xref: /llvm-project/llvm/lib/Target/AMDGPU/AMDGPULowerModuleLDSPass.cpp (revision 923b90bddbc39eb76fd8a303a16182cd518b3257)
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/Analysis/CallGraph.h"
36 #include "llvm/IR/Constants.h"
37 #include "llvm/IR/DerivedTypes.h"
38 #include "llvm/IR/IRBuilder.h"
39 #include "llvm/IR/InlineAsm.h"
40 #include "llvm/IR/Instructions.h"
41 #include "llvm/IR/MDBuilder.h"
42 #include "llvm/InitializePasses.h"
43 #include "llvm/Pass.h"
44 #include "llvm/Support/CommandLine.h"
45 #include "llvm/Support/Debug.h"
46 #include "llvm/Support/OptimizedStructLayout.h"
47 #include "llvm/Transforms/Utils/ModuleUtils.h"
48 #include <tuple>
49 #include <vector>
50 
51 #define DEBUG_TYPE "amdgpu-lower-module-lds"
52 
53 using namespace llvm;
54 
55 static cl::opt<bool> SuperAlignLDSGlobals(
56     "amdgpu-super-align-lds-globals",
57     cl::desc("Increase alignment of LDS if it is not on align boundary"),
58     cl::init(true), cl::Hidden);
59 
60 namespace {
61 class AMDGPULowerModuleLDS : public ModulePass {
62 
63   static void removeFromUsedList(Module &M, StringRef Name,
64                                  SmallPtrSetImpl<Constant *> &ToRemove) {
65     GlobalVariable *GV = M.getNamedGlobal(Name);
66     if (!GV || ToRemove.empty()) {
67       return;
68     }
69 
70     SmallVector<Constant *, 16> Init;
71     auto *CA = cast<ConstantArray>(GV->getInitializer());
72     for (auto &Op : CA->operands()) {
73       // ModuleUtils::appendToUsed only inserts Constants
74       Constant *C = cast<Constant>(Op);
75       if (!ToRemove.contains(C->stripPointerCasts())) {
76         Init.push_back(C);
77       }
78     }
79 
80     if (Init.size() == CA->getNumOperands()) {
81       return; // none to remove
82     }
83 
84     GV->eraseFromParent();
85 
86     for (Constant *C : ToRemove) {
87       C->removeDeadConstantUsers();
88     }
89 
90     if (!Init.empty()) {
91       ArrayType *ATy =
92           ArrayType::get(Type::getInt8PtrTy(M.getContext()), Init.size());
93       GV =
94           new llvm::GlobalVariable(M, ATy, false, GlobalValue::AppendingLinkage,
95                                    ConstantArray::get(ATy, Init), Name);
96       GV->setSection("llvm.metadata");
97     }
98   }
99 
100   static void
101   removeFromUsedLists(Module &M,
102                       const std::vector<GlobalVariable *> &LocalVars) {
103     SmallPtrSet<Constant *, 32> LocalVarsSet;
104     for (GlobalVariable *LocalVar : LocalVars)
105       if (Constant *C = dyn_cast<Constant>(LocalVar->stripPointerCasts()))
106         LocalVarsSet.insert(C);
107     removeFromUsedList(M, "llvm.used", LocalVarsSet);
108     removeFromUsedList(M, "llvm.compiler.used", LocalVarsSet);
109   }
110 
111   static void markUsedByKernel(IRBuilder<> &Builder, Function *Func,
112                                GlobalVariable *SGV) {
113     // The llvm.amdgcn.module.lds instance is implicitly used by all kernels
114     // that might call a function which accesses a field within it. This is
115     // presently approximated to 'all kernels' if there are any such functions
116     // in the module. This implicit use is redefined as an explicit use here so
117     // that later passes, specifically PromoteAlloca, account for the required
118     // memory without any knowledge of this transform.
119 
120     // An operand bundle on llvm.donothing works because the call instruction
121     // survives until after the last pass that needs to account for LDS. It is
122     // better than inline asm as the latter survives until the end of codegen. A
123     // totally robust solution would be a function with the same semantics as
124     // llvm.donothing that takes a pointer to the instance and is lowered to a
125     // no-op after LDS is allocated, but that is not presently necessary.
