LCOV - code coverage report
Current view: top level - lib/CodeGen - GlobalMerge.cpp (source / functions) Hit Total Coverage
Test: llvm-toolchain.info Lines: 209 213 98.1 %
Date: 2018-06-17 00:07:59 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             : //
      10             : // This pass merges globals with internal linkage into one. This way all the
      11             : // globals which were merged into a biggest one can be addressed using offsets
      12             : // from the same base pointer (no need for separate base pointer for each of the
      13             : // global). Such a transformation can significantly reduce the register pressure
      14             : // when many globals are involved.
      15             : //
      16             : // For example, consider the code which touches several global variables at
      17             : // once:
      18             : //
      19             : // static int foo[N], bar[N], baz[N];
      20             : //
      21             : // for (i = 0; i < N; ++i) {
      22             : //    foo[i] = bar[i] * baz[i];
      23             : // }
      24             : //
      25             : //  On ARM the addresses of 3 arrays should be kept in the registers, thus
      26             : //  this code has quite large register pressure (loop body):
      27             : //
      28             : //  ldr     r1, [r5], #4
      29             : //  ldr     r2, [r6], #4
      30             : //  mul     r1, r2, r1
      31             : //  str     r1, [r0], #4
      32             : //
      33             : //  Pass converts the code to something like:
      34             : //
      35             : //  static struct {
      36             : //    int foo[N];
      37             : //    int bar[N];
      38             : //    int baz[N];
      39             : //  } merged;
      40             : //
      41             : //  for (i = 0; i < N; ++i) {
      42             : //    merged.foo[i] = merged.bar[i] * merged.baz[i];
      43             : //  }
      44             : //
      45             : //  and in ARM code this becomes:
      46             : //
      47             : //  ldr     r0, [r5, #40]
      48             : //  ldr     r1, [r5, #80]
      49             : //  mul     r0, r1, r0
      50             : //  str     r0, [r5], #4
      51             : //
      52             : //  note that we saved 2 registers here almostly "for free".
      53             : //
      54             : // However, merging globals can have tradeoffs:
      55             : // - it confuses debuggers, tools, and users
      56             : // - it makes linker optimizations less useful (order files, LOHs, ...)
      57             : // - it forces usage of indexed addressing (which isn't necessarily "free")
      58             : // - it can increase register pressure when the uses are disparate enough.
      59             : // 
      60             : // We use heuristics to discover the best global grouping we can (cf cl::opts).
      61             : //
      62             : // ===---------------------------------------------------------------------===//
      63             : 
      64             : #include "llvm/ADT/BitVector.h"
      65             : #include "llvm/ADT/DenseMap.h"
      66             : #include "llvm/ADT/SmallPtrSet.h"
      67             : #include "llvm/ADT/SmallVector.h"
      68             : #include "llvm/ADT/Statistic.h"
      69             : #include "llvm/ADT/StringRef.h"
      70             : #include "llvm/ADT/Triple.h"
      71             : #include "llvm/ADT/Twine.h"
      72             : #include "llvm/CodeGen/Passes.h"
      73             : #include "llvm/IR/BasicBlock.h"
      74             : #include "llvm/IR/Constants.h"
      75             : #include "llvm/IR/DataLayout.h"
      76             : #include "llvm/IR/DerivedTypes.h"
      77             : #include "llvm/IR/Function.h"
      78             : #include "llvm/IR/GlobalAlias.h"
      79             : #include "llvm/IR/GlobalValue.h"
      80             : #include "llvm/IR/GlobalVariable.h"
      81             : #include "llvm/IR/Instruction.h"
      82             : #include "llvm/IR/Module.h"
      83             : #include "llvm/IR/Type.h"
      84             : #include "llvm/IR/Use.h"
      85             : #include "llvm/IR/User.h"
      86             : #include "llvm/Pass.h"
      87             : #include "llvm/Support/Casting.h"
      88             : #include "llvm/Support/CommandLine.h"
      89             : #include "llvm/Support/Debug.h"
      90             : #include "llvm/Support/raw_ostream.h"
      91             : #include "llvm/Target/TargetLoweringObjectFile.h"
      92             : #include "llvm/Target/TargetMachine.h"
      93             : #include <algorithm>
      94             : #include <cassert>
      95             : #include <cstddef>
      96             : #include <cstdint>
      97             : #include <string>
      98             : #include <vector>
      99             : 
     100             : using namespace llvm;
     101             : 
     102             : #define DEBUG_TYPE "global-merge"
     103             : 
     104             : // FIXME: This is only useful as a last-resort way to disable the pass.
