S - C L - C++ H F T: B B P H: EMI Static Onditions in OW Latency FOR IGH Requency Rading Etter Than Ranch Rediction Ints

S EMI - STATIC C ONDITIONS IN L OW- LATENCY C++ FOR H IGH
F REQUENCY T RADING :
B ETTER THAN B RANCH P REDICTION H INTS
A P REPRINT
Paul Alexander Bilokon Maximilian Lucuta Erez Shermer

arXiv:2308.14185v1 [cs.PF] 27 Aug 2023
Department of Computing Department of Computing qSpark LLC

Department of Mathematics Imperial College London 3422 Old Capitol Trl
Imperial College London South Kensington Campus Wilmington, DE, 19808-6124
South Kensington Campus Exhibition Road United States
Exhibition Road London SW7 2AZ [email protected]
London SW7 2AZ [email protected]
[email protected]
August 27, 2023
A BSTRACT
Conditional branches pose a challenge for code optimisation, particularly in low latency
settings. For better performance, processors leverage dedicated hardware to predict the
outcome of a branch and execute the following instructions speculatively, a powerful op-
timisation. Modern branch predictors employ sophisticated algorithms and heuristics that
utilise historical data and patterns to make predictions, and often, are extremely effec-
tive at doing so. Consequently, programmers may inadvertently underestimate the cost of
misprediction when benchmarking code with synthetic data that is either too short or too
predictable. While eliminating branches may not always be feasible, C++20 introduced the
[[likely]] and [[unlikely]] attributes that enable the compiler to perform spot optimisations
on assembly code associated with likely execution paths. Can we do better than this?
This work presents the development of a novel language construct, referred to as a semi-
static condition, which enables programmers to dynamically modify the direction of a
branch at run-time by modifying the assembly code within the underlying executable. Sub-
sequently, we explore scenarios where the use of semi-static conditions outperforms tradi-
tional conditional branching, highlighting their potential applications in real-time machine
learning and high-frequency trading. Throughout the development process, key consider-
ations of performance, portability, syntax, and security were taken into account. The re-
sulting construct is open source and can be accessed at https://github.com/maxlucuta/
semi-static-conditions.
Acknowledgments
Thank you to Jonathan Keinan, Lior Keren, Nataly Rasovsky, Nimrod Sapir, Michael Stevenson, and other
colleagues at qSpark for many constructive suggestions.
Semi-static Conditions A P REPRINT
1 Introduction
1.1 Aims and Approach
The implementation of semi-static conditions, a language construct that can programmatically alter the di-
rection of a branch at execution time, and the identification of applications where it outperforms conditional
branching constitute the primary objectives of this project. To achieve this goal, minimal run-time overhead
associated with branch-taking must be ensured. A key challenge in the development process is to create a
language construct that mimics the behavior of direct method invocations while simultaneously providing
the ability for dynamic switching, in order to offer competitive performance with existing software solutions.
The research contribution of this paper is structured into two stages. In the first stage, the focus is on the
development of the construct, with an emphasis on the strategies employed to facilitate low-latency branch-
taking through binary editing and addressing the associated considerations of syntax, performance and
portability. The second stage shifts the focus to exploring instances where this construct surpasses conditional
statements, and understanding the behaviour and implications of semi-static conditions at the hardware
level. Additionally, a comprehensive software archive showcasing examples of usage will be created and
made readily accessible.
1.2 Research Context
Current research in branch prediction optimization has predominantly concentrated on hardware-based so-
lutions that aim to enhance speculative efficiencies and overall performance of modern CPU’s. However, de-
spite these efforts, the problem of branch prediction remains unresolved [1], with limited attention given to
software-based optimizations. For typical commercial applications, the pursuit of such micro-optimizations is
often unnecessary and may introduce additional complexity. Nonetheless, in industries like High Frequency
Trading, even slight improvements in execution latencies on the clock cycle level are highly valued. Conse-
quently, these optimizations are highly sought-after and can provide significant competitive advantages. Due
to their crucial role in determining a firm’s profitability and success, cutting-edge research on software-based
optimizations in such industries is typically shrouded in secrecy.
Several books have attempted to bridge the divide between computational and financial research, aiming to
combine mathematical modeling and the development of algorithmic trading strategies (e.g., [2, 3]). Addi-
tionally, numerous public conferences are available, focusing on the development of low-latency execution
systems (e.g., [4, 5, 6, 7]), emphasizing topics such as data structures, atomics, and low-latency design
patterns. Notably, there has been some emphasis on branchless design [8], which showcases some common
alternatives to conditional statements with high misprediction rates. However, the strategies employed in
this context lack flexibility and rely on the assumption that branches can be pre-computed without incurring
significant costs.
In contrast, there is a wealth of literature and extensively documented resources available for C++ [9, 10],
which is widely used as the primary language in the development of low-latency trading systems. However,
when it comes to the development of semi-static conditions and strategies involving the modification of
running executables using C++, the scope becomes more specialized. Nevertheless, these techniques are
well-documented and find applications in various areas such as debuggers, profilers, hot patching software,
and security tools (e.g., [11, 12, 13]).
There is a significant scarcity of ultra-low latency C++ tools, particularly those specifically designed to
address control flow problems, offering both rigorous application verification and superior performance.
While it is possible to come across online posts outlining small-scale experiments that focus on minor branch
optimizations in specific scenarios (as demonstrated in [8]), such micro-optimizations are often overlooked
and left to the compiler and hardware to handle. Interestingly, extensive research has been conducted
on the true cost of branch misprediction [14, 15], highlighting its significant contribution to performance
bottlenecks in low-latency systems. While these articles provide in-depth analysis of benchmark data and
performance losses, the strategies proposed for preventing branch mispredictions remain lackluster or non-
existent.
In light of the existing literature and its insights into software-based branch optimization problems, it be-
comes apparent that a research gap exists, which this study aims to fill. The motivation behind this research
lies in providing solutions to the aforementioned void and contributing to the understanding and resolution
of software-based branch optimization.
2
1.3 Outline
In the research review portion of this work, the focus is on encapsulating and critically analyzing research
pertaining to the problem at hand. The section begins with an outline of modern CPU pipelined architec-
tures and the implications and cost of conditional branching. Next, attention is shifted to advancements
in hardware-based solutions, outlining the strengths and weaknesses of various schemes when encounter-
ing branches of different predictability. The focus is then redirected to software, providing an outline of
C++ and its importance in the development of low-latency trading systems. Language features that exist
exclusively to optimize branch prediction are also discussed. Finally, discussions are concluded with HFT,
examining economic effects, known technical advancements and the problems that remain to be solved in
the industry.
The next two sections dedicated to the research contribution, which can be conceptualized as consisting
of two stages: the development of semi-static conditions and data-backed applications. In the development
stage, a sequential approach is adopted to identify the requirements and challenges associated with designing
the language construct. The solution to the problem is outlined and demonstrated, providing an overview
of key theory with subsequent design decisions and optimizations. The proceeding section demonstrates
the instances where semi-static conditions offer superior performance compared to conditional branching,
accompanied by detailed analyses and supporting benchmark data to substantiate the findings, with novel
investigations into effects of binary editing on modern hardware. Furthermore, discussions are conducted
on how the outlined scenarios can be incorporated into a commercial trading system, considering their
suitability and practical implications.
Next, the software contribution alongside the various experimental methods employed to benchmark semi-
static checks are critically evaluated. Detailed examples of usage are provided, along with recommendations
for maximizing the security and reliability of the language construct. We then provide some concluding
remarks and future directions, discussing potential areas for further research and development.
1.4 Summary of Achievements
The project has achieved significant milestones across multiple dimensions. The research contribution offers
valuable insights and approaches for developing software-based branch optimizations, with comprehensive
and rigorous investigations into the resulting hardware level behaviours, filling an important research gap
in the academe. By employing unconventional yet effective binary editing strategies, the project enables
ultra-low latency branch execution through the decoupling of condition evaluation logic and branch tak-
ing, controlled directly by the programmer. Through extensive benchmarking and detailed performance
analysis in pseudo-realistic scenarios, the proof-of-concept semi-static checks demonstrate their real-world
applicability, particularly within real-time systems and high-frequency trading.
The research contribution is encapsulated within a open-source library that allows programmers to utilize
the language construct for both commercial and experimental purposes. With a focus on syntax, security,
efficiency, and portability, the semi-static conditions can be seamlessly integrated into high-performance real-
time systems, as exemplified in domains such as high-frequency trading and real-time machine learning.
The availability of this library further enhances the practical usability and potential impact of the project’s
achievements.
2 Background
2.1 Pipelining and Conditional Branches
The microprocessor is an integrated circuit responsible for executing arithmetic, logic, control, and in-
put/output operations in a digital system [16]. In the early 1970s, the Intel 4004 emerged as the first
commercially available microprocessor, initially designed as a 4-bit central processing unit (CPU) with a
clock speed of 740 kHz for early printing calculators [17, 18]. Over the past three decades, advancements in
integrated circuit technology have enabled microprocessor manufacturers to develop increasingly sophisti-
cated CPUs, driven by the exponential growth in transistor density dictated by Moore’s law [19]. Alongside
these developments, the introduction of modern instruction sets and standardized operating systems has
propelled the computational capabilities of contemporary computers to unprecedented heights.
Modern CPU’s utilize pipelining as an implementation technique to exploit the inherent parallelism in instruc-
tion execution to improve overall throughput [16]. Similar to cars on an assembly line, pipelining allows for
3
the overlapping execution of multiple instructions, with each step in the assembly line constituting a pipe
stage that represents a phase in the fetch-decode-execute cycle. If all stages take the same amount of time,
a pipeline with n stages will achieve a throughput n times faster than an un-pipelined counterpart, with the
bottleneck stage bounding the number of processor cycles required for a single execution [16]. Whilst we
provide a simplified schematic of instruction pipelining, it is important to note that modern processors vastly
more complex pipeline architectures, often super-scalar with complex instruction sets and addressing modes
to prevent memory-access conflicts in instruction/data memory, optimize register handling, and maximize
instruction level-parallelism [20].
Clock number
Instruction number 1 2 3 4 5 6 7 8 9
Instruction i + 1 IF ID EX MEM WB
Figure 1: Simplified representation of 5 stage pipeline using RISC instruction set. On each clock cycle, a new
instruction is fetched and begins the 5 stage fetch-decode-execute cycle, with older instructions in the 5th stage
being retired, maintaining a throughput five times greater than non-pipeline processor. IF = instruction fetch, ID =
instruction decode, EX = execution, MEM = memory access, and WB = write back. Table has been adapted from
[16].
Pipelining in CPUs, while a powerful optimization technique, introduces hazards that can impact the overall
performance. These hazards can be broadly categorized into three types: structural hazards, data hazards,
and control hazards. (1) Structural hazards occur when the hardware resources are incompatible with the se-
quence of instructions, leading to conflicts in resource allocation and pipeline stalls. However, advancements
in super-scalar architectures and out-of-order instruction execution have made structural hazards less preva-
lent to virtually non-existent [16]. (2) Data hazards arise when an instruction depends on the completion
of a previous instruction that has not finished executing. These dependencies can cause conflicts and hinder
parallel execution of instructions, but are broadly mitigated through the use of virtual registers (register
renaming) [20]. (3) Control hazards are caused by conditional branch instructions that change the program
counter. These instructions introduce uncertainty into the execution flow since the branch target needs to be
determined prior to the next instruction fetch [16]. Branches constitute around 12-30% of all instructions
executed on modern instruction sets and are widely regarded the most significant barrier to achieving single
cycle executions [21], and as a consequence, has become a large area of focus for optimisations by hardware
and software engineers.
Microprocessors employ various strategies to mitigate control hazards in the CPU instruction pipeline. The
simplest and potentially the most costly approach is a full pipeline stall/freeze, where instructions following
a branch instruction are ignored until the target of the branch is known, resulting in a fixed cycle penalty
[16, 22]. Improving upon this, processors can make static predictions about the branch target instead of
discarding subsequent instructions, maintaining sequential execution of instructions pertaining to either the
taken or not-taken branches. By leveraging compiler static analysis and optimizing likely paths of execution,
static prediction becomes a powerful optimization technique in pipelined processors, providing non-zero
probabilities of correctly predicted branch targets and thus minimises throughput loss from flushes [21, 23].
Though an improvement, this approach is rather inflexible particularly for branch targets that change. Static
prediction schemes lack the ability to adapt at runtime to changing patterns in branch target execution
which is a common theme for the majority of conditional branches. With modern CPU’s employing increas-
ingly more speculative architectures and deeper pipelines to maximise instruction throughput [16], branch
penalties become more significant, scaling monotonically with pipeline depth [24]. With this in account, it
is clear that processors require even more aggressive optimisation techniques beyond simple static analysis
to minimise idle cycles from branch mispredictions.
4
Untaken branch instruction IF ID EX MEM WB
Taken branch instruction IF ID EX MEM WB
Instruction i + 1 IF idle idle idle idle

Branch target IF ID EX MEM WB
Branch target + 1 IF ID EX MEM WB
Figure 2: Example of a predicted-not-taken scheme and the instances when the branch is taken (top) and the
branch is not taken (bottom). Correct prediction (top) results in subsequent instructions to fall through, wheres
misprediction (bottom) results instrcutions pertaining to the branch to be flushed, incurring a single cycle penalty.
Table has been adapted from [16].
The final mitigation technique utilised commonly on classical 5-stage pipeline MIPS architectures, but are
typically non-existent on modern processors, are branch delay slots which interleave instructions independent
to the branch prior to branch target deduction at decode or execute time [25]. The concept is rooted
in the observation that not all instructions in a program depend on the outcome of a branch instruction,
thereby in theory, allowing the compiler to schedule instructions before the branch is taken and hide some
of the branches latency. However, on modern processors, branch delay slots are generally avoided. Utilizing
branch delay slots effectively requires the compiler to identify and schedule instructions that can fill the
slots, adding complexity to the compiler design and optimization process. Compiler writers need to analyze
dependencies, identify independent instructions, and rearrange the code to take advantage of the delay slots.
This additional burden makes it more challenging to generate efficient code and can increase compilation
time, and are generally poor at doing so [21, 25]. With deeply complex instruction pipelines on modern
processors, the scheduling task becomes exponentially more complex and interferes with modern hardware
solutions, deeming the once performance enhancing technique as a complication in modern processors.
Whilst this section offers a fundamental overview of conditional branches and their implications at the pro-
cessor level, it is crucial to recognize the significance of primitive branch penalty mitigation techniques.
These early forms of mitigation laid the groundwork for the modern solutions discussed in the subsequent
sections, and how the evolution of modern speculative processors played a pivotal role in driving the devel-
opment of powerful hardware-based solutions, such as the branch predictor.
2.2 Dynamic Branch Prediction
The advancement of super-scalar processors has introduced intricate speculative architectures and deep in-
struction pipelines, which effectively maximize throughput to meet the performance requirements of contem-
porary computers. Consequently, the issue of branch misprediction has emerged as a significant impediment
to sequential instruction execution [16, 24]. As briefly mentioned earlier, static branch prediction techniques
rely on predetermined rules or assumptions rather than leveraging runtime information. To enhance static
prediction in modern architectures, compilers play a vital role in making profile-guided decisions based on
historical execution patterns. Notably, the binomial distribution of simple branches renders static prediction
an effective strategy [16, 26]. Supporting this notion, Fisher and Freudenberger’s work [27] demonstrates
that applications with statically predictable branches exhibit commendable performance under the current
paradigm. The overarching limitation of the former schemes however is the inability to adapt to runtime
changes or varying input conditions, and the lack of ability to capture complex patterns in branch behavior.
Whilst previous execution patterns can be used as a proxy to determine the likelihood of a branch, in a real
time system with non-deterministic data patterns, solely relying on such a scheme would likely lead to higher
mispredictions rates and performance degradation. In light of these issues, a plethora of dynamic branch
predictors where developed with the ability to adapt, leverage runtime information, and capture complex
branch patterns to improve prediction accuracy and runtime performance.
5
Instruction PHT
stream (2-bit counter)
Not taken
Predict taken Predict taken

