Remove some more CPU cycles (0xPolygonZero#1472)

* Speed-up some mload/mstore calls * Misc * Improve logs loop * Speed-up bloom * Speed-up access_lists loops * Fix * Speed up selfdestruct loop * Speed-up touched_addresses loop * Speed-up receipt loop * Skip rep loop * Fix * Misc * Review
topos-protocol · Jan 17, 2024 · 10fc966 · 10fc966
1 parent 990eb34
commit 10fc966
Show file tree

Hide file tree

Showing 18 changed files with 239 additions and 143 deletions.
diff --git a/evm/src/cpu/kernel/asm/bloom_filter.asm b/evm/src/cpu/kernel/asm/bloom_filter.asm
@@ -55,20 +55,21 @@ logs_bloom_loop:
     // Add address to bloom filter.
     %increment
     // stack: addr_ptr, i, logs_len, retdest
+    PUSH @SEGMENT_LOGS_DATA %build_kernel_address
     DUP1
-    %mload_kernel(@SEGMENT_LOGS_DATA)
-    // stack: addr, addr_ptr, i, logs_len, retdest
+    MLOAD_GENERAL
+    // stack: addr, full_addr_ptr, i, logs_len, retdest
     PUSH 0
-    // stack: is_topic, addr, addr_ptr, i, logs_len, retdest
+    // stack: is_topic, addr, full_addr_ptr, i, logs_len, retdest
     %add_to_bloom
-    // stack: addr_ptr, i, logs_len, retdest
+    // stack: full_addr_ptr, i, logs_len, retdest
     %increment
-    // stack: num_topics_ptr, i, logs_len, retdest
+    // stack: full_num_topics_ptr, i, logs_len, retdest
     DUP1
-    %mload_kernel(@SEGMENT_LOGS_DATA)
-    // stack: num_topics, num_topics_ptr, i, logs_len, retdest
+    MLOAD_GENERAL
+    // stack: num_topics, full_num_topics_ptr, i, logs_len, retdest
     SWAP1 %increment
-    // stack: topics_ptr, num_topics, i, logs_len, retdest
+    // stack: full_topics_ptr, num_topics, i, logs_len, retdest
     PUSH 0
 
 logs_bloom_topic_loop:
@@ -78,7 +79,7 @@ logs_bloom_topic_loop:
     %jumpi(logs_bloom_topic_end)
     DUP2 DUP2 ADD
     // stack: curr_topic_ptr, j, topics_ptr, num_topics, i, logs_len, retdest
-    %mload_kernel(@SEGMENT_LOGS_DATA)
+    MLOAD_GENERAL
     // stack: topic, j, topics_ptr, num_topics, i, logs_len, retdest
     PUSH 1
     // stack: is_topic, topic, j, topics_ptr, num_topics, i, logs_len, retdest
@@ -142,19 +143,20 @@ logs_bloom_end:
 // Also updates the block bloom filter.
 %macro bloom_write_bit
     // stack: byte_index, byte_bit_index
+    PUSH @SEGMENT_TXN_BLOOM
+    %build_kernel_address
     PUSH 1
     DUP3
-    // stack: byte_bit_index, 1, byte_index, byte_bit_index
+    // stack: byte_bit_index, 1, byte_addr, byte_bit_index
     PUSH 7 SUB
     SHL
     // Updates the current txn bloom filter.
     SWAP2 POP DUP1
-    %mload_kernel(@SEGMENT_TXN_BLOOM)
-    // stack: old_bloom_byte, byte_index, one_shifted_by_index
+    MLOAD_GENERAL
+    // stack: old_bloom_byte, byte_addr, one_shifted_by_index
     DUP3 OR
-    // stack: new_bloom_byte, byte_index, one_shifted_by_index
-    SWAP1
-    %mstore_kernel(@SEGMENT_TXN_BLOOM)
+    // stack: new_bloom_byte, byte_addr, one_shifted_by_index
+    MSTORE_GENERAL
     // stack: one_shifted_by_index
     POP
     // stack: empty

