1 //===- TargetLoweringBase.cpp - Implement the TargetLoweringBase class ----===// 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 implements the TargetLoweringBase class. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "llvm/ADT/BitVector.h" 15 #include "llvm/ADT/STLExtras.h" 16 #include "llvm/ADT/SmallVector.h" 17 #include "llvm/ADT/StringExtras.h" 18 #include "llvm/ADT/StringRef.h" 19 #include "llvm/ADT/Triple.h" 20 #include "llvm/ADT/Twine.h" 21 #include "llvm/CodeGen/Analysis.h" 22 #include "llvm/CodeGen/ISDOpcodes.h" 23 #include "llvm/CodeGen/MachineBasicBlock.h" 24 #include "llvm/CodeGen/MachineFrameInfo.h" 25 #include "llvm/CodeGen/MachineFunction.h" 26 #include "llvm/CodeGen/MachineInstr.h" 27 #include "llvm/CodeGen/MachineInstrBuilder.h" 28 #include "llvm/CodeGen/MachineMemOperand.h" 29 #include "llvm/CodeGen/MachineOperand.h" 30 #include "llvm/CodeGen/MachineRegisterInfo.h" 31 #include "llvm/CodeGen/MachineValueType.h" 32 #include "llvm/CodeGen/RuntimeLibcalls.h" 33 #include "llvm/CodeGen/StackMaps.h" 34 #include "llvm/CodeGen/ValueTypes.h" 35 #include "llvm/IR/Attributes.h" 36 #include "llvm/IR/CallingConv.h" 37 #include "llvm/IR/DataLayout.h" 38 #include "llvm/IR/DerivedTypes.h" 39 #include "llvm/IR/Function.h" 40 #include "llvm/IR/GlobalValue.h" 41 #include "llvm/IR/GlobalVariable.h" 42 #include "llvm/IR/IRBuilder.h" 43 #include "llvm/IR/Module.h" 44 #include "llvm/IR/Type.h" 45 #include "llvm/Support/BranchProbability.h" 46 #include "llvm/Support/Casting.h" 47 #include "llvm/Support/CommandLine.h" 48 #include "llvm/Support/Compiler.h" 49 #include "llvm/Support/ErrorHandling.h" 50 #include "llvm/Support/MathExtras.h" 51 #include "llvm/Target/TargetLowering.h" 52 #include "llvm/Target/TargetMachine.h" 53 #include "llvm/Target/TargetOpcodes.h" 54 #include "llvm/Target/TargetRegisterInfo.h" 55 #include <algorithm> 56 #include <cassert> 57 #include <cstring> 58 #include <cstddef> 59 #include <cstdint> 60 #include <iterator> 61 #include <string> 62 #include <tuple> 63 #include <utility> 64 65 using namespace llvm; 66 67 static cl::opt<bool> JumpIsExpensiveOverride( 68 "jump-is-expensive", cl::init(false), 69 cl::desc("Do not create extra branches to split comparison logic."), 70 cl::Hidden); 71 72 static cl::opt<unsigned> MinimumJumpTableEntries 73 ("min-jump-table-entries", cl::init(4), cl::Hidden, 74 cl::desc("Set minimum number of entries to use a jump table.")); 75 76 static cl::opt<unsigned> MaximumJumpTableSize 77 ("max-jump-table-size", cl::init(0), cl::Hidden, 78 cl::desc("Set maximum size of jump tables; zero for no limit.")); 79 80 /// Minimum jump table density for normal functions. 81 static cl::opt<unsigned> 82 JumpTableDensity("jump-table-density", cl::init(10), cl::Hidden, 83 cl::desc("Minimum density for building a jump table in " 84 "a normal function")); 85 86 /// Minimum jump table density for -Os or -Oz functions. 87 static cl::opt<unsigned> OptsizeJumpTableDensity( 88 "optsize-jump-table-density", cl::init(40), cl::Hidden, 89 cl::desc("Minimum density for building a jump table in " 90 "an optsize function")); 91 92 // Although this default value is arbitrary, it is not random. It is assumed 93 // that a condition that evaluates the same way by a higher percentage than this 94 // is best represented as control flow. Therefore, the default value N should be 95 // set such that the win from N% correct executions is greater than the loss 96 // from (100 - N)% mispredicted executions for the majority of intended targets. 97 static cl::opt<int> MinPercentageForPredictableBranch( 98 "min-predictable-branch", cl::init(99), 99 cl::desc("Minimum percentage (0-100) that a condition must be either true " 100 "or false to assume that the condition is predictable"), 101 cl::Hidden); 102 103 /// InitLibcallNames - Set default libcall names. 104 static void InitLibcallNames(const char **Names, const Triple &TT) { 105 #define HANDLE_LIBCALL(code, name) \ 106 Names[RTLIB::code] = name; 107 #include "llvm/CodeGen/RuntimeLibcalls.def" 108 #undef HANDLE_LIBCALL 109 110 // A few names are different on particular architectures or environments. 111 if (TT.isOSDarwin()) { 112 // For f16/f32 conversions, Darwin uses the standard naming scheme, instead 113 // of the gnueabi-style __gnu_*_ieee. 114 // FIXME: What about other targets? 115 Names[RTLIB::FPEXT_F16_F32] = "__extendhfsf2"; 116 Names[RTLIB::FPROUND_F32_F16] = "__truncsfhf2"; 117 } else { 118 Names[RTLIB::FPEXT_F16_F32] = "__gnu_h2f_ieee"; 119 Names[RTLIB::FPROUND_F32_F16] = "__gnu_f2h_ieee"; 120 } 121 122 if (TT.isGNUEnvironment() || TT.isOSFuchsia()) { 123 Names[RTLIB::SINCOS_F32] = "sincosf"; 124 Names[RTLIB::SINCOS_F64] = "sincos"; 125 Names[RTLIB::SINCOS_F80] = "sincosl"; 126 Names[RTLIB::SINCOS_F128] = "sincosl"; 127 Names[RTLIB::SINCOS_PPCF128] = "sincosl"; 128 } 129 130 if (TT.isOSOpenBSD()) { 131 Names[RTLIB::STACKPROTECTOR_CHECK_FAIL] = nullptr; 132 } 133 } 134 135 /// Set default libcall CallingConvs. 136 static void InitLibcallCallingConvs(CallingConv::ID *CCs) { 137 for (int LC = 0; LC < RTLIB::UNKNOWN_LIBCALL; ++LC) 138 CCs[LC] = CallingConv::C; 139 } 140 141 /// getFPEXT - Return the FPEXT_*_* value for the given types, or 142 /// UNKNOWN_LIBCALL if there is none. 143 RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) { 144 if (OpVT == MVT::f16) { 145 if (RetVT == MVT::f32) 146 return FPEXT_F16_F32; 147 } else if (OpVT == MVT::f32) { 148 if (RetVT == MVT::f64) 149 return FPEXT_F32_F64; 150 if (RetVT == MVT::f128) 151 return FPEXT_F32_F128; 152 if (RetVT == MVT::ppcf128) 153 return FPEXT_F32_PPCF128; 154 } else if (OpVT == MVT::f64) { 155 if (RetVT == MVT::f128) 156 return FPEXT_F64_F128; 157 else if (RetVT == MVT::ppcf128) 158 return FPEXT_F64_PPCF128; 159 } 160 161 return UNKNOWN_LIBCALL; 162 } 163 164 /// getFPROUND - Return the FPROUND_*_* value for the given types, or 165 /// UNKNOWN_LIBCALL if there is none. 166 RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) { 167 if (RetVT == MVT::f16) { 168 if (OpVT == MVT::f32) 169 return FPROUND_F32_F16; 170 if (OpVT == MVT::f64) 171 return FPROUND_F64_F16; 172 if (OpVT == MVT::f80) 173 return FPROUND_F80_F16; 174 if (OpVT == MVT::f128) 175 return FPROUND_F128_F16; 176 if (OpVT == MVT::ppcf128) 177 return FPROUND_PPCF128_F16; 178 } else if (RetVT == MVT::f32) { 179 if (OpVT == MVT::f64) 180 return FPROUND_F64_F32; 181 if (OpVT == MVT::f80) 182 return FPROUND_F80_F32; 183 if (OpVT == MVT::f128) 184 return FPROUND_F128_F32; 185 if (OpVT == MVT::ppcf128) 186 return FPROUND_PPCF128_F32; 187 } else if (RetVT == MVT::f64) { 188 if (OpVT == MVT::f80) 189 return FPROUND_F80_F64; 190 if (OpVT == MVT::f128) 191 return FPROUND_F128_F64; 192 if (OpVT == MVT::ppcf128) 193 return FPROUND_PPCF128_F64; 194 } 195 196 return UNKNOWN_LIBCALL; 197 } 198 199 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or 200 /// UNKNOWN_LIBCALL if there is none. 201 RTLIB::Libcall RTLIB::getFPTOSINT(EVT OpVT, EVT RetVT) { 202 if (OpVT == MVT::f32) { 203 if (RetVT == MVT::i32) 204 return FPTOSINT_F32_I32; 205 if (RetVT == MVT::i64) 206 return FPTOSINT_F32_I64; 207 if (RetVT == MVT::i128) 208 return FPTOSINT_F32_I128; 209 } else if (OpVT == MVT::f64) { 210 if (RetVT == MVT::i32) 211 return FPTOSINT_F64_I32; 212 if (RetVT == MVT::i64) 213 return FPTOSINT_F64_I64; 214 if (RetVT == MVT::i128) 215 return FPTOSINT_F64_I128; 216 } else if (OpVT == MVT::f80) { 217 if (RetVT == MVT::i32) 218 return FPTOSINT_F80_I32; 219 if (RetVT == MVT::i64) 220 return FPTOSINT_F80_I64; 221 if (RetVT == MVT::i128) 222 return FPTOSINT_F80_I128; 223 } else if (OpVT == MVT::f128) { 224 if (RetVT == MVT::i32) 225 return FPTOSINT_F128_I32; 226 if (RetVT == MVT::i64) 227 return FPTOSINT_F128_I64; 228 if (RetVT == MVT::i128) 229 return FPTOSINT_F128_I128; 230 } else if (OpVT == MVT::ppcf128) { 231 if (RetVT == MVT::i32) 232 return FPTOSINT_PPCF128_I32; 233 if (RetVT == MVT::i64) 234 return FPTOSINT_PPCF128_I64; 235 if (RetVT == MVT::i128) 236 return FPTOSINT_PPCF128_I128; 237 } 238 return UNKNOWN_LIBCALL; 239 } 240 241 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or 242 /// UNKNOWN_LIBCALL if there is none. 243 RTLIB::Libcall RTLIB::getFPTOUINT(EVT OpVT, EVT RetVT) { 244 if (OpVT == MVT::f32) { 245 if (RetVT == MVT::i32) 246 return FPTOUINT_F32_I32; 247 if (RetVT == MVT::i64) 248 return FPTOUINT_F32_I64; 249 if (RetVT == MVT::i128) 250 return FPTOUINT_F32_I128; 251 } else if (OpVT == MVT::f64) { 252 if (RetVT == MVT::i32) 253 return FPTOUINT_F64_I32; 254 if (RetVT == MVT::i64) 255 return FPTOUINT_F64_I64; 256 if (RetVT == MVT::i128) 257 return FPTOUINT_F64_I128; 258 } else if (OpVT == MVT::f80) { 259 if (RetVT == MVT::i32) 260 return FPTOUINT_F80_I32; 261 if (RetVT == MVT::i64) 262 return FPTOUINT_F80_I64; 263 if (RetVT == MVT::i128) 264 return FPTOUINT_F80_I128; 265 } else if (OpVT == MVT::f128) { 266 if (RetVT == MVT::i32) 267 return FPTOUINT_F128_I32; 268 if (RetVT == MVT::i64) 269 return FPTOUINT_F128_I64; 270 if (RetVT == MVT::i128) 271 return FPTOUINT_F128_I128; 272 } else if (OpVT == MVT::ppcf128) { 273 if (RetVT == MVT::i32) 274 return FPTOUINT_PPCF128_I32; 275 if (RetVT == MVT::i64) 276 return FPTOUINT_PPCF128_I64; 277 if (RetVT == MVT::i128) 278 return FPTOUINT_PPCF128_I128; 279 } 280 return UNKNOWN_LIBCALL; 281 } 282 283 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or 284 /// UNKNOWN_LIBCALL if there is none. 285 RTLIB::Libcall RTLIB::getSINTTOFP(EVT OpVT, EVT RetVT) { 286 if (OpVT == MVT::i32) { 287 if (RetVT == MVT::f32) 288 return SINTTOFP_I32_F32; 289 if (RetVT == MVT::f64) 290 return SINTTOFP_I32_F64; 291 if (RetVT == MVT::f80) 292 return SINTTOFP_I32_F80; 293 if (RetVT == MVT::f128) 294 return SINTTOFP_I32_F128; 295 if (RetVT == MVT::ppcf128) 296 return SINTTOFP_I32_PPCF128; 297 } else if (OpVT == MVT::i64) { 298 if (RetVT == MVT::f32) 299 return SINTTOFP_I64_F32; 300 if (RetVT == MVT::f64) 301 return SINTTOFP_I64_F64; 302 if (RetVT == MVT::f80) 303 return SINTTOFP_I64_F80; 304 if (RetVT == MVT::f128) 305 return SINTTOFP_I64_F128; 306 if (RetVT == MVT::ppcf128) 307 return SINTTOFP_I64_PPCF128; 308 } else if (OpVT == MVT::i128) { 309 if (RetVT == MVT::f32) 310 return SINTTOFP_I128_F32; 311 if (RetVT == MVT::f64) 312 return SINTTOFP_I128_F64; 313 if (RetVT == MVT::f80) 314 return SINTTOFP_I128_F80; 315 if (RetVT == MVT::f128) 316 return SINTTOFP_I128_F128; 317 if (RetVT == MVT::ppcf128) 318 return SINTTOFP_I128_PPCF128; 319 } 320 return UNKNOWN_LIBCALL; 321 } 322 323 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or 324 /// UNKNOWN_LIBCALL if there is none. 325 RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) { 326 if (OpVT == MVT::i32) { 327 if (RetVT == MVT::f32) 328 return UINTTOFP_I32_F32; 329 if (RetVT == MVT::f64) 330 return UINTTOFP_I32_F64; 331 if (RetVT == MVT::f80) 332 return UINTTOFP_I32_F80; 333 if (RetVT == MVT::f128) 334 return UINTTOFP_I32_F128; 335 if (RetVT == MVT::ppcf128) 336 return UINTTOFP_I32_PPCF128; 337 } else if (OpVT == MVT::i64) { 338 if (RetVT == MVT::f32) 339 return UINTTOFP_I64_F32; 340 if (RetVT == MVT::f64) 341 return UINTTOFP_I64_F64; 342 if (RetVT == MVT::f80) 343 return UINTTOFP_I64_F80; 344 if (RetVT == MVT::f128) 345 return UINTTOFP_I64_F128; 346 if (RetVT == MVT::ppcf128) 347 return UINTTOFP_I64_PPCF128; 348 } else if (OpVT == MVT::i128) { 349 if (RetVT == MVT::f32) 350 return UINTTOFP_I128_F32; 351 if (RetVT == MVT::f64) 352 return UINTTOFP_I128_F64; 353 if (RetVT == MVT::f80) 354 return UINTTOFP_I128_F80; 355 if (RetVT == MVT::f128) 356 return UINTTOFP_I128_F128; 357 if (RetVT == MVT::ppcf128) 358 return UINTTOFP_I128_PPCF128; 359 } 360 return UNKNOWN_LIBCALL; 361 } 362 363 RTLIB::Libcall RTLIB::getSYNC(unsigned Opc, MVT VT) { 364 #define OP_TO_LIBCALL(Name, Enum) \ 365 case Name: \ 366 switch (VT.SimpleTy) { \ 367 default: \ 368 return UNKNOWN_LIBCALL; \ 369 case MVT::i8: \ 370 return Enum##_1; \ 371 case MVT::i16: \ 372 return Enum##_2; \ 373 case MVT::i32: \ 374 return Enum##_4; \ 375 case MVT::i64: \ 376 return Enum##_8; \ 377 case MVT::i128: \ 378 return Enum##_16; \ 379 } 380 381 switch (Opc) { 382 OP_TO_LIBCALL(ISD::ATOMIC_SWAP, SYNC_LOCK_TEST_AND_SET) 383 OP_TO_LIBCALL(ISD::ATOMIC_CMP_SWAP, SYNC_VAL_COMPARE_AND_SWAP) 384 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_ADD, SYNC_FETCH_AND_ADD) 385 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_SUB, SYNC_FETCH_AND_SUB) 386 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_AND, SYNC_FETCH_AND_AND) 387 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_OR, SYNC_FETCH_AND_OR) 388 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_XOR, SYNC_FETCH_AND_XOR) 389 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_NAND, SYNC_FETCH_AND_NAND) 390 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MAX, SYNC_FETCH_AND_MAX) 391 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMAX, SYNC_FETCH_AND_UMAX) 392 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MIN, SYNC_FETCH_AND_MIN) 393 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMIN, SYNC_FETCH_AND_UMIN) 394 } 395 396 #undef OP_TO_LIBCALL 397 398 return UNKNOWN_LIBCALL; 399 } 400 401 RTLIB::Libcall RTLIB::getMEMCPY_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { 402 switch (ElementSize) { 403 case 1: 404 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_1; 405 case 2: 406 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_2; 407 case 4: 408 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_4; 409 case 8: 410 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_8; 411 case 16: 412 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_16; 413 default: 414 return UNKNOWN_LIBCALL; 415 } 416 } 417 418 RTLIB::Libcall RTLIB::getMEMMOVE_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { 419 switch (ElementSize) { 420 case 1: 421 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_1; 422 case 2: 423 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_2; 424 case 4: 425 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_4; 426 case 8: 427 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_8; 428 case 16: 429 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_16; 430 default: 431 return UNKNOWN_LIBCALL; 432 } 433 } 434 435 RTLIB::Libcall RTLIB::getMEMSET_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { 436 switch (ElementSize) { 437 case 1: 438 return MEMSET_ELEMENT_UNORDERED_ATOMIC_1; 439 case 2: 440 return MEMSET_ELEMENT_UNORDERED_ATOMIC_2; 441 case 4: 442 return MEMSET_ELEMENT_UNORDERED_ATOMIC_4; 443 case 8: 444 return MEMSET_ELEMENT_UNORDERED_ATOMIC_8; 445 case 16: 446 return MEMSET_ELEMENT_UNORDERED_ATOMIC_16; 447 default: 448 return UNKNOWN_LIBCALL; 449 } 450 } 451 452 /// InitCmpLibcallCCs - Set default comparison libcall CC. 453 static void InitCmpLibcallCCs(ISD::CondCode *CCs) { 454 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL); 455 CCs[RTLIB::OEQ_F32] = ISD::SETEQ; 456 CCs[RTLIB::OEQ_F64] = ISD::SETEQ; 457 CCs[RTLIB::OEQ_F128] = ISD::SETEQ; 458 CCs[RTLIB::OEQ_PPCF128] = ISD::SETEQ; 459 CCs[RTLIB::UNE_F32] = ISD::SETNE; 460 CCs[RTLIB::UNE_F64] = ISD::SETNE; 461 CCs[RTLIB::UNE_F128] = ISD::SETNE; 462 CCs[RTLIB::UNE_PPCF128] = ISD::SETNE; 463 CCs[RTLIB::OGE_F32] = ISD::SETGE; 464 CCs[RTLIB::OGE_F64] = ISD::SETGE; 465 CCs[RTLIB::OGE_F128] = ISD::SETGE; 466 CCs[RTLIB::OGE_PPCF128] = ISD::SETGE; 467 CCs[RTLIB::OLT_F32] = ISD::SETLT; 468 CCs[RTLIB::OLT_F64] = ISD::SETLT; 469 CCs[RTLIB::OLT_F128] = ISD::SETLT; 470 CCs[RTLIB::OLT_PPCF128] = ISD::SETLT; 471 CCs[RTLIB::OLE_F32] = ISD::SETLE; 472 CCs[RTLIB::OLE_F64] = ISD::SETLE; 473 CCs[RTLIB::OLE_F128] = ISD::SETLE; 474 CCs[RTLIB::OLE_PPCF128] = ISD::SETLE; 475 CCs[RTLIB::OGT_F32] = ISD::SETGT; 476 CCs[RTLIB::OGT_F64] = ISD::SETGT; 477 CCs[RTLIB::OGT_F128] = ISD::SETGT; 478 CCs[RTLIB::OGT_PPCF128] = ISD::SETGT; 479 CCs[RTLIB::UO_F32] = ISD::SETNE; 480 CCs[RTLIB::UO_F64] = ISD::SETNE; 481 CCs[RTLIB::UO_F128] = ISD::SETNE; 482 CCs[RTLIB::UO_PPCF128] = ISD::SETNE; 483 CCs[RTLIB::O_F32] = ISD::SETEQ; 484 CCs[RTLIB::O_F64] = ISD::SETEQ; 485 CCs[RTLIB::O_F128] = ISD::SETEQ; 486 CCs[RTLIB::O_PPCF128] = ISD::SETEQ; 487 } 488 489 /// NOTE: The TargetMachine owns TLOF. 490 TargetLoweringBase::TargetLoweringBase(const TargetMachine &tm) : TM(tm) { 491 initActions(); 492 493 // Perform these initializations only once. 494 MaxStoresPerMemset = MaxStoresPerMemcpy = MaxStoresPerMemmove = 495 MaxLoadsPerMemcmp = 8; 496 MaxStoresPerMemsetOptSize = MaxStoresPerMemcpyOptSize = 497 MaxStoresPerMemmoveOptSize = MaxLoadsPerMemcmpOptSize = 4; 498 UseUnderscoreSetJmp = false; 499 UseUnderscoreLongJmp = false; 500 HasMultipleConditionRegisters = false; 501 HasExtractBitsInsn = false; 502 JumpIsExpensive = JumpIsExpensiveOverride; 503 PredictableSelectIsExpensive = false; 504 EnableExtLdPromotion = false; 505 HasFloatingPointExceptions = true; 506 StackPointerRegisterToSaveRestore = 0; 507 BooleanContents = UndefinedBooleanContent; 508 BooleanFloatContents = UndefinedBooleanContent; 509 BooleanVectorContents = UndefinedBooleanContent; 510 SchedPreferenceInfo = Sched::ILP; 511 JumpBufSize = 0; 512 JumpBufAlignment = 0; 513 MinFunctionAlignment = 0; 514 PrefFunctionAlignment = 0; 515 PrefLoopAlignment = 0; 516 GatherAllAliasesMaxDepth = 18; 517 MinStackArgumentAlignment = 1; 518 // TODO: the default will be switched to 0 in the next commit, along 519 // with the Target-specific changes necessary. 