X-Git-Url: https://bearssl.org/gitweb//home/git/?p=BearSSL;a=blobdiff_plain;f=src%2Finner.h;h=1b5267380ff3d55e975fed85d5db698c048cd785;hp=bcf61b04d9bc0dbd8feff5dff392d2e3313177b4;hb=c6ffcd29381f555783a75d957c28c0ff861cd33a;hpb=5db2d48b128326704cd33aff56c8b46e463f3bf6

diff --git a/src/inner.h b/src/inner.h
index bcf61b0..1b52673 100644
--- a/src/inner.h
+++ b/src/inner.h
@@ -48,10 +48,22 @@
  * already set their root keys to RSA-4096, so we should be able to
  * process such keys.
  *
- * This value MUST be a multiple of 64.
+ * This value MUST be a multiple of 64. This value MUST NOT exceed 47666
+ * (some computations in RSA key generation rely on the factor size being
+ * no more than 23833 bits). RSA key sizes beyond 3072 bits don't make a
+ * lot of sense anyway.
  */
 #define BR_MAX_RSA_SIZE   4096
 
+/*
+ * Minimum size for a RSA modulus (in bits); this value is used only to
+ * filter out invalid parameters for key pair generation. Normally,
+ * applications should not use RSA keys smaller than 2048 bits; but some
+ * specific cases might need shorter keys, for legacy or research
+ * purposes.
+ */
+#define BR_MIN_RSA_SIZE   512
+
 /*
  * Maximum size for a RSA factor (in bits). This is for RSA private-key
  * operations. Default is to support factors up to a bit more than half
@@ -85,7 +97,7 @@
  *
  * The test on 'unsigned long' should already catch most cases, the one
  * notable exception being Windows code where 'unsigned long' is kept to
- * 32-bit for compatbility with all the legacy code that liberally uses
+ * 32-bit for compatibility with all the legacy code that liberally uses
  * the 'DWORD' type for 32-bit values.
  *
  * Macro names are taken from: http://nadeausoftware.com/articles/2012/02/c_c_tip_how_detect_processor_type_using_compiler_predefined_macros
@@ -109,60 +121,212 @@
  * Set BR_LOMUL on platforms where it makes sense.
  */
 #ifndef BR_LOMUL
-#if BR_ARMEL_CORTEX_GCC
+#if BR_ARMEL_CORTEXM_GCC
 #define BR_LOMUL   1
 #endif
 #endif
 
 /*
- * Determine whether x86 AES instructions are understood by the compiler.
+ * Architecture detection.
  */
-#ifndef BR_AES_X86NI
+#ifndef BR_i386
+#if __i386__ || _M_IX86
+#define BR_i386   1
+#endif
+#endif
+
+#ifndef BR_amd64
+#if __x86_64__ || _M_X64
+#define BR_amd64   1
+#endif
+#endif
+
+/*
+ * Compiler brand and version.
+ *
+ * Implementations that use intrinsics need to detect the compiler type
+ * and version because some specific actions may be needed to activate
+ * the corresponding opcodes, both for header inclusion, and when using
+ * them in a function.
+ *
+ * BR_GCC, BR_CLANG and BR_MSC will be set to 1 for, respectively, GCC,
+ * Clang and MS Visual C. For each of them, sub-macros will be defined
+ * for versions; each sub-macro is set whenever the compiler version is
+ * at least as recent as the one corresponding to the macro.
+ */
+
+/*
+ * GCC thresholds are on versions 4.4 to 4.9 and 5.0.
+ */
+#ifndef BR_GCC
+#if __GNUC__ && !__clang__
+#define BR_GCC   1
+
+#if __GNUC__ > 4
+#define BR_GCC_5_0   1
+#elif __GNUC__ == 4 && __GNUC_MINOR__ >= 9
+#define BR_GCC_4_9   1
+#elif __GNUC__ == 4 && __GNUC_MINOR__ >= 8
+#define BR_GCC_4_8   1
+#elif __GNUC__ == 4 && __GNUC_MINOR__ >= 7
+#define BR_GCC_4_7   1
+#elif __GNUC__ == 4 && __GNUC_MINOR__ >= 6
+#define BR_GCC_4_6   1
+#elif __GNUC__ == 4 && __GNUC_MINOR__ >= 5
+#define BR_GCC_4_5   1
+#elif __GNUC__ == 4 && __GNUC_MINOR__ >= 4
+#define BR_GCC_4_4   1
+#endif
+
+#if BR_GCC_5_0
+#define BR_GCC_4_9   1
+#endif
+#if BR_GCC_4_9
+#define BR_GCC_4_8   1
+#endif
+#if BR_GCC_4_8
+#define BR_GCC_4_7   1
+#endif
+#if BR_GCC_4_7
+#define BR_GCC_4_6   1
+#endif
+#if BR_GCC_4_6
+#define BR_GCC_4_5   1
+#endif
+#if BR_GCC_4_5
+#define BR_GCC_4_4   1
+#endif
+
+#endif
+#endif
+
+/*
+ * Clang thresholds are on versions 3.7.0 and 3.8.0.
+ */
+#ifndef BR_CLANG
+#if __clang__
+#define BR_CLANG   1
+
+#if __clang_major__ > 3 || (__clang_major__ == 3 && __clang_minor__ >= 8)
+#define BR_CLANG_3_8   1
+#elif __clang_major__ == 3 && __clang_minor__ >= 7
+#define BR_CLANG_3_7   1
+#endif
+
+#if BR_CLANG_3_8
+#define BR_CLANG_3_7   1
+#endif
 
