-//===- llvm/Pass.h - Base class for XForm Passes -----------------*- C++ -*--=//
+//===- llvm/Pass.h - Base class for Passes ----------------------*- C++ -*-===//
//
-// This file defines a base class that indicates that a specified class is a
-// transformation pass implementation.
+// The LLVM Compiler Infrastructure
//
-// Pass's are designed this way so that it is possible to apply N passes to a
-// module, by first doing N Pass specific initializations for the module, then
-// looping over all of the methods in the module, doing method specific work
-// N times for each method. Like this:
+// This file was developed by the LLVM research group and is distributed under
+// the University of Illinois Open Source License. See LICENSE.TXT for details.
//
-// for_each(Passes.begin(), Passes.end(), doPassInitialization(Module));
-// for_each(Method *M <- Module->begin(), Module->end())
-// for_each(Passes.begin(), Passes.end(), doPerMethodWork(M));
+//===----------------------------------------------------------------------===//
//
-// The other way to do things is like this:
-// for_each(Pass *P <- Passes.begin(), Passes.end()) {
-// Passes->doPassInitialization(Module)
-// for_each(Module->begin(), Module->end(), P->doPerMethodWork);
-// }
+// This file defines a base class that indicates that a specified class is a
+// transformation pass implementation.
//
-// But this can cause thrashing and poor cache performance, so we don't do it
-// that way.
+// Passes are designed this way so that it is possible to run passes in a cache
+// and organizationally optimal order without having to specify it at the front
+// end. This allows arbitrary passes to be strung together and have them
+// executed as effeciently as possible.
//
-// Because a transformation does not see all methods consecutively, it should
-// be careful about the state that it maintains... another pass may modify a
-// method between two invocatations of doPerMethodWork.
+// Passes should extend one of the classes below, depending on the guarantees
+// that it can make about what will be modified as it is run. For example, most
+// global optimizations should derive from FunctionPass, because they do not add
+// or delete functions, they operate on the internals of the function.
//
-// Also, implementations of doMethodWork should not remove any methods from the
-// module.
+// Note that this file #includes PassSupport.h and PassAnalysisSupport.h (at the
+// bottom), so the APIs exposed by these files are also automatically available
+// to all users of this file.
//
//===----------------------------------------------------------------------===//
#ifndef LLVM_PASS_H
#define LLVM_PASS_H
-#include "llvm/Module.h"
-#include "llvm/Method.h"
+#include <vector>
+#include <map>
+#include <iosfwd>
+#include <typeinfo>
+#include <cassert>
+
+namespace llvm {
+
+class Value;
+class BasicBlock;
+class Function;
+class Module;
+class AnalysisUsage;
+class PassInfo;
+class ImmutablePass;
+template<class Trait> class PassManagerT;
+class BasicBlockPassManager;
+class FunctionPassManagerT;
+class ModulePassManager;
+struct AnalysisResolver;
+
+// AnalysisID - Use the PassInfo to identify a pass...
+typedef const PassInfo* AnalysisID;
//===----------------------------------------------------------------------===//
-// Pass interface - Implemented by all 'passes'.
-//
-struct Pass {
- //===--------------------------------------------------------------------===//
- // The externally useful entry points
+/// Pass interface - Implemented by all 'passes'. Subclass this if you are an
+/// interprocedural optimization or you do not fit into any of the more
+/// constrained passes described below.
+///
+class Pass {
+ friend struct AnalysisResolver;
+ AnalysisResolver *Resolver; // AnalysisResolver this pass is owned by...
+ const PassInfo *PassInfoCache;
+
+ // AnalysisImpls - This keeps track of which passes implement the interfaces
+ // that are required by the current pass (to implement getAnalysis()).
//
+ std::vector<std::pair<const PassInfo*, Pass*> > AnalysisImpls;
+
+ void operator=(const Pass&); // DO NOT IMPLEMENT
+ Pass(const Pass &); // DO NOT IMPLEMENT
+public:
+ Pass() : Resolver(0), PassInfoCache(0) {}
+ virtual ~Pass() {} // Destructor is virtual so we can be subclassed
+
+ /// getPassName - Return a nice clean name for a pass. This usually
+ /// implemented in terms of the name that is registered by one of the
+ /// Registration templates, but can be overloaded directly, and if nothing
+ /// else is available, C++ RTTI will be consulted to get a SOMEWHAT
+ /// intelligible name for the pass.
+ ///
+ virtual const char *getPassName() const;
+
+ /// getPassInfo - Return the PassInfo data structure that corresponds to this
+ /// pass... If the pass has not been registered, this will return null.
+ ///
+ const PassInfo *getPassInfo() const;
+
+ /// runPass - Run this pass, returning true if a modification was made to the
+ /// module argument. This should be implemented by all concrete subclasses.
