protobuf-go/proto/encode.go
Damien Neil c37adefdac internal/impl: add fast-path marshal implementation
This is a port of the v1 table marshaler, with some substantial
cleanup and refactoring.

Benchstat results from the protobuf reference benchmark data comparing the
v1 package with v2, with AllowPartial:true set for the new package. This
is not an apples-to-apples comparison, since v1 doesn't have a way to
disable required field checks.  Required field checks in v2 package
currently go through reflection, which performs terribly; my initial
experimentation indicates that fast-path required field checks will
not add a large amount of cost; these results are incomplete but not
wholly inaccurate.

name                                           old time/op  new time/op  delta
/dataset.google_message3_1.pb/Marshal-12        219ms ± 1%   232ms ± 1%   +5.85%  (p=0.004 n=6+5)
/dataset.google_message2.pb/Marshal-12          261µs ± 3%   248µs ± 1%   -5.14%  (p=0.002 n=6+6)
/dataset.google_message1_proto2.pb/Marshal-12   681ns ± 2%   637ns ± 3%   -6.53%  (p=0.002 n=6+6)
/dataset.google_message1_proto3.pb/Marshal-12  1.10µs ± 8%  0.99µs ± 3%   -9.63%  (p=0.002 n=6+6)
/dataset.google_message3_3.pb/Marshal-12       44.2ms ± 3%  35.2ms ± 1%  -20.28%  (p=0.004 n=6+5)
/dataset.google_message4.pb/Marshal-12         91.4ms ± 2%  94.9ms ± 2%   +3.78%  (p=0.002 n=6+6)
/dataset.google_message3_2.pb/Marshal-12       78.7ms ± 6%  80.8ms ± 4%     ~     (p=0.310 n=6+6)
/dataset.google_message3_4.pb/Marshal-12       10.6ms ± 3%  10.6ms ± 8%     ~     (p=0.662 n=5+6)
/dataset.google_message3_5.pb/Marshal-12        675ms ± 4%   510ms ± 2%  -24.40%  (p=0.002 n=6+6)
/dataset.google_message3_1.pb/Marshal           219ms ± 1%   236ms ± 7%   +8.06%  (p=0.004 n=5+6)
/dataset.google_message2.pb/Marshal             257µs ± 1%   250µs ± 3%     ~     (p=0.052 n=5+6)
/dataset.google_message1_proto2.pb/Marshal      685ns ± 1%   628ns ± 1%   -8.41%  (p=0.008 n=5+5)
/dataset.google_message1_proto3.pb/Marshal     1.08µs ± 1%  0.98µs ± 2%   -9.31%  (p=0.004 n=5+6)
/dataset.google_message3_3.pb/Marshal          43.7ms ± 1%  35.1ms ± 1%  -19.76%  (p=0.002 n=6+6)
/dataset.google_message4.pb/Marshal            93.4ms ± 4%  94.9ms ± 2%     ~     (p=0.180 n=6+6)
/dataset.google_message3_2.pb/Marshal           105ms ± 2%    98ms ± 7%   -6.81%  (p=0.009 n=5+6)
/dataset.google_message3_4.pb/Marshal          16.3ms ± 6%  15.7ms ± 3%   -3.44%  (p=0.041 n=6+6)
/dataset.google_message3_5.pb/Marshal           676ms ± 4%   504ms ± 2%  -25.50%  (p=0.004 n=6+5)

Change-Id: I72cc4597117f4cf5d236ef505777d49dd4a5f75d
Reviewed-on: https://go-review.googlesource.com/c/protobuf/+/171020
Reviewed-by: Joe Tsai <thebrokentoaster@gmail.com>
2019-05-16 22:13:43 +00:00

