perf: reduce MemBefore initial size

evm_arithmetization/src/cpu/kernel/asm/cdk_pre_execution.asm

@@ -57,20 +57,19 @@ global update_scalable_l1blockhash:
  // stack: retdest
  PROVER_INPUT(ger)
  // stack: l1blockhash?, retdest
- DUP1 %eq_const(@U256_MAX) %jumpi(skip_and_exit)
- // stack: l1blockhash, retdest
+ PUSH @SEGMENT_KERNEL_GENERAL
+ DUP2 %eq_const(@U256_MAX) %jumpi(skip_and_exit)
+ // stack: addr, l1blockhash, retdest
  PUSH @GLOBAL_EXIT_ROOT_STORAGE_POS
  PROVER_INPUT(ger)
- // stack: root, GLOBAL_EXIT_ROOT_STORAGE_POS, l1blockhash, retdest
- PUSH @SEGMENT_KERNEL_GENERAL
- // stack: addr, root, GLOBAL_EXIT_ROOT_STORAGE_POS, l1blockhash, retdest
+ // stack: root, GLOBAL_EXIT_ROOT_STORAGE_POS, addr, l1blockhash, retdest
+ DUP3
+ // stack: addr', root, GLOBAL_EXIT_ROOT_STORAGE_POS, addr, l1blockhash, retdest
  MSTORE_32BYTES_32
- // stack: addr, GLOBAL_EXIT_ROOT_STORAGE_POS, l1blockhash, retdest
+ // stack: addr'', GLOBAL_EXIT_ROOT_STORAGE_POS, addr, l1blockhash, retdest
  MSTORE_32BYTES_32
- // stack: addr, l1blockhash, retdest
- POP
- // stack: l1blockhash, retdest
- PUSH 64 PUSH @SEGMENT_KERNEL_GENERAL
+ // stack: addr'', addr, l1blockhash, retdest
+ %stack (addr_2, addr) -> (addr, 64)
  // stack: addr, len, l1blockhash, retdest
  KECCAK_GENERAL
  // stack: slot, l1blockhash, retdest
@@ -83,8 +82,8 @@ global update_scalable_l1blockhash:
  JUMP
 
 skip_and_exit:
- // stack: null, retdest
- POP
+ // stack: garbage, null, retdest
+ %pop2
  JUMP
 
 global update_scalable_timestamp:

evm_arithmetization/src/cpu/kernel/asm/core/gas.asm

@@ -16,36 +16,16 @@ global sys_gas:
 %endmacro
 
 
-// TODO: `%refund_gas` and `refund_gas_hook` are hooks used for debugging. They should be removed at some point and `refund_gas_original` renamed to `refund_gas`.
 %macro refund_gas
- PUSH %%after %jump(refund_gas_hook)
-%%after:
- %refund_gas_original
-%endmacro
-
-global refund_gas_hook:
- JUMP
-
-%macro refund_gas_original
  // stack: amount
  DUP1 %journal_refund
  %mload_global_metadata(@GLOBAL_METADATA_REFUND_COUNTER)
  ADD
  %mstore_global_metadata(@GLOBAL_METADATA_REFUND_COUNTER)
 %endmacro
 
-// TODO: `%charge_gas` and `charge_gas_hook` are hooks used for debugging. They should be removed at some point and `charge_gas_original` renamed to `charge_gas`.
-%macro charge_gas
- PUSH %%after %jump(charge_gas_hook)
-%%after:
- %charge_gas_original
-%endmacro
-
-global charge_gas_hook:
- JUMP
-
 // Charge gas. Faults if we exceed the limit for the current context.
-%macro charge_gas_original
+%macro charge_gas
  // stack: gas, kexit_info
  %shl_const(192)
  ADD

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

@@ -151,137 +151,127 @@ global write_table_if_jumpdest:
  (ctx, proof_prefix_addr, 32, proof_prefix_addr, ctx)
  ADD // combine context and offset to make an address (SEGMENT_CODE == 0)
  MLOAD_32BYTES
- // packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
- DUP1 %shl_const(1)
- DUP2 %shl_const(2)
+ // stack: packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
+ PUSH 0x8080808080808080808080808080808080808080808080808080808080808080 // mask
+ // stack: mask, packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
+ DUP2 %shl_const(1)
+ DUP3 %shl_const(2)
  AND
- // stack: (is_1_at_pos_2_and_3|(X)⁷)³², packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
+ // stack: (is_1_at_pos_2_and_3|(X)⁷)³², mask, packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
  // X denotes any value in {0,1} and Z^i is Z repeated i times
  NOT
- // stack: (is_0_at_2_or_3|X⁷)³², packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
- DUP2
+ // stack: (is_0_at_2_or_3|X⁷)³², mask, packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
+ DUP3
  OR
- // stack: (is_1_at_1 or is_0_at_2_or_3|X⁷)³², packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
- // stack: (~has_prefix|X⁷)³², packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
+ // stack: (is_1_at_1 or is_0_at_2_or_3|X⁷)³², mask, packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
+ // stack: (~has_prefix|X⁷)³², mask, packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
 
