The LLVM Pass Framework is an important part of the LLVM system, because LLVM passes are where most of the interesting parts of the compiler exist. Passes perform the transformations and optimizations that make up the compiler, they build the analysis results that are used by these transformations, and they are, above all, a structuring technique for compiler code.

All LLVM passes are subclasses of the Pass class, which implement functionality by overriding virtual methods inherited from Pass. Depending on how your pass works, you should inherit from the ModulePass, CallGraphSCCPass, FunctionPass, or LoopPass, or BasicBlockPass classes, which gives the system more information about what your pass does, and how it can be combined with other passes. One of the main features of the LLVM Pass Framework is that it schedules passes to run in an efficient way based on the constraints that your pass meets (which are indicated by which class they derive from).

We start by showing you how to construct a pass, everything from setting up the code, to compiling, loading, and executing it. After the basics are down, more advanced features are discussed.

Here we describe how to write the "hello world" of passes. The "Hello" pass is designed to simply print out the name of non-external functions that exist in the program being compiled. It does not modify the program at all, it just inspects it. The source code and files for this pass are available in the LLVM source tree in the lib/Transforms/Hello directory.

First, you need to create a new directory somewhere in the LLVM source base. For this example, we'll assume that you made lib/Transforms/Hello. Next, you must set up a build script (Makefile) that will compile the source code for the new pass. To do this, copy the following into Makefile:

# Makefile for hello pass

# Path to top level of LLVM heirarchy
LEVEL = ../../..

# Name of the library to build
LIBRARYNAME = Hello

# Make the shared library become a loadable module so the tools can 
# dlopen/dlsym on the resulting library.
LOADABLE_MODULE = 1

# Tell the build system which LLVM libraries your pass needs. You'll probably
# need at least LLVMSystem.a, LLVMSupport.a, LLVMCore.a but possibly several
# others too.
LLVMLIBS = LLVMCore.a LLVMSupport.a LLVMSystem.a

# Include the makefile implementation stuff
include $(LEVEL)/Makefile.common

This makefile specifies that all of the .cpp files in the current directory are to be compiled and linked together into a Debug/lib/Hello.so shared object that can be dynamically loaded by the opt or bugpoint tools via their -load options. If your operating system uses a suffix other than .so (such as windows or Mac OS/X), the appropriate extension will be used.

Now that we have the build scripts set up, we just need to write the code for the pass itself.

Now that we have a way to compile our new pass, we just have to write it. Start out with:

#include "llvm/Pass.h"
#include "llvm/Function.h"

Which are needed because we are writing a Pass, and we are operating on Function's.

Next we have:

using namespace llvm;

... which is required because the functions from the include files live in the llvm namespace.

Next we have:

namespace {

... which starts out an anonymous namespace. Anonymous namespaces are to C++ what the "static" keyword is to C (at global scope). It makes the things declared inside of the anonymous namespace only visible to the current file. If you're not familiar with them, consult a decent C++ book for more information.

Next, we declare our pass itself:

  struct Hello : public FunctionPass {

This declares a "Hello" class that is a subclass of FunctionPass. The different builtin pass subclasses are described in detail later, but for now, know that FunctionPass's operate a function at a time.

     static const char ID;
     Hello() : FunctionPass((intptr_t)&ID) {}

This declares pass identifier used by LLVM to identify pass. This allows LLVM to avoid using expensive C++ runtime information.

    virtual bool runOnFunction(Function &F) {
      llvm::cerr << "Hello: " << F.getName() << "\n";
      return false;
    }
  };  // end of struct Hello

We declare a "runOnFunction" method, which overloads an abstract virtual method inherited from FunctionPass. This is where we are supposed to do our thing, so we just print out our message with the name of each function.

  const char Hello::ID = 0;

We initialize pass ID here. LLVM uses ID's address to identify pass so initialization value is not important.

  RegisterPass<Hello> X("hello", "Hello World Pass");
}  // end of anonymous namespace

Lastly, we register our class Hello, giving it a command line argument "hello", and a name "Hello World Pass".

As a whole, the .cpp file looks like:

#include "llvm/Pass.h"
#include "llvm/Function.h"

using namespace llvm;

namespace {
  struct Hello : public FunctionPass {
    
    static const char ID;
    Hello() : FunctionPass((intptr_t)&ID) {}

    virtual bool runOnFunction(Function &F) {
      llvm::cerr << "Hello: " << F.getName() << "\n";
      return false;
    }
  };
  
  RegisterPass<Hello> X("hello", "Hello World Pass");
}

Now that it's all together, compile the file with a simple "gmake" command in the local directory and you should get a new "Debug/lib/Hello.so file. Note that everything in this file is contained in an anonymous namespace: this reflects the fact that passes are self contained units that do not need external interfaces (although they can have them) to be useful.

