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 local data store, LDS, uses from non-kernel functions. 10 // LDS is contiguous memory allocated per kernel execution. 11 // 12 // Background. 13 // 14 // The programming model is global variables, or equivalently function local 15 // static variables, accessible from kernels or other functions. For uses from 16 // kernels this is straightforward - assign an integer to the kernel for the 17 // memory required by all the variables combined, allocate them within that. 18 // For uses from functions there are performance tradeoffs to choose between. 19 // 20 // This model means the GPU runtime can specify the amount of memory allocated. 21 // If this is more than the kernel assumed, the excess can be made available 22 // using a language specific feature, which IR represents as a variable with 23 // no initializer. This feature is referred to here as "Dynamic LDS" and is 24 // lowered slightly differently to the normal case. 25 // 26 // Consequences of this GPU feature: 27 // - memory is limited and exceeding it halts compilation 28 // - a global accessed by one kernel exists independent of other kernels 29 // - a global exists independent of simultaneous execution of the same kernel 30 // - the address of the global may be different from different kernels as they 31 // do not alias, which permits only allocating variables they use 32 // - if the address is allowed to differ, functions need help to find it 33 // 34 // Uses from kernels are implemented here by grouping them in a per-kernel 35 // struct instance. This duplicates the variables, accurately modelling their 36 // aliasing properties relative to a single global representation. It also 37 // permits control over alignment via padding. 38 // 39 // Uses from functions are more complicated and the primary purpose of this 40 // IR pass. Several different lowering are chosen between to meet requirements 41 // to avoid allocating any LDS where it is not necessary, as that impacts 42 // occupancy and may fail the compilation, while not imposing overhead on a 43 // feature whose primary advantage over global memory is performance. The basic 44 // design goal is to avoid one kernel imposing overhead on another. 45 // 46 // Implementation. 47 // 48 // LDS variables with constant annotation or non-undef initializer are passed 49 // through unchanged for simplification or error diagnostics in later passes. 50 // Non-undef initializers are not yet implemented for LDS. 51 // 52 // LDS variables that are always allocated at the same address can be found 53 // by lookup at that address. Otherwise runtime information/cost is required. 54 // 55 // The simplest strategy possible is to group all LDS variables in a single 56 // struct and allocate that struct in every kernel such that the original 57 // variables are always at the same address. LDS is however a limited resource 58 // so this strategy is unusable in practice. It is not implemented here. 59 // 60 // Strategy | Precise allocation | Zero runtime cost | General purpose | 61 // --------+--------------------+-------------------+-----------------+ 62 // Module | No | Yes | Yes | 63 // Table | Yes | No | Yes | 64 // Kernel | Yes | Yes | No | 65 // Hybrid | Yes | Partial | Yes | 66 // 67 // "Module" spends LDS memory to save cycles. "Table" spends cycles and global 68 // memory to save LDS. "Kernel" is as fast as kernel allocation but only works 69 // for variables that are known reachable from a single kernel. "Hybrid" picks 70 // between all three. When forced to choose between LDS and cycles we minimise 71 // LDS use. 72 73 // The "module" lowering implemented here finds LDS variables which are used by 74 // non-kernel functions and creates a new struct with a field for each of those 75 // LDS variables. Variables that are only used from kernels are excluded. 76 // 77 // The "table" lowering implemented here has three components. 78 // First kernels are assigned a unique integer identifier which is available in 79 // functions it calls through the intrinsic amdgcn_lds_kernel_id. The integer 80 // is passed through a specific SGPR, thus works with indirect calls. 81 // Second, each kernel allocates LDS variables independent of other kernels and 82 // writes the addresses it chose for each variable into an array in consistent 83 // order. If the kernel does not allocate a given variable, it writes undef to 84 // the corresponding array location. These arrays are written to a constant 85 // table in the order matching the kernel unique integer identifier. 86 // Third, uses from non-kernel functions are replaced with a table lookup using 87 // the intrinsic function to find the address of the variable. 88 // 89 // "Kernel" lowering is only applicable for variables that are unambiguously 90 // reachable from exactly one kernel. For those cases, accesses to the variable 91 // can be lowered to ConstantExpr address of a struct instance specific to that 92 // one kernel. This is zero cost in space and in compute. It will raise a fatal 93 // error on any variable that might be reachable from multiple kernels and is 94 // thus most easily used as part of the hybrid lowering strategy. 95 // 96 // Hybrid lowering is a mixture of the above. It uses the zero cost kernel 97 // lowering where it can. It lowers the variable accessed by the greatest 98 // number of kernels using the module strategy as that is free for the first 99 // variable. Any futher variables that can be lowered with the module strategy 100 // without incurring LDS memory overhead are. The remaining ones are lowered 101 // via table. 102 // 103 // Consequences 104 // - No heuristics or user controlled magic numbers, hybrid is the right choice 105 // - Kernels that don't use functions (or have had them all inlined) are not 106 // affected by any lowering for kernels that do. 107 // - Kernels that don't make indirect function calls are not affected by those 108 // that do. 109 // - Variables which are used by lots of kernels, e.g. those injected by a 110 // language runtime in most kernels, are expected to have no overhead 111 // - Implementations that instantiate templates per-kernel where those templates 112 // use LDS are expected to hit the "Kernel" lowering strategy 113 // - The runtime properties impose a cost in compiler implementation complexity 114 // 115 // Dynamic LDS implementation 116 // Dynamic LDS is lowered similarly to the "table" strategy above and uses the 117 // same intrinsic to identify which kernel is at the root of the dynamic call 118 // graph. This relies on the specified behaviour that all dynamic LDS variables 119 // alias one another, i.e. are at the same address, with respect to a given 120 // kernel. Therefore this pass creates new dynamic LDS variables for each kernel 121 // that allocates any dynamic LDS and builds a table of addresses out of those. 