<div dir="ltr">Hi Ben, (and hi all prospective VM hackers)<br><div class="gmail_quote"><div dir="ltr">On Tue, Aug 28, 2018 at 12:22 PM Ben Coman <<a href="mailto:btc@openinworld.com">btc@openinworld.com</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex"> <div dir="ltr">Hi Eliot, Thanks for the detailed response.<br><br><div class="gmail_quote"><div dir="ltr">On Tue, 28 Aug 2018 at 04:21, Eliot Miranda <<a href="mailto:eliot.miranda@gmail.com" target="_blank">eliot.miranda@gmail.com</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex"><div dir="ltr"><br><div class="gmail_quote"><div dir="ltr">On Mon, Aug 27, 2018 at 11:36 AM Ben Coman <<a href="mailto:btc@openinworld.com" target="_blank">btc@openinworld.com</a>> wrote:</div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div class="gmail_quote"><div>Back when I was having a go at new mutex primitives, </div><div>when a process "A" failed to lock a mutex, I wanted to return </div><div>a primitive-failed-code in addition to the usual context-change to different process "B"</div><div>However what I observed was that because of the context change </div><div>the primitive-failed-code incorrectly ended up returned on the stack of process "B".</div></div></div></blockquote><div><br></div><div>Your description doesn't match how (I understand) the VM works. The only way that Process A can initiate a process switch, mutex lock, et al, is by sending a message to some object (a process, mutex or semaphore). So we're talking about Process>>suspend & resume as well as the lock/unlock and wait/signal primitives. Primitive failure *always* delivers a primitive error to the method that contains the primitive and, in this case, initiated the process switch. Primitives validate their arguments and then succeed, or fail, leaving their arguments undisturbed (there is one regrettable exception to this in the BttF/Cog VM which is the segment loading primitive that leaves its input word array scrambled if a failure occurs, rather than incur the cost of cloning the array).</div></div></div></blockquote><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div class="gmail_quote"><div><br></div><div>So the only way that a primitive could fail and the error code end up on the wrong process's stack would be if the primitive was mis-designed to not validate before occurring. </div></div></div></blockquote><div><br></div><div><br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div class="gmail_quote"><div>Essentially it can not fail and cause a side effect. </div></div></div></blockquote><div><br></div><div>This was my first foray into writing a primitive (still on my backlog to be completed).</div><div>I was aware of validating the arguments and leaving them undisturbed for a failure, </div><div>but wasn't paying attention to primitive failure being completely free from side effects. </div><div><br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div class="gmail_quote"><div>Primitives should be side-effect free when they fail and hence if a process switch primitive fails, it cannot yet have caused a process switch and therefore the error code would have to be delivered to process A's stack.</div></div></div></blockquote><div><br></div><div><div>That is kind of the outcome of what I was proposing. I was trying for a mutex locking primitive that could fail without causing a side-effect "in the image". Maybe its not valid to distinguish between side-effects "inside" or "outside" the image, but I thought it might be reasonable for a flag hidden "in the VM" to be just another event checked by #checkForEventsMayContextSwitch: . Effectively the "in image" effect happens outside the primitive, a bit like reaching nextWakeupUsecs, or like I imagine a callback might work.</div><div><br></div><div>A side thought was that if context-changes occurred in a *single* location in #checkForEventsMayContextSwitch, </div><div>it might be easier to make an "Idle VM"</div></div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div class="gmail_quote"><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div class="gmail_quote"><div><br></div><div>I'm stretching my memory so there is a reasonable change this is misleading...</div><div>but I believe I observed this happening in CoInterpreter>>internalExecuteNewMethod </div><div>near this code..</div><div><br></div><div> "slowPrimitiveResponse may of course context-switch. ..."</div><div> succeeded := self slowPrimitiveResponse.</div><div> ...</div><div> succeeded ifTrue: [....