Protocol Buffers - Google's data interchange format (grpc依赖) https://developers.google.com/protocol-buffers/
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

1384 lines
50 KiB

// Protocol Buffers - Google's data interchange format
// Copyright 2023 Google LLC. All rights reserved.
//
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file or at
// https://developers.google.com/open-source/licenses/bsd
#include "upb/wire/decode.h"
#include <assert.h>
#include <stdbool.h>
#include <stddef.h>
#include <stdint.h>
6 years ago
#include <string.h>
#include "upb/base/descriptor_constants.h"
#include "upb/base/string_view.h"
#include "upb/hash/common.h"
#include "upb/mem/arena.h"
#include "upb/message/array.h"
#include "upb/message/internal/accessors.h"
#include "upb/message/internal/array.h"
#include "upb/message/internal/extension.h"
#include "upb/message/internal/map.h"
#include "upb/message/internal/map_entry.h"
#include "upb/message/internal/message.h"
#include "upb/message/map.h"
#include "upb/message/message.h"
#include "upb/message/tagged_ptr.h"
#include "upb/mini_table/enum.h"
#include "upb/mini_table/extension.h"
#include "upb/mini_table/extension_registry.h"
#include "upb/mini_table/field.h"
#include "upb/mini_table/internal/size_log2.h"
#include "upb/mini_table/message.h"
#include "upb/mini_table/sub.h"
#include "upb/port/atomic.h"
#include "upb/wire/encode.h"
#include "upb/wire/eps_copy_input_stream.h"
#include "upb/wire/internal/constants.h"
#include "upb/wire/internal/decode.h"
#include "upb/wire/internal/swap.h"
#include "upb/wire/reader.h"
// Must be last.
#include "upb/port/def.inc"
// A few fake field types for our tables.
enum {
kUpb_FakeFieldType_FieldNotFound = 0,
kUpb_FakeFieldType_MessageSetItem = 19,
};
// DecodeOp: an action to be performed for a wire-type/field-type combination.
enum {
// Special ops: we don't write data to regular fields for these.
kUpb_DecodeOp_UnknownField = -1,
kUpb_DecodeOp_MessageSetItem = -2,
// Scalar-only ops.
kUpb_DecodeOp_Scalar1Byte = 0,
kUpb_DecodeOp_Scalar4Byte = 2,
kUpb_DecodeOp_Scalar8Byte = 3,
kUpb_DecodeOp_Enum = 1,
// Scalar/repeated ops.
kUpb_DecodeOp_String = 4,
kUpb_DecodeOp_Bytes = 5,
kUpb_DecodeOp_SubMessage = 6,
// Repeated-only ops (also see macros below).
kUpb_DecodeOp_PackedEnum = 13,
};
// For packed fields it is helpful to be able to recover the lg2 of the data
// size from the op.
#define OP_FIXPCK_LG2(n) (n + 5) /* n in [2, 3] => op in [7, 8] */
#define OP_VARPCK_LG2(n) (n + 9) /* n in [0, 2, 3] => op in [9, 11, 12] */
typedef union {
bool bool_val;
uint32_t uint32_val;
uint64_t uint64_val;
uint32_t size;
} wireval;
// Ideally these two functions should take the owning MiniTable pointer as a
// first argument, then we could just put them in mini_table/message.h as nice
// clean getters. But we don't have that so instead we gotta write these
// Frankenfunctions that take an array of subtables.
// TODO: Move these to mini_table/ anyway since there are other places
// that could use them.
// Returns the MiniTable corresponding to a given MiniTableField
// from an array of MiniTableSubs.
static const upb_MiniTable* _upb_MiniTableSubs_MessageByField(
const upb_MiniTableSub* subs, const upb_MiniTableField* field) {
return upb_MiniTableSub_Message(subs[field->UPB_PRIVATE(submsg_index)]);
}
// Returns the MiniTableEnum corresponding to a given MiniTableField
// from an array of MiniTableSub.
static const upb_MiniTableEnum* _upb_MiniTableSubs_EnumByField(
const upb_MiniTableSub* subs, const upb_MiniTableField* field) {
return upb_MiniTableSub_Enum(subs[field->UPB_PRIVATE(submsg_index)]);
}
static const char* _upb_Decoder_DecodeMessage(upb_Decoder* d, const char* ptr,
upb_Message* msg,
const upb_MiniTable* layout);
UPB_NORETURN static void* _upb_Decoder_ErrorJmp(upb_Decoder* d,
upb_DecodeStatus status) {
UPB_ASSERT(status != kUpb_DecodeStatus_Ok);
d->status = status;
UPB_LONGJMP(d->err, 1);
}
const char* _upb_FastDecoder_ErrorJmp(upb_Decoder* d, int status) {
UPB_ASSERT(status != kUpb_DecodeStatus_Ok);
d->status = status;
UPB_LONGJMP(d->err, 1);
return NULL;
}
static void _upb_Decoder_VerifyUtf8(upb_Decoder* d, const char* buf, int len) {
if (!_upb_Decoder_VerifyUtf8Inline(buf, len)) {
_upb_Decoder_ErrorJmp(d, kUpb_DecodeStatus_BadUtf8);
}
}
static bool _upb_Decoder_Reserve(upb_Decoder* d, upb_Array* arr, size_t elem) {
bool need_realloc =
arr->UPB_PRIVATE(capacity) - arr->UPB_PRIVATE(size) < elem;
if (need_realloc && !UPB_PRIVATE(_upb_Array_Realloc)(
arr, arr->UPB_PRIVATE(size) + elem, &d->arena)) {
_upb_Decoder_ErrorJmp(d, kUpb_DecodeStatus_OutOfMemory);
}
return need_realloc;
}
typedef struct {
const char* ptr;
uint64_t val;
} _upb_DecodeLongVarintReturn;
UPB_NOINLINE
static _upb_DecodeLongVarintReturn _upb_Decoder_DecodeLongVarint(
const char* ptr, uint64_t val) {
_upb_DecodeLongVarintReturn ret = {NULL, 0};
uint64_t byte;
int i;
for (i = 1; i < 10; i++) {
byte = (uint8_t)ptr[i];
val += (byte - 1) << (i * 7);
if (!(byte & 0x80)) {
ret.ptr = ptr + i + 1;
ret.val = val;
return ret;
}
}
return ret;
}
UPB_FORCEINLINE
static const char* _upb_Decoder_DecodeVarint(upb_Decoder* d, const char* ptr,
uint64_t* val) {
uint64_t byte = (uint8_t)*ptr;
if (UPB_LIKELY((byte & 0x80) == 0)) {
*val = byte;
return ptr + 1;
} else {
_upb_DecodeLongVarintReturn res = _upb_Decoder_DecodeLongVarint(ptr, byte);
if (!res.ptr) _upb_Decoder_ErrorJmp(d, kUpb_DecodeStatus_Malformed);
*val = res.val;
return res.ptr;
}
}
UPB_FORCEINLINE
static const char* _upb_Decoder_DecodeTag(upb_Decoder* d, const char* ptr,
uint32_t* val) {
uint64_t byte = (uint8_t)*ptr;
if (UPB_LIKELY((byte & 0x80) == 0)) {
*val = byte;
return ptr + 1;
} else {
const char* start = ptr;
_upb_DecodeLongVarintReturn res = _upb_Decoder_DecodeLongVarint(ptr, byte);
if (!res.ptr || res.ptr - start > 5 || res.val > UINT32_MAX) {
_upb_Decoder_ErrorJmp(d, kUpb_DecodeStatus_Malformed);
}
*val = res.val;
return res.ptr;
}
}
UPB_FORCEINLINE
static const char* upb_Decoder_DecodeSize(upb_Decoder* d, const char* ptr,
uint32_t* size) {
uint64_t size64;
ptr = _upb_Decoder_DecodeVarint(d, ptr, &size64);
if (size64 >= INT32_MAX ||
!upb_EpsCopyInputStream_CheckSize(&d->input, ptr, (int)size64)) {
_upb_Decoder_ErrorJmp(d, kUpb_DecodeStatus_Malformed);
}
*size = size64;
return ptr;
}
static void _upb_Decoder_MungeInt32(wireval* val) {
if (!_upb_IsLittleEndian()) {
/* The next stage will memcpy(dst, &val, 4) */
val->uint32_val = val->uint64_val;
}
}
static void _upb_Decoder_Munge(int type, wireval* val) {
switch (type) {
case kUpb_FieldType_Bool:
val->bool_val = val->uint64_val != 0;
break;
case kUpb_FieldType_SInt32: {
uint32_t n = val->uint64_val;
val->uint32_val = (n >> 1) ^ -(int32_t)(n & 1);
break;
}
case kUpb_FieldType_SInt64: {
uint64_t n = val->uint64_val;
val->uint64_val = (n >> 1) ^ -(int64_t)(n & 1);
break;
}
case kUpb_FieldType_Int32:
case kUpb_FieldType_UInt32:
case kUpb_FieldType_Enum:
_upb_Decoder_MungeInt32(val);
break;
7 years ago
}
}
Added a new dynamic tree shaking model to upb, with the intention of removing the old model once YouTube has migrated. The `kUpb_DecodeOption_ExperimentalAllowUnlinked` flag to the decoder will enable the new behavior. When that flag is not passed, tree shaking with the old model will still be possible. "Dynamic tree shaking" in upb is a feature that allows messages to be parsed even if the MiniTables have not been fully linked. Unlinked sub-message fields can be parsed by preserving their data in the unknown fields. If the application later discovers that the message field is actually needed, the MiniTable can be patched to properly link that field, and existing message instances can "promote" the data from the unknown fields to an actual message of the correct type. Before this change, dynamic tree shaking stored unparsed message data in the unknown fields of the *parent*. In effect, we were treating the field as if it did not exist at all. This meant that parsing an unlinked field did not affect the hasbits or oneof cases of the parent, nor did it create a `upb_Array` or `upb_Map` for array/map fields. Only when a message was linked and promoted did any of these things occur. While this model had some amount of conceptual simplicity, it caused significant problems with oneofs. When multiple fields inside a single oneof are parsed from the wire, order matters, because later oneof fields must overwrite earlier ones. Dynamic tree shaking can mean that some fields in a oneof are linked while others are not. It is essential that we preserve this ordering semantic even when dynamic tree shaking is being used, but it is difficult to do if the oneof's data can be split between linked fields (which have been reified into parsed field data) and unlinked fields (whose data lives in the unknown fields of the parent). To solve this problem, this CL changes the representation for unlinked fields. Instead of being placed in the parent's unknown fields, we create an actual message instance for each unlinked message we parse, but we use a placeholder "empty message" MiniTable as the message's type. All of the message's data will therefore be placed into the "empty message's" unknown fields. But unlike before, this "empty message" is actually present according to the hasbits, oneof case, and `upb_Array`/`upb_Map` of the parent. This means that all of the oneof presence logic works as normal. Since the MiniTable can be patched at any time, we need a bit in the message instance itself to signal whether a pointer to a sub-message is an "empty message" or not. When dynamic tree shaking is in use, all users must be capable of recognizing an empty message and acting accordingly (promoting, etc) even if the MiniTable itself says that the field is linked. Because dynamic tree shaking imposes this extra requirement on users, we require that users pass an extra option to the decoder to allow parsing of unlinked sub-messages. Many existing users of upb (Ruby, PHP, Python, etc) will always have fully-linked MiniTables, so there is no reason for them to add extra logic to handle empty messages. By omitting the `kUpb_DecodeOption_ExperimentalAllowUnlinked` option, they will be relieved of the duty to check the tagged pointer that would indicate an empty, unlinked message. For existing users of dynamic tree shaking, there are three main changes: 1. The APIs in message/promote.h have changed, and users will need to update to the new interfaces. 2. The model for maps has changed slightly. Before, we required that map entries always had their values linked; for dynamic tree shaking to apply to maps, we required that the *entry* was left unlinked, not the entry's value. In the new model, that is reversed: map entries must always be linked, but a map entry's value can be unlinked. 3. The presence model for unlinked fields has changed. Unlinked fields will now register as "present" from the perspective of hasbits, oneof cases, and array/map entries. Users must test the tagged pointer to know if a message is of the correct, linked type or whether it is a placeholder "empty" message. There is a new function `upb_Message_GetTaggedMessagePtr()`, as well as a new accessor `upb_MessageValue.tagged_msg_val` that can be used to read and test the tagged pointer directly. PiperOrigin-RevId: 535288031
2 years ago
static upb_Message* _upb_Decoder_NewSubMessage(upb_Decoder* d,
const upb_MiniTableSub* subs,
const upb_MiniTableField* field,
upb_TaggedMessagePtr* target) {
const upb_MiniTable* subl = _upb_MiniTableSubs_MessageByField(subs, field);
UPB_ASSERT(subl);
upb_Message* msg = _upb_Message_New(subl, &d->arena);
if (!msg) _upb_Decoder_ErrorJmp(d, kUpb_DecodeStatus_OutOfMemory);
Added a new dynamic tree shaking model to upb, with the intention of removing the old model once YouTube has migrated. The `kUpb_DecodeOption_ExperimentalAllowUnlinked` flag to the decoder will enable the new behavior. When that flag is not passed, tree shaking with the old model will still be possible. "Dynamic tree shaking" in upb is a feature that allows messages to be parsed even if the MiniTables have not been fully linked. Unlinked sub-message fields can be parsed by preserving their data in the unknown fields. If the application later discovers that the message field is actually needed, the MiniTable can be patched to properly link that field, and existing message instances can "promote" the data from the unknown fields to an actual message of the correct type. Before this change, dynamic tree shaking stored unparsed message data in the unknown fields of the *parent*. In effect, we were treating the field as if it did not exist at all. This meant that parsing an unlinked field did not affect the hasbits or oneof cases of the parent, nor did it create a `upb_Array` or `upb_Map` for array/map fields. Only when a message was linked and promoted did any of these things occur. While this model had some amount of conceptual simplicity, it caused significant problems with oneofs. When multiple fields inside a single oneof are parsed from the wire, order matters, because later oneof fields must overwrite earlier ones. Dynamic tree shaking can mean that some fields in a oneof are linked while others are not. It is essential that we preserve this ordering semantic even when dynamic tree shaking is being used, but it is difficult to do if the oneof's data can be split between linked fields (which have been reified into parsed field data) and unlinked fields (whose data lives in the unknown fields of the parent). To solve this problem, this CL changes the representation for unlinked fields. Instead of being placed in the parent's unknown fields, we create an actual message instance for each unlinked message we parse, but we use a placeholder "empty message" MiniTable as the message's type. All of the message's data will therefore be placed into the "empty message's" unknown fields. But unlike before, this "empty message" is actually present according to the hasbits, oneof case, and `upb_Array`/`upb_Map` of the parent. This means that all of the oneof presence logic works as normal. Since the MiniTable can be patched at any time, we need a bit in the message instance itself to signal whether a pointer to a sub-message is an "empty message" or not. When dynamic tree shaking is in use, all users must be capable of recognizing an empty message and acting accordingly (promoting, etc) even if the MiniTable itself says that the field is linked. Because dynamic tree shaking imposes this extra requirement on users, we require that users pass an extra option to the decoder to allow parsing of unlinked sub-messages. Many existing users of upb (Ruby, PHP, Python, etc) will always have fully-linked MiniTables, so there is no reason for them to add extra logic to handle empty messages. By omitting the `kUpb_DecodeOption_ExperimentalAllowUnlinked` option, they will be relieved of the duty to check the tagged pointer that would indicate an empty, unlinked message. For existing users of dynamic tree shaking, there are three main changes: 1. The APIs in message/promote.h have changed, and users will need to update to the new interfaces. 2. The model for maps has changed slightly. Before, we required that map entries always had their values linked; for dynamic tree shaking to apply to maps, we required that the *entry* was left unlinked, not the entry's value. In the new model, that is reversed: map entries must always be linked, but a map entry's value can be unlinked. 3. The presence model for unlinked fields has changed. Unlinked fields will now register as "present" from the perspective of hasbits, oneof cases, and array/map entries. Users must test the tagged pointer to know if a message is of the correct, linked type or whether it is a placeholder "empty" message. There is a new function `upb_Message_GetTaggedMessagePtr()`, as well as a new accessor `upb_MessageValue.tagged_msg_val` that can be used to read and test the tagged pointer directly. PiperOrigin-RevId: 535288031
2 years ago
// Extensions should not be unlinked. A message extension should not be
Added a new dynamic tree shaking model to upb, with the intention of removing the old model once YouTube has migrated. The `kUpb_DecodeOption_ExperimentalAllowUnlinked` flag to the decoder will enable the new behavior. When that flag is not passed, tree shaking with the old model will still be possible. "Dynamic tree shaking" in upb is a feature that allows messages to be parsed even if the MiniTables have not been fully linked. Unlinked sub-message fields can be parsed by preserving their data in the unknown fields. If the application later discovers that the message field is actually needed, the MiniTable can be patched to properly link that field, and existing message instances can "promote" the data from the unknown fields to an actual message of the correct type. Before this change, dynamic tree shaking stored unparsed message data in the unknown fields of the *parent*. In effect, we were treating the field as if it did not exist at all. This meant that parsing an unlinked field did not affect the hasbits or oneof cases of the parent, nor did it create a `upb_Array` or `upb_Map` for array/map fields. Only when a message was linked and promoted did any of these things occur. While this model had some amount of conceptual simplicity, it caused significant problems with oneofs. When multiple fields inside a single oneof are parsed from the wire, order matters, because later oneof fields must overwrite earlier ones. Dynamic tree shaking can mean that some fields in a oneof are linked while others are not. It is essential that we preserve this ordering semantic even when dynamic tree shaking is being used, but it is difficult to do if the oneof's data can be split between linked fields (which have been reified into parsed field data) and unlinked fields (whose data lives in the unknown fields of the parent). To solve this problem, this CL changes the representation for unlinked fields. Instead of being placed in the parent's unknown fields, we create an actual message instance for each unlinked message we parse, but we use a placeholder "empty message" MiniTable as the message's type. All of the message's data will therefore be placed into the "empty message's" unknown fields. But unlike before, this "empty message" is actually present according to the hasbits, oneof case, and `upb_Array`/`upb_Map` of the parent. This means that all of the oneof presence logic works as normal. Since the MiniTable can be patched at any time, we need a bit in the message instance itself to signal whether a pointer to a sub-message is an "empty message" or not. When dynamic tree shaking is in use, all users must be capable of recognizing an empty message and acting accordingly (promoting, etc) even if the MiniTable itself says that the field is linked. Because dynamic tree shaking imposes this extra requirement on users, we require that users pass an extra option to the decoder to allow parsing of unlinked sub-messages. Many existing users of upb (Ruby, PHP, Python, etc) will always have fully-linked MiniTables, so there is no reason for them to add extra logic to handle empty messages. By omitting the `kUpb_DecodeOption_ExperimentalAllowUnlinked` option, they will be relieved of the duty to check the tagged pointer that would indicate an empty, unlinked message. For existing users of dynamic tree shaking, there are three main changes: 1. The APIs in message/promote.h have changed, and users will need to update to the new interfaces. 2. The model for maps has changed slightly. Before, we required that map entries always had their values linked; for dynamic tree shaking to apply to maps, we required that the *entry* was left unlinked, not the entry's value. In the new model, that is reversed: map entries must always be linked, but a map entry's value can be unlinked. 3. The presence model for unlinked fields has changed. Unlinked fields will now register as "present" from the perspective of hasbits, oneof cases, and array/map entries. Users must test the tagged pointer to know if a message is of the correct, linked type or whether it is a placeholder "empty" message. There is a new function `upb_Message_GetTaggedMessagePtr()`, as well as a new accessor `upb_MessageValue.tagged_msg_val` that can be used to read and test the tagged pointer directly. PiperOrigin-RevId: 535288031
2 years ago
// registered until its sub-message type is available to be linked.
bool is_empty = UPB_PRIVATE(_upb_MiniTable_IsEmpty)(subl);
bool is_extension = field->UPB_PRIVATE(mode) & kUpb_LabelFlags_IsExtension;
Added a new dynamic tree shaking model to upb, with the intention of removing the old model once YouTube has migrated. The `kUpb_DecodeOption_ExperimentalAllowUnlinked` flag to the decoder will enable the new behavior. When that flag is not passed, tree shaking with the old model will still be possible. "Dynamic tree shaking" in upb is a feature that allows messages to be parsed even if the MiniTables have not been fully linked. Unlinked sub-message fields can be parsed by preserving their data in the unknown fields. If the application later discovers that the message field is actually needed, the MiniTable can be patched to properly link that field, and existing message instances can "promote" the data from the unknown fields to an actual message of the correct type. Before this change, dynamic tree shaking stored unparsed message data in the unknown fields of the *parent*. In effect, we were treating the field as if it did not exist at all. This meant that parsing an unlinked field did not affect the hasbits or oneof cases of the parent, nor did it create a `upb_Array` or `upb_Map` for array/map fields. Only when a message was linked and promoted did any of these things occur. While this model had some amount of conceptual simplicity, it caused significant problems with oneofs. When multiple fields inside a single oneof are parsed from the wire, order matters, because later oneof fields must overwrite earlier ones. Dynamic tree shaking can mean that some fields in a oneof are linked while others are not. It is essential that we preserve this ordering semantic even when dynamic tree shaking is being used, but it is difficult to do if the oneof's data can be split between linked fields (which have been reified into parsed field data) and unlinked fields (whose data lives in the unknown fields of the parent). To solve this problem, this CL changes the representation for unlinked fields. Instead of being placed in the parent's unknown fields, we create an actual message instance for each unlinked message we parse, but we use a placeholder "empty message" MiniTable as the message's type. All of the message's data will therefore be placed into the "empty message's" unknown fields. But unlike before, this "empty message" is actually present according to the hasbits, oneof case, and `upb_Array`/`upb_Map` of the parent. This means that all of the oneof presence logic works as normal. Since the MiniTable can be patched at any time, we need a bit in the message instance itself to signal whether a pointer to a sub-message is an "empty message" or not. When dynamic tree shaking is in use, all users must be capable of recognizing an empty message and acting accordingly (promoting, etc) even if the MiniTable itself says that the field is linked. Because dynamic tree shaking imposes this extra requirement on users, we require that users pass an extra option to the decoder to allow parsing of unlinked sub-messages. Many existing users of upb (Ruby, PHP, Python, etc) will always have fully-linked MiniTables, so there is no reason for them to add extra logic to handle empty messages. By omitting the `kUpb_DecodeOption_ExperimentalAllowUnlinked` option, they will be relieved of the duty to check the tagged pointer that would indicate an empty, unlinked message. For existing users of dynamic tree shaking, there are three main changes: 1. The APIs in message/promote.h have changed, and users will need to update to the new interfaces. 2. The model for maps has changed slightly. Before, we required that map entries always had their values linked; for dynamic tree shaking to apply to maps, we required that the *entry* was left unlinked, not the entry's value. In the new model, that is reversed: map entries must always be linked, but a map entry's value can be unlinked. 3. The presence model for unlinked fields has changed. Unlinked fields will now register as "present" from the perspective of hasbits, oneof cases, and array/map entries. Users must test the tagged pointer to know if a message is of the correct, linked type or whether it is a placeholder "empty" message. There is a new function `upb_Message_GetTaggedMessagePtr()`, as well as a new accessor `upb_MessageValue.tagged_msg_val` that can be used to read and test the tagged pointer directly. PiperOrigin-RevId: 535288031
2 years ago
UPB_ASSERT(!(is_empty && is_extension));
if (is_empty && !(d->options & kUpb_DecodeOption_ExperimentalAllowUnlinked)) {
_upb_Decoder_ErrorJmp(d, kUpb_DecodeStatus_UnlinkedSubMessage);
}
upb_TaggedMessagePtr tagged = _upb_TaggedMessagePtr_Pack(msg, is_empty);
memcpy(target, &tagged, sizeof(tagged));
return msg;
}
Added a new dynamic tree shaking model to upb, with the intention of removing the old model once YouTube has migrated. The `kUpb_DecodeOption_ExperimentalAllowUnlinked` flag to the decoder will enable the new behavior. When that flag is not passed, tree shaking with the old model will still be possible. "Dynamic tree shaking" in upb is a feature that allows messages to be parsed even if the MiniTables have not been fully linked. Unlinked sub-message fields can be parsed by preserving their data in the unknown fields. If the application later discovers that the message field is actually needed, the MiniTable can be patched to properly link that field, and existing message instances can "promote" the data from the unknown fields to an actual message of the correct type. Before this change, dynamic tree shaking stored unparsed message data in the unknown fields of the *parent*. In effect, we were treating the field as if it did not exist at all. This meant that parsing an unlinked field did not affect the hasbits or oneof cases of the parent, nor did it create a `upb_Array` or `upb_Map` for array/map fields. Only when a message was linked and promoted did any of these things occur. While this model had some amount of conceptual simplicity, it caused significant problems with oneofs. When multiple fields inside a single oneof are parsed from the wire, order matters, because later oneof fields must overwrite earlier ones. Dynamic tree shaking can mean that some fields in a oneof are linked while others are not. It is essential that we preserve this ordering semantic even when dynamic tree shaking is being used, but it is difficult to do if the oneof's data can be split between linked fields (which have been reified into parsed field data) and unlinked fields (whose data lives in the unknown fields of the parent). To solve this problem, this CL changes the representation for unlinked fields. Instead of being placed in the parent's unknown fields, we create an actual message instance for each unlinked message we parse, but we use a placeholder "empty message" MiniTable as the message's type. All of the message's data will therefore be placed into the "empty message's" unknown fields. But unlike before, this "empty message" is actually present according to the hasbits, oneof case, and `upb_Array`/`upb_Map` of the parent. This means that all of the oneof presence logic works as normal. Since the MiniTable can be patched at any time, we need a bit in the message instance itself to signal whether a pointer to a sub-message is an "empty message" or not. When dynamic tree shaking is in use, all users must be capable of recognizing an empty message and acting accordingly (promoting, etc) even if the MiniTable itself says that the field is linked. Because dynamic tree shaking imposes this extra requirement on users, we require that users pass an extra option to the decoder to allow parsing of unlinked sub-messages. Many existing users of upb (Ruby, PHP, Python, etc) will always have fully-linked MiniTables, so there is no reason for them to add extra logic to handle empty messages. By omitting the `kUpb_DecodeOption_ExperimentalAllowUnlinked` option, they will be relieved of the duty to check the tagged pointer that would indicate an empty, unlinked message. For existing users of dynamic tree shaking, there are three main changes: 1. The APIs in message/promote.h have changed, and users will need to update to the new interfaces. 2. The model for maps has changed slightly. Before, we required that map entries always had their values linked; for dynamic tree shaking to apply to maps, we required that the *entry* was left unlinked, not the entry's value. In the new model, that is reversed: map entries must always be linked, but a map entry's value can be unlinked. 3. The presence model for unlinked fields has changed. Unlinked fields will now register as "present" from the perspective of hasbits, oneof cases, and array/map entries. Users must test the tagged pointer to know if a message is of the correct, linked type or whether it is a placeholder "empty" message. There is a new function `upb_Message_GetTaggedMessagePtr()`, as well as a new accessor `upb_MessageValue.tagged_msg_val` that can be used to read and test the tagged pointer directly. PiperOrigin-RevId: 535288031
2 years ago
static upb_Message* _upb_Decoder_ReuseSubMessage(
upb_Decoder* d, const upb_MiniTableSub* subs,
const upb_MiniTableField* field, upb_TaggedMessagePtr* target) {
upb_TaggedMessagePtr tagged = *target;
const upb_MiniTable* subl = _upb_MiniTableSubs_MessageByField(subs, field);
Added a new dynamic tree shaking model to upb, with the intention of removing the old model once YouTube has migrated. The `kUpb_DecodeOption_ExperimentalAllowUnlinked` flag to the decoder will enable the new behavior. When that flag is not passed, tree shaking with the old model will still be possible. "Dynamic tree shaking" in upb is a feature that allows messages to be parsed even if the MiniTables have not been fully linked. Unlinked sub-message fields can be parsed by preserving their data in the unknown fields. If the application later discovers that the message field is actually needed, the MiniTable can be patched to properly link that field, and existing message instances can "promote" the data from the unknown fields to an actual message of the correct type. Before this change, dynamic tree shaking stored unparsed message data in the unknown fields of the *parent*. In effect, we were treating the field as if it did not exist at all. This meant that parsing an unlinked field did not affect the hasbits or oneof cases of the parent, nor did it create a `upb_Array` or `upb_Map` for array/map fields. Only when a message was linked and promoted did any of these things occur. While this model had some amount of conceptual simplicity, it caused significant problems with oneofs. When multiple fields inside a single oneof are parsed from the wire, order matters, because later oneof fields must overwrite earlier ones. Dynamic tree shaking can mean that some fields in a oneof are linked while others are not. It is essential that we preserve this ordering semantic even when dynamic tree shaking is being used, but it is difficult to do if the oneof's data can be split between linked fields (which have been reified into parsed field data) and unlinked fields (whose data lives in the unknown fields of the parent). To solve this problem, this CL changes the representation for unlinked fields. Instead of being placed in the parent's unknown fields, we create an actual message instance for each unlinked message we parse, but we use a placeholder "empty message" MiniTable as the message's type. All of the message's data will therefore be placed into the "empty message's" unknown fields. But unlike before, this "empty message" is actually present according to the hasbits, oneof case, and `upb_Array`/`upb_Map` of the parent. This means that all of the oneof presence logic works as normal. Since the MiniTable can be patched at any time, we need a bit in the message instance itself to signal whether a pointer to a sub-message is an "empty message" or not. When dynamic tree shaking is in use, all users must be capable of recognizing an empty message and acting accordingly (promoting, etc) even if the MiniTable itself says that the field is linked. Because dynamic tree shaking imposes this extra requirement on users, we require that users pass an extra option to the decoder to allow parsing of unlinked sub-messages. Many existing users of upb (Ruby, PHP, Python, etc) will always have fully-linked MiniTables, so there is no reason for them to add extra logic to handle empty messages. By omitting the `kUpb_DecodeOption_ExperimentalAllowUnlinked` option, they will be relieved of the duty to check the tagged pointer that would indicate an empty, unlinked message. For existing users of dynamic tree shaking, there are three main changes: 1. The APIs in message/promote.h have changed, and users will need to update to the new interfaces. 2. The model for maps has changed slightly. Before, we required that map entries always had their values linked; for dynamic tree shaking to apply to maps, we required that the *entry* was left unlinked, not the entry's value. In the new model, that is reversed: map entries must always be linked, but a map entry's value can be unlinked. 3. The presence model for unlinked fields has changed. Unlinked fields will now register as "present" from the perspective of hasbits, oneof cases, and array/map entries. Users must test the tagged pointer to know if a message is of the correct, linked type or whether it is a placeholder "empty" message. There is a new function `upb_Message_GetTaggedMessagePtr()`, as well as a new accessor `upb_MessageValue.tagged_msg_val` that can be used to read and test the tagged pointer directly. PiperOrigin-RevId: 535288031
2 years ago
UPB_ASSERT(subl);
if (!upb_TaggedMessagePtr_IsEmpty(tagged) ||
UPB_PRIVATE(_upb_MiniTable_IsEmpty)(subl)) {
Added a new dynamic tree shaking model to upb, with the intention of removing the old model once YouTube has migrated. The `kUpb_DecodeOption_ExperimentalAllowUnlinked` flag to the decoder will enable the new behavior. When that flag is not passed, tree shaking with the old model will still be possible. "Dynamic tree shaking" in upb is a feature that allows messages to be parsed even if the MiniTables have not been fully linked. Unlinked sub-message fields can be parsed by preserving their data in the unknown fields. If the application later discovers that the message field is actually needed, the MiniTable can be patched to properly link that field, and existing message instances can "promote" the data from the unknown fields to an actual message of the correct type. Before this change, dynamic tree shaking stored unparsed message data in the unknown fields of the *parent*. In effect, we were treating the field as if it did not exist at all. This meant that parsing an unlinked field did not affect the hasbits or oneof cases of the parent, nor did it create a `upb_Array` or `upb_Map` for array/map fields. Only when a message was linked and promoted did any of these things occur. While this model had some amount of conceptual simplicity, it caused significant problems with oneofs. When multiple fields inside a single oneof are parsed from the wire, order matters, because later oneof fields must overwrite earlier ones. Dynamic tree shaking can mean that some fields in a oneof are linked while others are not. It is essential that we preserve this ordering semantic even when dynamic tree shaking is being used, but it is difficult to do if the oneof's data can be split between linked fields (which have been reified into parsed field data) and unlinked fields (whose data lives in the unknown fields of the parent). To solve this problem, this CL changes the representation for unlinked fields. Instead of being placed in the parent's unknown fields, we create an actual message instance for each unlinked message we parse, but we use a placeholder "empty message" MiniTable as the message's type. All of the message's data will therefore be placed into the "empty message's" unknown fields. But unlike before, this "empty message" is actually present according to the hasbits, oneof case, and `upb_Array`/`upb_Map` of the parent. This means that all of the oneof presence logic works as normal. Since the MiniTable can be patched at any time, we need a bit in the message instance itself to signal whether a pointer to a sub-message is an "empty message" or not. When dynamic tree shaking is in use, all users must be capable of recognizing an empty message and acting accordingly (promoting, etc) even if the MiniTable itself says that the field is linked. Because dynamic tree shaking imposes this extra requirement on users, we require that users pass an extra option to the decoder to allow parsing of unlinked sub-messages. Many existing users of upb (Ruby, PHP, Python, etc) will always have fully-linked MiniTables, so there is no reason for them to add extra logic to handle empty messages. By omitting the `kUpb_DecodeOption_ExperimentalAllowUnlinked` option, they will be relieved of the duty to check the tagged pointer that would indicate an empty, unlinked message. For existing users of dynamic tree shaking, there are three main changes: 1. The APIs in message/promote.h have changed, and users will need to update to the new interfaces. 2. The model for maps has changed slightly. Before, we required that map entries always had their values linked; for dynamic tree shaking to apply to maps, we required that the *entry* was left unlinked, not the entry's value. In the new model, that is reversed: map entries must always be linked, but a map entry's value can be unlinked. 3. The presence model for unlinked fields has changed. Unlinked fields will now register as "present" from the perspective of hasbits, oneof cases, and array/map entries. Users must test the tagged pointer to know if a message is of the correct, linked type or whether it is a placeholder "empty" message. There is a new function `upb_Message_GetTaggedMessagePtr()`, as well as a new accessor `upb_MessageValue.tagged_msg_val` that can be used to read and test the tagged pointer directly. PiperOrigin-RevId: 535288031
2 years ago
return _upb_TaggedMessagePtr_GetMessage(tagged);
}
// We found an empty message from a previous parse that was performed before
// this field was linked. But it is linked now, so we want to allocate a new
// message of the correct type and promote data into it before continuing.
upb_Message* existing = _upb_TaggedMessagePtr_GetEmptyMessage(tagged);
upb_Message* promoted = _upb_Decoder_NewSubMessage(d, subs, field, target);
size_t size;
const char* unknown = upb_Message_GetUnknown(existing, &size);
upb_DecodeStatus status = upb_Decode(unknown, size, promoted, subl, d->extreg,
d->options, &d->arena);
if (status != kUpb_DecodeStatus_Ok) _upb_Decoder_ErrorJmp(d, status);
return promoted;
}
static const char* _upb_Decoder_ReadString(upb_Decoder* d, const char* ptr,
int size, upb_StringView* str) {
const char* str_ptr = ptr;
ptr = upb_EpsCopyInputStream_ReadString(&d->input, &str_ptr, size, &d->arena);
if (!ptr) _upb_Decoder_ErrorJmp(d, kUpb_DecodeStatus_OutOfMemory);
str->data = str_ptr;
str->size = size;
return ptr;
}
UPB_FORCEINLINE
static const char* _upb_Decoder_RecurseSubMessage(upb_Decoder* d,
const char* ptr,
upb_Message* submsg,
const upb_MiniTable* subl,
uint32_t expected_end_group) {
if (--d->depth < 0) {
_upb_Decoder_ErrorJmp(d, kUpb_DecodeStatus_MaxDepthExceeded);
}
ptr = _upb_Decoder_DecodeMessage(d, ptr, submsg, subl);
d->depth++;
if (d->end_group != expected_end_group) {
_upb_Decoder_ErrorJmp(d, kUpb_DecodeStatus_Malformed);
}
return ptr;
}
UPB_FORCEINLINE
static const char* _upb_Decoder_DecodeSubMessage(
upb_Decoder* d, const char* ptr, upb_Message* submsg,
const upb_MiniTableSub* subs, const upb_MiniTableField* field, int size) {
int saved_delta = upb_EpsCopyInputStream_PushLimit(&d->input, ptr, size);
const upb_MiniTable* subl = _upb_MiniTableSubs_MessageByField(subs, field);
UPB_ASSERT(subl);
ptr = _upb_Decoder_RecurseSubMessage(d, ptr, submsg, subl, DECODE_NOGROUP);
upb_EpsCopyInputStream_PopLimit(&d->input, ptr, saved_delta);
return ptr;
}
UPB_FORCEINLINE
static const char* _upb_Decoder_DecodeGroup(upb_Decoder* d, const char* ptr,
upb_Message* submsg,
const upb_MiniTable* subl,
uint32_t number) {
if (_upb_Decoder_IsDone(d, &ptr)) {
_upb_Decoder_ErrorJmp(d, kUpb_DecodeStatus_Malformed);
}
ptr = _upb_Decoder_RecurseSubMessage(d, ptr, submsg, subl, number);
d->end_group = DECODE_NOGROUP;
return ptr;
}
UPB_FORCEINLINE
static const char* _upb_Decoder_DecodeUnknownGroup(upb_Decoder* d,
const char* ptr,
uint32_t number) {
return _upb_Decoder_DecodeGroup(d, ptr, NULL, NULL, number);
}
UPB_FORCEINLINE
static const char* _upb_Decoder_DecodeKnownGroup(
upb_Decoder* d, const char* ptr, upb_Message* submsg,
const upb_MiniTableSub* subs, const upb_MiniTableField* field) {
const upb_MiniTable* subl = _upb_MiniTableSubs_MessageByField(subs, field);
UPB_ASSERT(subl);
return _upb_Decoder_DecodeGroup(d, ptr, submsg, subl,
field->UPB_PRIVATE(number));
}
static char* upb_Decoder_EncodeVarint32(uint32_t val, char* ptr) {
do {
uint8_t byte = val & 0x7fU;
val >>= 7;
if (val) byte |= 0x80U;
*(ptr++) = byte;
} while (val);
return ptr;
}
static void _upb_Decoder_AddUnknownVarints(upb_Decoder* d, upb_Message* msg,
uint32_t val1, uint32_t val2) {
char buf[20];
char* end = buf;
end = upb_Decoder_EncodeVarint32(val1, end);
end = upb_Decoder_EncodeVarint32(val2, end);
if (!UPB_PRIVATE(_upb_Message_AddUnknown)(msg, buf, end - buf, &d->arena)) {
_upb_Decoder_ErrorJmp(d, kUpb_DecodeStatus_OutOfMemory);
}
}
UPB_FORCEINLINE
static bool _upb_Decoder_CheckEnum(upb_Decoder* d, const char* ptr,
upb_Message* msg, const upb_MiniTableEnum* e,
const upb_MiniTableField* field,
wireval* val) {
const uint32_t v = val->uint32_val;
if (UPB_LIKELY(upb_MiniTableEnum_CheckValue(e, v))) return true;
// Unrecognized enum goes into unknown fields.
