xref: /llvm-project/llvm/lib/Target/AMDGPU/AMDGPULowerModuleLDSPass.cpp (revision 35f2584ef9897ec9e5455ad2660dc1be5171c38c)
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         // The association between kernel function and LDS struct is done by
297         // symbol name, which only works if the function in question has a name
298         // This is not expected to be a problem in practice as kernels are
299         // called by name making anonymous ones (which are named by the backend)
300         // difficult to use. This does mean that llvm test cases need
301         // to name the kernels.
302         if (!F.hasName()) {
303           report_fatal_error("Anonymous kernels cannot use LDS variables");
304         }
305 
306         std::string VarName =
307             (Twine("llvm.amdgcn.kernel.") + F.getName() + ".lds").str();
308         GlobalVariable *SGV;
309         DenseMap<GlobalVariable *, Constant *> LDSVarsToConstantGEP;
310         std::tie(SGV, LDSVarsToConstantGEP) =
311             createLDSVariableReplacement(M, VarName, KernelUsedVariables);
312 
313         removeFromUsedLists(M, KernelUsedVariables);
314         replaceLDSVariablesWithStruct(
315             M, KernelUsedVariables, SGV, LDSVarsToConstantGEP, [&F](Use &U) {
316               Instruction *I = dyn_cast<Instruction>(U.getUser());
317               return I && I->getFunction() == &F;
318             });
319         Changed = true;
320       }
321     }
322 
323     for (auto &GV : make_early_inc_range(M.globals()))
324       if (AMDGPU::isLDSVariableToLower(GV)) {
325         GV.removeDeadConstantUsers();
326         if (GV.use_empty())
327           GV.eraseFromParent();
328       }
329 
330     return Changed;
331   }
332 
333 private:
334   // Increase the alignment of LDS globals if necessary to maximise the chance
335   // that we can use aligned LDS instructions to access them.
336   static bool superAlignLDSGlobals(Module &M) {
337     const DataLayout &DL = M.getDataLayout();
338     bool Changed = false;
339     if (!SuperAlignLDSGlobals) {
340       return Changed;
341     }
342 
343     for (auto &GV : M.globals()) {
344       if (GV.getType()->getPointerAddressSpace() != AMDGPUAS::LOCAL_ADDRESS) {
345         // Only changing alignment of LDS variables
346         continue;
347       }
348       if (!GV.hasInitializer()) {
349         // cuda/hip extern __shared__ variable, leave alignment alone
350         continue;
351       }
352 
353       Align Alignment = AMDGPU::getAlign(DL, &GV);
354       TypeSize GVSize = DL.getTypeAllocSize(GV.getValueType());
355 
356       if (GVSize > 8) {
357         // We might want to use a b96 or b128 load/store
358         Alignment = std::max(Alignment, Align(16));
359       } else if (GVSize > 4) {
360         // We might want to use a b64 load/store
361         Alignment = std::max(Alignment, Align(8));
362       } else if (GVSize > 2) {
363         // We might want to use a b32 load/store
364         Alignment = std::max(Alignment, Align(4));
365       } else if (GVSize > 1) {
366         // We might want to use a b16 load/store
367         Alignment = std::max(Alignment, Align(2));
368       }
369 
370       if (Alignment != AMDGPU::getAlign(DL, &GV)) {
371         Changed = true;
372         GV.setAlignment(Alignment);
373       }
374     }
375     return Changed;
376   }
377 
378   std::tuple<GlobalVariable *, DenseMap<GlobalVariable *, Constant *>>
379   createLDSVariableReplacement(
380       Module &M, std::string VarName,
381       std::vector<GlobalVariable *> const &LDSVarsToTransform) {
382     // Create a struct instance containing LDSVarsToTransform and map from those
383     // variables to ConstantExprGEP
384     // Variables may be introduced to meet alignment requirements. No aliasing
385     // metadata is useful for these as they have no uses. Erased before return.
