Merge bab101ce8d into 4ec3be8bfa

c/blake3_dispatch.c

@@ -89,6 +89,7 @@ static void cpuidex(uint32_t out[4], uint32_t id, uint32_t sid) {
 #endif
 
 enum cpu_feature {
+  NONE = 0,
   SSE2 = 1 << 0,
   SSSE3 = 1 << 1,
   SSE41 = 1 << 2,
@@ -265,11 +266,11 @@ void blake3_hash_many(const uint8_t *const *inputs, size_t num_inputs,
   blake3_hash_many_neon(inputs, num_inputs, blocks, key, counter,
                         increment_counter, flags, flags_start, flags_end, out);
   return;
-#endif
-
+#else
   blake3_hash_many_portable(inputs, num_inputs, blocks, key, counter,
                             increment_counter, flags, flags_start, flags_end,
                             out);
+#endif
 }
 
 // The dynamically detected SIMD degree of the current platform.

c/blake3_impl.h

@@ -279,6 +279,10 @@ void blake3_hash_many_neon(const uint8_t *const *inputs, size_t num_inputs,
                            uint64_t counter, bool increment_counter,
                            uint8_t flags, uint8_t flags_start,
                            uint8_t flags_end, uint8_t *out);
+void blake3_hash4_neon(const uint8_t *const *inputs, size_t blocks,
+                       const uint32_t key[8], uint64_t counter,
+                       bool increment_counter, uint8_t flags,
+                       uint8_t flags_start, uint8_t flags_end, uint8_t *out);
 #endif
 
 

reference_impl/reference_impl.rs

@@ -78,23 +78,14 @@ fn compress(
     block_len: u32,
     flags: u32,
 ) -> [u32; 16] {
+    let counter_low = counter as u32;
+    let counter_high = (counter >> 32) as u32;
+    #[rustfmt::skip]
     let mut state = [
-        chaining_value[0],
-        chaining_value[1],
-        chaining_value[2],
-        chaining_value[3],
-        chaining_value[4],
-        chaining_value[5],
-        chaining_value[6],
-        chaining_value[7],
-        IV[0],
-        IV[1],
-        IV[2],
-        IV[3],
-        counter as u32,
-        (counter >> 32) as u32,
-        block_len,
-        flags,
+        chaining_value[0], chaining_value[1], chaining_value[2], chaining_value[3],
+        chaining_value[4], chaining_value[5], chaining_value[6], chaining_value[7],
+        IV[0],             IV[1],             IV[2],             IV[3],
+        counter_low,       counter_high,      block_len,         flags,
     ];
     let mut block = *block_words;