ * @param  X(int)  The macro to expand 4 times
+ * @param  D       Code to insert between each expansion of `X`
+ */
+#define LIST_4(X, D)\
+	X(0) D X(1) D X(2) D X(3)
+
+/**
  * X-macro-enabled listing of all intergers in [0, 4]
  * 
  * @param  X(int)  The macro to expand 5 times
diff --git a/digest.c b/digest.c
index 7ee744a..384d9df 100644
--- a/digest.c
+++ b/digest.c
@@ -12,26 +12,63 @@ static const long int LANE_TRANSPOSE_MAP[] = { LIST_25(X, COMMA) };
 
 
 /**
- * Keccak-f round constants
- */
-static const uint_fast64_t RC[] = {
-	0x0000000000000001ULL, 0x0000000000008082ULL, 0x800000000000808AULL, 0x8000000080008000ULL,
-	0x000000000000808BULL, 0x0000000080000001ULL, 0x8000000080008081ULL, 0x8000000000008009ULL,
-	0x000000000000008AULL, 0x0000000000000088ULL, 0x0000000080008009ULL, 0x000000008000000AULL,
-	0x000000008000808BULL, 0x800000000000008BULL, 0x8000000000008089ULL, 0x8000000000008003ULL,
-	0x8000000000008002ULL, 0x8000000000000080ULL, 0x000000000000800AULL, 0x800000008000000AULL,
-	0x8000000080008081ULL, 0x8000000000008080ULL, 0x0000000080000001ULL, 0x8000000080008008ULL
+ * 64-bit Keccak-f round constants
+ */
+static const uint64_t rc64[] = {
+	UINT64_C(0x0000000000000001), UINT64_C(0x0000000000008082), UINT64_C(0x800000000000808A), UINT64_C(0x8000000080008000),
+	UINT64_C(0x000000000000808B), UINT64_C(0x0000000080000001), UINT64_C(0x8000000080008081), UINT64_C(0x8000000000008009),
+	UINT64_C(0x000000000000008A), UINT64_C(0x0000000000000088), UINT64_C(0x0000000080008009), UINT64_C(0x000000008000000A),
+	UINT64_C(0x000000008000808B), UINT64_C(0x800000000000008B), UINT64_C(0x8000000000008089), UINT64_C(0x8000000000008003),
+	UINT64_C(0x8000000000008002), UINT64_C(0x8000000000000080), UINT64_C(0x000000000000800A), UINT64_C(0x800000008000000A),
+	UINT64_C(0x8000000080008081), UINT64_C(0x8000000000008080), UINT64_C(0x0000000080000001), UINT64_C(0x8000000080008008)
+};
+
+
+/**
+ * 32-bit Keccak-f round constants
+ */
+static const uint32_t rc32[] = {
+	UINT32_C(0x00000001), UINT32_C(0x00008082), UINT32_C(0x0000808A), UINT32_C(0x80008000),
+	UINT32_C(0x0000808B), UINT32_C(0x80000001), UINT32_C(0x80008081), UINT32_C(0x00008009),
+	UINT32_C(0x0000008A), UINT32_C(0x00000088), UINT32_C(0x80008009), UINT32_C(0x8000000A),
+	UINT32_C(0x8000808B), UINT32_C(0x0000008B), UINT32_C(0x00008089), UINT32_C(0x00008003),
+	UINT32_C(0x00008002), UINT32_C(0x00000080), UINT32_C(0x0000800A), UINT32_C(0x8000000A),
+	UINT32_C(0x80008081), UINT32_C(0x00008080)
+};
+
+
+/**
+ * 16-bit Keccak-f round constants
+ */
+static const uint16_t rc16[] = {
+	UINT16_C(0x0001), UINT16_C(0x8082), UINT16_C(0x808A), UINT16_C(0x8000),
+	UINT16_C(0x808B), UINT16_C(0x0001), UINT16_C(0x8081), UINT16_C(0x8009),
+	UINT16_C(0x008A), UINT16_C(0x0088), UINT16_C(0x8009), UINT16_C(0x000A),
+	UINT16_C(0x808B), UINT16_C(0x008B), UINT16_C(0x8089), UINT16_C(0x8003),
+	UINT16_C(0x8002), UINT16_C(0x0080), UINT16_C(0x800A), UINT16_C(0x000A)
+};
+
+
+/**
+ * 8-bit Keccak-f round constants
+ */
+static const uint8_t rc8[] = {
+	UINT8_C(0x01), UINT8_C(0x82), UINT8_C(0x8A), UINT8_C(0x00),
+	UINT8_C(0x8B), UINT8_C(0x01), UINT8_C(0x81), UINT8_C(0x09),
+	UINT8_C(0x8A), UINT8_C(0x88), UINT8_C(0x09), UINT8_C(0x0A),
+	UINT8_C(0x8B), UINT8_C(0x8B), UINT8_C(0x89), UINT8_C(0x03),
+	UINT8_C(0x02), UINT8_C(0x80)
 };
 