122 // The AMDGPUPromoteAlloca pass skips kernels that use dynamic LDS. 123 // The corresponding optimisation for "kernel" lowering where the table lookup 124 // is elided is not implemented. 125 // 126 // 127 // Implementation notes / limitations 128 // A single LDS global variable represents an instance per kernel that can reach 129 // said variables. This pass essentially specialises said variables per kernel. 130 // Handling ConstantExpr during the pass complicated this significantly so now 131 // all ConstantExpr uses of LDS variables are expanded to instructions. This 132 // may need amending when implementing non-undef initialisers. 133 // 134 // Lowering is split between this IR pass and the back end. This pass chooses 135 // where given variables should be allocated and marks them with metadata, 136 // MD_absolute_symbol. The backend places the variables in coincidentally the 137 // same location and raises a fatal error if something has gone awry. This works 138 // in practice because the only pass between this one and the backend that 139 // changes LDS is PromoteAlloca and the changes it makes do not conflict. 140 // 141 // Addresses are written to constant global arrays based on the same metadata. 142 // 143 // The backend lowers LDS variables in the order of traversal of the function. 144 // This is at odds with the deterministic layout required. The workaround is to 145 // allocate the fixed-address variables immediately upon starting the function 146 // where they can be placed as intended. This requires a means of mapping from 147 // the function to the variables that it allocates. For the module scope lds, 148 // this is via metadata indicating whether the variable is not required. If a 149 // pass deletes that metadata, a fatal error on disagreement with the absolute 150 // symbol metadata will occur. For kernel scope and dynamic, this is by _name_ 151 // correspondence between the function and the variable. It requires the 152 // kernel to have a name (which is only a limitation for tests in practice) and 153 // for nothing to rename the corresponding symbols. This is a hazard if the pass 154 // is run multiple times during debugging. Alternative schemes considered all 155 // involve bespoke metadata. 156 // 157 // If the name correspondence can be replaced, multiple distinct kernels that 158 // have the same memory layout can map to the same kernel id (as the address 159 // itself is handled by the absolute symbol metadata) and that will allow more 160 // uses of the "kernel" style faster lowering and reduce the size of the lookup 161 // tables. 162 // 163 // There is a test that checks this does not fire for a graphics shader. This 164 // lowering is expected to work for graphics if the isKernel test is changed. 165 // 166 // The current markUsedByKernel is sufficient for PromoteAlloca but is elided 167 // before codegen. Replacing this with an equivalent intrinsic which lasts until 168 // shortly after the machine function lowering of LDS would help break the name 169 // mapping. The other part needed is probably to amend PromoteAlloca to embed 170 // the LDS variables it creates in the same struct created here. That avoids the 171 // current hazard where a PromoteAlloca LDS variable might be allocated before 172 // the kernel scope (and thus error on the address check). Given a new invariant 173 // that no LDS variables exist outside of the structs managed here, and an 174 // intrinsic that lasts until after the LDS frame lowering, it should be 175 // possible to drop the name mapping and fold equivalent memory layouts. 176 // 177 //===----------------------------------------------------------------------===// 178 179 #include "AMDGPU.h" 180 #include "AMDGPUMemoryUtils.h" 181 #include "AMDGPUTargetMachine.h" 182 #include "Utils/AMDGPUBaseInfo.h" 183 #include "llvm/ADT/BitVector.h" 184 #include "llvm/ADT/DenseMap.h" 185 #include "llvm/ADT/DenseSet.h" 186 #include "llvm/ADT/STLExtras.h" 187 #include "llvm/ADT/SetOperations.h" 188 #include "llvm/Analysis/CallGraph.h" 189 #include "llvm/CodeGen/TargetPassConfig.h" 190 #include "llvm/IR/Constants.h" 191 #include "llvm/IR/DerivedTypes.h" 192 #include "llvm/IR/IRBuilder.h" 193 #include "llvm/IR/InlineAsm.h" 194 #include "llvm/IR/Instructions.h" 195 #include "llvm/IR/IntrinsicsAMDGPU.h" 196 #include "llvm/IR/MDBuilder.h" 197 #include "llvm/IR/ReplaceConstant.h" 198 #include "llvm/InitializePasses.h" 199 #include "llvm/Pass.h" 200 #include "llvm/Support/CommandLine.h" 201 #include "llvm/Support/Debug.h" 202 #include "llvm/Support/Format.h" 203 #include "llvm/Support/OptimizedStructLayout.h" 204 #include "llvm/Support/raw_ostream.h" 205 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 206 #include "llvm/Transforms/Utils/ModuleUtils.h" 207 208 #include <vector> 209 210 #include <cstdio> 211 212 #define DEBUG_TYPE "amdgpu-lower-module-lds" 213 214 using namespace llvm; 215 using namespace AMDGPU; 216 217 namespace { 218 219 cl::opt<bool> SuperAlignLDSGlobals( 220 "amdgpu-super-align-lds-globals", 221 cl::desc("Increase alignment of LDS if it is not on align boundary"), 222 cl::init(true), cl::Hidden); 223 224 enum class LoweringKind { module, table, kernel, hybrid }; 225 cl::opt<LoweringKind> LoweringKindLoc( 226 "amdgpu-lower-module-lds-strategy", 227 cl::desc("Specify lowering strategy for function LDS access:"), cl::Hidden, 228 cl::init(LoweringKind::hybrid), 229 cl::values( 230 clEnumValN(LoweringKind::table, "table", "Lower via table lookup"), 231 clEnumValN(LoweringKind::module, "module", "Lower via module struct"), 232 clEnumValN( 233 LoweringKind::kernel, "kernel", 234 "Lower variables reachable from one kernel, otherwise abort"), 235 clEnumValN(LoweringKind::hybrid, "hybrid", 236 "Lower via mixture of above strategies"))); 237 238 template <typename T> std::vector<T> sortByName(std::vector<T> &&V) { 239 llvm::sort(V.begin(), V.end(), [](const auto *L, const auto *R) { 240 return L->getName() < R->getName(); 241 }); 242 return {std::move(V)}; 243 } 244 245 class AMDGPULowerModuleLDS { 246 const AMDGPUTargetMachine &TM; 247 248 static void 249 removeLocalVarsFromUsedLists(Module &M, 250 const DenseSet<GlobalVariable *> &LocalVars) { 251 // The verifier rejects used lists containing an inttoptr of a constant 252 // so remove the variables from these lists before replaceAllUsesWith 253 SmallPtrSet<Constant *, 8> LocalVarsSet; 254 for (GlobalVariable *LocalVar : LocalVars) 255 LocalVarsSet.insert(cast<Constant>(LocalVar->stripPointerCasts())); 256 257 removeFromUsedLists( 258 M, [&LocalVarsSet](Constant *C) { return LocalVarsSet.count(C); }); 259 260 for (GlobalVariable *LocalVar : LocalVars) 261 LocalVar->removeDeadConstantUsers(); 262 } 263 264 static void markUsedByKernel(Function *Func, GlobalVariable *SGV) { 265 // The llvm.amdgcn.module.lds instance is implicitly used by all kernels 266 // that might call a function which accesses a field within it. This is 267 // presently approximated to 'all kernels' if there are any such functions 268 // in the module. This implicit use is redefined as an explicit use here so 269 // that later passes, specifically PromoteAlloca, account for the required 270 // memory without any knowledge of this transform. 