protobuf-go/internal/impl/codec_extension.go

216 lines
5.8 KiB
Go
Raw Normal View History

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>
2019-04-01 20:49:56 +00:00
// 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 impl
import (
"sync"
"sync/atomic"
"google.golang.org/protobuf/encoding/protowire"
"google.golang.org/protobuf/internal/errors"
"google.golang.org/protobuf/reflect/protoreflect"
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>
2019-04-01 20:49:56 +00:00
)
type extensionFieldInfo struct {
internal/impl: add fast-path unmarshal Benchmarks run with: go test ./benchmarks/ -bench=Wire -benchtime=500ms -benchmem -count=8 Fast-path vs. parent commit: name old time/op new time/op delta Wire/Unmarshal/google_message1_proto2-12 1.35µs ± 2% 0.45µs ± 4% -67.01% (p=0.000 n=8+8) Wire/Unmarshal/google_message1_proto3-12 1.07µs ± 1% 0.31µs ± 1% -71.04% (p=0.000 n=8+8) Wire/Unmarshal/google_message2-12 691µs ± 2% 188µs ± 2% -72.78% (p=0.000 n=7+8) name old allocs/op new allocs/op delta Wire/Unmarshal/google_message1_proto2-12 60.0 ± 0% 25.0 ± 0% -58.33% (p=0.000 n=8+8) Wire/Unmarshal/google_message1_proto3-12 42.0 ± 0% 7.0 ± 0% -83.33% (p=0.000 n=8+8) Wire/Unmarshal/google_message2-12 28.6k ± 0% 8.5k ± 0% -70.34% (p=0.000 n=8+8) Fast-path vs. -v1: name old time/op new time/op delta Wire/Unmarshal/google_message1_proto2-12 702ns ± 1% 445ns ± 4% -36.58% (p=0.000 n=8+8) Wire/Unmarshal/google_message1_proto3-12 604ns ± 1% 311ns ± 1% -48.54% (p=0.000 n=8+8) Wire/Unmarshal/google_message2-12 179µs ± 3% 188µs ± 2% +5.30% (p=0.000 n=7+8) name old allocs/op new allocs/op delta Wire/Unmarshal/google_message1_proto2-12 26.0 ± 0% 25.0 ± 0% -3.85% (p=0.000 n=8+8) Wire/Unmarshal/google_message1_proto3-12 8.00 ± 0% 7.00 ± 0% -12.50% (p=0.000 n=8+8) Wire/Unmarshal/google_message2-12 8.49k ± 0% 8.49k ± 0% -0.01% (p=0.000 n=8+8) Change-Id: I6247ac3fd66a63d9acb902cbd192094ee3d151c3 Reviewed-on: https://go-review.googlesource.com/c/protobuf/+/185147 Reviewed-by: Joe Tsai <thebrokentoaster@gmail.com>
2019-06-27 17:54:42 +00:00
wiretag uint64
tagsize int
unmarshalNeedsValue bool
funcs valueCoderFuncs
validation validationInfo
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>
2019-04-01 20:49:56 +00:00
}
func getExtensionFieldInfo(xt protoreflect.ExtensionType) *extensionFieldInfo {
if xi, ok := xt.(*ExtensionInfo); ok {
xi.lazyInit()
return xi.info
}
protoreflect: remove unnecessary cache so extension types can be GC'ed Resolves golang/protobuf#1521 In the internal/impl package, a helper computes metadata used for serialization and deserialization of extension values. It features a package var of type sync.Map that is used as cache. This ostensibly was for performance, however it has never worked, because the code that updates the cache inserts entries into the wrong map. (These erroneous entries do not cause any issues because they are keys that never conflict with those used in valid queries.) Instead of a one-line fix to have this code update the correct cache, this change removes the cache altogether. The existence of the cache means that once a protoreflect.ExtensionType is used to serialize or deserialize a message, it can *never* be garbage collected. Workloads that are long-lived servers using dynamic messages and extensions based on user requests will exhibit unbounded growth in memory usage as the cache only gets larger and larger. Since the cache has never worked, any advantages it ostensibly conferred have not been missed. So this fixes the unbounded memory growth instead. Change-Id: I15957fd8521852f9f7f9f89db7ebfd7170d85202 Reviewed-on: https://go-review.googlesource.com/c/protobuf/+/560095 LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com> Auto-Submit: Damien Neil <dneil@google.com> Reviewed-by: Michael Stapelberg <stapelberg@google.com> Reviewed-by: Lasse Folger <lassefolger@google.com> Reviewed-by: Damien Neil <dneil@google.com>
2024-02-01 03:33:48 +00:00
// Ideally we'd cache the resulting *extensionFieldInfo so we don't have to
// recompute this metadata repeatedly. But without support for something like
// weak references, such a cache would pin temporary values (like dynamic
// extension types, constructed for the duration of a user request) to the
// heap forever, causing memory usage of the cache to grow unbounded.
