[Vm-dev] Re: [Pharo-project] Integrating Changes in 1.4 that
require a new VM
Eliot Miranda
eliot.miranda at gmail.com
Fri Sep 23 17:12:29 UTC 2011
On Fri, Sep 23, 2011 at 5:01 AM, Henrik Sperre Johansen <
henrik.s.johansen at veloxit.no> wrote:
>
> On 22.09.2011 21:57, Eliot Miranda wrote:
>
>
>
>
>
> On Thu, Sep 22, 2011 at 12:29 PM, Henrik Sperre Johansen <
> henrik.s.johansen at veloxit.no> wrote:
>
>
> On 22.09.2011 20:20, Eliot Miranda wrote:
>
>
>
>
>
> On Thu, Sep 22, 2011 at 11:06 AM, Eliot Miranda <eliot.miranda at gmail.com>wrote:
>
> Hi Igor,
>
> On Thu, Sep 22, 2011 at 10:53 AM, Igor Stasenko <siguctua at gmail.com>wrote:
>
> On 22 September 2011 19:16, Eliot Miranda <eliot.miranda at gmail.com>
> wrote:
> > (apologies for the duplicate reply; someone needs to sort out their
> > threading for the benefit of the community ;) )
> >
> > On Thu, Sep 22, 2011 at 2:36 AM, Marcus Denker <marcus.denker at inria.fr>
> > wrote:
> >>
> >> Hi,
> >>
> >> There are two changesets waiting for integrating in 1.4 that have
> serious
> >> consequences:
> >>
> >> - Ephemerons. The VM level changes are in the Cog VMs build on Jenkins,
> >> but have not
> >> been integrated in the VMMaker codebase.
> >>
> >> http://code.google.com/p/pharo/issues/detail?id=4265
> >
> > I would *really* like to back out these changes. The Ephemeron
> > implementation is very much a prototype, requiring a hack to determine
> > whether an object is an ephemeron (the presence of a marker class in the
> > first inst var) that I'm not at all happy with. There is a neater
> > implementation available via using an unused instSpec which IMO has
> > significant advantages (much simpler & faster, instSpec is valid at all
> > times, including during compaction, less overhead, doesn't require a
> marker
> > class), and is the route I'm taking with the new GC/object-representation
> > I'm working on now. Note that other than determining whether an object
> is
> > an ephemeron (instSpec/format vs inst var test) the rest of Igor's code
> > remains the same. I'd like to avoid too much VM forking. Would you all
> > consider putting these changes on hold for now?
> > If so, I'll make the effort to produce prototype changes (in the area of
> > ClassBuilder and class definition; no VM code necessary as yet) to allow
> > defining Ephemerons via the int spec route by next week at the latest.
> >
>
> i agree that in my implementation this is a weak point. But its hard
> to do anything without
> making changes to object format to identify these special objects.
>
> The main story behind this is can we afford to change the internals of
> VM without being beaten hard
> by "backwards compatibility" party? :)
>
>
> I don't think we get stuck in this at all. The instSpec/format field
> has an unused value (5 i believe) and this can easily be used for
> Ephemerons. All that is needed is a little image work on these methods:
>
> Behavior>>typeOfClass
> needs to answer e.g. #ephemeron for ephemeron classes
>
> ClassBuilder>>computeFormat:instSize:forSuper:ccIndex:
> needs to accept e.g. #ephemeron for type and pass variable: false
> and weak: true for ephemerons to format:variable:words:pointers:weak:.
>
> ClassBuilder>>format:variable:words:pointers:weak:
> needs to respond to variable: false and weak: true by computing the
> ephemeron instSpec.
>
>
> Class>>weakSubclass:instanceVariableNames:classVariableNames:poolDictionaries:category:
>
> ClassBuilder>>superclass:weakSubclass:instanceVariableNames:classVariableNames:poolDictionaries:category:
> need siblings, e.g.
>
> ephemeronSubclass:instanceVariableNames:classVariableNames:poolDictionaries:category
>
> superclass:ephemeronSubclass:instanceVariableNames:classVariableNames:poolDictionaries:category:
>
> Right? This is easy. Then in the VM there are a few places where
> pointer indexability (formats 3 and 4) need to be firmed up to exclude 5,
> but nothing difficult. We talked about this in email last week.
>
>
> Here's the format field (Behavior>instSpec at the image level) as
> currently populated:
>
> 0 = 0 sized objects (UndefinedObject True False et al)
> 1 = non-indexable objects with inst vars (Point et al)
> 2 = indexable objects with no inst vars (Array et al)
> 3 = indexable objects with inst vars (MethodContext AdditionalMethodState
> et al)
> 4 = weak indexable objects with inst vars (WeakArray et al)
> 6 = 32-bit indexable objects (Float, Bitmap ert al)
> 8 = 8-bit indexable objects (ByteString, ByteArray et al)
> 12 = CompiledMethod
>
> N.B. in the VM the least two bits of the format/instSpec for byte objects
> (formats 8 and 12) is used to encode the number of odd bytes in the object,
> so that a 1 character ByteString has a format of 11, = 8 + 3, size = 1 word
> - 3 bytes.
