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 "AMDGPUTargetMachine.h" 181 #include "Utils/AMDGPUBaseInfo.h" 182 #include "Utils/AMDGPUMemoryUtils.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; 215 216 namespace { 217 218 cl::opt<bool> SuperAlignLDSGlobals( 219 "amdgpu-super-align-lds-globals", 220 cl::desc("Increase alignment of LDS if it is not on align boundary"), 221 cl::init(true), cl::Hidden); 222 223 enum class LoweringKind { module, table, kernel, hybrid }; 224 cl::opt<LoweringKind> LoweringKindLoc( 225 "amdgpu-lower-module-lds-strategy", 226 cl::desc("Specify lowering strategy for function LDS access:"), cl::Hidden, 227 cl::init(LoweringKind::hybrid), 228 cl::values( 229 clEnumValN(LoweringKind::table, "table", "Lower via table lookup"), 230 clEnumValN(LoweringKind::module, "module", "Lower via module struct"), 231 clEnumValN( 232 LoweringKind::kernel, "kernel", 233 "Lower variables reachable from one kernel, otherwise abort"), 234 clEnumValN(LoweringKind::hybrid, "hybrid", 235 "Lower via mixture of above strategies"))); 236 237 bool isKernelLDS(const Function *F) { 238 // Some weirdness here. AMDGPU::isKernelCC does not call into 239 // AMDGPU::isKernel with the calling conv, it instead calls into 240 // isModuleEntryFunction which returns true for more calling conventions 241 // than AMDGPU::isKernel does. There's a FIXME on AMDGPU::isKernel. 242 // There's also a test that checks that the LDS lowering does not hit on 243 // a graphics shader, denoted amdgpu_ps, so stay with the limited case. 244 // Putting LDS in the name of the function to draw attention to this. 245 return AMDGPU::isKernel(F->getCallingConv()); 246 } 247 248 template <typename T> std::vector<T> sortByName(std::vector<T> &&V) { 249 llvm::sort(V.begin(), V.end(), [](const auto *L, const auto *R) { 250 return L->getName() < R->getName(); 251 }); 252 return {std::move(V)}; 253 } 254 255 class AMDGPULowerModuleLDS { 256 const AMDGPUTargetMachine &TM; 257 258 static void 259 removeLocalVarsFromUsedLists(Module &M, 260 const DenseSet<GlobalVariable *> &LocalVars) { 261 // The verifier rejects used lists containing an inttoptr of a constant 262 // so remove the variables from these lists before replaceAllUsesWith 263 SmallPtrSet<Constant *, 8> LocalVarsSet; 264 for (GlobalVariable *LocalVar : LocalVars) 265 LocalVarsSet.insert(cast<Constant>(LocalVar->stripPointerCasts())); 266 267 removeFromUsedLists( 268 M, [&LocalVarsSet](Constant *C) { return LocalVarsSet.count(C); }); 269 270 for (GlobalVariable *LocalVar : LocalVars) 271 LocalVar->removeDeadConstantUsers(); 272 } 273 274 static void markUsedByKernel(Function *Func, GlobalVariable *SGV) { 275 // The llvm.amdgcn.module.lds instance is implicitly used by all kernels 276 // that might call a function which accesses a field within it. This is 277 // presently approximated to 'all kernels' if there are any such functions 278 // in the module. This implicit use is redefined as an explicit use here so 279 // that later passes, specifically PromoteAlloca, account for the required 280 // memory without any knowledge of this transform. 281 282 // An operand bundle on llvm.donothing works because the call instruction 283 // survives until after the last pass that needs to account for LDS. It is 284 // better than inline asm as the latter survives until the end of codegen. A 285 // totally robust solution would be a function with the same semantics as 286 // llvm.donothing that takes a pointer to the instance and is lowered to a 287 // no-op after LDS is allocated, but that is not presently necessary. 288 289 // This intrinsic is eliminated shortly before instruction selection. It 290 // does not suffice to indicate to ISel that a given global which is not 291 // immediately used by the kernel must still be allocated by it. An 292 // equivalent target specific intrinsic which lasts until immediately after 293 // codegen would suffice for that, but one would still need to ensure that 294 // the variables are allocated in the anticpated order. 295 BasicBlock *Entry = &Func->getEntryBlock(); 296 IRBuilder<> Builder(Entry, Entry->getFirstNonPHIIt()); 297 298 Function *Decl = 299 Intrinsic::getDeclaration(Func->getParent(), Intrinsic::donothing, {}); 300 301 Value *UseInstance[1] = { 302 Builder.CreateConstInBoundsGEP1_32(SGV->getValueType(), SGV, 0)}; 303 304 Builder.CreateCall( 305 Decl, {}, {OperandBundleDefT<Value *>("ExplicitUse", UseInstance)}); 306 } 307 308 static bool eliminateConstantExprUsesOfLDSFromAllInstructions(Module &M) { 309 // Constants are uniqued within LLVM. A ConstantExpr referring to a LDS 310 // global may have uses from multiple different functions as a result. 311 // This pass specialises LDS variables with respect to the kernel that 312 // allocates them. 313 314 // This is semantically equivalent to (the unimplemented as slow): 315 // for (auto &F : M.functions()) 316 // for (auto &BB : F) 317 // for (auto &I : BB) 318 // for (Use &Op : I.operands()) 319 // if (constantExprUsesLDS(Op)) 320 // replaceConstantExprInFunction(I, Op); 321 322 SmallVector<Constant *> LDSGlobals; 323 for (auto &GV : M.globals()) 324 if (AMDGPU::isLDSVariableToLower(GV)) 325 LDSGlobals.push_back(&GV); 326 327 return convertUsersOfConstantsToInstructions(LDSGlobals); 328 } 329 330 public: 331 AMDGPULowerModuleLDS(const AMDGPUTargetMachine &TM_) : TM(TM_) {} 332 333 using FunctionVariableMap = DenseMap<Function *, DenseSet<GlobalVariable *>>; 334 335 using VariableFunctionMap = DenseMap<GlobalVariable *, DenseSet<Function *>>; 336 337 static void getUsesOfLDSByFunction(CallGraph const &CG, Module &M, 338 FunctionVariableMap &kernels, 339 FunctionVariableMap &functions) { 340 341 // Get uses from the current function, excluding uses by called functions 342 // Two output variables to avoid walking the globals list twice 343 for (auto &GV : M.globals()) { 344 if (!AMDGPU::isLDSVariableToLower(GV)) { 345 continue; 346 } 347 348 if (GV.isAbsoluteSymbolRef()) { 349 report_fatal_error( 350 "LDS variables with absolute addresses are unimplemented."); 351 } 352 353 for (User *V : GV.users()) { 354 if (auto *I = dyn_cast<Instruction>(V)) { 355 Function *F = I->getFunction(); 356 if (isKernelLDS(F)) { 357 kernels[F].insert(&GV); 358 } else { 359 functions[F].insert(&GV); 360 } 361 } 362 } 363 } 364 } 365 366 struct LDSUsesInfoTy { 367 FunctionVariableMap direct_access; 368 