// For packed fields the tag could be arbitrarily far in the past,
// so we just re-encode the tag and value here.
const uint32_t tag =
((uint32_t)field->UPB_PRIVATE(number) << 3) | kUpb_WireType_Varint;
upb_Message* unknown_msg =
field->UPB_PRIVATE(mode) & kUpb_LabelFlags_IsExtension ? d->unknown_msg
: msg;
_upb_Decoder_AddUnknownVarints(d, unknown_msg, tag, v);
return false;
}
UPB_NOINLINE
static const char* _upb_Decoder_DecodeEnumArray(upb_Decoder* d, const char* ptr,
upb_Message* msg,
upb_Array* arr,
const upb_MiniTableSub* subs,
const upb_MiniTableField* field,
wireval* val) {
const upb_MiniTableEnum* e = _upb_MiniTableSubs_EnumByField(subs, field);
if (!_upb_Decoder_CheckEnum(d, ptr, msg, e, field, val)) return ptr;
void* mem = UPB_PTR_AT(_upb_array_ptr(arr), arr->UPB_PRIVATE(size) * 4, void);
arr->UPB_PRIVATE(size)++;
memcpy(mem, val, 4);
return ptr;
}
UPB_FORCEINLINE
static const char* _upb_Decoder_DecodeFixedPacked(
upb_Decoder* d, const char* ptr, upb_Array* arr, wireval* val,
const upb_MiniTableField* field, int lg2) {
int mask = (1 << lg2) - 1;
size_t count = val->size >> lg2;
if ((val->size & mask) != 0) {
// Length isn't a round multiple of elem size.
_upb_Decoder_ErrorJmp(d, kUpb_DecodeStatus_Malformed);
}
_upb_Decoder_Reserve(d, arr, count);
void* mem =
UPB_PTR_AT(_upb_array_ptr(arr), arr->UPB_PRIVATE(size) << lg2, void);
arr->UPB_PRIVATE(size) += count;
// Note: if/when the decoder supports multi-buffer input, we will need to
// handle buffer seams here.
if (_upb_IsLittleEndian()) {
ptr = upb_EpsCopyInputStream_Copy(&d->input, ptr, mem, val->size);
} else {
int delta = upb_EpsCopyInputStream_PushLimit(&d->input, ptr, val->size);
char* dst = mem;
while (!_upb_Decoder_IsDone(d, &ptr)) {
if (lg2 == 2) {
ptr = upb_WireReader_ReadFixed32(ptr, dst);
dst += 4;
} else {
UPB_ASSERT(lg2 == 3);
ptr = upb_WireReader_ReadFixed64(ptr, dst);
dst += 8;
}
}
upb_EpsCopyInputStream_PopLimit(&d->input, ptr, delta);
}
return ptr;
}
UPB_FORCEINLINE
static const char* _upb_Decoder_DecodeVarintPacked(
upb_Decoder* d, const char* ptr, upb_Array* arr, wireval* val,
const upb_MiniTableField* field, int lg2) {
int scale = 1 << lg2;
int saved_limit = upb_EpsCopyInputStream_PushLimit(&d->input, ptr, val->size);
char* out =
UPB_PTR_AT(_upb_array_ptr(arr), arr->UPB_PRIVATE(size) << lg2, void);
while (!_upb_Decoder_IsDone(d, &ptr)) {
wireval elem;
ptr = _upb_Decoder_DecodeVarint(d, ptr, &elem.uint64_val);
_upb_Decoder_Munge(field->UPB_PRIVATE(descriptortype), &elem);
if (_upb_Decoder_Reserve(d, arr, 1)) {
out =
UPB_PTR_AT(_upb_array_ptr(arr), arr->UPB_PRIVATE(size) << lg2, void);
}
arr->UPB_PRIVATE(size)++;
memcpy(out, &elem, scale);
out += scale;
}
upb_EpsCopyInputStream_PopLimit(&d->input, ptr, saved_limit);
return ptr;
}
UPB_NOINLINE
static const char* _upb_Decoder_DecodeEnumPacked(
upb_Decoder* d, const char* ptr, upb_Message* msg, upb_Array* arr,
const upb_MiniTableSub* subs, const upb_MiniTableField* field,
wireval* val) {
const upb_MiniTableEnum* e = _upb_MiniTableSubs_EnumByField(subs, field);
int saved_limit = upb_EpsCopyInputStream_PushLimit(&d->input, ptr, val->size);
char* out = UPB_PTR_AT(_upb_array_ptr(arr), arr->UPB_PRIVATE(size) * 4, void);
while (!_upb_Decoder_IsDone(d, &ptr)) {
wireval elem;
ptr = _upb_Decoder_DecodeVarint(d, ptr, &elem.uint64_val);
_upb_Decoder_MungeInt32(&elem);
if (!_upb_Decoder_CheckEnum(d, ptr, msg, e, field, &elem)) {
continue;
}
if (_upb_Decoder_Reserve(d, arr, 1)) {
out = UPB_PTR_AT(_upb_array_ptr(arr), arr->UPB_PRIVATE(size) * 4, void);
}
arr->UPB_PRIVATE(size)++;
memcpy(out, &elem, 4);
out += 4;
}
upb_EpsCopyInputStream_PopLimit(&d->input, ptr, saved_limit);
return ptr;
}
upb_Array* _upb_Decoder_CreateArray(upb_Decoder* d,
const upb_MiniTableField* field) {
const upb_FieldType field_type = field->UPB_PRIVATE(descriptortype);
const size_t lg2 = UPB_PRIVATE(_upb_FieldType_SizeLg2)(field_type);
upb_Array* ret = UPB_PRIVATE(_upb_Array_New)(&d->arena, 4, lg2);
if (!ret) _upb_Decoder_ErrorJmp(d, kUpb_DecodeStatus_OutOfMemory);
return ret;
}
static const char* _upb_Decoder_DecodeToArray(upb_Decoder* d, const char* ptr,
upb_Message* msg,
const upb_MiniTableSub* subs,
const upb_MiniTableField* field,
wireval* val, int op) {
upb_Array** arrp = UPB_PTR_AT(msg, field->UPB_PRIVATE(offset), void);
upb_Array* arr = *arrp;
void* mem;
if (arr) {
_upb_Decoder_Reserve(d, arr, 1);
} else {
arr = _upb_Decoder_CreateArray(d, field);
*arrp = arr;
}
switch (op) {
case kUpb_DecodeOp_Scalar1Byte:
case kUpb_DecodeOp_Scalar4Byte:
case kUpb_DecodeOp_Scalar8Byte:
/* Append scalar value. */
mem = UPB_PTR_AT(_upb_array_ptr(arr), arr->UPB_PRIVATE(size) << op, void);
arr->UPB_PRIVATE(size)++;
Optimized decoder and paved the way for parsing extensions. The primary motivation for this change is to avoid referring to the `upb_msglayout` object when we are trying to fetch the `upb_msglayout` object for a sub-message. This will help pave the way for parsing extensions. We also implement several optimizations so that we can make this change without regressing performance. Normally we compute the layout for a sub-message field like so: ``` const upb_msglayout *get_submsg_layout( const upb_msglayout *layout, const upb_msglayout_field *field) { return layout->submsgs[field->submsg_index] } ``` The reason for this indirection is to avoid storing a pointer directly in `upb_msglayout_field`, as this would double its size (from 12 to 24 bytes on 64-bit architectures) which is wasteful as this pointer is only needed for message typed fields. However `get_submsg_layout` as written above does not work for extensions, as they will not have entries in the message's `layout->submsgs` array by nature, and we want to avoid creating an entire fake `upb_msglayout` for each such extension since that would also be wasteful. This change removes the dependency on `upb_msglayout` by passing down the `submsgs` array instead: ``` const upb_msglayout *get_submsg_layout( const upb_msglayout *const *submsgs, const upb_msglayout_field *field) { return submsgs[field->submsg_index] } ``` This will pave the way for parsing extensions, as we can more easily create an alternative `submsgs` array for extension fields without extra overhead or waste. Along the way several optimizations presented themselves that allow a nice increase in performance: 1. Passing the parsed `wireval` by address instead of by value ended up avoiding an expensive and useless stack copy (this is on Clang, which was used for all measurements). 2. When field numbers are densely packed, we can find a field by number with a single indexed lookup instead of linear search. At codegen time we can compute the maximum field number that will allow such an indexed lookup. 3. For fields that do require linear search, we can start the linear search at the location where we found the previous field, taking advantage of the fact that field numbers are generally increasing. 4. When the hasbit index is less than 32 (the common case) we can use a less expensive code sequence to set it. 5. We check for the hasbit case before the oneof case, as optional fields are more common than oneof fields. Benchmark results indicate a 20% improvement in parse speed with a small code size increase: ``` name old time/op new time/op delta ArenaOneAlloc 21.3ns ± 0% 21.5ns ± 0% +0.96% (p=0.000 n=12+12) ArenaInitialBlockOneAlloc 6.32ns ± 0% 6.32ns ± 0% +0.03% (p=0.000 n=12+10) LoadDescriptor_Upb 53.5µs ± 1% 51.5µs ± 2% -3.70% (p=0.000 n=12+12) LoadAdsDescriptor_Upb 2.78ms ± 2% 2.68ms ± 0% -3.57% (p=0.000 n=12+12) LoadDescriptor_Proto2 240µs ± 0% 240µs ± 0% +0.12% (p=0.001 n=12+12) LoadAdsDescriptor_Proto2 12.8ms ± 0% 12.7ms ± 0% -1.15% (p=0.000 n=12+10) Parse_Upb_FileDesc<UseArena,Copy> 13.2µs ± 2% 10.7µs ± 0% -18.49% (p=0.000 n=10+12) Parse_Upb_FileDesc<UseArena,Alias> 11.3µs ± 0% 9.6µs ± 0% -15.11% (p=0.000 n=12+11) Parse_Upb_FileDesc<InitBlock,Copy> 12.7µs ± 0% 10.3µs ± 0% -19.00% (p=0.000 n=10+12) Parse_Upb_FileDesc<InitBlock,Alias> 10.9µs ± 0% 9.2µs ± 0% -15.82% (p=0.000 n=12+12) Parse_Proto2<FileDesc,NoArena,Copy> 29.4µs ± 0% 29.5µs ± 0% +0.61% (p=0.000 n=12+12) Parse_Proto2<FileDesc,UseArena,Copy> 20.7µs ± 2% 20.6µs ± 2% ~ (p=0.260 n=12+11) Parse_Proto2<FileDesc,InitBlock,Copy> 16.7µs ± 1% 16.7µs ± 0% -0.25% (p=0.036 n=12+10) Parse_Proto2<FileDescSV,InitBlock,Alias> 16.5µs ± 0% 16.5µs ± 0% +0.20% (p=0.016 n=12+11) SerializeDescriptor_Proto2 5.30µs ± 1% 5.36µs ± 1% +1.09% (p=0.000 n=12+11) SerializeDescriptor_Upb 12.9µs ± 0% 13.0µs ± 0% +0.90% (p=0.000 n=12+11) FILE SIZE VM SIZE -------------- -------------- +1.5% +176 +1.6% +176 upb/decode.c +1.8% +176 +1.9% +176 decode_msg +0.4% +64 +0.4% +64 upb/def.c +1.4% +64 +1.4% +64 _upb_symtab_addfile +1.2% +48 +1.4% +48 upb/reflection.c +15% +32 +18% +32 upb_msg_set +2.9% +16 +3.1% +16 upb_msg_mutable -9.3% -288 [ = ] 0 [Unmapped] [ = ] 0 +0.2% +288 TOTAL ```
4 years ago
memcpy(mem, val, 1 << op);
return ptr;
case kUpb_DecodeOp_String:
_upb_Decoder_VerifyUtf8(d, ptr, val->size);
/* Fallthrough. */
case kUpb_DecodeOp_Bytes: {
/* Append bytes. */
upb_StringView* str =
(upb_StringView*)_upb_array_ptr(arr) + arr->UPB_PRIVATE(size);
arr->UPB_PRIVATE(size)++;
return _upb_Decoder_ReadString(d, ptr, val->size, str);
}
case kUpb_DecodeOp_SubMessage: {
/* Append submessage / group. */
Added a new dynamic tree shaking model to upb, with the intention of removing the old model once YouTube has migrated. The `kUpb_DecodeOption_ExperimentalAllowUnlinked` flag to the decoder will enable the new behavior. When that flag is not passed, tree shaking with the old model will still be possible. "Dynamic tree shaking" in upb is a feature that allows messages to be parsed even if the MiniTables have not been fully linked. Unlinked sub-message fields can be parsed by preserving their data in the unknown fields. If the application later discovers that the message field is actually needed, the MiniTable can be patched to properly link that field, and existing message instances can "promote" the data from the unknown fields to an actual message of the correct type. Before this change, dynamic tree shaking stored unparsed message data in the unknown fields of the *parent*. In effect, we were treating the field as if it did not exist at all. This meant that parsing an unlinked field did not affect the hasbits or oneof cases of the parent, nor did it create a `upb_Array` or `upb_Map` for array/map fields. Only when a message was linked and promoted did any of these things occur. While this model had some amount of conceptual simplicity, it caused significant problems with oneofs. When multiple fields inside a single oneof are parsed from the wire, order matters, because later oneof fields must overwrite earlier ones. Dynamic tree shaking can mean that some fields in a oneof are linked while others are not. It is essential that we preserve this ordering semantic even when dynamic tree shaking is being used, but it is difficult to do if the oneof's data can be split between linked fields (which have been reified into parsed field data) and unlinked fields (whose data lives in the unknown fields of the parent). To solve this problem, this CL changes the representation for unlinked fields. Instead of being placed in the parent's unknown fields, we create an actual message instance for each unlinked message we parse, but we use a placeholder "empty message" MiniTable as the message's type. All of the message's data will therefore be placed into the "empty message's" unknown fields. But unlike before, this "empty message" is actually present according to the hasbits, oneof case, and `upb_Array`/`upb_Map` of the parent. This means that all of the oneof presence logic works as normal. Since the MiniTable can be patched at any time, we need a bit in the message instance itself to signal whether a pointer to a sub-message is an "empty message" or not. When dynamic tree shaking is in use, all users must be capable of recognizing an empty message and acting accordingly (promoting, etc) even if the MiniTable itself says that the field is linked. Because dynamic tree shaking imposes this extra requirement on users, we require that users pass an extra option to the decoder to allow parsing of unlinked sub-messages. Many existing users of upb (Ruby, PHP, Python, etc) will always have fully-linked MiniTables, so there is no reason for them to add extra logic to handle empty messages. By omitting the `kUpb_DecodeOption_ExperimentalAllowUnlinked` option, they will be relieved of the duty to check the tagged pointer that would indicate an empty, unlinked message. For existing users of dynamic tree shaking, there are three main changes: 1. The APIs in message/promote.h have changed, and users will need to update to the new interfaces. 2. The model for maps has changed slightly. Before, we required that map entries always had their values linked; for dynamic tree shaking to apply to maps, we required that the *entry* was left unlinked, not the entry's value. In the new model, that is reversed: map entries must always be linked, but a map entry's value can be unlinked. 3. The presence model for unlinked fields has changed. Unlinked fields will now register as "present" from the perspective of hasbits, oneof cases, and array/map entries. Users must test the tagged pointer to know if a message is of the correct, linked type or whether it is a placeholder "empty" message. There is a new function `upb_Message_GetTaggedMessagePtr()`, as well as a new accessor `upb_MessageValue.tagged_msg_val` that can be used to read and test the tagged pointer directly. PiperOrigin-RevId: 535288031
2 years ago
upb_TaggedMessagePtr* target = UPB_PTR_AT(
_upb_array_ptr(arr), arr->UPB_PRIVATE(size) * sizeof(void*),
upb_TaggedMessagePtr);
Added a new dynamic tree shaking model to upb, with the intention of removing the old model once YouTube has migrated. The `kUpb_DecodeOption_ExperimentalAllowUnlinked` flag to the decoder will enable the new behavior. When that flag is not passed, tree shaking with the old model will still be possible. "Dynamic tree shaking" in upb is a feature that allows messages to be parsed even if the MiniTables have not been fully linked. Unlinked sub-message fields can be parsed by preserving their data in the unknown fields. If the application later discovers that the message field is actually needed, the MiniTable can be patched to properly link that field, and existing message instances can "promote" the data from the unknown fields to an actual message of the correct type. Before this change, dynamic tree shaking stored unparsed message data in the unknown fields of the *parent*. In effect, we were treating the field as if it did not exist at all. This meant that parsing an unlinked field did not affect the hasbits or oneof cases of the parent, nor did it create a `upb_Array` or `upb_Map` for array/map fields. Only when a message was linked and promoted did any of these things occur. While this model had some amount of conceptual simplicity, it caused significant problems with oneofs. When multiple fields inside a single oneof are parsed from the wire, order matters, because later oneof fields must overwrite earlier ones. Dynamic tree shaking can mean that some fields in a oneof are linked while others are not. It is essential that we preserve this ordering semantic even when dynamic tree shaking is being used, but it is difficult to do if the oneof's data can be split between linked fields (which have been reified into parsed field data) and unlinked fields (whose data lives in the unknown fields of the parent). To solve this problem, this CL changes the representation for unlinked fields. Instead of being placed in the parent's unknown fields, we create an actual message instance for each unlinked message we parse, but we use a placeholder "empty message" MiniTable as the message's type. All of the message's data will therefore be placed into the "empty message's" unknown fields. But unlike before, this "empty message" is actually present according to the hasbits, oneof case, and `upb_Array`/`upb_Map` of the parent. This means that all of the oneof presence logic works as normal. Since the MiniTable can be patched at any time, we need a bit in the message instance itself to signal whether a pointer to a sub-message is an "empty message" or not. When dynamic tree shaking is in use, all users must be capable of recognizing an empty message and acting accordingly (promoting, etc) even if the MiniTable itself says that the field is linked. Because dynamic tree shaking imposes this extra requirement on users, we require that users pass an extra option to the decoder to allow parsing of unlinked sub-messages. Many existing users of upb (Ruby, PHP, Python, etc) will always have fully-linked MiniTables, so there is no reason for them to add extra logic to handle empty messages. By omitting the `kUpb_DecodeOption_ExperimentalAllowUnlinked` option, they will be relieved of the duty to check the tagged pointer that would indicate an empty, unlinked message. For existing users of dynamic tree shaking, there are three main changes: 1. The APIs in message/promote.h have changed, and users will need to update to the new interfaces. 2. The model for maps has changed slightly. Before, we required that map entries always had their values linked; for dynamic tree shaking to apply to maps, we required that the *entry* was left unlinked, not the entry's value. In the new model, that is reversed: map entries must always be linked, but a map entry's value can be unlinked. 3. The presence model for unlinked fields has changed. Unlinked fields will now register as "present" from the perspective of hasbits, oneof cases, and array/map entries. Users must test the tagged pointer to know if a message is of the correct, linked type or whether it is a placeholder "empty" message. There is a new function `upb_Message_GetTaggedMessagePtr()`, as well as a new accessor `upb_MessageValue.tagged_msg_val` that can be used to read and test the tagged pointer directly. PiperOrigin-RevId: 535288031
2 years ago
upb_Message* submsg = _upb_Decoder_NewSubMessage(d, subs, field, target);
arr->UPB_PRIVATE(size)++;
if (UPB_UNLIKELY(field->UPB_PRIVATE(descriptortype) ==
kUpb_FieldType_Group)) {
return _upb_Decoder_DecodeKnownGroup(d, ptr, submsg, subs, field);
} else {
return _upb_Decoder_DecodeSubMessage(d, ptr, submsg, subs, field,
val->size);
}
}
case OP_FIXPCK_LG2(2):
case OP_FIXPCK_LG2(3):
return _upb_Decoder_DecodeFixedPacked(d, ptr, arr, val, field,
op - OP_FIXPCK_LG2(0));
case OP_VARPCK_LG2(0):
case OP_VARPCK_LG2(2):
case OP_VARPCK_LG2(3):
return _upb_Decoder_DecodeVarintPacked(d, ptr, arr, val, field,
op - OP_VARPCK_LG2(0));
case kUpb_DecodeOp_Enum:
return _upb_Decoder_DecodeEnumArray(d, ptr, msg, arr, subs, field, val);
case kUpb_DecodeOp_PackedEnum:
return _upb_Decoder_DecodeEnumPacked(d, ptr, msg, arr, subs, field, val);
default:
UPB_UNREACHABLE();
}
}
upb_Map* _upb_Decoder_CreateMap(upb_Decoder* d, const upb_MiniTable* entry) {
/* Maps descriptor type -> upb map size. */
static const uint8_t kSizeInMap[] = {
[0] = -1, // invalid descriptor type */
[kUpb_FieldType_Double] = 8,
[kUpb_FieldType_Float] = 4,
[kUpb_FieldType_Int64] = 8,
[kUpb_FieldType_UInt64] = 8,
[kUpb_FieldType_Int32] = 4,
[kUpb_FieldType_Fixed64] = 8,
[kUpb_FieldType_Fixed32] = 4,
[kUpb_FieldType_Bool] = 1,
[kUpb_FieldType_String] = UPB_MAPTYPE_STRING,
[kUpb_FieldType_Group] = sizeof(void*),
[kUpb_FieldType_Message] = sizeof(void*),
[kUpb_FieldType_Bytes] = UPB_MAPTYPE_STRING,
[kUpb_FieldType_UInt32] = 4,
[kUpb_FieldType_Enum] = 4,
[kUpb_FieldType_SFixed32] = 4,
[kUpb_FieldType_SFixed64] = 8,
[kUpb_FieldType_SInt32] = 4,
[kUpb_FieldType_SInt64] = 8,
};
const upb_MiniTableField* key_field = &entry->UPB_PRIVATE(fields)[0];
const upb_MiniTableField* val_field = &entry->UPB_PRIVATE(fields)[1];
char key_size = kSizeInMap[key_field->UPB_PRIVATE(descriptortype)];
char val_size = kSizeInMap[val_field->UPB_PRIVATE(descriptortype)];
UPB_ASSERT(key_field->UPB_PRIVATE(offset) == offsetof(upb_MapEntryData, k));
UPB_ASSERT(val_field->UPB_PRIVATE(offset) == offsetof(upb_MapEntryData, v));
upb_Map* ret = _upb_Map_New(&d->arena, key_size, val_size);
if (!ret) _upb_Decoder_ErrorJmp(d, kUpb_DecodeStatus_OutOfMemory);
return ret;
}
static const char* _upb_Decoder_DecodeToMap(upb_Decoder* d, const char* ptr,
upb_Message* msg,
const upb_MiniTableSub* subs,
const upb_MiniTableField* field,
wireval* val) {
upb_Map** map_p = UPB_PTR_AT(msg, field->UPB_PRIVATE(offset), upb_Map*);
upb_Map* map = *map_p;
upb_MapEntry ent;
UPB_ASSERT(upb_MiniTableField_Type(field) == kUpb_FieldType_Message);
const upb_MiniTable* entry = _upb_MiniTableSubs_MessageByField(subs, field);
Added a new dynamic tree shaking model to upb, with the intention of removing the old model once YouTube has migrated. The `kUpb_DecodeOption_ExperimentalAllowUnlinked` flag to the decoder will enable the new behavior. When that flag is not passed, tree shaking with the old model will still be possible. "Dynamic tree shaking" in upb is a feature that allows messages to be parsed even if the MiniTables have not been fully linked. Unlinked sub-message fields can be parsed by preserving their data in the unknown fields. If the application later discovers that the message field is actually needed, the MiniTable can be patched to properly link that field, and existing message instances can "promote" the data from the unknown fields to an actual message of the correct type. Before this change, dynamic tree shaking stored unparsed message data in the unknown fields of the *parent*. In effect, we were treating the field as if it did not exist at all. This meant that parsing an unlinked field did not affect the hasbits or oneof cases of the parent, nor did it create a `upb_Array` or `upb_Map` for array/map fields. Only when a message was linked and promoted did any of these things occur. While this model had some amount of conceptual simplicity, it caused significant problems with oneofs. When multiple fields inside a single oneof are parsed from the wire, order matters, because later oneof fields must overwrite earlier ones. Dynamic tree shaking can mean that some fields in a oneof are linked while others are not. It is essential that we preserve this ordering semantic even when dynamic tree shaking is being used, but it is difficult to do if the oneof's data can be split between linked fields (which have been reified into parsed field data) and unlinked fields (whose data lives in the unknown fields of the parent). To solve this problem, this CL changes the representation for unlinked fields. Instead of being placed in the parent's unknown fields, we create an actual message instance for each unlinked message we parse, but we use a placeholder "empty message" MiniTable as the message's type. All of the message's data will therefore be placed into the "empty message's" unknown fields. But unlike before, this "empty message" is actually present according to the hasbits, oneof case, and `upb_Array`/`upb_Map` of the parent. This means that all of the oneof presence logic works as normal. Since the MiniTable can be patched at any time, we need a bit in the message instance itself to signal whether a pointer to a sub-message is an "empty message" or not. When dynamic tree shaking is in use, all users must be capable of recognizing an empty message and acting accordingly (promoting, etc) even if the MiniTable itself says that the field is linked. Because dynamic tree shaking imposes this extra requirement on users, we require that users pass an extra option to the decoder to allow parsing of unlinked sub-messages. Many existing users of upb (Ruby, PHP, Python, etc) will always have fully-linked MiniTables, so there is no reason for them to add extra logic to handle empty messages. By omitting the `kUpb_DecodeOption_ExperimentalAllowUnlinked` option, they will be relieved of the duty to check the tagged pointer that would indicate an empty, unlinked message. For existing users of dynamic tree shaking, there are three main changes: 1. The APIs in message/promote.h have changed, and users will need to update to the new interfaces. 2. The model for maps has changed slightly. Before, we required that map entries always had their values linked; for dynamic tree shaking to apply to maps, we required that the *entry* was left unlinked, not the entry's value. In the new model, that is reversed: map entries must always be linked, but a map entry's value can be unlinked. 3. The presence model for unlinked fields has changed. Unlinked fields will now register as "present" from the perspective of hasbits, oneof cases, and array/map entries. Users must test the tagged pointer to know if a message is of the correct, linked type or whether it is a placeholder "empty" message. There is a new function `upb_Message_GetTaggedMessagePtr()`, as well as a new accessor `upb_MessageValue.tagged_msg_val` that can be used to read and test the tagged pointer directly. PiperOrigin-RevId: 535288031
2 years ago
UPB_ASSERT(entry);
UPB_ASSERT(entry->UPB_PRIVATE(field_count) == 2);
UPB_ASSERT(upb_MiniTableField_IsScalar(&entry->UPB_PRIVATE(fields)[0]));
UPB_ASSERT(upb_MiniTableField_IsScalar(&entry->UPB_PRIVATE(fields)[1]));
if (!map) {
map = _upb_Decoder_CreateMap(d, entry);
*map_p = map;
}
// Parse map entry.
memset(&ent, 0, sizeof(ent));
if (entry->UPB_PRIVATE(fields)[1].UPB_PRIVATE(descriptortype) ==
kUpb_FieldType_Message ||
entry->UPB_PRIVATE(fields)[1].UPB_PRIVATE(descriptortype) ==
kUpb_FieldType_Group) {
Added a new dynamic tree shaking model to upb, with the intention of removing the old model once YouTube has migrated. The `kUpb_DecodeOption_ExperimentalAllowUnlinked` flag to the decoder will enable the new behavior. When that flag is not passed, tree shaking with the old model will still be possible. "Dynamic tree shaking" in upb is a feature that allows messages to be parsed even if the MiniTables have not been fully linked. Unlinked sub-message fields can be parsed by preserving their data in the unknown fields. If the application later discovers that the message field is actually needed, the MiniTable can be patched to properly link that field, and existing message instances can "promote" the data from the unknown fields to an actual message of the correct type. Before this change, dynamic tree shaking stored unparsed message data in the unknown fields of the *parent*. In effect, we were treating the field as if it did not exist at all. This meant that parsing an unlinked field did not affect the hasbits or oneof cases of the parent, nor did it create a `upb_Array` or `upb_Map` for array/map fields. Only when a message was linked and promoted did any of these things occur. While this model had some amount of conceptual simplicity, it caused significant problems with oneofs. When multiple fields inside a single oneof are parsed from the wire, order matters, because later oneof fields must overwrite earlier ones. Dynamic tree shaking can mean that some fields in a oneof are linked while others are not. It is essential that we preserve this ordering semantic even when dynamic tree shaking is being used, but it is difficult to do if the oneof's data can be split between linked fields (which have been reified into parsed field data) and unlinked fields (whose data lives in the unknown fields of the parent). To solve this problem, this CL changes the representation for unlinked fields. Instead of being placed in the parent's unknown fields, we create an actual message instance for each unlinked message we parse, but we use a placeholder "empty message" MiniTable as the message's type. All of the message's data will therefore be placed into the "empty message's" unknown fields. But unlike before, this "empty message" is actually present according to the hasbits, oneof case, and `upb_Array`/`upb_Map` of the parent. This means that all of the oneof presence logic works as normal. Since the MiniTable can be patched at any time, we need a bit in the message instance itself to signal whether a pointer to a sub-message is an "empty message" or not. When dynamic tree shaking is in use, all users must be capable of recognizing an empty message and acting accordingly (promoting, etc) even if the MiniTable itself says that the field is linked. Because dynamic tree shaking imposes this extra requirement on users, we require that users pass an extra option to the decoder to allow parsing of unlinked sub-messages. Many existing users of upb (Ruby, PHP, Python, etc) will always have fully-linked MiniTables, so there is no reason for them to add extra logic to handle empty messages. By omitting the `kUpb_DecodeOption_ExperimentalAllowUnlinked` option, they will be relieved of the duty to check the tagged pointer that would indicate an empty, unlinked message. For existing users of dynamic tree shaking, there are three main changes: 1. The APIs in message/promote.h have changed, and users will need to update to the new interfaces. 2. The model for maps has changed slightly. Before, we required that map entries always had their values linked; for dynamic tree shaking to apply to maps, we required that the *entry* was left unlinked, not the entry's value. In the new model, that is reversed: map entries must always be linked, but a map entry's value can be unlinked. 3. The presence model for unlinked fields has changed. Unlinked fields will now register as "present" from the perspective of hasbits, oneof cases, and array/map entries. Users must test the tagged pointer to know if a message is of the correct, linked type or whether it is a placeholder "empty" message. There is a new function `upb_Message_GetTaggedMessagePtr()`, as well as a new accessor `upb_MessageValue.tagged_msg_val` that can be used to read and test the tagged pointer directly. PiperOrigin-RevId: 535288031
2 years ago
// Create proactively to handle the case where it doesn't appear.
