Evaluating high-level language pointer-expressions can require a
significant portion of the execution time and memory bandwidth in
programs, especially for pointers with multiple memory accesses and
address computations. Often such pointers repeatedly follow the same
evaluation paths and return the same values. We present a
micro-architecture technique that dynamically eliminates the execution
of pointers based on results obtained from previous evaluations,
without using prediction. We experiment with a benchmark set that
includes continuous-speech recognition, MPEG-1 decoding, and SPECint95
ijpeg, li and go programs. Simulations of Digital Alpha 21164
microprocessor show that the eliminated pointers account for up to
11.3% of execution time, and 26% of data memory bandwidth of an entire
program. Hence, the proposed microarchitecture provides substantial
performance gains by significantly reducing the number of executed
pointers.


<p><hr>

The author's web site:
<a href="mailto:nitzan@andrew.cmu.edu">nitzan@andrew.cmu.edu</a>