Use uint_leastN_t instead of uintN_t

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

4 files changed, 70 insertions, 49 deletions

diff --git a/common.h b/common.h
index c053fac..50e62ab 100644
--- a/common.h
+++ b/common.h
@@ -16,6 +16,23 @@
 
 
 /**
+ * Truncate an unsigned integer to an unsigned 32-bit integer
+ * 
+ * @param   X:uint_least32_t  The value to truncate
+ * @return  :uint_least32_t   The 32 lowest bits in `X`
+ */
+#define TRUNC32(X) ((X) & (uint_least32_t)0xFFFFFFFFUL)
+
+/**
+ * Truncate an unsigned integer to an unsigned 64-bit integer
+ * 
+ * @param   X:uint_least64_t  The value to truncate
+ * @return  :uint_least64_t   The 64 lowest bits in `X`
+ */
+#define TRUNC64(X) ((X) & (uint_least64_t)0xFFFFFFFFFFFFFFFFULL)
+
+
+/**
  * Process a chunk using SHA-2
  * 
  * @param  state  The hashing state
diff --git a/init.c b/init.c
index 3059ce9..50ec102 100644
--- a/init.c
+++ b/init.c
@@ -5,7 +5,7 @@
 /**
  * Round constants, SHA-256, should keep the 32 most significant bits of 64 first constants
  */
-static const uint64_t ROUND_CONSTANTS[] = {
+static const uint_least64_t ROUND_CONSTANTS[] = {
 	0x428A2F98D728AE22ULL, 0x7137449123EF65CDULL, 0xB5C0FBCFEC4D3B2FULL, 0xE9B5DBA58189DBBCULL,
 	0x3956C25BF348B538ULL, 0x59F111F1B605D019ULL, 0x923F82A4AF194F9BULL, 0xAB1C5ED5DA6D8118ULL,
 	0xD807AA98A3030242ULL, 0x12835B0145706FBEULL, 0x243185BE4EE4B28CULL, 0x550C7DC3D5FFB4E2ULL,
@@ -32,7 +32,7 @@ static const uint64_t ROUND_CONSTANTS[] = {
 /**
  * Initial state for SHA224
  */
-static const uint32_t H_224[] = {
+static const uint_least32_t H_224[] = {
 	0xC1059ED8UL, 0x367CD507UL, 0x3070DD17UL, 0xF70E5939UL,
 	0xFFC00B31UL, 0x68581511UL, 0x64F98FA7UL, 0xBEFA4FA4UL
 };
@@ -40,7 +40,7 @@ static const uint32_t H_224[] = {
 /**
  * Initial state for SHA256
  */
-static const uint32_t H_256[] = {
+static const uint_least32_t H_256[] = {
 	0x6A09E667UL, 0xBB67AE85UL, 0x3C6EF372UL, 0xA54FF53AUL,
 	0x510E527FUL, 0x9B05688CUL, 0x1F83D9ABUL, 0x5BE0CD19UL
 };
@@ -48,7 +48,7 @@ static const uint32_t H_256[] = {
 /**
  * Initial state for SHA384
  */
-static const uint64_t H_384[] = {
+static const uint_least64_t H_384[] = {
 	0xCBBB9D5DC1059ED8ULL, 0x629A292A367CD507ULL, 0x9159015A3070DD17ULL, 0x152FECD8F70E5939ULL,
 	0x67332667FFC00B31ULL, 0x8EB44A8768581511ULL, 0xDB0C2E0D64F98FA7ULL, 0x47B5481DBEFA4FA4ULL
 };
@@ -56,7 +56,7 @@ static const uint64_t H_384[] = {
 /**
  * Initial state for SHA512
  */
-static const uint64_t H_512[] = {
+static const uint_least64_t H_512[] = {
 	0x6A09E667F3BCC908ULL, 0xBB67AE8584CAA73BULL, 0x3C6EF372FE94F82BULL, 0xA54FF53A5F1D36F1ULL,
 	0x510E527FADE682D1ULL, 0x9B05688C2B3E6C1FULL, 0x1F83D9ABFB41BD6BULL, 0x5BE0CD19137E2179ULL
 };
@@ -64,7 +64,7 @@ static const uint64_t H_512[] = {
 /**
  * Initial state for SHA512/224
  */
-static const uint64_t H_512_224[] = {
+static const uint_least64_t H_512_224[] = {
 	0x8C3D37C819544DA2ULL, 0x73E1996689DCD4D6ULL, 0x1DFAB7AE32FF9C82ULL, 0x679DD514582F9FCFULL,
 	0x0F6D2B697BD44DA8ULL, 0x77E36F7304C48942ULL, 0x3F9D85A86A1D36C8ULL, 0x1112E6AD91D692A1ULL
 };
@@ -72,7 +72,7 @@ static const uint64_t H_512_224[] = {
 /**
  * Initial state for SHA512/256
  */
-static const uint64_t H_512_256[] = {
+static const uint_least64_t H_512_256[] = {
 	0x22312194FC2BF72CULL, 0x9F555FA3C84C64C2ULL, 0x2393B86B6F53B151ULL, 0x963877195940EABDULL,
 	0x96283EE2A88EFFE3ULL, 0xBE5E1E2553863992ULL, 0x2B0199FC2C85B8AAULL, 0x0EB72DDC81C52CA2ULL
 };
@@ -103,7 +103,7 @@ libsha2_init(struct libsha2_state *restrict state, enum libsha2_algorithm algori
 	/* Set round constants, and chunk size. */
 	if (algorithm <= LIBSHA2_256) {
 		for (i = 0; i < 64; i++)
-			state->k.b32[i] = (uint32_t)(ROUND_CONSTANTS[i] >> 32);
+			state->k.b32[i] = (uint_least32_t)(ROUND_CONSTANTS[i] >> 32);
 		state->chunk_size = 64;
 	} else {
 		memcpy(state->k.b64, ROUND_CONSTANTS, sizeof(ROUND_CONSTANTS));
diff --git a/libsha2.h b/libsha2.h
index 3cd19b9..87c8c20 100644
--- a/libsha2.h
+++ b/libsha2.h
@@ -66,12 +66,12 @@ struct libsha2_state {
 		/**
 		 * For 32-bit algorithms
 		 */
-		uint32_t b32[64];
+		uint_least32_t b32[64];
     