</div></div></div></blockquote><div><br></div><div>But internalExecuteNewMethod doesn't contain the switch code, internalActivateNewMethod does, and it does the process switch *after* delivering the primitive failure code, see reapAndResetErrorCodeTo:header: in the following:</div><div><br></div><div>internalActivateNewMethod<br></div><div><div><span class="gmail-m_7335283777981739541gmail-m_8629604044109620616gmail-Apple-tab-span" style="white-space:pre-wrap"> </span>...</div><div><span class="gmail-m_7335283777981739541gmail-m_8629604044109620616gmail-Apple-tab-span" style="white-space:pre-wrap"> </span>(self methodHeaderHasPrimitive: methodHeader) ifTrue:</div><div><span class="gmail-m_7335283777981739541gmail-m_8629604044109620616gmail-Apple-tab-span" style="white-space:pre-wrap"> </span>["Skip the CallPrimitive bytecode, if it's there, and store the error code if the method starts</div><div><span class="gmail-m_7335283777981739541gmail-m_8629604044109620616gmail-Apple-tab-span" style="white-space:pre-wrap"> </span> with a long store temp. Strictly no need to skip the store because it's effectively a noop."</div><div><span class="gmail-m_7335283777981739541gmail-m_8629604044109620616gmail-Apple-tab-span" style="white-space:pre-wrap"> </span> localIP := localIP + (self sizeOfCallPrimitiveBytecode: methodHeader).</div><div><span class="gmail-m_7335283777981739541gmail-m_8629604044109620616gmail-Apple-tab-span" style="white-space:pre-wrap"> </span> primFailCode ~= 0 ifTrue:</div><div><span class="gmail-m_7335283777981739541gmail-m_8629604044109620616gmail-Apple-tab-span" style="white-space:pre-wrap"> </span>[self reapAndResetErrorCodeTo: localSP header: methodHeader]].</div><div><br></div><div><span class="gmail-m_7335283777981739541gmail-m_8629604044109620616gmail-Apple-tab-span" style="white-space:pre-wrap"> </span>self assert: (self frameNumArgs: localFP) == argumentCount.</div><div><span class="gmail-m_7335283777981739541gmail-m_8629604044109620616gmail-Apple-tab-span" style="white-space:pre-wrap"> </span>self assert: (self frameIsBlockActivation: localFP) not.</div><div><span class="gmail-m_7335283777981739541gmail-m_8629604044109620616gmail-Apple-tab-span" style="white-space:pre-wrap"> </span>self assert: (self frameHasContext: localFP) not.</div><div><br></div><div><span class="gmail-m_7335283777981739541gmail-m_8629604044109620616gmail-Apple-tab-span" style="white-space:pre-wrap"> </span>"Now check for stack overflow or an event (interrupt, must scavenge, etc)."</div><div><span class="gmail-m_7335283777981739541gmail-m_8629604044109620616gmail-Apple-tab-span" style="white-space:pre-wrap"> </span>localSP < stackLimit ifTrue:</div><div><span class="gmail-m_7335283777981739541gmail-m_8629604044109620616gmail-Apple-tab-span" style="white-space:pre-wrap"> </span>[self externalizeIPandSP.</div><div><span class="gmail-m_7335283777981739541gmail-m_8629604044109620616gmail-Apple-tab-span" style="white-space:pre-wrap"> </span> switched := self handleStackOverflowOrEventAllowContextSwitch:</div><div><span class="gmail-m_7335283777981739541gmail-m_8629604044109620616gmail-Apple-tab-span" style="white-space:pre-wrap"> </span>(self canContextSwitchIfActivating: newMethod header: methodHeader).</div><div><span class="gmail-m_7335283777981739541gmail-m_8629604044109620616gmail-Apple-tab-span" style="white-space:pre-wrap"> </span> self returnToExecutive: true postContextSwitch: switched.</div><div><span class="gmail-m_7335283777981739541gmail-m_8629604044109620616gmail-Apple-tab-span" style="white-space:pre-wrap"> </span> self internalizeIPandSP]</div></div><div><br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div class="gmail_quote"><div><br></div><div>Though I can't exactly put my finger on explaining why, my intuition is that </div><div>changing threads "half way" through a bytecode is a bad thing. </div></div></div></blockquote><div><br></div><div>Indeed it is, and the VM does not do this. It is possible that the execution simulation machinery in Context, InstructionStream at al could have been written carelessly to allow this to occur, but it cannot and does not occur in the VM proper.</div></div></div></blockquote><div><br></div><div><div> I made a chart to understand this better. One thing first, I'm not sure I've correctly linked execution of the primitives into slowPrimitiveResponse. I'm not at all clear about how internalExecuteNewMethod selects between internalQuickPrimitiveResponse </div><div>and slowPrimitiveResponse, and what is the difference between them?