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), AMDGPUASI(ST.getAMDGPUAS()) {} 34 35 // FIXME: This behaves strangely. If, for example, you have 32 load + stores, 36 // the first 16 loads will be interleaved with the stores, and the next 16 will 37 // be clustered as expected. It should really split into 2 16 store batches. 38 // 39 // Loads are clustered until this returns false, rather than trying to schedule 40 // groups of stores. This also means we have to deal with saying different 41 // address space loads should be clustered, and ones which might cause bank 42 // conflicts. 43 // 44 // This might be deprecated so it might not be worth that much effort to fix. 45 bool AMDGPUInstrInfo::shouldScheduleLoadsNear(SDNode *Load0, SDNode *Load1, 46 int64_t Offset0, int64_t Offset1, 47 unsigned NumLoads) const { 48 assert(Offset1 > Offset0 && 49 "Second offset should be larger than first offset!"); 50 // If we have less than 16 loads in a row, and the offsets are within 64 51 // bytes, then schedule together. 52 53 // A cacheline is 64 bytes (for global memory). 54 return (NumLoads <= 16 && (Offset1 - Offset0) < 64); 55 } 56 57 int AMDGPUInstrInfo::getMaskedMIMGOp(uint16_t Opcode, unsigned Channels) const { 58 switch (Channels) { 59 default: return Opcode; 60 case 1: return AMDGPU::getMaskedMIMGOp(Opcode, AMDGPU::Channels_1); 61 case 2: return AMDGPU::getMaskedMIMGOp(Opcode, AMDGPU::Channels_2); 62 case 3: return AMDGPU::getMaskedMIMGOp(Opcode, AMDGPU::Channels_3); 63 } 64 } 65 66 // This must be kept in sync with the SIEncodingFamily class in SIInstrInfo.td 67 enum SIEncodingFamily { 68 SI = 0, 69 VI = 1, 70 SDWA = 2, 71 SDWA9 = 3 72 }; 73 74 // Wrapper for Tablegen'd function. enum Subtarget is not defined in any 75 // header files, so we need to wrap it in a function that takes unsigned 76 // instead. 77 namespace llvm { 78 namespace AMDGPU { 79 static int getMCOpcode(uint16_t Opcode, unsigned Gen) { 80 return getMCOpcodeGen(Opcode, static_cast<Subtarget>(Gen)); 81 } 82 } 83 } 84 85 static SIEncodingFamily subtargetEncodingFamily(const AMDGPUSubtarget &ST) { 86 switch (ST.getGeneration()) { 87 case AMDGPUSubtarget::SOUTHERN_ISLANDS: 88 case AMDGPUSubtarget::SEA_ISLANDS: 89 return SIEncodingFamily::SI; 90 case AMDGPUSubtarget::VOLCANIC_ISLANDS: 91 case AMDGPUSubtarget::GFX9: 92 return SIEncodingFamily::VI; 93 94 // FIXME: This should never be called for r600 GPUs. 95 case AMDGPUSubtarget::R600: 96 case AMDGPUSubtarget::R700: 97 case AMDGPUSubtarget::EVERGREEN: 98 case AMDGPUSubtarget::NORTHERN_ISLANDS: 99 return SIEncodingFamily::SI; 100 } 101 102 llvm_unreachable("Unknown subtarget generation!"); 103 } 104 105 int AMDGPUInstrInfo::pseudoToMCOpcode(int Opcode) const { 106 SIEncodingFamily Gen = subtargetEncodingFamily(ST); 107 if (get(Opcode).TSFlags & SIInstrFlags::SDWA) 108 Gen = ST.getGeneration() == AMDGPUSubtarget::GFX9 ? SIEncodingFamily::SDWA9 109 : SIEncodingFamily::SDWA; 110 111 int MCOp = AMDGPU::getMCOpcode(Opcode, Gen); 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