plonky2/evm/src/cpu/kernel/asm/hash/blake/compression.asm

327 lines
9.7 KiB
NASM
Raw Normal View History

2022-11-10 15:40:44 -08:00
// Load the initial hash value (the IV, but with params XOR'd into the first word).
2022-11-10 15:24:08 -08:00
%macro blake_initial_hash_value
2022-12-13 10:08:30 -08:00
%blake_iv_i(7)
%blake_iv_i(6)
%blake_iv_i(5)
%blake_iv_i(4)
%blake_iv_i(3)
%blake_iv_i(2)
%blake_iv_i(1)
2022-11-09 14:41:41 -08:00
// stack: IV_1, IV_2, IV_3, IV_4, IV_5, IV_6, IV_7
PUSH 0x01010040 // params: key = 00, digest_size = 64 = 0x40
2022-12-13 10:08:30 -08:00
%blake_iv_i(0)
2022-11-09 14:41:41 -08:00
XOR
// stack: IV_0 ^ params, IV_1, IV_2, IV_3, IV_4, IV_5, IV_6, IV_7
%endmacro
2022-11-10 15:40:44 -08:00
// Address where the working version of the hash value is stored.
2022-11-10 15:24:08 -08:00
%macro blake_hash_value_addr
2022-12-09 09:34:42 -08:00
PUSH 0
2022-11-07 15:47:17 -08:00
// stack: 0
%mload_kernel_general
// stack: num_blocks
%mul_const(128)
2022-11-29 16:22:23 -08:00
%add_const(2)
// stack: num_bytes+2
2022-12-09 09:34:42 -08:00
%endmacro
2022-11-10 15:40:44 -08:00
// Address where the working version of the compression internal state is stored.
2022-11-10 15:24:08 -08:00
%macro blake_internal_state_addr
%blake_hash_value_addr
%add_const(8)
%endmacro
2022-11-10 15:40:44 -08:00
// Address where the current message block is stored.
2022-11-07 15:47:17 -08:00
%macro blake_message_addr
%blake_internal_state_addr
%add_const(16)
2022-11-07 14:30:14 -08:00
%endmacro
2022-11-15 14:46:45 -08:00
%macro blake_generate_new_hash_value(i)
%blake_hash_value_addr
%add_const($i)
%mload_kernel_general
// stack: h_i, ...
%blake_internal_state_addr
%add_const($i)
%mload_kernel_general
// stack: v_i, h_i, ...
%blake_internal_state_addr
%add_const($i)
%add_const(8)
%mload_kernel_general
// stack: v_(i+8), v_i, h_i, ...
XOR
XOR
// stack: h_i' = v_(i+8) ^ v_i ^ h_i, ...
%endmacro
2022-11-29 16:22:23 -08:00
%macro invert_bytes_blake_word
// stack: word, ...
DUP1
%and_const(0xff)
%shl_const(56)
SWAP1
// stack: word, first_byte, ...
DUP1
%shr_const(8)
%and_const(0xff)
%shl_const(48)
SWAP1
// stack: word, second_byte, first_byte, ...
