|
|
|
|
/*
|
|
|
|
|
* Copyright (c) 2009-2022, Google LLC
|
|
|
|
|
* All rights reserved.
|
|
|
|
|
*
|
|
|
|
|
* Redistribution and use in source and binary forms, with or without
|
|
|
|
|
* modification, are permitted provided that the following conditions are met:
|
|
|
|
|
* * Redistributions of source code must retain the above copyright
|
|
|
|
|
* notice, this list of conditions and the following disclaimer.
|
|
|
|
|
* * Redistributions in binary form must reproduce the above copyright
|
|
|
|
|
* notice, this list of conditions and the following disclaimer in the
|
|
|
|
|
* documentation and/or other materials provided with the distribution.
|
|
|
|
|
* * Neither the name of Google LLC nor the
|
|
|
|
|
* names of its contributors may be used to endorse or promote products
|
|
|
|
|
* derived from this software without specific prior written permission.
|
|
|
|
|
*
|
|
|
|
|
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
|
|
|
|
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
|
|
|
|
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
|
|
|
|
|
* ARE DISCLAIMED. IN NO EVENT SHALL Google LLC BE LIABLE FOR ANY DIRECT,
|
|
|
|
|
* INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
|
|
|
|
|
* (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
|
|
|
|
|
* LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
|
|
|
|
|
* ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
|
|
|
|
|
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
|
|
|
|
|
* SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
#ifndef UPB_MESSAGE_ACCESSORS_H_
|
|
|
|
|
#define UPB_MESSAGE_ACCESSORS_H_
|
|
|
|
|
|
|
|
|
|
#include "upb/collections/array.h"
|
Refactored message accessors to share a common set of functions instead of duplicating logic.
Prior to this CL, there were several different code paths for reading/writing message data. Generated code, MiniTable accessors, and reflection all performed direct manipulation of the bits and bytes in a message, but they all had distinct implementations that did not share much of any code. This divergence meant that they could easily have different behavior, bugs could creep into one but not another, and we would need three different sets of tests to get full test coverage. This also made it very difficult to change the internal representation in any way, since it would require updating many places in the code.
With this CL, the three different APIs for accessing message data now all share a common set of functions. The common functions all take a `upb_MiniTableField` as the canonical description of a field's type and layout. The lowest-level functions are very branchy, as they must test for every possible variation in the field type (field vs oneof, hasbit vs no-hasbit, different field sizes, whether a nonzero default value exists, extension vs. regular field), however these functions are declared inline and designed to be very optimizable when values are known at compile time.
In generated accessors, for example, we can declare constant `upb_MiniTableField` instances so that all values can constant-propagate, and we can get fully specialized code even though we are calling a generic function. On the other hand, when we use the generic functions from reflection, we get runtime branches since values are not known at compile time. But even the function is written to still be as efficient as possible even when used from reflection. For example, we use memcpy() calls with constant length so that the compiler can optimize these into inline loads/stores without having to make an out-of-line call to memcpy().
In this way, this CL should be a benefit to both correctness and performance. It will also make it easier to change the message representation, for example to optimize the encoder by giving hasbits to all fields.
Note that we have not completely consolidated all access in this CL:
1. Some functions outside of get/set such as clear and hazzers are not yet unified.
2. The encoder and decoder still touch the message without going through the common functions. The encoder and decoder require a bit more specialized code to get good performance when reading/writing fields en masse.
PiperOrigin-RevId: 490016095
2 years ago
|
|
|
#include "upb/message/extension_internal.h"
|
|
|
|
|
#include "upb/message/internal.h"
|
|
|
|
|
#include "upb/mini_table/common.h"
|
|
|
|
|
#include "upb/mini_table/enum_internal.h"
|
|
|
|
|
#include "upb/mini_table/field_internal.h"
|
|
|
|
|
|
|
|
|
|
// Must be last.
|
|
|
|
|
#include "upb/port/def.inc"
|
|
|
|
|
|
|
|
|
|
#ifdef __cplusplus
|
|
|
|
|
extern "C" {
|
|
|
|
|
#endif
|
|
|
|
|
|
|
|
|
|
UPB_INLINE bool _upb_MiniTableField_InOneOf(const upb_MiniTableField* field) {
|
|
|
|
|
return field->presence < 0;
|
|
|
|
|
}
|
|
|
|
|
|
Refactored message accessors to share a common set of functions instead of duplicating logic.
Prior to this CL, there were several different code paths for reading/writing message data. Generated code, MiniTable accessors, and reflection all performed direct manipulation of the bits and bytes in a message, but they all had distinct implementations that did not share much of any code. This divergence meant that they could easily have different behavior, bugs could creep into one but not another, and we would need three different sets of tests to get full test coverage. This also made it very difficult to change the internal representation in any way, since it would require updating many places in the code.
With this CL, the three different APIs for accessing message data now all share a common set of functions. The common functions all take a `upb_MiniTableField` as the canonical description of a field's type and layout. The lowest-level functions are very branchy, as they must test for every possible variation in the field type (field vs oneof, hasbit vs no-hasbit, different field sizes, whether a nonzero default value exists, extension vs. regular field), however these functions are declared inline and designed to be very optimizable when values are known at compile time.
In generated accessors, for example, we can declare constant `upb_MiniTableField` instances so that all values can constant-propagate, and we can get fully specialized code even though we are calling a generic function. On the other hand, when we use the generic functions from reflection, we get runtime branches since values are not known at compile time. But even the function is written to still be as efficient as possible even when used from reflection. For example, we use memcpy() calls with constant length so that the compiler can optimize these into inline loads/stores without having to make an out-of-line call to memcpy().
In this way, this CL should be a benefit to both correctness and performance. It will also make it easier to change the message representation, for example to optimize the encoder by giving hasbits to all fields.
Note that we have not completely consolidated all access in this CL:
1. Some functions outside of get/set such as clear and hazzers are not yet unified.
2. The encoder and decoder still touch the message without going through the common functions. The encoder and decoder require a bit more specialized code to get good performance when reading/writing fields en masse.
PiperOrigin-RevId: 490016095
2 years ago
|
|
|
UPB_INLINE void* _upb_MiniTableField_GetPtr(upb_Message* msg,
|
|
|
|
|
const upb_MiniTableField* field) {
|
|
|
|
|
return (char*)msg + field->offset;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
UPB_INLINE const void* _upb_MiniTableField_GetConstPtr(
|
|
|
|
|
const upb_Message* msg, const upb_MiniTableField* field) {
|
|
|
|
|
return (char*)msg + field->offset;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
UPB_INLINE void _upb_MiniTable_SetPresence(upb_Message* msg,
|
|
|
|
|
const upb_MiniTableField* field) {
|
|
|
|
|
if (field->presence > 0) {
|
|
|
|
|
_upb_sethas_field(msg, field);
|
|
|
|
|
} else if (_upb_MiniTableField_InOneOf(field)) {
|
|
|
|
|
*_upb_oneofcase_field(msg, field) = field->number;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
Refactored message accessors to share a common set of functions instead of duplicating logic.
Prior to this CL, there were several different code paths for reading/writing message data. Generated code, MiniTable accessors, and reflection all performed direct manipulation of the bits and bytes in a message, but they all had distinct implementations that did not share much of any code. This divergence meant that they could easily have different behavior, bugs could creep into one but not another, and we would need three different sets of tests to get full test coverage. This also made it very difficult to change the internal representation in any way, since it would require updating many places in the code.
With this CL, the three different APIs for accessing message data now all share a common set of functions. The common functions all take a `upb_MiniTableField` as the canonical description of a field's type and layout. The lowest-level functions are very branchy, as they must test for every possible variation in the field type (field vs oneof, hasbit vs no-hasbit, different field sizes, whether a nonzero default value exists, extension vs. regular field), however these functions are declared inline and designed to be very optimizable when values are known at compile time.
In generated accessors, for example, we can declare constant `upb_MiniTableField` instances so that all values can constant-propagate, and we can get fully specialized code even though we are calling a generic function. On the other hand, when we use the generic functions from reflection, we get runtime branches since values are not known at compile time. But even the function is written to still be as efficient as possible even when used from reflection. For example, we use memcpy() calls with constant length so that the compiler can optimize these into inline loads/stores without having to make an out-of-line call to memcpy().
In this way, this CL should be a benefit to both correctness and performance. It will also make it easier to change the message representation, for example to optimize the encoder by giving hasbits to all fields.
Note that we have not completely consolidated all access in this CL:
1. Some functions outside of get/set such as clear and hazzers are not yet unified.
2. The encoder and decoder still touch the message without going through the common functions. The encoder and decoder require a bit more specialized code to get good performance when reading/writing fields en masse.
PiperOrigin-RevId: 490016095
2 years ago
|
|
|
UPB_INLINE bool upb_MiniTable_HasField(const upb_Message* msg,
|
|
|
|
|
const upb_MiniTableField* field);
|
|
|
|
|
|
|
|
|
|
UPB_INLINE bool _upb_MiniTable_ValueIsNonZero(const void* default_val,
|
|
|
|
|
const upb_MiniTableField* field) {
|
Refactored message accessors to share a common set of functions instead of duplicating logic.
Prior to this CL, there were several different code paths for reading/writing message data. Generated code, MiniTable accessors, and reflection all performed direct manipulation of the bits and bytes in a message, but they all had distinct implementations that did not share much of any code. This divergence meant that they could easily have different behavior, bugs could creep into one but not another, and we would need three different sets of tests to get full test coverage. This also made it very difficult to change the internal representation in any way, since it would require updating many places in the code.
With this CL, the three different APIs for accessing message data now all share a common set of functions. The common functions all take a `upb_MiniTableField` as the canonical description of a field's type and layout. The lowest-level functions are very branchy, as they must test for every possible variation in the field type (field vs oneof, hasbit vs no-hasbit, different field sizes, whether a nonzero default value exists, extension vs. regular field), however these functions are declared inline and designed to be very optimizable when values are known at compile time.
In generated accessors, for example, we can declare constant `upb_MiniTableField` instances so that all values can constant-propagate, and we can get fully specialized code even though we are calling a generic function. On the other hand, when we use the generic functions from reflection, we get runtime branches since values are not known at compile time. But even the function is written to still be as efficient as possible even when used from reflection. For example, we use memcpy() calls with constant length so that the compiler can optimize these into inline loads/stores without having to make an out-of-line call to memcpy().
In this way, this CL should be a benefit to both correctness and performance. It will also make it easier to change the message representation, for example to optimize the encoder by giving hasbits to all fields.
Note that we have not completely consolidated all access in this CL:
1. Some functions outside of get/set such as clear and hazzers are not yet unified.
2. The encoder and decoder still touch the message without going through the common functions. The encoder and decoder require a bit more specialized code to get good performance when reading/writing fields en masse.
