Difference between revisions of "SLICC"

From gem5
Jump to: navigation, search
(Add link to transition guide)
Line 163: Line 163:
 
   transition({IS, IM, MI, II}, {Load, Ifetch, Store, Replacement}) {
 
   transition({IS, IM, MI, II}, {Load, Ifetch, Store, Replacement}) {
 
     z_stall;
 
     z_stall;
 +
  }
 +
 +
 +
=== Special Functions ===
 +
 +
==== Stalling/Recycling/Waiting input ports ====
 +
 +
One of the more complicated internal features of SLICC and the resulting state machines is how the deal with the situation when events cannot be process due to the cache block being in a transient state.  There are several possible ways to deal with this situation and each solution has different tradeoffs.  This sub-section attempts to explain the differences.  Please email the gem5-user list for further follow-up.
 +
 +
* Stalling the input port.  The simpliest way to handle events that can't be processed is to simply stall the input port.  The correct way to do this is to include the "z_stall" action within the transition statement:
 +
 +
  transition({IS, IM, MI, II}, {Load, Ifetch, Store, Replacement}) {
 +
    z_stall;
 +
  }
 +
 +
Internally SLICC will return a ProtocolStall for this transition and no subsequent messages from the associated input port will be processed until the stalled message is processed.  However, the other input ports will be analyzed for ready messages and processed in parallel.  While this is a relatively simple solution, one may notice that stalling unrelated messages on the same input port will cause excessive and unecessary stalls. 
 +
 +
One thing to note is '''Do Not''' leave the transition statement blank like so:
 +
 +
  transition({IS, IM, MI, II}, {Load, Ifetch, Store, Replacement}) {
 +
    // stall the input port by simply not popping the message
 +
  }
 +
 +
This will cause SLICC to return success for this transition and SLICC will continue to repeatly analyze the same input port.  The result is eventual deadlock.
 +
 +
* Recycling the input port.  The better performat but more unrealistic solution is to recycle the stalled message on the input port.  The way to do this is to use the "zz_recycleMandatoryQueue" action:
 +
 +
  action(zz_recycleMandatoryQueue, "\z", desc="Send the head of the mandatory queue to the back of the queue.") {
 +
    mandatoryQueue_in.recycle();
 +
  }
 +
 +
  transition({IS, IM, MI, II}, {Load, Ifetch, Store, Replacement}) {
 +
    zz_recycleMandatoryQueue;
 +
  }
 +
 +
The result of this action is that the transition returns a Protocol Stall and the offending message moved to the back of the FIFO input port.  Therefore, other unrelated messages on the same input port can be processed.  The problem with this solution is that recycled messages may be analyzed and reanalyzed every cycle until an address changes state.
 +
 +
* Stall and wait the input port.  An even better, but more complicated solution is to "stall and wait" the offending input messge.  The way to do this is to use the "z_stallAndWaitMandatoryQueue" action:
 +
 +
  action(z_stallAndWaitMandatoryQueue, "\z", desc="recycle L1 request queue") {
 +
    stall_and_wait(mandatoryQueue_in, address);
 +
  }
 +
 +
  transition({IS, IM, IS_I, M_I, SM, SINK_WB_ACK}, {Load, Ifetch, Store, L1_Replacement}) {
 +
    z_stallAndWaitMandatoryQueue;
 +
  }
 +
 +
The result of this action is that the transition returns success, which is ok because stall_and_wait moves the offending message off the input port and to a side table associated with the input port.  The message will not be analyzed again until it is woken up.  In the meantime, other unrelated messages wil be processed.
 +
 +
The complicated part of stall and wait is that stalled messages must be explicitly woken up by other messages/transitions.  In particular, transitions that move an address to a base state should wake up potentially stalled messages waiting for that address:
 +
 +
  action(kd_wakeUpDependents, "kd", desc="wake-up dependents") {
 +
    wakeUpBuffers(address);
 +
  }
 +
 +
  transition(M_I, WB_Ack, I) {
 +
    s_deallocateTBE;
 +
    o_popIncomingResponseQueue;
 +
    kd_wakeUpDependents;
 +
  }
 +
 +
Replacements are particularly complicated since stalled addresses are not associated with the same address they are actually waiting to change.  In those situations all waiting messages must be woken up:
 +
 +
  action(ka_wakeUpAllDependents, "ka", desc="wake-up all dependents") {
 +
    wakeUpAllBuffers();
 +
  }
 +
 +
  transition(I, L2_Replacement) {
 +
    rr_deallocateL2CacheBlock;
 +
    ka_wakeUpAllDependents;
 
