All you can cache

layer 2cachingstorage

Intermediate

Ori Pomerantz

September 15, 2022

23 minute read minute read

When using rollups the cost of a byte in the transaction is a lot more expensive than the cost of a storage slot. Therefore, it makes sense to cache as much information as possible on chain.

In this article you learn how to create and use a caching contract in such a way that any parameter value that is likely to be used multiple times will be cached and available for use (after the first time) with a much smaller number of bytes, and how to write off chain code that uses this cache.

If you want to skip the article and just see the source code, it is here(opens in a new tab). The development stack is Foundry(opens in a new tab).

Overall design

For the sake of simplicity we'll assume all transaction parameters are uint256, 32 bytes long. When we receive a transaction, we'll parse each parameter like this:

If the first byte is 0xFF, take the next 32 bytes as a parameter value and write it to the cache.
If the first byte is 0xFE, take the next 32 bytes as a parameter value but do not write it to the cache.
For any other value, take the top four bits as the number of additional bytes, and the bottom four bits as the most significant bits of the cache key. Here are some examples:
Bytes in calldata Cache key
0x0F 0x0F
0x10,0x10 0x10
0x12,0xAC 0x02AC
0x2D,0xEA, 0xD6 0x0DEAD6

Bytes in calldata	Cache key
0x0F	0x0F
0x10,0x10	0x10
0x12,0xAC	0x02AC
0x2D,0xEA, 0xD6	0x0DEAD6

Cache manipulation

The cache is implemented in Cache.sol(opens in a new tab). Let's go over it line by line.

1// SPDX-License-Identifier: UNLICENSED
2pragma solidity ^0.8.13;
3
4
5contract Cache {
6
7    bytes1 public constant INTO_CACHE = 0xFF;
8    bytes1 public constant DONT_CACHE = 0xFE;
 Hwɛ so kyerɛw bi

These constants are used to interpret the special cases where we provide all the information and either want it written into the cache or not. Writing into the cache requires two SSTORE(opens in a new tab) operations into previously unused storage slots at a cost of 22100 gas each, so we make it optional.

1
2    mapping(uint => uint) public val2key;
 Hwɛ so kyerɛw bi

A mapping(opens in a new tab) between the values and their keys. This information is necessary to encode values before you send out the transaction.

1    // Location n has the value for key n+1, because we need to preserve
2    // zero as "not in the cache".
3    uint[] public key2val;
 Hwɛ so kyerɛw bi

We can use an array for the mapping from keys to values because we assign the keys, and for simplicity we do it sequentially.

1    function cacheRead(uint _key) public view returns (uint) {
2        require(_key <= key2val.length, "Reading uninitialize cache entry");
3        return key2val[_key-1];
4    }  // cacheRead
 Hwɛ so kyerɛw bi

Read a value from the cache.

1    // Write a value to the cache if it's not there already
2    // Only public to enable the test to work
3    function cacheWrite(uint _value) public returns (uint) {
4        // If the value is already in the cache, return the current key
5        if (val2key[_value] != 0) {
6            return val2key[_value];
7        }
 Hwɛ so kyerɛw bi

There is no point in putting the same value in the cache more than once. If the value is already there, just return the existing key.

1        // Since 0xFE is a special case, the largest key the cache can
2        // hold is 0x0D followed by 15 0xFF's. If the cache length is already that
3        // large, fail.
4        //                              1 2 3 4 5 6 7 8 9 A B C D E F
5        require(key2val.length+1 < 0x0DFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF,
6            "cache overflow");
 Hwɛ so kyerɛw bi

I don't think we'll ever get a cache that big (approximately 1.8*10³⁷ entries, which would require about 10²⁷ TB to store). However, I'm old enough to remember "640kB would always be enough"(opens in a new tab). This test is very cheap.

1        // Write the value using the next key
2        val2key[_value] = key2val.length+1;
 Hwɛ so kyerɛw bi

Add the reverse lookup (from the value to the key).

