Improve some calls to %mstore_rlp (#1452)

* Improve some calls to mstore_rlp

* Remove comment

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

@@ -409,8 +409,7 @@
     %mstore_u32
 %endmacro
 
-// Store a single byte to @SEGMENT_RLP_RAW.
-%macro mstore_rlp
+%macro swap_mstore
     // stack: addr, value
     SWAP1
     MSTORE_GENERAL

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

@@ -151,8 +151,8 @@ global encode_node_branch:
 
     // No value; append the empty string (0x80).
     // stack: value_ptr, rlp_pos', rlp_start, encode_value, cur_len, retdest
-    %stack (value_ptr, rlp_pos, rlp_start, encode_value) -> (rlp_pos, 0x80, rlp_pos, rlp_start)
-    %mstore_rlp
+    %stack (value_ptr, rlp_pos, rlp_start, encode_value) -> (0x80, rlp_pos, rlp_pos, rlp_start)
+    MSTORE_GENERAL
     // stack: rlp_pos', rlp_start, cur_len, retdest
     %increment
     // stack: rlp_pos'', rlp_start, cur_len, retdest
@@ -192,9 +192,9 @@ encode_node_branch_prepend_prefix:
     SWAP1 DUP1 %sub_const(32) %jumpi(%%unpack)
     // Otherwise, result is a hash, and we need to add the prefix 0x80 + 32 = 160.
     // stack: result_len, result, cur_len, rlp_pos, rlp_start, node_payload_ptr, encode_value, retdest
+    DUP4 // rlp_pos
     PUSH 160
-    DUP5 // rlp_pos
-    %mstore_rlp
+    MSTORE_GENERAL
     SWAP3 %increment SWAP3 // rlp_pos += 1
 %%unpack:
     %stack (result_len, result, cur_len, rlp_pos, rlp_start, node_payload_ptr, encode_value, retdest)
@@ -233,9 +233,9 @@ encode_node_extension_after_hex_prefix:
     // If result_len != 32, result is raw RLP, with an appropriate RLP prefix already.
     DUP4 %sub_const(32) %jumpi(encode_node_extension_unpack)
     // Otherwise, result is a hash, and we need to add the prefix 0x80 + 32 = 160.
+    DUP1 // rlp_pos
     PUSH 160
-    DUP2 // rlp_pos
-    %mstore_rlp
+    MSTORE_GENERAL
     %increment // rlp_pos += 1
 encode_node_extension_unpack:
     %stack (rlp_pos, rlp_start, result, result_len, node_payload_ptr, cur_len)

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

@@ -326,8 +326,8 @@ encode_nonzero_receipt_type:
     // stack: rlp_receipt_len, txn_type, rlp_addr, value_ptr, retdest
     DUP3 %encode_rlp_multi_byte_string_prefix
     // stack: rlp_addr, txn_type, old_rlp_addr, value_ptr, retdest
-    DUP2 DUP2
-    %mstore_rlp
+    DUP1 DUP3
+    MSTORE_GENERAL
     %increment
     // stack: rlp_addr, txn_type, old_rlp_addr, value_ptr, retdest
     %stack (rlp_addr, txn_type, old_rlp_addr, value_ptr, retdest) -> (rlp_addr, value_ptr, retdest)

evm/src/cpu/kernel/asm/mpt/hex_prefix.asm

@@ -27,7 +27,7 @@ first_byte:
     // stack: rlp_addr, num_nibbles, packed_nibbles, terminated, retdest
     // get the first nibble, if num_nibbles is odd, or zero otherwise
     SWAP2
-    // stack: packed_nibbles, num_nibbbles, rlp_addr, terminated, retdest
+    // stack: packed_nibbles, num_nibbles, rlp_addr, terminated, retdest
     DUP2 DUP1
     %mod_const(2)
     // stack: parity, num_nibbles, packed_nibbles, num_nibbles, rlp_addr, terminated, retdest
@@ -50,7 +50,7 @@ first_byte:
     ADD
     // stack: first_byte, rlp_addr, retdest
     DUP2
-    %mstore_rlp
+    %swap_mstore
     %increment
     // stack: rlp_addr', retdest
     SWAP1
@@ -88,7 +88,7 @@ rlp_header_medium:
     // stack: hp_len, rlp_addr, num_nibbles, packed_nibbles, terminated, retdest
     %add_const(0x80) // value = 0x80 + hp_len
     DUP2
-    %mstore_rlp
+    %swap_mstore
     // stack: rlp_addr, num_nibbles, packed_nibbles, terminated, retdest
     // rlp_addr += 1
     %increment
@@ -108,14 +108,14 @@ rlp_header_large:
     // In practice hex-prefix length will never exceed 256, so the length of the
     // length will always be 1 byte in this case.
 
