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 // Kernels that do not use this struct are annoteated with the attribute 77 // amdgpu-elide-module-lds which allows the back end to elide the allocation. 78 // 79 // The "table" lowering implemented here has three components. 80 // First kernels are assigned a unique integer identifier which is available in 81 // functions it calls through the intrinsic amdgcn_lds_kernel_id. The integer 82 // is passed through a specific SGPR, thus works with indirect calls. 83 // Second, each kernel allocates LDS variables independent of other kernels and 84 // writes the addresses it chose for each variable into an array in consistent 85 // order. If the kernel does not allocate a given variable, it writes undef to 86 // the corresponding array location. These arrays are written to a constant 87 // table in the order matching the kernel unique integer identifier. 88 // Third, uses from non-kernel functions are replaced with a table lookup using 89 // the intrinsic function to find the address of the variable. 90 // 91 // "Kernel" lowering is only applicable for variables that are unambiguously 92 // reachable from exactly one kernel. For those cases, accesses to the variable 93 // can be lowered to ConstantExpr address of a struct instance specific to that 94 // one kernel. This is zero cost in space and in compute. It will raise a fatal 95 // error on any variable that might be reachable from multiple kernels and is 96 // thus most easily used as part of the hybrid lowering strategy. 97 // 98 // Hybrid lowering is a mixture of the above. It uses the zero cost kernel 99 // lowering where it can. It lowers the variable accessed by the greatest 100 // number of kernels using the module strategy as that is free for the first 101 // variable. Any futher variables that can be lowered with the module strategy 102 // without incurring LDS memory overhead are. The remaining ones are lowered 103 // via table. 104 // 105 // Consequences 106 // - No heuristics or user controlled magic numbers, hybrid is the right choice 107 // - Kernels that don't use functions (or have had them all inlined) are not 108 // affected by any lowering for kernels that do. 109 // - Kernels that don't make indirect function calls are not affected by those 110 // that do. 111 // - Variables which are used by lots of kernels, e.g. those injected by a 112 // language runtime in most kernels, are expected to have no overhead 113 // - Implementations that instantiate templates per-kernel where those templates 114 // use LDS are expected to hit the "Kernel" lowering strategy 115 // - The runtime properties impose a cost in compiler implementation complexity 116 // 117 // Dynamic LDS implementation 118 // Dynamic LDS is lowered similarly to the "table" strategy above and uses the 119 // same intrinsic to identify which kernel is at the root of the dynamic call 120 // graph. This relies on the specified behaviour that all dynamic LDS variables 121 // alias one another, i.e. are at the same address, with respect to a given 122 // kernel. Therefore this pass creates new dynamic LDS variables for each kernel 123 // that allocates any dynamic LDS and builds a table of addresses out of those. 124 // The AMDGPUPromoteAlloca pass skips kernels that use dynamic LDS. 125 // The corresponding optimisation for "kernel" lowering where the table lookup 126 // is elided is not implemented. 127 // 128 // 129 // Implementation notes / limitations 130 // A single LDS global variable represents an instance per kernel that can reach 131 // said variables. This pass essentially specialises said variables per kernel. 132 // Handling ConstantExpr during the pass complicated this significantly so now 133 // all ConstantExpr uses of LDS variables are expanded to instructions. This 134 // may need amending when implementing non-undef initialisers. 135 // 136 // Lowering is split between this IR pass and the back end. This pass chooses 137 // where given variables should be allocated and marks them with metadata, 138 // MD_absolute_symbol. The backend places the variables in coincidentally the 139 // same location and raises a fatal error if something has gone awry. This works 140 // in practice because the only pass between this one and the backend that 141 // changes LDS is PromoteAlloca and the changes it makes do not conflict. 142 // 143 // Addresses are written to constant global arrays based on the same metadata. 144 // 145 // The backend lowers LDS variables in the order of traversal of the function. 146 // This is at odds with the deterministic layout required. The workaround is to 147 // allocate the fixed-address variables immediately upon starting the function 148 // where they can be placed as intended. This requires a means of mapping from 149 // the function to the variables that it allocates. For the module scope lds, 150 // this is via metadata indicating whether the variable is not required. If a 151 // pass deletes that metadata, a fatal error on disagreement with the absolute 152 // symbol metadata will occur. For kernel scope and dynamic, this is by _name_ 153 // correspondence between the function and the variable. It requires the 154 // kernel to have a name (which is only a limitation for tests in practice) and 155 // for nothing to rename the corresponding symbols. This is a hazard if the pass 156 // is run multiple times during debugging. Alternative schemes considered all 157 // involve bespoke metadata. 158 // 159 // If the name correspondence can be replaced, multiple distinct kernels that 160 // have the same memory layout can map to the same kernel id (as the address 161 // itself is handled by the absolute symbol metadata) and that will allow more 162 // uses of the "kernel" style faster lowering and reduce the size of the lookup 163 // tables. 164 // 165 // There is a test that checks this does not fire for a graphics shader. This 166 // lowering is expected to work for graphics if the isKernel test is changed. 