diff --git a/configure.ac b/configure.ac
index ce331a967b..580bdc9f21 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1645,6 +1645,19 @@ then
 fi
 
 
+# STACK usage logging
+AC_ARG_ENABLE([stacklog],
+    [AS_HELP_STRING([--enable-stacklog],[Enable stack logging (default: disabled)])],
+    [ ENABLED_STACKLOG=$enableval ],
+    [ ENABLED_STACKLOG=no ]
+    )
+
+if test "$ENABLED_STACKLOG" = "yes"
+then
+    AM_CFLAGS="$AM_CFLAGS -DWOLFSSL_STACK_LOG -finstrument-functions"
+fi
+
+
 
 # RSA
 AC_ARG_ENABLE([rsa],
diff --git a/src/internal.c b/src/internal.c
index eb7ed702ef..e62a43221c 100644
--- a/src/internal.c
+++ b/src/internal.c
@@ -8679,6 +8679,8 @@ int ProcessPeerCerts(WOLFSSL* ssl, byte* input, word32* inOutIdx,
     (void)sizeof(args_test);
 #elif defined(WOLFSSL_NONBLOCK_OCSP)
     ProcPeerCertArgs* args = ssl->nonblockarg;
+#elif defined(WOLFSSL_SMALL_STACK)
+    ProcPeerCertArgs* args = NULL;
 #else
     ProcPeerCertArgs  args[1];
 #endif
@@ -8707,6 +8709,12 @@ int ProcessPeerCerts(WOLFSSL* ssl, byte* input, word32* inOutIdx,
         }
     }
     if (ssl->nonblockarg == NULL) /* new args */
+#elif defined(WOLFSSL_SMALL_STACK)
+    args = (ProcPeerCertArgs*)XMALLOC(
+        sizeof(ProcPeerCertArgs), ssl->heap, DYNAMIC_TYPE_TMP_BUFFER);
+    if (args == NULL) {
+        ERROR_OUT(MEMORY_E, exit_ppc);
+    }
 #endif
     {
         /* Reset state */
@@ -9837,9 +9845,12 @@ exit_ppc:
 
     FreeProcPeerCertArgs(ssl, args);
 
-#if !defined(WOLFSSL_ASYNC_CRYPT) && defined(WOLFSSL_NONBLOCK_OCSP)
+#if defined(WOLFSSL_ASYNC_CRYPT)
+#elif defined(WOLFSSL_NONBLOCK_OCSP)
     XFREE(args, ssl->heap, DYNAMIC_TYPE_TMP_BUFFER);
     ssl->nonblockarg = NULL;
+#elif defined(WOLFSSL_SMALL_STACK)
+    XFREE(args, ssl->heap, DYNAMIC_TYPE_TMP_BUFFER);
 #endif
 
     FreeKeyExchange(ssl);
diff --git a/wolfcrypt/src/memory.c b/wolfcrypt/src/memory.c
index 87ed921f3d..55111662f3 100644
--- a/wolfcrypt/src/memory.c
+++ b/wolfcrypt/src/memory.c
@@ -1074,3 +1074,22 @@ void xfree(void *p, void* heap, int type, const char* func, const char* file,
 }
 #endif /* WOLFSSL_MEMORY_LOG */
 
+#ifdef WOLFSSL_STACK_LOG
+/* Note: this code only works with GCC using -finstrument-functions. */
+void __attribute__((no_instrument_function))
+     __cyg_profile_func_enter(void *func,  void *caller)
+{
+    register void* sp asm("sp");
+    fprintf(stderr, "ENTER: %016lx %p\n", (size_t)func, sp);
+    (void)caller;
+}
+
+void __attribute__((no_instrument_function))
+     __cyg_profile_func_exit(void *func, void *caller)
+{
+    register void* sp asm("sp");
+    fprintf(stderr, "EXIT: %016lx %p\n", (size_t)func, sp);
+    (void)caller;
+}
+#endif
+
diff --git a/wolfcrypt/src/sp_arm32.c b/wolfcrypt/src/sp_arm32.c
index 6510a22e46..ad12487160 100644
--- a/wolfcrypt/src/sp_arm32.c
+++ b/wolfcrypt/src/sp_arm32.c
@@ -19184,8 +19184,7 @@ static void sp_256_div2_8(sp_digit* r, sp_digit* a, sp_digit* m)
  */
 static void sp_256_proj_point_dbl_8(sp_point* r, sp_point* p, sp_digit* t)
 {
-    sp_point *rp[2];
-    sp_point tp;
+    sp_point* rp[2];
     sp_digit* t1 = t;
     sp_digit* t2 = t + 2*8;
     sp_digit* x;
@@ -19195,7 +19194,8 @@ static void sp_256_proj_point_dbl_8(sp_point* r, sp_point* p, sp_digit* t)
 
