Commit b1fa33d
perf(bal): verify-only fast path — skip end-of-block encode + Keccak (#11659)
* Add EIP-8037 and Amsterdam fixes
* Run all tests
* Align EIP-7928 BlockAccessIndex with uint32 spec, add validation (#11362)
* fix(bal): align EIP-7928 BlockAccessIndex with uint32 spec, add validation
Widens BlockAccessIndex from uint16 to uint32 per EIP-7928 (commit 645099785a)
and geth bal-devnet-4. Hardens the BAL decoder so truncated/malformed wire
bytes produce a clean RlpException at engine_newPayloadV5 instead of crashing.
Adds the missing validation rules geth bal-devnet-4 enforces: empty
storage_changes per slot is rejected, and BlockAccessIndex is bounded by
[0, txCount + 1].
Decoder primitives:
- New Rlp.ValueDecoderContext.DecodeUInt() / RlpStream.Encode(uint) /
Rlp.LengthOf(uint) helpers.
- Rlp.Decode<T> and Rlp.DecodeArrayPool<T> wrap IndexOutOfRangeException /
ArgumentOutOfRangeException as RlpException so any truncated-RLP primitive
read surfaces consistently.
- BlockAccessListDecoder rejects an empty AccountChanges entry (0xc0 inside
the outer list) and SlotChangesDecoder rejects an empty StorageChange list
for a slot — geth bal-devnet-4 "empty storage writes" parity.
Type widening:
- IIndexedChange.Index, BalanceChange/NonceChange/CodeChange/StorageChange.Index,
BlockAccessList.Index and BlockAccessIndex on the Change journal record all
become uint. SortedList<int, T> keys flip to SortedList<uint, T>; ushort
query parameters on AccountChanges become uint.
Prestate sentinel preservation:
- Replaces the legacy -1 int sentinel with Eip7928Constants.PrestateIndex
(uint.MaxValue). Adds PrestateAwareIndexComparer that orders the sentinel
before all real indices, restoring the iteration semantics that
AccountChanges.GetBalance/GetNonce/GetCode/AccountExists and
ApplyStateChanges' [^1].Index check rely on.
- Decoder-built SortedLists also use the prestate-aware comparer so that
LoadPreStateToSuggestedBlockAccessList grafting prestate onto the suggested
BAL after decode keeps it sorted first.
- Loop predicates that compare change.Key directly against blockAccessIndex
explicitly skip PrestateIndex so the raw uint comparison doesn't trigger
on the huge sentinel value.
Validation:
- BlockValidator gains ValidateBlockLevelAccessListIndexBounds enforcing
index <= txCount + 1 (mirrors geth's index < txCount + 2 check) with a new
BlockLevelAccessListIndexOutOfRange error message.
Tests:
- New PrestateAwareIndexComparerTests, AccountChangesPrestateTests covering
the comparer and prestate-fallback iteration semantics.
- BlockAccessListDecoderTests adds: empty-bytes / truncated-outer-list /
inner-empty-list throw RlpException; empty storage_changes per slot throw;
decoded SlotChanges accepts a later prestate graft as first; BalanceChange
round-trips for indices 0x10_0000 and uint.MaxValue.
- BlockValidatorTests adds tx-index bound cases (0, 1 valid; 2,
uint.MaxValue rejected for a 0-tx block).
- ExecutionPayloadV4Tests covers the engine-API decoding-error path for
malformed BAL bytes.
* style: drop unused usings flagged by IDE0005
CI surfaced four IDE0005 warnings (treated as errors):
- BlockAccessListManager: drop `using Nethermind.Crypto;` —
Keccak.OfAnEmptySequenceRlp comes from Nethermind.Core.Crypto, already imported.
- PrestateAwareIndexComparerTests / AccountChangesPrestateTests: drop
`using Nethermind.Core;` — Eip7928Constants resolves via the test's parent
namespace Nethermind.Core.Test.BlockAccessLists.
- Eip8037Tests: drop `using System;` — no System.* type referenced directly.
* fix(bal): record system pre/post-block SSTOREs as reads per EIP-7928 v5.7.0
EIP-7928 v5.7.0 specifies that SSTOREs performed during system pre-block calls
(EIP-2935 BlockHashHistory, EIP-4788 BeaconRootContract) and post-block calls
(EIP-7002 withdrawal requests, EIP-7251 consolidation requests) are recorded
in the BAL as storage reads on the touched slot — not as storage changes with
post-values. Same-value writes are skipped entirely. Without this, nethermind
generated a BAL whose Keccak256 differs from what eels-built fixtures expect,
so the BlockAccessListHash check fails for every Amsterdam block that touches
a system contract slot (most pyspec tests).
IWorldState gains an opt-in scope:
IDisposable? BeginSystemPreBlockScope()
TracedAccessWorldState implements it via an int depth counter. While depth > 0,
Set(storageCell, value) reclassifies the recording: AddStorageRead when the
slot value would change, no-op when unchanged. The state mutation still
applies via the inner world state.
BlockAccessListManager wraps the three system contract entry points with the
scope: StoreBeaconRoot (EIP-4788 system tx), ApplyBlockhashStateChanges
(EIP-2935 fast-path Set), and ProcessExecutionRequests (EIP-7002 / EIP-7251
post-block system txs).
Parallel-mode state application:
In parallel processing, non-system slots wrap stateProvider with
BlockAccessListBasedWorldState whose Set is a no-op — actual state mutation
relies on ApplyStateChanges replaying the suggested BAL's storage_changes.
With the spec-correct BAL containing reads instead of changes for the system
slots, ApplyStateChanges has nothing to replay for them, so the canonical
state would diverge.
TxProcessorWithWorldState gains an `isSystemSlot` parameter. The
ParallelTxProcessorWithWorldStateManager passes `isSystemSlot: i == 0 || i ==
_len - 1` (pre-execution and post-execution slots). For those slots, the
TracedAccessWorldState wraps stateProvider directly, so system pre/post-block
SSTOREs mutate the canonical state regardless of BAL recording. Tx slots
(1..n) keep the BAL-backed wrapping unchanged.
Sequential pyspec tests (which is what the Ethereum.Blockchain.Pyspec.Test
suite runs) are unaffected by the parallel slot change but benefit from the
BAL recording fix; the BlockAccessListHash check now passes for blocks that
previously failed solely on system pre-block storage encoding.
Note: a residual InvalidStateRoot mismatch remains on a subset of Amsterdam
pyspec tests (~70/360 ecrecover_weird_v + a similar slice of stMemoryStress).
These were previously masked by the BAL hash error firing first. The
remaining state divergence appears unrelated to BAL recording and is left
for follow-up.
* test(jsonrpc): update Eth_get_block_access_list_by_* fixtures for EIP-7928 v5.7.0
The Eth_get_block_access_list_by_hash and _by_number tests had hardcoded the
pre-v5.7.0 shape, recording the EIP-2935 BlockHashHistory system pre-block
SSTORE as a storageChanges entry with the post-value. v5.7.0 records system
pre-block SSTOREs as storageReads (slot key only).
Updated both expected JSON strings to match the new spec-compliant output.
* Revert "fix(bal): record system pre/post-block SSTOREs as reads per EIP-7928 v5.7.0"
This reverts commit 364f294826d448eb76bf61b3a80f42351a26ea0b.
* Revert "test(jsonrpc): update Eth_get_block_access_list_by_* fixtures for EIP-7928 v5.7.0"
This reverts commit 937ca5d1f970dab62e7d51d2ebff10e8d5151f15.
* review: address PR #11362 feedback
- Rlp.DecodeArrayPool<T>: dispose partially-allocated ArrayPoolList<T>
before wrapping IndexOutOfRangeException/ArgumentOutOfRangeException
as RlpException so the rented buffer is returned to the pool.
- Rlp.ValueDecoderContext.DecodeUInt(): collapse case 0 to a single
`return RlpHelpers.ThrowNonCanonicalInteger(Position)` (DoesNotReturn,
uint) to match DecodeInt and drop the dead `return default`.
- BlockAccessListManager.GetPostExecution(): use uint.MaxValue literal
to match the uint? balIndex parameter.
- AccountChanges.SlotChangesAtIndex: build the returned SortedList with
PrestateAwareIndexComparer.Instance so a later prestate graft sorts
first, matching the rest of the codebase.
- PrestateAwareIndexComparer xmldoc: clarify that decoded BALs also use
this comparer (so later prestate grafting preserves order).
* test(pyspec): skip EELS bal@v5.7.0 ported_static fixtures with legacy state-test conversion bug
EELS's `from_state_test` conversion for ported_static tests omits the
EIP-2935 / EIP-4788 system pre-block SSTORE entries from the suggested
BAL, while a real client (and Nethermind) executes them — so every such
fixture's BlockAccessListHash diverges from what we compute. 91 such
tests were the entire residual pyspec failure set on the bal-devnet-4a
branch.
Detect the conversion via the legacy difficulty value 0x20000 baked
into the post-merge mixHash field — real prevRandao would never be
exactly 0x...020000 — and Assert.Ignore those tests.
Track upstream EELS fix; remove the guard once bal@>v5.7.0 ships with
the system pre-block SSTOREs included in the BAL.
* fix(pyspec test): drop ?-annotation in non-nullable file
CS8632: PyspecTestFixture.cs is not under `#nullable enable`, so the
`string?` introduced in 6fb652762b broke the build of every pyspec
job. `string` works the same here — the value already gets a null
check on the next line.
