[Vm-dev] Primitive replaceFrom:to:with:startingAt: in the JIT

Clément Bera bera.clement at gmail.com
Wed Dec 27 22:56:14 UTC 2017

On Wed, Dec 27, 2017 at 10:37 PM, Eliot Miranda <eliot.miranda at gmail.com>

> Hi Clément,
> On Dec 27, 2017, at 12:14 PM, Clément Bera <bera.clement at gmail.com> wrote:
> Hi,
>> Indeed, I implemented the quick paths only for byte objects and pointer
>> objects. The reason for this is that I based the implementation on customer
>> benchmarks, and there was a lot of replacements for pointer objects
>> (typically, growing the inner array of OrderedCollection/Dictionaries)
>> and byte objects (typically bytestrings concatenation, such as 'a' , 'b'
>> which calls twice this primitive for only a couple characters and is now
>> much faster).
>> I wrote the quick path with short copies mind, in which case switching to
>> the C runtime is too expensive. Based on your analysis, for pointer objects
>> it looks good, but for byte objects it gets slower.
>> So I see two solutions for the slow down on byte copies:
>> 1) over a threshold of 1000 elements, fall back to the slang primitive
>> for byte objects
>> 2) Generate smarter copying machine code (for example with some kind of
>> duff device / unrolled loop / overlapping copies with 1 copy non aligned).
>> If possible nothing requiring to extend the RTL. Note that memcopy in C is
>> written as a per back-end files which is thousands of lines on x86 and we
>> don't want to go anywhere near this complexity, in our case the size of
>> generate machine code also matters and we want to keep it small. We could
>> use pointer size read / writes on byte objects too if we deal with aliasing
>> (if array == replArray just fall back to a slower path).
>> Let's look at the naive machine code in Cog's RTL (I could show the
>> actual assembly version if you like it better), in each case the loop does:
>> - compare and jumpBelow to see if the copy is finished
>> - read one field from repl
>> - write the read field in the array
>> - increment the 2 registers holding indexes for read/write
>> - jump back.
>> The copying machine code for byte object is as follow:
>>                        instr := cogit CmpR: startReg R: stopReg.
>> jumpFinished := cogit JumpBelow: 0.
>> cogit MoveXbr: repStartReg R: replReg R: TempReg.
>> cogit MoveR: TempReg Xbr: startReg R: arrayReg.
>> cogit AddCq: 1 R: startReg.
>> cogit AddCq: 1 R: repStartReg.
>> cogit Jump: instr.
> I think the thing to do is to implement a word copier for byte objects.
> It could use up to two registers for the source to assemble a word when the
> source is not aligned mod a word boundary.  The copying engine would copy
> bytes up to the start of a word, then copy words, then copy trailing bytes.
> Moving 4 or 8 bytes in the inner loop is probably more than good enough to
> exceed the interpreter primitive's speed, without the complexity of a
> duff's device dispatch.  Note that fetching the first word of the
> destination, when there are odd initial bytes is probably a good way of
> making the initial odd bytes write cheap.  Or fetch the destination, mask
> out the bytes that are written to, add/or in the source bytes.

Yes that's the kind of thing what I was thinking of. In sista image side I
do it for byte equality in some cases and it speeds things up significantly
(though I do it only when everything is aligned like I want, which is the
common case).

Assembling a word when reads are not aligned with the writes seemed tedious
- not sure if that's the best time invested/speed-up ratio. One could

> Another thing is to use derived pointers that are incremented instead of
> the Xbr form, which wastes a register better used for buffering bytes into
> whole word writes.

Yeah well I thought about it, just make sure to restore the register to
return correctly with the array. I've felt it was not worth it - we don't
need more registers at this point and the processor already internally
optimise those tights loops. I made the simplest thing that could possibly
work. The primitive code is quite complex already, mostly due to all the
range/type checks, but still...

We could experiment a bit and evaluate performance but I don't want to
waste too much time. I watched at google a couple weeks ago an hour talk on
building fast memcopy, the result of 4 people working for weeks if not
months, and I don't want to go there - their implementation has like 20
different ways of copying things depending on the number of bytes to copy. I
am not going to implement 20 versions and check they all work and that they
are all fast.

Maybe just falling back to C code if there are more than 1000 elements to
copy is good enough... At least on the short term...

Of course one alternative is extend the RTL... to use the REP prefix on
intel and bulk memory moves on ARM... I don't want to go there either but
someone could.

> Isn't it great to finally be at a stage where we can meaningfully have
> these JIT vs C compiler performance discussions?

So many things I would like to do.