 
 /**
  * Rotate a word
  * 
- * @param   x:int_fast64_t     The value to rotate
- * @param   n:long             Rotation steps, may be zero mod `w`
- * @param   w:long             `state->w`
- * @param   wmod:int_fast64_t  `state->wmod`
- * @return  :int_fast64_t      The value rotated
+ * @param   x:uint_fast64_t     The value to rotate
+ * @param   n:long int          Rotation steps, may be zero mod `w`
+ * @param   w:long int          `state->w`
+ * @param   wmod:uint_fast64_t  `state->wmod`
+ * @return  :uint_fast64_t      The value rotated
  */
 #define rotate(x, n, w, wmod) ((((x) >> ((w) - ((n) % (w)))) | ((x) << ((n) % (w)))) & (wmod))
 
@@ -39,11 +76,41 @@ static const uint_fast64_t RC[] = {
 /**
  * Rotate a 64-bit word
  * 
- * @param   x:int_fast64_t  The value to rotate
- * @param   n:long          Rotation steps, may not be zero
- * @return   :int_fast64_t  The value rotated
+ * @param   x:uint64_t  The value to rotate
+ * @param   n:long int  Rotation steps, may not be zero
+ * @return   :uint64_t  The value rotated
+ */
+#define rotate64(x, n) ((uint64_t)(((uint64_t)(x) >> (64L - (n))) | ((uint64_t)(x) << (n))))
+
+
+/**
+ * Rotate a 32-bit word
+ * 
+ * @param   x:uint32_t  The value to rotate
+ * @param   n:long int  Rotation steps, may not be zero
+ * @return   :uint32_t  The value rotated
+ */
+#define rotate32(x, n) ((uint32_t)(((uint32_t)(x) >> (32L - (n))) | ((uint32_t)(x) << (n))))
+
+
+/**
+ * Rotate a 16-bit word
+ * 
+ * @param   x:uint16_t  The value to rotate
+ * @param   n:long int  Rotation steps, may not be zero
+ * @return   :uint16_t  The value rotated
  */
-#define rotate64(x, n) ((int_fast64_t)(((uint64_t)(x) >> (64L - (n))) | ((uint64_t)(x) << (n))))
+#define rotate16(x, n) ((uint16_t)(((uint16_t)(x) >> (16L - (n))) | ((uint16_t)(x) << (n))))
+
+
+/**
+ * Rotate a 8-bit word
+ * 
+ * @param   x:uint8_t   The value to rotate
+ * @param   n:long int  Rotation steps, may not be zero
+ * @return   :uint8_t   The value rotated
+ */
+#define rotate8(x, n) ((uint8_t)(((uint8_t)(x) >> (8L - (n))) | ((uint8_t)(x) << (n))))
 
 
 /**
@@ -54,13 +121,13 @@ static const uint_fast64_t RC[] = {
  */
 LIBKECCAK_GCC_ONLY(__attribute__((__nonnull__, __nothrow__, __hot__)))
 static void
-libkeccak_f_round(register struct libkeccak_state *state, register int_fast64_t rc)
+libkeccak_f_round(register struct libkeccak_state *state, register uint64_t rc)
 {
-	int_fast64_t *restrict A = state->S;
-	int_fast64_t B[25];
-	int_fast64_t C[5];
-	int_fast64_t da, db, dc, dd, de;
-	int_fast64_t wmod = state->wmod;
+	uint64_t *restrict A = state->S.w64;
+	uint_fast64_t B[25];
+	uint_fast64_t C[5];
+	uint_fast64_t da, db, dc, dd, de;
+	uint_fast64_t wmod = state->wmod;
 	long int w = state->w;
 
