Communications - August 2010

Furthermore, because safe languages have a fully defined semantics, unlike languages like C, with one semantics if a program obeys the language rules and no guarantees if they don’t, program-analysis tools are not put in the untenable position of assuming a buggy program plays strictly according to the language definition.

Safe languages are far more popular since the introduction of Java but are generally considered inappropriate for systems code, which is usually written in a low-level, glorified assembly language like C or its more sophisticated cousin C++. The common belief is that safe languages are inefficient, due in part to the size and complexity of their runtime systems and reliance on garbage collection.

Singularity’s Bartok compiler provides language safety without the typical performance penalty by compiling C#’s Microsoft Intermediate Language representation to native (x86, x64, or ARM) code at installation time rather than at runtime. Bartok also links compiled code to a small runtime consisting of only a class library and a garbage collector, not a large runtime environment like the Common Language Runtime (the virtual machine component of Microsoft .NET) and the Java Virtual Machine. The Shared Source Common Language Infrastructure runtime and class library are more than five times larger than the Bartok runtime ( 64 thousand lines of code, or KLOC, vs. 350KLOC), roughly the same as the C runtime in the latest version of Windows (72KLOC). Moreover, Bartok is a highly optimizing compiler that generates high-quality code and reduces memory use through extensive tree shaking to discard unneeded class variables and method definitions.

Table 1 emphasizes this point by
outlining the memory footprint for a
small program written in C, C++, and
C# running on several different oper-
ating systems. The program outputs
“Hello World” using the standard I/O
libraries and APIs for each system—
printf for C and C++ and Console.
WriteLine for C#. The C# code on
Singularity is smaller than for all but
one other system—the statistically
linked code on Free BSD—in some
cases half to one-third the size of C++
code. Table 2 outlines the reduction
in memory footprint Bartok achieves
for a variety of programs. Much of the
code and data “shaken” out of these
programs comes from the unused por-
tions of general-purpose libraries.