EVM Optimisations - Improves 70% of EEST benchmarks by siladu · Pull Request #9775 · hyperledger/besu

siladu · 2026-02-10T06:56:04Z

Optimisations found while profiling REVERT benchmark:
https://github.com/ethereum/execution-specs/blob/forks/amsterdam/tests/benchmark/compute/instruction/test_system.py
https://github.com/ethereum/execution-spec-tests/releases/tag/benchmark%40v0.0.7

These are all minimally-invasive fast path optimisations:

AbstractMessageProcessor.clearAccumulatedStateBesidesGasAndOutput - use java 7-style loop instead of Stream API
AbstractMessageProcessor.clearAccumulatedStateBesidesGasAndOutput - revert fast path when touched accounts are empty
MessageFrame.build - Skip redundant warmUpAddress for contract
AbstractCallOperation.complete - Skip empty collection adds
StackedUpdater.getTouchedAccounts - Returns emptyList when no updates exist, avoiding ArrayList allocation for empty STATICCALL frames

Benchmarks

On m6a.2xlarge, ran 5 rounds of:

evmtool --repeat 10 block-test ~/2026-02-09_fixtures_benchmark_v0.0.7/blockchain_tests/benchmark/compute/*/*/* --test-name 100M > block-test-control-bf64be2_100M_1.log

versus

./build/install/besu/bin/evmtool --repeat 10 block-test ~/2026-02-09_fixtures_benchmark_v0.0.7/blockchain_tests/benchmark/compute/*/*/* --test-name 100M > block-test-evm-fast-paths_100M_1.log

Analysis:

python3 ../compare_benchmarks.py --control *control* --opt *evm-fast-paths* --output 2026-02-13_benchmark-comparison_evm-fast-paths.md

Tests compared: 940
Anomaly threshold: CV > 15.0%
Anomalous tests: 293 (31.2%)

Values shown are the mean across rounds.

Metric	All Tests	Stable Only (excl. anomalies)
Total	940	647
Improvements (>1%)	662 (70.4%)	379 (58.6%)
Regressions (<-1%)	90 (9.6%)	81 (12.5%)
Neutral (-1% to +1%)	188 (20.0%)	187 (28.9%)
Mean change	+6.46%	+3.28%
Median change	+6.08%	+1.80%

Slow test reports - where OPCODEs improved/regressed wrt 60MGas/s

python3 ../slow_tests_report.py --files *evm-fast-paths*.log --control *control*.log --output 2026-02-13_slow-tests_evm-fast-paths-60MGasps.md --threshold 60

  Threshold: 60.0 MGas/s
  Control slow: 46  |  Test slow: 45  |  Net: +1
  Escaped: 2  |  Still slow: 44  |  Newly slow: 1

  Escaped by opcode:
    SHL                         1 tests  (+5.42%)
    SHR                         1 tests  (+5.71%)

  Newly slow by opcode:
    ADD                         1 tests  (-2.79%)

Escaped Tests (2 tests now above 60.0 MGas/s)

By Opcode

Opcode	Count	Ctrl Range	Test Range	Mean Change
SHL	1	58.56-58.56	61.74-61.74	+5.42%
SHR	1	57.02-57.02	60.28-60.28	+5.71%

All Escaped Tests

Ctrl Mean	Test Mean	Change	Opcode	Test
57.02	60.28	+5.71%	SHR	`test_bitwise[opcode_SHR- @ 100M`
58.56	61.74	+5.42%	SHL	`test_bitwise[opcode_SHL- @ 100M`

Newly Slow Tests (1 tests now below 60.0 MGas/s)

By Opcode

Opcode	Count	Ctrl Range	Test Range	Mean Change
ADD	1	60.21-60.21	58.53-58.53	-2.79%

All Newly Slow Tests

Ctrl Mean	Test Mean	Change	Opcode	Test
60.21	58.53	-2.79%	ADD	`test_arithmetic[opcode_ADD- @ 100M`