 	/* θ step (step 1 of 3). */
@@ -102,12 +169,10 @@ libkeccak_f_round(register struct libkeccak_state *state, register int_fast64_t
  */
 LIBKECCAK_GCC_ONLY(__attribute__((__nonnull__, __nothrow__, __hot__)))
 static void
-libkeccak_f_round64(register struct libkeccak_state *state, register int_fast64_t rc)
+libkeccak_f_round64(register struct libkeccak_state *state, register uint64_t rc)
 {
-	int_fast64_t *restrict A = state->S;
-	int_fast64_t B[25];
-	int_fast64_t C[5];
-	int_fast64_t da, db, dc, dd, de;
+	uint64_t *restrict A = state->S.w64;
+	uint64_t B[25], C[5], da, db, dc, dd, de;
 
 	/* θ step (step 1 of 3). */
 #define X(N) C[N] = A[N * 5] ^ A[N * 5 + 1] ^ A[N * 5 + 2] ^ A[N * 5 + 3] ^ A[N * 5 + 4]
@@ -141,6 +206,138 @@ libkeccak_f_round64(register struct libkeccak_state *state, register int_fast64_
 
 
 /**
+ * 32-bit word version of `libkeccak_f_round`
+ * 
+ * @param  state  The hashing state
+ * @param  rc     The round contant for this round
+ */
+LIBKECCAK_GCC_ONLY(__attribute__((__nonnull__, __nothrow__, __hot__)))
+static void
+libkeccak_f_round32(register struct libkeccak_state *state, register uint32_t rc)
+{
+	uint32_t *restrict A = state->S.w32;
+	uint32_t B[25], C[5], da, db, dc, dd, de;
+
+	/* θ step (step 1 of 3). */
+#define X(N) C[N] = A[N * 5] ^ A[N * 5 + 1] ^ A[N * 5 + 2] ^ A[N * 5 + 3] ^ A[N * 5 + 4]
+	LIST_5(X, ;);
+#undef X
+
+	/* θ step (step 2 of 3). */
+	da = C[4] ^ rotate32(C[1], 1);
+	dd = C[2] ^ rotate32(C[4], 1);
+	db = C[0] ^ rotate32(C[2], 1);
+	de = C[3] ^ rotate32(C[0], 1);
+	dc = C[1] ^ rotate32(C[3], 1);
+
+	/* ρ and π steps, with last two part of θ. */
+#define X(bi, ai, dv, r) B[bi] = rotate32(A[ai] ^ dv, (r & 31))
+	B[0] = A[0] ^ da;   X( 1, 15, dd, 28);  X( 2,  5, db,  1);  X( 3, 20, de, 27);  X( 4, 10, dc, 62);
+	X( 5,  6, db, 44);  X( 6, 21, de, 20);  X( 7, 11, dc,  6);  X( 8,  1, da, 36);  X( 9, 16, dd, 55);
+	X(10, 12, dc, 43);  X(11,  2, da,  3);  X(12, 17, dd, 25);  X(13,  7, db, 10);  X(14, 22, de, 39);
+	X(15, 18, dd, 21);  X(16,  8, db, 45);  X(17, 23, de,  8);  X(18, 13, dc, 15);  X(19,  3, da, 41);
+	X(20, 24, de, 14);  X(21, 14, dc, 61);  X(22,  4, da, 18);  X(23, 19, dd, 56);  X(24,  9, db,  2);
+#undef X
+
+	/* ξ step. */
+#define X(N) A[N] = (uint32_t)(B[N] ^ ((~(B[(N + 5) % 25])) & B[(N + 10) % 25]))
+	LIST_25(X, ;);
+#undef X
+
+	/* ι step. */
+	A[0] ^= rc;
+}
+
+
+/**
+ * 16-bit word version of `libkeccak_f_round`
+ * 
+ * @param  state  The hashing state
+ * @param  rc     The round contant for this round
+ */
+LIBKECCAK_GCC_ONLY(__attribute__((__nonnull__, __nothrow__, __hot__)))
+static void
+libkeccak_f_round16(register struct libkeccak_state *state, register uint16_t rc)
+{
+	uint16_t *restrict A = state->S.w16;
+	uint16_t B[25], C[5], da, db, dc, dd, de;
+
+	/* θ step (step 1 of 3). */
+#define X(N) C[N] = A[N * 5] ^ A[N * 5 + 1] ^ A[N * 5 + 2] ^ A[N * 5 + 3] ^ A[N * 5 + 4]
+	LIST_5(X, ;);
+#undef X
+
+	/* θ step (step 2 of 3). */
+	da = C[4] ^ rotate16(C[1], 1);
+	dd = C[2] ^ rotate16(C[4], 1);
+	db = C[0] ^ rotate16(C[2], 1);
+	de = C[3] ^ rotate16(C[0], 1);
+	dc = C[1] ^ rotate16(C[3], 1);
+
+	/* ρ and π steps, with last two part of θ. */
