1 //===-- AMDGPULowerKernelArguments.cpp ------------------------------------------===// 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 /// \file This pass replaces accesses to kernel arguments with loads from 11 /// offsets from the kernarg base pointer. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #include "AMDGPU.h" 16 #include "AMDGPUSubtarget.h" 17 #include "AMDGPUTargetMachine.h" 18 #include "llvm/ADT/StringRef.h" 19 #include "llvm/Analysis/DivergenceAnalysis.h" 20 #include "llvm/Analysis/Loads.h" 21 #include "llvm/CodeGen/Passes.h" 22 #include "llvm/CodeGen/TargetPassConfig.h" 23 #include "llvm/IR/Attributes.h" 24 #include "llvm/IR/BasicBlock.h" 25 #include "llvm/IR/Constants.h" 26 #include "llvm/IR/DerivedTypes.h" 27 #include "llvm/IR/Function.h" 28 #include "llvm/IR/IRBuilder.h" 29 #include "llvm/IR/InstrTypes.h" 30 #include "llvm/IR/Instruction.h" 31 #include "llvm/IR/Instructions.h" 32 #include "llvm/IR/LLVMContext.h" 33 #include "llvm/IR/MDBuilder.h" 34 #include "llvm/IR/Metadata.h" 35 #include "llvm/IR/Operator.h" 36 #include "llvm/IR/Type.h" 37 #include "llvm/IR/Value.h" 38 #include "llvm/Pass.h" 39 #include "llvm/Support/Casting.h" 40 41 #define DEBUG_TYPE "amdgpu-lower-kernel-arguments" 42 43 using namespace llvm; 44 45 namespace { 46 47 class AMDGPULowerKernelArguments : public FunctionPass{ 48 public: 49 static char ID; 50 51 AMDGPULowerKernelArguments() : FunctionPass(ID) {} 52 53 bool runOnFunction(Function &F) override; 54 55 void getAnalysisUsage(AnalysisUsage &AU) const override { 56 AU.addRequired<TargetPassConfig>(); 57 AU.setPreservesAll(); 58 } 59 }; 60 61 } // end anonymous namespace 62 63 bool AMDGPULowerKernelArguments::runOnFunction(Function &F) { 64 CallingConv::ID CC = F.getCallingConv(); 65 if (CC != CallingConv::AMDGPU_KERNEL || F.arg_empty()) 66 return false; 67 68 auto &TPC = getAnalysis<TargetPassConfig>(); 69 70 const TargetMachine &TM = TPC.getTM<TargetMachine>(); 71 const SISubtarget &ST = TM.getSubtarget<SISubtarget>(F); 72 LLVMContext &Ctx = F.getParent()->getContext(); 73 const DataLayout &DL = F.getParent()->getDataLayout(); 74 BasicBlock &EntryBlock = *F.begin(); 75 IRBuilder<> Builder(&*EntryBlock.begin()); 76 77 const unsigned KernArgBaseAlign = 16; // FIXME: Increase if necessary 78 const uint64_t BaseOffset = ST.getExplicitKernelArgOffset(F); 79 80 // FIXME: Alignment is broken broken with explicit arg offset.; 81 const uint64_t TotalKernArgSize = ST.getKernArgSegmentSize(F); 82 if (TotalKernArgSize == 0) 83 return false; 84 85 CallInst *KernArgSegment = 86 Builder.CreateIntrinsic(Intrinsic::amdgcn_kernarg_segment_ptr, nullptr, 87 F.getName() + ".kernarg.segment"); 88 89 KernArgSegment->addAttribute(AttributeList::ReturnIndex, Attribute::NonNull); 90 KernArgSegment->addAttribute(AttributeList::ReturnIndex, 91 Attribute::getWithDereferenceableBytes(Ctx, TotalKernArgSize)); 92 93 unsigned AS = KernArgSegment->getType()->getPointerAddressSpace(); 94 unsigned MaxAlign = 1; 95 uint64_t ExplicitArgOffset = 0; 96 97 for (Argument &Arg : F.args()) { 98 Type *ArgTy = Arg.getType(); 99 unsigned Align = DL.getABITypeAlignment(ArgTy); 100 MaxAlign = std::max(Align, MaxAlign); 101 unsigned Size = DL.getTypeSizeInBits(ArgTy); 102 unsigned AllocSize = DL.getTypeAllocSize(ArgTy); 103 104 105 // Clover seems to always pad i8/i16 to i32, but doesn't properly align 106 // them? 107 // Make sure the struct elements have correct size and alignment for ext 108 // args. These seem to be padded up to 4-bytes but not correctly aligned. 109 bool IsExtArg = AllocSize < 32 && (Arg.hasZExtAttr() || Arg.hasSExtAttr()) && 110 !ST.isAmdHsaOS(); 111 if (IsExtArg) 112 AllocSize = 4; 113 114 uint64_t EltOffset = alignTo(ExplicitArgOffset, Align) + BaseOffset; 115 ExplicitArgOffset = alignTo(ExplicitArgOffset, Align) + AllocSize; 116 117 if (Arg.use_empty()) 118 continue; 119 120 if (PointerType *PT = dyn_cast<PointerType>(ArgTy)) { 121 // FIXME: Hack. We rely on AssertZext to be able to fold DS addressing 122 // modes on SI to know the high bits are 0 so pointer adds don't wrap. We 123 // can't represent this with range metadata because it's only allowed for 124 // integer types. 125 if (PT->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS && 126 ST.getGeneration() == AMDGPUSubtarget::SOUTHERN_ISLANDS) 127 continue; 128 129 // FIXME: We can replace this with equivalent alias.scope/noalias 130 // metadata, but this appears to be a lot of work. 131 if (Arg.hasNoAliasAttr()) 132 continue; 133 } 134 135 VectorType *VT = dyn_cast<VectorType>(ArgTy); 136 bool IsV3 = VT && VT->getNumElements() == 3; 137 VectorType *V4Ty = nullptr; 138 139 int64_t AlignDownOffset = alignDown(EltOffset, 4); 140 int64_t OffsetDiff = EltOffset - AlignDownOffset; 141 unsigned AdjustedAlign = MinAlign(KernArgBaseAlign, AlignDownOffset); 142 143 Value *ArgPtr; 144 if (Size < 32 && !ArgTy->isAggregateType()) { // FIXME: Handle aggregate types 145 // Since we don't have sub-dword scalar loads, avoid doing an extload by 146 // loading earlier than the argument address, and extracting the relevant 147 // bits. 148 // 149 // Additionally widen any sub-dword load to i32 even if suitably aligned, 150 // so that CSE between different argument loads works easily. 151 152 ArgPtr = Builder.CreateConstInBoundsGEP1_64( 153 KernArgSegment, 154 AlignDownOffset, 155 Arg.getName() + ".kernarg.offset.align.down"); 156 ArgPtr = Builder.CreateBitCast(ArgPtr, 157 Builder.getInt32Ty()->getPointerTo(AS), 158 ArgPtr->getName() + ".cast"); 159 } else { 160 ArgPtr = Builder.CreateConstInBoundsGEP1_64( 161 KernArgSegment, 162 AlignDownOffset, 163 Arg.getName() + ".kernarg.offset"); 164 ArgPtr = Builder.CreateBitCast(ArgPtr, ArgTy->getPointerTo(AS), 165 ArgPtr->getName() + ".cast"); 166 } 167 168 assert((!IsExtArg || !IsV3) && "incompatible situation"); 169 170 if (IsV3 && Size >= 32) { 171 V4Ty = VectorType::get(VT->getVectorElementType(), 4); 172 // Use the hack that