*
* 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
* 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
-#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
+#ifndef BR_amd64
+#if __x86_64__ || _M_X64
+#define BR_amd64 1
#endif
#endif
/*
- * If we use x86 AES instruction, determine the compiler brand.
+ * 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.
*/
-#if BR_AES_X86NI
-#ifndef BR_AES_X86NI_GCC
-#if __GNUC__
-#define BR_AES_X86NI_GCC 1
+#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
-#ifndef BR_AES_X86NI_MSC
-#if _MSC_VER >= 1700
-#define BR_AES_X86NI_MSC 1
+
#endif
+#endif
+
+/*
+ * MS Visual C thresholds are on Visual Studio 2005 to 2015.
+ */
+#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
+#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
+/*
+ * 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+.
+ */
+#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
+
+/*
+ * POWER8 crypto support. We rely on compiler macros for the
+ * architecture, since we do not have a reliable, simple way to detect
+ * the required support at runtime (we could try running an opcode, and
+ * trapping the exception or signal on illegal instruction, but this
+ * induces some non-trivial OS dependencies that we would prefer to
+ * avoid if possible).
+ */
+#ifndef BR_POWER8
+#if __GNUC__ && ((_ARCH_PWR8 || _ARCH_PPC) && __CRYPTO__)
+#define BR_POWER8 1
+#endif
+#endif
+
+/*
+ * Detect endinanness on POWER8.
+ */
+#if BR_POWER8
+#if defined BR_POWER8_LE
+#undef BR_POWER8_BE
+#if BR_POWER8_LE
+#define BR_POWER8_BE 0
+#else
+#define BR_POWER8_BE 1
+#endif
+#elif defined BR_POWER8_BE
+#undef BR_POWER8_LE
+#if BR_POWER8_BE
+#define BR_POWER8_LE 0
+#else
+#define BR_POWER8_LE 1
+#endif
+#else
+#if __LITTLE_ENDIAN__
+#define BR_POWER8_LE 1
+#define BR_POWER8_BE 0
+#else
+#define BR_POWER8_LE 0
+#define BR_POWER8_BE 1
+#endif
+#endif
+#endif
+
+/*
+ * Detect support for 128-bit integers.
+ */
+#if !defined BR_INT128 && !defined BR_UMUL128
+#ifdef __SIZEOF_INT128__
+#define BR_INT128 1
+#elif _M_X64
+#define BR_UMUL128 1
+#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;
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;
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;
| ((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;
| ((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
}
/*
*/
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
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
void br_i15_mulacc(uint16_t *d, const uint16_t *a, const uint16_t *b);
+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);
+
/* ==================================================================== */
static inline size_t
unsigned br_aes_x86ni_keysched_dec(unsigned char *skni,
const void *key, size_t len);
+/*
+ * Test support for AES POWER8 opcodes.
+ */
+int br_aes_pwr8_supported(void);
+
+/*
+ * AES key schedule, using POWER8 instructions. This yields the
+ * subkeys in the encryption direction. Number of rounds is returned.
+ * Key size MUST be 16, 24 or 32 bytes; otherwise, 0 is returned.
+ */
+unsigned br_aes_pwr8_keysched(unsigned char *skni,
+ const void *key, size_t len);
+
/* ==================================================================== */
/*
* RSA.
const unsigned char *hash_oid, size_t hash_len,
unsigned char *hash_out);
+/*
+ * 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);
+
/* ==================================================================== */
/*
* Elliptic curves.
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.
*/
/* ==================================================================== */
+/*
+ * PowerPC / POWER assembly stuff. The special BR_POWER_ASM_MACROS macro
+ * must be defined before including this file; this is done by source
+ * files that use some inline assembly for PowerPC / POWER machines.