+    DUP2 // rlp_addr
     PUSH 0xb8 // value = 0xb7 + len_of_len = 0xb8
-    DUP3
-    %mstore_rlp
+    MSTORE_GENERAL
     // stack: rlp_addr, value, hp_len, i, rlp_addr, num_nibbles, packed_nibbles, terminated, retdest
 
     // stack: hp_len, rlp_addr, num_nibbles, packed_nibbles, terminated, retdest
     DUP2 %increment
-    %mstore_rlp
+    %swap_mstore
 
     // stack: rlp_addr, num_nibbles, packed_nibbles, terminated, retdest
     // rlp_addr += 2

evm/src/cpu/kernel/asm/mpt/util.asm

@@ -11,9 +11,9 @@
 %endmacro
 
 %macro initialize_rlp_segment
-    PUSH 0x80
     PUSH @ENCODED_EMPTY_NODE_POS
-    %mstore_rlp
+    PUSH 0x80
+    MSTORE_GENERAL
 %endmacro
 
 %macro alloc_rlp_block

evm/src/cpu/kernel/asm/rlp/encode.asm

@@ -29,12 +29,11 @@ global encode_rlp_256:
 // RLP-encode a fixed-length string with the given byte length. Assumes string < 2^(8 * len).
 global encode_rlp_fixed:
     // stack: len, rlp_addr, string, retdest
-    DUP1
+    DUP2
+    DUP2
     %add_const(0x80)
-    // stack: first_byte, len, rlp_addr, string, retdest
-    DUP3
-    // stack: rlp_addr, first_byte, len, rlp_addr, string, retdest
-    %mstore_rlp
+    // stack: first_byte, rlp_addr, len, rlp_addr, string, retdest
+    MSTORE_GENERAL
     // stack: len, rlp_addr, string, retdest
     SWAP1
     %increment // increment rlp_addr
@@ -51,19 +50,17 @@ encode_rlp_fixed_finish:
 // I.e. writes encode(encode(string). Assumes string < 2^(8 * len).
 global doubly_encode_rlp_fixed:
     // stack: len, rlp_addr, string, retdest
-    DUP1
+    DUP2
+    DUP2
     %add_const(0x81)
-    // stack: first_byte, len, rlp_addr, string, retdest
-    DUP3
-    // stack: rlp_addr, first_byte, len, rlp_addr, string, retdest
-    %mstore_rlp
+    // stack: first_byte, rlp_addr, len, rlp_addr, string, retdest
+    MSTORE_GENERAL
     // stack: len, rlp_addr, string, retdest
-    DUP1
+    DUP2 %increment
+    DUP2
     %add_const(0x80)
-    // stack: second_byte, len, original_rlp_addr, string, retdest
-    DUP3 %increment
-    // stack: rlp_addr', second_byte, len, rlp_addr, string, retdest
-    %mstore_rlp
+    // stack: second_byte, rlp_addr', len, original_rlp_addr, string, retdest
+    MSTORE_GENERAL
     // stack: len, rlp_addr, string, retdest
     SWAP1
     %add_const(2) // advance past the two prefix bytes
@@ -87,11 +84,10 @@ global encode_rlp_multi_byte_string_prefix:
     %jumpi(encode_rlp_multi_byte_string_prefix_large)
     // Medium case; prefix is 0x80 + str_len.
     // stack: rlp_addr, str_len, retdest
-    SWAP1 %add_const(0x80)
-    // stack: prefix, rlp_addr, retdest
-    DUP2
-    // stack: rlp_addr, prefix, rlp_addr, retdest
-    %mstore_rlp
+    DUP1
+    SWAP2 %add_const(0x80)
+    // stack: prefix, rlp_addr, rlp_addr, retdest
+    MSTORE_GENERAL
     // stack: rlp_addr, retdest
     %increment
     // stack: rlp_addr', retdest
@@ -104,12 +100,11 @@ encode_rlp_multi_byte_string_prefix_large:
     %num_bytes
     // stack: len_of_len, rlp_addr, str_len, retdest
     SWAP1
-    DUP2 // len_of_len
+    DUP1 // rlp_addr
+    DUP3 // len_of_len
     %add_const(0xb7)
-    // stack: first_byte, rlp_addr, len_of_len, str_len, retdest
-    DUP2
-    // stack: rlp_addr, first_byte, rlp_addr, len_of_len, str_len, retdest
-    %mstore_rlp
+    // stack: first_byte, rlp_addr, rlp_addr, len_of_len, str_len, retdest
+    MSTORE_GENERAL
     // stack: rlp_addr, len_of_len, str_len, retdest
     %increment
     // stack: rlp_addr', len_of_len, str_len, retdest
@@ -132,12 +127,11 @@ global encode_rlp_list_prefix:
     %jumpi(encode_rlp_list_prefix_large)
     // Small case: prefix is just 0xc0 + length.
     // stack: rlp_addr, payload_len, retdest
-    SWAP1
+    DUP1
+    SWAP2
     %add_const(0xc0)
-    // stack: prefix, rlp_addr, retdest
-    DUP2
-    // stack: rlp_addr, prefix, rlp_addr, retdest
-    %mstore_rlp
+    // stack: prefix, rlp_addr, rlp_addr, retdest
+    MSTORE_GENERAL
     // stack: rlp_addr, retdest
     %increment
     SWAP1
@@ -147,10 +141,10 @@ encode_rlp_list_prefix_large:
     // stack: rlp_addr, payload_len, retdest
     DUP2 %num_bytes
     // stack: len_of_len, rlp_addr, payload_len, retdest
-    DUP1 %add_const(0xf7)
-    // stack: first_byte, len_of_len, rlp_addr, payload_len, retdest
-    DUP3 // rlp_addr
-    %mstore_rlp
+    DUP2
+    DUP2 %add_const(0xf7)
+    // stack: first_byte, rlp_addr, len_of_len, rlp_addr, payload_len, retdest
+    MSTORE_GENERAL
     // stack: len_of_len, rlp_addr, payload_len, retdest
     SWAP1 %increment
     // stack: rlp_addr', len_of_len, payload_len, retdest
@@ -184,10 +178,10 @@ global prepend_rlp_list_prefix:
 