386 
387     LLVMContext &Ctx = M.getContext();
388     const DataLayout &DL = M.getDataLayout();
389     assert(!LDSVarsToTransform.empty());
390 
391     SmallVector<OptimizedStructLayoutField, 8> LayoutFields;
392     LayoutFields.reserve(LDSVarsToTransform.size());
393     for (GlobalVariable *GV : LDSVarsToTransform) {
394       OptimizedStructLayoutField F(GV, DL.getTypeAllocSize(GV->getValueType()),
395                                    AMDGPU::getAlign(DL, GV));
396       LayoutFields.emplace_back(F);
397     }
398 
399     performOptimizedStructLayout(LayoutFields);
400 
401     std::vector<GlobalVariable *> LocalVars;
402     BitVector IsPaddingField;
403     LocalVars.reserve(LDSVarsToTransform.size()); // will be at least this large
404     IsPaddingField.reserve(LDSVarsToTransform.size());
405     {
406       uint64_t CurrentOffset = 0;
407       for (size_t I = 0; I < LayoutFields.size(); I++) {
408         GlobalVariable *FGV = static_cast<GlobalVariable *>(
409             const_cast<void *>(LayoutFields[I].Id));
410         Align DataAlign = LayoutFields[I].Alignment;
411 
412         uint64_t DataAlignV = DataAlign.value();
413         if (uint64_t Rem = CurrentOffset % DataAlignV) {
414           uint64_t Padding = DataAlignV - Rem;
415 
416           // Append an array of padding bytes to meet alignment requested
417           // Note (o +      (a - (o % a)) ) % a == 0
418           //      (offset + Padding       ) % align == 0
419 
420           Type *ATy = ArrayType::get(Type::getInt8Ty(Ctx), Padding);
421           LocalVars.push_back(new GlobalVariable(
422               M, ATy, false, GlobalValue::InternalLinkage, UndefValue::get(ATy),
423               "", nullptr, GlobalValue::NotThreadLocal, AMDGPUAS::LOCAL_ADDRESS,
424               false));
425           IsPaddingField.push_back(true);
426           CurrentOffset += Padding;
427         }
428 
429         LocalVars.push_back(FGV);
430         IsPaddingField.push_back(false);
431         CurrentOffset += LayoutFields[I].Size;
432       }
433     }
434 
435     std::vector<Type *> LocalVarTypes;
436     LocalVarTypes.reserve(LocalVars.size());
437     std::transform(
438         LocalVars.cbegin(), LocalVars.cend(), std::back_inserter(LocalVarTypes),
439         [](const GlobalVariable *V) -> Type * { return V->getValueType(); });
440 
441     StructType *LDSTy = StructType::create(Ctx, LocalVarTypes, VarName + ".t");
442 
443     Align StructAlign = AMDGPU::getAlign(DL, LocalVars[0]);
444 
445     GlobalVariable *SGV = new GlobalVariable(
446         M, LDSTy, false, GlobalValue::InternalLinkage, UndefValue::get(LDSTy),
447         VarName, nullptr, GlobalValue::NotThreadLocal, AMDGPUAS::LOCAL_ADDRESS,
448         false);
449     SGV->setAlignment(StructAlign);
450 
451     DenseMap<GlobalVariable *, Constant *> Map;
452     Type *I32 = Type::getInt32Ty(Ctx);
453     for (size_t I = 0; I < LocalVars.size(); I++) {
454       GlobalVariable *GV = LocalVars[I];
455       Constant *GEPIdx[] = {ConstantInt::get(I32, 0), ConstantInt::get(I32, I)};
456       Constant *GEP = ConstantExpr::getGetElementPtr(LDSTy, SGV, GEPIdx, true);
457       if (IsPaddingField[I]) {
458         assert(GV->use_empty());
459         GV->eraseFromParent();
460       } else {
461         Map[GV] = GEP;
462       }
463     }
464     assert(Map.size() == LDSVarsToTransform.size());
465     return std::make_tuple(SGV, std::move(Map));
466   }
467 
468   template <typename PredicateTy>
469   void replaceLDSVariablesWithStruct(
470       Module &M, std::vector<GlobalVariable *> const &LDSVarsToTransform,
471       GlobalVariable *SGV,
472       DenseMap<GlobalVariable *, Constant *> &LDSVarsToConstantGEP,
473       PredicateTy Predicate) {
474     LLVMContext &Ctx = M.getContext();
475     const DataLayout &DL = M.getDataLayout();
476 
477     // Create alias.scope and their lists. Each field in the new structure
478     // does not alias with all other fields.