Branch A’s 11 10
Branch A index Taken
00010 Taken Not taken
Not taken
00010
Predict not taken Predict not taken
Branch B Branch B’s 01 00
index Taken
Figure 3: Diagrammatic representation of a one-level branch history table with aliasing interference between two
branches, adapted from [28] (left). State machine for 2-bit branch prediction, adapted from [16] (right).
Dynamic branch predictors (BP) can be broadly categorized into one-level local BP’s and two-level global
BP’s, with more modern BP schemes incorporating the strengths of both. One-level BP’s typically utilize
a one-dimensional branch prediction buffer or branch history table, acting as a cache indexed by the lower
bits of the program counter related to a branch instruction in memory [16, 28]. 1-bit prediction schemes
have a single bit entry in the branch history table, indicating whether the branch has recently been taken
or not. In the event of a misprediction, the prediction bit is inverted. While this scheme offers simplicity
and low hardware overhead, the limited historical information stored in a single bit often leads to frequent
mispredictions [16]. To address this limitation, N-bit saturated counter schemes are statically assigned to
branches with distinct addresses. When a branch is about to be taken, the counter value associated with
the branch and the direction of the branch is determined based on a predetermined threshold [26]. The
count is incremented if the branch is taken, and vice versa. It may seem intuitive that increasing the number
of bits would improve prediction accuracy by storing more information about the branch history. However,
Smith [29] demonstrated that strategies employing larger counter sizes than classical 2-bit schemes do not
consistently result in higher prediction accuracies. For loop insensitive programs with biased branches, 2-bit
counter schemes have been found to be effective with branch misprediction rates averaging at 11%, but
performance was found to deteriorate with integer based programs with more complex branch dependen-
cies [16]. The limitations of one-level BP schemes are multifaceted. Relying on a single branch history
table restricts the ability to capture intricate patterns and dependencies between branches, hampering the
prediction accuracy of one-level BPs [29]. Additionally, interference can occur with branch buffer accesses,
as finite space necessitates the use of hashing schemes to access the bit-counters for predictions. This can
lead to collisions, with negative aliasing occurring more prominently than positive aliasing [28].
In light of these limitations, Yeh and Patt [30] proposed the first two-level adaptive branch prediction scheme
that utilises a global branch history table which maintains a shared history of branch outcomes, allowing it
to capture patterns and dependencies between branches. Implementation wise, two level BP’s comprise of a
branch history register (BHR) which tracks recent outcomes of branches and a global pattern history table
(PHT) which stores patterns and outcomes associated with specific branch instructions. In this scheme both
the BHR and PHT work in collaborative fashion; most recent branch results are shifted into the BHR with
branch instruction addresses used to index a BHR table, the content of the BHR is used to index the global
PHT for making predictions, with mispredictions updating both the PHT and BHR [30, 31, 28]. Whilst
the two-level adaptive predictor improves prediction accuracy and captures branch dependencies, it still
faces challenges related to aliasing conflicts similar to saturated counter-based branch predictors. Yeh and
Patt explored several alternative branch prediction schemes based on the original two-level approach [31],
and whilst branch correlation algorithms could be improved it was found that inherit trade-offs existed
between efficient storage capacity, memory access overhead and reduced interference, often bounded by
physical constraints. As a consequence, current state-of-the-art BP research is focused primarily on the
development and optimisation of the prediction algorithms that build off the foundational work originally
done by [30, 31], and have been immensely successful at doing so with the emergence of incredibly powerful
BP’s such as the competition winning TAGE-L with 3-4 mispredictions per 1000 instructions [32, 33].
6
hist[L1:0] hist[L2:0] hist[LN:0]
PC
Index Tag Index Tag Index Tag
Tag match? Tag match? Tag match?
Prediction
Figure 4: Organisation of the TAGE (Tagged Geometric) BP with N-tagged tables. TAGE features a base binomial
predictor which provides a basic prediction and a number of partially tagged tables that store branch history infor-
mation and associated predictions. At prediction time, each of the tagged tables are indexed using different history
lengths that form a geometric series, with the longest history generally being chosen [34]. The TAGE-L predictor
builds on this and uses a dynamic table organization with varying number of tables and table lengths which better
captures branch outcomes and history lengths [35]. TAGE-SC-L further builds on this and incorporates a statistical
correlator to further refine predictions [32]. Figure adapted from [34].
Whilst the current state of modern BP’s are capable of predicting the majority of branches to near perfect ac-
curacy, there exists classes of branches that are inherently hard to predict (HTP) by TAGE-like predictors and
hence the problem of branch prediction is still considered unsolved [1, 36]. The key weakness arises from
attempting to identify correlations between branches in a noisy global history, or the branches themselves are
uncorrelated and must rely solely on saturated counter schemes to make predictions[33]. Fundamentally,
noise caused by HTP branches introduce a number of redundant patterns which pollute the global branch
history causing TAGE-like predictors to struggle, and with non-deterministic ordering of historical patterns,
more exotic BP’s such as perceptron based BP’s which rely on the position of a branch in the global branch
history also struggle [28, 33, 37]. Problematically, HTP branches are common in most applications, but have
the more prevalent effect in performance critical fields such as games, streaming services, and HFT.
The weaknesses of modern BP’s stem from their need to be simple, computationally cheap and adaptive to
execution phase behaviour, preventing them from capturing complex patterns seen in HTP branches [33]. In
addition, in the worst case the storage requirements for capturing complex branch patterns are exponential,
7
and make traditional BP strategies infeasible. Whilst fundamental breakthoughs in novel BP’s that build off
TAGE and percepton-based BP’s have become rare, Zangeneh et al proposed a convolutional neural network
(CNN) based BP which can be trained offline and be capable of predicting complex and noisy branches
with improved accuracy [33]. Their work explored two CNN models, one of which was a pure software
solutions and the second being a smaller hardware optimised version capable of being implemented as a
practical BP, both of which showed vast reductions in MPKI on benchmarks run on TAGE-SC-L. Although
prediction accuracy greatly improved for correlated branches with noisy histories, BranchNet showed poor
performance on data-dependant branches and programs where mispredictions are spread across a number
of static branches. The reason for the former, is that data dependant branches are dependant on input data
and program phase behaviour which means there is little to no history of the branch that is correlated to
the data stored in memory, which is problematic for capturing training data to make predictions [33]. In
terms of mispredictions for static branches, this is a problem regarding the sparse storage requirements for
a practical CNN BP, though it may be possible to use larger models using proposed predictor virtualisation
techniques [38].
Though NN based BP’s have the potential to be the next standard for hardware-based BP’s, there is still much
work to be done before commercialisation and a number of problems to be accounted for. Whilst the work
by Zangeneh et al [33] demonstrated MPKI reductions on several HTP branches on a number of benchmarks,
the first challenge of using such CNN’s would be to achieve some sort of generalisation across all known HTP
branches, which in turn would require a substantially large model or a number of models which introduces
large administrative overhead by the OS. In addition, modern OS would need to be adapted to load these
CNN models on-chip prior to execution time and handle context switching, and associated context switching
penalties accordingly which in turn adds complexity and overhead.
At its current state, research pertaining to novel hardware based BP’s have slowed substantially, and it is
likely that modern BP’s have reached an asymptotic state of prediction accuracy versus complexity and
implementation overhead. Whilst there may be BP’s in the future that can have low misprediction rates
on noisy or inherently HTP branches, there will always exist some class of branches, especially in real-time
systems, that will always be completely non-deterministic and impossible to predict. Whilst this likely can
never be solved by hardware, it may be possible that programmatic approaches exist that are able to tame
these impossible-to-predict branches.
2.3 C++ and Compiler Hints
High-frequency trading (HFT) demands ultra-low latency and exceptional performance, necessitating the
selection of a programming language that offers efficient execution, deterministic behavior, and fine-grained
control over hardware resources. C++ has emerged as the primary language for developing critical path
(path of order action) components in automated trading systems, primarily due to its design philosophy.
One fundamental axiom associated with C++ is the ”zero overhead principle” [9, 10], which ensures that
developers only pay for what they use, resulting in a predictable and transparent performance model. For ex-
ample, unlike high-level languages that employ garbage collectors to manage memory, C++ utilizes manual
memory management. This approach avoids the unpredictable invocation times and latency associated with
garbage collectors, which is critical in low-latency applications that require deterministic performance. C++
provides developers with granular control over memory using abstractions such as pointers, references, heap
allocation operators, and standard library methods. However, the trade-off for low-level memory access is
increased complexity and the potential for memory leaks and errors. Whilst uncommon, other languages
that target the java virtual machine (JVM), for example Java, have become more popular in the HFT space
using optimised GC algorithms and compilation techniques to minimise latency cost [39]. Notably, this has
become popular with leading quantitative trading firm and liquidity provider, Jane Street.
C++ features also support the shifting of execution time operations to compile time, resulting in the defer-
ral of computational costs and the reduction of runtime latency when appropriately utilized. This capability
proves to be a powerful tool for applications with stringent requirements for low latency. Templates, as a
Turing-complete language feature, play a vital role in enabling this paradigm by providing type-safe param-
eterized blueprints, which allow the generation of specialized code by the compiler for each type-specific
instantiation. Consequently, compile-time polymorphism can be achieved, albeit at the expense of flexibil-
ity and maintainability [40, 41]. Moreover, the template system can be further leveraged for performing
recursive instantiations and type deductions, thereby facilitating the manipulation of types and values and
enabling the execution of complex computations at compile time. It is important to note that this ap-
8
i f ( condition ) 10 a f : je 10bd
function 1 (); 10b1 : call <f u n c t i o n 1 >
else (...)
function 2 (); 10bd : call <f u n c t i o n 2 >
Figure 5: Comparison of C++ code without branch prediction hints (compiled with GCC on x86-64). In this case
the forward branch (else) is assumed not taken and the backward branch (if) is assumed taken.
i f ( condition ) [[ unlikely ]] 10 a f : je 10bd

function 1 (); 10b1 : call <f u n c t i o n 2 >
else (...)
function 2 (); 10bd : call <f u n c t i o n 1 >
Figure 6: Comparison of C++ code with branch prediction hints (compiled with GCC on x86-64). Since the
backward branch is now deemed as ”unlikely”, the compiler will reorganise the ASM such that the backward
branch is now the forward branch and vice versa.
proach, known as ”template metaprogramming,” is extensively employed in low-latency settings such as

High-Frequency Trading [5, 42].
In the realm of optimizing branch prediction at the language level, specific extensions provided by compilers
have long existed, allowing programmers to provide hints for branch prediction. These hints enable targeted
optimizations on anticipated execution paths. In GCC and Clang, this capability is manifested in the form of
the builtin expect attributes, which allow programmers to specify conditions and associated probabilities
(fixed as either 0 or 1) for condition evaluation at runtime [43, 44]. It is worth noting that these so-called
branch prediction hints do not directly affect the hardware-based branch prediction of processors and have
not been effective since the release of Intel’s Pentium M and Core 2 processors [45]. Instead, compiler
built-ins optimize branch-taking by rearranging assembly code related to branches to exploit processor static
prediction schemes and instruction cache effects for improved performance.
Modern processors enhance execution performance by prefetching instructions sequentially from slower to
faster memory storage, such as high-speed caches in close proximity to the CPU. This prefetching is done to
avoid high memory access latencies during the fetch stages [46]. When executing a code segment contain-
ing conditional statements, blocks of sequential assembly instructions associated with different branches are
prefetched into the instruction cache, irrespective of how frequently the branches are executed. However,
prefetched code that remains infrequently accessed throughout the program’s lifespan can contaminate the
instruction cache and result in cache-line fragmentation of hot code segments that are frequently executed.
This can introduce jitter and latency costs [47, 48]. To optimize branch prediction, the compiler reorganizes
the underlying assembly code such that the conditional jump occurs on the least likely path. This is because
modern processors initially assume that forward branches are never taken and, therefore, avoid mispredic-
tion by fetching the likely branch to the branch target [45]. It is important to note that such static-prediction
schemes are employed when the BPU encounters a branch that has not been previously visited. The dynamic
prediction schemes outlined in the previous sections will begin to dictate which blocks are speculatively
fetched once a branch history has been established.
C++20 introduced the [[likely]] and [[unlikely]] attributes, which serve as wrappers around the orig-
inal builtin expect compiler attributes and function in the same manner. Apart from these attributes,
there have been no other language features attempting to optimize branch prediction [49]. There is limited
formal research investigating the performance impact using different benchmarks. However, some online
blogs report performance gains (e.g., [47] reports a 15% increase) primarily on branches specifically de-
signed to be highly predictable for demonstration purposes rather than from a research perspective. The
fundamental problem with these attributes is that they rely on the programmer’s accurate prediction of likely
execution paths. While profiling and synthetic benchmark data may offer insights into hot/cold branches,
programmers tend to significantly underestimate the cost of misprediction when misusing these attributes
(as programmers are notoriously poor at predicting branches) [15]. Furthermore, the capabilities of these at-
tributes are confined to compile time, rendering them inflexible during runtime. Suppose a branch is indeed
more likely to be taken at compile time, allowing the programmer to benefit from it if used correctly. In that
case, if the likelihood of the branch changes during execution, the programmer would have no control over it
9
and would suffer from increased latency through assembly reordering schematics. Real-time systems, which
constantly need to react to live events and data, inherently exhibit such variability. Consequently, static
language features like these are inadequate as effective branch misprediction mitigators. This inadequacy
forms the focus of the research conducted in this work.
2.4 High Frequency Trading
The evolution of computer-based trading dates back several decades, starting with the introduction of fully
electronic trading by NASDAQ [2]. With the decrease in regulation and advancements in electronic ex-
changes and telecommunications infrastructure, high-frequency trading (HFT) has gained significant pop-
ularity, accounting for over 50% of trading volume in equity markets [50]. Defining HFT itself poses
challenges. Haldane [51] emphasizes the use of sophisticated algorithms as its main characteristic, while
MacKenzie [52] and Arnoldi [53] highlight the importance of speed in data processing and execution rather
than the underlying strategies employed. This work primarily focuses on optimizations in proprietary auto-
mated trading systems, and hence considers HFT as a form of algorithmic trading who’s strategies rely low
order execution latencies to ensure profitability.
High-Frequency Trading (HFT) firms actively engage in generating trading signals, validating models, and
executing trades in order to exploit inefficiencies in market micro-structure within short time frames, with
the ultimate goal of achieving profitability [2]. These firms employ diverse strategies to generate profits in
financial markets. One prevalent strategy is market-making, where HFT participants continuously provide
liquidity by simultaneously placing buy and sell orders, aiming to profit from the bid-ask spread. Another
strategy, known as statistical arbitrage, involves capitalizing on transient deviations from fair value by iden-
tifying mispricing opportunities. Additionally, event-driven trading strategies focus on leveraging informa-
tional asymmetries that arise from significant market events [54]. While a significant body of literature
exists on the economic effects of HFT (e.g., [55, 56, 57]), information pertaining to the engineering aspects
of low-latency automated trading systems is often concealed. Nevertheless, some online conferences vaguely
present important features of HFT systems, such as the networking stack, kernel bypass, custom hardware,
and strategies for enhancing the speed and efficiency of production code [4, 5].
While this work primarily addresses software-based optimizations for latency-critical applications, it is crucial
to underscore the significance of the hardware stack in HFT firms for trade execution. Simple trading
strategies often rely heavily on the network stack, implemented in custom firmware and Field Programmable
Gate Arrays (FPGA) [59], to achieve execution latencies in the nanosecond range. As proprietary software
approaches maximum optimization, it is likely that the engineering focus will increasingly shift towards the
hardware stack, which is still in its early stages and has considerable room for growth.
3 Software Contribution
3.1 Outline
This section is dedicated to the idea and development of semi-static conditions, outlining theory, design phi-
losophy and optimisations that have been applied iteratively over the development process. The majority
of this section is focused on developing a prototype that emulates the behaviour of a simple if/else state-
ments, with some discussions given on generalisations to switch statements and non-static member functions
given in later sections. Whilst discussions on portability are reserved for the evaluation stage, there was an
important decision to be made about the choice of development environment for the prototype, more specif-
ically the choice of operating system, compiler and C++ version. Since the primary applications of this
language construct will find use in low-latency environments such as HFT systems, the choice of develop-
ment environment was curated to reflect an industry standard. In light of this, development was done under
a Linux OS (Ubuntu distribution) with GCC 13.1 and C++20. At certain stages of the development pro-
cess, attention will be brought on specific Linux system calls that involve manipulating page permissions
for running executables, however equivalent API’s exist in other OS’s (mentioned in later sections) and its
fundamental mechanics is common for OS’s that use paging for memory management.
3.2 Semi-static Conditions
Semi-static conditions can be defined as a language construct that emulate the behaviour of conditional
statements, but separates condition evaluation logic and branch taking (the subsequent machine code exe-
10
Network
UDP Network Stack
10G Ethernet
Network Switch &
Timestamper AXI-Stream
Network Layer
FAST FAST
Encoder Decoder Financial Protocol
Application Layer
Custom Order
App Book
Figure 7: Simplified anatomy of a HFT system. Exchanges broadcast ticker data along a 10 GiB Ethernet cable
to the HFT system, where the networking stack receives and processes packets in user space. Packets are typically
compressed in a domain specific format for bandwidth reasons, which are then parsed into meaningful market
orders by the financial protocol which are then ordered in the order book. The custom application then issues the
buy and sell orders back over the network, this is the area in which the optimisations presented in this work pertain
to[58]. Figure adapted from [58].
cuted on the pretext of the condition). At compile time, the underlying assembly of semi-static branch taking
would resemble that of a function call with no indirection, allowing for a deterministic flow of execution
with full support of compiler optimisations and hardware intrinsics. In the context of conditional branching,
semi-static branch taking behaves as a static conditional statement: the condition is evaluated at compile
time and does not change for the duration of program execution, which in a classical sense, be thought of as
compile time template polymorphism (the branch). The semi part is associated with the polymorphic nature
of the language construct: the ability to programatically change the direction of the branch at runtime, whilst
maintaining the deterministic compile-time behaviour already mentioned. With semi-static conditions, the
lines between the compilation and execution phases of a program become blurred, and the nature of the
executable shifts from static to somewhat polymorphic or self-modifying. This behaviour manifests itself
within the branch-taking mechanism as a single/sequence of assembly instructions that redirect control flow
to the respective if or else branches, controlled by branch-switching logic that performs the modification.
In the context of branch optimisation, the philosophy behind the construct is simple. The seperation of
branch-switching and branch-taking logic produces an important decoupling of relatively expensive and
cheap operations, allowing for more strategic and granular control over conditional branching. In this case,
the branch-switching logic can be considered as an expensive operation because it involves altering assembly
instructions in memory, whilst branch-taking logic can be considered cheap as it is simply a direct function
call. Isolating branch-switching logic in less performance critical code paths allow conditions in performance-
critical sections to be evaluated preemptively without interference, bypassing the need for branch prediction
(more specifically for conditional branches) and eliminating mispredictions in the latency-critical path. When
the hot-path is executed, no conditional checks are needed and the branch is executed as if it where always
11
perfectly predicted. In instances where code paths are infrequently executed, but contain branches that are
often mispredicted, semi-static conditions show promise for optimisations.
In order to realise this language construct, some key challenges are addressed in the development process,
which can be broadly split into branch-changing and branch-taking logic. Branch-changing logic needs to
be able to find the address of the assembly code instructions to edit in memory, and perform the editing in
way that calling the branch-taking method redirects program control flow to user-specified regions based
on a runtime condition. Branch-taking logic needs to ensure that control flow is redirected with minimal
overhead, so it becomes comparable with the execution latency of a perfectly predicted branch, or a direct
function call. At a language level, this is not only dependant upon the underlying assembly instructions
that are edited, but also being able to reap the full benefits of compiler optimisations without compromising
the safety of the program. On the hardware level, it is paramount that branch-taking code benefits from
the same caching, instruction pre-fetching and branch target resolution effects that regular function calls or
unconditional jumps do to ensure deterministic and low-execution latencies. Whilst it may be possible to
achieve near-identical execution schematics to direct function calls on a language level, the cost of cache
incoherence and instruction pipeline stalls can trump branch-misprediction by orders of magnitude, making
the construct infeasible in low-latency settings. Therefore, it is crucial that semi-static conditions are com-
patible in this way with modern hardware and processors. Lastly, and more broadly, the language construct
needs to be designed with ease of use in mind. This includes simple and elegant syntax, flexibility and a
design that allows it to be easily portable across different architectures, compilers and operating systems.
3.3 Prototype Development
The first step in the development process is to establish the desired syntax of the core branch-switching and
branch-taking functionality of the language construct. It is likely that this will have a significant influence on
the design of semi-static conditions, so establishing how the end product is desired to look in the preliminary
stages gives a clear direction in development goals. After careful consideration of simplicity and elegance,
the desired usage can be seen below.
void f u n c t i o n 1 ( ) { . . . }
void f u n c t i o n 2 ( ) { . . . }
(...)
BranchChanger branch ( f u n c t i o n 1 , f u n c t i o n 2 ) ;
branch . s e t d i r e c t i o n ( c o n d i t i o n ) ;
branch . branch ( ) ;
Semi-static conditions will manifest itself as the BranchChanger class which is instantiated by taking the
addresses of two functions as arguments. These functions represent the if and else branches respectively,
and their equivalent usage with conditional statements can be seen below.
void f u n c t i o n 1 ( ) { . . . }
void f u n c t i o n 2 ( ) { . . . }
(...)
i f ( condition )
function 1 ();
else
function 2 ();
The set direction method will be responsible for controlling which of the branches is executed based on
a user specified runtime condition, whilst the branch method will be responsible for executing the branch
with minimal overhead. The signature of the branch method will always be identical to the functions passed
as arguments when the class is instantiated, serving as a single entry and exit point for both branches.
Now that a clear high-level design has been established, the next course of action is to implement the
branch-taking functionality. Before delving into assembly instruction modification to facilitate the execution
of the if and else branches, some thought needs to be given in how the branch method can act as a single
entry and exit point for both branches. This method will not do any meaningful work, its sole purpose is to
12
behave as a trampoline to other areas of the code segment while still being able to propagate return values
and register data as if one of the branches where called directly. Its first clear responsibility is to set up the
call stack in the exact way that the branches would as if they where called in isolation; since control flow
will be redirected before the branch function has opportunity to manipulate the stack, in theory the target
branches will be able to use any caller-saved data as if it where called directly. To test this theory, we can
observe the disassembly for two functions with identical signatures under -O0 optimisations so any calling
behaviour is not omitted.
i n t add ( i n t a , i n t b ) { . . . }
i n t branch ( i n t a , i n t b ) { . . . }
...
mov esi , 2
mov edi , 1
call add ( i n t , i n t )
...
mov esi , 2
mov edi , 1
call branch ( i n t , i n t )
As expected, both instances follow a standardised calling convention resulting in identical caller behaviour:
arguments are pushed from right to left (in this case since there are less than 3 arguments, they are instead
moved into registers as per x86 calling conventions) before the subroutine is executed. While this may seem
trivial, the standardisation of calling conventions is an extremely important feature of modern compilers that
can be leveraged to generalise the branch method in a safe and portable manner.
Now that it’s clear that functions with identical signatures observe identical caller setup, and hence the
assembly generated on the callee side will be tailored to reflect this, the next course of action is to make the
branch entry point mimic the identical signature of the branches passed into the constructor. Using class
template deduction, return types and arguments of the branches passed into the constructor can be deduced
at compile time and leveraged to generate the correct assembly code for the branch method.
template <typename Ret , typename . . . Args>

