\section{Introduction}

High-level synthesis (HLS), which refers to the automatic translation of software into hardware, is becoming an important part of the computing landscape, even in such high-assurance settings as financial services~\cite{hls_fintech}, control systems~\cite{hls_controller}, and real-time object detection~\cite{hls_objdetect}.
The appeal of HLS is twofold: it promises hardware engineers an increase in productivity by raising the abstraction level of their designs, and it promises software engineers the ability to produce application-specific hardware accelerators without having to understand Verilog or VHDL.

As such, we are increasingly reliant on HLS tools. But are these tools reliable? Questions have been raised about the reliability of HLS before; for example, Andrew Canis, co-creator of the LegUp HLS tool, wrote that ``high-level synthesis research and development is inherently prone to introducing bugs or regressions in the final circuit functionality''~\cite[Section 3.4.6]{canis15_legup}. In this paper, we investigate whether there is substance to this concern by conducting an empirical evaluation of the reliability of several widely used HLS tools.

The approach we take is \emph{fuzzing}.
%To test the trustworthiness of HLS tools, we need a robust way of generating programs that both have good coverage and also explores various corner cases. 
%Therein lies the difficulty in testing HLS tools. 
%Human testing may not achieve both these objectives, as HLS tools are often require complex inputs to trigger wrong behaviour. 
%In this paper, we employ program fuzzing on HLS tools. 
This is an automated testing method in which randomly generated programs are given to compilers to test their robustness~\cite{fuzzing+chen+13+taming,fuzz+sun+16+toward,fuzzing+liang+18+survey,fuzzing+zhang+19,yang11_findin_under_bugs_c_compil,lidbury15_many_core_compil_fuzzin}.
The generated programs are typically large and rather complex, and they often combine language features in ways that are legal but counter-intuitive; hence they can be effective at exercising corner cases missed by human-designed test suites.
Fuzzing has been used extensively to test conventional compilers; for example, Yang \textit{et al.}~\cite{yang11_findin_under_bugs_c_compil} used it to reveal more than three hundred bugs in GCC and LLVM. 
%\JW{Clang or LLVM?}\YH{Hard to say actually, they just mention Clang, but I believe they go hand in hand.  If it was optimisations, it is likely LLVM, but Clang is still the front end compiler.} \JW{To my mind, Clang is part of LLVM, so a bug in Clang is necessarily a bug in LLVM. So I think we should say LLVM throughout.}
In this paper, we bring fuzzing to the HLS context.
%We specifically target HLS by restricting a fuzzer to generate programs within the subset of C supported by HLS.


% Most fuzzing tools randomly generate random C programs that are then provided to the compiler under test. 
% Furthermore, fuzzing tools can be configured by users to generate and avoid particular patterns, which is important since HLS tools typically support a subset of C.

% Compiler Fuzzing is a popular technique to find bugs in programs, and it is especially effective at finding compiler bugs, as   This technique can therefore also be used to find bugs in HLS tools.  There has also been some prior work in trying to find bugs in HLS tools and also ensuring that synthesis tools output a correct design.

% \NR{add a sentence about fuzzing}

% % Our method is brought over from the compiler testing literature. 
% Program fuzzing is bla.. 
% Fuzzing enables us to overcome

\begin{example}[A miscompilation bug in Vivado HLS]
\label{ex:vivado_miscomp}
Figure~\ref{fig:vivado_bug1} shows a program that produces the wrong result during RTL simulation in Xilinx Vivado HLS v2018.3, v2019.1 and v2019.2.\footnote{This program, like all the others in this paper, includes a \code{main} function, which means that it compiles straightforwardly with GCC. To compile it with an HLS tool, we rename \code{main} to \code{result}, synthesise that function, and then add a new \code{main} function as a testbench that calls \code{result}.} The program repeatedly shifts a large integer value \code{x} right by the values stored in array \code{arr}.
Vivado HLS returns \code{0x006535FF}, but the result returned by GCC (and subsequently confirmed manually to be the correct one) is \code{0x046535FF}. The bug was initially revealed by a randomly generated program of around 113 lines, which we were able to reduce to the minimal example shown in the figure.  We reported this issue to Xilinx, who confirmed it to be a bug.\footnote{\url{https://bit.ly/3mzfzgA}} 
\end{example}

\begin{figure}[t]
\begin{minted}{c}
unsigned int x = 0x1194D7FF;
int arr[6] = {1, 1, 1, 1, 1, 1};

int main() {
  for (int i = 0; i < 2; i++) 
    x = x >> arr[i];
  return x;
}
\end{minted}
  \caption[Miscompilation bug in Xilinx Vivado HLS. The generated RTL returns \code{0x006535FF} but the correct result is \code{0x046535FF}.]{Miscompilation bug in Xilinx Vivado HLS. The generated RTL returns \code{0x006535FF} but the correct result is \code{0x046535FF}.}
  \label{fig:vivado_bug1}
\end{figure}