-#if (__i386__ || __x86_64__) \
-	&& ((__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 8)) \
-	    || (__clang_major__ > 3 \
-	        || (__clang_major__ == 3 && __clang_minor__ >= 7)))
-#define BR_AES_X86NI   1
-#elif (_M_IX86 || _M_X64) && (_MSC_VER >= 1700)
-#define BR_AES_X86NI   1
 #endif
 #endif
 
 /*
- * If we use x86 AES instruction, determine the compiler brand.
+ * MS Visual C thresholds are on Visual Studio 2005 to 2015.
  */
-#if BR_AES_X86NI
-#ifndef BR_AES_X86NI_GCC
-#if __GNUC__
-#define BR_AES_X86NI_GCC   1
+#ifndef BR_MSC
+#if _MSC_VER
+#define BR_MSC   1
+
+#if _MSC_VER >= 1900
+#define BR_MSC_2015   1
+#elif _MSC_VER >= 1800
+#define BR_MSC_2013   1
+#elif _MSC_VER >= 1700
+#define BR_MSC_2012   1
+#elif _MSC_VER >= 1600
+#define BR_MSC_2010   1
+#elif _MSC_VER >= 1500
+#define BR_MSC_2008   1
+#elif _MSC_VER >= 1400
+#define BR_MSC_2005   1
+#endif
+
+#if BR_MSC_2015
+#define BR_MSC_2013   1
+#endif
+#if BR_MSC_2013
+#define BR_MSC_2012   1
 #endif
+#if BR_MSC_2012
+#define BR_MSC_2010   1
 #endif
-#ifndef BR_AES_X86NI_MSC
-#if _MSC_VER >= 1700
-#define BR_AES_X86NI_MSC   1
+#if BR_MSC_2010
+#define BR_MSC_2008   1
 #endif
+#if BR_MSC_2008
+#define BR_MSC_2005   1
+#endif
+
 #endif
 #endif
 
 /*
- * A macro to tag a function with a "target" attribute (for GCC and Clang).
+ * GCC 4.4+ and Clang 3.7+ allow tagging specific functions with a
+ * 'target' attribute that activates support for specific opcodes.
  */
-#if BR_AES_X86NI_GCC
+#if BR_GCC_4_4 || BR_CLANG_3_7
 #define BR_TARGET(x)   __attribute__((target(x)))
 #else
 #define BR_TARGET(x)
 #endif
 
 /*
- * GCC versions from 4.4 to 4.8 (inclusive) must use a special #pragma
- * to activate extra opcodes before including the relevant intrinsic
- * headers. But these don't work with Clang (which does not need them
- * either).
+ * AES-NI intrinsics are available on x86 (32-bit and 64-bit) with
+ * GCC 4.8+, Clang 3.7+ and MSC 2012+.
+ */
+#ifndef BR_AES_X86NI
+#if (BR_i386 || BR_amd64) && (BR_GCC_4_8 || BR_CLANG_3_7 || BR_MSC_2012)
+#define BR_AES_X86NI   1
+#endif
+#endif
+
+/*
+ * SSE2 intrinsics are available on x86 (32-bit and 64-bit) with
+ * GCC 4.4+, Clang 3.7+ and MSC 2005+.
+ */
+#ifndef BR_SSE2
+#if (BR_i386 || BR_amd64) && (BR_GCC_4_4 || BR_CLANG_3_7 || BR_MSC_2005)
+#define BR_SSE2   1
+#endif
+#endif
+
+/*
+ * RDRAND intrinsics are available on x86 (32-bit and 64-bit) with
+ * GCC 4.6+, Clang 3.7+ and MSC 2012+.
  */
-#if BR_AES_X86NI_GCC && !defined BR_AES_X86NI_GCC_OLD
-#if __GNUC__ == 4 && __GNUC_MINOR__ >= 4 && __GNUC_MINOR__ <= 8 && !__clang__
-#define BR_AES_X86NI_GCC_OLD   1
+#ifndef BR_RDRAND
+#if (BR_i386 || BR_amd64) && (BR_GCC_4_6 || BR_CLANG_3_7 || BR_MSC_2012)
+#define BR_RDRAND   1
+#endif
+#endif
+
+/*
+ * Determine type of OS for random number generation. Macro names and
+ * values are documented on:
+ *    https://sourceforge.net/p/predef/wiki/OperatingSystems/
+ *
+ * TODO: enrich the list of detected system. Also add detection for
+ * alternate system calls like getentropy(), which are usually
+ * preferable when available.
+ */
+
+#ifndef BR_USE_URANDOM
+#if defined _AIX \
+	|| defined __ANDROID__ \
+	|| defined __FreeBSD__ \
+	|| defined __NetBSD__ \
+	|| defined __OpenBSD__ \
+	|| defined __DragonFly__ \
+	|| defined __linux__ \
+	|| (defined __sun && (defined __SVR4 || defined __svr4__)) \
+	|| (defined __APPLE__ && defined __MACH__)
+#define BR_USE_URANDOM   1
+#endif
+#endif
+
+#ifndef BR_USE_WIN32_RAND
+#if defined _WIN32 || defined _WIN64
+#define BR_USE_WIN32_RAND   1
 #endif
 #endif
 
