Remove some CPU cycles (#1469)

* Amortize mload_packing

* Reduce stack overhead

* Amortize mstore_unpacking

* Speed-up stack operation in hash.asm

* Misc

* Small tweaks

* Misc small optims

* Fix comments

* Fix main access to withdrawals

* Fix stack description

* minor: rename label

* Comments

---------

Co-authored-by: Linda Guiga <lindaguiga3@gmail.com>

evm/src/cpu/kernel/asm/account_code.asm

@@ -2,13 +2,13 @@ global sys_extcodehash:
     // stack: kexit_info, address
     SWAP1 %u256_to_addr
     // stack: address, kexit_info
-    DUP1 %insert_accessed_addresses
-    // stack: cold_access, address, kexit_info
+    SWAP1
+    DUP2 %insert_accessed_addresses
+    // stack: cold_access, kexit_info, address
     PUSH @GAS_COLDACCOUNTACCESS_MINUS_WARMACCESS
     MUL
     PUSH @GAS_WARMACCESS
     ADD
-    %stack (gas, address, kexit_info) -> (gas, kexit_info, address)
     %charge_gas
     // stack: kexit_info, address
 
@@ -57,13 +57,13 @@ global sys_extcodesize:
     // stack: kexit_info, address
     SWAP1 %u256_to_addr
     // stack: address, kexit_info
-    DUP1 %insert_accessed_addresses
-    // stack: cold_access, address, kexit_info
+    SWAP1
+    DUP2 %insert_accessed_addresses
+    // stack: cold_access, kexit_info, address
     PUSH @GAS_COLDACCOUNTACCESS_MINUS_WARMACCESS
     MUL
     PUSH @GAS_WARMACCESS
     ADD
-    %stack (gas, address, kexit_info) -> (gas, kexit_info, address)
     %charge_gas
     // stack: kexit_info, address
 

evm/src/cpu/kernel/asm/balance.asm

@@ -2,13 +2,13 @@ global sys_balance:
     // stack: kexit_info, address
     SWAP1 %u256_to_addr
     // stack: address, kexit_info
-    DUP1 %insert_accessed_addresses
-    // stack: cold_access, address, kexit_info
+    SWAP1
+    DUP2 %insert_accessed_addresses
+    // stack: cold_access, kexit_info, address
     PUSH @GAS_COLDACCOUNTACCESS_MINUS_WARMACCESS
     MUL
     PUSH @GAS_WARMACCESS
     ADD
-    %stack (gas, address, kexit_info) -> (gas, kexit_info, address)
     %charge_gas
     // stack: kexit_info, address
 

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

@@ -38,20 +38,17 @@ global get_create_address:
 global get_create2_address:
     // stack: sender, code_hash, salt, retdest
     PUSH 0xff PUSH 0 %mstore_kernel_general
-    %stack (sender, code_hash, salt, retdest) -> (@SEGMENT_KERNEL_GENERAL, 1, sender, 20, get_create2_address_contd, salt, code_hash, retdest)
+    %stack (sender, code_hash, salt, retdest) -> (@SEGMENT_KERNEL_GENERAL, 1, sender, salt, code_hash, retdest)
     ADD
-    %jump(mstore_unpacking)
-get_create2_address_contd:
+    MSTORE_32BYTES_20
     POP
-    %stack (salt, code_hash, retdest) -> (@SEGMENT_KERNEL_GENERAL, 21, salt, 32, get_create2_address_contd2, code_hash, retdest)
+    %stack (salt, code_hash, retdest) -> (@SEGMENT_KERNEL_GENERAL, 21, salt, code_hash, retdest)
     ADD
-    %jump(mstore_unpacking)
-get_create2_address_contd2:
+    MSTORE_32BYTES_32
     POP
-    %stack (code_hash, retdest) -> (@SEGMENT_KERNEL_GENERAL, 53, code_hash, 32, get_create2_address_finish, retdest)
+    %stack (code_hash, retdest) -> (@SEGMENT_KERNEL_GENERAL, 53, code_hash, retdest)
     ADD
-    %jump(mstore_unpacking)
-get_create2_address_finish:
+    MSTORE_32BYTES_32
     POP
     %stack (retdest) -> (@SEGMENT_KERNEL_GENERAL, 85, retdest) // offset == context == 0
     // addr, len, retdest

