1 //===-- X86InstrFMA3Info.cpp - X86 FMA3 Instruction 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 // This file contains the implementation of the classes providing information 11 // about existing X86 FMA3 opcodes, classifying and grouping them. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #include "X86InstrFMA3Info.h" 16 #include "X86InstrInfo.h" 17 #include "llvm/Support/ManagedStatic.h" 18 #include "llvm/Support/Threading.h" 19 #include <cassert> 20 #include <cstdint> 21 22 using namespace llvm; 23 24 #define FMA3BASE(X132, X213, X231, Attrs) \ 25 { { X132, X213, X231 }, Attrs }, 26 27 #define FMA3RA(R132, R213, R231, Attrs) \ 28 FMA3BASE(X86::R132, X86::R213, X86::R231, Attrs) 29 30 #define FMA3R(R132, R213, R231) \ 31 FMA3RA(R132, R213, R231, 0) 32 33 #define FMA3MA(M132, M213, M231, Attrs) \ 34 FMA3BASE(X86::M132, X86::M213, X86::M231, Attrs) 35 36 #define FMA3M(M132, M213, M231) \ 37 FMA3MA(M132, M213, M231, 0) 38 39 #define FMA3RMA(R132, R213, R231, M132, M213, M231, Attrs) \ 40 FMA3RA(R132, R213, R231, Attrs) \ 41 FMA3MA(M132, M213, M231, Attrs) 42 43 #define FMA3RM(R132, R213, R231, M132, M213, M231) \ 44 FMA3RMA(R132, R213, R231, M132, M213, M231, 0) 45 46 #define FMA3_AVX2_VECTOR_GROUP(Name) \ 47 FMA3RM(Name##132PSr, Name##213PSr, Name##231PSr, \ 48 Name##132PSm, Name##213PSm, Name##231PSm) \ 49 FMA3RM(Name##132PDr, Name##213PDr, Name##231PDr, \ 50 Name##132PDm, Name##213PDm, Name##231PDm) \ 51 FMA3RM(Name##132PSYr, Name##213PSYr, Name##231PSYr, \ 52 Name##132PSYm, Name##213PSYm, Name##231PSYm) \ 53 FMA3RM(Name##132PDYr, Name##213PDYr, Name##231PDYr, \ 54 Name##132PDYm, Name##213PDYm, Name##231PDYm) 55 56 #define FMA3_AVX2_SCALAR_GROUP(Name) \ 57 FMA3RM(Name##132SSr, Name##213SSr, Name##231SSr, \ 58 Name##132SSm, Name##213SSm, Name##231SSm) \ 59 FMA3RM(Name##132SDr, Name##213SDr, Name##231SDr, \ 60 Name##132SDm, Name##213SDm, Name##231SDm) \ 61 FMA3RMA(Name##132SSr_Int, Name##213SSr_Int, Name##231SSr_Int, \ 62 Name##132SSm_Int, Name##213SSm_Int, Name##231SSm_Int, \ 63 X86InstrFMA3Group::X86FMA3Intrinsic) \ 64 FMA3RMA(Name##132SDr_Int, Name##213SDr_Int, Name##231SDr_Int, \ 65 Name##132SDm_Int, Name##213SDm_Int, Name##231SDm_Int, \ 66 X86InstrFMA3Group::X86FMA3Intrinsic) 67 68 #define FMA3_AVX2_FULL_GROUP(Name) \ 69 FMA3_AVX2_VECTOR_GROUP(Name) \ 70 FMA3_AVX2_SCALAR_GROUP(Name) 71 72 #define FMA3_AVX512_VECTOR_GROUP(Name) \ 73 FMA3RM(Name##132PSZ128r, Name##213PSZ128r, Name##231PSZ128r, \ 74 Name##132PSZ128m, Name##213PSZ128m, Name##231PSZ128m) \ 75 FMA3RM(Name##132PDZ128r, Name##213PDZ128r, Name##231PDZ128r, \ 76 Name##132PDZ128m, Name##213PDZ128m, Name##231PDZ128m) \ 77 FMA3RM(Name##132PSZ256r, Name##213PSZ256r, Name##231PSZ256r, \ 78 Name##132PSZ256m, Name##213PSZ256m, Name##231PSZ256m) \ 79 FMA3RM(Name##132PDZ256r, Name##213PDZ256r, Name##231PDZ256r, \ 80 Name##132PDZ256m, Name##213PDZ256m, Name##231PDZ256m) \ 81 