126 
127     LLVMContext &Ctx = Func->getContext();
128 
129     Builder.SetInsertPoint(Func->getEntryBlock().getFirstNonPHI());
130 
131     FunctionType *FTy = FunctionType::get(Type::getVoidTy(Ctx), {});
132 
133     Function *Decl =
134         Intrinsic::getDeclaration(Func->getParent(), Intrinsic::donothing, {});
135 
136     Value *UseInstance[1] = {Builder.CreateInBoundsGEP(
137         SGV->getValueType(), SGV, ConstantInt::get(Type::getInt32Ty(Ctx), 0))};
138 
139     Builder.CreateCall(FTy, Decl, {},
140                        {OperandBundleDefT<Value *>("ExplicitUse", UseInstance)},
141                        "");
142   }
143 
144 public:
145   static char ID;
146 
147   AMDGPULowerModuleLDS() : ModulePass(ID) {
148     initializeAMDGPULowerModuleLDSPass(*PassRegistry::getPassRegistry());
149   }
150 
151   bool runOnModule(Module &M) override {
152     CallGraph CG = CallGraph(M);
153     bool Changed = superAlignLDSGlobals(M);
154     std::vector<GlobalVariable *> ModuleScopeVariables =
155         AMDGPU::findVariablesToLower(M, nullptr);
156     Changed |= processUsedLDS(CG, M, ModuleScopeVariables);
157 
158     for (Function &F : M.functions()) {
159       if (F.isDeclaration())
160         continue;
161 
162       // Only lower compute kernels' LDS.
163       if (!AMDGPU::isKernel(F.getCallingConv()))
164         continue;
165       std::vector<GlobalVariable *> KernelUsedVariables =
166           AMDGPU::findVariablesToLower(M, &F);
167       Changed |= processUsedLDS(CG, M, KernelUsedVariables, &F);
168     }
169 
170     return Changed;
171   }
172 
173 private:
174   // Increase the alignment of LDS globals if necessary to maximise the chance
175   // that we can use aligned LDS instructions to access them.
176   static bool superAlignLDSGlobals(Module &M) {
177     const DataLayout &DL = M.getDataLayout();
178     bool Changed = false;
179     if (!SuperAlignLDSGlobals) {
180       return Changed;
181     }
182 
183     for (auto &GV : M.globals()) {
184       if (GV.getType()->getPointerAddressSpace() != AMDGPUAS::LOCAL_ADDRESS) {
185         // Only changing alignment of LDS variables
186         continue;
187       }
188       if (!GV.hasInitializer()) {
189         // cuda/hip extern __shared__ variable, leave alignment alone
190         continue;
191       }
192 
193       Align Alignment = AMDGPU::getAlign(DL, &GV);
194       TypeSize GVSize = DL.getTypeAllocSize(GV.getValueType());
195 
196       if (GVSize > 8) {
197         // We might want to use a b96 or b128 load/store
198         Alignment = std::max(Alignment, Align(16));
199       } else if (GVSize > 4) {
200         // We might want to use a b64 load/store
201         Alignment = std::max(Alignment, Align(8));
202       } else if (GVSize > 2) {
203         // We might want to use a b32 load/store
204         Alignment = std::max(Alignment, Align(4));
205       } else if (GVSize > 1) {
206         // We might want to use a b16 load/store
207         Alignment = std::max(Alignment, Align(2));
208       }
209 
210       if (Alignment != AMDGPU::getAlign(DL, &GV)) {
211         Changed = true;
212         GV.setAlignment(Alignment);
213       }
214     }
215     return Changed;
216   }
217 
218   std::tuple<GlobalVariable *, DenseMap<GlobalVariable *, Constant *>>
219   createLDSVariableReplacement(
220       Module &M, std::string VarName,
221       std::vector<GlobalVariable *> const &LDSVarsToTransform) {
222     // Create a struct instance containing LDSVarsToTransform and map from those
223     // variables to ConstantExprGEP
224     // Variables may be introduced to meet alignment requirements. No aliasing
225     // metadata is useful for these as they have no uses. Erased before return.