167 // 168 // The current markUsedByKernel is sufficient for PromoteAlloca but is elided 169 // before codegen. Replacing this with an equivalent intrinsic which lasts until 170 // shortly after the machine function lowering of LDS would help break the name 171 // mapping. The other part needed is probably to amend PromoteAlloca to embed 172 // the LDS variables it creates in the same struct created here. That avoids the 173 // current hazard where a PromoteAlloca LDS variable might be allocated before 174 // the kernel scope (and thus error on the address check). Given a new invariant 175 // that no LDS variables exist outside of the structs managed here, and an 176 // intrinsic that lasts until after the LDS frame lowering, it should be 177 // possible to drop the name mapping and fold equivalent memory layouts. 178 // 179 //===----------------------------------------------------------------------===// 180 181 #include "AMDGPU.h" 182 #include "Utils/AMDGPUBaseInfo.h" 183 #include "Utils/AMDGPUMemoryUtils.h" 184 #include "llvm/ADT/BitVector.h" 185 #include "llvm/ADT/DenseMap.h" 186 #include "llvm/ADT/DenseSet.h" 187 #include "llvm/ADT/STLExtras.h" 188 #include "llvm/ADT/SetOperations.h" 189 #include "llvm/ADT/SetVector.h" 190 #include "llvm/Analysis/CallGraph.h" 191 #include "llvm/IR/Constants.h" 192 #include "llvm/IR/DerivedTypes.h" 193 #include "llvm/IR/IRBuilder.h" 194 #include "llvm/IR/InlineAsm.h" 195 #include "llvm/IR/Instructions.h" 196 #include "llvm/IR/IntrinsicsAMDGPU.h" 197 #include "llvm/IR/MDBuilder.h" 198 #include "llvm/IR/ReplaceConstant.h" 199 #include "llvm/InitializePasses.h" 200 #include "llvm/Pass.h" 201 #include "llvm/Support/CommandLine.h" 202 #include "llvm/Support/Debug.h" 203 #include "llvm/Support/OptimizedStructLayout.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 class AMDGPULowerModuleLDS : public ModulePass { 249 250 static void 251 removeLocalVarsFromUsedLists(Module &M, 252 const DenseSet<GlobalVariable *> &LocalVars) { 253 // The verifier rejects used lists containing an inttoptr of a constant 254 // so remove the variables from these lists before replaceAllUsesWith 255 SmallPtrSet<Constant *, 8> LocalVarsSet; 256 for (GlobalVariable *LocalVar : LocalVars) 257 LocalVarsSet.insert(cast<Constant>(LocalVar->stripPointerCasts())); 258 259 removeFromUsedLists( 260 M, [&LocalVarsSet](Constant *C) { return LocalVarsSet.count(C); }); 261 262 for (GlobalVariable *LocalVar : LocalVars) 263 LocalVar->removeDeadConstantUsers(); 264 } 265 266 static void markUsedByKernel(IRBuilder<> &Builder, Function *Func, 267 GlobalVariable *SGV) { 268 // The llvm.amdgcn.module.lds instance is implicitly used by all kernels 269 // that might call a function which accesses a field within it. This is 270 // presently approximated to 'all kernels' if there are any such functions 271 // in the module. This implicit use is redefined as an explicit use here so 272 // that later passes, specifically PromoteAlloca, account for the required 273 // memory without any knowledge of this transform. 274 275 // An operand bundle on llvm.donothing works because the call instruction 276 // survives until after the last pass that needs to account for LDS. It is 277 // better than inline asm as the latter survives until the end of codegen. A 278 // totally robust solution would be a function with the same semantics as 279 // llvm.donothing that takes a pointer to the instance and is lowered to a 280 // no-op after LDS is allocated, but that is not presently necessary. 281 282 // This intrinsic is eliminated shortly before instruction selection. It 283 // does not suffice to indicate to ISel that a given global which is not 284 // immediately used by the kernel must still be allocated by it. An 285 // equivalent target specific intrinsic which lasts until immediately after 286 // codegen would suffice for that, but one would still need to ensure that 287 // the variables are allocated in the anticpated order. 288 289 LLVMContext &Ctx = Func->getContext(); 290 291 Builder.SetInsertPoint(Func->getEntryBlock().getFirstNonPHI()); 292 293 FunctionType *FTy = FunctionType::get(Type::getVoidTy(Ctx), {}); 294 295 Function *Decl = 296 Intrinsic::getDeclaration(Func->getParent(), Intrinsic::donothing, {}); 297 298 Value *UseInstance[1] = {Builder.CreateInBoundsGEP( 299 SGV->getValueType(), SGV, ConstantInt::get(Type::getInt32Ty(Ctx), 0))}; 300 301 Builder.CreateCall(FTy, Decl, {}, 302 {OperandBundleDefT<Value *>("ExplicitUse", UseInstance)}, 303 ""); 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 SmallVector<User *, 16> Stack(GV.users()); 351 for (User *V : GV.users()) { 352 if (auto *I = dyn_cast<Instruction>(V)) { 353 Function *F = I->getFunction(); 354 if (isKernelLDS(F)) { 355 kernels[F].insert(&GV); 356 } else { 357 functions[F].insert(&GV); 358 } 359 } 360 } 361 } 362 } 363 364 struct LDSUsesInfoTy { 365 FunctionVariableMap direct_access; 366 FunctionVariableMap indirect_access; 367 }; 368 369 static LDSUsesInfoTy getTransitiveUsesOfLDS(CallGraph const &CG, Module &M) { 370 371 FunctionVariableMap direct_map_kernel; 372 FunctionVariableMap direct_map_function; 373 getUsesOfLDSByFunction(CG, M, direct_map_kernel, direct_map_function); 374 375 // Collect variables that are used by functions whose address has escaped 376 DenseSet<GlobalVariable *> VariablesReachableThroughFunctionPointer; 377 for (Function &F : M.functions()) { 378 if (!isKernelLDS(&F)) 379 if (F.hasAddressTaken(nullptr, 380 /* IgnoreCallbackUses */ false, 381 /* IgnoreAssumeLikeCalls */ false, 382 /* IgnoreLLVMUsed */ true, 383 /* IgnoreArcAttachedCall */ false)) { 384 set_union(VariablesReachableThroughFunctionPointer, 385 direct_map_function[&F]); 386 } 387 } 388 389 auto functionMakesUnknownCall = [&](const Function *F) -> bool { 390 assert(!F->isDeclaration()); 391 for (CallGraphNode::CallRecord R : *CG[F]) { 392 if (!R.second->getFunction()) { 393 return true; 394 } 395 } 396 return false; 397 }; 398 399 // Work out which variables are reachable through function calls 400 FunctionVariableMap transitive_map_function = direct_map_function; 401 402 // If the function makes any unknown call, assume the worst case that it can 403 // access all variables accessed by functions whose address escaped 404 for (Function &F : M.functions()) { 405 if (!F.isDeclaration() && functionMakesUnknownCall(&F)) { 406 if (!isKernelLDS(&F)) { 407 set_union(transitive_map_function[&F], 408 VariablesReachableThroughFunctionPointer); 409 } 410 } 411 } 412 413 // Direct implementation of collecting all variables reachable from each 414 // function 415 for (Function &Func : M.functions()) { 416 if (Func.isDeclaration() || isKernelLDS(&Func)) 417 continue; 418 419 DenseSet<Function *> seen; // catches cycles 420 SmallVector<Function *, 4> wip{&Func}; 421 422 while (!wip.empty()) { 423 Function *F = wip.pop_back_val(); 424 425 // Can accelerate this by referring to transitive map for functions that 426 // have already been computed, with more care than this 427 set_union(transitive_map_function[&Func], direct_map_function[F]); 428 429 for (CallGraphNode::CallRecord R : *CG[F]) { 430 Function *ith = R.second->getFunction(); 431 if (ith) { 432 if (!seen.contains(ith)) { 433 seen.insert(ith); 434 wip.push_back(ith); 435 } 436 } 437 } 438 } 439 } 440 441 // direct_map_kernel lists which variables are used by the kernel 442 // find the variables which are used through a function call 443 FunctionVariableMap indirect_map_kernel; 444 445 for (Function &Func : M.functions()) { 446 if (Func.isDeclaration() || !isKernelLDS(&Func)) 447 continue; 448 449 for (CallGraphNode::CallRecord R : *CG[&Func]) { 450 Function *ith = R.second->getFunction(); 451 if (ith) { 452 set_union(indirect_map_kernel[&Func], transitive_map_function[ith]); 453 } else { 454 set_union(indirect_map_kernel[&Func], 455 VariablesReachableThroughFunctionPointer); 456 } 457 } 458 } 459 460 return {std::move(direct_map_kernel), std::move(indirect_map_kernel)}; 461 } 462 463 struct LDSVariableReplacement { 464 GlobalVariable *SGV = nullptr; 465 DenseMap<GlobalVariable *, Constant *> LDSVarsToConstantGEP; 466 }; 467 468 // remap from lds global to a constantexpr gep to where it has been moved to 469 // for each kernel 470 // an array with an element for each kernel containing where the corresponding 471 // variable was remapped to 472 473 static Constant *getAddressesOfVariablesInKernel( 474 LLVMContext &Ctx, ArrayRef<GlobalVariable *> Variables, 475 DenseMap<GlobalVariable *, Constant *> &LDSVarsToConstantGEP) { 476 // Create a ConstantArray containing the address of each Variable within the 477 // kernel corresponding to LDSVarsToConstantGEP, or poison if that kernel 478 // does not allocate it 479 // TODO: Drop the ptrtoint conversion 480 481 Type *I32 = Type::getInt32Ty(Ctx); 482 483 ArrayType *KernelOffsetsType = ArrayType::get(I32, Variables.size()); 484 485 SmallVector<Constant *> Elements; 486 for (size_t i = 0; i < Variables.size(); i++) { 487 GlobalVariable *GV = Variables[i]; 488 if (LDSVarsToConstantGEP.count(GV) != 0) { 489 auto elt = ConstantExpr::getPtrToInt(LDSVarsToConstantGEP[GV], I32); 490 Elements.push_back(elt); 491 } else { 492 Elements.push_back(PoisonValue::get(I32)); 493 } 494 } 495 return ConstantArray::get(KernelOffsetsType, Elements); 496 } 497 498 static GlobalVariable *buildLookupTable( 499 Module &M, ArrayRef<GlobalVariable *> Variables, 500 ArrayRef<Function *> kernels, 501 DenseMap<Function *, LDSVariableReplacement> &KernelToReplacement) { 502 if (Variables.empty()) { 503 return nullptr; 504 } 505 LLVMContext &Ctx = M.getContext(); 506 507 const size_t NumberVariables = Variables.size(); 508 const size_t NumberKernels = kernels.size(); 509 510 ArrayType *KernelOffsetsType = 511 ArrayType::get(Type::getInt32Ty(Ctx), NumberVariables); 512 513 ArrayType *AllKernelsOffsetsType = 514 ArrayType::get(KernelOffsetsType, NumberKernels); 515 516 std::vector<Constant *> overallConstantExprElts(NumberKernels); 517 for (size_t i = 0; i < NumberKernels; i++) { 518 LDSVariableReplacement Replacement = KernelToReplacement[kernels[i]]; 519 overallConstantExprElts[i] = getAddressesOfVariablesInKernel( 520 Ctx, Variables, Replacement.LDSVarsToConstantGEP); 521 } 522 523 Constant *init = 524 ConstantArray::get(AllKernelsOffsetsType, overallConstantExprElts); 525 526 return new GlobalVariable( 527 M, AllKernelsOffsetsType, true, GlobalValue::InternalLinkage, init, 528 "llvm.amdgcn.lds.offset.table", nullptr, GlobalValue::NotThreadLocal, 529 AMDGPUAS::CONSTANT_ADDRESS); 530 } 531 532 void replaceUseWithTableLookup(Module &M, IRBuilder<> &Builder, 533 GlobalVariable *LookupTable, 534 GlobalVariable *GV, Use &U, 535 Value *OptionalIndex) { 536 // Table is a constant array of the same length as OrderedKernels 537 LLVMContext &Ctx = M.getContext(); 538 Type *I32 = Type::getInt32Ty(Ctx); 539 auto *I = cast<Instruction>(U.getUser()); 540 541 Value *tableKernelIndex = getTableLookupKernelIndex(M, I->getFunction()); 542 543 if (auto *Phi = dyn_cast<PHINode>(I)) { 544 BasicBlock *BB = Phi->getIncomingBlock(U); 545 Builder.SetInsertPoint(&(*(BB->getFirstInsertionPt()))); 546 } else { 547 Builder.SetInsertPoint(I); 548 } 549 550 SmallVector<Value *, 3> GEPIdx = { 551 ConstantInt::get(I32, 0), 552 tableKernelIndex, 553 }; 554 if (OptionalIndex) 555 GEPIdx.push_back(OptionalIndex); 556 557 Value *Address = Builder.CreateInBoundsGEP( 558 LookupTable->getValueType(), LookupTable, GEPIdx, GV->getName()); 559 560 Value *loaded = Builder.CreateLoad(I32, Address); 561 562 Value *replacement = 563 Builder.CreateIntToPtr(loaded, GV->getType(), GV->getName()); 564 565 U.set(replacement); 566 } 567 568 void replaceUsesInInstructionsWithTableLookup( 569 Module &M, ArrayRef<GlobalVariable *> ModuleScopeVariables, 570 