implement WC_C_DYNAMIC_FALLBACK for SHA512.

wolfcrypt/src/sha512.c

@@ -96,7 +96,6 @@
     #include <wolfssl/wolfcrypt/port/nxp/se050_port.h>
 #endif
 
-
 #if defined(WOLFSSL_X86_64_BUILD) && defined(USE_INTEL_SPEEDUP)
     #if defined(__GNUC__) && ((__GNUC__ < 4) || \
                               (__GNUC__ == 4 && __GNUC_MINOR__ <= 8))
@@ -206,6 +205,17 @@
 
 #ifdef WOLFSSL_SHA512
 
+#if defined(WOLFSSL_X86_64_BUILD) && defined(USE_INTEL_SPEEDUP) && \
+    (defined(HAVE_INTEL_AVX1) || defined(HAVE_INTEL_AVX2))
+static void Sha512_SetTransform(
+#ifdef WC_C_DYNAMIC_FALLBACK
+        int *sha_method
+#else
+        void
+#endif
+    );
+#endif
+
 static int InitSha512(wc_Sha512* sha512)
 {
     if (sha512 == NULL)
@@ -224,6 +234,16 @@ static int InitSha512(wc_Sha512* sha512)
     sha512->loLen   = 0;
     sha512->hiLen   = 0;
 
+#if defined(WOLFSSL_X86_64_BUILD) && defined(USE_INTEL_SPEEDUP) && \
+    (defined(HAVE_INTEL_AVX1) || defined(HAVE_INTEL_AVX2))
+#ifdef WC_C_DYNAMIC_FALLBACK
+    sha512->sha_method = 0;
+    Sha512_SetTransform(&sha512->sha_method);
+#else
+    Sha512_SetTransform();
+#endif
+#endif
+
 #if defined(WOLFSSL_USE_ESP32_CRYPT_HASH_HW) && \
    !defined(NO_WOLFSSL_ESP32_CRYPT_HASH_SHA512)
 
@@ -265,6 +285,16 @@ static int InitSha512_224(wc_Sha512* sha512)
     sha512->loLen   = 0;
     sha512->hiLen   = 0;
 
+#if defined(WOLFSSL_X86_64_BUILD) && defined(USE_INTEL_SPEEDUP) && \
+    (defined(HAVE_INTEL_AVX1) || defined(HAVE_INTEL_AVX2))
+#ifdef WC_C_DYNAMIC_FALLBACK
+    sha512->sha_method = 0;
+    Sha512_SetTransform(&sha512->sha_method);
+#else
+    Sha512_SetTransform();
+#endif
+#endif
+
 #if defined(WOLFSSL_USE_ESP32_CRYPT_HASH_HW) && \
    !defined(NO_WOLFSSL_ESP32_CRYPT_HASH_SHA512)
     /* HW needs to be carefully initialized, taking into account soft copy.
@@ -308,6 +338,16 @@ static int InitSha512_256(wc_Sha512* sha512)
     sha512->loLen   = 0;
     sha512->hiLen   = 0;
 
+#if defined(WOLFSSL_X86_64_BUILD) && defined(USE_INTEL_SPEEDUP) && \
+    (defined(HAVE_INTEL_AVX1) || defined(HAVE_INTEL_AVX2))
+#ifdef WC_C_DYNAMIC_FALLBACK
+    sha512->sha_method = 0;
+    Sha512_SetTransform(&sha512->sha_method);
+#else
+    Sha512_SetTransform();
+#endif
+#endif
+
 #if defined(WOLFSSL_USE_ESP32_CRYPT_HASH_HW) && \
    !defined(NO_WOLFSSL_ESP32_CRYPT_HASH_SHA512)
     /* HW needs to be carefully initialized, taking into account soft copy.
@@ -425,36 +465,58 @@ static int InitSha512_256(wc_Sha512* sha512)
     }  /* extern "C" */
 #endif
 
+#if defined(WC_C_DYNAMIC_FALLBACK) && !defined(WC_NO_INTERNAL_FUNCTION_POINTERS)
+    #define WC_NO_INTERNAL_FUNCTION_POINTERS
+#endif
+
     static int _Transform_Sha512(wc_Sha512 *sha512);
-    static word32 intel_flags;
-    static int Transform_Sha512_is_vectorized = 0;
 
 #ifdef WC_NO_INTERNAL_FUNCTION_POINTERS
 
-    static enum { SHA512_UNSET, SHA512_AVX1, SHA512_AVX2, SHA512_AVX1_RORX,
+    enum sha_methods { SHA512_UNSET = 0, SHA512_AVX1, SHA512_AVX2,
-                  SHA512_AVX2_RORX, SHA512_C }
+                       SHA512_AVX1_RORX, SHA512_AVX2_RORX, SHA512_C };
-        sha_method = SHA512_UNSET;
 