PiperOrigin-RevId: 490016095
2 years ago
|
|
|
char zero[16] = {0};
|
|
|
|
|
switch (_upb_MiniTableField_GetRep(field)) {
|
|
|
|
|
case kUpb_FieldRep_1Byte:
|
|
|
|
|
return memcmp(&zero, default_val, 1) != 0;
|
|
|
|
|
case kUpb_FieldRep_4Byte:
|
|
|
|
|
return memcmp(&zero, default_val, 4) != 0;
|
|
|
|
|
case kUpb_FieldRep_8Byte:
|
|
|
|
|
return memcmp(&zero, default_val, 8) != 0;
|
|
|
|
|
case kUpb_FieldRep_StringView: {
|
|
|
|
|
const upb_StringView* sv = (const upb_StringView*)default_val;
|
|
|
|
|
return sv->size != 0;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
UPB_UNREACHABLE();
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
UPB_INLINE void _upb_MiniTable_CopyFieldData(void* to, const void* from,
|
|
|
|
|
const upb_MiniTableField* field) {
|
|
|
|
|
switch (_upb_MiniTableField_GetRep(field)) {
|
|
|
|
|
case kUpb_FieldRep_1Byte:
|
|
|
|
|
memcpy(to, from, 1);
|
|
|
|
|
return;
|
|
|
|
|
case kUpb_FieldRep_4Byte:
|
|
|
|
|
memcpy(to, from, 4);
|
|
|
|
|
return;
|
|
|
|
|
case kUpb_FieldRep_8Byte:
|
|
|
|
|
memcpy(to, from, 8);
|
|
|
|
|
return;
|
|
|
|
|
case kUpb_FieldRep_StringView: {
|
|
|
|
|
memcpy(to, from, sizeof(upb_StringView));
|
|
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
UPB_UNREACHABLE();
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Here we define universal getter/setter functions for message fields.
|
|
|
|
|
// These look very branchy and inefficient, but as long as the MiniTableField
|
|
|
|
|
// values are known at compile time, all the branches are optimized away and
|
|
|
|
|
// we are left with ideal code. This can happen either through through
|
|
|
|
|
// literals or UPB_ASSUME():
|
|
|
|
|
//
|
|
|
|
|
// // Via string literals.
|
|
|
|
|
// bool FooMessage_set_bool_field(const upb_Message* msg, bool val) {
|
|
|
|
|
// const upb_MiniTableField field = {1, 0, 0, /* etc... */};
|
|
|
|
|
// // All value in "field" are compile-time known.
|
|
|
|
|
// _upb_MiniTable_SetNonExtensionField(msg, &field, &value);
|
|
|
|
|
// }
|
|
|
|
|
//
|
|
|
|
|
// // Via UPB_ASSUME().
|
|
|
|
|
// UPB_INLINE void upb_MiniTable_SetBool(upb_Message* msg,
|
|
|
|
|
// const upb_MiniTableField* field,
|
|
|
|
|
// bool value) {
|
|
|
|
|
// UPB_ASSUME(field->descriptortype == kUpb_FieldType_Bool);
|
|
|
|
|
// UPB_ASSUME(!upb_IsRepeatedOrMap(field));
|
|
|
|
|
// UPB_ASSUME(_upb_MiniTableField_GetRep(field) == kUpb_FieldRep_1Byte);
|
|
|
|
|
// _upb_MiniTable_SetNonExtensionField(msg, field, &value);
|
|
|
|
|
// }
|
|
|
|
|
//
|
|
|
|
|
// As a result, we can use these universal getters/setters for *all* message
|
|
|
|
|
// accessors: generated code, MiniTable accessors, and reflection. The only
|
|
|
|
|
// exception is the binary encoder/decoder, which need to be a bit more clever
|
|
|
|
|
// about how the read/write the message data, for efficiency.
|
|
|
|
|
|
|
|
|
|
static UPB_FORCEINLINE void _upb_MiniTable_GetNonExtensionField(
|
|
|
|
|
const upb_Message* msg, const upb_MiniTableField* field,
|
|
|
|
|
const void* default_val, void* val) {
|
|
|
|
|
UPB_ASSUME(!upb_MiniTableField_IsExtension(field));
|
|
|
|
|
if ((_upb_MiniTableField_InOneOf(field) ||
|
|
|
|
|
_upb_MiniTable_ValueIsNonZero(default_val, field)) &&
|
Refactored message accessors to share a common set of functions instead of duplicating logic.
Prior to this CL, there were several different code paths for reading/writing message data. Generated code, MiniTable accessors, and reflection all performed direct manipulation of the bits and bytes in a message, but they all had distinct implementations that did not share much of any code. This divergence meant that they could easily have different behavior, bugs could creep into one but not another, and we would need three different sets of tests to get full test coverage. This also made it very difficult to change the internal representation in any way, since it would require updating many places in the code.
With this CL, the three different APIs for accessing message data now all share a common set of functions. The common functions all take a `upb_MiniTableField` as the canonical description of a field's type and layout. The lowest-level functions are very branchy, as they must test for every possible variation in the field type (field vs oneof, hasbit vs no-hasbit, different field sizes, whether a nonzero default value exists, extension vs. regular field), however these functions are declared inline and designed to be very optimizable when values are known at compile time.
In generated accessors, for example, we can declare constant `upb_MiniTableField` instances so that all values can constant-propagate, and we can get fully specialized code even though we are calling a generic function. On the other hand, when we use the generic functions from reflection, we get runtime branches since values are not known at compile time. But even the function is written to still be as efficient as possible even when used from reflection. For example, we use memcpy() calls with constant length so that the compiler can optimize these into inline loads/stores without having to make an out-of-line call to memcpy().
In this way, this CL should be a benefit to both correctness and performance. It will also make it easier to change the message representation, for example to optimize the encoder by giving hasbits to all fields.
Note that we have not completely consolidated all access in this CL:
1. Some functions outside of get/set such as clear and hazzers are not yet unified.
2. The encoder and decoder still touch the message without going through the common functions. The encoder and decoder require a bit more specialized code to get good performance when reading/writing fields en masse.
PiperOrigin-RevId: 490016095
2 years ago
|
|
|
!upb_MiniTable_HasField(msg, field)) {
|
|
|
|
|
_upb_MiniTable_CopyFieldData(val, default_val, field);
|
|
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
_upb_MiniTable_CopyFieldData(val, _upb_MiniTableField_GetConstPtr(msg, field),
|
|
|
|
|
field);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
UPB_INLINE void _upb_MiniTable_GetExtensionField(
|
|
|
|
|
const upb_Message* msg, const upb_MiniTableExtension* mt_ext,
|
|
|
|
|
const void* default_val, void* val) {
|
|
|
|
|
UPB_ASSUME(upb_MiniTableField_IsExtension(&mt_ext->field));
|
|
|
|
|
const upb_Message_Extension* ext = _upb_Message_Getext(msg, mt_ext);
|
|
|
|
|
if (ext) {
|
|
|
|
|
_upb_MiniTable_CopyFieldData(val, &ext->data, &mt_ext->field);
|
|
|
|
|
} else {
|
|
|
|
|
_upb_MiniTable_CopyFieldData(val, default_val, &mt_ext->field);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
UPB_INLINE void _upb_MiniTable_GetField(const upb_Message* msg,
|
|
|
|
|
const upb_MiniTableField* field,
|
|
|
|
|
const void* default_val, void* val) {
|
|
|
|
|
if (upb_MiniTableField_IsExtension(field)) {
|
|
|
|
|
_upb_MiniTable_GetExtensionField(msg, (upb_MiniTableExtension*)field,
|
|
|
|
|
default_val, val);
|
|
|
|
|
} else {
|
|
|
|
|
_upb_MiniTable_GetNonExtensionField(msg, field, default_val, val);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
UPB_INLINE void _upb_MiniTable_SetNonExtensionField(
|
|
|
|
|
upb_Message* msg, const upb_MiniTableField* field, const void* val) {
|
|
|
|
|
UPB_ASSUME(!upb_MiniTableField_IsExtension(field));
|
|
|
|
|
_upb_MiniTable_SetPresence(msg, field);
|
|
|
|
|
_upb_MiniTable_CopyFieldData(_upb_MiniTableField_GetPtr(msg, field), val,
|
|
|
|
|
field);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
UPB_INLINE bool _upb_MiniTable_SetExtensionField(
|
|
|
|
|
upb_Message* msg, const upb_MiniTableExtension* mt_ext, const void* val,
|
|
|
|
|
upb_Arena* a) {
|
|
|
|
|
upb_Message_Extension* ext =
|
|
|
|
|
_upb_Message_GetOrCreateExtension(msg, mt_ext, a);
|
|
|
|
|
if (!ext) return false;
|
|
|
|
|
_upb_MiniTable_CopyFieldData(&ext->data, val, &mt_ext->field);
|
|
|
|
|
return true;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
UPB_INLINE bool _upb_MiniTable_SetField(upb_Message* msg,
|
|
|
|
|
const upb_MiniTableField* field,
|
|
|
|
|
const void* val, upb_Arena* a) {
|
|
|
|
|
if (upb_MiniTableField_IsExtension(field)) {
|
|
|
|
|
return _upb_MiniTable_SetExtensionField(
|
|
|
|
|
msg, (const upb_MiniTableExtension*)field, val, a);
|
|
|
|
|
} else {
|
|
|
|
|
_upb_MiniTable_SetNonExtensionField(msg, field, val);
|
|
|
|
|
return true;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
UPB_INLINE bool _upb_MiniTable_HasExtensionField(
|
|
|
|
|
const upb_Message* msg, const upb_MiniTableExtension* ext) {
|
|
|
|
|
UPB_ASSERT(upb_MiniTableField_HasPresence(&ext->field));
|
|
|
|
|
return _upb_Message_Getext(msg, ext) != NULL;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
UPB_INLINE bool _upb_MiniTable_HasNonExtensionField(
|
|
|
|
|
const upb_Message* msg, const upb_MiniTableField* field) {
|
|
|
|
|
UPB_ASSERT(upb_MiniTableField_HasPresence(field));
|
|
|
|
|
UPB_ASSUME(!upb_MiniTableField_IsExtension(field));
|
|
|
|
|
if (_upb_MiniTableField_InOneOf(field)) {
|
|
|
|
|
return _upb_getoneofcase_field(msg, field) == field->number;
|
|
|
|
|
} else {
|
|
|
|
|
return _upb_hasbit_field(msg, field);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
UPB_INLINE bool _upb_MiniTable_HasField(const upb_Message* msg,
|
|
|
|
|
const upb_MiniTableField* field) {
|
|
|
|
|
if (upb_MiniTableField_IsExtension(field)) {
|
|
|
|
|
return _upb_MiniTable_HasExtensionField(
|
|
|
|
|
msg, (const upb_MiniTableExtension*)field);
|
|
|
|
|
} else {
|
|
|
|
|
return _upb_MiniTable_HasNonExtensionField(msg, field);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
UPB_INLINE void _upb_MiniTable_ClearExtensionField(
|
|
|
|
|
upb_Message* msg, const upb_MiniTableExtension* ext) {
|
|
|
|
|
_upb_Message_Clearext(msg, ext);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
UPB_INLINE void _upb_MiniTable_ClearNonExtensionField(
|
|
|
|
|
upb_Message* msg, const upb_MiniTableField* field) {
|
|
|
|
|
if (field->presence > 0) {
|
|
|
|
|
_upb_clearhas_field(msg, field);
|
|
|
|
|
} else if (_upb_MiniTableField_InOneOf(field)) {
|
|
|
|
|
uint32_t* oneof_case = _upb_oneofcase_field(msg, field);
|
|
|
|
|
if (*oneof_case != field->number) return;
|
|
|
|
|
*oneof_case = 0;
|
|
|
|
|
}
|
|
|
|
|
const char zeros[16] = {0};
|
|
|
|
|
_upb_MiniTable_CopyFieldData(_upb_MiniTableField_GetPtr(msg, field), zeros,
|
|
|
|
|
field);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
UPB_INLINE void _upb_MiniTable_ClearField(upb_Message* msg,
|
|
|
|
|
const upb_MiniTableField* field) {
|
|
|
|
|
if (upb_MiniTableField_IsExtension(field)) {
|
|
|
|
|
_upb_Message_Clearext(msg, (const upb_MiniTableExtension*)field);
|
|
|
|
|
} else {
|
|
|
|
|
_upb_MiniTable_ClearNonExtensionField(msg, field);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// EVERYTHING ABOVE THIS LINE IS INTERNAL - DO NOT USE /////////////////////////
|
|
|
|
|
|
|
|
|
|
UPB_API_INLINE void upb_MiniTable_ClearField(upb_Message* msg,
|
|
|
|
|
const upb_MiniTableField* field) {
|
|
|
|
|
_upb_MiniTable_ClearNonExtensionField(msg, field);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
UPB_API_INLINE bool upb_MiniTable_HasField(const upb_Message* msg,
|
|
|
|
|
const upb_MiniTableField* field) {
|
|
|
|
|
return _upb_MiniTable_HasNonExtensionField(msg, field);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
UPB_API_INLINE bool upb_MiniTable_GetBool(const upb_Message* msg,
|
|
|
|
|
const upb_MiniTableField* field,
|
|
|
|
|
bool default_val) {
|
Refactored message accessors to share a common set of functions instead of duplicating logic.