</div></div></div></div></blockquote><div><br></div><div>Well, the first thing to say is that this is a magnificent diagram; thank you. The problem is that the VM, and hence the diagram, is much more complex than the blue book specification, essentially because the VM is a highly optimized interpreter, whereas the specification is bare bones. So I would ask you, and anyone else who wants to understand a Smalltalk-80 VM (a VM that provides Context objects for method activations, rather than a Smalltalk that uses a more conventional stack model) to read the Blue Book Specification:<a href="http://www.mirandabanda.org/bluebook/bluebook_chapter28.html"> http://www.mirandabanda.org/bluebook/bluebook_chapter28.html</a> carefully and fully. This is the last section of Smalltalk-80: The Language and its Implementation, by Adele Goldberg and David Robson. The specification is well-written and clear and once digested serves as essential reference for understanding a more complex production VM.</div><div><br></div><div>Now to your question: "I'm not at all clear about how internalExecuteNewMethod selects between internalQuickPrimitiveResponse and slowPrimitiveResponse, and what is the difference between them?".</div><div><br></div><div>First, in a system that notionally allocates a context object to hold every activation, leaf routines are extremely expensive if all they do is answer a constant or an instance variable. Dan Ingall's optimization is to avoid activations by providing a set of quick primitives that answer an instance variable whose slot index is from 0 to 255, or self, nil, true, false, -1, 0, 1 & 2. The self, nil, true, false, -1, 0, 1 & 2 constants are derived from a static frequency analysis of literals and variable references in Smalltalk code and are echoed in the original bytecode set, byte codes 112-119 bering 01110iii Push (receiver, true, false, nil, -1, 0, 1, 2) [iii]. internalQuickPrimitiveResponse handles precisely these primitives, and these primitives only. These primitives have the property that they can never fail, so they are invoked along a path that does not reset the flag used to identify primitive failure, nor test it.</div><div><br></div><div>All other primitives are handled by slowPrimitiveResponse. This requires clearing primErrorCode and calling a function implementing the primitive and then testing primErrorCode before either continuing or building an activation for the failing primitive. There is another important task of internalExecuteNewMethod, which is to store the "internal" frame and stack pointers (localFP & localSP) into the global interpreter frame and stack pointers (framePointer and stackPointer) before invoking and then restoring localFP & localSP from framePointer and stackPointer after having invoked slowPrimitiveResponse. In the Back-to-the-Future (BTTF) VMs (the original Squeak VM and the Stack and Cog VMs) primitives access their receiver and arguments through framePointer and stackPointer. But these, being global, are slow and without compiler-specific hacks cannot be placed in registers. The Slang translator and the interpreter code collaborate to inline much of the interpreter, including every method beginning with internal, into the interpret routine in which localFP localSP and localIP are declared, hence allowing a C compiler to assign these variables to registers. So another reason slowPrimitiveResponse is slow is that it writes and reads localFP, localSP & localIP to/from framePointer, stackPointer & instructionPointer. But because it does so, primitives that change the execution context (process switch primitives that switch to another "stack" of contexts, or eval primitives such as perform:with:* and value:value* which build another frame) can be written and change the execution context at a send point (primitive invocation is always at a send).</div><div><br></div><div>Note that all the internal methods have non-internal duals that do the same thing but use framePointer, stackPointer & instructionPointer, not localFP, localSP & localIP. These are used to implement the eval primitives perform:* and withArgs:executeMethod: since these may also invoke primitives. And hence you might be able to get your head around my favorite Smalltalk construction:</div><div><div><span style="white-space:pre"> </span>| array |</div><div><span class="gmail-Apple-tab-span" style="white-space:pre"> </span>array := { #perform:withArguments:. nil }.</div><div><span class="gmail-Apple-tab-span" style="white-space:pre"> </span>array at: 2 put: array.