+#define X(bi, ai, dv, r) B[bi] = rotate16(A[ai] ^ dv, (r & 15))
+	B[0] = A[0] ^ da;   X( 1, 15, dd, 28);  X( 2,  5, db,  1);  X( 3, 20, de, 27);  X( 4, 10, dc, 62);
+	X( 5,  6, db, 44);  X( 6, 21, de, 20);  X( 7, 11, dc,  6);  X( 8,  1, da, 36);  X( 9, 16, dd, 55);
+	X(10, 12, dc, 43);  X(11,  2, da,  3);  X(12, 17, dd, 25);  X(13,  7, db, 10);  X(14, 22, de, 39);
+	X(15, 18, dd, 21);  X(16,  8, db, 45);  X(17, 23, de,  8);  X(18, 13, dc, 15);  X(19,  3, da, 41);
+	X(20, 24, de, 14);  X(21, 14, dc, 61);  X(22,  4, da, 18);  X(23, 19, dd, 56);  X(24,  9, db,  2);
+#undef X
+
+	/* ξ step. */
+#define X(N) A[N] = (uint16_t)(B[N] ^ ((~(B[(N + 5) % 25])) & B[(N + 10) % 25]))
+	LIST_25(X, ;);
+#undef X
+
+	/* ι step. */
+	A[0] ^= rc;
+}
+
+
+/**
+ * 8-bit word version of `libkeccak_f_round`
+ * 
+ * @param  state  The hashing state
+ * @param  rc     The round contant for this round
+ */
+LIBKECCAK_GCC_ONLY(__attribute__((__nonnull__, __nothrow__, __hot__)))
+static void
+libkeccak_f_round8(register struct libkeccak_state *state, register uint8_t rc)
+{
+	uint8_t *restrict A = state->S.w8;
+	uint8_t B[25], C[5], da, db, dc, dd, de;
+
+	/* θ step (step 1 of 3). */
+#define X(N) C[N] = A[N * 5] ^ A[N * 5 + 1] ^ A[N * 5 + 2] ^ A[N * 5 + 3] ^ A[N * 5 + 4]
+	LIST_5(X, ;);
+#undef X
+
+	/* θ step (step 2 of 3). */
+	da = C[4] ^ rotate8(C[1], 1);
+	dd = C[2] ^ rotate8(C[4], 1);
+	db = C[0] ^ rotate8(C[2], 1);
+	de = C[3] ^ rotate8(C[0], 1);
+	dc = C[1] ^ rotate8(C[3], 1);
+
+	/* ρ and π steps, with last two part of θ. */
+#define X(bi, ai, dv, r) B[bi] = rotate8(A[ai] ^ dv, (r & 7))
+	B[0] = A[0] ^ da;   X( 1, 15, dd, 28);  X( 2,  5, db,  1);  X( 3, 20, de, 27);  X( 4, 10, dc, 62);
+	X( 5,  6, db, 44);  X( 6, 21, de, 20);  X( 7, 11, dc,  6);  X( 8,  1, da, 36);  X( 9, 16, dd, 55);
+	X(10, 12, dc, 43);  X(11,  2, da,  3);  X(12, 17, dd, 25);  X(13,  7, db, 10);  X(14, 22, de, 39);
+	X(15, 18, dd, 21);  X(16,  8, db, 45);  X(17, 23, de,  8);  X(18, 13, dc, 15);  X(19,  3, da, 41);
+	X(20, 24, de, 14);  X(21, 14, dc, 61);  X(22,  4, da, 18);  X(23, 19, dd, 56);  X(24,  9, db,  2);
+#undef X
+
+	/* ξ step. */
+#define X(N) A[N] = (uint8_t)(B[N] ^ ((~(B[(N + 5) % 25])) & B[(N + 10) % 25]))
+	LIST_25(X, ;);
+#undef X
+
+	/* ι step. */
+	A[0] ^= rc;
+}
+
+
+/**
  * Convert a chunk of bytes to a lane
  * 
  * @param  state  The hashing state
@@ -151,14 +348,34 @@ libkeccak_f(register struct libkeccak_state *state)
 {
 	register long int i = 0;
 	register long int nr = state->nr;
-	register long int wmod = state->wmod;
+	register uint_fast64_t wmod = state->wmod;
+
 	if (nr == 24) {
 		for (; i < nr; i++)
-			libkeccak_f_round64(state, (int_fast64_t)(RC[i]));
-	} else {
+			libkeccak_f_round64(state, rc64[i]);
+		return;
+	}
+
+	if (nr == 22) {
 		for (; i < nr; i++)
-			libkeccak_f_round(state, (int_fast64_t)(RC[i] & (uint_fast64_t)wmod));
+			libkeccak_f_round32(state, rc32[i]);
+		return;
 	}
+
+	if (nr == 20) {
+		for (; i < nr; i++)
+			libkeccak_f_round16(state, rc16[i]);
+		return;
+	}
+
+	if (nr == 18) {
+		for (; i < nr; i++)
+			libkeccak_f_round8(state, rc8[i]);
+		return;
+	}
+
+	for (; i < nr; i++)
+		libkeccak_f_round(state, rc64[i] & wmod);
 }
 