@@ -220,58 +384,142 @@
 #endif
 #endif
 
+/*
+ * Detect support for unaligned accesses with known endianness.
+ *
+ *  x86 (both 32-bit and 64-bit) is little-endian and allows unaligned
+ *  accesses.
+ *
+ *  POWER/PowerPC allows unaligned accesses when big-endian. POWER8 and
+ *  later also allow unaligned accesses when little-endian.
+ */
+#if !defined BR_LE_UNALIGNED && !defined BR_BE_UNALIGNED
+
+#if __i386 || __i386__ || __x86_64__ || _M_IX86 || _M_X64
+#define BR_LE_UNALIGNED   1
+#elif BR_POWER8_BE
+#define BR_BE_UNALIGNED   1
+#elif BR_POWER8_LE
+#define BR_LE_UNALIGNED   1
+#elif (__powerpc__ || __powerpc64__ || _M_PPC || _ARCH_PPC || _ARCH_PPC64) \
+	&& __BIG_ENDIAN__
+#define BR_BE_UNALIGNED   1
+#endif
+
+#endif
+
+/*
+ * Detect support for an OS-provided time source.
+ */
+
+#ifndef BR_USE_UNIX_TIME
+#if defined __unix__ || defined __linux__ \
+	|| defined _POSIX_SOURCE || defined _POSIX_C_SOURCE \
+	|| (defined __APPLE__ && defined __MACH__)
+#define BR_USE_UNIX_TIME   1
+#endif
+#endif
+
+#ifndef BR_USE_WIN32_TIME
+#if defined _WIN32 || defined _WIN64
+#define BR_USE_WIN32_TIME   1
+#endif
+#endif
+
 /* ==================================================================== */
 /*
  * Encoding/decoding functions.
  *
  * 32-bit and 64-bit decoding, both little-endian and big-endian, is
- * implemented with the inline functions below. These functions are
- * generic: they don't depend on the architecture natural endianness,
- * and they can handle unaligned accesses. Optimized versions for some
- * specific architectures may be implemented at a later time.
- */
+ * implemented with the inline functions below.
+ *
+ * When allowed by some compile-time options (autodetected or provided),
+ * optimised code is used, to perform direct memory access when the
+ * underlying architecture supports it, both for endianness and
+ * alignment. This, however, may trigger strict aliasing issues; the
+ * code below uses unions to perform (supposedly) safe type punning.
+ * Since the C aliasing rules are relatively complex and were amended,
+ * or at least re-explained with different phrasing, in all successive
+ * versions of the C standard, it is always a bit risky to bet that any
+ * specific version of a C compiler got it right, for some notion of
+ * "right".
+ */
+
+typedef union {
+	uint16_t u;
+	unsigned char b[sizeof(uint16_t)];
+} br_union_u16;
+
+typedef union {
+	uint32_t u;
+	unsigned char b[sizeof(uint32_t)];
+} br_union_u32;
+
+typedef union {
+	uint64_t u;
+	unsigned char b[sizeof(uint64_t)];
+} br_union_u64;
 
 static inline void
 br_enc16le(void *dst, unsigned x)
 {
+#if BR_LE_UNALIGNED
+	((br_union_u16 *)dst)->u = x;
+#else
 	unsigned char *buf;
 
 	buf = dst;
 	buf[0] = (unsigned char)x;
 	buf[1] = (unsigned char)(x >> 8);
+#endif
 }
 