Now that you have a brand new shiny shared object file, we can use the opt command to run an LLVM program through your pass. Because you registered your pass with the RegisterPass template, you will be able to use the opt tool to access it, once loaded.

To test it, follow the example at the end of the Getting Started Guide to compile "Hello World" to LLVM. We can now run the bytecode file (hello.bc) for the program through our transformation like this (or course, any bytecode file will work):

$ opt -load ../../../Debug/lib/Hello.so -hello < hello.bc > /dev/null
Hello: __main
Hello: puts
Hello: main

The '-load' option specifies that 'opt' should load your pass as a shared object, which makes '-hello' a valid command line argument (which is one reason you need to register your pass). Because the hello pass does not modify the program in any interesting way, we just throw away the result of opt (sending it to /dev/null).

To see what happened to the other string you registered, try running opt with the --help option:

$ opt -load ../../../Debug/lib/Hello.so --help
OVERVIEW: llvm .bc -> .bc modular optimizer

USAGE: opt [options] <input bytecode>

OPTIONS:
  Optimizations available:
...
    -funcresolve    - Resolve Functions
    -gcse           - Global Common Subexpression Elimination
    -globaldce      - Dead Global Elimination
    -hello          - Hello World Pass
    -indvars        - Canonicalize Induction Variables
    -inline         - Function Integration/Inlining
    -instcombine    - Combine redundant instructions
...

The pass name get added as the information string for your pass, giving some documentation to users of opt. Now that you have a working pass, you would go ahead and make it do the cool transformations you want. Once you get it all working and tested, it may become useful to find out how fast your pass is. The PassManager provides a nice command line option (--time-passes) that allows you to get information about the execution time of your pass along with the other passes you queue up. For example:

$ opt -load ../../../Debug/lib/Hello.so -hello -time-passes < hello.bc > /dev/null
Hello: __main
Hello: puts
Hello: main
===============================================================================
                      ... Pass execution timing report ...
===============================================================================
  Total Execution Time: 0.02 seconds (0.0479059 wall clock)

   ---User Time---   --System Time--   --User+System--   ---Wall Time---  --- Pass Name ---
   0.0100 (100.0%)   0.0000 (  0.0%)   0.0100 ( 50.0%)   0.0402 ( 84.0%)  Bytecode Writer
   0.0000 (  0.0%)   0.0100 (100.0%)   0.0100 ( 50.0%)   0.0031 (  6.4%)  Dominator Set Construction
   0.0000 (  0.0%)   0.0000 (  0.0%)   0.0000 (  0.0%)   0.0013 (  2.7%)  Module Verifier
   0.0000 (  0.0%)   0.0000 (  0.0%)   0.0000 (  0.0%)   0.0033 (  6.9%)  Hello World Pass
   0.0100 (100.0%)   0.0100 (100.0%)   0.0200 (100.0%)   0.0479 (100.0%)  TOTAL

As you can see, our implementation above is pretty fast :). The additional passes listed are automatically inserted by the 'opt' tool to verify that the LLVM emitted by your pass is still valid and well formed LLVM, which hasn't been broken somehow.

Now that you have seen the basics of the mechanics behind passes, we can talk about some more details of how they work and how to use them.

One of the first things that you should do when designing a new pass is to decide what class you should subclass for your pass. The Hello World example uses the FunctionPass class for its implementation, but we did not discuss why or when this should occur. Here we talk about the classes available, from the most general to the most specific.

When choosing a superclass for your Pass, you should choose the most specific class possible, while still being able to meet the requirements listed. This gives the LLVM Pass Infrastructure information necessary to optimize how passes are run, so that the resultant compiler isn't unneccesarily slow.

The most plain and boring type of pass is the "ImmutablePass" class. This pass type is used for passes that do not have to be run, do not change state, and never need to be updated. This is not a normal type of transformation or analysis, but can provide information about the current compiler configuration.

Although this pass class is very infrequently used, it is important for providing information about the current target machine being compiled for, and other static information that can affect the various transformations.

ImmutablePasses never invalidate other transformations, are never invalidated, and are never "run".

The "ModulePass" class is the most general of all superclasses that you can use. Deriving from ModulePass indicates that your pass uses the entire program as a unit, refering to function bodies in no predictable order, or adding and removing functions. Because nothing is known about the behavior of ModulePass subclasses, no optimization can be done for their execution. A module pass can use function level passes (e.g. dominators) using getAnalysis interface getAnalysis(Function).

To write a correct ModulePass subclass, derive from ModulePass and overload the runOnModule method with the following signature:

  virtual bool runOnModule(Module &M) = 0;

The runOnModule method performs the interesting work of the pass. It should return true if the module was modified by the transformation and false otherwise.

