2013 TheChallengeofCrossLanguageInte

(Chisnall, 2013) ⇒ David Chisnall. (2013). “The Challenge of Cross-language Interoperability.” In: Queue Journal, 11(10). doi:10.1145/2542661.2543971

Subject Headings: Programming Language, Object Model.

Notes

Cited By

Quotes

Abstract

Interfacing between languages is increasingly important.

Introduction

Interoperability between languages has been a problem since the second programming language was invented. Solutions have ranged from language-independent object models such as COM (Component Object Model) and CORBA (Common Object Request Broker Architecture) to VMs (virtual machines) designed to integrate languages, such as the JVM (Java Virtual Machine) and CLR (Common Language Runtime). With software becoming ever more complex and hardware less homogeneous, the likelihood of a single language being the correct tool for an entire program is lower than ever. As modern compilers become more modular, there is potential for a new generation of interesting solutions.

In 1961 the British company Stantec released a computer called the ZEBRA, which was interesting for a number of reasons, not least of which was its data flow-based instruction set. The ZEBRA was quite difficult to program with the full form of its native instruction set, so it also included a more conventional version, called Simple Code. This form came with some restrictions, including a limit of 150 instructions per program. The manual helpfully informs users that this is not a severe limitation, as it is impossible that someone would write a working program so complex that it would need more than 150 instructions.

Today, this claim seems ludicrous. Even simple functions in a relatively low-level language such as C have more than 150 instructions once they are compiled, and most programs are far more than a single function. The shift from writing assembly code to writing in a higher-level language dramatically increased the complexity of programs that were possible, as did various software engineering practices.

The trend toward increased complexity in software shows no sign of abating, and modern hardware creates new challenges. Programmers in the late 1990s had to target PCs at the low end that had an abstract model a lot like a fast PDP-11. At the high end, they would have encountered an abstract model like a very fast PDP-11, possibly with two to four (identical) processors. Now, mobile phones are starting to appear with eight cores with the same ISA (instruction set architecture) but different speeds, some other streaming processors optimized for different workloads (DSPs, GPUs), and other specialized cores.

The traditional division between high-level languages representing the class that is similar to a human's understanding of the problem domain and low-level languages representing the class similar to the hardware no longer applies. No low-level language has semantics that are close to a programmable data-flow processor, an x86 CPU, a massively multithreaded GPU, and a VLIW (very long instruction word) DSP (digital signal processor). Programmers wanting to get the last bit of performance out of the available hardware no longer have a single language they can use for all probable targets.

Similarly, at the other end of the abstraction spectrum, domain-specific languages are growing more prevalent. High-level languages typically trade generality for the ability to represent a subset of algorithms efficiently. More general-purpose high-level languages such as Java sacrifice the ability to manipulate pointers directly in exchange for providing the programmer with a more abstract memory model. Specialized languages such as SQL make certain categories of algorithms impossible to implement but make common tasks within their domain possible to express in a few lines.

You can no longer expect a nontrivial application to be written in a single language. High-level languages typically call code written in lower-level languages as part of their standard libraries (for example, GUI rendering), but adding calls can be difficult.

In particular, interfaces between two languages that are not C are often difficult to construct. Even relatively simple examples, such as bridging between C++ and Java, are not typically handled automatically and require a C interface. The Kaffe Native Interface4 did provide a mechanism for doing this, but it was not widely adopted and had limitations.

The problem of interfacing between languages is going to become increasingly important to compiler writers over the coming years. It presents a number of challenges, detailed here.

Object Model Differences

Object-oriented languages bind some notion of code and data together. Alan Kay, who helped develop object-oriented programming while at Xerox PARC, described objects as "simple computers that communicate via message passing." This definition leaves a lot of leeway for different languages to fill in the details:

• Should there be factory objects (classes) as first-class constructs in the language?

• If there are classes, are they also objects?

• Should there be zero (e.g., Go), one (e.g., Smalltalk, Java, JavaScript, Objective-C), or many (e.g., C++, Self, Simula) superclasses or prototypes for an]]object]]?

• Is method lookup tied to the static type system (if there is one)?

• Is the data contained within an object of static or dynamic layout?

• Is it possible to modify method lookup at runtime?

The question of multiple inheritance is one of the most common areas of focus. Single inheritance is convenient, because it simplifies many aspects of the implementation. Objects can be extended just by appending fields; a cast to the supertype just involves ignoring the end, and a cast to a subtype just involves a check — the pointer values remain the same. Downcasting in C++ requires a complex search of the inheritance graph in the run-time type information via a runtime library function.

In isolation, both types of inheritance are possible to implement, but what happens if you want, for example, to expose a C++ object into Java? You could perhaps follow the .NET or Kaffe approach, and support direct interoperability with only a subset of C++ (Managed C++ or C++/CLI) that supports single inheritance only for classes that will be exposed on the Java side of the barrier.

