reserved. © 2024 Bloomberg Finance L.P. All rights reserved. Can you RVO? Using Return Value Optimization for Performance in Bloomberg’s C++ Codebases CppCon 2024 September 16, 2024 Michelle D’Souza RVO? How many people here have heard about “Return Value Optimization”? 2 How many people here are experts on “Return Value Optimization”?© 2018 Bloomberg Finance L.P. All rights reserved. © 2024 Finance L.P. All rights reserved. © 2024 Bloomberg Finance L.P. All rights reserved. Return Value Optimization (RVO) 5© 2018 Bloomberg Finance L.P. All rights reserved. Agenda Questions at the end Source:

 However, an optimizer might cause this code to fail. 17 Overly-Aggressive Optimization  To see why optimization might be a problem, consider this variation on the code, which sends a '\r' and Volatile Means (and Doesn't Mean) Copyright © 2024 by Ben Saks and Dan Saks 0-8 Overly-Aggressive Optimization  Although they’re mapped to memory locations, device registers aren’t ordinary memory.  effects.  Eliminating those side effects might cause device drivers to fail. 19 Overly-Aggressive Optimization  Unfortunately, to the compiler, USTAT0looks like an ordinary object.  Its state should change

BENCHMARK ● 03a, b – branch is not predicted, optimization works ● 03c, d – branch is well-predicted, no optimizationBranchless Computing 51 ADVANCED OPTIMIZATION – ALWAYS MEASURE ● Sometimes the compiler – Often using “conditional move” instructions (they are not branches) ● Compiler’s branchless optimization is usually better than yours ● In particular, this is almost always branchless in reality: return Branchless Computing 52 BENCHMARK ● 04c, d – optimization does not work with GCC ● with perf – no branchBranchless Computing 53 ADVANCED OPTIMIZATION – ALWAYS MEASURE ● Sometimes the compiler will

hardware. Static compilation can also devote an arbitrary amount of time to program analysis and optimization. This brings us to the primary constraint on JiT systems: speed A JiT system must not cause hardware. Static compilation can also devote an arbitrary amount of time to program analysis and optimization. This brings us to the primary constraint on JiT systems: speed A JiT system must not cause hardware. Static compilation can also devote an arbitrary amount of time to program analysis and optimization. This brings us to the primary constraint on JiT systems: speed A JiT system must not cause

= 0; j != n; j += 2) ++n2; } ● As written, loop overhead is incurred twice ● Very desirable optimization: void f(size_t n) { for (size_t j = 0; j != n; j += 2) ++n1, ++n2; }Undefined Behavior 19 Why Optimized code: for (size_t j = 0; j != n; j += 2) ++n1, ++n2; ● Is the optimization valid? ● If the loops terminate, then the optimization is valid ● If the first loop runs forever, n2 should not be incremented today ● Oversights or omissions in the standard – Some may be considered bugs in the standard ● Optimization-related reasons always remain – Compilers need to make assumptions they cannot proveUndefined

string_view • Using string_view safely • Intro to span • span vs. string_view • Case study of an optimization using string_view. 2Motivation • Consider a function foo which operates on an immutable string allocation. Ensure that string_view is not converted to string later. That will cause us to lose the optimization. 24 Where to use?• string_view can used in constexpr functions. • string constructors are constexpr (aka 'span>') return kKnownHosts; ^~~~~~~~~~~ <> Optimization: Using span, string_view in GetKnownHosts 48Using the above version, we immediately run into

== sizeof(void*)); Andreas Fertig v1.0 Fast and small C++ 12 Implementing the Small String Optimization 1 struct string { 2 size_t mSize{}; 3 size_t mCapacity{}; 4 char* mData{}; 5 char mSSOBuffer[16]{}; Andreas Fertig https://AndreasFertig.com post@AndreasFertig.com 9Implementing the Small String Optimization - libstdc++ 1 struct string { 2 char* mPtr; 3 size_t mSize{}; 4 union { 5 size_t mCapacity; sso_cap(); } 26 }; Andreas Fertig v1.0 Fast and small C++ 14 Implementing the Small String Optimization - MS STL 1 struct string { 2 union { 3 char* mPtr; 4 char mBuf[8]; 5 }; 6 size_t mSize{};

compile-time decision to use an alternative implementation to potentially get this optimization through the empty base optimization (EBO). 19 Copyright © 2020 by Stephen C. Dewhurst and Daniel Saks 18 19Class Saks 8 Empty Base Classes  The “Empty Base Class Optimization,” also known as the EBCO or EBO (if you prefer TLA’s) is a common compiler optimization.  In component design, base classes are often simply storage for an empty base class.  Sometimes it makes sense to help the compiler to perform the optimization by permuting the base classes. class Derived : public Empty, public NonEmpty { ~~~ }; NonEmpty

left out - Folly MPMC - moodycamel::ConcurrentQueue94 SeqLockQueue95 FastQueue – Baseline96 Optimization #1: Caching the Write Counter void QProducer::Write(std::span buffer) { const int32_t mQ->mReadCounter.store(mLocalCounter, std::memory_order_release); mNextElement += payloadSize; }97 Optimization #1: Caching the Write Counter void QProducer::Write(std::span buffer) { const int32_t template inline T Align(T value) { return (value + (S - 1)) & (~(S - 1)); }99 Optimization #2: Data Alignment void QProducer::Write(std::span buffer) { const int32_t payloadSize

memory algorithms from • std::flat_map and std::flat_set • [[likely]], [[unlikely]] • Optimization hints to use with care 18What is SG14? • There have already been some contributions to C++ due [[likely]], [[unlikely]] • [[no_unique_address]] • Non-inheritance-based alternative to the empty base optimization in some cases 19What is SG14? • There have already been some contributions to C++ due in part pitfall” with realloc(), while providing an attribute usable for user code wanting the same optimization opportunities and semantics • The desired result is a compile-time error • Note: this is currently