FunctionVariableMap indirect_access; 369 }; 370 371 static LDSUsesInfoTy getTransitiveUsesOfLDS(CallGraph const &CG, Module &M) { 372 373 FunctionVariableMap direct_map_kernel; 374 FunctionVariableMap direct_map_function; 375 getUsesOfLDSByFunction(CG, M, direct_map_kernel, direct_map_function); 376 377 // Collect variables that are used by functions whose address has escaped 378 DenseSet<GlobalVariable *> VariablesReachableThroughFunctionPointer; 379 for (Function &F : M.functions()) { 380 if (!isKernelLDS(&F)) 381 if (F.hasAddressTaken(nullptr, 382 /* IgnoreCallbackUses */ false, 383 /* IgnoreAssumeLikeCalls */ false, 384 /* IgnoreLLVMUsed */ true, 385 /* IgnoreArcAttachedCall */ false)) { 386 set_union(VariablesReachableThroughFunctionPointer, 387 direct_map_function[&F]); 388 } 389 } 390 391 auto functionMakesUnknownCall = [&](const Function *F) -> bool { 392 assert(!F->isDeclaration()); 393 for (const CallGraphNode::CallRecord &R : *CG[F]) { 394 if (!R.second->getFunction()) { 395 return true; 396 } 397 } 398 return false; 399 }; 400 401 // Work out which variables are reachable through function calls 402 FunctionVariableMap transitive_map_function = direct_map_function; 403 404 // If the function makes any unknown call, assume the worst case that it can 405 // access all variables accessed by functions whose address escaped 406 for (Function &F : M.functions()) { 407 if (!F.isDeclaration() && functionMakesUnknownCall(&F)) { 408 if (!isKernelLDS(&F)) { 409 set_union(transitive_map_function[&F], 410 VariablesReachableThroughFunctionPointer); 411 } 412 } 413 } 414 415 // Direct implementation of collecting all variables reachable from each 416 // function 417 for (Function &Func : M.functions()) { 418 if (Func.isDeclaration() || isKernelLDS(&Func)) 419 continue; 420 421 DenseSet<Function *> seen; // catches cycles 422 SmallVector<Function *, 4> wip{&Func}; 423 424 while (!wip.empty()) { 425 Function *F = wip.pop_back_val(); 426 427 // Can accelerate this by referring to transitive map for functions that 428 // have already been computed, with more care than this 429 set_union(transitive_map_function[&Func], direct_map_function[F]); 430 431 for (const CallGraphNode::CallRecord &R : *CG[F]) { 432 Function *ith = R.second->getFunction(); 433 if (ith) { 434 if (!seen.contains(ith)) { 435 seen.insert(ith); 436 wip.push_back(ith); 437 } 438 } 439 } 440 } 441 } 442 443 // direct_map_kernel lists which variables are used by the kernel 444 // find the variables which are used through a function call 445 FunctionVariableMap indirect_map_kernel; 446 447 for (Function &Func : M.functions()) { 448 if (Func.isDeclaration() || !isKernelLDS(&Func)) 449 continue; 450 451 for (const CallGraphNode::CallRecord &R : *CG[&Func]) { 452 Function *ith = R.second->getFunction(); 453 if (ith) { 454 set_union(indirect_map_kernel[&Func], transitive_map_function[ith]); 455 } else { 456 set_union(indirect_map_kernel[&Func], 457 VariablesReachableThroughFunctionPointer); 458 } 459 } 460 } 461 462 return {std::move(direct_map_kernel), std::move(indirect_map_kernel)}; 463 } 464 465 struct LDSVariableReplacement { 466 GlobalVariable *SGV = nullptr; 467 DenseMap<GlobalVariable *, Constant *> LDSVarsToConstantGEP; 468 }; 469 470 // remap from lds global to a constantexpr gep to where it has been moved to 471 // for each kernel 472 // an array with an element for each kernel containing where the corresponding 473 // variable was remapped to 474 475 static Constant *getAddressesOfVariablesInKernel( 476 LLVMContext &Ctx, ArrayRef<GlobalVariable *> Variables, 477 const DenseMap<GlobalVariable *, Constant *> &LDSVarsToConstantGEP) { 478 // Create a ConstantArray containing the address of each Variable within the 479 // kernel corresponding to LDSVarsToConstantGEP, or poison if that kernel 480 // does not allocate it 481 // TODO: Drop the ptrtoint conversion 482 483 Type *I32 = Type::getInt32Ty(Ctx); 484 485 ArrayType *KernelOffsetsType = ArrayType::get(I32, Variables.size()); 486 487 SmallVector<Constant *> Elements; 488 for (size_t i = 0; i < Variables.size(); i++) { 489 GlobalVariable *GV = Variables[i]; 490 auto ConstantGepIt = LDSVarsToConstantGEP.find(GV); 491 if (ConstantGepIt != LDSVarsToConstantGEP.end()) { 492 auto elt = ConstantExpr::getPtrToInt(ConstantGepIt->second, I32); 493 Elements.push_back(elt); 494 } else { 495 Elements.push_back(PoisonValue::get(I32)); 496 } 497 } 498 return ConstantArray::get(KernelOffsetsType, Elements); 499 } 500 501 static GlobalVariable *buildLookupTable( 502 Module &M, ArrayRef<GlobalVariable *> Variables, 503 ArrayRef<Function *> kernels, 504 DenseMap<Function *, LDSVariableReplacement> &KernelToReplacement) { 505 if (Variables.empty()) { 506 return nullptr; 507 } 508 LLVMContext &Ctx = M.getContext(); 509 510 const size_t NumberVariables = Variables.size(); 511 const size_t NumberKernels = kernels.size(); 512 513 ArrayType *KernelOffsetsType = 514 ArrayType::get(Type::getInt32Ty(Ctx), NumberVariables); 515 516 ArrayType *AllKernelsOffsetsType = 517 ArrayType::get(KernelOffsetsType, NumberKernels); 518 519 Constant *Missing = PoisonValue::get(KernelOffsetsType); 520 std::vector<Constant *> overallConstantExprElts(NumberKernels); 521 for (size_t i = 0; i < NumberKernels; i++) { 522 auto Replacement = KernelToReplacement.find(kernels[i]); 523 overallConstantExprElts[i] = 524 (Replacement == KernelToReplacement.end()) 525 ? Missing 526 : getAddressesOfVariablesInKernel( 527 Ctx, Variables, Replacement->second.LDSVarsToConstantGEP); 528 } 529 530 Constant *init = 531 ConstantArray::get(AllKernelsOffsetsType, overallConstantExprElts); 532 533 return new GlobalVariable( 534 M, AllKernelsOffsetsType, true, GlobalValue::InternalLinkage, init, 535 "llvm.amdgcn.lds.offset.table", nullptr, GlobalValue::NotThreadLocal, 536 AMDGPUAS::CONSTANT_ADDRESS); 537 } 538 539 void replaceUseWithTableLookup(Module &M, IRBuilder<> &Builder, 540 GlobalVariable *LookupTable, 541 GlobalVariable *GV, Use &U, 542 Value *OptionalIndex) { 543 // Table is a constant array of the same length as OrderedKernels 544 LLVMContext &Ctx = M.getContext(); 545 Type *I32 = Type::getInt32Ty(Ctx); 546 auto *I = cast<Instruction>(U.getUser()); 547 548 Value *tableKernelIndex = getTableLookupKernelIndex(M, I->getFunction()); 549 550 if (auto *Phi = dyn_cast<PHINode>(I)) { 551 BasicBlock *BB = Phi->getIncomingBlock(U); 552 