|
|
|
// Protocol Buffers - Google's data interchange format
|
|
|
|
// Copyright 2023 Google LLC. All rights reserved.
|
|
|
|
//
|
|
|
|
// Use of this source code is governed by a BSD-style
|
|
|
|
// license that can be found in the LICENSE file or at
|
|
|
|
// https://developers.google.com/open-source/licenses/bsd
|
|
|
|
|
|
|
|
/*
|
|
|
|
* This is where we define internal portability macros used across upb.
|
|
|
|
*
|
|
|
|
* All of these macros are undef'd in undef.inc to avoid leaking them to users.
|
|
|
|
*
|
|
|
|
* The correct usage is:
|
|
|
|
*
|
|
|
|
* #include "upb/foobar.h"
|
|
|
|
* #include "upb/baz.h"
|
|
|
|
*
|
|
|
|
* // MUST be last included header.
|
|
|
|
* #include "upb/port/def.inc"
|
|
|
|
*
|
|
|
|
* // Code for this file.
|
|
|
|
* // <...>
|
|
|
|
*
|
|
|
|
* // Can be omitted for .c files, required for .h.
|
|
|
|
* #include "upb/port/undef.inc"
|
|
|
|
*
|
|
|
|
* This file is private and must not be included by users!
|
|
|
|
*/
|
|
|
|
|
|
|
|
#if !((defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L) || \
|
|
|
|
(defined(__cplusplus) && __cplusplus >= 201402L) || \
|
|
|
|
(defined(_MSC_VER) && _MSC_VER >= 1900))
|
|
|
|
#error upb requires C99 or C++14 or MSVC >= 2015.
|
|
|
|
#endif
|
|
|
|
|
|
|
|
// Portable check for GCC minimum version:
|
|
|
|
// https://gcc.gnu.org/onlinedocs/cpp/Common-Predefined-Macros.html
|
|
|
|
#if defined(__GNUC__) && defined(__GNUC_MINOR__) && defined(__GNUC_PATCHLEVEL__)
|
|
|
|
#define UPB_GNUC_MIN(x, y) \
|
|
|
|
(__GNUC__ > (x) || __GNUC__ == (x) && __GNUC_MINOR__ >= (y))
|
|
|
|
#else
|
|
|
|
#define UPB_GNUC_MIN(x, y) 0
|
|
|
|
#endif
|
|
|
|
|
|
|
|
#include <assert.h>
|
|
|
|
#include <setjmp.h>
|
|
|
|
#include <stdbool.h>
|
|
|
|
#include <stdint.h>
|
|
|
|
#include <stdio.h>
|
|
|
|
|
|
|
|
#ifndef UINTPTR_MAX
|
|
|
|
Error, UINTPTR_MAX is undefined
|
|
|
|
#endif
|
|
|
|
|
|
|
|
#if UINTPTR_MAX == 0xffffffff
|
|
|
|
#define UPB_SIZE(size32, size64) size32
|
|
|
|
#else
|
|
|
|
#define UPB_SIZE(size32, size64) size64
|
|
|
|
#endif
|
|
|
|
|
|
|
|
/* If we always read/write as a consistent type to each address, this shouldn't
|
|
|
|
* violate aliasing.
|
|
|
|
*/
|
|
|
|
#define UPB_PTR_AT(msg, ofs, type) ((type*)((char*)(msg) + (ofs)))
|
|
|
|
|
|
|
|
#define UPB_MAPTYPE_STRING 0
|
|
|
|
|
|
|
|
// UPB_EXPORT: always generate a public symbol.
|
|
|
|
#if defined(__GNUC__) || defined(__clang__)
|
|
|
|
#define UPB_EXPORT __attribute__((visibility("default"))) __attribute__((used))
|
|
|
|
#else
|
|
|
|
#define UPB_EXPORT
|
|
|
|
#endif
|
|
|
|
|
|
|
|
// UPB_INLINE: inline if possible, emit standalone code if required.
|
|
|
|
#ifdef __cplusplus
|
|
|
|
#define UPB_INLINE inline
|
|
|
|
#elif defined (__GNUC__) || defined(__clang__)
|
|
|
|
#define UPB_INLINE static __inline__
|
|
|
|
#else
|
|
|
|
#define UPB_INLINE static
|
|
|
|
#endif
|
|
|
|
|
|
|
|
#ifdef UPB_BUILD_API
|
|
|
|
#define UPB_API UPB_EXPORT
|
|
|
|
#define UPB_API_INLINE UPB_EXPORT
|
|
|
|
#else
|
|
|
|
#define UPB_API
|
|
|
|
#define UPB_API_INLINE UPB_INLINE
|
|
|
|
#endif
|
|
|
|
|
|
|
|
#define UPB_MALLOC_ALIGN 8
|
|
|
|
#define UPB_ALIGN_UP(size, align) (((size) + (align) - 1) / (align) * (align))
|
|
|
|
#define UPB_ALIGN_DOWN(size, align) ((size) / (align) * (align))
|
|
|
|
#define UPB_ALIGN_MALLOC(size) UPB_ALIGN_UP(size, UPB_MALLOC_ALIGN)
|
|
|
|
#ifdef __clang__
|
|
|
|
#define UPB_ALIGN_OF(type) _Alignof(type)
|
|
|
|
#else
|
|
|
|
#define UPB_ALIGN_OF(type) offsetof (struct { char c; type member; }, member)
|
|
|
|
#endif
|
|
|
|
|
|
|
|
// Hints to the compiler about likely/unlikely branches.
|
|
|
|
#if defined (__GNUC__) || defined(__clang__)
|
|
|
|
#define UPB_LIKELY(x) __builtin_expect((bool)(x), 1)
|
|
|
|
#define UPB_UNLIKELY(x) __builtin_expect((bool)(x), 0)
|
|
|
|
#else
|
|
|
|
#define UPB_LIKELY(x) (x)
|
|
|
|
#define UPB_UNLIKELY(x) (x)
|
|
|
|
#endif
|
|
|
|
|
|
|
|
// Macros for function attributes on compilers that support them.
|
|
|
|
#ifdef __GNUC__
|
|
|
|
#define UPB_FORCEINLINE __inline__ __attribute__((always_inline))
|
|
|
|
#define UPB_NOINLINE __attribute__((noinline))
|
|
|
|
#define UPB_NORETURN __attribute__((__noreturn__))
|
|
|
|
#define UPB_PRINTF(str, first_vararg) __attribute__((format (printf, str, first_vararg)))
|
|
|
|
#elif defined(_MSC_VER)
|
|
|
|
#define UPB_NOINLINE
|
|
|
|
#define UPB_FORCEINLINE
|
|
|
|
#define UPB_NORETURN __declspec(noreturn)
|
|
|
|
#define UPB_PRINTF(str, first_vararg)
|
|
|
|
#else /* !defined(__GNUC__) */
|
|
|
|
#define UPB_FORCEINLINE
|
|
|
|
#define UPB_NOINLINE
|
|
|
|
#define UPB_NORETURN
|
|
|
|
#define UPB_PRINTF(str, first_vararg)
|
|
|
|
#endif
|
|
|
|
|
|
|
|
#define UPB_MAX(x, y) ((x) > (y) ? (x) : (y))
|
|
|
|
#define UPB_MIN(x, y) ((x) < (y) ? (x) : (y))
|
|
|
|
|
|
|
|
#define UPB_UNUSED(var) (void)var
|
|
|
|
|
|
|
|
// UPB_ASSUME(): in release mode, we tell the compiler to assume this is true.
|
|
|
|
#ifdef NDEBUG
|
|
|
|
#ifdef __GNUC__
|
|
|
|
#define UPB_ASSUME(expr) if (!(expr)) __builtin_unreachable()
|
|
|
|
#elif defined _MSC_VER
|
|
|
|
#define UPB_ASSUME(expr) if (!(expr)) __assume(0)
|
|
|
|
#else
|
|
|
|
#define UPB_ASSUME(expr) do {} while (false && (expr))
|
|
|
|
#endif
|
|
|
|
#else
|
|
|
|
#define UPB_ASSUME(expr) assert(expr)
|
|
|
|
#endif
|
|
|
|
|
|
|
|
/* UPB_ASSERT(): in release mode, we use the expression without letting it be
|
|
|
|
* evaluated. This prevents "unused variable" warnings. */
|
|
|
|
#ifdef NDEBUG
|
|
|
|
#define UPB_ASSERT(expr) do {} while (false && (expr))
|
|
|
|
#else
|
|
|
|
#define UPB_ASSERT(expr) assert(expr)
|
|
|
|
#endif
|
|
|
|
|
|
|
|
#if defined(__GNUC__) || defined(__clang__)
|
|
|
|
#define UPB_UNREACHABLE() do { assert(0); __builtin_unreachable(); } while(0)
|
|
|
|
#elif defined(_MSC_VER)
|
|
|
|
#define UPB_UNREACHABLE() \
|
|
|
|
do { \
|
|
|
|
assert(0); \
|
|
|
|
__assume(0); \
|
|
|
|
} while (0)
|
|
|
|
#else
|
|
|
|
#define UPB_UNREACHABLE() do { assert(0); } while(0)
|
|
|
|
#endif
|
|
|
|
|
|
|
|
/* UPB_SETJMP() / UPB_LONGJMP(): avoid setting/restoring signal mask. */
|
|
|
|
#ifdef __APPLE__
|
|
|
|
#define UPB_SETJMP(buf) _setjmp(buf)
|
|
|
|
#define UPB_LONGJMP(buf, val) _longjmp(buf, val)
|
|
|
|
#else
|
|
|
|
#define UPB_SETJMP(buf) setjmp(buf)
|
|
|
|
#define UPB_LONGJMP(buf, val) longjmp(buf, val)
|
|
|
|
#endif
|
|
|
|
|
Allow for fuse/free races in `upb_Arena`.
Implementation is by kfm@, I only added the portability code around it.
`upb_Arena` was designed to be only thread-compatible. However, fusing of arenas muddies the waters somewhat, because two distinct `upb_Arena` objects will end up sharing state when fused. This causes a `upb_Arena_Free(a)` to interfere with `upb_Arena_Fuse(b, c)` if `a` and `b` were previously fused.
It turns out that we can use atomics to fix this with about a 35% regression in fuse performance (see below). Arena create+free does not regress, thanks to special-case logic in Free().
`upb_Arena` is still a thread-compatible type, and it is still never safe to call `upb_Arena_xxx(a)` and `upb_Arena_yyy(a)` in parallel. However you can at least now call `upb_Arena_Free(a)` and `upb_Arena_Fuse(b, c)` in parallel, even if `a` and `b` were previously fused.
Note that `upb_Arena_Fuse(a, b)` and `upb_Arena_Fuse(c, d)` is still not allowed if `b` and `c` were previously fused. In practice this means that fuses must still be single-threaded within a single fused group.
