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 using namespace llvm;
  69
  70 namespace {
  71   class ARMGlobalMerge : public FunctionPass {
  72     /// TLI - Keep a pointer of a TargetLowering to consult for determining
  73     /// target type sizes.
  74     const TargetLowering *TLI;
  75
  76     bool doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
  77                  Module &M, bool) const;
  78
  79   public:
  80     static char ID;             // Pass identification, replacement for typeid.
  81     explicit ARMGlobalMerge(const TargetLowering *tli)
  82       : FunctionPass(ID), TLI(tli) {}
  83
  84     virtual bool doInitialization(Module &M);
  85     virtual bool runOnFunction(Function &F);
  86
  87     const char *getPassName() const {
  88       return "Merge internal globals";
  89     }
  90
  91     virtual void getAnalysisUsage(AnalysisUsage &AU) const {
  92       AU.setPreservesCFG();
  93       FunctionPass::getAnalysisUsage(AU);
  94     }
  95
  96     struct GlobalCmp {
  97       const TargetData *TD;
  98
  99       GlobalCmp(const TargetData *td) : TD(td) { }
 100
 101       bool operator()(const GlobalVariable *GV1, const GlobalVariable *GV2) {
 102         const Type *Ty1 = cast<PointerType>(GV1->getType())->getElementType();
 103         const Type *Ty2 = cast<PointerType>(GV2->getType())->getElementType();
 104
 105         return (TD->getTypeAllocSize(Ty1) < TD->getTypeAllocSize(Ty2));
 106       }
 107     };
 108   };
 109 } // end anonymous namespace
 110
 111 char ARMGlobalMerge::ID = 0;
 112
 113 bool ARMGlobalMerge::doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
 114                              Module &M, bool isConst) const {
 115   const TargetData *TD = TLI->getTargetData();
 116
 117   // FIXME: Infer the maximum possible offset depending on the actual users
 118   // (these max offsets are different for the users inside Thumb or ARM
 119   // functions)
 120   unsigned MaxOffset = TLI->getMaximalGlobalOffset();
 121
 122   // FIXME: Find better heuristics
 123   std::stable_sort(Globals.begin(), Globals.end(), GlobalCmp(TD));
 124
 125   const Type *Int32Ty = Type::getInt32Ty(M.getContext());
 126
 127   for (size_t i = 0, e = Globals.size(); i != e; ) {
 128     size_t j = 0;
 129     uint64_t MergedSize = 0;
 130     std::vector<const Type*> Tys;
 131     std::vector<Constant*> Inits;
 132     for (j = i; MergedSize < MaxOffset && j != e; ++j) {
 133       const Type *Ty = Globals[j]->getType()->getElementType();
 134       Tys.push_back(Ty);
 135       Inits.push_back(Globals[j]->getInitializer());
 136       MergedSize += TD->getTypeAllocSize(Ty);
 137     }
 138
 139     StructType *MergedTy = StructType::get(M.getContext(), Tys);
 140     Constant *MergedInit = ConstantStruct::get(MergedTy, Inits);
 141     GlobalVariable *MergedGV = new GlobalVariable(M, MergedTy, isConst,
 142                                                   GlobalValue::InternalLinkage,
 143                                                   MergedInit, "merged");
 144     for (size_t k = i; k < j; ++k) {
 145       Constant *Idx[2] = {
 146         ConstantInt::get(Int32Ty, 0),
 147         ConstantInt::get(Int32Ty, k-i)
 148       };
 149       Constant *GEP = ConstantExpr::getInBoundsGetElementPtr(MergedGV, Idx, 2);
 150       Globals[k]->replaceAllUsesWith(GEP);
 151       Globals[k]->eraseFromParent();
 152     }
 153     i = j;
 154   }
 155
 156   return true;
 157 }
 158
 159
 160 bool ARMGlobalMerge::doInitialization(Module &M) {
 161   SmallVector<GlobalVariable*, 16> Globals, ConstGlobals;
 162   const TargetData *TD = TLI->getTargetData();
 163   unsigned MaxOffset = TLI->getMaximalGlobalOffset();
 164   bool Changed = false;
 165
 166   // Grab all non-const globals.
 167   for (Module::global_iterator I = M.global_begin(),
 168          E = M.global_end(); I != E; ++I) {
 169     // Merge is safe for "normal" internal globals only
 170     if (!I->hasLocalLinkage() || I->isThreadLocal() || I->hasSection())
 171       continue;
 172
 173     // Ignore fancy-aligned globals for now.
 174     if (I->getAlignment() != 0)
 175       continue;
 176
 177     // Ignore all 'special' globals.
 178     if (I->getName().startswith("llvm.") ||
 179         I->getName().startswith(".llvm."))
 180       continue;
 181
 182     if (TD->getTypeAllocSize(I->getType()) < MaxOffset) {
 183       if (I->isConstant())
 184         ConstGlobals.push_back(I);
 185       else
 186         Globals.push_back(I);
 187     }
 188   }
 189
 190   if (Globals.size() > 1)
 191     Changed |= doMerge(Globals, M, false);
 192   // FIXME: This currently breaks the EH processing due to way how the
 193   // typeinfo detection works. We might want to detect the TIs and ignore
 194   // them in the future.
 195
 196   // if (ConstGlobals.size() > 1)
 197   //  Changed |= doMerge(ConstGlobals, M, true);
 198
 199   return Changed;
 200 }
 201
 202 bool ARMGlobalMerge::runOnFunction(Function &F) {
 203   return false;
 204 }
 205
 206 FunctionPass *llvm::createARMGlobalMergePass(const TargetLowering *tli) {
 207   return new ARMGlobalMerge(tli);
 208 }