xref: /llvm-project/llvm/lib/Target/AMDGPU/AMDGPULowerModuleLDSPass.cpp (revision cdb9738963a1584d6530bcf7b102423f54a6bdb1)
1 //===-- AMDGPULowerModuleLDSPass.cpp ------------------------------*- C++ -*-=//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This pass eliminates LDS uses from non-kernel functions.
10 //
11 // The strategy is to create a new struct with a field for each LDS variable
12 // and allocate that struct at the same address for every kernel. Uses of the
13 // original LDS variables are then replaced with compile time offsets from that
14 // known address. AMDGPUMachineFunction allocates the LDS global.
15 //
16 // Local variables with constant annotation or non-undef initializer are passed
17 // through unchanged for simplification or error diagnostics in later passes.
18 //
19 // To reduce the memory overhead variables that are only used by kernels are
20 // excluded from this transform. The analysis to determine whether a variable
21 // is only used by a kernel is cheap and conservative so this may allocate
22 // a variable in every kernel when it was not strictly necessary to do so.
23 //
24 // A possible future refinement is to specialise the structure per-kernel, so
25 // that fields can be elided based on more expensive analysis.
26 //
27 //===----------------------------------------------------------------------===//
28 
29 #include "AMDGPU.h"
30 #include "Utils/AMDGPUBaseInfo.h"
31 #include "Utils/AMDGPUMemoryUtils.h"
32 #include "llvm/ADT/BitVector.h"
33 #include "llvm/ADT/DenseMap.h"
34 #include "llvm/ADT/STLExtras.h"
35 #include "llvm/ADT/SetVector.h"
36 #include "llvm/Analysis/CallGraph.h"
37 #include "llvm/IR/Constants.h"
38 #include "llvm/IR/DerivedTypes.h"
39 #include "llvm/IR/IRBuilder.h"
40 #include "llvm/IR/InlineAsm.h"
41 #include "llvm/IR/Instructions.h"
42 #include "llvm/IR/MDBuilder.h"
43 #include "llvm/InitializePasses.h"
44 #include "llvm/Pass.h"
45 #include "llvm/Support/CommandLine.h"
46 #include "llvm/Support/Debug.h"
47 #include "llvm/Support/OptimizedStructLayout.h"
48 #include "llvm/Transforms/Utils/ModuleUtils.h"
49 #include <tuple>
50 #include <vector>
51 
52 #define DEBUG_TYPE "amdgpu-lower-module-lds"
53 
54 using namespace llvm;
55 
56 static cl::opt<bool> SuperAlignLDSGlobals(
57     "amdgpu-super-align-lds-globals",
58     cl::desc("Increase alignment of LDS if it is not on align boundary"),
59     cl::init(true), cl::Hidden);
60 
61 namespace {
62 class AMDGPULowerModuleLDS : public ModulePass {
63 
64   static void removeFromUsedList(Module &M, StringRef Name,
65                                  SmallPtrSetImpl<Constant *> &ToRemove) {
66     GlobalVariable *GV = M.getNamedGlobal(Name);
67     if (!GV || ToRemove.empty()) {
68       return;
69     }
70 
71     SmallVector<Constant *, 16> Init;
72     auto *CA = cast<ConstantArray>(GV->getInitializer());
73     for (auto &Op : CA->operands()) {
74       // ModuleUtils::appendToUsed only inserts Constants
75       Constant *C = cast<Constant>(Op);
76       if (!ToRemove.contains(C->stripPointerCasts())) {
77         Init.push_back(C);
78       }
79     }
80 
81     if (Init.size() == CA->getNumOperands()) {
82       return; // none to remove
83     }
84 
85     GV->eraseFromParent();
86 
87     for (Constant *C : ToRemove) {
88       C->removeDeadConstantUsers();
89     }
90 
91     if (!Init.empty()) {
92       ArrayType *ATy =
93           ArrayType::get(Type::getInt8PtrTy(M.getContext()), Init.size());
94       GV =
95           new llvm::GlobalVariable(M, ATy, false, GlobalValue::AppendingLinkage,
96                                    ConstantArray::get(ATy, Init), Name);
97       GV->setSection("llvm.metadata");
98     }
99   }
100 
101   static void
102   removeFromUsedLists(Module &M,
103                       const std::vector<GlobalVariable *> &LocalVars) {
104     // The verifier rejects used lists containing an inttoptr of a constant
105     // so remove the variables from these lists before replaceAllUsesWith
106 
107     SmallPtrSet<Constant *, 32> LocalVarsSet;
108     for (GlobalVariable *LocalVar : LocalVars)
109       if (Constant *C = dyn_cast<Constant>(LocalVar->stripPointerCasts()))
110         LocalVarsSet.insert(C);
111     removeFromUsedList(M, "llvm.used", LocalVarsSet);
112     removeFromUsedList(M, "llvm.compiler.used", LocalVarsSet);
113   }
114 
115   static void markUsedByKernel(IRBuilder<> &Builder, Function *Func,
116                                GlobalVariable *SGV) {
117     // The llvm.amdgcn.module.lds instance is implicitly used by all kernels
118     // that might call a function which accesses a field within it. This is
119     // presently approximated to 'all kernels' if there are any such functions
120     // in the module. This implicit use is redefined as an explicit use here so
121     // that later passes, specifically PromoteAlloca, account for the required
122     // memory without any knowledge of this transform.