   }
 
   }
  

Revision as of 23:55, 14 March 2012

SLICC is a domain specific language for specifying cache coherence protocols. The SLICC compiler generates C++ code for different controllers, which can work in tandem with other parts of Ruby. The compiler also generates an HTML specification of the protocol.

Input To the Compiler

The SLICC compiler takes as input files that specify the controllers involved in the protocol. The .slicc file specifies the different files used by the particular protocol under consideration. For example, if trying to specify the MI protocol using SLICC, then we may ues MI.slicc as the file that specifies all the files necessary for the protocol. The files necessary for specifying a protocol include the definitions of the state machines for different controllers, and of the network messages that are passed on between these controllers.

The files have a syntax similar to that of C++. The compiler, written using PLY (Python Lex-Yacc), parses these files to create an Abstract Syntax Tree (AST). The AST is then traversed to build some of the internal data structures. Finally the compiler outputs the C++ code by traversing the tree again. The AST represents the hierarchy of different structures present with in a state machine. We describe these structures next.

Protocol State Machines

In this section we take a closer look at what goes in to a file containing specification of a state machine.

Specifying Data Members

Each state machine is described using SLICC's machine datatype. Each machine has several different types of members. Machines for cache and directory controllers include cache memory and directory memory data members respectively. We will use the MI protocol available in src/mem/protocol as our running example. So here is how you might want to start writing a state machine

 machine(L1Cache, "MI Example L1 Cache")
  : Sequencer * sequencer,
    CacheMemory * cacheMemory,
    int cache_response_latency = 12,
    int issue_latency = 2 {
      // Add rest of the stuff
    }


  • In order to let the controller receive messages from different entities in the system, the machine has a number of Message Buffers. These act as input and output ports for the machine. Here is an example specifying the output ports.
 MessageBuffer requestFromCache, network="To", virtual_network="2", ordered="true";
 MessageBuffer responseFromCache, network="To", virtual_network="4", ordered="true";

Note that Message Buffers have some attributes that need to be specified correctly. Another example, this time for specifying the input ports.

 MessageBuffer forwardToCache, network="From", virtual_network="3", ordered="true";
 MessageBuffer responseToCache, network="From", virtual_network="4", ordered="true";


  • Next the machine includes a declaration of the states that machine can possibly reach. In cache coherence protocol, states can be of two types -- stable and transient. A cache block is said to be in a stable state if in the absence of any activity (in coming request for the block from another controller, for example), the cache block would remain in that state for ever. Transient states are required for transitioning between stable states. They are needed when ever the transition between two stable states can not be done in an atomic fashion. Next is an example that shows how states are declared. SLICC has a keyword state_declaration that has to be used for declaring states.
 state_declaration(State, desc="Cache states") {
   I, AccessPermission:Invalid, desc="Not Present/Invalid";
   II, AccessPermission:Busy, desc="Not Present/Invalid, issued PUT";
   M, AccessPermission:Read_Write, desc="Modified";
   MI, AccessPermission:Busy, desc="Modified, issued PUT";
   MII, AccessPermission:Busy, desc="Modified, issued PUTX, received nack";
   IS, AccessPermission:Busy, desc="Issued request for LOAD/IFETCH";
   IM, AccessPermission:Busy, desc="Issued request for STORE/ATOMIC";
 }

The states I and M are the only stable states in this example. Again note that certain attributes have to be specified with the states.