     105             : static cl::opt<bool>
     106      101169 : EnableGlobalMerge("enable-global-merge", cl::Hidden,
     107      101169 :                   cl::desc("Enable the global merge pass"),
     108      303507 :                   cl::init(true));
     109             : 
     110             : static cl::opt<unsigned>
     111      101169 : GlobalMergeMaxOffset("global-merge-max-offset", cl::Hidden,
     112      101169 :                      cl::desc("Set maximum offset for global merge pass"),
     113      303507 :                      cl::init(0));
     114             : 
     115      101169 : static cl::opt<bool> GlobalMergeGroupByUse(
     116             :     "global-merge-group-by-use", cl::Hidden,
     117      202338 :     cl::desc("Improve global merge pass to look at uses"), cl::init(true));
     118             : 
     119      101169 : static cl::opt<bool> GlobalMergeIgnoreSingleUse(
     120             :     "global-merge-ignore-single-use", cl::Hidden,
     121      101169 :     cl::desc("Improve global merge pass to ignore globals only used alone"),
     122      303507 :     cl::init(true));
     123             : 
     124             : static cl::opt<bool>
     125      101169 : EnableGlobalMergeOnConst("global-merge-on-const", cl::Hidden,
     126      101169 :                          cl::desc("Enable global merge pass on constants"),
     127      303507 :                          cl::init(false));
     128             : 
     129             : // FIXME: this could be a transitional option, and we probably need to remove
     130             : // it if only we are sure this optimization could always benefit all targets.
     131             : static cl::opt<cl::boolOrDefault>
     132      101169 : EnableGlobalMergeOnExternal("global-merge-on-external", cl::Hidden,
     133      101169 :      cl::desc("Enable global merge pass on external linkage"));
     134             : 
     135             : STATISTIC(NumMerged, "Number of globals merged");
     136             : 
     137             : namespace {
     138             : 
     139       10395 :   class GlobalMerge : public FunctionPass {
     140             :     const TargetMachine *TM = nullptr;
     141             : 
     142             :     // FIXME: Infer the maximum possible offset depending on the actual users
     143             :     // (these max offsets are different for the users inside Thumb or ARM
     144             :     // functions), see the code that passes in the offset in the ARM backend
     145             :     // for more information.
     146             :     unsigned MaxOffset;
     147             : 
     148             :     /// Whether we should try to optimize for size only.
     149             :     /// Currently, this applies a dead simple heuristic: only consider globals
     150             :     /// used in minsize functions for merging.
     151             :     /// FIXME: This could learn about optsize, and be used in the cost model.
     152             :     bool OnlyOptimizeForSize = false;
     153             : 
     154             :     /// Whether we should merge global variables that have external linkage.
     155             :     bool MergeExternalGlobals = false;
     156             : 
     157             :     bool IsMachO;
     158             : 
     159             :     bool doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
     160             :                  Module &M, bool isConst, unsigned AddrSpace) const;
     161             : 
     162             :     /// Merge everything in \p Globals for which the corresponding bit
     163             :     /// in \p GlobalSet is set.
     164             :     bool doMerge(const SmallVectorImpl<GlobalVariable *> &Globals,
     165             :                  const BitVector &GlobalSet, Module &M, bool isConst,
     166             :                  unsigned AddrSpace) const;
     167             : 
     168             :     /// Check if the given variable has been identified as must keep
     169             :     /// \pre setMustKeepGlobalVariables must have been called on the Module that
     170             :     ///      contains GV
     171             :     bool isMustKeepGlobalVariable(const GlobalVariable *GV) const {
     172         336 :       return MustKeepGlobalVariables.count(GV);
     173             :     }
     174             : 
     175             :     /// Collect every variables marked as "used" or used in a landing pad
     176             :     /// instruction for this Module.