123 
124     // An operand bundle on llvm.donothing works because the call instruction
125     // survives until after the last pass that needs to account for LDS. It is
126     // better than inline asm as the latter survives until the end of codegen. A
127     // totally robust solution would be a function with the same semantics as
128     // llvm.donothing that takes a pointer to the instance and is lowered to a
129     // no-op after LDS is allocated, but that is not presently necessary.
130 
131     LLVMContext &Ctx = Func->getContext();
132 
133     Builder.SetInsertPoint(Func->getEntryBlock().getFirstNonPHI());
134 
135     FunctionType *FTy = FunctionType::get(Type::getVoidTy(Ctx), {});
136 
137     Function *Decl =
138         Intrinsic::getDeclaration(Func->getParent(), Intrinsic::donothing, {});
139 
140     Value *UseInstance[1] = {Builder.CreateInBoundsGEP(
141         SGV->getValueType(), SGV, ConstantInt::get(Type::getInt32Ty(Ctx), 0))};
142 
143     Builder.CreateCall(FTy, Decl, {},
144                        {OperandBundleDefT<Value *>("ExplicitUse", UseInstance)},
145                        "");
146   }
147 
148   static bool eliminateConstantExprUsesOfLDSFromAllInstructions(Module &M) {
149     // Constants are uniqued within LLVM. A ConstantExpr referring to a LDS
150     // global may have uses from multiple different functions as a result.
151     // This pass specialises LDS variables with respect to the kernel that
152     // allocates them.
153 
154     // This is semantically equivalent to:
155     // for (auto &F : M.functions())
156     //   for (auto &BB : F)
157     //     for (auto &I : BB)
158     //       for (Use &Op : I.operands())
159     //         if (constantExprUsesLDS(Op))
160     //           replaceConstantExprInFunction(I, Op);
161 
162     bool Changed = false;
163 
164     // Find all ConstantExpr that are direct users of an LDS global
165     SmallVector<ConstantExpr *> Stack;
166     for (auto &GV : M.globals())
167       if (AMDGPU::isLDSVariableToLower(GV))
168         for (User *U : GV.users())
169           if (ConstantExpr *C = dyn_cast<ConstantExpr>(U))
170             Stack.push_back(C);
171 
172     // Expand to include constexpr users of direct users
173     SetVector<ConstantExpr *> ConstExprUsersOfLDS;
174     while (!Stack.empty()) {
175       ConstantExpr *V = Stack.pop_back_val();
176       if (ConstExprUsersOfLDS.contains(V))
177         continue;
178 
179       ConstExprUsersOfLDS.insert(V);
180 
181       for (auto *Nested : V->users())
182         if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Nested))
183           Stack.push_back(CE);
184     }
185 
186     // Find all instructions that use any of the ConstExpr users of LDS
187     SetVector<Instruction *> InstructionWorklist;
188     for (ConstantExpr *CE : ConstExprUsersOfLDS)
189       for (User *U : CE->users())
190         if (auto *I = dyn_cast<Instruction>(U))
191           InstructionWorklist.insert(I);
192 
193     // Replace those ConstExpr operands with instructions
194     while (!InstructionWorklist.empty()) {
195       Instruction *I = InstructionWorklist.pop_back_val();
196       for (Use &U : I->operands()) {
197 
198         auto *BI = I;
199         if (auto *Phi = dyn_cast<PHINode>(I)) {
200           BasicBlock *BB = Phi->getIncomingBlock(U);
201           BasicBlock::iterator It = BB->getFirstInsertionPt();
202           assert(It != BB->end() && "Unexpected empty basic block");
203           BI = &(*(It));
204         }
205 
206         if (ConstantExpr *C = dyn_cast<ConstantExpr>(U.get())) {
207           if (ConstExprUsersOfLDS.contains(C)) {
208             Changed = true;
209             Instruction *NI = C->getAsInstruction(BI);
210             InstructionWorklist.insert(NI);
211             U.set(NI);
212             C->removeDeadConstantUsers();
213           }
214         }
215       }
216     }
217 
218     return Changed;
219   }
220 
221 public:
222   static char ID;
223 
224   AMDGPULowerModuleLDS() : ModulePass(ID) {
225     initializeAMDGPULowerModuleLDSPass(*PassRegistry::getPassRegistry());
226   }
227 
