Check out the new USENIX Web site.

Home About USENIX Events Membership Publications Students
JVM 2001 Abstract

Mostly Accurate Stack Scanning

Katherine Barabash, Niv Buchbinder, Tamar Domani, Elliot K. Kolodner, Yoav Ossia, and Shlomit S. Pinter, IBM Haifa Research Laboratory; Janice Shepherd, IBM T.J. Watson Research Laboratory; and Ron Sivan and Victor Umansky, IBM Haifa Research Laboratory


The Java Virtual Machine (Jvm) needs, for the purpose of garbage collection (GC), to determine the data type stored in every memory location. Jvms that can do this reliably are said to be type-accurate (TA). Full type accuracy usually exacts a price in performance due to the need to scan stacks and registers accurately. The mostly accurate approach presented here can reduce the TA overhead significantly by sacrificing accuracy for the small minority of memory locations that add the most to the cost. Performance results show that mostly accurate stack scanning performs as well as conservative stack scanning and that relatively few objects are identified conservatively. In addition our implementation is designed to support and generate type maps for any verifiable bytecode stream (including combinations that are unlikely to be produced by a compiler) without requiring rewriting of the bytecode. We introduce a new compression technique for type maps that uses a program-friendly format for the maps; yet, achieves good compression and provides fast opening of compressed maps. We show how to apply systematic testing techniques and test coverage tools to an accurate stack scanner.
  • View the full text of this paper in PDF form.

  • If you need the latest Adobe Acrobat Reader, you can download it from Adobe's site.

  • To become a USENIX Member, please see our Membership Information.
?Need help? Use our Contacts page.

Last changed: 3 Jan. 2002 ml
Technical Program
JVM 2001 Home