Hook in GlobalMerge pass

llvm-svn: 109359

include/llvm/Target/TargetLowering.h

@@ -792,6 +792,12 @@ public:
     return false;
   }
 
+  /// getMaximalGlobalOffset - Returns the maximal possible offset which can be
+  /// used for loads / stores from the global.
+  virtual unsigned getMaximalGlobalOffset() const {
+    return 0;
+  }
+
   //===--------------------------------------------------------------------===//
   // TargetLowering Optimization Methods
   //

lib/Target/ARM/ARM.h

@@ -98,6 +98,7 @@ FunctionPass *createARMJITCodeEmitterPass(ARMBaseTargetMachine &TM,
 
 FunctionPass *createARMLoadStoreOptimizationPass(bool PreAlloc = false);
 FunctionPass *createARMExpandPseudoPass();
+FunctionPass *createARMGlobalMergePass(const TargetLowering* tli);
 FunctionPass *createARMConstantIslandPass();
 FunctionPass *createNEONPreAllocPass();
 FunctionPass *createNEONMoveFixPass();

lib/Target/ARM/ARMGlobalMerge.cpp

@@ -0,0 +1,203 @@
+//===-- ARMGlobalMerge.cpp - Internal globals merging  --------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+// This pass merges globals with internal linkage into one. This way all the
+// globals which were merged into a biggest one can be addressed using offsets
+// from the same base pointer (no need for separate base pointer for each of the
+// global). Such a transformation can significantly reduce the register pressure
+// when many globals are involved.
+//
+// For example, consider the code which touches several global variables at once:
+//
+// static int foo[N], bar[N], baz[N];
+//
+// for (i = 0; i < N; ++i) {
+//    foo[i] = bar[i] * baz[i];
+// }
+//
+//  On ARM the addresses of 3 arrays should be kept in the registers, thus
+//  this code has quite large register pressure (loop body):
+//
+//  ldr     r1, [r5], #4
+//  ldr     r2, [r6], #4
+//  mul     r1, r2, r1
+//  str     r1, [r0], #4
+//
+//  Pass converts the code to something like:
+//
+//  static struct {
+//    int foo[N];
+//    int bar[N];
+//    int baz[N];
+//  } merged;
+//
+//  for (i = 0; i < N; ++i) {
+//    merged.foo[i] = merged.bar[i] * merged.baz[i];
+//  }
+//
+//  and in ARM code this becomes:
+//
+//  ldr     r0, [r5, #40]
+//  ldr     r1, [r5, #80]
+//  mul     r0, r1, r0
+//  str     r0, [r5], #4
+//
+//  note that we saved 2 registers here almostly "for free".
+// ===----------------------------------------------------------------------===//
+
+#define DEBUG_TYPE "arm-global-merge"
+#include "ARM.h"
+#include "llvm/CodeGen/Passes.h"
+#include "llvm/Attributes.h"
+#include "llvm/Constants.h"
+#include "llvm/DerivedTypes.h"
+#include "llvm/Function.h"
+#include "llvm/GlobalVariable.h"
+#include "llvm/Instructions.h"
+#include "llvm/Intrinsics.h"
+#include "llvm/Module.h"
+#include "llvm/Pass.h"
+#include "llvm/Target/TargetData.h"
+#include "llvm/Target/TargetLowering.h"
+using namespace llvm;
+
+namespace {
+  class LLVM_LIBRARY_VISIBILITY ARMGlobalMerge : public FunctionPass {
+    /// TLI - Keep a pointer of a TargetLowering to consult for determining
+    /// target type sizes.
+    const TargetLowering *TLI;
+
+    bool doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
+                 Module &M, bool) const;
+
+  public:
+    static char ID;             // Pass identification, replacement for typeid.
+    explicit ARMGlobalMerge(const TargetLowering *tli)
+      : FunctionPass(&ID), TLI(tli) {}
+
+    virtual bool doInitialization(Module &M);
+    virtual bool runOnFunction(Function& F);
+
+    const char *getPassName() const {
+      return "Merge internal globals";
+    }
+
+    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+      AU.setPreservesCFG();
+      FunctionPass::getAnalysisUsage(AU);
+    }
+
+    struct GlobalCmp {
+      const TargetData *TD;
+
+      GlobalCmp(const TargetData *td):
+        TD(td) { };
+
+      bool operator() (const GlobalVariable* GV1,
+                       const GlobalVariable* GV2) {
+        const Type* Ty1 = cast<PointerType>(GV1->getType())->getElementType();
+        const Type* Ty2 = cast<PointerType>(GV2->getType())->getElementType();
+
+        return (TD->getTypeAllocSize(Ty1) < TD->getTypeAllocSize(Ty2));
+      }
+    };
+  };
+} // end anonymous namespace
+
+char ARMGlobalMerge::ID = 0;
+
+bool ARMGlobalMerge::doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
+                             Module &M, bool isConst) const {
+  const TargetData *TD = TLI->getTargetData();
+
+  // FIXME: Infer the maximum possible offset depending on the actual users
+  // (these max offsets are different for the users inside Thumb or ARM
+  // functions)
+  unsigned MaxOffset = TLI->getMaximalGlobalOffset();
+
+  // FIXME: Find better heuristics
+  std::stable_sort(Globals.begin(), Globals.end(), GlobalCmp(TD));
+
+  const Type *Int32Ty = Type::getInt32Ty(M.getContext());
+
+  for (size_t i = 0, e = Globals.size(); i != e; ) {
+    size_t j = 0;
+    uint64_t MergedSize = 0;
+    std::vector<const Type*> Tys;
+    std::vector<Constant*> Inits;
+    for (j = i; MergedSize < MaxOffset && j != e; ++j) {
+      const Type* Ty = Globals[j]->getType()->getElementType();
+      Tys.push_back(Ty);
+      Inits.push_back(Globals[j]->getInitializer());
+      MergedSize += TD->getTypeAllocSize(Ty);
+    }
+
+    StructType* MergedTy = StructType::get(M.getContext(), Tys);
+    Constant* MergedInit = ConstantStruct::get(MergedTy, Inits);
+    GlobalVariable* MergedGV = new GlobalVariable(M, MergedTy, isConst,
+                                                  GlobalValue::InternalLinkage,
+                                                  MergedInit, "merged");
+    for (size_t k = i; k < j; ++k) {
+      SmallVector<Constant*, 2> Idx;
+      Idx.push_back(ConstantInt::get(Int32Ty, 0));
+      Idx.push_back(ConstantInt::get(Int32Ty, k-i));
+
+      Constant* GEP =
+        ConstantExpr::getInBoundsGetElementPtr(MergedGV,
+                                               &Idx[0], Idx.size());
+
+      Globals[k]->replaceAllUsesWith(GEP);
+      Globals[k]->eraseFromParent();
+    }
+    i = j;
+  }
+
+  return true;
+}
+
+
+bool ARMGlobalMerge::doInitialization(Module& M) {
+  SmallVector<GlobalVariable*, 16> Globals, ConstGlobals;
+  const TargetData *TD = TLI->getTargetData();
+  unsigned MaxOffset = TLI->getMaximalGlobalOffset();
+  bool Changed = false;
+
+  // Grab all non-const globals.
+  for (Module::global_iterator I = M.global_begin(),
+         E = M.global_end(); I != E; ++I) {
+    // Merge is safe for "normal" internal globals only
+    if (!I->hasLocalLinkage() || I->isThreadLocal() || I->hasSection())
+      continue;
+
+    // Ignore fancy-aligned globals for now.
+    if (I->getAlignment() != 0)
+      continue;
+
+    if (TD->getTypeAllocSize(I->getType()) < MaxOffset) {
+      if (I->isConstant())
+        ConstGlobals.push_back(I);
+      else
+        Globals.push_back(I);
+    }
+  }
+
+  if (Globals.size() > 1)
+    Changed |= doMerge(Globals, M, false);
+  if (ConstGlobals.size() > 1)
+    Changed |= doMerge(ConstGlobals, M, true);
+
+  return Changed;
+}
+
+bool ARMGlobalMerge::runOnFunction(Function& F) {
+  return false;
+}
+
+FunctionPass *llvm::createARMGlobalMergePass(const TargetLowering *tli) {
+  return new ARMGlobalMerge(tli);
+}