Still Slow Tests (44 tests still below 60.0 MGas/s)

Summary by Opcode

Opcode	Count	Min MGas/s	Max MGas/s	Mean Change
MULMOD	4	29.02	35.01	+0.53%
SDIV	2	30.52	30.66	-0.47%
EQ	1	32.26	32.26	+3.15%
SMOD	4	33.42	42.09	+3.49%
MOD	5	34.33	52.50	+5.27%
SAR	2	37.28	44.70	+3.86%
POINT_EVALUATION	1	38.88	38.88	+0.29%
SUB	1	39.07	39.07	-1.87%
MUL	1	42.88	42.88	+3.39%
DIV	2	43.20	44.22	+5.09%
ADDMOD	4	46.06	55.98	+2.84%
KECCAK_MAX_PERMUTATIONS	1	53.78	53.78	+0.06%
KECCAK	3	54.92	55.14	+1.43%
KECCAK_DIFF_MEM_MSG_SIZES	8	57.03	58.29	+0.94%
REVERT	1	59.30	59.30	+10.27%
BLAKE2F	4	59.68	59.82	+1.45%

This PR helps with higher gas limits too:

100MGas/s

python3 ../slow_tests_report.py --files *evm-fast-paths*.log --control *control*.log --output 2026-02-13_slow-tests_evm-fast-paths-100MGasps.md --threshold 100

  Threshold: 100.0 MGas/s
  Control slow: 208  |  Test slow: 168  |  Net: +40
  Escaped: 40  |  Still slow: 168  |  Newly slow: 0

  Escaped by opcode:
    CALLDATACOPY_FROM_CALL      1 tests  (+6.88%)
    GASLIMIT                    1 tests  (+14.31%)
    GAS_OP                      1 tests  (+14.01%)
    MSIZE                       3 tests  (+13.58%)
    MSTORE                     30 tests  (+8.60%)
    NUMBER                      1 tests  (+14.11%)
    RETURNDATASIZE_ZERO         1 tests  (+13.77%)
    REVERT                      1 tests  (+9.12%)
    TIMESTAMP                   1 tests  (+13.76%)

180MGas/s

python3 ../slow_tests_report.py --files *evm-fast-paths*.log --control *control*.log --output 2026-02-13_slow-tests_evm-fast-paths-180MGasps.md --threshold 180

  Threshold: 180.0 MGas/s
  Control slow: 458  |  Test slow: 435  |  Net: +23
  Escaped: 23  |  Still slow: 435  |  Newly slow: 0

  Escaped by opcode:
    ALT_BN128                   2 tests  (+0.14%)
    CALLVALUE_FROM_ORIGIN       2 tests  (+25.83%)
    JUMPDESTS                   1 tests  (+47.10%)
    PUSH0                       1 tests  (+23.52%)
    PUSH6                       1 tests  (+11.94%)
    SWAP1                       1 tests  (+10.12%)
    SWAP10                      1 tests  (+10.55%)
    SWAP11                      1 tests  (+10.63%)
    SWAP12                      1 tests  (+10.33%)
    SWAP13                      1 tests  (+9.64%)
    SWAP14                      1 tests  (+10.11%)
    SWAP15                      1 tests  (+10.02%)
    SWAP16                      1 tests  (+10.23%)
    SWAP2                       1 tests  (+10.10%)
    SWAP3                       1 tests  (+10.52%)
    SWAP4                       1 tests  (+10.29%)
    SWAP5                       1 tests  (+9.98%)
    SWAP6                       1 tests  (+10.29%)
    SWAP7                       1 tests  (+10.37%)
    SWAP8                       1 tests  (+10.93%)
    SWAP9                       1 tests  (+10.25%)

Files

2026-02-13_benchmark-comparison_evm-fast-paths.md
2026-02-13_slow-tests_evm-fast-paths-60MGasps.md
2026-02-13_slow-tests_evm-fast-paths-100MGasps.md
2026-02-13_slow-tests_evm-fast-paths-180MGasps.md
compare_benchmarks.py
slow_tests_report.py

