Home | History | Annotate | Download | only in NVPTX
      1 //===-- NVPTXInferAddressSpace.cpp - ---------------------*- C++ -*-===//
      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 // CUDA C/C++ includes memory space designation as variable type qualifers (such
     11 // as __global__ and __shared__). Knowing the space of a memory access allows
     12 // CUDA compilers to emit faster PTX loads and stores. For example, a load from
     13 // shared memory can be translated to `ld.shared` which is roughly 10% faster
     14 // than a generic `ld` on an NVIDIA Tesla K40c.
     15 //
     16 // Unfortunately, type qualifiers only apply to variable declarations, so CUDA
     17 // compilers must infer the memory space of an address expression from
     18 // type-qualified variables.
     19 //
     20 // LLVM IR uses non-zero (so-called) specific address spaces to represent memory
     21 // spaces (e.g. addrspace(3) means shared memory). The Clang frontend
     22 // places only type-qualified variables in specific address spaces, and then
     23 // conservatively `addrspacecast`s each type-qualified variable to addrspace(0)
     24 // (so-called the generic address space) for other instructions to use.
     25 //
     26 // For example, the Clang translates the following CUDA code
     27 //   __shared__ float a[10];
     28 //   float v = a[i];
     29 // to
     30 //   %0 = addrspacecast [10 x float] addrspace(3)* @a to [10 x float]*
     31 //   %1 = gep [10 x float], [10 x float]* %0, i64 0, i64 %i
     32 //   %v = load float, float* %1 ; emits ld.f32
     33 // @a is in addrspace(3) since it's type-qualified, but its use from %1 is
     34 // redirected to %0 (the generic version of @a).
     35 //
     36 // The optimization implemented in this file propagates specific address spaces
     37 // from type-qualified variable declarations to its users. For example, it
     38 // optimizes the above IR to
     39 //   %1 = gep [10 x float] addrspace(3)* @a, i64 0, i64 %i
     40 //   %v = load float addrspace(3)* %1 ; emits ld.shared.f32
     41 // propagating the addrspace(3) from @a to %1. As the result, the NVPTX
     42 // codegen is able to emit ld.shared.f32 for %v.
     43 //
     44 // Address space inference works in two steps. First, it uses a data-flow
     45 // analysis to infer as many generic pointers as possible to point to only one
     46 // specific address space. In the above example, it can prove that %1 only
     47 // points to addrspace(3). This algorithm was published in
     48 //   CUDA: Compiling and optimizing for a GPU platform
     49 //   Chakrabarti, Grover, Aarts, Kong, Kudlur, Lin, Marathe, Murphy, Wang
     50 //   ICCS 2012
     51 //
     52 // Then, address space inference replaces all refinable generic pointers with
     53 // equivalent specific pointers.
     54 //
     55 // The major challenge of implementing this optimization is handling PHINodes,
     56 // which may create loops in the data flow graph. This brings two complications.
     57 //
     58 // First, the data flow analysis in Step 1 needs to be circular. For example,
     59 //     %generic.input = addrspacecast float addrspace(3)* %input to float*
     60 //   loop:
     61 //     %y = phi [ %generic.input, %y2 ]
     62 //     %y2 = getelementptr %y, 1
     63 //     %v = load %y2
     64 //     br ..., label %loop, ...
     65 // proving %y specific requires proving both %generic.input and %y2 specific,
     66 // but proving %y2 specific circles back to %y. To address this complication,
     67 // the data flow analysis operates on a lattice:
     68 //   uninitialized > specific address spaces > generic.
     69 // All address expressions (our implementation only considers phi, bitcast,
     70 // addrspacecast, and getelementptr) start with the uninitialized address space.
     71 // The monotone transfer function moves the address space of a pointer down a
     72 // lattice path from uninitialized to specific and then to generic. A join
     73 // operation of two different specific address spaces pushes the expression down
     74 // to the generic address space. The analysis completes once it reaches a fixed
     75 // point.