Prior to this CL, there were several different code paths for reading/writing message data. Generated code, MiniTable accessors, and reflection all performed direct manipulation of the bits and bytes in a message, but they all had distinct implementations that did not share much of any code. This divergence meant that they could easily have different behavior, bugs could creep into one but not another, and we would need three different sets of tests to get full test coverage. This also made it very difficult to change the internal representation in any way, since it would require updating many places in the code.
With this CL, the three different APIs for accessing message data now all share a common set of functions. The common functions all take a `upb_MiniTableField` as the canonical description of a field's type and layout. The lowest-level functions are very branchy, as they must test for every possible variation in the field type (field vs oneof, hasbit vs no-hasbit, different field sizes, whether a nonzero default value exists, extension vs. regular field), however these functions are declared inline and designed to be very optimizable when values are known at compile time.
In generated accessors, for example, we can declare constant `upb_MiniTableField` instances so that all values can constant-propagate, and we can get fully specialized code even though we are calling a generic function. On the other hand, when we use the generic functions from reflection, we get runtime branches since values are not known at compile time. But even the function is written to still be as efficient as possible even when used from reflection. For example, we use memcpy() calls with constant length so that the compiler can optimize these into inline loads/stores without having to make an out-of-line call to memcpy().
In this way, this CL should be a benefit to both correctness and performance. It will also make it easier to change the message representation, for example to optimize the encoder by giving hasbits to all fields.
Note that we have not completely consolidated all access in this CL:
1. Some functions outside of get/set such as clear and hazzers are not yet unified.
2. The encoder and decoder still touch the message without going through the common functions. The encoder and decoder require a bit more specialized code to get good performance when reading/writing fields en masse.
PiperOrigin-RevId: 490016095
2 years ago
|
|
|
UPB_ASSUME(field->descriptortype == kUpb_FieldType_Bool);
|
|
|
|
|
UPB_ASSUME(!upb_IsRepeatedOrMap(field));
|
|
|
|
|
UPB_ASSUME(_upb_MiniTableField_GetRep(field) == kUpb_FieldRep_1Byte);
|
|
|
|
|
bool ret;
|
|
|
|
|
_upb_MiniTable_GetNonExtensionField(msg, field, &default_val, &ret);
|
|
|
|
|
return ret;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
UPB_API_INLINE void upb_MiniTable_SetBool(upb_Message* msg,
|
|
|
|
|
const upb_MiniTableField* field,
|
|
|
|
|
bool value) {
|
Refactored message accessors to share a common set of functions instead of duplicating logic.
Prior to this CL, there were several different code paths for reading/writing message data. Generated code, MiniTable accessors, and reflection all performed direct manipulation of the bits and bytes in a message, but they all had distinct implementations that did not share much of any code. This divergence meant that they could easily have different behavior, bugs could creep into one but not another, and we would need three different sets of tests to get full test coverage. This also made it very difficult to change the internal representation in any way, since it would require updating many places in the code.
With this CL, the three different APIs for accessing message data now all share a common set of functions. The common functions all take a `upb_MiniTableField` as the canonical description of a field's type and layout. The lowest-level functions are very branchy, as they must test for every possible variation in the field type (field vs oneof, hasbit vs no-hasbit, different field sizes, whether a nonzero default value exists, extension vs. regular field), however these functions are declared inline and designed to be very optimizable when values are known at compile time.
In generated accessors, for example, we can declare constant `upb_MiniTableField` instances so that all values can constant-propagate, and we can get fully specialized code even though we are calling a generic function. On the other hand, when we use the generic functions from reflection, we get runtime branches since values are not known at compile time. But even the function is written to still be as efficient as possible even when used from reflection. For example, we use memcpy() calls with constant length so that the compiler can optimize these into inline loads/stores without having to make an out-of-line call to memcpy().
In this way, this CL should be a benefit to both correctness and performance. It will also make it easier to change the message representation, for example to optimize the encoder by giving hasbits to all fields.
Note that we have not completely consolidated all access in this CL:
1. Some functions outside of get/set such as clear and hazzers are not yet unified.
2. The encoder and decoder still touch the message without going through the common functions. The encoder and decoder require a bit more specialized code to get good performance when reading/writing fields en masse.
PiperOrigin-RevId: 490016095
2 years ago
|
|
|
UPB_ASSUME(field->descriptortype == kUpb_FieldType_Bool);
|
|
|
|
|
UPB_ASSUME(!upb_IsRepeatedOrMap(field));
|
|
|
|
|
UPB_ASSUME(_upb_MiniTableField_GetRep(field) == kUpb_FieldRep_1Byte);
|
|
|
|
|
_upb_MiniTable_SetNonExtensionField(msg, field, &value);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
UPB_API_INLINE int32_t upb_MiniTable_GetInt32(const upb_Message* msg,
|
|
|
|
|
const upb_MiniTableField* field,
|
|
|
|
|
int32_t default_val) {
|
Refactored message accessors to share a common set of functions instead of duplicating logic.
Prior to this CL, there were several different code paths for reading/writing message data. Generated code, MiniTable accessors, and reflection all performed direct manipulation of the bits and bytes in a message, but they all had distinct implementations that did not share much of any code. This divergence meant that they could easily have different behavior, bugs could creep into one but not another, and we would need three different sets of tests to get full test coverage. This also made it very difficult to change the internal representation in any way, since it would require updating many places in the code.
With this CL, the three different APIs for accessing message data now all share a common set of functions. The common functions all take a `upb_MiniTableField` as the canonical description of a field's type and layout. The lowest-level functions are very branchy, as they must test for every possible variation in the field type (field vs oneof, hasbit vs no-hasbit, different field sizes, whether a nonzero default value exists, extension vs. regular field), however these functions are declared inline and designed to be very optimizable when values are known at compile time.
In generated accessors, for example, we can declare constant `upb_MiniTableField` instances so that all values can constant-propagate, and we can get fully specialized code even though we are calling a generic function. On the other hand, when we use the generic functions from reflection, we get runtime branches since values are not known at compile time. But even the function is written to still be as efficient as possible even when used from reflection. For example, we use memcpy() calls with constant length so that the compiler can optimize these into inline loads/stores without having to make an out-of-line call to memcpy().
In this way, this CL should be a benefit to both correctness and performance. It will also make it easier to change the message representation, for example to optimize the encoder by giving hasbits to all fields.
Note that we have not completely consolidated all access in this CL:
1. Some functions outside of get/set such as clear and hazzers are not yet unified.
2. The encoder and decoder still touch the message without going through the common functions. The encoder and decoder require a bit more specialized code to get good performance when reading/writing fields en masse.
PiperOrigin-RevId: 490016095
2 years ago
|
|
|
UPB_ASSUME(field->descriptortype == kUpb_FieldType_Int32 ||
|
|
|
|
|
field->descriptortype == kUpb_FieldType_SInt32 ||
|
|
|
|
|
field->descriptortype == kUpb_FieldType_SFixed32 ||
|
|
|
|
|
field->descriptortype == kUpb_FieldType_Enum);
|
Refactored message accessors to share a common set of functions instead of duplicating logic.
Prior to this CL, there were several different code paths for reading/writing message data. Generated code, MiniTable accessors, and reflection all performed direct manipulation of the bits and bytes in a message, but they all had distinct implementations that did not share much of any code. This divergence meant that they could easily have different behavior, bugs could creep into one but not another, and we would need three different sets of tests to get full test coverage. This also made it very difficult to change the internal representation in any way, since it would require updating many places in the code.
With this CL, the three different APIs for accessing message data now all share a common set of functions. The common functions all take a `upb_MiniTableField` as the canonical description of a field's type and layout. The lowest-level functions are very branchy, as they must test for every possible variation in the field type (field vs oneof, hasbit vs no-hasbit, different field sizes, whether a nonzero default value exists, extension vs. regular field), however these functions are declared inline and designed to be very optimizable when values are known at compile time.
In generated accessors, for example, we can declare constant `upb_MiniTableField` instances so that all values can constant-propagate, and we can get fully specialized code even though we are calling a generic function. On the other hand, when we use the generic functions from reflection, we get runtime branches since values are not known at compile time. But even the function is written to still be as efficient as possible even when used from reflection. For example, we use memcpy() calls with constant length so that the compiler can optimize these into inline loads/stores without having to make an out-of-line call to memcpy().
In this way, this CL should be a benefit to both correctness and performance. It will also make it easier to change the message representation, for example to optimize the encoder by giving hasbits to all fields.
Note that we have not completely consolidated all access in this CL:
1. Some functions outside of get/set such as clear and hazzers are not yet unified.
2. The encoder and decoder still touch the message without going through the common functions. The encoder and decoder require a bit more specialized code to get good performance when reading/writing fields en masse.
PiperOrigin-RevId: 490016095
2 years ago
|
|
|
UPB_ASSUME(!upb_IsRepeatedOrMap(field));
|
|
|
|
|
UPB_ASSUME(_upb_MiniTableField_GetRep(field) == kUpb_FieldRep_4Byte);
|
|
|
|
|
int32_t ret;
|
|
|
|
|
_upb_MiniTable_GetNonExtensionField(msg, field, &default_val, &ret);
|
|
|
|
|
return ret;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
UPB_API_INLINE void upb_MiniTable_SetInt32(upb_Message* msg,
|
|
|
|
|
const upb_MiniTableField* field,
|
|
|
|
|
int32_t value) {
|
Refactored message accessors to share a common set of functions instead of duplicating logic.
Prior to this CL, there were several different code paths for reading/writing message data. Generated code, MiniTable accessors, and reflection all performed direct manipulation of the bits and bytes in a message, but they all had distinct implementations that did not share much of any code. This divergence meant that they could easily have different behavior, bugs could creep into one but not another, and we would need three different sets of tests to get full test coverage. This also made it very difficult to change the internal representation in any way, since it would require updating many places in the code.
With this CL, the three different APIs for accessing message data now all share a common set of functions. The common functions all take a `upb_MiniTableField` as the canonical description of a field's type and layout. The lowest-level functions are very branchy, as they must test for every possible variation in the field type (field vs oneof, hasbit vs no-hasbit, different field sizes, whether a nonzero default value exists, extension vs. regular field), however these functions are declared inline and designed to be very optimizable when values are known at compile time.
In generated accessors, for example, we can declare constant `upb_MiniTableField` instances so that all values can constant-propagate, and we can get fully specialized code even though we are calling a generic function. On the other hand, when we use the generic functions from reflection, we get runtime branches since values are not known at compile time. But even the function is written to still be as efficient as possible even when used from reflection. For example, we use memcpy() calls with constant length so that the compiler can optimize these into inline loads/stores without having to make an out-of-line call to memcpy().
In this way, this CL should be a benefit to both correctness and performance. It will also make it easier to change the message representation, for example to optimize the encoder by giving hasbits to all fields.
Note that we have not completely consolidated all access in this CL:
1. Some functions outside of get/set such as clear and hazzers are not yet unified.
2. The encoder and decoder still touch the message without going through the common functions. The encoder and decoder require a bit more specialized code to get good performance when reading/writing fields en masse.
