lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp

   1 //===-- AMDGPUPromoteAlloca.cpp - Promote Allocas -------------------------===//
   2 //
   3 //                     The LLVM Compiler Infrastructure
   4 //
   5 // This file is distributed under the University of Illinois Open Source
   6 // License. See LICENSE.TXT for details.
   7 //
   8 //===----------------------------------------------------------------------===//
   9 //
  10 // This pass eliminates allocas by either converting them into vectors or
  11 // by migrating them to local address space.
  12 //
  13 //===----------------------------------------------------------------------===//
  14
  15 #include "AMDGPU.h"
  16 #include "AMDGPUSubtarget.h"
  17 #include "Utils/AMDGPUBaseInfo.h"
  18 #include "llvm/ADT/APInt.h"
  19 #include "llvm/ADT/None.h"
  20 #include "llvm/ADT/STLExtras.h"
  21 #include "llvm/ADT/StringRef.h"
  22 #include "llvm/ADT/Triple.h"
  23 #include "llvm/ADT/Twine.h"
  24 #include "llvm/Analysis/CaptureTracking.h"
  25 #include "llvm/Analysis/ValueTracking.h"
  26 #include "llvm/CodeGen/TargetPassConfig.h"
  27 #include "llvm/IR/Attributes.h"
  28 #include "llvm/IR/BasicBlock.h"
  29 #include "llvm/IR/Constant.h"
  30 #include "llvm/IR/Constants.h"
  31 #include "llvm/IR/DataLayout.h"
  32 #include "llvm/IR/DerivedTypes.h"
  33 #include "llvm/IR/Function.h"
  34 #include "llvm/IR/GlobalValue.h"
  35 #include "llvm/IR/GlobalVariable.h"
  36 #include "llvm/IR/IRBuilder.h"
  37 #include "llvm/IR/Instruction.h"
  38 #include "llvm/IR/Instructions.h"
  39 #include "llvm/IR/IntrinsicInst.h"
  40 #include "llvm/IR/Intrinsics.h"
  41 #include "llvm/IR/LLVMContext.h"
  42 #include "llvm/IR/Metadata.h"
  43 #include "llvm/IR/Module.h"
  44 #include "llvm/IR/Type.h"
  45 #include "llvm/IR/User.h"
  46 #include "llvm/IR/Value.h"
  47 #include "llvm/Pass.h"
  48 #include "llvm/Support/Casting.h"
  49 #include "llvm/Support/Debug.h"
  50 #include "llvm/Support/ErrorHandling.h"
  51 #include "llvm/Support/MathExtras.h"
  52 #include "llvm/Support/raw_ostream.h"
  53 #include "llvm/Target/TargetMachine.h"
  54 #include <algorithm>
  55 #include <cassert>
  56 #include <cstdint>
  57 #include <map>
  58 #include <tuple>
  59 #include <utility>
  60 #include <vector>
  61
  62 #define DEBUG_TYPE "amdgpu-promote-alloca"
  63
  64 using namespace llvm;
  65
  66 namespace {
  67
  68 static cl::opt<bool> DisablePromoteAllocaToVector(
  69   "disable-promote-alloca-to-vector",
  70   cl::desc("Disable promote alloca to vector"),
  71   cl::init(false));
  72
  73 // FIXME: This can create globals so should be a module pass.
  74 class AMDGPUPromoteAlloca : public FunctionPass {
  75 private:
  76   const TargetMachine *TM;
  77   Module *Mod = nullptr;
  78   const DataLayout *DL = nullptr;
  79
  80   // FIXME: This should be per-kernel.
  81   uint32_t LocalMemLimit = 0;
  82   uint32_t CurrentLocalMemUsage = 0;
  83
  84   bool IsAMDGCN = false;
  85   bool IsAMDHSA = false;
  86
  87   std::pair<Value *, Value *> getLocalSizeYZ(IRBuilder<> &Builder);
  88   Value *getWorkitemID(IRBuilder<> &Builder, unsigned N);
  89
  90   /// BaseAlloca is the alloca root the search started from.
  91   /// Val may be that alloca or a recursive user of it.
  92   bool collectUsesWithPtrTypes(Value *BaseAlloca,
  93                                Value *Val,
  94                                std::vector<Value*> &WorkList) const;
  95
  96   /// Val is a derived pointer from Alloca. OpIdx0/OpIdx1 are the operand
  97   /// indices to an instruction with 2 pointer inputs (e.g. select, icmp).
  98   /// Returns true if both operands are derived from the same alloca. Val should
  99   /// be the same value as one of the input operands of UseInst.
 100   bool binaryOpIsDerivedFromSameAlloca(Value *Alloca, Value *Val,
 101                                        Instruction *UseInst,
 102                                        int OpIdx0, int OpIdx1) const;
 103
 104   /// Check whether we have enough local memory for promotion.
