Quantcast

Teleplace Cog VMs are now available

classic Classic list List threaded Threaded
41 messages Options
123
Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Teleplace Cog VMs are now available

Eliot Miranda-2

Hi All,


it gives me great pleasure to announce that the Teleplace Cog VMs are now available.  Huge thanks to all at Teleplace who have given me the opportunity to build Cog and release it as open source, been willing guinea pigs braving its bugs, and providing indispensable participation in getting Cog to its current state.  Huge thanks are also due to the original Back To The Future team whose VMMaker Cog extends to write the VM, and to Peter Deutsch from whom I've taken many ideas.


This release contains two VMs.  The Stack VM, is a cross-platform interpreter that uses context-to-stack mapping to achieve modest performance gains.  The Cog VM is a just-in-time compiler that currently supports only x86 that builds upon the Stack VM to achieve substantial performance improvements.  The release is in the form of a Monticello package containing the VMMaker source and a tarball containing the platform sources, the generated sources and a Squeak 4.1 image containing the VMMaker sources.  Download both at

http://ftp.squeak.org/Cog/VMMaker-oscog.11.mcz

http://ftp.squeak.org/Cog/OpenSourceCog.tar.gz


Cog VMs:


The Cog VMs are Squeak/Croquet VMs that run closure Squeak/Croquet/Pharo/Cuis images. The VMs support existing plugin source but will require plugins to be recompiled as the VM_PROXY_MAJOR plugin api has been extended.


This release contains two distinct VMs, the StackInterpreter and the Cogit.  The StackInterpreter is a fully-portable plug-in replacement for the current closure Squeak VMs and images.  The Stack VM uses context-to-stack mapping and a somewhat improved garbage collector to achieve modest but useful performance gains in the 10% to 15% range.  The StackInterpreter is intended to supersede the Squeak VM on platforms where the Cogit cannot be used.  The Cogit extends the StackInterpreter with a just-in-time compiler that uses aggressive inline caching techniques to deliver substantial performance gains in the 3x to 15x range, depending on benchmark.  The Cogit currently supports only x86 and the floating-point primitives and parts of the platform support code depend on SSE2.  I hope members of the community will attempt to port it, e.g. to ARM, PowerPC and x86-64.  The Cogit (excuse the pun) is so named because it is both an interpreter and a JIT, choosing not to generate machine code for large methods, interpreting them instead, the default policy being not to JIT methods with more than 60 literals.


The Cog VM requires a few minor image changes all in image/NecessaryImageChangesForCogToWork.1.cs.  The JIT's machine-code SmallInteger primitives insist on a SmallInteger receiver so the primitives in LargePositiveInteger = ~= bitAnd: bitOr: butShift: and bitXor: cannot be used and these methods must be deleted.  The Cogit inlines the address of the Character instance table, Smalltalk specialObjectsArray at: 25, into the machine-code at: primitive for faster ByteString>>at: and so the table cannot be rebuilt in SmalltalkImage>>recreateSpecialObjectsArray.  The new version preserves the existing table.  Both VMs maintain floats in platform order to ease implementation of machine code floating-point primitives, and hence internally are in little-endian order instead of big-endian in current Squeak images.  While the VMs convert float order automatically on load they do require special accessing primitives Float>>basicAt: & Float>>basicAt:put: that undo the reversal and answer Float contents in big-endian order so that e.g. Float>>hash is unchanged.  The methods assume these primitives can fail, allowing the code to be used on current Squeak VMs.


The image/VMMaker-Squeak4.1.image is a Squeak 4.1 image, runnable with the current Squeak VMs, that contains these changes, and can hence also be run with a Cog VM.  But beware, once an image has been saved on Cog it cannot be run by an existing Squeak VM, because existing VMs cannot undo the Float order change.



Platform Subsystem:


Most of the platform subsystem is unchanged but there are some important changes that need description.  The biggest change is the heartbeat and the clock in platforms/unix/vm/sqUnixHeartbeat.c and platforms/win32/vm/sqWin32Heartbeat.c.  The Cog VMs avoid the slow and variable interruptCheckCounter, folding the event check into the stack overflow check on frame build.  The heartbeat, typically 500Hz or 1KHz, changes the stackLimit to a value that will always fail.  On the next frame building send the VM will enter stack overflow handling that, as a side effect, will also check for events.  This is more efficient than the update of interruptCheckCounter and much more regular.  If one is running code that executes long-running primitives (e.g. large integer arithmetic) the counter approach will result in too low an interrupt check frequency, and conversely if one is running normal code the interrupt check frequency can be very high.


The heartbeat also maintains a 64-bit microsecond clock, UTC microseconds from 1901, from which the backward-compatible millisecond and second clocks are derived.  Primitives exist to answer UTC microseconds and local microseconds.  Updating the clock in the heartbeat results in a 1 or 2 millisecond resolution but avoids the cost of accessing the OS time on every prim tie which we've found important for performance at Teleplace.  The 64-bit microsecond clocks provide a unified time basis and eliminate wrapping (for the next 54,000 years at least).  I hope community images will move to these clocks.  It's worked well in VisualWorks.


Another significant change is in the external semaphore table support code.  This is now lock-free at the cost of having to specify a maximum number of external semaphores at start-up (default 256).  The support code for the lock-free data structures are processor-specific and is currently implemented only for x86 and gcc-compatible compilers; see platforms/Cross/vm/{sqAtomicOps.h,sqMemoryFence.h}.


There is also improved crash reporting code that prints a primitive log and a C backtrace in addition to the Smalltalk backtrace.  See platforms/Mac OS/vm/sqMacMain.c, platforms/unix/vm/sqUnixMain.c, platforms/win32/vm/sqWin32Intel.c & platforms/win32/vm/sqWin32Backtrace.c.


Finally there is support for the QVMProfiler, a pc-sampling profiler for profiling at the VM level.  See platforms/unix/vm/sqUnixVMProfile.c and platforms/win32/vm/sqWin32VMProfile.c.  The profiler itself is in the VMMaker image described below in Qwaq-VMProfiling.


There are also changes to do with Teleplace-specific extensions to the HostWindowPlugin but these are not essential to Cog.



VMMaker and Slang:


The image/VMMaker-Squeak4.1.image Squeak 4.1 image contains the complete Cog VMMaker with necessary support code for simulation. This image was used to generate the sources in the src and stacksrc directories.


Cog's VMMaker is substantially revised and extended from the current VMMaker.  It supports multiple classes, not just Interpreter and superclasses, because both context-to-stack mapping and the Cogit are too complex to write monolithically.  Classes can specify ancilliaryClasses and ancilliaryStructClasses, such as CoInterpreterStackPage, CogMethod and CogAbstractInstruction.  The Monticello package version is included in the header of all generated files and constitutes the version stamp for generated code.  Code is generated in sorted order so that minor changes in the Smalltalk source produce correspondingly minor changes in the generated code.  The gnuification step is built-in to VMMaker.  No effort has been made to maintain 64-bit compatibility.  Apologies, this was unaffordable.


The VMMaker generates a single source tree used by all platforms.  Instead of deciding at generation time whether to use the Interpreter struct the generated code depends on the SQ_USE_GLOBAL_STRUCT define which can be overridden in platform makefiles.  All plugins live in src/plugins and platform makefiles along with plugins.int and plugins.ext files in the build subdirectories decide which plugins are built as external or internal.  The VM Generation Workspace from Workspace.text workspace contains dots to generate the sources.  We no longer use the VMMakerTool since there should be nothing platform-specific in the generated sources (if we add ports to other ISAs all their source can be included and selected as required by the platform makefiles).


Since the Cogit generates x86 machine code simulation is much more complex.  There is a support plugin, platforms/Cross/plugins/BochsIA32Plugin that depends on a large simulation of the x86 family implemented in C++ (see  processors/IA32/bochs) and on Alien.  I use the simulator frequently (but note that I haven't had time to build a working version for Squeak 4.1).  I have tested Cog simulation in this image, running on the image/VMMaker-Squeak4.1.image itself.  The VM Simulation Workspace in the VMMaker image contains an example doit that starts the simulator. Be patient, even on a fast machine unhibernating the Squeak display background image takes nearly a minute.  Native fonts do not (yet) simulate correctly, but the system runs.  But note that I have only attempted to build and run the simulator on Mac OS X.  I expect Bochs can be built on linux and win32 but I have not tried.  By the way, I've not described how to run the Bochs simulator on the current Squeak VM.  That's because the plugin depends on the heartbeat to break out of simulation occasionally via a new interpreterProxy entry point setInterruptCheckChain.  As this isn't supported by the current Squeak VMs the plugin would require modification.  So to simulate first build either of the Cog VMs and then run the simulation with it.


There are a number of unpublished changes to the base other than those in NecessaryImageChangesForCogToWork.1.cs.  This is partly laziness on my part, partly avoiding publishing things in advance of Cog.  These changes are better motivated once Cog is in use.  There are changes to the "translated primitives" (see implementors of translatedPrimitives) which replace messages with method tags for generation directives.  The Cog VMMaker uses Object>>perform:with:with:with:with: & Object>>perform:with:with:with:with:with: during simulation, and Collection>>#fold: & SquenceableCollection>>#copyUpThrough: during generation.  Object>>inline: and Object var:declareC:, which are mispackaged in Kernel in Squeak 4.1 are obsolete (method tags being used instead) and have been removed. I have changed Integer>>hex and Integer>>hex8 back to their original semantics as of 3.8.  Backward compatibility is important and one can easily add new selectors if one wants different functionality.  VMMaker was here first ;)



Tarball:


The top-level directories in the tarball are


src

the tree for the Cog generated sources including all plugins

stacksrc/vm

the directory containing the Stack VM source (plugins can be taken from above)

platforms

the usual svn platform tree but including Cog specific changes such as the heartbeat

processors

the tree containing simulation support code, i.e. the bochs C++ x86 simulation library, along with a potential ARM, PowerPC & MIPS simulator, Skeye.


image

the Cog-prepared Squeak 4.1 VMMaker image

scripts

some svn scripts to revert unchanged plugins that haven't really changed


cygwinbuild

the win32 build directory

winbuild

the old win32 build directory for minnow gcc 2.95.  Not entirely obsolete as the cygwin build as yet fails to generate a functional FFIPlugin

macbuild

the CoreVM.xcodeproj and support build projects for Mac OS X 10.5 or better

unixbuild

the build directory for linux



Building Cog:


Each build directory above contains a HowToBuild file that describes building in more detail.  The build directories only contain Cogit makefiles.  f you want to build a Stack VM you're on your own but this is very close to the existing Squeak VM build.



Status:

The Cogit VM has been our sole VM at Teleplace for nearly a year.  We do occasionally find bugs and there are almost certainly areas of functionality that we have not touched (for example I know that co-routining does not yet work).  If you find a bug please try and create a reproducible test case and let me know.  I can't promise to take a look or fix it but I am motivated to do so and will try my best as time allows.  Better still if you find and fix bugs be sure to let me know.


License (MIT):

All contributions from Teleplace in this release are

Copyright (c) 2010 Teleplace, Inc.


Permission is hereby granted, free of charge, to any person obtaining a copy

of this software and associated documentation files (the "Software"), to deal

in the Software without restriction, including without limitation the rights

to use, copy, modify, merge, publish, distribute, sublicense, and/or sell

copies of the Software, and to permit persons to whom the Software is

furnished to do so, subject to the following conditions:


The above copyright notice and this permission notice shall be included in

all copies or substantial portions of the Software.


THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR

IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,

FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE

AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER

LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,

OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN

THE SOFTWARE.


Eliot Miranda

June 2010




Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: Teleplace Cog VMs are now available

Michael Haupt-3
Eliot,

this is absolutely wonderful news. Thank you and everyone at Teleplace
who made this possible.

Best wishes,

Michael

On Sun, Jun 20, 2010 at 10:11 PM, Eliot Miranda <[hidden email]> wrote:

> Hi All,
>
> it gives me great pleasure to announce that the Teleplace Cog VMs are now
> available.  Huge thanks to all at Teleplace who have given me the
> opportunity to build Cog and release it as open source, been willing guinea
> pigs braving its bugs, and providing indispensable participation in getting
> Cog to its current state.  Huge thanks are also due to the original Back To
> The Future team whose VMMaker Cog extends to write the VM, and to Peter
> Deutsch from whom I've taken many ideas.
>
> This release contains two VMs.  The Stack VM, is a cross-platform
> interpreter that uses context-to-stack mapping to achieve modest performance
> gains.  The Cog VM is a just-in-time compiler that currently supports only
> x86 that builds upon the Stack VM to achieve substantial performance
> improvements.  The release is in the form of a Monticello package containing
> the VMMaker source and a tarball containing the platform sources, the
> generated sources and a Squeak 4.1 image containing the VMMaker sources.
> Download both at
>
> http://ftp.squeak.org/Cog/VMMaker-oscog.11.mcz
>
> http://ftp.squeak.org/Cog/OpenSourceCog.tar.gz
>
> Cog VMs:
>
> The Cog VMs are Squeak/Croquet VMs that run closure
> Squeak/Croquet/Pharo/Cuis images. The VMs support existing plugin source but
> will require plugins to be recompiled as the VM_PROXY_MAJOR plugin api has
> been extended.
>
> This release contains two distinct VMs, the StackInterpreter and the Cogit.
> The StackInterpreter is a fully-portable plug-in replacement for the current
> closure Squeak VMs and images.  The Stack VM uses context-to-stack mapping
> and a somewhat improved garbage collector to achieve modest but useful
> performance gains in the 10% to 15% range.  The StackInterpreter is intended
> to supersede the Squeak VM on platforms where the Cogit cannot be used.  The
> Cogit extends the StackInterpreter with a just-in-time compiler that uses
> aggressive inline caching techniques to deliver substantial performance
> gains in the 3x to 15x range, depending on benchmark.  The Cogit currently
> supports only x86 and the floating-point primitives and parts of the
> platform support code depend on SSE2.  I hope members of the community will
> attempt to port it, e.g. to ARM, PowerPC and x86-64.  The Cogit (excuse the
> pun) is so named because it is both an interpreter and a JIT, choosing not
> to generate machine code for large methods, interpreting them instead, the
> default policy being not to JIT methods with more than 60 literals.
>
> The Cog VM requires a few minor image changes all in
> image/NecessaryImageChangesForCogToWork.1.cs.  The JIT's machine-code
> SmallInteger primitives insist on a SmallInteger receiver so the primitives
> in LargePositiveInteger = ~= bitAnd: bitOr: butShift: and bitXor: cannot be
> used and these methods must be deleted.  The Cogit inlines the address of
> the Character instance table, Smalltalk specialObjectsArray at: 25, into the
> machine-code at: primitive for faster ByteString>>at: and so the table
> cannot be rebuilt in SmalltalkImage>>recreateSpecialObjectsArray.  The new
> version preserves the existing table.  Both VMs maintain floats in platform
> order to ease implementation of machine code floating-point primitives, and
> hence internally are in little-endian order instead of big-endian in current
> Squeak images.  While the VMs convert float order automatically on load they
> do require special accessing primitives Float>>basicAt: &
> Float>>basicAt:put: that undo the reversal and answer Float contents in
> big-endian order so that e.g. Float>>hash is unchanged.  The methods assume
> these primitives can fail, allowing the code to be used on current Squeak
> VMs.
>
> The image/VMMaker-Squeak4.1.image is a Squeak 4.1 image, runnable with the
> current Squeak VMs, that contains these changes, and can hence also be run
> with a Cog VM.  But beware, once an image has been saved on Cog it cannot be
> run by an existing Squeak VM, because existing VMs cannot undo the Float
> order change.
>
> Platform Subsystem:
>
> Most of the platform subsystem is unchanged but there are some important
> changes that need description.  The biggest change is the heartbeat and the
> clock in platforms/unix/vm/sqUnixHeartbeat.c and
> platforms/win32/vm/sqWin32Heartbeat.c.  The Cog VMs avoid the slow and
> variable interruptCheckCounter, folding the event check into the stack
> overflow check on frame build.  The heartbeat, typically 500Hz or 1KHz,
> changes the stackLimit to a value that will always fail.  On the next frame
> building send the VM will enter stack overflow handling that, as a side
> effect, will also check for events.  This is more efficient than the update
> of interruptCheckCounter and much more regular.  If one is running code that
> executes long-running primitives (e.g. large integer arithmetic) the counter
> approach will result in too low an interrupt check frequency, and conversely
> if one is running normal code the interrupt check frequency can be very
> high.
>
> The heartbeat also maintains a 64-bit microsecond clock, UTC microseconds
> from 1901, from which the backward-compatible millisecond and second clocks
> are derived.  Primitives exist to answer UTC microseconds and local
> microseconds.  Updating the clock in the heartbeat results in a 1 or 2
> millisecond resolution but avoids the cost of accessing the OS time on every
> prim tie which we've found important for performance at Teleplace.  The
> 64-bit microsecond clocks provide a unified time basis and eliminate
> wrapping (for the next 54,000 years at least).  I hope community images will
> move to these clocks.  It's worked well in VisualWorks.
>
> Another significant change is in the external semaphore table support code.
> This is now lock-free at the cost of having to specify a maximum number of
> external semaphores at start-up (default 256).  The support code for the
> lock-free data structures are processor-specific and is currently
> implemented only for x86 and gcc-compatible compilers; see
> platforms/Cross/vm/{sqAtomicOps.h,sqMemoryFence.h}.
>
> There is also improved crash reporting code that prints a primitive log and
> a C backtrace in addition to the Smalltalk backtrace.  See platforms/Mac
> OS/vm/sqMacMain.c, platforms/unix/vm/sqUnixMain.c,
> platforms/win32/vm/sqWin32Intel.c & platforms/win32/vm/sqWin32Backtrace.c.
>
> Finally there is support for the QVMProfiler, a pc-sampling profiler for
> profiling at the VM level.  See platforms/unix/vm/sqUnixVMProfile.c and
> platforms/win32/vm/sqWin32VMProfile.c.  The profiler itself is in the
> VMMaker image described below in Qwaq-VMProfiling.
>
> There are also changes to do with Teleplace-specific extensions to the
> HostWindowPlugin but these are not essential to Cog.
>
> VMMaker and Slang:
>
> The image/VMMaker-Squeak4.1.image Squeak 4.1 image contains the complete Cog
> VMMaker with necessary support code for simulation. This image was used to
> generate the sources in the src and stacksrc directories.
>
> Cog's VMMaker is substantially revised and extended from the current
> VMMaker.  It supports multiple classes, not just Interpreter and
> superclasses, because both context-to-stack mapping and the Cogit are too
> complex to write monolithically.  Classes can specify ancilliaryClasses and
> ancilliaryStructClasses, such as CoInterpreterStackPage, CogMethod and
> CogAbstractInstruction.  The Monticello package version is included in the
> header of all generated files and constitutes the version stamp for
> generated code.  Code is generated in sorted order so that minor changes in
> the Smalltalk source produce correspondingly minor changes in the generated
> code.  The gnuification step is built-in to VMMaker.  No effort has been
> made to maintain 64-bit compatibility.  Apologies, this was unaffordable.
>
> The VMMaker generates a single source tree used by all platforms.  Instead
> of deciding at generation time whether to use the Interpreter struct the
> generated code depends on the SQ_USE_GLOBAL_STRUCT define which can be
> overridden in platform makefiles.  All plugins live in src/plugins and
> platform makefiles along with plugins.int and plugins.ext files in the build
> subdirectories decide which plugins are built as external or internal.  The
> VM Generation Workspace from Workspace.text workspace contains dots to
> generate the sources.  We no longer use the VMMakerTool since there should
> be nothing platform-specific in the generated sources (if we add ports to
> other ISAs all their source can be included and selected as required by the
> platform makefiles).
>
> Since the Cogit generates x86 machine code simulation is much more complex.
> There is a support plugin, platforms/Cross/plugins/BochsIA32Plugin that
> depends on a large simulation of the x86 family implemented in C++ (see
> processors/IA32/bochs) and on Alien.  I use the simulator frequently (but
> note that I haven't had time to build a working version for Squeak 4.1).  I
> have tested Cog simulation in this image, running on the
> image/VMMaker-Squeak4.1.image itself.  The VM Simulation Workspace in the
> VMMaker image contains an example doit that starts the simulator. Be
> patient, even on a fast machine unhibernating the Squeak display background
> image takes nearly a minute.  Native fonts do not (yet) simulate correctly,
> but the system runs.  But note that I have only attempted to build and run
> the simulator on Mac OS X.  I expect Bochs can be built on linux and win32
> but I have not tried.  By the way, I've not described how to run the Bochs
> simulator on the current Squeak VM.  That's because the plugin depends on
> the heartbeat to break out of simulation occasionally via a new
> interpreterProxy entry point setInterruptCheckChain.  As this isn't
> supported by the current Squeak VMs the plugin would require modification.
> So to simulate first build either of the Cog VMs and then run the simulation
> with it.
>
> There are a number of unpublished changes to the base other than those in
> NecessaryImageChangesForCogToWork.1.cs.  This is partly laziness on my part,
> partly avoiding publishing things in advance of Cog.  These changes are
> better motivated once Cog is in use.  There are changes to the "translated
> primitives" (see implementors of translatedPrimitives) which replace
> messages with method tags for generation directives.  The Cog VMMaker uses
> Object>>perform:with:with:with:with: &
> Object>>perform:with:with:with:with:with: during simulation, and
> Collection>>#fold: & SquenceableCollection>>#copyUpThrough: during
> generation.  Object>>inline: and Object var:declareC:, which are mispackaged
> in Kernel in Squeak 4.1 are obsolete (method tags being used instead) and
> have been removed. I have changed Integer>>hex and Integer>>hex8 back to
> their original semantics as of 3.8.  Backward compatibility is important and
> one can easily add new selectors if one wants different functionality.
> VMMaker was here first ;)
>
> Tarball:
>
> The top-level directories in the tarball are
>
> src
>
> the tree for the Cog generated sources including all plugins
>
> stacksrc/vm
>
> the directory containing the Stack VM source (plugins can be taken from
> above)
>
> platforms
>
> the usual svn platform tree but including Cog specific changes such as the
> heartbeat
>
> processors
>
> the tree containing simulation support code, i.e. the bochs C++ x86
> simulation library, along with a potential ARM, PowerPC & MIPS simulator,
> Skeye.
>
> image
>
> the Cog-prepared Squeak 4.1 VMMaker image
>
> scripts
>
> some svn scripts to revert unchanged plugins that haven't really changed
>
> cygwinbuild
>
> the win32 build directory
>
> winbuild
>
> the old win32 build directory for minnow gcc 2.95.  Not entirely obsolete as
> the cygwin build as yet fails to generate a functional FFIPlugin
>
> macbuild
>
> the CoreVM.xcodeproj and support build projects for Mac OS X 10.5 or better
>
> unixbuild
>
> the build directory for linux
>
> Building Cog:
>
> Each build directory above contains a HowToBuild file that describes
> building in more detail.  The build directories only contain Cogit
> makefiles.  f you want to build a Stack VM you're on your own but this is
> very close to the existing Squeak VM build.
>
> Status:
>
> The Cogit VM has been our sole VM at Teleplace for nearly a year.  We do
> occasionally find bugs and there are almost certainly areas of functionality
> that we have not touched (for example I know that co-routining does not yet
> work).  If you find a bug please try and create a reproducible test case and
> let me know.  I can't promise to take a look or fix it but I am motivated to
> do so and will try my best as time allows.  Better still if you find and fix
> bugs be sure to let me know.
>
> License (MIT):
>
> All contributions from Teleplace in this release are
>
> Copyright (c) 2010 Teleplace, Inc.
>
> Permission is hereby granted, free of charge, to any person obtaining a copy
>
> of this software and associated documentation files (the "Software"), to
> deal
>
> in the Software without restriction, including without limitation the rights
>
> to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
>
> copies of the Software, and to permit persons to whom the Software is
>
> furnished to do so, subject to the following conditions:
>
> The above copyright notice and this permission notice shall be included in
>
> all copies or substantial portions of the Software.
>
> THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
>
> IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
>
> FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
>
> AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
>
> LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
> FROM,
>
> OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
>
> THE SOFTWARE.
>
> Eliot Miranda
>
> June 2010
>
>
>
>

Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: Teleplace Cog VMs are now available

garduino
In reply to this post by Eliot Miranda-2
Hi Eliot:

Thanks you very much (and also to the Teleplace people) by share this
really big step in the VM area.

Excellent news to this sunday!!


2010/6/20 Eliot Miranda <[hidden email]>:

> Hi All,
>
> it gives me great pleasure to announce that the Teleplace Cog VMs are now
> available.  Huge thanks to all at Teleplace who have given me the
> opportunity to build Cog and release it as open source, been willing guinea
> pigs braving its bugs, and providing indispensable participation in getting
> Cog to its current state.  Huge thanks are also due to the original Back To
> The Future team whose VMMaker Cog extends to write the VM, and to Peter
> Deutsch from whom I've taken many ideas.
>
> This release contains two VMs.  The Stack VM, is a cross-platform
> interpreter that uses context-to-stack mapping to achieve modest performance
> gains.  The Cog VM is a just-in-time compiler that currently supports only
> x86 that builds upon the Stack VM to achieve substantial performance
> improvements.  The release is in the form of a Monticello package containing
> the VMMaker source and a tarball containing the platform sources, the
> generated sources and a Squeak 4.1 image containing the VMMaker sources.
> Download both at
>
> http://ftp.squeak.org/Cog/VMMaker-oscog.11.mcz
>
> http://ftp.squeak.org/Cog/OpenSourceCog.tar.gz
>
> Cog VMs:
>
> The Cog VMs are Squeak/Croquet VMs that run closure
> Squeak/Croquet/Pharo/Cuis images. The VMs support existing plugin source but
> will require plugins to be recompiled as the VM_PROXY_MAJOR plugin api has
> been extended.
>
> This release contains two distinct VMs, the StackInterpreter and the Cogit.
> The StackInterpreter is a fully-portable plug-in replacement for the current
> closure Squeak VMs and images.  The Stack VM uses context-to-stack mapping
> and a somewhat improved garbage collector to achieve modest but useful
> performance gains in the 10% to 15% range.  The StackInterpreter is intended
> to supersede the Squeak VM on platforms where the Cogit cannot be used.  The
> Cogit extends the StackInterpreter with a just-in-time compiler that uses
> aggressive inline caching techniques to deliver substantial performance
> gains in the 3x to 15x range, depending on benchmark.  The Cogit currently
> supports only x86 and the floating-point primitives and parts of the
> platform support code depend on SSE2.  I hope members of the community will
> attempt to port it, e.g. to ARM, PowerPC and x86-64.  The Cogit (excuse the
> pun) is so named because it is both an interpreter and a JIT, choosing not
> to generate machine code for large methods, interpreting them instead, the
> default policy being not to JIT methods with more than 60 literals.
>
> The Cog VM requires a few minor image changes all in
> image/NecessaryImageChangesForCogToWork.1.cs.  The JIT's machine-code
> SmallInteger primitives insist on a SmallInteger receiver so the primitives
> in LargePositiveInteger = ~= bitAnd: bitOr: butShift: and bitXor: cannot be
> used and these methods must be deleted.  The Cogit inlines the address of
> the Character instance table, Smalltalk specialObjectsArray at: 25, into the
> machine-code at: primitive for faster ByteString>>at: and so the table
> cannot be rebuilt in SmalltalkImage>>recreateSpecialObjectsArray.  The new
> version preserves the existing table.  Both VMs maintain floats in platform
> order to ease implementation of machine code floating-point primitives, and
> hence internally are in little-endian order instead of big-endian in current
> Squeak images.  While the VMs convert float order automatically on load they
> do require special accessing primitives Float>>basicAt: &
> Float>>basicAt:put: that undo the reversal and answer Float contents in
> big-endian order so that e.g. Float>>hash is unchanged.  The methods assume
> these primitives can fail, allowing the code to be used on current Squeak
> VMs.
>
> The image/VMMaker-Squeak4.1.image is a Squeak 4.1 image, runnable with the
> current Squeak VMs, that contains these changes, and can hence also be run
> with a Cog VM.  But beware, once an image has been saved on Cog it cannot be
> run by an existing Squeak VM, because existing VMs cannot undo the Float
> order change.
>
> Platform Subsystem:
>
> Most of the platform subsystem is unchanged but there are some important
> changes that need description.  The biggest change is the heartbeat and the
> clock in platforms/unix/vm/sqUnixHeartbeat.c and
> platforms/win32/vm/sqWin32Heartbeat.c.  The Cog VMs avoid the slow and
> variable interruptCheckCounter, folding the event check into the stack
> overflow check on frame build.  The heartbeat, typically 500Hz or 1KHz,
> changes the stackLimit to a value that will always fail.  On the next frame
> building send the VM will enter stack overflow handling that, as a side
> effect, will also check for events.  This is more efficient than the update
> of interruptCheckCounter and much more regular.  If one is running code that
> executes long-running primitives (e.g. large integer arithmetic) the counter
> approach will result in too low an interrupt check frequency, and conversely
> if one is running normal code the interrupt check frequency can be very
> high.
>
> The heartbeat also maintains a 64-bit microsecond clock, UTC microseconds
> from 1901, from which the backward-compatible millisecond and second clocks
> are derived.  Primitives exist to answer UTC microseconds and local
> microseconds.  Updating the clock in the heartbeat results in a 1 or 2
> millisecond resolution but avoids the cost of accessing the OS time on every
> prim tie which we've found important for performance at Teleplace.  The
> 64-bit microsecond clocks provide a unified time basis and eliminate
> wrapping (for the next 54,000 years at least).  I hope community images will
> move to these clocks.  It's worked well in VisualWorks.
>
> Another significant change is in the external semaphore table support code.
> This is now lock-free at the cost of having to specify a maximum number of
> external semaphores at start-up (default 256).  The support code for the
> lock-free data structures are processor-specific and is currently
> implemented only for x86 and gcc-compatible compilers; see
> platforms/Cross/vm/{sqAtomicOps.h,sqMemoryFence.h}.
>
> There is also improved crash reporting code that prints a primitive log and
> a C backtrace in addition to the Smalltalk backtrace.  See platforms/Mac
> OS/vm/sqMacMain.c, platforms/unix/vm/sqUnixMain.c,
> platforms/win32/vm/sqWin32Intel.c & platforms/win32/vm/sqWin32Backtrace.c.
>
> Finally there is support for the QVMProfiler, a pc-sampling profiler for
> profiling at the VM level.  See platforms/unix/vm/sqUnixVMProfile.c and
> platforms/win32/vm/sqWin32VMProfile.c.  The profiler itself is in the
> VMMaker image described below in Qwaq-VMProfiling.
>
> There are also changes to do with Teleplace-specific extensions to the
> HostWindowPlugin but these are not essential to Cog.
>
> VMMaker and Slang:
>
> The image/VMMaker-Squeak4.1.image Squeak 4.1 image contains the complete Cog
> VMMaker with necessary support code for simulation. This image was used to
> generate the sources in the src and stacksrc directories.
>
> Cog's VMMaker is substantially revised and extended from the current
> VMMaker.  It supports multiple classes, not just Interpreter and
> superclasses, because both context-to-stack mapping and the Cogit are too
> complex to write monolithically.  Classes can specify ancilliaryClasses and
> ancilliaryStructClasses, such as CoInterpreterStackPage, CogMethod and
> CogAbstractInstruction.  The Monticello package version is included in the
> header of all generated files and constitutes the version stamp for
> generated code.  Code is generated in sorted order so that minor changes in
> the Smalltalk source produce correspondingly minor changes in the generated
> code.  The gnuification step is built-in to VMMaker.  No effort has been
> made to maintain 64-bit compatibility.  Apologies, this was unaffordable.
>
> The VMMaker generates a single source tree used by all platforms.  Instead
> of deciding at generation time whether to use the Interpreter struct the
> generated code depends on the SQ_USE_GLOBAL_STRUCT define which can be
> overridden in platform makefiles.  All plugins live in src/plugins and
> platform makefiles along with plugins.int and plugins.ext files in the build
> subdirectories decide which plugins are built as external or internal.  The
> VM Generation Workspace from Workspace.text workspace contains dots to
> generate the sources.  We no longer use the VMMakerTool since there should
> be nothing platform-specific in the generated sources (if we add ports to
> other ISAs all their source can be included and selected as required by the
> platform makefiles).
>
> Since the Cogit generates x86 machine code simulation is much more complex.
> There is a support plugin, platforms/Cross/plugins/BochsIA32Plugin that
> depends on a large simulation of the x86 family implemented in C++ (see
> processors/IA32/bochs) and on Alien.  I use the simulator frequently (but
> note that I haven't had time to build a working version for Squeak 4.1).  I
> have tested Cog simulation in this image, running on the
> image/VMMaker-Squeak4.1.image itself.  The VM Simulation Workspace in the
> VMMaker image contains an example doit that starts the simulator. Be
> patient, even on a fast machine unhibernating the Squeak display background
> image takes nearly a minute.  Native fonts do not (yet) simulate correctly,
> but the system runs.  But note that I have only attempted to build and run
> the simulator on Mac OS X.  I expect Bochs can be built on linux and win32
> but I have not tried.  By the way, I've not described how to run the Bochs
> simulator on the current Squeak VM.  That's because the plugin depends on
> the heartbeat to break out of simulation occasionally via a new
> interpreterProxy entry point setInterruptCheckChain.  As this isn't
> supported by the current Squeak VMs the plugin would require modification.
> So to simulate first build either of the Cog VMs and then run the simulation
> with it.
>
> There are a number of unpublished changes to the base other than those in
> NecessaryImageChangesForCogToWork.1.cs.  This is partly laziness on my part,
> partly avoiding publishing things in advance of Cog.  These changes are
> better motivated once Cog is in use.  There are changes to the "translated
> primitives" (see implementors of translatedPrimitives) which replace
> messages with method tags for generation directives.  The Cog VMMaker uses
> Object>>perform:with:with:with:with: &
> Object>>perform:with:with:with:with:with: during simulation, and
> Collection>>#fold: & SquenceableCollection>>#copyUpThrough: during
> generation.  Object>>inline: and Object var:declareC:, which are mispackaged
> in Kernel in Squeak 4.1 are obsolete (method tags being used instead) and
> have been removed. I have changed Integer>>hex and Integer>>hex8 back to
> their original semantics as of 3.8.  Backward compatibility is important and
> one can easily add new selectors if one wants different functionality.
> VMMaker was here first ;)
>
> Tarball:
>
> The top-level directories in the tarball are
>
> src
>
> the tree for the Cog generated sources including all plugins
>
> stacksrc/vm
>
> the directory containing the Stack VM source (plugins can be taken from
> above)
>
> platforms
>
> the usual svn platform tree but including Cog specific changes such as the
> heartbeat
>
> processors
>
> the tree containing simulation support code, i.e. the bochs C++ x86
> simulation library, along with a potential ARM, PowerPC & MIPS simulator,
> Skeye.
>
> image
>
> the Cog-prepared Squeak 4.1 VMMaker image
>
> scripts
>
> some svn scripts to revert unchanged plugins that haven't really changed
>
> cygwinbuild
>
> the win32 build directory
>
> winbuild
>
> the old win32 build directory for minnow gcc 2.95.  Not entirely obsolete as
> the cygwin build as yet fails to generate a functional FFIPlugin
>
> macbuild
>
> the CoreVM.xcodeproj and support build projects for Mac OS X 10.5 or better
>
> unixbuild
>
> the build directory for linux
>
> Building Cog:
>
> Each build directory above contains a HowToBuild file that describes
> building in more detail.  The build directories only contain Cogit
> makefiles.  f you want to build a Stack VM you're on your own but this is
> very close to the existing Squeak VM build.
>
> Status:
>
> The Cogit VM has been our sole VM at Teleplace for nearly a year.  We do
> occasionally find bugs and there are almost certainly areas of functionality
> that we have not touched (for example I know that co-routining does not yet
> work).  If you find a bug please try and create a reproducible test case and
> let me know.  I can't promise to take a look or fix it but I am motivated to
> do so and will try my best as time allows.  Better still if you find and fix
> bugs be sure to let me know.
>
> License (MIT):
>
> All contributions from Teleplace in this release are
>
> Copyright (c) 2010 Teleplace, Inc.
>
> Permission is hereby granted, free of charge, to any person obtaining a copy
>
> of this software and associated documentation files (the "Software"), to
> deal
>
> in the Software without restriction, including without limitation the rights
>
> to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
>
> copies of the Software, and to permit persons to whom the Software is
>
> furnished to do so, subject to the following conditions:
>
> The above copyright notice and this permission notice shall be included in
>
> all copies or substantial portions of the Software.
>
> THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
>
> IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
>
> FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
>
> AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
>
> LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
> FROM,
>
> OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
>
> THE SOFTWARE.
>
> Eliot Miranda
>
> June 2010
>
>
>
>



--
=================================================
Germán S. Arduino  <gsa @ arsol.net>   Twitter: garduino
Arduino Software & Web Hosting   http://www.arduinosoftware.com
PasswordsPro  http://www.passwordspro.com
=================================================

Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: Teleplace Cog VMs are now available

stephane ducasse-4
In reply to this post by Eliot Miranda-2
This is really an excellent news. I open a lot of new futures!
Thanks teleplace, andreas and all the people that helped.

Stef
PS: you already answer my questions about platform were the jit cannot work so this is great.


On Jun 20, 2010, at 10:11 PM, Eliot Miranda wrote:

> Hi All,
>
> it gives me great pleasure to announce that the Teleplace Cog VMs are now available.  Huge thanks to all at Teleplace who have given me the opportunity to build Cog and release it as open source, been willing guinea pigs braving its bugs, and providing indispensable participation in getting Cog to its current state.  Huge thanks are also due to the original Back To The Future team whose VMMaker Cog extends to write the VM, and to Peter Deutsch from whom I've taken many ideas.
>
> This release contains two VMs.  The Stack VM, is a cross-platform interpreter that uses context-to-stack mapping to achieve modest performance gains.  The Cog VM is a just-in-time compiler that currently supports only x86 that builds upon the Stack VM to achieve substantial performance improvements.  The release is in the form of a Monticello package containing the VMMaker source and a tarball containing the platform sources, the generated sources and a Squeak 4.1 image containing the VMMaker sources.  Download both at
> http://ftp.squeak.org/Cog/VMMaker-oscog.11.mcz
> http://ftp.squeak.org/Cog/OpenSourceCog.tar.gz
>
> Cog VMs:
>
> The Cog VMs are Squeak/Croquet VMs that run closure Squeak/Croquet/Pharo/Cuis images. The VMs support existing plugin source but will require plugins to be recompiled as the VM_PROXY_MAJOR plugin api has been extended.
>
> This release contains two distinct VMs, the StackInterpreter and the Cogit.  The StackInterpreter is a fully-portable plug-in replacement for the current closure Squeak VMs and images.  The Stack VM uses context-to-stack mapping and a somewhat improved garbage collector to achieve modest but useful performance gains in the 10% to 15% range.  The StackInterpreter is intended to supersede the Squeak VM on platforms where the Cogit cannot be used.  The Cogit extends the StackInterpreter with a just-in-time compiler that uses aggressive inline caching techniques to deliver substantial performance gains in the 3x to 15x range, depending on benchmark.  The Cogit currently supports only x86 and the floating-point primitives and parts of the platform support code depend on SSE2.  I hope members of the community will attempt to port it, e.g. to ARM, PowerPC and x86-64.  The Cogit (excuse the pun) is so named because it is both an interpreter and a JIT, choosing not to generate machine code for large methods, interpreting them instead, the default policy being not to JIT methods with more than 60 literals.
>
> The Cog VM requires a few minor image changes all in image/NecessaryImageChangesForCogToWork.1.cs.  The JIT's machine-code SmallInteger primitives insist on a SmallInteger receiver so the primitives in LargePositiveInteger = ~= bitAnd: bitOr: butShift: and bitXor: cannot be used and these methods must be deleted.  The Cogit inlines the address of the Character instance table, Smalltalk specialObjectsArray at: 25, into the machine-code at: primitive for faster ByteString>>at: and so the table cannot be rebuilt in SmalltalkImage>>recreateSpecialObjectsArray.  The new version preserves the existing table.  Both VMs maintain floats in platform order to ease implementation of machine code floating-point primitives, and hence internally are in little-endian order instead of big-endian in current Squeak images.  While the VMs convert float order automatically on load they do require special accessing primitives Float>>basicAt: & Float>>basicAt:put: that undo the reversal and answer Float contents in big-endian order so that e.g. Float>>hash is unchanged.  The methods assume these primitives can fail, allowing the code to be used on current Squeak VMs.
>
> The image/VMMaker-Squeak4.1.image is a Squeak 4.1 image, runnable with the current Squeak VMs, that contains these changes, and can hence also be run with a Cog VM.  But beware, once an image has been saved on Cog it cannot be run by an existing Squeak VM, because existing VMs cannot undo the Float order change.
>
>
> Platform Subsystem:
>
> Most of the platform subsystem is unchanged but there are some important changes that need description.  The biggest change is the heartbeat and the clock in platforms/unix/vm/sqUnixHeartbeat.c and platforms/win32/vm/sqWin32Heartbeat.c.  The Cog VMs avoid the slow and variable interruptCheckCounter, folding the event check into the stack overflow check on frame build.  The heartbeat, typically 500Hz or 1KHz, changes the stackLimit to a value that will always fail.  On the next frame building send the VM will enter stack overflow handling that, as a side effect, will also check for events.  This is more efficient than the update of interruptCheckCounter and much more regular.  If one is running code that executes long-running primitives (e.g. large integer arithmetic) the counter approach will result in too low an interrupt check frequency, and conversely if one is running normal code the interrupt check frequency can be very high.
>
> The heartbeat also maintains a 64-bit microsecond clock, UTC microseconds from 1901, from which the backward-compatible millisecond and second clocks are derived.  Primitives exist to answer UTC microseconds and local microseconds.  Updating the clock in the heartbeat results in a 1 or 2 millisecond resolution but avoids the cost of accessing the OS time on every prim tie which we've found important for performance at Teleplace.  The 64-bit microsecond clocks provide a unified time basis and eliminate wrapping (for the next 54,000 years at least).  I hope community images will move to these clocks.  It's worked well in VisualWorks.
>
> Another significant change is in the external semaphore table support code.  This is now lock-free at the cost of having to specify a maximum number of external semaphores at start-up (default 256).  The support code for the lock-free data structures are processor-specific and is currently implemented only for x86 and gcc-compatible compilers; see platforms/Cross/vm/{sqAtomicOps.h,sqMemoryFence.h}.
>
> There is also improved crash reporting code that prints a primitive log and a C backtrace in addition to the Smalltalk backtrace.  See platforms/Mac OS/vm/sqMacMain.c, platforms/unix/vm/sqUnixMain.c, platforms/win32/vm/sqWin32Intel.c & platforms/win32/vm/sqWin32Backtrace.c.
>
> Finally there is support for the QVMProfiler, a pc-sampling profiler for profiling at the VM level.  See platforms/unix/vm/sqUnixVMProfile.c and platforms/win32/vm/sqWin32VMProfile.c.  The profiler itself is in the VMMaker image described below in Qwaq-VMProfiling.
>
> There are also changes to do with Teleplace-specific extensions to the HostWindowPlugin but these are not essential to Cog.
>
>
> VMMaker and Slang:
>
> The image/VMMaker-Squeak4.1.image Squeak 4.1 image contains the complete Cog VMMaker with necessary support code for simulation. This image was used to generate the sources in the src and stacksrc directories.
>
> Cog's VMMaker is substantially revised and extended from the current VMMaker.  It supports multiple classes, not just Interpreter and superclasses, because both context-to-stack mapping and the Cogit are too complex to write monolithically.  Classes can specify ancilliaryClasses and ancilliaryStructClasses, such as CoInterpreterStackPage, CogMethod and CogAbstractInstruction.  The Monticello package version is included in the header of all generated files and constitutes the version stamp for generated code.  Code is generated in sorted order so that minor changes in the Smalltalk source produce correspondingly minor changes in the generated code.  The gnuification step is built-in to VMMaker.  No effort has been made to maintain 64-bit compatibility.  Apologies, this was unaffordable.
>
> The VMMaker generates a single source tree used by all platforms.  Instead of deciding at generation time whether to use the Interpreter struct the generated code depends on the SQ_USE_GLOBAL_STRUCT define which can be overridden in platform makefiles.  All plugins live in src/plugins and platform makefiles along with plugins.int and plugins.ext files in the build subdirectories decide which plugins are built as external or internal.  The VM Generation Workspace from Workspace.text workspace contains dots to generate the sources.  We no longer use the VMMakerTool since there should be nothing platform-specific in the generated sources (if we add ports to other ISAs all their source can be included and selected as required by the platform makefiles).
>
> Since the Cogit generates x86 machine code simulation is much more complex.  There is a support plugin, platforms/Cross/plugins/BochsIA32Plugin that depends on a large simulation of the x86 family implemented in C++ (see  processors/IA32/bochs) and on Alien.  I use the simulator frequently (but note that I haven't had time to build a working version for Squeak 4.1).  I have tested Cog simulation in this image, running on the image/VMMaker-Squeak4.1.image itself.  The VM Simulation Workspace in the VMMaker image contains an example doit that starts the simulator. Be patient, even on a fast machine unhibernating the Squeak display background image takes nearly a minute.  Native fonts do not (yet) simulate correctly, but the system runs.  But note that I have only attempted to build and run the simulator on Mac OS X.  I expect Bochs can be built on linux and win32 but I have not tried.  By the way, I've not described how to run the Bochs simulator on the current Squeak VM.  That's because the plugin depends on the heartbeat to break out of simulation occasionally via a new interpreterProxy entry point setInterruptCheckChain.  As this isn't supported by the current Squeak VMs the plugin would require modification.  So to simulate first build either of the Cog VMs and then run the simulation with it.
>
> There are a number of unpublished changes to the base other than those in NecessaryImageChangesForCogToWork.1.cs.  This is partly laziness on my part, partly avoiding publishing things in advance of Cog.  These changes are better motivated once Cog is in use.  There are changes to the "translated primitives" (see implementors of translatedPrimitives) which replace messages with method tags for generation directives.  The Cog VMMaker uses Object>>perform:with:with:with:with: & Object>>perform:with:with:with:with:with: during simulation, and Collection>>#fold: & SquenceableCollection>>#copyUpThrough: during generation.  Object>>inline: and Object var:declareC:, which are mispackaged in Kernel in Squeak 4.1 are obsolete (method tags being used instead) and have been removed. I have changed Integer>>hex and Integer>>hex8 back to their original semantics as of 3.8.  Backward compatibility is important and one can easily add new selectors if one wants different functionality.  VMMaker was here first ;)
>
>
> Tarball:
>
> The top-level directories in the tarball are
>
> src
> the tree for the Cog generated sources including all plugins
> stacksrc/vm
> the directory containing the Stack VM source (plugins can be taken from above)
> platforms
> the usual svn platform tree but including Cog specific changes such as the heartbeat
> processors
> the tree containing simulation support code, i.e. the bochs C++ x86 simulation library, along with a potential ARM, PowerPC & MIPS simulator, Skeye.
>
> image
> the Cog-prepared Squeak 4.1 VMMaker image
> scripts
> some svn scripts to revert unchanged plugins that haven't really changed
>
> cygwinbuild
> the win32 build directory
> winbuild
> the old win32 build directory for minnow gcc 2.95.  Not entirely obsolete as the cygwin build as yet fails to generate a functional FFIPlugin
> macbuild
> the CoreVM.xcodeproj and support build projects for Mac OS X 10.5 or better
> unixbuild
> the build directory for linux
>
>
> Building Cog:
>
> Each build directory above contains a HowToBuild file that describes building in more detail.  The build directories only contain Cogit makefiles.  f you want to build a Stack VM you're on your own but this is very close to the existing Squeak VM build.
>
>
> Status:
> The Cogit VM has been our sole VM at Teleplace for nearly a year.  We do occasionally find bugs and there are almost certainly areas of functionality that we have not touched (for example I know that co-routining does not yet work).  If you find a bug please try and create a reproducible test case and let me know.  I can't promise to take a look or fix it but I am motivated to do so and will try my best as time allows.  Better still if you find and fix bugs be sure to let me know.
>
> License (MIT):
> All contributions from Teleplace in this release are
> Copyright (c) 2010 Teleplace, Inc.
>
> Permission is hereby granted, free of charge, to any person obtaining a copy
> of this software and associated documentation files (the "Software"), to deal
> in the Software without restriction, including without limitation the rights
> to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
> copies of the Software, and to permit persons to whom the Software is
> furnished to do so, subject to the following conditions:
>
> The above copyright notice and this permission notice shall be included in
> all copies or substantial portions of the Software.
>
> THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
> IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
> FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
> AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
> LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
> OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
> THE SOFTWARE.
>
> Eliot Miranda
> June 2010
>
>


Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: Teleplace Cog VMs are now available

Josh Gargus
In reply to this post by Eliot Miranda-2
Not much to say other than: Hooray Eliot!  

And thanks to everyone who played a part in convincing the decision-makers at Teleplace that releasing Cog is a win for everybody.

Cheers,
Josh


On Jun 20, 2010, at 1:11 PM, Eliot Miranda wrote:

Hi All,

it gives me great pleasure to announce that the Teleplace Cog VMs are now available.  Huge thanks to all at Teleplace who have given me the opportunity to build Cog and release it as open source, been willing guinea pigs braving its bugs, and providing indispensable participation in getting Cog to its current state.  Huge thanks are also due to the original Back To The Future team whose VMMaker Cog extends to write the VM, and to Peter Deutsch from whom I've taken many ideas.

This release contains two VMs.  The Stack VM, is a cross-platform interpreter that uses context-to-stack mapping to achieve modest performance gains.  The Cog VM is a just-in-time compiler that currently supports only x86 that builds upon the Stack VM to achieve substantial performance improvements.  The release is in the form of a Monticello package containing the VMMaker source and a tarball containing the platform sources, the generated sources and a Squeak 4.1 image containing the VMMaker sources.  Download both at

Cog VMs:

The Cog VMs are Squeak/Croquet VMs that run closure Squeak/Croquet/Pharo/Cuis images. The VMs support existing plugin source but will require plugins to be recompiled as the VM_PROXY_MAJOR plugin api has been extended.

This release contains two distinct VMs, the StackInterpreter and the Cogit.  The StackInterpreter is a fully-portable plug-in replacement for the current closure Squeak VMs and images.  The Stack VM uses context-to-stack mapping and a somewhat improved garbage collector to achieve modest but useful performance gains in the 10% to 15% range.  The StackInterpreter is intended to supersede the Squeak VM on platforms where the Cogit cannot be used.  The Cogit extends the StackInterpreter with a just-in-time compiler that uses aggressive inline caching techniques to deliver substantial performance gains in the 3x to 15x range, depending on benchmark.  The Cogit currently supports only x86 and the floating-point primitives and parts of the platform support code depend on SSE2.  I hope members of the community will attempt to port it, e.g. to ARM, PowerPC and x86-64.  The Cogit (excuse the pun) is so named because it is both an interpreter and a JIT, choosing not to generate machine code for large methods, interpreting them instead, the default policy being not to JIT methods with more than 60 literals.

The Cog VM requires a few minor image changes all in image/NecessaryImageChangesForCogToWork.1.cs.  The JIT's machine-code SmallInteger primitives insist on a SmallInteger receiver so the primitives in LargePositiveInteger = ~= bitAnd: bitOr: butShift: and bitXor: cannot be used and these methods must be deleted.  The Cogit inlines the address of the Character instance table, Smalltalk specialObjectsArray at: 25, into the machine-code at: primitive for faster ByteString>>at: and so the table cannot be rebuilt in SmalltalkImage>>recreateSpecialObjectsArray.  The new version preserves the existing table.  Both VMs maintain floats in platform order to ease implementation of machine code floating-point primitives, and hence internally are in little-endian order instead of big-endian in current Squeak images.  While the VMs convert float order automatically on load they do require special accessing primitives Float>>basicAt: & Float>>basicAt:put: that undo the reversal and answer Float contents in big-endian order so that e.g. Float>>hash is unchanged.  The methods assume these primitives can fail, allowing the code to be used on current Squeak VMs.

The image/VMMaker-Squeak4.1.image is a Squeak 4.1 image, runnable with the current Squeak VMs, that contains these changes, and can hence also be run with a Cog VM.  But beware, once an image has been saved on Cog it cannot be run by an existing Squeak VM, because existing VMs cannot undo the Float order change.


Platform Subsystem:

Most of the platform subsystem is unchanged but there are some important changes that need description.  The biggest change is the heartbeat and the clock in platforms/unix/vm/sqUnixHeartbeat.c and platforms/win32/vm/sqWin32Heartbeat.c.  The Cog VMs avoid the slow and variable interruptCheckCounter, folding the event check into the stack overflow check on frame build.  The heartbeat, typically 500Hz or 1KHz, changes the stackLimit to a value that will always fail.  On the next frame building send the VM will enter stack overflow handling that, as a side effect, will also check for events.  This is more efficient than the update of interruptCheckCounter and much more regular.  If one is running code that executes long-running primitives (e.g. large integer arithmetic) the counter approach will result in too low an interrupt check frequency, and conversely if one is running normal code the interrupt check frequency can be very high.

The heartbeat also maintains a 64-bit microsecond clock, UTC microseconds from 1901, from which the backward-compatible millisecond and second clocks are derived.  Primitives exist to answer UTC microseconds and local microseconds.  Updating the clock in the heartbeat results in a 1 or 2 millisecond resolution but avoids the cost of accessing the OS time on every prim tie which we've found important for performance at Teleplace.  The 64-bit microsecond clocks provide a unified time basis and eliminate wrapping (for the next 54,000 years at least).  I hope community images will move to these clocks.  It's worked well in VisualWorks.

Another significant change is in the external semaphore table support code.  This is now lock-free at the cost of having to specify a maximum number of external semaphores at start-up (default 256).  The support code for the lock-free data structures are processor-specific and is currently implemented only for x86 and gcc-compatible compilers; see platforms/Cross/vm/{sqAtomicOps.h,sqMemoryFence.h}.

There is also improved crash reporting code that prints a primitive log and a C backtrace in addition to the Smalltalk backtrace.  See platforms/Mac OS/vm/sqMacMain.c, platforms/unix/vm/sqUnixMain.c, platforms/win32/vm/sqWin32Intel.c & platforms/win32/vm/sqWin32Backtrace.c.

Finally there is support for the QVMProfiler, a pc-sampling profiler for profiling at the VM level.  See platforms/unix/vm/sqUnixVMProfile.c and platforms/win32/vm/sqWin32VMProfile.c.  The profiler itself is in the VMMaker image described below in Qwaq-VMProfiling.

There are also changes to do with Teleplace-specific extensions to the HostWindowPlugin but these are not essential to Cog.


VMMaker and Slang:

The image/VMMaker-Squeak4.1.image Squeak 4.1 image contains the complete Cog VMMaker with necessary support code for simulation. This image was used to generate the sources in the src and stacksrc directories.

Cog's VMMaker is substantially revised and extended from the current VMMaker.  It supports multiple classes, not just Interpreter and superclasses, because both context-to-stack mapping and the Cogit are too complex to write monolithically.  Classes can specify ancilliaryClasses and ancilliaryStructClasses, such as CoInterpreterStackPage, CogMethod and CogAbstractInstruction.  The Monticello package version is included in the header of all generated files and constitutes the version stamp for generated code.  Code is generated in sorted order so that minor changes in the Smalltalk source produce correspondingly minor changes in the generated code.  The gnuification step is built-in to VMMaker.  No effort has been made to maintain 64-bit compatibility.  Apologies, this was unaffordable.