</div><div><span class="gmail-Apple-tab-span" style="white-space:pre"> </span>array perform: array first withArguments: array</div></div><div>;-)</div><div><br></div><div>Given the primitives are always invoked at a send we can see how elegant Dan's invention of primitive failure is. Primitives are essentially transactional and atomic. They validate their arguments and if validation succeeds they carry out their action and answer a result as if from some normal send. But if validation fails, or if they are unimplemented, the method containing the primitive reference (<primitive: 61>, <primitive: 'primitiveSocketAccept' module: 'SocketPlugin'>) is simply activated as if the primitive didn't exist, or as if the method was a normal method. Hence primitives are optional, in that if the method body does what the primitive does then no one can tell if the primitive is doing the work or Smalltalk code, except by measuring performance. Hence for example large integer arithmetic and string display primitives are optional and serve to accelerate the system.</div><div><br></div><div>There is one other route by which primitives are executed, also at the point of send, and this is via the special selector bytecocdes. Another of Dan's excellent optimizations, the special selectors both save space and make the interpreter faster. They save space by encoding the 32 most frequently occurring sends as one byte bytecodes, hence saving the 2 (16-bit Smalltalk-80), 4 (32-bit Squeak) or 8 (64-bit Squeak) bytes to store the selector in a method's literal frame. But some of them also speed up the interpreter by statically predicting the receiver type. i.e. #+, #-, $/ #*, #<, #>, #<= et al are most often sent to integers, and hence these bytecodes, as specified in the Blue Book, check for the top two elements on the stack being SmallIntegers, and if so replace the two top elements by the result of the operation, avoiding a send and primitive dispatch. Note that in the BttF interpreter this checking is extended both to apply to Float, and to check for a fool,lowing branch after the conditionals #<, #<= et al, so that the result doesn't have to be reified into a boolean that is tested immediately; effectively the following branch gets folded into the relational special selector bytecode. The JIT uses this same technique, but is able to do a much better job because, for example, it can know if a relational special selector send is followed by a jump bytecode or not an JIT time.</div><div><br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div class="gmail_quote"><div><br></div><div><div><img src="cid:ii_jle2plu34" alt="ContextChange-Existing.png" width="472" height="260"><br></div></div><div><br></div><div><div>So I understand that checkForEventsMayContextSwitch: called at the end of internalActivateNewMethod<br>occurs after bytecode execution has completed, so that context switches made there are done "between" bytecodes.</div><div>However my perspective is that internalExecuteNewMethod is only half way through a bytecode execution when </div><div>the primitives effect context changes. So internalActivateNewMethod ends up working on a different Process than internalExecuteNewMethod started with. The bottom three red lines in the chart are what I considered to be changing threads "half way" through a bytecode.</div></div></div></div></blockquote><div><br></div><div>More accurately, checkForEventsMayContextSwitch: is called on activating a method, after the send has occurred, but before the first bytecode has executed. Hence primitive sends are not suspension points unless the primitive is a process switch or eval primitive. Instead, just as a non-primitive (or failing primitive) method is being activated checkForEventsMayContextSwitch: is invoked.</div><div><br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div class="gmail_quote"><div><div>Ahh, I'm slowly coming to grips with this. It was extremely confusing at the time why my failure code from the primitive was turning up in a different Process, though I then learnt a lot digging to discover why. In summary, if the primitive succeeds it simply returns from internalExecuteNewMethod and internalActivateNewMethod never sees the new Process B. My problem violating the primitive-failure side-effect rule was that internalActivateNewMethod trying to run Process A in-Image-primitive-failure code <br></div><div>instead ran Process B in-Image-primitive-failure code.</div></div></div></div></blockquote><div><br></div><div>Right. So that design requirement is key to the VM architecture. That was something I had to explain to Alistair when he did the first cut of the FileAttributesPlugin, which used to return failure codes on error instead of failing, leaving error recovery to clients. And so it underscores the importance of reading the blue book specification. [We really should do an up-to-date version that describes a simplified 32-bit implementation].