     76 //
     77 // Second, IR rewriting in Step 2 also needs to be circular. For example,
     78 // converting %y to addrspace(3) requires the compiler to know the converted
     79 // %y2, but converting %y2 needs the converted %y. To address this complication,
     80 // we break these cycles using "undef" placeholders. When converting an
     81 // instruction `I` to a new address space, if its operand `Op` is not converted
     82 // yet, we let `I` temporarily use `undef` and fix all the uses of undef later.
     83 // For instance, our algorithm first converts %y to
     84 //   %y' = phi float addrspace(3)* [ %input, undef ]
     85 // Then, it converts %y2 to
     86 //   %y2' = getelementptr %y', 1
     87 // Finally, it fixes the undef in %y' so that
     88 //   %y' = phi float addrspace(3)* [ %input, %y2' ]
     89 //
     90 // TODO: This pass is experimental and not enabled by default. Users can turn it
     91 // on by setting the -nvptx-use-infer-addrspace flag of llc. We plan to replace
     92 // NVPTXNonFavorGenericAddrSpaces with this pass shortly.
     93 //===----------------------------------------------------------------------===//
     94 
     95 #define DEBUG_TYPE "nvptx-infer-addrspace"
     96 
     97 #include "NVPTX.h"
     98 #include "MCTargetDesc/NVPTXBaseInfo.h"
     99 #include "llvm/ADT/DenseSet.h"
    100 #include "llvm/ADT/Optional.h"
    101 #include "llvm/ADT/SetVector.h"
    102 #include "llvm/IR/Function.h"
    103 #include "llvm/IR/InstIterator.h"
    104 #include "llvm/IR/Instructions.h"
    105 #include "llvm/IR/Operator.h"
    106 #include "llvm/Support/Debug.h"
    107 #include "llvm/Support/raw_ostream.h"
    108 #include "llvm/Transforms/Utils/Local.h"
    109 #include "llvm/Transforms/Utils/ValueMapper.h"
    110 
    111 using namespace llvm;
    112 
    113 namespace {
    114 const unsigned ADDRESS_SPACE_UNINITIALIZED = (unsigned)-1;
    115 
    116 using ValueToAddrSpaceMapTy = DenseMap<const Value *, unsigned>;
    117 
    118 /// \brief NVPTXInferAddressSpaces
    119 class NVPTXInferAddressSpaces: public FunctionPass {
    120 public:
    121   static char ID;
    122 
    123   NVPTXInferAddressSpaces() : FunctionPass(ID) {}
    124 
    125   bool runOnFunction(Function &F) override;
    126 
    127 private:
    128   // Returns the new address space of V if updated; otherwise, returns None.
    129   Optional<unsigned>
    130   updateAddressSpace(const Value &V,
    131                      const ValueToAddrSpaceMapTy &InferredAddrSpace);
    132 
    133   // Tries to infer the specific address space of each address expression in
    134   // Postorder.
    135   void inferAddressSpaces(const std::vector<Value *> &Postorder,
    136                           ValueToAddrSpaceMapTy *InferredAddrSpace);
    137 
    138   // Changes the generic address expressions in function F to point to specific
    139   // address spaces if InferredAddrSpace says so. Postorder is the postorder of
    140   // all generic address expressions in the use-def graph of function F.
    141   bool
    142   rewriteWithNewAddressSpaces(const std::vector<Value *> &Postorder,
    143                               const ValueToAddrSpaceMapTy &InferredAddrSpace,
    144                               Function *F);
    145 };
    146 } // end anonymous namespace
    147 
    148 char NVPTXInferAddressSpaces::ID = 0;
    149 
    150 namespace llvm {
    151 void initializeNVPTXInferAddressSpacesPass(PassRegistry &);
    152 }
    153 INITIALIZE_PASS(NVPTXInferAddressSpaces, "nvptx-infer-addrspace",
    154                 "Infer address spaces",
    155                 false, false)
    156 
    157 // Returns true if V is an address expression.
    158 // TODO: Currently, we consider only phi, bitcast, addrspacecast, and
    159 // getelementptr operators.
    160 static bool isAddressExpression(const Value &V) {
    161   if (!isa<Operator>(V))
    162     return false;
    163 
    164   switch (cast<Operator>(V).getOpcode()) {
    165   case Instruction::PHI:
    166   case Instruction::BitCast:
    167   case Instruction::AddrSpaceCast:
    168   case Instruction::GetElementPtr:
    169     return true;
    170   default:
    171     return false;
    172   }
    173 }
    174 
    175 // Returns the pointer operands of V.