The VMMaker generates a single source tree used by all platforms.  Instead of deciding at generation time whether to use the Interpreter struct the generated code depends on the SQ_USE_GLOBAL_STRUCT define which can be overridden in platform makefiles.  All plugins live in src/plugins and platform makefiles along with plugins.int and plugins.ext files in the build subdirectories decide which plugins are built as external or internal.  The VM Generation Workspace from Workspace.text workspace contains dots to generate the sources.  We no longer use the VMMakerTool since there should be nothing platform-specific in the generated sources (if we add ports to other ISAs all their source can be included and selected as required by the platform makefiles).

Since the Cogit generates x86 machine code simulation is much more complex.  There is a support plugin, platforms/Cross/plugins/BochsIA32Plugin that depends on a large simulation of the x86 family implemented in C++ (see  processors/IA32/bochs) and on Alien.  I use the simulator frequently (but note that I haven't had time to build a working version for Squeak 4.1).  I have tested Cog simulation in this image, running on the image/VMMaker-Squeak4.1.image itself.  The VM Simulation Workspace in the VMMaker image contains an example doit that starts the simulator. Be patient, even on a fast machine unhibernating the Squeak display background image takes nearly a minute.  Native fonts do not (yet) simulate correctly, but the system runs.  But note that I have only attempted to build and run the simulator on Mac OS X.  I expect Bochs can be built on linux and win32 but I have not tried.  By the way, I've not described how to run the Bochs simulator on the current Squeak VM.  That's because the plugin depends on the heartbeat to break out of simulation occasionally via a new interpreterProxy entry point setInterruptCheckChain.  As this isn't supported by the current Squeak VMs the plugin would require modification.  So to simulate first build either of the Cog VMs and then run the simulation with it.

There are a number of unpublished changes to the base other than those in NecessaryImageChangesForCogToWork.1.cs.  This is partly laziness on my part, partly avoiding publishing things in advance of Cog.  These changes are better motivated once Cog is in use.  There are changes to the "translated primitives" (see implementors of translatedPrimitives) which replace messages with method tags for generation directives.  The Cog VMMaker uses Object>>perform:with:with:with:with: & Object>>perform:with:with:with:with:with: during simulation, and Collection>>#fold: & SquenceableCollection>>#copyUpThrough: during generation.  Object>>inline: and Object var:declareC:, which are mispackaged in Kernel in Squeak 4.1 are obsolete (method tags being used instead) and have been removed. I have changed Integer>>hex and Integer>>hex8 back to their original semantics as of 3.8.  Backward compatibility is important and one can easily add new selectors if one wants different functionality.  VMMaker was here first ;)


Tarball:

The top-level directories in the tarball are

src
the tree for the Cog generated sources including all plugins
stacksrc/vm
the directory containing the Stack VM source (plugins can be taken from above)
platforms
the usual svn platform tree but including Cog specific changes such as the heartbeat
processors
the tree containing simulation support code, i.e. the bochs C++ x86 simulation library, along with a potential ARM, PowerPC & MIPS simulator, Skeye.

image
the Cog-prepared Squeak 4.1 VMMaker image
scripts
some svn scripts to revert unchanged plugins that haven't really changed

cygwinbuild
the win32 build directory
winbuild
the old win32 build directory for minnow gcc 2.95.  Not entirely obsolete as the cygwin build as yet fails to generate a functional FFIPlugin
macbuild
the CoreVM.xcodeproj and support build projects for Mac OS X 10.5 or better
unixbuild
the build directory for linux


Building Cog:

Each build directory above contains a HowToBuild file that describes building in more detail.  The build directories only contain Cogit makefiles.  f you want to build a Stack VM you're on your own but this is very close to the existing Squeak VM build.


Status:
The Cogit VM has been our sole VM at Teleplace for nearly a year.  We do occasionally find bugs and there are almost certainly areas of functionality that we have not touched (for example I know that co-routining does not yet work).  If you find a bug please try and create a reproducible test case and let me know.  I can't promise to take a look or fix it but I am motivated to do so and will try my best as time allows.  Better still if you find and fix bugs be sure to let me know.

License (MIT):
All contributions from Teleplace in this release are
Copyright (c) 2010 Teleplace, Inc.

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in
all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
THE SOFTWARE.

Eliot Miranda
June 2010





Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Squeak 4.1 on Cog (Re: Teleplace Cog VMs are now available)

Andreas.Raab
In reply to this post by Eliot Miranda-2
Folks -

I've prepared a version of Squeak 4.1 for those of you who are not into
recompiling VMs yourself. You can download it from:

        http://squeakvm.org/win32/experimental/CogSqueak4.1.zip

The package contains a freshly compiled Cog VM for Windows and an
updated Squeak4.1 image. Some things to note:

* Once you save the image with Cog, you can only run it with Cog. So
load your benchmarks into the image *before* you launch it with Cog
(i.e., to ensure you're comparing apples with apples you should use the
same image)

* There are some changes that are required in the image. Most
importantly, some LargeInt primitives need to be removed:
        #(= bitAnd: bitOr: bitShift: bitXor: ~=) do:[:sel|
                LargePositiveInteger removeSelector: sel.
        ].

This will be necessary for any images you want to run with Cog. There
are also some other changes but they are not critical (i.e., don't crash
the VM).

Enjoy,
   - Andreas

Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: Teleplace Cog VMs are now available

Lukas Renggli
In reply to this post by stephane ducasse-4
Thank you Eliot.

This is wonderful, it feels like I just got a new laptop :-)

I noticed some problems in the continuation tests of Seaside
(WAContinuationTest, WAFlowPlatformTest and
WAPartialContinuationTest):

   "Computation has been terminated" in MethodContext>>cannotReturn:

Is this a known problem? Or is this maybe specific to Pharo on Cog?

Cheers,
Lukas

On 20 June 2010 22:53, stephane ducasse <[hidden email]> wrote:

> This is really an excellent news. I open a lot of new futures!
> Thanks teleplace, andreas and all the people that helped.
>
> Stef
> PS: you already answer my questions about platform were the jit cannot work so this is great.
>
>
> On Jun 20, 2010, at 10:11 PM, Eliot Miranda wrote:
>
>> Hi All,
>>
>>       it gives me great pleasure to announce that the Teleplace Cog VMs are now available.  Huge thanks to all at Teleplace who have given me the opportunity to build Cog and release it as open source, been willing guinea pigs braving its bugs, and providing indispensable participation in getting Cog to its current state.  Huge thanks are also due to the original Back To The Future team whose VMMaker Cog extends to write the VM, and to Peter Deutsch from whom I've taken many ideas.
>>
>> This release contains two VMs.  The Stack VM, is a cross-platform interpreter that uses context-to-stack mapping to achieve modest performance gains.  The Cog VM is a just-in-time compiler that currently supports only x86 that builds upon the Stack VM to achieve substantial performance improvements.  The release is in the form of a Monticello package containing the VMMaker source and a tarball containing the platform sources, the generated sources and a Squeak 4.1 image containing the VMMaker sources.  Download both at
>>       http://ftp.squeak.org/Cog/VMMaker-oscog.11.mcz
>>       http://ftp.squeak.org/Cog/OpenSourceCog.tar.gz
>>
>> Cog VMs:
>>
>> The Cog VMs are Squeak/Croquet VMs that run closure Squeak/Croquet/Pharo/Cuis images. The VMs support existing plugin source but will require plugins to be recompiled as the VM_PROXY_MAJOR plugin api has been extended.
>>
>> This release contains two distinct VMs, the StackInterpreter and the Cogit.  The StackInterpreter is a fully-portable plug-in replacement for the current closure Squeak VMs and images.  The Stack VM uses context-to-stack mapping and a somewhat improved garbage collector to achieve modest but useful performance gains in the 10% to 15% range.  The StackInterpreter is intended to supersede the Squeak VM on platforms where the Cogit cannot be used.  The Cogit extends the StackInterpreter with a just-in-time compiler that uses aggressive inline caching techniques to deliver substantial performance gains in the 3x to 15x range, depending on benchmark.  The Cogit currently supports only x86 and the floating-point primitives and parts of the platform support code depend on SSE2.  I hope members of the community will attempt to port it, e.g. to ARM, PowerPC and x86-64.  The Cogit (excuse the pun) is so named because it is both an interpreter and a JIT, choosing not to generate machine code for large methods, interpreting them instead, the default policy being not to JIT methods with more than 60 literals.
>>
>> The Cog VM requires a few minor image changes all in image/NecessaryImageChangesForCogToWork.1.cs.  The JIT's machine-code SmallInteger primitives insist on a SmallInteger receiver so the primitives in LargePositiveInteger = ~= bitAnd: bitOr: butShift: and bitXor: cannot be used and these methods must be deleted.  The Cogit inlines the address of the Character instance table, Smalltalk specialObjectsArray at: 25, into the machine-code at: primitive for faster ByteString>>at: and so the table cannot be rebuilt in SmalltalkImage>>recreateSpecialObjectsArray.  The new version preserves the existing table.  Both VMs maintain floats in platform order to ease implementation of machine code floating-point primitives, and hence internally are in little-endian order instead of big-endian in current Squeak images.  While the VMs convert float order automatically on load they do require special accessing primitives Float>>basicAt: & Float>>basicAt:put: that undo the reversal and answer Float contents in big-endian order so that e.g. Float>>hash is unchanged.  The methods assume these primitives can fail, allowing the code to be used on current Squeak VMs.
>>
>> The image/VMMaker-Squeak4.1.image is a Squeak 4.1 image, runnable with the current Squeak VMs, that contains these changes, and can hence also be run with a Cog VM.  But beware, once an image has been saved on Cog it cannot be run by an existing Squeak VM, because existing VMs cannot undo the Float order change.
>>
>>
>> Platform Subsystem:
>>
>> Most of the platform subsystem is unchanged but there are some important changes that need description.  The biggest change is the heartbeat and the clock in platforms/unix/vm/sqUnixHeartbeat.c and platforms/win32/vm/sqWin32Heartbeat.c.  The Cog VMs avoid the slow and variable interruptCheckCounter, folding the event check into the stack overflow check on frame build.  The heartbeat, typically 500Hz or 1KHz, changes the stackLimit to a value that will always fail.  On the next frame building send the VM will enter stack overflow handling that, as a side effect, will also check for events.  This is more efficient than the update of interruptCheckCounter and much more regular.  If one is running code that executes long-running primitives (e.g. large integer arithmetic) the counter approach will result in too low an interrupt check frequency, and conversely if one is running normal code the interrupt check frequency can be very high.
>>
>> The heartbeat also maintains a 64-bit microsecond clock, UTC microseconds from 1901, from which the backward-compatible millisecond and second clocks are derived.  Primitives exist to answer UTC microseconds and local microseconds.  Updating the clock in the heartbeat results in a 1 or 2 millisecond resolution but avoids the cost of accessing the OS time on every prim tie which we've found important for performance at Teleplace.  The 64-bit microsecond clocks provide a unified time basis and eliminate wrapping (for the next 54,000 years at least).  I hope community images will move to these clocks.  It's worked well in VisualWorks.
>>
>> Another significant change is in the external semaphore table support code.  This is now lock-free at the cost of having to specify a maximum number of external semaphores at start-up (default 256).  The support code for the lock-free data structures are processor-specific and is currently implemented only for x86 and gcc-compatible compilers; see platforms/Cross/vm/{sqAtomicOps.h,sqMemoryFence.h}.
>>
>> There is also improved crash reporting code that prints a primitive log and a C backtrace in addition to the Smalltalk backtrace.  See platforms/Mac OS/vm/sqMacMain.c, platforms/unix/vm/sqUnixMain.c, platforms/win32/vm/sqWin32Intel.c & platforms/win32/vm/sqWin32Backtrace.c.
>>
>> Finally there is support for the QVMProfiler, a pc-sampling profiler for profiling at the VM level.  See platforms/unix/vm/sqUnixVMProfile.c and platforms/win32/vm/sqWin32VMProfile.c.  The profiler itself is in the VMMaker image described below in Qwaq-VMProfiling.
>>
>> There are also changes to do with Teleplace-specific extensions to the HostWindowPlugin but these are not essential to Cog.
>>
>>
>> VMMaker and Slang:
>>
>> The image/VMMaker-Squeak4.1.image Squeak 4.1 image contains the complete Cog VMMaker with necessary support code for simulation. This image was used to generate the sources in the src and stacksrc directories.
>>
>> Cog's VMMaker is substantially revised and extended from the current VMMaker.  It supports multiple classes, not just Interpreter and superclasses, because both context-to-stack mapping and the Cogit are too complex to write monolithically.  Classes can specify ancilliaryClasses and ancilliaryStructClasses, such as CoInterpreterStackPage, CogMethod and CogAbstractInstruction.  The Monticello package version is included in the header of all generated files and constitutes the version stamp for generated code.  Code is generated in sorted order so that minor changes in the Smalltalk source produce correspondingly minor changes in the generated code.  The gnuification step is built-in to VMMaker.  No effort has been made to maintain 64-bit compatibility.  Apologies, this was unaffordable.
>>
>> The VMMaker generates a single source tree used by all platforms.  Instead of deciding at generation time whether to use the Interpreter struct the generated code depends on the SQ_USE_GLOBAL_STRUCT define which can be overridden in platform makefiles.  All plugins live in src/plugins and platform makefiles along with plugins.int and plugins.ext files in the build subdirectories decide which plugins are built as external or internal.  The VM Generation Workspace from Workspace.text workspace contains dots to generate the sources.  We no longer use the VMMakerTool since there should be nothing platform-specific in the generated sources (if we add ports to other ISAs all their source can be included and selected as required by the platform makefiles).
>>
>> Since the Cogit generates x86 machine code simulation is much more complex.  There is a support plugin, platforms/Cross/plugins/BochsIA32Plugin that depends on a large simulation of the x86 family implemented in C++ (see  processors/IA32/bochs) and on Alien.  I use the simulator frequently (but note that I haven't had time to build a working version for Squeak 4.1).  I have tested Cog simulation in this image, running on the image/VMMaker-Squeak4.1.image itself.  The VM Simulation Workspace in the VMMaker image contains an example doit that starts the simulator. Be patient, even on a fast machine unhibernating the Squeak display background image takes nearly a minute.  Native fonts do not (yet) simulate correctly, but the system runs.  But note that I have only attempted to build and run the simulator on Mac OS X.  I expect Bochs can be built on linux and win32 but I have not tried.  By the way, I've not described how to run the Bochs simulator on the current Squeak VM.  That's because the plugin depends on the heartbeat to break out of simulation occasionally via a new interpreterProxy entry point setInterruptCheckChain.  As this isn't supported by the current Squeak VMs the plugin would require modification.  So to simulate first build either of the Cog VMs and then run the simulation with it.
>>
>> There are a number of unpublished changes to the base other than those in NecessaryImageChangesForCogToWork.1.cs.  This is partly laziness on my part, partly avoiding publishing things in advance of Cog.  These changes are better motivated once Cog is in use.  There are changes to the "translated primitives" (see implementors of translatedPrimitives) which replace messages with method tags for generation directives.  The Cog VMMaker uses Object>>perform:with:with:with:with: & Object>>perform:with:with:with:with:with: during simulation, and Collection>>#fold: & SquenceableCollection>>#copyUpThrough: during generation.  Object>>inline: and Object var:declareC:, which are mispackaged in Kernel in Squeak 4.1 are obsolete (method tags being used instead) and have been removed. I have changed Integer>>hex and Integer>>hex8 back to their original semantics as of 3.8.  Backward compatibility is important and one can easily add new selectors if one wants different functionality.  VMMaker was here first ;)
>>
>>
>> Tarball:
>>
>> The top-level directories in the tarball are
>>
>>       src
>>               the tree for the Cog generated sources including all plugins
>>       stacksrc/vm
>>               the directory containing the Stack VM source (plugins can be taken from above)
>>       platforms
>>               the usual svn platform tree but including Cog specific changes such as the heartbeat
>>       processors
>>               the tree containing simulation support code, i.e. the bochs C++ x86 simulation library, along with a potential ARM, PowerPC & MIPS simulator, Skeye.
>>
>>       image
>>               the Cog-prepared Squeak 4.1 VMMaker image
>>       scripts
>>               some svn scripts to revert unchanged plugins that haven't really changed
>>
>>       cygwinbuild
>>               the win32 build directory
>>       winbuild
>>               the old win32 build directory for minnow gcc 2.95.  Not entirely obsolete as the cygwin build as yet fails to generate a functional FFIPlugin
>>       macbuild
>>               the CoreVM.xcodeproj and support build projects for Mac OS X 10.5 or better
>>       unixbuild
>>               the build directory for linux
>>
>>
>> Building Cog:
>>
>> Each build directory above contains a HowToBuild file that describes building in more detail.  The build directories only contain Cogit makefiles.  f you want to build a Stack VM you're on your own but this is very close to the existing Squeak VM build.
>>
>>
>> Status:
>> The Cogit VM has been our sole VM at Teleplace for nearly a year.  We do occasionally find bugs and there are almost certainly areas of functionality that we have not touched (for example I know that co-routining does not yet work).  If you find a bug please try and create a reproducible test case and let me know.  I can't promise to take a look or fix it but I am motivated to do so and will try my best as time allows.  Better still if you find and fix bugs be sure to let me know.
>>
>> License (MIT):
>> All contributions from Teleplace in this release are
>> Copyright (c) 2010 Teleplace, Inc.
>>
>> Permission is hereby granted, free of charge, to any person obtaining a copy
>> of this software and associated documentation files (the "Software"), to deal
>> in the Software without restriction, including without limitation the rights
>> to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
>> copies of the Software, and to permit persons to whom the Software is
>> furnished to do so, subject to the following conditions:
>>
>> The above copyright notice and this permission notice shall be included in
>> all copies or substantial portions of the Software.
>>
>> THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
>> IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
>> FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
>> AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
>> LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
>> OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
>> THE SOFTWARE.
>>
>> Eliot Miranda
>> June 2010
>>
>>
>
>
>