1        key2val.push(_value);
 Hwɛ so kyerɛw bi

Add the forward lookup (from the key to the value). Because we assign values sequentially we can just add it after the last array value.

1        return key2val.length;
2    }  // cacheWrite
 Hwɛ so kyerɛw bi

Return the new length of key2val, which is the cell where the new value is stored.

1    function _calldataVal(uint startByte, uint length)
2        private pure returns (uint)
 Hwɛ so kyerɛw bi

This function reads a value from the calldata of arbitrary length (up to 32 bytes, the word size).

1    {
2        uint _retVal;
3
4        require(length < 0x21,
5            "_calldataVal length limit is 32 bytes");
6        require(length + startByte <= msg.data.length,
7            "_calldataVal trying to read beyond calldatasize");
 Hwɛ so kyerɛw bi

This function is internal, so if the rest of the code is written correctly these tests are not required. However, they don't cost much so we might as well have them.

1        assembly {
2            _retVal := calldataload(startByte)
3        }
 Hwɛ so kyerɛw bi

This code is in Yul(opens in a new tab). It reads a 32 byte value from the calldata. This works even if the calldata stops before startByte+32 because uninitialized space in EVM is considered to be zero.

1        _retVal = _retVal >> (256-length*8);
 Hwɛ so kyerɛw bi

We don't necessarily want a 32 byte value. This gets rid of the excess bytes.

1        return _retVal;
2    } // _calldataVal
3
4
5    // Read a single parameter from the calldata, starting at _fromByte
6    function _readParam(uint _fromByte) internal
7        returns (uint _nextByte, uint _parameterValue)
8    {
 Hwɛ so kyerɛw bi

Read a single parameter from the calldata. Note that we need to return not just the value we read, but also the location of the next byte because parameters can range from 1 byte long to 33 bytes.

1        // The first byte tells us how to interpret the rest
2        uint8 _firstByte;
3
4        _firstByte = uint8(_calldataVal(_fromByte, 1));
 Hwɛ so kyerɛw bi

Solidity tries to reduce the number of bugs by forbidding potentially dangerous implicit type conversions(opens in a new tab). A downgrade, for example from 256 bits to 8 bits, needs to be explicit.

1
2        // Read the value, but do not write it to the cache
3        if (_firstByte == uint8(DONT_CACHE))
4            return(_fromByte+33, _calldataVal(_fromByte+1, 32));
5
6        // Read the value, and write it to the cache
7        if (_firstByte == uint8(INTO_CACHE)) {
8            uint _param = _calldataVal(_fromByte+1, 32);
9            cacheWrite(_param);
10            return(_fromByte+33, _param);
11        }
12
13        // If we got here it means that we need to read from the cache
14
15        // Number of extra bytes to read
16        uint8 _extraBytes = _firstByte / 16;
Kyerɛ biribiara
 Hwɛ so kyerɛw bi

Take the lower nibble(opens in a new tab) and combine it with the other bytes to read the value from the cache.

1        uint _key = (uint256(_firstByte & 0x0F) << (8*_extraBytes)) +
2            _calldataVal(_fromByte+1, _extraBytes);
3
4        return (_fromByte+_extraBytes+1, cacheRead(_key));
5
6    }  // _readParam
7
8
9    // Read n parameters (functions know how many parameters they expect)
10    function _readParams(uint _paramNum) internal returns (uint[] memory) {
Kyerɛ biribiara
 Hwɛ so kyerɛw bi

We could get the number of parameters we have from the calldata itself, but the functions that call us know how many parameters they expect. It's easier to let them tell us.

1        // The parameters we read
2        uint[] memory params = new uint[](_paramNum);
3
4        // Parameters start at byte 4, before that it's the function signature
5        uint _atByte = 4;
6
7        for(uint i=0; i<_paramNum; i++) {
8            (_atByte, params[i]) = _readParam(_atByte);
9        }
Kyerɛ biribiara
 Hwɛ so kyerɛw bi

Read the parameters until you have the number you need. If we go past the end of the calldata, _readParams will revert the call.

