diff --git a/crypto/fipsmodule/aes/aes_nohw.c b/crypto/fipsmodule/aes/aes_nohw.c
index 308c7cca1c..a86f468cdf 100644
--- a/crypto/fipsmodule/aes/aes_nohw.c
+++ b/crypto/fipsmodule/aes/aes_nohw.c
@@ -912,29 +912,27 @@ void aes_nohw_ctr32_encrypt_blocks(const uint8_t *in, uint8_t *out,
   aes_nohw_expand_round_keys(&sched, key);
 
   // Make |AES_NOHW_BATCH_SIZE| copies of |ivec|.
-  alignas(AES_NOHW_WORD_SIZE) union {
-    uint32_t u32[AES_NOHW_BATCH_SIZE * 4];
-    uint8_t u8[AES_NOHW_BATCH_SIZE * 16];
-  } ivs, enc_ivs;
+  alignas(AES_NOHW_WORD_SIZE) uint8_t ivs[AES_NOHW_BATCH_SIZE * 16];
+  alignas(AES_NOHW_WORD_SIZE) uint8_t enc_ivs[AES_NOHW_BATCH_SIZE * 16];
   for (size_t i = 0; i < AES_NOHW_BATCH_SIZE; i++) {
-    OPENSSL_memcpy(ivs.u8 + 16 * i, ivec, 16);
+    OPENSSL_memcpy(ivs + 16 * i, ivec, 16);
   }
 
-  uint32_t ctr = CRYPTO_bswap4(ivs.u32[3]);
+  uint32_t ctr = CRYPTO_load_u32_be(ivs + 12);
   for (;;) {
     // Update counters.
     for (uint32_t i = 0; i < AES_NOHW_BATCH_SIZE; i++) {
-      ivs.u32[4 * i + 3] = CRYPTO_bswap4(ctr + i);
+      CRYPTO_store_u32_be(ivs + 16 * i + 12, ctr + i);
     }
 
     size_t todo = blocks >= AES_NOHW_BATCH_SIZE ? AES_NOHW_BATCH_SIZE : blocks;
     AES_NOHW_BATCH batch;
-    aes_nohw_to_batch(&batch, ivs.u8, todo);
+    aes_nohw_to_batch(&batch, ivs, todo);
     aes_nohw_encrypt_batch(&sched, key->rounds, &batch);
-    aes_nohw_from_batch(enc_ivs.u8, todo, &batch);
+    aes_nohw_from_batch(enc_ivs, todo, &batch);
 
     for (size_t i = 0; i < todo; i++) {
-      aes_nohw_xor_block(out + 16 * i, in + 16 * i, enc_ivs.u8 + 16 * i);
+      aes_nohw_xor_block(out + 16 * i, in + 16 * i, enc_ivs + 16 * i);
     }
 
     blocks -= todo;
diff --git a/crypto/internal.h b/crypto/internal.h
index dba229c399..cb2f703cc4 100644
--- a/crypto/internal.h
+++ b/crypto/internal.h
@@ -264,6 +264,10 @@ static inline crypto_word constant_time_select_w(crypto_word mask,
 static inline uint32_t CRYPTO_bswap4(uint32_t x) {
   return __builtin_bswap32(x);
 }
+
+static inline uint64_t CRYPTO_bswap8(uint64_t x) {
+  return __builtin_bswap64(x);
+}
 #elif defined(_MSC_VER)
 #pragma warning(push, 3)
 #include <stdlib.h>
@@ -272,6 +276,10 @@ static inline uint32_t CRYPTO_bswap4(uint32_t x) {
 static inline uint32_t CRYPTO_bswap4(uint32_t x) {
   return _byteswap_ulong(x);
 }
+
+static inline uint64_t CRYPTO_bswap8(uint64_t x) {
+  return _byteswap_uint64(x);
+}
 #endif
 
 #if !defined(RING_CORE_NOSTDLIBINC)
@@ -310,6 +318,65 @@ static inline void *OPENSSL_memset(void *dst, int c, size_t n) {
 }
 
 
+// Loads and stores.
+//
+// The following functions load and store sized integers with the specified
+// endianness. They use |memcpy|, and so avoid alignment or strict aliasing
+// requirements on the input and output pointers.
+
+static inline uint32_t CRYPTO_load_u32_le(const void *in) {
+  uint32_t v;
+  OPENSSL_memcpy(&v, in, sizeof(v));
+  return v;
+}
+
+static inline void CRYPTO_store_u32_le(void *out, uint32_t v) {
+  OPENSSL_memcpy(out, &v, sizeof(v));
+}
+
+static inline uint32_t CRYPTO_load_u32_be(const void *in) {
+  uint32_t v;
+  OPENSSL_memcpy(&v, in, sizeof(v));
+  return CRYPTO_bswap4(v);
+}
+
+static inline void CRYPTO_store_u32_be(void *out, uint32_t v) {
+  v = CRYPTO_bswap4(v);
+  OPENSSL_memcpy(out, &v, sizeof(v));
+}
+
+static inline uint64_t CRYPTO_load_u64_le(const void *in) {
+  uint64_t v;
+  OPENSSL_memcpy(&v, in, sizeof(v));
+  return v;
+}
+
+static inline void CRYPTO_store_u64_le(void *out, uint64_t v) {
+  OPENSSL_memcpy(out, &v, sizeof(v));
+}
+
+static inline uint64_t CRYPTO_load_u64_be(const void *ptr) {
+  uint64_t ret;
+  OPENSSL_memcpy(&ret, ptr, sizeof(ret));
+  return CRYPTO_bswap8(ret);
+}
+
+static inline void CRYPTO_store_u64_be(void *out, uint64_t v) {
+  v = CRYPTO_bswap8(v);
+  OPENSSL_memcpy(out, &v, sizeof(v));
+}
+
+static inline crypto_word CRYPTO_load_word_le(const void *in) {
+  crypto_word v;
+  OPENSSL_memcpy(&v, in, sizeof(v));
+  return v;
+}
+
+static inline void CRYPTO_store_word_le(void *out, crypto_word v) {
+  OPENSSL_memcpy(out, &v, sizeof(v));
+}
+
+
 // Runtime CPU feature support
 