>> The copying code for pointer object is as follow:
>> instr := cogit CmpR: startReg R: stopReg.
>> jumpFinished := cogit JumpBelow: 0.
>> cogit MoveXwr: repStartReg R: replReg R: TempReg.
>> cogit MoveR: TempReg Xwr: startReg R: arrayReg.
>> cogit AddCq: 1 R: startReg.
>> cogit AddCq: 1 R: repStartReg.
>> cogit Jump: instr.
>> Any idea what we could do smarter, but not too complex, without extending
>> the RTL and without generating let's say more than 20 extra instructions ?
>> I think checking for aliasing and using pointer size read/write may be
>> the way to go. Then if performance is not better we just fall back to slang
>> code.
>> On Mon, Dec 25, 2017 at 10:57 PM, Levente Uzonyi <leves at caesar.elte.hu>
>> wrote:
>>> Hi Clémont,
>>> I finally found the time to write some benchmarks.
>>> I compared the output of the script below on sqcogspur64linuxht vm
>>> 201710061559 and 201712221331 avaiable on bintray.
>>> result := { ByteArray. DoubleByteArray. WordArray. DoubleWordArray.
>>> ByteString. WideString. FloatArray. Array } collect: [ :class |
>>>         | collection |
>>>         Smalltalk garbageCollect.
>>>         collection := class basicNew: 10000.
>>>         class -> (#(0 1 2 5 10 20 50 100 200 500 1000 2000 5000 10000)
>>> collect: [ :size |
>>>                 | iterations time overhead |
>>>                 iterations := (40000000 // (size max: 1) sqrt) floor.
>>>                 overhead := [ 1 to: iterations do: [ :i | ] ] timeToRun.
>>>                 time := [ 1 to: iterations do: [ :i |
>>>                         collection replaceFrom: 1 to: size with:
>>> collection startingAt: 1 ] ] timeToRun.
>>>                 { size. iterations. time - overhead } ]) ].
>>> I found that the quick paths are probably only implented for bytes and
>>> pointers collections, because there was no significant difference for
>>> DoubleByteArray, WordArray, DoubleWordArray, WideString and FloatArray.
>>> For pointers and bytes collections, there's significant speedup when the
>>> copied portion is small. However, somewhere between 50 and 100 copied
>>> elements, the copying of bytes collections becomes slower (up to 1.5x @
>>> 100k elements) with the newer VM.
>>> It's interesting that this doesn't happen to pointers classes. Instead
>>> of slowdown there's still 1.5x speedup even at 100k elements.
>>> Levente
>>> On Mon, 23 Oct 2017, Clément Bera wrote:
>>> Hi all,
>>>> For a long time I was willing to add primitive
>>>> #replaceFrom:to:with:startingAt: in the JIT but did not take time to
>>>> do it. These days I am showing the JIT to one of my students and as an
>>>> example of how one would write code in the JIT we implemented this primitive
>>>> together, Spur-only. This is part of commit 2273.
>>>> I implemented quick paths for byte objects and array-like objects only.
>>>> The rationale behind this is that the most common cases I see in Pharo user
>>>> benchmarks in the profiler is copy of arrays and byteStrings. Typically
>>>> some application benchmarks would show 3-5% of
>>>> time spent in copying small things, and switching from the JIT runtime
>>>> to C runtime is an important part of the cost.
>>>> First evaluation shows the following speed-ups, but I've just done that
>>>> quickly in my machine:
>>>> Copy of size 0
>>>>     Array 2.85x
>>>>     ByteString 2.7x
>>>> Copy of size 1
>>>>     Array 2.1x
>>>>     ByteString 2x
>>>> Copy of size 3
>>>>     Array 2x
>>>>     ByteString 1.9x
>>>> Copy of size 8
>>>>     Array 1.8x
>>>>     ByteString 1.8x
>>>> Copy of size 64
>>>>    Array 1.1x
>>>>    ByteString 1.1x
>>>> Copy of size 1000
>>>>    Array 1x
>>>>    ByteString 1x
>>>> So I would expect some macro benchmarks to get 1 to 3% percent
>>>> speed-up. Not as much as I expected but it's there.
>>>> Can someone who is good at benchmarks such as Levente have a look and
>>>> provide us with a better evaluation of the performance difference ?
>>>> Thanks.
>>>> --
>>>> Clément Béra
>>>> https://clementbera.wordpress.com/
>>>> Bâtiment B 40, avenue Halley 59650 Villeneuve d'Ascq
>> --
>> Clément Béra
>> https://clementbera.wordpress.com/
>> Bâtiment B 40, avenue Halley 59650 *Villeneuve d'Ascq*

Clément Béra
Pharo consortium engineer
Bâtiment B 40, avenue Halley 59650 Villeneuve d'Ascq
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.squeakfoundation.org/pipermail/vm-dev/attachments/20171227/19ccb66d/attachment-0001.html>

More information about the Vm-dev mailing list