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 "Utils/AMDGPUBaseInfo.h" 181 #include "Utils/AMDGPUMemoryUtils.h" 182 #include "llvm/ADT/BitVector.h" 183 #include "llvm/ADT/DenseMap.h" 184 #include "llvm/ADT/DenseSet.h" 185 #include "llvm/ADT/STLExtras.h" 186 #include "llvm/ADT/SetOperations.h" 187 #include "llvm/ADT/SetVector.h" 188 #include "llvm/Analysis/CallGraph.h" 189 #include "llvm/IR/Constants.h" 190 #include "llvm/IR/DerivedTypes.h" 191 #include "llvm/IR/IRBuilder.h" 192 #include "llvm/IR/InlineAsm.h" 193 #include "llvm/IR/Instructions.h" 194 #include "llvm/IR/IntrinsicsAMDGPU.h" 195 #include "llvm/IR/MDBuilder.h" 196 #include "llvm/IR/ReplaceConstant.h" 197 #include "llvm/InitializePasses.h" 198 #include "llvm/Pass.h" 199 #include "llvm/Support/CommandLine.h" 200 #include "llvm/Support/Debug.h" 201 #include "llvm/Support/Format.h" 202 #include "llvm/Support/OptimizedStructLayout.h" 203 #include "llvm/Support/raw_ostream.h" 204 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 205 #include "llvm/Transforms/Utils/ModuleUtils.h" 206 207 #include <tuple> 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 : public ModulePass { 256 257 static void 258 removeLocalVarsFromUsedLists(Module &M, 259 const DenseSet<GlobalVariable *> &LocalVars) { 260 // The verifier rejects used lists containing an inttoptr of a constant 261 // so remove the variables from these lists before replaceAllUsesWith 262 SmallPtrSet<Constant *, 8> LocalVarsSet; 263 for (GlobalVariable *LocalVar : LocalVars) 264 LocalVarsSet.insert(cast<Constant>(LocalVar->stripPointerCasts())); 265 266 removeFromUsedLists( 267 M, [&LocalVarsSet](Constant *C) { return LocalVarsSet.count(C); }); 268 269 for (GlobalVariable *LocalVar : LocalVars) 270 LocalVar->removeDeadConstantUsers(); 271 } 272 273 static void markUsedByKernel(Function *Func, GlobalVariable *SGV) { 274 // The llvm.amdgcn.module.lds instance is implicitly used by all kernels 275 // that might call a function which accesses a field within it. This is 276 // presently approximated to 'all kernels' if there are any such functions 277 // in the module. This implicit use is redefined as an explicit use here so 278 // that later passes, specifically PromoteAlloca, account for the required 279 // memory without any knowledge of this transform. 280 281 // An operand bundle on llvm.donothing works because the call instruction 282 // survives until after the last pass that needs to account for LDS. It is 283 // better than inline asm as the latter survives until the end of codegen. A 284 // totally robust solution would be a function with the same semantics as 285 // llvm.donothing that takes a pointer to the instance and is lowered to a 286 // no-op after LDS is allocated, but that is not presently necessary. 287 288 // This intrinsic is eliminated shortly before instruction selection. It 289 // does not suffice to indicate to ISel that a given global which is not 290 // immediately used by the kernel must still be allocated by it. An 291 // equivalent target specific intrinsic which lasts until immediately after 292 // codegen would suffice for that, but one would still need to ensure that 293 // the variables are allocated in the anticpated order. 294 IRBuilder<> Builder(Func->getEntryBlock().getFirstNonPHI()); 295 296 Function *Decl = 297 Intrinsic::getDeclaration(Func->getParent(), Intrinsic::donothing, {}); 298 299 Value *UseInstance[1] = { 300 Builder.CreateConstInBoundsGEP1_32(SGV->getValueType(), SGV, 0)}; 301 302 Builder.CreateCall( 303 Decl, {}, {OperandBundleDefT<Value *>("ExplicitUse", UseInstance)}); 304 } 305 306 static bool eliminateConstantExprUsesOfLDSFromAllInstructions(Module &M) { 307 // Constants are uniqued within LLVM. A ConstantExpr referring to a LDS 308 // global may have uses from multiple different functions as a result. 309 // This pass specialises LDS variables with respect to the kernel that 310 // allocates them. 311 312 // This is semantically equivalent to (the unimplemented as slow): 313 // for (auto &F : M.functions()) 314 // for (auto &BB : F) 315 // for (auto &I : BB) 316 // for (Use &Op : I.operands()) 317 // if (constantExprUsesLDS(Op)) 318 // replaceConstantExprInFunction(I, Op); 319 320 SmallVector<Constant *> LDSGlobals; 321 for (auto &GV : M.globals()) 322 if (AMDGPU::isLDSVariableToLower(GV)) 323 LDSGlobals.push_back(&GV); 324 325 return convertUsersOfConstantsToInstructions(LDSGlobals); 326 } 327 328 public: 329 static char ID; 330 331 AMDGPULowerModuleLDS() : ModulePass(ID) { 332 initializeAMDGPULowerModuleLDSPass(*PassRegistry::getPassRegistry()); 333 } 334 335 using FunctionVariableMap = DenseMap<Function *, DenseSet<GlobalVariable *>>; 336 337 using VariableFunctionMap = DenseMap<GlobalVariable *, DenseSet<Function *>>; 338 339 static void getUsesOfLDSByFunction(CallGraph const &CG, Module &M, 340 FunctionVariableMap &kernels, 341 FunctionVariableMap &functions) { 342 343 // Get uses from the current function, excluding uses by called functions 344 // Two output variables to avoid walking the globals list twice 345 for (auto &GV : M.globals()) { 346 if (!AMDGPU::isLDSVariableToLower(GV)) { 347 continue; 348 } 349 350 if (GV.isAbsoluteSymbolRef()) { 351 report_fatal_error( 352 "LDS variables with absolute addresses are unimplemented."); 353 } 354 355 for (User *V : GV.users()) { 356 if (auto *I = dyn_cast<Instruction>(V)) { 357 Function *F = I->getFunction(); 358 if (isKernelLDS(F)) { 359 kernels[F].insert(&GV); 360 } else { 361 functions[F].insert(&GV); 362 } 363 } 364 } 365 } 366 } 367 368 struct LDSUsesInfoTy { 369 FunctionVariableMap direct_access; 370 FunctionVariableMap indirect_access; 371 }; 372 373 static LDSUsesInfoTy getTransitiveUsesOfLDS(CallGraph const &CG, Module &M) { 374 375 FunctionVariableMap direct_map_kernel; 376 FunctionVariableMap direct_map_function; 377 getUsesOfLDSByFunction(CG, M, direct_map_kernel, direct_map_function); 378 379 // Collect variables that are