feat: EVM execution opcode level optimization

several optimizations included:
- opcode fusion
- jumpdest calculation optimization
- SHA caching in interpreter

Co-authored-by: redhdx <[email protected]>

cmd/evm/main.go

@@ -134,6 +134,11 @@ var (
 		Usage:    "enable return data output",
 		Category: flags.VMCategory,
 	}
+	VMOpcodeOptimizeFlag = &cli.BoolFlag{
+		Name:  "vm.opcode.optimize",
+		Usage: "enable opcode optimization",
+		Value: true,
+	}
 )
 
 var stateTransitionCommand = &cli.Command{

cmd/evm/runner.go

@@ -20,6 +20,7 @@ import (
 	"bytes"
 	"encoding/json"
 	"fmt"
+	compiler2 "github.com/ethereum/go-ethereum/core/opcodeCompiler/compiler"
 	"io"
 	"math/big"
 	"os"
@@ -224,7 +225,8 @@ func runCmd(ctx *cli.Context) error {
 		BlobHashes:  blobHashes,
 		BlobBaseFee: blobBaseFee,
 		EVMConfig: vm.Config{
-			Tracer: tracer,
+			Tracer:                    tracer,
+			EnableOpcodeOptimizations: ctx.Bool(VMOpcodeOptimizeFlag.Name),
 		},
 	}
 
@@ -234,6 +236,10 @@ func runCmd(ctx *cli.Context) error {
 		runtimeConfig.ChainConfig = params.AllEthashProtocolChanges
 	}
 
+	if runtimeConfig.EVMConfig.EnableOpcodeOptimizations {
+		compiler2.EnableOptimization()
+	}
+
 	var hexInput []byte
 	if inputFileFlag := ctx.String(InputFileFlag.Name); inputFileFlag != "" {
 		var err error

cmd/geth/main.go

@@ -156,6 +156,7 @@ var (
 		utils.RollupHaltOnIncompatibleProtocolVersionFlag,
 		utils.RollupSuperchainUpgradesFlag,
 		configFileFlag,
+		utils.VMOpcodeOptimizeFlag,
 	}, utils.NetworkFlags, utils.DatabaseFlags)
 
 	rpcFlags = []cli.Flag{

cmd/utils/flags.go

@@ -23,6 +23,7 @@ import (
 	"encoding/hex"
 	"errors"
 	"fmt"
+	"github.com/ethereum/go-ethereum/core/opcodeCompiler/compiler"
 	"math"
 	"math/big"
 	"net"
@@ -1050,6 +1051,13 @@ Please note that --` + MetricsHTTPFlag.Name + ` must be set to start the server.
 		Value:    metrics.DefaultConfig.InfluxDBOrganization,
 		Category: flags.MetricsCategory,
 	}
+
+	VMOpcodeOptimizeFlag = &cli.BoolFlag{
+		Name:     "vm.opcode.optimize",
+		Usage:    "enable opcode optimization",
+		Value:    true,
+		Category: flags.VMCategory,
+	}
 )
 
 var (
@@ -1901,6 +1909,12 @@ func SetEthConfig(ctx *cli.Context, stack *node.Node, cfg *ethconfig.Config) {
 		cfg.EnablePreimageRecording = ctx.Bool(VMEnableDebugFlag.Name)
 	}
 
+	if ctx.IsSet(VMOpcodeOptimizeFlag.Name) {
+		// TODO(fjl): force-enable this in --dev mode
+		cfg.EnableOpcodeOptimizing = ctx.Bool(VMOpcodeOptimizeFlag.Name)
+		compiler.EnableOptimization()
+	}
+
 	if ctx.IsSet(RPCGlobalGasCapFlag.Name) {
 		cfg.RPCGasCap = ctx.Uint64(RPCGlobalGasCapFlag.Name)
 	}
@@ -2387,8 +2401,12 @@ func MakeChain(ctx *cli.Context, stack *node.Node, readonly bool) (*core.BlockCh
 	if ctx.IsSet(CacheFlag.Name) || ctx.IsSet(CacheGCFlag.Name) {
 		cache.TrieDirtyLimit = ctx.Int(CacheFlag.Name) * ctx.Int(CacheGCFlag.Name) / 100
 	}
-	vmcfg := vm.Config{EnablePreimageRecording: ctx.Bool(VMEnableDebugFlag.Name)}
+	vmcfg := vm.Config{EnablePreimageRecording: ctx.Bool(VMEnableDebugFlag.Name),
+		EnableOpcodeOptimizations: ctx.Bool(VMOpcodeOptimizeFlag.Name)}
 