1
2        return(params);
3    }   // readParams
4
5    // For testing _readParams, test reading four parameters
6    function fourParam() public
7        returns (uint256,uint256,uint256,uint256)
8    {
9        uint[] memory params;
10        params = _readParams(4);
11        return (params[0], params[1], params[2], params[3]);
12    }    // fourParam
Kyerɛ biribiara
 Hwɛ so kyerɛw bi

One big advantage of Foundry is that it allows tests to be written in Solidity (see Testing the cache below). This makes unit tests a lot easier. This is a function that reads four parameters and returns them so the test can verify they were correct.

1    // Get a value, return bytes that will encode it (using the cache if possible)
2    function encodeVal(uint _val) public view returns(bytes memory) {
 Hwɛ so kyerɛw bi

encodeVal is a function that off-chain code calls to help create calldata that uses the cache. It receives a single value and returns the bytes that encode it. This function is a view, so it does not require a transaction and when called externally does not cost any gas.

1        uint _key = val2key[_val];
2
3        // The value isn't in the cache yet, add it
4        if (_key == 0)
5            return bytes.concat(INTO_CACHE, bytes32(_val));
 Hwɛ so kyerɛw bi

In the EVM all uninitialized storage is assumed to be zeros. So if we look for the key for a value that isn't there, we get a zero. In that case the bytes that encode it are INTO_CACHE (so it will be cached next time), followed by the actual value.

1        // If the key is <0x10, return it as a single byte
2        if (_key < 0x10)
3            return bytes.concat(bytes1(uint8(_key)));
 Hwɛ so kyerɛw bi

Single bytes are the easiest. We just use bytes.concat(opens in a new tab) to turn a bytes<n> type into a byte array which can be any length. Despite the name, it works fine when provided with just one argument.

1        // Two byte value, encoded as 0x1vvv
2        if (_key < 0x1000)
3            return bytes.concat(bytes2(uint16(_key) | 0x1000));
 Hwɛ so kyerɛw bi

When we have a key that is less than 16³, we can express it in two bytes. We first convert _key, which is a 256 bit value, to a 16 bit value and use logical or to add the number of extra bytes to the first byte. Then we just it into a bytes2 value, which can be converted to bytes.

1        // There is probably a clever way to do the following lines as a loop,
2        // but it's a view function so I'm optimizing for programmer time and
3        // simplicity.
4
5        if (_key < 16*256**2)
6            return bytes.concat(bytes3(uint24(_key) | (0x2 * 16 * 256**2)));
7        if (_key < 16*256**3)
8            return bytes.concat(bytes4(uint32(_key) | (0x3 * 16 * 256**3)));
9             .
10             .
11             .
12        if (_key < 16*256**14)
13            return bytes.concat(bytes15(uint120(_key) | (0xE * 16 * 256**14)));
14        if (_key < 16*256**15)
15            return bytes.concat(bytes16(uint128(_key) | (0xF * 16 * 256**15)));
Kyerɛ biribiara
 Hwɛ so kyerɛw bi

The other values (3 bytes, 4 bytes, etc.) are handled the same way, just with different field sizes.

1        // If we get here, something is wrong.
2        revert("Error in encodeVal, should not happen");
 Hwɛ so kyerɛw bi

If we get here it means we got a key that's not less than 16*256¹⁵. But cacheWrite limits the keys so we can't even get up to 14*256¹⁶ (which would have a first byte of 0xFE, so it would look like DONT_CACHE). But it doesn't cost us much to add a test in case a future programmer introduces a bug.

1    } // encodeVal
2
3}  // Cache
 Hwɛ so kyerɛw bi

Testing the cache

One of the advantages of Foundry is that it lets you write tests in Solidity(opens in a new tab), which makes it easier to write unit tests. The tests for the Cache class are here(opens in a new tab). Because the testing code is repetitive, as tests tend to be, this article only explains the interesting parts.

