gosthash2012: Switch Streebog implementations dynamically

Currently, only switch between see2 and ref implementations. This should
affect only i686, since x86_64 always have SSE2 unless compiled with
`-mno-sse2'.

This version of dynamic dispatch would work only on GCC-10 / Clang-3,
otherwise fallback to static dispatch like before.

Signed-off-by: Vitaly Chikunov <vt@altlinux.org>

Author: vt-alt

gosthash2012.c

@@ -114,13 +114,24 @@ _internal
 void g(union uint512_u *h, const union uint512_u * RESTRICT N,
     const union uint512_u * RESTRICT m)
 {
-#if defined __GOST3411_HAS_SSE2__
+#ifdef __GOST3411_DISPATCH__
+# if defined __GOST3411_HAS_SSE2__
+    if (__builtin_cpu_supports("sse2"))
+        return g_sse2(h, N, m);
+# elif defined  __GOST3411_HAS_REF__
+    g_ref(h, N, m);
+# else
+#  error "No implementation of g() is selected."
+# endif
+#else /* !__GOST3411_DISPATCH__ */
+# if defined __GOST3411_HAS_SSE2__ && defined __SSE2__
     g_sse2(h, N, m);
-#elif defined  __GOST3411_HAS_REF__
+# elif defined  __GOST3411_HAS_REF__
     g_ref(h, N, m);
-#else
+# else
 #  error "No implementation of g() is selected."
-#endif
+# endif
+#endif /* !__GOST3411_DISPATCH__ */
 }
 
 static INLINE void stage2(gost2012_hash_ctx * CTX, const union uint512_u *data)

gosthash2012.h

@@ -13,9 +13,11 @@
 /* Can be undef'd to disable ref impl. */
 #define __GOST3411_HAS_REF__
 
-#if defined __SSE2__
+#if defined __x86_64__ || defined __i386__
 # define __GOST3411_HAS_SSE2__
-# if !defined __x86_64__ && !defined __e2k__
+#elif defined __SSE2__
+# define __GOST3411_HAS_SSE2__
+# if !defined __e2k__
 /*
  * x86-64 bit Linux and Windows ABIs provide malloc function that returns
  * 16-byte alignment memory buffers required by SSE load/store instructions.
@@ -32,6 +34,14 @@
 # if (__GNUC__ < 4) || (__GNUC__ == 4 && __GNUC_MINOR__ < 2)
 #  undef __GOST3411_HAS_SSE2__
 # endif
+# ifdef __x86_64__
+/*
+ * On x86_64 there is always SSE2, so no need to even build reference
+ * implementation. But only if SSE2 is actually compiled, since it could
+ * be disabled with -mno-sse2.
+ */
+#  undef __GOST3411_HAS_REF__
+# endif
 #endif
 
 #ifndef L_ENDIAN
@@ -58,6 +68,15 @@
 #  define _internal
 #endif
 
+/* '__has_builtin is supported on gcc >= 10, clang >= 3 and icc >= 21.' */
+#ifndef __has_builtin
+#  define __has_builtin(x) 0
+#else
+#  if __has_builtin(__builtin_cpu_supports)
+#    define __GOST3411_DISPATCH__
+#  endif
+#endif
+
 ALIGN(16)
 typedef union uint512_u {
     unsigned long long QWORD[8];

gosthash2012_sse2.c

@@ -31,7 +31,13 @@
 # define _mm_cvtm64_si64(v) (long long) v
 #endif
 
-#ifdef __SSE3__
+/*
+ * We cannot just use SSE3 instructions in SSE2 implementation if dynamic
+ * dispatch is used. SSE3 belongs to different microarchitecture level
+ * (x86_64-v2) than SSE2 (x86_64 baseline). If there is x86_64-v2 CPU then
+ * SSE4.1 implementation should be used.
+ */
+#if defined __SSE3__ && !defined __GOST3411_DISPATCH__
 /*
  * "This intrinsic may perform better than _mm_loadu_si128 when
  * the data crosses a cache line boundary."