 #if defined(OPENSSL_X86) || defined(OPENSSL_X86_64)
diff --git a/crypto/poly1305/poly1305_vec.c b/crypto/poly1305/poly1305_vec.c
index 610a8c9512..08ae68bad4 100644
--- a/crypto/poly1305/poly1305_vec.c
+++ b/crypto/poly1305/poly1305_vec.c
@@ -31,22 +31,6 @@
 
 #include <emmintrin.h>
 
-static uint32_t load_u32_le(const uint8_t in[4]) {
-  uint32_t ret;
-  OPENSSL_memcpy(&ret, in, 4);
-  return ret;
-}
-
-static uint64_t load_u64_le(const uint8_t in[8]) {
-  uint64_t ret;
-  OPENSSL_memcpy(&ret, in, 8);
-  return ret;
-}
-
-static void store_u64_le(uint8_t out[8], uint64_t v) {
-  OPENSSL_memcpy(out, &v, 8);
-}
-
 typedef __m128i xmmi;
 
 static const alignas(16) uint32_t poly1305_x64_sse2_message_mask[4] = {
@@ -117,8 +101,8 @@ void CRYPTO_poly1305_init(poly1305_state *state, const uint8_t key[32]) {
   uint64_t t0, t1;
 
   // clamp key
-  t0 = load_u64_le(key + 0);
-  t1 = load_u64_le(key + 8);
+  t0 = CRYPTO_load_u64_le(key + 0);
+  t1 = CRYPTO_load_u64_le(key + 8);
   r0 = t0 & 0xffc0fffffff;
   t0 >>= 44;
   t0 |= t1 << 20;
@@ -136,10 +120,10 @@ void CRYPTO_poly1305_init(poly1305_state *state, const uint8_t key[32]) {
   p->R22.d[3] = (uint32_t)(r2 >> 32);
 
   // store pad
-  p->R23.d[1] = load_u32_le(key + 16);
-  p->R23.d[3] = load_u32_le(key + 20);
-  p->R24.d[1] = load_u32_le(key + 24);
-  p->R24.d[3] = load_u32_le(key + 28);
+  p->R23.d[1] = CRYPTO_load_u32_le(key + 16);
+  p->R23.d[3] = CRYPTO_load_u32_le(key + 20);
+  p->R24.d[1] = CRYPTO_load_u32_le(key + 24);
+  p->R24.d[3] = CRYPTO_load_u32_le(key + 28);
 
   // H = 0
   st->H[0] = _mm_setzero_si128();
@@ -771,8 +755,8 @@ void CRYPTO_poly1305_finish(poly1305_state *state, uint8_t mac[16]) {
   }
 
 poly1305_donna_atleast16bytes:
-  t0 = load_u64_le(m + 0);
-  t1 = load_u64_le(m + 8);
+  t0 = CRYPTO_load_u64_le(m + 0);
+  t1 = CRYPTO_load_u64_le(m + 8);
   h0 += t0 & 0xfffffffffff;
   t0 = shr128_pair(t1, t0, 44);
   h1 += t0 & 0xfffffffffff;
@@ -811,8 +795,8 @@ void CRYPTO_poly1305_finish(poly1305_state *state, uint8_t mac[16]) {
   OPENSSL_memset(m + leftover, 0, 16 - leftover);
   leftover = 16;
 
-  t0 = load_u64_le(m + 0);
-  t1 = load_u64_le(m + 8);
+  t0 = CRYPTO_load_u64_le(m + 0);
+  t1 = CRYPTO_load_u64_le(m + 8);
   h0 += t0 & 0xfffffffffff;
   t0 = shr128_pair(t1, t0, 44);
   h1 += t0 & 0xfffffffffff;
@@ -858,8 +842,8 @@ void CRYPTO_poly1305_finish(poly1305_state *state, uint8_t mac[16]) {
   t1 = (t1 >> 24);
   h2 += (t1)+c;
 
-  store_u64_le(mac + 0, ((h0) | (h1 << 44)));
-  store_u64_le(mac + 8, ((h1 >> 20) | (h2 << 24)));
+  CRYPTO_store_u64_le(mac + 0, ((h0) | (h1 << 44)));
+  CRYPTO_store_u64_le(mac + 8, ((h1 >> 20) | (h2 << 24)));
 }
 
 #endif  // BORINGSSL_HAS_UINT128 && OPENSSL_X86_64