226 
227     LLVMContext &Ctx = M.getContext();
228     const DataLayout &DL = M.getDataLayout();
229     assert(!LDSVarsToTransform.empty());
230 
231     SmallVector<OptimizedStructLayoutField, 8> LayoutFields;
232     LayoutFields.reserve(LDSVarsToTransform.size());
233     for (GlobalVariable *GV : LDSVarsToTransform) {
234       OptimizedStructLayoutField F(GV, DL.getTypeAllocSize(GV->getValueType()),
235                                    AMDGPU::getAlign(DL, GV));
236       LayoutFields.emplace_back(F);
237     }
238 
239     performOptimizedStructLayout(LayoutFields);
240 
241     std::vector<GlobalVariable *> LocalVars;
242     BitVector IsPaddingField;
243     LocalVars.reserve(LDSVarsToTransform.size()); // will be at least this large
244     IsPaddingField.reserve(LDSVarsToTransform.size());
245     {
246       uint64_t CurrentOffset = 0;
247       for (size_t I = 0; I < LayoutFields.size(); I++) {
248         GlobalVariable *FGV = static_cast<GlobalVariable *>(
249             const_cast<void *>(LayoutFields[I].Id));
250         Align DataAlign = LayoutFields[I].Alignment;
251 
252         uint64_t DataAlignV = DataAlign.value();
253         if (uint64_t Rem = CurrentOffset % DataAlignV) {
254           uint64_t Padding = DataAlignV - Rem;
255 
256           // Append an array of padding bytes to meet alignment requested
257           // Note (o +      (a - (o % a)) ) % a == 0
258           //      (offset + Padding       ) % align == 0
259 
260           Type *ATy = ArrayType::get(Type::getInt8Ty(Ctx), Padding);
261           LocalVars.push_back(new GlobalVariable(
262               M, ATy, false, GlobalValue::InternalLinkage, UndefValue::get(ATy),
263               "", nullptr, GlobalValue::NotThreadLocal, AMDGPUAS::LOCAL_ADDRESS,
264               false));
265           IsPaddingField.push_back(true);
266           CurrentOffset += Padding;
267         }
268 
269         LocalVars.push_back(FGV);
270         IsPaddingField.push_back(false);
271         CurrentOffset += LayoutFields[I].Size;
272       }
273     }
274 
275     std::vector<Type *> LocalVarTypes;
276     LocalVarTypes.reserve(LocalVars.size());
277     std::transform(
278         LocalVars.cbegin(), LocalVars.cend(), std::back_inserter(LocalVarTypes),
279         [](const GlobalVariable *V) -> Type * { return V->getValueType(); });
280 
281     StructType *LDSTy = StructType::create(Ctx, LocalVarTypes, VarName + ".t");
282 
283     Align StructAlign =
284         AMDGPU::getAlign(DL, LocalVars[0]);
285 
286     GlobalVariable *SGV = new GlobalVariable(
287         M, LDSTy, false, GlobalValue::InternalLinkage, UndefValue::get(LDSTy),
288         VarName, nullptr, GlobalValue::NotThreadLocal, AMDGPUAS::LOCAL_ADDRESS,
289         false);
290     SGV->setAlignment(StructAlign);
291 
292     DenseMap<GlobalVariable *, Constant *> Map;
293     Type *I32 = Type::getInt32Ty(Ctx);
294     for (size_t I = 0; I < LocalVars.size(); I++) {
295       GlobalVariable *GV = LocalVars[I];
296       Constant *GEPIdx[] = {ConstantInt::get(I32, 0), ConstantInt::get(I32, I)};
297       Constant *GEP = ConstantExpr::getGetElementPtr(LDSTy, SGV, GEPIdx, true);
298       if (IsPaddingField[I]) {
299         assert(GV->use_empty());
300         GV->eraseFromParent();
301       } else {
302         Map[GV] = GEP;
303       }
304     }
305     assert(Map.size() == LDSVarsToTransform.size());
306     return {SGV, std::move(Map)};
307   }
308 
309   bool processUsedLDS(CallGraph const &CG, Module &M,
310                       std::vector<GlobalVariable *> const &LDSVarsToTransform,
311                       Function *F = nullptr) {
312     LLVMContext &Ctx = M.getContext();
313     const DataLayout &DL = M.getDataLayout();
314 
315     if (LDSVarsToTransform.empty()) {
316       // No variables to rewrite, no changes made.