lib/Target/ARM/ARMISelLowering.cpp

@@ -703,6 +703,12 @@ unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
   return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 1 : 2;
 }
 
+/// getMaximalGlobalOffset - Returns the maximal possible offset which can
+/// be used for loads / stores from the global.
+unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
+  return (Subtarget->isThumb1Only() ? 127 : 4095);
+}
+
 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
   unsigned NumVals = N->getNumValues();
   if (!NumVals)

lib/Target/ARM/ARMISelLowering.h

@@ -263,6 +263,10 @@ namespace llvm {
     /// getFunctionAlignment - Return the Log2 alignment of this function.
     virtual unsigned getFunctionAlignment(const Function *F) const;
 
+    /// getMaximalGlobalOffset - Returns the maximal possible offset which can
+    /// be used for loads / stores from the global.
+    virtual unsigned getMaximalGlobalOffset() const;
+
     /// createFastISel - This method returns a target specific FastISel object,
     /// or null if the target does not support "fast" ISel.
     virtual FastISel *createFastISel(FunctionLoweringInfo &funcInfo) const;

lib/Target/ARM/ARMTargetMachine.cpp

@@ -85,9 +85,15 @@ ThumbTargetMachine::ThumbTargetMachine(const Target &T, const std::string &TT,
     TSInfo(*this) {
 }
 
-
-
 // Pass Pipeline Configuration
+bool ARMBaseTargetMachine::addPreISel(PassManagerBase &PM,
+                                      CodeGenOpt::Level OptLevel) {
+  if (OptLevel != CodeGenOpt::None)
+    PM.add(createARMGlobalMergePass(getTargetLowering()));
+
+  return false;
+}
+
 bool ARMBaseTargetMachine::addInstSelector(PassManagerBase &PM,
                                            CodeGenOpt::Level OptLevel) {
   PM.add(createARMISelDag(*this, OptLevel));

lib/Target/ARM/ARMTargetMachine.h

@@ -50,6 +50,7 @@ public:
   }
 
   // Pass Pipeline Configuration
+  virtual bool addPreISel(PassManagerBase &PM, CodeGenOpt::Level OptLevel);
   virtual bool addInstSelector(PassManagerBase &PM, CodeGenOpt::Level OptLevel);
   virtual bool addPreRegAlloc(PassManagerBase &PM, CodeGenOpt::Level OptLevel);
   virtual bool addPreSched2(PassManagerBase &PM, CodeGenOpt::Level OptLevel);