520 MaxAtomicSizeInBitsSupported = 1024; 521 522 MinCmpXchgSizeInBits = 0; 523 524 std::fill(std::begin(LibcallRoutineNames), std::end(LibcallRoutineNames), nullptr); 525 526 InitLibcallNames(LibcallRoutineNames, TM.getTargetTriple()); 527 InitCmpLibcallCCs(CmpLibcallCCs); 528 InitLibcallCallingConvs(LibcallCallingConvs); 529 } 530 531 void TargetLoweringBase::initActions() { 532 // All operations default to being supported. 533 memset(OpActions, 0, sizeof(OpActions)); 534 memset(LoadExtActions, 0, sizeof(LoadExtActions)); 535 memset(TruncStoreActions, 0, sizeof(TruncStoreActions)); 536 memset(IndexedModeActions, 0, sizeof(IndexedModeActions)); 537 memset(CondCodeActions, 0, sizeof(CondCodeActions)); 538 std::fill(std::begin(RegClassForVT), std::end(RegClassForVT), nullptr); 539 std::fill(std::begin(TargetDAGCombineArray), 540 std::end(TargetDAGCombineArray), 0); 541 542 // Set default actions for various operations. 543 for (MVT VT : MVT::all_valuetypes()) { 544 // Default all indexed load / store to expand. 545 for (unsigned IM = (unsigned)ISD::PRE_INC; 546 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) { 547 setIndexedLoadAction(IM, VT, Expand); 548 setIndexedStoreAction(IM, VT, Expand); 549 } 550 551 // Most backends expect to see the node which just returns the value loaded. 552 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Expand); 553 554 // These operations default to expand. 555 setOperationAction(ISD::FGETSIGN, VT, Expand); 556 setOperationAction(ISD::CONCAT_VECTORS, VT, Expand); 557 setOperationAction(ISD::FMINNUM, VT, Expand); 558 setOperationAction(ISD::FMAXNUM, VT, Expand); 559 setOperationAction(ISD::FMINNAN, VT, Expand); 560 setOperationAction(ISD::FMAXNAN, VT, Expand); 561 setOperationAction(ISD::FMAD, VT, Expand); 562 setOperationAction(ISD::SMIN, VT, Expand); 563 setOperationAction(ISD::SMAX, VT, Expand); 564 setOperationAction(ISD::UMIN, VT, Expand); 565 setOperationAction(ISD::UMAX, VT, Expand); 566 setOperationAction(ISD::ABS, VT, Expand); 567 568 // Overflow operations default to expand 569 setOperationAction(ISD::SADDO, VT, Expand); 570 setOperationAction(ISD::SSUBO, VT, Expand); 571 setOperationAction(ISD::UADDO, VT, Expand); 572 setOperationAction(ISD::USUBO, VT, Expand); 573 setOperationAction(ISD::SMULO, VT, Expand); 574 setOperationAction(ISD::UMULO, VT, Expand); 575 576 // ADDCARRY operations default to expand 577 setOperationAction(ISD::ADDCARRY, VT, Expand); 578 setOperationAction(ISD::SUBCARRY, VT, Expand); 579 setOperationAction(ISD::SETCCCARRY, VT, Expand); 580 581 // These default to Expand so they will be expanded to CTLZ/CTTZ by default. 582 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand); 583 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand); 584 585 setOperationAction(ISD::BITREVERSE, VT, Expand); 586 587 // These library functions default to expand. 588 setOperationAction(ISD::FROUND, VT, Expand); 589 setOperationAction(ISD::FPOWI, VT, Expand); 590 591 // These operations default to expand for vector types. 592 if (VT.isVector()) { 593 setOperationAction(ISD::FCOPYSIGN, VT, Expand); 594 setOperationAction(ISD::ANY_EXTEND_VECTOR_INREG, VT, Expand); 595 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, VT, Expand); 596 setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, VT, Expand); 597 } 598 599 // For most targets @llvm.get.dynamic.area.offset just returns 0. 600 setOperationAction(ISD::GET_DYNAMIC_AREA_OFFSET, VT, Expand); 601 } 602 603 // Most targets ignore the @llvm.prefetch intrinsic. 604 setOperationAction(ISD::PREFETCH, MVT::Other, Expand); 605 606 // Most targets also ignore the @llvm.readcyclecounter intrinsic. 607 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Expand); 608 609 // ConstantFP nodes default to expand. Targets can either change this to 610 // Legal, in which case all fp constants are legal, or use isFPImmLegal() 611 // to optimize expansions for certain constants. 612 setOperationAction(ISD::ConstantFP, MVT::f16, Expand); 613 setOperationAction(ISD::ConstantFP, MVT::f32, Expand); 614 setOperationAction(ISD::ConstantFP, MVT::f64, Expand); 615 setOperationAction(ISD::ConstantFP, MVT::f80, Expand); 616 setOperationAction(ISD::ConstantFP, MVT::f128, Expand); 617 618 // These library functions default to expand. 619 for (MVT VT : {MVT::f32, MVT::f64, MVT::f128}) { 620 setOperationAction(ISD::FLOG , VT, Expand); 621 setOperationAction(ISD::FLOG2, VT, Expand); 622 setOperationAction(ISD::FLOG10, VT, Expand); 623 setOperationAction(ISD::FEXP , VT, Expand); 624 setOperationAction(ISD::FEXP2, VT, Expand); 625 setOperationAction(ISD::FFLOOR, VT, Expand); 626 setOperationAction(ISD::FNEARBYINT, VT, Expand); 627 setOperationAction(ISD::FCEIL, VT, Expand); 628 setOperationAction(ISD::FRINT, VT, Expand); 629 setOperationAction(ISD::FTRUNC, VT, Expand); 630 setOperationAction(ISD::FROUND, VT, Expand); 631 } 632 633 // Default ISD::TRAP to expand (which turns it into abort). 634 setOperationAction(ISD::TRAP, MVT::Other, Expand); 635 636 // On most systems, DEBUGTRAP and TRAP have no difference. The "Expand" 637 // here is to inform DAG Legalizer to replace DEBUGTRAP with TRAP. 638 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Expand); 639 } 640 641 MVT TargetLoweringBase::getScalarShiftAmountTy(const DataLayout &DL, 642 EVT) const { 643 return MVT::getIntegerVT(8 * DL.getPointerSize(0)); 644 } 645 646 EVT TargetLoweringBase::getShiftAmountTy(EVT LHSTy, 647 const DataLayout &DL) const { 648 assert(LHSTy.isInteger() && "Shift amount is not an integer type!"); 649 if (LHSTy.isVector()) 650 return LHSTy; 651 return getScalarShiftAmountTy(DL, LHSTy); 652 } 653 654 bool TargetLoweringBase::canOpTrap(unsigned Op, EVT VT) const { 655 assert(isTypeLegal(VT)); 656 switch (Op) { 657 default: 658 return false; 659 case ISD::SDIV: 660 case ISD::UDIV: 661 case ISD::SREM: 662 case ISD::UREM: 663 return true; 664 } 665 } 666 667 void TargetLoweringBase::setJumpIsExpensive(bool isExpensive) { 668 // If the command-line option was specified, ignore this request. 669 if (!JumpIsExpensiveOverride.getNumOccurrences()) 670 JumpIsExpensive = isExpensive; 671 } 672 673 TargetLoweringBase::LegalizeKind 674 TargetLoweringBase::getTypeConversion(LLVMContext &Context, EVT VT) const { 675 // If this is a simple type, use the ComputeRegisterProp mechanism. 676 if (VT.isSimple()) { 677 MVT SVT = VT.getSimpleVT(); 678 assert((unsigned)SVT.SimpleTy < array_lengthof(TransformToType)); 679 MVT NVT = TransformToType[SVT.SimpleTy]; 680 LegalizeTypeAction LA = ValueTypeActions.getTypeAction(SVT); 681 682 assert((LA == TypeLegal || LA == TypeSoftenFloat || 683 ValueTypeActions.getTypeAction(NVT) != TypePromoteInteger) && 684 "Promote may not follow Expand or Promote"); 685 686 if (LA == TypeSplitVector) 687 return LegalizeKind(LA, 688 EVT::getVectorVT(Context, SVT.getVectorElementType(), 689 SVT.getVectorNumElements() / 2)); 690 if (LA == TypeScalarizeVector) 691 return LegalizeKind(LA, SVT.getVectorElementType()); 692 return LegalizeKind(LA, NVT); 693 } 694 695 // Handle Extended Scalar Types. 696 if (!VT.isVector()) { 697 assert(VT.isInteger() && "Float types must be simple"); 698 unsigned BitSize = VT.getSizeInBits(); 699 // First promote to a power-of-two size, then expand if necessary. 700 if (BitSize < 8 || !isPowerOf2_32(BitSize)) { 701 EVT NVT = VT.getRoundIntegerType(Context); 702 assert(NVT != VT && "Unable to round integer VT"); 703 LegalizeKind NextStep = getTypeConversion(Context, NVT); 704 // Avoid multi-step promotion. 705 if (NextStep.first == TypePromoteInteger) 706 return NextStep; 707 // Return rounded integer type. 708 return LegalizeKind(TypePromoteInteger, NVT); 709 } 710 711 return LegalizeKind(TypeExpandInteger, 712 EVT::getIntegerVT(Context, VT.getSizeInBits() / 2)); 713 } 714 715 // Handle vector types. 