1// SPDX-License-Identifier: UNLICENSED
2pragma solidity ^0.8.13;
3
4import "forge-std/Test.sol";
5
6
7// Need to run `forge test -vv` for the console.
8import "forge-std/console.sol";
 Hwɛ so kyerɛw bi

This is just boilerplate that is necessary to use the test package and console.log.

1import "src/Cache.sol";
 Hwɛ so kyerɛw bi

We need to know the contract we are testing.

1contract CacheTest is Test {
2    Cache cache;
3
4    function setUp() public {
5        cache = new Cache();
6    }
 Hwɛ so kyerɛw bi

The setUp function is called before each test. In this case we just create a new cache, so that our tests won't affect each other.

1    function testCaching() public {
 Hwɛ so kyerɛw bi

Tests are functions whose names start with test. This function checks the basic cache functionality, writing values and reading them again.

1        for(uint i=1; i<5000; i++) {
2            cache.cacheWrite(i*i);
3        }
4
5        for(uint i=1; i<5000; i++) {
6            assertEq(cache.cacheRead(i), i*i);
 Hwɛ so kyerɛw bi

This is how you do the actual testing, using assert... functions(opens in a new tab). In this case, we check that the value we wrote is the one we read. We can discard the result of cache.cacheWrite because we know that cache keys are assigned linearly.

1        }
2    }    // testCaching
3
4
5    // Cache the same value multiple times, ensure that the key stays
6    // the same
7    function testRepeatCaching() public {
8        for(uint i=1; i<100; i++) {
9            uint _key1 = cache.cacheWrite(i);
10            uint _key2 = cache.cacheWrite(i);
11            assertEq(_key1, _key2);
12        }
Kyerɛ biribiara
 Hwɛ so kyerɛw bi

First we write each value twice to the cache and make sure the keys are the same (meaning the second write didn't really happen).

1        for(uint i=1; i<100; i+=3) {
2            uint _key = cache.cacheWrite(i);
3            assertEq(_key, i);
4        }
5    }    // testRepeatCaching
 Hwɛ so kyerɛw bi

In theory there could be a bug that doesn't affect consecutive cache writes. So here we do some writes that aren't consecutive and see the values are still not rewritten.

1    // Read a uint from a memory buffer (to make sure we get back the parameters
2    // we sent out)
3    function toUint256(bytes memory _bytes, uint256 _start) internal pure
4        returns (uint256)
 Hwɛ so kyerɛw bi

Read a 256 bit word from a bytes memory buffer. This utility function lets us verify that we receive the correct results when we run a function call that uses the cache.

1    {
2        require(_bytes.length >= _start + 32, "toUint256_outOfBounds");
3        uint256 tempUint;
4
5        assembly {
6            tempUint := mload(add(add(_bytes, 0x20), _start))
7        }
 Hwɛ so kyerɛw bi

Yul does not support data structures beyond uint256, so when you refer to a more sophisticated data structure, such as the memory buffer _bytes, you get the address of that structure. Solidity stores bytes memory values as a 32 byte word that contains the length, followed by the actual bytes, so to get byte number _start we need to calculate _bytes+32+_start.

1
2        return tempUint;
3    }     // toUint256
4
5    // Function signature for fourParams(), courtesy of
6    // https://www.4byte.directory/signatures/?bytes4_signature=0x3edc1e6d
7    bytes4 constant FOUR_PARAMS = 0x3edc1e6d;
8
9    // Just some constant values to see we're getting the correct values back
10    uint256 constant VAL_A = 0xDEAD60A7;
11    uint256 constant VAL_B =     0xBEEF;
12    uint256 constant VAL_C =     0x600D;
13    uint256 constant VAL_D = 0x600D60A7;
Kyerɛ biribiara
 Hwɛ so kyerɛw bi

Some constants we need for testing.

1    function testReadParam() public {
 Hwɛ so kyerɛw bi

Call fourParams(), a function that uses readParams, to test we can read parameters correctly.