GlobalVariable *LookupTable) { 571 572 LLVMContext &Ctx = M.getContext(); 573 IRBuilder<> Builder(Ctx); 574 Type *I32 = Type::getInt32Ty(Ctx); 575 576 for (size_t Index = 0; Index < ModuleScopeVariables.size(); Index++) { 577 auto *GV = ModuleScopeVariables[Index]; 578 579 for (Use &U : make_early_inc_range(GV->uses())) { 580 auto *I = dyn_cast<Instruction>(U.getUser()); 581 if (!I) 582 continue; 583 584 replaceUseWithTableLookup(M, Builder, LookupTable, GV, U, 585 ConstantInt::get(I32, Index)); 586 } 587 } 588 } 589 590 static DenseSet<Function *> kernelsThatIndirectlyAccessAnyOfPassedVariables( 591 Module &M, LDSUsesInfoTy &LDSUsesInfo, 592 DenseSet<GlobalVariable *> const &VariableSet) { 593 594 DenseSet<Function *> KernelSet; 595 596 if (VariableSet.empty()) 597 return KernelSet; 598 599 for (Function &Func : M.functions()) { 600 if (Func.isDeclaration() || !isKernelLDS(&Func)) 601 continue; 602 for (GlobalVariable *GV : LDSUsesInfo.indirect_access[&Func]) { 603 if (VariableSet.contains(GV)) { 604 KernelSet.insert(&Func); 605 break; 606 } 607 } 608 } 609 610 return KernelSet; 611 } 612 613 static GlobalVariable * 614 chooseBestVariableForModuleStrategy(const DataLayout &DL, 615 VariableFunctionMap &LDSVars) { 616 // Find the global variable with the most indirect uses from kernels 617 618 struct CandidateTy { 619 GlobalVariable *GV = nullptr; 620 size_t UserCount = 0; 621 size_t Size = 0; 622 623 CandidateTy() = default; 624 625 CandidateTy(GlobalVariable *GV, uint64_t UserCount, uint64_t AllocSize) 626 : GV(GV), UserCount(UserCount), Size(AllocSize) {} 627 628 bool operator<(const CandidateTy &Other) const { 629 // Fewer users makes module scope variable less attractive 630 if (UserCount < Other.UserCount) { 631 return true; 632 } 633 if (UserCount > Other.UserCount) { 634 return false; 635 } 636 637 // Bigger makes module scope variable less attractive 638 if (Size < Other.Size) { 639 return false; 640 } 641 642 if (Size > Other.Size) { 643 return true; 644 } 645 646 // Arbitrary but consistent 647 return GV->getName() < Other.GV->getName(); 648 } 649 }; 650 651 CandidateTy MostUsed; 652 653 for (auto &K : LDSVars) { 654 GlobalVariable *GV = K.first; 655 if (K.second.size() <= 1) { 656 // A variable reachable by only one kernel is best lowered with kernel 657 // strategy 658 continue; 659 } 660 CandidateTy Candidate( 661 GV, K.second.size(), 662 DL.getTypeAllocSize(GV->getValueType()).getFixedValue()); 663 if (MostUsed < Candidate) 664 MostUsed = Candidate; 665 } 666 667 return MostUsed.GV; 668 } 669 670 static void recordLDSAbsoluteAddress(Module *M, GlobalVariable *GV, 671 uint32_t Address) { 672 // Write the specified address into metadata where it can be retrieved by 673 // the assembler. Format is a half open range, [Address Address+1) 674 LLVMContext &Ctx = M->getContext(); 675 auto *IntTy = 676 M->getDataLayout().getIntPtrType(Ctx, AMDGPUAS::LOCAL_ADDRESS); 677 auto *MinC = ConstantAsMetadata::get(ConstantInt::get(IntTy, Address)); 678 auto *MaxC = ConstantAsMetadata::get(ConstantInt::get(IntTy, Address + 1)); 679 GV->setMetadata(LLVMContext::MD_absolute_symbol, 680 MDNode::get(Ctx, {MinC, MaxC})); 681 } 682 683 DenseMap<Function *, Value *> tableKernelIndexCache; 684 Value *getTableLookupKernelIndex(Module &M, Function *F) { 685 // Accesses from a function use the amdgcn_lds_kernel_id intrinsic which 686 // lowers to a read from a live in register. Emit it once in the entry 687 // block to spare deduplicating it later. 688 if (tableKernelIndexCache.count(F) == 0) { 689 LLVMContext &Ctx = M.getContext(); 690 IRBuilder<> Builder(Ctx); 691 FunctionType *FTy = FunctionType::get(Type::getInt32Ty(Ctx), {}); 692 Function *Decl = 693 Intrinsic::getDeclaration(&M, Intrinsic::amdgcn_lds_kernel_id, {}); 694 695 BasicBlock::iterator it = 696 F->getEntryBlock().getFirstNonPHIOrDbgOrAlloca(); 697 Instruction &i = *it; 698 Builder.SetInsertPoint(&i); 699 700 tableKernelIndexCache[F] = Builder.CreateCall(FTy, Decl, {}); 701 } 702 703 return tableKernelIndexCache[F]; 704 } 705 706 static std::vector<Function *> assignLDSKernelIDToEachKernel( 707 Module *M, DenseSet<Function *> const &KernelsThatAllocateTableLDS, 708 DenseSet<Function *> const &KernelsThatIndirectlyAllocateDynamicLDS) { 709 // Associate kernels in the set with an arbirary but reproducible order and 710 // annotate them with that order in metadata. This metadata is recognised by 711 // the backend and lowered to a SGPR which can be read from using 712 // amdgcn_lds_kernel_id. 713 714 std::vector<Function *> OrderedKernels; 715 if (!KernelsThatAllocateTableLDS.empty() || 716 !KernelsThatIndirectlyAllocateDynamicLDS.empty()) { 717 718 for (Function &Func : M->functions()) { 719 if (Func.isDeclaration()) 720 continue; 721 if (!isKernelLDS(&Func)) 722 continue; 723 724 if (KernelsThatAllocateTableLDS.contains(&Func) || 725 KernelsThatIndirectlyAllocateDynamicLDS.contains(&Func)) { 726 assert(Func.hasName()); // else fatal error earlier 727 OrderedKernels.push_back(&Func); 728 } 729 } 730 731 // Put them in an arbitrary but reproducible order 732 llvm::sort(OrderedKernels.begin(), OrderedKernels.end(), 733 [](const Function *lhs, const Function *rhs) -> bool { 734 return lhs->getName() < rhs->getName(); 735 }); 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 Type::getInt8PtrTy(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 IRBuilder<> Builder(Ctx); 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(Builder, &Func, ModuleScopeReplacement.SGV); 896 897 } else { 898 markElideModuleLDS(Func); 899 } 900 } 901 902 return ModuleScopeReplacement.SGV; 903 } 904 905 static DenseMap<Function *, LDSVariableReplacement> 906 lowerKernelScopeStructVariables( 907 Module &M, LDSUsesInfoTy &LDSUsesInfo, 908 DenseSet<GlobalVariable *> const &ModuleScopeVariables, 909 DenseSet<Function *> const &KernelsThatAllocateModuleLDS, 910 GlobalVariable *MaybeModuleScopeStruct) { 911 912 // Create a struct for each kernel for the non-module-scope variables. 