 static inline void
 br_enc16be(void *dst, unsigned x)
 {
+#if BR_BE_UNALIGNED
+	((br_union_u16 *)dst)->u = x;
+#else
 	unsigned char *buf;
 
 	buf = dst;
 	buf[0] = (unsigned char)(x >> 8);
 	buf[1] = (unsigned char)x;
+#endif
 }
 
 static inline unsigned
 br_dec16le(const void *src)
 {
+#if BR_LE_UNALIGNED
+	return ((const br_union_u16 *)src)->u;
+#else
 	const unsigned char *buf;
 
 	buf = src;
 	return (unsigned)buf[0] | ((unsigned)buf[1] << 8);
+#endif
 }
 
 static inline unsigned
 br_dec16be(const void *src)
 {
+#if BR_BE_UNALIGNED
+	return ((const br_union_u16 *)src)->u;
+#else
 	const unsigned char *buf;
 
 	buf = src;
 	return ((unsigned)buf[0] << 8) | (unsigned)buf[1];
+#endif
 }
 
 static inline void
 br_enc32le(void *dst, uint32_t x)
 {
+#if BR_LE_UNALIGNED
+	((br_union_u32 *)dst)->u = x;
+#else
 	unsigned char *buf;
 
 	buf = dst;
@@ -279,11 +527,15 @@ br_enc32le(void *dst, uint32_t x)
 	buf[1] = (unsigned char)(x >> 8);
 	buf[2] = (unsigned char)(x >> 16);
 	buf[3] = (unsigned char)(x >> 24);
+#endif
 }
 
 static inline void
 br_enc32be(void *dst, uint32_t x)
 {
+#if BR_BE_UNALIGNED
+	((br_union_u32 *)dst)->u = x;
+#else
 	unsigned char *buf;
 
 	buf = dst;
@@ -291,11 +543,15 @@ br_enc32be(void *dst, uint32_t x)
 	buf[1] = (unsigned char)(x >> 16);
 	buf[2] = (unsigned char)(x >> 8);
 	buf[3] = (unsigned char)x;
+#endif
 }
 
 static inline uint32_t
 br_dec32le(const void *src)
 {
+#if BR_LE_UNALIGNED
+	return ((const br_union_u32 *)src)->u;
+#else
 	const unsigned char *buf;
 
 	buf = src;
@@ -303,11 +559,15 @@ br_dec32le(const void *src)
 		| ((uint32_t)buf[1] << 8)
 		| ((uint32_t)buf[2] << 16)
 		| ((uint32_t)buf[3] << 24);
+#endif
 }
 
 static inline uint32_t
 br_dec32be(const void *src)
 {
+#if BR_BE_UNALIGNED
+	return ((const br_union_u32 *)src)->u;
+#else
 	const unsigned char *buf;
 
 	buf = src;
@@ -315,46 +575,63 @@ br_dec32be(const void *src)
 		| ((uint32_t)buf[1] << 16)
 		| ((uint32_t)buf[2] << 8)
 		| (uint32_t)buf[3];
+#endif
 }
 
 static inline void
 br_enc64le(void *dst, uint64_t x)
 {
+#if BR_LE_UNALIGNED
+	((br_union_u64 *)dst)->u = x;
+#else
 	unsigned char *buf;
 
 	buf = dst;
 	br_enc32le(buf, (uint32_t)x);
 	br_enc32le(buf + 4, (uint32_t)(x >> 32));
+#endif
 }
 
 static inline void
 br_enc64be(void *dst, uint64_t x)
 {
+#if BR_BE_UNALIGNED
+	((br_union_u64 *)dst)->u = x;
+#else
 	unsigned char *buf;
 
 	buf = dst;
 	br_enc32be(buf, (uint32_t)(x >> 32));
 	br_enc32be(buf + 4, (uint32_t)x);
+#endif
 }
 
 static inline uint64_t
 br_dec64le(const void *src)
 {
+#if BR_LE_UNALIGNED
+	return ((const br_union_u64 *)src)->u;
+#else
 	const unsigned char *buf;
 
 	buf = src;
 	return (uint64_t)br_dec32le(buf)
 		| ((uint64_t)br_dec32le(buf + 4) << 32);
+#endif
 }
 
 static inline uint64_t
 br_dec64be(const void *src)
 {
+#if BR_BE_UNALIGNED
+	return ((const br_union_u64 *)src)->u;
+#else
 	const unsigned char *buf;
 
 	buf = src;
 	return ((uint64_t)br_dec32be(buf) << 32)
 		| (uint64_t)br_dec32be(buf + 4);
+#endif
 }
 
 /*
@@ -413,8 +690,8 @@ void br_sha2small_round(const unsigned char *buf, uint32_t *val);
  */
 void br_tls_phash(void *dst, size_t len,
 	const br_hash_class *dig,
-	const void *secret, size_t secret_len,
-	const char *label, const void *seed, size_t seed_len);
+	const void *secret, size_t secret_len, const char *label,
+	size_t seed_num, const br_tls_prf_seed_chunk *seed);
 
