CppCon 2026

Python's Global Interpreter Lock, which determines which single thread can execute native Python code and call C API functions, simplifies writing multithreaded code. By default, the most popular C++ bindings to this API, pybind11 and Cython, implicitly enables the GIL by default. However, sticking with this execution model leaves out extra performance afforded by modern multicore CPUs with hyperthreading, as automatic locking and unlocking of the GIL does not scale well with thread counts, especially in performance-sensitive workloads.

The newfangled free-threaded interpreter promises salvation when running either pure Python code or with compiled extensions. General multithreading rules apply (prefer thread-local variables, using locks to prevent simultaneous access of shared data), but when dealing with projects containing compiled extensions that directly or indirectly interface with Python's C API, more porting rules also apply.

Key porting tips, including projects using the Limited API, include: port native code away from C API functions that avoid borrowed references because they aren't thread-safe; modify unit tests to catch concurrency bugs arising from assuming the presence of the GIL; and extend CI coverage of Python interpreters both for testing and to build free-threaded compatible wheels.

Tooling Track sessions are sponsored by Optiver.

In low-latency C++, correctness is table stakes. What separates good systems from great ones is predictability — the ability to hit your deadline not just on average, but at the 99th and 99.9th percentile, where real workloads live. Latency is a feature, and if you don't design for it explicitly, you lose it accidentally.

This talk is a practitioner's guide to writing C++ that behaves predictably under load, drawn from experience building and tuning low-latency systems on Microsoft's Azure Core platform. We start from the mental models that matter most — CPU-centric thinking, the latency stack from L1 to DRAM, and why minimizing work, unpredictability, and memory movement is the foundation of everything else — and then work through the layers of the stack where latency is won or lost in practice.

Through live demos and benchmark data on both x86 and ARM, we'll cover:

STL containers as latency contracts — why most latency bugs start with the wrong container, the real cost of dynamic reallocation, and the measurable wins from reserve() and custom allocators.

Memory layout and cache behavior — struct layout and alignment (and how it differs on x86-64 vs. ARM64), AoS vs. SoA trade-offs, hot/cold data separation, false sharing, NUMA, and TLB pressure.

Atomics vs. mutexes — when atomics actually lose to mutexes, why architecture and memory model dictate the answer, and how to choose between them in real code.

Threading and scheduling — fixed vs. dynamic thread pools, context-switch and migration costs, and why understanding the OS is half the battle.

Benchmarking and observability — why microbenchmarks lie, how to avoid measurement bias and jitter, why latency histograms beat averages, and why "absolute benchmark" is a myth.

You'll leave with a practical playbook for designing and measuring low-latency C++ systems, an honest understanding of why benchmarks from one machine rarely generalize to another, and a sharper instinct for the silent killers — allocations, branches, cache misses, and synchronization primitives — that don't show up in code review but do show up in production.

In CPython 3.14, we introduced an alternative interpreter implementation. This implementation used [[musttail]] , an experimental C/C++ feature that tells the compiler to enforce tail calls. Against the previous interpreter using computed gotos/switch-case, the new interpreter style shows 2%--15% performance improvement (i.e. geometric mean running time), and is more resilient against certain types of compiler bugs. This feature is already supported by Clang, GCC, and MSVC, and is also the center of a draft proposal targeting C++29. In this talk, I’ll cover the story of getting this feature into the Python interpreter, our collaboration with compiler developers, and the potential wider impact on the C++ community. I’ll also cover interesting use cases of this feature other than for writing interpreters, such as for automatically generating a simple Just in Time (JIT) compiler.

Tooling Track sessions are sponsored by Optiver.

Zero-cost abstractions are a foundational idea in C++, promising expressive, high-level code without sacrificing performance. However, developers often encounter unexpected costs when using modern abstractions in practice, even when the code appears idiomatic and well-designed.

This talk explores the assumptions behind zero-cost abstractions and examines what happens when those assumptions no longer hold. Through concrete examples drawn from modern C++ — including ranges and views, type erasure, allocators, and other common abstractions — we will examine how factors such as optimizer visibility, inlining boundaries, allocation behavior, and runtime flexibility influence performance.