317       return false;
318     }
319 
320     std::string VarName(
321         F ? (Twine("llvm.amdgcn.kernel.") + F->getName() + ".lds").str()
322           : "llvm.amdgcn.module.lds");
323 
324     GlobalVariable *SGV;
325     DenseMap<GlobalVariable *, Constant *> LDSVarToConstantGEP;
326     std::tie(SGV, LDSVarToConstantGEP) =
327         createLDSVariableReplacement(M, VarName, LDSVarsToTransform);
328 
329     if (!F) {
330       appendToCompilerUsed(
331           M, {static_cast<GlobalValue *>(
332                  ConstantExpr::getPointerBitCastOrAddrSpaceCast(
333                      cast<Constant>(SGV), Type::getInt8PtrTy(Ctx)))});
334     }
335 
336     // The verifier rejects used lists containing an inttoptr of a constant
337     // so remove the variables from these lists before replaceAllUsesWith
338     removeFromUsedLists(M, LDSVarsToTransform);
339 
340     // Create alias.scope and their lists. Each field in the new structure
341     // does not alias with all other fields.
342     SmallVector<MDNode *> AliasScopes;
343     SmallVector<Metadata *> NoAliasList;
344     const size_t NumberVars = LDSVarsToTransform.size();
345     if (NumberVars > 1) {
346       MDBuilder MDB(Ctx);
347       AliasScopes.reserve(NumberVars);
348       MDNode *Domain = MDB.createAnonymousAliasScopeDomain();
349       for (size_t I = 0; I < NumberVars; I++) {
350         MDNode *Scope = MDB.createAnonymousAliasScope(Domain);
351         AliasScopes.push_back(Scope);
352       }
353       NoAliasList.append(&AliasScopes[1], AliasScopes.end());
354     }
355 
356     // Replace uses of ith variable with a constantexpr to the corresponding
357     // field of the instance that will be allocated by AMDGPUMachineFunction
358     for (size_t I = 0; I < NumberVars; I++) {
359       GlobalVariable *GV = LDSVarsToTransform[I];
360       Constant *GEP = LDSVarToConstantGEP[GV];
361 
362       if (F) {
363         // Replace all constant uses with instructions if they belong to the
364         // current kernel.
365         for (User *U : make_early_inc_range(GV->users())) {
366           if (ConstantExpr *C = dyn_cast<ConstantExpr>(U))
367             AMDGPU::replaceConstantUsesInFunction(C, F);
368         }
369 
370         GV->removeDeadConstantUsers();
371 
372         GV->replaceUsesWithIf(GEP, [F](Use &U) {
373           Instruction *I = dyn_cast<Instruction>(U.getUser());
374           return I && I->getFunction() == F;
375         });
376       } else {
377         GV->replaceAllUsesWith(GEP);
378       }
379       if (GV->use_empty()) {
380         GV->eraseFromParent();
381       }
382 
383       APInt APOff(DL.getIndexTypeSizeInBits(GEP->getType()), 0);
384       GEP->stripAndAccumulateInBoundsConstantOffsets(DL, APOff);
385       uint64_t Offset = APOff.getZExtValue();
386 
387       Align A = commonAlignment(SGV->getAlign().valueOrOne(), Offset);
388 
389       if (I)
390         NoAliasList[I - 1] = AliasScopes[I - 1];
391       MDNode *NoAlias =
392           NoAliasList.empty() ? nullptr : MDNode::get(Ctx, NoAliasList);
393       MDNode *AliasScope =
394           AliasScopes.empty() ? nullptr : MDNode::get(Ctx, {AliasScopes[I]});
395 
396       refineUsesAlignmentAndAA(GEP, A, DL, AliasScope, NoAlias);
397     }
398 
399     // This ensures the variable is allocated when called functions access it.
400     // It also lets other passes, specifically PromoteAlloca, accurately
401     // calculate how much LDS will be used by the kernel after lowering.
402     if (!F) {
403       IRBuilder<> Builder(Ctx);
404       for (Function &Func : M.functions()) {
405         if (!Func.isDeclaration() && AMDGPU::isKernelCC(&Func)) {
406           const CallGraphNode *N = CG[&Func];
407           const bool CalleesRequireModuleLDS = N->size() > 0;
408 
409           if (CalleesRequireModuleLDS) {
410             // If a function this kernel might call requires module LDS,
411             // annotate the kernel to let later passes know it will allocate
412             // this structure, even if not apparent from the IR.