 /*
  * Copy all configured hash implementations from a multihash context
@@ -1177,6 +1454,25 @@ void br_i31_from_monty(uint32_t *x, const uint32_t *m, uint32_t m0i);
 void br_i31_modpow(uint32_t *x, const unsigned char *e, size_t elen,
 	const uint32_t *m, uint32_t m0i, uint32_t *t1, uint32_t *t2);
 
+/*
+ * Compute a modular exponentiation. x[] MUST be an integer modulo m[]
+ * (same announced bit length, lower value). m[] MUST be odd. The
+ * exponent is in big-endian unsigned notation, over 'elen' bytes. The
+ * "m0i" parameter is equal to -(1/m0) mod 2^31, where m0 is the least
+ * significant value word of m[] (this works only if m[] is an odd
+ * integer). The tmp[] array is used for temporaries, and has size
+ * 'twlen' words; it must be large enough to accommodate at least two
+ * temporary values with the same size as m[] (including the leading
+ * "bit length" word). If there is room for more temporaries, then this
+ * function may use the extra room for window-based optimisation,
+ * resulting in faster computations.
+ *
+ * Returned value is 1 on success, 0 on error. An error is reported if
+ * the provided tmp[] array is too short.
+ */
+uint32_t br_i31_modpow_opt(uint32_t *x, const unsigned char *e, size_t elen,
+	const uint32_t *m, uint32_t m0i, uint32_t *tmp, size_t twlen);
+
 /*
  * Compute d+a*b, result in d. The initial announced bit length of d[]
  * MUST match that of a[]. The d[] array MUST be large enough to
@@ -1190,6 +1486,23 @@ void br_i31_modpow(uint32_t *x, const unsigned char *e, size_t elen,
  */
 void br_i31_mulacc(uint32_t *d, const uint32_t *a, const uint32_t *b);
 
+/*
+ * Compute x/y mod m, result in x. Values x and y must be between 0 and
+ * m-1, and have the same announced bit length as m. Modulus m must be
+ * odd. The "m0i" parameter is equal to -1/m mod 2^31. The array 't'
+ * must point to a temporary area that can hold at least three integers
+ * of the size of m.
+ *
+ * m may not overlap x and y. x and y may overlap each other (this can
+ * be useful to test whether a value is invertible modulo m). t must be
+ * disjoint from all other arrays.
+ *
+ * Returned value is 1 on success, 0 otherwise. Success is attained if
+ * y is invertible modulo m.
+ */
+uint32_t br_i31_moddiv(uint32_t *x, const uint32_t *y,
+	const uint32_t *m, uint32_t m0i, uint32_t *t);
+
 /* ==================================================================== */
 
 /*
@@ -1244,6 +1557,36 @@ void br_i15_reduce(uint16_t *x, const uint16_t *a, const uint16_t *m);
 
 void br_i15_mulacc(uint16_t *d, const uint16_t *a, const uint16_t *b);
 
+uint32_t br_i15_moddiv(uint16_t *x, const uint16_t *y,
+	const uint16_t *m, uint16_t m0i, uint16_t *t);
+
+/*
+ * Variant of br_i31_modpow_opt() that internally uses 64x64->128
+ * multiplications. It expects the same parameters as br_i31_modpow_opt(),
+ * except that the temporaries should be 64-bit integers, not 32-bit
+ * integers.
+ */
+uint32_t br_i62_modpow_opt(uint32_t *x31, const unsigned char *e, size_t elen,
+	const uint32_t *m31, uint32_t m0i31, uint64_t *tmp, size_t twlen);
+
+/*
+ * Type for a function with the same API as br_i31_modpow_opt() (some
+ * implementations of this type may have stricter alignment requirements
+ * on the temporaries).
+ */
+typedef uint32_t (*br_i31_modpow_opt_type)(uint32_t *x,
+	const unsigned char *e, size_t elen,
+	const uint32_t *m, uint32_t m0i, uint32_t *tmp, size_t twlen);
+
+/*
+ * Wrapper for br_i62_modpow_opt() that uses the same type as
+ * br_i31_modpow_opt(); however, it requires its 'tmp' argument to the
+ * 64-bit aligned.
+ */
+uint32_t br_i62_modpow_opt_as_i31(uint32_t *x,
+	const unsigned char *e, size_t elen,
+	const uint32_t *m, uint32_t m0i, uint32_t *tmp, size_t twlen);
+
 /* ==================================================================== */
 
 static inline size_t
@@ -1600,6 +1943,61 @@ uint32_t br_rsa_pkcs1_sig_unpad(const unsigned char *sig, size_t sig_len,
 	const unsigned char *hash_oid, size_t hash_len,
 	unsigned char *hash_out);
 