+	if vmcfg.EnableOpcodeOptimizations {
+		compiler.EnableOptimization()
+	}
 	// Disable transaction indexing/unindexing by default.
 	chain, err := core.NewBlockChain(chainDb, cache, gspec, nil, engine, vmcfg, nil, nil)
 	if err != nil {

core/opcodeCompiler/compiler/OpCodeCache.go

@@ -0,0 +1,51 @@
+package compiler
+
+import (
+	"github.com/ethereum/go-ethereum/common"
+	"github.com/ethereum/go-ethereum/common/lru"
+)
+
+type OpCodeCache struct {
+	optimizedCodeCache *lru.Cache[common.Hash, []byte]
+	bitvecCache        *lru.Cache[common.Hash, []byte]
+}
+
+func (c *OpCodeCache) GetCachedBitvec(codeHash common.Hash) []byte {
+	bitvec, _ := c.bitvecCache.Get(codeHash)
+	return bitvec
+}
+
+func (c *OpCodeCache) AddBitvecCache(codeHash common.Hash, bitvec []byte) {
+	c.bitvecCache.Add(codeHash, bitvec)
+}
+
+func (c *OpCodeCache) RemoveCachedCode(hash common.Hash) {
+	c.optimizedCodeCache.Remove(hash)
+}
+
+func (c *OpCodeCache) GetCachedCode(hash common.Hash) []byte {
+	processedCode, _ := c.optimizedCodeCache.Get(hash)
+	return processedCode
+}
+
+func (c *OpCodeCache) AddCodeCache(hash common.Hash, optimizedCode []byte) {
+	c.optimizedCodeCache.Add(hash, optimizedCode)
+}
+
+var opcodeCache *OpCodeCache
+
+const (
+	optimizedCodeCacheCap = 1024
+	bitvecCacheCap        = 1024
+)
+
+func init() {
+	opcodeCache = &OpCodeCache{
+		optimizedCodeCache: lru.NewCache[common.Hash, []byte](optimizedCodeCacheCap),
+		bitvecCache:        lru.NewCache[common.Hash, []byte](bitvecCacheCap),
+	}
+}
+
+func getOpCodeCacheInstance() *OpCodeCache {
+	return opcodeCache
+}

