blit 0.2.1

Zero-copy Bytes/BytesMut and high-performance random-access file I/O for Dart and Flutter, with async file I/O, native FFI mmap, and Web storage.

Changelog

0.2.1

Added

• UnsafeBytes.alloc / free: fixed-length external memory outside the Dart GC.
• Native malloc/free and Web WebAssembly.Memory arena backends.
• Raw int API: unsafeAlloc, unsafeRead*, unsafeWrite*, unsafeFree.
• AsyncFile.readAtUnsafe / writeAtUnsafe with explicit length.

0.2.0

Added

• AsyncFile, BlitFile, FileOptions, and the BlitException hierarchy.
• blit simple API: open, read, write, append, copy, exists, delete.
• NativeBuffer, NativeBufferPool, and NativeMmapFile in package:blit/blit_native.dart.
• readMany/writeMany coalesce contiguous requests; clustered writes use one read-modify-write; readMany uses concurrent backends.
• AsyncFileNative pooled read/write via NativeBuffer.
• Web openIndexedDb; OPFS falls back to IndexedDB.
• Cooperative CancelToken; exclusive create; flushWith.
• close/delete drain pending operations before closing.
• direct mmap selection and PerformanceProfile worker tuning.
• Split public entries: blit.dart, blit_io.dart, blit_buffer.dart, blit_web.dart, blit_native.dart.

0.1.7

Added

• AsyncIO utilities in package:blit/blit_io.dart: ReadBuf, IoSlice, AsyncRead, AsyncWrite, AsyncBufRead, AsyncReadExt, AsyncWriteExt.
• BufReader, BufWriter, readLine, readerStream, copy.
• Uint8ListReader, BytesAsyncReader, Uint8ListWriter, BytesMutAsyncWriter, AsyncFileReader, AsyncFileWriter.

0.1.6

Added

• IndexedDB backend: openIdbBytesFile.
• Read-only Blob backend: openBlobBytesFile.

0.1.5

Added

• ReadFast.
• benchmark/rust with iocp / uring features.
• file.read.pointer8MiB / file.write.pointer8MiB benchmark rows.

Changed

• Rust benchmark replaces the C benchmark in CI.
• Added Linux UringBytesFile benchmarks.
• Benchmark chart uses benchcat and bundled fonts.

0.1.4

Changed

• Benchmark chart rendering moved to benchcat (benchcat.toml); CI uses unk-404/benchcat@master instead of benchmark_svg.

Fixed

• Shared io_uring loops are now explicitly closable through UringBytesFile.closeSharedLoopPool; the Linux test suite closes its pool in tearDownAll. This prevents an idle io_uring-reaper isolate from keeping dart test / CI alive after the shared-loop tests finish.

0.1.3

Changed

• Requires io_uring ^0.1.1, which provides UringLoopPool / UringLoopLease.

Fixed

• BytesMut.splitTo / splitOff now mark the source frozen, so the returned views survive later BytesMut mutations (copy-on-write), matching the README.

Performance

• Added BytesMut.spareView + advance and unchecked put variants for uint8/uint16/uint32/uint64/float64/bytes/slice/repeated/uint; Bytes gained unchecked uint64/float32/float64 reads.
• Windows IocpFile reuses OVERLAPPED blocks and its single-request scratch buffer instead of allocating per request.
• Linux UringBytesFile registers a small fixed-buffer set and uses READ_FIXED / WRITE_FIXED, falling back to normal ops when unsupported.
• Added openUringBytesFile(..., sharedLoop: true) using the io_uring 0.1.1 UringLoopPool / UringLoopLease; fixed buffers are unregistered before a lease is returned. Leases are exclusive, so this reuses reaper isolates for sequential open/close workloads rather than multiplexing concurrent files.
• Added a core Bytes/BytesMut/reader/writer benchmark suite, including a raw Uint8List baseline, merged into the main benchmark card.

Added

• MmapFile.readAtPointer / writeAtPointer for caller-owned native buffers; these bypass the typed-data scratch copy (and the transparent mmap cache).

0.1.2

Performance

• Buffered MmapFile.readAt / writeAt now use a transparent whole-file mmap cache by default; mmapCache: false keeps pool workers and direct-I/O files on the syscall path. File size and FFI argument storage are reused instead of queried/allocated per call.
• Added an isolate-local PooledNativeAllocator for uninitialized, reusable scratch blocks.
• UringBytesFile reuses UringBuffers across single and batched operations.