913 914 DenseMap<Function *, LDSVariableReplacement> KernelToReplacement; 915 for (Function &Func : M.functions()) { 916 if (Func.isDeclaration() || !isKernelLDS(&Func)) 917 continue; 918 919 DenseSet<GlobalVariable *> KernelUsedVariables; 920 // Allocating variables that are used directly in this struct to get 921 // alignment aware allocation and predictable frame size. 922 for (auto &v : LDSUsesInfo.direct_access[&Func]) { 923 if (!AMDGPU::isDynamicLDS(*v)) { 924 KernelUsedVariables.insert(v); 925 } 926 } 927 928 // Allocating variables that are accessed indirectly so that a lookup of 929 // this struct instance can find them from nested functions. 930 for (auto &v : LDSUsesInfo.indirect_access[&Func]) { 931 if (!AMDGPU::isDynamicLDS(*v)) { 932 KernelUsedVariables.insert(v); 933 } 934 } 935 936 // Variables allocated in module lds must all resolve to that struct, 937 // not to the per-kernel instance. 938 if (KernelsThatAllocateModuleLDS.contains(&Func)) { 939 for (GlobalVariable *v : ModuleScopeVariables) { 940 KernelUsedVariables.erase(v); 941 } 942 } 943 944 if (KernelUsedVariables.empty()) { 945 // Either used no LDS, or the LDS it used was all in the module struct 946 // or dynamically sized 947 continue; 948 } 949 950 // The association between kernel function and LDS struct is done by 951 // symbol name, which only works if the function in question has a 952 // name This is not expected to be a problem in practice as kernels 953 // are called by name making anonymous ones (which are named by the 954 // backend) difficult to use. This does mean that llvm test cases need 955 // to name the kernels. 956 if (!Func.hasName()) { 957 report_fatal_error("Anonymous kernels cannot use LDS variables"); 958 } 959 960 std::string VarName = 961 (Twine("llvm.amdgcn.kernel.") + Func.getName() + ".lds").str(); 962 963 auto Replacement = 964 createLDSVariableReplacement(M, VarName, KernelUsedVariables); 965 966 // remove preserves existing codegen 967 removeLocalVarsFromUsedLists(M, KernelUsedVariables); 968 KernelToReplacement[&Func] = Replacement; 969 970 // Rewrite uses within kernel to the new struct 971 replaceLDSVariablesWithStruct( 972 M, KernelUsedVariables, Replacement, [&Func](Use &U) { 973 Instruction *I = dyn_cast<Instruction>(U.getUser()); 974 return I && I->getFunction() == &Func; 975 }); 976 } 977 return KernelToReplacement; 978 } 979 980 static GlobalVariable * 981 buildRepresentativeDynamicLDSInstance(Module &M, LDSUsesInfoTy &LDSUsesInfo, 982 Function *func) { 983 // Create a dynamic lds variable with a name associated with the passed 984 // function that has the maximum alignment of any dynamic lds variable 985 // reachable from this kernel. Dynamic LDS is allocated after the static LDS 986 // allocation, possibly after alignment padding. The representative variable 987 // created here has the maximum alignment of any other dynamic variable 988 // reachable by that kernel. All dynamic LDS variables are allocated at the 989 // same address in each kernel in order to provide the documented aliasing 990 // semantics. Setting the alignment here allows this IR pass to accurately 991 // predict the exact constant at which it will be allocated. 992 993 assert(isKernelLDS(func)); 994 995 LLVMContext &Ctx = M.getContext(); 996 const DataLayout &DL = M.getDataLayout(); 997 Align MaxDynamicAlignment(1); 998 999 auto UpdateMaxAlignment = [&MaxDynamicAlignment, &DL](GlobalVariable *GV) { 1000 if (AMDGPU::isDynamicLDS(*GV)) { 1001 MaxDynamicAlignment = 1002 std::max(MaxDynamicAlignment, AMDGPU::getAlign(DL, GV)); 1003 } 1004 }; 1005 1006 for (GlobalVariable *GV : LDSUsesInfo.indirect_access[func]) { 1007 UpdateMaxAlignment(GV); 1008 } 1009 1010 for (GlobalVariable *GV : LDSUsesInfo.direct_access[func]) { 1011 UpdateMaxAlignment(GV); 1012 } 1013 1014 assert(func->hasName()); // Checked by caller 1015 auto emptyCharArray = ArrayType::get(Type::getInt8Ty(Ctx), 0); 1016 GlobalVariable *N = new GlobalVariable( 1017 M, emptyCharArray, false, GlobalValue::ExternalLinkage, nullptr, 1018 Twine("llvm.amdgcn." + func->getName() + ".dynlds"), nullptr, GlobalValue::NotThreadLocal, AMDGPUAS::LOCAL_ADDRESS, 1019 false); 1020 N->setAlignment(MaxDynamicAlignment); 1021 1022 assert(AMDGPU::isDynamicLDS(*N)); 1023 return N; 1024 } 1025 1026 DenseMap<Function *, GlobalVariable *> lowerDynamicLDSVariables( 1027 Module &M, LDSUsesInfoTy &LDSUsesInfo, 1028 DenseSet<Function *> const &KernelsThatIndirectlyAllocateDynamicLDS, 1029 DenseSet<GlobalVariable *> const &DynamicVariables, 1030 std::vector<Function *> const &OrderedKernels) { 1031 DenseMap<Function *, GlobalVariable *> KernelToCreatedDynamicLDS; 1032 if (!KernelsThatIndirectlyAllocateDynamicLDS.empty()) { 1033 LLVMContext &Ctx = M.getContext(); 1034 IRBuilder<> Builder(Ctx); 1035 Type *I32 = Type::getInt32Ty(Ctx); 1036 1037 std::vector<Constant *> newDynamicLDS; 1038 1039 // Table is built in the same order as OrderedKernels 1040 for (auto &func : OrderedKernels) { 1041 1042 if (KernelsThatIndirectlyAllocateDynamicLDS.contains(func)) { 1043 assert(isKernelLDS(func)); 1044 if (!func->hasName()) { 1045 report_fatal_error("Anonymous kernels cannot use LDS variables"); 1046 } 1047 1048 GlobalVariable *N = 1049 buildRepresentativeDynamicLDSInstance(M, LDSUsesInfo, func); 1050 1051 KernelToCreatedDynamicLDS[func] = N; 1052 1053 markUsedByKernel(Builder, func, N); 1054 1055 auto emptyCharArray = ArrayType::get(Type::getInt8Ty(Ctx), 0); 1056 auto GEP = ConstantExpr::getGetElementPtr( 1057 emptyCharArray, N, ConstantInt::get(I32, 0), true); 1058 newDynamicLDS.push_back(ConstantExpr::getPtrToInt(GEP, I32)); 1059 } else { 1060 newDynamicLDS.push_back(PoisonValue::get(I32)); 1061 } 1062 } 1063 assert(OrderedKernels.size() == newDynamicLDS.size()); 1064 1065 ArrayType *t = ArrayType::get(I32, newDynamicLDS.size()); 1066 Constant *init = ConstantArray::get(t, newDynamicLDS); 