     /* When infinity don't double point passed in - constant time. */
     rp[0] = r;
-    rp[1] = &tp;
+    rp[1] = (sp_point*)t;
+    XMEMSET(rp[1], 0, sizeof(sp_point));
     x = rp[p->infinity]->x;
     y = rp[p->infinity]->y;
     z = rp[p->infinity]->z;
@@ -19370,9 +19370,8 @@ static int sp_256_cmp_equal_8(const sp_digit* a, const sp_digit* b)
 static void sp_256_proj_point_add_8(sp_point* r, sp_point* p, sp_point* q,
         sp_digit* t)
 {
-    sp_point *ap[2];
-    sp_point *rp[2];
-    sp_point tp;
+    sp_point* ap[2];
+    sp_point* rp[2];
     sp_digit* t1 = t;
     sp_digit* t2 = t + 2*8;
     sp_digit* t3 = t + 4*8;
@@ -19399,8 +19398,8 @@ static void sp_256_proj_point_add_8(sp_point* r, sp_point* p, sp_point* q,
     }
     else {
         rp[0] = r;
-        rp[1] = &tp;
-        XMEMSET(&tp, 0, sizeof(tp));
+        rp[1] = (sp_point*)t;
+        XMEMSET(rp[1], 0, sizeof(sp_point));
         x = rp[p->infinity | q->infinity]->x;
         y = rp[p->infinity | q->infinity]->y;
         z = rp[p->infinity | q->infinity]->z;
@@ -19595,8 +19594,7 @@ typedef struct sp_table_entry {
 static void sp_256_proj_point_dbl_n_8(sp_point* r, sp_point* p, int n,
         sp_digit* t)
 {
-    sp_point *rp[2];
-    sp_point tp;
+    sp_point* rp[2];
     sp_digit* w = t;
     sp_digit* a = t + 2*8;
     sp_digit* b = t + 4*8;
@@ -19608,7 +19606,8 @@ static void sp_256_proj_point_dbl_n_8(sp_point* r, sp_point* p, int n,
     int i;
 
     rp[0] = r;
-    rp[1] = &tp;
+    rp[1] = (sp_point*)t;
+    XMEMSET(rp[1], 0, sizeof(sp_point));
     x = rp[p->infinity]->x;
     y = rp[p->infinity]->y;
     z = rp[p->infinity]->z;
@@ -19670,9 +19669,8 @@ static void sp_256_proj_point_dbl_n_8(sp_point* r, sp_point* p, int n,
 static void sp_256_proj_point_add_qz1_8(sp_point* r, sp_point* p,
         sp_point* q, sp_digit* t)
 {
-    sp_point *ap[2];
-    sp_point *rp[2];
-    sp_point tp;
+    sp_point* ap[2];
+    sp_point* rp[2];
     sp_digit* t1 = t;
     sp_digit* t2 = t + 2*8;
     sp_digit* t3 = t + 4*8;
@@ -19692,8 +19690,8 @@ static void sp_256_proj_point_add_qz1_8(sp_point* r, sp_point* p,
     }
     else {
         rp[0] = r;
-        rp[1] = &tp;
-        XMEMSET(&tp, 0, sizeof(tp));
+        rp[1] = (sp_point*)t;
+        XMEMSET(rp[1], 0, sizeof(sp_point));
         x = rp[p->infinity | q->infinity]->x;
         y = rp[p->infinity | q->infinity]->y;
         z = rp[p->infinity | q->infinity]->z;
diff --git a/wolfcrypt/src/sp_arm64.c b/wolfcrypt/src/sp_arm64.c
index 03e5823c57..1efb881083 100644
--- a/wolfcrypt/src/sp_arm64.c
+++ b/wolfcrypt/src/sp_arm64.c
@@ -13407,8 +13407,7 @@ static void sp_256_div2_4(sp_digit* r, sp_digit* a, sp_digit* m)
  */
 static void sp_256_proj_point_dbl_4(sp_point* r, sp_point* p, sp_digit* t)
 {
-    sp_point *rp[2];
-    sp_point tp;
+    sp_point* rp[2];
     sp_digit* t1 = t;
     sp_digit* t2 = t + 2*4;
     sp_digit* x;
@@ -13418,7 +13417,8 @@ static void sp_256_proj_point_dbl_4(sp_point* r, sp_point* p, sp_digit* t)
 