    176 //
    177 // Precondition: V is an address expression.
    178 static SmallVector<Value *, 2> getPointerOperands(const Value &V) {
    179   assert(isAddressExpression(V));
    180   const Operator& Op = cast<Operator>(V);
    181   switch (Op.getOpcode()) {
    182   case Instruction::PHI: {
    183     auto IncomingValues = cast<PHINode>(Op).incoming_values();
    184     return SmallVector<Value *, 2>(IncomingValues.begin(),
    185                                    IncomingValues.end());
    186   }
    187   case Instruction::BitCast:
    188   case Instruction::AddrSpaceCast:
    189   case Instruction::GetElementPtr:
    190     return {Op.getOperand(0)};
    191   default:
    192     llvm_unreachable("Unexpected instruction type.");
    193   }
    194 }
    195 
    196 // If V is an unvisited generic address expression, appends V to PostorderStack
    197 // and marks it as visited.
    198 static void appendsGenericAddressExpressionToPostorderStack(
    199     Value *V, std::vector<std::pair<Value *, bool>> *PostorderStack,
    200     DenseSet<Value *> *Visited) {
    201   assert(V->getType()->isPointerTy());
    202   if (isAddressExpression(*V) &&
    203       V->getType()->getPointerAddressSpace() ==
    204           AddressSpace::ADDRESS_SPACE_GENERIC) {
    205     if (Visited->insert(V).second)
    206       PostorderStack->push_back(std::make_pair(V, false));
    207   }
    208 }
    209 
    210 // Returns all generic address expressions in function F. The elements are
    211 // ordered in postorder.
    212 static std::vector<Value *> collectGenericAddressExpressions(Function &F) {
    213   // This function implements a non-recursive postorder traversal of a partial
    214   // use-def graph of function F.
    215   std::vector<std::pair<Value*, bool>> PostorderStack;
    216   // The set of visited expressions.
    217   DenseSet<Value*> Visited;
    218   // We only explore address expressions that are reachable from loads and
    219   // stores for now because we aim at generating faster loads and stores.
    220   for (Instruction &I : instructions(F)) {
    221     if (isa<LoadInst>(I)) {
    222       appendsGenericAddressExpressionToPostorderStack(
    223           I.getOperand(0), &PostorderStack, &Visited);
    224     } else if (isa<StoreInst>(I)) {
    225       appendsGenericAddressExpressionToPostorderStack(
    226           I.getOperand(1), &PostorderStack, &Visited);
    227     }
    228   }
    229 
    230   std::vector<Value *> Postorder; // The resultant postorder.
    231   while (!PostorderStack.empty()) {
    232     // If the operands of the expression on the top are already explored,
    233     // adds that expression to the resultant postorder.
    234     if (PostorderStack.back().second) {
    235       Postorder.push_back(PostorderStack.back().first);
    236       PostorderStack.pop_back();
    237       continue;
    238     }
    239     // Otherwise, adds its operands to the stack and explores them.
    240     PostorderStack.back().second = true;
    241     for (Value *PtrOperand : getPointerOperands(*PostorderStack.back().first)) {
    242       appendsGenericAddressExpressionToPostorderStack(
    243           PtrOperand, &PostorderStack, &Visited);
    244     }
    245   }
    246   return Postorder;
    247 }
    248 
    249 // A helper function for cloneInstructionWithNewAddressSpace. Returns the clone
    250 // of OperandUse.get() in the new address space. If the clone is not ready yet,
    251 // returns an undef in the new address space as a placeholder.
    252 static Value *operandWithNewAddressSpaceOrCreateUndef(
    253     const Use &OperandUse, unsigned NewAddrSpace,
    254     const ValueToValueMapTy &ValueWithNewAddrSpace,
    255     SmallVectorImpl<const Use *> *UndefUsesToFix) {
    256   Value *Operand = OperandUse.get();
    257   if (Value *NewOperand = ValueWithNewAddrSpace.lookup(Operand))
    258     return NewOperand;
    259 
    260   UndefUsesToFix->push_back(&OperandUse);
    261   return UndefValue::get(
    262       Operand->getType()->getPointerElementType()->getPointerTo(NewAddrSpace));
    263 }
    264 
    265 // Returns a clone of `I` with its operands converted to those specified in
    266 // ValueWithNewAddrSpace. Due to potential cycles in the data flow graph, an
    267 // operand whose address space needs to be modified might not exist in
    268 // ValueWithNewAddrSpace. In that case, uses undef as a placeholder operand and
    269 // adds that operand use to UndefUsesToFix so that caller can fix them later.
    270 //
    271 // Note that we do not necessarily clone `I`, e.g., if it is an addrspacecast
    272 // from a pointer whose type already matches. Therefore, this function returns a
    273 // Value* instead of an Instruction*.
    274 static Value *cloneInstructionWithNewAddressSpace(
    275     Instruction *I, unsigned NewAddrSpace,
    276     const ValueToValueMapTy &ValueWithNewAddrSpace,
    277     SmallVectorImpl<const Use *> *UndefUsesToFix) {
    278   Type *NewPtrType =
    279       I->getType()->getPointerElementType()->getPointerTo(NewAddrSpace);
    280 
    281   if (I->getOpcode() == Instruction::AddrSpaceCast) {
    282     Value *Src = I->getOperand(0);
    283     // Because `I` is generic, the source address space must be specific.
    284     // Therefore, the inferred address space must be the source space, according
    285     // to our algorithm.
    286     assert(Src->getType()->getPointerAddressSpace() == NewAddrSpace);
    287     if (Src->getType() != NewPtrType)
    288       return new BitCastInst(Src, NewPtrType);
    289     return Src;
    290   }
    291 
    292   // Computes the converted pointer operands.
    293   SmallVector<Value *, 4> NewPointerOperands;
    294   for (const Use &OperandUse : I->operands()) {
    295     if (!OperandUse.get()->getType()->isPointerTy())
    296       NewPointerOperands.push_back(nullptr);
    297     else
    298       NewPointerOperands.push_back(operandWithNewAddressSpaceOrCreateUndef(
    299           OperandUse, NewAddrSpace, ValueWithNewAddrSpace, UndefUsesToFix));
    300   }
    301 
    302   switch (I->getOpcode()) {
    303   case Instruction::BitCast:
    304     return new BitCastInst(NewPointerOperands[0], NewPtrType);
    305   case Instruction::PHI: {
    306     assert(I->getType()->isPointerTy());
    307     PHINode *PHI = cast<PHINode>(I);
    308     PHINode *NewPHI = PHINode::Create(NewPtrType, PHI->getNumIncomingValues());
    309     for (unsigned Index = 0; Index < PHI->getNumIncomingValues(); ++Index) {
    310       unsigned OperandNo = PHINode::getOperandNumForIncomingValue(Index);
    311       NewPHI->addIncoming(NewPointerOperands[OperandNo],
    312                           PHI->getIncomingBlock(Index));
    313     }
    314     return NewPHI;
    315   }
    316   case Instruction::GetElementPtr: {
    317     GetElementPtrInst *GEP = cast<GetElementPtrInst>(I);
    318     GetElementPtrInst *NewGEP = GetElementPtrInst::Create(
    319         GEP->getSourceElementType(), NewPointerOperands[0],
    320         SmallVector<Value *, 4>(GEP->idx_begin(), GEP->idx_end()));
    321     NewGEP->setIsInBounds(GEP->isInBounds());
    322     return NewGEP;
    323   }
    324   default:
    325     llvm_unreachable("Unexpected opcode");
    326   }
    327 }
    328 
    329 // Similar to cloneInstructionWithNewAddressSpace, returns a clone of the
    330 // constant expression `CE` with its operands replaced as specified in
    331 // ValueWithNewAddrSpace.
    332 static Value *cloneConstantExprWithNewAddressSpace(
    333     ConstantExpr *CE, unsigned NewAddrSpace,
    334     const ValueToValueMapTy &ValueWithNewAddrSpace) {
    335   Type *TargetType =
    336       CE->getType()->getPointerElementType()->getPointerTo(NewAddrSpace);
    337 
    338   if (CE->getOpcode() == Instruction::AddrSpaceCast) {
    339     // Because CE is generic, the source address space must be specific.
    340     // Therefore, the inferred address space must be the source space according
    341     // to our algorithm.
    342     assert(CE->getOperand(0)->getType()->getPointerAddressSpace() ==
    343            NewAddrSpace);
    344     return ConstantExpr::getBitCast(CE->getOperand(0), TargetType);
    345   }
    346 
    347   // Computes the operands of the new constant expression.
    348   SmallVector<Constant *, 4> NewOperands;
    349   for (unsigned Index = 0; Index < CE->getNumOperands(); ++Index) {
    350     Constant *Operand = CE->getOperand(Index);
    351     // If the address space of `Operand` needs to be modified, the new operand
    352     // with the new address space should already be in ValueWithNewAddrSpace
    353     // because (1) the constant expressions we consider (i.e. addrspacecast,
    354     // bitcast, and getelementptr) do not incur cycles in the data flow graph
    355     // and (2) this function is called on constant expressions in postorder.
    356     if (Value *NewOperand = ValueWithNewAddrSpace.lookup(Operand)) {
    357       NewOperands.push_back(cast<Constant>(NewOperand));
    358     } else {
    359       // Otherwise, reuses the old operand.
    360       NewOperands.push_back(Operand);
    361     }
    362   }
    363 
    364   if (CE->getOpcode() == Instruction::GetElementPtr) {
    365     // Needs to specify the source type while constructing a getelementptr
    366     // constant expression.
    367     return CE->getWithOperands(
    368         NewOperands, TargetType, /*OnlyIfReduced=*/false,
    369         NewOperands[0]->getType()->getPointerElementType());
    370   }
    371 
    372   return CE->getWithOperands(NewOperands, TargetType);
    373 }
    374 
    375 // Returns a clone of the value `V`, with its operands replaced as specified in
    376 // ValueWithNewAddrSpace. This function is called on every generic address
    377 // expression whose address space needs to be modified, in postorder.
    378 //
    379 // See cloneInstructionWithNewAddressSpace for the meaning of UndefUsesToFix.
    380 static Value *
    381 cloneValueWithNewAddressSpace(Value *V, unsigned NewAddrSpace,
    382                               const ValueToValueMapTy &ValueWithNewAddrSpace,
    383                               SmallVectorImpl<const Use *> *UndefUsesToFix) {
    384   // All values in Postorder are generic address expressions.
    385   assert(isAddressExpression(*V) &&
    386          V->getType()->getPointerAddressSpace() ==
    387              AddressSpace::ADDRESS_SPACE_GENERIC);
    388 
    389   if (Instruction *I = dyn_cast<Instruction>(V)) {
    390     Value *NewV = cloneInstructionWithNewAddressSpace(
    391         I, NewAddrSpace, ValueWithNewAddrSpace, UndefUsesToFix);
    392     if (Instruction *NewI = dyn_cast<Instruction>(NewV)) {
    393       if (NewI->getParent() == nullptr) {
    394         NewI->insertBefore(I);
    395         NewI->takeName(I);
    396       }
    397     }
    398     return NewV;
    399   }
    400 
    401   return cloneConstantExprWithNewAddressSpace(
    402       cast<ConstantExpr>(V), NewAddrSpace, ValueWithNewAddrSpace);
    403 }
    404 
    405 // Defines the join operation on the address space lattice (see the file header
    406 // comments).
    407 static unsigned joinAddressSpaces(unsigned AS1, unsigned AS2) {
    408   if (AS1 == AddressSpace::ADDRESS_SPACE_GENERIC ||
    409       AS2 == AddressSpace::ADDRESS_SPACE_GENERIC)
    410     return AddressSpace::ADDRESS_SPACE_GENERIC;
    411 
    412   if (AS1 == ADDRESS_SPACE_UNINITIALIZED)
    413     return AS2;
    414   if (AS2 == ADDRESS_SPACE_UNINITIALIZED)
    415     return AS1;
    416 
    417   // The join of two different specific address spaces is generic.