* test(pyspec): also skip blockchain_test_engine_from_state_test variant
The first guard only walked test.Blocks, which is null for engine
fixtures. Engine fixtures keep the same legacy 0x...020000 sentinel,
just on the JSON `prevRandao` field of the engine_newPayload params.
Walk EngineNewPayloads too.
* Update tests
* fix(eip-8037): pin cost_per_state_byte at static 1174 for bal-devnet-4
bal-devnet-4 keeps cost_per_state_byte static at 1174 (carried over from
bal-devnet-3), removing the per-block-gas-limit scaling formula that an
earlier draft of EIP-8037 specified. snøbal-devnet-4 fixtures encode the
same gas usage (63574) at 1M / 5M / 10M / 30M block gas limits, confirming
the value is now invariant across blockGasLimit.
Reduce CalculateCostPerStateByte to a direct return of CostPerStateByte
and drop the dynamic-quantization helpers and BitOperations dependency.
The blockGasLimit parameter is retained on call sites in case a future
devnet revisits per-block scaling. Update the EIP-8037 unit test to pin
the static behaviour rather than asserting quantized values.
* fix(test): match TCS Exception type to IncrementalValidation signature
* fix test
* fixes
* fix
* Fixes
* lint
* update tests
* fix tests
* fix ci
* perf: optimize BAL lookups and eliminate redundant GetCodeHash calls
- Replace SortedDictionary with Dictionary in BlockAccessList for O(1)
account lookups on the EVM hot path (was O(log n) with 20-byte span
comparisons). Sorted enumeration preserved for encoding/validation.
- Merge TryGetDelegation + GetCachedCodeInfo into a single call in
InstructionCall, eliminating a redundant GetCodeHash per CALL opcode.
- Inline IsDeadAccount in EXTCODEHASH to avoid a second GetCodeHash
call for the same address.
* remove skips
* More pyspec test chunks
* Squashed commit of the following:
commit 3a3078f428e24435464d33676e50c50eacb9644e
Author: Marc Harvey-Hill <10379486+Marchhill@users.noreply.github.com>
Date: Tue Apr 28 18:23:03 2026 +0200
delete old tests
commit 95de888a18a16a38ed425bda0a70c4235de96d9c
Merge: 244c2c60b1 31a673a6a4
Author: MarekM25 <marekm2504@gmail.com>
Date: Tue Apr 28 18:08:12 2026 +0200
Merge branch 'master' into engine-api-glamsterdam-cleanup
commit 244c2c60b1873b1d732025b0240c154ecbbe6dd0
Author: MarekM25 <marekm2504@gmail.com>
Date: Tue Apr 28 18:05:32 2026 +0200
fix lint
commit 9194b8f54322915a57ecbc0acbb9b4b71a7508b5
Author: MarekM25 <marekm2504@gmail.com>
Date: Tue Apr 28 18:04:19 2026 +0200
remove tests
commit eedfbb775a018860ef0a6b51822589752ac5ea76
Author: MarekM25 <marekm2504@gmail.com>
Date: Tue Apr 28 17:57:46 2026 +0200
revert formatting changes and zero hash stuff
commit dc71aa5093665c75a5fb126ffad835cc0b81ee3a
Author: MarekM25 <marekm2504@gmail.com>
Date: Tue Apr 28 17:38:42 2026 +0200
cleanups
commit ec2fce3609d2e06fa856cb0c545d6fb434900391
Author: MarekM25 <marekm2504@gmail.com>
Date: Tue Apr 28 17:33:11 2026 +0200
fixing
commit 90a1da8c5f816cc3965cadd98f28ceb31b236bf6
Author: MarekM25 <marekm2504@gmail.com>
Date: Tue Apr 28 17:15:58 2026 +0200
remove test
commit 4451656d1c79855c0d6c23c0c1d9a3f0b9755139
Author: MarekM25 <marekm2504@gmail.com>
Date: Tue Apr 28 17:12:32 2026 +0200
Cleanup mess
commit 2425748d949b3ac4364b549d66abbb0427d65289
Merge: 4a904608e4 a36154c39a
Author: Stavros Vlachakis <89769224+svlachakis@users.noreply.github.com>
Date: Mon Apr 27 14:33:36 2026 +0300
Merge branch 'master' into engine-api-glamsterdam
commit 4a904608e48750d3f20cbf2554294f7e64969a0a
Merge: 35497680be 18d60a482b
Author: Stavros Vlachakis <89769224+svlachakis@users.noreply.github.com>
Date: Mon Apr 27 14:08:25 2026 +0300
Merge branch 'master' into engine-api-glamsterdam
commit 35497680be5f1d60b672e72f66d2949833dc82d0
Merge: 30fcc367cd b71c3528d0
Author: Stavros Vlachakis <89769224+svlachakis@users.noreply.github.com>
Date: Mon Apr 27 14:07:39 2026 +0300
Merge branch 'master' into engine-api-glamsterdam
commit 30fcc367cd486bbd0a5611d943bdf2eff2956a9b
Merge: 540e4a2fd1 ed6a354e6a
Author: Stavros Vlachakis <89769224+svlachakis@users.noreply.github.com>
Date: Mon Apr 27 13:09:48 2026 +0300
Merge branch 'master' into engine-api-glamsterdam
commit 540e4a2fd1774064c1cff747118a3ae4a7644be0
Author: stavrosvl7 <stavrosvl7@gmail.com>
Date: Sun Apr 26 04:52:49 2026 +0300
minor fixes
commit 010547be9ef97dcf853a2b239b7b0bb7a210879a
Author: stavrosvl7 <stavrosvl7@gmail.com>
Date: Sun Apr 26 03:22:39 2026 +0300
test fixes
commit 6fcc136f54a8acc1b1266103230b772f302f7c09
Author: stavrosvl7 <stavrosvl7@gmail.com>
Date: Sun Apr 26 02:11:12 2026 +0300
fixes
commit 0bccb0649f8c9d6c7c689d61b5635aae0e9bc191
Author: stavrosvl7 <stavrosvl7@gmail.com>
Date: Sun Apr 26 01:52:02 2026 +0300
test fix
commit 830c578eb2388140e2141b4639313fa83f48d854
Author: stavrosvl7 <stavrosvl7@gmail.com>
Date: Sun Apr 26 01:38:55 2026 +0300
fix tests
commit 1f5fd2cff7221d0f72dfe0f994567ef883710f4e
Author: stavrosvl7 <stavrosvl7@gmail.com>
Date: Sun Apr 26 01:35:17 2026 +0300
claude review again
commit d87491f353d6525dc50c398df01b2990f4c692f2
Author: stavrosvl7 <stavrosvl7@gmail.com>
Date: Sun Apr 26 00:57:57 2026 +0300
fixes
commit dd431e29428ae75d0586af2981f8764510ba3b47
Author: stavrosvl7 <stavrosvl7@gmail.com>
Date: Sun Apr 26 00:51:52 2026 +0300
fix tests
commit 71dc99f55d9d7055f05ea5bb9bc89a35954e42c0
Author: stavrosvl7 <stavrosvl7@gmail.com>
Date: Sat Apr 25 22:40:50 2026 +0300
fixes
commit 9c03d12606e21bbc6306d70419437a7c55d535e1
Author: stavrosvl7 <stavrosvl7@gmail.com>
Date: Sat Apr 25 22:32:58 2026 +0300
update per 786
commit a4c4f3c4134f6d0c18c7ed3da162079a794b3435
Merge: fc49e7e1fe 9cba44cda1
Author: stavrosvl7 <stavrosvl7@gmail.com>
Date: Sat Apr 25 21:38:25 2026 +0300
Merge remote-tracking branch 'origin/master' into engine-api-glamsterdam
commit fc49e7e1fe858eafdc895483189e17da9751a237
Merge: bca0c63446 42c551f4ab
Author: Stavros Vlachakis <89769224+svlachakis@users.noreply.github.com>
Date: Mon Apr 20 16:14:33 2026 +0300
Merge branch 'master' into engine-api-glamsterdam
commit bca0c634462fadb7a479dcb2e63786b70845ede1
Author: stavrosvl7 <stavrosvl7@gmail.com>
Date: Mon Apr 20 16:14:10 2026 +0300
remove unused import
commit 3d2c973a68c5ce443a468790766905dfcbf0d7f0
Author: stavrosvl7 <stavrosvl7@gmail.com>
Date: Mon Apr 20 16:09:17 2026 +0300
fix taiko tests
commit b30ef5c436dd6a2e6c7f352245205a41e0b26cd7
Author: stavrosvl7 <stavrosvl7@gmail.com>
Date: Mon Apr 20 15:34:35 2026 +0300
fix tests
commit 6f131cd7afd097785a24635a5bed232a3b836064
Merge: a8c884a092 0fe41f4176
Author: Stavros Vlachakis <89769224+svlachakis@users.noreply.github.com>
Date: Mon Apr 20 14:55:46 2026 +0300
Merge branch 'master' into engine-api-glamsterdam
commit a8c884a092cc7e7fcd4f38a7bd9d30e0e5836b06
Author: stavrosvl7 <stavrosvl7@gmail.com>
Date: Mon Apr 20 14:54:40 2026 +0300
conflicts fix
commit 96eb9124bbeac213070c5d29759e79010df4b310
Merge: 82a1f70572 425a2a3fe8
Author: stavrosvl7 <stavrosvl7@gmail.com>
Date: Mon Apr 20 14:52:33 2026 +0300
Merge remote-tracking branch 'origin/master' into engine-api-glamsterdam
# Conflicts:
# src/Nethermind/Nethermind.Merge.Plugin/BlockTreeExtensions.cs
# src/Nethermind/Nethermind.Merge.Plugin/Handlers/ForkchoiceUpdatedHandler.cs
# src/Nethermind/Nethermind.Taiko/Rpc/TaikoForkchoiceUpdatedHandler.cs
commit 82a1f705728f739c85b650ff8316f2a2ebd1f7e6