-    static void Sha512_SetTransform(void)
+#ifndef WC_C_DYNAMIC_FALLBACK
+    static enum sha_methods sha_method = SHA512_UNSET;
+    static word32 intel_flags;
+#endif
+
+    static void Sha512_SetTransform(
+#ifdef WC_C_DYNAMIC_FALLBACK
+        int *sha_method
+#else
+        void
+#endif
+        )
     {
-
+#ifdef WC_C_DYNAMIC_FALLBACK
-        if (sha_method != SHA512_UNSET)
+    #define SHA_METHOD (*sha_method)
+        word32 intel_flags;
+#else
+    #define SHA_METHOD sha_method
+#endif
+        if (SHA_METHOD != SHA512_UNSET)
             return;
 
+#ifdef WC_C_DYNAMIC_FALLBACK
+        if (! CAN_SAVE_VECTOR_REGISTERS()) {
+            SHA_METHOD = SHA512_C;
+            return;
+        }
+#endif
+
         intel_flags = cpuid_get_flags();
 
     #if defined(HAVE_INTEL_AVX2)
         if (IS_INTEL_AVX2(intel_flags)) {
         #ifdef HAVE_INTEL_RORX
             if (IS_INTEL_BMI2(intel_flags)) {
-                sha_method = SHA512_AVX2_RORX;
+                SHA_METHOD = SHA512_AVX2_RORX;
-                Transform_Sha512_is_vectorized = 1;
             }
             else
         #endif
             {
-                sha_method = SHA512_AVX2;
+                SHA_METHOD = SHA512_AVX2;
-                Transform_Sha512_is_vectorized = 1;
             }
         }
         else
@@ -463,30 +525,33 @@ static int InitSha512_256(wc_Sha512* sha512)
         if (IS_INTEL_AVX1(intel_flags)) {
         #ifdef HAVE_INTEL_RORX
             if (IS_INTEL_BMI2(intel_flags)) {
-                sha_method = SHA512_AVX1_RORX;
+                SHA_METHOD = SHA512_AVX1_RORX;
-                Transform_Sha512_is_vectorized = 1;
             }
             else
         #endif
             {
-                sha_method = SHA512_AVX1;
+                SHA_METHOD = SHA512_AVX1;
-                Transform_Sha512_is_vectorized = 1;
             }
         }
         else
     #endif
         {
-            sha_method = SHA512_C;
+            SHA_METHOD = SHA512_C;
-            Transform_Sha512_is_vectorized = 0;
         }
+#undef SHA_METHOD
     }
 
     static WC_INLINE int Transform_Sha512(wc_Sha512 *sha512) {
+#ifdef WC_C_DYNAMIC_FALLBACK
+    #define SHA_METHOD (sha512->sha_method)
+#else
+    #define SHA_METHOD sha_method
+#endif
         int ret;
-        if (sha_method == SHA512_C)
+        if (SHA_METHOD == SHA512_C)
             return _Transform_Sha512(sha512);
         SAVE_VECTOR_REGISTERS(return _svr_ret;);
-        switch (sha_method) {
+        switch (SHA_METHOD) {
         case SHA512_AVX2:
             ret = Transform_Sha512_AVX2(sha512);
             break;
@@ -507,13 +572,19 @@ static int InitSha512_256(wc_Sha512* sha512)
         }
         RESTORE_VECTOR_REGISTERS();
         return ret;
+#undef SHA_METHOD
     }
 #define XTRANSFORM(...) inline_XTRANSFORM(__VA_ARGS__)
 
     static WC_INLINE int Transform_Sha512_Len(wc_Sha512 *sha512, word32 len) {
+#ifdef WC_C_DYNAMIC_FALLBACK
+    #define SHA_METHOD (sha512->sha_method)
+#else
+    #define SHA_METHOD sha_method
+#endif
         int ret;
         SAVE_VECTOR_REGISTERS(return _svr_ret;);
-        switch (sha_method) {
+        switch (SHA_METHOD) {
         case SHA512_AVX2:
             ret = Transform_Sha512_AVX2_Len(sha512, len);
             break;
@@ -534,6 +605,7 @@ static int InitSha512_256(wc_Sha512* sha512)
         }
         RESTORE_VECTOR_REGISTERS();
         return ret;
+#undef SHA_METHOD
     }
 #define XTRANSFORM_LEN(...) inline_XTRANSFORM_LEN(__VA_ARGS__)
 