716 unsigned NumElts = VT.getVectorNumElements(); 717 EVT EltVT = VT.getVectorElementType(); 718 719 // Vectors with only one element are always scalarized. 720 if (NumElts == 1) 721 return LegalizeKind(TypeScalarizeVector, EltVT); 722 723 // Try to widen vector elements until the element type is a power of two and 724 // promote it to a legal type later on, for example: 725 // <3 x i8> -> <4 x i8> -> <4 x i32> 726 if (EltVT.isInteger()) { 727 // Vectors with a number of elements that is not a power of two are always 728 // widened, for example <3 x i8> -> <4 x i8>. 729 if (!VT.isPow2VectorType()) { 730 NumElts = (unsigned)NextPowerOf2(NumElts); 731 EVT NVT = EVT::getVectorVT(Context, EltVT, NumElts); 732 return LegalizeKind(TypeWidenVector, NVT); 733 } 734 735 // Examine the element type. 736 LegalizeKind LK = getTypeConversion(Context, EltVT); 737 738 // If type is to be expanded, split the vector. 739 // <4 x i140> -> <2 x i140> 740 if (LK.first == TypeExpandInteger) 741 return LegalizeKind(TypeSplitVector, 742 EVT::getVectorVT(Context, EltVT, NumElts / 2)); 743 744 // Promote the integer element types until a legal vector type is found 745 // or until the element integer type is too big. If a legal type was not 746 // found, fallback to the usual mechanism of widening/splitting the 747 // vector. 748 EVT OldEltVT = EltVT; 749 while (true) { 750 // Increase the bitwidth of the element to the next pow-of-two 751 // (which is greater than 8 bits). 752 EltVT = EVT::getIntegerVT(Context, 1 + EltVT.getSizeInBits()) 753 .getRoundIntegerType(Context); 754 755 // Stop trying when getting a non-simple element type. 756 // Note that vector elements may be greater than legal vector element 757 // types. Example: X86 XMM registers hold 64bit element on 32bit 758 // systems. 759 if (!EltVT.isSimple()) 760 break; 761 762 // Build a new vector type and check if it is legal. 763 MVT NVT = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 764 // Found a legal promoted vector type. 765 if (NVT != MVT() && ValueTypeActions.getTypeAction(NVT) == TypeLegal) 766 return LegalizeKind(TypePromoteInteger, 767 EVT::getVectorVT(Context, EltVT, NumElts)); 768 } 769 770 // Reset the type to the unexpanded type if we did not find a legal vector 771 // type with a promoted vector element type. 772 EltVT = OldEltVT; 773 } 774 775 // Try to widen the vector until a legal type is found. 776 // If there is no wider legal type, split the vector. 777 while (true) { 778 // Round up to the next power of 2. 779 NumElts = (unsigned)NextPowerOf2(NumElts); 780 781 // If there is no simple vector type with this many elements then there 782 // cannot be a larger legal vector type. Note that this assumes that 783 // there are no skipped intermediate vector types in the simple types. 784 if (!EltVT.isSimple()) 785 break; 786 MVT LargerVector = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 787 if (LargerVector == MVT()) 788 break; 789 790 // If this type is legal then widen the vector. 791 if (ValueTypeActions.getTypeAction(LargerVector) == TypeLegal) 792 return LegalizeKind(TypeWidenVector, LargerVector); 793 } 794 795 // Widen odd vectors to next power of two. 796 if (!VT.isPow2VectorType()) { 797 EVT NVT = VT.getPow2VectorType(Context); 798 return LegalizeKind(TypeWidenVector, NVT); 799 } 800 801 // Vectors with illegal element types are expanded. 802 EVT NVT = EVT::getVectorVT(Context, EltVT, VT.getVectorNumElements() / 2); 803 return LegalizeKind(TypeSplitVector, NVT); 804 } 805 806 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT, 807 unsigned &NumIntermediates, 808 MVT &RegisterVT, 809 TargetLoweringBase *TLI) { 810 // Figure out the right, legal destination reg to copy into. 811 unsigned NumElts = VT.getVectorNumElements(); 812 MVT EltTy = VT.getVectorElementType(); 813 814 unsigned NumVectorRegs = 1; 815 816 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 817 // could break down into LHS/RHS like LegalizeDAG does. 818 if (!isPowerOf2_32(NumElts)) { 819 NumVectorRegs = NumElts; 820 NumElts = 1; 821 } 822 823 // Divide the input until we get to a supported size. This will always 824 // end with a scalar if the target doesn't support vectors. 825 while (NumElts > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) { 826 NumElts >>= 1; 827 NumVectorRegs <<= 1; 828 } 829 830 NumIntermediates = NumVectorRegs; 831 832 MVT NewVT = MVT::getVectorVT(EltTy, NumElts); 833 if (!TLI->isTypeLegal(NewVT)) 834 NewVT = EltTy; 835 IntermediateVT = NewVT; 836 837 unsigned NewVTSize = NewVT.getSizeInBits(); 838 839 // Convert sizes such as i33 to i64. 840 if (!isPowerOf2_32(NewVTSize)) 841 NewVTSize = NextPowerOf2(NewVTSize); 842 843 MVT DestVT = TLI->getRegisterType(NewVT); 844 RegisterVT = DestVT; 845 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 846 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 847 848 // Otherwise, promotion or legal types use the same number of registers as 849 // the vector decimated to the appropriate level. 850 return NumVectorRegs; 851 } 852 853 /// isLegalRC - Return true if the value types that can be represented by the 854 /// specified register class are all legal. 855 bool TargetLoweringBase::isLegalRC(const TargetRegisterInfo &TRI, 856 const TargetRegisterClass &RC) const { 857 for (auto I = TRI.legalclasstypes_begin(RC); *I != MVT::Other; ++I) 858 if (isTypeLegal(*I)) 859 return true; 860 return false; 861 } 862 863 /// Replace/modify any TargetFrameIndex operands with a targte-dependent 864 /// sequence of memory operands that is recognized by PrologEpilogInserter. 865 MachineBasicBlock * 866 TargetLoweringBase::emitPatchPoint(MachineInstr &InitialMI, 867 MachineBasicBlock *MBB) const { 868 MachineInstr *MI = &InitialMI; 869 MachineFunction &MF = *MI->getMF(); 870 MachineFrameInfo &MFI = MF.getFrameInfo(); 871 872 // We're handling multiple types of operands here: 873 // PATCHPOINT MetaArgs - live-in, read only, direct 874 // STATEPOINT Deopt Spill - live-through, read only, indirect 875 // STATEPOINT Deopt Alloca - live-through, read only, direct 876 // (We're currently conservative and mark the deopt slots read/write in 877 // practice.) 878 // STATEPOINT GC Spill - live-through, read/write, indirect 879 // STATEPOINT GC Alloca - live-through, read/write, direct 880 // The live-in vs live-through is handled already (the live through ones are 881 // all stack slots), but we need to handle the different type of stackmap 882 // operands and memory effects here. 883 884 // MI changes inside this loop as we grow operands. 885 for(unsigned OperIdx = 0; OperIdx != MI->getNumOperands(); ++OperIdx) { 886 MachineOperand &MO = MI->getOperand(OperIdx); 887 if (!MO.isFI()) 888 continue; 889 890 // foldMemoryOperand builds a new MI after replacing a single FI operand 891 // with the canonical set of five x86 addressing-mode operands. 892 int FI = MO.getIndex(); 893 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), MI->getDesc()); 894 895 // Copy operands before the frame-index. 896 for (unsigned i = 0; i < OperIdx; ++i) 897 MIB.add(MI->getOperand(i)); 898 // Add frame index operands recognized by stackmaps.cpp 899 if (MFI.isStatepointSpillSlotObjectIndex(FI)) { 900 // indirect-mem-ref tag, size, #FI, offset. 901 // Used for spills inserted by StatepointLowering. This codepath is not 902 // used for patchpoints/stackmaps at all, for these spilling is done via 903 // foldMemoryOperand callback only. 904 assert(MI->getOpcode() == TargetOpcode::STATEPOINT && "sanity"); 905 MIB.addImm(StackMaps::IndirectMemRefOp); 906 MIB.addImm(MFI.getObjectSize(FI)); 907 MIB.add(MI->getOperand(OperIdx)); 908 MIB.addImm(0); 909 } else { 910 // direct-mem-ref tag, #FI, offset. 911 // Used by patchpoint, and direct alloca arguments to statepoints 912 MIB.addImm(StackMaps::DirectMemRefOp); 913 MIB.add(MI->getOperand(OperIdx)); 914 MIB.addImm(0); 915 } 916 // Copy the operands after the frame index. 917 for (unsigned i = OperIdx + 1; i != MI->getNumOperands(); ++i) 918 MIB.add(MI->getOperand(i)); 919 920 // Inherit previous memory operands. 