Merge: 7b6133f3a8 436c65bbc3
Author: stavrosvl7 <stavrosvl7@gmail.com>
Date: Mon Apr 20 13:16:55 2026 +0300
Merge remote-tracking branch 'origin/master' into engine-api-glamsterdam
commit 7b6133f3a89ffc60106e88b7c5b4f95d4feae20f
Merge: 2f068ec5b3 02c202601b
Author: Stavros Vlachakis <89769224+svlachakis@users.noreply.github.com>
Date: Thu Apr 16 14:49:46 2026 +0300
Merge branch 'master' into engine-api-glamsterdam
commit 2f068ec5b3f8e1a4742bfd29fbafbf620609e40a
Merge: bdd7ee59c4 5464c8ba2b
Author: Stavros Vlachakis <89769224+svlachakis@users.noreply.github.com>
Date: Thu Apr 16 14:10:02 2026 +0300
Merge branch 'master' into engine-api-glamsterdam
commit bdd7ee59c40d65d664cbadbada4e6708581eceb7
Merge: f09cb41ddd bfaaeb0f53
Author: Stavros Vlachakis <89769224+svlachakis@users.noreply.github.com>
Date: Wed Apr 15 16:45:19 2026 +0300
Merge branch 'master' into engine-api-glamsterdam
commit f09cb41ddd3293a1269d137796187cdb1babcaf7
Merge: 370acdf4c4 cf03d184f5
Author: stavrosvl7 <stavrosvl7@gmail.com>
Date: Wed Apr 15 16:24:17 2026 +0300
Merge remote-tracking branch 'origin/master' into engine-api-glamsterdam
commit 370acdf4c4a587c46022feb6f16837422dfaa5f2
Merge: d1d0370429 6cd0ea49db
Author: Stavros Vlachakis <89769224+svlachakis@users.noreply.github.com>
Date: Mon Apr 13 14:28:21 2026 +0300
Merge branch 'master' into engine-api-glamsterdam
commit d1d0370429045dc3819dbbe4c09785227526f21e
Author: stavrosvl7 <stavrosvl7@gmail.com>
Date: Fri Apr 10 17:30:54 2026 +0300
more fixes
commit c99d6cc7fb4391d9f37862ea47ffa0b830c5cf51
Merge: 4d0f215ad0 a52cb90edc
Author: Stavros Vlachakis <89769224+svlachakis@users.noreply.github.com>
Date: Fri Apr 10 17:24:06 2026 +0300
Merge branch 'master' into engine-api-glamsterdam
commit 4d0f215ad0e184ac4e4e73a2411eae0d80b69100
Author: stavrosvl7 <stavrosvl7@gmail.com>
Date: Fri Apr 10 17:23:42 2026 +0300
claude review
commit cb6defe8ce8196fe7fdda28b35a1376e3e72f5e1
Merge: af63142ba1 0057bd83c2
Author: Stavros Vlachakis <89769224+svlachakis@users.noreply.github.com>
Date: Fri Apr 10 12:34:37 2026 +0300
Merge branch 'master' into engine-api-glamsterdam
commit af63142ba12aaddd443cb124ddea4af41f4785e4
Merge: 1b338f380d 2f24891849
Author: Stavros Vlachakis <89769224+svlachakis@users.noreply.github.com>
Date: Fri Apr 10 12:08:36 2026 +0300
Merge branch 'master' into engine-api-glamsterdam
commit 1b338f380ddff161e66a20a7fad584578d9214ed
Author: stavrosvl7 <stavrosvl7@gmail.com>
Date: Fri Apr 10 12:08:13 2026 +0300
claude review
commit 619259b75cb57403ab826708b1127f65412905a5
Author: stavrosvl7 <stavrosvl7@gmail.com>
Date: Thu Apr 9 18:32:59 2026 +0300
taiko fix
commit d06779353879b10f287b7405c319f23540125a52
Author: stavrosvl7 <stavrosvl7@gmail.com>
Date: Thu Apr 9 18:19:14 2026 +0300
cleaner design
commit 1e27be19466204dc0bfbb515edde6dbd90053d52
Author: stavrosvl7 <stavrosvl7@gmail.com>
Date: Thu Apr 9 18:12:04 2026 +0300
improve tests matching the specs
commit a77e1827d4fb7d5e0a3a8739c8747a9a43279b7e
Author: stavrosvl7 <stavrosvl7@gmail.com>
Date: Thu Apr 9 18:04:58 2026 +0300
improve test
commit 4ba069c9c65ce7e7b4922d63c4e5341c17a80f71
Author: stavrosvl7 <stavrosvl7@gmail.com>
Date: Thu Apr 9 18:01:59 2026 +0300
fix comment
commit fe6f9b094eeeccc97158d1652dfa5e03275d78ca
Author: stavrosvl7 <stavrosvl7@gmail.com>
Date: Thu Apr 9 17:53:57 2026 +0300
engine api changes as per https://github.com/ethereum/execution-apis/pull/770
https://github.com/ethereum/execution-apis/pull/760
* update tests
* remove skips
* halt changes
* fixes
* fix
* Bump pyspec fixtures to snobal-devnet-6@v1.1.0
* fix spill
* Don't refund spilled state gas on top-level halt
* Refund full state gas on top-level revert; only reservoir-portion on halt
Top-level REVERT preserves gas_left, so the spilled portion of state_gas_used
(originally drawn from gas_left) is still in the user's pocket and must be
refunded to the reservoir alongside the reservoir-funded portion. Top-level
halt burns gas_left, so the spilled portion was paid out of gas the user can
no longer reclaim — only the reservoir-funded portion is restorable.
Splits RefundRevertedTopLevelStateGas into a revert path (full refund) and a
new RefundHaltedTopLevelStateGas (reservoir-only, spill discarded) and wires
the halt error sites to the latter.
* fix: address review - replace LINQ OrderBy with List.Sort, consistent GetContentLength
* fix
* fix
* Cover missing pyspec fork dirs + add sync and transaction test types
The pyspec fixture archive ships three uncovered directories: a stray state-
tests transition fork, plus two test types we never wired in. Adds:
- CancunToPragueAtTime15kStateTests — fills a one-class gap in Tests.cs
- PyspecSyncBlockchainTestFixture + AmsterdamSyncBlockchainTests,
OsakaSyncBlockchainTests — runs blockchain_tests_sync fixtures through the
engine harness; the additional syncPayload field is left for a follow-up
- TestType.Transaction + TransactionTest/Json/Base + ConvertTransactionTests
+ PyspecTransactionTestFixture — decodes raw txbytes via Rlp.Decode and
matches expected EEST exception tokens (TYPE_4_*) against the validator's
ValidationResult.Error or RLP decode message; covers AmsterdamTxTests,
OsakaTxTests, PragueTxTests fork directories
Bumps FlatDB pyspec chunking from 4 to 16 to match the regular workflow —
~860 tests/shard instead of ~3,437/shard, well under the 256/matrix cap and
the 20-minute job timeout.
* Moar tests
* fx
* Filter known-broken from_state_test BAL revert tests; relax BAL error match
Two CI failures in the new EEST snobal-devnet-6 fixtures around BAL
validation in negative and revert scenarios:
- Class B (negative tests, 15 cases): EEST's *_from_state_test synthesized
blockchain test versions of state tests ship an incomplete suggested BAL
for transactions expected to fail at tx-level. Nethermind correctly
rejects the block but via "InvalidBlockLevelAccessList: missing/surplus
account changes" rather than the expected TransactionException. The
block IS invalid; only the failure mode differs. AssertValidationError
now accepts BAL missing/surplus/incorrect-changes errors as a valid
alternative outcome.
- Class A (positive tests, 4 cases): Nethermind's generated BAL for
REVERT/TSTORE scenarios disagrees with the suggested BAL on intermediate
storage values, even though the final post-state matches. Filter these
four specific from_state_test variants out of Pyspec test loading until
the BAL-on-revert behavior is investigated. Original state tests in
PyspecStateTestFixture still cover the same scenarios.
* optimise allocations of worldstates, txprocessors, bals
* optimise bal structures
* fix tests
* perf(bal): zero-alloc dict-lookup ValidateBlockAccessList (#11448)
* Optimize early validation
* Drop redundant `using` on SortedDictionary value-enumerator
The Dispose chain through `SortedDictionary.ValueCollection.Enumerator`
→ `SortedDictionary.Enumerator` → `TreeSet<KVP>.Enumerator` bottoms out
at an empty Dispose, so the `using` only adds visual noise around the
manual MoveNext/Current control. Concrete struct type still keeps the
enumerator unboxed.
