Review Board 2.0.15


alpha: Remove ALPHA support and supplemental files

Review Request #3684 - Created Oct. 24, 2016 and updated

Information
Andreas Hansson
gem5
default
Reviewers
Default
Changeset 11689:e1835e5846b9
---------------------------
alpha: Remove ALPHA support and supplemental files

This patch removes support for ALPHA, and correspondingly also removes
any devices, tests, supporting functions, and configurations unique to
ALPHA.

R.I.P.

A few things that are worth re-adding using something other than
ALPHA:

1) A two-system regresssion using Ethernet

2) Ruby tester regressions using NULL for the variuos Ruby protocols

   

Issue Summary

3 1 0 2
Posted (Oct. 24, 2016, 1:51 p.m.)



  
src/cpu/o3/rename_map.cc (Diff revision 1)
 
 

It would be great if someone familiar with the o3 model could double check this modification.

Posted (Oct. 25, 2016, 3:13 p.m.)

I'm a little surprised at how much coupling there is between the CPU models and the ISAs. I didn't expect this to require so many modifications.

Thanks for putting this together, Andreas, other than a couple of small things below it looks good. I'm fully on board with dropping Alpha support.

SConstruct (Diff revision 1)
 
 

Clearly X86 should be the default, not ARM! ;)

Joking aside, maybe the default should be the same as the host system? Or we could remove the default entirely. I didn't even know a default existed.

  1. It was based on ASCII distance (and the fact that I think it's fair to say that ARM is the best supported ISA).

util/regress (Diff revision 1)
 
 

I think this has been pointed out by Brad in the past, but we need to keep these tests for other Ruby protocols around. I think it's fine to just switch them to using x86 instead of Alpha. As far as I know, most Ruby users are using x86 anyway.

  1. Why X86 and not NULL? I would suggest to let all the ruby testers simply be part of NULL (as suggested in the patch summary). It would be even nicer if we didn't have to compile gem5 for every protocol, but that may be asking too much.

  2. There's two components to testing the protocols. First, do they compile, which the NULL ISA would test perfectly well. Second, do they work, which we need to run some applications on the protocol (Linux boot is actually a pretty good application for this).

    I agree that it would be good if Ruby was a library, but I do think it's too much to ask :).

  3. Feel free to add Linux boot to Ruby+X86. Today all we do is "hello" and rubytest. For the latter we don't need an ISA, NULL will do just fine (we use NULL for all the memtest and traffic-gen tests of the classic memory system). Running "hello" seems fairly redundant to me. I would think adding some more synthetic tests should be just as good as "hello" if not better. No?

  4. Yeah, you're probably right. You've convinced me it's fine to just use the NULL ISA for most of the protocols. We can revisit adding other tests later.