Performance results:
```
name old cpu/op new cpu/op delta
BM_ArenaOneAlloc 18.6ns ± 1% 18.6ns ± 1% ~ (p=0.726 n=18+17)
BM_ArenaInitialBlockOneAlloc 6.28ns ± 1% 5.73ns ± 1% -8.68% (p=0.000 n=17+20)
BM_ArenaFuseUnbalanced/2 44.1ns ± 2% 60.4ns ± 1% +37.05% (p=0.000 n=18+19)
BM_ArenaFuseUnbalanced/8 370ns ± 2% 500ns ± 1% +35.12% (p=0.000 n=19+20)
BM_ArenaFuseUnbalanced/64 3.52µs ± 1% 4.71µs ± 1% +33.80% (p=0.000 n=18+19)
BM_ArenaFuseUnbalanced/128 7.20µs ± 1% 9.72µs ± 2% +34.93% (p=0.000 n=16+19)
BM_ArenaFuseBalanced/2 44.4ns ± 2% 61.4ns ± 1% +38.23% (p=0.000 n=20+17)
BM_ArenaFuseBalanced/8 373ns ± 2% 509ns ± 1% +36.57% (p=0.000 n=19+17)
BM_ArenaFuseBalanced/64 3.55µs ± 2% 4.79µs ± 1% +34.80% (p=0.000 n=19+19)
BM_ArenaFuseBalanced/128 7.26µs ± 1% 9.76µs ± 1% +34.45% (p=0.000 n=17+19)
BM_LoadAdsDescriptor_Upb<NoLayout> 5.66ms ± 1% 5.69ms ± 1% +0.57% (p=0.013 n=18+20)
BM_LoadAdsDescriptor_Upb<WithLayout> 6.30ms ± 1% 6.36ms ± 1% +0.90% (p=0.000 n=19+18)
BM_LoadAdsDescriptor_Proto2<NoLayout> 12.1ms ± 1% 12.1ms ± 1% ~ (p=0.118 n=18+18)
BM_LoadAdsDescriptor_Proto2<WithLayout> 12.2ms ± 1% 12.3ms ± 1% +0.50% (p=0.006 n=18+18)
BM_Parse_Upb_FileDesc<UseArena, Copy> 12.7µs ± 1% 12.7µs ± 1% ~ (p=0.194 n=20+19)
BM_Parse_Upb_FileDesc<UseArena, Alias> 11.6µs ± 1% 11.6µs ± 1% ~ (p=0.192 n=20+20)
BM_Parse_Upb_FileDesc<InitBlock, Copy> 12.5µs ± 1% 12.5µs ± 0% ~ (p=0.750 n=18+14)
BM_Parse_Upb_FileDesc<InitBlock, Alias> 11.4µs ± 1% 11.3µs ± 1% -0.34% (p=0.046 n=19+19)
BM_Parse_Proto2<FileDesc, NoArena, Copy> 25.4µs ± 1% 25.7µs ± 2% +1.37% (p=0.000 n=18+18)
BM_Parse_Proto2<FileDesc, UseArena, Copy> 12.1µs ± 2% 12.1µs ± 1% ~ (p=0.143 n=18+18)
BM_Parse_Proto2<FileDesc, InitBlock, Copy> 11.9µs ± 3% 11.9µs ± 1% ~ (p=0.076 n=17+19)
BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 13.2µs ± 1% 13.2µs ± 1% ~ (p=0.053 n=19+19)
BM_SerializeDescriptor_Proto2 5.97µs ± 4% 5.90µs ± 4% ~ (p=0.093 n=17+19)
BM_SerializeDescriptor_Upb 10.4µs ± 1% 10.4µs ± 1% ~ (p=0.909 n=17+18)
name old time/op new time/op delta
BM_ArenaOneAlloc 18.7ns ± 2% 18.6ns ± 0% ~ (p=0.607 n=18+17)
BM_ArenaInitialBlockOneAlloc 6.29ns ± 1% 5.74ns ± 1% -8.71% (p=0.000 n=17+19)
BM_ArenaFuseUnbalanced/2 44.1ns ± 1% 60.6ns ± 1% +37.21% (p=0.000 n=17+19)
BM_ArenaFuseUnbalanced/8 371ns ± 2% 500ns ± 1% +35.02% (p=0.000 n=19+16)
BM_ArenaFuseUnbalanced/64 3.53µs ± 1% 4.72µs ± 1% +33.85% (p=0.000 n=18+19)
BM_ArenaFuseUnbalanced/128 7.22µs ± 1% 9.73µs ± 2% +34.87% (p=0.000 n=16+19)
BM_ArenaFuseBalanced/2 44.5ns ± 2% 61.5ns ± 1% +38.22% (p=0.000 n=20+17)
BM_ArenaFuseBalanced/8 373ns ± 2% 510ns ± 1% +36.58% (p=0.000 n=19+16)
BM_ArenaFuseBalanced/64 3.56µs ± 2% 4.80µs ± 1% +34.87% (p=0.000 n=19+19)
BM_ArenaFuseBalanced/128 7.27µs ± 1% 9.77µs ± 1% +34.40% (p=0.000 n=17+19)
BM_LoadAdsDescriptor_Upb<NoLayout> 5.67ms ± 1% 5.71ms ± 1% +0.60% (p=0.011 n=18+20)
BM_LoadAdsDescriptor_Upb<WithLayout> 6.32ms ± 1% 6.37ms ± 1% +0.87% (p=0.000 n=19+18)
BM_LoadAdsDescriptor_Proto2<NoLayout> 12.1ms ± 1% 12.2ms ± 1% ~ (p=0.126 n=18+19)
BM_LoadAdsDescriptor_Proto2<WithLayout> 12.2ms ± 1% 12.3ms ± 1% +0.51% (p=0.002 n=18+18)
BM_Parse_Upb_FileDesc<UseArena, Copy> 12.7µs ± 1% 12.7µs ± 1% ~ (p=0.149 n=20+19)
BM_Parse_Upb_FileDesc<UseArena, Alias> 11.6µs ± 1% 11.6µs ± 1% ~ (p=0.211 n=20+20)
BM_Parse_Upb_FileDesc<InitBlock, Copy> 12.5µs ± 1% 12.5µs ± 1% ~ (p=0.986 n=18+15)
BM_Parse_Upb_FileDesc<InitBlock, Alias> 11.4µs ± 1% 11.3µs ± 1% ~ (p=0.081 n=19+18)
BM_Parse_Proto2<FileDesc, NoArena, Copy> 25.4µs ± 1% 25.8µs ± 2% +1.41% (p=0.000 n=18+18)
BM_Parse_Proto2<FileDesc, UseArena, Copy> 12.1µs ± 2% 12.1µs ± 1% ~ (p=0.558 n=19+18)
BM_Parse_Proto2<FileDesc, InitBlock, Copy> 12.0µs ± 3% 11.9µs ± 1% ~ (p=0.165 n=17+19)
BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 13.2µs ± 1% 13.2µs ± 1% ~ (p=0.070 n=19+19)
BM_SerializeDescriptor_Proto2 5.98µs ± 4% 5.92µs ± 3% ~ (p=0.138 n=17+19)
BM_SerializeDescriptor_Upb 10.4µs ± 1% 10.4µs ± 1% ~ (p=0.858 n=17+18)
```
PiperOrigin-RevId: 518573683
2 years ago
|
|
|
#ifdef __GNUC__
|
|
|
|
#define UPB_USE_C11_ATOMICS
|
Changed Arena representation so that fusing links arenas together instead of blocks.
Previously when fusing, we would concatenate all blocks into a single list that lived in the arena root. From then on, all arenas would add their blocks to this single unified list.
After this CL, arenas keep their distinct list of blocks even after being fused. Instead of unifying the block list, fuse now puts the arenas themselves into a list, so all arenas in the fused group can be iterated over at any time.
This design makes it easier to keep each individual arena thread-compatible, because fuse and free are now the only mutating operations that touch state that is shared with the entire group. Read-only operations like `SpaceAllocated()` also iterate the list of arenas, but in a read-only fashion.
(Note: we need tests for SpaceAllocated(), both single-threaded for correctness and multi-threaded for resilience to crashes and data races).
Performance of fuse regresses by 5-20%. This is somewhat expected as we are performing more atomic operations during a fuse.