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

@@ -42,7 +42,7 @@ continue:
 proof_ok:
     // stack: i, ctx, final_pos, retdest
     // We already know final_pos is a jumpdest
-    %stack (i, ctx, final_pos) -> (ctx, @SEGMENT_JUMPDEST_BITS, i)
+    %stack (i, ctx, final_pos) -> (ctx, @SEGMENT_JUMPDEST_BITS, final_pos)
     %build_address
     PUSH 1
     MSTORE_GENERAL
@@ -145,7 +145,7 @@ global write_table_if_jumpdest:
         (proof_prefix_addr, ctx) ->
         (ctx, proof_prefix_addr, 32, proof_prefix_addr, ctx)
     ADD // combine context and offset to make an address (SEGMENT_CODE == 0)
-    %mload_packing
+    MLOAD_32BYTES
     // packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
     DUP1 %shl_const(1)
     DUP2 %shl_const(2)

evm/src/cpu/kernel/asm/core/precompiles/blake2_f.asm

@@ -105,7 +105,7 @@ global precompile_blake2_f:
     GET_CONTEXT
     // stack: ctx, @SEGMENT_CALLDATA, 4, h_0..h_7, m_0..m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
     %build_address_no_offset
-    %mload_packing
+    MLOAD_32BYTES
     // stack: rounds, h_0..h_7, m_0..m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
     
     DUP1

evm/src/cpu/kernel/asm/core/precompiles/bn_add.asm

@@ -14,32 +14,32 @@ global precompile_bn_add:
 
     %charge_gas_const(@BN_ADD_GAS)
 
-    // Load x0, y0, x1, y1 from the call data using `mload_packing`.
+    // Load x0, y0, x1, y1 from the call data using `MLOAD_32BYTES`.
     PUSH bn_add_return
     // stack: bn_add_return, kexit_info
     %stack () -> (@SEGMENT_CALLDATA, 96, 32)
     GET_CONTEXT
     // stack: ctx, @SEGMENT_CALLDATA, 96, 32, bn_add_return, kexit_info
     %build_address
-    %mload_packing
+    MLOAD_32BYTES
     // stack: y1, bn_add_return, kexit_info
     %stack () -> (@SEGMENT_CALLDATA, 64, 32)
     GET_CONTEXT
     // stack: ctx, @SEGMENT_CALLDATA, 64, 32, y1, bn_add_return, kexit_info
     %build_address
-    %mload_packing
+    MLOAD_32BYTES
     // stack: x1, y1, bn_add_return, kexit_info
     %stack () -> (@SEGMENT_CALLDATA, 32, 32)
     GET_CONTEXT
     // stack: ctx, @SEGMENT_CALLDATA, 32, 32, x1, y1, bn_add_return, kexit_info
     %build_address
-    %mload_packing
+    MLOAD_32BYTES
     // stack: y0, x1, y1, bn_add_return, kexit_info
     %stack () -> (@SEGMENT_CALLDATA, 32)
     GET_CONTEXT
     // stack: ctx, @SEGMENT_CALLDATA, 32, y0, x1, y1, bn_add_return, kexit_info
     %build_address_no_offset
-    %mload_packing
+    MLOAD_32BYTES
     // stack: x0, y0, x1, y1, bn_add_return, kexit_info
     %jump(bn_add)
 bn_add_return:
@@ -53,11 +53,11 @@ bn_add_return:
     // Store the result (x, y) to the parent's return data using `mstore_unpacking`.
     %mstore_parent_context_metadata(@CTX_METADATA_RETURNDATA_SIZE, 64)
     %mload_context_metadata(@CTX_METADATA_PARENT_CONTEXT)
-    %stack (parent_ctx, x, y) -> (parent_ctx, @SEGMENT_RETURNDATA, x, 32, bn_add_contd6, parent_ctx, y)
+    %stack (parent_ctx, x, y) -> (parent_ctx, @SEGMENT_RETURNDATA, x, parent_ctx, y)
     %build_address_no_offset
-    %jump(mstore_unpacking)
-bn_add_contd6:
+    MSTORE_32BYTES_32
     POP
-    %stack (parent_ctx, y) -> (parent_ctx, @SEGMENT_RETURNDATA, 32, y, 32, pop_and_return_success)
+    %stack (parent_ctx, y) -> (parent_ctx, @SEGMENT_RETURNDATA, 32, y)
     %build_address
-    %jump(mstore_unpacking)
+    MSTORE_32BYTES_32
+    %jump(pop_and_return_success)

evm/src/cpu/kernel/asm/core/precompiles/bn_mul.asm

@@ -14,26 +14,26 @@ global precompile_bn_mul:
 
     %charge_gas_const(@BN_MUL_GAS)
 