// See discussion in https://github.com/golang/protobuf/issues/1521.
return makeExtensionFieldInfo(xt.TypeDescriptor())
}
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>
2019-04-01 20:49:56 +00:00
func makeExtensionFieldInfo(xd protoreflect.ExtensionDescriptor) *extensionFieldInfo {
var wiretag uint64
if !xd.IsPacked() {
wiretag = protowire.EncodeTag(xd.Number(), wireTypes[xd.Kind()])
} else {
wiretag = protowire.EncodeTag(xd.Number(), protowire.BytesType)
}
e := &extensionFieldInfo{
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>
2019-04-01 20:49:56 +00:00
wiretag: wiretag,
tagsize: protowire.SizeVarint(wiretag),
funcs: encoderFuncsForValue(xd),
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>
2019-04-01 20:49:56 +00:00
}
internal/impl: add fast-path unmarshal Benchmarks run with: go test ./benchmarks/ -bench=Wire -benchtime=500ms -benchmem -count=8 Fast-path vs. parent commit: name old time/op new time/op delta Wire/Unmarshal/google_message1_proto2-12 1.35µs ± 2% 0.45µs ± 4% -67.01% (p=0.000 n=8+8) Wire/Unmarshal/google_message1_proto3-12 1.07µs ± 1% 0.31µs ± 1% -71.04% (p=0.000 n=8+8) Wire/Unmarshal/google_message2-12 691µs ± 2% 188µs ± 2% -72.78% (p=0.000 n=7+8) name old allocs/op new allocs/op delta Wire/Unmarshal/google_message1_proto2-12 60.0 ± 0% 25.0 ± 0% -58.33% (p=0.000 n=8+8) Wire/Unmarshal/google_message1_proto3-12 42.0 ± 0% 7.0 ± 0% -83.33% (p=0.000 n=8+8) Wire/Unmarshal/google_message2-12 28.6k ± 0% 8.5k ± 0% -70.34% (p=0.000 n=8+8) Fast-path vs. -v1: name old time/op new time/op delta Wire/Unmarshal/google_message1_proto2-12 702ns ± 1% 445ns ± 4% -36.58% (p=0.000 n=8+8) Wire/Unmarshal/google_message1_proto3-12 604ns ± 1% 311ns ± 1% -48.54% (p=0.000 n=8+8) Wire/Unmarshal/google_message2-12 179µs ± 3% 188µs ± 2% +5.30% (p=0.000 n=7+8) name old allocs/op new allocs/op delta Wire/Unmarshal/google_message1_proto2-12 26.0 ± 0% 25.0 ± 0% -3.85% (p=0.000 n=8+8) Wire/Unmarshal/google_message1_proto3-12 8.00 ± 0% 7.00 ± 0% -12.50% (p=0.000 n=8+8) Wire/Unmarshal/google_message2-12 8.49k ± 0% 8.49k ± 0% -0.01% (p=0.000 n=8+8) Change-Id: I6247ac3fd66a63d9acb902cbd192094ee3d151c3 Reviewed-on: https://go-review.googlesource.com/c/protobuf/+/185147 Reviewed-by: Joe Tsai <thebrokentoaster@gmail.com>
2019-06-27 17:54:42 +00:00
// Does the unmarshal function need a value passed to it?
// This is true for composite types, where we pass in a message, list, or map to fill in,
// and for enums, where we pass in a prototype value to specify the concrete enum type.
