bal: lazy scope allocation in access list builder (#34019)

## Summary - Avoid allocating a `map[common.Address]*constructionAccountAccess` on every `enterScope()` call in the BAL access list builder - Push `nil` onto the stack and lazily allocate via `currentScope()` only when a state change occurs - `exitScope()` short-circuits when child scope is nil (no merge needed) - Adds correctness test (`TestLazyScopeCorrectness`) and benchmark (`BenchmarkPrecompileScopes`) ## Context The [execution-specs benchmark tests](https://github.com/ethereum/execution-specs/blob/tests-benchmark@v0.0.7/tests/benchmark/compute/precompile/test_alt_bn128.py) for `bn128_add` (line 19) and `bn128_pairing` (line 496) create blocks with transactions that loop `STATICCALL` to precompiles thousands of times. Each STATICCALL triggers `EnterScope()`/`ExitScope()` in the BAL tracer, but precompile calls produce empty scopes since they don't touch state (`AddBalance(0)` is filtered by `!amount.IsZero()`). The previous implementation eagerly allocated a map on every `enterScope()`, creating ~200K unnecessary heap allocations per transaction for `bn128_add` (150 gas, ~200K calls per 30M gas block). **Note:** For benchmark-only workloads that don't need BAL validation, `--bal.executionmode=sequential` can be used to skip the parallel processor entirely and avoid all tracing overhead. ## Benchmark results Precompile calls with no state changes per scope: | Calls | Before (ns/op) | After (ns/op) | Speedup | Allocs Before → After | |---|---|---|---|---| | 100 | 5,073 | 1,499 | **3.4x** | 119 → 19 | | 1,000 | 47,392 | 3,199 | **14.8x** | 1,019 → 19 | | 10,000 | 493,651 | 20,644 | **23.9x** | 10,019 → 19 | | 100,000 | 4,526,165 | 192,502 | **23.5x** | 100,019 → 19 | ## Test plan - [x] `TestLazyScopeCorrectness` — mixed workload: precompile scopes + state-changing scopes + reverted scopes - [x] `BenchmarkPrecompileScopes` — measures scope tracking overhead for 100 to 100K empty scopes - [x] All `core/vm/` tests pass - [x] All `core/types/bal/` tests pass (pre-existing failures in `TestBALEncoding` and `TestBlockAccessListValidation` unrelated) 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-03-20 01:50:40 +00:00 · 2026-03-17 14:08:06 +01:00 · 2026-03-17 14:08:06 +01:00 · ff772bfa08
commit ff772bfa08
parent 956f55fe83
2 changed files with 167 additions and 29 deletions
--- a/core/types/bal/bal.go
+++ b/core/types/bal/bal.go
@ -50,17 +50,30 @@ func newAccessListBuilder() *idxAccessListBuilder {
 	}
 }

-func (c *idxAccessListBuilder) storageRead(address common.Address, key common.Hash) {
-	if _, ok := c.accessesStack[len(c.accessesStack)-1][address]; !ok {
-		c.accessesStack[len(c.accessesStack)-1][address] = &constructionAccountAccess{}
+// currentScope returns the current (topmost) scope, lazily allocating if needed.
+// This avoids allocating a map for scopes that never record any state changes,
+// which is critical for precompile-heavy blocks where STATICCALL creates/destroys
+// thousands of empty scopes per transaction.
+func (c *idxAccessListBuilder) currentScope() map[common.Address]*constructionAccountAccess {
+	top := len(c.accessesStack) - 1
+	if c.accessesStack[top] == nil {
+		c.accessesStack[top] = make(map[common.Address]*constructionAccountAccess)
 	}
-	acctAccesses := c.accessesStack[len(c.accessesStack)-1][address]
-	acctAccesses.StorageRead(key)
+	return c.accessesStack[top]
+}
+
+func (c *idxAccessListBuilder) storageRead(address common.Address, key common.Hash) {
+	scope := c.currentScope()
+	if _, ok := scope[address]; !ok {
+		scope[address] = &constructionAccountAccess{}
+	}
+	scope[address].StorageRead(key)
 }