upb_TaggedMessagePtr msg;
_upb_Decoder_NewSubMessage(d, entry->UPB_PRIVATE(subs),
&entry->UPB_PRIVATE(fields)[1], &msg);
Added a new dynamic tree shaking model to upb, with the intention of removing the old model once YouTube has migrated. The `kUpb_DecodeOption_ExperimentalAllowUnlinked` flag to the decoder will enable the new behavior. When that flag is not passed, tree shaking with the old model will still be possible. "Dynamic tree shaking" in upb is a feature that allows messages to be parsed even if the MiniTables have not been fully linked. Unlinked sub-message fields can be parsed by preserving their data in the unknown fields. If the application later discovers that the message field is actually needed, the MiniTable can be patched to properly link that field, and existing message instances can "promote" the data from the unknown fields to an actual message of the correct type. Before this change, dynamic tree shaking stored unparsed message data in the unknown fields of the *parent*. In effect, we were treating the field as if it did not exist at all. This meant that parsing an unlinked field did not affect the hasbits or oneof cases of the parent, nor did it create a `upb_Array` or `upb_Map` for array/map fields. Only when a message was linked and promoted did any of these things occur. While this model had some amount of conceptual simplicity, it caused significant problems with oneofs. When multiple fields inside a single oneof are parsed from the wire, order matters, because later oneof fields must overwrite earlier ones. Dynamic tree shaking can mean that some fields in a oneof are linked while others are not. It is essential that we preserve this ordering semantic even when dynamic tree shaking is being used, but it is difficult to do if the oneof's data can be split between linked fields (which have been reified into parsed field data) and unlinked fields (whose data lives in the unknown fields of the parent). To solve this problem, this CL changes the representation for unlinked fields. Instead of being placed in the parent's unknown fields, we create an actual message instance for each unlinked message we parse, but we use a placeholder "empty message" MiniTable as the message's type. All of the message's data will therefore be placed into the "empty message's" unknown fields. But unlike before, this "empty message" is actually present according to the hasbits, oneof case, and `upb_Array`/`upb_Map` of the parent. This means that all of the oneof presence logic works as normal. Since the MiniTable can be patched at any time, we need a bit in the message instance itself to signal whether a pointer to a sub-message is an "empty message" or not. When dynamic tree shaking is in use, all users must be capable of recognizing an empty message and acting accordingly (promoting, etc) even if the MiniTable itself says that the field is linked. Because dynamic tree shaking imposes this extra requirement on users, we require that users pass an extra option to the decoder to allow parsing of unlinked sub-messages. Many existing users of upb (Ruby, PHP, Python, etc) will always have fully-linked MiniTables, so there is no reason for them to add extra logic to handle empty messages. By omitting the `kUpb_DecodeOption_ExperimentalAllowUnlinked` option, they will be relieved of the duty to check the tagged pointer that would indicate an empty, unlinked message. For existing users of dynamic tree shaking, there are three main changes: 1. The APIs in message/promote.h have changed, and users will need to update to the new interfaces. 2. The model for maps has changed slightly. Before, we required that map entries always had their values linked; for dynamic tree shaking to apply to maps, we required that the *entry* was left unlinked, not the entry's value. In the new model, that is reversed: map entries must always be linked, but a map entry's value can be unlinked. 3. The presence model for unlinked fields has changed. Unlinked fields will now register as "present" from the perspective of hasbits, oneof cases, and array/map entries. Users must test the tagged pointer to know if a message is of the correct, linked type or whether it is a placeholder "empty" message. There is a new function `upb_Message_GetTaggedMessagePtr()`, as well as a new accessor `upb_MessageValue.tagged_msg_val` that can be used to read and test the tagged pointer directly. PiperOrigin-RevId: 535288031
2 years ago
ent.data.v.val = upb_value_uintptr(msg);
}
ptr = _upb_Decoder_DecodeSubMessage(d, ptr, (upb_Message*)&ent.data, subs,
field, val->size);
3 years ago
// check if ent had any unknown fields
size_t size;
upb_Message_GetUnknown((upb_Message*)&ent.data, &size);
3 years ago
if (size != 0) {
char* buf;
size_t size;
uint32_t tag =
((uint32_t)field->UPB_PRIVATE(number) << 3) | kUpb_WireType_Delimited;
upb_EncodeStatus status =
upb_Encode((upb_Message*)&ent.data, entry, 0, &d->arena, &buf, &size);
if (status != kUpb_EncodeStatus_Ok) {
_upb_Decoder_ErrorJmp(d, kUpb_DecodeStatus_OutOfMemory);
}
_upb_Decoder_AddUnknownVarints(d, msg, tag, size);
if (!UPB_PRIVATE(_upb_Message_AddUnknown)(msg, buf, size, &d->arena)) {
_upb_Decoder_ErrorJmp(d, kUpb_DecodeStatus_OutOfMemory);
}
3 years ago
} else {
if (_upb_Map_Insert(map, &ent.data.k, map->key_size, &ent.data.v,
map->val_size,
&d->arena) == kUpb_MapInsertStatus_OutOfMemory) {
_upb_Decoder_ErrorJmp(d, kUpb_DecodeStatus_OutOfMemory);
}
3 years ago
}
return ptr;
}
static const char* _upb_Decoder_DecodeToSubMessage(
upb_Decoder* d, const char* ptr, upb_Message* msg,
const upb_MiniTableSub* subs, const upb_MiniTableField* field, wireval* val,
int op) {
void* mem = UPB_PTR_AT(msg, field->UPB_PRIVATE(offset), void);
int type = field->UPB_PRIVATE(descriptortype);
if (UPB_UNLIKELY(op == kUpb_DecodeOp_Enum) &&
!_upb_Decoder_CheckEnum(d, ptr, msg,
_upb_MiniTableSubs_EnumByField(subs, field),
field, val)) {
return ptr;
}
/* Set presence if necessary. */
Optimized decoder and paved the way for parsing extensions. The primary motivation for this change is to avoid referring to the `upb_msglayout` object when we are trying to fetch the `upb_msglayout` object for a sub-message. This will help pave the way for parsing extensions. We also implement several optimizations so that we can make this change without regressing performance. Normally we compute the layout for a sub-message field like so: ``` const upb_msglayout *get_submsg_layout( const upb_msglayout *layout, const upb_msglayout_field *field) { return layout->submsgs[field->submsg_index] } ``` The reason for this indirection is to avoid storing a pointer directly in `upb_msglayout_field`, as this would double its size (from 12 to 24 bytes on 64-bit architectures) which is wasteful as this pointer is only needed for message typed fields. However `get_submsg_layout` as written above does not work for extensions, as they will not have entries in the message's `layout->submsgs` array by nature, and we want to avoid creating an entire fake `upb_msglayout` for each such extension since that would also be wasteful. This change removes the dependency on `upb_msglayout` by passing down the `submsgs` array instead: ``` const upb_msglayout *get_submsg_layout( const upb_msglayout *const *submsgs, const upb_msglayout_field *field) { return submsgs[field->submsg_index] } ``` This will pave the way for parsing extensions, as we can more easily create an alternative `submsgs` array for extension fields without extra overhead or waste. Along the way several optimizations presented themselves that allow a nice increase in performance: 1. Passing the parsed `wireval` by address instead of by value ended up avoiding an expensive and useless stack copy (this is on Clang, which was used for all measurements). 2. When field numbers are densely packed, we can find a field by number with a single indexed lookup instead of linear search. At codegen time we can compute the maximum field number that will allow such an indexed lookup. 3. For fields that do require linear search, we can start the linear search at the location where we found the previous field, taking advantage of the fact that field numbers are generally increasing. 4. When the hasbit index is less than 32 (the common case) we can use a less expensive code sequence to set it. 5. We check for the hasbit case before the oneof case, as optional fields are more common than oneof fields. Benchmark results indicate a 20% improvement in parse speed with a small code size increase: ``` name old time/op new time/op delta ArenaOneAlloc 21.3ns ± 0% 21.5ns ± 0% +0.96% (p=0.000 n=12+12) ArenaInitialBlockOneAlloc 6.32ns ± 0% 6.32ns ± 0% +0.03% (p=0.000 n=12+10) LoadDescriptor_Upb 53.5µs ± 1% 51.5µs ± 2% -3.70% (p=0.000 n=12+12) LoadAdsDescriptor_Upb 2.78ms ± 2% 2.68ms ± 0% -3.57% (p=0.000 n=12+12) LoadDescriptor_Proto2 240µs ± 0% 240µs ± 0% +0.12% (p=0.001 n=12+12) LoadAdsDescriptor_Proto2 12.8ms ± 0% 12.7ms ± 0% -1.15% (p=0.000 n=12+10) Parse_Upb_FileDesc<UseArena,Copy> 13.2µs ± 2% 10.7µs ± 0% -18.49% (p=0.000 n=10+12) Parse_Upb_FileDesc<UseArena,Alias> 11.3µs ± 0% 9.6µs ± 0% -15.11% (p=0.000 n=12+11) Parse_Upb_FileDesc<InitBlock,Copy> 12.7µs ± 0% 10.3µs ± 0% -19.00% (p=0.000 n=10+12) Parse_Upb_FileDesc<InitBlock,Alias> 10.9µs ± 0% 9.2µs ± 0% -15.82% (p=0.000 n=12+12) Parse_Proto2<FileDesc,NoArena,Copy> 29.4µs ± 0% 29.5µs ± 0% +0.61% (p=0.000 n=12+12) Parse_Proto2<FileDesc,UseArena,Copy> 20.7µs ± 2% 20.6µs ± 2% ~ (p=0.260 n=12+11) Parse_Proto2<FileDesc,InitBlock,Copy> 16.7µs ± 1% 16.7µs ± 0% -0.25% (p=0.036 n=12+10) Parse_Proto2<FileDescSV,InitBlock,Alias> 16.5µs ± 0% 16.5µs ± 0% +0.20% (p=0.016 n=12+11) SerializeDescriptor_Proto2 5.30µs ± 1% 5.36µs ± 1% +1.09% (p=0.000 n=12+11) SerializeDescriptor_Upb 12.9µs ± 0% 13.0µs ± 0% +0.90% (p=0.000 n=12+11) FILE SIZE VM SIZE -------------- -------------- +1.5% +176 +1.6% +176 upb/decode.c +1.8% +176 +1.9% +176 decode_msg +0.4% +64 +0.4% +64 upb/def.c +1.4% +64 +1.4% +64 _upb_symtab_addfile +1.2% +48 +1.4% +48 upb/reflection.c +15% +32 +18% +32 upb_msg_set +2.9% +16 +3.1% +16 upb_msg_mutable -9.3% -288 [ = ] 0 [Unmapped] [ = ] 0 +0.2% +288 TOTAL ```
4 years ago
if (field->presence > 0) {
UPB_PRIVATE(_upb_Message_SetHasbit)(msg, field);
Optimized decoder and paved the way for parsing extensions. The primary motivation for this change is to avoid referring to the `upb_msglayout` object when we are trying to fetch the `upb_msglayout` object for a sub-message. This will help pave the way for parsing extensions. We also implement several optimizations so that we can make this change without regressing performance. Normally we compute the layout for a sub-message field like so: ``` const upb_msglayout *get_submsg_layout( const upb_msglayout *layout, const upb_msglayout_field *field) { return layout->submsgs[field->submsg_index] } ``` The reason for this indirection is to avoid storing a pointer directly in `upb_msglayout_field`, as this would double its size (from 12 to 24 bytes on 64-bit architectures) which is wasteful as this pointer is only needed for message typed fields. However `get_submsg_layout` as written above does not work for extensions, as they will not have entries in the message's `layout->submsgs` array by nature, and we want to avoid creating an entire fake `upb_msglayout` for each such extension since that would also be wasteful. This change removes the dependency on `upb_msglayout` by passing down the `submsgs` array instead: ``` const upb_msglayout *get_submsg_layout( const upb_msglayout *const *submsgs, const upb_msglayout_field *field) { return submsgs[field->submsg_index] } ``` This will pave the way for parsing extensions, as we can more easily create an alternative `submsgs` array for extension fields without extra overhead or waste. Along the way several optimizations presented themselves that allow a nice increase in performance: 1. Passing the parsed `wireval` by address instead of by value ended up avoiding an expensive and useless stack copy (this is on Clang, which was used for all measurements). 2. When field numbers are densely packed, we can find a field by number with a single indexed lookup instead of linear search. At codegen time we can compute the maximum field number that will allow such an indexed lookup. 3. For fields that do require linear search, we can start the linear search at the location where we found the previous field, taking advantage of the fact that field numbers are generally increasing. 4. When the hasbit index is less than 32 (the common case) we can use a less expensive code sequence to set it. 5. We check for the hasbit case before the oneof case, as optional fields are more common than oneof fields. Benchmark results indicate a 20% improvement in parse speed with a small code size increase: ``` name old time/op new time/op delta ArenaOneAlloc 21.3ns ± 0% 21.5ns ± 0% +0.96% (p=0.000 n=12+12) ArenaInitialBlockOneAlloc 6.32ns ± 0% 6.32ns ± 0% +0.03% (p=0.000 n=12+10) LoadDescriptor_Upb 53.5µs ± 1% 51.5µs ± 2% -3.70% (p=0.000 n=12+12) LoadAdsDescriptor_Upb 2.78ms ± 2% 2.68ms ± 0% -3.57% (p=0.000 n=12+12) LoadDescriptor_Proto2 240µs ± 0% 240µs ± 0% +0.12% (p=0.001 n=12+12) LoadAdsDescriptor_Proto2 12.8ms ± 0% 12.7ms ± 0% -1.15% (p=0.000 n=12+10) Parse_Upb_FileDesc<UseArena,Copy> 13.2µs ± 2% 10.7µs ± 0% -18.49% (p=0.000 n=10+12) Parse_Upb_FileDesc<UseArena,Alias> 11.3µs ± 0% 9.6µs ± 0% -15.11% (p=0.000 n=12+11) Parse_Upb_FileDesc<InitBlock,Copy> 12.7µs ± 0% 10.3µs ± 0% -19.00% (p=0.000 n=10+12) Parse_Upb_FileDesc<InitBlock,Alias> 10.9µs ± 0% 9.2µs ± 0% -15.82% (p=0.000 n=12+12) Parse_Proto2<FileDesc,NoArena,Copy> 29.4µs ± 0% 29.5µs ± 0% +0.61% (p=0.000 n=12+12) Parse_Proto2<FileDesc,UseArena,Copy> 20.7µs ± 2% 20.6µs ± 2% ~ (p=0.260 n=12+11) Parse_Proto2<FileDesc,InitBlock,Copy> 16.7µs ± 1% 16.7µs ± 0% -0.25% (p=0.036 n=12+10) Parse_Proto2<FileDescSV,InitBlock,Alias> 16.5µs ± 0% 16.5µs ± 0% +0.20% (p=0.016 n=12+11) SerializeDescriptor_Proto2 5.30µs ± 1% 5.36µs ± 1% +1.09% (p=0.000 n=12+11) SerializeDescriptor_Upb 12.9µs ± 0% 13.0µs ± 0% +0.90% (p=0.000 n=12+11) FILE SIZE VM SIZE -------------- -------------- +1.5% +176 +1.6% +176 upb/decode.c +1.8% +176 +1.9% +176 decode_msg +0.4% +64 +0.4% +64 upb/def.c +1.4% +64 +1.4% +64 _upb_symtab_addfile +1.2% +48 +1.4% +48 upb/reflection.c +15% +32 +18% +32 upb_msg_set +2.9% +16 +3.1% +16 upb_msg_mutable -9.3% -288 [ = ] 0 [Unmapped] [ = ] 0 +0.2% +288 TOTAL ```
4 years ago
} else if (field->presence < 0) {
/* Oneof case */
uint32_t* oneof_case = UPB_PRIVATE(_upb_Message_OneofCasePtr)(msg, field);
if (op == kUpb_DecodeOp_SubMessage &&
*oneof_case != field->UPB_PRIVATE(number)) {
Fixes for PHP. (#286) - A new PHP-specific upb amalgamation. It contains everything related to upb_msg, but leaves out all of the old handlers-related interfaces and encoders/decoders. # Schema/Defs Changes - Changed `upb_fielddef_msgsubdef()` and `upb_fielddef_enumsubdef()` to return `NULL` instead of assert-failing if the field is not a message or enum. - Added `upb_msgdef_iswrapper()`, to test whether this is a wrapper well-known type. # Decoder - Decoder bugfix: when we parse a submessage inside a oneof, we need to clear out any previous data, so we don't misinterpret it as a pointer to an existing submessage. # JSON Decoder - Allowed well-known types at the top level to have their special processing. - Fixed a bug that could occur when parsing nested empty lists/objects, eg `[[]]`. - Made the "ignore unknown" option also be permissive about unknown enumerators by setting them to 0. # JSON Encoder - Allowed well-known types at the top level to have their special processing. - Removed all spaces after `:` and `,` characters, to match the old encoder and pass goldenfile tests. # Message / Reflection - Changed `upb_msg_hasoneof()` -> `upb_msg_whichoneof()`. The new function returns the `upb_fielddef*` of whichever oneof is set. - Implemented `upb_msg_clearfield()` and added/implemented `upb_msg_clear()`. - Added `upb_msg_discardunknown()`. Part of me thinks this should go in a util library instead of core reflection since it is a recursive algorithm. # Compiler - Always emit descriptors as an array instead of as a string, to avoid exceeding maximum string lengths. If this becomes a speed issue later we can go back to two separate paths.
5 years ago
memset(mem, 0, sizeof(void*));
}
*oneof_case = field->UPB_PRIVATE(number);
}
/* Store into message. */
switch (op) {
case kUpb_DecodeOp_SubMessage: {
Added a new dynamic tree shaking model to upb, with the intention of removing the old model once YouTube has migrated. The `kUpb_DecodeOption_ExperimentalAllowUnlinked` flag to the decoder will enable the new behavior. When that flag is not passed, tree shaking with the old model will still be possible. "Dynamic tree shaking" in upb is a feature that allows messages to be parsed even if the MiniTables have not been fully linked. Unlinked sub-message fields can be parsed by preserving their data in the unknown fields. If the application later discovers that the message field is actually needed, the MiniTable can be patched to properly link that field, and existing message instances can "promote" the data from the unknown fields to an actual message of the correct type. Before this change, dynamic tree shaking stored unparsed message data in the unknown fields of the *parent*. In effect, we were treating the field as if it did not exist at all. This meant that parsing an unlinked field did not affect the hasbits or oneof cases of the parent, nor did it create a `upb_Array` or `upb_Map` for array/map fields. Only when a message was linked and promoted did any of these things occur. While this model had some amount of conceptual simplicity, it caused significant problems with oneofs. When multiple fields inside a single oneof are parsed from the wire, order matters, because later oneof fields must overwrite earlier ones. Dynamic tree shaking can mean that some fields in a oneof are linked while others are not. It is essential that we preserve this ordering semantic even when dynamic tree shaking is being used, but it is difficult to do if the oneof's data can be split between linked fields (which have been reified into parsed field data) and unlinked fields (whose data lives in the unknown fields of the parent). To solve this problem, this CL changes the representation for unlinked fields. Instead of being placed in the parent's unknown fields, we create an actual message instance for each unlinked message we parse, but we use a placeholder "empty message" MiniTable as the message's type. All of the message's data will therefore be placed into the "empty message's" unknown fields. But unlike before, this "empty message" is actually present according to the hasbits, oneof case, and `upb_Array`/`upb_Map` of the parent. This means that all of the oneof presence logic works as normal. Since the MiniTable can be patched at any time, we need a bit in the message instance itself to signal whether a pointer to a sub-message is an "empty message" or not. When dynamic tree shaking is in use, all users must be capable of recognizing an empty message and acting accordingly (promoting, etc) even if the MiniTable itself says that the field is linked. Because dynamic tree shaking imposes this extra requirement on users, we require that users pass an extra option to the decoder to allow parsing of unlinked sub-messages. Many existing users of upb (Ruby, PHP, Python, etc) will always have fully-linked MiniTables, so there is no reason for them to add extra logic to handle empty messages. By omitting the `kUpb_DecodeOption_ExperimentalAllowUnlinked` option, they will be relieved of the duty to check the tagged pointer that would indicate an empty, unlinked message. For existing users of dynamic tree shaking, there are three main changes: 1. The APIs in message/promote.h have changed, and users will need to update to the new interfaces. 2. The model for maps has changed slightly. Before, we required that map entries always had their values linked; for dynamic tree shaking to apply to maps, we required that the *entry* was left unlinked, not the entry's value. In the new model, that is reversed: map entries must always be linked, but a map entry's value can be unlinked. 3. The presence model for unlinked fields has changed. Unlinked fields will now register as "present" from the perspective of hasbits, oneof cases, and array/map entries. Users must test the tagged pointer to know if a message is of the correct, linked type or whether it is a placeholder "empty" message. There is a new function `upb_Message_GetTaggedMessagePtr()`, as well as a new accessor `upb_MessageValue.tagged_msg_val` that can be used to read and test the tagged pointer directly. PiperOrigin-RevId: 535288031
2 years ago
upb_TaggedMessagePtr* submsgp = mem;
upb_Message* submsg;
if (*submsgp) {
submsg = _upb_Decoder_ReuseSubMessage(d, subs, field, submsgp);
} else {
submsg = _upb_Decoder_NewSubMessage(d, subs, field, submsgp);
}
if (UPB_UNLIKELY(type == kUpb_FieldType_Group)) {
ptr = _upb_Decoder_DecodeKnownGroup(d, ptr, submsg, subs, field);
} else {
ptr = _upb_Decoder_DecodeSubMessage(d, ptr, submsg, subs, field,
val->size);
}
break;
}
case kUpb_DecodeOp_String:
_upb_Decoder_VerifyUtf8(d, ptr, val->size);
/* Fallthrough. */
case kUpb_DecodeOp_Bytes:
return _upb_Decoder_ReadString(d, ptr, val->size, mem);
case kUpb_DecodeOp_Scalar8Byte:
Optimized decoder and paved the way for parsing extensions. The primary motivation for this change is to avoid referring to the `upb_msglayout` object when we are trying to fetch the `upb_msglayout` object for a sub-message. This will help pave the way for parsing extensions. We also implement several optimizations so that we can make this change without regressing performance. Normally we compute the layout for a sub-message field like so: ``` const upb_msglayout *get_submsg_layout( const upb_msglayout *layout, const upb_msglayout_field *field) { return layout->submsgs[field->submsg_index] } ``` The reason for this indirection is to avoid storing a pointer directly in `upb_msglayout_field`, as this would double its size (from 12 to 24 bytes on 64-bit architectures) which is wasteful as this pointer is only needed for message typed fields. However `get_submsg_layout` as written above does not work for extensions, as they will not have entries in the message's `layout->submsgs` array by nature, and we want to avoid creating an entire fake `upb_msglayout` for each such extension since that would also be wasteful. This change removes the dependency on `upb_msglayout` by passing down the `submsgs` array instead: ``` const upb_msglayout *get_submsg_layout( const upb_msglayout *const *submsgs, const upb_msglayout_field *field) { return submsgs[field->submsg_index] } ``` This will pave the way for parsing extensions, as we can more easily create an alternative `submsgs` array for extension fields without extra overhead or waste. Along the way several optimizations presented themselves that allow a nice increase in performance: 1. Passing the parsed `wireval` by address instead of by value ended up avoiding an expensive and useless stack copy (this is on Clang, which was used for all measurements). 2. When field numbers are densely packed, we can find a field by number with a single indexed lookup instead of linear search. At codegen time we can compute the maximum field number that will allow such an indexed lookup. 3. For fields that do require linear search, we can start the linear search at the location where we found the previous field, taking advantage of the fact that field numbers are generally increasing. 4. When the hasbit index is less than 32 (the common case) we can use a less expensive code sequence to set it. 5. We check for the hasbit case before the oneof case, as optional fields are more common than oneof fields. Benchmark results indicate a 20% improvement in parse speed with a small code size increase: ``` name old time/op new time/op delta ArenaOneAlloc 21.3ns ± 0% 21.5ns ± 0% +0.96% (p=0.000 n=12+12) ArenaInitialBlockOneAlloc 6.32ns ± 0% 6.32ns ± 0% +0.03% (p=0.000 n=12+10) LoadDescriptor_Upb 53.5µs ± 1% 51.5µs ± 2% -3.70% (p=0.000 n=12+12) LoadAdsDescriptor_Upb 2.78ms ± 2% 2.68ms ± 0% -3.57% (p=0.000 n=12+12) LoadDescriptor_Proto2 240µs ± 0% 240µs ± 0% +0.12% (p=0.001 n=12+12) LoadAdsDescriptor_Proto2 12.8ms ± 0% 12.7ms ± 0% -1.15% (p=0.000 n=12+10) Parse_Upb_FileDesc<UseArena,Copy> 13.2µs ± 2% 10.7µs ± 0% -18.49% (p=0.000 n=10+12) Parse_Upb_FileDesc<UseArena,Alias> 11.3µs ± 0% 9.6µs ± 0% -15.11% (p=0.000 n=12+11) Parse_Upb_FileDesc<InitBlock,Copy> 12.7µs ± 0% 10.3µs ± 0% -19.00% (p=0.000 n=10+12) Parse_Upb_FileDesc<InitBlock,Alias> 10.9µs ± 0% 9.2µs ± 0% -15.82% (p=0.000 n=12+12) Parse_Proto2<FileDesc,NoArena,Copy> 29.4µs ± 0% 29.5µs ± 0% +0.61% (p=0.000 n=12+12) Parse_Proto2<FileDesc,UseArena,Copy> 20.7µs ± 2% 20.6µs ± 2% ~ (p=0.260 n=12+11) Parse_Proto2<FileDesc,InitBlock,Copy> 16.7µs ± 1% 16.7µs ± 0% -0.25% (p=0.036 n=12+10) Parse_Proto2<FileDescSV,InitBlock,Alias> 16.5µs ± 0% 16.5µs ± 0% +0.20% (p=0.016 n=12+11) SerializeDescriptor_Proto2 5.30µs ± 1% 5.36µs ± 1% +1.09% (p=0.000 n=12+11) SerializeDescriptor_Upb 12.9µs ± 0% 13.0µs ± 0% +0.90% (p=0.000 n=12+11) FILE SIZE VM SIZE -------------- -------------- +1.5% +176 +1.6% +176 upb/decode.c +1.8% +176 +1.9% +176 decode_msg +0.4% +64 +0.4% +64 upb/def.c +1.4% +64 +1.4% +64 _upb_symtab_addfile +1.2% +48 +1.4% +48 upb/reflection.c +15% +32 +18% +32 upb_msg_set +2.9% +16 +3.1% +16 upb_msg_mutable -9.3% -288 [ = ] 0 [Unmapped] [ = ] 0 +0.2% +288 TOTAL ```
4 years ago
memcpy(mem, val, 8);
break;
case kUpb_DecodeOp_Enum:
case kUpb_DecodeOp_Scalar4Byte:
Optimized decoder and paved the way for parsing extensions. The primary motivation for this change is to avoid referring to the `upb_msglayout` object when we are trying to fetch the `upb_msglayout` object for a sub-message. This will help pave the way for parsing extensions. We also implement several optimizations so that we can make this change without regressing performance. Normally we compute the layout for a sub-message field like so: ``` const upb_msglayout *get_submsg_layout( const upb_msglayout *layout, const upb_msglayout_field *field) { return layout->submsgs[field->submsg_index] } ``` The reason for this indirection is to avoid storing a pointer directly in `upb_msglayout_field`, as this would double its size (from 12 to 24 bytes on 64-bit architectures) which is wasteful as this pointer is only needed for message typed fields. However `get_submsg_layout` as written above does not work for extensions, as they will not have entries in the message's `layout->submsgs` array by nature, and we want to avoid creating an entire fake `upb_msglayout` for each such extension since that would also be wasteful. This change removes the dependency on `upb_msglayout` by passing down the `submsgs` array instead: ``` const upb_msglayout *get_submsg_layout( const upb_msglayout *const *submsgs, const upb_msglayout_field *field) { return submsgs[field->submsg_index] } ``` This will pave the way for parsing extensions, as we can more easily create an alternative `submsgs` array for extension fields without extra overhead or waste. Along the way several optimizations presented themselves that allow a nice increase in performance: 1. Passing the parsed `wireval` by address instead of by value ended up avoiding an expensive and useless stack copy (this is on Clang, which was used for all measurements). 2. When field numbers are densely packed, we can find a field by number with a single indexed lookup instead of linear search. At codegen time we can compute the maximum field number that will allow such an indexed lookup. 3. For fields that do require linear search, we can start the linear search at the location where we found the previous field, taking advantage of the fact that field numbers are generally increasing. 4. When the hasbit index is less than 32 (the common case) we can use a less expensive code sequence to set it. 5. We check for the hasbit case before the oneof case, as optional fields are more common than oneof fields. Benchmark results indicate a 20% improvement in parse speed with a small code size increase: ``` name old time/op new time/op delta ArenaOneAlloc 21.3ns ± 0% 21.5ns ± 0% +0.96% (p=0.000 n=12+12) ArenaInitialBlockOneAlloc 6.32ns ± 0% 6.32ns ± 0% +0.03% (p=0.000 n=12+10) LoadDescriptor_Upb 53.5µs ± 1% 51.5µs ± 2% -3.70% (p=0.000 n=12+12) LoadAdsDescriptor_Upb 2.78ms ± 2% 2.68ms ± 0% -3.57% (p=0.000 n=12+12) LoadDescriptor_Proto2 240µs ± 0% 240µs ± 0% +0.12% (p=0.001 n=12+12) LoadAdsDescriptor_Proto2 12.8ms ± 0% 12.7ms ± 0% -1.15% (p=0.000 n=12+10) Parse_Upb_FileDesc<UseArena,Copy> 13.2µs ± 2% 10.7µs ± 0% -18.49% (p=0.000 n=10+12) Parse_Upb_FileDesc<UseArena,Alias> 11.3µs ± 0% 9.6µs ± 0% -15.11% (p=0.000 n=12+11) Parse_Upb_FileDesc<InitBlock,Copy> 12.7µs ± 0% 10.3µs ± 0% -19.00% (p=0.000 n=10+12) Parse_Upb_FileDesc<InitBlock,Alias> 10.9µs ± 0% 9.2µs ± 0% -15.82% (p=0.000 n=12+12) Parse_Proto2<FileDesc,NoArena,Copy> 29.4µs ± 0% 29.5µs ± 0% +0.61% (p=0.000 n=12+12) Parse_Proto2<FileDesc,UseArena,Copy> 20.7µs ± 2% 20.6µs ± 2% ~ (p=0.260 n=12+11) Parse_Proto2<FileDesc,InitBlock,Copy> 16.7µs ± 1% 16.7µs ± 0% -0.25% (p=0.036 n=12+10) Parse_Proto2<FileDescSV,InitBlock,Alias> 16.5µs ± 0% 16.5µs ± 0% +0.20% (p=0.016 n=12+11) SerializeDescriptor_Proto2 5.30µs ± 1% 5.36µs ± 1% +1.09% (p=0.000 n=12+11) SerializeDescriptor_Upb 12.9µs ± 0% 13.0µs ± 0% +0.90% (p=0.000 n=12+11) FILE SIZE VM SIZE -------------- -------------- +1.5% +176 +1.6% +176 upb/decode.c +1.8% +176 +1.9% +176 decode_msg +0.4% +64 +0.4% +64 upb/def.c +1.4% +64 +1.4% +64 _upb_symtab_addfile +1.2% +48 +1.4% +48 upb/reflection.c +15% +32 +18% +32 upb_msg_set +2.9% +16 +3.1% +16 upb_msg_mutable -9.3% -288 [ = ] 0 [Unmapped] [ = ] 0 +0.2% +288 TOTAL ```
4 years ago
memcpy(mem, val, 4);
break;
case kUpb_DecodeOp_Scalar1Byte:
Optimized decoder and paved the way for parsing extensions. The primary motivation for this change is to avoid referring to the `upb_msglayout` object when we are trying to fetch the `upb_msglayout` object for a sub-message. This will help pave the way for parsing extensions. We also implement several optimizations so that we can make this change without regressing performance. Normally we compute the layout for a sub-message field like so: ``` const upb_msglayout *get_submsg_layout( const upb_msglayout *layout, const upb_msglayout_field *field) { return layout->submsgs[field->submsg_index] } ``` The reason for this indirection is to avoid storing a pointer directly in `upb_msglayout_field`, as this would double its size (from 12 to 24 bytes on 64-bit architectures) which is wasteful as this pointer is only needed for message typed fields. However `get_submsg_layout` as written above does not work for extensions, as they will not have entries in the message's `layout->submsgs` array by nature, and we want to avoid creating an entire fake `upb_msglayout` for each such extension since that would also be wasteful. This change removes the dependency on `upb_msglayout` by passing down the `submsgs` array instead: ``` const upb_msglayout *get_submsg_layout( const upb_msglayout *const *submsgs, const upb_msglayout_field *field) { return submsgs[field->submsg_index] } ``` This will pave the way for parsing extensions, as we can more easily create an alternative `submsgs` array for extension fields without extra overhead or waste. Along the way several optimizations presented themselves that allow a nice increase in performance: 1. Passing the parsed `wireval` by address instead of by value ended up avoiding an expensive and useless stack copy (this is on Clang, which was used for all measurements). 2. When field numbers are densely packed, we can find a field by number with a single indexed lookup instead of linear search. At codegen time we can compute the maximum field number that will allow such an indexed lookup. 3. For fields that do require linear search, we can start the linear search at the location where we found the previous field, taking advantage of the fact that field numbers are generally increasing. 4. When the hasbit index is less than 32 (the common case) we can use a less expensive code sequence to set it. 5. We check for the hasbit case before the oneof case, as optional fields are more common than oneof fields. Benchmark results indicate a 20% improvement in parse speed with a small code size increase: ``` name old time/op new time/op delta ArenaOneAlloc 21.3ns ± 0% 21.5ns ± 0% +0.96% (p=0.000 n=12+12) ArenaInitialBlockOneAlloc 6.32ns ± 0% 6.32ns ± 0% +0.03% (p=0.000 n=12+10) LoadDescriptor_Upb 53.5µs ± 1% 51.5µs ± 2% -3.70% (p=0.000 n=12+12) LoadAdsDescriptor_Upb 2.78ms ± 2% 2.68ms ± 0% -3.57% (p=0.000 n=12+12) LoadDescriptor_Proto2 240µs ± 0% 240µs ± 0% +0.12% (p=0.001 n=12+12) LoadAdsDescriptor_Proto2 12.8ms ± 0% 12.7ms ± 0% -1.15% (p=0.000 n=12+10) Parse_Upb_FileDesc<UseArena,Copy> 13.2µs ± 2% 10.7µs ± 0% -18.49% (p=0.000 n=10+12) Parse_Upb_FileDesc<UseArena,Alias> 11.3µs ± 0% 9.6µs ± 0% -15.11% (p=0.000 n=12+11) Parse_Upb_FileDesc<InitBlock,Copy> 12.7µs ± 0% 10.3µs ± 0% -19.00% (p=0.000 n=10+12) Parse_Upb_FileDesc<InitBlock,Alias> 10.9µs ± 0% 9.2µs ± 0% -15.82% (p=0.000 n=12+12) Parse_Proto2<FileDesc,NoArena,Copy> 29.4µs ± 0% 29.5µs ± 0% +0.61% (p=0.000 n=12+12) Parse_Proto2<FileDesc,UseArena,Copy> 20.7µs ± 2% 20.6µs ± 2% ~ (p=0.260 n=12+11) Parse_Proto2<FileDesc,InitBlock,Copy> 16.7µs ± 1% 16.7µs ± 0% -0.25% (p=0.036 n=12+10) Parse_Proto2<FileDescSV,InitBlock,Alias> 16.5µs ± 0% 16.5µs ± 0% +0.20% (p=0.016 n=12+11) SerializeDescriptor_Proto2 5.30µs ± 1% 5.36µs ± 1% +1.09% (p=0.000 n=12+11) SerializeDescriptor_Upb 12.9µs ± 0% 13.0µs ± 0% +0.90% (p=0.000 n=12+11) FILE SIZE VM SIZE -------------- -------------- +1.5% +176 +1.6% +176 upb/decode.c +1.8% +176 +1.9% +176 decode_msg +0.4% +64 +0.4% +64 upb/def.c +1.4% +64 +1.4% +64 _upb_symtab_addfile +1.2% +48 +1.4% +48 upb/reflection.c +15% +32 +18% +32 upb_msg_set +2.9% +16 +3.1% +16 upb_msg_mutable -9.3% -288 [ = ] 0 [Unmapped] [ = ] 0 +0.2% +288 TOTAL ```