-    // Load x, y, n from the call data using `mload_packing`.
+    // Load x, y, n from the call data using `MLOAD_32BYTES`.
     PUSH bn_mul_return
     // stack: bn_mul_return, kexit_info
     %stack () -> (@SEGMENT_CALLDATA, 64, 32)
     GET_CONTEXT
     // stack: ctx, @SEGMENT_CALLDATA, 64, 32, bn_mul_return, kexit_info
     %build_address
-    %mload_packing
+    MLOAD_32BYTES
     // stack: n, bn_mul_return, kexit_info
     %stack () -> (@SEGMENT_CALLDATA, 32, 32)
     GET_CONTEXT
     // stack: ctx, @SEGMENT_CALLDATA, 32, 32, n, bn_mul_return, kexit_info
     %build_address
-    %mload_packing
+    MLOAD_32BYTES
     // stack: y, n, bn_mul_return, kexit_info
     %stack () -> (@SEGMENT_CALLDATA, 32)
     GET_CONTEXT
     // stack: ctx, @SEGMENT_CALLDATA, 32, y, n, bn_mul_return, kexit_info
     %build_address_no_offset
-    %mload_packing
+    MLOAD_32BYTES
     // stack: x, y, n, bn_mul_return, kexit_info
     %jump(bn_mul)
 bn_mul_return:
@@ -47,11 +47,12 @@ bn_mul_return:
     // Store the result (Px, Py) to the parent's return data using `mstore_unpacking`.
     %mstore_parent_context_metadata(@CTX_METADATA_RETURNDATA_SIZE, 64)
     %mload_context_metadata(@CTX_METADATA_PARENT_CONTEXT)
-    %stack (parent_ctx, Px, Py) -> (parent_ctx, @SEGMENT_RETURNDATA, Px, 32, bn_mul_contd6, parent_ctx, Py)
+    %stack (parent_ctx, Px, Py) -> (parent_ctx, @SEGMENT_RETURNDATA, Px, parent_ctx, Py)
     %build_address_no_offset
-    %jump(mstore_unpacking)
+    MSTORE_32BYTES_32
 bn_mul_contd6:
     POP
-    %stack (parent_ctx, Py) -> (parent_ctx, @SEGMENT_RETURNDATA, 32, Py, 32, pop_and_return_success)
+    %stack (parent_ctx, Py) -> (parent_ctx, @SEGMENT_RETURNDATA, 32, Py)
     %build_address
-    %jump(mstore_unpacking)
+    MSTORE_32BYTES_32
+    %jump(pop_and_return_success)

evm/src/cpu/kernel/asm/core/precompiles/ecrec.asm

@@ -14,32 +14,32 @@ global precompile_ecrec:
 
     %charge_gas_const(@ECREC_GAS)
 
-    // Load hash, v, r, s from the call data using `mload_packing`.
+    // Load hash, v, r, s from the call data using `MLOAD_32BYTES`.
     PUSH ecrec_return
     // stack: ecrec_return, kexit_info
     %stack () -> (@SEGMENT_CALLDATA, 96, 32)
     GET_CONTEXT
     // stack: ctx, @SEGMENT_CALLDATA, 96, 32, ecrec_return, kexit_info
     %build_address
-    %mload_packing
+    MLOAD_32BYTES
     // stack: s, ecrec_return, kexit_info
     %stack () -> (@SEGMENT_CALLDATA, 64, 32)
     GET_CONTEXT
     // stack: ctx, @SEGMENT_CALLDATA, 64, 32, s, ecrec_return, kexit_info
     %build_address
-    %mload_packing
+    MLOAD_32BYTES
     // stack: r, s, ecrec_return, kexit_info
     %stack () -> (@SEGMENT_CALLDATA, 32, 32)
     GET_CONTEXT
     // stack: ctx, @SEGMENT_CALLDATA, 32, 32, r, s, ecrec_return, kexit_info
     %build_address
-    %mload_packing
+    MLOAD_32BYTES
     // stack: v, r, s, ecrec_return, kexit_info
     %stack () -> (@SEGMENT_CALLDATA, 32)
     GET_CONTEXT
     // stack: ctx, @SEGMENT_CALLDATA, 32, v, r, s, ecrec_return, kexit_info
     %build_address_no_offset
-    %mload_packing
+    MLOAD_32BYTES
     // stack: hash, v, r, s, ecrec_return, kexit_info
     %jump(ecrecover)
 ecrec_return:
@@ -49,9 +49,10 @@ ecrec_return:
     // Store the result address to the parent's return data using `mstore_unpacking`.
     %mstore_parent_context_metadata(@CTX_METADATA_RETURNDATA_SIZE, 32)
     %mload_context_metadata(@CTX_METADATA_PARENT_CONTEXT)
-    %stack (parent_ctx, address) -> (parent_ctx, @SEGMENT_RETURNDATA, address, 32, pop_and_return_success)
+    %stack (parent_ctx, address) -> (parent_ctx, @SEGMENT_RETURNDATA, address)
     %build_address_no_offset
-    %jump(mstore_unpacking)
+    MSTORE_32BYTES_32
+    %jump(pop_and_return_success)
 