921 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end()); 922 assert(MIB->mayLoad() && "Folded a stackmap use to a non-load!"); 923 924 // Add a new memory operand for this FI. 925 assert(MFI.getObjectOffset(FI) != -1); 926 927 auto Flags = MachineMemOperand::MOLoad; 928 if (MI->getOpcode() == TargetOpcode::STATEPOINT) { 929 Flags |= MachineMemOperand::MOStore; 930 Flags |= MachineMemOperand::MOVolatile; 931 } 932 MachineMemOperand *MMO = MF.getMachineMemOperand( 933 MachinePointerInfo::getFixedStack(MF, FI), Flags, 934 MF.getDataLayout().getPointerSize(), MFI.getObjectAlignment(FI)); 935 MIB->addMemOperand(MF, MMO); 936 937 // Replace the instruction and update the operand index. 938 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 939 OperIdx += (MIB->getNumOperands() - MI->getNumOperands()) - 1; 940 MI->eraseFromParent(); 941 MI = MIB; 942 } 943 return MBB; 944 } 945 946 /// findRepresentativeClass - Return the largest legal super-reg register class 947 /// of the register class for the specified type and its associated "cost". 948 // This function is in TargetLowering because it uses RegClassForVT which would 949 // need to be moved to TargetRegisterInfo and would necessitate moving 950 // isTypeLegal over as well - a massive change that would just require 951 // TargetLowering having a TargetRegisterInfo class member that it would use. 952 std::pair<const TargetRegisterClass *, uint8_t> 953 TargetLoweringBase::findRepresentativeClass(const TargetRegisterInfo *TRI, 954 MVT VT) const { 955 const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy]; 956 if (!RC) 957 return std::make_pair(RC, 0); 958 959 // Compute the set of all super-register classes. 960 BitVector SuperRegRC(TRI->getNumRegClasses()); 961 for (SuperRegClassIterator RCI(RC, TRI); RCI.isValid(); ++RCI) 962 SuperRegRC.setBitsInMask(RCI.getMask()); 963 964 // Find the first legal register class with the largest spill size. 965 const TargetRegisterClass *BestRC = RC; 966 for (unsigned i : SuperRegRC.set_bits()) { 967 const TargetRegisterClass *SuperRC = TRI->getRegClass(i); 968 // We want the largest possible spill size. 969 if (TRI->getSpillSize(*SuperRC) <= TRI->getSpillSize(*BestRC)) 970 continue; 971 if (!isLegalRC(*TRI, *SuperRC)) 972 continue; 973 BestRC = SuperRC; 974 } 975 return std::make_pair(BestRC, 1); 976 } 977 978 /// computeRegisterProperties - Once all of the register classes are added, 979 /// this allows us to compute derived properties we expose. 980 void TargetLoweringBase::computeRegisterProperties( 981 const TargetRegisterInfo *TRI) { 982 static_assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE, 983 "Too many value types for ValueTypeActions to hold!"); 984 985 // Everything defaults to needing one register. 986 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 987 NumRegistersForVT[i] = 1; 988 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i; 989 } 990 // ...except isVoid, which doesn't need any registers. 991 NumRegistersForVT[MVT::isVoid] = 0; 992 993 // Find the largest integer register class. 994 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE; 995 for (; RegClassForVT[LargestIntReg] == nullptr; --LargestIntReg) 996 assert(LargestIntReg != MVT::i1 && "No integer registers defined!"); 997 998 // Every integer value type larger than this largest register takes twice as 999 // many registers to represent as the previous ValueType. 1000 for (unsigned ExpandedReg = LargestIntReg + 1; 1001 ExpandedReg <= MVT::LAST_INTEGER_VALUETYPE; ++ExpandedReg) { 1002 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1]; 1003 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg; 1004 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1); 1005 ValueTypeActions.setTypeAction((MVT::SimpleValueType)ExpandedReg, 1006 TypeExpandInteger); 1007 } 1008 1009 // Inspect all of the ValueType's smaller than the largest integer 1010 // register to see which ones need promotion. 1011 unsigned LegalIntReg = LargestIntReg; 1012 for (unsigned IntReg = LargestIntReg - 1; 1013 IntReg >= (unsigned)MVT::i1; --IntReg) { 1014 MVT IVT = (MVT::SimpleValueType)IntReg; 1015 if (isTypeLegal(IVT)) { 1016 LegalIntReg = IntReg; 1017 } else { 1018 RegisterTypeForVT[IntReg] = TransformToType[IntReg] = 1019 (const MVT::SimpleValueType)LegalIntReg; 1020 ValueTypeActions.setTypeAction(IVT, TypePromoteInteger); 1021 } 1022 } 1023 1024 // ppcf128 type is really two f64's. 1025 if (!isTypeLegal(MVT::ppcf128)) { 1026 if (isTypeLegal(MVT::f64)) { 1027 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64]; 1028 RegisterTypeForVT[MVT::ppcf128] = MVT::f64; 1029 TransformToType[MVT::ppcf128] = MVT::f64; 1030 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeExpandFloat); 1031 } else { 1032 NumRegistersForVT[MVT::ppcf128] = NumRegistersForVT[MVT::i128]; 1033 RegisterTypeForVT[MVT::ppcf128] = RegisterTypeForVT[MVT::i128]; 1034 TransformToType[MVT::ppcf128] = MVT::i128; 1035 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeSoftenFloat); 1036 } 1037 } 1038 1039 // Decide how to handle f128. If the target does not have native f128 support, 1040 // expand it to i128 and we will be generating soft float library calls. 1041 if (!isTypeLegal(MVT::f128)) { 1042 NumRegistersForVT[MVT::f128] = NumRegistersForVT[MVT::i128]; 1043 RegisterTypeForVT[MVT::f128] = RegisterTypeForVT[MVT::i128]; 1044 TransformToType[MVT::f128] = MVT::i128; 1045 ValueTypeActions.setTypeAction(MVT::f128, TypeSoftenFloat); 1046 } 1047 1048 // Decide how to handle f64. If the target does not have native f64 support, 1049 // expand it to i64 and we will be generating soft float library calls. 1050 if (!isTypeLegal(MVT::f64)) { 1051 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64]; 1052 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64]; 1053 TransformToType[MVT::f64] = MVT::i64; 1054 ValueTypeActions.setTypeAction(MVT::f64, TypeSoftenFloat); 1055 } 1056 1057 // Decide how to handle f32. If the target does not have native f32 support, 1058 // expand it to i32 and we will be generating soft float library calls. 1059 if (!isTypeLegal(MVT::f32)) { 1060 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32]; 1061 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32]; 1062 TransformToType[MVT::f32] = MVT::i32; 1063 ValueTypeActions.setTypeAction(MVT::f32, TypeSoftenFloat); 1064 } 1065 1066 // Decide how to handle f16. If the target does not have native f16 support, 1067 // promote it to f32, because there are no f16 library calls (except for 1068 // conversions). 1069 if (!isTypeLegal(MVT::f16)) { 1070 NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::f32]; 1071 RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::f32]; 1072 TransformToType[MVT::f16] = MVT::f32; 1073 ValueTypeActions.setTypeAction(MVT::f16, TypePromoteFloat); 1074 } 1075 1076 // Loop over all of the vector value types to see which need transformations. 1077 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE; 1078 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) { 1079 MVT VT = (MVT::SimpleValueType) i; 1080 if (isTypeLegal(VT)) 1081 continue; 1082 1083 MVT EltVT = VT.getVectorElementType(); 1084 unsigned NElts = VT.getVectorNumElements(); 1085 bool IsLegalWiderType = false; 1086 LegalizeTypeAction PreferredAction = getPreferredVectorAction(VT); 1087 switch (PreferredAction) { 1088 case TypePromoteInteger: 1089 // Try to promote the elements of integer vectors. If no legal 1090 // promotion was found, fall through to the widen-vector method. 1091 for (unsigned nVT = i + 1; nVT <= MVT::LAST_INTEGER_VECTOR_VALUETYPE; ++nVT) { 1092 MVT SVT = (MVT::SimpleValueType) nVT; 1093 // Promote vectors of integers to vectors with the same number 1094 // of elements, with a wider element type. 1095 if (SVT.getScalarSizeInBits() > EltVT.getSizeInBits() && 1096 SVT.getVectorNumElements() == NElts && isTypeLegal(SVT)) { 1097 TransformToType[i] = SVT; 1098 RegisterTypeForVT[i] = SVT; 1099 NumRegistersForVT[i] = 1; 1100 ValueTypeActions.setTypeAction(VT, TypePromoteInteger); 1101 IsLegalWiderType = true; 1102 break; 1103 } 1104 } 1105 if (IsLegalWiderType) 1106 break; 1107 LLVM_FALLTHROUGH; 1108 1109 case TypeWidenVector: 1110 // Try to widen the vector. 