+/*
+ * Apply proper PSS padding. The 'x' buffer is output only: it
+ * receives the value that is to be exponentiated.
+ */
+uint32_t br_rsa_pss_sig_pad(const br_prng_class **rng,
+	const br_hash_class *hf_data, const br_hash_class *hf_mgf1,
+	const unsigned char *hash, size_t salt_len,
+	uint32_t n_bitlen, unsigned char *x);
+
+/*
+ * Check PSS padding. The provided value is the one _after_
+ * the modular exponentiation; it is modified by this function.
+ * This function infers the signature length from the public key
+ * size, i.e. it assumes that this has already been verified (as
+ * part of the exponentiation).
+ */
+uint32_t br_rsa_pss_sig_unpad(
+	const br_hash_class *hf_data, const br_hash_class *hf_mgf1,
+	const unsigned char *hash, size_t salt_len,
+	const br_rsa_public_key *pk, unsigned char *x);
+
+/*
+ * Apply OAEP padding. Returned value is the actual padded string length,
+ * or zero on error.
+ */
+size_t br_rsa_oaep_pad(const br_prng_class **rnd, const br_hash_class *dig,
+	const void *label, size_t label_len, const br_rsa_public_key *pk,
+	void *dst, size_t dst_nax_len, const void *src, size_t src_len);
+
+/*
+ * Unravel and check OAEP padding. If the padding is correct, then 1 is
+ * returned, '*len' is adjusted to the length of the message, and the
+ * data is moved to the start of the 'data' buffer. If the padding is
+ * incorrect, then 0 is returned and '*len' is untouched. Either way,
+ * the complete buffer contents are altered.
+ */
+uint32_t br_rsa_oaep_unpad(const br_hash_class *dig,
+	const void *label, size_t label_len, void *data, size_t *len);
+
+/*
+ * Compute MGF1 for a given seed, and XOR the output into the provided
+ * buffer.
+ */
+void br_mgf1_xor(void *data, size_t len,
+	const br_hash_class *dig, const void *seed, size_t seed_len);
+
+/*
+ * Inner function for RSA key generation; used by the "i31" and "i62"
+ * implementations.
+ */
+uint32_t br_rsa_i31_keygen_inner(const br_prng_class **rng,
+	br_rsa_private_key *sk, void *kbuf_priv,
+	br_rsa_public_key *pk, void *kbuf_pub,
+	unsigned size, uint32_t pubexp, br_i31_modpow_opt_type mp31);
+
 /* ==================================================================== */
 /*
  * Elliptic curves.
@@ -1650,6 +2048,72 @@ void br_ecdsa_i31_bits2int(uint32_t *x,
 void br_ecdsa_i15_bits2int(uint16_t *x,
 	const void *src, size_t len, uint32_t ebitlen);
 
+/* ==================================================================== */
+/*
+ * ASN.1 support functions.
+ */
+
+/*
+ * A br_asn1_uint structure contains encoding information about an
+ * INTEGER nonnegative value: pointer to the integer contents (unsigned
+ * big-endian representation), length of the integer contents,
+ * and length of the encoded value. The data shall have minimal length:
+ *  - If the integer value is zero, then 'len' must be zero.
+ *  - If the integer value is not zero, then data[0] must be non-zero.
+ *
+ * Under these conditions, 'asn1len' is necessarily equal to either len
+ * or len+1.
+ */
+typedef struct {
+	const unsigned char *data;
+	size_t len;
+	size_t asn1len;
+} br_asn1_uint;
+
+/*
+ * Given an encoded integer (unsigned big-endian, with possible leading
+ * bytes of value 0), returned the "prepared INTEGER" structure.
+ */
+br_asn1_uint br_asn1_uint_prepare(const void *xdata, size_t xlen);
+
+/*
+ * Encode an ASN.1 length. The length of the encoded length is returned.
+ * If 'dest' is NULL, then no encoding is performed, but the length of
+ * the encoded length is still computed and returned.
+ */
+size_t br_asn1_encode_length(void *dest, size_t len);
+
+/*
+ * Convenient macro for computing lengths of lengths.
+ */
+#define len_of_len(len)   br_asn1_encode_length(NULL, len)
+
+/*
+ * Encode a (prepared) ASN.1 INTEGER. The encoded length is returned.
+ * If 'dest' is NULL, then no encoding is performed, but the length of
+ * the encoded integer is still computed and returned.
+ */
+size_t br_asn1_encode_uint(void *dest, br_asn1_uint pp);
+
+/*
+ * Get the OID that identifies an elliptic curve. Returned value is
+ * the DER-encoded OID, with the length (always one byte) but without
+ * the tag. Thus, the first byte of the returned buffer contains the
+ * number of subsequent bytes in the value. If the curve is not
+ * recognised, NULL is returned.
+ */
+const unsigned char *br_get_curve_OID(int curve);
+
+/*
+ * Inner function for EC private key encoding. This is equivalent to
+ * the API function br_encode_ec_raw_der(), except for an extra
+ * parameter: if 'include_curve_oid' is zero, then the curve OID is
+ * _not_ included in the output blob (this is for PKCS#8 support).
+ */
+size_t br_encode_ec_raw_der_inner(void *dest,
+	const br_ec_private_key *sk, const br_ec_public_key *pk,
+	int include_curve_oid);
+
 /* ==================================================================== */
 /*
  * SSL/TLS support functions.
@@ -1835,6 +2299,34 @@ void br_ssl_engine_switch_chapol_in(br_ssl_engine_context *cc,
 void br_ssl_engine_switch_chapol_out(br_ssl_engine_context *cc,
 	int is_client, int prf_id);
 