 105   bool hasSufficientLocalMem(const Function &F);
 106
 107 public:
 108   static char ID;
 109
 110   AMDGPUPromoteAlloca() : FunctionPass(ID) {}
 111
 112   bool doInitialization(Module &M) override;
 113   bool runOnFunction(Function &F) override;
 114
 115   StringRef getPassName() const override { return "AMDGPU Promote Alloca"; }
 116
 117   bool handleAlloca(AllocaInst &I, bool SufficientLDS);
 118
 119   void getAnalysisUsage(AnalysisUsage &AU) const override {
 120     AU.setPreservesCFG();
 121     FunctionPass::getAnalysisUsage(AU);
 122   }
 123 };
 124
 125 } // end anonymous namespace
 126
 127 char AMDGPUPromoteAlloca::ID = 0;
 128
 129 INITIALIZE_PASS(AMDGPUPromoteAlloca, DEBUG_TYPE,
 130                 "AMDGPU promote alloca to vector or LDS", false, false)
 131
 132 char &llvm::AMDGPUPromoteAllocaID = AMDGPUPromoteAlloca::ID;
 133
 134 bool AMDGPUPromoteAlloca::doInitialization(Module &M) {
 135   Mod = &M;
 136   DL = &Mod->getDataLayout();
 137
 138   return false;
 139 }
 140
 141 bool AMDGPUPromoteAlloca::runOnFunction(Function &F) {
 142   if (skipFunction(F))
 143     return false;
 144
 145   if (auto *TPC = getAnalysisIfAvailable<TargetPassConfig>())
 146     TM = &TPC->getTM<TargetMachine>();
 147   else
 148     return false;
 149
 150   const Triple &TT = TM->getTargetTriple();
 151   IsAMDGCN = TT.getArch() == Triple::amdgcn;
 152   IsAMDHSA = TT.getOS() == Triple::AMDHSA;
 153
 154   const AMDGPUSubtarget &ST = AMDGPUSubtarget::get(*TM, F);
 155   if (!ST.isPromoteAllocaEnabled())
 156     return false;
 157
 158   bool SufficientLDS = hasSufficientLocalMem(F);
 159   bool Changed = false;
 160   BasicBlock &EntryBB = *F.begin();
 161   for (auto I = EntryBB.begin(), E = EntryBB.end(); I != E; ) {
 162     AllocaInst *AI = dyn_cast<AllocaInst>(I);
 163
 164     ++I;
 165     if (AI)
 166       Changed |= handleAlloca(*AI, SufficientLDS);
 167   }
 168
 169   return Changed;
 170 }
 171
 172 std::pair<Value *, Value *>
 173 AMDGPUPromoteAlloca::getLocalSizeYZ(IRBuilder<> &Builder) {
 174   const Function &F = *Builder.GetInsertBlock()->getParent();
 175   const AMDGPUSubtarget &ST = AMDGPUSubtarget::get(*TM, F);
 176
 177   if (!IsAMDHSA) {
 178     Function *LocalSizeYFn
 179       = Intrinsic::getDeclaration(Mod, Intrinsic::r600_read_local_size_y);
 180     Function *LocalSizeZFn
 181       = Intrinsic::getDeclaration(Mod, Intrinsic::r600_read_local_size_z);
 182
 183     CallInst *LocalSizeY = Builder.CreateCall(LocalSizeYFn, {});
 184     CallInst *LocalSizeZ = Builder.CreateCall(LocalSizeZFn, {});
 185
 186     ST.makeLIDRangeMetadata(LocalSizeY);
 187     ST.makeLIDRangeMetadata(LocalSizeZ);
 188
 189     return std::make_pair(LocalSizeY, LocalSizeZ);
 190   }
 191
 192   // We must read the size out of the dispatch pointer.
 193   assert(IsAMDGCN);
 194
 195   // We are indexing into this struct, and want to extract the workgroup_size_*
 196   // fields.
 197   //
 198   //   typedef struct hsa_kernel_dispatch_packet_s {
 199   //     uint16_t header;
 200   //     uint16_t setup;
 201   //     uint16_t workgroup_size_x ;
 202   //     uint16_t workgroup_size_y;
 203   //     uint16_t workgroup_size_z;
 204   //     uint16_t reserved0;
 205   //     uint32_t grid_size_x ;
 206   //     uint32_t grid_size_y ;
 207   //     uint32_t grid_size_z;
 208   //
 209   //     uint32_t private_segment_size;
 210   //     uint32_t group_segment_size;
 211   //     uint64_t kernel_object;
 212   //
 213   // #ifdef HSA_LARGE_MODEL
 214   //     void *kernarg_address;
 215   // #elif defined HSA_LITTLE_ENDIAN
 216   //     void *kernarg_address;
 217   //     uint32_t reserved1;
 218   // #else
 219   //     uint32_t reserved1;
 220   //     void *kernarg_address;
 221   // #endif
 222   //     uint64_t reserved2;
 223   //     hsa_signal_t completion_signal; // uint64_t wrapper
 224   //   } hsa_kernel_dispatch_packet_t
 225   //
 226   Function *DispatchPtrFn
 227     = Intrinsic::getDeclaration(Mod, Intrinsic::amdgcn_dispatch_ptr);
 228
 229   CallInst *DispatchPtr = Builder.CreateCall(DispatchPtrFn, {});
 230   DispatchPtr->addAttribute(AttributeList::ReturnIndex, Attribute::NoAlias);
 231   DispatchPtr->addAttribute(AttributeList::ReturnIndex, Attribute::NonNull);
 232
 233   // Size of the dispatch packet struct.
