1 //===-- NVPTXInferAddressSpace.cpp - ---------------------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
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.
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.
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.
26 // For example, the Clang translates the following CUDA code
27 // __shared__ float a[10];
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).
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.
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
52 // Then, address space inference replaces all refinable generic pointers with
53 // equivalent specific pointers.
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.
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*
61 // %y = phi [ %generic.input, %y2 ]
62 // %y2 = getelementptr %y, 1
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
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' ]
90 //===----------------------------------------------------------------------===//
92 #include "llvm/ADT/DenseSet.h"
93 #include "llvm/ADT/Optional.h"
94 #include "llvm/ADT/SetVector.h"
95 #include "llvm/Analysis/TargetTransformInfo.h"
96 #include "llvm/IR/Function.h"
97 #include "llvm/IR/InstIterator.h"
98 #include "llvm/IR/Instructions.h"
99 #include "llvm/IR/Operator.h"
100 #include "llvm/Support/Debug.h"
101 #include "llvm/Support/raw_ostream.h"
102 #include "llvm/Transforms/Scalar.h"
103 #include "llvm/Transforms/Utils/Local.h"
104 #include "llvm/Transforms/Utils/ValueMapper.h"
106 #define DEBUG_TYPE "infer-address-spaces"
108 using namespace llvm;
111 static const unsigned UninitializedAddressSpace = ~0u;
113 using ValueToAddrSpaceMapTy = DenseMap<const Value *, unsigned>;
115 /// \brief InferAddressSpaces
116 class InferAddressSpaces : public FunctionPass {
117 /// Target specific address space which uses of should be replaced if
119 unsigned FlatAddrSpace;
124 InferAddressSpaces() : FunctionPass(ID) {}
126 void getAnalysisUsage(AnalysisUsage &AU) const override {
127 AU.setPreservesCFG();
128 AU.addRequired<TargetTransformInfoWrapperPass>();
131 bool runOnFunction(Function &F) override;
134 // Returns the new address space of V if updated; otherwise, returns None.
136 updateAddressSpace(const Value &V,
137 const ValueToAddrSpaceMapTy &InferredAddrSpace) const;
139 // Tries to infer the specific address space of each address expression in
141 void inferAddressSpaces(ArrayRef<WeakTrackingVH> Postorder,
142 ValueToAddrSpaceMapTy *InferredAddrSpace) const;
144 bool isSafeToCastConstAddrSpace(Constant *C, unsigned NewAS) const;
146 // Changes the flat address expressions in function F to point to specific
147 // address spaces if InferredAddrSpace says so. Postorder is the postorder of
148 // all flat expressions in the use-def graph of function F.
150 rewriteWithNewAddressSpaces(ArrayRef<WeakTrackingVH> Postorder,
151 const ValueToAddrSpaceMapTy &InferredAddrSpace,
154 void appendsFlatAddressExpressionToPostorderStack(
155 Value *V, std::vector<std::pair<Value *, bool>> &PostorderStack,
156 DenseSet<Value *> &Visited) const;
158 bool rewriteIntrinsicOperands(IntrinsicInst *II,
159 Value *OldV, Value *NewV) const;
160 void collectRewritableIntrinsicOperands(
162 std::vector<std::pair<Value *, bool>> &PostorderStack,
163 DenseSet<Value *> &Visited) const;
165 std::vector<WeakTrackingVH> collectFlatAddressExpressions(Function &F) const;
167 Value *cloneValueWithNewAddressSpace(
168 Value *V, unsigned NewAddrSpace,
169 const ValueToValueMapTy &ValueWithNewAddrSpace,
170 SmallVectorImpl<const Use *> *UndefUsesToFix) const;
171 unsigned joinAddressSpaces(unsigned AS1, unsigned AS2) const;
173 } // end anonymous namespace
175 char InferAddressSpaces::ID = 0;
178 void initializeInferAddressSpacesPass(PassRegistry &);
181 INITIALIZE_PASS(InferAddressSpaces, DEBUG_TYPE, "Infer address spaces",
184 // Returns true if V is an address expression.
185 // TODO: Currently, we consider only phi, bitcast, addrspacecast, and
186 // getelementptr operators.
187 static bool isAddressExpression(const Value &V) {
188 if (!isa<Operator>(V))
191 switch (cast<Operator>(V).getOpcode()) {
192 case Instruction::PHI:
193 case Instruction::BitCast:
194 case Instruction::AddrSpaceCast:
195 case Instruction::GetElementPtr:
196 case Instruction::Select:
203 // Returns the pointer operands of V.
205 // Precondition: V is an address expression.
206 static SmallVector<Value *, 2> getPointerOperands(const Value &V) {
207 const Operator &Op = cast<Operator>(V);
208 switch (Op.getOpcode()) {
209 case Instruction::PHI: {
210 auto IncomingValues = cast<PHINode>(Op).incoming_values();
211 return SmallVector<Value *, 2>(IncomingValues.begin(),
212 IncomingValues.end());
214 case Instruction::BitCast:
215 case Instruction::AddrSpaceCast:
216 case Instruction::GetElementPtr:
217 return {Op.getOperand(0)};
218 case Instruction::Select:
219 return {Op.getOperand(1), Op.getOperand(2)};
221 llvm_unreachable("Unexpected instruction type.");
225 // TODO: Move logic to TTI?
226 bool InferAddressSpaces::rewriteIntrinsicOperands(IntrinsicInst *II,
229 Module *M = II->getParent()->getParent()->getParent();
231 switch (II->getIntrinsicID()) {
232 case Intrinsic::amdgcn_atomic_inc:
233 case Intrinsic::amdgcn_atomic_dec:{
234 const ConstantInt *IsVolatile = dyn_cast<ConstantInt>(II->getArgOperand(4));
235 if (!IsVolatile || !IsVolatile->isZero())
240 case Intrinsic::objectsize: {
241 Type *DestTy = II->getType();
242 Type *SrcTy = NewV->getType();
244 Intrinsic::getDeclaration(M, II->getIntrinsicID(), {DestTy, SrcTy});
245 II->setArgOperand(0, NewV);
246 II->setCalledFunction(NewDecl);
254 // TODO: Move logic to TTI?
255 void InferAddressSpaces::collectRewritableIntrinsicOperands(
256 IntrinsicInst *II, std::vector<std::pair<Value *, bool>> &PostorderStack,
257 DenseSet<Value *> &Visited) const {
258 switch (II->getIntrinsicID()) {
259 case Intrinsic::objectsize:
260 case Intrinsic::amdgcn_atomic_inc:
261 case Intrinsic::amdgcn_atomic_dec:
262 appendsFlatAddressExpressionToPostorderStack(II->getArgOperand(0),
263 PostorderStack, Visited);
270 // Returns all flat address expressions in function F. The elements are
271 // If V is an unvisited flat address expression, appends V to PostorderStack
272 // and marks it as visited.
273 void InferAddressSpaces::appendsFlatAddressExpressionToPostorderStack(
274 Value *V, std::vector<std::pair<Value *, bool>> &PostorderStack,
275 DenseSet<Value *> &Visited) const {
276 assert(V->getType()->isPointerTy());
278 // Generic addressing expressions may be hidden in nested constant
280 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
281 // TODO: Look in non-address parts, like icmp operands.
282 if (isAddressExpression(*CE) && Visited.insert(CE).second)
283 PostorderStack.push_back(std::make_pair(CE, false));
288 if (isAddressExpression(*V) &&
289 V->getType()->getPointerAddressSpace() == FlatAddrSpace) {
290 if (Visited.insert(V).second) {
291 PostorderStack.push_back(std::make_pair(V, false));
293 Operator *Op = cast<Operator>(V);
294 for (unsigned I = 0, E = Op->getNumOperands(); I != E; ++I) {
295 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op->getOperand(I))) {
296 if (isAddressExpression(*CE) && Visited.insert(CE).second)
297 PostorderStack.emplace_back(CE, false);
304 // Returns all flat address expressions in function F. The elements are ordered
305 // ordered in postorder.
306 std::vector<WeakTrackingVH>
307 InferAddressSpaces::collectFlatAddressExpressions(Function &F) const {
308 // This function implements a non-recursive postorder traversal of a partial
309 // use-def graph of function F.
310 std::vector<std::pair<Value *, bool>> PostorderStack;
311 // The set of visited expressions.
312 DenseSet<Value *> Visited;
314 auto PushPtrOperand = [&](Value *Ptr) {
315 appendsFlatAddressExpressionToPostorderStack(Ptr, PostorderStack,
319 // Look at operations that may be interesting accelerate by moving to a known
320 // address space. We aim at generating after loads and stores, but pure
321 // addressing calculations may also be faster.
322 for (Instruction &I : instructions(F)) {
323 if (auto *GEP = dyn_cast<GetElementPtrInst>(&I)) {
324 if (!GEP->getType()->isVectorTy())
325 PushPtrOperand(GEP->getPointerOperand());
326 } else if (auto *LI = dyn_cast<LoadInst>(&I))
327 PushPtrOperand(LI->getPointerOperand());
328 else if (auto *SI = dyn_cast<StoreInst>(&I))
329 PushPtrOperand(SI->getPointerOperand());
330 else if (auto *RMW = dyn_cast<AtomicRMWInst>(&I))
331 PushPtrOperand(RMW->getPointerOperand());
332 else if (auto *CmpX = dyn_cast<AtomicCmpXchgInst>(&I))
333 PushPtrOperand(CmpX->getPointerOperand());
334 else if (auto *MI = dyn_cast<MemIntrinsic>(&I)) {
335 // For memset/memcpy/memmove, any pointer operand can be replaced.
336 PushPtrOperand(MI->getRawDest());
338 // Handle 2nd operand for memcpy/memmove.
339 if (auto *MTI = dyn_cast<MemTransferInst>(MI))
340 PushPtrOperand(MTI->getRawSource());
341 } else if (auto *II = dyn_cast<IntrinsicInst>(&I))
342 collectRewritableIntrinsicOperands(II, PostorderStack, Visited);
343 else if (ICmpInst *Cmp = dyn_cast<ICmpInst>(&I)) {
344 // FIXME: Handle vectors of pointers
345 if (Cmp->getOperand(0)->getType()->isPointerTy()) {
346 PushPtrOperand(Cmp->getOperand(0));
347 PushPtrOperand(Cmp->getOperand(1));
349 } else if (auto *ASC = dyn_cast<AddrSpaceCastInst>(&I)) {
350 if (!ASC->getType()->isVectorTy())
351 PushPtrOperand(ASC->getPointerOperand());
355 std::vector<WeakTrackingVH> Postorder; // The resultant postorder.
356 while (!PostorderStack.empty()) {
357 Value *TopVal = PostorderStack.back().first;
358 // If the operands of the expression on the top are already explored,
359 // adds that expression to the resultant postorder.
360 if (PostorderStack.back().second) {
361 if (TopVal->getType()->getPointerAddressSpace() == FlatAddrSpace)
362 Postorder.push_back(TopVal);
363 PostorderStack.pop_back();
366 // Otherwise, adds its operands to the stack and explores them.
367 PostorderStack.back().second = true;
368 for (Value *PtrOperand : getPointerOperands(*TopVal)) {
369 appendsFlatAddressExpressionToPostorderStack(PtrOperand, PostorderStack,
376 // A helper function for cloneInstructionWithNewAddressSpace. Returns the clone
377 // of OperandUse.get() in the new address space. If the clone is not ready yet,
378 // returns an undef in the new address space as a placeholder.
379 static Value *operandWithNewAddressSpaceOrCreateUndef(
380 const Use &OperandUse, unsigned NewAddrSpace,
381 const ValueToValueMapTy &ValueWithNewAddrSpace,
382 SmallVectorImpl<const Use *> *UndefUsesToFix) {
383 Value *Operand = OperandUse.get();
386 Operand->getType()->getPointerElementType()->getPointerTo(NewAddrSpace);
388 if (Constant *C = dyn_cast<Constant>(Operand))
389 return ConstantExpr::getAddrSpaceCast(C, NewPtrTy);
391 if (Value *NewOperand = ValueWithNewAddrSpace.lookup(Operand))
394 UndefUsesToFix->push_back(&OperandUse);
395 return UndefValue::get(NewPtrTy);
398 // Returns a clone of `I` with its operands converted to those specified in
399 // ValueWithNewAddrSpace. Due to potential cycles in the data flow graph, an
400 // operand whose address space needs to be modified might not exist in
401 // ValueWithNewAddrSpace. In that case, uses undef as a placeholder operand and
402 // adds that operand use to UndefUsesToFix so that caller can fix them later.
404 // Note that we do not necessarily clone `I`, e.g., if it is an addrspacecast
405 // from a pointer whose type already matches. Therefore, this function returns a
406 // Value* instead of an Instruction*.
407 static Value *cloneInstructionWithNewAddressSpace(
408 Instruction *I, unsigned NewAddrSpace,
409 const ValueToValueMapTy &ValueWithNewAddrSpace,
410 SmallVectorImpl<const Use *> *UndefUsesToFix) {
412 I->getType()->getPointerElementType()->getPointerTo(NewAddrSpace);
414 if (I->getOpcode() == Instruction::AddrSpaceCast) {
415 Value *Src = I->getOperand(0);
416 // Because `I` is flat, the source address space must be specific.
417 // Therefore, the inferred address space must be the source space, according
419 assert(Src->getType()->getPointerAddressSpace() == NewAddrSpace);
420 if (Src->getType() != NewPtrType)
421 return new BitCastInst(Src, NewPtrType);
425 // Computes the converted pointer operands.
426 SmallVector<Value *, 4> NewPointerOperands;
427 for (const Use &OperandUse : I->operands()) {
428 if (!OperandUse.get()->getType()->isPointerTy())
429 NewPointerOperands.push_back(nullptr);
431 NewPointerOperands.push_back(operandWithNewAddressSpaceOrCreateUndef(
432 OperandUse, NewAddrSpace, ValueWithNewAddrSpace, UndefUsesToFix));
435 switch (I->getOpcode()) {
436 case Instruction::BitCast:
437 return new BitCastInst(NewPointerOperands[0], NewPtrType);
438 case Instruction::PHI: {
439 assert(I->getType()->isPointerTy());
440 PHINode *PHI = cast<PHINode>(I);
441 PHINode *NewPHI = PHINode::Create(NewPtrType, PHI->getNumIncomingValues());
442 for (unsigned Index = 0; Index < PHI->getNumIncomingValues(); ++Index) {
443 unsigned OperandNo = PHINode::getOperandNumForIncomingValue(Index);
444 NewPHI->addIncoming(NewPointerOperands[OperandNo],
445 PHI->getIncomingBlock(Index));
449 case Instruction::GetElementPtr: {
450 GetElementPtrInst *GEP = cast<GetElementPtrInst>(I);
451 GetElementPtrInst *NewGEP = GetElementPtrInst::Create(
452 GEP->getSourceElementType(), NewPointerOperands[0],
453 SmallVector<Value *, 4>(GEP->idx_begin(), GEP->idx_end()));
454 NewGEP->setIsInBounds(GEP->isInBounds());
457 case Instruction::Select: {
458 assert(I->getType()->isPointerTy());
459 return SelectInst::Create(I->getOperand(0), NewPointerOperands[1],
460 NewPointerOperands[2], "", nullptr, I);
463 llvm_unreachable("Unexpected opcode");
467 // Similar to cloneInstructionWithNewAddressSpace, returns a clone of the
468 // constant expression `CE` with its operands replaced as specified in
469 // ValueWithNewAddrSpace.
470 static Value *cloneConstantExprWithNewAddressSpace(
471 ConstantExpr *CE, unsigned NewAddrSpace,
472 const ValueToValueMapTy &ValueWithNewAddrSpace) {
474 CE->getType()->getPointerElementType()->getPointerTo(NewAddrSpace);
476 if (CE->getOpcode() == Instruction::AddrSpaceCast) {
477 // Because CE is flat, the source address space must be specific.
478 // Therefore, the inferred address space must be the source space according
480 assert(CE->getOperand(0)->getType()->getPointerAddressSpace() ==
482 return ConstantExpr::getBitCast(CE->getOperand(0), TargetType);
485 if (CE->getOpcode() == Instruction::BitCast) {
486 if (Value *NewOperand = ValueWithNewAddrSpace.lookup(CE->getOperand(0)))
487 return ConstantExpr::getBitCast(cast<Constant>(NewOperand), TargetType);
488 return ConstantExpr::getAddrSpaceCast(CE, TargetType);
491 if (CE->getOpcode() == Instruction::Select) {
492 Constant *Src0 = CE->getOperand(1);
493 Constant *Src1 = CE->getOperand(2);
494 if (Src0->getType()->getPointerAddressSpace() ==
495 Src1->getType()->getPointerAddressSpace()) {
497 return ConstantExpr::getSelect(
498 CE->getOperand(0), ConstantExpr::getAddrSpaceCast(Src0, TargetType),
499 ConstantExpr::getAddrSpaceCast(Src1, TargetType));
503 // Computes the operands of the new constant expression.
505 SmallVector<Constant *, 4> NewOperands;
506 for (unsigned Index = 0; Index < CE->getNumOperands(); ++Index) {
507 Constant *Operand = CE->getOperand(Index);
508 // If the address space of `Operand` needs to be modified, the new operand
509 // with the new address space should already be in ValueWithNewAddrSpace
510 // because (1) the constant expressions we consider (i.e. addrspacecast,
511 // bitcast, and getelementptr) do not incur cycles in the data flow graph
512 // and (2) this function is called on constant expressions in postorder.
513 if (Value *NewOperand = ValueWithNewAddrSpace.lookup(Operand)) {
515 NewOperands.push_back(cast<Constant>(NewOperand));
517 // Otherwise, reuses the old operand.
518 NewOperands.push_back(Operand);
522 // If !IsNew, we will replace the Value with itself. However, replaced values
523 // are assumed to wrapped in a addrspace cast later so drop it now.
527 if (CE->getOpcode() == Instruction::GetElementPtr) {
528 // Needs to specify the source type while constructing a getelementptr
529 // constant expression.
530 return CE->getWithOperands(
531 NewOperands, TargetType, /*OnlyIfReduced=*/false,
532 NewOperands[0]->getType()->getPointerElementType());
535 return CE->getWithOperands(NewOperands, TargetType);
538 // Returns a clone of the value `V`, with its operands replaced as specified in
539 // ValueWithNewAddrSpace. This function is called on every flat address
540 // expression whose address space needs to be modified, in postorder.
542 // See cloneInstructionWithNewAddressSpace for the meaning of UndefUsesToFix.
543 Value *InferAddressSpaces::cloneValueWithNewAddressSpace(
544 Value *V, unsigned NewAddrSpace,
545 const ValueToValueMapTy &ValueWithNewAddrSpace,
546 SmallVectorImpl<const Use *> *UndefUsesToFix) const {
547 // All values in Postorder are flat address expressions.
548 assert(isAddressExpression(*V) &&
549 V->getType()->getPointerAddressSpace() == FlatAddrSpace);
551 if (Instruction *I = dyn_cast<Instruction>(V)) {
552 Value *NewV = cloneInstructionWithNewAddressSpace(
553 I, NewAddrSpace, ValueWithNewAddrSpace, UndefUsesToFix);
554 if (Instruction *NewI = dyn_cast<Instruction>(NewV)) {
555 if (NewI->getParent() == nullptr) {
556 NewI->insertBefore(I);
563 return cloneConstantExprWithNewAddressSpace(
564 cast<ConstantExpr>(V), NewAddrSpace, ValueWithNewAddrSpace);
567 // Defines the join operation on the address space lattice (see the file header
569 unsigned InferAddressSpaces::joinAddressSpaces(unsigned AS1,
570 unsigned AS2) const {
571 if (AS1 == FlatAddrSpace || AS2 == FlatAddrSpace)
572 return FlatAddrSpace;
574 if (AS1 == UninitializedAddressSpace)
576 if (AS2 == UninitializedAddressSpace)
579 // The join of two different specific address spaces is flat.
580 return (AS1 == AS2) ? AS1 : FlatAddrSpace;
583 bool InferAddressSpaces::runOnFunction(Function &F) {
587 const TargetTransformInfo &TTI =
588 getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
589 FlatAddrSpace = TTI.getFlatAddressSpace();
590 if (FlatAddrSpace == UninitializedAddressSpace)
593 // Collects all flat address expressions in postorder.
594 std::vector<WeakTrackingVH> Postorder = collectFlatAddressExpressions(F);
596 // Runs a data-flow analysis to refine the address spaces of every expression
598 ValueToAddrSpaceMapTy InferredAddrSpace;
599 inferAddressSpaces(Postorder, &InferredAddrSpace);
601 // Changes the address spaces of the flat address expressions who are inferred
602 // to point to a specific address space.
603 return rewriteWithNewAddressSpaces(Postorder, InferredAddrSpace, &F);
606 // Constants need to be tracked through RAUW to handle cases with nested
607 // constant expressions, so wrap values in WeakTrackingVH.
608 void InferAddressSpaces::inferAddressSpaces(
609 ArrayRef<WeakTrackingVH> Postorder,
610 ValueToAddrSpaceMapTy *InferredAddrSpace) const {
611 SetVector<Value *> Worklist(Postorder.begin(), Postorder.end());
612 // Initially, all expressions are in the uninitialized address space.
613 for (Value *V : Postorder)
614 (*InferredAddrSpace)[V] = UninitializedAddressSpace;
616 while (!Worklist.empty()) {
617 Value *V = Worklist.pop_back_val();
619 // Tries to update the address space of the stack top according to the
620 // address spaces of its operands.
621 DEBUG(dbgs() << "Updating the address space of\n " << *V << '\n');
622 Optional<unsigned> NewAS = updateAddressSpace(*V, *InferredAddrSpace);
623 if (!NewAS.hasValue())
625 // If any updates are made, grabs its users to the worklist because
626 // their address spaces can also be possibly updated.
627 DEBUG(dbgs() << " to " << NewAS.getValue() << '\n');
628 (*InferredAddrSpace)[V] = NewAS.getValue();
630 for (Value *User : V->users()) {
631 // Skip if User is already in the worklist.
632 if (Worklist.count(User))
635 auto Pos = InferredAddrSpace->find(User);
636 // Our algorithm only updates the address spaces of flat address
637 // expressions, which are those in InferredAddrSpace.
638 if (Pos == InferredAddrSpace->end())
641 // Function updateAddressSpace moves the address space down a lattice
642 // path. Therefore, nothing to do if User is already inferred as flat (the
643 // bottom element in the lattice).
644 if (Pos->second == FlatAddrSpace)
647 Worklist.insert(User);
652 Optional<unsigned> InferAddressSpaces::updateAddressSpace(
653 const Value &V, const ValueToAddrSpaceMapTy &InferredAddrSpace) const {
654 assert(InferredAddrSpace.count(&V));
656 // The new inferred address space equals the join of the address spaces
657 // of all its pointer operands.
658 unsigned NewAS = UninitializedAddressSpace;
660 const Operator &Op = cast<Operator>(V);
661 if (Op.getOpcode() == Instruction::Select) {
662 Value *Src0 = Op.getOperand(1);
663 Value *Src1 = Op.getOperand(2);
665 auto I = InferredAddrSpace.find(Src0);
666 unsigned Src0AS = (I != InferredAddrSpace.end()) ?
667 I->second : Src0->getType()->getPointerAddressSpace();
669 auto J = InferredAddrSpace.find(Src1);
670 unsigned Src1AS = (J != InferredAddrSpace.end()) ?
671 J->second : Src1->getType()->getPointerAddressSpace();
673 auto *C0 = dyn_cast<Constant>(Src0);
674 auto *C1 = dyn_cast<Constant>(Src1);
676 // If one of the inputs is a constant, we may be able to do a constant
677 // addrspacecast of it. Defer inferring the address space until the input
678 // address space is known.
679 if ((C1 && Src0AS == UninitializedAddressSpace) ||
680 (C0 && Src1AS == UninitializedAddressSpace))
683 if (C0 && isSafeToCastConstAddrSpace(C0, Src1AS))
685 else if (C1 && isSafeToCastConstAddrSpace(C1, Src0AS))
688 NewAS = joinAddressSpaces(Src0AS, Src1AS);
690 for (Value *PtrOperand : getPointerOperands(V)) {
691 auto I = InferredAddrSpace.find(PtrOperand);
692 unsigned OperandAS = I != InferredAddrSpace.end() ?
693 I->second : PtrOperand->getType()->getPointerAddressSpace();
695 // join(flat, *) = flat. So we can break if NewAS is already flat.
696 NewAS = joinAddressSpaces(NewAS, OperandAS);
697 if (NewAS == FlatAddrSpace)
702 unsigned OldAS = InferredAddrSpace.lookup(&V);
703 assert(OldAS != FlatAddrSpace);
709 /// \p returns true if \p U is the pointer operand of a memory instruction with
710 /// a single pointer operand that can have its address space changed by simply
711 /// mutating the use to a new value.
712 static bool isSimplePointerUseValidToReplace(Use &U) {
713 User *Inst = U.getUser();
714 unsigned OpNo = U.getOperandNo();
716 if (auto *LI = dyn_cast<LoadInst>(Inst))
717 return OpNo == LoadInst::getPointerOperandIndex() && !LI->isVolatile();
719 if (auto *SI = dyn_cast<StoreInst>(Inst))
720 return OpNo == StoreInst::getPointerOperandIndex() && !SI->isVolatile();
722 if (auto *RMW = dyn_cast<AtomicRMWInst>(Inst))
723 return OpNo == AtomicRMWInst::getPointerOperandIndex() && !RMW->isVolatile();
725 if (auto *CmpX = dyn_cast<AtomicCmpXchgInst>(Inst)) {
726 return OpNo == AtomicCmpXchgInst::getPointerOperandIndex() &&
733 /// Update memory intrinsic uses that require more complex processing than
734 /// simple memory instructions. Thse require re-mangling and may have multiple
735 /// pointer operands.
736 static bool handleMemIntrinsicPtrUse(MemIntrinsic *MI, Value *OldV,
739 MDNode *TBAA = MI->getMetadata(LLVMContext::MD_tbaa);
740 MDNode *ScopeMD = MI->getMetadata(LLVMContext::MD_alias_scope);
741 MDNode *NoAliasMD = MI->getMetadata(LLVMContext::MD_noalias);
743 if (auto *MSI = dyn_cast<MemSetInst>(MI)) {
744 B.CreateMemSet(NewV, MSI->getValue(),
745 MSI->getLength(), MSI->getAlignment(),
747 TBAA, ScopeMD, NoAliasMD);
748 } else if (auto *MTI = dyn_cast<MemTransferInst>(MI)) {
749 Value *Src = MTI->getRawSource();
750 Value *Dest = MTI->getRawDest();
752 // Be careful in case this is a self-to-self copy.
759 if (isa<MemCpyInst>(MTI)) {
760 MDNode *TBAAStruct = MTI->getMetadata(LLVMContext::MD_tbaa_struct);
761 B.CreateMemCpy(Dest, Src, MTI->getLength(),
764 TBAA, TBAAStruct, ScopeMD, NoAliasMD);
766 assert(isa<MemMoveInst>(MTI));
767 B.CreateMemMove(Dest, Src, MTI->getLength(),
770 TBAA, ScopeMD, NoAliasMD);
773 llvm_unreachable("unhandled MemIntrinsic");
775 MI->eraseFromParent();
779 // \p returns true if it is OK to change the address space of constant \p C with
780 // a ConstantExpr addrspacecast.
781 bool InferAddressSpaces::isSafeToCastConstAddrSpace(Constant *C, unsigned NewAS) const {
782 assert(NewAS != UninitializedAddressSpace);
784 unsigned SrcAS = C->getType()->getPointerAddressSpace();
785 if (SrcAS == NewAS || isa<UndefValue>(C))
788 // Prevent illegal casts between different non-flat address spaces.
789 if (SrcAS != FlatAddrSpace && NewAS != FlatAddrSpace)
792 if (isa<ConstantPointerNull>(C))
795 if (auto *Op = dyn_cast<Operator>(C)) {
796 // If we already have a constant addrspacecast, it should be safe to cast it
798 if (Op->getOpcode() == Instruction::AddrSpaceCast)
799 return isSafeToCastConstAddrSpace(cast<Constant>(Op->getOperand(0)), NewAS);
801 if (Op->getOpcode() == Instruction::IntToPtr &&
802 Op->getType()->getPointerAddressSpace() == FlatAddrSpace)
809 static Value::use_iterator skipToNextUser(Value::use_iterator I,
810 Value::use_iterator End) {
811 User *CurUser = I->getUser();
814 while (I != End && I->getUser() == CurUser)
820 bool InferAddressSpaces::rewriteWithNewAddressSpaces(
821 ArrayRef<WeakTrackingVH> Postorder,
822 const ValueToAddrSpaceMapTy &InferredAddrSpace, Function *F) const {
823 // For each address expression to be modified, creates a clone of it with its
824 // pointer operands converted to the new address space. Since the pointer
825 // operands are converted, the clone is naturally in the new address space by
827 ValueToValueMapTy ValueWithNewAddrSpace;
828 SmallVector<const Use *, 32> UndefUsesToFix;
829 for (Value* V : Postorder) {
830 unsigned NewAddrSpace = InferredAddrSpace.lookup(V);
831 if (V->getType()->getPointerAddressSpace() != NewAddrSpace) {
832 ValueWithNewAddrSpace[V] = cloneValueWithNewAddressSpace(
833 V, NewAddrSpace, ValueWithNewAddrSpace, &UndefUsesToFix);
837 if (ValueWithNewAddrSpace.empty())
840 // Fixes all the undef uses generated by cloneInstructionWithNewAddressSpace.
841 for (const Use *UndefUse : UndefUsesToFix) {
842 User *V = UndefUse->getUser();
843 User *NewV = cast<User>(ValueWithNewAddrSpace.lookup(V));
844 unsigned OperandNo = UndefUse->getOperandNo();
845 assert(isa<UndefValue>(NewV->getOperand(OperandNo)));
846 NewV->setOperand(OperandNo, ValueWithNewAddrSpace.lookup(UndefUse->get()));
849 SmallVector<Instruction *, 16> DeadInstructions;
851 // Replaces the uses of the old address expressions with the new ones.
852 for (const WeakTrackingVH &WVH : Postorder) {
853 assert(WVH && "value was unexpectedly deleted");
855 Value *NewV = ValueWithNewAddrSpace.lookup(V);
859 DEBUG(dbgs() << "Replacing the uses of " << *V
860 << "\n with\n " << *NewV << '\n');
862 if (Constant *C = dyn_cast<Constant>(V)) {
863 Constant *Replace = ConstantExpr::getAddrSpaceCast(cast<Constant>(NewV),
866 DEBUG(dbgs() << "Inserting replacement const cast: "
867 << Replace << ": " << *Replace << '\n');
868 C->replaceAllUsesWith(Replace);
873 Value::use_iterator I, E, Next;
874 for (I = V->use_begin(), E = V->use_end(); I != E; ) {
877 // Some users may see the same pointer operand in multiple operands. Skip
878 // to the next instruction.
879 I = skipToNextUser(I, E);
881 if (isSimplePointerUseValidToReplace(U)) {
882 // If V is used as the pointer operand of a compatible memory operation,
883 // sets the pointer operand to NewV. This replacement does not change
884 // the element type, so the resultant load/store is still valid.
889 User *CurUser = U.getUser();
890 // Handle more complex cases like intrinsic that need to be remangled.
891 if (auto *MI = dyn_cast<MemIntrinsic>(CurUser)) {
892 if (!MI->isVolatile() && handleMemIntrinsicPtrUse(MI, V, NewV))
896 if (auto *II = dyn_cast<IntrinsicInst>(CurUser)) {
897 if (rewriteIntrinsicOperands(II, V, NewV))
901 if (isa<Instruction>(CurUser)) {
902 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(CurUser)) {
903 // If we can infer that both pointers are in the same addrspace,
905 // %cmp = icmp eq float* %p, %q
907 // %cmp = icmp eq float addrspace(3)* %new_p, %new_q
909 unsigned NewAS = NewV->getType()->getPointerAddressSpace();
910 int SrcIdx = U.getOperandNo();
911 int OtherIdx = (SrcIdx == 0) ? 1 : 0;
912 Value *OtherSrc = Cmp->getOperand(OtherIdx);
914 if (Value *OtherNewV = ValueWithNewAddrSpace.lookup(OtherSrc)) {
915 if (OtherNewV->getType()->getPointerAddressSpace() == NewAS) {
916 Cmp->setOperand(OtherIdx, OtherNewV);
917 Cmp->setOperand(SrcIdx, NewV);
922 // Even if the type mismatches, we can cast the constant.
923 if (auto *KOtherSrc = dyn_cast<Constant>(OtherSrc)) {
924 if (isSafeToCastConstAddrSpace(KOtherSrc, NewAS)) {
925 Cmp->setOperand(SrcIdx, NewV);
926 Cmp->setOperand(OtherIdx,
927 ConstantExpr::getAddrSpaceCast(KOtherSrc, NewV->getType()));
933 if (AddrSpaceCastInst *ASC = dyn_cast<AddrSpaceCastInst>(CurUser)) {
934 unsigned NewAS = NewV->getType()->getPointerAddressSpace();
935 if (ASC->getDestAddressSpace() == NewAS) {
936 ASC->replaceAllUsesWith(NewV);
937 DeadInstructions.push_back(ASC);
942 // Otherwise, replaces the use with flat(NewV).
943 if (Instruction *I = dyn_cast<Instruction>(V)) {
944 BasicBlock::iterator InsertPos = std::next(I->getIterator());
945 while (isa<PHINode>(InsertPos))
947 U.set(new AddrSpaceCastInst(NewV, V->getType(), "", &*InsertPos));
949 U.set(ConstantExpr::getAddrSpaceCast(cast<Constant>(NewV),
955 if (V->use_empty()) {
956 if (Instruction *I = dyn_cast<Instruction>(V))
957 DeadInstructions.push_back(I);
961 for (Instruction *I : DeadInstructions)
962 RecursivelyDeleteTriviallyDeadInstructions(I);
967 FunctionPass *llvm::createInferAddressSpacesPass() {
968 return new InferAddressSpaces();