    418   return AS1 == AS2 ? AS1 : (unsigned)AddressSpace::ADDRESS_SPACE_GENERIC;
    419 }
    420 
    421 bool NVPTXInferAddressSpaces::runOnFunction(Function &F) {
    422   if (skipFunction(F))
    423     return false;
    424 
    425   // Collects all generic address expressions in postorder.
    426   std::vector<Value *> Postorder = collectGenericAddressExpressions(F);
    427 
    428   // Runs a data-flow analysis to refine the address spaces of every expression
    429   // in Postorder.
    430   ValueToAddrSpaceMapTy InferredAddrSpace;
    431   inferAddressSpaces(Postorder, &InferredAddrSpace);
    432 
    433   // Changes the address spaces of the generic address expressions who are
    434   // inferred to point to a specific address space.
    435   return rewriteWithNewAddressSpaces(Postorder, InferredAddrSpace, &F);
    436 }
    437 
    438 void NVPTXInferAddressSpaces::inferAddressSpaces(
    439     const std::vector<Value *> &Postorder,
    440     ValueToAddrSpaceMapTy *InferredAddrSpace) {
    441   SetVector<Value *> Worklist(Postorder.begin(), Postorder.end());
    442   // Initially, all expressions are in the uninitialized address space.
    443   for (Value *V : Postorder)
    444     (*InferredAddrSpace)[V] = ADDRESS_SPACE_UNINITIALIZED;
    445 
    446   while (!Worklist.empty()) {
    447     Value* V = Worklist.pop_back_val();
    448 
    449     // Tries to update the address space of the stack top according to the
    450     // address spaces of its operands.
    451     DEBUG(dbgs() << "Updating the address space of\n"
    452                  << "  " << *V << "\n");
    453     Optional<unsigned> NewAS = updateAddressSpace(*V, *InferredAddrSpace);
    454     if (!NewAS.hasValue())
    455       continue;
    456     // If any updates are made, grabs its users to the worklist because
    457     // their address spaces can also be possibly updated.
    458     DEBUG(dbgs() << "  to " << NewAS.getValue() << "\n");
    459     (*InferredAddrSpace)[V] = NewAS.getValue();
    460 
    461     for (Value *User : V->users()) {
    462       // Skip if User is already in the worklist.
    463       if (Worklist.count(User))
    464         continue;
    465 
    466       auto Pos = InferredAddrSpace->find(User);
    467       // Our algorithm only updates the address spaces of generic address
    468       // expressions, which are those in InferredAddrSpace.
    469       if (Pos == InferredAddrSpace->end())
    470         continue;
    471 
    472       // Function updateAddressSpace moves the address space down a lattice
    473       // path. Therefore, nothing to do if User is already inferred as
    474       // generic (the bottom element in the lattice).
    475       if (Pos->second == AddressSpace::ADDRESS_SPACE_GENERIC)
    476         continue;
    477 
    478       Worklist.insert(User);
    479     }
    480   }
    481 }
    482 
    483 Optional<unsigned> NVPTXInferAddressSpaces::updateAddressSpace(
    484     const Value &V, const ValueToAddrSpaceMapTy &InferredAddrSpace) {
    485   assert(InferredAddrSpace.count(&V));
    486 
    487   // The new inferred address space equals the join of the address spaces
    488   // of all its pointer operands.
    489   unsigned NewAS = ADDRESS_SPACE_UNINITIALIZED;
    490   for (Value *PtrOperand : getPointerOperands(V)) {
    491     unsigned OperandAS;
    492     if (InferredAddrSpace.count(PtrOperand))
    493       OperandAS = InferredAddrSpace.lookup(PtrOperand);
    494     else
    495       OperandAS = PtrOperand->getType()->getPointerAddressSpace();
    496     NewAS = joinAddressSpaces(NewAS, OperandAS);
    497     // join(generic, *) = generic. So we can break if NewAS is already generic.