Testing

Mainnet sync
Hoodi full sync
ARM benchmarks

When a child frame reverts with no touched accounts (common for STATICCALL with empty REVERT), skip the stream/filter/collect pipeline and go straight to revert+commit. Profiler shows ~5.9% of CPU in the revert path: - ConcurrentHashMap$Traverser.advance (from getTouchedAccounts) - ConcurrentHashMap.clear (from revert) - MessageFrame.clearGasRefund Most of this work is wasted when collections are empty. Signed-off-by: Simon Dudley <simon.dudley@consensys.net>

Signed-off-by: Simon Dudley <simon.dudley@consensys.net>

…when no updates exist, avoiding ArrayList allocation for empty STATICCALL frames Signed-off-by: Simon Dudley <simon.dudley@consensys.net>

Signed-off-by: Simon Dudley <simon.dudley@consensys.net>

evm/src/main/java/org/hyperledger/besu/evm/processor/AbstractMessageProcessor.java

evm/src/main/java/org/hyperledger/besu/evm/frame/MessageFrame.java

evm/src/main/java/org/hyperledger/besu/evm/operation/AbstractCallOperation.java

evm/src/main/java/org/hyperledger/besu/evm/processor/AbstractMessageProcessor.java

lu-pinto

At least we should check while creating the MessageFrame if we could preload it with the warmed up addresses instead of doing another null check.
Also investigate if we really need those isEmpty checks in AbstractCallOperation, like doing some bench runs solely without them

… Set Signed-off-by: Luis Pinto <luis.pinto@consensys.net>

Signed-off-by: Simon Dudley <simon.dudley@consensys.net>

This is a bug as was left over from BytesTrieSet which had different semantics Signed-off-by: Simon Dudley <simon.dudley@consensys.net>

lu-pinto

LGTM

ahamlat

LGTM

matkt

Checked the updater part and it is ok . I don't see any issues

siladu · 2026-02-13T01:53:11Z

ADD -2.79%

Confirmed this regression is noise with more targetted and better warmed up runs (on VM) and JMH (on Mac):

Before this PR:

Benchmark                               Mode  Cnt   Score   Error  Units
AddOperationBenchmark.executeOperation  avgt   15  37.448 ± 0.843  ns/op

This PR:

Benchmark                               Mode  Cnt   Score   Error  Units
AddOperationBenchmark.executeOperation  avgt   15  37.306 ± 0.823  ns/op

Also profiling analysis (on Mac):

JAVA_OPTS=-agentpath:~/tools/async-profiler-4.2.1-macos/lib/libasyncProfiler.dylib=start,event=cpu,file=bf64be2-control-ADD.coll ./build/install/besu/bin/evmtool --repeat 100 block-test ~/Documents/eest-benchmark-fixtures/2026-02-09_fixtures_benchmark_v0.0.7/blockchain_tests/benchmark/compute/instruction/arithmetic/arithmetic.json --test-name "*ADD-*100M*"

Profile Comparison: Control (bf64be2) vs Fast-paths (evm-fast-paths) — ADD workload

  Overview
  ┌───────────────┬─────────┬────────────┐
  │    Metric     │ Control │ Fast-paths │
  ├───────────────┼─────────┼────────────┤
  │ Total samples │ 5,596   │ 8,250      │
  ├───────────────┼─────────┼────────────┤
  │ Sample ratio  │ 1.0x    │ 1.47x      │
  └───────────────┴─────────┴────────────┘
  Since total samples differ, percentage share is the correct metric for comparison.