The "CallGraphSCCPass" is used by passes that need to traverse the program bottom-up on the call graph (callees before callers). Deriving from CallGraphSCCPass provides some mechanics for building and traversing the CallGraph, but also allows the system to optimize execution of CallGraphSCCPass's. If your pass meets the requirements outlined below, and doesn't meet the requirements of a FunctionPass or BasicBlockPass, you should derive from CallGraphSCCPass.

TODO: explain briefly what SCC, Tarjan's algo, and B-U mean.

To be explicit, CallGraphSCCPass subclasses are:

1. ... not allowed to modify any Functions that are not in the current SCC.
2. ... allowed to inspect any Function's other than those in the current SCC and the direct callees of the SCC.
3. ... required to preserve the current CallGraph object, updating it to reflect any changes made to the program.
4. ... not allowed to add or remove SCC's from the current Module, though they may change the contents of an SCC.
5. ... allowed to add or remove global variables from the current Module.
6. ... allowed to maintain state across invocations of runOnSCC (including global data).

Implementing a CallGraphSCCPass is slightly tricky in some cases because it has to handle SCCs with more than one node in it. All of the virtual methods described below should return true if they modified the program, or false if they didn't.

  virtual bool doInitialization(CallGraph &CG);

The doIninitialize method is allowed to do most of the things that CallGraphSCCPass's are not allowed to do. They can add and remove functions, get pointers to functions, etc. The doInitialization method is designed to do simple initialization type of stuff that does not depend on the SCCs being processed. The doInitialization method call is not scheduled to overlap with any other pass executions (thus it should be very fast).

  virtual bool runOnSCC(const std::vector<CallGraphNode *> &SCCM) = 0;

The runOnSCC method performs the interesting work of the pass, and should return true if the module was modified by the transformation, false otherwise.

  virtual bool doFinalization(CallGraph &CG);

The doFinalization method is an infrequently used method that is called when the pass framework has finished calling runOnFunction for every function in the program being compiled.

In contrast to ModulePass subclasses, FunctionPass subclasses do have a predictable, local behavior that can be expected by the system. All FunctionPass execute on each function in the program independent of all of the other functions in the program. FunctionPass's do not require that they are executed in a particular order, and FunctionPass's do not modify external functions.

To be explicit, FunctionPass subclasses are not allowed to:

1. Modify a Function other than the one currently being processed.
2. Add or remove Function's from the current Module.
3. Add or remove global variables from the current Module.
4. Maintain state across invocations of runOnFunction (including global data)

Implementing a FunctionPass is usually straightforward (See the Hello World pass for example). FunctionPass's may overload three virtual methods to do their work. All of these methods should return true if they modified the program, or false if they didn't.

  virtual bool doInitialization(Module &M);

The doIninitialize method is allowed to do most of the things that FunctionPass's are not allowed to do. They can add and remove functions, get pointers to functions, etc. The doInitialization method is designed to do simple initialization type of stuff that does not depend on the functions being processed. The doInitialization method call is not scheduled to overlap with any other pass executions (thus it should be very fast).

A good example of how this method should be used is the LowerAllocations pass. This pass converts malloc and free instructions into platform dependent malloc() and free() function calls. It uses the doInitialization method to get a reference to the malloc and free functions that it needs, adding prototypes to the module if necessary.

  virtual bool runOnFunction(Function &F) = 0;

The runOnFunction method must be implemented by your subclass to do the transformation or analysis work of your pass. As usual, a true value should be returned if the function is modified.

  virtual bool doFinalization(Module &M);

The doFinalization method is an infrequently used method that is called when the pass framework has finished calling runOnFunction for every function in the program being compiled.

All LoopPass execute on each loop in the function independent of all of the other loops in the function. LoopPass processes loops in loop nest order such that outer most loop is processed last.

LoopPass subclasses are allowed to update loop nest using LPPassManager interface. Implementing a loop pass is usually straightforward. Looppass's may overload three virtual methods to do their work. All these methods should return true if they modified the program, or false if they didn't.

  virtual bool runOnLoop(Loop *, LPPassManager &LPM) = 0;

The runOnLoop method must be implemented by your subclass to do the transformation or analysis work of your pass. As usual, a true value should be returned if the function is modified. LPPassManager interface should be used to update loop nest.

  virtual bool doFinalization();

The doFinalization method is an infrequently used method that is called when the pass framework has finished calling runOnLoop for every loop in the program being compiled.

BasicBlockPass's are just like FunctionPass's, except that they must limit their scope of inspection and modification to a single basic block at a time. As such, they are not allowed to do any of the following:

1. Modify or inspect any basic blocks outside of the current one
2. Maintain state across invocations of runOnBasicBlock
3. Modify the control flow graph (by altering terminator instructions)
4. Any of the things forbidden for FunctionPasses.