 234   DispatchPtr->addDereferenceableAttr(AttributeList::ReturnIndex, 64);
 235
 236   Type *I32Ty = Type::getInt32Ty(Mod->getContext());
 237   Value *CastDispatchPtr = Builder.CreateBitCast(
 238     DispatchPtr, PointerType::get(I32Ty, AMDGPUAS::CONSTANT_ADDRESS));
 239
 240   // We could do a single 64-bit load here, but it's likely that the basic
 241   // 32-bit and extract sequence is already present, and it is probably easier
 242   // to CSE this. The loads should be mergable later anyway.
 243   Value *GEPXY = Builder.CreateConstInBoundsGEP1_64(CastDispatchPtr, 1);
 244   LoadInst *LoadXY = Builder.CreateAlignedLoad(GEPXY, 4);
 245
 246   Value *GEPZU = Builder.CreateConstInBoundsGEP1_64(CastDispatchPtr, 2);
 247   LoadInst *LoadZU = Builder.CreateAlignedLoad(GEPZU, 4);
 248
 249   MDNode *MD = MDNode::get(Mod->getContext(), None);
 250   LoadXY->setMetadata(LLVMContext::MD_invariant_load, MD);
 251   LoadZU->setMetadata(LLVMContext::MD_invariant_load, MD);
 252   ST.makeLIDRangeMetadata(LoadZU);
 253
 254   // Extract y component. Upper half of LoadZU should be zero already.
 255   Value *Y = Builder.CreateLShr(LoadXY, 16);
 256
 257   return std::make_pair(Y, LoadZU);
 258 }
 259
 260 Value *AMDGPUPromoteAlloca::getWorkitemID(IRBuilder<> &Builder, unsigned N) {
 261   const AMDGPUSubtarget &ST =
 262       AMDGPUSubtarget::get(*TM, *Builder.GetInsertBlock()->getParent());
 263   Intrinsic::ID IntrID = Intrinsic::ID::not_intrinsic;
 264
 265   switch (N) {
 266   case 0:
 267     IntrID = IsAMDGCN ? Intrinsic::amdgcn_workitem_id_x
 268       : Intrinsic::r600_read_tidig_x;
 269     break;
 270   case 1:
 271     IntrID = IsAMDGCN ? Intrinsic::amdgcn_workitem_id_y
 272       : Intrinsic::r600_read_tidig_y;
 273     break;
 274
 275   case 2:
 276     IntrID = IsAMDGCN ? Intrinsic::amdgcn_workitem_id_z
 277       : Intrinsic::r600_read_tidig_z;
 278     break;
 279   default:
 280     llvm_unreachable("invalid dimension");
 281   }
 282
 283   Function *WorkitemIdFn = Intrinsic::getDeclaration(Mod, IntrID);
 284   CallInst *CI = Builder.CreateCall(WorkitemIdFn);
 285   ST.makeLIDRangeMetadata(CI);
 286
 287   return CI;
 288 }
 289
 290 static VectorType *arrayTypeToVecType(ArrayType *ArrayTy) {
 291   return VectorType::get(ArrayTy->getElementType(),
 292                          ArrayTy->getNumElements());
 293 }
 294
 295 static Value *
 296 calculateVectorIndex(Value *Ptr,
 297                      const std::map<GetElementPtrInst *, Value *> &GEPIdx) {
 298   GetElementPtrInst *GEP = cast<GetElementPtrInst>(Ptr);
 299
 300   auto I = GEPIdx.find(GEP);
 301   return I == GEPIdx.end() ? nullptr : I->second;
 302 }
 303
 304 static Value* GEPToVectorIndex(GetElementPtrInst *GEP) {
 305   // FIXME we only support simple cases
 306   if (GEP->getNumOperands() != 3)
 307     return nullptr;
 308
 309   ConstantInt *I0 = dyn_cast<ConstantInt>(GEP->getOperand(1));
 310   if (!I0 || !I0->isZero())
 311     return nullptr;
 312
 313   return GEP->getOperand(2);
 314 }
 315
 316 // Not an instruction handled below to turn into a vector.
 317 //
 318 // TODO: Check isTriviallyVectorizable for calls and handle other
 319 // instructions.
 320 static bool canVectorizeInst(Instruction *Inst, User *User) {
 321   switch (Inst->getOpcode()) {
 322   case Instruction::Load: {
 323     // Currently only handle the case where the Pointer Operand is a GEP.
 324     // Also we could not vectorize volatile or atomic loads.