1111 for (unsigned nVT = i + 1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) { 1112 MVT SVT = (MVT::SimpleValueType) nVT; 1113 if (SVT.getVectorElementType() == EltVT 1114 && SVT.getVectorNumElements() > NElts && isTypeLegal(SVT)) { 1115 TransformToType[i] = SVT; 1116 RegisterTypeForVT[i] = SVT; 1117 NumRegistersForVT[i] = 1; 1118 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1119 IsLegalWiderType = true; 1120 break; 1121 } 1122 } 1123 if (IsLegalWiderType) 1124 break; 1125 LLVM_FALLTHROUGH; 1126 1127 case TypeSplitVector: 1128 case TypeScalarizeVector: { 1129 MVT IntermediateVT; 1130 MVT RegisterVT; 1131 unsigned NumIntermediates; 1132 NumRegistersForVT[i] = getVectorTypeBreakdownMVT(VT, IntermediateVT, 1133 NumIntermediates, RegisterVT, this); 1134 RegisterTypeForVT[i] = RegisterVT; 1135 1136 MVT NVT = VT.getPow2VectorType(); 1137 if (NVT == VT) { 1138 // Type is already a power of 2. The default action is to split. 1139 TransformToType[i] = MVT::Other; 1140 if (PreferredAction == TypeScalarizeVector) 1141 ValueTypeActions.setTypeAction(VT, TypeScalarizeVector); 1142 else if (PreferredAction == TypeSplitVector) 1143 ValueTypeActions.setTypeAction(VT, TypeSplitVector); 1144 else 1145 // Set type action according to the number of elements. 1146 ValueTypeActions.setTypeAction(VT, NElts == 1 ? TypeScalarizeVector 1147 : TypeSplitVector); 1148 } else { 1149 TransformToType[i] = NVT; 1150 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1151 } 1152 break; 1153 } 1154 default: 1155 llvm_unreachable("Unknown vector legalization action!"); 1156 } 1157 } 1158 1159 // Determine the 'representative' register class for each value type. 1160 // An representative register class is the largest (meaning one which is 1161 // not a sub-register class / subreg register class) legal register class for 1162 // a group of value types. For example, on i386, i8, i16, and i32 1163 // representative would be GR32; while on x86_64 it's GR64. 1164 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 1165 const TargetRegisterClass* RRC; 1166 uint8_t Cost; 1167 std::tie(RRC, Cost) = findRepresentativeClass(TRI, (MVT::SimpleValueType)i); 1168 RepRegClassForVT[i] = RRC; 1169 RepRegClassCostForVT[i] = Cost; 1170 } 1171 } 1172 1173 EVT TargetLoweringBase::getSetCCResultType(const DataLayout &DL, LLVMContext &, 1174 EVT VT) const { 1175 assert(!VT.isVector() && "No default SetCC type for vectors!"); 1176 return getPointerTy(DL).SimpleTy; 1177 } 1178 1179 MVT::SimpleValueType TargetLoweringBase::getCmpLibcallReturnType() const { 1180 return MVT::i32; // return the default value 1181 } 1182 1183 /// getVectorTypeBreakdown - Vector types are broken down into some number of 1184 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32 1185 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack. 1186 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86. 1187 /// 1188 /// This method returns the number of registers needed, and the VT for each 1189 /// register. It also returns the VT and quantity of the intermediate values 1190 /// before they are promoted/expanded. 1191 unsigned TargetLoweringBase::getVectorTypeBreakdown(LLVMContext &Context, EVT VT, 1192 EVT &IntermediateVT, 1193 unsigned &NumIntermediates, 1194 MVT &RegisterVT) const { 1195 unsigned NumElts = VT.getVectorNumElements(); 1196 1197 // If there is a wider vector type with the same element type as this one, 1198 // or a promoted vector type that has the same number of elements which 1199 // are wider, then we should convert to that legal vector type. 1200 // This handles things like <2 x float> -> <4 x float> and 1201 // <4 x i1> -> <4 x i32>. 1202 LegalizeTypeAction TA = getTypeAction(Context, VT); 1203 if (NumElts != 1 && (TA == TypeWidenVector || TA == TypePromoteInteger)) { 1204 EVT RegisterEVT = getTypeToTransformTo(Context, VT); 1205 if (isTypeLegal(RegisterEVT)) { 1206 IntermediateVT = RegisterEVT; 1207 RegisterVT = RegisterEVT.getSimpleVT(); 1208 NumIntermediates = 1; 1209 return 1; 1210 } 1211 } 1212 1213 // Figure out the right, legal destination reg to copy into. 1214 EVT EltTy = VT.getVectorElementType(); 1215 1216 unsigned NumVectorRegs = 1; 1217 1218 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 1219 // could break down into LHS/RHS like LegalizeDAG does. 1220 if (!isPowerOf2_32(NumElts)) { 1221 NumVectorRegs = NumElts; 1222 NumElts = 1; 1223 } 1224 1225 // Divide the input until we get to a supported size. This will always 1226 // end with a scalar if the target doesn't support vectors. 1227 while (NumElts > 1 && !isTypeLegal( 1228 EVT::getVectorVT(Context, EltTy, NumElts))) { 1229 NumElts >>= 1; 1230 NumVectorRegs <<= 1; 1231 } 1232 1233 NumIntermediates = NumVectorRegs; 1234 1235 EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts); 1236 if (!isTypeLegal(NewVT)) 1237 NewVT = EltTy; 1238 IntermediateVT = NewVT; 1239 1240 MVT DestVT = getRegisterType(Context, NewVT); 1241 RegisterVT = DestVT; 1242 unsigned NewVTSize = NewVT.getSizeInBits(); 1243 1244 // Convert sizes such as i33 to i64. 1245 if (!isPowerOf2_32(NewVTSize)) 1246 NewVTSize = NextPowerOf2(NewVTSize); 1247 1248 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 1249 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 1250 1251 // Otherwise, promotion or legal types use the same number of registers as 1252 // the vector decimated to the appropriate level. 1253 return NumVectorRegs; 1254 } 1255 1256 /// Get the EVTs and ArgFlags collections that represent the legalized return 1257 /// type of the given function. This does not require a DAG or a return value, 1258 /// and is suitable for use before any DAGs for the function are constructed. 1259 /// TODO: Move this out of TargetLowering.cpp. 1260 void llvm::GetReturnInfo(Type *ReturnType, AttributeList attr, 1261 SmallVectorImpl<ISD::OutputArg> &Outs, 1262 const TargetLowering &TLI, const DataLayout &DL) { 1263 SmallVector<EVT, 4> ValueVTs; 1264 ComputeValueVTs(TLI, DL, ReturnType, ValueVTs); 1265 unsigned NumValues = ValueVTs.size(); 1266 if (NumValues == 0) return; 1267 1268 for (unsigned j = 0, f = NumValues; j != f; ++j) { 1269 EVT VT = ValueVTs[j]; 1270 ISD::NodeType ExtendKind = ISD::ANY_EXTEND; 1271 1272 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt)) 1273 ExtendKind = ISD::SIGN_EXTEND; 1274 else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt)) 1275 ExtendKind = ISD::ZERO_EXTEND; 1276 1277 // FIXME: C calling convention requires the return type to be promoted to 1278 // at least 32-bit. But this is not necessary for non-C calling 1279 // conventions. The frontend should mark functions whose return values 1280 // require promoting with signext or zeroext attributes. 1281 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) { 1282 MVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32); 1283 if (VT.bitsLT(MinVT)) 1284 VT = MinVT; 1285 } 1286 1287 unsigned NumParts = 1288 TLI.getNumRegistersForCallingConv(ReturnType->getContext(), VT); 1289 MVT PartVT = 1290 TLI.getRegisterTypeForCallingConv(ReturnType->getContext(), VT); 1291 1292 // 'inreg' on function refers to return value 1293 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy(); 1294 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::InReg)) 1295 Flags.setInReg(); 1296 1297 // Propagate extension type if any 1298 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt)) 1299 Flags.setSExt(); 1300 else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt)) 1301 Flags.setZExt(); 1302 1303 for (unsigned i = 0; i < NumParts; ++i) 1304 Outs.push_back(ISD::OutputArg(Flags, PartVT, VT, /*isFixed=*/true, 0, 0)); 1305 } 1306 } 1307 1308 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate 1309 /// function arguments in the caller parameter area. This is the actual 1310 /// alignment, not its logarithm. 1311 unsigned TargetLoweringBase::getByValTypeAlignment(Type *Ty, 1312 const DataLayout &DL) const { 1313 return DL.getABITypeAlignment(Ty); 1314 } 1315 1316 bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context, 1317 const DataLayout &DL, EVT VT, 1318 unsigned AddrSpace, 1319 unsigned Alignment, 1320 bool *Fast) const { 1321 // Check if the specified alignment is sufficient based on the data layout. 1322 // TODO: While using the data layout works in practice, a better solution 1323 // would be to implement this check directly (make this a virtual function). 1324 // For example, the ABI alignment may change based on software platform while 1325 // this function should only be affected by hardware implementation. 