     /* When infinity don't double point passed in - constant time. */
     rp[0] = r;
-    rp[1] = &tp;
+    rp[1] = (sp_point*)t;
+    XMEMSET(rp[1], 0, sizeof(sp_point));
     x = rp[p->infinity]->x;
     y = rp[p->infinity]->y;
     z = rp[p->infinity]->z;
@@ -13482,8 +13482,7 @@ static void sp_256_proj_point_dbl_4(sp_point* r, sp_point* p, sp_digit* t)
 static void sp_256_proj_point_dbl_n_4(sp_point* r, sp_point* p, int n,
         sp_digit* t)
 {
-    sp_point *rp[2];
-    sp_point tp;
+    sp_point* rp[2];
     sp_digit* w = t;
     sp_digit* a = t + 2*4;
     sp_digit* b = t + 4*4;
@@ -13495,7 +13494,8 @@ static void sp_256_proj_point_dbl_n_4(sp_point* r, sp_point* p, int n,
     int i;
 
     rp[0] = r;
-    rp[1] = &tp;
+    rp[1] = (sp_point*)t;
+    XMEMSET(rp[1], 0, sizeof(sp_point));
     x = rp[p->infinity]->x;
     y = rp[p->infinity]->y;
     z = rp[p->infinity]->z;
@@ -13566,9 +13566,8 @@ static int sp_256_cmp_equal_4(const sp_digit* a, const sp_digit* b)
 static void sp_256_proj_point_add_4(sp_point* r, sp_point* p, sp_point* q,
         sp_digit* t)
 {
-    sp_point *ap[2];
-    sp_point *rp[2];
-    sp_point tp;
+    sp_point* ap[2];
+    sp_point* rp[2];
     sp_digit* t1 = t;
     sp_digit* t2 = t + 2*4;
     sp_digit* t3 = t + 4*4;
@@ -13595,8 +13594,8 @@ static void sp_256_proj_point_add_4(sp_point* r, sp_point* p, sp_point* q,
     }
     else {
         rp[0] = r;
-        rp[1] = &tp;
-        XMEMSET(&tp, 0, sizeof(tp));
+        rp[1] = (sp_point*)t;
+        XMEMSET(rp[1], 0, sizeof(sp_point));
         x = rp[p->infinity | q->infinity]->x;
         y = rp[p->infinity | q->infinity]->y;
         z = rp[p->infinity | q->infinity]->z;
@@ -13996,9 +13995,8 @@ typedef struct sp_table_entry {
 static void sp_256_proj_point_add_qz1_4(sp_point* r, sp_point* p,
         sp_point* q, sp_digit* t)
 {
-    sp_point *ap[2];
-    sp_point *rp[2];
-    sp_point tp;
+    sp_point* ap[2];
+    sp_point* rp[2];
     sp_digit* t1 = t;
     sp_digit* t2 = t + 2*4;
     sp_digit* t3 = t + 4*4;
@@ -14018,8 +14016,8 @@ static void sp_256_proj_point_add_qz1_4(sp_point* r, sp_point* p,
     }
     else {
         rp[0] = r;
-        rp[1] = &tp;
-        XMEMSET(&tp, 0, sizeof(tp));
+        rp[1] = (sp_point*)t;
+        XMEMSET(rp[1], 0, sizeof(sp_point));
         x = rp[p->infinity | q->infinity]->x;
         y = rp[p->infinity | q->infinity]->y;
         z = rp[p->infinity | q->infinity]->z;
diff --git a/wolfcrypt/src/sp_c32.c b/wolfcrypt/src/sp_c32.c
index 82f638d370..4c7a076b7b 100644
--- a/wolfcrypt/src/sp_c32.c
+++ b/wolfcrypt/src/sp_c32.c
@@ -8503,8 +8503,7 @@ static void sp_256_div2_10(sp_digit* r, sp_digit* a, sp_digit* m)
  */
 static void sp_256_proj_point_dbl_10(sp_point* r, sp_point* p, sp_digit* t)
 {
-    sp_point *rp[2];
-    sp_point tp;
+    sp_point* rp[2];
     sp_digit* t1 = t;
     sp_digit* t2 = t + 2*10;
     sp_digit* x;
@@ -8514,7 +8513,8 @@ static void sp_256_proj_point_dbl_10(sp_point* r, sp_point* p, sp_digit* t)
 