0.1.1

Added

• TypedData interop: Bytes.fromTypedData, TypedData.asBytes(), Bytes.fromBytesLike, and the BytesLike boundary interface; Bytes and BytesMut implement BytesLike without copying.
• MmapFile.punchHole (Windows FSCTL_SET_ZERO_DATA, Linux fallocate(PUNCH_HOLE)), Windows cloneRangeFrom (FSCTL_DUPLICATE_EXTENTS_TO_FILE), macOS preallocate (F_PREALLOCATE) / adviseRead (F_RDADVISE), and cloneFileSync (clonefile on macOS/iOS, copy fallback elsewhere).
• Linux UringBytesFile / openUringBytesFile: pure-Dart package:io_uring SQ/CQ rings, batched readMany/writeMany, no liburing or C shim.
• Windows MmapPlaceholder via VirtualAlloc2 + MapViewOfFile3 (MEM_RESERVE_PLACEHOLDER / MEM_REPLACE_PLACEHOLDER), for gap-free window replacement on Windows 10 1803+.
• Windows direct-I/O largePages: true opt-in (MEM_LARGE_PAGES scratch, privilege-checked with normal-page fallback).
• ParallelBytesFile.maxQueue backpressure for queued + in-flight jobs.
• Web OPFS worker-per-chunk pool: OpfsWorkerPool routes chunk i to workers[i % N] (default 4), keeps the manifest single-writer, and caches each chunk handle in its owning worker.
• Optional Web SAB ring for small read/write calls when crossOriginIsolated and Atomics.waitAsync are available: 8 闂?64 KiB slots, Atomics wakeups, postMessage fallback for larger payloads and non-isolated pages.
• Benchmark SVG rendering now uses bench-svg (bench-svg.toml + the unk-404/benchmark_svg action) instead of a repo-local SVG generator.
• Benchmark chart compares Dart with a C counterpart implementing the same single-thread 8 MiB sequential and 4 KiB random/mmap operations; Dart-only engine rows use bench-svg placeholder tracks.

0.1.0

First development release.

Added

• Zero-copy Bytes views (slice/sublist O(1)) and BytesMut with freeze/splitTo/splitOff copy-on-write semantics.
• Exact Uint64 (hi/lo limbs) for VM, dart2js and dart2wasm, including arithmetic, bitwise, shifts, parse and toRadixString.
• BytesReader/BytesWriter fixed-width cursors with BE/LE support.
• BytesFile/SyncBytesFile and the native dart:io backend.
• Native mmap engine: Windows CreateFileMappingW, POSIX mmap/pread; sparse files, zero-copy windows, zeroize, read-only mappings.
• Windows IocpFile: overlapped I/O + IOCP, QD 1闂?28 batch reads/writes.
• MappedRegion.prefetch() / discard(): Windows PrefetchVirtualMemory (discard is a no-op for file-backed maps), POSIX madvise(MADV_WILLNEED / MADV_DONTNEED).
• Bytes.getUint8/16/32Unchecked and BytesReader.readUint8/16/32/64Unchecked hot-path methods (caller checks bounds once).
• Direct I/O: openMmapFileSync(..., direct: true) / ParallelBytesFile.open(..., direct: true) via Windows FILE_FLAG_NO_BUFFERING + VirtualAlloc, Linux/Android O_DIRECT, and macOS/iOS F_NOCACHE; aligned scratch buffers and 4 KiB-aligned streaming reads/writes.
• ParallelBytesFile: isolate worker pool (default up to 32), automatic FIFO dispatch, batch 32, readMany/writeMany, watermark barriers, adjacent-range coalescing, and a read-only full-file mmap cache for hot files (default limit 1 GiB).
• Web OPFS backend: chunk files, double-slot manifest, CRC-32, dedicated worker, capability/quota probing.
• Benchmark suite (benchmark_test, JIT/AOT + JSONL), a generated assets/benchmarks.svg, and Dart/Web CI workflows.
• Tests: differential fuzz, lifetime/error paths, 100k-op stress, 32 GiB sparse.

Fixed during development

• putUint64 little-endian limb encoding and atomic failure on short buffers.
• ParallelBytesFile barrier deadlock when queued work already existed.
• Windows read-only mapping zeroize guard and empty-file map consistency.