  // Compute in_range and has_prefix' = 
  // - in_range = (0xFF|X⁷)³² and ~has_prefix' = ~has_prefix OR is_0_at_4, for the first 15 bytes
  // - in_range = (has_prefix => is_0_at_4 |X⁷)³² and ~has_prefix' = ~has_prefix, for the next 15 bytes
  // - in_range = (~has_prefix|X⁷)³² and ~has_prefix' = ~has_prefix, for the last byte.
- DUP2 %shl_const(3)
+ DUP3 %shl_const(3)
  NOT
- // stack: (is_0_at_4|X⁷)³²,  (~has_prefix|X⁷)³², packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
+ // stack: (is_0_at_4|X⁷)³², (~has_prefix|X⁷)³², mask, packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
  // pos 0102030405060708091011121314151617181920212223242526272829303132
  PUSH 0xFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF00
  AND
- // stack: (is_0_at_4|X⁷)³¹|0⁸, (~has_prefix|X⁷)³², packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
- DUP1
+ // stack: (is_0_at_4|X⁷)³¹|0⁸, (~has_prefix|X⁷)³², mask, packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
  // pos 0102030405060708091011121314151617181920212223242526272829303132
  PUSH 0xFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF0000000000000000000000000000000000
+ DUP1 DUP3
  AND
- // stack: (is_0_at_4|X⁷)¹⁵|(0⁸)¹⁷, (is_0_at_4|X⁷)³¹|0⁸, (~has_prefix|X⁷)³², packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
- DUP3
+ // stack: (is_0_at_4|X⁷)¹⁵|(0⁸)¹⁷, 0xFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF0000000000000000000000000000000000, (is_0_at_4|X⁷)³¹|0⁸, (~has_prefix|X⁷)³², mask, packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
+ DUP4
  OR
- // (~has_prefix'|X⁷)³², (is_0_at_4|X⁷)³¹|0⁸, (~has_prefix|X⁷)³², packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
- SWAP2
+ // (~has_prefix'|X⁷)³², 0xFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF0000000000000000000000000000000000, (is_0_at_4|X⁷)³¹|0⁸, (~has_prefix|X⁷)³², mask, packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
+ SWAP3
  OR
- // pos 0102030405060708091011121314151617181920212223242526272829303132
- PUSH 0xFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF0000000000000000000000000000000000
  OR
- // stack: (in_range|X⁷)³², (~has_prefix'|X⁷)³², packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
+ // stack: (in_range|X⁷)³², (~has_prefix'|X⁷)³², mask, packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
 
  // Compute in_range' and ~has_prefix as
  // - in_range' = in_range and has_prefix' = ~has_prefix OR is_0_at_5, for bytes in positions 1-7 and 16-23 
  // - in_range' = in_range AND (has_prefix => is_0_at_5 |X⁷)³² and has_prefix' = ~has_prefix, for the rest.
 
