64 bit cleanup completion?

It's been a while since the 64 bit cleanup work was started and  
almost as long since anything much was done with it. There are a  
number of plugins that need some work to get rid of things like  
cCoerce:to: in favour of oopForPointer: etc.

There are places with #flag: Dan lying around and Dan has expressed a  
lack of time to do anything with them but a willingness to sit with  
someone and a beer to explain what he was thinking of. I'm not going  
to be heading down to silicon valley anytime soon so I certainly  
can't do that but anyone that can take notes could do it and report  
back. Volunteers?

tim
--
tim Rowledge; [hidden email]; http://www.rowledge.org/tim
"Bother" said Pooh, as he flunked the the sobriety test.

On Fri, Apr 21, 2006 at 04:49:01PM -0700, tim Rowledge wrote:
All we need here is a bit of coordination. Maybe just a swiki page.
Way back when, Ian posted a list of "who seems to own what plugin"
and status with respect to 64 bit cleanliness. This is a perfectly
good starting point if we can just put it on a swiki page and update
it from there.

Also needed is a "64 bit Squeak" category for bug tracking on
Mantis. There are image *and* VM issues that will need to be
resolved, but these are of no interest to most folks with 32 bit
Squeak image/VM, and should be separately tracked if anything
is going to be done about them.

The original 64-bit motivators (Dan and Ian) did the hard bit. Now
all we need is to put a little bit of structure around the follow
up tasks. Yes, I will volunteer to pick up one or two tasks, and
I'm sure that other folks will do so also.

Dave

p.s. I think this power of two business is a bit over-hyped. The
correct machine word size is and always has been 24 bits.

David,

> p.s. I think this power of two business is a bit over-hyped. The
> correct machine word size is and always has been 24 bits.

24?  24???!!!  What kind of number is that?!?  You can't even fit an  
entire file name into it (sixbit uppercase ASCII, naturally)!

Noooo....  You need 36 bits, my man.  Split down the middle for those  
ordinary everyday tasks where 18 bit halfwords will do just fine, and  
in only half the resources too.

A PDP-10 with a KL10 core driving 40 Newbury terminals.  That's what  
Squeak really needs!

(Seriously: AFAIK DEC never finished the design and implementation of  
the PDP-2 which would have added a 24-bit machine to the most elegant  
range of CPUs ever designed, IMO.  That left either a few 24-bit  
offerings from Honeywell and the like, or move up to glorious 36-bit  
PDP-/DECsystem-10s.)

>> There are places with #flag: Dan lying around and Dan has expressed a
>> lack of time to do anything with them but a willingness to sit with
>> someone and a beer to explain what he was thinking of. I'm not going
>> to be heading down to silicon valley anytime soon so I certainly
>> can't do that but anyone that can take notes could do it and report
>> back. Volunteers?

I could maybe do that next week.  But here's what I remember as the  
last thoughts on the matter:

Plugins:  This is mostly going to be tidying up declarations making  
sure int means machine width, foo * means machine pointer, sqOop  
means Squeak width pointer and sqInt means integer with same width as  
sqOop.  Gratuitous casts between pointer and int all have to be  
replaced with appropriate macros/inline functions.  (In fact I think  
all declarations should honestly and truthfully declare precisely the  
type of their value, including function pointers, with _no_  
exceptions.  While we're at it, functions returning nothing should be  
'void', not 'int'.)  Beyond that is hunting down and repairing 32-bit  
assumptions: divide/multiply by 4, shift by 2, mask by [~]3, etc.  
LargeIntegerPlugin might have some unique surprises waiting there.  
The parts of BitBlt and B3D that were never tested (NOTE: these  
plugins are NOT entirely converted yet) should be tested and repaired  
where needed.  In some cases it's necessary to truncate to 32 bits to  
avoid unwanted carry out of 32 bits or to force sign extension out of  
bit 31.  IIRC, graphics code had a few very hard to find issues along  
those lines.  This is mostly mindless tedium with the occasional need  
to think seriously about the right type to use, made even more  
tedious by having to regenerate the plugin from the image ever few  
minutes having correlated a C line number with a statement in a  
Squeak method somewhere.

