C++ 优化点
C++ 高性能
经验法则
- reserve memory for std::vector
- 编译器优化
- 函数传参时避免 copy
- 减少shared_ptr 的 copy
- Avoid the compiler ever having to consider a potential call to malloc(), or any extern function, in hot loops.
- avoid thread synchronisation anywhere inside a hot loop.
- Keep collections of data in a single contiguous block of memory as much as possible
- In multi-threading programming, sometimes copying blocks of data can be faster than sharing via a mutex. There are other benefits to this, such as avoiding data synchronisation pitfalls, deadlocks, priority inversions, etc.
- Choose std::unordered_set/map over std::set/map
- Choose std::streambuf and their variants over std::i/ostream when reading/writing lots of data.
- std::unique_ptr over std::shared_ptr.
- Evaluate things in constexpr if you can.
- Make as many types as possible TriviallyCopyable.
- 用 lambda 代替 std function https://quick-bench.com/q/4IKpxAA5VEbXVziV1G2Po-ZupSE
- std::fill 要注意模板的类型匹配 https://travisdowns.github.io/blog/2020/01/20/zero.html
- Multiplying and shifting is a standard technique to emulate the division.
Final 去虚拟化
https://gcc.godbolt.org/z/16zMdjEjY
Virtual functions do not incur too much additional cost by themselves, The number one factor that is responsible for bad performance are data cache misses
Avoiding vector of pointers on a hot path is a must
Most overhead of virtual functions comes from small functions, they cost more to call than to execute
虚函数通常通过虚函数表来实现，在虚表中存储函数指针，实际调用时需要间接访问，这需要多一点时间。
然而这并不是虚函数速度慢的主要原因，真正原因是编译器在编译时通常并不知道它将要调用哪个函数，所以它不能被内联优化和其它很多优化，因此就会增加很多无意义的指令（准备寄存器、调用函数、保存状态等），而且如果虚函数有很多实现方法，那分支预测的成功率也会降低很多，分支预测错误也会导致程序性能下降。
Inline
Memory Manage 内存管理
memcpy
内存相关的操作对性能影响会比较大，memcpy 基础操作的优化点：
- Prefetch：小数据量时使用
- 区分内存对齐和非内存对齐的 Copy，内存对齐的simd 指令会更快
- Simd 指令加速：sse,avx,avx2,avx512 等指令
- Overlapped： to copy the last 4 <= n <= 7 bytes: copy the first & last 4 bytes。 非对齐字节，
- 不同 szie 使用不同的策略：小于8，小于16，小于32，小于256，大于256 做不同的处理
- 绕过 CPU Cache，内存直接写，避免影响 CPU Cache
- 查表法：拷贝不同小尺寸内存，直接跳转到相应地址解除循环。
- 循环展开
- 汇编优化
几个系统的优化实现，代码里面的注释可以细读：
- StarRocks: Improve performance of strings::memcpy_inlined https://github.com/StarRocks/starrocks/pull/13330/files
- ClickHouse: https://github.com/ClickHouse/ClickHouse/pull/21520/files
- folly: https://github.com/facebook/folly/blob/master/folly/memcpy.S
- 标准库： https://www.zhihu.com/question/35172305
memchr
https://gms.tf/stdfind-and-memchr-optimizations.html
memcmp
dpdk Implement memcmp using AVX/SSE instructio : http://patchwork.dpdk.org/project/dpdk/patch/1429716828-19012-2-git-send-email-rkerur@gmail.com/
+/**
+ * Compare 16 bytes between two locations.
+ * locations should not overlap.
+ */
+static inline bool
+rte_cmp16(const uint8_t *src_1, const uint8_t *src_2)
+{
+ __m128i xmm0;
+ __m128i xmm1;
+ __m128i vcmp;
+ uint32_t vmask;
+
+ xmm0 = _mm_loadu_si128((const __m128i *)src_1);
+ xmm1 = _mm_loadu_si128((const __m128i *)src_2);
+
+ vcmp = _mm_cmpeq_epi16(xmm0, xmm1);
+ vmask = _mm_movemask_epi8(vcmp);
+ return (!(vmask == 0xffffU));
+}
memequal
类似 memcmp
Loop Optimization
Unrolling
Resize No Initialize
Copy To Move
Std::vector
Code Layout
Compile-time Computation
查表法 Table Lookup
- optimize harmonic mean evaluation in hll::estimate_cardinality https://github.com/StarRocks/starrocks/pull/16351
- Make your lookup table do more https://commaok.xyz/post/lookup_tables/
编译期运算 constexpr
std::inplace_merge 代替 std::merge
std::inplace_merge 是原地 Merge，相比 std::merge 更省内存
https://github.com/StarRocks/starrocks/pull/14609
用 std::variant 代替 polymorphism
memory_resource 优化内存分配
http://quick-bench.com/kW0AAZlrPK0ubds9_ZMsEYvFnlI
高性能第三方库
fmt
https://www.zverovich.net/slides/2017-cppcon.pdf
llfio
https://github.com/ned14/llfio
asio
zeromq
https://aosabook.org/en/v2/zeromq.html
nanomsg
https://github.com/nanomsg/nanomsg
phmap
https://github.com/greg7mdp/parallel-hashmap
roaringbitmap
simdjson
https://github.com/simdjson/simdjson
不同的指令集
High-performance load-time implementation Selection:
https://github.com/CppCon/CppCon2022/blob/main/Presentations/Load-Time-Function-Selection.pdf
数学相关
- Computing the number of digits of an integer even faster: https://lemire.me/blog/2021/06/03/computing-the-number-of-digits-of-an-integer-even-faster/