- DUP3 %shl_const(4)
+ DUP4 %shl_const(4)
  NOT
- // stack: (is_0_at_5|X⁷)³², (in_range|X⁷)³², (~has_prefix|X⁷)³², packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
- DUP1
+ // stack: (is_0_at_5|X⁷)³², (in_range|X⁷)³², (~has_prefix|X⁷)³², mask, packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
  // pos 0102030405060708091011121314151617181920212223242526272829303132
  PUSH 0xFFFFFFFFFFFFFF0000000000000000FFFFFFFFFFFFFFFF000000000000000000
- AND
- // stack: (is_0_at_5|X⁷)⁷|(0⁸)⁸|(is_0_at_5|X⁷)⁸|(0⁸)⁸, (is_0_at_5|X⁷)³²,  (in_range|X⁷)³², (~has_prefix|X⁷)³², packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
- DUP4
+ DUP2 DUP2 AND
+ // stack: (is_0_at_5|X⁷)⁷|(0⁸)⁸|(is_0_at_5|X⁷)⁸|(0⁸)⁸, 0xFFFFFFFFFFFFFF0000000000000000FFFFFFFFFFFFFFFF000000000000000000, (is_0_at_5|X⁷)³², (in_range|X⁷)³², (~has_prefix|X⁷)³², mask, packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
+ DUP5
  OR
- // stack: (~has_prefix'|X⁷)³², (is_0_at_5|X⁷)³²,  (in_range|X⁷)³², (~has_prefix|X⁷)³², packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
- SWAP3
+ // stack: (~has_prefix'|X⁷)³², 0xFFFFFFFFFFFFFF0000000000000000FFFFFFFFFFFFFFFF000000000000000000, (is_0_at_5|X⁷)³², (in_range|X⁷)³², (~has_prefix|X⁷)³², mask, packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
+ SWAP4
  OR
- // pos 0102030405060708091011121314151617181920212223242526272829303132
- PUSH 0xFFFFFFFFFFFFFF0000000000000000FFFFFFFFFFFFFFFF000000000000000000
  OR
  AND
- // stack: (in_range'|X⁷)³²,  (~has_prefix'|X⁷)³², packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
+ // stack: (in_range'|X⁷)³², (~has_prefix'|X⁷)³², mask, packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
 
  // Compute in_range' and ~has_prefix' as
  // - in_range' = in_range and ~has_prefix' = ~has_prefix OR is_0_at_6, for bytes in positions 1-3, 8-11, 16-19, and 24-27 
  // - in_range' = in_range AND (has_prefix => is_0_at_6 |X⁷)³² and ~has_prefix' = has_prefix, for the rest.
- DUP3 %shl_const(5)
+ DUP4 %shl_const(5)
  NOT
- // stack: (is_0_at_6|X⁷)³², (in_range|X⁷)³², (~has_prefix|X⁷)³², packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
- DUP1
+ // stack: (is_0_at_6|X⁷)³², (in_range|X⁷)³², (~has_prefix|X⁷)³², mask, packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
  // pos 0102030405060708091011121314151617181920212223242526272829303132
  PUSH 0xFFFFFF00000000FFFFFFFF00000000FFFFFFFF00000000FFFFFFFF0000000000
- AND
- // stack: (is_0_at_6|X⁷)³|(0⁸)⁴|((is_0_at_6|X⁷)⁴|(0⁸)⁴)³, (is_0_at_6|X⁷)³²,  (in_range|X⁷)³², (~has_prefix|X⁷)³², packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
- DUP4
+ DUP2 DUP2 AND
+ // stack: (is_0_at_6|X⁷)³|(0⁸)⁴|((is_0_at_6|X⁷)⁴|(0⁸)⁴)³, 0xFFFFFF00000000FFFFFFFF00000000FFFFFFFF00000000FFFFFFFF0000000000, (is_0_at_6|X⁷)³², (in_range|X⁷)³², (~has_prefix|X⁷)³², mask, packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
+ DUP5
  OR
- // stack: (~has_prefix'|X⁷)³², (is_0_at_6|X⁷)³²,  (in_range|X⁷)³², (~has_prefix|X⁷)³², packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
- SWAP3
+ // stack: (~has_prefix'|X⁷)³², 0xFFFFFF00000000FFFFFFFF00000000FFFFFFFF00000000FFFFFFFF0000000000, (is_0_at_6|X⁷)³², (in_range|X⁷)³², (~has_prefix|X⁷)³², mask, packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
+ SWAP4
  OR
- // pos 0102030405060708091011121314151617181920212223242526272829303132
- PUSH 0xFFFFFF00000000FFFFFFFF00000000FFFFFFFF00000000FFFFFFFF0000000000
  OR
  AND
- // stack: (in_range'|X⁷)³², (~has_prefix'|X⁷)³², (in_range|X⁷)³², packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
+ // stack: (in_range'|X⁷)³², (~has_prefix'|X⁷)³², (in_range|X⁷)³², mask, packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
 