Taking advantage of 64-bit 'long's:  We also discussed the two other  
obvious changes: SmallIntegers being represented as 63-bits + tag in  
64-bit images, and 64-bit Bitmaps storing two 32-bit pixels in each  
word -- with a couple of additional primitives for 64-bit accesses.  
It could be the case that many of the 'flag:'s in the image are  
related to those two issues.  Both of these go deep, and profoundly  
affect things both in the Interpreter and in various plugins.  Not to  
be undertaken lightly.

I know Dan also wanted to clean up the base and method headers, in  
particular the split fields, in an incompatible image format change.  
Some 'flag:'s might be attached to those too.

If there was anything else then I'm afraid I can't remember it  
offhand now.

I think I remember Andreas working on the win32 support code to make  
it 64-bit clean(er), but I'm not certain.

Frankly I think we'd be making huge progress if we can just get the  
rest of the plugins to compile and run in both 64- and 32-bits, with  
sufficient confidence in their correctness to fold them back into the  
trunk.

I'd be more than happy to set up a repository for 64-bit support code  
while it's unstable.  I could also, in case of dire need, offer  
accounts on 64-bit hardware for anyone who is eager to work on a  
particular plugin but has no 64-bit hardware to test their changes.

FWIW: one carrot to dangle in front of people as a motivator is that  
this work is only half concerned with running 64-bit images.  The  
other half is about running 32-bit images on 64-bit hardware, which  
more and more people are finding to be a frustrating limitation.

Cheers,
Ian

On 21-Apr-06, at 8:39 PM, Ian Piumarta wrote:

Fie on that nonsense. What y'need is 33bits. 32 for a decent size  
integer (who*needs* 64 bit values, anyway) and one bit for a tag.  
Then you  need an asynchronous ARM-like architecture with the TLC  
(sorta WCS) and floating pint hardware connected to fullspeed mram so  
that there is no need for data cache. Then you have a few thousand of  
those in a single machine. Easy. Now gimme One Billion Euros to  
implement it.

tim
--
tim Rowledge; [hidden email]; http://www.rowledge.org/tim
State-of-the-art: What we could do with enough money.

On Apr 21, 2006, at 8:39 PM, Ian Piumarta wrote:

The only thing Dan wanted to add immediately (without going back and  
looking at his image) to my recollections was that storing 64-bit  
words in Bitmap-like objects also required a new format ("indexable  
64-bit words only") and class (variableLongWordSubclass:... or  
something akin).

He also reminded me that I'd been wanting to generalise the code  
generation such that the choice of image word size could be made at C  
compile time, rather than when translating the sources.  (The various  
constants appearing by value and that are currently set depending on  
the word size would instead appear by name in the code, and the build  
environment would be responsible for switching between 32- or 64-bit  
definitions for them.)  This would allow identical sources to  
generate all 4 permutations of image width and hardware pointer width.

(The beer in fact turned out to be fish and wine at Joanie's. :)

Cheers,
Ian

On Thu, Apr 27, 2006 at 01:03:58AM -0700, Ian Piumarta wrote:
>
> He also reminded me that I'd been wanting to generalise the code  
> generation such that the choice of image word size could be made at C  
> compile time, rather than when translating the sources.

That would be a nice improvement. It would eliminate the need for
separate ./src32 and ./src64 directories, and would allow VMMaker to
go back to generating the sources to ./src in either case.

Dave

On 27-Apr-06, at 1:03 AM, Ian Piumarta wrote:

This would be very helpful I suspect. Do you
a) have time to make those changes yourself, or
b) have particular ideas on what to do about them if bar-a.


tim
--
tim Rowledge; [hidden email]; http://www.rowledge.org/tim
There's a guy works down the fish shop swears he's elvish...

Any reply?

Begin forwarded message:


tim
--
tim Rowledge; [hidden email]; http://www.rowledge.org/tim
"Bollocks," said Pooh being more forthright than usual