```
name old cpu/op new cpu/op delta
BM_ArenaOneAlloc 18.4ns ± 6% 18.7ns ± 4% +2.00% (p=0.016 n=18+18)
BM_ArenaInitialBlockOneAlloc 5.50ns ± 4% 6.57ns ± 4% +19.42% (p=0.000 n=16+17)
BM_ArenaFuseUnbalanced/2 59.3ns ±10% 68.7ns ± 4% +15.85% (p=0.000 n=19+19)
BM_ArenaFuseUnbalanced/8 479ns ± 5% 540ns ± 8% +12.57% (p=0.000 n=18+19)
BM_ArenaFuseUnbalanced/64 4.50µs ± 4% 4.93µs ± 8% +9.59% (p=0.000 n=17+17)
BM_ArenaFuseUnbalanced/128 9.24µs ± 3% 9.96µs ± 3% +7.81% (p=0.000 n=17+17)
BM_ArenaFuseBalanced/2 63.3ns ±18% 71.0ns ± 4% +12.14% (p=0.000 n=19+18)
BM_ArenaFuseBalanced/8 484ns ± 9% 543ns ±10% +12.11% (p=0.000 n=17+16)
BM_ArenaFuseBalanced/64 4.50µs ± 6% 4.94µs ± 4% +9.62% (p=0.000 n=19+17)
BM_ArenaFuseBalanced/128 9.20µs ± 4% 9.95µs ± 4% +8.12% (p=0.000 n=16+19)
BM_LoadAdsDescriptor_Upb<NoLayout> 5.50ms ± 8% 5.69ms ±17% ~ (p=0.189 n=18+19)
BM_LoadAdsDescriptor_Upb<WithLayout> 6.10ms ± 5% 6.05ms ± 4% ~ (p=0.258 n=17+18)
BM_LoadAdsDescriptor_Proto2<NoLayout> 11.9ms ±15% 11.6ms ± 5% ~ (p=0.589 n=19+16)
BM_LoadAdsDescriptor_Proto2<WithLayout> 11.8ms ± 5% 12.4ms ±17% ~ (p=0.604 n=16+20)
BM_Parse_Upb_FileDesc<UseArena, Copy> 12.1µs ± 8% 12.1µs ± 4% ~ (p=1.000 n=18+18)
BM_Parse_Upb_FileDesc<UseArena, Alias> 11.8µs ±17% 11.1µs ± 4% ~ (p=0.104 n=20+17)
BM_Parse_Upb_FileDesc<InitBlock, Copy> 12.0µs ± 5% 11.9µs ± 4% ~ (p=0.134 n=18+19)
BM_Parse_Upb_FileDesc<InitBlock, Alias> 10.9µs ± 7% 11.0µs ± 4% ~ (p=0.195 n=17+18)
BM_Parse_Proto2<FileDesc, NoArena, Copy> 24.2µs ± 4% 24.4µs ± 7% ~ (p=0.767 n=18+18)
BM_Parse_Proto2<FileDesc, UseArena, Copy> 11.6µs ± 5% 11.6µs ± 4% ~ (p=0.621 n=18+16)
BM_Parse_Proto2<FileDesc, InitBlock, Copy> 11.3µs ± 3% 11.3µs ± 3% ~ (p=0.743 n=18+18)
BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 12.7µs ± 8% 12.7µs ± 4% ~ (p=0.988 n=18+19)
BM_SerializeDescriptor_Proto2 5.77µs ± 5% 5.71µs ± 5% ~ (p=0.433 n=17+17)
BM_SerializeDescriptor_Upb 10.0µs ± 5% 10.1µs ± 7% ~ (p=0.102 n=19+16)
name old time/op new time/op delta
BM_ArenaOneAlloc 18.4ns ± 6% 18.8ns ± 4% +1.98% (p=0.019 n=18+18)
BM_ArenaInitialBlockOneAlloc 5.51ns ± 4% 6.58ns ± 4% +19.42% (p=0.000 n=16+17)
BM_ArenaFuseUnbalanced/2 59.5ns ±10% 68.9ns ± 4% +15.83% (p=0.000 n=19+19)
BM_ArenaFuseUnbalanced/8 481ns ± 5% 541ns ± 8% +12.54% (p=0.000 n=18+19)
BM_ArenaFuseUnbalanced/64 4.51µs ± 4% 4.94µs ± 8% +9.53% (p=0.000 n=17+17)
BM_ArenaFuseUnbalanced/128 9.26µs ± 3% 9.98µs ± 3% +7.79% (p=0.000 n=17+17)
BM_ArenaFuseBalanced/2 63.5ns ±19% 71.1ns ± 3% +12.07% (p=0.000 n=19+18)
BM_ArenaFuseBalanced/8 485ns ± 9% 551ns ±20% +13.47% (p=0.000 n=17+17)
BM_ArenaFuseBalanced/64 4.51µs ± 6% 4.95µs ± 4% +9.62% (p=0.000 n=19+17)
BM_ArenaFuseBalanced/128 9.22µs ± 4% 9.97µs ± 4% +8.12% (p=0.000 n=16+19)
BM_LoadAdsDescriptor_Upb<NoLayout> 5.52ms ± 8% 5.72ms ±18% ~ (p=0.199 n=18+19)
BM_LoadAdsDescriptor_Upb<WithLayout> 6.12ms ± 5% 6.07ms ± 4% ~ (p=0.273 n=17+18)
BM_LoadAdsDescriptor_Proto2<NoLayout> 11.9ms ±15% 11.6ms ± 5% ~ (p=0.589 n=19+16)
BM_LoadAdsDescriptor_Proto2<WithLayout> 11.9ms ± 5% 12.5ms ±18% ~ (p=0.582 n=16+20)
BM_Parse_Upb_FileDesc<UseArena, Copy> 12.2µs ± 8% 12.1µs ± 3% ~ (p=0.963 n=18+18)
BM_Parse_Upb_FileDesc<UseArena, Alias> 11.8µs ±17% 11.1µs ± 4% ~ (p=0.104 n=20+17)
BM_Parse_Upb_FileDesc<InitBlock, Copy> 12.0µs ± 5% 11.9µs ± 4% ~ (p=0.126 n=18+19)
BM_Parse_Upb_FileDesc<InitBlock, Alias> 11.0µs ± 6% 11.1µs ± 4% ~ (p=0.195 n=17+18)
BM_Parse_Proto2<FileDesc, NoArena, Copy> 24.3µs ± 4% 24.5µs ± 6% ~ (p=0.743 n=18+18)
BM_Parse_Proto2<FileDesc, UseArena, Copy> 11.7µs ± 5% 11.6µs ± 4% ~ (p=0.574 n=18+16)
BM_Parse_Proto2<FileDesc, InitBlock, Copy> 11.3µs ± 3% 11.3µs ± 3% ~ (p=0.743 n=18+18)
BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 12.7µs ± 8% 12.7µs ± 4% ~ (p=0.988 n=18+19)
BM_SerializeDescriptor_Proto2 5.78µs ± 5% 5.73µs ± 5% ~ (p=0.357 n=17+17)
BM_SerializeDescriptor_Upb 10.0µs ± 5% 10.1µs ± 7% ~ (p=0.117 n=19+16)
name old allocs/op new allocs/op delta
BM_ArenaOneAlloc 1.00 ± 0% 1.00 ± 0% ~ (all samples are equal)
BM_ArenaFuseUnbalanced/2 2.00 ± 0% 2.00 ± 0% ~ (all samples are equal)
BM_ArenaFuseUnbalanced/8 8.00 ± 0% 8.00 ± 0% ~ (all samples are equal)
BM_ArenaFuseUnbalanced/64 64.0 ± 0% 64.0 ± 0% ~ (all samples are equal)
BM_ArenaFuseUnbalanced/128 128 ± 0% 128 ± 0% ~ (all samples are equal)
BM_ArenaFuseBalanced/2 2.00 ± 0% 2.00 ± 0% ~ (all samples are equal)
BM_ArenaFuseBalanced/8 8.00 ± 0% 8.00 ± 0% ~ (all samples are equal)
BM_ArenaFuseBalanced/64 64.0 ± 0% 64.0 ± 0% ~ (all samples are equal)
BM_ArenaFuseBalanced/128 128 ± 0% 128 ± 0% ~ (all samples are equal)
BM_LoadAdsDescriptor_Upb<NoLayout> 6.08k ± 0% 6.05k ± 0% -0.54% (p=0.000 n=20+20)
BM_LoadAdsDescriptor_Upb<WithLayout> 6.39k ± 0% 6.36k ± 0% -0.55% (p=0.000 n=20+20)
BM_LoadAdsDescriptor_Proto2<NoLayout> 83.4k ± 0% 83.4k ± 0% ~ (p=0.800 n=20+20)
BM_LoadAdsDescriptor_Proto2<WithLayout> 84.4k ± 0% 84.4k ± 0% ~ (p=0.752 n=20+20)
BM_Parse_Upb_FileDesc<UseArena, Copy> 7.00 ± 0% 7.00 ± 0% ~ (all samples are equal)
BM_Parse_Upb_FileDesc<UseArena, Alias> 7.00 ± 0% 7.00 ± 0% ~ (all samples are equal)
BM_Parse_Proto2<FileDesc, NoArena, Copy> 765 ± 0% 765 ± 0% ~ (all samples are equal)
BM_Parse_Proto2<FileDesc, UseArena, Copy> 8.00 ± 0% 8.00 ± 0% ~ (all samples are equal)
name old peak-mem(Bytes)/op new peak-mem(Bytes)/op delta
BM_ArenaOneAlloc 336 ± 0% 336 ± 0% ~ (all samples are equal)
BM_ArenaFuseUnbalanced/2 672 ± 0% 672 ± 0% ~ (all samples are equal)
BM_ArenaFuseUnbalanced/8 2.69k ± 0% 2.69k ± 0% ~ (all samples are equal)
BM_ArenaFuseUnbalanced/64 21.5k ± 0% 21.5k ± 0% ~ (all samples are equal)
BM_ArenaFuseUnbalanced/128 43.0k ± 0% 43.0k ± 0% ~ (all samples are equal)
BM_ArenaFuseBalanced/2 672 ± 0% 672 ± 0% ~ (all samples are equal)
BM_ArenaFuseBalanced/8 2.69k ± 0% 2.69k ± 0% ~ (all samples are equal)
BM_ArenaFuseBalanced/64 21.5k ± 0% 21.5k ± 0% ~ (all samples are equal)
BM_ArenaFuseBalanced/128 43.0k ± 0% 43.0k ± 0% ~ (all samples are equal)
BM_LoadAdsDescriptor_Upb<NoLayout> 9.89M ± 0% 9.95M ± 0% +0.65% (p=0.000 n=20+20)
BM_LoadAdsDescriptor_Upb<WithLayout> 9.95M ± 0% 10.02M ± 0% +0.70% (p=0.000 n=20+20)
BM_LoadAdsDescriptor_Proto2<NoLayout> 6.62M ± 0% 6.62M ± 0% ~ (p=0.800 n=20+20)
BM_LoadAdsDescriptor_Proto2<WithLayout> 6.66M ± 0% 6.66M ± 0% ~ (p=0.752 n=20+20)
BM_Parse_Upb_FileDesc<UseArena, Copy> 36.5k ± 0% 36.5k ± 0% ~ (all samples are equal)
BM_Parse_Upb_FileDesc<UseArena, Alias> 36.5k ± 0% 36.5k ± 0% ~ (all samples are equal)
BM_Parse_Proto2<FileDesc, NoArena, Copy> 35.8k ± 0% 35.8k ± 0% ~ (all samples are equal)
BM_Parse_Proto2<FileDesc, UseArena, Copy> 65.3k ± 0% 65.3k ± 0% ~ (all samples are equal)
name old speed new speed delta
BM_LoadAdsDescriptor_Upb<NoLayout> 138MB/s ± 7% 132MB/s ±15% ~ (p=0.126 n=18+20)
BM_LoadAdsDescriptor_Upb<WithLayout> 124MB/s ± 5% 125MB/s ± 4% ~ (p=0.258 n=17+18)
BM_LoadAdsDescriptor_Proto2<NoLayout> 63.9MB/s ±13% 65.2MB/s ± 5% ~ (p=0.589 n=19+16)
BM_LoadAdsDescriptor_Proto2<WithLayout> 64.0MB/s ± 5% 61.3MB/s ±15% ~ (p=0.604 n=16+20)
BM_Parse_Upb_FileDesc<UseArena, Copy> 620MB/s ± 8% 622MB/s ± 4% ~ (p=1.000 n=18+18)
BM_Parse_Upb_FileDesc<UseArena, Alias> 644MB/s ±15% 679MB/s ± 4% ~ (p=0.104 n=20+17)
BM_Parse_Upb_FileDesc<InitBlock, Copy> 627MB/s ± 4% 633MB/s ± 4% ~ (p=0.134 n=18+19)
BM_Parse_Upb_FileDesc<InitBlock, Alias> 688MB/s ± 6% 682MB/s ± 4% ~ (p=0.195 n=17+18)
BM_Parse_Proto2<FileDesc, NoArena, Copy> 310MB/s ± 4% 309MB/s ± 6% ~ (p=0.767 n=18+18)
BM_Parse_Proto2<FileDesc, UseArena, Copy> 646MB/s ± 4% 649MB/s ± 4% ~ (p=0.621 n=18+16)
BM_Parse_Proto2<FileDesc, InitBlock, Copy> 666MB/s ± 3% 666MB/s ± 3% ~ (p=0.743 n=18+18)
BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 592MB/s ± 7% 593MB/s ± 4% ~ (p=0.988 n=18+19)
BM_SerializeDescriptor_Proto2 1.30GB/s ± 5% 1.32GB/s ± 5% ~ (p=0.433 n=17+17)
BM_SerializeDescriptor_Upb 756MB/s ± 5% 745MB/s ± 6% ~ (p=0.102 n=19+16)
```
PiperOrigin-RevId: 520144430
2 years ago
|
|
|
#define UPB_ATOMIC(T) _Atomic(T)
|
Allow for fuse/free races in `upb_Arena`.