1067 GlobalVariable *table = new GlobalVariable( 1068 M, t, true, GlobalValue::InternalLinkage, init, 1069 "llvm.amdgcn.dynlds.offset.table", nullptr, 1070 GlobalValue::NotThreadLocal, AMDGPUAS::CONSTANT_ADDRESS); 1071 1072 for (GlobalVariable *GV : DynamicVariables) { 1073 for (Use &U : make_early_inc_range(GV->uses())) { 1074 auto *I = dyn_cast<Instruction>(U.getUser()); 1075 if (!I) 1076 continue; 1077 if (isKernelLDS(I->getFunction())) 1078 continue; 1079 1080 replaceUseWithTableLookup(M, Builder, table, GV, U, nullptr); 1081 } 1082 } 1083 } 1084 return KernelToCreatedDynamicLDS; 1085 } 1086 1087 static bool canElideModuleLDS(const Function &F) { 1088 return F.hasFnAttribute("amdgpu-elide-module-lds"); 1089 } 1090 1091 static void markElideModuleLDS(Function &F) { 1092 F.addFnAttr("amdgpu-elide-module-lds"); 1093 } 1094 1095 bool runOnModule(Module &M) override { 1096 CallGraph CG = CallGraph(M); 1097 bool Changed = superAlignLDSGlobals(M); 1098 1099 Changed |= eliminateConstantExprUsesOfLDSFromAllInstructions(M); 1100 1101 Changed = true; // todo: narrow this down 1102 1103 // For each kernel, what variables does it access directly or through 1104 // callees 1105 LDSUsesInfoTy LDSUsesInfo = getTransitiveUsesOfLDS(CG, M); 1106 1107 // For each variable accessed through callees, which kernels access it 1108 VariableFunctionMap LDSToKernelsThatNeedToAccessItIndirectly; 1109 for (auto &K : LDSUsesInfo.indirect_access) { 1110 Function *F = K.first; 1111 assert(isKernelLDS(F)); 1112 for (GlobalVariable *GV : K.second) { 1113 LDSToKernelsThatNeedToAccessItIndirectly[GV].insert(F); 1114 } 1115 } 1116 1117 // Partition variables accessed indirectly into the different strategies 1118 DenseSet<GlobalVariable *> ModuleScopeVariables; 1119 DenseSet<GlobalVariable *> TableLookupVariables; 1120 DenseSet<GlobalVariable *> KernelAccessVariables; 1121 DenseSet<GlobalVariable *> DynamicVariables; 1122 partitionVariablesIntoIndirectStrategies( 1123 M, LDSUsesInfo, LDSToKernelsThatNeedToAccessItIndirectly, 1124 ModuleScopeVariables, TableLookupVariables, KernelAccessVariables, 1125 DynamicVariables); 1126 1127 // If the kernel accesses a variable that is going to be stored in the 1128 // module instance through a call then that kernel needs to allocate the 1129 // module instance 1130 DenseSet<Function *> KernelsThatAllocateModuleLDS = 1131 kernelsThatIndirectlyAccessAnyOfPassedVariables(M, LDSUsesInfo, 1132 ModuleScopeVariables); 1133 DenseSet<Function *> KernelsThatAllocateTableLDS = 1134 kernelsThatIndirectlyAccessAnyOfPassedVariables(M, LDSUsesInfo, 1135 TableLookupVariables); 1136 1137 DenseSet<Function *> KernelsThatIndirectlyAllocateDynamicLDS = 1138 kernelsThatIndirectlyAccessAnyOfPassedVariables(M, LDSUsesInfo, 1139 DynamicVariables); 1140 1141 GlobalVariable *MaybeModuleScopeStruct = lowerModuleScopeStructVariables( 1142 M, ModuleScopeVariables, KernelsThatAllocateModuleLDS); 1143 1144 DenseMap<Function *, LDSVariableReplacement> KernelToReplacement = 1145 lowerKernelScopeStructVariables(M, LDSUsesInfo, ModuleScopeVariables, 1146 KernelsThatAllocateModuleLDS, 1147 MaybeModuleScopeStruct); 1148 1149 // Lower zero cost accesses to the kernel instances just created 1150 for (auto &GV : KernelAccessVariables) { 1151 auto &funcs = LDSToKernelsThatNeedToAccessItIndirectly[GV]; 1152 assert(funcs.size() == 1); // Only one kernel can access it 1153 LDSVariableReplacement Replacement = 1154 KernelToReplacement[*(funcs.begin())]; 1155 1156 DenseSet<GlobalVariable *> Vec; 1157 Vec.insert(GV); 1158 1159 // TODO: Looks like a latent bug, Replacement may not be marked 1160 // UsedByKernel here 1161 replaceLDSVariablesWithStruct(M, Vec, Replacement, [](Use &U) { 1162 return isa<Instruction>(U.getUser()); 1163 }); 1164 } 1165 1166 // The ith element of this vector is kernel id i 1167 std::vector<Function *> OrderedKernels = 1168 assignLDSKernelIDToEachKernel(&M, KernelsThatAllocateTableLDS, 1169 KernelsThatIndirectlyAllocateDynamicLDS); 1170 1171 if (!KernelsThatAllocateTableLDS.empty()) { 1172 LLVMContext &Ctx = M.getContext(); 1173 IRBuilder<> Builder(Ctx); 1174 1175 for (size_t i = 0; i < OrderedKernels.size(); i++) { 1176 markUsedByKernel(Builder, OrderedKernels[i], 1177 KernelToReplacement[OrderedKernels[i]].SGV); 1178 } 1179 1180 // The order must be consistent between lookup table and accesses to 1181 // lookup table 1182 std::vector<GlobalVariable *> TableLookupVariablesOrdered( 1183 TableLookupVariables.begin(), TableLookupVariables.end()); 1184 llvm::sort(TableLookupVariablesOrdered.begin(), 1185 TableLookupVariablesOrdered.end(), 1186 [](const GlobalVariable *lhs, const GlobalVariable *rhs) { 1187 return lhs->getName() < rhs->getName(); 1188 }); 1189 1190 GlobalVariable *LookupTable = buildLookupTable( 1191 M, TableLookupVariablesOrdered, OrderedKernels, KernelToReplacement); 1192 replaceUsesInInstructionsWithTableLookup(M, TableLookupVariablesOrdered, 1193 LookupTable); 1194 } 1195 1196 DenseMap<Function *, GlobalVariable *> KernelToCreatedDynamicLDS = 1197 lowerDynamicLDSVariables(M, LDSUsesInfo, 1198 KernelsThatIndirectlyAllocateDynamicLDS, 1199 DynamicVariables, OrderedKernels); 1200 1201 // All kernel frames have been allocated. Calculate and record the 1202 // addresses. 