diff --git a/evm/src/cpu/kernel/asm/core/access_lists.asm b/evm/src/cpu/kernel/asm/core/access_lists.asm
@@ -25,12 +25,15 @@ global insert_accessed_addresses:
     // stack: addr, retdest
     %mload_global_metadata(@GLOBAL_METADATA_ACCESSED_ADDRESSES_LEN)
     // stack: len, addr, retdest
-    PUSH 0
+    PUSH @SEGMENT_ACCESSED_ADDRESSES ADD
+    PUSH @SEGMENT_ACCESSED_ADDRESSES
 insert_accessed_addresses_loop:
+    // `i` and `len` are both scaled by SEGMENT_ACCESSED_ADDRESSES
     %stack (i, len, addr, retdest) -> (i, len, i, len, addr, retdest)
     EQ %jumpi(insert_address)
     // stack: i, len, addr, retdest
-    DUP1 %mload_kernel(@SEGMENT_ACCESSED_ADDRESSES)
+    DUP1
+    MLOAD_GENERAL
     // stack: loaded_addr, i, len, addr, retdest
     DUP4
     // stack: addr, loaded_addr, i, len, addr, retdest
@@ -42,9 +45,10 @@ insert_accessed_addresses_loop:
 insert_address:
     %stack (i, len, addr, retdest) -> (i, addr, len, retdest)
     DUP2 %journal_add_account_loaded // Add a journal entry for the loaded account.
-    %mstore_kernel(@SEGMENT_ACCESSED_ADDRESSES) // Store new address at the end of the array.
+    %swap_mstore // Store new address at the end of the array.
     // stack: len, retdest
     %increment
+    %sub_const(@SEGMENT_ACCESSED_ADDRESSES) // unscale `len`
     %mstore_global_metadata(@GLOBAL_METADATA_ACCESSED_ADDRESSES_LEN) // Store new length.
     PUSH 1 // Return 1 to indicate that the address was inserted.
     SWAP1 JUMP
@@ -59,12 +63,14 @@ global remove_accessed_addresses:
     // stack: addr, retdest
     %mload_global_metadata(@GLOBAL_METADATA_ACCESSED_ADDRESSES_LEN)
     // stack: len, addr, retdest
-    PUSH 0
+    PUSH @SEGMENT_ACCESSED_ADDRESSES ADD
+    PUSH @SEGMENT_ACCESSED_ADDRESSES
 remove_accessed_addresses_loop:
+    // `i` and `len` are both scaled by SEGMENT_ACCESSED_ADDRESSES
     %stack (i, len, addr, retdest) -> (i, len, i, len, addr, retdest)
     EQ %jumpi(panic)
     // stack: i, len, addr, retdest
-    DUP1 %mload_kernel(@SEGMENT_ACCESSED_ADDRESSES)
+    DUP1 MLOAD_GENERAL
     // stack: loaded_addr, i, len, addr, retdest
     DUP4
     // stack: addr, loaded_addr, i, len, addr, retdest
@@ -74,12 +80,15 @@ remove_accessed_addresses_loop:
     %jump(remove_accessed_addresses_loop)
 remove_accessed_addresses_found:
     %stack (i, len, addr, retdest) -> (len, 1, i, retdest)
-    SUB DUP1 %mstore_global_metadata(@GLOBAL_METADATA_ACCESSED_ADDRESSES_LEN) // Decrement the access list length.
+    SUB  // len -= 1
+    PUSH @SEGMENT_ACCESSED_ADDRESSES
+    DUP2 SUB // unscale `len`
+    %mstore_global_metadata(@GLOBAL_METADATA_ACCESSED_ADDRESSES_LEN) // Decrement the access list length.
     // stack: len-1, i, retdest
-    %mload_kernel(@SEGMENT_ACCESSED_ADDRESSES) // Load the last address in the access list.
+    MLOAD_GENERAL // Load the last address in the access list.
     // stack: last_addr, i, retdest
-    SWAP1
-    %mstore_kernel(@SEGMENT_ACCESSED_ADDRESSES) // Store the last address at the position of the removed address.
+    MSTORE_GENERAL
+    // Store the last address at the position of the removed address.
     JUMP
 