 325     LoadInst *LI = cast<LoadInst>(Inst);
 326     return isa<GetElementPtrInst>(LI->getPointerOperand()) && LI->isSimple();
 327   }
 328   case Instruction::BitCast:
 329     return true;
 330   case Instruction::Store: {
 331     // Must be the stored pointer operand, not a stored value, plus
 332     // since it should be canonical form, the User should be a GEP.
 333     // Also we could not vectorize volatile or atomic stores.
 334     StoreInst *SI = cast<StoreInst>(Inst);
 335     return (SI->getPointerOperand() == User) && isa<GetElementPtrInst>(User) && SI->isSimple();
 336   }
 337   default:
 338     return false;
 339   }
 340 }
 341
 342 static bool tryPromoteAllocaToVector(AllocaInst *Alloca) {
 343
 344   if (DisablePromoteAllocaToVector) {
 345     LLVM_DEBUG(dbgs() << "  Promotion alloca to vector is disabled\n");
 346     return false;
 347   }
 348
 349   ArrayType *AllocaTy = dyn_cast<ArrayType>(Alloca->getAllocatedType());
 350
 351   LLVM_DEBUG(dbgs() << "Alloca candidate for vectorization\n");
 352
 353   // FIXME: There is no reason why we can't support larger arrays, we
 354   // are just being conservative for now.
 355   // FIXME: We also reject alloca's of the form [ 2 x [ 2 x i32 ]] or equivalent. Potentially these
 356   // could also be promoted but we don't currently handle this case
 357   if (!AllocaTy ||
 358       AllocaTy->getNumElements() > 16 ||
 359       AllocaTy->getNumElements() < 2 ||
 360       !VectorType::isValidElementType(AllocaTy->getElementType())) {
 361     LLVM_DEBUG(dbgs() << "  Cannot convert type to vector\n");
 362     return false;
 363   }
 364
 365   std::map<GetElementPtrInst*, Value*> GEPVectorIdx;
 366   std::vector<Value*> WorkList;
 367   for (User *AllocaUser : Alloca->users()) {
 368     GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(AllocaUser);
 369     if (!GEP) {
 370       if (!canVectorizeInst(cast<Instruction>(AllocaUser), Alloca))
 371         return false;
 372
 373       WorkList.push_back(AllocaUser);
 374       continue;
 375     }
 376
 377     Value *Index = GEPToVectorIndex(GEP);
 378
 379     // If we can't compute a vector index from this GEP, then we can't
 380     // promote this alloca to vector.
 381     if (!Index) {
 382       LLVM_DEBUG(dbgs() << "  Cannot compute vector index for GEP " << *GEP
 383                         << '\n');
 384       return false;
 385     }
 386
 387     GEPVectorIdx[GEP] = Index;
 388     for (User *GEPUser : AllocaUser->users()) {
 389       if (!canVectorizeInst(cast<Instruction>(GEPUser), AllocaUser))
 390         return false;
 391
 392       WorkList.push_back(GEPUser);
 393     }
 394   }
 395
 396   VectorType *VectorTy = arrayTypeToVecType(AllocaTy);
 397
 398   LLVM_DEBUG(dbgs() << "  Converting alloca to vector " << *AllocaTy << " -> "
 399                     << *VectorTy << '\n');
 400
 401   for (Value *V : WorkList) {
 402     Instruction *Inst = cast<Instruction>(V);
 403     IRBuilder<> Builder(Inst);
 404     switch (Inst->getOpcode()) {
 405     case Instruction::Load: {
 406       Type *VecPtrTy = VectorTy->getPointerTo(AMDGPUAS::PRIVATE_ADDRESS);
 407       Value *Ptr = cast<LoadInst>(Inst)->getPointerOperand();
 408       Value *Index = calculateVectorIndex(Ptr, GEPVectorIdx);
 409
 410       Value *BitCast = Builder.CreateBitCast(Alloca, VecPtrTy);
 411       Value *VecValue = Builder.CreateLoad(BitCast);
 412       Value *ExtractElement = Builder.CreateExtractElement(VecValue, Index);
 413       Inst->replaceAllUsesWith(ExtractElement);
 414       Inst->eraseFromParent();
 415       break;
 416     }
 417     case Instruction::Store: {
 418       Type *VecPtrTy = VectorTy->getPointerTo(AMDGPUAS::PRIVATE_ADDRESS);
 419
 420       StoreInst *SI = cast<StoreInst>(Inst);
 421       Value *Ptr = SI->getPointerOperand();
 422       Value *Index = calculateVectorIndex(Ptr, GEPVectorIdx);
 423       Value *BitCast = Builder.CreateBitCast(Alloca, VecPtrTy);
 424       Value *VecValue = Builder.CreateLoad(BitCast);
 425       Value *NewVecValue = Builder.CreateInsertElement(VecValue,
 426                                                        SI->getValueOperand(),
 427                                                        Index);
 428       Builder.CreateStore(NewVecValue, BitCast);
 429       Inst->eraseFromParent();
 430       break;
 431     }
 432     case Instruction::BitCast:
 433     case Instruction::AddrSpaceCast:
 434       break;
 435
 436     default:
 437       llvm_unreachable("Inconsistency in instructions promotable to vector");
 438     }
 439   }
 440   return true;
 441 }
 442
 443 static bool isCallPromotable(CallInst *CI) {
 444   IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
 445   if (!II)
 446     return false;
 447
 448   switch (II->getIntrinsicID()) {
 449   case Intrinsic::memcpy:
 450   case Intrinsic::memmove:
 451   case Intrinsic::memset:
 452   case Intrinsic::lifetime_start:
 453   case Intrinsic::lifetime_end:
 454   case Intrinsic::invariant_start:
 455   case Intrinsic::invariant_end:
 456   case Intrinsic::launder_invariant_group:
 457   case Intrinsic::strip_invariant_group:
 458   case Intrinsic::objectsize:
 459     return true;
 460   default:
 461     return false;
 462   }
 463 }
 464
 465 bool AMDGPUPromoteAlloca::binaryOpIsDerivedFromSameAlloca(Value *BaseAlloca,
 466                                                           Value *Val,
 467                                                           Instruction *Inst,
 468                                                           int OpIdx0,
 469                                                           int OpIdx1) const {
 470   // Figure out which operand is the one we might not be promoting.