  Key Wins — Stack Manipulation Overhead Reduced
  ┌───────────────────────────────────────┬───────────┬──────────────┬─────────┐
  │                Method                 │ Control % │ Fast-paths % │ Change  │
  ├───────────────────────────────────────┼───────────┼──────────────┼─────────┤
  │ FlexStack.* (all)                     │ 14.49%    │ 9.08%        │ -5.41pp │
  ├───────────────────────────────────────┼───────────┼──────────────┼─────────┤
  │ FlexStack.pop                         │ 5.61%     │ 2.85%        │ -2.76pp │
  ├───────────────────────────────────────┼───────────┼──────────────┼─────────┤
  │ FlexStack.push                        │ 6.77%     │ 4.78%        │ -1.99pp │
  ├───────────────────────────────────────┼───────────┼──────────────┼─────────┤
  │ FlexStack.get                         │ 2.07%     │ 1.26%        │ -0.81pp │
  ├───────────────────────────────────────┼───────────┼──────────────┼─────────┤
  │ EVM.runToHalt                         │ 12.42%    │ 10.12%       │ -2.30pp │
  ├───────────────────────────────────────┼───────────┼──────────────┼─────────┤
  │ AddOperationOptimized.staticOperation │ 5.02%     │ 3.79%        │ -1.23pp │
  ├───────────────────────────────────────┼───────────┼──────────────┼─────────┤
  │ MessageFrame.popStackItem             │ 2.07%     │ 1.33%        │ -0.74pp │
  └───────────────────────────────────────┴───────────┴──────────────┴─────────┘
  The fast-path is successfully reducing stack manipulation overhead — FlexStack operations dropped 5.41pp collectively.

  Concerning — UInt256 byte conversions now dominate
  ┌──────────────────────────────────┬───────────┬──────────────┬──────────┐
  │              Method              │ Control % │ Fast-paths % │  Change  │
  ├──────────────────────────────────┼───────────┼──────────────┼──────────┤
  │ UInt256.* (all)                  │ 39.33%    │ 49.95%       │ +10.62pp │
  ├──────────────────────────────────┼───────────┼──────────────┼──────────┤
  │ UInt256.fromBytesBE              │ 12.76%    │ 17.53%       │ +4.77pp  │
  ├──────────────────────────────────┼───────────┼──────────────┼──────────┤
  │ UInt256.getIntBE                 │ 13.21%    │ 17.35%       │ +4.14pp  │
  ├──────────────────────────────────┼───────────┼──────────────┼──────────┤
  │ UInt256.toBytesBE                │ 0.98%     │ 2.00%        │ +1.02pp  │
  ├──────────────────────────────────┼───────────┼──────────────┼──────────┤
  │ UInt256.addWithCarry             │ 10.85%    │ 11.47%       │ +0.62pp  │
  ├──────────────────────────────────┼───────────┼──────────────┼──────────┤
  │ MessageFrame.setCurrentOperation │ 4.13%     │ 8.22%        │ +4.09pp  │
  └──────────────────────────────────┴───────────┴──────────────┴──────────┘
  With the stack overhead removed, UInt256.fromBytesBE/getIntBE/toBytesBE byte serialization is now the dominant bottleneck (~36% of the fast-paths profile).

  EVM vs Non-EVM
  ┌───────────────────────┬─────────┬────────────┐
  │       Category        │ Control │ Fast-paths │
  ├───────────────────────┼─────────┼────────────┤
  │ EVM-related           │ 80.81%  │ 84.93%     │
  ├───────────────────────┼─────────┼────────────┤
  │ Non-EVM (JIT, OS, GC) │ 16.19%  │ 13.33%     │
  └───────────────────────┴─────────┴────────────┘
  Interpretation

  1. Stack overhead successfully reduced — FlexStack ops down 5.41pp, confirming the fast-path bypasses MessageFrame.popStackItem -> FlexStack.pop for ADD.
  2. UInt256 byte conversion is the new bottleneck — fromBytesBE+getIntBE+toBytesBE went from 26.95% to 36.88%. These proportional increases likely reflect the removal of other overhead making these stand out more.
  3. MessageFrame.setCurrentOperation doubled (4.13% → 8.22%) — possibly because the main loop runs faster now, so more time is proportionally spent on per-iteration bookkeeping.
  4. Next optimization target: If the ADD fast-path could work with native long values directly (avoiding the bytes-to-int round-trip in UInt256), that could yield further gains.