PiperOrigin-RevId: 490016095
2 years ago
|
|
|
UPB_ASSUME(field->descriptortype == kUpb_FieldType_Int32 ||
|
|
|
|
|
field->descriptortype == kUpb_FieldType_SInt32 ||
|
|
|
|
|
field->descriptortype == kUpb_FieldType_SFixed32);
|
Refactored message accessors to share a common set of functions instead of duplicating logic.
Prior to this CL, there were several different code paths for reading/writing message data. Generated code, MiniTable accessors, and reflection all performed direct manipulation of the bits and bytes in a message, but they all had distinct implementations that did not share much of any code. This divergence meant that they could easily have different behavior, bugs could creep into one but not another, and we would need three different sets of tests to get full test coverage. This also made it very difficult to change the internal representation in any way, since it would require updating many places in the code.
With this CL, the three different APIs for accessing message data now all share a common set of functions. The common functions all take a `upb_MiniTableField` as the canonical description of a field's type and layout. The lowest-level functions are very branchy, as they must test for every possible variation in the field type (field vs oneof, hasbit vs no-hasbit, different field sizes, whether a nonzero default value exists, extension vs. regular field), however these functions are declared inline and designed to be very optimizable when values are known at compile time.
In generated accessors, for example, we can declare constant `upb_MiniTableField` instances so that all values can constant-propagate, and we can get fully specialized code even though we are calling a generic function. On the other hand, when we use the generic functions from reflection, we get runtime branches since values are not known at compile time. But even the function is written to still be as efficient as possible even when used from reflection. For example, we use memcpy() calls with constant length so that the compiler can optimize these into inline loads/stores without having to make an out-of-line call to memcpy().
In this way, this CL should be a benefit to both correctness and performance. It will also make it easier to change the message representation, for example to optimize the encoder by giving hasbits to all fields.
Note that we have not completely consolidated all access in this CL:
1. Some functions outside of get/set such as clear and hazzers are not yet unified.
2. The encoder and decoder still touch the message without going through the common functions. The encoder and decoder require a bit more specialized code to get good performance when reading/writing fields en masse.
PiperOrigin-RevId: 490016095
2 years ago
|
|
|
UPB_ASSUME(!upb_IsRepeatedOrMap(field));
|
|
|
|
|
UPB_ASSUME(_upb_MiniTableField_GetRep(field) == kUpb_FieldRep_4Byte);
|
|
|
|
|
_upb_MiniTable_SetNonExtensionField(msg, field, &value);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
UPB_API_INLINE uint32_t upb_MiniTable_GetUInt32(const upb_Message* msg,
|
|
|
|
|
const upb_MiniTableField* field,
|
|
|
|
|
uint32_t default_val) {
|
Refactored message accessors to share a common set of functions instead of duplicating logic.
Prior to this CL, there were several different code paths for reading/writing message data. Generated code, MiniTable accessors, and reflection all performed direct manipulation of the bits and bytes in a message, but they all had distinct implementations that did not share much of any code. This divergence meant that they could easily have different behavior, bugs could creep into one but not another, and we would need three different sets of tests to get full test coverage. This also made it very difficult to change the internal representation in any way, since it would require updating many places in the code.
With this CL, the three different APIs for accessing message data now all share a common set of functions. The common functions all take a `upb_MiniTableField` as the canonical description of a field's type and layout. The lowest-level functions are very branchy, as they must test for every possible variation in the field type (field vs oneof, hasbit vs no-hasbit, different field sizes, whether a nonzero default value exists, extension vs. regular field), however these functions are declared inline and designed to be very optimizable when values are known at compile time.
In generated accessors, for example, we can declare constant `upb_MiniTableField` instances so that all values can constant-propagate, and we can get fully specialized code even though we are calling a generic function. On the other hand, when we use the generic functions from reflection, we get runtime branches since values are not known at compile time. But even the function is written to still be as efficient as possible even when used from reflection. For example, we use memcpy() calls with constant length so that the compiler can optimize these into inline loads/stores without having to make an out-of-line call to memcpy().
In this way, this CL should be a benefit to both correctness and performance. It will also make it easier to change the message representation, for example to optimize the encoder by giving hasbits to all fields.
Note that we have not completely consolidated all access in this CL:
1. Some functions outside of get/set such as clear and hazzers are not yet unified.
2. The encoder and decoder still touch the message without going through the common functions. The encoder and decoder require a bit more specialized code to get good performance when reading/writing fields en masse.
PiperOrigin-RevId: 490016095
2 years ago
|
|
|
UPB_ASSUME(field->descriptortype == kUpb_FieldType_UInt32 ||
|
|
|
|
|
field->descriptortype == kUpb_FieldType_Fixed32);
|
Refactored message accessors to share a common set of functions instead of duplicating logic.
Prior to this CL, there were several different code paths for reading/writing message data. Generated code, MiniTable accessors, and reflection all performed direct manipulation of the bits and bytes in a message, but they all had distinct implementations that did not share much of any code. This divergence meant that they could easily have different behavior, bugs could creep into one but not another, and we would need three different sets of tests to get full test coverage. This also made it very difficult to change the internal representation in any way, since it would require updating many places in the code.
With this CL, the three different APIs for accessing message data now all share a common set of functions. The common functions all take a `upb_MiniTableField` as the canonical description of a field's type and layout. The lowest-level functions are very branchy, as they must test for every possible variation in the field type (field vs oneof, hasbit vs no-hasbit, different field sizes, whether a nonzero default value exists, extension vs. regular field), however these functions are declared inline and designed to be very optimizable when values are known at compile time.
In generated accessors, for example, we can declare constant `upb_MiniTableField` instances so that all values can constant-propagate, and we can get fully specialized code even though we are calling a generic function. On the other hand, when we use the generic functions from reflection, we get runtime branches since values are not known at compile time. But even the function is written to still be as efficient as possible even when used from reflection. For example, we use memcpy() calls with constant length so that the compiler can optimize these into inline loads/stores without having to make an out-of-line call to memcpy().
In this way, this CL should be a benefit to both correctness and performance. It will also make it easier to change the message representation, for example to optimize the encoder by giving hasbits to all fields.
Note that we have not completely consolidated all access in this CL:
1. Some functions outside of get/set such as clear and hazzers are not yet unified.
2. The encoder and decoder still touch the message without going through the common functions. The encoder and decoder require a bit more specialized code to get good performance when reading/writing fields en masse.
PiperOrigin-RevId: 490016095
2 years ago
|
|
|
UPB_ASSUME(!upb_IsRepeatedOrMap(field));
|
|
|
|
|
UPB_ASSUME(_upb_MiniTableField_GetRep(field) == kUpb_FieldRep_4Byte);
|
|
|
|
|
uint32_t ret;
|
|
|
|
|
_upb_MiniTable_GetNonExtensionField(msg, field, &default_val, &ret);
|
|
|
|
|
return ret;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
UPB_API_INLINE void upb_MiniTable_SetUInt32(upb_Message* msg,
|
|
|
|
|
const upb_MiniTableField* field,
|
|
|
|
|
uint32_t value) {
|
Refactored message accessors to share a common set of functions instead of duplicating logic.
Prior to this CL, there were several different code paths for reading/writing message data. Generated code, MiniTable accessors, and reflection all performed direct manipulation of the bits and bytes in a message, but they all had distinct implementations that did not share much of any code. This divergence meant that they could easily have different behavior, bugs could creep into one but not another, and we would need three different sets of tests to get full test coverage. This also made it very difficult to change the internal representation in any way, since it would require updating many places in the code.
With this CL, the three different APIs for accessing message data now all share a common set of functions. The common functions all take a `upb_MiniTableField` as the canonical description of a field's type and layout. The lowest-level functions are very branchy, as they must test for every possible variation in the field type (field vs oneof, hasbit vs no-hasbit, different field sizes, whether a nonzero default value exists, extension vs. regular field), however these functions are declared inline and designed to be very optimizable when values are known at compile time.
In generated accessors, for example, we can declare constant `upb_MiniTableField` instances so that all values can constant-propagate, and we can get fully specialized code even though we are calling a generic function. On the other hand, when we use the generic functions from reflection, we get runtime branches since values are not known at compile time. But even the function is written to still be as efficient as possible even when used from reflection. For example, we use memcpy() calls with constant length so that the compiler can optimize these into inline loads/stores without having to make an out-of-line call to memcpy().
In this way, this CL should be a benefit to both correctness and performance. It will also make it easier to change the message representation, for example to optimize the encoder by giving hasbits to all fields.
Note that we have not completely consolidated all access in this CL:
1. Some functions outside of get/set such as clear and hazzers are not yet unified.
2. The encoder and decoder still touch the message without going through the common functions. The encoder and decoder require a bit more specialized code to get good performance when reading/writing fields en masse.
PiperOrigin-RevId: 490016095
2 years ago
|
|
|
UPB_ASSUME(field->descriptortype == kUpb_FieldType_UInt32 ||
|
|
|
|
|
field->descriptortype == kUpb_FieldType_Fixed32);
|
Refactored message accessors to share a common set of functions instead of duplicating logic.
Prior to this CL, there were several different code paths for reading/writing message data. Generated code, MiniTable accessors, and reflection all performed direct manipulation of the bits and bytes in a message, but they all had distinct implementations that did not share much of any code. This divergence meant that they could easily have different behavior, bugs could creep into one but not another, and we would need three different sets of tests to get full test coverage. This also made it very difficult to change the internal representation in any way, since it would require updating many places in the code.
With this CL, the three different APIs for accessing message data now all share a common set of functions. The common functions all take a `upb_MiniTableField` as the canonical description of a field's type and layout. The lowest-level functions are very branchy, as they must test for every possible variation in the field type (field vs oneof, hasbit vs no-hasbit, different field sizes, whether a nonzero default value exists, extension vs. regular field), however these functions are declared inline and designed to be very optimizable when values are known at compile time.
In generated accessors, for example, we can declare constant `upb_MiniTableField` instances so that all values can constant-propagate, and we can get fully specialized code even though we are calling a generic function. On the other hand, when we use the generic functions from reflection, we get runtime branches since values are not known at compile time. But even the function is written to still be as efficient as possible even when used from reflection. For example, we use memcpy() calls with constant length so that the compiler can optimize these into inline loads/stores without having to make an out-of-line call to memcpy().
In this way, this CL should be a benefit to both correctness and performance. It will also make it easier to change the message representation, for example to optimize the encoder by giving hasbits to all fields.
Note that we have not completely consolidated all access in this CL:
1. Some functions outside of get/set such as clear and hazzers are not yet unified.
2. The encoder and decoder still touch the message without going through the common functions. The encoder and decoder require a bit more specialized code to get good performance when reading/writing fields en masse.
PiperOrigin-RevId: 490016095
2 years ago
|
|
|
UPB_ASSUME(!upb_IsRepeatedOrMap(field));
|
|
|
|
|
UPB_ASSUME(_upb_MiniTableField_GetRep(field) == kUpb_FieldRep_4Byte);
|
|
|
|
|
_upb_MiniTable_SetNonExtensionField(msg, field, &value);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
UPB_API_INLINE void upb_MiniTable_SetEnumProto2(
|
|
|
|
|
upb_Message* msg, const upb_MiniTable* msg_mini_table,
|
|
|
|
|
const upb_MiniTableField* field, int32_t value) {
|
|
|
|
|
UPB_ASSERT(field->descriptortype == kUpb_FieldType_Enum);
|
Refactored message accessors to share a common set of functions instead of duplicating logic.
Prior to this CL, there were several different code paths for reading/writing message data. Generated code, MiniTable accessors, and reflection all performed direct manipulation of the bits and bytes in a message, but they all had distinct implementations that did not share much of any code. This divergence meant that they could easily have different behavior, bugs could creep into one but not another, and we would need three different sets of tests to get full test coverage. This also made it very difficult to change the internal representation in any way, since it would require updating many places in the code.