Builder.SetInsertPoint(&(*(BB->getFirstInsertionPt()))); 553 } else { 554 Builder.SetInsertPoint(I); 555 } 556 557 SmallVector<Value *, 3> GEPIdx = { 558 ConstantInt::get(I32, 0), 559 tableKernelIndex, 560 }; 561 if (OptionalIndex) 562 GEPIdx.push_back(OptionalIndex); 563 564 Value *Address = Builder.CreateInBoundsGEP( 565 LookupTable->getValueType(), LookupTable, GEPIdx, GV->getName()); 566 567 Value *loaded = Builder.CreateLoad(I32, Address); 568 569 Value *replacement = 570 Builder.CreateIntToPtr(loaded, GV->getType(), GV->getName()); 571 572 U.set(replacement); 573 } 574 575 void replaceUsesInInstructionsWithTableLookup( 576 Module &M, ArrayRef<GlobalVariable *> ModuleScopeVariables, 577 GlobalVariable *LookupTable) { 578 579 LLVMContext &Ctx = M.getContext(); 580 IRBuilder<> Builder(Ctx); 581 Type *I32 = Type::getInt32Ty(Ctx); 582 583 for (size_t Index = 0; Index < ModuleScopeVariables.size(); Index++) { 584 auto *GV = ModuleScopeVariables[Index]; 585 586 for (Use &U : make_early_inc_range(GV->uses())) { 587 auto *I = dyn_cast<Instruction>(U.getUser()); 588 if (!I) 589 continue; 590 591 replaceUseWithTableLookup(M, Builder, LookupTable, GV, U, 592 ConstantInt::get(I32, Index)); 593 } 594 } 595 } 596 597 static DenseSet<Function *> kernelsThatIndirectlyAccessAnyOfPassedVariables( 598 Module &M, LDSUsesInfoTy &LDSUsesInfo, 599 DenseSet<GlobalVariable *> const &VariableSet) { 600 601 DenseSet<Function *> KernelSet; 602 603 if (VariableSet.empty()) 604 return KernelSet; 605 606 for (Function &Func : M.functions()) { 607 if (Func.isDeclaration() || !isKernelLDS(&Func)) 608 continue; 609 for (GlobalVariable *GV : LDSUsesInfo.indirect_access[&Func]) { 610 if (VariableSet.contains(GV)) { 611 KernelSet.insert(&Func); 612 break; 613 } 614 } 615 } 616 617 return KernelSet; 618 } 619 620 static GlobalVariable * 621 chooseBestVariableForModuleStrategy(const DataLayout &DL, 622 VariableFunctionMap &LDSVars) { 623 // Find the global variable with the most indirect uses from kernels 624 625 struct CandidateTy { 626 GlobalVariable *GV = nullptr; 627 size_t UserCount = 0; 628 size_t Size = 0; 629 630 CandidateTy() = default; 631 632 CandidateTy(GlobalVariable *GV, uint64_t UserCount, uint64_t AllocSize) 633 : GV(GV), UserCount(UserCount), Size(AllocSize) {} 634 635 bool operator<(const CandidateTy &Other) const { 636 // Fewer users makes module scope variable less attractive 637 if (UserCount < Other.UserCount) { 638 return true; 639 } 640 if (UserCount > Other.UserCount) { 641 return false; 642 } 643 644 // Bigger makes module scope variable less attractive 645 if (Size < Other.Size) { 646 return false; 647 } 648 649 if (Size > Other.Size) { 650 return true; 651 } 652 653 // Arbitrary but consistent 654 return GV->getName() < Other.GV->getName(); 655 } 656 }; 657 658 CandidateTy MostUsed; 659 660 for (auto &K : LDSVars) { 661 GlobalVariable *GV = K.first; 662 if (K.second.size() <= 1) { 663 // A variable reachable by only one kernel is best lowered with kernel 664 // strategy 665 continue; 666 } 667 CandidateTy Candidate( 668 GV, K.second.size(), 669 DL.getTypeAllocSize(GV->getValueType()).getFixedValue()); 670 if (MostUsed < Candidate) 671 MostUsed = Candidate; 672 } 673 674 return MostUsed.GV; 675 } 676 677 static void recordLDSAbsoluteAddress(Module *M, GlobalVariable *GV, 678 uint32_t Address) { 679 // Write the specified address into metadata where it can be retrieved by 680 // the assembler. Format is a half open range, [Address Address+1) 681 LLVMContext &Ctx = M->getContext(); 682 auto *IntTy = 683 M->getDataLayout().getIntPtrType(Ctx, AMDGPUAS::LOCAL_ADDRESS); 684 auto *MinC = ConstantAsMetadata::get(ConstantInt::get(IntTy, Address)); 685 auto *MaxC = ConstantAsMetadata::get(ConstantInt::get(IntTy, Address + 1)); 686 GV->setMetadata(LLVMContext::MD_absolute_symbol, 687 MDNode::get(Ctx, {MinC, MaxC})); 688 } 689 690 DenseMap<Function *, Value *> tableKernelIndexCache; 691 Value *getTableLookupKernelIndex(Module &M, Function *F) { 692 // Accesses from a function use the amdgcn_lds_kernel_id intrinsic which 693 // lowers to a read from a live in register. Emit it once in the entry 694 // block to spare deduplicating it later. 695 auto [It, Inserted] = tableKernelIndexCache.try_emplace(F); 696 if (Inserted) { 697 Function *Decl = 698 Intrinsic::getDeclaration(&M, Intrinsic::amdgcn_lds_kernel_id, {}); 699 700 auto InsertAt = F->getEntryBlock().getFirstNonPHIOrDbgOrAlloca(); 701 IRBuilder<> Builder(&*InsertAt); 702 703 It->second = Builder.CreateCall(Decl, {}); 704 } 705 706 return It->second; 707 } 708 709 static std::vector<Function *> assignLDSKernelIDToEachKernel( 710 Module *M, DenseSet<Function *> const &KernelsThatAllocateTableLDS, 711 DenseSet<Function *> const &KernelsThatIndirectlyAllocateDynamicLDS) { 712 // Associate kernels in the set with an arbirary but reproducible order and 713 // annotate them with that order in metadata. This metadata is recognised by 714 // the backend and lowered to a SGPR which can be read from using 715 // amdgcn_lds_kernel_id. 716 717 std::vector<Function *> OrderedKernels; 718 if (!KernelsThatAllocateTableLDS.empty() || 719 !KernelsThatIndirectlyAllocateDynamicLDS.empty()) { 720 721 for (Function &Func : M->functions()) { 722 if (Func.isDeclaration()) 723 continue; 724 if (!isKernelLDS(&Func)) 725 continue; 726 727 if (KernelsThatAllocateTableLDS.contains(&Func) || 728 KernelsThatIndirectlyAllocateDynamicLDS.contains(&Func)) { 729 assert(Func.hasName()); // else fatal error earlier 730 OrderedKernels.push_back(&Func); 731 } 732 } 733 734 // Put them in an arbitrary but reproducible order 735 OrderedKernels = sortByName(std::move(OrderedKernels)); 736 737 // Annotate the kernels with their order in this vector 738 LLVMContext &Ctx = M->getContext(); 739 IRBuilder<> Builder(Ctx); 740 741 if (OrderedKernels.size() > UINT32_MAX) { 742 // 32 bit keeps it in one SGPR. > 2**32 kernels won't fit on the GPU 743 report_fatal_error("Unimplemented LDS lowering for > 2**32 kernels"); 744 } 745 746 for (size_t i = 0; i < OrderedKernels.size(); i++) { 747 Metadata *AttrMDArgs[1] = { 748 ConstantAsMetadata::get(Builder.getInt32(i)), 749 }; 750 OrderedKernels[i]->setMetadata("llvm.amdgcn.lds.kernel.id", 751 MDNode::get(Ctx, AttrMDArgs)); 752 } 753 } 754 return OrderedKernels; 755 } 756 757 static void partitionVariablesIntoIndirectStrategies( 758 Module &M, LDSUsesInfoTy const &LDSUsesInfo, 759 VariableFunctionMap &LDSToKernelsThatNeedToAccessItIndirectly, 760 DenseSet<GlobalVariable *> &ModuleScopeVariables, 761 DenseSet<GlobalVariable *> &TableLookupVariables, 762 DenseSet<GlobalVariable *> &KernelAccessVariables, 763 DenseSet<GlobalVariable *> &DynamicVariables) { 764 765 GlobalVariable *HybridModuleRoot = 766 LoweringKindLoc != LoweringKind::hybrid 767 ? nullptr 768 : chooseBestVariableForModuleStrategy( 769 M.getDataLayout(), LDSToKernelsThatNeedToAccessItIndirectly); 770 771 DenseSet<Function *> const EmptySet; 772 DenseSet<Function *> const &HybridModuleRootKernels = 773 HybridModuleRoot 774 ? LDSToKernelsThatNeedToAccessItIndirectly[HybridModuleRoot] 775 : EmptySet; 776 777 for (auto &K : LDSToKernelsThatNeedToAccessItIndirectly) { 778 // Each iteration of this loop assigns exactly one global variable to 779 // exactly one of the implementation strategies. 