     /* When infinity don't double point passed in - constant time. */
     rp[0] = r;
-    rp[1] = &tp;
+    rp[1] = (sp_point*)t;
+    XMEMSET(rp[1], 0, sizeof(sp_point));
     x = rp[p->infinity]->x;
     y = rp[p->infinity]->y;
     z = rp[p->infinity]->z;
@@ -8592,9 +8592,8 @@ static int sp_256_cmp_equal_10(const sp_digit* a, const sp_digit* b)
 static void sp_256_proj_point_add_10(sp_point* r, sp_point* p, sp_point* q,
         sp_digit* t)
 {
-    sp_point *ap[2];
-    sp_point *rp[2];
-    sp_point tp;
+    sp_point* ap[2];
+    sp_point* rp[2];
     sp_digit* t1 = t;
     sp_digit* t2 = t + 2*10;
     sp_digit* t3 = t + 4*10;
@@ -8621,8 +8620,8 @@ static void sp_256_proj_point_add_10(sp_point* r, sp_point* p, sp_point* q,
     }
     else {
         rp[0] = r;
-        rp[1] = &tp;
-        XMEMSET(&tp, 0, sizeof(tp));
+        rp[1] = (sp_point*)t;
+        XMEMSET(rp[1], 0, sizeof(sp_point));
         x = rp[p->infinity | q->infinity]->x;
         y = rp[p->infinity | q->infinity]->y;
         z = rp[p->infinity | q->infinity]->z;
@@ -9017,8 +9016,7 @@ static int sp_256_ecc_mulmod_fast_10(sp_point* r, sp_point* g, sp_digit* k,
 static void sp_256_proj_point_dbl_n_10(sp_point* r, sp_point* p, int n,
         sp_digit* t)
 {
-    sp_point *rp[2];
-    sp_point tp;
+    sp_point* rp[2];
     sp_digit* w = t;
     sp_digit* a = t + 2*10;
     sp_digit* b = t + 4*10;
@@ -9030,7 +9028,8 @@ static void sp_256_proj_point_dbl_n_10(sp_point* r, sp_point* p, int n,
     int i;
 
     rp[0] = r;
-    rp[1] = &tp;
+    rp[1] = (sp_point*)t;
+    XMEMSET(rp[1], 0, sizeof(sp_point));
     x = rp[p->infinity]->x;
     y = rp[p->infinity]->y;
     z = rp[p->infinity]->z;
@@ -9092,9 +9091,8 @@ static void sp_256_proj_point_dbl_n_10(sp_point* r, sp_point* p, int n,
 static void sp_256_proj_point_add_qz1_10(sp_point* r, sp_point* p,
         sp_point* q, sp_digit* t)
 {
-    sp_point *ap[2];
-    sp_point *rp[2];
-    sp_point tp;
+    sp_point* ap[2];
+    sp_point* rp[2];
     sp_digit* t1 = t;
     sp_digit* t2 = t + 2*10;
     sp_digit* t3 = t + 4*10;
@@ -9114,8 +9112,8 @@ static void sp_256_proj_point_add_qz1_10(sp_point* r, sp_point* p,
     }
     else {
         rp[0] = r;
-        rp[1] = &tp;
-        XMEMSET(&tp, 0, sizeof(tp));
+        rp[1] = (sp_point*)t;
+        XMEMSET(rp[1], 0, sizeof(sp_point));
         x = rp[p->infinity | q->infinity]->x;
         y = rp[p->infinity | q->infinity]->y;
         z = rp[p->infinity | q->infinity]->z;
diff --git a/wolfcrypt/src/sp_c64.c b/wolfcrypt/src/sp_c64.c
index b906b039d8..c0823eba66 100644
--- a/wolfcrypt/src/sp_c64.c
+++ b/wolfcrypt/src/sp_c64.c
@@ -7726,8 +7726,7 @@ static void sp_256_div2_5(sp_digit* r, sp_digit* a, sp_digit* m)
  */
 static void sp_256_proj_point_dbl_5(sp_point* r, sp_point* p, sp_digit* t)
 {
-    sp_point *rp[2];
-    sp_point tp;
+    sp_point* rp[2];
     sp_digit* t1 = t;
     sp_digit* t2 = t + 2*5;
     sp_digit* x;
@@ -7737,7 +7736,8 @@ static void sp_256_proj_point_dbl_5(sp_point* r, sp_point* p, sp_digit* t)
 
