[Search for users] [Overall Top Noters] [List of all Conferences] [Download this site]

Conference ricks::dechips

Title:Hudson VLSI
Notice:For Digital Chip Data - CHIPBZ::PRODUCTION$:[DS_INFO...]
Moderator:RICKS::PHIPPS
Created:Wed Feb 12 1986
Last Modified:Fri Jun 06 1997
Last Successful Update:Fri Jun 06 1997
Number of topics:701
Total number of notes:4658

690.0. "Another Math bug." by RDGENG::ATTWOOL () Tue May 06 1997 13:31

    
    Do Alpha's suffer from Math bugs ??
    
    See http://www.x86.org/secrets/Dan0411.html for infomation on the
    new Pentium II / Pro math bug.
    
    
T.RTitleUserPersonal
Name
DateLines
690.1Anyone else seen this one ?BBPBV1::WALLACEjohn wallace @ bbp. +44 860 675093Tue May 06 1997 14:228
    That URL may need a - in it: try dan-0411.html
    
    And while we're on the subject, anyone seen anything definitive on the
    alleged MMX "feature" which screws up systems where MMX apps and FPU
    apps timeslicing allegedly overwrite each other's MMX/FPU context ?
    
    regards
    john
690.2DIGITAL hardware has had a few bugs, tooXDELTA::HOFFMANSteve, OpenVMS EngineeringTue May 06 1997 14:4318
:    Do Alpha's suffer from Math bugs ??

    Please see HUMANE::DIGITAL note 5276.* for a discussion of this.

    I would tend to avoid making any public comment about any of Intel's
    microprocessor problems, as these public comments could easily come
    back to haunt DIGITAL later.

    DIGITAL has had cases of problems in a processor or microprocessor
    design -- I know of various examples over the years, which I will
    refrain from mentioning here for obvious reasons.  (These problems
    typically have hardware or software fixes or workarounds when the
    problem is identified, and are thus not visible to users.)
    
    The current Intel Pentium-II bug appears fairly obscure when compared
    with the original Pentium FP bug.

690.3it's all 1's and 0's to me.RDGENG::ATTWOOLTue May 06 1997 15:113
    
    ..I guess they do. Now what was that date ? ;^(
    
690.4BIGUN::nessus.cao.dec.com::MayneA wretched hive of scum and villainyTue May 06 1997 22:1816
Re .1:

MMX uses the same registers as the FPU. The reasoning is that if you put more 
registers in, you'd lose backwards compatibility in a big way, like for instance 
a context swap not saving the MMX registers. If the same registers are used, 
then a context swap which currently saves the FPU registers will also 
conveniently save the MMX registers with no extra work needed, and the OS 
doesn't know or care whether you were doing FPU or MMX stuff at the time. This 
specifically avoids the problem you stated.

The obvious downside is that your code can't do MMX and FPU work at the same 
time. This may or may not be a problem.

This is what I've figured out from reading popular magazines.

PJDM