used by functions whose address has escaped 380 DenseSet<GlobalVariable *> VariablesReachableThroughFunctionPointer; 381 for (Function &F : M.functions()) { 382 if (!isKernelLDS(&F)) 383 if (F.hasAddressTaken(nullptr, 384 /* IgnoreCallbackUses */ false, 385 /* IgnoreAssumeLikeCalls */ false, 386 /* IgnoreLLVMUsed */ true, 387 /* IgnoreArcAttachedCall */ false)) { 388 set_union(VariablesReachableThroughFunctionPointer, 389 direct_map_function[&F]); 390 } 391 } 392 393 auto functionMakesUnknownCall = [&](const Function *F) -> bool { 394 assert(!F->isDeclaration()); 395 for (const CallGraphNode::CallRecord &R : *CG[F]) { 396 if (!R.second->getFunction()) { 397 return true; 398 } 399 } 400 return false; 401 }; 402 403 // Work out which variables are reachable through function calls 404 FunctionVariableMap transitive_map_function = direct_map_function; 405 406 // If the function makes any unknown call, assume the worst case that it can 407 // access all variables accessed by functions whose address escaped 408 for (Function &F : M.functions()) { 409 if (!F.isDeclaration() && functionMakesUnknownCall(&F)) { 410 if (!isKernelLDS(&F)) { 411 set_union(transitive_map_function[&F], 412 VariablesReachableThroughFunctionPointer); 413 } 414 } 415 } 416 417 // Direct implementation of collecting all variables reachable from each 418 // function 419 for (Function &Func : M.functions()) { 420 if (Func.isDeclaration() || isKernelLDS(&Func)) 421 continue; 422 423 DenseSet<Function *> seen; // catches cycles 424 SmallVector<Function *, 4> wip{&Func}; 425 426 while (!wip.empty()) { 427 Function *F = wip.pop_back_val(); 428 429 // Can accelerate this by referring to transitive map for functions that 430 // have already been computed, with more care than this 431 set_union(transitive_map_function[&Func], direct_map_function[F]); 432 433 for (const CallGraphNode::CallRecord &R : *CG[F]) { 434 Function *ith = R.second->getFunction(); 435 if (ith) { 436 if (!seen.contains(ith)) { 437 seen.insert(ith); 438 wip.push_back(ith); 439 } 440 } 441 } 442 } 443 } 444 445 // direct_map_kernel lists which variables are used by the kernel 446 // find the variables which are used through a function call 447 FunctionVariableMap indirect_map_kernel; 448 449 for (Function &Func : M.functions()) { 450 if (Func.isDeclaration() || !isKernelLDS(&Func)) 451 continue; 452 453 for (const CallGraphNode::CallRecord &R : *CG[&Func]) { 454 Function *ith = R.second->getFunction(); 455 if (ith) { 456 set_union(indirect_map_kernel[&Func], transitive_map_function[ith]); 457 } else { 458 set_union(indirect_map_kernel[&Func], 459 VariablesReachableThroughFunctionPointer); 460 } 461 } 462 } 463 464 return {std::move(direct_map_kernel), std::move(indirect_map_kernel)}; 465 } 466 467 struct LDSVariableReplacement { 468 GlobalVariable *SGV = nullptr; 469 DenseMap<GlobalVariable *, Constant *> LDSVarsToConstantGEP; 470 }; 471 472 // remap from lds global to a constantexpr gep to where it has been moved to 473 // for each kernel 474 // an array with an element for each kernel containing where the corresponding 475 // variable was remapped to 476 477 static Constant *getAddressesOfVariablesInKernel( 478 LLVMContext &Ctx, ArrayRef<GlobalVariable *> Variables, 479 const DenseMap<GlobalVariable *, Constant *> &LDSVarsToConstantGEP) { 480 // Create a ConstantArray containing the address of each Variable within the 481 // kernel corresponding to LDSVarsToConstantGEP, or poison if that kernel 482 // does not allocate it 483 // TODO: Drop the ptrtoint conversion 484 485 Type *I32 = Type::getInt32Ty(Ctx); 486 487 ArrayType *KernelOffsetsType = ArrayType::get(I32, Variables.size()); 488 489 SmallVector<Constant *> Elements; 490 for (size_t i = 0; i < Variables.size(); i++) { 491 GlobalVariable *GV = Variables[i]; 492 auto ConstantGepIt = LDSVarsToConstantGEP.find(GV); 493 if (ConstantGepIt != LDSVarsToConstantGEP.end()) { 494 auto elt = ConstantExpr::getPtrToInt(ConstantGepIt->second, I32); 495 Elements.push_back(elt); 496 } else { 497 Elements.push_back(PoisonValue::get(I32)); 498 } 499 } 500 return ConstantArray::get(KernelOffsetsType, Elements); 501 } 502 503 static GlobalVariable *buildLookupTable( 504 Module &M, ArrayRef<GlobalVariable *> Variables, 505 ArrayRef<Function *> kernels, 506 DenseMap<Function *, LDSVariableReplacement> &KernelToReplacement) { 507 if (Variables.empty()) { 508 return nullptr; 509 } 510 LLVMContext &Ctx = M.getContext(); 511 512 const size_t NumberVariables = Variables.size(); 513 const size_t NumberKernels = kernels.size(); 514 515 ArrayType *KernelOffsetsType = 516 ArrayType::get(Type::getInt32Ty(Ctx), NumberVariables); 517 518 ArrayType *AllKernelsOffsetsType = 519 ArrayType::get(KernelOffsetsType, NumberKernels); 520 521 Constant *Missing = PoisonValue::get(KernelOffsetsType); 522 std::vector<Constant *> overallConstantExprElts(NumberKernels); 523 for (size_t i = 0; i < NumberKernels; i++) { 524 auto Replacement = KernelToReplacement.find(kernels[i]); 525 overallConstantExprElts[i] = 526 (Replacement == KernelToReplacement.end()) 527 ? Missing 528 : getAddressesOfVariablesInKernel( 529 Ctx, Variables, Replacement->second.LDSVarsToConstantGEP); 530 } 531 532 Constant *init = 533 ConstantArray::get(AllKernelsOffsetsType, overallConstantExprElts); 534 535 return new GlobalVariable( 536 M, AllKernelsOffsetsType, true, GlobalValue::InternalLinkage, init, 537 "llvm.amdgcn.lds.offset.table", nullptr, GlobalValue::NotThreadLocal, 538 AMDGPUAS::CONSTANT_ADDRESS); 539 } 540 541 void replaceUseWithTableLookup(Module &M, IRBuilder<> &Builder, 542 GlobalVariable *LookupTable, 543 GlobalVariable *GV, Use &U, 544 Value *OptionalIndex) { 545 // Table is a constant array of the same length as OrderedKernels 546 LLVMContext &Ctx = M.getContext(); 547 Type *I32 = Type::getInt32Ty(Ctx); 548 auto *I = cast<Instruction>(U.getUser()); 549 550 Value *tableKernelIndex = getTableLookupKernelIndex(M, I->getFunction()); 551 