--
Lukas Renggli
www.lukas-renggli.ch

Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: Teleplace Cog VMs are now available

johnmci
In reply to this post by Eliot Miranda-2
Wonderful news.


On 2010-06-20, at 1:11 PM, Eliot Miranda wrote:

 The Cogit currently supports only x86 and the floating-point primitives and parts of the platform support code depend on SSE2.  I hope members of the community will attempt to port it, e.g. to ARM, PowerPC and x86-64. 

Now before anyone asks, it's very doubtful you will see Cogit on the iOS series of operating systems (iphone/ipad). I have had detailed discussions with the Apple Security team since their goal is to prevent user level code from doing dynamic code generation. 

mmap for example won't return a page of memory that is executable. 

Let's say the probability of getting Apple to change their minds on this is zero. 

However I'll look into the StackInterpreter for the Macintosh 5.x VM and the iOS flavour, mind if anyone wants to do that chore early then they should ping me. 


--
===========================================================================
John M. McIntosh <[hidden email]>   Twitter:  squeaker68882
Corporate Smalltalk Consulting Ltd.  http://www.smalltalkconsulting.com
===========================================================================







smime.p7s (3K) Download Attachment
Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: [Vm-dev] Re: [squeak-dev] Teleplace Cog VMs are now available

Rob Rothwell
For the binary-challenged, I am sure some kind soul will post compiled versions somewhere...!

(I ALWAYS struggle with build.bat, missing files, environment variables, and the like...)

Take care,

Rob

On Sun, Jun 20, 2010 at 5:37 PM, John M McIntosh <[hidden email]> wrote:
 
Wonderful news.


On 2010-06-20, at 1:11 PM, Eliot Miranda wrote:

 The Cogit currently supports only x86 and the floating-point primitives and parts of the platform support code depend on SSE2.  I hope members of the community will attempt to port it, e.g. to ARM, PowerPC and x86-64. 

Now before anyone asks, it's very doubtful you will see Cogit on the iOS series of operating systems (iphone/ipad). I have had detailed discussions with the Apple Security team since their goal is to prevent user level code from doing dynamic code generation. 

mmap for example won't return a page of memory that is executable. 

Let's say the probability of getting Apple to change their minds on this is zero. 

However I'll look into the StackInterpreter for the Macintosh 5.x VM and the iOS flavour, mind if anyone wants to do that chore early then they should ping me. 


--
===========================================================================
John M. McIntosh <[hidden email]>   Twitter:  squeaker68882
Corporate Smalltalk Consulting Ltd.  http://www.smalltalkconsulting.com
===========================================================================








Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: Teleplace Cog VMs are now available

Andreas.Raab
On 6/20/2010 2:42 PM, Rob Rothwell wrote:
> For the binary-challenged, I am sure some kind soul will post compiled
> versions somewhere...!

For Windows, get yours here:

     http://squeakvm.org/win32/experimental/CogSqueak4.1.zip

Cheers,
   - Andreas

> (I ALWAYS struggle with build.bat, missing files, environment variables,
> and the like...)
>
> Take care,
>
> Rob
>
> On Sun, Jun 20, 2010 at 5:37 PM, John M McIntosh
> <[hidden email]
> <mailto:[hidden email]>> wrote:
>
>
>     Wonderful news.
>
>
>     On 2010-06-20, at 1:11 PM, Eliot Miranda wrote:
>
>>      The Cogit currently supports only x86 and the floating-point
>>     primitives and parts of the platform support code depend on SSE2.
>>     I hope members of the community will attempt to port it, e.g. to
>>     ARM, PowerPC and x86-64.
>
>     Now before anyone asks, it's very doubtful you will see Cogit on the
>     iOS series of operating systems (iphone/ipad). I have had detailed
>     discussions with the Apple Security team since their goal is to
>     prevent user level code from doing dynamic code generation.
>
>     mmap for example won't return a page of memory that is executable.
>
>     Let's say the probability of getting Apple to change their minds on
>     this is zero.
>
>     However I'll look into the StackInterpreter for the Macintosh 5.x VM
>     and the iOS flavour, mind if anyone wants to do that chore early
>     then they should ping me.
>
>
>     --
>     ===========================================================================
>     John M. McIntosh <[hidden email]
>     <mailto:[hidden email]>>   Twitter:  squeaker68882
>     Corporate Smalltalk Consulting Ltd. http://www.smalltalkconsulting.com
>     ===========================================================================
>
>
>
>
>
>


Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: Teleplace Cog VMs are now available

Eliot Miranda-2
In reply to this post by Lukas Renggli


On Sun, Jun 20, 2010 at 2:23 PM, Lukas Renggli <[hidden email]> wrote:
Thank you Eliot.

This is wonderful, it feels like I just got a new laptop :-)

:)
 
I noticed some problems in the continuation tests of Seaside
(WAContinuationTest, WAFlowPlatformTest and
WAPartialContinuationTest):

  "Computation has been terminated" in MethodContext>>cannotReturn:

Is this a known problem? Or is this maybe specific to Pharo on Cog?

This is a known area of problems.  From my announcement:
Status:
The Cogit VM has been our sole VM at Teleplace for nearly a year.  We do 
occasionally find bugs and there are almost certainly areas of functionality 
that we have not touched (for example I know that co-routining does not yet 
work).  If you find a bug please try and create a reproducible test case and 
let me know.  I can't promise to take a look or fix it but I am motivated to do 
so and will try my best as time allows.  Better still if you find and fix bugs 
be sure to let me know.
The use of context-to-stack mapping is fundamental to the performance improvements in Cog and it should be transparent, but so far I've only fixed what issues we've encountered in our usage at Teleplace.  I expect the next few weeks to involve some bug fixing as it is used in a much broader context (no pun intended).  Y'all can help by giving me reproducible test cases.  This can be in the form of a precise specification of an image configuration and a doit, but even easier is to save an image that exhibits the crash on start-up.  e.g.
    Smalltalk saveAs: 'crash.image' thenQuit: true.
    MyClass crashTheVM

I also need to write the sequence of blog posts on the JIT I've been meaning to write for over a year now, as others will want to understand and, I hope, port Cog to other processors.

There's also the need to improve performance.  Right now the code generator is a very naive JIT where every there is no inter-bytecode optimization and hence every push or pop in the bytecode is reflected in a real push or pop in the native code.  This makes loops such as
    1 to: n do: [:i|... ]
very slow because the send: #<= (1 arg); jumpFalseTo: L2 in the underlying code actually runs a non-inlined SmallInteger <= primitive which pushes the actual true object, which the jump bytecode must compare against the actual true false object.  In the VisualWorks code generator the JIT spits out two paths here, one that tests or SmallIntegers and then does the comparison and jump directly branching on the condition codes.  Consequently loops in VsualWorks are much faster.  Also, the object representation is convoluted, compact classes being OK for reducing space overhead but bad at creating overhead and complexity, so much so that Cog as yet lacks machine code primitives for instance creation and at:put:.  

So I hope that soon we'll upgrade the code generator and the object representation and will see another significant jump in compute performance to bring us up to VisualWorks levels.  I hope that we'll also start working on inlining and get the fastest Smalltalk VM ever.  We have a really exciting opportunity here with Exupery and Igor's Native Boost to do inlining in a Klein-like way, i.e. having SMalltalk code do inlining and code optimization.  The only concern here is portability; keeping an intermediate bytecode might be wise.  But we'll see.  I'm also going to be working on the FFI, in part with my GSoC student, but also in Teleplace where I have threaded/non-blocking/overlapped calls woking in prototype form, but as yet lack proper calbacks.


So that speculation aside please do prepare reproducible cases of crashes or differences in behaviour between the current VM and Cog (I've just remembered that another difference is that Cog will produce Float infinity for 1.0/0.0 instead of failing).  Crashes need to be fixed.  Some semantic issues may need to change (for example I think the floating-point behaviour is better, does this need a flag to control as in VisualWorks's case, etc).  But Cog being available is only the beginning, and now we can all be involved.

cheers
Eliot


Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: [Vm-dev] Re: [squeak-dev] Teleplace Cog VMs are now available

Andreas.Raab
In reply to this post by Lukas Renggli
On 6/20/2010 2:23 PM, Lukas Renggli wrote:

>
> Thank you Eliot.
>
> This is wonderful, it feels like I just got a new laptop :-)
>
> I noticed some problems in the continuation tests of Seaside
> (WAContinuationTest, WAFlowPlatformTest and
> WAPartialContinuationTest):
>
>     "Computation has been terminated" in MethodContext>>cannotReturn:
>
> Is this a known problem? Or is this maybe specific to Pharo on Cog?

It's a Cog bug. GeneratorTest dies with the same symptoms in 4.1.