Implementation is by kfm@, I only added the portability code around it.
`upb_Arena` was designed to be only thread-compatible. However, fusing of arenas muddies the waters somewhat, because two distinct `upb_Arena` objects will end up sharing state when fused. This causes a `upb_Arena_Free(a)` to interfere with `upb_Arena_Fuse(b, c)` if `a` and `b` were previously fused.
It turns out that we can use atomics to fix this with about a 35% regression in fuse performance (see below). Arena create+free does not regress, thanks to special-case logic in Free().
`upb_Arena` is still a thread-compatible type, and it is still never safe to call `upb_Arena_xxx(a)` and `upb_Arena_yyy(a)` in parallel. However you can at least now call `upb_Arena_Free(a)` and `upb_Arena_Fuse(b, c)` in parallel, even if `a` and `b` were previously fused.
Note that `upb_Arena_Fuse(a, b)` and `upb_Arena_Fuse(c, d)` is still not allowed if `b` and `c` were previously fused. In practice this means that fuses must still be single-threaded within a single fused group.
Performance results:
```
name old cpu/op new cpu/op delta
BM_ArenaOneAlloc 18.6ns ± 1% 18.6ns ± 1% ~ (p=0.726 n=18+17)
BM_ArenaInitialBlockOneAlloc 6.28ns ± 1% 5.73ns ± 1% -8.68% (p=0.000 n=17+20)
BM_ArenaFuseUnbalanced/2 44.1ns ± 2% 60.4ns ± 1% +37.05% (p=0.000 n=18+19)
BM_ArenaFuseUnbalanced/8 370ns ± 2% 500ns ± 1% +35.12% (p=0.000 n=19+20)
BM_ArenaFuseUnbalanced/64 3.52µs ± 1% 4.71µs ± 1% +33.80% (p=0.000 n=18+19)
BM_ArenaFuseUnbalanced/128 7.20µs ± 1% 9.72µs ± 2% +34.93% (p=0.000 n=16+19)
BM_ArenaFuseBalanced/2 44.4ns ± 2% 61.4ns ± 1% +38.23% (p=0.000 n=20+17)
BM_ArenaFuseBalanced/8 373ns ± 2% 509ns ± 1% +36.57% (p=0.000 n=19+17)
BM_ArenaFuseBalanced/64 3.55µs ± 2% 4.79µs ± 1% +34.80% (p=0.000 n=19+19)
BM_ArenaFuseBalanced/128 7.26µs ± 1% 9.76µs ± 1% +34.45% (p=0.000 n=17+19)
BM_LoadAdsDescriptor_Upb<NoLayout> 5.66ms ± 1% 5.69ms ± 1% +0.57% (p=0.013 n=18+20)
BM_LoadAdsDescriptor_Upb<WithLayout> 6.30ms ± 1% 6.36ms ± 1% +0.90% (p=0.000 n=19+18)
BM_LoadAdsDescriptor_Proto2<NoLayout> 12.1ms ± 1% 12.1ms ± 1% ~ (p=0.118 n=18+18)
BM_LoadAdsDescriptor_Proto2<WithLayout> 12.2ms ± 1% 12.3ms ± 1% +0.50% (p=0.006 n=18+18)
BM_Parse_Upb_FileDesc<UseArena, Copy> 12.7µs ± 1% 12.7µs ± 1% ~ (p=0.194 n=20+19)
BM_Parse_Upb_FileDesc<UseArena, Alias> 11.6µs ± 1% 11.6µs ± 1% ~ (p=0.192 n=20+20)
BM_Parse_Upb_FileDesc<InitBlock, Copy> 12.5µs ± 1% 12.5µs ± 0% ~ (p=0.750 n=18+14)
BM_Parse_Upb_FileDesc<InitBlock, Alias> 11.4µs ± 1% 11.3µs ± 1% -0.34% (p=0.046 n=19+19)
BM_Parse_Proto2<FileDesc, NoArena, Copy> 25.4µs ± 1% 25.7µs ± 2% +1.37% (p=0.000 n=18+18)
BM_Parse_Proto2<FileDesc, UseArena, Copy> 12.1µs ± 2% 12.1µs ± 1% ~ (p=0.143 n=18+18)
BM_Parse_Proto2<FileDesc, InitBlock, Copy> 11.9µs ± 3% 11.9µs ± 1% ~ (p=0.076 n=17+19)
BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 13.2µs ± 1% 13.2µs ± 1% ~ (p=0.053 n=19+19)
BM_SerializeDescriptor_Proto2 5.97µs ± 4% 5.90µs ± 4% ~ (p=0.093 n=17+19)
BM_SerializeDescriptor_Upb 10.4µs ± 1% 10.4µs ± 1% ~ (p=0.909 n=17+18)
name old time/op new time/op delta
BM_ArenaOneAlloc 18.7ns ± 2% 18.6ns ± 0% ~ (p=0.607 n=18+17)
BM_ArenaInitialBlockOneAlloc 6.29ns ± 1% 5.74ns ± 1% -8.71% (p=0.000 n=17+19)
BM_ArenaFuseUnbalanced/2 44.1ns ± 1% 60.6ns ± 1% +37.21% (p=0.000 n=17+19)
BM_ArenaFuseUnbalanced/8 371ns ± 2% 500ns ± 1% +35.02% (p=0.000 n=19+16)
BM_ArenaFuseUnbalanced/64 3.53µs ± 1% 4.72µs ± 1% +33.85% (p=0.000 n=18+19)
BM_ArenaFuseUnbalanced/128 7.22µs ± 1% 9.73µs ± 2% +34.87% (p=0.000 n=16+19)
BM_ArenaFuseBalanced/2 44.5ns ± 2% 61.5ns ± 1% +38.22% (p=0.000 n=20+17)
BM_ArenaFuseBalanced/8 373ns ± 2% 510ns ± 1% +36.58% (p=0.000 n=19+16)
BM_ArenaFuseBalanced/64 3.56µs ± 2% 4.80µs ± 1% +34.87% (p=0.000 n=19+19)
BM_ArenaFuseBalanced/128 7.27µs ± 1% 9.77µs ± 1% +34.40% (p=0.000 n=17+19)
BM_LoadAdsDescriptor_Upb<NoLayout> 5.67ms ± 1% 5.71ms ± 1% +0.60% (p=0.011 n=18+20)
BM_LoadAdsDescriptor_Upb<WithLayout> 6.32ms ± 1% 6.37ms ± 1% +0.87% (p=0.000 n=19+18)
BM_LoadAdsDescriptor_Proto2<NoLayout> 12.1ms ± 1% 12.2ms ± 1% ~ (p=0.126 n=18+19)
BM_LoadAdsDescriptor_Proto2<WithLayout> 12.2ms ± 1% 12.3ms ± 1% +0.51% (p=0.002 n=18+18)
BM_Parse_Upb_FileDesc<UseArena, Copy> 12.7µs ± 1% 12.7µs ± 1% ~ (p=0.149 n=20+19)
BM_Parse_Upb_FileDesc<UseArena, Alias> 11.6µs ± 1% 11.6µs ± 1% ~ (p=0.211 n=20+20)
BM_Parse_Upb_FileDesc<InitBlock, Copy> 12.5µs ± 1% 12.5µs ± 1% ~ (p=0.986 n=18+15)
BM_Parse_Upb_FileDesc<InitBlock, Alias> 11.4µs ± 1% 11.3µs ± 1% ~ (p=0.081 n=19+18)
BM_Parse_Proto2<FileDesc, NoArena, Copy> 25.4µs ± 1% 25.8µs ± 2% +1.41% (p=0.000 n=18+18)
BM_Parse_Proto2<FileDesc, UseArena, Copy> 12.1µs ± 2% 12.1µs ± 1% ~ (p=0.558 n=19+18)
BM_Parse_Proto2<FileDesc, InitBlock, Copy> 12.0µs ± 3% 11.9µs ± 1% ~ (p=0.165 n=17+19)
BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 13.2µs ± 1% 13.2µs ± 1% ~ (p=0.070 n=19+19)
BM_SerializeDescriptor_Proto2 5.98µs ± 4% 5.92µs ± 3% ~ (p=0.138 n=17+19)
BM_SerializeDescriptor_Upb 10.4µs ± 1% 10.4µs ± 1% ~ (p=0.858 n=17+18)
```
PiperOrigin-RevId: 518573683
2 years ago
|
|
|
#else
|
Changed Arena representation so that fusing links arenas together instead of blocks.
Previously when fusing, we would concatenate all blocks into a single list that lived in the arena root. From then on, all arenas would add their blocks to this single unified list.
After this CL, arenas keep their distinct list of blocks even after being fused. Instead of unifying the block list, fuse now puts the arenas themselves into a list, so all arenas in the fused group can be iterated over at any time.
This design makes it easier to keep each individual arena thread-compatible, because fuse and free are now the only mutating operations that touch state that is shared with the entire group. Read-only operations like `SpaceAllocated()` also iterate the list of arenas, but in a read-only fashion.
(Note: we need tests for SpaceAllocated(), both single-threaded for correctness and multi-threaded for resilience to crashes and data races).
Performance of fuse regresses by 5-20%. This is somewhat expected as we are performing more atomic operations during a fuse.