780 781 GlobalVariable *GV = K.first; 782 assert(AMDGPU::isLDSVariableToLower(*GV)); 783 assert(K.second.size() != 0); 784 785 if (AMDGPU::isDynamicLDS(*GV)) { 786 DynamicVariables.insert(GV); 787 continue; 788 } 789 790 switch (LoweringKindLoc) { 791 case LoweringKind::module: 792 ModuleScopeVariables.insert(GV); 793 break; 794 795 case LoweringKind::table: 796 TableLookupVariables.insert(GV); 797 break; 798 799 case LoweringKind::kernel: 800 if (K.second.size() == 1) { 801 KernelAccessVariables.insert(GV); 802 } else { 803 report_fatal_error( 804 "cannot lower LDS '" + GV->getName() + 805 "' to kernel access as it is reachable from multiple kernels"); 806 } 807 break; 808 809 case LoweringKind::hybrid: { 810 if (GV == HybridModuleRoot) { 811 assert(K.second.size() != 1); 812 ModuleScopeVariables.insert(GV); 813 } else if (K.second.size() == 1) { 814 KernelAccessVariables.insert(GV); 815 } else if (set_is_subset(K.second, HybridModuleRootKernels)) { 816 ModuleScopeVariables.insert(GV); 817 } else { 818 TableLookupVariables.insert(GV); 819 } 820 break; 821 } 822 } 823 } 824 825 // All LDS variables accessed indirectly have now been partitioned into 826 // the distinct lowering strategies. 827 assert(ModuleScopeVariables.size() + TableLookupVariables.size() + 828 KernelAccessVariables.size() + DynamicVariables.size() == 829 LDSToKernelsThatNeedToAccessItIndirectly.size()); 830 } 831 832 static GlobalVariable *lowerModuleScopeStructVariables( 833 Module &M, DenseSet<GlobalVariable *> const &ModuleScopeVariables, 834 DenseSet<Function *> const &KernelsThatAllocateModuleLDS) { 835 // Create a struct to hold the ModuleScopeVariables 836 // Replace all uses of those variables from non-kernel functions with the 837 // new struct instance Replace only the uses from kernel functions that will 838 // allocate this instance. That is a space optimisation - kernels that use a 839 // subset of the module scope struct and do not need to allocate it for 840 // indirect calls will only allocate the subset they use (they do so as part 841 // of the per-kernel lowering). 842 if (ModuleScopeVariables.empty()) { 843 return nullptr; 844 } 845 846 LLVMContext &Ctx = M.getContext(); 847 848 LDSVariableReplacement ModuleScopeReplacement = 849 createLDSVariableReplacement(M, "llvm.amdgcn.module.lds", 850 ModuleScopeVariables); 851 852 appendToCompilerUsed(M, {static_cast<GlobalValue *>( 853 ConstantExpr::getPointerBitCastOrAddrSpaceCast( 854 cast<Constant>(ModuleScopeReplacement.SGV), 855 PointerType::getUnqual(Ctx)))}); 856 857 // module.lds will be allocated at zero in any kernel that allocates it 858 recordLDSAbsoluteAddress(&M, ModuleScopeReplacement.SGV, 0); 859 860 // historic 861 removeLocalVarsFromUsedLists(M, ModuleScopeVariables); 862 863 // Replace all uses of module scope variable from non-kernel functions 864 replaceLDSVariablesWithStruct( 865 M, ModuleScopeVariables, ModuleScopeReplacement, [&](Use &U) { 866 Instruction *I = dyn_cast<Instruction>(U.getUser()); 867 if (!I) { 868 return false; 869 } 870 Function *F = I->getFunction(); 871 return !isKernelLDS(F); 872 }); 873 874 // Replace uses of module scope variable from kernel functions that 875 // allocate the module scope variable, otherwise leave them unchanged 876 // Record on each kernel whether the module scope global is used by it 877 878 for (Function &Func : M.functions()) { 879 if (Func.isDeclaration() || !isKernelLDS(&Func)) 880 continue; 881 882 if (KernelsThatAllocateModuleLDS.contains(&Func)) { 883 replaceLDSVariablesWithStruct( 884 M, ModuleScopeVariables, ModuleScopeReplacement, [&](Use &U) { 885 Instruction *I = dyn_cast<Instruction>(U.getUser()); 886 if (!I) { 887 return false; 888 } 889 Function *F = I->getFunction(); 890 return F == &Func; 891 }); 892 893 markUsedByKernel(&Func, ModuleScopeReplacement.SGV); 894 } 895 } 896 897 return ModuleScopeReplacement.SGV; 898 } 899 900 static DenseMap<Function *, LDSVariableReplacement> 901 lowerKernelScopeStructVariables( 902 Module &M, LDSUsesInfoTy &LDSUsesInfo, 903 DenseSet<GlobalVariable *> const &ModuleScopeVariables, 904 DenseSet<Function *> const &KernelsThatAllocateModuleLDS, 905 GlobalVariable *MaybeModuleScopeStruct) { 906 907 // Create a struct for each kernel for the non-module-scope variables. 908 909 DenseMap<Function *, LDSVariableReplacement> KernelToReplacement; 910 for (Function &Func : M.functions()) { 911 if (Func.isDeclaration() || !isKernelLDS(&Func)) 912 continue; 913 914 DenseSet<GlobalVariable *> KernelUsedVariables; 915 // Allocating variables that are used directly in this struct to get 916 // alignment aware allocation and predictable frame size. 917 for (auto &v : LDSUsesInfo.direct_access[&Func]) { 918 if (!AMDGPU::isDynamicLDS(*v)) { 919 KernelUsedVariables.insert(v); 920 } 921 } 922 923 // Allocating variables that are accessed indirectly so that a lookup of 924 // this struct instance can find them from nested functions. 925 for (auto &v : LDSUsesInfo.indirect_access[&Func]) { 926 if (!AMDGPU::isDynamicLDS(*v)) { 927 KernelUsedVariables.insert(v); 928 } 929 } 930 931 // Variables allocated in module lds must all resolve to that struct, 932 // not to the per-kernel instance. 