1326 Type *Ty = VT.getTypeForEVT(Context); 1327 if (Alignment >= DL.getABITypeAlignment(Ty)) { 1328 // Assume that an access that meets the ABI-specified alignment is fast. 1329 if (Fast != nullptr) 1330 *Fast = true; 1331 return true; 1332 } 1333 1334 // This is a misaligned access. 1335 return allowsMisalignedMemoryAccesses(VT, AddrSpace, Alignment, Fast); 1336 } 1337 1338 BranchProbability TargetLoweringBase::getPredictableBranchThreshold() const { 1339 return BranchProbability(MinPercentageForPredictableBranch, 100); 1340 } 1341 1342 //===----------------------------------------------------------------------===// 1343 // TargetTransformInfo Helpers 1344 //===----------------------------------------------------------------------===// 1345 1346 int TargetLoweringBase::InstructionOpcodeToISD(unsigned Opcode) const { 1347 enum InstructionOpcodes { 1348 #define HANDLE_INST(NUM, OPCODE, CLASS) OPCODE = NUM, 1349 #define LAST_OTHER_INST(NUM) InstructionOpcodesCount = NUM 1350 #include "llvm/IR/Instruction.def" 1351 }; 1352 switch (static_cast<InstructionOpcodes>(Opcode)) { 1353 case Ret: return 0; 1354 case Br: return 0; 1355 case Switch: return 0; 1356 case IndirectBr: return 0; 1357 case Invoke: return 0; 1358 case Resume: return 0; 1359 case Unreachable: return 0; 1360 case CleanupRet: return 0; 1361 case CatchRet: return 0; 1362 case CatchPad: return 0; 1363 case CatchSwitch: return 0; 1364 case CleanupPad: return 0; 1365 case Add: return ISD::ADD; 1366 case FAdd: return ISD::FADD; 1367 case Sub: return ISD::SUB; 1368 case FSub: return ISD::FSUB; 1369 case Mul: return ISD::MUL; 1370 case FMul: return ISD::FMUL; 1371 case UDiv: return ISD::UDIV; 1372 case SDiv: return ISD::SDIV; 1373 case FDiv: return ISD::FDIV; 1374 case URem: return ISD::UREM; 1375 case SRem: return ISD::SREM; 1376 case FRem: return ISD::FREM; 1377 case Shl: return ISD::SHL; 1378 case LShr: return ISD::SRL; 1379 case AShr: return ISD::SRA; 1380 case And: return ISD::AND; 1381 case Or: return ISD::OR; 1382 case Xor: return ISD::XOR; 1383 case Alloca: return 0; 1384 case Load: return ISD::LOAD; 1385 case Store: return ISD::STORE; 1386 case GetElementPtr: return 0; 1387 case Fence: return 0; 1388 case AtomicCmpXchg: return 0; 1389 case AtomicRMW: return 0; 1390 case Trunc: return ISD::TRUNCATE; 1391 case ZExt: return ISD::ZERO_EXTEND; 1392 case SExt: return ISD::SIGN_EXTEND; 1393 case FPToUI: return ISD::FP_TO_UINT; 1394 case FPToSI: return ISD::FP_TO_SINT; 1395 case UIToFP: return ISD::UINT_TO_FP; 1396 case SIToFP: return ISD::SINT_TO_FP; 1397 case FPTrunc: return ISD::FP_ROUND; 1398 case FPExt: return ISD::FP_EXTEND; 1399 case PtrToInt: return ISD::BITCAST; 1400 case IntToPtr: return ISD::BITCAST; 1401 case BitCast: return ISD::BITCAST; 1402 case AddrSpaceCast: return ISD::ADDRSPACECAST; 1403 case ICmp: return ISD::SETCC; 1404 case FCmp: return ISD::SETCC; 1405 case PHI: return 0; 1406 case Call: return 0; 1407 case Select: return ISD::SELECT; 1408 case UserOp1: return 0; 1409 case UserOp2: return 0; 1410 case VAArg: return 0; 1411 case ExtractElement: return ISD::EXTRACT_VECTOR_ELT; 1412 case InsertElement: return ISD::INSERT_VECTOR_ELT; 1413 case ShuffleVector: return ISD::VECTOR_SHUFFLE; 1414 case ExtractValue: return ISD::MERGE_VALUES; 1415 case InsertValue: return ISD::MERGE_VALUES; 1416 case LandingPad: return 0; 1417 } 1418 1419 llvm_unreachable("Unknown instruction type encountered!"); 1420 } 1421 1422 std::pair<int, MVT> 1423 TargetLoweringBase::getTypeLegalizationCost(const DataLayout &DL, 1424 Type *Ty) const { 1425 LLVMContext &C = Ty->getContext(); 1426 EVT MTy = getValueType(DL, Ty); 1427 1428 int Cost = 1; 1429 // We keep legalizing the type until we find a legal kind. We assume that 1430 // the only operation that costs anything is the split. After splitting 1431 // we need to handle two types. 1432 while (true) { 1433 LegalizeKind LK = getTypeConversion(C, MTy); 1434 1435 if (LK.first == TypeLegal) 1436 return std::make_pair(Cost, MTy.getSimpleVT()); 1437 1438 if (LK.first == TypeSplitVector || LK.first == TypeExpandInteger) 1439 Cost *= 2; 1440 1441 // Do not loop with f128 type. 1442 if (MTy == LK.second) 1443 return std::make_pair(Cost, MTy.getSimpleVT()); 1444 1445 // Keep legalizing the type. 1446 MTy = LK.second; 1447 } 1448 } 1449 1450 Value *TargetLoweringBase::getDefaultSafeStackPointerLocation(IRBuilder<> &IRB, 1451 bool UseTLS) const { 1452 // compiler-rt provides a variable with a magic name. Targets that do not 1453 // link with compiler-rt may also provide such a variable. 1454 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 1455 const char *UnsafeStackPtrVar = "__safestack_unsafe_stack_ptr"; 1456 auto UnsafeStackPtr = 1457 dyn_cast_or_null<GlobalVariable>(M->getNamedValue(UnsafeStackPtrVar)); 1458 1459 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext()); 1460 1461 if (!UnsafeStackPtr) { 1462 auto TLSModel = UseTLS ? 1463 GlobalValue::InitialExecTLSModel : 1464 GlobalValue::NotThreadLocal; 1465 // The global variable is not defined yet, define it ourselves. 1466 // We use the initial-exec TLS model because we do not support the 1467 // variable living anywhere other than in the main executable. 1468 UnsafeStackPtr = new GlobalVariable( 1469 *M, StackPtrTy, false, GlobalValue::ExternalLinkage, nullptr, 1470 UnsafeStackPtrVar, nullptr, TLSModel); 1471 } else { 1472 // The variable exists, check its type and attributes. 1473 if (UnsafeStackPtr->getValueType() != StackPtrTy) 1474 report_fatal_error(Twine(UnsafeStackPtrVar) + " must have void* type"); 1475 if (UseTLS != UnsafeStackPtr->isThreadLocal()) 1476 report_fatal_error(Twine(UnsafeStackPtrVar) + " must " + 1477 (UseTLS ? "" : "not ") + "be thread-local"); 1478 } 1479 return UnsafeStackPtr; 1480 } 1481 1482 Value *TargetLoweringBase::getSafeStackPointerLocation(IRBuilder<> &IRB) const { 1483 if (!TM.getTargetTriple().isAndroid()) 1484 return getDefaultSafeStackPointerLocation(IRB, true); 1485 1486 // Android provides a libc function to retrieve the address of the current 1487 // thread's unsafe stack pointer. 1488 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 1489 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext()); 1490 Value *Fn = M->getOrInsertFunction("__safestack_pointer_address", 1491 StackPtrTy->getPointerTo(0)); 1492 return IRB.CreateCall(Fn); 1493 } 1494 1495 //===----------------------------------------------------------------------===// 1496 // Loop Strength Reduction hooks 1497 //===----------------------------------------------------------------------===// 1498 1499 /// isLegalAddressingMode - Return true if the addressing mode represented 1500 /// by AM is legal for this target, for a load/store of the specified type. 1501 bool TargetLoweringBase::isLegalAddressingMode(const DataLayout &DL, 1502 const AddrMode &AM, Type *Ty, 1503 unsigned AS, Instruction *I) const { 1504 // The default implementation of this implements a conservative RISCy, r+r and 1505 // r+i addr mode. 1506 1507 // Allows a sign-extended 16-bit immediate field. 1508 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1) 1509 return false; 1510 1511 // No global is ever allowed as a base. 1512 if (AM.BaseGV) 1513 return false; 1514 1515 // Only support r+r, 1516 switch (AM.Scale) { 1517 case 0: // "r+i" or just "i", depending on HasBaseReg. 1518 break; 1519 case 1: 1520 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed. 1521 return false; 1522 // Otherwise we have r+r or r+i. 1523 break; 1524 case 2: 1525 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed. 1526 return false; 1527 // Allow 2*r as r+r. 1528 break; 1529 default: // Don't allow n * r 1530 return false; 1531 } 1532 1533 return true; 1534 } 1535 1536 //===----------------------------------------------------------------------===// 1537 // Stack Protector 1538 //===----------------------------------------------------------------------===// 1539 1540 // For OpenBSD return its special guard variable. Otherwise return nullptr, 1541 // so that SelectionDAG handle SSP. 1542 Value *TargetLoweringBase::getIRStackGuard(IRBuilder<> &IRB) const { 1543 if (getTargetMachine().getTargetTriple().isOSOpenBSD()) { 1544 Module &M = *IRB.GetInsertBlock()->getParent()->getParent(); 1545 PointerType *PtrTy = Type::getInt8PtrTy(M.getContext()); 1546 return M.getOrInsertGlobal("__guard_local", PtrTy); 1547 } 1548 return nullptr; 1549 } 1550 1551 // Currently only support "standard" __stack_chk_guard. 