```
name old cpu/op new cpu/op delta
BM_ArenaOneAlloc 18.4ns ± 6% 18.7ns ± 4% +2.00% (p=0.016 n=18+18)
BM_ArenaInitialBlockOneAlloc 5.50ns ± 4% 6.57ns ± 4% +19.42% (p=0.000 n=16+17)
BM_ArenaFuseUnbalanced/2 59.3ns ±10% 68.7ns ± 4% +15.85% (p=0.000 n=19+19)
BM_ArenaFuseUnbalanced/8 479ns ± 5% 540ns ± 8% +12.57% (p=0.000 n=18+19)
BM_ArenaFuseUnbalanced/64 4.50µs ± 4% 4.93µs ± 8% +9.59% (p=0.000 n=17+17)
BM_ArenaFuseUnbalanced/128 9.24µs ± 3% 9.96µs ± 3% +7.81% (p=0.000 n=17+17)
BM_ArenaFuseBalanced/2 63.3ns ±18% 71.0ns ± 4% +12.14% (p=0.000 n=19+18)
BM_ArenaFuseBalanced/8 484ns ± 9% 543ns ±10% +12.11% (p=0.000 n=17+16)
BM_ArenaFuseBalanced/64 4.50µs ± 6% 4.94µs ± 4% +9.62% (p=0.000 n=19+17)
BM_ArenaFuseBalanced/128 9.20µs ± 4% 9.95µs ± 4% +8.12% (p=0.000 n=16+19)
BM_LoadAdsDescriptor_Upb<NoLayout> 5.50ms ± 8% 5.69ms ±17% ~ (p=0.189 n=18+19)
BM_LoadAdsDescriptor_Upb<WithLayout> 6.10ms ± 5% 6.05ms ± 4% ~ (p=0.258 n=17+18)
BM_LoadAdsDescriptor_Proto2<NoLayout> 11.9ms ±15% 11.6ms ± 5% ~ (p=0.589 n=19+16)
BM_LoadAdsDescriptor_Proto2<WithLayout> 11.8ms ± 5% 12.4ms ±17% ~ (p=0.604 n=16+20)
BM_Parse_Upb_FileDesc<UseArena, Copy> 12.1µs ± 8% 12.1µs ± 4% ~ (p=1.000 n=18+18)
BM_Parse_Upb_FileDesc<UseArena, Alias> 11.8µs ±17% 11.1µs ± 4% ~ (p=0.104 n=20+17)
BM_Parse_Upb_FileDesc<InitBlock, Copy> 12.0µs ± 5% 11.9µs ± 4% ~ (p=0.134 n=18+19)
BM_Parse_Upb_FileDesc<InitBlock, Alias> 10.9µs ± 7% 11.0µs ± 4% ~ (p=0.195 n=17+18)
BM_Parse_Proto2<FileDesc, NoArena, Copy> 24.2µs ± 4% 24.4µs ± 7% ~ (p=0.767 n=18+18)
BM_Parse_Proto2<FileDesc, UseArena, Copy> 11.6µs ± 5% 11.6µs ± 4% ~ (p=0.621 n=18+16)
BM_Parse_Proto2<FileDesc, InitBlock, Copy> 11.3µs ± 3% 11.3µs ± 3% ~ (p=0.743 n=18+18)
BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 12.7µs ± 8% 12.7µs ± 4% ~ (p=0.988 n=18+19)
BM_SerializeDescriptor_Proto2 5.77µs ± 5% 5.71µs ± 5% ~ (p=0.433 n=17+17)
BM_SerializeDescriptor_Upb 10.0µs ± 5% 10.1µs ± 7% ~ (p=0.102 n=19+16)
name old time/op new time/op delta
BM_ArenaOneAlloc 18.4ns ± 6% 18.8ns ± 4% +1.98% (p=0.019 n=18+18)
BM_ArenaInitialBlockOneAlloc 5.51ns ± 4% 6.58ns ± 4% +19.42% (p=0.000 n=16+17)
BM_ArenaFuseUnbalanced/2 59.5ns ±10% 68.9ns ± 4% +15.83% (p=0.000 n=19+19)
BM_ArenaFuseUnbalanced/8 481ns ± 5% 541ns ± 8% +12.54% (p=0.000 n=18+19)
BM_ArenaFuseUnbalanced/64 4.51µs ± 4% 4.94µs ± 8% +9.53% (p=0.000 n=17+17)
BM_ArenaFuseUnbalanced/128 9.26µs ± 3% 9.98µs ± 3% +7.79% (p=0.000 n=17+17)
BM_ArenaFuseBalanced/2 63.5ns ±19% 71.1ns ± 3% +12.07% (p=0.000 n=19+18)
BM_ArenaFuseBalanced/8 485ns ± 9% 551ns ±20% +13.47% (p=0.000 n=17+17)
BM_ArenaFuseBalanced/64 4.51µs ± 6% 4.95µs ± 4% +9.62% (p=0.000 n=19+17)
BM_ArenaFuseBalanced/128 9.22µs ± 4% 9.97µs ± 4% +8.12% (p=0.000 n=16+19)
BM_LoadAdsDescriptor_Upb<NoLayout> 5.52ms ± 8% 5.72ms ±18% ~ (p=0.199 n=18+19)
BM_LoadAdsDescriptor_Upb<WithLayout> 6.12ms ± 5% 6.07ms ± 4% ~ (p=0.273 n=17+18)
BM_LoadAdsDescriptor_Proto2<NoLayout> 11.9ms ±15% 11.6ms ± 5% ~ (p=0.589 n=19+16)
BM_LoadAdsDescriptor_Proto2<WithLayout> 11.9ms ± 5% 12.5ms ±18% ~ (p=0.582 n=16+20)
BM_Parse_Upb_FileDesc<UseArena, Copy> 12.2µs ± 8% 12.1µs ± 3% ~ (p=0.963 n=18+18)
BM_Parse_Upb_FileDesc<UseArena, Alias> 11.8µs ±17% 11.1µs ± 4% ~ (p=0.104 n=20+17)
BM_Parse_Upb_FileDesc<InitBlock, Copy> 12.0µs ± 5% 11.9µs ± 4% ~ (p=0.126 n=18+19)
BM_Parse_Upb_FileDesc<InitBlock, Alias> 11.0µs ± 6% 11.1µs ± 4% ~ (p=0.195 n=17+18)
BM_Parse_Proto2<FileDesc, NoArena, Copy> 24.3µs ± 4% 24.5µs ± 6% ~ (p=0.743 n=18+18)
BM_Parse_Proto2<FileDesc, UseArena, Copy> 11.7µs ± 5% 11.6µs ± 4% ~ (p=0.574 n=18+16)
BM_Parse_Proto2<FileDesc, InitBlock, Copy> 11.3µs ± 3% 11.3µs ± 3% ~ (p=0.743 n=18+18)
BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 12.7µs ± 8% 12.7µs ± 4% ~ (p=0.988 n=18+19)
BM_SerializeDescriptor_Proto2 5.78µs ± 5% 5.73µs ± 5% ~ (p=0.357 n=17+17)
BM_SerializeDescriptor_Upb 10.0µs ± 5% 10.1µs ± 7% ~ (p=0.117 n=19+16)
name old allocs/op new allocs/op delta
BM_ArenaOneAlloc 1.00 ± 0% 1.00 ± 0% ~ (all samples are equal)
BM_ArenaFuseUnbalanced/2 2.00 ± 0% 2.00 ± 0% ~ (all samples are equal)
BM_ArenaFuseUnbalanced/8 8.00 ± 0% 8.00 ± 0% ~ (all samples are equal)
BM_ArenaFuseUnbalanced/64 64.0 ± 0% 64.0 ± 0% ~ (all samples are equal)
BM_ArenaFuseUnbalanced/128 128 ± 0% 128 ± 0% ~ (all samples are equal)
BM_ArenaFuseBalanced/2 2.00 ± 0% 2.00 ± 0% ~ (all samples are equal)
BM_ArenaFuseBalanced/8 8.00 ± 0% 8.00 ± 0% ~ (all samples are equal)
BM_ArenaFuseBalanced/64 64.0 ± 0% 64.0 ± 0% ~ (all samples are equal)
BM_ArenaFuseBalanced/128 128 ± 0% 128 ± 0% ~ (all samples are equal)
BM_LoadAdsDescriptor_Upb<NoLayout> 6.08k ± 0% 6.05k ± 0% -0.54% (p=0.000 n=20+20)
BM_LoadAdsDescriptor_Upb<WithLayout> 6.39k ± 0% 6.36k ± 0% -0.55% (p=0.000 n=20+20)
BM_LoadAdsDescriptor_Proto2<NoLayout> 83.4k ± 0% 83.4k ± 0% ~ (p=0.800 n=20+20)
BM_LoadAdsDescriptor_Proto2<WithLayout> 84.4k ± 0% 84.4k ± 0% ~ (p=0.752 n=20+20)
BM_Parse_Upb_FileDesc<UseArena, Copy> 7.00 ± 0% 7.00 ± 0% ~ (all samples are equal)
BM_Parse_Upb_FileDesc<UseArena, Alias> 7.00 ± 0% 7.00 ± 0% ~ (all samples are equal)
BM_Parse_Proto2<FileDesc, NoArena, Copy> 765 ± 0% 765 ± 0% ~ (all samples are equal)
BM_Parse_Proto2<FileDesc, UseArena, Copy> 8.00 ± 0% 8.00 ± 0% ~ (all samples are equal)
name old peak-mem(Bytes)/op new peak-mem(Bytes)/op delta
BM_ArenaOneAlloc 336 ± 0% 336 ± 0% ~ (all samples are equal)
BM_ArenaFuseUnbalanced/2 672 ± 0% 672 ± 0% ~ (all samples are equal)
BM_ArenaFuseUnbalanced/8 2.69k ± 0% 2.69k ± 0% ~ (all samples are equal)
BM_ArenaFuseUnbalanced/64 21.5k ± 0% 21.5k ± 0% ~ (all samples are equal)
BM_ArenaFuseUnbalanced/128 43.0k ± 0% 43.0k ± 0% ~ (all samples are equal)
BM_ArenaFuseBalanced/2 672 ± 0% 672 ± 0% ~ (all samples are equal)
BM_ArenaFuseBalanced/8 2.69k ± 0% 2.69k ± 0% ~ (all samples are equal)
BM_ArenaFuseBalanced/64 21.5k ± 0% 21.5k ± 0% ~ (all samples are equal)
BM_ArenaFuseBalanced/128 43.0k ± 0% 43.0k ± 0% ~ (all samples are equal)
BM_LoadAdsDescriptor_Upb<NoLayout> 9.89M ± 0% 9.95M ± 0% +0.65% (p=0.000 n=20+20)
BM_LoadAdsDescriptor_Upb<WithLayout> 9.95M ± 0% 10.02M ± 0% +0.70% (p=0.000 n=20+20)
BM_LoadAdsDescriptor_Proto2<NoLayout> 6.62M ± 0% 6.62M ± 0% ~ (p=0.800 n=20+20)
BM_LoadAdsDescriptor_Proto2<WithLayout> 6.66M ± 0% 6.66M ± 0% ~ (p=0.752 n=20+20)
BM_Parse_Upb_FileDesc<UseArena, Copy> 36.5k ± 0% 36.5k ± 0% ~ (all samples are equal)
BM_Parse_Upb_FileDesc<UseArena, Alias> 36.5k ± 0% 36.5k ± 0% ~ (all samples are equal)
BM_Parse_Proto2<FileDesc, NoArena, Copy> 35.8k ± 0% 35.8k ± 0% ~ (all samples are equal)
BM_Parse_Proto2<FileDesc, UseArena, Copy> 65.3k ± 0% 65.3k ± 0% ~ (all samples are equal)
name old speed new speed delta
BM_LoadAdsDescriptor_Upb<NoLayout> 138MB/s ± 7% 132MB/s ±15% ~ (p=0.126 n=18+20)
BM_LoadAdsDescriptor_Upb<WithLayout> 124MB/s ± 5% 125MB/s ± 4% ~ (p=0.258 n=17+18)
BM_LoadAdsDescriptor_Proto2<NoLayout> 63.9MB/s ±13% 65.2MB/s ± 5% ~ (p=0.589 n=19+16)
BM_LoadAdsDescriptor_Proto2<WithLayout> 64.0MB/s ± 5% 61.3MB/s ±15% ~ (p=0.604 n=16+20)
BM_Parse_Upb_FileDesc<UseArena, Copy> 620MB/s ± 8% 622MB/s ± 4% ~ (p=1.000 n=18+18)
BM_Parse_Upb_FileDesc<UseArena, Alias> 644MB/s ±15% 679MB/s ± 4% ~ (p=0.104 n=20+17)
BM_Parse_Upb_FileDesc<InitBlock, Copy> 627MB/s ± 4% 633MB/s ± 4% ~ (p=0.134 n=18+19)
BM_Parse_Upb_FileDesc<InitBlock, Alias> 688MB/s ± 6% 682MB/s ± 4% ~ (p=0.195 n=17+18)
BM_Parse_Proto2<FileDesc, NoArena, Copy> 310MB/s ± 4% 309MB/s ± 6% ~ (p=0.767 n=18+18)
BM_Parse_Proto2<FileDesc, UseArena, Copy> 646MB/s ± 4% 649MB/s ± 4% ~ (p=0.621 n=18+16)
BM_Parse_Proto2<FileDesc, InitBlock, Copy> 666MB/s ± 3% 666MB/s ± 3% ~ (p=0.743 n=18+18)
BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 592MB/s ± 7% 593MB/s ± 4% ~ (p=0.988 n=18+19)
BM_SerializeDescriptor_Proto2 1.30GB/s ± 5% 1.32GB/s ± 5% ~ (p=0.433 n=17+17)
BM_SerializeDescriptor_Upb 756MB/s ± 5% 745MB/s ± 6% ~ (p=0.102 n=19+16)
```
PiperOrigin-RevId: 520144430
2 years ago
|
|
|
#define UPB_ATOMIC(T) T
|
Allow for fuse/free races in `upb_Arena`.