  // Compute in_range' and ~has_prefix' as
  // - in_range' = in_range and ~has_prefix' = has_prefix OR is_0_at_7, for bytes in 1, 4-5, 8-9, 12-13, 16-17, 20-21, 24-25, 28-29
  // - in_range' = in_range AND (has_prefix => is_0_at_7 |X⁷)³² and ~has_prefix' = ~has_prefix, for the rest.
- DUP3 %shl_const(6)
+ DUP4 %shl_const(6)
  NOT
- // stack: (is_0_at_7|X⁷)³², (in_range|X⁷)³², (~has_prefix|X⁷)³², packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
- DUP1
+ // stack: (is_0_at_7|X⁷)³², (in_range|X⁷)³², (~has_prefix|X⁷)³², mask, packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
  // pos 0102030405060708091011121314151617181920212223242526272829303132
  PUSH 0xFF0000FFFF0000FFFF0000FFFF0000FFFF0000FFFF0000FFFF0000FFFF000000
- AND
- // stack: is_0_at_7|X⁷|(0⁸)²|((is_0_at_7|X⁷)²|(0⁸)²)⁷, (is_0_at_7|X⁷)³², (in_range|X⁷)³², (~has_prefix|X⁷)³², packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
- DUP4
+ DUP2 DUP2 AND
+ // stack: is_0_at_7|X⁷|(0⁸)²|((is_0_at_7|X⁷)²|(0⁸)²)⁷, 0xFF0000FFFF0000FFFF0000FFFF0000FFFF0000FFFF0000FFFF0000FFFF000000, (is_0_at_7|X⁷)³², (in_range|X⁷)³², (~has_prefix|X⁷)³², mask, packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
+ DUP5
  OR
- // (~has_prefix'|X⁷)³², (is_0_at_7|X⁷)³²,  (in_range|X⁷)³², (~has_prefix|X⁷)³², packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
- SWAP3
+ // (~has_prefix'|X⁷)³², 0xFF0000FFFF0000FFFF0000FFFF0000FFFF0000FFFF0000FFFF0000FFFF000000, (is_0_at_7|X⁷)³², (in_range|X⁷)³², (~has_prefix|X⁷)³², mask, packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
+ SWAP4
  OR
- // pos 0102030405060708091011121314151617181920212223242526272829303132
- PUSH 0xFF0000FFFF0000FFFF0000FFFF0000FFFF0000FFFF0000FFFF0000FFFF000000
  OR
  AND
- // stack: (in_range'|X⁷)³², (~has_prefix'|X⁷)³², packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
+ // stack: (in_range'|X⁷)³², (~has_prefix'|X⁷)³², mask, packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
 
  // Compute in_range' as
  // - in_range' = in_range, for odd positions
  // - in_range' = in_range AND (has_prefix => is_0_at_8 |X⁷)³², for the rest
 
  SWAP1
- // stack: (~has_prefix|X⁷)³², (in_range|X⁷)³², packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
- DUP3 %shl_const(7)
+ // stack: (~has_prefix|X⁷)³², (in_range|X⁷)³², mask, packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
+ DUP4 %shl_const(7)
  NOT
- // stack: (is_0_at_8|X⁷)³², (~has_prefix|X⁷)³², (in_range|X⁷)³², packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
+ // stack: (is_0_at_8|X⁷)³², (~has_prefix|X⁷)³², (in_range|X⁷)³², mask, packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
  OR
  // pos 0102030405060708091011121314151617181920212223242526272829303132
  PUSH 0x00FF00FF00FF00FF00FF00FF00FF00FF00FF00FF00FF00FF00FF00FF00FF00FF
  OR
  AND
- // stack: (in_range|X⁷)³², packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
+ // stack: (in_range|X⁷)³², mask, packed_opcodes, proof_prefix_addr, ctx, jumpdest, retdest
 
  // Get rid of the irrelevant bits
- // pos 0102030405060708091011121314151617181920212223242526272829303132
- PUSH 0x8080808080808080808080808080808080808080808080808080808080808080
- AND
+ DUP2 AND
+
  // If we received a proof it MUST be valid or we abort immediately. This
  // is especially important for non-jumpdest proofs. Otherwise a malicious
  // prover might mark a valid jumpdest as invalid by providing an invalid proof
  // that makes verify_non_jumpdest return prematurely.
- %assert_eq_const(0x8080808080808080808080808080808080808080808080808080808080808080)
+ %assert_eq
  POP
  %add_const(32)
 

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

@@ -75,47 +75,39 @@ global precompile_blake2_f:
  SWAP1
  // stack: t0_addr = m0_addr + 8 * 16, t_0, t_1, flag, blake2_f_contd, kexit_info
 