552 if (auto *Phi = dyn_cast<PHINode>(I)) { 553 BasicBlock *BB = Phi->getIncomingBlock(U); 554 Builder.SetInsertPoint(&(*(BB->getFirstInsertionPt()))); 555 } else { 556 Builder.SetInsertPoint(I); 557 } 558 559 SmallVector<Value *, 3> GEPIdx = { 560 ConstantInt::get(I32, 0), 561 tableKernelIndex, 562 }; 563 if (OptionalIndex) 564 GEPIdx.push_back(OptionalIndex); 565 566 Value *Address = Builder.CreateInBoundsGEP( 567 LookupTable->getValueType(), LookupTable, GEPIdx, GV->getName()); 568 569 Value *loaded = Builder.CreateLoad(I32, Address); 570 571 Value *replacement = 572 Builder.CreateIntToPtr(loaded, GV->getType(), GV->getName()); 573 574 U.set(replacement); 575 } 576 577 void replaceUsesInInstructionsWithTableLookup( 578 Module &M, ArrayRef<GlobalVariable *> ModuleScopeVariables, 579 GlobalVariable *LookupTable) { 580 581 LLVMContext &Ctx = M.getContext(); 582 IRBuilder<> Builder(Ctx); 583 Type *I32 = Type::getInt32Ty(Ctx); 584 585 for (size_t Index = 0; Index < ModuleScopeVariables.size(); Index++) { 586 auto *GV = ModuleScopeVariables[Index]; 587 588 for (Use &U : make_early_inc_range(GV->uses())) { 589 auto *I = dyn_cast<Instruction>(U.getUser()); 590 if (!I) 591 continue; 592 593 replaceUseWithTableLookup(M, Builder, LookupTable, GV, U, 594 ConstantInt::get(I32, Index)); 595 } 596 } 597 } 598 599 static DenseSet<Function *> kernelsThatIndirectlyAccessAnyOfPassedVariables( 600 Module &M, LDSUsesInfoTy &LDSUsesInfo, 601 DenseSet<GlobalVariable *> const &VariableSet) { 602 603 DenseSet<Function *> KernelSet; 604 605 if (VariableSet.empty()) 606 return KernelSet; 607 608 for (Function &Func : M.functions()) { 609 if (Func.isDeclaration() || !isKernelLDS(&Func)) 610 continue; 611 for (GlobalVariable *GV : LDSUsesInfo.indirect_access[&Func]) { 612 if (VariableSet.contains(GV)) { 613 KernelSet.insert(&Func); 614 break; 615 } 616 } 617 } 618 619 return KernelSet; 620 } 621 622 static GlobalVariable * 623 chooseBestVariableForModuleStrategy(const DataLayout &DL, 624 VariableFunctionMap &LDSVars) { 625 // Find the global variable with the most indirect uses from kernels 626 627 struct CandidateTy { 628 GlobalVariable *GV = nullptr; 629 size_t UserCount = 0; 630 size_t Size = 0; 631 632 CandidateTy() = default; 633 634 CandidateTy(GlobalVariable *GV, uint64_t UserCount, uint64_t AllocSize) 635 : GV(GV), UserCount(UserCount), Size(AllocSize) {} 636 637 bool operator<(const CandidateTy &Other) const { 638 // Fewer users makes module scope variable less attractive 639 if (UserCount < Other.UserCount) { 640 return true; 641 } 642 if (UserCount > Other.UserCount) { 643 return false; 644 } 645 646 // Bigger makes module scope variable less attractive 647 if (Size < Other.Size) { 648 return false; 649 } 650 651 if (Size > Other.Size) { 652 return true; 653 } 654 655 // Arbitrary but consistent 656 return GV->getName() < Other.GV->getName(); 657 } 658 }; 659 660 CandidateTy MostUsed; 661 662 for (auto &K : LDSVars) { 663 GlobalVariable *GV = K.first; 664 if (K.second.size() <= 1) { 665 // A variable reachable by only one kernel is best lowered with kernel 666 // strategy 667 continue; 668 } 669 CandidateTy Candidate( 670 GV, K.second.size(), 671 DL.getTypeAllocSize(GV->getValueType()).getFixedValue()); 672 if (MostUsed < Candidate) 673 MostUsed = Candidate; 674 } 675 676 return MostUsed.GV; 677 } 678 679 static void recordLDSAbsoluteAddress(Module *M, GlobalVariable *GV, 680 uint32_t Address) { 681 // Write the specified address into metadata where it can be retrieved by 682 // the assembler. Format is a half open range, [Address Address+1) 683 LLVMContext &Ctx = M->getContext(); 684 auto *IntTy = 685 M->getDataLayout().getIntPtrType(Ctx, AMDGPUAS::LOCAL_ADDRESS); 686 auto *MinC = ConstantAsMetadata::get(ConstantInt::get(IntTy, Address)); 687 auto *MaxC = ConstantAsMetadata::get(ConstantInt::get(IntTy, Address + 1)); 688 GV->setMetadata(LLVMContext::MD_absolute_symbol, 689 MDNode::get(Ctx, {MinC, MaxC})); 690 } 691 692 DenseMap<Function *, Value *> tableKernelIndexCache; 693 Value *getTableLookupKernelIndex(Module &M, Function *F) { 694 // Accesses from a function use the amdgcn_lds_kernel_id intrinsic which 695 // lowers to a read from a live in register. Emit it once in the entry 696 // block to spare deduplicating it later. 697 auto [It, Inserted] = tableKernelIndexCache.try_emplace(F); 698 if (Inserted) { 699 Function *Decl = 700 Intrinsic::getDeclaration(&M, Intrinsic::amdgcn_lds_kernel_id, {}); 701 702 auto InsertAt = F->getEntryBlock().getFirstNonPHIOrDbgOrAlloca(); 703 IRBuilder<> Builder(&*InsertAt); 704 705 It->second = Builder.CreateCall(Decl, {}); 706 } 707 708 return It->second; 709 } 710 711 static std::vector<Function *> assignLDSKernelIDToEachKernel( 712 Module *M, DenseSet<Function *> const &KernelsThatAllocateTableLDS, 713 DenseSet<Function *> const &KernelsThatIndirectlyAllocateDynamicLDS) { 714 // Associate kernels in the set with an arbirary but reproducible order and 715 // annotate them with that order in metadata. This metadata is recognised by 716 // the backend and lowered to a SGPR which can be read from using 717 // amdgcn_lds_kernel_id. 718 719 std::vector<Function *> OrderedKernels; 720 if (!KernelsThatAllocateTableLDS.empty() || 721 !KernelsThatIndirectlyAllocateDynamicLDS.empty()) { 722 723 for (Function &Func : M->functions()) { 724 if (Func.isDeclaration()) 725 continue; 726 if (!isKernelLDS(&Func)) 727 continue; 728 729 if (KernelsThatAllocateTableLDS.contains(&Func) || 730 KernelsThatIndirectlyAllocateDynamicLDS.contains(&Func)) { 731 assert(Func.hasName()); // else fatal error earlier 732 OrderedKernels.push_back(&Func); 733 } 734 } 735 736 // Put them in an arbitrary but reproducible order 737 OrderedKernels = sortByName(std::move(OrderedKernels)); 738 739 // Annotate the kernels with their order in this vector 740 LLVMContext &Ctx = M->getContext(); 741 IRBuilder<> Builder(Ctx); 742 743 if (OrderedKernels.size() > UINT32_MAX) { 744 // 