lib/Target/ARM/ARMGlobalMerge.cpp

   1 //===-- ARMGlobalMerge.cpp - Internal globals merging  --------------------===//
   2 //
   3 //                     The LLVM Compiler Infrastructure
   4 //
   5 // This file is distributed under the University of Illinois Open Source
   6 // License. See LICENSE.TXT for details.
   7 //
   8 //===----------------------------------------------------------------------===//
   9 // This pass merges globals with internal linkage into one. This way all the
  10 // globals which were merged into a biggest one can be addressed using offsets
  11 // from the same base pointer (no need for separate base pointer for each of the
  12 // global). Such a transformation can significantly reduce the register pressure
  13 // when many globals are involved.
  14 //
  15 // For example, consider the code which touches several global variables at
  16 // once:
  17 //
  18 // static int foo[N], bar[N], baz[N];
  19 //
  20 // for (i = 0; i < N; ++i) {
  21 //    foo[i] = bar[i] * baz[i];
  22 // }
  23 //
  24 //  On ARM the addresses of 3 arrays should be kept in the registers, thus
  25 //  this code has quite large register pressure (loop body):
  26 //
  27 //  ldr     r1, [r5], #4
  28 //  ldr     r2, [r6], #4
  29 //  mul     r1, r2, r1
  30 //  str     r1, [r0], #4
  31 //
  32 //  Pass converts the code to something like:
  33 //
  34 //  static struct {
  35 //    int foo[N];
  36 //    int bar[N];
  37 //    int baz[N];
  38 //  } merged;
  39 //
  40 //  for (i = 0; i < N; ++i) {
  41 //    merged.foo[i] = merged.bar[i] * merged.baz[i];
  42 //  }
  43 //
  44 //  and in ARM code this becomes:
  45 //
  46 //  ldr     r0, [r5, #40]
  47 //  ldr     r1, [r5, #80]
  48 //  mul     r0, r1, r0
  49 //  str     r0, [r5], #4
  50 //
  51 //  note that we saved 2 registers here almostly "for free".
  52 // ===---------------------------------------------------------------------===//
  53
  54 #define DEBUG_TYPE "arm-global-merge"
  55 #include "ARM.h"
  56 #include "llvm/CodeGen/Passes.h"
  57 #include "llvm/Attributes.h"
  58 #include "llvm/Constants.h"
  59 #include "llvm/DerivedTypes.h"
  60 #include "llvm/Function.h"
  61 #include "llvm/GlobalVariable.h"
  62 #include "llvm/Instructions.h"
  63 #include "llvm/Intrinsics.h"
  64 #include "llvm/Module.h"
  65 #include "llvm/Pass.h"
  66 #include "llvm/Target/TargetData.h"
  67 #include "llvm/Target/TargetLowering.h"
  68 #include "llvm/Target/TargetLoweringObjectFile.h"
  69 using namespace llvm;
  70
  71 namespace {
  72   class ARMGlobalMerge : public FunctionPass {
  73     /// TLI - Keep a pointer of a TargetLowering to consult for determining
  74     /// target type sizes.
  75     const TargetLowering *TLI;
  76
  77     bool doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
  78                  Module &M, bool isConst) const;
  79
  80   public:
  81     static char ID;             // Pass identification, replacement for typeid.