c l a s s BranchChanger
{
using fun c = Ret ( * ) ( Args . . . ) ;
...
BranchChanger ( f unc i f b r a n c h , fu nc e l s e b r a n c h ) ;
...
Ret branch ( Args . . . a r g s ) ;
}
The templating splits the signature into arguments and return type, where the arguments are represented
by a variadic parameter pack, which can be deduced through the pointer types passed into the constructor
(represented with the type alias func for readability). These types are then used to declare the signature of
the branch member function, ensuring that it is identical to the if and else branches. For implementation,
virtually anything can be placed inside branch and as long as the return type matches the signature, it
will compile. Here there are two main cases to disginuish between; void and non-void return types. If the
return type is non-void, returning a brace initialised object of type Ret will suffice for compilation, whereas
void return types must be absent of non-void return values. Using std::is void v<Ret>, we can perform
compile time type checking to circumvent this edge case and always ensure compilation for both void and
non-void return types.
Ret branch ( Args . . . a r g s )

{
i f c o n s t e x p r ( ! s t d : : i s v o i d v <Ret >)
{
13
return Ret { } ;
}
}
Now that the entry point is functional, we can double check the underlying assembly to ensure it has
identical caller behavior to the branches. For this example, the branches used to instantiate the construct
will be addition and subtraction functions with two integer arguments and an integer return type.
lea rax , [ rbp −16]

mov edx , 2
mov esi , 1
mov rdi , rax
call BranchChanger:: branch ( i n t , i n t )
From the demangled function call it appears that the correct template is generated, however additional
instructions have been added on the caller side. In addition to the integer arguments, an effective address is
computed based on an offset from the frame pointer, which is moved into the rdi register after all previous
arguments have been set up. From first glance it may be unclear why this is occurring, however when delving
deeper into C++ calling conventions, what is happening is that an implicit this pointer belonging to the
specific BranchChanger instance is being pushed onto the stack [60]. If the branch member function utilised
data members specific to the parent instance then this would be necessary, however this not the case and
all it does is disrupt register offsets making trampolining to regular functions infeasible. Whilst it may be
possible to rectify this programmatically using inline assembly, a safer and more portable solution would be
to declare the branch method as static since static member functions are not associated with any particular
instance of a class.
mov esi , 2
mov edi , 1
call BranchChanger:: branch ( i n t , i n t )
Static declarations allow the branch entry point to work seamlessly, however this limits the number of
BranchChanger constructs that can be instantiated per function signature. Template specialisation allows
for the differentiation of static branch entry points if the branches passed into the constructor have different
signatures, which allows for multiple instances of semi-static conditions in a single program. However if
more than one BranchChanger instance exists for a specific signature, this means they will share a common
entry point which is problematic: two instances will be performing assembly modification on a single branch
method which will have undefined behaviour. A way to circumvent is would be to alter the return types of
the branches with other built-in or custom types to ensure that different templates are generated. Given
these trade-offs, the priority for development is to work as much as possible with the compiler to avoid
writing inline assembly for safety and portability reasons, and hence the final decision was to keep branch
declarations as static.
Before moving onto implementing assembly modification, an important caveat to consider is compiler
optimisations. From the compilers perspective, the branch method is a small function that does not produce
any meaningful work making it susceptible to inlining or dead code elimination [61]. Obviously, if this
happens then the construct will be unusable, but limiting its usage to programs that are intended to be run
on -O0 defeats the purpose of it being used in high performance applications. The simplest solution would
be to disable optimisations specifically on the branch method, which can be achieved on GCC using pragma
directives or more elegantly using attributes.
a t t r i b u t e ( ( o p t i m i z e ( ”O0” ) ) )
s t a t i c Ret branch ( Args . . . a r g s )
{
{
return Ret { } ;
14
Virtual Address Space
Stack
Executable Pages
Heap
Data mov eax ...
Text
Figure 8: Simplified representation of virtual address space segments, with lower segments residing at lower
memory addresses. Blue segment highlighted in executable page represents an offset where a specific instruction
can be found.
}
}
Assembly Editing Now that the entry point is fully functional, development can start for the core assembly
modification functionality. The target for this editing will the prologue instructions of the branch method,
specific to each instance produced by template specialisation of the parent BranchChanger class.
The first course of action is identifying the addresses of the machine code instructions to edit. In terms of the
virtual address space, the machine code instructions of interest pertaining to the executable reside in the text
segment, which itself is mapped by pages with read-only and execute permissions. Before we can perform
any modifications, the page where the function prologue resides must be located and its permissions must
be changed to read/write, otherwise the processor will raise a segmentation fault if any memory stores are
attempted. Modern operating systems employ address space layout randomisation (ASLR) by randomising
the base address of the virtual address space to prevent attackers from exploiting known memory addresses
in executable, so locating executable pages must be deferred to runtime [62]. Upon construct instantiation,
we can generate a pointer to the template specialised branch method which is essentially the logical address
of the first instruction pertaining to the function. To obtain the address of the page boundary in which the
function resides, we can compute the page offset modding the logical address by the page size of the system
(which can be obtained using the C standard function getpagesize), and then subtracting this offset from
the original address to align it with the lowest multiple of the page size. To change the page permissions,
we can use the mprotect system call to alter the flags of the VMA (virtual memory area, a kernel data
structure which describes a continuous section in the processes memory) corresponding to the page address
previously computed [63].
uint64 t page size = getpagesize ();

a d d r e s s −= ( u i n t 6 4 t ) a d d r e s s % p a g e s i z e ;
mprotect (
address , p a g e s i z e , PROT READ | PROT WRITE | PROT EXEC
15
);
Now that we have located the instructions in memory to edit (through the pointer to the branch function),
and made this editing permissible through altering the page permissions where the function resides, we can
start adding instructions to redirect control flow to the branches. With latency in mind, the scope of control
flow instructions that can be used become limited to direct jumps or calls, which conventionally cannot be
polymorphic without employing assembly editing. On x86 architectures, jumps and calls redirect control
flow by supplying a relative 32-bit offset from the current program counter, which is reduced to a simple
signed displacement arithmetic operation.
00000000000011a9 <foo >:

00000000000011a9 : f 3 0 f 1e f a endbr64
00000000000011ad : 55 push rbp
00000000000011 ae : 48 89 e5 mov rbp , r s p
...
00000000000011e3 : 55 push rbp
00000000000011e4 : 48 89 e5 mov rbp , r s p
00000000000011e7 : e8 bd f f f f f f call 11a9 <foo>
00000000000011 ec : b8 00 00 00 00 mov eax , 0 x0
Above is an example of the machine code generated for a call instruction (achieved with objdump -d -M
intel) along with the program counter (left) and equivalent assembly (right). Focusing on the call instruction
with opcode e8, we can see the following 4-byte displacement encoded as a signed hexadecimal value in
little endian format (architecture specific). The signed 2’s compliment equivalent of this displacement is
-67 bytes, which is 5 bytes smaller than the displacement from the PC of the call instruction (11e7) to the
function entry point (11a9). The reason for this is because the offset of relative jumps are computed from
the last byte of the instruction to the first byte of the target address, which can be formally described with
the equation:
Jump Offset = Target Address - Entry Point - Size of Instruction
Fundamentally, relative jumps are indeed branches and introduce control hazards in pipelined processors,
but have subtly different prediction schemes and penalties on the micro-architectural level in comparison
to conditional jumps and indirect jumps (via register values). Prediction schemes for direct jumps occur
relatively early in the pipeline (processor front-end), whereas branches that have data dependencies (such
as indirect jumps/calls) or are conditional on FLAGS can get deep in the pipeline (back-end) execute stages
before mispredictions are discovered, and hence incur a higher penalty when previous instructions need to
be flushed. For relative branches, the first line of prediction occurs in the branch target buffer (BTB) which
acts as a specialised cache who’s role is to predict weather the PC resolves to a branching instruction, and
if so, what block to fetch next [16]. This is especially important for speculative prefetching: the instruction
prefetcher needs to know in advance which blocks to fetch next, so if the PC is a branch, it can can steer
the prefetcher to the predicted branch target and begin bringing the associated instructions into lower level
caches. If predicted correctly then virtually no penalties are incurred as for conditional and data dependant
branches, however the distinction occurs at the pipeline stage where mispredictions are detected and re-
solved. When an instruction reaches the decode stage, more information is attained surrounding the nature
of the instruction. The branch address calculator (BAC) will ensure that the branches have the correct target
by computing the absolute address of the PC and comparing it with the supplied target. If a direct jump is
mispredicted at this stage, this means that the supplied branch target does not match the predicted, and pro-
ceeding instructions that have been incorrectly fetched will be flushed and the prefetcher will be re-steered
to the correct branch target [64]. In regards to conditional branches, the same applies with the BAC however
mispredictions are ultimately detected later in the execute stages which result in more severe pipeline stalls.
On processors with branch order buffers (BOB), the recovery process can start before the processor pipeline
has been flushed, but nevertheless the relative cost for mispredicted unconditional branches is much lower
than conditional branches [64].
It is clear that relative jump/calls provide the cheapest means of control flow alteration, providing that the
size of the jump does not exceed 232 bytes. The question that remains is which instruction would be most
suitable from a latency and implementation perspective. Call’s and jumps are very similar mechanically,
16
Instruction / Data Caches
Front-end
Fetch Decode Execute Retire
BTB BAC
BPU / Branch Prediction
Figure 9: Simplified representation of the interaction of caches and branch prediction schemes on the instruction
pipeline. Unconditional branch mispredictions are resolved at the decode stage by the BAC, whereas conditional
branches are resolved at the execute stage. Information of retired predicted and mispredicted branches are fed into
the BPU to update the predictor.
with calls being a two-part atomic operation which jumps to an offset while pushing the return address
onto the stack. From a latency perspective, while minimal, the additional use of the call stack pollutes
data caches unnecessarily and the additional return instruction introduces more branching which is prone
to mispredictions. However the main challenge comes from a development perspective; after the call is
complete (within the branch method), the return address will be the proceeding instruction which may
involve manipulating data on the stack/registers which have already been dealt with. Ensuring this does not
occur for all possible signatures is tedious, and may limit future optimisations on the language construct.
Using a jump will be much simpler; we can simply go straight to the branch without needing to ever complete
execution of the entry point, and when the branch has finished executing, control flow will be redirected to
the original calling site of the branch method (recall that a ret instruction is essentially a jmp [reg]).
To implement the jump, the first thing we do is alter the opcode of the first instruction pertaining to the
branch method to e9 (jmp opcode on x86) through its pointer, and then increment it. The following 4 bytes
will be reserved for the relative offsets from the current program counter. To compute the offsets, we simply
use pointer arithmetic to compute the displacement in memory from the branch method to the respective
if/else branches, then subtracting the length of the instruction from the offset as specified in the formula
mentioned earlier. Next, the the integer offsets are converted into a 4-byte representation and stored in two
dimensional member array (total of 8-bytes), accounting for architectural byte ordering.
unsigned char o f f s e t i n b y t e s [DWORD] = {

s t a t i c c a s t <unsigned char >( o f f s e t & 0 x f f ) ,
s t a t i c c a s t <unsigned char >(( o f f s e t >> 8) & 0 x f f ) ,
s t a t i c c a s t <unsigned char >(( o f f s e t >> 16) & 0 x f f ) ,
s t a t i c c a s t <unsigned char >(( o f f s e t >> 24) & 0 x f f )
};
#i f BYTE ORDER == ORDER BIG ENDIAN

change byte ordering ( o f f s e t i n b y t e s );
#e n d i f
17
s t d : : memcpy( d e s t a r r a y , o f f s e t i n b y t e s , DWORD) ;
Changing Branch Directions At this point the development of the construct is nearly complete, leaving
only the direction setting method for development. With the offsets computed and stored in a class member
array, setting the branch direction would simply involve a memcpy of these bytes to the branch pointer as
the destination address. Direction setting must be initially done upon instantiation, since altering the first
byte of the branch entry point will result in a jump to an undefined location, likely causing a segmentation
fault. The class was adapted to have an optional parameter in the constructor to specify the initial direction
of the branch, with the default condition being true. This can be though of similar scheme to compiler
branch prediction hints; the programmer can specify the likely direction which the branch would first be
taken, but will still have the control to change this at any given time. The actual set direction method
will use boolean indexing (the boolean being the runtime condition passed to the method) to access the
bytes pertaining to the correct branch, which are copied into the 4 byte slot next to the jmp opcode. The
boolean indexing approach is simple, and allows for active cache warming if desired.
void s e t D i r e c t i o n ( bool c o n d i t i o n )
{
s t d : : memcpy( d e s t , s r c [ c o n d i t i o n ] , DWORD) ;
}
Concluding Remarks Upon testing, the semi-static conditions prototype appears to work seamlessly for
varying branch signatures on all optimisation levels. Whilst it is not possible to observe the assembly editing
in real time without specialised disassemblers (for example, objdump only shows the contents of the object
file which is not edited, but rather the pages that are mapped by it), using perf record it is possible to
observe it indirectly through the percentage cycles spent in the branch method.
Percent |
| BranchChanger:: branch ( i n t , i n t )
100.00 | push %rbp
| mov %rsp ,% rbp
| mov %edi ,−0x4(%rbp )
| mov %e s i ,−0x8(%rbp )
| mov $0x0,%eax
| pop %rbp
| ret
Above shows the disassembly of the branch method with the percentage cycles spent on each instruction on
the left hand side, obtained using perf record. The branches used in this example are simple addition and
subtraction functions. The data shows that the first instruction within the branch constitutes all the cycles
spent in the function entirely; this is the instruction that is edited to a jump and hence it is expected that this
is the only instruction that executes for the duration of the program. The branches themselves have a small
percentage of cycles spent relative to all other methods in the test program, which supports that the branches
are in fact executed and control flow is redirected accurately, this is also confirmed by simply printing the
return values to the standard output. All the above, along with correct program behaviour, suggest that
the language construct works as intended. Now the core prototype is complete, thought can be given into
optimisations for branch taking and branch setting, as well as additional features that expand from this
core concept (switch statements, also class member functions which observe different calling behaviour than
conventional functions).
3.4 Optimisations
So far, the prototype demonstrates a proof-of-concept, but not a final product. The overarching goal of this
language construct is to provide deterministic (low standard deviation) and low latency branch taking in
scenarios where misprediction rates are high. Although mispredictions on the processor level are expensive,
if the branch-taking component does not perform at a similar level to perfectly predicted branches, the
18
construct will not find use in performance sensitive environments. Prior to running benchmarks, it is crucial
that we level the playing field as much as possible and make semi-static conditions competitive.
Branch Taking Optimisations Naturally the most obvious place to start is the branch-taking method itself.
In the development process, we ensured that the first instruction executed within branch is the jump that
detours execution to one of the branches, so the processor does not waste any time executing instructions
it does not need to. This is fine, however the glaring bottleneck arises from having to prevent all compiler
optimisations on the method, which was implemented as a one-hot fix from preventing the function from
being eliminated. Ideally, the entry point should benefit from all optimisations that regular functions do, but
have the minimum amount of optimisations disabled that prevent the construct from working as intended.
The first obvious approach is do only disable in-lining for the entry point; this is destructive as the compiler
will place the body of the function pre-editing within the calling site which essentially does nothing. Even
if it managed to inline the edited assembly, it will be completely infeasible to target the in-lined instruction
within the code segment. Replacing the compiler attribute on branch with attribute ((noinline))
generated the following assembly under -O3 optimisations.
000000000000118c : lea rdx , [ r i p+0x26d ]