1        address _cacheAddr = address(cache);
2        bool _success;
3        bytes memory _callInput;
4        bytes memory _callOutput;
 Hwɛ so kyerɛw bi

We can't use the normal ABI mechanism to call a function using the cache, so we need to use the low level <address>.call()(opens in a new tab) mechanism. That mechanism takes a bytes memory as input, and returns that (as well as a Boolean value) as output.

1        // First call, the cache is empty
2        _callInput = bytes.concat(
3            FOUR_PARAMS,
 Hwɛ so kyerɛw bi

It is useful for the same contract to support both cached functions (for calls directly from transactions) and non-cached functions (for calls from other smart contracts). To do that we need to continue to rely on the Solidity mechanism to call the correct function, instead of putting everything in a fallback function(opens in a new tab). Doing this makes composability a lot easier. A single byte would be enough to identify the function in most cases, so we are wasting three bytes (16*3=48 gas). However, as I'm writing this those 48 gas cost 0.07 cents, which is a reasonable cost of simpler, less bug prone, code.

1            // First value, add it to the cache
2            cache.INTO_CACHE(),
3            bytes32(VAL_A),
 Hwɛ so kyerɛw bi

The first value: A flag saying it's a full value that needs to be written to the cache, followed by the 32 bytes of the value. The other three values are similar, except that VAL_B isn't written to the cache and VAL_C is both the third parameter and the fourth one.

1             .
2             .
3             .
4        );
5        (_success, _callOutput) = _cacheAddr.call(_callInput);
 Hwɛ so kyerɛw bi

This is where we actually call the Cache contract.

1        assertEq(_success, true);
 Hwɛ so kyerɛw bi

We expect the call to be successful.

1        assertEq(cache.cacheRead(1), VAL_A);
2        assertEq(cache.cacheRead(2), VAL_C);
 Hwɛ so kyerɛw bi

We start with an empty cache and then add VAL_A followed by VAL_C. We'd expect the first one to have the key 1, and the second one to have 2.

1        assertEq(toUint256(_callOutput,0), VAL_A);
2        assertEq(toUint256(_callOutput,32), VAL_B);
3        assertEq(toUint256(_callOutput,64), VAL_C);
4        assertEq(toUint256(_callOutput,96), VAL_C);

The output is the four parameters. Here we verify it is correct.

1        // Second call, we can use the cache
2        _callInput = bytes.concat(
3            FOUR_PARAMS,
4
5            // First value in the Cache
6            bytes1(0x01),
 Hwɛ so kyerɛw bi

Cache keys below 16 are just one byte.

1            // Second value, don't add it to the cache
2            cache.DONT_CACHE(),
3            bytes32(VAL_B),
4
5            // Third and fourth values, same value
6            bytes1(0x02),
7            bytes1(0x02)
8        );
9        .
10        .
11        .
12    }   // testReadParam
Kyerɛ biribiara
 Hwɛ so kyerɛw bi

The tests after the call are identical to those after the first call.

1    function testEncodeVal() public {
 Hwɛ so kyerɛw bi

This function is similar to testReadParam, except that instead of writing the parameters explicitly we use encodeVal().

1        .
2        .
3        .
4        _callInput = bytes.concat(
5            FOUR_PARAMS,
6            cache.encodeVal(VAL_A),
7            cache.encodeVal(VAL_B),
8            cache.encodeVal(VAL_C),
9            cache.encodeVal(VAL_D)
10        );
11        .
12        .
13        .
14        assertEq(_callInput.length, 4+1*4);
15    }   // testEncodeVal
Kyerɛ biribiara
 Hwɛ so kyerɛw bi

The only additional test in testEncodeVal() is to verify that the length of _callInput is correct. For the first call it is 4+33*4. For the second, where every value is already in the cache, it is 4+1*4.