With this CL, the three different APIs for accessing message data now all share a common set of functions. The common functions all take a `upb_MiniTableField` as the canonical description of a field's type and layout. The lowest-level functions are very branchy, as they must test for every possible variation in the field type (field vs oneof, hasbit vs no-hasbit, different field sizes, whether a nonzero default value exists, extension vs. regular field), however these functions are declared inline and designed to be very optimizable when values are known at compile time.
In generated accessors, for example, we can declare constant `upb_MiniTableField` instances so that all values can constant-propagate, and we can get fully specialized code even though we are calling a generic function. On the other hand, when we use the generic functions from reflection, we get runtime branches since values are not known at compile time. But even the function is written to still be as efficient as possible even when used from reflection. For example, we use memcpy() calls with constant length so that the compiler can optimize these into inline loads/stores without having to make an out-of-line call to memcpy().
In this way, this CL should be a benefit to both correctness and performance. It will also make it easier to change the message representation, for example to optimize the encoder by giving hasbits to all fields.
Note that we have not completely consolidated all access in this CL:
1. Some functions outside of get/set such as clear and hazzers are not yet unified.
2. The encoder and decoder still touch the message without going through the common functions. The encoder and decoder require a bit more specialized code to get good performance when reading/writing fields en masse.
PiperOrigin-RevId: 490016095
2 years ago
|
|
|
UPB_ASSUME(!upb_IsRepeatedOrMap(field));
|
|
|
|
|
UPB_ASSUME(_upb_MiniTableField_GetRep(field) == kUpb_FieldRep_4Byte);
|
|
|
|
|
UPB_ASSERT(upb_MiniTableEnum_CheckValue(
|
|
|
|
|
upb_MiniTable_GetSubEnumTable(msg_mini_table, field), value));
|
Refactored message accessors to share a common set of functions instead of duplicating logic.
Prior to this CL, there were several different code paths for reading/writing message data. Generated code, MiniTable accessors, and reflection all performed direct manipulation of the bits and bytes in a message, but they all had distinct implementations that did not share much of any code. This divergence meant that they could easily have different behavior, bugs could creep into one but not another, and we would need three different sets of tests to get full test coverage. This also made it very difficult to change the internal representation in any way, since it would require updating many places in the code.
With this CL, the three different APIs for accessing message data now all share a common set of functions. The common functions all take a `upb_MiniTableField` as the canonical description of a field's type and layout. The lowest-level functions are very branchy, as they must test for every possible variation in the field type (field vs oneof, hasbit vs no-hasbit, different field sizes, whether a nonzero default value exists, extension vs. regular field), however these functions are declared inline and designed to be very optimizable when values are known at compile time.
In generated accessors, for example, we can declare constant `upb_MiniTableField` instances so that all values can constant-propagate, and we can get fully specialized code even though we are calling a generic function. On the other hand, when we use the generic functions from reflection, we get runtime branches since values are not known at compile time. But even the function is written to still be as efficient as possible even when used from reflection. For example, we use memcpy() calls with constant length so that the compiler can optimize these into inline loads/stores without having to make an out-of-line call to memcpy().
In this way, this CL should be a benefit to both correctness and performance. It will also make it easier to change the message representation, for example to optimize the encoder by giving hasbits to all fields.
Note that we have not completely consolidated all access in this CL:
1. Some functions outside of get/set such as clear and hazzers are not yet unified.
2. The encoder and decoder still touch the message without going through the common functions. The encoder and decoder require a bit more specialized code to get good performance when reading/writing fields en masse.
PiperOrigin-RevId: 490016095
2 years ago
|
|
|
_upb_MiniTable_SetNonExtensionField(msg, field, &value);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
UPB_API_INLINE int64_t upb_MiniTable_GetInt64(const upb_Message* msg,
|
|
|
|
|
const upb_MiniTableField* field,
|
|
|
|
|
uint64_t default_val) {
|
|
|
|
|
UPB_ASSERT(field->descriptortype == kUpb_FieldType_Int64 ||
|
|
|
|
|
field->descriptortype == kUpb_FieldType_SInt64 ||
|
|
|
|
|
field->descriptortype == kUpb_FieldType_SFixed64);
|
Refactored message accessors to share a common set of functions instead of duplicating logic.
Prior to this CL, there were several different code paths for reading/writing message data. Generated code, MiniTable accessors, and reflection all performed direct manipulation of the bits and bytes in a message, but they all had distinct implementations that did not share much of any code. This divergence meant that they could easily have different behavior, bugs could creep into one but not another, and we would need three different sets of tests to get full test coverage. This also made it very difficult to change the internal representation in any way, since it would require updating many places in the code.
With this CL, the three different APIs for accessing message data now all share a common set of functions. The common functions all take a `upb_MiniTableField` as the canonical description of a field's type and layout. The lowest-level functions are very branchy, as they must test for every possible variation in the field type (field vs oneof, hasbit vs no-hasbit, different field sizes, whether a nonzero default value exists, extension vs. regular field), however these functions are declared inline and designed to be very optimizable when values are known at compile time.
In generated accessors, for example, we can declare constant `upb_MiniTableField` instances so that all values can constant-propagate, and we can get fully specialized code even though we are calling a generic function. On the other hand, when we use the generic functions from reflection, we get runtime branches since values are not known at compile time. But even the function is written to still be as efficient as possible even when used from reflection. For example, we use memcpy() calls with constant length so that the compiler can optimize these into inline loads/stores without having to make an out-of-line call to memcpy().
In this way, this CL should be a benefit to both correctness and performance. It will also make it easier to change the message representation, for example to optimize the encoder by giving hasbits to all fields.
Note that we have not completely consolidated all access in this CL:
1. Some functions outside of get/set such as clear and hazzers are not yet unified.
2. The encoder and decoder still touch the message without going through the common functions. The encoder and decoder require a bit more specialized code to get good performance when reading/writing fields en masse.
PiperOrigin-RevId: 490016095
2 years ago
|
|
|
UPB_ASSUME(!upb_IsRepeatedOrMap(field));
|
|
|
|
|
UPB_ASSUME(_upb_MiniTableField_GetRep(field) == kUpb_FieldRep_8Byte);
|
|
|
|
|
int64_t ret;
|
|
|
|
|
_upb_MiniTable_GetNonExtensionField(msg, field, &default_val, &ret);
|
|
|
|
|
return ret;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
UPB_API_INLINE void upb_MiniTable_SetInt64(upb_Message* msg,
|
|
|
|
|
const upb_MiniTableField* field,
|
|
|
|
|
int64_t value) {
|
|
|
|
|
UPB_ASSERT(field->descriptortype == kUpb_FieldType_Int64 ||
|
|
|
|
|
field->descriptortype == kUpb_FieldType_SInt64 ||
|
|
|
|
|
field->descriptortype == kUpb_FieldType_SFixed64);
|
Refactored message accessors to share a common set of functions instead of duplicating logic.
Prior to this CL, there were several different code paths for reading/writing message data. Generated code, MiniTable accessors, and reflection all performed direct manipulation of the bits and bytes in a message, but they all had distinct implementations that did not share much of any code. This divergence meant that they could easily have different behavior, bugs could creep into one but not another, and we would need three different sets of tests to get full test coverage. This also made it very difficult to change the internal representation in any way, since it would require updating many places in the code.
With this CL, the three different APIs for accessing message data now all share a common set of functions. The common functions all take a `upb_MiniTableField` as the canonical description of a field's type and layout. The lowest-level functions are very branchy, as they must test for every possible variation in the field type (field vs oneof, hasbit vs no-hasbit, different field sizes, whether a nonzero default value exists, extension vs. regular field), however these functions are declared inline and designed to be very optimizable when values are known at compile time.
In generated accessors, for example, we can declare constant `upb_MiniTableField` instances so that all values can constant-propagate, and we can get fully specialized code even though we are calling a generic function. On the other hand, when we use the generic functions from reflection, we get runtime branches since values are not known at compile time. But even the function is written to still be as efficient as possible even when used from reflection. For example, we use memcpy() calls with constant length so that the compiler can optimize these into inline loads/stores without having to make an out-of-line call to memcpy().
In this way, this CL should be a benefit to both correctness and performance. It will also make it easier to change the message representation, for example to optimize the encoder by giving hasbits to all fields.
Note that we have not completely consolidated all access in this CL:
1. Some functions outside of get/set such as clear and hazzers are not yet unified.
2. The encoder and decoder still touch the message without going through the common functions. The encoder and decoder require a bit more specialized code to get good performance when reading/writing fields en masse.
PiperOrigin-RevId: 490016095
2 years ago
|
|
|
UPB_ASSUME(!upb_IsRepeatedOrMap(field));
|
|
|
|
|
UPB_ASSUME(_upb_MiniTableField_GetRep(field) == kUpb_FieldRep_8Byte);
|
|
|
|
|
_upb_MiniTable_SetNonExtensionField(msg, field, &value);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
UPB_API_INLINE uint64_t upb_MiniTable_GetUInt64(const upb_Message* msg,
|
|
|
|
|
const upb_MiniTableField* field,
|
|
|
|
|
uint64_t default_val) {
|
|
|
|
|
UPB_ASSERT(field->descriptortype == kUpb_FieldType_UInt64 ||
|
|
|
|
|
field->descriptortype == kUpb_FieldType_Fixed64);
|
Refactored message accessors to share a common set of functions instead of duplicating logic.
Prior to this CL, there were several different code paths for reading/writing message data. Generated code, MiniTable accessors, and reflection all performed direct manipulation of the bits and bytes in a message, but they all had distinct implementations that did not share much of any code. This divergence meant that they could easily have different behavior, bugs could creep into one but not another, and we would need three different sets of tests to get full test coverage. This also made it very difficult to change the internal representation in any way, since it would require updating many places in the code.
With this CL, the three different APIs for accessing message data now all share a common set of functions. The common functions all take a `upb_MiniTableField` as the canonical description of a field's type and layout. The lowest-level functions are very branchy, as they must test for every possible variation in the field type (field vs oneof, hasbit vs no-hasbit, different field sizes, whether a nonzero default value exists, extension vs. regular field), however these functions are declared inline and designed to be very optimizable when values are known at compile time.
In generated accessors, for example, we can declare constant `upb_MiniTableField` instances so that all values can constant-propagate, and we can get fully specialized code even though we are calling a generic function. On the other hand, when we use the generic functions from reflection, we get runtime branches since values are not known at compile time. But even the function is written to still be as efficient as possible even when used from reflection. For example, we use memcpy() calls with constant length so that the compiler can optimize these into inline loads/stores without having to make an out-of-line call to memcpy().
In this way, this CL should be a benefit to both correctness and performance. It will also make it easier to change the message representation, for example to optimize the encoder by giving hasbits to all fields.
Note that we have not completely consolidated all access in this CL:
1. Some functions outside of get/set such as clear and hazzers are not yet unified.
2. The encoder and decoder still touch the message without going through the common functions. The encoder and decoder require a bit more specialized code to get good performance when reading/writing fields en masse.
PiperOrigin-RevId: 490016095
2 years ago
|
|
|
UPB_ASSUME(!upb_IsRepeatedOrMap(field));
|
|
|
|
|
UPB_ASSUME(_upb_MiniTableField_GetRep(field) == kUpb_FieldRep_8Byte);
|
|
|
|
|
uint64_t ret;
|
|
|
|
|
_upb_MiniTable_GetNonExtensionField(msg, field, &default_val, &ret);
|
|
|
|
|
return ret;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
UPB_API_INLINE void upb_MiniTable_SetUInt64(upb_Message* msg,
|
|
|
|
|
const upb_MiniTableField* field,
|
|
|
|
|
uint64_t value) {
|
|
|
|
|
UPB_ASSERT(field->descriptortype == kUpb_FieldType_UInt64 ||
|
|
|
|
|
field->descriptortype == kUpb_FieldType_Fixed64);
|
Refactored message accessors to share a common set of functions instead of duplicating logic.