Files

profile_compare.py
evm-fast-paths-ADD.coll.txt
bf64be2-control-ADD.coll.txt

ahamlat · 2026-02-13T20:22:46Z

Since total samples differ, percentage share is the correct metric for comparison.

Why are there more samples in the optimized version ? in general that indicates that the program takes more time to execute.

siladu · 2026-02-16T07:08:44Z

ARM results

On m8g.2xlarge, ran 5 rounds of:

evmtool --repeat 10 block-test ~/2026-02-09_fixtures_benchmark_v0.0.7/blockchain_tests/benchmark/compute/*/*/* --test-name 100M > ~/block-test_benchmark_compute_control_bf64be2_100M_arm1.log

versus

evmtool --repeat 10 block-test ~/2026-02-09_fixtures_benchmark_v0.0.7/blockchain_tests/benchmark/compute/*/*/* --test-name 100M > ~/block-test_benchmark_compute_evm-fast-paths_100M_arm1.log

Analysis:

python3 ../compare_benchmarks.py --control *control*arm*log --opt *paths*arm*log --output 2026-02-16_benchmark-comparison_control-vs-evm-fast-paths-arm.md

Tests compared: 940
Anomaly threshold: CV > 15.0%
Anomalous tests: 293 (31.2%)

Values shown are the mean across rounds.

Metric	All Tests	Stable Only (excl. anomalies)
Total	940	852
Improvements (>1%)	239 (25.4%)	234 (27.5%)
Regressions (<-1%)	472 (50.2%)	389 (45.7%)
Neutral (-1% to +1%)	229 (24.4%)	229 (26.9%)
Mean change	-1.39%	-0.66%
Median change	-1.02%	-0.37%

The biggest regressions are all well above 60MGas/s.

Slow test reports - where OPCODEs improved/regressed wrt 60MGas/s

python3 ../slow_tests_report.py --files *paths*arm*log --control *control*arm*.log --output 2026-02-16_slow-tests_control-vs-evm-fast-paths-60MGasps-arm.md --threshold 60

Threshold: 60.0 MGas/s
  Control slow: 64  |  Test slow: 66  |  Net: -2
  Escaped: 4  |  Still slow: 60  |  Newly slow: 6

  Escaped by opcode:
    CALL                        3 tests  (+3.24%)
    DELEGATECALL                1 tests  (+3.87%)

  Newly slow by opcode:
    GASLIMIT                    1 tests  (-1.23%)
    MSIZE                       2 tests  (-1.73%)
    NUMBER                      1 tests  (-0.99%)
    RETURNDATASIZE_ZERO         1 tests  (-1.94%)
    TIMESTAMP                   1 tests  (-1.41%)

Escaped Tests (2 tests now above 60.0 MGas/s)

By Opcode

Opcode	Count	Ctrl Range	Test Range	Mean Change
SHL	1	58.56-58.56	61.74-61.74	+5.42%
SHR	1	57.02-57.02	60.28-60.28	+5.71%

All Escaped Tests

Ctrl Mean	Test Mean	Change	Opcode	Test
59.83	62.15	+3.87%	DELEGATECALL	`test_ext_account_query_warm[initial_storage_True-initial_balance_True-empty_code_False-opcode_DELEGATECALL @ 100M`
59.77	61.79	+3.38%	CALL	`test_ext_account_query_warm[initial_storage_False-initial_balance_False-empty_code_True-opcode_CALL @ 100M`
58.72	60.64	+3.26%	CALL	`test_ext_account_query_warm[initial_storage_True-initial_balance_True-empty_code_True-opcode_CALL @ 100M`
58.72	60.52	+3.07%	CALL	`test_ext_account_query_warm[initial_storage_False-initial_balance_True-empty_code_True-opcode_CALL @ 100M`