* Drop unused System.Collections.Generic / System.Linq usings (IDE0005)
* Increase tests
* fix
* optimise bal generation (#11452)
* optimise bal generation
* change to ref readonly
* fix
---------
Co-authored-by: Marc Harvey-Hill <10379486+Marchhill@users.noreply.github.com>
* fix
* Revert "fix"
This reverts commit 4767a18e6e6c20516a1494f07fea397b1a39ad3e.
* Revert "fix"
This reverts commit 9870f9732db82393fce8468c04c6d2c9ce804c09.
* EIP-8037: stop subtracting state-gas spill from block regular-dim
Calculate8037BlockRegularGas was subtracting StateGasSpill from the
regular dim's per-tx contribution. Per EELS amsterdam/fork.py:1167-1172,
spilled state gas reduces gas_left directly and must remain counted in
block_regular_gas; max(sum_regular, sum_state) at block level single-
counts it via the dim that wins. Subtracting it under-counted block.gasUsed
when sum_regular dominated (HeaderGasUsedMismatch on devnet block 1788:
expected 28551136, computed 28438432, diff 112704 = 3×SSetState worth of
spill across the block's CREATE-tx init-code paths).
Updates 4 unit tests with hardcoded expectations from the previous model
and adds a regression test mirroring block 1788 (sub-cap CREATE tx whose
init code spills state gas via cold SSTOREs before top-level REVERT, so
sum_regular dominates and the spill must remain in the regular dim).
Verified against 1369 EIP-8037 / BAL / state-gas pyspec fixtures and 201
EIP-7928/8037/6780 unit tests.
* update tests
* fx
* Revert pyspec fixtures to execution-spec-tests snobal-devnet-6@v1.1.0
The tests-snobal-devnet-6@v1.1.1 tag in ethereum/execution-specs has no
fixtures_snobal-devnet-6.tar.gz asset attached (only auto-generated source
archives), so every Pyspec shard 404s on download. Point back at
ethereum/execution-spec-tests where the fixture artifact is published.
* Revert "EIP-8037: stop subtracting state-gas spill from block regular-dim"
This reverts commit 3758ae4288.
EELS reference (devnets/snobal/6 amsterdam/vm/__init__.py + vm/gas.py)
shows regular_gas_used is incremented only by charge_gas (regular ops);
charge_state_gas spilling into gas_left increments only state_gas_used,
never regular_gas_used. Therefore tx_regular_gas at line 1168
(intrinsic_regular + regular_gas_used) excludes spill. The pre-revert
'- stateGasSpill' subtraction is required to recover regular_ops_only
from (initial_gas_left - final_gas_left), which includes spill.
The original block-1788 mismatch this commit tried to address
(HeaderGasUsedMismatch -112704) is a separate halt/revert state-gas
restoration bug specific to a CREATE scenario. EELS team confirmed the
spec-side test had the same gap and clients (geth, besu, nethermind
pre-fix) were aligned via convergent behaviour; spec-side fix is
expected next week. Reverting to keep CI green and consensus-aligned
with the rest of the network. Block 1788 root-cause investigation moves
to a follow-up.
* add mapping
* Wire transient storage snapshot/restore through BlockAccessListBasedWorldState
The parallel BAL builder uses BlockAccessListBasedWorldState as the inner
world state, which owns its own TransientStorageProvider but had stub
TakeSnapshot/Restore (returned Snapshot.Empty / no-op). On REVERT the
EVM calls IWorldState.Restore(snapshot), which in parallel mode reached
BALWS and silently dropped the transient-storage rollback — TSTORE
writes inside reverting frames leaked into TLOADs after the revert,
producing wrong storage / fee values in the generated BAL versus EELS'
suggested BAL.
Forward TakeSnapshot/Restore to the inner _transientStorageProvider,
matching how WorldState wires the same field. Resolves the
KnownPyspecBalRevertBugs skip list (deleted): all six failing
[ParallelEngine] tests now pass — test_tstore_reentrancy variants,
test_subcall[delegatecall_with_invalid], test_trans_storage_reset,
test_set_code_max_depth_call_stack, test_10_revert_undoes_store_after_return.
Verified locally: 1562 of 1562 pyspec tests in the affected families pass,
194 of 194 EIP-7928/8037/BAL unit tests pass, 72 of 73 State.Test pass
(1 pre-existing skip).
* Tag sequential Pyspec jobs
* Tag regular Pyspec jobs
* Use bracket tags for extra test variants
* EIP-8037: replace worst-case-OR pre-validation with intrinsic-floor check
ValidateTransactionGasAllowance (BlockAccessListManager and TransactionProcessor)
rejected a tx whenever its worst-case contribution to EITHER dim exceeded that
dim's remaining capacity. Per EIP-8037 the tx contributes (r,s) with r+s ≤
tx.gasLimit and routes between dims; the block fits iff some split satisfies
both per-dim caps. Worst-case-OR rejected blocks where each dim's worst-case
alone overflowed but the actual two-dim split fit — surfaced on bal-devnet-6
via kurtosis: a 60M block with 4×~16M txs to a state-heavy contract had
sum_regular=37.7M and sum_state=57.8M, so block.gasUsed = max() = 57.8M ≤ 60M,
but the OR check rejected tx[3] because either worst-case alone exceeded
remaining headroom in its dim. NM rejected → all NM nodes diverged from
geth+besu. Same shape as the prior 'all prysm-nm nodes struggle' devnet report.
Replaces with the spec-correct condition: a tx unavoidably contributes its
intrinsic to each dim, so reject only when intrinsic_regular alone overflows
the regular dim, intrinsic_state alone overflows the state dim, or the minimum
required execution gas exceeds the combined remaining capacity. Anything
beyond intrinsic that overflows is caught post-execution by CheckGasUsed
using the proper max(Σregular, Σstate) ≤ block.gasLimit formula.
Updates Eip7928Tests:
- Tx_exceeding_block_gas_limit_rejected_in_parallel_mode renamed and
retargeted at the intrinsic-overflow path (a 20k block where the 21k
intrinsic alone overflows). The original assertion (tx.gasLimit
> block.gasLimit triggers rejection) was itself the bug — under EIP-8037
such a tx is valid as long as its actual execution fits.
- New regression Tx_with_gaslimit_above_block_remaining_but_intrinsic_fits_accepted_under_eip8037
pins the now-correct behaviour.
BlockProcessorTests.IncrementalValidation_rejects_eip8037_tx_when_worst_case_exceeds_ordered_remaining_gas
keeps passing — its scenario also trips the intrinsic-overflow path (after
firstTx claims 80k regular, secondTx's 21k intrinsic_regular alone overflows
the 20k regular headroom).
* EIP-8037: avoid long overflow in combined-capacity admission check
The combined-capacity branch `minGasRequired > regularAvailable + stateAvailable`
overflowed when pyspec fixtures used block gas_limit near i64.MaxValue
(e.g. test_call_bounds: gas_limit=9_223_372_036_854_775_807). The sum
wrapped to a negative long, the comparison fired truthy, and tx[0] was
rejected with 'Block gas limit exceeded' even though both dims had
effectively unbounded headroom. 18 pyspec tests in the (13of16) chunk
regressed on PR #11436.
Reformulated as: 'tx unavoidably needs minGasRequired across both dims;
the regular dim alone covers it iff regularAvailable >= minGasRequired,
otherwise the residual minGasRequired - regularAvailable must fit in
stateAvailable'. Same semantics, no overflow.
* Revert "EIP-8037: avoid long overflow in combined-capacity admission check"
This reverts commit 60aa7ef00dc897dd70f1f3f1f9847a8d2c423237.
* Revert "EIP-8037: replace worst-case-OR pre-validation with intrinsic-floor check"
This reverts commit 83dcde59bf54dede0bc5b97c10bee36a19fd32f4.
* Reapply "EIP-8037: replace worst-case-OR pre-validation with intrinsic-floor check"
This reverts commit d2410313642445d69c4ef787ee87fd539e9e72a0.
* Reapply "EIP-8037: avoid long overflow in combined-capacity admission check"
This reverts commit b7c2a725d5e2f304989eb337ce84b59d26f17221.
* Revert ValidateTransactionGasAllowance to spec-compliant worst-case-OR
The intrinsic-floor / AND variants of this check made block-125-shape kurtosis
devnet traffic stay in lockstep, but they diverge from the EELS amsterdam
spec text (fork.py:540-560 codifies worst-case-OR exactly) and break
test_low_gas_limit[fork_Amsterdam-state_test--g0] which asserts rejection
of tx.gasLimit > block.gasLimit on a fresh block.
The original devnet-6 block-125 divergence (NM rejects, geth+besu accept)
is therefore not in the admission rule itself but in NM's per-tx (regular,
state) accumulator that feeds it. Investigating that as a separate fix:
geth+besu accept block 125 with worst-case-OR, so their per-tx values must
keep totalRegular AND totalState below 60M-15.86M=44.14M after 3 txs while
NM's evidently exceeds that threshold.
* Surface parallel worker tx-rejection cause without masking
When a parallel BAL worker rejects a tx with InvalidBlockException, the
incremental validator was running the admission rule and CheckGasUsed
before checking the worker's exception slot. With the worker reporting
tx.GasLimit on rejection, the cumulative-gas check could trip a follow-on
"block gas limit exceeded" that masked the original cause and diverged
from the sequential path. Rethrow ParallelExecutionException as soon as
the worker's slot carries an exception, and have the worker report (0, 0)
so any future consumer of the tuple agrees with sequential semantics.