FMA3RM(Name##132PSZr, Name##213PSZr, Name##231PSZr, \ 82 Name##132PSZm, Name##213PSZm, Name##231PSZm) \ 83 FMA3RM(Name##132PDZr, Name##213PDZr, Name##231PDZr, \ 84 Name##132PDZm, Name##213PDZm, Name##231PDZm) \ 85 FMA3RMA(Name##132PSZ128rk, Name##213PSZ128rk, Name##231PSZ128rk, \ 86 Name##132PSZ128mk, Name##213PSZ128mk, Name##231PSZ128mk, \ 87 X86InstrFMA3Group::X86FMA3KMergeMasked) \ 88 FMA3RMA(Name##132PDZ128rk, Name##213PDZ128rk, Name##231PDZ128rk, \ 89 Name##132PDZ128mk, Name##213PDZ128mk, Name##231PDZ128mk, \ 90 X86InstrFMA3Group::X86FMA3KMergeMasked) \ 91 FMA3RMA(Name##132PSZ256rk, Name##213PSZ256rk, Name##231PSZ256rk, \ 92 Name##132PSZ256mk, Name##213PSZ256mk, Name##231PSZ256mk, \ 93 X86InstrFMA3Group::X86FMA3KMergeMasked) \ 94 FMA3RMA(Name##132PDZ256rk, Name##213PDZ256rk, Name##231PDZ256rk, \ 95 Name##132PDZ256mk, Name##213PDZ256mk, Name##231PDZ256mk, \ 96 X86InstrFMA3Group::X86FMA3KMergeMasked) \ 97 FMA3RMA(Name##132PSZrk, Name##213PSZrk, Name##231PSZrk, \ 98 Name##132PSZmk, Name##213PSZmk, Name##231PSZmk, \ 99 X86InstrFMA3Group::X86FMA3KMergeMasked) \ 100 FMA3RMA(Name##132PDZrk, Name##213PDZrk, Name##231PDZrk, \ 101 Name##132PDZmk, Name##213PDZmk, Name##231PDZmk, \ 102 X86InstrFMA3Group::X86FMA3KMergeMasked) \ 103 FMA3RMA(Name##132PSZ128rkz, Name##213PSZ128rkz, Name##231PSZ128rkz, \ 104 Name##132PSZ128mkz, Name##213PSZ128mkz, Name##231PSZ128mkz, \ 105 X86InstrFMA3Group::X86FMA3KZeroMasked) \ 106 FMA3RMA(Name##132PDZ128rkz, Name##213PDZ128rkz, Name##231PDZ128rkz, \ 107 Name##132PDZ128mkz, Name##213PDZ128mkz, Name##231PDZ128mkz, \ 108 X86InstrFMA3Group::X86FMA3KZeroMasked) \ 109 FMA3RMA(Name##132PSZ256rkz, Name##213PSZ256rkz, Name##231PSZ256rkz, \ 110 Name##132PSZ256mkz, Name##213PSZ256mkz, Name##231PSZ256mkz, \ 111 X86InstrFMA3Group::X86FMA3KZeroMasked) \ 112 FMA3RMA(Name##132PDZ256rkz, Name##213PDZ256rkz, Name##231PDZ256rkz, \ 113 Name##132PDZ256mkz, Name##213PDZ256mkz, Name##231PDZ256mkz, \ 114 X86InstrFMA3Group::X86FMA3KZeroMasked) \ 115 FMA3RMA(Name##132PSZrkz, Name##213PSZrkz, Name##231PSZrkz, \ 116 Name##132PSZmkz, Name##213PSZmkz, Name##231PSZmkz, \ 117 X86InstrFMA3Group::X86FMA3KZeroMasked) \ 118 FMA3RMA(Name##132PDZrkz, Name##213PDZrkz, Name##231PDZrkz, \ 119 Name##132PDZmkz, Name##213PDZmkz, Name##231PDZmkz, \ 120 X86InstrFMA3Group::X86FMA3KZeroMasked) \ 121 FMA3R(Name##132PSZrb, Name##213PSZrb, Name##231PSZrb) \ 122 FMA3R(Name##132PDZrb, Name##213PDZrb, Name##231PDZrb) \ 123 FMA3RA(Name##132PSZrbk, Name##213PSZrbk, Name##231PSZrbk, \ 124 X86InstrFMA3Group::X86FMA3KMergeMasked) \ 125 FMA3RA(Name##132PDZrbk, Name##213PDZrbk, Name##231PDZrbk, \ 126 X86InstrFMA3Group::X86FMA3KMergeMasked) \ 127 FMA3RA(Name##132PSZrbkz, Name##213PSZrbkz, Name##231PSZrbkz, \ 128 X86InstrFMA3Group::X86FMA3KZeroMasked) \ 129 FMA3RA(Name##132PDZrbkz, Name##213PDZrbkz, Name##231PDZrbkz, \ 130 X86InstrFMA3Group::X86FMA3KZeroMasked) \ 131 FMA3M(Name##132PSZ128mb, Name##213PSZ128mb, Name##231PSZ128mb) \ 