  • The state machine needs to specify the events it can handle and thus transition from one state to another. SLICC provides the keyword enumeration which can be used for specifying the set of possible events. An example to shed more light on this -
 enumeration(Event, desc="Cache events") {
   // From processor
   Load,       desc="Load request from processor";
   Ifetch,     desc="Ifetch request from processor";
   Store,      desc="Store request from processor";
   Data,       desc="Data from network";
   Fwd_GETX,        desc="Forward from network";
   Inv,        desc="Invalidate request from dir";
   Replacement,  desc="Replace a block";
   Writeback_Ack,   desc="Ack from the directory for a writeback";
   Writeback_Nack,   desc="Nack from the directory for a writeback";
 }


  • While developing a protocol machine, we may need to define structures that represent different entities in a memory system. SLICC provides the keyword structure for this purpose. An example follows
 structure(Entry, desc="...", interface="AbstractCacheEntry") {
   State CacheState,        desc="cache state";
   bool Dirty,              desc="Is the data dirty (different than memory)?";
   DataBlock DataBlk,       desc="Data in the block";
 }

The cool thing about using SLICC's structure is that it automatically generates for you the get and set functions on different fields. It also writes a nice print function and overloads the << operator. But in case you would prefer do everything on your own, you can make use of the keyword external in the declaration of the structure. This would prevent SLICC from generating C++ code for this structure.

 structure(TBETable, external="yes") {
   TBE lookup(Address);
   void allocate(Address);
   void deallocate(Address);
   bool isPresent(Address);
 }

In fact many predefined types exist in src/mem/protocol/RubySlicc_*.sm files. You can make use of them, or if you need new types, you can define new ones as well. You can also use the keyword interface to make use of inheritance features available in C++. Note that currently SLICC supports public inheritance only.

  • We can also declare and define functions as we do in C++. There are certain functions that the compiler expects would always be defined by the controller. These include
    • getState()
    • setState()


Input for the Machine

Since protocol is state machine, we need to specify how to machine transitions from one state to another on receiving inputs. As mentioned before, each machine has several input and output ports. For each input port, the in_port keyword is used for specifying the behavior of the machine, when a message is received on that input port. An example follows that shows the syntax for declaring an input port.


 in_port(mandatoryQueue_in, RubyRequest, mandatoryQueue, desc="...") {
   if (mandatoryQueue_in.isReady()) {
     peek(mandatoryQueue_in, RubyRequest, block_on="LineAddress") {
       Entry cache_entry := getCacheEntry(in_msg.LineAddress);
       if (is_invalid(cache_entry) &&
           cacheMemory.cacheAvail(in_msg.LineAddress) == false ) {
         // make room for the block
         trigger(Event:Replacement, cacheMemory.cacheProbe(in_msg.LineAddress),
                 getCacheEntry(cacheMemory.cacheProbe(in_msg.LineAddress)),
                 TBEs[cacheMemory.cacheProbe(in_msg.LineAddress)]);
       }
       else {
         trigger(mandatory_request_type_to_event(in_msg.Type), in_msg.LineAddress,
                 cache_entry, TBEs[in_msg.LineAddress]);
       }
     }
   }
 }


  • As you can see, in_port takes in multiple arguments. The first argument, mandatoryQueue_in, is the identifier for the in_port that is used in the file. The next argument, RubyRequest, is the type of the messages that this input port receives. Each input port uses a queue to store the messages, the name of the queue is the third argument.


  • The keyword peek is used to extract messages from the queue of the input port. The use of this keyword implicitly declares a variable in_msg which is of the same type as specified in the input port's declaration. This variable points to the message at the head of the queue. It can be used for accessing the fields of the message as shown in the code above.