     /* When infinity don't double point passed in - constant time. */
     rp[0] = r;
-    rp[1] = &tp;
+    rp[1] = (sp_point*)t;
+    XMEMSET(rp[1], 0, sizeof(sp_point));
     x = rp[p->infinity]->x;
     y = rp[p->infinity]->y;
     z = rp[p->infinity]->z;
@@ -7814,9 +7814,8 @@ static int sp_256_cmp_equal_5(const sp_digit* a, const sp_digit* b)
 static void sp_256_proj_point_add_5(sp_point* r, sp_point* p, sp_point* q,
         sp_digit* t)
 {
-    sp_point *ap[2];
-    sp_point *rp[2];
-    sp_point tp;
+    sp_point* ap[2];
+    sp_point* rp[2];
     sp_digit* t1 = t;
     sp_digit* t2 = t + 2*5;
     sp_digit* t3 = t + 4*5;
@@ -7843,8 +7842,8 @@ static void sp_256_proj_point_add_5(sp_point* r, sp_point* p, sp_point* q,
     }
     else {
         rp[0] = r;
-        rp[1] = &tp;
-        XMEMSET(&tp, 0, sizeof(tp));
+        rp[1] = (sp_point*)t;
+        XMEMSET(rp[1], 0, sizeof(sp_point));
         x = rp[p->infinity | q->infinity]->x;
         y = rp[p->infinity | q->infinity]->y;
         z = rp[p->infinity | q->infinity]->z;
@@ -8239,8 +8238,7 @@ static int sp_256_ecc_mulmod_fast_5(sp_point* r, sp_point* g, sp_digit* k,
 static void sp_256_proj_point_dbl_n_5(sp_point* r, sp_point* p, int n,
         sp_digit* t)
 {
-    sp_point *rp[2];
-    sp_point tp;
+    sp_point* rp[2];
     sp_digit* w = t;
     sp_digit* a = t + 2*5;
     sp_digit* b = t + 4*5;
@@ -8252,7 +8250,8 @@ static void sp_256_proj_point_dbl_n_5(sp_point* r, sp_point* p, int n,
     int i;
 
     rp[0] = r;
-    rp[1] = &tp;
+    rp[1] = (sp_point*)t;
+    XMEMSET(rp[1], 0, sizeof(sp_point));
     x = rp[p->infinity]->x;
     y = rp[p->infinity]->y;
     z = rp[p->infinity]->z;
@@ -8314,9 +8313,8 @@ static void sp_256_proj_point_dbl_n_5(sp_point* r, sp_point* p, int n,
 static void sp_256_proj_point_add_qz1_5(sp_point* r, sp_point* p,
         sp_point* q, sp_digit* t)
 {
-    sp_point *ap[2];
-    sp_point *rp[2];
-    sp_point tp;
+    sp_point* ap[2];
+    sp_point* rp[2];
     sp_digit* t1 = t;
     sp_digit* t2 = t + 2*5;
     sp_digit* t3 = t + 4*5;
@@ -8336,8 +8334,8 @@ static void sp_256_proj_point_add_qz1_5(sp_point* r, sp_point* p,
     }
     else {
         rp[0] = r;
-        rp[1] = &tp;
-        XMEMSET(&tp, 0, sizeof(tp));
+        rp[1] = (sp_point*)t;
+        XMEMSET(rp[1], 0, sizeof(sp_point));
         x = rp[p->infinity | q->infinity]->x;
         y = rp[p->infinity | q->infinity]->y;
         z = rp[p->infinity | q->infinity]->z;
diff --git a/wolfcrypt/src/sp_x86_64.c b/wolfcrypt/src/sp_x86_64.c
index 9a42a1343e..7d869c5416 100644
--- a/wolfcrypt/src/sp_x86_64.c
+++ b/wolfcrypt/src/sp_x86_64.c
@@ -28685,8 +28685,7 @@ SP_NOINLINE static void sp_256_div2_4(sp_digit* r, sp_digit* a, sp_digit* m)
  */
 static void sp_256_proj_point_dbl_4(sp_point* r, sp_point* p, sp_digit* t)
 {
-    sp_point *rp[2];
-    sp_point tp;
+    sp_point* rp[2];
     sp_digit* t1 = t;
     sp_digit* t2 = t + 2*4;
     sp_digit* x;
@@ -28696,7 +28695,8 @@ static void sp_256_proj_point_dbl_4(sp_point* r, sp_point* p, sp_digit* t)
 