1    // Test encodeVal when the key is more than a single byte
2    // Maximum three bytes because filling the cache to four bytes takes
3    // too long.
4    function testEncodeValBig() public {
5        // Put a number of values in the cache.
6        // To keep things simple, use key n for value n.
7        for(uint i=1; i<0x1FFF; i++) {
8            cache.cacheWrite(i);
9        }
Kyerɛ biribiara
 Hwɛ so kyerɛw bi

The testEncodeVal function above only writes four values into the cache, so the part of the function that deals with multi-byte values(opens in a new tab) is not checked. But that code is complicated and error-prone.

The first part of this function is a loop that writes all the values from 1 to 0x1FFF to the cache in order, so we'll be able to encode those values and know where they are going.

1        .
2        .
3        .
4
5        _callInput = bytes.concat(
6            FOUR_PARAMS,
7            cache.encodeVal(0x000F),   // One byte        0x0F
8            cache.encodeVal(0x0010),   // Two bytes     0x1010
9            cache.encodeVal(0x0100),   // Two bytes     0x1100
10            cache.encodeVal(0x1000)    // Three bytes 0x201000
11        );
Kyerɛ biribiara
 Hwɛ so kyerɛw bi

Test one byte, two byte, and three byte values. We don't test beyond that because it would take too long to write enough stack entries (at least 0x10000000, approximately a quarter of a billion).

1        .
2        .
3        .
4        .
5    }    // testEncodeValBig
6
7
8    // Test what with an excessively small buffer we get a revert
9    function testShortCalldata() public {
Kyerɛ biribiara
 Hwɛ so kyerɛw bi

Test what happens in the abnormal case where there aren't enough parameters.

1        .
2        .
3        .
4        (_success, _callOutput) = _cacheAddr.call(_callInput);
5        assertEq(_success, false);
6    }   // testShortCalldata
 Hwɛ so kyerɛw bi

Since it reverts, the result we should get is false.

1    // Call with cache keys that aren't there
2    function testNoCacheKey() public {
3        .
4        .
5        .
6        _callInput = bytes.concat(
7            FOUR_PARAMS,
8
9            // First value, add it to the cache
10            cache.INTO_CACHE(),
11            bytes32(VAL_A),
12
13            // Second value
14            bytes1(0x0F),
15            bytes2(0x1234),
16            bytes11(0xA10102030405060708090A)
17        );
Kyerɛ biribiara

This function gets four perfectly legitimate parameters, except that the cache is empty so there are no values there to read.

1        .
2        .
3        .
4    // Test what with an excessively long buffer everything works file
5    function testLongCalldata() public {
6        address _cacheAddr = address(cache);
7        bool _success;
8        bytes memory _callInput;
9        bytes memory _callOutput;
10
11        // First call, the cache is empty
12        _callInput = bytes.concat(
13            FOUR_PARAMS,
14
15            // First value, add it to the cache
16            cache.INTO_CACHE(), bytes32(VAL_A),
17
18            // Second value, add it to the cache
19            cache.INTO_CACHE(), bytes32(VAL_B),
20
21            // Third value, add it to the cache
22            cache.INTO_CACHE(), bytes32(VAL_C),
23
24            // Fourth value, add it to the cache
25            cache.INTO_CACHE(), bytes32(VAL_D),
26
27            // And another value for "good luck"
28            bytes4(0x31112233)
29        );
Kyerɛ biribiara
 Hwɛ so kyerɛw bi

This function sends five values. We know that the fifth value is ignored because it is not a valid cache entry, which would have caused a revert had it not been included.

1        (_success, _callOutput) = _cacheAddr.call(_callInput);
2        assertEq(_success, true);
3        .
4        .
5        .
6    }   // testLongCalldata
7
8}        // CacheTest
9
Kyerɛ biribiara
 Hwɛ so kyerɛw bi

A sample application

Writing tests in Solidity is all very well, but at the end of the day a dapp needs to be able to process requests from outside the chain to be useful. This article demonstrates how to use caching in a dapp with WORM, which stands for "Write Once, Read Many". If a key is not yet written, you can write a value to it. If the key is written already, you get a revert.