* Invalidate BlockAccessList ItemCount cache on mutations
BlockAccessListManager reuses one GeneratedBlockAccessList instance
across blocks, while BlockAccessList cached its computed ItemCount
without invalidation on Reset, Clear, Merge, or any Add/Restore path.
After one validation cached a small count, a later oversized generated
BAL would pass the EIP-7928 item-limit check on RLP-imported blocks
(and the reverse — a smaller BAL after a larger one — would be rejected).
Null _itemCount at the entry of every mutating method on BlockAccessList.
The RLP decoder's init-set value still survives for the lifetime of
freshly decoded (and never mutated) BALs. Add unit tests covering each
mutator and a two-block ValidateProcessedBlock regression that processes
the same BAL instance at the floor, then over the floor.
* fixes
* fix
* fixes
* fixesfix
* Drop duplicate Rlp.ValueDecoderContext.DecodeUInt from #11362
The merge of master left two DecodeUInt() definitions on
ValueDecoderContext: one from #11362 (bal-devnet-6) and one from
#11479 (master). Keep #11479's version (captures position before
ReadByte advances; adds the result < 128 non-canonical-integer
check) and remove #11362's. The Encode(uint)/LengthOf(uint) helpers
from #11362 are kept since master did not add equivalents.
* fixes
* tests
* Fixes
* Align error messages
* fix
* lint
* sentinel change
* Fix receipts
* CheckPerTxInclusion
* initial impl
* use task
* fix
* fix one more time
* fix
* attempt fix
* Feedback
* fix
* tidy
* Revert "attempt fix"
This reverts commit 9f78de89669c14c6a5659d0a7a0aadfc83bb5038.
* tidy
* fix
* Fixes
* pre-execution in parallel thread
* fix
* add comment
* feedback
* Feedback
* Feedback
* Fix processing stats for parallel
* lint
* Harden
* Update deps
* Optimize
* Super-optimize
* Ultra-optimize
* Turbo-optimize
* lint
* Eldritch-optimize
* Forbidden optimizations
* fix
* Feedback
* merge conflict
* Hive fixes
* fix(bal): defer GetCachedCodeInfo past CALL OOG checks to keep delegation target out of BAL on revert
cea517aa20 (perf: optimize BAL lookups...) collapsed the two-step CALL
delegation lookup into a single GetCachedCodeInfo call placed before the
delegated cold-account-access gas charge. That moved InternalGetCodeInfo
of the delegation target ahead of the OOG point, so _worldState.GetCodeHash
recorded the target in the BAL even when the CALL OOG'd before its frame
was entered. Per EIP-7928 / EELS, the delegation target must only appear
in the BAL when the CALL frame actually executes — so its recording must
happen after all CALL-level OOG checks pass.
Restore upstream's two-step ordering: TryGetDelegation early (records
codeSource only by parsing its code), then GetCachedCodeInfo after the
delegated cold-access charge and new-account-creation check (records the
delegation target only when the call is sure to enter the frame).
Surfaced by snobal-devnet-6@v1.1.0 fixtures that exercise CALL/CALLCODE/
DELEGATECALL/STATICCALL × cold/warm × oog_success_minus_1 / oog_after_target_access.
Cleared all ~50 Amsterdam blockchain Pyspec failures across 8 chunks.
Adds a focused EvmInstructions regression test that constructs an outer CALL
into an EIP-7702-delegated EOA with gas exactly one short of the delegation
target's cold-access charge and asserts the BAL contains the call target
(the EOA) but not the delegation target. The test fails without this fix.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feedback
* Improve validation performance
* MOAR
* MOAR (fix)
* fix lint
* Revert "MOAR (fix)"
This reverts commit 597551877ff10dc0637953a27059f24cbc20ca54.
* Revert "MOAR"
This reverts commit b88a14fcad8e317212379b772fdd094e610cc160.
* perf(bal): defer slot-array rebuild in ReadOnlyAccountChanges
Replace the per-call O(n) InsertSorted in LoadPreStateStorage with a dirty
flag and a single O(n log n) Array.Sort triggered lazily on first read of
StorageChanges or ChangedSlots. Cumulative cost for a block touching n unique
slots drops from O(n²) to O(n log n).
Adapted from upstream PR #11455 with two changes:
1. Thread-safety: getters call WaitForPrestate() then EnsureSorted(), which
double-checks _sortedDirty under _sortLock. Volatile.Write on the flag
publishes the new array references with release semantics so a reader
observing dirty=false is also guaranteed to see the updated arrays.
Required because parallel tx workers on this branch all wake on the
prestate gate and hit StorageChanges concurrently.
2. Sentinel: tests use Eip7928Constants.PrestateIndex (uint.MaxValue) /
uint literals instead of -1; the upstream PR predates the int->uint
widening of IIndexedChange.Index.
Co-authored-by: kamilchodola
* Reduce sorting
* Partial feedback
* Partial feedback
* Improve test runners
* Add pyspec memory monitor
* More stable tests
* Kamil fix
* fix
* Revert "Kamil fix"
This reverts commit 7954eaa489a0a83195eba4ac2b9d5a24787135b3.
* Reduce code changes
* Reduce code
* dedupe tests
* docs(bal): address Claude review feedback on PR #11511
Seven of eight review comments addressed; the leftover IBlockAccessListManager
TODO is intentionally kept as-is.
- BlockAccessListBasedWorldState.{Get,GetOriginal}: extract shared
GetAtCurrentIndex and document that EIP-2200 "original" and "current"
collapse to the same BAL slot here (intra-tx writes go through the
per-tx journal, not back into this state).
- ReadOnlyAccountChanges.LoadPreState*: document the per-account
realloc cost (~3 small arrays per account) and the trade-off against
the cross-cutting reads that a separate-prestate-field design would
impose; kept simple deliberately.
- ReadOnlyAccountChanges.WaitForPrestate: spell out the two scheduling
invariants (loader must not call WaitForPrestate; ParallelUnbalancedWork
must guarantee slot 0 a thread). Mirrored at the parallel-loop call site.
- GeneratedAccountChanges.SlotChangesAtIndexEqual: dispose the
SortedDictionary value-enumerator in finally so a future BCL change to
its Dispose semantics doesn't leak.
- BlockAccessListAtIndex.AddNonceChange: document the BAL convention that
newNonce == 0 means "no recorded nonce change", not "set to 0".
- BlockAccessListManager.ApplyStateChanges: document the precondition
that the BAL has been prestate-loaded, and explain the GetBalance(0)
fallback to zero (created-mid-block accounts).
- TracedAccessWorldState: class-level remarks documenting the
SetGeneratingBlockAccessList setup contract; field comment noting the
deliberate fail-fast on null.
* fix lint
* Fix bad merge
* fix lint
* perf(bal): cut per-SLOAD allocations and SortedDictionary insert cost
Three review comments on the parallel BAL hot path:
- ReadOnlySlotChanges.Get(uint blockAccessIndex) used to return a fresh
leading-zero-stripped byte[] every call (one alloc per SLOAD through
BlockAccessListBasedWorldState). Now takes a caller-owned Span<byte>
buffer and returns a slice. The single caller
(BlockAccessListBasedWorldState.GetAtCurrentIndex) owns a 32-byte
instance scratch buffer — one per parallel worker, single-threaded use.
- TracedAccessWorldState.GetInternal did the same thing for intra-tx
SLOADs: MemoryMarshal.CreateReadOnlySpan(...).ToArray() allocated a
new byte[32] per call. Replaced with a 32-byte instance scratch buffer
on TracedAccessWorldState (also rented per-worker from the pool).
- BlockAccessListAtIndex._accountChanges was a SortedDictionary keyed on
Address. The sorted property is only consumed once at merge time, by
GeneratedBlockAccessList which has its own SortedDictionary doing the
re-sort anyway. Swapped for a plain Dictionary so AddBalanceChange /
AddNonceChange / AddStorageChange / AddAccountRead / etc. are O(1)
instead of O(log n) on the per-tx hot path.
Test update: AccountChangesPrestateTests.Slot_get_returns_prestate_value
adopts the new ReadOnlySlotChanges.Get(uint, Span<byte>) signature.
* test(pyspec): bump fixtures to bal@v7.0.0
Point pyspec fixture download at the bal-devnet-7 mirror release on
execution-spec-tests (bal@v7.0.0 / fixtures_bal.tar.gz), tracking the
tests-bal@v7.0.0 tag on execution-specs.
* Update EIP-8037 gas constants
* Add EIP-7928 raw block access list RPC
* Fix EIP-8037 reverted state gas accounting
* Fix EIP-8037 block state gas accounting for EIP-7702 refunds
* Add EIP-8037 CREATE collision gas regressions
* Extend EIP-8037 same-tx selfdestruct regressions
* Add EIP-8037 selfdestruct beneficiary charging regression
* Clean up EIP-7928 decoder test comments
* Cover eth/71 protocol inheritance
* Verify post-merge eth capabilities
* Add EIP-8037 selfdestruct beneficiary variants
* Update EIP-8037 block gas budget regression
* Update EIP-7928 BAL expectations for EIP-8037
* Cover EIP-7702 BAL authorization rejections
* Cover EIP-7702 BAL delegation chains
* Cover EIP-7702 BAL precompile delegation
* Cover EIP-7928 same-tx selfdestruct storage
* Cover EIP-7928 CREATE2 recreation BAL
* Cover EIP-7928 selfdestruct sender BAL
* Cover EIP-7928 EXTCODEHASH boundaries
* Error message consistency
* Fix EIP-8037 reverted state gas accounting
* Update errors
* Update messages
* Simplify EIP-8037 state gas constants
* Return stored raw block access list RLP
* Clean up EIP-8037 state gas helpers
* Shorten blob gas validation note
* refactor(bal): use GasValidationResultSlot in IncrementalValidation
The GasValidationResultSlot type from master already lives in the tree but
wasn't being used — every IncrementalValidation signature carried a verbose
TaskCompletionSource<(long BlockGasUsed, long BlockStateGasUsed,
IntrinsicGas<EthereumGasPolicy> IntrinsicGas, InvalidBlockException? Exception)>[]
instead.