Newly Slow Tests (6 tests now below 60.0 MGas/s)

By Opcode

Opcode	Count	Ctrl Range	Test Range	Mean Change
GASLIMIT	1	60.35-60.35	59.61-59.61	-1.23%
MSIZE	2	60.75-60.83	59.66-59.81	-1.73%
NUMBER	1	60.33-60.33	59.73-59.73	-0.99%
RETURNDATASIZE_ZERO	1	60.55-60.55	59.37-59.37	-1.94%
TIMESTAMP	1	60.51-60.51	59.66-59.66	-1.41%

All Newly Slow Tests

Ctrl Mean	Test Mean	Change	Opcode	Test
60.55	59.37	-1.94%	RETURNDATASIZE_ZERO	`test_returndatasize_zero[benchmark-gas-value_100M`
60.35	59.61	-1.23%	GASLIMIT	`test_block_context_ops[opcode_GASLIMIT @ 100M`
60.51	59.66	-1.41%	TIMESTAMP	`test_block_context_ops[opcode_TIMESTAMP @ 100M`
60.75	59.66	-1.78%	MSIZE	`test_msize[mem_size_1 @ 100M`
60.33	59.73	-0.99%	NUMBER	`test_block_context_ops[opcode_NUMBER @ 100M`
60.83	59.81	-1.68%	MSIZE	`test_msize[mem_size_0 @ 100M`

Still Slow Tests (60 tests still below 60.0 MGas/s)

Summary by Opcode

Opcode	Count	Min MGas/s	Max MGas/s	Mean Change
MULMOD	4	28.55	32.90	-1.75%
SMOD	4	30.11	36.45	-2.38%
EQ	1	32.09	32.09	-7.11%
POINT_EVALUATION	1	34.19	34.19	-0.19%
MOD	4	35.85	44.96	-3.10%
SDIV	2	36.80	37.30	-1.36%
SUB	1	42.42	42.42	-6.31%
ADDMOD	4	42.75	48.49	-2.51%
BLAKE2F	5	43.07	54.50	+2.92%
SAR	2	43.38	49.17	-2.54%
DIV	2	43.57	43.93	-1.77%
MUL	1	45.13	45.13	-2.49%
ADD	1	51.25	51.25	-6.20%
CODESIZE	1	51.43	51.43	+2.68%
REVERT	1	51.57	51.57	+13.20%
CALLDATACOPY_FROM_ORIGIN	6	54.50	57.41	-2.03%
CALLDATACOPY_FROM_CALL	1	55.14	55.14	-2.45%
CALLCODE	7	56.79	57.96	+3.13%
CALLDATASIZE	8	57.94	58.06	-2.71%
CALL	4	59.27	59.87	+3.14%

Files

2026-02-16_benchmark-comparison_control-vs-evm-fast-paths-arm.md
2026-02-16_slow-tests_control-vs-evm-fast-paths-60MGasps-arm.md

siladu · 2026-02-16T07:46:48Z

Why are there more samples in the optimized version ? in general that indicates that the program takes more time to execute.

Good spot. This was an analysis error as the script normalised the percentages. Attached the updated script.

I ran profiling on my mac so probably too noisy for ADD. Reran on m6a.2xlarge and think it's within acceptable noise...

python3 ../profile_compare_updated.py bf64be2-control-ADD.coll evm-fast-paths-ADD.coll
Parsing profiles...
  Control:    bf64be2-control-ADD.coll
  Fast-paths: evm-fast-paths-ADD.coll

==================================================================================================================================
  PROFILE OVERVIEW
==================================================================================================================================
  Control:      12473 samples  ~124.7s wall time   100 unique stacks
  Fast-paths:   12710 samples  ~127.1s wall time   100 unique stacks
  Difference:    +237 samples  (+1.9%)
  ==> REGRESSION: fast-paths is ~2% SLOWER (more samples = more wall time)

  NOTE: Both profiles run the same workload. Absolute sample counts are directly
  comparable. More samples = more time spent. Percentage share is secondary.