 471   Value *OtherOp = Inst->getOperand(OpIdx0);
 472   if (Val == OtherOp)
 473     OtherOp = Inst->getOperand(OpIdx1);
 474
 475   if (isa<ConstantPointerNull>(OtherOp))
 476     return true;
 477
 478   Value *OtherObj = GetUnderlyingObject(OtherOp, *DL);
 479   if (!isa<AllocaInst>(OtherObj))
 480     return false;
 481
 482   // TODO: We should be able to replace undefs with the right pointer type.
 483
 484   // TODO: If we know the other base object is another promotable
 485   // alloca, not necessarily this alloca, we can do this. The
 486   // important part is both must have the same address space at
 487   // the end.
 488   if (OtherObj != BaseAlloca) {
 489     LLVM_DEBUG(
 490         dbgs() << "Found a binary instruction with another alloca object\n");
 491     return false;
 492   }
 493
 494   return true;
 495 }
 496
 497 bool AMDGPUPromoteAlloca::collectUsesWithPtrTypes(
 498   Value *BaseAlloca,
 499   Value *Val,
 500   std::vector<Value*> &WorkList) const {
 501
 502   for (User *User : Val->users()) {
 503     if (is_contained(WorkList, User))
 504       continue;
 505
 506     if (CallInst *CI = dyn_cast<CallInst>(User)) {
 507       if (!isCallPromotable(CI))
 508         return false;
 509
 510       WorkList.push_back(User);
 511       continue;
 512     }
 513
 514     Instruction *UseInst = cast<Instruction>(User);
 515     if (UseInst->getOpcode() == Instruction::PtrToInt)
 516       return false;
 517
 518     if (LoadInst *LI = dyn_cast<LoadInst>(UseInst)) {
 519       if (LI->isVolatile())
 520         return false;
 521
 522       continue;
 523     }
 524
 525     if (StoreInst *SI = dyn_cast<StoreInst>(UseInst)) {
 526       if (SI->isVolatile())
 527         return false;
 528
 529       // Reject if the stored value is not the pointer operand.
 530       if (SI->getPointerOperand() != Val)
 531         return false;
 532     } else if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(UseInst)) {
 533       if (RMW->isVolatile())
 534         return false;
 535     } else if (AtomicCmpXchgInst *CAS = dyn_cast<AtomicCmpXchgInst>(UseInst)) {
 536       if (CAS->isVolatile())
 537         return false;
 538     }
 539
 540     // Only promote a select if we know that the other select operand
 541     // is from another pointer that will also be promoted.
 542     if (ICmpInst *ICmp = dyn_cast<ICmpInst>(UseInst)) {
 543       if (!binaryOpIsDerivedFromSameAlloca(BaseAlloca, Val, ICmp, 0, 1))
 544         return false;
 545
 546       // May need to rewrite constant operands.
 547       WorkList.push_back(ICmp);
 548     }
 549
 550     if (UseInst->getOpcode() == Instruction::AddrSpaceCast) {
 551       // Give up if the pointer may be captured.
 552       if (PointerMayBeCaptured(UseInst, true, true))
 553         return false;
 554       // Don't collect the users of this.
 555       WorkList.push_back(User);
 556       continue;
 557     }
 558
 559     if (!User->getType()->isPointerTy())
 560       continue;
 561
 562     if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(UseInst)) {
 563       // Be conservative if an address could be computed outside the bounds of
 564       // the alloca.
 565       if (!GEP->isInBounds())
 566         return false;
 567     }
 568
 569     // Only promote a select if we know that the other select operand is from
 570     // another pointer that will also be promoted.
 571     if (SelectInst *SI = dyn_cast<SelectInst>(UseInst)) {
 572       if (!binaryOpIsDerivedFromSameAlloca(BaseAlloca, Val, SI, 1, 2))
 573         return false;
 574     }
 575
 576     // Repeat for phis.
