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
54#define DEBUG_TYPE "global-merge"
55#include "llvm/Transforms/Scalar.h"
56#include "llvm/ADT/SmallPtrSet.h"
57#include "llvm/ADT/Statistic.h"
58#include "llvm/IR/Attributes.h"
59#include "llvm/IR/Constants.h"
60#include "llvm/IR/DataLayout.h"
61#include "llvm/IR/DerivedTypes.h"
62#include "llvm/IR/Function.h"
63#include "llvm/IR/GlobalVariable.h"
64#include "llvm/IR/Instructions.h"
65#include "llvm/IR/Intrinsics.h"
66#include "llvm/IR/Module.h"
67#include "llvm/Pass.h"
68#include "llvm/Support/CommandLine.h"
69#include "llvm/Target/TargetLowering.h"
70#include "llvm/Target/TargetLoweringObjectFile.h"
71using namespace llvm;
72
73static cl::opt<bool>
74EnableGlobalMergeOnConst("global-merge-on-const", cl::Hidden,
75                  	cl::desc("Enable global merge pass on constants"),
76                  	cl::init(false));
77
78STATISTIC(NumMerged      , "Number of globals merged");
79namespace {
80  class GlobalMerge : public FunctionPass {
81    /// TLI - Keep a pointer of a TargetLowering to consult for determining
82    /// target type sizes.
83    const TargetLowering *TLI;
84
85    bool doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
86                 Module &M, bool isConst, unsigned AddrSpace) const;
87
88    /// \brief Check if the given variable has been identified as must keep
89    /// \pre setMustKeepGlobalVariables must have been called on the Module that
90    ///      contains GV
91    bool isMustKeepGlobalVariable(const GlobalVariable *GV) const {
92      return MustKeepGlobalVariables.count(GV);
93    }
94
95    /// Collect every variables marked as "used" or used in a landing pad
96    /// instruction for this Module.
97    void setMustKeepGlobalVariables(Module &M);
98
99    /// Collect every variables marked as "used"
100    void collectUsedGlobalVariables(Module &M);
101
102    /// Keep track of the GlobalVariable that must not be merged away
103    SmallPtrSet<const GlobalVariable *, 16> MustKeepGlobalVariables;
104
105  public:
106    static char ID;             // Pass identification, replacement for typeid.
107    explicit GlobalMerge(const TargetLowering *tli = 0)
108      : FunctionPass(ID), TLI(tli) {
109      initializeGlobalMergePass(*PassRegistry::getPassRegistry());
110    }
111
112    virtual bool doInitialization(Module &M);
113    virtual bool runOnFunction(Function &F);
114    virtual bool doFinalization(Module &M);
115
116    const char *getPassName() const {
117      return "Merge internal globals";
118    }
119
120    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
121      AU.setPreservesCFG();
122      FunctionPass::getAnalysisUsage(AU);
123    }
124
125    struct GlobalCmp {
126      const DataLayout *TD;
127
128      GlobalCmp(const DataLayout *td) : TD(td) { }
129
130      bool operator()(const GlobalVariable *GV1, const GlobalVariable *GV2) {
131        Type *Ty1 = cast<PointerType>(GV1->getType())->getElementType();
132        Type *Ty2 = cast<PointerType>(GV2->getType())->getElementType();
133
134        return (TD->getTypeAllocSize(Ty1) < TD->getTypeAllocSize(Ty2));
135      }
136    };
137  };
138} // end anonymous namespace
139
140char GlobalMerge::ID = 0;
141INITIALIZE_PASS(GlobalMerge, "global-merge",
142                "Global Merge", false, false)
143
144
145bool GlobalMerge::doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
146                          Module &M, bool isConst, unsigned AddrSpace) const {
147  const DataLayout *TD = TLI->getDataLayout();
148
149  // FIXME: Infer the maximum possible offset depending on the actual users
150  // (these max offsets are different for the users inside Thumb or ARM
151  // functions)
152  unsigned MaxOffset = TLI->getMaximalGlobalOffset();
153
154  // FIXME: Find better heuristics
155  std::stable_sort(Globals.begin(), Globals.end(), GlobalCmp(TD));
156
157  Type *Int32Ty = Type::getInt32Ty(M.getContext());
158
159  for (size_t i = 0, e = Globals.size(); i != e; ) {
160    size_t j = 0;
161    uint64_t MergedSize = 0;
162    std::vector<Type*> Tys;
163    std::vector<Constant*> Inits;
164    for (j = i; j != e; ++j) {
165      Type *Ty = Globals[j]->getType()->getElementType();
166      MergedSize += TD->getTypeAllocSize(Ty);
167      if (MergedSize > MaxOffset) {
168        break;
169      }
170      Tys.push_back(Ty);
171      Inits.push_back(Globals[j]->getInitializer());
172    }
173
174    StructType *MergedTy = StructType::get(M.getContext(), Tys);
175    Constant *MergedInit = ConstantStruct::get(MergedTy, Inits);
176    GlobalVariable *MergedGV = new GlobalVariable(M, MergedTy, isConst,
177                                                  GlobalValue::InternalLinkage,
178                                                  MergedInit, "_MergedGlobals",
179                                                  0, GlobalVariable::NotThreadLocal,
180                                                  AddrSpace);
181    for (size_t k = i; k < j; ++k) {
182      Constant *Idx[2] = {
183        ConstantInt::get(Int32Ty, 0),
184        ConstantInt::get(Int32Ty, k-i)
185      };
186      Constant *GEP = ConstantExpr::getInBoundsGetElementPtr(MergedGV, Idx);
187      Globals[k]->replaceAllUsesWith(GEP);
188      Globals[k]->eraseFromParent();
189      NumMerged++;
190    }
191    i = j;
192  }
193
194  return true;
195}
196
197void GlobalMerge::collectUsedGlobalVariables(Module &M) {
198  // Extract global variables from llvm.used array
199  const GlobalVariable *GV = M.getGlobalVariable("llvm.used");
200  if (!GV || !GV->hasInitializer()) return;
201
202  // Should be an array of 'i8*'.
