Add macro for copying sequences of bytes

evm/src/cpu/kernel/asm/memory/memcpy.asm

@@ -53,3 +53,74 @@ memcpy_finish:
     %jump(memcpy)
 %%after:
 %endmacro
+
+// Similar logic than memcpy, but optimized for copying sequences of bytes.
+global memcpy_bytes:
+    // stack: DST, SRC, count, retdest
+    DUP7
+    // stack: count, DST, SRC, count, retdest
+    %lt_const(0x20)
+    // stack: count < 32, DST, SRC, count, retdest
+    %jumpi(memcpy_bytes_finish)
+    
+    // We will pack 32 bytes into a U256 from the source, and then unpack it at the destination.
+    // Copy the next chunk of bytes.
+    PUSH 32
+    DUP1
+    DUP8
+    DUP8
+    DUP8
+    // stack: SRC, 32, 32, DST, SRC, count, retdest
+    MLOAD_32BYTES
+    // stack: value, 32, DST, SRC, count, retdest
+    DUP5
+    DUP5
+    DUP5
+    // stack: DST, value, 32, DST, SRC, count, retdest
+    MSTORE_32BYTES
+    // stack: DST, SRC, count, retdest
+
+    // Increment dst_addr by 32.
+    SWAP2
+    %add_const(0x20)
+    SWAP2
+    // Increment src_addr by 32.
+    SWAP5
+    %add_const(0x20)
+    SWAP5
+    // Decrement count by 32.
+    SWAP6
+    %sub_const(0x20)
+    SWAP6
+
+    // Continue the loop.
+    %jump(memcpy_bytes)
+
+memcpy_bytes_finish:
+    // stack: DST, SRC, count, retdest
+
+    // Copy the last chunk of `count` bytes.
+    DUP7
+    DUP1
+    DUP8
+    DUP8
+    DUP8
+    // stack: SRC, count, count, DST, SRC, count, retdest
+    MLOAD_32BYTES
+    // stack: value, count, DST, SRC, count, retdest
+    DUP5
+    DUP5
+    DUP5
+    // stack: DST, value, count, DST, SRC, count, retdest
+    MSTORE_32BYTES
+    // stack: DST, SRC, count, retdest
+
+    %pop7
+    // stack: retdest
+    JUMP
+
+%macro memcpy_bytes
+    %stack (dst: 3, src: 3, count) -> (dst, src, count, %%after)
+    %jump(memcpy_bytes)
+%%after:
+%endmacro

evm/src/cpu/kernel/asm/mpt/hash/hash_trie_specific.asm

@@ -116,7 +116,7 @@ global encode_txn:
         0, @SEGMENT_TRIE_DATA, txn_rlp_ptr, // src addr. Kernel has context 0
         txn_rlp_len, // mcpy len
         txn_rlp_len, rlp_pos)
-    %memcpy
+    %memcpy_bytes
     ADD
     // stack new_rlp_pos, retdest
     SWAP1

evm/src/cpu/kernel/asm/transactions/router.asm

@@ -57,7 +57,7 @@ global update_txn_trie:
         0, @SEGMENT_RLP_RAW, 0, // src addr. Kernel has context 0
         txn_rlp_len, // mcpy len
         txn_rlp_len, rlp_start, txn_counter, num_nibbles, value_ptr)
-    %memcpy
+    %memcpy_bytes
     ADD
     %set_trie_data_size
     // stack: txn_counter, num_nibbles, value_ptr, retdest