... AIX\texttrademark.¹

There was also a partial port available for Linux/PowerPC, provided by collaborators from the University of Massachusetts.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... registers.²

These registers are treated as dedicated, scratch, volatile (caller-save), or non-volatile (callee-save). Parameters are passed in volatile registers. Intermediate results can be accumulated in volatile or scratch registers. Non-volatile registers retain their value across method calls.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... problems³

PowerPC is ``big endian'' -- the high-order byte of a word is at the lowest address within it -- while IA32 is ``little endian''.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... Java.⁴

In addition, the Jikes RVM source tree contains several tools used in the build process (5,600 lines) and the jdp debugger (33,300 lines).

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... text.⁵

To keep down instances of this rather ugly feature, we try to limit it usage to three cases: to define a control, to reference a class that would not otherwise be loaded, or to define or reference a field that would not otherwise be needed. In the 162,000 lines of platform independent Java code, there are 38 preprocessor blocks impacting 479 lines of source code that are related to the choice between AIX/PowerPC or Linux/IA32.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... backend,⁶

The original PowerPC compilers each had their own assemblers.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... instruction.⁷

On the PowerPC architecture, the baseline compiler expends unnecessary extra work pushing the unused middle parameter onto the stack. However, the optimizing compiler identifies the computation as dead and eliminates it. Thus, the optimizing compiler produces code for the prepare and attempt primitives that is as efficient as architecture-specific primitives.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... problem.⁸

It remains in the 2.4 Linux/PowerPC release.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... performance⁹

The Jikes RVM FastAdaptiveSemispace images, used for Figures 2 and 3, gives the best overall performance due to feedback-directed optimization and delayed compilation effects [4]. However, to reduce the impact of timer-driven non-deterministic actions, the remainder of this paper reports results using a (non adaptive) FastSemispace image which compiles each method the first time it is invoked at optimization level O2. In all cases, two virtual processors and a 400 MB heap were used. All runs are on a 4-way IBM Netfinity with 700MHz Pentium III processors and 3 GB of memory. The Linux installation is a customized RedHat version with a 2.4.12 kernel and GNU Libc version 2.2.4; the libc and kernel versions support the version of LinuxThreads that uses the GS segment register for thread-local storage enabling it to support Jikes RVM's user-level multithreading mechanism.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... benchmarks¹⁰

These benchmarks were developed by the Standard Performance Evaluation Corporation [6]. The performance numbers reported in this paper are the best run of 5 on each individual SPECjvm98 benchmark. These runs do not conform to the official SPEC run rules, so our results do not directly or indirectly represent a SPECjvm98 metric, and are not comparable with a SPECjvm98 metric.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... performance.¹¹

During the six month period shown in Figure 2 several new platform independent optimizations were added to the optimizing compiler and existing optimizations were enhanced. Although these contributed to the performance improvements shown in the graph, most of the gain was caused by IA32 specific improvements (during the same period AIX/PowerPC performance only improved by about 10%).

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... trees¹²

A tree is ready if all of the trees on which it is dependent have already been emitted.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... ways.¹³

We have not performed an apples-to-apples comparison of our algorithm compared to Traub et al.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.