  • Once the incoming message has been analyzed, it is time for using this message for taking some appropriate action and changing the state of the machine. This done using the keyword trigger. The trigger function is actually used only in SLICC code and is not present in the generated code. Instead this call is converted in to a call to the doTransition() function which appears in the generated code. The doTransition() function is automatically generated by SLICC for each of the state machines. The number of arguments to trigger depend on the machine itself. In general, the input arguments for trigger are the type of the message that needs to processed, the address for which this message is meant for, the cache and the transaction buffer entries for that address.


  • trigger also increments a counter that is checked before a transition is made. In one ruby cycle, there is a limit on the number of transitions that can be carried out. This is done to resemble more closely to a hardware based state machine. @TODO: What happens if there are no more transitions left? Does the wakeup abort?

Actions

In this section we will go over how the actions that a state machine can carry out are defined. These actions will be called in to action when the state machine receives some input message which is then used to make a transition. Let's go over an example on how the key word action can be made use of.

 action(a_issueRequest, "a", desc="Issue a request") {
   enqueue(requestNetwork_out, RequestMsg, latency=issue_latency) {
   out_msg.Address := address;
     out_msg.Type := CoherenceRequestType:GETX;
     out_msg.Requestor := machineID;
     out_msg.Destination.add(map_Address_to_Directory(address));
     out_msg.MessageSize := MessageSizeType:Control;
   }
 }


  • The first input argument is the name of the action, the next argument is the abbreviation used for generating the documentation and last one is the description of the action which used in the HTML documentation and as a comment in the C++ code.
  • Each action is converted in to a C++ function of that name. The generated C++ code implicitly includes up to three input parameters in the function header, again depending on the machine. These arguments are the memory address on which the action is being taken, the cache and transaction buffer entries pertaining to this address.
  • Next useful thing to look at is the enqueue keyword. This keyword is used for queuing a message, generated as a result of the action, to an output port. The keyword takes three input arguments, namely, the name of the output port, the type of the message to be queued and the latency after which this message can be dequeued. Note that in case randomization is enabled, the specified latency is ignored. The use of the keyword implicitly declares a variable out_msg which is populated by the follow on statements.

Transitions

A transition function is a mapping from the cross product of set of states and set of events to the set of states. SLICC provides the keyword transition for specifying the transition function for state machines. An example follows --

 transition(IM, Data, M) {
   u_writeDataToCache;
   sx_store_hit;
   w_deallocateTBE;
   n_popResponseQueue;
 }

In this example, the initial state is IM. If an event of type Data occurs in that state, then final state would be M. Before making the transition, the state machine can perform certain actions on the structures that it maintains. In the given example, u_writeDataToCache is an action. All these operations are performed in an atomic fashion, i.e. no other event can occur before the set of actions specified with the transition has been completed.

  • For ease of use, sets of events and states can be provided as input to transition. The cross product of these sets will map to the same final state. Note that the final state cannot be a set. If for a particular event, the final state is same as the initial state, then the final state can be omitted.
 transition({IS, IM, MI, II}, {Load, Ifetch, Store, Replacement}) {
   z_stall;
 }


Special Functions

Stalling/Recycling/Waiting input ports

One of the more complicated internal features of SLICC and the resulting state machines is how the deal with the situation when events cannot be process due to the cache block being in a transient state. There are several possible ways to deal with this situation and each solution has different tradeoffs. This sub-section attempts to explain the differences. Please email the gem5-user list for further follow-up.

  • Stalling the input port. The simpliest way to handle events that can't be processed is to simply stall the input port. The correct way to do this is to include the "z_stall" action within the transition statement:
 transition({IS, IM, MI, II}, {Load, Ifetch, Store, Replacement}) {
   z_stall;
 }

Internally SLICC will return a ProtocolStall for this transition and no subsequent messages from the associated input port will be processed until the stalled message is processed. However, the other input ports will be analyzed for ready messages and processed in parallel. While this is a relatively simple solution, one may notice that stalling unrelated messages on the same input port will cause excessive and unecessary stalls.