 
@@ -173,18 +390,18 @@ libkeccak_f(register struct libkeccak_state *state)
  * @return           The lane
  */
 LIBKECCAK_GCC_ONLY(__attribute__((__nonnull__, __nothrow__, __pure__, __warn_unused_result__, __gnu_inline__)))
-static inline int_fast64_t
+static inline uint64_t
 libkeccak_to_lane(register const unsigned char *restrict message, register size_t msglen,
                   register long int rr, register long int ww, size_t off)
 {
 	register long int n = (long)((msglen < (size_t)rr ? msglen : (size_t)rr) - off);
-	int_fast64_t rc = 0;
+	uint_fast64_t rc = 0;
 	message += off;
 	while (ww--) {
 		rc <<= 8;
-		rc |= __builtin_expect(ww < n, 1) ? (int_fast64_t)(unsigned char)(message[ww]) : 0L;
+		rc |= __builtin_expect(ww < n, 1) ? (uint_fast64_t)(unsigned char)message[ww] : 0L;
 	}
-	return rc;
+	return (uint64_t)rc;
 }
 
 
@@ -198,13 +415,13 @@ libkeccak_to_lane(register const unsigned char *restrict message, register size_
  * @return           The lane
  */
 LIBKECCAK_GCC_ONLY(__attribute__((__nonnull__, __nothrow__, __pure__, __hot__, __warn_unused_result__, __gnu_inline__)))
-static inline int_fast64_t
+static inline uint64_t
 libkeccak_to_lane64(register const unsigned char *message, register size_t msglen, register long int rr, size_t off)
 {
 	register long int n = (long)((msglen < (size_t)rr ? msglen : (size_t)rr) - off);
-	int_fast64_t rc = 0;
+	uint64_t rc = 0;
 	message += off;
-#define X(N) if (__builtin_expect(N < n, 1)) rc |= (int_fast64_t)(unsigned char)(message[N]) << (N * 8);\
+#define X(N) if (__builtin_expect(N < n, 1)) rc |= (uint64_t)message[N] << (N * 8);\
              else return rc
 	LIST_8(X, ;);
 #undef X
@@ -213,6 +430,78 @@ libkeccak_to_lane64(register const unsigned char *message, register size_t msgle
 
