Spaghetti stack?

Spaghetti stack?

[Qiantan Hong]

[-- Attachment #1: Type: text/plain, Size: 523 bytes --]

Hi,

I’m trying to find a way to implement/hack
a runtime using spaghetti stack with libgccjit, 
I think I need to tell libgccjit to use a different
calling convention, either implemented by
myself, or somehow let it use the “bare”
calling convention and generate prologue/
epilogue myself. However, I didn’t find
any API about calling convention — neither
do I have any idea how to express the
prologue/epilogue with libgccjit because they’re
at register level.

Any idea?


Best,
Qiantan


[-- Attachment #2: smime.p7s --]
[-- Type: application/pkcs7-signature, Size: 1858 bytes --]

Re: Spaghetti stack?

[David Malcolm]

On Sat, 2020-10-31 at 05:58 +0000, Qiantan Hong wrote:
> Hi,
> 
> I’m trying to find a way to implement/hack
> a runtime using spaghetti stack with libgccjit, 

I have to confess that I wasn't familiar with the term "spaghetti
stack".

> I think I need to tell libgccjit to use a different
> calling convention, either implemented by
> myself, or somehow let it use the “bare”
> calling convention and generate prologue/
> epilogue myself. However, I didn’t find
> any API about calling convention — neither
> do I have any idea how to express the
> prologue/epilogue with libgccjit because they’re
> at register level.
> 
> Any idea?

That sounds lower level that the approach libgccjit takes; libgccjit's
API tends to mimic what's expressible by the C frontend.

I don't think libgccjit currently exposes a way to set the calling
convention, but IIRC the C frontend does.

Does an existing GCC frontend exposes a way of implementing any of
this?  If so, it can probably be exposed by libgccjit, otherwise we'd
have to add support to GCC for it, which makes it much harder.

Another approach would be to do it "by hand", by implementing the
calling convention by injecting calls to dynamically allocate the
frame, populate it, and then make a call with the parameters suitably
modified - if that makes sense.

Hope this is helpful
Dave

Re: Spaghetti stack?

[Basile Starynkevitch]

On 10/31/20 6:33 PM, David Malcolm via Jit wrote:
> On Sat, 2020-10-31 at 05:58 +0000, Qiantan Hong wrote:
>> Hi,
>>
>> I’m trying to find a way to implement/hack
>> a runtime using spaghetti stack with libgccjit,
> I have to confess that I wasn't familiar with the term "spaghetti
> stack".


IIRS, the Go language and execution model has it. This is how you could 
have millions of goroutines with only a few dozen threads (and call stacks)


Regards.

-- 
Basile STARYNKEVITCH   == http://starynkevitch.net/Basile
opinions are mine only - les opinions sont seulement miennes
Bourg La Reine, France; <basile@starynkevitch.net>
(mobile phone: cf my web page / voir ma page web...)

Re: Spaghetti stack?

[Qiantan Hong]

[-- Attachment #1: Type: text/plain, Size: 1001 bytes --]

Ah, that’s a great hint. I want to use
it to get better call/cc, but that’s kind of similar to
stackless coroutines.

Any idea how Go did it using GCC?

> On Oct 31, 2020, at 1:45 PM, Basile Starynkevitch <basile@starynkevitch.net> wrote:
> 
> 
> On 10/31/20 6:33 PM, David Malcolm via Jit wrote:
>> On Sat, 2020-10-31 at 05:58 +0000, Qiantan Hong wrote:
>>> Hi,
>>> 
>>> I’m trying to find a way to implement/hack
>>> a runtime using spaghetti stack with libgccjit,
>> I have to confess that I wasn't familiar with the term "spaghetti
>> stack".
> 
> 
> IIRS, the Go language and execution model has it. This is how you could have millions of goroutines with only a few dozen threads (and call stacks)
> 
> 
> Regards.
> 
> -- 
> Basile STARYNKEVITCH   == http://starynkevitch.net/Basile
> opinions are mine only - les opinions sont seulement miennes
> Bourg La Reine, France; <basile@starynkevitch.net>
> (mobile phone: cf my web page / voir ma page web...)
> 


[-- Attachment #2: smime.p7s --]
[-- Type: application/pkcs7-signature, Size: 1858 bytes --]

Re: Spaghetti stack?

[David Malcolm]

On Sat, 2020-10-31 at 19:52 +0000, Qiantan Hong wrote:
> Ah, that’s a great hint. I want to use
> it to get better call/cc, but that’s kind of similar to
> stackless coroutines.
> 
> Any idea how Go did it using GCC?

Caveat: I'm not an expert at go

GCC's go frontend is split into two parts:
* gcc/go/gofrontend is shared with the golang and is C++ code that
implements parsing Go to an AST representation
* gcc/go implements a GCC frontend for go using the above, calling it
in the correct place within GCC's framework, and converting the AST
into GCC's IR so that the rest of GCC can optimize it and generate
assembler.

Looking at gcc/go/gofrontend and grepping for "routine" I see there are
Go_statement and Thunk_statement classes that seem to implement
goroutines and these non-trivial calling conventions in terms of
lowering to a more typical to C level of IR - if I'm reading it right.

For example, runtime.def has:

// Start a new goroutine.
DEF_GO_RUNTIME(GO, "__go_go", P2(UINTPTR, POINTER), R1(POINTER))

// Defer a function.
DEF_GO_RUNTIME(DEFERPROC, "runtime.deferproc", P3(BOOLPTR, UINTPTR,
POINTER),
	       R0())

// Defer a function, with stack-allocated defer structure.
DEF_GO_RUNTIME(DEFERPROCSTACK, "runtime.deferprocStack",
               P4(POINTER, BOOLPTR, UINTPTR, POINTER), R0())

which would seem to be relevant.

So I *think* the gofrontend code shared with golang is doing this in a
go-specific way.

Hope this is helpful
Dave

> > On Oct 31, 2020, at 1:45 PM, Basile Starynkevitch <
> > basile@starynkevitch.net> wrote:
> > 
> > 
> > On 10/31/20 6:33 PM, David Malcolm via Jit wrote:
> > > On Sat, 2020-10-31 at 05:58 +0000, Qiantan Hong wrote:
> > > > Hi,
> > > > 
> > > > I’m trying to find a way to implement/hack
> > > > a runtime using spaghetti stack with libgccjit,
> > > I have to confess that I wasn't familiar with the term "spaghetti
> > > stack".
> > 
> > IIRS, the Go language and execution model has it. This is how you
> > could have millions of goroutines with only a few dozen threads
> > (and call stacks)
> > 
> > 
> > Regards.
> > 
> > -- 
> > Basile STARYNKEVITCH   == http://starynkevitch.net/Basile
> > opinions are mine only - les opinions sont seulement miennes
> > Bourg La Reine, France; <basile@starynkevitch.net>
> > (mobile phone: cf my web page / voir ma page web...)
> >