 func (c *idxAccessListBuilder) accountRead(address common.Address) {
-	if _, ok := c.accessesStack[len(c.accessesStack)-1][address]; !ok {
-		c.accessesStack[len(c.accessesStack)-1][address] = &constructionAccountAccess{}
+	scope := c.currentScope()
+	if _, ok := scope[address]; !ok {
+		scope[address] = &constructionAccountAccess{}
 	}
 }

@ -75,11 +88,11 @@ func (c *idxAccessListBuilder) storageWrite(address common.Address, key, prevVal
 		c.prestates[address].storage[key] = prevVal
 	}

-	if _, ok := c.accessesStack[len(c.accessesStack)-1][address]; !ok {
-		c.accessesStack[len(c.accessesStack)-1][address] = &constructionAccountAccess{}
+	scope := c.currentScope()
+	if _, ok := scope[address]; !ok {
+		scope[address] = &constructionAccountAccess{}
 	}
-	acctAccesses := c.accessesStack[len(c.accessesStack)-1][address]
-	acctAccesses.StorageWrite(key, prevVal, newVal)
+	scope[address].StorageWrite(key, prevVal, newVal)
 }

 func (c *idxAccessListBuilder) balanceChange(address common.Address, prev, cur *uint256.Int) {
@ -89,11 +102,11 @@ func (c *idxAccessListBuilder) balanceChange(address common.Address, prev, cur *
 	if c.prestates[address].balance == nil {
 		c.prestates[address].balance = prev
 	}
-	if _, ok := c.accessesStack[len(c.accessesStack)-1][address]; !ok {
-		c.accessesStack[len(c.accessesStack)-1][address] = &constructionAccountAccess{}
+	scope := c.currentScope()
+	if _, ok := scope[address]; !ok {
+		scope[address] = &constructionAccountAccess{}
 	}
-	acctAccesses := c.accessesStack[len(c.accessesStack)-1][address]
-	acctAccesses.BalanceChange(cur)
+	scope[address].BalanceChange(cur)
 }

 func (c *idxAccessListBuilder) codeChange(address common.Address, prev, cur []byte) {
@ -113,12 +126,11 @@ func (c *idxAccessListBuilder) codeChange(address common.Address, prev, cur []by
 		}
 		c.prestates[address].code = prev
 	}
-	if _, ok := c.accessesStack[len(c.accessesStack)-1][address]; !ok {
-		c.accessesStack[len(c.accessesStack)-1][address] = &constructionAccountAccess{}
+	scope := c.currentScope()
+	if _, ok := scope[address]; !ok {
+		scope[address] = &constructionAccountAccess{}
 	}
-	acctAccesses := c.accessesStack[len(c.accessesStack)-1][address]
-
-	acctAccesses.CodeChange(cur)
+	scope[address].CodeChange(cur)
 }

 // selfDestruct is invoked when an account which has been created and invoked
@ -127,7 +139,8 @@ func (c *idxAccessListBuilder) codeChange(address common.Address, prev, cur []by
 // Any storage accesses/modifications performed at the contract during execution
 // of the current call are retained in the block access list as state reads.
 func (c *idxAccessListBuilder) selfDestruct(address common.Address) {
-	access := c.accessesStack[len(c.accessesStack)-1][address]
+	scope := c.currentScope()
+	access := scope[address]
 	if len(access.storageMutations) != 0 && access.storageReads == nil {
 		access.storageReads = make(map[common.Hash]struct{})
 	}
@ -144,16 +157,19 @@ func (c *idxAccessListBuilder) nonceChange(address common.Address, prev, cur uin
 	if c.prestates[address].nonce == nil {
 		c.prestates[address].nonce = &prev
 	}
-	if _, ok := c.accessesStack[len(c.accessesStack)-1][address]; !ok {
-		c.accessesStack[len(c.accessesStack)-1][address] = &constructionAccountAccess{}
+	scope := c.currentScope()
+	if _, ok := scope[address]; !ok {
+		scope[address] = &constructionAccountAccess{}
 	}
-	acctAccesses := c.accessesStack[len(c.accessesStack)-1][address]
-	acctAccesses.NonceChange(cur)
+	scope[address].NonceChange(cur)
 }

