Sign In

Communications of the ACM

Research highlights

Technical Perspective: Building Bug-Free Compilers


View as: Print Mobile App ACM Digital Library Full Text (PDF) In the Digital Edition Share: Send by email Share on reddit Share on StumbleUpon Share on Hacker News Share on Tweeter Share on Facebook

Software engineersand, by extension, anyone who uses the software they create (that is, nearly everyone)rely critically on compilers. These ubiquitous tools, familiar even to the most novice programmer, translate high-level ideas, expressed as code, into the low-level instructions understood by computer hardware.

Using compilers is so commonplace and transparent that most of us soon forget about the near-miracles of code analysis and optimization they perform at a keystroke. Today's compilers produce marvelously tuned and optimized code that almost always performs better than what could be achieved painstakingly by hand.

Optimizing compilers are an essential part of our computing infrastructure.

To achieve these near-miracles, compilers themselves are very complex beasts, comprising tens- or hundreds-of-thousands of lines of code, developed over many years by many people. And, just as with any software, they too are potentially buggy. A broken compiler might produce erroneous machine code, thereby converting the programmer's intended algorithm into gibberish, or, worse, into a subtly flawed product that sometimes gets the wrong answer or has a pitfall waiting for just the right input to be triggered.

Fortunately, thanks to their frequent use, compilers tend to be very well-tested software: they do their job correctly most of the time for most programs. Nevertheless, compiler bugs do existas demonstrated by Regehr and his collaborators in prior work using their Csmith tool. They can often be provoked by surprisingly small (but tricky!) programs that involve the corner-cases of fixed-bitwidth arithmetic calculations, which are incorrectly optimized when generating low-level code. These kinds of compiler bugs are potentially catastrophic, notoriously difficult for programmers to diagnose, and require real expertise to track down and fix.

So, what do we do about them?

The authors of the following paper give us a compelling and practical answer. Their main idea is to raise the level of abstraction used by compiler implementors, allowing them to express optimizations by using a domain-specific language suited perfectly for the task.

The authors explain their ideas in the context of the LLVM intermediate representation, which, owing to the widespread use of LLVM in both academia and industry (most prominently by Apple), makes the results in this paper immediately applicableindeed the authors found numerous bugs in the LLVM implementation, and have been working with LLVM developers to find more.

The paper focuses on peephole optimizations, in which the compiler rewrites a short sequence of instructions into a more efficient sequence, for instance, by replacing multiplication by bitshift and addition operations. The traditional way to implement this kind of thing in LLVM is for the compiler implementor to manually write a bunch of C++ code that case-analyzes the LLVM instructions looking for a specific pattern that should be replaced.

Why is it difficult for compiler implementors to get this right? At a superficial level, such code consists of a lot of fiddly boilerplate, and, when there are hundreds of peephole optimizations, as in an industrial-strength compiler like LLVM, bugs are bound to creep in.

But the problem is actually much deeper. One difficulty is the semantics for even "simple" arithmetic operations is not so straightforward. These operations include many corner cases: underflows, overflows, and undefined behaviors, all of which must be properly accounted for to ensure an optimization is correct. Moreover, not all optimizations apply in all situations. The compiler might need to know facts about the program (for example, that a variable is nonzero) to ensure an optimization applies. When hand-implementing an optimization, the programmer must carefully follow all of these constraints.

The paper describes Alive, a tool that realizes the domain-specific language for describing peephole optimizations. Instead of writing an optimization by hand, the compiler writer expresses it using natural LLVM syntax. Alive generates a C++ implementation, thereby eliminating the ugly boilerplate. More importantly, Alive also verifies the proposed optimization is indeed correct: the tool understands the semantics of LLVM (including the tricky corner cases), and checks that the optimized code agrees with the original. When an optimization is broken, Alive produces a counterexample showing the problem. The tool can even infer parts of the optimization, which lets compiler writers be more aggressive about producing good code, without the fear of bugs.

To show that Alive works in practice, the authors reimplemented hundreds of peephole optimizations from the LLVM suite and stress-tested them to look for miscompilation bugs. They found none.

This paper is not the end of the story for compiler correctnessthere are other kinds of optimizations, analyses, and transformations not covered herebut it does provide a large step forward by showing how to build a tool that is practical yet backed by modern verification technology. In short, this paper suggests a way to make better compilers that are easier to build and far less buggy than they are today.

Back to Top

Author

Steve Zdancewic is a professor in the Department of Computer Science and Information Science at the University of Pennsylvania, Philadelphia, PA, USA.

Back to Top

Footnotes

To view the accompanying paper, visit doi.acm.org/10.1145/3166064


Copyright held by author.
Request permission to (re)publish from the owner/author

The Digital Library is published by the Association for Computing Machinery. Copyright © 2018 ACM, Inc.


 

No entries found