Implementation is by kfm@, I only added the portability code around it.
`upb_Arena` was designed to be only thread-compatible. However, fusing of arenas muddies the waters somewhat, because two distinct `upb_Arena` objects will end up sharing state when fused. This causes a `upb_Arena_Free(a)` to interfere with `upb_Arena_Fuse(b, c)` if `a` and `b` were previously fused.
It turns out that we can use atomics to fix this with about a 35% regression in fuse performance (see below). Arena create+free does not regress, thanks to special-case logic in Free().
`upb_Arena` is still a thread-compatible type, and it is still never safe to call `upb_Arena_xxx(a)` and `upb_Arena_yyy(a)` in parallel. However you can at least now call `upb_Arena_Free(a)` and `upb_Arena_Fuse(b, c)` in parallel, even if `a` and `b` were previously fused.
Note that `upb_Arena_Fuse(a, b)` and `upb_Arena_Fuse(c, d)` is still not allowed if `b` and `c` were previously fused. In practice this means that fuses must still be single-threaded within a single fused group.
Performance results:
```
name old cpu/op new cpu/op delta
BM_ArenaOneAlloc 18.6ns ± 1% 18.6ns ± 1% ~ (p=0.726 n=18+17)
BM_ArenaInitialBlockOneAlloc 6.28ns ± 1% 5.73ns ± 1% -8.68% (p=0.000 n=17+20)
BM_ArenaFuseUnbalanced/2 44.1ns ± 2% 60.4ns ± 1% +37.05% (p=0.000 n=18+19)
BM_ArenaFuseUnbalanced/8 370ns ± 2% 500ns ± 1% +35.12% (p=0.000 n=19+20)
BM_ArenaFuseUnbalanced/64 3.52µs ± 1% 4.71µs ± 1% +33.80% (p=0.000 n=18+19)
BM_ArenaFuseUnbalanced/128 7.20µs ± 1% 9.72µs ± 2% +34.93% (p=0.000 n=16+19)
BM_ArenaFuseBalanced/2 44.4ns ± 2% 61.4ns ± 1% +38.23% (p=0.000 n=20+17)
BM_ArenaFuseBalanced/8 373ns ± 2% 509ns ± 1% +36.57% (p=0.000 n=19+17)
BM_ArenaFuseBalanced/64 3.55µs ± 2% 4.79µs ± 1% +34.80% (p=0.000 n=19+19)
BM_ArenaFuseBalanced/128 7.26µs ± 1% 9.76µs ± 1% +34.45% (p=0.000 n=17+19)
BM_LoadAdsDescriptor_Upb<NoLayout> 5.66ms ± 1% 5.69ms ± 1% +0.57% (p=0.013 n=18+20)
BM_LoadAdsDescriptor_Upb<WithLayout> 6.30ms ± 1% 6.36ms ± 1% +0.90% (p=0.000 n=19+18)
BM_LoadAdsDescriptor_Proto2<NoLayout> 12.1ms ± 1% 12.1ms ± 1% ~ (p=0.118 n=18+18)
BM_LoadAdsDescriptor_Proto2<WithLayout> 12.2ms ± 1% 12.3ms ± 1% +0.50% (p=0.006 n=18+18)
BM_Parse_Upb_FileDesc<UseArena, Copy> 12.7µs ± 1% 12.7µs ± 1% ~ (p=0.194 n=20+19)
BM_Parse_Upb_FileDesc<UseArena, Alias> 11.6µs ± 1% 11.6µs ± 1% ~ (p=0.192 n=20+20)
BM_Parse_Upb_FileDesc<InitBlock, Copy> 12.5µs ± 1% 12.5µs ± 0% ~ (p=0.750 n=18+14)
BM_Parse_Upb_FileDesc<InitBlock, Alias> 11.4µs ± 1% 11.3µs ± 1% -0.34% (p=0.046 n=19+19)
BM_Parse_Proto2<FileDesc, NoArena, Copy> 25.4µs ± 1% 25.7µs ± 2% +1.37% (p=0.000 n=18+18)
BM_Parse_Proto2<FileDesc, UseArena, Copy> 12.1µs ± 2% 12.1µs ± 1% ~ (p=0.143 n=18+18)
BM_Parse_Proto2<FileDesc, InitBlock, Copy> 11.9µs ± 3% 11.9µs ± 1% ~ (p=0.076 n=17+19)
BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 13.2µs ± 1% 13.2µs ± 1% ~ (p=0.053 n=19+19)
BM_SerializeDescriptor_Proto2 5.97µs ± 4% 5.90µs ± 4% ~ (p=0.093 n=17+19)
BM_SerializeDescriptor_Upb 10.4µs ± 1% 10.4µs ± 1% ~ (p=0.909 n=17+18)
name old time/op new time/op delta
BM_ArenaOneAlloc 18.7ns ± 2% 18.6ns ± 0% ~ (p=0.607 n=18+17)
BM_ArenaInitialBlockOneAlloc 6.29ns ± 1% 5.74ns ± 1% -8.71% (p=0.000 n=17+19)
BM_ArenaFuseUnbalanced/2 44.1ns ± 1% 60.6ns ± 1% +37.21% (p=0.000 n=17+19)
BM_ArenaFuseUnbalanced/8 371ns ± 2% 500ns ± 1% +35.02% (p=0.000 n=19+16)
BM_ArenaFuseUnbalanced/64 3.53µs ± 1% 4.72µs ± 1% +33.85% (p=0.000 n=18+19)
BM_ArenaFuseUnbalanced/128 7.22µs ± 1% 9.73µs ± 2% +34.87% (p=0.000 n=16+19)
BM_ArenaFuseBalanced/2 44.5ns ± 2% 61.5ns ± 1% +38.22% (p=0.000 n=20+17)
BM_ArenaFuseBalanced/8 373ns ± 2% 510ns ± 1% +36.58% (p=0.000 n=19+16)
BM_ArenaFuseBalanced/64 3.56µs ± 2% 4.80µs ± 1% +34.87% (p=0.000 n=19+19)
BM_ArenaFuseBalanced/128 7.27µs ± 1% 9.77µs ± 1% +34.40% (p=0.000 n=17+19)
BM_LoadAdsDescriptor_Upb<NoLayout> 5.67ms ± 1% 5.71ms ± 1% +0.60% (p=0.011 n=18+20)
BM_LoadAdsDescriptor_Upb<WithLayout> 6.32ms ± 1% 6.37ms ± 1% +0.87% (p=0.000 n=19+18)
BM_LoadAdsDescriptor_Proto2<NoLayout> 12.1ms ± 1% 12.2ms ± 1% ~ (p=0.126 n=18+19)
BM_LoadAdsDescriptor_Proto2<WithLayout> 12.2ms ± 1% 12.3ms ± 1% +0.51% (p=0.002 n=18+18)
BM_Parse_Upb_FileDesc<UseArena, Copy> 12.7µs ± 1% 12.7µs ± 1% ~ (p=0.149 n=20+19)
BM_Parse_Upb_FileDesc<UseArena, Alias> 11.6µs ± 1% 11.6µs ± 1% ~ (p=0.211 n=20+20)
BM_Parse_Upb_FileDesc<InitBlock, Copy> 12.5µs ± 1% 12.5µs ± 1% ~ (p=0.986 n=18+15)
BM_Parse_Upb_FileDesc<InitBlock, Alias> 11.4µs ± 1% 11.3µs ± 1% ~ (p=0.081 n=19+18)
BM_Parse_Proto2<FileDesc, NoArena, Copy> 25.4µs ± 1% 25.8µs ± 2% +1.41% (p=0.000 n=18+18)
BM_Parse_Proto2<FileDesc, UseArena, Copy> 12.1µs ± 2% 12.1µs ± 1% ~ (p=0.558 n=19+18)
BM_Parse_Proto2<FileDesc, InitBlock, Copy> 12.0µs ± 3% 11.9µs ± 1% ~ (p=0.165 n=17+19)
BM_Parse_Proto2<FileDescSV, InitBlock, Alias> 13.2µs ± 1% 13.2µs ± 1% ~ (p=0.070 n=19+19)
BM_SerializeDescriptor_Proto2 5.98µs ± 4% 5.92µs ± 3% ~ (p=0.138 n=17+19)
BM_SerializeDescriptor_Upb 10.4µs ± 1% 10.4µs ± 1% ~ (p=0.858 n=17+18)
```
PiperOrigin-RevId: 518573683
2 years ago
|
|
|
#endif
|
|
|
|
|
|
|
|
/* UPB_PTRADD(ptr, ofs): add pointer while avoiding "NULL + 0" UB */
|
|
|
|
#define UPB_PTRADD(ptr, ofs) ((ofs) ? (ptr) + (ofs) : (ptr))
|
|
|
|
|
|
|
|
#define UPB_PRIVATE(x) x##_dont_copy_me__upb_internal_use_only
|
|
|
|
|
|
|
|
/* Configure whether fasttable is switched on or not. *************************/
|
|
|
|
|
|
|
|
#ifdef __has_attribute
|
|
|
|
#define UPB_HAS_ATTRIBUTE(x) __has_attribute(x)
|
|
|
|
#else
|
|
|
|
#define UPB_HAS_ATTRIBUTE(x) 0
|
|
|
|
#endif
|
|
|
|
|
|
|
|
#if UPB_HAS_ATTRIBUTE(musttail)
|
|
|
|
#define UPB_MUSTTAIL __attribute__((musttail))
|
|
|
|
#else
|
|
|
|
#define UPB_MUSTTAIL
|
|
|
|
#endif
|
|
|
|
|
|
|
|
#undef UPB_HAS_ATTRIBUTE
|
|
|
|
|
|
|
|
/* This check is not fully robust: it does not require that we have "musttail"
|
|
|
|
* support available. We need tail calls to avoid consuming arbitrary amounts
|
|
|
|
* of stack space.
|
|
|
|
*
|
|
|
|
* GCC/Clang can mostly be trusted to generate tail calls as long as
|
|
|
|
* optimization is enabled, but, debug builds will not generate tail calls
|
|
|
|
* unless "musttail" is available.
|
|
|
|
*
|
|
|
|
* We should probably either:
|
|
|
|
* 1. require that the compiler supports musttail.
|
|
|
|
* 2. add some fallback code for when musttail isn't available (ie. return
|
|
|
|
* instead of tail calling). This is safe and portable, but this comes at
|
|
|
|
* a CPU cost.
|
|
|
|
*/
|
|
|
|
#if (defined(__x86_64__) || defined(__aarch64__)) && defined(__GNUC__)
|
|
|
|
#define UPB_FASTTABLE_SUPPORTED 1
|
|
|
|
#else
|
|
|
|
#define UPB_FASTTABLE_SUPPORTED 0
|
|
|
|
#endif
|
|
|
|
|
|
|
|
/* define UPB_ENABLE_FASTTABLE to force fast table support.
|
|
|
|
* This is useful when we want to ensure we are really getting fasttable,
|
|
|
|
* for example for testing or benchmarking. */
|
|
|
|
#if defined(UPB_ENABLE_FASTTABLE)
|
|
|
|
#if !UPB_FASTTABLE_SUPPORTED
|
|
|
|
#error fasttable is x86-64/ARM64 only and requires GCC or Clang.
|
|
|
|
#endif
|
|
|
|
#define UPB_FASTTABLE 1
|
|
|
|
/* Define UPB_TRY_ENABLE_FASTTABLE to use fasttable if possible.
|
|
|
|
* This is useful for releasing code that might be used on multiple platforms,
|
|
|
|
* for example the PHP or Ruby C extensions. */
|
|
|
|
#elif defined(UPB_TRY_ENABLE_FASTTABLE)
|
|
|
|
#define UPB_FASTTABLE UPB_FASTTABLE_SUPPORTED
|
|
|
|
#else
|
|
|
|
#define UPB_FASTTABLE 0
|
|
|
|
#endif
|
|
|
|
|
|
|
|
/* UPB_FASTTABLE_INIT() allows protos compiled for fasttable to gracefully
|
|
|
|
* degrade to non-fasttable if the runtime or platform do not support it. */
|
|
|
|
#if !UPB_FASTTABLE
|
|
|
|
#define UPB_FASTTABLE_INIT(...)
|
upb is self-hosting!
This CL changes the upb compiler to no longer depend on C++ protobuf libraries. upb now uses its own reflection libraries to implement its code generator.
# Key Benefits
1. upb can now use its own reflection libraries throughout the compiler. This makes upb more consistent and principled, and gives us more chances to dogfood our own C++ reflection API. This highlighted several parts of the C++ reflection API that were incomplete.
2. This CL removes code duplication that previously existed in the compiler. The upb reflection library has code to build MiniDescriptors and MiniTables out of descriptors, but prior to this CL the upb compiler could not use it. The upb compiler had a separate copy of this logic, and the compiler's copy of this logic was especially tricky and hard to maintain. This CL removes the separate copy of that logic.
3. This CL (mostly) removes upb's dependency on the C++ protobuf library. We still depend on `protoc` (the binary), but the runtime and compiler no longer link against C++'s libraries. This opens up the possibility of speeding up some builds significantly if we can use a prebuilt `protoc` binary.
# Bootstrap Stages
To bootstrap, we check in a copy of our generated code for `descriptor.proto` and `plugin.proto`. This allows the compiler to depend on the generated code for these two protos without creating a circular dependency. This code is checked in to the `stage0` directory.
The bootstrapping process is divided into a few stages. All `cc_library()`, `upb_proto_library()`, and `cc_binary()` targets that would otherwise be circular participate in this staging process. That currently includes:
* `//third_party/upb:descriptor_upb_proto`
* `//third_party/upb:plugin_upb_proto`
* `//third_party/upb:reflection`
* `//third_party/upb:reflection_internal`
* `//third_party/upbc:common`
* `//third_party/upbc:file_layout`
* `//third_party/upbc:plugin`
* `//third_party/upbc:protoc-gen-upb`
For each of these targets, we produce a rule for each stage (the logic for this is nicely encapsulated in Blaze/Bazel macros like `bootstrap_cc_library()` and `bootstrap_upb_proto_library()`, so the `BUILD` file remains readable). For example:
* `//third_party/upb:descriptor_upb_proto_stage0`
* `//third_party/upb:descriptor_upb_proto_stage1`
* `//third_party/upb:descriptor_upb_proto`
The stages are:
1. `stage0`: This uses the checked-in version of the generated code. The stage0 compiler is correct and outputs the same code as all other compilers, but it is unnecessarily slow because its protos were compiled in bootstrap mode. The stage0 compiler is used to generate protos for stage1.
2. `stage1`: The stage1 compiler is correct and fast, and therefore we use it in almost all cases (eg. `upb_proto_library()`). However its own protos were not generated using `upb_proto_library()`, so its `cc_library()` targets cannot be safely mixed with `upb_proto_library()`, as this would lead to duplicate symbols.
3. final (no stage): The final compiler is identical to the `stage1` compiler. The only difference is that its protos were built with `upb_proto_library()`. This doesn't matter very much for the compiler binary, but for the `cc_library()` targets like `//third_party/upb:reflection`, only the final targets can be safely linked in by other applications.
# "Bootstrap Mode" Protos
The checked-in generated code is generated in a special "bootstrap" mode that is a bit different than normal generated code. Bootstrap mode avoids depending on the internal representation of MiniTables or the messages, at the cost of slower runtime performance.
Bootstrap mode only interacts with MiniTables and messages using public APIs such as `upb_MiniTable_Build()`, `upb_Message_GetInt32()`, etc. This is very important as it allows us to change the internal representation without needing to regenerate our bootstrap protos. This will make it far easier to write CLs that change the internal representation, because it avoids the awkward dance of trying to regenerate the bootstrap protos when the compiler itself is broken due to bootstrap protos being out of date.
The bootstrap generated code does have two downsides:
1. The accessors are less efficient, because they look up MiniTable fields by number instead of hard-coding the MiniTableField into the generated code.
2. It requires runtime initialization of the MiniTables, which costs CPU cycles at startup, and also allocates memory which is never freed. Per google3 rules this is not really a leak, since this memory is still reachable via static variables, but it is undesirable in many contexts. We could fix this part by introducing the equivalent of `google::protobuf::ShutdownProtobufLibrary()`).
These downsides are fine for the bootstrapping process, but they are reason enough not to enable bootstrap mode in general for all protos.
# Bootstrapping Always Uses OSS Protos
To enable smooth syncing between Google3 and OSS, we always use an OSS version of the checked in generated code for `stage0`, even in google3.
This requires that the google3 code can be switched to reference the OSS proto names using a preprocessor define. We introduce the `UPB_DESC(xyz)` macro for this, which will expand into either `proto2_xyz` or `google_protobuf_xyz`. Any libraries used in `stage0` must use `UPB_DESC(xyz)` rather than refer to the symbol names directly.
PiperOrigin-RevId: 501458451
2 years ago
|
|
|
#define UPB_FASTTABLE_MASK(mask) -1
|
|
|
|
#else
|
|
|
|
#define UPB_FASTTABLE_INIT(...) __VA_ARGS__
|
upb is self-hosting!
This CL changes the upb compiler to no longer depend on C++ protobuf libraries. upb now uses its own reflection libraries to implement its code generator.
# Key Benefits
1. upb can now use its own reflection libraries throughout the compiler. This makes upb more consistent and principled, and gives us more chances to dogfood our own C++ reflection API. This highlighted several parts of the C++ reflection API that were incomplete.
2. This CL removes code duplication that previously existed in the compiler. The upb reflection library has code to build MiniDescriptors and MiniTables out of descriptors, but prior to this CL the upb compiler could not use it. The upb compiler had a separate copy of this logic, and the compiler's copy of this logic was especially tricky and hard to maintain. This CL removes the separate copy of that logic.
3. This CL (mostly) removes upb's dependency on the C++ protobuf library. We still depend on `protoc` (the binary), but the runtime and compiler no longer link against C++'s libraries. This opens up the possibility of speeding up some builds significantly if we can use a prebuilt `protoc` binary.
# Bootstrap Stages
To bootstrap, we check in a copy of our generated code for `descriptor.proto` and `plugin.proto`. This allows the compiler to depend on the generated code for these two protos without creating a circular dependency. This code is checked in to the `stage0` directory.
The bootstrapping process is divided into a few stages. All `cc_library()`, `upb_proto_library()`, and `cc_binary()` targets that would otherwise be circular participate in this staging process. That currently includes:
* `//third_party/upb:descriptor_upb_proto`
* `//third_party/upb:plugin_upb_proto`
* `//third_party/upb:reflection`
* `//third_party/upb:reflection_internal`
* `//third_party/upbc:common`
* `//third_party/upbc:file_layout`
* `//third_party/upbc:plugin`
* `//third_party/upbc:protoc-gen-upb`
For each of these targets, we produce a rule for each stage (the logic for this is nicely encapsulated in Blaze/Bazel macros like `bootstrap_cc_library()` and `bootstrap_upb_proto_library()`, so the `BUILD` file remains readable). For example:
* `//third_party/upb:descriptor_upb_proto_stage0`
* `//third_party/upb:descriptor_upb_proto_stage1`
* `//third_party/upb:descriptor_upb_proto`
The stages are:
1. `stage0`: This uses the checked-in version of the generated code. The stage0 compiler is correct and outputs the same code as all other compilers, but it is unnecessarily slow because its protos were compiled in bootstrap mode. The stage0 compiler is used to generate protos for stage1.
2. `stage1`: The stage1 compiler is correct and fast, and therefore we use it in almost all cases (eg. `upb_proto_library()`). However its own protos were not generated using `upb_proto_library()`, so its `cc_library()` targets cannot be safely mixed with `upb_proto_library()`, as this would lead to duplicate symbols.
3. final (no stage): The final compiler is identical to the `stage1` compiler. The only difference is that its protos were built with `upb_proto_library()`. This doesn't matter very much for the compiler binary, but for the `cc_library()` targets like `//third_party/upb:reflection`, only the final targets can be safely linked in by other applications.
