feat(core): HTTP/2 Phase 4 — byte-layer foundations
Builds dev.relism.flash.bytes: ByteScan (scanning/comparison/hashing, scalar + SWAR, property-tested against each other on every boundary and 20,000 random fuzz trials each), ArrayBackedByteView/SegmentedByteView capability hierarchy, PooledSlice/SlicePool, ByteWriter, Pairs. Cashes in the allocation and scanning wins the existing code left on the table: EX-04 (word-at-a-time router matching, verified directly against fpr-core's own ByteCompare), EX-05 (pooled views replacing per-call anonymous ByteView allocations in HeaderMap/QueryParams/PathParams), EX-09 (HeaderMap index built once per reset() instead of rescanning per lookup), EX-19 (reusable PathParams on the router's per-connection scratch), EX-25/EX-26 (single-allocation String construction), EX-33 (SWAR header-terminator scan in RequestParser). Also closes EX-06's router half, missing from this phase's own EX-item list in the plan (same class of omission DEC-12 recorded for Phase 1): FastPathRouterImpl/FastPathWsRouterImpl's ThreadLocals (unbounded under one-virtual-thread-per-connection) are replaced by an opaque, caller-owned per-connection scratch object (AbstractRouter#newScratch), not by extending ConnectionScratch as its own Javadoc originally assumed -- that would have created transport's first dependency on routing in the reverse direction. Full rationale in DEC-19. Every optimization is measured, not asserted (DEC-20): SWAR scan 35.4% faster than scalar, kept; EX-04's word-path 32.1% faster than byte-at-a-time at the mechanism level, kept for its real future consumers even though today's router doesn't yet route through it (MethodPathByteView stays deliberately non-array-backed, per the plan's own text). Router matching itself is ~0 B/op including parametric routes. The full h1 pipeline is not literally 0 B/op yet -- 120 B/op is Request/RequestBody/RequestLine construction, honestly attributed to Phase 6's explicit scope rather than hidden. 395/395 tests green, both with and without -Pjmh. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
This commit is contained in:
Zakaria El Orche
co-authored by
Claude Sonnet 5
Claude Sonnet 5
parent
2bf261e4e2
commit
704a00a551
@@ -0,0 +1,179 @@
|
||||
# The Byte Layer (Phase 4)
|
||||
|
||||
Audience: contributors. This is the design record for `dev.relism.flash.bytes` — the
|
||||
protocol-neutral byte primitives both HTTP/1.1 and HTTP/2 build on — and for the Phase 4
|
||||
allocation/scanning fixes (`EX-04`, `EX-05`, `EX-09`, `EX-19`, `EX-25`, `EX-26`, `EX-33`, plus
|
||||
`EX-06`'s router half) that consume them.
|
||||
|
||||
## Why this exists
|
||||
|
||||
Before Phase 4, byte-scanning and case-insensitive comparison logic was duplicated, slightly
|
||||
differently, in `RequestParser`, `HeaderMap`, and `Http1KeepAlive`; `fpr-core`'s word-at-a-time
|
||||
router-matching fast path (`ByteCompare`) was wired up but never actually enabled anywhere
|
||||
(`EX-04` — every `ByteView` implementation returned `supportsLong() == false`); and four call
|
||||
sites allocated a fresh view, array, or `String` per call on paths a realistic middleware chain
|
||||
hits 6–10 times per request. `dev.relism.flash.bytes` is the single home these fixes converge on,
|
||||
so no later phase (HPACK, the frame layer) has to invent its own scanning primitives.