32 bit keeps it in one SGPR. > 2**32 kernels won't fit on the GPU 745 report_fatal_error("Unimplemented LDS lowering for > 2**32 kernels"); 746 } 747 748 for (size_t i = 0; i < OrderedKernels.size(); i++) { 749 Metadata *AttrMDArgs[1] = { 750 ConstantAsMetadata::get(Builder.getInt32(i)), 751 }; 752 OrderedKernels[i]->setMetadata("llvm.amdgcn.lds.kernel.id", 753 MDNode::get(Ctx, AttrMDArgs)); 754 } 755 } 756 return OrderedKernels; 757 } 758 759 static void partitionVariablesIntoIndirectStrategies( 760 Module &M, LDSUsesInfoTy const &LDSUsesInfo, 761 VariableFunctionMap &LDSToKernelsThatNeedToAccessItIndirectly, 762 DenseSet<GlobalVariable *> &ModuleScopeVariables, 763 DenseSet<GlobalVariable *> &TableLookupVariables, 764 DenseSet<GlobalVariable *> &KernelAccessVariables, 765 DenseSet<GlobalVariable *> &DynamicVariables) { 766 767 GlobalVariable *HybridModuleRoot = 768 LoweringKindLoc != LoweringKind::hybrid 769 ? nullptr 770 : chooseBestVariableForModuleStrategy( 771 M.getDataLayout(), LDSToKernelsThatNeedToAccessItIndirectly); 772 773 DenseSet<Function *> const EmptySet; 774 DenseSet<Function *> const &HybridModuleRootKernels = 775 HybridModuleRoot 776 ? LDSToKernelsThatNeedToAccessItIndirectly[HybridModuleRoot] 777 : EmptySet; 778 779 for (auto &K : LDSToKernelsThatNeedToAccessItIndirectly) { 780 // Each iteration of this loop assigns exactly one global variable to 781 // exactly one of the implementation strategies. 782 783 GlobalVariable *GV = K.first; 784 assert(AMDGPU::isLDSVariableToLower(*GV)); 785 assert(K.second.size() != 0); 786 787 if (AMDGPU::isDynamicLDS(*GV)) { 788 DynamicVariables.insert(GV); 789 continue; 790 } 791 792 switch (LoweringKindLoc) { 793 case LoweringKind::module: 794 ModuleScopeVariables.insert(GV); 795 break; 796 797 case LoweringKind::table: 798 TableLookupVariables.insert(GV); 799 break; 800 801 case LoweringKind::kernel: 802 if (K.second.size() == 1) { 803 KernelAccessVariables.insert(GV); 804 } else { 805 report_fatal_error( 806 "cannot lower LDS '" + GV->getName() + 807 "' to kernel access as it is reachable from multiple kernels"); 808 } 809 break; 810 811 case LoweringKind::hybrid: { 812 if (GV == HybridModuleRoot) { 813 assert(K.second.size() != 1); 814 ModuleScopeVariables.insert(GV); 815 } else if (K.second.size() == 1) { 816 KernelAccessVariables.insert(GV); 817 } else if (set_is_subset(K.second, HybridModuleRootKernels)) { 818 ModuleScopeVariables.insert(GV); 819 } else { 820 TableLookupVariables.insert(GV); 821 } 822 break; 823 } 824 } 825 } 826 827 // All LDS variables accessed indirectly have now been partitioned into 828 // the distinct lowering strategies. 829 assert(ModuleScopeVariables.size() + TableLookupVariables.size() + 830 KernelAccessVariables.size() + DynamicVariables.size() == 831 LDSToKernelsThatNeedToAccessItIndirectly.size()); 832 } 833 834 static GlobalVariable *lowerModuleScopeStructVariables( 835 Module &M, DenseSet<GlobalVariable *> const &ModuleScopeVariables, 836 DenseSet<Function *> const &KernelsThatAllocateModuleLDS) { 837 // Create a struct to hold the ModuleScopeVariables 838 // Replace all uses of those variables from non-kernel functions with the 839 // new struct instance Replace only the uses from kernel functions that will 840 // allocate this instance. That is a space optimisation - kernels that use a 841 // subset of the module scope struct and do not need to allocate it for 842 // indirect calls will only allocate the subset they use (they do so as part 843 // of the per-kernel lowering). 844 if (ModuleScopeVariables.empty()) { 845 return nullptr; 846 } 847 848 LLVMContext &Ctx = M.getContext(); 849 850 LDSVariableReplacement ModuleScopeReplacement = 851 createLDSVariableReplacement(M, "llvm.amdgcn.module.lds", 852 ModuleScopeVariables); 853 854 appendToCompilerUsed(M, {static_cast<GlobalValue *>( 855 ConstantExpr::getPointerBitCastOrAddrSpaceCast( 856 cast<Constant>(ModuleScopeReplacement.SGV), 857 Type::getInt8PtrTy(Ctx)))}); 858 859 // module.lds will be allocated at zero in any kernel that allocates it 860 recordLDSAbsoluteAddress(&M, ModuleScopeReplacement.SGV, 0); 861 862 // historic 863 removeLocalVarsFromUsedLists(M, ModuleScopeVariables); 864 865 // Replace all uses of module scope variable from non-kernel functions 866 replaceLDSVariablesWithStruct( 867 M, ModuleScopeVariables, ModuleScopeReplacement, [&](Use &U) { 868 Instruction *I = dyn_cast<Instruction>(U.getUser()); 869 if (!I) { 870 return false; 871 } 872 Function *F = I->getFunction(); 873 return !isKernelLDS(F); 874 }); 875 876 // Replace uses of module scope variable from kernel functions that 877 // allocate the module scope variable, otherwise leave them unchanged 878 // Record on each kernel whether the module scope global is used by it 879 880 for (Function &Func : M.functions()) { 881 if (Func.isDeclaration() || !isKernelLDS(&Func)) 882 continue; 883 884 if (KernelsThatAllocateModuleLDS.contains(&Func)) { 885 replaceLDSVariablesWithStruct( 886 M, ModuleScopeVariables, ModuleScopeReplacement, [&](Use &U) { 887 Instruction *I = dyn_cast<Instruction>(U.getUser()); 888 if (!I) { 889 return false; 890 } 891 Function *F = I->getFunction(); 892 return F == &Func; 893 }); 894 895 markUsedByKernel(&Func, ModuleScopeReplacement.SGV); 896 } 897 } 898 899 return ModuleScopeReplacement.SGV; 900 } 901 902 static DenseMap<Function *, LDSVariableReplacement> 903 lowerKernelScopeStructVariables( 904 Module &M, LDSUsesInfoTy &LDSUsesInfo, 905 DenseSet<GlobalVariable *> const &ModuleScopeVariables, 906 DenseSet<Function *> const &KernelsThatAllocateModuleLDS, 907 GlobalVariable *MaybeModuleScopeStruct) { 908 909 // Create a struct for each kernel for the non-module-scope variables. 