413             markUsedByKernel(Builder, &Func, SGV);
414           } else {
415             // However if we are certain this kernel cannot call a function that
416             // requires module LDS, annotate the kernel so the backend can elide
417             // the allocation without repeating callgraph walks.
418             Func.addFnAttr("amdgpu-elide-module-lds");
419           }
420         }
421       }
422     }
423     return true;
424   }
425 
426   void refineUsesAlignmentAndAA(Value *Ptr, Align A, const DataLayout &DL,
427                                 MDNode *AliasScope, MDNode *NoAlias,
428                                 unsigned MaxDepth = 5) {
429     if (!MaxDepth || (A == 1 && !AliasScope))
430       return;
431 
432     for (User *U : Ptr->users()) {
433       if (auto *I = dyn_cast<Instruction>(U)) {
434         if (AliasScope && I->mayReadOrWriteMemory()) {
435           MDNode *AS = I->getMetadata(LLVMContext::MD_alias_scope);
436           AS = (AS ? MDNode::getMostGenericAliasScope(AS, AliasScope)
437                    : AliasScope);
438           I->setMetadata(LLVMContext::MD_alias_scope, AS);
439 
440           MDNode *NA = I->getMetadata(LLVMContext::MD_noalias);
441           NA = (NA ? MDNode::intersect(NA, NoAlias) : NoAlias);
442           I->setMetadata(LLVMContext::MD_noalias, NA);
443         }
444       }
445 
446       if (auto *LI = dyn_cast<LoadInst>(U)) {
447         LI->setAlignment(std::max(A, LI->getAlign()));
448         continue;
449       }
450       if (auto *SI = dyn_cast<StoreInst>(U)) {
451         if (SI->getPointerOperand() == Ptr)
452           SI->setAlignment(std::max(A, SI->getAlign()));
453         continue;
454       }
455       if (auto *AI = dyn_cast<AtomicRMWInst>(U)) {
456         // None of atomicrmw operations can work on pointers, but let's
457         // check it anyway in case it will or we will process ConstantExpr.
458         if (AI->getPointerOperand() == Ptr)
459           AI->setAlignment(std::max(A, AI->getAlign()));
460         continue;
461       }
462       if (auto *AI = dyn_cast<AtomicCmpXchgInst>(U)) {
463         if (AI->getPointerOperand() == Ptr)
464           AI->setAlignment(std::max(A, AI->getAlign()));
465         continue;
466       }
467       if (auto *GEP = dyn_cast<GetElementPtrInst>(U)) {
468         unsigned BitWidth = DL.getIndexTypeSizeInBits(GEP->getType());
469         APInt Off(BitWidth, 0);
470         if (GEP->getPointerOperand() == Ptr) {
471           Align GA;
472           if (GEP->accumulateConstantOffset(DL, Off))
473             GA = commonAlignment(A, Off.getLimitedValue());
474           refineUsesAlignmentAndAA(GEP, GA, DL, AliasScope, NoAlias,
475                                    MaxDepth - 1);
476         }
477         continue;
478       }
479       if (auto *I = dyn_cast<Instruction>(U)) {
480         if (I->getOpcode() == Instruction::BitCast ||
481             I->getOpcode() == Instruction::AddrSpaceCast)
482           refineUsesAlignmentAndAA(I, A, DL, AliasScope, NoAlias, MaxDepth - 1);
483       }
484     }
485   }
486 };
487 
488 } // namespace
489 char AMDGPULowerModuleLDS::ID = 0;
490 
491 char &llvm::AMDGPULowerModuleLDSID = AMDGPULowerModuleLDS::ID;
492 
493 INITIALIZE_PASS(AMDGPULowerModuleLDS, DEBUG_TYPE,
494                 "Lower uses of LDS variables from non-kernel functions", false,
495                 false)
496 
497 ModulePass *llvm::createAMDGPULowerModuleLDSPass() {
498   return new AMDGPULowerModuleLDS();
499 }
500 
501 PreservedAnalyses AMDGPULowerModuleLDSPass::run(Module &M,
502                                                 ModuleAnalysisManager &) {
503   return AMDGPULowerModuleLDS().runOnModule(M) ? PreservedAnalyses::none()
504                                                : PreservedAnalyses::all();
505 }
506