1552 // TODO: add LOAD_STACK_GUARD support. 1553 void TargetLoweringBase::insertSSPDeclarations(Module &M) const { 1554 M.getOrInsertGlobal("__stack_chk_guard", Type::getInt8PtrTy(M.getContext())); 1555 } 1556 1557 // Currently only support "standard" __stack_chk_guard. 1558 // TODO: add LOAD_STACK_GUARD support. 1559 Value *TargetLoweringBase::getSDagStackGuard(const Module &M) const { 1560 return M.getGlobalVariable("__stack_chk_guard", true); 1561 } 1562 1563 Value *TargetLoweringBase::getSSPStackGuardCheck(const Module &M) const { 1564 return nullptr; 1565 } 1566 1567 unsigned TargetLoweringBase::getMinimumJumpTableEntries() const { 1568 return MinimumJumpTableEntries; 1569 } 1570 1571 void TargetLoweringBase::setMinimumJumpTableEntries(unsigned Val) { 1572 MinimumJumpTableEntries = Val; 1573 } 1574 1575 unsigned TargetLoweringBase::getMinimumJumpTableDensity(bool OptForSize) const { 1576 return OptForSize ? OptsizeJumpTableDensity : JumpTableDensity; 1577 } 1578 1579 unsigned TargetLoweringBase::getMaximumJumpTableSize() const { 1580 return MaximumJumpTableSize; 1581 } 1582 1583 void TargetLoweringBase::setMaximumJumpTableSize(unsigned Val) { 1584 MaximumJumpTableSize = Val; 1585 } 1586 1587 //===----------------------------------------------------------------------===// 1588 // Reciprocal Estimates 1589 //===----------------------------------------------------------------------===// 1590 1591 /// Get the reciprocal estimate attribute string for a function that will 1592 /// override the target defaults. 1593 static StringRef getRecipEstimateForFunc(MachineFunction &MF) { 1594 const Function *F = MF.getFunction(); 1595 return F->getFnAttribute("reciprocal-estimates").getValueAsString(); 1596 } 1597 1598 /// Construct a string for the given reciprocal operation of the given type. 1599 /// This string should match the corresponding option to the front-end's 1600 /// "-mrecip" flag assuming those strings have been passed through in an 1601 /// attribute string. For example, "vec-divf" for a division of a vXf32. 1602 static std::string getReciprocalOpName(bool IsSqrt, EVT VT) { 1603 std::string Name = VT.isVector() ? "vec-" : ""; 1604 1605 Name += IsSqrt ? "sqrt" : "div"; 1606 1607 // TODO: Handle "half" or other float types? 1608 if (VT.getScalarType() == MVT::f64) { 1609 Name += "d"; 1610 } else { 1611 assert(VT.getScalarType() == MVT::f32 && 1612 "Unexpected FP type for reciprocal estimate"); 1613 Name += "f"; 1614 } 1615 1616 return Name; 1617 } 1618 1619 /// Return the character position and value (a single numeric character) of a 1620 /// customized refinement operation in the input string if it exists. Return 1621 /// false if there is no customized refinement step count. 1622 static bool parseRefinementStep(StringRef In, size_t &Position, 1623 uint8_t &Value) { 1624 const char RefStepToken = ':'; 1625 Position = In.find(RefStepToken); 1626 if (Position == StringRef::npos) 1627 return false; 1628 1629 StringRef RefStepString = In.substr(Position + 1); 1630 // Allow exactly one numeric character for the additional refinement 1631 // step parameter. 1632 if (RefStepString.size() == 1) { 1633 char RefStepChar = RefStepString[0]; 1634 if (RefStepChar >= '0' && RefStepChar <= '9') { 1635 Value = RefStepChar - '0'; 1636 return true; 1637 } 1638 } 1639 report_fatal_error("Invalid refinement step for -recip."); 1640 } 1641 1642 /// For the input attribute string, return one of the ReciprocalEstimate enum 1643 /// status values (enabled, disabled, or not specified) for this operation on 1644 /// the specified data type. 1645 static int getOpEnabled(bool IsSqrt, EVT VT, StringRef Override) { 1646 if (Override.empty()) 1647 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1648 1649 SmallVector<StringRef, 4> OverrideVector; 1650 SplitString(Override, OverrideVector, ","); 1651 unsigned NumArgs = OverrideVector.size(); 1652 1653 // Check if "all", "none", or "default" was specified. 1654 if (NumArgs == 1) { 1655 // Look for an optional setting of the number of refinement steps needed 1656 // for this type of reciprocal operation. 1657 size_t RefPos; 1658 uint8_t RefSteps; 1659 if (parseRefinementStep(Override, RefPos, RefSteps)) { 1660 // Split the string for further processing. 1661 Override = Override.substr(0, RefPos); 1662 } 1663 1664 // All reciprocal types are enabled. 1665 if (Override == "all") 1666 return TargetLoweringBase::ReciprocalEstimate::Enabled; 1667 1668 // All reciprocal types are disabled. 1669 if (Override == "none") 1670 return TargetLoweringBase::ReciprocalEstimate::Disabled; 1671 1672 // Target defaults for enablement are used. 1673 if (Override == "default") 1674 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1675 } 1676 1677 // The attribute string may omit the size suffix ('f'/'d'). 1678 std::string VTName = getReciprocalOpName(IsSqrt, VT); 1679 std::string VTNameNoSize = VTName; 1680 VTNameNoSize.pop_back(); 1681 static const char DisabledPrefix = '!'; 1682 1683 for (StringRef RecipType : OverrideVector) { 1684 size_t RefPos; 1685 uint8_t RefSteps; 1686 if (parseRefinementStep(RecipType, RefPos, RefSteps)) 1687 RecipType = RecipType.substr(0, RefPos); 1688 1689 // Ignore the disablement token for string matching. 1690 bool IsDisabled = RecipType[0] == DisabledPrefix; 1691 if (IsDisabled) 1692 RecipType = RecipType.substr(1); 1693 1694 if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize)) 1695 return IsDisabled ? TargetLoweringBase::ReciprocalEstimate::Disabled 1696 : TargetLoweringBase::ReciprocalEstimate::Enabled; 1697 } 1698 1699 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1700 } 1701 1702 /// For the input attribute string, return the customized refinement step count 1703 /// for this operation on the specified data type. If the step count does not 1704 /// exist, return the ReciprocalEstimate enum value for unspecified. 1705 static int getOpRefinementSteps(bool IsSqrt, EVT VT, StringRef Override) { 1706 if (Override.empty()) 1707 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1708 1709 SmallVector<StringRef, 4> OverrideVector; 1710 SplitString(Override, OverrideVector, ","); 1711 unsigned NumArgs = OverrideVector.size(); 1712 1713 // Check if "all", "default", or "none" was specified. 1714 if (NumArgs == 1) { 1715 // Look for an optional setting of the number of refinement steps needed 1716 // for this type of reciprocal operation. 1717 size_t RefPos; 1718 uint8_t RefSteps; 1719 if (!parseRefinementStep(Override, RefPos, RefSteps)) 1720 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1721 1722 // Split the string for further processing. 1723 Override = Override.substr(0, RefPos); 1724 assert(Override != "none" && 1725 "Disabled reciprocals, but specifed refinement steps?"); 1726 1727 // If this is a general override, return the specified number of steps. 1728 if (Override == "all" || Override == "default") 1729 return RefSteps; 1730 } 1731 1732 // The attribute string may omit the size suffix ('f'/'d'). 1733 std::string VTName = getReciprocalOpName(IsSqrt, VT); 1734 std::string VTNameNoSize = VTName; 1735 VTNameNoSize.pop_back(); 1736 1737 for (StringRef RecipType : OverrideVector) { 1738 size_t RefPos; 1739 uint8_t RefSteps; 1740 if (!parseRefinementStep(RecipType, RefPos, RefSteps)) 1741 continue; 1742 1743 RecipType = RecipType.substr(0, RefPos); 1744 if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize)) 1745 return RefSteps; 1746 } 1747 1748 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1749 } 1750 1751 int TargetLoweringBase::getRecipEstimateSqrtEnabled(EVT VT, 1752 MachineFunction &MF) const { 1753 return getOpEnabled(true, VT, getRecipEstimateForFunc(MF)); 1754 } 1755 1756 int TargetLoweringBase::getRecipEstimateDivEnabled(EVT VT, 1757 MachineFunction &MF) const { 1758 return getOpEnabled(false, VT, getRecipEstimateForFunc(MF)); 1759 } 1760 1761 int TargetLoweringBase::getSqrtRefinementSteps(EVT VT, 1762 MachineFunction &MF) const { 1763 return getOpRefinementSteps(true, VT, getRecipEstimateForFunc(MF)); 1764 } 1765 1766 int TargetLoweringBase::getDivRefinementSteps(EVT VT, 1767 MachineFunction &MF) const { 1768 return getOpRefinementSteps(false, VT, getRecipEstimateForFunc(MF)); 1769 } 1770 1771 void TargetLoweringBase::finalizeLowering(MachineFunction &MF) const { 1772 MF.getRegInfo().freezeReservedRegs(MF); 1773 } 1774