perf: Add arm64 decompression assembly

[klauspost]

Only tested on QEMU.

See perf on real HW below.

[Copilot]

Pull request overview

This PR adds ARM64 assembly implementation for decompression in the minlz package. The main goal is to provide optimized decompression on ARM64 architectures to match the existing AMD64 assembly support. The implementation includes comprehensive test coverage and updates CI configuration to test on ARM64 runners.

Key changes:

• New ARM64 assembly decoder with fast and slow loops, NEON SIMD optimizations, and special handling for overlapping copies
• Comprehensive test suite with edge cases for overlapping copies, long offsets, and various data patterns
• CI updates to include ARM64 testing and newer toolchain versions

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 18 comments.

Show a summary per file

File	Description
decode_test.go	Adds comprehensive tests for ARM64 decoder covering edge cases, overlapping copies, random data, and regression testing
decode_other.go	Updates build tags to exclude ARM64 from the Go fallback implementation
decode_arm64.go	ARM64-specific wrapper that calls the assembly implementation with race detection support
asm_arm64.go	Go stub declaration for the ARM64 assembly decoder function
asm_arm64.s	Complete ARM64 assembly implementation with ~986 lines covering all decompression tag types and copy operations
.github/workflows/release.yml	Updates Go version to 1.25.x and goreleaser to 2.13.2
.github/workflows/go.yml	Adds ARM64 runner, updates Go versions to 1.25.x, updates goreleaser, and expands fuzz test matrix

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[RaduBerinde]

This is great! I will do some testing with CockroachDB.

[RaduBerinde]

Also, I think GitHub actions support arm: https://github.com/orgs/community/discussions/148648

[klauspost]

Also, I think GitHub actions support arm: https://github.com/orgs/community/discussions/148648

Already added: https://github.com/minio/minlz/pull/29/files#diff-678682767f2477de3e3c546746f8568b9a1942b2c647d32331d7e774b8ff8d9fR16

[klauspost]

@RaduBerinde I would definitely be happy if you could verify that it is at least on par with the Go code.

Probably the easiest would be to build the mz tool.

For example using this as a testset:
7mb-cockroach-db.log.mzb.gz.

λ go build -tags=noasm&&./mz d -bench=10 -block 7mb-cockroach-db.log.mzb
Reading 7mb-cockroach-db.log.mzb...

Decompressing Block using 1 thread...
 * 568610 -> 8247494 bytes [1450.47%]; 9.907s, 5044.8MB/s

Decompressing block (16 threads)...
 * 568610 -> 8247494 bytes [6.89%]; 9.969s, 29949.9MB/s (5.9x)

λ go build -tags=&&./mz d -bench=10 -block 7mb-cockroach-db.log.mzb
Reading 7mb-cockroach-db.log.mzb...

Decompressing Block using 1 thread...
 * 568610 -> 8247494 bytes [1450.47%]; 9.883s, 8231.6MB/s

Decompressing block (16 threads)...
 * 568610 -> 8247494 bytes [6.89%]; 9.845s, 30352.4MB/s (3.7x)

(here multithreaded just hits memory bandwidth, but single thread should be fine).

[RaduBerinde]

This is on an Apple M1 laptop:

❯ go run -tags=noasm ./cmd/mz d -bench=10 -block 7mb-cockroach-db.log.mzb
Reading 7mb-cockroach-db.log.mzb...

Decompressing Block using 1 thread...
 * 568610 -> 8247494 bytes [1450.47%]; 9.885s, 4160.9MB/s

Decompressing block (10 threads)...
 * 568610 -> 8247494 bytes [6.89%]; 9.901s, 32239.8MB/s (7.7x)

❯ go run ./cmd/mz d -bench=10 -block 7mb-cockroach-db.log.mzb
Reading 7mb-cockroach-db.log.mzb...

Decompressing Block using 1 thread...
 * 568610 -> 8247494 bytes [1450.47%]; 9.878s, 4673.1MB/s

Decompressing block (10 threads)...
 * 568610 -> 8247494 bytes [6.89%]; 9.972s, 35766.4MB/s (7.7x)

I will also check on a GCE arm machine.

[klauspost]

👍🏼 Small improvement. I will have some ARM hardware available for further investigation - I can test on a wider data set.

Mostly just making sure it wasn't a regression.

[RaduBerinde]

On T2A (older):

ubuntu@radu-minlz-t2a-standard-4-0001:~/minlz$ go run -tags=noasm ./cmd/mz d -bench=10 -block 7mb-cockroach-db.log.mzb
go: downloading github.com/klauspost/compress v1.17.11
Reading 7mb-cockroach-db.log.mzb...

Decompressing Block using 1 thread...
 * 568610 -> 8247494 bytes [1450.47%]; 9.941s, 2618.3MB/s

Decompressing block (4 threads)...
 * 568610 -> 8247494 bytes [6.89%]; 9.849s, 10043.5MB/s (3.8x)
ubuntu@radu-minlz-t2a-standard-4-0001:~/minlz$ go run ./cmd/mz d -bench=10 -block 7mb-cockroach-db.log.mzb
Reading 7mb-cockroach-db.log.mzb...

Decompressing Block using 1 thread...
 * 568610 -> 8247494 bytes [1450.47%]; 9.913s, 2947.0MB/s

Decompressing block (4 threads)...
 * 568610 -> 8247494 bytes [6.89%]; 9.834s, 11338.6MB/s (3.8x)

On C4A (newer):

ubuntu@radu-minlz-c4a-0001:~/minlz$ go run -tags=noasm ./cmd/mz d -bench=10 -block 7mb-cockroach-db.log.mzb
go: downloading github.com/klauspost/compress v1.17.11
Reading 7mb-cockroach-db.log.mzb...

Decompressing Block using 1 thread...
 * 568610 -> 8247494 bytes [1450.47%]; 9.913s, 4054.3MB/s

Decompressing block (4 threads)...
 * 568610 -> 8247494 bytes [6.89%]; 9.839s, 15934.1MB/s (3.9x)
ubuntu@radu-minlz-c4a-0001:~/minlz$ go run ./cmd/mz d -bench=10 -block 7mb-cockroach-db.log.mzb
Reading 7mb-cockroach-db.log.mzb...

Decompressing Block using 1 thread...
 * 568610 -> 8247494 bytes [1450.47%]; 9.881s, 4955.4MB/s

Decompressing block (4 threads)...
 * 568610 -> 8247494 bytes [6.89%]; 9.84s, 19486.9MB/s (3.9x)

For comparison, this is what a recent x86 (C4) looks like:

ubuntu@radu-tpcc2-0001:~/minlz$ go run -tags=noasm ./cmd/mz d -bench=10 -block 7mb-cockroach-db.log.mzb
go: downloading github.com/klauspost/compress v1.17.11
Reading 7mb-cockroach-db.log.mzb...

Decompressing Block using 1 thread...
 * 568610 -> 8247494 bytes [1450.47%]; 9.929s, 3959.6MB/s

Decompressing block (32 threads)...
 * 568610 -> 8247494 bytes [6.89%]; 9.871s, 80850.6MB/s (20.4x)
ubuntu@radu-tpcc2-0001:~/minlz$ go run ./cmd/mz d -bench=10 -block 7mb-cockroach-db.log.mzb
Reading 7mb-cockroach-db.log.mzb...

Decompressing Block using 1 thread...
 * 568610 -> 8247494 bytes [1450.47%]; 9.881s, 5833.4MB/s

Decompressing block (32 threads)...
 * 568610 -> 8247494 bytes [6.89%]; 9.913s, 112799.6MB/s (19.3x)