     /* When infinity don't double point passed in - constant time. */
     rp[0] = r;
-    rp[1] = &tp;
+    rp[1] = (sp_point*)t;
+    XMEMSET(rp[1], 0, sizeof(sp_point));
     x = rp[p->infinity]->x;
     y = rp[p->infinity]->y;
     z = rp[p->infinity]->z;
@@ -28760,8 +28760,7 @@ static void sp_256_proj_point_dbl_4(sp_point* r, sp_point* p, sp_digit* t)
 static void sp_256_proj_point_dbl_n_4(sp_point* r, sp_point* p, int n,
         sp_digit* t)
 {
-    sp_point *rp[2];
-    sp_point tp;
+    sp_point* rp[2];
     sp_digit* w = t;
     sp_digit* a = t + 2*4;
     sp_digit* b = t + 4*4;
@@ -28773,7 +28772,8 @@ static void sp_256_proj_point_dbl_n_4(sp_point* r, sp_point* p, int n,
     int i;
 
     rp[0] = r;
-    rp[1] = &tp;
+    rp[1] = (sp_point*)t;
+    XMEMSET(rp[1], 0, sizeof(sp_point));
     x = rp[p->infinity]->x;
     y = rp[p->infinity]->y;
     z = rp[p->infinity]->z;
@@ -28844,9 +28844,8 @@ static int sp_256_cmp_equal_4(const sp_digit* a, const sp_digit* b)
 static void sp_256_proj_point_add_4(sp_point* r, sp_point* p, sp_point* q,
         sp_digit* t)
 {
-    sp_point *ap[2];
-    sp_point *rp[2];
-    sp_point tp;
+    sp_point* ap[2];
+    sp_point* rp[2];
     sp_digit* t1 = t;
     sp_digit* t2 = t + 2*4;
     sp_digit* t3 = t + 4*4;
@@ -28873,8 +28872,8 @@ static void sp_256_proj_point_add_4(sp_point* r, sp_point* p, sp_point* q,
     }
     else {
         rp[0] = r;
-        rp[1] = &tp;
-        XMEMSET(&tp, 0, sizeof(tp));
+        rp[1] = (sp_point*)t;
+        XMEMSET(rp[1], 0, sizeof(sp_point));
         x = rp[p->infinity | q->infinity]->x;
         y = rp[p->infinity | q->infinity]->y;
         z = rp[p->infinity | q->infinity]->z;
@@ -29707,8 +29706,7 @@ static void sp_256_map_avx2_4(sp_point* r, sp_point* p, sp_digit* t)
  */
 static void sp_256_proj_point_dbl_avx2_4(sp_point* r, sp_point* p, sp_digit* t)
 {
-    sp_point *rp[2];
-    sp_point tp;
+    sp_point* rp[2];
     sp_digit* t1 = t;
     sp_digit* t2 = t + 2*4;
     sp_digit* x;
@@ -29718,7 +29716,8 @@ static void sp_256_proj_point_dbl_avx2_4(sp_point* r, sp_point* p, sp_digit* t)
 