The contract

This is the contract(opens in a new tab). It mostly repeats what we have already done with Cache and CacheTest, so we only cover the parts that are interesting.

1import "./Cache.sol";
2
3contract WORM is Cache {
 Hwɛ so kyerɛw bi

The easiest way to use Cache is to inherit it in our own contract.

1    function writeEntryCached() external {
2        uint[] memory params = _readParams(2);
3        writeEntry(params[0], params[1]);
4    }    // writeEntryCached
 Hwɛ so kyerɛw bi

This function is similar to fourParam in CacheTest above. Because we don't follow the ABI specifications, it is best not to declare any parameters into the function.

1    // Make it easier to call us
2    // Function signature for writeEntryCached(), courtesy of
3    // https://www.4byte.directory/signatures/?bytes4_signature=0xe4e4f2d3
4    bytes4 constant public WRITE_ENTRY_CACHED = 0xe4e4f2d3;
 Hwɛ so kyerɛw bi

The external code that calls writeEntryCached will need to manually build the calldata, instead of using worm.writeEntryCached, because we do not follow the ABI specifications. Having this constant value just makes it easier to write it.

Note that even though we define WRITE_ENTRY_CACHED as a state variable, to read it externally it is necessary to use the getter function for it, worm.WRITE_ENTRY_CACHED().

1    function readEntry(uint key) public view
2        returns (uint _value, address _writtenBy, uint _writtenAtBlock)
 Hwɛ so kyerɛw bi

The read function is a view, so it does not require a transaction and does not cost gas. As a result, there is no benefit to using the cache for the parameter. With view functions it is best to use the standard mechanism that is simpler.

The testing code

This is the testing code for the contract(opens in a new tab). Again, let us look only at what is interesting.

1    function testWReadWrite() public {
2        worm.writeEntry(0xDEAD, 0x60A7);
3
4        vm.expectRevert(bytes("entry already written"));
5        worm.writeEntry(0xDEAD, 0xBEEF);
 Hwɛ so kyerɛw bi

This (vm.expectRevert)(opens in a new tab) is how we specify in a Foundry test that the next call should fail, and the reported reason for a failure. This applies when we use the syntax <contract>.<function name>() rather than building the calldata and calling the contract using the low level interface (<contract>.call(), etc.).

1    function testReadWriteCached() public {
2        uint cacheGoat = worm.cacheWrite(0x60A7);
 Hwɛ so kyerɛw bi

Here we use the fact that cacheWrite returns the cache key. This is not something we'd expect to use in production, because cacheWrite changes the state, and therefore can only be called during a transaction. Transactions don't have return values, if they have results those results are supposed to be emitted as events. So the cacheWrite return value is only accessible from on-chain code, and on-chain code does not need parameter caching.

1        (_success,) = address(worm).call(_callInput);
 Hwɛ so kyerɛw bi

This is how we tell Solidity that while <contract address>.call() has two return values, we only care about the first.

1        (_success,) = address(worm).call(_callInput);
2        assertEq(_success, false);
 Hwɛ so kyerɛw bi

Since we use the low level <address>.call() function, we can't use vm.expectRevert() and have to look at the boolean success value we get from the call.

1    event EntryWritten(uint indexed key, uint indexed value);
2
3        .
4        .
5        .
6
7        _callInput = bytes.concat(
8            worm.WRITE_ENTRY_CACHED(), worm.encodeVal(a), worm.encodeVal(b));
9        vm.expectEmit(true, true, false, false);
10        emit EntryWritten(a, b);
11        (_success,) = address(worm).call(_callInput);
Kyerɛ biribiara
 Hwɛ so kyerɛw bi

This is the way we verify that code emits an event correctly(opens in a new tab) in Foundry.

The client

One thing you don't get with Solidity tests is JavaScript code you can cut and paste into your own application. To write that code I deployed WORM to Optimism Goerli(opens in a new tab), Optimism's(opens in a new tab) new testnet. It is at address 0xd34335b1d818cee54e3323d3246bd31d94e6a78a(opens in a new tab).