==================================================================================================================================
  FINAL SUMMARY
==================================================================================================================================

  Wall time:  Control ~124.7s  vs  Fast-paths ~127.1s  (+1.9%)

  Category                                          Ctrl       FP    Delta    % Chg
  --------------------------------------------------------------------------------
  AddOperation*                                      153      151       -2    -1.3%
  UInt256.*                                         5169     5192      +23    +0.4%
  FlexStack.*                                       1978     1831     -147    -7.4%
  MessageFrame.*                                    2095     2300     +205    +9.8%
  EVM.runToHalt                                      340      758     +418  +122.9%
  Non-EVM (JIT/OS/GC)                               2057     1902     -155    -7.5%
  TOTAL                                            12473    12710     +237 +1.9%

Files

profile_compare_updated.py
bf64be2-control-ADD.coll.txt
evm-fast-paths-ADD.coll.txt

Signed-off-by: Simon Dudley <simon.dudley@consensys.net>

siladu added 5 commits February 10, 2026 16:56

Invert if else for readability

100b630

Signed-off-by: Simon Dudley <simon.dudley@consensys.net>

Skip redundant warmUpAddress for contract in MessageFrame

d077be1

Signed-off-by: Simon Dudley <simon.dudley@consensys.net>

Skip empty collection adds in AbstractCallOperation.complete()

38d2468

Signed-off-by: Simon Dudley <simon.dudley@consensys.net>

StackedUpdater.getTouchedAccounts(): Returns Collections.emptyList() …

7b284dc

…when no updates exist, avoiding ArrayList allocation for empty STATICCALL frames Signed-off-by: Simon Dudley <simon.dudley@consensys.net>

siladu force-pushed the evm-fast-paths branch from 303cfee to 7b284dc Compare February 10, 2026 06:57

siladu added 2 commits February 10, 2026 16:57

Merge branch 'main' into evm-fast-paths

7392b50

spotless

cfde3b7

Signed-off-by: Simon Dudley <simon.dudley@consensys.net>

ahamlat reviewed Feb 10, 2026

View reviewed changes

evm/src/main/java/org/hyperledger/besu/evm/processor/AbstractMessageProcessor.java Outdated Show resolved Hide resolved

evm/src/main/java/org/hyperledger/besu/evm/processor/AbstractMessageProcessor.java Outdated Show resolved Hide resolved

lu-pinto reviewed Feb 10, 2026

View reviewed changes

evm/src/main/java/org/hyperledger/besu/evm/frame/MessageFrame.java Outdated Show resolved Hide resolved

lu-pinto reviewed Feb 10, 2026

View reviewed changes

evm/src/main/java/org/hyperledger/besu/evm/operation/AbstractCallOperation.java Show resolved Hide resolved

lu-pinto reviewed Feb 10, 2026

View reviewed changes

evm/src/main/java/org/hyperledger/besu/evm/processor/AbstractMessageProcessor.java Show resolved Hide resolved

lu-pinto requested changes Feb 10, 2026

View reviewed changes

Enforce AbstractMessageProcessor.forceDeleteAccountsWhenEmpty to be a…

3f18df6

… Set Signed-off-by: Luis Pinto <luis.pinto@consensys.net>

lu-pinto force-pushed the evm-fast-paths branch from a03a944 to 3f18df6 Compare February 10, 2026 11:53

siladu added 5 commits February 11, 2026 15:57

Merge branch 'main' into evm-fast-paths

6872c1d

Java-7-ify clearAccumulatedStateBesidesGasAndOutput

c6e1110

Signed-off-by: Simon Dudley <simon.dudley@consensys.net>

spotless

a9f1f10

Signed-off-by: Simon Dudley <simon.dudley@consensys.net>

Move contract warming into newTxValues creation

0ffeafc

Signed-off-by: Simon Dudley <simon.dudley@consensys.net>

Remove initialCapacity from HashMap

2d9ed45

This is a bug as was left over from BytesTrieSet which had different semantics Signed-off-by: Simon Dudley <simon.dudley@consensys.net>

