1 //===----------------------------------------------------------------------===// 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 #include "AMDGPUArgumentUsageInfo.h" 10 #include "AMDGPU.h" 11 #include "AMDGPUTargetMachine.h" 12 #include "MCTargetDesc/AMDGPUMCTargetDesc.h" 13 #include "SIRegisterInfo.h" 14 #include "llvm/CodeGen/TargetRegisterInfo.h" 15 #include "llvm/IR/Function.h" 16 #include "llvm/Support/NativeFormatting.h" 17 #include "llvm/Support/raw_ostream.h" 18 19 using namespace llvm; 20 21 #define DEBUG_TYPE "amdgpu-argument-reg-usage-info" 22 23 INITIALIZE_PASS(AMDGPUArgumentUsageInfo, DEBUG_TYPE, 24 "Argument Register Usage Information Storage", false, true) 25 26 void ArgDescriptor::print(raw_ostream &OS, 27 const TargetRegisterInfo *TRI) const { 28 if (!isSet()) { 29 OS << "<not set>\n"; 30 return; 31 } 32 33 if (isRegister()) 34 OS << "Reg " << printReg(getRegister(), TRI); 35 else 36 OS << "Stack offset " << getStackOffset(); 37 38 if (isMasked()) { 39 OS << " & "; 40 llvm::write_hex(OS, Mask, llvm::HexPrintStyle::PrefixLower); 41 } 42 43 OS << '\n'; 44 } 45 46 char AMDGPUArgumentUsageInfo::ID = 0; 47 48 const AMDGPUFunctionArgInfo AMDGPUArgumentUsageInfo::ExternFunctionInfo{}; 49 50 // Hardcoded registers from fixed function ABI 51 const AMDGPUFunctionArgInfo AMDGPUArgumentUsageInfo::FixedABIFunctionInfo 52 = AMDGPUFunctionArgInfo::fixedABILayout(); 53 54 bool AMDGPUArgumentUsageInfo::doInitialization(Module &M) { 55 return false; 56 } 57 58 bool AMDGPUArgumentUsageInfo::doFinalization(Module &M) { 59 ArgInfoMap.clear(); 60 return false; 61 } 62 63 void AMDGPUArgumentUsageInfo::print(raw_ostream &OS, const Module *M) const { 64 for (const auto &FI : ArgInfoMap) { 65 OS << "Arguments for " << FI.first->getName() << '\n' 66 << " PrivateSegmentBuffer: " << FI.second.PrivateSegmentBuffer 67 << " DispatchPtr: " << FI.second.DispatchPtr 68 << " QueuePtr: " << FI.second.QueuePtr 69 << " KernargSegmentPtr: " << FI.second.KernargSegmentPtr 70 << " DispatchID: " << FI.second.DispatchID 71 << " FlatScratchInit: " << FI.second.FlatScratchInit 72 << " PrivateSegmentSize: " << FI.second.PrivateSegmentSize 73 << " WorkGroupIDX: " << FI.second.WorkGroupIDX 74 << " WorkGroupIDY: " << FI.second.WorkGroupIDY 75 << " WorkGroupIDZ: " << FI.second.WorkGroupIDZ 76 << " WorkGroupInfo: " << FI.second.WorkGroupInfo 77 << " PrivateSegmentWaveByteOffset: " 78 << FI.second.PrivateSegmentWaveByteOffset 79 << " ImplicitBufferPtr: " << FI.second.ImplicitBufferPtr 80 << " ImplicitArgPtr: " << FI.second.ImplicitArgPtr 81 << " WorkItemIDX " << FI.second.WorkItemIDX 82 << " WorkItemIDY " << FI.second.WorkItemIDY 83 << " WorkItemIDZ " << FI.second.WorkItemIDZ 84 << '\n'; 85 } 86 } 87 88 std::tuple<const ArgDescriptor *, const TargetRegisterClass *, LLT> 89 AMDGPUFunctionArgInfo::getPreloadedValue( 90 AMDGPUFunctionArgInfo::PreloadedValue Value) const { 91 switch (Value) { 92 case AMDGPUFunctionArgInfo::PRIVATE_SEGMENT_BUFFER: { 93 return std::make_tuple(PrivateSegmentBuffer ? &PrivateSegmentBuffer 94 : nullptr, 95 &AMDGPU::SGPR_128RegClass, LLT::vector(4, 32)); 96 } 97 case AMDGPUFunctionArgInfo::IMPLICIT_BUFFER_PTR: 98 return std::make_tuple(ImplicitBufferPtr ? &ImplicitBufferPtr : nullptr, 99 &AMDGPU::SGPR_64RegClass, 100 LLT::pointer(AMDGPUAS::CONSTANT_ADDRESS, 64)); 101 case AMDGPUFunctionArgInfo::WORKGROUP_ID_X: 102 return std::make_tuple(WorkGroupIDX ? &WorkGroupIDX : nullptr, 103 &AMDGPU::SGPR_32RegClass, LLT::scalar(32)); 104 case AMDGPUFunctionArgInfo::WORKGROUP_ID_Y: 105 return std::make_tuple(WorkGroupIDY ? &WorkGroupIDY : nullptr, 106 &AMDGPU::SGPR_32RegClass, LLT::scalar(32)); 