next up previous
Next: The Bytecode IR Up: Code Analysis and Optimization Previous: Code Analysis and Optimization


The Quad IR

Figure 2: Overview of the operator hierarchy. Visitors visit in most-specific to least-specific order.
\includegraphics[width=6in]{operators}

The major intermediate representation in Joeq is the Quad format. The Quad format is a set of instructions, called Quads, that are organized into a control flow graph. Each Quad consists of an operator and up to four operands. For efficiency, Quads are implemented as a final class; polymorphism is implemented in the different operators and operands. However, we retain most of the benefit of polymorphism on quads by utilizing strict type checking on the operator type. Operators are implemented using the singleton pattern, so there is only one instance of each operator, which is shared across multiple Quads.

The control flow graph in the Quad format does not explicitly represent control flow edges due to exceptions. Instead, edges due to exceptional control flow are factored so that there is only one ``exception'' edge for each basic block[10]. This exception edge points to the exception handlers that can catch exceptions thrown in the block. This means that, due to exceptions, control flow can potentially exit from the middle of a basic block. However, this greatly reduces the number of basic blocks and control flow graph edges and thereby makes compilation more efficient.

Operators in the Quad format fall into three classes: high, middle, and low. High-level operators correspond to complicated, often language-dependent operations, such as array accesses, method invocations, or dynamic type checks. These operations operate solely on symbolic references, and are therefore independent of the virtual machine and object memory layout. Middle-level operators correspond to more basic operations, such as loads and stores to calculated addresses. These operations are dependent on the memory layout, but are still independent of CPU type. Finally, low-level operators correspond to machine code instructions. At this low level, pseudo-registers are replaced by physical registers, and code can be generated.

As compilation proceeds, translation passes on the Quad format replace higher-level operations with equivalent sequences of lower-level operations. However, the basic data structures stay the same, so one can uniformly execute compiler passes on the IR regardless of its level. This maximizes code reuse in the compiler.

The different types of operators in the Quad format form a hierarchy, as seen in Figure 2. Analyses can use instanceof tests or the visitor pattern[6,15] to select the Quads that match certain characteristics.

The Quad IR supports the inclusion of optional supplementary information such as profile data and mappings back to line numbers in the source code. These are implemented as extensions to the IR, and are not required for correct operation. Extensions implement a standard interface that specifies how to update the extra data when performing various IR transformations. This allows the compiler to automatically and implicitly update data even across IR transformations.


next up previous
Next: The Bytecode IR Up: Code Analysis and Optimization Previous: Code Analysis and Optimization
John Whaley 2003-03-15