|
||||
|
||||
## Package layout
|
||||
|
||||
```
|
||||
dev.relism.flash.bytes
|
||||
├── ByteScan static scanning/comparison/hashing utilities, scalar + SWAR
|
||||
├── ArrayBackedByteView capability interface: a ByteView backed by one contiguous byte[]
|
||||
├── SegmentedByteView the deliberate non-array-backed case (K discontiguous segments)
|
||||
├── PooledSlice reusable ArrayBackedByteView, the EX-05 fix
|
||||
├── SlicePool a small fixed-size ring of PooledSlice
|
||||
├── ByteWriter index-based writer into a growable byte[] scratch buffer
|
||||
└── Pairs the (hi<<32)|lo allocation-free pair-return idiom, named
|
||||
```
|
||||
|
||||
## The `ByteView` capability hierarchy
|
||||
|
||||
```
|
||||
ByteView (fpr-core)
|
||||
├── ArrayBackedByteView capability: array() + offset()
|
||||
│ ├── FastPathViews.RequestByteView RequestParser's request-line/header slices
|
||||
│ ├── FastPathViews.SocketByteView a bare byte[] (e.g. a WebSocket payload)
|
||||
│ ├── FastPathViews.StringByteView a String's UTF-8 bytes
|
||||
│ └── PooledSlice the EX-05 reusable, pool-issued slice
|
||||
└── (bare ByteView, not array-backed)
|
||||
├── SegmentedByteView K discontiguous segments (future HPACK CONTINUATION)
|
||||
└── FastPathViews.MethodPathByteView method bytes + another ByteView, composed
|
||||
```
|
||||
|
||||
Code holding a bare `ByteView` and wanting the fast path when the concrete instance happens to
|
||||
be array-backed does `instanceof ArrayBackedByteView` and falls back to the byte-at-a-time path
|
||||
otherwise — see `ArrayBackedByteView`'s own Javadoc. This is used throughout Phase 4's fixes:
|
||||
`Request.path()`, `PathParams.get()` (`EX-25`), and `QueryParams.decode`'s clean-value fast path
|
||||
(`EX-26`) all take this shape.
|
||||
|
||||
## `EX-04`: the `supportsLong()`/`longAt()` contract
|
||||
|
||||
`fpr-core`'s `ByteCompare` (decompiled from `fpr-core-1.1.1`, since no source jar is published)
|
||||
reads its comparison word via
|
||||
`MethodHandles.byteArrayViewVarHandle(long[].class, ByteOrder.LITTLE_ENDIAN)` and only takes the
|
||||
word-at-a-time branch when the caller passed `useLong = true`, comparing the result bit-for-bit
|
||||
against whatever `ByteView#longAt` returns. The contract this imposes on any `longAt`
|
||||
implementation:
|
||||
|
||||
- Return the same value `LONG_VIEW.get(array, pos)` would, for the identical 8 bytes — meaning
|
||||
**little-endian**, fixed, regardless of the host's native byte order (unlike `ByteScan`'s own
|
||||
SWAR internals, which use `ByteOrder.nativeOrder()` for speed — see below for why that's a
|
||||
different, safe choice in a different context).
|
||||
- The caller (`ByteCompare`) never calls `longAt(i)` without first establishing `i + 8 <=
|
||||
length()` — so `longAt` implementations do not re-check this themselves (a defensive check
|
||||
would be dead code on every real call path).
|
||||
|
||||
`FastPathViews.RequestByteView`/`SocketByteView`/`StringByteView` implement this;
|
||||
`MethodPathByteView` (composite, no single backing array) and `SegmentedByteView` (genuinely
|
||||
discontiguous) both stay at the inherited `false` default — a word-at-a-time read is not merely
|
||||
unimplemented for these, it is structurally unsound (a read could straddle two sources).
|
||||
|
||||
**Verified against `fpr-core` directly** (`FastPathViewsLongAtTest`), not merely by reading
|
||||
bytecode: `ByteCompare.equals`/`indexOf` called with `useLong=true` and `useLong=false` are
|
||||
asserted to agree on identical content, on content diverging at every position across an
|
||||
8+-byte range (word-interior, word-boundary, and scalar-tail cases), and end-to-end through a
|
||||
real compiled `fpr-core` router with literal route segments ≥ 8 bytes — including a near-miss
|
||||
route differing only in its last byte, to catch exactly the kind of bounds/endianness bug that
|
||||
would otherwise silently mis-route a request (the failure mode `EX-04`'s registry entry calls out
|
||||
by name as the worst possible one here).