 		/**
 		 * For 64-bit algorithms
 		 */
-		uint64_t b64[80];
+		uint_least64_t b64[80];
 	} k;
 
 	/**
@@ -83,12 +83,12 @@ struct libsha2_state {
 		/**
 		 * For 32-bit algorithms
 		 */
-		uint32_t b32[64];
+		uint_least32_t b32[64];
 
 		/**
 		 * For 64-bit algorithms
 		 */
-		uint64_t b64[80];
+		uint_least64_t b64[80];
 	} w;
 
 	/**
@@ -98,12 +98,12 @@ struct libsha2_state {
 		/**
 		 * For 32-bit algorithms
 		 */
-		uint32_t b32[8];
+		uint_least32_t b32[8];
 
 		/**
 		 * For 64-bit algorithms
 		 */
-		uint64_t b64[8];
+		uint_least64_t b64[8];
 	} h;
 
 	/**
@@ -115,12 +115,12 @@ struct libsha2_state {
 		/**
 		 * For 32-bit algorithms
 		 */
-		uint32_t b32[8];
+		uint_least32_t b32[8];
 
 		/**
 		 * For 64-bit algorithms
 		 */
-		uint64_t b64[8];
+		uint_least64_t b64[8];
 	} work_h;
 
 	/**
@@ -153,6 +153,7 @@ struct libsha2_state {
  * `explicit_bzero` (or `memset`) when you are done.
  */
 struct libsha2_hmac_state {
+
 	/**
 	 * State of the underlaying hash function
 	 */
diff --git a/process.c b/process.c
index 7e3d807..a309628 100644
--- a/process.c
+++ b/process.c
@@ -6,37 +6,40 @@
  * Unified implementation (what can unified without performance impact)
  * of the chunk processing for all SHA-2 functions
  * 
- * @param  chunk   The data to process
- * @param  A       Wordsize-dependent constant, take a look at the code
- * @param  B       Wordsize-dependent constant, take a look at the code
- * @param  C       Wordsize-dependent constant, take a look at the code
- * @param  D       Wordsize-dependent constant, take a look at the code
- * @param  E       Wordsize-dependent constant, take a look at the code
- * @param  F       Wordsize-dependent constant, take a look at the code
- * @param  G       Wordsize-dependent constant, take a look at the code
- * @param  H       Wordsize-dependent constant, take a look at the code
- * @param  I       Wordsize-dependent constant, take a look at the code
- * @param  J       Wordsize-dependent constant, take a look at the code
- * @param  K       Wordsize-dependent constant, take a look at the code
- * @param  L       Wordsize-dependent constant, take a look at the code
- * @param  WORD_T  `__typeof()` on any wordsize-dependent variable, with exact size
- * @param  k       Round constants
- * @param  w       Words
- * @param  h       Hash values
- * @param  work_h  Space for temporary hash values
+ * @param  chunk      The data to process
+ * @param  A          Wordsize-dependent constant, take a look at the code
+ * @param  B          Wordsize-dependent constant, take a look at the code
+ * @param  C          Wordsize-dependent constant, take a look at the code
+ * @param  D          Wordsize-dependent constant, take a look at the code
+ * @param  E          Wordsize-dependent constant, take a look at the code
+ * @param  F          Wordsize-dependent constant, take a look at the code
+ * @param  G          Wordsize-dependent constant, take a look at the code
+ * @param  H          Wordsize-dependent constant, take a look at the code
+ * @param  I          Wordsize-dependent constant, take a look at the code
+ * @param  J          Wordsize-dependent constant, take a look at the code
+ * @param  K          Wordsize-dependent constant, take a look at the code
+ * @param  L          Wordsize-dependent constant, take a look at the code