One thing to note is Do Not leave the transition statement blank like so:

 transition({IS, IM, MI, II}, {Load, Ifetch, Store, Replacement}) {
   // stall the input port by simply not popping the message
 }

This will cause SLICC to return success for this transition and SLICC will continue to repeatly analyze the same input port. The result is eventual deadlock.

  • Recycling the input port. The better performat but more unrealistic solution is to recycle the stalled message on the input port. The way to do this is to use the "zz_recycleMandatoryQueue" action:
 action(zz_recycleMandatoryQueue, "\z", desc="Send the head of the mandatory queue to the back of the queue.") {
   mandatoryQueue_in.recycle();
 }
 transition({IS, IM, MI, II}, {Load, Ifetch, Store, Replacement}) {
   zz_recycleMandatoryQueue;
 }

The result of this action is that the transition returns a Protocol Stall and the offending message moved to the back of the FIFO input port. Therefore, other unrelated messages on the same input port can be processed. The problem with this solution is that recycled messages may be analyzed and reanalyzed every cycle until an address changes state.

  • Stall and wait the input port. An even better, but more complicated solution is to "stall and wait" the offending input messge. The way to do this is to use the "z_stallAndWaitMandatoryQueue" action:
 action(z_stallAndWaitMandatoryQueue, "\z", desc="recycle L1 request queue") {
   stall_and_wait(mandatoryQueue_in, address);
 }
 transition({IS, IM, IS_I, M_I, SM, SINK_WB_ACK}, {Load, Ifetch, Store, L1_Replacement}) {
   z_stallAndWaitMandatoryQueue;
 }

The result of this action is that the transition returns success, which is ok because stall_and_wait moves the offending message off the input port and to a side table associated with the input port. The message will not be analyzed again until it is woken up. In the meantime, other unrelated messages wil be processed.

The complicated part of stall and wait is that stalled messages must be explicitly woken up by other messages/transitions. In particular, transitions that move an address to a base state should wake up potentially stalled messages waiting for that address:

 action(kd_wakeUpDependents, "kd", desc="wake-up dependents") {
   wakeUpBuffers(address);
 }
 transition(M_I, WB_Ack, I) {
   s_deallocateTBE;
   o_popIncomingResponseQueue;
   kd_wakeUpDependents;
 }

Replacements are particularly complicated since stalled addresses are not associated with the same address they are actually waiting to change. In those situations all waiting messages must be woken up:

 action(ka_wakeUpAllDependents, "ka", desc="wake-up all dependents") {
   wakeUpAllBuffers();
 }
 transition(I, L2_Replacement) {
   rr_deallocateL2CacheBlock;
   ka_wakeUpAllDependents;
 }


Other Compiler Features

  • SLICC supports conditional statements in form of if and else. Note that SLICC does not support else if.
  • Each function has return type which can be void as well. Returned values cannot be ignored.
  • SLICC has limited support for pointer variables. is_valid() and is_invalid() operations are supported for testing whether a given pointer 'is not NULL' and 'is NULL' respectively. The keyword OOD, which stands for Out of Domain, plays the role of keyword NULL used in C++.
  • Static type casting is supported in SLICC. The keyword static_cast has been provided for this purpose. For example, in the following piece of code, a variable of type AbstractCacheEntry is being casted in to a variable of type Entry.
   Entry L1Dcache_entry := static_cast(Entry, "pointer", L1DcacheMemory[addr]);

SLICC Internals

C++ to Slicc Interface - @note: What do each of these files do/define???

  • src/mem/protocol/RubySlicc_interaces.sm
    • RubySlicc_Exports.sm
    • RubySlicc_Defines.sm
    • RubySlicc_Profiler.sm
    • RubySlicc_Types.sm
    • RubySlicc_MemControl.sm
    • RubySlicc_ComponentMapping.sm

Variable Assignments

  • Use the := operator to assign members in class (e.g. a member defined in RubySlicc_Types.sm):
    • an automatic "m_" is added to the name mentioned in the SLICC file.

See Also

GEMS-gem5 SLICC Transition Guide