     177             :     void setMustKeepGlobalVariables(Module &M);
     178             : 
     179             :     /// Collect every variables marked as "used"
     180             :     void collectUsedGlobalVariables(Module &M);
     181             : 
     182             :     /// Keep track of the GlobalVariable that must not be merged away
     183             :     SmallPtrSet<const GlobalVariable *, 16> MustKeepGlobalVariables;
     184             : 
     185             :   public:
     186             :     static char ID;             // Pass identification, replacement for typeid.
     187             : 
     188           2 :     explicit GlobalMerge()
     189           6 :         : FunctionPass(ID), MaxOffset(GlobalMergeMaxOffset) {
     190           2 :       initializeGlobalMergePass(*PassRegistry::getPassRegistry());
     191           2 :     }
     192             : 
     193        3497 :     explicit GlobalMerge(const TargetMachine *TM, unsigned MaximalOffset,
     194             :                          bool OnlyOptimizeForSize, bool MergeExternalGlobals)
     195        3497 :         : FunctionPass(ID), TM(TM), MaxOffset(MaximalOffset),
     196             :           OnlyOptimizeForSize(OnlyOptimizeForSize),
     197        3497 :           MergeExternalGlobals(MergeExternalGlobals) {
     198        3497 :       initializeGlobalMergePass(*PassRegistry::getPassRegistry());
     199        3497 :     }
     200             : 
     201             :     bool doInitialization(Module &M) override;
     202             :     bool runOnFunction(Function &F) override;
     203             :     bool doFinalization(Module &M) override;
     204             : 
     205           7 :     StringRef getPassName() const override { return "Merge internal globals"; }
     206             : 
     207        3488 :     void getAnalysisUsage(AnalysisUsage &AU) const override {
     208        3488 :       AU.setPreservesCFG();
     209        3488 :       FunctionPass::getAnalysisUsage(AU);
     210        3488 :     }
     211             :   };
     212             : 
     213             : } // end anonymous namespace
     214             : 
     215             : char GlobalMerge::ID = 0;
     216             : 
     217       54980 : INITIALIZE_PASS(GlobalMerge, DEBUG_TYPE, "Merge global variables", false, false)
     218             : 
     219          81 : bool GlobalMerge::doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
     220             :                           Module &M, bool isConst, unsigned AddrSpace) const {
     221          81 :   auto &DL = M.getDataLayout();
     222             :   // FIXME: Find better heuristics
     223             :   std::stable_sort(Globals.begin(), Globals.end(),
     224         482 :                    [&DL](const GlobalVariable *GV1, const GlobalVariable *GV2) {
     225         241 :                      return DL.getTypeAllocSize(GV1->getValueType()) <
     226         241 :                             DL.getTypeAllocSize(GV2->getValueType());
     227         241 :                    });
     228             : 
     229             :   // If we want to just blindly group all globals together, do so.
     230          81 :   if (!GlobalMergeGroupByUse) {
     231           8 :     BitVector AllGlobals(Globals.size());
     232           8 :     AllGlobals.set();
     233           8 :     return doMerge(Globals, AllGlobals, M, isConst, AddrSpace);
     234             :   }
     235             : 
     236             :   // If we want to be smarter, look at all uses of each global, to try to
     237             :   // discover all sets of globals used together, and how many times each of
     238             :   // these sets occurred.
     239             :   //
     240             :   // Keep this reasonably efficient, by having an append-only list of all sets
     241             :   // discovered so far (UsedGlobalSet), and mapping each "together-ness" unit of
     242             :   // code (currently, a Function) to the set of globals seen so far that are
     243             :   // used together in that unit (GlobalUsesByFunction).
     244             :   //
     245             :   // When we look at the Nth global, we know that any new set is either:
     246             :   // - the singleton set {N}, containing this global only, or
     247             :   // - the union of {N} and a previously-discovered set, containing some
     248             :   //   combination of the previous N-1 globals.
     249             :   // Using that knowledge, when looking at the Nth global, we can keep:
     250             :   // - a reference to the singleton set {N} (CurGVOnlySetIdx)
     251             :   // - a list mapping each previous set to its union with {N} (EncounteredUGS),
     252             :   //   if it actually occurs.
     253             : 
     254             :   // We keep track of the sets of globals used together "close enough".
     255        1440 :   struct UsedGlobalSet {
     256             :     BitVector Globals;
     257             :     unsigned UsageCount = 1;
     258             : 
     259         176 :     UsedGlobalSet(size_t Size) : Globals(Size) {}
     260             :   };
     261             : 
     262             :   // Each set is unique in UsedGlobalSets.