 // enterScope is called after a new EVM call frame has been entered.
+// The scope map is lazily allocated by currentScope() only when a state
+// change occurs, avoiding heap allocations for precompile calls that
+// don't touch state.
 func (c *idxAccessListBuilder) enterScope() {
-	c.accessesStack = append(c.accessesStack, make(map[common.Address]*constructionAccountAccess))
+	c.accessesStack = append(c.accessesStack, nil)
 }

 // exitScope is called after an EVM call scope terminates.  If the call scope
@ -162,8 +178,14 @@ func (c *idxAccessListBuilder) enterScope() {
 // * mutated accounts/storage are added into the calling scope's access list as state accesses
 func (c *idxAccessListBuilder) exitScope(evmErr bool) {
 	childAccessList := c.accessesStack[len(c.accessesStack)-1]
-	parentAccessList := c.accessesStack[len(c.accessesStack)-2]
+	c.accessesStack = c.accessesStack[:len(c.accessesStack)-1]

+	// If no state was accessed in this scope, nothing to merge.
+	if childAccessList == nil {
+		return
+	}
+
+	parentAccessList := c.currentScope()
 	for addr, childAccess := range childAccessList {
 		if _, ok := parentAccessList[addr]; ok {
 		} else {
@ -177,8 +199,6 @@ func (c *idxAccessListBuilder) exitScope(evmErr bool) {
 			parentAccessList[addr].Merge(childAccess)
 		}
 	}
-
-	c.accessesStack = c.accessesStack[:len(c.accessesStack)-1]
 }

 // finalise returns the net state mutations at the access list index as well as
@ -188,6 +208,11 @@ func (a *idxAccessListBuilder) finalise() (*StateDiff, StateAccesses) {
 	diff := &StateDiff{make(map[common.Address]*AccountMutations)}
 	stateAccesses := make(StateAccesses)

+	// Root scope may be nil if no state changes occurred at all.
+	if a.accessesStack[0] == nil {
+		return diff, stateAccesses
+	}
+
 	for addr, access := range a.accessesStack[0] {
 		// remove any reported mutations from the access list with no net difference vs the index prestate value
 		if access.nonce != nil && *a.prestates[addr].nonce == *access.nonce {
--- a/core/types/bal/bal_test.go
+++ b/core/types/bal/bal_test.go
@ -19,6 +19,7 @@ package bal
 import (
 	"bytes"
 	"cmp"
+	"fmt"
 	"reflect"
 	"slices"
 	"testing"
@ -257,5 +258,117 @@ func TestBlockAccessListValidation(t *testing.T) {
 	}
 }