Cheers,
   - Andreas

> On 20 June 2010 22:53, stephane ducasse<[hidden email]>  wrote:
>> This is really an excellent news. I open a lot of new futures!
>> Thanks teleplace, andreas and all the people that helped.
>>
>> Stef
>> PS: you already answer my questions about platform were the jit cannot work so this is great.
>>
>>
>> On Jun 20, 2010, at 10:11 PM, Eliot Miranda wrote:
>>
>>> Hi All,
>>>
>>>        it gives me great pleasure to announce that the Teleplace Cog VMs are now available.  Huge thanks to all at Teleplace who have given me the opportunity to build Cog and release it as open source, been willing guinea pigs braving its bugs, and providing indispensable participation in getting Cog to its current state.  Huge thanks are also due to the original Back To The Future team whose VMMaker Cog extends to write the VM, and to Peter Deutsch from whom I've taken many ideas.
>>>
>>> This release contains two VMs.  The Stack VM, is a cross-platform interpreter that uses context-to-stack mapping to achieve modest performance gains.  The Cog VM is a just-in-time compiler that currently supports only x86 that builds upon the Stack VM to achieve substantial performance improvements.  The release is in the form of a Monticello package containing the VMMaker source and a tarball containing the platform sources, the generated sources and a Squeak 4.1 image containing the VMMaker sources.  Download both at
>>>        http://ftp.squeak.org/Cog/VMMaker-oscog.11.mcz
>>>        http://ftp.squeak.org/Cog/OpenSourceCog.tar.gz
>>>
>>> Cog VMs:
>>>
>>> The Cog VMs are Squeak/Croquet VMs that run closure Squeak/Croquet/Pharo/Cuis images. The VMs support existing plugin source but will require plugins to be recompiled as the VM_PROXY_MAJOR plugin api has been extended.
>>>
>>> This release contains two distinct VMs, the StackInterpreter and the Cogit.  The StackInterpreter is a fully-portable plug-in replacement for the current closure Squeak VMs and images.  The Stack VM uses context-to-stack mapping and a somewhat improved garbage collector to achieve modest but useful performance gains in the 10% to 15% range.  The StackInterpreter is intended to supersede the Squeak VM on platforms where the Cogit cannot be used.  The Cogit extends the StackInterpreter with a just-in-time compiler that uses aggressive inline caching techniques to deliver substantial performance gains in the 3x to 15x range, depending on benchmark.  The Cogit currently supports only x86 and the floating-point primitives and parts of the platform support code depend on SSE2.  I hope members of the community will attempt to port it, e.g. to ARM, PowerPC and x86-64.  The Cogit (excuse the pun) is so named because it is both an interpreter and a JIT, choosing not to generate ma
chine code for large methods, interpreting them instead, the default policy being not to JIT methods with more than 60 literals.
>>>
>>> The Cog VM requires a few minor image changes all in image/NecessaryImageChangesForCogToWork.1.cs.  The JIT's machine-code SmallInteger primitives insist on a SmallInteger receiver so the primitives in LargePositiveInteger = ~= bitAnd: bitOr: butShift: and bitXor: cannot be used and these methods must be deleted.  The Cogit inlines the address of the Character instance table, Smalltalk specialObjectsArray at: 25, into the machine-code at: primitive for faster ByteString>>at: and so the table cannot be rebuilt in SmalltalkImage>>recreateSpecialObjectsArray.  The new version preserves the existing table.  Both VMs maintain floats in platform order to ease implementation of machine code floating-point primitives, and hence internally are in little-endian order instead of big-endian in current Squeak images.  While the VMs convert float order automatically on load they do require special accessing primitives Float>>basicAt:&  Float>>basicAt:put: that undo the reversal and an
swer Float contents in big-endian order so that e.g. Float>>hash is unchanged.  The methods assume these primitives can fail, allowing the code to be used on current Squeak VMs.

>>>
>>> The image/VMMaker-Squeak4.1.image is a Squeak 4.1 image, runnable with the current Squeak VMs, that contains these changes, and can hence also be run with a Cog VM.  But beware, once an image has been saved on Cog it cannot be run by an existing Squeak VM, because existing VMs cannot undo the Float order change.
>>>
>>>
>>> Platform Subsystem:
>>>
>>> Most of the platform subsystem is unchanged but there are some important changes that need description.  The biggest change is the heartbeat and the clock in platforms/unix/vm/sqUnixHeartbeat.c and platforms/win32/vm/sqWin32Heartbeat.c.  The Cog VMs avoid the slow and variable interruptCheckCounter, folding the event check into the stack overflow check on frame build.  The heartbeat, typically 500Hz or 1KHz, changes the stackLimit to a value that will always fail.  On the next frame building send the VM will enter stack overflow handling that, as a side effect, will also check for events.  This is more efficient than the update of interruptCheckCounter and much more regular.  If one is running code that executes long-running primitives (e.g. large integer arithmetic) the counter approach will result in too low an interrupt check frequency, and conversely if one is running normal code the interrupt check frequency can be very high.
>>>
>>> The heartbeat also maintains a 64-bit microsecond clock, UTC microseconds from 1901, from which the backward-compatible millisecond and second clocks are derived.  Primitives exist to answer UTC microseconds and local microseconds.  Updating the clock in the heartbeat results in a 1 or 2 millisecond resolution but avoids the cost of accessing the OS time on every prim tie which we've found important for performance at Teleplace.  The 64-bit microsecond clocks provide a unified time basis and eliminate wrapping (for the next 54,000 years at least).  I hope community images will move to these clocks.  It's worked well in VisualWorks.
>>>
>>> Another significant change is in the external semaphore table support code.  This is now lock-free at the cost of having to specify a maximum number of external semaphores at start-up (default 256).  The support code for the lock-free data structures are processor-specific and is currently implemented only for x86 and gcc-compatible compilers; see platforms/Cross/vm/{sqAtomicOps.h,sqMemoryFence.h}.
>>>
>>> There is also improved crash reporting code that prints a primitive log and a C backtrace in addition to the Smalltalk backtrace.  See platforms/Mac OS/vm/sqMacMain.c, platforms/unix/vm/sqUnixMain.c, platforms/win32/vm/sqWin32Intel.c&  platforms/win32/vm/sqWin32Backtrace.c.
>>>
>>> Finally there is support for the QVMProfiler, a pc-sampling profiler for profiling at the VM level.  See platforms/unix/vm/sqUnixVMProfile.c and platforms/win32/vm/sqWin32VMProfile.c.  The profiler itself is in the VMMaker image described below in Qwaq-VMProfiling.
>>>
>>> There are also changes to do with Teleplace-specific extensions to the HostWindowPlugin but these are not essential to Cog.
>>>
>>>
>>> VMMaker and Slang:
>>>
>>> The image/VMMaker-Squeak4.1.image Squeak 4.1 image contains the complete Cog VMMaker with necessary support code for simulation. This image was used to generate the sources in the src and stacksrc directories.
>>>
>>> Cog's VMMaker is substantially revised and extended from the current VMMaker.  It supports multiple classes, not just Interpreter and superclasses, because both context-to-stack mapping and the Cogit are too complex to write monolithically.  Classes can specify ancilliaryClasses and ancilliaryStructClasses, such as CoInterpreterStackPage, CogMethod and CogAbstractInstruction.  The Monticello package version is included in the header of all generated files and constitutes the version stamp for generated code.  Code is generated in sorted order so that minor changes in the Smalltalk source produce correspondingly minor changes in the generated code.  The gnuification step is built-in to VMMaker.  No effort has been made to maintain 64-bit compatibility.  Apologies, this was unaffordable.
>>>
>>> The VMMaker generates a single source tree used by all platforms.  Instead of deciding at generation time whether to use the Interpreter struct the generated code depends on the SQ_USE_GLOBAL_STRUCT define which can be overridden in platform makefiles.  All plugins live in src/plugins and platform makefiles along with plugins.int and plugins.ext files in the build subdirectories decide which plugins are built as external or internal.  The VM Generation Workspace from Workspace.text workspace contains dots to generate the sources.  We no longer use the VMMakerTool since there should be nothing platform-specific in the generated sources (if we add ports to other ISAs all their source can be included and selected as required by the platform makefiles).
>>>
>>> Since the Cogit generates x86 machine code simulation is much more complex.  There is a support plugin, platforms/Cross/plugins/BochsIA32Plugin that depends on a large simulation of the x86 family implemented in C++ (see  processors/IA32/bochs) and on Alien.  I use the simulator frequently (but note that I haven't had time to build a working version for Squeak 4.1).  I have tested Cog simulation in this image, running on the image/VMMaker-Squeak4.1.image itself.  The VM Simulation Workspace in the VMMaker image contains an example doit that starts the simulator. Be patient, even on a fast machine unhibernating the Squeak display background image takes nearly a minute.  Native fonts do not (yet) simulate correctly, but the system runs.  But note that I have only attempted to build and run the simulator on Mac OS X.  I expect Bochs can be built on linux and win32 but I have not tried.  By the way, I've not described how to run the Bochs simulator on the current Squeak VM.
 That's because the plugin depends on the heartbeat to break out of simulation occasionally via a new interpreterProxy entry point setInterruptCheckChain.  As this isn't supported by the current Squeak VMs the plugin would require modification.  So to simulate first build either of the Cog VMs and then run the simulation with it.
>>>
>>> There are a number of unpublished changes to the base other than those in NecessaryImageChangesForCogToWork.1.cs.  This is partly laziness on my part, partly avoiding publishing things in advance of Cog.  These changes are better motivated once Cog is in use.  There are changes to the "translated primitives" (see implementors of translatedPrimitives) which replace messages with method tags for generation directives.  The Cog VMMaker uses Object>>perform:with:with:with:with:&  Object>>perform:with:with:with:with:with: during simulation, and Collection>>#fold:&  SquenceableCollection>>#copyUpThrough: during generation.  Object>>inline: and Object var:declareC:, which are mispackaged in Kernel in Squeak 4.1 are obsolete (method tags being used instead) and have been removed. I have changed Integer>>hex and Integer>>hex8 back to their original semantics as of 3.8.  Backward compatibility is important and one can easily add new selectors if one wants different functionality.
 VMMaker was here first ;)

>>>
>>>
>>> Tarball:
>>>
>>> The top-level directories in the tarball are
>>>
>>>        src
>>>                the tree for the Cog generated sources including all plugins
>>>        stacksrc/vm
>>>                the directory containing the Stack VM source (plugins can be taken from above)
>>>        platforms
>>>                the usual svn platform tree but including Cog specific changes such as the heartbeat
>>>        processors
>>>                the tree containing simulation support code, i.e. the bochs C++ x86 simulation library, along with a potential ARM, PowerPC&  MIPS simulator, Skeye.
>>>
>>>        image
>>>                the Cog-prepared Squeak 4.1 VMMaker image
>>>        scripts
>>>                some svn scripts to revert unchanged plugins that haven't really changed
>>>
>>>        cygwinbuild
>>>                the win32 build directory
>>>        winbuild
>>>                the old win32 build directory for minnow gcc 2.95.  Not entirely obsolete as the cygwin build as yet fails to generate a functional FFIPlugin
>>>        macbuild
>>>                the CoreVM.xcodeproj and support build projects for Mac OS X 10.5 or better
>>>        unixbuild
>>>                the build directory for linux
>>>
>>>
>>> Building Cog:
>>>
>>> Each build directory above contains a HowToBuild file that describes building in more detail.  The build directories only contain Cogit makefiles.  f you want to build a Stack VM you're on your own but this is very close to the existing Squeak VM build.
>>>
>>>
>>> Status:
>>> The Cogit VM has been our sole VM at Teleplace for nearly a year.  We do occasionally find bugs and there are almost certainly areas of functionality that we have not touched (for example I know that co-routining does not yet work).  If you find a bug please try and create a reproducible test case and let me know.  I can't promise to take a look or fix it but I am motivated to do so and will try my best as time allows.  Better still if you find and fix bugs be sure to let me know.
>>>
>>> License (MIT):
>>> All contributions from Teleplace in this release are
>>> Copyright (c) 2010 Teleplace, Inc.
>>>
>>> Permission is hereby granted, free of charge, to any person obtaining a copy
>>> of this software and associated documentation files (the "Software"), to deal
>>> in the Software without restriction, including without limitation the rights
>>> to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
>>> copies of the Software, and to permit persons to whom the Software is
>>> furnished to do so, subject to the following conditions:
>>>
>>> The above copyright notice and this permission notice shall be included in
>>> all copies or substantial portions of the Software.
>>>
>>> THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
>>> IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
>>> FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
>>> AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
>>> LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
>>> OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
>>> THE SOFTWARE.
>>>
>>> Eliot Miranda
>>> June 2010
>>>
>>>
>>
>>
>>
>
>
>

Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: Teleplace Cog VMs are now available

Stéphane Rollandin
In reply to this post by Eliot Miranda-2
Wonderful !

A naive question: are there now specific coding guidelines for getting
the best out of the JIT, or can we go and code as usual and expect all
methods to run much faster automatically ?


Stef



Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Performance (Re: Teleplace Cog VMs are now available)

Andreas.Raab
On 6/20/2010 3:11 PM, Stéphane Rollandin wrote:
> A naive question: are there now specific coding guidelines for getting
> the best out of the JIT, or can we go and code as usual and expect all
> methods to run much faster automatically ?

The latter. You should expect a 2-3x performance improvement in pretty
much any kind of code (more towards 2x when running primitive and I/O
bound stuff; more torwards 3x when running "pure" Smalltalk code).

Examples of non-I/O bound stuff:

[Smalltalk browseAllCallsOn: #do:] timeToRun.

    Squeak: 779
    CogVM:  258

[Morph methodsDo:[:meth| meth decompileString]] timeToRun.

    Squeak: 537
    CogVM:  166

Examples of I/O heavy stuff:

[Morph compileAll] timeToRun.

    Squeak: 1510
    CogVM:   766

[saveMorphs := World submorphs.
World removeAllMorphs."heh, heh"
time := [1 to: 10 do: [:i |
        Browser fullOnClass: SystemDictionary selector: #macroBenchmarks].
        World submorphs do: [:m | m delete. self currentWorld doOneCycle].
] timeToRun ] ensure:[World addAllMorphs: saveMorphs].

     Squeak:  2633
     CogVM:   1445

As usual for benchmarks YMMV slightly depending on what you're doing but
I'd be surprised if you wouldn't get a 2-3x performance boost regardless
of what you're doing.

Cheers,
   - Andreas

Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: Teleplace Cog VMs are now available

Eliot Miranda-2
In reply to this post by Stéphane Rollandin


2010/6/20 Stéphane Rollandin <[hidden email]>
Wonderful !

A naive question: are there now specific coding guidelines for getting the best out of the JIT, or can we go and code as usual and expect all methods to run much faster automatically ?

Since Cog is a work in progress, and since future performance work should address performance needs of current code I would not want to recommend any changes in style.   Cog is intended to make Squeak code run faster, and hopefully it'll get better at doing so.

best
Eliot


Stef






Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: Teleplace Cog VMs are now available

David T. Lewis
In reply to this post by Eliot Miranda-2
On Sun, Jun 20, 2010 at 01:11:38PM -0700, Eliot Miranda wrote:
> Hi All,
>
>  it gives me great pleasure to announce that the Teleplace Cog VMs are now
> available.  Huge thanks to all at Teleplace who have given me the
> opportunity to build Cog and release it as open source, been willing guinea
> pigs braving its bugs, and providing indispensable participation in getting
> Cog to its current state.  Huge thanks are also due to the original Back To
> The Future team whose VMMaker Cog extends to write the VM, and to Peter
> Deutsch from whom I've taken many ideas.

Congratulations Eliot! And many thanks to those who supported your
work and had the wisdom and foresight to make it publicly available.

Dave


Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: Teleplace Cog VMs are now available

Nikolay Suslov
In reply to this post by Eliot Miranda-2
Congratulations!