4 years ago
memcpy(mem, val, 1);
break;
default:
UPB_UNREACHABLE();
}
return ptr;
}
UPB_NOINLINE
const char* _upb_Decoder_CheckRequired(upb_Decoder* d, const char* ptr,
const upb_Message* msg,
const upb_MiniTable* m) {
UPB_ASSERT(m->UPB_PRIVATE(required_count));
if (UPB_LIKELY((d->options & kUpb_DecodeOption_CheckRequired) == 0)) {
return ptr;
}
uint64_t msg_head;
memcpy(&msg_head, msg, 8);
msg_head = _upb_BigEndian_Swap64(msg_head);
if (UPB_PRIVATE(_upb_MiniTable_RequiredMask)(m) & ~msg_head) {
d->missing_required = true;
}
return ptr;
}
UPB_FORCEINLINE
static bool _upb_Decoder_TryFastDispatch(upb_Decoder* d, const char** ptr,
upb_Message* msg,
const upb_MiniTable* m) {
#if UPB_FASTTABLE
if (m && m->UPB_PRIVATE(table_mask) != (unsigned char)-1) {
uint16_t tag = _upb_FastDecoder_LoadTag(*ptr);
intptr_t table = decode_totable(m);
*ptr = _upb_FastDecoder_TagDispatch(d, *ptr, msg, table, 0, tag);
return true;
4 years ago
}
#endif
return false;
}
4 years ago
static const char* upb_Decoder_SkipField(upb_Decoder* d, const char* ptr,
uint32_t tag) {
int field_number = tag >> 3;
int wire_type = tag & 7;
switch (wire_type) {
case kUpb_WireType_Varint: {
uint64_t val;
return _upb_Decoder_DecodeVarint(d, ptr, &val);
}
case kUpb_WireType_64Bit:
return ptr + 8;
case kUpb_WireType_32Bit:
return ptr + 4;
case kUpb_WireType_Delimited: {
uint32_t size;
ptr = upb_Decoder_DecodeSize(d, ptr, &size);
return ptr + size;
}
case kUpb_WireType_StartGroup:
return _upb_Decoder_DecodeUnknownGroup(d, ptr, field_number);
default:
_upb_Decoder_ErrorJmp(d, kUpb_DecodeStatus_Malformed);
}
}
enum {
kStartItemTag = ((kUpb_MsgSet_Item << 3) | kUpb_WireType_StartGroup),
kEndItemTag = ((kUpb_MsgSet_Item << 3) | kUpb_WireType_EndGroup),
kTypeIdTag = ((kUpb_MsgSet_TypeId << 3) | kUpb_WireType_Varint),
kMessageTag = ((kUpb_MsgSet_Message << 3) | kUpb_WireType_Delimited),
};
static void upb_Decoder_AddKnownMessageSetItem(
upb_Decoder* d, upb_Message* msg, const upb_MiniTableExtension* item_mt,
const char* data, uint32_t size) {
upb_Extension* ext =
_upb_Message_GetOrCreateExtension(msg, item_mt, &d->arena);
if (UPB_UNLIKELY(!ext)) {
_upb_Decoder_ErrorJmp(d, kUpb_DecodeStatus_OutOfMemory);
}
Added a new dynamic tree shaking model to upb, with the intention of removing the old model once YouTube has migrated. The `kUpb_DecodeOption_ExperimentalAllowUnlinked` flag to the decoder will enable the new behavior. When that flag is not passed, tree shaking with the old model will still be possible. "Dynamic tree shaking" in upb is a feature that allows messages to be parsed even if the MiniTables have not been fully linked. Unlinked sub-message fields can be parsed by preserving their data in the unknown fields. If the application later discovers that the message field is actually needed, the MiniTable can be patched to properly link that field, and existing message instances can "promote" the data from the unknown fields to an actual message of the correct type. Before this change, dynamic tree shaking stored unparsed message data in the unknown fields of the *parent*. In effect, we were treating the field as if it did not exist at all. This meant that parsing an unlinked field did not affect the hasbits or oneof cases of the parent, nor did it create a `upb_Array` or `upb_Map` for array/map fields. Only when a message was linked and promoted did any of these things occur. While this model had some amount of conceptual simplicity, it caused significant problems with oneofs. When multiple fields inside a single oneof are parsed from the wire, order matters, because later oneof fields must overwrite earlier ones. Dynamic tree shaking can mean that some fields in a oneof are linked while others are not. It is essential that we preserve this ordering semantic even when dynamic tree shaking is being used, but it is difficult to do if the oneof's data can be split between linked fields (which have been reified into parsed field data) and unlinked fields (whose data lives in the unknown fields of the parent). To solve this problem, this CL changes the representation for unlinked fields. Instead of being placed in the parent's unknown fields, we create an actual message instance for each unlinked message we parse, but we use a placeholder "empty message" MiniTable as the message's type. All of the message's data will therefore be placed into the "empty message's" unknown fields. But unlike before, this "empty message" is actually present according to the hasbits, oneof case, and `upb_Array`/`upb_Map` of the parent. This means that all of the oneof presence logic works as normal. Since the MiniTable can be patched at any time, we need a bit in the message instance itself to signal whether a pointer to a sub-message is an "empty message" or not. When dynamic tree shaking is in use, all users must be capable of recognizing an empty message and acting accordingly (promoting, etc) even if the MiniTable itself says that the field is linked. Because dynamic tree shaking imposes this extra requirement on users, we require that users pass an extra option to the decoder to allow parsing of unlinked sub-messages. Many existing users of upb (Ruby, PHP, Python, etc) will always have fully-linked MiniTables, so there is no reason for them to add extra logic to handle empty messages. By omitting the `kUpb_DecodeOption_ExperimentalAllowUnlinked` option, they will be relieved of the duty to check the tagged pointer that would indicate an empty, unlinked message. For existing users of dynamic tree shaking, there are three main changes: 1. The APIs in message/promote.h have changed, and users will need to update to the new interfaces. 2. The model for maps has changed slightly. Before, we required that map entries always had their values linked; for dynamic tree shaking to apply to maps, we required that the *entry* was left unlinked, not the entry's value. In the new model, that is reversed: map entries must always be linked, but a map entry's value can be unlinked. 3. The presence model for unlinked fields has changed. Unlinked fields will now register as "present" from the perspective of hasbits, oneof cases, and array/map entries. Users must test the tagged pointer to know if a message is of the correct, linked type or whether it is a placeholder "empty" message. There is a new function `upb_Message_GetTaggedMessagePtr()`, as well as a new accessor `upb_MessageValue.tagged_msg_val` that can be used to read and test the tagged pointer directly. PiperOrigin-RevId: 535288031
2 years ago
upb_Message* submsg = _upb_Decoder_NewSubMessage(
d, &ext->ext->UPB_PRIVATE(sub), upb_MiniTableExtension_AsField(ext->ext),
(upb_TaggedMessagePtr*)&ext->data);
upb_DecodeStatus status = upb_Decode(
data, size, submsg, upb_MiniTableExtension_GetSubMessage(item_mt),
d->extreg, d->options, &d->arena);
if (status != kUpb_DecodeStatus_Ok) _upb_Decoder_ErrorJmp(d, status);
}
static void upb_Decoder_AddUnknownMessageSetItem(upb_Decoder* d,
upb_Message* msg,
uint32_t type_id,
const char* message_data,
uint32_t message_size) {
char buf[60];
char* ptr = buf;
ptr = upb_Decoder_EncodeVarint32(kStartItemTag, ptr);
ptr = upb_Decoder_EncodeVarint32(kTypeIdTag, ptr);
ptr = upb_Decoder_EncodeVarint32(type_id, ptr);
ptr = upb_Decoder_EncodeVarint32(kMessageTag, ptr);
ptr = upb_Decoder_EncodeVarint32(message_size, ptr);
char* split = ptr;
ptr = upb_Decoder_EncodeVarint32(kEndItemTag, ptr);
char* end = ptr;
if (!UPB_PRIVATE(_upb_Message_AddUnknown)(msg, buf, split - buf, &d->arena) ||
!UPB_PRIVATE(_upb_Message_AddUnknown)(msg, message_data, message_size,
&d->arena) ||
!UPB_PRIVATE(_upb_Message_AddUnknown)(msg, split, end - split,
&d->arena)) {
_upb_Decoder_ErrorJmp(d, kUpb_DecodeStatus_OutOfMemory);
}
}
static void upb_Decoder_AddMessageSetItem(upb_Decoder* d, upb_Message* msg,
const upb_MiniTable* t,
uint32_t type_id, const char* data,
uint32_t size) {
const upb_MiniTableExtension* item_mt =
upb_ExtensionRegistry_Lookup(d->extreg, t, type_id);
if (item_mt) {
upb_Decoder_AddKnownMessageSetItem(d, msg, item_mt, data, size);
} else {
upb_Decoder_AddUnknownMessageSetItem(d, msg, type_id, data, size);
}
}
static const char* upb_Decoder_DecodeMessageSetItem(
upb_Decoder* d, const char* ptr, upb_Message* msg,
const upb_MiniTable* layout) {
uint32_t type_id = 0;
upb_StringView preserved = {NULL, 0};
typedef enum {
kUpb_HaveId = 1 << 0,
kUpb_HavePayload = 1 << 1,
} StateMask;
StateMask state_mask = 0;
while (!_upb_Decoder_IsDone(d, &ptr)) {
uint32_t tag;
ptr = _upb_Decoder_DecodeTag(d, ptr, &tag);
switch (tag) {
case kEndItemTag:
return ptr;
case kTypeIdTag: {
uint64_t tmp;
ptr = _upb_Decoder_DecodeVarint(d, ptr, &tmp);
if (state_mask & kUpb_HaveId) break; // Ignore dup.
state_mask |= kUpb_HaveId;
type_id = tmp;
if (state_mask & kUpb_HavePayload) {
upb_Decoder_AddMessageSetItem(d, msg, layout, type_id, preserved.data,
preserved.size);
}
break;
}
case kMessageTag: {
uint32_t size;
ptr = upb_Decoder_DecodeSize(d, ptr, &size);
const char* data = ptr;
ptr += size;
if (state_mask & kUpb_HavePayload) break; // Ignore dup.
state_mask |= kUpb_HavePayload;
if (state_mask & kUpb_HaveId) {
upb_Decoder_AddMessageSetItem(d, msg, layout, type_id, data, size);
} else {
// Out of order, we must preserve the payload.
preserved.data = data;
preserved.size = size;
}
break;
}
default:
// We do not preserve unexpected fields inside a message set item.
ptr = upb_Decoder_SkipField(d, ptr, tag);
break;
}
}
_upb_Decoder_ErrorJmp(d, kUpb_DecodeStatus_Malformed);
}
static const upb_MiniTableField* _upb_Decoder_FindField(upb_Decoder* d,
const upb_MiniTable* t,
uint32_t field_number,
int* last_field_index) {
static upb_MiniTableField none = {
0, 0, 0, 0, kUpb_FakeFieldType_FieldNotFound, 0};
if (t == NULL) return &none;
size_t idx = ((size_t)field_number) - 1; // 0 wraps to SIZE_MAX
if (idx < t->UPB_PRIVATE(dense_below)) {
/* Fastest case: index into dense fields. */
goto found;
}
if (t->UPB_PRIVATE(dense_below) < t->UPB_PRIVATE(field_count)) {
/* Linear search non-dense fields. Resume scanning from last_field_index
* since fields are usually in order. */
size_t last = *last_field_index;
for (idx = last; idx < t->UPB_PRIVATE(field_count); idx++) {
if (t->UPB_PRIVATE(fields)[idx].UPB_PRIVATE(number) == field_number) {
goto found;
}
}
for (idx = t->UPB_PRIVATE(dense_below); idx < last; idx++) {
if (t->UPB_PRIVATE(fields)[idx].UPB_PRIVATE(number) == field_number) {
goto found;
}
}
}
if (d->extreg) {
switch (t->UPB_PRIVATE(ext)) {
case kUpb_ExtMode_Extendable: {
const upb_MiniTableExtension* ext =
upb_ExtensionRegistry_Lookup(d->extreg, t, field_number);
if (ext) return upb_MiniTableExtension_AsField(ext);
break;
}
case kUpb_ExtMode_IsMessageSet:
if (field_number == kUpb_MsgSet_Item) {
static upb_MiniTableField item = {
0, 0, 0, 0, kUpb_FakeFieldType_MessageSetItem, 0};
return &item;
}
break;
}
}
return &none; /* Unknown field. */
found:
UPB_ASSERT(t->UPB_PRIVATE(fields)[idx].UPB_PRIVATE(number) == field_number);
*last_field_index = idx;
return &t->UPB_PRIVATE(fields)[idx];
}
int _upb_Decoder_GetVarintOp(const upb_MiniTableField* field) {
static const int8_t kVarintOps[] = {
[kUpb_FakeFieldType_FieldNotFound] = kUpb_DecodeOp_UnknownField,
[kUpb_FieldType_Double] = kUpb_DecodeOp_UnknownField,
[kUpb_FieldType_Float] = kUpb_DecodeOp_UnknownField,
[kUpb_FieldType_Int64] = kUpb_DecodeOp_Scalar8Byte,
[kUpb_FieldType_UInt64] = kUpb_DecodeOp_Scalar8Byte,
[kUpb_FieldType_Int32] = kUpb_DecodeOp_Scalar4Byte,
[kUpb_FieldType_Fixed64] = kUpb_DecodeOp_UnknownField,
[kUpb_FieldType_Fixed32] = kUpb_DecodeOp_UnknownField,
[kUpb_FieldType_Bool] = kUpb_DecodeOp_Scalar1Byte,
[kUpb_FieldType_String] = kUpb_DecodeOp_UnknownField,
[kUpb_FieldType_Group] = kUpb_DecodeOp_UnknownField,
[kUpb_FieldType_Message] = kUpb_DecodeOp_UnknownField,
[kUpb_FieldType_Bytes] = kUpb_DecodeOp_UnknownField,
[kUpb_FieldType_UInt32] = kUpb_DecodeOp_Scalar4Byte,
[kUpb_FieldType_Enum] = kUpb_DecodeOp_Enum,
[kUpb_FieldType_SFixed32] = kUpb_DecodeOp_UnknownField,
[kUpb_FieldType_SFixed64] = kUpb_DecodeOp_UnknownField,
[kUpb_FieldType_SInt32] = kUpb_DecodeOp_Scalar4Byte,
[kUpb_FieldType_SInt64] = kUpb_DecodeOp_Scalar8Byte,
[kUpb_FakeFieldType_MessageSetItem] = kUpb_DecodeOp_UnknownField,
};
return kVarintOps[field->UPB_PRIVATE(descriptortype)];
}
UPB_FORCEINLINE
Added a new dynamic tree shaking model to upb, with the intention of removing the old model once YouTube has migrated. The `kUpb_DecodeOption_ExperimentalAllowUnlinked` flag to the decoder will enable the new behavior. When that flag is not passed, tree shaking with the old model will still be possible. "Dynamic tree shaking" in upb is a feature that allows messages to be parsed even if the MiniTables have not been fully linked. Unlinked sub-message fields can be parsed by preserving their data in the unknown fields. If the application later discovers that the message field is actually needed, the MiniTable can be patched to properly link that field, and existing message instances can "promote" the data from the unknown fields to an actual message of the correct type. Before this change, dynamic tree shaking stored unparsed message data in the unknown fields of the *parent*. In effect, we were treating the field as if it did not exist at all. This meant that parsing an unlinked field did not affect the hasbits or oneof cases of the parent, nor did it create a `upb_Array` or `upb_Map` for array/map fields. Only when a message was linked and promoted did any of these things occur. While this model had some amount of conceptual simplicity, it caused significant problems with oneofs. When multiple fields inside a single oneof are parsed from the wire, order matters, because later oneof fields must overwrite earlier ones. Dynamic tree shaking can mean that some fields in a oneof are linked while others are not. It is essential that we preserve this ordering semantic even when dynamic tree shaking is being used, but it is difficult to do if the oneof's data can be split between linked fields (which have been reified into parsed field data) and unlinked fields (whose data lives in the unknown fields of the parent). To solve this problem, this CL changes the representation for unlinked fields. Instead of being placed in the parent's unknown fields, we create an actual message instance for each unlinked message we parse, but we use a placeholder "empty message" MiniTable as the message's type. All of the message's data will therefore be placed into the "empty message's" unknown fields. But unlike before, this "empty message" is actually present according to the hasbits, oneof case, and `upb_Array`/`upb_Map` of the parent. This means that all of the oneof presence logic works as normal. Since the MiniTable can be patched at any time, we need a bit in the message instance itself to signal whether a pointer to a sub-message is an "empty message" or not. When dynamic tree shaking is in use, all users must be capable of recognizing an empty message and acting accordingly (promoting, etc) even if the MiniTable itself says that the field is linked. Because dynamic tree shaking imposes this extra requirement on users, we require that users pass an extra option to the decoder to allow parsing of unlinked sub-messages. Many existing users of upb (Ruby, PHP, Python, etc) will always have fully-linked MiniTables, so there is no reason for them to add extra logic to handle empty messages. By omitting the `kUpb_DecodeOption_ExperimentalAllowUnlinked` option, they will be relieved of the duty to check the tagged pointer that would indicate an empty, unlinked message. For existing users of dynamic tree shaking, there are three main changes: 1. The APIs in message/promote.h have changed, and users will need to update to the new interfaces. 2. The model for maps has changed slightly. Before, we required that map entries always had their values linked; for dynamic tree shaking to apply to maps, we required that the *entry* was left unlinked, not the entry's value. In the new model, that is reversed: map entries must always be linked, but a map entry's value can be unlinked. 3. The presence model for unlinked fields has changed. Unlinked fields will now register as "present" from the perspective of hasbits, oneof cases, and array/map entries. Users must test the tagged pointer to know if a message is of the correct, linked type or whether it is a placeholder "empty" message. There is a new function `upb_Message_GetTaggedMessagePtr()`, as well as a new accessor `upb_MessageValue.tagged_msg_val` that can be used to read and test the tagged pointer directly. PiperOrigin-RevId: 535288031
2 years ago
static void _upb_Decoder_CheckUnlinked(upb_Decoder* d, const upb_MiniTable* mt,
const upb_MiniTableField* field,
int* op) {
// If sub-message is not linked, treat as unknown.
if (field->UPB_PRIVATE(mode) & kUpb_LabelFlags_IsExtension) return;
const upb_MiniTable* mt_sub =
_upb_MiniTableSubs_MessageByField(mt->UPB_PRIVATE(subs), field);
Added a new dynamic tree shaking model to upb, with the intention of removing the old model once YouTube has migrated. The `kUpb_DecodeOption_ExperimentalAllowUnlinked` flag to the decoder will enable the new behavior. When that flag is not passed, tree shaking with the old model will still be possible. "Dynamic tree shaking" in upb is a feature that allows messages to be parsed even if the MiniTables have not been fully linked. Unlinked sub-message fields can be parsed by preserving their data in the unknown fields. If the application later discovers that the message field is actually needed, the MiniTable can be patched to properly link that field, and existing message instances can "promote" the data from the unknown fields to an actual message of the correct type. Before this change, dynamic tree shaking stored unparsed message data in the unknown fields of the *parent*. In effect, we were treating the field as if it did not exist at all. This meant that parsing an unlinked field did not affect the hasbits or oneof cases of the parent, nor did it create a `upb_Array` or `upb_Map` for array/map fields. Only when a message was linked and promoted did any of these things occur. While this model had some amount of conceptual simplicity, it caused significant problems with oneofs. When multiple fields inside a single oneof are parsed from the wire, order matters, because later oneof fields must overwrite earlier ones. Dynamic tree shaking can mean that some fields in a oneof are linked while others are not. It is essential that we preserve this ordering semantic even when dynamic tree shaking is being used, but it is difficult to do if the oneof's data can be split between linked fields (which have been reified into parsed field data) and unlinked fields (whose data lives in the unknown fields of the parent). To solve this problem, this CL changes the representation for unlinked fields. Instead of being placed in the parent's unknown fields, we create an actual message instance for each unlinked message we parse, but we use a placeholder "empty message" MiniTable as the message's type. All of the message's data will therefore be placed into the "empty message's" unknown fields. But unlike before, this "empty message" is actually present according to the hasbits, oneof case, and `upb_Array`/`upb_Map` of the parent. This means that all of the oneof presence logic works as normal. Since the MiniTable can be patched at any time, we need a bit in the message instance itself to signal whether a pointer to a sub-message is an "empty message" or not. When dynamic tree shaking is in use, all users must be capable of recognizing an empty message and acting accordingly (promoting, etc) even if the MiniTable itself says that the field is linked. Because dynamic tree shaking imposes this extra requirement on users, we require that users pass an extra option to the decoder to allow parsing of unlinked sub-messages. Many existing users of upb (Ruby, PHP, Python, etc) will always have fully-linked MiniTables, so there is no reason for them to add extra logic to handle empty messages. By omitting the `kUpb_DecodeOption_ExperimentalAllowUnlinked` option, they will be relieved of the duty to check the tagged pointer that would indicate an empty, unlinked message. For existing users of dynamic tree shaking, there are three main changes: 1. The APIs in message/promote.h have changed, and users will need to update to the new interfaces. 2. The model for maps has changed slightly. Before, we required that map entries always had their values linked; for dynamic tree shaking to apply to maps, we required that the *entry* was left unlinked, not the entry's value. In the new model, that is reversed: map entries must always be linked, but a map entry's value can be unlinked. 3. The presence model for unlinked fields has changed. Unlinked fields will now register as "present" from the perspective of hasbits, oneof cases, and array/map entries. Users must test the tagged pointer to know if a message is of the correct, linked type or whether it is a placeholder "empty" message. There is a new function `upb_Message_GetTaggedMessagePtr()`, as well as a new accessor `upb_MessageValue.tagged_msg_val` that can be used to read and test the tagged pointer directly. PiperOrigin-RevId: 535288031
2 years ago
if ((d->options & kUpb_DecodeOption_ExperimentalAllowUnlinked) ||
!UPB_PRIVATE(_upb_MiniTable_IsEmpty)(mt_sub)) {
Added a new dynamic tree shaking model to upb, with the intention of removing the old model once YouTube has migrated. The `kUpb_DecodeOption_ExperimentalAllowUnlinked` flag to the decoder will enable the new behavior. When that flag is not passed, tree shaking with the old model will still be possible. "Dynamic tree shaking" in upb is a feature that allows messages to be parsed even if the MiniTables have not been fully linked. Unlinked sub-message fields can be parsed by preserving their data in the unknown fields. If the application later discovers that the message field is actually needed, the MiniTable can be patched to properly link that field, and existing message instances can "promote" the data from the unknown fields to an actual message of the correct type. Before this change, dynamic tree shaking stored unparsed message data in the unknown fields of the *parent*. In effect, we were treating the field as if it did not exist at all. This meant that parsing an unlinked field did not affect the hasbits or oneof cases of the parent, nor did it create a `upb_Array` or `upb_Map` for array/map fields. Only when a message was linked and promoted did any of these things occur. While this model had some amount of conceptual simplicity, it caused significant problems with oneofs. When multiple fields inside a single oneof are parsed from the wire, order matters, because later oneof fields must overwrite earlier ones. Dynamic tree shaking can mean that some fields in a oneof are linked while others are not. It is essential that we preserve this ordering semantic even when dynamic tree shaking is being used, but it is difficult to do if the oneof's data can be split between linked fields (which have been reified into parsed field data) and unlinked fields (whose data lives in the unknown fields of the parent). To solve this problem, this CL changes the representation for unlinked fields. Instead of being placed in the parent's unknown fields, we create an actual message instance for each unlinked message we parse, but we use a placeholder "empty message" MiniTable as the message's type. All of the message's data will therefore be placed into the "empty message's" unknown fields. But unlike before, this "empty message" is actually present according to the hasbits, oneof case, and `upb_Array`/`upb_Map` of the parent. This means that all of the oneof presence logic works as normal. Since the MiniTable can be patched at any time, we need a bit in the message instance itself to signal whether a pointer to a sub-message is an "empty message" or not. When dynamic tree shaking is in use, all users must be capable of recognizing an empty message and acting accordingly (promoting, etc) even if the MiniTable itself says that the field is linked. Because dynamic tree shaking imposes this extra requirement on users, we require that users pass an extra option to the decoder to allow parsing of unlinked sub-messages. Many existing users of upb (Ruby, PHP, Python, etc) will always have fully-linked MiniTables, so there is no reason for them to add extra logic to handle empty messages. By omitting the `kUpb_DecodeOption_ExperimentalAllowUnlinked` option, they will be relieved of the duty to check the tagged pointer that would indicate an empty, unlinked message. For existing users of dynamic tree shaking, there are three main changes: 1. The APIs in message/promote.h have changed, and users will need to update to the new interfaces. 2. The model for maps has changed slightly. Before, we required that map entries always had their values linked; for dynamic tree shaking to apply to maps, we required that the *entry* was left unlinked, not the entry's value. In the new model, that is reversed: map entries must always be linked, but a map entry's value can be unlinked. 3. The presence model for unlinked fields has changed. Unlinked fields will now register as "present" from the perspective of hasbits, oneof cases, and array/map entries. Users must test the tagged pointer to know if a message is of the correct, linked type or whether it is a placeholder "empty" message. There is a new function `upb_Message_GetTaggedMessagePtr()`, as well as a new accessor `upb_MessageValue.tagged_msg_val` that can be used to read and test the tagged pointer directly. PiperOrigin-RevId: 535288031
2 years ago
return;
}
#ifndef NDEBUG
const upb_MiniTableField* oneof = upb_MiniTable_GetOneof(mt, field);
if (oneof) {
// All other members of the oneof must be message fields that are also
// unlinked.
do {
UPB_ASSERT(upb_MiniTableField_CType(oneof) == kUpb_CType_Message);
const upb_MiniTableSub* oneof_sub =
&mt->UPB_PRIVATE(subs)[oneof->UPB_PRIVATE(submsg_index)];
UPB_ASSERT(!oneof_sub);
} while (upb_MiniTable_NextOneofField(mt, &oneof));
}
#endif // NDEBUG
*op = kUpb_DecodeOp_UnknownField;
}
Added a new dynamic tree shaking model to upb, with the intention of removing the old model once YouTube has migrated. The `kUpb_DecodeOption_ExperimentalAllowUnlinked` flag to the decoder will enable the new behavior. When that flag is not passed, tree shaking with the old model will still be possible. "Dynamic tree shaking" in upb is a feature that allows messages to be parsed even if the MiniTables have not been fully linked. Unlinked sub-message fields can be parsed by preserving their data in the unknown fields. If the application later discovers that the message field is actually needed, the MiniTable can be patched to properly link that field, and existing message instances can "promote" the data from the unknown fields to an actual message of the correct type. Before this change, dynamic tree shaking stored unparsed message data in the unknown fields of the *parent*. In effect, we were treating the field as if it did not exist at all. This meant that parsing an unlinked field did not affect the hasbits or oneof cases of the parent, nor did it create a `upb_Array` or `upb_Map` for array/map fields. Only when a message was linked and promoted did any of these things occur. While this model had some amount of conceptual simplicity, it caused significant problems with oneofs. When multiple fields inside a single oneof are parsed from the wire, order matters, because later oneof fields must overwrite earlier ones. Dynamic tree shaking can mean that some fields in a oneof are linked while others are not. It is essential that we preserve this ordering semantic even when dynamic tree shaking is being used, but it is difficult to do if the oneof's data can be split between linked fields (which have been reified into parsed field data) and unlinked fields (whose data lives in the unknown fields of the parent). To solve this problem, this CL changes the representation for unlinked fields. Instead of being placed in the parent's unknown fields, we create an actual message instance for each unlinked message we parse, but we use a placeholder "empty message" MiniTable as the message's type. All of the message's data will therefore be placed into the "empty message's" unknown fields. But unlike before, this "empty message" is actually present according to the hasbits, oneof case, and `upb_Array`/`upb_Map` of the parent. This means that all of the oneof presence logic works as normal. Since the MiniTable can be patched at any time, we need a bit in the message instance itself to signal whether a pointer to a sub-message is an "empty message" or not. When dynamic tree shaking is in use, all users must be capable of recognizing an empty message and acting accordingly (promoting, etc) even if the MiniTable itself says that the field is linked. Because dynamic tree shaking imposes this extra requirement on users, we require that users pass an extra option to the decoder to allow parsing of unlinked sub-messages. Many existing users of upb (Ruby, PHP, Python, etc) will always have fully-linked MiniTables, so there is no reason for them to add extra logic to handle empty messages. By omitting the `kUpb_DecodeOption_ExperimentalAllowUnlinked` option, they will be relieved of the duty to check the tagged pointer that would indicate an empty, unlinked message. For existing users of dynamic tree shaking, there are three main changes: 1. The APIs in message/promote.h have changed, and users will need to update to the new interfaces. 2. The model for maps has changed slightly. Before, we required that map entries always had their values linked; for dynamic tree shaking to apply to maps, we required that the *entry* was left unlinked, not the entry's value. In the new model, that is reversed: map entries must always be linked, but a map entry's value can be unlinked. 3. The presence model for unlinked fields has changed. Unlinked fields will now register as "present" from the perspective of hasbits, oneof cases, and array/map entries. Users must test the tagged pointer to know if a message is of the correct, linked type or whether it is a placeholder "empty" message. There is a new function `upb_Message_GetTaggedMessagePtr()`, as well as a new accessor `upb_MessageValue.tagged_msg_val` that can be used to read and test the tagged pointer directly. PiperOrigin-RevId: 535288031
2 years ago
int _upb_Decoder_GetDelimitedOp(upb_Decoder* d, const upb_MiniTable* mt,
const upb_MiniTableField* field) {
enum { kRepeatedBase = 19 };
static const int8_t kDelimitedOps[] = {
/* For non-repeated field type. */
[kUpb_FakeFieldType_FieldNotFound] =
kUpb_DecodeOp_UnknownField, // Field not found.
