Hi Igor,

On Fri, Nov 15, 2013 at 5:04 AM, Igor Stasenko <siguctua@gmail.com> wrote:

 

On 14 November 2013 17:44, Eliot Miranda <eliot.miranda@gmail.com> wrote:

 

On Thu, Nov 14, 2013 at 3:43 AM, Igor Stasenko <siguctua@gmail.com> wrote:

 

On 12 November 2013 19:23, Eliot Miranda <eliot.miranda@gmail.com> wrote:

 

Hi Igor,

On Tue, Nov 12, 2013 at 3:37 AM, Igor Stasenko <siguctua@gmail.com> wrote:

 

And let me remind you what i proposed back then:

:)
===========
About immediates zoo.

Keep in mind, that the more immediates we have, the more complex implementation
tends to be.

I would just keep 2 data types:
 - integers
 - floats

and third, special 'arbitrary' immediate , which seen by VM as a 60-bit value.
The interpretation of this value depends on lookup in range-table,
where developer specifying the correspondence between the value
interval and class:
[min .. max] -> class

for this idea to go anywhere you'd have to show at least the pseudo-code for the inline cache test in machine code methods.  These schemes seem great in theory but in practice end up with a long, complex and slow fetchClassOf: and/or inline cache test.  To remind you, you have to compete with the following in Spur Cog:

Limm:

andl $0x1, %eax

j Lcmp

Lentry:

movl %edx, %eax

andl $0x3, %eax

jnz Limm

movl 0(%edx), %eax

andl $0x3fffff, %eax

Lcmp:

It is extra comparison for immediate case:

cmp eax, LowRange

jl miss

cmp eax, HighRange

jg miss

I don't understand.  Does that mean there are two cache entries, one for HighRange one for LowRange?  Spell it out please.  Give the pseudo-code.  Right now it is too vague for me to see how it is supposed to work.

Two entries where? If i understand, in COG you using generated code for inline cache, by jumping onto PIC entry for checking if it hits or misses. So, you just generate different code for immediate case(s), to check that receiver oop value is tagged and within certain range (cache hit) , or not (cache miss).

That's not right.  PICs are only used at send sites that prove to be polymorphic, i.e. become PICs after a send failure.  At first they are simply monomorphic inline caches with a single entry.  Only 9% of active send sites fail to be rebound to PICs.

   

A pseudo-code for bit-pattern matching is this:

"presuming least significant bit is tag"

(oop bitAnd: 2r101) == 2r101 ifTrue: [ hit ] ifFalse: [ miss ]

And code for range matching is this:

"presuming least significant bit is tag"

(oop odd and: [ oop between: lowValue and: highValue]) ifTrue: [ hit ] ifFalse: [ miss ]

"or presuming most significant bit is tag"

(oop between: lowValue and: highValue) ifTrue: [ hit ] ifFalse: [ miss ]

and how are these two pieces of code related?

 so, if value between low & high range, it is already known what class it is otherwise

you got cache miss.

And this scheme works even better if you pick highest bit for tagging, like that you can get rid of testing tag bit completely:

andl $0x1, %eax

j Lcmp

and start straight from comparing with high and low values.

That of course, if we can use highest bit for tagging.

The class lookup then will be binary search in table of ranges, which is O(log n).

So the inline cache becomes a range of entries?  That will cause a significant code bloat.

i'm not sure i understood, what that extra inline cache you talking about?
presuming that each PIC is generated code, you just generate different code for entry which specialized to detect
that given oop is of certain immediate class.

But if its not, then yes, you need to hold these two values somewhere to be able to compare them with input oop.

But even in that case, i fail to see how single extra comparison could cause so much bloat.

And it always good to compare, how much bloat will cause bit pattern matching code.

What is clearly beneficial for immediate ranges that you can redefine them at any moment by introducing

new kinds of immediates , without need to even touch VM: language can easily control that.

And the potential number of immediate classes is much bigger , if you encode them in ranges, because:

2^31 (smallint) + 2^32 (short float) + 2^24 (character unicode)  = 6459228160

which leaves us:

2^63 - 6459228160 =  9223372030395547648

space for other kinds of immediates, which we can introduce later !!without!! need to ever bother VM again.

--
Best regards,
Igor Stasenko.

--
best,

Eliot