228   bool runOnModule(Module &M) override {
229     LLVMContext &Ctx = M.getContext();
230     CallGraph CG = CallGraph(M);
231     bool Changed = superAlignLDSGlobals(M);
232 
233     Changed |= eliminateConstantExprUsesOfLDSFromAllInstructions(M);
234 
235     // Move variables used by functions into amdgcn.module.lds
236     std::vector<GlobalVariable *> ModuleScopeVariables =
237         AMDGPU::findLDSVariablesToLower(M, nullptr);
238     if (!ModuleScopeVariables.empty()) {
239       std::string VarName = "llvm.amdgcn.module.lds";
240 
241       GlobalVariable *SGV;
242       DenseMap<GlobalVariable *, Constant *> LDSVarsToConstantGEP;
243       std::tie(SGV, LDSVarsToConstantGEP) =
244           createLDSVariableReplacement(M, VarName, ModuleScopeVariables);
245 
246       appendToCompilerUsed(
247           M, {static_cast<GlobalValue *>(
248                  ConstantExpr::getPointerBitCastOrAddrSpaceCast(
249                      cast<Constant>(SGV), Type::getInt8PtrTy(Ctx)))});
250 
251       removeFromUsedLists(M, ModuleScopeVariables);
252       replaceLDSVariablesWithStruct(M, ModuleScopeVariables, SGV,
253                                     LDSVarsToConstantGEP,
254                                     [](Use &) { return true; });
255 
256       // This ensures the variable is allocated when called functions access it.
257       // It also lets other passes, specifically PromoteAlloca, accurately
258       // calculate how much LDS will be used by the kernel after lowering.
259 
260       IRBuilder<> Builder(Ctx);
261       for (Function &Func : M.functions()) {
262         if (!Func.isDeclaration() && AMDGPU::isKernelCC(&Func)) {
263           const CallGraphNode *N = CG[&Func];
264           const bool CalleesRequireModuleLDS = N->size() > 0;
265 
266           if (CalleesRequireModuleLDS) {
267             // If a function this kernel might call requires module LDS,
268             // annotate the kernel to let later passes know it will allocate
269             // this structure, even if not apparent from the IR.
270             markUsedByKernel(Builder, &Func, SGV);
271           } else {
272             // However if we are certain this kernel cannot call a function that
273             // requires module LDS, annotate the kernel so the backend can elide
274             // the allocation without repeating callgraph walks.
275             Func.addFnAttr("amdgpu-elide-module-lds");
276           }
277         }
278       }
279 
280       Changed = true;
281     }
282 
283     // Move variables used by kernels into per-kernel instances
284     for (Function &F : M.functions()) {
285       if (F.isDeclaration())
286         continue;
287 
288       // Only lower compute kernels' LDS.
289       if (!AMDGPU::isKernel(F.getCallingConv()))
290         continue;
291 
292       std::vector<GlobalVariable *> KernelUsedVariables =
293           AMDGPU::findLDSVariablesToLower(M, &F);
294 
295       if (!KernelUsedVariables.empty()) {
296         std::string VarName =
297             (Twine("llvm.amdgcn.kernel.") + F.getName() + ".lds").str();
298         GlobalVariable *SGV;
299         DenseMap<GlobalVariable *, Constant *> LDSVarsToConstantGEP;
300         std::tie(SGV, LDSVarsToConstantGEP) =
301             createLDSVariableReplacement(M, VarName, KernelUsedVariables);
302 
303         removeFromUsedLists(M, KernelUsedVariables);
304         replaceLDSVariablesWithStruct(
305             M, KernelUsedVariables, SGV, LDSVarsToConstantGEP, [&F](Use &U) {
306               Instruction *I = dyn_cast<Instruction>(U.getUser());
307               return I && I->getFunction() == &F;
308             });
309         Changed = true;
310       }
311     }
312 
313     for (auto &GV : make_early_inc_range(M.globals()))
314       if (AMDGPU::isLDSVariableToLower(GV)) {
315         GV.removeDeadConstantUsers();
316         if (GV.use_empty())
317           GV.eraseFromParent();
318       }
319 
320     return Changed;
321   }
322 
323 private:
324   // Increase the alignment of LDS globals if necessary to maximise the chance
325   // that we can use aligned LDS instructions to access them.