 577     if (PHINode *Phi = dyn_cast<PHINode>(UseInst)) {
 578       // TODO: Handle more complex cases. We should be able to replace loops
 579       // over arrays.
 580       switch (Phi->getNumIncomingValues()) {
 581       case 1:
 582         break;
 583       case 2:
 584         if (!binaryOpIsDerivedFromSameAlloca(BaseAlloca, Val, Phi, 0, 1))
 585           return false;
 586         break;
 587       default:
 588         return false;
 589       }
 590     }
 591
 592     WorkList.push_back(User);
 593     if (!collectUsesWithPtrTypes(BaseAlloca, User, WorkList))
 594       return false;
 595   }
 596
 597   return true;
 598 }
 599
 600 bool AMDGPUPromoteAlloca::hasSufficientLocalMem(const Function &F) {
 601
 602   FunctionType *FTy = F.getFunctionType();
 603   const AMDGPUSubtarget &ST = AMDGPUSubtarget::get(*TM, F);
 604
 605   // If the function has any arguments in the local address space, then it's
 606   // possible these arguments require the entire local memory space, so
 607   // we cannot use local memory in the pass.
 608   for (Type *ParamTy : FTy->params()) {
 609     PointerType *PtrTy = dyn_cast<PointerType>(ParamTy);
 610     if (PtrTy && PtrTy->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS) {
 611       LocalMemLimit = 0;
 612       LLVM_DEBUG(dbgs() << "Function has local memory argument. Promoting to "
 613                            "local memory disabled.\n");
 614       return false;
 615     }
 616   }
 617
 618   LocalMemLimit = ST.getLocalMemorySize();
 619   if (LocalMemLimit == 0)
 620     return false;
 621
 622   const DataLayout &DL = Mod->getDataLayout();
 623
 624   // Check how much local memory is being used by global objects
 625   CurrentLocalMemUsage = 0;
 626   for (GlobalVariable &GV : Mod->globals()) {
 627     if (GV.getType()->getAddressSpace() != AMDGPUAS::LOCAL_ADDRESS)
 628       continue;
 629
 630     for (const User *U : GV.users()) {
 631       const Instruction *Use = dyn_cast<Instruction>(U);
 632       if (!Use)
 633         continue;
 634
 635       if (Use->getParent()->getParent() == &F) {
 636         unsigned Align = GV.getAlignment();
 637         if (Align == 0)
 638           Align = DL.getABITypeAlignment(GV.getValueType());
 639
 640         // FIXME: Try to account for padding here. The padding is currently
 641         // determined from the inverse order of uses in the function. I'm not
 642         // sure if the use list order is in any way connected to this, so the
 643         // total reported size is likely incorrect.
 644         uint64_t AllocSize = DL.getTypeAllocSize(GV.getValueType());
 645         CurrentLocalMemUsage = alignTo(CurrentLocalMemUsage, Align);
 646         CurrentLocalMemUsage += AllocSize;
 647         break;
 648       }
 649     }
 650   }
 651
 652   unsigned MaxOccupancy = ST.getOccupancyWithLocalMemSize(CurrentLocalMemUsage,
 653                                                           F);
 654
 655   // Restrict local memory usage so that we don't drastically reduce occupancy,
 656   // unless it is already significantly reduced.
 657
 658   // TODO: Have some sort of hint or other heuristics to guess occupancy based
 659   // on other factors..
 660   unsigned OccupancyHint = ST.getWavesPerEU(F).second;
 661   if (OccupancyHint == 0)
 662     OccupancyHint = 7;
 663
 664   // Clamp to max value.
 665   OccupancyHint = std::min(OccupancyHint, ST.getMaxWavesPerEU());
 666
 667   // Check the hint but ignore it if it's obviously wrong from the existing LDS
 668   // usage.
 669   MaxOccupancy = std::min(OccupancyHint, MaxOccupancy);
 670
 671
 672   // Round up to the next tier of usage.
 673   unsigned MaxSizeWithWaveCount
 674     = ST.getMaxLocalMemSizeWithWaveCount(MaxOccupancy, F);
 675
 676   // Program is possibly broken by using more local mem than available.
 677   if (CurrentLocalMemUsage > MaxSizeWithWaveCount)
 678     return false;
 679
 680   LocalMemLimit = MaxSizeWithWaveCount;
 681
 682   LLVM_DEBUG(dbgs() << F.getName() << " uses " << CurrentLocalMemUsage
 683                     << " bytes of LDS\n"
 684                     << "  Rounding size to " << MaxSizeWithWaveCount
 685                     << " with a maximum occupancy of " << MaxOccupancy << '\n'
 686                     << " and " << (LocalMemLimit - CurrentLocalMemUsage)
 687                     << " available for promotion\n");
 688
 689   return true;
 690 }
 691
 692 // FIXME: Should try to pick the most likely to be profitable allocas first.
 693 bool AMDGPUPromoteAlloca::handleAlloca(AllocaInst &I, bool SufficientLDS) {
 694   // Array allocations are probably not worth handling, since an allocation of
 695   // the array type is the canonical form.