@@ -542,6 +614,7 @@ static int InitSha512_256(wc_Sha512* sha512)
     static int (*Transform_Sha512_p)(wc_Sha512* sha512) = _Transform_Sha512;
     static int (*Transform_Sha512_Len_p)(wc_Sha512* sha512, word32 len) = NULL;
     static int transform_check = 0;
+    static int Transform_Sha512_is_vectorized = 0;
 
     static WC_INLINE int Transform_Sha512(wc_Sha512 *sha512) {
         int ret;
@@ -634,7 +707,7 @@ static int InitSha512_256(wc_Sha512* sha512)
 static int InitSha512_Family(wc_Sha512* sha512, void* heap, int devId,
                              int (*initfp)(wc_Sha512*))
 {
-   int ret = 0;
+    int ret = 0;
 
     if (sha512 == NULL) {
         return BAD_FUNC_ARG;
@@ -655,10 +728,6 @@ static int InitSha512_Family(wc_Sha512* sha512, void* heap, int devId,
     if (ret != 0)
         return ret;
 
-#if defined(WOLFSSL_X86_64_BUILD) && defined(USE_INTEL_SPEEDUP) && \
-    (defined(HAVE_INTEL_AVX1) || defined(HAVE_INTEL_AVX2))
-    Sha512_SetTransform();
-#endif
 #ifdef WOLFSSL_HASH_KEEP
     sha512->msg  = NULL;
     sha512->len  = 0;
@@ -898,7 +967,11 @@ static WC_INLINE int Sha512Update(wc_Sha512* sha512, const byte* data, word32 le
     #if defined(LITTLE_ENDIAN_ORDER)
         #if defined(WOLFSSL_X86_64_BUILD) && defined(USE_INTEL_SPEEDUP) && \
             (defined(HAVE_INTEL_AVX1) || defined(HAVE_INTEL_AVX2))
+            #ifdef WC_C_DYNAMIC_FALLBACK
+            if (sha512->sha_method == SHA512_C)
+            #else
             if (!IS_INTEL_AVX1(intel_flags) && !IS_INTEL_AVX2(intel_flags))
+            #endif
         #endif
             {
         #if !defined(WOLFSSL_ESP32_CRYPT) || \
@@ -934,11 +1007,14 @@ static WC_INLINE int Sha512Update(wc_Sha512* sha512, const byte* data, word32 le
 #if defined(WOLFSSL_X86_64_BUILD) && defined(USE_INTEL_SPEEDUP) && \
     (defined(HAVE_INTEL_AVX1) || defined(HAVE_INTEL_AVX2))
 
-    #ifdef WC_NO_INTERNAL_FUNCTION_POINTERS
+    #ifdef WC_C_DYNAMIC_FALLBACK
+    if (sha512->sha_method != SHA512_C)
+    #elif defined(WC_NO_INTERNAL_FUNCTION_POINTERS)
     if (sha_method != SHA512_C)
     #else
     if (Transform_Sha512_Len_p != NULL)
     #endif
+
     {
         word32 blocksLen = len & ~((word32)WC_SHA512_BLOCK_SIZE-1);
 
@@ -964,7 +1040,11 @@ static WC_INLINE int Sha512Update(wc_Sha512* sha512, const byte* data, word32 le
 
         #if defined(WOLFSSL_X86_64_BUILD) && defined(USE_INTEL_SPEEDUP) && \
             (defined(HAVE_INTEL_AVX1) || defined(HAVE_INTEL_AVX2))
+            #ifdef WC_C_DYNAMIC_FALLBACK
+            if (sha512->sha_method == SHA512_C)
+            #else
             if (!IS_INTEL_AVX1(intel_flags) && !IS_INTEL_AVX2(intel_flags))
+            #endif
             {
                 ByteReverseWords64(sha512->buffer, sha512->buffer,
                                                           WC_SHA512_BLOCK_SIZE);
@@ -1094,7 +1174,11 @@ static WC_INLINE int Sha512Final(wc_Sha512* sha512)
 #if defined(LITTLE_ENDIAN_ORDER)
     #if defined(WOLFSSL_X86_64_BUILD) && defined(USE_INTEL_SPEEDUP) && \
         (defined(HAVE_INTEL_AVX1) || defined(HAVE_INTEL_AVX2))
+        #ifdef WC_C_DYNAMIC_FALLBACK
+        if (sha512->sha_method == SHA512_C)
+        #else
         if (!IS_INTEL_AVX1(intel_flags) && !IS_INTEL_AVX2(intel_flags))
+        #endif
     #endif
         {
 