DUP1
%shr_const(16)
%and_const(0xff)
%shl_const(40)
SWAP1
DUP1
%shr_const(24)
%and_const(0xff)
%shl_const(32)
SWAP1
DUP1
%shr_const(32)
%and_const(0xff)
%shl_const(24)
SWAP1
DUP1
%shr_const(40)
%and_const(0xff)
%shl_const(16)
SWAP1
DUP1
%shr_const(48)
%and_const(0xff)
%shl_const(8)
SWAP1
%shr_const(56)
%and_const(0xff)
%rep 7
OR
%endrep
%endmacro
2022-11-03 16:01:08 -07:00
global blake_compression:
2022-11-10 15:24:08 -08:00
// stack: retdest
2022-11-29 16:22:23 -08:00
PUSH 0
// stack: cur_block = 0, retdest
2022-11-14 12:33:14 -08:00
%blake_initial_hash_value
2022-11-29 16:22:23 -08:00
// stack: h_0, ..., h_7, cur_block, retdest
2022-11-14 12:33:14 -08:00
%blake_hash_value_addr
2022-11-29 16:22:23 -08:00
// stack: addr, h_0, ..., h_7, cur_block, retdest
2022-11-14 12:33:14 -08:00
%rep 8
2022-11-15 14:46:45 -08:00
SWAP1
2022-11-14 12:33:14 -08:00
DUP2
%mstore_kernel_general
%increment
%endrep
2022-11-29 16:22:23 -08:00
// stack: addr, cur_block, retdest
2022-11-14 12:33:14 -08:00
POP
2022-11-29 16:22:23 -08:00
// stack: cur_block, retdest
2022-11-10 15:24:08 -08:00
compression_loop:
2022-11-29 16:22:23 -08:00
// stack: cur_block, retdest
2022-11-10 15:24:08 -08:00
PUSH 0
%mload_kernel_general
2022-11-29 16:22:23 -08:00
// stack: num_blocks, cur_block, retdest
2022-11-10 15:24:08 -08:00
%decrement
2022-11-29 16:22:23 -08:00
// stack: num_blocks - 1, cur_block, retdest
2022-11-10 15:24:08 -08:00
DUP2
2022-11-29 16:22:23 -08:00
// stack: cur_block, num_blocks - 1, cur_block, retdest
2022-11-10 15:24:08 -08:00
EQ
2022-11-29 16:22:23 -08:00
// stack: is_last_block, cur_block, retdest
2022-11-10 15:24:08 -08:00
SWAP1
2022-11-29 16:22:23 -08:00
// stack: cur_block, is_last_block, retdest
PUSH 1
%mload_kernel_general
// stack: num_bytes, cur_block, is_last_block, retdest
DUP3
// stack: is_last_block, num_bytes, cur_block, is_last_block, retdest
MUL
// stack: is_last_block * num_bytes, cur_block, is_last_block, retdest
DUP2
// stack: cur_block, is_last_block * num_bytes, cur_block, is_last_block, retdest
2022-11-10 15:24:08 -08:00
%mul_const(128)
2022-11-29 16:22:23 -08:00
// stack: cur_block * 128, is_last_block * num_bytes, cur_block, is_last_block, retdest
DUP4
// stack: is_last_block, cur_block * 128, is_last_block * num_bytes, cur_block, is_last_block, retdest
ISZERO
// stack: not_last_block, cur_block * 128, is_last_block * num_bytes, cur_block, is_last_block, retdest
MUL
// stack: not_last_block * (cur_block * 128), is_last_block * num_bytes, cur_block, is_last_block, retdest
ADD
// stack: t = not_last_block * (cur_block * 128) + is_last_block * num_bytes, cur_block, is_last_block, retdest
SWAP1
// stack: cur_block, t, is_last_block, retdest
// stack: cur_block, t, is_last_block, retdest
%mul_const(128)
%add_const(2)
// stack: cur_block_start_byte, t, is_last_block, retdest
2022-11-10 15:24:08 -08:00
%blake_message_addr
2022-11-29 16:22:23 -08:00
// stack: message_addr, cur_block_start_byte, t, is_last_block, retdest
2022-11-10 15:24:08 -08:00
%rep 16
// stack: cur_message_addr, cur_block_byte, ...
DUP2
// stack: cur_block_byte, cur_message_addr, cur_block_byte, ...
%mload_blake_word
// stack: m_i, cur_message_addr, cur_block_byte, ...
DUP2
// stack: cur_message_addr, m_i, cur_message_addr, cur_block_byte, ...
%mstore_kernel_general
// stack: cur_message_addr, cur_block_byte, ...
%increment
// stack: cur_message_addr + 1, cur_block_byte, ...
SWAP1
// stack: cur_block_byte, cur_message_addr + 1, ...
2022-11-29 16:22:23 -08:00
%add_const(8)
// stack: cur_block_byte + 8, cur_message_addr + 1, ...
2022-11-10 15:24:08 -08:00
SWAP1
2022-11-29 16:22:23 -08:00
// stack: cur_message_addr + 1, cur_block_byte + 8, ...