+ %sub_const(8)
+ // stack: m0_addr + 8 * (16 - 1), t_0, t_1, flag, blake2_f_contd, kexit_info
+
+ PUSH @SEGMENT_CALLDATA
+ GET_CONTEXT
+ %build_address_no_offset
+
  %rep 16
- // stack: m0_addr + 8 * (16 - i), m_(i+1), ..., m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
- %sub_const(8)
- // stack: m0_addr + 8 * (16 - i - 1), m_(i+1), ..., m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
- DUP1
- // stack: m0_addr + 8 * (16 - i - 1), m0_addr + 8 * (16 - i - 1), m_(i+1), ..., m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
- PUSH @SEGMENT_CALLDATA
- // stack: @SEGMENT_CALLDATA, m0_addr + 8 * (16 - i - 1), m0_addr + 8 * (16 - i - 1), m_(i+1), ..., m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
- GET_CONTEXT
- // stack: ctx, @SEGMENT_CALLDATA, m0_addr + 8 * (16 - i - 1), m0_addr + 8 * (16 - i - 1), m_(i+1), ..., m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
- %build_address
+ // stack: base_addr, m0_addr + 8 * (16 - i - 1), m_(i+1), ..., m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
+ DUP2 DUP2
+ // stack: base_addr, m0_addr + 8 * (16 - i - 1), base_addr, m0_addr + 8 * (16 - i - 1), m_(i+1), ..., m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
+ ADD // base_addr + offset
  %mload_packing_u64_LE
- // stack: m_i, m0_addr + 8 * (16 - i - 1), m_(i+1), ..., m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
- SWAP1
- // stack: m0_addr + 8 * (16 - i - 1), m_i, m_(i+1), ..., m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
+ // stack: m_i, base_addr, m0_addr + 8 * (16 - i - 1), m_(i+1), ..., m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
+ SWAP2 %sub_const(8) SWAP1
+ // stack: base_addr, m0_addr + 8 * (16 - i - 2), m_i, m_(i+1), ..., m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
  %endrep
- // stack: m0_addr = h0_addr + 8 * 8, m_0, ..., m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
+ // stack: base_addr, m0_addr = h0_addr + 8 * 8, m_0, ..., m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
 
  %rep 8
- // stack: h0_addr + 8 * (8 - i), h_(i+1), ..., h_7, m_0..m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
- %sub_const(8)
- // stack: h0_addr + 8 * (8 - i - 1), h_(i+1), ..., h_7, m_0..m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
- DUP1
- // stack: h0_addr + 8 * (8 - i), h0_addr + 8 * (8 - i), h_(i+1), ..., h_7, m_0..m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
- PUSH @SEGMENT_CALLDATA
- // stack: @SEGMENT_CALLDATA, h0_addr + 8 * (8 - i), h0_addr + 8 * (8 - i), h_(i+1), ..., h_7, m_0..m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
- GET_CONTEXT
- // stack: ctx, @SEGMENT_CALLDATA, h0_addr + 8 * (8 - i), h0_addr + 8 * (8 - i), h_(i+1), ..., h_7, m_0..m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
- %build_address
+ // stack: base_addr, h0_addr + 8 * (8 - i), h_(i+1), ..., h_7, m_0..m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
+ DUP2 DUP2
+ // stack: base_addr, h0_addr + 8 * (8 - i), base_addr, h0_addr + 8 * (8 - i), h_(i+1), ..., h_7, m_0..m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
+ ADD // base_addr + offset
  %mload_packing_u64_LE
- // stack: h_i, h0_addr + 8 * (8 - i), h_(i+1), ..., h_7, m_0..m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
- SWAP1
- // stack: h0_addr + 8 * (8 - i), h_i, h_(i+1), ..., h_7, m_0..m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
+ // stack: h_i, base_addr, h0_addr + 8 * (8 - i), h_(i+1), ..., h_7, m_0..m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
+ SWAP2 %sub_const(8) SWAP1
+ // stack: base_addr, h0_addr + 8 * (8 - i), h_i, h_(i+1), ..., h_7, m_0..m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
  %endrep
- // stack: h0_addr + 8 * 8 = 68, h_0, ..., h_7, m_0..m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
- POP
-
- %stack () -> (@SEGMENT_CALLDATA, 4)
- 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
+ // stack: base_addr, garbage, h_0, ..., h_7, m_0..m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
+
+ PUSH 4 SWAP2 POP
+ // stack: base_addr, 4, h_0, ..., h_7, m_0..m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
  MLOAD_32BYTES
 
  // stack: rounds, h_0..h_7, m_0..m_15, t_0, t_1, flag, blake2_f_contd, kexit_info