diff --git a/src/crypto/slow-hash.c b/src/crypto/slow-hash.c
index 53628ab18..38aeeee54 100644
--- a/src/crypto/slow-hash.c
+++ b/src/crypto/slow-hash.c
@@ -51,6 +51,12 @@
 #define INIT_SIZE_BLK   8
 #define INIT_SIZE_BYTE (INIT_SIZE_BLK * AES_BLOCK_SIZE)
 
+#if defined(_MSC_VER)
+#define THREADV __declspec(thread)
+#else
+#define THREADV __thread
+#endif
+
 extern void aesb_single_round(const uint8_t *in, uint8_t *out, const uint8_t *expandedKey);
 extern void aesb_pseudo_round(const uint8_t *in, uint8_t *out, const uint8_t *expandedKey);
 
@@ -459,12 +465,6 @@ static inline int force_software_aes(void)
   _b1 = _b; \
   _b = _c; \
 
-#if defined(_MSC_VER)
-#define THREADV __declspec(thread)
-#else
-#define THREADV __thread
-#endif
-
 #pragma pack(push, 1)
 union cn_slow_hash_state
 {
@@ -1012,6 +1012,44 @@ void cn_slow_hash(const void *data, size_t length, char *hash, int variant, int
 }
 
 #elif !defined NO_AES && (defined(__arm__) || defined(__aarch64__))
+#ifdef __aarch64__
+#include <sys/mman.h>
+THREADV uint8_t *hp_state = NULL;
+THREADV int hp_malloced = 0;
+
+void cn_slow_hash_allocate_state(void)
+{
+    if(hp_state != NULL)
+        return;
+
+#ifndef	MAP_HUGETLB
+#define	MAP_HUGETLB	0
+#endif
+    hp_state = mmap(0, MEMORY, PROT_READ | PROT_WRITE,
+                    MAP_PRIVATE | MAP_ANON | MAP_HUGETLB, -1, 0);
+
+    if(hp_state == MAP_FAILED)
+        hp_state = NULL;
+    if(hp_state == NULL)
+    {
+        hp_malloced = 1;
+        hp_state = (uint8_t *) malloc(MEMORY);
+    }
+}
+
+void cn_slow_hash_free_state(void)
+{
+    if(hp_state == NULL)
+        return;
+
+    if (hp_malloced)
+        free(hp_state);
+    else
+        munmap(hp_state, MEMORY);
+    hp_state = NULL;
+    hp_malloced = 0;
+}
+#else
 void cn_slow_hash_allocate_state(void)
 {
   // Do nothing, this is just to maintain compatibility with the upgraded slow-hash.c
@@ -1023,6 +1061,7 @@ void cn_slow_hash_free_state(void)
   // As above
   return;
 }
+#endif
 
 #if defined(__GNUC__)
 #define RDATA_ALIGN16 __attribute__ ((aligned(16)))
@@ -1272,12 +1311,7 @@ void cn_slow_hash(const void *data, size_t length, char *hash, int variant, int
 {
     RDATA_ALIGN16 uint8_t expandedKey[240];
 
-#ifndef FORCE_USE_HEAP
-    RDATA_ALIGN16 uint8_t local_hp_state[MEMORY];
-#else
-    uint8_t *local_hp_state = (uint8_t *)aligned_malloc(MEMORY,16);
-#endif
-
+    uint8_t *local_hp_state;
     uint8_t text[INIT_SIZE_BYTE];
     RDATA_ALIGN16 uint64_t a[2];
     RDATA_ALIGN16 uint64_t b[4];
@@ -1296,6 +1330,14 @@ void cn_slow_hash(const void *data, size_t length, char *hash, int variant, int
         hash_extra_blake, hash_extra_groestl, hash_extra_jh, hash_extra_skein
     };
 
+    // this isn't supposed to happen, but guard against it for now.
+    if(hp_state == NULL)
+        cn_slow_hash_allocate_state();
+
+    // locals to avoid constant TLS dereferencing
+    local_hp_state = hp_state;
+
+    // locals to avoid constant TLS dereferencing
     /* CryptoNight Step 1:  Use Keccak1600 to initialize the 'state' (and 'text') buffers from the data. */
 
     if (prehashed) {
@@ -1409,10 +1451,6 @@ void cn_slow_hash(const void *data, size_t length, char *hash, int variant, int
     memcpy(state.init, text, INIT_SIZE_BYTE);
     hash_permutation(&state.hs);
     extra_hashes[state.hs.b[0] & 3](&state, 200, hash);
-
-#ifdef FORCE_USE_HEAP
-    aligned_free(local_hp_state);
-#endif
 }
 #else /* aarch64 && crypto */