Swap to GasValidationResultSlot[] across IBlockAccessListManager and its two
implementations, the parallel executor, and the tests. Worker sites become
TrySetResult(new GasValidationResult(...)) and the validator destructures
gasResults[j].GetResult() into the typed record.
Functionally equivalent — GasValidationResultSlot.GetResult blocks via
Monitor.Wait the same way TaskCompletionSource.Task.GetAwaiter().GetResult
did, and TrySetCanceled likewise dispatches a TaskCanceledException via
ExceptionDispatchInfo. The Task preExecutionTask parameter is kept since
our parallel-loop has a separate pre-execution iteration.
* test(eip8037): replace opaque PR 2703 references with descriptive names
Renamed Spec_pr2703_* tests to describe the scenario they pin, and rephrased
docstrings / inline comments / failure messages that anchored to the
upstream PR number. Reader no longer needs to open execution-specs PR 2703
to know what each test covers.
Eip8037BlockGasIntegrationTests:
Spec_pr2703_boundary_state_exact_fit_accepts
-> State_dimension_exact_fit_at_block_gas_limit_accepts
Spec_pr2703_boundary_state_exceeded_by_one_rejects
-> State_dimension_one_over_block_gas_limit_rejects
Spec_pr2703_creation_tx_regular_check_actual_usage_modest_accepts
-> Creation_tx_intrinsic_state_excluded_from_regular_worst_case_accepts
Spec_pr2703_single_tx_state_check_exceeds_block_limit_rejects
-> Single_tx_state_worst_case_over_block_gas_limit_rejects_at_inclusion
Spec_pr2703_creation_tx_state_check_exceeded_rejects
-> Creation_tx_state_worst_case_over_remaining_state_budget_rejects_at_inclusion
Spec_pr2703_eip7825_cap_with_modest_actual_gas_accepts
-> Regular_worst_case_capped_by_eip7825_with_modest_post_exec_gas_accepts
Eip8037BlockGasInclusionCheckTests + BlockProcessorTests: rephrased "PR 2703
test_..." section comments into prose describing what the case pins.
Production-code spec anchors (BlockAccessListManager, Eip8037BlockGasInclusionCheck,
EthereumGasPolicy) keep their "execution-specs PR 2703" references — they were
deliberately added as spec citations and are useful for reviewers tracking the
rule back to upstream.
* perf(bal): port master's BlockAccessListValidationIndex for fast incremental validation
Ports upstream's column-oriented validation index (lost to merge-conflict
"take ours" resolution when bal-devnet-6 merged into master, since master's
implementation was bound to the unified BlockAccessList / AccountChanges
types we'd already split into ReadOnly*/Generated*/*AtIndex on this branch).
The index flattens the suggested BAL into 4 column-oriented lanes
(balance/nonce/code/storage) keyed by (accountOrdinal, key) at the row of
each tx index. ChangesEqual(other, index) then compares two indexes
row-by-row via ReadOnlySpan<T>.SequenceEqual — no per-account dict lookups,
no merge-walks — which is what ValidateBlockAccessList runs once per tx.
Wiring (BlockAccessListManager):
- PrepareForProcessing builds the suggested index once and pairs a
mutable generated index laid out identically. Also computes the
suggested chargeable-storage-reads tally once for the fast-path
surplus-reads gas check.
- MergeAndReturnBal grows an optional Action<BlockAccessListAtIndex>
callback; the parallel impl invokes it with the live slice between
target.Merge() and pool-return, the sequential impl with the slice
held on the worldstate. The manager-side hook (RegisterGeneratedSlice)
pushes the slice's rows into the mutable index and rolls the
generated-side read counter forward.
- ValidateBlockAccessList tries the index first: a single ChangesEqual
call plus the precomputed surplus-reads check. On mismatch (or before
the index is populated) it falls through to the existing streaming
walk that produces precise error diagnostics.
Adaptations from master's version:
- Build / Count / Fill bind to ReadOnlyBlockAccessList +
ReadOnlyAccountChanges, and read change arrays directly (no
ChangeSet.BlockAccessChanges indirection).
- Add(BlockAccessList) replaced with Add(BlockAccessListAtIndex slice):
our generated rows arrive as per-tx slices, one push per tx, each
contributing balance/nonce/code/storage at its own .Index.
- StorageLane values are EvmWord (matches StorageChange.Value's wire
type on this branch).
Tests: 12 new BlockAccessListValidationIndexTests covering exact match,
order-insensitive match (by address, by slot), large-row sort path,
overflow isolation, and mismatch detection on each lane. Suite passes
along with all existing BAL tests (244 total).
* Deduplicate BAL/EIP-8037 test bodies
- Collapse the two Eip8037BlockGasInclusionCheck boundary tests and the
two CalculateBlockRegularGas floor tests into single [TestCase]s.
- Collapse the three Eip7702 pre-validation authorization rejection tests
(max_nonce zero/max, high_s) into one [TestCaseSource]; the
post-validation existing_code variant stays separate because its
assertion shape differs.
- Fold the three error variants of debug_getRawBlockAccessList into one
[TestCaseSource] keyed on a setup callback and expected error code.
- Extract a BuildCreateFactory helper for the CREATE/CREATE2 factory
pattern repeated across three Eip8037Regression selfdestruct tests.
- Extract a SetupPrecompileBalScenario helper for the shared init ritual
used by the four *_under_PrecompileCachedCodeInfoRepository tests.
- Drop the spurious [Test] on the CodeInfoRepository_getcachedcodeinfo
parameterized test (combined with two [TestCase]s, NUnit would also
invoke it with no arguments).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* Drop redundant comments from dedup commit
Helper signatures and TestCase names already convey what the removed
comments restated.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* opts
* feedback
* refactor(bal): split BlockAccessListManager into partial files by concern
The class had grown to 919 lines across half a dozen concerns; primary
constructor parameters were buried under fields and four levels of nested
types. Split into five partial files, one concern each, with a docstring
on each partial declaration so a reader scanning the directory sees the
boundary before opening the file.
BlockAccessListManager.cs (215)
Class skeleton + primary constructor, fields, ParallelExecutionException,
public properties, lifecycle (PrepareForProcessing / Setup / Reset /
SpendGas / SetBlockExecutionContext / CheckInitialized), and the per-tx
hot path (GetTxProcessor / NextTransaction / Rollback / ReturnTxProcessor).
BlockAccessListManager.Validation.cs (258)
IncrementalValidation, the two CheckPerTxInclusion overloads,
ValidateBlockAccessList (fast-path + streaming slow-path), the
RegisterGeneratedSlice merge-hook, and IsSystemContract.
BlockAccessListManager.PrestateAndStateChanges.cs (174)
LoadPreStateToSuggestedBlockAccessList (populates the suggested BAL with
start-of-block values), the static ApplyStateChanges (writes BAL deltas
onto stateProvider), and SetBlockAccessList (finalises the produced block
with GeneratedBlockAccessList + RLP + hash).
BlockAccessListManager.SystemContracts.cs (74)
Bridge methods that route through pre-/post-execution slots of the tx
processor pool: StoreBeaconRoot, ApplyBlockhashStateChanges,
ApplyAuRaPreprocessingChanges, ProcessWithdrawals, ProcessExecutionRequests.
BlockAccessListManager.TxProcessorPool.cs (304)
Nested ITxProcessorWithWorldStateManager interface plus its
ParallelTxProcessorWithWorldStateManager and
SequentialTxProcessorWithWorldStateManager implementations and the
TxProcessorWithWorldState bundle. No behavioural change — just a move.
No public API changes, no behaviour changes. Verified via full
solution build, dotnet format clean, and the BAL-area test sweep:
50 Core BAL + 62 Consensus + 42 Blockchain + 17 BalRecorder +
72/73 State + 239 EVM tests all pass.
* Address Claude review on PR #11573 + skip per-tx BAL validation during sequential block building
- High (BlockAccessListManager.PrestateAndStateChanges.cs): _lastLoadedBal
was set before the try/catch, so a partial-load failure poisoned the dedup
key. A retry against the same hash silently skipped the load and workers
saw partial / default prestate. Moved the assignment to after the try
block; on success only.
- Medium (ReadOnlyAccountChanges.cs): replace LINQ Enumerable.SequenceEqual
on four arrays with MemoryExtensions.SequenceEqual over ReadOnlySpan<T>
(zero-alloc, no iterator, BCL non-LINQ — see coding-style.md). Drop the
now-unused `using System.Linq;`.