lu-pinto approved these changes Feb 11, 2026

View reviewed changes

ahamlat approved these changes Feb 11, 2026

View reviewed changes

siladu added the performance label Feb 12, 2026

github-project-automation bot added this to Performance Feb 12, 2026

siladu moved this to In Progress in Performance Feb 12, 2026

siladu moved this from In Progress to In Review in Performance Feb 12, 2026

siladu self-assigned this Feb 12, 2026

siladu added 2 commits February 12, 2026 15:14

Merge branch 'main' into evm-fast-paths

eb1a324

Merge branch 'main' into evm-fast-paths

d3ee24c

matkt approved these changes Feb 12, 2026

View reviewed changes

siladu changed the title ~~EVM fast path Optimisations - Improves STATICCALL and REVERT benchmarks~~ EVM fast path Optimisations - Improves 70% of EEST benchmarks Feb 12, 2026

Merge branch 'main' into evm-fast-paths

1c503e0

siladu changed the title ~~EVM fast path Optimisations - Improves 70% of EEST benchmarks~~ EVM Optimisations - Improves 70% of EEST benchmarks Feb 16, 2026

changelog

abdf9c4

Signed-off-by: Simon Dudley <simon.dudley@consensys.net>

siladu enabled auto-merge (squash) February 16, 2026 10:19

siladu merged commit c4ccec8 into hyperledger:main Feb 16, 2026
46 checks passed

github-project-automation bot moved this from In Review to Done in Performance Feb 16, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

EVM Optimisations - Improves 70% of EEST benchmarks#9775

EVM Optimisations - Improves 70% of EEST benchmarks#9775
siladu merged 17 commits intohyperledger:mainfrom
siladu:evm-fast-paths

siladu commented Feb 10, 2026 •

edited

Loading

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

lu-pinto left a comment •

edited

Loading

Uh oh!

lu-pinto left a comment

Uh oh!

ahamlat left a comment

Uh oh!

matkt left a comment

Uh oh!

siladu commented Feb 13, 2026 •

edited

Loading

Uh oh!

ahamlat commented Feb 13, 2026

Uh oh!

siladu commented Feb 16, 2026

Uh oh!

siladu commented Feb 16, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Conversation

siladu commented Feb 10, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Benchmarks

Slow test reports - where OPCODEs improved/regressed wrt 60MGas/s

Escaped Tests (2 tests now above 60.0 MGas/s)

By Opcode

All Escaped Tests

Newly Slow Tests (1 tests now below 60.0 MGas/s)

By Opcode

All Newly Slow Tests

Still Slow Tests (44 tests still below 60.0 MGas/s)

Summary by Opcode

Files

Testing

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

lu-pinto left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

lu-pinto left a comment

Choose a reason for hiding this comment

Uh oh!

ahamlat left a comment

Choose a reason for hiding this comment

Uh oh!

matkt left a comment

Choose a reason for hiding this comment

Uh oh!

siladu commented Feb 13, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Files

Uh oh!

ahamlat commented Feb 13, 2026

Uh oh!

siladu commented Feb 16, 2026

ARM results

Slow test reports - where OPCODEs improved/regressed wrt 60MGas/s

Escaped Tests (2 tests now above 60.0 MGas/s)

By Opcode

All Escaped Tests

Newly Slow Tests (6 tests now below 60.0 MGas/s)

By Opcode

All Newly Slow Tests

Still Slow Tests (60 tests still below 60.0 MGas/s)

Summary by Opcode

Files

Uh oh!

siladu commented Feb 16, 2026

Files

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

siladu commented Feb 10, 2026 •

edited

Loading

lu-pinto left a comment •

edited

Loading

siladu commented Feb 13, 2026 •

edited

Loading