     // If we got here, we have a small list, so we prepend 0xc0 + len at rlp_address 8.
     // stack: payload_len, end_rlp_addr, start_rlp_addr, retdest
-    DUP1 %add_const(0xc0)
-    // stack: prefix_byte, payload_len, end_rlp_addr, start_rlp_addr, retdest
-    DUP4 %decrement // offset of prefix
-    %mstore_rlp
+    DUP3 %decrement // offset of prefix
+    DUP2 %add_const(0xc0)
+    // stack: prefix_byte, start_rlp_addr-1, payload_len, end_rlp_addr, start_rlp_addr, retdest
+    MSTORE_GENERAL
     // stack: payload_len, end_rlp_addr, start_rlp_addr, retdest
     %increment
     // stack: rlp_len, end_rlp_addr, start_rlp_addr, retdest
@@ -207,7 +201,7 @@ prepend_rlp_list_prefix_big:
     DUP5 %decrement // start_rlp_addr - 1
     SUB
     // stack: prefix_start_rlp_addr, len_of_len, payload_len, end_rlp_addr, start_rlp_addr, retdest
-    DUP2 %add_const(0xf7) DUP2 %mstore_rlp // rlp[prefix_start_rlp_addr] = 0xf7 + len_of_len
+    DUP2 %add_const(0xf7) DUP2 %swap_mstore // rlp[prefix_start_rlp_addr] = 0xf7 + len_of_len
     // stack: prefix_start_rlp_addr, len_of_len, payload_len, end_rlp_addr, start_rlp_addr, retdest
     DUP1 %increment // start_len_rlp_addr = prefix_start_rlp_addr + 1
     %stack (start_len_rlp_addr, prefix_start_rlp_addr, len_of_len, payload_len, end_rlp_addr, start_rlp_addr, retdest)