 
 /**
+ * 32-bit lane version of `libkeccak_to_lane`
+ * 
+ * @param   message  The message
+ * @param   msglen   The length of the message
+ * @param   rr       Bitrate in bytes
+ * @param   off      The offset in the message
+ * @return           The lane
+ */
+LIBKECCAK_GCC_ONLY(__attribute__((__nonnull__, __nothrow__, __pure__, __hot__, __warn_unused_result__, __gnu_inline__)))
+static inline uint32_t
+libkeccak_to_lane32(register const unsigned char *message, register size_t msglen, register long int rr, size_t off)
+{
+	register long int n = (long)((msglen < (size_t)rr ? msglen : (size_t)rr) - off);
+	uint32_t rc = 0;
+	message += off;
+#define X(N) if (__builtin_expect(N < n, 1)) rc |= (uint32_t)message[N] << (N * 8);\
+             else return rc
+	LIST_4(X, ;);
+#undef X
+	return rc;
+}
+
+
+/**
+ * 16-bit lane version of `libkeccak_to_lane`
+ * 
+ * @param   message  The message
+ * @param   msglen   The length of the message
+ * @param   rr       Bitrate in bytes
+ * @param   off      The offset in the message
+ * @return           The lane
+ */
+LIBKECCAK_GCC_ONLY(__attribute__((__nonnull__, __nothrow__, __pure__, __hot__, __warn_unused_result__, __gnu_inline__)))
+static inline uint16_t
+libkeccak_to_lane16(register const unsigned char *message, register size_t msglen, register long int rr, size_t off)
+{
+	register long int n = (long)((msglen < (size_t)rr ? msglen : (size_t)rr) - off);
+	uint16_t rc = 0;
+	message += off;
+#define X(N) if (__builtin_expect(N < n, 1)) rc |= (uint16_t)message[N] << (N * 8);\
+             else return rc
+	LIST_2(X, ;);
+#undef X
+	return rc;
+}
+
+
+/**
+ * 8-bit lane version of `libkeccak_to_lane`
+ * 
+ * @param   message  The message
+ * @param   msglen   The length of the message
+ * @param   rr       Bitrate in bytes
+ * @param   off      The offset in the message
+ * @return           The lane
+ */
+LIBKECCAK_GCC_ONLY(__attribute__((__nonnull__, __nothrow__, __pure__, __hot__, __warn_unused_result__, __gnu_inline__)))
+static inline uint8_t
+libkeccak_to_lane8(register const unsigned char *message, register size_t msglen, register long int rr, size_t off)
+{
+	register long int n = (long)((msglen < (size_t)rr ? msglen : (size_t)rr) - off);
+	uint8_t rc = 0;
+	message += off;
+#define X(N) if (__builtin_expect(N < n, 1)) rc |= (uint8_t)(unsigned char)message[N] << (N * 8);\
+             else return (uint8_t)rc
+	LIST_1(X, ;);
+#undef X
+	return (uint8_t)rc;
+}
+
+
+/**
  * Right-pad message with a 10*1-pad
  * 
  * @param   r       Should be `state->r` where `state` is the hashing state
@@ -263,24 +552,62 @@ libkeccak_absorption_phase(register struct libkeccak_state *restrict state,
 	register long int rr = state->r >> 3;
 	register long int ww = state->w >> 3;
 	register long int n = (long)len / rr;
-	if (__builtin_expect(ww >= 8, 1)) { /* ww > 8 is impossible, it is just for optimisation possibilities. */
+
+	if (__builtin_expect(ww == 8, 1)) {
 		while (n--) {
-#define X(N) state->S[N] ^= libkeccak_to_lane64(message, len, rr, (size_t)(LANE_TRANSPOSE_MAP[N] * 8))
+#define X(N) state->S.w64[N] ^= libkeccak_to_lane64(message, len, rr, (size_t)(LANE_TRANSPOSE_MAP[N] * 8))
 			LIST_25(X, ;);
 #undef X
 			libkeccak_f(state);
 			message += (size_t)rr;
 			len -= (size_t)rr;
 		}
-	} else {
+		return;
+	}
+
+	if (__builtin_expect(ww == 4, 1)) {
+		while (n--) {
+#define X(N) state->S.w32[N] ^= libkeccak_to_lane32(message, len, rr, (size_t)(LANE_TRANSPOSE_MAP[N] * 4))
+			LIST_25(X, ;);
+#undef X
+			libkeccak_f(state);
+			message += (size_t)rr;
+			len -= (size_t)rr;
+		}
+		return;
+	}
+
+	if (__builtin_expect(ww == 2, 1)) {
+		while (n--) {
+#define X(N) state->S.w16[N] ^= libkeccak_to_lane16(message, len, rr, (size_t)(LANE_TRANSPOSE_MAP[N] * 2))
+			LIST_25(X, ;);
+#undef X
+			libkeccak_f(state);
+			message += (size_t)rr;
+			len -= (size_t)rr;
+		}
+		return;
+	}
+
+	if (__builtin_expect(ww == 1, 1)) {
 		while (n--) {
-#define X(N) state->S[N] ^= libkeccak_to_lane(message, len, rr, ww, (size_t)(LANE_TRANSPOSE_MAP[N] * ww))
+#define X(N) state->S.w8[N] ^= libkeccak_to_lane8(message, len, rr, (size_t)(LANE_TRANSPOSE_MAP[N] * 1))
 			LIST_25(X, ;);
 #undef X
 			libkeccak_f(state);
 			message += (size_t)rr;
 			len -= (size_t)rr;
 		}
+		return;
+	}
+
+	while (n--) {
+#define X(N) state->S.w64[N] ^= libkeccak_to_lane(message, len, rr, ww, (size_t)(LANE_TRANSPOSE_MAP[N] * ww))
+		LIST_25(X, ;);
+#undef X
+		libkeccak_f(state);
+		message += (size_t)rr;
+		len -= (size_t)rr;
 	}
 }
 