     263          73 :   std::vector<UsedGlobalSet> UsedGlobalSets;
     264             : 
     265             :   // Avoid repeating the create-global-set pattern.
     266         176 :   auto CreateGlobalSet = [&]() -> UsedGlobalSet & {
     267         352 :     UsedGlobalSets.emplace_back(Globals.size());
     268         176 :     return UsedGlobalSets.back();
     269          73 :   };
     270             : 
     271             :   // The first set is the empty set.
     272          73 :   CreateGlobalSet().UsageCount = 0;
     273             : 
     274             :   // We define "close enough" to be "in the same function".
     275             :   // FIXME: Grouping uses by function is way too aggressive, so we should have
     276             :   // a better metric for distance between uses.
     277             :   // The obvious alternative would be to group by BasicBlock, but that's in
     278             :   // turn too conservative..
     279             :   // Anything in between wouldn't be trivial to compute, so just stick with
     280             :   // per-function grouping.
     281             : 
     282             :   // The value type is an index into UsedGlobalSets.
     283             :   // The default (0) conveniently points to the empty set.
     284             :   DenseMap<Function *, size_t /*UsedGlobalSetIdx*/> GlobalUsesByFunction;
     285             : 
     286             :   // Now, look at each merge-eligible global in turn.
     287             : 
     288             :   // Keep track of the sets we already encountered to which we added the
     289             :   // current global.
     290             :   // Each element matches the same-index element in UsedGlobalSets.
     291             :   // This lets us efficiently tell whether a set has already been expanded to
     292             :   // include the current global.
     293             :   std::vector<size_t> EncounteredUGS;
     294             : 
     295         489 :   for (size_t GI = 0, GE = Globals.size(); GI != GE; ++GI) {
     296         208 :     GlobalVariable *GV = Globals[GI];
     297             : 
     298             :     // Reset the encountered sets for this global...
     299             :     std::fill(EncounteredUGS.begin(), EncounteredUGS.end(), 0);
     300             :     // ...and grow it in case we created new sets for the previous global.
     301         416 :     EncounteredUGS.resize(UsedGlobalSets.size());
     302             : 
     303             :     // We might need to create a set that only consists of the current global.
     304             :     // Keep track of its index into UsedGlobalSets.
     305             :     size_t CurGVOnlySetIdx = 0;
     306             : 
     307             :     // For each global, look at all its Uses.
     308         840 :     for (auto &U : GV->uses()) {
     309             :       // This Use might be a ConstantExpr.  We're interested in Instruction
     310             :       // users, so look through ConstantExpr...
     311             :       Use *UI, *UE;
     312         632 :       if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U.getUser())) {
     313         107 :         if (CE->use_empty())
     314           0 :           continue;
     315             :         UI = &*CE->use_begin();
     316             :         UE = nullptr;
     317         525 :       } else if (isa<Instruction>(U.getUser())) {
     318             :         UI = &U;
     319         525 :         UE = UI->getNext();
     320             :       } else {
     321           0 :         continue;
     322             :       }
     323             : 
     324             :       // ...to iterate on all the instruction users of the global.
     325             :       // Note that we iterate on Uses and not on Users to be able to getNext().
     326        1938 :       for (; UI != UE; UI = UI->getNext()) {
     327         653 :         Instruction *I = dyn_cast<Instruction>(UI->getUser());
     328           0 :         if (!I)
     329         600 :           continue;
     330             : 
     331         653 :         Function *ParentFn = I->getParent()->getParent();
     332             : 
     333             :         // If we're only optimizing for size, ignore non-minsize functions.
     334        1605 :         if (OnlyOptimizeForSize && !ParentFn->optForMinSize())
     335         472 :           continue;
     336             : 
     337         181 :         size_t UGSIdx = GlobalUsesByFunction[ParentFn];
     338             : 
     339             :         // If this is the first global the basic block uses, map it to the set
     340             :         // consisting of this global only.
     341         237 :         if (!UGSIdx) {
     342             :           // If that set doesn't exist yet, create it.