     /* When infinity don't double point passed in - constant time. */
     rp[0] = r;
-    rp[1] = &tp;
+    rp[1] = (sp_point*)t;
+    XMEMSET(rp[1], 0, sizeof(sp_point));
     x = rp[p->infinity]->x;
     y = rp[p->infinity]->y;
     z = rp[p->infinity]->z;
@@ -29782,8 +29781,7 @@ static void sp_256_proj_point_dbl_avx2_4(sp_point* r, sp_point* p, sp_digit* t)
 static void sp_256_proj_point_dbl_n_avx2_4(sp_point* r, sp_point* p, int n,
         sp_digit* t)
 {
-    sp_point *rp[2];
-    sp_point tp;
+    sp_point* rp[2];
     sp_digit* w = t;
     sp_digit* a = t + 2*4;
     sp_digit* b = t + 4*4;
@@ -29795,7 +29793,8 @@ static void sp_256_proj_point_dbl_n_avx2_4(sp_point* r, sp_point* p, int n,
     int i;
 
     rp[0] = r;
-    rp[1] = &tp;
+    rp[1] = (sp_point*)t;
+    XMEMSET(rp[1], 0, sizeof(sp_point));
     x = rp[p->infinity]->x;
     y = rp[p->infinity]->y;
     z = rp[p->infinity]->z;
@@ -29854,9 +29853,8 @@ static void sp_256_proj_point_dbl_n_avx2_4(sp_point* r, sp_point* p, int n,
 static void sp_256_proj_point_add_avx2_4(sp_point* r, sp_point* p, sp_point* q,
         sp_digit* t)
 {
-    sp_point *ap[2];
-    sp_point *rp[2];
-    sp_point tp;
+    sp_point* ap[2];
+    sp_point* rp[2];
     sp_digit* t1 = t;
     sp_digit* t2 = t + 2*4;
     sp_digit* t3 = t + 4*4;
@@ -29883,8 +29881,8 @@ static void sp_256_proj_point_add_avx2_4(sp_point* r, sp_point* p, sp_point* q,
     }
     else {
         rp[0] = r;
-        rp[1] = &tp;
-        XMEMSET(&tp, 0, sizeof(tp));
+        rp[1] = (sp_point*)t;
+        XMEMSET(rp[1], 0, sizeof(sp_point));
         x = rp[p->infinity | q->infinity]->x;
         y = rp[p->infinity | q->infinity]->y;
         z = rp[p->infinity | q->infinity]->z;
@@ -30215,9 +30213,8 @@ typedef struct sp_table_entry {
 static void sp_256_proj_point_add_qz1_4(sp_point* r, sp_point* p,
         sp_point* q, sp_digit* t)
 {
-    sp_point *ap[2];
-    sp_point *rp[2];
-    sp_point tp;
+    sp_point* ap[2];
+    sp_point* rp[2];
     sp_digit* t1 = t;
     sp_digit* t2 = t + 2*4;
     sp_digit* t3 = t + 4*4;
@@ -30237,8 +30234,8 @@ static void sp_256_proj_point_add_qz1_4(sp_point* r, sp_point* p,
     }
     else {
         rp[0] = r;
-        rp[1] = &tp;
-        XMEMSET(&tp, 0, sizeof(tp));
+        rp[1] = (sp_point*)t;
+        XMEMSET(rp[1], 0, sizeof(sp_point));
         x = rp[p->infinity | q->infinity]->x;
         y = rp[p->infinity | q->infinity]->y;
         z = rp[p->infinity | q->infinity]->z;
@@ -30608,9 +30605,8 @@ static int sp_256_ecc_mulmod_4(sp_point* r, sp_point* g, sp_digit* k,
 static void sp_256_proj_point_add_qz1_avx2_4(sp_point* r, sp_point* p,
         sp_point* q, sp_digit* t)
 {
-    sp_point *ap[2];
-    sp_point *rp[2];
-    sp_point tp;
+    sp_point* ap[2];
+    sp_point* rp[2];
     sp_digit* t1 = t;
     sp_digit* t2 = t + 2*4;
     sp_digit* t3 = t + 4*4;
@@ -30630,8 +30626,8 @@ static void sp_256_proj_point_add_qz1_avx2_4(sp_point* r, sp_point* p,
     }
     else {
         rp[0] = r;
-        rp[1] = &tp;
-        XMEMSET(&tp, 0, sizeof(tp));
+        rp[1] = (sp_point*)t;
+        XMEMSET(rp[1], 0, sizeof(sp_point));
         x = rp[p->infinity | q->infinity]->x;
         y = rp[p->infinity | q->infinity]->y;
         z = rp[p->infinity | q->infinity]->z;