Do not pack blake2_state structs

While the param structs need to be packed, because they're compressed as
a byte stream, the state structs do not have this requirement. As such,
the compiler likely can emit more efficient code in certain
circumstances when it can apply the ordinary struct alignment rules.

1 files changed, 32 insertions, 31 deletions

diff --git a/ref/blake2.h b/ref/blake2.h
index cc625cb..abd1f35 100644
--- a/ref/blake2.h
+++ b/ref/blake2.h
@@ -39,22 +39,6 @@ extern "C" {
     BLAKE2B_PERSONALBYTES = 16
   };
 
-#pragma pack(push, 1)
-  typedef struct __blake2s_param
-  {
-    uint8_t  digest_length; // 1
-    uint8_t  key_length;    // 2
-    uint8_t  fanout;        // 3
-    uint8_t  depth;         // 4
-    uint32_t leaf_length;   // 8
-    uint8_t  node_offset[6];// 14
-    uint8_t  node_depth;    // 15
-    uint8_t  inner_length;  // 16
-    // uint8_t  reserved[0];
-    uint8_t  salt[BLAKE2S_SALTBYTES]; // 24
-    uint8_t  personal[BLAKE2S_PERSONALBYTES];  // 32
-  } blake2s_param;
-
   typedef struct __blake2s_state
   {
     uint32_t h[8];
@@ -65,21 +49,6 @@ extern "C" {
     uint8_t  last_node;
   } blake2s_state;
 
-  typedef struct __blake2b_param
-  {
-    uint8_t  digest_length; // 1
-    uint8_t  key_length;    // 2
-    uint8_t  fanout;        // 3
-    uint8_t  depth;         // 4
-    uint32_t leaf_length;   // 8
-    uint64_t node_offset;   // 16
-    uint8_t  node_depth;    // 17
-    uint8_t  inner_length;  // 18
-    uint8_t  reserved[14];  // 32
-    uint8_t  salt[BLAKE2B_SALTBYTES]; // 48
-    uint8_t  personal[BLAKE2B_PERSONALBYTES];  // 64
-  } blake2b_param;
-
   typedef struct __blake2b_state
   {
     uint64_t h[8];
@@ -105,6 +74,38 @@ extern "C" {
     uint8_t buf[4 * BLAKE2B_BLOCKBYTES];
     size_t  buflen;
   } blake2bp_state;
+
+
+#pragma pack(push, 1)
+  typedef struct __blake2s_param
+  {
+    uint8_t  digest_length; // 1
+    uint8_t  key_length;    // 2
+    uint8_t  fanout;        // 3
+    uint8_t  depth;         // 4
+    uint32_t leaf_length;   // 8
+    uint8_t  node_offset[6];// 14
+    uint8_t  node_depth;    // 15
+    uint8_t  inner_length;  // 16
+    // uint8_t  reserved[0];
+    uint8_t  salt[BLAKE2S_SALTBYTES]; // 24
+    uint8_t  personal[BLAKE2S_PERSONALBYTES];  // 32
+  } blake2s_param;
+
+  typedef struct __blake2b_param
+  {
+    uint8_t  digest_length; // 1
+    uint8_t  key_length;    // 2
+    uint8_t  fanout;        // 3
+    uint8_t  depth;         // 4
+    uint32_t leaf_length;   // 8
+    uint64_t node_offset;   // 16
+    uint8_t  node_depth;    // 17
+    uint8_t  inner_length;  // 18
+    uint8_t  reserved[14];  // 32
+    uint8_t  salt[BLAKE2B_SALTBYTES]; // 48
+    uint8_t  personal[BLAKE2B_PERSONALBYTES];  // 64
+  } blake2b_param;
 #pragma pack(pop)
 
   // Streaming API