BasicBlockPasses are useful for traditional local and "peephole" optimizations. They may override the same doInitialization(Module &) and doFinalization(Module &) methods that FunctionPass's have, but also have the following virtual methods that may also be implemented:

  virtual bool doInitialization(Function &F);

The doIninitialize method is allowed to do most of the things that BasicBlockPass's are not allowed to do, but that FunctionPass's can. The doInitialization method is designed to do simple initialization that does not depend on the BasicBlocks being processed. The doInitialization method call is not scheduled to overlap with any other pass executions (thus it should be very fast).

  virtual bool runOnBasicBlock(BasicBlock &BB) = 0;

Override this function to do the work of the BasicBlockPass. This function is not allowed to inspect or modify basic blocks other than the parameter, and are not allowed to modify the CFG. A true value must be returned if the basic block is modified.

  virtual bool doFinalization(Function &F);

The doFinalization method is an infrequently used method that is called when the pass framework has finished calling runOnBasicBlock for every BasicBlock in the program being compiled. This can be used to perform per-function finalization.

A MachineFunctionPass is a part of the LLVM code generator that executes on the machine-dependent representation of each LLVM function in the program. A MachineFunctionPass is also a FunctionPass, so all the restrictions that apply to a FunctionPass also apply to it. MachineFunctionPasses also have additional restrictions. In particular, MachineFunctionPasses are not allowed to do any of the following:

1. Modify any LLVM Instructions, BasicBlocks or Functions.
2. Modify a MachineFunction other than the one currently being processed.
3. Add or remove MachineFunctions from the current Module.
4. Add or remove global variables from the current Module.
5. Maintain state across invocations of runOnMachineFunction (including global data)

  virtual bool runOnMachineFunction(MachineFunction &MF) = 0;

runOnMachineFunction can be considered the main entry point of a MachineFunctionPass; that is, you should override this method to do the work of your MachineFunctionPass.

The runOnMachineFunction method is called on every MachineFunction in a Module, so that the MachineFunctionPass may perform optimizations on the machine-dependent representation of the function. If you want to get at the LLVM Function for the MachineFunction you're working on, use MachineFunction's getFunction() accessor method -- but remember, you may not modify the LLVM Function or its contents from a MachineFunctionPass.

In the Hello World example pass we illustrated how pass registration works, and discussed some of the reasons that it is used and what it does. Here we discuss how and why passes are registered.

As we saw above, passes are registered with the RegisterPass template, which requires you to pass at least two parameters. The first parameter is the name of the pass that is to be used on the command line to specify that the pass should be added to a program (for example, with opt or bugpoint). The second argument is the name of the pass, which is to be used for the --help output of programs, as well as for debug output generated by the --debug-pass option.

If you want your pass to be easily dumpable, you should implement the virtual print method:

  virtual void print(llvm::OStream &O, const Module *M) const;

The print method must be implemented by "analyses" in order to print a human readable version of the analysis results. This is useful for debugging an analysis itself, as well as for other people to figure out how an analysis works. Use the opt -analyze argument to invoke this method.

The llvm::OStream parameter specifies the stream to write the results on, and the Module parameter gives a pointer to the top level module of the program that has been analyzed. Note however that this pointer may be null in certain circumstances (such as calling the Pass::dump() from a debugger), so it should only be used to enhance debug output, it should not be depended on.

One of the main responsibilities of the PassManager is the make sure that passes interact with each other correctly. Because PassManager tries to optimize the execution of passes it must know how the passes interact with each other and what dependencies exist between the various passes. To track this, each pass can declare the set of passes that are required to be executed before the current pass, and the passes which are invalidated by the current pass.

Typically this functionality is used to require that analysis results are computed before your pass is run. Running arbitrary transformation passes can invalidate the computed analysis results, which is what the invalidation set specifies. If a pass does not implement the getAnalysisUsage method, it defaults to not having any prerequisite passes, and invalidating all other passes.

  virtual void getAnalysisUsage(AnalysisUsage &Info) const;

By implementing the getAnalysisUsage method, the required and invalidated sets may be specified for your transformation. The implementation should fill in the AnalysisUsage object with information about which passes are required and not invalidated. To do this, a pass may call any of the following methods on the AnalysisUsage object:

If your pass requires a previous pass to be executed (an analysis for example), it can use one of these methods to arrange for it to be run before your pass. LLVM has many different types of analyses and passes that can be required, spanning the range from DominatorSet to BreakCriticalEdges. Requiring BreakCriticalEdges, for example, guarantees that there will be no critical edges in the CFG when your pass has been run.

Some analyses chain to other analyses to do their job. For example, an AliasAnalysis implementation is required to chain to other alias analysis passes. In cases where analyses chain, the addRequiredTransitive method should be used instead of the addRequired method. This informs the PassManager that the transitively required pass should be alive as long as the requiring pass is.