0000000000001193: lea rax , [ r i p+0x136 ]
000000000000119a : sub rax , rdx
000000000000119d : sub rax , 0 x1
00000000000011a1 : mov DWORD PTR [ r i p+0x25a ] , eax
Surprisingly, even with the noinline directive the compiler still reduced the branch call to the lea instruc-
tions highlighted in bold. Upon further research into the effects of GCC compiler attributes, what appears
to be happening is that the inlining prevention does in fact take place, but the compiler deems the function
to have no side effects and as a result optimises out the call completely. A simple way to control this opti-
misation is by inlining assembly within the function body; inline assembly adds uncertainty to the compiler
optimiser as it cannot determine if it has side effects on register or memory values. Adding a simple asm("")
which does not produce any meaningful work is sufficient to prevent optimising out the function call in
addition to using the noinline attribute.
Further testing revealed an interesting yet problematic optimisation, calls to the original branch method
where replaced with calls to a different branch method which the compiler duplicated and altered a number
of instructions within the body. Below is an example of both instances of the method, with demangled
function names simplified for readability.
0000000000001280: 

0000000000001280: ret
0000000000001281: cs nop WORD PTR [ r a x+r a x *1+0x0 ]
0000000000001288: nop
000000000000128b : nop DWORD PTR [ r a x+r a x *1+0x0 ]
...
0000000000001290: < BranchChanger branch>
0000000000001290: endbr64
0000000000001294: xor eax , eax
0000000000001296: ret
The first example of the branch method represents the duplicated form which is called, whereas the sec-
ond example represents the method which is needed to be called to allow semi-static conditions to work.
Inspecting the demangled name of the duplicated function reveals the optimisation that has been applied:
interprocedual constant propagation (ICP). This optimisation is multifaceted; when the compiler recognises
that a function call has some arguments passed as constants, it creates a spot-optimised clone of the func-
tion which can involve removing redundant computations and memory accesses. This can be seen in the
constprop version; the primary instruction becomes a ret because the compiler can see that the branch
method does not produce any meaningful work, the remaining instructions are included as padding to align
the function on a 16-byte boundary. This padding is important especially for procedural calls since most
19
modern processors fetch instructions on aligned 16-32 byte boundaries; fetching code after after an uncon-
ditional jump costs a few clock cycles however this delay is worsened if the branch target does not lie on a
16-32 byte boundary [60]. Following this, an interesting observation can be made in regards to the original
branch method. In many instances, the function itself does not follow alignment and as a result the compiler
seems to always place it at the bottom of the text segment to prevent misalignment of all other procedures
in the executable. Another interesting observation is that the constprop version is often placed close to hot
code paths in the text segment (often very close to main), which can reduce instruction cache fragmentation
by placing contiguous subroutines relatively close to one another.
The issue of preventing constant propagation and function cloning can be easily solved by including the
optimize("no-ipa-cp-clone") attribute in the function header. However prior analysis into the effects
of ICP and procedural reordering opens some interesting avenues into possible improvements. Taking a
page out of the compilers book, the first observation of ICP was the reordering of instructions such that
the most important instructions reside at the function entry point with no wasted work in between. In
the improved branch method, the preliminary instruction is typically a 4-byte endbr64 on Intel CPU’s that
employ control flow enforcement technology (present on Linux with GCC and Clang), which ensures that
indirect jumps/calls can only be made to functions which start in this instruction [65]. In the case of
semi-static conditions, it is highly unlikely that the branch method will find use in indirect calls due to the
associated costs with indirection in general, especially in low latency settings. Given this, overwriting this
preliminary instruction with a 5-byte jump was the direction taken in development, however a key thing
to note is that this editing overwrites the opcode the proceeding instruction given its greater length. While
it is unclear what ramifications this has on variable instruction length pre-fetching, perhaps hard coding a
5-byte jump in the entry point (and editing will not alter the length of the instruction) will be more ”friendly”
towards hardware semantics. In addition this may have positive implications on the BTB; from compile time
the PC associated with the preliminary instruction of the branch method will always be a jump, meaning
that in theory the BTB should always predict that a control flow instruction is present within branch which
can save some cycles associated with mispredictions from preliminary calls. Even if later benchmarks reveal
there is no observable performance gain from doing this, it does defer some work on construct instantiation.
To ensure that a unconditional jump always resides at the branch entry point, the endbr64 instruction will
need to be omitted by the compiler which can be done with the nocf check attribute. Since there is already
an assembly instruction present within branch to prevent the compiler from optimising out the call, this can
be simply changed to asm("jmp 0x0") which hard-codes a jump to an arbitrary 4-byte offset, this will be
edited exclusively. Upon inspecting the disassembly, the branch method starts to resemble its constprop
counterpart even more:

0000000000001280: jmp 0 < a b i t a g −0x38c>
0000000000001285: xor eax , eax
0000000000001287: ret
Interestingly, this assembly ordering seems to be maintained regardless of the function signature; the
compiler seems to understand to not optimise the function call so there is full benefit of caller setup and
teardown, but it also understands that the function does no useful work and optimises accordingly. The only
differences that remain now between the ICP counterpart is 16-byte alignment and procedural reordering.
A simple way to ensure this is including the hot attribute which instructs the compiler to optimise the
function more aggressively and places it in a subsection of the text segment where hot code lies. This is
typically done automatically with the –vprofile-use flag to which the compiler uses profile feedback from
previous executions to determine which functions can benefit from reordering. A caveat with using this
approach is it relinquishes the programmers ability to decide which functions should have priority in the hot
text segment, which may decrease performance depending on the application this is integrated in. However
hot attributes are far more common across compilers than byte-alignment directives, so from a portability
standpoint it would be easier to generate the desired assembly using this method. Given these alterations,
the final disassembly can be seen below:

0000000000001170: jmp 0 < a b i t a g −0x38c>
0000000000001175: xor eax , eax
20
0000000000001177: ret
0000000000001178: nop DWORD PTR [ r a x+r a x *1+0x0 ]
000000000000117 f : nop
The altered version of the branch method, including the hot attribute is also shown below. Note that the
noinline attribute is omitted since no-ipa-cp-clone includes this implicitly, and the function is always
optimised on -O3 to ensure the preliminary instruction is always a jump.
attribute
( ( hot , n o c f c h e c k , o p t i m i z e ( ” no−ipa−cp−c l o n e ” , ”O3” ) ) )
s t a t i c Ret branch ( Args . . . a r g s )
{
asm ( ” jmp 0x00000000 ” ) ;
{
return Ret { } ;
}
}
The improved branch method was benchmarked against the prototype version with various suites, broadly
split into instruction-level benchmarks with in-lined perf events and Intel cycle counters, as well as micro-
benchmarks with google benchmark which involved measurements of more computationally expensive situa-
tions. More detailed methodology is explained in later sections, the purpose of these preliminary benchmarks
is to ensure the proposed changes do not incur adverse effects on the language construct. Broadly on the
instruction level, the improved version improved performance by several cycles across different branches,
with the most signifcant contribution coming from procedual reordering associated with the hot attribute.
For higher level measurements, some benchmarks showed performance gain by 5-10% whereas others had
identical execution times. On the instruction level it is difficult to speculate the source of this performance
gain; at runtime both instances have identical execution pathways in terms of instructions so a reasonable
explanation would be to improved locality between the entry point and branch targets. In larger systems
with increased cache contention, the effects of alignment and locality have more of a prevalent effect on
caching and prefetching which is reflected in some of the larger micro-benchmarks. Given these optimi-
sations showed no adverse effects and showed marginal performance gain in some scenarios, they where
incorporated into the final artefact.
Branch Changing Optimisations The current state of the direction changing method is already in quite an
optimised state. There is an implicit branch for accessing the correct byte offset using a boolean index, how-
ever this is unavoidable. Fundamentally the performance of branch changing is not as important as branch
taking; the whole reason for this separation is to isolate this more expensive operation from performance
critical code to facilitate condition evaluation preemptively.
3.5 Generalisations
The focus of development has been primarily on semi-static conditions that emulate the behaviour of
two-way conditional statements for conventional functions and static class member functions, and has
been success-full thus fair in exploiting calling conventions to facilitate safe branch-taking. Nevertheless,
the current prototype has the capability to be further generalised to work for a larger scope of branches
without needing to alter the core branch-changing logic. Whilst these extensions may not find as much use
for the specialised case (branch optimisation in HFT environments), they will improve the flexibility of the
language construct for more general use cases outside the field of low-latency development.
Class Member Functions The current state of semi-static conditions rely on standardised calling conventions
to facilitate stack setup/teardown and function argument passing by the compiler, without needing to write
inline assembly code. Before extending this to class member functions, one must examine the disassembly
pertaining to these invocations to understand the necessary changes that need to be implemented.
21
0000000000002436: lea rax , [ rbp−0x60 ]

000000000000243a : mov rdi , rax
000000000000243d : call 263a < ZN9SomeClass3fooEv>
0000000000002442: lea rax , [ rbp−0x60 ]
0000000000002446: mov rdi , rax
0000000000002449: call 268a < ZN9SomeClass3barEv>
The underlying assembly shows similar behaviour encountered during the development of the branch entry
point; the effective addresses being computed (highlighted in bold) represents an implicit this pointer to
the parent class which is the first parameter moved onto the stack. In this example both member functions
are being invoked from the same instance, hence the identical offsets represented in the lea instruction.
Propagating this behaviour to the entry point is simply the case of altering the class template to deduce the
member function pointer type, and then updating the signature of the branch method to include the class
instance within the signature, prior to the parameter pack that represents the functions arguments.
template <typename C l a s s , typename Ret , typename . . . Args>

c l a s s BranchChanger
{
using fun c = Ret ( C l a s s : : * ) ( Args . . . ) ;
...
attribute
( ( hot , n o c f c h e c k , o p t i m i z e ( ” no−ipa−cp−c l o n e ” , ”O3” ) ) )
s t a t i c Ret branch ( const C l a s s& i n s t a n c e , Args . . . a r g s ) ;
}
These alterations are sufficient for making semi-static conditions work for non-static member functions with-
out needing to change any optimisations on branch or the core assembly editing logic. This will become a
reoccurring theme in further generalisations. This extension is contrived to work only for member functions
that belong to the same class, which is expected considering the approach used to deduce the class type
through templating. Nevertheless multiple instances are able to share the same entry point and have their
member functions invoked through branch, with support for derived class methods as long as they are not
overloaded (if derived class overloads a method from the base class and the base class pointers are passed
to the constructor, only the base methods will be executed).
Switch Statements The design of the language construct make it seem that generalisation to n-ary condi-
tional statements would be simple; simply change the template parameters and the offset storage array to
reflect the number of branches. However the syntactic requirements complicate template deduction. If it
was possible to alias parameter packs directly, this would be a simple task of deducing the function pointer
signature from a variadic pack of pointers, extracting the return types and arguments externally and aliasing
them from within the class. Unfortunately C++20 does not support this directly, and trying to work around
this by using containers such as std::tuple to hold the arguments is non-trivial, since there is also the task
of extracting these types and forwarding them to either a pointer or the branch template declaration.
The solution to this is to break the BranchChanger class into a base class and derived class. The base class
will be partially specialised to extract the regular or class member function signature (as seen so far), with
the sole purpose of generating the branch method through template deduction and hense locating the bytes
to edit.
template <typename T>

class branch changer aux {};
template <typename Ret , typename . . . Args>

class branch changer aux { . . . };
template <typename C l a s s , typename Ret , typename . . . Args>

class branch changer aux { . . . };
22
The derived class will be the actual BranchChanger which the programmer will interact with. The class itself
has variadic template parameters representing a number of function pointers with identical signatures, these
will be the branches which can range from 2 to ∞. Using std::common type, we can deduce the actual func-
tion signature type from the parameter pack which is used to instantiate the correct base class through CRTP.
template <typename . . . Funcs>

BranchChanger : p u b l i c b r a n c h c h a n g e r a u x
<typename s t d : : common type<Funcs . . . > : : type> { . . . }
The only adaptation needed will be the constructor, which will need to expand the parameter pack and
iterate over all pointers passed to the constructor, computing relative offsets and storing them element-wise.
C++20 supports unpacking these types into into a std::vector directly using brace-initialised fold
expressions:
BranchChanger ( const Funcs . . . f u n c s )

{
using p t r t = typename s t d : : common type<Funcs . . . > : : t y p e
s t d : : v e c t o r pack = { f u n c s . . . } ;
f o r ( i n t i = 0 ; i < pack . s i z e ( ) ; i++) { . . . }
(...)
}
After a bit of hideous template meta-programming, the language construct becomes fully generalised to work
for any number of branches, ans both regular and member functions. Template deduction is completely
abstracted from the programmer without the need of manually writing out types, providing and elegant and
affluent interface for easy use and integration. This concludes the development of semi-static conditions,
the remainder of development time was focused on productionising the construct into a library, with a focus
on portability across different operating systems and compilers. This stage is rather dull and not worth
discussing, most of the complexity arose from creating pre-processor macros to ensure different system calls
and optimisations are enabled based on the users OS and compiler flags specified.
4 Benchmarks and Applications
4.1 Outline
This section is dedicated to benchmarking the core operations that comprise semi-static conditions, explor-
ing the effects self-modifying assembly instructions on performance, and investigating applications in both
HFT and more general use cases. The experiments outlined will leverage a number of benchmarking suites
ranging from Google Benchmark to custom performance counters that offer the fine granularity required to
measure instruction level effects. The proceeding section is dedicated to outlining the experimental methods
employed in obtaining these results for transparency and reproduciblity purposes. All measurements have
been collected on an Intel(R) Core(TM) i7-10700 CPU (2.90GHz) with 256-kilobyte L1 instruction and data
caches, 2-megabyte L2 caches and 16-megabyte L3 caches. Measurements collected will be architecture spe-
cific but have nevertheless been tested on similar architectures with Intel processors and have had consistent
performance patterns with varying numbers. Tests have primarily been focused on semi-static conditions
with regular or static member functions as branches given the large search space that exists with these kind
of experiments.
4.2 Experimental Method
Conventional microbenchmarking frameworks such as Google Benchmark are useful for high level perfor-
mance measurements where test cases are sufficiently long enough to measure observable differences in
latency. However they often fall short for higher resolution measurements involving instruction level micro-
benchmarks. When expected differences in performance manifest at the cycle level, the overhead associated
with running these frameworks in conjunction with timer resolutions and standard errors mean that any
23
observable differences in latency are hidden by background noise, and often results become more influenced
by the measurement taking rather than the actual measurements. This section is dedicated to outlining the
experimental methods employed in capturing these sensitive measurements, based heavily on work done by
Agner Fog, Matt Godbolt, and Intel as part of their microbenchmarking guides for i7 processors [66, 67, 68].
Clock Cycle Measurements Benchmarks for code comprised of small numbers of assembly instructions
where conducted using architecture specific timestamp counters, in this case RDTSC was used. Mea-
surements are taken by reading the processors timestamp counter at two intervals with the code to
benchmark in between, it is important to note that RDTSC counts reference cycles rather than core clock
cycles due to CPU throttling effects. On super-scalar processors instructions are executed out-of-order and
in-parallel to optimise penalties associated with different instruction latencies. This is problematic for such
measurements; there is no guarantee that the RDTSC instruction is called in the precise temporal order
that is specified programtically, and measurements may include other assembly instructions that are not
intended to be measured. To resolve this, serialising instructions can be used in conjunction with RDTSC to
force the CPU to complete all preceding instructions before continuing execution. Examples of serialising
instructions are CPUID and LFENCE, in all tests LFENCE is used as it has a lower overhead and does not
clobber the output registers of RDTSC. An example setup can be seen below, this snippet has been adapted
from the official Intel microbenchmarking guide for i7 processors [68].
mm lfence ( ) ;
uint64 t start = rdtsc ();
mm lfence ( ) ;
code to measure ( ) ;
mm lfence ( ) ;
u i n t 6 4 t end = rdtsc ();
mm lfence ( ) ;
u i n t 6 4 t c y c l e s = end − s t a r t ;
Compiler optimisations often reorder assembly which complicates measurement taking. While there is no
set way to ensure this, a trial and error approach was taken by padding instructions before the measurement
taking and cross checking the disassembly to see if the necessary code is placed between the RDTSC calls.
Measuring taking itself does incur some overhead. To account for this, prior to benchmarking a background
measurement is taken by running the above code with no instructions in between the RDTSC calls for many
iterations (often in the order of 107 ), from which a mean latency is computed and subtracted from all
proceeding benchmarks (excluding outliers). The benchmarks themselves are also run for many iterations
since measurements tend to fluctuate around a mean value after sufficient warm-up time, due to variance in
CPU frequency and individual instruction latencies. Therefore, data collected for benchmarks are processed
as distributions rather than fixed computed values, which is beneficial for reasoning about latency standard
deviations (important for HFT) and observing hardware level effects that contribute to this (e.g. branch
mispredictions).
Profiling CPU performance counters are incredibly useful in identifying sources for particular hot-spots
during program execution, and in the context of this research, identifying granular hardware effects that
contribute to observed latencies. Perf was primarily used for performance profiling. A downside to this
is that perf traditionally profiles the entire executable, rather than small subsets of it, meaning that any
small observable changes hardware counters that are expected become enveloped in the overall noise of
the system. Luckily, linux offers a API to access a subset of perf performance counters inside the executable
using the perf event open system call, allowing for small pieces of code to be profiled in isolation (similar
to RDTSC). Events are set up using the following code:
struct perf event attr attr ;

a t t r . t y p e = PERF TYPE HARDWARE ;
a t t r . c o n f i g = PERF COUNT HW INSTRUCTIONS ;
a t t r . disabled = 0;
a t t r . exclude kernel = 1;
24
(a) Measurement fluctuations (b) Sample distribution