[kUpb_FieldType_Double] = kUpb_DecodeOp_UnknownField,
[kUpb_FieldType_Float] = kUpb_DecodeOp_UnknownField,
[kUpb_FieldType_Int64] = kUpb_DecodeOp_UnknownField,
[kUpb_FieldType_UInt64] = kUpb_DecodeOp_UnknownField,
[kUpb_FieldType_Int32] = kUpb_DecodeOp_UnknownField,
[kUpb_FieldType_Fixed64] = kUpb_DecodeOp_UnknownField,
[kUpb_FieldType_Fixed32] = kUpb_DecodeOp_UnknownField,
[kUpb_FieldType_Bool] = kUpb_DecodeOp_UnknownField,
[kUpb_FieldType_String] = kUpb_DecodeOp_String,
[kUpb_FieldType_Group] = kUpb_DecodeOp_UnknownField,
[kUpb_FieldType_Message] = kUpb_DecodeOp_SubMessage,
[kUpb_FieldType_Bytes] = kUpb_DecodeOp_Bytes,
[kUpb_FieldType_UInt32] = kUpb_DecodeOp_UnknownField,
[kUpb_FieldType_Enum] = kUpb_DecodeOp_UnknownField,
[kUpb_FieldType_SFixed32] = kUpb_DecodeOp_UnknownField,
[kUpb_FieldType_SFixed64] = kUpb_DecodeOp_UnknownField,
[kUpb_FieldType_SInt32] = kUpb_DecodeOp_UnknownField,
[kUpb_FieldType_SInt64] = kUpb_DecodeOp_UnknownField,
[kUpb_FakeFieldType_MessageSetItem] = kUpb_DecodeOp_UnknownField,
// For repeated field type. */
[kRepeatedBase + kUpb_FieldType_Double] = OP_FIXPCK_LG2(3),
[kRepeatedBase + kUpb_FieldType_Float] = OP_FIXPCK_LG2(2),
[kRepeatedBase + kUpb_FieldType_Int64] = OP_VARPCK_LG2(3),
[kRepeatedBase + kUpb_FieldType_UInt64] = OP_VARPCK_LG2(3),
[kRepeatedBase + kUpb_FieldType_Int32] = OP_VARPCK_LG2(2),
[kRepeatedBase + kUpb_FieldType_Fixed64] = OP_FIXPCK_LG2(3),
[kRepeatedBase + kUpb_FieldType_Fixed32] = OP_FIXPCK_LG2(2),
[kRepeatedBase + kUpb_FieldType_Bool] = OP_VARPCK_LG2(0),
[kRepeatedBase + kUpb_FieldType_String] = kUpb_DecodeOp_String,
[kRepeatedBase + kUpb_FieldType_Group] = kUpb_DecodeOp_SubMessage,
[kRepeatedBase + kUpb_FieldType_Message] = kUpb_DecodeOp_SubMessage,
[kRepeatedBase + kUpb_FieldType_Bytes] = kUpb_DecodeOp_Bytes,
[kRepeatedBase + kUpb_FieldType_UInt32] = OP_VARPCK_LG2(2),
[kRepeatedBase + kUpb_FieldType_Enum] = kUpb_DecodeOp_PackedEnum,
[kRepeatedBase + kUpb_FieldType_SFixed32] = OP_FIXPCK_LG2(2),
[kRepeatedBase + kUpb_FieldType_SFixed64] = OP_FIXPCK_LG2(3),
[kRepeatedBase + kUpb_FieldType_SInt32] = OP_VARPCK_LG2(2),
[kRepeatedBase + kUpb_FieldType_SInt64] = OP_VARPCK_LG2(3),
// Omitting kUpb_FakeFieldType_MessageSetItem, because we never emit a
// repeated msgset type
};
int ndx = field->UPB_PRIVATE(descriptortype);
if (upb_MiniTableField_IsArray(field)) ndx += kRepeatedBase;
Treat unlinked sub-messages in the MiniTable as unknown This is an observable behavior change in the decoder. After submitting this CL, clients of the decoder can assume that any unlinked sub-messages will be treated as unknown, rather than crashing. Unlinked sub-messages must never have values present in the message. We can verify this with asserts. Since the values are never set, the encoder should never encounter data for any unlinked sub-message. ``` name old cpu/op new cpu/op delta BM_ArenaOneAlloc 18.3ns ± 9% 17.9ns ± 2% ~ (p=0.690 n=5+5) BM_ArenaInitialBlockOneAlloc 6.40ns ± 1% 6.68ns ±10% ~ (p=0.730 n=4+5) BM_LoadAdsDescriptor_Upb<NoLayout> 5.09ms ± 2% 5.03ms ± 3% ~ (p=0.222 n=5+5) BM_LoadAdsDescriptor_Upb<WithLayout> 5.45ms ± 3% 5.43ms ± 1% ~ (p=0.905 n=5+4) BM_LoadAdsDescriptor_Proto2<NoLayout> 10.9ms ± 1% 10.8ms ± 1% -1.09% (p=0.016 n=5+4) BM_LoadAdsDescriptor_Proto2<WithLayout> 11.3ms ± 9% 11.1ms ± 3% ~ (p=0.841 n=5+5) BM_Parse_Upb_FileDesc<UseArena, Copy> 11.2µs ± 3% 11.3µs ± 3% ~ (p=0.222 n=5+5) BM_Parse_Upb_FileDesc<UseArena, Alias> 10.3µs ± 5% 10.5µs ± 5% ~ (p=0.310 n=5+5) BM_Parse_Upb_FileDesc<InitBlock, Copy> 11.4µs ±18% 11.0µs ± 2% ~ (p=1.000 n=5+5) BM_Parse_Upb_FileDesc<InitBlock, Alias> 10.5µs ±17% 10.6µs ±19% ~ (p=0.421 n=5+5) BM_Parse_Proto2<FileDesc, NoArena, Copy> 20.5µs ± 2% 20.2µs ± 2% ~ (p=0.222 n=5+5) BM_Parse_Proto2<FileDesc, UseArena, Copy> 10.8µs ± 2% 10.9µs ± 4% ~ (p=0.841 n=5+5) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 10.5µs ± 3% 10.6µs ± 3% ~ (p=0.690 n=5+5) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 9.22µs ± 2% 9.23µs ± 3% ~ (p=1.000 n=5+5) BM_SerializeDescriptor_Proto2 6.05µs ± 3% 5.90µs ± 3% ~ (p=0.222 n=5+5) BM_SerializeDescriptor_Upb 10.2µs ± 3% 10.6µs ±14% ~ (p=0.841 n=5+5) name old time/op new time/op delta BM_ArenaOneAlloc 18.3ns ± 9% 17.9ns ± 2% ~ (p=0.841 n=5+5) BM_ArenaInitialBlockOneAlloc 6.42ns ± 1% 6.69ns ±10% ~ (p=0.730 n=4+5) BM_LoadAdsDescriptor_Upb<NoLayout> 5.10ms ± 2% 5.05ms ± 3% ~ (p=0.222 n=5+5) BM_LoadAdsDescriptor_Upb<WithLayout> 5.47ms ± 3% 5.45ms ± 1% ~ (p=0.905 n=5+4) BM_LoadAdsDescriptor_Proto2<NoLayout> 10.9ms ± 1% 10.8ms ± 1% -1.11% (p=0.016 n=5+4) BM_LoadAdsDescriptor_Proto2<WithLayout> 11.4ms ± 9% 11.1ms ± 3% ~ (p=0.841 n=5+5) BM_Parse_Upb_FileDesc<UseArena, Copy> 11.2µs ± 3% 11.3µs ± 3% ~ (p=0.222 n=5+5) BM_Parse_Upb_FileDesc<UseArena, Alias> 10.3µs ± 5% 10.5µs ± 5% ~ (p=0.151 n=5+5) BM_Parse_Upb_FileDesc<InitBlock, Copy> 11.5µs ±18% 11.0µs ± 2% ~ (p=1.000 n=5+5) BM_Parse_Upb_FileDesc<InitBlock, Alias> 10.5µs ±17% 10.7µs ±19% ~ (p=0.421 n=5+5) BM_Parse_Proto2<FileDesc, NoArena, Copy> 20.6µs ± 2% 20.3µs ± 2% ~ (p=0.222 n=5+5) BM_Parse_Proto2<FileDesc, UseArena, Copy> 10.9µs ± 2% 10.9µs ± 4% ~ (p=0.841 n=5+5) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 10.6µs ± 3% 10.6µs ± 3% ~ (p=0.690 n=5+5) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 9.24µs ± 2% 9.25µs ± 3% ~ (p=1.000 n=5+5) BM_SerializeDescriptor_Proto2 6.07µs ± 3% 5.91µs ± 3% ~ (p=0.222 n=5+5) BM_SerializeDescriptor_Upb 10.3µs ± 3% 10.6µs ±14% ~ (p=0.841 n=5+5) name old INSTRUCTIONS/op new INSTRUCTIONS/op delta BM_ArenaOneAlloc 201 ± 0% 201 ± 0% ~ (p=0.841 n=5+5) BM_ArenaInitialBlockOneAlloc 69.0 ± 0% 69.0 ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Upb<NoLayout> 33.9M ± 0% 34.1M ± 0% +0.66% (p=0.008 n=5+5) BM_LoadAdsDescriptor_Upb<WithLayout> 35.6M ± 0% 35.8M ± 0% +0.64% (p=0.008 n=5+5) BM_LoadAdsDescriptor_Proto2<NoLayout> 70.8M ± 0% 70.8M ± 0% ~ (p=0.548 n=5+5) BM_LoadAdsDescriptor_Proto2<WithLayout> 71.6M ± 0% 71.6M ± 0% ~ (p=0.151 n=5+5) BM_Parse_Upb_FileDesc<UseArena, Copy> 137k ± 0% 141k ± 0% +2.87% (p=0.008 n=5+5) BM_Parse_Upb_FileDesc<UseArena, Alias> 125k ± 0% 128k ± 0% +2.83% (p=0.008 n=5+5) BM_Parse_Upb_FileDesc<InitBlock, Copy> 135k ± 0% 139k ± 0% +2.89% (p=0.008 n=5+5) BM_Parse_Upb_FileDesc<InitBlock, Alias> 124k ± 0% 127k ± 0% +2.85% (p=0.016 n=5+4) BM_Parse_Proto2<FileDesc, NoArena, Copy> 201k ± 0% 201k ± 0% ~ (p=0.222 n=5+5) BM_Parse_Proto2<FileDesc, UseArena, Copy> 107k ± 0% 107k ± 0% ~ (p=1.000 n=5+5) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 105k ± 0% 105k ± 0% ~ (p=0.286 n=5+4) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 86.5k ± 0% 86.5k ± 0% ~ (p=0.222 n=5+5) BM_SerializeDescriptor_Proto2 60.3k ± 0% 60.3k ± 0% ~ (p=0.071 n=5+5) BM_SerializeDescriptor_Upb 111k ± 0% 111k ± 0% ~ (p=0.841 n=5+5) name old CYCLES/op new CYCLES/op delta BM_ArenaOneAlloc 60.0 ± 7% 58.8 ± 0% -2.15% (p=0.016 n=5+5) BM_ArenaInitialBlockOneAlloc 21.0 ± 0% 21.0 ± 0% ~ (p=1.000 n=5+5) BM_LoadAdsDescriptor_Upb<NoLayout> 16.9M ± 0% 16.9M ± 0% ~ (p=0.056 n=5+5) BM_LoadAdsDescriptor_Upb<WithLayout> 17.9M ± 1% 18.0M ± 1% ~ (p=0.095 n=5+5) BM_LoadAdsDescriptor_Proto2<NoLayout> 35.9M ± 1% 35.8M ± 1% ~ (p=0.421 n=5+5) BM_LoadAdsDescriptor_Proto2<WithLayout> 36.5M ± 0% 36.5M ± 0% ~ (p=0.841 n=5+5) BM_Parse_Upb_FileDesc<UseArena, Copy> 37.2k ± 0% 37.3k ± 0% ~ (p=0.222 n=5+5) BM_Parse_Upb_FileDesc<UseArena, Alias> 34.1k ± 0% 34.7k ± 0% +1.66% (p=0.008 n=5+5) BM_Parse_Upb_FileDesc<InitBlock, Copy> 36.4k ± 0% 36.7k ± 0% +0.83% (p=0.008 n=5+5) BM_Parse_Upb_FileDesc<InitBlock, Alias> 33.3k ± 1% 34.1k ± 1% +2.39% (p=0.008 n=5+5) BM_Parse_Proto2<FileDesc, NoArena, Copy> 68.1k ± 1% 68.0k ± 1% ~ (p=0.421 n=5+5) BM_Parse_Proto2<FileDesc, UseArena, Copy> 36.0k ± 1% 36.1k ± 1% ~ (p=0.841 n=5+5) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 35.3k ± 1% 35.5k ± 1% ~ (p=0.151 n=5+5) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 30.7k ± 0% 30.9k ± 1% ~ (p=0.151 n=5+5) BM_SerializeDescriptor_Proto2 20.3k ± 2% 19.7k ± 3% ~ (p=0.151 n=5+5) BM_SerializeDescriptor_Upb 33.6k ± 0% 33.7k ± 2% ~ (p=1.000 n=5+5) name old allocs/op new allocs/op delta BM_ArenaOneAlloc 1.19 ± 0% 1.19 ± 0% ~ (all samples are equal) BM_ArenaInitialBlockOneAlloc 0.19 ± 0% 0.19 ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Upb<NoLayout> 6.00k ± 0% 6.00k ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Upb<WithLayout> 5.99k ± 0% 5.99k ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Proto2<NoLayout> 77.8k ± 0% 77.8k ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Proto2<WithLayout> 79.0k ± 0% 79.0k ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<UseArena, Copy> 7.19 ± 0% 7.19 ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<UseArena, Alias> 7.19 ± 0% 7.19 ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<InitBlock, Copy> 0.19 ± 0% 0.19 ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<InitBlock, Alias> 0.19 ± 0% 0.19 ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, NoArena, Copy> 765 ± 0% 765 ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, UseArena, Copy> 10.2 ± 0% 10.2 ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 1.19 ± 0% 1.19 ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 1.19 ± 0% 1.19 ± 0% ~ (all samples are equal) BM_SerializeDescriptor_Proto2 0.19 ± 0% 0.19 ± 0% ~ (all samples are equal) BM_SerializeDescriptor_Upb 0.19 ± 0% 0.19 ± 0% ~ (all samples are equal) name old peak-mem(Bytes)/op new peak-mem(Bytes)/op delta BM_ArenaOneAlloc 344 ± 0% 344 ± 0% ~ (all samples are equal) BM_ArenaInitialBlockOneAlloc 112 ± 0% 112 ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Upb<NoLayout> 9.64M ± 0% 9.64M ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Upb<WithLayout> 9.70M ± 0% 9.70M ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Proto2<NoLayout> 6.38M ± 0% 6.38M ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Proto2<WithLayout> 6.44M ± 0% 6.44M ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<UseArena, Copy> 36.5k ± 0% 36.5k ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<UseArena, Alias> 36.5k ± 0% 36.5k ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<InitBlock, Copy> 112 ± 0% 112 ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<InitBlock, Alias> 112 ± 0% 112 ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, NoArena, Copy> 35.8k ± 0% 35.8k ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, UseArena, Copy> 40.8k ± 0% 40.8k ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 112 ± 0% 112 ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 112 ± 0% 112 ± 0% ~ (all samples are equal) BM_SerializeDescriptor_Proto2 112 ± 0% 112 ± 0% ~ (all samples are equal) BM_SerializeDescriptor_Upb 112 ± 0% 112 ± 0% ~ (all samples are equal) name old speed new speed delta BM_LoadAdsDescriptor_Upb<NoLayout> 147MB/s ± 2% 148MB/s ± 3% ~ (p=0.222 n=5+5) BM_LoadAdsDescriptor_Upb<WithLayout> 137MB/s ± 3% 137MB/s ± 1% ~ (p=0.905 n=5+4) BM_LoadAdsDescriptor_Proto2<NoLayout> 68.6MB/s ± 1% 69.3MB/s ± 1% +1.10% (p=0.016 n=5+4) BM_LoadAdsDescriptor_Proto2<WithLayout> 66.0MB/s ± 9% 67.4MB/s ± 3% ~ (p=0.841 n=5+5) BM_Parse_Upb_FileDesc<UseArena, Copy> 675MB/s ± 3% 667MB/s ± 3% ~ (p=0.222 n=5+5) BM_Parse_Upb_FileDesc<UseArena, Alias> 730MB/s ± 5% 718MB/s ± 5% ~ (p=0.310 n=5+5) BM_Parse_Upb_FileDesc<InitBlock, Copy> 663MB/s ±16% 685MB/s ± 2% ~ (p=1.000 n=5+5) BM_Parse_Upb_FileDesc<InitBlock, Alias> 723MB/s ±15% 712MB/s ±16% ~ (p=0.421 n=5+5) BM_Parse_Proto2<FileDesc, NoArena, Copy> 367MB/s ± 2% 372MB/s ± 2% ~ (p=0.222 n=5+5) BM_Parse_Proto2<FileDesc, UseArena, Copy> 694MB/s ± 2% 691MB/s ± 4% ~ (p=0.841 n=5+5) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 714MB/s ± 3% 709MB/s ± 3% ~ (p=0.690 n=5+5) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 816MB/s ± 2% 816MB/s ± 3% ~ (p=1.000 n=5+5) BM_SerializeDescriptor_Proto2 1.24GB/s ± 3% 1.28GB/s ± 3% ~ (p=0.222 n=5+5) BM_SerializeDescriptor_Upb 734MB/s ± 3% 713MB/s ±13% ~ (p=0.841 n=5+5) ``` PiperOrigin-RevId: 477770562
2 years ago
int op = kDelimitedOps[ndx];
if (op == kUpb_DecodeOp_SubMessage) {
Added a new dynamic tree shaking model to upb, with the intention of removing the old model once YouTube has migrated. The `kUpb_DecodeOption_ExperimentalAllowUnlinked` flag to the decoder will enable the new behavior. When that flag is not passed, tree shaking with the old model will still be possible. "Dynamic tree shaking" in upb is a feature that allows messages to be parsed even if the MiniTables have not been fully linked. Unlinked sub-message fields can be parsed by preserving their data in the unknown fields. If the application later discovers that the message field is actually needed, the MiniTable can be patched to properly link that field, and existing message instances can "promote" the data from the unknown fields to an actual message of the correct type. Before this change, dynamic tree shaking stored unparsed message data in the unknown fields of the *parent*. In effect, we were treating the field as if it did not exist at all. This meant that parsing an unlinked field did not affect the hasbits or oneof cases of the parent, nor did it create a `upb_Array` or `upb_Map` for array/map fields. Only when a message was linked and promoted did any of these things occur. While this model had some amount of conceptual simplicity, it caused significant problems with oneofs. When multiple fields inside a single oneof are parsed from the wire, order matters, because later oneof fields must overwrite earlier ones. Dynamic tree shaking can mean that some fields in a oneof are linked while others are not. It is essential that we preserve this ordering semantic even when dynamic tree shaking is being used, but it is difficult to do if the oneof's data can be split between linked fields (which have been reified into parsed field data) and unlinked fields (whose data lives in the unknown fields of the parent). To solve this problem, this CL changes the representation for unlinked fields. Instead of being placed in the parent's unknown fields, we create an actual message instance for each unlinked message we parse, but we use a placeholder "empty message" MiniTable as the message's type. All of the message's data will therefore be placed into the "empty message's" unknown fields. But unlike before, this "empty message" is actually present according to the hasbits, oneof case, and `upb_Array`/`upb_Map` of the parent. This means that all of the oneof presence logic works as normal. Since the MiniTable can be patched at any time, we need a bit in the message instance itself to signal whether a pointer to a sub-message is an "empty message" or not. When dynamic tree shaking is in use, all users must be capable of recognizing an empty message and acting accordingly (promoting, etc) even if the MiniTable itself says that the field is linked. Because dynamic tree shaking imposes this extra requirement on users, we require that users pass an extra option to the decoder to allow parsing of unlinked sub-messages. Many existing users of upb (Ruby, PHP, Python, etc) will always have fully-linked MiniTables, so there is no reason for them to add extra logic to handle empty messages. By omitting the `kUpb_DecodeOption_ExperimentalAllowUnlinked` option, they will be relieved of the duty to check the tagged pointer that would indicate an empty, unlinked message. For existing users of dynamic tree shaking, there are three main changes: 1. The APIs in message/promote.h have changed, and users will need to update to the new interfaces. 2. The model for maps has changed slightly. Before, we required that map entries always had their values linked; for dynamic tree shaking to apply to maps, we required that the *entry* was left unlinked, not the entry's value. In the new model, that is reversed: map entries must always be linked, but a map entry's value can be unlinked. 3. The presence model for unlinked fields has changed. Unlinked fields will now register as "present" from the perspective of hasbits, oneof cases, and array/map entries. Users must test the tagged pointer to know if a message is of the correct, linked type or whether it is a placeholder "empty" message. There is a new function `upb_Message_GetTaggedMessagePtr()`, as well as a new accessor `upb_MessageValue.tagged_msg_val` that can be used to read and test the tagged pointer directly. PiperOrigin-RevId: 535288031
2 years ago
_upb_Decoder_CheckUnlinked(d, mt, field, &op);
Treat unlinked sub-messages in the MiniTable as unknown This is an observable behavior change in the decoder. After submitting this CL, clients of the decoder can assume that any unlinked sub-messages will be treated as unknown, rather than crashing. Unlinked sub-messages must never have values present in the message. We can verify this with asserts. Since the values are never set, the encoder should never encounter data for any unlinked sub-message. ``` name old cpu/op new cpu/op delta BM_ArenaOneAlloc 18.3ns ± 9% 17.9ns ± 2% ~ (p=0.690 n=5+5) BM_ArenaInitialBlockOneAlloc 6.40ns ± 1% 6.68ns ±10% ~ (p=0.730 n=4+5) BM_LoadAdsDescriptor_Upb<NoLayout> 5.09ms ± 2% 5.03ms ± 3% ~ (p=0.222 n=5+5) BM_LoadAdsDescriptor_Upb<WithLayout> 5.45ms ± 3% 5.43ms ± 1% ~ (p=0.905 n=5+4) BM_LoadAdsDescriptor_Proto2<NoLayout> 10.9ms ± 1% 10.8ms ± 1% -1.09% (p=0.016 n=5+4) BM_LoadAdsDescriptor_Proto2<WithLayout> 11.3ms ± 9% 11.1ms ± 3% ~ (p=0.841 n=5+5) BM_Parse_Upb_FileDesc<UseArena, Copy> 11.2µs ± 3% 11.3µs ± 3% ~ (p=0.222 n=5+5) BM_Parse_Upb_FileDesc<UseArena, Alias> 10.3µs ± 5% 10.5µs ± 5% ~ (p=0.310 n=5+5) BM_Parse_Upb_FileDesc<InitBlock, Copy> 11.4µs ±18% 11.0µs ± 2% ~ (p=1.000 n=5+5) BM_Parse_Upb_FileDesc<InitBlock, Alias> 10.5µs ±17% 10.6µs ±19% ~ (p=0.421 n=5+5) BM_Parse_Proto2<FileDesc, NoArena, Copy> 20.5µs ± 2% 20.2µs ± 2% ~ (p=0.222 n=5+5) BM_Parse_Proto2<FileDesc, UseArena, Copy> 10.8µs ± 2% 10.9µs ± 4% ~ (p=0.841 n=5+5) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 10.5µs ± 3% 10.6µs ± 3% ~ (p=0.690 n=5+5) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 9.22µs ± 2% 9.23µs ± 3% ~ (p=1.000 n=5+5) BM_SerializeDescriptor_Proto2 6.05µs ± 3% 5.90µs ± 3% ~ (p=0.222 n=5+5) BM_SerializeDescriptor_Upb 10.2µs ± 3% 10.6µs ±14% ~ (p=0.841 n=5+5) name old time/op new time/op delta BM_ArenaOneAlloc 18.3ns ± 9% 17.9ns ± 2% ~ (p=0.841 n=5+5) BM_ArenaInitialBlockOneAlloc 6.42ns ± 1% 6.69ns ±10% ~ (p=0.730 n=4+5) BM_LoadAdsDescriptor_Upb<NoLayout> 5.10ms ± 2% 5.05ms ± 3% ~ (p=0.222 n=5+5) BM_LoadAdsDescriptor_Upb<WithLayout> 5.47ms ± 3% 5.45ms ± 1% ~ (p=0.905 n=5+4) BM_LoadAdsDescriptor_Proto2<NoLayout> 10.9ms ± 1% 10.8ms ± 1% -1.11% (p=0.016 n=5+4) BM_LoadAdsDescriptor_Proto2<WithLayout> 11.4ms ± 9% 11.1ms ± 3% ~ (p=0.841 n=5+5) BM_Parse_Upb_FileDesc<UseArena, Copy> 11.2µs ± 3% 11.3µs ± 3% ~ (p=0.222 n=5+5) BM_Parse_Upb_FileDesc<UseArena, Alias> 10.3µs ± 5% 10.5µs ± 5% ~ (p=0.151 n=5+5) BM_Parse_Upb_FileDesc<InitBlock, Copy> 11.5µs ±18% 11.0µs ± 2% ~ (p=1.000 n=5+5) BM_Parse_Upb_FileDesc<InitBlock, Alias> 10.5µs ±17% 10.7µs ±19% ~ (p=0.421 n=5+5) BM_Parse_Proto2<FileDesc, NoArena, Copy> 20.6µs ± 2% 20.3µs ± 2% ~ (p=0.222 n=5+5) BM_Parse_Proto2<FileDesc, UseArena, Copy> 10.9µs ± 2% 10.9µs ± 4% ~ (p=0.841 n=5+5) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 10.6µs ± 3% 10.6µs ± 3% ~ (p=0.690 n=5+5) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 9.24µs ± 2% 9.25µs ± 3% ~ (p=1.000 n=5+5) BM_SerializeDescriptor_Proto2 6.07µs ± 3% 5.91µs ± 3% ~ (p=0.222 n=5+5) BM_SerializeDescriptor_Upb 10.3µs ± 3% 10.6µs ±14% ~ (p=0.841 n=5+5) name old INSTRUCTIONS/op new INSTRUCTIONS/op delta BM_ArenaOneAlloc 201 ± 0% 201 ± 0% ~ (p=0.841 n=5+5) BM_ArenaInitialBlockOneAlloc 69.0 ± 0% 69.0 ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Upb<NoLayout> 33.9M ± 0% 34.1M ± 0% +0.66% (p=0.008 n=5+5) BM_LoadAdsDescriptor_Upb<WithLayout> 35.6M ± 0% 35.8M ± 0% +0.64% (p=0.008 n=5+5) BM_LoadAdsDescriptor_Proto2<NoLayout> 70.8M ± 0% 70.8M ± 0% ~ (p=0.548 n=5+5) BM_LoadAdsDescriptor_Proto2<WithLayout> 71.6M ± 0% 71.6M ± 0% ~ (p=0.151 n=5+5) BM_Parse_Upb_FileDesc<UseArena, Copy> 137k ± 0% 141k ± 0% +2.87% (p=0.008 n=5+5) BM_Parse_Upb_FileDesc<UseArena, Alias> 125k ± 0% 128k ± 0% +2.83% (p=0.008 n=5+5) BM_Parse_Upb_FileDesc<InitBlock, Copy> 135k ± 0% 139k ± 0% +2.89% (p=0.008 n=5+5) BM_Parse_Upb_FileDesc<InitBlock, Alias> 124k ± 0% 127k ± 0% +2.85% (p=0.016 n=5+4) BM_Parse_Proto2<FileDesc, NoArena, Copy> 201k ± 0% 201k ± 0% ~ (p=0.222 n=5+5) BM_Parse_Proto2<FileDesc, UseArena, Copy> 107k ± 0% 107k ± 0% ~ (p=1.000 n=5+5) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 105k ± 0% 105k ± 0% ~ (p=0.286 n=5+4) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 86.5k ± 0% 86.5k ± 0% ~ (p=0.222 n=5+5) BM_SerializeDescriptor_Proto2 60.3k ± 0% 60.3k ± 0% ~ (p=0.071 n=5+5) BM_SerializeDescriptor_Upb 111k ± 0% 111k ± 0% ~ (p=0.841 n=5+5) name old CYCLES/op new CYCLES/op delta BM_ArenaOneAlloc 60.0 ± 7% 58.8 ± 0% -2.15% (p=0.016 n=5+5) BM_ArenaInitialBlockOneAlloc 21.0 ± 0% 21.0 ± 0% ~ (p=1.000 n=5+5) BM_LoadAdsDescriptor_Upb<NoLayout> 16.9M ± 0% 16.9M ± 0% ~ (p=0.056 n=5+5) BM_LoadAdsDescriptor_Upb<WithLayout> 17.9M ± 1% 18.0M ± 1% ~ (p=0.095 n=5+5) BM_LoadAdsDescriptor_Proto2<NoLayout> 35.9M ± 1% 35.8M ± 1% ~ (p=0.421 n=5+5) BM_LoadAdsDescriptor_Proto2<WithLayout> 36.5M ± 0% 36.5M ± 0% ~ (p=0.841 n=5+5) BM_Parse_Upb_FileDesc<UseArena, Copy> 37.2k ± 0% 37.3k ± 0% ~ (p=0.222 n=5+5) BM_Parse_Upb_FileDesc<UseArena, Alias> 34.1k ± 0% 34.7k ± 0% +1.66% (p=0.008 n=5+5) BM_Parse_Upb_FileDesc<InitBlock, Copy> 36.4k ± 0% 36.7k ± 0% +0.83% (p=0.008 n=5+5) BM_Parse_Upb_FileDesc<InitBlock, Alias> 33.3k ± 1% 34.1k ± 1% +2.39% (p=0.008 n=5+5) BM_Parse_Proto2<FileDesc, NoArena, Copy> 68.1k ± 1% 68.0k ± 1% ~ (p=0.421 n=5+5) BM_Parse_Proto2<FileDesc, UseArena, Copy> 36.0k ± 1% 36.1k ± 1% ~ (p=0.841 n=5+5) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 35.3k ± 1% 35.5k ± 1% ~ (p=0.151 n=5+5) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 30.7k ± 0% 30.9k ± 1% ~ (p=0.151 n=5+5) BM_SerializeDescriptor_Proto2 20.3k ± 2% 19.7k ± 3% ~ (p=0.151 n=5+5) BM_SerializeDescriptor_Upb 33.6k ± 0% 33.7k ± 2% ~ (p=1.000 n=5+5) name old allocs/op new allocs/op delta BM_ArenaOneAlloc 1.19 ± 0% 1.19 ± 0% ~ (all samples are equal) BM_ArenaInitialBlockOneAlloc 0.19 ± 0% 0.19 ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Upb<NoLayout> 6.00k ± 0% 6.00k ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Upb<WithLayout> 5.99k ± 0% 5.99k ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Proto2<NoLayout> 77.8k ± 0% 77.8k ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Proto2<WithLayout> 79.0k ± 0% 79.0k ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<UseArena, Copy> 7.19 ± 0% 7.19 ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<UseArena, Alias> 7.19 ± 0% 7.19 ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<InitBlock, Copy> 0.19 ± 0% 0.19 ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<InitBlock, Alias> 0.19 ± 0% 0.19 ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, NoArena, Copy> 765 ± 0% 765 ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, UseArena, Copy> 10.2 ± 0% 10.2 ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 1.19 ± 0% 1.19 ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 1.19 ± 0% 1.19 ± 0% ~ (all samples are equal) BM_SerializeDescriptor_Proto2 0.19 ± 0% 0.19 ± 0% ~ (all samples are equal) BM_SerializeDescriptor_Upb 0.19 ± 0% 0.19 ± 0% ~ (all samples are equal) name old peak-mem(Bytes)/op new peak-mem(Bytes)/op delta BM_ArenaOneAlloc 344 ± 0% 344 ± 0% ~ (all samples are equal) BM_ArenaInitialBlockOneAlloc 112 ± 0% 112 ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Upb<NoLayout> 9.64M ± 0% 9.64M ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Upb<WithLayout> 9.70M ± 0% 9.70M ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Proto2<NoLayout> 6.38M ± 0% 6.38M ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Proto2<WithLayout> 6.44M ± 0% 6.44M ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<UseArena, Copy> 36.5k ± 0% 36.5k ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<UseArena, Alias> 36.5k ± 0% 36.5k ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<InitBlock, Copy> 112 ± 0% 112 ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<InitBlock, Alias> 112 ± 0% 112 ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, NoArena, Copy> 35.8k ± 0% 35.8k ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, UseArena, Copy> 40.8k ± 0% 40.8k ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 112 ± 0% 112 ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 112 ± 0% 112 ± 0% ~ (all samples are equal) BM_SerializeDescriptor_Proto2 112 ± 0% 112 ± 0% ~ (all samples are equal) BM_SerializeDescriptor_Upb 112 ± 0% 112 ± 0% ~ (all samples are equal) name old speed new speed delta BM_LoadAdsDescriptor_Upb<NoLayout> 147MB/s ± 2% 148MB/s ± 3% ~ (p=0.222 n=5+5) BM_LoadAdsDescriptor_Upb<WithLayout> 137MB/s ± 3% 137MB/s ± 1% ~ (p=0.905 n=5+4) BM_LoadAdsDescriptor_Proto2<NoLayout> 68.6MB/s ± 1% 69.3MB/s ± 1% +1.10% (p=0.016 n=5+4) BM_LoadAdsDescriptor_Proto2<WithLayout> 66.0MB/s ± 9% 67.4MB/s ± 3% ~ (p=0.841 n=5+5) BM_Parse_Upb_FileDesc<UseArena, Copy> 675MB/s ± 3% 667MB/s ± 3% ~ (p=0.222 n=5+5) BM_Parse_Upb_FileDesc<UseArena, Alias> 730MB/s ± 5% 718MB/s ± 5% ~ (p=0.310 n=5+5) BM_Parse_Upb_FileDesc<InitBlock, Copy> 663MB/s ±16% 685MB/s ± 2% ~ (p=1.000 n=5+5) BM_Parse_Upb_FileDesc<InitBlock, Alias> 723MB/s ±15% 712MB/s ±16% ~ (p=0.421 n=5+5) BM_Parse_Proto2<FileDesc, NoArena, Copy> 367MB/s ± 2% 372MB/s ± 2% ~ (p=0.222 n=5+5) BM_Parse_Proto2<FileDesc, UseArena, Copy> 694MB/s ± 2% 691MB/s ± 4% ~ (p=0.841 n=5+5) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 714MB/s ± 3% 709MB/s ± 3% ~ (p=0.690 n=5+5) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 816MB/s ± 2% 816MB/s ± 3% ~ (p=1.000 n=5+5) BM_SerializeDescriptor_Proto2 1.24GB/s ± 3% 1.28GB/s ± 3% ~ (p=0.222 n=5+5) BM_SerializeDescriptor_Upb 734MB/s ± 3% 713MB/s ±13% ~ (p=0.841 n=5+5) ``` PiperOrigin-RevId: 477770562
2 years ago
}
return op;
}
UPB_FORCEINLINE
static const char* _upb_Decoder_DecodeWireValue(upb_Decoder* d, const char* ptr,
const upb_MiniTable* mt,
const upb_MiniTableField* field,
int wire_type, wireval* val,
int* op) {
static const unsigned kFixed32OkMask = (1 << kUpb_FieldType_Float) |
(1 << kUpb_FieldType_Fixed32) |
(1 << kUpb_FieldType_SFixed32);
static const unsigned kFixed64OkMask = (1 << kUpb_FieldType_Double) |
(1 << kUpb_FieldType_Fixed64) |
(1 << kUpb_FieldType_SFixed64);
switch (wire_type) {
case kUpb_WireType_Varint:
ptr = _upb_Decoder_DecodeVarint(d, ptr, &val->uint64_val);
*op = _upb_Decoder_GetVarintOp(field);
_upb_Decoder_Munge(field->UPB_PRIVATE(descriptortype), val);
return ptr;
case kUpb_WireType_32Bit:
*op = kUpb_DecodeOp_Scalar4Byte;
if (((1 << field->UPB_PRIVATE(descriptortype)) & kFixed32OkMask) == 0) {
*op = kUpb_DecodeOp_UnknownField;
}
return upb_WireReader_ReadFixed32(ptr, &val->uint32_val);
case kUpb_WireType_64Bit:
*op = kUpb_DecodeOp_Scalar8Byte;
if (((1 << field->UPB_PRIVATE(descriptortype)) & kFixed64OkMask) == 0) {
*op = kUpb_DecodeOp_UnknownField;
}
return upb_WireReader_ReadFixed64(ptr, &val->uint64_val);
case kUpb_WireType_Delimited:
ptr = upb_Decoder_DecodeSize(d, ptr, &val->size);