# "Bootstrap Mode" Protos
The checked-in generated code is generated in a special "bootstrap" mode that is a bit different than normal generated code. Bootstrap mode avoids depending on the internal representation of MiniTables or the messages, at the cost of slower runtime performance.
Bootstrap mode only interacts with MiniTables and messages using public APIs such as `upb_MiniTable_Build()`, `upb_Message_GetInt32()`, etc. This is very important as it allows us to change the internal representation without needing to regenerate our bootstrap protos. This will make it far easier to write CLs that change the internal representation, because it avoids the awkward dance of trying to regenerate the bootstrap protos when the compiler itself is broken due to bootstrap protos being out of date.
The bootstrap generated code does have two downsides:
1. The accessors are less efficient, because they look up MiniTable fields by number instead of hard-coding the MiniTableField into the generated code.
2. It requires runtime initialization of the MiniTables, which costs CPU cycles at startup, and also allocates memory which is never freed. Per google3 rules this is not really a leak, since this memory is still reachable via static variables, but it is undesirable in many contexts. We could fix this part by introducing the equivalent of `google::protobuf::ShutdownProtobufLibrary()`).
These downsides are fine for the bootstrapping process, but they are reason enough not to enable bootstrap mode in general for all protos.
# Bootstrapping Always Uses OSS Protos
To enable smooth syncing between Google3 and OSS, we always use an OSS version of the checked in generated code for `stage0`, even in google3.
This requires that the google3 code can be switched to reference the OSS proto names using a preprocessor define. We introduce the `UPB_DESC(xyz)` macro for this, which will expand into either `proto2_xyz` or `google_protobuf_xyz`. Any libraries used in `stage0` must use `UPB_DESC(xyz)` rather than refer to the symbol names directly.
PiperOrigin-RevId: 501458451
2 years ago
|
|
|
#define UPB_FASTTABLE_MASK(mask) mask
|
|
|
|
#endif
|
|
|
|
|
|
|
|
#undef UPB_FASTTABLE_SUPPORTED
|
|
|
|
|
|
|
|
/* ASAN poisoning (for arena).
|
|
|
|
* If using UPB from an interpreted language like Ruby, a build of the
|
|
|
|
* interpreter compiled with ASAN enabled must be used in order to get sane and
|
|
|
|
* expected behavior.
|
|
|
|
*/
|
|
|
|
|
|
|
|
/* Due to preprocessor limitations, the conditional logic for setting
|
|
|
|
* UPN_CLANG_ASAN below cannot be consolidated into a portable one-liner.
|
|
|
|
* See https://gcc.gnu.org/onlinedocs/cpp/_005f_005fhas_005fattribute.html.
|
|
|
|
*/
|
|
|
|
#if defined(__has_feature)
|
|
|
|
#if __has_feature(address_sanitizer)
|
|
|
|
#define UPB_CLANG_ASAN 1
|
|
|
|
#else
|
|
|
|
#define UPB_CLANG_ASAN 0
|
|
|
|
#endif
|
|
|
|
#else
|
|
|
|
#define UPB_CLANG_ASAN 0
|
|
|
|
#endif
|
|
|
|
|
|
|
|
#if defined(__SANITIZE_ADDRESS__) || UPB_CLANG_ASAN
|
|
|
|
#define UPB_ASAN 1
|
|
|
|
#define UPB_ASAN_GUARD_SIZE 32
|
|
|
|
#ifdef __cplusplus
|
|
|
|
extern "C" {
|
|
|
|
#endif
|
|
|
|
void __asan_poison_memory_region(void const volatile *addr, size_t size);
|
|
|
|
void __asan_unpoison_memory_region(void const volatile *addr, size_t size);
|
|
|
|
#ifdef __cplusplus
|
|
|
|
} /* extern "C" */
|
|
|
|
#endif
|
|
|
|
#define UPB_POISON_MEMORY_REGION(addr, size) \
|
|
|
|
__asan_poison_memory_region((addr), (size))
|
|
|
|
#define UPB_UNPOISON_MEMORY_REGION(addr, size) \
|
|
|
|
__asan_unpoison_memory_region((addr), (size))
|
|
|
|
#else
|
|
|
|
#define UPB_ASAN 0
|
|
|
|
#define UPB_ASAN_GUARD_SIZE 0
|
|
|
|
#define UPB_POISON_MEMORY_REGION(addr, size) \
|
|
|
|
((void)(addr), (void)(size))
|
|
|
|
#define UPB_UNPOISON_MEMORY_REGION(addr, size) \
|
|
|
|
((void)(addr), (void)(size))
|
|
|
|
#endif
|
|
|
|
|
|
|
|
/* Disable proto2 arena behavior (TEMPORARY) **********************************/
|
|
|
|
|
|
|
|
#ifdef UPB_DISABLE_PROTO2_ENUM_CHECKING
|
|
|
|
#define UPB_TREAT_PROTO2_ENUMS_LIKE_PROTO3 1
|
|
|
|
#else
|
|
|
|
#define UPB_TREAT_PROTO2_ENUMS_LIKE_PROTO3 0
|
|
|
|
#endif
|
|
|
|
|
|
|
|
#if defined(__cplusplus)
|
|
|
|
#if defined(__clang__) || UPB_GNUC_MIN(6, 0)
|
|
|
|
// https://gcc.gnu.org/gcc-6/changes.html
|
|
|
|
#if __cplusplus >= 201402L
|
|
|
|
#define UPB_DEPRECATED [[deprecated]]
|
|
|
|
#else
|
|
|
|
#define UPB_DEPRECATED __attribute__((deprecated))
|
|
|
|
#endif
|
|
|
|
#else
|
|
|
|
#define UPB_DEPRECATED
|
|
|
|
#endif
|
|
|
|
#else
|
|
|
|
#define UPB_DEPRECATED
|
|
|
|
#endif
|
|
|
|
|
|
|
|
// begin:google_only
|
|
|
|
// #define UPB_IS_GOOGLE3
|
|
|
|
// end:google_only
|
|
|
|
|
upb is self-hosting!
This CL changes the upb compiler to no longer depend on C++ protobuf libraries. upb now uses its own reflection libraries to implement its code generator.
# Key Benefits
1. upb can now use its own reflection libraries throughout the compiler. This makes upb more consistent and principled, and gives us more chances to dogfood our own C++ reflection API. This highlighted several parts of the C++ reflection API that were incomplete.
2. This CL removes code duplication that previously existed in the compiler. The upb reflection library has code to build MiniDescriptors and MiniTables out of descriptors, but prior to this CL the upb compiler could not use it. The upb compiler had a separate copy of this logic, and the compiler's copy of this logic was especially tricky and hard to maintain. This CL removes the separate copy of that logic.
3. This CL (mostly) removes upb's dependency on the C++ protobuf library. We still depend on `protoc` (the binary), but the runtime and compiler no longer link against C++'s libraries. This opens up the possibility of speeding up some builds significantly if we can use a prebuilt `protoc` binary.
# Bootstrap Stages
To bootstrap, we check in a copy of our generated code for `descriptor.proto` and `plugin.proto`. This allows the compiler to depend on the generated code for these two protos without creating a circular dependency. This code is checked in to the `stage0` directory.
The bootstrapping process is divided into a few stages. All `cc_library()`, `upb_proto_library()`, and `cc_binary()` targets that would otherwise be circular participate in this staging process. That currently includes:
* `//third_party/upb:descriptor_upb_proto`
* `//third_party/upb:plugin_upb_proto`
* `//third_party/upb:reflection`
* `//third_party/upb:reflection_internal`
* `//third_party/upbc:common`
* `//third_party/upbc:file_layout`
* `//third_party/upbc:plugin`
* `//third_party/upbc:protoc-gen-upb`
For each of these targets, we produce a rule for each stage (the logic for this is nicely encapsulated in Blaze/Bazel macros like `bootstrap_cc_library()` and `bootstrap_upb_proto_library()`, so the `BUILD` file remains readable). For example:
* `//third_party/upb:descriptor_upb_proto_stage0`
* `//third_party/upb:descriptor_upb_proto_stage1`
* `//third_party/upb:descriptor_upb_proto`
The stages are:
1. `stage0`: This uses the checked-in version of the generated code. The stage0 compiler is correct and outputs the same code as all other compilers, but it is unnecessarily slow because its protos were compiled in bootstrap mode. The stage0 compiler is used to generate protos for stage1.
2. `stage1`: The stage1 compiler is correct and fast, and therefore we use it in almost all cases (eg. `upb_proto_library()`). However its own protos were not generated using `upb_proto_library()`, so its `cc_library()` targets cannot be safely mixed with `upb_proto_library()`, as this would lead to duplicate symbols.
3. final (no stage): The final compiler is identical to the `stage1` compiler. The only difference is that its protos were built with `upb_proto_library()`. This doesn't matter very much for the compiler binary, but for the `cc_library()` targets like `//third_party/upb:reflection`, only the final targets can be safely linked in by other applications.
# "Bootstrap Mode" Protos
The checked-in generated code is generated in a special "bootstrap" mode that is a bit different than normal generated code. Bootstrap mode avoids depending on the internal representation of MiniTables or the messages, at the cost of slower runtime performance.
Bootstrap mode only interacts with MiniTables and messages using public APIs such as `upb_MiniTable_Build()`, `upb_Message_GetInt32()`, etc. This is very important as it allows us to change the internal representation without needing to regenerate our bootstrap protos. This will make it far easier to write CLs that change the internal representation, because it avoids the awkward dance of trying to regenerate the bootstrap protos when the compiler itself is broken due to bootstrap protos being out of date.
The bootstrap generated code does have two downsides:
1. The accessors are less efficient, because they look up MiniTable fields by number instead of hard-coding the MiniTableField into the generated code.
2. It requires runtime initialization of the MiniTables, which costs CPU cycles at startup, and also allocates memory which is never freed. Per google3 rules this is not really a leak, since this memory is still reachable via static variables, but it is undesirable in many contexts. We could fix this part by introducing the equivalent of `google::protobuf::ShutdownProtobufLibrary()`).
These downsides are fine for the bootstrapping process, but they are reason enough not to enable bootstrap mode in general for all protos.
# Bootstrapping Always Uses OSS Protos
To enable smooth syncing between Google3 and OSS, we always use an OSS version of the checked in generated code for `stage0`, even in google3.
This requires that the google3 code can be switched to reference the OSS proto names using a preprocessor define. We introduce the `UPB_DESC(xyz)` macro for this, which will expand into either `proto2_xyz` or `google_protobuf_xyz`. Any libraries used in `stage0` must use `UPB_DESC(xyz)` rather than refer to the symbol names directly.
PiperOrigin-RevId: 501458451
2 years ago
|
|
|
#if defined(UPB_IS_GOOGLE3) && !defined(UPB_BOOTSTRAP_STAGE0)
|
|
|
|
#define UPB_DESC(sym) proto2_##sym
|
|
|
|
#else
|
|
|
|
#define UPB_DESC(sym) google_protobuf_##sym
|
|
|
|
#endif
|