271 272 // An operand bundle on llvm.donothing works because the call instruction 273 // survives until after the last pass that needs to account for LDS. It is 274 // better than inline asm as the latter survives until the end of codegen. A 275 // totally robust solution would be a function with the same semantics as 276 // llvm.donothing that takes a pointer to the instance and is lowered to a 277 // no-op after LDS is allocated, but that is not presently necessary. 278 279 // This intrinsic is eliminated shortly before instruction selection. It 280 // does not suffice to indicate to ISel that a given global which is not 281 // immediately used by the kernel must still be allocated by it. An 282 // equivalent target specific intrinsic which lasts until immediately after 283 // codegen would suffice for that, but one would still need to ensure that 284 // the variables are allocated in the anticipated order. 285 BasicBlock *Entry = &Func->getEntryBlock(); 286 IRBuilder<> Builder(Entry, Entry->getFirstNonPHIIt()); 287 288 Function *Decl = 289 Intrinsic::getDeclaration(Func->getParent(), Intrinsic::donothing, {}); 290 291 Value *UseInstance[1] = { 292 Builder.CreateConstInBoundsGEP1_32(SGV->getValueType(), SGV, 0)}; 293 294 Builder.CreateCall( 295 Decl, {}, {OperandBundleDefT<Value *>("ExplicitUse", UseInstance)}); 296 } 297 298 public: 299 AMDGPULowerModuleLDS(const AMDGPUTargetMachine &TM_) : TM(TM_) {} 300 301 struct LDSVariableReplacement { 302 GlobalVariable *SGV = nullptr; 303 DenseMap<GlobalVariable *, Constant *> LDSVarsToConstantGEP; 304 }; 305 306 // remap from lds global to a constantexpr gep to where it has been moved to 307 // for each kernel 308 // an array with an element for each kernel containing where the corresponding 309 // variable was remapped to 310 311 static Constant *getAddressesOfVariablesInKernel( 312 LLVMContext &Ctx, ArrayRef<GlobalVariable *> Variables, 313 const DenseMap<GlobalVariable *, Constant *> &LDSVarsToConstantGEP) { 314 // Create a ConstantArray containing the address of each Variable within the 315 // kernel corresponding to LDSVarsToConstantGEP, or poison if that kernel 316 // does not allocate it 317 // TODO: Drop the ptrtoint conversion 318 319 Type *I32 = Type::getInt32Ty(Ctx); 320 321 ArrayType *KernelOffsetsType = ArrayType::get(I32, Variables.size()); 322 323 SmallVector<Constant *> Elements; 324 for (GlobalVariable *GV : Variables) { 325 auto ConstantGepIt = LDSVarsToConstantGEP.find(GV); 326 if (ConstantGepIt != LDSVarsToConstantGEP.end()) { 327 auto elt = ConstantExpr::getPtrToInt(ConstantGepIt->second, I32); 328 Elements.push_back(elt); 329 } else { 330 Elements.push_back(PoisonValue::get(I32)); 331 } 332 } 333 return ConstantArray::get(KernelOffsetsType, Elements); 334 } 335 336 static GlobalVariable *buildLookupTable( 337 Module &M, ArrayRef<GlobalVariable *> Variables, 338 ArrayRef<Function *> kernels, 339 DenseMap<Function *, LDSVariableReplacement> &KernelToReplacement) { 340 if (Variables.empty()) { 341 return nullptr; 342 } 343 LLVMContext &Ctx = M.getContext(); 344 345 const size_t NumberVariables = Variables.size(); 346 const size_t NumberKernels = kernels.size(); 347 348 ArrayType *KernelOffsetsType = 349 ArrayType::get(Type::getInt32Ty(Ctx), NumberVariables); 350 351 ArrayType *AllKernelsOffsetsType = 352 ArrayType::get(KernelOffsetsType, NumberKernels); 353 354 Constant *Missing = PoisonValue::get(KernelOffsetsType); 355 std::vector<Constant *> overallConstantExprElts(NumberKernels); 356 for (size_t i = 0; i < NumberKernels; i++) { 357 auto Replacement = KernelToReplacement.find(kernels[i]); 358 overallConstantExprElts[i] = 359 (Replacement == KernelToReplacement.end()) 360 ? Missing 361 : getAddressesOfVariablesInKernel( 362 Ctx, Variables, Replacement->second.LDSVarsToConstantGEP); 363 } 364 365 Constant *init = 366 ConstantArray::get(AllKernelsOffsetsType, overallConstantExprElts); 367 368 return new GlobalVariable( 369 M, AllKernelsOffsetsType, true, GlobalValue::InternalLinkage, init, 370 "llvm.amdgcn.lds.offset.table", nullptr, GlobalValue::NotThreadLocal, 371 AMDGPUAS::CONSTANT_ADDRESS); 372 } 373 374 void replaceUseWithTableLookup(Module &M, IRBuilder<> &Builder, 375 GlobalVariable *LookupTable, 376 GlobalVariable *GV, Use &U, 377 Value *OptionalIndex) { 378 // Table is a constant array of the same length as OrderedKernels 379 LLVMContext &Ctx = M.getContext(); 380 Type *I32 = Type::getInt32Ty(Ctx); 381 auto *I = cast<Instruction>(U.getUser()); 382 383 Value *tableKernelIndex = getTableLookupKernelIndex(M, I->getFunction()); 384 385 if (auto *Phi = dyn_cast<PHINode>(I)) { 386 BasicBlock *BB = Phi->getIncomingBlock(U); 387 Builder.SetInsertPoint(&(*(BB->getFirstInsertionPt()))); 388 } else { 389 Builder.SetInsertPoint(I); 390 } 391 392 SmallVector<Value *, 3> GEPIdx = { 393 ConstantInt::get(I32, 0), 394 tableKernelIndex, 395 }; 396 if (OptionalIndex) 397 GEPIdx.push_back(OptionalIndex); 398 399 Value *Address = Builder.CreateInBoundsGEP( 400 LookupTable->getValueType(), LookupTable, GEPIdx, GV->getName()); 401 402 Value *loaded = Builder.CreateLoad(I32, Address); 403 404 Value *replacement = 405 Builder.CreateIntToPtr(loaded, GV->getType(), GV->getName()); 406 407 U.set(replacement); 408 } 409 410 void replaceUsesInInstructionsWithTableLookup( 411 Module &M, ArrayRef<GlobalVariable *> ModuleScopeVariables, 412 GlobalVariable *LookupTable) { 413 414 LLVMContext &Ctx = M.getContext(); 415 IRBuilder<> Builder(Ctx); 416 Type *I32 = Type::getInt32Ty(Ctx); 417 418 for (size_t Index = 0; Index < ModuleScopeVariables.size(); Index++) { 419 auto *GV = ModuleScopeVariables[Index]; 420 421 for (Use &U : make_early_inc_range(GV->uses())) { 422 auto *I = dyn_cast<Instruction>(U.getUser()); 423 if (!I) 424 continue; 425 426 replaceUseWithTableLookup(M, Builder, LookupTable, GV, U, 427 ConstantInt::get(I32, Index)); 428 } 429 } 430 } 431 432 static DenseSet<Function *> kernelsThatIndirectlyAccessAnyOfPassedVariables( 433 Module &M, LDSUsesInfoTy &LDSUsesInfo, 434 DenseSet<GlobalVariable *> const &VariableSet) { 435 436 DenseSet<Function *> KernelSet; 437 438 if (VariableSet.empty()) 439 return KernelSet; 440 441 for (Function &Func : M.functions()) { 442 if (Func.isDeclaration() || !isKernelLDS(&Func)) 443 continue; 444 for (GlobalVariable *GV : LDSUsesInfo.indirect_access[&Func]) { 445 if (VariableSet.contains(GV)) { 446 KernelSet.insert(&Func); 447 break; 448 } 449 } 450 } 451 452 return KernelSet; 453 } 454 455 static GlobalVariable * 456 chooseBestVariableForModuleStrategy(const DataLayout &DL, 457 VariableFunctionMap &LDSVars) { 458 // Find the global variable with the most indirect uses from kernels 459 460 struct CandidateTy { 461 GlobalVariable *GV = nullptr; 462 size_t UserCount = 0; 463 size_t Size = 0; 464 465 CandidateTy() = default; 