Added a new dynamic tree shaking model to upb, with the intention of removing the old model once YouTube has migrated. The `kUpb_DecodeOption_ExperimentalAllowUnlinked` flag to the decoder will enable the new behavior. When that flag is not passed, tree shaking with the old model will still be possible. "Dynamic tree shaking" in upb is a feature that allows messages to be parsed even if the MiniTables have not been fully linked. Unlinked sub-message fields can be parsed by preserving their data in the unknown fields. If the application later discovers that the message field is actually needed, the MiniTable can be patched to properly link that field, and existing message instances can "promote" the data from the unknown fields to an actual message of the correct type. Before this change, dynamic tree shaking stored unparsed message data in the unknown fields of the *parent*. In effect, we were treating the field as if it did not exist at all. This meant that parsing an unlinked field did not affect the hasbits or oneof cases of the parent, nor did it create a `upb_Array` or `upb_Map` for array/map fields. Only when a message was linked and promoted did any of these things occur. While this model had some amount of conceptual simplicity, it caused significant problems with oneofs. When multiple fields inside a single oneof are parsed from the wire, order matters, because later oneof fields must overwrite earlier ones. Dynamic tree shaking can mean that some fields in a oneof are linked while others are not. It is essential that we preserve this ordering semantic even when dynamic tree shaking is being used, but it is difficult to do if the oneof's data can be split between linked fields (which have been reified into parsed field data) and unlinked fields (whose data lives in the unknown fields of the parent). To solve this problem, this CL changes the representation for unlinked fields. Instead of being placed in the parent's unknown fields, we create an actual message instance for each unlinked message we parse, but we use a placeholder "empty message" MiniTable as the message's type. All of the message's data will therefore be placed into the "empty message's" unknown fields. But unlike before, this "empty message" is actually present according to the hasbits, oneof case, and `upb_Array`/`upb_Map` of the parent. This means that all of the oneof presence logic works as normal. Since the MiniTable can be patched at any time, we need a bit in the message instance itself to signal whether a pointer to a sub-message is an "empty message" or not. When dynamic tree shaking is in use, all users must be capable of recognizing an empty message and acting accordingly (promoting, etc) even if the MiniTable itself says that the field is linked. Because dynamic tree shaking imposes this extra requirement on users, we require that users pass an extra option to the decoder to allow parsing of unlinked sub-messages. Many existing users of upb (Ruby, PHP, Python, etc) will always have fully-linked MiniTables, so there is no reason for them to add extra logic to handle empty messages. By omitting the `kUpb_DecodeOption_ExperimentalAllowUnlinked` option, they will be relieved of the duty to check the tagged pointer that would indicate an empty, unlinked message. For existing users of dynamic tree shaking, there are three main changes: 1. The APIs in message/promote.h have changed, and users will need to update to the new interfaces. 2. The model for maps has changed slightly. Before, we required that map entries always had their values linked; for dynamic tree shaking to apply to maps, we required that the *entry* was left unlinked, not the entry's value. In the new model, that is reversed: map entries must always be linked, but a map entry's value can be unlinked. 3. The presence model for unlinked fields has changed. Unlinked fields will now register as "present" from the perspective of hasbits, oneof cases, and array/map entries. Users must test the tagged pointer to know if a message is of the correct, linked type or whether it is a placeholder "empty" message. There is a new function `upb_Message_GetTaggedMessagePtr()`, as well as a new accessor `upb_MessageValue.tagged_msg_val` that can be used to read and test the tagged pointer directly. PiperOrigin-RevId: 535288031
2 years ago
*op = _upb_Decoder_GetDelimitedOp(d, mt, field);
return ptr;
case kUpb_WireType_StartGroup:
val->uint32_val = field->UPB_PRIVATE(number);
if (field->UPB_PRIVATE(descriptortype) == kUpb_FieldType_Group) {
*op = kUpb_DecodeOp_SubMessage;
Added a new dynamic tree shaking model to upb, with the intention of removing the old model once YouTube has migrated. The `kUpb_DecodeOption_ExperimentalAllowUnlinked` flag to the decoder will enable the new behavior. When that flag is not passed, tree shaking with the old model will still be possible. "Dynamic tree shaking" in upb is a feature that allows messages to be parsed even if the MiniTables have not been fully linked. Unlinked sub-message fields can be parsed by preserving their data in the unknown fields. If the application later discovers that the message field is actually needed, the MiniTable can be patched to properly link that field, and existing message instances can "promote" the data from the unknown fields to an actual message of the correct type. Before this change, dynamic tree shaking stored unparsed message data in the unknown fields of the *parent*. In effect, we were treating the field as if it did not exist at all. This meant that parsing an unlinked field did not affect the hasbits or oneof cases of the parent, nor did it create a `upb_Array` or `upb_Map` for array/map fields. Only when a message was linked and promoted did any of these things occur. While this model had some amount of conceptual simplicity, it caused significant problems with oneofs. When multiple fields inside a single oneof are parsed from the wire, order matters, because later oneof fields must overwrite earlier ones. Dynamic tree shaking can mean that some fields in a oneof are linked while others are not. It is essential that we preserve this ordering semantic even when dynamic tree shaking is being used, but it is difficult to do if the oneof's data can be split between linked fields (which have been reified into parsed field data) and unlinked fields (whose data lives in the unknown fields of the parent). To solve this problem, this CL changes the representation for unlinked fields. Instead of being placed in the parent's unknown fields, we create an actual message instance for each unlinked message we parse, but we use a placeholder "empty message" MiniTable as the message's type. All of the message's data will therefore be placed into the "empty message's" unknown fields. But unlike before, this "empty message" is actually present according to the hasbits, oneof case, and `upb_Array`/`upb_Map` of the parent. This means that all of the oneof presence logic works as normal. Since the MiniTable can be patched at any time, we need a bit in the message instance itself to signal whether a pointer to a sub-message is an "empty message" or not. When dynamic tree shaking is in use, all users must be capable of recognizing an empty message and acting accordingly (promoting, etc) even if the MiniTable itself says that the field is linked. Because dynamic tree shaking imposes this extra requirement on users, we require that users pass an extra option to the decoder to allow parsing of unlinked sub-messages. Many existing users of upb (Ruby, PHP, Python, etc) will always have fully-linked MiniTables, so there is no reason for them to add extra logic to handle empty messages. By omitting the `kUpb_DecodeOption_ExperimentalAllowUnlinked` option, they will be relieved of the duty to check the tagged pointer that would indicate an empty, unlinked message. For existing users of dynamic tree shaking, there are three main changes: 1. The APIs in message/promote.h have changed, and users will need to update to the new interfaces. 2. The model for maps has changed slightly. Before, we required that map entries always had their values linked; for dynamic tree shaking to apply to maps, we required that the *entry* was left unlinked, not the entry's value. In the new model, that is reversed: map entries must always be linked, but a map entry's value can be unlinked. 3. The presence model for unlinked fields has changed. Unlinked fields will now register as "present" from the perspective of hasbits, oneof cases, and array/map entries. Users must test the tagged pointer to know if a message is of the correct, linked type or whether it is a placeholder "empty" message. There is a new function `upb_Message_GetTaggedMessagePtr()`, as well as a new accessor `upb_MessageValue.tagged_msg_val` that can be used to read and test the tagged pointer directly. PiperOrigin-RevId: 535288031
2 years ago
_upb_Decoder_CheckUnlinked(d, mt, field, op);
} else if (field->UPB_PRIVATE(descriptortype) ==
kUpb_FakeFieldType_MessageSetItem) {
*op = kUpb_DecodeOp_MessageSetItem;
} else {
*op = kUpb_DecodeOp_UnknownField;
}
return ptr;
default:
break;
}
_upb_Decoder_ErrorJmp(d, kUpb_DecodeStatus_Malformed);
}
UPB_FORCEINLINE
static const char* _upb_Decoder_DecodeKnownField(
upb_Decoder* d, const char* ptr, upb_Message* msg,
const upb_MiniTable* layout, const upb_MiniTableField* field, int op,
wireval* val) {
const upb_MiniTableSub* subs = layout->UPB_PRIVATE(subs);
uint8_t mode = field->UPB_PRIVATE(mode);
if (UPB_UNLIKELY(mode & kUpb_LabelFlags_IsExtension)) {
const upb_MiniTableExtension* ext_layout =
(const upb_MiniTableExtension*)field;
upb_Extension* ext =
_upb_Message_GetOrCreateExtension(msg, ext_layout, &d->arena);
if (UPB_UNLIKELY(!ext)) {
_upb_Decoder_ErrorJmp(d, kUpb_DecodeStatus_OutOfMemory);
}
Fixed bug when parsing an unknown value in a proto2 enum extension. #fuzzing Proto2 enum parsing is the only case where we have to look at the wire value (not merely the tag) to decide whether the field is known or unknown. If the value is unknown, we need to put the value in the Unknown Fields, but for an extension we no longer have easy access to the message, because for extensions we replace the `msg` pointer with a pointer to the extension. The bug occurred when we were treating the fake `upb_Message*` (which was actually a pointer to an extension) as a real `upb_Message*` that can have unknown fields. This CL fixes the problem by preserving the true message pointer in `d->unknown_msg` when we are parsing an extension. This also required fixing a bug in MiniTable building when fasttables are enabled. We need to set the table_mask to `-1` to disable fasttable parsing, not `0`. For unknown reasons, this CL appears to speed up parsing somewhat significantly. Ideally we should be tracking parsing performance better over time, as it is possible this is merely regaining performance that was lost at a different time: ``` benchy --reference=srcfs third_party/upb/benchmarks:benchmark 10 / 10 [=================================================================================================================] 100.00% 2m32s (Generated by http://go/benchy. Settings: --runs 5 --reference "srcfs") name old cpu/op new cpu/op delta BM_ArenaOneAlloc 23.9ns ± 6% 23.7ns ± 4% ~ (p=0.180 n=53+51) BM_ArenaInitialBlockOneAlloc 7.62ns ± 4% 7.70ns ± 5% +0.99% (p=0.024 n=59+60) BM_LoadAdsDescriptor_Upb<NoLayout> 6.60ms ±10% 6.57ms ± 8% ~ (p=0.607 n=47+54) BM_LoadAdsDescriptor_Upb<WithLayout> 6.92ms ± 5% 6.88ms ± 8% ~ (p=0.257 n=54+54) BM_LoadAdsDescriptor_Proto2<NoLayout> 14.2ms ± 8% 14.0ms ± 7% -1.38% (p=0.025 n=58+59) BM_LoadAdsDescriptor_Proto2<WithLayout> 14.3ms ± 8% 14.2ms ± 8% -1.16% (p=0.031 n=58+57) BM_Parse_Upb_FileDesc<UseArena, Copy> 15.9µs ± 4% 14.6µs ± 4% -7.85% (p=0.000 n=57+59) BM_Parse_Upb_FileDesc<UseArena, Alias> 14.5µs ± 4% 13.3µs ± 5% -8.50% (p=0.000 n=57+60) BM_Parse_Upb_FileDesc<InitBlock, Copy> 15.7µs ± 4% 14.4µs ± 5% -7.99% (p=0.000 n=59+60) BM_Parse_Upb_FileDesc<InitBlock, Alias> 14.2µs ± 5% 13.0µs ± 4% -8.56% (p=0.000 n=57+58) BM_Parse_Proto2<FileDesc, NoArena, Copy> 26.3µs ± 4% 26.2µs ± 4% ~ (p=0.195 n=55+53) BM_Parse_Proto2<FileDesc, UseArena, Copy> 13.3µs ± 5% 13.2µs ± 4% ~ (p=0.085 n=59+59) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 12.9µs ± 4% 12.8µs ± 3% -0.66% (p=0.023 n=60+58) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 10.9µs ± 6% 10.9µs ± 4% ~ (p=0.063 n=59+58) BM_SerializeDescriptor_Proto2 7.57µs ± 6% 7.62µs ± 6% ~ (p=0.147 n=57+58) BM_SerializeDescriptor_Upb 12.8µs ± 4% 12.8µs ± 4% ~ (p=0.163 n=59+56) name old time/op new time/op delta BM_ArenaOneAlloc 23.9ns ± 5% 23.7ns ± 4% ~ (p=0.172 n=53+51) BM_ArenaInitialBlockOneAlloc 7.62ns ± 4% 7.70ns ± 5% +1.02% (p=0.017 n=59+60) BM_LoadAdsDescriptor_Upb<NoLayout> 6.60ms ±10% 6.58ms ± 8% ~ (p=0.727 n=47+55) BM_LoadAdsDescriptor_Upb<WithLayout> 6.92ms ± 5% 6.88ms ± 8% ~ (p=0.260 n=54+54) BM_LoadAdsDescriptor_Proto2<NoLayout> 14.2ms ± 7% 14.0ms ± 7% -1.40% (p=0.019 n=58+59) BM_LoadAdsDescriptor_Proto2<WithLayout> 14.3ms ± 8% 14.2ms ± 8% -1.13% (p=0.037 n=58+57) BM_Parse_Upb_FileDesc<UseArena, Copy> 15.9µs ± 4% 14.6µs ± 3% -7.88% (p=0.000 n=57+59) BM_Parse_Upb_FileDesc<UseArena, Alias> 14.5µs ± 4% 13.3µs ± 5% -8.46% (p=0.000 n=57+60) BM_Parse_Upb_FileDesc<InitBlock, Copy> 15.7µs ± 4% 14.4µs ± 5% -7.99% (p=0.000 n=59+60) BM_Parse_Upb_FileDesc<InitBlock, Alias> 14.2µs ± 5% 13.0µs ± 4% -8.56% (p=0.000 n=57+58) BM_Parse_Proto2<FileDesc, NoArena, Copy> 26.3µs ± 4% 26.2µs ± 4% ~ (p=0.224 n=55+53) BM_Parse_Proto2<FileDesc, UseArena, Copy> 13.3µs ± 5% 13.2µs ± 4% ~ (p=0.098 n=59+59) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 12.9µs ± 4% 12.8µs ± 3% -0.68% (p=0.015 n=60+58) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 10.9µs ± 6% 10.9µs ± 4% ~ (p=0.052 n=59+58) BM_SerializeDescriptor_Proto2 7.56µs ± 6% 7.62µs ± 6% ~ (p=0.111 n=58+58) BM_SerializeDescriptor_Upb 12.8µs ± 4% 12.8µs ± 4% ~ (p=0.241 n=56+56) name old allocs/op new allocs/op delta BM_ArenaOneAlloc 1.00 ± 0% 1.00 ± 0% ~ (all samples are equal) BM_ArenaInitialBlockOneAlloc 0.00 ±NaN% 0.00 ±NaN% ~ (all samples are equal) BM_LoadAdsDescriptor_Upb<NoLayout> 5.98k ± 0% 5.98k ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Upb<WithLayout> 5.98k ± 0% 5.98k ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Proto2<NoLayout> 80.9k ± 0% 80.9k ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Proto2<WithLayout> 82.1k ± 0% 82.1k ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<UseArena, Copy> 7.00 ± 0% 7.00 ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<UseArena, Alias> 7.00 ± 0% 7.00 ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<InitBlock, Copy> 0.00 ±NaN% 0.00 ±NaN% ~ (all samples are equal) BM_Parse_Upb_FileDesc<InitBlock, Alias> 0.00 ±NaN% 0.00 ±NaN% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, NoArena, Copy> 765 ± 0% 765 ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, UseArena, Copy> 9.00 ± 0% 9.00 ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 0.00 ±NaN% 0.00 ±NaN% ~ (all samples are equal) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 0.00 ±NaN% 0.00 ±NaN% ~ (all samples are equal) BM_SerializeDescriptor_Proto2 0.00 ±NaN% 0.00 ±NaN% ~ (all samples are equal) BM_SerializeDescriptor_Upb 0.00 ±NaN% 0.00 ±NaN% ~ (all samples are equal) name old peak-mem(Bytes)/op new peak-mem(Bytes)/op delta BM_ArenaOneAlloc 344 ± 0% 344 ± 0% ~ (all samples are equal) BM_ArenaInitialBlockOneAlloc 0.00 ±NaN% 0.00 ±NaN% ~ (all samples are equal) BM_LoadAdsDescriptor_Upb<NoLayout> 9.60M ± 0% 9.60M ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Upb<WithLayout> 9.68M ± 0% 9.68M ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Proto2<NoLayout> 6.41M ± 0% 6.41M ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Proto2<WithLayout> 6.44M ± 0% 6.44M ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<UseArena, Copy> 36.5k ± 0% 36.5k ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<UseArena, Alias> 36.5k ± 0% 36.5k ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<InitBlock, Copy> 0.00 ±NaN% 0.00 ±NaN% ~ (all samples are equal) BM_Parse_Upb_FileDesc<InitBlock, Alias> 0.00 ±NaN% 0.00 ±NaN% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, NoArena, Copy> 35.8k ± 0% 35.8k ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, UseArena, Copy> 40.7k ± 0% 40.7k ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 0.00 ±NaN% 0.00 ±NaN% ~ (all samples are equal) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 0.00 ±NaN% 0.00 ±NaN% ~ (all samples are equal) BM_SerializeDescriptor_Proto2 0.00 ±NaN% 0.00 ±NaN% ~ (all samples are equal) BM_SerializeDescriptor_Upb 0.00 ±NaN% 0.00 ±NaN% ~ (all samples are equal) name old speed new speed delta BM_LoadAdsDescriptor_Upb<NoLayout> 113MB/s ± 9% 113MB/s ± 8% ~ (p=0.712 n=47+55) BM_LoadAdsDescriptor_Upb<WithLayout> 107MB/s ± 8% 108MB/s ± 8% ~ (p=0.200 n=55+54) BM_LoadAdsDescriptor_Proto2<NoLayout> 52.5MB/s ± 8% 53.3MB/s ± 7% +1.51% (p=0.018 n=59+59) BM_LoadAdsDescriptor_Proto2<WithLayout> 51.9MB/s ± 7% 52.4MB/s ± 8% +1.01% (p=0.050 n=58+58) BM_Parse_Upb_FileDesc<UseArena, Copy> 473MB/s ± 4% 514MB/s ± 4% +8.52% (p=0.000 n=57+59) BM_Parse_Upb_FileDesc<UseArena, Alias> 518MB/s ± 4% 566MB/s ± 5% +9.30% (p=0.000 n=57+60) BM_Parse_Upb_FileDesc<InitBlock, Copy> 480MB/s ± 4% 521MB/s ± 5% +8.69% (p=0.000 n=59+60) BM_Parse_Upb_FileDesc<InitBlock, Alias> 528MB/s ± 4% 578MB/s ± 4% +9.36% (p=0.000 n=57+58) BM_Parse_Proto2<FileDesc, NoArena, Copy> 286MB/s ± 4% 287MB/s ± 4% ~ (p=0.195 n=55+53) BM_Parse_Proto2<FileDesc, UseArena, Copy> 566MB/s ± 5% 570MB/s ± 4% ~ (p=0.085 n=59+59) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 583MB/s ± 5% 587MB/s ± 3% +0.64% (p=0.023 n=60+58) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 688MB/s ± 6% 693MB/s ± 4% ~ (p=0.063 n=59+58) BM_SerializeDescriptor_Proto2 995MB/s ± 6% 988MB/s ± 5% ~ (p=0.147 n=57+58) BM_SerializeDescriptor_Upb 586MB/s ± 4% 589MB/s ± 4% ~ (p=0.163 n=59+56) ``` PiperOrigin-RevId: 462022073
2 years ago
d->unknown_msg = msg;
msg = (upb_Message*)&ext->data;
subs = &ext->ext->UPB_PRIVATE(sub);
}
switch (mode & kUpb_FieldMode_Mask) {
case kUpb_FieldMode_Array:
return _upb_Decoder_DecodeToArray(d, ptr, msg, subs, field, val, op);
case kUpb_FieldMode_Map:
return _upb_Decoder_DecodeToMap(d, ptr, msg, subs, field, val);
case kUpb_FieldMode_Scalar:
return _upb_Decoder_DecodeToSubMessage(d, ptr, msg, subs, field, val, op);
default:
UPB_UNREACHABLE();
}
}
static const char* _upb_Decoder_ReverseSkipVarint(const char* ptr,
uint32_t val) {
uint32_t seen = 0;
do {
ptr--;
seen <<= 7;
seen |= *ptr & 0x7f;
} while (seen != val);
return ptr;
}
static const char* _upb_Decoder_DecodeUnknownField(upb_Decoder* d,
const char* ptr,
upb_Message* msg,
int field_number,
int wire_type, wireval val) {
if (field_number == 0) _upb_Decoder_ErrorJmp(d, kUpb_DecodeStatus_Malformed);
// Since unknown fields are the uncommon case, we do a little extra work here
// to walk backwards through the buffer to find the field start. This frees
// up a register in the fast paths (when the field is known), which leads to
// significant speedups in benchmarks.
const char* start = ptr;
if (wire_type == kUpb_WireType_Delimited) ptr += val.size;
if (msg) {
switch (wire_type) {
case kUpb_WireType_Varint:
case kUpb_WireType_Delimited:
start--;
while (start[-1] & 0x80) start--;
break;
case kUpb_WireType_32Bit:
start -= 4;
break;
case kUpb_WireType_64Bit:
start -= 8;
break;
default:
break;
}
assert(start == d->debug_valstart);
uint32_t tag = ((uint32_t)field_number << 3) | wire_type;
start = _upb_Decoder_ReverseSkipVarint(start, tag);
assert(start == d->debug_tagstart);
if (wire_type == kUpb_WireType_StartGroup) {
d->unknown = start;
d->unknown_msg = msg;
ptr = _upb_Decoder_DecodeUnknownGroup(d, ptr, field_number);
start = d->unknown;
d->unknown = NULL;
}
if (!UPB_PRIVATE(_upb_Message_AddUnknown)(msg, start, ptr - start,
&d->arena)) {
_upb_Decoder_ErrorJmp(d, kUpb_DecodeStatus_OutOfMemory);
}
} else if (wire_type == kUpb_WireType_StartGroup) {
ptr = _upb_Decoder_DecodeUnknownGroup(d, ptr, field_number);
}
return ptr;
}
UPB_NOINLINE
static const char* _upb_Decoder_DecodeMessage(upb_Decoder* d, const char* ptr,
upb_Message* msg,
const upb_MiniTable* layout) {
Optimized decoder and paved the way for parsing extensions. The primary motivation for this change is to avoid referring to the `upb_msglayout` object when we are trying to fetch the `upb_msglayout` object for a sub-message. This will help pave the way for parsing extensions. We also implement several optimizations so that we can make this change without regressing performance. Normally we compute the layout for a sub-message field like so: ``` const upb_msglayout *get_submsg_layout( const upb_msglayout *layout, const upb_msglayout_field *field) { return layout->submsgs[field->submsg_index] } ``` The reason for this indirection is to avoid storing a pointer directly in `upb_msglayout_field`, as this would double its size (from 12 to 24 bytes on 64-bit architectures) which is wasteful as this pointer is only needed for message typed fields. However `get_submsg_layout` as written above does not work for extensions, as they will not have entries in the message's `layout->submsgs` array by nature, and we want to avoid creating an entire fake `upb_msglayout` for each such extension since that would also be wasteful. This change removes the dependency on `upb_msglayout` by passing down the `submsgs` array instead: ``` const upb_msglayout *get_submsg_layout( const upb_msglayout *const *submsgs, const upb_msglayout_field *field) { return submsgs[field->submsg_index] } ``` This will pave the way for parsing extensions, as we can more easily create an alternative `submsgs` array for extension fields without extra overhead or waste. Along the way several optimizations presented themselves that allow a nice increase in performance: 1. Passing the parsed `wireval` by address instead of by value ended up avoiding an expensive and useless stack copy (this is on Clang, which was used for all measurements). 2. When field numbers are densely packed, we can find a field by number with a single indexed lookup instead of linear search. At codegen time we can compute the maximum field number that will allow such an indexed lookup. 3. For fields that do require linear search, we can start the linear search at the location where we found the previous field, taking advantage of the fact that field numbers are generally increasing. 4. When the hasbit index is less than 32 (the common case) we can use a less expensive code sequence to set it. 5. We check for the hasbit case before the oneof case, as optional fields are more common than oneof fields. Benchmark results indicate a 20% improvement in parse speed with a small code size increase: ``` name old time/op new time/op delta ArenaOneAlloc 21.3ns ± 0% 21.5ns ± 0% +0.96% (p=0.000 n=12+12) ArenaInitialBlockOneAlloc 6.32ns ± 0% 6.32ns ± 0% +0.03% (p=0.000 n=12+10) LoadDescriptor_Upb 53.5µs ± 1% 51.5µs ± 2% -3.70% (p=0.000 n=12+12) LoadAdsDescriptor_Upb 2.78ms ± 2% 2.68ms ± 0% -3.57% (p=0.000 n=12+12) LoadDescriptor_Proto2 240µs ± 0% 240µs ± 0% +0.12% (p=0.001 n=12+12) LoadAdsDescriptor_Proto2 12.8ms ± 0% 12.7ms ± 0% -1.15% (p=0.000 n=12+10) Parse_Upb_FileDesc<UseArena,Copy> 13.2µs ± 2% 10.7µs ± 0% -18.49% (p=0.000 n=10+12) Parse_Upb_FileDesc<UseArena,Alias> 11.3µs ± 0% 9.6µs ± 0% -15.11% (p=0.000 n=12+11) Parse_Upb_FileDesc<InitBlock,Copy> 12.7µs ± 0% 10.3µs ± 0% -19.00% (p=0.000 n=10+12) Parse_Upb_FileDesc<InitBlock,Alias> 10.9µs ± 0% 9.2µs ± 0% -15.82% (p=0.000 n=12+12) Parse_Proto2<FileDesc,NoArena,Copy> 29.4µs ± 0% 29.5µs ± 0% +0.61% (p=0.000 n=12+12) Parse_Proto2<FileDesc,UseArena,Copy> 20.7µs ± 2% 20.6µs ± 2% ~ (p=0.260 n=12+11) Parse_Proto2<FileDesc,InitBlock,Copy> 16.7µs ± 1% 16.7µs ± 0% -0.25% (p=0.036 n=12+10) Parse_Proto2<FileDescSV,InitBlock,Alias> 16.5µs ± 0% 16.5µs ± 0% +0.20% (p=0.016 n=12+11) SerializeDescriptor_Proto2 5.30µs ± 1% 5.36µs ± 1% +1.09% (p=0.000 n=12+11) SerializeDescriptor_Upb 12.9µs ± 0% 13.0µs ± 0% +0.90% (p=0.000 n=12+11) FILE SIZE VM SIZE -------------- -------------- +1.5% +176 +1.6% +176 upb/decode.c +1.8% +176 +1.9% +176 decode_msg +0.4% +64 +0.4% +64 upb/def.c +1.4% +64 +1.4% +64 _upb_symtab_addfile +1.2% +48 +1.4% +48 upb/reflection.c +15% +32 +18% +32 upb_msg_set +2.9% +16 +3.1% +16 upb_msg_mutable -9.3% -288 [ = ] 0 [Unmapped] [ = ] 0 +0.2% +288 TOTAL ```
4 years ago
int last_field_index = 0;
#if UPB_FASTTABLE
// The first time we want to skip fast dispatch, because we may have just been
// invoked by the fast parser to handle a case that it bailed on.
if (!_upb_Decoder_IsDone(d, &ptr)) goto nofast;
#endif
while (!_upb_Decoder_IsDone(d, &ptr)) {
uint32_t tag;
const upb_MiniTableField* field;
int field_number;
int wire_type;
wireval val;
int op;
if (_upb_Decoder_TryFastDispatch(d, &ptr, msg, layout)) break;
#if UPB_FASTTABLE
nofast:
#endif
#ifndef NDEBUG
d->debug_tagstart = ptr;
#endif
UPB_ASSERT(ptr < d->input.limit_ptr);
ptr = _upb_Decoder_DecodeTag(d, ptr, &tag);
field_number = tag >> 3;
wire_type = tag & 7;
#ifndef NDEBUG
d->debug_valstart = ptr;
#endif
if (wire_type == kUpb_WireType_EndGroup) {
d->end_group = field_number;
return ptr;
}
field = _upb_Decoder_FindField(d, layout, field_number, &last_field_index);