And is it possible to run a Croquet image on it, which is based on Squeak 3.8?
What is needed for that? Will closure bootstrap from 2008 (http://www.mirandabanda.org/files/Cog/Closures0811/Bootstrap/ClosureBootstrap-08-11-14.tar.gz) be enough?

Best regards,
Nikolay

On Mon, Jun 21, 2010 at 12:11 AM, Eliot Miranda <[hidden email]> wrote:

Hi All,


it gives me great pleasure to announce that the Teleplace Cog VMs are now available.  Huge thanks to all at Teleplace who have given me the opportunity to build Cog and release it as open source, been willing guinea pigs braving its bugs, and providing indispensable participation in getting Cog to its current state.  Huge thanks are also due to the original Back To The Future team whose VMMaker Cog extends to write the VM, and to Peter Deutsch from whom I've taken many ideas.


This release contains two VMs.  The Stack VM, is a cross-platform interpreter that uses context-to-stack mapping to achieve modest performance gains.  The Cog VM is a just-in-time compiler that currently supports only x86 that builds upon the Stack VM to achieve substantial performance improvements.  The release is in the form of a Monticello package containing the VMMaker source and a tarball containing the platform sources, the generated sources and a Squeak 4.1 image containing the VMMaker sources.  Download both at

http://ftp.squeak.org/Cog/VMMaker-oscog.11.mcz

http://ftp.squeak.org/Cog/OpenSourceCog.tar.gz


Cog VMs:


The Cog VMs are Squeak/Croquet VMs that run closure Squeak/Croquet/Pharo/Cuis images. The VMs support existing plugin source but will require plugins to be recompiled as the VM_PROXY_MAJOR plugin api has been extended.


This release contains two distinct VMs, the StackInterpreter and the Cogit.  The StackInterpreter is a fully-portable plug-in replacement for the current closure Squeak VMs and images.  The Stack VM uses context-to-stack mapping and a somewhat improved garbage collector to achieve modest but useful performance gains in the 10% to 15% range.  The StackInterpreter is intended to supersede the Squeak VM on platforms where the Cogit cannot be used.  The Cogit extends the StackInterpreter with a just-in-time compiler that uses aggressive inline caching techniques to deliver substantial performance gains in the 3x to 15x range, depending on benchmark.  The Cogit currently supports only x86 and the floating-point primitives and parts of the platform support code depend on SSE2.  I hope members of the community will attempt to port it, e.g. to ARM, PowerPC and x86-64.  The Cogit (excuse the pun) is so named because it is both an interpreter and a JIT, choosing not to generate machine code for large methods, interpreting them instead, the default policy being not to JIT methods with more than 60 literals.


The Cog VM requires a few minor image changes all in image/NecessaryImageChangesForCogToWork.1.cs.  The JIT's machine-code SmallInteger primitives insist on a SmallInteger receiver so the primitives in LargePositiveInteger = ~= bitAnd: bitOr: butShift: and bitXor: cannot be used and these methods must be deleted.  The Cogit inlines the address of the Character instance table, Smalltalk specialObjectsArray at: 25, into the machine-code at: primitive for faster ByteString>>at: and so the table cannot be rebuilt in SmalltalkImage>>recreateSpecialObjectsArray.  The new version preserves the existing table.  Both VMs maintain floats in platform order to ease implementation of machine code floating-point primitives, and hence internally are in little-endian order instead of big-endian in current Squeak images.  While the VMs convert float order automatically on load they do require special accessing primitives Float>>basicAt: & Float>>basicAt:put: that undo the reversal and answer Float contents in big-endian order so that e.g. Float>>hash is unchanged.  The methods assume these primitives can fail, allowing the code to be used on current Squeak VMs.


The image/VMMaker-Squeak4.1.image is a Squeak 4.1 image, runnable with the current Squeak VMs, that contains these changes, and can hence also be run with a Cog VM.  But beware, once an image has been saved on Cog it cannot be run by an existing Squeak VM, because existing VMs cannot undo the Float order change.



Platform Subsystem:


Most of the platform subsystem is unchanged but there are some important changes that need description.  The biggest change is the heartbeat and the clock in platforms/unix/vm/sqUnixHeartbeat.c and platforms/win32/vm/sqWin32Heartbeat.c.  The Cog VMs avoid the slow and variable interruptCheckCounter, folding the event check into the stack overflow check on frame build.  The heartbeat, typically 500Hz or 1KHz, changes the stackLimit to a value that will always fail.  On the next frame building send the VM will enter stack overflow handling that, as a side effect, will also check for events.  This is more efficient than the update of interruptCheckCounter and much more regular.  If one is running code that executes long-running primitives (e.g. large integer arithmetic) the counter approach will result in too low an interrupt check frequency, and conversely if one is running normal code the interrupt check frequency can be very high.


The heartbeat also maintains a 64-bit microsecond clock, UTC microseconds from 1901, from which the backward-compatible millisecond and second clocks are derived.  Primitives exist to answer UTC microseconds and local microseconds.  Updating the clock in the heartbeat results in a 1 or 2 millisecond resolution but avoids the cost of accessing the OS time on every prim tie which we've found important for performance at Teleplace.  The 64-bit microsecond clocks provide a unified time basis and eliminate wrapping (for the next 54,000 years at least).  I hope community images will move to these clocks.  It's worked well in VisualWorks.


Another significant change is in the external semaphore table support code.  This is now lock-free at the cost of having to specify a maximum number of external semaphores at start-up (default 256).  The support code for the lock-free data structures are processor-specific and is currently implemented only for x86 and gcc-compatible compilers; see platforms/Cross/vm/{sqAtomicOps.h,sqMemoryFence.h}.


There is also improved crash reporting code that prints a primitive log and a C backtrace in addition to the Smalltalk backtrace.  See platforms/Mac OS/vm/sqMacMain.c, platforms/unix/vm/sqUnixMain.c, platforms/win32/vm/sqWin32Intel.c & platforms/win32/vm/sqWin32Backtrace.c.


Finally there is support for the QVMProfiler, a pc-sampling profiler for profiling at the VM level.  See platforms/unix/vm/sqUnixVMProfile.c and platforms/win32/vm/sqWin32VMProfile.c.  The profiler itself is in the VMMaker image described below in Qwaq-VMProfiling.


There are also changes to do with Teleplace-specific extensions to the HostWindowPlugin but these are not essential to Cog.



VMMaker and Slang:


The image/VMMaker-Squeak4.1.image Squeak 4.1 image contains the complete Cog VMMaker with necessary support code for simulation. This image was used to generate the sources in the src and stacksrc directories.


Cog's VMMaker is substantially revised and extended from the current VMMaker.  It supports multiple classes, not just Interpreter and superclasses, because both context-to-stack mapping and the Cogit are too complex to write monolithically.  Classes can specify ancilliaryClasses and ancilliaryStructClasses, such as CoInterpreterStackPage, CogMethod and CogAbstractInstruction.  The Monticello package version is included in the header of all generated files and constitutes the version stamp for generated code.  Code is generated in sorted order so that minor changes in the Smalltalk source produce correspondingly minor changes in the generated code.  The gnuification step is built-in to VMMaker.  No effort has been made to maintain 64-bit compatibility.  Apologies, this was unaffordable.


The VMMaker generates a single source tree used by all platforms.  Instead of deciding at generation time whether to use the Interpreter struct the generated code depends on the SQ_USE_GLOBAL_STRUCT define which can be overridden in platform makefiles.  All plugins live in src/plugins and platform makefiles along with plugins.int and plugins.ext files in the build subdirectories decide which plugins are built as external or internal.  The VM Generation Workspace from Workspace.text workspace contains dots to generate the sources.  We no longer use the VMMakerTool since there should be nothing platform-specific in the generated sources (if we add ports to other ISAs all their source can be included and selected as required by the platform makefiles).


Since the Cogit generates x86 machine code simulation is much more complex.  There is a support plugin, platforms/Cross/plugins/BochsIA32Plugin that depends on a large simulation of the x86 family implemented in C++ (see  processors/IA32/bochs) and on Alien.  I use the simulator frequently (but note that I haven't had time to build a working version for Squeak 4.1).  I have tested Cog simulation in this image, running on the image/VMMaker-Squeak4.1.image itself.  The VM Simulation Workspace in the VMMaker image contains an example doit that starts the simulator. Be patient, even on a fast machine unhibernating the Squeak display background image takes nearly a minute.  Native fonts do not (yet) simulate correctly, but the system runs.  But note that I have only attempted to build and run the simulator on Mac OS X.  I expect Bochs can be built on linux and win32 but I have not tried.  By the way, I've not described how to run the Bochs simulator on the current Squeak VM.  That's because the plugin depends on the heartbeat to break out of simulation occasionally via a new interpreterProxy entry point setInterruptCheckChain.  As this isn't supported by the current Squeak VMs the plugin would require modification.  So to simulate first build either of the Cog VMs and then run the simulation with it.


There are a number of unpublished changes to the base other than those in NecessaryImageChangesForCogToWork.1.cs.  This is partly laziness on my part, partly avoiding publishing things in advance of Cog.  These changes are better motivated once Cog is in use.  There are changes to the "translated primitives" (see implementors of translatedPrimitives) which replace messages with method tags for generation directives.  The Cog VMMaker uses Object>>perform:with:with:with:with: & Object>>perform:with:with:with:with:with: during simulation, and Collection>>#fold: & SquenceableCollection>>#copyUpThrough: during generation.  Object>>inline: and Object var:declareC:, which are mispackaged in Kernel in Squeak 4.1 are obsolete (method tags being used instead) and have been removed. I have changed Integer>>hex and Integer>>hex8 back to their original semantics as of 3.8.  Backward compatibility is important and one can easily add new selectors if one wants different functionality.  VMMaker was here first ;)



Tarball:


The top-level directories in the tarball are


src

the tree for the Cog generated sources including all plugins

stacksrc/vm

the directory containing the Stack VM source (plugins can be taken from above)

platforms

the usual svn platform tree but including Cog specific changes such as the heartbeat

processors

the tree containing simulation support code, i.e. the bochs C++ x86 simulation library, along with a potential ARM, PowerPC & MIPS simulator, Skeye.


image

the Cog-prepared Squeak 4.1 VMMaker image

scripts

some svn scripts to revert unchanged plugins that haven't really changed


cygwinbuild

the win32 build directory

winbuild

the old win32 build directory for minnow gcc 2.95.  Not entirely obsolete as the cygwin build as yet fails to generate a functional FFIPlugin

macbuild

the CoreVM.xcodeproj and support build projects for Mac OS X 10.5 or better

unixbuild

the build directory for linux



Building Cog:


Each build directory above contains a HowToBuild file that describes building in more detail.  The build directories only contain Cogit makefiles.  f you want to build a Stack VM you're on your own but this is very close to the existing Squeak VM build.



Status:

The Cogit VM has been our sole VM at Teleplace for nearly a year.  We do occasionally find bugs and there are almost certainly areas of functionality that we have not touched (for example I know that co-routining does not yet work).  If you find a bug please try and create a reproducible test case and let me know.  I can't promise to take a look or fix it but I am motivated to do so and will try my best as time allows.  Better still if you find and fix bugs be sure to let me know.


License (MIT):

All contributions from Teleplace in this release are

Copyright (c) 2010 Teleplace, Inc.


Permission is hereby granted, free of charge, to any person obtaining a copy

of this software and associated documentation files (the "Software"), to deal

in the Software without restriction, including without limitation the rights

to use, copy, modify, merge, publish, distribute, sublicense, and/or sell

copies of the Software, and to permit persons to whom the Software is

furnished to do so, subject to the following conditions:


The above copyright notice and this permission notice shall be included in

all copies or substantial portions of the Software.


THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR

IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,

FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE

AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER

LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,

OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN

THE SOFTWARE.


Eliot Miranda

June 2010








Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: Squeak 4.1 on Cog (Re: Teleplace Cog VMs are now available)

laza
In reply to this post by Andreas.Raab
Just for the heck of it:

Squeak = Squeak-4.0.3.2202-linux_i386 on Ubuntu 10.4
CogVM = wine and win32 CogVM on Ubuntu 10.4 same image, same system

[Smalltalk browseAllCallsOn: #do:] timeToRun.

 Squeak: 1984
 CogVM:   877

[Morph methodsDo:[:meth| meth decompileString]] timeToRun.

 Squeak: 1206
 CogVM:   580

Examples of I/O heavy stuff:

[Morph compileAll] timeToRun.

 Squeak: 3289
 CogVM:  3076

[saveMorphs := World submorphs.

World removeAllMorphs."heh, heh"
time := [1 to: 10 do: [:i |
      Browser fullOnClass: SystemDictionary selector: #macroBenchmarks].
      World submorphs do: [:m | m delete. self currentWorld doOneCycle].
] timeToRun ] ensure:[World addAllMorphs: saveMorphs].

  Squeak:  6586
  CogVM:   5743

Alex

PS: Running all tests with the win32 CogVM using wine will produce a fatal vm crash at some point


Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: Teleplace Cog VMs are now available

Paolo Bonzini-2
In reply to this post by Eliot Miranda-2
On 06/20/2010 10:11 PM, Eliot Miranda wrote:
> Hi All,
>
> it gives me great pleasure to announce that the Teleplace Cog VMs are
> now available.  Huge thanks to all at Teleplace who have given me the
> opportunity to build Cog and release it as open source

Great news, thanks Eliot for your work!

Paolo

Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: Squeak 4.1 on Cog (Re: Teleplace Cog VMs are now available)

Göran Krampe
In reply to this post by laza
On 06/21/2010 10:49 AM, Alexander Lazarević wrote:
> Just for the heck of it:
>
> Squeak = Squeak-4.0.3.2202-linux_i386 on Ubuntu 10.4
> CogVM = wine and win32 CogVM on Ubuntu 10.4 same image, same system
>
> [Smalltalk browseAllCallsOn: #do:] timeToRun.
>
>   Squeak: 1984
>   CogVM:   877
[SNIP]

Interesting, you don't seem to get as good numbers as Andreas did? Do
you guys have different CPUs too?

regards, Göran


123
Loading...