+/*
+ * Switch to CCM decryption for incoming records.
+ *    cc               the engine context
+ *    is_client        non-zero for a client, zero for a server
+ *    prf_id           id of hash function for PRF
+ *    bc_impl          block cipher implementation (CTR+CBC)
+ *    cipher_key_len   block cipher key length (in bytes)
+ *    tag_len          tag length (in bytes)
+ */
+void br_ssl_engine_switch_ccm_in(br_ssl_engine_context *cc,
+	int is_client, int prf_id,
+	const br_block_ctrcbc_class *bc_impl,
+	size_t cipher_key_len, size_t tag_len);
+
+/*
+ * Switch to GCM encryption for outgoing records.
+ *    cc               the engine context
+ *    is_client        non-zero for a client, zero for a server
+ *    prf_id           id of hash function for PRF
+ *    bc_impl          block cipher implementation (CTR+CBC)
+ *    cipher_key_len   block cipher key length (in bytes)
+ *    tag_len          tag length (in bytes)
+ */
+void br_ssl_engine_switch_ccm_out(br_ssl_engine_context *cc,
+	int is_client, int prf_id,
+	const br_block_ctrcbc_class *bc_impl,
+	size_t cipher_key_len, size_t tag_len);
+
 /*
  * Calls to T0-generated code.
  */
@@ -1866,6 +2358,7 @@ int br_ssl_choose_hash(unsigned bf);
 #define stxvw4x(xt, ra, rb)       stxvw4x_(xt, ra, rb)
 
 #define bdnz(foo)                 bdnz_(foo)
+#define bdz(foo)                  bdz_(foo)
 #define beq(foo)                  beq_(foo)
 
 #define li(rx, value)             li_(rx, value)
@@ -1884,6 +2377,7 @@ int br_ssl_choose_hash(unsigned bf);
 #define vsl(vrt, vra, vrb)        vsl_(vrt, vra, vrb)
 #define vsldoi(vt, va, vb, sh)    vsldoi_(vt, va, vb, sh)
 #define vsr(vrt, vra, vrb)        vsr_(vrt, vra, vrb)
+#define vaddcuw(vrt, vra, vrb)    vaddcuw_(vrt, vra, vrb)
 #define vadduwm(vrt, vra, vrb)    vadduwm_(vrt, vra, vrb)
 #define vsububm(vrt, vra, vrb)    vsububm_(vrt, vra, vrb)
 #define vsubuwm(vrt, vra, vrb)    vsubuwm_(vrt, vra, vrb)
@@ -1901,6 +2395,7 @@ int br_ssl_choose_hash(unsigned bf);
 
 #define label(foo)                #foo "%=:\n"
 #define bdnz_(foo)                "\tbdnz\t" #foo "%=\n"
+#define bdz_(foo)                 "\tbdz\t" #foo "%=\n"
 #define beq_(foo)                 "\tbeq\t" #foo "%=\n"
 
 #define li_(rx, value)            "\tli\t" #rx "," #value "\n"
@@ -1919,6 +2414,7 @@ int br_ssl_choose_hash(unsigned bf);
 #define vsl_(vrt, vra, vrb)       "\tvsl\t" #vrt "," #vra "," #vrb "\n"
 #define vsldoi_(vt, va, vb, sh)   "\tvsldoi\t" #vt "," #va "," #vb "," #sh "\n"
 #define vsr_(vrt, vra, vrb)       "\tvsr\t" #vrt "," #vra "," #vrb "\n"
+#define vaddcuw_(vrt, vra, vrb)   "\tvaddcuw\t" #vrt "," #vra "," #vrb "\n"
 #define vadduwm_(vrt, vra, vrb)   "\tvadduwm\t" #vrt "," #vra "," #vrb "\n"
 #define vsububm_(vrt, vra, vrb)   "\tvsububm\t" #vrt "," #vra "," #vrb "\n"
 #define vsubuwm_(vrt, vra, vrb)   "\tvsubuwm\t" #vrt "," #vra "," #vrb "\n"
@@ -1933,6 +2429,125 @@ int br_ssl_choose_hash(unsigned bf);
 
