LCOV - code coverage report
Current view: top level - lib/CodeGen - GlobalMerge.cpp (source / functions) Hit Total Coverage
Test: llvm-toolchain.info Lines: 237 241 98.3 %
Date: 2017-09-14 15:23:50 Functions: 21 21 100.0 %
Legend: Lines: hit not hit

          Line data    Source code
       1             : //===-- GlobalMerge.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             : // However, merging globals can have tradeoffs:
      54             : // - it confuses debuggers, tools, and users
      55             : // - it makes linker optimizations less useful (order files, LOHs, ...)
      56             : // - it forces usage of indexed addressing (which isn't necessarily "free")
      57             : // - it can increase register pressure when the uses are disparate enough.
      58             : // 
      59             : // We use heuristics to discover the best global grouping we can (cf cl::opts).
      60             : // ===---------------------------------------------------------------------===//
      61             : 
      62             : #include "llvm/ADT/DenseMap.h"
      63             : #include "llvm/ADT/SmallBitVector.h"
      64             : #include "llvm/ADT/SmallPtrSet.h"
      65             : #include "llvm/ADT/Statistic.h"
      66             : #include "llvm/CodeGen/Passes.h"
      67             : #include "llvm/IR/Attributes.h"
      68             : #include "llvm/IR/Constants.h"
      69             : #include "llvm/IR/DataLayout.h"
      70             : #include "llvm/IR/DerivedTypes.h"
      71             : #include "llvm/IR/Function.h"
      72             : #include "llvm/IR/GlobalVariable.h"
      73             : #include "llvm/IR/Instructions.h"
      74             : #include "llvm/IR/Intrinsics.h"
      75             : #include "llvm/IR/Module.h"
      76             : #include "llvm/Pass.h"
      77             : #include "llvm/Support/CommandLine.h"
      78             : #include "llvm/Support/Debug.h"
      79             : #include "llvm/Support/raw_ostream.h"
      80             : #include "llvm/Target/TargetLowering.h"
      81             : #include "llvm/Target/TargetLoweringObjectFile.h"
      82             : #include "llvm/Target/TargetSubtargetInfo.h"
      83             : #include <algorithm>
      84             : using namespace llvm;
      85             : 
      86             : #define DEBUG_TYPE "global-merge"
      87             : 
      88             : // FIXME: This is only useful as a last-resort way to disable the pass.
      89             : static cl::opt<bool>
      90       72306 : EnableGlobalMerge("enable-global-merge", cl::Hidden,
      91      216918 :                   cl::desc("Enable the global merge pass"),
      92      289224 :                   cl::init(true));
      93             : 
      94             : static cl::opt<unsigned>
      95       72306 : GlobalMergeMaxOffset("global-merge-max-offset", cl::Hidden,
      96      216918 :                      cl::desc("Set maximum offset for global merge pass"),
      97      289224 :                      cl::init(0));
      98             : 
      99       72306 : static cl::opt<bool> GlobalMergeGroupByUse(
     100             :     "global-merge-group-by-use", cl::Hidden,
     101      216918 :     cl::desc("Improve global merge pass to look at uses"), cl::init(true));
     102             : 
     103       72306 : static cl::opt<bool> GlobalMergeIgnoreSingleUse(
     104             :     "global-merge-ignore-single-use", cl::Hidden,
     105      216918 :     cl::desc("Improve global merge pass to ignore globals only used alone"),
     106      289224 :     cl::init(true));
     107             : 
     108             : static cl::opt<bool>
     109       72306 : EnableGlobalMergeOnConst("global-merge-on-const", cl::Hidden,
     110      216918 :                          cl::desc("Enable global merge pass on constants"),
     111      289224 :                          cl::init(false));
     112             : 
     113             : // FIXME: this could be a transitional option, and we probably need to remove
     114             : // it if only we are sure this optimization could always benefit all targets.
