Merge 36ecfb829b into 4ec3be8bfa

c/blake3_avx2_x86-64_unix.S

@@ -1,3 +1,4 @@
+#if defined(__x86_64__)
 #if defined(__ELF__) && defined(__linux__)
 .section .note.GNU-stack,"",%progbits
 #endif
@@ -1813,3 +1814,4 @@ CMP_MSB_MASK:
 BLAKE3_IV:
         .long  0x6A09E667, 0xBB67AE85, 0x3C6EF372, 0xA54FF53A
 
+#endif

c/blake3_avx512_x86-64_unix.S

@@ -1,3 +1,4 @@
+#if defined(__x86_64__)
 #if defined(__ELF__) && defined(__linux__)
 .section .note.GNU-stack,"",%progbits
 #endif
@@ -2583,3 +2584,5 @@ BLAKE3_IV_2:
         .long   0x3C6EF372
 BLAKE3_IV_3:
         .long   0xA54FF53A
+
+#endif

c/blake3_impl.h

@@ -38,7 +38,7 @@ enum blake3_flags {
 #define IS_X86_32
 #endif
 
-#if defined(__aarch64__) || defined(_M_ARM64)
+#if defined(__aarch64__) || defined(_M_ARM64) || defined(__arm64__)
 #define IS_AARCH64
 #endif
 

c/blake3_neon.c

@@ -1,4 +1,5 @@
 #include "blake3_impl.h"
+#if BLAKE3_USE_NEON
 
 #include <arm_neon.h>
 
@@ -364,3 +365,5 @@ void blake3_hash_many_neon(const uint8_t *const *inputs, size_t num_inputs,
     out = &out[BLAKE3_OUT_LEN];
   }
 }
+
+#endif

c/blake3_sse2_x86-64_unix.S

@@ -1,3 +1,4 @@
+#if defined(__x86_64__)
 #if defined(__ELF__) && defined(__linux__)
 .section .note.GNU-stack,"",%progbits
 #endif
@@ -2289,3 +2290,5 @@ PBLENDW_0x3F_MASK:
 	.long  0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0x00000000
 PBLENDW_0xC0_MASK:
 	.long  0x00000000, 0x00000000, 0x00000000, 0xFFFFFFFF
+
+#endif

c/blake3_sse41_x86-64_unix.S

@@ -1,3 +1,4 @@
+#if defined(__x86_64__)
 #if defined(__ELF__) && defined(__linux__)
 .section .note.GNU-stack,"",%progbits
 #endif
@@ -2026,3 +2027,4 @@ BLAKE3_BLOCK_LEN:
 	.long  64, 64, 64, 64
 CMP_MSB_MASK:
 	.long  0x80000000, 0x80000000, 0x80000000, 0x80000000
+#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;