1 //===-- AMDGPUInstrInfo.cpp - Base class for AMD GPU InstrInfo ------------===// 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 Implementation of the TargetInstrInfo class that is common to all 12 /// AMD GPUs. 13 // 14 //===----------------------------------------------------------------------===// 15 16 #include "AMDGPUInstrInfo.h" 17 #include "AMDGPURegisterInfo.h" 18 #include "AMDGPUTargetMachine.h" 19 #include "llvm/CodeGen/MachineFrameInfo.h" 20 #include "llvm/CodeGen/MachineInstrBuilder.h" 21 #include "llvm/CodeGen/MachineRegisterInfo.h" 22 23 using namespace llvm; 24 25 #define GET_INSTRINFO_CTOR_DTOR 26 #define GET_INSTRMAP_INFO 27 #include "AMDGPUGenInstrInfo.inc" 28 29 // Pin the vtable to this file. 30 void AMDGPUInstrInfo::anchor() {} 31 32 AMDGPUInstrInfo::AMDGPUInstrInfo(const AMDGPUSubtarget &ST) 33 : AMDGPUGenInstrInfo(-1, -1), ST(ST) {} 34 35 bool AMDGPUInstrInfo::enableClusterLoads() const { 36 return true; 37 } 38 39 bool AMDGPUInstrInfo::enableClusterStores() const { 40 return true; 41 } 42 43 // FIXME: This behaves strangely. If, for example, you have 32 load + stores, 44 // the first 16 loads will be interleaved with the stores, and the next 16 will 45 // be clustered as expected. It should really split into 2 16 store batches. 46 // 47 // Loads are clustered until this returns false, rather than trying to schedule 48 // groups of stores. This also means we have to deal with saying different 49 // address space loads should be clustered, and ones which might cause bank 50 // conflicts. 51 // 52 // This might be deprecated so it might not be worth that much effort to fix. 53 bool AMDGPUInstrInfo::shouldScheduleLoadsNear(SDNode *Load0, SDNode *Load1, 54 int64_t Offset0, int64_t Offset1, 55 unsigned NumLoads) const { 56 assert(Offset1 > Offset0 && 57 "Second offset should be larger than first offset!"); 58 // If we have less than 16 loads in a row, and the offsets are within 64 59 // bytes, then schedule together. 60 61 // A cacheline is 64 bytes (for global memory). 62 return (NumLoads <= 16 && (Offset1 - Offset0) < 64); 63 } 64 65 int AMDGPUInstrInfo::getMaskedMIMGOp(uint16_t Opcode, unsigned Channels) const { 66 switch (Channels) { 67 default: return Opcode; 68 case 1: return AMDGPU::getMaskedMIMGOp(Opcode, AMDGPU::Channels_1); 69 case 2: return AMDGPU::getMaskedMIMGOp(Opcode, AMDGPU::Channels_2); 70 case 3: return AMDGPU::getMaskedMIMGOp(Opcode, AMDGPU::Channels_3); 71 } 72 } 73 74 // This must be kept in sync with the SIEncodingFamily class in SIInstrInfo.td 75 enum SIEncodingFamily { 76 SI = 0, 77 VI = 1 78 }; 79 80 // Wrapper for Tablegen'd function. enum Subtarget is not defined in any 81 // header files, so we need to wrap it in a function that takes unsigned 82 // instead. 83 namespace llvm { 84 namespace AMDGPU { 85 static int getMCOpcode(uint16_t Opcode, unsigned Gen) { 86 return getMCOpcodeGen(Opcode, static_cast<Subtarget>(Gen)); 87 } 88 } 89 } 90 91 static SIEncodingFamily subtargetEncodingFamily(const AMDGPUSubtarget &ST) { 92 switch (ST.getGeneration()) { 93 case AMDGPUSubtarget::SOUTHERN_ISLANDS: 94 case AMDGPUSubtarget::SEA_ISLANDS: 95 return SIEncodingFamily::SI; 96 case AMDGPUSubtarget::VOLCANIC_ISLANDS: 97 return SIEncodingFamily::VI; 98 99 // FIXME: This should never be called for r600 GPUs. 100 case AMDGPUSubtarget::R600: 101 case AMDGPUSubtarget::R700: 102 case AMDGPUSubtarget::EVERGREEN: 103 case AMDGPUSubtarget::NORTHERN_ISLANDS: 104 return SIEncodingFamily::SI; 105 } 106 107 llvm_unreachable("Unknown subtarget generation!"); 108 } 109 110 int AMDGPUInstrInfo::pseudoToMCOpcode(int Opcode) const { 111 int MCOp = AMDGPU::getMCOpcode(Opcode, subtargetEncodingFamily(ST)); 112 113 // -1 means that Opcode is already a native instruction. 114 if (MCOp == -1) 115 return Opcode; 116 117 // (uint16_t)-1 means that Opcode is a pseudo instruction that has 118 // no encoding in the given subtarget generation. 119 if (MCOp == (uint16_t)-1) 120 return -1; 121 122 return MCOp; 123 } 124