1203 1204 { 1205 const DataLayout &DL = M.getDataLayout(); 1206 1207 for (Function &Func : M.functions()) { 1208 if (Func.isDeclaration() || !isKernelLDS(&Func)) 1209 continue; 1210 1211 // All three of these are optional. The first variable is allocated at 1212 // zero. They are allocated by allocateKnownAddressLDSGlobal in the 1213 // following order: 1214 //{ 1215 // module.lds 1216 // alignment padding 1217 // kernel instance 1218 // alignment padding 1219 // dynamic lds variables 1220 //} 1221 1222 const bool AllocateModuleScopeStruct = 1223 MaybeModuleScopeStruct && !canElideModuleLDS(Func); 1224 1225 const bool AllocateKernelScopeStruct = 1226 KernelToReplacement.contains(&Func); 1227 1228 const bool AllocateDynamicVariable = 1229 KernelToCreatedDynamicLDS.contains(&Func); 1230 1231 uint32_t Offset = 0; 1232 1233 if (AllocateModuleScopeStruct) { 1234 // Allocated at zero, recorded once on construction, not once per 1235 // kernel 1236 Offset += DL.getTypeAllocSize(MaybeModuleScopeStruct->getValueType()); 1237 } 1238 1239 if (AllocateKernelScopeStruct) { 1240 GlobalVariable *KernelStruct = KernelToReplacement[&Func].SGV; 1241 1242 Offset = alignTo(Offset, AMDGPU::getAlign(DL, KernelStruct)); 1243 1244 recordLDSAbsoluteAddress(&M, KernelStruct, Offset); 1245 1246 Offset += DL.getTypeAllocSize(KernelStruct->getValueType()); 1247 1248 } 1249 1250 if (AllocateDynamicVariable) { 1251 GlobalVariable *DynamicVariable = KernelToCreatedDynamicLDS[&Func]; 1252 1253 Offset = alignTo(Offset, AMDGPU::getAlign(DL, DynamicVariable)); 1254 1255 recordLDSAbsoluteAddress(&M, DynamicVariable, Offset); 1256 } 1257 } 1258 } 1259 1260 for (auto &GV : make_early_inc_range(M.globals())) 1261 if (AMDGPU::isLDSVariableToLower(GV)) { 1262 // probably want to remove from used lists 1263 GV.removeDeadConstantUsers(); 1264 if (GV.use_empty()) 1265 GV.eraseFromParent(); 1266 } 1267 1268 return Changed; 1269 } 1270 1271 private: 1272 // Increase the alignment of LDS globals if necessary to maximise the chance 1273 // that we can use aligned LDS instructions to access them. 1274 static bool superAlignLDSGlobals(Module &M) { 1275 const DataLayout &DL = M.getDataLayout(); 1276 bool Changed = false; 1277 if (!SuperAlignLDSGlobals) { 1278 return Changed; 1279 } 1280 1281 for (auto &GV : M.globals()) { 1282 if (GV.getType()->getPointerAddressSpace() != AMDGPUAS::LOCAL_ADDRESS) { 1283 // Only changing alignment of LDS variables 1284 continue; 1285 } 1286 if (!GV.hasInitializer()) { 1287 // cuda/hip extern __shared__ variable, leave alignment alone 1288 continue; 1289 } 1290 1291 Align Alignment = AMDGPU::getAlign(DL, &GV); 1292 TypeSize GVSize = DL.getTypeAllocSize(GV.getValueType()); 1293 1294 if (GVSize > 8) { 1295 // We might want to use a b96 or b128 load/store 1296 Alignment = std::max(Alignment, Align(16)); 1297 } else if (GVSize > 4) { 1298 // We might want to use a b64 load/store 1299 Alignment = std::max(Alignment, Align(8)); 1300 } else if (GVSize > 2) { 1301 // We might want to use a b32 load/store 1302 Alignment = std::max(Alignment, Align(4)); 1303 } else if (GVSize > 1) { 1304 // We might want to use a b16 load/store 1305 Alignment = std::max(Alignment, Align(2)); 1306 } 1307 1308 if (Alignment != AMDGPU::getAlign(DL, &GV)) { 1309 Changed = true; 1310 GV.setAlignment(Alignment); 1311 } 1312 } 1313 return Changed; 1314 } 1315 1316 static LDSVariableReplacement createLDSVariableReplacement( 1317 Module &M, std::string VarName, 1318 DenseSet<GlobalVariable *> const &LDSVarsToTransform) { 1319 // Create a struct instance containing LDSVarsToTransform and map from those 1320 // variables to ConstantExprGEP 1321 // Variables may be introduced to meet alignment requirements. No aliasing 1322 // metadata is useful for these as they have no uses. Erased before return. 1323 1324 LLVMContext &Ctx = M.getContext(); 1325 const DataLayout &DL = M.getDataLayout(); 1326 assert(!LDSVarsToTransform.empty()); 1327 1328 SmallVector<OptimizedStructLayoutField, 8> LayoutFields; 1329 LayoutFields.reserve(LDSVarsToTransform.size()); 1330 { 1331 // The order of fields in this struct depends on the order of 1332 // varables in the argument which varies when changing how they 1333 // are identified, leading to spurious test breakage. 1334 std::vector<GlobalVariable *> Sorted(LDSVarsToTransform.begin(), 1335 LDSVarsToTransform.end()); 1336 llvm::sort(Sorted.begin(), Sorted.end(), 1337 [](const GlobalVariable *lhs, const GlobalVariable *rhs) { 1338 return lhs->getName() < rhs->getName(); 1339 }); 1340 for (GlobalVariable *GV : Sorted) { 1341 OptimizedStructLayoutField F(GV, 1342 DL.getTypeAllocSize(GV->getValueType()), 1343 AMDGPU::getAlign(DL, GV)); 1344 LayoutFields.emplace_back(F); 1345 } 1346 } 1347 1348 performOptimizedStructLayout(LayoutFields); 1349 1350 std::vector<GlobalVariable *> LocalVars; 1351 BitVector IsPaddingField; 1352 LocalVars.reserve(LDSVarsToTransform.size()); // will be at least this large 1353 IsPaddingField.reserve(LDSVarsToTransform.size()); 1354 { 1355 uint64_t CurrentOffset = 0; 1356 for (size_t I = 0; I < LayoutFields.size(); I++) { 1357 GlobalVariable *FGV = static_cast<GlobalVariable *>( 1358 const_cast<void *>(LayoutFields[I].Id)); 1359 Align DataAlign = LayoutFields[I].Alignment; 1360 1361 uint64_t DataAlignV = DataAlign.value(); 1362 if (uint64_t Rem = CurrentOffset % DataAlignV) { 1363 uint64_t Padding = DataAlignV - Rem; 1364 1365 // Append an array of padding bytes to meet alignment requested 1366 // Note (o + (a - (o % a)) ) % a == 0 1367 // (offset + Padding ) % align == 0 1368 1369 Type *ATy = ArrayType::get(Type::getInt8Ty(Ctx), Padding); 1370 LocalVars.push_back(new GlobalVariable( 1371 M, ATy, false, GlobalValue::InternalLinkage, UndefValue::get(ATy), 1372 "", nullptr, GlobalValue::NotThreadLocal, AMDGPUAS::LOCAL_ADDRESS, 1373 false)); 1374 IsPaddingField.push_back(true); 1375 CurrentOffset += Padding; 1376 } 1377 1378 LocalVars.push_back(FGV); 1379 IsPaddingField.push_back(false); 1380 CurrentOffset += LayoutFields[I].Size; 1381 } 1382 } 1383 1384 std::vector<Type *> LocalVarTypes; 1385 LocalVarTypes.reserve(LocalVars.size()); 1386 std::transform( 1387 LocalVars.cbegin(), LocalVars.cend(), std::back_inserter(LocalVarTypes), 1388 [](const GlobalVariable *V) -> Type * { return V->getValueType(); }); 1389 1390 StructType *LDSTy = StructType::create(Ctx, LocalVarTypes, VarName + ".t"); 1391 1392 Align StructAlign = AMDGPU::getAlign(DL, LocalVars[0]); 1393 1394 GlobalVariable *SGV = new GlobalVariable( 1395 M, LDSTy, false, GlobalValue::InternalLinkage, UndefValue::get(LDSTy), 1396 