One of the jobs of the PassManager is to optimize how and when analyses are run. In particular, it attempts to avoid recomputing data unless it needs to. For this reason, passes are allowed to declare that they preserve (i.e., they don't invalidate) an existing analysis if it's available. For example, a simple constant folding pass would not modify the CFG, so it can't possibly affect the results of dominator analysis. By default, all passes are assumed to invalidate all others.

The AnalysisUsage class provides several methods which are useful in certain circumstances that are related to addPreserved. In particular, the setPreservesAll method can be called to indicate that the pass does not modify the LLVM program at all (which is true for analyses), and the setPreservesCFG method can be used by transformations that change instructions in the program but do not modify the CFG or terminator instructions (note that this property is implicitly set for BasicBlockPass's).

addPreserved is particularly useful for transformations like BreakCriticalEdges. This pass knows how to update a small set of loop and dominator related analyses if they exist, so it can preserve them, despite the fact that it hacks on the CFG.

  // This is an example implementation from an analysis, which does not modify
  // the program at all, yet has a prerequisite.
  void PostDominanceFrontier::getAnalysisUsage(AnalysisUsage &AU) const {
    AU.setPreservesAll();
    AU.addRequired<PostDominatorTree>();
  }

and:

  // This example modifies the program, but does not modify the CFG
  void LICM::getAnalysisUsage(AnalysisUsage &AU) const {
    AU.setPreservesCFG();
    AU.addRequired<LoopInfo>();
  }

The Pass::getAnalysis<> method is automatically inherited by your class, providing you with access to the passes that you declared that you required with the getAnalysisUsage method. It takes a single template argument that specifies which pass class you want, and returns a reference to that pass. For example:

   bool LICM::runOnFunction(Function &F) {
     LoopInfo &LI = getAnalysis<LoopInfo>();
     ...
   }

This method call returns a reference to the pass desired. You may get a runtime assertion failure if you attempt to get an analysis that you did not declare as required in your getAnalysisUsage implementation. This method can be called by your run* method implementation, or by any other local method invoked by your run* method. A module level pass can use function level analysis info using this interface. For example:

   bool ModuleLevelPass::runOnModule(Module &M) {
     ...
     DominatorTree &DT = getAnalysis<DominatorTree>(Func);
     ...
   }

If your pass is capable of updating analyses if they exist (e.g., BreakCriticalEdges, as described above), you can use the getAnalysisToUpdate method, which returns a pointer to the analysis if it is active. For example:

  ...
  if (DominatorSet *DS = getAnalysisToUpdate<DominatorSet>()) {
    // A DominatorSet is active.  This code will update it.
  }
  ...

Now that we understand the basics of how passes are defined, how the are used, and how they are required from other passes, it's time to get a little bit fancier. All of the pass relationships that we have seen so far are very simple: one pass depends on one other specific pass to be run before it can run. For many applications, this is great, for others, more flexibility is required.

In particular, some analyses are defined such that there is a single simple interface to the analysis results, but multiple ways of calculating them. Consider alias analysis for example. The most trivial alias analysis returns "may alias" for any alias query. The most sophisticated analysis a flow-sensitive, context-sensitive interprocedural analysis that can take a significant amount of time to execute (and obviously, there is a lot of room between these two extremes for other implementations). To cleanly support situations like this, the LLVM Pass Infrastructure supports the notion of Analysis Groups.

An Analysis Group is a single simple interface that may be implemented by multiple different passes. Analysis Groups can be given human readable names just like passes, but unlike passes, they need not derive from the Pass class. An analysis group may have one or more implementations, one of which is the "default" implementation.

Analysis groups are used by client passes just like other passes are: the AnalysisUsage::addRequired() and Pass::getAnalysis() methods. In order to resolve this requirement, the PassManager scans the available passes to see if any implementations of the analysis group are available. If none is available, the default implementation is created for the pass to use. All standard rules for interaction between passes still apply.

Although Pass Registration is optional for normal passes, all analysis group implementations must be registered, and must use the RegisterAnalysisGroup template to join the implementation pool. Also, a default implementation of the interface must be registered with RegisterAnalysisGroup.

As a concrete example of an Analysis Group in action, consider the AliasAnalysis analysis group. The default implementation of the alias analysis interface (the basicaa pass) just does a few simple checks that don't require significant analysis to compute (such as: two different globals can never alias each other, etc). Passes that use the AliasAnalysis interface (for example the gcse pass), do not care which implementation of alias analysis is actually provided, they just use the designated interface.