910 911 DenseMap<Function *, LDSVariableReplacement> KernelToReplacement; 912 for (Function &Func : M.functions()) { 913 if (Func.isDeclaration() || !isKernelLDS(&Func)) 914 continue; 915 916 DenseSet<GlobalVariable *> KernelUsedVariables; 917 // Allocating variables that are used directly in this struct to get 918 // alignment aware allocation and predictable frame size. 919 for (auto &v : LDSUsesInfo.direct_access[&Func]) { 920 if (!AMDGPU::isDynamicLDS(*v)) { 921 KernelUsedVariables.insert(v); 922 } 923 } 924 925 // Allocating variables that are accessed indirectly so that a lookup of 926 // this struct instance can find them from nested functions. 927 for (auto &v : LDSUsesInfo.indirect_access[&Func]) { 928 if (!AMDGPU::isDynamicLDS(*v)) { 929 KernelUsedVariables.insert(v); 930 } 931 } 932 933 // Variables allocated in module lds must all resolve to that struct, 934 // not to the per-kernel instance. 935 if (KernelsThatAllocateModuleLDS.contains(&Func)) { 936 for (GlobalVariable *v : ModuleScopeVariables) { 937 KernelUsedVariables.erase(v); 938 } 939 } 940 941 if (KernelUsedVariables.empty()) { 942 // Either used no LDS, or the LDS it used was all in the module struct 943 // or dynamically sized 944 continue; 945 } 946 947 // The association between kernel function and LDS struct is done by 948 // symbol name, which only works if the function in question has a 949 // name This is not expected to be a problem in practice as kernels 950 // are called by name making anonymous ones (which are named by the 951 // backend) difficult to use. This does mean that llvm test cases need 952 // to name the kernels. 953 if (!Func.hasName()) { 954 report_fatal_error("Anonymous kernels cannot use LDS variables"); 955 } 956 957 std::string VarName = 958 (Twine("llvm.amdgcn.kernel.") + Func.getName() + ".lds").str(); 959 960 auto Replacement = 961 createLDSVariableReplacement(M, VarName, KernelUsedVariables); 962 963 // If any indirect uses, create a direct use to ensure allocation 964 // TODO: Simpler to unconditionally mark used but that regresses 965 // codegen in test/CodeGen/AMDGPU/noclobber-barrier.ll 966 auto Accesses = LDSUsesInfo.indirect_access.find(&Func); 967 if ((Accesses != LDSUsesInfo.indirect_access.end()) && 968 !Accesses->second.empty()) 969 markUsedByKernel(&Func, Replacement.SGV); 970 971 // remove preserves existing codegen 972 removeLocalVarsFromUsedLists(M, KernelUsedVariables); 973 KernelToReplacement[&Func] = Replacement; 974 975 // Rewrite uses within kernel to the new struct 976 replaceLDSVariablesWithStruct( 977 M, KernelUsedVariables, Replacement, [&Func](Use &U) { 978 Instruction *I = dyn_cast<Instruction>(U.getUser()); 979 return I && I->getFunction() == &Func; 980 }); 981 } 982 return KernelToReplacement; 983 } 984 985 static GlobalVariable * 986 buildRepresentativeDynamicLDSInstance(Module &M, LDSUsesInfoTy &LDSUsesInfo, 987 Function *func) { 988 // Create a dynamic lds variable with a name associated with the passed 989 // function that has the maximum alignment of any dynamic lds variable 990 // reachable from this kernel. Dynamic LDS is allocated after the static LDS 991 // allocation, possibly after alignment padding. The representative variable 992 // created here has the maximum alignment of any other dynamic variable 993 // reachable by that kernel. All dynamic LDS variables are allocated at the 994 // same address in each kernel in order to provide the documented aliasing 995 // semantics. Setting the alignment here allows this IR pass to accurately 996 // predict the exact constant at which it will be allocated. 997 998 assert(isKernelLDS(func)); 999 1000 LLVMContext &Ctx = M.getContext(); 1001 const DataLayout &DL = M.getDataLayout(); 1002 Align MaxDynamicAlignment(1); 1003 1004 auto UpdateMaxAlignment = [&MaxDynamicAlignment, &DL](GlobalVariable *GV) { 1005 if (AMDGPU::isDynamicLDS(*GV)) { 1006 MaxDynamicAlignment = 1007 std::max(MaxDynamicAlignment, AMDGPU::getAlign(DL, GV)); 1008 } 1009 }; 1010 1011 for (GlobalVariable *GV : LDSUsesInfo.indirect_access[func]) { 1012 UpdateMaxAlignment(GV); 1013 } 1014 1015 for (GlobalVariable *GV : LDSUsesInfo.direct_access[func]) { 1016 UpdateMaxAlignment(GV); 1017 } 1018 1019 assert(func->hasName()); // Checked by caller 1020 auto emptyCharArray = ArrayType::get(Type::getInt8Ty(Ctx), 0); 1021 GlobalVariable *N = new GlobalVariable( 1022 M, emptyCharArray, false, GlobalValue::ExternalLinkage, nullptr, 1023 Twine("llvm.amdgcn." + func->getName() + ".dynlds"), nullptr, GlobalValue::NotThreadLocal, AMDGPUAS::LOCAL_ADDRESS, 1024 false); 1025 N->setAlignment(MaxDynamicAlignment); 1026 1027 assert(AMDGPU::isDynamicLDS(*N)); 1028 return N; 1029 } 1030 1031 DenseMap<Function *, GlobalVariable *> lowerDynamicLDSVariables( 1032 Module &M, LDSUsesInfoTy &LDSUsesInfo, 1033 DenseSet<Function *> const &KernelsThatIndirectlyAllocateDynamicLDS, 1034 DenseSet<GlobalVariable *> const &DynamicVariables, 1035 std::vector<Function *> const &OrderedKernels) { 1036 DenseMap<Function *, GlobalVariable *> KernelToCreatedDynamicLDS; 1037 if (!KernelsThatIndirectlyAllocateDynamicLDS.empty()) { 1038 LLVMContext &Ctx = M.getContext(); 1039 IRBuilder<> Builder(Ctx); 1040 Type *I32 = Type::getInt32Ty(Ctx); 1041 1042 std::vector<Constant *> newDynamicLDS; 1043 1044 // Table is built in the same order as OrderedKernels 1045 for (auto &func : OrderedKernels) { 1046 1047 if (KernelsThatIndirectlyAllocateDynamicLDS.contains(func)) { 1048 assert(isKernelLDS(func)); 1049 if (!func->hasName()) { 1050 report_fatal_error("Anonymous kernels cannot use LDS variables"); 1051 } 1052 1053 GlobalVariable *N = 1054 buildRepresentativeDynamicLDSInstance(M, LDSUsesInfo, func); 1055 1056 KernelToCreatedDynamicLDS[func] = N; 1057 1058 markUsedByKernel(func, N); 1059 1060 auto emptyCharArray = ArrayType::get(Type::getInt8Ty(Ctx), 0); 1061 auto GEP = ConstantExpr::getGetElementPtr( 1062 emptyCharArray, N, ConstantInt::get(I32, 0), true); 1063 newDynamicLDS.push_back(ConstantExpr::getPtrToInt(GEP, I32)); 1064 } else { 1065 newDynamicLDS.push_back(PoisonValue::get(I32)); 1066 } 1067 } 1068 assert(OrderedKernels.size() == newDynamicLDS.size()); 1069 1070 ArrayType *t = ArrayType::get(I32, newDynamicLDS.size()); 