Prior to this CL, there were several different code paths for reading/writing message data. Generated code, MiniTable accessors, and reflection all performed direct manipulation of the bits and bytes in a message, but they all had distinct implementations that did not share much of any code. This divergence meant that they could easily have different behavior, bugs could creep into one but not another, and we would need three different sets of tests to get full test coverage. This also made it very difficult to change the internal representation in any way, since it would require updating many places in the code.
With this CL, the three different APIs for accessing message data now all share a common set of functions. The common functions all take a `upb_MiniTableField` as the canonical description of a field's type and layout. The lowest-level functions are very branchy, as they must test for every possible variation in the field type (field vs oneof, hasbit vs no-hasbit, different field sizes, whether a nonzero default value exists, extension vs. regular field), however these functions are declared inline and designed to be very optimizable when values are known at compile time.
In generated accessors, for example, we can declare constant `upb_MiniTableField` instances so that all values can constant-propagate, and we can get fully specialized code even though we are calling a generic function. On the other hand, when we use the generic functions from reflection, we get runtime branches since values are not known at compile time. But even the function is written to still be as efficient as possible even when used from reflection. For example, we use memcpy() calls with constant length so that the compiler can optimize these into inline loads/stores without having to make an out-of-line call to memcpy().
In this way, this CL should be a benefit to both correctness and performance. It will also make it easier to change the message representation, for example to optimize the encoder by giving hasbits to all fields.
Note that we have not completely consolidated all access in this CL:
1. Some functions outside of get/set such as clear and hazzers are not yet unified.
2. The encoder and decoder still touch the message without going through the common functions. The encoder and decoder require a bit more specialized code to get good performance when reading/writing fields en masse.
PiperOrigin-RevId: 490016095
2 years ago
|
|
|
UPB_ASSUME(!upb_IsRepeatedOrMap(field));
|
|
|
|
|
UPB_ASSUME(_upb_MiniTableField_GetRep(field) == kUpb_FieldRep_8Byte);
|
|
|
|
|
_upb_MiniTable_SetNonExtensionField(msg, field, &value);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
UPB_API_INLINE float upb_MiniTable_GetFloat(const upb_Message* msg,
|
|
|
|
|
const upb_MiniTableField* field,
|
|
|
|
|
float default_val) {
|
|
|
|
|
UPB_ASSERT(field->descriptortype == kUpb_FieldType_Float);
|
Refactored message accessors to share a common set of functions instead of duplicating logic.
Prior to this CL, there were several different code paths for reading/writing message data. Generated code, MiniTable accessors, and reflection all performed direct manipulation of the bits and bytes in a message, but they all had distinct implementations that did not share much of any code. This divergence meant that they could easily have different behavior, bugs could creep into one but not another, and we would need three different sets of tests to get full test coverage. This also made it very difficult to change the internal representation in any way, since it would require updating many places in the code.
With this CL, the three different APIs for accessing message data now all share a common set of functions. The common functions all take a `upb_MiniTableField` as the canonical description of a field's type and layout. The lowest-level functions are very branchy, as they must test for every possible variation in the field type (field vs oneof, hasbit vs no-hasbit, different field sizes, whether a nonzero default value exists, extension vs. regular field), however these functions are declared inline and designed to be very optimizable when values are known at compile time.
In generated accessors, for example, we can declare constant `upb_MiniTableField` instances so that all values can constant-propagate, and we can get fully specialized code even though we are calling a generic function. On the other hand, when we use the generic functions from reflection, we get runtime branches since values are not known at compile time. But even the function is written to still be as efficient as possible even when used from reflection. For example, we use memcpy() calls with constant length so that the compiler can optimize these into inline loads/stores without having to make an out-of-line call to memcpy().
In this way, this CL should be a benefit to both correctness and performance. It will also make it easier to change the message representation, for example to optimize the encoder by giving hasbits to all fields.
Note that we have not completely consolidated all access in this CL:
1. Some functions outside of get/set such as clear and hazzers are not yet unified.
2. The encoder and decoder still touch the message without going through the common functions. The encoder and decoder require a bit more specialized code to get good performance when reading/writing fields en masse.
PiperOrigin-RevId: 490016095
2 years ago
|
|
|
UPB_ASSUME(!upb_IsRepeatedOrMap(field));
|
|
|
|
|
UPB_ASSUME(_upb_MiniTableField_GetRep(field) == kUpb_FieldRep_4Byte);
|
|
|
|
|
float ret;
|
|
|
|
|
_upb_MiniTable_GetNonExtensionField(msg, field, &default_val, &ret);
|
|
|
|
|
return ret;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
UPB_API_INLINE void upb_MiniTable_SetFloat(upb_Message* msg,
|
|
|
|
|
const upb_MiniTableField* field,
|
|
|
|
|
float value) {
|
|
|
|
|
UPB_ASSERT(field->descriptortype == kUpb_FieldType_Float);
|
Refactored message accessors to share a common set of functions instead of duplicating logic.
Prior to this CL, there were several different code paths for reading/writing message data. Generated code, MiniTable accessors, and reflection all performed direct manipulation of the bits and bytes in a message, but they all had distinct implementations that did not share much of any code. This divergence meant that they could easily have different behavior, bugs could creep into one but not another, and we would need three different sets of tests to get full test coverage. This also made it very difficult to change the internal representation in any way, since it would require updating many places in the code.
With this CL, the three different APIs for accessing message data now all share a common set of functions. The common functions all take a `upb_MiniTableField` as the canonical description of a field's type and layout. The lowest-level functions are very branchy, as they must test for every possible variation in the field type (field vs oneof, hasbit vs no-hasbit, different field sizes, whether a nonzero default value exists, extension vs. regular field), however these functions are declared inline and designed to be very optimizable when values are known at compile time.
In generated accessors, for example, we can declare constant `upb_MiniTableField` instances so that all values can constant-propagate, and we can get fully specialized code even though we are calling a generic function. On the other hand, when we use the generic functions from reflection, we get runtime branches since values are not known at compile time. But even the function is written to still be as efficient as possible even when used from reflection. For example, we use memcpy() calls with constant length so that the compiler can optimize these into inline loads/stores without having to make an out-of-line call to memcpy().
In this way, this CL should be a benefit to both correctness and performance. It will also make it easier to change the message representation, for example to optimize the encoder by giving hasbits to all fields.
Note that we have not completely consolidated all access in this CL:
1. Some functions outside of get/set such as clear and hazzers are not yet unified.
2. The encoder and decoder still touch the message without going through the common functions. The encoder and decoder require a bit more specialized code to get good performance when reading/writing fields en masse.
PiperOrigin-RevId: 490016095
2 years ago
|
|
|
UPB_ASSUME(!upb_IsRepeatedOrMap(field));
|
|
|
|
|
UPB_ASSUME(_upb_MiniTableField_GetRep(field) == kUpb_FieldRep_4Byte);
|
|
|
|
|
_upb_MiniTable_SetNonExtensionField(msg, field, &value);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
UPB_API_INLINE double upb_MiniTable_GetDouble(const upb_Message* msg,
|
|
|
|
|
const upb_MiniTableField* field,
|
|
|
|
|
double default_val) {
|
|
|
|
|
UPB_ASSERT(field->descriptortype == kUpb_FieldType_Double);
|
Refactored message accessors to share a common set of functions instead of duplicating logic.
Prior to this CL, there were several different code paths for reading/writing message data. Generated code, MiniTable accessors, and reflection all performed direct manipulation of the bits and bytes in a message, but they all had distinct implementations that did not share much of any code. This divergence meant that they could easily have different behavior, bugs could creep into one but not another, and we would need three different sets of tests to get full test coverage. This also made it very difficult to change the internal representation in any way, since it would require updating many places in the code.
With this CL, the three different APIs for accessing message data now all share a common set of functions. The common functions all take a `upb_MiniTableField` as the canonical description of a field's type and layout. The lowest-level functions are very branchy, as they must test for every possible variation in the field type (field vs oneof, hasbit vs no-hasbit, different field sizes, whether a nonzero default value exists, extension vs. regular field), however these functions are declared inline and designed to be very optimizable when values are known at compile time.
In generated accessors, for example, we can declare constant `upb_MiniTableField` instances so that all values can constant-propagate, and we can get fully specialized code even though we are calling a generic function. On the other hand, when we use the generic functions from reflection, we get runtime branches since values are not known at compile time. But even the function is written to still be as efficient as possible even when used from reflection. For example, we use memcpy() calls with constant length so that the compiler can optimize these into inline loads/stores without having to make an out-of-line call to memcpy().
In this way, this CL should be a benefit to both correctness and performance. It will also make it easier to change the message representation, for example to optimize the encoder by giving hasbits to all fields.
Note that we have not completely consolidated all access in this CL:
1. Some functions outside of get/set such as clear and hazzers are not yet unified.
2. The encoder and decoder still touch the message without going through the common functions. The encoder and decoder require a bit more specialized code to get good performance when reading/writing fields en masse.
PiperOrigin-RevId: 490016095
2 years ago
|
|
|
UPB_ASSUME(!upb_IsRepeatedOrMap(field));
|
|
|
|
|
UPB_ASSUME(_upb_MiniTableField_GetRep(field) == kUpb_FieldRep_8Byte);
|
|
|
|
|
double ret;
|
|
|
|
|
_upb_MiniTable_GetNonExtensionField(msg, field, &default_val, &ret);
|
|
|
|
|
return ret;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
UPB_API_INLINE void upb_MiniTable_SetDouble(upb_Message* msg,
|
|
|
|
|
const upb_MiniTableField* field,
|
|
|
|
|
double value) {
|
|
|
|
|
UPB_ASSERT(field->descriptortype == kUpb_FieldType_Double);
|
Refactored message accessors to share a common set of functions instead of duplicating logic.
Prior to this CL, there were several different code paths for reading/writing message data. Generated code, MiniTable accessors, and reflection all performed direct manipulation of the bits and bytes in a message, but they all had distinct implementations that did not share much of any code. This divergence meant that they could easily have different behavior, bugs could creep into one but not another, and we would need three different sets of tests to get full test coverage. This also made it very difficult to change the internal representation in any way, since it would require updating many places in the code.
With this CL, the three different APIs for accessing message data now all share a common set of functions. The common functions all take a `upb_MiniTableField` as the canonical description of a field's type and layout. The lowest-level functions are very branchy, as they must test for every possible variation in the field type (field vs oneof, hasbit vs no-hasbit, different field sizes, whether a nonzero default value exists, extension vs. regular field), however these functions are declared inline and designed to be very optimizable when values are known at compile time.
In generated accessors, for example, we can declare constant `upb_MiniTableField` instances so that all values can constant-propagate, and we can get fully specialized code even though we are calling a generic function. On the other hand, when we use the generic functions from reflection, we get runtime branches since values are not known at compile time. But even the function is written to still be as efficient as possible even when used from reflection. For example, we use memcpy() calls with constant length so that the compiler can optimize these into inline loads/stores without having to make an out-of-line call to memcpy().
In this way, this CL should be a benefit to both correctness and performance. It will also make it easier to change the message representation, for example to optimize the encoder by giving hasbits to all fields.
Note that we have not completely consolidated all access in this CL:
1. Some functions outside of get/set such as clear and hazzers are not yet unified.
2. The encoder and decoder still touch the message without going through the common functions. The encoder and decoder require a bit more specialized code to get good performance when reading/writing fields en masse.