 #endif
 
+/* ==================================================================== */
+/*
+ * Special "activate intrinsics" code, needed for some compiler versions.
+ * This is defined at the end of this file, so that it won't impact any
+ * of the inline functions defined previously; and it is controlled by
+ * a specific macro defined in the caller code.
+ *
+ * Calling code conventions:
+ *
+ *  - Caller must define BR_ENABLE_INTRINSICS before including "inner.h".
+ *  - Functions that use intrinsics must be enclosed in an "enabled"
+ *    region (between BR_TARGETS_X86_UP and BR_TARGETS_X86_DOWN).
+ *  - Functions that use intrinsics must be tagged with the appropriate
+ *    BR_TARGET().
+ */
+
+#if BR_ENABLE_INTRINSICS && (BR_GCC_4_4 || BR_CLANG_3_7 || BR_MSC_2005)
+
+/*
+ * x86 intrinsics (both 32-bit and 64-bit).
+ */
+#if BR_i386 || BR_amd64
+
+/*
+ * On GCC before version 5.0, we need to use the pragma to enable the
+ * target options globally, because the 'target' function attribute
+ * appears to be unreliable. Before 4.6 we must also avoid the
+ * push_options / pop_options mechanism, because it tends to trigger
+ * some internal compiler errors.
+ */
+#if BR_GCC && !BR_GCC_5_0
+#if BR_GCC_4_6
+#define BR_TARGETS_X86_UP \
+	_Pragma("GCC push_options") \
+	_Pragma("GCC target(\"sse2,ssse3,sse4.1,aes,pclmul,rdrnd\")")
+#define BR_TARGETS_X86_DOWN \
+	_Pragma("GCC pop_options")
+#else
+#define BR_TARGETS_X86_UP \
+	_Pragma("GCC target(\"sse2,ssse3,sse4.1,aes,pclmul\")")
+#define BR_TARGETS_X86_DOWN
+#endif
+#pragma GCC diagnostic ignored "-Wpsabi"
+#endif
+
+#if BR_CLANG && !BR_CLANG_3_8
+#undef __SSE2__
+#undef __SSE3__
+#undef __SSSE3__
+#undef __SSE4_1__
+#undef __AES__
+#undef __PCLMUL__
+#undef __RDRND__
+#define __SSE2__     1
+#define __SSE3__     1
+#define __SSSE3__    1
+#define __SSE4_1__   1
+#define __AES__      1
+#define __PCLMUL__   1
+#define __RDRND__    1
+#endif
+
+#ifndef BR_TARGETS_X86_UP
+#define BR_TARGETS_X86_UP
+#endif
+#ifndef BR_TARGETS_X86_DOWN
+#define BR_TARGETS_X86_DOWN
+#endif
+
+#if BR_GCC || BR_CLANG
+BR_TARGETS_X86_UP
+#include <x86intrin.h>
+#include <cpuid.h>
+#define br_bswap32   __builtin_bswap32
+BR_TARGETS_X86_DOWN
+#endif
+
+#if BR_MSC
+#include <stdlib.h>
+#include <intrin.h>
+#include <immintrin.h>
+#define br_bswap32   _byteswap_ulong
+#endif
+
+static inline int
+br_cpuid(uint32_t mask_eax, uint32_t mask_ebx,
+	uint32_t mask_ecx, uint32_t mask_edx)
+{
+#if BR_GCC || BR_CLANG
+	unsigned eax, ebx, ecx, edx;
+
+	if (__get_cpuid(1, &eax, &ebx, &ecx, &edx)) {
+		if ((eax & mask_eax) == mask_eax
+			&& (ebx & mask_ebx) == mask_ebx
+			&& (ecx & mask_ecx) == mask_ecx
+			&& (edx & mask_edx) == mask_edx)
+		{
+			return 1;
+		}
+	}
+#elif BR_MSC
+	int info[4];
+
+	__cpuid(info, 1);
+	if (((uint32_t)info[0] & mask_eax) == mask_eax
+		&& ((uint32_t)info[1] & mask_ebx) == mask_ebx
+		&& ((uint32_t)info[2] & mask_ecx) == mask_ecx
+		&& ((uint32_t)info[3] & mask_edx) == mask_edx)
+	{
+		return 1;
+	}
+#endif
+	return 0;
+}
+
+#endif
+
+#endif
+
 /* ==================================================================== */
 
 #endif