933 if (KernelsThatAllocateModuleLDS.contains(&Func)) { 934 for (GlobalVariable *v : ModuleScopeVariables) { 935 KernelUsedVariables.erase(v); 936 } 937 } 938 939 if (KernelUsedVariables.empty()) { 940 // Either used no LDS, or the LDS it used was all in the module struct 941 // or dynamically sized 942 continue; 943 } 944 945 // The association between kernel function and LDS struct is done by 946 // symbol name, which only works if the function in question has a 947 // name This is not expected to be a problem in practice as kernels 948 // are called by name making anonymous ones (which are named by the 949 // backend) difficult to use. This does mean that llvm test cases need 950 // to name the kernels. 951 if (!Func.hasName()) { 952 report_fatal_error("Anonymous kernels cannot use LDS variables"); 953 } 954 955 std::string VarName = 956 (Twine("llvm.amdgcn.kernel.") + Func.getName() + ".lds").str(); 957 958 auto Replacement = 959 createLDSVariableReplacement(M, VarName, KernelUsedVariables); 960 961 // If any indirect uses, create a direct use to ensure allocation 962 // TODO: Simpler to unconditionally mark used but that regresses 963 // codegen in test/CodeGen/AMDGPU/noclobber-barrier.ll 964 auto Accesses = LDSUsesInfo.indirect_access.find(&Func); 965 if ((Accesses != LDSUsesInfo.indirect_access.end()) && 966 !Accesses->second.empty()) 967 markUsedByKernel(&Func, Replacement.SGV); 968 969 // remove preserves existing codegen 970 removeLocalVarsFromUsedLists(M, KernelUsedVariables); 971 KernelToReplacement[&Func] = Replacement; 972 973 // Rewrite uses within kernel to the new struct 974 replaceLDSVariablesWithStruct( 975 M, KernelUsedVariables, Replacement, [&Func](Use &U) { 976 Instruction *I = dyn_cast<Instruction>(U.getUser()); 977 return I && I->getFunction() == &Func; 978 }); 979 } 980 return KernelToReplacement; 981 } 982 983 static GlobalVariable * 984 buildRepresentativeDynamicLDSInstance(Module &M, LDSUsesInfoTy &LDSUsesInfo, 985 Function *func) { 986 // Create a dynamic lds variable with a name associated with the passed 987 // function that has the maximum alignment of any dynamic lds variable 988 // reachable from this kernel. Dynamic LDS is allocated after the static LDS 989 // allocation, possibly after alignment padding. The representative variable 990 // created here has the maximum alignment of any other dynamic variable 991 // reachable by that kernel. All dynamic LDS variables are allocated at the 992 // same address in each kernel in order to provide the documented aliasing 993 // semantics. Setting the alignment here allows this IR pass to accurately 994 // predict the exact constant at which it will be allocated. 995 996 assert(isKernelLDS(func)); 997 998 LLVMContext &Ctx = M.getContext(); 999 const DataLayout &DL = M.getDataLayout(); 1000 Align MaxDynamicAlignment(1); 1001 1002 auto UpdateMaxAlignment = [&MaxDynamicAlignment, &DL](GlobalVariable *GV) { 1003 if (AMDGPU::isDynamicLDS(*GV)) { 1004 MaxDynamicAlignment = 1005 std::max(MaxDynamicAlignment, AMDGPU::getAlign(DL, GV)); 1006 } 1007 }; 1008 1009 for (GlobalVariable *GV : LDSUsesInfo.indirect_access[func]) { 1010 UpdateMaxAlignment(GV); 1011 } 1012 1013 for (GlobalVariable *GV : LDSUsesInfo.direct_access[func]) { 1014 UpdateMaxAlignment(GV); 1015 } 1016 1017 assert(func->hasName()); // Checked by caller 1018 auto emptyCharArray = ArrayType::get(Type::getInt8Ty(Ctx), 0); 1019 GlobalVariable *N = new GlobalVariable( 1020 M, emptyCharArray, false, GlobalValue::ExternalLinkage, nullptr, 1021 Twine("llvm.amdgcn." + func->getName() + ".dynlds"), nullptr, GlobalValue::NotThreadLocal, AMDGPUAS::LOCAL_ADDRESS, 1022 false); 1023 N->setAlignment(MaxDynamicAlignment); 1024 1025 assert(AMDGPU::isDynamicLDS(*N)); 1026 return N; 1027 } 1028 1029 /// Strip "amdgpu-no-lds-kernel-id" from any functions where we may have 1030 /// introduced its use. If AMDGPUAttributor ran prior to the pass, we inferred 1031 /// the lack of llvm.amdgcn.lds.kernel.id calls. 1032 void removeNoLdsKernelIdFromReachable(CallGraph &CG, Function *KernelRoot) { 1033 KernelRoot->removeFnAttr("amdgpu-no-lds-kernel-id"); 1034 1035 SmallVector<Function *> Tmp({CG[KernelRoot]->getFunction()}); 1036 if (!Tmp.back()) 1037 return; 1038 1039 SmallPtrSet<Function *, 8> Visited; 1040 bool SeenUnknownCall = false; 1041 1042 do { 1043 Function *F = Tmp.pop_back_val(); 1044 1045 for (auto &N : *CG[F]) { 1046 if (!N.second) 1047 continue; 1048 1049 Function *Callee = N.second->getFunction(); 1050 if (!Callee) { 1051 if (!SeenUnknownCall) { 1052 SeenUnknownCall = true; 1053 1054 // If we see any indirect calls, assume nothing about potential 1055 // targets. 1056 // TODO: This could be refined to possible LDS global users. 1057 for (auto &N : *CG.getExternalCallingNode()) { 1058 Function *PotentialCallee = N.second->getFunction(); 1059 if (!isKernelLDS(PotentialCallee)) 1060 PotentialCallee->removeFnAttr("amdgpu-no-lds-kernel-id"); 1061 } 1062 1063 continue; 1064 } 1065 } 1066 1067 Callee->removeFnAttr("amdgpu-no-lds-kernel-id"); 1068 if (Visited.insert(Callee).second) 1069 Tmp.push_back(Callee); 1070 } 1071 } while (!Tmp.empty()); 1072 } 1073 1074 DenseMap<Function *, GlobalVariable *> lowerDynamicLDSVariables( 1075 Module &M, LDSUsesInfoTy &LDSUsesInfo, 1076 DenseSet<Function *> const &KernelsThatIndirectlyAllocateDynamicLDS, 1077 DenseSet<GlobalVariable *> const &DynamicVariables, 1078 std::vector<Function *> const &OrderedKernels) { 1079 DenseMap<Function *, GlobalVariable *> KernelToCreatedDynamicLDS; 1080 if (!KernelsThatIndirectlyAllocateDynamicLDS.empty()) { 1081 LLVMContext &Ctx = M.getContext(); 1082 IRBuilder<> Builder(Ctx); 1083 Type *I32 = Type::getInt32Ty(Ctx); 1084 1085 std::vector<Constant *> newDynamicLDS; 1086 1087 // Table is built in the same order as OrderedKernels 1088 for (auto &func : OrderedKernels) { 1089 1090 if (KernelsThatIndirectlyAllocateDynamicLDS.contains(func)) { 1091 assert(isKernelLDS(func)); 1092 if (!func->hasName()) { 1093 report_fatal_error("Anonymous kernels cannot use LDS variables"); 1094 } 1095 1096 GlobalVariable *N = 1097 buildRepresentativeDynamicLDSInstance(M, LDSUsesInfo, func); 1098 1099 KernelToCreatedDynamicLDS[func] = N; 1100 1101 markUsedByKernel(func, N); 1102 1103 auto emptyCharArray = ArrayType::get(Type::getInt8Ty(Ctx), 0); 1104 auto GEP = ConstantExpr::getGetElementPtr( 1105 emptyCharArray, N, ConstantInt::get(I32, 0), true); 1106 newDynamicLDS.push_back(ConstantExpr::getPtrToInt(GEP, I32)); 1107 } else { 1108 newDynamicLDS.push_back(PoisonValue::get(I32)); 1109 } 1110 } 1111 assert(OrderedKernels.size() == newDynamicLDS.size()); 1112 1113 ArrayType *t = ArrayType::get(I32, newDynamicLDS.size()); 1114 Constant *init = ConstantArray::get(t, newDynamicLDS); 1115 GlobalVariable *table = new GlobalVariable( 1116 M, t, true, GlobalValue::InternalLinkage, init, 1117 "llvm.amdgcn.dynlds.offset.table", nullptr, 1118 GlobalValue::NotThreadLocal, AMDGPUAS::CONSTANT_ADDRESS); 1119 1120 for (GlobalVariable *GV : DynamicVariables) { 1121 for (Use &U : make_early_inc_range(GV->uses())) { 1122 auto *I = dyn_cast<Instruction>(U.getUser()); 1123 if (!I) 1124 continue; 1125 if (isKernelLDS(I->getFunction())) 1126 continue; 1127 1128 replaceUseWithTableLookup(M, Builder, table, GV, U, nullptr); 1129 } 1130 } 1131 } 1132 return KernelToCreatedDynamicLDS; 1133 } 1134 1135 bool runOnModule(Module &M) { 1136 CallGraph CG = CallGraph(M); 1137 bool Changed = superAlignLDSGlobals(M); 1138 1139 Changed |= eliminateConstantExprUsesOfLDSFromAllInstructions(M); 1140 1141 Changed = true; // todo: narrow this down 1142 1143 // For each kernel, what variables does it access directly or through 1144 // callees 1145 LDSUsesInfoTy LDSUsesInfo = getTransitiveUsesOfLDS(CG, M); 1146 1147 // For each variable accessed through callees, which kernels access it 1148 VariableFunctionMap LDSToKernelsThatNeedToAccessItIndirectly; 1149 for (auto &K : LDSUsesInfo.indirect_access) { 1150 Function *F = K.first; 1151 assert(isKernelLDS(F)); 1152 for (GlobalVariable *GV : K.second) { 1153 LDSToKernelsThatNeedToAccessItIndirectly[GV].insert(F); 1154 } 1155 } 1156 1157 // Partition variables accessed indirectly into the different strategies 1158 DenseSet<GlobalVariable *> ModuleScopeVariables; 1159 DenseSet<GlobalVariable *> TableLookupVariables; 1160 DenseSet<GlobalVariable *> KernelAccessVariables; 1161 DenseSet<GlobalVariable *> DynamicVariables; 1162 partitionVariablesIntoIndirectStrategies( 1163 M, LDSUsesInfo, LDSToKernelsThatNeedToAccessItIndirectly, 1164 ModuleScopeVariables, TableLookupVariables, KernelAccessVariables, 1165 DynamicVariables); 1166 1167 // If the kernel accesses a variable that is going to be stored in the 1168 // module instance through a call then that kernel needs to allocate the 1169 // module instance 1170 const DenseSet<Function *> KernelsThatAllocateModuleLDS = 1171 kernelsThatIndirectlyAccessAnyOfPassedVariables(M, LDSUsesInfo, 1172 ModuleScopeVariables); 1173 const DenseSet<Function *> KernelsThatAllocateTableLDS = 1174 kernelsThatIndirectlyAccessAnyOfPassedVariables(M, LDSUsesInfo, 1175 TableLookupVariables); 1176 1177 const DenseSet<Function *> KernelsThatIndirectlyAllocateDynamicLDS = 1178 kernelsThatIndirectlyAccessAnyOfPassedVariables(M, LDSUsesInfo, 1179 DynamicVariables); 1180 1181 GlobalVariable *MaybeModuleScopeStruct = lowerModuleScopeStructVariables( 1182 M, ModuleScopeVariables, KernelsThatAllocateModuleLDS); 1183 1184 DenseMap<Function *, LDSVariableReplacement> KernelToReplacement = 1185 lowerKernelScopeStructVariables(M, LDSUsesInfo, ModuleScopeVariables, 1186 KernelsThatAllocateModuleLDS, 1187 MaybeModuleScopeStruct); 1188 1189 // Lower zero cost accesses to the kernel instances just created 1190 for (auto &GV : KernelAccessVariables) { 1191 auto &funcs = LDSToKernelsThatNeedToAccessItIndirectly[GV]; 1192 assert(funcs.size() == 1); // Only one kernel can access it 1193 LDSVariableReplacement Replacement = 1194 KernelToReplacement[*(funcs.begin())]; 1195 1196 DenseSet<GlobalVariable *> Vec; 1197 Vec.insert(GV); 1198 1199 replaceLDSVariablesWithStruct(M, Vec, Replacement, [](Use &U) { 1200 return isa<Instruction>(U.getUser()); 1201 }); 1202 } 1203 1204 // The ith element of this vector is kernel id i 1205 std::vector<Function *> OrderedKernels = 1206 assignLDSKernelIDToEachKernel(&M, KernelsThatAllocateTableLDS, 1207 KernelsThatIndirectlyAllocateDynamicLDS); 1208 1209 if (!KernelsThatAllocateTableLDS.empty()) { 1210 LLVMContext &Ctx = M.getContext(); 1211 IRBuilder<> Builder(Ctx); 1212 1213 // The order must be consistent between lookup table and accesses to 1214 // lookup table 1215 auto TableLookupVariablesOrdered = 1216 sortByName(std::vector<GlobalVariable *>(TableLookupVariables.begin(), 1217 TableLookupVariables.end())); 1218 1219 GlobalVariable *LookupTable = buildLookupTable( 1220 M, TableLookupVariablesOrdered, OrderedKernels, KernelToReplacement); 1221 replaceUsesInInstructionsWithTableLookup(M, TableLookupVariablesOrdered, 1222 LookupTable); 1223 1224 // Strip amdgpu-no-lds-kernel-id from all functions reachable from the 1225 // kernel. We may have inferred this wasn't used prior to the pass. 1226 // 1227 // TODO: We could filter out subgraphs that do not access LDS globals. 1228 for (Function *F : KernelsThatAllocateTableLDS) 1229 removeNoLdsKernelIdFromReachable(CG, F); 1230 } 1231 1232 DenseMap<Function *, GlobalVariable *> KernelToCreatedDynamicLDS = 1233 lowerDynamicLDSVariables(M, LDSUsesInfo, 1234 KernelsThatIndirectlyAllocateDynamicLDS, 1235 DynamicVariables, OrderedKernels); 1236 1237 // All kernel frames have been allocated. Calculate and record the 1238 // addresses. 1239 { 1240 const DataLayout &DL = M.getDataLayout(); 1241 1242 for (Function &Func : M.functions()) { 1243 if (Func.isDeclaration() || !isKernelLDS(&Func)) 1244 continue; 1245 1246 // All three of these are optional. The first variable is allocated at 1247 // zero. They are allocated by AMDGPUMachineFunction as one block. 1248 // Layout: 1249 //{ 1250 // module.lds 1251 // alignment padding 1252 // kernel instance 1253 // alignment padding 1254 // dynamic lds variables 1255 //} 1256 1257 const bool AllocateModuleScopeStruct = 1258 MaybeModuleScopeStruct && 1259 KernelsThatAllocateModuleLDS.contains(&Func); 1260 1261 auto Replacement = KernelToReplacement.find(&Func); 1262 const bool AllocateKernelScopeStruct = 1263 Replacement != KernelToReplacement.end(); 1264 1265 const bool AllocateDynamicVariable = 1266 KernelToCreatedDynamicLDS.contains(&Func); 1267 1268 uint32_t Offset = 0; 1269 1270 if (AllocateModuleScopeStruct) { 1271 // Allocated at zero, recorded once on construction, not once per 1272 // kernel 1273 Offset += DL.getTypeAllocSize(MaybeModuleScopeStruct->getValueType()); 1274 } 1275 1276 if (AllocateKernelScopeStruct) { 1277 GlobalVariable *KernelStruct = Replacement->second.SGV; 1278 Offset = alignTo(Offset, AMDGPU::getAlign(DL, KernelStruct)); 1279 recordLDSAbsoluteAddress(&M, KernelStruct, Offset); 1280 Offset += DL.getTypeAllocSize(KernelStruct->getValueType()); 1281 } 1282 1283 // If there is dynamic allocation, the alignment needed is included in 1284 // the static frame size. There may be no reference to the dynamic 1285 // variable in the kernel itself, so without including it here, that 1286 // alignment padding could be missed. 