 
@@ -97,14 +106,16 @@ global insert_accessed_storage_keys:
     // stack: addr, key, value, retdest
     %mload_global_metadata(@GLOBAL_METADATA_ACCESSED_STORAGE_KEYS_LEN)
     // stack: len, addr, key, value, retdest
-    PUSH 0
+    PUSH @SEGMENT_ACCESSED_STORAGE_KEYS ADD
+    PUSH @SEGMENT_ACCESSED_STORAGE_KEYS
 insert_accessed_storage_keys_loop:
+    // `i` and `len` are both scaled by SEGMENT_ACCESSED_STORAGE_KEYS
     %stack (i, len, addr, key, value, retdest) -> (i, len, i, len, addr, key, value, retdest)
     EQ %jumpi(insert_storage_key)
     // stack: i, len, addr, key, value, retdest
-    DUP1 %increment %mload_kernel(@SEGMENT_ACCESSED_STORAGE_KEYS)
+    DUP1 %increment MLOAD_GENERAL
     // stack: loaded_key, i, len, addr, key, value, retdest
-    DUP2 %mload_kernel(@SEGMENT_ACCESSED_STORAGE_KEYS)
+    DUP2 MLOAD_GENERAL
     // stack: loaded_addr, loaded_key, i, len, addr, key, value, retdest
     DUP5 EQ
     // stack: loaded_addr==addr, loaded_key, i, len, addr, key, value, retdest
@@ -120,28 +131,26 @@ insert_storage_key:
     // stack: i, len, addr, key, value, retdest
     DUP4 DUP4 %journal_add_storage_loaded // Add a journal entry for the loaded storage key.
     // stack: i, len, addr, key, value, retdest
-    DUP1
-    PUSH @SEGMENT_ACCESSED_STORAGE_KEYS
-    %build_kernel_address
 
-    %stack(dst, i, len, addr, key, value) -> (addr, dst, dst, key, dst, value, i, value)
+    %stack(dst, len, addr, key, value) -> (addr, dst, dst, key, dst, value, dst, @SEGMENT_ACCESSED_STORAGE_KEYS, value)
     MSTORE_GENERAL // Store new address at the end of the array.
-    // stack: dst, key, dst, value, i, value, retdest
+    // stack: dst, key, dst, value, dst, segment, value, retdest
     %increment SWAP1
     MSTORE_GENERAL // Store new key after that
-    // stack: dst, value, i, value, retdest
+    // stack: dst, value, dst, segment, value, retdest
     %add_const(2) SWAP1
     MSTORE_GENERAL // Store new value after that
-    // stack: i, value, retdest
+    // stack: dst, segment, value, retdest
     %add_const(3)
+    SUB // unscale dst
     %mstore_global_metadata(@GLOBAL_METADATA_ACCESSED_STORAGE_KEYS_LEN) // Store new length.
     %stack (value, retdest) -> (retdest, 1, value) // Return 1 to indicate that the storage key was inserted.
     JUMP
 
 insert_accessed_storage_keys_found:
     // stack: i, len, addr, key, value, retdest
     %add_const(2)
-    %mload_kernel(@SEGMENT_ACCESSED_STORAGE_KEYS)
+    MLOAD_GENERAL
     %stack (original_value, len, addr, key, value, retdest) -> (retdest, 0, original_value) // Return 0 to indicate that the storage key was already present.
     JUMP
 
@@ -151,14 +160,16 @@ global remove_accessed_storage_keys:
     // stack: addr, key, retdest
     %mload_global_metadata(@GLOBAL_METADATA_ACCESSED_STORAGE_KEYS_LEN)
     // stack: len, addr, key, retdest
-    PUSH 0
+    PUSH @SEGMENT_ACCESSED_STORAGE_KEYS ADD
+    PUSH @SEGMENT_ACCESSED_STORAGE_KEYS
 remove_accessed_storage_keys_loop:
+    // `i` and `len` are both scaled by SEGMENT_ACCESSED_STORAGE_KEYS
     %stack (i, len, addr, key, retdest) -> (i, len, i, len, addr, key, retdest)
     EQ %jumpi(panic)
     // stack: i, len, addr, key, retdest
-    DUP1 %increment %mload_kernel(@SEGMENT_ACCESSED_STORAGE_KEYS)
+    DUP1 %increment MLOAD_GENERAL
     // stack: loaded_key, i, len, addr, key, retdest
-    DUP2 %mload_kernel(@SEGMENT_ACCESSED_STORAGE_KEYS)
+    DUP2 MLOAD_GENERAL
     // stack: loaded_addr, loaded_key, i, len, addr, key, retdest
     DUP5 EQ
     // stack: loaded_addr==addr, loaded_key, i, len, addr, key, retdest
@@ -172,18 +183,21 @@ remove_accessed_storage_keys_loop:
 