@@ -299,16 +626,20 @@ static void
 libkeccak_squeezing_phase(register struct libkeccak_state *restrict state, long int rr,
                           long int nn, long int ww, register unsigned char *restrict hashsum)
 {
-	register int_fast64_t v;
+	register uint64_t v;
 	register long int ni = rr / ww + !!(rr % ww);
 	auto long int olen = state->n;
 	auto long int i, j = 0;
 	register long int k;
 	while (olen > 0) {
 		for (i = 0; i < ni && j < nn; i++) {
-			v = state->S[LANE_TRANSPOSE_MAP[i]];
+			if (__builtin_expect(ww == 8, 1)) v = state->S.w64[LANE_TRANSPOSE_MAP[i]]; else
+			if (__builtin_expect(ww == 4, 1)) v = state->S.w32[LANE_TRANSPOSE_MAP[i]]; else
+			if (__builtin_expect(ww == 2, 1)) v = state->S.w16[LANE_TRANSPOSE_MAP[i]]; else
+			if (__builtin_expect(ww == 1, 1)) v = state->S.w8[LANE_TRANSPOSE_MAP[i]]; else
+			v = state->S.w64[LANE_TRANSPOSE_MAP[i]];
 			for (k = 0; k++ < ww && j++ < nn; v >>= 8)
-				*hashsum++ = (unsigned char)v;
+				*hashsum++ = (unsigned char)(v & 0xFFU);
 		}
 		olen -= state->r;
 		if (olen > 0)
@@ -657,13 +988,7 @@ libkeccak_simple_squeeze(register struct libkeccak_state *state, register long i
  * @param  state  The hashing state
  * @param  times  The number of digests
  */
-void
-libkeccak_fast_squeeze(register struct libkeccak_state *state, register long int times)
-{
-	times *= (state->n - 1) / state->r + 1;
-	while (times--)
-		libkeccak_f(state);
-}
+extern inline void libkeccak_fast_squeeze(register struct libkeccak_state *state, register long int times);
 
 
 /**
diff --git a/extra/libkeccak_state_marshal.c b/extra/libkeccak_state_marshal.c
index 7541164..f4fd4b5 100644
--- a/extra/libkeccak_state_marshal.c
+++ b/extra/libkeccak_state_marshal.c
@@ -28,7 +28,7 @@ libkeccak_state_marshal(const struct libkeccak_state *restrict state, void *rest
 
 	if (!data) {
 		return 7 * sizeof(long int) +
-		       1 * sizeof(int64_t) +
+		       1 * sizeof(uint64_t) +
 		       sizeof(state->S) +
 		       2 * sizeof(size_t) +
 		       state->mptr;
@@ -42,7 +42,7 @@ libkeccak_state_marshal(const struct libkeccak_state *restrict state, void *rest
 	set(wmod);
 	set(l);
 	set(nr);
-	__builtin_memcpy(data, state->S, sizeof(state->S));
+	__builtin_memcpy(data, &state->S, sizeof(state->S));
 	data += sizeof(state->S);
 	set(mptr);
 	set(mlen);
diff --git a/extra/libkeccak_state_unmarshal.c b/extra/libkeccak_state_unmarshal.c
index 4714566..0938ba1 100644
--- a/extra/libkeccak_state_unmarshal.c
+++ b/extra/libkeccak_state_unmarshal.c
@@ -30,7 +30,7 @@ libkeccak_state_unmarshal(struct libkeccak_state *restrict state, const void *re
 