203  const ConstantArray *InitList = cast<ConstantArray>(GV->getInitializer());
204
205  for (unsigned i = 0, e = InitList->getNumOperands(); i != e; ++i)
206    if (const GlobalVariable *G =
207        dyn_cast<GlobalVariable>(InitList->getOperand(i)->stripPointerCasts()))
208      MustKeepGlobalVariables.insert(G);
209}
210
211void GlobalMerge::setMustKeepGlobalVariables(Module &M) {
212  collectUsedGlobalVariables(M);
213
214  for (Module::iterator IFn = M.begin(), IEndFn = M.end(); IFn != IEndFn;
215       ++IFn) {
216    for (Function::iterator IBB = IFn->begin(), IEndBB = IFn->end();
217         IBB != IEndBB; ++IBB) {
218      // Follow the inwoke link to find the landing pad instruction
219      const InvokeInst *II = dyn_cast<InvokeInst>(IBB->getTerminator());
220      if (!II) continue;
221
222      const LandingPadInst *LPInst = II->getUnwindDest()->getLandingPadInst();
223      // Look for globals in the clauses of the landing pad instruction
224      for (unsigned Idx = 0, NumClauses = LPInst->getNumClauses();
225           Idx != NumClauses; ++Idx)
226        if (const GlobalVariable *GV =
227            dyn_cast<GlobalVariable>(LPInst->getClause(Idx)
228                                     ->stripPointerCasts()))
229          MustKeepGlobalVariables.insert(GV);
230    }
231  }
232}
233
234bool GlobalMerge::doInitialization(Module &M) {
235  DenseMap<unsigned, SmallVector<GlobalVariable*, 16> > Globals, ConstGlobals,
236                                                        BSSGlobals;
237  const DataLayout *TD = TLI->getDataLayout();
238  unsigned MaxOffset = TLI->getMaximalGlobalOffset();
239  bool Changed = false;
240  setMustKeepGlobalVariables(M);
241
242  // Grab all non-const globals.
243  for (Module::global_iterator I = M.global_begin(),
244         E = M.global_end(); I != E; ++I) {
245    // Merge is safe for "normal" internal globals only
246    if (!I->hasLocalLinkage() || I->isThreadLocal() || I->hasSection())
247      continue;
248
249    PointerType *PT = dyn_cast<PointerType>(I->getType());
250    assert(PT && "Global variable is not a pointer!");
251
252    unsigned AddressSpace = PT->getAddressSpace();
253
254    // Ignore fancy-aligned globals for now.
255    unsigned Alignment = TD->getPreferredAlignment(I);
256    Type *Ty = I->getType()->getElementType();
257    if (Alignment > TD->getABITypeAlignment(Ty))
258      continue;
259
260    // Ignore all 'special' globals.
261    if (I->getName().startswith("llvm.") ||
262        I->getName().startswith(".llvm."))
263      continue;
264
265    // Ignore all "required" globals:
266    if (isMustKeepGlobalVariable(I))
267      continue;
268
269    if (TD->getTypeAllocSize(Ty) < MaxOffset) {
270      if (TargetLoweringObjectFile::getKindForGlobal(I, TLI->getTargetMachine())
271          .isBSSLocal())
272        BSSGlobals[AddressSpace].push_back(I);
273      else if (I->isConstant())
274        ConstGlobals[AddressSpace].push_back(I);
275      else
276        Globals[AddressSpace].push_back(I);
277    }
278  }
279
280  for (DenseMap<unsigned, SmallVector<GlobalVariable*, 16> >::iterator
281       I = Globals.begin(), E = Globals.end(); I != E; ++I)
282    if (I->second.size() > 1)
283      Changed |= doMerge(I->second, M, false, I->first);
284
285  for (DenseMap<unsigned, SmallVector<GlobalVariable*, 16> >::iterator
286       I = BSSGlobals.begin(), E = BSSGlobals.end(); I != E; ++I)
287    if (I->second.size() > 1)
288      Changed |= doMerge(I->second, M, false, I->first);
289
290  if (EnableGlobalMergeOnConst)
291    for (DenseMap<unsigned, SmallVector<GlobalVariable*, 16> >::iterator
292         I = ConstGlobals.begin(), E = ConstGlobals.end(); I != E; ++I)
293      if (I->second.size() > 1)
294        Changed |= doMerge(I->second, M, true, I->first);
295
296  return Changed;
297}
298
299bool GlobalMerge::runOnFunction(Function &F) {
300  return false;
301}
302
303bool GlobalMerge::doFinalization(Module &M) {
304  MustKeepGlobalVariables.clear();
305  return false;
306}
307
308Pass *llvm::createGlobalMergePass(const TargetLowering *tli) {
309  return new GlobalMerge(tli);
310}
311