</div><div><br></div><div>The design requirement that primitives validate their arguments and complete atomically or fail without side-effects is also key to Spur. Spur speeds up become by using transparent forwarders (any object can become a forwarder) but reduces the cost of transparent forwarders by arranging the forwarders only have to be checked for during a send or during primitive argument validation. The receiver has to be accessed during a send anyway, and so Spur is able to move the check for a forwarder to the lookup side of a send, only checking for forwarders if the method cache probe failed, which will always be the case for forwarders. Likewise, primitive argument validation always fails for forwarders, and hence in the Spur slowPrimitiveResponse is a check on failure for a primitive's "accessor depth", the depth of the graph of arguments it validates is. If a primitive has a non-negative accessor depth then on failure the arguments are traversed to that depth, and any forwarders encountered are followed, fixing up that part of the object graph, and the primitive retried. Without Dan's primitive design, Spur could not work well.</div><div><br></div><div><br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div class="gmail_quote"><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div class="gmail_quote"><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div class="gmail_quote"><div>Interestingly the comment in #transferTo:from says... <br></div><div> "Record a process to be awoken on the next interpreter cycle."</div><div>which sounds like what I'd propose, but actually it doesn't wait for </div><div>the next interpreter cycle and instead immediately changes context.</div></div></div></blockquote><div><br></div><div>No it doesn't. It effects the process change, but control continues with the caller, allowing the caller to do other things before the process resumes. </div></div></div></blockquote><div><br></div><div>In my case I believe "control continues with the caller" was not true since internalActivateNewMethod </div><div>was trying to run the in-Image-primitive-failure code after the process changed.</div><div>But that was because I violated the side-effect rule.</div><div><br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div class="gmail_quote"><div>For example, if in checkForEventsMayContextSwitch: more than one semaphore is signaled (it could initiate any combination of signals of the low space semaphore, the input semaphore, external semaphores (associated with file descriptors & sockets), and the delay semaphore) then only the highest priority process would be runnable after the sequence, and several transferTo:[from:]'s could have been initiated from these signals, depending on process priority. But checkForEventsMayContextSwitch: will not finish mid-sequence. It will always complete all of its signals before returning to code that can then resume the newly activated process.</div></div></div></blockquote><div><br></div><div>Just to summarise to check I understood this correctly, no bytecode is executed during checkForEventsMayContextSwitch:. </div><div>That is, its multiple transferTo: calls don't re-enter the interpreter? It just which Process is set to run changes,</div><div>until at the end of checkForEventsMayContextSwitch: it returns to the interpreter to pick up the next bytecode of the active process.</div></div></div></blockquote><div><br></div><div>That's right. transferTo: merely sets framePointer, stackPointer & instructionPointer to those of the highest priority runnable process. Execution resumes at that point once checkForEventsMayContextSwitch returns to its caller. For details to do with mixing JITted code and interpreted code checkForEventsMayContextSwitch: answers a variable indicating if a switch actually took place, and returnToExecutive:postContextSwitch: may or may not have to longjmp back into the interpreter or jump into machine code. But that's just an optimization. Execution could always resume in the interpreter; the VM would simply be a little slower, and arguably a lot less complicated.</div><div><br></div><div>So the moral is, both the BttF and Cog VMs are complex, because they are optimized, Cog adding an entirely new level of complexity over the BttF interpreter VM. If you want to see the wood for the trees first read the Blue Book spec <a href="http://www.mirandabanda.org/bluebook/bluebook_chapter28.html">http://www.mirandabanda.org/bluebook/bluebook_chapter28.html</a>.</div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div class="gmail_quote"><div>cheers -ben. </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex">
</blockquote></div></div>
</blockquote></div><div><br></div><div dir="ltr" class="gmail_signature"><div dir="ltr"><div><span style="font-size:small;border-collapse:separate"><div>_,,,^..^,,,_<br></div><div>best, Eliot</div></span></div></div></div></div>