 	if (!state) {
 		data += 7U * sizeof(long int);
-		data += 1U * sizeof(int64_t);
+		data += 1U * sizeof(uint64_t);
 		data += sizeof(state->S);
 		mptr = *(const size_t *)data;
 		data += 2U * sizeof(size_t);
@@ -46,7 +46,7 @@ libkeccak_state_unmarshal(struct libkeccak_state *restrict state, const void *re
 	get(wmod);
 	get(l);
 	get(nr);
-	memcpy(state->S, data, sizeof(state->S));
+	memcpy(&state->S, data, sizeof(state->S));
 	data += sizeof(state->S);
 	get(mptr);
 	get(mlen);
diff --git a/extra/libkeccak_state_wipe_sponge.c b/extra/libkeccak_state_wipe_sponge.c
index 91e845a..69aa46a 100644
--- a/extra/libkeccak_state_wipe_sponge.c
+++ b/extra/libkeccak_state_wipe_sponge.c
@@ -10,7 +10,7 @@
 void
 libkeccak_state_wipe_sponge(volatile struct libkeccak_state *state)
 {
-	volatile int64_t *restrict S = state->S;
+	volatile uint64_t *restrict S = state->S.w64;
 	size_t i;
 
 	for (i = 0; i < 25; i++)
diff --git a/libkeccak.h b/libkeccak.h
index 670c69e..814883b 100644
--- a/libkeccak.h
+++ b/libkeccak.h
@@ -14,6 +14,10 @@
 # pragma clang diagnostic ignored "-Wdocumentation"
 # pragma clang diagnostic ignored "-Wunknown-attributes"
 #endif
+#if defined(__GNUC__)
+# pragma GCC diagnostic push
+# pragma GCC diagnostic ignored "-Winline"
+#endif
 
 
 /**
@@ -59,7 +63,12 @@ struct libkeccak_state {
 	/**
 	 * The lanes (state/sponge)
 	 */
-	int64_t S[25];
+	union {
+		uint64_t w64[25];
+		uint32_t w32[25];
+		uint16_t w16[25];
+		uint8_t w8[25];
+	} S;
 
 	/**
 	 * The bitrate
@@ -89,7 +98,7 @@ struct libkeccak_state {
 	/**
 	 * The word mask
 	 */
-	int64_t wmod;
+	uint64_t wmod;
 
 	/**
 	 * ℓ, the binary logarithm of the word size
@@ -255,7 +264,12 @@ void libkeccak_simple_squeeze(register struct libkeccak_state *, register long i
  * @param  times  The number of digests
  */
 LIBKECCAK_GCC_ONLY(__attribute__((__nonnull__, __nothrow__)))
-void libkeccak_fast_squeeze(register struct libkeccak_state *, register long int);
+inline void
+libkeccak_fast_squeeze(register struct libkeccak_state *state, register long int times)
+{
+	times *= (state->n - 1) / state->r + 1;
+	libkeccak_simple_squeeze(state, times);
+}
 
 /**
  * Squeeze out another digest
@@ -311,6 +325,9 @@ libkeccak_state_destroy(volatile struct libkeccak_state *state)
 
 
 
+#if defined(__GNUC__)
+# pragma GCC diagnostic pop
+#endif
 #if defined(__clang__)
 # pragma clang diagnostic pop
 #endif
diff --git a/libkeccak/extra.h b/libkeccak/extra.h
index 60a2601..b9734b3 100644
--- a/libkeccak/extra.h
+++ b/libkeccak/extra.h
@@ -11,7 +11,7 @@ inline void
 libkeccak_state_reset(struct libkeccak_state *state)
 {
 	state->mptr = 0;
-	memset(state->S, 0, sizeof(state->S));
+	memset(&state->S, 0, sizeof(state->S));
 }
 
 
diff --git a/libkeccak_state_initialise.c b/libkeccak_state_initialise.c
index 7644ff7..545d3df 100644
--- a/libkeccak_state_initialise.c
+++ b/libkeccak_state_initialise.c
@@ -34,9 +34,9 @@ libkeccak_state_initialise(struct libkeccak_state *restrict state, const struct
 	}
 
 	state->nr = 12 + (state->l << 1);
-	state->wmod = (state->w == 64) ? ~0LL : (int64_t)((1ULL << state->w) - 1);
+	state->wmod = (state->w == 64) ? ~0ULL : (uint64_t)((1ULL << state->w) - 1);
 	for (x = 0; x < 25; x++)
-		state->S[x] = 0;
+		state->S.w64[x] = 0;
 	state->mptr = 0;
 	state->mlen = (size_t)(state->r * state->b) >> 2;