 remove_accessed_storage_keys_found:
     %stack (i, len, addr, key, retdest) -> (len, 3, i, retdest)
-    SUB DUP1 %mstore_global_metadata(@GLOBAL_METADATA_ACCESSED_STORAGE_KEYS_LEN) // Decrease the access list length.
+    SUB 
+    PUSH @SEGMENT_ACCESSED_STORAGE_KEYS
+    DUP2 SUB // unscale
+    %mstore_global_metadata(@GLOBAL_METADATA_ACCESSED_STORAGE_KEYS_LEN) // Decrease the access list length.
     // stack: len-3, i, retdest
-    DUP1 %add_const(2) %mload_kernel(@SEGMENT_ACCESSED_STORAGE_KEYS)
+    DUP1 %add_const(2) MLOAD_GENERAL
     // stack: last_value, len-3, i, retdest
-    DUP2 %add_const(1) %mload_kernel(@SEGMENT_ACCESSED_STORAGE_KEYS)
+    DUP2 %add_const(1) MLOAD_GENERAL
     // stack: last_key, last_value, len-3, i, retdest
-    DUP3 %mload_kernel(@SEGMENT_ACCESSED_STORAGE_KEYS)
+    DUP3 MLOAD_GENERAL
     // stack: last_addr, last_key, last_value, len-3, i, retdest
-    DUP5 %mstore_kernel(@SEGMENT_ACCESSED_STORAGE_KEYS) // Move the last tuple to the position of the removed tuple.
+    DUP5 %swap_mstore // Move the last tuple to the position of the removed tuple.
     // stack: last_key, last_value, len-3, i, retdest
-    DUP4 %add_const(1) %mstore_kernel(@SEGMENT_ACCESSED_STORAGE_KEYS)
+    DUP4 %add_const(1) %swap_mstore
     // stack: last_value, len-3, i, retdest
-    DUP3 %add_const(2) %mstore_kernel(@SEGMENT_ACCESSED_STORAGE_KEYS)
+    DUP3 %add_const(2) %swap_mstore
     // stack: len-3, i, retdest
     %pop2 JUMP
diff --git a/evm/src/cpu/kernel/asm/core/create_addresses.asm b/evm/src/cpu/kernel/asm/core/create_addresses.asm
@@ -37,17 +37,17 @@ global get_create_address:
 // Post stack: 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, salt, code_hash, retdest)
-    ADD
+    PUSH @SEGMENT_KERNEL_GENERAL
+    DUP1
+    PUSH 0xff
+    MSTORE_GENERAL
+    // stack: addr, sender, code_hash, salt, retdest
+    %increment
+    %stack (addr, sender, code_hash, salt, retdest) -> (addr, sender, salt, code_hash, retdest)
     MSTORE_32BYTES_20
-    POP
-    %stack (salt, code_hash, retdest) -> (@SEGMENT_KERNEL_GENERAL, 21, salt, code_hash, retdest)
-    ADD
+    // stack: addr, salt, code_hash, retdest
     MSTORE_32BYTES_32
-    POP
-    %stack (code_hash, retdest) -> (@SEGMENT_KERNEL_GENERAL, 53, code_hash, retdest)
-    ADD
+    // stack: addr, code_hash, retdest
     MSTORE_32BYTES_32
     POP
     %stack (retdest) -> (@SEGMENT_KERNEL_GENERAL, 85, retdest) // offset == context == 0