Figure 10: CPU cycle measurements of RDTSC overhead.
a t t r . e x c l u d e i d l e = 1;
a t t r . exclude hv = 1;
a t t r . exclude guest = 1;
The type and config attributes are used to select the performance counters which are generally the only
parameters that are changed between tests. The remaining attributes offer finer control over sampling and
are configured to exclude external noise from measurement taking. The actual profiling code can be shown
below which has been adapted from the linux documentation of perf event:
f d = p e r f e v e n t o p e n (& a t t r , g e t p i d ( ) , −1, −1, 0 ) ;

i o c t l ( fd , PERF EVENT IOC RESET , 0 ) ;
i o c t l ( fd , PERF EVENT IOC ENABLE , 0 ) ;
code to profile ();
i o c t l ( fd , PERF EVENT IOC DISABLE , 0 ) ;

r c = read ( fd , &count , s i z e o f ( count ) ) ;
Whilst this approach offers the best solution to granular profiling, the drawback is that the perf event API
only offers a small subset of performance counters that perf offers. In experiments that utilise profiling,
this inline approach is used when applicable, whilst the command line approach is used when performance
counters that cannot be obtained using perf event are needed. In this case, code is often kept to a minimum
to reduce measurement noise and often run alongside a baseline to extrapolate differences in performance
counters.
Microbenchmarking Frameworks Where applicable, Google benchmark was used to gather latency data for
less fine grained events. Google benchmark automatically configures the number of iterations the benchmark
is run to get a stable estimate.
4.3 Benchmarks
This section is dedicated to benchmarking the branch-changing (set direction) and branch-taking
(branch) methods for semi-static conditions, exploring the effects of self-modifying code and deducing op-
timal usage. Often, measurement distributions do not follow standard distributions due to skewness, so
non-parametric tests are conducted on small subsets of the samples where applicable.
25
Branch-changing Benchmarks This set of tests is concerned with exploring the instances where altering
assembly instructions in memory cause performance degradation and how they can be avoided.
The first test benchmarks the performance of set direction versus an equivalent 4-byte memcpy to non-
executable memory. For fairness, a class was created with identical data members to semi-static conditions
which where initialised with random bytes to represent some runtime deduced data. The set direction
method in the baseline class is identical to the one in BranchChanger.
class Baseline
{
private :
unsigned char * b y t e c o d e ;
unsigned char b y t e s [ 2 ] [DWORD] ;
public :
Baseline ()
{
b y t e c o d e = new unsigned char [DWORD] ;
random bytes ( bytes [ 0 ] ) ;
random bytes ( bytes [ 1 ] ) ;
}
void s e t d i r e c t i o n ( bool c o n d i t i o n )
{
s t d : : memcpy( bytecode , b y t e s [ c o n d i t i o n ] , DWORD) ;
}
};
Interestingly, writing to executable memory on its own does not incur any additional penalties with respect to
the baseline, which is reflected by the near identical distributions in execution latencies. It is understood that
modern processors tolerate self-modifying code (SMC) but are in no way friendly towards it, often initiating
full pipeline and trace cache clears which can cause penalties in the hundreds of cycles [69]. The actual
semantics of how processors detect SMC is unclear, however the general consensus in architecture forums
and patents point towards a ”snooping” mechanism which is initiated by store instructions into executable
memory addresses. These snoops compare physical addresses of in-flight store instructions with entries in
instruction cache-lines to see if the store location corresponds to instructions in executable memory. If there
is an address match, the SMC clear is initiated and new instructions are fetched from memory to lower level
instruction caches [70]. Understanding this the results make sense; since there is no branch-taking occurring
(where SMC occurs) there are no traces of the associated instructions in instruction cache lines, pre-fetch
queues, or i-TLB and hence the physical address check fails to initiate SMC clears.
Following these observations, the next test involved benchmarking the semi-static conditions set direction
method followed by branch-taking, with the baseline having branch replaced with a direct call to one of the
functions passed to the constructor. The goal is to try and trigger SMC machine clears following previous
discussions, and measure their associated penalties.
void i f b r a n c h ( ) { return ; }
void e l s e b r a n c h ( ) { return ; }
(...)
f o r ( i n t i = 0 ; i 0.5)

Figure 11: Benchmark results in CPU cycles for branch-changing overhead versus an equivalent 4-byte memcpy to
non-executable memory.
As expected, the presence of branch-taking close to the branch-chaining method triggers large numbers of
SMC machine clears. The penalty of this is quite severe; extrapolating the execution latencies reveal that
SMC machine clears multiply running times by 30-40x in this benchmark, with approximately 2 clears oc-
curring per iteration on average. The additional overhead of SMC machine clears seems to be approximately
100 cycles which is consistent with approximations made by Agner and Intel, and are in agreement with
various benchmarks made on architectural forums. Nevertheless the machine clear trigger seems to have
deterministic behaviour; the number of SMC clears scale linearly with iterations, which is reflected in the
overall execution latency. It can be said for certain that executing modified assembly instructions relatively
soon after editing initiate these clears, which outlaw the use of semi-static conditions inside tight loops even
if branches are poorly predicted.
Whilst the cost of such machine clears and their trigger are easy to reason about and are deterministic, the
actual segments of assembly where these penalties manifest are not. This is problematic; the total cost of
the branch-changing method are propagated to areas of code outside of itself in the form of SMC penal-
ties, which introduces uncertainty in execution latencies for code that may be performance critical (such as
branch!). An interesting observation is that there seems to a ”lag” period from when the set direction
method is executed to where the SMC cost starts to manifest, assuming that instructions are executed in
the temporal order in which they appear in the disassembly, which is consistent with behaviour observed
27
(a) Execution latency (b) SMC machine clears count

Figure 12: Latency and SMC machine clear count measurements for branch-changing followed by immediate
branch-taking using semi-static conditions.
(a) SMC penalty (M=111, SD=2) (b) Comparison (P=0)

Figure 13: CPU cycle measurements for branch-changing with SMC machine clear penalty.
by Ragab et al. The process of initiating the pipeline snoop to triggering the SMC clear takes several cycles
since it involves instruction stream walks and i-TLB checks, resulting in a transient execution window of
stale instructions caused by the de-synchronisation of the store buffer and instruction queue [71]. Ideally,
there should be some large enough buffer within the branch-changing method to contain this cost exclusively
within set direction for more deterministic execution latencies. From a prevention standpoint, strong se-
rialising instructions such as CPUID and SERIALISE where inserted into set direction to see if they where
capable of preventing the SMC trigger, however this was not the case. In previous discussions we estab-
lished that SMC triggers are caused by comparisons of in-flight or executed store instructions with physical
addresses in instruction caches, so the ineffectiveness of serialising instructions is clear. Whilst CPUID and
SERIALISE force the processor to complete all previous instructions and even drain the store buffer to pre-
vent reordering, they do not have influence on instruction caches from which SMC checks are conducted,
further supporting the presence of such ”snooping” mechanism.
It is possible to manually flush instruction cache lines pertaining to branch using the mm clflush intrinsic
which shows some promise in minimising SMC clears when assembly modification is temporally closer to
branch-taking, reducing such clears to approximately one per edit. Localility in this sence is very important;
the closer assembly editing is to branch-taking, the more prominent the effect of SMC clears since the
stale instructions have now polluted the pipeline, caches and instruction queues and thus require more
28
machine clears to rectify. In the event that branch-changing occurs right before branch-taking, cache flushes
do not seem to have a net positive effect which supports the notion that locality (in terms of instructions
queued from the point to modification to the point where the modified code is executed) of SMC is the
determining factor for the severity of associated penalties. Interestingly, Intel’s optimization manuals do
in fact recommend that SMC should not share the same 1-2Kib sub-page for speculative prefetching and
execution reasons, further supporting this notion [69]. To test this hypothesis further, an artificial buffer
was created which comprised of a cache flush followed by some computational work, acting as a temporal
barrier between assembly modification and branch taking.
void s m c b u f f e r ( )
{
mm clflush ( address of branch ) ;
u i n t 6 4 t b u f f e r [DWORD * 4 ] ;
f o r ( i n t i = 0 ; i < DWORD * 4 ; i++)
b u f f e r [ i ]++;
}
Executing the following code directly after assembly modification is indeed sufficient in halving SMC clears,
the same behaviour observed when adding cache flushes relatively close to branch-taking. The cache flush
prevents additional machine clears due to physical address matches between store instructions and instruc-
tion cache data, whilst the computational work provides a sufficient buffer within the instruction pipeline
to prevent similar matches with in-flight instructions. Usage of such a buffer will be optional to the pro-
grammer, but would also require some sort of active cache warming with branch to ensure that memory
access penalties do not propagate to the hot-path. A simple optimisation to minimise clears would be to
conditionally check if condition passed to set direction is the same to the current direction already set; in
the current state, assembly modification is performed even when it isn’t needed and always initiates machine
clears. The associated overhead with this buffer incorporated within set direction prevents it from being
used tight loops where branch taking also occurs; even if these penalties did not exist, the 9 cycle latency of
copying bytes would likely see no performance benefit even if branches are often mispredicted. Whilst it is
hard to quantify the amount of computational work required between assembly editing and branch-taking to
minimise these penalties, if branch-changing occurs sufficiently far from the hot-path, SMC clears are likely
to be minimal.
Branch-taking Benchmarks This set of tests is concerned with comparing the efficiency of the branch-taking
method with conventional direct function calls, and exploring its implications on the BTB.
The first test investigated the overhead of branch versus a direct function call. The function call latency
served as the baseline and was the same function that was executed through the branch entry point, so the
branch direction was never changed.
void i f b r a n c h ( ) { return ; }
void e l s e b r a n c h ( ) { return ; }
Measurements where taken in the following manner, repeated for 107 iterations.
void measurement ( )
{
start measurement ( ) ;
stop measurement ( ) ;
}
29
(a) Baseline (M=9, SD=1) (b) Branch (M=10, SD=1)
(c) Comparison (P<0.000001)

Figure 14: Benchmark results in CPU cycles for branch-taking overhead versus a conventional direct function call.
Under the same conditions branch-taking has virtually identical overhead to a regular function call, with
equal standard deviations in execution latency. The small difference in overhead may be attributed to the
additional jmp instruction that resides within the prologue of branch, which is the only difference between
the execution pathways of the baseline and semi-static conditions. An experiment to measure the latency
of a single relative jmp on this specific architecture would validate the hypothesis, however this is tedious
and unnecessary. Using Agner’s instruction benchmarks for Intel and AMD CPU’s (note that the CPU that
the benchmarks where run on is part of the 10th generation Comet Lake family, this was unavailable in
Agner’s instruction tables so reference latencies from Ice Lake where used), the typical latency of a relative
jump is 1-2 cycles which falls within the range of the observed differences [72]. Given the large number
of iterations that the benchmark was run and that the branch direction was never changed, it is likely that
the instructions that where measured where hot in lower-level caches with optimal usage of other hardware
semantics, skewing the latency of the additional jmp towards the minimum value, and thus supporting the
hypothesis.
The implications of these results are multifaceted. The current implementation of branch-taking seems to be
optimised to the theoretical limit; the additional overhead incurred seems to be caused exclusively by the
additional control flow instruction that is inserted, which fulfills the overarching goal for its implementation.
In low-latency environments such as HFT, low standard deviations are important for deterministic execution
times and hence are a focus in development. The branch-taking method has shown that it has virtually
30
(a) BAC clears count (b) Overhead of BAC clears