+ ///
+ virtual bool runPass(Module &M) { return false; }
+ virtual bool runPass(BasicBlock&) { return false; }
+
+ /// print - Print out the internal state of the pass. This is called by
+ /// Analyze to print out the contents of an analysis. Otherwise it is not
+ /// necessary to implement this method. Beware that the module pointer MAY be
+ /// null. This automatically forwards to a virtual function that does not
+ /// provide the Module* in case the analysis doesn't need it it can just be
+ /// ignored.
+ ///
+ virtual void print(std::ostream &O, const Module *M) const;
+ void dump() const; // dump - call print(std::cerr, 0);
+
- // runAllPasses - Run a bunch of passes on the specified module, efficiently.
- static bool runAllPasses(Module *M, std::vector<Pass*> &Passes) {
- bool MadeChanges = false;
- // Run all of the pass initializers
- for (unsigned i = 0; i < Passes.size(); ++i)
- MadeChanges |= Passes[i]->doPassInitialization(M);
-
- // Loop over all of the methods, applying all of the passes to them
- for (unsigned m = 0; m < M->size(); ++m)
- for (unsigned i = 0; i < Passes.size(); ++i)
- MadeChanges |= Passes[i]->doPerMethodWork(*(M->begin()+m));
-
- // Run all of the pass finalizers...
- for (unsigned i = 0; i < Passes.size(); ++i)
- MadeChanges |= Passes[i]->doPassFinalization(M);
- return MadeChanges;
+ /// getAnalysisUsage - This function should be overriden by passes that need
+ /// analysis information to do their job. If a pass specifies that it uses a
+ /// particular analysis result to this function, it can then use the
+ /// getAnalysis<AnalysisType>() function, below.
+ ///
+ virtual void getAnalysisUsage(AnalysisUsage &Info) const {
+ // By default, no analysis results are used, all are invalidated.
}
- // runAllPassesAndFree - Run a bunch of passes on the specified module,
- // efficiently. When done, delete all of the passes.
- //
- static bool runAllPassesAndFree(Module *M, std::vector<Pass*> &Passes) {
- // First run all of the passes
- bool MadeChanges = runAllPasses(M, Passes);
-
- // Free all of the passes.
- for (unsigned i = 0; i < Passes.size(); ++i)
- delete Passes[i];
- return MadeChanges;
+ /// releaseMemory() - This member can be implemented by a pass if it wants to
+ /// be able to release its memory when it is no longer needed. The default
+ /// behavior of passes is to hold onto memory for the entire duration of their
+ /// lifetime (which is the entire compile time). For pipelined passes, this
+ /// is not a big deal because that memory gets recycled every time the pass is
+ /// invoked on another program unit. For IP passes, it is more important to
+ /// free memory when it is unused.
+ ///
+ /// Optionally implement this function to release pass memory when it is no
+ /// longer used.
+ ///
+ virtual void releaseMemory() {}
+
+ // dumpPassStructure - Implement the -debug-passes=PassStructure option
+ virtual void dumpPassStructure(unsigned Offset = 0);
+
+
+ // getPassInfo - Static method to get the pass information from a class name.
+ template<typename AnalysisClass>
+ static const PassInfo *getClassPassInfo() {
+ return lookupPassInfo(typeid(AnalysisClass));
}
+ // lookupPassInfo - Return the pass info object for the specified pass class,
+ // or null if it is not known.
+ static const PassInfo *lookupPassInfo(const std::type_info &TI);
- // run(Module*) - Run this pass on a module and all of the methods contained
- // within it. Returns true if any of the contained passes returned true.
- //
- bool run(Module *M) {
- bool MadeChanges = doPassInitialization(M);
+ /// getAnalysisToUpdate<AnalysisType>() - This function is used by subclasses
+ /// to get to the analysis information that might be around that needs to be
+ /// updated. This is different than getAnalysis in that it can fail (ie the
+ /// analysis results haven't been computed), so should only be used if you
+ /// provide the capability to update an analysis that exists. This method is
+ /// often used by transformation APIs to update analysis results for a pass
+ /// automatically as the transform is performed.
+ ///
+ template<typename AnalysisType>
+ AnalysisType *getAnalysisToUpdate() const; // Defined in PassAnalysisSupport.h
- // Loop over methods in the module. doPerMethodWork could add a method to
- // the Module, so we have to keep checking for end of method list condition.
- //
- for (unsigned m = 0; m < M->size(); ++m)
- MadeChanges |= doPerMethodWork(*(M->begin()+m));
- return MadeChanges | doPassFinalization(M);
+ /// mustPreserveAnalysisID - This method serves the same function as
+ /// getAnalysisToUpdate, but works if you just have an AnalysisID. This
+ /// obviously cannot give you a properly typed instance of the class if you
+ /// don't have the class name available (use getAnalysisToUpdate if you do),
+ /// but it can tell you if you need to preserve the pass at least.