 696   if (!I.isStaticAlloca() || I.isArrayAllocation())
 697     return false;
 698
 699   IRBuilder<> Builder(&I);
 700
 701   // First try to replace the alloca with a vector
 702   Type *AllocaTy = I.getAllocatedType();
 703
 704   LLVM_DEBUG(dbgs() << "Trying to promote " << I << '\n');
 705
 706   if (tryPromoteAllocaToVector(&I))
 707     return true; // Promoted to vector.
 708
 709   const Function &ContainingFunction = *I.getParent()->getParent();
 710   CallingConv::ID CC = ContainingFunction.getCallingConv();
 711
 712   // Don't promote the alloca to LDS for shader calling conventions as the work
 713   // item ID intrinsics are not supported for these calling conventions.
 714   // Furthermore not all LDS is available for some of the stages.
 715   switch (CC) {
 716   case CallingConv::AMDGPU_KERNEL:
 717   case CallingConv::SPIR_KERNEL:
 718     break;
 719   default:
 720     LLVM_DEBUG(
 721         dbgs()
 722         << " promote alloca to LDS not supported with calling convention.\n");
 723     return false;
 724   }
 725
 726   // Not likely to have sufficient local memory for promotion.
 727   if (!SufficientLDS)
 728     return false;
 729
 730   const AMDGPUSubtarget &ST = AMDGPUSubtarget::get(*TM, ContainingFunction);
 731   unsigned WorkGroupSize = ST.getFlatWorkGroupSizes(ContainingFunction).second;
 732
 733   const DataLayout &DL = Mod->getDataLayout();
 734
 735   unsigned Align = I.getAlignment();
 736   if (Align == 0)
 737     Align = DL.getABITypeAlignment(I.getAllocatedType());
 738
 739   // FIXME: This computed padding is likely wrong since it depends on inverse
 740   // usage order.
 741   //
 742   // FIXME: It is also possible that if we're allowed to use all of the memory
 743   // could could end up using more than the maximum due to alignment padding.
 744
 745   uint32_t NewSize = alignTo(CurrentLocalMemUsage, Align);
 746   uint32_t AllocSize = WorkGroupSize * DL.getTypeAllocSize(AllocaTy);
 747   NewSize += AllocSize;
 748
 749   if (NewSize > LocalMemLimit) {
 750     LLVM_DEBUG(dbgs() << "  " << AllocSize
 751                       << " bytes of local memory not available to promote\n");
 752     return false;
 753   }
 754
 755   CurrentLocalMemUsage = NewSize;
 756
 757   std::vector<Value*> WorkList;
 758
 759   if (!collectUsesWithPtrTypes(&I, &I, WorkList)) {
 760     LLVM_DEBUG(dbgs() << " Do not know how to convert all uses\n");
 761     return false;
 762   }
 763
 764   LLVM_DEBUG(dbgs() << "Promoting alloca to local memory\n");
 765
 766   Function *F = I.getParent()->getParent();
 767
 768   Type *GVTy = ArrayType::get(I.getAllocatedType(), WorkGroupSize);
 769   GlobalVariable *GV = new GlobalVariable(
 770       *Mod, GVTy, false, GlobalValue::InternalLinkage,
 771       UndefValue::get(GVTy),
 772       Twine(F->getName()) + Twine('.') + I.getName(),
 773       nullptr,
 774       GlobalVariable::NotThreadLocal,
 775       AMDGPUAS::LOCAL_ADDRESS);
 776   GV->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
 777   GV->setAlignment(I.getAlignment());
 778
 779   Value *TCntY, *TCntZ;
 780
 781   std::tie(TCntY, TCntZ) = getLocalSizeYZ(Builder);
 782   Value *TIdX = getWorkitemID(Builder, 0);
 783   Value *TIdY = getWorkitemID(Builder, 1);
 784   Value *TIdZ = getWorkitemID(Builder, 2);
 785
 786   Value *Tmp0 = Builder.CreateMul(TCntY, TCntZ, "", true, true);
 787   Tmp0 = Builder.CreateMul(Tmp0, TIdX);
 788   Value *Tmp1 = Builder.CreateMul(TIdY, TCntZ, "", true, true);
 789   Value *TID = Builder.CreateAdd(Tmp0, Tmp1);
 790   TID = Builder.CreateAdd(TID, TIdZ);
 791
 792   Value *Indices[] = {
 793     Constant::getNullValue(Type::getInt32Ty(Mod->getContext())),
 794     TID
 795   };
 796
 797   Value *Offset = Builder.CreateInBoundsGEP(GVTy, GV, Indices);
 798   I.mutateType(Offset->getType());
 799   I.replaceAllUsesWith(Offset);
 800   I.eraseFromParent();
 801
 802   for (Value *V : WorkList) {
 803     CallInst *Call = dyn_cast<CallInst>(V);
 804     if (!Call) {
 805       if (ICmpInst *CI = dyn_cast<ICmpInst>(V)) {
 806         Value *Src0 = CI->getOperand(0);
 807         Type *EltTy = Src0->getType()->getPointerElementType();
 808         PointerType *NewTy = PointerType::get(EltTy, AMDGPUAS::LOCAL_ADDRESS);
 809
 810         if (isa<ConstantPointerNull>(CI->getOperand(0)))
 811           CI->setOperand(0, ConstantPointerNull::get(NewTy));
 812
 813         if (isa<ConstantPointerNull>(CI->getOperand(1)))
 814           CI->setOperand(1, ConstantPointerNull::get(NewTy));
 815
 816         continue;
 817       }
 818
 819       // The operand's value should be corrected on its own and we don't want to
 820       // touch the users.
