I’m Done Simulating: Now what? verification coverage analysis and correctness checking of the DEC chip 21164 Alpha microprocessor

Digital's Alpha-based DECchip 21 164 processor was verified extensively prior to fabrication of silicon. This simulation-based verification effort used implementation-directed, pseudorandom excrchsers which were supplemented with implementation-specific, hand-generated tests. Special emphasis was placed on the tasks of checking for correct operation and functional wverage analysis. Coverage analysis shows where testing is incomplete, under the assumption that untested logic often contains bugs. Correctness checkers are various mechanisms (both during and after simulation) that monitor a test to detennine if it was successful. This paper details the coverage analysis and correctness checking techniques that were used. We show how our methodology and its implementation was successful, and we discuss the reasons why this methodology allowed several minor bugs to escape detection until the f i s t prototype systems were available. These bugs were corrected before any chips were shipped to customers. OVERVIEW The DECchip 21164 CPU chip is a quad-issue, super-scalar implementation of the Alpha architecture which required a rigorlms verification effort to ensure that there were no logical bugs. World-class performance dictated the use of many advanced micro-architectural features, such as a virtual instniction cache with seven-bit Address Space Numbers, a dual-read-ported data cache, out-of-order instruction completion, on-chip three-way set-associative write-back second-level cache, module-level cache control, branch prediction, demand-paged mem'ory management unit, write buffer unit, miss-address file unit, and a complicated Bus Interface Unit with support for various CPU-system clock ratios, system configurations, and module-level cache parameters. [ 11 Increasingly, functional verification efforts are relying on pseudorandom test generation to improve the quality of functional coverage. These techniques have been in use at Digil.al for more than seven years and are also used elsewhere in the industry and in academia.[2-51 However, the heavy use of pseudorandom testing increases the need for new ways o