+ */
+
+#if BR_POWER_ASM_MACROS
+
+#define lxvw4x(xt, ra, rb) lxvw4x_(xt, ra, rb)
+#define stxvw4x(xt, ra, rb) stxvw4x_(xt, ra, rb)
+
+#define bdnz(foo) bdnz_(foo)
+#define beq(foo) beq_(foo)
+
+#define li(rx, value) li_(rx, value)
+#define addi(rx, ra, imm) addi_(rx, ra, imm)
+#define cmpldi(rx, imm) cmpldi_(rx, imm)
+#define mtctr(rx) mtctr_(rx)
+#define vspltb(vrt, vrb, uim) vspltb_(vrt, vrb, uim)
+#define vspltw(vrt, vrb, uim) vspltw_(vrt, vrb, uim)
+#define vspltisb(vrt, imm) vspltisb_(vrt, imm)
+#define vspltisw(vrt, imm) vspltisw_(vrt, imm)
+#define vrlw(vrt, vra, vrb) vrlw_(vrt, vra, vrb)
+#define vsbox(vrt, vra) vsbox_(vrt, vra)
+#define vxor(vrt, vra, vrb) vxor_(vrt, vra, vrb)
+#define vand(vrt, vra, vrb) vand_(vrt, vra, vrb)
+#define vsro(vrt, vra, vrb) vsro_(vrt, vra, vrb)
+#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 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)
+#define vsrw(vrt, vra, vrb) vsrw_(vrt, vra, vrb)
+#define vcipher(vt, va, vb) vcipher_(vt, va, vb)
+#define vcipherlast(vt, va, vb) vcipherlast_(vt, va, vb)
+#define vncipher(vt, va, vb) vncipher_(vt, va, vb)
+#define vncipherlast(vt, va, vb) vncipherlast_(vt, va, vb)
+#define vperm(vt, va, vb, vc) vperm_(vt, va, vb, vc)
+#define vpmsumd(vt, va, vb) vpmsumd_(vt, va, vb)
+#define xxpermdi(vt, va, vb, d) xxpermdi_(vt, va, vb, d)
+
+#define lxvw4x_(xt, ra, rb) "\tlxvw4x\t" #xt "," #ra "," #rb "\n"
+#define stxvw4x_(xt, ra, rb) "\tstxvw4x\t" #xt "," #ra "," #rb "\n"
+
+#define label(foo) #foo "%=:\n"
+#define bdnz_(foo) "\tbdnz\t" #foo "%=\n"
+#define beq_(foo) "\tbeq\t" #foo "%=\n"
+
+#define li_(rx, value) "\tli\t" #rx "," #value "\n"
+#define addi_(rx, ra, imm) "\taddi\t" #rx "," #ra "," #imm "\n"
+#define cmpldi_(rx, imm) "\tcmpldi\t" #rx "," #imm "\n"
+#define mtctr_(rx) "\tmtctr\t" #rx "\n"
+#define vspltb_(vrt, vrb, uim) "\tvspltb\t" #vrt "," #vrb "," #uim "\n"
+#define vspltw_(vrt, vrb, uim) "\tvspltw\t" #vrt "," #vrb "," #uim "\n"
+#define vspltisb_(vrt, imm) "\tvspltisb\t" #vrt "," #imm "\n"
+#define vspltisw_(vrt, imm) "\tvspltisw\t" #vrt "," #imm "\n"
+#define vrlw_(vrt, vra, vrb) "\tvrlw\t" #vrt "," #vra "," #vrb "\n"
+#define vsbox_(vrt, vra) "\tvsbox\t" #vrt "," #vra "\n"
+#define vxor_(vrt, vra, vrb) "\tvxor\t" #vrt "," #vra "," #vrb "\n"
+#define vand_(vrt, vra, vrb) "\tvand\t" #vrt "," #vra "," #vrb "\n"
+#define vsro_(vrt, vra, vrb) "\tvsro\t" #vrt "," #vra "," #vrb "\n"
+#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 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"
+#define vsrw_(vrt, vra, vrb) "\tvsrw\t" #vrt "," #vra "," #vrb "\n"
+#define vcipher_(vt, va, vb) "\tvcipher\t" #vt "," #va "," #vb "\n"
+#define vcipherlast_(vt, va, vb) "\tvcipherlast\t" #vt "," #va "," #vb "\n"
+#define vncipher_(vt, va, vb) "\tvncipher\t" #vt "," #va "," #vb "\n"
+#define vncipherlast_(vt, va, vb) "\tvncipherlast\t" #vt "," #va "," #vb "\n"
+#define vperm_(vt, va, vb, vc) "\tvperm\t" #vt "," #va "," #vb "," #vc "\n"
+#define vpmsumd_(vt, va, vb) "\tvpmsumd\t" #vt "," #va "," #vb "\n"
+#define xxpermdi_(vt, va, vb, d) "\txxpermdi\t" #vt "," #va "," #vb "," #d "\n"
+
+#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\")")
+#endif
+#define BR_TARGETS_X86_DOWN
+#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