Add a new MergeFunctions pass. It finds identical functions and merges them.

This triggers only 60 times in llvm-test (look at .llvm.bc, not .linked.rbc)
and so it probably wont be turned on by default. Also, may of those are likely
to go away when PR2973 is fixed.

llvm-svn: 58557

llvm/include/llvm/LinkAllPasses.h

@@ -121,6 +121,7 @@ namespace {
       (void) llvm::createInstructionNamerPass();
       (void) llvm::createPartialSpecializationPass();
       (void) llvm::createAddReadAttrsPass();
+      (void) llvm::createMergeFunctionsPass();
       (void) llvm::createPrintModulePass(0);
       (void) llvm::createPrintFunctionPass("", 0);
 

llvm/include/llvm/Transforms/IPO.h

@@ -38,7 +38,7 @@ ModulePass *createStripSymbolsPass(bool OnlyDebugInfo = false);
 /// to invoke/unwind instructions.  This should really be part of the C/C++
 /// front-end, but it's so much easier to write transformations in LLVM proper.
 ///
-ModulePass* createLowerSetJmpPass();
+ModulePass *createLowerSetJmpPass();
 
 //===----------------------------------------------------------------------===//
 /// createConstantMergePass - This function returns a new pass that merges
@@ -186,13 +186,19 @@ ModulePass *createStripDeadPrototypesPass();
 /// createPartialSpecializationPass - This pass specializes functions for
 /// constant arguments.
 ///
-ModulePass* createPartialSpecializationPass();
+ModulePass *createPartialSpecializationPass();
 
 //===----------------------------------------------------------------------===//
 /// createAddReadAttrsPass - This pass discovers functions that do not access
 /// memory, or only read memory, and gives them the readnone/readonly attribute.
 ///
-Pass* createAddReadAttrsPass();
+Pass *createAddReadAttrsPass();
+
+//===----------------------------------------------------------------------===//
+/// createMergeFunctionsPass - This pass discovers identical functions and
+/// collapses them.
+///
+ModulePass *createMergeFunctionsPass();
 