 821       if (isa<AddrSpaceCastInst>(V))
 822         continue;
 823
 824       Type *EltTy = V->getType()->getPointerElementType();
 825       PointerType *NewTy = PointerType::get(EltTy, AMDGPUAS::LOCAL_ADDRESS);
 826
 827       // FIXME: It doesn't really make sense to try to do this for all
 828       // instructions.
 829       V->mutateType(NewTy);
 830
 831       // Adjust the types of any constant operands.
 832       if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
 833         if (isa<ConstantPointerNull>(SI->getOperand(1)))
 834           SI->setOperand(1, ConstantPointerNull::get(NewTy));
 835
 836         if (isa<ConstantPointerNull>(SI->getOperand(2)))
 837           SI->setOperand(2, ConstantPointerNull::get(NewTy));
 838       } else if (PHINode *Phi = dyn_cast<PHINode>(V)) {
 839         for (unsigned I = 0, E = Phi->getNumIncomingValues(); I != E; ++I) {
 840           if (isa<ConstantPointerNull>(Phi->getIncomingValue(I)))
 841             Phi->setIncomingValue(I, ConstantPointerNull::get(NewTy));
 842         }
 843       }
 844
 845       continue;
 846     }
 847
 848     IntrinsicInst *Intr = cast<IntrinsicInst>(Call);
 849     Builder.SetInsertPoint(Intr);
 850     switch (Intr->getIntrinsicID()) {
 851     case Intrinsic::lifetime_start:
 852     case Intrinsic::lifetime_end:
 853       // These intrinsics are for address space 0 only
 854       Intr->eraseFromParent();
 855       continue;
 856     case Intrinsic::memcpy: {
 857       MemCpyInst *MemCpy = cast<MemCpyInst>(Intr);
 858       Builder.CreateMemCpy(MemCpy->getRawDest(), MemCpy->getDestAlignment(),
 859                            MemCpy->getRawSource(), MemCpy->getSourceAlignment(),
 860                            MemCpy->getLength(), MemCpy->isVolatile());
 861       Intr->eraseFromParent();
 862       continue;
 863     }
 864     case Intrinsic::memmove: {
 865       MemMoveInst *MemMove = cast<MemMoveInst>(Intr);
 866       Builder.CreateMemMove(MemMove->getRawDest(), MemMove->getDestAlignment(),
 867                             MemMove->getRawSource(), MemMove->getSourceAlignment(),
 868                             MemMove->getLength(), MemMove->isVolatile());
 869       Intr->eraseFromParent();
 870       continue;
 871     }
 872     case Intrinsic::memset: {
 873       MemSetInst *MemSet = cast<MemSetInst>(Intr);
 874       Builder.CreateMemSet(MemSet->getRawDest(), MemSet->getValue(),
 875                            MemSet->getLength(), MemSet->getDestAlignment(),
 876                            MemSet->isVolatile());
 877       Intr->eraseFromParent();
 878       continue;
 879     }
 880     case Intrinsic::invariant_start:
 881     case Intrinsic::invariant_end:
 882     case Intrinsic::launder_invariant_group:
 883     case Intrinsic::strip_invariant_group:
 884       Intr->eraseFromParent();
 885       // FIXME: I think the invariant marker should still theoretically apply,
 886       // but the intrinsics need to be changed to accept pointers with any
 887       // address space.
 888       continue;
 889     case Intrinsic::objectsize: {
 890       Value *Src = Intr->getOperand(0);
 891       Type *SrcTy = Src->getType()->getPointerElementType();
 892       Function *ObjectSize = Intrinsic::getDeclaration(Mod,
 893         Intrinsic::objectsize,
 894         { Intr->getType(), PointerType::get(SrcTy, AMDGPUAS::LOCAL_ADDRESS) }
 895       );
 896
 897       CallInst *NewCall = Builder.CreateCall(
 898           ObjectSize, {Src, Intr->getOperand(1), Intr->getOperand(2)});
 899       Intr->replaceAllUsesWith(NewCall);
 900       Intr->eraseFromParent();
 901       continue;
 902     }
 903     default:
 904       Intr->print(errs());
 905       llvm_unreachable("Don't know how to promote alloca intrinsic use.");
 906     }
 907   }
 908   return true;
 909 }
 910
 911 FunctionPass *llvm::createAMDGPUPromoteAlloca() {
 912   return new AMDGPUPromoteAlloca();
 913 }