+// TestLazyScopeCorrectness verifies that lazy scope allocation produces
+// identical results to the previous eager allocation for mixed workloads:
+// precompile calls (empty scopes) interspersed with state-changing calls.
+func TestLazyScopeCorrectness(t *testing.T) {
+	builder := newAccessListBuilder()
+	sender := common.HexToAddress("0x1234")
+	contract := common.HexToAddress("0x5678")
+	precompile := common.HexToAddress("0x06")
+
+	// Tx-level: sender balance/nonce
+	builder.balanceChange(sender, uint256.NewInt(1000), uint256.NewInt(900))
+	builder.nonceChange(sender, 0, 1)
+
+	// Enter contract scope
+	builder.enterScope()
+	builder.storageRead(contract, common.HexToHash("0x01"))
+
+	// Precompile STATICCALL (empty scope)
+	builder.enterScope()
+	builder.exitScope(false)
+
+	// Another precompile call
+	builder.enterScope()
+	builder.exitScope(false)
+
+	// Contract writes storage
+	builder.storageWrite(contract, common.HexToHash("0x02"), common.Hash{}, common.HexToHash("0xff"))
+
+	// Precompile that reverts (still empty scope, reverted)
+	builder.enterScope()
+	builder.exitScope(true)
+
+	// Nested call to another contract
+	builder.enterScope()
+	builder.balanceChange(precompile, uint256.NewInt(0), uint256.NewInt(100))
+	builder.exitScope(false)
+
+	// Exit contract scope
+	builder.exitScope(false)
+
+	diff, accesses := builder.finalise()
+
+	// Verify sender mutations
+	senderMut, ok := diff.Mutations[sender]
+	if !ok {
+		t.Fatal("sender not in mutations")
+	}
+	if senderMut.Balance == nil || !senderMut.Balance.Eq(uint256.NewInt(900)) {
+		t.Fatalf("sender balance mismatch: got %v", senderMut.Balance)
+	}
+	if senderMut.Nonce == nil || *senderMut.Nonce != 1 {
+		t.Fatalf("sender nonce mismatch: got %v", senderMut.Nonce)
+	}
+
+	// Verify contract mutations (storage write)
+	contractMut, ok := diff.Mutations[contract]
+	if !ok {
+		t.Fatal("contract not in mutations")
+	}
+	if contractMut.StorageWrites == nil {
+		t.Fatal("contract has no storage writes")
+	}
+	if contractMut.StorageWrites[common.HexToHash("0x02")] != common.HexToHash("0xff") {
+		t.Fatal("contract storage write mismatch")
+	}
+
+	// Verify precompile balance change
+	precompileMut, ok := diff.Mutations[precompile]
+	if !ok {
+		t.Fatal("precompile not in mutations")
+	}
+	if precompileMut.Balance == nil || !precompileMut.Balance.Eq(uint256.NewInt(100)) {
+		t.Fatalf("precompile balance mismatch: got %v", precompileMut.Balance)
+	}
+
+	// Verify contract storage read is in accesses
+	contractAccesses, ok := accesses[contract]
+	if !ok {
+		t.Fatal("contract not in accesses")
+	}
+	if _, ok := contractAccesses[common.HexToHash("0x01")]; !ok {
+		t.Fatal("contract storage read not in accesses")
+	}
+}
+
+// BenchmarkPrecompileScopes simulates a precompile-heavy transaction where
+// STATICCALL is invoked thousands of times against a precompile (e.g. bn128_add).
+// Each call creates a scope (EnterScope) that records no state changes (precompiles
+// don't touch state), then exits (ExitScope). This benchmark measures the overhead
+// of scope tracking for such workloads.
+func BenchmarkPrecompileScopes(b *testing.B) {
+	for _, numCalls := range []int{100, 1000, 10000, 100000} {
+		b.Run(fmt.Sprintf("calls=%d", numCalls), func(b *testing.B) {
+			for i := 0; i < b.N; i++ {
+				builder := newAccessListBuilder()
+				// Simulate a transaction: sender balance/nonce change at depth 0,
+				// then thousands of precompile STATICCALL scopes that touch no state.
+				sender := common.HexToAddress("0x1234")
+				builder.balanceChange(sender, uint256.NewInt(1000), uint256.NewInt(900))
+				builder.nonceChange(sender, 0, 1)
+
+				for j := 0; j < numCalls; j++ {
+					builder.enterScope()
+					// Precompile call: no state hooks fire
+					builder.exitScope(false)
+				}
+				builder.finalise()
+			}
+		})
+	}
+}
+
 // BALReader test ideas
 // * BAL which doesn't have any pre-tx system contracts should return an empty state diff at idx 0