 } // End llvm namespace
 

llvm/lib/Transforms/IPO/MergeFunctions.cpp

@@ -0,0 +1,358 @@
+//===- MergeFunctions.cpp - Merge identical functions ---------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This pass looks for equivalent functions that are mergable and folds them.
+//
+// A Function will not be analyzed if:
+// * it is overridable at runtime (except for weak linkage), or
+// * it is used by anything other than the callee parameter of a call/invoke
+//
+// A hash is computed from the function, based on its type and number of
+// basic blocks.
+//
+// Once all hashes are computed, we perform an expensive equality comparison
+// on each function pair. This takes n^2/2 comparisons per bucket, so it's
+// important that the hash function be high quality. The equality comparison
+// iterates through each instruction in each basic block.
+//
+// When a match is found, the functions are folded. We can only fold two
+// functions when we know that the definition of one of them is not
+// overridable.
+// * fold a function marked internal by replacing all of its users.
+// * fold extern or weak functions by replacing them with a global alias
+//
+//===----------------------------------------------------------------------===//
+//
+// Future work:
+//
+// * fold vector<T*>::push_back and vector<S*>::push_back.
+//
+// These two functions have different types, but in a way that doesn't matter
+// to us. As long as we never see an S or T itself, using S* and S** is the
+// same as using a T* and T**.
+//
+// * virtual functions.
+//
+// Many functions have their address taken by the virtual function table for
+// the object they belong to. However, as long as it's only used for a lookup
+// and call, this is irrelevant, and we'd like to fold such implementations.
+//
+//===----------------------------------------------------------------------===//
+
+#define DEBUG_TYPE "mergefunc"
+#include "llvm/Transforms/IPO.h"
+#include "llvm/ADT/DenseMap.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/Constants.h"
+#include "llvm/InlineAsm.h"
+#include "llvm/Instructions.h"
+#include "llvm/Module.h"
+#include "llvm/Pass.h"
+#include "llvm/Support/Compiler.h"
+#include "llvm/Support/Debug.h"
+#include <map>
+#include <vector>
+using namespace llvm;
+
+STATISTIC(NumFunctionsMerged, "Number of functions merged");
+STATISTIC(NumMergeFails, "Number of identical function pairings not merged");
+
+namespace {
+  struct VISIBILITY_HIDDEN MergeFunctions : public ModulePass {
+    static char ID; // Pass identification, replacement for typeid
+    MergeFunctions() : ModulePass((intptr_t)&ID) {}
+
+    bool runOnModule(Module &M);
+  };
+}
+
+char MergeFunctions::ID = 0;
+static RegisterPass<MergeFunctions>
+X("mergefunc", "Merge Functions");
+
+ModulePass *llvm::createMergeFunctionsPass() {
+  return new MergeFunctions();
+}
+
+static unsigned hash(const Function *F) {
+  return F->size() ^ reinterpret_cast<unsigned>(F->getType());
+  //return F->size() ^ F->arg_size() ^ F->getReturnType();
+}
+
+static bool compare(const Value *V, const Value *U) {
+  assert(!isa<BasicBlock>(V) && !isa<BasicBlock>(U) &&
+         "Must not compare basic blocks.");
+
+  assert(V->getType() == U->getType() &&
+        "Two of the same operation have operands of different type.");
+
+  // TODO: If the constant is an expression of F, we should accept that it's
+  // equal to the same expression in terms of G.
+  if (isa<Constant>(V))
+    return V == U;
+
+  // The caller has ensured that ValueMap[V] != U. Since Arguments are
+  // pre-loaded into the ValueMap, and Instructions are added as we go, we know
+  // that this can only be a mis-match.
+  if (isa<Instruction>(V) || isa<Argument>(V))
+    return false;
+
+  if (isa<InlineAsm>(V) && isa<InlineAsm>(U)) {
+    const InlineAsm *IAF = cast<InlineAsm>(V);
+    const InlineAsm *IAG = cast<InlineAsm>(U);
+    return IAF->getAsmString() == IAG->getAsmString() &&
+           IAF->getConstraintString() == IAG->getConstraintString();
+  }
+
+  return false;
+}
+
+static bool equals(const BasicBlock *BB1, const BasicBlock *BB2,
+                   DenseMap<const Value *, const Value *> &ValueMap,
+                   DenseMap<const Value *, const Value *> &SpeculationMap) {
+  // Specutively add it anyways. If it's false, we'll notice a difference later, and
+  // this won't matter.
+  ValueMap[BB1] = BB2;
+
+  BasicBlock::const_iterator FI = BB1->begin(), FE = BB1->end();
+  BasicBlock::const_iterator GI = BB2->begin(), GE = BB2->end();
+
+  do {
+    if (!FI->isSameOperationAs(const_cast<Instruction *>(&*GI)))
+      return false;
+
+    if (FI->getNumOperands() != GI->getNumOperands())
+      return false;
+
+    if (ValueMap[FI] == GI) {
+      ++FI, ++GI;
+      continue;
+    }
+
+    if (ValueMap[FI] != NULL)
+      return false;
+
+    for (unsigned i = 0, e = FI->getNumOperands(); i != e; ++i) {
+      Value *OpF = FI->getOperand(i);
+      Value *OpG = GI->getOperand(i);
+
+      if (ValueMap[OpF] == OpG)
+        continue;
+
+      if (ValueMap[OpF] != NULL)
+        return false;
+
+      assert(OpF->getType() == OpG->getType() &&
+             "Two of the same operation has operands of different type.");
+
+      if (OpF->getValueID() != OpG->getValueID())
+        return false;
+
+      if (isa<PHINode>(FI)) {
+        if (SpeculationMap[OpF] == NULL)
+          SpeculationMap[OpF] = OpG;
+        else if (SpeculationMap[OpF] != OpG)
+          return false;
+        continue;
+      } else if (isa<BasicBlock>(OpF)) {
+        assert(isa<TerminatorInst>(FI) &&
+               "BasicBlock referenced by non-Terminator non-PHI");