This talk treats abstraction as a powerful engineering tool, examining the limits of its zero-cost guarantees in real-world systems. We will look at how seemingly small design choices can introduce hidden costs, why those costs are often difficult to spot through inspection alone, and how to recover performance without abandoning good design or readability.

Attendees will leave with a clearer understanding of when abstractions are truly zero-cost, how to recognize situations where they are not, and how to make informed tradeoffs between expressiveness, flexibility, and performance in modern C++.

Notifiers are a critical synchronization primitive in task-parallel programming systems such as Intel TBB and Taskflow, responsible for efficiently sleeping and waking worker threads as tasks become unavailable and available over and over again, directly impacting scheduler throughput and latency. Correctness here is non-negotiable: a single missed wakeup can significantly hamper the performance of an entire program. Yet writing strong unit tests for notifiers is notoriously difficult, because the bugs they target, lost wakeups, spurious wakes, race conditions, are timing-dependent, non-deterministic, and often only surface under specific thread interleavings that are hard to force reliably.

The problem is compounded in practice. Notifier implementations evolve constantly: small algorithmic tweaks, memory ordering changes, and refactors across systems demand a fresh round of carefully constructed tests. This is tedious, expertise-heavy work that takes a lot of time and engineering effort. In this talk, we explore using Large Language Models (LLMs) to automate the generation of the unit tests for notifiers. Specifically, we will demonstrate how LLM-generated tests, guided by proper prompts can systematically stress the two-phase wait protocol across Notifiers in Taskflow. We will show this in a widely used Notifier implemented in Taskflow. We are able to find an undiscovered bug that has been existing in the project.

Range views enable a "no raw loops" style of programming — not as a matter of taste, but as a pragmatic way to reuse well-tested algorithms to write code faster, express intent clearly, and avoid common bugs. In practice, however, developers often run into issues that may lead them to abandon ranges altogether: - unintuitive behavior and surprising limitations - slower compilation and runtime performance compared to raw loops - high complexity when implementing custom views

In this talk, we'll distill the core design choices behind these problems — and the alternatives that avoid them. In particular, we'll compare: - iterators and indices - external and internal iteration - transformations of nested views that overcome limitations of internal iteration

Attendees will leave with practical insights for designing and using range abstractions, along with examples of libraries that embody these ideas.

Benchmarks shouldn't live in a different world from tests. Most C++ projects already use Google Test, but the moment performance enters the picture, developers reach for separate tools, separate workflows, and ad-hoc timing loops that don't survive contact with CI. This talk argues for a different default: treat performance measurements as first-class tests. They are written in the same harness, run in the same suite, and fail the same builds.

Using a real C++23 framework built on top of Google Test, the session shows what that looks like in practice. Semantic macros distinguish throughput, latency, and contention tests; built-in statistical analysis tracks medians, percentiles, and coefficient of variation. Adaptive thresholds scale with payload size, so tiny operations and multi-megabyte workloads aren't held to the same stability expectations.

Once benchmarking lives inside the test framework, the rest of the performance toolchain follows. Five CPU profiler backends drop in behind a single --profile flag: perf, gperftools, bpftrace, RAPL, and callgrind. Attaching a memory profile to a test prints bandwidth, efficiency, and a CPU-bound or memory-bound classification with no extra code. The same harness extends to CUDA through a fluent kernel builder that captures launch configuration, achieved occupancy, transfer overhead, multi-GPU scaling efficiency, and thermal throttling. Nsight Compute drops in through the same --profile flag.

The payoff is CI integration that's nearly automatic. A companion CLI tool compares baseline and candidate runs, applies statistical thresholds, posts markdown reports on pull requests, and fails the build on regressions. The result is a single workflow for performance verification that scales from a developer's laptop to a CI runner to a Jetson on a workbench, across hardware ranging from x86 to ARM to RISC-V.

Tooling Track sessions are sponsored by Optiver.

Starting from the simple objective of writing an optimized lexer we will grapple with their fundamental performance factor: branch prediction. We will investigate how lookup tables can improve (or hurt) the CPUs prediction capabilities with some unexpected results. Over time our design evolves by combining parsing and lexing into a single step in a way that is as fast as just a standalone lexer. A central aspect of the design will be a primitive parsing state machine from which the parser source code is generated.

After the talk attendees will have a better understanding of some advanced branch prediction techniques and should have some new ideas for their present or future parsing endeavors.