Files in this item
|(no description provided)|
|Title:||Automatic array privatization and demand-driven symbolic analysis|
|Doctoral Committee Chair(s):||Padua, David A.|
|Department / Program:||Computer Science|
|Degree Granting Institution:||University of Illinois at Urbana-Champaign|
|Abstract:||Memory-related anti- and output dependences are false dependences because they do not represent the flow of data but rather only the collisions caused by memory location reuse. Privatization is a technique to eliminate cross-iteration anti-dependences. A variable is privatized if each processor allocates a local copy of the variable such that each processor accesses only its local copy.
To determine whether an array is privatizable in a loop, we analyze the array sections that are defined and used in each iteration. A use section of an array is dominated by a definition section if the definition section covers the use section and if the definition site dominates the use site in the program control flow graph. To enhance the effectiveness of array section analysis, we develop several new techniques that can analyze conditionally defined (used) variables and symbolic expressions in array sections.
Gated Single Assignment (GSA) form of program is used for conditional data flow analysis and demand-driven symbolic analysis. We developed an almost-linear time algorithm for building and placing gating functions to convert program into GSA form.
We introduce a technique that exploits the single assignment property of GSA to perform demand-driven symbolic analysis using the use-def chains embedded in the variable names. The symbolic analysis technique uses the gating functions of GSA to represent the possible values of symbolic variable and the predicates guarding the choice of the possible values. We present a technique that can obtain more precise results from the predicated possible values using path conditions.
We also present a technique that uses the gating functions as join operators at the join nodes of control flow graphs to perform conditional data flow analysis.
We evaluate the effectiveness of these techniques using programs from the Perfect Benchmarks and NCSA users. After array privatization, our compiler is able to parallelize many outer loops. The resulting programs show speedups around 5 on an 8 processor SGI Challenge system.
|Rights Information:||Copyright 1995 Tu, Peng|
|Date Available in IDEALS:||2011-05-07|
|Identifier in Online Catalog:||AAI9624520|