@@ -1140,7 +1224,11 @@ static WC_INLINE int Sha512Final(wc_Sha512* sha512)
 #if defined(LITTLE_ENDIAN_ORDER)
     #if defined(WOLFSSL_X86_64_BUILD) && defined(USE_INTEL_SPEEDUP) && \
         (defined(HAVE_INTEL_AVX1) || defined(HAVE_INTEL_AVX2))
+        #ifdef WC_C_DYNAMIC_FALLBACK
+        if (sha512->sha_method == SHA512_C)
+        #else
         if (!IS_INTEL_AVX1(intel_flags) && !IS_INTEL_AVX2(intel_flags))
+        #endif
     #endif
     #if !defined(WOLFSSL_ESP32_CRYPT) || \
          defined(NO_WOLFSSL_ESP32_CRYPT_HASH) || \
@@ -1159,7 +1247,11 @@ static WC_INLINE int Sha512Final(wc_Sha512* sha512)
 
 #if defined(WOLFSSL_X86_64_BUILD) && defined(USE_INTEL_SPEEDUP) && \
     (defined(HAVE_INTEL_AVX1) || defined(HAVE_INTEL_AVX2))
+    #ifdef WC_C_DYNAMIC_FALLBACK
+    if (sha512->sha_method != SHA512_C)
+    #else
     if (IS_INTEL_AVX1(intel_flags) || IS_INTEL_AVX2(intel_flags))
+    #endif
         ByteReverseWords64(&(sha512->buffer[WC_SHA512_BLOCK_SIZE / sizeof(word64) - 2]),
                            &(sha512->buffer[WC_SHA512_BLOCK_SIZE / sizeof(word64) - 2]),
                            WC_SHA512_BLOCK_SIZE - WC_SHA512_PAD_SIZE);
@@ -1362,15 +1454,14 @@ int wc_Sha512Transform(wc_Sha512* sha, const unsigned char* data)
         return MEMORY_E;
 #endif
 
-#if defined(WOLFSSL_X86_64_BUILD) && defined(USE_INTEL_SPEEDUP) && \
-    (defined(HAVE_INTEL_AVX1) || defined(HAVE_INTEL_AVX2))
-    Sha512_SetTransform();
-#endif
-
 #if defined(LITTLE_ENDIAN_ORDER)
 #if defined(WOLFSSL_X86_64_BUILD) && defined(USE_INTEL_SPEEDUP) && \
     (defined(HAVE_INTEL_AVX1) || defined(HAVE_INTEL_AVX2))
+    #ifdef WC_C_DYNAMIC_FALLBACK
+    if (sha->sha_method == SHA512_C)
+    #else
     if (!IS_INTEL_AVX1(intel_flags) && !IS_INTEL_AVX2(intel_flags))
+    #endif
 #endif
     {
         ByteReverseWords64((word64*)data, (word64*)data,
@@ -1463,6 +1554,16 @@ static int InitSha384(wc_Sha384* sha384)
     sha384->loLen   = 0;
     sha384->hiLen   = 0;
 
+#if defined(WOLFSSL_X86_64_BUILD) && defined(USE_INTEL_SPEEDUP) && \
+    (defined(HAVE_INTEL_AVX1) || defined(HAVE_INTEL_AVX2))
+#ifdef WC_C_DYNAMIC_FALLBACK
+    sha384->sha_method = 0;
+    Sha512_SetTransform(&sha384->sha_method);
+#else
+    Sha512_SetTransform();
+#endif
+#endif
+
 #if defined(WOLFSSL_USE_ESP32_CRYPT_HASH_HW)  && \
    !defined(NO_WOLFSSL_ESP32_CRYPT_HASH_SHA384)
     /* HW needs to be carefully initialized, taking into account soft copy.
@@ -1605,11 +1706,6 @@ int wc_InitSha384_ex(wc_Sha384* sha384, void* heap, int devId)
         return ret;
     }
 
-#if defined(WOLFSSL_X86_64_BUILD) && defined(USE_INTEL_SPEEDUP) && \
-    (defined(HAVE_INTEL_AVX1) || defined(HAVE_INTEL_AVX2))
-    Sha512_SetTransform();
-#endif
-
 #if defined(WOLFSSL_ASYNC_CRYPT) && defined(WC_ASYNC_ENABLE_SHA384)
     ret = wolfAsync_DevCtxInit(&sha384->asyncDev, WOLFSSL_ASYNC_MARKER_SHA384,
                                                            sha384->heap, devId);

wolfssl/wolfcrypt/sha512.h

@@ -151,6 +151,9 @@ struct wc_Sha512 {
 #ifdef USE_INTEL_SPEEDUP
     const byte* data;
 #endif
+#ifdef WC_C_DYNAMIC_FALLBACK
+    int sha_method;
+#endif
 #ifdef WOLFSSL_ASYNC_CRYPT
     WC_ASYNC_DEV asyncDev;
 #endif /* WOLFSSL_ASYNC_CRYPT */