switch xd.Kind() {
case protoreflect.MessageKind, protoreflect.GroupKind, protoreflect.EnumKind:
internal/impl: add fast-path unmarshal Benchmarks run with: go test ./benchmarks/ -bench=Wire -benchtime=500ms -benchmem -count=8 Fast-path vs. parent commit: name old time/op new time/op delta Wire/Unmarshal/google_message1_proto2-12 1.35µs ± 2% 0.45µs ± 4% -67.01% (p=0.000 n=8+8) Wire/Unmarshal/google_message1_proto3-12 1.07µs ± 1% 0.31µs ± 1% -71.04% (p=0.000 n=8+8) Wire/Unmarshal/google_message2-12 691µs ± 2% 188µs ± 2% -72.78% (p=0.000 n=7+8) name old allocs/op new allocs/op delta Wire/Unmarshal/google_message1_proto2-12 60.0 ± 0% 25.0 ± 0% -58.33% (p=0.000 n=8+8) Wire/Unmarshal/google_message1_proto3-12 42.0 ± 0% 7.0 ± 0% -83.33% (p=0.000 n=8+8) Wire/Unmarshal/google_message2-12 28.6k ± 0% 8.5k ± 0% -70.34% (p=0.000 n=8+8) Fast-path vs. -v1: name old time/op new time/op delta Wire/Unmarshal/google_message1_proto2-12 702ns ± 1% 445ns ± 4% -36.58% (p=0.000 n=8+8) Wire/Unmarshal/google_message1_proto3-12 604ns ± 1% 311ns ± 1% -48.54% (p=0.000 n=8+8) Wire/Unmarshal/google_message2-12 179µs ± 3% 188µs ± 2% +5.30% (p=0.000 n=7+8) name old allocs/op new allocs/op delta Wire/Unmarshal/google_message1_proto2-12 26.0 ± 0% 25.0 ± 0% -3.85% (p=0.000 n=8+8) Wire/Unmarshal/google_message1_proto3-12 8.00 ± 0% 7.00 ± 0% -12.50% (p=0.000 n=8+8) Wire/Unmarshal/google_message2-12 8.49k ± 0% 8.49k ± 0% -0.01% (p=0.000 n=8+8) Change-Id: I6247ac3fd66a63d9acb902cbd192094ee3d151c3 Reviewed-on: https://go-review.googlesource.com/c/protobuf/+/185147 Reviewed-by: Joe Tsai <thebrokentoaster@gmail.com>
2019-06-27 17:54:42 +00:00
e.unmarshalNeedsValue = true
default:
if xd.Cardinality() == protoreflect.Repeated {
internal/impl: add fast-path unmarshal Benchmarks run with: go test ./benchmarks/ -bench=Wire -benchtime=500ms -benchmem -count=8 Fast-path vs. parent commit: name old time/op new time/op delta Wire/Unmarshal/google_message1_proto2-12 1.35µs ± 2% 0.45µs ± 4% -67.01% (p=0.000 n=8+8) Wire/Unmarshal/google_message1_proto3-12 1.07µs ± 1% 0.31µs ± 1% -71.04% (p=0.000 n=8+8) Wire/Unmarshal/google_message2-12 691µs ± 2% 188µs ± 2% -72.78% (p=0.000 n=7+8) name old allocs/op new allocs/op delta Wire/Unmarshal/google_message1_proto2-12 60.0 ± 0% 25.0 ± 0% -58.33% (p=0.000 n=8+8) Wire/Unmarshal/google_message1_proto3-12 42.0 ± 0% 7.0 ± 0% -83.33% (p=0.000 n=8+8) Wire/Unmarshal/google_message2-12 28.6k ± 0% 8.5k ± 0% -70.34% (p=0.000 n=8+8) Fast-path vs. -v1: name old time/op new time/op delta Wire/Unmarshal/google_message1_proto2-12 702ns ± 1% 445ns ± 4% -36.58% (p=0.000 n=8+8) Wire/Unmarshal/google_message1_proto3-12 604ns ± 1% 311ns ± 1% -48.54% (p=0.000 n=8+8) Wire/Unmarshal/google_message2-12 179µs ± 3% 188µs ± 2% +5.30% (p=0.000 n=7+8) name old allocs/op new allocs/op delta Wire/Unmarshal/google_message1_proto2-12 26.0 ± 0% 25.0 ± 0% -3.85% (p=0.000 n=8+8) Wire/Unmarshal/google_message1_proto3-12 8.00 ± 0% 7.00 ± 0% -12.50% (p=0.000 n=8+8) Wire/Unmarshal/google_message2-12 8.49k ± 0% 8.49k ± 0% -0.01% (p=0.000 n=8+8) Change-Id: I6247ac3fd66a63d9acb902cbd192094ee3d151c3 Reviewed-on: https://go-review.googlesource.com/c/protobuf/+/185147 Reviewed-by: Joe Tsai <thebrokentoaster@gmail.com>
2019-06-27 17:54:42 +00:00
e.unmarshalNeedsValue = true
}
}
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>
2019-04-01 20:49:56 +00:00
return e
}
type lazyExtensionValue struct {
atomicOnce uint32 // atomically set if value is valid
mu sync.Mutex
xi *extensionFieldInfo
value protoreflect.Value
b []byte
fn func() protoreflect.Value
}
type ExtensionField struct {
typ protoreflect.ExtensionType
// value is either the value of GetValue,
// or a *lazyExtensionValue that then returns the value of GetValue.