diff --git a/evm/src/cpu/kernel/asm/core/create_receipt.asm b/evm/src/cpu/kernel/asm/core/create_receipt.asm
@@ -116,22 +116,23 @@ process_receipt_logs_loop:
     %mload_kernel(@SEGMENT_LOGS)
     // stack: log_ptr, i, num_logs, receipt_ptr, txn_nb, new_cum_gas, txn_nb, num_nibbles, retdest
     // Write payload_len.
+    PUSH @SEGMENT_LOGS_DATA %build_kernel_address
     DUP1
-    %mload_kernel(@SEGMENT_LOGS_DATA)
+    MLOAD_GENERAL
     %append_to_trie_data
     // stack: log_ptr, i, num_logs, receipt_ptr, txn_nb, new_cum_gas, txn_nb, num_nibbles, retdest
     // Write address.
     %increment
     // stack: addr_ptr, i, num_logs, receipt_ptr, txn_nb, new_cum_gas, txn_nb, num_nibbles, retdest
     DUP1
-    %mload_kernel(@SEGMENT_LOGS_DATA)
+    MLOAD_GENERAL
     %append_to_trie_data
     // stack: addr_ptr, i, num_logs, receipt_ptr, txn_nb, new_cum_gas, txn_nb, num_nibbles, retdest
     //Write num_topics.
     %increment
     // stack: num_topics_ptr, i, num_logs, receipt_ptr, txn_nb, new_cum_gas, txn_nb, num_nibbles, retdest
     DUP1
-    %mload_kernel(@SEGMENT_LOGS_DATA)
+    MLOAD_GENERAL
     // stack: num_topics, num_topics_ptr, i, num_logs, receipt_ptr, txn_nb, new_cum_gas, txn_nb, num_nibbles, retdest
     DUP1
     %append_to_trie_data
@@ -151,7 +152,7 @@ process_receipt_topics_loop:
     DUP3 DUP2
     ADD
     // stack: cur_topic_ptr, j, num_topics, topics_ptr, i, num_logs, receipt_ptr, txn_nb, new_cum_gas, txn_nb, num_nibbles, retdest
-    %mload_kernel(@SEGMENT_LOGS_DATA)
+    MLOAD_GENERAL
     %append_to_trie_data
     // stack: j, num_topics, topics_ptr, i, num_logs, receipt_ptr, txn_nb, new_cum_gas, txn_nb, num_nibbles, retdest
     %increment
@@ -164,7 +165,7 @@ process_receipt_topics_end:
     // stack: data_len_ptr, i, num_logs, receipt_ptr, txn_nb, new_cum_gas, txn_nb, num_nibbles, retdest
     // Write data_len
     DUP1
-    %mload_kernel(@SEGMENT_LOGS_DATA)
+    MLOAD_GENERAL
     // stack: data_len, data_len_ptr, i, num_logs, receipt_ptr, txn_nb, new_cum_gas, txn_nb, num_nibbles, retdest
     DUP1
     %append_to_trie_data
@@ -184,7 +185,7 @@ process_receipt_data_loop:
     DUP3 DUP2
     ADD
     // stack: cur_data_ptr, j, data_len, data_ptr, i, num_logs, receipt_ptr, txn_nb, new_cum_gas, txn_nb, num_nibbles, retdest
-    %mload_kernel(@SEGMENT_LOGS_DATA)
+    MLOAD_GENERAL
     %append_to_trie_data
     // stack: j, data_len, data_ptr, i, num_logs, receipt_ptr, txn_nb, new_cum_gas, txn_nb, num_nibbles, retdest
     %increment
@@ -205,19 +206,10 @@ process_receipt_after_write:
     DUP5
     %mpt_insert_receipt_trie
     // stack: new_cum_gas, txn_nb, num_nibbles, retdest
-    // Now, we set the Bloom filter back to 0. We proceed by chunks of 32 bytes.
-    PUSH @SEGMENT_TXN_BLOOM // ctx == offset == 0
-    %rep 8
-        // stack: addr, new_cum_gas, txn_nb, num_nibbles, retdest
-        PUSH 0 // we will fill the memory segment with zeroes
-        DUP2
-        // stack: addr, 0, addr, new_cum_gas, txn_nb, num_nibbles, retdest
-        MSTORE_32BYTES_32
-        // stack: new_addr, addr, new_cum_gas, txn_nb, num_nibbles, retdest
-        SWAP1 POP
-    %endrep
-    POP
-    // stack: new_cum_gas, txn_nb, num_nibbles, retdest
+
+    // We don't need to reset the bloom filter segment as we only process a single transaction.
+    // TODO: Revert in case we add back support for multi-txn proofs.
+
     %stack (new_cum_gas, txn_nb, num_nibbles, retdest) -> (retdest, new_cum_gas)
     JUMP