    498     if (NewAS == AddressSpace::ADDRESS_SPACE_GENERIC)
    499       break;
    500   }
    501 
    502   unsigned OldAS = InferredAddrSpace.lookup(&V);
    503   assert(OldAS != AddressSpace::ADDRESS_SPACE_GENERIC);
    504   if (OldAS == NewAS)
    505     return None;
    506   return NewAS;
    507 }
    508 
    509 bool NVPTXInferAddressSpaces::rewriteWithNewAddressSpaces(
    510     const std::vector<Value *> &Postorder,
    511     const ValueToAddrSpaceMapTy &InferredAddrSpace, Function *F) {
    512   // For each address expression to be modified, creates a clone of it with its
    513   // pointer operands converted to the new address space. Since the pointer
    514   // operands are converted, the clone is naturally in the new address space by
    515   // construction.
    516   ValueToValueMapTy ValueWithNewAddrSpace;
    517   SmallVector<const Use *, 32> UndefUsesToFix;
    518   for (Value* V : Postorder) {
    519     unsigned NewAddrSpace = InferredAddrSpace.lookup(V);
    520     if (V->getType()->getPointerAddressSpace() != NewAddrSpace) {
    521       ValueWithNewAddrSpace[V] = cloneValueWithNewAddressSpace(
    522           V, NewAddrSpace, ValueWithNewAddrSpace, &UndefUsesToFix);
    523     }
    524   }
    525 
    526   if (ValueWithNewAddrSpace.empty())
    527     return false;
    528 
    529   // Fixes all the undef uses generated by cloneInstructionWithNewAddressSpace.
    530   for (const Use* UndefUse : UndefUsesToFix) {
    531     User *V = UndefUse->getUser();
    532     User *NewV = cast<User>(ValueWithNewAddrSpace.lookup(V));
    533     unsigned OperandNo = UndefUse->getOperandNo();
    534     assert(isa<UndefValue>(NewV->getOperand(OperandNo)));
    535     NewV->setOperand(OperandNo, ValueWithNewAddrSpace.lookup(UndefUse->get()));
    536   }
    537 
    538   // Replaces the uses of the old address expressions with the new ones.
    539   for (Value *V : Postorder) {
    540     Value *NewV = ValueWithNewAddrSpace.lookup(V);
    541     if (NewV == nullptr)
    542       continue;
    543 
    544     SmallVector<Use *, 4> Uses;
    545     for (Use &U : V->uses())
    546       Uses.push_back(&U);
    547     DEBUG(dbgs() << "Replacing the uses of " << *V << "\n  to\n  " << *NewV
    548                  << "\n");
    549     for (Use *U : Uses) {
    550       if (isa<LoadInst>(U->getUser()) ||
    551           (isa<StoreInst>(U->getUser()) && U->getOperandNo() == 1)) {
    552         // If V is used as the pointer operand of a load/store, sets the pointer
    553         // operand to NewV. This replacement does not change the element type,
    554         // so the resultant load/store is still valid.
    555         U->set(NewV);
    556       } else if (isa<Instruction>(U->getUser())) {
    557         // Otherwise, replaces the use with generic(NewV).
    558         // TODO: Some optimization opportunities are missed. For example, in
    559         //   %0 = icmp eq float* %p, %q
    560         // if both p and q are inferred to be shared, we can rewrite %0 as
    561         //   %0 = icmp eq float addrspace(3)* %new_p, %new_q
    562         // instead of currently
    563         //   %generic_p = addrspacecast float addrspace(3)* %new_p to float*
    564         //   %generic_q = addrspacecast float addrspace(3)* %new_q to float*
    565         //   %0 = icmp eq float* %generic_p, %generic_q
    566         if (Instruction *I = dyn_cast<Instruction>(V)) {
    567           BasicBlock::iterator InsertPos = std::next(I->getIterator());
    568           while (isa<PHINode>(InsertPos))
    569             ++InsertPos;
    570           U->set(new AddrSpaceCastInst(NewV, V->getType(), "", &*InsertPos));
    571         } else {
    572           U->set(ConstantExpr::getAddrSpaceCast(cast<Constant>(NewV),
    573                                                 V->getType()));
    574         }
    575       }
    576     }
    577     if (V->use_empty())
    578       RecursivelyDeleteTriviallyDeadInstructions(V);
    579   }
    580 
    581   return true;
    582 }
    583 
    584 FunctionPass *llvm::createNVPTXInferAddressSpacesPass() {
    585   return new NVPTXInferAddressSpaces();
    586 }
    587