Figure 15: BAC clear counters for continuously changing branch targets (branch) versus static branch targets
(baseline, set direction is always true), total overhead is calculated by subtracting the baseline latency from the
benchmark. Branch (buffered) in (a) represents some computational work between set direction and branch.
identical standard deviations in execution latency as an isolated function call, offering programmers the
assurance of determinism with respect to the current state-of-the-art.
In the development portion of this work, we discussed the difference in prediction schemes for conditional
and unconditional branches. In theory, when the branch direction is changed, the BTB entry corresponding
the jump instruction within the branch prologue becomes invalidated due to an incorrect branch target.
Whilst the BTB will still accurately predict if the PC is indeed a jump, the stale branch target will likely
be detected by the BAC which will initiate a pipeline flush and re-steer the prefetcher. To observe this
behavior, we run the following testing suite with perf stat -e baclears.any:u which counts the number of
BPU front-end re-steers initiated by user-space code:
f o r ( i n t i = 0 ; i < i t e r a t i o n s ; i++)
{
condition = ! condition ;
}
Similar to SMC machine clears, BAC corrections appear to increase linearly with iterations when the branch
direction is continuously changed in deterministic fashion. An interesting observation is that introducing a
buffer of computation between branch-changing and branch-taking calls halves the number of corrections,
averaging one clear per iteration. This suggests that Intel BTB’s are updated atomically upon the retirement
of branch macro-instructions; without the buffer the measurement loop is sufficiently small in terms of
computation such that the modified jmp from the next iteration enters the pipeline before the current jmp
gets retired. Since the BAC does not update BTB entries, but rather re-steers the prefetcher to the correct
branch target, it initiates two re-steers per iteration due to stale BTB entries. The penalty of the correction
also scales linearly, averaging an additional 2.2ns per iteration which equates to approximately 6 cycles on
this particular architecture, less than half the cost of a conditional branch misprediction (presumed around
13 cycles on Skylake CPU’s) [72].
Though less severe, branch-taking using semi-static conditions does in-fact impose misprediction penalties,
however it can be mitigated from the hot-path. The misprediction is essentially a one time cost when
switching branch-directions, once the BTB has been corrected with the updated branch target, all successive
calls branch will incur zero penalties and be executed with minimal latencies. This allows programmer to
do effective ’warming’ in the cold-path which is not possible with conditional statements; branch prediction
is facilitated through capturing histories of branches based on their program counter and correlating them
31
with global patterns. Adding conditional statements in the cold path in an attempt to ’warm’ the BPU for
the identical hot-path code will likely have little since the BPU treats both branches separately based on
PC. Whilst it may capture some correlation, introducing more branches pollutes the global history can also
introduce more noise. Conversely, calling branch in the cold path to warm the BTB works since control
flow is redirected to the PC with the stale jmp, facilitating the correction preemptively and also brings the
associated branch target (the functions being branched to) into lower level instruction caches. This forming
of warming can also be used to isolate the SMC clear penalty discussed in the set direction benchmarks
since the modified code is being executed temporally close to assembly editing. Using HFT as an example,
this can be realised by sending ’dummy orders’ through the branch method after the branch-direction has
been set to ensure warming occurs.
void c o l d p a t h ( )
{
(...)
do some work ( ) ;
branch . branch ( dummy order ) ;
do some more work ( ) ;
(...)
}
Interim conclusions Results obtained from the preceding test suite provide valuable insight into the impli-
cations of using semi-static conditions on the hardware level, which in turn help deduce optimal usage. The
majority of the cost with using this language construct is manifested within the branch changer method in
terms of machine clears caused by self modifying code, whilst branch-taking (the latency critical part) has ex-
tremely low overhead with deterministic execution latencies. In the preliminary sections of this work, it was
hypothesised that optimal usage would involve separating branch-changing and branch-taking into cold and
hot paths respectively, providing a means for pre-empative condition evaluation and branchless execution.
The benchmarking results validate this use case, showing immense promise for branch optimisation in code
paths that are infrequently executed, but contain branches that are poorly predicted. In addition, the tests
brought novel investigations into modern Intel processor semantics, exploring behaviour relating to SMC
and branch-prediction pertaining to unconditional branches, which in turn can help low-latency developers
optimise their code paths respectively.
4.4 Applications
This section builds off the previous findings, exploring applications where semi-static conditions outperforms
the current state of the art: conditional statements with branch prediction. Tests are primarily focused on
HFT applications, which can be applied conceptually to more general use cases.
Hot-path optimisation in HFT In contrary to traditional engineering terminology, the ”hot-path” in the
context of HFT refers to a section of code which is executed relatively infrequently, but when it is executed it
needs to be extremely fast. This is typically the time taken from receiving market data to sending an order,
with the whole process taking just several microseconds to execute. Lying at the heart of the trading system,
and probably the most critical piece of code in terms profitability, this area is the target for most optimisation.
Since it executed relatively infrequently, the remainder of the trading system is often designed with cache
warming measures that keep data used in the critical path hot in low level caches to avoid memory access
penalties. However this poses a challenge for branch prediction; branches in the hot path may have limited
histories with complex patterns, resulting in mispredictions. Here, semi-static conditions are applied to
optimise branch-taking for this use case.
The test suite comprised of a tight measurement loop using RDTSC counters where runtime conditions
where randomly generated using the Mersenne Twister Engine. For an infinite series of randomly generated
booleans, it is expected that roughly half will be true and false, however since measurements are non-infinite
there will likely be skews towards one boolean in benchmarks. To ensure testing is fair, both branches need
to have the same execution latencies. The choice of branches to benchmark initially where 64 byte memory
copies and bit flips to a volatile struct, the scenario representing message passing to custom firmware
32
(such as network cards and FPGA’s) in a HFT system.
a t t r i b u t e (( noinline ))
void s e n d o r d e r ( unsigned char * message )
{
s t d : : copy ( message , message + 64 , FPGA . payload ) ;
FPGA . f l a g = ! FPGA . f l a g ;
}
Emulating the hot-path in terms of infrequent execution is especially challenging from a benchmarking point
of view. Percentage cycles spent in the measurement zone (the hot-path) where used as a proxy, and this
was minimised by adding com-
putational work in the form of randomly generating messages and running pricing calculations outside the
hot-path. The randomly generated messages where passed to the branches in attempt to emulate real-time
message passing, and measurements where taken with and without cache warming.
The results shown are quite remarkable. Across the board semi-static conditions produce a tight uni-modal
distribution with median latencies and standard deviations much lower than conditional branching. In terms
of conditional branches, the sample densities elegantly model the latencies of predicted and mispredicted
branches as a bimodal distribution which contribute to the larger median and standard deviation in cycles.
In tests without cache warming, both conditional branching and semi-static conditions have a small distribu-
tion of measurements in the 70-90 cycle range presumably due to cache effects caused by message passing,
which seems to disappear completely when cache warming is employed. Looking closer at the bimodal na-
ture of conditional statement execution latencies (M=65 SD=2, M=78 SD=2 without cache warming and
M=64 SD=2, M=80 SD=2 with cache warming respectively), the difference in median values between both
distributions is 13-16 cycles which are in good agreement with estimations made by Agner for i7 processors
which seem to have a penalty of 13-18 cycles [66]. Through using semi-static conditions, programmers
gain the benefit branch-taking latencies comparable (in this case even slightly better!) with perfectly pre-
dicted branches, saving 2-4ns on average and up to 6ns if the branch is always mispredicted, with more
deterministic execution latencies manifesting with low standard deviations.
Subsequent tests with other branches reveal the same behaviour for both conditional statements and
semi-static conditions; whilst the distributions where centred around different medians due to the vary-
ing execution of different branches, the relative offsets between the distributions remained the same.
Interestingly usage of the [[likely]] and [[unlikely]] branch prediction hints had no effect in mitigating
misprediction rate for conditional branching. The reason for this is clear, compiler hints simply reorganise
the assembly code to aid the processors static predictor in taking the more likely execution path, but since
conditions are random and neither path is more likely being taken it has no net positive effect. That being
said, it does seem that semi-static conditions are better than branch prediction hints for this use case.
Further tests added more computational work to the hot path to observe this behaviour when branching is
surrounded by more comlex logic, an example of the measuring suite for conditional branches is seen below.
void m e a s u r e h o t p a t h ( )
{
begin measurement ( ) ;
(...)
do some calculations ();
i f ( condition )
s e n d o r d e r ( message ) ;
else
a d j u s t o r d e r ( message ) ;
do some work ( ) ;
(...)
end measurement ( ) ;
}
Cycle distributions seem to follow an identical pattern to the previous results regardless of the additional of
extra computational work. It seems that misprediction cost for this case increased slightly for conditional
33
branches to 18 cycles (M=108 SD=2 for predicted and M=126 SD=4 for mispredicted) which suggested
that branch mispredictions can impact surrounding code resulting in higher penalties.
In the case of n-ary conditional statements in the form of if-else chains or switch statements, unpredictable
conditions yield similar cycle distributions to the preceding examples. Using the current methodology where
random conditions are generated in the range of 0 to n − 1 where n is the number of branches, as n → ∞
the misprediction rate tends to 1 and hence distributions become uni-modal and skewed to higher cycle
numbers. This is entirely expected: if conditions are random and constantly changing, then the probability
of predicting the correct branch is inversely proportional to the number of branches, which tends to zero
as the number of branches tends to infinity. Typically on GCC, large if-else chains or switch statements are
optimised into jump tables, which organise branches at distinct memory locations in a particular structure
(binary tree for example), and utilise indirect jumps to computed offsets to facilitate branch taking. This
additional indirection does incur more significant penalties when mispredicted owing to the additional data
dependencies in the pipeline; extrapolating the median latencies from the predicted (M=13, SD=2) and
mispredicted (M=31, SD=3) branches, the misprediction cost is estimated to be 18 cycles in this example!
It is clear that for this particular use case, semi-static conditions offer superior performance to conditional
branching when misprediction rates are high and branch-changing can be isolated in cold code paths. In
the context of HFT, speed is paramount, and shaving off several nanoseconds per branch in the critical path
offers an edge in order execution latency against competitors and can result in more profitable trading.
General use cases Whilst hot-path optimisation of mispredicted branches is not necessarily tied to HFT, but
can be expanded to general performance critical sections of code in various applications (gaming, aerospace,
infrastructure ect), an interesting investigation would be for general use cases where branches are not neces-
sarily unpredictable (in terms of conditions). Earlier investigations revealed that branch-taking has compa-
rable latency with isolated function calls, and in comparison to conditional statements, have marginally less
assembly to execute to facilitate branch taking. The following investigations outline instances where these
subtle differences manifest in increased performance, using the same methodology of separating branch-
changing from hot-path measurements. To simulate predictable branches, the same test suite used for spe-
cialised applications can be adapted to changing the boolean conditions based on a regular interval, rather
than randomly generating them.
Even when branches are predictable, semi-static conditions are slightly more performant in terms of branch-
taking. In contrast with earlier tests, the latency of conditional statements for, a uni-modal distribution
since the misprediction rate is close to zero, nevertheless the extension of the trend-line to the 80-85ns
region shows there are mispredictions, but are rare. What is fascinating is that the core distribution of
measurements for conditional branching is shifted to higher latencies in comparison to semi-static conditions,
when execution latencies for predicted branches should be the same. The key questions that arise are what is
the cause of this shift, and how significant is the contribution of mispredictions to the overall median latency
at different switching intervals. To answer the former, a subset of measurements for both semi-static and
conditional branches can be extracted and plotted to visualise the execution latencies per iteration.
The results reveal some remarkable behaviour. Firstly, the misprediction penalty can be seen distinctly by
sharp increases in latency at regular 1000 iteration intervals, this is indeed directly caused by changing
the branch direction and seems to be corrected relatively quickly, which is indicative of an n-bit prediction
scheme. Every time the branch direction changes, the median shifts by 2 or 3 cycles forming the observed
the observed saw-tooth pattern. Upon inspection of the underlying assembly, this behaviour is an artefact of
the compiler reordering the conditional statement’s assembly such that the forward branch (if) experiences
slightly lower execution latencies than the backward branch (else) due to additional jumps around the code
segment.
000000000000305 c : mov r d i ,QWORD PTR [ r i p+0x42e5 ]

0000000000003063: shl rdx , 0 x20
0000000000003067: mov r s i , rax
000000000000306a : or r s i , rdx
000000000000306d : cmp BYTE PTR [ r i p+0x42dc ] , 0 x0
0000000000003074: je 3090 < measure+0x40>
0000000000003076: call 2c00 <send order>
(...)
0000000000003090: call 2c40 < a d j u s t p r i c i n g >
0000000000003095: jmp 307b < measure+0x2b>
34
These kind of assembly ordering semantics have even more prevalent effects for switch statements (5-6
cycles faster!) which can be seen by comparing the distributions of semi-static conditions and predicted
branches for switch statements. This is mainly due to the additional assembly required perform the
necessary computations when preparing to traverse jump tables:
00000000000041 f d : shl rdx , 0 x20

0000000000004201: mov rbx , r a x
0000000000004204: or rbx , rdx
0000000000004207: cmp QWORD PTR [ r i p+0x5161 ] , 0 x4
000000000000420 f : ja 4235 < measure+0x45>
0000000000004211: mov rax ,QWORD PTR [ r i p+0x5158 ]
0000000000004218: lea rdx , [ r i p+0x1e05 ]
000000000000421 f : mov r d i ,QWORD PTR [ r i p+0x5142 ]
0000000000004226: movsxd rax ,DWORD PTR [ rdx+r a x * 4]
000000000000422a : add rax , rdx
000000000000422d : jmp * rax
0000000000004230: call 3ed0 < send order 1>
(...)
0000000000004249: nop DWORD PTR [ r a x+0x0 ]
0000000000004250: call 3eb0 < send order n>
0000000000004255: jmp 4235 < measurev+0x45>
0000000000004257: nop WORD PTR [ r a x+r a x *1+0x0 ]
(...)
When comparing this to the underlying assembly generated for semi-static conditions, it becomes very
apparent why branch-taking for the latter exhibits lower and more deterministic execution latencies.
000000000000309d : mov r d i ,QWORD PTR [ r i p+0x42a4 ]

00000000000030a4 : mov rbx , r a x
00000000000030a7 : shl rdx , 0 x20
00000000000030ab : or rbx , rdx
00000000000030ae : call 2b00 < branch >
In the context of general usage for ’static branches’ (conditions that change infrequently), the results have
several implications. The power of semi-static conditions has always been the ability to change the direction
of a branch programatically. At compile time, the programmer has little influence over such re-orderings,
and although an informed prediction can be made about the more likely direction of the branch (forward or
backward), if this where to change at runtime, the programmer will be paying at least 2-3 cycles per iteration
which will start to add up. This allows for the set direction method to be used more freely; rather than
isolating it in the cold path, if the branch direction is changed relatively infrequently then the cost of code
modification will amortise itself over many iterations of cheap branch taking.
In regards to contributions from branch mispredictions, the median latency and standard deviations for
conditional branching where measured for varying branch changing frequencies (in terms of number of iter-
ations passed before the condition is changed), along with the associated misprediction rates. These results
are quite unexpected; even for the branch changing every every iteration one would expect the BPU to spot
this relatively simple pattern of taken not-taken, however this is not the case. Even when the conditions
change at regular intervals in a predictable manner, the organisation of where these instructions lie within
the executable effect the BP’s ability to make predictions based on history correlations. In these benchmarks,
the actual ’path’ that is being measured resides in a function to help maintain assembly ordering for fair and
precise measurements, which is called within a tight loop which performs a significant amount of computa-
tion per iteration. This can add noise, however the more likely reason is that the BP is unable to correlate
the iteration count with the condition being evaluated (at regular intervals). The actual predictive mecha-
nism can be deduced through the results; when branches are changed every iteration, misprediction rates
are close to 100% which results in significantly higher latencies but tighter standard distributions since all
35
branches are mispredicted. As branch changing frequency decreases, the misprediction rate follows as well
as the median latency, but standard deviations increase increase since measurement distributions contain
a mix of predicted and mispredicted branches. From the data, it appears that there are 1.5-2 mispredic-
tions per condition change, which is synonymous with a 2-bit saturated counter prediction scheme (hinting
towards TAGE-like predictors used in modern Intel processors)! When misprediction rates are high, the as-
sociated penalties are the major contributor to increased latencies and standard deviations. When branches
are better predicted and conditions change relatively infrequently, differences in latency are attributed to the
underlying assembly generated by the compiler, resulting in more subtle performance differences.
So far general application benchmarks have been conducted on the cycle level, an interesting investigation
is to see how they manifest in larger systems. The next test involves a multi-threaded benchmark where
branch-directions are changed at regular time intervals, the branches perform a relatively simple computa-
tion and store the result in an array to prevent optimisation. Whilst the example is simplistic, it represents a
system that relies polling events on a worker thread which effect conditions that are evaluated in continuous
loops, an example of such application is feature-flag selection for larger code-bases.
s t a t i c void benchmark ( benchmark : : S t a t e& s)