From the user's perspective, commands work just like normal. Issuing the command 'opt -gcse ...' will cause the basicaa class to be instantiated and added to the pass sequence. Issuing the command 'opt -somefancyaa -gcse ...' will cause the gcse pass to use the somefancyaa alias analysis (which doesn't actually exist, it's just a hypothetical example) instead.

The RegisterAnalysisGroup template is used to register the analysis group itself as well as add pass implementations to the analysis group. First, an analysis should be registered, with a human readable name provided for it. Unlike registration of passes, there is no command line argument to be specified for the Analysis Group Interface itself, because it is "abstract":

  static RegisterAnalysisGroup<AliasAnalysis> A("Alias Analysis");

Once the analysis is registered, passes can declare that they are valid implementations of the interface by using the following code:

namespace {
  // Analysis Group implementations must be registered normally...
  RegisterPass<FancyAA>
  B("somefancyaa", "A more complex alias analysis implementation");

  // Declare that we implement the AliasAnalysis interface
  RegisterAnalysisGroup<AliasAnalysis> C(B);
}

This just shows a class FancyAA that is registered normally, then uses the RegisterAnalysisGroup template to "join" the AliasAnalysis analysis group. Every implementation of an analysis group should join using this template. A single pass may join multiple different analysis groups with no problem.

namespace {
  // Analysis Group implementations must be registered normally...
  RegisterPass<BasicAliasAnalysis>
  D("basicaa", "Basic Alias Analysis (default AA impl)");

  // Declare that we implement the AliasAnalysis interface
  RegisterAnalysisGroup<AliasAnalysis, true> E(D);
}

Here we show how the default implementation is specified (using the extra argument to the RegisterAnalysisGroup template). There must be exactly one default implementation available at all times for an Analysis Group to be used. Here we declare that the BasicAliasAnalysis pass is the default implementation for the interface.

The Statistic class is designed to be an easy way to expose various success metrics from passes. These statistics are printed at the end of a run, when the -stats command line option is enabled on the command line. See the Statistics section in the Programmer's Manual for details.

The PassManager class takes a list of passes, ensures their prerequisites are set up correctly, and then schedules passes to run efficiently. All of the LLVM tools that run passes use the PassManager for execution of these passes.

The PassManager does two main things to try to reduce the execution time of a series of passes:

1. Share analysis results - The PassManager attempts to avoid recomputing analysis results as much as possible. This means keeping track of which analyses are available already, which analyses get invalidated, and which analyses are needed to be run for a pass. An important part of work is that the PassManager tracks the exact lifetime of all analysis results, allowing it to free memory allocated to holding analysis results as soon as they are no longer needed.
2. Pipeline the execution of passes on the program - The PassManager attempts to get better cache and memory usage behavior out of a series of passes by pipelining the passes together. This means that, given a series of consequtive FunctionPass's, it will execute all of the FunctionPass's on the first function, then all of the FunctionPasses on the second function, etc... until the entire program has been run through the passes.

   This improves the cache behavior of the compiler, because it is only touching the LLVM program representation for a single function at a time, instead of traversing the entire program. It reduces the memory consumption of compiler, because, for example, only one DominatorSet needs to be calculated at a time. This also makes it possible some interesting enhancements in the future.

The effectiveness of the PassManager is influenced directly by how much information it has about the behaviors of the passes it is scheduling. For example, the "preserved" set is intentionally conservative in the face of an unimplemented getAnalysisUsage method. Not implementing when it should be implemented will have the effect of not allowing any analysis results to live across the execution of your pass.

The PassManager class exposes a --debug-pass command line options that is useful for debugging pass execution, seeing how things work, and diagnosing when you should be preserving more analyses than you currently are (To get information about all of the variants of the --debug-pass option, just type 'opt --help-hidden').

By using the --debug-pass=Structure option, for example, we can see how our Hello World pass interacts with other passes. Lets try it out with the gcse and licm passes:

$ opt -load ../../../Debug/lib/Hello.so -gcse -licm --debug-pass=Structure < hello.bc > /dev/null
Module Pass Manager
  Function Pass Manager
    Dominator Set Construction
    Immediate Dominators Construction
    Global Common Subexpression Elimination
--  Immediate Dominators Construction
--  Global Common Subexpression Elimination
    Natural Loop Construction
    Loop Invariant Code Motion
--  Natural Loop Construction
--  Loop Invariant Code Motion
    Module Verifier
--  Dominator Set Construction
--  Module Verifier
  Bytecode Writer
--Bytecode Writer

This output shows us when passes are constructed and when the analysis results are known to be dead (prefixed with '--'). Here we see that GCSE uses dominator and immediate dominator information to do its job. The LICM pass uses natural loop information, which uses dominator sets, but not immediate dominators. Because immediate dominators are no longer useful after the GCSE pass, it is immediately destroyed. The dominator sets are then reused to compute natural loop information, which is then used by the LICM pass.