1071 Constant *init = ConstantArray::get(t, newDynamicLDS); 1072 GlobalVariable *table = new GlobalVariable( 1073 M, t, true, GlobalValue::InternalLinkage, init, 1074 "llvm.amdgcn.dynlds.offset.table", nullptr, 1075 GlobalValue::NotThreadLocal, AMDGPUAS::CONSTANT_ADDRESS); 1076 1077 for (GlobalVariable *GV : DynamicVariables) { 1078 for (Use &U : make_early_inc_range(GV->uses())) { 1079 auto *I = dyn_cast<Instruction>(U.getUser()); 1080 if (!I) 1081 continue; 1082 if (isKernelLDS(I->getFunction())) 1083 continue; 1084 1085 replaceUseWithTableLookup(M, Builder, table, GV, U, nullptr); 1086 } 1087 } 1088 } 1089 return KernelToCreatedDynamicLDS; 1090 } 1091 1092 bool runOnModule(Module &M) override { 1093 CallGraph CG = CallGraph(M); 1094 bool Changed = superAlignLDSGlobals(M); 1095 1096 Changed |= eliminateConstantExprUsesOfLDSFromAllInstructions(M); 1097 1098 Changed = true; // todo: narrow this down 1099 1100 // For each kernel, what variables does it access directly or through 1101 // callees 1102 LDSUsesInfoTy LDSUsesInfo = getTransitiveUsesOfLDS(CG, M); 1103 1104 // For each variable accessed through callees, which kernels access it 1105 VariableFunctionMap LDSToKernelsThatNeedToAccessItIndirectly; 1106 for (auto &K : LDSUsesInfo.indirect_access) { 1107 Function *F = K.first; 1108 assert(isKernelLDS(F)); 1109 for (GlobalVariable *GV : K.second) { 1110 LDSToKernelsThatNeedToAccessItIndirectly[GV].insert(F); 1111 } 1112 } 1113 1114 // Partition variables accessed indirectly into the different strategies 1115 DenseSet<GlobalVariable *> ModuleScopeVariables; 1116 DenseSet<GlobalVariable *> TableLookupVariables; 1117 DenseSet<GlobalVariable *> KernelAccessVariables; 1118 DenseSet<GlobalVariable *> DynamicVariables; 1119 partitionVariablesIntoIndirectStrategies( 1120 M, LDSUsesInfo, LDSToKernelsThatNeedToAccessItIndirectly, 1121 ModuleScopeVariables, TableLookupVariables, KernelAccessVariables, 1122 DynamicVariables); 1123 1124 // If the kernel accesses a variable that is going to be stored in the 1125 // module instance through a call then that kernel needs to allocate the 1126 // module instance 1127 const DenseSet<Function *> KernelsThatAllocateModuleLDS = 1128 kernelsThatIndirectlyAccessAnyOfPassedVariables(M, LDSUsesInfo, 1129 ModuleScopeVariables); 1130 const DenseSet<Function *> KernelsThatAllocateTableLDS = 1131 kernelsThatIndirectlyAccessAnyOfPassedVariables(M, LDSUsesInfo, 1132 TableLookupVariables); 1133 1134 const DenseSet<Function *> KernelsThatIndirectlyAllocateDynamicLDS = 1135 kernelsThatIndirectlyAccessAnyOfPassedVariables(M, LDSUsesInfo, 1136 DynamicVariables); 1137 1138 GlobalVariable *MaybeModuleScopeStruct = lowerModuleScopeStructVariables( 1139 M, ModuleScopeVariables, KernelsThatAllocateModuleLDS); 1140 1141 DenseMap<Function *, LDSVariableReplacement> KernelToReplacement = 1142 lowerKernelScopeStructVariables(M, LDSUsesInfo, ModuleScopeVariables, 1143 KernelsThatAllocateModuleLDS, 1144 MaybeModuleScopeStruct); 1145 1146 // Lower zero cost accesses to the kernel instances just created 1147 for (auto &GV : KernelAccessVariables) { 1148 auto &funcs = LDSToKernelsThatNeedToAccessItIndirectly[GV]; 1149 assert(funcs.size() == 1); // Only one kernel can access it 1150 LDSVariableReplacement Replacement = 1151 KernelToReplacement[*(funcs.begin())]; 1152 1153 DenseSet<GlobalVariable *> Vec; 1154 Vec.insert(GV); 1155 1156 replaceLDSVariablesWithStruct(M, Vec, Replacement, [](Use &U) { 1157 return isa<Instruction>(U.getUser()); 1158 }); 1159 } 1160 1161 // The ith element of this vector is kernel id i 1162 std::vector<Function *> OrderedKernels = 1163 assignLDSKernelIDToEachKernel(&M, KernelsThatAllocateTableLDS, 1164 KernelsThatIndirectlyAllocateDynamicLDS); 1165 1166 if (!KernelsThatAllocateTableLDS.empty()) { 1167 LLVMContext &Ctx = M.getContext(); 1168 IRBuilder<> Builder(Ctx); 1169 1170 // The order must be consistent between lookup table and accesses to 1171 // lookup table 1172 auto TableLookupVariablesOrdered = 1173 sortByName(std::vector<GlobalVariable *>(TableLookupVariables.begin(), 1174 TableLookupVariables.end())); 1175 1176 GlobalVariable *LookupTable = buildLookupTable( 1177 M, TableLookupVariablesOrdered, OrderedKernels, KernelToReplacement); 1178 replaceUsesInInstructionsWithTableLookup(M, TableLookupVariablesOrdered, 1179 LookupTable); 1180 } 1181 1182 DenseMap<Function *, GlobalVariable *> KernelToCreatedDynamicLDS = 1183 lowerDynamicLDSVariables(M, LDSUsesInfo, 1184 KernelsThatIndirectlyAllocateDynamicLDS, 1185 DynamicVariables, OrderedKernels); 1186 1187 // All kernel frames have been allocated. Calculate and record the 1188 // addresses. 1189 { 1190 const DataLayout &DL = M.getDataLayout(); 1191 1192 for (Function &Func : M.functions()) { 1193 if (Func.isDeclaration() || !isKernelLDS(&Func)) 1194 continue; 1195 1196 // All three of these are optional. The first variable is allocated at 1197 // zero. They are allocated by AMDGPUMachineFunction as one block. 1198 // Layout: 1199 //{ 1200 // module.lds 1201 // alignment padding 1202 // kernel instance 1203 // alignment padding 1204 // dynamic lds variables 1205 //} 1206 1207 const bool AllocateModuleScopeStruct = 1208 MaybeModuleScopeStruct && 1209 KernelsThatAllocateModuleLDS.contains(&Func); 1210 1211 auto Replacement = KernelToReplacement.find(&Func); 1212 const bool AllocateKernelScopeStruct = 1213 Replacement != KernelToReplacement.end(); 1214 1215 const bool AllocateDynamicVariable = 1216 KernelToCreatedDynamicLDS.contains(&Func); 1217 1218 uint32_t Offset = 0; 1219 1220 if (AllocateModuleScopeStruct) { 1221 // Allocated at zero, recorded once on construction, not once per 1222 // kernel 1223 Offset += DL.getTypeAllocSize(MaybeModuleScopeStruct->getValueType()); 1224 } 1225 1226 if (AllocateKernelScopeStruct) { 1227 GlobalVariable *KernelStruct = Replacement->second.SGV; 1228 Offset = alignTo(Offset, AMDGPU::getAlign(DL, KernelStruct)); 1229 recordLDSAbsoluteAddress(&M, KernelStruct, Offset); 1230 Offset += DL.getTypeAllocSize(KernelStruct->getValueType()); 1231 } 1232 1233 // If there is dynamic allocation, the alignment needed is included in 1234 // the static frame size. There may be no reference to the dynamic 1235 // variable in the kernel itself, so without including it here, that 1236 // alignment padding could be missed. 