466 467 CandidateTy(GlobalVariable *GV, uint64_t UserCount, uint64_t AllocSize) 468 : GV(GV), UserCount(UserCount), Size(AllocSize) {} 469 470 bool operator<(const CandidateTy &Other) const { 471 // Fewer users makes module scope variable less attractive 472 if (UserCount < Other.UserCount) { 473 return true; 474 } 475 if (UserCount > Other.UserCount) { 476 return false; 477 } 478 479 // Bigger makes module scope variable less attractive 480 if (Size < Other.Size) { 481 return false; 482 } 483 484 if (Size > Other.Size) { 485 return true; 486 } 487 488 // Arbitrary but consistent 489 return GV->getName() < Other.GV->getName(); 490 } 491 }; 492 493 CandidateTy MostUsed; 494 495 for (auto &K : LDSVars) { 496 GlobalVariable *GV = K.first; 497 if (K.second.size() <= 1) { 498 // A variable reachable by only one kernel is best lowered with kernel 499 // strategy 500 continue; 501 } 502 CandidateTy Candidate( 503 GV, K.second.size(), 504 DL.getTypeAllocSize(GV->getValueType()).getFixedValue()); 505 if (MostUsed < Candidate) 506 MostUsed = Candidate; 507 } 508 509 return MostUsed.GV; 510 } 511 512 static void recordLDSAbsoluteAddress(Module *M, GlobalVariable *GV, 513 uint32_t Address) { 514 // Write the specified address into metadata where it can be retrieved by 515 // the assembler. Format is a half open range, [Address Address+1) 516 LLVMContext &Ctx = M->getContext(); 517 auto *IntTy = 518 M->getDataLayout().getIntPtrType(Ctx, AMDGPUAS::LOCAL_ADDRESS); 519 auto *MinC = ConstantAsMetadata::get(ConstantInt::get(IntTy, Address)); 520 auto *MaxC = ConstantAsMetadata::get(ConstantInt::get(IntTy, Address + 1)); 521 GV->setMetadata(LLVMContext::MD_absolute_symbol, 522 MDNode::get(Ctx, {MinC, MaxC})); 523 } 524 525 DenseMap<Function *, Value *> tableKernelIndexCache; 526 Value *getTableLookupKernelIndex(Module &M, Function *F) { 527 // Accesses from a function use the amdgcn_lds_kernel_id intrinsic which 528 // lowers to a read from a live in register. Emit it once in the entry 529 // block to spare deduplicating it later. 530 auto [It, Inserted] = tableKernelIndexCache.try_emplace(F); 531 if (Inserted) { 532 Function *Decl = 533 Intrinsic::getDeclaration(&M, Intrinsic::amdgcn_lds_kernel_id, {}); 534 535 auto InsertAt = F->getEntryBlock().getFirstNonPHIOrDbgOrAlloca(); 536 IRBuilder<> Builder(&*InsertAt); 537 538 It->second = Builder.CreateCall(Decl, {}); 539 } 540 541 return It->second; 542 } 543 544 static std::vector<Function *> assignLDSKernelIDToEachKernel( 545 Module *M, DenseSet<Function *> const &KernelsThatAllocateTableLDS, 546 DenseSet<Function *> const &KernelsThatIndirectlyAllocateDynamicLDS) { 547 // Associate kernels in the set with an arbitrary but reproducible order and 548 // annotate them with that order in metadata. This metadata is recognised by 549 // the backend and lowered to a SGPR which can be read from using 550 // amdgcn_lds_kernel_id. 551 552 std::vector<Function *> OrderedKernels; 553 if (!KernelsThatAllocateTableLDS.empty() || 554 !KernelsThatIndirectlyAllocateDynamicLDS.empty()) { 555 556 for (Function &Func : M->functions()) { 557 if (Func.isDeclaration()) 558 continue; 559 if (!isKernelLDS(&Func)) 560 continue; 561 562 if (KernelsThatAllocateTableLDS.contains(&Func) || 563 KernelsThatIndirectlyAllocateDynamicLDS.contains(&Func)) { 564 assert(Func.hasName()); // else fatal error earlier 565 OrderedKernels.push_back(&Func); 566 } 567 } 568 569 // Put them in an arbitrary but reproducible order 570 OrderedKernels = sortByName(std::move(OrderedKernels)); 571 572 // Annotate the kernels with their order in this vector 573 LLVMContext &Ctx = M->getContext(); 574 IRBuilder<> Builder(Ctx); 575 576 if (OrderedKernels.size() > UINT32_MAX) { 577 // 32 bit keeps it in one SGPR. > 2**32 kernels won't fit on the GPU 578 report_fatal_error("Unimplemented LDS lowering for > 2**32 kernels"); 579 } 580 581 for (size_t i = 0; i < OrderedKernels.size(); i++) { 582 Metadata *AttrMDArgs[1] = { 583 ConstantAsMetadata::get(Builder.getInt32(i)), 584 }; 585 OrderedKernels[i]->setMetadata("llvm.amdgcn.lds.kernel.id", 586 MDNode::get(Ctx, AttrMDArgs)); 587 } 588 } 589 return OrderedKernels; 590 } 591 592 static void partitionVariablesIntoIndirectStrategies( 593 Module &M, LDSUsesInfoTy const &LDSUsesInfo, 594 VariableFunctionMap &LDSToKernelsThatNeedToAccessItIndirectly, 595 DenseSet<GlobalVariable *> &ModuleScopeVariables, 596 DenseSet<GlobalVariable *> &TableLookupVariables, 597 DenseSet<GlobalVariable *> &KernelAccessVariables, 598 DenseSet<GlobalVariable *> &DynamicVariables) { 599 600 GlobalVariable *HybridModuleRoot = 601 LoweringKindLoc != LoweringKind::hybrid 602 ? nullptr 603 : chooseBestVariableForModuleStrategy( 604 M.getDataLayout(), LDSToKernelsThatNeedToAccessItIndirectly); 605 606 DenseSet<Function *> const EmptySet; 607 DenseSet<Function *> const &HybridModuleRootKernels = 608 HybridModuleRoot 609 ? LDSToKernelsThatNeedToAccessItIndirectly[HybridModuleRoot] 610 : EmptySet; 611 612 for (auto &K : LDSToKernelsThatNeedToAccessItIndirectly) { 613 // Each iteration of this loop assigns exactly one global variable to 614 // exactly one of the implementation strategies. 615 616 GlobalVariable *GV = K.first; 617 assert(AMDGPU::isLDSVariableToLower(*GV)); 618 assert(K.second.size() != 0); 619 620 if (AMDGPU::isDynamicLDS(*GV)) { 621 DynamicVariables.insert(GV); 622 continue; 623 } 624 625 switch (LoweringKindLoc) { 626 case LoweringKind::module: 627 ModuleScopeVariables.insert(GV); 628 break; 629 630 case LoweringKind::table: 631 TableLookupVariables.insert(GV); 632 break; 633 634 case LoweringKind::kernel: 635 if (K.second.size() == 1) { 636 KernelAccessVariables.insert(GV); 637 } else { 638 report_fatal_error( 639 "cannot lower LDS '" + GV->getName() + 640 "' to kernel access as it is reachable from multiple kernels"); 641 } 642 break; 643 644 case LoweringKind::hybrid: { 645 if (GV == HybridModuleRoot) { 646 assert(K.second.size() != 1); 647 ModuleScopeVariables.insert(GV); 648 } else if (K.second.size() == 1) { 649 KernelAccessVariables.insert(GV); 650 } else if (set_is_subset(K.second, HybridModuleRootKernels)) { 651 ModuleScopeVariables.insert(GV); 652 } else { 653 TableLookupVariables.insert(GV); 654 } 655 break; 656 } 657 } 658 } 659 660 // All LDS variables accessed indirectly have now been partitioned into 661 // the distinct lowering strategies. 662 assert(ModuleScopeVariables.size() + TableLookupVariables.size() + 663 KernelAccessVariables.size() + DynamicVariables.size() == 664 LDSToKernelsThatNeedToAccessItIndirectly.size()); 665 } 666 667 static GlobalVariable *lowerModuleScopeStructVariables( 668 Module &M, DenseSet<GlobalVariable *> const &ModuleScopeVariables, 669 DenseSet<Function *> const &KernelsThatAllocateModuleLDS) { 670 // Create a struct to hold the ModuleScopeVariables 671 // Replace all uses of those variables from non-kernel functions with the 672 // new struct instance Replace only the uses from kernel functions that will 673 // allocate this instance. That is a space optimisation - kernels that use a 674 // subset of the module scope struct and do not need to allocate it for 675 // indirect calls will only allocate the subset they use (they do so as part 676 // of the per-kernel lowering). 