479     SmallVector<MDNode *> AliasScopes;
480     SmallVector<Metadata *> NoAliasList;
481     const size_t NumberVars = LDSVarsToTransform.size();
482     if (NumberVars > 1) {
483       MDBuilder MDB(Ctx);
484       AliasScopes.reserve(NumberVars);
485       MDNode *Domain = MDB.createAnonymousAliasScopeDomain();
486       for (size_t I = 0; I < NumberVars; I++) {
487         MDNode *Scope = MDB.createAnonymousAliasScope(Domain);
488         AliasScopes.push_back(Scope);
489       }
490       NoAliasList.append(&AliasScopes[1], AliasScopes.end());
491     }
492 
493     // Replace uses of ith variable with a constantexpr to the corresponding
494     // field of the instance that will be allocated by AMDGPUMachineFunction
495     for (size_t I = 0; I < NumberVars; I++) {
496       GlobalVariable *GV = LDSVarsToTransform[I];
497       Constant *GEP = LDSVarsToConstantGEP[GV];
498 
499       GV->replaceUsesWithIf(GEP, Predicate);
500       if (GV->use_empty()) {
501         GV->eraseFromParent();
502       }
503 
504       APInt APOff(DL.getIndexTypeSizeInBits(GEP->getType()), 0);
505       GEP->stripAndAccumulateInBoundsConstantOffsets(DL, APOff);
506       uint64_t Offset = APOff.getZExtValue();
507 
508       Align A = commonAlignment(SGV->getAlign().valueOrOne(), Offset);
509 
510       if (I)
511         NoAliasList[I - 1] = AliasScopes[I - 1];
512       MDNode *NoAlias =
513           NoAliasList.empty() ? nullptr : MDNode::get(Ctx, NoAliasList);
514       MDNode *AliasScope =
515           AliasScopes.empty() ? nullptr : MDNode::get(Ctx, {AliasScopes[I]});
516 
517       refineUsesAlignmentAndAA(GEP, A, DL, AliasScope, NoAlias);
518     }
519   }
520 
521   void refineUsesAlignmentAndAA(Value *Ptr, Align A, const DataLayout &DL,
522                                 MDNode *AliasScope, MDNode *NoAlias,
523                                 unsigned MaxDepth = 5) {
524     if (!MaxDepth || (A == 1 && !AliasScope))
525       return;
526 
527     for (User *U : Ptr->users()) {
528       if (auto *I = dyn_cast<Instruction>(U)) {
529         if (AliasScope && I->mayReadOrWriteMemory()) {
530           MDNode *AS = I->getMetadata(LLVMContext::MD_alias_scope);
531           AS = (AS ? MDNode::getMostGenericAliasScope(AS, AliasScope)
532                    : AliasScope);
533           I->setMetadata(LLVMContext::MD_alias_scope, AS);
534 
535           MDNode *NA = I->getMetadata(LLVMContext::MD_noalias);
536           NA = (NA ? MDNode::intersect(NA, NoAlias) : NoAlias);
537           I->setMetadata(LLVMContext::MD_noalias, NA);
538         }
539       }
540 
541       if (auto *LI = dyn_cast<LoadInst>(U)) {
542         LI->setAlignment(std::max(A, LI->getAlign()));
543         continue;
544       }
545       if (auto *SI = dyn_cast<StoreInst>(U)) {
546         if (SI->getPointerOperand() == Ptr)
547           SI->setAlignment(std::max(A, SI->getAlign()));
548         continue;
549       }
550       if (auto *AI = dyn_cast<AtomicRMWInst>(U)) {
551         // None of atomicrmw operations can work on pointers, but let's
552         // check it anyway in case it will or we will process ConstantExpr.
553         if (AI->getPointerOperand() == Ptr)
554           AI->setAlignment(std::max(A, AI->getAlign()));
555         continue;
556       }
557       if (auto *AI = dyn_cast<AtomicCmpXchgInst>(U)) {
558         if (AI->getPointerOperand() == Ptr)
559           AI->setAlignment(std::max(A, AI->getAlign()));
560         continue;
561       }
562       if (auto *GEP = dyn_cast<GetElementPtrInst>(U)) {
563         unsigned BitWidth = DL.getIndexTypeSizeInBits(GEP->getType());
564         APInt Off(BitWidth, 0);
565         if (GEP->getPointerOperand() == Ptr) {
566           Align GA;
567           if (GEP->accumulateConstantOffset(DL, Off))
568             GA = commonAlignment(A, Off.getLimitedValue());
569           refineUsesAlignmentAndAA(GEP, GA, DL, AliasScope, NoAlias,
570                                    MaxDepth - 1);
571         }
572         continue;
573       }
574       if (auto *I = dyn_cast<Instruction>(U)) {
575         if (I->getOpcode() == Instruction::BitCast ||
576             I->getOpcode() == Instruction::AddrSpaceCast)
577           refineUsesAlignmentAndAA(I, A, DL, AliasScope, NoAlias, MaxDepth - 1);
578       }
579     }
580   }
581 };
582 
583 } // namespace
584 char AMDGPULowerModuleLDS::ID = 0;
585 
586 char &llvm::AMDGPULowerModuleLDSID = AMDGPULowerModuleLDS::ID;
587 
588 INITIALIZE_PASS(AMDGPULowerModuleLDS, DEBUG_TYPE,
589                 "Lower uses of LDS variables from non-kernel functions", false,
590                 false)
591 
592 ModulePass *llvm::createAMDGPULowerModuleLDSPass() {
593   return new AMDGPULowerModuleLDS();
594 }
595 
596 PreservedAnalyses AMDGPULowerModuleLDSPass::run(Module &M,
597                                                 ModuleAnalysisManager &) {
598   return AMDGPULowerModuleLDS().runOnModule(M) ? PreservedAnalyses::none()
599                                                : PreservedAnalyses::all();
600 }
601