2022-11-10 15:24:08 -08:00
%endrep
2022-11-29 16:22:23 -08:00
// stack: end_message_addr, end_block_start_byte, t, is_last_block, retdest
2022-11-10 15:24:08 -08:00
POP
POP
2022-11-29 16:22:23 -08:00
// stack: t, is_last_block, retdest
SWAP1
// stack: is_last_block, t, retdest
2022-11-16 14:23:15 -08:00
%mul_const(0xFFFFFFFFFFFFFFFF)
2022-11-29 16:22:23 -08:00
// stack: invert_if_last_block, t, retdest
2022-11-14 12:33:14 -08:00
%blake_hash_value_addr
2022-11-16 14:23:15 -08:00
%add_const(7)
2022-11-14 12:33:14 -08:00
%rep 8
// stack: addr, ...
DUP1
// stack: addr, addr, ...
%mload_kernel_general
// stack: val, addr, ...
SWAP1
// stack: addr, val, ...
2022-11-16 14:23:15 -08:00
%decrement
2022-11-14 12:33:14 -08:00
%endrep
2022-11-29 16:22:23 -08:00
// stack: addr, h_0, ..., h_7, invert_if_last_block, t, retdest
2022-11-15 14:46:45 -08:00
POP
2022-11-29 16:22:23 -08:00
// stack: h_0, ..., h_7, invert_if_last_block, t, retdest
2022-11-09 13:48:56 -08:00
%blake_internal_state_addr
2022-11-29 16:22:23 -08:00
// stack: start, h_0, ..., h_7, invert_if_last_block, t, retdest
2022-11-10 15:24:08 -08:00
// First eight words of compression state: current state h_0, ..., h_7.
2022-11-04 15:27:34 -07:00
%rep 8
SWAP1
DUP2
%mstore_kernel_general
%increment
%endrep
2022-11-29 16:22:23 -08:00
// stack: start + 8, invert_if_last_block, t, retdest
2022-11-04 15:27:34 -07:00
PUSH 0
2022-11-29 16:22:23 -08:00
// stack: 0, start + 8, invert_if_last_block, t, retdest
2022-11-04 15:27:34 -07:00
%rep 4
2022-11-07 14:30:14 -08:00
// stack: i, loc, ...
DUP2
DUP2
// stack: i, loc, i, loc,...
%blake_iv
// stack: IV_i, loc, i, loc,...
SWAP1
// stack: loc, IV_i, i, loc,...
%mstore_kernel_general
// stack: i, loc,...
%increment
SWAP1
%increment
SWAP1
// stack: i + 1, loc + 1,...
%endrep
2022-11-29 16:22:23 -08:00
// stack: 4, start + 12, invert_if_last_block, t, retdest
%stack (i, loc, inv, t) -> (t, t, i, loc, inv)
// stack: t, t, 4, start + 12, invert_if_last_block, retdest
%shr_const(64)
// stack: t >> 64, t, 4, start + 12, invert_if_last_block, retdest
SWAP1
// stack: t, t >> 64, 4, start + 12, invert_if_last_block, retdest
PUSH 1
%shl_const(64)
// stack: 1 << 64, t, t >> 64, 4, start + 12, invert_if_last_block, retdest
SWAP1
MOD
// stack: t_lo = t % (1 << 64), t_hi = t >> 64, 4, start + 12, invert_if_last_block, retdest
%stack (t_lo, t_hi, i, loc, inv) -> (i, loc, t_lo, t_hi, inv, 0)
// stack: 4, start + 12, t_lo, t_hi, invert_if_last_block, 0, retdest
// XOR the values (t % 2**64, t >> 64, invert_if, 0) into the last four IV values.
2022-11-07 14:30:14 -08:00
%rep 4
2022-11-15 14:46:45 -08:00
// stack: i, loc, val, next_val,...
2022-11-07 14:30:14 -08:00
%stack (i, loc, val) -> (i, val, loc, i, loc)
// stack: i, val, loc, i, loc, next_val,...
%blake_iv
// stack: IV_i, val, loc, i, loc, next_val,...
XOR
// stack: val ^ IV_i, loc, i, loc, next_val,...
SWAP1
// stack: loc, val ^ IV_i, i, loc, next_val,...
%mstore_kernel_general
// stack: i, loc, next_val,...
%increment
SWAP1
%increment
SWAP1
// stack: i + 1, loc + 1, next_val,...