After the LICM pass, the module verifier runs (which is automatically added by the 'opt' tool), which uses the dominator set to check that the resultant LLVM code is well formed. After it finishes, the dominator set information is destroyed, after being computed once, and shared by three passes.

Lets see how this changes when we run the Hello World pass in between the two passes:

$ opt -load ../../../Debug/lib/Hello.so -gcse -hello -licm --debug-pass=Structure < hello.bc > /dev/null
Module Pass Manager
  Function Pass Manager
    Dominator Set Construction
    Immediate Dominators Construction
    Global Common Subexpression Elimination
--  Dominator Set Construction
--  Immediate Dominators Construction
--  Global Common Subexpression Elimination
    Hello World Pass
--  Hello World Pass
    Dominator Set Construction
    Natural Loop Construction
    Loop Invariant Code Motion
--  Natural Loop Construction
--  Loop Invariant Code Motion
    Module Verifier
--  Dominator Set Construction
--  Module Verifier
  Bytecode Writer
--Bytecode Writer
Hello: __main
Hello: puts
Hello: main

Here we see that the Hello World pass has killed the Dominator Set pass, even though it doesn't modify the code at all! To fix this, we need to add the following getAnalysisUsage method to our pass:

    // We don't modify the program, so we preserve all analyses
    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
      AU.setPreservesAll();
    }

Now when we run our pass, we get this output:

$ opt -load ../../../Debug/lib/Hello.so -gcse -hello -licm --debug-pass=Structure < hello.bc > /dev/null
Pass Arguments:  -gcse -hello -licm
Module Pass Manager
  Function Pass Manager
    Dominator Set Construction
    Immediate Dominators Construction
    Global Common Subexpression Elimination
--  Immediate Dominators Construction
--  Global Common Subexpression Elimination
    Hello World Pass
--  Hello World Pass
    Natural Loop Construction
    Loop Invariant Code Motion
--  Loop Invariant Code Motion
--  Natural Loop Construction
    Module Verifier
--  Dominator Set Construction
--  Module Verifier
  Bytecode Writer
--Bytecode Writer
Hello: __main
Hello: puts
Hello: main

Which shows that we don't accidentally invalidate dominator information anymore, and therefore do not have to compute it twice.

  virtual void releaseMemory();

The PassManager automatically determines when to compute analysis results, and how long to keep them around for. Because the lifetime of the pass object itself is effectively the entire duration of the compilation process, we need some way to free analysis results when they are no longer useful. The releaseMemory virtual method is the way to do this.

If you are writing an analysis or any other pass that retains a significant amount of state (for use by another pass which "requires" your pass and uses the getAnalysis method) you should implement releaseMEmory to, well, release the memory allocated to maintain this internal state. This method is called after the run* method for the class, before the next call of run* in your pass.

Size matters when constructing production quality tools using llvm, both for the purposes of distribution, and for regulating the resident code size when running on the target system. Therefore, it becomes desirable to selectively use some passes, while omitting others and maintain the flexibility to change configurations later on. You want to be able to do all this, and, provide feedback to the user. This is where pass registration comes into play.

The fundamental mechanisms for pass registration are the MachinePassRegistry class and subclasses of MachinePassRegistryNode.

An instance of MachinePassRegistry is used to maintain a list of MachinePassRegistryNode objects. This instance maintains the list and communicates additions and deletions to the command line interface.

An instance of MachinePassRegistryNode subclass is used to maintain information provided about a particular pass. This information includes the command line name, the command help string and the address of the function used to create an instance of the pass. A global static constructor of one of these instances registers with a corresponding MachinePassRegistry, the static destructor unregisters. Thus a pass that is statically linked in the tool will be registered at start up. A dynamically loaded pass will register on load and unregister at unload.

There are predefined registries to track instruction scheduling (RegisterScheduler) and register allocation (RegisterRegAlloc) machine passes. Here we will describe how to register a register allocator machine pass.

Implement your register allocator machine pass. In your register allocator .cpp file add the following include;

  #include "llvm/CodeGen/RegAllocRegistry.h"

Also in your register allocator .cpp file, define a creator function in the form;

  FunctionPass *createMyRegisterAllocator() {
    return new MyRegisterAllocator();
  }

Note that the signature of this function should match the type of RegisterRegAlloc::FunctionPassCtor. In the same file add the "installing" declaration, in the form;

  static RegisterRegAlloc myRegAlloc("myregalloc",
    "  my register allocator help string",
    createMyRegisterAllocator);

Note the two spaces prior to the help string produces a tidy result on the --help query.

$ llc --help
  ...
  -regalloc                    - Register allocator to use: (default = linearscan)
    =linearscan                -   linear scan register allocator
    =local                     -   local register allocator
    =simple                    -   simple register allocator
    =myregalloc                -   my register allocator help string
  ...