677 if (ModuleScopeVariables.empty()) { 678 return nullptr; 679 } 680 681 LLVMContext &Ctx = M.getContext(); 682 683 LDSVariableReplacement ModuleScopeReplacement = 684 createLDSVariableReplacement(M, "llvm.amdgcn.module.lds", 685 ModuleScopeVariables); 686 687 appendToCompilerUsed(M, {static_cast<GlobalValue *>( 688 ConstantExpr::getPointerBitCastOrAddrSpaceCast( 689 cast<Constant>(ModuleScopeReplacement.SGV), 690 PointerType::getUnqual(Ctx)))}); 691 692 // module.lds will be allocated at zero in any kernel that allocates it 693 recordLDSAbsoluteAddress(&M, ModuleScopeReplacement.SGV, 0); 694 695 // historic 696 removeLocalVarsFromUsedLists(M, ModuleScopeVariables); 697 698 // Replace all uses of module scope variable from non-kernel functions 699 replaceLDSVariablesWithStruct( 700 M, ModuleScopeVariables, ModuleScopeReplacement, [&](Use &U) { 701 Instruction *I = dyn_cast<Instruction>(U.getUser()); 702 if (!I) { 703 return false; 704 } 705 Function *F = I->getFunction(); 706 return !isKernelLDS(F); 707 }); 708 709 // Replace uses of module scope variable from kernel functions that 710 // allocate the module scope variable, otherwise leave them unchanged 711 // Record on each kernel whether the module scope global is used by it 712 713 for (Function &Func : M.functions()) { 714 if (Func.isDeclaration() || !isKernelLDS(&Func)) 715 continue; 716 717 if (KernelsThatAllocateModuleLDS.contains(&Func)) { 718 replaceLDSVariablesWithStruct( 719 M, ModuleScopeVariables, ModuleScopeReplacement, [&](Use &U) { 720 Instruction *I = dyn_cast<Instruction>(U.getUser()); 721 if (!I) { 722 return false; 723 } 724 Function *F = I->getFunction(); 725 return F == &Func; 726 }); 727 728 markUsedByKernel(&Func, ModuleScopeReplacement.SGV); 729 } 730 } 731 732 return ModuleScopeReplacement.SGV; 733 } 734 735 static DenseMap<Function *, LDSVariableReplacement> 736 lowerKernelScopeStructVariables( 737 Module &M, LDSUsesInfoTy &LDSUsesInfo, 738 DenseSet<GlobalVariable *> const &ModuleScopeVariables, 739 DenseSet<Function *> const &KernelsThatAllocateModuleLDS, 740 GlobalVariable *MaybeModuleScopeStruct) { 741 742 // Create a struct for each kernel for the non-module-scope variables. 743 744 DenseMap<Function *, LDSVariableReplacement> KernelToReplacement; 745 for (Function &Func : M.functions()) { 746 if (Func.isDeclaration() || !isKernelLDS(&Func)) 747 continue; 748 749 DenseSet<GlobalVariable *> KernelUsedVariables; 750 // Allocating variables that are used directly in this struct to get 751 // alignment aware allocation and predictable frame size. 752 for (auto &v : LDSUsesInfo.direct_access[&Func]) { 753 if (!AMDGPU::isDynamicLDS(*v)) { 754 KernelUsedVariables.insert(v); 755 } 756 } 757 758 // Allocating variables that are accessed indirectly so that a lookup of 759 // this struct instance can find them from nested functions. 760 for (auto &v : LDSUsesInfo.indirect_access[&Func]) { 761 if (!AMDGPU::isDynamicLDS(*v)) { 762 KernelUsedVariables.insert(v); 763 } 764 } 765 766 // Variables allocated in module lds must all resolve to that struct, 767 // not to the per-kernel instance. 768 if (KernelsThatAllocateModuleLDS.contains(&Func)) { 769 for (GlobalVariable *v : ModuleScopeVariables) { 770 KernelUsedVariables.erase(v); 771 } 772 } 773 774 if (KernelUsedVariables.empty()) { 775 // Either used no LDS, or the LDS it used was all in the module struct 776 // or dynamically sized 777 continue; 778 } 779 780 // The association between kernel function and LDS struct is done by 781 // symbol name, which only works if the function in question has a 782 // name This is not expected to be a problem in practice as kernels 783 // are called by name making anonymous ones (which are named by the 784 // backend) difficult to use. This does mean that llvm test cases need 785 // to name the kernels. 786 if (!Func.hasName()) { 787 report_fatal_error("Anonymous kernels cannot use LDS variables"); 788 } 789 790 std::string VarName = 791 (Twine("llvm.amdgcn.kernel.") + Func.getName() + ".lds").str(); 792 793 auto Replacement = 794 createLDSVariableReplacement(M, VarName, KernelUsedVariables); 795 796 // If any indirect uses, create a direct use to ensure allocation 797 // TODO: Simpler to unconditionally mark used but that regresses 798 // codegen in test/CodeGen/AMDGPU/noclobber-barrier.ll 799 auto Accesses = LDSUsesInfo.indirect_access.find(&Func); 800 if ((Accesses != LDSUsesInfo.indirect_access.end()) && 801 !Accesses->second.empty()) 802 markUsedByKernel(&Func, Replacement.SGV); 803 804 // remove preserves existing codegen 805 removeLocalVarsFromUsedLists(M, KernelUsedVariables); 806 KernelToReplacement[&Func] = Replacement; 807 808 // Rewrite uses within kernel to the new struct 809 replaceLDSVariablesWithStruct( 810 M, KernelUsedVariables, Replacement, [&Func](Use &U) { 811 Instruction *I = dyn_cast<Instruction>(U.getUser()); 812 return I && I->getFunction() == &Func; 813 }); 814 } 815 return KernelToReplacement; 816 } 817 818 static GlobalVariable * 819 buildRepresentativeDynamicLDSInstance(Module &M, LDSUsesInfoTy &LDSUsesInfo, 820 Function *func) { 821 // Create a dynamic lds variable with a name associated with the passed 822 // function that has the maximum alignment of any dynamic lds variable 823 // reachable from this kernel. Dynamic LDS is allocated after the static LDS 824 // allocation, possibly after alignment padding. The representative variable 825 // created here has the maximum alignment of any other dynamic variable 826 // reachable by that kernel. All dynamic LDS variables are allocated at the 827 // same address in each kernel in order to provide the documented aliasing 828 // semantics. Setting the alignment here allows this IR pass to accurately 829 // predict the exact constant at which it will be allocated. 