Fuse XOR-and-permute in BMI2 assembly and minor cleanups#8
Merged
Conversation
Combine the XOR of rate bytes into state with the Keccak-f[1600] permutation in a single assembly call, eliminating one full memory pass over the 200-byte state. Remove the Go-level xorIn helper and its unsafe usage. Also: reuse s.buf directly in padAndSqueeze (avoids lastBlock stack alloc), remove redundant BenchmarkSum256_500K, simplify benchName, and clarify D-value register comments in the generator. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
There was a problem hiding this comment.
Pull request overview
This PR updates the accelerated Keccak-256 implementation to reduce overhead on amd64 BMI2 by allowing the assembly permutation to optionally XOR a full rate block into the state before permuting, and performs a few related cleanups in padding and benchmarks.
Changes:
- Extend
keccakF1600BMI2(amd64/BMI2) to accept an optionalbuf *byteand XOR 136 bytes into the state before running the 24 rounds when provided. - Refactor final-block padding to avoid
unsafe-based XOR helpers and reuse the sponge buffer inpadAndSqueeze. - Remove a redundant benchmark and simplify benchmark naming logic; update generator comments accordingly.
Reviewed changes
Copilot reviewed 5 out of 5 changed files in this pull request and generated no comments.
Show a summary per file
| File | Description |
|---|---|
keccakf_amd64_bmi2.s |
Adds optional pre-round XOR of 17 lanes (rate) from buf into the state before permutation. |
keccak_amd64.go |
Updates the BMI2 function signature and routes keccakF1600 / xorAndPermute through the new assembly entrypoint. |
keccak_asm.go |
Removes unsafe xor helper; updates final padding paths to use xorAndPermute and in-place buffer padding for squeezing. |
gen_keccakf_bmi2.go |
Updates the generator to emit the new BMI2 entrypoint signature and the optional XOR prelude; clarifies comments. |
keccak_test.go |
Removes a duplicate benchmark and simplifies benchName. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
This was referenced Apr 7, 2026
Merged
There was a problem hiding this comment.
Pull request overview
Copilot reviewed 7 out of 7 changed files in this pull request and generated 1 comment.
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
After Read is called, padAndSqueeze mutates the sponge state (XORs padding, permutes). A subsequent Sum256 would re-apply padding and permute a copy of this already-finalized state, producing an incorrect digest. Panic to make this misuse visible. Note: this differs from x/crypto/sha3, where Sum() works after Read() by cloning the full sponge state before finalizing. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Re-add xorIn using binary.LittleEndian.Uint64 (no unsafe) for the padding/finalization path (Sum256, padAndSqueeze, sum256Sponge). Keep fused xorAndPermute assembly only for full blocks in Write. This eliminates the ~2% ARM regression from the previous approach (clear + full-rate xorAndPermute for partial blocks) while aligning the ARM and amd64 code paths. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
keccakF1600Sha3 now takes an optional buf *byte parameter, matching the amd64 keccakF1600BMI2 pattern. When buf != nil, XORs rate bytes into state before permuting; when nil, just permutes. The round loop and store are shared between both paths. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
4129299 to
db74886
Compare
There was a problem hiding this comment.
Pull request overview
Copilot reviewed 7 out of 7 changed files in this pull request and generated 1 comment.
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
AskAlexSharov
approved these changes
Apr 8, 2026
github-merge-queue Bot
pushed a commit
to erigontech/erigon
that referenced
this pull request
Apr 10, 2026
Update `github.com/erigontech/fastkeccak` to latest master (`08e7b660`) to pick up erigontech/fastkeccak#8
github-merge-queue Bot
pushed a commit
to erigontech/erigon
that referenced
this pull request
Apr 10, 2026
Update `github.com/erigontech/fastkeccak` to latest master (`08e7b660`) to pick up erigontech/fastkeccak#8
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
keccakF1600BMI2now accepts an optionalbuf *byteparameter. When non-nil, it XORs 17 lanes (136 bytes) into state before permuting — eliminating one full memory pass. Removes the Go-levelxorInhelper and itsunsafeusage.keccakF1600andxorAndPermuteinto a singlekeccakF1600Sha3(a, buf)function with the same optional-buf pattern as amd64. Deduplicates the 24-round loop (-143 lines).xorInfor finalization: Re-addxorInusingencoding/binary.LittleEndian.Uint64(nounsafe) for the padding/finalization path (Sum256, padAndSqueeze, sum256Sponge). Full blocks still use fused assemblyxorAndPermute. This avoids an ARM regression from always-XOR-136-bytes on partial blocks.Sumpanics once squeezing has started.useASMon BMI1 && BMI2: The assembly usesANDNQ(BMI1) in addition toRORXQ(BMI2).BenchmarkSum256_500K, simplifybenchName, clarify D-value register comments in the generator.Benchmark results (amd64)
Benchmark results (arm64)
Zero allocations maintained across all benchmarks (0 B/op, 0 allocs/op).
Test plan
go test -v ./...) on both amd64 and arm64go test -fuzz=FuzzSum256 -fuzztime=10s)benchstat(6 runs, 500ms each) on both platforms🤖 Generated with Claude Code