 // On bad input, return empty return data but still return success.
 ecrec_bad_input:

evm/src/cpu/kernel/asm/core/precompiles/expmod.asm

@@ -18,7 +18,7 @@ mload_bytes_as_limbs:
     // stack: min(16, num_bytes), addr, num_bytes, retdest, total_num_limbs, len, ..limbs
     DUP2
     // stack: addr, min(16, num_bytes), addr, num_bytes, retdest, total_num_limbs, len, ..limbs
-    %mload_packing
+    MLOAD_32BYTES
     // stack: new_limb, addr, num_bytes, retdest, total_num_limbs, len, ..limbs
     %stack (new, addr, numb, ret, tot, len) -> (numb, addr, ret, tot, len, new)
     // stack: num_bytes, addr, retdest, total_num_limbs, len, new_limb, ..limbs
@@ -113,7 +113,7 @@ calculate_l_E_prime:
     PUSH @SEGMENT_CALLDATA
     GET_CONTEXT
     %build_address
-    %mload_packing
+    MLOAD_32BYTES
     // stack: i[96 + l_B..128 + l_B], l_E, l_B, retdest
     %log2_floor
     // stack: log2(i[96 + l_B..128 + l_B]), l_E, l_B, retdest
@@ -144,7 +144,7 @@ case_le_32:
     PUSH @SEGMENT_CALLDATA
     GET_CONTEXT
     %build_address
-    %mload_packing
+    MLOAD_32BYTES
     // stack: E, retdest
     %log2_floor
     // stack: log2(E), retdest
@@ -172,21 +172,21 @@ global precompile_expmod:
     GET_CONTEXT
     // stack: ctx, @SEGMENT_CALLDATA, 32, kexit_info
     %build_address_no_offset
-    %mload_packing
+    MLOAD_32BYTES
     // stack: l_B, kexit_info
 
     // Load l_E from i[32..64].
     %stack () -> (@SEGMENT_CALLDATA, 32, 32)
     GET_CONTEXT
     %build_address
-    %mload_packing
+    MLOAD_32BYTES
     // stack: l_E, l_B, kexit_info
 
     // Load l_M from i[64..96].
     %stack () -> (@SEGMENT_CALLDATA, 64, 32)
     GET_CONTEXT
     %build_address
-    %mload_packing
+    MLOAD_32BYTES
     // stack: l_M, l_E, l_B, kexit_info
     DUP3 ISZERO DUP2 ISZERO
     MUL // AND

evm/src/cpu/kernel/asm/core/precompiles/rip160.asm

@@ -44,6 +44,7 @@ rip160_contd:
     // Store the result hash to the parent's return data using `mstore_unpacking`.
     %mstore_parent_context_metadata(@CTX_METADATA_RETURNDATA_SIZE, 32)
     %mload_context_metadata(@CTX_METADATA_PARENT_CONTEXT)
-    %stack (parent_ctx, hash) -> (parent_ctx, @SEGMENT_RETURNDATA, hash, 32, pop_and_return_success)
+    %stack (parent_ctx, hash) -> (parent_ctx, @SEGMENT_RETURNDATA, hash)
     %build_address_no_offset
-    %jump(mstore_unpacking)
+    MSTORE_32BYTES_32
+    %jump(pop_and_return_success)

evm/src/cpu/kernel/asm/core/precompiles/sha256.asm

@@ -44,6 +44,7 @@ sha256_contd:
     // Store the result hash to the parent's return data using `mstore_unpacking`.
     %mstore_parent_context_metadata(@CTX_METADATA_RETURNDATA_SIZE, 32)
     %mload_context_metadata(@CTX_METADATA_PARENT_CONTEXT)
-    %stack (parent_ctx, hash) -> (parent_ctx, @SEGMENT_RETURNDATA, hash, 32, pop_and_return_success)
+    %stack (parent_ctx, hash) -> (parent_ctx, @SEGMENT_RETURNDATA, hash)
     %build_address_no_offset
-    %jump(mstore_unpacking)
+    MSTORE_32BYTES_32
+    %jump(pop_and_return_success)