830 831 assert(isKernelLDS(func)); 832 833 LLVMContext &Ctx = M.getContext(); 834 const DataLayout &DL = M.getDataLayout(); 835 Align MaxDynamicAlignment(1); 836 837 auto UpdateMaxAlignment = [&MaxDynamicAlignment, &DL](GlobalVariable *GV) { 838 if (AMDGPU::isDynamicLDS(*GV)) { 839 MaxDynamicAlignment = 840 std::max(MaxDynamicAlignment, AMDGPU::getAlign(DL, GV)); 841 } 842 }; 843 844 for (GlobalVariable *GV : LDSUsesInfo.indirect_access[func]) { 845 UpdateMaxAlignment(GV); 846 } 847 848 for (GlobalVariable *GV : LDSUsesInfo.direct_access[func]) { 849 UpdateMaxAlignment(GV); 850 } 851 852 assert(func->hasName()); // Checked by caller 853 auto emptyCharArray = ArrayType::get(Type::getInt8Ty(Ctx), 0); 854 GlobalVariable *N = new GlobalVariable( 855 M, emptyCharArray, false, GlobalValue::ExternalLinkage, nullptr, 856 Twine("llvm.amdgcn." + func->getName() + ".dynlds"), nullptr, GlobalValue::NotThreadLocal, AMDGPUAS::LOCAL_ADDRESS, 857 false); 858 N->setAlignment(MaxDynamicAlignment); 859 860 assert(AMDGPU::isDynamicLDS(*N)); 861 return N; 862 } 863 864 DenseMap<Function *, GlobalVariable *> lowerDynamicLDSVariables( 865 Module &M, LDSUsesInfoTy &LDSUsesInfo, 866 DenseSet<Function *> const &KernelsThatIndirectlyAllocateDynamicLDS, 867 DenseSet<GlobalVariable *> const &DynamicVariables, 868 std::vector<Function *> const &OrderedKernels) { 869 DenseMap<Function *, GlobalVariable *> KernelToCreatedDynamicLDS; 870 if (!KernelsThatIndirectlyAllocateDynamicLDS.empty()) { 871 LLVMContext &Ctx = M.getContext(); 872 IRBuilder<> Builder(Ctx); 873 Type *I32 = Type::getInt32Ty(Ctx); 874 875 std::vector<Constant *> newDynamicLDS; 876 877 // Table is built in the same order as OrderedKernels 878 for (auto &func : OrderedKernels) { 879 880 if (KernelsThatIndirectlyAllocateDynamicLDS.contains(func)) { 881 assert(isKernelLDS(func)); 882 if (!func->hasName()) { 883 report_fatal_error("Anonymous kernels cannot use LDS variables"); 884 } 885 886 GlobalVariable *N = 887 buildRepresentativeDynamicLDSInstance(M, LDSUsesInfo, func); 888 889 KernelToCreatedDynamicLDS[func] = N; 890 891 markUsedByKernel(func, N); 892 893 auto emptyCharArray = ArrayType::get(Type::getInt8Ty(Ctx), 0); 894 auto GEP = ConstantExpr::getGetElementPtr( 895 emptyCharArray, N, ConstantInt::get(I32, 0), true); 896 newDynamicLDS.push_back(ConstantExpr::getPtrToInt(GEP, I32)); 897 } else { 898 newDynamicLDS.push_back(PoisonValue::get(I32)); 899 } 900 } 901 assert(OrderedKernels.size() == newDynamicLDS.size()); 902 903 ArrayType *t = ArrayType::get(I32, newDynamicLDS.size()); 904 Constant *init = ConstantArray::get(t, newDynamicLDS); 905 GlobalVariable *table = new GlobalVariable( 906 M, t, true, GlobalValue::InternalLinkage, init, 907 "llvm.amdgcn.dynlds.offset.table", nullptr, 908 GlobalValue::NotThreadLocal, AMDGPUAS::CONSTANT_ADDRESS); 909 910 for (GlobalVariable *GV : DynamicVariables) { 911 for (Use &U : make_early_inc_range(GV->uses())) { 912 auto *I = dyn_cast<Instruction>(U.getUser()); 913 if (!I) 914 continue; 915 if (isKernelLDS(I->getFunction())) 916 continue; 917 918 replaceUseWithTableLookup(M, Builder, table, GV, U, nullptr); 919 } 920 } 921 } 922 return KernelToCreatedDynamicLDS; 923 } 924 925 bool runOnModule(Module &M) { 926 CallGraph CG = CallGraph(M); 927 bool Changed = superAlignLDSGlobals(M); 928 929 Changed |= eliminateConstantExprUsesOfLDSFromAllInstructions(M); 930 931 Changed = true; // todo: narrow this down 932 933 // For each kernel, what variables does it access directly or through 934 // callees 935 LDSUsesInfoTy LDSUsesInfo = getTransitiveUsesOfLDS(CG, M); 936 937 // For each variable accessed through callees, which kernels access it 938 VariableFunctionMap LDSToKernelsThatNeedToAccessItIndirectly; 939 for (auto &K : LDSUsesInfo.indirect_access) { 940 Function *F = K.first; 941 assert(isKernelLDS(F)); 942 for (GlobalVariable *GV : K.second) { 943 LDSToKernelsThatNeedToAccessItIndirectly[GV].insert(F); 944 } 945 } 946 947 // Partition variables accessed indirectly into the different strategies 948 DenseSet<GlobalVariable *> ModuleScopeVariables; 949 DenseSet<GlobalVariable *> TableLookupVariables; 950 DenseSet<GlobalVariable *> KernelAccessVariables; 951 DenseSet<GlobalVariable *> DynamicVariables; 952 partitionVariablesIntoIndirectStrategies( 953 M, LDSUsesInfo, LDSToKernelsThatNeedToAccessItIndirectly, 954 ModuleScopeVariables, TableLookupVariables, KernelAccessVariables, 955 DynamicVariables); 956 957 // If the kernel accesses a variable that is going to be stored in the 958 // module instance through a call then that kernel needs to allocate the 959 // module instance 960 const DenseSet<Function *> KernelsThatAllocateModuleLDS = 961 kernelsThatIndirectlyAccessAnyOfPassedVariables(M, LDSUsesInfo, 962 ModuleScopeVariables); 963 const DenseSet<Function *> KernelsThatAllocateTableLDS = 964 kernelsThatIndirectlyAccessAnyOfPassedVariables(M, LDSUsesInfo, 965 TableLookupVariables); 966 967 const DenseSet<Function *> KernelsThatIndirectlyAllocateDynamicLDS = 968 kernelsThatIndirectlyAccessAnyOfPassedVariables(M, LDSUsesInfo, 969 DynamicVariables); 970 971 GlobalVariable *MaybeModuleScopeStruct = lowerModuleScopeStructVariables( 972 M, ModuleScopeVariables, KernelsThatAllocateModuleLDS); 973 974 DenseMap<Function *, LDSVariableReplacement> KernelToReplacement = 975 lowerKernelScopeStructVariables(M, LDSUsesInfo, ModuleScopeVariables, 976 KernelsThatAllocateModuleLDS, 977 MaybeModuleScopeStruct); 978 979 // Lower zero cost accesses to the kernel instances just created 980 for (auto &GV : KernelAccessVariables) { 981 auto &funcs = LDSToKernelsThatNeedToAccessItIndirectly[GV]; 982 assert(funcs.size() == 1); // Only one kernel can access it 983 LDSVariableReplacement Replacement = 984 KernelToReplacement[*(funcs.begin())]; 985 986 DenseSet<GlobalVariable *> Vec; 987 Vec.insert(GV); 988 989 replaceLDSVariablesWithStruct(M, Vec, Replacement, [](Use &U) { 990 return isa<Instruction>(U.getUser()); 991 }); 992 } 993 994 // The ith element of this vector is kernel id i 995 std::vector<Function *> OrderedKernels = 996 assignLDSKernelIDToEachKernel(&M, KernelsThatAllocateTableLDS, 997 KernelsThatIndirectlyAllocateDynamicLDS); 998 999 if (!KernelsThatAllocateTableLDS.empty()) { 1000 LLVMContext &Ctx = M.getContext(); 1001 IRBuilder<> Builder(Ctx); 1002 1003 // The order must be consistent between lookup table and accesses to 1004 // lookup table 1005 auto TableLookupVariablesOrdered = 1006 sortByName(std::vector<GlobalVariable *>(TableLookupVariables.begin(), 1007 TableLookupVariables.end())); 1008 1009 GlobalVariable *LookupTable = buildLookupTable( 1010 M, TableLookupVariablesOrdered, OrderedKernels, KernelToReplacement); 1011 replaceUsesInInstructionsWithTableLookup(M, TableLookupVariablesOrdered, 1012 LookupTable); 1013 } 1014 1015 DenseMap<Function *, GlobalVariable *> KernelToCreatedDynamicLDS = 1016 lowerDynamicLDSVariables(M, LDSUsesInfo, 1017 KernelsThatIndirectlyAllocateDynamicLDS, 1018 DynamicVariables, OrderedKernels); 1019 1020 // Strip amdgpu-no-lds-kernel-id from all functions reachable from the 1021 // kernel. We may have inferred this wasn't used prior to the pass. 1022 // TODO: We could filter out subgraphs that do not access LDS globals. 1023 for (auto *KernelSet : {&KernelsThatIndirectlyAllocateDynamicLDS, 1024 &KernelsThatAllocateTableLDS}) 1025 for (Function *F : *KernelSet) 1026 removeFnAttrFromReachable(CG, F, {"amdgpu-no-lds-kernel-id"}); 1027 1028 // All kernel frames have been allocated. Calculate and record the 1029 // addresses. 1030 { 1031 const DataLayout &DL = M.getDataLayout(); 1032 1033 for (Function &Func : M.functions()) { 1034 if (Func.isDeclaration() || !isKernelLDS(&Func)) 1035 continue; 1036 1037 // All three of these are optional. The first variable is allocated at 1038 // zero. They are allocated by AMDGPUMachineFunction as one block. 