{
int results [2];
s t d : : t h r e a d worker ( p o l l e v e n t s ) ;
f o r ( auto : s)
{
for ( int i = 0; i < i t e r a t i o n s ; i++)
{
i f ( condition )
r e s u l t s [ f l a g ] += a c t i o n 1 ();
else
r e s u l t s [ f l a g ] += a c t i o n 2 ();
}
}
worker . j o i n ( ) ;
}
Even though the observed changes are relatively small, occasionally they can manifest themselves as large
performance gains, however this will be very dependant on the system in which semi-static conditions is
integrated in. When using the language construct in a multi-threaded environments, there is a chance that
the wrong branch is executed since assembly editing is not thread safe. Using synchronisation will prevent
this, however it results in large performance degradation. Regardless, the proposed language construct
shows immense promise in performance optimisation in general settings; any places where misprediction
rates cause performance bottlenecks or branch-taking is slow due to surrounding code, semi-static conditions
offer a convenient alternative to more efficient branch-taking.
4.5 Summary of Conclusions
The foregoing investigations have not only provided a transparent performance analysis of the various meth-
ods that comprise semi-static conditions, but have been successful identifying numerous use cases where
mispredicted branches can be optimised heavily with the use of the language construct.
By employing a variety of sophisticated testing suites originating from architectural forums and prior lit-
erature, we effectively delved into the underlying factors driving the performance of branch-changing and
branch-taking methods. Concerning branch-changing, the minimal isolated cost of assembly editing was
evident; however, an important discovery was the substantial impact of the code’s locality that undergoes
modification on the hardware penalties induced by SMC detection. The mechanics behind this phenomenon
are well comprehended, along with the associated penalties and strategies for alleviation. This comprehen-
sion led to the realization that, for branch-changing, it is optimal to confine such modifications to less critical,
cold code paths. Leveraging preemptive conditional evaluation, the control over branch directions can be
guided away from performance-sensitive code sections. As for branch-taking, which invariably exists within
the critical path, performance was optimized to the theoretical limit. The only disparities observed where in
additional instruction latencies when compared to conventional calls. While there is a slight initial penalty
36
for branch-taking after assembly modification, it was discerned that this setback could be mitigated through
BTB warming, a possibility not readily achievable in the same manner with conditional branch prediction.
Using these newfound insights to compare against the current state-of-the-art practices, it becomes evident
that under circumstances involving occasional branch execution combined with frequent mispredictions,
the approach of semi-static conditions emerges as the preferred choice. The strategy of employing semi-
static conditions demonstrates advantages in terms of both execution speed and consistency across all such
scenarios, yielding the possibility of notable performance improvements in real-world application due to
the influence of neighboring logic. In a broader context, the proposed methodology once again displays
potential by offering slightly improved performance even when branches are largely predictable. However,
more substantial discrepancies in performance become apparent in more extensive conditional constructs,
which could translate to significant gains in larger-scale systems.
5 Evaluation
This section will evaluate the software contribution based on the prototype that we have developed, along
with the experimental methods employed to accurately benchmark the efficiency of semi-static checks. We
then proceed to evaluate the overall approach taken to developing semi-static checks in comparison to al-
ternative viable solutions. We then analyse the safety of the language construct through a combination of
static and runtime analysis approaches, these methods are also employed to focus more on reliability in
terms of behaviour and synchronisation. We then touch on portability to different architectures and operat-
ing systems, and focus on the experimental methods employed for efficiency benchmarks and outlines more
appropriate tests for industry settings.
5.1 Overall Approach
The goal of semi-static checks is to provide programmers control over conditional branching, reaping the
performance benefit of direct method invocation whilst being able to change which method is called at run-
time based on a condition. This kind of behaviour cannot be attained through conventional means available
in the C++ standard; particular ”branches” can be generated at compile-time based on conditions through
template instantiation, but the ability to do this with runtime generated conditions is fundamentally impossi-
ble since templates are a compile-time phenomenon. Mechanisms for switching between function calls exist
without explicit conditional branching exist in the form virtual inheritance facilitated by v-table look-ups
and function pointer de-referencing. However in reality this is slower than conditional branching in most
scenarios even when branches are often mispredicted, and in the case for virtual functions, a good optimiser
will often speculatively de-virtualise calls which reduce to conditionals. At the hardware level, these types of
indirect calls require not only predictions for conditions but also data in the form of memory addresses (for
example, a v-table lookup takes roughly 3 memory accesses before the call address is resolved), and as a
result suffer from higher misprediction penalties and are more vulnerable to adverse cache-related effects in
oppose to regular conditional statements which have a smaller instruction cache footprint.
This lack of flexibility meant that only assembly editing can facilitate the desired behaviour: fast determin-
istic branch-taking controlled by the programmer through a slower auxiliary interface. The reality of this
is multifaceted. SMC is non-standard compliant and is widely considered a poor programming practice
owing to complex maintenance, debugging and portability, despite it being used abundantly in debuggers
and the Linux kernel. Given these concerns, the goal of development was to minimise the assembly editing
component whilst maximising the simplicity of design for maintenance and portability reasons. However
this is not the only viable approach.
Run-time code generation (JIT) An alternative approach to SMC facilitated branch-changing is a just-in-
time (JIT) approach, which has become popular in modern compilers and interpreters. Whilst JIT is generally
more widely accepted form of runtime assembly manipulation, there are some inherent issues that make it
inferior to our approach. The first is complexity. We show the process of preparing the byte-code associated
with the branches to actually executing them, with the majority of the complexity residing in the first two
stages. Copying the byte-code is error prone: whilst it is simple to find the preliminary instructions for the
function, being able to accurately determine when the function is ”finished” in memory requires architecture
specific complex logic. For example, using a ret opcode as a proxy would not work as there can be multiple
exit points, and differentiating an opcode from a byte offset would require accounting for instruction length
which introduces a substantial amount of administrative overhead. Then comes the challenge of adjusting
37
position dependant instructions within this byte-code to work, such as PC relative instructions, which further
increases development complexity. The best way to do this would be to have an in house JIT compiler as
part of the language construct, which is impractical.
Even if hypothetically the former challenges where addressed, the eminent problem that would prevent JIT
from being used in low-latency settings is the means of which it can be executed in modern C++.
void * new page = mmap( n u l l p t r , 4096 ,
PROT READ | PROT WRITE | PROT EXEC ,
MAP PRIVATE | MAP ANONYMOUS,
);
s t d : : memcpy( new page , bytecode , s i z e o f ( b y t e c o d e ) ) ;
(...)
i n t ( * add ) ( i n t , i n t ) = ( i n t ( * ) ( i n t , i n t ) ) new page ;
i n t r e s u l t = add ( 5 , 3 ) ;
Above is a sample of two necessary components required required to facilitate JIT-style assembly execution
in C++: memory allocation and execution. From user space, the only way to execute this newly generated
byte-code is through casting the address of the newly created executable page to a function pointer and
defencing it, which is much more expensive at runtime than conditional branching. Whilst runtime code
generation could offset SMC machine clears since instructions are not being modified and executed, rather
just generated, the overhead of pointer de-referencing in latency critical paths defeats the whole purpose of
the language construct.
Assembly editing In comparison to JIT, the method used to develop semi-static conditions has many advan-
tages. The first is clearly simplicity as runtime assembly editing using an intermediary trampoline function
(branch method) is facilitated through a 4-byte memcpy at runtime. The simplicity of the concept directly
translates to simplicity in development which is reflected by the overall size of software artefact. In terms
of maintainability, which is important for a library which employs architecture specific optimisations, a sim-
ple implementation will easier adapt to future changes. From a low-latency application perspective, less
code translates to less assembly instructions which have smaller instruction cache footprints, a favourable
property for systems that prioritise keeping as much latency-critical code in lower level caches. The second
advantage is branch-taking speed; using the assembly editing approach the branch-taking overhead is simply
the overhead of a relative jump, which has superior prediction schemes from pointer de-referencing in JIT.
A downside of the current approach is SMC penalties incurred during branch direction changing. Even in
a hybrid approach where the trampoline can be hard-coded on separate executable pages, editing existing
assembly is unavoidable. Whilst avoiding SMC penalties seems impossible using this scheme on modern ar-
chitectures, it does open some interesting investigations in mitigation schemes, but nevertheless the trade-off
for superior branch-taking ought to be favourable for these sorts of applications.
5.2 Safety
Assembly editing gives rise to undefined behaviour with respect to the C++ standard, and can bring forth
security vulnerabilities. The first issue that can arise is that one or more of the branches lies at a signed
displacement greater than 232 bytes from the jmp instruction within the branch method. In this instance
the program will redirect control flow to an area in the code segment which does not belong to any of the
branches, resulting in adverse behaviour. Instances like this ought to be caught out as early as possible. To
simulate this, two arbitrary function pointers with a displacement greater than 232 are created and passed
into the BranchChanger constructor as such:
using p t r = i n t ( * ) ( i n t , i n t ) ;
ptr func 1 ;
p t r f u n c 2 = r e i n t e r p r e t c a s t (
r e i n t e r p r e t c a s t ( f u n c 1 ) +
( s t a t i c c a s t (1) << 34)
);
BranchChanger branch ( func 1 , f u n c 2 ) ;
38
As a result the following runtime exception is raised upon instantiation:
terminate called after throwing an instance of ’branch changer error’
what(): Supplied branch targets (as function pointers) exceed a 2GiB displacement from the entry
point in the text segment, and cannot be reached with a 32-bit relative jump. Consider moving the
entry point to different areas in the text segment by altering hot/cold attributes.
Aborted (core dumped)
Another form of undefined behaviour stems from more than one instance of semi-static conditions be-
ing present at any given time during program execution. In this scenario, the conflicting instances of
BranchChanger will share the same branch entry point due to template specialisation and compete for
assembly editing of a single jmp. As a result, branches that do not belong to the immediate BranchChanger
instance may be executed. When multiple instances do exist, the following error is raised:
terminate called after throwing an instance of ’branch changer error’
what(): More than once instance for template specialised semi-static conditions detected. Program
terminated as multiple instances sharing the same entry point is dangerous and results in undefined
behaviour (multiple instances write to same function.
Aborted (core dumped)
The former exception handling measures are effective at detecting the main behaviours associated with
assembly editing that can result in detrimental effects later on in program execution. This, along with other
higher level behaviour exceptions, have been integrated into a rigorous automated testing suite part of the
library, allowing programmers to test the build prior to usage for peace of mind.
In the context of security, the primary vulnerability that arises using semi-static conditions is through chang-
ing executable page permissions to read/write/execute. Often times during program execution, multiple
pages and hence multiple address ranges are vulnerable to memory modification from other processes,
allowing attackers to inject their own code, access sensitive data or tamper with sensitive information. From
a mitigation standpoint, eliminating this vulnerability completely is not possible, since page permissions
need to be altered at least temporarily to facilitate assembly editing. What can be done is minimise the
transient window of vulnerable address ranges by performing page permission modification and reversion
exclusively within the set direction method, in this case an auxiliary version named set direction safe.
void s e t d i r e c t i o n s a f e ( bool c o n d i t i o n )
{
i f ( c o n d i t i o n != d i r e c t i o n )
{
c h a n g e p a g e p e r i m i s s i o n s ( bytecode ,
PROT READ | PROT WRITE | PROT EXEC
);
s t d : : memcpy( s r c , d e s t [ c o n d i t i o n ] , DWORD) ;
c h a n g e p a g e p e r i m i s s i o n s ( bytecode ,
PROT READ | PROT EXEC
);
direction = condition ;
}
}
39
Using this method, programmers can be ensured that the risk of such exploits are minimised. To ensure
that safe-mode is respected throughout the program lifetime, programmers can add the following flag upon
compilation:
−DSAFE MODE
However this comes with a cost. In a low-latency setting, using the secure branch-changing method intro-
duces higher execution times and larger standard deviations owing to the two system calls used to alter
page permissions, in addition to increased pressure on instruction-caches. When thinking of general cases,
the more expensive cost of branch-changing has negative implications on amortisation, and will likely limit
the use in scenarios where branches are well predicted. This is the inherit trade off in low-latency settings;
security versus speed. In commercial software that is exposed directly to the user, such security ought to
be included, however in the case of secretive proprietary software (such as trading systems) such security
measures may not be necessary. In light of this, both API’s are exposed to the programmer, whom can make
an informed decision what to include.
5.3 Reliability
In terms of reliability, the main areas of focus in the context of semi-static conditions are correctness and
consistency of behaviour. From a high-level perspective, an exhaustive formal proof that shows semi-static
conditions has equivalent behaviour to conditional statements is unnecessary. In the context of semi-static
conditions, static analysis of program behaviour does not bear much meaning since the program state itself
is not static due to self-modifying assembly instructions! To evaluate correctness a test suite can be set up
by running a tight loop that (1) changes the branch direction and (2) takes the branch successively and see
if the wrong branch is taken based on the runtime condition.
while ( run )
{
branch . branch ( c o u n t e r ) ;
condition = ! condition ;
}
The branches will have a similar function prologue which determines if the correct branch is executed using
logical expressions:
void t r u e b r a n c h ( i n t& c o u n t e r )
{
c o u n t e r += ( c o n d i t i o n ˆ tr ue ) ;
(...)
}
Under these conditions, in a single threaded environment semi-static conditions will always exhibit correct
behaviour, however this is not always the case in multi-threaded setting. Fundamentally, what is occurring
are write operations to memory (producer) which are executed by the CPU (consumer) in some interleaved
order giving rise to race conditions. The extent of how often the wrong branch is executed is heavily depen-
dant upon how often context switching occurs followed by branch taking and how large the branches are in
terms of assembly instructions, which is often very difficult to predict. For general use cases, wrong branches
are executed relatively rarely (close to 0), whereas more chaotic systems where context switches occur every
several microseconds or so result in much more inconsistent behaviour.
The absence of thread safety without synchronisation is a key drawback of the assembly editing approach.
With the use of synchronisation, correct behaviour is assured however performance degrades significantly
which can be seen from the examples. Although the overall focus in multi-threaded environments where
40
limited, it would be interesting to observe if serialising instructions such as lfence and mfence can be used
as a synchronisation mechanism unilaterally without compromising the performance of branch-taking in
future investigations.
5.4 Usage and Portability
The software artifact is packaged as a static library which can be incorporated into project using the CMake
build system. The decision to package semi-static checks as a static library over a dynamic library is due
to performance reasons; static libraries present all code in the executable at compile time whereas dynamic
libraries need to be loaded by the OS at runtime. Using the dynamic library methods require an extra layer of
indirection through symbol table look-ups, and parts of the library themselves are brought into memory on-
demand which is prone to cache misses and page faults, all of which contribute to jitter which is undesirable
in low-latency setting. The drawbacks of using static libraries are increased compilation times and the size
of the executable, which is made more prominent given the template-heavy implementation of semi-static
conditions.
Cross platform builds on C++ are notoriously messy. CMake abstracts the differences between various
build systems and compilers on different platforms which simplifies development and usage greatly. From
a portability standpoint, the CMake pre-processor can be configured to selectively include architectural
specific headers into the final build which is extremely beneficial given the low-level nature of the library.
To begin the build (assuming the library is cloned into the same directory), users simply first specify a build
directory as such:
$ cmake -E make directory ”build”
Next, the build system files can be generated in the newly created directory with
$ cmake -E chdir ”build” cmake ../
Then finally built with
$ cmake –build ”build”
Subsequent tests can be run to validate the build system using
$ cmake -E chdir ”build” ctest
To use the library, the generated archive must be compiled and linked against the branch library (lib-
branch.a) as such
$ g++ mycode.cpp -std=c++17 -isystem semi-static-conditions/include -Lsemi-static-

conditions/build -lbranch -o mycode
This necessitates the entire setup required to use the library, the only requirements being a relatively new
version of CMake (version 3.2 and above) to facilitate the building! The core BranchChanger construct can
be used through including the following header:
#i n c l u d e <branch . h>
Portability was a challenge from a development perspective. The implementation of semi-static checks
utilises OS specific system calls, compiler specific attributes, inline assembly, and architecture specific offset
41
Windows MAC Linux

