1 //===-- AMDGPUSubtarget.cpp - AMDGPU Subtarget Information ----------------===// 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 11 /// \brief Implements the AMDGPU specific subclass of TargetSubtarget. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #include "AMDGPUSubtarget.h" 16 #include "R600ISelLowering.h" 17 #include "R600InstrInfo.h" 18 #include "R600MachineScheduler.h" 19 #include "SIFrameLowering.h" 20 #include "SIISelLowering.h" 21 #include "SIInstrInfo.h" 22 #include "SIMachineFunctionInfo.h" 23 #include "llvm/ADT/SmallString.h" 24 #include "llvm/CodeGen/MachineScheduler.h" 25 26 using namespace llvm; 27 28 #define DEBUG_TYPE "amdgpu-subtarget" 29 30 #define GET_SUBTARGETINFO_ENUM 31 #define GET_SUBTARGETINFO_TARGET_DESC 32 #define GET_SUBTARGETINFO_CTOR 33 #include "AMDGPUGenSubtargetInfo.inc" 34 35 AMDGPUSubtarget & 36 AMDGPUSubtarget::initializeSubtargetDependencies(const Triple &TT, 37 StringRef GPU, StringRef FS) { 38 // Determine default and user-specified characteristics 39 // On SI+, we want FP64 denormals to be on by default. FP32 denormals can be 40 // enabled, but some instructions do not respect them and they run at the 41 // double precision rate, so don't enable by default. 42 // 43 // We want to be able to turn these off, but making this a subtarget feature 44 // for SI has the unhelpful behavior that it unsets everything else if you 45 // disable it. 46 47 SmallString<256> FullFS("+promote-alloca,+fp64-denormals,"); 48 if (isAmdHsaOS()) // Turn on FlatForGlobal for HSA. 49 FullFS += "+flat-for-global,"; 50 FullFS += FS; 51 52 if (GPU == "" && TT.getArch() == Triple::amdgcn) 53 GPU = "SI"; 54 55 ParseSubtargetFeatures(GPU, FullFS); 56 57 // FIXME: I don't think think Evergreen has any useful support for 58 // denormals, but should be checked. Should we issue a warning somewhere 59 // if someone tries to enable these? 60 if (getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) { 61 FP32Denormals = false; 62 FP64Denormals = false; 63 } 64 return *this; 65 } 66 67 AMDGPUSubtarget::AMDGPUSubtarget(const Triple &TT, StringRef GPU, StringRef FS, 68 TargetMachine &TM) 69 : AMDGPUGenSubtargetInfo(TT, GPU, FS), DevName(GPU), 70 DumpCode(false), R600ALUInst(false), HasVertexCache(false), 71 TexVTXClauseSize(0), Gen(AMDGPUSubtarget::R600), FP64(false), 72 FP64Denormals(false), FP32Denormals(false), FastFMAF32(false), 73 HalfRate64Ops(false), CaymanISA(false), FlatAddressSpace(false), 74 FlatForGlobal(false), EnableIRStructurizer(true), 75 EnablePromoteAlloca(false), 76 EnableIfCvt(true), EnableLoadStoreOpt(false), 77 EnableUnsafeDSOffsetFolding(false), 78 EnableXNACK(false), 79 WavefrontSize(0), CFALUBug(false), 80 LocalMemorySize(0), 81 EnableVGPRSpilling(false), SGPRInitBug(false), IsGCN(false), 82 GCN1Encoding(false), GCN3Encoding(false), CIInsts(false), LDSBankCount(0), 83 IsaVersion(ISAVersion0_0_0), EnableHugeScratchBuffer(false), 84 EnableSIScheduler(false), FrameLowering(nullptr), 85 InstrItins(getInstrItineraryForCPU(GPU)), TargetTriple(TT) { 86 87 initializeSubtargetDependencies(TT, GPU, FS); 88 89 const unsigned MaxStackAlign = 64 * 16; // Maximum stack alignment (long16) 90 91 if (getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) { 92 InstrInfo.reset(new R600InstrInfo(*this)); 93 TLInfo.reset(new R600TargetLowering(TM, *this)); 94 95 // FIXME: Should have R600 specific FrameLowering 96 FrameLowering.reset(new AMDGPUFrameLowering( 97 TargetFrameLowering::StackGrowsUp, 98 MaxStackAlign, 99 0)); 100 } else { 101 InstrInfo.reset(new SIInstrInfo(*this)); 102 TLInfo.reset(new SITargetLowering(TM, *this)); 103 FrameLowering.reset(new SIFrameLowering( 104 TargetFrameLowering::StackGrowsUp, 105 MaxStackAlign, 106 0)); 107 } 108 } 109 110 unsigned AMDGPUSubtarget::getStackEntrySize() const { 111 assert(getGeneration() <= NORTHERN_ISLANDS); 112 switch(getWavefrontSize()) { 113 case 16: 114 return 8; 115 case 32: 116 return hasCaymanISA() ? 4 : 8; 117 case 64: 118 return 4; 119 default: 120 llvm_unreachable("Illegal wavefront size."); 121 } 122 } 123 124 unsigned AMDGPUSubtarget::getAmdKernelCodeChipID() const { 125 switch(getGeneration()) { 126 default: llvm_unreachable("ChipID unknown"); 127 case SEA_ISLANDS: return 12; 128 } 129 } 130 131 AMDGPU::IsaVersion AMDGPUSubtarget::getIsaVersion() const { 132 return AMDGPU::getIsaVersion(getFeatureBits()); 133 } 134 135 bool AMDGPUSubtarget::isVGPRSpillingEnabled( 136 const SIMachineFunctionInfo *MFI) const { 137 return MFI->getShaderType() == ShaderType::COMPUTE || EnableVGPRSpilling; 138 } 139 140 void AMDGPUSubtarget::overrideSchedPolicy(MachineSchedPolicy &Policy, 141 MachineInstr *begin, 142 MachineInstr *end, 143 unsigned NumRegionInstrs) const { 144 if (getGeneration() >= SOUTHERN_ISLANDS) { 145 146 // Track register pressure so the scheduler can try to decrease 147 // pressure once register usage is above the threshold defined by 148 // SIRegisterInfo::getRegPressureSetLimit() 149 Policy.ShouldTrackPressure = true; 150 151 // Enabling both top down and bottom up scheduling seems to give us less 152 // register spills than just using one of these approaches on its own. 153 Policy.OnlyTopDown = false; 154 Policy.OnlyBottomUp = false; 155 } 156 } 157 158