     115             : static cl::opt<cl::boolOrDefault>
     116       72306 : EnableGlobalMergeOnExternal("global-merge-on-external", cl::Hidden,
     117      144612 :      cl::desc("Enable global merge pass on external linkage"));
     118             : 
     119             : STATISTIC(NumMerged, "Number of globals merged");
     120             : namespace {
     121       12812 :   class GlobalMerge : public FunctionPass {
     122             :     const TargetMachine *TM;
     123             :     // FIXME: Infer the maximum possible offset depending on the actual users
     124             :     // (these max offsets are different for the users inside Thumb or ARM
     125             :     // functions), see the code that passes in the offset in the ARM backend
     126             :     // for more information.
     127             :     unsigned MaxOffset;
     128             : 
     129             :     /// Whether we should try to optimize for size only.
     130             :     /// Currently, this applies a dead simple heuristic: only consider globals
     131             :     /// used in minsize functions for merging.
     132             :     /// FIXME: This could learn about optsize, and be used in the cost model.
     133             :     bool OnlyOptimizeForSize;
     134             : 
     135             :     /// Whether we should merge global variables that have external linkage.
     136             :     bool MergeExternalGlobals;
     137             : 
     138             :     bool IsMachO;
     139             : 
     140             :     bool doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
     141             :                  Module &M, bool isConst, unsigned AddrSpace) const;
     142             :     /// \brief Merge everything in \p Globals for which the corresponding bit
     143             :     /// in \p GlobalSet is set.
     144             :     bool doMerge(const SmallVectorImpl<GlobalVariable *> &Globals,
     145             :                  const BitVector &GlobalSet, Module &M, bool isConst,
     146             :                  unsigned AddrSpace) const;
     147             : 
     148             :     /// \brief Check if the given variable has been identified as must keep
     149             :     /// \pre setMustKeepGlobalVariables must have been called on the Module that
     150             :     ///      contains GV
     151             :     bool isMustKeepGlobalVariable(const GlobalVariable *GV) const {
     152         320 :       return MustKeepGlobalVariables.count(GV);
     153             :     }
     154             : 
     155             :     /// Collect every variables marked as "used" or used in a landing pad
     156             :     /// instruction for this Module.
     157             :     void setMustKeepGlobalVariables(Module &M);
     158             : 
     159             :     /// Collect every variables marked as "used"
     160             :     void collectUsedGlobalVariables(Module &M);
     161             : 
     162             :     /// Keep track of the GlobalVariable that must not be merged away
     163             :     SmallPtrSet<const GlobalVariable *, 16> MustKeepGlobalVariables;
     164             : 
     165             :   public:
     166             :     static char ID;             // Pass identification, replacement for typeid.
     167           2 :     explicit GlobalMerge()
     168           2 :         : FunctionPass(ID), TM(nullptr), MaxOffset(GlobalMergeMaxOffset),
     169           8 :           OnlyOptimizeForSize(false), MergeExternalGlobals(false) {
     170           2 :       initializeGlobalMergePass(*PassRegistry::getPassRegistry());
     171           2 :     }
     172             : 
     173        3235 :     explicit GlobalMerge(const TargetMachine *TM, unsigned MaximalOffset,
     174             :                          bool OnlyOptimizeForSize, bool MergeExternalGlobals)
     175        3235 :         : FunctionPass(ID), TM(TM), MaxOffset(MaximalOffset),
     176             :           OnlyOptimizeForSize(OnlyOptimizeForSize),
     177        9705 :           MergeExternalGlobals(MergeExternalGlobals) {
     178        3235 :       initializeGlobalMergePass(*PassRegistry::getPassRegistry());
     179        3235 :     }
     180             : 
     181             :     bool doInitialization(Module &M) override;
     182             :     bool runOnFunction(Function &F) override;
     183             :     bool doFinalization(Module &M) override;
     184             : 
     185           6 :     StringRef getPassName() const override { return "Merge internal globals"; }
     186             : 
     187        3229 :     void getAnalysisUsage(AnalysisUsage &AU) const override {
     188        3229 :       AU.setPreservesCFG();
     189        3229 :       FunctionPass::getAnalysisUsage(AU);
     190        3229 :     }
     191             :   };
     192             : } // end anonymous namespace
     193             : 
     194             : char GlobalMerge::ID = 0;
     195       61921 : INITIALIZE_PASS(GlobalMerge, DEBUG_TYPE, "Merge global variables", false, false)
     196             : 
     197          76 : bool GlobalMerge::doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
     198             :                           Module &M, bool isConst, unsigned AddrSpace) const {
     199          76 :   auto &DL = M.getDataLayout();
     200             :   // FIXME: Find better heuristics
     201         304 :   std::stable_sort(Globals.begin(), Globals.end(),
     202         458 :                    [&DL](const GlobalVariable *GV1, const GlobalVariable *GV2) {
     203         229 :                      return DL.getTypeAllocSize(GV1->getValueType()) <
     204         229 :                             DL.getTypeAllocSize(GV2->getValueType());
     205         229 :                    });
     206             : 
     207             :   // If we want to just blindly group all globals together, do so.