1039 // Layout: 1040 //{ 1041 // module.lds 1042 // alignment padding 1043 // kernel instance 1044 // alignment padding 1045 // dynamic lds variables 1046 //} 1047 1048 const bool AllocateModuleScopeStruct = 1049 MaybeModuleScopeStruct && 1050 KernelsThatAllocateModuleLDS.contains(&Func); 1051 1052 auto Replacement = KernelToReplacement.find(&Func); 1053 const bool AllocateKernelScopeStruct = 1054 Replacement != KernelToReplacement.end(); 1055 1056 const bool AllocateDynamicVariable = 1057 KernelToCreatedDynamicLDS.contains(&Func); 1058 1059 uint32_t Offset = 0; 1060 1061 if (AllocateModuleScopeStruct) { 1062 // Allocated at zero, recorded once on construction, not once per 1063 // kernel 1064 Offset += DL.getTypeAllocSize(MaybeModuleScopeStruct->getValueType()); 1065 } 1066 1067 if (AllocateKernelScopeStruct) { 1068 GlobalVariable *KernelStruct = Replacement->second.SGV; 1069 Offset = alignTo(Offset, AMDGPU::getAlign(DL, KernelStruct)); 1070 recordLDSAbsoluteAddress(&M, KernelStruct, Offset); 1071 Offset += DL.getTypeAllocSize(KernelStruct->getValueType()); 1072 } 1073 1074 // If there is dynamic allocation, the alignment needed is included in 1075 // the static frame size. There may be no reference to the dynamic 1076 // variable in the kernel itself, so without including it here, that 1077 // alignment padding could be missed. 1078 if (AllocateDynamicVariable) { 1079 GlobalVariable *DynamicVariable = KernelToCreatedDynamicLDS[&Func]; 1080 Offset = alignTo(Offset, AMDGPU::getAlign(DL, DynamicVariable)); 1081 recordLDSAbsoluteAddress(&M, DynamicVariable, Offset); 1082 } 1083 1084 if (Offset != 0) { 1085 (void)TM; // TODO: Account for target maximum LDS 1086 std::string Buffer; 1087 raw_string_ostream SS{Buffer}; 1088 SS << format("%u", Offset); 1089 1090 // Instead of explicitly marking kernels that access dynamic variables 1091 // using special case metadata, annotate with min-lds == max-lds, i.e. 1092 // that there is no more space available for allocating more static 1093 // LDS variables. That is the right condition to prevent allocating 1094 // more variables which would collide with the addresses assigned to 1095 // dynamic variables. 1096 if (AllocateDynamicVariable) 1097 SS << format(",%u", Offset); 1098 1099 Func.addFnAttr("amdgpu-lds-size", Buffer); 1100 } 1101 } 1102 } 1103 1104 for (auto &GV : make_early_inc_range(M.globals())) 1105 if (AMDGPU::isLDSVariableToLower(GV)) { 1106 // probably want to remove from used lists 1107 GV.removeDeadConstantUsers(); 1108 if (GV.use_empty()) 1109 GV.eraseFromParent(); 1110 } 1111 1112 return Changed; 1113 } 1114 1115 private: 1116 // Increase the alignment of LDS globals if necessary to maximise the chance 1117 // that we can use aligned LDS instructions to access them. 1118 static bool superAlignLDSGlobals(Module &M) { 1119 const DataLayout &DL = M.getDataLayout(); 1120 bool Changed = false; 1121 if (!SuperAlignLDSGlobals) { 1122 return Changed; 1123 } 1124 1125 for (auto &GV : M.globals()) { 1126 if (GV.getType()->getPointerAddressSpace() != AMDGPUAS::LOCAL_ADDRESS) { 1127 // Only changing alignment of LDS variables 1128 continue; 1129 } 1130 if (!GV.hasInitializer()) { 1131 // cuda/hip extern __shared__ variable, leave alignment alone 1132 continue; 1133 } 1134 1135 if (GV.isAbsoluteSymbolRef()) { 1136 // If the variable is already allocated, don't change the alignment 1137 continue; 1138 } 1139 1140 Align Alignment = AMDGPU::getAlign(DL, &GV); 1141 TypeSize GVSize = DL.getTypeAllocSize(GV.getValueType()); 1142 1143 if (GVSize > 8) { 1144 // We might want to use a b96 or b128 load/store 1145 Alignment = std::max(Alignment, Align(16)); 1146 } else if (GVSize > 4) { 1147 // We might want to use a b64 load/store 1148 Alignment = std::max(Alignment, Align(8)); 1149 } else if (GVSize > 2) { 1150 // We might want to use a b32 load/store 1151 Alignment = std::max(Alignment, Align(4)); 1152 } else if (GVSize > 1) { 1153 // We might want to use a b16 load/store 1154 Alignment = std::max(Alignment, Align(2)); 1155 } 1156 1157 if (Alignment != AMDGPU::getAlign(DL, &GV)) { 1158 Changed = true; 1159 GV.setAlignment(Alignment); 1160 } 1161 } 1162 return Changed; 1163 } 1164 1165 static LDSVariableReplacement createLDSVariableReplacement( 1166 Module &M, std::string VarName, 1167 DenseSet<GlobalVariable *> const &LDSVarsToTransform) { 1168 // Create a struct instance containing LDSVarsToTransform and map from those 1169 // variables to ConstantExprGEP 1170 // Variables may be introduced to meet alignment requirements. No aliasing 1171 // metadata is useful for these as they have no uses. Erased before return. 1172 1173 LLVMContext &Ctx = M.getContext(); 1174 const DataLayout &DL = M.getDataLayout(); 1175 assert(!LDSVarsToTransform.empty()); 1176 1177 SmallVector<OptimizedStructLayoutField, 8> LayoutFields; 1178 LayoutFields.reserve(LDSVarsToTransform.size()); 1179 { 1180 // The order of fields in this struct depends on the order of 1181 // variables in the argument which varies when changing how they 1182 // are identified, leading to spurious test breakage. 1183 auto Sorted = sortByName(std::vector<GlobalVariable *>( 1184 LDSVarsToTransform.begin(), LDSVarsToTransform.end())); 1185 1186 for (GlobalVariable *GV : Sorted) { 1187 OptimizedStructLayoutField F(GV, 1188 DL.getTypeAllocSize(GV->getValueType()), 1189 AMDGPU::getAlign(DL, GV)); 1190 LayoutFields.emplace_back(F); 1191 } 1192 } 1193 1194 performOptimizedStructLayout(LayoutFields); 1195 1196 std::vector<GlobalVariable *> LocalVars; 1197 BitVector IsPaddingField; 1198 LocalVars.reserve(LDSVarsToTransform.size()); // will be at least this large 1199 IsPaddingField.reserve(LDSVarsToTransform.size()); 1200 { 1201 uint64_t CurrentOffset = 0; 1202 for (auto &F : LayoutFields) { 1203 GlobalVariable *FGV = 1204 static_cast<GlobalVariable *>(const_cast<void *>(F.Id)); 1205 Align DataAlign = F.Alignment; 1206 1207 uint64_t DataAlignV = DataAlign.value(); 1208 if (uint64_t Rem = CurrentOffset % DataAlignV) { 1209 uint64_t Padding = DataAlignV - Rem; 1210 1211 // Append an array of padding bytes to meet alignment requested 1212 // Note (o + (a - (o % a)) ) % a == 0 1213 // (offset + Padding ) % align == 0 1214 1215 Type *ATy = ArrayType::get(Type::getInt8Ty(Ctx), Padding); 1216 LocalVars.push_back(new GlobalVariable( 1217 M, ATy, false, GlobalValue::InternalLinkage, 1218 PoisonValue::get(ATy), "", nullptr, GlobalValue::NotThreadLocal, 1219 AMDGPUAS::LOCAL_ADDRESS, false)); 1220 IsPaddingField.push_back(true); 1221 CurrentOffset += Padding; 1222 } 1223 1224 LocalVars.push_back(FGV); 1225 IsPaddingField.push_back(false); 1226 CurrentOffset += F.Size; 1227 } 1228 } 1229 1230 std::vector<Type *> LocalVarTypes; 1231 LocalVarTypes.reserve(LocalVars.size()); 1232 std::transform( 1233 LocalVars.cbegin(), LocalVars.cend(), std::back_inserter(LocalVarTypes), 1234 [](const GlobalVariable *V) -> Type * { return V->getValueType(); }); 1235 1236 StructType *LDSTy = StructType::create(Ctx, LocalVarTypes, VarName + ".t"); 1237 1238 Align