+        // This call changes the ValueMap, hence we can't use
+        // Value *& = ValueMap[...]
+        if (!equals(cast<BasicBlock>(OpF), cast<BasicBlock>(OpG), ValueMap,
+                    SpeculationMap))
+          return false;
+      } else {
+        if (!compare(OpF, OpG))
+          return false;
+      }
+
+      ValueMap[OpF] = OpG;
+    }
+
+    ValueMap[FI] = GI;
+    ++FI, ++GI;
+  } while (FI != FE && GI != GE);
+
+  return FI == FE && GI == GE;
+}
+
+static bool equals(const Function *F, const Function *G) {
+  // We need to recheck everything, but check the things that weren't included
+  // in the hash first.
+
+  if (F->getAttributes() != G->getAttributes())
+    return false;
+
+  if (F->hasGC() != G->hasGC())
+    return false;
+
+  if (F->hasGC() && F->getGC() != G->getGC())
+    return false;
+
+  if (F->hasSection() != G->hasSection())
+    return false;
+
+  if (F->hasSection() && F->getSection() != G->getSection())
+    return false;
+
+  // TODO: if it's internal and only used in direct calls, we could handle this
+  // case too.
+  if (F->getCallingConv() != G->getCallingConv())
+    return false;
+
+  // TODO: We want to permit cases where two functions take T* and S* but
+  // only load or store them into T** and S**.
+  if (F->getType() != G->getType())
+    return false;
+
+  DenseMap<const Value *, const Value *> ValueMap;
+  DenseMap<const Value *, const Value *> SpeculationMap;
+  ValueMap[F] = G;
+
+  assert(F->arg_size() == G->arg_size() &&
+         "Identical functions have a different number of args.");
+
+  for (Function::const_arg_iterator fi = F->arg_begin(), gi = G->arg_begin(),
+         fe = F->arg_end(); fi != fe; ++fi, ++gi)
+    ValueMap[fi] = gi;
+
+  if (!equals(&F->getEntryBlock(), &G->getEntryBlock(), ValueMap,
+              SpeculationMap))
+    return false;
+
+  for (DenseMap<const Value *, const Value *>::iterator
+         I = SpeculationMap.begin(), E = SpeculationMap.end(); I != E; ++I) {
+    if (ValueMap[I->first] != I->second)
+      return false;
+  }
+
+  return true;
+}
+
+static bool fold(std::vector<Function *> &FnVec, unsigned i, unsigned j) {
+  if (FnVec[i]->mayBeOverridden() && !FnVec[j]->mayBeOverridden())
+    std::swap(FnVec[i], FnVec[j]);
+
+  Function *F = FnVec[i];
+  Function *G = FnVec[j];
+
+  if (!F->mayBeOverridden()) {
+    if (G->hasInternalLinkage()) {
+      F->setAlignment(std::max(F->getAlignment(), G->getAlignment()));
+      G->replaceAllUsesWith(F);
+      G->eraseFromParent();
+      ++NumFunctionsMerged;
+      return true;
+    }
+
+    if (G->hasExternalLinkage() || G->hasWeakLinkage()) {
+      GlobalAlias *GA = new GlobalAlias(G->getType(), G->getLinkage(), "",
+                                        F, G->getParent());
+      F->setAlignment(std::max(F->getAlignment(), G->getAlignment()));
+      GA->takeName(G);
+      GA->setVisibility(G->getVisibility());
+      G->replaceAllUsesWith(GA);
+      G->eraseFromParent();
+      ++NumFunctionsMerged;
+      return true;
+    }
+  }
+
+  DOUT << "Failed on " << F->getName() << " and " << G->getName() << "\n";
+
+  ++NumMergeFails;
+  return false;
+}
+
+static bool hasAddressTaken(User *U) {
+  for (User::use_iterator I = U->use_begin(), E = U->use_end(); I != E; ++I) {
+    User *Use = *I;
+
+    // 'call (bitcast @F to ...)' happens a lot.
+    while (isa<ConstantExpr>(Use) && Use->hasOneUse()) {
+      Use = *Use->use_begin();
+    }
+
+    if (isa<ConstantExpr>(Use)) {
+      if (hasAddressTaken(Use))
+        return true;
+    }
+
+    if (!isa<CallInst>(Use) && !isa<InvokeInst>(Use))
+      return true;
+
+    // Make sure we aren't passing U as a parameter to call instead of the
+    // callee. getOperand(0) is the callee for both CallInst and InvokeInst.
+    // Check the other operands to see if any of them is F.
+    for (User::op_iterator OI = I->op_begin() + 1, OE = I->op_end(); OI != OE;
+         ++OI) {
+      if (*OI == U)
+        return true;
+    }
+  }
+
+  return false;
+}
+
+bool MergeFunctions::runOnModule(Module &M) {
+  bool Changed = false;
+
+  std::map<unsigned, std::vector<Function *> > FnMap;
+
+  for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
+    if (F->isDeclaration() || F->isIntrinsic())
+      continue;
+
+    if (F->hasLinkOnceLinkage() || F->hasCommonLinkage() ||
+        F->hasDLLImportLinkage() || F->hasDLLExportLinkage())
+      continue;
+
+    if (hasAddressTaken(F))
+      continue;
+
+    FnMap[hash(F)].push_back(F);
+  }
+
+  // TODO: instead of running in a loop, we could also fold functions in callgraph
+  // order. Constructing the CFG probably isn't cheaper than just running in a loop.
+
+  bool LocalChanged;
+  do {
+    LocalChanged = false;
+    for (std::map<unsigned, std::vector<Function *> >::iterator I = FnMap.begin(),
+           E = FnMap.end(); I != E; ++I) {
+      DOUT << "size: " << FnMap.size() << "\n";
+      std::vector<Function *> &FnVec = I->second;
+      DOUT << "hash (" << I->first << "): " << FnVec.size() << "\n";
+
+      for (int i = 0, e = FnVec.size(); i != e; ++i) {
+        for (int j = i + 1; j != e; ++j) {
+          bool isEqual = equals(FnVec[i], FnVec[j]);
+
+          DOUT << "  " << FnVec[i]->getName()
+               << (isEqual ? " == " : " != ")
+               << FnVec[j]->getName() << "\n";
+
+          if (isEqual) {
+            if (fold(FnVec, i, j)) {
+              LocalChanged = true;
+              FnVec.erase(FnVec.begin() + j);
+              --j, --e;
+            }
+          }
+        }
+      }
+
+    }
+    Changed |= LocalChanged;
+  } while (LocalChanged);
+
+  return Changed;
+}