Compiler
x86-64 ARM x86-64 ARM x86-64 ARM
GCC ✓ ✗ ✗ ✗ ✓ ✓
MSVC ✓ ✗ ✗ ✗ ✓ ✗
Clang ✓ ✗ ✗ ✗ ✓ ✓
Table 1: Compatibility matrix for the semi-static conditions library. Ticks are given to platforms where the library
has been tested and is functional, and crosses are given to untested platforms and/or no functionality.
computations which can all manifest in different combinations depending on the users system configuration.
In terms of achieving the correct build, a dedicated single header file was utilized to define the platform
using a series of pre-processor directives. These directives are employed throughout the library to enable the
incorporation of architecture-specific, compiler-specific, and OS-specific code. It is possible to delegate this
process entirely to CMake, however such an approach would necessitate users to specify the platform using
a series of compiler flags which is rather tedious, and therefore avoided.
For this initial version, popular compilers and operating systems where targeted for compatibility, on both
x86-64 and ARM architectures. Builds where tested by invoking the library using the CMake steps shown
above on different systems, results of the tests (including if the library was unable to tested) are shown
in Table 5.1. Unfortunately, semi-static conditions is not portable on Apple silicon owing to the Hardened
Runtime OS security feature which prevents page permissions from being changed to write/execute. Some
of these security features can be disabled allowing for binary editing of runtime-allocated pages using the
JIT approach, however as discussed, does not meet the latency requirements for this construct.
5.5 Experimental Method
A great focus of this work was to evaluate the performance of semi-static conditions against the current state
of the art and rationalise the observed behaviour and inefficiencies observed from a micro-architectural per-
spective. Due to the sensitivity of the measurements, careful consideration was taken in developing the mea-
suring suite that prioritises the accuracy and reproducibility of measurements. The choice of measurement
instruments where sensible; reference cycle counters where sufficient in producing accurate measurements
allowing for the observation of subtle behaviours on the hardware level owing to its low usage overhead.
However a number of improvements can be made to the overall production setup in the context of OS tuning
and networking.
Given the primary objective of this work, which centers around exploring innovative branch optimization
techniques in a HFT context, it becomes imperative that the OS chosen for conducting performance bench-
marks mirrors an industrial setup to the greatest extent possible. The selection process for the system was
driven by a preference for a Linux-based High-Performance Computing (HPC) environment featuring a Intel
i7 processor. Additionally, a cloud-based Virtual Machine (VM) with an Intel Xeon chip was utilized. It is,
however, important to note that in both cases, complete root privileges were not attainable. This limitation
had significant implications, particularly in the domain of kernel tuning. Specific adjustments related to CPU
scaling and scheduling could not be replicated to the same extent as they would be in a genuine HFT system.
This discrepancy is noticeable in the variations and distributions observed in the recorded measurements.
To address this challenge, a nuanced approach was adopted. The measurement suite was meticulously tai-
lored to execute numerous iterations, allowing the CPU to stabilize at an optimal operating frequency. This
strategic approach effectively minimized measurement variance, enabling subsequent interpretation through
statistical tests. While it remains true that with meticulous kernel tuning, the necessity for extensive statis-
tical treatment could be substantially diminished, leading to clearer and more refined data, the outcomes
presented successfully fulfilled the intended communication goals. Despite the existing constraints, the data
provided here aptly conveyed the essential insights and findings.
The second improvement is of a more comprehensive nature, and its attribution extends to the entire system.
An inherent limitation associated with microbenchmarking in isolation, particularly when examining minute
differences at the low-nanosecond scale, is that the tests themselves possess an artificial nature, deviating
from the essence of a genuine production environment. Despite the endeavors to craft test suites that emulate
a pseudo-realistic production environment in terms of computational workload, the actual representation of
behavioral changes remains somewhat imprecise when contrasted with a fully functional High-Frequency
42
Trading (HFT) system. Yet, it’s important to acknowledge the inherent limitations in addressing this issue.
The proprietary nature of trading firms’ code renders it a closely guarded trade secret, largely due to its direct
influence on profitability. The epitome of an ideal experimental arrangement, a concept initially proposed by
Carl Cook as the most robust methodology for micro-benchmarking latency enhancements within a trading
system [5]. It’s worth noting, however, that such a setup warrants its own comprehensive analysis and
entails substantial complexity and associated costs.
6 Conclusions
We have shown that software-level branch optimisation can be achieved using a novel language construct,
semi-static conditions, offering superior execution latencies to the current state of the art in both specialised
and generalised scenarios. There is no doubt that the methodologies applied to facilitate this optimisation
are unconventional. The use of assembly editing has long been shunned from a software development and
micro-architectural perspective, however it seems that revisiting this old yet interesting nuance of low-level
development has opened new doors for low-latency optimisation.
The notion of semi-static checks is simple, separate branch-taking from condition evaluation, however as we
have seen throughout this work this is quite a surface level interpretation. What really is happening is a trade
of branch-prediction schemes on the hardware level which is facilitated through assembly editing. In semi-
static checks, the conditional statement is re-engineered as polymorphic relative jump instruction which can
be controlled directly by the programmer through an auxiliary interface. As an artefact of this modification,
the idea of execute-stage branch mispredictions are completely eliminated, and replaced with less severe
BAC correction ocurring earlier in the decode stage. Unlike conditional branching, BAC corrections caused
by stale BTB entries can be mitigated from latency critical code through pre-empative branching, a new type
of active ’BPU warming’ that guarantees a BTB correction upon instruction retirement and circumvents any
associated penalties for the current branch direction indefinitely. The result? A powerful decoupling that
optimises branch-taking to the theoretical limit. Our results showcased the superior performance of semi-
static conditions against conditional branching when often mispredicted, with lower execution latencies and
more deterministic standard deviations, an desirable property for HFT. Even when branches are predicted
well, semi-static conditions show faster branch-taking as a consequence of a cleaner control path and fewer
instructions on the machine code level, allowing the expensive branch-changing method to be amortised in
single and multi-threaded scenarios.
The phenomenon of semi-static checks brings forth many avenues for further investigation, particularly
around the application of self-modifying binaries for program optimisation, and minimising any adverse
hardware effects. We took a detailed look at the behaviour of writing instructions into executable mem-
ory and revealed that locality between assembly editing and the assembly being executed results in severe
processor penalties in the form of SMC machine clears. Though it is understood that locality in terms of
caching, prefetching and paging has a proportional effect on the severity of SMC penalties, this investigation
was unable to properly quantify the effect. Further investigations into understanding this would allow the
branch-changing portion of the construct to be further optimised and make it more flexible to use cases,
in addition this would be beneficial from a fundamental perspective in understanding the effects of SMC
fully. In addition, the general idea of making targeted and granular changes in assembly to optimise vari-
ous language level intrinsics should be investigated more broadly. Perhaps the general idea of substituting
more BPU-friendly control flow instructions and modifying them can be applied to many different types of
branches, more specifically dynamic dispatching and any form of register jumps.
In sum, the absence of language level-optimisations for hardware based branch prediction has birthed a
new toolkit in tackling optimisation problems in low-latency setting. With this work pioneering the use
of assembly editing for branch optimisation, we hope it serves as a benchmark for future investigations
that apply this long-neglected and interesting programming practice to more complex problems surrounding
optimisation.
References
[1] Lin CK, Tarsa SJ. Branch prediction is not a solved problem: Measurements, opportunities, and future
directions. arXiv preprint arXiv:190608170. 2019. pages 2, 7
[2] Aldridge I. High-frequency Trading: A Practical Guide to Algorithmic Strategies and Trading Systems.
Wiley trading series. John Wiley & Sons, Incorporated; 2013. Available from: https://books.google.
43
co.uk/books?id=kHPwjgEACAAJ. pages 2, 10
[3] Cartea Á, Jaimungal S, Penalva J. Algorithmic and High-Frequency Trading. Cambridge University
Press; 2015. Available from: https://books.google.co.uk/books?id=5dMmCgAAQBAJ. pages 2
[4] Core C++ 2019 :: Nimrod Sapir :: High Frequency Trading and Ultra Low Latency development
techniques;. Available from: https://www.youtube.com/watch?v=_0aU8S-hFQI. pages 2, 10
[5] CppCon 2017: Carl Cook “When a Microsecond Is an Eternity: High Performance Trading Systems in
C++; 2017. Available from: https://www.youtube.com/watch?v=NH1Tta7purM. pages 2, 9, 10, 43,
53
[6] CppCon 2014: Chandler Carruth ”Efficiency with Algorithms, Performance with Data Structures”;
2014. Available from: https://www.youtube.com/watch?v=fHNmRkzxHWs. pages 2
[7] Trading at light speed: designing low latency systems in C++ - David Gross - Meeting C++ 2022;
2022. Available from: https://www.youtube.com/watch?v=8uAW5FQtcvE&t=2875s. pages 2
[8] Branchless Programming in C++ - Fedor Pikus - CppCon 2021; 2021. Available from: https://www.
youtube.com/watch?v=g-WPhYREFjk&t=1723s. pages 2
[9] Stroustrup B. An overview of C++. In: Proceedings of the 1986 SIGPLAN workshop on Object-oriented
programming; 1986. p. 7-18. pages 2, 8
[10] Stroustrup B. The design and evolution of C++. Pearson Education India; 1994. pages 2, 8
[11] Hunt G, Brubacher D. Detours: Binaryinterception ofwin 3 2 functions. In: 3rd usenix windows nt
symposium; 1999. . pages 2
[12] Giffin JT, Christodorescu M, Kruger L. Strengthening software self-checksumming via self-modifying
code. In: 21st Annual Computer Security Applications Conference (ACSAC’05); 2005. p. 10 pp.-32.
pages 2
[13] GDB: The GNU Project Debugger; 2023. Available from: https://www.sourceware.org/gdb/. pages
2
[14] Eyerman S, Smith JE, Eeckhout L. Characterizing the branch misprediction penalty. In: 2006 IEEE
International Symposium on Performance Analysis of Systems and Software; 2006. p. 48-58. pages 2
[15] Making Your Code Faster by Taming Branches; 2023. Available from: https://www.infoq.com/
articles/making-code-faster-taming-branches/. pages 2, 9
[16] Hennessy JL, Patterson DA. Computer architecture: a quantitative approach. Elsevier; 2017. pages 3,
4, 5, 6, 16
[17] Noyce R, Hoff M. A History of Microprocessor Development at Intel. IEEE Micro. 1981 jan;1(01):8-21.
pages 3
[18] S TANENBAUM A, Bos H. Modern operating systems; 2015. pages 3
[19] Moore GE, et al.. Cramming more components onto integrated circuits. McGraw-Hill New York; 1965.
pages 3
[20] Stallings W. Computer organization. Architecture,:, Designing for performance. 2017;10. pages 4
[21] McFarling S, Hennesey J. Reducing the cost of branches. ACM SIGARCH Computer Architecture News.
1986;14(2):396-403. pages 4, 5
[22] Smith JE, Sohi GS. The microarchitecture of superscalar processors. Proceedings of the IEEE.
1995;83(12):1609-24. pages 4
[23] Deitrich BL, Chen BC, Hwu W. Improving static branch prediction in a compiler. In: Proceedings. 1998
International Conference on Parallel Architectures and Compilation Techniques (Cat. No. 98EX192).
IEEE; 1998. p. 214-21. pages 4
[24] Emma. Characterization of branch and data dependencies in programs for evaluating pipeline perfor-
mance. IEEE Transactions on Computers. 1987;100(7):859-75. pages 4, 5
[25] Flynn MJ. Computer Architecture Pipelined and Parallel Processor Design, 1995. Jones and Bartiett
Publishers. pages 5
[26] Pan ST, So K, Rahmeh JT. Improving the accuracy of dynamic branch prediction using branch corre-
lation. In: Proceedings of the fifth international conference on Architectural support for programming
languages and operating systems; 1992. p. 76-84. pages 5, 6
44
[27] Fisher JA, Freudenberger SM. Predicting conditional branch directions from previous runs of a pro-
gram. ACM SIGPLAN Notices. 1992;27(9):85-95. pages 5
[28] Mittal S. A survey of techniques for dynamic branch prediction. Concurrency and Computation: Prac-
tice and Experience. 2019;31(1):e4666. pages 6, 7
[29] Smith JE. A study of branch prediction strategies. In: 25 years of the international symposia on
Computer architecture (selected papers); 1998. p. 202-15. pages 6
[30] Yeh TY, Patt YN. Two-level adaptive training branch prediction. In: Proceedings of the 24th annual
international symposium on Microarchitecture; 1991. p. 51-61. pages 6
[31] Yeh TY, Patt YN. Alternative Implementations of Two-Level Adaptive Branch Prediction. 1992;20(2).
Available from: https://doi.org/10.1145/146628.139709. pages 6
[32] Seznec A. TAGE-SC-L branch predictors again. In: 5th JILP Workshop on Computer Architecture
Competitions (JWAC-5): Championship Branch Prediction (CBP-5); 2016. . pages 6, 7
[33] Zangeneh S, Pruett S, Lym S, Patt YN. Branchnet: A convolutional neural network to predict hard-
to-predict branches. In: 2020 53rd Annual IEEE/ACM International Symposium on Microarchitecture
(MICRO). IEEE; 2020. p. 118-30. pages 6, 7, 8
[34] Sadooghi-Alvandi M, Aasaraai K, Moshovos A. Toward virtualizing branch direction prediction. In:
2012 Design, Automation & Test in Europe Conference & Exhibition (DATE). IEEE; 2012. p. 455-60.
pages 7
[35] Seznec A. A 256 kbits l-tage branch predictor. Journal of Instruction-Level Parallelism (JILP) Special
Issue: The Second Championship Branch Prediction Competition (CBP-2). 2007;9:1-6. pages 7
[36] Ozturk C, Sendag R. An analysis of hard to predict branches. In: 2010 IEEE International Symposium
on Performance Analysis of Systems & Software (ISPASS). IEEE; 2010. p. 213-22. pages 7
[37] Jimenez DA, Lin C. Dynamic branch prediction with perceptrons. In: Proceedings HPCA Seventh
International Symposium on High-Performance Computer Architecture; 2001. p. 197-206. pages 7
[38] Burcea I, Somogyi S, Moshovos A, Falsafi B. Predictor virtualization. ACM SIGOPS Operating Systems
Review. 2008;42(2):157-67. pages 8
[39] Donadio S, Ghosh S, Rossier R. Developing High-Frequency Trading Systems: Learn how to implement
high-frequency trading from scratch with C++ or Java basics. Packt Publishing; 2022. Available from:
https://books.google.co.uk/books?id=HBp2EAAAQBAJ. pages 8
[40] Veldhuizen TL. C++ templates are turing complete. Available at citeseer ist psu edu/581150 html.
2003. pages 8
[41] Vandevoorde D, Josuttis NM. C++ Templates: The Complete Guide, Portable Documents. Addison-
Wesley Professional; 2002. pages 8
[42] Abrahams D, Gurtovoy A. C++ template metaprogramming: concepts, tools, and techniques from
Boost and beyond. Pearson Education; 2004. pages 9
[43] Clang Language Extensions; 2023. Available from: https://clang.llvm.org/docs/
LanguageExtensions.html. pages 9
[44] Other Built-in Functions Provided by GCC; 2023. Available from: https://gcc.gnu.org/onlinedocs/
gcc/Other-Builtins.html. pages 9
[45] Fog A. The microarchitecture of Intel, AMD and VIA CPUs: An optimization guide for assembly pro-
grammers and compiler makers. Copenhagen University College of Engineering. 2012;2. pages 9
[46] Smith AJ. Cache Memories. ACM Comput Surv. 1982 sep;14(3):473–530. Available from: https:
//doi.org/10.1145/356887.356892. pages 9
[47] Improving Performance by Better Code Locality; 2023. Available from: https://easyperf.net/blog/
2018/07/09/Improving-performance-by-better-code-locality. pages 9
[48] Memory Part 5: What Programmers Can Do; 2023. Available from: https://lwn.net/Articles/
255364/. pages 9
[49] C++ attribute: likely, unlikely; 2023. Available from: https://en.cppreference.com/w/cpp/
language/attributes/likely. pages 9
[50] Menkveld AJ. High frequency trading and the new market makers. Journal of Financial Markets.
2013;16(4):712-40. High-Frequency Trading. Available from: https://www.sciencedirect.com/
science/article/pii/S1386418113000281. pages 10
45
[51] The race to zero: Speech held at the International Economic Association Sixteenth World Congress.
Bank of England; 2011. . pages 10
[52] MacKenzie D. A sociology of algorithms: High-frequency trading and the shaping of markets. Preprint
School of Social and Political Science, University of Edinburgh. 2014. pages 10
[53] Arnoldi J. Computer algorithms, market manipulation and the institutionalization of high frequency
trading. Theory, Culture & Society. 2016;33(1):29-52. pages 10
[54] Hasbrouck J, Saar G. Low-latency trading. Journal of Financial Markets. 2013;16(4):646-79.
High-Frequency Trading. Available from: https://www.sciencedirect.com/science/article/pii/
S1386418113000165. pages 10
[55] Zhang F. High-frequency trading, stock volatility, and price discovery. Available at SSRN 1691679.
2010. pages 10
[56] Brogaard J, Hendershott T, Riordan R. High frequency trading and the 2008 short-sale ban. Journal of
Financial Economics. 2017;124(1):22-42. pages 10
[57] Brogaard J, et al. High frequency trading and its impact on market quality. Northwestern University
Kellogg School of Management Working Paper. 2010;66. pages 10
[58] Boutros A, Grady B, Abbas M, Chow P. Build fast, trade fast: FPGA-based high-frequency trading using
high-level synthesis. In: 2017 International Conference on ReConFigurable Computing and FPGAs
(ReConFig); 2017. p. 1-6. pages 11
[59] Leber C, Geib B, Litz H. High frequency trading acceleration using FPGAs. In: 2011 21st International
Conference on Field Programmable Logic and Applications. IEEE; 2011. p. 317-22. pages 10
[60] Fog A. Calling conventions for different C++ compilers and operating systems. Technical University
of Denmark; 2023. pages 14, 20
[61] Knoop J, Rüthing O, Steffen B. Partial dead code elimination. ACM Sigplan Notices. 1994;29(6):147-
58. pages 14
[62] Aga MT, Austin T. Smokestack: thwarting DOP attacks with runtime stack layout randomization. In:
2019 IEEE/ACM International Symposium on Code Generation and Optimization (CGO). IEEE; 2019.
p. 26-36. pages 15
[63] Corbet J. Kernel Development - Multi-protection VMA’s. 2006. Available from: https://lwn.net/
Articles/182495/. pages 15
[64] Corp I. Method for pipeline processing of instructions by controlling access to a reorder buffer using a
register file outside the reorder buffer. 1996. Available from: https://patents.google.com/patent/
US5721855A/en. pages 16
[65] Intel. A Technical Look at Intel’s Control-flow Enforcement Technology. 2020. Avail-
able from: https://www.intel.com/content/www/us/en/developer/articles/technical/
technical-look-control-flow-enforcement-technology.html. pages 20
[66] Fog A. Test programs for measuring clock cycles and performance monitoring. Technical University of
Denmark; 2023. pages 24, 33
[67] Godbolt M. Microarchitecture; 2016. Available from: https://xania.org/
Microarchitecture-archive. pages 24
[68] Gabriele Paoloni IC. How to Benchmark Code Execution Times on Intel® IA-32 and IA-64 Instruction
Set Architectures; 2010. Available from: https://www.intel.com/content/dam/www/public/us/en/
documents/white-papers/ia-32-ia-64-benchmark-code-execution-paper.pdf. pages 24
[69] Corporation I. Intel® 64 and IA-32 Architectures Optimization Reference Man-
ual; 2012. Available from: https://www.intel.com/content/dam/doc/manual/
64-ia-32-architectures-optimization-manual.pdf. pages 26, 29
[70] Corporation I. Method and apparatus for self modifying code detection using a translation lookaside
buffer; 1999. Available from: https://patents.google.com/patent/US6594734. pages 26
[71] Ragab H, Barberis E, Bos H, Giuffrida C. Rage against the machine clear: A systematic analysis of
machine clears and their implications for transient execution attacks. In: 30th USENIX Security Sym-
posium (USENIX Security 21); 2021. p. 1451-68. pages 28
[72] Fog A. Instruction tables. Technical University of Denmark; 2023. Available from: https://www.
agner.org/optimize/instruction_tables.pdf. pages 30, 31
46
(a) Conditional statements without cache warming (b) Semi-static conditions without cache warming
(M=75, SD=10) (M=63, SD=3)
(c) Conditional statements with cache warming (d) Semi-static conditions with cache warming
(M=68, SD=8) (M=62, SD=2)
(e) Comparison without cache warming (f) Comparison with cache warming
Figure 16: CPU cycle measurements of conditional branching (branch) versus semi-static conditions (branchless)
with and without cache warming.
47
(a) Conditional statements (M=120, SD=10) (b) Semi-static conditions (M=104, SD=3)
(c) Comparison
Figure 17: CPU cycle measurements for conditional branching versus semi-static conditions with updated hot-path.
48
(a) Switch statements (M=30, SD=8) (b) Semi-static conditions (M=8, SD=1)
(c) Comparison
Figure 18: CPU cycle measurements for a 5 case switch statement versus semi-static conditions with unpredictable
conditions in the hot-path. Empty functions used as branches for ease of testing.
49
(a) Conditional statements (M=64, SD=3) (b) Semi-static conditions (M=62, SD=2)
(c) Comparison (P<0.000001)

Figure 19: CPU cycle measurements for conditional branching versus semi-static conditions with predictable
branching, conditions change every 1000 iterations.
50
(a) Conditional statements
(b) Semi-static conditions

Figure 20: CPU cycle measurements per iteration for conditional branching and semi-static conditions.
51
(a) Latency (b) Misprediction rates

Figure 21: CPU cycle measurements and respective misprediction rates for conditional branching at varying branch
changing frequencies (number of iterations per condition change).
(a) Without std::mutex (b) With std::mutex

Figure 22: Benchmark results for semi-static conditions versus conditional branching in multi-threaded application
where branches change at regular time intervals. Mutex’s have been applied to semi-static conditions exclusively.
52
set direction branch
Allocate Execute
Process and
Copy bytecode executable
adjust offsets bytecode
memory
Branches
Write bytecode Free memory
to memory
Figure 23: State diagram of semi-static conditions internals for both setup and execution using JIT-style runtime
assembly generation.
Figure 24: Ideal productionised setup of benchmarking suite for HFT applications. Left server replays market data
across a high speed ethernet cable, the switch in the middle is equipped with high precision time-stamps, server
on the right is the production system to be measured, and server in the middle computes response times. Adapted
from [5].
53

S - C L - C++ H F T: B B P H: EMI Static Onditions in OW Latency FOR IGH Requency Rading Etter Than Ranch Rediction Ints

Uploaded by

Copyright:

Available Formats

S - C L - C++ H F T: B B P H: EMI Static Onditions in OW Latency FOR IGH Requency Rading Etter Than Ranch Rediction Ints

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

S - C L - C++ H F T: B B P H: EMI Static Onditions in OW Latency FOR IGH Requency Rading Etter Than Ranch Rediction Ints

Uploaded by

Copyright:

Available Formats

S EMI - STATIC C ONDITIONS IN L OW- LATENCY C++ FOR H IGH

Paul Alexander Bilokon Maximilian Lucuta Erez Shermer

Department of Computing Department of Computing qSpark LLC

August 27, 2023

1.1 Aims and Approach

1.2 Research Context

1.4 Summary of Achievements

Untaken branch instruction IF ID EX MEM WB

Taken branch instruction IF ID EX MEM WB

Instruction i + 1 IF idle idle idle idle

2.2 Dynamic Branch Prediction

Predict taken Predict taken

00010 Taken Not taken

hist[L1:0] hist[L2:0] hist[LN:0]

Index Tag Index Tag Index Tag

Tag match? Tag match? Tag match?

2.3 C++ and Compiler Hints

i f ( condition ) [[ unlikely ]] 10 a f : je 10bd

proach, known as ”template metaprogramming,” is extensively employed in low-latency settings such as

2.4 High Frequency Trading

3.2 Semi-static Conditions

UDP Network Stack

3.3 Prototype Development

template <typename Ret , typename . . . Args>

Ret branch ( Args . . . a r g s )

lea rax , [ rbp −16]

Virtual Address Space

Data mov eax ...

uint64 t page size = getpagesize ();

00000000000011a9 <foo >:

Jump Offset = Target Address - Entry Point - Size of Instruction

Instruction / Data Caches

Fetch Decode Execute Retire

BPU / Branch Prediction

unsigned char o f f s e t i n b y t e s [DWORD] = {

#i f BYTE ORDER == ORDER BIG ENDIAN

000000000000118c : lea rdx , [ r i p+0x26d ]

0000000000001280: < B r a n c h C h a n g e r b r a n c h . c o n s t p r o p . 0 . i s r a . 0 >

0000000000001280: < BranchChanger branch>

0000000000001170: < BranchChanger branch>

0000000000002436: lea rax , [ rbp−0x60 ]

template <typename C l a s s , typename Ret , typename . . . Args>

template <typename T>

template <typename Ret , typename . . . Args>

template <typename C l a s s , typename Ret , typename . . . Args>

template <typename . . . Funcs>

BranchChanger ( const Funcs . . . f u n c s )

4 Benchmarks and Applications

4.2 Experimental Method

struct perf event attr attr ;

(a) Measurement fluctuations (b) Sample distribution

f d = p e r f e v e n t o p e n (& a t t r , g e t p i d ( ) , −1, −1, 0 ) ;

code to profile ();

i o c t l ( fd , PERF EVENT IOC DISABLE , 0 ) ;

(a) Baseline (M=9, SD=1) (b) Branch (M=9, SD=1)

(c) Comparison (P>0.5)

(a) Execution latency (b) SMC machine clears count

(a) SMC penalty (M=111, SD=2) (b) Comparison (P=0)

(a) Baseline (M=9, SD=1) (b) Branch (M=10, SD=1)