107 case AMDGPUFunctionArgInfo::WORKGROUP_ID_Z: 108 return std::make_tuple(WorkGroupIDZ ? &WorkGroupIDZ : nullptr, 109 &AMDGPU::SGPR_32RegClass, LLT::scalar(32)); 110 case AMDGPUFunctionArgInfo::PRIVATE_SEGMENT_WAVE_BYTE_OFFSET: 111 return std::make_tuple( 112 PrivateSegmentWaveByteOffset ? &PrivateSegmentWaveByteOffset : nullptr, 113 &AMDGPU::SGPR_32RegClass, LLT::scalar(32)); 114 case AMDGPUFunctionArgInfo::KERNARG_SEGMENT_PTR: 115 return std::make_tuple(KernargSegmentPtr ? &KernargSegmentPtr : nullptr, 116 &AMDGPU::SGPR_64RegClass, 117 LLT::pointer(AMDGPUAS::CONSTANT_ADDRESS, 64)); 118 case AMDGPUFunctionArgInfo::IMPLICIT_ARG_PTR: 119 return std::make_tuple(ImplicitArgPtr ? &ImplicitArgPtr : nullptr, 120 &AMDGPU::SGPR_64RegClass, 121 LLT::pointer(AMDGPUAS::CONSTANT_ADDRESS, 64)); 122 case AMDGPUFunctionArgInfo::DISPATCH_ID: 123 return std::make_tuple(DispatchID ? &DispatchID : nullptr, 124 &AMDGPU::SGPR_64RegClass, LLT::scalar(64)); 125 case AMDGPUFunctionArgInfo::FLAT_SCRATCH_INIT: 126 return std::make_tuple(FlatScratchInit ? &FlatScratchInit : nullptr, 127 &AMDGPU::SGPR_64RegClass, LLT::scalar(64)); 128 case AMDGPUFunctionArgInfo::DISPATCH_PTR: 129 return std::make_tuple(DispatchPtr ? &DispatchPtr : nullptr, 130 &AMDGPU::SGPR_64RegClass, 131 LLT::pointer(AMDGPUAS::CONSTANT_ADDRESS, 64)); 132 case AMDGPUFunctionArgInfo::QUEUE_PTR: 133 return std::make_tuple(QueuePtr ? &QueuePtr : nullptr, 134 &AMDGPU::SGPR_64RegClass, 135 LLT::pointer(AMDGPUAS::CONSTANT_ADDRESS, 64)); 136 case AMDGPUFunctionArgInfo::WORKITEM_ID_X: 137 return std::make_tuple(WorkItemIDX ? &WorkItemIDX : nullptr, 138 &AMDGPU::VGPR_32RegClass, LLT::scalar(32)); 139 case AMDGPUFunctionArgInfo::WORKITEM_ID_Y: 140 return std::make_tuple(WorkItemIDY ? &WorkItemIDY : nullptr, 141 &AMDGPU::VGPR_32RegClass, LLT::scalar(32)); 142 case AMDGPUFunctionArgInfo::WORKITEM_ID_Z: 143 return std::make_tuple(WorkItemIDZ ? &WorkItemIDZ : nullptr, 144 &AMDGPU::VGPR_32RegClass, LLT::scalar(32)); 145 } 146 llvm_unreachable("unexpected preloaded value type"); 147 } 148 149 constexpr AMDGPUFunctionArgInfo AMDGPUFunctionArgInfo::fixedABILayout() { 150 AMDGPUFunctionArgInfo AI; 151 AI.PrivateSegmentBuffer 152 = ArgDescriptor::createRegister(AMDGPU::SGPR0_SGPR1_SGPR2_SGPR3); 153 AI.DispatchPtr = ArgDescriptor::createRegister(AMDGPU::SGPR4_SGPR5); 154 AI.QueuePtr = ArgDescriptor::createRegister(AMDGPU::SGPR6_SGPR7); 155 156 // Do not pass kernarg segment pointer, only pass increment version in its 157 // place. 158 AI.ImplicitArgPtr = ArgDescriptor::createRegister(AMDGPU::SGPR8_SGPR9); 159 AI.DispatchID = ArgDescriptor::createRegister(AMDGPU::SGPR10_SGPR11); 160 161 // Skip FlatScratchInit/PrivateSegmentSize 162 AI.WorkGroupIDX = ArgDescriptor::createRegister(AMDGPU::SGPR12); 163 AI.WorkGroupIDY = ArgDescriptor::createRegister(AMDGPU::SGPR13); 164 AI.WorkGroupIDZ = ArgDescriptor::createRegister(AMDGPU::SGPR14); 165 166 const unsigned Mask = 0x3ff; 167 AI.WorkItemIDX = ArgDescriptor::createRegister(AMDGPU::VGPR31, Mask); 168 AI.WorkItemIDY = ArgDescriptor::createRegister(AMDGPU::VGPR31, Mask << 10); 169 AI.WorkItemIDZ = ArgDescriptor::createRegister(AMDGPU::VGPR31, Mask << 20); 170 return AI; 171 } 172 173 const AMDGPUFunctionArgInfo & 174 AMDGPUArgumentUsageInfo::lookupFuncArgInfo(const Function &F) const { 175 auto I = ArgInfoMap.find(&F); 176 if (I == ArgInfoMap.end()) { 177 if (AMDGPUTargetMachine::EnableFixedFunctionABI) 178 return FixedABIFunctionInfo; 179 180 // Without the fixed ABI, we assume no function has special inputs. 181 assert(F.isDeclaration()); 182 return ExternFunctionInfo; 183 } 184 185 return I->second; 186 } 187