Treat unlinked sub-messages in the MiniTable as unknown This is an observable behavior change in the decoder. After submitting this CL, clients of the decoder can assume that any unlinked sub-messages will be treated as unknown, rather than crashing. Unlinked sub-messages must never have values present in the message. We can verify this with asserts. Since the values are never set, the encoder should never encounter data for any unlinked sub-message. ``` name old cpu/op new cpu/op delta BM_ArenaOneAlloc 18.3ns ± 9% 17.9ns ± 2% ~ (p=0.690 n=5+5) BM_ArenaInitialBlockOneAlloc 6.40ns ± 1% 6.68ns ±10% ~ (p=0.730 n=4+5) BM_LoadAdsDescriptor_Upb<NoLayout> 5.09ms ± 2% 5.03ms ± 3% ~ (p=0.222 n=5+5) BM_LoadAdsDescriptor_Upb<WithLayout> 5.45ms ± 3% 5.43ms ± 1% ~ (p=0.905 n=5+4) BM_LoadAdsDescriptor_Proto2<NoLayout> 10.9ms ± 1% 10.8ms ± 1% -1.09% (p=0.016 n=5+4) BM_LoadAdsDescriptor_Proto2<WithLayout> 11.3ms ± 9% 11.1ms ± 3% ~ (p=0.841 n=5+5) BM_Parse_Upb_FileDesc<UseArena, Copy> 11.2µs ± 3% 11.3µs ± 3% ~ (p=0.222 n=5+5) BM_Parse_Upb_FileDesc<UseArena, Alias> 10.3µs ± 5% 10.5µs ± 5% ~ (p=0.310 n=5+5) BM_Parse_Upb_FileDesc<InitBlock, Copy> 11.4µs ±18% 11.0µs ± 2% ~ (p=1.000 n=5+5) BM_Parse_Upb_FileDesc<InitBlock, Alias> 10.5µs ±17% 10.6µs ±19% ~ (p=0.421 n=5+5) BM_Parse_Proto2<FileDesc, NoArena, Copy> 20.5µs ± 2% 20.2µs ± 2% ~ (p=0.222 n=5+5) BM_Parse_Proto2<FileDesc, UseArena, Copy> 10.8µs ± 2% 10.9µs ± 4% ~ (p=0.841 n=5+5) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 10.5µs ± 3% 10.6µs ± 3% ~ (p=0.690 n=5+5) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 9.22µs ± 2% 9.23µs ± 3% ~ (p=1.000 n=5+5) BM_SerializeDescriptor_Proto2 6.05µs ± 3% 5.90µs ± 3% ~ (p=0.222 n=5+5) BM_SerializeDescriptor_Upb 10.2µs ± 3% 10.6µs ±14% ~ (p=0.841 n=5+5) name old time/op new time/op delta BM_ArenaOneAlloc 18.3ns ± 9% 17.9ns ± 2% ~ (p=0.841 n=5+5) BM_ArenaInitialBlockOneAlloc 6.42ns ± 1% 6.69ns ±10% ~ (p=0.730 n=4+5) BM_LoadAdsDescriptor_Upb<NoLayout> 5.10ms ± 2% 5.05ms ± 3% ~ (p=0.222 n=5+5) BM_LoadAdsDescriptor_Upb<WithLayout> 5.47ms ± 3% 5.45ms ± 1% ~ (p=0.905 n=5+4) BM_LoadAdsDescriptor_Proto2<NoLayout> 10.9ms ± 1% 10.8ms ± 1% -1.11% (p=0.016 n=5+4) BM_LoadAdsDescriptor_Proto2<WithLayout> 11.4ms ± 9% 11.1ms ± 3% ~ (p=0.841 n=5+5) BM_Parse_Upb_FileDesc<UseArena, Copy> 11.2µs ± 3% 11.3µs ± 3% ~ (p=0.222 n=5+5) BM_Parse_Upb_FileDesc<UseArena, Alias> 10.3µs ± 5% 10.5µs ± 5% ~ (p=0.151 n=5+5) BM_Parse_Upb_FileDesc<InitBlock, Copy> 11.5µs ±18% 11.0µs ± 2% ~ (p=1.000 n=5+5) BM_Parse_Upb_FileDesc<InitBlock, Alias> 10.5µs ±17% 10.7µs ±19% ~ (p=0.421 n=5+5) BM_Parse_Proto2<FileDesc, NoArena, Copy> 20.6µs ± 2% 20.3µs ± 2% ~ (p=0.222 n=5+5) BM_Parse_Proto2<FileDesc, UseArena, Copy> 10.9µs ± 2% 10.9µs ± 4% ~ (p=0.841 n=5+5) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 10.6µs ± 3% 10.6µs ± 3% ~ (p=0.690 n=5+5) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 9.24µs ± 2% 9.25µs ± 3% ~ (p=1.000 n=5+5) BM_SerializeDescriptor_Proto2 6.07µs ± 3% 5.91µs ± 3% ~ (p=0.222 n=5+5) BM_SerializeDescriptor_Upb 10.3µs ± 3% 10.6µs ±14% ~ (p=0.841 n=5+5) name old INSTRUCTIONS/op new INSTRUCTIONS/op delta BM_ArenaOneAlloc 201 ± 0% 201 ± 0% ~ (p=0.841 n=5+5) BM_ArenaInitialBlockOneAlloc 69.0 ± 0% 69.0 ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Upb<NoLayout> 33.9M ± 0% 34.1M ± 0% +0.66% (p=0.008 n=5+5) BM_LoadAdsDescriptor_Upb<WithLayout> 35.6M ± 0% 35.8M ± 0% +0.64% (p=0.008 n=5+5) BM_LoadAdsDescriptor_Proto2<NoLayout> 70.8M ± 0% 70.8M ± 0% ~ (p=0.548 n=5+5) BM_LoadAdsDescriptor_Proto2<WithLayout> 71.6M ± 0% 71.6M ± 0% ~ (p=0.151 n=5+5) BM_Parse_Upb_FileDesc<UseArena, Copy> 137k ± 0% 141k ± 0% +2.87% (p=0.008 n=5+5) BM_Parse_Upb_FileDesc<UseArena, Alias> 125k ± 0% 128k ± 0% +2.83% (p=0.008 n=5+5) BM_Parse_Upb_FileDesc<InitBlock, Copy> 135k ± 0% 139k ± 0% +2.89% (p=0.008 n=5+5) BM_Parse_Upb_FileDesc<InitBlock, Alias> 124k ± 0% 127k ± 0% +2.85% (p=0.016 n=5+4) BM_Parse_Proto2<FileDesc, NoArena, Copy> 201k ± 0% 201k ± 0% ~ (p=0.222 n=5+5) BM_Parse_Proto2<FileDesc, UseArena, Copy> 107k ± 0% 107k ± 0% ~ (p=1.000 n=5+5) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 105k ± 0% 105k ± 0% ~ (p=0.286 n=5+4) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 86.5k ± 0% 86.5k ± 0% ~ (p=0.222 n=5+5) BM_SerializeDescriptor_Proto2 60.3k ± 0% 60.3k ± 0% ~ (p=0.071 n=5+5) BM_SerializeDescriptor_Upb 111k ± 0% 111k ± 0% ~ (p=0.841 n=5+5) name old CYCLES/op new CYCLES/op delta BM_ArenaOneAlloc 60.0 ± 7% 58.8 ± 0% -2.15% (p=0.016 n=5+5) BM_ArenaInitialBlockOneAlloc 21.0 ± 0% 21.0 ± 0% ~ (p=1.000 n=5+5) BM_LoadAdsDescriptor_Upb<NoLayout> 16.9M ± 0% 16.9M ± 0% ~ (p=0.056 n=5+5) BM_LoadAdsDescriptor_Upb<WithLayout> 17.9M ± 1% 18.0M ± 1% ~ (p=0.095 n=5+5) BM_LoadAdsDescriptor_Proto2<NoLayout> 35.9M ± 1% 35.8M ± 1% ~ (p=0.421 n=5+5) BM_LoadAdsDescriptor_Proto2<WithLayout> 36.5M ± 0% 36.5M ± 0% ~ (p=0.841 n=5+5) BM_Parse_Upb_FileDesc<UseArena, Copy> 37.2k ± 0% 37.3k ± 0% ~ (p=0.222 n=5+5) BM_Parse_Upb_FileDesc<UseArena, Alias> 34.1k ± 0% 34.7k ± 0% +1.66% (p=0.008 n=5+5) BM_Parse_Upb_FileDesc<InitBlock, Copy> 36.4k ± 0% 36.7k ± 0% +0.83% (p=0.008 n=5+5) BM_Parse_Upb_FileDesc<InitBlock, Alias> 33.3k ± 1% 34.1k ± 1% +2.39% (p=0.008 n=5+5) BM_Parse_Proto2<FileDesc, NoArena, Copy> 68.1k ± 1% 68.0k ± 1% ~ (p=0.421 n=5+5) BM_Parse_Proto2<FileDesc, UseArena, Copy> 36.0k ± 1% 36.1k ± 1% ~ (p=0.841 n=5+5) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 35.3k ± 1% 35.5k ± 1% ~ (p=0.151 n=5+5) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 30.7k ± 0% 30.9k ± 1% ~ (p=0.151 n=5+5) BM_SerializeDescriptor_Proto2 20.3k ± 2% 19.7k ± 3% ~ (p=0.151 n=5+5) BM_SerializeDescriptor_Upb 33.6k ± 0% 33.7k ± 2% ~ (p=1.000 n=5+5) name old allocs/op new allocs/op delta BM_ArenaOneAlloc 1.19 ± 0% 1.19 ± 0% ~ (all samples are equal) BM_ArenaInitialBlockOneAlloc 0.19 ± 0% 0.19 ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Upb<NoLayout> 6.00k ± 0% 6.00k ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Upb<WithLayout> 5.99k ± 0% 5.99k ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Proto2<NoLayout> 77.8k ± 0% 77.8k ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Proto2<WithLayout> 79.0k ± 0% 79.0k ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<UseArena, Copy> 7.19 ± 0% 7.19 ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<UseArena, Alias> 7.19 ± 0% 7.19 ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<InitBlock, Copy> 0.19 ± 0% 0.19 ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<InitBlock, Alias> 0.19 ± 0% 0.19 ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, NoArena, Copy> 765 ± 0% 765 ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, UseArena, Copy> 10.2 ± 0% 10.2 ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 1.19 ± 0% 1.19 ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 1.19 ± 0% 1.19 ± 0% ~ (all samples are equal) BM_SerializeDescriptor_Proto2 0.19 ± 0% 0.19 ± 0% ~ (all samples are equal) BM_SerializeDescriptor_Upb 0.19 ± 0% 0.19 ± 0% ~ (all samples are equal) name old peak-mem(Bytes)/op new peak-mem(Bytes)/op delta BM_ArenaOneAlloc 344 ± 0% 344 ± 0% ~ (all samples are equal) BM_ArenaInitialBlockOneAlloc 112 ± 0% 112 ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Upb<NoLayout> 9.64M ± 0% 9.64M ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Upb<WithLayout> 9.70M ± 0% 9.70M ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Proto2<NoLayout> 6.38M ± 0% 6.38M ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Proto2<WithLayout> 6.44M ± 0% 6.44M ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<UseArena, Copy> 36.5k ± 0% 36.5k ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<UseArena, Alias> 36.5k ± 0% 36.5k ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<InitBlock, Copy> 112 ± 0% 112 ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<InitBlock, Alias> 112 ± 0% 112 ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, NoArena, Copy> 35.8k ± 0% 35.8k ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, UseArena, Copy> 40.8k ± 0% 40.8k ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 112 ± 0% 112 ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 112 ± 0% 112 ± 0% ~ (all samples are equal) BM_SerializeDescriptor_Proto2 112 ± 0% 112 ± 0% ~ (all samples are equal) BM_SerializeDescriptor_Upb 112 ± 0% 112 ± 0% ~ (all samples are equal) name old speed new speed delta BM_LoadAdsDescriptor_Upb<NoLayout> 147MB/s ± 2% 148MB/s ± 3% ~ (p=0.222 n=5+5) BM_LoadAdsDescriptor_Upb<WithLayout> 137MB/s ± 3% 137MB/s ± 1% ~ (p=0.905 n=5+4) BM_LoadAdsDescriptor_Proto2<NoLayout> 68.6MB/s ± 1% 69.3MB/s ± 1% +1.10% (p=0.016 n=5+4) BM_LoadAdsDescriptor_Proto2<WithLayout> 66.0MB/s ± 9% 67.4MB/s ± 3% ~ (p=0.841 n=5+5) BM_Parse_Upb_FileDesc<UseArena, Copy> 675MB/s ± 3% 667MB/s ± 3% ~ (p=0.222 n=5+5) BM_Parse_Upb_FileDesc<UseArena, Alias> 730MB/s ± 5% 718MB/s ± 5% ~ (p=0.310 n=5+5) BM_Parse_Upb_FileDesc<InitBlock, Copy> 663MB/s ±16% 685MB/s ± 2% ~ (p=1.000 n=5+5) BM_Parse_Upb_FileDesc<InitBlock, Alias> 723MB/s ±15% 712MB/s ±16% ~ (p=0.421 n=5+5) BM_Parse_Proto2<FileDesc, NoArena, Copy> 367MB/s ± 2% 372MB/s ± 2% ~ (p=0.222 n=5+5) BM_Parse_Proto2<FileDesc, UseArena, Copy> 694MB/s ± 2% 691MB/s ± 4% ~ (p=0.841 n=5+5) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 714MB/s ± 3% 709MB/s ± 3% ~ (p=0.690 n=5+5) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 816MB/s ± 2% 816MB/s ± 3% ~ (p=1.000 n=5+5) BM_SerializeDescriptor_Proto2 1.24GB/s ± 3% 1.28GB/s ± 3% ~ (p=0.222 n=5+5) BM_SerializeDescriptor_Upb 734MB/s ± 3% 713MB/s ±13% ~ (p=0.841 n=5+5) ``` PiperOrigin-RevId: 477770562
2 years ago
ptr = _upb_Decoder_DecodeWireValue(d, ptr, layout, field, wire_type, &val,
&op);
if (op >= 0) {
ptr = _upb_Decoder_DecodeKnownField(d, ptr, msg, layout, field, op, &val);
} else {
switch (op) {
case kUpb_DecodeOp_UnknownField:
ptr = _upb_Decoder_DecodeUnknownField(d, ptr, msg, field_number,
wire_type, val);
break;
case kUpb_DecodeOp_MessageSetItem:
ptr = upb_Decoder_DecodeMessageSetItem(d, ptr, msg, layout);
break;
}
}
}
return UPB_UNLIKELY(layout && layout->UPB_PRIVATE(required_count))
? _upb_Decoder_CheckRequired(d, ptr, msg, layout)
: ptr;
4 years ago
}
const char* _upb_FastDecoder_DecodeGeneric(struct upb_Decoder* d,
const char* ptr, upb_Message* msg,
intptr_t table, uint64_t hasbits,
uint64_t data) {
(void)data;
*(uint32_t*)msg |= hasbits;
return _upb_Decoder_DecodeMessage(d, ptr, msg, decode_totablep(table));
}
static upb_DecodeStatus _upb_Decoder_DecodeTop(struct upb_Decoder* d,
const char* buf, void* msg,
const upb_MiniTable* l) {
if (!_upb_Decoder_TryFastDispatch(d, &buf, msg, l)) {
_upb_Decoder_DecodeMessage(d, buf, msg, l);
}
if (d->end_group != DECODE_NOGROUP) return kUpb_DecodeStatus_Malformed;
if (d->missing_required) return kUpb_DecodeStatus_MissingRequired;
return kUpb_DecodeStatus_Ok;
}
UPB_NOINLINE
const char* _upb_Decoder_IsDoneFallback(upb_EpsCopyInputStream* e,
const char* ptr, int overrun) {
return _upb_EpsCopyInputStream_IsDoneFallbackInline(
e, ptr, overrun, _upb_Decoder_BufferFlipCallback);
}
static upb_DecodeStatus upb_Decoder_Decode(upb_Decoder* const decoder,
const char* const buf,
void* const msg,
const upb_MiniTable* const l,
upb_Arena* const arena) {
if (UPB_SETJMP(decoder->err) == 0) {
decoder->status = _upb_Decoder_DecodeTop(decoder, buf, msg, l);
} else {
UPB_ASSERT(decoder->status != kUpb_DecodeStatus_Ok);
}
UPB_PRIVATE(_upb_Arena_SwapOut)(arena, &decoder->arena);
return decoder->status;
}
upb_DecodeStatus upb_Decode(const char* buf, size_t size, upb_Message* msg,
const upb_MiniTable* l,
const upb_ExtensionRegistry* extreg, int options,
upb_Arena* arena) {
Changed Arena representation so that fusing links arenas together instead of blocks. Previously when fusing, we would concatenate all blocks into a single list that lived in the arena root. From then on, all arenas would add their blocks to this single unified list. After this CL, arenas keep their distinct list of blocks even after being fused. Instead of unifying the block list, fuse now puts the arenas themselves into a list, so all arenas in the fused group can be iterated over at any time. This design makes it easier to keep each individual arena thread-compatible, because fuse and free are now the only mutating operations that touch state that is shared with the entire group. Read-only operations like `SpaceAllocated()` also iterate the list of arenas, but in a read-only fashion. (Note: we need tests for SpaceAllocated(), both single-threaded for correctness and multi-threaded for resilience to crashes and data races). Performance of fuse regresses by 5-20%. This is somewhat expected as we are performing more atomic operations during a fuse. ``` name old cpu/op new cpu/op delta BM_ArenaOneAlloc 18.4ns ± 6% 18.7ns ± 4% +2.00% (p=0.016 n=18+18) BM_ArenaInitialBlockOneAlloc 5.50ns ± 4% 6.57ns ± 4% +19.42% (p=0.000 n=16+17) BM_ArenaFuseUnbalanced/2 59.3ns ±10% 68.7ns ± 4% +15.85% (p=0.000 n=19+19) BM_ArenaFuseUnbalanced/8 479ns ± 5% 540ns ± 8% +12.57% (p=0.000 n=18+19) BM_ArenaFuseUnbalanced/64 4.50µs ± 4% 4.93µs ± 8% +9.59% (p=0.000 n=17+17) BM_ArenaFuseUnbalanced/128 9.24µs ± 3% 9.96µs ± 3% +7.81% (p=0.000 n=17+17) BM_ArenaFuseBalanced/2 63.3ns ±18% 71.0ns ± 4% +12.14% (p=0.000 n=19+18) BM_ArenaFuseBalanced/8 484ns ± 9% 543ns ±10% +12.11% (p=0.000 n=17+16) BM_ArenaFuseBalanced/64 4.50µs ± 6% 4.94µs ± 4% +9.62% (p=0.000 n=19+17) BM_ArenaFuseBalanced/128 9.20µs ± 4% 9.95µs ± 4% +8.12% (p=0.000 n=16+19) BM_LoadAdsDescriptor_Upb<NoLayout> 5.50ms ± 8% 5.69ms ±17% ~ (p=0.189 n=18+19) BM_LoadAdsDescriptor_Upb<WithLayout> 6.10ms ± 5% 6.05ms ± 4% ~ (p=0.258 n=17+18) BM_LoadAdsDescriptor_Proto2<NoLayout> 11.9ms ±15% 11.6ms ± 5% ~ (p=0.589 n=19+16) BM_LoadAdsDescriptor_Proto2<WithLayout> 11.8ms ± 5% 12.4ms ±17% ~ (p=0.604 n=16+20) BM_Parse_Upb_FileDesc<UseArena, Copy> 12.1µs ± 8% 12.1µs ± 4% ~ (p=1.000 n=18+18) BM_Parse_Upb_FileDesc<UseArena, Alias> 11.8µs ±17% 11.1µs ± 4% ~ (p=0.104 n=20+17) BM_Parse_Upb_FileDesc<InitBlock, Copy> 12.0µs ± 5% 11.9µs ± 4% ~ (p=0.134 n=18+19) BM_Parse_Upb_FileDesc<InitBlock, Alias> 10.9µs ± 7% 11.0µs ± 4% ~ (p=0.195 n=17+18) BM_Parse_Proto2<FileDesc, NoArena, Copy> 24.2µs ± 4% 24.4µs ± 7% ~ (p=0.767 n=18+18) BM_Parse_Proto2<FileDesc, UseArena, Copy> 11.6µs ± 5% 11.6µs ± 4% ~ (p=0.621 n=18+16) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 11.3µs ± 3% 11.3µs ± 3% ~ (p=0.743 n=18+18) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 12.7µs ± 8% 12.7µs ± 4% ~ (p=0.988 n=18+19) BM_SerializeDescriptor_Proto2 5.77µs ± 5% 5.71µs ± 5% ~ (p=0.433 n=17+17) BM_SerializeDescriptor_Upb 10.0µs ± 5% 10.1µs ± 7% ~ (p=0.102 n=19+16) name old time/op new time/op delta BM_ArenaOneAlloc 18.4ns ± 6% 18.8ns ± 4% +1.98% (p=0.019 n=18+18) BM_ArenaInitialBlockOneAlloc 5.51ns ± 4% 6.58ns ± 4% +19.42% (p=0.000 n=16+17) BM_ArenaFuseUnbalanced/2 59.5ns ±10% 68.9ns ± 4% +15.83% (p=0.000 n=19+19) BM_ArenaFuseUnbalanced/8 481ns ± 5% 541ns ± 8% +12.54% (p=0.000 n=18+19) BM_ArenaFuseUnbalanced/64 4.51µs ± 4% 4.94µs ± 8% +9.53% (p=0.000 n=17+17) BM_ArenaFuseUnbalanced/128 9.26µs ± 3% 9.98µs ± 3% +7.79% (p=0.000 n=17+17) BM_ArenaFuseBalanced/2 63.5ns ±19% 71.1ns ± 3% +12.07% (p=0.000 n=19+18) BM_ArenaFuseBalanced/8 485ns ± 9% 551ns ±20% +13.47% (p=0.000 n=17+17) BM_ArenaFuseBalanced/64 4.51µs ± 6% 4.95µs ± 4% +9.62% (p=0.000 n=19+17) BM_ArenaFuseBalanced/128 9.22µs ± 4% 9.97µs ± 4% +8.12% (p=0.000 n=16+19) BM_LoadAdsDescriptor_Upb<NoLayout> 5.52ms ± 8% 5.72ms ±18% ~ (p=0.199 n=18+19) BM_LoadAdsDescriptor_Upb<WithLayout> 6.12ms ± 5% 6.07ms ± 4% ~ (p=0.273 n=17+18) BM_LoadAdsDescriptor_Proto2<NoLayout> 11.9ms ±15% 11.6ms ± 5% ~ (p=0.589 n=19+16) BM_LoadAdsDescriptor_Proto2<WithLayout> 11.9ms ± 5% 12.5ms ±18% ~ (p=0.582 n=16+20) BM_Parse_Upb_FileDesc<UseArena, Copy> 12.2µs ± 8% 12.1µs ± 3% ~ (p=0.963 n=18+18) BM_Parse_Upb_FileDesc<UseArena, Alias> 11.8µs ±17% 11.1µs ± 4% ~ (p=0.104 n=20+17) BM_Parse_Upb_FileDesc<InitBlock, Copy> 12.0µs ± 5% 11.9µs ± 4% ~ (p=0.126 n=18+19) BM_Parse_Upb_FileDesc<InitBlock, Alias> 11.0µs ± 6% 11.1µs ± 4% ~ (p=0.195 n=17+18) BM_Parse_Proto2<FileDesc, NoArena, Copy> 24.3µs ± 4% 24.5µs ± 6% ~ (p=0.743 n=18+18) BM_Parse_Proto2<FileDesc, UseArena, Copy> 11.7µs ± 5% 11.6µs ± 4% ~ (p=0.574 n=18+16) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 11.3µs ± 3% 11.3µs ± 3% ~ (p=0.743 n=18+18) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 12.7µs ± 8% 12.7µs ± 4% ~ (p=0.988 n=18+19) BM_SerializeDescriptor_Proto2 5.78µs ± 5% 5.73µs ± 5% ~ (p=0.357 n=17+17) BM_SerializeDescriptor_Upb 10.0µs ± 5% 10.1µs ± 7% ~ (p=0.117 n=19+16) name old allocs/op new allocs/op delta BM_ArenaOneAlloc 1.00 ± 0% 1.00 ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/2 2.00 ± 0% 2.00 ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/8 8.00 ± 0% 8.00 ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/64 64.0 ± 0% 64.0 ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/128 128 ± 0% 128 ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/2 2.00 ± 0% 2.00 ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/8 8.00 ± 0% 8.00 ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/64 64.0 ± 0% 64.0 ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/128 128 ± 0% 128 ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Upb<NoLayout> 6.08k ± 0% 6.05k ± 0% -0.54% (p=0.000 n=20+20) BM_LoadAdsDescriptor_Upb<WithLayout> 6.39k ± 0% 6.36k ± 0% -0.55% (p=0.000 n=20+20) BM_LoadAdsDescriptor_Proto2<NoLayout> 83.4k ± 0% 83.4k ± 0% ~ (p=0.800 n=20+20) BM_LoadAdsDescriptor_Proto2<WithLayout> 84.4k ± 0% 84.4k ± 0% ~ (p=0.752 n=20+20) BM_Parse_Upb_FileDesc<UseArena, Copy> 7.00 ± 0% 7.00 ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<UseArena, Alias> 7.00 ± 0% 7.00 ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, NoArena, Copy> 765 ± 0% 765 ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, UseArena, Copy> 8.00 ± 0% 8.00 ± 0% ~ (all samples are equal) name old peak-mem(Bytes)/op new peak-mem(Bytes)/op delta BM_ArenaOneAlloc 336 ± 0% 336 ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/2 672 ± 0% 672 ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/8 2.69k ± 0% 2.69k ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/64 21.5k ± 0% 21.5k ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/128 43.0k ± 0% 43.0k ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/2 672 ± 0% 672 ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/8 2.69k ± 0% 2.69k ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/64 21.5k ± 0% 21.5k ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/128 43.0k ± 0% 43.0k ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Upb<NoLayout> 9.89M ± 0% 9.95M ± 0% +0.65% (p=0.000 n=20+20) BM_LoadAdsDescriptor_Upb<WithLayout> 9.95M ± 0% 10.02M ± 0% +0.70% (p=0.000 n=20+20) BM_LoadAdsDescriptor_Proto2<NoLayout> 6.62M ± 0% 6.62M ± 0% ~ (p=0.800 n=20+20) BM_LoadAdsDescriptor_Proto2<WithLayout> 6.66M ± 0% 6.66M ± 0% ~ (p=0.752 n=20+20) BM_Parse_Upb_FileDesc<UseArena, Copy> 36.5k ± 0% 36.5k ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<UseArena, Alias> 36.5k ± 0% 36.5k ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, NoArena, Copy> 35.8k ± 0% 35.8k ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, UseArena, Copy> 65.3k ± 0% 65.3k ± 0% ~ (all samples are equal) name old speed new speed delta BM_LoadAdsDescriptor_Upb<NoLayout> 138MB/s ± 7% 132MB/s ±15% ~ (p=0.126 n=18+20) BM_LoadAdsDescriptor_Upb<WithLayout> 124MB/s ± 5% 125MB/s ± 4% ~ (p=0.258 n=17+18) BM_LoadAdsDescriptor_Proto2<NoLayout> 63.9MB/s ±13% 65.2MB/s ± 5% ~ (p=0.589 n=19+16) BM_LoadAdsDescriptor_Proto2<WithLayout> 64.0MB/s ± 5% 61.3MB/s ±15% ~ (p=0.604 n=16+20) BM_Parse_Upb_FileDesc<UseArena, Copy> 620MB/s ± 8% 622MB/s ± 4% ~ (p=1.000 n=18+18) BM_Parse_Upb_FileDesc<UseArena, Alias> 644MB/s ±15% 679MB/s ± 4% ~ (p=0.104 n=20+17) BM_Parse_Upb_FileDesc<InitBlock, Copy> 627MB/s ± 4% 633MB/s ± 4% ~ (p=0.134 n=18+19) BM_Parse_Upb_FileDesc<InitBlock, Alias> 688MB/s ± 6% 682MB/s ± 4% ~ (p=0.195 n=17+18) BM_Parse_Proto2<FileDesc, NoArena, Copy> 310MB/s ± 4% 309MB/s ± 6% ~ (p=0.767 n=18+18) BM_Parse_Proto2<FileDesc, UseArena, Copy> 646MB/s ± 4% 649MB/s ± 4% ~ (p=0.621 n=18+16) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 666MB/s ± 3% 666MB/s ± 3% ~ (p=0.743 n=18+18) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 592MB/s ± 7% 593MB/s ± 4% ~ (p=0.988 n=18+19) BM_SerializeDescriptor_Proto2 1.30GB/s ± 5% 1.32GB/s ± 5% ~ (p=0.433 n=17+17) BM_SerializeDescriptor_Upb 756MB/s ± 5% 745MB/s ± 6% ~ (p=0.102 n=19+16) ``` PiperOrigin-RevId: 520144430
2 years ago
upb_Decoder decoder;
unsigned depth = (unsigned)options >> 16;
Changed Arena representation so that fusing links arenas together instead of blocks. Previously when fusing, we would concatenate all blocks into a single list that lived in the arena root. From then on, all arenas would add their blocks to this single unified list. After this CL, arenas keep their distinct list of blocks even after being fused. Instead of unifying the block list, fuse now puts the arenas themselves into a list, so all arenas in the fused group can be iterated over at any time. This design makes it easier to keep each individual arena thread-compatible, because fuse and free are now the only mutating operations that touch state that is shared with the entire group. Read-only operations like `SpaceAllocated()` also iterate the list of arenas, but in a read-only fashion. (Note: we need tests for SpaceAllocated(), both single-threaded for correctness and multi-threaded for resilience to crashes and data races). Performance of fuse regresses by 5-20%. This is somewhat expected as we are performing more atomic operations during a fuse. ``` name old cpu/op new cpu/op delta BM_ArenaOneAlloc 18.4ns ± 6% 18.7ns ± 4% +2.00% (p=0.016 n=18+18) BM_ArenaInitialBlockOneAlloc 5.50ns ± 4% 6.57ns ± 4% +19.42% (p=0.000 n=16+17) BM_ArenaFuseUnbalanced/2 59.3ns ±10% 68.7ns ± 4% +15.85% (p=0.000 n=19+19) BM_ArenaFuseUnbalanced/8 479ns ± 5% 540ns ± 8% +12.57% (p=0.000 n=18+19) BM_ArenaFuseUnbalanced/64 4.50µs ± 4% 4.93µs ± 8% +9.59% (p=0.000 n=17+17) BM_ArenaFuseUnbalanced/128 9.24µs ± 3% 9.96µs ± 3% +7.81% (p=0.000 n=17+17) BM_ArenaFuseBalanced/2 63.3ns ±18% 71.0ns ± 4% +12.14% (p=0.000 n=19+18) BM_ArenaFuseBalanced/8 484ns ± 9% 543ns ±10% +12.11% (p=0.000 n=17+16) BM_ArenaFuseBalanced/64 4.50µs ± 6% 4.94µs ± 4% +9.62% (p=0.000 n=19+17) BM_ArenaFuseBalanced/128 9.20µs ± 4% 9.95µs ± 4% +8.12% (p=0.000 n=16+19) BM_LoadAdsDescriptor_Upb<NoLayout> 5.50ms ± 8% 5.69ms ±17% ~ (p=0.189 n=18+19) BM_LoadAdsDescriptor_Upb<WithLayout> 6.10ms ± 5% 6.05ms ± 4% ~ (p=0.258 n=17+18) BM_LoadAdsDescriptor_Proto2<NoLayout> 11.9ms ±15% 11.6ms ± 5% ~ (p=0.589 n=19+16) BM_LoadAdsDescriptor_Proto2<WithLayout> 11.8ms ± 5% 12.4ms ±17% ~ (p=0.604 n=16+20) BM_Parse_Upb_FileDesc<UseArena, Copy> 12.1µs ± 8% 12.1µs ± 4% ~ (p=1.000 n=18+18) BM_Parse_Upb_FileDesc<UseArena, Alias> 11.8µs ±17% 11.1µs ± 4% ~ (p=0.104 n=20+17) BM_Parse_Upb_FileDesc<InitBlock, Copy> 12.0µs ± 5% 11.9µs ± 4% ~ (p=0.134 n=18+19) BM_Parse_Upb_FileDesc<InitBlock, Alias> 10.9µs ± 7% 11.0µs ± 4% ~ (p=0.195 n=17+18) BM_Parse_Proto2<FileDesc, NoArena, Copy> 24.2µs ± 4% 24.4µs ± 7% ~ (p=0.767 n=18+18) BM_Parse_Proto2<FileDesc, UseArena, Copy> 11.6µs ± 5% 11.6µs ± 4% ~ (p=0.621 n=18+16) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 11.3µs ± 3% 11.3µs ± 3% ~ (p=0.743 n=18+18) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 12.7µs ± 8% 12.7µs ± 4% ~ (p=0.988 n=18+19) BM_SerializeDescriptor_Proto2 5.77µs ± 5% 5.71µs ± 5% ~ (p=0.433 n=17+17) BM_SerializeDescriptor_Upb 10.0µs ± 5% 10.1µs ± 7% ~ (p=0.102 n=19+16) name old time/op new time/op delta BM_ArenaOneAlloc 18.4ns ± 6% 18.8ns ± 4% +1.98% (p=0.019 n=18+18) BM_ArenaInitialBlockOneAlloc 5.51ns ± 4% 6.58ns ± 4% +19.42% (p=0.000 n=16+17) BM_ArenaFuseUnbalanced/2 59.5ns ±10% 68.9ns ± 4% +15.83% (p=0.000 n=19+19) BM_ArenaFuseUnbalanced/8 481ns ± 5% 541ns ± 8% +12.54% (p=0.000 n=18+19) BM_ArenaFuseUnbalanced/64 4.51µs ± 4% 4.94µs ± 8% +9.53% (p=0.000 n=17+17) BM_ArenaFuseUnbalanced/128 9.26µs ± 3% 9.98µs ± 3% +7.79% (p=0.000 n=17+17) BM_ArenaFuseBalanced/2 63.5ns ±19% 71.1ns ± 3% +12.07% (p=0.000 n=19+18) BM_ArenaFuseBalanced/8 485ns ± 9% 551ns ±20% +13.47% (p=0.000 n=17+17) BM_ArenaFuseBalanced/64 4.51µs ± 6% 4.95µs ± 4% +9.62% (p=0.000 n=19+17) BM_ArenaFuseBalanced/128 9.22µs ± 4% 9.97µs ± 4% +8.12% (p=0.000 n=16+19) BM_LoadAdsDescriptor_Upb<NoLayout> 5.52ms ± 8% 5.72ms ±18% ~ (p=0.199 n=18+19) BM_LoadAdsDescriptor_Upb<WithLayout> 6.12ms ± 5% 6.07ms ± 4% ~ (p=0.273 n=17+18) BM_LoadAdsDescriptor_Proto2<NoLayout> 11.9ms ±15% 11.6ms ± 5% ~ (p=0.589 n=19+16) BM_LoadAdsDescriptor_Proto2<WithLayout> 11.9ms ± 5% 12.5ms ±18% ~ (p=0.582 n=16+20) BM_Parse_Upb_FileDesc<UseArena, Copy> 12.2µs ± 8% 12.1µs ± 3% ~ (p=0.963 n=18+18) BM_Parse_Upb_FileDesc<UseArena, Alias> 11.8µs ±17% 11.1µs ± 4% ~ (p=0.104 n=20+17) BM_Parse_Upb_FileDesc<InitBlock, Copy> 12.0µs ± 5% 11.9µs ± 4% ~ (p=0.126 n=18+19) BM_Parse_Upb_FileDesc<InitBlock, Alias> 11.0µs ± 6% 11.1µs ± 4% ~ (p=0.195 n=17+18) BM_Parse_Proto2<FileDesc, NoArena, Copy> 24.3µs ± 4% 24.5µs ± 6% ~ (p=0.743 n=18+18) BM_Parse_Proto2<FileDesc, UseArena, Copy> 11.7µs ± 5% 11.6µs ± 4% ~ (p=0.574 n=18+16) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 11.3µs ± 3% 11.3µs ± 3% ~ (p=0.743 n=18+18) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 12.7µs ± 8% 12.7µs ± 4% ~ (p=0.988 n=18+19) BM_SerializeDescriptor_Proto2 5.78µs ± 5% 5.73µs ± 5% ~ (p=0.357 n=17+17) BM_SerializeDescriptor_Upb 10.0µs ± 5% 10.1µs ± 7% ~ (p=0.117 n=19+16) name old allocs/op new allocs/op delta BM_ArenaOneAlloc 1.00 ± 0% 1.00 ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/2 2.00 ± 0% 2.00 ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/8 8.00 ± 0% 8.00 ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/64 64.0 ± 0% 64.0 ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/128 128 ± 0% 128 ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/2 2.00 ± 0% 2.00 ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/8 8.00 ± 0% 8.00 ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/64 64.0 ± 0% 64.0 ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/128 128 ± 0% 128 ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Upb<NoLayout> 6.08k ± 0% 6.05k ± 0% -0.54% (p=0.000 n=20+20) BM_LoadAdsDescriptor_Upb<WithLayout> 6.39k ± 0% 6.36k ± 0% -0.55% (p=0.000 n=20+20) BM_LoadAdsDescriptor_Proto2<NoLayout> 83.4k ± 0% 83.4k ± 0% ~ (p=0.800 n=20+20) BM_LoadAdsDescriptor_Proto2<WithLayout> 84.4k ± 0% 84.4k ± 0% ~ (p=0.752 n=20+20) BM_Parse_Upb_FileDesc<UseArena, Copy> 7.00 ± 0% 7.00 ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<UseArena, Alias> 7.00 ± 0% 7.00 ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, NoArena, Copy> 765 ± 0% 765 ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, UseArena, Copy> 8.00 ± 0% 8.00 ± 0% ~ (all samples are equal) name old peak-mem(Bytes)/op new peak-mem(Bytes)/op delta BM_ArenaOneAlloc 336 ± 0% 336 ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/2 672 ± 0% 672 ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/8 2.69k ± 0% 2.69k ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/64 21.5k ± 0% 21.5k ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/128 43.0k ± 0% 43.0k ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/2 672 ± 0% 672 ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/8 2.69k ± 0% 2.69k ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/64 21.5k ± 0% 21.5k ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/128 43.0k ± 0% 43.0k ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Upb<NoLayout> 9.89M ± 0% 9.95M ± 0% +0.65% (p=0.000 n=20+20) BM_LoadAdsDescriptor_Upb<WithLayout> 9.95M ± 0% 10.02M ± 0% +0.70% (p=0.000 n=20+20) BM_LoadAdsDescriptor_Proto2<NoLayout> 6.62M ± 0% 6.62M ± 0% ~ (p=0.800 n=20+20) BM_LoadAdsDescriptor_Proto2<WithLayout> 6.66M ± 0% 6.66M ± 0% ~ (p=0.752 n=20+20) BM_Parse_Upb_FileDesc<UseArena, Copy> 36.5k ± 0% 36.5k ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<UseArena, Alias> 36.5k ± 0% 36.5k ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, NoArena, Copy> 35.8k ± 0% 35.8k ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, UseArena, Copy> 65.3k ± 0% 65.3k ± 0% ~ (all samples are equal) name old speed new speed delta BM_LoadAdsDescriptor_Upb<NoLayout> 138MB/s ± 7% 132MB/s ±15% ~ (p=0.126 n=18+20) BM_LoadAdsDescriptor_Upb<WithLayout> 124MB/s ± 5% 125MB/s ± 4% ~ (p=0.258 n=17+18) BM_LoadAdsDescriptor_Proto2<NoLayout> 63.9MB/s ±13% 65.2MB/s ± 5% ~ (p=0.589 n=19+16) BM_LoadAdsDescriptor_Proto2<WithLayout> 64.0MB/s ± 5% 61.3MB/s ±15% ~ (p=0.604 n=16+20) BM_Parse_Upb_FileDesc<UseArena, Copy> 620MB/s ± 8% 622MB/s ± 4% ~ (p=1.000 n=18+18) BM_Parse_Upb_FileDesc<UseArena, Alias> 644MB/s ±15% 679MB/s ± 4% ~ (p=0.104 n=20+17) BM_Parse_Upb_FileDesc<InitBlock, Copy> 627MB/s ± 4% 633MB/s ± 4% ~ (p=0.134 n=18+19) BM_Parse_Upb_FileDesc<InitBlock, Alias> 688MB/s ± 6% 682MB/s ± 4% ~ (p=0.195 n=17+18) BM_Parse_Proto2<FileDesc, NoArena, Copy> 310MB/s ± 4% 309MB/s ± 6% ~ (p=0.767 n=18+18) BM_Parse_Proto2<FileDesc, UseArena, Copy> 646MB/s ± 4% 649MB/s ± 4% ~ (p=0.621 n=18+16) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 666MB/s ± 3% 666MB/s ± 3% ~ (p=0.743 n=18+18) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 592MB/s ± 7% 593MB/s ± 4% ~ (p=0.988 n=18+19) BM_SerializeDescriptor_Proto2 1.30GB/s ± 5% 1.32GB/s ± 5% ~ (p=0.433 n=17+17) BM_SerializeDescriptor_Upb 756MB/s ± 5% 745MB/s ± 6% ~ (p=0.102 n=19+16) ``` PiperOrigin-RevId: 520144430
2 years ago
upb_EpsCopyInputStream_Init(&decoder.input, &buf, size,
options & kUpb_DecodeOption_AliasString);