     208          76 :   if (!GlobalMergeGroupByUse) {
     209          24 :     BitVector AllGlobals(Globals.size());
     210           8 :     AllGlobals.set();
     211           8 :     return doMerge(Globals, AllGlobals, M, isConst, AddrSpace);
     212             :   }
     213             : 
     214             :   // If we want to be smarter, look at all uses of each global, to try to
     215             :   // discover all sets of globals used together, and how many times each of
     216             :   // these sets occurred.
     217             :   //
     218             :   // Keep this reasonably efficient, by having an append-only list of all sets
     219             :   // discovered so far (UsedGlobalSet), and mapping each "together-ness" unit of
     220             :   // code (currently, a Function) to the set of globals seen so far that are
     221             :   // used together in that unit (GlobalUsesByFunction).
     222             :   //
     223             :   // When we look at the Nth global, we now that any new set is either:
     224             :   // - the singleton set {N}, containing this global only, or
     225             :   // - the union of {N} and a previously-discovered set, containing some
     226             :   //   combination of the previous N-1 globals.
     227             :   // Using that knowledge, when looking at the Nth global, we can keep:
     228             :   // - a reference to the singleton set {N} (CurGVOnlySetIdx)
     229             :   // - a list mapping each previous set to its union with {N} (EncounteredUGS),
     230             :   //   if it actually occurs.
     231             : 
     232             :   // We keep track of the sets of globals used together "close enough".
     233        1377 :   struct UsedGlobalSet {
     234         161 :     UsedGlobalSet(size_t Size) : Globals(Size), UsageCount(1) {}
     235             :     BitVector Globals;
     236             :     unsigned UsageCount;
     237             :   };
     238             : 
     239             :   // Each set is unique in UsedGlobalSets.
     240         136 :   std::vector<UsedGlobalSet> UsedGlobalSets;
     241             : 
     242             :   // Avoid repeating the create-global-set pattern.
     243         161 :   auto CreateGlobalSet = [&]() -> UsedGlobalSet & {
     244         322 :     UsedGlobalSets.emplace_back(Globals.size());
     245         161 :     return UsedGlobalSets.back();
     246          68 :   };
     247             : 
     248             :   // The first set is the empty set.
     249          68 :   CreateGlobalSet().UsageCount = 0;
     250             : 
     251             :   // We define "close enough" to be "in the same function".
     252             :   // FIXME: Grouping uses by function is way too aggressive, so we should have
     253             :   // a better metric for distance between uses.
     254             :   // The obvious alternative would be to group by BasicBlock, but that's in
     255             :   // turn too conservative..
     256             :   // Anything in between wouldn't be trivial to compute, so just stick with
     257             :   // per-function grouping.
     258             : 
     259             :   // The value type is an index into UsedGlobalSets.
     260             :   // The default (0) conveniently points to the empty set.