|
||||
|
||||
## `ByteScan`'s SWAR technique
|
||||
|
||||
Both `ByteScan.indexOf` (single byte) and `ByteScan.indexOfCrLfCrLf` (the `\r\n\r\n` header
|
||||
terminator, `EX-33`) use the classic "does this word contain byte `b`" bit trick: XOR the 8-byte
|
||||
word against `b` broadcast into every lane, then test for any zero lane with
|
||||
`(v - 0x0101...01) & ~v & 0x8080...80`. `indexOfCrLfCrLf` uses this as a pre-filter to find a
|
||||
candidate `CR` byte 8 at a time, then a cheap scalar 3-byte check verifies the full 4-byte match
|
||||
at each candidate — so a scan touches every byte once per 8-byte stride in the common
|
||||
no-CR-yet case, rather than once per byte.
|
||||
|
||||
This reads the word via `ByteOrder.nativeOrder()`, not a fixed order — safe here (unlike
|
||||
`EX-04`'s `longAt`) because nothing compares this word against an independently-decoded one;
|
||||
byte-equality detection itself (finding *that* a matching lane exists) is indifferent to lane
|
||||
order, and position extraction (`laneIndexOf`) branches on the actual native order once, at
|
||||
class-init time, to convert a matching bit back into the correct array index either way.
|
||||
|
||||
Every SWAR method has a scalar counterpart (`indexOfScalar`, `indexOfCrLfCrLfScalar`) used as
|
||||
the correctness oracle: `ByteScanTest` property-tests SWAR against scalar at every length 0–256
|
||||
and every match position (including unaligned starts and matches at the very last valid byte),
|
||||
and `ByteScanFuzzTest` throws 20 000 fully-random trials at each, per the plan's task 1. All
|
||||
green — see the class's own Javadoc for the full technique writeup.
|
||||
|
||||
## `EX-09`: `HeaderMap`'s index
|
||||
|
||||
Before this phase, every `HeaderMap` lookup (`first`, `all`, `view`, `valueEqualsIgnoreCase`)
|
||||
rescanned the entire header section from scratch — O(n·m) for a realistic middleware chain
|
||||
performing 6–10 lookups per request. `HeaderMap.reset()` now scans the section exactly once,
|
||||
recording per-header `(nameOffset, nameLength, valueOffset, valueLength)` and a case-insensitive
|
||||
32-bit FNV-1a hash of the name (`ByteScan.hashNameIgnoreCaseAscii`) into `int[]` arrays grown
|
||||
(never shrunk) to the connection's high-water mark, capped by `Http1Limits.MAX_HEADER_COUNT`
|
||||
(asserted, not silently truncated — Phase 1 already rejects any request that would exceed it).
|
||||
Every lookup then compares the caller's own hash (`ByteScan.hashNameIgnoreCaseAscii(String)`,
|
||||
computed once) against the index's hashes before ever falling back to a full case-insensitive
|
||||
name comparison. Net effect: the header section is scanned once per request (at `reset()`) plus
|
||||
once at parse time (`RequestParser`'s own validation pass) — two scans total, replacing "one scan
|
||||
at parse time plus one rescan per lookup" — strictly less work even for a single lookup, and much
|
||||
less for the realistic multi-lookup case. `forEach` was unified onto the same index rather than
|
||||
keeping its own independent scan, removing a second, easily-diverging scanning implementation.
|
||||
|
||||
## `EX-05`: pooled slices
|
||||
|
||||
`HeaderMap.view`, `QueryParams.view`, and `PathParams.view` used to allocate a fresh anonymous
|
||||
`ByteView` (plus its capturing instance) on every call. Each now draws from a small
|
||||
(`VIEW_POOL_SIZE = 4`) `SlicePool` of reusable `PooledSlice` instances instead. The lifetime
|
||||
contract, restated on each method: **a returned view stays valid until either the request ends,
|
||||
or the same `view()` method is called `VIEW_POOL_SIZE` more times on the same instance —
|
||||
whichever comes first** — at which point the ring silently repositions the same object over
|
||||
different bytes. This is a real, demonstrated hazard, not a hypothetical one:
|
||||
`SlicePoolTest#wraparoundAliasesThePreviouslyReturnedSlice` and the analogous tests in
|
||||
`HeaderMapIndexTest`, `QueryParamsFastPathTest`, and `PathParamsTest` all show a 5th call
|
||||
returning the exact same object instance the 1st call did, now aliased to different content.