StructAlign = AMDGPU::getAlign(DL, LocalVars[0]); 1239 1240 GlobalVariable *SGV = new GlobalVariable( 1241 M, LDSTy, false, GlobalValue::InternalLinkage, PoisonValue::get(LDSTy), 1242 VarName, nullptr, GlobalValue::NotThreadLocal, AMDGPUAS::LOCAL_ADDRESS, 1243 false); 1244 SGV->setAlignment(StructAlign); 1245 1246 DenseMap<GlobalVariable *, Constant *> Map; 1247 Type *I32 = Type::getInt32Ty(Ctx); 1248 for (size_t I = 0; I < LocalVars.size(); I++) { 1249 GlobalVariable *GV = LocalVars[I]; 1250 Constant *GEPIdx[] = {ConstantInt::get(I32, 0), ConstantInt::get(I32, I)}; 1251 Constant *GEP = ConstantExpr::getGetElementPtr(LDSTy, SGV, GEPIdx, true); 1252 if (IsPaddingField[I]) { 1253 assert(GV->use_empty()); 1254 GV->eraseFromParent(); 1255 } else { 1256 Map[GV] = GEP; 1257 } 1258 } 1259 assert(Map.size() == LDSVarsToTransform.size()); 1260 return {SGV, std::move(Map)}; 1261 } 1262 1263 template <typename PredicateTy> 1264 static void replaceLDSVariablesWithStruct( 1265 Module &M, DenseSet<GlobalVariable *> const &LDSVarsToTransformArg, 1266 const LDSVariableReplacement &Replacement, PredicateTy Predicate) { 1267 LLVMContext &Ctx = M.getContext(); 1268 const DataLayout &DL = M.getDataLayout(); 1269 1270 // A hack... we need to insert the aliasing info in a predictable order for 1271 // lit tests. Would like to have them in a stable order already, ideally the 1272 // same order they get allocated, which might mean an ordered set container 1273 auto LDSVarsToTransform = sortByName(std::vector<GlobalVariable *>( 1274 LDSVarsToTransformArg.begin(), LDSVarsToTransformArg.end())); 1275 1276 // Create alias.scope and their lists. Each field in the new structure 1277 // does not alias with all other fields. 1278 SmallVector<MDNode *> AliasScopes; 1279 SmallVector<Metadata *> NoAliasList; 1280 const size_t NumberVars = LDSVarsToTransform.size(); 1281 if (NumberVars > 1) { 1282 MDBuilder MDB(Ctx); 1283 AliasScopes.reserve(NumberVars); 1284 MDNode *Domain = MDB.createAnonymousAliasScopeDomain(); 1285 for (size_t I = 0; I < NumberVars; I++) { 1286 MDNode *Scope = MDB.createAnonymousAliasScope(Domain); 1287 AliasScopes.push_back(Scope); 1288 } 1289 NoAliasList.append(&AliasScopes[1], AliasScopes.end()); 1290 } 1291 1292 // Replace uses of ith variable with a constantexpr to the corresponding 1293 // field of the instance that will be allocated by AMDGPUMachineFunction 1294 for (size_t I = 0; I < NumberVars; I++) { 1295 GlobalVariable *GV = LDSVarsToTransform[I]; 1296 Constant *GEP = Replacement.LDSVarsToConstantGEP.at(GV); 1297 1298 GV->replaceUsesWithIf(GEP, Predicate); 1299 1300 APInt APOff(DL.getIndexTypeSizeInBits(GEP->getType()), 0); 1301 GEP->stripAndAccumulateInBoundsConstantOffsets(DL, APOff); 1302 uint64_t Offset = APOff.getZExtValue(); 1303 1304 Align A = 1305 commonAlignment(Replacement.SGV->getAlign().valueOrOne(), Offset); 1306 1307 if (I) 1308 NoAliasList[I - 1] = AliasScopes[I - 1]; 1309 MDNode *NoAlias = 1310 NoAliasList.empty() ? nullptr : MDNode::get(Ctx, NoAliasList); 1311 MDNode *AliasScope = 1312 AliasScopes.empty() ? nullptr : MDNode::get(Ctx, {AliasScopes[I]}); 1313 1314 refineUsesAlignmentAndAA(GEP, A, DL, AliasScope, NoAlias); 1315 } 1316 } 1317 1318 static void refineUsesAlignmentAndAA(Value *Ptr, Align A, 1319 const DataLayout &DL, MDNode *AliasScope, 1320 MDNode *NoAlias, unsigned MaxDepth = 5) { 1321 if (!MaxDepth || (A == 1 && !AliasScope)) 1322 return; 1323 1324 for (User *U : Ptr->users()) { 1325 if (auto *I = dyn_cast<Instruction>(U)) { 1326 if (AliasScope && I->mayReadOrWriteMemory()) { 1327 MDNode *AS = I->getMetadata(LLVMContext::MD_alias_scope); 1328 AS = (AS ? MDNode::getMostGenericAliasScope(AS, AliasScope) 1329 : AliasScope); 1330 I->setMetadata(LLVMContext::MD_alias_scope, AS); 1331 1332 MDNode *NA = I->getMetadata(LLVMContext::MD_noalias); 1333 NA = (NA ? MDNode::intersect(NA, NoAlias) : NoAlias); 1334 I->setMetadata(LLVMContext::MD_noalias, NA); 1335 } 1336 } 1337 1338 if (auto *LI = dyn_cast<LoadInst>(U)) { 1339 LI->setAlignment(std::max(A, LI->getAlign())); 1340 continue; 1341 } 1342 if (auto *SI = dyn_cast<StoreInst>(U)) { 1343 if (SI->getPointerOperand() == Ptr) 1344 SI->setAlignment(std::max(A, SI->getAlign())); 1345 continue; 1346 } 1347 if (auto *AI = dyn_cast<AtomicRMWInst>(U)) { 1348 // None of atomicrmw operations can work on pointers, but let's 1349 // check it anyway in case it will or we will process ConstantExpr. 1350 if (AI->getPointerOperand() == Ptr) 1351 AI->setAlignment(std::max(A, AI->getAlign())); 1352 continue; 1353 } 1354 if (auto *AI = dyn_cast<AtomicCmpXchgInst>(U)) { 1355 if (AI->getPointerOperand() == Ptr) 1356 AI->setAlignment(std::max(A, AI->getAlign())); 1357 continue; 1358 } 1359 if (auto *GEP = dyn_cast<GetElementPtrInst>(U)) { 1360 unsigned BitWidth = DL.getIndexTypeSizeInBits(GEP->getType()); 1361 APInt Off(BitWidth, 0); 1362 if (GEP->getPointerOperand() == Ptr) { 1363 Align GA; 1364 if (GEP->accumulateConstantOffset(DL, Off)) 1365 GA = commonAlignment(A, Off.getLimitedValue()); 1366 refineUsesAlignmentAndAA(GEP, GA, DL, AliasScope, NoAlias, 1367 MaxDepth - 1); 1368 } 1369 continue; 1370 } 1371 if (auto *I = dyn_cast<Instruction>(U)) { 1372 if (I->getOpcode() == Instruction::BitCast || 1373 I->getOpcode() == Instruction::AddrSpaceCast) 1374 refineUsesAlignmentAndAA(I, A, DL, AliasScope, NoAlias, MaxDepth - 1); 1375 } 1376 } 1377 } 1378 }; 1379 1380 class AMDGPULowerModuleLDSLegacy : public ModulePass { 1381 public: 1382 const AMDGPUTargetMachine *TM; 1383 static char ID; 1384 1385 AMDGPULowerModuleLDSLegacy(const AMDGPUTargetMachine *TM_ = nullptr) 1386 : ModulePass(ID), TM(TM_) { 1387 initializeAMDGPULowerModuleLDSLegacyPass(*PassRegistry::getPassRegistry()); 1388 } 1389 1390 void getAnalysisUsage(AnalysisUsage &AU) const override { 1391 if (!TM) 1392 AU.addRequired<TargetPassConfig>(); 1393 } 1394 1395 bool runOnModule(Module &M) override { 1396 if (!TM) { 1397 auto &TPC = getAnalysis<TargetPassConfig>(); 1398 TM = &TPC.getTM<AMDGPUTargetMachine>(); 1399 } 1400 1401 return AMDGPULowerModuleLDS(*TM).runOnModule(M); 1402 } 1403 }; 1404 1405 } // namespace 1406 char AMDGPULowerModuleLDSLegacy::ID = 0; 1407 1408 char &llvm::AMDGPULowerModuleLDSLegacyPassID = AMDGPULowerModuleLDSLegacy::ID; 1409 1410 INITIALIZE_PASS_BEGIN(AMDGPULowerModuleLDSLegacy, DEBUG_TYPE, 1411 "Lower uses of LDS variables from non-kernel functions", 1412 false, false) 1413 INITIALIZE_PASS_DEPENDENCY(TargetPassConfig) 1414 INITIALIZE_PASS_END(AMDGPULowerModuleLDSLegacy, DEBUG_TYPE, 1415 "Lower uses of LDS variables from non-kernel functions", 1416 false, false) 1417 1418 ModulePass * 1419 llvm::createAMDGPULowerModuleLDSLegacyPass(const AMDGPUTargetMachine *TM) { 1420 return new AMDGPULowerModuleLDSLegacy(TM); 1421 } 1422 1423 PreservedAnalyses AMDGPULowerModuleLDSPass::run(Module &M, 1424 ModuleAnalysisManager &) { 1425 return AMDGPULowerModuleLDS(TM).runOnModule(M) ? PreservedAnalyses::none() 1426 : PreservedAnalyses::all(); 1427 } 1428