     261         136 :   DenseMap<Function *, size_t /*UsedGlobalSetIdx*/> GlobalUsesByFunction;
     262             : 
     263             :   // Now, look at each merge-eligible global in turn.
     264             : 
     265             :   // Keep track of the sets we already encountered to which we added the
     266             :   // current global.
     267             :   // Each element matches the same-index element in UsedGlobalSets.
     268             :   // This lets us efficiently tell whether a set has already been expanded to
     269             :   // include the current global.
     270         136 :   std::vector<size_t> EncounteredUGS;
     271             : 
     272         331 :   for (size_t GI = 0, GE = Globals.size(); GI != GE; ++GI) {
     273         390 :     GlobalVariable *GV = Globals[GI];
     274             : 
     275             :     // Reset the encountered sets for this global...
     276         780 :     std::fill(EncounteredUGS.begin(), EncounteredUGS.end(), 0);
     277             :     // ...and grow it in case we created new sets for the previous global.
     278         390 :     EncounteredUGS.resize(UsedGlobalSets.size());
     279             : 
     280             :     // We might need to create a set that only consists of the current global.
     281             :     // Keep track of its index into UsedGlobalSets.
     282         195 :     size_t CurGVOnlySetIdx = 0;
     283             : 
     284             :     // For each global, look at all its Uses.
     285        1203 :     for (auto &U : GV->uses()) {
     286             :       // This Use might be a ConstantExpr.  We're interested in Instruction
     287             :       // users, so look through ConstantExpr...
     288             :       Use *UI, *UE;
     289         723 :       if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U.getUser())) {
     290         210 :         if (CE->use_empty())
     291           0 :           continue;
     292             :         UI = &*CE->use_begin();
     293             :         UE = nullptr;
     294        1026 :       } else if (isa<Instruction>(U.getUser())) {
     295         513 :         UI = &U;
     296         513 :         UE = UI->getNext();
     297             :       } else {
     298           0 :         continue;
     299             :       }
     300             : 
     301             :       // ...to iterate on all the instruction users of the global.
     302             :       // Note that we iterate on Uses and not on Users to be able to getNext().
     303        1896 :       for (; UI != UE; UI = UI->getNext()) {
     304        1278 :         Instruction *I = dyn_cast<Instruction>(UI->getUser());
     305           0 :         if (!I)
     306         591 :           continue;
     307             : 
     308         639 :         Function *ParentFn = I->getParent()->getParent();
     309             : 
     310             :         // If we're only optimizing for size, ignore non-minsize functions.
     311        1583 :         if (OnlyOptimizeForSize && !ParentFn->optForMinSize())
     312         468 :           continue;
     313             : 
     314         171 :         size_t UGSIdx = GlobalUsesByFunction[ParentFn];
     315             : 
     316             :         // If this is the first global the basic block uses, map it to the set
     317             :         // consisting of this global only.
     318         222 :         if (!UGSIdx) {
     319             :           // If that set doesn't exist yet, create it.
     320          51 :           if (!CurGVOnlySetIdx) {
     321          90 :             CurGVOnlySetIdx = UsedGlobalSets.size();
     322          45 :             CreateGlobalSet().Globals.set(GI);
     323             :           } else {
     324          12 :             ++UsedGlobalSets[CurGVOnlySetIdx].UsageCount;
     325             :           }
     326             : 
     327          51 :           GlobalUsesByFunction[ParentFn] = CurGVOnlySetIdx;
     328          51 :           continue;
     329             :         }
     330             : 
     331             :         // If we already encountered this BB, just increment the counter.
     332         426 :         if (UsedGlobalSets[UGSIdx].Globals.test(GI)) {
     333          66 :           ++UsedGlobalSets[UGSIdx].UsageCount;
     334          66 :           continue;
     335             :         }
     336             : 
     337             :         // If not, the previous set wasn't actually used in this function.