clang uses to avoid SelectionDAG ruining v3 loads 173 ArgPtr = Builder.CreateBitCast(ArgPtr, V4Ty->getPointerTo(AS)); 174 } 175 176 LoadInst *Load = Builder.CreateAlignedLoad(ArgPtr, AdjustedAlign); 177 Load->setMetadata(LLVMContext::MD_invariant_load, MDNode::get(Ctx, {})); 178 179 MDBuilder MDB(Ctx); 180 181 if (isa<PointerType>(ArgTy)) { 182 if (Arg.hasNonNullAttr()) 183 Load->setMetadata(LLVMContext::MD_nonnull, MDNode::get(Ctx, {})); 184 185 uint64_t DerefBytes = Arg.getDereferenceableBytes(); 186 if (DerefBytes != 0) { 187 Load->setMetadata( 188 LLVMContext::MD_dereferenceable, 189 MDNode::get(Ctx, 190 MDB.createConstant( 191 ConstantInt::get(Builder.getInt64Ty(), DerefBytes)))); 192 } 193 194 uint64_t DerefOrNullBytes = Arg.getDereferenceableOrNullBytes(); 195 if (DerefOrNullBytes != 0) { 196 Load->setMetadata( 197 LLVMContext::MD_dereferenceable_or_null, 198 MDNode::get(Ctx, 199 MDB.createConstant(ConstantInt::get(Builder.getInt64Ty(), 200 DerefOrNullBytes)))); 201 } 202 203 unsigned ParamAlign = Arg.getParamAlignment(); 204 if (ParamAlign != 0) { 205 Load->setMetadata( 206 LLVMContext::MD_align, 207 MDNode::get(Ctx, 208 MDB.createConstant(ConstantInt::get(Builder.getInt64Ty(), 209 ParamAlign)))); 210 } 211 } 212 213 // TODO: Convert noalias arg to !noalias 214 215 if (Size < 32 && !ArgTy->isAggregateType()) { 216 if (IsExtArg && OffsetDiff == 0) { 217 Type *I32Ty = Builder.getInt32Ty(); 218 bool IsSext = Arg.hasSExtAttr(); 219 Metadata *LowAndHigh[] = { 220 ConstantAsMetadata::get( 221 ConstantInt::get(I32Ty, IsSext ? minIntN(Size) : 0)), 222 ConstantAsMetadata::get( 223 ConstantInt::get(I32Ty, 224 IsSext ? maxIntN(Size) + 1 : maxUIntN(Size) + 1)) 225 }; 226 227 Load->setMetadata(LLVMContext::MD_range, MDNode::get(Ctx, LowAndHigh)); 228 } 229 230 Value *ExtractBits = OffsetDiff == 0 ? 231 Load : Builder.CreateLShr(Load, OffsetDiff * 8); 232 233 IntegerType *ArgIntTy = Builder.getIntNTy(Size); 234 Value *Trunc = Builder.CreateTrunc(ExtractBits, ArgIntTy); 235 Value *NewVal = Builder.CreateBitCast(Trunc, ArgTy, 236 Arg.getName() + ".load"); 237 Arg.replaceAllUsesWith(NewVal); 238 } else if (IsV3) { 239 Value *Shuf = Builder.CreateShuffleVector(Load, UndefValue::get(V4Ty), 240 {0, 1, 2}, 241 Arg.getName() + ".load"); 242 Arg.replaceAllUsesWith(Shuf); 243 } else { 244 Load->setName(Arg.getName() + ".load"); 245 Arg.replaceAllUsesWith(Load); 246 } 247 } 248 249 KernArgSegment->addAttribute( 250 AttributeList::ReturnIndex, 251 Attribute::getWithAlignment(Ctx, std::max(KernArgBaseAlign, MaxAlign))); 252 253 return true; 254 } 255 256 INITIALIZE_PASS_BEGIN(AMDGPULowerKernelArguments, DEBUG_TYPE, 257 "AMDGPU Lower Kernel Arguments", false, false) 258 INITIALIZE_PASS_END(AMDGPULowerKernelArguments, DEBUG_TYPE, "AMDGPU Lower Kernel Arguments", 259 false, false) 260 261 char AMDGPULowerKernelArguments::ID = 0; 262 263 FunctionPass *llvm::createAMDGPULowerKernelArgumentsPass() { 264 return new AMDGPULowerKernelArguments(); 265 } 266