llvm/test/Transforms/MergeFunc/dg.exp

@@ -0,0 +1,3 @@
+load_lib llvm.exp
+
+RunLLVMTests [lsort [glob -nocomplain $srcdir/$subdir/*.{ll,c,cpp}]]

llvm/test/Transforms/MergeFunc/phi-speculation1.ll

@@ -0,0 +1,29 @@
+; RUN: llvm-as < %s | opt -mergefunc -stats | not grep {functions merged}
+
+define i32 @foo1(i32 %x) {
+entry:
+  %A = add i32 %x, 1
+  %B = call i32 @foo1(i32 %A)
+  br label %loop
+loop:
+  %C = phi i32 [%B, %entry], [%D, %loop]
+  %D = add i32 %x, 2
+  %E = icmp ugt i32 %D, 10000
+  br i1 %E, label %loopexit, label %loop
+loopexit:
+  ret i32 %D
+}
+
+define i32 @foo2(i32 %x) {
+entry:
+  %0 = add i32 %x, 1
+  %1 = call i32 @foo2(i32 %0)
+  br label %loop
+loop:
+  %2 = phi i32 [%1, %entry], [%3, %loop]
+  %3 = add i32 %2, 2
+  %4 = icmp ugt i32 %3, 10000
+  br i1 %4, label %loopexit, label %loop
+loopexit:
+  ret i32 %3
+}

llvm/test/Transforms/MergeFunc/phi-speculation2.ll

@@ -0,0 +1,29 @@
+; RUN: llvm-as < %s | opt -mergefunc -stats |& grep {functions merged}
+
+define i32 @foo1(i32 %x) {
+entry:
+  %A = add i32 %x, 1
+  %B = call i32 @foo1(i32 %A)
+  br label %loop
+loop:
+  %C = phi i32 [%B, %entry], [%D, %loop]
+  %D = add i32 %C, 2
+  %E = icmp ugt i32 %D, 10000
+  br i1 %E, label %loopexit, label %loop
+loopexit:
+  ret i32 %D
+}
+
+define i32 @foo2(i32 %x) {
+entry:
+  %0 = add i32 %x, 1
+  %1 = call i32 @foo2(i32 %0)
+  br label %loop
+loop:
+  %2 = phi i32 [%1, %entry], [%3, %loop]
+  %3 = add i32 %2, 2
+  %4 = icmp ugt i32 %3, 10000
+  br i1 %4, label %loopexit, label %loop
+loopexit:
+  ret i32 %3
+}