1237 if (AllocateDynamicVariable) { 1238 GlobalVariable *DynamicVariable = KernelToCreatedDynamicLDS[&Func]; 1239 Offset = alignTo(Offset, AMDGPU::getAlign(DL, DynamicVariable)); 1240 recordLDSAbsoluteAddress(&M, DynamicVariable, Offset); 1241 } 1242 1243 if (Offset != 0) { 1244 std::string Buffer; 1245 raw_string_ostream SS{Buffer}; 1246 SS << format("%u", Offset); 1247 1248 // Instead of explictly marking kernels that access dynamic variables 1249 // using special case metadata, annotate with min-lds == max-lds, i.e. 1250 // that there is no more space available for allocating more static 1251 // LDS variables. That is the right condition to prevent allocating 1252 // more variables which would collide with the addresses assigned to 1253 // dynamic variables. 1254 if (AllocateDynamicVariable) 1255 SS << format(",%u", Offset); 1256 1257 Func.addFnAttr("amdgpu-lds-size", Buffer); 1258 } 1259 } 1260 } 1261 1262 for (auto &GV : make_early_inc_range(M.globals())) 1263 if (AMDGPU::isLDSVariableToLower(GV)) { 1264 // probably want to remove from used lists 1265 GV.removeDeadConstantUsers(); 1266 if (GV.use_empty()) 1267 GV.eraseFromParent(); 1268 } 1269 1270 return Changed; 1271 } 1272 1273 private: 1274 // Increase the alignment of LDS globals if necessary to maximise the chance 1275 // that we can use aligned LDS instructions to access them. 1276 static bool superAlignLDSGlobals(Module &M) { 1277 const DataLayout &DL = M.getDataLayout(); 1278 bool Changed = false; 1279 if (!SuperAlignLDSGlobals) { 1280 return Changed; 1281 } 1282 1283 for (auto &GV : M.globals()) { 1284 if (GV.getType()->getPointerAddressSpace() != AMDGPUAS::LOCAL_ADDRESS) { 1285 // Only changing alignment of LDS variables 1286 continue; 1287 } 1288 if (!GV.hasInitializer()) { 1289 // cuda/hip extern __shared__ variable, leave alignment alone 1290 continue; 1291 } 1292 1293 Align Alignment = AMDGPU::getAlign(DL, &GV); 1294 TypeSize GVSize = DL.getTypeAllocSize(GV.getValueType()); 1295 1296 if (GVSize > 8) { 1297 // We might want to use a b96 or b128 load/store 1298 Alignment = std::max(Alignment, Align(16)); 1299 } else if (GVSize > 4) { 1300 // We might want to use a b64 load/store 1301 Alignment = std::max(Alignment, Align(8)); 1302 } else if (GVSize > 2) { 1303 // We might want to use a b32 load/store 1304 Alignment = std::max(Alignment, Align(4)); 1305 } else if (GVSize > 1) { 1306 // We might want to use a b16 load/store 1307 Alignment = std::max(Alignment, Align(2)); 1308 } 1309 1310 if (Alignment != AMDGPU::getAlign(DL, &GV)) { 1311 Changed = true; 1312 GV.setAlignment(Alignment); 1313 } 1314 } 1315 return Changed; 1316 } 1317 1318 static LDSVariableReplacement createLDSVariableReplacement( 1319 Module &M, std::string VarName, 1320 DenseSet<GlobalVariable *> const &LDSVarsToTransform) { 1321 // Create a struct instance containing LDSVarsToTransform and map from those 1322 // variables to ConstantExprGEP 1323 // Variables may be introduced to meet alignment requirements. No aliasing 1324 // metadata is useful for these as they have no uses. Erased before return. 1325 1326 LLVMContext &Ctx = M.getContext(); 1327 const DataLayout &DL = M.getDataLayout(); 1328 assert(!LDSVarsToTransform.empty()); 1329 1330 SmallVector<OptimizedStructLayoutField, 8> LayoutFields; 1331 LayoutFields.reserve(LDSVarsToTransform.size()); 1332 { 1333 // The order of fields in this struct depends on the order of 1334 // varables in the argument which varies when changing how they 1335 // are identified, leading to spurious test breakage. 1336 auto Sorted = sortByName(std::vector<GlobalVariable *>( 1337 LDSVarsToTransform.begin(), LDSVarsToTransform.end())); 1338 1339 for (GlobalVariable *GV : Sorted) { 1340 OptimizedStructLayoutField F(GV, 1341 DL.getTypeAllocSize(GV->getValueType()), 1342 AMDGPU::getAlign(DL, GV)); 1343 LayoutFields.emplace_back(F); 1344 } 1345 } 1346 1347 performOptimizedStructLayout(LayoutFields); 1348 1349 std::vector<GlobalVariable *> LocalVars; 1350 BitVector IsPaddingField; 1351 LocalVars.reserve(LDSVarsToTransform.size()); // will be at least this large 1352 IsPaddingField.reserve(LDSVarsToTransform.size()); 1353 { 1354 uint64_t CurrentOffset = 0; 1355 for (size_t I = 0; I < LayoutFields.size(); I++) { 1356 GlobalVariable *FGV = static_cast<GlobalVariable *>( 1357 const_cast<void *>(LayoutFields[I].Id)); 1358 Align DataAlign = LayoutFields[I].Alignment; 1359 1360 uint64_t DataAlignV = DataAlign.value(); 1361 if (uint64_t Rem = CurrentOffset % DataAlignV) { 1362 uint64_t Padding = DataAlignV - Rem; 1363 1364 // Append an array of padding bytes to meet alignment requested 1365 // Note (o + (a - (o % a)) ) % a == 0 1366 // (offset + Padding ) % align == 0 1367 1368 Type *ATy = ArrayType::get(Type::getInt8Ty(Ctx), Padding); 1369 LocalVars.push_back(new GlobalVariable( 1370 M, ATy, false, GlobalValue::InternalLinkage, UndefValue::get(ATy), 1371 "", nullptr, GlobalValue::NotThreadLocal, AMDGPUAS::LOCAL_ADDRESS, 1372 false)); 1373 IsPaddingField.push_back(true); 1374 CurrentOffset += Padding; 1375 } 1376 1377 LocalVars.push_back(FGV); 1378 IsPaddingField.push_back(false); 1379 CurrentOffset += LayoutFields[I].Size; 1380 } 1381 } 1382 1383 std::vector<Type *> LocalVarTypes; 1384 LocalVarTypes.reserve(LocalVars.size()); 1385 std::transform( 1386 LocalVars.cbegin(), LocalVars.cend(), std::back_inserter(LocalVarTypes), 1387 [](const GlobalVariable *V) -> Type * { return V->getValueType(); }); 1388 1389 StructType *LDSTy = StructType::create(Ctx, LocalVarTypes, VarName + ".t"); 1390 1391 Align StructAlign = AMDGPU::getAlign(DL, LocalVars[0]); 