- Medium (BlockAccessListManager.cs): only the parallel path feeds the
generated validation index (RegisterGeneratedSlice is wired into the
parallel MergeAndReturnBal callback; the sequential NextTransaction
merges through WorldState.MergeGeneratingBal without the hook). So
_hasGeneratedValidationIndexUpdates never flips in sequential mode and
the fast path never triggers. Gate the index build on ParallelExecutionEnabled
to skip the O(n) BAL walk + lane sort entirely in sequential mode.
- High (ReadOnlyAccountChanges.cs): document why slot 0 cannot be starved by
ThreadPool pressure. The structural enforcement is in ParallelUnbalancedWork.For:
the calling thread is one of the workers (it runs InitProcessor.Execute()
inline at InitProcessor.For:305 rather than queueing it) and indices are
drawn from an atomic counter starting at fromInclusive, so the very first
GetNext() across all workers returns 0 — the calling thread can't be
pre-empted out of existing, so slot 0 always begins executing.
Plus: ParallelBlockValidationTransactionsExecutor.ProcessTransactionsSequential
now skips the per-tx ValidateBlockAccessList calls when building a block
(ProcessingOptions.ProducingBlock). There is no suggested BAL to compare
against during building — ValidateBlockAccessList would early-return on
`BlockAccessList is null` anyway, but skipping the call removes the
NextTransaction → Validate dance and makes the intent explicit on this
hot path.
* restore some balstore changes, remove workflows
* test(bal): restore inlined helpers in BlockValidatorTests
Three helpers got deleted during the BAL split refactor — only one of them
was actually tied to the unified BlockAccessList type. The other two were
deleted along with it and their bodies inlined into every test.
Restored:
- AmsterdamSut(ITxValidator? tx): factory for BlockValidator wired to
Amsterdam. Used 5x.
- AssertValidation(expected, actual, error, failPrefix): canonical
`isValid + error.StartsWith` assertion pattern. Used 5x.
- WithBal(this BlockBuilder, ReadOnlyBlockAccessList bal): file-scoped
extension that encodes the BAL, computes the hash, and chains the three
.With… calls. Retyped from `BlockAccessList` to `ReadOnlyBlockAccessList`
for our split refactor; used 4x.
Net delta in the file vs master shrinks from +85/-112 to +29/-71 — the
remainder is the necessary type renames (BlockAccessList → ReadOnly…,
AccountChanges → ReadOnly…), one rewritten test that has to build a
GeneratedBlockAccessList via BlockAccessListAtIndex (our split-refactor
generated side has no AddStorageChange on the immutable read-only type),
and the deletion of `..._fresh_item_count..._reused` (it relied on
`AddAccountRead` mutation which our types don't expose, and the cache-
invalidation concern doesn't apply because GeneratedBlockAccessList.ItemCount
computes fresh on every get).
49 BlockValidatorTests pass, lint clean.
* test(bal): restore inlined helpers in BlockProcessorTests
Re-introduces PrepareSetup, BuildGasResults, and ResultsForCount, and
collapses the two PrepareForProcessing_keeps_parallel_bal_execution_*
tests back into a single [TestCase(1)] [TestCase(2)] parametrized form,
matching the structure on master. The deletion of the helpers and the
parametrized test was unrelated to the BAL split refactor and only added
noise to the diff.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* refactor(bal): replace prestate gating with per-worker parent reader
Adopts master's parallel-execution architecture (PR #11436) for parent-state
reads while keeping our split BAL-type design:
* Parallel workers now rent an IReadOnlyTxProcessingScope from a pooled
IReadOnlyTxProcessingEnvFactory, scoped against the parent state root
captured at PrepareForProcessing. BlockAccessListBasedWorldState falls
through to that snapshot reader for any (address, slot) the suggested
BAL doesn't carry at the current block-access index.
* BAL completeness is preserved: reads for an account that isn't declared
in the suggested BAL at all throw InvalidBlockLevelAccessListException
with the same message format the sequential validator produces from
ValidateBlockAccessList.
* Parallel execution now requires stateProvider.IsInScope so a parent
state root is always capturable; sequential path takes over otherwise.
Drops prestate-loading machinery in favour of the parent reader:
* LoadPreStateToSuggestedBlockAccessList + per-account TaskCompletionSource
gate (EnablePrestateGate / SignalPrestateLoaded / WaitForPrestate) are
gone — workers no longer block on a per-account gate before reading
prestate-dependent state.
* ReadOnlyAccountChanges drops LoadPreStateBalance/Nonce/Code/Storage,
GetSlotsForPreStateLoad, RecordWasChanged, ExistedBeforeBlock, and
EmptyBeforeBlock; ReadOnlySlotChanges drops LoadPreStateChange.
* Adds HasStateChanges, IsStorageRead, and TryGetLast{Balance,Nonce,Code}
ChangeBefore convenience APIs the new world-state read paths need.
* IndexedChanges keeps its prestate-slot storage and the wire-level
PrestateIndex sentinel guard in IndexedChangeDecoder — both are dormant
at runtime now but still defend against a malicious peer.
The slot-0 loader iteration in
BlockProcessor.ParallelBlockValidationTransactionsExecutor disappears (only
ApplyStateChanges remains in iteration 0); slot 1 keeps the pre-execution
StoreBeaconRoot / ApplyBlockhashStateChanges step and continues to
synchronize with IncrementalValidation via preExecutionDoneTcs.
Test updates:
* BlockAccessListBasedWorldStateTests + the relevant TracedAccessWorldState
test now wire up a parent reader (the inner state itself, scoped against
genesis) and stop populating BAL prestate sentinels — covered values
live in the parent state instead.
* Removes AccountChangesPrestateTests and ReadOnlyAccountChangesTests
entirely (every test exercised a deleted API).
* Restores the parametrized
PrepareForProcessing_keeps_parallel_bal_execution_for_validated_eip8037_blocks
test and adds back the
PrepareForProcessing_disables_parallel_bal_execution_when_state_provider_is_not_scoped
test now that the IsInScope gate is in place.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* test(bal): restore tx-scheduling, parent-reader, and ApplyStateChanges tests
Tests deleted earlier as collateral damage from the BAL split refactor
that didn't carry the master-only features they exercised:
* Parallel_validation_execution_order_keeps_canonical_lead_and_sorts_tail_by_gas_limit
* Parallel_validation_execution_order_uses_stable_estimated_work_tie_breakers
* Parallel_validation_cancel_incomplete_gas_results_preserves_completed_slots
* Parallel_validation_uses_canonical_receipt_and_bal_indexes_with_scheduled_work_order
* Parallel_validation_parent_reader_scope_is_per_worker_and_disposed_on_return
* Parallel_validation_parent_reader_uses_parent_root_captured_before_pre_block_changes
* ValidateBlockAccessList_matches_accounts_by_address_when_insertion_order_differs
* ApplyStateChanges_uses_parent_state_without_prestate_sentinels (renamed
to ApplyStateChanges_replays_balance_nonce_and_storage_onto_parent_state)
* ApplyStateChanges_creates_missing_account_from_balance_change
The tx-scheduling, parent-reader, and ApplyStateChanges tests come back
verbatim against the now-available APIs. The two formerly-mutation-driven
tests (ValidateBlockAccessList + ApplyStateChanges) are ported to the
split BAL types: the suggested side is built immutably via the
Build.A.BlockAccessList / Build.An.AccountChanges builders, and any
generated-side mutation goes through BlockAccessListAtIndex.AddBalanceChange
followed by GeneratedBlockAccessList.Merge — preserving the original test
intents (insertion-order tolerance, parent-state replay, missing-account
materialisation) under the new architecture.
Re-introduces TrackingReadOnlyTxProcessingEnvFactory,
BalIndexRecordingTransactionProcessorAdapter, CreateTxForExecutionOrder,
CreateAuthorizationList, CreateAccessList, and the factory-overload
ctor on ParallelTestBlockAccessListManager that the scheduling test needs.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(bal): hide computed HasStateChanges from JSON serialization
ReadOnlyAccountChanges.HasStateChanges is a computed read-side helper
introduced for BlockAccessListBasedWorldState.GetAccountChanges; it must
not appear in the eth_getBlockAccessList* wire payload alongside the
real BAL fields, otherwise the EthRpcModule serialization tests diverge
from the spec-shaped output.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* test(bal): minimize BlockProcessorTests diff against master
Rebases BlockProcessorTests onto master's structure verbatim, applying
only the changes the BAL split refactor strictly requires:
* Test ordering matches master exactly.
* Test names match master exactly (no more "_replays_*" rename).
* Helper layout matches master.
* All ~12 `new BlockAccessList()` literals → `new ReadOnlyBlockAccessList()`.
* 5 IncrementalValidation call sites pass `Task.CompletedTask` (the extra
pre-execution gate param our branch's signature carries).
* `using Nethermind.Blockchain;` for IReadOnlyTxProcessor* interfaces.
* ParallelTestBlockAccessListManager: BlockAccessList → GeneratedBlockAccessList
type on the mock property, and the IncrementalValidation signature adds the
same `Task preExecutionTask` parameter.
Four tests are rebodied because their master forms mutate the unified
BlockAccessList type (AddBalanceChange/AddAccountRead/AddStorageRead) which
our split types don't expose on the suggested-side:
* ApplyStateChanges_uses_parent_state_without_prestate_sentinels
* ApplyStateChanges_creates_missing_account_from_balance_change
* ValidateBlockAccessList_storage_read_budget_uses_ItemCost
* ValidateBlockAccessList_matches_accounts_by_address_when_insertion_order_differs
On our branch each builds the suggested side immutably via
Build.A.BlockAccessList / Build.An.AccountChanges and (for the ValidateBlockAccessList
tests) emits the generated side via BlockAccessListAtIndex + GeneratedBlockAccessList.Merge.