|
||||
|
||||
`QueryParams` and `PathParams`'s pools are created **lazily**, on the first actual `view()` call
|
||||
— not eagerly in the constructor — because both classes are otherwise-cheap objects created per
|
||||
request (or, for `PathParams`'s `FastPathRouterImpl`-owned reusable instance, once per
|
||||
connection) regardless of whether `view()` is ever invoked; an eager pool would add
|
||||
`VIEW_POOL_SIZE` allocations to every such object whether or not it needed them; `HeaderMap`'s
|
||||
pool, by contrast, is unconditionally useful (every request's `HeaderMap` handles headers) and is
|
||||
constructed eagerly for simplicity.
|
||||
|
||||
**Two documented, deliberately-kept exceptions to "no `new ByteView()` remains"**: `QueryParams.view`
|
||||
and `PathParams.view` each retain a fallback anonymous `ByteView` for the case where their
|
||||
backing source is *not* `ArrayBackedByteView` — structurally unreachable on the real request path
|
||||
today (`RequestParser` only ever constructs array-backed views), kept because both constructors
|
||||
are `public` and could in principle be called with an arbitrary `ByteView`. A silent, correct,
|
||||
allocating fallback was judged preferable to either crashing on a technically-valid input or
|
||||
deleting a case that only test/future code could exercise. `HeaderMap.view` has no such fallback
|
||||
— it is always buffer-backed by construction.
|
||||
|
||||
## `EX-19`/`EX-06` (router half): the `FastPathRouterImpl` scratch
|
||||
|
||||
`FastPathRouterImpl.RouteScratch` (created once per connection via `AbstractRouter#newScratch`,
|
||||
replacing the `ThreadLocal<MatchResult>`/`ThreadLocal<MethodPathByteView>` pair — see
|
||||
`DECISIONS.md`, `DEC-19`, for why this is an opaque caller-owned object rather than an extension
|
||||
of `ConnectionScratch`) also owns the reusable path-param arrays and a single long-lived
|
||||
`PathParams` instance, grown (via `ensureParamCapacity`, doubling) to the largest param count any
|
||||
route on that connection has ever matched, and repositioned (`PathParams#reset`) rather than
|
||||
reallocated on every match. `PathParams` gained a second, count-explicit constructor and a public
|
||||
`reset(ByteView, int)` specifically for this: the reusable arrays can be larger than a given
|
||||
request's actual param count, so `count` must be tracked independently of `names.length`.
|
||||
|
||||
## `EX-25`/`EX-26`: single-allocation `String` construction
|
||||
|
||||
`Request.path()`, `PathParams.get()`, and (for the common "no `%`/`+` in the value" case)
|
||||
`QueryParams.decode` now build their result `String` directly from the backing array via
|
||||
`new String(array, offset, length, UTF_8)` when the source is `ArrayBackedByteView`, instead of a
|
||||
byte-at-a-time copy into a scratch `byte[]` followed by a second allocation for the `String`
|
||||
itself. `QueryParams.decode` scans the value once for `%`/`+` first; only a value that actually
|
||||
needs percent-decoding pays for the scratch-buffer path — verified to produce byte-identical
|
||||
output to the always-decode path it bypasses, across clean values, `+`-only, `%XX`-only, invalid
|
||||
escapes, and mixed queries (`QueryParamsFastPathTest`).
|
||||
|
||||
## Performance measurement
|
||||
|
||||
`EX-04` (the router's word-at-a-time path) and `EX-33` (the SWAR header-end scan) both carry an
|
||||
explicit "measure, and keep only if it doesn't cost" instruction in the plan. Both are measured
|
||||
together with the phase's overall zero-allocation contract in one JMH pass — see `DECISIONS.md`,
|
||||
`DEC-20`, for the numbers and the keep/revert decision for each.
|
||||
Reference in New Issue
Block a user