     338          54 :         --UsedGlobalSets[UGSIdx].UsageCount;
     339             : 
     340             :         // If we already expanded the previous set to include this global, just
     341             :         // reuse that expanded set.
     342         114 :         if (size_t ExpandedIdx = EncounteredUGS[UGSIdx]) {
     343          12 :           ++UsedGlobalSets[ExpandedIdx].UsageCount;
     344           6 :           GlobalUsesByFunction[ParentFn] = ExpandedIdx;
     345           6 :           continue;
     346             :         }
     347             : 
     348             :         // If not, create a new set consisting of the union of the previous set
     349             :         // and this global.  Mark it as encountered, so we can reuse it later.
     350         144 :         GlobalUsesByFunction[ParentFn] = EncounteredUGS[UGSIdx] =
     351          48 :             UsedGlobalSets.size();
     352             : 
     353          48 :         UsedGlobalSet &NewUGS = CreateGlobalSet();
     354          96 :         NewUGS.Globals.set(GI);
     355          96 :         NewUGS.Globals |= UsedGlobalSets[UGSIdx].Globals;
     356             :       }
     357             :     }
     358             :   }
     359             : 
     360             :   // Now we found a bunch of sets of globals used together.  We accumulated
     361             :   // the number of times we encountered the sets (i.e., the number of blocks
     362             :   // that use that exact set of globals).
     363             :   //
     364             :   // Multiply that by the size of the set to give us a crude profitability
     365             :   // metric.
     366         204 :   std::sort(UsedGlobalSets.begin(), UsedGlobalSets.end(),
     367         223 :             [](const UsedGlobalSet &UGS1, const UsedGlobalSet &UGS2) {
     368         223 :               return UGS1.Globals.count() * UGS1.UsageCount <
     369         446 :                      UGS2.Globals.count() * UGS2.UsageCount;
     370         223 :             });
     371             : 
     372             :   // We can choose to merge all globals together, but ignore globals never used
     373             :   // with another global.  This catches the obviously non-profitable cases of
     374             :   // having a single global, but is aggressive enough for any other case.
     375          68 :   if (GlobalMergeIgnoreSingleUse) {
     376         201 :     BitVector AllGlobals(Globals.size());
     377         283 :     for (size_t i = 0, e = UsedGlobalSets.size(); i != e; ++i) {
     378         298 :       const UsedGlobalSet &UGS = UsedGlobalSets[e - i - 1];
     379         149 :       if (UGS.UsageCount == 0)
     380          98 :         continue;
     381          51 :       if (UGS.Globals.count() > 1)
     382          38 :         AllGlobals |= UGS.Globals;
     383             :     }
     384          67 :     return doMerge(Globals, AllGlobals, M, isConst, AddrSpace);
     385             :   }
     386             : 
     387             :   // Starting from the sets with the best (=biggest) profitability, find a
     388             :   // good combination.
     389             :   // The ideal (and expensive) solution can only be found by trying all
     390             :   // combinations, looking for the one with the best profitability.
     391             :   // Don't be smart about it, and just pick the first compatible combination,
     392             :   // starting with the sets with the best profitability.
     393           2 :   BitVector PickedGlobals(Globals.size());
     394           1 :   bool Changed = false;
     395             : 
     396          14 :   for (size_t i = 0, e = UsedGlobalSets.size(); i != e; ++i) {
     397          24 :     const UsedGlobalSet &UGS = UsedGlobalSets[e - i - 1];
     398          12 :     if (UGS.UsageCount == 0)
     399           7 :       continue;
     400           5 :     if (PickedGlobals.anyCommon(UGS.Globals))
     401           1 :       continue;
     402           4 :     PickedGlobals |= UGS.Globals;
     403             :     // If the set only contains one global, there's no point in merging.
     404             :     // Ignore the global for inclusion in other sets though, so keep it in
     405             :     // PickedGlobals.