326   static bool superAlignLDSGlobals(Module &M) {
327     const DataLayout &DL = M.getDataLayout();
328     bool Changed = false;
329     if (!SuperAlignLDSGlobals) {
330       return Changed;
331     }
332 
333     for (auto &GV : M.globals()) {
334       if (GV.getType()->getPointerAddressSpace() != AMDGPUAS::LOCAL_ADDRESS) {
335         // Only changing alignment of LDS variables
336         continue;
337       }
338       if (!GV.hasInitializer()) {
339         // cuda/hip extern __shared__ variable, leave alignment alone
340         continue;
341       }
342 
343       Align Alignment = AMDGPU::getAlign(DL, &GV);
344       TypeSize GVSize = DL.getTypeAllocSize(GV.getValueType());
345 
346       if (GVSize > 8) {
347         // We might want to use a b96 or b128 load/store
348         Alignment = std::max(Alignment, Align(16));
349       } else if (GVSize > 4) {
350         // We might want to use a b64 load/store
351         Alignment = std::max(Alignment, Align(8));
352       } else if (GVSize > 2) {
353         // We might want to use a b32 load/store
354         Alignment = std::max(Alignment, Align(4));
355       } else if (GVSize > 1) {
356         // We might want to use a b16 load/store
357         Alignment = std::max(Alignment, Align(2));
358       }
359 
360       if (Alignment != AMDGPU::getAlign(DL, &GV)) {
361         Changed = true;
362         GV.setAlignment(Alignment);
363       }
364     }
365     return Changed;
366   }
367 
368   std::tuple<GlobalVariable *, DenseMap<GlobalVariable *, Constant *>>
369   createLDSVariableReplacement(
370       Module &M, std::string VarName,
371       std::vector<GlobalVariable *> const &LDSVarsToTransform) {
372     // Create a struct instance containing LDSVarsToTransform and map from those
373     // variables to ConstantExprGEP
374     // Variables may be introduced to meet alignment requirements. No aliasing
375     // metadata is useful for these as they have no uses. Erased before return.
376 
377     LLVMContext &Ctx = M.getContext();
378     const DataLayout &DL = M.getDataLayout();
379     assert(!LDSVarsToTransform.empty());
380 
381     SmallVector<OptimizedStructLayoutField, 8> LayoutFields;
382     LayoutFields.reserve(LDSVarsToTransform.size());
383     for (GlobalVariable *GV : LDSVarsToTransform) {
384       OptimizedStructLayoutField F(GV, DL.getTypeAllocSize(GV->getValueType()),
385                                    AMDGPU::getAlign(DL, GV));
386       LayoutFields.emplace_back(F);
387     }
388 
389     performOptimizedStructLayout(LayoutFields);
390 
391     std::vector<GlobalVariable *> LocalVars;
392     BitVector IsPaddingField;
393     LocalVars.reserve(LDSVarsToTransform.size()); // will be at least this large
394     IsPaddingField.reserve(LDSVarsToTransform.size());
395     {
396       uint64_t CurrentOffset = 0;
397       for (size_t I = 0; I < LayoutFields.size(); I++) {
398         GlobalVariable *FGV = static_cast<GlobalVariable *>(
399             const_cast<void *>(LayoutFields[I].Id));
400         Align DataAlign = LayoutFields[I].Alignment;
401 
402         uint64_t DataAlignV = DataAlign.value();
403         if (uint64_t Rem = CurrentOffset % DataAlignV) {
404           uint64_t Padding = DataAlignV - Rem;
405 
406           // Append an array of padding bytes to meet alignment requested
407           // Note (o +      (a - (o % a)) ) % a == 0
408           //      (offset + Padding       ) % align == 0
409 
410           Type *ATy = ArrayType::get(Type::getInt8Ty(Ctx), Padding);
411           LocalVars.push_back(new GlobalVariable(
412               M, ATy, false, GlobalValue::InternalLinkage, UndefValue::get(ATy),