%endrep
2022-11-15 14:46:45 -08:00
// stack: 8, loc + 16, retdest
2022-11-07 14:30:14 -08:00
POP
POP
2022-11-15 14:46:45 -08:00
// stack: retdest
2022-11-09 13:48:56 -08:00
%blake_internal_state_addr
2022-11-15 14:46:45 -08:00
// stack: start, retdest
2022-11-07 14:30:14 -08:00
PUSH 0
2022-11-15 14:46:45 -08:00
// stack: round=0, start, retdest
2022-11-07 15:47:17 -08:00
%rep 12
2022-11-15 14:46:45 -08:00
// stack: round, start, retdest
2022-11-07 15:47:17 -08:00
%call_blake_g_function(0, 4, 8, 12, 0, 1)
%call_blake_g_function(1, 5, 9, 13, 2, 3)
%call_blake_g_function(2, 6, 10, 14, 4, 5)
%call_blake_g_function(3, 7, 11, 15, 6, 7)
%call_blake_g_function(0, 5, 10, 15, 8, 9)
%call_blake_g_function(1, 6, 11, 12, 10, 11)
%call_blake_g_function(2, 7, 8, 13, 12, 13)
%call_blake_g_function(3, 4, 9, 14, 14, 15)
2022-11-15 14:46:45 -08:00
// stack: round, start, retdest
2022-11-07 15:47:17 -08:00
%increment
2022-11-15 14:46:45 -08:00
// stack: round + 1, start, retdest
2022-11-07 15:47:17 -08:00
%endrep
2022-11-15 14:46:45 -08:00
// stack: 12, start, retdest
POP
POP
// stack: retdest
%blake_generate_new_hash_value(7)
2022-11-29 16:22:23 -08:00
%invert_bytes_blake_word
2022-11-15 14:46:45 -08:00
%blake_generate_new_hash_value(6)
2022-11-29 16:22:23 -08:00
%invert_bytes_blake_word
2022-11-15 14:46:45 -08:00
%blake_generate_new_hash_value(5)
2022-11-29 16:22:23 -08:00
%invert_bytes_blake_word
2022-11-15 14:46:45 -08:00
%blake_generate_new_hash_value(4)
2022-11-29 16:22:23 -08:00
%invert_bytes_blake_word
2022-11-15 14:46:45 -08:00
%blake_generate_new_hash_value(3)
2022-11-29 16:22:23 -08:00
%invert_bytes_blake_word
2022-11-15 14:46:45 -08:00
%blake_generate_new_hash_value(2)
2022-11-29 16:22:23 -08:00
%invert_bytes_blake_word
2022-11-15 14:46:45 -08:00
%blake_generate_new_hash_value(1)
2022-11-29 16:22:23 -08:00
%invert_bytes_blake_word
2022-11-15 14:46:45 -08:00
%blake_generate_new_hash_value(0)
2022-11-29 16:22:23 -08:00
%invert_bytes_blake_word
2022-11-15 14:46:45 -08:00
// stack: h_0', h_1', h_2', h_3', h_4', h_5', h_6', h_7', retdest
2022-11-16 14:23:15 -08:00
%shl_const(64)
2022-11-15 15:40:56 -08:00
OR
2022-11-16 14:23:15 -08:00
%shl_const(64)
2022-11-15 15:40:56 -08:00
OR
2022-11-16 14:23:15 -08:00
%shl_const(64)
2022-11-15 15:40:56 -08:00
OR
2022-11-16 14:23:15 -08:00
// stack: h_0' || h_1' || h_2' || h_3', h_4', h_5', h_6', h_7', retdest
%stack (first, second: 4) -> (second, first)
// stack: h_4', h_5', h_6', h_7', h_0' || h_1' || h_2' || h_3', retdest
%shl_const(64)
2022-11-15 15:40:56 -08:00
OR
2022-11-16 14:23:15 -08:00
%shl_const(64)
2022-11-15 15:40:56 -08:00
OR
2022-11-16 14:23:15 -08:00
%shl_const(64)
2022-11-15 15:40:56 -08:00
OR
2022-11-29 16:22:23 -08:00
// stack: hash_second = h_4' || h_5' || h_6' || h_7', hash_first = h_0' || h_1' || h_2' || h_3', retdest
%stack (second, first, ret) -> (ret, second, first)
2022-11-16 14:23:15 -08:00
// stack: retdest, hash_first, hash_second
2022-11-15 15:36:15 -08:00
JUMP