The intents — apply replays balance/nonce/storage onto the parent state, BAL
size check uses fresh ItemCost, validator matches by address despite insertion
order — are preserved.
The previously-renamed `ApplyStateChanges_replays_balance_nonce_and_storage_onto_parent_state`
is renamed back to master's `ApplyStateChanges_uses_parent_state_without_prestate_sentinels`.
The `AddAccountRead` helper is deleted (it relied on unified-BAL mutation
and only existed to compose those four tests).
Verified all 28 master test names are present in our file. Diff vs master
shrinks from +361/-340 to +62/-42 — all remaining changes are the type and
signature deltas above. 1476 tests pass.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* refactor(bal): run BAL system contracts sequentially, drop preExecutionTask
Reverts to master's parallel-execution shape: StoreBeaconRoot and
ApplyBlockhashStateChanges run sequentially in BlockProcessor.ProcessBlock
*before* ParallelBlockValidationTransactionsExecutor.ProcessTransactions,
instead of as a dedicated iteration inside the parallel For.
Removes the extra synchronization the overlapped layout required:
* IBlockAccessListManager.IncrementalValidation drops Task preExecutionTask
(NullBlockAccessListManager and BlockAccessListManager.Validation follow).
The validator can merge balIndex=0 immediately on entry — the
pre-execution writes are already in the slice by then.
* ParallelBlockValidationTransactionsExecutor:
- parallel loop range len+2 → len+1; iteration 1 (the pre-execution
block + preExecutionDoneTcs SetResult/TrySetException) deleted.
- scheduled tx index goes back to txExecutionOrder[i-1] (was i-2);
balIndex remains txIndex+1.
- state tuple drops preExecutionDoneTcs.
- outer catch drops preExecutionDoneTcs.TrySetCanceled() and the
catch(TaskCanceledException) widens to master's
catch(OperationCanceledException ex) when (ex is TaskCanceledException
|| token.IsCancellationRequested).
* BlockProcessor.cs: the `if (!_balManager.ParallelExecutionEnabled)`
gate around the StoreBeaconRoot/ApplyBlockhashStateChanges/Commit
triple is removed — both paths now invoke pre-execution sequentially.
On the BAL path, _systemContractHandler is BlockAccessListSystemContractHandler,
which routes through balManager → GetPreExecution() → BAL slice for
balIndex=0; on the standard path, it routes through BeaconBlockRootHandler
+ BlockhashStore against stateProvider, same as master.
* BlockProcessorTests.cs + Eip8037BlockGasIntegrationTests.cs: drop the
extra Task.CompletedTask argument from every IncrementalValidation call
(6 + 6 sites) and from the ParallelTestBlockAccessListManager mock
signature.
Trade-off: pre-execution no longer overlaps with parallel tx execution.
For Amsterdam that's two system-contract calls per block — small enough
that diff parity with master wins on maintainability. Easy to revert if
benchmarks ever show pre-execution on the critical path.
Test sweep: Blockchain (1476), State (782), EVM (3579), Merge.Plugin
(1006) all pass.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthr…1 parent 0d11efa commit b1fa33d
7 files changed
Lines changed: 140 additions & 23 deletions
File tree
- src/Nethermind
- Nethermind.Blockchain.Test
- Nethermind.Consensus/Processing
- Nethermind.Core/BlockAccessLists
- Nethermind.Evm/Instructions
- Nethermind.Merge.Plugin.Test
Lines changed: 55 additions & 0 deletions
| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
| |||
597 | 597 | | |
598 | 598 | | |
599 | 599 | | |
| 600 | + | |
| 601 | + | |
| 602 | + | |
| 603 | + | |
| 604 | + | |
| 605 | + | |
| 606 | + | |
| 607 | + | |
| 608 | + | |
| 609 | + | |
| 610 | + | |
| 611 | + | |
| 612 | + | |
| 613 | + | |
| 614 | + | |
| 615 | + | |
| 616 | + | |
| 617 | + | |
| 618 | + | |
| 619 | + | |
| 620 | + | |
| 621 | + | |
| 622 | + | |
| 623 | + | |
| 624 | + | |
| 625 | + | |
| 626 | + | |
| 627 | + | |
| 628 | + | |
| 629 | + | |
| 630 | + | |
| 631 | + | |
| 632 | + | |
| 633 | + | |
| 634 | + | |
| 635 | + | |
| 636 | + | |
| 637 | + | |
| 638 | + | |
| 639 | + | |
| 640 | + | |
| 641 | + | |
| 642 | + | |
| 643 | + | |
| 644 | + | |
| 645 | + | |
| 646 | + | |
| 647 | + | |
| 648 | + | |
| 649 | + | |
| 650 | + | |
| 651 | + | |
| 652 | + | |
| 653 | + | |
| 654 | + | |
600 | 655 | | |
601 | 656 | | |
602 | 657 | | |
| |||
Lines changed: 15 additions & 10 deletions
| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
| |||
93 | 93 | | |
94 | 94 | | |
95 | 95 | | |
| 96 | + | |
96 | 97 | | |
97 | | - | |
98 | | - | |
99 | | - | |
100 | 98 | | |
101 | | - | |
102 | | - | |
103 | | - | |
104 | | - | |
105 | | - | |
106 | | - | |
107 | | - | |
| 99 | + | |
| 100 | + | |
| 101 | + | |
| 102 | + | |
| 103 | + | |
| 104 | + | |
| 105 | + | |
| 106 | + | |
| 107 | + | |
| 108 | + | |
| 109 | + | |
108 | 110 | | |
| 111 | + | |
| 112 | + | |
| 113 | + | |
109 | 114 | | |
110 | 115 | | |
Lines changed: 52 additions & 0 deletions
| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
| |||
8 | 8 | | |
9 | 9 | | |
10 | 10 | | |
| 11 | + | |
11 | 12 | | |
12 | 13 | | |
13 | 14 | | |
| |||
286 | 287 | | |
287 | 288 | | |
288 | 289 | | |
| 290 | + | |
| 291 | + | |
| 292 | + | |
| 293 | + | |
| 294 | + | |
| 295 | + | |
| 296 | + | |
| 297 | + | |
| 298 | + | |
| 299 | + | |
| 300 | + | |
| 301 | + | |
| 302 | + | |
| 303 | + | |
| 304 | + | |
| 305 | + | |
| 306 | + | |
| 307 | + | |
| 308 | + | |
| 309 | + | |
| 310 | + | |
| 311 | + | |
| 312 | + | |
| 313 | + | |
| 314 | + | |
| 315 | + | |
| 316 | + | |
| 317 | + | |
| 318 | + | |
| 319 | + | |
| 320 | + | |
| 321 | + | |
| 322 | + | |
| 323 | + | |
| 324 | + | |
| 325 | + | |
| 326 | + | |
| 327 | + | |
| 328 | + | |
| 329 | + | |
| 330 | + | |
| 331 | + | |
| 332 | + | |
| 333 | + | |
| 334 | + | |
| 335 | + | |
| 336 | + | |
| 337 | + | |
| 338 | + | |
| 339 | + | |
| 340 | + | |
289 | 341 | | |
Lines changed: 4 additions & 2 deletions
| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
| |||
101 | 101 | | |
102 | 102 | | |
103 | 103 | | |
104 | | - | |
| 104 | + | |
105 | 105 | | |
106 | 106 | | |
107 | 107 | | |
| |||
243 | 243 | | |
244 | 244 | | |
245 | 245 | | |
246 | | - | |
| 246 | + | |
| 247 | + | |
| 248 | + | |
247 | 249 | | |
248 | 250 | | |
249 | 251 | | |
| |||
Lines changed: 9 additions & 0 deletions
| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
| |||
19 | 19 | | |
20 | 20 | | |
21 | 21 | | |
| 22 | + | |
22 | 23 | | |
23 | 24 | | |
24 | 25 | | |
25 | 26 | | |
26 | 27 | | |
27 | 28 | | |
28 | 29 | | |
| 30 | + | |
| 31 | + | |
| 32 | + | |
| 33 | + | |
| 34 | + | |
| 35 | + | |
| 36 | + | |
29 | 37 | | |
30 | 38 | | |
31 | 39 | | |
| |||
41 | 49 | | |
42 | 50 | | |
43 | 51 | | |
| 52 | + | |
44 | 53 | | |
45 | 54 | | |
46 | 55 | | |
| |||
Lines changed: 0 additions & 1 deletion
| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
| |||
618 | 618 | | |
619 | 619 | | |
620 | 620 | | |
621 | | - | |
622 | 621 | | |
623 | 622 | | |
624 | 623 | | |
| |||
Lines changed: 5 additions & 10 deletions
| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
| |||
291 | 291 | | |
292 | 292 | | |
293 | 293 | | |
294 | | - | |
295 | | - | |
296 | | - | |
297 | | - | |
298 | | - | |
299 | | - | |
300 | | - | |
301 | | - | |
302 | | - | |
303 | | - | |
| 294 | + | |
| 295 | + | |
| 296 | + | |
| 297 | + | |
| 298 | + | |
304 | 299 | | |
305 | 300 | | |
306 | 301 | | |
| |||
0 commit comments