     343          56 :           if (!CurGVOnlySetIdx) {
     344          50 :             CurGVOnlySetIdx = UsedGlobalSets.size();
     345          50 :             CreateGlobalSet().Globals.set(GI);
     346             :           } else {
     347          12 :             ++UsedGlobalSets[CurGVOnlySetIdx].UsageCount;
     348             :           }
     349             : 
     350          56 :           GlobalUsesByFunction[ParentFn] = CurGVOnlySetIdx;
     351          56 :           continue;
     352             :         }
     353             : 
     354             :         // If we already encountered this BB, just increment the counter.
     355         441 :         if (UsedGlobalSets[UGSIdx].Globals.test(GI)) {
     356          66 :           ++UsedGlobalSets[UGSIdx].UsageCount;
     357          66 :           continue;
     358             :         }
     359             : 
     360             :         // If not, the previous set wasn't actually used in this function.
     361          59 :         --UsedGlobalSets[UGSIdx].UsageCount;
     362             : 
     363             :         // If we already expanded the previous set to include this global, just
     364             :         // reuse that expanded set.
     365         124 :         if (size_t ExpandedIdx = EncounteredUGS[UGSIdx]) {
     366           6 :           ++UsedGlobalSets[ExpandedIdx].UsageCount;
     367           6 :           GlobalUsesByFunction[ParentFn] = ExpandedIdx;
     368           6 :           continue;
     369             :         }
     370             : 
     371             :         // If not, create a new set consisting of the union of the previous set
     372             :         // and this global.  Mark it as encountered, so we can reuse it later.
     373         106 :         GlobalUsesByFunction[ParentFn] = EncounteredUGS[UGSIdx] =
     374          53 :             UsedGlobalSets.size();
     375             : 
     376          53 :         UsedGlobalSet &NewUGS = CreateGlobalSet();
     377             :         NewUGS.Globals.set(GI);
     378         106 :         NewUGS.Globals |= UsedGlobalSets[UGSIdx].Globals;
     379             :       }
     380             :     }
     381             :   }
     382             : 
     383             :   // Now we found a bunch of sets of globals used together.  We accumulated
     384             :   // the number of times we encountered the sets (i.e., the number of blocks
     385             :   // that use that exact set of globals).
     386             :   //
     387             :   // Multiply that by the size of the set to give us a crude profitability
     388             :   // metric.
     389             :   std::stable_sort(UsedGlobalSets.begin(), UsedGlobalSets.end(),
     390         221 :             [](const UsedGlobalSet &UGS1, const UsedGlobalSet &UGS2) {
     391         221 :               return UGS1.Globals.count() * UGS1.UsageCount <
     392         221 :                      UGS2.Globals.count() * UGS2.UsageCount;
     393         221 :             });
     394             : 
     395             :   // We can choose to merge all globals together, but ignore globals never used
     396             :   // with another global.  This catches the obviously non-profitable cases of
     397             :   // having a single global, but is aggressive enough for any other case.
     398          73 :   if (GlobalMergeIgnoreSingleUse) {
     399          72 :     BitVector AllGlobals(Globals.size());
     400         472 :     for (size_t i = 0, e = UsedGlobalSets.size(); i != e; ++i) {
     401         164 :       const UsedGlobalSet &UGS = UsedGlobalSets[e - i - 1];
     402         164 :       if (UGS.UsageCount == 0)
     403         108 :         continue;
     404          56 :       if (UGS.Globals.count() > 1)
     405          43 :         AllGlobals |= UGS.Globals;
     406             :     }
     407          72 :     return doMerge(Globals, AllGlobals, M, isConst, AddrSpace);
     408             :   }
     409             : 
     410             :   // Starting from the sets with the best (=biggest) profitability, find a
     411             :   // good combination.
     412             :   // The ideal (and expensive) solution can only be found by trying all
     413             :   // combinations, looking for the one with the best profitability.
     414             :   // Don't be smart about it, and just pick the first compatible combination,
     415             :   // starting with the sets with the best profitability.
     416           1 :   BitVector PickedGlobals(Globals.size());
     417             :   bool Changed = false;
     418             : 
     419          26 :   for (size_t i = 0, e = UsedGlobalSets.size(); i != e; ++i) {
     420          12 :     const UsedGlobalSet &UGS = UsedGlobalSets[e - i - 1];
     421          12 :     if (UGS.UsageCount == 0)
     422           7 :       continue;
     423           5 :     if (PickedGlobals.anyCommon(UGS.Globals))
     424           1 :       continue;
     425           4 :     PickedGlobals |= UGS.Globals;
     426             :     // If the set only contains one global, there's no point in merging.