+ ///
+ bool mustPreserveAnalysisID(const PassInfo *AnalysisID) const;
+
+ /// getAnalysis<AnalysisType>() - This function is used by subclasses to get
+ /// to the analysis information that they claim to use by overriding the
+ /// getAnalysisUsage function.
+ ///
+ template<typename AnalysisType>
+ AnalysisType &getAnalysis() const {
+ assert(Resolver && "Pass has not been inserted into a PassManager object!");
+ const PassInfo *PI = getClassPassInfo<AnalysisType>();
+ return getAnalysisID<AnalysisType>(PI);
}
- // run(Method*) - Run this pass on a module and one specific method. Returns
- // false on success.
- //
- bool run(Method *M) {
- return doPassInitialization(M->getParent()) | doPerMethodWork(M) |
- doPassFinalization(M->getParent());
+ template<typename AnalysisType>
+ AnalysisType &getAnalysisID(const PassInfo *PI) const {
+ assert(Resolver && "Pass has not been inserted into a PassManager object!");
+ assert(PI && "getAnalysis for unregistered pass!");
+
+ // PI *must* appear in AnalysisImpls. Because the number of passes used
+ // should be a small number, we just do a linear search over a (dense)
+ // vector.
+ Pass *ResultPass = 0;
+ for (unsigned i = 0; ; ++i) {
+ assert(i != AnalysisImpls.size() &&
+ "getAnalysis*() called on an analysis that was not "
+ "'required' by pass!");
+ if (AnalysisImpls[i].first == PI) {
+ ResultPass = AnalysisImpls[i].second;
+ break;
+ }
+ }
+
+ // Because the AnalysisType may not be a subclass of pass (for
+ // AnalysisGroups), we must use dynamic_cast here to potentially adjust the
+ // return pointer (because the class may multiply inherit, once from pass,
+ // once from AnalysisType).
+ //
+ AnalysisType *Result = dynamic_cast<AnalysisType*>(ResultPass);
+ assert(Result && "Pass does not implement interface required!");
+ return *Result;
}
+private:
+ template<typename Trait> friend class PassManagerT;
+ friend class ModulePassManager;
+ friend class FunctionPassManagerT;
+ friend class BasicBlockPassManager;
+};
- //===--------------------------------------------------------------------===//
- // Functions to be implemented by subclasses
- //
+inline std::ostream &operator<<(std::ostream &OS, const Pass &P) {
+ P.print(OS, 0); return OS;
+}
- // Destructor - Virtual so we can be subclassed
- inline virtual ~Pass() {}
+//===----------------------------------------------------------------------===//
+/// ModulePass class - This class is used to implement unstructured
+/// interprocedural optimizations and analyses. ModulePasses may do anything
+/// they want to the program.
+///
+class ModulePass : public Pass {
+public:
+ /// runOnModule - Virtual method overriden by subclasses to process the module
+ /// being operated on.
+ virtual bool runOnModule(Module &M) = 0;
- // doPassInitialization - Virtual method overridden by subclasses to do
- // any neccesary per-module initialization.
- //
- virtual bool doPassInitialization(Module *M) { return false; }
+ virtual bool runPass(Module &M) { return runOnModule(M); }
+ virtual bool runPass(BasicBlock&) { return false; }
+
+ virtual void addToPassManager(ModulePassManager *PM, AnalysisUsage &AU);
+};
+
+
+//===----------------------------------------------------------------------===//
+/// ImmutablePass class - This class is used to provide information that does
+/// not need to be run. This is useful for things like target information and
+/// "basic" versions of AnalysisGroups.
+///
+class ImmutablePass : public ModulePass {
+public:
+ /// initializePass - This method may be overriden by immutable passes to allow
+ /// them to perform various initialization actions they require. This is
+ /// primarily because an ImmutablePass can "require" another ImmutablePass,
+ /// and if it does, the overloaded version of initializePass may get access to
+ /// these passes with getAnalysis<>.
+ ///
+ virtual void initializePass() {}
+
+ /// ImmutablePasses are never run.
+ ///
+ virtual bool runOnModule(Module &M) { return false; }
+
+private:
+ template<typename Trait> friend class PassManagerT;
+ friend class ModulePassManager;
+ virtual void addToPassManager(ModulePassManager *PM, AnalysisUsage &AU);
+};
+
+//===----------------------------------------------------------------------===//
+/// FunctionPass class - This class is used to implement most global
+/// optimizations. Optimizations should subclass this class if they meet the
+/// following constraints:
+///
+/// 1. Optimizations are organized globally, i.e., a function at a time
+/// 2. Optimizing a function does not cause the addition or removal of any
+/// functions in the module
+///
+class FunctionPass : public ModulePass {
+public:
+ /// doInitialization - Virtual method overridden by subclasses to do
+ /// any necessary per-module initialization.