132 FMA3M(Name##132PDZ128mb, Name##213PDZ128mb, Name##231PDZ128mb) \ 133 FMA3M(Name##132PSZ256mb, Name##213PSZ256mb, Name##231PSZ256mb) \ 134 FMA3M(Name##132PDZ256mb, Name##213PDZ256mb, Name##231PDZ256mb) \ 135 FMA3M(Name##132PSZmb, Name##213PSZmb, Name##231PSZmb) \ 136 FMA3M(Name##132PDZmb, Name##213PDZmb, Name##231PDZmb) \ 137 FMA3MA(Name##132PSZ128mbk, Name##213PSZ128mbk, Name##231PSZ128mbk, \ 138 X86InstrFMA3Group::X86FMA3KMergeMasked) \ 139 FMA3MA(Name##132PDZ128mbk, Name##213PDZ128mbk, Name##231PDZ128mbk, \ 140 X86InstrFMA3Group::X86FMA3KMergeMasked) \ 141 FMA3MA(Name##132PSZ256mbk, Name##213PSZ256mbk, Name##231PSZ256mbk, \ 142 X86InstrFMA3Group::X86FMA3KMergeMasked) \ 143 FMA3MA(Name##132PDZ256mbk, Name##213PDZ256mbk, Name##231PDZ256mbk, \ 144 X86InstrFMA3Group::X86FMA3KMergeMasked) \ 145 FMA3MA(Name##132PSZmbk, Name##213PSZmbk, Name##231PSZmbk, \ 146 X86InstrFMA3Group::X86FMA3KMergeMasked) \ 147 FMA3MA(Name##132PDZmbk, Name##213PDZmbk, Name##231PDZmbk, \ 148 X86InstrFMA3Group::X86FMA3KMergeMasked) \ 149 FMA3MA(Name##132PSZ128mbkz, Name##213PSZ128mbkz, Name##231PSZ128mbkz, \ 150 X86InstrFMA3Group::X86FMA3KZeroMasked) \ 151 FMA3MA(Name##132PDZ128mbkz, Name##213PDZ128mbkz, Name##231PDZ128mbkz, \ 152 X86InstrFMA3Group::X86FMA3KZeroMasked) \ 153 FMA3MA(Name##132PSZ256mbkz, Name##213PSZ256mbkz, Name##231PSZ256mbkz, \ 154 X86InstrFMA3Group::X86FMA3KZeroMasked) \ 155 FMA3MA(Name##132PDZ256mbkz, Name##213PDZ256mbkz, Name##231PDZ256mbkz, \ 156 X86InstrFMA3Group::X86FMA3KZeroMasked) \ 157 FMA3MA(Name##132PSZmbkz, Name##213PSZmbkz, Name##231PSZmbkz, \ 158 X86InstrFMA3Group::X86FMA3KZeroMasked) \ 159 FMA3MA(Name##132PDZmbkz, Name##213PDZmbkz, Name##231PDZmbkz, \ 160 X86InstrFMA3Group::X86FMA3KZeroMasked) 161 162 #define FMA3_AVX512_SCALAR_GROUP(Name) \ 163 FMA3RM(Name##132SSZr, Name##213SSZr, Name##231SSZr, \ 164 Name##132SSZm, Name##213SSZm, Name##231SSZm) \ 165 FMA3RM(Name##132SDZr, Name##213SDZr, Name##231SDZr, \ 166 Name##132SDZm, Name##213SDZm, Name##231SDZm) \ 167 FMA3RMA(Name##132SSZr_Int, Name##213SSZr_Int, Name##231SSZr_Int, \ 168 Name##132SSZm_Int, Name##213SSZm_Int, Name##231SSZm_Int, \ 169 X86InstrFMA3Group::X86FMA3Intrinsic) \ 170 FMA3RMA(Name##132SDZr_Int, Name##213SDZr_Int, Name##231SDZr_Int, \ 171 Name##132SDZm_Int, Name##213SDZm_Int, Name##231SDZm_Int, \ 172 X86InstrFMA3Group::X86FMA3Intrinsic) \ 173 FMA3RMA(Name##132SSZr_Intk, Name##213SSZr_Intk, Name##231SSZr_Intk, \ 174 Name##132SSZm_Intk, Name##213SSZm_Intk, Name##231SSZm_Intk, \ 175 X86InstrFMA3Group::X86FMA3Intrinsic | \ 176 X86InstrFMA3Group::X86FMA3KMergeMasked) \ 177 FMA3RMA(Name##132SDZr_Intk, Name##213SDZr_Intk, Name##231SDZr_Intk, \ 178 Name##132SDZm_Intk, Name##213SDZm_Intk, Name##231SDZm_Intk, \ 179 X86InstrFMA3Group::X86FMA3Intrinsic | \ 180 X86InstrFMA3Group::X86FMA3KMergeMasked) \ 181 FMA3RMA(Name##132SSZr_Intkz, Name##213SSZr_Intkz, Name##231SSZr_Intkz, \ 182 