The example above demonstrates the effectiveness of fuzzing. It seems unlikely that a human-written test-suite would discover this particular bug, given that it requires several components all to coincide before the bug is revealed. If the loop is unrolled, or the seemingly random value of \code{b} is simplified, or the array is declared with fewer than six elements (even though only two are accessed), then the bug goes away.

Yet this example also begs the question: do bugs found by fuzzers really \emph{matter}, given that they are usually found by combining language features in ways that are vanishingly unlikely to happen `in the real world'~\cite{marcozzi+19}. This question is especially pertinent for our particular context of HLS tools, which are well-known to have restrictions on the language features they support. Nevertheless, although the \emph{test-cases} we generated do not resemble the programs that humans write, the \emph{bugs} that we exposed using those test-cases are real, and \emph{could also be exposed by realistic programs}.

%Moreover, it is worth noting that HLS tools are not exclusively provided with human-written programs to compile: they are often fed programs that have been automatically generated by another compiler.
Ultimately, we believe that any errors in an HLS tool are worth identifying because they have the potential to cause problems -- either now or in the future. And problems caused by HLS tools going wrong (or indeed any sort of compiler for that matter) are particularly egregious, because it is so difficult for end-users to identify whether the fault lies with their design or the HLS tool.

\subsection{Our approach and results}

Our approach to fuzzing HLS tools comprises three steps. 
First, we use Csmith~\cite{yang11_findin_under_bugs_c_compil} to generate thousands of valid C programs within the subset of the C language that is supported by all the HLS tools we test. We also augment each program with a random selection of HLS-specific directives. Second, we give these programs to four widely used HLS tools: Xilinx Vivado HLS~\cite{xilinx20_vivad_high_synth}, LegUp HLS~\cite{canis13_legup}, the Intel HLS Compiler, also known as i++~\cite{intel20_sdk_openc_applic}, and finally Bambu~\cite{pilato13_bambu}. Third, if we find a program that causes an HLS tool to crash or to generate hardware that produces a different result from GCC, we reduce it to a minimal example with the help of \creduce{}~\cite{creduce}.

Our testing campaign revealed that all four tools could be made to generate an incorrect design.  In total, \totaltestcases{} test-cases were run through each tool, of which \totaltestcasefailures{} failed in at least one of the tools.  Test-case reduction was then performed on some of these failing test-cases to obtain at least \numuniquebugs{} unique failing test-cases, detailed on our companion webpage: \begin{center}
    \url{https://ymherklotz.github.io/fuzzing-hls/}
\end{center}
    
To investigate whether HLS tools are getting more or less reliable, we also tested three different versions of Vivado HLS (v2018.3, v2019.1, and v2019.2).  We found fewer failures in v2019.1 and v2019.2 compared to v2018.3, but also identified a few test-cases that only failed in v2019.1 and v2019.2; this suggests that new features may have introduced bugs.

In summary, the overall aim of our paper is to raise awareness about the reliability (or lack thereof) of current HLS tools, and to serve as a call-to-arms for investment in better-engineered tools. We hope that future work on developing more reliable HLS tools will find our empirical study a valuable source of motivation.

%  we test, and then augment each program with randomly chosen HLS-specific directives. We synthesise each C program to RTL, and use a Verilog simulator to calculate its return value. If synthesis crashes, or if this return value differs from the return value obtained by executing a binary compiled from the C program by GCC, then we have found a candidate bug. We then use trial-and-error to reduce the C program to a minimal version that still triggers a bug.

% We have tested three widely used HLS tools: LegUp~\cite{canis13_legup}, Xilinx Vivado HLS~\cite{xilinx20_vivad_high_synth}, and the Intel HLS Compiler~\cite{?}. For all three tools, we were able to find valid C programs that cause crashes while compiling and valid C programs that cause wrong RTL to be generated. We have submitted a total of \ref{?} bug reports to the developers, \ref{?} of which have been confirmed and \ref{?} of which have now been fixed at the time of writing. 
% We hope that our work serves to stimulate efforts to improve the quality of HLS tools.

%%% Local Variables:
%%% mode: latex
%%% TeX-master: "main"
%%% End: