[Vm-dev] goto instruction with Cog VM

Eliot Miranda eliot.miranda at gmail.com
Mon Nov 3 19:34:48 UTC 2014

Hi Ralph,

On Sun, Nov 2, 2014 at 9:48 PM, Ralph Boland <rpboland at gmail.com> wrote:

> >>
> >> I am working on a parser generator tool (a replacement for SmaCC) and
> >> one of the things I a m interested in is the direct translation of
> >> language specifications into the virtual machine code (SmaCC and my
> >> current version of it use Squeak as the target language).
> >>
> > First, a different approach than compiling to Smalltalk is to compile to
> a
> > parse tree.  We do this in the pseudo JavaScript compiler we've
> implemented
> > at Cadence, translating an AST into a Squeak compiler parse tree for code
> > generation.  Targeting a parse tree gives you much more freedom; you can
> > express things that aren't expressible in Smalltalk.  And if you target
> > bytecodes you can do even more.
> I never considered using a parse tree as the target.  An interesting idea
> which in many instances may be the best approach.  But for my regular
> expression
> example I would still want to generate byte codes.  In any case I wouldn't
> want to
> restrict users of my parser generator tool to any one of the three options
> (Smalltalk code,
> parse tree, byte code).  It is my responsibility to make all three options
> as easy and
> efficient as reasonably possible for users of the parser generator tool.
> Haven't put
> much thought into this yet though.   So far Smalltalk (Squeak) is the only
> option.
> >> One of the problems I have is that, for some languages, the natural
> >> translation
> >> into VM code uses computed gotos.
> >> There are two scenarios here:
> >>
> >>      1) goto X  where X is a variable.
> >>      2) goto  (coll at: y)  where coll is a Collection.
> >>
> > There are several ways of implementing this without computed bytecodes in
> > the instruction set, but there is also the possibility of implementing it
> > directly in the instruction set.
> > Off the top of my head one could
> > - map to perform: using some mangled selector.  Yes this is problematic
> > because one has to split the scope between the two methods, so in general
> > it's not a solution
> Doesn't appeal to me.
> > - map to a case statement, which is what Squeak does. Just map it to a
> > sequence of compare and branches.  Or better still, to a binary tree.
> > Coincidentally this is used by the JIT to implement block dispatch in
> > methods that contain more than one block.  I know of other VM
> > implementations using it for PIC dispatch with really good performance.
> I don't know what you mean my Squeak mapping to a case statement since
> there is no case statement in  Squeak/Smalltalk and I can't off hand think
> of
> where one is needed (some Squeak users might feel they need one but that
> is a
> different matter).

But there is.  And it is very convenient when one doesn't want to spread a
case over different classes, or when the cases distribute over values of
the same class (e.g. integer values).  People often claim that a case
statement isn't necessary because one has polymorphism, but unless the
language supports singletons (which of course Smalltalk does not) a case
statement is much more readable than e.g. an if-then-else tree or a
"Dictionary mapping values to blocks or selectors" solution.

Since you're unaware of the case statement I suggest you browse senders of
caseOf: and caseOf:otherwise: in Squeak trunk (which includes selectors of
optimized selectors that end up not being sent) or caseError, which is sent
when there's no otherwise clause.  Here's an example:

menuHook: aMenu named: aSymbol shifted: aBool
"Enhance aMenu with registered services."
{ [ #classListMenu ] -> [ ServiceGui browser: self classMenu: aMenu ].
[ #codePaneMenu ] -> [ ServiceGui browser: self codePaneMenu: aMenu ].
[ #messageCategoryMenu] -> [ ServiceGui browser: self messageCategoryMenu:
aMenu ].
[ #messageListMenu ] -> [ ServiceGui browser: self messageListMenu: aMenu ].
[ #systemCategoryMenu ] -> [ ServiceGui browser: self classCategoryMenu:
aMenu ] }
otherwise: [ "do nothing" ]

This compiles down to a sequence of comparisons.

The use of compare and branches might be OK in some cases
> but a mess for the finite state machines generated from regular
> expressions.
> Actually, even with computed gotos  FSMs are somewhat messy but without
> them
> it's worse.  I don't know what  'PIC dispatch' is.

Polymorphic inline cache dispatch.  In JIT VMs such as Cog polymorphic send
sites are optimized using jump tables, one jump for each class encountered
at the send site, up tio some small limit such as 6 cases.  Right now in
Cog these jump tables are  sequential comparisons.  But in some VMs, where
the degree of polymorphism may be much higher (think prototype languages
such as JavaScript) a binary tree may be much miore efficient, if harder to

To use a binary tree don't I need some kind of computed goto for when I
> reach
> a leaf of the tree????

No.  Once you're at the leaf you just include the bytecodes.  So for
example take the following case statement:

quickPrimitiveGeneratorFor: aQuickPrimitiveIndex
<returnTypeC: 'int (*quickPrimitiveGeneratorFor(sqInt
caseOf: {
[256] -> [#genQuickReturnSelf].
[257] -> [#genQuickReturnConstNil].
[258] -> [#genQuickReturnConstTrue].
[259] -> [#genQuickReturnConstFalse].
[260] -> [#genQuickReturnConstMinusOne].
[261] -> [#genQuickReturnConstZero].
[262] -> [#genQuickReturnConstOne].
[263] -> [#genQuickReturnConstTwo] }
otherwise: [#genQuickReturnInstVar]

This is compiled in the current Squeak compiler as

61 <10> pushTemp: 0
62 <88> dup
63 <2A> pushConstant: 256
64 <B6> send: =
65 <9B> jumpFalse: 70
66 <87> pop
67 <29> pushConstant: #genQuickReturnSelf
68 <A4 35> jumpTo: 123
70 <88> dup
71 <28> pushConstant: 257
72 <B6> send: =
73 <9B> jumpFalse: 78
74 <87> pop
75 <21> pushConstant: #genQuickReturnConstNil
76 <A4 2D> jumpTo: 123
117 <22> pushConstant: 263
118 <B6> send: =
119 <99> jumpFalse: 122
120 <21> pushConstant: #genQuickReturnConstTwo
121 <90> jumpTo: 123
122 <20> pushConstant: #genQuickReturnInstVar
123 <7C> returnTop

but it could be more efficient if the code were

    pushTemp: 0
    pushConstant: 256
    send: <
    jumpFalse: L1
    pushConstant: 260
    send: <
    jumpFalse: L2
    pushConstant: 258
    send: <
    jumpFalse: L3
    pushConstant: 257
    send: <
    jumpFalse: L4
    pushConstant: #genQuickReturnSelf
    jumpTo: L0
    pushConstant: #genQuickReturnConstNil
    jumpTo: L0
    pushConstant: #genQuickReturnInstVar

> - use thisContext pc: value.
> This would be a possibility for me to experiment with for now.  When I
> have a working
> parser generator tool I could campaign for my computed goto instructions
> to be added
> to the VM.
> > This /should/ be fine in the stack VM, but
> > slooooow in the JIT because internally mapping bytecode pcs to machine
> code
> > pcs is slow, and currently slower still because the frame will be
> converted
> > to a pure context and then converted back into a frame on the return from
> > pc:.  But this solution isn't to be rejected out-of-hand.  It can be
> > optimized to avoid the frame conversion and the JIT might be able to
> > optimize it.
> I assume that if computed gotos were used the translation to machine code
> would require a direct
> mapping of (virtually labeled) bytecode locations to machine code
> locations.  I think this can be done
> in a reasonable amount of time but others such as yourself clearly
> understand the issues far better than
> I do.  The dirty solution to start would be to simply not JIT the code
> that uses computed gotos.

Yes, it could be.  The JIT would generate a jump table from the literal
containing the bytecoded pcs, and there would be no conversion; only
indexing the jump table and jumping.  Again, organizing that table as a
binary switch may well be faster on modern architectures.  Indirect jumps
typically involve pipeline stalls, whereas binary jump trees don't.

> The main problem is the compiler has no support for labels so
> > there would be work here.
> I don't mind doing the work but to my way of thinking  "goto X" is pretty
> basic
> and is thus best handled at the VM/byte code level.  Anything else is doing
> in a complicated way something that is fairly simple.  Of course changing
> the VM/byte codes by even a single byte code is a major deal unless done
> when the VM/byte codes are initially created.  Alas I must deal with what
> already exists.  Even so, my preference is to work with the VM if at all
> possible.
> >> For example, one such language is that of regular expressions, which I
> >> wish to translate into finite state machines implemented in VM code.
> >> In this case I need case 2) gotos where coll is a collection of
> >> associations, possibly a
> >> Dictionary. I also plan to write a debugger for this (and other
> languages)
> >> but that is another story.
> >>
> >> I realize that the Cog VM is being built for Smalltalk (Squeak? Pharo?)
> >> for which the goto instructions are not needed and thus I assume
> >> unavailable. But there is something to
> >> viewing a virtual machine as general purpose and thus the target of
> >> multiple languages as is
> >> the case for the Java virtual machine.
> >> If the Cog VM is viewed this way then I argue there is a need for my
> goto
> >> instructions
> >> because some languages have need for them.
> >> For example, many languages have case statements.  (I am all for object
> >> oriented
> >> but I would be willing to accept a case statement in Smalltalk too;  the
> >> Squeak code
> >> implemented one in Squeak doesn't cut it).
> >
> > I've occasionally thought about this for many years.  A computed jump
> might
> > be nice.  Eg index an Array literal of pcs with the integer on top of
> > stack, falling through on bad type or out of range.
> This is the way I am thinking.  If there are other reasons for a computed
> jumpTo
> as well all the better.
> >> Anyway, I am not arguing to Change Squeak or Smalltalk but I am arguing
> >> to have my goto instructions in Cog VM. Is there any chance of this?????
> >>
> > There's no chance of me spending time implementing this any time soon.  I
> > have too much high-priority tasks to tackle this.  But I want to
> encourage
> > you or others to have a go implementing it.  It's fun!
> I understand and am willing to be the one to add one or more computed jump
> instructions, including working on the JIT code generator if needed.
> As you say it should be fun (and also educational).  But
>    1)  I am pretty busy now too and probably won't get to this for a year.
>    2)  If I am to do this it would be great if someone can write a
> specification as to
>         what is to be done.  If someone can write this now that would be
> great but
>         if they write it when I post that I am ready to do the work that
> would also
>         be fine.
>    3)  I don't want to just have my own private VM/byte codes.  I want
> users of my
>         parser generator tool to be able to load it into a standard
> version of Squeak
>         and run it there including the possible generation of compilers
> for compiling
>         their domain specific language programs into byte codes if desired.

Sure.  Get back to me when you're ready to work on it and I'll write the
specification, and help you with whatever info you need.  There should be
room in the bytecode sets we'll likely be using next year.

> >> I don't know the Squeak VM or the Cog VM either but I assume these
> >> instructions don't exist because I see no need of them when the source
> >> language is
> >> Squeak or any version of Smalltalk for that matter. I also assume that
> >> there is already
> >> a full list of 256 instructions in the Cog VM and thus no room for my
> goto
> >> instructions
> >> unless some instructions are removed.
> >>
> >> Are there Cog VM instructions that are so rarely used that they could be
> >> removed without
> >> unreasonably slowing down the Cog VM interpretation of byte codes
> >> generated from Squeak source code?????
> >>
> > The current set has 3 unused bytecodes, one of which Spur uses, so
> > effectively there are two unused bytecodes.
> Levente Uzonyi  in his posting pointed out that only one instruction is
> needed.
> I don't like having to push the address to jump to onto the stack,
> preferring a byte
> code with an argument, but I could live with his solution if that is what
> is decided.
> In the case of  goto  coll at: X  the address is likely to end up on top
> of the stack
> anyway so Levente's  jumpToTop instruction looks good in any case.

If the bytecode is one that takes an integer on top of stack, and an Array
literal containing bytecode pcs, falling through on out of range, then
nothing other than the index need be pushed on top of stack.  That would be
my preference.

> The Cog VMs support multiple bytecode sets.  If you look at the
> > BytecodeSets package on VMMaker you can read the class comments of the
> > BytecodeEncoder subclasses such as EncoderForSistaV1.  These bytecode
> sets
> > have a few more unused bytecodes.  This multiple bytecode set support is
> > better implemented in Spur where there is only one compiled method header
> > format and support for 64k literals.  So let me encourage you to move to
> > Spur and to look at the Sista set.  The class comment of each encoder
> class
> > specifies the instruction set it targets.
> I am prepared to work with Spur and the Sista set.  I am looking for
> someone to
> say that if I do this work that incorporating the work into Spur will be
> seriously
> considered.

I can say that, yes.  I will seriously consider adding it.  I think its a
useful thing in the bytecode set, precisely to allow compiling other
languages to the VM.

> Ralph Boland

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.squeakfoundation.org/pipermail/vm-dev/attachments/20141103/469369c9/attachment-0001.htm

More information about the Vm-dev mailing list