     427             :     // Ignore the global for inclusion in other sets though, so keep it in
     428             :     // PickedGlobals.
     429           4 :     if (UGS.Globals.count() < 2)
     430           1 :       continue;
     431           3 :     Changed |= doMerge(Globals, UGS.Globals, M, isConst, AddrSpace);
     432             :   }
     433             : 
     434             :   return Changed;
     435             : }
     436             : 
     437          83 : bool GlobalMerge::doMerge(const SmallVectorImpl<GlobalVariable *> &Globals,
     438             :                           const BitVector &GlobalSet, Module &M, bool isConst,
     439             :                           unsigned AddrSpace) const {
     440             :   assert(Globals.size() > 1);
     441             : 
     442          83 :   Type *Int32Ty = Type::getInt32Ty(M.getContext());
     443          83 :   auto &DL = M.getDataLayout();
     444             : 
     445             :   LLVM_DEBUG(dbgs() << " Trying to merge set, starts with #"
     446             :                     << GlobalSet.find_first() << "\n");
     447             : 
     448             :   bool Changed = false;
     449          83 :   ssize_t i = GlobalSet.find_first();
     450         127 :   while (i != -1) {
     451             :     ssize_t j = 0;
     452             :     uint64_t MergedSize = 0;
     453             :     std::vector<Type*> Tys;
     454             :     std::vector<Constant*> Inits;
     455             : 
     456             :     bool HasExternal = false;
     457          44 :     StringRef FirstExternalName;
     458         371 :     for (j = i; j != -1; j = GlobalSet.find_next(j)) {
     459         226 :       Type *Ty = Globals[j]->getValueType();
     460         113 :       MergedSize += DL.getTypeAllocSize(Ty);
     461         113 :       if (MergedSize > MaxOffset) {
     462             :         break;
     463             :       }
     464         109 :       Tys.push_back(Ty);
     465         327 :       Inits.push_back(Globals[j]->getInitializer());
     466             : 
     467         218 :       if (Globals[j]->hasExternalLinkage() && !HasExternal) {
     468             :         HasExternal = true;
     469          15 :         FirstExternalName = Globals[j]->getName();
     470             :       }
     471             :     }
     472             : 
     473             :     // Exit early if there is only one global to merge.
     474          88 :     if (Tys.size() < 2) {
     475             :       i = j;
     476             :       continue;
     477             :     }
     478             : 
     479             :     // If merged variables doesn't have external linkage, we needn't to expose
     480             :     // the symbol after merging.
     481             :     GlobalValue::LinkageTypes Linkage = HasExternal
     482          40 :                                             ? GlobalValue::ExternalLinkage
     483             :                                             : GlobalValue::InternalLinkage;
     484          40 :     StructType *MergedTy = StructType::get(M.getContext(), Tys);
     485          40 :     Constant *MergedInit = ConstantStruct::get(MergedTy, Inits);
     486             : 
     487             :     // On Darwin external linkage needs to be preserved, otherwise
     488             :     // dsymutil cannot preserve the debug info for the merged
     489             :     // variables.  If they have external linkage, use the symbol name
     490             :     // of the first variable merged as the suffix of global symbol
     491             :     // name.  This avoids a link-time naming conflict for the
     492             :     // _MergedGlobals symbols.
     493             :     Twine MergedName =
     494          62 :         (IsMachO && HasExternal)
     495          40 :             ? "_MergedGlobals_" + FirstExternalName
     496             :             : "_MergedGlobals";
     497          40 :     auto MergedLinkage = IsMachO ? Linkage : GlobalValue::PrivateLinkage;
     498             :     auto *MergedGV = new GlobalVariable(
     499             :         M, MergedTy, isConst, MergedLinkage, MergedInit, MergedName, nullptr,
     500          40 :         GlobalVariable::NotThreadLocal, AddrSpace);
     501             : 
     502          40 :     const StructLayout *MergedLayout = DL.getStructLayout(MergedTy);
     503             :     // Set the alignment of the merged struct as the maximum alignment of the
     504             :     // globals to prevent over-alignment. We don't handle globals that are not
     505             :     // default aligned, so the alignment of the MergedLayout struct is
     506             :     // equivalent.
