poly1305_sse2.c in rbnacl-libsodium-1.0.15

- old
+ new

@@ -38,18 +38,18 @@
 enum poly1305_state_flags_t {
     poly1305_started       = 1,
     poly1305_final_shift8  = 4,
     poly1305_final_shift16 = 8,
     poly1305_final_r2_r    = 16, /* use [r^2,r] for the final block */
-    poly1305_final_r_1     = 32, /* use [r,1] for the final block */
+    poly1305_final_r_1     = 32  /* use [r,1] for the final block */
 };
 
 typedef struct poly1305_state_internal_t {
     union {
         uint64_t h[3];
         uint32_t hh[10];
-    };                                              /*  40 bytes  */
+    } H;                                            /*  40 bytes  */
     uint32_t           R[5];                        /*  20 bytes  */
     uint32_t           R2[5];                       /*  20 bytes  */
     uint32_t           R4[5];                       /*  20 bytes  */
     uint64_t           pad[2];                      /*  16 bytes  */
     uint64_t           flags;                       /*   8 bytes  */
@@ -118,13 +118,13 @@
 
     if (!bytes) {
         bytes = ~(unsigned long long) 0;
     }
     /* H = 0 */
-    _mm_storeu_si128((xmmi *) (void *) &st->hh[0], _mm_setzero_si128());
-    _mm_storeu_si128((xmmi *) (void *) &st->hh[4], _mm_setzero_si128());
-    _mm_storeu_si128((xmmi *) (void *) &st->hh[8], _mm_setzero_si128());
+    _mm_storeu_si128((xmmi *) (void *) &st->H.hh[0], _mm_setzero_si128());
+    _mm_storeu_si128((xmmi *) (void *) &st->H.hh[4], _mm_setzero_si128());
+    _mm_storeu_si128((xmmi *) (void *) &st->H.hh[8], _mm_setzero_si128());
 
     /* clamp key */
     memcpy(&t0, key, 8);
     memcpy(&t1, key + 8, 8);
     r0 = t0 & 0xffc0fffffff;
@@ -240,13 +240,13 @@
         H4 = _mm_or_si128(H4, HIBIT);
         m += 32;
         bytes -= 32;
         st->flags |= poly1305_started;
     } else {
-        T0 = _mm_loadu_si128((const xmmi *) (const void *) &st->hh[0]);
-        T1 = _mm_loadu_si128((const xmmi *) (const void *) &st->hh[4]);
-        T2 = _mm_loadu_si128((const xmmi *) (const void *) &st->hh[8]);
+        T0 = _mm_loadu_si128((const xmmi *) (const void *) &st->H.hh[0]);
+        T1 = _mm_loadu_si128((const xmmi *) (const void *) &st->H.hh[4]);
+        T2 = _mm_loadu_si128((const xmmi *) (const void *) &st->H.hh[8]);
         H0 = _mm_shuffle_epi32(T0, _MM_SHUFFLE(1, 1, 0, 0));
         H1 = _mm_shuffle_epi32(T0, _MM_SHUFFLE(3, 3, 2, 2));
         H2 = _mm_shuffle_epi32(T1, _MM_SHUFFLE(1, 1, 0, 0));
         H3 = _mm_shuffle_epi32(T1, _MM_SHUFFLE(3, 3, 2, 2));
         H4 = _mm_shuffle_epi32(T2, _MM_SHUFFLE(1, 1, 0, 0));
@@ -682,13 +682,13 @@
         T2 = _mm_shuffle_epi32(H2, _MM_SHUFFLE(0, 0, 2, 0));
         T3 = _mm_shuffle_epi32(H3, _MM_SHUFFLE(0, 0, 2, 0));
         T4 = _mm_shuffle_epi32(H4, _MM_SHUFFLE(0, 0, 2, 0));
         T0 = _mm_unpacklo_epi64(T0, T1);
         T1 = _mm_unpacklo_epi64(T2, T3);
-        _mm_storeu_si128((xmmi *) (void *) &st->hh[0], T0);
-        _mm_storeu_si128((xmmi *) (void *) &st->hh[4], T1);
-        _mm_storel_epi64((xmmi *) (void *) &st->hh[8], T4);
+        _mm_storeu_si128((xmmi *) (void *) &st->H.hh[0], T0);
+        _mm_storeu_si128((xmmi *) (void *) &st->H.hh[4], T1);
+        _mm_storel_epi64((xmmi *) (void *) &st->H.hh[8], T4);
     } else {
         uint32_t t0, t1, t2, t3, t4, b;
         uint64_t h0, h1, h2, g0, g1, g2, c, nc;
 
         /* H = H[0]+H[1] */
@@ -753,13 +753,13 @@
         nc = ~c;
         h0 = (h0 & nc) | (g0 & c);
         h1 = (h1 & nc) | (g1 & c);
         h2 = (h2 & nc) | (g2 & c);
 
-        st->h[0] = h0;
-        st->h[1] = h1;
-        st->h[2] = h2;
+        st->H.h[0] = h0;
+        st->H.h[1] = h1;
+        st->H.h[2] = h2;
     }
 }
 
 static void
 poly1305_update(poly1305_state_internal_t *st, const unsigned char *m,
@@ -831,12 +831,12 @@
             st->flags |= poly1305_final_r_1;
         }
         poly1305_blocks(st, NULL, 32);
     }
 
-    h0 = st->h[0];
-    h1 = st->h[1];
-    h2 = st->h[2];
+    h0 = st->H.h[0];
+    h1 = st->H.h[1];
+    h2 = st->H.h[2];
 
     /* pad */
     h0 = ((h0) | (h1 << 44));
     h1 = ((h1 >> 20) | (h2 << 24));
 #ifdef HAVE_AMD64_ASM