Tier-1 disasm of a generated-style Serialize showed PGO's guarded
devirtualization already inlines every IGenericWriter.Write body (interface
vs concrete-typed callsites measured identical), so no API or generator
changes are needed. What remained as a real call per primitive write was the
Index property setter - too large to inline due to its range-check throw path,
plus per-write high-water tracking, an AsSpan bounds check, and a
BinaryPrimitives span check.
The write path now reserves capacity once (the existing grow-on-Flush check),
then does an unaligned store through a ref with a raw index increment - the
capacity check proves the store in-bounds, and the index only moves forward
between Seeks. The _bytesWritten high-water mark (used only by
SeekOrigin.End) folds at Seek/Resize instead of per write, which is
equivalent because writes are monotonic between seeks. The Index property
keeps its validating semantics for Seek and subclasses.
BufferWriter also gains class-level implementations of the hottest
IGenericWriter default interface methods (WriteEncodedInt, DateTime,
TimeSpan, Point2D/3D, Rectangle2D/3D, Map, Race): a DIM dispatches again on
`this` for every nested Write even at a devirtualized callsite, and the class
overloads keep the whole write inlined.
Strings (arbitrary UTF-16) previously walked every string twice
(GetByteCount then GetBytes) because the variable-width length prefix
precedes the bytes. Strings of 85 chars or fewer (any content, incl.
surrogate pairs - 85 * 3 = 255 bytes max) now encode once into a 256-byte
stack scratch, then write the prefix and copy. Byte output is identical.
Measured: 34.4 -> 15.7 ns/entity on a generated-style write mix (~20 writes,
58 bytes), 22.5 -> 11.8 ns per short-string write, and the end-to-end freeze
benchmark (10M entities / 1.7GB / 24 cores, dense profile) drops from ~99ms
to ~74-82ms. Combined with the earlier pipeline commits: ~740ms -> ~78ms.
New tests pin byte-level output and position semantics: primitive
little-endian layouts, Seek(End) high-water behavior, growth preservation,
span writes across growth, encoded-int formats, and string equivalence
across the scratch/two-pass boundary with mixed-width UTF-16 content.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
f7b835d7e9