     507          40 :     MergedGV->setAlignment(MergedLayout->getAlignment());
     508             : 
     509         355 :     for (ssize_t k = i, idx = 0; k != j; k = GlobalSet.find_next(k), ++idx) {
     510         210 :       GlobalValue::LinkageTypes Linkage = Globals[k]->getLinkage();
     511         210 :       std::string Name = Globals[k]->getName();
     512             :       GlobalValue::DLLStorageClassTypes DLLStorage =
     513         105 :           Globals[k]->getDLLStorageClass();
     514             : 
     515             :       // Copy metadata while adjusting any debug info metadata by the original
     516             :       // global's offset within the merged global.
     517         210 :       MergedGV->copyMetadata(Globals[k], MergedLayout->getElementOffset(idx));
     518             : 
     519             :       Constant *Idx[2] = {
     520         105 :         ConstantInt::get(Int32Ty, 0),
     521         105 :         ConstantInt::get(Int32Ty, idx),
     522         210 :       };
     523             :       Constant *GEP =
     524             :           ConstantExpr::getInBoundsGetElementPtr(MergedTy, MergedGV, Idx);
     525         105 :       Globals[k]->replaceAllUsesWith(GEP);
     526         105 :       Globals[k]->eraseFromParent();
     527             : 
     528             :       // When the linkage is not internal we must emit an alias for the original
     529             :       // variable name as it may be accessed from another object. On non-Mach-O
     530             :       // we can also emit an alias for internal linkage as it's safe to do so.
     531             :       // It's not safe on Mach-O as the alias (and thus the portion of the
     532             :       // MergedGlobals variable) may be dead stripped at link time.
     533         105 :       if (Linkage != GlobalValue::InternalLinkage || !IsMachO) {
     534             :         GlobalAlias *GA =
     535          94 :             GlobalAlias::create(Tys[idx], AddrSpace, Linkage, Name, GEP, &M);
     536             :         GA->setDLLStorageClass(DLLStorage);
     537             :       }
     538             : 
     539             :       NumMerged++;
     540             :     }
     541             :     Changed = true;
     542             :     i = j;
     543             :   }
     544             : 
     545          83 :   return Changed;
     546             : }
     547             : 
     548        3489 : void GlobalMerge::collectUsedGlobalVariables(Module &M) {
     549             :   // Extract global variables from llvm.used array
     550             :   const GlobalVariable *GV = M.getGlobalVariable("llvm.used");
     551        3502 :   if (!GV || !GV->hasInitializer()) return;
     552             : 
     553             :   // Should be an array of 'i8*'.
     554             :   const ConstantArray *InitList = cast<ConstantArray>(GV->getInitializer());
     555             : 
     556          40 :   for (unsigned i = 0, e = InitList->getNumOperands(); i != e; ++i)
     557             :     if (const GlobalVariable *G =
     558          14 :         dyn_cast<GlobalVariable>(InitList->getOperand(i)->stripPointerCasts()))
     559           4 :       MustKeepGlobalVariables.insert(G);
     560             : }
     561             : 
     562        3489 : void GlobalMerge::setMustKeepGlobalVariables(Module &M) {
     563        3489 :   collectUsedGlobalVariables(M);
     564             : 
     565       36958 :   for (Function &F : M) {
     566       66877 :     for (BasicBlock &BB : F) {
     567             :       Instruction *Pad = BB.getFirstNonPHI();
     568       33212 :       if (!Pad->isEHPad())
     569       33212 :         continue;
     570             : 
     571             :       // Keep globals used by landingpads and catchpads.
     572         600 :       for (const Use &U : Pad->operands()) {
     573             :         if (const GlobalVariable *GV =
     574         104 :                 dyn_cast<GlobalVariable>(U->stripPointerCasts()))
     575          18 :           MustKeepGlobalVariables.insert(GV);
     576             :       }
     577             :     }
     578             :   }
     579        3489 : }
     580             : 
     581        3489 : bool GlobalMerge::doInitialization(Module &M) {
     582        3489 :   if (!EnableGlobalMerge)
     583             :     return false;
     584             : 
     585       10467 :   IsMachO = Triple(M.getTargetTriple()).isOSBinFormatMachO();
     586             : 
     587        3489 :   auto &DL = M.getDataLayout();
     588             :   DenseMap<unsigned, SmallVector<GlobalVariable *, 16>> Globals, ConstGlobals,
     589             :                                                         BSSGlobals;
     590             :   bool Changed = false;
     591        3489 :   setMustKeepGlobalVariables(M);
     592             : 
     593             :   // Grab all non-const globals.