PiperOrigin-RevId: 490016095
2 years ago
|
|
|
UPB_ASSUME(!upb_IsRepeatedOrMap(field));
|
|
|
|
|
UPB_ASSUME(_upb_MiniTableField_GetRep(field) == kUpb_FieldRep_8Byte);
|
|
|
|
|
_upb_MiniTable_SetNonExtensionField(msg, field, &value);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
UPB_API_INLINE upb_StringView
|
|
|
|
|
upb_MiniTable_GetString(const upb_Message* msg, const upb_MiniTableField* field,
|
|
|
|
|
upb_StringView def_val) {
|
|
|
|
|
UPB_ASSERT(field->descriptortype == kUpb_FieldType_Bytes ||
|
|
|
|
|
field->descriptortype == kUpb_FieldType_String);
|
Refactored message accessors to share a common set of functions instead of duplicating logic.
Prior to this CL, there were several different code paths for reading/writing message data. Generated code, MiniTable accessors, and reflection all performed direct manipulation of the bits and bytes in a message, but they all had distinct implementations that did not share much of any code. This divergence meant that they could easily have different behavior, bugs could creep into one but not another, and we would need three different sets of tests to get full test coverage. This also made it very difficult to change the internal representation in any way, since it would require updating many places in the code.
With this CL, the three different APIs for accessing message data now all share a common set of functions. The common functions all take a `upb_MiniTableField` as the canonical description of a field's type and layout. The lowest-level functions are very branchy, as they must test for every possible variation in the field type (field vs oneof, hasbit vs no-hasbit, different field sizes, whether a nonzero default value exists, extension vs. regular field), however these functions are declared inline and designed to be very optimizable when values are known at compile time.
In generated accessors, for example, we can declare constant `upb_MiniTableField` instances so that all values can constant-propagate, and we can get fully specialized code even though we are calling a generic function. On the other hand, when we use the generic functions from reflection, we get runtime branches since values are not known at compile time. But even the function is written to still be as efficient as possible even when used from reflection. For example, we use memcpy() calls with constant length so that the compiler can optimize these into inline loads/stores without having to make an out-of-line call to memcpy().
In this way, this CL should be a benefit to both correctness and performance. It will also make it easier to change the message representation, for example to optimize the encoder by giving hasbits to all fields.
Note that we have not completely consolidated all access in this CL:
1. Some functions outside of get/set such as clear and hazzers are not yet unified.
2. The encoder and decoder still touch the message without going through the common functions. The encoder and decoder require a bit more specialized code to get good performance when reading/writing fields en masse.
PiperOrigin-RevId: 490016095
2 years ago
|
|
|
UPB_ASSUME(!upb_IsRepeatedOrMap(field));
|
|
|
|
|
UPB_ASSUME(_upb_MiniTableField_GetRep(field) == kUpb_FieldRep_StringView);
|
|
|
|
|
upb_StringView ret;
|
|
|
|
|
_upb_MiniTable_GetNonExtensionField(msg, field, &def_val, &ret);
|
|
|
|
|
return ret;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
UPB_API_INLINE void upb_MiniTable_SetString(upb_Message* msg,
|
|
|
|
|
const upb_MiniTableField* field,
|
|
|
|
|
upb_StringView value) {
|
|
|
|
|
UPB_ASSERT(field->descriptortype == kUpb_FieldType_Bytes ||
|
|
|
|
|
field->descriptortype == kUpb_FieldType_String);
|
Refactored message accessors to share a common set of functions instead of duplicating logic.
Prior to this CL, there were several different code paths for reading/writing message data. Generated code, MiniTable accessors, and reflection all performed direct manipulation of the bits and bytes in a message, but they all had distinct implementations that did not share much of any code. This divergence meant that they could easily have different behavior, bugs could creep into one but not another, and we would need three different sets of tests to get full test coverage. This also made it very difficult to change the internal representation in any way, since it would require updating many places in the code.
With this CL, the three different APIs for accessing message data now all share a common set of functions. The common functions all take a `upb_MiniTableField` as the canonical description of a field's type and layout. The lowest-level functions are very branchy, as they must test for every possible variation in the field type (field vs oneof, hasbit vs no-hasbit, different field sizes, whether a nonzero default value exists, extension vs. regular field), however these functions are declared inline and designed to be very optimizable when values are known at compile time.
In generated accessors, for example, we can declare constant `upb_MiniTableField` instances so that all values can constant-propagate, and we can get fully specialized code even though we are calling a generic function. On the other hand, when we use the generic functions from reflection, we get runtime branches since values are not known at compile time. But even the function is written to still be as efficient as possible even when used from reflection. For example, we use memcpy() calls with constant length so that the compiler can optimize these into inline loads/stores without having to make an out-of-line call to memcpy().
In this way, this CL should be a benefit to both correctness and performance. It will also make it easier to change the message representation, for example to optimize the encoder by giving hasbits to all fields.
Note that we have not completely consolidated all access in this CL:
1. Some functions outside of get/set such as clear and hazzers are not yet unified.
2. The encoder and decoder still touch the message without going through the common functions. The encoder and decoder require a bit more specialized code to get good performance when reading/writing fields en masse.
PiperOrigin-RevId: 490016095
2 years ago
|
|
|
UPB_ASSUME(!upb_IsRepeatedOrMap(field));
|
|
|
|
|
UPB_ASSUME(_upb_MiniTableField_GetRep(field) == kUpb_FieldRep_StringView);
|
|
|
|
|
_upb_MiniTable_SetNonExtensionField(msg, field, &value);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
UPB_API_INLINE const upb_Message* upb_MiniTable_GetMessage(
|
|
|
|
|
const upb_Message* msg, const upb_MiniTableField* field,
|
|
|
|
|
upb_Message* default_val) {
|
|
|
|
|
UPB_ASSERT(field->descriptortype == kUpb_FieldType_Message ||
|
|
|
|
|
field->descriptortype == kUpb_FieldType_Group);
|
Refactored message accessors to share a common set of functions instead of duplicating logic.
Prior to this CL, there were several different code paths for reading/writing message data. Generated code, MiniTable accessors, and reflection all performed direct manipulation of the bits and bytes in a message, but they all had distinct implementations that did not share much of any code. This divergence meant that they could easily have different behavior, bugs could creep into one but not another, and we would need three different sets of tests to get full test coverage. This also made it very difficult to change the internal representation in any way, since it would require updating many places in the code.
With this CL, the three different APIs for accessing message data now all share a common set of functions. The common functions all take a `upb_MiniTableField` as the canonical description of a field's type and layout. The lowest-level functions are very branchy, as they must test for every possible variation in the field type (field vs oneof, hasbit vs no-hasbit, different field sizes, whether a nonzero default value exists, extension vs. regular field), however these functions are declared inline and designed to be very optimizable when values are known at compile time.
In generated accessors, for example, we can declare constant `upb_MiniTableField` instances so that all values can constant-propagate, and we can get fully specialized code even though we are calling a generic function. On the other hand, when we use the generic functions from reflection, we get runtime branches since values are not known at compile time. But even the function is written to still be as efficient as possible even when used from reflection. For example, we use memcpy() calls with constant length so that the compiler can optimize these into inline loads/stores without having to make an out-of-line call to memcpy().
In this way, this CL should be a benefit to both correctness and performance. It will also make it easier to change the message representation, for example to optimize the encoder by giving hasbits to all fields.
Note that we have not completely consolidated all access in this CL:
1. Some functions outside of get/set such as clear and hazzers are not yet unified.
2. The encoder and decoder still touch the message without going through the common functions. The encoder and decoder require a bit more specialized code to get good performance when reading/writing fields en masse.
PiperOrigin-RevId: 490016095
2 years ago
|
|
|
UPB_ASSUME(!upb_IsRepeatedOrMap(field));
|
|
|
|
|
UPB_ASSUME(_upb_MiniTableField_GetRep(field) ==
|
|
|
|
|
UPB_SIZE(kUpb_FieldRep_4Byte, kUpb_FieldRep_8Byte));
|
|
|
|
|
upb_Message* ret;
|
|
|
|
|
_upb_MiniTable_GetNonExtensionField(msg, field, &default_val, &ret);
|
|
|
|
|
return ret;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
UPB_API_INLINE void upb_MiniTable_SetMessage(upb_Message* msg,
|
|
|
|
|
const upb_MiniTable* mini_table,
|
|
|
|
|
const upb_MiniTableField* field,
|
|
|
|
|
upb_Message* sub_message) {
|
|
|
|
|
UPB_ASSERT(field->descriptortype == kUpb_FieldType_Message ||
|
|
|
|
|
field->descriptortype == kUpb_FieldType_Group);
|
Refactored message accessors to share a common set of functions instead of duplicating logic.
Prior to this CL, there were several different code paths for reading/writing message data. Generated code, MiniTable accessors, and reflection all performed direct manipulation of the bits and bytes in a message, but they all had distinct implementations that did not share much of any code. This divergence meant that they could easily have different behavior, bugs could creep into one but not another, and we would need three different sets of tests to get full test coverage. This also made it very difficult to change the internal representation in any way, since it would require updating many places in the code.
With this CL, the three different APIs for accessing message data now all share a common set of functions. The common functions all take a `upb_MiniTableField` as the canonical description of a field's type and layout. The lowest-level functions are very branchy, as they must test for every possible variation in the field type (field vs oneof, hasbit vs no-hasbit, different field sizes, whether a nonzero default value exists, extension vs. regular field), however these functions are declared inline and designed to be very optimizable when values are known at compile time.
In generated accessors, for example, we can declare constant `upb_MiniTableField` instances so that all values can constant-propagate, and we can get fully specialized code even though we are calling a generic function. On the other hand, when we use the generic functions from reflection, we get runtime branches since values are not known at compile time. But even the function is written to still be as efficient as possible even when used from reflection. For example, we use memcpy() calls with constant length so that the compiler can optimize these into inline loads/stores without having to make an out-of-line call to memcpy().
In this way, this CL should be a benefit to both correctness and performance. It will also make it easier to change the message representation, for example to optimize the encoder by giving hasbits to all fields.
Note that we have not completely consolidated all access in this CL:
1. Some functions outside of get/set such as clear and hazzers are not yet unified.
2. The encoder and decoder still touch the message without going through the common functions. The encoder and decoder require a bit more specialized code to get good performance when reading/writing fields en masse.
PiperOrigin-RevId: 490016095
2 years ago
|
|
|
UPB_ASSUME(!upb_IsRepeatedOrMap(field));
|
|
|
|
|
UPB_ASSUME(_upb_MiniTableField_GetRep(field) ==
|
|
|
|
|
UPB_SIZE(kUpb_FieldRep_4Byte, kUpb_FieldRep_8Byte));
|
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
|
|
|
UPB_ASSERT(mini_table->subs[field->submsg_index].submsg);
|
Refactored message accessors to share a common set of functions instead of duplicating logic.
Prior to this CL, there were several different code paths for reading/writing message data. Generated code, MiniTable accessors, and reflection all performed direct manipulation of the bits and bytes in a message, but they all had distinct implementations that did not share much of any code. This divergence meant that they could easily have different behavior, bugs could creep into one but not another, and we would need three different sets of tests to get full test coverage. This also made it very difficult to change the internal representation in any way, since it would require updating many places in the code.
With this CL, the three different APIs for accessing message data now all share a common set of functions. The common functions all take a `upb_MiniTableField` as the canonical description of a field's type and layout. The lowest-level functions are very branchy, as they must test for every possible variation in the field type (field vs oneof, hasbit vs no-hasbit, different field sizes, whether a nonzero default value exists, extension vs. regular field), however these functions are declared inline and designed to be very optimizable when values are known at compile time.
In generated accessors, for example, we can declare constant `upb_MiniTableField` instances so that all values can constant-propagate, and we can get fully specialized code even though we are calling a generic function. On the other hand, when we use the generic functions from reflection, we get runtime branches since values are not known at compile time. But even the function is written to still be as efficient as possible even when used from reflection. For example, we use memcpy() calls with constant length so that the compiler can optimize these into inline loads/stores without having to make an out-of-line call to memcpy().