413               "", nullptr, GlobalValue::NotThreadLocal, AMDGPUAS::LOCAL_ADDRESS,
414               false));
415           IsPaddingField.push_back(true);
416           CurrentOffset += Padding;
417         }
418 
419         LocalVars.push_back(FGV);
420         IsPaddingField.push_back(false);
421         CurrentOffset += LayoutFields[I].Size;
422       }
423     }
424 
425     std::vector<Type *> LocalVarTypes;
426     LocalVarTypes.reserve(LocalVars.size());
427     std::transform(
428         LocalVars.cbegin(), LocalVars.cend(), std::back_inserter(LocalVarTypes),
429         [](const GlobalVariable *V) -> Type * { return V->getValueType(); });
430 
431     StructType *LDSTy = StructType::create(Ctx, LocalVarTypes, VarName + ".t");
432 
433     Align StructAlign = AMDGPU::getAlign(DL, LocalVars[0]);
434 
435     GlobalVariable *SGV = new GlobalVariable(
436         M, LDSTy, false, GlobalValue::InternalLinkage, UndefValue::get(LDSTy),
437         VarName, nullptr, GlobalValue::NotThreadLocal, AMDGPUAS::LOCAL_ADDRESS,
438         false);
439     SGV->setAlignment(StructAlign);
440 
441     DenseMap<GlobalVariable *, Constant *> Map;
442     Type *I32 = Type::getInt32Ty(Ctx);
443     for (size_t I = 0; I < LocalVars.size(); I++) {
444       GlobalVariable *GV = LocalVars[I];
445       Constant *GEPIdx[] = {ConstantInt::get(I32, 0), ConstantInt::get(I32, I)};
446       Constant *GEP = ConstantExpr::getGetElementPtr(LDSTy, SGV, GEPIdx, true);
447       if (IsPaddingField[I]) {
448         assert(GV->use_empty());
449         GV->eraseFromParent();
450       } else {
451         Map[GV] = GEP;
452       }
453     }
454     assert(Map.size() == LDSVarsToTransform.size());
455     return std::make_tuple(SGV, std::move(Map));
456   }
457 
458   template <typename PredicateTy>
459   void replaceLDSVariablesWithStruct(
460       Module &M, std::vector<GlobalVariable *> const &LDSVarsToTransform,
461       GlobalVariable *SGV,
462       DenseMap<GlobalVariable *, Constant *> &LDSVarsToConstantGEP,
463       PredicateTy Predicate) {
464     LLVMContext &Ctx = M.getContext();
465     const DataLayout &DL = M.getDataLayout();
466 
467     // Create alias.scope and their lists. Each field in the new structure
468     // does not alias with all other fields.
469     SmallVector<MDNode *> AliasScopes;
470     SmallVector<Metadata *> NoAliasList;
471     const size_t NumberVars = LDSVarsToTransform.size();
472     if (NumberVars > 1) {
473       MDBuilder MDB(Ctx);
474       AliasScopes.reserve(NumberVars);
475       MDNode *Domain = MDB.createAnonymousAliasScopeDomain();
476       for (size_t I = 0; I < NumberVars; I++) {
477         MDNode *Scope = MDB.createAnonymousAliasScope(Domain);
478         AliasScopes.push_back(Scope);
479       }
480       NoAliasList.append(&AliasScopes[1], AliasScopes.end());
481     }
482 
483     // Replace uses of ith variable with a constantexpr to the corresponding
484     // field of the instance that will be allocated by AMDGPUMachineFunction
485     for (size_t I = 0; I < NumberVars; I++) {
486       GlobalVariable *GV = LDSVarsToTransform[I];
487       Constant *GEP = LDSVarsToConstantGEP[GV];
488 
489       GV->replaceUsesWithIf(GEP, Predicate);
490       if (GV->use_empty()) {
491         GV->eraseFromParent();
492       }
493 
494       APInt APOff(DL.getIndexTypeSizeInBits(GEP->getType()), 0);
495       GEP->stripAndAccumulateInBoundsConstantOffsets(DL, APOff);
496       uint64_t Offset = APOff.getZExtValue();
497 
498       Align A = commonAlignment(SGV->getAlign().valueOrOne(), Offset);
499 
500       if (I)
501         NoAliasList[I - 1] = AliasScopes[I - 1];
502       MDNode *NoAlias =
503           NoAliasList.empty() ? nullptr : MDNode::get(Ctx, NoAliasList);
504       MDNode *AliasScope =
505           AliasScopes.empty() ? nullptr : MDNode::get(Ctx, {AliasScopes[I]});
506 
507       refineUsesAlignmentAndAA(GEP, A, DL, AliasScope, NoAlias);
508     }
509   }
510 
511   void refineUsesAlignmentAndAA(Value *Ptr, Align A, const DataLayout &DL,
512                                 MDNode *AliasScope, MDNode *NoAlias,
513                                 unsigned MaxDepth = 5) {
514     if (!MaxDepth || (A == 1 && !AliasScope))
515       return;
516 
517     for (User *U : Ptr->users()) {
518       if (auto *I = dyn_cast<Instruction>(U)) {
519         if (AliasScope && I->mayReadOrWriteMemory()) {
520           MDNode *AS = I->getMetadata(LLVMContext::MD_alias_scope);
521           AS = (AS ? MDNode::getMostGenericAliasScope(AS, AliasScope)
522                    : AliasScope);
523           I->setMetadata(LLVMContext::MD_alias_scope, AS);
524 
525           MDNode *NA = I->getMetadata(LLVMContext::MD_noalias);
526           NA = (NA ? MDNode::intersect(NA, NoAlias) : NoAlias);
527           I->setMetadata(LLVMContext::MD_noalias, NA);
528         }
529       }
530 
531       if (auto *LI = dyn_cast<LoadInst>(U)) {
532         LI->setAlignment(std::max(A, LI->getAlign()));
533         continue;
534       }
535       if (auto *SI = dyn_cast<StoreInst>(U)) {
536         if (SI->getPointerOperand() == Ptr)
537           SI->setAlignment(std::max(A, SI->getAlign()));
538         continue;
539       }
540       if (auto *AI = dyn_cast<AtomicRMWInst>(U)) {
541         // None of atomicrmw operations can work on pointers, but let's
542         // check it anyway in case it will or we will process ConstantExpr.
543         if (AI->getPointerOperand() == Ptr)
544           AI->setAlignment(std::max(A, AI->getAlign()));
545         continue;
546       }
547       if (auto *AI = dyn_cast<AtomicCmpXchgInst>(U)) {
548         if (AI->getPointerOperand() == Ptr)
549           AI->setAlignment(std::max(A, AI->getAlign()));
550         continue;
551       }
552       if (auto *GEP = dyn_cast<GetElementPtrInst>(U)) {
553         unsigned BitWidth = DL.getIndexTypeSizeInBits(GEP->getType());
554         APInt Off(BitWidth, 0);
555         if (GEP->getPointerOperand() == Ptr) {
556           Align GA;
557           if (GEP->accumulateConstantOffset(DL, Off))
558             GA = commonAlignment(A, Off.getLimitedValue());
559           refineUsesAlignmentAndAA(GEP, GA, DL, AliasScope, NoAlias,
560                                    MaxDepth - 1);
561         }
562         continue;
563       }
564       if (auto *I = dyn_cast<Instruction>(U)) {
565         if (I->getOpcode() == Instruction::BitCast ||
566             I->getOpcode() == Instruction::AddrSpaceCast)
567           refineUsesAlignmentAndAA(I, A, DL, AliasScope, NoAlias, MaxDepth - 1);
568       }
569     }
570   }
571 };
572 
573 } // namespace
574 char AMDGPULowerModuleLDS::ID = 0;
575 
576 char &llvm::AMDGPULowerModuleLDSID = AMDGPULowerModuleLDS::ID;
577 
578 INITIALIZE_PASS(AMDGPULowerModuleLDS, DEBUG_TYPE,
579                 "Lower uses of LDS variables from non-kernel functions", false,
580                 false)
581 
582 ModulePass *llvm::createAMDGPULowerModuleLDSPass() {
583   return new AMDGPULowerModuleLDS();
584 }
585 
586 PreservedAnalyses AMDGPULowerModuleLDSPass::run(Module &M,
587                                                 ModuleAnalysisManager &) {
588   return AMDGPULowerModuleLDS().runOnModule(M) ? PreservedAnalyses::none()
589                                                : PreservedAnalyses::all();
590 }
591