     406           8 :     if (UGS.Globals.count() < 2)
     407           1 :       continue;
     408           3 :     Changed |= doMerge(Globals, UGS.Globals, M, isConst, AddrSpace);
     409             :   }
     410             : 
     411           1 :   return Changed;
     412             : }
     413             : 
     414          78 : bool GlobalMerge::doMerge(const SmallVectorImpl<GlobalVariable *> &Globals,
     415             :                           const BitVector &GlobalSet, Module &M, bool isConst,
     416             :                           unsigned AddrSpace) const {
     417             :   assert(Globals.size() > 1);
     418             : 
     419          78 :   Type *Int32Ty = Type::getInt32Ty(M.getContext());
     420          78 :   auto &DL = M.getDataLayout();
     421             : 
     422             :   DEBUG(dbgs() << " Trying to merge set, starts with #"
     423             :                << GlobalSet.find_first() << "\n");
     424             : 
     425          78 :   ssize_t i = GlobalSet.find_first();
     426         158 :   while (i != -1) {
     427          40 :     ssize_t j = 0;
     428          40 :     uint64_t MergedSize = 0;
     429          80 :     std::vector<Type*> Tys;
     430          80 :     std::vector<Constant*> Inits;
     431             : 
     432          40 :     bool HasExternal = false;
     433          40 :     StringRef FirstExternalName;
     434         240 :     for (j = i; j != -1; j = GlobalSet.find_next(j)) {
     435         208 :       Type *Ty = Globals[j]->getValueType();
     436         104 :       MergedSize += DL.getTypeAllocSize(Ty);
     437         104 :       if (MergedSize > MaxOffset) {
     438             :         break;
     439             :       }
     440         100 :       Tys.push_back(Ty);
     441         400 :       Inits.push_back(Globals[j]->getInitializer());
     442             : 
     443         300 :       if (Globals[j]->hasExternalLinkage() && !HasExternal) {
     444          12 :         HasExternal = true;
     445          24 :         FirstExternalName = Globals[j]->getName();
     446             :       }
     447             :     }
     448             : 
     449             :     // If merged variables doesn't have external linkage, we needn't to expose
     450             :     // the symbol after merging.
     451          40 :     GlobalValue::LinkageTypes Linkage = HasExternal
     452          40 :                                             ? GlobalValue::ExternalLinkage
     453             :                                             : GlobalValue::InternalLinkage;
     454          40 :     StructType *MergedTy = StructType::get(M.getContext(), Tys);
     455          40 :     Constant *MergedInit = ConstantStruct::get(MergedTy, Inits);
     456             : 
     457             :     // On Darwin external linkage needs to be preserved, otherwise
     458             :     // dsymutil cannot preserve the debug info for the merged
     459             :     // variables.  If they have external linkage, use the symbol name
     460             :     // of the first variable merged as the suffix of global symbol
     461             :     // name.  This avoids a link-time naming conflict for the
     462             :     // _MergedGlobals symbols.
     463             :     Twine MergedName =
     464          64 :         (IsMachO && HasExternal)
     465          40 :             ? "_MergedGlobals_" + FirstExternalName
     466             :             : "_MergedGlobals";
     467          40 :     auto MergedLinkage = IsMachO ? Linkage : GlobalValue::PrivateLinkage;
     468             :     auto *MergedGV = new GlobalVariable(
     469             :         M, MergedTy, isConst, MergedLinkage, MergedInit, MergedName, nullptr,
     470          40 :         GlobalVariable::NotThreadLocal, AddrSpace);
     471             : 
     472          40 :     const StructLayout *MergedLayout = DL.getStructLayout(MergedTy);
     473             : 
     474         240 :     for (ssize_t k = i, idx = 0; k != j; k = GlobalSet.find_next(k), ++idx) {
     475         300 :       GlobalValue::LinkageTypes Linkage = Globals[k]->getLinkage();
     476         400 :       std::string Name = Globals[k]->getName();
     477             : 
     478             :       // Copy metadata while adjusting any debug info metadata by the original
     479             :       // global's offset within the merged global.