     594        5757 :   for (auto &GV : M.globals()) {
     595             :     // Merge is safe for "normal" internal or external globals only
     596        6089 :     if (GV.isDeclaration() || GV.isThreadLocal() ||
     597        3586 :         GV.hasSection() || GV.hasImplicitSection())
     598        2901 :       continue;
     599             : 
     600             :     // It's not safe to merge globals that may be preempted
     601        1390 :     if (TM && !TM->shouldAssumeDSOLocal(M, &GV))
     602          86 :       continue;
     603             : 
     604        3631 :     if (!(MergeExternalGlobals && GV.hasExternalLinkage()) &&
     605             :         !GV.hasInternalLinkage())
     606         830 :       continue;
     607             : 
     608             :     PointerType *PT = dyn_cast<PointerType>(GV.getType());
     609             :     assert(PT && "Global variable is not a pointer!");
     610             : 
     611         388 :     unsigned AddressSpace = PT->getAddressSpace();
     612             : 
     613             :     // Ignore fancy-aligned globals for now.
     614         388 :     unsigned Alignment = DL.getPreferredAlignment(&GV);
     615         388 :     Type *Ty = GV.getValueType();
     616         440 :     if (Alignment > DL.getABITypeAlignment(Ty))
     617          52 :       continue;
     618             : 
     619             :     // Ignore all 'special' globals.
     620         672 :     if (GV.getName().startswith("llvm.") ||
     621         672 :         GV.getName().startswith(".llvm."))
     622           0 :       continue;
     623             : 
     624             :     // Ignore all "required" globals:
     625         341 :     if (isMustKeepGlobalVariable(&GV))
     626           5 :       continue;
     627             : 
     628         331 :     if (DL.getTypeAllocSize(Ty) < MaxOffset) {
     629         658 :       if (TM &&
     630         658 :           TargetLoweringObjectFile::getKindForGlobal(&GV, *TM).isBSSLocal())
     631          93 :         BSSGlobals[AddressSpace].push_back(&GV);
     632         238 :       else if (GV.isConstant())
     633          41 :         ConstGlobals[AddressSpace].push_back(&GV);
     634             :       else
     635         197 :         Globals[AddressSpace].push_back(&GV);
     636             :     }
     637             :   }
     638             : 
     639        7078 :   for (auto &P : Globals)
     640         100 :     if (P.second.size() > 1)
     641          56 :       Changed |= doMerge(P.second, M, false, P.first);
     642             : 
     643        7019 :   for (auto &P : BSSGlobals)
     644          41 :     if (P.second.size() > 1)
     645          24 :       Changed |= doMerge(P.second, M, false, P.first);
     646             : 
     647        3489 :   if (EnableGlobalMergeOnConst)
     648           3 :     for (auto &P : ConstGlobals)
     649           1 :       if (P.second.size() > 1)
     650           1 :         Changed |= doMerge(P.second, M, true, P.first);
     651             : 
     652             :   return Changed;
     653             : }
     654             : 
     655       25851 : bool GlobalMerge::runOnFunction(Function &F) {
     656       25851 :   return false;
     657             : }
     658             : 
     659        3455 : bool GlobalMerge::doFinalization(Module &M) {
     660        3455 :   MustKeepGlobalVariables.clear();
     661        3455 :   return false;
     662             : }
     663             : 
     664        3497 : Pass *llvm::createGlobalMergePass(const TargetMachine *TM, unsigned Offset,
     665             :                                   bool OnlyOptimizeForSize,
     666             :                                   bool MergeExternalByDefault) {
     667        3497 :   bool MergeExternal = (EnableGlobalMergeOnExternal == cl::BOU_UNSET) ?
     668             :     MergeExternalByDefault : (EnableGlobalMergeOnExternal == cl::BOU_TRUE);
     669        3497 :   return new GlobalMerge(TM, Offset, OnlyOptimizeForSize, MergeExternal);
     670      303507 : }

Generated by: LCOV version 1.13