Name##132SSZm_Intkz, Name##213SSZm_Intkz, Name##231SSZm_Intkz, \ 183 X86InstrFMA3Group::X86FMA3Intrinsic | \ 184 X86InstrFMA3Group::X86FMA3KZeroMasked) \ 185 FMA3RMA(Name##132SDZr_Intkz, Name##213SDZr_Intkz, Name##231SDZr_Intkz, \ 186 Name##132SDZm_Intkz, Name##213SDZm_Intkz, Name##231SDZm_Intkz, \ 187 X86InstrFMA3Group::X86FMA3Intrinsic | \ 188 X86InstrFMA3Group::X86FMA3KZeroMasked) \ 189 FMA3RA(Name##132SSZrb_Int, Name##213SSZrb_Int, Name##231SSZrb_Int, \ 190 X86InstrFMA3Group::X86FMA3Intrinsic) \ 191 FMA3RA(Name##132SDZrb_Int, Name##213SDZrb_Int, Name##231SDZrb_Int, \ 192 X86InstrFMA3Group::X86FMA3Intrinsic) \ 193 FMA3RA(Name##132SSZrb_Intk, Name##213SSZrb_Intk, Name##231SSZrb_Intk, \ 194 X86InstrFMA3Group::X86FMA3Intrinsic | \ 195 X86InstrFMA3Group::X86FMA3KMergeMasked) \ 196 FMA3RA(Name##132SDZrb_Intk, Name##213SDZrb_Intk, Name##231SDZrb_Intk, \ 197 X86InstrFMA3Group::X86FMA3Intrinsic | \ 198 X86InstrFMA3Group::X86FMA3KMergeMasked) \ 199 FMA3RA(Name##132SSZrb_Intkz, Name##213SSZrb_Intkz, Name##231SSZrb_Intkz, \ 200 X86InstrFMA3Group::X86FMA3Intrinsic | \ 201 X86InstrFMA3Group::X86FMA3KZeroMasked) \ 202 FMA3RA(Name##132SDZrb_Intkz, Name##213SDZrb_Intkz, Name##231SDZrb_Intkz, \ 203 X86InstrFMA3Group::X86FMA3Intrinsic | \ 204 X86InstrFMA3Group::X86FMA3KZeroMasked) 205 206 #define FMA3_AVX512_FULL_GROUP(Name) \ 207 FMA3_AVX512_VECTOR_GROUP(Name) \ 208 FMA3_AVX512_SCALAR_GROUP(Name) 209 210 static const X86InstrFMA3Group Groups[] = { 211 FMA3_AVX2_FULL_GROUP(VFMADD) 212 FMA3_AVX2_FULL_GROUP(VFMSUB) 213 FMA3_AVX2_FULL_GROUP(VFNMADD) 214 FMA3_AVX2_FULL_GROUP(VFNMSUB) 215 216 FMA3_AVX2_VECTOR_GROUP(VFMADDSUB) 217 FMA3_AVX2_VECTOR_GROUP(VFMSUBADD) 218 219 FMA3_AVX512_FULL_GROUP(VFMADD) 220 FMA3_AVX512_FULL_GROUP(VFMSUB) 221 FMA3_AVX512_FULL_GROUP(VFNMADD) 222 FMA3_AVX512_FULL_GROUP(VFNMSUB) 223 224 FMA3_AVX512_VECTOR_GROUP(VFMADDSUB) 225 FMA3_AVX512_VECTOR_GROUP(VFMSUBADD) 226 }; 227 228 namespace { 229 230 struct X86InstrFMA3Info { 231 /// A map that is used to find the group of FMA opcodes using any FMA opcode 232 /// from the group. 233 DenseMap<unsigned, const X86InstrFMA3Group *> OpcodeToGroup; 234 235 /// Constructor. Just creates an object of the class. 236 X86InstrFMA3Info() { 237 for (const X86InstrFMA3Group &G : Groups) { 238 OpcodeToGroup[G.Opcodes[0]] = &G; 239 OpcodeToGroup[G.Opcodes[1]] = &G; 240 OpcodeToGroup[G.Opcodes[2]] = &G; 241 } 242 } 243 }; 244 245 } 246 247 static ManagedStatic<X86InstrFMA3Info> X86FMA3InfoObj; 248 249 /// Returns a reference to a group of FMA3 opcodes to where the given 250 /// \p Opcode is included. If the given \p Opcode is not recognized as FMA3 251 /// and not included into any FMA3 group, then nullptr is returned. 252 const X86InstrFMA3Group *llvm::getFMA3Group(unsigned Opcode) { 253 auto &Map = X86FMA3InfoObj->OpcodeToGroup; 254 auto I = Map.find(Opcode); 255 if (I != Map.end()) 256 return I->second; 257 258 return nullptr; 259 } 260