You can see JavaScript code for the client here(opens in a new tab). To use it:

Clone the git repository:

1git clone https://github.com/qbzzt/20220915-all-you-can-cache.git

Install the necessary packages:
```
1cd javascript
2yarn
```
Copy the configuration file:
```
1cp .env.example .env
```

Edit .env for your configuration:

Parameter	Value
MNEMONIC	The mnemonic for an account that has enough ETH to pay for a transaction. You can get free ETH for the Optimism Goerli network here(opens in a new tab).
OPTIMISM_GOERLI_URL	URL to Optimism Goerli. The public endpoint, `https://goerli.optimism.io`, is rate limited but sufficient to what we need here

Run index.js.
```
1node index.js
```
This sample application first writes an entry to WORM, displaying the calldata and a link to the transaction on Etherscan. Then it reads back that entry, and displays the key it uses and the values in the entry (value, block number, and author).

Most of the client is normal Dapp JavaScript. So again we'll only go over the interesting parts.

1.
2.
3.
4const main = async () => {
5    const func = await worm.WRITE_ENTRY_CACHED()
6
7    // Need a new key every time
8    const key = await worm.encodeVal(Number(new Date()))

A given slot can only be written into once, so we use the timestamp to make sure we don't reuse slots.

1const val = await worm.encodeVal("0x600D")
2
3// Write an entry
4const calldata = func + key.slice(2) + val.slice(2)

Ethers expects the call data to be a hex string, 0x followed by an even number of hexadecimal digits. As key and val both start with 0x, we need to remove those headers.

1const tx = await worm.populateTransaction.writeEntryCached()
2tx.data = calldata
3
4sentTx = await wallet.sendTransaction(tx)

As with the Solidity testing code, we cannot call a cached function normally. Instead, we need to use a lower level mechanism.

1    .
2    .
3    .
4    // Read the entry just written
5    const realKey = '0x' + key.slice(4)  // remove the FF flag
6    const entryRead = await worm.readEntry(realKey)
7    .
8    .
9    .
Kyerɛ biribiara

For reading entries we can use the normal mechanism. There's no need to use parameter caching with view functions.

Conclusion

The code in this article is a proof of concept, the purpose is to make the idea easy to understand. For a production-ready system you might want to implement some additional functionality:

Handle values that aren't uint256. For example, strings.
Instead of a global cache, maybe have a mapping between users and caches. Different users use different values.
Values used for addresses are distinct from those used for other purposes. It might make sense to have a separate cache just for addresses.
Currently, the cache keys are on a "first come, smallest key" algorithm. The first sixteen values can be sent as a single byte. The next 4080 values can be sent as two bytes. The next approximately million values are three bytes, etc. A production system should keep usage counters on cache entries and reorganize them so that the sixteen most common values are one byte, the next 4080 most common values two bytes, etc.
However, that is a potentially dangerous operation. Imagine the following sequence of events:
1. Noam Naive calls encodeVal to encode the address to which he wants to send tokens. That address is one of the first used on the application, so the encoded value is 0x06. This is a view function, not a transaction, so it's between Noam and the node he uses, and nobody else knows about it
2. Owen Owner runs the cache reordering operation. Very few people actually use that address, so it is now encoded as 0x201122. A different value, 10¹⁸, is assigned 0x06.
3. Noam Naive sends his tokens to 0x06. They go to the address 0x0000000000000000000000000de0b6b3a7640000, and since nobody knows the private key for that address, they are just stuck there. Noam is not happy.
There are ways to solve this problem, and the related problem of transactions that are in the mempool during the cache reorder, but you must be aware of it.

I demonstrated caching here with Optimism, because I'm an Optimism employee and this is the rollup I know best. But it should work with any rollup that charges a minimal cost for internal processing, so that in comparison writing the transaction data to L1 is the major expense.

Nsakraeɛ edi akyire: @nhsz(opens in a new tab), August 15, 2023