     480         300 :       MergedGV->copyMetadata(Globals[k], MergedLayout->getElementOffset(idx));
     481             : 
     482             :       Constant *Idx[2] = {
     483         100 :         ConstantInt::get(Int32Ty, 0),
     484         100 :         ConstantInt::get(Int32Ty, idx),
     485         200 :       };
     486             :       Constant *GEP =
     487         200 :           ConstantExpr::getInBoundsGetElementPtr(MergedTy, MergedGV, Idx);
     488         200 :       Globals[k]->replaceAllUsesWith(GEP);
     489         200 :       Globals[k]->eraseFromParent();
     490             : 
     491             :       // When the linkage is not internal we must emit an alias for the original
     492             :       // variable name as it may be accessed from another object. On non-Mach-O
     493             :       // we can also emit an alias for internal linkage as it's safe to do so.
     494             :       // It's not safe on Mach-O as the alias (and thus the portion of the
     495             :       // MergedGlobals variable) may be dead stripped at link time.
     496         100 :       if (Linkage != GlobalValue::InternalLinkage || !IsMachO) {
     497          82 :         GlobalAlias::create(Tys[idx], AddrSpace, Linkage, Name, GEP, &M);
     498             :       }
     499             : 
     500         100 :       NumMerged++;
     501             :     }
     502          40 :     i = j;
     503             :   }
     504             : 
     505          78 :   return true;
     506             : }
     507             : 
     508        3230 : void GlobalMerge::collectUsedGlobalVariables(Module &M) {
     509             :   // Extract global variables from llvm.used array
     510        6460 :   const GlobalVariable *GV = M.getGlobalVariable("llvm.used");
     511        3242 :   if (!GV || !GV->hasInitializer()) return;
     512             : 
     513             :   // Should be an array of 'i8*'.
     514          22 :   const ConstantArray *InitList = cast<ConstantArray>(GV->getInitializer());
     515             : 
     516          33 :   for (unsigned i = 0, e = InitList->getNumOperands(); i != e; ++i)
     517             :     if (const GlobalVariable *G =
     518          34 :         dyn_cast<GlobalVariable>(InitList->getOperand(i)->stripPointerCasts()))
     519           1 :       MustKeepGlobalVariables.insert(G);
     520             : }
     521             : 
     522        3230 : void GlobalMerge::setMustKeepGlobalVariables(Module &M) {
     523        3230 :   collectUsedGlobalVariables(M);
     524             : 
     525       38728 :   for (Function &F : M) {
     526      116337 :     for (BasicBlock &BB : F) {
     527       29223 :       Instruction *Pad = BB.getFirstNonPHI();
     528       29223 :       if (!Pad->isEHPad())
     529       29037 :         continue;
     530             : 
     531             :       // Keep globals used by landingpads and catchpads.
     532         470 :       for (const Use &U : Pad->operands()) {
     533             :         if (const GlobalVariable *GV =
     534         113 :                 dyn_cast<GlobalVariable>(U->stripPointerCasts()))
     535          15 :           MustKeepGlobalVariables.insert(GV);
     536             :       }
     537             :     }
     538             :   }
     539        3230 : }
     540             : 
     541        3230 : bool GlobalMerge::doInitialization(Module &M) {
     542        3230 :   if (!EnableGlobalMerge)
     543             :     return false;
     544             : 
     545       12920 :   IsMachO = Triple(M.getTargetTriple()).isOSBinFormatMachO();
     546             : 
     547        3230 :   auto &DL = M.getDataLayout();
     548        9690 :   DenseMap<unsigned, SmallVector<GlobalVariable*, 16> > Globals, ConstGlobals,
     549        6460 :                                                         BSSGlobals;
     550        3230 :   bool Changed = false;
     551        3230 :   setMustKeepGlobalVariables(M);
     552             : 
     553             :   // Grab all non-const globals.