287 lines
8.8 KiB
Go

// Copyright 2019 The Go Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
package proto
import (
"fmt"
"sort"
"google.golang.org/protobuf/internal/encoding/wire"
"google.golang.org/protobuf/internal/errors"
"google.golang.org/protobuf/internal/mapsort"
"google.golang.org/protobuf/internal/pragma"
"google.golang.org/protobuf/reflect/protoreflect"
"google.golang.org/protobuf/runtime/protoiface"
)
// MarshalOptions configures the marshaler.
//
// Example usage:
// b, err := MarshalOptions{Deterministic: true}.Marshal(m)
type MarshalOptions struct {
// AllowPartial allows messages that have missing required fields to marshal
// without returning an error. If AllowPartial is false (the default),
// Marshal will return an error if there are any missing required fields.
AllowPartial bool
// Deterministic controls whether the same message will always be
// serialized to the same bytes within the same binary.
//
// Setting this option guarantees that repeated serialization of
// the same message will return the same bytes, and that different
// processes of the same binary (which may be executing on different
// machines) will serialize equal messages to the same bytes.
//
// Note that the deterministic serialization is NOT canonical across
// languages. It is not guaranteed to remain stable over time. It is
// unstable across different builds with schema changes due to unknown
// fields. Users who need canonical serialization (e.g., persistent
// storage in a canonical form, fingerprinting, etc.) must define
// their own canonicalization specification and implement their own
// serializer rather than relying on this API.
//
// If deterministic serialization is requested, map entries will be
// sorted by keys in lexographical order. This is an implementation
// detail and subject to change.
Deterministic bool
// UseCachedSize indicates that the result of a previous Size call
// may be reused.
//
// Setting this option asserts that:
//
// 1. Size has previously been called on this message with identical
// options (except for UseCachedSize itself).
//
// 2. The message and all its submessages have not changed in any
// way since the Size call.
//
// If either of these invariants is broken, the results are undefined
// but may include panics or invalid output.
//
// Implementations MAY take this option into account to provide
// better performance, but there is no guarantee that they will do so.
// There is absolutely no guarantee that Size followed by Marshal with
// UseCachedSize set will perform equivalently to Marshal alone.
UseCachedSize bool
pragma.NoUnkeyedLiterals
}
var _ = protoiface.MarshalOptions(MarshalOptions{})
// Marshal returns the wire-format encoding of m.
func Marshal(m Message) ([]byte, error) {
return MarshalOptions{}.MarshalAppend(nil, m)
}
// Marshal returns the wire-format encoding of m.
func (o MarshalOptions) Marshal(m Message) ([]byte, error) {
return o.MarshalAppend(nil, m)
}
// MarshalAppend appends the wire-format encoding of m to b,
// returning the result.
func (o MarshalOptions) MarshalAppend(b []byte, m Message) ([]byte, error) {
// Set AllowPartial in recursive calls to marshal to avoid duplicating
// effort with the single initialization check below.
allowPartial := o.AllowPartial
o.AllowPartial = true
out, err := o.marshalMessageFast(b, m)
if err == errInternalNoFast {
out, err = o.marshalMessage(b, m.ProtoReflect())
}
var nerr errors.NonFatal
if !nerr.Merge(err) {
return out, err
}
if !allowPartial {
nerr.Merge(IsInitialized(m))
}
return out, nerr.E
}
func (o MarshalOptions) marshalMessageFast(b []byte, m Message) ([]byte, error) {
methods := protoMethods(m)
if methods == nil ||
methods.MarshalAppend == nil ||
(o.Deterministic && methods.Flags&protoiface.MethodFlagDeterministicMarshal == 0) {
return nil, errInternalNoFast
}
if methods.Size != nil {
sz := methods.Size(m)
if cap(b) < len(b)+sz {
x := make([]byte, len(b), len(b)+sz)
copy(x, b)
b = x
}
o.UseCachedSize = true
}
return methods.MarshalAppend(b, m, protoiface.MarshalOptions(o))
}
func (o MarshalOptions) marshalMessage(b []byte, m protoreflect.Message) ([]byte, error) {
// There are many choices for what order we visit fields in. The default one here
// is chosen for reasonable efficiency and simplicity given the protoreflect API.
