Browse Prior Art Database

Hold-Offs On Loads With Uncorrected Data With a Load-Thru Cache

IP.com Disclosure Number: IPCOM000100712D
Original Publication Date: 1990-May-01
Included in the Prior Art Database: 2005-Mar-16
Document File: 2 page(s) / 65K

Publishing Venue

IBM

Related People

Kahle, JA: AUTHOR

Abstract

This article describes a way to increase system performance by eliminating one cycle from the reload time of the data cache. This is achieved by tracking where 'dirty' data is in the system.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 52% of the total text.

Hold-Offs On Loads With Uncorrected Data With a Load-Thru Cache

       This article describes a way to increase system
performance by eliminating one cycle from the reload time of the data
cache.  This is achieved by tracking where 'dirty' data is in the
system.

      The load-thru path of a data cache is activated when there is a
data cache miss on a load.  A data cache miss will cause a read of
main memory.  The data cache chip will read this data and Error
Correction Codes (ECCs).  It will take one cycle for the data cache
chip to determine if there is a data bit error and correct it.  In
most cases no correction of the data will be required.  The load-thru
path of the data cache will bypass this correction logic and feed
directly to the fixed point unit.  Thus, the fixed point unit can
receive uncorrected data when there is a load-thru. The data cache
chip will signal the ECC error to the fixed point unit, late in the
cycle, that the data came over.  The data cache might have to wait a
cycle to send the error signal over if the ECC logic is not fast.
When the data cache sends uncorrected data from the load-thru path,
it sends corrected data over to the fixed point unit the next cycle.

      The fixed point unit has a choice of which data it can use.
The uncorrected data is received by the fixed point chip one cycle
earlier than the corrected data.  In the majority of the cases when
system performance should be high, no correction is needed on the
data.

      To achieve high performance, the fixed point unit uses the
uncorrected data and starts the execution of the instruction using
the uncorr...