From ba1cfccb66d38f15cb044207b40c31bbbdda5ace Mon Sep 17 00:00:00 2001
From: Mike <tankf33der@disroot.org>
Date: Sun, 1 Feb 2026 10:20:32 +0200
Subject: [PATCH] crypto.blake3: add @[direct_array_access] and improve f()
 performance (#26480)

---
 vlib/crypto/blake3/blake3.v               |  5 +++--
 vlib/crypto/blake3/blake3_block_generic.v | 23 +++++++++++------------
 vlib/crypto/blake3/blake3_chunk.v         |  3 ++-
 3 files changed, 16 insertions(+), 15 deletions(-)

diff --git a/vlib/crypto/blake3/blake3.v b/vlib/crypto/blake3/blake3.v
index cfd66551b..4f7f3994f 100644
--- a/vlib/crypto/blake3/blake3.v
+++ b/vlib/crypto/blake3/blake3.v
@@ -137,7 +137,7 @@ pub fn Digest.new_hash() !Digest {
 // Digest.new_keyed_hash initializes a Digest structure for a Blake3 keyed hash
 pub fn Digest.new_keyed_hash(key []u8) !Digest {
 	// treat the key bytes as little endian u32 values
-	mut key_words := []u32{len: 8, cap: 8}
+	mut key_words := []u32{len: 8}
 	for i in 0 .. 8 {
 		key_words[i] = binary.little_endian_u32_at(key, i * 4)
 	}
@@ -153,7 +153,7 @@ pub fn Digest.new_derive_key_hash(context []u8) !Digest {
 	context_key := context_digest.checksum_internal(key_length)
 
 	// treat the context key bytes as little endian u32 values
-	mut key_words := []u32{len: 8, cap: 8}
+	mut key_words := []u32{len: 8}
 	for i in 0 .. 8 {
 		key_words[i] = binary.little_endian_u32_at(context_key, i * 4)
 	}
@@ -309,6 +309,7 @@ fn root_output_bytes(state HashState, size u64) []u8 {
 	return output
 }
 
+@[direct_array_access]
 fn (mut d Digest) add_node(node Node, level u8) {
 	// if we are above the highst level,
 	// just add the node at the top
diff --git a/vlib/crypto/blake3/blake3_block_generic.v b/vlib/crypto/blake3/blake3_block_generic.v
index d10782c2e..6e7ae0bdb 100644
--- a/vlib/crypto/blake3/blake3_block_generic.v
+++ b/vlib/crypto/blake3/blake3_block_generic.v
@@ -8,10 +8,11 @@
 
 module blake3
 
+import arrays
 import math.bits
 
 // mixing function g
-@[inline]
+@[direct_array_access; inline]
 fn g(mut v []u32, a u8, b u8, c u8, d u8, x u32, y u32) {
 	v[a] = v[a] + v[b] + x
 	v[d] = bits.rotate_left_32((v[d] ^ v[a]), nr1)
@@ -24,7 +25,7 @@ fn g(mut v []u32, a u8, b u8, c u8, d u8, x u32, y u32) {
 }
 
 // one complete mixing round with the function g
-@[inline]
+@[direct_array_access; inline]
 fn mixing_round(mut v []u32, m []u32, s []u8) {
 	g(mut v, 0, 4, 8, 12, m[s[0]], m[s[1]])
 	g(mut v, 1, 5, 9, 13, m[s[2]], m[s[3]])
@@ -38,19 +39,17 @@ fn mixing_round(mut v []u32, m []u32, s []u8) {
 }
 
 // compression function f
+@[direct_array_access]
 fn f(h []u32, m []u32, counter u64, input_bytes u32, flags u32) []u32 {
-	mut v := []u32{len: 0, cap: 16}
+	mut v := []u32{len: 16}
 
 	// initialize the working vector
-	v << h[..8]
-	v << iv[..4]
-
-	v << u32(counter & 0x00000000ffffffff)
-	v << u32(counter >> 32)
-
-	v << input_bytes
-
-	v << flags
+	arrays.copy[u32](mut v, h[..8])
+	arrays.copy[u32](mut v[8..], iv[..4])
+	v[12] = u32(counter)
+	v[13] = u32(counter >> 32)
+	v[14] = input_bytes
+	v[15] = flags
 
 	// go 7 rounds of cryptographic mixing
 	//
diff --git a/vlib/crypto/blake3/blake3_chunk.v b/vlib/crypto/blake3/blake3_chunk.v
index a4aecbecc..c534e9803 100644
--- a/vlib/crypto/blake3/blake3_chunk.v
+++ b/vlib/crypto/blake3/blake3_chunk.v
@@ -50,6 +50,7 @@ fn (c Chunk) str() string {
 //
 // As a potential speed up, we could try spawning this function
 // in a concurrent task and see if it is worth the overhead.
+@[direct_array_access]
 fn (mut c Chunk) process_input(input []u8, key_words []u32, counter u64, flags u32, root bool) []u32 {
 	mut remaining_input := unsafe { input[..] }
 
@@ -63,7 +64,7 @@ fn (mut c Chunk) process_input(input []u8, key_words []u32, counter u64, flags u
 
 	c.chunk_number = counter
 	c.chaining_value = key_words.clone()
-	c.block_words = []u32{len: 16, cap: 16, init: 0}
+	c.block_words = []u32{len: 16}
 
 	for i in 0 .. 16 {
 		c.block_len = u32(block_size)
-- 
2.39.5