core/opcodeCompiler/compiler/evmByteCode.go

@@ -0,0 +1,225 @@
+package compiler
+
+// This is copied from vm/opcodes.go.
+
+// ByteCode is an EVM ByteCode
+type ByteCode byte
+
+// 0x0 range - arithmetic ops.
+const (
+	STOP       ByteCode = 0x0
+	ADD        ByteCode = 0x1
+	MUL        ByteCode = 0x2
+	SUB        ByteCode = 0x3
+	DIV        ByteCode = 0x4
+	SDIV       ByteCode = 0x5
+	MOD        ByteCode = 0x6
+	SMOD       ByteCode = 0x7
+	ADDMOD     ByteCode = 0x8
+	MULMOD     ByteCode = 0x9
+	EXP        ByteCode = 0xa
+	SIGNEXTEND ByteCode = 0xb
+)
+
+// 0x10 range - comparison ops.
+const (
+	LT     ByteCode = 0x10
+	GT     ByteCode = 0x11
+	SLT    ByteCode = 0x12
+	SGT    ByteCode = 0x13
+	EQ     ByteCode = 0x14
+	ISZERO ByteCode = 0x15
+	AND    ByteCode = 0x16
+	OR     ByteCode = 0x17
+	XOR    ByteCode = 0x18
+	NOT    ByteCode = 0x19
+	BYTE   ByteCode = 0x1a
+	SHL    ByteCode = 0x1b
+	SHR    ByteCode = 0x1c
+	SAR    ByteCode = 0x1d
+)
+
+// 0x20 range - crypto.
+const (
+	KECCAK256 ByteCode = 0x20
+)
+
+// 0x30 range - closure state.
+const (
+	ADDRESS        ByteCode = 0x30
+	BALANCE        ByteCode = 0x31
+	ORIGIN         ByteCode = 0x32
+	CALLER         ByteCode = 0x33
+	CALLVALUE      ByteCode = 0x34
+	CALLDATALOAD   ByteCode = 0x35
+	CALLDATASIZE   ByteCode = 0x36
+	CALLDATACOPY   ByteCode = 0x37
+	CODESIZE       ByteCode = 0x38
+	CODECOPY       ByteCode = 0x39
+	GASPRICE       ByteCode = 0x3a
+	EXTCODESIZE    ByteCode = 0x3b
+	EXTCODECOPY    ByteCode = 0x3c
+	RETURNDATASIZE ByteCode = 0x3d
+	RETURNDATACOPY ByteCode = 0x3e
+	EXTCODEHASH    ByteCode = 0x3f
+)
+
+// 0x40 range - block operations.
+const (
+	BLOCKHASH   ByteCode = 0x40
+	COINBASE    ByteCode = 0x41
+	TIMESTAMP   ByteCode = 0x42
+	NUMBER      ByteCode = 0x43
+	DIFFICULTY  ByteCode = 0x44
+	RANDOM      ByteCode = 0x44 // Same as DIFFICULTY
+	PREVRANDAO  ByteCode = 0x44 // Same as DIFFICULTY
+	GASLIMIT    ByteCode = 0x45
+	CHAINID     ByteCode = 0x46
+	SELFBALANCE ByteCode = 0x47
+	BASEFEE     ByteCode = 0x48
+)
+
+// 0x50 range - 'storage' and execution.
+const (
+	POP      ByteCode = 0x50
+	MLOAD    ByteCode = 0x51
+	MSTORE   ByteCode = 0x52
+	MSTORE8  ByteCode = 0x53
+	SLOAD    ByteCode = 0x54
+	SSTORE   ByteCode = 0x55
+	JUMP     ByteCode = 0x56
+	JUMPI    ByteCode = 0x57
+	PC       ByteCode = 0x58
+	MSIZE    ByteCode = 0x59
+	GAS      ByteCode = 0x5a
+	JUMPDEST ByteCode = 0x5b
+	PUSH0    ByteCode = 0x5f
+)
+
+// 0x60 range - pushes.
+const (
+	PUSH1 ByteCode = 0x60 + iota
+	PUSH2
+	PUSH3
+	PUSH4
+	PUSH5
+	PUSH6
+	PUSH7
+	PUSH8
+	PUSH9
+	PUSH10
+	PUSH11
+	PUSH12
+	PUSH13
+	PUSH14
+	PUSH15
+	PUSH16
+	PUSH17
+	PUSH18
+	PUSH19
+	PUSH20
+	PUSH21
+	PUSH22
+	PUSH23
+	PUSH24
+	PUSH25
+	PUSH26
+	PUSH27
+	PUSH28
+	PUSH29
+	PUSH30
+	PUSH31
+	PUSH32
+)
+
+// 0x80 range - dups.
+const (
+	DUP1 = 0x80 + iota
+	DUP2
+	DUP3
+	DUP4
+	DUP5
+	DUP6
+	DUP7
+	DUP8
+	DUP9
+	DUP10
+	DUP11
+	DUP12
+	DUP13
+	DUP14
+	DUP15
+	DUP16
+)
+
+// 0x90 range - swaps.
+const (
+	SWAP1 = 0x90 + iota
+	SWAP2
+	SWAP3
+	SWAP4
+	SWAP5
+	SWAP6
+	SWAP7
+	SWAP8
+	SWAP9
+	SWAP10
+	SWAP11
+	SWAP12
+	SWAP13
+	SWAP14
+	SWAP15
+	SWAP16
+)
+
+// 0xa0 range - logging ops.
+const (
+	LOG0 ByteCode = 0xa0 + iota
+	LOG1
+	LOG2
+	LOG3
+	LOG4
+)
+
+// 0xd0 range - customized instructions.
+const (
+	Nop ByteCode = 0xd0 + iota
+	AndSwap1PopSwap2Swap1
+	Swap2Swap1PopJump
+	Swap1PopSwap2Swap1
+	PopSwap2Swap1Pop
+	Push2Jump
+	Push2JumpI
+	Push1Push1
+	Push1Add
+	Push1Shl
+	Push1Dup1
+	Swap1Pop
+	PopJump
+	Pop2
+	Swap2Swap1
+	Swap2Pop
+	Dup2LT
+	JumpIfZero // 0xe2
+)
+
+// 0xf0 range - closures.
+const (
+	CREATE       ByteCode = 0xf0
+	CALL         ByteCode = 0xf1
+	CALLCODE     ByteCode = 0xf2
+	RETURN       ByteCode = 0xf3
+	DELEGATECALL ByteCode = 0xf4
+	CREATE2      ByteCode = 0xf5
+
+	STATICCALL   ByteCode = 0xfa
+	REVERT       ByteCode = 0xfd
+	INVALID      ByteCode = 0xfe
+	SELFDESTRUCT ByteCode = 0xff
+)
+
+// 0xb0 range.
+const (
+	TLOAD  ByteCode = 0xb3
+	TSTORE ByteCode = 0xb4
+)