1287 if (AllocateDynamicVariable) { 1288 GlobalVariable *DynamicVariable = KernelToCreatedDynamicLDS[&Func]; 1289 Offset = alignTo(Offset, AMDGPU::getAlign(DL, DynamicVariable)); 1290 recordLDSAbsoluteAddress(&M, DynamicVariable, Offset); 1291 } 1292 1293 if (Offset != 0) { 1294 (void)TM; // TODO: Account for target maximum LDS 1295 std::string Buffer; 1296 raw_string_ostream SS{Buffer}; 1297 SS << format("%u", Offset); 1298 1299 // Instead of explictly marking kernels that access dynamic variables 1300 // using special case metadata, annotate with min-lds == max-lds, i.e. 1301 // that there is no more space available for allocating more static 1302 // LDS variables. That is the right condition to prevent allocating 1303 // more variables which would collide with the addresses assigned to 1304 // dynamic variables. 1305 if (AllocateDynamicVariable) 1306 SS << format(",%u", Offset); 1307 1308 Func.addFnAttr("amdgpu-lds-size", Buffer); 1309 } 1310 } 1311 } 1312 1313 for (auto &GV : make_early_inc_range(M.globals())) 1314 if (AMDGPU::isLDSVariableToLower(GV)) { 1315 // probably want to remove from used lists 1316 GV.removeDeadConstantUsers(); 1317 if (GV.use_empty()) 1318 GV.eraseFromParent(); 1319 } 1320 1321 return Changed; 1322 } 1323 1324 private: 1325 // Increase the alignment of LDS globals if necessary to maximise the chance 1326 // that we can use aligned LDS instructions to access them. 1327 static bool superAlignLDSGlobals(Module &M) { 1328 const DataLayout &DL = M.getDataLayout(); 1329 bool Changed = false; 1330 if (!SuperAlignLDSGlobals) { 1331 return Changed; 1332 } 1333 1334 for (auto &GV : M.globals()) { 1335 if (GV.getType()->getPointerAddressSpace() != AMDGPUAS::LOCAL_ADDRESS) { 1336 // Only changing alignment of LDS variables 1337 continue; 1338 } 1339 if (!GV.hasInitializer()) { 1340 // cuda/hip extern __shared__ variable, leave alignment alone 1341 continue; 1342 } 1343 1344 Align Alignment = AMDGPU::getAlign(DL, &GV); 1345 TypeSize GVSize = DL.getTypeAllocSize(GV.getValueType()); 1346 1347 if (GVSize > 8) { 1348 // We might want to use a b96 or b128 load/store 1349 Alignment = std::max(Alignment, Align(16)); 1350 } else if (GVSize > 4) { 1351 // We might want to use a b64 load/store 1352 Alignment = std::max(Alignment, Align(8)); 1353 } else if (GVSize > 2) { 1354 // We might want to use a b32 load/store 1355 Alignment = std::max(Alignment, Align(4)); 1356 } else if (GVSize > 1) { 1357 // We might want to use a b16 load/store 1358 Alignment = std::max(Alignment, Align(2)); 1359 } 1360 1361 if (Alignment != AMDGPU::getAlign(DL, &GV)) { 1362 Changed = true; 1363 GV.setAlignment(Alignment); 1364 } 1365 } 1366 return Changed; 1367 } 1368 1369 static LDSVariableReplacement createLDSVariableReplacement( 1370 Module &M, std::string VarName, 1371 DenseSet<GlobalVariable *> const &LDSVarsToTransform) { 1372 // Create a struct instance containing LDSVarsToTransform and map from those 1373 // variables to ConstantExprGEP 1374 // Variables may be introduced to meet alignment requirements. No aliasing 1375 // metadata is useful for these as they have no uses. Erased before return. 1376 1377 LLVMContext &Ctx = M.getContext(); 1378 const DataLayout &DL = M.getDataLayout(); 1379 assert(!LDSVarsToTransform.empty()); 1380 1381 SmallVector<OptimizedStructLayoutField, 8> LayoutFields; 1382 LayoutFields.reserve(LDSVarsToTransform.size()); 1383 { 1384 // The order of fields in this struct depends on the order of 1385 // varables in the argument which varies when changing how they 1386 // are identified, leading to spurious test breakage. 1387 auto Sorted = sortByName(std::vector<GlobalVariable *>( 1388 LDSVarsToTransform.begin(), LDSVarsToTransform.end())); 1389 1390 for (GlobalVariable *GV : Sorted) { 1391 OptimizedStructLayoutField F(GV, 1392 DL.getTypeAllocSize(GV->getValueType()), 1393 AMDGPU::getAlign(DL, GV)); 1394 LayoutFields.emplace_back(F); 1395 } 1396 } 1397 1398 performOptimizedStructLayout(LayoutFields); 1399 1400 std::vector<GlobalVariable *> LocalVars; 1401 BitVector IsPaddingField; 1402 LocalVars.reserve(LDSVarsToTransform.size()); // will be at least this large 1403 IsPaddingField.reserve(LDSVarsToTransform.size()); 1404 { 1405 uint64_t CurrentOffset = 0; 1406 for (size_t I = 0; I < LayoutFields.size(); I++) { 1407 GlobalVariable *FGV = static_cast<GlobalVariable *>( 1408 const_cast<void *>(LayoutFields[I].Id)); 1409 Align DataAlign = LayoutFields[I].Alignment; 1410 1411 uint64_t DataAlignV = DataAlign.value(); 1412 if (uint64_t Rem = CurrentOffset % DataAlignV) { 1413 uint64_t Padding = DataAlignV - Rem; 1414 1415 // Append an array of padding bytes to meet alignment requested 1416 // Note (o + (a - (o % a)) ) % a == 0 1417 // (offset + Padding ) % align == 0 1418 1419 Type *ATy = ArrayType::get(Type::getInt8Ty(Ctx), Padding); 1420 LocalVars.push_back(new GlobalVariable( 1421 M, ATy, false, GlobalValue::InternalLinkage, 1422 PoisonValue::get(ATy), "", nullptr, GlobalValue::NotThreadLocal, 1423 AMDGPUAS::LOCAL_ADDRESS, false)); 1424 IsPaddingField.push_back(true); 1425 CurrentOffset += Padding; 1426 } 1427 1428 LocalVars.push_back(FGV); 1429 IsPaddingField.push_back(false); 1430 CurrentOffset += LayoutFields[I].Size; 1431 } 1432 } 1433 1434 std::vector<Type *> LocalVarTypes; 1435 LocalVarTypes.reserve(LocalVars.size()); 1436 std::transform( 1437 LocalVars.cbegin(), LocalVars.cend(), std::back_inserter(LocalVarTypes), 1438 [](const GlobalVariable *V) -> Type * { return V->getValueType(); }); 1439 1440 StructType *LDSTy = StructType::create(Ctx, LocalVarTypes, VarName + ".t"); 1441 1442 Align StructAlign = AMDGPU::getAlign(DL, LocalVars[0]); 1443 1444 GlobalVariable *SGV = new GlobalVariable( 1445 M, LDSTy, false, GlobalValue::InternalLinkage, PoisonValue::get(LDSTy), 1446 VarName, nullptr, GlobalValue::NotThreadLocal, AMDGPUAS::LOCAL_ADDRESS, 1447 false); 1448 SGV->setAlignment(StructAlign); 1449 1450 DenseMap<GlobalVariable *, Constant *> Map; 1451 Type *I32 = Type::getInt32Ty(Ctx); 1452 for (size_t I = 0; I < LocalVars.size(); I++) { 1453 GlobalVariable *GV = LocalVars[I]; 1454 Constant *GEPIdx[] = {ConstantInt::get(I32, 0), ConstantInt::get(I32, I)}; 1455 Constant *GEP = ConstantExpr::getGetElementPtr(LDSTy, SGV, GEPIdx, true); 1456 if (IsPaddingField[I]) { 1457 assert(GV->use_empty()); 1458 GV->eraseFromParent(); 1459 } else { 1460 Map[GV] = GEP; 1461 } 1462 } 1463 assert(Map.size() == LDSVarsToTransform.size()); 1464 return {SGV, std::move(Map)}; 1465 } 1466 1467 template <typename PredicateTy> 1468 static void replaceLDSVariablesWithStruct( 1469 Module &M, DenseSet<GlobalVariable *> const &LDSVarsToTransformArg, 1470 const LDSVariableReplacement &Replacement, PredicateTy Predicate) { 1471 LLVMContext &Ctx = M.getContext(); 1472 const DataLayout &DL = M.getDataLayout(); 1473 1474 // A hack... we need to insert the aliasing info in a predictable order for 1475 // lit tests. Would like to have them in a stable order already, ideally the 1476 // same order they get allocated, which might mean an ordered set container 1477 auto LDSVarsToTransform = sortByName(std::vector<GlobalVariable *>( 1478 LDSVarsToTransformArg.begin(), LDSVarsToTransformArg.end())); 1479 1480 // Create alias.scope and their lists. Each field in the new structure 1481 // does not alias with all other fields. 1482 SmallVector<MDNode *> AliasScopes; 1483 SmallVector<Metadata *> NoAliasList; 1484 const size_t NumberVars = LDSVarsToTransform.size(); 1485 if (NumberVars > 1) { 1486 MDBuilder MDB(Ctx); 1487 AliasScopes.reserve(NumberVars); 1488 MDNode *Domain = MDB.createAnonymousAliasScopeDomain(); 1489 for (size_t I = 0; I < NumberVars; I++) { 1490 MDNode *Scope = MDB.createAnonymousAliasScope(Domain); 1491 AliasScopes.push_back(Scope); 1492 } 1493 NoAliasList.append(&AliasScopes[1], AliasScopes.end()); 1494 } 1495 1496 // Replace uses of ith variable with a constantexpr to the corresponding 1497 // field of the instance that will be allocated by AMDGPUMachineFunction 1498 for (size_t I = 0; I < NumberVars; I++) { 1499 GlobalVariable *GV = LDSVarsToTransform[I]; 1500 Constant *GEP = Replacement.LDSVarsToConstantGEP.at(GV); 1501 1502 GV->replaceUsesWithIf(GEP, Predicate); 1503 1504 APInt APOff(DL.getIndexTypeSizeInBits(GEP->getType()), 0); 1505 GEP->stripAndAccumulateInBoundsConstantOffsets(DL, APOff); 1506 uint64_t Offset = APOff.getZExtValue(); 1507 1508 Align A = 1509 commonAlignment(Replacement.SGV->getAlign().valueOrOne(), Offset); 1510 1511 if (I) 1512 NoAliasList[I - 1] = AliasScopes[I - 1]; 1513 MDNode *NoAlias = 1514 NoAliasList.empty() ? nullptr : MDNode::get(Ctx, NoAliasList); 1515 MDNode *AliasScope = 1516 AliasScopes.empty() ? nullptr : MDNode::get(Ctx, {AliasScopes[I]}); 1517 1518 refineUsesAlignmentAndAA(GEP, A, DL, AliasScope, NoAlias); 1519 } 1520 } 1521 1522 static void refineUsesAlignmentAndAA(Value *Ptr, Align A, 1523 const DataLayout &DL, MDNode *AliasScope, 1524 MDNode *NoAlias, unsigned MaxDepth = 5) { 1525 if (!MaxDepth || (A == 1 && !AliasScope)) 1526 return; 1527 1528 for (User *U : Ptr->users()) { 1529 if (auto *I = dyn_cast<Instruction>(U)) { 1530 if (AliasScope && I->mayReadOrWriteMemory()) { 1531 MDNode *AS = I->getMetadata(LLVMContext::MD_alias_scope); 1532 AS = (AS ? MDNode::getMostGenericAliasScope(AS, AliasScope) 1533 : AliasScope); 1534 I->setMetadata(LLVMContext::MD_alias_scope, AS); 1535 1536 MDNode *NA = I->getMetadata(LLVMContext::MD_noalias); 1537 NA = (NA ? MDNode::intersect(NA, NoAlias) : NoAlias); 1538 I->setMetadata(LLVMContext::MD_noalias, NA); 1539 } 1540 } 1541 1542 if (auto *LI = dyn_cast<LoadInst>(U)) { 1543 LI->setAlignment(std::max(A, LI->getAlign())); 1544 continue; 1545 } 1546 if (auto *SI = dyn_cast<StoreInst>(U)) { 1547 if (SI->getPointerOperand() == Ptr) 1548 SI->setAlignment(std::max(A, SI->getAlign())); 1549 continue; 1550 } 1551 if (auto *AI = dyn_cast<AtomicRMWInst>(U)) { 1552 // None of atomicrmw operations can work on pointers, but let's 1553 // check it anyway in case it will or we will process ConstantExpr. 1554 if (AI->getPointerOperand() == Ptr) 1555 AI->setAlignment(std::max(A, AI->getAlign())); 1556 continue; 1557 } 1558 if (auto *AI = dyn_cast<AtomicCmpXchgInst>(U)) { 1559 if (AI->getPointerOperand() == Ptr) 1560 AI->setAlignment(std::max(A, AI->getAlign())); 1561 continue; 1562 } 1563 if (auto *GEP = dyn_cast<GetElementPtrInst>(U)) { 1564 unsigned BitWidth = DL.getIndexTypeSizeInBits(GEP->getType()); 1565 APInt Off(BitWidth, 0); 1566 if (GEP->getPointerOperand() == Ptr) { 1567 Align GA; 1568 if (GEP->accumulateConstantOffset(DL, Off)) 1569 GA = commonAlignment(A, Off.getLimitedValue()); 1570 refineUsesAlignmentAndAA(GEP, GA, DL, AliasScope, NoAlias, 1571 MaxDepth - 1); 1572 } 1573 continue; 1574 } 1575 if (auto *I = dyn_cast<Instruction>(U)) { 1576 if (I->getOpcode() == Instruction::BitCast || 1577 I->getOpcode() == Instruction::AddrSpaceCast) 1578 refineUsesAlignmentAndAA(I, A, DL, AliasScope, NoAlias, MaxDepth - 1); 1579 } 1580 } 1581 } 1582 }; 1583 1584 class AMDGPULowerModuleLDSLegacy : public ModulePass { 1585 public: 1586 const AMDGPUTargetMachine *TM; 1587 static char ID; 1588 1589 AMDGPULowerModuleLDSLegacy(const AMDGPUTargetMachine *TM_ = nullptr) 1590 : ModulePass(ID), TM(TM_) { 1591 initializeAMDGPULowerModuleLDSLegacyPass(*PassRegistry::getPassRegistry()); 1592 } 1593 1594 void getAnalysisUsage(AnalysisUsage &AU) const override { 1595 if (!TM) 1596 AU.addRequired<TargetPassConfig>(); 1597 } 1598 1599 bool runOnModule(Module &M) override { 1600 if (!TM) { 1601 auto &TPC = getAnalysis<TargetPassConfig>(); 1602 TM = &TPC.getTM<AMDGPUTargetMachine>(); 1603 } 1604 1605 return AMDGPULowerModuleLDS(*TM).runOnModule(M); 1606 } 1607 }; 1608 1609 } // namespace 1610 char AMDGPULowerModuleLDSLegacy::ID = 0; 1611 1612 char &llvm::AMDGPULowerModuleLDSLegacyPassID = AMDGPULowerModuleLDSLegacy::ID; 1613 1614 INITIALIZE_PASS_BEGIN(AMDGPULowerModuleLDSLegacy, DEBUG_TYPE, 1615 "Lower uses of LDS variables from non-kernel functions", 1616 false, false) 1617 INITIALIZE_PASS_DEPENDENCY(TargetPassConfig) 1618 INITIALIZE_PASS_END(AMDGPULowerModuleLDSLegacy, DEBUG_TYPE, 1619 "Lower uses of LDS variables from non-kernel functions", 1620 false, false) 1621 1622 ModulePass * 1623 llvm::createAMDGPULowerModuleLDSLegacyPass(const AMDGPUTargetMachine *TM) { 1624 return new AMDGPULowerModuleLDSLegacy(TM); 1625 } 1626 1627 PreservedAnalyses AMDGPULowerModuleLDSPass::run(Module &M, 1628 ModuleAnalysisManager &) { 1629 return AMDGPULowerModuleLDS(TM).runOnModule(M) ? PreservedAnalyses::none() 1630 : PreservedAnalyses::all(); 1631 } 1632