This is a good solution in general: define a subset of one language that maps cleanly to the other but can understand the full power of the other. This is the approach taken in Pragmatic Smalltalk:5 allow Objective-C++ objects (which can have C++ objects as instance variables and invoke their methods) to be exposed directly as if they were Smalltalk objects, sharing the same underlying representation.

This approach still provides a cognitive barrier, however. If you want to use a C++ framework directly, such as LLVM from Pragmatic Smalltalk or .NET, then you will need to write single-inheritance classes that encapsulate the multiple-inheritance classes that the library uses for most of its core types.

Another possible approach would be to avoid exposing any fields within the objects and just expose each C++ class as an interface. This would, however, make it impossible to inherit from the bridged classes without special compiler support to understand that some interfaces came along with implementation.

Although complex, this is a simpler system than interfacing between languages that differ on what method lookup means. For example, Java and Smalltalk have almost identical object and memory models, but Java ties the notion of method dispatch to the class hierarchy, whereas in Smalltalk two objects can be used interchangeably if they implement methods with the same names.

This is a problem encountered by RedLine Smalltalk,1 which compiles Smalltalk to run on JVM. Its mechanism for implementing Smalltalk method dispatch involves generating a Java interface for each method and then performing a cast of the receiver to the relevant interface type before dispatch. Sending messages to Java classes requires extra information, because existing Java classes don't implement this; thus, RedLine Smalltalk must fall back to using Java's Reflection APIs.

The method lookup for Smalltalk (and Objective-C) is more complex, because there are a number of second-chance dispatch mechanisms that are either missing or limited in other languages. When compiling Objective-C to JavaScript, rather than using the JavaScript method invocation, you must wrap each Objective-C message send in a small function that first checks if the method actually exists and, if it doesn't, calls some lookup code.

This is relatively simple in JavaScript because it handles variadic functions in a convenient way: if a function or method is called with more arguments than it expects, then it receives the remainder as an array that it can expect. Go does something similar. C-like languages just put them on the stack and expect the programmer to do the write with no error checking.

Memory Models

The obvious dichotomy in memory models is between automatic and manual deallocation. A slightly more important concern is the difference between deterministic and nondeterministic destruction.

…

The Path Forward

Many years ago the big interoperability question of the day was C and Pascal — two languages with an almost identical abstract machine model. The problem was that Pascal compilers pushed their parameters onto the stack left to right (because that required fewer temporaries), whereas C compilers pushed them right to left (to ensure that the first ones were at the top of the stack for variadic functions).

This interoperability problem was largely solved by the simple expedient of defining calling conventions as part of the platform ABI (application binary interface). No virtual machine or intermediate target was required, nor was any source-to-source translation. The equivalent of the virtual machine is defined by the ABI and the target machine's ISA.

Objective-C provides another useful case study. Methods in Objective-C use the C calling convention, with two hidden parameters (the object and the selector, which is an abstract form of the method name) passed first. All parts of the language that don't trivially map to the target ABI or ISA are factored out into library calls. A method invocation is implemented as a call to the objc_msgSend() function, which is implemented as a short assembly routine. All of the introspection works via the mechanism of calls to the runtime library.

We've used GNUstep's Objective-C runtime to implement front ends for dialects of Smalltalk and JavaScript in LanguageKit. This uses LLVM, but only because having a low-level intermediate representation permits optimizations to be reused between compilers: the interoperability happens in the native code. This runtime also supports the blocks ABI defined by Apple; therefore, closures can be passed between Smalltalk and C code.

Boehm GC (garbage collector) and Apple AutoZone both aimed to provide garbage collection in a library form, with different requirements. Can concurrent compacting collectors be exposed as libraries, with objects individually marked as nonmovable when they are passed out to low-level code? Is it possible to enforce mutability and concurrency guarantees in an ABI or library? These are open problems, and the availability of mature libraries for compiler design makes them interesting research questions.

Perhaps more interesting is the question of how many of these can be sunk down into the hardware. In CTSRD (Crash-worthy Trustworthy Systems R&D), a joint project between SRI International and the University of Cambridge Computer Laboratory, researchers have been experimenting with putting fine-grained memory protection into the hardware, which they hope will provide more efficient ways of expressing certain language memory models. This is a start, but there is a lot more potential for providing richer feature sets for high-level languages in silicon, something that was avoided in the 1980s because transistors were scarce and expensive resources. Now transistors are plentiful but power is scarce, so the tradeoffs in CPU design are very different.

The industry has spent the past 30 years building CPUs optimized for running languages such as C, because people who needed fast code used C (because people who designed processors optimized them for C, because...). Maybe the time has come to start exploring better built-in support for common operations in other languages. The RISC project was born from looking at the instructions that a primitive compiler generated from compiling C code. What would we end up with if we started by looking at what a native JavaScript or Haskell compiler would emit?

References

;

	Author	volume	Date Value	title	type	journal	titleUrl	doi	note	year
2013 TheChallengeofCrossLanguageInte	David Chisnall			The Challenge of Cross-language Interoperability				10.1145/2542661.2543971