Changed Arena representation so that fusing links arenas together instead of blocks. Previously when fusing, we would concatenate all blocks into a single list that lived in the arena root. From then on, all arenas would add their blocks to this single unified list. After this CL, arenas keep their distinct list of blocks even after being fused. Instead of unifying the block list, fuse now puts the arenas themselves into a list, so all arenas in the fused group can be iterated over at any time. This design makes it easier to keep each individual arena thread-compatible, because fuse and free are now the only mutating operations that touch state that is shared with the entire group. Read-only operations like `SpaceAllocated()` also iterate the list of arenas, but in a read-only fashion. (Note: we need tests for SpaceAllocated(), both single-threaded for correctness and multi-threaded for resilience to crashes and data races). Performance of fuse regresses by 5-20%. This is somewhat expected as we are performing more atomic operations during a fuse. ``` name old cpu/op new cpu/op delta BM_ArenaOneAlloc 18.4ns ± 6% 18.7ns ± 4% +2.00% (p=0.016 n=18+18) BM_ArenaInitialBlockOneAlloc 5.50ns ± 4% 6.57ns ± 4% +19.42% (p=0.000 n=16+17) BM_ArenaFuseUnbalanced/2 59.3ns ±10% 68.7ns ± 4% +15.85% (p=0.000 n=19+19) BM_ArenaFuseUnbalanced/8 479ns ± 5% 540ns ± 8% +12.57% (p=0.000 n=18+19) BM_ArenaFuseUnbalanced/64 4.50µs ± 4% 4.93µs ± 8% +9.59% (p=0.000 n=17+17) BM_ArenaFuseUnbalanced/128 9.24µs ± 3% 9.96µs ± 3% +7.81% (p=0.000 n=17+17) BM_ArenaFuseBalanced/2 63.3ns ±18% 71.0ns ± 4% +12.14% (p=0.000 n=19+18) BM_ArenaFuseBalanced/8 484ns ± 9% 543ns ±10% +12.11% (p=0.000 n=17+16) BM_ArenaFuseBalanced/64 4.50µs ± 6% 4.94µs ± 4% +9.62% (p=0.000 n=19+17) BM_ArenaFuseBalanced/128 9.20µs ± 4% 9.95µs ± 4% +8.12% (p=0.000 n=16+19) BM_LoadAdsDescriptor_Upb<NoLayout> 5.50ms ± 8% 5.69ms ±17% ~ (p=0.189 n=18+19) BM_LoadAdsDescriptor_Upb<WithLayout> 6.10ms ± 5% 6.05ms ± 4% ~ (p=0.258 n=17+18) BM_LoadAdsDescriptor_Proto2<NoLayout> 11.9ms ±15% 11.6ms ± 5% ~ (p=0.589 n=19+16) BM_LoadAdsDescriptor_Proto2<WithLayout> 11.8ms ± 5% 12.4ms ±17% ~ (p=0.604 n=16+20) BM_Parse_Upb_FileDesc<UseArena, Copy> 12.1µs ± 8% 12.1µs ± 4% ~ (p=1.000 n=18+18) BM_Parse_Upb_FileDesc<UseArena, Alias> 11.8µs ±17% 11.1µs ± 4% ~ (p=0.104 n=20+17) BM_Parse_Upb_FileDesc<InitBlock, Copy> 12.0µs ± 5% 11.9µs ± 4% ~ (p=0.134 n=18+19) BM_Parse_Upb_FileDesc<InitBlock, Alias> 10.9µs ± 7% 11.0µs ± 4% ~ (p=0.195 n=17+18) BM_Parse_Proto2<FileDesc, NoArena, Copy> 24.2µs ± 4% 24.4µs ± 7% ~ (p=0.767 n=18+18) BM_Parse_Proto2<FileDesc, UseArena, Copy> 11.6µs ± 5% 11.6µs ± 4% ~ (p=0.621 n=18+16) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 11.3µs ± 3% 11.3µs ± 3% ~ (p=0.743 n=18+18) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 12.7µs ± 8% 12.7µs ± 4% ~ (p=0.988 n=18+19) BM_SerializeDescriptor_Proto2 5.77µs ± 5% 5.71µs ± 5% ~ (p=0.433 n=17+17) BM_SerializeDescriptor_Upb 10.0µs ± 5% 10.1µs ± 7% ~ (p=0.102 n=19+16) name old time/op new time/op delta BM_ArenaOneAlloc 18.4ns ± 6% 18.8ns ± 4% +1.98% (p=0.019 n=18+18) BM_ArenaInitialBlockOneAlloc 5.51ns ± 4% 6.58ns ± 4% +19.42% (p=0.000 n=16+17) BM_ArenaFuseUnbalanced/2 59.5ns ±10% 68.9ns ± 4% +15.83% (p=0.000 n=19+19) BM_ArenaFuseUnbalanced/8 481ns ± 5% 541ns ± 8% +12.54% (p=0.000 n=18+19) BM_ArenaFuseUnbalanced/64 4.51µs ± 4% 4.94µs ± 8% +9.53% (p=0.000 n=17+17) BM_ArenaFuseUnbalanced/128 9.26µs ± 3% 9.98µs ± 3% +7.79% (p=0.000 n=17+17) BM_ArenaFuseBalanced/2 63.5ns ±19% 71.1ns ± 3% +12.07% (p=0.000 n=19+18) BM_ArenaFuseBalanced/8 485ns ± 9% 551ns ±20% +13.47% (p=0.000 n=17+17) BM_ArenaFuseBalanced/64 4.51µs ± 6% 4.95µs ± 4% +9.62% (p=0.000 n=19+17) BM_ArenaFuseBalanced/128 9.22µs ± 4% 9.97µs ± 4% +8.12% (p=0.000 n=16+19) BM_LoadAdsDescriptor_Upb<NoLayout> 5.52ms ± 8% 5.72ms ±18% ~ (p=0.199 n=18+19) BM_LoadAdsDescriptor_Upb<WithLayout> 6.12ms ± 5% 6.07ms ± 4% ~ (p=0.273 n=17+18) BM_LoadAdsDescriptor_Proto2<NoLayout> 11.9ms ±15% 11.6ms ± 5% ~ (p=0.589 n=19+16) BM_LoadAdsDescriptor_Proto2<WithLayout> 11.9ms ± 5% 12.5ms ±18% ~ (p=0.582 n=16+20) BM_Parse_Upb_FileDesc<UseArena, Copy> 12.2µs ± 8% 12.1µs ± 3% ~ (p=0.963 n=18+18) BM_Parse_Upb_FileDesc<UseArena, Alias> 11.8µs ±17% 11.1µs ± 4% ~ (p=0.104 n=20+17) BM_Parse_Upb_FileDesc<InitBlock, Copy> 12.0µs ± 5% 11.9µs ± 4% ~ (p=0.126 n=18+19) BM_Parse_Upb_FileDesc<InitBlock, Alias> 11.0µs ± 6% 11.1µs ± 4% ~ (p=0.195 n=17+18) BM_Parse_Proto2<FileDesc, NoArena, Copy> 24.3µs ± 4% 24.5µs ± 6% ~ (p=0.743 n=18+18) BM_Parse_Proto2<FileDesc, UseArena, Copy> 11.7µs ± 5% 11.6µs ± 4% ~ (p=0.574 n=18+16) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 11.3µs ± 3% 11.3µs ± 3% ~ (p=0.743 n=18+18) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 12.7µs ± 8% 12.7µs ± 4% ~ (p=0.988 n=18+19) BM_SerializeDescriptor_Proto2 5.78µs ± 5% 5.73µs ± 5% ~ (p=0.357 n=17+17) BM_SerializeDescriptor_Upb 10.0µs ± 5% 10.1µs ± 7% ~ (p=0.117 n=19+16) name old allocs/op new allocs/op delta BM_ArenaOneAlloc 1.00 ± 0% 1.00 ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/2 2.00 ± 0% 2.00 ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/8 8.00 ± 0% 8.00 ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/64 64.0 ± 0% 64.0 ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/128 128 ± 0% 128 ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/2 2.00 ± 0% 2.00 ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/8 8.00 ± 0% 8.00 ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/64 64.0 ± 0% 64.0 ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/128 128 ± 0% 128 ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Upb<NoLayout> 6.08k ± 0% 6.05k ± 0% -0.54% (p=0.000 n=20+20) BM_LoadAdsDescriptor_Upb<WithLayout> 6.39k ± 0% 6.36k ± 0% -0.55% (p=0.000 n=20+20) BM_LoadAdsDescriptor_Proto2<NoLayout> 83.4k ± 0% 83.4k ± 0% ~ (p=0.800 n=20+20) BM_LoadAdsDescriptor_Proto2<WithLayout> 84.4k ± 0% 84.4k ± 0% ~ (p=0.752 n=20+20) BM_Parse_Upb_FileDesc<UseArena, Copy> 7.00 ± 0% 7.00 ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<UseArena, Alias> 7.00 ± 0% 7.00 ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, NoArena, Copy> 765 ± 0% 765 ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, UseArena, Copy> 8.00 ± 0% 8.00 ± 0% ~ (all samples are equal) name old peak-mem(Bytes)/op new peak-mem(Bytes)/op delta BM_ArenaOneAlloc 336 ± 0% 336 ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/2 672 ± 0% 672 ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/8 2.69k ± 0% 2.69k ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/64 21.5k ± 0% 21.5k ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/128 43.0k ± 0% 43.0k ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/2 672 ± 0% 672 ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/8 2.69k ± 0% 2.69k ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/64 21.5k ± 0% 21.5k ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/128 43.0k ± 0% 43.0k ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Upb<NoLayout> 9.89M ± 0% 9.95M ± 0% +0.65% (p=0.000 n=20+20) BM_LoadAdsDescriptor_Upb<WithLayout> 9.95M ± 0% 10.02M ± 0% +0.70% (p=0.000 n=20+20) BM_LoadAdsDescriptor_Proto2<NoLayout> 6.62M ± 0% 6.62M ± 0% ~ (p=0.800 n=20+20) BM_LoadAdsDescriptor_Proto2<WithLayout> 6.66M ± 0% 6.66M ± 0% ~ (p=0.752 n=20+20) BM_Parse_Upb_FileDesc<UseArena, Copy> 36.5k ± 0% 36.5k ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<UseArena, Alias> 36.5k ± 0% 36.5k ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, NoArena, Copy> 35.8k ± 0% 35.8k ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, UseArena, Copy> 65.3k ± 0% 65.3k ± 0% ~ (all samples are equal) name old speed new speed delta BM_LoadAdsDescriptor_Upb<NoLayout> 138MB/s ± 7% 132MB/s ±15% ~ (p=0.126 n=18+20) BM_LoadAdsDescriptor_Upb<WithLayout> 124MB/s ± 5% 125MB/s ± 4% ~ (p=0.258 n=17+18) BM_LoadAdsDescriptor_Proto2<NoLayout> 63.9MB/s ±13% 65.2MB/s ± 5% ~ (p=0.589 n=19+16) BM_LoadAdsDescriptor_Proto2<WithLayout> 64.0MB/s ± 5% 61.3MB/s ±15% ~ (p=0.604 n=16+20) BM_Parse_Upb_FileDesc<UseArena, Copy> 620MB/s ± 8% 622MB/s ± 4% ~ (p=1.000 n=18+18) BM_Parse_Upb_FileDesc<UseArena, Alias> 644MB/s ±15% 679MB/s ± 4% ~ (p=0.104 n=20+17) BM_Parse_Upb_FileDesc<InitBlock, Copy> 627MB/s ± 4% 633MB/s ± 4% ~ (p=0.134 n=18+19) BM_Parse_Upb_FileDesc<InitBlock, Alias> 688MB/s ± 6% 682MB/s ± 4% ~ (p=0.195 n=17+18) BM_Parse_Proto2<FileDesc, NoArena, Copy> 310MB/s ± 4% 309MB/s ± 6% ~ (p=0.767 n=18+18) BM_Parse_Proto2<FileDesc, UseArena, Copy> 646MB/s ± 4% 649MB/s ± 4% ~ (p=0.621 n=18+16) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 666MB/s ± 3% 666MB/s ± 3% ~ (p=0.743 n=18+18) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 592MB/s ± 7% 593MB/s ± 4% ~ (p=0.988 n=18+19) BM_SerializeDescriptor_Proto2 1.30GB/s ± 5% 1.32GB/s ± 5% ~ (p=0.433 n=17+17) BM_SerializeDescriptor_Upb 756MB/s ± 5% 745MB/s ± 6% ~ (p=0.102 n=19+16) ``` PiperOrigin-RevId: 520144430
2 years ago
decoder.extreg = extreg;
decoder.unknown = NULL;
decoder.depth = depth ? depth : kUpb_WireFormat_DefaultDepthLimit;
Changed Arena representation so that fusing links arenas together instead of blocks. Previously when fusing, we would concatenate all blocks into a single list that lived in the arena root. From then on, all arenas would add their blocks to this single unified list. After this CL, arenas keep their distinct list of blocks even after being fused. Instead of unifying the block list, fuse now puts the arenas themselves into a list, so all arenas in the fused group can be iterated over at any time. This design makes it easier to keep each individual arena thread-compatible, because fuse and free are now the only mutating operations that touch state that is shared with the entire group. Read-only operations like `SpaceAllocated()` also iterate the list of arenas, but in a read-only fashion. (Note: we need tests for SpaceAllocated(), both single-threaded for correctness and multi-threaded for resilience to crashes and data races). Performance of fuse regresses by 5-20%. This is somewhat expected as we are performing more atomic operations during a fuse. ``` name old cpu/op new cpu/op delta BM_ArenaOneAlloc 18.4ns ± 6% 18.7ns ± 4% +2.00% (p=0.016 n=18+18) BM_ArenaInitialBlockOneAlloc 5.50ns ± 4% 6.57ns ± 4% +19.42% (p=0.000 n=16+17) BM_ArenaFuseUnbalanced/2 59.3ns ±10% 68.7ns ± 4% +15.85% (p=0.000 n=19+19) BM_ArenaFuseUnbalanced/8 479ns ± 5% 540ns ± 8% +12.57% (p=0.000 n=18+19) BM_ArenaFuseUnbalanced/64 4.50µs ± 4% 4.93µs ± 8% +9.59% (p=0.000 n=17+17) BM_ArenaFuseUnbalanced/128 9.24µs ± 3% 9.96µs ± 3% +7.81% (p=0.000 n=17+17) BM_ArenaFuseBalanced/2 63.3ns ±18% 71.0ns ± 4% +12.14% (p=0.000 n=19+18) BM_ArenaFuseBalanced/8 484ns ± 9% 543ns ±10% +12.11% (p=0.000 n=17+16) BM_ArenaFuseBalanced/64 4.50µs ± 6% 4.94µs ± 4% +9.62% (p=0.000 n=19+17) BM_ArenaFuseBalanced/128 9.20µs ± 4% 9.95µs ± 4% +8.12% (p=0.000 n=16+19) BM_LoadAdsDescriptor_Upb<NoLayout> 5.50ms ± 8% 5.69ms ±17% ~ (p=0.189 n=18+19) BM_LoadAdsDescriptor_Upb<WithLayout> 6.10ms ± 5% 6.05ms ± 4% ~ (p=0.258 n=17+18) BM_LoadAdsDescriptor_Proto2<NoLayout> 11.9ms ±15% 11.6ms ± 5% ~ (p=0.589 n=19+16) BM_LoadAdsDescriptor_Proto2<WithLayout> 11.8ms ± 5% 12.4ms ±17% ~ (p=0.604 n=16+20) BM_Parse_Upb_FileDesc<UseArena, Copy> 12.1µs ± 8% 12.1µs ± 4% ~ (p=1.000 n=18+18) BM_Parse_Upb_FileDesc<UseArena, Alias> 11.8µs ±17% 11.1µs ± 4% ~ (p=0.104 n=20+17) BM_Parse_Upb_FileDesc<InitBlock, Copy> 12.0µs ± 5% 11.9µs ± 4% ~ (p=0.134 n=18+19) BM_Parse_Upb_FileDesc<InitBlock, Alias> 10.9µs ± 7% 11.0µs ± 4% ~ (p=0.195 n=17+18) BM_Parse_Proto2<FileDesc, NoArena, Copy> 24.2µs ± 4% 24.4µs ± 7% ~ (p=0.767 n=18+18) BM_Parse_Proto2<FileDesc, UseArena, Copy> 11.6µs ± 5% 11.6µs ± 4% ~ (p=0.621 n=18+16) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 11.3µs ± 3% 11.3µs ± 3% ~ (p=0.743 n=18+18) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 12.7µs ± 8% 12.7µs ± 4% ~ (p=0.988 n=18+19) BM_SerializeDescriptor_Proto2 5.77µs ± 5% 5.71µs ± 5% ~ (p=0.433 n=17+17) BM_SerializeDescriptor_Upb 10.0µs ± 5% 10.1µs ± 7% ~ (p=0.102 n=19+16) name old time/op new time/op delta BM_ArenaOneAlloc 18.4ns ± 6% 18.8ns ± 4% +1.98% (p=0.019 n=18+18) BM_ArenaInitialBlockOneAlloc 5.51ns ± 4% 6.58ns ± 4% +19.42% (p=0.000 n=16+17) BM_ArenaFuseUnbalanced/2 59.5ns ±10% 68.9ns ± 4% +15.83% (p=0.000 n=19+19) BM_ArenaFuseUnbalanced/8 481ns ± 5% 541ns ± 8% +12.54% (p=0.000 n=18+19) BM_ArenaFuseUnbalanced/64 4.51µs ± 4% 4.94µs ± 8% +9.53% (p=0.000 n=17+17) BM_ArenaFuseUnbalanced/128 9.26µs ± 3% 9.98µs ± 3% +7.79% (p=0.000 n=17+17) BM_ArenaFuseBalanced/2 63.5ns ±19% 71.1ns ± 3% +12.07% (p=0.000 n=19+18) BM_ArenaFuseBalanced/8 485ns ± 9% 551ns ±20% +13.47% (p=0.000 n=17+17) BM_ArenaFuseBalanced/64 4.51µs ± 6% 4.95µs ± 4% +9.62% (p=0.000 n=19+17) BM_ArenaFuseBalanced/128 9.22µs ± 4% 9.97µs ± 4% +8.12% (p=0.000 n=16+19) BM_LoadAdsDescriptor_Upb<NoLayout> 5.52ms ± 8% 5.72ms ±18% ~ (p=0.199 n=18+19) BM_LoadAdsDescriptor_Upb<WithLayout> 6.12ms ± 5% 6.07ms ± 4% ~ (p=0.273 n=17+18) BM_LoadAdsDescriptor_Proto2<NoLayout> 11.9ms ±15% 11.6ms ± 5% ~ (p=0.589 n=19+16) BM_LoadAdsDescriptor_Proto2<WithLayout> 11.9ms ± 5% 12.5ms ±18% ~ (p=0.582 n=16+20) BM_Parse_Upb_FileDesc<UseArena, Copy> 12.2µs ± 8% 12.1µs ± 3% ~ (p=0.963 n=18+18) BM_Parse_Upb_FileDesc<UseArena, Alias> 11.8µs ±17% 11.1µs ± 4% ~ (p=0.104 n=20+17) BM_Parse_Upb_FileDesc<InitBlock, Copy> 12.0µs ± 5% 11.9µs ± 4% ~ (p=0.126 n=18+19) BM_Parse_Upb_FileDesc<InitBlock, Alias> 11.0µs ± 6% 11.1µs ± 4% ~ (p=0.195 n=17+18) BM_Parse_Proto2<FileDesc, NoArena, Copy> 24.3µs ± 4% 24.5µs ± 6% ~ (p=0.743 n=18+18) BM_Parse_Proto2<FileDesc, UseArena, Copy> 11.7µs ± 5% 11.6µs ± 4% ~ (p=0.574 n=18+16) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 11.3µs ± 3% 11.3µs ± 3% ~ (p=0.743 n=18+18) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 12.7µs ± 8% 12.7µs ± 4% ~ (p=0.988 n=18+19) BM_SerializeDescriptor_Proto2 5.78µs ± 5% 5.73µs ± 5% ~ (p=0.357 n=17+17) BM_SerializeDescriptor_Upb 10.0µs ± 5% 10.1µs ± 7% ~ (p=0.117 n=19+16) name old allocs/op new allocs/op delta BM_ArenaOneAlloc 1.00 ± 0% 1.00 ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/2 2.00 ± 0% 2.00 ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/8 8.00 ± 0% 8.00 ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/64 64.0 ± 0% 64.0 ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/128 128 ± 0% 128 ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/2 2.00 ± 0% 2.00 ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/8 8.00 ± 0% 8.00 ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/64 64.0 ± 0% 64.0 ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/128 128 ± 0% 128 ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Upb<NoLayout> 6.08k ± 0% 6.05k ± 0% -0.54% (p=0.000 n=20+20) BM_LoadAdsDescriptor_Upb<WithLayout> 6.39k ± 0% 6.36k ± 0% -0.55% (p=0.000 n=20+20) BM_LoadAdsDescriptor_Proto2<NoLayout> 83.4k ± 0% 83.4k ± 0% ~ (p=0.800 n=20+20) BM_LoadAdsDescriptor_Proto2<WithLayout> 84.4k ± 0% 84.4k ± 0% ~ (p=0.752 n=20+20) BM_Parse_Upb_FileDesc<UseArena, Copy> 7.00 ± 0% 7.00 ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<UseArena, Alias> 7.00 ± 0% 7.00 ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, NoArena, Copy> 765 ± 0% 765 ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, UseArena, Copy> 8.00 ± 0% 8.00 ± 0% ~ (all samples are equal) name old peak-mem(Bytes)/op new peak-mem(Bytes)/op delta BM_ArenaOneAlloc 336 ± 0% 336 ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/2 672 ± 0% 672 ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/8 2.69k ± 0% 2.69k ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/64 21.5k ± 0% 21.5k ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/128 43.0k ± 0% 43.0k ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/2 672 ± 0% 672 ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/8 2.69k ± 0% 2.69k ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/64 21.5k ± 0% 21.5k ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/128 43.0k ± 0% 43.0k ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Upb<NoLayout> 9.89M ± 0% 9.95M ± 0% +0.65% (p=0.000 n=20+20) BM_LoadAdsDescriptor_Upb<WithLayout> 9.95M ± 0% 10.02M ± 0% +0.70% (p=0.000 n=20+20) BM_LoadAdsDescriptor_Proto2<NoLayout> 6.62M ± 0% 6.62M ± 0% ~ (p=0.800 n=20+20) BM_LoadAdsDescriptor_Proto2<WithLayout> 6.66M ± 0% 6.66M ± 0% ~ (p=0.752 n=20+20) BM_Parse_Upb_FileDesc<UseArena, Copy> 36.5k ± 0% 36.5k ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<UseArena, Alias> 36.5k ± 0% 36.5k ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, NoArena, Copy> 35.8k ± 0% 35.8k ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, UseArena, Copy> 65.3k ± 0% 65.3k ± 0% ~ (all samples are equal) name old speed new speed delta BM_LoadAdsDescriptor_Upb<NoLayout> 138MB/s ± 7% 132MB/s ±15% ~ (p=0.126 n=18+20) BM_LoadAdsDescriptor_Upb<WithLayout> 124MB/s ± 5% 125MB/s ± 4% ~ (p=0.258 n=17+18) BM_LoadAdsDescriptor_Proto2<NoLayout> 63.9MB/s ±13% 65.2MB/s ± 5% ~ (p=0.589 n=19+16) BM_LoadAdsDescriptor_Proto2<WithLayout> 64.0MB/s ± 5% 61.3MB/s ±15% ~ (p=0.604 n=16+20) BM_Parse_Upb_FileDesc<UseArena, Copy> 620MB/s ± 8% 622MB/s ± 4% ~ (p=1.000 n=18+18) BM_Parse_Upb_FileDesc<UseArena, Alias> 644MB/s ±15% 679MB/s ± 4% ~ (p=0.104 n=20+17) BM_Parse_Upb_FileDesc<InitBlock, Copy> 627MB/s ± 4% 633MB/s ± 4% ~ (p=0.134 n=18+19) BM_Parse_Upb_FileDesc<InitBlock, Alias> 688MB/s ± 6% 682MB/s ± 4% ~ (p=0.195 n=17+18) BM_Parse_Proto2<FileDesc, NoArena, Copy> 310MB/s ± 4% 309MB/s ± 6% ~ (p=0.767 n=18+18) BM_Parse_Proto2<FileDesc, UseArena, Copy> 646MB/s ± 4% 649MB/s ± 4% ~ (p=0.621 n=18+16) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 666MB/s ± 3% 666MB/s ± 3% ~ (p=0.743 n=18+18) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 592MB/s ± 7% 593MB/s ± 4% ~ (p=0.988 n=18+19) BM_SerializeDescriptor_Proto2 1.30GB/s ± 5% 1.32GB/s ± 5% ~ (p=0.433 n=17+17) BM_SerializeDescriptor_Upb 756MB/s ± 5% 745MB/s ± 6% ~ (p=0.102 n=19+16) ``` PiperOrigin-RevId: 520144430
2 years ago
decoder.end_group = DECODE_NOGROUP;
decoder.options = (uint16_t)options;
decoder.missing_required = false;
decoder.status = kUpb_DecodeStatus_Ok;
// Violating the encapsulation of the arena for performance reasons.
// This is a temporary arena that we swap into and swap out of when we are
// done. The temporary arena only needs to be able to handle allocation,
// not fuse or free, so it does not need many of the members to be initialized
// (particularly parent_or_count).
UPB_PRIVATE(_upb_Arena_SwapIn)(&decoder.arena, arena);
Changed Arena representation so that fusing links arenas together instead of blocks. Previously when fusing, we would concatenate all blocks into a single list that lived in the arena root. From then on, all arenas would add their blocks to this single unified list. After this CL, arenas keep their distinct list of blocks even after being fused. Instead of unifying the block list, fuse now puts the arenas themselves into a list, so all arenas in the fused group can be iterated over at any time. This design makes it easier to keep each individual arena thread-compatible, because fuse and free are now the only mutating operations that touch state that is shared with the entire group. Read-only operations like `SpaceAllocated()` also iterate the list of arenas, but in a read-only fashion. (Note: we need tests for SpaceAllocated(), both single-threaded for correctness and multi-threaded for resilience to crashes and data races). Performance of fuse regresses by 5-20%. This is somewhat expected as we are performing more atomic operations during a fuse. ``` name old cpu/op new cpu/op delta BM_ArenaOneAlloc 18.4ns ± 6% 18.7ns ± 4% +2.00% (p=0.016 n=18+18) BM_ArenaInitialBlockOneAlloc 5.50ns ± 4% 6.57ns ± 4% +19.42% (p=0.000 n=16+17) BM_ArenaFuseUnbalanced/2 59.3ns ±10% 68.7ns ± 4% +15.85% (p=0.000 n=19+19) BM_ArenaFuseUnbalanced/8 479ns ± 5% 540ns ± 8% +12.57% (p=0.000 n=18+19) BM_ArenaFuseUnbalanced/64 4.50µs ± 4% 4.93µs ± 8% +9.59% (p=0.000 n=17+17) BM_ArenaFuseUnbalanced/128 9.24µs ± 3% 9.96µs ± 3% +7.81% (p=0.000 n=17+17) BM_ArenaFuseBalanced/2 63.3ns ±18% 71.0ns ± 4% +12.14% (p=0.000 n=19+18) BM_ArenaFuseBalanced/8 484ns ± 9% 543ns ±10% +12.11% (p=0.000 n=17+16) BM_ArenaFuseBalanced/64 4.50µs ± 6% 4.94µs ± 4% +9.62% (p=0.000 n=19+17) BM_ArenaFuseBalanced/128 9.20µs ± 4% 9.95µs ± 4% +8.12% (p=0.000 n=16+19) BM_LoadAdsDescriptor_Upb<NoLayout> 5.50ms ± 8% 5.69ms ±17% ~ (p=0.189 n=18+19) BM_LoadAdsDescriptor_Upb<WithLayout> 6.10ms ± 5% 6.05ms ± 4% ~ (p=0.258 n=17+18) BM_LoadAdsDescriptor_Proto2<NoLayout> 11.9ms ±15% 11.6ms ± 5% ~ (p=0.589 n=19+16) BM_LoadAdsDescriptor_Proto2<WithLayout> 11.8ms ± 5% 12.4ms ±17% ~ (p=0.604 n=16+20) BM_Parse_Upb_FileDesc<UseArena, Copy> 12.1µs ± 8% 12.1µs ± 4% ~ (p=1.000 n=18+18) BM_Parse_Upb_FileDesc<UseArena, Alias> 11.8µs ±17% 11.1µs ± 4% ~ (p=0.104 n=20+17) BM_Parse_Upb_FileDesc<InitBlock, Copy> 12.0µs ± 5% 11.9µs ± 4% ~ (p=0.134 n=18+19) BM_Parse_Upb_FileDesc<InitBlock, Alias> 10.9µs ± 7% 11.0µs ± 4% ~ (p=0.195 n=17+18) BM_Parse_Proto2<FileDesc, NoArena, Copy> 24.2µs ± 4% 24.4µs ± 7% ~ (p=0.767 n=18+18) BM_Parse_Proto2<FileDesc, UseArena, Copy> 11.6µs ± 5% 11.6µs ± 4% ~ (p=0.621 n=18+16) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 11.3µs ± 3% 11.3µs ± 3% ~ (p=0.743 n=18+18) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 12.7µs ± 8% 12.7µs ± 4% ~ (p=0.988 n=18+19) BM_SerializeDescriptor_Proto2 5.77µs ± 5% 5.71µs ± 5% ~ (p=0.433 n=17+17) BM_SerializeDescriptor_Upb 10.0µs ± 5% 10.1µs ± 7% ~ (p=0.102 n=19+16) name old time/op new time/op delta BM_ArenaOneAlloc 18.4ns ± 6% 18.8ns ± 4% +1.98% (p=0.019 n=18+18) BM_ArenaInitialBlockOneAlloc 5.51ns ± 4% 6.58ns ± 4% +19.42% (p=0.000 n=16+17) BM_ArenaFuseUnbalanced/2 59.5ns ±10% 68.9ns ± 4% +15.83% (p=0.000 n=19+19) BM_ArenaFuseUnbalanced/8 481ns ± 5% 541ns ± 8% +12.54% (p=0.000 n=18+19) BM_ArenaFuseUnbalanced/64 4.51µs ± 4% 4.94µs ± 8% +9.53% (p=0.000 n=17+17) BM_ArenaFuseUnbalanced/128 9.26µs ± 3% 9.98µs ± 3% +7.79% (p=0.000 n=17+17) BM_ArenaFuseBalanced/2 63.5ns ±19% 71.1ns ± 3% +12.07% (p=0.000 n=19+18) BM_ArenaFuseBalanced/8 485ns ± 9% 551ns ±20% +13.47% (p=0.000 n=17+17) BM_ArenaFuseBalanced/64 4.51µs ± 6% 4.95µs ± 4% +9.62% (p=0.000 n=19+17) BM_ArenaFuseBalanced/128 9.22µs ± 4% 9.97µs ± 4% +8.12% (p=0.000 n=16+19) BM_LoadAdsDescriptor_Upb<NoLayout> 5.52ms ± 8% 5.72ms ±18% ~ (p=0.199 n=18+19) BM_LoadAdsDescriptor_Upb<WithLayout> 6.12ms ± 5% 6.07ms ± 4% ~ (p=0.273 n=17+18) BM_LoadAdsDescriptor_Proto2<NoLayout> 11.9ms ±15% 11.6ms ± 5% ~ (p=0.589 n=19+16) BM_LoadAdsDescriptor_Proto2<WithLayout> 11.9ms ± 5% 12.5ms ±18% ~ (p=0.582 n=16+20) BM_Parse_Upb_FileDesc<UseArena, Copy> 12.2µs ± 8% 12.1µs ± 3% ~ (p=0.963 n=18+18) BM_Parse_Upb_FileDesc<UseArena, Alias> 11.8µs ±17% 11.1µs ± 4% ~ (p=0.104 n=20+17) BM_Parse_Upb_FileDesc<InitBlock, Copy> 12.0µs ± 5% 11.9µs ± 4% ~ (p=0.126 n=18+19) BM_Parse_Upb_FileDesc<InitBlock, Alias> 11.0µs ± 6% 11.1µs ± 4% ~ (p=0.195 n=17+18) BM_Parse_Proto2<FileDesc, NoArena, Copy> 24.3µs ± 4% 24.5µs ± 6% ~ (p=0.743 n=18+18) BM_Parse_Proto2<FileDesc, UseArena, Copy> 11.7µs ± 5% 11.6µs ± 4% ~ (p=0.574 n=18+16) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 11.3µs ± 3% 11.3µs ± 3% ~ (p=0.743 n=18+18) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 12.7µs ± 8% 12.7µs ± 4% ~ (p=0.988 n=18+19) BM_SerializeDescriptor_Proto2 5.78µs ± 5% 5.73µs ± 5% ~ (p=0.357 n=17+17) BM_SerializeDescriptor_Upb 10.0µs ± 5% 10.1µs ± 7% ~ (p=0.117 n=19+16) name old allocs/op new allocs/op delta BM_ArenaOneAlloc 1.00 ± 0% 1.00 ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/2 2.00 ± 0% 2.00 ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/8 8.00 ± 0% 8.00 ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/64 64.0 ± 0% 64.0 ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/128 128 ± 0% 128 ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/2 2.00 ± 0% 2.00 ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/8 8.00 ± 0% 8.00 ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/64 64.0 ± 0% 64.0 ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/128 128 ± 0% 128 ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Upb<NoLayout> 6.08k ± 0% 6.05k ± 0% -0.54% (p=0.000 n=20+20) BM_LoadAdsDescriptor_Upb<WithLayout> 6.39k ± 0% 6.36k ± 0% -0.55% (p=0.000 n=20+20) BM_LoadAdsDescriptor_Proto2<NoLayout> 83.4k ± 0% 83.4k ± 0% ~ (p=0.800 n=20+20) BM_LoadAdsDescriptor_Proto2<WithLayout> 84.4k ± 0% 84.4k ± 0% ~ (p=0.752 n=20+20) BM_Parse_Upb_FileDesc<UseArena, Copy> 7.00 ± 0% 7.00 ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<UseArena, Alias> 7.00 ± 0% 7.00 ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, NoArena, Copy> 765 ± 0% 765 ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, UseArena, Copy> 8.00 ± 0% 8.00 ± 0% ~ (all samples are equal) name old peak-mem(Bytes)/op new peak-mem(Bytes)/op delta BM_ArenaOneAlloc 336 ± 0% 336 ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/2 672 ± 0% 672 ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/8 2.69k ± 0% 2.69k ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/64 21.5k ± 0% 21.5k ± 0% ~ (all samples are equal) BM_ArenaFuseUnbalanced/128 43.0k ± 0% 43.0k ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/2 672 ± 0% 672 ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/8 2.69k ± 0% 2.69k ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/64 21.5k ± 0% 21.5k ± 0% ~ (all samples are equal) BM_ArenaFuseBalanced/128 43.0k ± 0% 43.0k ± 0% ~ (all samples are equal) BM_LoadAdsDescriptor_Upb<NoLayout> 9.89M ± 0% 9.95M ± 0% +0.65% (p=0.000 n=20+20) BM_LoadAdsDescriptor_Upb<WithLayout> 9.95M ± 0% 10.02M ± 0% +0.70% (p=0.000 n=20+20) BM_LoadAdsDescriptor_Proto2<NoLayout> 6.62M ± 0% 6.62M ± 0% ~ (p=0.800 n=20+20) BM_LoadAdsDescriptor_Proto2<WithLayout> 6.66M ± 0% 6.66M ± 0% ~ (p=0.752 n=20+20) BM_Parse_Upb_FileDesc<UseArena, Copy> 36.5k ± 0% 36.5k ± 0% ~ (all samples are equal) BM_Parse_Upb_FileDesc<UseArena, Alias> 36.5k ± 0% 36.5k ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, NoArena, Copy> 35.8k ± 0% 35.8k ± 0% ~ (all samples are equal) BM_Parse_Proto2<FileDesc, UseArena, Copy> 65.3k ± 0% 65.3k ± 0% ~ (all samples are equal) name old speed new speed delta BM_LoadAdsDescriptor_Upb<NoLayout> 138MB/s ± 7% 132MB/s ±15% ~ (p=0.126 n=18+20) BM_LoadAdsDescriptor_Upb<WithLayout> 124MB/s ± 5% 125MB/s ± 4% ~ (p=0.258 n=17+18) BM_LoadAdsDescriptor_Proto2<NoLayout> 63.9MB/s ±13% 65.2MB/s ± 5% ~ (p=0.589 n=19+16) BM_LoadAdsDescriptor_Proto2<WithLayout> 64.0MB/s ± 5% 61.3MB/s ±15% ~ (p=0.604 n=16+20) BM_Parse_Upb_FileDesc<UseArena, Copy> 620MB/s ± 8% 622MB/s ± 4% ~ (p=1.000 n=18+18) BM_Parse_Upb_FileDesc<UseArena, Alias> 644MB/s ±15% 679MB/s ± 4% ~ (p=0.104 n=20+17) BM_Parse_Upb_FileDesc<InitBlock, Copy> 627MB/s ± 4% 633MB/s ± 4% ~ (p=0.134 n=18+19) BM_Parse_Upb_FileDesc<InitBlock, Alias> 688MB/s ± 6% 682MB/s ± 4% ~ (p=0.195 n=17+18) BM_Parse_Proto2<FileDesc, NoArena, Copy> 310MB/s ± 4% 309MB/s ± 6% ~ (p=0.767 n=18+18) BM_Parse_Proto2<FileDesc, UseArena, Copy> 646MB/s ± 4% 649MB/s ± 4% ~ (p=0.621 n=18+16) BM_Parse_Proto2<FileDesc, InitBlock, Copy> 666MB/s ± 3% 666MB/s ± 3% ~ (p=0.743 n=18+18) BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 592MB/s ± 7% 593MB/s ± 4% ~ (p=0.988 n=18+19) BM_SerializeDescriptor_Proto2 1.30GB/s ± 5% 1.32GB/s ± 5% ~ (p=0.433 n=17+17) BM_SerializeDescriptor_Upb 756MB/s ± 5% 745MB/s ± 6% ~ (p=0.102 n=19+16) ``` PiperOrigin-RevId: 520144430
2 years ago
return upb_Decoder_Decode(&decoder, buf, msg, l, arena);
}
#undef OP_FIXPCK_LG2
#undef OP_VARPCK_LG2