VarName, nullptr, GlobalValue::NotThreadLocal, AMDGPUAS::LOCAL_ADDRESS, 1397 false); 1398 SGV->setAlignment(StructAlign); 1399 1400 DenseMap<GlobalVariable *, Constant *> Map; 1401 Type *I32 = Type::getInt32Ty(Ctx); 1402 for (size_t I = 0; I < LocalVars.size(); I++) { 1403 GlobalVariable *GV = LocalVars[I]; 1404 Constant *GEPIdx[] = {ConstantInt::get(I32, 0), ConstantInt::get(I32, I)}; 1405 Constant *GEP = ConstantExpr::getGetElementPtr(LDSTy, SGV, GEPIdx, true); 1406 if (IsPaddingField[I]) { 1407 assert(GV->use_empty()); 1408 GV->eraseFromParent(); 1409 } else { 1410 Map[GV] = GEP; 1411 } 1412 } 1413 assert(Map.size() == LDSVarsToTransform.size()); 1414 return {SGV, std::move(Map)}; 1415 } 1416 1417 template <typename PredicateTy> 1418 static void replaceLDSVariablesWithStruct( 1419 Module &M, DenseSet<GlobalVariable *> const &LDSVarsToTransformArg, 1420 LDSVariableReplacement Replacement, PredicateTy Predicate) { 1421 LLVMContext &Ctx = M.getContext(); 1422 const DataLayout &DL = M.getDataLayout(); 1423 1424 // A hack... we need to insert the aliasing info in a predictable order for 1425 // lit tests. Would like to have them in a stable order already, ideally the 1426 // same order they get allocated, which might mean an ordered set container 1427 std::vector<GlobalVariable *> LDSVarsToTransform( 1428 LDSVarsToTransformArg.begin(), LDSVarsToTransformArg.end()); 1429 llvm::sort(LDSVarsToTransform.begin(), LDSVarsToTransform.end(), 1430 [](const GlobalVariable *lhs, const GlobalVariable *rhs) { 1431 return lhs->getName() < rhs->getName(); 1432 }); 1433 1434 // Create alias.scope and their lists. Each field in the new structure 1435 // does not alias with all other fields. 1436 SmallVector<MDNode *> AliasScopes; 1437 SmallVector<Metadata *> NoAliasList; 1438 const size_t NumberVars = LDSVarsToTransform.size(); 1439 if (NumberVars > 1) { 1440 MDBuilder MDB(Ctx); 1441 AliasScopes.reserve(NumberVars); 1442 MDNode *Domain = MDB.createAnonymousAliasScopeDomain(); 1443 for (size_t I = 0; I < NumberVars; I++) { 1444 MDNode *Scope = MDB.createAnonymousAliasScope(Domain); 1445 AliasScopes.push_back(Scope); 1446 } 1447 NoAliasList.append(&AliasScopes[1], AliasScopes.end()); 1448 } 1449 1450 // Replace uses of ith variable with a constantexpr to the corresponding 1451 // field of the instance that will be allocated by AMDGPUMachineFunction 1452 for (size_t I = 0; I < NumberVars; I++) { 1453 GlobalVariable *GV = LDSVarsToTransform[I]; 1454 Constant *GEP = Replacement.LDSVarsToConstantGEP[GV]; 1455 1456 GV->replaceUsesWithIf(GEP, Predicate); 1457 1458 APInt APOff(DL.getIndexTypeSizeInBits(GEP->getType()), 0); 1459 GEP->stripAndAccumulateInBoundsConstantOffsets(DL, APOff); 1460 uint64_t Offset = APOff.getZExtValue(); 1461 1462 Align A = 1463 commonAlignment(Replacement.SGV->getAlign().valueOrOne(), Offset); 1464 1465 if (I) 1466 NoAliasList[I - 1] = AliasScopes[I - 1]; 1467 MDNode *NoAlias = 1468 NoAliasList.empty() ? nullptr : MDNode::get(Ctx, NoAliasList); 1469 MDNode *AliasScope = 1470 AliasScopes.empty() ? nullptr : MDNode::get(Ctx, {AliasScopes[I]}); 1471 1472 refineUsesAlignmentAndAA(GEP, A, DL, AliasScope, NoAlias); 1473 } 1474 } 1475 1476 static void refineUsesAlignmentAndAA(Value *Ptr, Align A, 1477 const DataLayout &DL, MDNode *AliasScope, 1478 MDNode *NoAlias, unsigned MaxDepth = 5) { 1479 if (!MaxDepth || (A == 1 && !AliasScope)) 1480 return; 1481 1482 for (User *U : Ptr->users()) { 1483 if (auto *I = dyn_cast<Instruction>(U)) { 1484 if (AliasScope && I->mayReadOrWriteMemory()) { 1485 MDNode *AS = I->getMetadata(LLVMContext::MD_alias_scope); 1486 AS = (AS ? MDNode::getMostGenericAliasScope(AS, AliasScope) 1487 : AliasScope); 1488 I->setMetadata(LLVMContext::MD_alias_scope, AS); 1489 1490 MDNode *NA = I->getMetadata(LLVMContext::MD_noalias); 1491 NA = (NA ? MDNode::intersect(NA, NoAlias) : NoAlias); 1492 I->setMetadata(LLVMContext::MD_noalias, NA); 1493 } 1494 } 1495 1496 if (auto *LI = dyn_cast<LoadInst>(U)) { 1497 LI->setAlignment(std::max(A, LI->getAlign())); 1498 continue; 1499 } 1500 if (auto *SI = dyn_cast<StoreInst>(U)) { 1501 if (SI->getPointerOperand() == Ptr) 1502 SI->setAlignment(std::max(A, SI->getAlign())); 1503 continue; 1504 } 1505 if (auto *AI = dyn_cast<AtomicRMWInst>(U)) { 1506 // None of atomicrmw operations can work on pointers, but let's 1507 // check it anyway in case it will or we will process ConstantExpr. 1508 if (AI->getPointerOperand() == Ptr) 1509 AI->setAlignment(std::max(A, AI->getAlign())); 1510 continue; 1511 } 1512 if (auto *AI = dyn_cast<AtomicCmpXchgInst>(U)) { 1513 if (AI->getPointerOperand() == Ptr) 1514 AI->setAlignment(std::max(A, AI->getAlign())); 1515 continue; 1516 } 1517 if (auto *GEP = dyn_cast<GetElementPtrInst>(U)) { 1518 unsigned BitWidth = DL.getIndexTypeSizeInBits(GEP->getType()); 1519 APInt Off(BitWidth, 0); 1520 if (GEP->getPointerOperand() == Ptr) { 1521 Align GA; 1522 if (GEP->accumulateConstantOffset(DL, Off)) 1523 GA = commonAlignment(A, Off.getLimitedValue()); 1524 refineUsesAlignmentAndAA(GEP, GA, DL, AliasScope, NoAlias, 1525 MaxDepth - 1); 1526 } 1527 continue; 1528 } 1529 if (auto *I = dyn_cast<Instruction>(U)) { 1530 if (I->getOpcode() == Instruction::BitCast || 1531 I->getOpcode() == Instruction::AddrSpaceCast) 1532 refineUsesAlignmentAndAA(I, A, DL, AliasScope, NoAlias, MaxDepth - 1); 1533 } 1534 } 1535 } 1536 }; 1537 1538 } // namespace 1539 char AMDGPULowerModuleLDS::ID = 0; 1540 1541 char &llvm::AMDGPULowerModuleLDSID = AMDGPULowerModuleLDS::ID; 1542 1543 INITIALIZE_PASS(AMDGPULowerModuleLDS, DEBUG_TYPE, 1544 "Lower uses of LDS variables from non-kernel functions", false, 1545 false) 1546 1547 ModulePass *llvm::createAMDGPULowerModuleLDSPass() { 1548 return new AMDGPULowerModuleLDS(); 1549 } 1550 1551 PreservedAnalyses AMDGPULowerModuleLDSPass::run(Module &M, 1552 ModuleAnalysisManager &) { 1553 return AMDGPULowerModuleLDS().runOnModule(M) ? PreservedAnalyses::none() 1554 : PreservedAnalyses::all(); 1555 } 1556