And that's it. The user is now free to use -regalloc=myregalloc as an option. Registering instruction schedulers is similar except use the RegisterScheduler class. Note that the RegisterScheduler::FunctionPassCtor is significantly different from RegisterRegAlloc::FunctionPassCtor.

To force the load/linking of your register allocator into the llc/lli tools, add your creator function's global declaration to "Passes.h" and add a "pseudo" call line to llvm/Codegen/LinkAllCodegenComponents.h.

The easiest way to get started is to clone one of the existing registries; we recommend llvm/CodeGen/RegAllocRegistry.h. The key things to modify are the class name and the FunctionPassCtor type.

Then you need to declare the registry. Example: if your pass registry is RegisterMyPasses then define;

MachinePassRegistry RegisterMyPasses::Registry;

And finally, declare the command line option for your passes. Example:

  cl::opt<RegisterMyPasses::FunctionPassCtor, false,
          RegisterPassParser<RegisterMyPasses> >
  MyPassOpt("mypass",
            cl::init(&createDefaultMyPass),
            cl::desc("my pass option help")); 

Here the command option is "mypass", with createDefaultMyPass as the default creator.

Unfortunately, using GDB with dynamically loaded passes is not as easy as it should be. First of all, you can't set a breakpoint in a shared object that has not been loaded yet, and second of all there are problems with inlined functions in shared objects. Here are some suggestions to debugging your pass with GDB.

For sake of discussion, I'm going to assume that you are debugging a transformation invoked by opt, although nothing described here depends on that.

First thing you do is start gdb on the opt process:

$ gdb opt
GNU gdb 5.0
Copyright 2000 Free Software Foundation, Inc.
GDB is free software, covered by the GNU General Public License, and you are
welcome to change it and/or distribute copies of it under certain conditions.
Type "show copying" to see the conditions.
There is absolutely no warranty for GDB.  Type "show warranty" for details.
This GDB was configured as "sparc-sun-solaris2.6"...
(gdb)

Note that opt has a lot of debugging information in it, so it takes time to load. Be patient. Since we cannot set a breakpoint in our pass yet (the shared object isn't loaded until runtime), we must execute the process, and have it stop before it invokes our pass, but after it has loaded the shared object. The most foolproof way of doing this is to set a breakpoint in PassManager::run and then run the process with the arguments you want:

(gdb) break llvm::PassManager::run
Breakpoint 1 at 0x2413bc: file Pass.cpp, line 70.
(gdb) run test.bc -load $(LLVMTOP)/llvm/Debug/lib/[libname].so -[passoption]
Starting program: opt test.bc -load $(LLVMTOP)/llvm/Debug/lib/[libname].so -[passoption]
Breakpoint 1, PassManager::run (this=0xffbef174, M=@0x70b298) at Pass.cpp:70
70      bool PassManager::run(Module &M) { return PM->run(M); }
(gdb)

Once the opt stops in the PassManager::run method you are now free to set breakpoints in your pass so that you can trace through execution or do other standard debugging stuff.

Once you have the basics down, there are a couple of problems that GDB has, some with solutions, some without.

• Inline functions have bogus stack information. In general, GDB does a pretty good job getting stack traces and stepping through inline functions. When a pass is dynamically loaded however, it somehow completely loses this capability. The only solution I know of is to de-inline a function (move it from the body of a class to a .cpp file).
• Restarting the program breaks breakpoints. After following the information above, you have succeeded in getting some breakpoints planted in your pass. Nex thing you know, you restart the program (i.e., you type 'run' again), and you start getting errors about breakpoints being unsettable. The only way I have found to "fix" this problem is to delete the breakpoints that are already set in your pass, run the program, and re-set the breakpoints once execution stops in PassManager::run.

Hopefully these tips will help with common case debugging situations. If you'd like to contribute some tips of your own, just contact Chris.

Although the LLVM Pass Infrastructure is very capable as it stands, and does some nifty stuff, there are things we'd like to add in the future. Here is where we are going:

Multiple CPU machines are becoming more common and compilation can never be fast enough: obviously we should allow for a multithreaded compiler. Because of the semantics defined for passes above (specifically they cannot maintain state across invocations of their run* methods), a nice clean way to implement a multithreaded compiler would be for the PassManager class to create multiple instances of each pass object, and allow the separate instances to be hacking on different parts of the program at the same time.

This implementation would prevent each of the passes from having to implement multithreaded constructs, requiring only the LLVM core to have locking in a few places (for global resources). Although this is a simple extension, we simply haven't had time (or multiprocessor machines, thus a reason) to implement this. Despite that, we have kept the LLVM passes SMP ready, and you should too.

Chris Lattner
The LLVM Compiler Infrastructure
Last modified: $Date$