In this way, this CL should be a benefit to both correctness and performance. It will also make it easier to change the message representation, for example to optimize the encoder by giving hasbits to all fields.
Note that we have not completely consolidated all access in this CL:
1. Some functions outside of get/set such as clear and hazzers are not yet unified.
2. The encoder and decoder still touch the message without going through the common functions. The encoder and decoder require a bit more specialized code to get good performance when reading/writing fields en masse.
PiperOrigin-RevId: 490016095
2 years ago
|
|
|
_upb_MiniTable_SetNonExtensionField(msg, field, &sub_message);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
UPB_API_INLINE upb_Message* upb_MiniTable_GetMutableMessage(
|
|
|
|
|
upb_Message* msg, const upb_MiniTable* mini_table,
|
|
|
|
|
const upb_MiniTableField* field, upb_Arena* arena) {
|
|
|
|
|
UPB_ASSERT(field->descriptortype == kUpb_FieldType_Message ||
|
|
|
|
|
field->descriptortype == kUpb_FieldType_Group);
|
|
|
|
|
upb_Message* sub_message = *UPB_PTR_AT(msg, field->offset, upb_Message*);
|
|
|
|
|
if (!sub_message) {
|
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
|
|
|
const upb_MiniTable* sub_mini_table =
|
|
|
|
|
mini_table->subs[field->submsg_index].submsg;
|
|
|
|
|
UPB_ASSERT(sub_mini_table);
|
|
|
|
|
sub_message = _upb_Message_New(sub_mini_table, arena);
|
|
|
|
|
*UPB_PTR_AT(msg, field->offset, upb_Message*) = sub_message;
|
|
|
|
|
_upb_MiniTable_SetPresence(msg, field);
|
|
|
|
|
}
|
|
|
|
|
return sub_message;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
UPB_API_INLINE const upb_Array* upb_MiniTable_GetArray(
|
|
|
|
|
const upb_Message* msg, const upb_MiniTableField* field) {
|
Refactored message accessors to share a common set of functions instead of duplicating logic.
Prior to this CL, there were several different code paths for reading/writing message data. Generated code, MiniTable accessors, and reflection all performed direct manipulation of the bits and bytes in a message, but they all had distinct implementations that did not share much of any code. This divergence meant that they could easily have different behavior, bugs could creep into one but not another, and we would need three different sets of tests to get full test coverage. This also made it very difficult to change the internal representation in any way, since it would require updating many places in the code.
With this CL, the three different APIs for accessing message data now all share a common set of functions. The common functions all take a `upb_MiniTableField` as the canonical description of a field's type and layout. The lowest-level functions are very branchy, as they must test for every possible variation in the field type (field vs oneof, hasbit vs no-hasbit, different field sizes, whether a nonzero default value exists, extension vs. regular field), however these functions are declared inline and designed to be very optimizable when values are known at compile time.
In generated accessors, for example, we can declare constant `upb_MiniTableField` instances so that all values can constant-propagate, and we can get fully specialized code even though we are calling a generic function. On the other hand, when we use the generic functions from reflection, we get runtime branches since values are not known at compile time. But even the function is written to still be as efficient as possible even when used from reflection. For example, we use memcpy() calls with constant length so that the compiler can optimize these into inline loads/stores without having to make an out-of-line call to memcpy().
In this way, this CL should be a benefit to both correctness and performance. It will also make it easier to change the message representation, for example to optimize the encoder by giving hasbits to all fields.
Note that we have not completely consolidated all access in this CL:
1. Some functions outside of get/set such as clear and hazzers are not yet unified.
2. The encoder and decoder still touch the message without going through the common functions. The encoder and decoder require a bit more specialized code to get good performance when reading/writing fields en masse.
PiperOrigin-RevId: 490016095
2 years ago
|
|
|
const upb_Array* ret;
|
|
|
|
|
const upb_Array* default_val = NULL;
|
|
|
|
|
_upb_MiniTable_GetNonExtensionField(msg, field, &default_val, &ret);
|
|
|
|
|
return ret;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
UPB_API_INLINE upb_Array* upb_MiniTable_GetMutableArray(
|
|
|
|
|
upb_Message* msg, const upb_MiniTableField* field) {
|
Refactored message accessors to share a common set of functions instead of duplicating logic.
Prior to this CL, there were several different code paths for reading/writing message data. Generated code, MiniTable accessors, and reflection all performed direct manipulation of the bits and bytes in a message, but they all had distinct implementations that did not share much of any code. This divergence meant that they could easily have different behavior, bugs could creep into one but not another, and we would need three different sets of tests to get full test coverage. This also made it very difficult to change the internal representation in any way, since it would require updating many places in the code.
With this CL, the three different APIs for accessing message data now all share a common set of functions. The common functions all take a `upb_MiniTableField` as the canonical description of a field's type and layout. The lowest-level functions are very branchy, as they must test for every possible variation in the field type (field vs oneof, hasbit vs no-hasbit, different field sizes, whether a nonzero default value exists, extension vs. regular field), however these functions are declared inline and designed to be very optimizable when values are known at compile time.
In generated accessors, for example, we can declare constant `upb_MiniTableField` instances so that all values can constant-propagate, and we can get fully specialized code even though we are calling a generic function. On the other hand, when we use the generic functions from reflection, we get runtime branches since values are not known at compile time. But even the function is written to still be as efficient as possible even when used from reflection. For example, we use memcpy() calls with constant length so that the compiler can optimize these into inline loads/stores without having to make an out-of-line call to memcpy().
In this way, this CL should be a benefit to both correctness and performance. It will also make it easier to change the message representation, for example to optimize the encoder by giving hasbits to all fields.
Note that we have not completely consolidated all access in this CL:
1. Some functions outside of get/set such as clear and hazzers are not yet unified.
2. The encoder and decoder still touch the message without going through the common functions. The encoder and decoder require a bit more specialized code to get good performance when reading/writing fields en masse.
PiperOrigin-RevId: 490016095
2 years ago
|
|
|
return (upb_Array*)upb_MiniTable_GetArray(msg, field);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void* upb_MiniTable_ResizeArray(upb_Message* msg,
|
|
|
|
|
const upb_MiniTableField* field, size_t len,
|
|
|
|
|
upb_Arena* arena);
|
|
|
|
|
typedef enum {
|
|
|
|
|
kUpb_GetExtension_Ok,
|
|
|
|
|
kUpb_GetExtension_NotPresent,
|
|
|
|
|
kUpb_GetExtension_ParseError,
|
|
|
|
|
kUpb_GetExtension_OutOfMemory,
|
|
|
|
|
} upb_GetExtension_Status;
|
|
|
|
|
|
|
|
|
|
typedef enum {
|
|
|
|
|
kUpb_GetExtensionAsBytes_Ok,
|
|
|
|
|
kUpb_GetExtensionAsBytes_NotPresent,
|
|
|
|
|
kUpb_GetExtensionAsBytes_EncodeError,
|
|
|
|
|
} upb_GetExtensionAsBytes_Status;
|
|
|
|
|
|
|
|
|
|
// Returns a message extension or promotes an unknown field to
|
|
|
|
|
// an extension.
|
|
|
|
|
//
|
|
|
|
|
// TODO(ferhat): Only supports extension fields that are messages,
|
|
|
|
|
// expand support to include non-message types.
|
|
|
|
|
upb_GetExtension_Status upb_MiniTable_GetOrPromoteExtension(
|
|
|
|
|
upb_Message* msg, const upb_MiniTableExtension* ext_table,
|
|
|
|
|
int decode_options, upb_Arena* arena,
|
|
|
|
|
const upb_Message_Extension** extension);
|
|
|
|
|
|
|
|
|
|
// Returns a message extension or unknown field matching the extension
|
|
|
|
|
// data as bytes.
|
|
|
|
|
//
|
|
|
|
|
// If an extension has already been decoded it will be re-encoded
|
|
|
|
|
// to bytes.
|
|
|
|
|
upb_GetExtensionAsBytes_Status upb_MiniTable_GetExtensionAsBytes(
|
|
|
|
|
const upb_Message* msg, const upb_MiniTableExtension* ext_table,
|
|
|
|
|
int encode_options, upb_Arena* arena, const char** extension_data,
|
|
|
|
|
size_t* len);
|
|
|
|
|
|
|
|
|
|
typedef enum {
|
|
|
|
|
kUpb_FindUnknown_Ok,
|
|
|
|
|
kUpb_FindUnknown_NotPresent,
|
|
|
|
|
kUpb_FindUnknown_ParseError,
|
|
|
|
|
} upb_FindUnknown_Status;
|
|
|
|
|
|
|
|
|
|
typedef struct {
|
|
|
|
|
upb_FindUnknown_Status status;
|
|
|
|
|
// Start of unknown field data in message arena.
|
|
|
|
|
const char* ptr;
|
|
|
|
|
// Size of unknown field data.
|
|
|
|
|
size_t len;
|
|
|
|
|
} upb_FindUnknownRet;
|
|
|
|
|
|
|
|
|
|
// Finds first occurrence of unknown data by tag id in message.
|
|
|
|
|
upb_FindUnknownRet upb_MiniTable_FindUnknown(const upb_Message* msg,
|
|
|
|
|
uint32_t field_number);
|
|
|
|
|
|
|
|
|
|
typedef enum {
|
|
|
|
|
kUpb_UnknownToMessage_Ok,
|
|
|
|
|
kUpb_UnknownToMessage_ParseError,
|
|
|
|
|
kUpb_UnknownToMessage_OutOfMemory,
|
|
|
|
|
kUpb_UnknownToMessage_NotFound,
|
|
|
|
|
} upb_UnknownToMessage_Status;
|
|
|
|
|
|
|
|
|
|
typedef struct {
|
|
|
|
|
upb_UnknownToMessage_Status status;
|
|
|
|
|
upb_Message* message;
|
|
|
|
|
} upb_UnknownToMessageRet;
|
|
|
|
|
|
|
|
|
|
// Promotes unknown data inside message to a upb_Message parsing the unknown.
|
|
|
|
|
//
|
|
|
|
|
// The unknown data is removed from message after field value is set
|
|
|
|
|
// using upb_MiniTable_SetMessage.
|
|
|
|
|
upb_UnknownToMessageRet upb_MiniTable_PromoteUnknownToMessage(
|
|
|
|
|
upb_Message* msg, const upb_MiniTable* mini_table,
|
|
|
|
|
const upb_MiniTableField* field, const upb_MiniTable* sub_mini_table,
|
|
|
|
|
int decode_options, upb_Arena* arena);
|
|
|
|
|
|
|
|
|
|
// Promotes all unknown data that matches field tag id to repeated messages
|
|
|
|
|
// in upb_Array.
|
|
|
|
|
//
|
|
|
|
|
// The unknown data is removed from message after upb_Array is populated.
|
|
|
|
|
// Since repeated messages can't be packed we remove each unknown that
|
|
|
|
|
// contains the target tag id.
|
|
|
|
|
upb_UnknownToMessage_Status upb_MiniTable_PromoteUnknownToMessageArray(
|
|
|
|
|
upb_Message* msg, const upb_MiniTableField* field,
|
|
|
|
|
const upb_MiniTable* mini_table, int decode_options, upb_Arena* arena);
|
|
|
|
|
|
|
|
|
|
#ifdef __cplusplus
|
|
|
|
|
} /* extern "C" */
|
|
|
|
|
#endif
|
|
|
|
|
|
|
|
|
|
#include "upb/port/undef.inc"
|
|
|
|
|
|
|
|
|
|
#endif // UPB_MESSAGE_ACCESSORS_H_
|