  82     explicit ARMGlobalMerge(const TargetLowering *tli)
  83       : FunctionPass(ID), TLI(tli) {}
  84
  85     virtual bool doInitialization(Module &M);
  86     virtual bool runOnFunction(Function &F);
  87
  88     const char *getPassName() const {
  89       return "Merge internal globals";
  90     }
  91
  92     virtual void getAnalysisUsage(AnalysisUsage &AU) const {
  93       AU.setPreservesCFG();
  94       FunctionPass::getAnalysisUsage(AU);
  95     }
  96
  97     struct GlobalCmp {
  98       const TargetData *TD;
  99
 100       GlobalCmp(const TargetData *td) : TD(td) { }
 101
 102       bool operator()(const GlobalVariable *GV1, const GlobalVariable *GV2) {
 103         const Type *Ty1 = cast<PointerType>(GV1->getType())->getElementType();
 104         const Type *Ty2 = cast<PointerType>(GV2->getType())->getElementType();
 105
 106         return (TD->getTypeAllocSize(Ty1) < TD->getTypeAllocSize(Ty2));
 107       }
 108     };
 109   };
 110 } // end anonymous namespace
 111
 112 char ARMGlobalMerge::ID = 0;
 113
 114 bool ARMGlobalMerge::doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
 115                              Module &M, bool isConst) const {
 116   const TargetData *TD = TLI->getTargetData();
 117
 118   // FIXME: Infer the maximum possible offset depending on the actual users
 119   // (these max offsets are different for the users inside Thumb or ARM
 120   // functions)
 121   unsigned MaxOffset = TLI->getMaximalGlobalOffset();
 122
 123   // FIXME: Find better heuristics
 124   std::stable_sort(Globals.begin(), Globals.end(), GlobalCmp(TD));
 125
 126   const Type *Int32Ty = Type::getInt32Ty(M.getContext());
 127
 128   for (size_t i = 0, e = Globals.size(); i != e; ) {
 129     size_t j = 0;
 130     uint64_t MergedSize = 0;
 131     std::vector<const Type*> Tys;
 132     std::vector<Constant*> Inits;
 133     for (j = i; j != e; ++j) {
 134       const Type *Ty = Globals[j]->getType()->getElementType();
 135       MergedSize += TD->getTypeAllocSize(Ty);
 136       if (MergedSize > MaxOffset) {
 137         break;
 138       }
 139       Tys.push_back(Ty);
 140       Inits.push_back(Globals[j]->getInitializer());
 141     }
 142
 143     StructType *MergedTy = StructType::get(M.getContext(), Tys);
 144     Constant *MergedInit = ConstantStruct::get(MergedTy, Inits);
 145     GlobalVariable *MergedGV = new GlobalVariable(M, MergedTy, isConst,
 146                                                   GlobalValue::InternalLinkage,
 147                                                   MergedInit, "_MergedGlobals");
 148     for (size_t k = i; k < j; ++k) {
 149       Constant *Idx[2] = {
 150         ConstantInt::get(Int32Ty, 0),
 151         ConstantInt::get(Int32Ty, k-i)
 152       };
 153       Constant *GEP = ConstantExpr::getInBoundsGetElementPtr(MergedGV, Idx, 2);
 154       Globals[k]->replaceAllUsesWith(GEP);
 155       Globals[k]->eraseFromParent();
 156     }
 157     i = j;
 158   }
 159
 160   return true;
 161 }
 162
 163
 164 bool ARMGlobalMerge::doInitialization(Module &M) {
 165   SmallVector<GlobalVariable*, 16> Globals, ConstGlobals, BSSGlobals;
 166   const TargetData *TD = TLI->getTargetData();
 167   unsigned MaxOffset = TLI->getMaximalGlobalOffset();
 168   bool Changed = false;
 169
 170   // Grab all non-const globals.
 171   for (Module::global_iterator I = M.global_begin(),
 172          E = M.global_end(); I != E; ++I) {
 173     // Merge is safe for "normal" internal globals only
 174     if (!I->hasLocalLinkage() || I->isThreadLocal() || I->hasSection())
 175       continue;
 176
 177     // Ignore fancy-aligned globals for now.
 178     unsigned Alignment = I->getAlignment();
 179     unsigned AllocSize = TD->getTypeAllocSize(I->getType()->getElementType());
 180     if (Alignment > AllocSize)
 181       continue;
 182
 183     // Ignore all 'special' globals.
 184     if (I->getName().startswith("llvm.") ||
 185         I->getName().startswith(".llvm."))
 186       continue;
 187
 188     if (AllocSize < MaxOffset) {
 189       const TargetLoweringObjectFile &TLOF = TLI->getObjFileLowering();
 190       if (TLOF.getKindForGlobal(I, TLI->getTargetMachine()).isBSSLocal())
 191         BSSGlobals.push_back(I);
 192       else if (I->isConstant())
 193         ConstGlobals.push_back(I);
 194       else
 195         Globals.push_back(I);
 196     }
 197   }
 198
 199   if (Globals.size() > 1)
 200     Changed |= doMerge(Globals, M, false);
 201   if (BSSGlobals.size() > 1)
 202     Changed |= doMerge(BSSGlobals, M, false);
 203
 204   // FIXME: This currently breaks the EH processing due to way how the
 205   // typeinfo detection works. We might want to detect the TIs and ignore
 206   // them in the future.
 207   // if (ConstGlobals.size() > 1)
 208   //  Changed |= doMerge(ConstGlobals, M, true);
 209
 210   return Changed;
 211 }
 212
 213 bool ARMGlobalMerge::runOnFunction(Function &F) {
 214   return false;
 215 }
 216
 217 FunctionPass *llvm::createARMGlobalMergePass(const TargetLowering *tli) {
 218   return new ARMGlobalMerge(tli);
 219 }