1392 1393 GlobalVariable *SGV = new GlobalVariable( 1394 M, LDSTy, false, GlobalValue::InternalLinkage, UndefValue::get(LDSTy), 1395 VarName, nullptr, GlobalValue::NotThreadLocal, AMDGPUAS::LOCAL_ADDRESS, 1396 false); 1397 SGV->setAlignment(StructAlign); 1398 1399 DenseMap<GlobalVariable *, Constant *> Map; 1400 Type *I32 = Type::getInt32Ty(Ctx); 1401 for (size_t I = 0; I < LocalVars.size(); I++) { 1402 GlobalVariable *GV = LocalVars[I]; 1403 Constant *GEPIdx[] = {ConstantInt::get(I32, 0), ConstantInt::get(I32, I)}; 1404 Constant *GEP = ConstantExpr::getGetElementPtr(LDSTy, SGV, GEPIdx, true); 1405 if (IsPaddingField[I]) { 1406 assert(GV->use_empty()); 1407 GV->eraseFromParent(); 1408 } else { 1409 Map[GV] = GEP; 1410 } 1411 } 1412 assert(Map.size() == LDSVarsToTransform.size()); 1413 return {SGV, std::move(Map)}; 1414 } 1415 1416 template <typename PredicateTy> 1417 static void replaceLDSVariablesWithStruct( 1418 Module &M, DenseSet<GlobalVariable *> const &LDSVarsToTransformArg, 1419 const LDSVariableReplacement &Replacement, PredicateTy Predicate) { 1420 LLVMContext &Ctx = M.getContext(); 1421 const DataLayout &DL = M.getDataLayout(); 1422 1423 // A hack... we need to insert the aliasing info in a predictable order for 1424 // lit tests. Would like to have them in a stable order already, ideally the 1425 // same order they get allocated, which might mean an ordered set container 1426 auto LDSVarsToTransform = sortByName(std::vector<GlobalVariable *>( 1427 LDSVarsToTransformArg.begin(), LDSVarsToTransformArg.end())); 1428 1429 // Create alias.scope and their lists. Each field in the new structure 1430 // does not alias with all other fields. 1431 SmallVector<MDNode *> AliasScopes; 1432 SmallVector<Metadata *> NoAliasList; 1433 const size_t NumberVars = LDSVarsToTransform.size(); 1434 if (NumberVars > 1) { 1435 MDBuilder MDB(Ctx); 1436 AliasScopes.reserve(NumberVars); 1437 MDNode *Domain = MDB.createAnonymousAliasScopeDomain(); 1438 for (size_t I = 0; I < NumberVars; I++) { 1439 MDNode *Scope = MDB.createAnonymousAliasScope(Domain); 1440 AliasScopes.push_back(Scope); 1441 } 1442 NoAliasList.append(&AliasScopes[1], AliasScopes.end()); 1443 } 1444 1445 // Replace uses of ith variable with a constantexpr to the corresponding 1446 // field of the instance that will be allocated by AMDGPUMachineFunction 1447 for (size_t I = 0; I < NumberVars; I++) { 1448 GlobalVariable *GV = LDSVarsToTransform[I]; 1449 Constant *GEP = Replacement.LDSVarsToConstantGEP.at(GV); 1450 1451 GV->replaceUsesWithIf(GEP, Predicate); 1452 1453 APInt APOff(DL.getIndexTypeSizeInBits(GEP->getType()), 0); 1454 GEP->stripAndAccumulateInBoundsConstantOffsets(DL, APOff); 1455 uint64_t Offset = APOff.getZExtValue(); 1456 1457 Align A = 1458 commonAlignment(Replacement.SGV->getAlign().valueOrOne(), Offset); 1459 1460 if (I) 1461 NoAliasList[I - 1] = AliasScopes[I - 1]; 1462 MDNode *NoAlias = 1463 NoAliasList.empty() ? nullptr : MDNode::get(Ctx, NoAliasList); 1464 MDNode *AliasScope = 1465 AliasScopes.empty() ? nullptr : MDNode::get(Ctx, {AliasScopes[I]}); 1466 1467 refineUsesAlignmentAndAA(GEP, A, DL, AliasScope, NoAlias); 1468 } 1469 } 1470 1471 static void refineUsesAlignmentAndAA(Value *Ptr, Align A, 1472 const DataLayout &DL, MDNode *AliasScope, 1473 MDNode *NoAlias, unsigned MaxDepth = 5) { 1474 if (!MaxDepth || (A == 1 && !AliasScope)) 1475 return; 1476 1477 for (User *U : Ptr->users()) { 1478 if (auto *I = dyn_cast<Instruction>(U)) { 1479 if (AliasScope && I->mayReadOrWriteMemory()) { 1480 MDNode *AS = I->getMetadata(LLVMContext::MD_alias_scope); 1481 AS = (AS ? MDNode::getMostGenericAliasScope(AS, AliasScope) 1482 : AliasScope); 1483 I->setMetadata(LLVMContext::MD_alias_scope, AS); 1484 1485 MDNode *NA = I->getMetadata(LLVMContext::MD_noalias); 1486 NA = (NA ? MDNode::intersect(NA, NoAlias) : NoAlias); 1487 I->setMetadata(LLVMContext::MD_noalias, NA); 1488 } 1489 } 1490 1491 if (auto *LI = dyn_cast<LoadInst>(U)) { 1492 LI->setAlignment(std::max(A, LI->getAlign())); 1493 continue; 1494 } 1495 if (auto *SI = dyn_cast<StoreInst>(U)) { 1496 if (SI->getPointerOperand() == Ptr) 1497 SI->setAlignment(std::max(A, SI->getAlign())); 1498 continue; 1499 } 1500 if (auto *AI = dyn_cast<AtomicRMWInst>(U)) { 1501 // None of atomicrmw operations can work on pointers, but let's 1502 // check it anyway in case it will or we will process ConstantExpr. 1503 if (AI->getPointerOperand() == Ptr) 1504 AI->setAlignment(std::max(A, AI->getAlign())); 1505 continue; 1506 } 1507 if (auto *AI = dyn_cast<AtomicCmpXchgInst>(U)) { 1508 if (AI->getPointerOperand() == Ptr) 1509 AI->setAlignment(std::max(A, AI->getAlign())); 1510 continue; 1511 } 1512 if (auto *GEP = dyn_cast<GetElementPtrInst>(U)) { 1513 unsigned BitWidth = DL.getIndexTypeSizeInBits(GEP->getType()); 1514 APInt Off(BitWidth, 0); 1515 if (GEP->getPointerOperand() == Ptr) { 1516 Align GA; 1517 if (GEP->accumulateConstantOffset(DL, Off)) 1518 GA = commonAlignment(A, Off.getLimitedValue()); 1519 refineUsesAlignmentAndAA(GEP, GA, DL, AliasScope, NoAlias, 1520 MaxDepth - 1); 1521 } 1522 continue; 1523 } 1524 if (auto *I = dyn_cast<Instruction>(U)) { 1525 if (I->getOpcode() == Instruction::BitCast || 1526 I->getOpcode() == Instruction::AddrSpaceCast) 1527 refineUsesAlignmentAndAA(I, A, DL, AliasScope, NoAlias, MaxDepth - 1); 1528 } 1529 } 1530 } 1531 }; 1532 1533 } // namespace 1534 char AMDGPULowerModuleLDS::ID = 0; 1535 1536 char &llvm::AMDGPULowerModuleLDSID = AMDGPULowerModuleLDS::ID; 1537 1538 INITIALIZE_PASS(AMDGPULowerModuleLDS, DEBUG_TYPE, 1539 "Lower uses of LDS variables from non-kernel functions", false, 1540 false) 1541 1542 ModulePass *llvm::createAMDGPULowerModuleLDSPass() { 1543 return new AMDGPULowerModuleLDS(); 1544 } 1545 1546 PreservedAnalyses AMDGPULowerModuleLDSPass::run(Module &M, 1547 ModuleAnalysisManager &) { 1548 return AMDGPULowerModuleLDS().runOnModule(M) ? PreservedAnalyses::none() 1549 : PreservedAnalyses::all(); 1550 } 1551