// It is not deterministic, since KnownFields.Range does not return fields in any
// defined order.
//
// When using deterministic serialization, we sort the known fields by field number.
fieldDescs := m.Descriptor().Fields()
knownFields := m.KnownFields()
var err error
var nerr errors.NonFatal
o.rangeKnown(knownFields, func(num protoreflect.FieldNumber, value protoreflect.Value) bool {
field := fieldDescs.ByNumber(num)
if field == nil {
field = knownFields.ExtensionTypes().ByNumber(num).Descriptor()
if field == nil {
panic(fmt.Errorf("no descriptor for field %d in %q", num, m.Descriptor().FullName()))
}
}
b, err = o.marshalField(b, field, value)
if nerr.Merge(err) {
err = nil
return true
}
return false
})
if err != nil {
return b, err
}
m.UnknownFields().Range(func(_ protoreflect.FieldNumber, raw protoreflect.RawFields) bool {
b = append(b, raw...)
return true
})
return b, nerr.E
}
// rangeKnown visits known fields in field number order when deterministic
// serialization is enabled.
func (o MarshalOptions) rangeKnown(knownFields protoreflect.KnownFields, f func(protoreflect.FieldNumber, protoreflect.Value) bool) {
if !o.Deterministic {
knownFields.Range(f)
return
}
nums := make([]protoreflect.FieldNumber, 0, knownFields.Len())
knownFields.Range(func(num protoreflect.FieldNumber, _ protoreflect.Value) bool {
nums = append(nums, num)
return true
})
sort.Slice(nums, func(a, b int) bool {
return nums[a] < nums[b]
})
for _, num := range nums {
if !f(num, knownFields.Get(num)) {
break
}
}
}
func (o MarshalOptions) marshalField(b []byte, fd protoreflect.FieldDescriptor, value protoreflect.Value) ([]byte, error) {
num := fd.Number()
kind := fd.Kind()
switch {
case fd.IsList():
return o.marshalList(b, num, fd, value.List())
case fd.IsMap():
return o.marshalMap(b, num, fd, value.Map())
default:
b = wire.AppendTag(b, num, wireTypes[kind])
return o.marshalSingular(b, num, fd, value)
}
}
func (o MarshalOptions) marshalList(b []byte, num wire.Number, fd protoreflect.FieldDescriptor, list protoreflect.List) ([]byte, error) {
if fd.IsPacked() {
b = wire.AppendTag(b, num, wire.BytesType)
b, pos := appendSpeculativeLength(b)
var nerr errors.NonFatal
for i, llen := 0, list.Len(); i < llen; i++ {
var err error
b, err = o.marshalSingular(b, num, fd, list.Get(i))
if !nerr.Merge(err) {
return b, err
}
}
b = finishSpeculativeLength(b, pos)
return b, nerr.E
}
kind := fd.Kind()
var nerr errors.NonFatal
for i, llen := 0, list.Len(); i < llen; i++ {
var err error
b = wire.AppendTag(b, num, wireTypes[kind])
b, err = o.marshalSingular(b, num, fd, list.Get(i))
if !nerr.Merge(err) {
return b, err
}
}
return b, nerr.E
}
func (o MarshalOptions) marshalMap(b []byte, num wire.Number, fd protoreflect.FieldDescriptor, mapv protoreflect.Map) ([]byte, error) {
keyf := fd.MapKey()
valf := fd.MapValue()
var nerr errors.NonFatal
var err error
o.rangeMap(mapv, keyf.Kind(), func(key protoreflect.MapKey, value protoreflect.Value) bool {
b = wire.AppendTag(b, num, wire.BytesType)
var pos int
b, pos = appendSpeculativeLength(b)
b, err = o.marshalField(b, keyf, key.Value())
if !nerr.Merge(err) {
return false
}
b, err = o.marshalField(b, valf, value)
if !nerr.Merge(err) {
return false
}
err = nil
b = finishSpeculativeLength(b, pos)
return true
})
if err != nil {
return b, err
}
return b, nerr.E
}
func (o MarshalOptions) rangeMap(mapv protoreflect.Map, kind protoreflect.Kind, f func(protoreflect.MapKey, protoreflect.Value) bool) {
if !o.Deterministic {
mapv.Range(f)
return
}
mapsort.Range(mapv, kind, f)
}
// When encoding length-prefixed fields, we speculatively set aside some number of bytes
// for the length, encode the data, and then encode the length (shifting the data if necessary
// to make room).
const speculativeLength = 1
func appendSpeculativeLength(b []byte) ([]byte, int) {
pos := len(b)
b = append(b, "\x00\x00\x00\x00"[:speculativeLength]...)
return b, pos
}
func finishSpeculativeLength(b []byte, pos int) []byte {
mlen := len(b) - pos - speculativeLength
msiz := wire.SizeVarint(uint64(mlen))
if msiz != speculativeLength {
for i := 0; i < msiz-speculativeLength; i++ {
b = append(b, 0)
}
copy(b[pos+msiz:], b[pos+speculativeLength:])
b = b[:pos+msiz+mlen]
}
wire.AppendVarint(b[:pos], uint64(mlen))
return b
}