+ * @param  WORD_T     `__typeof()` on any wordsize-dependent variable
+ * @param  WORD_SIZE  4 for 32-bit algorithms and 8 for 64-bit algorithms
+ * @param  TRUNC      `TRUNC32` for 32-bit algorithms and `TRUNC64` for 64-bit algorithms
+ * @param  k          Round constants
+ * @param  w          Words
+ * @param  h          Hash values
+ * @param  work_h     Space for temporary hash values
  */
-#define SHA2_IMPLEMENTATION(chunk, A, B, C, D, E, F, G, H, I, J, K, L, WORD_T, k, w, h, work_h)\
+#define SHA2_IMPLEMENTATION(chunk, A, B, C, D, E, F, G, H, I, J, K, L, WORD_T, WORD_SIZE, TRUNC, k, w, h, work_h) \
 	memcpy(work_h, h, sizeof(work_h));\
 	\
 	memset(w, 0, 16 * sizeof(*(w)));\
 	for (i = 0; i < 16; i++)\
-		for (j = 0; j < sizeof(WORD_T); j++)\
-			w[i] |= ((WORD_T)(chunk[(i + 1) * sizeof(WORD_T) - j - 1])) << (j << 3);\
+		for (j = 0; j < WORD_SIZE; j++)\
+			w[i] |= ((WORD_T)(chunk[(i + 1) * WORD_SIZE - j - 1])) << (j << 3);\
 	\
 	for (i = 16; i < sizeof(k) / sizeof(*(k)); i++)	{\
 		w[i] = w[i - 16] + w[i - 7];\
 		w[i] += ROTR(w[i - 15], A) ^ ROTR(w[i - 15], B) ^ (w[i - 15] >> (C));\
 		w[i] += ROTR(w[i - 2], D) ^ ROTR(w[i - 2], E) ^ (w[i - 2] >> (F));\
+		w[i] = TRUNC(w[i]);\
 	}\
 	\
 	for (i = 0; i < sizeof(k) / sizeof(*(k)); i++) {\
@@ -46,30 +49,30 @@
 		s1 += ROTR(work_h[4], G) ^ ROTR(work_h[4], H) ^ ROTR(work_h[4], I);\
 		s0 += ROTR(work_h[0], J) ^ ROTR(work_h[0], K) ^ ROTR(work_h[0], L);\
 		\
-		memmove(work_h + 1, work_h, 7 * sizeof(*(work_h)));\
-		work_h[4] += s1;\
-		work_h[0] = s1 + s0;\
+		memmove(&work_h[1], work_h, 7 * sizeof(*(work_h)));\
+		work_h[4] = TRUNC(work_h[4] + s1);\
+		work_h[0] = TRUNC(s1 + s0);\
 	}\
 	\
 	for (i = 0; i < 8; i++)\
-		h[i] += work_h[i]
+		h[i] = TRUNC(h[i] + work_h[i]);
 
 
 void
 libsha2_process(struct libsha2_state *restrict state, const unsigned char *restrict chunk)
 {
 	if (state->algorithm <= LIBSHA2_256) {
-		uint32_t s0, s1;
+		uint_least32_t s0, s1;
 		size_t i, j;
 
 #if defined(__GNUC__)
 # pragma GCC diagnostic push
 # pragma GCC diagnostic ignored "-Wmemset-elt-size"
 #endif
-#define ROTR(X, N) (((X) >> (N)) | ((X) << ((sizeof(uint32_t) * 8) - (N))))
+#define ROTR(X, N) TRUNC32(((X) >> (N)) | ((X) << (32 - (N))))
 
-		SHA2_IMPLEMENTATION(chunk, 7, 18, 3, 17, 19, 10, 6, 11, 25, 2, 13, 22, uint32_t,
-		                    state->k.b32, state->w.b32, state->h.b32, state->work_h.b32);
+		SHA2_IMPLEMENTATION(chunk, 7, 18, 3, 17, 19, 10, 6, 11, 25, 2, 13, 22, uint_least32_t,
+		                    4, TRUNC32, state->k.b32, state->w.b32, state->h.b32, state->work_h.b32);
 
 #undef ROTR
 #if defined(__GNUC__)
@@ -77,13 +80,13 @@ libsha2_process(struct libsha2_state *restrict state, const unsigned char *restr
 #endif
 
 	} else {
-		uint64_t s0, s1;
+		uint_least64_t s0, s1;
 		size_t i, j;
 
-#define ROTR(X, N) (((X) >> (N)) | ((X) << ((sizeof(uint64_t) * 8) - (N))))
+#define ROTR(X, N) TRUNC64(((X) >> (N)) | ((X) << (64 - (N))))
 
-		SHA2_IMPLEMENTATION(chunk, 1, 8, 7, 19, 61, 6, 14, 18, 41, 28, 34, 39, uint64_t,
-		                    state->k.b64, state->w.b64, state->h.b64, state->work_h.b64);
+		SHA2_IMPLEMENTATION(chunk, 1, 8, 7, 19, 61, 6, 14, 18, 41, 28, 34, 39, uint_least64_t,
+		                    8, TRUNC64, state->k.b64, state->w.b64, state->h.b64, state->work_h.b64);
 
 #undef ROTR
 	}