     554        5280 :   for (auto &GV : M.globals()) {
     555             :     // Merge is safe for "normal" internal or external globals only
     556        6901 :     if (GV.isDeclaration() || GV.isThreadLocal() ||
     557        5808 :         GV.hasSection() || GV.hasImplicitSection())
     558        2569 :       continue;
     559             : 
     560             :     // It's not safe to merge globals that may be preempted
     561        1297 :     if (TM && !TM->shouldAssumeDSOLocal(M, &GV))
     562          83 :       continue;
     563             : 
     564        3888 :     if (!(MergeExternalGlobals && GV.hasExternalLinkage()) &&
     565        1924 :         !GV.hasInternalLinkage())
     566         759 :       continue;
     567             : 
     568        1116 :     PointerType *PT = dyn_cast<PointerType>(GV.getType());
     569             :     assert(PT && "Global variable is not a pointer!");
     570             : 
     571         372 :     unsigned AddressSpace = PT->getAddressSpace();
     572             : 
     573             :     // Ignore fancy-aligned globals for now.
     574         372 :     unsigned Alignment = DL.getPreferredAlignment(&GV);
     575         372 :     Type *Ty = GV.getValueType();
     576         424 :     if (Alignment > DL.getABITypeAlignment(Ty))
     577          52 :       continue;
     578             : 
     579             :     // Ignore all 'special' globals.
     580         640 :     if (GV.getName().startswith("llvm.") ||
     581         640 :         GV.getName().startswith(".llvm."))
     582           0 :       continue;
     583             : 
     584             :     // Ignore all "required" globals:
     585         323 :     if (isMustKeepGlobalVariable(&GV))
     586           3 :       continue;
     587             : 
     588         317 :     if (DL.getTypeAllocSize(Ty) < MaxOffset) {
     589         630 :       if (TM &&
     590         943 :           TargetLoweringObjectFile::getKindForGlobal(&GV, *TM).isBSSLocal())
     591          91 :         BSSGlobals[AddressSpace].push_back(&GV);
     592         226 :       else if (GV.isConstant())
     593          41 :         ConstGlobals[AddressSpace].push_back(&GV);
     594             :       else
     595         185 :         Globals[AddressSpace].push_back(&GV);
     596             :     }
     597             :   }
     598             : 
     599        9785 :   for (auto &P : Globals)
     600         190 :     if (P.second.size() > 1)
     601          52 :       Changed |= doMerge(P.second, M, false, P.first);
     602             : 
     603        9730 :   for (auto &P : BSSGlobals)
     604          80 :     if (P.second.size() > 1)
     605          23 :       Changed |= doMerge(P.second, M, false, P.first);
     606             : 
     607        3230 :   if (EnableGlobalMergeOnConst)
     608           4 :     for (auto &P : ConstGlobals)
     609           2 :       if (P.second.size() > 1)
     610           1 :         Changed |= doMerge(P.second, M, true, P.first);
     611             : 
     612        3230 :   return Changed;
     613             : }
     614             : 
     615       21984 : bool GlobalMerge::runOnFunction(Function &F) {
     616       21984 :   return false;
     617             : }
     618             : 
     619        3196 : bool GlobalMerge::doFinalization(Module &M) {
     620        3196 :   MustKeepGlobalVariables.clear();
     621        3196 :   return false;
     622             : }
     623             : 
     624        3235 : Pass *llvm::createGlobalMergePass(const TargetMachine *TM, unsigned Offset,
     625             :                                   bool OnlyOptimizeForSize,
     626             :                                   bool MergeExternalByDefault) {
     627        3246 :   bool MergeExternal = (EnableGlobalMergeOnExternal == cl::BOU_UNSET) ?
     628        3246 :     MergeExternalByDefault : (EnableGlobalMergeOnExternal == cl::BOU_TRUE);
     629        3235 :   return new GlobalMerge(TM, Offset, OnlyOptimizeForSize, MergeExternal);
     630      216918 : }

Generated by: LCOV version 1.13