value protoreflect.Value
lazy *lazyExtensionValue
}
func (f *ExtensionField) appendLazyBytes(xt protoreflect.ExtensionType, xi *extensionFieldInfo, num protowire.Number, wtyp protowire.Type, b []byte) {
if f.lazy == nil {
f.lazy = &lazyExtensionValue{xi: xi}
}
f.typ = xt
f.lazy.xi = xi
f.lazy.b = protowire.AppendTag(f.lazy.b, num, wtyp)
f.lazy.b = append(f.lazy.b, b...)
}
func (f *ExtensionField) canLazy(xt protoreflect.ExtensionType) bool {
if f.typ == nil {
return true
}
if f.typ == xt && f.lazy != nil && atomic.LoadUint32(&f.lazy.atomicOnce) == 0 {
return true
}
return false
}
func (f *ExtensionField) lazyInit() {
f.lazy.mu.Lock()
defer f.lazy.mu.Unlock()
if atomic.LoadUint32(&f.lazy.atomicOnce) == 1 {
return
}
if f.lazy.xi != nil {
b := f.lazy.b
val := f.typ.New()
for len(b) > 0 {
var tag uint64
if b[0] < 0x80 {
tag = uint64(b[0])
b = b[1:]
} else if len(b) >= 2 && b[1] < 128 {
tag = uint64(b[0]&0x7f) + uint64(b[1])<<7
b = b[2:]
} else {
var n int
tag, n = protowire.ConsumeVarint(b)
if n < 0 {
panic(errors.New("bad tag in lazy extension decoding"))
}
b = b[n:]
}
num := protowire.Number(tag >> 3)
wtyp := protowire.Type(tag & 7)
var out unmarshalOutput
var err error
val, out, err = f.lazy.xi.funcs.unmarshal(b, val, num, wtyp, lazyUnmarshalOptions)
if err != nil {
panic(errors.New("decode failure in lazy extension decoding: %v", err))
}
b = b[out.n:]
}
f.lazy.value = val
} else {
f.lazy.value = f.lazy.fn()
}
f.lazy.xi = nil
f.lazy.fn = nil
f.lazy.b = nil
atomic.StoreUint32(&f.lazy.atomicOnce, 1)
}
// Set sets the type and value of the extension field.
// This must not be called concurrently.
func (f *ExtensionField) Set(t protoreflect.ExtensionType, v protoreflect.Value) {
f.typ = t
f.value = v
f.lazy = nil
}
// SetLazy sets the type and a value that is to be lazily evaluated upon first use.
// This must not be called concurrently.
func (f *ExtensionField) SetLazy(t protoreflect.ExtensionType, fn func() protoreflect.Value) {
f.typ = t
f.lazy = &lazyExtensionValue{fn: fn}
}
// Value returns the value of the extension field.
// This may be called concurrently.
func (f *ExtensionField) Value() protoreflect.Value {
if f.lazy != nil {
if atomic.LoadUint32(&f.lazy.atomicOnce) == 0 {
f.lazyInit()
}
return f.lazy.value
}
return f.value
}
// Type returns the type of the extension field.
// This may be called concurrently.
func (f ExtensionField) Type() protoreflect.ExtensionType {
return f.typ
}
// IsSet returns whether the extension field is set.
// This may be called concurrently.
func (f ExtensionField) IsSet() bool {
return f.typ != nil
}
// IsLazy reports whether a field is lazily encoded.
// It is exported for testing.
func IsLazy(m protoreflect.Message, fd protoreflect.FieldDescriptor) bool {
var mi *MessageInfo
var p pointer
switch m := m.(type) {
case *messageState:
mi = m.messageInfo()
p = m.pointer()
case *messageReflectWrapper:
mi = m.messageInfo()
p = m.pointer()
default:
return false
}
xd, ok := fd.(protoreflect.ExtensionTypeDescriptor)
if !ok {
return false
}
xt := xd.Type()
ext := mi.extensionMap(p)
if ext == nil {
return false
}
f, ok := (*ext)[int32(fd.Number())]
if !ok {
return false
}
return f.typ == xt && f.lazy != nil && atomic.LoadUint32(&f.lazy.atomicOnce) == 0
}