+ ///
+ virtual bool doInitialization(Module &M) { return false; }
+
+ /// runOnFunction - Virtual method overriden by subclasses to do the
+ /// per-function processing of the pass.
+ ///
+ virtual bool runOnFunction(Function &F) = 0;
+
+ /// doFinalization - Virtual method overriden by subclasses to do any post
+ /// processing needed after all passes have run.
+ ///
+ virtual bool doFinalization(Module &M) { return false; }
+
+ /// runOnModule - On a module, we run this pass by initializing,
+ /// ronOnFunction'ing once for every function in the module, then by
+ /// finalizing.
+ ///
+ virtual bool runOnModule(Module &M);
+
+ /// run - On a function, we simply initialize, run the function, then
+ /// finalize.
+ ///
+ bool run(Function &F);
+
+protected:
+ template<typename Trait> friend class PassManagerT;
+ friend class ModulePassManager;
+ friend class FunctionPassManagerT;
+ friend class BasicBlockPassManager;
+ virtual void addToPassManager(ModulePassManager *PM, AnalysisUsage &AU);
+ virtual void addToPassManager(FunctionPassManagerT *PM, AnalysisUsage &AU);
+};
- // doPerMethodWork - Virtual method overriden by subclasses to do the
- // per-method processing of the pass.
- //
- virtual bool doPerMethodWork(Method *M) { return false; }
- // doPassFinalization - Virtual method overriden by subclasses to do any post
- // processing needed after all passes have run.
+
+//===----------------------------------------------------------------------===//
+/// BasicBlockPass class - This class is used to implement most local
+/// optimizations. Optimizations should subclass this class if they
+/// meet the following constraints:
+/// 1. Optimizations are local, operating on either a basic block or
+/// instruction at a time.
+/// 2. Optimizations do not modify the CFG of the contained function, or any
+/// other basic block in the function.
+/// 3. Optimizations conform to all of the constraints of FunctionPasses.
+///
+class BasicBlockPass : public FunctionPass {
+public:
+ /// doInitialization - Virtual method overridden by subclasses to do
+ /// any necessary per-module initialization.
+ ///
+ virtual bool doInitialization(Module &M) { return false; }
+
+ /// doInitialization - Virtual method overridden by BasicBlockPass subclasses
+ /// to do any necessary per-function initialization.
+ ///
+ virtual bool doInitialization(Function &F) { return false; }
+
+ /// runOnBasicBlock - Virtual method overriden by subclasses to do the
+ /// per-basicblock processing of the pass.
+ ///
+ virtual bool runOnBasicBlock(BasicBlock &BB) = 0;
+
+ /// doFinalization - Virtual method overriden by BasicBlockPass subclasses to
+ /// do any post processing needed after all passes have run.
+ ///
+ virtual bool doFinalization(Function &F) { return false; }
+
+ /// doFinalization - Virtual method overriden by subclasses to do any post
+ /// processing needed after all passes have run.
+ ///
+ virtual bool doFinalization(Module &M) { return false; }
+
+
+ // To run this pass on a function, we simply call runOnBasicBlock once for
+ // each function.
//
- virtual bool doPassFinalization(Module *M) { return false; }
+ bool runOnFunction(Function &F);
+
+ /// To run directly on the basic block, we initialize, runOnBasicBlock, then
+ /// finalize.
+ ///
+ virtual bool runPass(Module &M) { return false; }
+ virtual bool runPass(BasicBlock &BB);
+
+private:
+ template<typename Trait> friend class PassManagerT;
+ friend class FunctionPassManagerT;
+ friend class BasicBlockPassManager;
+ virtual void addToPassManager(ModulePassManager *PM, AnalysisUsage &AU) {
+ FunctionPass::addToPassManager(PM, AU);
+ }
+ virtual void addToPassManager(FunctionPassManagerT *PM, AnalysisUsage &AU);
+ virtual void addToPassManager(BasicBlockPassManager *PM,AnalysisUsage &AU);
};
-#endif
+/// If the user specifies the -time-passes argument on an LLVM tool command line
+/// then the value of this boolean will be true, otherwise false.
+/// @brief This is the storage for the -time-passes option.
+extern bool TimePassesIsEnabled;
+} // End llvm namespace
+
+// Include support files that contain important APIs commonly used by Passes,
+// but that we want to separate out to make it easier to read the header files.
+//
+#include "llvm/PassSupport.h"
+#include "llvm/PassAnalysisSupport.h"
+
+#endif