>
>
> For the future (i.e. the new GC/object representation, /not/ for the
> first implementation of ephemerons which we can do now, for Pharo 1.4 or
> 1.5) we need to extend format/instSpec to support 64 bits. I think format
> needs to be made a 5 bit field with room for 4 bits of odd bytes for 64-bit
> images. [For VMers, the Size4Bit is a horrible hack). So then
>
> 0 = 0 sized objects (UndefinedObject True False et al)
> 1 = non-indexable objects with inst vars (Point et al)
> 2 = indexable objects with no inst vars (Array et al)
> 3 = indexable objects with inst vars (MethodContext AdditionalMethodState
> et al)
> 4 = weak indexable objects with inst vars (WeakArray et al)
> 5 = weak non-indexable objects with inst vars (ephemerons) (Ephemeron)
>
> and we need 8 CompiledMethod values, 8 byte values, 4 16-bit values, 2
> 32-bit values and a 64-bit value, = 23 values, 23 + 5 = 30, so there is
> room, e.g.
>
> 9 (?) 64-bit indexable
> 10 - 11 32-bit indexable
> 12 - 15 16-bit indexable
> 16 - 23 byte indexable
> 24 - 31 compiled method
>
> In 32-bit images only the least significant 2 bits would be used for
> formats 16 & 24, and the least significant bit for format 12.
>
> If we are changing the format for 64bit images anyways, why not simplify
> it/ be more consistent by spending a full byte?
>
> Bit: 8 7 6 5 4 3
> 2 1
> | 64bit | 32bit |16bit | 8bit |compiled | weak | indexable | instVars |
>
> (Odd number encoded in remaining indexable bit fields)
>
>
> I used to prefer this approach but I've realised that the format/instSpec
> approach (I think Dan came up with) makes better use of bits because so many
> of the bit combinations are mutually exclusive. For example, pointers
> excludes all the byte/short/32-bit/64-bit indexability combinations. Also,
> see below...
>
>
>
> Could get away with 7 if you put f.ex. the unused indexable weak
> combination (6) as compiled method/8bit
>
> Or is the header space in your new 64bit format already quite filled, so
> this is a bad idea?
>
>
> Yes, ish. But they're scarce, and very useful for experiments etc.
> Right now I have
>
> typedef struct {
> unsigned short classIndex;
> unsigned unused0 : 6;
> unsigned isPinned : 1;
> unsigned isImmutable : 1;
> unsigned format : 5; /* on a byte boundary */
> unsigned isMarked : 1;
> unsigned isGrey : 1;
> unsigned isRemembered : 1;
> unsigned objHash : 24; /* on a 32-bit word boundary */
> unsigned char slotSize; /* on a byte boundary */
> } CogObjectHeader;
>
> Where classIndex is 16-bits simply for efficiency and will grow to 20 or
> 22 bits as needed. So one could steal one or two bits from unused0 and two
> bits from objHash, and give these to format, but it would be a waste.
> Better keep these back for other uses.
>
> Also, can I ask the assembled company exactly how many bits you'd spend
> on the objHash (identityHash)? Think forward to 64-bits. Is 24 bits about
> all we can afford or still too generous? Anybody have any data to
> contribute?
>
> This is probably a stupid question, but where is the variable size in words
> stored?
>
So the size of an object in slots (32-bit or 64-bit) is the byte slotSize
field, allowing an object to have 254 slots before it is too big for this
representation. If an object has more than 254 slots its slotSize has the
value 255 and an additional 64-bit word (64-bit to guarantee 64-bit
alignment of all objects) preceeding contains the actual slot size (even if
only 32-bits are used). This means that there is a maximum of 2/255 words
of overhead for this extra word, or < 1%.
So to compute the size of an indexable object one determines the slot size
and subtracts the number of named inst vars. For format 4 there are no
named inst vars, so the computation is something like
| ss |
numIndexableFields := (ss := obj slotSize > 254 ifTrue: [obj
overflowSlotSize] ifFalse: [ss]) - (obj format = 3 ifTrue: [self
numIndexableFields: obj] ifFalse: [0])
In 64-bit VisualWorks I took a different approach, using an 8 bit field to
hold either the number of named inst vars or the number of odd bytes. But
this meant an 8 bit slot size, an 8 bit nun named inst vars/odd bytes and 8
header flags, leaving only 20 buts each for the class index and obj hash
fields. With Squeak this isn't compelling because the collection hierarchy
has been reworked to use Arrays instead of the older flat format (e.g. where
OrderedCollection was indexable and has firstIndex and lastIndex).
In a quadword preceding the header like it is in 32bit format?
>
Yes.
> The reason I'm asking is that to me, the main application of identityHash
> is for HashedCollection's, and the max size of those thus impact what a
> reasonable answer would be...
>
and what kinds of sizes have you seen?
>
> Cheers,
> Henry
>
>
--
best,
Eliot
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.squeakfoundation.org/pipermail/vm-dev/attachments/20110923/d225be80/attachment-0001.htm
More information about the Vm-dev
mailing list